diff --git a/.gitattributes b/.gitattributes index 5be54c12713972d4233f5cfe068dfad82aa9af46..9669dd6619c004574f0d34a3c1894ee177b11059 100644 --- a/.gitattributes +++ b/.gitattributes @@ -1589,3 +1589,67 @@ saved_model/**/* filter=lfs diff=lfs merge=lfs -text 2024/PALM_[[:space:]]Predicting[[:space:]]Actions[[:space:]]through[[:space:]]Language[[:space:]]Models/26a4ce29-0668-4f26-8004-06463aacfbd9_origin.pdf filter=lfs diff=lfs merge=lfs -text 2024/PARE-Net_[[:space:]]Position-Aware[[:space:]]Rotation-Equivariant[[:space:]]Networks[[:space:]]for[[:space:]]Robust[[:space:]]Point[[:space:]]Cloud[[:space:]]Registration/bb52733c-5c36-49b5-8e14-12ac007a5cf4_origin.pdf filter=lfs diff=lfs merge=lfs -text 2024/PARIS3D_[[:space:]]Reasoning-based[[:space:]]3D[[:space:]]Part[[:space:]]Segmentation[[:space:]]Using[[:space:]]Large[[:space:]]Multimodal[[:space:]]Model/ee219343-d433-449b-a58a-eb679a19a121_origin.pdf filter=lfs diff=lfs merge=lfs -text +2024/PAV_[[:space:]]Personalized[[:space:]]Head[[:space:]]Avatar[[:space:]]from[[:space:]]Unstructured[[:space:]]Video[[:space:]]Collection/75c53aba-b5dd-481e-8738-bfaa232bbce0_origin.pdf filter=lfs diff=lfs merge=lfs -text +2024/PCF-Lift_[[:space:]]Panoptic[[:space:]]Lifting[[:space:]]by[[:space:]]Probabilistic[[:space:]]Contrastive[[:space:]]Fusion/43912c86-ca82-4896-afdb-6d1a9980542c_origin.pdf filter=lfs diff=lfs merge=lfs -text +2024/PDT[[:space:]]Uav[[:space:]]Target[[:space:]]Detection[[:space:]]Dataset[[:space:]]for[[:space:]]Pests[[:space:]]and[[:space:]]Diseases[[:space:]]Tree/77c7722b-38da-423b-b56a-11a16b09eb35_origin.pdf filter=lfs diff=lfs merge=lfs -text +2024/PDiscoFormer_[[:space:]]Relaxing[[:space:]]Part[[:space:]]Discovery[[:space:]]Constraints[[:space:]]with[[:space:]]Vision[[:space:]]Transformers/aa33e6f8-0640-410e-98c0-983fb97aea81_origin.pdf filter=lfs diff=lfs merge=lfs -text +2024/PEA-Diffusion_[[:space:]]Parameter-Efficient[[:space:]]Adapter[[:space:]]with[[:space:]]Knowledge[[:space:]]Distillation[[:space:]]in[[:space:]]non-English[[:space:]]Text-to-Image[[:space:]]Generation/e154c47c-6f75-47c7-a454-a9dcc00f12d4_origin.pdf filter=lfs diff=lfs merge=lfs -text +2024/PFGS_[[:space:]]High[[:space:]]Fidelity[[:space:]]Point[[:space:]]Cloud[[:space:]]Rendering[[:space:]]via[[:space:]]Feature[[:space:]]Splatting/685f2bf6-434e-4901-92a2-92656824eb88_origin.pdf filter=lfs diff=lfs merge=lfs -text +2024/PFedEdit_[[:space:]]Personalized[[:space:]]Federated[[:space:]]Learning[[:space:]]via[[:space:]]Automated[[:space:]]Model[[:space:]]Editing/984c6b27-188f-4dc6-b40f-18bc786fe91d_origin.pdf filter=lfs diff=lfs merge=lfs -text +2024/PILoRA_[[:space:]]Prototype[[:space:]]Guided[[:space:]]Incremental[[:space:]]LoRA[[:space:]]for[[:space:]]Federated[[:space:]]Class-Incremental[[:space:]]Learning/b10fddc1-55da-4bf4-bfbe-8037f4bdbf91_origin.pdf filter=lfs diff=lfs merge=lfs -text +2024/PISR_[[:space:]]Polarimetric[[:space:]]Neural[[:space:]]Implicit[[:space:]]Surface[[:space:]]Reconstruction[[:space:]]for[[:space:]]Textureless[[:space:]]and[[:space:]]Specular[[:space:]]Objects/949f096c-d4bf-4988-bb1a-9febc3a46def_origin.pdf filter=lfs diff=lfs merge=lfs -text +2024/PLOT_[[:space:]]Text-based[[:space:]]Person[[:space:]]Search[[:space:]]with[[:space:]]Part[[:space:]]Slot[[:space:]]Attention[[:space:]]for[[:space:]]Corresponding[[:space:]]Part[[:space:]]Discovery/aafb2e8f-8788-410d-8a77-952604d06d27_origin.pdf filter=lfs diff=lfs merge=lfs -text +2024/PMT_[[:space:]]Progressive[[:space:]]Mean[[:space:]]Teacher[[:space:]]via[[:space:]]Exploring[[:space:]]Temporal[[:space:]]Consistency[[:space:]]for[[:space:]]Semi-Supervised[[:space:]]Medical[[:space:]]Image[[:space:]]Segmentation/049d7a98-0307-4aa9-ae68-729601a51f9f_origin.pdf filter=lfs diff=lfs merge=lfs -text +2024/POA_[[:space:]]Pre-training[[:space:]]Once[[:space:]]for[[:space:]]Models[[:space:]]of[[:space:]]All[[:space:]]Sizes/9b5de88d-0728-41f9-a801-73fd67e5b8a0_origin.pdf filter=lfs diff=lfs merge=lfs -text +2024/POCA_[[:space:]]Post-training[[:space:]]Quantization[[:space:]]with[[:space:]]Temporal[[:space:]]Alignment[[:space:]]for[[:space:]]Codec[[:space:]]Avatars/f5688d31-6662-4aa8-b308-c286c399e498_origin.pdf filter=lfs diff=lfs merge=lfs -text +2024/POET_[[:space:]]Prompt[[:space:]]Offset[[:space:]]Tuning[[:space:]]for[[:space:]]Continual[[:space:]]Human[[:space:]]Action[[:space:]]Adaptation/9942db11-78b9-4c6d-815f-c44748e2ea7c_origin.pdf filter=lfs diff=lfs merge=lfs -text +2024/PPAD_[[:space:]]Iterative[[:space:]]Interactions[[:space:]]of[[:space:]]Prediction[[:space:]]and[[:space:]]Planning[[:space:]]for[[:space:]]End-to-end[[:space:]]Autonomous[[:space:]]Driving/2548ffc2-3e35-4690-aa4a-52b1b19d7a46_origin.pdf filter=lfs diff=lfs merge=lfs -text +2024/PQ-SAM_[[:space:]]Post-training[[:space:]]Quantization[[:space:]]for[[:space:]]Segment[[:space:]]Anything[[:space:]]Model/5e412ebd-1f25-44d2-98a8-cb85fe86acb8_origin.pdf filter=lfs diff=lfs merge=lfs -text +2024/PRET_[[:space:]]Planning[[:space:]]with[[:space:]]Directed[[:space:]]Fidelity[[:space:]]Trajectory[[:space:]]for[[:space:]]Vision[[:space:]]and[[:space:]]Language[[:space:]]Navigation/4e64d425-40e4-4880-bd30-e7c4e06b2cbc_origin.pdf filter=lfs diff=lfs merge=lfs -text +2024/PSALM_[[:space:]]Pixelwise[[:space:]]Segmentation[[:space:]]with[[:space:]]Large[[:space:]]Multi-modal[[:space:]]Model/6e437084-134d-4cf9-a26b-095cfa791f83_origin.pdf filter=lfs diff=lfs merge=lfs -text +2024/PYRA_[[:space:]]Parallel[[:space:]]Yielding[[:space:]]Re-Activation[[:space:]]for[[:space:]]Training-Inference[[:space:]]Efficient[[:space:]]Task[[:space:]]Adaptation/a96361d9-106b-4325-b9d1-39f6bd731ebb_origin.pdf filter=lfs diff=lfs merge=lfs -text +2024/PaPr_[[:space:]]Training-Free[[:space:]]One-Step[[:space:]]Patch[[:space:]]Pruning[[:space:]]with[[:space:]]Lightweight[[:space:]]ConvNets[[:space:]]for[[:space:]]Faster[[:space:]]Inference/f59c8716-3e2f-464a-8e1c-4f54e686a19f_origin.pdf filter=lfs diff=lfs merge=lfs -text +2024/PairingNet_[[:space:]]A[[:space:]]Learning-based[[:space:]]Pair-searching[[:space:]]and[[:space:]]-matching[[:space:]]Network[[:space:]]for[[:space:]]Image[[:space:]]Fragments/3b568531-7960-492d-9c00-7cccb8dd34e1_origin.pdf filter=lfs diff=lfs merge=lfs -text +2024/Pairwise[[:space:]]Distance[[:space:]]Distillation[[:space:]]for[[:space:]]Unsupervised[[:space:]]Real-World[[:space:]]Image[[:space:]]Super-Resolution/821e13b2-5f5a-4742-b51d-28c69e766f13_origin.pdf filter=lfs diff=lfs merge=lfs -text +2024/PanGu-Draw_[[:space:]]Advancing[[:space:]]Resource-Efficient[[:space:]]Text-to-Image[[:space:]]Synthesis[[:space:]]with[[:space:]]Time-Decoupled[[:space:]]Training[[:space:]]and[[:space:]]Reusable[[:space:]]Coop-Diffusion/c75d5546-18a8-4ad8-982c-ff094dd050a6_origin.pdf filter=lfs diff=lfs merge=lfs -text +2024/Panel-Specific[[:space:]]Degradation[[:space:]]Representation[[:space:]]for[[:space:]]Raw[[:space:]]Under-Display[[:space:]]Camera[[:space:]]Image[[:space:]]Restoration/c089de88-5e00-46c3-9141-1e16159753ca_origin.pdf filter=lfs diff=lfs merge=lfs -text +2024/PanoFree_[[:space:]]Tuning-Free[[:space:]]Holistic[[:space:]]Multi-view[[:space:]]Image[[:space:]]Generation[[:space:]]with[[:space:]]Cross-view[[:space:]]Self-Guidance/cbd590c4-f320-4c08-83b4-3c8681128a51_origin.pdf filter=lfs diff=lfs merge=lfs -text +2024/PanoVOS_[[:space:]]Bridging[[:space:]]Non-panoramic[[:space:]]and[[:space:]]Panoramic[[:space:]]Views[[:space:]]with[[:space:]]Transformer[[:space:]]for[[:space:]]Video[[:space:]]Segmentation/5c945af1-cf63-4f10-9827-c4fd88e7d8a2_origin.pdf filter=lfs diff=lfs merge=lfs -text +2024/PapMOT_[[:space:]]Exploring[[:space:]]Adversarial[[:space:]]Patch[[:space:]]Attack[[:space:]]against[[:space:]]Multiple[[:space:]]Object[[:space:]]Tracking/e6b4ef0f-e80b-4da1-8f00-4210586c7699_origin.pdf filter=lfs diff=lfs merge=lfs -text +2024/ParCo_[[:space:]]Part-Coordinating[[:space:]]Text-to-Motion[[:space:]]Synthesis/5d0214df-d42e-4bdf-a60a-2b013c23d49a_origin.pdf filter=lfs diff=lfs merge=lfs -text +2024/Parameter-Efficient[[:space:]]and[[:space:]]Memory-Efficient[[:space:]]Tuning[[:space:]]for[[:space:]]Vision[[:space:]]Transformer_[[:space:]]A[[:space:]]Disentangled[[:space:]]Approach/e724c6a7-eae9-4271-a988-9b07cf4a65fc_origin.pdf filter=lfs diff=lfs merge=lfs -text +2024/Parameterization-driven[[:space:]]Neural[[:space:]]Surface[[:space:]]Reconstruction[[:space:]]for[[:space:]]Object-oriented[[:space:]]Editing[[:space:]]in[[:space:]]Neural[[:space:]]Rendering/20a75bf0-0a6a-4ea4-b1e9-a52604587383_origin.pdf filter=lfs diff=lfs merge=lfs -text +2024/Parameterized[[:space:]]Quasi-Physical[[:space:]]Simulators[[:space:]]for[[:space:]]Dexterous[[:space:]]Manipulations[[:space:]]Transfer/ae282149-332d-459a-aebe-a693d1ac89fd_origin.pdf filter=lfs diff=lfs merge=lfs -text +2024/Parrot[[:space:]]Captions[[:space:]]Teach[[:space:]]CLIP[[:space:]]to[[:space:]]Spot[[:space:]]Text/b85a274a-d711-419a-aeba-49fd9715a806_origin.pdf filter=lfs diff=lfs merge=lfs -text +2024/Parrot_[[:space:]]Pareto-optimal[[:space:]]Multi-Reward[[:space:]]Reinforcement[[:space:]]Learning[[:space:]]Framework[[:space:]]for[[:space:]]Text-to-Image[[:space:]]Generation/938566d1-b778-4bb0-b908-cbc4a31c6445_origin.pdf filter=lfs diff=lfs merge=lfs -text +2024/Part2Object_[[:space:]]Hierarchical[[:space:]]Unsupervised[[:space:]]3D[[:space:]]Instance[[:space:]]Segmentation/918bb93b-c2e7-4e11-b966-4fcc81659f0e_origin.pdf filter=lfs diff=lfs merge=lfs -text +2024/PartCraft_[[:space:]]Crafting[[:space:]]Creative[[:space:]]Objects[[:space:]]by[[:space:]]Parts/4747571e-3f49-4aac-b94c-924e39815a4f_origin.pdf filter=lfs diff=lfs merge=lfs -text +2024/PartGLEE_[[:space:]]A[[:space:]]Foundation[[:space:]]Model[[:space:]]for[[:space:]]Recognizing[[:space:]]and[[:space:]]Parsing[[:space:]]Any[[:space:]]Objects/a546ffc1-52e4-4141-be35-2478ed6c5802_origin.pdf filter=lfs diff=lfs merge=lfs -text +2024/PartImageNet++[[:space:]]Dataset_[[:space:]]Scaling[[:space:]]up[[:space:]]Part-based[[:space:]]Models[[:space:]]for[[:space:]]Robust[[:space:]]Recognition/0d30f97b-2cbe-4f46-92c7-0bf8c016e1e3_origin.pdf filter=lfs diff=lfs merge=lfs -text +2024/PartSTAD_[[:space:]]2D-to-3D[[:space:]]Part[[:space:]]Segmentation[[:space:]]Task[[:space:]]Adaptation/bbdaa935-8663-4268-b38a-384a769ccb82_origin.pdf filter=lfs diff=lfs merge=lfs -text +2024/PatchRefiner_[[:space:]]Leveraging[[:space:]]Synthetic[[:space:]]Data[[:space:]]for[[:space:]]Real-Domain[[:space:]]High-Resolution[[:space:]]Monocular[[:space:]]Metric[[:space:]]Depth[[:space:]]Estimation/3be463cc-22b0-43d6-8e8c-f91c9fa34d5b_origin.pdf filter=lfs diff=lfs merge=lfs -text +2024/PathMMU_[[:space:]]A[[:space:]]Massive[[:space:]]Multimodal[[:space:]]Expert-Level[[:space:]]Benchmark[[:space:]]for[[:space:]]Understanding[[:space:]]and[[:space:]]Reasoning[[:space:]]in[[:space:]]Pathology/0e28fdb1-202a-41b0-8b8d-6b1d1b08e3fe_origin.pdf filter=lfs diff=lfs merge=lfs -text +2024/Pathformer3D_[[:space:]]A[[:space:]]3D[[:space:]]Scanpath[[:space:]]Transformer[[:space:]]for[[:space:]]360°[[:space:]]Images/8acb38cc-358d-422c-8280-ca63ea0a1521_origin.pdf filter=lfs diff=lfs merge=lfs -text +2024/Pathology-knowledge[[:space:]]Enhanced[[:space:]]Multi-instance[[:space:]]Prompt[[:space:]]Learning[[:space:]]for[[:space:]]Few-shot[[:space:]]Whole[[:space:]]Slide[[:space:]]Image[[:space:]]Classification/3e473ac8-930f-46c0-b97b-e8e52a7c0bfd_origin.pdf filter=lfs diff=lfs merge=lfs -text +2024/Paying[[:space:]]More[[:space:]]Attention[[:space:]]to[[:space:]]Images_[[:space:]]A[[:space:]]Training-Free[[:space:]]Method[[:space:]]for[[:space:]]Alleviating[[:space:]]Hallucination[[:space:]]in[[:space:]]LVLMs/71543d84-0242-4633-b85d-0ad4feeefd05_origin.pdf filter=lfs diff=lfs merge=lfs -text +2024/Per-Gaussian[[:space:]]Embedding-Based[[:space:]]Deformation[[:space:]]for[[:space:]]Deformable[[:space:]]3D[[:space:]]Gaussian[[:space:]]Splatting/8af14cc6-a852-4c45-9f34-dffb20b17a0a_origin.pdf filter=lfs diff=lfs merge=lfs -text +2024/Perceptual[[:space:]]Evaluation[[:space:]]of[[:space:]]Audio-Visual[[:space:]]Synchrony[[:space:]]Grounded[[:space:]]in[[:space:]]Viewers’[[:space:]]Opinion[[:space:]]Scores/59607ef8-05c8-452e-b964-a6375c1e0fbc_origin.pdf filter=lfs diff=lfs merge=lfs -text +2024/Personalized[[:space:]]Federated[[:space:]]Domain-Incremental[[:space:]]Learning[[:space:]]based[[:space:]]on[[:space:]]Adaptive[[:space:]]Knowledge[[:space:]]Matching/ea77b70c-2acd-46a7-86ea-e1ba20b951f3_origin.pdf filter=lfs diff=lfs merge=lfs -text +2024/Personalized[[:space:]]Privacy[[:space:]]Protection[[:space:]]Mask[[:space:]]Against[[:space:]]Unauthorized[[:space:]]Facial[[:space:]]Recognition/40580728-0a4c-4c74-a026-5407ddf870b6_origin.pdf filter=lfs diff=lfs merge=lfs -text +2024/Personalized[[:space:]]Video[[:space:]]Relighting[[:space:]]With[[:space:]]an[[:space:]]At-Home[[:space:]]Light[[:space:]]Stage/0751f00d-698a-4afe-b1e5-471a9991f11a_origin.pdf filter=lfs diff=lfs merge=lfs -text +2024/PetFace_[[:space:]]A[[:space:]]Large-Scale[[:space:]]Dataset[[:space:]]and[[:space:]]Benchmark[[:space:]]for[[:space:]]Animal[[:space:]]Identification/79b48743-23c5-4624-b76c-ce49a285a1a7_origin.pdf filter=lfs diff=lfs merge=lfs -text +2024/Phase[[:space:]]Concentration[[:space:]]and[[:space:]]Shortcut[[:space:]]Suppression[[:space:]]for[[:space:]]Weakly[[:space:]]Supervised[[:space:]]Semantic[[:space:]]Segmentation/8ce12d89-9c27-46eb-bd86-41483594c43e_origin.pdf filter=lfs diff=lfs merge=lfs -text +2024/Photon[[:space:]]Inhibition[[:space:]]for[[:space:]]Energy-Efficient[[:space:]]Single-Photon[[:space:]]Imaging/03673fae-a9a6-4975-8cf4-3d5bc33fd48c_origin.pdf filter=lfs diff=lfs merge=lfs -text +2024/Photorealistic[[:space:]]Object[[:space:]]Insertion[[:space:]]with[[:space:]]Diffusion-Guided[[:space:]]Inverse[[:space:]]Rendering/27eaedd8-53d6-44df-a63a-fb92723dff8a_origin.pdf filter=lfs diff=lfs merge=lfs -text +2024/Photorealistic[[:space:]]Video[[:space:]]Generation[[:space:]]with[[:space:]]Diffusion[[:space:]]Models/327f0bbf-9985-4862-ae62-9cd9ec84c015_origin.pdf filter=lfs diff=lfs merge=lfs -text +2024/PhysAvatar_[[:space:]]Learning[[:space:]]the[[:space:]]Physics[[:space:]]of[[:space:]]Dressed[[:space:]]3D[[:space:]]Avatars[[:space:]]from[[:space:]]Visual[[:space:]]Observations/958c8811-3fcb-45f8-9957-bab313d0628e_origin.pdf filter=lfs diff=lfs merge=lfs -text +2024/PhysGen_[[:space:]]Rigid-Body[[:space:]]Physics-Grounded[[:space:]]Image-to-Video[[:space:]]Generation/63954133-cd49-423f-97fe-3962651db98c_origin.pdf filter=lfs diff=lfs merge=lfs -text +2024/Physical-Based[[:space:]]Event[[:space:]]Camera[[:space:]]Simulator/0986a08f-4884-431b-9c20-c64d0c63357b_origin.pdf filter=lfs diff=lfs merge=lfs -text +2024/Physically[[:space:]]Plausible[[:space:]]Color[[:space:]]Correction[[:space:]]for[[:space:]]Neural[[:space:]]Radiance[[:space:]]Fields/88cfdd54-68fa-4f55-bb61-e981b6d01cd5_origin.pdf filter=lfs diff=lfs merge=lfs -text +2024/Physics-Based[[:space:]]Interaction[[:space:]]with[[:space:]]3D[[:space:]]Objects[[:space:]]via[[:space:]]Video[[:space:]]Generation/017e5389-f4df-4a9c-a529-0be7433a235c_origin.pdf filter=lfs diff=lfs merge=lfs -text +2024/Physics-Free[[:space:]]Spectrally[[:space:]]Multiplexed[[:space:]]Photometric[[:space:]]Stereo[[:space:]]under[[:space:]]Unknown[[:space:]]Spectral[[:space:]]Composition/09ecf602-4895-4520-8961-a33ebbe3e42d_origin.pdf filter=lfs diff=lfs merge=lfs -text +2024/Physics-informed[[:space:]]Knowledge[[:space:]]Transfer[[:space:]]for[[:space:]]Underwater[[:space:]]Monocular[[:space:]]Depth[[:space:]]Estimation/58639476-9724-4b09-9db4-e436240a5f18_origin.pdf filter=lfs diff=lfs merge=lfs -text +2024/PiTe_[[:space:]]Pixel-Temporal[[:space:]]Alignment[[:space:]]for[[:space:]]Large[[:space:]]Video-Language[[:space:]]Model/3468332a-054d-4ee9-a3ba-3d30cf6d6594_origin.pdf filter=lfs diff=lfs merge=lfs -text +2024/Pick-a-back_[[:space:]]Selective[[:space:]]Device-to-Device[[:space:]]Knowledge[[:space:]]Transfer[[:space:]]in[[:space:]]Federated[[:space:]]Continual[[:space:]]Learning/ecda4e75-a956-4af4-a55c-23aa50728c18_origin.pdf filter=lfs diff=lfs merge=lfs -text +2024/Pix2Gif_[[:space:]]Motion-Guided[[:space:]]Diffusion[[:space:]]for[[:space:]]GIF[[:space:]]Generation/b48548a6-50ce-46d1-9de4-333ad980d37c_origin.pdf filter=lfs diff=lfs merge=lfs -text +2024/PixArt-Sigma_[[:space:]]Weak-to-Strong[[:space:]]Training[[:space:]]of[[:space:]]Diffusion[[:space:]]Transformer[[:space:]]for[[:space:]]4K[[:space:]]Text-to-Image[[:space:]]Generation/56a1c2d9-7461-45b0-bc11-505740d1999b_origin.pdf filter=lfs diff=lfs merge=lfs -text diff --git a/2024/PAV_ Personalized Head Avatar from Unstructured Video Collection/75c53aba-b5dd-481e-8738-bfaa232bbce0_content_list.json b/2024/PAV_ Personalized Head Avatar from Unstructured Video Collection/75c53aba-b5dd-481e-8738-bfaa232bbce0_content_list.json new file mode 100644 index 0000000000000000000000000000000000000000..0985d62e626d6b8482a5d330f05b548a23f8c9c2 --- /dev/null +++ b/2024/PAV_ Personalized Head Avatar from Unstructured Video Collection/75c53aba-b5dd-481e-8738-bfaa232bbce0_content_list.json @@ -0,0 +1,1474 @@ +[ + { + "type": "text", + "text": "PAV: Personalized Head Avatar from Unstructured Video Collection", + "text_level": 1, + "bbox": [ + 289, + 140, + 712, + 183 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Akin Caliskan1, Berkay Kicanaoglu1, and Hyeongwoo Kim2", + "bbox": [ + 269, + 210, + 732, + 227 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "1 Flawless AI", + "bbox": [ + 455, + 239, + 549, + 252 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "2 Imperial College London", + "bbox": [ + 411, + 253, + 591, + 267 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Abstract. We propose PAV, Personalized Head Avatar for the synthesis of human faces under arbitrary viewpoints and facial expressions. PAV introduces a method that learns a dynamic deformable neural radiance field (NeRF), in particular from a collection of monocular talking face videos of the same character under various appearance and shape changes. Unlike existing head NeRF methods that are limited to modeling such input videos on a per-appearance basis, our method allows for learning multi-appearance NeRFs, introducing appearance embedding for each input video via learnable latent neural features attached to the underlying geometry. Furthermore, the proposed appearance-conditioned density formulation facilitates the shape variation of the character, such as facial hair and soft tissues, in the radiance field prediction. To the best of our knowledge, our approach is the first dynamic deformable NeRF framework to model appearance and shape variations in a single unified network for multi-appearances of the same subject. We demonstrate experimentally that PAV outperforms the baseline method in terms of visual rendering quality in our quantitative and qualitative studies on various subjects.", + "bbox": [ + 261, + 301, + 738, + 551 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "1 Introduction", + "text_level": 1, + "bbox": [ + 217, + 598, + 374, + 614 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Digital human avatars enabling the editing of facial expressions and head motion have broad applications across telepresence, animation, and digital content creation. The demands for personalized human avatars such as digital doubles have also increased over the past few years not only in the film industry but in the research community. To enable the widespread adaptation of these neural avatars, they should be easy to generate and animate under novel poses and facial expressions. A vast majority of existing methods aim to learn neural avatars from monocular videos to make it more accessible and easy to use [1,10,12-14,16,34,42,44]. Early approaches [12,13,16,34] mainly rely on 3D morphable models (3DMM) to guide the modeling process and to provide an intuitive control mechanism in novel view synthesis. Despite their success, they do not model the 3D face geometry. To address this issue, neural radiance fields (NeRFs) [23] have demonstrated volumetric 3D modeling of face geometry and appearance [1,10,14,42,44] under novel poses and expressions. While this allows", + "bbox": [ + 212, + 628, + 787, + 843 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/8c0e0140fb6c0cdb30c0c9fe988d7f54fe92f483434bea1be89f987b0d6d9713.jpg", + "image_caption": [ + "Fig. 1: We present an approach for learning personalized head avatars from an unstructured video collection of the same subject. The videos are taken in different environments, with diverse appearances and possible shape changes of a character, as shown on the left. Our method leverages all of the input clips into a single unified dynamic deformable NeRF framework and offers high-quality video replays in novel head poses and facial expressions. On the right, we show some rendering results of each input appearance of the same subject aligned in head pose and facial expression that are not seen during training. Notably, our method can synthesize identity-specific details while showing coherent expressions across different appearances." + ], + "image_footnote": [], + "bbox": [ + 253, + 152, + 750, + 378 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "photo-realistic image quality and animation, NeRF-based approaches require per-appearance optimization, maintaining the consistent appearance of the subject, which is constrained to the pre-recording of characters in controlled settings. However, this limits the potential for practical application scenarios where we want to generate a 3D head avatar from a collection of prerecorded video clips of a subject. In this work, we present the Personalized Head Avatar method transforming an unstructured personal video collection, containing images spanning multiple years with different appearances, head poses, and expressions, into a 3D representation of the subject. To create an avatar, we only require an unstructured collection of video clips of the person in arbitrary poses, expressions, and appearances. Our novel framework enables creation of avatars from sparse observations, while yielding high reconstruction quality in novel combinations of pose and appearance, and opens up new applications.", + "bbox": [ + 212, + 550, + 787, + 747 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Learning dynamic deformable head NeRF from unstructured video clips is a challenging problem. First, video collections can contain frames of the same identity at different appearances, poses, and expressions. Second, we only have limited pose and expression variation for each appearance, so it is unlikely that all poses and expressions of the face would be well-observed for any given appearance. Given the data, existing NeRF approaches' appearance-specific nature", + "bbox": [ + 212, + 750, + 787, + 840 + ], + "page_idx": 1 + }, + { + "type": "page_number", + "text": "2", + "bbox": [ + 217, + 114, + 228, + 126 + ], + "page_idx": 1 + }, + { + "type": "header", + "text": "A. Caliskan et al.", + "bbox": [ + 271, + 114, + 388, + 127 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "largely limits their potential for practical application scenarios. They require an independent model that is trained from scratch for every appearance from each monocular video. To address these challenges, we model shared deformation field and canonical space neural volumetric representation for all appearances with appearance-dependent latent features. Our key insight is that although observed faces have different appearances across video clips, they should all be explained by a common deformation model with small differences since they all originate from the same person. Furthermore, we assert that every video clip specific to an appearance contains both general and unique information. A single unified network could collaboratively learn the generic information from various appearances, thus this benefits rendering performance as well. The single network is designed to learn a deformation-aware dynamic head NeRF while disentangling appearance-related details from those that are not appearance-specific. For appearance conditioning, we propose employing learnable appearance embedding attached to vertices of the tracked mesh and passing them to the dynamic head NeRF network. By adopting this process, PAV is able to learn discriminative latent neural features that are specific for each appearance, resulting in better rendering of appearance-specific details for arbitrary poses and expressions. Unlike existing NeRF methods [1, 10, 14, 42, 44], PAV simultaneously trains on all appearances of the same subject, eliminating the need for separate training sessions for each appearance and reducing the overall training duration. Learning of personalized head avatars in a single unified network from an unstructured collection of videos has not been addressed before and is a non-trivial extension of existing monocular head avatar generation methods since they can not handle multiple appearances in a single network. PAV makes this possible through a carefully designed appearance conditioning representation, geometry-attached trainable latent neural texture to learn appearance-specific features, appearance conditioned implicit functions to predict variations in color and density. We conducted a variety of experiments to show these improvements over the baseline approach. Since there is no such public dataset to evaluate our method, we curate our own in-the-wild video collections for multiple subjects. Overall, we present Personalized Head Avatar with the following contributions:", + "bbox": [ + 212, + 145, + 787, + 631 + ], + "page_idx": 2 + }, + { + "type": "list", + "sub_type": "text", + "list_items": [ + "- We present, PAV, a framework for learning controllable head avatars from unconstrained short video collections where the same person shows different appearances and facial expressions. To the best of our knowledge, this is the first method learning appearance controllable head avatars from in-the-wild monocular video collections of the same subject.", + "- The proposed method can synthesise images of head avatar under target appearance, head pose and facial expressions.", + "- PAV learns modeling local changes of geometry and color across different appearances of the same identity." + ], + "bbox": [ + 225, + 686, + 787, + 840 + ], + "page_idx": 2 + }, + { + "type": "header", + "text": "PAV: Personalized Head Avatar from Video Collection", + "bbox": [ + 369, + 114, + 730, + 127 + ], + "page_idx": 2 + }, + { + "type": "page_number", + "text": "3", + "bbox": [ + 774, + 116, + 784, + 126 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "2 Related Work", + "text_level": 1, + "bbox": [ + 215, + 143, + 387, + 160 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Monocular Neural Head Avatars. More recent methods learn neural representations of the person from monocular videos and use neural renderers [31, 32] to directly generate photo-realistic images in the target head pose and facial expression [1-3, 5, 10, 11, 14, 16, 21, 29, 33, 40, 42, 44]. These methods are generally classified into image-based or 3D-based neural rendering methods [31, 32]. The image-based models synthesize the face of a subject without relying on any representation of 3D space. Existing image-based approaches either employ learned warping fields [2, 29] to deform an input image or encoder-decoder architecture where decoder synthesize the output image [21, 25, 35, 39]. Even though these methods produce high-quality results, they suffer from artifacts for strong head pose and expression changes due to the lack of 3D geometry. The 3D-based neural rendering methods, on the other hand, rely on 3D Morphable Models (3DMM) [9]. These methods are generally classified into explicit and implicit models. Explicit models utilize 2D neural rendering methods [31, 32] and rely on neural networks which can be conditioned on coarse RGB renderings based on a linear texture model [16], uv-maps [4] or latent feature maps [33]. Although these techniques produce high-quality renderings, they are confined to operating within a 2D image space and are limited to craniofacial structures. To tackle this, implicit neural representations for modeling 3D faces have recently become a major focus of attention. In particular, NeRFs [23] shows photorealistic face image synthesis in novel poses and expressions [1, 10, 14, 42, 44]. Implicit models represent the geometry using implicit surface functions or volumetric representations and they are optimized via volumetric rendering [23]. Early approaches [23] and follow-up works synthesize novel views of complex static scenes using differentiable volumetric rendering. After introduction of NeRF for static scenes, a natural research direction was to generalize it to dynamic and time-varying ones [26, 28]. The reconstruction problem is divided into two different spaces, the deformed space and the canonical space. For human body, a series of approaches have been proposed to target this challenge [20, 27]. In particular, these methods are differentiated from each other based on deformation field modeling. Gafni et al. [10] introduced implicit deformation modeling conditioned on facial expression and head pose. In contrast to modeling the deformation field implicitly, IMAvatar [42], RigNerf [1] and Insta [44] propose explicit modeling of the deformation field to improve control and rendering quality. Our method falls into this paradigm as we build our system on top of dynamic deformable head NeRF. In particular, we take inspiration from Insta [44] for designing our framework. However, existing deformable dynamic head NeRF methods including Insta [44] require per-identity optimization, maintaining the consistent appearance of the same subject. To the best of our knowledge, PAV is the first method to learn a single unified dynamic deformable head NeRF from monocular video of multiple appearances of the same subject.", + "bbox": [ + 212, + 172, + 787, + 796 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Geometry-anchored Features for NeRFs. PAV introduces appearance-conditioned head NeRF by using geometry-attached features. By using these features, exist", + "bbox": [ + 215, + 809, + 790, + 840 + ], + "page_idx": 3 + }, + { + "type": "page_number", + "text": "4", + "bbox": [ + 217, + 114, + 228, + 126 + ], + "page_idx": 3 + }, + { + "type": "header", + "text": "A. Caliskan et al.", + "bbox": [ + 271, + 114, + 388, + 127 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "ing works show improved rendering quality for static objects [38] and human bodies [3, 20]. However, none of the existing methods [3, 20] has yet addressed multi-appearance learning for dynamic neural radiance fields.", + "bbox": [ + 212, + 146, + 782, + 191 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "NeRF from Unstructured Data Collections. Learning neural radiance field from unstructured data collections has recently gained much attention. Existing works have enabled photorealistic view synthesis of challenging static scenes, including tourist sites [22] and city-scale scenes [30], and static objects [18] from internet photo collections. For human subjects, Weng et al. [37] introduce free-viewpoint human body rendering from an unstructured photo collection. Despite the photorealistic rendering quality of PersonNerf [37], they do not model animatable faces due to the static training images. In our work, we learn animatable 3D head avatars from video collections.", + "bbox": [ + 212, + 205, + 784, + 340 + ], + "page_idx": 4 + }, + { + "type": "image", + "img_path": "images/198902175a488fa1839da38f7a9d3bdbf7958a8df09f4111f5c5b2903b36a25a.jpg", + "image_caption": [ + "Fig. 2: Overview of the proposed approach, PAV. (a) shows some training videos in various appearances of the same subject. (b) We estimate the 3D head geometry of the person using FLAME head model. (c) shows the proposed personalized deformable neural radiance field from video collection (d) We learn appearance-specific neural features and attach them to vertices. (d) This feature is passed to fully fused multi-layer perceptrons with additional conditioning on the facial expressions $e$ and the encoded view direction $d$ ." + ], + "image_footnote": [], + "bbox": [ + 218, + 366, + 784, + 523 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "3 Method", + "text_level": 1, + "bbox": [ + 215, + 686, + 328, + 702 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Our goal is to learn a personalized head avatar, PAV, from a monocular video collection of a subject under different appearances, poses, and facial expressions. To create the avatar, we use a small collection $\\mathbf{V} = \\{V_{i}\\}_{i = 1}^{N}$ with $N$ videos. We parameterize PAV using FLAME [19] face model to control head poses, facial expressions and use neural texture to model appearance. During inference, we only need the target FLAME parameters and neural texture. Learning PAV from a collection of monocular videos requires capturing the geometry and appearance of the dynamic head while also disentangling different appearances of the same", + "bbox": [ + 212, + 719, + 787, + 840 + ], + "page_idx": 4 + }, + { + "type": "header", + "text": "PAV: Personalized Head Avatar from Video Collection", + "bbox": [ + 367, + 114, + 730, + 127 + ], + "page_idx": 4 + }, + { + "type": "page_number", + "text": "5", + "bbox": [ + 774, + 116, + 784, + 126 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "identity. In order to tackle this problem, we first capture the dynamic 3D geometry by employing a parametric head model [19] (Sec. 3.1). We use the coarse geometry to model geometry-guided deformable neural radiance field (Sec. 3.2) and vertex-attached neural features to condition appearance for rendering (Sec. 3.3). We also propose estimating appearance-conditioned neural radiance field, aimed at accurately capturing distinct color and geometry with high fidelity. (Sec. 3.4). An overview of our method can be seen in Fig. 2. In the following, we describe each of these modules in greater detail.", + "bbox": [ + 212, + 146, + 787, + 268 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "3.1 Head Geometry Estimation", + "text_level": 1, + "bbox": [ + 214, + 287, + 488, + 303 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Given the training frames $\\mathbf{I} = \\{I_{\\mathrm{i}}\\}$ , from all appearances, we first estimate the optimized camera parameters $\\mathbf{K} \\in \\mathbb{R}^{3 \\times 3}$ , tracked FLAME [19] meshes $\\mathbf{M} = \\{M_{\\mathrm{i}}\\}$ with corresponding facial expression coefficients $\\mathbf{E} = \\{E_{\\mathrm{i}}\\}$ and poses $\\mathbf{P} = \\{P_{\\mathrm{i}}\\}$ . We follow an optimization-based method, that minimizes an objective function with photo-consistency and landmark terms [43]. Different from the original work [43] which is designed for a single appearance, in PAV, we compute the average shape parameters from all appearances of the same subject to represent identity shape during optimization.", + "bbox": [ + 212, + 309, + 787, + 431 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "3.2 Personalized Deformable NeRF", + "text_level": 1, + "bbox": [ + 214, + 450, + 522, + 465 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Our goal is to create a head avatar that can be learned from frames in video collection of various appearance of the same subject. We parameterize PAV using FLAME model and learned appearance conditioning. Specifically, PAV takes target head pose $\\theta$ , facial expression $e$ and latent neural features $(LNF)$ $Z_{j} \\in \\mathbb{R}^{512 \\times 512 \\times 16}$ in the texture atlas space (UV space) for an appearance $j$ as inputs and synthesizes the target image $I^{\\theta, e, Z_{j}}$ as:", + "bbox": [ + 212, + 472, + 787, + 564 + ], + "page_idx": 5 + }, + { + "type": "equation", + "text": "\n$$\nI ^ {\\theta , e, Z _ {j}} = \\operatorname {P A V} (\\theta , e, Z _ {j}, K), \\tag {1}\n$$\n", + "text_format": "latex", + "bbox": [ + 411, + 571, + 785, + 588 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "where $K$ corresponds to the intrinsic parameters of the target camera viewpoint. For this purpose, we are using a geometry-guided dynamic deformable neural radiance field. However, deformable head NeRF models cannot disentangle appearance from the neural radiance field (See Sec. 2). To tackle this problem, we employ shared canonical space for all appearances where a shared neural radiance field is constructed. With this shared radiance field at canonical space, to render the target head pose and facial expression using volumetric rendering, we canon- icalize the sample points $x$ on a ray from the observation space to query points $x_{c}$ in the shared canonical space radiance field by using deformation gradient $\\hat{F}$ .. $x_{c} = \\hat{F}\\cdot x$ where $\\hat{F}\\in \\mathbb{R}^{4\\times 4}$ and $\\{x,x_c\\} \\in \\mathbb{R}^4$ . We employ the nearest triangle search for a given point to compute the deformation gradient. Specifically, the deformation gradient is computed between the deformed triangle $\\hat{T}_{def}\\in M_i$ in the observed mesh and the canonical triangle $\\hat{T}_{canon}\\in M^{canon}$ in the canonical mesh. To increase the speed of nearest triangle search, we employ bounding volume hierarchy (BVH) [6]. We refer readers to [44] for a detailed description of deformation gradient computation.", + "bbox": [ + 212, + 595, + 787, + 840 + ], + "page_idx": 5 + }, + { + "type": "page_number", + "text": "6", + "bbox": [ + 217, + 114, + 228, + 126 + ], + "page_idx": 5 + }, + { + "type": "header", + "text": "A. Caliskan et al.", + "bbox": [ + 271, + 114, + 388, + 127 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "After canonicalization, given the predicted appearance-conditioned radiance $c$ and density $\\sigma$ (Sec. 3.4) for every point on each ray, the representation of the avatar is optimized by using the differentiable volumetric rendering. For each camera ray $r(t) = o + td$ with camera center $o$ and viewing direction $d$ , the color of the corresponding pixel can be computed by accumulating the predicting radiance and densities of the sample points along the ray:", + "bbox": [ + 212, + 146, + 782, + 237 + ], + "page_idx": 6 + }, + { + "type": "equation", + "text": "\n$$\n\\mathcal {C} (\\mathbf {r}) = \\int_ {t _ {n}} ^ {t _ {f}} \\sigma (\\mathbf {r} (t)) \\cdot \\mathbf {c} (\\mathbf {r} (t)) \\cdot \\mathcal {T} (t) d t \\tag {2}\n$$\n", + "text_format": "latex", + "bbox": [ + 375, + 250, + 785, + 282 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "where $t_n$ and $t_f$ are the near and far bounds correspondingly, and $\\mathcal{T}(t_n) = \\exp \\left(-\\int_{t_n}^{t_f}\\sigma (t)dt\\right)$ is the accumulated transmittance along the ray from $t_n$ to $t_f$ , $\\sigma (t)$ is the density and $c(t)$ it the radiance at position $x_t$ .", + "bbox": [ + 214, + 287, + 782, + 338 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "3.3 Learned Appearance Embedding", + "text_level": 1, + "bbox": [ + 214, + 359, + 532, + 376 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "In this section, PAV introduces a learnable appearance embedding that is essential and effective for enabling discriminative multi-appearance rendering using single unified network. To achieve this goal (See Fig. 2), a set of $N$ learnable latent features $\\mathbf{Z} = \\{Z_{\\mathrm{i}}\\}_{i = 1}^{N}$ , $Z_{i}\\in \\mathbb{R}^{512\\times 512\\times 16}$ are defined in UV space corresponding to $N$ appearances of the same subject in the video collection. We impose high-dimensional learnable latent features to condition implicit representations for both density $\\sigma$ and radiance $c$ values in the canonical space. To this end, these latent neural features are attached to the vertices of 3DMM mesh, which enables PAV to memorize local variations. For each point $x$ sampled on a ray, we find the nearest triangle $\\mathcal{T} = \\{v_{\\mathrm{i}}\\}_{i = 1}^{3}, v_{i}\\in \\mathbb{R}^{3}$ (sec. 3.2) and aggregate the features $z(v_{i})$ of vertices $v_{i}$ from the vertices of the nearest triangle to get the appearance embedding $Z^x$ . Then we directly use the appearance embedding to condition for appearance-dependent neural radiance field estimation (Sec. 3.4). During training, these features are randomly initialized and optimized jointly with the objectives of the entire network.", + "bbox": [ + 212, + 383, + 787, + 609 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "3.4 Appearance Conditioned NeRF", + "text_level": 1, + "bbox": [ + 214, + 632, + 522, + 648 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "To model the appearance-conditioned dynamic deformable head NeRF, PAV incorporates learned appearance-embedding (Sec. 3.3) into neural radiance field in canonical space (See Fig. 2). For a point $x$ sampled on a ray, after canonicalization, the point $x_{c}$ is encoded using multi-resolution hash encoding [24]. This feature $\\phi(x_{c})$ is passed to multi-layer perceptron for geometric feature estimation. Following [10] and [44], we conditioned every point features $\\phi(x_{c})$ with the facial expressions $e$ of frame $i$ . Without per-frame latent code, for shared volume density, we learn implicit representation $\\mathcal{F}_{\\text{density}}$ using multi-layer perceptron (MLP-D in Fig. 2):", + "bbox": [ + 212, + 657, + 787, + 792 + ], + "page_idx": 6 + }, + { + "type": "equation", + "text": "\n$$\n\\mathcal {F} _ {\\text {d e n s i t y}}: \\left(\\phi \\left(x _ {c}\\right), e _ {i}\\right)\\rightarrow \\sigma \\tag {3}\n$$\n", + "text_format": "latex", + "bbox": [ + 413, + 821, + 784, + 835 + ], + "page_idx": 6 + }, + { + "type": "header", + "text": "PAV: Personalized Head Avatar from Video Collection", + "bbox": [ + 367, + 114, + 730, + 127 + ], + "page_idx": 6 + }, + { + "type": "page_number", + "text": "7", + "bbox": [ + 774, + 116, + 784, + 126 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "A primary challenge lies in learning unique density estimations for each appearance while separating the shared density aspects from the appearance. Hence we design PAV such that appearance-based density offset is predicted from appearance embedding $Z^{x}$ sampled from the nearest triangle to the point $x$ on the mesh $M_{i}$ and facial expression $e_{i}$ of frame $i$ . For density offset, we learn implicit representation $\\mathcal{F}_{offset}$ using multi-layer perceptron (MLP-O in Fig. 2):", + "bbox": [ + 212, + 146, + 787, + 238 + ], + "page_idx": 7 + }, + { + "type": "equation", + "text": "\n$$\n\\mathcal {F} _ {\\text {o f f s e t}}: \\left(Z ^ {x}, e _ {i}\\right)\\rightarrow \\Delta_ {\\sigma} \\tag {4}\n$$\n", + "text_format": "latex", + "bbox": [ + 418, + 250, + 785, + 266 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Final density value $\\sigma(x) = \\sigma + \\Delta_{\\sigma}$ for the point $x$ is the summation of shared density features $\\sigma$ and density offset $\\Delta_{\\sigma}$ . For emitted radiance at point $x$ , we learn implicit representation $\\mathcal{F}_{color}$ using multi-layer perceptron (MLP-C in Fig. 2):", + "bbox": [ + 212, + 277, + 787, + 337 + ], + "page_idx": 7 + }, + { + "type": "equation", + "text": "\n$$\n\\mathcal {F} _ {\\text {c o l o r}}: \\left(Z ^ {x}, d _ {i}, \\sigma (x)\\right)\\rightarrow c \\tag {5}\n$$\n", + "text_format": "latex", + "bbox": [ + 410, + 349, + 785, + 367 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "where $d_{i}$ is the encoded view direction of training frame $i$ . Then, final color for the pixel is computed by using differentiable volume rendering (Eq. 2)", + "bbox": [ + 214, + 377, + 785, + 409 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "3.5 Optimization", + "text_level": 1, + "bbox": [ + 214, + 430, + 372, + 446 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Our model is trained on a collection of monocular RGB videos with photometric and depth losses. For photometric loss $\\mathcal{L}_{color}$ , we employ $L_{2}$ distance between the renderings $C$ and ground-truth images $I$ , $\\mathcal{L}_{color} = \\sum_{i}\\sum_{r}\\| C_{i}(r) - I_{i}(r)\\|$ , where $r$ denotes the camera ray of each pixel and $i$ denotes frame index. For depth loss $\\mathcal{L}_{depth}$ , we rasterize the depth of tracking mesh $M_{i}$ and apply $L_{1}$ distance between this map $D$ and the ray termination $H$ of the volumetric rendering only for face region as FLAME model is restricted to model details like hair: $\\mathcal{L}_{depth} = \\sum_{i}\\sum_{r}\\mathcal{M}_{i}(r)\\| H_{i}(r) - D_{i}(r)\\|$ , where $\\mathcal{M}$ denotes face binary mask. The total loss is defined as:", + "bbox": [ + 212, + 455, + 787, + 590 + ], + "page_idx": 7 + }, + { + "type": "equation", + "text": "\n$$\n\\mathcal {L} = \\lambda_ {\\text {c o l o r}} \\mathcal {L} _ {\\text {c o l o r}} + \\lambda_ {\\text {d e p t h}} \\mathcal {L} _ {\\text {d e p t h}} \\tag {6}\n$$\n", + "text_format": "latex", + "bbox": [ + 390, + 604, + 785, + 619 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "where we empirically set $\\lambda_{color} = 50$ and $\\lambda_{depth} = 10$ .", + "bbox": [ + 214, + 631, + 596, + 648 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "4 Experiments", + "text_level": 1, + "bbox": [ + 214, + 670, + 375, + 686 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "In this section, we evaluate the performance of PAV using two different datasets. We perform an ablation study to validate our design choices and a comparison against the state-of-the-art monocular head avatar method in various configurations.", + "bbox": [ + 212, + 700, + 785, + 762 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "4.1 Dataset", + "text_level": 1, + "bbox": [ + 214, + 784, + 326, + 797 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Our method takes a video collection of a subject containing different appearances, poses, and expressions. Since there is no existing dataset to evaluate", + "bbox": [ + 212, + 809, + 785, + 840 + ], + "page_idx": 7 + }, + { + "type": "page_number", + "text": "8", + "bbox": [ + 217, + 114, + 228, + 126 + ], + "page_idx": 7 + }, + { + "type": "header", + "text": "A. Caliskan et al.", + "bbox": [ + 271, + 114, + 388, + 127 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "learning 3D head avatar from video collections of the same subjects with various appearances, poses, and expressions, we compiled a new Video Collection (VidCol) dataset from talking face videos of Roger Federer, King Charles and Tom Hanks across years. This dataset consists of in-the-wild videos recorded in uncontrolled environments. In VidCol dataset, we curated videos from 5 different appearances for each subject. We randomly sampled 500 frames at $512^2$ resolution from each video (or appearance) and we used a total of 2500 frames to train and test a single unified network per subject. We will share this dataset with annotations of semantic masks, tracked 3DMM models, and camera calibrations for further research in this direction.", + "bbox": [ + 212, + 146, + 782, + 295 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "For our experiments, we evaluate all characters and train a separate PAV model for each subject. We develop two different protocols for evaluation:", + "bbox": [ + 212, + 297, + 782, + 328 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "a) Unified Network for Novel Pose/Exp. Synthesis This protocol evaluates the performance in terms of novel appearance, head pose, and facial expression synthesis when training a single unified network from all appearances. For training, we select 400 frames for each appearance of a subject talking in front of a camera with various head poses. It mimics a training video commonly used for monocular neural head avatar methods (Sec. 2). For testing, we use the remaining 100 frames for each appearance with unseen poses and expressions.", + "bbox": [ + 212, + 340, + 782, + 446 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "b) Separated Networks for Novel Pose/Exp. Synthesis This protocol evaluates the novel appearance, head pose, and facial expression generalization ability of the methods when training a separate network for each appearance. We select 400 training images for each appearance of a subject talking of a camera, same as protocol-a. For testing, we choose the remaining 100 unseen pose/expression combinations with ground truth, same as protocol-a, and another 300 unseen pose/expression sets without ground truth image data.", + "bbox": [ + 212, + 460, + 784, + 568 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "4.2 Metrics", + "text_level": 1, + "bbox": [ + 214, + 588, + 326, + 602 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "We report several metrics to evaluate the quality of synthesized images as well as the ablation study of our method. For synthesized images with provided ground-truth, we use Learned Perceptual Patch Similarity (LPIPS) [41], Deep Image Structure and Texture Similarity (DISTS) [8], Structural Similarity Index (SSIM) [36], Peak Signal-to-Noise Ratio (PSNR) and Fréchet Inception Distance (FID) [15]. When there is an absence of ground truth, we adopt FID metric for quantitative comparison.", + "bbox": [ + 212, + 612, + 782, + 719 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "4.3 Implementation Details", + "text_level": 1, + "bbox": [ + 214, + 739, + 455, + 753 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "We optimize Eq. 6 using Adam optimizer [17] with hyperparameters $\\beta_{1} = 0.9$ and $\\beta_{2} = 0.99$ . We set the learning rate to $2.5 \\times 10^{-3}$ for $F_{density}$ , $F_{color}$ and $Z$ , and to $2.5 \\times 10^{-4}$ for $F_{offset}$ . We sample 1024 points along each ray for rendering. Multi-resolution hash encoding used in the method is set to 16 number of levels, 18 hash map size, and 2048 finest resolution. The optimization takes 15K iterations", + "bbox": [ + 212, + 763, + 782, + 839 + ], + "page_idx": 8 + }, + { + "type": "header", + "text": "PAV: Personalized Head Avatar from Video Collection", + "bbox": [ + 369, + 114, + 730, + 127 + ], + "page_idx": 8 + }, + { + "type": "page_number", + "text": "9", + "bbox": [ + 774, + 116, + 784, + 126 + ], + "page_idx": 8 + }, + { + "type": "image", + "img_path": "images/fb9dbf79da6fd0ec4d6e8efc8a9696e8bd0953409ef827f3544c570740b3e090.jpg", + "image_caption": [ + "Fig. 3: Qualitative Ablation Study on VidCol Dataset. This figure shows synthesized images and pixel error maps against ground-truth. Please note that brighter pixel colors denote higher pixel errors. Without appearance embedding with latent neural features and appearance-conditioned density offset (a), the model fails to learn appearance-based distinct details. The latent neural features (LNF) (c) resolve the issue of high-fidelity rendering of texture (blue arrow). The appearance-conditioned density offset field (d) allows the model to learn more accurate geometry and texture that corresponds to the face, hair, and neck regions (red arrow)" + ], + "image_footnote": [], + "bbox": [ + 233, + 141, + 789, + 455 + ], + "page_idx": 9 + }, + { + "type": "text", + "text": "to converge when training network with all appearances of single subjects. More details on the network are provided in the supplementary document.", + "bbox": [ + 212, + 607, + 784, + 638 + ], + "page_idx": 9 + }, + { + "type": "text", + "text": "4.4 Ablation Study", + "text_level": 1, + "bbox": [ + 214, + 662, + 390, + 678 + ], + "page_idx": 9 + }, + { + "type": "text", + "text": "We evaluate different design choices of PAV in Tab. 1 and Fig. 3. In this study, we use protocol-a of the VidCol dataset for all experiments. We first report the results of the final model which includes latent neural features (LNF) (Sec. 3.3) and density offset $\\varDelta_{\\sigma}$ (Sec. 3.4). The full model achieves PSNR score of 25.15 and SSIM score of 0.91 for image synthesis. If we remove the density offset from our model while employing appearance-conditioned radiance estimation, PSNR and SSIM scores for image synthesis decrease to 24.14 and 0.90, respectively. This indicates that without appearance-conditioned density offset $\\varDelta_{\\sigma}$ , density estimation MLP ( $\\mathcal{F}_{\\text {density }}$ in Sec. 3.4) can not learn to synthesize appearance-specific geometry and texture details, thus rendering fails there as well. Even", + "bbox": [ + 212, + 688, + 787, + 840 + ], + "page_idx": 9 + }, + { + "type": "page_number", + "text": "10", + "bbox": [ + 217, + 114, + 235, + 126 + ], + "page_idx": 9 + }, + { + "type": "header", + "text": "A. Caliskan et al.", + "bbox": [ + 271, + 114, + 388, + 126 + ], + "page_idx": 9 + }, + { + "type": "table", + "img_path": "images/e3d1a0f121851fe7cf617751938e6d1d4f58f36ed68d2122d9e405cc3f9464e2.jpg", + "table_caption": [], + "table_footnote": [], + "table_body": "
PSNR ↑SSIM ↑DISTS ↓LPIPS ↓FID ↓
Insta [44]21.030.8490.1850.169104.28
PAV without LNF24.190.9000.1450.11954.39
PAV without Δσ24.140.9060.1340.10847.71
PAV (Full Model)25.150.9100.1330.10746.67
", + "bbox": [ + 250, + 143, + 759, + 229 + ], + "page_idx": 10 + }, + { + "type": "text", + "text": "Table 1: Ablation Study: We evaluate the impact of different components of the proposed method. See Fig. 3 for qualitative comparison.", + "bbox": [ + 215, + 231, + 784, + 258 + ], + "page_idx": 10 + }, + { + "type": "text", + "text": "though the model $\\mathrm{w / o}\\varDelta_{\\sigma}$ synthesise images that look on par with ground truth in terms of texture, it generates severe artifacts which can be seen in pixel error maps. An example of this behavior can also be seen in Fig. 3 $(\\mathrm{w / o}\\varDelta_{\\sigma})$ Rendering failures on face, head, and neck regions are indicated by red arrow and error map.", + "bbox": [ + 212, + 303, + 782, + 378 + ], + "page_idx": 10 + }, + { + "type": "text", + "text": "Next, we evaluate the impact of latent neural features (LNF) as an appearance embedding on PAV. In this experiment, to condition the canonical MLPs in PAV, inspired by \"Nerf In-the-wild\" [22, 37], we adopt Generative Latent Optimization (GLO) for appearance embedding. Specifically, this approach defines a single real-valued appearance embedding vector $l$ for each appearance in the dataset. To this end, we employ a pre-trained encoder from in-the-wild face detector RetinaFace [7] for GLO. During training, if we replace $\\mathbf{Z} = \\{Z_{\\mathrm{i}}\\}_{i=1}^{N}$ , $Z_{i} \\in \\mathbb{R}^{512 \\times 512 \\times 16}$ with $\\mathbf{L} = \\{l_{\\mathrm{i}}\\}_{i=1}^{N}$ , $l_{i} \\in \\mathbb{R}^{512 \\times 1}$ for $N$ appearances of each subject, and pass it to appearance conditioned NeRF (See Sec. 3.4), PSNR and SSIM scores for image synthesis decrease to 24.19 and 0.90, respectively. This is also evident from the qualitative results shown in Fig. 3 (w/o LNF), indicating that latent neural features help in improved image synthesis quality with their local features attached to 3D geometry. Rendering failures on head, neck, and skin texture are shown by red arrow, blue arrow and error maps in Fig. 3. Finally, as discussed in Sec. 3, we keep the network design as we explain after the ablation study.", + "bbox": [ + 212, + 378, + 784, + 604 + ], + "page_idx": 10 + }, + { + "type": "text", + "text": "4.5 Comparison", + "text_level": 1, + "bbox": [ + 215, + 631, + 361, + 645 + ], + "page_idx": 10 + }, + { + "type": "text", + "text": "Since our method is the first work that learns dynamic deformable head NeRF from multiple appearances of the same subjects using a unified network, we ourselves design some baseline comparisons in this section. First, we compare our method against Insta [44] using protocol-a (Sec. 4.1) of the VidCol dataset. Both methods are trained using the frames sampled from all appearances of the same subject. For testing, both methods synthesize images for target head pose and expression. The results are summarized in Tab. 2 and Fig. 4. For protocol-a, our method clearly outperforms the baseline for final image synthesis results across all metrics. It should be noted that PAV framework jointly learns all different appearances via one-time training using a unified network and it is able to preserve appearance-specific details for novel poses and expressions. On the other hand, as shown in Fig. 4, Insta can not disentangle appearance from the", + "bbox": [ + 212, + 657, + 787, + 840 + ], + "page_idx": 10 + }, + { + "type": "header", + "text": "PAV: Personalized Head Avatar from Video Collection", + "bbox": [ + 369, + 114, + 730, + 127 + ], + "page_idx": 10 + }, + { + "type": "page_number", + "text": "11", + "bbox": [ + 767, + 116, + 782, + 126 + ], + "page_idx": 10 + }, + { + "type": "image", + "img_path": "images/862f9ccf65ef075cc91ec5ff1d498e4c0b5dbe4a2ed0739cf4064be36fd6ad56.jpg", + "image_caption": [ + "Fig. 4: Qualitative comparison of our method (PAV) against baseline method (Insta [44]. Columns $\\{2,6\\}$ show target head poses and facial expressions for synthesized images at columns $\\{3,4,7,8\\}$ ." + ], + "image_footnote": [], + "bbox": [ + 222, + 142, + 803, + 492 + ], + "page_idx": 11 + }, + { + "type": "text", + "text": "neural radiance field, which results in a mixture of appearances in synthesized images and inaccurate geometry predictions. This is the limitation of Insta we are addressing in the paper.", + "bbox": [ + 212, + 579, + 785, + 625 + ], + "page_idx": 11 + }, + { + "type": "text", + "text": "Second, we train separate Insta [44] models for each appearance of the same subject using protocol-b while training a single unified model for our method using the frames sampled from all appearances of the same subject. For testing, we use the same test data from protocol-b for both PAV and Insta [44] and synthesize images for novel poses and expressions. Note that, we added more novel expressions and poses in protocol-b (Sec. 4.1) to test the generalization of our method under a unified training schema, which results in increasing FID score for both methods in Tab. 3. Even though the results show that artifacts (in Fig. 4) of the Insta renders are decreased in Fig. 5, our method demonstrates better-rendering results (Fig. 5) and lower FID score (Tab. 3). This indicates that PAV's single unified model helps in learning generic neural radiance field representation across all appearances of the same subject while preserving appearance-based details. Please note that we are demonstrating Insta's performance under a challenging pose/expression in Fig. 5 which can also be seen in", + "bbox": [ + 212, + 628, + 787, + 840 + ], + "page_idx": 11 + }, + { + "type": "page_number", + "text": "12", + "bbox": [ + 217, + 114, + 235, + 126 + ], + "page_idx": 11 + }, + { + "type": "header", + "text": "A. Caliskan et al.", + "bbox": [ + 271, + 114, + 388, + 127 + ], + "page_idx": 11 + }, + { + "type": "table", + "img_path": "images/60e1dae0bc93fd572ccbadc1d3b35012533f73f2df59aae16b9fe6acf9f1d9ae.jpg", + "table_caption": [], + "table_footnote": [], + "table_body": "
MethodVideo Collection 1Video Collection 2
PSNR ↑SSIM ↑DISTS ↓LPIPS ↓FID ↓PSNR ↑SSIM ↑DISTS ↓LPIPS ↓FID ↓
Protocol (a): Novel Head Pose and Expression Synthesis
Insta [44]21.030.840.1850.169104.2823.670.860.1870.16398.87
PAV (Ours)25.150.910.1330.10746.6727.090.890.1440.12053.91
", + "bbox": [ + 218, + 145, + 787, + 213 + ], + "page_idx": 12 + }, + { + "type": "text", + "text": "Table 2: Comparison of PAV against Insta [44]. See Fig. 4 for qualitative comparison. Video Collection 1 and 2 refer to Federer and King Charles splits of VidCol dataset, respectively.", + "bbox": [ + 215, + 214, + 782, + 253 + ], + "page_idx": 12 + }, + { + "type": "text", + "text": "the original paper results under extreme poses/expressions. In Fig. 5, the reference image is for photoconsistency check only since there is no ground truth image for that pose/expression. Finally, we provide additional qualitative results of avatars animated in the supplementary video. We also compared PAV against Insta for model efficiency using protocol-b dataset and recorded values in Tab. 3. Although we train our model with multiple appearances jointly, PAV is more efficient than Insta in terms of memory footprint and training time", + "bbox": [ + 212, + 297, + 784, + 402 + ], + "page_idx": 12 + }, + { + "type": "image", + "img_path": "images/06e6d4e9ac32ff115f6293a48eda57c3754b4c7e80c1df691cc1c39f984f2b04.jpg", + "image_caption": [ + "Reference" + ], + "image_footnote": [], + "bbox": [ + 302, + 426, + 403, + 585 + ], + "page_idx": 12 + }, + { + "type": "image", + "img_path": "images/785981cd406aa46df4b499f550ab7df46574d32d2a2a328b131545f1839849bb.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 413, + 435, + 491, + 507 + ], + "page_idx": 12 + }, + { + "type": "image", + "img_path": "images/e7b2a0970d0ea6029004843b50749a6242a6eb6f832be788d1c488f5bc0c3603.jpg", + "image_caption": [ + "Target Pose/Exp.", + "Fig. 5: Qualitative comparison of PAV against single-appearance traines Insta [44]. Please note that the reference image is for photoconsistency check only since there is no ground truth image for that pose/expression" + ], + "image_footnote": [], + "bbox": [ + 421, + 511, + 493, + 584 + ], + "page_idx": 12 + }, + { + "type": "image", + "img_path": "images/3bbf67cab567bb2963ea47f86303776b311718312d4ff11c2d718ec5bf22872b.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 514, + 429, + 584, + 506 + ], + "page_idx": 12 + }, + { + "type": "image", + "img_path": "images/6cefb32e6d820a2f9a30c6c8c622afccf81ce2c5b1b1b9c762d8d4a5cb7c6d02.jpg", + "image_caption": [ + "Insta [44]" + ], + "image_footnote": [], + "bbox": [ + 514, + 507, + 586, + 584 + ], + "page_idx": 12 + }, + { + "type": "image", + "img_path": "images/45aee65b3f48520f21bf405d7f9cd85f8772b2c46159bfe22a07bd61d3e665c9.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 614, + 430, + 679, + 505 + ], + "page_idx": 12 + }, + { + "type": "image", + "img_path": "images/ccb511cdbc55f7c76327c5111109e121b8472b176573cbe83398ed8bcecb894c.jpg", + "image_caption": [ + "PAV" + ], + "image_footnote": [], + "bbox": [ + 617, + 507, + 679, + 584 + ], + "page_idx": 12 + }, + { + "type": "text", + "text": "5 Discussion", + "text_level": 1, + "bbox": [ + 215, + 702, + 354, + 718 + ], + "page_idx": 12 + }, + { + "type": "text", + "text": "We presented PAV which is a novel method for learning personalized head avatars from monocular and in-the-wild video collections of a person. We demonstrated that PAV can model a 3D head avatar from an unstructured video collection of a subject in a single unified network while disentangling face appearances from the radiance field. PAV can be trained simultaneously for multiple appearances of the same subject, allowing novel pose and expression generalization and leading to a superior rendering quality of our method compared to the baseline.", + "bbox": [ + 212, + 733, + 787, + 840 + ], + "page_idx": 12 + }, + { + "type": "header", + "text": "PAV: Personalized Head Avatar from Video Collection", + "bbox": [ + 369, + 114, + 730, + 127 + ], + "page_idx": 12 + }, + { + "type": "page_number", + "text": "13", + "bbox": [ + 767, + 116, + 784, + 126 + ], + "page_idx": 12 + }, + { + "type": "table", + "img_path": "images/ea68bdeb1ba7a7a738820dc72b53e940306cf5081ab4370c59e5f66756b37cba.jpg", + "table_caption": [], + "table_footnote": [], + "table_body": "
Train Time MinutesNumber of Param. MillionModel Size (MB)FID ↓
Multiple Insta [44]13579.71932109.22
PAV (Ours)4526.865091.01
", + "bbox": [ + 220, + 143, + 787, + 218 + ], + "page_idx": 13 + }, + { + "type": "text", + "text": "Table 3: Efficiency and performance comparison between PAV and Insta [44] trained for each appearance (Single Appearance Training) in dataset collection 1. FID score is computed from synthetized images under novel head poses and expressions where there are no existing ground-truth images.", + "bbox": [ + 215, + 220, + 787, + 275 + ], + "page_idx": 13 + }, + { + "type": "image", + "img_path": "images/631be0a96611e488f6eb28929bbf6b526136f5c25d8411e93327cabe30d5e28e.jpg", + "image_caption": [ + "Fig. 6: Multi-identity and multi-appearance extension of PAV. The learned appearance (middle) rendered in a new viewpoint (left) does not match the visual quality of input images (right). Please note that the reference image is for photoconsistency check only since there is no ground truth image for that pose/expression." + ], + "image_footnote": [], + "bbox": [ + 349, + 301, + 651, + 468 + ], + "page_idx": 13 + }, + { + "type": "text", + "text": "Limitation for Multi-Identity. We extended our method for multiple identities and multiple appearances for each identity. To this end, we train a single unified PAV model using frames smaped from all appearances of Federer and King subjects from VidCol dataset with shared canonical space. We evaluate novel head pose and facial expression rendering of this model and Fig. 6 illustrates the results. Even though this model produces reasonable results, synthesized images demonstrate a lack of subject-specific shape and appearance details with rendering artifacts. In the paper, we focus on multiple appearances of a single subject. Scaling up this model to a wide range of subjects and exploring generalized dynamic deformable NeRF would be interesting future work.", + "bbox": [ + 212, + 565, + 787, + 718 + ], + "page_idx": 13 + }, + { + "type": "text", + "text": "Ethical Considerations. Although the rapid progress on digital human avatars facilitates numerous downstream applications, it also raises concerns about ethical misuse. On the synthesis side, it would be ideal to actively use watermarking technologies and avoid driving avatars with different identities. In this work, we aim to faithfully reproduce images of a person with the capability of rendering unseen poses and expressions and switching appearance within their own set of appearances. The work does not intend to create animations that didn't happen.", + "bbox": [ + 212, + 733, + 787, + 840 + ], + "page_idx": 13 + }, + { + "type": "page_number", + "text": "14", + "bbox": [ + 217, + 114, + 235, + 126 + ], + "page_idx": 13 + }, + { + "type": "header", + "text": "A. Caliskan et al.", + "bbox": [ + 271, + 114, + 388, + 127 + ], + "page_idx": 13 + }, + { + "type": "text", + "text": "References", + "text_level": 1, + "bbox": [ + 215, + 143, + 323, + 160 + ], + "page_idx": 14 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "1. Athar, S., Xu, Z., Sunkavalli, K., Shechtman, E., Shu, Z.: Rignerf: Fully controllable neural 3d portraits. In: Proceedings of the IEEE/CVF conference on Computer Vision and Pattern Recognition. pp. 20364-20373 (2022)", + "2. Averbuch-Elor, H., Cohen-Or, D., Kopf, J., Cohen, M.F.: Bringing portraits to life. ACM transactions on graphics (TOG) 36(6), 1-13 (2017)", + "3. Bai, Z., Tan, F., Huang, Z., Sarkar, K., Tang, D., Qiu, D., Meka, A., Du, R., Dou, M., Orts-Escalano, S., et al.: Learning personalized high quality volumetric head avatars from monocular rgb videos. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 16890-16900 (2023)", + "4. Bühler, M.C., Meka, A., Li, G., Beeler, T., Hilliges, O.: Varitex: Variational neural face textures. In: Proceedings of the IEEE/CVF International Conference on Computer Vision. pp. 13890-13899 (2021)", + "5. Chen, C., O'Toole, M., Bharaj, G., Garrido, P.: Implicit neural head synthesis via controllable local deformation fields. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 416-426 (2023)", + "6. Clark, J.H.: Hierarchical geometric models for visible surface algorithms. Communications of the ACM 19(10), 547-554 (1976)", + "7. Deng, J., Guo, J., Zhou, Y., Yu, J., Kotsia, I., Zafeiriou, S.: Retinaface: Single-stage dense face localisation in the wild. arXiv preprint arXiv:1905.00641 (2019)", + "8. Ding, K., Ma, K., Wang, S., Simoncelli, E.P.: Image Quality Assessment: Unifying Structure and Texture Similarity. TPAMI (2022)", + "9. Egger, B., Smith, W.A., Tewari, A., Wuhrer, S., Zollhoefer, M., Beeler, T., Bernard, F., Bolkart, T., Kortylewski, A., Romdhani, S., et al.: 3d morphable face models—past, present, and future. ACM Transactions on Graphics (ToG) 39(5), 1-38 (2020)", + "0. Gafni, G., Thies, J., Zollhöfer, M., Nießner, M.: Dynamic neural radiance fields for monocular 4d facial avatar reconstruction. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR). pp. 8649-8658 (June 2021)", + "1. Gao, X., Zhong, C., Xiang, J., Hong, Y., Guo, Y., Zhang, J.: Reconstructing personalized semantic facial nerf models from monocular video. ACM Transactions on Graphics (TOG) 41(6), 1-12 (2022)", + "2. Garrido, P., Valgaerts, L., Rehmsen, O., Thormahlen, T., Perez, P., Theobalt, C.: Automatic face reenactment. In: Proceedings of the IEEE conference on computer vision and pattern recognition. pp. 4217-4224 (2014)", + "3. Garrido, P., Valgaerts, L., Sarmadi, H., Steiner, I., Varanasi, K., Perez, P., Theobalt, C.: Vdub: Modifying face video of actors for plausible visual alignment to a dubbed audio track. In: Computer graphics forum. vol. 34, pp. 193-204. Wiley Online Library (2015)", + "4. Grassal, P.W., Prinzler, M., Leistner, T., Rother, C., Nießner, M., Thies, J.: Neural head avatars from monocular rgb videos. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 18653-18664 (2022)", + "5. Heusel, M., Ramsauer, H., Unterthiner, T., Nessler, B., Hochreiter, S.: Gans trained by a two time-scale update rule converge to a local nash equilibrium. Advances in neural information processing systems 30 (2017)", + "6. Kim, H., Garrido, P., Tewari, A., Xu, W., Thies, J., Niessner, M., Pérez, P., Richardt, C., Zollhöfer, M., Theobalt, C.: Deep video portraits. ACM transactions on graphics (TOG) 37(4), 1-14 (2018)" + ], + "bbox": [ + 225, + 175, + 785, + 840 + ], + "page_idx": 14 + }, + { + "type": "header", + "text": "PAV: Personalized Head Avatar from Video Collection", + "bbox": [ + 367, + 114, + 730, + 127 + ], + "page_idx": 14 + }, + { + "type": "page_number", + "text": "15", + "bbox": [ + 767, + 116, + 785, + 126 + ], + "page_idx": 14 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "17. Kingma, D.P., Ba, J.: Adam: A method for stochastic optimization. arXiv preprint arXiv:1412.6980 (2014)", + "18. Kuang, Z., Olszewski, K., Chai, M., Huang, Z., Achlioptas, P., Tulyakov, S.: Neroic: Neural rendering of objects from online image collections. ACM Transactions on Graphics (TOG) 41(4), 1-12 (2022)", + "19. Li, T., Bolkart, T., Black, M.J., Li, H., Romero, J.: Learning a model of facial shape and expression from 4D scans. ACM Transactions on Graphics, (Proc. SIGGRAPH Asia) 36(6), 194:1-194:17 (2017), https://doi.org/10.1145/3130800.3130813", + "20. Liu, L., Habermann, M., Rudnev, V., Sarkar, K., Gu, J., Theobalt, C.: Neural actor: Neural free-view synthesis of human actors with pose control. ACM transactions on graphics (TOG) 40(6), 1-16 (2021)", + "21. Mallya, A., Wang, T.C., Liu, M.Y.: Implicit warping for animation with image sets. Advances in Neural Information Processing Systems 35, 22438-22450 (2022)", + "22. Martin-Brualla, R., Radwan, N., Sajjadi, M.S., Barron, J.T., Dosovitskiy, A., Duckworth, D.: Nerf in the wild: Neural radiance fields for unconstrained photo collections. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 7210-7219 (2021)", + "23. Mildenhall, B., Srinivasan, P.P., Tancik, M., Barron, J.T., Ramamoorthi, R., Ng, R.: Nerf: Representing scenes as neural radiance fields for view synthesis. Communications of the ACM 65(1), 99-106 (2021)", + "24. Müller, T., Evans, A., Schied, C., Keller, A.: Instant neural graphics primitives with a multiresolution hash encoding. ACM Transactions on Graphics (ToG) 41(4), 1-15 (2022)", + "25. Nirkin, Y., Keller, Y., Hassner, T.: Fsgan: Subject agnostic face swapping and reenactment. In: Proceedings of the IEEE/CVF international conference on computer vision. pp. 7184-7193 (2019)", + "26. Park, K., Sinha, U., Hedman, P., Barron, J.T., Bouaziz, S., Goldman, D.B., Martin-Brualla, R., Seitz, S.M.: Hypernerf: A higher-dimensional representation for topologically varying neural radiance fields. arXiv preprint arXiv:2106.13228 (2021)", + "27. Peng, S., Dong, J., Wang, Q., Zhang, S., Shuai, Q., Zhou, X., Bao, H.: Animatable neural radiance fields for modeling dynamic human bodies. In: Proceedings of the IEEE/CVF International Conference on Computer Vision. pp. 14314-14323 (2021)", + "28. Pumarola, A., Corona, E., Pons-Moll, G., Moreno-Noguer, F.: D-nerf: Neural radiance fields for dynamic scenes. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 10318-10327 (2021)", + "29. Siarohin, A., Lathuilière, S., Tulyakov, S., Ricci, E., Sebe, N.: First order motion model for image animation. Advances in neural information processing systems 32 (2019)", + "30. Tancik, M., Casser, V., Yan, X., Pradhan, S., Mildenhall, B., Srinivasan, P.P., Barron, J.T., Kretzschmar, H.: Block-nerf: Scalable large scene neural view synthesis. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 8248-8258 (2022)", + "31. Tewari, A., Fried, O., Thies, J., Sitzmann, V., Lombardi, S., Sunkavalli, K., MartinBrualla, R., Simon, T., Saragih, J., Nießner, M., et al.: State of the art on neural rendering. In: Computer Graphics Forum. vol. 39, pp. 701-727. Wiley Online Library (2020)", + "32. Tewari, A., Thies, J., Mildenhall, B., Srinivasan, P., Tretschk, E., Yifan, W., Lassner, C., Sitzmann, V., Martin-Brualla, R., Lombardi, S., et al.: Advances in neural rendering. In: Computer Graphics Forum. vol. 41, pp. 703-735. Wiley Online Library (2022)" + ], + "bbox": [ + 215, + 147, + 784, + 839 + ], + "page_idx": 15 + }, + { + "type": "page_number", + "text": "16", + "bbox": [ + 217, + 114, + 235, + 126 + ], + "page_idx": 15 + }, + { + "type": "header", + "text": "A. Caliskan et al.", + "bbox": [ + 271, + 114, + 388, + 126 + ], + "page_idx": 15 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "33. Thies, J., Zollhöfer, M., Nießner, M.: Deferred neural rendering: Image synthesis using neural textures. Acm Transactions on Graphics (TOG) 38(4), 1-12 (2019)", + "34. Thies, J., Zollhofer, M., Stamminger, M., Theobalt, C., Nieflner, M.: Face2face: Real-time face capture and reenactment of rgb videos. In: Proceedings of the IEEE conference on computer vision and pattern recognition. pp. 2387-2395 (2016)", + "35. Wang, T.C., Liu, M.Y., Zhu, J.Y., Tao, A., Kautz, J., Catanzaro, B.: High-resolution image synthesis and semantic manipulation with conditional gans. In: Proceedings of the IEEE conference on computer vision and pattern recognition. pp. 8798-8807 (2018)", + "36. Wang, Z., Bovik, A.C., Sheikh, H.R., Simoncelli, E.P.: Image quality assessment: From error visibility to structural similarity. TIP (2004)", + "37. Weng, C.Y., Srinivasan, P.P., Curless, B., Kemelmacher-Shlizerman, I.: Person-nerf: Personalized reconstruction from photo collections. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 524-533 (2023)", + "38. Yang, B., Bao, C., Zeng, J., Bao, H., Zhang, Y., Cui, Z., Zhang, G.: Neumesh: Learning disentangled neural mesh-based implicit field for geometry and texture editing. In: European Conference on Computer Vision. pp. 597-614. Springer (2022)", + "39. Zakharov, E., Ivakhnenko, A., Shysheya, A., Lempitsky, V.: Fast bi-layer neural synthesis of one-shot realistic head avatars. In: Computer Vision-ECCV 2020: 16th European Conference, Glasgow, UK, August 23-28, 2020, Proceedings, Part XII 16. pp. 524-540. Springer (2020)", + "40. Zhang, B., Qi, C., Zhang, P., Zhang, B., Wu, H., Chen, D., Chen, Q., Wang, Y., Wen, F.: Metaportrait: Identity-preserving talking head generation with fast personalized adaptation. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR). pp. 22096-22105 (June 2023)", + "41. Zhang, R., Isola, P., Efros, A.A., Shechtman, E., Wang, O.: The unreasonable effectiveness of deep features as a perceptual metric. In: CVPR (2018)", + "42. Zheng, Y., Abrevaya, V.F., Bühler, M.C., Chen, X., Black, M.J., Hilliges, O.: Im avatar: Implicit morphable head avatars from videos. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 13545-13555 (2022)", + "43. Zielonka, W., Bolkart, T., Thies, J.: Towards metrical reconstruction of human faces. In: European Conference on Computer Vision. pp. 250-269. Springer (2022)", + "44. Zielonka, W., Bolkart, T., Thies, J.: Instant volumetric head avatars. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 4574-4584 (2023)" + ], + "bbox": [ + 212, + 146, + 787, + 674 + ], + "page_idx": 16 + }, + { + "type": "header", + "text": "PAV: Personalized Head Avatar from Video Collection", + "bbox": [ + 369, + 114, + 730, + 127 + ], + "page_idx": 16 + }, + { + "type": "page_number", + "text": "17", + "bbox": [ + 767, + 116, + 784, + 126 + ], + "page_idx": 16 + } +] \ No newline at end of file diff --git a/2024/PAV_ Personalized Head Avatar from Unstructured Video Collection/75c53aba-b5dd-481e-8738-bfaa232bbce0_model.json b/2024/PAV_ Personalized Head Avatar from Unstructured Video Collection/75c53aba-b5dd-481e-8738-bfaa232bbce0_model.json new file mode 100644 index 0000000000000000000000000000000000000000..9272ae105e718852c2c1ec1a7829a68c8834f274 --- /dev/null +++ b/2024/PAV_ Personalized Head Avatar from Unstructured Video Collection/75c53aba-b5dd-481e-8738-bfaa232bbce0_model.json @@ -0,0 +1,2005 @@ +[ + [ + { + "type": "title", + "bbox": [ + 0.29, + 0.141, + 0.713, + 0.184 + ], + "angle": 0, + "content": "PAV: Personalized Head Avatar from Unstructured Video Collection" + }, + { + "type": "text", + "bbox": [ + 0.271, + 0.212, + 0.733, + 0.228 + ], + "angle": 0, + "content": "Akin Caliskan1, Berkay Kicanaoglu1, and Hyeongwoo Kim2" + }, + { + "type": "text", + "bbox": [ + 0.456, + 0.24, + 0.55, + 0.253 + ], + "angle": 0, + "content": "1 Flawless AI" + }, + { + "type": "text", + "bbox": [ + 0.412, + 0.254, + 0.592, + 0.268 + ], + "angle": 0, + "content": "2 Imperial College London" + }, + { + "type": "text", + "bbox": [ + 0.263, + 0.302, + 0.74, + 0.552 + ], + "angle": 0, + "content": "Abstract. We propose PAV, Personalized Head Avatar for the synthesis of human faces under arbitrary viewpoints and facial expressions. PAV introduces a method that learns a dynamic deformable neural radiance field (NeRF), in particular from a collection of monocular talking face videos of the same character under various appearance and shape changes. Unlike existing head NeRF methods that are limited to modeling such input videos on a per-appearance basis, our method allows for learning multi-appearance NeRFs, introducing appearance embedding for each input video via learnable latent neural features attached to the underlying geometry. Furthermore, the proposed appearance-conditioned density formulation facilitates the shape variation of the character, such as facial hair and soft tissues, in the radiance field prediction. To the best of our knowledge, our approach is the first dynamic deformable NeRF framework to model appearance and shape variations in a single unified network for multi-appearances of the same subject. We demonstrate experimentally that PAV outperforms the baseline method in terms of visual rendering quality in our quantitative and qualitative studies on various subjects." + }, + { + "type": "title", + "bbox": [ + 0.218, + 0.599, + 0.375, + 0.615 + ], + "angle": 0, + "content": "1 Introduction" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.629, + 0.788, + 0.844 + ], + "angle": 0, + "content": "Digital human avatars enabling the editing of facial expressions and head motion have broad applications across telepresence, animation, and digital content creation. The demands for personalized human avatars such as digital doubles have also increased over the past few years not only in the film industry but in the research community. To enable the widespread adaptation of these neural avatars, they should be easy to generate and animate under novel poses and facial expressions. A vast majority of existing methods aim to learn neural avatars from monocular videos to make it more accessible and easy to use [1,10,12-14,16,34,42,44]. Early approaches [12,13,16,34] mainly rely on 3D morphable models (3DMM) to guide the modeling process and to provide an intuitive control mechanism in novel view synthesis. Despite their success, they do not model the 3D face geometry. To address this issue, neural radiance fields (NeRFs) [23] have demonstrated volumetric 3D modeling of face geometry and appearance [1,10,14,42,44] under novel poses and expressions. While this allows" + } + ], + [ + { + "type": "page_number", + "bbox": [ + 0.218, + 0.116, + 0.23, + 0.127 + ], + "angle": 0, + "content": "2" + }, + { + "type": "header", + "bbox": [ + 0.272, + 0.115, + 0.39, + 0.128 + ], + "angle": 0, + "content": "A. Caliskan et al." + }, + { + "type": "image", + "bbox": [ + 0.254, + 0.153, + 0.75, + 0.379 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.214, + 0.392, + 0.788, + 0.518 + ], + "angle": 0, + "content": "Fig. 1: We present an approach for learning personalized head avatars from an unstructured video collection of the same subject. The videos are taken in different environments, with diverse appearances and possible shape changes of a character, as shown on the left. Our method leverages all of the input clips into a single unified dynamic deformable NeRF framework and offers high-quality video replays in novel head poses and facial expressions. On the right, we show some rendering results of each input appearance of the same subject aligned in head pose and facial expression that are not seen during training. Notably, our method can synthesize identity-specific details while showing coherent expressions across different appearances." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.551, + 0.788, + 0.748 + ], + "angle": 0, + "content": "photo-realistic image quality and animation, NeRF-based approaches require per-appearance optimization, maintaining the consistent appearance of the subject, which is constrained to the pre-recording of characters in controlled settings. However, this limits the potential for practical application scenarios where we want to generate a 3D head avatar from a collection of prerecorded video clips of a subject. In this work, we present the Personalized Head Avatar method transforming an unstructured personal video collection, containing images spanning multiple years with different appearances, head poses, and expressions, into a 3D representation of the subject. To create an avatar, we only require an unstructured collection of video clips of the person in arbitrary poses, expressions, and appearances. Our novel framework enables creation of avatars from sparse observations, while yielding high reconstruction quality in novel combinations of pose and appearance, and opens up new applications." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.75, + 0.788, + 0.841 + ], + "angle": 0, + "content": "Learning dynamic deformable head NeRF from unstructured video clips is a challenging problem. First, video collections can contain frames of the same identity at different appearances, poses, and expressions. Second, we only have limited pose and expression variation for each appearance, so it is unlikely that all poses and expressions of the face would be well-observed for any given appearance. Given the data, existing NeRF approaches' appearance-specific nature" + } + ], + [ + { + "type": "header", + "bbox": [ + 0.37, + 0.115, + 0.732, + 0.128 + ], + "angle": 0, + "content": "PAV: Personalized Head Avatar from Video Collection" + }, + { + "type": "page_number", + "bbox": [ + 0.775, + 0.117, + 0.785, + 0.127 + ], + "angle": 0, + "content": "3" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.146, + 0.788, + 0.632 + ], + "angle": 0, + "content": "largely limits their potential for practical application scenarios. They require an independent model that is trained from scratch for every appearance from each monocular video. To address these challenges, we model shared deformation field and canonical space neural volumetric representation for all appearances with appearance-dependent latent features. Our key insight is that although observed faces have different appearances across video clips, they should all be explained by a common deformation model with small differences since they all originate from the same person. Furthermore, we assert that every video clip specific to an appearance contains both general and unique information. A single unified network could collaboratively learn the generic information from various appearances, thus this benefits rendering performance as well. The single network is designed to learn a deformation-aware dynamic head NeRF while disentangling appearance-related details from those that are not appearance-specific. For appearance conditioning, we propose employing learnable appearance embedding attached to vertices of the tracked mesh and passing them to the dynamic head NeRF network. By adopting this process, PAV is able to learn discriminative latent neural features that are specific for each appearance, resulting in better rendering of appearance-specific details for arbitrary poses and expressions. Unlike existing NeRF methods [1, 10, 14, 42, 44], PAV simultaneously trains on all appearances of the same subject, eliminating the need for separate training sessions for each appearance and reducing the overall training duration. Learning of personalized head avatars in a single unified network from an unstructured collection of videos has not been addressed before and is a non-trivial extension of existing monocular head avatar generation methods since they can not handle multiple appearances in a single network. PAV makes this possible through a carefully designed appearance conditioning representation, geometry-attached trainable latent neural texture to learn appearance-specific features, appearance conditioned implicit functions to predict variations in color and density. We conducted a variety of experiments to show these improvements over the baseline approach. Since there is no such public dataset to evaluate our method, we curate our own in-the-wild video collections for multiple subjects. Overall, we present Personalized Head Avatar with the following contributions:" + }, + { + "type": "text", + "bbox": [ + 0.226, + 0.687, + 0.788, + 0.765 + ], + "angle": 0, + "content": "- We present, PAV, a framework for learning controllable head avatars from unconstrained short video collections where the same person shows different appearances and facial expressions. To the best of our knowledge, this is the first method learning appearance controllable head avatars from in-the-wild monocular video collections of the same subject." + }, + { + "type": "text", + "bbox": [ + 0.226, + 0.771, + 0.788, + 0.802 + ], + "angle": 0, + "content": "- The proposed method can synthesise images of head avatar under target appearance, head pose and facial expressions." + }, + { + "type": "text", + "bbox": [ + 0.226, + 0.81, + 0.788, + 0.842 + ], + "angle": 0, + "content": "- PAV learns modeling local changes of geometry and color across different appearances of the same identity." + }, + { + "type": "list", + "bbox": [ + 0.226, + 0.687, + 0.788, + 0.842 + ], + "angle": 0, + "content": null + } + ], + [ + { + "type": "page_number", + "bbox": [ + 0.218, + 0.116, + 0.23, + 0.127 + ], + "angle": 0, + "content": "4" + }, + { + "type": "header", + "bbox": [ + 0.272, + 0.115, + 0.39, + 0.128 + ], + "angle": 0, + "content": "A. Caliskan et al." + }, + { + "type": "title", + "bbox": [ + 0.217, + 0.145, + 0.388, + 0.161 + ], + "angle": 0, + "content": "2 Related Work" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.174, + 0.789, + 0.797 + ], + "angle": 0, + "content": "Monocular Neural Head Avatars. More recent methods learn neural representations of the person from monocular videos and use neural renderers [31, 32] to directly generate photo-realistic images in the target head pose and facial expression [1-3, 5, 10, 11, 14, 16, 21, 29, 33, 40, 42, 44]. These methods are generally classified into image-based or 3D-based neural rendering methods [31, 32]. The image-based models synthesize the face of a subject without relying on any representation of 3D space. Existing image-based approaches either employ learned warping fields [2, 29] to deform an input image or encoder-decoder architecture where decoder synthesize the output image [21, 25, 35, 39]. Even though these methods produce high-quality results, they suffer from artifacts for strong head pose and expression changes due to the lack of 3D geometry. The 3D-based neural rendering methods, on the other hand, rely on 3D Morphable Models (3DMM) [9]. These methods are generally classified into explicit and implicit models. Explicit models utilize 2D neural rendering methods [31, 32] and rely on neural networks which can be conditioned on coarse RGB renderings based on a linear texture model [16], uv-maps [4] or latent feature maps [33]. Although these techniques produce high-quality renderings, they are confined to operating within a 2D image space and are limited to craniofacial structures. To tackle this, implicit neural representations for modeling 3D faces have recently become a major focus of attention. In particular, NeRFs [23] shows photorealistic face image synthesis in novel poses and expressions [1, 10, 14, 42, 44]. Implicit models represent the geometry using implicit surface functions or volumetric representations and they are optimized via volumetric rendering [23]. Early approaches [23] and follow-up works synthesize novel views of complex static scenes using differentiable volumetric rendering. After introduction of NeRF for static scenes, a natural research direction was to generalize it to dynamic and time-varying ones [26, 28]. The reconstruction problem is divided into two different spaces, the deformed space and the canonical space. For human body, a series of approaches have been proposed to target this challenge [20, 27]. In particular, these methods are differentiated from each other based on deformation field modeling. Gafni et al. [10] introduced implicit deformation modeling conditioned on facial expression and head pose. In contrast to modeling the deformation field implicitly, IMAvatar [42], RigNerf [1] and Insta [44] propose explicit modeling of the deformation field to improve control and rendering quality. Our method falls into this paradigm as we build our system on top of dynamic deformable head NeRF. In particular, we take inspiration from Insta [44] for designing our framework. However, existing deformable dynamic head NeRF methods including Insta [44] require per-identity optimization, maintaining the consistent appearance of the same subject. To the best of our knowledge, PAV is the first method to learn a single unified dynamic deformable head NeRF from monocular video of multiple appearances of the same subject." + }, + { + "type": "text", + "bbox": [ + 0.216, + 0.81, + 0.792, + 0.842 + ], + "angle": 0, + "content": "Geometry-anchored Features for NeRFs. PAV introduces appearance-conditioned head NeRF by using geometry-attached features. By using these features, exist" + } + ], + [ + { + "type": "header", + "bbox": [ + 0.369, + 0.115, + 0.732, + 0.128 + ], + "angle": 0, + "content": "PAV: Personalized Head Avatar from Video Collection" + }, + { + "type": "page_number", + "bbox": [ + 0.775, + 0.117, + 0.785, + 0.127 + ], + "angle": 0, + "content": "5" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.147, + 0.784, + 0.193 + ], + "angle": 0, + "content": "ing works show improved rendering quality for static objects [38] and human bodies [3, 20]. However, none of the existing methods [3, 20] has yet addressed multi-appearance learning for dynamic neural radiance fields." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.207, + 0.785, + 0.342 + ], + "angle": 0, + "content": "NeRF from Unstructured Data Collections. Learning neural radiance field from unstructured data collections has recently gained much attention. Existing works have enabled photorealistic view synthesis of challenging static scenes, including tourist sites [22] and city-scale scenes [30], and static objects [18] from internet photo collections. For human subjects, Weng et al. [37] introduce free-viewpoint human body rendering from an unstructured photo collection. Despite the photorealistic rendering quality of PersonNerf [37], they do not model animatable faces due to the static training images. In our work, we learn animatable 3D head avatars from video collections." + }, + { + "type": "image", + "bbox": [ + 0.22, + 0.367, + 0.785, + 0.525 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.214, + 0.535, + 0.788, + 0.631 + ], + "angle": 0, + "content": "Fig. 2: Overview of the proposed approach, PAV. (a) shows some training videos in various appearances of the same subject. (b) We estimate the 3D head geometry of the person using FLAME head model. (c) shows the proposed personalized deformable neural radiance field from video collection (d) We learn appearance-specific neural features and attach them to vertices. (d) This feature is passed to fully fused multi-layer perceptrons with additional conditioning on the facial expressions \\( e \\) and the encoded view direction \\( d \\)." + }, + { + "type": "title", + "bbox": [ + 0.216, + 0.688, + 0.33, + 0.703 + ], + "angle": 0, + "content": "3 Method" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.72, + 0.788, + 0.841 + ], + "angle": 0, + "content": "Our goal is to learn a personalized head avatar, PAV, from a monocular video collection of a subject under different appearances, poses, and facial expressions. To create the avatar, we use a small collection \\(\\mathbf{V} = \\{V_{i}\\}_{i = 1}^{N}\\) with \\(N\\) videos. We parameterize PAV using FLAME [19] face model to control head poses, facial expressions and use neural texture to model appearance. During inference, we only need the target FLAME parameters and neural texture. Learning PAV from a collection of monocular videos requires capturing the geometry and appearance of the dynamic head while also disentangling different appearances of the same" + } + ], + [ + { + "type": "page_number", + "bbox": [ + 0.218, + 0.116, + 0.23, + 0.127 + ], + "angle": 0, + "content": "6" + }, + { + "type": "header", + "bbox": [ + 0.272, + 0.115, + 0.39, + 0.128 + ], + "angle": 0, + "content": "A. Caliskan et al." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.147, + 0.788, + 0.269 + ], + "angle": 0, + "content": "identity. In order to tackle this problem, we first capture the dynamic 3D geometry by employing a parametric head model [19] (Sec. 3.1). We use the coarse geometry to model geometry-guided deformable neural radiance field (Sec. 3.2) and vertex-attached neural features to condition appearance for rendering (Sec. 3.3). We also propose estimating appearance-conditioned neural radiance field, aimed at accurately capturing distinct color and geometry with high fidelity. (Sec. 3.4). An overview of our method can be seen in Fig. 2. In the following, we describe each of these modules in greater detail." + }, + { + "type": "title", + "bbox": [ + 0.215, + 0.288, + 0.49, + 0.304 + ], + "angle": 0, + "content": "3.1 Head Geometry Estimation" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.31, + 0.789, + 0.432 + ], + "angle": 0, + "content": "Given the training frames \\(\\mathbf{I} = \\{I_{\\mathrm{i}}\\}\\), from all appearances, we first estimate the optimized camera parameters \\(\\mathbf{K} \\in \\mathbb{R}^{3 \\times 3}\\), tracked FLAME [19] meshes \\(\\mathbf{M} = \\{M_{\\mathrm{i}}\\}\\) with corresponding facial expression coefficients \\(\\mathbf{E} = \\{E_{\\mathrm{i}}\\}\\) and poses \\(\\mathbf{P} = \\{P_{\\mathrm{i}}\\}\\). We follow an optimization-based method, that minimizes an objective function with photo-consistency and landmark terms [43]. Different from the original work [43] which is designed for a single appearance, in PAV, we compute the average shape parameters from all appearances of the same subject to represent identity shape during optimization." + }, + { + "type": "title", + "bbox": [ + 0.215, + 0.451, + 0.524, + 0.466 + ], + "angle": 0, + "content": "3.2 Personalized Deformable NeRF" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.473, + 0.789, + 0.565 + ], + "angle": 0, + "content": "Our goal is to create a head avatar that can be learned from frames in video collection of various appearance of the same subject. We parameterize PAV using FLAME model and learned appearance conditioning. Specifically, PAV takes target head pose \\(\\theta\\), facial expression \\(e\\) and latent neural features \\((LNF)\\) \\(Z_{j} \\in \\mathbb{R}^{512 \\times 512 \\times 16}\\) in the texture atlas space (UV space) for an appearance \\(j\\) as inputs and synthesizes the target image \\(I^{\\theta, e, Z_{j}}\\) as:" + }, + { + "type": "equation", + "bbox": [ + 0.413, + 0.572, + 0.786, + 0.589 + ], + "angle": 0, + "content": "\\[\nI ^ {\\theta , e, Z _ {j}} = \\operatorname {P A V} (\\theta , e, Z _ {j}, K), \\tag {1}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.597, + 0.789, + 0.842 + ], + "angle": 0, + "content": "where \\(K\\) corresponds to the intrinsic parameters of the target camera viewpoint. For this purpose, we are using a geometry-guided dynamic deformable neural radiance field. However, deformable head NeRF models cannot disentangle appearance from the neural radiance field (See Sec. 2). To tackle this problem, we employ shared canonical space for all appearances where a shared neural radiance field is constructed. With this shared radiance field at canonical space, to render the target head pose and facial expression using volumetric rendering, we canon- icalize the sample points \\(x\\) on a ray from the observation space to query points \\(x_{c}\\) in the shared canonical space radiance field by using deformation gradient \\(\\hat{F}\\) .. \\(x_{c} = \\hat{F}\\cdot x\\) where \\(\\hat{F}\\in \\mathbb{R}^{4\\times 4}\\) and \\(\\{x,x_c\\} \\in \\mathbb{R}^4\\) . We employ the nearest triangle search for a given point to compute the deformation gradient. Specifically, the deformation gradient is computed between the deformed triangle \\(\\hat{T}_{def}\\in M_i\\) in the observed mesh and the canonical triangle \\(\\hat{T}_{canon}\\in M^{canon}\\) in the canonical mesh. To increase the speed of nearest triangle search, we employ bounding volume hierarchy (BVH) [6]. We refer readers to [44] for a detailed description of deformation gradient computation." + } + ], + [ + { + "type": "header", + "bbox": [ + 0.369, + 0.115, + 0.731, + 0.128 + ], + "angle": 0, + "content": "PAV: Personalized Head Avatar from Video Collection" + }, + { + "type": "page_number", + "bbox": [ + 0.775, + 0.117, + 0.785, + 0.127 + ], + "angle": 0, + "content": "7" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.147, + 0.784, + 0.238 + ], + "angle": 0, + "content": "After canonicalization, given the predicted appearance-conditioned radiance \\( c \\) and density \\( \\sigma \\) (Sec. 3.4) for every point on each ray, the representation of the avatar is optimized by using the differentiable volumetric rendering. For each camera ray \\( r(t) = o + td \\) with camera center \\( o \\) and viewing direction \\( d \\), the color of the corresponding pixel can be computed by accumulating the predicting radiance and densities of the sample points along the ray:" + }, + { + "type": "equation", + "bbox": [ + 0.376, + 0.25, + 0.786, + 0.284 + ], + "angle": 0, + "content": "\\[\n\\mathcal {C} (\\mathbf {r}) = \\int_ {t _ {n}} ^ {t _ {f}} \\sigma (\\mathbf {r} (t)) \\cdot \\mathbf {c} (\\mathbf {r} (t)) \\cdot \\mathcal {T} (t) d t \\tag {2}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.215, + 0.289, + 0.784, + 0.339 + ], + "angle": 0, + "content": "where \\( t_n \\) and \\( t_f \\) are the near and far bounds correspondingly, and \\( \\mathcal{T}(t_n) = \\exp \\left(-\\int_{t_n}^{t_f}\\sigma (t)dt\\right) \\) is the accumulated transmittance along the ray from \\( t_n \\) to \\( t_f \\), \\( \\sigma (t) \\) is the density and \\( c(t) \\) it the radiance at position \\( x_t \\)." + }, + { + "type": "title", + "bbox": [ + 0.215, + 0.36, + 0.533, + 0.377 + ], + "angle": 0, + "content": "3.3 Learned Appearance Embedding" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.385, + 0.788, + 0.611 + ], + "angle": 0, + "content": "In this section, PAV introduces a learnable appearance embedding that is essential and effective for enabling discriminative multi-appearance rendering using single unified network. To achieve this goal (See Fig. 2), a set of \\(N\\) learnable latent features \\(\\mathbf{Z} = \\{Z_{\\mathrm{i}}\\}_{i = 1}^{N}\\), \\(Z_{i}\\in \\mathbb{R}^{512\\times 512\\times 16}\\) are defined in UV space corresponding to \\(N\\) appearances of the same subject in the video collection. We impose high-dimensional learnable latent features to condition implicit representations for both density \\(\\sigma\\) and radiance \\(c\\) values in the canonical space. To this end, these latent neural features are attached to the vertices of 3DMM mesh, which enables PAV to memorize local variations. For each point \\(x\\) sampled on a ray, we find the nearest triangle \\(\\mathcal{T} = \\{v_{\\mathrm{i}}\\}_{i = 1}^{3}, v_{i}\\in \\mathbb{R}^{3}\\) (sec. 3.2) and aggregate the features \\(z(v_{i})\\) of vertices \\(v_{i}\\) from the vertices of the nearest triangle to get the appearance embedding \\(Z^x\\). Then we directly use the appearance embedding to condition for appearance-dependent neural radiance field estimation (Sec. 3.4). During training, these features are randomly initialized and optimized jointly with the objectives of the entire network." + }, + { + "type": "title", + "bbox": [ + 0.215, + 0.633, + 0.523, + 0.649 + ], + "angle": 0, + "content": "3.4 Appearance Conditioned NeRF" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.658, + 0.788, + 0.794 + ], + "angle": 0, + "content": "To model the appearance-conditioned dynamic deformable head NeRF, PAV incorporates learned appearance-embedding (Sec. 3.3) into neural radiance field in canonical space (See Fig. 2). For a point \\( x \\) sampled on a ray, after canonicalization, the point \\( x_{c} \\) is encoded using multi-resolution hash encoding [24]. This feature \\( \\phi(x_{c}) \\) is passed to multi-layer perceptron for geometric feature estimation. Following [10] and [44], we conditioned every point features \\( \\phi(x_{c}) \\) with the facial expressions \\( e \\) of frame \\( i \\). Without per-frame latent code, for shared volume density, we learn implicit representation \\( \\mathcal{F}_{\\text{density}} \\) using multi-layer perceptron (MLP-D in Fig. 2):" + }, + { + "type": "equation", + "bbox": [ + 0.414, + 0.822, + 0.785, + 0.837 + ], + "angle": 0, + "content": "\\[\n\\mathcal {F} _ {\\text {d e n s i t y}}: \\left(\\phi \\left(x _ {c}\\right), e _ {i}\\right)\\rightarrow \\sigma \\tag {3}\n\\]" + } + ], + [ + { + "type": "page_number", + "bbox": [ + 0.218, + 0.116, + 0.23, + 0.127 + ], + "angle": 0, + "content": "8" + }, + { + "type": "header", + "bbox": [ + 0.272, + 0.115, + 0.39, + 0.128 + ], + "angle": 0, + "content": "A. Caliskan et al." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.147, + 0.788, + 0.239 + ], + "angle": 0, + "content": "A primary challenge lies in learning unique density estimations for each appearance while separating the shared density aspects from the appearance. Hence we design PAV such that appearance-based density offset is predicted from appearance embedding \\( Z^{x} \\) sampled from the nearest triangle to the point \\( x \\) on the mesh \\( M_{i} \\) and facial expression \\( e_{i} \\) of frame \\( i \\). For density offset, we learn implicit representation \\( \\mathcal{F}_{offset} \\) using multi-layer perceptron (MLP-O in Fig. 2):" + }, + { + "type": "equation", + "bbox": [ + 0.419, + 0.251, + 0.786, + 0.267 + ], + "angle": 0, + "content": "\\[\n\\mathcal {F} _ {\\text {o f f s e t}}: \\left(Z ^ {x}, e _ {i}\\right)\\rightarrow \\Delta_ {\\sigma} \\tag {4}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.278, + 0.788, + 0.338 + ], + "angle": 0, + "content": "Final density value \\(\\sigma(x) = \\sigma + \\Delta_{\\sigma}\\) for the point \\(x\\) is the summation of shared density features \\(\\sigma\\) and density offset \\(\\Delta_{\\sigma}\\). For emitted radiance at point \\(x\\), we learn implicit representation \\(\\mathcal{F}_{color}\\) using multi-layer perceptron (MLP-C in Fig. 2):" + }, + { + "type": "equation", + "bbox": [ + 0.411, + 0.351, + 0.786, + 0.368 + ], + "angle": 0, + "content": "\\[\n\\mathcal {F} _ {\\text {c o l o r}}: \\left(Z ^ {x}, d _ {i}, \\sigma (x)\\right)\\rightarrow c \\tag {5}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.215, + 0.378, + 0.787, + 0.41 + ], + "angle": 0, + "content": "where \\( d_{i} \\) is the encoded view direction of training frame \\( i \\). Then, final color for the pixel is computed by using differentiable volume rendering (Eq. 2)" + }, + { + "type": "title", + "bbox": [ + 0.215, + 0.431, + 0.373, + 0.447 + ], + "angle": 0, + "content": "3.5 Optimization" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.456, + 0.788, + 0.592 + ], + "angle": 0, + "content": "Our model is trained on a collection of monocular RGB videos with photometric and depth losses. For photometric loss \\(\\mathcal{L}_{color}\\), we employ \\(L_{2}\\) distance between the renderings \\(C\\) and ground-truth images \\(I\\), \\(\\mathcal{L}_{color} = \\sum_{i}\\sum_{r}\\| C_{i}(r) - I_{i}(r)\\|\\), where \\(r\\) denotes the camera ray of each pixel and \\(i\\) denotes frame index. For depth loss \\(\\mathcal{L}_{depth}\\), we rasterize the depth of tracking mesh \\(M_{i}\\) and apply \\(L_{1}\\) distance between this map \\(D\\) and the ray termination \\(H\\) of the volumetric rendering only for face region as FLAME model is restricted to model details like hair: \\(\\mathcal{L}_{depth} = \\sum_{i}\\sum_{r}\\mathcal{M}_{i}(r)\\| H_{i}(r) - D_{i}(r)\\|\\), where \\(\\mathcal{M}\\) denotes face binary mask. The total loss is defined as:" + }, + { + "type": "equation", + "bbox": [ + 0.392, + 0.606, + 0.786, + 0.621 + ], + "angle": 0, + "content": "\\[\n\\mathcal {L} = \\lambda_ {\\text {c o l o r}} \\mathcal {L} _ {\\text {c o l o r}} + \\lambda_ {\\text {d e p t h}} \\mathcal {L} _ {\\text {d e p t h}} \\tag {6}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.215, + 0.632, + 0.597, + 0.649 + ], + "angle": 0, + "content": "where we empirically set \\(\\lambda_{color} = 50\\) and \\(\\lambda_{depth} = 10\\)." + }, + { + "type": "title", + "bbox": [ + 0.215, + 0.671, + 0.377, + 0.687 + ], + "angle": 0, + "content": "4 Experiments" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.702, + 0.787, + 0.763 + ], + "angle": 0, + "content": "In this section, we evaluate the performance of PAV using two different datasets. We perform an ablation study to validate our design choices and a comparison against the state-of-the-art monocular head avatar method in various configurations." + }, + { + "type": "title", + "bbox": [ + 0.215, + 0.785, + 0.328, + 0.799 + ], + "angle": 0, + "content": "4.1 Dataset" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.81, + 0.787, + 0.842 + ], + "angle": 0, + "content": "Our method takes a video collection of a subject containing different appearances, poses, and expressions. Since there is no existing dataset to evaluate" + } + ], + [ + { + "type": "header", + "bbox": [ + 0.37, + 0.115, + 0.732, + 0.128 + ], + "angle": 0, + "content": "PAV: Personalized Head Avatar from Video Collection" + }, + { + "type": "page_number", + "bbox": [ + 0.775, + 0.117, + 0.785, + 0.127 + ], + "angle": 0, + "content": "9" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.147, + 0.784, + 0.296 + ], + "angle": 0, + "content": "learning 3D head avatar from video collections of the same subjects with various appearances, poses, and expressions, we compiled a new Video Collection (VidCol) dataset from talking face videos of Roger Federer, King Charles and Tom Hanks across years. This dataset consists of in-the-wild videos recorded in uncontrolled environments. In VidCol dataset, we curated videos from 5 different appearances for each subject. We randomly sampled 500 frames at \\(512^2\\) resolution from each video (or appearance) and we used a total of 2500 frames to train and test a single unified network per subject. We will share this dataset with annotations of semantic masks, tracked 3DMM models, and camera calibrations for further research in this direction." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.298, + 0.784, + 0.329 + ], + "angle": 0, + "content": "For our experiments, we evaluate all characters and train a separate PAV model for each subject. We develop two different protocols for evaluation:" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.342, + 0.784, + 0.448 + ], + "angle": 0, + "content": "a) Unified Network for Novel Pose/Exp. Synthesis This protocol evaluates the performance in terms of novel appearance, head pose, and facial expression synthesis when training a single unified network from all appearances. For training, we select 400 frames for each appearance of a subject talking in front of a camera with various head poses. It mimics a training video commonly used for monocular neural head avatar methods (Sec. 2). For testing, we use the remaining 100 frames for each appearance with unseen poses and expressions." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.461, + 0.785, + 0.569 + ], + "angle": 0, + "content": "b) Separated Networks for Novel Pose/Exp. Synthesis This protocol evaluates the novel appearance, head pose, and facial expression generalization ability of the methods when training a separate network for each appearance. We select 400 training images for each appearance of a subject talking of a camera, same as protocol-a. For testing, we choose the remaining 100 unseen pose/expression combinations with ground truth, same as protocol-a, and another 300 unseen pose/expression sets without ground truth image data." + }, + { + "type": "title", + "bbox": [ + 0.215, + 0.589, + 0.327, + 0.603 + ], + "angle": 0, + "content": "4.2 Metrics" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.613, + 0.784, + 0.72 + ], + "angle": 0, + "content": "We report several metrics to evaluate the quality of synthesized images as well as the ablation study of our method. For synthesized images with provided ground-truth, we use Learned Perceptual Patch Similarity (LPIPS) [41], Deep Image Structure and Texture Similarity (DISTS) [8], Structural Similarity Index (SSIM) [36], Peak Signal-to-Noise Ratio (PSNR) and Fréchet Inception Distance (FID) [15]. When there is an absence of ground truth, we adopt FID metric for quantitative comparison." + }, + { + "type": "title", + "bbox": [ + 0.215, + 0.741, + 0.457, + 0.755 + ], + "angle": 0, + "content": "4.3 Implementation Details" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.765, + 0.784, + 0.84 + ], + "angle": 0, + "content": "We optimize Eq. 6 using Adam optimizer [17] with hyperparameters \\(\\beta_{1} = 0.9\\) and \\(\\beta_{2} = 0.99\\). We set the learning rate to \\(2.5 \\times 10^{-3}\\) for \\(F_{density}\\), \\(F_{color}\\) and \\(Z\\), and to \\(2.5 \\times 10^{-4}\\) for \\(F_{offset}\\). We sample 1024 points along each ray for rendering. Multi-resolution hash encoding used in the method is set to 16 number of levels, 18 hash map size, and 2048 finest resolution. The optimization takes 15K iterations" + } + ], + [ + { + "type": "page_number", + "bbox": [ + 0.218, + 0.116, + 0.236, + 0.127 + ], + "angle": 0, + "content": "10" + }, + { + "type": "header", + "bbox": [ + 0.272, + 0.115, + 0.39, + 0.127 + ], + "angle": 0, + "content": "A. Caliskan et al." + }, + { + "type": "image", + "bbox": [ + 0.235, + 0.142, + 0.79, + 0.457 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.214, + 0.466, + 0.788, + 0.577 + ], + "angle": 0, + "content": "Fig. 3: Qualitative Ablation Study on VidCol Dataset. This figure shows synthesized images and pixel error maps against ground-truth. Please note that brighter pixel colors denote higher pixel errors. Without appearance embedding with latent neural features and appearance-conditioned density offset (a), the model fails to learn appearance-based distinct details. The latent neural features (LNF) (c) resolve the issue of high-fidelity rendering of texture (blue arrow). The appearance-conditioned density offset field (d) allows the model to learn more accurate geometry and texture that corresponds to the face, hair, and neck regions (red arrow)" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.608, + 0.785, + 0.639 + ], + "angle": 0, + "content": "to converge when training network with all appearances of single subjects. More details on the network are provided in the supplementary document." + }, + { + "type": "title", + "bbox": [ + 0.215, + 0.663, + 0.391, + 0.679 + ], + "angle": 0, + "content": "4.4 Ablation Study" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.689, + 0.788, + 0.841 + ], + "angle": 0, + "content": "We evaluate different design choices of PAV in Tab. 1 and Fig. 3. In this study, we use protocol-a of the VidCol dataset for all experiments. We first report the results of the final model which includes latent neural features (LNF) (Sec. 3.3) and density offset \\(\\varDelta_{\\sigma}\\) (Sec. 3.4). The full model achieves PSNR score of 25.15 and SSIM score of 0.91 for image synthesis. If we remove the density offset from our model while employing appearance-conditioned radiance estimation, PSNR and SSIM scores for image synthesis decrease to 24.14 and 0.90, respectively. This indicates that without appearance-conditioned density offset \\(\\varDelta_{\\sigma}\\), density estimation MLP (\\(\\mathcal{F}_{\\text {density }}\\) in Sec. 3.4) can not learn to synthesize appearance-specific geometry and texture details, thus rendering fails there as well. Even" + } + ], + [ + { + "type": "header", + "bbox": [ + 0.37, + 0.115, + 0.732, + 0.128 + ], + "angle": 0, + "content": "PAV: Personalized Head Avatar from Video Collection" + }, + { + "type": "page_number", + "bbox": [ + 0.769, + 0.117, + 0.784, + 0.127 + ], + "angle": 0, + "content": "11" + }, + { + "type": "table", + "bbox": [ + 0.25, + 0.144, + 0.76, + 0.23 + ], + "angle": 0, + "content": "
PSNR ↑SSIM ↑DISTS ↓LPIPS ↓FID ↓
Insta [44]21.030.8490.1850.169104.28
PAV without LNF24.190.9000.1450.11954.39
PAV without Δσ24.140.9060.1340.10847.71
PAV (Full Model)25.150.9100.1330.10746.67
" + }, + { + "type": "table_caption", + "bbox": [ + 0.216, + 0.232, + 0.785, + 0.26 + ], + "angle": 0, + "content": "Table 1: Ablation Study: We evaluate the impact of different components of the proposed method. See Fig. 3 for qualitative comparison." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.304, + 0.784, + 0.379 + ], + "angle": 0, + "content": "though the model \\(\\mathrm{w / o}\\varDelta_{\\sigma}\\) synthesise images that look on par with ground truth in terms of texture, it generates severe artifacts which can be seen in pixel error maps. An example of this behavior can also be seen in Fig. 3 \\((\\mathrm{w / o}\\varDelta_{\\sigma})\\) Rendering failures on face, head, and neck regions are indicated by red arrow and error map." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.38, + 0.785, + 0.606 + ], + "angle": 0, + "content": "Next, we evaluate the impact of latent neural features (LNF) as an appearance embedding on PAV. In this experiment, to condition the canonical MLPs in PAV, inspired by \"Nerf In-the-wild\" [22, 37], we adopt Generative Latent Optimization (GLO) for appearance embedding. Specifically, this approach defines a single real-valued appearance embedding vector \\( l \\) for each appearance in the dataset. To this end, we employ a pre-trained encoder from in-the-wild face detector RetinaFace [7] for GLO. During training, if we replace \\( \\mathbf{Z} = \\{Z_{\\mathrm{i}}\\}_{i=1}^{N} \\), \\( Z_{i} \\in \\mathbb{R}^{512 \\times 512 \\times 16} \\) with \\( \\mathbf{L} = \\{l_{\\mathrm{i}}\\}_{i=1}^{N} \\), \\( l_{i} \\in \\mathbb{R}^{512 \\times 1} \\) for \\( N \\) appearances of each subject, and pass it to appearance conditioned NeRF (See Sec. 3.4), PSNR and SSIM scores for image synthesis decrease to 24.19 and 0.90, respectively. This is also evident from the qualitative results shown in Fig. 3 (w/o LNF), indicating that latent neural features help in improved image synthesis quality with their local features attached to 3D geometry. Rendering failures on head, neck, and skin texture are shown by red arrow, blue arrow and error maps in Fig. 3. Finally, as discussed in Sec. 3, we keep the network design as we explain after the ablation study." + }, + { + "type": "title", + "bbox": [ + 0.216, + 0.632, + 0.362, + 0.646 + ], + "angle": 0, + "content": "4.5 Comparison" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.659, + 0.788, + 0.841 + ], + "angle": 0, + "content": "Since our method is the first work that learns dynamic deformable head NeRF from multiple appearances of the same subjects using a unified network, we ourselves design some baseline comparisons in this section. First, we compare our method against Insta [44] using protocol-a (Sec. 4.1) of the VidCol dataset. Both methods are trained using the frames sampled from all appearances of the same subject. For testing, both methods synthesize images for target head pose and expression. The results are summarized in Tab. 2 and Fig. 4. For protocol-a, our method clearly outperforms the baseline for final image synthesis results across all metrics. It should be noted that PAV framework jointly learns all different appearances via one-time training using a unified network and it is able to preserve appearance-specific details for novel poses and expressions. On the other hand, as shown in Fig. 4, Insta can not disentangle appearance from the" + } + ], + [ + { + "type": "page_number", + "bbox": [ + 0.218, + 0.116, + 0.236, + 0.127 + ], + "angle": 0, + "content": "12" + }, + { + "type": "header", + "bbox": [ + 0.272, + 0.115, + 0.39, + 0.128 + ], + "angle": 0, + "content": "A. Caliskan et al." + }, + { + "type": "image", + "bbox": [ + 0.223, + 0.143, + 0.805, + 0.493 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.216, + 0.502, + 0.788, + 0.545 + ], + "angle": 0, + "content": "Fig. 4: Qualitative comparison of our method (PAV) against baseline method (Insta [44]. Columns \\(\\{2,6\\}\\) show target head poses and facial expressions for synthesized images at columns \\(\\{3,4,7,8\\}\\)." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.58, + 0.787, + 0.626 + ], + "angle": 0, + "content": "neural radiance field, which results in a mixture of appearances in synthesized images and inaccurate geometry predictions. This is the limitation of Insta we are addressing in the paper." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.629, + 0.789, + 0.842 + ], + "angle": 0, + "content": "Second, we train separate Insta [44] models for each appearance of the same subject using protocol-b while training a single unified model for our method using the frames sampled from all appearances of the same subject. For testing, we use the same test data from protocol-b for both PAV and Insta [44] and synthesize images for novel poses and expressions. Note that, we added more novel expressions and poses in protocol-b (Sec. 4.1) to test the generalization of our method under a unified training schema, which results in increasing FID score for both methods in Tab. 3. Even though the results show that artifacts (in Fig. 4) of the Insta renders are decreased in Fig. 5, our method demonstrates better-rendering results (Fig. 5) and lower FID score (Tab. 3). This indicates that PAV's single unified model helps in learning generic neural radiance field representation across all appearances of the same subject while preserving appearance-based details. Please note that we are demonstrating Insta's performance under a challenging pose/expression in Fig. 5 which can also be seen in" + } + ], + [ + { + "type": "header", + "bbox": [ + 0.37, + 0.115, + 0.732, + 0.128 + ], + "angle": 0, + "content": "PAV: Personalized Head Avatar from Video Collection" + }, + { + "type": "page_number", + "bbox": [ + 0.769, + 0.117, + 0.785, + 0.127 + ], + "angle": 0, + "content": "13" + }, + { + "type": "table", + "bbox": [ + 0.22, + 0.146, + 0.788, + 0.214 + ], + "angle": 0, + "content": "
MethodVideo Collection 1Video Collection 2
PSNR ↑SSIM ↑DISTS ↓LPIPS ↓FID ↓PSNR ↑SSIM ↑DISTS ↓LPIPS ↓FID ↓
Protocol (a): Novel Head Pose and Expression Synthesis
Insta [44]21.030.840.1850.169104.2823.670.860.1870.16398.87
PAV (Ours)25.150.910.1330.10746.6727.090.890.1440.12053.91
" + }, + { + "type": "table_caption", + "bbox": [ + 0.216, + 0.215, + 0.784, + 0.255 + ], + "angle": 0, + "content": "Table 2: Comparison of PAV against Insta [44]. See Fig. 4 for qualitative comparison. Video Collection 1 and 2 refer to Federer and King Charles splits of VidCol dataset, respectively." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.298, + 0.785, + 0.404 + ], + "angle": 0, + "content": "the original paper results under extreme poses/expressions. In Fig. 5, the reference image is for photoconsistency check only since there is no ground truth image for that pose/expression. Finally, we provide additional qualitative results of avatars animated in the supplementary video. We also compared PAV against Insta for model efficiency using protocol-b dataset and recorded values in Tab. 3. Although we train our model with multiple appearances jointly, PAV is more efficient than Insta in terms of memory footprint and training time" + }, + { + "type": "image", + "bbox": [ + 0.303, + 0.428, + 0.404, + 0.586 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.33, + 0.588, + 0.377, + 0.597 + ], + "angle": 0, + "content": "Reference" + }, + { + "type": "image", + "bbox": [ + 0.414, + 0.436, + 0.493, + 0.508 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.423, + 0.512, + 0.495, + 0.585 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.411, + 0.588, + 0.493, + 0.597 + ], + "angle": 0, + "content": "Target Pose/Exp." + }, + { + "type": "image", + "bbox": [ + 0.515, + 0.43, + 0.585, + 0.507 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.515, + 0.508, + 0.588, + 0.585 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.528, + 0.588, + 0.573, + 0.598 + ], + "angle": 0, + "content": "Insta [44]" + }, + { + "type": "image", + "bbox": [ + 0.616, + 0.431, + 0.681, + 0.506 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.619, + 0.508, + 0.681, + 0.585 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.639, + 0.588, + 0.662, + 0.597 + ], + "angle": 0, + "content": "PAV" + }, + { + "type": "image_caption", + "bbox": [ + 0.215, + 0.609, + 0.785, + 0.652 + ], + "angle": 0, + "content": "Fig. 5: Qualitative comparison of PAV against single-appearance traines Insta [44]. Please note that the reference image is for photoconsistency check only since there is no ground truth image for that pose/expression" + }, + { + "type": "title", + "bbox": [ + 0.216, + 0.703, + 0.355, + 0.719 + ], + "angle": 0, + "content": "5 Discussion" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.734, + 0.788, + 0.841 + ], + "angle": 0, + "content": "We presented PAV which is a novel method for learning personalized head avatars from monocular and in-the-wild video collections of a person. We demonstrated that PAV can model a 3D head avatar from an unstructured video collection of a subject in a single unified network while disentangling face appearances from the radiance field. PAV can be trained simultaneously for multiple appearances of the same subject, allowing novel pose and expression generalization and leading to a superior rendering quality of our method compared to the baseline." + } + ], + [ + { + "type": "page_number", + "bbox": [ + 0.218, + 0.116, + 0.236, + 0.127 + ], + "angle": 0, + "content": "14" + }, + { + "type": "header", + "bbox": [ + 0.272, + 0.115, + 0.39, + 0.128 + ], + "angle": 0, + "content": "A. Caliskan et al." + }, + { + "type": "table", + "bbox": [ + 0.221, + 0.145, + 0.788, + 0.219 + ], + "angle": 0, + "content": "
Train Time MinutesNumber of Param. MillionModel Size (MB)FID ↓
Multiple Insta [44]13579.71932109.22
PAV (Ours)4526.865091.01
" + }, + { + "type": "table_caption", + "bbox": [ + 0.216, + 0.221, + 0.788, + 0.276 + ], + "angle": 0, + "content": "Table 3: Efficiency and performance comparison between PAV and Insta [44] trained for each appearance (Single Appearance Training) in dataset collection 1. FID score is computed from synthetized images under novel head poses and expressions where there are no existing ground-truth images." + }, + { + "type": "image", + "bbox": [ + 0.351, + 0.303, + 0.652, + 0.469 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.215, + 0.479, + 0.788, + 0.536 + ], + "angle": 0, + "content": "Fig. 6: Multi-identity and multi-appearance extension of PAV. The learned appearance (middle) rendered in a new viewpoint (left) does not match the visual quality of input images (right). Please note that the reference image is for photoconsistency check only since there is no ground truth image for that pose/expression." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.566, + 0.789, + 0.719 + ], + "angle": 0, + "content": "Limitation for Multi-Identity. We extended our method for multiple identities and multiple appearances for each identity. To this end, we train a single unified PAV model using frames smaped from all appearances of Federer and King subjects from VidCol dataset with shared canonical space. We evaluate novel head pose and facial expression rendering of this model and Fig. 6 illustrates the results. Even though this model produces reasonable results, synthesized images demonstrate a lack of subject-specific shape and appearance details with rendering artifacts. In the paper, we focus on multiple appearances of a single subject. Scaling up this model to a wide range of subjects and exploring generalized dynamic deformable NeRF would be interesting future work." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.734, + 0.789, + 0.841 + ], + "angle": 0, + "content": "Ethical Considerations. Although the rapid progress on digital human avatars facilitates numerous downstream applications, it also raises concerns about ethical misuse. On the synthesis side, it would be ideal to actively use watermarking technologies and avoid driving avatars with different identities. In this work, we aim to faithfully reproduce images of a person with the capability of rendering unseen poses and expressions and switching appearance within their own set of appearances. The work does not intend to create animations that didn't happen." + } + ], + [ + { + "type": "header", + "bbox": [ + 0.369, + 0.115, + 0.732, + 0.128 + ], + "angle": 0, + "content": "PAV: Personalized Head Avatar from Video Collection" + }, + { + "type": "page_number", + "bbox": [ + 0.768, + 0.117, + 0.786, + 0.127 + ], + "angle": 0, + "content": "15" + }, + { + "type": "title", + "bbox": [ + 0.217, + 0.145, + 0.324, + 0.161 + ], + "angle": 0, + "content": "References" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.176, + 0.786, + 0.218 + ], + "angle": 0, + "content": "1. Athar, S., Xu, Z., Sunkavalli, K., Shechtman, E., Shu, Z.: Rignerf: Fully controllable neural 3d portraits. In: Proceedings of the IEEE/CVF conference on Computer Vision and Pattern Recognition. pp. 20364-20373 (2022)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.219, + 0.786, + 0.247 + ], + "angle": 0, + "content": "2. Averbuch-Elor, H., Cohen-Or, D., Kopf, J., Cohen, M.F.: Bringing portraits to life. ACM transactions on graphics (TOG) 36(6), 1-13 (2017)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.247, + 0.786, + 0.302 + ], + "angle": 0, + "content": "3. Bai, Z., Tan, F., Huang, Z., Sarkar, K., Tang, D., Qiu, D., Meka, A., Du, R., Dou, M., Orts-Escalano, S., et al.: Learning personalized high quality volumetric head avatars from monocular rgb videos. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 16890-16900 (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.302, + 0.786, + 0.342 + ], + "angle": 0, + "content": "4. Bühler, M.C., Meka, A., Li, G., Beeler, T., Hilliges, O.: Varitex: Variational neural face textures. In: Proceedings of the IEEE/CVF International Conference on Computer Vision. pp. 13890-13899 (2021)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.343, + 0.786, + 0.385 + ], + "angle": 0, + "content": "5. Chen, C., O'Toole, M., Bharaj, G., Garrido, P.: Implicit neural head synthesis via controllable local deformation fields. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 416-426 (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.385, + 0.786, + 0.412 + ], + "angle": 0, + "content": "6. Clark, J.H.: Hierarchical geometric models for visible surface algorithms. Communications of the ACM 19(10), 547-554 (1976)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.412, + 0.786, + 0.439 + ], + "angle": 0, + "content": "7. Deng, J., Guo, J., Zhou, Y., Yu, J., Kotsia, I., Zafeiriou, S.: Retinaface: Single-stage dense face localisation in the wild. arXiv preprint arXiv:1905.00641 (2019)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.44, + 0.786, + 0.467 + ], + "angle": 0, + "content": "8. Ding, K., Ma, K., Wang, S., Simoncelli, E.P.: Image Quality Assessment: Unifying Structure and Texture Similarity. TPAMI (2022)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.468, + 0.786, + 0.522 + ], + "angle": 0, + "content": "9. Egger, B., Smith, W.A., Tewari, A., Wuhrer, S., Zollhoefer, M., Beeler, T., Bernard, F., Bolkart, T., Kortylewski, A., Romdhani, S., et al.: 3d morphable face models—past, present, and future. ACM Transactions on Graphics (ToG) 39(5), 1-38 (2020)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.522, + 0.786, + 0.578 + ], + "angle": 0, + "content": "0. Gafni, G., Thies, J., Zollhöfer, M., Nießner, M.: Dynamic neural radiance fields for monocular 4d facial avatar reconstruction. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR). pp. 8649-8658 (June 2021)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.578, + 0.786, + 0.619 + ], + "angle": 0, + "content": "1. Gao, X., Zhong, C., Xiang, J., Hong, Y., Guo, Y., Zhang, J.: Reconstructing personalized semantic facial nerf models from monocular video. ACM Transactions on Graphics (TOG) 41(6), 1-12 (2022)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.62, + 0.786, + 0.66 + ], + "angle": 0, + "content": "2. Garrido, P., Valgaerts, L., Rehmsen, O., Thormahlen, T., Perez, P., Theobalt, C.: Automatic face reenactment. In: Proceedings of the IEEE conference on computer vision and pattern recognition. pp. 4217-4224 (2014)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.661, + 0.786, + 0.716 + ], + "angle": 0, + "content": "3. Garrido, P., Valgaerts, L., Sarmadi, H., Steiner, I., Varanasi, K., Perez, P., Theobalt, C.: Vdub: Modifying face video of actors for plausible visual alignment to a dubbed audio track. In: Computer graphics forum. vol. 34, pp. 193-204. Wiley Online Library (2015)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.717, + 0.786, + 0.758 + ], + "angle": 0, + "content": "4. Grassal, P.W., Prinzler, M., Leistner, T., Rother, C., Nießner, M., Thies, J.: Neural head avatars from monocular rgb videos. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 18653-18664 (2022)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.759, + 0.786, + 0.799 + ], + "angle": 0, + "content": "5. Heusel, M., Ramsauer, H., Unterthiner, T., Nessler, B., Hochreiter, S.: Gans trained by a two time-scale update rule converge to a local nash equilibrium. Advances in neural information processing systems 30 (2017)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.799, + 0.786, + 0.841 + ], + "angle": 0, + "content": "6. Kim, H., Garrido, P., Tewari, A., Xu, W., Thies, J., Niessner, M., Pérez, P., Richardt, C., Zollhöfer, M., Theobalt, C.: Deep video portraits. ACM transactions on graphics (TOG) 37(4), 1-14 (2018)" + }, + { + "type": "list", + "bbox": [ + 0.226, + 0.176, + 0.786, + 0.841 + ], + "angle": 0, + "content": null + } + ], + [ + { + "type": "page_number", + "bbox": [ + 0.218, + 0.116, + 0.236, + 0.127 + ], + "angle": 0, + "content": "16" + }, + { + "type": "header", + "bbox": [ + 0.272, + 0.115, + 0.389, + 0.127 + ], + "angle": 0, + "content": "A. Caliskan et al." + }, + { + "type": "ref_text", + "bbox": [ + 0.218, + 0.148, + 0.785, + 0.175 + ], + "angle": 0, + "content": "17. Kingma, D.P., Ba, J.: Adam: A method for stochastic optimization. arXiv preprint arXiv:1412.6980 (2014)" + }, + { + "type": "ref_text", + "bbox": [ + 0.218, + 0.176, + 0.785, + 0.217 + ], + "angle": 0, + "content": "18. Kuang, Z., Olszewski, K., Chai, M., Huang, Z., Achlioptas, P., Tulyakov, S.: Neroic: Neural rendering of objects from online image collections. ACM Transactions on Graphics (TOG) 41(4), 1-12 (2022)" + }, + { + "type": "ref_text", + "bbox": [ + 0.218, + 0.218, + 0.785, + 0.259 + ], + "angle": 0, + "content": "19. Li, T., Bolkart, T., Black, M.J., Li, H., Romero, J.: Learning a model of facial shape and expression from 4D scans. ACM Transactions on Graphics, (Proc. SIGGRAPH Asia) 36(6), 194:1-194:17 (2017), https://doi.org/10.1145/3130800.3130813" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.26, + 0.785, + 0.3 + ], + "angle": 0, + "content": "20. Liu, L., Habermann, M., Rudnev, V., Sarkar, K., Gu, J., Theobalt, C.: Neural actor: Neural free-view synthesis of human actors with pose control. ACM transactions on graphics (TOG) 40(6), 1-16 (2021)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.301, + 0.785, + 0.327 + ], + "angle": 0, + "content": "21. Mallya, A., Wang, T.C., Liu, M.Y.: Implicit warping for animation with image sets. Advances in Neural Information Processing Systems 35, 22438-22450 (2022)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.328, + 0.785, + 0.383 + ], + "angle": 0, + "content": "22. Martin-Brualla, R., Radwan, N., Sajjadi, M.S., Barron, J.T., Dosovitskiy, A., Duckworth, D.: Nerf in the wild: Neural radiance fields for unconstrained photo collections. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 7210-7219 (2021)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.384, + 0.785, + 0.424 + ], + "angle": 0, + "content": "23. Mildenhall, B., Srinivasan, P.P., Tancik, M., Barron, J.T., Ramamoorthi, R., Ng, R.: Nerf: Representing scenes as neural radiance fields for view synthesis. Communications of the ACM 65(1), 99-106 (2021)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.425, + 0.785, + 0.466 + ], + "angle": 0, + "content": "24. Müller, T., Evans, A., Schied, C., Keller, A.: Instant neural graphics primitives with a multiresolution hash encoding. ACM Transactions on Graphics (ToG) 41(4), 1-15 (2022)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.467, + 0.785, + 0.508 + ], + "angle": 0, + "content": "25. Nirkin, Y., Keller, Y., Hassner, T.: Fsgan: Subject agnostic face swapping and reenactment. In: Proceedings of the IEEE/CVF international conference on computer vision. pp. 7184-7193 (2019)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.509, + 0.785, + 0.55 + ], + "angle": 0, + "content": "26. Park, K., Sinha, U., Hedman, P., Barron, J.T., Bouaziz, S., Goldman, D.B., Martin-Brualla, R., Seitz, S.M.: Hypernerf: A higher-dimensional representation for topologically varying neural radiance fields. arXiv preprint arXiv:2106.13228 (2021)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.551, + 0.785, + 0.591 + ], + "angle": 0, + "content": "27. Peng, S., Dong, J., Wang, Q., Zhang, S., Shuai, Q., Zhou, X., Bao, H.: Animatable neural radiance fields for modeling dynamic human bodies. In: Proceedings of the IEEE/CVF International Conference on Computer Vision. pp. 14314-14323 (2021)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.592, + 0.785, + 0.633 + ], + "angle": 0, + "content": "28. Pumarola, A., Corona, E., Pons-Moll, G., Moreno-Noguer, F.: D-nerf: Neural radiance fields for dynamic scenes. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 10318-10327 (2021)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.634, + 0.785, + 0.673 + ], + "angle": 0, + "content": "29. Siarohin, A., Lathuilière, S., Tulyakov, S., Ricci, E., Sebe, N.: First order motion model for image animation. Advances in neural information processing systems 32 (2019)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.675, + 0.785, + 0.729 + ], + "angle": 0, + "content": "30. Tancik, M., Casser, V., Yan, X., Pradhan, S., Mildenhall, B., Srinivasan, P.P., Barron, J.T., Kretzschmar, H.: Block-nerf: Scalable large scene neural view synthesis. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 8248-8258 (2022)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.729, + 0.785, + 0.785 + ], + "angle": 0, + "content": "31. Tewari, A., Fried, O., Thies, J., Sitzmann, V., Lombardi, S., Sunkavalli, K., MartinBrualla, R., Simon, T., Saragih, J., Nießner, M., et al.: State of the art on neural rendering. In: Computer Graphics Forum. vol. 39, pp. 701-727. Wiley Online Library (2020)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.786, + 0.785, + 0.84 + ], + "angle": 0, + "content": "32. Tewari, A., Thies, J., Mildenhall, B., Srinivasan, P., Tretschk, E., Yifan, W., Lassner, C., Sitzmann, V., Martin-Brualla, R., Lombardi, S., et al.: Advances in neural rendering. In: Computer Graphics Forum. vol. 41, pp. 703-735. Wiley Online Library (2022)" + }, + { + "type": "list", + "bbox": [ + 0.217, + 0.148, + 0.785, + 0.84 + ], + "angle": 0, + "content": null + } + ], + [ + { + "type": "header", + "bbox": [ + 0.37, + 0.115, + 0.732, + 0.128 + ], + "angle": 0, + "content": "PAV: Personalized Head Avatar from Video Collection" + }, + { + "type": "page_number", + "bbox": [ + 0.769, + 0.117, + 0.785, + 0.127 + ], + "angle": 0, + "content": "17" + }, + { + "type": "ref_text", + "bbox": [ + 0.214, + 0.147, + 0.788, + 0.175 + ], + "angle": 0, + "content": "33. Thies, J., Zollhöfer, M., Nießner, M.: Deferred neural rendering: Image synthesis using neural textures. Acm Transactions on Graphics (TOG) 38(4), 1-12 (2019)" + }, + { + "type": "ref_text", + "bbox": [ + 0.214, + 0.177, + 0.788, + 0.217 + ], + "angle": 0, + "content": "34. Thies, J., Zollhofer, M., Stamminger, M., Theobalt, C., Nieflner, M.: Face2face: Real-time face capture and reenactment of rgb videos. In: Proceedings of the IEEE conference on computer vision and pattern recognition. pp. 2387-2395 (2016)" + }, + { + "type": "ref_text", + "bbox": [ + 0.216, + 0.218, + 0.787, + 0.272 + ], + "angle": 0, + "content": "35. Wang, T.C., Liu, M.Y., Zhu, J.Y., Tao, A., Kautz, J., Catanzaro, B.: High-resolution image synthesis and semantic manipulation with conditional gans. In: Proceedings of the IEEE conference on computer vision and pattern recognition. pp. 8798-8807 (2018)" + }, + { + "type": "ref_text", + "bbox": [ + 0.216, + 0.273, + 0.787, + 0.3 + ], + "angle": 0, + "content": "36. Wang, Z., Bovik, A.C., Sheikh, H.R., Simoncelli, E.P.: Image quality assessment: From error visibility to structural similarity. TIP (2004)" + }, + { + "type": "ref_text", + "bbox": [ + 0.216, + 0.301, + 0.787, + 0.355 + ], + "angle": 0, + "content": "37. Weng, C.Y., Srinivasan, P.P., Curless, B., Kemelmacher-Shlizerman, I.: Person-nerf: Personalized reconstruction from photo collections. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 524-533 (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.216, + 0.357, + 0.787, + 0.411 + ], + "angle": 0, + "content": "38. Yang, B., Bao, C., Zeng, J., Bao, H., Zhang, Y., Cui, Z., Zhang, G.: Neumesh: Learning disentangled neural mesh-based implicit field for geometry and texture editing. In: European Conference on Computer Vision. pp. 597-614. Springer (2022)" + }, + { + "type": "ref_text", + "bbox": [ + 0.216, + 0.412, + 0.787, + 0.467 + ], + "angle": 0, + "content": "39. Zakharov, E., Ivakhnenko, A., Shysheya, A., Lempitsky, V.: Fast bi-layer neural synthesis of one-shot realistic head avatars. In: Computer Vision-ECCV 2020: 16th European Conference, Glasgow, UK, August 23-28, 2020, Proceedings, Part XII 16. pp. 524-540. Springer (2020)" + }, + { + "type": "ref_text", + "bbox": [ + 0.216, + 0.468, + 0.787, + 0.522 + ], + "angle": 0, + "content": "40. Zhang, B., Qi, C., Zhang, P., Zhang, B., Wu, H., Chen, D., Chen, Q., Wang, Y., Wen, F.: Metaportrait: Identity-preserving talking head generation with fast personalized adaptation. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR). pp. 22096-22105 (June 2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.216, + 0.523, + 0.787, + 0.55 + ], + "angle": 0, + "content": "41. Zhang, R., Isola, P., Efros, A.A., Shechtman, E., Wang, O.: The unreasonable effectiveness of deep features as a perceptual metric. In: CVPR (2018)" + }, + { + "type": "ref_text", + "bbox": [ + 0.216, + 0.551, + 0.787, + 0.605 + ], + "angle": 0, + "content": "42. Zheng, Y., Abrevaya, V.F., Bühler, M.C., Chen, X., Black, M.J., Hilliges, O.: Im avatar: Implicit morphable head avatars from videos. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 13545-13555 (2022)" + }, + { + "type": "ref_text", + "bbox": [ + 0.216, + 0.606, + 0.787, + 0.633 + ], + "angle": 0, + "content": "43. Zielonka, W., Bolkart, T., Thies, J.: Towards metrical reconstruction of human faces. In: European Conference on Computer Vision. pp. 250-269. Springer (2022)" + }, + { + "type": "ref_text", + "bbox": [ + 0.216, + 0.634, + 0.787, + 0.675 + ], + "angle": 0, + "content": "44. Zielonka, W., Bolkart, T., Thies, J.: Instant volumetric head avatars. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 4574-4584 (2023)" + }, + { + "type": "list", + "bbox": [ + 0.214, + 0.147, + 0.788, + 0.675 + ], + "angle": 0, + "content": null + } + ] +] \ No newline at end of file diff --git a/2024/PAV_ Personalized Head Avatar from Unstructured Video Collection/75c53aba-b5dd-481e-8738-bfaa232bbce0_origin.pdf b/2024/PAV_ Personalized Head Avatar from Unstructured Video Collection/75c53aba-b5dd-481e-8738-bfaa232bbce0_origin.pdf new file mode 100644 index 0000000000000000000000000000000000000000..57881476003a5be844500417fcfd23d22689c5db --- /dev/null +++ b/2024/PAV_ Personalized Head Avatar from Unstructured Video Collection/75c53aba-b5dd-481e-8738-bfaa232bbce0_origin.pdf @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:46f1b4ae439345cc71fbdc52be81f218ddce6e250fd2633dc6ba78c44d583d9a +size 9987771 diff --git a/2024/PAV_ Personalized Head Avatar from Unstructured Video Collection/full.md b/2024/PAV_ Personalized Head Avatar from Unstructured Video Collection/full.md new file mode 100644 index 0000000000000000000000000000000000000000..ae058a9d59e4e75a981e519fae985ea6c2110e51 --- /dev/null +++ b/2024/PAV_ Personalized Head Avatar from Unstructured Video Collection/full.md @@ -0,0 +1,244 @@ +# PAV: Personalized Head Avatar from Unstructured Video Collection + +Akin Caliskan1, Berkay Kicanaoglu1, and Hyeongwoo Kim2 + +1 Flawless AI + +2 Imperial College London + +Abstract. We propose PAV, Personalized Head Avatar for the synthesis of human faces under arbitrary viewpoints and facial expressions. PAV introduces a method that learns a dynamic deformable neural radiance field (NeRF), in particular from a collection of monocular talking face videos of the same character under various appearance and shape changes. Unlike existing head NeRF methods that are limited to modeling such input videos on a per-appearance basis, our method allows for learning multi-appearance NeRFs, introducing appearance embedding for each input video via learnable latent neural features attached to the underlying geometry. Furthermore, the proposed appearance-conditioned density formulation facilitates the shape variation of the character, such as facial hair and soft tissues, in the radiance field prediction. To the best of our knowledge, our approach is the first dynamic deformable NeRF framework to model appearance and shape variations in a single unified network for multi-appearances of the same subject. We demonstrate experimentally that PAV outperforms the baseline method in terms of visual rendering quality in our quantitative and qualitative studies on various subjects. + +# 1 Introduction + +Digital human avatars enabling the editing of facial expressions and head motion have broad applications across telepresence, animation, and digital content creation. The demands for personalized human avatars such as digital doubles have also increased over the past few years not only in the film industry but in the research community. To enable the widespread adaptation of these neural avatars, they should be easy to generate and animate under novel poses and facial expressions. A vast majority of existing methods aim to learn neural avatars from monocular videos to make it more accessible and easy to use [1,10,12-14,16,34,42,44]. Early approaches [12,13,16,34] mainly rely on 3D morphable models (3DMM) to guide the modeling process and to provide an intuitive control mechanism in novel view synthesis. Despite their success, they do not model the 3D face geometry. To address this issue, neural radiance fields (NeRFs) [23] have demonstrated volumetric 3D modeling of face geometry and appearance [1,10,14,42,44] under novel poses and expressions. While this allows + +![](images/8c0e0140fb6c0cdb30c0c9fe988d7f54fe92f483434bea1be89f987b0d6d9713.jpg) +Fig. 1: We present an approach for learning personalized head avatars from an unstructured video collection of the same subject. The videos are taken in different environments, with diverse appearances and possible shape changes of a character, as shown on the left. Our method leverages all of the input clips into a single unified dynamic deformable NeRF framework and offers high-quality video replays in novel head poses and facial expressions. On the right, we show some rendering results of each input appearance of the same subject aligned in head pose and facial expression that are not seen during training. Notably, our method can synthesize identity-specific details while showing coherent expressions across different appearances. + +photo-realistic image quality and animation, NeRF-based approaches require per-appearance optimization, maintaining the consistent appearance of the subject, which is constrained to the pre-recording of characters in controlled settings. However, this limits the potential for practical application scenarios where we want to generate a 3D head avatar from a collection of prerecorded video clips of a subject. In this work, we present the Personalized Head Avatar method transforming an unstructured personal video collection, containing images spanning multiple years with different appearances, head poses, and expressions, into a 3D representation of the subject. To create an avatar, we only require an unstructured collection of video clips of the person in arbitrary poses, expressions, and appearances. Our novel framework enables creation of avatars from sparse observations, while yielding high reconstruction quality in novel combinations of pose and appearance, and opens up new applications. + +Learning dynamic deformable head NeRF from unstructured video clips is a challenging problem. First, video collections can contain frames of the same identity at different appearances, poses, and expressions. Second, we only have limited pose and expression variation for each appearance, so it is unlikely that all poses and expressions of the face would be well-observed for any given appearance. Given the data, existing NeRF approaches' appearance-specific nature + +largely limits their potential for practical application scenarios. They require an independent model that is trained from scratch for every appearance from each monocular video. To address these challenges, we model shared deformation field and canonical space neural volumetric representation for all appearances with appearance-dependent latent features. Our key insight is that although observed faces have different appearances across video clips, they should all be explained by a common deformation model with small differences since they all originate from the same person. Furthermore, we assert that every video clip specific to an appearance contains both general and unique information. A single unified network could collaboratively learn the generic information from various appearances, thus this benefits rendering performance as well. The single network is designed to learn a deformation-aware dynamic head NeRF while disentangling appearance-related details from those that are not appearance-specific. For appearance conditioning, we propose employing learnable appearance embedding attached to vertices of the tracked mesh and passing them to the dynamic head NeRF network. By adopting this process, PAV is able to learn discriminative latent neural features that are specific for each appearance, resulting in better rendering of appearance-specific details for arbitrary poses and expressions. Unlike existing NeRF methods [1, 10, 14, 42, 44], PAV simultaneously trains on all appearances of the same subject, eliminating the need for separate training sessions for each appearance and reducing the overall training duration. Learning of personalized head avatars in a single unified network from an unstructured collection of videos has not been addressed before and is a non-trivial extension of existing monocular head avatar generation methods since they can not handle multiple appearances in a single network. PAV makes this possible through a carefully designed appearance conditioning representation, geometry-attached trainable latent neural texture to learn appearance-specific features, appearance conditioned implicit functions to predict variations in color and density. We conducted a variety of experiments to show these improvements over the baseline approach. Since there is no such public dataset to evaluate our method, we curate our own in-the-wild video collections for multiple subjects. Overall, we present Personalized Head Avatar with the following contributions: + +- We present, PAV, a framework for learning controllable head avatars from unconstrained short video collections where the same person shows different appearances and facial expressions. To the best of our knowledge, this is the first method learning appearance controllable head avatars from in-the-wild monocular video collections of the same subject. +- The proposed method can synthesise images of head avatar under target appearance, head pose and facial expressions. +- PAV learns modeling local changes of geometry and color across different appearances of the same identity. + +# 2 Related Work + +Monocular Neural Head Avatars. More recent methods learn neural representations of the person from monocular videos and use neural renderers [31, 32] to directly generate photo-realistic images in the target head pose and facial expression [1-3, 5, 10, 11, 14, 16, 21, 29, 33, 40, 42, 44]. These methods are generally classified into image-based or 3D-based neural rendering methods [31, 32]. The image-based models synthesize the face of a subject without relying on any representation of 3D space. Existing image-based approaches either employ learned warping fields [2, 29] to deform an input image or encoder-decoder architecture where decoder synthesize the output image [21, 25, 35, 39]. Even though these methods produce high-quality results, they suffer from artifacts for strong head pose and expression changes due to the lack of 3D geometry. The 3D-based neural rendering methods, on the other hand, rely on 3D Morphable Models (3DMM) [9]. These methods are generally classified into explicit and implicit models. Explicit models utilize 2D neural rendering methods [31, 32] and rely on neural networks which can be conditioned on coarse RGB renderings based on a linear texture model [16], uv-maps [4] or latent feature maps [33]. Although these techniques produce high-quality renderings, they are confined to operating within a 2D image space and are limited to craniofacial structures. To tackle this, implicit neural representations for modeling 3D faces have recently become a major focus of attention. In particular, NeRFs [23] shows photorealistic face image synthesis in novel poses and expressions [1, 10, 14, 42, 44]. Implicit models represent the geometry using implicit surface functions or volumetric representations and they are optimized via volumetric rendering [23]. Early approaches [23] and follow-up works synthesize novel views of complex static scenes using differentiable volumetric rendering. After introduction of NeRF for static scenes, a natural research direction was to generalize it to dynamic and time-varying ones [26, 28]. The reconstruction problem is divided into two different spaces, the deformed space and the canonical space. For human body, a series of approaches have been proposed to target this challenge [20, 27]. In particular, these methods are differentiated from each other based on deformation field modeling. Gafni et al. [10] introduced implicit deformation modeling conditioned on facial expression and head pose. In contrast to modeling the deformation field implicitly, IMAvatar [42], RigNerf [1] and Insta [44] propose explicit modeling of the deformation field to improve control and rendering quality. Our method falls into this paradigm as we build our system on top of dynamic deformable head NeRF. In particular, we take inspiration from Insta [44] for designing our framework. However, existing deformable dynamic head NeRF methods including Insta [44] require per-identity optimization, maintaining the consistent appearance of the same subject. To the best of our knowledge, PAV is the first method to learn a single unified dynamic deformable head NeRF from monocular video of multiple appearances of the same subject. + +Geometry-anchored Features for NeRFs. PAV introduces appearance-conditioned head NeRF by using geometry-attached features. By using these features, exist + +ing works show improved rendering quality for static objects [38] and human bodies [3, 20]. However, none of the existing methods [3, 20] has yet addressed multi-appearance learning for dynamic neural radiance fields. + +NeRF from Unstructured Data Collections. Learning neural radiance field from unstructured data collections has recently gained much attention. Existing works have enabled photorealistic view synthesis of challenging static scenes, including tourist sites [22] and city-scale scenes [30], and static objects [18] from internet photo collections. For human subjects, Weng et al. [37] introduce free-viewpoint human body rendering from an unstructured photo collection. Despite the photorealistic rendering quality of PersonNerf [37], they do not model animatable faces due to the static training images. In our work, we learn animatable 3D head avatars from video collections. + +![](images/198902175a488fa1839da38f7a9d3bdbf7958a8df09f4111f5c5b2903b36a25a.jpg) +Fig. 2: Overview of the proposed approach, PAV. (a) shows some training videos in various appearances of the same subject. (b) We estimate the 3D head geometry of the person using FLAME head model. (c) shows the proposed personalized deformable neural radiance field from video collection (d) We learn appearance-specific neural features and attach them to vertices. (d) This feature is passed to fully fused multi-layer perceptrons with additional conditioning on the facial expressions $e$ and the encoded view direction $d$ . + +# 3 Method + +Our goal is to learn a personalized head avatar, PAV, from a monocular video collection of a subject under different appearances, poses, and facial expressions. To create the avatar, we use a small collection $\mathbf{V} = \{V_{i}\}_{i = 1}^{N}$ with $N$ videos. We parameterize PAV using FLAME [19] face model to control head poses, facial expressions and use neural texture to model appearance. During inference, we only need the target FLAME parameters and neural texture. Learning PAV from a collection of monocular videos requires capturing the geometry and appearance of the dynamic head while also disentangling different appearances of the same + +identity. In order to tackle this problem, we first capture the dynamic 3D geometry by employing a parametric head model [19] (Sec. 3.1). We use the coarse geometry to model geometry-guided deformable neural radiance field (Sec. 3.2) and vertex-attached neural features to condition appearance for rendering (Sec. 3.3). We also propose estimating appearance-conditioned neural radiance field, aimed at accurately capturing distinct color and geometry with high fidelity. (Sec. 3.4). An overview of our method can be seen in Fig. 2. In the following, we describe each of these modules in greater detail. + +# 3.1 Head Geometry Estimation + +Given the training frames $\mathbf{I} = \{I_{\mathrm{i}}\}$ , from all appearances, we first estimate the optimized camera parameters $\mathbf{K} \in \mathbb{R}^{3 \times 3}$ , tracked FLAME [19] meshes $\mathbf{M} = \{M_{\mathrm{i}}\}$ with corresponding facial expression coefficients $\mathbf{E} = \{E_{\mathrm{i}}\}$ and poses $\mathbf{P} = \{P_{\mathrm{i}}\}$ . We follow an optimization-based method, that minimizes an objective function with photo-consistency and landmark terms [43]. Different from the original work [43] which is designed for a single appearance, in PAV, we compute the average shape parameters from all appearances of the same subject to represent identity shape during optimization. + +# 3.2 Personalized Deformable NeRF + +Our goal is to create a head avatar that can be learned from frames in video collection of various appearance of the same subject. We parameterize PAV using FLAME model and learned appearance conditioning. Specifically, PAV takes target head pose $\theta$ , facial expression $e$ and latent neural features $(LNF)$ $Z_{j} \in \mathbb{R}^{512 \times 512 \times 16}$ in the texture atlas space (UV space) for an appearance $j$ as inputs and synthesizes the target image $I^{\theta, e, Z_{j}}$ as: + +$$ +I ^ {\theta , e, Z _ {j}} = \operatorname {P A V} (\theta , e, Z _ {j}, K), \tag {1} +$$ + +where $K$ corresponds to the intrinsic parameters of the target camera viewpoint. For this purpose, we are using a geometry-guided dynamic deformable neural radiance field. However, deformable head NeRF models cannot disentangle appearance from the neural radiance field (See Sec. 2). To tackle this problem, we employ shared canonical space for all appearances where a shared neural radiance field is constructed. With this shared radiance field at canonical space, to render the target head pose and facial expression using volumetric rendering, we canon- icalize the sample points $x$ on a ray from the observation space to query points $x_{c}$ in the shared canonical space radiance field by using deformation gradient $\hat{F}$ .. $x_{c} = \hat{F}\cdot x$ where $\hat{F}\in \mathbb{R}^{4\times 4}$ and $\{x,x_c\} \in \mathbb{R}^4$ . We employ the nearest triangle search for a given point to compute the deformation gradient. Specifically, the deformation gradient is computed between the deformed triangle $\hat{T}_{def}\in M_i$ in the observed mesh and the canonical triangle $\hat{T}_{canon}\in M^{canon}$ in the canonical mesh. To increase the speed of nearest triangle search, we employ bounding volume hierarchy (BVH) [6]. We refer readers to [44] for a detailed description of deformation gradient computation. + +After canonicalization, given the predicted appearance-conditioned radiance $c$ and density $\sigma$ (Sec. 3.4) for every point on each ray, the representation of the avatar is optimized by using the differentiable volumetric rendering. For each camera ray $r(t) = o + td$ with camera center $o$ and viewing direction $d$ , the color of the corresponding pixel can be computed by accumulating the predicting radiance and densities of the sample points along the ray: + +$$ +\mathcal {C} (\mathbf {r}) = \int_ {t _ {n}} ^ {t _ {f}} \sigma (\mathbf {r} (t)) \cdot \mathbf {c} (\mathbf {r} (t)) \cdot \mathcal {T} (t) d t \tag {2} +$$ + +where $t_n$ and $t_f$ are the near and far bounds correspondingly, and $\mathcal{T}(t_n) = \exp \left(-\int_{t_n}^{t_f}\sigma (t)dt\right)$ is the accumulated transmittance along the ray from $t_n$ to $t_f$ , $\sigma (t)$ is the density and $c(t)$ it the radiance at position $x_t$ . + +# 3.3 Learned Appearance Embedding + +In this section, PAV introduces a learnable appearance embedding that is essential and effective for enabling discriminative multi-appearance rendering using single unified network. To achieve this goal (See Fig. 2), a set of $N$ learnable latent features $\mathbf{Z} = \{Z_{\mathrm{i}}\}_{i = 1}^{N}$ , $Z_{i}\in \mathbb{R}^{512\times 512\times 16}$ are defined in UV space corresponding to $N$ appearances of the same subject in the video collection. We impose high-dimensional learnable latent features to condition implicit representations for both density $\sigma$ and radiance $c$ values in the canonical space. To this end, these latent neural features are attached to the vertices of 3DMM mesh, which enables PAV to memorize local variations. For each point $x$ sampled on a ray, we find the nearest triangle $\mathcal{T} = \{v_{\mathrm{i}}\}_{i = 1}^{3}, v_{i}\in \mathbb{R}^{3}$ (sec. 3.2) and aggregate the features $z(v_{i})$ of vertices $v_{i}$ from the vertices of the nearest triangle to get the appearance embedding $Z^x$ . Then we directly use the appearance embedding to condition for appearance-dependent neural radiance field estimation (Sec. 3.4). During training, these features are randomly initialized and optimized jointly with the objectives of the entire network. + +# 3.4 Appearance Conditioned NeRF + +To model the appearance-conditioned dynamic deformable head NeRF, PAV incorporates learned appearance-embedding (Sec. 3.3) into neural radiance field in canonical space (See Fig. 2). For a point $x$ sampled on a ray, after canonicalization, the point $x_{c}$ is encoded using multi-resolution hash encoding [24]. This feature $\phi(x_{c})$ is passed to multi-layer perceptron for geometric feature estimation. Following [10] and [44], we conditioned every point features $\phi(x_{c})$ with the facial expressions $e$ of frame $i$ . Without per-frame latent code, for shared volume density, we learn implicit representation $\mathcal{F}_{\text{density}}$ using multi-layer perceptron (MLP-D in Fig. 2): + +$$ +\mathcal {F} _ {\text {d e n s i t y}}: \left(\phi \left(x _ {c}\right), e _ {i}\right)\rightarrow \sigma \tag {3} +$$ + +A primary challenge lies in learning unique density estimations for each appearance while separating the shared density aspects from the appearance. Hence we design PAV such that appearance-based density offset is predicted from appearance embedding $Z^{x}$ sampled from the nearest triangle to the point $x$ on the mesh $M_{i}$ and facial expression $e_{i}$ of frame $i$ . For density offset, we learn implicit representation $\mathcal{F}_{offset}$ using multi-layer perceptron (MLP-O in Fig. 2): + +$$ +\mathcal {F} _ {\text {o f f s e t}}: \left(Z ^ {x}, e _ {i}\right)\rightarrow \Delta_ {\sigma} \tag {4} +$$ + +Final density value $\sigma(x) = \sigma + \Delta_{\sigma}$ for the point $x$ is the summation of shared density features $\sigma$ and density offset $\Delta_{\sigma}$ . For emitted radiance at point $x$ , we learn implicit representation $\mathcal{F}_{color}$ using multi-layer perceptron (MLP-C in Fig. 2): + +$$ +\mathcal {F} _ {\text {c o l o r}}: \left(Z ^ {x}, d _ {i}, \sigma (x)\right)\rightarrow c \tag {5} +$$ + +where $d_{i}$ is the encoded view direction of training frame $i$ . Then, final color for the pixel is computed by using differentiable volume rendering (Eq. 2) + +# 3.5 Optimization + +Our model is trained on a collection of monocular RGB videos with photometric and depth losses. For photometric loss $\mathcal{L}_{color}$ , we employ $L_{2}$ distance between the renderings $C$ and ground-truth images $I$ , $\mathcal{L}_{color} = \sum_{i}\sum_{r}\| C_{i}(r) - I_{i}(r)\|$ , where $r$ denotes the camera ray of each pixel and $i$ denotes frame index. For depth loss $\mathcal{L}_{depth}$ , we rasterize the depth of tracking mesh $M_{i}$ and apply $L_{1}$ distance between this map $D$ and the ray termination $H$ of the volumetric rendering only for face region as FLAME model is restricted to model details like hair: $\mathcal{L}_{depth} = \sum_{i}\sum_{r}\mathcal{M}_{i}(r)\| H_{i}(r) - D_{i}(r)\|$ , where $\mathcal{M}$ denotes face binary mask. The total loss is defined as: + +$$ +\mathcal {L} = \lambda_ {\text {c o l o r}} \mathcal {L} _ {\text {c o l o r}} + \lambda_ {\text {d e p t h}} \mathcal {L} _ {\text {d e p t h}} \tag {6} +$$ + +where we empirically set $\lambda_{color} = 50$ and $\lambda_{depth} = 10$ . + +# 4 Experiments + +In this section, we evaluate the performance of PAV using two different datasets. We perform an ablation study to validate our design choices and a comparison against the state-of-the-art monocular head avatar method in various configurations. + +# 4.1 Dataset + +Our method takes a video collection of a subject containing different appearances, poses, and expressions. Since there is no existing dataset to evaluate + +learning 3D head avatar from video collections of the same subjects with various appearances, poses, and expressions, we compiled a new Video Collection (VidCol) dataset from talking face videos of Roger Federer, King Charles and Tom Hanks across years. This dataset consists of in-the-wild videos recorded in uncontrolled environments. In VidCol dataset, we curated videos from 5 different appearances for each subject. We randomly sampled 500 frames at $512^2$ resolution from each video (or appearance) and we used a total of 2500 frames to train and test a single unified network per subject. We will share this dataset with annotations of semantic masks, tracked 3DMM models, and camera calibrations for further research in this direction. + +For our experiments, we evaluate all characters and train a separate PAV model for each subject. We develop two different protocols for evaluation: + +a) Unified Network for Novel Pose/Exp. Synthesis This protocol evaluates the performance in terms of novel appearance, head pose, and facial expression synthesis when training a single unified network from all appearances. For training, we select 400 frames for each appearance of a subject talking in front of a camera with various head poses. It mimics a training video commonly used for monocular neural head avatar methods (Sec. 2). For testing, we use the remaining 100 frames for each appearance with unseen poses and expressions. + +b) Separated Networks for Novel Pose/Exp. Synthesis This protocol evaluates the novel appearance, head pose, and facial expression generalization ability of the methods when training a separate network for each appearance. We select 400 training images for each appearance of a subject talking of a camera, same as protocol-a. For testing, we choose the remaining 100 unseen pose/expression combinations with ground truth, same as protocol-a, and another 300 unseen pose/expression sets without ground truth image data. + +# 4.2 Metrics + +We report several metrics to evaluate the quality of synthesized images as well as the ablation study of our method. For synthesized images with provided ground-truth, we use Learned Perceptual Patch Similarity (LPIPS) [41], Deep Image Structure and Texture Similarity (DISTS) [8], Structural Similarity Index (SSIM) [36], Peak Signal-to-Noise Ratio (PSNR) and Fréchet Inception Distance (FID) [15]. When there is an absence of ground truth, we adopt FID metric for quantitative comparison. + +# 4.3 Implementation Details + +We optimize Eq. 6 using Adam optimizer [17] with hyperparameters $\beta_{1} = 0.9$ and $\beta_{2} = 0.99$ . We set the learning rate to $2.5 \times 10^{-3}$ for $F_{density}$ , $F_{color}$ and $Z$ , and to $2.5 \times 10^{-4}$ for $F_{offset}$ . We sample 1024 points along each ray for rendering. Multi-resolution hash encoding used in the method is set to 16 number of levels, 18 hash map size, and 2048 finest resolution. The optimization takes 15K iterations + +![](images/fb9dbf79da6fd0ec4d6e8efc8a9696e8bd0953409ef827f3544c570740b3e090.jpg) +Fig. 3: Qualitative Ablation Study on VidCol Dataset. This figure shows synthesized images and pixel error maps against ground-truth. Please note that brighter pixel colors denote higher pixel errors. Without appearance embedding with latent neural features and appearance-conditioned density offset (a), the model fails to learn appearance-based distinct details. The latent neural features (LNF) (c) resolve the issue of high-fidelity rendering of texture (blue arrow). The appearance-conditioned density offset field (d) allows the model to learn more accurate geometry and texture that corresponds to the face, hair, and neck regions (red arrow) + +to converge when training network with all appearances of single subjects. More details on the network are provided in the supplementary document. + +# 4.4 Ablation Study + +We evaluate different design choices of PAV in Tab. 1 and Fig. 3. In this study, we use protocol-a of the VidCol dataset for all experiments. We first report the results of the final model which includes latent neural features (LNF) (Sec. 3.3) and density offset $\varDelta_{\sigma}$ (Sec. 3.4). The full model achieves PSNR score of 25.15 and SSIM score of 0.91 for image synthesis. If we remove the density offset from our model while employing appearance-conditioned radiance estimation, PSNR and SSIM scores for image synthesis decrease to 24.14 and 0.90, respectively. This indicates that without appearance-conditioned density offset $\varDelta_{\sigma}$ , density estimation MLP ( $\mathcal{F}_{\text {density }}$ in Sec. 3.4) can not learn to synthesize appearance-specific geometry and texture details, thus rendering fails there as well. Even + +
PSNR ↑SSIM ↑DISTS ↓LPIPS ↓FID ↓
Insta [44]21.030.8490.1850.169104.28
PAV without LNF24.190.9000.1450.11954.39
PAV without Δσ24.140.9060.1340.10847.71
PAV (Full Model)25.150.9100.1330.10746.67
+ +Table 1: Ablation Study: We evaluate the impact of different components of the proposed method. See Fig. 3 for qualitative comparison. + +though the model $\mathrm{w / o}\varDelta_{\sigma}$ synthesise images that look on par with ground truth in terms of texture, it generates severe artifacts which can be seen in pixel error maps. An example of this behavior can also be seen in Fig. 3 $(\mathrm{w / o}\varDelta_{\sigma})$ Rendering failures on face, head, and neck regions are indicated by red arrow and error map. + +Next, we evaluate the impact of latent neural features (LNF) as an appearance embedding on PAV. In this experiment, to condition the canonical MLPs in PAV, inspired by "Nerf In-the-wild" [22, 37], we adopt Generative Latent Optimization (GLO) for appearance embedding. Specifically, this approach defines a single real-valued appearance embedding vector $l$ for each appearance in the dataset. To this end, we employ a pre-trained encoder from in-the-wild face detector RetinaFace [7] for GLO. During training, if we replace $\mathbf{Z} = \{Z_{\mathrm{i}}\}_{i=1}^{N}$ , $Z_{i} \in \mathbb{R}^{512 \times 512 \times 16}$ with $\mathbf{L} = \{l_{\mathrm{i}}\}_{i=1}^{N}$ , $l_{i} \in \mathbb{R}^{512 \times 1}$ for $N$ appearances of each subject, and pass it to appearance conditioned NeRF (See Sec. 3.4), PSNR and SSIM scores for image synthesis decrease to 24.19 and 0.90, respectively. This is also evident from the qualitative results shown in Fig. 3 (w/o LNF), indicating that latent neural features help in improved image synthesis quality with their local features attached to 3D geometry. Rendering failures on head, neck, and skin texture are shown by red arrow, blue arrow and error maps in Fig. 3. Finally, as discussed in Sec. 3, we keep the network design as we explain after the ablation study. + +# 4.5 Comparison + +Since our method is the first work that learns dynamic deformable head NeRF from multiple appearances of the same subjects using a unified network, we ourselves design some baseline comparisons in this section. First, we compare our method against Insta [44] using protocol-a (Sec. 4.1) of the VidCol dataset. Both methods are trained using the frames sampled from all appearances of the same subject. For testing, both methods synthesize images for target head pose and expression. The results are summarized in Tab. 2 and Fig. 4. For protocol-a, our method clearly outperforms the baseline for final image synthesis results across all metrics. It should be noted that PAV framework jointly learns all different appearances via one-time training using a unified network and it is able to preserve appearance-specific details for novel poses and expressions. On the other hand, as shown in Fig. 4, Insta can not disentangle appearance from the + +![](images/862f9ccf65ef075cc91ec5ff1d498e4c0b5dbe4a2ed0739cf4064be36fd6ad56.jpg) +Fig. 4: Qualitative comparison of our method (PAV) against baseline method (Insta [44]. Columns $\{2,6\}$ show target head poses and facial expressions for synthesized images at columns $\{3,4,7,8\}$ . + +neural radiance field, which results in a mixture of appearances in synthesized images and inaccurate geometry predictions. This is the limitation of Insta we are addressing in the paper. + +Second, we train separate Insta [44] models for each appearance of the same subject using protocol-b while training a single unified model for our method using the frames sampled from all appearances of the same subject. For testing, we use the same test data from protocol-b for both PAV and Insta [44] and synthesize images for novel poses and expressions. Note that, we added more novel expressions and poses in protocol-b (Sec. 4.1) to test the generalization of our method under a unified training schema, which results in increasing FID score for both methods in Tab. 3. Even though the results show that artifacts (in Fig. 4) of the Insta renders are decreased in Fig. 5, our method demonstrates better-rendering results (Fig. 5) and lower FID score (Tab. 3). This indicates that PAV's single unified model helps in learning generic neural radiance field representation across all appearances of the same subject while preserving appearance-based details. Please note that we are demonstrating Insta's performance under a challenging pose/expression in Fig. 5 which can also be seen in + +
MethodVideo Collection 1Video Collection 2
PSNR ↑SSIM ↑DISTS ↓LPIPS ↓FID ↓PSNR ↑SSIM ↑DISTS ↓LPIPS ↓FID ↓
Protocol (a): Novel Head Pose and Expression Synthesis
Insta [44]21.030.840.1850.169104.2823.670.860.1870.16398.87
PAV (Ours)25.150.910.1330.10746.6727.090.890.1440.12053.91
+ +Table 2: Comparison of PAV against Insta [44]. See Fig. 4 for qualitative comparison. Video Collection 1 and 2 refer to Federer and King Charles splits of VidCol dataset, respectively. + +the original paper results under extreme poses/expressions. In Fig. 5, the reference image is for photoconsistency check only since there is no ground truth image for that pose/expression. Finally, we provide additional qualitative results of avatars animated in the supplementary video. We also compared PAV against Insta for model efficiency using protocol-b dataset and recorded values in Tab. 3. Although we train our model with multiple appearances jointly, PAV is more efficient than Insta in terms of memory footprint and training time + +![](images/06e6d4e9ac32ff115f6293a48eda57c3754b4c7e80c1df691cc1c39f984f2b04.jpg) +Reference + +![](images/785981cd406aa46df4b499f550ab7df46574d32d2a2a328b131545f1839849bb.jpg) + +![](images/e7b2a0970d0ea6029004843b50749a6242a6eb6f832be788d1c488f5bc0c3603.jpg) +Target Pose/Exp. +Fig. 5: Qualitative comparison of PAV against single-appearance traines Insta [44]. Please note that the reference image is for photoconsistency check only since there is no ground truth image for that pose/expression + +![](images/3bbf67cab567bb2963ea47f86303776b311718312d4ff11c2d718ec5bf22872b.jpg) + +![](images/6cefb32e6d820a2f9a30c6c8c622afccf81ce2c5b1b1b9c762d8d4a5cb7c6d02.jpg) +Insta [44] + +![](images/45aee65b3f48520f21bf405d7f9cd85f8772b2c46159bfe22a07bd61d3e665c9.jpg) + +![](images/ccb511cdbc55f7c76327c5111109e121b8472b176573cbe83398ed8bcecb894c.jpg) +PAV + +# 5 Discussion + +We presented PAV which is a novel method for learning personalized head avatars from monocular and in-the-wild video collections of a person. We demonstrated that PAV can model a 3D head avatar from an unstructured video collection of a subject in a single unified network while disentangling face appearances from the radiance field. PAV can be trained simultaneously for multiple appearances of the same subject, allowing novel pose and expression generalization and leading to a superior rendering quality of our method compared to the baseline. + +
Train Time MinutesNumber of Param. MillionModel Size (MB)FID ↓
Multiple Insta [44]13579.71932109.22
PAV (Ours)4526.865091.01
+ +Table 3: Efficiency and performance comparison between PAV and Insta [44] trained for each appearance (Single Appearance Training) in dataset collection 1. FID score is computed from synthetized images under novel head poses and expressions where there are no existing ground-truth images. + +![](images/631be0a96611e488f6eb28929bbf6b526136f5c25d8411e93327cabe30d5e28e.jpg) +Fig. 6: Multi-identity and multi-appearance extension of PAV. The learned appearance (middle) rendered in a new viewpoint (left) does not match the visual quality of input images (right). Please note that the reference image is for photoconsistency check only since there is no ground truth image for that pose/expression. + +Limitation for Multi-Identity. We extended our method for multiple identities and multiple appearances for each identity. To this end, we train a single unified PAV model using frames smaped from all appearances of Federer and King subjects from VidCol dataset with shared canonical space. We evaluate novel head pose and facial expression rendering of this model and Fig. 6 illustrates the results. Even though this model produces reasonable results, synthesized images demonstrate a lack of subject-specific shape and appearance details with rendering artifacts. In the paper, we focus on multiple appearances of a single subject. Scaling up this model to a wide range of subjects and exploring generalized dynamic deformable NeRF would be interesting future work. + +Ethical Considerations. Although the rapid progress on digital human avatars facilitates numerous downstream applications, it also raises concerns about ethical misuse. On the synthesis side, it would be ideal to actively use watermarking technologies and avoid driving avatars with different identities. In this work, we aim to faithfully reproduce images of a person with the capability of rendering unseen poses and expressions and switching appearance within their own set of appearances. The work does not intend to create animations that didn't happen. + +# References + +1. Athar, S., Xu, Z., Sunkavalli, K., Shechtman, E., Shu, Z.: Rignerf: Fully controllable neural 3d portraits. In: Proceedings of the IEEE/CVF conference on Computer Vision and Pattern Recognition. pp. 20364-20373 (2022) +2. Averbuch-Elor, H., Cohen-Or, D., Kopf, J., Cohen, M.F.: Bringing portraits to life. ACM transactions on graphics (TOG) 36(6), 1-13 (2017) +3. Bai, Z., Tan, F., Huang, Z., Sarkar, K., Tang, D., Qiu, D., Meka, A., Du, R., Dou, M., Orts-Escalano, S., et al.: Learning personalized high quality volumetric head avatars from monocular rgb videos. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 16890-16900 (2023) +4. Bühler, M.C., Meka, A., Li, G., Beeler, T., Hilliges, O.: Varitex: Variational neural face textures. In: Proceedings of the IEEE/CVF International Conference on Computer Vision. pp. 13890-13899 (2021) +5. Chen, C., O'Toole, M., Bharaj, G., Garrido, P.: Implicit neural head synthesis via controllable local deformation fields. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 416-426 (2023) +6. Clark, J.H.: Hierarchical geometric models for visible surface algorithms. Communications of the ACM 19(10), 547-554 (1976) +7. Deng, J., Guo, J., Zhou, Y., Yu, J., Kotsia, I., Zafeiriou, S.: Retinaface: Single-stage dense face localisation in the wild. arXiv preprint arXiv:1905.00641 (2019) +8. Ding, K., Ma, K., Wang, S., Simoncelli, E.P.: Image Quality Assessment: Unifying Structure and Texture Similarity. TPAMI (2022) +9. Egger, B., Smith, W.A., Tewari, A., Wuhrer, S., Zollhoefer, M., Beeler, T., Bernard, F., Bolkart, T., Kortylewski, A., Romdhani, S., et al.: 3d morphable face models—past, present, and future. ACM Transactions on Graphics (ToG) 39(5), 1-38 (2020) +0. Gafni, G., Thies, J., Zollhöfer, M., Nießner, M.: Dynamic neural radiance fields for monocular 4d facial avatar reconstruction. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR). pp. 8649-8658 (June 2021) +1. Gao, X., Zhong, C., Xiang, J., Hong, Y., Guo, Y., Zhang, J.: Reconstructing personalized semantic facial nerf models from monocular video. ACM Transactions on Graphics (TOG) 41(6), 1-12 (2022) +2. Garrido, P., Valgaerts, L., Rehmsen, O., Thormahlen, T., Perez, P., Theobalt, C.: Automatic face reenactment. In: Proceedings of the IEEE conference on computer vision and pattern recognition. pp. 4217-4224 (2014) +3. Garrido, P., Valgaerts, L., Sarmadi, H., Steiner, I., Varanasi, K., Perez, P., Theobalt, C.: Vdub: Modifying face video of actors for plausible visual alignment to a dubbed audio track. In: Computer graphics forum. vol. 34, pp. 193-204. Wiley Online Library (2015) +4. Grassal, P.W., Prinzler, M., Leistner, T., Rother, C., Nießner, M., Thies, J.: Neural head avatars from monocular rgb videos. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 18653-18664 (2022) +5. Heusel, M., Ramsauer, H., Unterthiner, T., Nessler, B., Hochreiter, S.: Gans trained by a two time-scale update rule converge to a local nash equilibrium. Advances in neural information processing systems 30 (2017) +6. Kim, H., Garrido, P., Tewari, A., Xu, W., Thies, J., Niessner, M., Pérez, P., Richardt, C., Zollhöfer, M., Theobalt, C.: Deep video portraits. ACM transactions on graphics (TOG) 37(4), 1-14 (2018) + +17. Kingma, D.P., Ba, J.: Adam: A method for stochastic optimization. arXiv preprint arXiv:1412.6980 (2014) +18. Kuang, Z., Olszewski, K., Chai, M., Huang, Z., Achlioptas, P., Tulyakov, S.: Neroic: Neural rendering of objects from online image collections. ACM Transactions on Graphics (TOG) 41(4), 1-12 (2022) +19. Li, T., Bolkart, T., Black, M.J., Li, H., Romero, J.: Learning a model of facial shape and expression from 4D scans. ACM Transactions on Graphics, (Proc. SIGGRAPH Asia) 36(6), 194:1-194:17 (2017), https://doi.org/10.1145/3130800.3130813 +20. Liu, L., Habermann, M., Rudnev, V., Sarkar, K., Gu, J., Theobalt, C.: Neural actor: Neural free-view synthesis of human actors with pose control. ACM transactions on graphics (TOG) 40(6), 1-16 (2021) +21. Mallya, A., Wang, T.C., Liu, M.Y.: Implicit warping for animation with image sets. Advances in Neural Information Processing Systems 35, 22438-22450 (2022) +22. Martin-Brualla, R., Radwan, N., Sajjadi, M.S., Barron, J.T., Dosovitskiy, A., Duckworth, D.: Nerf in the wild: Neural radiance fields for unconstrained photo collections. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 7210-7219 (2021) +23. Mildenhall, B., Srinivasan, P.P., Tancik, M., Barron, J.T., Ramamoorthi, R., Ng, R.: Nerf: Representing scenes as neural radiance fields for view synthesis. Communications of the ACM 65(1), 99-106 (2021) +24. Müller, T., Evans, A., Schied, C., Keller, A.: Instant neural graphics primitives with a multiresolution hash encoding. ACM Transactions on Graphics (ToG) 41(4), 1-15 (2022) +25. Nirkin, Y., Keller, Y., Hassner, T.: Fsgan: Subject agnostic face swapping and reenactment. In: Proceedings of the IEEE/CVF international conference on computer vision. pp. 7184-7193 (2019) +26. Park, K., Sinha, U., Hedman, P., Barron, J.T., Bouaziz, S., Goldman, D.B., Martin-Brualla, R., Seitz, S.M.: Hypernerf: A higher-dimensional representation for topologically varying neural radiance fields. arXiv preprint arXiv:2106.13228 (2021) +27. Peng, S., Dong, J., Wang, Q., Zhang, S., Shuai, Q., Zhou, X., Bao, H.: Animatable neural radiance fields for modeling dynamic human bodies. In: Proceedings of the IEEE/CVF International Conference on Computer Vision. pp. 14314-14323 (2021) +28. Pumarola, A., Corona, E., Pons-Moll, G., Moreno-Noguer, F.: D-nerf: Neural radiance fields for dynamic scenes. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 10318-10327 (2021) +29. Siarohin, A., Lathuilière, S., Tulyakov, S., Ricci, E., Sebe, N.: First order motion model for image animation. Advances in neural information processing systems 32 (2019) +30. Tancik, M., Casser, V., Yan, X., Pradhan, S., Mildenhall, B., Srinivasan, P.P., Barron, J.T., Kretzschmar, H.: Block-nerf: Scalable large scene neural view synthesis. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 8248-8258 (2022) +31. Tewari, A., Fried, O., Thies, J., Sitzmann, V., Lombardi, S., Sunkavalli, K., MartinBrualla, R., Simon, T., Saragih, J., Nießner, M., et al.: State of the art on neural rendering. In: Computer Graphics Forum. vol. 39, pp. 701-727. Wiley Online Library (2020) +32. Tewari, A., Thies, J., Mildenhall, B., Srinivasan, P., Tretschk, E., Yifan, W., Lassner, C., Sitzmann, V., Martin-Brualla, R., Lombardi, S., et al.: Advances in neural rendering. In: Computer Graphics Forum. vol. 41, pp. 703-735. Wiley Online Library (2022) + +33. Thies, J., Zollhöfer, M., Nießner, M.: Deferred neural rendering: Image synthesis using neural textures. Acm Transactions on Graphics (TOG) 38(4), 1-12 (2019) +34. Thies, J., Zollhofer, M., Stamminger, M., Theobalt, C., Nieflner, M.: Face2face: Real-time face capture and reenactment of rgb videos. In: Proceedings of the IEEE conference on computer vision and pattern recognition. pp. 2387-2395 (2016) +35. Wang, T.C., Liu, M.Y., Zhu, J.Y., Tao, A., Kautz, J., Catanzaro, B.: High-resolution image synthesis and semantic manipulation with conditional gans. In: Proceedings of the IEEE conference on computer vision and pattern recognition. pp. 8798-8807 (2018) +36. Wang, Z., Bovik, A.C., Sheikh, H.R., Simoncelli, E.P.: Image quality assessment: From error visibility to structural similarity. TIP (2004) +37. Weng, C.Y., Srinivasan, P.P., Curless, B., Kemelmacher-Shlizerman, I.: Person-nerf: Personalized reconstruction from photo collections. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 524-533 (2023) +38. Yang, B., Bao, C., Zeng, J., Bao, H., Zhang, Y., Cui, Z., Zhang, G.: Neumesh: Learning disentangled neural mesh-based implicit field for geometry and texture editing. In: European Conference on Computer Vision. pp. 597-614. Springer (2022) +39. Zakharov, E., Ivakhnenko, A., Shysheya, A., Lempitsky, V.: Fast bi-layer neural synthesis of one-shot realistic head avatars. In: Computer Vision-ECCV 2020: 16th European Conference, Glasgow, UK, August 23-28, 2020, Proceedings, Part XII 16. pp. 524-540. Springer (2020) +40. Zhang, B., Qi, C., Zhang, P., Zhang, B., Wu, H., Chen, D., Chen, Q., Wang, Y., Wen, F.: Metaportrait: Identity-preserving talking head generation with fast personalized adaptation. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR). pp. 22096-22105 (June 2023) +41. Zhang, R., Isola, P., Efros, A.A., Shechtman, E., Wang, O.: The unreasonable effectiveness of deep features as a perceptual metric. In: CVPR (2018) +42. Zheng, Y., Abrevaya, V.F., Bühler, M.C., Chen, X., Black, M.J., Hilliges, O.: Im avatar: Implicit morphable head avatars from videos. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 13545-13555 (2022) +43. Zielonka, W., Bolkart, T., Thies, J.: Towards metrical reconstruction of human faces. In: European Conference on Computer Vision. pp. 250-269. Springer (2022) +44. Zielonka, W., Bolkart, T., Thies, J.: Instant volumetric head avatars. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 4574-4584 (2023) \ No newline at end of file diff --git a/2024/PAV_ Personalized Head Avatar from Unstructured Video Collection/images.zip b/2024/PAV_ Personalized Head Avatar from Unstructured Video Collection/images.zip new file mode 100644 index 0000000000000000000000000000000000000000..879aa4500c519f3ddc829f6115290b069970e85c --- /dev/null +++ b/2024/PAV_ Personalized Head Avatar from Unstructured Video Collection/images.zip @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:9d4b648bb9509054b9850bf70519e4065af36d06fd92326f5da3c9704683f196 +size 433437 diff --git a/2024/PAV_ Personalized Head Avatar from Unstructured Video Collection/layout.json b/2024/PAV_ Personalized Head Avatar from Unstructured Video Collection/layout.json new file mode 100644 index 0000000000000000000000000000000000000000..c0ee4eed385a84fd876bb20ba4c0b21f0660488d --- /dev/null +++ b/2024/PAV_ Personalized Head Avatar from Unstructured Video Collection/layout.json @@ -0,0 +1,8342 @@ +{ + "pdf_info": [ + { + "para_blocks": [ + { + "bbox": [ + 177, + 111, + 436, + 145 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 177, + 111, + 436, + 145 + ], + "spans": [ + { + "bbox": [ + 177, + 111, + 436, + 145 + ], + "type": "text", + "content": "PAV: Personalized Head Avatar from Unstructured Video Collection" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 165, + 167, + 448, + 180 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 165, + 167, + 448, + 180 + ], + "spans": [ + { + "bbox": [ + 165, + 167, + 448, + 180 + ], + "type": "text", + "content": "Akin Caliskan1, Berkay Kicanaoglu1, and Hyeongwoo Kim2" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 279, + 190, + 336, + 200 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 279, + 190, + 336, + 200 + ], + "spans": [ + { + "bbox": [ + 279, + 190, + 336, + 200 + ], + "type": "text", + "content": "1 Flawless AI" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 252, + 201, + 362, + 212 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 252, + 201, + 362, + 212 + ], + "spans": [ + { + "bbox": [ + 252, + 201, + 362, + 212 + ], + "type": "text", + "content": "2 Imperial College London" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 160, + 239, + 452, + 437 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 160, + 239, + 452, + 437 + ], + "spans": [ + { + "bbox": [ + 160, + 239, + 452, + 437 + ], + "type": "text", + "content": "Abstract. We propose PAV, Personalized Head Avatar for the synthesis of human faces under arbitrary viewpoints and facial expressions. PAV introduces a method that learns a dynamic deformable neural radiance field (NeRF), in particular from a collection of monocular talking face videos of the same character under various appearance and shape changes. Unlike existing head NeRF methods that are limited to modeling such input videos on a per-appearance basis, our method allows for learning multi-appearance NeRFs, introducing appearance embedding for each input video via learnable latent neural features attached to the underlying geometry. Furthermore, the proposed appearance-conditioned density formulation facilitates the shape variation of the character, such as facial hair and soft tissues, in the radiance field prediction. To the best of our knowledge, our approach is the first dynamic deformable NeRF framework to model appearance and shape variations in a single unified network for multi-appearances of the same subject. We demonstrate experimentally that PAV outperforms the baseline method in terms of visual rendering quality in our quantitative and qualitative studies on various subjects." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 133, + 474, + 229, + 487 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 474, + 229, + 487 + ], + "spans": [ + { + "bbox": [ + 133, + 474, + 229, + 487 + ], + "type": "text", + "content": "1 Introduction" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 130, + 498, + 482, + 668 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 498, + 482, + 668 + ], + "spans": [ + { + "bbox": [ + 130, + 498, + 482, + 668 + ], + "type": "text", + "content": "Digital human avatars enabling the editing of facial expressions and head motion have broad applications across telepresence, animation, and digital content creation. The demands for personalized human avatars such as digital doubles have also increased over the past few years not only in the film industry but in the research community. To enable the widespread adaptation of these neural avatars, they should be easy to generate and animate under novel poses and facial expressions. A vast majority of existing methods aim to learn neural avatars from monocular videos to make it more accessible and easy to use [1,10,12-14,16,34,42,44]. Early approaches [12,13,16,34] mainly rely on 3D morphable models (3DMM) to guide the modeling process and to provide an intuitive control mechanism in novel view synthesis. Despite their success, they do not model the 3D face geometry. To address this issue, neural radiance fields (NeRFs) [23] have demonstrated volumetric 3D modeling of face geometry and appearance [1,10,14,42,44] under novel poses and expressions. While this allows" + } + ] + } + ], + "index": 6 + } + ], + "discarded_blocks": [], + "page_size": [ + 612, + 792 + ], + "page_idx": 0 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 155, + 121, + 459, + 300 + ], + "blocks": [ + { + "bbox": [ + 155, + 121, + 459, + 300 + ], + "lines": [ + { + "bbox": [ + 155, + 121, + 459, + 300 + ], + "spans": [ + { + "bbox": [ + 155, + 121, + 459, + 300 + ], + "type": "image", + "image_path": "8c0e0140fb6c0cdb30c0c9fe988d7f54fe92f483434bea1be89f987b0d6d9713.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 130, + 310, + 482, + 410 + ], + "lines": [ + { + "bbox": [ + 130, + 310, + 482, + 410 + ], + "spans": [ + { + "bbox": [ + 130, + 310, + 482, + 410 + ], + "type": "text", + "content": "Fig. 1: We present an approach for learning personalized head avatars from an unstructured video collection of the same subject. The videos are taken in different environments, with diverse appearances and possible shape changes of a character, as shown on the left. Our method leverages all of the input clips into a single unified dynamic deformable NeRF framework and offers high-quality video replays in novel head poses and facial expressions. On the right, we show some rendering results of each input appearance of the same subject aligned in head pose and facial expression that are not seen during training. Notably, our method can synthesize identity-specific details while showing coherent expressions across different appearances." + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_caption" + } + ], + "index": 2 + }, + { + "bbox": [ + 130, + 436, + 482, + 592 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 436, + 482, + 592 + ], + "spans": [ + { + "bbox": [ + 130, + 436, + 482, + 592 + ], + "type": "text", + "content": "photo-realistic image quality and animation, NeRF-based approaches require per-appearance optimization, maintaining the consistent appearance of the subject, which is constrained to the pre-recording of characters in controlled settings. However, this limits the potential for practical application scenarios where we want to generate a 3D head avatar from a collection of prerecorded video clips of a subject. In this work, we present the Personalized Head Avatar method transforming an unstructured personal video collection, containing images spanning multiple years with different appearances, head poses, and expressions, into a 3D representation of the subject. To create an avatar, we only require an unstructured collection of video clips of the person in arbitrary poses, expressions, and appearances. Our novel framework enables creation of avatars from sparse observations, while yielding high reconstruction quality in novel combinations of pose and appearance, and opens up new applications." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 130, + 594, + 482, + 666 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 594, + 482, + 666 + ], + "spans": [ + { + "bbox": [ + 130, + 594, + 482, + 666 + ], + "type": "text", + "content": "Learning dynamic deformable head NeRF from unstructured video clips is a challenging problem. First, video collections can contain frames of the same identity at different appearances, poses, and expressions. Second, we only have limited pose and expression variation for each appearance, so it is unlikely that all poses and expressions of the face would be well-observed for any given appearance. Given the data, existing NeRF approaches' appearance-specific nature" + } + ] + } + ], + "index": 5 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "spans": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "type": "text", + "content": "2" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 166, + 91, + 238, + 101 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 166, + 91, + 238, + 101 + ], + "spans": [ + { + "bbox": [ + 166, + 91, + 238, + 101 + ], + "type": "text", + "content": "A. Caliskan et al." + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 1 + }, + { + "para_blocks": [ + { + "bbox": [ + 130, + 115, + 482, + 500 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 115, + 482, + 500 + ], + "spans": [ + { + "bbox": [ + 130, + 115, + 482, + 500 + ], + "type": "text", + "content": "largely limits their potential for practical application scenarios. They require an independent model that is trained from scratch for every appearance from each monocular video. To address these challenges, we model shared deformation field and canonical space neural volumetric representation for all appearances with appearance-dependent latent features. Our key insight is that although observed faces have different appearances across video clips, they should all be explained by a common deformation model with small differences since they all originate from the same person. Furthermore, we assert that every video clip specific to an appearance contains both general and unique information. A single unified network could collaboratively learn the generic information from various appearances, thus this benefits rendering performance as well. The single network is designed to learn a deformation-aware dynamic head NeRF while disentangling appearance-related details from those that are not appearance-specific. For appearance conditioning, we propose employing learnable appearance embedding attached to vertices of the tracked mesh and passing them to the dynamic head NeRF network. By adopting this process, PAV is able to learn discriminative latent neural features that are specific for each appearance, resulting in better rendering of appearance-specific details for arbitrary poses and expressions. Unlike existing NeRF methods [1, 10, 14, 42, 44], PAV simultaneously trains on all appearances of the same subject, eliminating the need for separate training sessions for each appearance and reducing the overall training duration. Learning of personalized head avatars in a single unified network from an unstructured collection of videos has not been addressed before and is a non-trivial extension of existing monocular head avatar generation methods since they can not handle multiple appearances in a single network. PAV makes this possible through a carefully designed appearance conditioning representation, geometry-attached trainable latent neural texture to learn appearance-specific features, appearance conditioned implicit functions to predict variations in color and density. We conducted a variety of experiments to show these improvements over the baseline approach. Since there is no such public dataset to evaluate our method, we curate our own in-the-wild video collections for multiple subjects. Overall, we present Personalized Head Avatar with the following contributions:" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 138, + 544, + 482, + 666 + ], + "type": "list", + "angle": 0, + "index": 6, + "blocks": [ + { + "bbox": [ + 138, + 544, + 482, + 605 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 544, + 482, + 605 + ], + "spans": [ + { + "bbox": [ + 138, + 544, + 482, + 605 + ], + "type": "text", + "content": "- We present, PAV, a framework for learning controllable head avatars from unconstrained short video collections where the same person shows different appearances and facial expressions. To the best of our knowledge, this is the first method learning appearance controllable head avatars from in-the-wild monocular video collections of the same subject." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 138, + 610, + 482, + 635 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 610, + 482, + 635 + ], + "spans": [ + { + "bbox": [ + 138, + 610, + 482, + 635 + ], + "type": "text", + "content": "- The proposed method can synthesise images of head avatar under target appearance, head pose and facial expressions." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 138, + 641, + 482, + 666 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 641, + 482, + 666 + ], + "spans": [ + { + "bbox": [ + 138, + 641, + 482, + 666 + ], + "type": "text", + "content": "- PAV learns modeling local changes of geometry and color across different appearances of the same identity." + } + ] + } + ], + "index": 5 + } + ], + "sub_type": "text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 226, + 91, + 447, + 101 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 226, + 91, + 447, + 101 + ], + "spans": [ + { + "bbox": [ + 226, + 91, + 447, + 101 + ], + "type": "text", + "content": "PAV: Personalized Head Avatar from Video Collection" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 474, + 92, + 480, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 474, + 92, + 480, + 100 + ], + "spans": [ + { + "bbox": [ + 474, + 92, + 480, + 100 + ], + "type": "text", + "content": "3" + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 2 + }, + { + "para_blocks": [ + { + "bbox": [ + 132, + 114, + 237, + 127 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 114, + 237, + 127 + ], + "spans": [ + { + "bbox": [ + 132, + 114, + 237, + 127 + ], + "type": "text", + "content": "2 Related Work" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 130, + 137, + 482, + 631 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 137, + 482, + 631 + ], + "spans": [ + { + "bbox": [ + 130, + 137, + 482, + 631 + ], + "type": "text", + "content": "Monocular Neural Head Avatars. More recent methods learn neural representations of the person from monocular videos and use neural renderers [31, 32] to directly generate photo-realistic images in the target head pose and facial expression [1-3, 5, 10, 11, 14, 16, 21, 29, 33, 40, 42, 44]. These methods are generally classified into image-based or 3D-based neural rendering methods [31, 32]. The image-based models synthesize the face of a subject without relying on any representation of 3D space. Existing image-based approaches either employ learned warping fields [2, 29] to deform an input image or encoder-decoder architecture where decoder synthesize the output image [21, 25, 35, 39]. Even though these methods produce high-quality results, they suffer from artifacts for strong head pose and expression changes due to the lack of 3D geometry. The 3D-based neural rendering methods, on the other hand, rely on 3D Morphable Models (3DMM) [9]. These methods are generally classified into explicit and implicit models. Explicit models utilize 2D neural rendering methods [31, 32] and rely on neural networks which can be conditioned on coarse RGB renderings based on a linear texture model [16], uv-maps [4] or latent feature maps [33]. Although these techniques produce high-quality renderings, they are confined to operating within a 2D image space and are limited to craniofacial structures. To tackle this, implicit neural representations for modeling 3D faces have recently become a major focus of attention. In particular, NeRFs [23] shows photorealistic face image synthesis in novel poses and expressions [1, 10, 14, 42, 44]. Implicit models represent the geometry using implicit surface functions or volumetric representations and they are optimized via volumetric rendering [23]. Early approaches [23] and follow-up works synthesize novel views of complex static scenes using differentiable volumetric rendering. After introduction of NeRF for static scenes, a natural research direction was to generalize it to dynamic and time-varying ones [26, 28]. The reconstruction problem is divided into two different spaces, the deformed space and the canonical space. For human body, a series of approaches have been proposed to target this challenge [20, 27]. In particular, these methods are differentiated from each other based on deformation field modeling. Gafni et al. [10] introduced implicit deformation modeling conditioned on facial expression and head pose. In contrast to modeling the deformation field implicitly, IMAvatar [42], RigNerf [1] and Insta [44] propose explicit modeling of the deformation field to improve control and rendering quality. Our method falls into this paradigm as we build our system on top of dynamic deformable head NeRF. In particular, we take inspiration from Insta [44] for designing our framework. However, existing deformable dynamic head NeRF methods including Insta [44] require per-identity optimization, maintaining the consistent appearance of the same subject. To the best of our knowledge, PAV is the first method to learn a single unified dynamic deformable head NeRF from monocular video of multiple appearances of the same subject." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 132, + 641, + 484, + 666 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 641, + 484, + 666 + ], + "spans": [ + { + "bbox": [ + 132, + 641, + 484, + 666 + ], + "type": "text", + "content": "Geometry-anchored Features for NeRFs. PAV introduces appearance-conditioned head NeRF by using geometry-attached features. By using these features, exist" + } + ] + } + ], + "index": 4 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "spans": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "type": "text", + "content": "4" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 166, + 91, + 238, + 101 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 166, + 91, + 238, + 101 + ], + "spans": [ + { + "bbox": [ + 166, + 91, + 238, + 101 + ], + "type": "text", + "content": "A. Caliskan et al." + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 3 + }, + { + "para_blocks": [ + { + "bbox": [ + 130, + 116, + 479, + 152 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 116, + 479, + 152 + ], + "spans": [ + { + "bbox": [ + 130, + 116, + 479, + 152 + ], + "type": "text", + "content": "ing works show improved rendering quality for static objects [38] and human bodies [3, 20]. However, none of the existing methods [3, 20] has yet addressed multi-appearance learning for dynamic neural radiance fields." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 130, + 163, + 480, + 270 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 163, + 480, + 270 + ], + "spans": [ + { + "bbox": [ + 130, + 163, + 480, + 270 + ], + "type": "text", + "content": "NeRF from Unstructured Data Collections. Learning neural radiance field from unstructured data collections has recently gained much attention. Existing works have enabled photorealistic view synthesis of challenging static scenes, including tourist sites [22] and city-scale scenes [30], and static objects [18] from internet photo collections. For human subjects, Weng et al. [37] introduce free-viewpoint human body rendering from an unstructured photo collection. Despite the photorealistic rendering quality of PersonNerf [37], they do not model animatable faces due to the static training images. In our work, we learn animatable 3D head avatars from video collections." + } + ] + } + ], + "index": 3 + }, + { + "type": "image", + "bbox": [ + 134, + 290, + 480, + 415 + ], + "blocks": [ + { + "bbox": [ + 134, + 290, + 480, + 415 + ], + "lines": [ + { + "bbox": [ + 134, + 290, + 480, + 415 + ], + "spans": [ + { + "bbox": [ + 134, + 290, + 480, + 415 + ], + "type": "image", + "image_path": "198902175a488fa1839da38f7a9d3bdbf7958a8df09f4111f5c5b2903b36a25a.jpg" + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 130, + 423, + 482, + 499 + ], + "lines": [ + { + "bbox": [ + 130, + 423, + 482, + 499 + ], + "spans": [ + { + "bbox": [ + 130, + 423, + 482, + 499 + ], + "type": "text", + "content": "Fig. 2: Overview of the proposed approach, PAV. (a) shows some training videos in various appearances of the same subject. (b) We estimate the 3D head geometry of the person using FLAME head model. (c) shows the proposed personalized deformable neural radiance field from video collection (d) We learn appearance-specific neural features and attach them to vertices. (d) This feature is passed to fully fused multi-layer perceptrons with additional conditioning on the facial expressions " + }, + { + "bbox": [ + 130, + 423, + 482, + 499 + ], + "type": "inline_equation", + "content": "e" + }, + { + "bbox": [ + 130, + 423, + 482, + 499 + ], + "type": "text", + "content": " and the encoded view direction " + }, + { + "bbox": [ + 130, + 423, + 482, + 499 + ], + "type": "inline_equation", + "content": "d" + }, + { + "bbox": [ + 130, + 423, + 482, + 499 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "image_caption" + } + ], + "index": 4 + }, + { + "bbox": [ + 132, + 544, + 201, + 556 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 544, + 201, + 556 + ], + "spans": [ + { + "bbox": [ + 132, + 544, + 201, + 556 + ], + "type": "text", + "content": "3 Method" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 130, + 570, + 482, + 666 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 570, + 482, + 666 + ], + "spans": [ + { + "bbox": [ + 130, + 570, + 482, + 666 + ], + "type": "text", + "content": "Our goal is to learn a personalized head avatar, PAV, from a monocular video collection of a subject under different appearances, poses, and facial expressions. To create the avatar, we use a small collection " + }, + { + "bbox": [ + 130, + 570, + 482, + 666 + ], + "type": "inline_equation", + "content": "\\mathbf{V} = \\{V_{i}\\}_{i = 1}^{N}" + }, + { + "bbox": [ + 130, + 570, + 482, + 666 + ], + "type": "text", + "content": " with " + }, + { + "bbox": [ + 130, + 570, + 482, + 666 + ], + "type": "inline_equation", + "content": "N" + }, + { + "bbox": [ + 130, + 570, + 482, + 666 + ], + "type": "text", + "content": " videos. We parameterize PAV using FLAME [19] face model to control head poses, facial expressions and use neural texture to model appearance. During inference, we only need the target FLAME parameters and neural texture. Learning PAV from a collection of monocular videos requires capturing the geometry and appearance of the dynamic head while also disentangling different appearances of the same" + } + ] + } + ], + "index": 7 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 225, + 91, + 447, + 101 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 225, + 91, + 447, + 101 + ], + "spans": [ + { + "bbox": [ + 225, + 91, + 447, + 101 + ], + "type": "text", + "content": "PAV: Personalized Head Avatar from Video Collection" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 474, + 92, + 480, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 474, + 92, + 480, + 100 + ], + "spans": [ + { + "bbox": [ + 474, + 92, + 480, + 100 + ], + "type": "text", + "content": "5" + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 4 + }, + { + "para_blocks": [ + { + "bbox": [ + 130, + 116, + 482, + 213 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 116, + 482, + 213 + ], + "spans": [ + { + "bbox": [ + 130, + 116, + 482, + 213 + ], + "type": "text", + "content": "identity. In order to tackle this problem, we first capture the dynamic 3D geometry by employing a parametric head model [19] (Sec. 3.1). We use the coarse geometry to model geometry-guided deformable neural radiance field (Sec. 3.2) and vertex-attached neural features to condition appearance for rendering (Sec. 3.3). We also propose estimating appearance-conditioned neural radiance field, aimed at accurately capturing distinct color and geometry with high fidelity. (Sec. 3.4). An overview of our method can be seen in Fig. 2. In the following, we describe each of these modules in greater detail." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 131, + 228, + 299, + 240 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 131, + 228, + 299, + 240 + ], + "spans": [ + { + "bbox": [ + 131, + 228, + 299, + 240 + ], + "type": "text", + "content": "3.1 Head Geometry Estimation" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 130, + 245, + 482, + 342 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 245, + 482, + 342 + ], + "spans": [ + { + "bbox": [ + 130, + 245, + 482, + 342 + ], + "type": "text", + "content": "Given the training frames " + }, + { + "bbox": [ + 130, + 245, + 482, + 342 + ], + "type": "inline_equation", + "content": "\\mathbf{I} = \\{I_{\\mathrm{i}}\\}" + }, + { + "bbox": [ + 130, + 245, + 482, + 342 + ], + "type": "text", + "content": ", from all appearances, we first estimate the optimized camera parameters " + }, + { + "bbox": [ + 130, + 245, + 482, + 342 + ], + "type": "inline_equation", + "content": "\\mathbf{K} \\in \\mathbb{R}^{3 \\times 3}" + }, + { + "bbox": [ + 130, + 245, + 482, + 342 + ], + "type": "text", + "content": ", tracked FLAME [19] meshes " + }, + { + "bbox": [ + 130, + 245, + 482, + 342 + ], + "type": "inline_equation", + "content": "\\mathbf{M} = \\{M_{\\mathrm{i}}\\}" + }, + { + "bbox": [ + 130, + 245, + 482, + 342 + ], + "type": "text", + "content": " with corresponding facial expression coefficients " + }, + { + "bbox": [ + 130, + 245, + 482, + 342 + ], + "type": "inline_equation", + "content": "\\mathbf{E} = \\{E_{\\mathrm{i}}\\}" + }, + { + "bbox": [ + 130, + 245, + 482, + 342 + ], + "type": "text", + "content": " and poses " + }, + { + "bbox": [ + 130, + 245, + 482, + 342 + ], + "type": "inline_equation", + "content": "\\mathbf{P} = \\{P_{\\mathrm{i}}\\}" + }, + { + "bbox": [ + 130, + 245, + 482, + 342 + ], + "type": "text", + "content": ". We follow an optimization-based method, that minimizes an objective function with photo-consistency and landmark terms [43]. Different from the original work [43] which is designed for a single appearance, in PAV, we compute the average shape parameters from all appearances of the same subject to represent identity shape during optimization." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 131, + 357, + 320, + 369 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 131, + 357, + 320, + 369 + ], + "spans": [ + { + "bbox": [ + 131, + 357, + 320, + 369 + ], + "type": "text", + "content": "3.2 Personalized Deformable NeRF" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 130, + 374, + 482, + 447 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 374, + 482, + 447 + ], + "spans": [ + { + "bbox": [ + 130, + 374, + 482, + 447 + ], + "type": "text", + "content": "Our goal is to create a head avatar that can be learned from frames in video collection of various appearance of the same subject. We parameterize PAV using FLAME model and learned appearance conditioning. Specifically, PAV takes target head pose " + }, + { + "bbox": [ + 130, + 374, + 482, + 447 + ], + "type": "inline_equation", + "content": "\\theta" + }, + { + "bbox": [ + 130, + 374, + 482, + 447 + ], + "type": "text", + "content": ", facial expression " + }, + { + "bbox": [ + 130, + 374, + 482, + 447 + ], + "type": "inline_equation", + "content": "e" + }, + { + "bbox": [ + 130, + 374, + 482, + 447 + ], + "type": "text", + "content": " and latent neural features " + }, + { + "bbox": [ + 130, + 374, + 482, + 447 + ], + "type": "inline_equation", + "content": "(LNF)" + }, + { + "bbox": [ + 130, + 374, + 482, + 447 + ], + "type": "inline_equation", + "content": "Z_{j} \\in \\mathbb{R}^{512 \\times 512 \\times 16}" + }, + { + "bbox": [ + 130, + 374, + 482, + 447 + ], + "type": "text", + "content": " in the texture atlas space (UV space) for an appearance " + }, + { + "bbox": [ + 130, + 374, + 482, + 447 + ], + "type": "inline_equation", + "content": "j" + }, + { + "bbox": [ + 130, + 374, + 482, + 447 + ], + "type": "text", + "content": " as inputs and synthesizes the target image " + }, + { + "bbox": [ + 130, + 374, + 482, + 447 + ], + "type": "inline_equation", + "content": "I^{\\theta, e, Z_{j}}" + }, + { + "bbox": [ + 130, + 374, + 482, + 447 + ], + "type": "text", + "content": " as:" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 252, + 453, + 481, + 466 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 252, + 453, + 481, + 466 + ], + "spans": [ + { + "bbox": [ + 252, + 453, + 481, + 466 + ], + "type": "interline_equation", + "content": "I ^ {\\theta , e, Z _ {j}} = \\operatorname {P A V} (\\theta , e, Z _ {j}, K), \\tag {1}", + "image_path": "2ccef1ce0bcaf96dc1fcb96cfd38bf4bc770d9f59b53712e374d1044203a6d94.jpg" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 130, + 472, + 482, + 666 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 472, + 482, + 666 + ], + "spans": [ + { + "bbox": [ + 130, + 472, + 482, + 666 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 130, + 472, + 482, + 666 + ], + "type": "inline_equation", + "content": "K" + }, + { + "bbox": [ + 130, + 472, + 482, + 666 + ], + "type": "text", + "content": " corresponds to the intrinsic parameters of the target camera viewpoint. For this purpose, we are using a geometry-guided dynamic deformable neural radiance field. However, deformable head NeRF models cannot disentangle appearance from the neural radiance field (See Sec. 2). To tackle this problem, we employ shared canonical space for all appearances where a shared neural radiance field is constructed. With this shared radiance field at canonical space, to render the target head pose and facial expression using volumetric rendering, we canon- icalize the sample points " + }, + { + "bbox": [ + 130, + 472, + 482, + 666 + ], + "type": "inline_equation", + "content": "x" + }, + { + "bbox": [ + 130, + 472, + 482, + 666 + ], + "type": "text", + "content": " on a ray from the observation space to query points " + }, + { + "bbox": [ + 130, + 472, + 482, + 666 + ], + "type": "inline_equation", + "content": "x_{c}" + }, + { + "bbox": [ + 130, + 472, + 482, + 666 + ], + "type": "text", + "content": " in the shared canonical space radiance field by using deformation gradient " + }, + { + "bbox": [ + 130, + 472, + 482, + 666 + ], + "type": "inline_equation", + "content": "\\hat{F}" + }, + { + "bbox": [ + 130, + 472, + 482, + 666 + ], + "type": "text", + "content": " .. " + }, + { + "bbox": [ + 130, + 472, + 482, + 666 + ], + "type": "inline_equation", + "content": "x_{c} = \\hat{F}\\cdot x" + }, + { + "bbox": [ + 130, + 472, + 482, + 666 + ], + "type": "text", + "content": " where " + }, + { + "bbox": [ + 130, + 472, + 482, + 666 + ], + "type": "inline_equation", + "content": "\\hat{F}\\in \\mathbb{R}^{4\\times 4}" + }, + { + "bbox": [ + 130, + 472, + 482, + 666 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 130, + 472, + 482, + 666 + ], + "type": "inline_equation", + "content": "\\{x,x_c\\} \\in \\mathbb{R}^4" + }, + { + "bbox": [ + 130, + 472, + 482, + 666 + ], + "type": "text", + "content": " . We employ the nearest triangle search for a given point to compute the deformation gradient. Specifically, the deformation gradient is computed between the deformed triangle " + }, + { + "bbox": [ + 130, + 472, + 482, + 666 + ], + "type": "inline_equation", + "content": "\\hat{T}_{def}\\in M_i" + }, + { + "bbox": [ + 130, + 472, + 482, + 666 + ], + "type": "text", + "content": " in the observed mesh and the canonical triangle " + }, + { + "bbox": [ + 130, + 472, + 482, + 666 + ], + "type": "inline_equation", + "content": "\\hat{T}_{canon}\\in M^{canon}" + }, + { + "bbox": [ + 130, + 472, + 482, + 666 + ], + "type": "text", + "content": " in the canonical mesh. To increase the speed of nearest triangle search, we employ bounding volume hierarchy (BVH) [6]. We refer readers to [44] for a detailed description of deformation gradient computation." + } + ] + } + ], + "index": 8 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "spans": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "type": "text", + "content": "6" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 166, + 91, + 238, + 101 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 166, + 91, + 238, + 101 + ], + "spans": [ + { + "bbox": [ + 166, + 91, + 238, + 101 + ], + "type": "text", + "content": "A. Caliskan et al." + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 5 + }, + { + "para_blocks": [ + { + "bbox": [ + 130, + 116, + 479, + 188 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 116, + 479, + 188 + ], + "spans": [ + { + "bbox": [ + 130, + 116, + 479, + 188 + ], + "type": "text", + "content": "After canonicalization, given the predicted appearance-conditioned radiance " + }, + { + "bbox": [ + 130, + 116, + 479, + 188 + ], + "type": "inline_equation", + "content": "c" + }, + { + "bbox": [ + 130, + 116, + 479, + 188 + ], + "type": "text", + "content": " and density " + }, + { + "bbox": [ + 130, + 116, + 479, + 188 + ], + "type": "inline_equation", + "content": "\\sigma" + }, + { + "bbox": [ + 130, + 116, + 479, + 188 + ], + "type": "text", + "content": " (Sec. 3.4) for every point on each ray, the representation of the avatar is optimized by using the differentiable volumetric rendering. For each camera ray " + }, + { + "bbox": [ + 130, + 116, + 479, + 188 + ], + "type": "inline_equation", + "content": "r(t) = o + td" + }, + { + "bbox": [ + 130, + 116, + 479, + 188 + ], + "type": "text", + "content": " with camera center " + }, + { + "bbox": [ + 130, + 116, + 479, + 188 + ], + "type": "inline_equation", + "content": "o" + }, + { + "bbox": [ + 130, + 116, + 479, + 188 + ], + "type": "text", + "content": " and viewing direction " + }, + { + "bbox": [ + 130, + 116, + 479, + 188 + ], + "type": "inline_equation", + "content": "d" + }, + { + "bbox": [ + 130, + 116, + 479, + 188 + ], + "type": "text", + "content": ", the color of the corresponding pixel can be computed by accumulating the predicting radiance and densities of the sample points along the ray:" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 230, + 198, + 481, + 224 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 230, + 198, + 481, + 224 + ], + "spans": [ + { + "bbox": [ + 230, + 198, + 481, + 224 + ], + "type": "interline_equation", + "content": "\\mathcal {C} (\\mathbf {r}) = \\int_ {t _ {n}} ^ {t _ {f}} \\sigma (\\mathbf {r} (t)) \\cdot \\mathbf {c} (\\mathbf {r} (t)) \\cdot \\mathcal {T} (t) d t \\tag {2}", + "image_path": "b0487494cac465260fd9bce0665f89255af8ba88669180bfacedd88456b658bc.jpg" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 131, + 228, + 479, + 268 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 131, + 228, + 479, + 268 + ], + "spans": [ + { + "bbox": [ + 131, + 228, + 479, + 268 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 131, + 228, + 479, + 268 + ], + "type": "inline_equation", + "content": "t_n" + }, + { + "bbox": [ + 131, + 228, + 479, + 268 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 131, + 228, + 479, + 268 + ], + "type": "inline_equation", + "content": "t_f" + }, + { + "bbox": [ + 131, + 228, + 479, + 268 + ], + "type": "text", + "content": " are the near and far bounds correspondingly, and " + }, + { + "bbox": [ + 131, + 228, + 479, + 268 + ], + "type": "inline_equation", + "content": "\\mathcal{T}(t_n) = \\exp \\left(-\\int_{t_n}^{t_f}\\sigma (t)dt\\right)" + }, + { + "bbox": [ + 131, + 228, + 479, + 268 + ], + "type": "text", + "content": " is the accumulated transmittance along the ray from " + }, + { + "bbox": [ + 131, + 228, + 479, + 268 + ], + "type": "inline_equation", + "content": "t_n" + }, + { + "bbox": [ + 131, + 228, + 479, + 268 + ], + "type": "text", + "content": " to " + }, + { + "bbox": [ + 131, + 228, + 479, + 268 + ], + "type": "inline_equation", + "content": "t_f" + }, + { + "bbox": [ + 131, + 228, + 479, + 268 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 131, + 228, + 479, + 268 + ], + "type": "inline_equation", + "content": "\\sigma (t)" + }, + { + "bbox": [ + 131, + 228, + 479, + 268 + ], + "type": "text", + "content": " is the density and " + }, + { + "bbox": [ + 131, + 228, + 479, + 268 + ], + "type": "inline_equation", + "content": "c(t)" + }, + { + "bbox": [ + 131, + 228, + 479, + 268 + ], + "type": "text", + "content": " it the radiance at position " + }, + { + "bbox": [ + 131, + 228, + 479, + 268 + ], + "type": "inline_equation", + "content": "x_t" + }, + { + "bbox": [ + 131, + 228, + 479, + 268 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 131, + 285, + 326, + 298 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 131, + 285, + 326, + 298 + ], + "spans": [ + { + "bbox": [ + 131, + 285, + 326, + 298 + ], + "type": "text", + "content": "3.3 Learned Appearance Embedding" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 130, + 304, + 482, + 483 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 304, + 482, + 483 + ], + "spans": [ + { + "bbox": [ + 130, + 304, + 482, + 483 + ], + "type": "text", + "content": "In this section, PAV introduces a learnable appearance embedding that is essential and effective for enabling discriminative multi-appearance rendering using single unified network. To achieve this goal (See Fig. 2), a set of " + }, + { + "bbox": [ + 130, + 304, + 482, + 483 + ], + "type": "inline_equation", + "content": "N" + }, + { + "bbox": [ + 130, + 304, + 482, + 483 + ], + "type": "text", + "content": " learnable latent features " + }, + { + "bbox": [ + 130, + 304, + 482, + 483 + ], + "type": "inline_equation", + "content": "\\mathbf{Z} = \\{Z_{\\mathrm{i}}\\}_{i = 1}^{N}" + }, + { + "bbox": [ + 130, + 304, + 482, + 483 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 130, + 304, + 482, + 483 + ], + "type": "inline_equation", + "content": "Z_{i}\\in \\mathbb{R}^{512\\times 512\\times 16}" + }, + { + "bbox": [ + 130, + 304, + 482, + 483 + ], + "type": "text", + "content": " are defined in UV space corresponding to " + }, + { + "bbox": [ + 130, + 304, + 482, + 483 + ], + "type": "inline_equation", + "content": "N" + }, + { + "bbox": [ + 130, + 304, + 482, + 483 + ], + "type": "text", + "content": " appearances of the same subject in the video collection. We impose high-dimensional learnable latent features to condition implicit representations for both density " + }, + { + "bbox": [ + 130, + 304, + 482, + 483 + ], + "type": "inline_equation", + "content": "\\sigma" + }, + { + "bbox": [ + 130, + 304, + 482, + 483 + ], + "type": "text", + "content": " and radiance " + }, + { + "bbox": [ + 130, + 304, + 482, + 483 + ], + "type": "inline_equation", + "content": "c" + }, + { + "bbox": [ + 130, + 304, + 482, + 483 + ], + "type": "text", + "content": " values in the canonical space. To this end, these latent neural features are attached to the vertices of 3DMM mesh, which enables PAV to memorize local variations. For each point " + }, + { + "bbox": [ + 130, + 304, + 482, + 483 + ], + "type": "inline_equation", + "content": "x" + }, + { + "bbox": [ + 130, + 304, + 482, + 483 + ], + "type": "text", + "content": " sampled on a ray, we find the nearest triangle " + }, + { + "bbox": [ + 130, + 304, + 482, + 483 + ], + "type": "inline_equation", + "content": "\\mathcal{T} = \\{v_{\\mathrm{i}}\\}_{i = 1}^{3}, v_{i}\\in \\mathbb{R}^{3}" + }, + { + "bbox": [ + 130, + 304, + 482, + 483 + ], + "type": "text", + "content": " (sec. 3.2) and aggregate the features " + }, + { + "bbox": [ + 130, + 304, + 482, + 483 + ], + "type": "inline_equation", + "content": "z(v_{i})" + }, + { + "bbox": [ + 130, + 304, + 482, + 483 + ], + "type": "text", + "content": " of vertices " + }, + { + "bbox": [ + 130, + 304, + 482, + 483 + ], + "type": "inline_equation", + "content": "v_{i}" + }, + { + "bbox": [ + 130, + 304, + 482, + 483 + ], + "type": "text", + "content": " from the vertices of the nearest triangle to get the appearance embedding " + }, + { + "bbox": [ + 130, + 304, + 482, + 483 + ], + "type": "inline_equation", + "content": "Z^x" + }, + { + "bbox": [ + 130, + 304, + 482, + 483 + ], + "type": "text", + "content": ". Then we directly use the appearance embedding to condition for appearance-dependent neural radiance field estimation (Sec. 3.4). During training, these features are randomly initialized and optimized jointly with the objectives of the entire network." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 131, + 501, + 320, + 514 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 131, + 501, + 320, + 514 + ], + "spans": [ + { + "bbox": [ + 131, + 501, + 320, + 514 + ], + "type": "text", + "content": "3.4 Appearance Conditioned NeRF" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 130, + 521, + 482, + 628 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 521, + 482, + 628 + ], + "spans": [ + { + "bbox": [ + 130, + 521, + 482, + 628 + ], + "type": "text", + "content": "To model the appearance-conditioned dynamic deformable head NeRF, PAV incorporates learned appearance-embedding (Sec. 3.3) into neural radiance field in canonical space (See Fig. 2). For a point " + }, + { + "bbox": [ + 130, + 521, + 482, + 628 + ], + "type": "inline_equation", + "content": "x" + }, + { + "bbox": [ + 130, + 521, + 482, + 628 + ], + "type": "text", + "content": " sampled on a ray, after canonicalization, the point " + }, + { + "bbox": [ + 130, + 521, + 482, + 628 + ], + "type": "inline_equation", + "content": "x_{c}" + }, + { + "bbox": [ + 130, + 521, + 482, + 628 + ], + "type": "text", + "content": " is encoded using multi-resolution hash encoding [24]. This feature " + }, + { + "bbox": [ + 130, + 521, + 482, + 628 + ], + "type": "inline_equation", + "content": "\\phi(x_{c})" + }, + { + "bbox": [ + 130, + 521, + 482, + 628 + ], + "type": "text", + "content": " is passed to multi-layer perceptron for geometric feature estimation. Following [10] and [44], we conditioned every point features " + }, + { + "bbox": [ + 130, + 521, + 482, + 628 + ], + "type": "inline_equation", + "content": "\\phi(x_{c})" + }, + { + "bbox": [ + 130, + 521, + 482, + 628 + ], + "type": "text", + "content": " with the facial expressions " + }, + { + "bbox": [ + 130, + 521, + 482, + 628 + ], + "type": "inline_equation", + "content": "e" + }, + { + "bbox": [ + 130, + 521, + 482, + 628 + ], + "type": "text", + "content": " of frame " + }, + { + "bbox": [ + 130, + 521, + 482, + 628 + ], + "type": "inline_equation", + "content": "i" + }, + { + "bbox": [ + 130, + 521, + 482, + 628 + ], + "type": "text", + "content": ". Without per-frame latent code, for shared volume density, we learn implicit representation " + }, + { + "bbox": [ + 130, + 521, + 482, + 628 + ], + "type": "inline_equation", + "content": "\\mathcal{F}_{\\text{density}}" + }, + { + "bbox": [ + 130, + 521, + 482, + 628 + ], + "type": "text", + "content": " using multi-layer perceptron (MLP-D in Fig. 2):" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 253, + 651, + 480, + 662 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 253, + 651, + 480, + 662 + ], + "spans": [ + { + "bbox": [ + 253, + 651, + 480, + 662 + ], + "type": "interline_equation", + "content": "\\mathcal {F} _ {\\text {d e n s i t y}}: \\left(\\phi \\left(x _ {c}\\right), e _ {i}\\right)\\rightarrow \\sigma \\tag {3}", + "image_path": "d0541a994c5f1126daf339faf4c61e5a349556e5aa9e6612d81e485686707aad.jpg" + } + ] + } + ], + "index": 9 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 225, + 91, + 447, + 101 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 225, + 91, + 447, + 101 + ], + "spans": [ + { + "bbox": [ + 225, + 91, + 447, + 101 + ], + "type": "text", + "content": "PAV: Personalized Head Avatar from Video Collection" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 474, + 92, + 480, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 474, + 92, + 480, + 100 + ], + "spans": [ + { + "bbox": [ + 474, + 92, + 480, + 100 + ], + "type": "text", + "content": "7" + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 6 + }, + { + "para_blocks": [ + { + "bbox": [ + 130, + 116, + 482, + 189 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 116, + 482, + 189 + ], + "spans": [ + { + "bbox": [ + 130, + 116, + 482, + 189 + ], + "type": "text", + "content": "A primary challenge lies in learning unique density estimations for each appearance while separating the shared density aspects from the appearance. Hence we design PAV such that appearance-based density offset is predicted from appearance embedding " + }, + { + "bbox": [ + 130, + 116, + 482, + 189 + ], + "type": "inline_equation", + "content": "Z^{x}" + }, + { + "bbox": [ + 130, + 116, + 482, + 189 + ], + "type": "text", + "content": " sampled from the nearest triangle to the point " + }, + { + "bbox": [ + 130, + 116, + 482, + 189 + ], + "type": "inline_equation", + "content": "x" + }, + { + "bbox": [ + 130, + 116, + 482, + 189 + ], + "type": "text", + "content": " on the mesh " + }, + { + "bbox": [ + 130, + 116, + 482, + 189 + ], + "type": "inline_equation", + "content": "M_{i}" + }, + { + "bbox": [ + 130, + 116, + 482, + 189 + ], + "type": "text", + "content": " and facial expression " + }, + { + "bbox": [ + 130, + 116, + 482, + 189 + ], + "type": "inline_equation", + "content": "e_{i}" + }, + { + "bbox": [ + 130, + 116, + 482, + 189 + ], + "type": "text", + "content": " of frame " + }, + { + "bbox": [ + 130, + 116, + 482, + 189 + ], + "type": "inline_equation", + "content": "i" + }, + { + "bbox": [ + 130, + 116, + 482, + 189 + ], + "type": "text", + "content": ". For density offset, we learn implicit representation " + }, + { + "bbox": [ + 130, + 116, + 482, + 189 + ], + "type": "inline_equation", + "content": "\\mathcal{F}_{offset}" + }, + { + "bbox": [ + 130, + 116, + 482, + 189 + ], + "type": "text", + "content": " using multi-layer perceptron (MLP-O in Fig. 2):" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 256, + 198, + 481, + 211 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 256, + 198, + 481, + 211 + ], + "spans": [ + { + "bbox": [ + 256, + 198, + 481, + 211 + ], + "type": "interline_equation", + "content": "\\mathcal {F} _ {\\text {o f f s e t}}: \\left(Z ^ {x}, e _ {i}\\right)\\rightarrow \\Delta_ {\\sigma} \\tag {4}", + "image_path": "62cd2c6c406bcf260322686b67e37524655dc14a43feac53d3b197b410365b89.jpg" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 130, + 220, + 482, + 267 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 220, + 482, + 267 + ], + "spans": [ + { + "bbox": [ + 130, + 220, + 482, + 267 + ], + "type": "text", + "content": "Final density value " + }, + { + "bbox": [ + 130, + 220, + 482, + 267 + ], + "type": "inline_equation", + "content": "\\sigma(x) = \\sigma + \\Delta_{\\sigma}" + }, + { + "bbox": [ + 130, + 220, + 482, + 267 + ], + "type": "text", + "content": " for the point " + }, + { + "bbox": [ + 130, + 220, + 482, + 267 + ], + "type": "inline_equation", + "content": "x" + }, + { + "bbox": [ + 130, + 220, + 482, + 267 + ], + "type": "text", + "content": " is the summation of shared density features " + }, + { + "bbox": [ + 130, + 220, + 482, + 267 + ], + "type": "inline_equation", + "content": "\\sigma" + }, + { + "bbox": [ + 130, + 220, + 482, + 267 + ], + "type": "text", + "content": " and density offset " + }, + { + "bbox": [ + 130, + 220, + 482, + 267 + ], + "type": "inline_equation", + "content": "\\Delta_{\\sigma}" + }, + { + "bbox": [ + 130, + 220, + 482, + 267 + ], + "type": "text", + "content": ". For emitted radiance at point " + }, + { + "bbox": [ + 130, + 220, + 482, + 267 + ], + "type": "inline_equation", + "content": "x" + }, + { + "bbox": [ + 130, + 220, + 482, + 267 + ], + "type": "text", + "content": ", we learn implicit representation " + }, + { + "bbox": [ + 130, + 220, + 482, + 267 + ], + "type": "inline_equation", + "content": "\\mathcal{F}_{color}" + }, + { + "bbox": [ + 130, + 220, + 482, + 267 + ], + "type": "text", + "content": " using multi-layer perceptron (MLP-C in Fig. 2):" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 251, + 277, + 481, + 291 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 251, + 277, + 481, + 291 + ], + "spans": [ + { + "bbox": [ + 251, + 277, + 481, + 291 + ], + "type": "interline_equation", + "content": "\\mathcal {F} _ {\\text {c o l o r}}: \\left(Z ^ {x}, d _ {i}, \\sigma (x)\\right)\\rightarrow c \\tag {5}", + "image_path": "0253fb7c920f059fef5c0bea583e958cb4b643a8a0fa40b1eb87e14e6514b79c.jpg" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 131, + 299, + 481, + 324 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 131, + 299, + 481, + 324 + ], + "spans": [ + { + "bbox": [ + 131, + 299, + 481, + 324 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 131, + 299, + 481, + 324 + ], + "type": "inline_equation", + "content": "d_{i}" + }, + { + "bbox": [ + 131, + 299, + 481, + 324 + ], + "type": "text", + "content": " is the encoded view direction of training frame " + }, + { + "bbox": [ + 131, + 299, + 481, + 324 + ], + "type": "inline_equation", + "content": "i" + }, + { + "bbox": [ + 131, + 299, + 481, + 324 + ], + "type": "text", + "content": ". Then, final color for the pixel is computed by using differentiable volume rendering (Eq. 2)" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 131, + 341, + 228, + 354 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 131, + 341, + 228, + 354 + ], + "spans": [ + { + "bbox": [ + 131, + 341, + 228, + 354 + ], + "type": "text", + "content": "3.5 Optimization" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 130, + 361, + 482, + 468 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 361, + 482, + 468 + ], + "spans": [ + { + "bbox": [ + 130, + 361, + 482, + 468 + ], + "type": "text", + "content": "Our model is trained on a collection of monocular RGB videos with photometric and depth losses. For photometric loss " + }, + { + "bbox": [ + 130, + 361, + 482, + 468 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_{color}" + }, + { + "bbox": [ + 130, + 361, + 482, + 468 + ], + "type": "text", + "content": ", we employ " + }, + { + "bbox": [ + 130, + 361, + 482, + 468 + ], + "type": "inline_equation", + "content": "L_{2}" + }, + { + "bbox": [ + 130, + 361, + 482, + 468 + ], + "type": "text", + "content": " distance between the renderings " + }, + { + "bbox": [ + 130, + 361, + 482, + 468 + ], + "type": "inline_equation", + "content": "C" + }, + { + "bbox": [ + 130, + 361, + 482, + 468 + ], + "type": "text", + "content": " and ground-truth images " + }, + { + "bbox": [ + 130, + 361, + 482, + 468 + ], + "type": "inline_equation", + "content": "I" + }, + { + "bbox": [ + 130, + 361, + 482, + 468 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 130, + 361, + 482, + 468 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_{color} = \\sum_{i}\\sum_{r}\\| C_{i}(r) - I_{i}(r)\\|" + }, + { + "bbox": [ + 130, + 361, + 482, + 468 + ], + "type": "text", + "content": ", where " + }, + { + "bbox": [ + 130, + 361, + 482, + 468 + ], + "type": "inline_equation", + "content": "r" + }, + { + "bbox": [ + 130, + 361, + 482, + 468 + ], + "type": "text", + "content": " denotes the camera ray of each pixel and " + }, + { + "bbox": [ + 130, + 361, + 482, + 468 + ], + "type": "inline_equation", + "content": "i" + }, + { + "bbox": [ + 130, + 361, + 482, + 468 + ], + "type": "text", + "content": " denotes frame index. For depth loss " + }, + { + "bbox": [ + 130, + 361, + 482, + 468 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_{depth}" + }, + { + "bbox": [ + 130, + 361, + 482, + 468 + ], + "type": "text", + "content": ", we rasterize the depth of tracking mesh " + }, + { + "bbox": [ + 130, + 361, + 482, + 468 + ], + "type": "inline_equation", + "content": "M_{i}" + }, + { + "bbox": [ + 130, + 361, + 482, + 468 + ], + "type": "text", + "content": " and apply " + }, + { + "bbox": [ + 130, + 361, + 482, + 468 + ], + "type": "inline_equation", + "content": "L_{1}" + }, + { + "bbox": [ + 130, + 361, + 482, + 468 + ], + "type": "text", + "content": " distance between this map " + }, + { + "bbox": [ + 130, + 361, + 482, + 468 + ], + "type": "inline_equation", + "content": "D" + }, + { + "bbox": [ + 130, + 361, + 482, + 468 + ], + "type": "text", + "content": " and the ray termination " + }, + { + "bbox": [ + 130, + 361, + 482, + 468 + ], + "type": "inline_equation", + "content": "H" + }, + { + "bbox": [ + 130, + 361, + 482, + 468 + ], + "type": "text", + "content": " of the volumetric rendering only for face region as FLAME model is restricted to model details like hair: " + }, + { + "bbox": [ + 130, + 361, + 482, + 468 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_{depth} = \\sum_{i}\\sum_{r}\\mathcal{M}_{i}(r)\\| H_{i}(r) - D_{i}(r)\\|" + }, + { + "bbox": [ + 130, + 361, + 482, + 468 + ], + "type": "text", + "content": ", where " + }, + { + "bbox": [ + 130, + 361, + 482, + 468 + ], + "type": "inline_equation", + "content": "\\mathcal{M}" + }, + { + "bbox": [ + 130, + 361, + 482, + 468 + ], + "type": "text", + "content": " denotes face binary mask. The total loss is defined as:" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 239, + 479, + 481, + 491 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 239, + 479, + 481, + 491 + ], + "spans": [ + { + "bbox": [ + 239, + 479, + 481, + 491 + ], + "type": "interline_equation", + "content": "\\mathcal {L} = \\lambda_ {\\text {c o l o r}} \\mathcal {L} _ {\\text {c o l o r}} + \\lambda_ {\\text {d e p t h}} \\mathcal {L} _ {\\text {d e p t h}} \\tag {6}", + "image_path": "def607f7cf480636fccb3031a8d1209f0fa1dae4e323b03fe0321f83bf5563ac.jpg" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 131, + 500, + 365, + 514 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 131, + 500, + 365, + 514 + ], + "spans": [ + { + "bbox": [ + 131, + 500, + 365, + 514 + ], + "type": "text", + "content": "where we empirically set " + }, + { + "bbox": [ + 131, + 500, + 365, + 514 + ], + "type": "inline_equation", + "content": "\\lambda_{color} = 50" + }, + { + "bbox": [ + 131, + 500, + 365, + 514 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 131, + 500, + 365, + 514 + ], + "type": "inline_equation", + "content": "\\lambda_{depth} = 10" + }, + { + "bbox": [ + 131, + 500, + 365, + 514 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 131, + 531, + 230, + 544 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 131, + 531, + 230, + 544 + ], + "spans": [ + { + "bbox": [ + 131, + 531, + 230, + 544 + ], + "type": "text", + "content": "4 Experiments" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 130, + 555, + 481, + 604 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 555, + 481, + 604 + ], + "spans": [ + { + "bbox": [ + 130, + 555, + 481, + 604 + ], + "type": "text", + "content": "In this section, we evaluate the performance of PAV using two different datasets. We perform an ablation study to validate our design choices and a comparison against the state-of-the-art monocular head avatar method in various configurations." + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 131, + 621, + 200, + 632 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 131, + 621, + 200, + 632 + ], + "spans": [ + { + "bbox": [ + 131, + 621, + 200, + 632 + ], + "type": "text", + "content": "4.1 Dataset" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 130, + 641, + 481, + 666 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 641, + 481, + 666 + ], + "spans": [ + { + "bbox": [ + 130, + 641, + 481, + 666 + ], + "type": "text", + "content": "Our method takes a video collection of a subject containing different appearances, poses, and expressions. Since there is no existing dataset to evaluate" + } + ] + } + ], + "index": 14 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "spans": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "type": "text", + "content": "8" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 166, + 91, + 238, + 101 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 166, + 91, + 238, + 101 + ], + "spans": [ + { + "bbox": [ + 166, + 91, + 238, + 101 + ], + "type": "text", + "content": "A. Caliskan et al." + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 7 + }, + { + "para_blocks": [ + { + "bbox": [ + 130, + 116, + 479, + 234 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 116, + 479, + 234 + ], + "spans": [ + { + "bbox": [ + 130, + 116, + 479, + 234 + ], + "type": "text", + "content": "learning 3D head avatar from video collections of the same subjects with various appearances, poses, and expressions, we compiled a new Video Collection (VidCol) dataset from talking face videos of Roger Federer, King Charles and Tom Hanks across years. This dataset consists of in-the-wild videos recorded in uncontrolled environments. In VidCol dataset, we curated videos from 5 different appearances for each subject. We randomly sampled 500 frames at " + }, + { + "bbox": [ + 130, + 116, + 479, + 234 + ], + "type": "inline_equation", + "content": "512^2" + }, + { + "bbox": [ + 130, + 116, + 479, + 234 + ], + "type": "text", + "content": " resolution from each video (or appearance) and we used a total of 2500 frames to train and test a single unified network per subject. We will share this dataset with annotations of semantic masks, tracked 3DMM models, and camera calibrations for further research in this direction." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 130, + 236, + 479, + 260 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 236, + 479, + 260 + ], + "spans": [ + { + "bbox": [ + 130, + 236, + 479, + 260 + ], + "type": "text", + "content": "For our experiments, we evaluate all characters and train a separate PAV model for each subject. We develop two different protocols for evaluation:" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 130, + 270, + 479, + 354 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 270, + 479, + 354 + ], + "spans": [ + { + "bbox": [ + 130, + 270, + 479, + 354 + ], + "type": "text", + "content": "a) Unified Network for Novel Pose/Exp. Synthesis This protocol evaluates the performance in terms of novel appearance, head pose, and facial expression synthesis when training a single unified network from all appearances. For training, we select 400 frames for each appearance of a subject talking in front of a camera with various head poses. It mimics a training video commonly used for monocular neural head avatar methods (Sec. 2). For testing, we use the remaining 100 frames for each appearance with unseen poses and expressions." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 130, + 365, + 480, + 450 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 365, + 480, + 450 + ], + "spans": [ + { + "bbox": [ + 130, + 365, + 480, + 450 + ], + "type": "text", + "content": "b) Separated Networks for Novel Pose/Exp. Synthesis This protocol evaluates the novel appearance, head pose, and facial expression generalization ability of the methods when training a separate network for each appearance. We select 400 training images for each appearance of a subject talking of a camera, same as protocol-a. For testing, we choose the remaining 100 unseen pose/expression combinations with ground truth, same as protocol-a, and another 300 unseen pose/expression sets without ground truth image data." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 131, + 466, + 200, + 477 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 131, + 466, + 200, + 477 + ], + "spans": [ + { + "bbox": [ + 131, + 466, + 200, + 477 + ], + "type": "text", + "content": "4.2 Metrics" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 130, + 485, + 479, + 570 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 485, + 479, + 570 + ], + "spans": [ + { + "bbox": [ + 130, + 485, + 479, + 570 + ], + "type": "text", + "content": "We report several metrics to evaluate the quality of synthesized images as well as the ablation study of our method. For synthesized images with provided ground-truth, we use Learned Perceptual Patch Similarity (LPIPS) [41], Deep Image Structure and Texture Similarity (DISTS) [8], Structural Similarity Index (SSIM) [36], Peak Signal-to-Noise Ratio (PSNR) and Fréchet Inception Distance (FID) [15]. When there is an absence of ground truth, we adopt FID metric for quantitative comparison." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 131, + 586, + 279, + 597 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 131, + 586, + 279, + 597 + ], + "spans": [ + { + "bbox": [ + 131, + 586, + 279, + 597 + ], + "type": "text", + "content": "4.3 Implementation Details" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 130, + 605, + 479, + 665 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 605, + 479, + 665 + ], + "spans": [ + { + "bbox": [ + 130, + 605, + 479, + 665 + ], + "type": "text", + "content": "We optimize Eq. 6 using Adam optimizer [17] with hyperparameters " + }, + { + "bbox": [ + 130, + 605, + 479, + 665 + ], + "type": "inline_equation", + "content": "\\beta_{1} = 0.9" + }, + { + "bbox": [ + 130, + 605, + 479, + 665 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 130, + 605, + 479, + 665 + ], + "type": "inline_equation", + "content": "\\beta_{2} = 0.99" + }, + { + "bbox": [ + 130, + 605, + 479, + 665 + ], + "type": "text", + "content": ". We set the learning rate to " + }, + { + "bbox": [ + 130, + 605, + 479, + 665 + ], + "type": "inline_equation", + "content": "2.5 \\times 10^{-3}" + }, + { + "bbox": [ + 130, + 605, + 479, + 665 + ], + "type": "text", + "content": " for " + }, + { + "bbox": [ + 130, + 605, + 479, + 665 + ], + "type": "inline_equation", + "content": "F_{density}" + }, + { + "bbox": [ + 130, + 605, + 479, + 665 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 130, + 605, + 479, + 665 + ], + "type": "inline_equation", + "content": "F_{color}" + }, + { + "bbox": [ + 130, + 605, + 479, + 665 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 130, + 605, + 479, + 665 + ], + "type": "inline_equation", + "content": "Z" + }, + { + "bbox": [ + 130, + 605, + 479, + 665 + ], + "type": "text", + "content": ", and to " + }, + { + "bbox": [ + 130, + 605, + 479, + 665 + ], + "type": "inline_equation", + "content": "2.5 \\times 10^{-4}" + }, + { + "bbox": [ + 130, + 605, + 479, + 665 + ], + "type": "text", + "content": " for " + }, + { + "bbox": [ + 130, + 605, + 479, + 665 + ], + "type": "inline_equation", + "content": "F_{offset}" + }, + { + "bbox": [ + 130, + 605, + 479, + 665 + ], + "type": "text", + "content": ". We sample 1024 points along each ray for rendering. Multi-resolution hash encoding used in the method is set to 16 number of levels, 18 hash map size, and 2048 finest resolution. The optimization takes 15K iterations" + } + ] + } + ], + "index": 9 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 226, + 91, + 447, + 101 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 226, + 91, + 447, + 101 + ], + "spans": [ + { + "bbox": [ + 226, + 91, + 447, + 101 + ], + "type": "text", + "content": "PAV: Personalized Head Avatar from Video Collection" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 474, + 92, + 480, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 474, + 92, + 480, + 100 + ], + "spans": [ + { + "bbox": [ + 474, + 92, + 480, + 100 + ], + "type": "text", + "content": "9" + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 8 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 143, + 112, + 483, + 361 + ], + "blocks": [ + { + "bbox": [ + 143, + 112, + 483, + 361 + ], + "lines": [ + { + "bbox": [ + 143, + 112, + 483, + 361 + ], + "spans": [ + { + "bbox": [ + 143, + 112, + 483, + 361 + ], + "type": "image", + "image_path": "fb9dbf79da6fd0ec4d6e8efc8a9696e8bd0953409ef827f3544c570740b3e090.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 130, + 369, + 482, + 456 + ], + "lines": [ + { + "bbox": [ + 130, + 369, + 482, + 456 + ], + "spans": [ + { + "bbox": [ + 130, + 369, + 482, + 456 + ], + "type": "text", + "content": "Fig. 3: Qualitative Ablation Study on VidCol Dataset. This figure shows synthesized images and pixel error maps against ground-truth. Please note that brighter pixel colors denote higher pixel errors. Without appearance embedding with latent neural features and appearance-conditioned density offset (a), the model fails to learn appearance-based distinct details. The latent neural features (LNF) (c) resolve the issue of high-fidelity rendering of texture (blue arrow). The appearance-conditioned density offset field (d) allows the model to learn more accurate geometry and texture that corresponds to the face, hair, and neck regions (red arrow)" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_caption" + } + ], + "index": 2 + }, + { + "bbox": [ + 130, + 481, + 480, + 506 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 481, + 480, + 506 + ], + "spans": [ + { + "bbox": [ + 130, + 481, + 480, + 506 + ], + "type": "text", + "content": "to converge when training network with all appearances of single subjects. More details on the network are provided in the supplementary document." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 131, + 525, + 239, + 537 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 131, + 525, + 239, + 537 + ], + "spans": [ + { + "bbox": [ + 131, + 525, + 239, + 537 + ], + "type": "text", + "content": "4.4 Ablation Study" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 130, + 545, + 482, + 666 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 545, + 482, + 666 + ], + "spans": [ + { + "bbox": [ + 130, + 545, + 482, + 666 + ], + "type": "text", + "content": "We evaluate different design choices of PAV in Tab. 1 and Fig. 3. In this study, we use protocol-a of the VidCol dataset for all experiments. We first report the results of the final model which includes latent neural features (LNF) (Sec. 3.3) and density offset " + }, + { + "bbox": [ + 130, + 545, + 482, + 666 + ], + "type": "inline_equation", + "content": "\\varDelta_{\\sigma}" + }, + { + "bbox": [ + 130, + 545, + 482, + 666 + ], + "type": "text", + "content": " (Sec. 3.4). The full model achieves PSNR score of 25.15 and SSIM score of 0.91 for image synthesis. If we remove the density offset from our model while employing appearance-conditioned radiance estimation, PSNR and SSIM scores for image synthesis decrease to 24.14 and 0.90, respectively. This indicates that without appearance-conditioned density offset " + }, + { + "bbox": [ + 130, + 545, + 482, + 666 + ], + "type": "inline_equation", + "content": "\\varDelta_{\\sigma}" + }, + { + "bbox": [ + 130, + 545, + 482, + 666 + ], + "type": "text", + "content": ", density estimation MLP (" + }, + { + "bbox": [ + 130, + 545, + 482, + 666 + ], + "type": "inline_equation", + "content": "\\mathcal{F}_{\\text {density }}" + }, + { + "bbox": [ + 130, + 545, + 482, + 666 + ], + "type": "text", + "content": " in Sec. 3.4) can not learn to synthesize appearance-specific geometry and texture details, thus rendering fails there as well. Even" + } + ] + } + ], + "index": 6 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "spans": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "type": "text", + "content": "10" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 166, + 91, + 238, + 100 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 166, + 91, + 238, + 100 + ], + "spans": [ + { + "bbox": [ + 166, + 91, + 238, + 100 + ], + "type": "text", + "content": "A. Caliskan et al." + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 9 + }, + { + "para_blocks": [ + { + "type": "table", + "bbox": [ + 153, + 114, + 465, + 182 + ], + "blocks": [ + { + "bbox": [ + 153, + 114, + 465, + 182 + ], + "lines": [ + { + "bbox": [ + 153, + 114, + 465, + 182 + ], + "spans": [ + { + "bbox": [ + 153, + 114, + 465, + 182 + ], + "type": "table", + "html": "
PSNR ↑SSIM ↑DISTS ↓LPIPS ↓FID ↓
Insta [44]21.030.8490.1850.169104.28
PAV without LNF24.190.9000.1450.11954.39
PAV without Δσ24.140.9060.1340.10847.71
PAV (Full Model)25.150.9100.1330.10746.67
", + "image_path": "e3d1a0f121851fe7cf617751938e6d1d4f58f36ed68d2122d9e405cc3f9464e2.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "table_body" + } + ], + "index": 2 + }, + { + "bbox": [ + 132, + 183, + 480, + 205 + ], + "lines": [ + { + "bbox": [ + 132, + 183, + 480, + 205 + ], + "spans": [ + { + "bbox": [ + 132, + 183, + 480, + 205 + ], + "type": "text", + "content": "Table 1: Ablation Study: We evaluate the impact of different components of the proposed method. See Fig. 3 for qualitative comparison." + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "text" + }, + { + "bbox": [ + 130, + 240, + 479, + 300 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 240, + 479, + 300 + ], + "spans": [ + { + "bbox": [ + 130, + 240, + 479, + 300 + ], + "type": "text", + "content": "though the model " + }, + { + "bbox": [ + 130, + 240, + 479, + 300 + ], + "type": "inline_equation", + "content": "\\mathrm{w / o}\\varDelta_{\\sigma}" + }, + { + "bbox": [ + 130, + 240, + 479, + 300 + ], + "type": "text", + "content": " synthesise images that look on par with ground truth in terms of texture, it generates severe artifacts which can be seen in pixel error maps. An example of this behavior can also be seen in Fig. 3 " + }, + { + "bbox": [ + 130, + 240, + 479, + 300 + ], + "type": "inline_equation", + "content": "(\\mathrm{w / o}\\varDelta_{\\sigma})" + }, + { + "bbox": [ + 130, + 240, + 479, + 300 + ], + "type": "text", + "content": " Rendering failures on face, head, and neck regions are indicated by red arrow and error map." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 130, + 300, + 480, + 479 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 300, + 480, + 479 + ], + "spans": [ + { + "bbox": [ + 130, + 300, + 480, + 479 + ], + "type": "text", + "content": "Next, we evaluate the impact of latent neural features (LNF) as an appearance embedding on PAV. In this experiment, to condition the canonical MLPs in PAV, inspired by \"Nerf In-the-wild\" [22, 37], we adopt Generative Latent Optimization (GLO) for appearance embedding. Specifically, this approach defines a single real-valued appearance embedding vector " + }, + { + "bbox": [ + 130, + 300, + 480, + 479 + ], + "type": "inline_equation", + "content": "l" + }, + { + "bbox": [ + 130, + 300, + 480, + 479 + ], + "type": "text", + "content": " for each appearance in the dataset. To this end, we employ a pre-trained encoder from in-the-wild face detector RetinaFace [7] for GLO. During training, if we replace " + }, + { + "bbox": [ + 130, + 300, + 480, + 479 + ], + "type": "inline_equation", + "content": "\\mathbf{Z} = \\{Z_{\\mathrm{i}}\\}_{i=1}^{N}" + }, + { + "bbox": [ + 130, + 300, + 480, + 479 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 130, + 300, + 480, + 479 + ], + "type": "inline_equation", + "content": "Z_{i} \\in \\mathbb{R}^{512 \\times 512 \\times 16}" + }, + { + "bbox": [ + 130, + 300, + 480, + 479 + ], + "type": "text", + "content": " with " + }, + { + "bbox": [ + 130, + 300, + 480, + 479 + ], + "type": "inline_equation", + "content": "\\mathbf{L} = \\{l_{\\mathrm{i}}\\}_{i=1}^{N}" + }, + { + "bbox": [ + 130, + 300, + 480, + 479 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 130, + 300, + 480, + 479 + ], + "type": "inline_equation", + "content": "l_{i} \\in \\mathbb{R}^{512 \\times 1}" + }, + { + "bbox": [ + 130, + 300, + 480, + 479 + ], + "type": "text", + "content": " for " + }, + { + "bbox": [ + 130, + 300, + 480, + 479 + ], + "type": "inline_equation", + "content": "N" + }, + { + "bbox": [ + 130, + 300, + 480, + 479 + ], + "type": "text", + "content": " appearances of each subject, and pass it to appearance conditioned NeRF (See Sec. 3.4), PSNR and SSIM scores for image synthesis decrease to 24.19 and 0.90, respectively. This is also evident from the qualitative results shown in Fig. 3 (w/o LNF), indicating that latent neural features help in improved image synthesis quality with their local features attached to 3D geometry. Rendering failures on head, neck, and skin texture are shown by red arrow, blue arrow and error maps in Fig. 3. Finally, as discussed in Sec. 3, we keep the network design as we explain after the ablation study." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 132, + 500, + 221, + 511 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 500, + 221, + 511 + ], + "spans": [ + { + "bbox": [ + 132, + 500, + 221, + 511 + ], + "type": "text", + "content": "4.5 Comparison" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 130, + 521, + 482, + 666 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 521, + 482, + 666 + ], + "spans": [ + { + "bbox": [ + 130, + 521, + 482, + 666 + ], + "type": "text", + "content": "Since our method is the first work that learns dynamic deformable head NeRF from multiple appearances of the same subjects using a unified network, we ourselves design some baseline comparisons in this section. First, we compare our method against Insta [44] using protocol-a (Sec. 4.1) of the VidCol dataset. Both methods are trained using the frames sampled from all appearances of the same subject. For testing, both methods synthesize images for target head pose and expression. The results are summarized in Tab. 2 and Fig. 4. For protocol-a, our method clearly outperforms the baseline for final image synthesis results across all metrics. It should be noted that PAV framework jointly learns all different appearances via one-time training using a unified network and it is able to preserve appearance-specific details for novel poses and expressions. On the other hand, as shown in Fig. 4, Insta can not disentangle appearance from the" + } + ] + } + ], + "index": 7 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 226, + 91, + 447, + 101 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 226, + 91, + 447, + 101 + ], + "spans": [ + { + "bbox": [ + 226, + 91, + 447, + 101 + ], + "type": "text", + "content": "PAV: Personalized Head Avatar from Video Collection" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 470, + 92, + 479, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 470, + 92, + 479, + 100 + ], + "spans": [ + { + "bbox": [ + 470, + 92, + 479, + 100 + ], + "type": "text", + "content": "11" + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 10 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 136, + 113, + 492, + 390 + ], + "blocks": [ + { + "bbox": [ + 136, + 113, + 492, + 390 + ], + "lines": [ + { + "bbox": [ + 136, + 113, + 492, + 390 + ], + "spans": [ + { + "bbox": [ + 136, + 113, + 492, + 390 + ], + "type": "image", + "image_path": "862f9ccf65ef075cc91ec5ff1d498e4c0b5dbe4a2ed0739cf4064be36fd6ad56.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 132, + 397, + 482, + 431 + ], + "lines": [ + { + "bbox": [ + 132, + 397, + 482, + 431 + ], + "spans": [ + { + "bbox": [ + 132, + 397, + 482, + 431 + ], + "type": "text", + "content": "Fig. 4: Qualitative comparison of our method (PAV) against baseline method (Insta [44]. Columns " + }, + { + "bbox": [ + 132, + 397, + 482, + 431 + ], + "type": "inline_equation", + "content": "\\{2,6\\}" + }, + { + "bbox": [ + 132, + 397, + 482, + 431 + ], + "type": "text", + "content": " show target head poses and facial expressions for synthesized images at columns " + }, + { + "bbox": [ + 132, + 397, + 482, + 431 + ], + "type": "inline_equation", + "content": "\\{3,4,7,8\\}" + }, + { + "bbox": [ + 132, + 397, + 482, + 431 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_caption" + } + ], + "index": 2 + }, + { + "bbox": [ + 130, + 459, + 481, + 495 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 459, + 481, + 495 + ], + "spans": [ + { + "bbox": [ + 130, + 459, + 481, + 495 + ], + "type": "text", + "content": "neural radiance field, which results in a mixture of appearances in synthesized images and inaccurate geometry predictions. This is the limitation of Insta we are addressing in the paper." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 130, + 498, + 482, + 666 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 498, + 482, + 666 + ], + "spans": [ + { + "bbox": [ + 130, + 498, + 482, + 666 + ], + "type": "text", + "content": "Second, we train separate Insta [44] models for each appearance of the same subject using protocol-b while training a single unified model for our method using the frames sampled from all appearances of the same subject. For testing, we use the same test data from protocol-b for both PAV and Insta [44] and synthesize images for novel poses and expressions. Note that, we added more novel expressions and poses in protocol-b (Sec. 4.1) to test the generalization of our method under a unified training schema, which results in increasing FID score for both methods in Tab. 3. Even though the results show that artifacts (in Fig. 4) of the Insta renders are decreased in Fig. 5, our method demonstrates better-rendering results (Fig. 5) and lower FID score (Tab. 3). This indicates that PAV's single unified model helps in learning generic neural radiance field representation across all appearances of the same subject while preserving appearance-based details. Please note that we are demonstrating Insta's performance under a challenging pose/expression in Fig. 5 which can also be seen in" + } + ] + } + ], + "index": 5 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "spans": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "type": "text", + "content": "12" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 166, + 91, + 238, + 101 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 166, + 91, + 238, + 101 + ], + "spans": [ + { + "bbox": [ + 166, + 91, + 238, + 101 + ], + "type": "text", + "content": "A. Caliskan et al." + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 11 + }, + { + "para_blocks": [ + { + "type": "table", + "bbox": [ + 134, + 115, + 482, + 169 + ], + "blocks": [ + { + "bbox": [ + 134, + 115, + 482, + 169 + ], + "lines": [ + { + "bbox": [ + 134, + 115, + 482, + 169 + ], + "spans": [ + { + "bbox": [ + 134, + 115, + 482, + 169 + ], + "type": "table", + "html": "
MethodVideo Collection 1Video Collection 2
PSNR ↑SSIM ↑DISTS ↓LPIPS ↓FID ↓PSNR ↑SSIM ↑DISTS ↓LPIPS ↓FID ↓
Protocol (a): Novel Head Pose and Expression Synthesis
Insta [44]21.030.840.1850.169104.2823.670.860.1870.16398.87
PAV (Ours)25.150.910.1330.10746.6727.090.890.1440.12053.91
", + "image_path": "60e1dae0bc93fd572ccbadc1d3b35012533f73f2df59aae16b9fe6acf9f1d9ae.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "table_body" + } + ], + "index": 2 + }, + { + "bbox": [ + 132, + 170, + 479, + 201 + ], + "lines": [ + { + "bbox": [ + 132, + 170, + 479, + 201 + ], + "spans": [ + { + "bbox": [ + 132, + 170, + 479, + 201 + ], + "type": "text", + "content": "Table 2: Comparison of PAV against Insta [44]. See Fig. 4 for qualitative comparison. Video Collection 1 and 2 refer to Federer and King Charles splits of VidCol dataset, respectively." + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "text" + }, + { + "bbox": [ + 130, + 236, + 480, + 319 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 236, + 480, + 319 + ], + "spans": [ + { + "bbox": [ + 130, + 236, + 480, + 319 + ], + "type": "text", + "content": "the original paper results under extreme poses/expressions. In Fig. 5, the reference image is for photoconsistency check only since there is no ground truth image for that pose/expression. Finally, we provide additional qualitative results of avatars animated in the supplementary video. We also compared PAV against Insta for model efficiency using protocol-b dataset and recorded values in Tab. 3. Although we train our model with multiple appearances jointly, PAV is more efficient than Insta in terms of memory footprint and training time" + } + ] + } + ], + "index": 4 + }, + { + "type": "image", + "bbox": [ + 185, + 338, + 247, + 464 + ], + "blocks": [ + { + "bbox": [ + 185, + 338, + 247, + 464 + ], + "lines": [ + { + "bbox": [ + 185, + 338, + 247, + 464 + ], + "spans": [ + { + "bbox": [ + 185, + 338, + 247, + 464 + ], + "type": "image", + "image_path": "06e6d4e9ac32ff115f6293a48eda57c3754b4c7e80c1df691cc1c39f984f2b04.jpg" + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 201, + 465, + 230, + 472 + ], + "lines": [ + { + "bbox": [ + 201, + 465, + 230, + 472 + ], + "spans": [ + { + "bbox": [ + 201, + 465, + 230, + 472 + ], + "type": "text", + "content": "Reference" + } + ] + } + ], + "index": 6, + "angle": 0, + "type": "image_caption" + } + ], + "index": 5 + }, + { + "type": "image", + "bbox": [ + 253, + 345, + 301, + 402 + ], + "blocks": [ + { + "bbox": [ + 253, + 345, + 301, + 402 + ], + "lines": [ + { + "bbox": [ + 253, + 345, + 301, + 402 + ], + "spans": [ + { + "bbox": [ + 253, + 345, + 301, + 402 + ], + "type": "image", + "image_path": "785981cd406aa46df4b499f550ab7df46574d32d2a2a328b131545f1839849bb.jpg" + } + ] + } + ], + "index": 7, + "angle": 0, + "type": "image_body" + } + ], + "index": 7 + }, + { + "type": "image", + "bbox": [ + 258, + 405, + 302, + 463 + ], + "blocks": [ + { + "bbox": [ + 258, + 405, + 302, + 463 + ], + "lines": [ + { + "bbox": [ + 258, + 405, + 302, + 463 + ], + "spans": [ + { + "bbox": [ + 258, + 405, + 302, + 463 + ], + "type": "image", + "image_path": "e7b2a0970d0ea6029004843b50749a6242a6eb6f832be788d1c488f5bc0c3603.jpg" + } + ] + } + ], + "index": 8, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 251, + 465, + 301, + 472 + ], + "lines": [ + { + "bbox": [ + 251, + 465, + 301, + 472 + ], + "spans": [ + { + "bbox": [ + 251, + 465, + 301, + 472 + ], + "type": "text", + "content": "Target Pose/Exp." + } + ] + } + ], + "index": 9, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 131, + 482, + 480, + 516 + ], + "lines": [ + { + "bbox": [ + 131, + 482, + 480, + 516 + ], + "spans": [ + { + "bbox": [ + 131, + 482, + 480, + 516 + ], + "type": "text", + "content": "Fig. 5: Qualitative comparison of PAV against single-appearance traines Insta [44]. Please note that the reference image is for photoconsistency check only since there is no ground truth image for that pose/expression" + } + ] + } + ], + "index": 16, + "angle": 0, + "type": "image_caption" + } + ], + "index": 8 + }, + { + "type": "image", + "bbox": [ + 315, + 340, + 358, + 401 + ], + "blocks": [ + { + "bbox": [ + 315, + 340, + 358, + 401 + ], + "lines": [ + { + "bbox": [ + 315, + 340, + 358, + 401 + ], + "spans": [ + { + "bbox": [ + 315, + 340, + 358, + 401 + ], + "type": "image", + "image_path": "3bbf67cab567bb2963ea47f86303776b311718312d4ff11c2d718ec5bf22872b.jpg" + } + ] + } + ], + "index": 10, + "angle": 0, + "type": "image_body" + } + ], + "index": 10 + }, + { + "type": "image", + "bbox": [ + 315, + 402, + 359, + 463 + ], + "blocks": [ + { + "bbox": [ + 315, + 402, + 359, + 463 + ], + "lines": [ + { + "bbox": [ + 315, + 402, + 359, + 463 + ], + "spans": [ + { + "bbox": [ + 315, + 402, + 359, + 463 + ], + "type": "image", + "image_path": "6cefb32e6d820a2f9a30c6c8c622afccf81ce2c5b1b1b9c762d8d4a5cb7c6d02.jpg" + } + ] + } + ], + "index": 11, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 323, + 465, + 350, + 473 + ], + "lines": [ + { + "bbox": [ + 323, + 465, + 350, + 473 + ], + "spans": [ + { + "bbox": [ + 323, + 465, + 350, + 473 + ], + "type": "text", + "content": "Insta [44]" + } + ] + } + ], + "index": 12, + "angle": 0, + "type": "image_caption" + } + ], + "index": 11 + }, + { + "type": "image", + "bbox": [ + 376, + 341, + 416, + 400 + ], + "blocks": [ + { + "bbox": [ + 376, + 341, + 416, + 400 + ], + "lines": [ + { + "bbox": [ + 376, + 341, + 416, + 400 + ], + "spans": [ + { + "bbox": [ + 376, + 341, + 416, + 400 + ], + "type": "image", + "image_path": "45aee65b3f48520f21bf405d7f9cd85f8772b2c46159bfe22a07bd61d3e665c9.jpg" + } + ] + } + ], + "index": 13, + "angle": 0, + "type": "image_body" + } + ], + "index": 13 + }, + { + "type": "image", + "bbox": [ + 378, + 402, + 416, + 463 + ], + "blocks": [ + { + "bbox": [ + 378, + 402, + 416, + 463 + ], + "lines": [ + { + "bbox": [ + 378, + 402, + 416, + 463 + ], + "spans": [ + { + "bbox": [ + 378, + 402, + 416, + 463 + ], + "type": "image", + "image_path": "ccb511cdbc55f7c76327c5111109e121b8472b176573cbe83398ed8bcecb894c.jpg" + } + ] + } + ], + "index": 14, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 391, + 465, + 405, + 472 + ], + "lines": [ + { + "bbox": [ + 391, + 465, + 405, + 472 + ], + "spans": [ + { + "bbox": [ + 391, + 465, + 405, + 472 + ], + "type": "text", + "content": "PAV" + } + ] + } + ], + "index": 15, + "angle": 0, + "type": "image_caption" + } + ], + "index": 14 + }, + { + "bbox": [ + 132, + 556, + 217, + 569 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 556, + 217, + 569 + ], + "spans": [ + { + "bbox": [ + 132, + 556, + 217, + 569 + ], + "type": "text", + "content": "5 Discussion" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 130, + 581, + 482, + 666 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 581, + 482, + 666 + ], + "spans": [ + { + "bbox": [ + 130, + 581, + 482, + 666 + ], + "type": "text", + "content": "We presented PAV which is a novel method for learning personalized head avatars from monocular and in-the-wild video collections of a person. We demonstrated that PAV can model a 3D head avatar from an unstructured video collection of a subject in a single unified network while disentangling face appearances from the radiance field. PAV can be trained simultaneously for multiple appearances of the same subject, allowing novel pose and expression generalization and leading to a superior rendering quality of our method compared to the baseline." + } + ] + } + ], + "index": 18 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 226, + 91, + 447, + 101 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 226, + 91, + 447, + 101 + ], + "spans": [ + { + "bbox": [ + 226, + 91, + 447, + 101 + ], + "type": "text", + "content": "PAV: Personalized Head Avatar from Video Collection" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 470, + 92, + 480, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 470, + 92, + 480, + 100 + ], + "spans": [ + { + "bbox": [ + 470, + 92, + 480, + 100 + ], + "type": "text", + "content": "13" + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 12 + }, + { + "para_blocks": [ + { + "type": "table", + "bbox": [ + 135, + 114, + 482, + 173 + ], + "blocks": [ + { + "bbox": [ + 135, + 114, + 482, + 173 + ], + "lines": [ + { + "bbox": [ + 135, + 114, + 482, + 173 + ], + "spans": [ + { + "bbox": [ + 135, + 114, + 482, + 173 + ], + "type": "table", + "html": "
Train Time MinutesNumber of Param. MillionModel Size (MB)FID ↓
Multiple Insta [44]13579.71932109.22
PAV (Ours)4526.865091.01
", + "image_path": "ea68bdeb1ba7a7a738820dc72b53e940306cf5081ab4370c59e5f66756b37cba.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "table_body" + } + ], + "index": 2 + }, + { + "bbox": [ + 132, + 175, + 482, + 218 + ], + "lines": [ + { + "bbox": [ + 132, + 175, + 482, + 218 + ], + "spans": [ + { + "bbox": [ + 132, + 175, + 482, + 218 + ], + "type": "text", + "content": "Table 3: Efficiency and performance comparison between PAV and Insta [44] trained for each appearance (Single Appearance Training) in dataset collection 1. FID score is computed from synthetized images under novel head poses and expressions where there are no existing ground-truth images." + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "text" + }, + { + "type": "image", + "bbox": [ + 214, + 239, + 399, + 371 + ], + "blocks": [ + { + "bbox": [ + 214, + 239, + 399, + 371 + ], + "lines": [ + { + "bbox": [ + 214, + 239, + 399, + 371 + ], + "spans": [ + { + "bbox": [ + 214, + 239, + 399, + 371 + ], + "type": "image", + "image_path": "631be0a96611e488f6eb28929bbf6b526136f5c25d8411e93327cabe30d5e28e.jpg" + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 131, + 379, + 482, + 424 + ], + "lines": [ + { + "bbox": [ + 131, + 379, + 482, + 424 + ], + "spans": [ + { + "bbox": [ + 131, + 379, + 482, + 424 + ], + "type": "text", + "content": "Fig. 6: Multi-identity and multi-appearance extension of PAV. The learned appearance (middle) rendered in a new viewpoint (left) does not match the visual quality of input images (right). Please note that the reference image is for photoconsistency check only since there is no ground truth image for that pose/expression." + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "image_caption" + } + ], + "index": 4 + }, + { + "bbox": [ + 130, + 448, + 482, + 569 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 448, + 482, + 569 + ], + "spans": [ + { + "bbox": [ + 130, + 448, + 482, + 569 + ], + "type": "text", + "content": "Limitation for Multi-Identity. We extended our method for multiple identities and multiple appearances for each identity. To this end, we train a single unified PAV model using frames smaped from all appearances of Federer and King subjects from VidCol dataset with shared canonical space. We evaluate novel head pose and facial expression rendering of this model and Fig. 6 illustrates the results. Even though this model produces reasonable results, synthesized images demonstrate a lack of subject-specific shape and appearance details with rendering artifacts. In the paper, we focus on multiple appearances of a single subject. Scaling up this model to a wide range of subjects and exploring generalized dynamic deformable NeRF would be interesting future work." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 130, + 581, + 482, + 666 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 581, + 482, + 666 + ], + "spans": [ + { + "bbox": [ + 130, + 581, + 482, + 666 + ], + "type": "text", + "content": "Ethical Considerations. Although the rapid progress on digital human avatars facilitates numerous downstream applications, it also raises concerns about ethical misuse. On the synthesis side, it would be ideal to actively use watermarking technologies and avoid driving avatars with different identities. In this work, we aim to faithfully reproduce images of a person with the capability of rendering unseen poses and expressions and switching appearance within their own set of appearances. The work does not intend to create animations that didn't happen." + } + ] + } + ], + "index": 7 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "spans": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "type": "text", + "content": "14" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 166, + 91, + 238, + 101 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 166, + 91, + 238, + 101 + ], + "spans": [ + { + "bbox": [ + 166, + 91, + 238, + 101 + ], + "type": "text", + "content": "A. Caliskan et al." + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 13 + }, + { + "para_blocks": [ + { + "bbox": [ + 132, + 114, + 198, + 127 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 114, + 198, + 127 + ], + "spans": [ + { + "bbox": [ + 132, + 114, + 198, + 127 + ], + "type": "text", + "content": "References" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 138, + 139, + 481, + 666 + ], + "type": "list", + "angle": 0, + "index": 19, + "blocks": [ + { + "bbox": [ + 138, + 139, + 481, + 172 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 139, + 481, + 172 + ], + "spans": [ + { + "bbox": [ + 138, + 139, + 481, + 172 + ], + "type": "text", + "content": "1. Athar, S., Xu, Z., Sunkavalli, K., Shechtman, E., Shu, Z.: Rignerf: Fully controllable neural 3d portraits. In: Proceedings of the IEEE/CVF conference on Computer Vision and Pattern Recognition. pp. 20364-20373 (2022)" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 138, + 173, + 481, + 195 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 173, + 481, + 195 + ], + "spans": [ + { + "bbox": [ + 138, + 173, + 481, + 195 + ], + "type": "text", + "content": "2. Averbuch-Elor, H., Cohen-Or, D., Kopf, J., Cohen, M.F.: Bringing portraits to life. ACM transactions on graphics (TOG) 36(6), 1-13 (2017)" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 138, + 195, + 481, + 239 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 195, + 481, + 239 + ], + "spans": [ + { + "bbox": [ + 138, + 195, + 481, + 239 + ], + "type": "text", + "content": "3. Bai, Z., Tan, F., Huang, Z., Sarkar, K., Tang, D., Qiu, D., Meka, A., Du, R., Dou, M., Orts-Escalano, S., et al.: Learning personalized high quality volumetric head avatars from monocular rgb videos. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 16890-16900 (2023)" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 138, + 239, + 481, + 270 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 239, + 481, + 270 + ], + "spans": [ + { + "bbox": [ + 138, + 239, + 481, + 270 + ], + "type": "text", + "content": "4. Bühler, M.C., Meka, A., Li, G., Beeler, T., Hilliges, O.: Varitex: Variational neural face textures. In: Proceedings of the IEEE/CVF International Conference on Computer Vision. pp. 13890-13899 (2021)" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 138, + 271, + 481, + 304 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 271, + 481, + 304 + ], + "spans": [ + { + "bbox": [ + 138, + 271, + 481, + 304 + ], + "type": "text", + "content": "5. Chen, C., O'Toole, M., Bharaj, G., Garrido, P.: Implicit neural head synthesis via controllable local deformation fields. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 416-426 (2023)" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 138, + 304, + 481, + 326 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 304, + 481, + 326 + ], + "spans": [ + { + "bbox": [ + 138, + 304, + 481, + 326 + ], + "type": "text", + "content": "6. Clark, J.H.: Hierarchical geometric models for visible surface algorithms. Communications of the ACM 19(10), 547-554 (1976)" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 138, + 326, + 481, + 347 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 326, + 481, + 347 + ], + "spans": [ + { + "bbox": [ + 138, + 326, + 481, + 347 + ], + "type": "text", + "content": "7. Deng, J., Guo, J., Zhou, Y., Yu, J., Kotsia, I., Zafeiriou, S.: Retinaface: Single-stage dense face localisation in the wild. arXiv preprint arXiv:1905.00641 (2019)" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 138, + 348, + 481, + 369 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 348, + 481, + 369 + ], + "spans": [ + { + "bbox": [ + 138, + 348, + 481, + 369 + ], + "type": "text", + "content": "8. Ding, K., Ma, K., Wang, S., Simoncelli, E.P.: Image Quality Assessment: Unifying Structure and Texture Similarity. TPAMI (2022)" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 138, + 370, + 481, + 413 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 370, + 481, + 413 + ], + "spans": [ + { + "bbox": [ + 138, + 370, + 481, + 413 + ], + "type": "text", + "content": "9. Egger, B., Smith, W.A., Tewari, A., Wuhrer, S., Zollhoefer, M., Beeler, T., Bernard, F., Bolkart, T., Kortylewski, A., Romdhani, S., et al.: 3d morphable face models—past, present, and future. ACM Transactions on Graphics (ToG) 39(5), 1-38 (2020)" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 138, + 413, + 481, + 457 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 413, + 481, + 457 + ], + "spans": [ + { + "bbox": [ + 138, + 413, + 481, + 457 + ], + "type": "text", + "content": "0. Gafni, G., Thies, J., Zollhöfer, M., Nießner, M.: Dynamic neural radiance fields for monocular 4d facial avatar reconstruction. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR). pp. 8649-8658 (June 2021)" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 138, + 457, + 481, + 490 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 457, + 481, + 490 + ], + "spans": [ + { + "bbox": [ + 138, + 457, + 481, + 490 + ], + "type": "text", + "content": "1. Gao, X., Zhong, C., Xiang, J., Hong, Y., Guo, Y., Zhang, J.: Reconstructing personalized semantic facial nerf models from monocular video. ACM Transactions on Graphics (TOG) 41(6), 1-12 (2022)" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 138, + 491, + 481, + 522 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 491, + 481, + 522 + ], + "spans": [ + { + "bbox": [ + 138, + 491, + 481, + 522 + ], + "type": "text", + "content": "2. Garrido, P., Valgaerts, L., Rehmsen, O., Thormahlen, T., Perez, P., Theobalt, C.: Automatic face reenactment. In: Proceedings of the IEEE conference on computer vision and pattern recognition. pp. 4217-4224 (2014)" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 138, + 523, + 481, + 567 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 523, + 481, + 567 + ], + "spans": [ + { + "bbox": [ + 138, + 523, + 481, + 567 + ], + "type": "text", + "content": "3. Garrido, P., Valgaerts, L., Sarmadi, H., Steiner, I., Varanasi, K., Perez, P., Theobalt, C.: Vdub: Modifying face video of actors for plausible visual alignment to a dubbed audio track. In: Computer graphics forum. vol. 34, pp. 193-204. Wiley Online Library (2015)" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 138, + 567, + 481, + 600 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 567, + 481, + 600 + ], + "spans": [ + { + "bbox": [ + 138, + 567, + 481, + 600 + ], + "type": "text", + "content": "4. Grassal, P.W., Prinzler, M., Leistner, T., Rother, C., Nießner, M., Thies, J.: Neural head avatars from monocular rgb videos. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 18653-18664 (2022)" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 138, + 601, + 481, + 632 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 601, + 481, + 632 + ], + "spans": [ + { + "bbox": [ + 138, + 601, + 481, + 632 + ], + "type": "text", + "content": "5. Heusel, M., Ramsauer, H., Unterthiner, T., Nessler, B., Hochreiter, S.: Gans trained by a two time-scale update rule converge to a local nash equilibrium. Advances in neural information processing systems 30 (2017)" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 138, + 632, + 481, + 666 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 632, + 481, + 666 + ], + "spans": [ + { + "bbox": [ + 138, + 632, + 481, + 666 + ], + "type": "text", + "content": "6. Kim, H., Garrido, P., Tewari, A., Xu, W., Thies, J., Niessner, M., Pérez, P., Richardt, C., Zollhöfer, M., Theobalt, C.: Deep video portraits. ACM transactions on graphics (TOG) 37(4), 1-14 (2018)" + } + ] + } + ], + "index": 18 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 225, + 91, + 447, + 101 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 225, + 91, + 447, + 101 + ], + "spans": [ + { + "bbox": [ + 225, + 91, + 447, + 101 + ], + "type": "text", + "content": "PAV: Personalized Head Avatar from Video Collection" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 470, + 92, + 481, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 470, + 92, + 481, + 100 + ], + "spans": [ + { + "bbox": [ + 470, + 92, + 481, + 100 + ], + "type": "text", + "content": "15" + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 14 + }, + { + "para_blocks": [ + { + "bbox": [ + 132, + 117, + 480, + 665 + ], + "type": "list", + "angle": 0, + "index": 18, + "blocks": [ + { + "bbox": [ + 133, + 117, + 480, + 138 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 117, + 480, + 138 + ], + "spans": [ + { + "bbox": [ + 133, + 117, + 480, + 138 + ], + "type": "text", + "content": "17. Kingma, D.P., Ba, J.: Adam: A method for stochastic optimization. arXiv preprint arXiv:1412.6980 (2014)" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 133, + 139, + 480, + 171 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 139, + 480, + 171 + ], + "spans": [ + { + "bbox": [ + 133, + 139, + 480, + 171 + ], + "type": "text", + "content": "18. Kuang, Z., Olszewski, K., Chai, M., Huang, Z., Achlioptas, P., Tulyakov, S.: Neroic: Neural rendering of objects from online image collections. ACM Transactions on Graphics (TOG) 41(4), 1-12 (2022)" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 133, + 172, + 480, + 205 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 172, + 480, + 205 + ], + "spans": [ + { + "bbox": [ + 133, + 172, + 480, + 205 + ], + "type": "text", + "content": "19. Li, T., Bolkart, T., Black, M.J., Li, H., Romero, J.: Learning a model of facial shape and expression from 4D scans. ACM Transactions on Graphics, (Proc. SIGGRAPH Asia) 36(6), 194:1-194:17 (2017), https://doi.org/10.1145/3130800.3130813" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 132, + 205, + 480, + 237 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 205, + 480, + 237 + ], + "spans": [ + { + "bbox": [ + 132, + 205, + 480, + 237 + ], + "type": "text", + "content": "20. Liu, L., Habermann, M., Rudnev, V., Sarkar, K., Gu, J., Theobalt, C.: Neural actor: Neural free-view synthesis of human actors with pose control. ACM transactions on graphics (TOG) 40(6), 1-16 (2021)" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 132, + 238, + 480, + 258 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 238, + 480, + 258 + ], + "spans": [ + { + "bbox": [ + 132, + 238, + 480, + 258 + ], + "type": "text", + "content": "21. Mallya, A., Wang, T.C., Liu, M.Y.: Implicit warping for animation with image sets. Advances in Neural Information Processing Systems 35, 22438-22450 (2022)" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 132, + 259, + 480, + 303 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 259, + 480, + 303 + ], + "spans": [ + { + "bbox": [ + 132, + 259, + 480, + 303 + ], + "type": "text", + "content": "22. Martin-Brualla, R., Radwan, N., Sajjadi, M.S., Barron, J.T., Dosovitskiy, A., Duckworth, D.: Nerf in the wild: Neural radiance fields for unconstrained photo collections. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 7210-7219 (2021)" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 132, + 304, + 480, + 335 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 304, + 480, + 335 + ], + "spans": [ + { + "bbox": [ + 132, + 304, + 480, + 335 + ], + "type": "text", + "content": "23. Mildenhall, B., Srinivasan, P.P., Tancik, M., Barron, J.T., Ramamoorthi, R., Ng, R.: Nerf: Representing scenes as neural radiance fields for view synthesis. Communications of the ACM 65(1), 99-106 (2021)" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 132, + 336, + 480, + 369 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 336, + 480, + 369 + ], + "spans": [ + { + "bbox": [ + 132, + 336, + 480, + 369 + ], + "type": "text", + "content": "24. Müller, T., Evans, A., Schied, C., Keller, A.: Instant neural graphics primitives with a multiresolution hash encoding. ACM Transactions on Graphics (ToG) 41(4), 1-15 (2022)" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 132, + 369, + 480, + 402 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 369, + 480, + 402 + ], + "spans": [ + { + "bbox": [ + 132, + 369, + 480, + 402 + ], + "type": "text", + "content": "25. Nirkin, Y., Keller, Y., Hassner, T.: Fsgan: Subject agnostic face swapping and reenactment. In: Proceedings of the IEEE/CVF international conference on computer vision. pp. 7184-7193 (2019)" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 132, + 403, + 480, + 435 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 403, + 480, + 435 + ], + "spans": [ + { + "bbox": [ + 132, + 403, + 480, + 435 + ], + "type": "text", + "content": "26. Park, K., Sinha, U., Hedman, P., Barron, J.T., Bouaziz, S., Goldman, D.B., Martin-Brualla, R., Seitz, S.M.: Hypernerf: A higher-dimensional representation for topologically varying neural radiance fields. arXiv preprint arXiv:2106.13228 (2021)" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 132, + 436, + 480, + 468 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 436, + 480, + 468 + ], + "spans": [ + { + "bbox": [ + 132, + 436, + 480, + 468 + ], + "type": "text", + "content": "27. Peng, S., Dong, J., Wang, Q., Zhang, S., Shuai, Q., Zhou, X., Bao, H.: Animatable neural radiance fields for modeling dynamic human bodies. In: Proceedings of the IEEE/CVF International Conference on Computer Vision. pp. 14314-14323 (2021)" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 132, + 468, + 480, + 501 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 468, + 480, + 501 + ], + "spans": [ + { + "bbox": [ + 132, + 468, + 480, + 501 + ], + "type": "text", + "content": "28. Pumarola, A., Corona, E., Pons-Moll, G., Moreno-Noguer, F.: D-nerf: Neural radiance fields for dynamic scenes. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 10318-10327 (2021)" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 132, + 502, + 480, + 533 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 502, + 480, + 533 + ], + "spans": [ + { + "bbox": [ + 132, + 502, + 480, + 533 + ], + "type": "text", + "content": "29. Siarohin, A., Lathuilière, S., Tulyakov, S., Ricci, E., Sebe, N.: First order motion model for image animation. Advances in neural information processing systems 32 (2019)" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 132, + 534, + 480, + 577 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 534, + 480, + 577 + ], + "spans": [ + { + "bbox": [ + 132, + 534, + 480, + 577 + ], + "type": "text", + "content": "30. Tancik, M., Casser, V., Yan, X., Pradhan, S., Mildenhall, B., Srinivasan, P.P., Barron, J.T., Kretzschmar, H.: Block-nerf: Scalable large scene neural view synthesis. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 8248-8258 (2022)" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 132, + 577, + 480, + 621 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 577, + 480, + 621 + ], + "spans": [ + { + "bbox": [ + 132, + 577, + 480, + 621 + ], + "type": "text", + "content": "31. Tewari, A., Fried, O., Thies, J., Sitzmann, V., Lombardi, S., Sunkavalli, K., MartinBrualla, R., Simon, T., Saragih, J., Nießner, M., et al.: State of the art on neural rendering. In: Computer Graphics Forum. vol. 39, pp. 701-727. Wiley Online Library (2020)" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 132, + 622, + 480, + 665 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 622, + 480, + 665 + ], + "spans": [ + { + "bbox": [ + 132, + 622, + 480, + 665 + ], + "type": "text", + "content": "32. Tewari, A., Thies, J., Mildenhall, B., Srinivasan, P., Tretschk, E., Yifan, W., Lassner, C., Sitzmann, V., Martin-Brualla, R., Lombardi, S., et al.: Advances in neural rendering. In: Computer Graphics Forum. vol. 41, pp. 703-735. Wiley Online Library (2022)" + } + ] + } + ], + "index": 17 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "spans": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "type": "text", + "content": "16" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 166, + 91, + 238, + 100 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 166, + 91, + 238, + 100 + ], + "spans": [ + { + "bbox": [ + 166, + 91, + 238, + 100 + ], + "type": "text", + "content": "A. Caliskan et al." + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 15 + }, + { + "para_blocks": [ + { + "bbox": [ + 130, + 116, + 482, + 534 + ], + "type": "list", + "angle": 0, + "index": 14, + "blocks": [ + { + "bbox": [ + 130, + 116, + 482, + 138 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 116, + 482, + 138 + ], + "spans": [ + { + "bbox": [ + 130, + 116, + 482, + 138 + ], + "type": "text", + "content": "33. Thies, J., Zollhöfer, M., Nießner, M.: Deferred neural rendering: Image synthesis using neural textures. Acm Transactions on Graphics (TOG) 38(4), 1-12 (2019)" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 130, + 140, + 482, + 171 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 140, + 482, + 171 + ], + "spans": [ + { + "bbox": [ + 130, + 140, + 482, + 171 + ], + "type": "text", + "content": "34. Thies, J., Zollhofer, M., Stamminger, M., Theobalt, C., Nieflner, M.: Face2face: Real-time face capture and reenactment of rgb videos. In: Proceedings of the IEEE conference on computer vision and pattern recognition. pp. 2387-2395 (2016)" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 132, + 172, + 481, + 215 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 172, + 481, + 215 + ], + "spans": [ + { + "bbox": [ + 132, + 172, + 481, + 215 + ], + "type": "text", + "content": "35. Wang, T.C., Liu, M.Y., Zhu, J.Y., Tao, A., Kautz, J., Catanzaro, B.: High-resolution image synthesis and semantic manipulation with conditional gans. In: Proceedings of the IEEE conference on computer vision and pattern recognition. pp. 8798-8807 (2018)" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 132, + 216, + 481, + 237 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 216, + 481, + 237 + ], + "spans": [ + { + "bbox": [ + 132, + 216, + 481, + 237 + ], + "type": "text", + "content": "36. Wang, Z., Bovik, A.C., Sheikh, H.R., Simoncelli, E.P.: Image quality assessment: From error visibility to structural similarity. TIP (2004)" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 132, + 238, + 481, + 281 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 238, + 481, + 281 + ], + "spans": [ + { + "bbox": [ + 132, + 238, + 481, + 281 + ], + "type": "text", + "content": "37. Weng, C.Y., Srinivasan, P.P., Curless, B., Kemelmacher-Shlizerman, I.: Person-nerf: Personalized reconstruction from photo collections. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 524-533 (2023)" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 132, + 282, + 481, + 325 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 282, + 481, + 325 + ], + "spans": [ + { + "bbox": [ + 132, + 282, + 481, + 325 + ], + "type": "text", + "content": "38. Yang, B., Bao, C., Zeng, J., Bao, H., Zhang, Y., Cui, Z., Zhang, G.: Neumesh: Learning disentangled neural mesh-based implicit field for geometry and texture editing. In: European Conference on Computer Vision. pp. 597-614. Springer (2022)" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 132, + 326, + 481, + 369 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 326, + 481, + 369 + ], + "spans": [ + { + "bbox": [ + 132, + 326, + 481, + 369 + ], + "type": "text", + "content": "39. Zakharov, E., Ivakhnenko, A., Shysheya, A., Lempitsky, V.: Fast bi-layer neural synthesis of one-shot realistic head avatars. In: Computer Vision-ECCV 2020: 16th European Conference, Glasgow, UK, August 23-28, 2020, Proceedings, Part XII 16. pp. 524-540. Springer (2020)" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 132, + 370, + 481, + 413 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 370, + 481, + 413 + ], + "spans": [ + { + "bbox": [ + 132, + 370, + 481, + 413 + ], + "type": "text", + "content": "40. Zhang, B., Qi, C., Zhang, P., Zhang, B., Wu, H., Chen, D., Chen, Q., Wang, Y., Wen, F.: Metaportrait: Identity-preserving talking head generation with fast personalized adaptation. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR). pp. 22096-22105 (June 2023)" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 132, + 414, + 481, + 435 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 414, + 481, + 435 + ], + "spans": [ + { + "bbox": [ + 132, + 414, + 481, + 435 + ], + "type": "text", + "content": "41. Zhang, R., Isola, P., Efros, A.A., Shechtman, E., Wang, O.: The unreasonable effectiveness of deep features as a perceptual metric. In: CVPR (2018)" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 132, + 436, + 481, + 479 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 436, + 481, + 479 + ], + "spans": [ + { + "bbox": [ + 132, + 436, + 481, + 479 + ], + "type": "text", + "content": "42. Zheng, Y., Abrevaya, V.F., Bühler, M.C., Chen, X., Black, M.J., Hilliges, O.: Im avatar: Implicit morphable head avatars from videos. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 13545-13555 (2022)" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 132, + 479, + 481, + 501 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 479, + 481, + 501 + ], + "spans": [ + { + "bbox": [ + 132, + 479, + 481, + 501 + ], + "type": "text", + "content": "43. Zielonka, W., Bolkart, T., Thies, J.: Towards metrical reconstruction of human faces. In: European Conference on Computer Vision. pp. 250-269. Springer (2022)" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 132, + 502, + 481, + 534 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 502, + 481, + 534 + ], + "spans": [ + { + "bbox": [ + 132, + 502, + 481, + 534 + ], + "type": "text", + "content": "44. Zielonka, W., Bolkart, T., Thies, J.: Instant volumetric head avatars. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 4574-4584 (2023)" + } + ] + } + ], + "index": 13 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 226, + 91, + 447, + 101 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 226, + 91, + 447, + 101 + ], + "spans": [ + { + "bbox": [ + 226, + 91, + 447, + 101 + ], + "type": "text", + "content": "PAV: Personalized Head Avatar from Video Collection" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 470, + 92, + 480, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 470, + 92, + 480, + 100 + ], + "spans": [ + { + "bbox": [ + 470, + 92, + 480, + 100 + ], + "type": "text", + "content": "17" + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 16 + } + ], + "_backend": "vlm", + "_version_name": "2.6.4" +} \ No newline at end of file diff --git a/2024/PCF-Lift_ Panoptic Lifting by Probabilistic Contrastive Fusion/43912c86-ca82-4896-afdb-6d1a9980542c_content_list.json b/2024/PCF-Lift_ Panoptic Lifting by Probabilistic Contrastive Fusion/43912c86-ca82-4896-afdb-6d1a9980542c_content_list.json new file mode 100644 index 0000000000000000000000000000000000000000..ffa1707bc6025550b75d80f7b08512e244da65f2 --- /dev/null +++ b/2024/PCF-Lift_ Panoptic Lifting by Probabilistic Contrastive Fusion/43912c86-ca82-4896-afdb-6d1a9980542c_content_list.json @@ -0,0 +1,1542 @@ +[ + { + "type": "text", + "text": "PCF-Lift: Panoptic Lifting by Probabilistic Contrastive Fusion", + "text_level": 1, + "bbox": [ + 316, + 140, + 684, + 183 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Runsong Zhu $^{1}$ , Shi Qiu $^{1}$ , Qianyi Wu $^{2}$ , Ka-Hei Hui $^{1}$ , Pheng-Ann Heng $^{1}$ , and Chi-Wing Fu $^{1}$", + "bbox": [ + 313, + 210, + 687, + 243 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "1 The Chinese University of Hong Kong \n2 Monash University", + "bbox": [ + 366, + 253, + 635, + 282 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Abstract. Panoptic lifting is an effective technique to address the 3D panoptic segmentation task by unprojecting 2D panoptic segmentations from multi-views to 3D scene. However, the quality of its results largely depends on the 2D segmentations, which could be noisy and error-prone, so its performance often drops significantly for complex scenes. In this work, we design a new pipeline coined PCF-Lift based on our Probabilistic Contrastive Fusion (PCF) to learn and embed probabilistic features throughout our pipeline to actively consider inaccurate segmentations and inconsistent instance IDs. Technical-wise, we first model the probabilistic feature embeddings through multivariate Gaussian distributions. To fuse the probabilistic features, we incorporate the probability product kernel into the contrastive loss formulation and design a cross-view constraint to enhance the feature consistency across different views. For the inference, we introduce a new probabilistic clustering method to effectively associate prototype features with the underlying 3D object instances for the generation of consistent panoptic segmentation results. Further, we provide a theoretical analysis to justify the superiority of the proposed probabilistic solution. By conducting extensive experiments, our PCF-lift not only significantly outperforms the state-of-the-art methods on widely used benchmarks including the ScanNet dataset and the challenging Messy Room dataset (4.4% improvement of scene-level PQ), but also demonstrates strong robustness when incorporating various 2D segmentation models or different levels of hand-crafted noise.", + "bbox": [ + 261, + 316, + 738, + 635 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Keywords: Panoptic Lifting $\\cdot$ Probabilistic Contrastive Fusion $\\cdot$ Probabilistic Feature Embeddings", + "bbox": [ + 261, + 650, + 740, + 678 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "1 Introduction", + "text_level": 1, + "bbox": [ + 215, + 703, + 374, + 718 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "3D panoptic segmentation [11,15,32,33,39,41,51] is a challenging 3D vision task, requiring the prediction of both semantic segmentation labels and instance segmentation labels. This task enables a comprehensive understanding of 3D scenes, thus facilitating many downstream applications, e.g., VR/AR, robotics, etc.", + "bbox": [ + 212, + 733, + 785, + 794 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Observing that the generalizability and applicability of existing 3D panoptic segmentation methods are limited by the scarcity of 3D training data, recent studies [3,25,40,43,49] suggest the idea of leveraging 2D panoptic segmentation", + "bbox": [ + 212, + 795, + 785, + 839 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/d2a718e81741a93182b97968c5df4432a04166f085d48d112925528e3b039880.jpg", + "image_caption": [ + "Fig. 1: Our PCF-Lift method unprojects 2D panoptic segmentation predictions to 3D domain, facilitating the generation of consistent panoptic segmentation masks. For simplicity and clarity, we highlight instance segmentation masks." + ], + "image_footnote": [], + "bbox": [ + 246, + 141, + 751, + 200 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "information predicted by foundation models [9, 24, 50]. Although 2D panoptic segmentation only provides image-based understanding, panoptic lifting [3, 40] is an effective technique to learn implicit 3D panoptic fields from 2D panoptic predictions, supporting the generation of coherent and view-consistent panoptic segmentation across different views.", + "bbox": [ + 212, + 303, + 782, + 378 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Recent works [3,40,44] on panoptic lifting mainly focus on solving the challenging instance-related issues, given that semantic predictions can be effectively associated across different views [49]. In practice, as shown in Fig. 1, one direct issue is inconsistent IDs that the same 3D object is assigned to different instance IDs in two different views for 2D panoptic prediction, which complicates the straightforward fusion process. To bypass the inconsistent IDs issue, Panoptic Lifting [40] directly fits instance ID permutations during the training process to learn deterministic feature embeddings for instance representation. Furthermore, Contrastive Lift [3] extends the scalability of Panoptic Lifting by using the contrastive learning technique to optimize the deterministic feature embeddings. However, existing methods still struggle to achieve satisfactory performance on complex scenes, as they overlook another interrelated issue: inconsistent segmentation. Specifically, due to the presence of inaccurate segmentation, the same object can be segmented inconsistently across two views: e.g., the chair in Fig. 1 is segmented into two parts in \"view 1\" but as a whole in \"view 2\". As existing methods use deterministic feature embeddings to learn 3D instance segmentation, the issue of inconsistent segmentation inherently introduces noise to training data and thus poses a robustness challenge for the models.", + "bbox": [ + 212, + 383, + 787, + 654 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "In this paper, we introduce an effective probabilistic contrastive fusion solution to collectively address the two issues. For the issue of inconsistent segmentation, we propose to learn probabilistic feature embeddings rather than deterministic feature embeddings used in [3,40]. Our key insight is the development of probabilistic features, which correspond to distributions, allows for the incorporation of uncertainty modeling and enhances robustness to noise. This leads to a stable model optimization and, ultimately, results in a more reliable representation of instance information. Given these benefits, we develop probabilistic feature embeddings based on multivariate Gaussian distributions. Particularly, to enable contrastive learning among different Gaussian distributions, we devise the Probability Product (PP) Kernel [18] to measure the probabilistic feature similarities. Accordingly, a probabilistic clustering algorithm is introduced", + "bbox": [ + 212, + 659, + 787, + 840 + ], + "page_idx": 1 + }, + { + "type": "page_number", + "text": "2", + "bbox": [ + 217, + 114, + 228, + 126 + ], + "page_idx": 1 + }, + { + "type": "header", + "text": "R. Zhu, S. Qiu, Q. Wu, et al.", + "bbox": [ + 271, + 114, + 465, + 128 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "in the inference phase to generate consistent panoptic segmentation, using the measured probabilistic feature similarities. For the issue of inconsistent IDs, we devise the contrastive loss to fuse the probabilistic feature embeddings. In addition to formulating the loss term with each single view observation as inspired by the prior work [3], we introduce a novel cross-view constraint to facilitate an effective model training from segmentation results with inconsistent instance IDs. By dynamically exploiting feature pairs from different views, our proposed constraint offers an effective formulation to further enhance the feature consistency of the same 3D object instance across multiple views, and thus improves the panoptic lifting performance.", + "bbox": [ + 212, + 146, + 787, + 297 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "To evaluate the effectiveness of our method, we conduct experiments on the ScanNet dataset [12] and the Messy Rooms dataset [3]. Furthermore, we conduct experiments to demonstrate the robustness of our probabilistic method to variations in segmentation models and different levels of noise. In addition to the experimental outcomes, our theoretical analysis from an optimization perspective demonstrates that the proposed probabilistic representation can be seen as a more flexible and generalized form compared to the previous deterministic representation [3].", + "bbox": [ + 212, + 297, + 787, + 417 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Our contributions are summarized as follows:", + "bbox": [ + 238, + 417, + 566, + 431 + ], + "page_idx": 2 + }, + { + "type": "list", + "sub_type": "text", + "list_items": [ + "- We introduce a probabilistic contrastive fusion solution (PCF-Lift) to effectively unproject the 2D panoptic segmentations to the 3D domain by collectively considering the issues of inconsistent segmentation and inconsistent IDs.", + "- To fuse the probabilistic feature embeddings modeled by multivariate Gaussian distributions, we reformulate the contrastive loss with the Probability Product kernel and propose a novel cross-view constraint to further encourage the multi-view consistency.", + "- Coupled with a new probabilistic clustering algorithm, our proposed method outperforms the state-of-the-art methods consistently on the ScanNet dataset and the Messy Room dataset." + ], + "bbox": [ + 225, + 443, + 785, + 608 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "2 Related Works", + "text_level": 1, + "bbox": [ + 215, + 632, + 395, + 648 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "2.1 Traditional 2D and 3D Panoptic Segmentation", + "text_level": 1, + "bbox": [ + 215, + 664, + 645, + 680 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "The 2D Panoptic segmentation task was initially introduced in [23]. Despite the notable advancements made by subsequent works [8-10,37,47], it remains challenging to panoptically understand individual images, while avoiding inconsistent instance recognitions across different image views of the scene.", + "bbox": [ + 212, + 688, + 785, + 748 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "To enhance the panoptic understanding of the real world, 3D panoptic segmentation focuses on segmenting pre-computed 3D structures [15,32,41,51] (e.g., point clouds, voxels) or performing simultaneous 3D reconstruction or segmentation from 2D images [11,33,39]. However, its generalizability is rather limited, largely due to the significant difference in scale between 2D and 3D training data. Given the capabilities of recent 3D reconstruction techniques [6,20,31] and 2D", + "bbox": [ + 212, + 750, + 787, + 840 + ], + "page_idx": 2 + }, + { + "type": "header", + "text": "PCF-Lift: Panoptic Lifting by Probabilistic Contrastive Fusion", + "bbox": [ + 313, + 114, + 730, + 128 + ], + "page_idx": 2 + }, + { + "type": "page_number", + "text": "3", + "bbox": [ + 774, + 116, + 784, + 126 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "panoptic segmentation models [9, 50], in this work, we explore 3D panoptic segmentation by leveraging multi-view images without explicit 3D input data.", + "bbox": [ + 215, + 146, + 784, + 176 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "2.2 Multi-view Fusion", + "text_level": 1, + "bbox": [ + 215, + 199, + 410, + 213 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Recently, researches in 3D reconstruction have made tremendous progress [2,6, 20, 29, 31]. Beyond novel view synthesis, we can utilize 3D reconstruction techniques as a tool to fuse 2D information like semantics and features in the 3D space [5, 14, 17, 21, 25, 35, 43, 46, 48]. For example, Semantic-NeRF [49] learns a semantic field from the 2D semantic segmentation, demonstrating the robustness and effectiveness of the semantic fusion. Later works [21, 25, 43] propose to fuse the unsupervised 2D dense feature for various segmentation and editing applications, showcasing the potential to adapt the 2D zero-shot models to 3D domain. In this work, we study the task of panoptic fusion, integrating both instance and semantic information to obtain a comprehensive understanding of the 3D scene. This task is significantly more challenging than the previous semantic fusion task [49], since we need to incorporate additional instance fusion and achieve instance label consistency. Unlike recent methods (e.g., Panoptic-Lifting [40] and Contrastive Lift [3]), which utilize deterministic feature embeddings to represent instance information, we develop a novel probabilistic solution, actively considering the error-prone and noisy nature of the 2D segmentations, such that we can duly enhance the effectiveness and robustness of instance fusion.", + "bbox": [ + 215, + 223, + 785, + 481 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "2.3 Probabilistic Representation", + "text_level": 1, + "bbox": [ + 215, + 503, + 495, + 517 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Probabilistic feature embedding is a popular tool employed in various tasks, e.g., image generation [22, 38], normal estimation [1], video understanding [36], point cloud understanding [4], and prototype embeddings for few-shot detection [42]. Motivated by its high capabilities in estimating aleatoric uncertainty [13] and addressing data noise, we innovate a new panoptic-lifting pipeline for 3D panoptic segmentation by formulating new modules based on probabilistic feature embeddings, including a reformulated contrastive loss with the probability product kernel [18] and an effective cross-view constraint to enhance the multi-view consistency. Further, we also design a novel probabilistic clustering algorithm to facilitate the generation of consistent panoptic segmentation.", + "bbox": [ + 215, + 527, + 785, + 679 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "3 Method", + "text_level": 1, + "bbox": [ + 215, + 702, + 328, + 717 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Given a set of posed images $\\{\\mathcal{I}\\}$ associated with 2D panoptic segmentation predictions (i.e., semantic masks $\\{\\mathcal{H}\\}$ and instance masks $\\{\\mathcal{K}\\}$ ) generated by 2D segmentation models, our goal is to learn accurate 3D panoptic fields that can be rendered into consistent panoptic segmentation results over different views.", + "bbox": [ + 215, + 734, + 784, + 792 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "The overview of PCF-Lift is illustrated in Fig. 2. Specifically, the 3D panoptic fields include a semantic field, an instance field, a density field, and a color field. Particularly, considering the significance of instance-related issues, we propose", + "bbox": [ + 215, + 795, + 784, + 839 + ], + "page_idx": 3 + }, + { + "type": "page_number", + "text": "4", + "bbox": [ + 217, + 114, + 228, + 126 + ], + "page_idx": 3 + }, + { + "type": "header", + "text": "R. Zhu, S. Qiu, Q. Wu, et al.", + "bbox": [ + 271, + 114, + 465, + 128 + ], + "page_idx": 3 + }, + { + "type": "image", + "img_path": "images/f6f4e7af24ecfec46155369fb1480767636138e609fe34faa13cdac3e3f754a0.jpg", + "image_caption": [ + "Fig. 2: Overview of PCF-Lift. The 3D panoptic fields include a semantic field, an instance field, a density field, and a color field. To solve the instance-related issues, we propose to learn probabilistic feature embeddings in the instance field (see Sec. 3.1). During the training phase, given two camera views, we can render the probabilistic feature maps from the instance field via volume rendering. To optimize the probabilistic instance field, we devise the probabilistic contrastive loss with Probability Product (PP) kernel [18], and propose a cross-view constraint to further enhance the feature consistency from different views (see Sec. 3.2). Similarly, we can render the semantic and color predictions, and adopt photometric loss and cross-entropy loss to optimize the semantic field, the density field, and the color field. During the inference phase, we design a novel multi-view object association (MVOA) algorithm for the generation of consistent panoptic segmentations (see Sec. 3.3)." + ], + "image_footnote": [], + "bbox": [ + 250, + 143, + 754, + 345 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "to learn probabilistic feature embeddings in the instances field (see Sec. 3.1). During the training phase, we jointly train the whole panoptic fields. Given two camera views, we can render the probabilistic feature maps from the instance field via volume rendering [19]. To optimize the probabilistic feature embeddings, we develop the contrastive loss with the Probability Product kernel [18], and propose a cross-view constraint to further enhance the feature consistency from different views (see Sec. 3.2). Similarly, the semantic predictions and color predictions are also rendered via volume rendering, where the semantic field, the density field, and the color field are optimized by the photometric loss and cross-entropy loss [3,6,40,49]. In the inference phase, we design a probabilistic clustering (i.e., multi-view object association) algorithm to effectively identify the prototype features of underlying 3D object instances for the generation of consistent panoptic segmentation results across any given views (see Sec. 3.3).", + "bbox": [ + 212, + 566, + 787, + 763 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "3.1 Probabilistic Feature Embeddings", + "text_level": 1, + "bbox": [ + 214, + 785, + 540, + 800 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "To provide a robust instance representation, we propose to learn probabilistic feature embeddings in the instance field, which maps each 3D point to a random", + "bbox": [ + 212, + 809, + 785, + 839 + ], + "page_idx": 4 + }, + { + "type": "header", + "text": "PCF-Lift: Panoptic Lifting by Probabilistic Contrastive Fusion", + "bbox": [ + 313, + 114, + 730, + 128 + ], + "page_idx": 4 + }, + { + "type": "page_number", + "text": "5", + "bbox": [ + 774, + 116, + 784, + 126 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "variable $\\mathcal{F}$ , modeled as an $N$ -dimensional multivariate Gaussian distribution $\\mathcal{F} \\sim \\mathcal{N}(\\mu, \\Sigma)$ . Here, $\\mu$ is the mean vector, indicating the central feature values and $\\Sigma$ is a diagonal covariance matrix $\\mathrm{diag}(\\sigma^2)$ with $\\sigma^2 = (\\sigma^{(1)^2}, \\sigma^{(2)^2}, \\dots, \\sigma^{(N)^2})$ . Concretely, for any given query point $\\mathbf{x} \\in \\mathbb{R}^3$ , the instance field predicts $(\\mu, \\sigma^2) \\in \\mathbb{R}^{2N}$ . Similar to the rendering of the color field, for each pixel in a given camera view, we can render its corresponding Gaussian distribution feature via volume rendering. A significant advantage of using probabilistic features rather than deterministic features is the ability of Gaussian distributions that assign different covariance values for uncertainty modeling. This property is crucial as it aids in reducing the impact of noise, thereby enhancing the robustness and accuracy of feature embeddings in representing complex 3D scenes.", + "bbox": [ + 212, + 146, + 787, + 313 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Intuitively, the similarity between two rendered Gaussian distributions indicates whether the corresponding pixels belong to the same instance. To quantify the similarities between different Gaussian distributions, we employ the Probability Product (PP) kernel [18] $K_{\\rho}$ . Specifically, given two Gaussian distributions $\\mathcal{F}_i \\sim \\mathcal{N}(\\mu_i,\\Sigma_i)$ and $\\mathcal{F}_j \\sim \\mathcal{N}(\\mu_j,\\Sigma_j)$ , the corresponding kernel can be explicitly formulated as", + "bbox": [ + 212, + 314, + 787, + 404 + ], + "page_idx": 5 + }, + { + "type": "equation", + "text": "\n$$\nK _ {\\rho} \\left(\\mathcal {F} _ {i}, \\mathcal {F} _ {j}\\right) = \\left(\\prod_ {d = 1} ^ {N} \\left(\\frac {\\sigma_ {i} ^ {(d) ^ {2}}}{\\sigma_ {j} ^ {(d) ^ {2}}} + \\frac {\\sigma_ {j} ^ {(d) ^ {2}}}{\\sigma_ {i} ^ {(d) ^ {2}}}\\right) / 2\\right) ^ {- \\frac {1}{2}} \\exp \\left(- \\sum_ {d = 1} ^ {N} \\left(\\mu_ {i} ^ {(d)} - \\mu_ {j} ^ {(d)}\\right) ^ {2} / 4 \\left(\\sigma_ {i} ^ {(d) ^ {2}} + \\sigma_ {j} ^ {(d) ^ {2}}\\right)\\right). \\tag {1}\n$$\n", + "text_format": "latex", + "bbox": [ + 222, + 411, + 792, + 455 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Note that the results produced by the PP kernel fall within the range of 0 (minimal similarity) to 1 (maximal similarity), the same as the Radial Basis Function (RBF) kernel used in the deterministic method [3].", + "bbox": [ + 212, + 457, + 787, + 503 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "3.2 Training Probabilistic Feature Embeddings", + "text_level": 1, + "bbox": [ + 214, + 523, + 617, + 539 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Probabilistic Contrastive Loss. During the training phase, we leverage the contrastive learning technique to optimize the probabilistic feature embeddings. Specifically, the contrastive loss used in [7, 27, 34, 45] tends to increase the predicted feature similarities of predefined positive pairs, while decreasing the similarities of negative pairs for training. In our work, a positive pair is defined as two pixels belonging to the same instance, while a negative pair denotes two pixels of different instances within a single image, following the previous method [3]. Since we basically learn probabilistic features, the PP kernel is exploited to measure the features similarity. In general, the loss can be formulated as", + "bbox": [ + 212, + 546, + 787, + 681 + ], + "page_idx": 5 + }, + { + "type": "equation", + "text": "\n$$\n\\mathcal {L} _ {\\mathrm {p i x e l - c o n t r a}} = - \\frac {1}{| \\varOmega |} \\sum_ {u \\in \\varOmega} \\log \\frac {\\sum_ {u ^ {\\prime} \\in \\varOmega} \\mathbf {1} _ {(u , u ^ {\\prime})} \\exp \\left(K _ {\\rho} \\left(\\mathcal {F} _ {u} , \\mathcal {F} _ {u ^ {\\prime}}\\right)\\right)}{\\sum_ {u ^ {\\prime} \\in \\varOmega} \\exp \\left(K _ {\\rho} \\left(\\mathcal {F} _ {u} , \\mathcal {F} _ {u ^ {\\prime}}\\right)\\right)}, \\tag {2}\n$$\n", + "text_format": "latex", + "bbox": [ + 285, + 691, + 785, + 729 + ], + "page_idx": 5 + }, + { + "type": "equation", + "text": "\n$$\n\\mathcal {L} _ {\\text {c o n c e n}} = - \\frac {1}{| \\varOmega |} \\sum_ {u \\in \\varOmega} \\log \\left(K _ {\\rho} \\left(\\mathcal {F} _ {u}, \\frac {\\sum_ {u ^ {\\prime} \\in \\varOmega} \\mathbf {1} _ {(u , u ^ {\\prime})} \\mathcal {F} _ {u ^ {\\prime}}}{\\sum_ {u ^ {\\prime} \\in \\varOmega} \\mathbf {1} _ {(u , u ^ {\\prime})}}\\right)\\right), \\tag {3}\n$$\n", + "text_format": "latex", + "bbox": [ + 302, + 739, + 785, + 777 + ], + "page_idx": 5 + }, + { + "type": "equation", + "text": "\n$$\n\\text {a n d} \\mathcal {L} _ {\\text {c o n t r a}} = \\mathcal {L} _ {\\text {p i x e l - c o n t r a}} + \\mathcal {L} _ {\\text {c o n c e n}}, \\tag {4}\n$$\n", + "text_format": "latex", + "bbox": [ + 372, + 789, + 785, + 804 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "where $\\mathcal{L}_{\\mathrm{pixel - contra}}$ is the pixel-wise contrastive loss, $\\mathcal{L}_{\\mathrm{concen}}$ is the concentrate loss term, $\\mathbf{1}$ is the indicator function of positive pairs, $\\Omega$ is the set of pixel sample, $\\mathcal{F}_u$", + "bbox": [ + 212, + 809, + 787, + 840 + ], + "page_idx": 5 + }, + { + "type": "page_number", + "text": "6", + "bbox": [ + 217, + 114, + 228, + 126 + ], + "page_idx": 5 + }, + { + "type": "header", + "text": "R. Zhu, S. Qiu, Q. Wu, et al.", + "bbox": [ + 271, + 114, + 467, + 128 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "and $\\mathcal{F}_{u'}$ are the rendered features for pixels $u$ and $u' \\in \\Omega$ via volume rendering, respectively. Different from previous method [3], we devise probabilistic similarity kernels to calculate a more effective contrastive loss for model training. By optimizing the loss Eq. (4), we can learn a more expressive 3D instance field for robust feature representations, as further analyzed in Sec. 3.4. To avoid the neural network from generating large covariances everywhere, we use an additional regularization term, $\\mathcal{L}_{\\mathrm{reg}} := \\log \\left( \\prod_{d=1}^{N} \\sigma^{(d)^2} \\right)$ , to penalize the large covariances.", + "bbox": [ + 212, + 146, + 787, + 253 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Cross-view Constraint. To better train the probabilistic feature embeddings, we propose a cross-view constraint to enhance feature consistency for the same object across different views. Given the rendered learned probabilistic feature embedding sets $\\{\\mathcal{F}^m\\}$ and $\\{\\mathcal{F}^n\\}$ for two different views, we devise the PP kernel $K_{\\rho}$ and a predefined threshold $\\tau$ to collect positive pairs $\\mathcal{P}$ that belong to the same object: $\\mathcal{P} = \\left\\{(\\mathcal{F}_r,\\mathcal{F}_s)\\mid K_\\rho (\\mathcal{F}_r,\\mathcal{F}_s) > \\tau ,\\mathcal{F}_r\\in \\{\\mathcal{F}^m\\} ,\\mathcal{F}_s\\in \\{\\mathcal{F}^n\\} \\right\\}$ . Moreover, as we expect to maximize the similarities between the features that belong to the same 3D object in different views, the cross-view constraint is defined as:", + "bbox": [ + 212, + 273, + 787, + 395 + ], + "page_idx": 6 + }, + { + "type": "equation", + "text": "\n$$\n\\mathcal {L} _ {\\text {c r o s s}} = - \\frac {1}{| \\mathcal {P} |} \\sum_ {\\left(\\mathcal {F} _ {r}, \\mathcal {F} _ {s}\\right) \\in \\mathcal {P}} \\log \\left(K _ {\\rho} \\left(\\mathcal {F} _ {r}, \\mathcal {F} _ {s}\\right)\\right). \\tag {5}\n$$\n", + "text_format": "latex", + "bbox": [ + 354, + 406, + 785, + 444 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "In practice, we use the cross-view constraint only in a later optimization stage (i.e., the last few epochs), where the feature space is sufficiently meaningful to provide reliable positive pairs. The overall loss is formulated as", + "bbox": [ + 214, + 448, + 787, + 494 + ], + "page_idx": 6 + }, + { + "type": "equation", + "text": "\n$$\n\\mathcal {L} = \\mathcal {L} _ {\\text {c o n t r a}} + w _ {\\text {c r o s s}} \\mathcal {L} _ {\\text {c r o s s}} + w _ {\\text {r e g}} \\mathcal {L} _ {\\text {r e g}}, \\tag {6}\n$$\n", + "text_format": "latex", + "bbox": [ + 369, + 505, + 785, + 522 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "where $w_{\\mathrm{cross}}$ and $w_{\\mathrm{reg}}$ are weight hyper-parameters.", + "bbox": [ + 214, + 531, + 588, + 547 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "3.3 Multi-view Object Association", + "text_level": 1, + "bbox": [ + 214, + 566, + 514, + 583 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "During the inference phase, to facilitate consistent panoptic segmentations, we introduce a novel clustering algorithm, named the multi-view object association (MVOA) algorithm. The algorithm aims to extract the prototype feature set from the learned instance field for the assignment of instance labels. In general, through volume rendering, we obtain a large-scale rendered per-pixel probabilistic feature set $\\{\\mathcal{F}\\}$ of the training views, as the input of our algorithm. Considering the efficiency issue, we first conduct an instance grouping operation to gather a smaller feature set $\\mathcal{C}$ , with the help of inconsistent instance segmentation masks $\\{\\mathcal{K}\\}$ obtained from training views. Then, we construct the probabilistic similarity graph based on $\\mathcal{C}$ and design a multi-view matching process to collect the prototype feature set $\\mathcal{D}$ . The details of instance grouping and multi-view matching operations are introduced as follows.", + "bbox": [ + 212, + 592, + 787, + 773 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Instance Grouping. We group the feature set $\\{\\mathcal{F}\\}$ from each individual view. For the view of $l$ -th image, we collect the pixels that belong to instance $p$ into the set $\\{i : \\mathcal{K}_l^i = p\\}$ , and the corresponding probabilistic feature set is denoted as", + "bbox": [ + 212, + 794, + 787, + 842 + ], + "page_idx": 6 + }, + { + "type": "header", + "text": "PCF-Lift: Panoptic Lifting by Probabilistic Contrastive Fusion", + "bbox": [ + 313, + 113, + 732, + 128 + ], + "page_idx": 6 + }, + { + "type": "page_number", + "text": "7", + "bbox": [ + 774, + 114, + 785, + 126 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "$\\{\\mathcal{F}_l^i:\\mathcal{K}_l^i = p\\}$ . Then, we group $\\{\\mathcal{F}_l^i:\\mathcal{K}_l^i = p\\}$ to a single probabilistic feature by calculating the average: $\\mathcal{C}_l^p = \\sum_i\\mathcal{F}_l^i /|\\{i:\\mathcal{K}_l^i = p\\} |.$ Concurrently, we model the score quantity $\\mathcal{S}_l^p$ for $\\mathcal{C}_l^p$ as an indicator of feature concentration, which aids in the subsequent multi-view matching process. The indicator function $\\varPhi(\\mathcal{C}_l^p)$ averages the PP kernel similarities between $\\mathcal{C}_l^p$ and all features in $\\{\\mathcal{F}_l^i:\\mathcal{K}_l^i = p\\}$ . Mathematically, $\\varPhi(\\mathcal{C}_l^p)$ follows: $\\mathcal{S}_l^p = \\varPhi(\\mathcal{C}_l^p) = \\sum_{\\{i:\\mathcal{K}_l^i = p\\}}K_\\rho (\\mathcal{C}_l^p,\\mathcal{F}_l^i) / |\\{i:\\mathcal{K}_l^i = p\\} |.$ By applying this across all observed images, we obtain the grouped feature set $\\mathcal{C}$ and the corresponding score set $\\mathcal{S}$ .", + "bbox": [ + 212, + 145, + 787, + 272 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Multi-view Matching. To extract the structural relationships among the features, we construct an unoriented similarity graph $G = (\\mathcal{C}, E)$ , where the grouped feature set $\\mathcal{C}$ represents the nodes and $E$ encompasses the edges that indicate quantitative feature similarities measured by the PP Kernel: $E_{} = E_{} = K_{\\rho}(\\mathcal{C}_g, \\mathcal{C}_h)$ , where $\\mathcal{C}_g, \\mathcal{C}_h \\in \\mathcal{C}$ . Then, we extract the prototype feature set $\\mathcal{D}$ from the graph $G$ and score set $S$ by a greedy procedure, which is akin to the classical greedy algorithm known as non-maximum suppression (NMS). The pseudo-code of our proposed MVOA algorithm is presented in Algo. 1. In practice, the selection of hyper-parameter $\\mathcal{T}$ is based on the average similarity computed across grouped feature pairs identified within each view.", + "bbox": [ + 212, + 291, + 789, + 444 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Generating Panoptic Segmentation Masks. We first follow the MVOA algorithm to collect the prototype feature set $\\mathcal{D}$ . For each view, including the novel view, we then generate the semantic labels from the learned semantic field to differentiate between the background and foreground. Coupled with a feature map rendered from the probabilistic instance field, for each foreground pixel, we finally determine its corresponding instance label by assigning the candidate index from the set $\\mathcal{D}$ that exhibits the highest similarity to the rendered feature. In practice, the MVOA algorithm needs to be conducted only once, while the extracted prototype features $\\mathcal{D}$ will be used for generating the panoptic segmentation masks of all test views.", + "bbox": [ + 212, + 464, + 789, + 617 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "3.4 Theoretical Analysis", + "text_level": 1, + "bbox": [ + 214, + 637, + 431, + 652 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Corollary 1 If the covariances of the given Gaussian distributions are isotropic and fixed, i.e., $\\Sigma_{i} = \\Sigma_{j} = \\sigma I$ , where $\\sigma$ is a constant scalar, the probability product kernel can be simplified to an RBF kernel.", + "bbox": [ + 212, + 662, + 787, + 708 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "The primary distinction between our probabilistic method and the prior deterministic method [3] is the choice of a similarity kernel. Specifically, our probabilistic method exploits the PP kernel, which offers higher flexibility and stronger expressive capability by adjusting the Gaussian mean and covariance compared to the RBF kernel employed in the deterministic method, as demonstrated in Fig 3. The RBF kernel can be considered as a degenerate PP kernel when applied to a Gaussian distribution with an isotropic and fixed covariance, as stated in Corollary 1. From this perspective, the proposed probabilistic method is a", + "bbox": [ + 212, + 719, + 787, + 842 + ], + "page_idx": 7 + }, + { + "type": "page_number", + "text": "8", + "bbox": [ + 217, + 114, + 228, + 126 + ], + "page_idx": 7 + }, + { + "type": "header", + "text": "R. Zhu, S. Qiu, Q. Wu, et al.", + "bbox": [ + 271, + 114, + 465, + 128 + ], + "page_idx": 7 + }, + { + "type": "code", + "sub_type": "algorithm", + "code_caption": [ + "Algorithm 1: Multi-view object association algorithm (MVOA)" + ], + "code_body": "Data: Inconsistent instance mask $\\{\\mathcal{K}\\}$ , input feature set $\\{\\mathcal{F}\\}$ , Threshold T \nResult: Prototype features set $\\mathcal{D}$ \n// Instance Grouping // Multi-view matching \n2 $\\mathcal{C} = \\{\\}$ , $\\mathcal{S} = \\{\\}$ 10 $\\mathcal{D} \\gets \\{\\}$ \n3 for $\\mathcal{K}_l$ in $\\{\\mathcal{K}\\}$ do 11 while $\\mathcal{S} \\neq empty$ do \n4 IDs = {unique_ID(\\mathcal{K}_l)} 12 m ← argmax S \n5 for $p$ in $IDs$ do D ← D ∪ $\\mathcal{C}_m$ \n6 $\\mathcal{C}_l^p = \\sum_i\\mathcal{F}_l^i /|\\{i:\\mathcal{K}_l^i = p\\}|$ C ← C - Cm; S ← S - Sm \n7 $\\mathcal{S}_l^p = \\Phi (\\mathcal{C}_l^p)$ 13 if Kρ(Cm, Ci) ≥ T then 14 | C ← C - Ci; S ← S - Si \n8 end 16 end \n9 end 17 end", + "bbox": [ + 228, + 167, + 735, + 364 + ], + "page_idx": 8 + }, + { + "type": "image", + "img_path": "images/5eb27910e9a7bbe8a14f61f75c1d7a5cd71931a08d3f1e524505cba76cef2509.jpg", + "image_caption": [ + "Fig. 3: (a) Flexibility of adjusting covariances. Contour plot of the PP kernel similarity for two Gaussians with different covariance values $(\\sigma_{1}$ and $\\sigma_{2})$ and fixed mean values in a 1-dimensional case. (b) Anisotropy. Contour plot of the PP kernel similarity for two Gaussians with different Gaussian mean offsets $(d_{x}$ and $d_{y})$ and fixed covariances in a 2-dimensional case. (c) Isotropy. Contour plot of the RBF kernel similarity for two deterministic features with different offsets $(d_{x}$ and $d_{y})$ in a 2-dimensional case." + ], + "image_footnote": [], + "bbox": [ + 246, + 383, + 751, + 491 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "more general framework, while the deterministic method is a subclass within the broader probabilistic paradigm. The experimental results also verify the effectiveness of our probabilistic solution in the ablation studies of Sec. 4.3.", + "bbox": [ + 212, + 627, + 787, + 674 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "4 Experiment", + "text_level": 1, + "bbox": [ + 214, + 694, + 366, + 712 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "4.1 Experimental Settings", + "text_level": 1, + "bbox": [ + 214, + 724, + 447, + 742 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "Implementation Details. For fair comparison, we adopt the same architecture of TensoRF [6] together with the same layer parameters used in previous works [3,40]. Particularly, our instance field is constructed following the slow-fast architecture of Contrastive Lift [3], where a shallow 5-layer MLP is applied to predict an probabilistic feature embedding for a 3D coordinate input. In practice, the dimension $N$ for the probabilistic feature embedding is 3 in our experiments.", + "bbox": [ + 212, + 750, + 787, + 840 + ], + "page_idx": 8 + }, + { + "type": "header", + "text": "PCF-Lift: Panoptic Lifting by Probabilistic Contrastive Fusion", + "bbox": [ + 313, + 114, + 730, + 128 + ], + "page_idx": 8 + }, + { + "type": "page_number", + "text": "9", + "bbox": [ + 774, + 116, + 784, + 126 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "For the training of the instance field, we sample rays from two different views in the last few epochs for optimization (Sec. 3.2), using the cross-view constraint $\\mathcal{L}_{\\mathrm{cross}}$ with a predefined threshold value $\\tau = 0.9$ and weight $w_{\\mathrm{cross}} = 0.05$ ; while in other training epochs, we set the weight $w_{\\mathrm{cross}}$ to 0 and only sample the rays from single views. Besides, $w_{\\mathrm{reg}}$ in Eq. (6) is set to 0.001 throughout the whole training process. For the training of the color, density and semantic fields, we adopt the same loss terms and training strategies from previous works [3, 40]. More implementation details are provided in the supplementary material.", + "bbox": [ + 212, + 146, + 787, + 268 + ], + "page_idx": 9 + }, + { + "type": "text", + "text": "Metrics. Since our method is proposed to solve the panoptic-lifting task, we employ scene-level Panoptic Quality (PQ $^{\\text{scene}}$ ) for evaluation, which was introduced in Panoptic Lifting [40] and widely used in the related works [3,40]. Unlike the standard PQ metric [23], this metric particularly considers the consistency of instance IDs across multiple views. By merging predictions and ground truths with consistent instance IDs into subsets for PQ $^{\\text{scene}}$ -based evaluation, we determine a match of subset pair when the intersection over union (IoU) exceeds 0.5, in line with prior baselines [3,40]. Since PQ $^{\\text{scene}}$ is a product of scene-level segmentation quality (SQ $^{\\text{scene}}$ ) and recognition quality (RQ $^{\\text{scene}}$ ), we also provide the SQ $^{\\text{scene}}$ and RQ $^{\\text{scene}}$ metrics for more detailed comparisons.", + "bbox": [ + 212, + 287, + 789, + 441 + ], + "page_idx": 9 + }, + { + "type": "text", + "text": "Baselines. We mainly compare the proposed method with the current state-of-the-art methods that target lifting 2D panoptic predictions to 3D, i.e., Contrastive Lift [3] and Panoptic Lifting [40]. Moreover, we compare our method with the recent NeRF-based 3D panoptic segmentation approaches, i.e., Panoptic Neural Fields [26] (PNF) and DM-NeRF [44].", + "bbox": [ + 212, + 458, + 787, + 536 + ], + "page_idx": 9 + }, + { + "type": "text", + "text": "Datasets. Following the previous works [3,40], we conduct experiments on two public datasets, the ScanNet dataset [12] and the Messy Room dataset [3], for both quantitative and qualitative evaluations. For fair comparisons on ScanNet, following the other state-of-the-art approaches [3,40], we adopt Mask2Former (M2F) [9] to generate 2D panoptic segmentation predictions, coupled with the same protocol in [40] that maps the COCO [28] vocabulary to 21 classes. For the experiments on Messy Rooms, we use the LVIS [16] vocabulary and the Detic [50] 2D panoptic segmentations, which are also utilized in Contrastive Lift [3].", + "bbox": [ + 212, + 555, + 787, + 678 + ], + "page_idx": 9 + }, + { + "type": "text", + "text": "4.2 Main Experiments", + "text_level": 1, + "bbox": [ + 214, + 696, + 416, + 710 + ], + "page_idx": 9 + }, + { + "type": "text", + "text": "ScanNet Dataset. To verify the performance on real data, we conduct experiments on the ScanNet dataset [12] of 12 scenes. Quantitative comparisons provided in Tab. 1 demonstrate that our method consistently outperforms previous baselines, including both the 3D panoptic segmentation approaches [26,44] and the recent state-of-the-art methods for lifting 2D panoptic segmentation [3,40]. Moreover, visual comparisons between our method and the latest state-of-the-art method [3] are presented in Fig. 4, which exhibits our method's capability of achieving a consistent and accurate 3D panoptic segmentation.", + "bbox": [ + 212, + 719, + 787, + 840 + ], + "page_idx": 9 + }, + { + "type": "page_number", + "text": "10", + "bbox": [ + 217, + 114, + 235, + 126 + ], + "page_idx": 9 + }, + { + "type": "header", + "text": "R. Zhu, S. Qiu, Q. Wu, et al.", + "bbox": [ + 271, + 114, + 465, + 128 + ], + "page_idx": 9 + }, + { + "type": "table", + "img_path": "images/42be9442a4b28aa70ba6117a17b80ad9e93f70a6a4b52e77398911d8fc8a40bd.jpg", + "table_caption": [ + "Table 1: Results on the ScanNet dataset. We report the PQscene, SQscene, and RQscene metrics. Since Contrastive Lift [3] does not report the performance using the SQscene and RQscene metrics, we apply the officially-released pre-trained model and re-run the clustering algorithm to obtain the values reported in this table. For the other metric values, we directly report the ones in previous papers [3,40]." + ], + "table_footnote": [], + "table_body": "
MethodVenueType\\(SQ^{scene}(\\%)\\)\\(RQ^{scene}(\\%)\\)\\(PQ^{scene}(\\%)\\)
DM-NeRF [44]ICLR'233D panoptic segmentation53.346.141.7
PNF [26]CVPR'223D panoptic segmentation63.050.748.3
PNF [26] + GT BBoxesCVPR'223D panoptic segmentation70.055.954.3
Panoptic Lifting [40]CVPR'232D panoptic Lifting73.565.058.9
Contrastive Lift [3]NeurIPS'232D panoptic Lifting75.763.662.0
Ours-2D panoptic Lifting78.565.463.5
", + "bbox": [ + 245, + 226, + 753, + 308 + ], + "page_idx": 10 + }, + { + "type": "table", + "img_path": "images/bdb5fa822af6124b8e21b8085593e21e193ed2665a169eb83e8c414113b41e04.jpg", + "table_caption": [ + "Table 2: Results on the Messy Rooms dataset [3]. Following [3], the PQscene metric is reported on both \"old room\" and \"large corridor\" environments with an increasing number of objects in the scene (25, 50, 100, 500)." + ], + "table_footnote": [], + "table_body": "
Method/ NumberOld Room Environment (%)Large Corridor Environment(%)Mean(%)
25501005002550100500
Panoptic Lifting [40]73.269.964.351.065.571.061.849.063.2
Contrastive Lift [3]78.975.869.155.076.575.568.752.569.0
Ours80.978.374.860.381.079.474.058.873.4
", + "bbox": [ + 245, + 375, + 753, + 444 + ], + "page_idx": 10 + }, + { + "type": "text", + "text": "Messy Rooms Dataset. We also conduct experiments on the challenging Messy Room dataset [3], which is provided by Contrastive Lift [3], containing up to 500 objects in each scene. We present the quantitative results in Tab. 2, showing that our method (73.4%) achieves significant improvements in terms of the mean PQ $^{\\text{scene}}$ results compared to the current state-of-the-art methods of Contrastive Lift [3] (69.0%) and Panoptic Lifting [40] (63.2%). The overall performance highlights the advantages of our probabilistic approach over the previous deterministic methods, particularly in segmenting complex scenes with hundreds of objects. Further, our method achieves a mean SQ $^{\\text{scene}}$ of 82.2% and a mean RQ $^{\\text{scene}}$ of 86.9%, surpassing the SQ $^{\\text{scene}}$ of 77.7% and RQ $^{\\text{scene}}$ of 86.6% obtained by Contrastive Lift [3]. Moreover, the visual comparisons in Fig. 4 further indicate that our method can accurately segment the small objects in the red frames, whereas Contrastive Lift [3] struggles to distinguish such instances.", + "bbox": [ + 212, + 470, + 787, + 667 + ], + "page_idx": 10 + }, + { + "type": "text", + "text": "4.3 Ablation Study", + "text_level": 1, + "bbox": [ + 214, + 691, + 388, + 708 + ], + "page_idx": 10 + }, + { + "type": "text", + "text": "Effectiveness of Each Component. We conduct an ablation study on the Messy Room dataset [3], which contains more challenging scenes with hundreds of instances. As shown in Tab. 3 (b) and (d), the proposed probabilistic feature embeddings greatly benefit our method, while replacing it with deterministic one leads to significant performance drop. Furthermore, the effectiveness of our proposed cross-view constraint is verified in Tab. 3 (e) and (f). For the proposed multi-view object association (MVOA) algorithm, the results in Tab. 3 (c), (d), (e), and (f) show that it particularly benefits the clustering of our proposed", + "bbox": [ + 212, + 718, + 787, + 842 + ], + "page_idx": 10 + }, + { + "type": "header", + "text": "PCF-Lift: Panoptic Lifting by Probabilistic Contrastive Fusion", + "bbox": [ + 313, + 113, + 730, + 128 + ], + "page_idx": 10 + }, + { + "type": "page_number", + "text": "11", + "bbox": [ + 767, + 114, + 782, + 126 + ], + "page_idx": 10 + }, + { + "type": "image", + "img_path": "images/39c375f5c14bb29d59de648735d61e2302061b04ed2f694827fcd1b78f3d147b.jpg", + "image_caption": [ + "Fig. 4: Visual comparison of the latest state-of-the-art method Contrastive Lift [3] and our method on the ScanNet [12] dataset and the Messy Room [3] dataset." + ], + "image_footnote": [], + "bbox": [ + 305, + 143, + 702, + 556 + ], + "page_idx": 11 + }, + { + "type": "table", + "img_path": "images/802b967888807f6914d8e20d9960a392ccd2d843989901e8353ca5d24548ee43.jpg", + "table_caption": [ + "Table 3: Ablation study on the Messy Room dataset [3]. The model (a) corresponds to Contrastive Lift [3] and the model (f) corresponds to our full method (PCF)." + ], + "table_footnote": [], + "table_body": "
Model Feature spaceClustering\\( {\\mathrm{{SQ}}}^{\\text{scene }}\\left( \\% \\right) \\)\\( {\\mathrm{{RQ}}}^{\\text{scene }}\\left( \\% \\right) \\)\\( {\\mathrm{{PQ}}}^{\\text{scene }}\\left( \\% \\right) \\)
(a)Deterministic [3]HDBSCAN [30]77.786.669.0
(b)Deterministic [3]MVOA79.386.270.4
(c)Learned Gaussian distributionHDBSCAN [30]78.086.669.6
(d)Learned Gaussian distributionMVOA81.386.872.3
(e)Learned Gaussian distribution (+ Cross-view constraint)HDBSCAN [30]78.886.970.4
(f)Learned Gaussian distribution (+ Cross-view constraint)MVOA82.286.973.4
", + "bbox": [ + 272, + 689, + 723, + 757 + ], + "page_idx": 11 + }, + { + "type": "text", + "text": "probabilistic representation. Also, MVOA can be effectively employed to the deterministic method [3] for a performance boost, as Tab. 3 (a) and (b) show.", + "bbox": [ + 212, + 809, + 785, + 840 + ], + "page_idx": 11 + }, + { + "type": "page_number", + "text": "12", + "bbox": [ + 217, + 114, + 235, + 126 + ], + "page_idx": 11 + }, + { + "type": "header", + "text": "R. Zhu, S. Qiu, Q. Wu, et al.", + "bbox": [ + 271, + 114, + 465, + 128 + ], + "page_idx": 11 + }, + { + "type": "image", + "img_path": "images/3b8cf2d43fca2acbdd6d2c4b542b71d6913358443b2654678a90e96ce0c07721.jpg", + "image_caption": [ + "Fig. 5: The visualization results of learned covariance components and the statistical results of covariances in two scenes of the Messy Room dataset [3]. For the histograms, the horizontal axis denotes the range of covariance magnitudes, while the vertical axis corresponds to the frequency statistics for those magnitudes. We calculate and plot the covariance magnitudes $(\\sigma^{(1)^2}*\\sigma^{(2)^2}*\\sigma^{(3)^2})$ of two distance image regions, the boundary areas and the internal areas of object instances, across all observed views." + ], + "image_footnote": [], + "bbox": [ + 282, + 141, + 718, + 250 + ], + "page_idx": 12 + }, + { + "type": "text", + "text": "Uncertainty Analysis. To verify whether our method could provide meaningful modeling for uncertainty, we provide the rendered uncertainty maps on two scenes of the Messy Room dataset [3] in Fig. 5. The figure clearly illustrates that the regions of high covariance are mainly located near the boundaries of the object instances, due to the inconsistent segmentation issue that leaves severe ambiguity around the instance boundaries. Moreover, we statistically analyze the learned covariances in two distinct image regions, namely the boundary areas and the internal areas of object instances, across all observed views. These results further demonstrate our method's ability to model high uncertainty in areas where the instance boundaries are ambiguous. We provide more details in the supplementary material.", + "bbox": [ + 212, + 383, + 787, + 551 + ], + "page_idx": 12 + }, + { + "type": "text", + "text": "4.4 Robustness Experiments", + "text_level": 1, + "bbox": [ + 215, + 570, + 465, + 585 + ], + "page_idx": 12 + }, + { + "type": "text", + "text": "2D Backbones. In practice, the quality of panoptic segmentation generated by different 2D models may vary a lot. To study the robustness of our probabilistic method and the deterministic method [3] when incorporating with different 2D models, we select four different models from the model zoo of Detic [3] and utilize the LVIS [16] vocabulary to generate the 2D panoptic segmentation predictions for lifting, as shown in Fig. 6 (a). The quantitative results are presented in Fig. 6 (b), where our probabilistic methods consistently outperform the baseline method by a large margin. Particularly, we observe that the proposed MVOA algorithm can consistently boost the performance of a deterministic method [3].", + "bbox": [ + 212, + 593, + 787, + 729 + ], + "page_idx": 12 + }, + { + "type": "text", + "text": "Hand-crafted Noise. We study the robustness by adding hand-crafted noise. Specifically, for all given panoptic segmentation masks, we randomly select 100 pixels as anchors. Then, for each anchor, we randomly choose a pixel within a $W * W$ window centered on the anchor and assign its instance ID to all pixels within this window, to simulate inaccurate segmentation predictions around object boundaries. As shown in Fig. 6 (c), as W gradually increases, the instance", + "bbox": [ + 212, + 750, + 787, + 840 + ], + "page_idx": 12 + }, + { + "type": "header", + "text": "PCF-Lift: Panoptic Lifting by Probabilistic Contrastive Fusion", + "bbox": [ + 313, + 114, + 730, + 128 + ], + "page_idx": 12 + }, + { + "type": "page_number", + "text": "13", + "bbox": [ + 767, + 114, + 784, + 126 + ], + "page_idx": 12 + }, + { + "type": "image", + "img_path": "images/36d2f150e399caa373b87d78d087fc90d1ac5b9babcbd02df12f42b5e2bf4583.jpg", + "image_caption": [ + "Fig. 6: Quantitative comparisons when using different 2D models or adding various levels of hand-crafted noise. We compare our probabilistic method (\"Probabilistic + MVOA\" in red) with the deterministic method [3] (\"Deterministic\" in dark blue) and a variant of deterministic method using our proposed MVOA algorithm (\"Deterministic + MVOA\" in light blue). Besides, the four tested models are from the model zoo of Detic $[50]^1$ , and the two tested scenes are from the Messy Room dataset [3]." + ], + "image_footnote": [], + "bbox": [ + 310, + 142, + 692, + 375 + ], + "page_idx": 13 + }, + { + "type": "text", + "text": "boundaries are continuously blurred. Our probabilistic approach consistently achieves the best performance as presented in Fig. 6 (d) and the results demonstrate that the proposed MVOA algorithm benefits the deterministic method [3].", + "bbox": [ + 212, + 517, + 784, + 564 + ], + "page_idx": 13 + }, + { + "type": "text", + "text": "5 Conclusion", + "text_level": 1, + "bbox": [ + 215, + 613, + 359, + 628 + ], + "page_idx": 13 + }, + { + "type": "text", + "text": "We present PCF-Lift for the panoptic lifting task. First, we propose to learn the probabilistic feature embeddings through a multivariate Gaussian distribution for instance representation. For training, we reformulate the contrastive loss with the Probability Product kernel and propose a novel cross-view constraint to enhance the feature consistency across different views. During the inference phase, we propose a novel multi-view object association algorithm to effectively identify the prototype features representing underlying 3D object instances. We verify the effectiveness and robustness of PCF-Lift by extensive experiments.", + "bbox": [ + 212, + 657, + 787, + 779 + ], + "page_idx": 13 + }, + { + "type": "page_number", + "text": "14", + "bbox": [ + 217, + 114, + 235, + 126 + ], + "page_idx": 13 + }, + { + "type": "header", + "text": "R. Zhu, S. Qiu, Q. Wu, et al.", + "bbox": [ + 271, + 114, + 465, + 128 + ], + "page_idx": 13 + }, + { + "type": "page_footnote", + "text": "1 We use the official pre-trained models provided in https://github.com/facebookresearch/Detic/tree/main. Please refer to our supplementary material for the details of the four models.", + "bbox": [ + 217, + 796, + 787, + 839 + ], + "page_idx": 13 + }, + { + "type": "text", + "text": "Acknowledgements", + "text_level": 1, + "bbox": [ + 217, + 143, + 401, + 162 + ], + "page_idx": 14 + }, + { + "type": "text", + "text": "This work is supported by the InnoHK of the Government of Hong Kong via the Hong Kong Centre for Logistics Robotics, the Shenzhen Portion of Shenzhen-Hong Kong Science and Technology Innovation Cooperation Zone under HZQB-KCZYB-20200089, and the CUHK T Stone Robotics Institute.", + "bbox": [ + 212, + 176, + 787, + 236 + ], + "page_idx": 14 + }, + { + "type": "text", + "text": "References", + "text_level": 1, + "bbox": [ + 217, + 258, + 321, + 273 + ], + "page_idx": 14 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "1. Bae, G., Budvytis, I., Cipolla, R.: Estimating and exploiting the aleatoric uncertainty in surface normal estimation. In: Proceedings of the IEEE/CVF International Conference on Computer Vision. pp. 13137-13146 (2021)", + "2. Barron, J.T., Mildenhall, B., Tancik, M., Hedman, P., Martin-Brualla, R., Srinivasan, P.P.: Mip-NeRF: A multiscale representation for anti-aliasing neural radiance fields. In: Proceedings of the IEEE/CVF International Conference on Computer Vision. pp. 5855-5864 (2021)", + "3. Bhalgat, Y., Laina, I., Henriques, J.F., Zisserman, A., Vedaldi, A.: Contrastive Lift: 3D object instance segmentation by slow-fast contrastive fusion. arXiv preprint arXiv:2306.04633 (2023)", + "4. Cai, K., Lu, C.X., Huang, X.: Uncertainty estimation for 3D dense prediction via cross-point embeddings. IEEE Robotics and Automation Letters 8(5), 2558-2565 (2023)", + "5. Cen, J., Zhou, Z., Fang, J., Shen, W., Xie, L., Jiang, D., Zhang, X., Tian, Q., et al.: Segment anything in 3D with NeRFs. Advances in Neural Information Processing Systems 36 (2024)", + "6. Chen, A., Xu, Z., Geiger, A., Yu, J., Su, H.: Tensorf: Tensorial radiance fields. In: European Conference on Computer Vision. pp. 333-350. Springer (2022)", + "7. Chen, T., Kornblith, S., Norouzi, M., Hinton, G.: A simple framework for contrastive learning of visual representations. In: International conference on machine learning. pp. 1597-1607. PMLR (2020)", + "8. Cheng, B., Collins, M.D., Zhu, Y., Liu, T., Huang, T.S., Adam, H., Chen, L.C.: Panoptic-deeplab: A simple, strong, and fast baseline for bottom-up panoptic segmentation. In: Proceedings of the IEEE/CVF conference on computer vision and pattern recognition. pp. 12475-12485 (2020)", + "9. Cheng, B., Misra, I., Schwing, A.G., Kirillov, A., Girdhar, R.: Masked-attention mask transformer for universal image segmentation. In: Proceedings of the IEEE/CVF conference on computer vision and pattern recognition. pp. 1290-1299 (2022)", + "0. Cheng, B., Schwing, A., Kirillov, A.: Per-pixel classification is not all you need for semantic segmentation. Advances in Neural Information Processing Systems 34, 17864-17875 (2021)", + "1. Dahnert, M., Hou, J., Nießner, M., Dai, A.: Panoptic 3D scene reconstruction from a single rgb image. Advances in Neural Information Processing Systems 34, 8282-8293 (2021)", + "12. Dai, A., Chang, A.X., Savva, M., Halber, M., Funkhouser, T., Nießner, M.: Scannet: Richly-annotated 3D reconstructions of indoor scenes. In: Proceedings of the IEEE conference on computer vision and pattern recognition. pp. 5828-5839 (2017)", + "13. Der Kiureghian, A., Ditlevsen, O.: Aleatory or epistemic? does it matter? Structural safety 31(2), 105-112 (2009)" + ], + "bbox": [ + 218, + 287, + 784, + 839 + ], + "page_idx": 14 + }, + { + "type": "header", + "text": "PCF-Lift: Panoptic Lifting by Probabilistic Contrastive Fusion", + "bbox": [ + 313, + 114, + 730, + 128 + ], + "page_idx": 14 + }, + { + "type": "page_number", + "text": "15", + "bbox": [ + 767, + 116, + 785, + 126 + ], + "page_idx": 14 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "14. Fan, Z., Wang, P., Jiang, Y., Gong, X., Xu, D., Wang, Z.: NeRF-SOS: Any-iew self-supervised object segmentation on complex scenes. arXiv preprint arXiv:2209.08776 (2022)", + "15. Gasperini, S., Mahani, M.A.N., Marcos-Ramiro, A., Navab, N., Tombari, F.: Panostter: End-to-end panoptic segmentation of LiDAR point clouds. IEEE Robotics and Automation Letters 6(2), 3216-3223 (2021)", + "16. Gupta, A., Dollar, P., Girshick, R.: LVIS: A dataset for large vocabulary instance segmentation. In: Proceedings of the IEEE/CVF conference on computer vision and pattern recognition. pp. 5356-5364 (2019)", + "17. Hamilton, M., Zhang, Z., Hariharan, B., Snavely, N., Freeman, W.T.: Unsupervised semantic segmentation by distilling feature correspondences. arXiv preprint arXiv:2203.08414 (2022)", + "18. Jebara, T., Kondor, R., Howard, A.: Probability product kernels. The Journal of Machine Learning Research 5, 819-844 (2004)", + "19. Kajiya, J.T., Von Herzen, B.P.: Ray tracing volume densities. ACM SIGGRAPH computer graphics 18(3), 165-174 (1984)", + "20. Kerbl, B., Kopanas, G., Leimkuhler, T., Drettakis, G.: 3D gaussian splatting for real-time radiance field rendering. ACM Transactions on Graphics 42(4) (2023)", + "21. Kerr, J., Kim, C.M., Goldberg, K., Kanazawa, A., Tancik, M.: LERF: Language embedded radiance fields. In: Proceedings of the IEEE/CVF International Conference on Computer Vision. pp. 19729-19739 (2023)", + "22. Kingma, D.P., Welling, M.: Auto-encoding variational bayes. arXiv preprint arXiv:1312.6114 (2013)", + "23. Kirillov, A., He, K., Girshick, R., Rother, C., Dollar, P.: Panoptic segmentation. In: Proceedings of the IEEE/CVF conference on computer vision and pattern recognition. pp. 9404-9413 (2019)", + "24. Kirillov, A., Mintun, E., Ravi, N., Mao, H., Rolland, C., Gustafson, L., Xiao, T., Whitehead, S., Berg, A.C., Lo, W.Y., et al.: Segment anything. arXiv preprint arXiv:2304.02643 (2023)", + "25. Kobayashi, S., Matsumoto, E., Sitzmann, V.: Decomposing NeRF for editing via feature field distillation. Advances in Neural Information Processing Systems 35, 23311-23330 (2022)", + "26. Kundu, A., Genova, K., Yin, X., Fathi, A., Pantofaru, C., Guibas, L.J., Tagliasacchi, A., Dellaert, F., Funkhouser, T.: Panoptic neural fields: A semantic object-aware neural scene representation. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 12871-12881 (2022)", + "27. Le-Khac, P.H., Healy, G., Smeaton, A.F.: Contrastive representation learning: A framework and review. IEEE Access 8, 193907-193934 (2020)", + "28. Lin, T.Y., Maire, M., Belongie, S., Hays, J., Perona, P., Ramanan, D., Dollar, P., Zitnick, C.L.: Microsoft COCO: Common objects in context. In: Computer Vision-ECCV 2014: 13th European Conference, Zurich, Switzerland, September 6-12, 2014, Proceedings, Part V 13. pp. 740-755. Springer (2014)", + "29. Liu, L., Gu, J., Zaw Lin, K., Chua, T.S., Theobalt, C.: Neural sparse voxel fields. Advances in Neural Information Processing Systems 33, 15651-15663 (2020)", + "30. McInnes, L., Healy, J., Astels, S.: hdbscan: Hierarchical density based clustering. J. Open Source Softw. 2(11), 205 (2017)", + "31. Mildenhall, B., Srinivasan, P.P., Tancik, M., Barron, J.T., Ramamoorthi, R., Ng, R.: NeRF: Representing scenes as neural radiance fields for view synthesis. Communications of the ACM 65(1), 99-106 (2021)" + ], + "bbox": [ + 215, + 146, + 785, + 839 + ], + "page_idx": 15 + }, + { + "type": "page_number", + "text": "16", + "bbox": [ + 217, + 114, + 235, + 126 + ], + "page_idx": 15 + }, + { + "type": "header", + "text": "R. Zhu, S. Qiu, Q. Wu, et al.", + "bbox": [ + 271, + 114, + 465, + 128 + ], + "page_idx": 15 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "32. Milioto, A., Behley, J., McCool, C., Stachniss, C.: LiDAR panoptic segmentation for autonomous driving. In: 2020 IEEE/RSJ International Conference on Intelligent Robots and Systems (IROS). pp. 8505-8512. IEEE (2020)", + "33. Narita, G., Seno, T., Ishikawa, T., Kaji, Y.: Panopticfusion: Online volumetric semantic mapping at the level of stuff and things. In: 2019 IEEE/RSJ International Conference on Intelligent Robots and Systems (IROS). pp. 4205-4212. IEEE (2019)", + "34. Newell, A., Huang, Z., Deng, J.: Associative embedding: End-to-end learning for joint detection and grouping. Advances in neural information processing systems 30 (2017)", + "35. Nie, Y., Han, X., Guo, S., Zheng, Y., Chang, J., Zhang, J.J.: Total3D Understanding: Joint layout, object pose and mesh reconstruction for indoor scenes from a single image. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 55-64 (2020)", + "36. Park, J., Lee, J., Kim, I.J., Sohn, K.: Probabilistic representations for video contrastive learning. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 14711-14721 (2022)", + "37. Porzi, L., Bulo, S.R., Colovic, A., Kontschieder, P.: Seamless scene segmentation. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 8277-8286 (2019)", + "38. Rezende, D.J., Mohamed, S., Wierstra, D.: Stochastic backpropagation and approximate inference in deep generative models. In: International conference on machine learning. pp. 1278-1286. PMLR (2014)", + "39. Rosinol, A., Gupta, A., Abate, M., Shi, J., Carlone, L.: 3D dynamic scene graphs: Actionable spatial perception with places, objects, and humans. arXiv preprint arXiv:2002.06289 (2020)", + "40. Siddiqui, Y., Porzi, L., Bulò, S.R., Müller, N., Nießner, M., Dai, A., Kontschieder, P.: Panoptic lifting for 3D scene understanding with neural fields. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 9043-9052 (2023)", + "41. Sirohi, K., Mohan, R., Buscher, D., Burgard, W., Valada, A.: Efficientlps: Efficient LiDAR panoptic segmentation. IEEE Transactions on Robotics 38(3), 1894-1914 (2021)", + "42. Tang, W., Biqi, Y., Li, X., Liu, Y.H., Heng, P.A., Fu, C.W.: Prototypical variational autoencoder for 3D few-shot object detection. In: Thirty-seventh Conference on Neural Information Processing Systems (2023)", + "43. Tschernezki, V., Laina, I., Larlus, D., Vedaldi, A.: Neural feature fusion fields: 3d distillation of self-supervised 2d image representations. In: 2022 International Conference on 3D Vision (3DV). pp. 443-453. IEEE (2022)", + "44. Wang, B., Chen, L., Yang, B.: Dm-NeRF: 3D scene geometry decomposition and manipulation from 2D images. arXiv preprint arXiv:2208.07227 (2022)", + "45. Xie, S., Gu, J., Guo, D., Qi, C.R., Guibas, L., Litany, O.: Pointcontrast: Unsupervised pre-training for 3D point cloud understanding. In: Computer Vision-ECCV 2020: 16th European Conference, Glasgow, UK, August 23-28, 2020, Proceedings, Part III 16. pp. 574-591. Springer (2020)", + "46. Ye, M., Danelljan, M., Yu, F., Ke, L.: Gaussian grouping: Segment and edit anything in 3D scenes. arXiv preprint arXiv:2312.00732 (2023)", + "47. Zhang, W., Pang, J., Chen, K., Loy, C.C.: K-net: Towards unified image segmentation. Advances in Neural Information Processing Systems 34, 10326-10338 (2021)", + "48. Zhang, X., Chen, Z., Wei, F., Tu, Z.: Uni-3D: A universal model for panoptic 3D scene reconstruction. In: Proceedings of the IEEE/CVF International Conference on Computer Vision. pp. 9256-9266 (2023)" + ], + "bbox": [ + 212, + 146, + 787, + 840 + ], + "page_idx": 16 + }, + { + "type": "header", + "text": "PCF-Lift: Panoptic Lifting by Probabilistic Contrastive Fusion", + "bbox": [ + 313, + 114, + 730, + 128 + ], + "page_idx": 16 + }, + { + "type": "page_number", + "text": "17", + "bbox": [ + 767, + 116, + 785, + 126 + ], + "page_idx": 16 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "49. Zhi, S., Laidlow, T., Leutenegger, S., Davison, A.J.: In-place scene labelling and understanding with implicit scene representation. In: Proceedings of the IEEE/CVF International Conference on Computer Vision. pp. 15838-15847 (2021)", + "50. Zhou, X., Girdhar, R., Joulin, A., Krahenbuhl, P., Misra, I.: Detecting twenty-thousand classes using image-level supervision. In: European Conference on Computer Vision. pp. 350-368. Springer (2022)", + "51. Zhou, Z., Zhang, Y., Foroosh, H.: Panoptic-polarnet: Proposal-free LiDAR point cloud panoptic segmentation. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 13194-13203 (2021)" + ], + "bbox": [ + 215, + 146, + 784, + 272 + ], + "page_idx": 17 + }, + { + "type": "page_number", + "text": "18", + "bbox": [ + 217, + 114, + 235, + 126 + ], + "page_idx": 17 + }, + { + "type": "header", + "text": "R. Zhu, S. Qiu, Q. Wu, et al.", + "bbox": [ + 271, + 114, + 465, + 128 + ], + "page_idx": 17 + } +] \ No newline at end of file diff --git a/2024/PCF-Lift_ Panoptic Lifting by Probabilistic Contrastive Fusion/43912c86-ca82-4896-afdb-6d1a9980542c_model.json b/2024/PCF-Lift_ Panoptic Lifting by Probabilistic Contrastive Fusion/43912c86-ca82-4896-afdb-6d1a9980542c_model.json new file mode 100644 index 0000000000000000000000000000000000000000..c47b5ff3989656a194c1bde811131299f0f4c48c --- /dev/null +++ b/2024/PCF-Lift_ Panoptic Lifting by Probabilistic Contrastive Fusion/43912c86-ca82-4896-afdb-6d1a9980542c_model.json @@ -0,0 +1,2150 @@ +[ + [ + { + "type": "title", + "bbox": [ + 0.318, + 0.141, + 0.685, + 0.184 + ], + "angle": 0, + "content": "PCF-Lift: Panoptic Lifting by Probabilistic Contrastive Fusion" + }, + { + "type": "text", + "bbox": [ + 0.315, + 0.212, + 0.689, + 0.244 + ], + "angle": 0, + "content": "Runsong Zhu\\(^{1}\\), Shi Qiu\\(^{1}\\), Qianyi Wu\\(^{2}\\), Ka-Hei Hui\\(^{1}\\), Pheng-Ann Heng\\(^{1}\\), and Chi-Wing Fu\\(^{1}\\)" + }, + { + "type": "text", + "bbox": [ + 0.367, + 0.255, + 0.637, + 0.283 + ], + "angle": 0, + "content": "1 The Chinese University of Hong Kong \n2 Monash University" + }, + { + "type": "text", + "bbox": [ + 0.263, + 0.318, + 0.74, + 0.636 + ], + "angle": 0, + "content": "Abstract. Panoptic lifting is an effective technique to address the 3D panoptic segmentation task by unprojecting 2D panoptic segmentations from multi-views to 3D scene. However, the quality of its results largely depends on the 2D segmentations, which could be noisy and error-prone, so its performance often drops significantly for complex scenes. In this work, we design a new pipeline coined PCF-Lift based on our Probabilistic Contrastive Fusion (PCF) to learn and embed probabilistic features throughout our pipeline to actively consider inaccurate segmentations and inconsistent instance IDs. Technical-wise, we first model the probabilistic feature embeddings through multivariate Gaussian distributions. To fuse the probabilistic features, we incorporate the probability product kernel into the contrastive loss formulation and design a cross-view constraint to enhance the feature consistency across different views. For the inference, we introduce a new probabilistic clustering method to effectively associate prototype features with the underlying 3D object instances for the generation of consistent panoptic segmentation results. Further, we provide a theoretical analysis to justify the superiority of the proposed probabilistic solution. By conducting extensive experiments, our PCF-lift not only significantly outperforms the state-of-the-art methods on widely used benchmarks including the ScanNet dataset and the challenging Messy Room dataset (4.4% improvement of scene-level PQ), but also demonstrates strong robustness when incorporating various 2D segmentation models or different levels of hand-crafted noise." + }, + { + "type": "text", + "bbox": [ + 0.262, + 0.651, + 0.741, + 0.679 + ], + "angle": 0, + "content": "Keywords: Panoptic Lifting \\(\\cdot\\) Probabilistic Contrastive Fusion \\(\\cdot\\) Probabilistic Feature Embeddings" + }, + { + "type": "title", + "bbox": [ + 0.217, + 0.704, + 0.375, + 0.719 + ], + "angle": 0, + "content": "1 Introduction" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.734, + 0.786, + 0.795 + ], + "angle": 0, + "content": "3D panoptic segmentation [11,15,32,33,39,41,51] is a challenging 3D vision task, requiring the prediction of both semantic segmentation labels and instance segmentation labels. This task enables a comprehensive understanding of 3D scenes, thus facilitating many downstream applications, e.g., VR/AR, robotics, etc." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.796, + 0.787, + 0.84 + ], + "angle": 0, + "content": "Observing that the generalizability and applicability of existing 3D panoptic segmentation methods are limited by the scarcity of 3D training data, recent studies [3,25,40,43,49] suggest the idea of leveraging 2D panoptic segmentation" + } + ], + [ + { + "type": "page_number", + "bbox": [ + 0.218, + 0.116, + 0.23, + 0.127 + ], + "angle": 0, + "content": "2" + }, + { + "type": "header", + "bbox": [ + 0.272, + 0.115, + 0.467, + 0.129 + ], + "angle": 0, + "content": "R. Zhu, S. Qiu, Q. Wu, et al." + }, + { + "type": "image", + "bbox": [ + 0.247, + 0.142, + 0.753, + 0.202 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.214, + 0.225, + 0.785, + 0.268 + ], + "angle": 0, + "content": "Fig. 1: Our PCF-Lift method unprojects 2D panoptic segmentation predictions to 3D domain, facilitating the generation of consistent panoptic segmentation masks. For simplicity and clarity, we highlight instance segmentation masks." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.304, + 0.784, + 0.38 + ], + "angle": 0, + "content": "information predicted by foundation models [9, 24, 50]. Although 2D panoptic segmentation only provides image-based understanding, panoptic lifting [3, 40] is an effective technique to learn implicit 3D panoptic fields from 2D panoptic predictions, supporting the generation of coherent and view-consistent panoptic segmentation across different views." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.384, + 0.788, + 0.655 + ], + "angle": 0, + "content": "Recent works [3,40,44] on panoptic lifting mainly focus on solving the challenging instance-related issues, given that semantic predictions can be effectively associated across different views [49]. In practice, as shown in Fig. 1, one direct issue is inconsistent IDs that the same 3D object is assigned to different instance IDs in two different views for 2D panoptic prediction, which complicates the straightforward fusion process. To bypass the inconsistent IDs issue, Panoptic Lifting [40] directly fits instance ID permutations during the training process to learn deterministic feature embeddings for instance representation. Furthermore, Contrastive Lift [3] extends the scalability of Panoptic Lifting by using the contrastive learning technique to optimize the deterministic feature embeddings. However, existing methods still struggle to achieve satisfactory performance on complex scenes, as they overlook another interrelated issue: inconsistent segmentation. Specifically, due to the presence of inaccurate segmentation, the same object can be segmented inconsistently across two views: e.g., the chair in Fig. 1 is segmented into two parts in \"view 1\" but as a whole in \"view 2\". As existing methods use deterministic feature embeddings to learn 3D instance segmentation, the issue of inconsistent segmentation inherently introduces noise to training data and thus poses a robustness challenge for the models." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.66, + 0.788, + 0.841 + ], + "angle": 0, + "content": "In this paper, we introduce an effective probabilistic contrastive fusion solution to collectively address the two issues. For the issue of inconsistent segmentation, we propose to learn probabilistic feature embeddings rather than deterministic feature embeddings used in [3,40]. Our key insight is the development of probabilistic features, which correspond to distributions, allows for the incorporation of uncertainty modeling and enhances robustness to noise. This leads to a stable model optimization and, ultimately, results in a more reliable representation of instance information. Given these benefits, we develop probabilistic feature embeddings based on multivariate Gaussian distributions. Particularly, to enable contrastive learning among different Gaussian distributions, we devise the Probability Product (PP) Kernel [18] to measure the probabilistic feature similarities. Accordingly, a probabilistic clustering algorithm is introduced" + } + ], + [ + { + "type": "header", + "bbox": [ + 0.314, + 0.115, + 0.732, + 0.129 + ], + "angle": 0, + "content": "PCF-Lift: Panoptic Lifting by Probabilistic Contrastive Fusion" + }, + { + "type": "page_number", + "bbox": [ + 0.775, + 0.117, + 0.785, + 0.127 + ], + "angle": 0, + "content": "3" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.147, + 0.788, + 0.298 + ], + "angle": 0, + "content": "in the inference phase to generate consistent panoptic segmentation, using the measured probabilistic feature similarities. For the issue of inconsistent IDs, we devise the contrastive loss to fuse the probabilistic feature embeddings. In addition to formulating the loss term with each single view observation as inspired by the prior work [3], we introduce a novel cross-view constraint to facilitate an effective model training from segmentation results with inconsistent instance IDs. By dynamically exploiting feature pairs from different views, our proposed constraint offers an effective formulation to further enhance the feature consistency of the same 3D object instance across multiple views, and thus improves the panoptic lifting performance." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.299, + 0.788, + 0.419 + ], + "angle": 0, + "content": "To evaluate the effectiveness of our method, we conduct experiments on the ScanNet dataset [12] and the Messy Rooms dataset [3]. Furthermore, we conduct experiments to demonstrate the robustness of our probabilistic method to variations in segmentation models and different levels of noise. In addition to the experimental outcomes, our theoretical analysis from an optimization perspective demonstrates that the proposed probabilistic representation can be seen as a more flexible and generalized form compared to the previous deterministic representation [3]." + }, + { + "type": "text", + "bbox": [ + 0.24, + 0.419, + 0.568, + 0.433 + ], + "angle": 0, + "content": "Our contributions are summarized as follows:" + }, + { + "type": "text", + "bbox": [ + 0.226, + 0.444, + 0.787, + 0.503 + ], + "angle": 0, + "content": "- We introduce a probabilistic contrastive fusion solution (PCF-Lift) to effectively unproject the 2D panoptic segmentations to the 3D domain by collectively considering the issues of inconsistent segmentation and inconsistent IDs." + }, + { + "type": "text", + "bbox": [ + 0.226, + 0.504, + 0.787, + 0.564 + ], + "angle": 0, + "content": "- To fuse the probabilistic feature embeddings modeled by multivariate Gaussian distributions, we reformulate the contrastive loss with the Probability Product kernel and propose a novel cross-view constraint to further encourage the multi-view consistency." + }, + { + "type": "text", + "bbox": [ + 0.226, + 0.565, + 0.787, + 0.609 + ], + "angle": 0, + "content": "- Coupled with a new probabilistic clustering algorithm, our proposed method outperforms the state-of-the-art methods consistently on the ScanNet dataset and the Messy Room dataset." + }, + { + "type": "list", + "bbox": [ + 0.226, + 0.444, + 0.787, + 0.609 + ], + "angle": 0, + "content": null + }, + { + "type": "title", + "bbox": [ + 0.216, + 0.633, + 0.396, + 0.649 + ], + "angle": 0, + "content": "2 Related Works" + }, + { + "type": "title", + "bbox": [ + 0.216, + 0.665, + 0.646, + 0.681 + ], + "angle": 0, + "content": "2.1 Traditional 2D and 3D Panoptic Segmentation" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.689, + 0.787, + 0.749 + ], + "angle": 0, + "content": "The 2D Panoptic segmentation task was initially introduced in [23]. Despite the notable advancements made by subsequent works [8-10,37,47], it remains challenging to panoptically understand individual images, while avoiding inconsistent instance recognitions across different image views of the scene." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.75, + 0.788, + 0.841 + ], + "angle": 0, + "content": "To enhance the panoptic understanding of the real world, 3D panoptic segmentation focuses on segmenting pre-computed 3D structures [15,32,41,51] (e.g., point clouds, voxels) or performing simultaneous 3D reconstruction or segmentation from 2D images [11,33,39]. However, its generalizability is rather limited, largely due to the significant difference in scale between 2D and 3D training data. Given the capabilities of recent 3D reconstruction techniques [6,20,31] and 2D" + } + ], + [ + { + "type": "page_number", + "bbox": [ + 0.218, + 0.116, + 0.23, + 0.127 + ], + "angle": 0, + "content": "4" + }, + { + "type": "header", + "bbox": [ + 0.272, + 0.115, + 0.467, + 0.129 + ], + "angle": 0, + "content": "R. Zhu, S. Qiu, Q. Wu, et al." + }, + { + "type": "text", + "bbox": [ + 0.217, + 0.147, + 0.785, + 0.178 + ], + "angle": 0, + "content": "panoptic segmentation models [9, 50], in this work, we explore 3D panoptic segmentation by leveraging multi-view images without explicit 3D input data." + }, + { + "type": "title", + "bbox": [ + 0.217, + 0.2, + 0.411, + 0.214 + ], + "angle": 0, + "content": "2.2 Multi-view Fusion" + }, + { + "type": "text", + "bbox": [ + 0.216, + 0.224, + 0.787, + 0.482 + ], + "angle": 0, + "content": "Recently, researches in 3D reconstruction have made tremendous progress [2,6, 20, 29, 31]. Beyond novel view synthesis, we can utilize 3D reconstruction techniques as a tool to fuse 2D information like semantics and features in the 3D space [5, 14, 17, 21, 25, 35, 43, 46, 48]. For example, Semantic-NeRF [49] learns a semantic field from the 2D semantic segmentation, demonstrating the robustness and effectiveness of the semantic fusion. Later works [21, 25, 43] propose to fuse the unsupervised 2D dense feature for various segmentation and editing applications, showcasing the potential to adapt the 2D zero-shot models to 3D domain. In this work, we study the task of panoptic fusion, integrating both instance and semantic information to obtain a comprehensive understanding of the 3D scene. This task is significantly more challenging than the previous semantic fusion task [49], since we need to incorporate additional instance fusion and achieve instance label consistency. Unlike recent methods (e.g., Panoptic-Lifting [40] and Contrastive Lift [3]), which utilize deterministic feature embeddings to represent instance information, we develop a novel probabilistic solution, actively considering the error-prone and noisy nature of the 2D segmentations, such that we can duly enhance the effectiveness and robustness of instance fusion." + }, + { + "type": "title", + "bbox": [ + 0.217, + 0.504, + 0.496, + 0.518 + ], + "angle": 0, + "content": "2.3 Probabilistic Representation" + }, + { + "type": "text", + "bbox": [ + 0.216, + 0.529, + 0.787, + 0.68 + ], + "angle": 0, + "content": "Probabilistic feature embedding is a popular tool employed in various tasks, e.g., image generation [22, 38], normal estimation [1], video understanding [36], point cloud understanding [4], and prototype embeddings for few-shot detection [42]. Motivated by its high capabilities in estimating aleatoric uncertainty [13] and addressing data noise, we innovate a new panoptic-lifting pipeline for 3D panoptic segmentation by formulating new modules based on probabilistic feature embeddings, including a reformulated contrastive loss with the probability product kernel [18] and an effective cross-view constraint to enhance the multi-view consistency. Further, we also design a novel probabilistic clustering algorithm to facilitate the generation of consistent panoptic segmentation." + }, + { + "type": "title", + "bbox": [ + 0.217, + 0.703, + 0.33, + 0.718 + ], + "angle": 0, + "content": "3 Method" + }, + { + "type": "text", + "bbox": [ + 0.216, + 0.735, + 0.785, + 0.794 + ], + "angle": 0, + "content": "Given a set of posed images \\(\\{\\mathcal{I}\\}\\) associated with 2D panoptic segmentation predictions (i.e., semantic masks \\(\\{\\mathcal{H}\\}\\) and instance masks \\(\\{\\mathcal{K}\\}\\)) generated by 2D segmentation models, our goal is to learn accurate 3D panoptic fields that can be rendered into consistent panoptic segmentation results over different views." + }, + { + "type": "text", + "bbox": [ + 0.216, + 0.796, + 0.785, + 0.84 + ], + "angle": 0, + "content": "The overview of PCF-Lift is illustrated in Fig. 2. Specifically, the 3D panoptic fields include a semantic field, an instance field, a density field, and a color field. Particularly, considering the significance of instance-related issues, we propose" + } + ], + [ + { + "type": "header", + "bbox": [ + 0.314, + 0.115, + 0.731, + 0.129 + ], + "angle": 0, + "content": "PCF-Lift: Panoptic Lifting by Probabilistic Contrastive Fusion" + }, + { + "type": "page_number", + "bbox": [ + 0.775, + 0.117, + 0.785, + 0.127 + ], + "angle": 0, + "content": "5" + }, + { + "type": "image", + "bbox": [ + 0.251, + 0.145, + 0.756, + 0.347 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.214, + 0.37, + 0.788, + 0.538 + ], + "angle": 0, + "content": "Fig. 2: Overview of PCF-Lift. The 3D panoptic fields include a semantic field, an instance field, a density field, and a color field. To solve the instance-related issues, we propose to learn probabilistic feature embeddings in the instance field (see Sec. 3.1). During the training phase, given two camera views, we can render the probabilistic feature maps from the instance field via volume rendering. To optimize the probabilistic instance field, we devise the probabilistic contrastive loss with Probability Product (PP) kernel [18], and propose a cross-view constraint to further enhance the feature consistency from different views (see Sec. 3.2). Similarly, we can render the semantic and color predictions, and adopt photometric loss and cross-entropy loss to optimize the semantic field, the density field, and the color field. During the inference phase, we design a novel multi-view object association (MVOA) algorithm for the generation of consistent panoptic segmentations (see Sec. 3.3)." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.567, + 0.788, + 0.765 + ], + "angle": 0, + "content": "to learn probabilistic feature embeddings in the instances field (see Sec. 3.1). During the training phase, we jointly train the whole panoptic fields. Given two camera views, we can render the probabilistic feature maps from the instance field via volume rendering [19]. To optimize the probabilistic feature embeddings, we develop the contrastive loss with the Probability Product kernel [18], and propose a cross-view constraint to further enhance the feature consistency from different views (see Sec. 3.2). Similarly, the semantic predictions and color predictions are also rendered via volume rendering, where the semantic field, the density field, and the color field are optimized by the photometric loss and cross-entropy loss [3,6,40,49]. In the inference phase, we design a probabilistic clustering (i.e., multi-view object association) algorithm to effectively identify the prototype features of underlying 3D object instances for the generation of consistent panoptic segmentation results across any given views (see Sec. 3.3)." + }, + { + "type": "title", + "bbox": [ + 0.215, + 0.786, + 0.542, + 0.801 + ], + "angle": 0, + "content": "3.1 Probabilistic Feature Embeddings" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.81, + 0.787, + 0.84 + ], + "angle": 0, + "content": "To provide a robust instance representation, we propose to learn probabilistic feature embeddings in the instance field, which maps each 3D point to a random" + } + ], + [ + { + "type": "page_number", + "bbox": [ + 0.218, + 0.116, + 0.23, + 0.127 + ], + "angle": 0, + "content": "6" + }, + { + "type": "header", + "bbox": [ + 0.272, + 0.115, + 0.468, + 0.129 + ], + "angle": 0, + "content": "R. Zhu, S. Qiu, Q. Wu, et al." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.147, + 0.788, + 0.314 + ], + "angle": 0, + "content": "variable \\(\\mathcal{F}\\), modeled as an \\(N\\)-dimensional multivariate Gaussian distribution \\(\\mathcal{F} \\sim \\mathcal{N}(\\mu, \\Sigma)\\). Here, \\(\\mu\\) is the mean vector, indicating the central feature values and \\(\\Sigma\\) is a diagonal covariance matrix \\(\\mathrm{diag}(\\sigma^2)\\) with \\(\\sigma^2 = (\\sigma^{(1)^2}, \\sigma^{(2)^2}, \\dots, \\sigma^{(N)^2})\\). Concretely, for any given query point \\(\\mathbf{x} \\in \\mathbb{R}^3\\), the instance field predicts \\((\\mu, \\sigma^2) \\in \\mathbb{R}^{2N}\\). Similar to the rendering of the color field, for each pixel in a given camera view, we can render its corresponding Gaussian distribution feature via volume rendering. A significant advantage of using probabilistic features rather than deterministic features is the ability of Gaussian distributions that assign different covariance values for uncertainty modeling. This property is crucial as it aids in reducing the impact of noise, thereby enhancing the robustness and accuracy of feature embeddings in representing complex 3D scenes." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.315, + 0.788, + 0.405 + ], + "angle": 0, + "content": "Intuitively, the similarity between two rendered Gaussian distributions indicates whether the corresponding pixels belong to the same instance. To quantify the similarities between different Gaussian distributions, we employ the Probability Product (PP) kernel [18] \\( K_{\\rho} \\). Specifically, given two Gaussian distributions \\( \\mathcal{F}_i \\sim \\mathcal{N}(\\mu_i,\\Sigma_i) \\) and \\( \\mathcal{F}_j \\sim \\mathcal{N}(\\mu_j,\\Sigma_j) \\), the corresponding kernel can be explicitly formulated as" + }, + { + "type": "equation", + "bbox": [ + 0.223, + 0.412, + 0.793, + 0.457 + ], + "angle": 0, + "content": "\\[\nK _ {\\rho} \\left(\\mathcal {F} _ {i}, \\mathcal {F} _ {j}\\right) = \\left(\\prod_ {d = 1} ^ {N} \\left(\\frac {\\sigma_ {i} ^ {(d) ^ {2}}}{\\sigma_ {j} ^ {(d) ^ {2}}} + \\frac {\\sigma_ {j} ^ {(d) ^ {2}}}{\\sigma_ {i} ^ {(d) ^ {2}}}\\right) / 2\\right) ^ {- \\frac {1}{2}} \\exp \\left(- \\sum_ {d = 1} ^ {N} \\left(\\mu_ {i} ^ {(d)} - \\mu_ {j} ^ {(d)}\\right) ^ {2} / 4 \\left(\\sigma_ {i} ^ {(d) ^ {2}} + \\sigma_ {j} ^ {(d) ^ {2}}\\right)\\right). \\tag {1}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.458, + 0.788, + 0.504 + ], + "angle": 0, + "content": "Note that the results produced by the PP kernel fall within the range of 0 (minimal similarity) to 1 (maximal similarity), the same as the Radial Basis Function (RBF) kernel used in the deterministic method [3]." + }, + { + "type": "title", + "bbox": [ + 0.215, + 0.524, + 0.618, + 0.54 + ], + "angle": 0, + "content": "3.2 Training Probabilistic Feature Embeddings" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.547, + 0.788, + 0.683 + ], + "angle": 0, + "content": "Probabilistic Contrastive Loss. During the training phase, we leverage the contrastive learning technique to optimize the probabilistic feature embeddings. Specifically, the contrastive loss used in [7, 27, 34, 45] tends to increase the predicted feature similarities of predefined positive pairs, while decreasing the similarities of negative pairs for training. In our work, a positive pair is defined as two pixels belonging to the same instance, while a negative pair denotes two pixels of different instances within a single image, following the previous method [3]. Since we basically learn probabilistic features, the PP kernel is exploited to measure the features similarity. In general, the loss can be formulated as" + }, + { + "type": "equation", + "bbox": [ + 0.286, + 0.692, + 0.786, + 0.73 + ], + "angle": 0, + "content": "\\[\n\\mathcal {L} _ {\\mathrm {p i x e l - c o n t r a}} = - \\frac {1}{| \\varOmega |} \\sum_ {u \\in \\varOmega} \\log \\frac {\\sum_ {u ^ {\\prime} \\in \\varOmega} \\mathbf {1} _ {(u , u ^ {\\prime})} \\exp \\left(K _ {\\rho} \\left(\\mathcal {F} _ {u} , \\mathcal {F} _ {u ^ {\\prime}}\\right)\\right)}{\\sum_ {u ^ {\\prime} \\in \\varOmega} \\exp \\left(K _ {\\rho} \\left(\\mathcal {F} _ {u} , \\mathcal {F} _ {u ^ {\\prime}}\\right)\\right)}, \\tag {2}\n\\]" + }, + { + "type": "equation", + "bbox": [ + 0.303, + 0.741, + 0.786, + 0.779 + ], + "angle": 0, + "content": "\\[\n\\mathcal {L} _ {\\text {c o n c e n}} = - \\frac {1}{| \\varOmega |} \\sum_ {u \\in \\varOmega} \\log \\left(K _ {\\rho} \\left(\\mathcal {F} _ {u}, \\frac {\\sum_ {u ^ {\\prime} \\in \\varOmega} \\mathbf {1} _ {(u , u ^ {\\prime})} \\mathcal {F} _ {u ^ {\\prime}}}{\\sum_ {u ^ {\\prime} \\in \\varOmega} \\mathbf {1} _ {(u , u ^ {\\prime})}}\\right)\\right), \\tag {3}\n\\]" + }, + { + "type": "equation", + "bbox": [ + 0.374, + 0.79, + 0.786, + 0.805 + ], + "angle": 0, + "content": "\\[\n\\text {a n d} \\mathcal {L} _ {\\text {c o n t r a}} = \\mathcal {L} _ {\\text {p i x e l - c o n t r a}} + \\mathcal {L} _ {\\text {c o n c e n}}, \\tag {4}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.81, + 0.788, + 0.841 + ], + "angle": 0, + "content": "where \\(\\mathcal{L}_{\\mathrm{pixel - contra}}\\) is the pixel-wise contrastive loss, \\(\\mathcal{L}_{\\mathrm{concen}}\\) is the concentrate loss term, \\(\\mathbf{1}\\) is the indicator function of positive pairs, \\(\\Omega\\) is the set of pixel sample, \\(\\mathcal{F}_u\\)" + } + ], + [ + { + "type": "header", + "bbox": [ + 0.314, + 0.114, + 0.733, + 0.129 + ], + "angle": 0, + "content": "PCF-Lift: Panoptic Lifting by Probabilistic Contrastive Fusion" + }, + { + "type": "page_number", + "bbox": [ + 0.775, + 0.116, + 0.787, + 0.127 + ], + "angle": 0, + "content": "7" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.147, + 0.788, + 0.254 + ], + "angle": 0, + "content": "and \\(\\mathcal{F}_{u'}\\) are the rendered features for pixels \\(u\\) and \\(u' \\in \\Omega\\) via volume rendering, respectively. Different from previous method [3], we devise probabilistic similarity kernels to calculate a more effective contrastive loss for model training. By optimizing the loss Eq. (4), we can learn a more expressive 3D instance field for robust feature representations, as further analyzed in Sec. 3.4. To avoid the neural network from generating large covariances everywhere, we use an additional regularization term, \\(\\mathcal{L}_{\\mathrm{reg}} := \\log \\left( \\prod_{d=1}^{N} \\sigma^{(d)^2} \\right)\\), to penalize the large covariances." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.274, + 0.789, + 0.396 + ], + "angle": 0, + "content": "Cross-view Constraint. To better train the probabilistic feature embeddings, we propose a cross-view constraint to enhance feature consistency for the same object across different views. Given the rendered learned probabilistic feature embedding sets \\(\\{\\mathcal{F}^m\\}\\) and \\(\\{\\mathcal{F}^n\\}\\) for two different views, we devise the PP kernel \\(K_{\\rho}\\) and a predefined threshold \\(\\tau\\) to collect positive pairs \\(\\mathcal{P}\\) that belong to the same object: \\(\\mathcal{P} = \\left\\{(\\mathcal{F}_r,\\mathcal{F}_s)\\mid K_\\rho (\\mathcal{F}_r,\\mathcal{F}_s) > \\tau ,\\mathcal{F}_r\\in \\{\\mathcal{F}^m\\} ,\\mathcal{F}_s\\in \\{\\mathcal{F}^n\\} \\right\\}\\). Moreover, as we expect to maximize the similarities between the features that belong to the same 3D object in different views, the cross-view constraint is defined as:" + }, + { + "type": "equation", + "bbox": [ + 0.355, + 0.407, + 0.787, + 0.445 + ], + "angle": 0, + "content": "\\[\n\\mathcal {L} _ {\\text {c r o s s}} = - \\frac {1}{| \\mathcal {P} |} \\sum_ {\\left(\\mathcal {F} _ {r}, \\mathcal {F} _ {s}\\right) \\in \\mathcal {P}} \\log \\left(K _ {\\rho} \\left(\\mathcal {F} _ {r}, \\mathcal {F} _ {s}\\right)\\right). \\tag {5}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.215, + 0.449, + 0.788, + 0.495 + ], + "angle": 0, + "content": "In practice, we use the cross-view constraint only in a later optimization stage (i.e., the last few epochs), where the feature space is sufficiently meaningful to provide reliable positive pairs. The overall loss is formulated as" + }, + { + "type": "equation", + "bbox": [ + 0.37, + 0.506, + 0.786, + 0.523 + ], + "angle": 0, + "content": "\\[\n\\mathcal {L} = \\mathcal {L} _ {\\text {c o n t r a}} + w _ {\\text {c r o s s}} \\mathcal {L} _ {\\text {c r o s s}} + w _ {\\text {r e g}} \\mathcal {L} _ {\\text {r e g}}, \\tag {6}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.215, + 0.532, + 0.589, + 0.549 + ], + "angle": 0, + "content": "where \\(w_{\\mathrm{cross}}\\) and \\(w_{\\mathrm{reg}}\\) are weight hyper-parameters." + }, + { + "type": "title", + "bbox": [ + 0.215, + 0.568, + 0.515, + 0.584 + ], + "angle": 0, + "content": "3.3 Multi-view Object Association" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.593, + 0.789, + 0.775 + ], + "angle": 0, + "content": "During the inference phase, to facilitate consistent panoptic segmentations, we introduce a novel clustering algorithm, named the multi-view object association (MVOA) algorithm. The algorithm aims to extract the prototype feature set from the learned instance field for the assignment of instance labels. In general, through volume rendering, we obtain a large-scale rendered per-pixel probabilistic feature set \\(\\{\\mathcal{F}\\}\\) of the training views, as the input of our algorithm. Considering the efficiency issue, we first conduct an instance grouping operation to gather a smaller feature set \\(\\mathcal{C}\\), with the help of inconsistent instance segmentation masks \\(\\{\\mathcal{K}\\}\\) obtained from training views. Then, we construct the probabilistic similarity graph based on \\(\\mathcal{C}\\) and design a multi-view matching process to collect the prototype feature set \\(\\mathcal{D}\\). The details of instance grouping and multi-view matching operations are introduced as follows." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.795, + 0.789, + 0.843 + ], + "angle": 0, + "content": "Instance Grouping. We group the feature set \\(\\{\\mathcal{F}\\}\\) from each individual view. For the view of \\(l\\)-th image, we collect the pixels that belong to instance \\(p\\) into the set \\(\\{i : \\mathcal{K}_l^i = p\\}\\), and the corresponding probabilistic feature set is denoted as" + } + ], + [ + { + "type": "page_number", + "bbox": [ + 0.218, + 0.116, + 0.23, + 0.127 + ], + "angle": 0, + "content": "8" + }, + { + "type": "header", + "bbox": [ + 0.272, + 0.115, + 0.467, + 0.129 + ], + "angle": 0, + "content": "R. Zhu, S. Qiu, Q. Wu, et al." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.146, + 0.788, + 0.273 + ], + "angle": 0, + "content": "\\(\\{\\mathcal{F}_l^i:\\mathcal{K}_l^i = p\\}\\). Then, we group \\(\\{\\mathcal{F}_l^i:\\mathcal{K}_l^i = p\\}\\) to a single probabilistic feature by calculating the average: \\(\\mathcal{C}_l^p = \\sum_i\\mathcal{F}_l^i /|\\{i:\\mathcal{K}_l^i = p\\} |.\\) Concurrently, we model the score quantity \\(\\mathcal{S}_l^p\\) for \\(\\mathcal{C}_l^p\\) as an indicator of feature concentration, which aids in the subsequent multi-view matching process. The indicator function \\(\\varPhi(\\mathcal{C}_l^p)\\) averages the PP kernel similarities between \\(\\mathcal{C}_l^p\\) and all features in \\(\\{\\mathcal{F}_l^i:\\mathcal{K}_l^i = p\\}\\). Mathematically, \\(\\varPhi(\\mathcal{C}_l^p)\\) follows: \\(\\mathcal{S}_l^p = \\varPhi(\\mathcal{C}_l^p) = \\sum_{\\{i:\\mathcal{K}_l^i = p\\}}K_\\rho (\\mathcal{C}_l^p,\\mathcal{F}_l^i) / |\\{i:\\mathcal{K}_l^i = p\\} |.\\) By applying this across all observed images, we obtain the grouped feature set \\(\\mathcal{C}\\) and the corresponding score set \\(\\mathcal{S}\\)." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.292, + 0.79, + 0.445 + ], + "angle": 0, + "content": "Multi-view Matching. To extract the structural relationships among the features, we construct an unoriented similarity graph \\( G = (\\mathcal{C}, E) \\), where the grouped feature set \\( \\mathcal{C} \\) represents the nodes and \\( E \\) encompasses the edges that indicate quantitative feature similarities measured by the PP Kernel: \\( E_{} = E_{} = K_{\\rho}(\\mathcal{C}_g, \\mathcal{C}_h) \\), where \\( \\mathcal{C}_g, \\mathcal{C}_h \\in \\mathcal{C} \\). Then, we extract the prototype feature set \\( \\mathcal{D} \\) from the graph \\( G \\) and score set \\( S \\) by a greedy procedure, which is akin to the classical greedy algorithm known as non-maximum suppression (NMS). The pseudo-code of our proposed MVOA algorithm is presented in Algo. 1. In practice, the selection of hyper-parameter \\( \\mathcal{T} \\) is based on the average similarity computed across grouped feature pairs identified within each view." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.465, + 0.79, + 0.618 + ], + "angle": 0, + "content": "Generating Panoptic Segmentation Masks. We first follow the MVOA algorithm to collect the prototype feature set \\(\\mathcal{D}\\). For each view, including the novel view, we then generate the semantic labels from the learned semantic field to differentiate between the background and foreground. Coupled with a feature map rendered from the probabilistic instance field, for each foreground pixel, we finally determine its corresponding instance label by assigning the candidate index from the set \\(\\mathcal{D}\\) that exhibits the highest similarity to the rendered feature. In practice, the MVOA algorithm needs to be conducted only once, while the extracted prototype features \\(\\mathcal{D}\\) will be used for generating the panoptic segmentation masks of all test views." + }, + { + "type": "title", + "bbox": [ + 0.215, + 0.638, + 0.433, + 0.654 + ], + "angle": 0, + "content": "3.4 Theoretical Analysis" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.663, + 0.789, + 0.709 + ], + "angle": 0, + "content": "Corollary 1 If the covariances of the given Gaussian distributions are isotropic and fixed, i.e., \\(\\Sigma_{i} = \\Sigma_{j} = \\sigma I\\), where \\(\\sigma\\) is a constant scalar, the probability product kernel can be simplified to an RBF kernel." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.72, + 0.789, + 0.843 + ], + "angle": 0, + "content": "The primary distinction between our probabilistic method and the prior deterministic method [3] is the choice of a similarity kernel. Specifically, our probabilistic method exploits the PP kernel, which offers higher flexibility and stronger expressive capability by adjusting the Gaussian mean and covariance compared to the RBF kernel employed in the deterministic method, as demonstrated in Fig 3. The RBF kernel can be considered as a degenerate PP kernel when applied to a Gaussian distribution with an isotropic and fixed covariance, as stated in Corollary 1. From this perspective, the proposed probabilistic method is a" + } + ], + [ + { + "type": "header", + "bbox": [ + 0.314, + 0.115, + 0.731, + 0.129 + ], + "angle": 0, + "content": "PCF-Lift: Panoptic Lifting by Probabilistic Contrastive Fusion" + }, + { + "type": "page_number", + "bbox": [ + 0.775, + 0.117, + 0.785, + 0.127 + ], + "angle": 0, + "content": "9" + }, + { + "type": "code_caption", + "bbox": [ + 0.23, + 0.151, + 0.696, + 0.166 + ], + "angle": 0, + "content": "Algorithm 1: Multi-view object association algorithm (MVOA)" + }, + { + "type": "algorithm", + "bbox": [ + 0.23, + 0.169, + 0.736, + 0.365 + ], + "angle": 0, + "content": "Data: Inconsistent instance mask \\(\\{\\mathcal{K}\\}\\), input feature set \\(\\{\\mathcal{F}\\}\\), Threshold T \nResult: Prototype features set \\(\\mathcal{D}\\) \n// Instance Grouping // Multi-view matching \n2 \\(\\mathcal{C} = \\{\\}\\), \\(\\mathcal{S} = \\{\\}\\) 10 \\(\\mathcal{D} \\gets \\{\\}\\) \n3 for \\(\\mathcal{K}_l\\) in \\(\\{\\mathcal{K}\\}\\) do 11 while \\(\\mathcal{S} \\neq empty\\) do \n4 IDs = {unique_ID(\\mathcal{K}_l)} 12 m ← argmax S \n5 for \\(p\\) in \\(IDs\\) do D ← D ∪ \\(\\mathcal{C}_m\\) \n6 \\(\\mathcal{C}_l^p = \\sum_i\\mathcal{F}_l^i /|\\{i:\\mathcal{K}_l^i = p\\}|\\) C ← C - Cm; S ← S - Sm \n7 \\(\\mathcal{S}_l^p = \\Phi (\\mathcal{C}_l^p)\\) 13 if Kρ(Cm, Ci) ≥ T then 14 | C ← C - Ci; S ← S - Si \n8 end 16 end \n9 end 17 end" + }, + { + "type": "image", + "bbox": [ + 0.248, + 0.384, + 0.753, + 0.492 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.214, + 0.515, + 0.788, + 0.6 + ], + "angle": 0, + "content": "Fig. 3: (a) Flexibility of adjusting covariances. Contour plot of the PP kernel similarity for two Gaussians with different covariance values \\((\\sigma_{1}\\) and \\(\\sigma_{2})\\) and fixed mean values in a 1-dimensional case. (b) Anisotropy. Contour plot of the PP kernel similarity for two Gaussians with different Gaussian mean offsets \\((d_{x}\\) and \\(d_{y})\\) and fixed covariances in a 2-dimensional case. (c) Isotropy. Contour plot of the RBF kernel similarity for two deterministic features with different offsets \\((d_{x}\\) and \\(d_{y})\\) in a 2-dimensional case." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.628, + 0.788, + 0.675 + ], + "angle": 0, + "content": "more general framework, while the deterministic method is a subclass within the broader probabilistic paradigm. The experimental results also verify the effectiveness of our probabilistic solution in the ablation studies of Sec. 4.3." + }, + { + "type": "title", + "bbox": [ + 0.215, + 0.695, + 0.367, + 0.713 + ], + "angle": 0, + "content": "4 Experiment" + }, + { + "type": "title", + "bbox": [ + 0.215, + 0.726, + 0.448, + 0.743 + ], + "angle": 0, + "content": "4.1 Experimental Settings" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.75, + 0.788, + 0.841 + ], + "angle": 0, + "content": "Implementation Details. For fair comparison, we adopt the same architecture of TensoRF [6] together with the same layer parameters used in previous works [3,40]. Particularly, our instance field is constructed following the slow-fast architecture of Contrastive Lift [3], where a shallow 5-layer MLP is applied to predict an probabilistic feature embedding for a 3D coordinate input. In practice, the dimension \\(N\\) for the probabilistic feature embedding is 3 in our experiments." + } + ], + [ + { + "type": "page_number", + "bbox": [ + 0.218, + 0.116, + 0.236, + 0.127 + ], + "angle": 0, + "content": "10" + }, + { + "type": "header", + "bbox": [ + 0.272, + 0.115, + 0.467, + 0.129 + ], + "angle": 0, + "content": "R. Zhu, S. Qiu, Q. Wu, et al." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.147, + 0.788, + 0.269 + ], + "angle": 0, + "content": "For the training of the instance field, we sample rays from two different views in the last few epochs for optimization (Sec. 3.2), using the cross-view constraint \\(\\mathcal{L}_{\\mathrm{cross}}\\) with a predefined threshold value \\(\\tau = 0.9\\) and weight \\(w_{\\mathrm{cross}} = 0.05\\); while in other training epochs, we set the weight \\(w_{\\mathrm{cross}}\\) to 0 and only sample the rays from single views. Besides, \\(w_{\\mathrm{reg}}\\) in Eq. (6) is set to 0.001 throughout the whole training process. For the training of the color, density and semantic fields, we adopt the same loss terms and training strategies from previous works [3, 40]. More implementation details are provided in the supplementary material." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.288, + 0.79, + 0.442 + ], + "angle": 0, + "content": "Metrics. Since our method is proposed to solve the panoptic-lifting task, we employ scene-level Panoptic Quality (PQ\\(^{\\text{scene}}\\)) for evaluation, which was introduced in Panoptic Lifting [40] and widely used in the related works [3,40]. Unlike the standard PQ metric [23], this metric particularly considers the consistency of instance IDs across multiple views. By merging predictions and ground truths with consistent instance IDs into subsets for PQ\\(^{\\text{scene}}\\)-based evaluation, we determine a match of subset pair when the intersection over union (IoU) exceeds 0.5, in line with prior baselines [3,40]. Since PQ\\(^{\\text{scene}}\\) is a product of scene-level segmentation quality (SQ\\(^{\\text{scene}}\\)) and recognition quality (RQ\\(^{\\text{scene}}\\)), we also provide the SQ\\(^{\\text{scene}}\\) and RQ\\(^{\\text{scene}}\\) metrics for more detailed comparisons." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.459, + 0.788, + 0.537 + ], + "angle": 0, + "content": "Baselines. We mainly compare the proposed method with the current state-of-the-art methods that target lifting 2D panoptic predictions to 3D, i.e., Contrastive Lift [3] and Panoptic Lifting [40]. Moreover, we compare our method with the recent NeRF-based 3D panoptic segmentation approaches, i.e., Panoptic Neural Fields [26] (PNF) and DM-NeRF [44]." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.556, + 0.788, + 0.679 + ], + "angle": 0, + "content": "Datasets. Following the previous works [3,40], we conduct experiments on two public datasets, the ScanNet dataset [12] and the Messy Room dataset [3], for both quantitative and qualitative evaluations. For fair comparisons on ScanNet, following the other state-of-the-art approaches [3,40], we adopt Mask2Former (M2F) [9] to generate 2D panoptic segmentation predictions, coupled with the same protocol in [40] that maps the COCO [28] vocabulary to 21 classes. For the experiments on Messy Rooms, we use the LVIS [16] vocabulary and the Detic [50] 2D panoptic segmentations, which are also utilized in Contrastive Lift [3]." + }, + { + "type": "title", + "bbox": [ + 0.215, + 0.697, + 0.418, + 0.712 + ], + "angle": 0, + "content": "4.2 Main Experiments" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.72, + 0.788, + 0.841 + ], + "angle": 0, + "content": "ScanNet Dataset. To verify the performance on real data, we conduct experiments on the ScanNet dataset [12] of 12 scenes. Quantitative comparisons provided in Tab. 1 demonstrate that our method consistently outperforms previous baselines, including both the 3D panoptic segmentation approaches [26,44] and the recent state-of-the-art methods for lifting 2D panoptic segmentation [3,40]. Moreover, visual comparisons between our method and the latest state-of-the-art method [3] are presented in Fig. 4, which exhibits our method's capability of achieving a consistent and accurate 3D panoptic segmentation." + } + ], + [ + { + "type": "header", + "bbox": [ + 0.314, + 0.114, + 0.732, + 0.129 + ], + "angle": 0, + "content": "PCF-Lift: Panoptic Lifting by Probabilistic Contrastive Fusion" + }, + { + "type": "page_number", + "bbox": [ + 0.768, + 0.116, + 0.784, + 0.127 + ], + "angle": 0, + "content": "11" + }, + { + "type": "table_caption", + "bbox": [ + 0.214, + 0.144, + 0.788, + 0.215 + ], + "angle": 0, + "content": "Table 1: Results on the ScanNet dataset. We report the PQscene, SQscene, and RQscene metrics. Since Contrastive Lift [3] does not report the performance using the SQscene and RQscene metrics, we apply the officially-released pre-trained model and re-run the clustering algorithm to obtain the values reported in this table. For the other metric values, we directly report the ones in previous papers [3,40]." + }, + { + "type": "table", + "bbox": [ + 0.246, + 0.227, + 0.754, + 0.309 + ], + "angle": 0, + "content": "
MethodVenueType\\(SQ^{scene}(\\%)\\)\\(RQ^{scene}(\\%)\\)\\(PQ^{scene}(\\%)\\)
DM-NeRF [44]ICLR'233D panoptic segmentation53.346.141.7
PNF [26]CVPR'223D panoptic segmentation63.050.748.3
PNF [26] + GT BBoxesCVPR'223D panoptic segmentation70.055.954.3
Panoptic Lifting [40]CVPR'232D panoptic Lifting73.565.058.9
Contrastive Lift [3]NeurIPS'232D panoptic Lifting75.763.662.0
Ours-2D panoptic Lifting78.565.463.5
" + }, + { + "type": "table_caption", + "bbox": [ + 0.214, + 0.321, + 0.788, + 0.364 + ], + "angle": 0, + "content": "Table 2: Results on the Messy Rooms dataset [3]. Following [3], the PQscene metric is reported on both \"old room\" and \"large corridor\" environments with an increasing number of objects in the scene (25, 50, 100, 500)." + }, + { + "type": "table", + "bbox": [ + 0.246, + 0.376, + 0.754, + 0.445 + ], + "angle": 0, + "content": "
Method/ NumberOld Room Environment (%)Large Corridor Environment(%)Mean(%)
25501005002550100500
Panoptic Lifting [40]73.269.964.351.065.571.061.849.063.2
Contrastive Lift [3]78.975.869.155.076.575.568.752.569.0
Ours80.978.374.860.381.079.474.058.873.4
" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.472, + 0.788, + 0.669 + ], + "angle": 0, + "content": "Messy Rooms Dataset. We also conduct experiments on the challenging Messy Room dataset [3], which is provided by Contrastive Lift [3], containing up to 500 objects in each scene. We present the quantitative results in Tab. 2, showing that our method (73.4%) achieves significant improvements in terms of the mean PQ\\(^{\\text{scene}}\\) results compared to the current state-of-the-art methods of Contrastive Lift [3] (69.0%) and Panoptic Lifting [40] (63.2%). The overall performance highlights the advantages of our probabilistic approach over the previous deterministic methods, particularly in segmenting complex scenes with hundreds of objects. Further, our method achieves a mean SQ\\(^{\\text{scene}}\\) of 82.2% and a mean RQ\\(^{\\text{scene}}\\) of 86.9%, surpassing the SQ\\(^{\\text{scene}}\\) of 77.7% and RQ\\(^{\\text{scene}}\\) of 86.6% obtained by Contrastive Lift [3]. Moreover, the visual comparisons in Fig. 4 further indicate that our method can accurately segment the small objects in the red frames, whereas Contrastive Lift [3] struggles to distinguish such instances." + }, + { + "type": "title", + "bbox": [ + 0.215, + 0.692, + 0.39, + 0.709 + ], + "angle": 0, + "content": "4.3 Ablation Study" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.719, + 0.788, + 0.843 + ], + "angle": 0, + "content": "Effectiveness of Each Component. We conduct an ablation study on the Messy Room dataset [3], which contains more challenging scenes with hundreds of instances. As shown in Tab. 3 (b) and (d), the proposed probabilistic feature embeddings greatly benefit our method, while replacing it with deterministic one leads to significant performance drop. Furthermore, the effectiveness of our proposed cross-view constraint is verified in Tab. 3 (e) and (f). For the proposed multi-view object association (MVOA) algorithm, the results in Tab. 3 (c), (d), (e), and (f) show that it particularly benefits the clustering of our proposed" + } + ], + [ + { + "type": "page_number", + "bbox": [ + 0.218, + 0.116, + 0.236, + 0.127 + ], + "angle": 0, + "content": "12" + }, + { + "type": "header", + "bbox": [ + 0.272, + 0.115, + 0.467, + 0.129 + ], + "angle": 0, + "content": "R. Zhu, S. Qiu, Q. Wu, et al." + }, + { + "type": "image", + "bbox": [ + 0.307, + 0.145, + 0.703, + 0.558 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.214, + 0.583, + 0.788, + 0.613 + ], + "angle": 0, + "content": "Fig. 4: Visual comparison of the latest state-of-the-art method Contrastive Lift [3] and our method on the ScanNet [12] dataset and the Messy Room [3] dataset." + }, + { + "type": "table_caption", + "bbox": [ + 0.214, + 0.65, + 0.788, + 0.679 + ], + "angle": 0, + "content": "Table 3: Ablation study on the Messy Room dataset [3]. The model (a) corresponds to Contrastive Lift [3] and the model (f) corresponds to our full method (PCF)." + }, + { + "type": "table", + "bbox": [ + 0.274, + 0.69, + 0.725, + 0.758 + ], + "angle": 0, + "content": "
Model Feature spaceClustering\\( {\\mathrm{{SQ}}}^{\\text{scene }}\\left( \\% \\right) \\)\\( {\\mathrm{{RQ}}}^{\\text{scene }}\\left( \\% \\right) \\)\\( {\\mathrm{{PQ}}}^{\\text{scene }}\\left( \\% \\right) \\)
(a)Deterministic [3]HDBSCAN [30]77.786.669.0
(b)Deterministic [3]MVOA79.386.270.4
(c)Learned Gaussian distributionHDBSCAN [30]78.086.669.6
(d)Learned Gaussian distributionMVOA81.386.872.3
(e)Learned Gaussian distribution (+ Cross-view constraint)HDBSCAN [30]78.886.970.4
(f)Learned Gaussian distribution (+ Cross-view constraint)MVOA82.286.973.4
" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.81, + 0.787, + 0.842 + ], + "angle": 0, + "content": "probabilistic representation. Also, MVOA can be effectively employed to the deterministic method [3] for a performance boost, as Tab. 3 (a) and (b) show." + } + ], + [ + { + "type": "header", + "bbox": [ + 0.314, + 0.115, + 0.731, + 0.129 + ], + "angle": 0, + "content": "PCF-Lift: Panoptic Lifting by Probabilistic Contrastive Fusion" + }, + { + "type": "page_number", + "bbox": [ + 0.769, + 0.116, + 0.785, + 0.127 + ], + "angle": 0, + "content": "13" + }, + { + "type": "image", + "bbox": [ + 0.283, + 0.142, + 0.72, + 0.251 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.214, + 0.271, + 0.788, + 0.357 + ], + "angle": 0, + "content": "Fig. 5: The visualization results of learned covariance components and the statistical results of covariances in two scenes of the Messy Room dataset [3]. For the histograms, the horizontal axis denotes the range of covariance magnitudes, while the vertical axis corresponds to the frequency statistics for those magnitudes. We calculate and plot the covariance magnitudes \\((\\sigma^{(1)^2}*\\sigma^{(2)^2}*\\sigma^{(3)^2})\\) of two distance image regions, the boundary areas and the internal areas of object instances, across all observed views." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.385, + 0.788, + 0.552 + ], + "angle": 0, + "content": "Uncertainty Analysis. To verify whether our method could provide meaningful modeling for uncertainty, we provide the rendered uncertainty maps on two scenes of the Messy Room dataset [3] in Fig. 5. The figure clearly illustrates that the regions of high covariance are mainly located near the boundaries of the object instances, due to the inconsistent segmentation issue that leaves severe ambiguity around the instance boundaries. Moreover, we statistically analyze the learned covariances in two distinct image regions, namely the boundary areas and the internal areas of object instances, across all observed views. These results further demonstrate our method's ability to model high uncertainty in areas where the instance boundaries are ambiguous. We provide more details in the supplementary material." + }, + { + "type": "title", + "bbox": [ + 0.216, + 0.571, + 0.466, + 0.586 + ], + "angle": 0, + "content": "4.4 Robustness Experiments" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.594, + 0.788, + 0.731 + ], + "angle": 0, + "content": "2D Backbones. In practice, the quality of panoptic segmentation generated by different 2D models may vary a lot. To study the robustness of our probabilistic method and the deterministic method [3] when incorporating with different 2D models, we select four different models from the model zoo of Detic [3] and utilize the LVIS [16] vocabulary to generate the 2D panoptic segmentation predictions for lifting, as shown in Fig. 6 (a). The quantitative results are presented in Fig. 6 (b), where our probabilistic methods consistently outperform the baseline method by a large margin. Particularly, we observe that the proposed MVOA algorithm can consistently boost the performance of a deterministic method [3]." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.75, + 0.788, + 0.841 + ], + "angle": 0, + "content": "Hand-crafted Noise. We study the robustness by adding hand-crafted noise. Specifically, for all given panoptic segmentation masks, we randomly select 100 pixels as anchors. Then, for each anchor, we randomly choose a pixel within a \\( W * W \\) window centered on the anchor and assign its instance ID to all pixels within this window, to simulate inaccurate segmentation predictions around object boundaries. As shown in Fig. 6 (c), as W gradually increases, the instance" + } + ], + [ + { + "type": "page_number", + "bbox": [ + 0.218, + 0.116, + 0.236, + 0.127 + ], + "angle": 0, + "content": "14" + }, + { + "type": "header", + "bbox": [ + 0.272, + 0.115, + 0.467, + 0.129 + ], + "angle": 0, + "content": "R. Zhu, S. Qiu, Q. Wu, et al." + }, + { + "type": "image", + "bbox": [ + 0.312, + 0.143, + 0.693, + 0.376 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.215, + 0.399, + 0.788, + 0.484 + ], + "angle": 0, + "content": "Fig. 6: Quantitative comparisons when using different 2D models or adding various levels of hand-crafted noise. We compare our probabilistic method (\"Probabilistic + MVOA\" in red) with the deterministic method [3] (\"Deterministic\" in dark blue) and a variant of deterministic method using our proposed MVOA algorithm (\"Deterministic + MVOA\" in light blue). Besides, the four tested models are from the model zoo of Detic \\([50]^1\\), and the two tested scenes are from the Messy Room dataset [3]." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.518, + 0.785, + 0.565 + ], + "angle": 0, + "content": "boundaries are continuously blurred. Our probabilistic approach consistently achieves the best performance as presented in Fig. 6 (d) and the results demonstrate that the proposed MVOA algorithm benefits the deterministic method [3]." + }, + { + "type": "title", + "bbox": [ + 0.216, + 0.614, + 0.36, + 0.63 + ], + "angle": 0, + "content": "5 Conclusion" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.658, + 0.788, + 0.78 + ], + "angle": 0, + "content": "We present PCF-Lift for the panoptic lifting task. First, we propose to learn the probabilistic feature embeddings through a multivariate Gaussian distribution for instance representation. For training, we reformulate the contrastive loss with the Probability Product kernel and propose a novel cross-view constraint to enhance the feature consistency across different views. During the inference phase, we propose a novel multi-view object association algorithm to effectively identify the prototype features representing underlying 3D object instances. We verify the effectiveness and robustness of PCF-Lift by extensive experiments." + }, + { + "type": "page_footnote", + "bbox": [ + 0.218, + 0.797, + 0.788, + 0.84 + ], + "angle": 0, + "content": "1 We use the official pre-trained models provided in https://github.com/facebookresearch/Detic/tree/main. Please refer to our supplementary material for the details of the four models." + } + ], + [ + { + "type": "header", + "bbox": [ + 0.314, + 0.115, + 0.732, + 0.129 + ], + "angle": 0, + "content": "PCF-Lift: Panoptic Lifting by Probabilistic Contrastive Fusion" + }, + { + "type": "page_number", + "bbox": [ + 0.768, + 0.117, + 0.786, + 0.127 + ], + "angle": 0, + "content": "15" + }, + { + "type": "title", + "bbox": [ + 0.218, + 0.145, + 0.403, + 0.163 + ], + "angle": 0, + "content": "Acknowledgements" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.177, + 0.788, + 0.237 + ], + "angle": 0, + "content": "This work is supported by the InnoHK of the Government of Hong Kong via the Hong Kong Centre for Logistics Robotics, the Shenzhen Portion of Shenzhen-Hong Kong Science and Technology Innovation Cooperation Zone under HZQB-KCZYB-20200089, and the CUHK T Stone Robotics Institute." + }, + { + "type": "title", + "bbox": [ + 0.218, + 0.259, + 0.323, + 0.275 + ], + "angle": 0, + "content": "References" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.289, + 0.785, + 0.331 + ], + "angle": 0, + "content": "1. Bae, G., Budvytis, I., Cipolla, R.: Estimating and exploiting the aleatoric uncertainty in surface normal estimation. In: Proceedings of the IEEE/CVF International Conference on Computer Vision. pp. 13137-13146 (2021)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.332, + 0.785, + 0.386 + ], + "angle": 0, + "content": "2. Barron, J.T., Mildenhall, B., Tancik, M., Hedman, P., Martin-Brualla, R., Srinivasan, P.P.: Mip-NeRF: A multiscale representation for anti-aliasing neural radiance fields. In: Proceedings of the IEEE/CVF International Conference on Computer Vision. pp. 5855-5864 (2021)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.387, + 0.785, + 0.428 + ], + "angle": 0, + "content": "3. Bhalgat, Y., Laina, I., Henriques, J.F., Zisserman, A., Vedaldi, A.: Contrastive Lift: 3D object instance segmentation by slow-fast contrastive fusion. arXiv preprint arXiv:2306.04633 (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.429, + 0.785, + 0.468 + ], + "angle": 0, + "content": "4. Cai, K., Lu, C.X., Huang, X.: Uncertainty estimation for 3D dense prediction via cross-point embeddings. IEEE Robotics and Automation Letters 8(5), 2558-2565 (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.469, + 0.785, + 0.51 + ], + "angle": 0, + "content": "5. Cen, J., Zhou, Z., Fang, J., Shen, W., Xie, L., Jiang, D., Zhang, X., Tian, Q., et al.: Segment anything in 3D with NeRFs. Advances in Neural Information Processing Systems 36 (2024)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.511, + 0.785, + 0.538 + ], + "angle": 0, + "content": "6. Chen, A., Xu, Z., Geiger, A., Yu, J., Su, H.: Tensorf: Tensorial radiance fields. In: European Conference on Computer Vision. pp. 333-350. Springer (2022)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.539, + 0.785, + 0.579 + ], + "angle": 0, + "content": "7. Chen, T., Kornblith, S., Norouzi, M., Hinton, G.: A simple framework for contrastive learning of visual representations. In: International conference on machine learning. pp. 1597-1607. PMLR (2020)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.58, + 0.785, + 0.634 + ], + "angle": 0, + "content": "8. Cheng, B., Collins, M.D., Zhu, Y., Liu, T., Huang, T.S., Adam, H., Chen, L.C.: Panoptic-deeplab: A simple, strong, and fast baseline for bottom-up panoptic segmentation. In: Proceedings of the IEEE/CVF conference on computer vision and pattern recognition. pp. 12475-12485 (2020)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.635, + 0.785, + 0.689 + ], + "angle": 0, + "content": "9. Cheng, B., Misra, I., Schwing, A.G., Kirillov, A., Girdhar, R.: Masked-attention mask transformer for universal image segmentation. In: Proceedings of the IEEE/CVF conference on computer vision and pattern recognition. pp. 1290-1299 (2022)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.69, + 0.785, + 0.731 + ], + "angle": 0, + "content": "0. Cheng, B., Schwing, A., Kirillov, A.: Per-pixel classification is not all you need for semantic segmentation. Advances in Neural Information Processing Systems 34, 17864-17875 (2021)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.731, + 0.785, + 0.771 + ], + "angle": 0, + "content": "1. Dahnert, M., Hou, J., Nießner, M., Dai, A.: Panoptic 3D scene reconstruction from a single rgb image. Advances in Neural Information Processing Systems 34, 8282-8293 (2021)" + }, + { + "type": "ref_text", + "bbox": [ + 0.22, + 0.772, + 0.785, + 0.813 + ], + "angle": 0, + "content": "12. Dai, A., Chang, A.X., Savva, M., Halber, M., Funkhouser, T., Nießner, M.: Scannet: Richly-annotated 3D reconstructions of indoor scenes. In: Proceedings of the IEEE conference on computer vision and pattern recognition. pp. 5828-5839 (2017)" + }, + { + "type": "ref_text", + "bbox": [ + 0.22, + 0.814, + 0.785, + 0.84 + ], + "angle": 0, + "content": "13. Der Kiureghian, A., Ditlevsen, O.: Aleatory or epistemic? does it matter? Structural safety 31(2), 105-112 (2009)" + }, + { + "type": "list", + "bbox": [ + 0.22, + 0.289, + 0.785, + 0.84 + ], + "angle": 0, + "content": null + } + ], + [ + { + "type": "page_number", + "bbox": [ + 0.218, + 0.116, + 0.236, + 0.127 + ], + "angle": 0, + "content": "16" + }, + { + "type": "header", + "bbox": [ + 0.272, + 0.115, + 0.467, + 0.129 + ], + "angle": 0, + "content": "R. Zhu, S. Qiu, Q. Wu, et al." + }, + { + "type": "ref_text", + "bbox": [ + 0.218, + 0.147, + 0.785, + 0.189 + ], + "angle": 0, + "content": "14. Fan, Z., Wang, P., Jiang, Y., Gong, X., Xu, D., Wang, Z.: NeRF-SOS: Any-iew self-supervised object segmentation on complex scenes. arXiv preprint arXiv:2209.08776 (2022)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.19, + 0.787, + 0.233 + ], + "angle": 0, + "content": "15. Gasperini, S., Mahani, M.A.N., Marcos-Ramiro, A., Navab, N., Tombari, F.: Panostter: End-to-end panoptic segmentation of LiDAR point clouds. IEEE Robotics and Automation Letters 6(2), 3216-3223 (2021)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.234, + 0.785, + 0.275 + ], + "angle": 0, + "content": "16. Gupta, A., Dollar, P., Girshick, R.: LVIS: A dataset for large vocabulary instance segmentation. In: Proceedings of the IEEE/CVF conference on computer vision and pattern recognition. pp. 5356-5364 (2019)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.276, + 0.785, + 0.317 + ], + "angle": 0, + "content": "17. Hamilton, M., Zhang, Z., Hariharan, B., Snavely, N., Freeman, W.T.: Unsupervised semantic segmentation by distilling feature correspondences. arXiv preprint arXiv:2203.08414 (2022)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.318, + 0.787, + 0.346 + ], + "angle": 0, + "content": "18. Jebara, T., Kondor, R., Howard, A.: Probability product kernels. The Journal of Machine Learning Research 5, 819-844 (2004)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.346, + 0.785, + 0.374 + ], + "angle": 0, + "content": "19. Kajiya, J.T., Von Herzen, B.P.: Ray tracing volume densities. ACM SIGGRAPH computer graphics 18(3), 165-174 (1984)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.375, + 0.785, + 0.403 + ], + "angle": 0, + "content": "20. Kerbl, B., Kopanas, G., Leimkuhler, T., Drettakis, G.: 3D gaussian splatting for real-time radiance field rendering. ACM Transactions on Graphics 42(4) (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.403, + 0.785, + 0.445 + ], + "angle": 0, + "content": "21. Kerr, J., Kim, C.M., Goldberg, K., Kanazawa, A., Tancik, M.: LERF: Language embedded radiance fields. In: Proceedings of the IEEE/CVF International Conference on Computer Vision. pp. 19729-19739 (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.445, + 0.785, + 0.473 + ], + "angle": 0, + "content": "22. Kingma, D.P., Welling, M.: Auto-encoding variational bayes. arXiv preprint arXiv:1312.6114 (2013)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.474, + 0.785, + 0.515 + ], + "angle": 0, + "content": "23. Kirillov, A., He, K., Girshick, R., Rother, C., Dollar, P.: Panoptic segmentation. In: Proceedings of the IEEE/CVF conference on computer vision and pattern recognition. pp. 9404-9413 (2019)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.516, + 0.785, + 0.558 + ], + "angle": 0, + "content": "24. Kirillov, A., Mintun, E., Ravi, N., Mao, H., Rolland, C., Gustafson, L., Xiao, T., Whitehead, S., Berg, A.C., Lo, W.Y., et al.: Segment anything. arXiv preprint arXiv:2304.02643 (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.559, + 0.785, + 0.6 + ], + "angle": 0, + "content": "25. Kobayashi, S., Matsumoto, E., Sitzmann, V.: Decomposing NeRF for editing via feature field distillation. Advances in Neural Information Processing Systems 35, 23311-23330 (2022)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.601, + 0.785, + 0.657 + ], + "angle": 0, + "content": "26. Kundu, A., Genova, K., Yin, X., Fathi, A., Pantofaru, C., Guibas, L.J., Tagliasacchi, A., Dellaert, F., Funkhouser, T.: Panoptic neural fields: A semantic object-aware neural scene representation. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 12871-12881 (2022)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.657, + 0.785, + 0.685 + ], + "angle": 0, + "content": "27. Le-Khac, P.H., Healy, G., Smeaton, A.F.: Contrastive representation learning: A framework and review. IEEE Access 8, 193907-193934 (2020)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.686, + 0.785, + 0.741 + ], + "angle": 0, + "content": "28. Lin, T.Y., Maire, M., Belongie, S., Hays, J., Perona, P., Ramanan, D., Dollar, P., Zitnick, C.L.: Microsoft COCO: Common objects in context. In: Computer Vision-ECCV 2014: 13th European Conference, Zurich, Switzerland, September 6-12, 2014, Proceedings, Part V 13. pp. 740-755. Springer (2014)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.741, + 0.785, + 0.77 + ], + "angle": 0, + "content": "29. Liu, L., Gu, J., Zaw Lin, K., Chua, T.S., Theobalt, C.: Neural sparse voxel fields. Advances in Neural Information Processing Systems 33, 15651-15663 (2020)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.771, + 0.785, + 0.798 + ], + "angle": 0, + "content": "30. McInnes, L., Healy, J., Astels, S.: hdbscan: Hierarchical density based clustering. J. Open Source Softw. 2(11), 205 (2017)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.799, + 0.785, + 0.84 + ], + "angle": 0, + "content": "31. Mildenhall, B., Srinivasan, P.P., Tancik, M., Barron, J.T., Ramamoorthi, R., Ng, R.: NeRF: Representing scenes as neural radiance fields for view synthesis. Communications of the ACM 65(1), 99-106 (2021)" + }, + { + "type": "list", + "bbox": [ + 0.217, + 0.147, + 0.787, + 0.84 + ], + "angle": 0, + "content": null + } + ], + [ + { + "type": "header", + "bbox": [ + 0.314, + 0.115, + 0.732, + 0.129 + ], + "angle": 0, + "content": "PCF-Lift: Panoptic Lifting by Probabilistic Contrastive Fusion" + }, + { + "type": "page_number", + "bbox": [ + 0.768, + 0.117, + 0.786, + 0.127 + ], + "angle": 0, + "content": "17" + }, + { + "type": "ref_text", + "bbox": [ + 0.214, + 0.147, + 0.788, + 0.189 + ], + "angle": 0, + "content": "32. Milioto, A., Behley, J., McCool, C., Stachniss, C.: LiDAR panoptic segmentation for autonomous driving. In: 2020 IEEE/RSJ International Conference on Intelligent Robots and Systems (IROS). pp. 8505-8512. IEEE (2020)" + }, + { + "type": "ref_text", + "bbox": [ + 0.214, + 0.189, + 0.788, + 0.23 + ], + "angle": 0, + "content": "33. Narita, G., Seno, T., Ishikawa, T., Kaji, Y.: Panopticfusion: Online volumetric semantic mapping at the level of stuff and things. In: 2019 IEEE/RSJ International Conference on Intelligent Robots and Systems (IROS). pp. 4205-4212. IEEE (2019)" + }, + { + "type": "ref_text", + "bbox": [ + 0.214, + 0.23, + 0.788, + 0.27 + ], + "angle": 0, + "content": "34. Newell, A., Huang, Z., Deng, J.: Associative embedding: End-to-end learning for joint detection and grouping. Advances in neural information processing systems 30 (2017)" + }, + { + "type": "ref_text", + "bbox": [ + 0.214, + 0.27, + 0.788, + 0.325 + ], + "angle": 0, + "content": "35. Nie, Y., Han, X., Guo, S., Zheng, Y., Chang, J., Zhang, J.J.: Total3D Understanding: Joint layout, object pose and mesh reconstruction for indoor scenes from a single image. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 55-64 (2020)" + }, + { + "type": "ref_text", + "bbox": [ + 0.214, + 0.325, + 0.788, + 0.366 + ], + "angle": 0, + "content": "36. Park, J., Lee, J., Kim, I.J., Sohn, K.: Probabilistic representations for video contrastive learning. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 14711-14721 (2022)" + }, + { + "type": "ref_text", + "bbox": [ + 0.214, + 0.366, + 0.788, + 0.407 + ], + "angle": 0, + "content": "37. Porzi, L., Bulo, S.R., Colovic, A., Kontschieder, P.: Seamless scene segmentation. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 8277-8286 (2019)" + }, + { + "type": "ref_text", + "bbox": [ + 0.214, + 0.407, + 0.788, + 0.447 + ], + "angle": 0, + "content": "38. Rezende, D.J., Mohamed, S., Wierstra, D.: Stochastic backpropagation and approximate inference in deep generative models. In: International conference on machine learning. pp. 1278-1286. PMLR (2014)" + }, + { + "type": "ref_text", + "bbox": [ + 0.214, + 0.447, + 0.788, + 0.488 + ], + "angle": 0, + "content": "39. Rosinol, A., Gupta, A., Abate, M., Shi, J., Carlone, L.: 3D dynamic scene graphs: Actionable spatial perception with places, objects, and humans. arXiv preprint arXiv:2002.06289 (2020)" + }, + { + "type": "ref_text", + "bbox": [ + 0.214, + 0.488, + 0.788, + 0.542 + ], + "angle": 0, + "content": "40. Siddiqui, Y., Porzi, L., Bulò, S.R., Müller, N., Nießner, M., Dai, A., Kontschieder, P.: Panoptic lifting for 3D scene understanding with neural fields. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 9043-9052 (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.214, + 0.542, + 0.788, + 0.583 + ], + "angle": 0, + "content": "41. Sirohi, K., Mohan, R., Buscher, D., Burgard, W., Valada, A.: Efficientlps: Efficient LiDAR panoptic segmentation. IEEE Transactions on Robotics 38(3), 1894-1914 (2021)" + }, + { + "type": "ref_text", + "bbox": [ + 0.214, + 0.583, + 0.788, + 0.624 + ], + "angle": 0, + "content": "42. Tang, W., Biqi, Y., Li, X., Liu, Y.H., Heng, P.A., Fu, C.W.: Prototypical variational autoencoder for 3D few-shot object detection. In: Thirty-seventh Conference on Neural Information Processing Systems (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.214, + 0.624, + 0.788, + 0.664 + ], + "angle": 0, + "content": "43. Tschernezki, V., Laina, I., Larlus, D., Vedaldi, A.: Neural feature fusion fields: 3d distillation of self-supervised 2d image representations. In: 2022 International Conference on 3D Vision (3DV). pp. 443-453. IEEE (2022)" + }, + { + "type": "ref_text", + "bbox": [ + 0.214, + 0.664, + 0.788, + 0.692 + ], + "angle": 0, + "content": "44. Wang, B., Chen, L., Yang, B.: Dm-NeRF: 3D scene geometry decomposition and manipulation from 2D images. arXiv preprint arXiv:2208.07227 (2022)" + }, + { + "type": "ref_text", + "bbox": [ + 0.214, + 0.692, + 0.788, + 0.746 + ], + "angle": 0, + "content": "45. Xie, S., Gu, J., Guo, D., Qi, C.R., Guibas, L., Litany, O.: Pointcontrast: Unsupervised pre-training for 3D point cloud understanding. In: Computer Vision-ECCV 2020: 16th European Conference, Glasgow, UK, August 23-28, 2020, Proceedings, Part III 16. pp. 574-591. Springer (2020)" + }, + { + "type": "ref_text", + "bbox": [ + 0.214, + 0.746, + 0.788, + 0.773 + ], + "angle": 0, + "content": "46. Ye, M., Danelljan, M., Yu, F., Ke, L.: Gaussian grouping: Segment and edit anything in 3D scenes. arXiv preprint arXiv:2312.00732 (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.214, + 0.773, + 0.788, + 0.799 + ], + "angle": 0, + "content": "47. Zhang, W., Pang, J., Chen, K., Loy, C.C.: K-net: Towards unified image segmentation. Advances in Neural Information Processing Systems 34, 10326-10338 (2021)" + }, + { + "type": "ref_text", + "bbox": [ + 0.214, + 0.799, + 0.788, + 0.841 + ], + "angle": 0, + "content": "48. Zhang, X., Chen, Z., Wei, F., Tu, Z.: Uni-3D: A universal model for panoptic 3D scene reconstruction. In: Proceedings of the IEEE/CVF International Conference on Computer Vision. pp. 9256-9266 (2023)" + }, + { + "type": "list", + "bbox": [ + 0.214, + 0.147, + 0.788, + 0.841 + ], + "angle": 0, + "content": null + } + ], + [ + { + "type": "page_number", + "bbox": [ + 0.218, + 0.116, + 0.236, + 0.127 + ], + "angle": 0, + "content": "18" + }, + { + "type": "header", + "bbox": [ + 0.272, + 0.115, + 0.467, + 0.129 + ], + "angle": 0, + "content": "R. Zhu, S. Qiu, Q. Wu, et al." + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.147, + 0.785, + 0.189 + ], + "angle": 0, + "content": "49. Zhi, S., Laidlow, T., Leutenegger, S., Davison, A.J.: In-place scene labelling and understanding with implicit scene representation. In: Proceedings of the IEEE/CVF International Conference on Computer Vision. pp. 15838-15847 (2021)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.19, + 0.785, + 0.231 + ], + "angle": 0, + "content": "50. Zhou, X., Girdhar, R., Joulin, A., Krahenbuhl, P., Misra, I.: Detecting twenty-thousand classes using image-level supervision. In: European Conference on Computer Vision. pp. 350-368. Springer (2022)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.231, + 0.785, + 0.273 + ], + "angle": 0, + "content": "51. Zhou, Z., Zhang, Y., Foroosh, H.: Panoptic-polarnet: Proposal-free LiDAR point cloud panoptic segmentation. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 13194-13203 (2021)" + }, + { + "type": "list", + "bbox": [ + 0.217, + 0.147, + 0.785, + 0.273 + ], + "angle": 0, + "content": null + } + ] +] \ No newline at end of file diff --git a/2024/PCF-Lift_ Panoptic Lifting by Probabilistic Contrastive Fusion/43912c86-ca82-4896-afdb-6d1a9980542c_origin.pdf b/2024/PCF-Lift_ Panoptic Lifting by Probabilistic Contrastive Fusion/43912c86-ca82-4896-afdb-6d1a9980542c_origin.pdf new file mode 100644 index 0000000000000000000000000000000000000000..a55fb785ea886c3b3951df7056a09a4b493693e7 --- /dev/null +++ b/2024/PCF-Lift_ Panoptic Lifting by Probabilistic Contrastive Fusion/43912c86-ca82-4896-afdb-6d1a9980542c_origin.pdf @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:afe5e9b63d4b06aa702461726d59415123f408cb4e1512f2e51b683092092f9a +size 9109555 diff --git a/2024/PCF-Lift_ Panoptic Lifting by Probabilistic Contrastive Fusion/full.md b/2024/PCF-Lift_ Panoptic Lifting by Probabilistic Contrastive Fusion/full.md new file mode 100644 index 0000000000000000000000000000000000000000..df4292457ce66892ab816901130c0c2de65394ec --- /dev/null +++ b/2024/PCF-Lift_ Panoptic Lifting by Probabilistic Contrastive Fusion/full.md @@ -0,0 +1,270 @@ +# PCF-Lift: Panoptic Lifting by Probabilistic Contrastive Fusion + +Runsong Zhu $^{1}$ , Shi Qiu $^{1}$ , Qianyi Wu $^{2}$ , Ka-Hei Hui $^{1}$ , Pheng-Ann Heng $^{1}$ , and Chi-Wing Fu $^{1}$ + +1 The Chinese University of Hong Kong +2 Monash University + +Abstract. Panoptic lifting is an effective technique to address the 3D panoptic segmentation task by unprojecting 2D panoptic segmentations from multi-views to 3D scene. However, the quality of its results largely depends on the 2D segmentations, which could be noisy and error-prone, so its performance often drops significantly for complex scenes. In this work, we design a new pipeline coined PCF-Lift based on our Probabilistic Contrastive Fusion (PCF) to learn and embed probabilistic features throughout our pipeline to actively consider inaccurate segmentations and inconsistent instance IDs. Technical-wise, we first model the probabilistic feature embeddings through multivariate Gaussian distributions. To fuse the probabilistic features, we incorporate the probability product kernel into the contrastive loss formulation and design a cross-view constraint to enhance the feature consistency across different views. For the inference, we introduce a new probabilistic clustering method to effectively associate prototype features with the underlying 3D object instances for the generation of consistent panoptic segmentation results. Further, we provide a theoretical analysis to justify the superiority of the proposed probabilistic solution. By conducting extensive experiments, our PCF-lift not only significantly outperforms the state-of-the-art methods on widely used benchmarks including the ScanNet dataset and the challenging Messy Room dataset (4.4% improvement of scene-level PQ), but also demonstrates strong robustness when incorporating various 2D segmentation models or different levels of hand-crafted noise. + +Keywords: Panoptic Lifting $\cdot$ Probabilistic Contrastive Fusion $\cdot$ Probabilistic Feature Embeddings + +# 1 Introduction + +3D panoptic segmentation [11,15,32,33,39,41,51] is a challenging 3D vision task, requiring the prediction of both semantic segmentation labels and instance segmentation labels. This task enables a comprehensive understanding of 3D scenes, thus facilitating many downstream applications, e.g., VR/AR, robotics, etc. + +Observing that the generalizability and applicability of existing 3D panoptic segmentation methods are limited by the scarcity of 3D training data, recent studies [3,25,40,43,49] suggest the idea of leveraging 2D panoptic segmentation + +![](images/d2a718e81741a93182b97968c5df4432a04166f085d48d112925528e3b039880.jpg) +Fig. 1: Our PCF-Lift method unprojects 2D panoptic segmentation predictions to 3D domain, facilitating the generation of consistent panoptic segmentation masks. For simplicity and clarity, we highlight instance segmentation masks. + +information predicted by foundation models [9, 24, 50]. Although 2D panoptic segmentation only provides image-based understanding, panoptic lifting [3, 40] is an effective technique to learn implicit 3D panoptic fields from 2D panoptic predictions, supporting the generation of coherent and view-consistent panoptic segmentation across different views. + +Recent works [3,40,44] on panoptic lifting mainly focus on solving the challenging instance-related issues, given that semantic predictions can be effectively associated across different views [49]. In practice, as shown in Fig. 1, one direct issue is inconsistent IDs that the same 3D object is assigned to different instance IDs in two different views for 2D panoptic prediction, which complicates the straightforward fusion process. To bypass the inconsistent IDs issue, Panoptic Lifting [40] directly fits instance ID permutations during the training process to learn deterministic feature embeddings for instance representation. Furthermore, Contrastive Lift [3] extends the scalability of Panoptic Lifting by using the contrastive learning technique to optimize the deterministic feature embeddings. However, existing methods still struggle to achieve satisfactory performance on complex scenes, as they overlook another interrelated issue: inconsistent segmentation. Specifically, due to the presence of inaccurate segmentation, the same object can be segmented inconsistently across two views: e.g., the chair in Fig. 1 is segmented into two parts in "view 1" but as a whole in "view 2". As existing methods use deterministic feature embeddings to learn 3D instance segmentation, the issue of inconsistent segmentation inherently introduces noise to training data and thus poses a robustness challenge for the models. + +In this paper, we introduce an effective probabilistic contrastive fusion solution to collectively address the two issues. For the issue of inconsistent segmentation, we propose to learn probabilistic feature embeddings rather than deterministic feature embeddings used in [3,40]. Our key insight is the development of probabilistic features, which correspond to distributions, allows for the incorporation of uncertainty modeling and enhances robustness to noise. This leads to a stable model optimization and, ultimately, results in a more reliable representation of instance information. Given these benefits, we develop probabilistic feature embeddings based on multivariate Gaussian distributions. Particularly, to enable contrastive learning among different Gaussian distributions, we devise the Probability Product (PP) Kernel [18] to measure the probabilistic feature similarities. Accordingly, a probabilistic clustering algorithm is introduced + +in the inference phase to generate consistent panoptic segmentation, using the measured probabilistic feature similarities. For the issue of inconsistent IDs, we devise the contrastive loss to fuse the probabilistic feature embeddings. In addition to formulating the loss term with each single view observation as inspired by the prior work [3], we introduce a novel cross-view constraint to facilitate an effective model training from segmentation results with inconsistent instance IDs. By dynamically exploiting feature pairs from different views, our proposed constraint offers an effective formulation to further enhance the feature consistency of the same 3D object instance across multiple views, and thus improves the panoptic lifting performance. + +To evaluate the effectiveness of our method, we conduct experiments on the ScanNet dataset [12] and the Messy Rooms dataset [3]. Furthermore, we conduct experiments to demonstrate the robustness of our probabilistic method to variations in segmentation models and different levels of noise. In addition to the experimental outcomes, our theoretical analysis from an optimization perspective demonstrates that the proposed probabilistic representation can be seen as a more flexible and generalized form compared to the previous deterministic representation [3]. + +Our contributions are summarized as follows: + +- We introduce a probabilistic contrastive fusion solution (PCF-Lift) to effectively unproject the 2D panoptic segmentations to the 3D domain by collectively considering the issues of inconsistent segmentation and inconsistent IDs. +- To fuse the probabilistic feature embeddings modeled by multivariate Gaussian distributions, we reformulate the contrastive loss with the Probability Product kernel and propose a novel cross-view constraint to further encourage the multi-view consistency. +- Coupled with a new probabilistic clustering algorithm, our proposed method outperforms the state-of-the-art methods consistently on the ScanNet dataset and the Messy Room dataset. + +# 2 Related Works + +# 2.1 Traditional 2D and 3D Panoptic Segmentation + +The 2D Panoptic segmentation task was initially introduced in [23]. Despite the notable advancements made by subsequent works [8-10,37,47], it remains challenging to panoptically understand individual images, while avoiding inconsistent instance recognitions across different image views of the scene. + +To enhance the panoptic understanding of the real world, 3D panoptic segmentation focuses on segmenting pre-computed 3D structures [15,32,41,51] (e.g., point clouds, voxels) or performing simultaneous 3D reconstruction or segmentation from 2D images [11,33,39]. However, its generalizability is rather limited, largely due to the significant difference in scale between 2D and 3D training data. Given the capabilities of recent 3D reconstruction techniques [6,20,31] and 2D + +panoptic segmentation models [9, 50], in this work, we explore 3D panoptic segmentation by leveraging multi-view images without explicit 3D input data. + +# 2.2 Multi-view Fusion + +Recently, researches in 3D reconstruction have made tremendous progress [2,6, 20, 29, 31]. Beyond novel view synthesis, we can utilize 3D reconstruction techniques as a tool to fuse 2D information like semantics and features in the 3D space [5, 14, 17, 21, 25, 35, 43, 46, 48]. For example, Semantic-NeRF [49] learns a semantic field from the 2D semantic segmentation, demonstrating the robustness and effectiveness of the semantic fusion. Later works [21, 25, 43] propose to fuse the unsupervised 2D dense feature for various segmentation and editing applications, showcasing the potential to adapt the 2D zero-shot models to 3D domain. In this work, we study the task of panoptic fusion, integrating both instance and semantic information to obtain a comprehensive understanding of the 3D scene. This task is significantly more challenging than the previous semantic fusion task [49], since we need to incorporate additional instance fusion and achieve instance label consistency. Unlike recent methods (e.g., Panoptic-Lifting [40] and Contrastive Lift [3]), which utilize deterministic feature embeddings to represent instance information, we develop a novel probabilistic solution, actively considering the error-prone and noisy nature of the 2D segmentations, such that we can duly enhance the effectiveness and robustness of instance fusion. + +# 2.3 Probabilistic Representation + +Probabilistic feature embedding is a popular tool employed in various tasks, e.g., image generation [22, 38], normal estimation [1], video understanding [36], point cloud understanding [4], and prototype embeddings for few-shot detection [42]. Motivated by its high capabilities in estimating aleatoric uncertainty [13] and addressing data noise, we innovate a new panoptic-lifting pipeline for 3D panoptic segmentation by formulating new modules based on probabilistic feature embeddings, including a reformulated contrastive loss with the probability product kernel [18] and an effective cross-view constraint to enhance the multi-view consistency. Further, we also design a novel probabilistic clustering algorithm to facilitate the generation of consistent panoptic segmentation. + +# 3 Method + +Given a set of posed images $\{\mathcal{I}\}$ associated with 2D panoptic segmentation predictions (i.e., semantic masks $\{\mathcal{H}\}$ and instance masks $\{\mathcal{K}\}$ ) generated by 2D segmentation models, our goal is to learn accurate 3D panoptic fields that can be rendered into consistent panoptic segmentation results over different views. + +The overview of PCF-Lift is illustrated in Fig. 2. Specifically, the 3D panoptic fields include a semantic field, an instance field, a density field, and a color field. Particularly, considering the significance of instance-related issues, we propose + +![](images/f6f4e7af24ecfec46155369fb1480767636138e609fe34faa13cdac3e3f754a0.jpg) +Fig. 2: Overview of PCF-Lift. The 3D panoptic fields include a semantic field, an instance field, a density field, and a color field. To solve the instance-related issues, we propose to learn probabilistic feature embeddings in the instance field (see Sec. 3.1). During the training phase, given two camera views, we can render the probabilistic feature maps from the instance field via volume rendering. To optimize the probabilistic instance field, we devise the probabilistic contrastive loss with Probability Product (PP) kernel [18], and propose a cross-view constraint to further enhance the feature consistency from different views (see Sec. 3.2). Similarly, we can render the semantic and color predictions, and adopt photometric loss and cross-entropy loss to optimize the semantic field, the density field, and the color field. During the inference phase, we design a novel multi-view object association (MVOA) algorithm for the generation of consistent panoptic segmentations (see Sec. 3.3). + +to learn probabilistic feature embeddings in the instances field (see Sec. 3.1). During the training phase, we jointly train the whole panoptic fields. Given two camera views, we can render the probabilistic feature maps from the instance field via volume rendering [19]. To optimize the probabilistic feature embeddings, we develop the contrastive loss with the Probability Product kernel [18], and propose a cross-view constraint to further enhance the feature consistency from different views (see Sec. 3.2). Similarly, the semantic predictions and color predictions are also rendered via volume rendering, where the semantic field, the density field, and the color field are optimized by the photometric loss and cross-entropy loss [3,6,40,49]. In the inference phase, we design a probabilistic clustering (i.e., multi-view object association) algorithm to effectively identify the prototype features of underlying 3D object instances for the generation of consistent panoptic segmentation results across any given views (see Sec. 3.3). + +# 3.1 Probabilistic Feature Embeddings + +To provide a robust instance representation, we propose to learn probabilistic feature embeddings in the instance field, which maps each 3D point to a random + +variable $\mathcal{F}$ , modeled as an $N$ -dimensional multivariate Gaussian distribution $\mathcal{F} \sim \mathcal{N}(\mu, \Sigma)$ . Here, $\mu$ is the mean vector, indicating the central feature values and $\Sigma$ is a diagonal covariance matrix $\mathrm{diag}(\sigma^2)$ with $\sigma^2 = (\sigma^{(1)^2}, \sigma^{(2)^2}, \dots, \sigma^{(N)^2})$ . Concretely, for any given query point $\mathbf{x} \in \mathbb{R}^3$ , the instance field predicts $(\mu, \sigma^2) \in \mathbb{R}^{2N}$ . Similar to the rendering of the color field, for each pixel in a given camera view, we can render its corresponding Gaussian distribution feature via volume rendering. A significant advantage of using probabilistic features rather than deterministic features is the ability of Gaussian distributions that assign different covariance values for uncertainty modeling. This property is crucial as it aids in reducing the impact of noise, thereby enhancing the robustness and accuracy of feature embeddings in representing complex 3D scenes. + +Intuitively, the similarity between two rendered Gaussian distributions indicates whether the corresponding pixels belong to the same instance. To quantify the similarities between different Gaussian distributions, we employ the Probability Product (PP) kernel [18] $K_{\rho}$ . Specifically, given two Gaussian distributions $\mathcal{F}_i \sim \mathcal{N}(\mu_i,\Sigma_i)$ and $\mathcal{F}_j \sim \mathcal{N}(\mu_j,\Sigma_j)$ , the corresponding kernel can be explicitly formulated as + +$$ +K _ {\rho} \left(\mathcal {F} _ {i}, \mathcal {F} _ {j}\right) = \left(\prod_ {d = 1} ^ {N} \left(\frac {\sigma_ {i} ^ {(d) ^ {2}}}{\sigma_ {j} ^ {(d) ^ {2}}} + \frac {\sigma_ {j} ^ {(d) ^ {2}}}{\sigma_ {i} ^ {(d) ^ {2}}}\right) / 2\right) ^ {- \frac {1}{2}} \exp \left(- \sum_ {d = 1} ^ {N} \left(\mu_ {i} ^ {(d)} - \mu_ {j} ^ {(d)}\right) ^ {2} / 4 \left(\sigma_ {i} ^ {(d) ^ {2}} + \sigma_ {j} ^ {(d) ^ {2}}\right)\right). \tag {1} +$$ + +Note that the results produced by the PP kernel fall within the range of 0 (minimal similarity) to 1 (maximal similarity), the same as the Radial Basis Function (RBF) kernel used in the deterministic method [3]. + +# 3.2 Training Probabilistic Feature Embeddings + +Probabilistic Contrastive Loss. During the training phase, we leverage the contrastive learning technique to optimize the probabilistic feature embeddings. Specifically, the contrastive loss used in [7, 27, 34, 45] tends to increase the predicted feature similarities of predefined positive pairs, while decreasing the similarities of negative pairs for training. In our work, a positive pair is defined as two pixels belonging to the same instance, while a negative pair denotes two pixels of different instances within a single image, following the previous method [3]. Since we basically learn probabilistic features, the PP kernel is exploited to measure the features similarity. In general, the loss can be formulated as + +$$ +\mathcal {L} _ {\mathrm {p i x e l - c o n t r a}} = - \frac {1}{| \varOmega |} \sum_ {u \in \varOmega} \log \frac {\sum_ {u ^ {\prime} \in \varOmega} \mathbf {1} _ {(u , u ^ {\prime})} \exp \left(K _ {\rho} \left(\mathcal {F} _ {u} , \mathcal {F} _ {u ^ {\prime}}\right)\right)}{\sum_ {u ^ {\prime} \in \varOmega} \exp \left(K _ {\rho} \left(\mathcal {F} _ {u} , \mathcal {F} _ {u ^ {\prime}}\right)\right)}, \tag {2} +$$ + +$$ +\mathcal {L} _ {\text {c o n c e n}} = - \frac {1}{| \varOmega |} \sum_ {u \in \varOmega} \log \left(K _ {\rho} \left(\mathcal {F} _ {u}, \frac {\sum_ {u ^ {\prime} \in \varOmega} \mathbf {1} _ {(u , u ^ {\prime})} \mathcal {F} _ {u ^ {\prime}}}{\sum_ {u ^ {\prime} \in \varOmega} \mathbf {1} _ {(u , u ^ {\prime})}}\right)\right), \tag {3} +$$ + +$$ +\text {a n d} \mathcal {L} _ {\text {c o n t r a}} = \mathcal {L} _ {\text {p i x e l - c o n t r a}} + \mathcal {L} _ {\text {c o n c e n}}, \tag {4} +$$ + +where $\mathcal{L}_{\mathrm{pixel - contra}}$ is the pixel-wise contrastive loss, $\mathcal{L}_{\mathrm{concen}}$ is the concentrate loss term, $\mathbf{1}$ is the indicator function of positive pairs, $\Omega$ is the set of pixel sample, $\mathcal{F}_u$ + +and $\mathcal{F}_{u'}$ are the rendered features for pixels $u$ and $u' \in \Omega$ via volume rendering, respectively. Different from previous method [3], we devise probabilistic similarity kernels to calculate a more effective contrastive loss for model training. By optimizing the loss Eq. (4), we can learn a more expressive 3D instance field for robust feature representations, as further analyzed in Sec. 3.4. To avoid the neural network from generating large covariances everywhere, we use an additional regularization term, $\mathcal{L}_{\mathrm{reg}} := \log \left( \prod_{d=1}^{N} \sigma^{(d)^2} \right)$ , to penalize the large covariances. + +Cross-view Constraint. To better train the probabilistic feature embeddings, we propose a cross-view constraint to enhance feature consistency for the same object across different views. Given the rendered learned probabilistic feature embedding sets $\{\mathcal{F}^m\}$ and $\{\mathcal{F}^n\}$ for two different views, we devise the PP kernel $K_{\rho}$ and a predefined threshold $\tau$ to collect positive pairs $\mathcal{P}$ that belong to the same object: $\mathcal{P} = \left\{(\mathcal{F}_r,\mathcal{F}_s)\mid K_\rho (\mathcal{F}_r,\mathcal{F}_s) > \tau ,\mathcal{F}_r\in \{\mathcal{F}^m\} ,\mathcal{F}_s\in \{\mathcal{F}^n\} \right\}$ . Moreover, as we expect to maximize the similarities between the features that belong to the same 3D object in different views, the cross-view constraint is defined as: + +$$ +\mathcal {L} _ {\text {c r o s s}} = - \frac {1}{| \mathcal {P} |} \sum_ {\left(\mathcal {F} _ {r}, \mathcal {F} _ {s}\right) \in \mathcal {P}} \log \left(K _ {\rho} \left(\mathcal {F} _ {r}, \mathcal {F} _ {s}\right)\right). \tag {5} +$$ + +In practice, we use the cross-view constraint only in a later optimization stage (i.e., the last few epochs), where the feature space is sufficiently meaningful to provide reliable positive pairs. The overall loss is formulated as + +$$ +\mathcal {L} = \mathcal {L} _ {\text {c o n t r a}} + w _ {\text {c r o s s}} \mathcal {L} _ {\text {c r o s s}} + w _ {\text {r e g}} \mathcal {L} _ {\text {r e g}}, \tag {6} +$$ + +where $w_{\mathrm{cross}}$ and $w_{\mathrm{reg}}$ are weight hyper-parameters. + +# 3.3 Multi-view Object Association + +During the inference phase, to facilitate consistent panoptic segmentations, we introduce a novel clustering algorithm, named the multi-view object association (MVOA) algorithm. The algorithm aims to extract the prototype feature set from the learned instance field for the assignment of instance labels. In general, through volume rendering, we obtain a large-scale rendered per-pixel probabilistic feature set $\{\mathcal{F}\}$ of the training views, as the input of our algorithm. Considering the efficiency issue, we first conduct an instance grouping operation to gather a smaller feature set $\mathcal{C}$ , with the help of inconsistent instance segmentation masks $\{\mathcal{K}\}$ obtained from training views. Then, we construct the probabilistic similarity graph based on $\mathcal{C}$ and design a multi-view matching process to collect the prototype feature set $\mathcal{D}$ . The details of instance grouping and multi-view matching operations are introduced as follows. + +Instance Grouping. We group the feature set $\{\mathcal{F}\}$ from each individual view. For the view of $l$ -th image, we collect the pixels that belong to instance $p$ into the set $\{i : \mathcal{K}_l^i = p\}$ , and the corresponding probabilistic feature set is denoted as + +$\{\mathcal{F}_l^i:\mathcal{K}_l^i = p\}$ . Then, we group $\{\mathcal{F}_l^i:\mathcal{K}_l^i = p\}$ to a single probabilistic feature by calculating the average: $\mathcal{C}_l^p = \sum_i\mathcal{F}_l^i /|\{i:\mathcal{K}_l^i = p\} |.$ Concurrently, we model the score quantity $\mathcal{S}_l^p$ for $\mathcal{C}_l^p$ as an indicator of feature concentration, which aids in the subsequent multi-view matching process. The indicator function $\varPhi(\mathcal{C}_l^p)$ averages the PP kernel similarities between $\mathcal{C}_l^p$ and all features in $\{\mathcal{F}_l^i:\mathcal{K}_l^i = p\}$ . Mathematically, $\varPhi(\mathcal{C}_l^p)$ follows: $\mathcal{S}_l^p = \varPhi(\mathcal{C}_l^p) = \sum_{\{i:\mathcal{K}_l^i = p\}}K_\rho (\mathcal{C}_l^p,\mathcal{F}_l^i) / |\{i:\mathcal{K}_l^i = p\} |.$ By applying this across all observed images, we obtain the grouped feature set $\mathcal{C}$ and the corresponding score set $\mathcal{S}$ . + +Multi-view Matching. To extract the structural relationships among the features, we construct an unoriented similarity graph $G = (\mathcal{C}, E)$ , where the grouped feature set $\mathcal{C}$ represents the nodes and $E$ encompasses the edges that indicate quantitative feature similarities measured by the PP Kernel: $E_{} = E_{} = K_{\rho}(\mathcal{C}_g, \mathcal{C}_h)$ , where $\mathcal{C}_g, \mathcal{C}_h \in \mathcal{C}$ . Then, we extract the prototype feature set $\mathcal{D}$ from the graph $G$ and score set $S$ by a greedy procedure, which is akin to the classical greedy algorithm known as non-maximum suppression (NMS). The pseudo-code of our proposed MVOA algorithm is presented in Algo. 1. In practice, the selection of hyper-parameter $\mathcal{T}$ is based on the average similarity computed across grouped feature pairs identified within each view. + +Generating Panoptic Segmentation Masks. We first follow the MVOA algorithm to collect the prototype feature set $\mathcal{D}$ . For each view, including the novel view, we then generate the semantic labels from the learned semantic field to differentiate between the background and foreground. Coupled with a feature map rendered from the probabilistic instance field, for each foreground pixel, we finally determine its corresponding instance label by assigning the candidate index from the set $\mathcal{D}$ that exhibits the highest similarity to the rendered feature. In practice, the MVOA algorithm needs to be conducted only once, while the extracted prototype features $\mathcal{D}$ will be used for generating the panoptic segmentation masks of all test views. + +# 3.4 Theoretical Analysis + +Corollary 1 If the covariances of the given Gaussian distributions are isotropic and fixed, i.e., $\Sigma_{i} = \Sigma_{j} = \sigma I$ , where $\sigma$ is a constant scalar, the probability product kernel can be simplified to an RBF kernel. + +The primary distinction between our probabilistic method and the prior deterministic method [3] is the choice of a similarity kernel. Specifically, our probabilistic method exploits the PP kernel, which offers higher flexibility and stronger expressive capability by adjusting the Gaussian mean and covariance compared to the RBF kernel employed in the deterministic method, as demonstrated in Fig 3. The RBF kernel can be considered as a degenerate PP kernel when applied to a Gaussian distribution with an isotropic and fixed covariance, as stated in Corollary 1. From this perspective, the proposed probabilistic method is a + +Algorithm 1: Multi-view object association algorithm (MVOA) +Data: Inconsistent instance mask $\{\mathcal{K}\}$ , input feature set $\{\mathcal{F}\}$ , Threshold T +Result: Prototype features set $\mathcal{D}$ +// Instance Grouping // Multi-view matching +2 $\mathcal{C} = \{\}$ , $\mathcal{S} = \{\}$ 10 $\mathcal{D} \gets \{\}$ +3 for $\mathcal{K}_l$ in $\{\mathcal{K}\}$ do 11 while $\mathcal{S} \neq empty$ do +4 IDs = {unique_ID(\mathcal{K}_l)} 12 m ← argmax S +5 for $p$ in $IDs$ do D ← D ∪ $\mathcal{C}_m$ +6 $\mathcal{C}_l^p = \sum_i\mathcal{F}_l^i /|\{i:\mathcal{K}_l^i = p\}|$ C ← C - Cm; S ← S - Sm +7 $\mathcal{S}_l^p = \Phi (\mathcal{C}_l^p)$ 13 if Kρ(Cm, Ci) ≥ T then 14 | C ← C - Ci; S ← S - Si +8 end 16 end +9 end 17 end + +![](images/5eb27910e9a7bbe8a14f61f75c1d7a5cd71931a08d3f1e524505cba76cef2509.jpg) +Fig. 3: (a) Flexibility of adjusting covariances. Contour plot of the PP kernel similarity for two Gaussians with different covariance values $(\sigma_{1}$ and $\sigma_{2})$ and fixed mean values in a 1-dimensional case. (b) Anisotropy. Contour plot of the PP kernel similarity for two Gaussians with different Gaussian mean offsets $(d_{x}$ and $d_{y})$ and fixed covariances in a 2-dimensional case. (c) Isotropy. Contour plot of the RBF kernel similarity for two deterministic features with different offsets $(d_{x}$ and $d_{y})$ in a 2-dimensional case. + +more general framework, while the deterministic method is a subclass within the broader probabilistic paradigm. The experimental results also verify the effectiveness of our probabilistic solution in the ablation studies of Sec. 4.3. + +# 4 Experiment + +# 4.1 Experimental Settings + +Implementation Details. For fair comparison, we adopt the same architecture of TensoRF [6] together with the same layer parameters used in previous works [3,40]. Particularly, our instance field is constructed following the slow-fast architecture of Contrastive Lift [3], where a shallow 5-layer MLP is applied to predict an probabilistic feature embedding for a 3D coordinate input. In practice, the dimension $N$ for the probabilistic feature embedding is 3 in our experiments. + +For the training of the instance field, we sample rays from two different views in the last few epochs for optimization (Sec. 3.2), using the cross-view constraint $\mathcal{L}_{\mathrm{cross}}$ with a predefined threshold value $\tau = 0.9$ and weight $w_{\mathrm{cross}} = 0.05$ ; while in other training epochs, we set the weight $w_{\mathrm{cross}}$ to 0 and only sample the rays from single views. Besides, $w_{\mathrm{reg}}$ in Eq. (6) is set to 0.001 throughout the whole training process. For the training of the color, density and semantic fields, we adopt the same loss terms and training strategies from previous works [3, 40]. More implementation details are provided in the supplementary material. + +Metrics. Since our method is proposed to solve the panoptic-lifting task, we employ scene-level Panoptic Quality (PQ $^{\text{scene}}$ ) for evaluation, which was introduced in Panoptic Lifting [40] and widely used in the related works [3,40]. Unlike the standard PQ metric [23], this metric particularly considers the consistency of instance IDs across multiple views. By merging predictions and ground truths with consistent instance IDs into subsets for PQ $^{\text{scene}}$ -based evaluation, we determine a match of subset pair when the intersection over union (IoU) exceeds 0.5, in line with prior baselines [3,40]. Since PQ $^{\text{scene}}$ is a product of scene-level segmentation quality (SQ $^{\text{scene}}$ ) and recognition quality (RQ $^{\text{scene}}$ ), we also provide the SQ $^{\text{scene}}$ and RQ $^{\text{scene}}$ metrics for more detailed comparisons. + +Baselines. We mainly compare the proposed method with the current state-of-the-art methods that target lifting 2D panoptic predictions to 3D, i.e., Contrastive Lift [3] and Panoptic Lifting [40]. Moreover, we compare our method with the recent NeRF-based 3D panoptic segmentation approaches, i.e., Panoptic Neural Fields [26] (PNF) and DM-NeRF [44]. + +Datasets. Following the previous works [3,40], we conduct experiments on two public datasets, the ScanNet dataset [12] and the Messy Room dataset [3], for both quantitative and qualitative evaluations. For fair comparisons on ScanNet, following the other state-of-the-art approaches [3,40], we adopt Mask2Former (M2F) [9] to generate 2D panoptic segmentation predictions, coupled with the same protocol in [40] that maps the COCO [28] vocabulary to 21 classes. For the experiments on Messy Rooms, we use the LVIS [16] vocabulary and the Detic [50] 2D panoptic segmentations, which are also utilized in Contrastive Lift [3]. + +# 4.2 Main Experiments + +ScanNet Dataset. To verify the performance on real data, we conduct experiments on the ScanNet dataset [12] of 12 scenes. Quantitative comparisons provided in Tab. 1 demonstrate that our method consistently outperforms previous baselines, including both the 3D panoptic segmentation approaches [26,44] and the recent state-of-the-art methods for lifting 2D panoptic segmentation [3,40]. Moreover, visual comparisons between our method and the latest state-of-the-art method [3] are presented in Fig. 4, which exhibits our method's capability of achieving a consistent and accurate 3D panoptic segmentation. + +Table 1: Results on the ScanNet dataset. We report the PQscene, SQscene, and RQscene metrics. Since Contrastive Lift [3] does not report the performance using the SQscene and RQscene metrics, we apply the officially-released pre-trained model and re-run the clustering algorithm to obtain the values reported in this table. For the other metric values, we directly report the ones in previous papers [3,40]. + +
MethodVenueType\(SQ^{scene}(\%)\)\(RQ^{scene}(\%)\)\(PQ^{scene}(\%)\)
DM-NeRF [44]ICLR'233D panoptic segmentation53.346.141.7
PNF [26]CVPR'223D panoptic segmentation63.050.748.3
PNF [26] + GT BBoxesCVPR'223D panoptic segmentation70.055.954.3
Panoptic Lifting [40]CVPR'232D panoptic Lifting73.565.058.9
Contrastive Lift [3]NeurIPS'232D panoptic Lifting75.763.662.0
Ours-2D panoptic Lifting78.565.463.5
+ +Table 2: Results on the Messy Rooms dataset [3]. Following [3], the PQscene metric is reported on both "old room" and "large corridor" environments with an increasing number of objects in the scene (25, 50, 100, 500). + +
Method/ NumberOld Room Environment (%)Large Corridor Environment(%)Mean(%)
25501005002550100500
Panoptic Lifting [40]73.269.964.351.065.571.061.849.063.2
Contrastive Lift [3]78.975.869.155.076.575.568.752.569.0
Ours80.978.374.860.381.079.474.058.873.4
+ +Messy Rooms Dataset. We also conduct experiments on the challenging Messy Room dataset [3], which is provided by Contrastive Lift [3], containing up to 500 objects in each scene. We present the quantitative results in Tab. 2, showing that our method (73.4%) achieves significant improvements in terms of the mean PQ $^{\text{scene}}$ results compared to the current state-of-the-art methods of Contrastive Lift [3] (69.0%) and Panoptic Lifting [40] (63.2%). The overall performance highlights the advantages of our probabilistic approach over the previous deterministic methods, particularly in segmenting complex scenes with hundreds of objects. Further, our method achieves a mean SQ $^{\text{scene}}$ of 82.2% and a mean RQ $^{\text{scene}}$ of 86.9%, surpassing the SQ $^{\text{scene}}$ of 77.7% and RQ $^{\text{scene}}$ of 86.6% obtained by Contrastive Lift [3]. Moreover, the visual comparisons in Fig. 4 further indicate that our method can accurately segment the small objects in the red frames, whereas Contrastive Lift [3] struggles to distinguish such instances. + +# 4.3 Ablation Study + +Effectiveness of Each Component. We conduct an ablation study on the Messy Room dataset [3], which contains more challenging scenes with hundreds of instances. As shown in Tab. 3 (b) and (d), the proposed probabilistic feature embeddings greatly benefit our method, while replacing it with deterministic one leads to significant performance drop. Furthermore, the effectiveness of our proposed cross-view constraint is verified in Tab. 3 (e) and (f). For the proposed multi-view object association (MVOA) algorithm, the results in Tab. 3 (c), (d), (e), and (f) show that it particularly benefits the clustering of our proposed + +![](images/39c375f5c14bb29d59de648735d61e2302061b04ed2f694827fcd1b78f3d147b.jpg) +Fig. 4: Visual comparison of the latest state-of-the-art method Contrastive Lift [3] and our method on the ScanNet [12] dataset and the Messy Room [3] dataset. + +Table 3: Ablation study on the Messy Room dataset [3]. The model (a) corresponds to Contrastive Lift [3] and the model (f) corresponds to our full method (PCF). + +
Model Feature spaceClustering\( {\mathrm{{SQ}}}^{\text{scene }}\left( \% \right) \)\( {\mathrm{{RQ}}}^{\text{scene }}\left( \% \right) \)\( {\mathrm{{PQ}}}^{\text{scene }}\left( \% \right) \)
(a)Deterministic [3]HDBSCAN [30]77.786.669.0
(b)Deterministic [3]MVOA79.386.270.4
(c)Learned Gaussian distributionHDBSCAN [30]78.086.669.6
(d)Learned Gaussian distributionMVOA81.386.872.3
(e)Learned Gaussian distribution (+ Cross-view constraint)HDBSCAN [30]78.886.970.4
(f)Learned Gaussian distribution (+ Cross-view constraint)MVOA82.286.973.4
+ +probabilistic representation. Also, MVOA can be effectively employed to the deterministic method [3] for a performance boost, as Tab. 3 (a) and (b) show. + +![](images/3b8cf2d43fca2acbdd6d2c4b542b71d6913358443b2654678a90e96ce0c07721.jpg) +Fig. 5: The visualization results of learned covariance components and the statistical results of covariances in two scenes of the Messy Room dataset [3]. For the histograms, the horizontal axis denotes the range of covariance magnitudes, while the vertical axis corresponds to the frequency statistics for those magnitudes. We calculate and plot the covariance magnitudes $(\sigma^{(1)^2}*\sigma^{(2)^2}*\sigma^{(3)^2})$ of two distance image regions, the boundary areas and the internal areas of object instances, across all observed views. + +Uncertainty Analysis. To verify whether our method could provide meaningful modeling for uncertainty, we provide the rendered uncertainty maps on two scenes of the Messy Room dataset [3] in Fig. 5. The figure clearly illustrates that the regions of high covariance are mainly located near the boundaries of the object instances, due to the inconsistent segmentation issue that leaves severe ambiguity around the instance boundaries. Moreover, we statistically analyze the learned covariances in two distinct image regions, namely the boundary areas and the internal areas of object instances, across all observed views. These results further demonstrate our method's ability to model high uncertainty in areas where the instance boundaries are ambiguous. We provide more details in the supplementary material. + +# 4.4 Robustness Experiments + +2D Backbones. In practice, the quality of panoptic segmentation generated by different 2D models may vary a lot. To study the robustness of our probabilistic method and the deterministic method [3] when incorporating with different 2D models, we select four different models from the model zoo of Detic [3] and utilize the LVIS [16] vocabulary to generate the 2D panoptic segmentation predictions for lifting, as shown in Fig. 6 (a). The quantitative results are presented in Fig. 6 (b), where our probabilistic methods consistently outperform the baseline method by a large margin. Particularly, we observe that the proposed MVOA algorithm can consistently boost the performance of a deterministic method [3]. + +Hand-crafted Noise. We study the robustness by adding hand-crafted noise. Specifically, for all given panoptic segmentation masks, we randomly select 100 pixels as anchors. Then, for each anchor, we randomly choose a pixel within a $W * W$ window centered on the anchor and assign its instance ID to all pixels within this window, to simulate inaccurate segmentation predictions around object boundaries. As shown in Fig. 6 (c), as W gradually increases, the instance + +![](images/36d2f150e399caa373b87d78d087fc90d1ac5b9babcbd02df12f42b5e2bf4583.jpg) +Fig. 6: Quantitative comparisons when using different 2D models or adding various levels of hand-crafted noise. We compare our probabilistic method ("Probabilistic + MVOA" in red) with the deterministic method [3] ("Deterministic" in dark blue) and a variant of deterministic method using our proposed MVOA algorithm ("Deterministic + MVOA" in light blue). Besides, the four tested models are from the model zoo of Detic $[50]^1$ , and the two tested scenes are from the Messy Room dataset [3]. + +boundaries are continuously blurred. Our probabilistic approach consistently achieves the best performance as presented in Fig. 6 (d) and the results demonstrate that the proposed MVOA algorithm benefits the deterministic method [3]. + +# 5 Conclusion + +We present PCF-Lift for the panoptic lifting task. First, we propose to learn the probabilistic feature embeddings through a multivariate Gaussian distribution for instance representation. For training, we reformulate the contrastive loss with the Probability Product kernel and propose a novel cross-view constraint to enhance the feature consistency across different views. During the inference phase, we propose a novel multi-view object association algorithm to effectively identify the prototype features representing underlying 3D object instances. We verify the effectiveness and robustness of PCF-Lift by extensive experiments. + +# Acknowledgements + +This work is supported by the InnoHK of the Government of Hong Kong via the Hong Kong Centre for Logistics Robotics, the Shenzhen Portion of Shenzhen-Hong Kong Science and Technology Innovation Cooperation Zone under HZQB-KCZYB-20200089, and the CUHK T Stone Robotics Institute. + +# References + +1. Bae, G., Budvytis, I., Cipolla, R.: Estimating and exploiting the aleatoric uncertainty in surface normal estimation. In: Proceedings of the IEEE/CVF International Conference on Computer Vision. pp. 13137-13146 (2021) +2. Barron, J.T., Mildenhall, B., Tancik, M., Hedman, P., Martin-Brualla, R., Srinivasan, P.P.: Mip-NeRF: A multiscale representation for anti-aliasing neural radiance fields. In: Proceedings of the IEEE/CVF International Conference on Computer Vision. pp. 5855-5864 (2021) +3. Bhalgat, Y., Laina, I., Henriques, J.F., Zisserman, A., Vedaldi, A.: Contrastive Lift: 3D object instance segmentation by slow-fast contrastive fusion. arXiv preprint arXiv:2306.04633 (2023) +4. Cai, K., Lu, C.X., Huang, X.: Uncertainty estimation for 3D dense prediction via cross-point embeddings. IEEE Robotics and Automation Letters 8(5), 2558-2565 (2023) +5. Cen, J., Zhou, Z., Fang, J., Shen, W., Xie, L., Jiang, D., Zhang, X., Tian, Q., et al.: Segment anything in 3D with NeRFs. Advances in Neural Information Processing Systems 36 (2024) +6. Chen, A., Xu, Z., Geiger, A., Yu, J., Su, H.: Tensorf: Tensorial radiance fields. In: European Conference on Computer Vision. pp. 333-350. Springer (2022) +7. Chen, T., Kornblith, S., Norouzi, M., Hinton, G.: A simple framework for contrastive learning of visual representations. In: International conference on machine learning. pp. 1597-1607. PMLR (2020) +8. Cheng, B., Collins, M.D., Zhu, Y., Liu, T., Huang, T.S., Adam, H., Chen, L.C.: Panoptic-deeplab: A simple, strong, and fast baseline for bottom-up panoptic segmentation. In: Proceedings of the IEEE/CVF conference on computer vision and pattern recognition. pp. 12475-12485 (2020) +9. Cheng, B., Misra, I., Schwing, A.G., Kirillov, A., Girdhar, R.: Masked-attention mask transformer for universal image segmentation. In: Proceedings of the IEEE/CVF conference on computer vision and pattern recognition. pp. 1290-1299 (2022) +0. Cheng, B., Schwing, A., Kirillov, A.: Per-pixel classification is not all you need for semantic segmentation. Advances in Neural Information Processing Systems 34, 17864-17875 (2021) +1. Dahnert, M., Hou, J., Nießner, M., Dai, A.: Panoptic 3D scene reconstruction from a single rgb image. Advances in Neural Information Processing Systems 34, 8282-8293 (2021) +12. Dai, A., Chang, A.X., Savva, M., Halber, M., Funkhouser, T., Nießner, M.: Scannet: Richly-annotated 3D reconstructions of indoor scenes. In: Proceedings of the IEEE conference on computer vision and pattern recognition. pp. 5828-5839 (2017) +13. Der Kiureghian, A., Ditlevsen, O.: Aleatory or epistemic? does it matter? Structural safety 31(2), 105-112 (2009) + +14. Fan, Z., Wang, P., Jiang, Y., Gong, X., Xu, D., Wang, Z.: NeRF-SOS: Any-iew self-supervised object segmentation on complex scenes. arXiv preprint arXiv:2209.08776 (2022) +15. Gasperini, S., Mahani, M.A.N., Marcos-Ramiro, A., Navab, N., Tombari, F.: Panostter: End-to-end panoptic segmentation of LiDAR point clouds. IEEE Robotics and Automation Letters 6(2), 3216-3223 (2021) +16. Gupta, A., Dollar, P., Girshick, R.: LVIS: A dataset for large vocabulary instance segmentation. In: Proceedings of the IEEE/CVF conference on computer vision and pattern recognition. pp. 5356-5364 (2019) +17. Hamilton, M., Zhang, Z., Hariharan, B., Snavely, N., Freeman, W.T.: Unsupervised semantic segmentation by distilling feature correspondences. arXiv preprint arXiv:2203.08414 (2022) +18. Jebara, T., Kondor, R., Howard, A.: Probability product kernels. The Journal of Machine Learning Research 5, 819-844 (2004) +19. Kajiya, J.T., Von Herzen, B.P.: Ray tracing volume densities. ACM SIGGRAPH computer graphics 18(3), 165-174 (1984) +20. Kerbl, B., Kopanas, G., Leimkuhler, T., Drettakis, G.: 3D gaussian splatting for real-time radiance field rendering. ACM Transactions on Graphics 42(4) (2023) +21. Kerr, J., Kim, C.M., Goldberg, K., Kanazawa, A., Tancik, M.: LERF: Language embedded radiance fields. In: Proceedings of the IEEE/CVF International Conference on Computer Vision. pp. 19729-19739 (2023) +22. Kingma, D.P., Welling, M.: Auto-encoding variational bayes. arXiv preprint arXiv:1312.6114 (2013) +23. Kirillov, A., He, K., Girshick, R., Rother, C., Dollar, P.: Panoptic segmentation. In: Proceedings of the IEEE/CVF conference on computer vision and pattern recognition. pp. 9404-9413 (2019) +24. Kirillov, A., Mintun, E., Ravi, N., Mao, H., Rolland, C., Gustafson, L., Xiao, T., Whitehead, S., Berg, A.C., Lo, W.Y., et al.: Segment anything. arXiv preprint arXiv:2304.02643 (2023) +25. Kobayashi, S., Matsumoto, E., Sitzmann, V.: Decomposing NeRF for editing via feature field distillation. Advances in Neural Information Processing Systems 35, 23311-23330 (2022) +26. Kundu, A., Genova, K., Yin, X., Fathi, A., Pantofaru, C., Guibas, L.J., Tagliasacchi, A., Dellaert, F., Funkhouser, T.: Panoptic neural fields: A semantic object-aware neural scene representation. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 12871-12881 (2022) +27. Le-Khac, P.H., Healy, G., Smeaton, A.F.: Contrastive representation learning: A framework and review. IEEE Access 8, 193907-193934 (2020) +28. Lin, T.Y., Maire, M., Belongie, S., Hays, J., Perona, P., Ramanan, D., Dollar, P., Zitnick, C.L.: Microsoft COCO: Common objects in context. In: Computer Vision-ECCV 2014: 13th European Conference, Zurich, Switzerland, September 6-12, 2014, Proceedings, Part V 13. pp. 740-755. Springer (2014) +29. Liu, L., Gu, J., Zaw Lin, K., Chua, T.S., Theobalt, C.: Neural sparse voxel fields. Advances in Neural Information Processing Systems 33, 15651-15663 (2020) +30. McInnes, L., Healy, J., Astels, S.: hdbscan: Hierarchical density based clustering. J. Open Source Softw. 2(11), 205 (2017) +31. Mildenhall, B., Srinivasan, P.P., Tancik, M., Barron, J.T., Ramamoorthi, R., Ng, R.: NeRF: Representing scenes as neural radiance fields for view synthesis. Communications of the ACM 65(1), 99-106 (2021) + +32. Milioto, A., Behley, J., McCool, C., Stachniss, C.: LiDAR panoptic segmentation for autonomous driving. In: 2020 IEEE/RSJ International Conference on Intelligent Robots and Systems (IROS). pp. 8505-8512. IEEE (2020) +33. Narita, G., Seno, T., Ishikawa, T., Kaji, Y.: Panopticfusion: Online volumetric semantic mapping at the level of stuff and things. In: 2019 IEEE/RSJ International Conference on Intelligent Robots and Systems (IROS). pp. 4205-4212. IEEE (2019) +34. Newell, A., Huang, Z., Deng, J.: Associative embedding: End-to-end learning for joint detection and grouping. Advances in neural information processing systems 30 (2017) +35. Nie, Y., Han, X., Guo, S., Zheng, Y., Chang, J., Zhang, J.J.: Total3D Understanding: Joint layout, object pose and mesh reconstruction for indoor scenes from a single image. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 55-64 (2020) +36. Park, J., Lee, J., Kim, I.J., Sohn, K.: Probabilistic representations for video contrastive learning. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 14711-14721 (2022) +37. Porzi, L., Bulo, S.R., Colovic, A., Kontschieder, P.: Seamless scene segmentation. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 8277-8286 (2019) +38. Rezende, D.J., Mohamed, S., Wierstra, D.: Stochastic backpropagation and approximate inference in deep generative models. In: International conference on machine learning. pp. 1278-1286. PMLR (2014) +39. Rosinol, A., Gupta, A., Abate, M., Shi, J., Carlone, L.: 3D dynamic scene graphs: Actionable spatial perception with places, objects, and humans. arXiv preprint arXiv:2002.06289 (2020) +40. Siddiqui, Y., Porzi, L., Bulò, S.R., Müller, N., Nießner, M., Dai, A., Kontschieder, P.: Panoptic lifting for 3D scene understanding with neural fields. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 9043-9052 (2023) +41. Sirohi, K., Mohan, R., Buscher, D., Burgard, W., Valada, A.: Efficientlps: Efficient LiDAR panoptic segmentation. IEEE Transactions on Robotics 38(3), 1894-1914 (2021) +42. Tang, W., Biqi, Y., Li, X., Liu, Y.H., Heng, P.A., Fu, C.W.: Prototypical variational autoencoder for 3D few-shot object detection. In: Thirty-seventh Conference on Neural Information Processing Systems (2023) +43. Tschernezki, V., Laina, I., Larlus, D., Vedaldi, A.: Neural feature fusion fields: 3d distillation of self-supervised 2d image representations. In: 2022 International Conference on 3D Vision (3DV). pp. 443-453. IEEE (2022) +44. Wang, B., Chen, L., Yang, B.: Dm-NeRF: 3D scene geometry decomposition and manipulation from 2D images. arXiv preprint arXiv:2208.07227 (2022) +45. Xie, S., Gu, J., Guo, D., Qi, C.R., Guibas, L., Litany, O.: Pointcontrast: Unsupervised pre-training for 3D point cloud understanding. In: Computer Vision-ECCV 2020: 16th European Conference, Glasgow, UK, August 23-28, 2020, Proceedings, Part III 16. pp. 574-591. Springer (2020) +46. Ye, M., Danelljan, M., Yu, F., Ke, L.: Gaussian grouping: Segment and edit anything in 3D scenes. arXiv preprint arXiv:2312.00732 (2023) +47. Zhang, W., Pang, J., Chen, K., Loy, C.C.: K-net: Towards unified image segmentation. Advances in Neural Information Processing Systems 34, 10326-10338 (2021) +48. Zhang, X., Chen, Z., Wei, F., Tu, Z.: Uni-3D: A universal model for panoptic 3D scene reconstruction. In: Proceedings of the IEEE/CVF International Conference on Computer Vision. pp. 9256-9266 (2023) + +49. Zhi, S., Laidlow, T., Leutenegger, S., Davison, A.J.: In-place scene labelling and understanding with implicit scene representation. In: Proceedings of the IEEE/CVF International Conference on Computer Vision. pp. 15838-15847 (2021) +50. Zhou, X., Girdhar, R., Joulin, A., Krahenbuhl, P., Misra, I.: Detecting twenty-thousand classes using image-level supervision. In: European Conference on Computer Vision. pp. 350-368. Springer (2022) +51. Zhou, Z., Zhang, Y., Foroosh, H.: Panoptic-polarnet: Proposal-free LiDAR point cloud panoptic segmentation. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 13194-13203 (2021) \ No newline at end of file diff --git a/2024/PCF-Lift_ Panoptic Lifting by Probabilistic Contrastive Fusion/images.zip b/2024/PCF-Lift_ Panoptic Lifting by Probabilistic Contrastive Fusion/images.zip new file mode 100644 index 0000000000000000000000000000000000000000..69e29879f9d0a53175ccf83d5a5804428c27e577 --- /dev/null +++ b/2024/PCF-Lift_ Panoptic Lifting by Probabilistic Contrastive Fusion/images.zip @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:233ebd1c7e0c50c52ce02709bc5cba2a391c81afa43d0d8bebab18c807df767a +size 486096 diff --git a/2024/PCF-Lift_ Panoptic Lifting by Probabilistic Contrastive Fusion/layout.json b/2024/PCF-Lift_ Panoptic Lifting by Probabilistic Contrastive Fusion/layout.json new file mode 100644 index 0000000000000000000000000000000000000000..3ca1ff37923a7e70928b1cedb8805800f464e1d4 --- /dev/null +++ b/2024/PCF-Lift_ Panoptic Lifting by Probabilistic Contrastive Fusion/layout.json @@ -0,0 +1,8864 @@ +{ + "pdf_info": [ + { + "para_blocks": [ + { + "bbox": [ + 194, + 111, + 419, + 145 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 194, + 111, + 419, + 145 + ], + "spans": [ + { + "bbox": [ + 194, + 111, + 419, + 145 + ], + "type": "text", + "content": "PCF-Lift: Panoptic Lifting by Probabilistic Contrastive Fusion" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 192, + 167, + 421, + 193 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 192, + 167, + 421, + 193 + ], + "spans": [ + { + "bbox": [ + 192, + 167, + 421, + 193 + ], + "type": "text", + "content": "Runsong Zhu" + }, + { + "bbox": [ + 192, + 167, + 421, + 193 + ], + "type": "inline_equation", + "content": "^{1}" + }, + { + "bbox": [ + 192, + 167, + 421, + 193 + ], + "type": "text", + "content": ", Shi Qiu" + }, + { + "bbox": [ + 192, + 167, + 421, + 193 + ], + "type": "inline_equation", + "content": "^{1}" + }, + { + "bbox": [ + 192, + 167, + 421, + 193 + ], + "type": "text", + "content": ", Qianyi Wu" + }, + { + "bbox": [ + 192, + 167, + 421, + 193 + ], + "type": "inline_equation", + "content": "^{2}" + }, + { + "bbox": [ + 192, + 167, + 421, + 193 + ], + "type": "text", + "content": ", Ka-Hei Hui" + }, + { + "bbox": [ + 192, + 167, + 421, + 193 + ], + "type": "inline_equation", + "content": "^{1}" + }, + { + "bbox": [ + 192, + 167, + 421, + 193 + ], + "type": "text", + "content": ", Pheng-Ann Heng" + }, + { + "bbox": [ + 192, + 167, + 421, + 193 + ], + "type": "inline_equation", + "content": "^{1}" + }, + { + "bbox": [ + 192, + 167, + 421, + 193 + ], + "type": "text", + "content": ", and Chi-Wing Fu" + }, + { + "bbox": [ + 192, + 167, + 421, + 193 + ], + "type": "inline_equation", + "content": "^{1}" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 224, + 201, + 389, + 224 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 224, + 201, + 389, + 224 + ], + "spans": [ + { + "bbox": [ + 224, + 201, + 389, + 224 + ], + "type": "text", + "content": "1 The Chinese University of Hong Kong \n2 Monash University" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 160, + 251, + 452, + 503 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 160, + 251, + 452, + 503 + ], + "spans": [ + { + "bbox": [ + 160, + 251, + 452, + 503 + ], + "type": "text", + "content": "Abstract. Panoptic lifting is an effective technique to address the 3D panoptic segmentation task by unprojecting 2D panoptic segmentations from multi-views to 3D scene. However, the quality of its results largely depends on the 2D segmentations, which could be noisy and error-prone, so its performance often drops significantly for complex scenes. In this work, we design a new pipeline coined PCF-Lift based on our Probabilistic Contrastive Fusion (PCF) to learn and embed probabilistic features throughout our pipeline to actively consider inaccurate segmentations and inconsistent instance IDs. Technical-wise, we first model the probabilistic feature embeddings through multivariate Gaussian distributions. To fuse the probabilistic features, we incorporate the probability product kernel into the contrastive loss formulation and design a cross-view constraint to enhance the feature consistency across different views. For the inference, we introduce a new probabilistic clustering method to effectively associate prototype features with the underlying 3D object instances for the generation of consistent panoptic segmentation results. Further, we provide a theoretical analysis to justify the superiority of the proposed probabilistic solution. By conducting extensive experiments, our PCF-lift not only significantly outperforms the state-of-the-art methods on widely used benchmarks including the ScanNet dataset and the challenging Messy Room dataset (4.4% improvement of scene-level PQ), but also demonstrates strong robustness when incorporating various 2D segmentation models or different levels of hand-crafted noise." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 160, + 515, + 453, + 537 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 160, + 515, + 453, + 537 + ], + "spans": [ + { + "bbox": [ + 160, + 515, + 453, + 537 + ], + "type": "text", + "content": "Keywords: Panoptic Lifting " + }, + { + "bbox": [ + 160, + 515, + 453, + 537 + ], + "type": "inline_equation", + "content": "\\cdot" + }, + { + "bbox": [ + 160, + 515, + 453, + 537 + ], + "type": "text", + "content": " Probabilistic Contrastive Fusion " + }, + { + "bbox": [ + 160, + 515, + 453, + 537 + ], + "type": "inline_equation", + "content": "\\cdot" + }, + { + "bbox": [ + 160, + 515, + 453, + 537 + ], + "type": "text", + "content": " Probabilistic Feature Embeddings" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 132, + 557, + 229, + 569 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 557, + 229, + 569 + ], + "spans": [ + { + "bbox": [ + 132, + 557, + 229, + 569 + ], + "type": "text", + "content": "1 Introduction" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 130, + 581, + 481, + 629 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 581, + 481, + 629 + ], + "spans": [ + { + "bbox": [ + 130, + 581, + 481, + 629 + ], + "type": "text", + "content": "3D panoptic segmentation [11,15,32,33,39,41,51] is a challenging 3D vision task, requiring the prediction of both semantic segmentation labels and instance segmentation labels. This task enables a comprehensive understanding of 3D scenes, thus facilitating many downstream applications, e.g., VR/AR, robotics, etc." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 130, + 630, + 481, + 665 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 630, + 481, + 665 + ], + "spans": [ + { + "bbox": [ + 130, + 630, + 481, + 665 + ], + "type": "text", + "content": "Observing that the generalizability and applicability of existing 3D panoptic segmentation methods are limited by the scarcity of 3D training data, recent studies [3,25,40,43,49] suggest the idea of leveraging 2D panoptic segmentation" + } + ] + } + ], + "index": 7 + } + ], + "discarded_blocks": [], + "page_size": [ + 612, + 792 + ], + "page_idx": 0 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 151, + 112, + 460, + 159 + ], + "blocks": [ + { + "bbox": [ + 151, + 112, + 460, + 159 + ], + "lines": [ + { + "bbox": [ + 151, + 112, + 460, + 159 + ], + "spans": [ + { + "bbox": [ + 151, + 112, + 460, + 159 + ], + "type": "image", + "image_path": "d2a718e81741a93182b97968c5df4432a04166f085d48d112925528e3b039880.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 130, + 178, + 480, + 212 + ], + "lines": [ + { + "bbox": [ + 130, + 178, + 480, + 212 + ], + "spans": [ + { + "bbox": [ + 130, + 178, + 480, + 212 + ], + "type": "text", + "content": "Fig. 1: Our PCF-Lift method unprojects 2D panoptic segmentation predictions to 3D domain, facilitating the generation of consistent panoptic segmentation masks. For simplicity and clarity, we highlight instance segmentation masks." + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_caption" + } + ], + "index": 2 + }, + { + "bbox": [ + 130, + 240, + 479, + 300 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 240, + 479, + 300 + ], + "spans": [ + { + "bbox": [ + 130, + 240, + 479, + 300 + ], + "type": "text", + "content": "information predicted by foundation models [9, 24, 50]. Although 2D panoptic segmentation only provides image-based understanding, panoptic lifting [3, 40] is an effective technique to learn implicit 3D panoptic fields from 2D panoptic predictions, supporting the generation of coherent and view-consistent panoptic segmentation across different views." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 130, + 304, + 482, + 518 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 304, + 482, + 518 + ], + "spans": [ + { + "bbox": [ + 130, + 304, + 482, + 518 + ], + "type": "text", + "content": "Recent works [3,40,44] on panoptic lifting mainly focus on solving the challenging instance-related issues, given that semantic predictions can be effectively associated across different views [49]. In practice, as shown in Fig. 1, one direct issue is inconsistent IDs that the same 3D object is assigned to different instance IDs in two different views for 2D panoptic prediction, which complicates the straightforward fusion process. To bypass the inconsistent IDs issue, Panoptic Lifting [40] directly fits instance ID permutations during the training process to learn deterministic feature embeddings for instance representation. Furthermore, Contrastive Lift [3] extends the scalability of Panoptic Lifting by using the contrastive learning technique to optimize the deterministic feature embeddings. However, existing methods still struggle to achieve satisfactory performance on complex scenes, as they overlook another interrelated issue: inconsistent segmentation. Specifically, due to the presence of inaccurate segmentation, the same object can be segmented inconsistently across two views: e.g., the chair in Fig. 1 is segmented into two parts in \"view 1\" but as a whole in \"view 2\". As existing methods use deterministic feature embeddings to learn 3D instance segmentation, the issue of inconsistent segmentation inherently introduces noise to training data and thus poses a robustness challenge for the models." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 130, + 522, + 482, + 666 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 522, + 482, + 666 + ], + "spans": [ + { + "bbox": [ + 130, + 522, + 482, + 666 + ], + "type": "text", + "content": "In this paper, we introduce an effective probabilistic contrastive fusion solution to collectively address the two issues. For the issue of inconsistent segmentation, we propose to learn probabilistic feature embeddings rather than deterministic feature embeddings used in [3,40]. Our key insight is the development of probabilistic features, which correspond to distributions, allows for the incorporation of uncertainty modeling and enhances robustness to noise. This leads to a stable model optimization and, ultimately, results in a more reliable representation of instance information. Given these benefits, we develop probabilistic feature embeddings based on multivariate Gaussian distributions. Particularly, to enable contrastive learning among different Gaussian distributions, we devise the Probability Product (PP) Kernel [18] to measure the probabilistic feature similarities. Accordingly, a probabilistic clustering algorithm is introduced" + } + ] + } + ], + "index": 6 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "spans": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "type": "text", + "content": "2" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 166, + 91, + 285, + 102 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 166, + 91, + 285, + 102 + ], + "spans": [ + { + "bbox": [ + 166, + 91, + 285, + 102 + ], + "type": "text", + "content": "R. Zhu, S. Qiu, Q. Wu, et al." + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 1 + }, + { + "para_blocks": [ + { + "bbox": [ + 130, + 116, + 482, + 236 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 116, + 482, + 236 + ], + "spans": [ + { + "bbox": [ + 130, + 116, + 482, + 236 + ], + "type": "text", + "content": "in the inference phase to generate consistent panoptic segmentation, using the measured probabilistic feature similarities. For the issue of inconsistent IDs, we devise the contrastive loss to fuse the probabilistic feature embeddings. In addition to formulating the loss term with each single view observation as inspired by the prior work [3], we introduce a novel cross-view constraint to facilitate an effective model training from segmentation results with inconsistent instance IDs. By dynamically exploiting feature pairs from different views, our proposed constraint offers an effective formulation to further enhance the feature consistency of the same 3D object instance across multiple views, and thus improves the panoptic lifting performance." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 130, + 236, + 482, + 331 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 236, + 482, + 331 + ], + "spans": [ + { + "bbox": [ + 130, + 236, + 482, + 331 + ], + "type": "text", + "content": "To evaluate the effectiveness of our method, we conduct experiments on the ScanNet dataset [12] and the Messy Rooms dataset [3]. Furthermore, we conduct experiments to demonstrate the robustness of our probabilistic method to variations in segmentation models and different levels of noise. In addition to the experimental outcomes, our theoretical analysis from an optimization perspective demonstrates that the proposed probabilistic representation can be seen as a more flexible and generalized form compared to the previous deterministic representation [3]." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 146, + 331, + 347, + 342 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 146, + 331, + 347, + 342 + ], + "spans": [ + { + "bbox": [ + 146, + 331, + 347, + 342 + ], + "type": "text", + "content": "Our contributions are summarized as follows:" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 138, + 351, + 481, + 482 + ], + "type": "list", + "angle": 0, + "index": 8, + "blocks": [ + { + "bbox": [ + 138, + 351, + 481, + 398 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 351, + 481, + 398 + ], + "spans": [ + { + "bbox": [ + 138, + 351, + 481, + 398 + ], + "type": "text", + "content": "- We introduce a probabilistic contrastive fusion solution (PCF-Lift) to effectively unproject the 2D panoptic segmentations to the 3D domain by collectively considering the issues of inconsistent segmentation and inconsistent IDs." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 138, + 399, + 481, + 446 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 399, + 481, + 446 + ], + "spans": [ + { + "bbox": [ + 138, + 399, + 481, + 446 + ], + "type": "text", + "content": "- To fuse the probabilistic feature embeddings modeled by multivariate Gaussian distributions, we reformulate the contrastive loss with the Probability Product kernel and propose a novel cross-view constraint to further encourage the multi-view consistency." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 138, + 447, + 481, + 482 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 447, + 481, + 482 + ], + "spans": [ + { + "bbox": [ + 138, + 447, + 481, + 482 + ], + "type": "text", + "content": "- Coupled with a new probabilistic clustering algorithm, our proposed method outperforms the state-of-the-art methods consistently on the ScanNet dataset and the Messy Room dataset." + } + ] + } + ], + "index": 7 + } + ], + "sub_type": "text" + }, + { + "bbox": [ + 132, + 501, + 242, + 514 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 501, + 242, + 514 + ], + "spans": [ + { + "bbox": [ + 132, + 501, + 242, + 514 + ], + "type": "text", + "content": "2 Related Works" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 132, + 526, + 395, + 539 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 526, + 395, + 539 + ], + "spans": [ + { + "bbox": [ + 132, + 526, + 395, + 539 + ], + "type": "text", + "content": "2.1 Traditional 2D and 3D Panoptic Segmentation" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 130, + 545, + 481, + 593 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 545, + 481, + 593 + ], + "spans": [ + { + "bbox": [ + 130, + 545, + 481, + 593 + ], + "type": "text", + "content": "The 2D Panoptic segmentation task was initially introduced in [23]. Despite the notable advancements made by subsequent works [8-10,37,47], it remains challenging to panoptically understand individual images, while avoiding inconsistent instance recognitions across different image views of the scene." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 130, + 594, + 482, + 666 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 594, + 482, + 666 + ], + "spans": [ + { + "bbox": [ + 130, + 594, + 482, + 666 + ], + "type": "text", + "content": "To enhance the panoptic understanding of the real world, 3D panoptic segmentation focuses on segmenting pre-computed 3D structures [15,32,41,51] (e.g., point clouds, voxels) or performing simultaneous 3D reconstruction or segmentation from 2D images [11,33,39]. However, its generalizability is rather limited, largely due to the significant difference in scale between 2D and 3D training data. Given the capabilities of recent 3D reconstruction techniques [6,20,31] and 2D" + } + ] + } + ], + "index": 12 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 192, + 91, + 447, + 102 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 192, + 91, + 447, + 102 + ], + "spans": [ + { + "bbox": [ + 192, + 91, + 447, + 102 + ], + "type": "text", + "content": "PCF-Lift: Panoptic Lifting by Probabilistic Contrastive Fusion" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 474, + 92, + 480, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 474, + 92, + 480, + 100 + ], + "spans": [ + { + "bbox": [ + 474, + 92, + 480, + 100 + ], + "type": "text", + "content": "3" + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 2 + }, + { + "para_blocks": [ + { + "bbox": [ + 132, + 116, + 480, + 140 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 116, + 480, + 140 + ], + "spans": [ + { + "bbox": [ + 132, + 116, + 480, + 140 + ], + "type": "text", + "content": "panoptic segmentation models [9, 50], in this work, we explore 3D panoptic segmentation by leveraging multi-view images without explicit 3D input data." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 132, + 158, + 251, + 169 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 158, + 251, + 169 + ], + "spans": [ + { + "bbox": [ + 132, + 158, + 251, + 169 + ], + "type": "text", + "content": "2.2 Multi-view Fusion" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 132, + 177, + 481, + 381 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 177, + 481, + 381 + ], + "spans": [ + { + "bbox": [ + 132, + 177, + 481, + 381 + ], + "type": "text", + "content": "Recently, researches in 3D reconstruction have made tremendous progress [2,6, 20, 29, 31]. Beyond novel view synthesis, we can utilize 3D reconstruction techniques as a tool to fuse 2D information like semantics and features in the 3D space [5, 14, 17, 21, 25, 35, 43, 46, 48]. For example, Semantic-NeRF [49] learns a semantic field from the 2D semantic segmentation, demonstrating the robustness and effectiveness of the semantic fusion. Later works [21, 25, 43] propose to fuse the unsupervised 2D dense feature for various segmentation and editing applications, showcasing the potential to adapt the 2D zero-shot models to 3D domain. In this work, we study the task of panoptic fusion, integrating both instance and semantic information to obtain a comprehensive understanding of the 3D scene. This task is significantly more challenging than the previous semantic fusion task [49], since we need to incorporate additional instance fusion and achieve instance label consistency. Unlike recent methods (e.g., Panoptic-Lifting [40] and Contrastive Lift [3]), which utilize deterministic feature embeddings to represent instance information, we develop a novel probabilistic solution, actively considering the error-prone and noisy nature of the 2D segmentations, such that we can duly enhance the effectiveness and robustness of instance fusion." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 132, + 399, + 303, + 410 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 399, + 303, + 410 + ], + "spans": [ + { + "bbox": [ + 132, + 399, + 303, + 410 + ], + "type": "text", + "content": "2.3 Probabilistic Representation" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 132, + 418, + 481, + 538 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 418, + 481, + 538 + ], + "spans": [ + { + "bbox": [ + 132, + 418, + 481, + 538 + ], + "type": "text", + "content": "Probabilistic feature embedding is a popular tool employed in various tasks, e.g., image generation [22, 38], normal estimation [1], video understanding [36], point cloud understanding [4], and prototype embeddings for few-shot detection [42]. Motivated by its high capabilities in estimating aleatoric uncertainty [13] and addressing data noise, we innovate a new panoptic-lifting pipeline for 3D panoptic segmentation by formulating new modules based on probabilistic feature embeddings, including a reformulated contrastive loss with the probability product kernel [18] and an effective cross-view constraint to enhance the multi-view consistency. Further, we also design a novel probabilistic clustering algorithm to facilitate the generation of consistent panoptic segmentation." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 132, + 556, + 201, + 568 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 556, + 201, + 568 + ], + "spans": [ + { + "bbox": [ + 132, + 556, + 201, + 568 + ], + "type": "text", + "content": "3 Method" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 132, + 582, + 480, + 628 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 582, + 480, + 628 + ], + "spans": [ + { + "bbox": [ + 132, + 582, + 480, + 628 + ], + "type": "text", + "content": "Given a set of posed images " + }, + { + "bbox": [ + 132, + 582, + 480, + 628 + ], + "type": "inline_equation", + "content": "\\{\\mathcal{I}\\}" + }, + { + "bbox": [ + 132, + 582, + 480, + 628 + ], + "type": "text", + "content": " associated with 2D panoptic segmentation predictions (i.e., semantic masks " + }, + { + "bbox": [ + 132, + 582, + 480, + 628 + ], + "type": "inline_equation", + "content": "\\{\\mathcal{H}\\}" + }, + { + "bbox": [ + 132, + 582, + 480, + 628 + ], + "type": "text", + "content": " and instance masks " + }, + { + "bbox": [ + 132, + 582, + 480, + 628 + ], + "type": "inline_equation", + "content": "\\{\\mathcal{K}\\}" + }, + { + "bbox": [ + 132, + 582, + 480, + 628 + ], + "type": "text", + "content": ") generated by 2D segmentation models, our goal is to learn accurate 3D panoptic fields that can be rendered into consistent panoptic segmentation results over different views." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 132, + 630, + 480, + 665 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 630, + 480, + 665 + ], + "spans": [ + { + "bbox": [ + 132, + 630, + 480, + 665 + ], + "type": "text", + "content": "The overview of PCF-Lift is illustrated in Fig. 2. Specifically, the 3D panoptic fields include a semantic field, an instance field, a density field, and a color field. Particularly, considering the significance of instance-related issues, we propose" + } + ] + } + ], + "index": 9 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "spans": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "type": "text", + "content": "4" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 166, + 91, + 285, + 102 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 166, + 91, + 285, + 102 + ], + "spans": [ + { + "bbox": [ + 166, + 91, + 285, + 102 + ], + "type": "text", + "content": "R. Zhu, S. Qiu, Q. Wu, et al." + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 3 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 153, + 114, + 462, + 274 + ], + "blocks": [ + { + "bbox": [ + 153, + 114, + 462, + 274 + ], + "lines": [ + { + "bbox": [ + 153, + 114, + 462, + 274 + ], + "spans": [ + { + "bbox": [ + 153, + 114, + 462, + 274 + ], + "type": "image", + "image_path": "f6f4e7af24ecfec46155369fb1480767636138e609fe34faa13cdac3e3f754a0.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 130, + 293, + 482, + 426 + ], + "lines": [ + { + "bbox": [ + 130, + 293, + 482, + 426 + ], + "spans": [ + { + "bbox": [ + 130, + 293, + 482, + 426 + ], + "type": "text", + "content": "Fig. 2: Overview of PCF-Lift. The 3D panoptic fields include a semantic field, an instance field, a density field, and a color field. To solve the instance-related issues, we propose to learn probabilistic feature embeddings in the instance field (see Sec. 3.1). During the training phase, given two camera views, we can render the probabilistic feature maps from the instance field via volume rendering. To optimize the probabilistic instance field, we devise the probabilistic contrastive loss with Probability Product (PP) kernel [18], and propose a cross-view constraint to further enhance the feature consistency from different views (see Sec. 3.2). Similarly, we can render the semantic and color predictions, and adopt photometric loss and cross-entropy loss to optimize the semantic field, the density field, and the color field. During the inference phase, we design a novel multi-view object association (MVOA) algorithm for the generation of consistent panoptic segmentations (see Sec. 3.3)." + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_caption" + } + ], + "index": 2 + }, + { + "bbox": [ + 130, + 449, + 482, + 605 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 449, + 482, + 605 + ], + "spans": [ + { + "bbox": [ + 130, + 449, + 482, + 605 + ], + "type": "text", + "content": "to learn probabilistic feature embeddings in the instances field (see Sec. 3.1). During the training phase, we jointly train the whole panoptic fields. Given two camera views, we can render the probabilistic feature maps from the instance field via volume rendering [19]. To optimize the probabilistic feature embeddings, we develop the contrastive loss with the Probability Product kernel [18], and propose a cross-view constraint to further enhance the feature consistency from different views (see Sec. 3.2). Similarly, the semantic predictions and color predictions are also rendered via volume rendering, where the semantic field, the density field, and the color field are optimized by the photometric loss and cross-entropy loss [3,6,40,49]. In the inference phase, we design a probabilistic clustering (i.e., multi-view object association) algorithm to effectively identify the prototype features of underlying 3D object instances for the generation of consistent panoptic segmentation results across any given views (see Sec. 3.3)." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 131, + 622, + 331, + 634 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 131, + 622, + 331, + 634 + ], + "spans": [ + { + "bbox": [ + 131, + 622, + 331, + 634 + ], + "type": "text", + "content": "3.1 Probabilistic Feature Embeddings" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 130, + 641, + 481, + 665 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 641, + 481, + 665 + ], + "spans": [ + { + "bbox": [ + 130, + 641, + 481, + 665 + ], + "type": "text", + "content": "To provide a robust instance representation, we propose to learn probabilistic feature embeddings in the instance field, which maps each 3D point to a random" + } + ] + } + ], + "index": 6 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 192, + 91, + 447, + 102 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 192, + 91, + 447, + 102 + ], + "spans": [ + { + "bbox": [ + 192, + 91, + 447, + 102 + ], + "type": "text", + "content": "PCF-Lift: Panoptic Lifting by Probabilistic Contrastive Fusion" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 474, + 92, + 480, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 474, + 92, + 480, + 100 + ], + "spans": [ + { + "bbox": [ + 474, + 92, + 480, + 100 + ], + "type": "text", + "content": "5" + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 4 + }, + { + "para_blocks": [ + { + "bbox": [ + 130, + 116, + 482, + 248 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 116, + 482, + 248 + ], + "spans": [ + { + "bbox": [ + 130, + 116, + 482, + 248 + ], + "type": "text", + "content": "variable " + }, + { + "bbox": [ + 130, + 116, + 482, + 248 + ], + "type": "inline_equation", + "content": "\\mathcal{F}" + }, + { + "bbox": [ + 130, + 116, + 482, + 248 + ], + "type": "text", + "content": ", modeled as an " + }, + { + "bbox": [ + 130, + 116, + 482, + 248 + ], + "type": "inline_equation", + "content": "N" + }, + { + "bbox": [ + 130, + 116, + 482, + 248 + ], + "type": "text", + "content": "-dimensional multivariate Gaussian distribution " + }, + { + "bbox": [ + 130, + 116, + 482, + 248 + ], + "type": "inline_equation", + "content": "\\mathcal{F} \\sim \\mathcal{N}(\\mu, \\Sigma)" + }, + { + "bbox": [ + 130, + 116, + 482, + 248 + ], + "type": "text", + "content": ". Here, " + }, + { + "bbox": [ + 130, + 116, + 482, + 248 + ], + "type": "inline_equation", + "content": "\\mu" + }, + { + "bbox": [ + 130, + 116, + 482, + 248 + ], + "type": "text", + "content": " is the mean vector, indicating the central feature values and " + }, + { + "bbox": [ + 130, + 116, + 482, + 248 + ], + "type": "inline_equation", + "content": "\\Sigma" + }, + { + "bbox": [ + 130, + 116, + 482, + 248 + ], + "type": "text", + "content": " is a diagonal covariance matrix " + }, + { + "bbox": [ + 130, + 116, + 482, + 248 + ], + "type": "inline_equation", + "content": "\\mathrm{diag}(\\sigma^2)" + }, + { + "bbox": [ + 130, + 116, + 482, + 248 + ], + "type": "text", + "content": " with " + }, + { + "bbox": [ + 130, + 116, + 482, + 248 + ], + "type": "inline_equation", + "content": "\\sigma^2 = (\\sigma^{(1)^2}, \\sigma^{(2)^2}, \\dots, \\sigma^{(N)^2})" + }, + { + "bbox": [ + 130, + 116, + 482, + 248 + ], + "type": "text", + "content": ". Concretely, for any given query point " + }, + { + "bbox": [ + 130, + 116, + 482, + 248 + ], + "type": "inline_equation", + "content": "\\mathbf{x} \\in \\mathbb{R}^3" + }, + { + "bbox": [ + 130, + 116, + 482, + 248 + ], + "type": "text", + "content": ", the instance field predicts " + }, + { + "bbox": [ + 130, + 116, + 482, + 248 + ], + "type": "inline_equation", + "content": "(\\mu, \\sigma^2) \\in \\mathbb{R}^{2N}" + }, + { + "bbox": [ + 130, + 116, + 482, + 248 + ], + "type": "text", + "content": ". Similar to the rendering of the color field, for each pixel in a given camera view, we can render its corresponding Gaussian distribution feature via volume rendering. A significant advantage of using probabilistic features rather than deterministic features is the ability of Gaussian distributions that assign different covariance values for uncertainty modeling. This property is crucial as it aids in reducing the impact of noise, thereby enhancing the robustness and accuracy of feature embeddings in representing complex 3D scenes." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 130, + 249, + 482, + 320 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 249, + 482, + 320 + ], + "spans": [ + { + "bbox": [ + 130, + 249, + 482, + 320 + ], + "type": "text", + "content": "Intuitively, the similarity between two rendered Gaussian distributions indicates whether the corresponding pixels belong to the same instance. To quantify the similarities between different Gaussian distributions, we employ the Probability Product (PP) kernel [18] " + }, + { + "bbox": [ + 130, + 249, + 482, + 320 + ], + "type": "inline_equation", + "content": "K_{\\rho}" + }, + { + "bbox": [ + 130, + 249, + 482, + 320 + ], + "type": "text", + "content": ". Specifically, given two Gaussian distributions " + }, + { + "bbox": [ + 130, + 249, + 482, + 320 + ], + "type": "inline_equation", + "content": "\\mathcal{F}_i \\sim \\mathcal{N}(\\mu_i,\\Sigma_i)" + }, + { + "bbox": [ + 130, + 249, + 482, + 320 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 130, + 249, + 482, + 320 + ], + "type": "inline_equation", + "content": "\\mathcal{F}_j \\sim \\mathcal{N}(\\mu_j,\\Sigma_j)" + }, + { + "bbox": [ + 130, + 249, + 482, + 320 + ], + "type": "text", + "content": ", the corresponding kernel can be explicitly formulated as" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 136, + 326, + 485, + 361 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 136, + 326, + 485, + 361 + ], + "spans": [ + { + "bbox": [ + 136, + 326, + 485, + 361 + ], + "type": "interline_equation", + "content": "K _ {\\rho} \\left(\\mathcal {F} _ {i}, \\mathcal {F} _ {j}\\right) = \\left(\\prod_ {d = 1} ^ {N} \\left(\\frac {\\sigma_ {i} ^ {(d) ^ {2}}}{\\sigma_ {j} ^ {(d) ^ {2}}} + \\frac {\\sigma_ {j} ^ {(d) ^ {2}}}{\\sigma_ {i} ^ {(d) ^ {2}}}\\right) / 2\\right) ^ {- \\frac {1}{2}} \\exp \\left(- \\sum_ {d = 1} ^ {N} \\left(\\mu_ {i} ^ {(d)} - \\mu_ {j} ^ {(d)}\\right) ^ {2} / 4 \\left(\\sigma_ {i} ^ {(d) ^ {2}} + \\sigma_ {j} ^ {(d) ^ {2}}\\right)\\right). \\tag {1}", + "image_path": "8066d549b358e3bb84e1b4b3e8dbc46dbd9a46861677600817a9a3e149221766.jpg" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 130, + 362, + 482, + 399 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 362, + 482, + 399 + ], + "spans": [ + { + "bbox": [ + 130, + 362, + 482, + 399 + ], + "type": "text", + "content": "Note that the results produced by the PP kernel fall within the range of 0 (minimal similarity) to 1 (maximal similarity), the same as the Radial Basis Function (RBF) kernel used in the deterministic method [3]." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 131, + 415, + 378, + 427 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 131, + 415, + 378, + 427 + ], + "spans": [ + { + "bbox": [ + 131, + 415, + 378, + 427 + ], + "type": "text", + "content": "3.2 Training Probabilistic Feature Embeddings" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 130, + 433, + 482, + 540 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 433, + 482, + 540 + ], + "spans": [ + { + "bbox": [ + 130, + 433, + 482, + 540 + ], + "type": "text", + "content": "Probabilistic Contrastive Loss. During the training phase, we leverage the contrastive learning technique to optimize the probabilistic feature embeddings. Specifically, the contrastive loss used in [7, 27, 34, 45] tends to increase the predicted feature similarities of predefined positive pairs, while decreasing the similarities of negative pairs for training. In our work, a positive pair is defined as two pixels belonging to the same instance, while a negative pair denotes two pixels of different instances within a single image, following the previous method [3]. Since we basically learn probabilistic features, the PP kernel is exploited to measure the features similarity. In general, the loss can be formulated as" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 175, + 548, + 481, + 578 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 175, + 548, + 481, + 578 + ], + "spans": [ + { + "bbox": [ + 175, + 548, + 481, + 578 + ], + "type": "interline_equation", + "content": "\\mathcal {L} _ {\\mathrm {p i x e l - c o n t r a}} = - \\frac {1}{| \\varOmega |} \\sum_ {u \\in \\varOmega} \\log \\frac {\\sum_ {u ^ {\\prime} \\in \\varOmega} \\mathbf {1} _ {(u , u ^ {\\prime})} \\exp \\left(K _ {\\rho} \\left(\\mathcal {F} _ {u} , \\mathcal {F} _ {u ^ {\\prime}}\\right)\\right)}{\\sum_ {u ^ {\\prime} \\in \\varOmega} \\exp \\left(K _ {\\rho} \\left(\\mathcal {F} _ {u} , \\mathcal {F} _ {u ^ {\\prime}}\\right)\\right)}, \\tag {2}", + "image_path": "748c376cc903ad8a4bb65b19279f35d0403282ab68a6f6126e3e65f948117bc0.jpg" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 185, + 586, + 481, + 616 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 185, + 586, + 481, + 616 + ], + "spans": [ + { + "bbox": [ + 185, + 586, + 481, + 616 + ], + "type": "interline_equation", + "content": "\\mathcal {L} _ {\\text {c o n c e n}} = - \\frac {1}{| \\varOmega |} \\sum_ {u \\in \\varOmega} \\log \\left(K _ {\\rho} \\left(\\mathcal {F} _ {u}, \\frac {\\sum_ {u ^ {\\prime} \\in \\varOmega} \\mathbf {1} _ {(u , u ^ {\\prime})} \\mathcal {F} _ {u ^ {\\prime}}}{\\sum_ {u ^ {\\prime} \\in \\varOmega} \\mathbf {1} _ {(u , u ^ {\\prime})}}\\right)\\right), \\tag {3}", + "image_path": "d1c8ccba74964377aa1859306ad433750ecedea52634273c5d3491cee7d2fc85.jpg" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 228, + 625, + 481, + 637 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 228, + 625, + 481, + 637 + ], + "spans": [ + { + "bbox": [ + 228, + 625, + 481, + 637 + ], + "type": "interline_equation", + "content": "\\text {a n d} \\mathcal {L} _ {\\text {c o n t r a}} = \\mathcal {L} _ {\\text {p i x e l - c o n t r a}} + \\mathcal {L} _ {\\text {c o n c e n}}, \\tag {4}", + "image_path": "ac33812ae6bf92a73c3dab5c65b795ad78aca9c4be1f081be99e1b2308ebacba.jpg" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 130, + 641, + 482, + 666 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 641, + 482, + 666 + ], + "spans": [ + { + "bbox": [ + 130, + 641, + 482, + 666 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 130, + 641, + 482, + 666 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_{\\mathrm{pixel - contra}}" + }, + { + "bbox": [ + 130, + 641, + 482, + 666 + ], + "type": "text", + "content": " is the pixel-wise contrastive loss, " + }, + { + "bbox": [ + 130, + 641, + 482, + 666 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_{\\mathrm{concen}}" + }, + { + "bbox": [ + 130, + 641, + 482, + 666 + ], + "type": "text", + "content": " is the concentrate loss term, " + }, + { + "bbox": [ + 130, + 641, + 482, + 666 + ], + "type": "inline_equation", + "content": "\\mathbf{1}" + }, + { + "bbox": [ + 130, + 641, + 482, + 666 + ], + "type": "text", + "content": " is the indicator function of positive pairs, " + }, + { + "bbox": [ + 130, + 641, + 482, + 666 + ], + "type": "inline_equation", + "content": "\\Omega" + }, + { + "bbox": [ + 130, + 641, + 482, + 666 + ], + "type": "text", + "content": " is the set of pixel sample, " + }, + { + "bbox": [ + 130, + 641, + 482, + 666 + ], + "type": "inline_equation", + "content": "\\mathcal{F}_u" + } + ] + } + ], + "index": 11 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "spans": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "type": "text", + "content": "6" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 166, + 91, + 286, + 102 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 166, + 91, + 286, + 102 + ], + "spans": [ + { + "bbox": [ + 166, + 91, + 286, + 102 + ], + "type": "text", + "content": "R. Zhu, S. Qiu, Q. Wu, et al." + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 5 + }, + { + "para_blocks": [ + { + "bbox": [ + 130, + 116, + 482, + 201 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 116, + 482, + 201 + ], + "spans": [ + { + "bbox": [ + 130, + 116, + 482, + 201 + ], + "type": "text", + "content": "and " + }, + { + "bbox": [ + 130, + 116, + 482, + 201 + ], + "type": "inline_equation", + "content": "\\mathcal{F}_{u'}" + }, + { + "bbox": [ + 130, + 116, + 482, + 201 + ], + "type": "text", + "content": " are the rendered features for pixels " + }, + { + "bbox": [ + 130, + 116, + 482, + 201 + ], + "type": "inline_equation", + "content": "u" + }, + { + "bbox": [ + 130, + 116, + 482, + 201 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 130, + 116, + 482, + 201 + ], + "type": "inline_equation", + "content": "u' \\in \\Omega" + }, + { + "bbox": [ + 130, + 116, + 482, + 201 + ], + "type": "text", + "content": " via volume rendering, respectively. Different from previous method [3], we devise probabilistic similarity kernels to calculate a more effective contrastive loss for model training. By optimizing the loss Eq. (4), we can learn a more expressive 3D instance field for robust feature representations, as further analyzed in Sec. 3.4. To avoid the neural network from generating large covariances everywhere, we use an additional regularization term, " + }, + { + "bbox": [ + 130, + 116, + 482, + 201 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_{\\mathrm{reg}} := \\log \\left( \\prod_{d=1}^{N} \\sigma^{(d)^2} \\right)" + }, + { + "bbox": [ + 130, + 116, + 482, + 201 + ], + "type": "text", + "content": ", to penalize the large covariances." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 130, + 217, + 482, + 313 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 217, + 482, + 313 + ], + "spans": [ + { + "bbox": [ + 130, + 217, + 482, + 313 + ], + "type": "text", + "content": "Cross-view Constraint. To better train the probabilistic feature embeddings, we propose a cross-view constraint to enhance feature consistency for the same object across different views. Given the rendered learned probabilistic feature embedding sets " + }, + { + "bbox": [ + 130, + 217, + 482, + 313 + ], + "type": "inline_equation", + "content": "\\{\\mathcal{F}^m\\}" + }, + { + "bbox": [ + 130, + 217, + 482, + 313 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 130, + 217, + 482, + 313 + ], + "type": "inline_equation", + "content": "\\{\\mathcal{F}^n\\}" + }, + { + "bbox": [ + 130, + 217, + 482, + 313 + ], + "type": "text", + "content": " for two different views, we devise the PP kernel " + }, + { + "bbox": [ + 130, + 217, + 482, + 313 + ], + "type": "inline_equation", + "content": "K_{\\rho}" + }, + { + "bbox": [ + 130, + 217, + 482, + 313 + ], + "type": "text", + "content": " and a predefined threshold " + }, + { + "bbox": [ + 130, + 217, + 482, + 313 + ], + "type": "inline_equation", + "content": "\\tau" + }, + { + "bbox": [ + 130, + 217, + 482, + 313 + ], + "type": "text", + "content": " to collect positive pairs " + }, + { + "bbox": [ + 130, + 217, + 482, + 313 + ], + "type": "inline_equation", + "content": "\\mathcal{P}" + }, + { + "bbox": [ + 130, + 217, + 482, + 313 + ], + "type": "text", + "content": " that belong to the same object: " + }, + { + "bbox": [ + 130, + 217, + 482, + 313 + ], + "type": "inline_equation", + "content": "\\mathcal{P} = \\left\\{(\\mathcal{F}_r,\\mathcal{F}_s)\\mid K_\\rho (\\mathcal{F}_r,\\mathcal{F}_s) > \\tau ,\\mathcal{F}_r\\in \\{\\mathcal{F}^m\\} ,\\mathcal{F}_s\\in \\{\\mathcal{F}^n\\} \\right\\}" + }, + { + "bbox": [ + 130, + 217, + 482, + 313 + ], + "type": "text", + "content": ". Moreover, as we expect to maximize the similarities between the features that belong to the same 3D object in different views, the cross-view constraint is defined as:" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 217, + 322, + 481, + 352 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 217, + 322, + 481, + 352 + ], + "spans": [ + { + "bbox": [ + 217, + 322, + 481, + 352 + ], + "type": "interline_equation", + "content": "\\mathcal {L} _ {\\text {c r o s s}} = - \\frac {1}{| \\mathcal {P} |} \\sum_ {\\left(\\mathcal {F} _ {r}, \\mathcal {F} _ {s}\\right) \\in \\mathcal {P}} \\log \\left(K _ {\\rho} \\left(\\mathcal {F} _ {r}, \\mathcal {F} _ {s}\\right)\\right). \\tag {5}", + "image_path": "308c115daca8d7b7d5091584b437363a6341438754e83b119c9189ffc2f01766.jpg" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 131, + 355, + 482, + 392 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 131, + 355, + 482, + 392 + ], + "spans": [ + { + "bbox": [ + 131, + 355, + 482, + 392 + ], + "type": "text", + "content": "In practice, we use the cross-view constraint only in a later optimization stage (i.e., the last few epochs), where the feature space is sufficiently meaningful to provide reliable positive pairs. The overall loss is formulated as" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 226, + 400, + 481, + 414 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 226, + 400, + 481, + 414 + ], + "spans": [ + { + "bbox": [ + 226, + 400, + 481, + 414 + ], + "type": "interline_equation", + "content": "\\mathcal {L} = \\mathcal {L} _ {\\text {c o n t r a}} + w _ {\\text {c r o s s}} \\mathcal {L} _ {\\text {c r o s s}} + w _ {\\text {r e g}} \\mathcal {L} _ {\\text {r e g}}, \\tag {6}", + "image_path": "f6c0bdd8cc7a70eb1e6afb38e0c1962943ca6636b7245440116b0eca1d3bc68f.jpg" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 131, + 421, + 360, + 434 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 131, + 421, + 360, + 434 + ], + "spans": [ + { + "bbox": [ + 131, + 421, + 360, + 434 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 131, + 421, + 360, + 434 + ], + "type": "inline_equation", + "content": "w_{\\mathrm{cross}}" + }, + { + "bbox": [ + 131, + 421, + 360, + 434 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 131, + 421, + 360, + 434 + ], + "type": "inline_equation", + "content": "w_{\\mathrm{reg}}" + }, + { + "bbox": [ + 131, + 421, + 360, + 434 + ], + "type": "text", + "content": " are weight hyper-parameters." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 131, + 449, + 315, + 462 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 131, + 449, + 315, + 462 + ], + "spans": [ + { + "bbox": [ + 131, + 449, + 315, + 462 + ], + "type": "text", + "content": "3.3 Multi-view Object Association" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 130, + 469, + 482, + 613 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 469, + 482, + 613 + ], + "spans": [ + { + "bbox": [ + 130, + 469, + 482, + 613 + ], + "type": "text", + "content": "During the inference phase, to facilitate consistent panoptic segmentations, we introduce a novel clustering algorithm, named the multi-view object association (MVOA) algorithm. The algorithm aims to extract the prototype feature set from the learned instance field for the assignment of instance labels. In general, through volume rendering, we obtain a large-scale rendered per-pixel probabilistic feature set " + }, + { + "bbox": [ + 130, + 469, + 482, + 613 + ], + "type": "inline_equation", + "content": "\\{\\mathcal{F}\\}" + }, + { + "bbox": [ + 130, + 469, + 482, + 613 + ], + "type": "text", + "content": " of the training views, as the input of our algorithm. Considering the efficiency issue, we first conduct an instance grouping operation to gather a smaller feature set " + }, + { + "bbox": [ + 130, + 469, + 482, + 613 + ], + "type": "inline_equation", + "content": "\\mathcal{C}" + }, + { + "bbox": [ + 130, + 469, + 482, + 613 + ], + "type": "text", + "content": ", with the help of inconsistent instance segmentation masks " + }, + { + "bbox": [ + 130, + 469, + 482, + 613 + ], + "type": "inline_equation", + "content": "\\{\\mathcal{K}\\}" + }, + { + "bbox": [ + 130, + 469, + 482, + 613 + ], + "type": "text", + "content": " obtained from training views. Then, we construct the probabilistic similarity graph based on " + }, + { + "bbox": [ + 130, + 469, + 482, + 613 + ], + "type": "inline_equation", + "content": "\\mathcal{C}" + }, + { + "bbox": [ + 130, + 469, + 482, + 613 + ], + "type": "text", + "content": " and design a multi-view matching process to collect the prototype feature set " + }, + { + "bbox": [ + 130, + 469, + 482, + 613 + ], + "type": "inline_equation", + "content": "\\mathcal{D}" + }, + { + "bbox": [ + 130, + 469, + 482, + 613 + ], + "type": "text", + "content": ". The details of instance grouping and multi-view matching operations are introduced as follows." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 130, + 629, + 482, + 667 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 629, + 482, + 667 + ], + "spans": [ + { + "bbox": [ + 130, + 629, + 482, + 667 + ], + "type": "text", + "content": "Instance Grouping. We group the feature set " + }, + { + "bbox": [ + 130, + 629, + 482, + 667 + ], + "type": "inline_equation", + "content": "\\{\\mathcal{F}\\}" + }, + { + "bbox": [ + 130, + 629, + 482, + 667 + ], + "type": "text", + "content": " from each individual view. For the view of " + }, + { + "bbox": [ + 130, + 629, + 482, + 667 + ], + "type": "inline_equation", + "content": "l" + }, + { + "bbox": [ + 130, + 629, + 482, + 667 + ], + "type": "text", + "content": "-th image, we collect the pixels that belong to instance " + }, + { + "bbox": [ + 130, + 629, + 482, + 667 + ], + "type": "inline_equation", + "content": "p" + }, + { + "bbox": [ + 130, + 629, + 482, + 667 + ], + "type": "text", + "content": " into the set " + }, + { + "bbox": [ + 130, + 629, + 482, + 667 + ], + "type": "inline_equation", + "content": "\\{i : \\mathcal{K}_l^i = p\\}" + }, + { + "bbox": [ + 130, + 629, + 482, + 667 + ], + "type": "text", + "content": ", and the corresponding probabilistic feature set is denoted as" + } + ] + } + ], + "index": 10 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 192, + 90, + 448, + 102 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 192, + 90, + 448, + 102 + ], + "spans": [ + { + "bbox": [ + 192, + 90, + 448, + 102 + ], + "type": "text", + "content": "PCF-Lift: Panoptic Lifting by Probabilistic Contrastive Fusion" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 474, + 91, + 481, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 474, + 91, + 481, + 100 + ], + "spans": [ + { + "bbox": [ + 474, + 91, + 481, + 100 + ], + "type": "text", + "content": "7" + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 6 + }, + { + "para_blocks": [ + { + "bbox": [ + 130, + 115, + 482, + 216 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 115, + 482, + 216 + ], + "spans": [ + { + "bbox": [ + 130, + 115, + 482, + 216 + ], + "type": "inline_equation", + "content": "\\{\\mathcal{F}_l^i:\\mathcal{K}_l^i = p\\}" + }, + { + "bbox": [ + 130, + 115, + 482, + 216 + ], + "type": "text", + "content": ". Then, we group " + }, + { + "bbox": [ + 130, + 115, + 482, + 216 + ], + "type": "inline_equation", + "content": "\\{\\mathcal{F}_l^i:\\mathcal{K}_l^i = p\\}" + }, + { + "bbox": [ + 130, + 115, + 482, + 216 + ], + "type": "text", + "content": " to a single probabilistic feature by calculating the average: " + }, + { + "bbox": [ + 130, + 115, + 482, + 216 + ], + "type": "inline_equation", + "content": "\\mathcal{C}_l^p = \\sum_i\\mathcal{F}_l^i /|\\{i:\\mathcal{K}_l^i = p\\} |." + }, + { + "bbox": [ + 130, + 115, + 482, + 216 + ], + "type": "text", + "content": " Concurrently, we model the score quantity " + }, + { + "bbox": [ + 130, + 115, + 482, + 216 + ], + "type": "inline_equation", + "content": "\\mathcal{S}_l^p" + }, + { + "bbox": [ + 130, + 115, + 482, + 216 + ], + "type": "text", + "content": " for " + }, + { + "bbox": [ + 130, + 115, + 482, + 216 + ], + "type": "inline_equation", + "content": "\\mathcal{C}_l^p" + }, + { + "bbox": [ + 130, + 115, + 482, + 216 + ], + "type": "text", + "content": " as an indicator of feature concentration, which aids in the subsequent multi-view matching process. The indicator function " + }, + { + "bbox": [ + 130, + 115, + 482, + 216 + ], + "type": "inline_equation", + "content": "\\varPhi(\\mathcal{C}_l^p)" + }, + { + "bbox": [ + 130, + 115, + 482, + 216 + ], + "type": "text", + "content": " averages the PP kernel similarities between " + }, + { + "bbox": [ + 130, + 115, + 482, + 216 + ], + "type": "inline_equation", + "content": "\\mathcal{C}_l^p" + }, + { + "bbox": [ + 130, + 115, + 482, + 216 + ], + "type": "text", + "content": " and all features in " + }, + { + "bbox": [ + 130, + 115, + 482, + 216 + ], + "type": "inline_equation", + "content": "\\{\\mathcal{F}_l^i:\\mathcal{K}_l^i = p\\}" + }, + { + "bbox": [ + 130, + 115, + 482, + 216 + ], + "type": "text", + "content": ". Mathematically, " + }, + { + "bbox": [ + 130, + 115, + 482, + 216 + ], + "type": "inline_equation", + "content": "\\varPhi(\\mathcal{C}_l^p)" + }, + { + "bbox": [ + 130, + 115, + 482, + 216 + ], + "type": "text", + "content": " follows: " + }, + { + "bbox": [ + 130, + 115, + 482, + 216 + ], + "type": "inline_equation", + "content": "\\mathcal{S}_l^p = \\varPhi(\\mathcal{C}_l^p) = \\sum_{\\{i:\\mathcal{K}_l^i = p\\}}K_\\rho (\\mathcal{C}_l^p,\\mathcal{F}_l^i) / |\\{i:\\mathcal{K}_l^i = p\\} |." + }, + { + "bbox": [ + 130, + 115, + 482, + 216 + ], + "type": "text", + "content": " By applying this across all observed images, we obtain the grouped feature set " + }, + { + "bbox": [ + 130, + 115, + 482, + 216 + ], + "type": "inline_equation", + "content": "\\mathcal{C}" + }, + { + "bbox": [ + 130, + 115, + 482, + 216 + ], + "type": "text", + "content": " and the corresponding score set " + }, + { + "bbox": [ + 130, + 115, + 482, + 216 + ], + "type": "inline_equation", + "content": "\\mathcal{S}" + }, + { + "bbox": [ + 130, + 115, + 482, + 216 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 130, + 231, + 483, + 352 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 231, + 483, + 352 + ], + "spans": [ + { + "bbox": [ + 130, + 231, + 483, + 352 + ], + "type": "text", + "content": "Multi-view Matching. To extract the structural relationships among the features, we construct an unoriented similarity graph " + }, + { + "bbox": [ + 130, + 231, + 483, + 352 + ], + "type": "inline_equation", + "content": "G = (\\mathcal{C}, E)" + }, + { + "bbox": [ + 130, + 231, + 483, + 352 + ], + "type": "text", + "content": ", where the grouped feature set " + }, + { + "bbox": [ + 130, + 231, + 483, + 352 + ], + "type": "inline_equation", + "content": "\\mathcal{C}" + }, + { + "bbox": [ + 130, + 231, + 483, + 352 + ], + "type": "text", + "content": " represents the nodes and " + }, + { + "bbox": [ + 130, + 231, + 483, + 352 + ], + "type": "inline_equation", + "content": "E" + }, + { + "bbox": [ + 130, + 231, + 483, + 352 + ], + "type": "text", + "content": " encompasses the edges that indicate quantitative feature similarities measured by the PP Kernel: " + }, + { + "bbox": [ + 130, + 231, + 483, + 352 + ], + "type": "inline_equation", + "content": "E_{} = E_{} = K_{\\rho}(\\mathcal{C}_g, \\mathcal{C}_h)" + }, + { + "bbox": [ + 130, + 231, + 483, + 352 + ], + "type": "text", + "content": ", where " + }, + { + "bbox": [ + 130, + 231, + 483, + 352 + ], + "type": "inline_equation", + "content": "\\mathcal{C}_g, \\mathcal{C}_h \\in \\mathcal{C}" + }, + { + "bbox": [ + 130, + 231, + 483, + 352 + ], + "type": "text", + "content": ". Then, we extract the prototype feature set " + }, + { + "bbox": [ + 130, + 231, + 483, + 352 + ], + "type": "inline_equation", + "content": "\\mathcal{D}" + }, + { + "bbox": [ + 130, + 231, + 483, + 352 + ], + "type": "text", + "content": " from the graph " + }, + { + "bbox": [ + 130, + 231, + 483, + 352 + ], + "type": "inline_equation", + "content": "G" + }, + { + "bbox": [ + 130, + 231, + 483, + 352 + ], + "type": "text", + "content": " and score set " + }, + { + "bbox": [ + 130, + 231, + 483, + 352 + ], + "type": "inline_equation", + "content": "S" + }, + { + "bbox": [ + 130, + 231, + 483, + 352 + ], + "type": "text", + "content": " by a greedy procedure, which is akin to the classical greedy algorithm known as non-maximum suppression (NMS). The pseudo-code of our proposed MVOA algorithm is presented in Algo. 1. In practice, the selection of hyper-parameter " + }, + { + "bbox": [ + 130, + 231, + 483, + 352 + ], + "type": "inline_equation", + "content": "\\mathcal{T}" + }, + { + "bbox": [ + 130, + 231, + 483, + 352 + ], + "type": "text", + "content": " is based on the average similarity computed across grouped feature pairs identified within each view." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 130, + 368, + 483, + 489 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 368, + 483, + 489 + ], + "spans": [ + { + "bbox": [ + 130, + 368, + 483, + 489 + ], + "type": "text", + "content": "Generating Panoptic Segmentation Masks. We first follow the MVOA algorithm to collect the prototype feature set " + }, + { + "bbox": [ + 130, + 368, + 483, + 489 + ], + "type": "inline_equation", + "content": "\\mathcal{D}" + }, + { + "bbox": [ + 130, + 368, + 483, + 489 + ], + "type": "text", + "content": ". For each view, including the novel view, we then generate the semantic labels from the learned semantic field to differentiate between the background and foreground. Coupled with a feature map rendered from the probabilistic instance field, for each foreground pixel, we finally determine its corresponding instance label by assigning the candidate index from the set " + }, + { + "bbox": [ + 130, + 368, + 483, + 489 + ], + "type": "inline_equation", + "content": "\\mathcal{D}" + }, + { + "bbox": [ + 130, + 368, + 483, + 489 + ], + "type": "text", + "content": " that exhibits the highest similarity to the rendered feature. In practice, the MVOA algorithm needs to be conducted only once, while the extracted prototype features " + }, + { + "bbox": [ + 130, + 368, + 483, + 489 + ], + "type": "inline_equation", + "content": "\\mathcal{D}" + }, + { + "bbox": [ + 130, + 368, + 483, + 489 + ], + "type": "text", + "content": " will be used for generating the panoptic segmentation masks of all test views." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 131, + 505, + 264, + 517 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 131, + 505, + 264, + 517 + ], + "spans": [ + { + "bbox": [ + 131, + 505, + 264, + 517 + ], + "type": "text", + "content": "3.4 Theoretical Analysis" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 130, + 525, + 482, + 561 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 525, + 482, + 561 + ], + "spans": [ + { + "bbox": [ + 130, + 525, + 482, + 561 + ], + "type": "text", + "content": "Corollary 1 If the covariances of the given Gaussian distributions are isotropic and fixed, i.e., " + }, + { + "bbox": [ + 130, + 525, + 482, + 561 + ], + "type": "inline_equation", + "content": "\\Sigma_{i} = \\Sigma_{j} = \\sigma I" + }, + { + "bbox": [ + 130, + 525, + 482, + 561 + ], + "type": "text", + "content": ", where " + }, + { + "bbox": [ + 130, + 525, + 482, + 561 + ], + "type": "inline_equation", + "content": "\\sigma" + }, + { + "bbox": [ + 130, + 525, + 482, + 561 + ], + "type": "text", + "content": " is a constant scalar, the probability product kernel can be simplified to an RBF kernel." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 130, + 570, + 482, + 667 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 570, + 482, + 667 + ], + "spans": [ + { + "bbox": [ + 130, + 570, + 482, + 667 + ], + "type": "text", + "content": "The primary distinction between our probabilistic method and the prior deterministic method [3] is the choice of a similarity kernel. Specifically, our probabilistic method exploits the PP kernel, which offers higher flexibility and stronger expressive capability by adjusting the Gaussian mean and covariance compared to the RBF kernel employed in the deterministic method, as demonstrated in Fig 3. The RBF kernel can be considered as a degenerate PP kernel when applied to a Gaussian distribution with an isotropic and fixed covariance, as stated in Corollary 1. From this perspective, the proposed probabilistic method is a" + } + ] + } + ], + "index": 7 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "spans": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "type": "text", + "content": "8" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 166, + 91, + 285, + 102 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 166, + 91, + 285, + 102 + ], + "spans": [ + { + "bbox": [ + 166, + 91, + 285, + 102 + ], + "type": "text", + "content": "R. Zhu, S. Qiu, Q. Wu, et al." + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 7 + }, + { + "para_blocks": [ + { + "type": "code", + "bbox": [ + 140, + 133, + 450, + 289 + ], + "blocks": [ + { + "bbox": [ + 140, + 119, + 425, + 131 + ], + "lines": [ + { + "bbox": [ + 140, + 119, + 425, + 131 + ], + "spans": [ + { + "bbox": [ + 140, + 119, + 425, + 131 + ], + "type": "text", + "content": "Algorithm 1: Multi-view object association algorithm (MVOA)" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "code_caption" + }, + { + "bbox": [ + 140, + 133, + 450, + 289 + ], + "lines": [ + { + "bbox": [ + 140, + 133, + 450, + 289 + ], + "spans": [ + { + "bbox": [ + 140, + 133, + 450, + 289 + ], + "type": "text", + "content": "Data: Inconsistent instance mask " + }, + { + "bbox": [ + 140, + 133, + 450, + 289 + ], + "type": "inline_equation", + "content": "\\{\\mathcal{K}\\}" + }, + { + "bbox": [ + 140, + 133, + 450, + 289 + ], + "type": "text", + "content": ", input feature set " + }, + { + "bbox": [ + 140, + 133, + 450, + 289 + ], + "type": "inline_equation", + "content": "\\{\\mathcal{F}\\}" + }, + { + "bbox": [ + 140, + 133, + 450, + 289 + ], + "type": "text", + "content": ", Threshold T \nResult: Prototype features set " + }, + { + "bbox": [ + 140, + 133, + 450, + 289 + ], + "type": "inline_equation", + "content": "\\mathcal{D}" + }, + { + "bbox": [ + 140, + 133, + 450, + 289 + ], + "type": "text", + "content": " \n// Instance Grouping // Multi-view matching \n2 " + }, + { + "bbox": [ + 140, + 133, + 450, + 289 + ], + "type": "inline_equation", + "content": "\\mathcal{C} = \\{\\}" + }, + { + "bbox": [ + 140, + 133, + 450, + 289 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 140, + 133, + 450, + 289 + ], + "type": "inline_equation", + "content": "\\mathcal{S} = \\{\\}" + }, + { + "bbox": [ + 140, + 133, + 450, + 289 + ], + "type": "text", + "content": " 10 " + }, + { + "bbox": [ + 140, + 133, + 450, + 289 + ], + "type": "inline_equation", + "content": "\\mathcal{D} \\gets \\{\\}" + }, + { + "bbox": [ + 140, + 133, + 450, + 289 + ], + "type": "text", + "content": " \n3 for " + }, + { + "bbox": [ + 140, + 133, + 450, + 289 + ], + "type": "inline_equation", + "content": "\\mathcal{K}_l" + }, + { + "bbox": [ + 140, + 133, + 450, + 289 + ], + "type": "text", + "content": " in " + }, + { + "bbox": [ + 140, + 133, + 450, + 289 + ], + "type": "inline_equation", + "content": "\\{\\mathcal{K}\\}" + }, + { + "bbox": [ + 140, + 133, + 450, + 289 + ], + "type": "text", + "content": " do 11 while " + }, + { + "bbox": [ + 140, + 133, + 450, + 289 + ], + "type": "inline_equation", + "content": "\\mathcal{S} \\neq empty" + }, + { + "bbox": [ + 140, + 133, + 450, + 289 + ], + "type": "text", + "content": " do \n4 IDs = {unique_ID(\\mathcal{K}_l)} 12 m ← argmax S \n5 for " + }, + { + "bbox": [ + 140, + 133, + 450, + 289 + ], + "type": "inline_equation", + "content": "p" + }, + { + "bbox": [ + 140, + 133, + 450, + 289 + ], + "type": "text", + "content": " in " + }, + { + "bbox": [ + 140, + 133, + 450, + 289 + ], + "type": "inline_equation", + "content": "IDs" + }, + { + "bbox": [ + 140, + 133, + 450, + 289 + ], + "type": "text", + "content": " do D ← D ∪ " + }, + { + "bbox": [ + 140, + 133, + 450, + 289 + ], + "type": "inline_equation", + "content": "\\mathcal{C}_m" + }, + { + "bbox": [ + 140, + 133, + 450, + 289 + ], + "type": "text", + "content": " \n6 " + }, + { + "bbox": [ + 140, + 133, + 450, + 289 + ], + "type": "inline_equation", + "content": "\\mathcal{C}_l^p = \\sum_i\\mathcal{F}_l^i /|\\{i:\\mathcal{K}_l^i = p\\}|" + }, + { + "bbox": [ + 140, + 133, + 450, + 289 + ], + "type": "text", + "content": " C ← C - Cm; S ← S - Sm \n7 " + }, + { + "bbox": [ + 140, + 133, + 450, + 289 + ], + "type": "inline_equation", + "content": "\\mathcal{S}_l^p = \\Phi (\\mathcal{C}_l^p)" + }, + { + "bbox": [ + 140, + 133, + 450, + 289 + ], + "type": "text", + "content": " 13 if Kρ(Cm, Ci) ≥ T then 14 | C ← C - Ci; S ← S - Si \n8 end 16 end \n9 end 17 end" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "code_body" + } + ], + "index": 3, + "sub_type": "algorithm" + }, + { + "type": "image", + "bbox": [ + 151, + 304, + 460, + 389 + ], + "blocks": [ + { + "bbox": [ + 151, + 304, + 460, + 389 + ], + "lines": [ + { + "bbox": [ + 151, + 304, + 460, + 389 + ], + "spans": [ + { + "bbox": [ + 151, + 304, + 460, + 389 + ], + "type": "image", + "image_path": "5eb27910e9a7bbe8a14f61f75c1d7a5cd71931a08d3f1e524505cba76cef2509.jpg" + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 130, + 407, + 482, + 475 + ], + "lines": [ + { + "bbox": [ + 130, + 407, + 482, + 475 + ], + "spans": [ + { + "bbox": [ + 130, + 407, + 482, + 475 + ], + "type": "text", + "content": "Fig. 3: (a) Flexibility of adjusting covariances. Contour plot of the PP kernel similarity for two Gaussians with different covariance values " + }, + { + "bbox": [ + 130, + 407, + 482, + 475 + ], + "type": "inline_equation", + "content": "(\\sigma_{1}" + }, + { + "bbox": [ + 130, + 407, + 482, + 475 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 130, + 407, + 482, + 475 + ], + "type": "inline_equation", + "content": "\\sigma_{2})" + }, + { + "bbox": [ + 130, + 407, + 482, + 475 + ], + "type": "text", + "content": " and fixed mean values in a 1-dimensional case. (b) Anisotropy. Contour plot of the PP kernel similarity for two Gaussians with different Gaussian mean offsets " + }, + { + "bbox": [ + 130, + 407, + 482, + 475 + ], + "type": "inline_equation", + "content": "(d_{x}" + }, + { + "bbox": [ + 130, + 407, + 482, + 475 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 130, + 407, + 482, + 475 + ], + "type": "inline_equation", + "content": "d_{y})" + }, + { + "bbox": [ + 130, + 407, + 482, + 475 + ], + "type": "text", + "content": " and fixed covariances in a 2-dimensional case. (c) Isotropy. Contour plot of the RBF kernel similarity for two deterministic features with different offsets " + }, + { + "bbox": [ + 130, + 407, + 482, + 475 + ], + "type": "inline_equation", + "content": "(d_{x}" + }, + { + "bbox": [ + 130, + 407, + 482, + 475 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 130, + 407, + 482, + 475 + ], + "type": "inline_equation", + "content": "d_{y})" + }, + { + "bbox": [ + 130, + 407, + 482, + 475 + ], + "type": "text", + "content": " in a 2-dimensional case." + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "image_caption" + } + ], + "index": 4 + }, + { + "bbox": [ + 130, + 497, + 482, + 534 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 497, + 482, + 534 + ], + "spans": [ + { + "bbox": [ + 130, + 497, + 482, + 534 + ], + "type": "text", + "content": "more general framework, while the deterministic method is a subclass within the broader probabilistic paradigm. The experimental results also verify the effectiveness of our probabilistic solution in the ablation studies of Sec. 4.3." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 131, + 550, + 224, + 564 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 131, + 550, + 224, + 564 + ], + "spans": [ + { + "bbox": [ + 131, + 550, + 224, + 564 + ], + "type": "text", + "content": "4 Experiment" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 131, + 574, + 274, + 588 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 131, + 574, + 274, + 588 + ], + "spans": [ + { + "bbox": [ + 131, + 574, + 274, + 588 + ], + "type": "text", + "content": "4.1 Experimental Settings" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 130, + 594, + 482, + 666 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 594, + 482, + 666 + ], + "spans": [ + { + "bbox": [ + 130, + 594, + 482, + 666 + ], + "type": "text", + "content": "Implementation Details. For fair comparison, we adopt the same architecture of TensoRF [6] together with the same layer parameters used in previous works [3,40]. Particularly, our instance field is constructed following the slow-fast architecture of Contrastive Lift [3], where a shallow 5-layer MLP is applied to predict an probabilistic feature embedding for a 3D coordinate input. In practice, the dimension " + }, + { + "bbox": [ + 130, + 594, + 482, + 666 + ], + "type": "inline_equation", + "content": "N" + }, + { + "bbox": [ + 130, + 594, + 482, + 666 + ], + "type": "text", + "content": " for the probabilistic feature embedding is 3 in our experiments." + } + ] + } + ], + "index": 9 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 192, + 91, + 447, + 102 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 192, + 91, + 447, + 102 + ], + "spans": [ + { + "bbox": [ + 192, + 91, + 447, + 102 + ], + "type": "text", + "content": "PCF-Lift: Panoptic Lifting by Probabilistic Contrastive Fusion" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 474, + 92, + 480, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 474, + 92, + 480, + 100 + ], + "spans": [ + { + "bbox": [ + 474, + 92, + 480, + 100 + ], + "type": "text", + "content": "9" + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 8 + }, + { + "para_blocks": [ + { + "bbox": [ + 130, + 116, + 482, + 213 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 116, + 482, + 213 + ], + "spans": [ + { + "bbox": [ + 130, + 116, + 482, + 213 + ], + "type": "text", + "content": "For the training of the instance field, we sample rays from two different views in the last few epochs for optimization (Sec. 3.2), using the cross-view constraint " + }, + { + "bbox": [ + 130, + 116, + 482, + 213 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_{\\mathrm{cross}}" + }, + { + "bbox": [ + 130, + 116, + 482, + 213 + ], + "type": "text", + "content": " with a predefined threshold value " + }, + { + "bbox": [ + 130, + 116, + 482, + 213 + ], + "type": "inline_equation", + "content": "\\tau = 0.9" + }, + { + "bbox": [ + 130, + 116, + 482, + 213 + ], + "type": "text", + "content": " and weight " + }, + { + "bbox": [ + 130, + 116, + 482, + 213 + ], + "type": "inline_equation", + "content": "w_{\\mathrm{cross}} = 0.05" + }, + { + "bbox": [ + 130, + 116, + 482, + 213 + ], + "type": "text", + "content": "; while in other training epochs, we set the weight " + }, + { + "bbox": [ + 130, + 116, + 482, + 213 + ], + "type": "inline_equation", + "content": "w_{\\mathrm{cross}}" + }, + { + "bbox": [ + 130, + 116, + 482, + 213 + ], + "type": "text", + "content": " to 0 and only sample the rays from single views. Besides, " + }, + { + "bbox": [ + 130, + 116, + 482, + 213 + ], + "type": "inline_equation", + "content": "w_{\\mathrm{reg}}" + }, + { + "bbox": [ + 130, + 116, + 482, + 213 + ], + "type": "text", + "content": " in Eq. (6) is set to 0.001 throughout the whole training process. For the training of the color, density and semantic fields, we adopt the same loss terms and training strategies from previous works [3, 40]. More implementation details are provided in the supplementary material." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 130, + 228, + 483, + 350 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 228, + 483, + 350 + ], + "spans": [ + { + "bbox": [ + 130, + 228, + 483, + 350 + ], + "type": "text", + "content": "Metrics. Since our method is proposed to solve the panoptic-lifting task, we employ scene-level Panoptic Quality (PQ" + }, + { + "bbox": [ + 130, + 228, + 483, + 350 + ], + "type": "inline_equation", + "content": "^{\\text{scene}}" + }, + { + "bbox": [ + 130, + 228, + 483, + 350 + ], + "type": "text", + "content": ") for evaluation, which was introduced in Panoptic Lifting [40] and widely used in the related works [3,40]. Unlike the standard PQ metric [23], this metric particularly considers the consistency of instance IDs across multiple views. By merging predictions and ground truths with consistent instance IDs into subsets for PQ" + }, + { + "bbox": [ + 130, + 228, + 483, + 350 + ], + "type": "inline_equation", + "content": "^{\\text{scene}}" + }, + { + "bbox": [ + 130, + 228, + 483, + 350 + ], + "type": "text", + "content": "-based evaluation, we determine a match of subset pair when the intersection over union (IoU) exceeds 0.5, in line with prior baselines [3,40]. Since PQ" + }, + { + "bbox": [ + 130, + 228, + 483, + 350 + ], + "type": "inline_equation", + "content": "^{\\text{scene}}" + }, + { + "bbox": [ + 130, + 228, + 483, + 350 + ], + "type": "text", + "content": " is a product of scene-level segmentation quality (SQ" + }, + { + "bbox": [ + 130, + 228, + 483, + 350 + ], + "type": "inline_equation", + "content": "^{\\text{scene}}" + }, + { + "bbox": [ + 130, + 228, + 483, + 350 + ], + "type": "text", + "content": ") and recognition quality (RQ" + }, + { + "bbox": [ + 130, + 228, + 483, + 350 + ], + "type": "inline_equation", + "content": "^{\\text{scene}}" + }, + { + "bbox": [ + 130, + 228, + 483, + 350 + ], + "type": "text", + "content": "), we also provide the SQ" + }, + { + "bbox": [ + 130, + 228, + 483, + 350 + ], + "type": "inline_equation", + "content": "^{\\text{scene}}" + }, + { + "bbox": [ + 130, + 228, + 483, + 350 + ], + "type": "text", + "content": " and RQ" + }, + { + "bbox": [ + 130, + 228, + 483, + 350 + ], + "type": "inline_equation", + "content": "^{\\text{scene}}" + }, + { + "bbox": [ + 130, + 228, + 483, + 350 + ], + "type": "text", + "content": " metrics for more detailed comparisons." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 130, + 363, + 482, + 425 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 363, + 482, + 425 + ], + "spans": [ + { + "bbox": [ + 130, + 363, + 482, + 425 + ], + "type": "text", + "content": "Baselines. We mainly compare the proposed method with the current state-of-the-art methods that target lifting 2D panoptic predictions to 3D, i.e., Contrastive Lift [3] and Panoptic Lifting [40]. Moreover, we compare our method with the recent NeRF-based 3D panoptic segmentation approaches, i.e., Panoptic Neural Fields [26] (PNF) and DM-NeRF [44]." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 130, + 440, + 482, + 537 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 440, + 482, + 537 + ], + "spans": [ + { + "bbox": [ + 130, + 440, + 482, + 537 + ], + "type": "text", + "content": "Datasets. Following the previous works [3,40], we conduct experiments on two public datasets, the ScanNet dataset [12] and the Messy Room dataset [3], for both quantitative and qualitative evaluations. For fair comparisons on ScanNet, following the other state-of-the-art approaches [3,40], we adopt Mask2Former (M2F) [9] to generate 2D panoptic segmentation predictions, coupled with the same protocol in [40] that maps the COCO [28] vocabulary to 21 classes. For the experiments on Messy Rooms, we use the LVIS [16] vocabulary and the Detic [50] 2D panoptic segmentations, which are also utilized in Contrastive Lift [3]." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 131, + 552, + 255, + 563 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 131, + 552, + 255, + 563 + ], + "spans": [ + { + "bbox": [ + 131, + 552, + 255, + 563 + ], + "type": "text", + "content": "4.2 Main Experiments" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 130, + 570, + 482, + 666 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 570, + 482, + 666 + ], + "spans": [ + { + "bbox": [ + 130, + 570, + 482, + 666 + ], + "type": "text", + "content": "ScanNet Dataset. To verify the performance on real data, we conduct experiments on the ScanNet dataset [12] of 12 scenes. Quantitative comparisons provided in Tab. 1 demonstrate that our method consistently outperforms previous baselines, including both the 3D panoptic segmentation approaches [26,44] and the recent state-of-the-art methods for lifting 2D panoptic segmentation [3,40]. Moreover, visual comparisons between our method and the latest state-of-the-art method [3] are presented in Fig. 4, which exhibits our method's capability of achieving a consistent and accurate 3D panoptic segmentation." + } + ] + } + ], + "index": 7 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "spans": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "type": "text", + "content": "10" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 166, + 91, + 285, + 102 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 166, + 91, + 285, + 102 + ], + "spans": [ + { + "bbox": [ + 166, + 91, + 285, + 102 + ], + "type": "text", + "content": "R. Zhu, S. Qiu, Q. Wu, et al." + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 9 + }, + { + "para_blocks": [ + { + "type": "table", + "bbox": [ + 150, + 179, + 461, + 244 + ], + "blocks": [ + { + "bbox": [ + 130, + 114, + 482, + 170 + ], + "lines": [ + { + "bbox": [ + 130, + 114, + 482, + 170 + ], + "spans": [ + { + "bbox": [ + 130, + 114, + 482, + 170 + ], + "type": "text", + "content": "Table 1: Results on the ScanNet dataset. We report the PQscene, SQscene, and RQscene metrics. Since Contrastive Lift [3] does not report the performance using the SQscene and RQscene metrics, we apply the officially-released pre-trained model and re-run the clustering algorithm to obtain the values reported in this table. For the other metric values, we directly report the ones in previous papers [3,40]." + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 150, + 179, + 461, + 244 + ], + "lines": [ + { + "bbox": [ + 150, + 179, + 461, + 244 + ], + "spans": [ + { + "bbox": [ + 150, + 179, + 461, + 244 + ], + "type": "table", + "html": "
MethodVenueType\\(SQ^{scene}(\\%)\\)\\(RQ^{scene}(\\%)\\)\\(PQ^{scene}(\\%)\\)
DM-NeRF [44]ICLR'233D panoptic segmentation53.346.141.7
PNF [26]CVPR'223D panoptic segmentation63.050.748.3
PNF [26] + GT BBoxesCVPR'223D panoptic segmentation70.055.954.3
Panoptic Lifting [40]CVPR'232D panoptic Lifting73.565.058.9
Contrastive Lift [3]NeurIPS'232D panoptic Lifting75.763.662.0
Ours-2D panoptic Lifting78.565.463.5
", + "image_path": "42be9442a4b28aa70ba6117a17b80ad9e93f70a6a4b52e77398911d8fc8a40bd.jpg" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "table_body" + } + ], + "index": 3 + }, + { + "type": "table", + "bbox": [ + 150, + 297, + 461, + 352 + ], + "blocks": [ + { + "bbox": [ + 130, + 254, + 482, + 288 + ], + "lines": [ + { + "bbox": [ + 130, + 254, + 482, + 288 + ], + "spans": [ + { + "bbox": [ + 130, + 254, + 482, + 288 + ], + "type": "text", + "content": "Table 2: Results on the Messy Rooms dataset [3]. Following [3], the PQscene metric is reported on both \"old room\" and \"large corridor\" environments with an increasing number of objects in the scene (25, 50, 100, 500)." + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 150, + 297, + 461, + 352 + ], + "lines": [ + { + "bbox": [ + 150, + 297, + 461, + 352 + ], + "spans": [ + { + "bbox": [ + 150, + 297, + 461, + 352 + ], + "type": "table", + "html": "
Method/ NumberOld Room Environment (%)Large Corridor Environment(%)Mean(%)
25501005002550100500
Panoptic Lifting [40]73.269.964.351.065.571.061.849.063.2
Contrastive Lift [3]78.975.869.155.076.575.568.752.569.0
Ours80.978.374.860.381.079.474.058.873.4
", + "image_path": "bdb5fa822af6124b8e21b8085593e21e193ed2665a169eb83e8c414113b41e04.jpg" + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "table_body" + } + ], + "index": 5 + }, + { + "bbox": [ + 130, + 373, + 482, + 529 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 373, + 482, + 529 + ], + "spans": [ + { + "bbox": [ + 130, + 373, + 482, + 529 + ], + "type": "text", + "content": "Messy Rooms Dataset. We also conduct experiments on the challenging Messy Room dataset [3], which is provided by Contrastive Lift [3], containing up to 500 objects in each scene. We present the quantitative results in Tab. 2, showing that our method (73.4%) achieves significant improvements in terms of the mean PQ" + }, + { + "bbox": [ + 130, + 373, + 482, + 529 + ], + "type": "inline_equation", + "content": "^{\\text{scene}}" + }, + { + "bbox": [ + 130, + 373, + 482, + 529 + ], + "type": "text", + "content": " results compared to the current state-of-the-art methods of Contrastive Lift [3] (69.0%) and Panoptic Lifting [40] (63.2%). The overall performance highlights the advantages of our probabilistic approach over the previous deterministic methods, particularly in segmenting complex scenes with hundreds of objects. Further, our method achieves a mean SQ" + }, + { + "bbox": [ + 130, + 373, + 482, + 529 + ], + "type": "inline_equation", + "content": "^{\\text{scene}}" + }, + { + "bbox": [ + 130, + 373, + 482, + 529 + ], + "type": "text", + "content": " of 82.2% and a mean RQ" + }, + { + "bbox": [ + 130, + 373, + 482, + 529 + ], + "type": "inline_equation", + "content": "^{\\text{scene}}" + }, + { + "bbox": [ + 130, + 373, + 482, + 529 + ], + "type": "text", + "content": " of 86.9%, surpassing the SQ" + }, + { + "bbox": [ + 130, + 373, + 482, + 529 + ], + "type": "inline_equation", + "content": "^{\\text{scene}}" + }, + { + "bbox": [ + 130, + 373, + 482, + 529 + ], + "type": "text", + "content": " of 77.7% and RQ" + }, + { + "bbox": [ + 130, + 373, + 482, + 529 + ], + "type": "inline_equation", + "content": "^{\\text{scene}}" + }, + { + "bbox": [ + 130, + 373, + 482, + 529 + ], + "type": "text", + "content": " of 86.6% obtained by Contrastive Lift [3]. Moreover, the visual comparisons in Fig. 4 further indicate that our method can accurately segment the small objects in the red frames, whereas Contrastive Lift [3] struggles to distinguish such instances." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 131, + 548, + 238, + 561 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 131, + 548, + 238, + 561 + ], + "spans": [ + { + "bbox": [ + 131, + 548, + 238, + 561 + ], + "type": "text", + "content": "4.3 Ablation Study" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 130, + 569, + 482, + 667 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 569, + 482, + 667 + ], + "spans": [ + { + "bbox": [ + 130, + 569, + 482, + 667 + ], + "type": "text", + "content": "Effectiveness of Each Component. We conduct an ablation study on the Messy Room dataset [3], which contains more challenging scenes with hundreds of instances. As shown in Tab. 3 (b) and (d), the proposed probabilistic feature embeddings greatly benefit our method, while replacing it with deterministic one leads to significant performance drop. Furthermore, the effectiveness of our proposed cross-view constraint is verified in Tab. 3 (e) and (f). For the proposed multi-view object association (MVOA) algorithm, the results in Tab. 3 (c), (d), (e), and (f) show that it particularly benefits the clustering of our proposed" + } + ] + } + ], + "index": 8 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 192, + 90, + 447, + 102 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 192, + 90, + 447, + 102 + ], + "spans": [ + { + "bbox": [ + 192, + 90, + 447, + 102 + ], + "type": "text", + "content": "PCF-Lift: Panoptic Lifting by Probabilistic Contrastive Fusion" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 470, + 91, + 479, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 470, + 91, + 479, + 100 + ], + "spans": [ + { + "bbox": [ + 470, + 91, + 479, + 100 + ], + "type": "text", + "content": "11" + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 10 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 187, + 114, + 430, + 441 + ], + "blocks": [ + { + "bbox": [ + 187, + 114, + 430, + 441 + ], + "lines": [ + { + "bbox": [ + 187, + 114, + 430, + 441 + ], + "spans": [ + { + "bbox": [ + 187, + 114, + 430, + 441 + ], + "type": "image", + "image_path": "39c375f5c14bb29d59de648735d61e2302061b04ed2f694827fcd1b78f3d147b.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 130, + 461, + 482, + 485 + ], + "lines": [ + { + "bbox": [ + 130, + 461, + 482, + 485 + ], + "spans": [ + { + "bbox": [ + 130, + 461, + 482, + 485 + ], + "type": "text", + "content": "Fig. 4: Visual comparison of the latest state-of-the-art method Contrastive Lift [3] and our method on the ScanNet [12] dataset and the Messy Room [3] dataset." + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_caption" + } + ], + "index": 2 + }, + { + "type": "table", + "bbox": [ + 167, + 546, + 443, + 600 + ], + "blocks": [ + { + "bbox": [ + 130, + 514, + 482, + 537 + ], + "lines": [ + { + "bbox": [ + 130, + 514, + 482, + 537 + ], + "spans": [ + { + "bbox": [ + 130, + 514, + 482, + 537 + ], + "type": "text", + "content": "Table 3: Ablation study on the Messy Room dataset [3]. The model (a) corresponds to Contrastive Lift [3] and the model (f) corresponds to our full method (PCF)." + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 167, + 546, + 443, + 600 + ], + "lines": [ + { + "bbox": [ + 167, + 546, + 443, + 600 + ], + "spans": [ + { + "bbox": [ + 167, + 546, + 443, + 600 + ], + "type": "table", + "html": "
Model Feature spaceClustering\\( {\\mathrm{{SQ}}}^{\\text{scene }}\\left( \\% \\right) \\)\\( {\\mathrm{{RQ}}}^{\\text{scene }}\\left( \\% \\right) \\)\\( {\\mathrm{{PQ}}}^{\\text{scene }}\\left( \\% \\right) \\)
(a)Deterministic [3]HDBSCAN [30]77.786.669.0
(b)Deterministic [3]MVOA79.386.270.4
(c)Learned Gaussian distributionHDBSCAN [30]78.086.669.6
(d)Learned Gaussian distributionMVOA81.386.872.3
(e)Learned Gaussian distribution (+ Cross-view constraint)HDBSCAN [30]78.886.970.4
(f)Learned Gaussian distribution (+ Cross-view constraint)MVOA82.286.973.4
", + "image_path": "802b967888807f6914d8e20d9960a392ccd2d843989901e8353ca5d24548ee43.jpg" + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "table_body" + } + ], + "index": 5 + }, + { + "bbox": [ + 130, + 641, + 481, + 666 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 641, + 481, + 666 + ], + "spans": [ + { + "bbox": [ + 130, + 641, + 481, + 666 + ], + "type": "text", + "content": "probabilistic representation. Also, MVOA can be effectively employed to the deterministic method [3] for a performance boost, as Tab. 3 (a) and (b) show." + } + ] + } + ], + "index": 6 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "spans": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "type": "text", + "content": "12" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 166, + 91, + 285, + 102 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 166, + 91, + 285, + 102 + ], + "spans": [ + { + "bbox": [ + 166, + 91, + 285, + 102 + ], + "type": "text", + "content": "R. Zhu, S. Qiu, Q. Wu, et al." + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 11 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 173, + 112, + 440, + 198 + ], + "blocks": [ + { + "bbox": [ + 173, + 112, + 440, + 198 + ], + "lines": [ + { + "bbox": [ + 173, + 112, + 440, + 198 + ], + "spans": [ + { + "bbox": [ + 173, + 112, + 440, + 198 + ], + "type": "image", + "image_path": "3b8cf2d43fca2acbdd6d2c4b542b71d6913358443b2654678a90e96ce0c07721.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 130, + 214, + 482, + 282 + ], + "lines": [ + { + "bbox": [ + 130, + 214, + 482, + 282 + ], + "spans": [ + { + "bbox": [ + 130, + 214, + 482, + 282 + ], + "type": "text", + "content": "Fig. 5: The visualization results of learned covariance components and the statistical results of covariances in two scenes of the Messy Room dataset [3]. For the histograms, the horizontal axis denotes the range of covariance magnitudes, while the vertical axis corresponds to the frequency statistics for those magnitudes. We calculate and plot the covariance magnitudes " + }, + { + "bbox": [ + 130, + 214, + 482, + 282 + ], + "type": "inline_equation", + "content": "(\\sigma^{(1)^2}*\\sigma^{(2)^2}*\\sigma^{(3)^2})" + }, + { + "bbox": [ + 130, + 214, + 482, + 282 + ], + "type": "text", + "content": " of two distance image regions, the boundary areas and the internal areas of object instances, across all observed views." + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_caption" + } + ], + "index": 2 + }, + { + "bbox": [ + 130, + 304, + 482, + 437 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 304, + 482, + 437 + ], + "spans": [ + { + "bbox": [ + 130, + 304, + 482, + 437 + ], + "type": "text", + "content": "Uncertainty Analysis. To verify whether our method could provide meaningful modeling for uncertainty, we provide the rendered uncertainty maps on two scenes of the Messy Room dataset [3] in Fig. 5. The figure clearly illustrates that the regions of high covariance are mainly located near the boundaries of the object instances, due to the inconsistent segmentation issue that leaves severe ambiguity around the instance boundaries. Moreover, we statistically analyze the learned covariances in two distinct image regions, namely the boundary areas and the internal areas of object instances, across all observed views. These results further demonstrate our method's ability to model high uncertainty in areas where the instance boundaries are ambiguous. We provide more details in the supplementary material." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 132, + 452, + 285, + 464 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 452, + 285, + 464 + ], + "spans": [ + { + "bbox": [ + 132, + 452, + 285, + 464 + ], + "type": "text", + "content": "4.4 Robustness Experiments" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 130, + 470, + 482, + 578 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 470, + 482, + 578 + ], + "spans": [ + { + "bbox": [ + 130, + 470, + 482, + 578 + ], + "type": "text", + "content": "2D Backbones. In practice, the quality of panoptic segmentation generated by different 2D models may vary a lot. To study the robustness of our probabilistic method and the deterministic method [3] when incorporating with different 2D models, we select four different models from the model zoo of Detic [3] and utilize the LVIS [16] vocabulary to generate the 2D panoptic segmentation predictions for lifting, as shown in Fig. 6 (a). The quantitative results are presented in Fig. 6 (b), where our probabilistic methods consistently outperform the baseline method by a large margin. Particularly, we observe that the proposed MVOA algorithm can consistently boost the performance of a deterministic method [3]." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 130, + 594, + 482, + 666 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 594, + 482, + 666 + ], + "spans": [ + { + "bbox": [ + 130, + 594, + 482, + 666 + ], + "type": "text", + "content": "Hand-crafted Noise. We study the robustness by adding hand-crafted noise. Specifically, for all given panoptic segmentation masks, we randomly select 100 pixels as anchors. Then, for each anchor, we randomly choose a pixel within a " + }, + { + "bbox": [ + 130, + 594, + 482, + 666 + ], + "type": "inline_equation", + "content": "W * W" + }, + { + "bbox": [ + 130, + 594, + 482, + 666 + ], + "type": "text", + "content": " window centered on the anchor and assign its instance ID to all pixels within this window, to simulate inaccurate segmentation predictions around object boundaries. As shown in Fig. 6 (c), as W gradually increases, the instance" + } + ] + } + ], + "index": 7 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 192, + 91, + 447, + 102 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 192, + 91, + 447, + 102 + ], + "spans": [ + { + "bbox": [ + 192, + 91, + 447, + 102 + ], + "type": "text", + "content": "PCF-Lift: Panoptic Lifting by Probabilistic Contrastive Fusion" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 470, + 91, + 480, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 470, + 91, + 480, + 100 + ], + "spans": [ + { + "bbox": [ + 470, + 91, + 480, + 100 + ], + "type": "text", + "content": "13" + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 12 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 190, + 113, + 424, + 297 + ], + "blocks": [ + { + "bbox": [ + 190, + 113, + 424, + 297 + ], + "lines": [ + { + "bbox": [ + 190, + 113, + 424, + 297 + ], + "spans": [ + { + "bbox": [ + 190, + 113, + 424, + 297 + ], + "type": "image", + "image_path": "36d2f150e399caa373b87d78d087fc90d1ac5b9babcbd02df12f42b5e2bf4583.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 131, + 316, + 482, + 383 + ], + "lines": [ + { + "bbox": [ + 131, + 316, + 482, + 383 + ], + "spans": [ + { + "bbox": [ + 131, + 316, + 482, + 383 + ], + "type": "text", + "content": "Fig. 6: Quantitative comparisons when using different 2D models or adding various levels of hand-crafted noise. We compare our probabilistic method (\"Probabilistic + MVOA\" in red) with the deterministic method [3] (\"Deterministic\" in dark blue) and a variant of deterministic method using our proposed MVOA algorithm (\"Deterministic + MVOA\" in light blue). Besides, the four tested models are from the model zoo of Detic " + }, + { + "bbox": [ + 131, + 316, + 482, + 383 + ], + "type": "inline_equation", + "content": "[50]^1" + }, + { + "bbox": [ + 131, + 316, + 482, + 383 + ], + "type": "text", + "content": ", and the two tested scenes are from the Messy Room dataset [3]." + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_caption" + } + ], + "index": 2 + }, + { + "bbox": [ + 130, + 410, + 480, + 447 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 410, + 480, + 447 + ], + "spans": [ + { + "bbox": [ + 130, + 410, + 480, + 447 + ], + "type": "text", + "content": "boundaries are continuously blurred. Our probabilistic approach consistently achieves the best performance as presented in Fig. 6 (d) and the results demonstrate that the proposed MVOA algorithm benefits the deterministic method [3]." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 132, + 486, + 220, + 498 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 486, + 220, + 498 + ], + "spans": [ + { + "bbox": [ + 132, + 486, + 220, + 498 + ], + "type": "text", + "content": "5 Conclusion" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 130, + 521, + 482, + 617 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 521, + 482, + 617 + ], + "spans": [ + { + "bbox": [ + 130, + 521, + 482, + 617 + ], + "type": "text", + "content": "We present PCF-Lift for the panoptic lifting task. First, we propose to learn the probabilistic feature embeddings through a multivariate Gaussian distribution for instance representation. For training, we reformulate the contrastive loss with the Probability Product kernel and propose a novel cross-view constraint to enhance the feature consistency across different views. During the inference phase, we propose a novel multi-view object association algorithm to effectively identify the prototype features representing underlying 3D object instances. We verify the effectiveness and robustness of PCF-Lift by extensive experiments." + } + ] + } + ], + "index": 6 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "spans": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "type": "text", + "content": "14" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 166, + 91, + 285, + 102 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 166, + 91, + 285, + 102 + ], + "spans": [ + { + "bbox": [ + 166, + 91, + 285, + 102 + ], + "type": "text", + "content": "R. Zhu, S. Qiu, Q. Wu, et al." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 133, + 631, + 482, + 665 + ], + "type": "page_footnote", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 631, + 482, + 665 + ], + "spans": [ + { + "bbox": [ + 133, + 631, + 482, + 665 + ], + "type": "text", + "content": "1 We use the official pre-trained models provided in https://github.com/facebookresearch/Detic/tree/main. Please refer to our supplementary material for the details of the four models." + } + ] + } + ], + "index": 7 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 13 + }, + { + "para_blocks": [ + { + "bbox": [ + 133, + 114, + 246, + 129 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 114, + 246, + 129 + ], + "spans": [ + { + "bbox": [ + 133, + 114, + 246, + 129 + ], + "type": "text", + "content": "Acknowledgements" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 130, + 140, + 482, + 187 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 140, + 482, + 187 + ], + "spans": [ + { + "bbox": [ + 130, + 140, + 482, + 187 + ], + "type": "text", + "content": "This work is supported by the InnoHK of the Government of Hong Kong via the Hong Kong Centre for Logistics Robotics, the Shenzhen Portion of Shenzhen-Hong Kong Science and Technology Innovation Cooperation Zone under HZQB-KCZYB-20200089, and the CUHK T Stone Robotics Institute." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 133, + 205, + 197, + 217 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 205, + 197, + 217 + ], + "spans": [ + { + "bbox": [ + 133, + 205, + 197, + 217 + ], + "type": "text", + "content": "References" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 134, + 228, + 480, + 665 + ], + "type": "list", + "angle": 0, + "index": 18, + "blocks": [ + { + "bbox": [ + 138, + 228, + 480, + 262 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 228, + 480, + 262 + ], + "spans": [ + { + "bbox": [ + 138, + 228, + 480, + 262 + ], + "type": "text", + "content": "1. Bae, G., Budvytis, I., Cipolla, R.: Estimating and exploiting the aleatoric uncertainty in surface normal estimation. In: Proceedings of the IEEE/CVF International Conference on Computer Vision. pp. 13137-13146 (2021)" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 138, + 262, + 480, + 305 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 262, + 480, + 305 + ], + "spans": [ + { + "bbox": [ + 138, + 262, + 480, + 305 + ], + "type": "text", + "content": "2. Barron, J.T., Mildenhall, B., Tancik, M., Hedman, P., Martin-Brualla, R., Srinivasan, P.P.: Mip-NeRF: A multiscale representation for anti-aliasing neural radiance fields. In: Proceedings of the IEEE/CVF International Conference on Computer Vision. pp. 5855-5864 (2021)" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 138, + 306, + 480, + 338 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 306, + 480, + 338 + ], + "spans": [ + { + "bbox": [ + 138, + 306, + 480, + 338 + ], + "type": "text", + "content": "3. Bhalgat, Y., Laina, I., Henriques, J.F., Zisserman, A., Vedaldi, A.: Contrastive Lift: 3D object instance segmentation by slow-fast contrastive fusion. arXiv preprint arXiv:2306.04633 (2023)" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 138, + 339, + 480, + 370 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 339, + 480, + 370 + ], + "spans": [ + { + "bbox": [ + 138, + 339, + 480, + 370 + ], + "type": "text", + "content": "4. Cai, K., Lu, C.X., Huang, X.: Uncertainty estimation for 3D dense prediction via cross-point embeddings. IEEE Robotics and Automation Letters 8(5), 2558-2565 (2023)" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 138, + 371, + 480, + 403 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 371, + 480, + 403 + ], + "spans": [ + { + "bbox": [ + 138, + 371, + 480, + 403 + ], + "type": "text", + "content": "5. Cen, J., Zhou, Z., Fang, J., Shen, W., Xie, L., Jiang, D., Zhang, X., Tian, Q., et al.: Segment anything in 3D with NeRFs. Advances in Neural Information Processing Systems 36 (2024)" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 138, + 404, + 480, + 426 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 404, + 480, + 426 + ], + "spans": [ + { + "bbox": [ + 138, + 404, + 480, + 426 + ], + "type": "text", + "content": "6. Chen, A., Xu, Z., Geiger, A., Yu, J., Su, H.: Tensorf: Tensorial radiance fields. In: European Conference on Computer Vision. pp. 333-350. Springer (2022)" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 138, + 426, + 480, + 458 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 426, + 480, + 458 + ], + "spans": [ + { + "bbox": [ + 138, + 426, + 480, + 458 + ], + "type": "text", + "content": "7. Chen, T., Kornblith, S., Norouzi, M., Hinton, G.: A simple framework for contrastive learning of visual representations. In: International conference on machine learning. pp. 1597-1607. PMLR (2020)" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 138, + 459, + 480, + 502 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 459, + 480, + 502 + ], + "spans": [ + { + "bbox": [ + 138, + 459, + 480, + 502 + ], + "type": "text", + "content": "8. Cheng, B., Collins, M.D., Zhu, Y., Liu, T., Huang, T.S., Adam, H., Chen, L.C.: Panoptic-deeplab: A simple, strong, and fast baseline for bottom-up panoptic segmentation. In: Proceedings of the IEEE/CVF conference on computer vision and pattern recognition. pp. 12475-12485 (2020)" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 138, + 502, + 480, + 545 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 502, + 480, + 545 + ], + "spans": [ + { + "bbox": [ + 138, + 502, + 480, + 545 + ], + "type": "text", + "content": "9. Cheng, B., Misra, I., Schwing, A.G., Kirillov, A., Girdhar, R.: Masked-attention mask transformer for universal image segmentation. In: Proceedings of the IEEE/CVF conference on computer vision and pattern recognition. pp. 1290-1299 (2022)" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 138, + 546, + 480, + 578 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 546, + 480, + 578 + ], + "spans": [ + { + "bbox": [ + 138, + 546, + 480, + 578 + ], + "type": "text", + "content": "0. Cheng, B., Schwing, A., Kirillov, A.: Per-pixel classification is not all you need for semantic segmentation. Advances in Neural Information Processing Systems 34, 17864-17875 (2021)" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 138, + 578, + 480, + 610 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 578, + 480, + 610 + ], + "spans": [ + { + "bbox": [ + 138, + 578, + 480, + 610 + ], + "type": "text", + "content": "1. Dahnert, M., Hou, J., Nießner, M., Dai, A.: Panoptic 3D scene reconstruction from a single rgb image. Advances in Neural Information Processing Systems 34, 8282-8293 (2021)" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 134, + 611, + 480, + 643 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 134, + 611, + 480, + 643 + ], + "spans": [ + { + "bbox": [ + 134, + 611, + 480, + 643 + ], + "type": "text", + "content": "12. Dai, A., Chang, A.X., Savva, M., Halber, M., Funkhouser, T., Nießner, M.: Scannet: Richly-annotated 3D reconstructions of indoor scenes. In: Proceedings of the IEEE conference on computer vision and pattern recognition. pp. 5828-5839 (2017)" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 134, + 644, + 480, + 665 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 134, + 644, + 480, + 665 + ], + "spans": [ + { + "bbox": [ + 134, + 644, + 480, + 665 + ], + "type": "text", + "content": "13. Der Kiureghian, A., Ditlevsen, O.: Aleatory or epistemic? does it matter? Structural safety 31(2), 105-112 (2009)" + } + ] + } + ], + "index": 17 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 192, + 91, + 447, + 102 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 192, + 91, + 447, + 102 + ], + "spans": [ + { + "bbox": [ + 192, + 91, + 447, + 102 + ], + "type": "text", + "content": "PCF-Lift: Panoptic Lifting by Probabilistic Contrastive Fusion" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 470, + 92, + 481, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 470, + 92, + 481, + 100 + ], + "spans": [ + { + "bbox": [ + 470, + 92, + 481, + 100 + ], + "type": "text", + "content": "15" + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 14 + }, + { + "para_blocks": [ + { + "bbox": [ + 132, + 116, + 481, + 665 + ], + "type": "list", + "angle": 0, + "index": 20, + "blocks": [ + { + "bbox": [ + 133, + 116, + 480, + 149 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 116, + 480, + 149 + ], + "spans": [ + { + "bbox": [ + 133, + 116, + 480, + 149 + ], + "type": "text", + "content": "14. Fan, Z., Wang, P., Jiang, Y., Gong, X., Xu, D., Wang, Z.: NeRF-SOS: Any-iew self-supervised object segmentation on complex scenes. arXiv preprint arXiv:2209.08776 (2022)" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 132, + 150, + 481, + 184 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 150, + 481, + 184 + ], + "spans": [ + { + "bbox": [ + 132, + 150, + 481, + 184 + ], + "type": "text", + "content": "15. Gasperini, S., Mahani, M.A.N., Marcos-Ramiro, A., Navab, N., Tombari, F.: Panostter: End-to-end panoptic segmentation of LiDAR point clouds. IEEE Robotics and Automation Letters 6(2), 3216-3223 (2021)" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 132, + 185, + 480, + 217 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 185, + 480, + 217 + ], + "spans": [ + { + "bbox": [ + 132, + 185, + 480, + 217 + ], + "type": "text", + "content": "16. Gupta, A., Dollar, P., Girshick, R.: LVIS: A dataset for large vocabulary instance segmentation. In: Proceedings of the IEEE/CVF conference on computer vision and pattern recognition. pp. 5356-5364 (2019)" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 132, + 218, + 480, + 251 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 218, + 480, + 251 + ], + "spans": [ + { + "bbox": [ + 132, + 218, + 480, + 251 + ], + "type": "text", + "content": "17. Hamilton, M., Zhang, Z., Hariharan, B., Snavely, N., Freeman, W.T.: Unsupervised semantic segmentation by distilling feature correspondences. arXiv preprint arXiv:2203.08414 (2022)" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 132, + 251, + 481, + 274 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 251, + 481, + 274 + ], + "spans": [ + { + "bbox": [ + 132, + 251, + 481, + 274 + ], + "type": "text", + "content": "18. Jebara, T., Kondor, R., Howard, A.: Probability product kernels. The Journal of Machine Learning Research 5, 819-844 (2004)" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 132, + 274, + 480, + 296 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 274, + 480, + 296 + ], + "spans": [ + { + "bbox": [ + 132, + 274, + 480, + 296 + ], + "type": "text", + "content": "19. Kajiya, J.T., Von Herzen, B.P.: Ray tracing volume densities. ACM SIGGRAPH computer graphics 18(3), 165-174 (1984)" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 132, + 297, + 480, + 319 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 297, + 480, + 319 + ], + "spans": [ + { + "bbox": [ + 132, + 297, + 480, + 319 + ], + "type": "text", + "content": "20. Kerbl, B., Kopanas, G., Leimkuhler, T., Drettakis, G.: 3D gaussian splatting for real-time radiance field rendering. ACM Transactions on Graphics 42(4) (2023)" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 132, + 319, + 480, + 352 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 319, + 480, + 352 + ], + "spans": [ + { + "bbox": [ + 132, + 319, + 480, + 352 + ], + "type": "text", + "content": "21. Kerr, J., Kim, C.M., Goldberg, K., Kanazawa, A., Tancik, M.: LERF: Language embedded radiance fields. In: Proceedings of the IEEE/CVF International Conference on Computer Vision. pp. 19729-19739 (2023)" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 132, + 352, + 480, + 374 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 352, + 480, + 374 + ], + "spans": [ + { + "bbox": [ + 132, + 352, + 480, + 374 + ], + "type": "text", + "content": "22. Kingma, D.P., Welling, M.: Auto-encoding variational bayes. arXiv preprint arXiv:1312.6114 (2013)" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 132, + 375, + 480, + 407 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 375, + 480, + 407 + ], + "spans": [ + { + "bbox": [ + 132, + 375, + 480, + 407 + ], + "type": "text", + "content": "23. Kirillov, A., He, K., Girshick, R., Rother, C., Dollar, P.: Panoptic segmentation. In: Proceedings of the IEEE/CVF conference on computer vision and pattern recognition. pp. 9404-9413 (2019)" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 132, + 408, + 480, + 441 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 408, + 480, + 441 + ], + "spans": [ + { + "bbox": [ + 132, + 408, + 480, + 441 + ], + "type": "text", + "content": "24. Kirillov, A., Mintun, E., Ravi, N., Mao, H., Rolland, C., Gustafson, L., Xiao, T., Whitehead, S., Berg, A.C., Lo, W.Y., et al.: Segment anything. arXiv preprint arXiv:2304.02643 (2023)" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 132, + 442, + 480, + 475 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 442, + 480, + 475 + ], + "spans": [ + { + "bbox": [ + 132, + 442, + 480, + 475 + ], + "type": "text", + "content": "25. Kobayashi, S., Matsumoto, E., Sitzmann, V.: Decomposing NeRF for editing via feature field distillation. Advances in Neural Information Processing Systems 35, 23311-23330 (2022)" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 132, + 475, + 480, + 520 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 475, + 480, + 520 + ], + "spans": [ + { + "bbox": [ + 132, + 475, + 480, + 520 + ], + "type": "text", + "content": "26. Kundu, A., Genova, K., Yin, X., Fathi, A., Pantofaru, C., Guibas, L.J., Tagliasacchi, A., Dellaert, F., Funkhouser, T.: Panoptic neural fields: A semantic object-aware neural scene representation. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 12871-12881 (2022)" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 132, + 520, + 480, + 542 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 520, + 480, + 542 + ], + "spans": [ + { + "bbox": [ + 132, + 520, + 480, + 542 + ], + "type": "text", + "content": "27. Le-Khac, P.H., Healy, G., Smeaton, A.F.: Contrastive representation learning: A framework and review. IEEE Access 8, 193907-193934 (2020)" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 132, + 543, + 480, + 586 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 543, + 480, + 586 + ], + "spans": [ + { + "bbox": [ + 132, + 543, + 480, + 586 + ], + "type": "text", + "content": "28. Lin, T.Y., Maire, M., Belongie, S., Hays, J., Perona, P., Ramanan, D., Dollar, P., Zitnick, C.L.: Microsoft COCO: Common objects in context. In: Computer Vision-ECCV 2014: 13th European Conference, Zurich, Switzerland, September 6-12, 2014, Proceedings, Part V 13. pp. 740-755. Springer (2014)" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 132, + 586, + 480, + 609 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 586, + 480, + 609 + ], + "spans": [ + { + "bbox": [ + 132, + 586, + 480, + 609 + ], + "type": "text", + "content": "29. Liu, L., Gu, J., Zaw Lin, K., Chua, T.S., Theobalt, C.: Neural sparse voxel fields. Advances in Neural Information Processing Systems 33, 15651-15663 (2020)" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 132, + 610, + 480, + 632 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 610, + 480, + 632 + ], + "spans": [ + { + "bbox": [ + 132, + 610, + 480, + 632 + ], + "type": "text", + "content": "30. McInnes, L., Healy, J., Astels, S.: hdbscan: Hierarchical density based clustering. J. Open Source Softw. 2(11), 205 (2017)" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 132, + 632, + 480, + 665 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 632, + 480, + 665 + ], + "spans": [ + { + "bbox": [ + 132, + 632, + 480, + 665 + ], + "type": "text", + "content": "31. Mildenhall, B., Srinivasan, P.P., Tancik, M., Barron, J.T., Ramamoorthi, R., Ng, R.: NeRF: Representing scenes as neural radiance fields for view synthesis. Communications of the ACM 65(1), 99-106 (2021)" + } + ] + } + ], + "index": 19 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "spans": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "type": "text", + "content": "16" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 166, + 91, + 285, + 102 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 166, + 91, + 285, + 102 + ], + "spans": [ + { + "bbox": [ + 166, + 91, + 285, + 102 + ], + "type": "text", + "content": "R. Zhu, S. Qiu, Q. Wu, et al." + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 15 + }, + { + "para_blocks": [ + { + "bbox": [ + 130, + 116, + 482, + 666 + ], + "type": "list", + "angle": 0, + "index": 19, + "blocks": [ + { + "bbox": [ + 130, + 116, + 482, + 149 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 116, + 482, + 149 + ], + "spans": [ + { + "bbox": [ + 130, + 116, + 482, + 149 + ], + "type": "text", + "content": "32. Milioto, A., Behley, J., McCool, C., Stachniss, C.: LiDAR panoptic segmentation for autonomous driving. In: 2020 IEEE/RSJ International Conference on Intelligent Robots and Systems (IROS). pp. 8505-8512. IEEE (2020)" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 130, + 149, + 482, + 182 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 149, + 482, + 182 + ], + "spans": [ + { + "bbox": [ + 130, + 149, + 482, + 182 + ], + "type": "text", + "content": "33. Narita, G., Seno, T., Ishikawa, T., Kaji, Y.: Panopticfusion: Online volumetric semantic mapping at the level of stuff and things. In: 2019 IEEE/RSJ International Conference on Intelligent Robots and Systems (IROS). pp. 4205-4212. IEEE (2019)" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 130, + 182, + 482, + 213 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 182, + 482, + 213 + ], + "spans": [ + { + "bbox": [ + 130, + 182, + 482, + 213 + ], + "type": "text", + "content": "34. Newell, A., Huang, Z., Deng, J.: Associative embedding: End-to-end learning for joint detection and grouping. Advances in neural information processing systems 30 (2017)" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 130, + 213, + 482, + 257 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 213, + 482, + 257 + ], + "spans": [ + { + "bbox": [ + 130, + 213, + 482, + 257 + ], + "type": "text", + "content": "35. Nie, Y., Han, X., Guo, S., Zheng, Y., Chang, J., Zhang, J.J.: Total3D Understanding: Joint layout, object pose and mesh reconstruction for indoor scenes from a single image. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 55-64 (2020)" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 130, + 257, + 482, + 289 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 257, + 482, + 289 + ], + "spans": [ + { + "bbox": [ + 130, + 257, + 482, + 289 + ], + "type": "text", + "content": "36. Park, J., Lee, J., Kim, I.J., Sohn, K.: Probabilistic representations for video contrastive learning. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 14711-14721 (2022)" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 130, + 289, + 482, + 322 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 289, + 482, + 322 + ], + "spans": [ + { + "bbox": [ + 130, + 289, + 482, + 322 + ], + "type": "text", + "content": "37. Porzi, L., Bulo, S.R., Colovic, A., Kontschieder, P.: Seamless scene segmentation. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 8277-8286 (2019)" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 130, + 322, + 482, + 354 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 322, + 482, + 354 + ], + "spans": [ + { + "bbox": [ + 130, + 322, + 482, + 354 + ], + "type": "text", + "content": "38. Rezende, D.J., Mohamed, S., Wierstra, D.: Stochastic backpropagation and approximate inference in deep generative models. In: International conference on machine learning. pp. 1278-1286. PMLR (2014)" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 130, + 354, + 482, + 386 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 354, + 482, + 386 + ], + "spans": [ + { + "bbox": [ + 130, + 354, + 482, + 386 + ], + "type": "text", + "content": "39. Rosinol, A., Gupta, A., Abate, M., Shi, J., Carlone, L.: 3D dynamic scene graphs: Actionable spatial perception with places, objects, and humans. arXiv preprint arXiv:2002.06289 (2020)" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 130, + 386, + 482, + 429 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 386, + 482, + 429 + ], + "spans": [ + { + "bbox": [ + 130, + 386, + 482, + 429 + ], + "type": "text", + "content": "40. Siddiqui, Y., Porzi, L., Bulò, S.R., Müller, N., Nießner, M., Dai, A., Kontschieder, P.: Panoptic lifting for 3D scene understanding with neural fields. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 9043-9052 (2023)" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 130, + 429, + 482, + 461 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 429, + 482, + 461 + ], + "spans": [ + { + "bbox": [ + 130, + 429, + 482, + 461 + ], + "type": "text", + "content": "41. Sirohi, K., Mohan, R., Buscher, D., Burgard, W., Valada, A.: Efficientlps: Efficient LiDAR panoptic segmentation. IEEE Transactions on Robotics 38(3), 1894-1914 (2021)" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 130, + 461, + 482, + 494 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 461, + 482, + 494 + ], + "spans": [ + { + "bbox": [ + 130, + 461, + 482, + 494 + ], + "type": "text", + "content": "42. Tang, W., Biqi, Y., Li, X., Liu, Y.H., Heng, P.A., Fu, C.W.: Prototypical variational autoencoder for 3D few-shot object detection. In: Thirty-seventh Conference on Neural Information Processing Systems (2023)" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 130, + 494, + 482, + 525 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 494, + 482, + 525 + ], + "spans": [ + { + "bbox": [ + 130, + 494, + 482, + 525 + ], + "type": "text", + "content": "43. Tschernezki, V., Laina, I., Larlus, D., Vedaldi, A.: Neural feature fusion fields: 3d distillation of self-supervised 2d image representations. In: 2022 International Conference on 3D Vision (3DV). pp. 443-453. IEEE (2022)" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 130, + 525, + 482, + 548 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 525, + 482, + 548 + ], + "spans": [ + { + "bbox": [ + 130, + 525, + 482, + 548 + ], + "type": "text", + "content": "44. Wang, B., Chen, L., Yang, B.: Dm-NeRF: 3D scene geometry decomposition and manipulation from 2D images. arXiv preprint arXiv:2208.07227 (2022)" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 130, + 548, + 482, + 590 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 548, + 482, + 590 + ], + "spans": [ + { + "bbox": [ + 130, + 548, + 482, + 590 + ], + "type": "text", + "content": "45. Xie, S., Gu, J., Guo, D., Qi, C.R., Guibas, L., Litany, O.: Pointcontrast: Unsupervised pre-training for 3D point cloud understanding. In: Computer Vision-ECCV 2020: 16th European Conference, Glasgow, UK, August 23-28, 2020, Proceedings, Part III 16. pp. 574-591. Springer (2020)" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 130, + 590, + 482, + 612 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 590, + 482, + 612 + ], + "spans": [ + { + "bbox": [ + 130, + 590, + 482, + 612 + ], + "type": "text", + "content": "46. Ye, M., Danelljan, M., Yu, F., Ke, L.: Gaussian grouping: Segment and edit anything in 3D scenes. arXiv preprint arXiv:2312.00732 (2023)" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 130, + 612, + 482, + 632 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 612, + 482, + 632 + ], + "spans": [ + { + "bbox": [ + 130, + 612, + 482, + 632 + ], + "type": "text", + "content": "47. Zhang, W., Pang, J., Chen, K., Loy, C.C.: K-net: Towards unified image segmentation. Advances in Neural Information Processing Systems 34, 10326-10338 (2021)" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 130, + 632, + 482, + 666 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 632, + 482, + 666 + ], + "spans": [ + { + "bbox": [ + 130, + 632, + 482, + 666 + ], + "type": "text", + "content": "48. Zhang, X., Chen, Z., Wei, F., Tu, Z.: Uni-3D: A universal model for panoptic 3D scene reconstruction. In: Proceedings of the IEEE/CVF International Conference on Computer Vision. pp. 9256-9266 (2023)" + } + ] + } + ], + "index": 18 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 192, + 91, + 447, + 102 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 192, + 91, + 447, + 102 + ], + "spans": [ + { + "bbox": [ + 192, + 91, + 447, + 102 + ], + "type": "text", + "content": "PCF-Lift: Panoptic Lifting by Probabilistic Contrastive Fusion" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 470, + 92, + 481, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 470, + 92, + 481, + 100 + ], + "spans": [ + { + "bbox": [ + 470, + 92, + 481, + 100 + ], + "type": "text", + "content": "17" + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 16 + }, + { + "para_blocks": [ + { + "bbox": [ + 132, + 116, + 480, + 216 + ], + "type": "list", + "angle": 0, + "index": 5, + "blocks": [ + { + "bbox": [ + 132, + 116, + 480, + 149 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 116, + 480, + 149 + ], + "spans": [ + { + "bbox": [ + 132, + 116, + 480, + 149 + ], + "type": "text", + "content": "49. Zhi, S., Laidlow, T., Leutenegger, S., Davison, A.J.: In-place scene labelling and understanding with implicit scene representation. In: Proceedings of the IEEE/CVF International Conference on Computer Vision. pp. 15838-15847 (2021)" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 132, + 150, + 480, + 182 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 150, + 480, + 182 + ], + "spans": [ + { + "bbox": [ + 132, + 150, + 480, + 182 + ], + "type": "text", + "content": "50. Zhou, X., Girdhar, R., Joulin, A., Krahenbuhl, P., Misra, I.: Detecting twenty-thousand classes using image-level supervision. In: European Conference on Computer Vision. pp. 350-368. Springer (2022)" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 132, + 182, + 480, + 216 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 182, + 480, + 216 + ], + "spans": [ + { + "bbox": [ + 132, + 182, + 480, + 216 + ], + "type": "text", + "content": "51. Zhou, Z., Zhang, Y., Foroosh, H.: Panoptic-polarnet: Proposal-free LiDAR point cloud panoptic segmentation. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 13194-13203 (2021)" + } + ] + } + ], + "index": 4 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "spans": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "type": "text", + "content": "18" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 166, + 91, + 285, + 102 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 166, + 91, + 285, + 102 + ], + "spans": [ + { + "bbox": [ + 166, + 91, + 285, + 102 + ], + "type": "text", + "content": "R. Zhu, S. Qiu, Q. Wu, et al." + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 17 + } + ], + "_backend": "vlm", + "_version_name": "2.6.4" +} \ No newline at end of file diff --git a/2024/PDT Uav Target Detection Dataset for Pests and Diseases Tree/77c7722b-38da-423b-b56a-11a16b09eb35_content_list.json b/2024/PDT Uav Target Detection Dataset for Pests and Diseases Tree/77c7722b-38da-423b-b56a-11a16b09eb35_content_list.json new file mode 100644 index 0000000000000000000000000000000000000000..9503923bef3e438bd25dd2f959c12446559fac26 --- /dev/null +++ b/2024/PDT Uav Target Detection Dataset for Pests and Diseases Tree/77c7722b-38da-423b-b56a-11a16b09eb35_content_list.json @@ -0,0 +1,1548 @@ +[ + { + "type": "text", + "text": "PDT: Uav Target Detection Dataset for Pests and Diseases Tree", + "text_level": 1, + "bbox": [ + 238, + 140, + 763, + 183 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Mingle Zhou $^{1,2,4}$ , Rui Xing $^{1,2}$ , Delong Han $^{1,2}$ , Zhiyong Qi $^{3}$ , and Gang Li $^{1,2(\\boxtimes)}$", + "bbox": [ + 316, + 210, + 684, + 244 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "1 Key Laboratory of Computing Power Network and Information Security, Ministry of Education, Shandong Computer Science Center (National Supercomputer Center in Jinan), Qilu University of Technology (Shandong Academy of Sciences), Jinan, China lig@qlu.edu.cn", + "bbox": [ + 212, + 253, + 785, + 310 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "$^{2}$ Shandong Provincial Key Laboratory of Computer Networks, Shandong Fundamental Research Center for Computer Science, Jinan, China", + "bbox": [ + 251, + 309, + 746, + 338 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "3 Institute of Remote Sensing and Geographic Information System, School of Earth and Space Sciences, Peking University, Beijing, China", + "bbox": [ + 220, + 337, + 779, + 366 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "4 SHANDONG SCICOM Information and Economy Research Institute Co., Ltd.", + "bbox": [ + 228, + 364, + 769, + 378 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Abstract. UAVs emerge as the optimal carriers for visual weed identification and integrated pest and disease management in crops. However, the absence of specialized datasets impedes the advancement of model development in this domain. To address this, we have developed the Pests and Diseases Tree dataset (PDT dataset). PDT dataset represents the first high-precision UAV-based dataset for targeted detection of tree pests and diseases, which is collected in real-world operational environments and aims to fill the gap in available datasets for this field. Moreover, by aggregating public datasets and network data, we further introduced the Common Weed and Crop dataset (CWC dataset) to address the challenge of inadequate classification capabilities of test models within datasets for this field. Finally, we propose the YOLO-Dense Pest (YOLO-DP) model for high-precision object detection of weed, pest, and disease crop images. We re-evaluate the state-of-the-art detection models with our proposed PDT dataset and CWC dataset, showing the completeness of the dataset and the effectiveness of the YOLO-DP. The proposed PDT dataset, CWC dataset, and YOLO-DP model are presented at https://github.com/RuiXing123/PDT_CWC_YOLO-DP.", + "bbox": [ + 259, + 415, + 743, + 666 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Keywords: Datasets $\\cdot$ UAVs $\\cdot$ Object detection", + "bbox": [ + 261, + 678, + 588, + 691 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "1 Introduction", + "text_level": 1, + "bbox": [ + 215, + 718, + 375, + 734 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "The efficacy of intelligent UAV-based plant protection operations hinges on precisely identifying weeds and pests within imagery, representing a critical challenge in computer vision [6,9,13]. As computer vision and UAV technologies have advanced swiftly, the adoption of automated UAV plant protection operations is on the rise [3,14]. The working mode of plant protection UAV has significant defects: non-intelligent work can result in the misuse or waste of agrochemicals,", + "bbox": [ + 212, + 750, + 787, + 840 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/80bc44f7affd18d99c46ff7d88fc98a204464af39040f5714c04d1070d202eff.jpg", + "image_caption": [ + "Fig. 1: Dataset comparison. (a) shows the PDT dataset (Low Resolution(LL) and High Resolution(LH)): $640 \\times 640$ , $5472 \\times 3648$ . (b) shows the characteristics of the CWC dataset: it contains 11 different similar plants. (c), (d) and (e) are the public datasets." + ], + "image_footnote": [], + "bbox": [ + 246, + 143, + 754, + 320 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "and the high-intensity, repetitive mechanical tasks can diminish their operational lifespan. Enhancing the precision of weed and pest target recognition is essential for the effectiveness of intelligent plant protection UAVs. The absence of specialized datasets and the limitations of existing detection models pose significant constraints on advancing research in this domain.", + "bbox": [ + 212, + 406, + 782, + 479 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Drones typically take images at both high and low altitudes when outdoors. Existing datasets are usually taken from indoor greenhouses (Fig. 1 (c)), where the majority of the data consists of target images captured at close range (Fig. 1 (d)). Such data struggles to incorporate environmental factors like changes in lighting, and the target sizes significantly deviate from real-world conditions, failing to satisfy practical requirements. Further, the majority of existing datasets are limited to single or double classes, which cannot meet the training of model classification ability, as illustrated in Fig. 1 (c), (d), and (e).", + "bbox": [ + 212, + 482, + 784, + 602 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "On the other hand, the conventional detection models utilized in UAV-based intelligent plant protection could be more challenging in accurately identifying targets. The deficiency in this field is underscored by the absence of tailored baseline models, resulting in unmet demands for detecting and managing diverse weeds, crops, pests, and diseases. This paper aims to address the lack of dedicated datasets and algorithms for weeds, pests and diseases in agricultural drones. Overall, the main contributions of this paper are as follows:", + "bbox": [ + 212, + 602, + 784, + 708 + ], + "page_idx": 1 + }, + { + "type": "list", + "sub_type": "text", + "list_items": [ + "- We have developed the UAV target detection agricultural Pests and Diseases Tree dataset, known as PDT dataset.", + "- We collated public data and provided Common Weed and Crop dataset (CWC dataset) to train the classification capabilities of the model.", + "- We have designed a specialized detection model, YOLO-DP, tailored for the dense object detection of tree crop pests and diseases.", + "- We reevaluated and analyzed existing generic and specific detection models on PDT and CWC datasets to establish benchmarks for the field." + ], + "bbox": [ + 223, + 718, + 782, + 838 + ], + "page_idx": 1 + }, + { + "type": "page_number", + "text": "2", + "bbox": [ + 217, + 114, + 228, + 126 + ], + "page_idx": 1 + }, + { + "type": "header", + "text": "M. Zhou et al.", + "bbox": [ + 271, + 114, + 369, + 127 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "2 Related Work", + "text_level": 1, + "bbox": [ + 215, + 143, + 387, + 160 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Agricultural UAV Detection Dataset. In the realm of agricultural robotics, acquiring pertinent datasets frequently presents a challenge, primarily due to the necessity of maintaining dedicated experimental fields and the critical timing of data collection, which is pivotal to the research outcomes. Chebrolu et al. presented SugarBeet2016 large agricultural robot dataset and subsequently introduced the SugarBeet2017 dataset tailored for the detection task [8,11]. The limitation of this dataset is that the data, captured from the vantage point of a ground robot, significantly diverges from the actual detection scenarios encountered by UAVs. Vakanski et al. introduced a dataset of multispectral potato plant images taken by drones [5]. The limitation of this dataset is the lack of raw data, consisting of only 360 images and a total of 1,500 images produced by data enhancement, which cannot effectively guide the model to learn the details of the crop. The University of Burgundy has launched the Ribworth Dataset, designed for crops and weeds, which uses data augmentation to simulate real-world environmental conditions [4]. The limitation of this dataset is that its acquisition target is relatively close, and the factors such as light change are not included, which is not enough to meet the requirements of accurate detection tasks of agricultural plant protection UAVs.", + "bbox": [ + 212, + 175, + 787, + 445 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "To foster the advancement of precision agriculture, we introduce the PDT dataset and CWC dataset, aimed at addressing the requirements for detecting agricultural weeds, crops, and pests.", + "bbox": [ + 212, + 446, + 787, + 492 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Agricultural UAV Detection Model. The agricultural UAV target detection task has garnered significant attention within the realms of machine learning and computer vision, owing to its attributes of precise localization, efficient operation, and minimal environmental impact. Among them, the weed detection models based on YOLO deep learning model include: Zhang et al. developed a detection model for weed location and identification in wheat fields by using YOLOv3-tiny [31]. Gao et al. modified the YOLOv3-tiny model and applied it to weed detection in sugar beet fields. Jabir and Falih proposed an intelligent wheat field weed detection system based on YOLOv5 [15]. Furthermore, a variety of neural network architectures have also been recognized by researchers and implemented in the field of precision agriculture. Guo et al. proposed WeedNet-R based on the RetinaNet architecture for weed identification and location in sugar beet fields [11]. Tang et al. employed Kmeans feature learning in conjunction with a convolutional neural network for the recognition of weeds [29]. Agarwal et al. utilized the Kmeans clustering algorithm to analyze color space features in multispectral weed images for effective weed detection [1].", + "bbox": [ + 212, + 492, + 787, + 734 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "It is evident that there is a scarcity of studies focusing on baseline models for the detection of tree diseases and pests, particularly those related to the invasion of the exotic species Red Turpentine Beetle. Furthermore, the majority of these baseline models are trained using single-class or two-class detection datasets, and the classification ability of the models is limited by the diversity of the data. However, for the detection tasks in large-scale, high-precision agricultural UAV operations, precise detection, accurate classification, and the expansion of", + "bbox": [ + 212, + 734, + 787, + 839 + ], + "page_idx": 2 + }, + { + "type": "header", + "text": "PDT Dataset", + "bbox": [ + 638, + 114, + 730, + 126 + ], + "page_idx": 2 + }, + { + "type": "page_number", + "text": "3", + "bbox": [ + 774, + 116, + 784, + 126 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "crop detection varieties are essential. Consequently, we propose the YOLO-DP detection model to address these requirements.", + "bbox": [ + 212, + 146, + 782, + 176 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "3 PDT Dataset", + "text_level": 1, + "bbox": [ + 215, + 200, + 382, + 215 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "In this study, we introduced a PDT dataset with both high and low resolution versions, specifically for the detection of red turpentine beetle natural pest targets. As most existing public datasets are gathered indoors or via simulated drones, they fail to capture the characteristics of outdoor UAV detection environments. Consequently, there is an immediate demand for an agricultural dataset that is tailored to real-world operational environments and encompasses the distribution of both high and low-altitude data, thereby fulfilling the high-precision, large-scale target detection requirements from a UAV perspective. This section delineates the data acquisition process for the PDT dataset, encompassing the selection of the data collection domain, the equipment used for acquisition, and the criteria for sample definition.", + "bbox": [ + 212, + 232, + 787, + 398 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "3.1 Data Acquisition", + "text_level": 1, + "bbox": [ + 215, + 421, + 401, + 436 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Selection of the Data Collection Domain. The Red Turpentine Beetle, a member of the Coleopteridae family, not only invades extensive pine forest areas but also contributes to the widespread mortality of trees, exerting a profound impact on the regional ecological environment and forestry. Despite the implementation of preventive and control strategies, including monitoring and trapping, effectively managing its continued proliferation remains a formidable challenge. The precision spraying platform of plant protection UAV is an effective solution, but the public UAV data for the invasive species Red Turpentine Beetle is blank. Consequently, we opted to collect data from a cluster of dead pine trees that have been infested by the Red Turpentine Beetle.", + "bbox": [ + 212, + 446, + 787, + 598 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "High-resolution Drone Camera Equipment. To obtain accurate sample data, we use a UAV mapping camera called DJI-ChanSi L2. This advanced system enhances the UAV flight platform's capabilities for acquiring accurate, efficient, and reliable 2D and 3D data. The LiDAR components have excellent performance, ranging accuracy of $2\\mathrm{cm}$ at a height of 150 meters, laser wavelength of $905~\\mathrm{nm}$ , and laser pulse emission frequency of $240\\mathrm{kHz}$ . The system supports the exFAT file system and can capture images in JPEG or PNG formats. In essence, the DJI-ChanSi L2 is a comprehensive platform that integrates lidar, a high-precision navigation system, and a mapping camera, specifically designed for high-precision data acquisition in mapping, forestry, power grid, and engineering and construction industries.", + "bbox": [ + 212, + 598, + 787, + 763 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Definition of Detection Target. Distinct from close-range pest detection, the task of pest detection at low or high altitudes fundamentally involves capturing the differential characteristics between the affected vector and the healthy state. Consequently, the detection samples for the PDT dataset are sourced from both healthy and unhealthy trees within pine forests. As depicted in Fig. 2, (a)", + "bbox": [ + 212, + 763, + 787, + 840 + ], + "page_idx": 3 + }, + { + "type": "page_number", + "text": "4", + "bbox": [ + 217, + 114, + 228, + 126 + ], + "page_idx": 3 + }, + { + "type": "header", + "text": "M. Zhou et al.", + "bbox": [ + 271, + 114, + 367, + 126 + ], + "page_idx": 3 + }, + { + "type": "image", + "img_path": "images/5bbf9b916d924007185e084ba54e0f49e4f09471caa70f75bcaff0f65546d111.jpg", + "image_caption": [ + "(a): healthy" + ], + "image_footnote": [], + "bbox": [ + 287, + 147, + 383, + 209 + ], + "page_idx": 4 + }, + { + "type": "image", + "img_path": "images/ee7711eb2a5d834cf5ed6528fe0ac0d83ba72e8187011d979422b09c67bc5361.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 401, + 147, + 486, + 208 + ], + "page_idx": 4 + }, + { + "type": "image", + "img_path": "images/d587f16990c6e14f0c0d356df49c131eb76bf54aad1e2fe7cbcf6f1104420253.jpg", + "image_caption": [ + "(b): unhealthy" + ], + "image_footnote": [], + "bbox": [ + 509, + 147, + 588, + 208 + ], + "page_idx": 4 + }, + { + "type": "image", + "img_path": "images/6c094c81bda8c2ebf78c1b9ceeeb225096f2034298407262f7dd3159b417e666.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 609, + 147, + 709, + 208 + ], + "page_idx": 4 + }, + { + "type": "image", + "img_path": "images/c3450a34a4019f05335b2fd838608c5690e342ce7305db900ff3b4b8d7aae41c.jpg", + "image_caption": [ + "Fig. 2: Data example. (a) is a healthy goal and (b) is a unhealthy goal. The PDT dataset takes (b) as the category.", + "Fig. 3: PDT dataset generation and detection process. (a) represents the sliding window method, and (b) represents the \"Human-in-the-loop\" data annotation method. (c) means that a LL image is sent to the neural network for training, and LL and LH dual-resolution images are detected at the same time." + ], + "image_footnote": [], + "bbox": [ + 245, + 282, + 754, + 406 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "illustrates a pine tree in a healthy state, while (b) displays three pine trees in a unhealthy state that have been infested by pests, exhibiting progressively severe symptoms from left to right.", + "bbox": [ + 212, + 507, + 784, + 551 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "3.2 Data Preprocessing and Statistics", + "text_level": 1, + "bbox": [ + 214, + 574, + 539, + 589 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "This section delves into the acquisition, processing, and production procedures of the original image data about the PDT dataset. We set up an isometric photo mode at an altitude of 150 meters to acquire 105 sets of raw image data and 3D point cloud data. Owing to the substantial size of the original images, they are unsuitable for training with the current mainstream object detection neural networks. We introduce a training approach that applies to the majority of neural network data. As illustrated in Fig. 3, we employ a methodology that encompasses \"data preprocessing - 'Human-in-the-loop' data labeling - manual correction\". The essence of this method is to leverage the preliminary analysis capabilities of artificial intelligence to aid human annotators, thereby enhancing the efficiency and accuracy of data annotation.", + "bbox": [ + 212, + 598, + 785, + 763 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Data Preprocessing. Fig. 3 (a) demonstrates the application of the sliding window technique on the original image data, which allows for the extraction of a usable, standard-sized image through regional cropping of the high-resolution image. To prevent information loss during the sliding window process, the window dimensions are set to $640 \\times 640$ pixels, with a step size of $630 \\times 630$ pixels.", + "bbox": [ + 212, + 763, + 785, + 839 + ], + "page_idx": 4 + }, + { + "type": "header", + "text": "PDT Dataset", + "bbox": [ + 638, + 114, + 730, + 126 + ], + "page_idx": 4 + }, + { + "type": "page_number", + "text": "5", + "bbox": [ + 774, + 116, + 784, + 126 + ], + "page_idx": 4 + }, + { + "type": "table", + "img_path": "images/f4064f094581d8868ecd333df2773089d9c97fa7abadbfb9794325f598a64bbd.jpg", + "table_caption": [ + "Table 1: Structure of PDT dataset and its sample." + ], + "table_footnote": [], + "table_body": "
EditionClassesStructureTargeted imagesUntargeted imagesImage sizeInstancesTarget Amount
S(Small)M(Medium)L(Large)
SampleunhealthyTrain811640×64025691896548179
Val1969152813825
LLunhealthyTrain316613709029070418163423530
Val395172640×6401252399262165432
Test3901771149489492095450
LHunhealthy-10505472×3648934749347400
", + "bbox": [ + 274, + 183, + 733, + 272 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Following the sliding window operation, 53 LL images were derived from a single LH image. Furthermore, due to the high similarity between the unhealthy pine tree depicted in Fig. 2 and the ground background, and to enhance the effectiveness of neural network training, we opt to retain images that do not contain targets. Upon generating an output from these data during training, a corresponding loss is produced, which aids the neural network in learning and reduces the rate of false positives.", + "bbox": [ + 212, + 311, + 785, + 417 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Data Annotation. We proceed to annotate the preprocessed data. To ensure the data's validity, we utilize the widely-used Labelme ${}^{1}$ data labeling software. For high-resolution datasets with dense, small targets, the sheer volume of targets per image makes manual annotation a tedious and time-intensive process. As illustrated in Fig. 3 (b), the \"Human-in-the-loop\" data annotation approach effectively addresses this challenge. Initially, we acquire the LL version samples of the PDT dataset through manual annotation. The structure of the sample dataset is depicted in Tab. 1, which comprises a training set and a validation set. Subsequently, the sample dataset is input into the YOLOv5s model for training over 300 epochs, with the optimal weights being saved. To ensure that no label data is missed, a confidence threshold of 0.25 and an NMS IOU (Intersection over Union) threshold of 0.45 were set to automatically label the LL data and obtain the original label set.", + "bbox": [ + 212, + 419, + 787, + 616 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Manual Check. The labeled dataset underwent manual filtering to eliminate incorrect annotations, rectify erroneous ones, and add any omitted annotations. The LL datasets were generated in two widely recognized formats, YOLO txt and VOCXml, by automatically partitioning the data structure into an 8:1:1 ratio using scripts. For the LH data, we input the completed LL data into the YOLO-DP model for training to acquire the optimal weights. The \"Human-in-the-loop\" data annotation and subsequent manual verification were conducted on the LH data to obtain the final LH version dataset.", + "bbox": [ + 212, + 616, + 787, + 736 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Data Statistics. The statistics of the PDT dataset are shown in Tab. 1. LL and LH versions consist of 5,670 and 105 images, respectively, with 114,307 and 93,474 samples, and feature a single class labeled as unhealthy. The training, validation, and test sets of the LL version include 1,370, 172, and 144 target-free images, respectively. On average, the LL and LH versions contain 29 and 890", + "bbox": [ + 212, + 737, + 787, + 814 + ], + "page_idx": 5 + }, + { + "type": "page_number", + "text": "6", + "bbox": [ + 217, + 114, + 228, + 126 + ], + "page_idx": 5 + }, + { + "type": "header", + "text": "M. Zhou et al.", + "bbox": [ + 271, + 114, + 369, + 127 + ], + "page_idx": 5 + }, + { + "type": "page_footnote", + "text": "1 https://github.com labelmeai/labelme", + "bbox": [ + 202, + 824, + 477, + 839 + ], + "page_idx": 5 + }, + { + "type": "table", + "img_path": "images/2c20a19bc947c83bcb5b5bbe0004cf2654986182c548df647ba3afdeef82c8b3.jpg", + "table_caption": [ + "Table 2: CWC dataset sources." + ], + "table_footnote": [], + "table_body": "
DatasetsCorn weed datasets [25]lettuce weed radish weed datasets [25] datasets [25]Fresh-weed-data [18]
Classesbluegrass, corn, sedge, chenopodium album, circium_setosumlettuceradishnightshadetomatocottonvelvet
Number2502002011151162438
Image Size800×600800×600800×600800×600800×600586×444643×500
", + "bbox": [ + 276, + 183, + 730, + 263 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "samples per image, respectively, indicating that the PDT dataset is suitable for the task of high-density and high-precision inspection of plant protection drones.", + "bbox": [ + 212, + 304, + 782, + 335 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "4 CWC Dataset", + "text_level": 1, + "bbox": [ + 215, + 359, + 388, + 376 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "We compile the available data to construct the Common Weed and Crop datasets. In special cases, there may be crop crossovers (usually in small trial plots) where weeds and crop targets exhibit multi-class structural characteristics and similarities. The majority of existing public datasets are limited to single or double classes, which do not suffice for the training and classification capabilities required by models. Consequently, when training the neural network, it is essential to supply detailed texture data that differentiates between various objects, thereby equipping the model with robust classification capabilities. This enables the model to effectively sift through and eliminate highly similar negative samples during detection, thereby enhancing the model's detection accuracy. This section discusses the acquisition and processing of CWC original image data.", + "bbox": [ + 212, + 392, + 787, + 559 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "4.1 Data Sources", + "text_level": 1, + "bbox": [ + 215, + 582, + 372, + 595 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "In Tab. 2, we present the source, categories, initial quantities of the CWC raw data, as well as their respective resolutions. As depicted in Fig. 1 (b), the CWC original data comprises a total of 11 categories, characterized by similar size, texture, and color. Please note that we manually annotated the raw Corn weed datasets, Lettuce weed datasets, and Radish weed datasets, while the Fresh-weed-data was already pre-labeled and available in YOLO txt format.", + "bbox": [ + 212, + 608, + 787, + 700 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "4.2 Data Preprocessing and Statistics", + "text_level": 1, + "bbox": [ + 215, + 722, + 540, + 738 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Data Preprocessing. Owing to the imbalance between positive and negative samples in the original CWC data, there is a risk that the model may become biased towards the majority class, leading to a diminished recognition capability for the minority class. We used the oversampling strategy, using the data augmentation method to equalize the number of samples. The methods include random rotation, random translation, random brightness change, random noise", + "bbox": [ + 212, + 750, + 787, + 840 + ], + "page_idx": 6 + }, + { + "type": "header", + "text": "PDT Dataset", + "bbox": [ + 638, + 114, + 730, + 126 + ], + "page_idx": 6 + }, + { + "type": "page_number", + "text": "7", + "bbox": [ + 774, + 114, + 784, + 125 + ], + "page_idx": 6 + }, + { + "type": "table", + "img_path": "images/25c882fbb2aa0801f7b2a8179d116f69c7070878f2e2943ba7ee05d7608782a4.jpg", + "table_caption": [ + "Table 3: CWC dataset structure." + ], + "table_footnote": [], + "table_body": "
Classesbluegrasschenopodium-album-cirsium-setosumcornsedgelettuceradishnightshadetomatocottonvelvet
Targeted ImagesTrain200200200200200200200200200200200200
Val404040404040404040404040
Test101010101010101010101010
Targeted AmountS100500000003
M000900000000
L249250250236250444326250210268248
Image Size800×600800×600800×600800×600800×600800×600800×600800×600800×600586×444643×500
", + "bbox": [ + 246, + 183, + 759, + 280 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "addition, flipping and cutout. This enhances the robustness of the dataset, simulating the illumination formula: $\\pmb{I}_{out} = \\pmb{I}_{in} \\times w + [255] \\times (1 - w)$ , $w \\in [0.35, 1]$ . Where $\\pmb{I}_{out}$ , $\\pmb{I}_{in}$ is the output and input images. $w$ is the random weight.", + "bbox": [ + 212, + 320, + 782, + 364 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Data Statistics. The statistics for the CWC dataset are presented in Tab. 3. The CWC dataset comprises 2,750 images and 2,599 data entries, with the majority featuring large-sized objects. Data enhancement makes the sample reach a balanced distribution, with an average image containing 1 to 2 objects.", + "bbox": [ + 212, + 366, + 784, + 428 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "5 Dataset Comparison", + "text_level": 1, + "bbox": [ + 215, + 450, + 452, + 467 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Tab. 4 provides a thorough comparison of the datasets. The LH version of the PDT dataset is exceptionally clear, offering image quality that is 35 to 200 times superior to other publicly available datasets. The PDT dataset is characterized by density, small target and rich real environmental details, which accords with the typical environmental factors of high and low altitude UAV work, and is suitable for training the special UAV detection model for plant protection. Furthermore, the PDT dataset includes a 3D point cloud version. While not highlighted in this study, a 3D data version is planned for development and will be released to the public in the future.", + "bbox": [ + 212, + 481, + 787, + 618 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "On the other hand, the CWC dataset demonstrates outstanding classification performance, particularly in terms of category diversity, detailed texture information, and dataset authenticity. CWC dataset has a large number of categories, 2 to 5 times more than other publicly available datasets, which indicates that the CWC dataset is well suited for tasks involving high-precision classification of plant protection UAVs.", + "bbox": [ + 212, + 619, + 787, + 709 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "6 YOLO-DP Model", + "text_level": 1, + "bbox": [ + 215, + 732, + 426, + 748 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "To expedite research efforts in UAV detection models for intelligent agricultural plant protection, we have designed and proposed a model, YOLO-DP (Fig. 4), specifically dedicated to the detection of tree crop diseases and pests, building upon the YOLOv5s foundatin [16]. To ensure the scalability of the model, the Kmeans clustering algorithm is employed to determine the distribution of the", + "bbox": [ + 212, + 763, + 787, + 840 + ], + "page_idx": 7 + }, + { + "type": "page_number", + "text": "8", + "bbox": [ + 217, + 114, + 228, + 126 + ], + "page_idx": 7 + }, + { + "type": "header", + "text": "M. Zhou et al.", + "bbox": [ + 271, + 114, + 369, + 127 + ], + "page_idx": 7 + }, + { + "type": "table", + "img_path": "images/ba80e14297bccb544bcbcc3c2542f2626cef5097b1203f37c6a1c003c7a09e47.jpg", + "table_caption": [ + "Table 4: Dataset comparison." + ], + "table_footnote": [], + "table_body": "
DatasetResolutionClassesRich DetailsHigh DefinitionScaleDense TargetTarget ScaleUav CollectionNo Target Image3D Point Cloud DataAnnotation Quality
PDT dataset5472×36481-✓100+✓S✓✓✓“Human-in-the-loop”
640×6401--5K+✓S/M/L✓✓✓
CWC dataset800×60011✓-2K+-L---Manually
SugarBeet2017 [11]1296×9662--5K+-S/M--✓Manually
Multispectral Potato Plants Images (MPP) [5]750×7502--1K+✓L✓--Manually
Ribworth Dataset (RI) [4]320×2401✓-5K+-M/L---Manually
320×320
Corn weed Dataset [25]800×6005✓-6K+-L---No
lettuce weed Dataset [25]800×6002✓-700+-L---No
radish weed Dataset [25]800×6002✓-400+-L---No
Fresh -weed-data [29]800×6004✓-200+-L---Manually
586×444
643×500
crop and weed detection data (CAW) [10]512×5122✓-1K+-L---Manually
Weeds [7]480×4801--600+-S/M✓--Manually
", + "bbox": [ + 246, + 183, + 759, + 402 + ], + "page_idx": 8 + }, + { + "type": "image", + "img_path": "images/7ef984eff2dea40bd293b288036c7be84dffa1767f08d087465ad7e40587b066.jpg", + "image_caption": [ + "Fig. 4: YOLO-DP baseline model architecture. The FPN [21] + PAN [23] module consists of GhostConv [12], Upsample, Concat, and C3. C stands for Concat, S for Sigmoid, P for channel number dilation, $\\times$ for matrix multiplication, and + for matrix addition." + ], + "image_footnote": [], + "bbox": [ + 246, + 431, + 756, + 535 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "actual bounding box data, allowing for the dynamic adjustment of the anchor box sizes. In light of the UAV's large-scale detection capabilities, an adaptive Large Scale Selective Kernel is incorporated into the Backbone network to capture the location information of dense, small target pest-infested trees [20]. Taking into account the efficient detection mode of UAVs, GhostConv is utilized in the Neck network to reduce the model size and computational complexity. Concurrently, the receptive field is enlarged to capture more texture feature information, thereby enhancing the model's classification capabilities. A version with decoupled detection heads is provided to minimize the interference between classification and regression losses, allowing the model to focus on specific tasks, improving the accuracy and accuracy of the detection, and also improving the model's generalization ability.", + "bbox": [ + 212, + 625, + 787, + 806 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "Adaptive Large Scale Selective Kernel. To tackle the demanding detection task involving a broad range of dense, small targets, we gather pest and", + "bbox": [ + 212, + 809, + 787, + 840 + ], + "page_idx": 8 + }, + { + "type": "header", + "text": "PDT Dataset", + "bbox": [ + 638, + 114, + 730, + 126 + ], + "page_idx": 8 + }, + { + "type": "page_number", + "text": "9", + "bbox": [ + 774, + 114, + 785, + 126 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "disease target information by designing a Convolutional group with an extensive receptive field. Firstly, the matrix M obtains the shallow range information M1 by Conv with $\\text{Kernelsize}(k) = 5$ and $\\text{Padding}(p) = 2$ . The depth information is explored for M1, and the deep range information M2 is obtained using Conv with $k = 7$ and $p = 9$ . Feature selection was performed on shallow and depth range information, and M3 and M4 were obtained by grouping convolution on M1 and M2 using GhostConv with $k = 1$ . The advantage of using GhostConv is that half of the range information is retained and the other half is processed to prevent excessive loss of information. Next, the spatial attention is calculated for M3 and M4 large-scale information. M3 and M4 are concatenated in the channel dimension to obtain M5, and M6 and M7 are obtained by averaging and maximizing the channel dimensions, respectively. M6 and M7 are concatenated in the channel dimension to obtain M8 with scale $b \\times 1 \\times w \\times h$ , which is fed into Conv with $k = 7$ and $p = 4$ for range attention collection to obtain M9. After Sigmoid activation, the channel dimension is expanded to obtain M10 and M11. Matrix multiplication is performed with M3 and M4 to obtain M12 and M13, respectively. After bitwise addition, the final spatial attention matrix M14 is obtained after Conv ( $k = 1$ ). Finally, the output is obtained by matrix multiplication with the input matrix.", + "bbox": [ + 212, + 146, + 787, + 434 + ], + "page_idx": 9 + }, + { + "type": "text", + "text": "7 Experiment", + "text_level": 1, + "bbox": [ + 215, + 455, + 366, + 472 + ], + "page_idx": 9 + }, + { + "type": "text", + "text": "7.1 Experimental Conditions", + "text_level": 1, + "bbox": [ + 215, + 484, + 468, + 500 + ], + "page_idx": 9 + }, + { + "type": "text", + "text": "The experimental environment of this paper is ubuntu 18.04, Python3.9, PyTorch 1.9.1 framework, CUDA 11.4 version, and cuDNN 8.0 version. This paper's model training and inference are performed on NVIDIA A100-SXM4 with 40GB GPU memory and 16 GB CPU memory on the experimental platform. The training metrics in this paper are: 300 epochs of iterative training, the early stopping mechanism is set to 100 rounds, the initial learning rate is 0.01, the weight decay of 0.001 and momentum of $0.9\\%$ are used, the Batch-size is 16.", + "bbox": [ + 212, + 508, + 785, + 613 + ], + "page_idx": 9 + }, + { + "type": "text", + "text": "Experiments were carried out on the PDT dataset and CWC dataset presented in this paper, as well as on the public dataset SugarBeet2017. We test the performance of the model YOLO-DP according to the characteristics of the dataset. First, the PDT dataset is used to test the detection ability of the model for dense small targets in the UAV perspective. Secondly, the ability of the model to extract fine-grained texture information is measured on the CWC dataset to verify its classification ability. It is reflected in the P metrics. Finally, the detection accuracy was measured on the SugarBeet2017 dataset. It is reflected in the R metrics. In order not to lose the image information, the input sizes are $640 \\times 640$ , $800 \\times 800$ , $1296 \\times 1296$ respectively.", + "bbox": [ + 212, + 614, + 785, + 765 + ], + "page_idx": 9 + }, + { + "type": "text", + "text": "7.2 Experimental Analysis", + "text_level": 1, + "bbox": [ + 215, + 786, + 449, + 801 + ], + "page_idx": 9 + }, + { + "type": "text", + "text": "Dataset Validation. We evaluated 6 detection models on seven datasets (2 of ours and 5 public datasets mentioned in the paper). Based on the dataset's", + "bbox": [ + 212, + 809, + 785, + 839 + ], + "page_idx": 9 + }, + { + "type": "page_number", + "text": "10", + "bbox": [ + 217, + 114, + 235, + 126 + ], + "page_idx": 9 + }, + { + "type": "header", + "text": "M. Zhou et al.", + "bbox": [ + 271, + 114, + 367, + 126 + ], + "page_idx": 9 + }, + { + "type": "text", + "text": "characteristics, we choose different metrics for the ranking model (Rank (Metrics)). Specifically, dense target datasets collected in natural environments use the comprehensive F1 metrics, datasets with target size changes use the R metrics, and multiclass datasets use the P metrics. We sort models with the same metrics score again using Gflops. Our model has excellent performance on all 7 datasets, and the performance of the proposed 2 datasets differs from that of existing datasets, indicating that our dataset can provide unique data distribution to the research field (Tab. 5).", + "bbox": [ + 212, + 146, + 787, + 268 + ], + "page_idx": 10 + }, + { + "type": "table", + "img_path": "images/a2501b3721929a165d350bb084ca6fa6cf6d0fda1883aa19cd1a1cb89fb22b5f.jpg", + "table_caption": [ + "Table 5: Datasets effect ranking." + ], + "table_footnote": [], + "table_body": "
MethodGflopsRank (F1)Rank (P)Rank (R)
PDT (LL) (our)MPP [5]CWC (our)Weeds [7]SugarBeet2017 [11]CAW [10]RI [4]
YOLO-DP11.71 (0.89)1 (0.42)2 (92.9%)2 (76.5%)1 (73.8%)2 (89.4%)1 (82.3%)
YOLOv3155.35 (0.88)5 (0.38)6 (86.6%)1 (83.0%)6 (46.2%)6 (73.1%)6 (74.0%)
YOLOv4s20.83 (0.88)2 (0.42)5 (87.3%)3 (75.6%)3 (60.3%)3 (86.5%)2 (81.7%)
YOLOv5s16.02 (0.89)3 (0.38)4 (88.6%)4 (75.3%)4 (58.3%)5 (82.5%)3 (81.5%)
YOLOv7105.16 (0.85)6 (0.24)1 (93.1%)5 (74.4%)5 (48.1%)4 (83.3%)4 (80.3%)
YOLOv8s28.64 (0.88)4 (0.38)3 (92.0%)6 (70.4%)2 (65.0%)1 (90.1%)5 (79.3%)
", + "bbox": [ + 305, + 321, + 702, + 386 + ], + "page_idx": 10 + }, + { + "type": "text", + "text": "Comparative Experiment. We perform a comparative analysis of single-stage detectors including the full YOLO family, SSD, EfficientDet, and RetinaNet, the two-stage detector Fast-RCNN, anchor-free based CenterNet, and WeedNet-R, a specialized model for weed detection. Among them, some models use pre-trained weights. As some models do not provide the computation results for certain metrics, these are denoted by “-” in the table.", + "bbox": [ + 212, + 415, + 787, + 505 + ], + "page_idx": 10 + }, + { + "type": "text", + "text": "In Tab. 6, we present the comparative evaluations of the proposed YOLO-DP baseline model alongside other models across the three datasets. The YOLO-DP model does not need to load pre-trained weights, and the detection speed is significantly ahead of other models. Under this premise, YOLO-DP demonstrates the most outstanding overall performance on the PDT dataset, with its F1 composite metrics surpassing all other models. This shows that the YOLO-DP model can well adapt to the detection scenario of UAV with large-scale, high-precision, dense and small targets. On the CWC dataset, its P metrics is at the first-class level, which is $4.3\\%$ higher than the YOLOv5s baseline model. This demonstrates that the YOLO-DP model possesses an effective capability for extracting fine-grained texture information and is capable of fulfilling the classification task requirements within this domain. On the authoritative dataset SugarBeet2017, while the overall performance does not match that of models with loaded and trained weights, the R metrics of YOLO-DP is comparable to the pre-trained model's performance. This substantiates that the YOLO-DP model possesses precise detection capabilities and can accurately pinpoint the detection targets.", + "bbox": [ + 212, + 506, + 787, + 748 + ], + "page_idx": 10 + }, + { + "type": "text", + "text": "The comparative experimental results show that the YOLO-DP model is suitable for the application of plant protection UAV in precision agriculture detection fields such as weeds, pests and diseases, and crops.", + "bbox": [ + 212, + 748, + 787, + 792 + ], + "page_idx": 10 + }, + { + "type": "text", + "text": "Ablation Experiment. We chose the YOLOv5s that performed well on the PDT dataset as the benchmark. We use the popular attention mechanism to ablate the adaptive large-scale selection kernel to verify its detection performance", + "bbox": [ + 212, + 794, + 787, + 839 + ], + "page_idx": 10 + }, + { + "type": "header", + "text": "PDT Dataset", + "bbox": [ + 638, + 114, + 730, + 126 + ], + "page_idx": 10 + }, + { + "type": "page_number", + "text": "11", + "bbox": [ + 767, + 114, + 784, + 126 + ], + "page_idx": 10 + }, + { + "type": "table", + "img_path": "images/5ee361efae815e1cad7c68eb32e7b983cea315df510350bb2b1a09573339eab8.jpg", + "table_caption": [ + "Table 6: Comparative experiment." + ], + "table_footnote": [], + "table_body": "
DatasetsApproachPRmAP@.5mAP@.5::.95F1GflopsParametersFPS Pre-training
PDT dataset (LL)SSD [24]84.5%87.7%85.1%-0.86273.623.7M37✓
EfficientDet [28]92.6%73.4%72.3%-0.8211.56.7M12✓
RetinaNet [22]93.3%65.3%64.2%-0.79109.732.7M32✓
CenterNet [32]95.2%67.4%66.5%-0.79109.732.7M32✓
Faster-RCNN [27]57.8%70.5%61.7%-0.64401.7136.7M13✓
YOLOv3 [26]88.5%88.1%93.4%65.7%0.88155.361.5M41-
YOLOv4s [2]88.8%88.2%94.7%66.1%0.8820.89.1M51-
YOLOv5s_7_0 [16]88.9%88.5%94.2%67.0%0.8916.07.0M93-
YOLOv6s [19]--91.4%63.2%-44.117.2M43-
YOLOv7 [30]87.4%82.6%90.1%55.5%0.85105.137.2M32-
YOLOv8s [17]88.7%87.5%94.0%67.9%0.8828.611.1M60-
WeedNet-R [11]87.7%48.1%70.4%-0.6219.025.6M0.5-
YOLO-DP (our)90.2%88.0%94.5%67.5%0.8911.75.2M109-
CWC datasetSSD [24]97.7%77.6%85.7%-0.91426.923.7M29✓
EfficientDet [28]97.2%98.6%98.6%-0.9011.56.7M13✓
RetinaNet [22]95.1%98.3%98.0%-0.97261.336.4M24✓
CenterNet [32]96.6%73.8%73.3%-0.80171.432.7M27✓
YOLOv3 [26]86.8%89.4%93.2%82.3%0.88154.761.5M30-
YOLOv4s [2]87.3%87.9%91.9%81.5%0.8820.89.1M43-
YOLOv5s_7_0 [16]88.6%88.7%93.0%81.2%0.8916.07.0M65-
YOLOv6s [19]--92.7%84.3%-68.917.2M31-
YOLOv7 [30]93.1%76.4%88.1%75.6%0.84105.137.2M21-
YOLOv8s [17]92.0%89.1%94.0%86.2%0.9128.611.1M38-
WeedNet-R [11]86.1%51.8%71.6%-0.6519.025.6M0.5-
YOLO-DP (our)92.9%87.5%91.8%81.0%0.9011.75.2M72-
Sugar-Beet2017SSD [24]85.0%83.6%79.3%-0.85112023.7M19✓
EfficientDet [28]93.3%79.8%77.8%-0.8611.56.7M16✓
RetinaNet [22]91.7%78.8%76.6%-0.84256.436.3M23✓
CenterNet [32]97.9%51.2%51.0%-0.62117.432.7M41✓
Faster-RCNN [27]63.6%87.4%80.0%-0.73546.9136.7M25✓
YOLOv3 [26]34.8%46.2%39.4%25.6%0.40155.361.5M28-
YOLOv4s [2]28.1%60.3%41.1%26.4%0.3820.89.1M28-
YOLOv5s_7_0 [16]25.0%58.3%40.6%26.7%0.3516.07.0M50-
YOLOv6s [19]--24.6%15.0%-185.217.2M49-
YOLOv7 [30]34.2%48.1%38.6%24.9%0.40105.137.2M18-
YOLOv8s [17]23.9%65.0%39.1%26.1%0.3528.611.1M33-
WeedNet-R [11]90.1%68.4%84.8%-0.7819.025.6M0.5-
YOLO-DP (our)23.1%73.8%38.3%25.0%0.3511.75.2M62-
", + "bbox": [ + 289, + 183, + 715, + 513 + ], + "page_idx": 11 + }, + { + "type": "text", + "text": "on PDT dataset. The attention modules are designed on Backbone's floors 3, 5, 7, and 9. Since the number of C3x and C3TR parameters is too large, we only replace them at the 9th layer in order to carry out the experiment. \"v5s\" means YOLOv5s_7.0. \"v5s__our\" stands for the use of Adaptive Large Scale Selective Kernel in the YOLOv5s_7.0. It can be seen from the results in Tab. 7 that the performance of YOLOv5s is greatly improved after Large Scale Selective Kernel is used. The mAP@.5, mAP@.5::.95 and F1 are the best metrics. This indicates that the Large Scale Selective Kernel used in this paper is suitable for dense object detection tasks.", + "bbox": [ + 212, + 551, + 787, + 690 + ], + "page_idx": 11 + }, + { + "type": "text", + "text": "7.3 Visualization Research", + "text_level": 1, + "bbox": [ + 215, + 710, + 450, + 724 + ], + "page_idx": 11 + }, + { + "type": "text", + "text": "In Fig. 5, we show the detection visualization results on the PDT dataset. We have selected the YOLOv5s and YOLOv8s models, which exhibited superior overall performance in the comparative experiment, to demonstrate and contrast with the YOLO-DP model. (a) and (e) illustrate the Ground Truth for the LL and LH versions, respectively. Observations indicate that the \"Human-in-the-loop\" data annotation approach is practical for weed and pest detection data annotation. The detection results in (b) and (c) suggest that the LL version", + "bbox": [ + 212, + 733, + 787, + 840 + ], + "page_idx": 11 + }, + { + "type": "page_number", + "text": "12", + "bbox": [ + 217, + 114, + 235, + 126 + ], + "page_idx": 11 + }, + { + "type": "header", + "text": "M. Zhou et al.", + "bbox": [ + 271, + 114, + 367, + 126 + ], + "page_idx": 11 + }, + { + "type": "table", + "img_path": "images/3d3c9ec97bdbf9a28d94cb28f7c6ee7a97598eb8334b1e7a8ab40ae226403984.jpg", + "table_caption": [ + "Table 7: Ablation experiment." + ], + "table_footnote": [], + "table_body": "
DatasetsApproachPRmAP@.5mAP@.5:95F1GflopsParameters
PDT dataset (LL)v5s_C188.2%88.5%93.9%67.1%0.8825.310.0M
v5s_C288.8%88.4%94.1%67.0%0.8817.27.3M
v5s_C2f88.6%88.5%93.8%67.1%0.8817.47.5M
v5s_C388.9%88.5%94.2%67.0%0.8916.07.0M
v5s_C3x88.7%81.5%87.4%63.2%0.8514.56.5M
v5s_C3TR88.3%89.0%94.1%67.1%0.8915.77.0M
v5s_C3Ghost88.9%88.2%94.2%66.7%0.8812.55.9M
v5s_SE88.8%88.3%94.4%66.6%0.8810.65.1M
v5s_CBAM89.8%87.5%94.4%66.5%0.8910.95.6M
v5s_GAM89.2%87.7%94.0%67.1%0.8816.47.5M
v5s_ECA89.7%87.0%94.3%66.2%0.8810.55.1M
v5s_our89.1%88.5%94.4%67.2%0.8912.26.1M
", + "bbox": [ + 331, + 183, + 674, + 303 + ], + "page_idx": 12 + }, + { + "type": "image", + "img_path": "images/ceafb2fd488380bbfcc29dd894e742119ccdd0771e8ed2c3cf8fc56452afe56a.jpg", + "image_caption": [ + "(a): Ground truth (LL)" + ], + "image_footnote": [], + "bbox": [ + 232, + 334, + 338, + 416 + ], + "page_idx": 12 + }, + { + "type": "image", + "img_path": "images/a041a01751df4760b48b9e8f4ba36bca4d1a15ec659587807ca7715750dd4908.jpg", + "image_caption": [ + "(b):YOLOv5s" + ], + "image_footnote": [], + "bbox": [ + 341, + 334, + 485, + 416 + ], + "page_idx": 12 + }, + { + "type": "image", + "img_path": "images/06fd8ca766369c50b1ce147af41b00eab3a5a3180ac70164c0b7da0b845fdaae.jpg", + "image_caption": [ + "(c):YOLOv8s" + ], + "image_footnote": [], + "bbox": [ + 486, + 335, + 656, + 416 + ], + "page_idx": 12 + }, + { + "type": "image", + "img_path": "images/739026f48194d4c770fe7097cea99b6aced284cd23277737612851d892eb8385.jpg", + "image_caption": [ + "(d):YOLO-DP" + ], + "image_footnote": [], + "bbox": [ + 663, + 335, + 771, + 416 + ], + "page_idx": 12 + }, + { + "type": "image", + "img_path": "images/3e3c91b210d3821aa0e1f7a0143f1686de5175e051793e82af5bc03f5822a233.jpg", + "image_caption": [ + "(e): Ground truth (LH)", + "Fig. 5: Visualization of PDT dataset detection." + ], + "image_footnote": [], + "bbox": [ + 232, + 428, + 500, + 565 + ], + "page_idx": 12 + }, + { + "type": "image", + "img_path": "images/e33321c0f7850da74d1c51d2d9fb75d56c7599a9367a12a18332ac327c07fa7c.jpg", + "image_caption": [ + "(f):YOLO-DP" + ], + "image_footnote": [], + "bbox": [ + 501, + 428, + 769, + 565 + ], + "page_idx": 12 + }, + { + "type": "text", + "text": "of PDT dataset is compatible with widely used detection models, confirming its utility. When compared to (d), it is evident that YOLO-DP delivers outstanding detection performance on LL, with no instances of missing, false, or duplicate detections. Examining the LH version's detection results in (f), we find no missed detections, and the critical issue of current detectors' inability to distinguish dead pine trees from the ground has been overcome. This leads us to conclude that the YOLO-DP model is well-suited for large-scale, high-precision, small-target UAV detection tasks. Moreover, the detection strategy depicted in Fig. 3 of this paper has been validated as effective.", + "bbox": [ + 212, + 638, + 787, + 775 + ], + "page_idx": 12 + }, + { + "type": "text", + "text": "In Fig. 6, we presents a comparison between YOLO-DP and the Confusion Matrix of the high-performing classification models YOLOv7 and YOLOv8s on the CWC dataset. It is evident that YOLO-DP's classification performance surpasses that of YOLOv7. Furthermore, YOLO-DP's classification capability is", + "bbox": [ + 212, + 779, + 787, + 840 + ], + "page_idx": 12 + }, + { + "type": "header", + "text": "PDT Dataset", + "bbox": [ + 638, + 114, + 730, + 126 + ], + "page_idx": 12 + }, + { + "type": "page_number", + "text": "13", + "bbox": [ + 767, + 114, + 785, + 126 + ], + "page_idx": 12 + }, + { + "type": "image", + "img_path": "images/81072e389a695241e9a2bbcfb447b15e823afacfdc8252a4a78a3ceed113b550.jpg", + "image_caption": [ + "Fig. 6: Confusion matrix. The rows represent the true class, the columns represent the predicted class, and the confidence value is [0,1]. The format of the matrix is: YOLOv7/YOLOv8s/YOLO-DP." + ], + "image_footnote": [], + "bbox": [ + 217, + 142, + 785, + 335 + ], + "page_idx": 13 + }, + { + "type": "text", + "text": "comparable to YOLOv8s, especially when its overall performance exceeds that of YOLOv8s.", + "bbox": [ + 215, + 422, + 784, + 450 + ], + "page_idx": 13 + }, + { + "type": "text", + "text": "8 Conclusion", + "text_level": 1, + "bbox": [ + 217, + 477, + 356, + 493 + ], + "page_idx": 13 + }, + { + "type": "text", + "text": "In this study, we introduce the PDT dataset to address the challenging task of UAV pest detection, with the goal of advancing research in UAV detection tasks within the realm of precision agriculture. We also present the CWC dataset for agricultural weed and crop classification, which compensates for the field's deficiency in training model classification capabilities. To tackle the scarcity of baseline model research in this domain, we propose the YOLO-DP model for dense, small-target UAV pest detection and validate its efficacy across three datasets. It is noteworthy that the PDT dataset features a dual-resolution version, and the CWC dataset boasts 11 detailed, texture-similar plant classes. Moreover, we provide a comprehensive evaluation of both datasets, objectively assessing their value. This work has its limitations: the 3D point cloud version of the PDT dataset is not discussed, and there is a size inconsistency issue within the CWC dataset. We plan to address these issues in future endeavors. Lastly, to ensure the continuity of this research, we make the datasets and associated code designed in this paper available on our website.", + "bbox": [ + 215, + 508, + 785, + 736 + ], + "page_idx": 13 + }, + { + "type": "text", + "text": "Acknowledgments", + "text_level": 1, + "bbox": [ + 217, + 761, + 390, + 777 + ], + "page_idx": 13 + }, + { + "type": "text", + "text": "This work was supported by Key R&D Program of Shandong Province, China (2023CXGC010112), the Taishan Scholars Program (NO. tsqn202103097, NO. tscy20221110).", + "bbox": [ + 215, + 794, + 784, + 839 + ], + "page_idx": 13 + }, + { + "type": "page_number", + "text": "14", + "bbox": [ + 217, + 114, + 235, + 126 + ], + "page_idx": 13 + }, + { + "type": "header", + "text": "M. Zhou et al.", + "bbox": [ + 271, + 114, + 367, + 126 + ], + "page_idx": 13 + }, + { + "type": "text", + "text": "References", + "text_level": 1, + "bbox": [ + 217, + 143, + 321, + 159 + ], + "page_idx": 14 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "1. Agarwal, R., Hariharan, S., Rao, M.N., Agarwal, A.: Weed identification using k-means clustering with color spaces features in multi-spectral images taken by UAV. In: IGARSS. pp. 7047-7050. IEEE (2021)", + "2. Bochkovskiy, A., Wang, C., Liao, H.M.: YOLOv4: Optimal speed and accuracy of object detection. CoRR abs/2004.10934 (2020)", + "3. Boursianis, A.D., Papadopoulou, M.S., Diamantoulakis, P.D., Liopa-Tsakalidi, A., Barouchas, P., Salahas, G., Karagiannidis, G.K., Wan, S., Goudos, S.K.: Internet of things (IoT) and agricultural unmanned aerial vehicles (UAVs) in smart farming: A comprehensive review. Internet Things 18, 100187 (2022)", + "4. university of burgandy: Ribworth dataset. https://universe.roboflow.com/university-of-burgandy-zowkw/ribworth (nov 2022), https://universe.roboflow.com/university-of-burgandy-zowkw/ribworth", + "5. Butte, S., Vakanski, A., Duellman, K., Wang, H., Mirkouei, A.: Potato crop stress identification in aerial images using deep learning-based object detection. CoRR abs/2106.07770 (2021)", + "6. Caras, T., Lati, R.N., Holland, D., Dubinin, V.M., Hatib, K., Shulner, I., Keiesar, O., Liddor, G., Paz-Kagan, T.: Monitoring the effects of weed management strategies on tree canopy structure and growth using uav-lidar in a young almond orchard. Comput. Electron. Agric. 216, 108467 (2024)", + "7. Carboni: weeds pytorch (2023), https://github.com/carboni123/weeds-pytorch", + "8. Chebrolu, N., Lottes, P., Schaefer, A., Winterhalter, W., Burgard, W., Stachniss, C.: Agricultural robot dataset for plant classification, localization and mapping on sugar beet fields. Int. J. Robotics Res. 36(10), 1045-1052 (2017)", + "9. Costa, L.S., Sano, E.E., Ferreira, M.E., Munhoz, C.B.R., Costa, J.V.S., Junior, L.R.A., de Mello, T.R.B., da Cunha Bustamante, M.M.: Woody plant encroachment in a seasonal tropical savanna: Lessons about classifiers and accuracy from UAV images. Remote. Sens. 15(9), 2342 (2023)", + "0. Dabhi: Crop and weed detection (2021), https://github.com/ravirajsinh45/Crop_and_weed_detection", + "1. Guo, Z., Goh, H.H., Li, X., Zhang, M., Li, Y.: WeedNet-R: a sugar beet field weed detection algorithm based on enhanced retinanet and context semantic fusion. Frontiers in Plant Science 14, 1226329 (2023)", + "2. Han, K., Wang, Y., Tian, Q., Guo, J., Xu, C., Xu, C.: GhostNet: more features from cheap operations. In: CVPR. pp. 1577-1586. Computer Vision Foundation / IEEE (2020)", + "3. Huang, J., Luo, Y., Quan, Q., Wang, B., Xue, X., Zhang, Y.: An autonomous task assignment and decision-making method for coverage path planning of multiple pesticide spraying UAVs. Comput. Electron. Agric. 212, 108128 (2023)", + "4. Innani, S., Dutande, P., Baheti, B., Talbar, S.N., Baid, U.: Fuse-PN: A novel architecture for anomaly pattern segmentation in aerial agricultural images. In: CVPR Workshops. pp. 2960-2968. Computer Vision Foundation / IEEE (2021)", + "5. Jabir, B., Falih, N.: Deep learning-based decision support system for weeds detection in wheat fields. International Journal of Electrical and Computer Engineering 12(1), 816 (2022)", + "6. Jocher, G.: YOLOv5 by Ultralytics (May 2020). https://doi.org/10.5281/ zenodo.3908559, https://github.com/ultralytics/yolov5", + "7. Jocher, G., Chaurasia, A., Qiu, J.: Ultralytics YOLO (Jan 2023), https://github.com/ultralytics/ultralytics" + ], + "bbox": [ + 225, + 176, + 784, + 839 + ], + "page_idx": 14 + }, + { + "type": "header", + "text": "PDT Dataset", + "bbox": [ + 638, + 114, + 730, + 126 + ], + "page_idx": 14 + }, + { + "type": "page_number", + "text": "15", + "bbox": [ + 767, + 116, + 785, + 126 + ], + "page_idx": 14 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "18. Lameski, P., Zdravevski, E., Trajkovik, V., Kulakov, A.: Weed detection dataset with RGB images taken under variable light conditions. In: ICT Innovations. Communications in Computer and Information Science, vol. 778, pp. 112-119. Springer (2017)", + "19. Li, C., Li, L., Jiang, H., Weng, K., Geng, Y., Li, L., Ke, Z., Li, Q., Cheng, M., Nie, W., Li, Y., Zhang, B., Liang, Y., Zhou, L., Xu, X., Chu, X., Wei, X., Wei, X.: YOLOv6: A single-stage object detection framework for industrial applications. CoRR abs/2209.02976 (2022)", + "20. Li, Y., Hou, Q., Zheng, Z., Cheng, M., Yang, J., Li, X.: Large selective kernel network for remote sensing object detection. In: ICCV. pp. 16748-16759. IEEE (2023)", + "21. Lin, T., Dólar, P., Girshick, R.B., He, K., Hariharan, B., Belongie, S.J.: Feature pyramid networks for object detection. In: CVPR. pp. 936-944. IEEE Computer Society (2017)", + "22. Lin, T., Goyal, P., Girshick, R.B., He, K., Dollar, P.: Focal loss for dense object detection. In: ICCV. pp. 2999-3007. IEEE Computer Society (2017)", + "23. Liu, S., Qi, L., Qin, H., Shi, J., Jia, J.: Path aggregation network for instance segmentation. In: CVPR. pp. 8759-8768. Computer Vision Foundation / IEEE Computer Society (2018)", + "24. Liu, W., Anguelov, D., Erhan, D., Szegedy, C., Reed, S.E., Fu, C., Berg, A.C.: SSD: single shot multibox detector. In: ECCV (1). Lecture Notes in Computer Science, vol. 9905, pp. 21-37. Springer (2016)", + "25. Monster: weed dataset (2019), https://gitee.com/Monster7/weed-database/tree/master/", + "26. Redmon, J., Farhadi, A.: YOLOv3: An incremental improvement. CoRR abs/1804.02767 (2018)", + "27. Ren, S., He, K., Girshick, R.B., Sun, J.: Faster R-CNN: towards real-time object detection with region proposal networks. IEEE Trans. Pattern Anal. Mach. Intell. 39(6), 1137-1149 (2017)", + "28. Tan, M., Pang, R., Le, Q.V.: EfficientDet: scalable and efficient object detection. In: CVPR. pp. 10778-10787. Computer Vision Foundation / IEEE (2020)", + "29. Tang, J., Wang, D., Zhang, Z., He, L., Xin, J., Xu, Y.: Weed identification based on k-means feature learning combined with convolutional neural network. Comput. Electron. Agric. 135, 63-70 (2017)", + "30. Wang, C., Bochkovskiy, A., Liao, H.M.: YOLOv7: Trainable bag-of-freebies sets new state-of-the-art for real-time object detectors. In: CVPR. pp. 7464-7475. IEEE (2023)", + "31. Zhang, R., Wang, C., Hu, X., Liu, Y., Chen, S., et al.: Weed location and recognition based on UAV imaging and deep learning. International Journal of Precision Agricultural Aviation 3(1) (2020)", + "32. Zhou, X., Wang, D., Krahenbuhl, P.: Objects as points. CoRR abs/1904.07850 (2019)" + ], + "bbox": [ + 215, + 146, + 785, + 729 + ], + "page_idx": 15 + }, + { + "type": "page_number", + "text": "16", + "bbox": [ + 217, + 114, + 235, + 126 + ], + "page_idx": 15 + }, + { + "type": "header", + "text": "M. Zhou et al.", + "bbox": [ + 271, + 114, + 367, + 126 + ], + "page_idx": 15 + } +] \ No newline at end of file diff --git a/2024/PDT Uav Target Detection Dataset for Pests and Diseases Tree/77c7722b-38da-423b-b56a-11a16b09eb35_model.json b/2024/PDT Uav Target Detection Dataset for Pests and Diseases Tree/77c7722b-38da-423b-b56a-11a16b09eb35_model.json new file mode 100644 index 0000000000000000000000000000000000000000..b2a693c46e427ffd71918874f9e5ac062109ff8d --- /dev/null +++ b/2024/PDT Uav Target Detection Dataset for Pests and Diseases Tree/77c7722b-38da-423b-b56a-11a16b09eb35_model.json @@ -0,0 +1,2058 @@ +[ + [ + { + "type": "title", + "bbox": [ + 0.24, + 0.141, + 0.764, + 0.184 + ], + "angle": 0, + "content": "PDT: Uav Target Detection Dataset for Pests and Diseases Tree" + }, + { + "type": "text", + "bbox": [ + 0.318, + 0.212, + 0.685, + 0.245 + ], + "angle": 0, + "content": "Mingle Zhou\\(^{1,2,4}\\), Rui Xing\\(^{1,2}\\), Delong Han\\(^{1,2}\\), Zhiyong Qi\\(^{3}\\), and Gang Li\\(^{1,2(\\boxtimes)}\\)" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.255, + 0.787, + 0.311 + ], + "angle": 0, + "content": "1 Key Laboratory of Computing Power Network and Information Security, Ministry of Education, Shandong Computer Science Center (National Supercomputer Center in Jinan), Qilu University of Technology (Shandong Academy of Sciences), Jinan, China lig@qlu.edu.cn" + }, + { + "type": "text", + "bbox": [ + 0.253, + 0.31, + 0.747, + 0.339 + ], + "angle": 0, + "content": "\\(^{2}\\) Shandong Provincial Key Laboratory of Computer Networks, Shandong Fundamental Research Center for Computer Science, Jinan, China" + }, + { + "type": "text", + "bbox": [ + 0.221, + 0.338, + 0.78, + 0.367 + ], + "angle": 0, + "content": "3 Institute of Remote Sensing and Geographic Information System, School of Earth and Space Sciences, Peking University, Beijing, China" + }, + { + "type": "text", + "bbox": [ + 0.229, + 0.366, + 0.771, + 0.38 + ], + "angle": 0, + "content": "4 SHANDONG SCICOM Information and Economy Research Institute Co., Ltd." + }, + { + "type": "text", + "bbox": [ + 0.261, + 0.416, + 0.744, + 0.667 + ], + "angle": 0, + "content": "Abstract. UAVs emerge as the optimal carriers for visual weed identification and integrated pest and disease management in crops. However, the absence of specialized datasets impedes the advancement of model development in this domain. To address this, we have developed the Pests and Diseases Tree dataset (PDT dataset). PDT dataset represents the first high-precision UAV-based dataset for targeted detection of tree pests and diseases, which is collected in real-world operational environments and aims to fill the gap in available datasets for this field. Moreover, by aggregating public datasets and network data, we further introduced the Common Weed and Crop dataset (CWC dataset) to address the challenge of inadequate classification capabilities of test models within datasets for this field. Finally, we propose the YOLO-Dense Pest (YOLO-DP) model for high-precision object detection of weed, pest, and disease crop images. We re-evaluate the state-of-the-art detection models with our proposed PDT dataset and CWC dataset, showing the completeness of the dataset and the effectiveness of the YOLO-DP. The proposed PDT dataset, CWC dataset, and YOLO-DP model are presented at https://github.com/RuiXing123/PDT_CWC_YOLO-DP." + }, + { + "type": "text", + "bbox": [ + 0.262, + 0.679, + 0.589, + 0.693 + ], + "angle": 0, + "content": "Keywords: Datasets \\(\\cdot\\) UAVs \\(\\cdot\\) Object detection" + }, + { + "type": "title", + "bbox": [ + 0.216, + 0.719, + 0.377, + 0.735 + ], + "angle": 0, + "content": "1 Introduction" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.75, + 0.788, + 0.841 + ], + "angle": 0, + "content": "The efficacy of intelligent UAV-based plant protection operations hinges on precisely identifying weeds and pests within imagery, representing a critical challenge in computer vision [6,9,13]. As computer vision and UAV technologies have advanced swiftly, the adoption of automated UAV plant protection operations is on the rise [3,14]. The working mode of plant protection UAV has significant defects: non-intelligent work can result in the misuse or waste of agrochemicals," + } + ], + [ + { + "type": "page_number", + "bbox": [ + 0.218, + 0.116, + 0.23, + 0.127 + ], + "angle": 0, + "content": "2" + }, + { + "type": "header", + "bbox": [ + 0.272, + 0.115, + 0.37, + 0.128 + ], + "angle": 0, + "content": "M. Zhou et al." + }, + { + "type": "image", + "bbox": [ + 0.248, + 0.145, + 0.756, + 0.321 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.214, + 0.334, + 0.788, + 0.378 + ], + "angle": 0, + "content": "Fig. 1: Dataset comparison. (a) shows the PDT dataset (Low Resolution(LL) and High Resolution(LH)): \\(640 \\times 640\\), \\(5472 \\times 3648\\). (b) shows the characteristics of the CWC dataset: it contains 11 different similar plants. (c), (d) and (e) are the public datasets." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.407, + 0.784, + 0.481 + ], + "angle": 0, + "content": "and the high-intensity, repetitive mechanical tasks can diminish their operational lifespan. Enhancing the precision of weed and pest target recognition is essential for the effectiveness of intelligent plant protection UAVs. The absence of specialized datasets and the limitations of existing detection models pose significant constraints on advancing research in this domain." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.483, + 0.785, + 0.603 + ], + "angle": 0, + "content": "Drones typically take images at both high and low altitudes when outdoors. Existing datasets are usually taken from indoor greenhouses (Fig. 1 (c)), where the majority of the data consists of target images captured at close range (Fig. 1 (d)). Such data struggles to incorporate environmental factors like changes in lighting, and the target sizes significantly deviate from real-world conditions, failing to satisfy practical requirements. Further, the majority of existing datasets are limited to single or double classes, which cannot meet the training of model classification ability, as illustrated in Fig. 1 (c), (d), and (e)." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.603, + 0.785, + 0.709 + ], + "angle": 0, + "content": "On the other hand, the conventional detection models utilized in UAV-based intelligent plant protection could be more challenging in accurately identifying targets. The deficiency in this field is underscored by the absence of tailored baseline models, resulting in unmet demands for detecting and managing diverse weeds, crops, pests, and diseases. This paper aims to address the lack of dedicated datasets and algorithms for weeds, pests and diseases in agricultural drones. Overall, the main contributions of this paper are as follows:" + }, + { + "type": "text", + "bbox": [ + 0.226, + 0.719, + 0.784, + 0.748 + ], + "angle": 0, + "content": "- We have developed the UAV target detection agricultural Pests and Diseases Tree dataset, known as PDT dataset." + }, + { + "type": "text", + "bbox": [ + 0.225, + 0.75, + 0.784, + 0.779 + ], + "angle": 0, + "content": "- We collated public data and provided Common Weed and Crop dataset (CWC dataset) to train the classification capabilities of the model." + }, + { + "type": "text", + "bbox": [ + 0.225, + 0.78, + 0.784, + 0.809 + ], + "angle": 0, + "content": "- We have designed a specialized detection model, YOLO-DP, tailored for the dense object detection of tree crop pests and diseases." + }, + { + "type": "text", + "bbox": [ + 0.225, + 0.811, + 0.784, + 0.839 + ], + "angle": 0, + "content": "- We reevaluated and analyzed existing generic and specific detection models on PDT and CWC datasets to establish benchmarks for the field." + }, + { + "type": "list", + "bbox": [ + 0.225, + 0.719, + 0.784, + 0.839 + ], + "angle": 0, + "content": null + } + ], + [ + { + "type": "header", + "bbox": [ + 0.64, + 0.115, + 0.732, + 0.127 + ], + "angle": 0, + "content": "PDT Dataset" + }, + { + "type": "page_number", + "bbox": [ + 0.775, + 0.117, + 0.785, + 0.127 + ], + "angle": 0, + "content": "3" + }, + { + "type": "title", + "bbox": [ + 0.217, + 0.145, + 0.388, + 0.161 + ], + "angle": 0, + "content": "2 Related Work" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.176, + 0.788, + 0.446 + ], + "angle": 0, + "content": "Agricultural UAV Detection Dataset. In the realm of agricultural robotics, acquiring pertinent datasets frequently presents a challenge, primarily due to the necessity of maintaining dedicated experimental fields and the critical timing of data collection, which is pivotal to the research outcomes. Chebrolu et al. presented SugarBeet2016 large agricultural robot dataset and subsequently introduced the SugarBeet2017 dataset tailored for the detection task [8,11]. The limitation of this dataset is that the data, captured from the vantage point of a ground robot, significantly diverges from the actual detection scenarios encountered by UAVs. Vakanski et al. introduced a dataset of multispectral potato plant images taken by drones [5]. The limitation of this dataset is the lack of raw data, consisting of only 360 images and a total of 1,500 images produced by data enhancement, which cannot effectively guide the model to learn the details of the crop. The University of Burgundy has launched the Ribworth Dataset, designed for crops and weeds, which uses data augmentation to simulate real-world environmental conditions [4]. The limitation of this dataset is that its acquisition target is relatively close, and the factors such as light change are not included, which is not enough to meet the requirements of accurate detection tasks of agricultural plant protection UAVs." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.448, + 0.788, + 0.493 + ], + "angle": 0, + "content": "To foster the advancement of precision agriculture, we introduce the PDT dataset and CWC dataset, aimed at addressing the requirements for detecting agricultural weeds, crops, and pests." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.493, + 0.788, + 0.735 + ], + "angle": 0, + "content": "Agricultural UAV Detection Model. The agricultural UAV target detection task has garnered significant attention within the realms of machine learning and computer vision, owing to its attributes of precise localization, efficient operation, and minimal environmental impact. Among them, the weed detection models based on YOLO deep learning model include: Zhang et al. developed a detection model for weed location and identification in wheat fields by using YOLOv3-tiny [31]. Gao et al. modified the YOLOv3-tiny model and applied it to weed detection in sugar beet fields. Jabir and Falih proposed an intelligent wheat field weed detection system based on YOLOv5 [15]. Furthermore, a variety of neural network architectures have also been recognized by researchers and implemented in the field of precision agriculture. Guo et al. proposed WeedNet-R based on the RetinaNet architecture for weed identification and location in sugar beet fields [11]. Tang et al. employed Kmeans feature learning in conjunction with a convolutional neural network for the recognition of weeds [29]. Agarwal et al. utilized the Kmeans clustering algorithm to analyze color space features in multispectral weed images for effective weed detection [1]." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.735, + 0.788, + 0.84 + ], + "angle": 0, + "content": "It is evident that there is a scarcity of studies focusing on baseline models for the detection of tree diseases and pests, particularly those related to the invasion of the exotic species Red Turpentine Beetle. Furthermore, the majority of these baseline models are trained using single-class or two-class detection datasets, and the classification ability of the models is limited by the diversity of the data. However, for the detection tasks in large-scale, high-precision agricultural UAV operations, precise detection, accurate classification, and the expansion of" + } + ], + [ + { + "type": "page_number", + "bbox": [ + 0.218, + 0.116, + 0.23, + 0.127 + ], + "angle": 0, + "content": "4" + }, + { + "type": "header", + "bbox": [ + 0.272, + 0.115, + 0.369, + 0.127 + ], + "angle": 0, + "content": "M. Zhou et al." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.147, + 0.784, + 0.178 + ], + "angle": 0, + "content": "crop detection varieties are essential. Consequently, we propose the YOLO-DP detection model to address these requirements." + }, + { + "type": "title", + "bbox": [ + 0.216, + 0.201, + 0.383, + 0.217 + ], + "angle": 0, + "content": "3 PDT Dataset" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.233, + 0.788, + 0.4 + ], + "angle": 0, + "content": "In this study, we introduced a PDT dataset with both high and low resolution versions, specifically for the detection of red turpentine beetle natural pest targets. As most existing public datasets are gathered indoors or via simulated drones, they fail to capture the characteristics of outdoor UAV detection environments. Consequently, there is an immediate demand for an agricultural dataset that is tailored to real-world operational environments and encompasses the distribution of both high and low-altitude data, thereby fulfilling the high-precision, large-scale target detection requirements from a UAV perspective. This section delineates the data acquisition process for the PDT dataset, encompassing the selection of the data collection domain, the equipment used for acquisition, and the criteria for sample definition." + }, + { + "type": "title", + "bbox": [ + 0.216, + 0.422, + 0.403, + 0.438 + ], + "angle": 0, + "content": "3.1 Data Acquisition" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.448, + 0.788, + 0.599 + ], + "angle": 0, + "content": "Selection of the Data Collection Domain. The Red Turpentine Beetle, a member of the Coleopteridae family, not only invades extensive pine forest areas but also contributes to the widespread mortality of trees, exerting a profound impact on the regional ecological environment and forestry. Despite the implementation of preventive and control strategies, including monitoring and trapping, effectively managing its continued proliferation remains a formidable challenge. The precision spraying platform of plant protection UAV is an effective solution, but the public UAV data for the invasive species Red Turpentine Beetle is blank. Consequently, we opted to collect data from a cluster of dead pine trees that have been infested by the Red Turpentine Beetle." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.599, + 0.788, + 0.765 + ], + "angle": 0, + "content": "High-resolution Drone Camera Equipment. To obtain accurate sample data, we use a UAV mapping camera called DJI-ChanSi L2. This advanced system enhances the UAV flight platform's capabilities for acquiring accurate, efficient, and reliable 2D and 3D data. The LiDAR components have excellent performance, ranging accuracy of \\(2\\mathrm{cm}\\) at a height of 150 meters, laser wavelength of \\(905~\\mathrm{nm}\\), and laser pulse emission frequency of \\(240\\mathrm{kHz}\\). The system supports the exFAT file system and can capture images in JPEG or PNG formats. In essence, the DJI-ChanSi L2 is a comprehensive platform that integrates lidar, a high-precision navigation system, and a mapping camera, specifically designed for high-precision data acquisition in mapping, forestry, power grid, and engineering and construction industries." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.765, + 0.788, + 0.842 + ], + "angle": 0, + "content": "Definition of Detection Target. Distinct from close-range pest detection, the task of pest detection at low or high altitudes fundamentally involves capturing the differential characteristics between the affected vector and the healthy state. Consequently, the detection samples for the PDT dataset are sourced from both healthy and unhealthy trees within pine forests. As depicted in Fig. 2, (a)" + } + ], + [ + { + "type": "header", + "bbox": [ + 0.64, + 0.115, + 0.732, + 0.127 + ], + "angle": 0, + "content": "PDT Dataset" + }, + { + "type": "page_number", + "bbox": [ + 0.775, + 0.117, + 0.785, + 0.127 + ], + "angle": 0, + "content": "5" + }, + { + "type": "image", + "bbox": [ + 0.289, + 0.148, + 0.384, + 0.21 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.308, + 0.214, + 0.362, + 0.223 + ], + "angle": 0, + "content": "(a): healthy" + }, + { + "type": "image", + "bbox": [ + 0.403, + 0.148, + 0.488, + 0.209 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.511, + 0.148, + 0.589, + 0.209 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.515, + 0.212, + 0.582, + 0.222 + ], + "angle": 0, + "content": "(b): unhealthy" + }, + { + "type": "image", + "bbox": [ + 0.61, + 0.148, + 0.71, + 0.209 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.215, + 0.24, + 0.785, + 0.268 + ], + "angle": 0, + "content": "Fig. 2: Data example. (a) is a healthy goal and (b) is a unhealthy goal. The PDT dataset takes (b) as the category." + }, + { + "type": "image", + "bbox": [ + 0.246, + 0.284, + 0.756, + 0.407 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.214, + 0.423, + 0.785, + 0.478 + ], + "angle": 0, + "content": "Fig. 3: PDT dataset generation and detection process. (a) represents the sliding window method, and (b) represents the \"Human-in-the-loop\" data annotation method. (c) means that a LL image is sent to the neural network for training, and LL and LH dual-resolution images are detected at the same time." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.508, + 0.785, + 0.553 + ], + "angle": 0, + "content": "illustrates a pine tree in a healthy state, while (b) displays three pine trees in a unhealthy state that have been infested by pests, exhibiting progressively severe symptoms from left to right." + }, + { + "type": "title", + "bbox": [ + 0.215, + 0.575, + 0.54, + 0.59 + ], + "angle": 0, + "content": "3.2 Data Preprocessing and Statistics" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.599, + 0.787, + 0.764 + ], + "angle": 0, + "content": "This section delves into the acquisition, processing, and production procedures of the original image data about the PDT dataset. We set up an isometric photo mode at an altitude of 150 meters to acquire 105 sets of raw image data and 3D point cloud data. Owing to the substantial size of the original images, they are unsuitable for training with the current mainstream object detection neural networks. We introduce a training approach that applies to the majority of neural network data. As illustrated in Fig. 3, we employ a methodology that encompasses \"data preprocessing - 'Human-in-the-loop' data labeling - manual correction\". The essence of this method is to leverage the preliminary analysis capabilities of artificial intelligence to aid human annotators, thereby enhancing the efficiency and accuracy of data annotation." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.765, + 0.787, + 0.84 + ], + "angle": 0, + "content": "Data Preprocessing. Fig. 3 (a) demonstrates the application of the sliding window technique on the original image data, which allows for the extraction of a usable, standard-sized image through regional cropping of the high-resolution image. To prevent information loss during the sliding window process, the window dimensions are set to \\(640 \\times 640\\) pixels, with a step size of \\(630 \\times 630\\) pixels." + } + ], + [ + { + "type": "page_number", + "bbox": [ + 0.218, + 0.116, + 0.23, + 0.127 + ], + "angle": 0, + "content": "6" + }, + { + "type": "header", + "bbox": [ + 0.272, + 0.115, + 0.37, + 0.128 + ], + "angle": 0, + "content": "M. Zhou et al." + }, + { + "type": "table_caption", + "bbox": [ + 0.329, + 0.145, + 0.674, + 0.16 + ], + "angle": 0, + "content": "Table 1: Structure of PDT dataset and its sample." + }, + { + "type": "table", + "bbox": [ + 0.276, + 0.184, + 0.734, + 0.273 + ], + "angle": 0, + "content": "
EditionClassesStructureTargeted imagesUntargeted imagesImage sizeInstancesTarget Amount
S(Small)M(Medium)L(Large)
SampleunhealthyTrain811640×64025691896548179
Val1969152813825
LLunhealthyTrain316613709029070418163423530
Val395172640×6401252399262165432
Test3901771149489492095450
LHunhealthy-10505472×3648934749347400
" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.313, + 0.787, + 0.419 + ], + "angle": 0, + "content": "Following the sliding window operation, 53 LL images were derived from a single LH image. Furthermore, due to the high similarity between the unhealthy pine tree depicted in Fig. 2 and the ground background, and to enhance the effectiveness of neural network training, we opt to retain images that do not contain targets. Upon generating an output from these data during training, a corresponding loss is produced, which aids the neural network in learning and reduces the rate of false positives." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.42, + 0.788, + 0.617 + ], + "angle": 0, + "content": "Data Annotation. We proceed to annotate the preprocessed data. To ensure the data's validity, we utilize the widely-used Labelme \\( {}^{1} \\) data labeling software. For high-resolution datasets with dense, small targets, the sheer volume of targets per image makes manual annotation a tedious and time-intensive process. As illustrated in Fig. 3 (b), the \"Human-in-the-loop\" data annotation approach effectively addresses this challenge. Initially, we acquire the LL version samples of the PDT dataset through manual annotation. The structure of the sample dataset is depicted in Tab. 1, which comprises a training set and a validation set. Subsequently, the sample dataset is input into the YOLOv5s model for training over 300 epochs, with the optimal weights being saved. To ensure that no label data is missed, a confidence threshold of 0.25 and an NMS IOU (Intersection over Union) threshold of 0.45 were set to automatically label the LL data and obtain the original label set." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.617, + 0.788, + 0.737 + ], + "angle": 0, + "content": "Manual Check. The labeled dataset underwent manual filtering to eliminate incorrect annotations, rectify erroneous ones, and add any omitted annotations. The LL datasets were generated in two widely recognized formats, YOLO txt and VOCXml, by automatically partitioning the data structure into an 8:1:1 ratio using scripts. For the LH data, we input the completed LL data into the YOLO-DP model for training to acquire the optimal weights. The \"Human-in-the-loop\" data annotation and subsequent manual verification were conducted on the LH data to obtain the final LH version dataset." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.738, + 0.788, + 0.815 + ], + "angle": 0, + "content": "Data Statistics. The statistics of the PDT dataset are shown in Tab. 1. LL and LH versions consist of 5,670 and 105 images, respectively, with 114,307 and 93,474 samples, and feature a single class labeled as unhealthy. The training, validation, and test sets of the LL version include 1,370, 172, and 144 target-free images, respectively. On average, the LL and LH versions contain 29 and 890" + }, + { + "type": "page_footnote", + "bbox": [ + 0.204, + 0.825, + 0.478, + 0.84 + ], + "angle": 0, + "content": "1 https://github.com labelmeai/labelme" + } + ], + [ + { + "type": "header", + "bbox": [ + 0.64, + 0.115, + 0.732, + 0.127 + ], + "angle": 0, + "content": "PDT Dataset" + }, + { + "type": "page_number", + "bbox": [ + 0.775, + 0.116, + 0.785, + 0.126 + ], + "angle": 0, + "content": "7" + }, + { + "type": "table_caption", + "bbox": [ + 0.393, + 0.145, + 0.609, + 0.158 + ], + "angle": 0, + "content": "Table 2: CWC dataset sources." + }, + { + "type": "table", + "bbox": [ + 0.277, + 0.184, + 0.732, + 0.264 + ], + "angle": 0, + "content": "
DatasetsCorn weed datasets [25]lettuce weed radish weed datasets [25] datasets [25]Fresh-weed-data [18]
Classesbluegrass, corn, sedge, chenopodium album, circium_setosumlettuceradishnightshadetomatocottonvelvet
Number2502002011151162438
Image Size800×600800×600800×600800×600800×600586×444643×500
" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.305, + 0.784, + 0.337 + ], + "angle": 0, + "content": "samples per image, respectively, indicating that the PDT dataset is suitable for the task of high-density and high-precision inspection of plant protection drones." + }, + { + "type": "title", + "bbox": [ + 0.216, + 0.36, + 0.389, + 0.377 + ], + "angle": 0, + "content": "4 CWC Dataset" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.393, + 0.788, + 0.56 + ], + "angle": 0, + "content": "We compile the available data to construct the Common Weed and Crop datasets. In special cases, there may be crop crossovers (usually in small trial plots) where weeds and crop targets exhibit multi-class structural characteristics and similarities. The majority of existing public datasets are limited to single or double classes, which do not suffice for the training and classification capabilities required by models. Consequently, when training the neural network, it is essential to supply detailed texture data that differentiates between various objects, thereby equipping the model with robust classification capabilities. This enables the model to effectively sift through and eliminate highly similar negative samples during detection, thereby enhancing the model's detection accuracy. This section discusses the acquisition and processing of CWC original image data." + }, + { + "type": "title", + "bbox": [ + 0.216, + 0.583, + 0.373, + 0.597 + ], + "angle": 0, + "content": "4.1 Data Sources" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.609, + 0.788, + 0.701 + ], + "angle": 0, + "content": "In Tab. 2, we present the source, categories, initial quantities of the CWC raw data, as well as their respective resolutions. As depicted in Fig. 1 (b), the CWC original data comprises a total of 11 categories, characterized by similar size, texture, and color. Please note that we manually annotated the raw Corn weed datasets, Lettuce weed datasets, and Radish weed datasets, while the Fresh-weed-data was already pre-labeled and available in YOLO txt format." + }, + { + "type": "title", + "bbox": [ + 0.216, + 0.723, + 0.542, + 0.739 + ], + "angle": 0, + "content": "4.2 Data Preprocessing and Statistics" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.75, + 0.788, + 0.841 + ], + "angle": 0, + "content": "Data Preprocessing. Owing to the imbalance between positive and negative samples in the original CWC data, there is a risk that the model may become biased towards the majority class, leading to a diminished recognition capability for the minority class. We used the oversampling strategy, using the data augmentation method to equalize the number of samples. The methods include random rotation, random translation, random brightness change, random noise" + } + ], + [ + { + "type": "page_number", + "bbox": [ + 0.218, + 0.116, + 0.23, + 0.127 + ], + "angle": 0, + "content": "8" + }, + { + "type": "header", + "bbox": [ + 0.272, + 0.115, + 0.37, + 0.128 + ], + "angle": 0, + "content": "M. Zhou et al." + }, + { + "type": "table_caption", + "bbox": [ + 0.386, + 0.145, + 0.616, + 0.159 + ], + "angle": 0, + "content": "Table 3: CWC dataset structure." + }, + { + "type": "table", + "bbox": [ + 0.248, + 0.184, + 0.761, + 0.281 + ], + "angle": 0, + "content": "
Classesbluegrasschenopodium-album-cirsium-setosumcornsedgelettuceradishnightshadetomatocottonvelvet
Targeted ImagesTrain200200200200200200200200200200200200
Val404040404040404040404040
Test101010101010101010101010
Targeted AmountS100500000003
M000900000000
L249250250236250444326250210268248
Image Size800×600800×600800×600800×600800×600800×600800×600800×600800×600586×444643×500
" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.321, + 0.784, + 0.366 + ], + "angle": 0, + "content": "addition, flipping and cutout. This enhances the robustness of the dataset, simulating the illumination formula: \\( \\pmb{I}_{out} = \\pmb{I}_{in} \\times w + [255] \\times (1 - w) \\), \\( w \\in [0.35, 1] \\). Where \\( \\pmb{I}_{out} \\), \\( \\pmb{I}_{in} \\) is the output and input images. \\( w \\) is the random weight." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.367, + 0.785, + 0.429 + ], + "angle": 0, + "content": "Data Statistics. The statistics for the CWC dataset are presented in Tab. 3. The CWC dataset comprises 2,750 images and 2,599 data entries, with the majority featuring large-sized objects. Data enhancement makes the sample reach a balanced distribution, with an average image containing 1 to 2 objects." + }, + { + "type": "title", + "bbox": [ + 0.216, + 0.451, + 0.453, + 0.468 + ], + "angle": 0, + "content": "5 Dataset Comparison" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.482, + 0.788, + 0.619 + ], + "angle": 0, + "content": "Tab. 4 provides a thorough comparison of the datasets. The LH version of the PDT dataset is exceptionally clear, offering image quality that is 35 to 200 times superior to other publicly available datasets. The PDT dataset is characterized by density, small target and rich real environmental details, which accords with the typical environmental factors of high and low altitude UAV work, and is suitable for training the special UAV detection model for plant protection. Furthermore, the PDT dataset includes a 3D point cloud version. While not highlighted in this study, a 3D data version is planned for development and will be released to the public in the future." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.62, + 0.788, + 0.71 + ], + "angle": 0, + "content": "On the other hand, the CWC dataset demonstrates outstanding classification performance, particularly in terms of category diversity, detailed texture information, and dataset authenticity. CWC dataset has a large number of categories, 2 to 5 times more than other publicly available datasets, which indicates that the CWC dataset is well suited for tasks involving high-precision classification of plant protection UAVs." + }, + { + "type": "title", + "bbox": [ + 0.216, + 0.733, + 0.427, + 0.749 + ], + "angle": 0, + "content": "6 YOLO-DP Model" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.765, + 0.788, + 0.842 + ], + "angle": 0, + "content": "To expedite research efforts in UAV detection models for intelligent agricultural plant protection, we have designed and proposed a model, YOLO-DP (Fig. 4), specifically dedicated to the detection of tree crop diseases and pests, building upon the YOLOv5s foundatin [16]. To ensure the scalability of the model, the Kmeans clustering algorithm is employed to determine the distribution of the" + } + ], + [ + { + "type": "header", + "bbox": [ + 0.64, + 0.115, + 0.732, + 0.127 + ], + "angle": 0, + "content": "PDT Dataset" + }, + { + "type": "page_number", + "bbox": [ + 0.775, + 0.116, + 0.787, + 0.127 + ], + "angle": 0, + "content": "9" + }, + { + "type": "table_caption", + "bbox": [ + 0.399, + 0.145, + 0.604, + 0.16 + ], + "angle": 0, + "content": "Table 4: Dataset comparison." + }, + { + "type": "table", + "bbox": [ + 0.248, + 0.184, + 0.761, + 0.403 + ], + "angle": 0, + "content": "
DatasetResolutionClassesRich DetailsHigh DefinitionScaleDense TargetTarget ScaleUav CollectionNo Target Image3D Point Cloud DataAnnotation Quality
PDT dataset5472×36481-✓100+✓S✓✓✓“Human-in-the-loop”
640×6401--5K+✓S/M/L✓✓✓
CWC dataset800×60011✓-2K+-L---Manually
SugarBeet2017 [11]1296×9662--5K+-S/M--✓Manually
Multispectral Potato Plants Images (MPP) [5]750×7502--1K+✓L✓--Manually
Ribworth Dataset (RI) [4]320×2401✓-5K+-M/L---Manually
320×320
Corn weed Dataset [25]800×6005✓-6K+-L---No
lettuce weed Dataset [25]800×6002✓-700+-L---No
radish weed Dataset [25]800×6002✓-400+-L---No
Fresh -weed-data [29]800×6004✓-200+-L---Manually
586×444
643×500
crop and weed detection data (CAW) [10]512×5122✓-1K+-L---Manually
Weeds [7]480×4801--600+-S/M✓--Manually
" + }, + { + "type": "image", + "bbox": [ + 0.248, + 0.433, + 0.758, + 0.536 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.214, + 0.548, + 0.788, + 0.593 + ], + "angle": 0, + "content": "Fig. 4: YOLO-DP baseline model architecture. The FPN [21] + PAN [23] module consists of GhostConv [12], Upsample, Concat, and C3. C stands for Concat, S for Sigmoid, P for channel number dilation, \\(\\times\\) for matrix multiplication, and + for matrix addition." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.626, + 0.788, + 0.808 + ], + "angle": 0, + "content": "actual bounding box data, allowing for the dynamic adjustment of the anchor box sizes. In light of the UAV's large-scale detection capabilities, an adaptive Large Scale Selective Kernel is incorporated into the Backbone network to capture the location information of dense, small target pest-infested trees [20]. Taking into account the efficient detection mode of UAVs, GhostConv is utilized in the Neck network to reduce the model size and computational complexity. Concurrently, the receptive field is enlarged to capture more texture feature information, thereby enhancing the model's classification capabilities. A version with decoupled detection heads is provided to minimize the interference between classification and regression losses, allowing the model to focus on specific tasks, improving the accuracy and accuracy of the detection, and also improving the model's generalization ability." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.81, + 0.788, + 0.842 + ], + "angle": 0, + "content": "Adaptive Large Scale Selective Kernel. To tackle the demanding detection task involving a broad range of dense, small targets, we gather pest and" + } + ], + [ + { + "type": "page_number", + "bbox": [ + 0.218, + 0.116, + 0.236, + 0.127 + ], + "angle": 0, + "content": "10" + }, + { + "type": "header", + "bbox": [ + 0.272, + 0.115, + 0.369, + 0.127 + ], + "angle": 0, + "content": "M. Zhou et al." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.147, + 0.788, + 0.435 + ], + "angle": 0, + "content": "disease target information by designing a Convolutional group with an extensive receptive field. Firstly, the matrix M obtains the shallow range information M1 by Conv with \\( \\text{Kernelsize}(k) = 5 \\) and \\( \\text{Padding}(p) = 2 \\). The depth information is explored for M1, and the deep range information M2 is obtained using Conv with \\( k = 7 \\) and \\( p = 9 \\). Feature selection was performed on shallow and depth range information, and M3 and M4 were obtained by grouping convolution on M1 and M2 using GhostConv with \\( k = 1 \\). The advantage of using GhostConv is that half of the range information is retained and the other half is processed to prevent excessive loss of information. Next, the spatial attention is calculated for M3 and M4 large-scale information. M3 and M4 are concatenated in the channel dimension to obtain M5, and M6 and M7 are obtained by averaging and maximizing the channel dimensions, respectively. M6 and M7 are concatenated in the channel dimension to obtain M8 with scale \\( b \\times 1 \\times w \\times h \\), which is fed into Conv with \\( k = 7 \\) and \\( p = 4 \\) for range attention collection to obtain M9. After Sigmoid activation, the channel dimension is expanded to obtain M10 and M11. Matrix multiplication is performed with M3 and M4 to obtain M12 and M13, respectively. After bitwise addition, the final spatial attention matrix M14 is obtained after Conv (\\( k = 1 \\)). Finally, the output is obtained by matrix multiplication with the input matrix." + }, + { + "type": "title", + "bbox": [ + 0.216, + 0.456, + 0.367, + 0.473 + ], + "angle": 0, + "content": "7 Experiment" + }, + { + "type": "title", + "bbox": [ + 0.216, + 0.486, + 0.469, + 0.501 + ], + "angle": 0, + "content": "7.1 Experimental Conditions" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.509, + 0.787, + 0.614 + ], + "angle": 0, + "content": "The experimental environment of this paper is ubuntu 18.04, Python3.9, PyTorch 1.9.1 framework, CUDA 11.4 version, and cuDNN 8.0 version. This paper's model training and inference are performed on NVIDIA A100-SXM4 with 40GB GPU memory and 16 GB CPU memory on the experimental platform. The training metrics in this paper are: 300 epochs of iterative training, the early stopping mechanism is set to 100 rounds, the initial learning rate is 0.01, the weight decay of 0.001 and momentum of \\(0.9\\%\\) are used, the Batch-size is 16." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.615, + 0.787, + 0.766 + ], + "angle": 0, + "content": "Experiments were carried out on the PDT dataset and CWC dataset presented in this paper, as well as on the public dataset SugarBeet2017. We test the performance of the model YOLO-DP according to the characteristics of the dataset. First, the PDT dataset is used to test the detection ability of the model for dense small targets in the UAV perspective. Secondly, the ability of the model to extract fine-grained texture information is measured on the CWC dataset to verify its classification ability. It is reflected in the P metrics. Finally, the detection accuracy was measured on the SugarBeet2017 dataset. It is reflected in the R metrics. In order not to lose the image information, the input sizes are \\(640 \\times 640\\), \\(800 \\times 800\\), \\(1296 \\times 1296\\) respectively." + }, + { + "type": "title", + "bbox": [ + 0.216, + 0.787, + 0.45, + 0.803 + ], + "angle": 0, + "content": "7.2 Experimental Analysis" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.81, + 0.787, + 0.84 + ], + "angle": 0, + "content": "Dataset Validation. We evaluated 6 detection models on seven datasets (2 of ours and 5 public datasets mentioned in the paper). Based on the dataset's" + } + ], + [ + { + "type": "header", + "bbox": [ + 0.64, + 0.115, + 0.732, + 0.127 + ], + "angle": 0, + "content": "PDT Dataset" + }, + { + "type": "page_number", + "bbox": [ + 0.769, + 0.116, + 0.785, + 0.127 + ], + "angle": 0, + "content": "11" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.147, + 0.788, + 0.269 + ], + "angle": 0, + "content": "characteristics, we choose different metrics for the ranking model (Rank (Metrics)). Specifically, dense target datasets collected in natural environments use the comprehensive F1 metrics, datasets with target size changes use the R metrics, and multiclass datasets use the P metrics. We sort models with the same metrics score again using Gflops. Our model has excellent performance on all 7 datasets, and the performance of the proposed 2 datasets differs from that of existing datasets, indicating that our dataset can provide unique data distribution to the research field (Tab. 5)." + }, + { + "type": "table_caption", + "bbox": [ + 0.388, + 0.297, + 0.616, + 0.311 + ], + "angle": 0, + "content": "Table 5: Datasets effect ranking." + }, + { + "type": "table", + "bbox": [ + 0.306, + 0.323, + 0.703, + 0.387 + ], + "angle": 0, + "content": "
MethodGflopsRank (F1)Rank (P)Rank (R)
PDT (LL) (our)MPP [5]CWC (our)Weeds [7]SugarBeet2017 [11]CAW [10]RI [4]
YOLO-DP11.71 (0.89)1 (0.42)2 (92.9%)2 (76.5%)1 (73.8%)2 (89.4%)1 (82.3%)
YOLOv3155.35 (0.88)5 (0.38)6 (86.6%)1 (83.0%)6 (46.2%)6 (73.1%)6 (74.0%)
YOLOv4s20.83 (0.88)2 (0.42)5 (87.3%)3 (75.6%)3 (60.3%)3 (86.5%)2 (81.7%)
YOLOv5s16.02 (0.89)3 (0.38)4 (88.6%)4 (75.3%)4 (58.3%)5 (82.5%)3 (81.5%)
YOLOv7105.16 (0.85)6 (0.24)1 (93.1%)5 (74.4%)5 (48.1%)4 (83.3%)4 (80.3%)
YOLOv8s28.64 (0.88)4 (0.38)3 (92.0%)6 (70.4%)2 (65.0%)1 (90.1%)5 (79.3%)
" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.416, + 0.788, + 0.506 + ], + "angle": 0, + "content": "Comparative Experiment. We perform a comparative analysis of single-stage detectors including the full YOLO family, SSD, EfficientDet, and RetinaNet, the two-stage detector Fast-RCNN, anchor-free based CenterNet, and WeedNet-R, a specialized model for weed detection. Among them, some models use pre-trained weights. As some models do not provide the computation results for certain metrics, these are denoted by “-” in the table." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.507, + 0.788, + 0.749 + ], + "angle": 0, + "content": "In Tab. 6, we present the comparative evaluations of the proposed YOLO-DP baseline model alongside other models across the three datasets. The YOLO-DP model does not need to load pre-trained weights, and the detection speed is significantly ahead of other models. Under this premise, YOLO-DP demonstrates the most outstanding overall performance on the PDT dataset, with its F1 composite metrics surpassing all other models. This shows that the YOLO-DP model can well adapt to the detection scenario of UAV with large-scale, high-precision, dense and small targets. On the CWC dataset, its P metrics is at the first-class level, which is \\(4.3\\%\\) higher than the YOLOv5s baseline model. This demonstrates that the YOLO-DP model possesses an effective capability for extracting fine-grained texture information and is capable of fulfilling the classification task requirements within this domain. On the authoritative dataset SugarBeet2017, while the overall performance does not match that of models with loaded and trained weights, the R metrics of YOLO-DP is comparable to the pre-trained model's performance. This substantiates that the YOLO-DP model possesses precise detection capabilities and can accurately pinpoint the detection targets." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.749, + 0.788, + 0.794 + ], + "angle": 0, + "content": "The comparative experimental results show that the YOLO-DP model is suitable for the application of plant protection UAV in precision agriculture detection fields such as weeds, pests and diseases, and crops." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.795, + 0.788, + 0.84 + ], + "angle": 0, + "content": "Ablation Experiment. We chose the YOLOv5s that performed well on the PDT dataset as the benchmark. We use the popular attention mechanism to ablate the adaptive large-scale selection kernel to verify its detection performance" + } + ], + [ + { + "type": "page_number", + "bbox": [ + 0.218, + 0.116, + 0.236, + 0.127 + ], + "angle": 0, + "content": "12" + }, + { + "type": "header", + "bbox": [ + 0.272, + 0.115, + 0.369, + 0.127 + ], + "angle": 0, + "content": "M. Zhou et al." + }, + { + "type": "table_caption", + "bbox": [ + 0.383, + 0.145, + 0.62, + 0.16 + ], + "angle": 0, + "content": "Table 6: Comparative experiment." + }, + { + "type": "table", + "bbox": [ + 0.29, + 0.184, + 0.716, + 0.514 + ], + "angle": 0, + "content": "
DatasetsApproachPRmAP@.5mAP@.5::.95F1GflopsParametersFPS Pre-training
PDT dataset (LL)SSD [24]84.5%87.7%85.1%-0.86273.623.7M37✓
EfficientDet [28]92.6%73.4%72.3%-0.8211.56.7M12✓
RetinaNet [22]93.3%65.3%64.2%-0.79109.732.7M32✓
CenterNet [32]95.2%67.4%66.5%-0.79109.732.7M32✓
Faster-RCNN [27]57.8%70.5%61.7%-0.64401.7136.7M13✓
YOLOv3 [26]88.5%88.1%93.4%65.7%0.88155.361.5M41-
YOLOv4s [2]88.8%88.2%94.7%66.1%0.8820.89.1M51-
YOLOv5s_7_0 [16]88.9%88.5%94.2%67.0%0.8916.07.0M93-
YOLOv6s [19]--91.4%63.2%-44.117.2M43-
YOLOv7 [30]87.4%82.6%90.1%55.5%0.85105.137.2M32-
YOLOv8s [17]88.7%87.5%94.0%67.9%0.8828.611.1M60-
WeedNet-R [11]87.7%48.1%70.4%-0.6219.025.6M0.5-
YOLO-DP (our)90.2%88.0%94.5%67.5%0.8911.75.2M109-
CWC datasetSSD [24]97.7%77.6%85.7%-0.91426.923.7M29✓
EfficientDet [28]97.2%98.6%98.6%-0.9011.56.7M13✓
RetinaNet [22]95.1%98.3%98.0%-0.97261.336.4M24✓
CenterNet [32]96.6%73.8%73.3%-0.80171.432.7M27✓
YOLOv3 [26]86.8%89.4%93.2%82.3%0.88154.761.5M30-
YOLOv4s [2]87.3%87.9%91.9%81.5%0.8820.89.1M43-
YOLOv5s_7_0 [16]88.6%88.7%93.0%81.2%0.8916.07.0M65-
YOLOv6s [19]--92.7%84.3%-68.917.2M31-
YOLOv7 [30]93.1%76.4%88.1%75.6%0.84105.137.2M21-
YOLOv8s [17]92.0%89.1%94.0%86.2%0.9128.611.1M38-
WeedNet-R [11]86.1%51.8%71.6%-0.6519.025.6M0.5-
YOLO-DP (our)92.9%87.5%91.8%81.0%0.9011.75.2M72-
Sugar-Beet2017SSD [24]85.0%83.6%79.3%-0.85112023.7M19✓
EfficientDet [28]93.3%79.8%77.8%-0.8611.56.7M16✓
RetinaNet [22]91.7%78.8%76.6%-0.84256.436.3M23✓
CenterNet [32]97.9%51.2%51.0%-0.62117.432.7M41✓
Faster-RCNN [27]63.6%87.4%80.0%-0.73546.9136.7M25✓
YOLOv3 [26]34.8%46.2%39.4%25.6%0.40155.361.5M28-
YOLOv4s [2]28.1%60.3%41.1%26.4%0.3820.89.1M28-
YOLOv5s_7_0 [16]25.0%58.3%40.6%26.7%0.3516.07.0M50-
YOLOv6s [19]--24.6%15.0%-185.217.2M49-
YOLOv7 [30]34.2%48.1%38.6%24.9%0.40105.137.2M18-
YOLOv8s [17]23.9%65.0%39.1%26.1%0.3528.611.1M33-
WeedNet-R [11]90.1%68.4%84.8%-0.7819.025.6M0.5-
YOLO-DP (our)23.1%73.8%38.3%25.0%0.3511.75.2M62-
" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.553, + 0.788, + 0.691 + ], + "angle": 0, + "content": "on PDT dataset. The attention modules are designed on Backbone's floors 3, 5, 7, and 9. Since the number of C3x and C3TR parameters is too large, we only replace them at the 9th layer in order to carry out the experiment. \"v5s\" means YOLOv5s_7.0. \"v5s__our\" stands for the use of Adaptive Large Scale Selective Kernel in the YOLOv5s_7.0. It can be seen from the results in Tab. 7 that the performance of YOLOv5s is greatly improved after Large Scale Selective Kernel is used. The mAP@.5, mAP@.5::.95 and F1 are the best metrics. This indicates that the Large Scale Selective Kernel used in this paper is suitable for dense object detection tasks." + }, + { + "type": "title", + "bbox": [ + 0.216, + 0.711, + 0.451, + 0.725 + ], + "angle": 0, + "content": "7.3 Visualization Research" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.734, + 0.788, + 0.842 + ], + "angle": 0, + "content": "In Fig. 5, we show the detection visualization results on the PDT dataset. We have selected the YOLOv5s and YOLOv8s models, which exhibited superior overall performance in the comparative experiment, to demonstrate and contrast with the YOLO-DP model. (a) and (e) illustrate the Ground Truth for the LL and LH versions, respectively. Observations indicate that the \"Human-in-the-loop\" data annotation approach is practical for weed and pest detection data annotation. The detection results in (b) and (c) suggest that the LL version" + } + ], + [ + { + "type": "header", + "bbox": [ + 0.64, + 0.115, + 0.732, + 0.127 + ], + "angle": 0, + "content": "PDT Dataset" + }, + { + "type": "page_number", + "bbox": [ + 0.769, + 0.116, + 0.786, + 0.127 + ], + "angle": 0, + "content": "13" + }, + { + "type": "table_caption", + "bbox": [ + 0.396, + 0.145, + 0.605, + 0.159 + ], + "angle": 0, + "content": "Table 7: Ablation experiment." + }, + { + "type": "table", + "bbox": [ + 0.333, + 0.184, + 0.675, + 0.304 + ], + "angle": 0, + "content": "
DatasetsApproachPRmAP@.5mAP@.5:95F1GflopsParameters
PDT dataset (LL)v5s_C188.2%88.5%93.9%67.1%0.8825.310.0M
v5s_C288.8%88.4%94.1%67.0%0.8817.27.3M
v5s_C2f88.6%88.5%93.8%67.1%0.8817.47.5M
v5s_C388.9%88.5%94.2%67.0%0.8916.07.0M
v5s_C3x88.7%81.5%87.4%63.2%0.8514.56.5M
v5s_C3TR88.3%89.0%94.1%67.1%0.8915.77.0M
v5s_C3Ghost88.9%88.2%94.2%66.7%0.8812.55.9M
v5s_SE88.8%88.3%94.4%66.6%0.8810.65.1M
v5s_CBAM89.8%87.5%94.4%66.5%0.8910.95.6M
v5s_GAM89.2%87.7%94.0%67.1%0.8816.47.5M
v5s_ECA89.7%87.0%94.3%66.2%0.8810.55.1M
v5s_our89.1%88.5%94.4%67.2%0.8912.26.1M
" + }, + { + "type": "image", + "bbox": [ + 0.233, + 0.335, + 0.339, + 0.417 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.241, + 0.419, + 0.331, + 0.427 + ], + "angle": 0, + "content": "(a): Ground truth (LL)" + }, + { + "type": "image", + "bbox": [ + 0.342, + 0.335, + 0.486, + 0.417 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.406, + 0.419, + 0.462, + 0.427 + ], + "angle": 0, + "content": "(b):YOLOv5s" + }, + { + "type": "image", + "bbox": [ + 0.488, + 0.336, + 0.657, + 0.417 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.549, + 0.419, + 0.605, + 0.427 + ], + "angle": 0, + "content": "(c):YOLOv8s" + }, + { + "type": "image", + "bbox": [ + 0.665, + 0.336, + 0.772, + 0.417 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.69, + 0.418, + 0.748, + 0.425 + ], + "angle": 0, + "content": "(d):YOLO-DP" + }, + { + "type": "image", + "bbox": [ + 0.233, + 0.429, + 0.501, + 0.566 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.321, + 0.567, + 0.413, + 0.577 + ], + "angle": 0, + "content": "(e): Ground truth (LH)" + }, + { + "type": "image", + "bbox": [ + 0.503, + 0.429, + 0.771, + 0.566 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.609, + 0.568, + 0.666, + 0.577 + ], + "angle": 0, + "content": "(f):YOLO-DP" + }, + { + "type": "image_caption", + "bbox": [ + 0.341, + 0.589, + 0.66, + 0.603 + ], + "angle": 0, + "content": "Fig. 5: Visualization of PDT dataset detection." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.64, + 0.788, + 0.776 + ], + "angle": 0, + "content": "of PDT dataset is compatible with widely used detection models, confirming its utility. When compared to (d), it is evident that YOLO-DP delivers outstanding detection performance on LL, with no instances of missing, false, or duplicate detections. Examining the LH version's detection results in (f), we find no missed detections, and the critical issue of current detectors' inability to distinguish dead pine trees from the ground has been overcome. This leads us to conclude that the YOLO-DP model is well-suited for large-scale, high-precision, small-target UAV detection tasks. Moreover, the detection strategy depicted in Fig. 3 of this paper has been validated as effective." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.78, + 0.788, + 0.841 + ], + "angle": 0, + "content": "In Fig. 6, we presents a comparison between YOLO-DP and the Confusion Matrix of the high-performing classification models YOLOv7 and YOLOv8s on the CWC dataset. It is evident that YOLO-DP's classification performance surpasses that of YOLOv7. Furthermore, YOLO-DP's classification capability is" + } + ], + [ + { + "type": "page_number", + "bbox": [ + 0.218, + 0.116, + 0.236, + 0.127 + ], + "angle": 0, + "content": "14" + }, + { + "type": "header", + "bbox": [ + 0.272, + 0.116, + 0.369, + 0.127 + ], + "angle": 0, + "content": "M. Zhou et al." + }, + { + "type": "image", + "bbox": [ + 0.218, + 0.143, + 0.787, + 0.337 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.217, + 0.35, + 0.785, + 0.392 + ], + "angle": 0, + "content": "Fig. 6: Confusion matrix. The rows represent the true class, the columns represent the predicted class, and the confidence value is [0,1]. The format of the matrix is: YOLOv7/YOLOv8s/YOLO-DP." + }, + { + "type": "text", + "bbox": [ + 0.217, + 0.423, + 0.785, + 0.452 + ], + "angle": 0, + "content": "comparable to YOLOv8s, especially when its overall performance exceeds that of YOLOv8s." + }, + { + "type": "title", + "bbox": [ + 0.218, + 0.478, + 0.357, + 0.494 + ], + "angle": 0, + "content": "8 Conclusion" + }, + { + "type": "text", + "bbox": [ + 0.217, + 0.51, + 0.787, + 0.737 + ], + "angle": 0, + "content": "In this study, we introduce the PDT dataset to address the challenging task of UAV pest detection, with the goal of advancing research in UAV detection tasks within the realm of precision agriculture. We also present the CWC dataset for agricultural weed and crop classification, which compensates for the field's deficiency in training model classification capabilities. To tackle the scarcity of baseline model research in this domain, we propose the YOLO-DP model for dense, small-target UAV pest detection and validate its efficacy across three datasets. It is noteworthy that the PDT dataset features a dual-resolution version, and the CWC dataset boasts 11 detailed, texture-similar plant classes. Moreover, we provide a comprehensive evaluation of both datasets, objectively assessing their value. This work has its limitations: the 3D point cloud version of the PDT dataset is not discussed, and there is a size inconsistency issue within the CWC dataset. We plan to address these issues in future endeavors. Lastly, to ensure the continuity of this research, we make the datasets and associated code designed in this paper available on our website." + }, + { + "type": "title", + "bbox": [ + 0.218, + 0.762, + 0.392, + 0.779 + ], + "angle": 0, + "content": "Acknowledgments" + }, + { + "type": "text", + "bbox": [ + 0.217, + 0.795, + 0.785, + 0.84 + ], + "angle": 0, + "content": "This work was supported by Key R&D Program of Shandong Province, China (2023CXGC010112), the Taishan Scholars Program (NO. tsqn202103097, NO. tscy20221110)." + } + ], + [ + { + "type": "header", + "bbox": [ + 0.64, + 0.115, + 0.732, + 0.127 + ], + "angle": 0, + "content": "PDT Dataset" + }, + { + "type": "page_number", + "bbox": [ + 0.769, + 0.117, + 0.786, + 0.127 + ], + "angle": 0, + "content": "15" + }, + { + "type": "title", + "bbox": [ + 0.218, + 0.145, + 0.323, + 0.16 + ], + "angle": 0, + "content": "References" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.177, + 0.785, + 0.218 + ], + "angle": 0, + "content": "1. Agarwal, R., Hariharan, S., Rao, M.N., Agarwal, A.: Weed identification using k-means clustering with color spaces features in multi-spectral images taken by UAV. In: IGARSS. pp. 7047-7050. IEEE (2021)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.219, + 0.785, + 0.245 + ], + "angle": 0, + "content": "2. Bochkovskiy, A., Wang, C., Liao, H.M.: YOLOv4: Optimal speed and accuracy of object detection. CoRR abs/2004.10934 (2020)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.247, + 0.785, + 0.301 + ], + "angle": 0, + "content": "3. Boursianis, A.D., Papadopoulou, M.S., Diamantoulakis, P.D., Liopa-Tsakalidi, A., Barouchas, P., Salahas, G., Karagiannidis, G.K., Wan, S., Goudos, S.K.: Internet of things (IoT) and agricultural unmanned aerial vehicles (UAVs) in smart farming: A comprehensive review. Internet Things 18, 100187 (2022)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.302, + 0.785, + 0.342 + ], + "angle": 0, + "content": "4. university of burgandy: Ribworth dataset. https://universe.roboflow.com/university-of-burgandy-zowkw/ribworth (nov 2022), https://universe.roboflow.com/university-of-burgandy-zowkw/ribworth" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.343, + 0.785, + 0.384 + ], + "angle": 0, + "content": "5. Butte, S., Vakanski, A., Duellman, K., Wang, H., Mirkouei, A.: Potato crop stress identification in aerial images using deep learning-based object detection. CoRR abs/2106.07770 (2021)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.385, + 0.785, + 0.439 + ], + "angle": 0, + "content": "6. Caras, T., Lati, R.N., Holland, D., Dubinin, V.M., Hatib, K., Shulner, I., Keiesar, O., Liddor, G., Paz-Kagan, T.: Monitoring the effects of weed management strategies on tree canopy structure and growth using uav-lidar in a young almond orchard. Comput. Electron. Agric. 216, 108467 (2024)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.44, + 0.785, + 0.453 + ], + "angle": 0, + "content": "7. Carboni: weeds pytorch (2023), https://github.com/carboni123/weeds-pytorch" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.455, + 0.785, + 0.494 + ], + "angle": 0, + "content": "8. Chebrolu, N., Lottes, P., Schaefer, A., Winterhalter, W., Burgard, W., Stachniss, C.: Agricultural robot dataset for plant classification, localization and mapping on sugar beet fields. Int. J. Robotics Res. 36(10), 1045-1052 (2017)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.496, + 0.785, + 0.55 + ], + "angle": 0, + "content": "9. Costa, L.S., Sano, E.E., Ferreira, M.E., Munhoz, C.B.R., Costa, J.V.S., Junior, L.R.A., de Mello, T.R.B., da Cunha Bustamante, M.M.: Woody plant encroachment in a seasonal tropical savanna: Lessons about classifiers and accuracy from UAV images. Remote. Sens. 15(9), 2342 (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.551, + 0.785, + 0.577 + ], + "angle": 0, + "content": "0. Dabhi: Crop and weed detection (2021), https://github.com/ravirajsinh45/Crop_and_weed_detection" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.578, + 0.785, + 0.619 + ], + "angle": 0, + "content": "1. Guo, Z., Goh, H.H., Li, X., Zhang, M., Li, Y.: WeedNet-R: a sugar beet field weed detection algorithm based on enhanced retinanet and context semantic fusion. Frontiers in Plant Science 14, 1226329 (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.62, + 0.785, + 0.66 + ], + "angle": 0, + "content": "2. Han, K., Wang, Y., Tian, Q., Guo, J., Xu, C., Xu, C.: GhostNet: more features from cheap operations. In: CVPR. pp. 1577-1586. Computer Vision Foundation / IEEE (2020)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.661, + 0.785, + 0.702 + ], + "angle": 0, + "content": "3. Huang, J., Luo, Y., Quan, Q., Wang, B., Xue, X., Zhang, Y.: An autonomous task assignment and decision-making method for coverage path planning of multiple pesticide spraying UAVs. Comput. Electron. Agric. 212, 108128 (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.703, + 0.785, + 0.744 + ], + "angle": 0, + "content": "4. Innani, S., Dutande, P., Baheti, B., Talbar, S.N., Baid, U.: Fuse-PN: A novel architecture for anomaly pattern segmentation in aerial agricultural images. In: CVPR Workshops. pp. 2960-2968. Computer Vision Foundation / IEEE (2021)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.745, + 0.785, + 0.785 + ], + "angle": 0, + "content": "5. Jabir, B., Falih, N.: Deep learning-based decision support system for weeds detection in wheat fields. International Journal of Electrical and Computer Engineering 12(1), 816 (2022)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.786, + 0.785, + 0.813 + ], + "angle": 0, + "content": "6. Jocher, G.: YOLOv5 by Ultralytics (May 2020). https://doi.org/10.5281/ zenodo.3908559, https://github.com/ultralytics/yolov5" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.814, + 0.785, + 0.84 + ], + "angle": 0, + "content": "7. Jocher, G., Chaurasia, A., Qiu, J.: Ultralytics YOLO (Jan 2023), https://github.com/ultralytics/ultralytics" + }, + { + "type": "list", + "bbox": [ + 0.226, + 0.177, + 0.785, + 0.84 + ], + "angle": 0, + "content": null + } + ], + [ + { + "type": "page_number", + "bbox": [ + 0.218, + 0.116, + 0.236, + 0.127 + ], + "angle": 0, + "content": "16" + }, + { + "type": "header", + "bbox": [ + 0.272, + 0.115, + 0.369, + 0.127 + ], + "angle": 0, + "content": "M. Zhou et al." + }, + { + "type": "ref_text", + "bbox": [ + 0.218, + 0.147, + 0.787, + 0.203 + ], + "angle": 0, + "content": "18. Lameski, P., Zdravevski, E., Trajkovik, V., Kulakov, A.: Weed detection dataset with RGB images taken under variable light conditions. In: ICT Innovations. Communications in Computer and Information Science, vol. 778, pp. 112-119. Springer (2017)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.204, + 0.787, + 0.259 + ], + "angle": 0, + "content": "19. Li, C., Li, L., Jiang, H., Weng, K., Geng, Y., Li, L., Ke, Z., Li, Q., Cheng, M., Nie, W., Li, Y., Zhang, B., Liang, Y., Zhou, L., Xu, X., Chu, X., Wei, X., Wei, X.: YOLOv6: A single-stage object detection framework for industrial applications. CoRR abs/2209.02976 (2022)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.26, + 0.787, + 0.3 + ], + "angle": 0, + "content": "20. Li, Y., Hou, Q., Zheng, Z., Cheng, M., Yang, J., Li, X.: Large selective kernel network for remote sensing object detection. In: ICCV. pp. 16748-16759. IEEE (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.301, + 0.787, + 0.342 + ], + "angle": 0, + "content": "21. Lin, T., Dólar, P., Girshick, R.B., He, K., Hariharan, B., Belongie, S.J.: Feature pyramid networks for object detection. In: CVPR. pp. 936-944. IEEE Computer Society (2017)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.343, + 0.787, + 0.37 + ], + "angle": 0, + "content": "22. Lin, T., Goyal, P., Girshick, R.B., He, K., Dollar, P.: Focal loss for dense object detection. In: ICCV. pp. 2999-3007. IEEE Computer Society (2017)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.371, + 0.787, + 0.412 + ], + "angle": 0, + "content": "23. Liu, S., Qi, L., Qin, H., Shi, J., Jia, J.: Path aggregation network for instance segmentation. In: CVPR. pp. 8759-8768. Computer Vision Foundation / IEEE Computer Society (2018)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.412, + 0.787, + 0.453 + ], + "angle": 0, + "content": "24. Liu, W., Anguelov, D., Erhan, D., Szegedy, C., Reed, S.E., Fu, C., Berg, A.C.: SSD: single shot multibox detector. In: ECCV (1). Lecture Notes in Computer Science, vol. 9905, pp. 21-37. Springer (2016)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.454, + 0.787, + 0.479 + ], + "angle": 0, + "content": "25. Monster: weed dataset (2019), https://gitee.com/Monster7/weed-database/tree/master/" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.48, + 0.787, + 0.508 + ], + "angle": 0, + "content": "26. Redmon, J., Farhadi, A.: YOLOv3: An incremental improvement. CoRR abs/1804.02767 (2018)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.509, + 0.787, + 0.55 + ], + "angle": 0, + "content": "27. Ren, S., He, K., Girshick, R.B., Sun, J.: Faster R-CNN: towards real-time object detection with region proposal networks. IEEE Trans. Pattern Anal. Mach. Intell. 39(6), 1137-1149 (2017)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.55, + 0.787, + 0.577 + ], + "angle": 0, + "content": "28. Tan, M., Pang, R., Le, Q.V.: EfficientDet: scalable and efficient object detection. In: CVPR. pp. 10778-10787. Computer Vision Foundation / IEEE (2020)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.578, + 0.787, + 0.619 + ], + "angle": 0, + "content": "29. Tang, J., Wang, D., Zhang, Z., He, L., Xin, J., Xu, Y.: Weed identification based on k-means feature learning combined with convolutional neural network. Comput. Electron. Agric. 135, 63-70 (2017)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.62, + 0.787, + 0.66 + ], + "angle": 0, + "content": "30. Wang, C., Bochkovskiy, A., Liao, H.M.: YOLOv7: Trainable bag-of-freebies sets new state-of-the-art for real-time object detectors. In: CVPR. pp. 7464-7475. IEEE (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.661, + 0.787, + 0.702 + ], + "angle": 0, + "content": "31. Zhang, R., Wang, C., Hu, X., Liu, Y., Chen, S., et al.: Weed location and recognition based on UAV imaging and deep learning. International Journal of Precision Agricultural Aviation 3(1) (2020)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.702, + 0.787, + 0.73 + ], + "angle": 0, + "content": "32. Zhou, X., Wang, D., Krahenbuhl, P.: Objects as points. CoRR abs/1904.07850 (2019)" + }, + { + "type": "list", + "bbox": [ + 0.217, + 0.147, + 0.787, + 0.73 + ], + "angle": 0, + "content": null + } + ] +] \ No newline at end of file diff --git a/2024/PDT Uav Target Detection Dataset for Pests and Diseases Tree/77c7722b-38da-423b-b56a-11a16b09eb35_origin.pdf b/2024/PDT Uav Target Detection Dataset for Pests and Diseases Tree/77c7722b-38da-423b-b56a-11a16b09eb35_origin.pdf new file mode 100644 index 0000000000000000000000000000000000000000..88bcbf02337dc7f95065be88049f6118d5086661 --- /dev/null +++ b/2024/PDT Uav Target Detection Dataset for Pests and Diseases Tree/77c7722b-38da-423b-b56a-11a16b09eb35_origin.pdf @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:71efa78ad559e036d9f3af66072cadb773c22b0a2a1b7f34e71412f116a1f687 +size 5165441 diff --git a/2024/PDT Uav Target Detection Dataset for Pests and Diseases Tree/full.md b/2024/PDT Uav Target Detection Dataset for Pests and Diseases Tree/full.md new file mode 100644 index 0000000000000000000000000000000000000000..b858bb03dee87566bd4ccf50c5b4d03783f87474 --- /dev/null +++ b/2024/PDT Uav Target Detection Dataset for Pests and Diseases Tree/full.md @@ -0,0 +1,252 @@ +# PDT: Uav Target Detection Dataset for Pests and Diseases Tree + +Mingle Zhou $^{1,2,4}$ , Rui Xing $^{1,2}$ , Delong Han $^{1,2}$ , Zhiyong Qi $^{3}$ , and Gang Li $^{1,2(\boxtimes)}$ + +1 Key Laboratory of Computing Power Network and Information Security, Ministry of Education, Shandong Computer Science Center (National Supercomputer Center in Jinan), Qilu University of Technology (Shandong Academy of Sciences), Jinan, China lig@qlu.edu.cn + +$^{2}$ Shandong Provincial Key Laboratory of Computer Networks, Shandong Fundamental Research Center for Computer Science, Jinan, China + +3 Institute of Remote Sensing and Geographic Information System, School of Earth and Space Sciences, Peking University, Beijing, China + +4 SHANDONG SCICOM Information and Economy Research Institute Co., Ltd. + +Abstract. UAVs emerge as the optimal carriers for visual weed identification and integrated pest and disease management in crops. However, the absence of specialized datasets impedes the advancement of model development in this domain. To address this, we have developed the Pests and Diseases Tree dataset (PDT dataset). PDT dataset represents the first high-precision UAV-based dataset for targeted detection of tree pests and diseases, which is collected in real-world operational environments and aims to fill the gap in available datasets for this field. Moreover, by aggregating public datasets and network data, we further introduced the Common Weed and Crop dataset (CWC dataset) to address the challenge of inadequate classification capabilities of test models within datasets for this field. Finally, we propose the YOLO-Dense Pest (YOLO-DP) model for high-precision object detection of weed, pest, and disease crop images. We re-evaluate the state-of-the-art detection models with our proposed PDT dataset and CWC dataset, showing the completeness of the dataset and the effectiveness of the YOLO-DP. The proposed PDT dataset, CWC dataset, and YOLO-DP model are presented at https://github.com/RuiXing123/PDT_CWC_YOLO-DP. + +Keywords: Datasets $\cdot$ UAVs $\cdot$ Object detection + +# 1 Introduction + +The efficacy of intelligent UAV-based plant protection operations hinges on precisely identifying weeds and pests within imagery, representing a critical challenge in computer vision [6,9,13]. As computer vision and UAV technologies have advanced swiftly, the adoption of automated UAV plant protection operations is on the rise [3,14]. The working mode of plant protection UAV has significant defects: non-intelligent work can result in the misuse or waste of agrochemicals, + +![](images/80bc44f7affd18d99c46ff7d88fc98a204464af39040f5714c04d1070d202eff.jpg) +Fig. 1: Dataset comparison. (a) shows the PDT dataset (Low Resolution(LL) and High Resolution(LH)): $640 \times 640$ , $5472 \times 3648$ . (b) shows the characteristics of the CWC dataset: it contains 11 different similar plants. (c), (d) and (e) are the public datasets. + +and the high-intensity, repetitive mechanical tasks can diminish their operational lifespan. Enhancing the precision of weed and pest target recognition is essential for the effectiveness of intelligent plant protection UAVs. The absence of specialized datasets and the limitations of existing detection models pose significant constraints on advancing research in this domain. + +Drones typically take images at both high and low altitudes when outdoors. Existing datasets are usually taken from indoor greenhouses (Fig. 1 (c)), where the majority of the data consists of target images captured at close range (Fig. 1 (d)). Such data struggles to incorporate environmental factors like changes in lighting, and the target sizes significantly deviate from real-world conditions, failing to satisfy practical requirements. Further, the majority of existing datasets are limited to single or double classes, which cannot meet the training of model classification ability, as illustrated in Fig. 1 (c), (d), and (e). + +On the other hand, the conventional detection models utilized in UAV-based intelligent plant protection could be more challenging in accurately identifying targets. The deficiency in this field is underscored by the absence of tailored baseline models, resulting in unmet demands for detecting and managing diverse weeds, crops, pests, and diseases. This paper aims to address the lack of dedicated datasets and algorithms for weeds, pests and diseases in agricultural drones. Overall, the main contributions of this paper are as follows: + +- We have developed the UAV target detection agricultural Pests and Diseases Tree dataset, known as PDT dataset. +- We collated public data and provided Common Weed and Crop dataset (CWC dataset) to train the classification capabilities of the model. +- We have designed a specialized detection model, YOLO-DP, tailored for the dense object detection of tree crop pests and diseases. +- We reevaluated and analyzed existing generic and specific detection models on PDT and CWC datasets to establish benchmarks for the field. + +# 2 Related Work + +Agricultural UAV Detection Dataset. In the realm of agricultural robotics, acquiring pertinent datasets frequently presents a challenge, primarily due to the necessity of maintaining dedicated experimental fields and the critical timing of data collection, which is pivotal to the research outcomes. Chebrolu et al. presented SugarBeet2016 large agricultural robot dataset and subsequently introduced the SugarBeet2017 dataset tailored for the detection task [8,11]. The limitation of this dataset is that the data, captured from the vantage point of a ground robot, significantly diverges from the actual detection scenarios encountered by UAVs. Vakanski et al. introduced a dataset of multispectral potato plant images taken by drones [5]. The limitation of this dataset is the lack of raw data, consisting of only 360 images and a total of 1,500 images produced by data enhancement, which cannot effectively guide the model to learn the details of the crop. The University of Burgundy has launched the Ribworth Dataset, designed for crops and weeds, which uses data augmentation to simulate real-world environmental conditions [4]. The limitation of this dataset is that its acquisition target is relatively close, and the factors such as light change are not included, which is not enough to meet the requirements of accurate detection tasks of agricultural plant protection UAVs. + +To foster the advancement of precision agriculture, we introduce the PDT dataset and CWC dataset, aimed at addressing the requirements for detecting agricultural weeds, crops, and pests. + +Agricultural UAV Detection Model. The agricultural UAV target detection task has garnered significant attention within the realms of machine learning and computer vision, owing to its attributes of precise localization, efficient operation, and minimal environmental impact. Among them, the weed detection models based on YOLO deep learning model include: Zhang et al. developed a detection model for weed location and identification in wheat fields by using YOLOv3-tiny [31]. Gao et al. modified the YOLOv3-tiny model and applied it to weed detection in sugar beet fields. Jabir and Falih proposed an intelligent wheat field weed detection system based on YOLOv5 [15]. Furthermore, a variety of neural network architectures have also been recognized by researchers and implemented in the field of precision agriculture. Guo et al. proposed WeedNet-R based on the RetinaNet architecture for weed identification and location in sugar beet fields [11]. Tang et al. employed Kmeans feature learning in conjunction with a convolutional neural network for the recognition of weeds [29]. Agarwal et al. utilized the Kmeans clustering algorithm to analyze color space features in multispectral weed images for effective weed detection [1]. + +It is evident that there is a scarcity of studies focusing on baseline models for the detection of tree diseases and pests, particularly those related to the invasion of the exotic species Red Turpentine Beetle. Furthermore, the majority of these baseline models are trained using single-class or two-class detection datasets, and the classification ability of the models is limited by the diversity of the data. However, for the detection tasks in large-scale, high-precision agricultural UAV operations, precise detection, accurate classification, and the expansion of + +crop detection varieties are essential. Consequently, we propose the YOLO-DP detection model to address these requirements. + +# 3 PDT Dataset + +In this study, we introduced a PDT dataset with both high and low resolution versions, specifically for the detection of red turpentine beetle natural pest targets. As most existing public datasets are gathered indoors or via simulated drones, they fail to capture the characteristics of outdoor UAV detection environments. Consequently, there is an immediate demand for an agricultural dataset that is tailored to real-world operational environments and encompasses the distribution of both high and low-altitude data, thereby fulfilling the high-precision, large-scale target detection requirements from a UAV perspective. This section delineates the data acquisition process for the PDT dataset, encompassing the selection of the data collection domain, the equipment used for acquisition, and the criteria for sample definition. + +# 3.1 Data Acquisition + +Selection of the Data Collection Domain. The Red Turpentine Beetle, a member of the Coleopteridae family, not only invades extensive pine forest areas but also contributes to the widespread mortality of trees, exerting a profound impact on the regional ecological environment and forestry. Despite the implementation of preventive and control strategies, including monitoring and trapping, effectively managing its continued proliferation remains a formidable challenge. The precision spraying platform of plant protection UAV is an effective solution, but the public UAV data for the invasive species Red Turpentine Beetle is blank. Consequently, we opted to collect data from a cluster of dead pine trees that have been infested by the Red Turpentine Beetle. + +High-resolution Drone Camera Equipment. To obtain accurate sample data, we use a UAV mapping camera called DJI-ChanSi L2. This advanced system enhances the UAV flight platform's capabilities for acquiring accurate, efficient, and reliable 2D and 3D data. The LiDAR components have excellent performance, ranging accuracy of $2\mathrm{cm}$ at a height of 150 meters, laser wavelength of $905~\mathrm{nm}$ , and laser pulse emission frequency of $240\mathrm{kHz}$ . The system supports the exFAT file system and can capture images in JPEG or PNG formats. In essence, the DJI-ChanSi L2 is a comprehensive platform that integrates lidar, a high-precision navigation system, and a mapping camera, specifically designed for high-precision data acquisition in mapping, forestry, power grid, and engineering and construction industries. + +Definition of Detection Target. Distinct from close-range pest detection, the task of pest detection at low or high altitudes fundamentally involves capturing the differential characteristics between the affected vector and the healthy state. Consequently, the detection samples for the PDT dataset are sourced from both healthy and unhealthy trees within pine forests. As depicted in Fig. 2, (a) + +![](images/5bbf9b916d924007185e084ba54e0f49e4f09471caa70f75bcaff0f65546d111.jpg) +(a): healthy + +![](images/ee7711eb2a5d834cf5ed6528fe0ac0d83ba72e8187011d979422b09c67bc5361.jpg) + +![](images/d587f16990c6e14f0c0d356df49c131eb76bf54aad1e2fe7cbcf6f1104420253.jpg) +(b): unhealthy + +![](images/6c094c81bda8c2ebf78c1b9ceeeb225096f2034298407262f7dd3159b417e666.jpg) + +![](images/c3450a34a4019f05335b2fd838608c5690e342ce7305db900ff3b4b8d7aae41c.jpg) +Fig. 2: Data example. (a) is a healthy goal and (b) is a unhealthy goal. The PDT dataset takes (b) as the category. +Fig. 3: PDT dataset generation and detection process. (a) represents the sliding window method, and (b) represents the "Human-in-the-loop" data annotation method. (c) means that a LL image is sent to the neural network for training, and LL and LH dual-resolution images are detected at the same time. + +illustrates a pine tree in a healthy state, while (b) displays three pine trees in a unhealthy state that have been infested by pests, exhibiting progressively severe symptoms from left to right. + +# 3.2 Data Preprocessing and Statistics + +This section delves into the acquisition, processing, and production procedures of the original image data about the PDT dataset. We set up an isometric photo mode at an altitude of 150 meters to acquire 105 sets of raw image data and 3D point cloud data. Owing to the substantial size of the original images, they are unsuitable for training with the current mainstream object detection neural networks. We introduce a training approach that applies to the majority of neural network data. As illustrated in Fig. 3, we employ a methodology that encompasses "data preprocessing - 'Human-in-the-loop' data labeling - manual correction". The essence of this method is to leverage the preliminary analysis capabilities of artificial intelligence to aid human annotators, thereby enhancing the efficiency and accuracy of data annotation. + +Data Preprocessing. Fig. 3 (a) demonstrates the application of the sliding window technique on the original image data, which allows for the extraction of a usable, standard-sized image through regional cropping of the high-resolution image. To prevent information loss during the sliding window process, the window dimensions are set to $640 \times 640$ pixels, with a step size of $630 \times 630$ pixels. + +Table 1: Structure of PDT dataset and its sample. + +
EditionClassesStructureTargeted imagesUntargeted imagesImage sizeInstancesTarget Amount
S(Small)M(Medium)L(Large)
SampleunhealthyTrain811640×64025691896548179
Val1969152813825
LLunhealthyTrain316613709029070418163423530
Val395172640×6401252399262165432
Test3901771149489492095450
LHunhealthy-10505472×3648934749347400
+ +Following the sliding window operation, 53 LL images were derived from a single LH image. Furthermore, due to the high similarity between the unhealthy pine tree depicted in Fig. 2 and the ground background, and to enhance the effectiveness of neural network training, we opt to retain images that do not contain targets. Upon generating an output from these data during training, a corresponding loss is produced, which aids the neural network in learning and reduces the rate of false positives. + +Data Annotation. We proceed to annotate the preprocessed data. To ensure the data's validity, we utilize the widely-used Labelme ${}^{1}$ data labeling software. For high-resolution datasets with dense, small targets, the sheer volume of targets per image makes manual annotation a tedious and time-intensive process. As illustrated in Fig. 3 (b), the "Human-in-the-loop" data annotation approach effectively addresses this challenge. Initially, we acquire the LL version samples of the PDT dataset through manual annotation. The structure of the sample dataset is depicted in Tab. 1, which comprises a training set and a validation set. Subsequently, the sample dataset is input into the YOLOv5s model for training over 300 epochs, with the optimal weights being saved. To ensure that no label data is missed, a confidence threshold of 0.25 and an NMS IOU (Intersection over Union) threshold of 0.45 were set to automatically label the LL data and obtain the original label set. + +Manual Check. The labeled dataset underwent manual filtering to eliminate incorrect annotations, rectify erroneous ones, and add any omitted annotations. The LL datasets were generated in two widely recognized formats, YOLO txt and VOCXml, by automatically partitioning the data structure into an 8:1:1 ratio using scripts. For the LH data, we input the completed LL data into the YOLO-DP model for training to acquire the optimal weights. The "Human-in-the-loop" data annotation and subsequent manual verification were conducted on the LH data to obtain the final LH version dataset. + +Data Statistics. The statistics of the PDT dataset are shown in Tab. 1. LL and LH versions consist of 5,670 and 105 images, respectively, with 114,307 and 93,474 samples, and feature a single class labeled as unhealthy. The training, validation, and test sets of the LL version include 1,370, 172, and 144 target-free images, respectively. On average, the LL and LH versions contain 29 and 890 + +Table 2: CWC dataset sources. + +
DatasetsCorn weed datasets [25]lettuce weed radish weed datasets [25] datasets [25]Fresh-weed-data [18]
Classesbluegrass, corn, sedge, chenopodium album, circium_setosumlettuceradishnightshadetomatocottonvelvet
Number2502002011151162438
Image Size800×600800×600800×600800×600800×600586×444643×500
+ +samples per image, respectively, indicating that the PDT dataset is suitable for the task of high-density and high-precision inspection of plant protection drones. + +# 4 CWC Dataset + +We compile the available data to construct the Common Weed and Crop datasets. In special cases, there may be crop crossovers (usually in small trial plots) where weeds and crop targets exhibit multi-class structural characteristics and similarities. The majority of existing public datasets are limited to single or double classes, which do not suffice for the training and classification capabilities required by models. Consequently, when training the neural network, it is essential to supply detailed texture data that differentiates between various objects, thereby equipping the model with robust classification capabilities. This enables the model to effectively sift through and eliminate highly similar negative samples during detection, thereby enhancing the model's detection accuracy. This section discusses the acquisition and processing of CWC original image data. + +# 4.1 Data Sources + +In Tab. 2, we present the source, categories, initial quantities of the CWC raw data, as well as their respective resolutions. As depicted in Fig. 1 (b), the CWC original data comprises a total of 11 categories, characterized by similar size, texture, and color. Please note that we manually annotated the raw Corn weed datasets, Lettuce weed datasets, and Radish weed datasets, while the Fresh-weed-data was already pre-labeled and available in YOLO txt format. + +# 4.2 Data Preprocessing and Statistics + +Data Preprocessing. Owing to the imbalance between positive and negative samples in the original CWC data, there is a risk that the model may become biased towards the majority class, leading to a diminished recognition capability for the minority class. We used the oversampling strategy, using the data augmentation method to equalize the number of samples. The methods include random rotation, random translation, random brightness change, random noise + +Table 3: CWC dataset structure. + +
Classesbluegrasschenopodium-album-cirsium-setosumcornsedgelettuceradishnightshadetomatocottonvelvet
Targeted ImagesTrain200200200200200200200200200200200200
Val404040404040404040404040
Test101010101010101010101010
Targeted AmountS100500000003
M000900000000
L249250250236250444326250210268248
Image Size800×600800×600800×600800×600800×600800×600800×600800×600800×600586×444643×500
+ +addition, flipping and cutout. This enhances the robustness of the dataset, simulating the illumination formula: $\pmb{I}_{out} = \pmb{I}_{in} \times w + [255] \times (1 - w)$ , $w \in [0.35, 1]$ . Where $\pmb{I}_{out}$ , $\pmb{I}_{in}$ is the output and input images. $w$ is the random weight. + +Data Statistics. The statistics for the CWC dataset are presented in Tab. 3. The CWC dataset comprises 2,750 images and 2,599 data entries, with the majority featuring large-sized objects. Data enhancement makes the sample reach a balanced distribution, with an average image containing 1 to 2 objects. + +# 5 Dataset Comparison + +Tab. 4 provides a thorough comparison of the datasets. The LH version of the PDT dataset is exceptionally clear, offering image quality that is 35 to 200 times superior to other publicly available datasets. The PDT dataset is characterized by density, small target and rich real environmental details, which accords with the typical environmental factors of high and low altitude UAV work, and is suitable for training the special UAV detection model for plant protection. Furthermore, the PDT dataset includes a 3D point cloud version. While not highlighted in this study, a 3D data version is planned for development and will be released to the public in the future. + +On the other hand, the CWC dataset demonstrates outstanding classification performance, particularly in terms of category diversity, detailed texture information, and dataset authenticity. CWC dataset has a large number of categories, 2 to 5 times more than other publicly available datasets, which indicates that the CWC dataset is well suited for tasks involving high-precision classification of plant protection UAVs. + +# 6 YOLO-DP Model + +To expedite research efforts in UAV detection models for intelligent agricultural plant protection, we have designed and proposed a model, YOLO-DP (Fig. 4), specifically dedicated to the detection of tree crop diseases and pests, building upon the YOLOv5s foundatin [16]. To ensure the scalability of the model, the Kmeans clustering algorithm is employed to determine the distribution of the + +Table 4: Dataset comparison. + +
DatasetResolutionClassesRich DetailsHigh DefinitionScaleDense TargetTarget ScaleUav CollectionNo Target Image3D Point Cloud DataAnnotation Quality
PDT dataset5472×36481-✓100+✓S✓✓✓“Human-in-the-loop”
640×6401--5K+✓S/M/L✓✓✓
CWC dataset800×60011✓-2K+-L---Manually
SugarBeet2017 [11]1296×9662--5K+-S/M--✓Manually
Multispectral Potato Plants Images (MPP) [5]750×7502--1K+✓L✓--Manually
Ribworth Dataset (RI) [4]320×2401✓-5K+-M/L---Manually
320×320
Corn weed Dataset [25]800×6005✓-6K+-L---No
lettuce weed Dataset [25]800×6002✓-700+-L---No
radish weed Dataset [25]800×6002✓-400+-L---No
Fresh -weed-data [29]800×6004✓-200+-L---Manually
586×444
643×500
crop and weed detection data (CAW) [10]512×5122✓-1K+-L---Manually
Weeds [7]480×4801--600+-S/M✓--Manually
+ +![](images/7ef984eff2dea40bd293b288036c7be84dffa1767f08d087465ad7e40587b066.jpg) +Fig. 4: YOLO-DP baseline model architecture. The FPN [21] + PAN [23] module consists of GhostConv [12], Upsample, Concat, and C3. C stands for Concat, S for Sigmoid, P for channel number dilation, $\times$ for matrix multiplication, and + for matrix addition. + +actual bounding box data, allowing for the dynamic adjustment of the anchor box sizes. In light of the UAV's large-scale detection capabilities, an adaptive Large Scale Selective Kernel is incorporated into the Backbone network to capture the location information of dense, small target pest-infested trees [20]. Taking into account the efficient detection mode of UAVs, GhostConv is utilized in the Neck network to reduce the model size and computational complexity. Concurrently, the receptive field is enlarged to capture more texture feature information, thereby enhancing the model's classification capabilities. A version with decoupled detection heads is provided to minimize the interference between classification and regression losses, allowing the model to focus on specific tasks, improving the accuracy and accuracy of the detection, and also improving the model's generalization ability. + +Adaptive Large Scale Selective Kernel. To tackle the demanding detection task involving a broad range of dense, small targets, we gather pest and + +disease target information by designing a Convolutional group with an extensive receptive field. Firstly, the matrix M obtains the shallow range information M1 by Conv with $\text{Kernelsize}(k) = 5$ and $\text{Padding}(p) = 2$ . The depth information is explored for M1, and the deep range information M2 is obtained using Conv with $k = 7$ and $p = 9$ . Feature selection was performed on shallow and depth range information, and M3 and M4 were obtained by grouping convolution on M1 and M2 using GhostConv with $k = 1$ . The advantage of using GhostConv is that half of the range information is retained and the other half is processed to prevent excessive loss of information. Next, the spatial attention is calculated for M3 and M4 large-scale information. M3 and M4 are concatenated in the channel dimension to obtain M5, and M6 and M7 are obtained by averaging and maximizing the channel dimensions, respectively. M6 and M7 are concatenated in the channel dimension to obtain M8 with scale $b \times 1 \times w \times h$ , which is fed into Conv with $k = 7$ and $p = 4$ for range attention collection to obtain M9. After Sigmoid activation, the channel dimension is expanded to obtain M10 and M11. Matrix multiplication is performed with M3 and M4 to obtain M12 and M13, respectively. After bitwise addition, the final spatial attention matrix M14 is obtained after Conv ( $k = 1$ ). Finally, the output is obtained by matrix multiplication with the input matrix. + +# 7 Experiment + +# 7.1 Experimental Conditions + +The experimental environment of this paper is ubuntu 18.04, Python3.9, PyTorch 1.9.1 framework, CUDA 11.4 version, and cuDNN 8.0 version. This paper's model training and inference are performed on NVIDIA A100-SXM4 with 40GB GPU memory and 16 GB CPU memory on the experimental platform. The training metrics in this paper are: 300 epochs of iterative training, the early stopping mechanism is set to 100 rounds, the initial learning rate is 0.01, the weight decay of 0.001 and momentum of $0.9\%$ are used, the Batch-size is 16. + +Experiments were carried out on the PDT dataset and CWC dataset presented in this paper, as well as on the public dataset SugarBeet2017. We test the performance of the model YOLO-DP according to the characteristics of the dataset. First, the PDT dataset is used to test the detection ability of the model for dense small targets in the UAV perspective. Secondly, the ability of the model to extract fine-grained texture information is measured on the CWC dataset to verify its classification ability. It is reflected in the P metrics. Finally, the detection accuracy was measured on the SugarBeet2017 dataset. It is reflected in the R metrics. In order not to lose the image information, the input sizes are $640 \times 640$ , $800 \times 800$ , $1296 \times 1296$ respectively. + +# 7.2 Experimental Analysis + +Dataset Validation. We evaluated 6 detection models on seven datasets (2 of ours and 5 public datasets mentioned in the paper). Based on the dataset's + +characteristics, we choose different metrics for the ranking model (Rank (Metrics)). Specifically, dense target datasets collected in natural environments use the comprehensive F1 metrics, datasets with target size changes use the R metrics, and multiclass datasets use the P metrics. We sort models with the same metrics score again using Gflops. Our model has excellent performance on all 7 datasets, and the performance of the proposed 2 datasets differs from that of existing datasets, indicating that our dataset can provide unique data distribution to the research field (Tab. 5). + +Table 5: Datasets effect ranking. + +
MethodGflopsRank (F1)Rank (P)Rank (R)
PDT (LL) (our)MPP [5]CWC (our)Weeds [7]SugarBeet2017 [11]CAW [10]RI [4]
YOLO-DP11.71 (0.89)1 (0.42)2 (92.9%)2 (76.5%)1 (73.8%)2 (89.4%)1 (82.3%)
YOLOv3155.35 (0.88)5 (0.38)6 (86.6%)1 (83.0%)6 (46.2%)6 (73.1%)6 (74.0%)
YOLOv4s20.83 (0.88)2 (0.42)5 (87.3%)3 (75.6%)3 (60.3%)3 (86.5%)2 (81.7%)
YOLOv5s16.02 (0.89)3 (0.38)4 (88.6%)4 (75.3%)4 (58.3%)5 (82.5%)3 (81.5%)
YOLOv7105.16 (0.85)6 (0.24)1 (93.1%)5 (74.4%)5 (48.1%)4 (83.3%)4 (80.3%)
YOLOv8s28.64 (0.88)4 (0.38)3 (92.0%)6 (70.4%)2 (65.0%)1 (90.1%)5 (79.3%)
+ +Comparative Experiment. We perform a comparative analysis of single-stage detectors including the full YOLO family, SSD, EfficientDet, and RetinaNet, the two-stage detector Fast-RCNN, anchor-free based CenterNet, and WeedNet-R, a specialized model for weed detection. Among them, some models use pre-trained weights. As some models do not provide the computation results for certain metrics, these are denoted by “-” in the table. + +In Tab. 6, we present the comparative evaluations of the proposed YOLO-DP baseline model alongside other models across the three datasets. The YOLO-DP model does not need to load pre-trained weights, and the detection speed is significantly ahead of other models. Under this premise, YOLO-DP demonstrates the most outstanding overall performance on the PDT dataset, with its F1 composite metrics surpassing all other models. This shows that the YOLO-DP model can well adapt to the detection scenario of UAV with large-scale, high-precision, dense and small targets. On the CWC dataset, its P metrics is at the first-class level, which is $4.3\%$ higher than the YOLOv5s baseline model. This demonstrates that the YOLO-DP model possesses an effective capability for extracting fine-grained texture information and is capable of fulfilling the classification task requirements within this domain. On the authoritative dataset SugarBeet2017, while the overall performance does not match that of models with loaded and trained weights, the R metrics of YOLO-DP is comparable to the pre-trained model's performance. This substantiates that the YOLO-DP model possesses precise detection capabilities and can accurately pinpoint the detection targets. + +The comparative experimental results show that the YOLO-DP model is suitable for the application of plant protection UAV in precision agriculture detection fields such as weeds, pests and diseases, and crops. + +Ablation Experiment. We chose the YOLOv5s that performed well on the PDT dataset as the benchmark. We use the popular attention mechanism to ablate the adaptive large-scale selection kernel to verify its detection performance + +Table 6: Comparative experiment. + +
DatasetsApproachPRmAP@.5mAP@.5::.95F1GflopsParametersFPS Pre-training
PDT dataset (LL)SSD [24]84.5%87.7%85.1%-0.86273.623.7M37✓
EfficientDet [28]92.6%73.4%72.3%-0.8211.56.7M12✓
RetinaNet [22]93.3%65.3%64.2%-0.79109.732.7M32✓
CenterNet [32]95.2%67.4%66.5%-0.79109.732.7M32✓
Faster-RCNN [27]57.8%70.5%61.7%-0.64401.7136.7M13✓
YOLOv3 [26]88.5%88.1%93.4%65.7%0.88155.361.5M41-
YOLOv4s [2]88.8%88.2%94.7%66.1%0.8820.89.1M51-
YOLOv5s_7_0 [16]88.9%88.5%94.2%67.0%0.8916.07.0M93-
YOLOv6s [19]--91.4%63.2%-44.117.2M43-
YOLOv7 [30]87.4%82.6%90.1%55.5%0.85105.137.2M32-
YOLOv8s [17]88.7%87.5%94.0%67.9%0.8828.611.1M60-
WeedNet-R [11]87.7%48.1%70.4%-0.6219.025.6M0.5-
YOLO-DP (our)90.2%88.0%94.5%67.5%0.8911.75.2M109-
CWC datasetSSD [24]97.7%77.6%85.7%-0.91426.923.7M29✓
EfficientDet [28]97.2%98.6%98.6%-0.9011.56.7M13✓
RetinaNet [22]95.1%98.3%98.0%-0.97261.336.4M24✓
CenterNet [32]96.6%73.8%73.3%-0.80171.432.7M27✓
YOLOv3 [26]86.8%89.4%93.2%82.3%0.88154.761.5M30-
YOLOv4s [2]87.3%87.9%91.9%81.5%0.8820.89.1M43-
YOLOv5s_7_0 [16]88.6%88.7%93.0%81.2%0.8916.07.0M65-
YOLOv6s [19]--92.7%84.3%-68.917.2M31-
YOLOv7 [30]93.1%76.4%88.1%75.6%0.84105.137.2M21-
YOLOv8s [17]92.0%89.1%94.0%86.2%0.9128.611.1M38-
WeedNet-R [11]86.1%51.8%71.6%-0.6519.025.6M0.5-
YOLO-DP (our)92.9%87.5%91.8%81.0%0.9011.75.2M72-
Sugar-Beet2017SSD [24]85.0%83.6%79.3%-0.85112023.7M19✓
EfficientDet [28]93.3%79.8%77.8%-0.8611.56.7M16✓
RetinaNet [22]91.7%78.8%76.6%-0.84256.436.3M23✓
CenterNet [32]97.9%51.2%51.0%-0.62117.432.7M41✓
Faster-RCNN [27]63.6%87.4%80.0%-0.73546.9136.7M25✓
YOLOv3 [26]34.8%46.2%39.4%25.6%0.40155.361.5M28-
YOLOv4s [2]28.1%60.3%41.1%26.4%0.3820.89.1M28-
YOLOv5s_7_0 [16]25.0%58.3%40.6%26.7%0.3516.07.0M50-
YOLOv6s [19]--24.6%15.0%-185.217.2M49-
YOLOv7 [30]34.2%48.1%38.6%24.9%0.40105.137.2M18-
YOLOv8s [17]23.9%65.0%39.1%26.1%0.3528.611.1M33-
WeedNet-R [11]90.1%68.4%84.8%-0.7819.025.6M0.5-
YOLO-DP (our)23.1%73.8%38.3%25.0%0.3511.75.2M62-
+ +on PDT dataset. The attention modules are designed on Backbone's floors 3, 5, 7, and 9. Since the number of C3x and C3TR parameters is too large, we only replace them at the 9th layer in order to carry out the experiment. "v5s" means YOLOv5s_7.0. "v5s__our" stands for the use of Adaptive Large Scale Selective Kernel in the YOLOv5s_7.0. It can be seen from the results in Tab. 7 that the performance of YOLOv5s is greatly improved after Large Scale Selective Kernel is used. The mAP@.5, mAP@.5::.95 and F1 are the best metrics. This indicates that the Large Scale Selective Kernel used in this paper is suitable for dense object detection tasks. + +# 7.3 Visualization Research + +In Fig. 5, we show the detection visualization results on the PDT dataset. We have selected the YOLOv5s and YOLOv8s models, which exhibited superior overall performance in the comparative experiment, to demonstrate and contrast with the YOLO-DP model. (a) and (e) illustrate the Ground Truth for the LL and LH versions, respectively. Observations indicate that the "Human-in-the-loop" data annotation approach is practical for weed and pest detection data annotation. The detection results in (b) and (c) suggest that the LL version + +Table 7: Ablation experiment. + +
DatasetsApproachPRmAP@.5mAP@.5:95F1GflopsParameters
PDT dataset (LL)v5s_C188.2%88.5%93.9%67.1%0.8825.310.0M
v5s_C288.8%88.4%94.1%67.0%0.8817.27.3M
v5s_C2f88.6%88.5%93.8%67.1%0.8817.47.5M
v5s_C388.9%88.5%94.2%67.0%0.8916.07.0M
v5s_C3x88.7%81.5%87.4%63.2%0.8514.56.5M
v5s_C3TR88.3%89.0%94.1%67.1%0.8915.77.0M
v5s_C3Ghost88.9%88.2%94.2%66.7%0.8812.55.9M
v5s_SE88.8%88.3%94.4%66.6%0.8810.65.1M
v5s_CBAM89.8%87.5%94.4%66.5%0.8910.95.6M
v5s_GAM89.2%87.7%94.0%67.1%0.8816.47.5M
v5s_ECA89.7%87.0%94.3%66.2%0.8810.55.1M
v5s_our89.1%88.5%94.4%67.2%0.8912.26.1M
+ +![](images/ceafb2fd488380bbfcc29dd894e742119ccdd0771e8ed2c3cf8fc56452afe56a.jpg) +(a): Ground truth (LL) + +![](images/a041a01751df4760b48b9e8f4ba36bca4d1a15ec659587807ca7715750dd4908.jpg) +(b):YOLOv5s + +![](images/06fd8ca766369c50b1ce147af41b00eab3a5a3180ac70164c0b7da0b845fdaae.jpg) +(c):YOLOv8s + +![](images/739026f48194d4c770fe7097cea99b6aced284cd23277737612851d892eb8385.jpg) +(d):YOLO-DP + +![](images/3e3c91b210d3821aa0e1f7a0143f1686de5175e051793e82af5bc03f5822a233.jpg) +(e): Ground truth (LH) +Fig. 5: Visualization of PDT dataset detection. + +![](images/e33321c0f7850da74d1c51d2d9fb75d56c7599a9367a12a18332ac327c07fa7c.jpg) +(f):YOLO-DP + +of PDT dataset is compatible with widely used detection models, confirming its utility. When compared to (d), it is evident that YOLO-DP delivers outstanding detection performance on LL, with no instances of missing, false, or duplicate detections. Examining the LH version's detection results in (f), we find no missed detections, and the critical issue of current detectors' inability to distinguish dead pine trees from the ground has been overcome. This leads us to conclude that the YOLO-DP model is well-suited for large-scale, high-precision, small-target UAV detection tasks. Moreover, the detection strategy depicted in Fig. 3 of this paper has been validated as effective. + +In Fig. 6, we presents a comparison between YOLO-DP and the Confusion Matrix of the high-performing classification models YOLOv7 and YOLOv8s on the CWC dataset. It is evident that YOLO-DP's classification performance surpasses that of YOLOv7. Furthermore, YOLO-DP's classification capability is + +![](images/81072e389a695241e9a2bbcfb447b15e823afacfdc8252a4a78a3ceed113b550.jpg) +Fig. 6: Confusion matrix. The rows represent the true class, the columns represent the predicted class, and the confidence value is [0,1]. The format of the matrix is: YOLOv7/YOLOv8s/YOLO-DP. + +comparable to YOLOv8s, especially when its overall performance exceeds that of YOLOv8s. + +# 8 Conclusion + +In this study, we introduce the PDT dataset to address the challenging task of UAV pest detection, with the goal of advancing research in UAV detection tasks within the realm of precision agriculture. We also present the CWC dataset for agricultural weed and crop classification, which compensates for the field's deficiency in training model classification capabilities. To tackle the scarcity of baseline model research in this domain, we propose the YOLO-DP model for dense, small-target UAV pest detection and validate its efficacy across three datasets. It is noteworthy that the PDT dataset features a dual-resolution version, and the CWC dataset boasts 11 detailed, texture-similar plant classes. Moreover, we provide a comprehensive evaluation of both datasets, objectively assessing their value. This work has its limitations: the 3D point cloud version of the PDT dataset is not discussed, and there is a size inconsistency issue within the CWC dataset. We plan to address these issues in future endeavors. Lastly, to ensure the continuity of this research, we make the datasets and associated code designed in this paper available on our website. + +# Acknowledgments + +This work was supported by Key R&D Program of Shandong Province, China (2023CXGC010112), the Taishan Scholars Program (NO. tsqn202103097, NO. tscy20221110). + +# References + +1. Agarwal, R., Hariharan, S., Rao, M.N., Agarwal, A.: Weed identification using k-means clustering with color spaces features in multi-spectral images taken by UAV. In: IGARSS. pp. 7047-7050. IEEE (2021) +2. Bochkovskiy, A., Wang, C., Liao, H.M.: YOLOv4: Optimal speed and accuracy of object detection. CoRR abs/2004.10934 (2020) +3. Boursianis, A.D., Papadopoulou, M.S., Diamantoulakis, P.D., Liopa-Tsakalidi, A., Barouchas, P., Salahas, G., Karagiannidis, G.K., Wan, S., Goudos, S.K.: Internet of things (IoT) and agricultural unmanned aerial vehicles (UAVs) in smart farming: A comprehensive review. Internet Things 18, 100187 (2022) +4. university of burgandy: Ribworth dataset. https://universe.roboflow.com/university-of-burgandy-zowkw/ribworth (nov 2022), https://universe.roboflow.com/university-of-burgandy-zowkw/ribworth +5. Butte, S., Vakanski, A., Duellman, K., Wang, H., Mirkouei, A.: Potato crop stress identification in aerial images using deep learning-based object detection. CoRR abs/2106.07770 (2021) +6. Caras, T., Lati, R.N., Holland, D., Dubinin, V.M., Hatib, K., Shulner, I., Keiesar, O., Liddor, G., Paz-Kagan, T.: Monitoring the effects of weed management strategies on tree canopy structure and growth using uav-lidar in a young almond orchard. Comput. Electron. Agric. 216, 108467 (2024) +7. Carboni: weeds pytorch (2023), https://github.com/carboni123/weeds-pytorch +8. Chebrolu, N., Lottes, P., Schaefer, A., Winterhalter, W., Burgard, W., Stachniss, C.: Agricultural robot dataset for plant classification, localization and mapping on sugar beet fields. Int. J. Robotics Res. 36(10), 1045-1052 (2017) +9. Costa, L.S., Sano, E.E., Ferreira, M.E., Munhoz, C.B.R., Costa, J.V.S., Junior, L.R.A., de Mello, T.R.B., da Cunha Bustamante, M.M.: Woody plant encroachment in a seasonal tropical savanna: Lessons about classifiers and accuracy from UAV images. Remote. Sens. 15(9), 2342 (2023) +0. Dabhi: Crop and weed detection (2021), https://github.com/ravirajsinh45/Crop_and_weed_detection +1. Guo, Z., Goh, H.H., Li, X., Zhang, M., Li, Y.: WeedNet-R: a sugar beet field weed detection algorithm based on enhanced retinanet and context semantic fusion. Frontiers in Plant Science 14, 1226329 (2023) +2. Han, K., Wang, Y., Tian, Q., Guo, J., Xu, C., Xu, C.: GhostNet: more features from cheap operations. In: CVPR. pp. 1577-1586. Computer Vision Foundation / IEEE (2020) +3. Huang, J., Luo, Y., Quan, Q., Wang, B., Xue, X., Zhang, Y.: An autonomous task assignment and decision-making method for coverage path planning of multiple pesticide spraying UAVs. Comput. Electron. Agric. 212, 108128 (2023) +4. Innani, S., Dutande, P., Baheti, B., Talbar, S.N., Baid, U.: Fuse-PN: A novel architecture for anomaly pattern segmentation in aerial agricultural images. In: CVPR Workshops. pp. 2960-2968. Computer Vision Foundation / IEEE (2021) +5. Jabir, B., Falih, N.: Deep learning-based decision support system for weeds detection in wheat fields. International Journal of Electrical and Computer Engineering 12(1), 816 (2022) +6. Jocher, G.: YOLOv5 by Ultralytics (May 2020). https://doi.org/10.5281/ zenodo.3908559, https://github.com/ultralytics/yolov5 +7. Jocher, G., Chaurasia, A., Qiu, J.: Ultralytics YOLO (Jan 2023), https://github.com/ultralytics/ultralytics + +18. Lameski, P., Zdravevski, E., Trajkovik, V., Kulakov, A.: Weed detection dataset with RGB images taken under variable light conditions. In: ICT Innovations. Communications in Computer and Information Science, vol. 778, pp. 112-119. Springer (2017) +19. Li, C., Li, L., Jiang, H., Weng, K., Geng, Y., Li, L., Ke, Z., Li, Q., Cheng, M., Nie, W., Li, Y., Zhang, B., Liang, Y., Zhou, L., Xu, X., Chu, X., Wei, X., Wei, X.: YOLOv6: A single-stage object detection framework for industrial applications. CoRR abs/2209.02976 (2022) +20. Li, Y., Hou, Q., Zheng, Z., Cheng, M., Yang, J., Li, X.: Large selective kernel network for remote sensing object detection. In: ICCV. pp. 16748-16759. IEEE (2023) +21. Lin, T., Dólar, P., Girshick, R.B., He, K., Hariharan, B., Belongie, S.J.: Feature pyramid networks for object detection. In: CVPR. pp. 936-944. IEEE Computer Society (2017) +22. Lin, T., Goyal, P., Girshick, R.B., He, K., Dollar, P.: Focal loss for dense object detection. In: ICCV. pp. 2999-3007. IEEE Computer Society (2017) +23. Liu, S., Qi, L., Qin, H., Shi, J., Jia, J.: Path aggregation network for instance segmentation. In: CVPR. pp. 8759-8768. Computer Vision Foundation / IEEE Computer Society (2018) +24. Liu, W., Anguelov, D., Erhan, D., Szegedy, C., Reed, S.E., Fu, C., Berg, A.C.: SSD: single shot multibox detector. In: ECCV (1). Lecture Notes in Computer Science, vol. 9905, pp. 21-37. Springer (2016) +25. Monster: weed dataset (2019), https://gitee.com/Monster7/weed-database/tree/master/ +26. Redmon, J., Farhadi, A.: YOLOv3: An incremental improvement. CoRR abs/1804.02767 (2018) +27. Ren, S., He, K., Girshick, R.B., Sun, J.: Faster R-CNN: towards real-time object detection with region proposal networks. IEEE Trans. Pattern Anal. Mach. Intell. 39(6), 1137-1149 (2017) +28. Tan, M., Pang, R., Le, Q.V.: EfficientDet: scalable and efficient object detection. In: CVPR. pp. 10778-10787. Computer Vision Foundation / IEEE (2020) +29. Tang, J., Wang, D., Zhang, Z., He, L., Xin, J., Xu, Y.: Weed identification based on k-means feature learning combined with convolutional neural network. Comput. Electron. Agric. 135, 63-70 (2017) +30. Wang, C., Bochkovskiy, A., Liao, H.M.: YOLOv7: Trainable bag-of-freebies sets new state-of-the-art for real-time object detectors. In: CVPR. pp. 7464-7475. IEEE (2023) +31. Zhang, R., Wang, C., Hu, X., Liu, Y., Chen, S., et al.: Weed location and recognition based on UAV imaging and deep learning. International Journal of Precision Agricultural Aviation 3(1) (2020) +32. Zhou, X., Wang, D., Krahenbuhl, P.: Objects as points. CoRR abs/1904.07850 (2019) \ No newline at end of file diff --git a/2024/PDT Uav Target Detection Dataset for Pests and Diseases Tree/images.zip b/2024/PDT Uav Target Detection Dataset for Pests and Diseases Tree/images.zip new file mode 100644 index 0000000000000000000000000000000000000000..0cfa2c9492912ffa67d9096b4a8fcef5fbc44084 --- /dev/null +++ b/2024/PDT Uav Target Detection Dataset for Pests and Diseases Tree/images.zip @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:6f5cb15ddf9d4d21b429223c25e44ba0b6ea497c3e566b0cf983219915151ef5 +size 657142 diff --git a/2024/PDT Uav Target Detection Dataset for Pests and Diseases Tree/layout.json b/2024/PDT Uav Target Detection Dataset for Pests and Diseases Tree/layout.json new file mode 100644 index 0000000000000000000000000000000000000000..19a471052cb6b6a603e80c6f732ff88d37f1f666 --- /dev/null +++ b/2024/PDT Uav Target Detection Dataset for Pests and Diseases Tree/layout.json @@ -0,0 +1,7155 @@ +{ + "pdf_info": [ + { + "para_blocks": [ + { + "bbox": [ + 146, + 111, + 467, + 145 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 146, + 111, + 467, + 145 + ], + "spans": [ + { + "bbox": [ + 146, + 111, + 467, + 145 + ], + "type": "text", + "content": "PDT: Uav Target Detection Dataset for Pests and Diseases Tree" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 194, + 167, + 419, + 194 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 194, + 167, + 419, + 194 + ], + "spans": [ + { + "bbox": [ + 194, + 167, + 419, + 194 + ], + "type": "text", + "content": "Mingle Zhou" + }, + { + "bbox": [ + 194, + 167, + 419, + 194 + ], + "type": "inline_equation", + "content": "^{1,2,4}" + }, + { + "bbox": [ + 194, + 167, + 419, + 194 + ], + "type": "text", + "content": ", Rui Xing" + }, + { + "bbox": [ + 194, + 167, + 419, + 194 + ], + "type": "inline_equation", + "content": "^{1,2}" + }, + { + "bbox": [ + 194, + 167, + 419, + 194 + ], + "type": "text", + "content": ", Delong Han" + }, + { + "bbox": [ + 194, + 167, + 419, + 194 + ], + "type": "inline_equation", + "content": "^{1,2}" + }, + { + "bbox": [ + 194, + 167, + 419, + 194 + ], + "type": "text", + "content": ", Zhiyong Qi" + }, + { + "bbox": [ + 194, + 167, + 419, + 194 + ], + "type": "inline_equation", + "content": "^{3}" + }, + { + "bbox": [ + 194, + 167, + 419, + 194 + ], + "type": "text", + "content": ", and Gang Li" + }, + { + "bbox": [ + 194, + 167, + 419, + 194 + ], + "type": "inline_equation", + "content": "^{1,2(\\boxtimes)}" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 130, + 201, + 481, + 246 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 201, + 481, + 246 + ], + "spans": [ + { + "bbox": [ + 130, + 201, + 481, + 246 + ], + "type": "text", + "content": "1 Key Laboratory of Computing Power Network and Information Security, Ministry of Education, Shandong Computer Science Center (National Supercomputer Center in Jinan), Qilu University of Technology (Shandong Academy of Sciences), Jinan, China lig@qlu.edu.cn" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 154, + 245, + 457, + 268 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 154, + 245, + 457, + 268 + ], + "spans": [ + { + "bbox": [ + 154, + 245, + 457, + 268 + ], + "type": "inline_equation", + "content": "^{2}" + }, + { + "bbox": [ + 154, + 245, + 457, + 268 + ], + "type": "text", + "content": " Shandong Provincial Key Laboratory of Computer Networks, Shandong Fundamental Research Center for Computer Science, Jinan, China" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 135, + 267, + 477, + 290 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 135, + 267, + 477, + 290 + ], + "spans": [ + { + "bbox": [ + 135, + 267, + 477, + 290 + ], + "type": "text", + "content": "3 Institute of Remote Sensing and Geographic Information System, School of Earth and Space Sciences, Peking University, Beijing, China" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 140, + 289, + 471, + 300 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 140, + 289, + 471, + 300 + ], + "spans": [ + { + "bbox": [ + 140, + 289, + 471, + 300 + ], + "type": "text", + "content": "4 SHANDONG SCICOM Information and Economy Research Institute Co., Ltd." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 159, + 329, + 455, + 528 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 159, + 329, + 455, + 528 + ], + "spans": [ + { + "bbox": [ + 159, + 329, + 455, + 528 + ], + "type": "text", + "content": "Abstract. UAVs emerge as the optimal carriers for visual weed identification and integrated pest and disease management in crops. However, the absence of specialized datasets impedes the advancement of model development in this domain. To address this, we have developed the Pests and Diseases Tree dataset (PDT dataset). PDT dataset represents the first high-precision UAV-based dataset for targeted detection of tree pests and diseases, which is collected in real-world operational environments and aims to fill the gap in available datasets for this field. Moreover, by aggregating public datasets and network data, we further introduced the Common Weed and Crop dataset (CWC dataset) to address the challenge of inadequate classification capabilities of test models within datasets for this field. Finally, we propose the YOLO-Dense Pest (YOLO-DP) model for high-precision object detection of weed, pest, and disease crop images. We re-evaluate the state-of-the-art detection models with our proposed PDT dataset and CWC dataset, showing the completeness of the dataset and the effectiveness of the YOLO-DP. The proposed PDT dataset, CWC dataset, and YOLO-DP model are presented at https://github.com/RuiXing123/PDT_CWC_YOLO-DP." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 160, + 537, + 360, + 548 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 160, + 537, + 360, + 548 + ], + "spans": [ + { + "bbox": [ + 160, + 537, + 360, + 548 + ], + "type": "text", + "content": "Keywords: Datasets " + }, + { + "bbox": [ + 160, + 537, + 360, + 548 + ], + "type": "inline_equation", + "content": "\\cdot" + }, + { + "bbox": [ + 160, + 537, + 360, + 548 + ], + "type": "text", + "content": " UAVs " + }, + { + "bbox": [ + 160, + 537, + 360, + 548 + ], + "type": "inline_equation", + "content": "\\cdot" + }, + { + "bbox": [ + 160, + 537, + 360, + 548 + ], + "type": "text", + "content": " Object detection" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 132, + 569, + 230, + 582 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 569, + 230, + 582 + ], + "spans": [ + { + "bbox": [ + 132, + 569, + 230, + 582 + ], + "type": "text", + "content": "1 Introduction" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 130, + 594, + 482, + 666 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 594, + 482, + 666 + ], + "spans": [ + { + "bbox": [ + 130, + 594, + 482, + 666 + ], + "type": "text", + "content": "The efficacy of intelligent UAV-based plant protection operations hinges on precisely identifying weeds and pests within imagery, representing a critical challenge in computer vision [6,9,13]. As computer vision and UAV technologies have advanced swiftly, the adoption of automated UAV plant protection operations is on the rise [3,14]. The working mode of plant protection UAV has significant defects: non-intelligent work can result in the misuse or waste of agrochemicals," + } + ] + } + ], + "index": 9 + } + ], + "discarded_blocks": [], + "page_size": [ + 612, + 792 + ], + "page_idx": 0 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 151, + 114, + 462, + 254 + ], + "blocks": [ + { + "bbox": [ + 151, + 114, + 462, + 254 + ], + "lines": [ + { + "bbox": [ + 151, + 114, + 462, + 254 + ], + "spans": [ + { + "bbox": [ + 151, + 114, + 462, + 254 + ], + "type": "image", + "image_path": "80bc44f7affd18d99c46ff7d88fc98a204464af39040f5714c04d1070d202eff.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 130, + 264, + 482, + 299 + ], + "lines": [ + { + "bbox": [ + 130, + 264, + 482, + 299 + ], + "spans": [ + { + "bbox": [ + 130, + 264, + 482, + 299 + ], + "type": "text", + "content": "Fig. 1: Dataset comparison. (a) shows the PDT dataset (Low Resolution(LL) and High Resolution(LH)): " + }, + { + "bbox": [ + 130, + 264, + 482, + 299 + ], + "type": "inline_equation", + "content": "640 \\times 640" + }, + { + "bbox": [ + 130, + 264, + 482, + 299 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 130, + 264, + 482, + 299 + ], + "type": "inline_equation", + "content": "5472 \\times 3648" + }, + { + "bbox": [ + 130, + 264, + 482, + 299 + ], + "type": "text", + "content": ". (b) shows the characteristics of the CWC dataset: it contains 11 different similar plants. (c), (d) and (e) are the public datasets." + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_caption" + } + ], + "index": 2 + }, + { + "bbox": [ + 130, + 322, + 479, + 380 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 322, + 479, + 380 + ], + "spans": [ + { + "bbox": [ + 130, + 322, + 479, + 380 + ], + "type": "text", + "content": "and the high-intensity, repetitive mechanical tasks can diminish their operational lifespan. Enhancing the precision of weed and pest target recognition is essential for the effectiveness of intelligent plant protection UAVs. The absence of specialized datasets and the limitations of existing detection models pose significant constraints on advancing research in this domain." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 130, + 382, + 480, + 477 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 382, + 480, + 477 + ], + "spans": [ + { + "bbox": [ + 130, + 382, + 480, + 477 + ], + "type": "text", + "content": "Drones typically take images at both high and low altitudes when outdoors. Existing datasets are usually taken from indoor greenhouses (Fig. 1 (c)), where the majority of the data consists of target images captured at close range (Fig. 1 (d)). Such data struggles to incorporate environmental factors like changes in lighting, and the target sizes significantly deviate from real-world conditions, failing to satisfy practical requirements. Further, the majority of existing datasets are limited to single or double classes, which cannot meet the training of model classification ability, as illustrated in Fig. 1 (c), (d), and (e)." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 130, + 477, + 480, + 561 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 477, + 480, + 561 + ], + "spans": [ + { + "bbox": [ + 130, + 477, + 480, + 561 + ], + "type": "text", + "content": "On the other hand, the conventional detection models utilized in UAV-based intelligent plant protection could be more challenging in accurately identifying targets. The deficiency in this field is underscored by the absence of tailored baseline models, resulting in unmet demands for detecting and managing diverse weeds, crops, pests, and diseases. This paper aims to address the lack of dedicated datasets and algorithms for weeds, pests and diseases in agricultural drones. Overall, the main contributions of this paper are as follows:" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 137, + 569, + 479, + 664 + ], + "type": "list", + "angle": 0, + "index": 11, + "blocks": [ + { + "bbox": [ + 138, + 569, + 479, + 592 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 569, + 479, + 592 + ], + "spans": [ + { + "bbox": [ + 138, + 569, + 479, + 592 + ], + "type": "text", + "content": "- We have developed the UAV target detection agricultural Pests and Diseases Tree dataset, known as PDT dataset." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 137, + 594, + 479, + 616 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 137, + 594, + 479, + 616 + ], + "spans": [ + { + "bbox": [ + 137, + 594, + 479, + 616 + ], + "type": "text", + "content": "- We collated public data and provided Common Weed and Crop dataset (CWC dataset) to train the classification capabilities of the model." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 137, + 617, + 479, + 640 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 137, + 617, + 479, + 640 + ], + "spans": [ + { + "bbox": [ + 137, + 617, + 479, + 640 + ], + "type": "text", + "content": "- We have designed a specialized detection model, YOLO-DP, tailored for the dense object detection of tree crop pests and diseases." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 137, + 642, + 479, + 664 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 137, + 642, + 479, + 664 + ], + "spans": [ + { + "bbox": [ + 137, + 642, + 479, + 664 + ], + "type": "text", + "content": "- We reevaluated and analyzed existing generic and specific detection models on PDT and CWC datasets to establish benchmarks for the field." + } + ] + } + ], + "index": 10 + } + ], + "sub_type": "text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "spans": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "type": "text", + "content": "2" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 166, + 91, + 226, + 101 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 166, + 91, + 226, + 101 + ], + "spans": [ + { + "bbox": [ + 166, + 91, + 226, + 101 + ], + "type": "text", + "content": "M. Zhou et al." + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 1 + }, + { + "para_blocks": [ + { + "bbox": [ + 132, + 114, + 237, + 127 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 114, + 237, + 127 + ], + "spans": [ + { + "bbox": [ + 132, + 114, + 237, + 127 + ], + "type": "text", + "content": "2 Related Work" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 130, + 139, + 482, + 353 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 139, + 482, + 353 + ], + "spans": [ + { + "bbox": [ + 130, + 139, + 482, + 353 + ], + "type": "text", + "content": "Agricultural UAV Detection Dataset. In the realm of agricultural robotics, acquiring pertinent datasets frequently presents a challenge, primarily due to the necessity of maintaining dedicated experimental fields and the critical timing of data collection, which is pivotal to the research outcomes. Chebrolu et al. presented SugarBeet2016 large agricultural robot dataset and subsequently introduced the SugarBeet2017 dataset tailored for the detection task [8,11]. The limitation of this dataset is that the data, captured from the vantage point of a ground robot, significantly diverges from the actual detection scenarios encountered by UAVs. Vakanski et al. introduced a dataset of multispectral potato plant images taken by drones [5]. The limitation of this dataset is the lack of raw data, consisting of only 360 images and a total of 1,500 images produced by data enhancement, which cannot effectively guide the model to learn the details of the crop. The University of Burgundy has launched the Ribworth Dataset, designed for crops and weeds, which uses data augmentation to simulate real-world environmental conditions [4]. The limitation of this dataset is that its acquisition target is relatively close, and the factors such as light change are not included, which is not enough to meet the requirements of accurate detection tasks of agricultural plant protection UAVs." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 130, + 354, + 482, + 390 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 354, + 482, + 390 + ], + "spans": [ + { + "bbox": [ + 130, + 354, + 482, + 390 + ], + "type": "text", + "content": "To foster the advancement of precision agriculture, we introduce the PDT dataset and CWC dataset, aimed at addressing the requirements for detecting agricultural weeds, crops, and pests." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 130, + 390, + 482, + 582 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 390, + 482, + 582 + ], + "spans": [ + { + "bbox": [ + 130, + 390, + 482, + 582 + ], + "type": "text", + "content": "Agricultural UAV Detection Model. The agricultural UAV target detection task has garnered significant attention within the realms of machine learning and computer vision, owing to its attributes of precise localization, efficient operation, and minimal environmental impact. Among them, the weed detection models based on YOLO deep learning model include: Zhang et al. developed a detection model for weed location and identification in wheat fields by using YOLOv3-tiny [31]. Gao et al. modified the YOLOv3-tiny model and applied it to weed detection in sugar beet fields. Jabir and Falih proposed an intelligent wheat field weed detection system based on YOLOv5 [15]. Furthermore, a variety of neural network architectures have also been recognized by researchers and implemented in the field of precision agriculture. Guo et al. proposed WeedNet-R based on the RetinaNet architecture for weed identification and location in sugar beet fields [11]. Tang et al. employed Kmeans feature learning in conjunction with a convolutional neural network for the recognition of weeds [29]. Agarwal et al. utilized the Kmeans clustering algorithm to analyze color space features in multispectral weed images for effective weed detection [1]." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 130, + 582, + 482, + 665 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 582, + 482, + 665 + ], + "spans": [ + { + "bbox": [ + 130, + 582, + 482, + 665 + ], + "type": "text", + "content": "It is evident that there is a scarcity of studies focusing on baseline models for the detection of tree diseases and pests, particularly those related to the invasion of the exotic species Red Turpentine Beetle. Furthermore, the majority of these baseline models are trained using single-class or two-class detection datasets, and the classification ability of the models is limited by the diversity of the data. However, for the detection tasks in large-scale, high-precision agricultural UAV operations, precise detection, accurate classification, and the expansion of" + } + ] + } + ], + "index": 6 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 391, + 91, + 447, + 100 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 391, + 91, + 447, + 100 + ], + "spans": [ + { + "bbox": [ + 391, + 91, + 447, + 100 + ], + "type": "text", + "content": "PDT Dataset" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 474, + 92, + 480, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 474, + 92, + 480, + 100 + ], + "spans": [ + { + "bbox": [ + 474, + 92, + 480, + 100 + ], + "type": "text", + "content": "3" + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 2 + }, + { + "para_blocks": [ + { + "bbox": [ + 130, + 116, + 479, + 140 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 116, + 479, + 140 + ], + "spans": [ + { + "bbox": [ + 130, + 116, + 479, + 140 + ], + "type": "text", + "content": "crop detection varieties are essential. Consequently, we propose the YOLO-DP detection model to address these requirements." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 132, + 159, + 234, + 171 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 159, + 234, + 171 + ], + "spans": [ + { + "bbox": [ + 132, + 159, + 234, + 171 + ], + "type": "text", + "content": "3 PDT Dataset" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 130, + 184, + 482, + 316 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 184, + 482, + 316 + ], + "spans": [ + { + "bbox": [ + 130, + 184, + 482, + 316 + ], + "type": "text", + "content": "In this study, we introduced a PDT dataset with both high and low resolution versions, specifically for the detection of red turpentine beetle natural pest targets. As most existing public datasets are gathered indoors or via simulated drones, they fail to capture the characteristics of outdoor UAV detection environments. Consequently, there is an immediate demand for an agricultural dataset that is tailored to real-world operational environments and encompasses the distribution of both high and low-altitude data, thereby fulfilling the high-precision, large-scale target detection requirements from a UAV perspective. This section delineates the data acquisition process for the PDT dataset, encompassing the selection of the data collection domain, the equipment used for acquisition, and the criteria for sample definition." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 132, + 334, + 246, + 346 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 334, + 246, + 346 + ], + "spans": [ + { + "bbox": [ + 132, + 334, + 246, + 346 + ], + "type": "text", + "content": "3.1 Data Acquisition" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 130, + 354, + 482, + 474 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 354, + 482, + 474 + ], + "spans": [ + { + "bbox": [ + 130, + 354, + 482, + 474 + ], + "type": "text", + "content": "Selection of the Data Collection Domain. The Red Turpentine Beetle, a member of the Coleopteridae family, not only invades extensive pine forest areas but also contributes to the widespread mortality of trees, exerting a profound impact on the regional ecological environment and forestry. Despite the implementation of preventive and control strategies, including monitoring and trapping, effectively managing its continued proliferation remains a formidable challenge. The precision spraying platform of plant protection UAV is an effective solution, but the public UAV data for the invasive species Red Turpentine Beetle is blank. Consequently, we opted to collect data from a cluster of dead pine trees that have been infested by the Red Turpentine Beetle." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 130, + 474, + 482, + 605 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 474, + 482, + 605 + ], + "spans": [ + { + "bbox": [ + 130, + 474, + 482, + 605 + ], + "type": "text", + "content": "High-resolution Drone Camera Equipment. To obtain accurate sample data, we use a UAV mapping camera called DJI-ChanSi L2. This advanced system enhances the UAV flight platform's capabilities for acquiring accurate, efficient, and reliable 2D and 3D data. The LiDAR components have excellent performance, ranging accuracy of " + }, + { + "bbox": [ + 130, + 474, + 482, + 605 + ], + "type": "inline_equation", + "content": "2\\mathrm{cm}" + }, + { + "bbox": [ + 130, + 474, + 482, + 605 + ], + "type": "text", + "content": " at a height of 150 meters, laser wavelength of " + }, + { + "bbox": [ + 130, + 474, + 482, + 605 + ], + "type": "inline_equation", + "content": "905~\\mathrm{nm}" + }, + { + "bbox": [ + 130, + 474, + 482, + 605 + ], + "type": "text", + "content": ", and laser pulse emission frequency of " + }, + { + "bbox": [ + 130, + 474, + 482, + 605 + ], + "type": "inline_equation", + "content": "240\\mathrm{kHz}" + }, + { + "bbox": [ + 130, + 474, + 482, + 605 + ], + "type": "text", + "content": ". The system supports the exFAT file system and can capture images in JPEG or PNG formats. In essence, the DJI-ChanSi L2 is a comprehensive platform that integrates lidar, a high-precision navigation system, and a mapping camera, specifically designed for high-precision data acquisition in mapping, forestry, power grid, and engineering and construction industries." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 130, + 605, + 482, + 666 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 605, + 482, + 666 + ], + "spans": [ + { + "bbox": [ + 130, + 605, + 482, + 666 + ], + "type": "text", + "content": "Definition of Detection Target. Distinct from close-range pest detection, the task of pest detection at low or high altitudes fundamentally involves capturing the differential characteristics between the affected vector and the healthy state. Consequently, the detection samples for the PDT dataset are sourced from both healthy and unhealthy trees within pine forests. As depicted in Fig. 2, (a)" + } + ] + } + ], + "index": 8 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "spans": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "type": "text", + "content": "4" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 166, + 91, + 225, + 100 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 166, + 91, + 225, + 100 + ], + "spans": [ + { + "bbox": [ + 166, + 91, + 225, + 100 + ], + "type": "text", + "content": "M. Zhou et al." + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 3 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 176, + 117, + 235, + 166 + ], + "blocks": [ + { + "bbox": [ + 176, + 117, + 235, + 166 + ], + "lines": [ + { + "bbox": [ + 176, + 117, + 235, + 166 + ], + "spans": [ + { + "bbox": [ + 176, + 117, + 235, + 166 + ], + "type": "image", + "image_path": "5bbf9b916d924007185e084ba54e0f49e4f09471caa70f75bcaff0f65546d111.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 188, + 169, + 221, + 176 + ], + "lines": [ + { + "bbox": [ + 188, + 169, + 221, + 176 + ], + "spans": [ + { + "bbox": [ + 188, + 169, + 221, + 176 + ], + "type": "text", + "content": "(a): healthy" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_caption" + } + ], + "index": 2 + }, + { + "type": "image", + "bbox": [ + 246, + 117, + 298, + 165 + ], + "blocks": [ + { + "bbox": [ + 246, + 117, + 298, + 165 + ], + "lines": [ + { + "bbox": [ + 246, + 117, + 298, + 165 + ], + "spans": [ + { + "bbox": [ + 246, + 117, + 298, + 165 + ], + "type": "image", + "image_path": "ee7711eb2a5d834cf5ed6528fe0ac0d83ba72e8187011d979422b09c67bc5361.jpg" + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "image_body" + } + ], + "index": 4 + }, + { + "type": "image", + "bbox": [ + 312, + 117, + 360, + 165 + ], + "blocks": [ + { + "bbox": [ + 312, + 117, + 360, + 165 + ], + "lines": [ + { + "bbox": [ + 312, + 117, + 360, + 165 + ], + "spans": [ + { + "bbox": [ + 312, + 117, + 360, + 165 + ], + "type": "image", + "image_path": "d587f16990c6e14f0c0d356df49c131eb76bf54aad1e2fe7cbcf6f1104420253.jpg" + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 315, + 167, + 356, + 175 + ], + "lines": [ + { + "bbox": [ + 315, + 167, + 356, + 175 + ], + "spans": [ + { + "bbox": [ + 315, + 167, + 356, + 175 + ], + "type": "text", + "content": "(b): unhealthy" + } + ] + } + ], + "index": 6, + "angle": 0, + "type": "image_caption" + } + ], + "index": 5 + }, + { + "type": "image", + "bbox": [ + 373, + 117, + 434, + 165 + ], + "blocks": [ + { + "bbox": [ + 373, + 117, + 434, + 165 + ], + "lines": [ + { + "bbox": [ + 373, + 117, + 434, + 165 + ], + "spans": [ + { + "bbox": [ + 373, + 117, + 434, + 165 + ], + "type": "image", + "image_path": "6c094c81bda8c2ebf78c1b9ceeeb225096f2034298407262f7dd3159b417e666.jpg" + } + ] + } + ], + "index": 7, + "angle": 0, + "type": "image_body" + } + ], + "index": 7 + }, + { + "type": "image", + "bbox": [ + 150, + 224, + 462, + 322 + ], + "blocks": [ + { + "bbox": [ + 131, + 190, + 480, + 212 + ], + "lines": [ + { + "bbox": [ + 131, + 190, + 480, + 212 + ], + "spans": [ + { + "bbox": [ + 131, + 190, + 480, + 212 + ], + "type": "text", + "content": "Fig. 2: Data example. (a) is a healthy goal and (b) is a unhealthy goal. The PDT dataset takes (b) as the category." + } + ] + } + ], + "index": 8, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 150, + 224, + 462, + 322 + ], + "lines": [ + { + "bbox": [ + 150, + 224, + 462, + 322 + ], + "spans": [ + { + "bbox": [ + 150, + 224, + 462, + 322 + ], + "type": "image", + "image_path": "c3450a34a4019f05335b2fd838608c5690e342ce7305db900ff3b4b8d7aae41c.jpg" + } + ] + } + ], + "index": 9, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 130, + 335, + 480, + 378 + ], + "lines": [ + { + "bbox": [ + 130, + 335, + 480, + 378 + ], + "spans": [ + { + "bbox": [ + 130, + 335, + 480, + 378 + ], + "type": "text", + "content": "Fig. 3: PDT dataset generation and detection process. (a) represents the sliding window method, and (b) represents the \"Human-in-the-loop\" data annotation method. (c) means that a LL image is sent to the neural network for training, and LL and LH dual-resolution images are detected at the same time." + } + ] + } + ], + "index": 10, + "angle": 0, + "type": "image_caption" + } + ], + "index": 9 + }, + { + "bbox": [ + 130, + 402, + 480, + 437 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 402, + 480, + 437 + ], + "spans": [ + { + "bbox": [ + 130, + 402, + 480, + 437 + ], + "type": "text", + "content": "illustrates a pine tree in a healthy state, while (b) displays three pine trees in a unhealthy state that have been infested by pests, exhibiting progressively severe symptoms from left to right." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 131, + 455, + 330, + 467 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 131, + 455, + 330, + 467 + ], + "spans": [ + { + "bbox": [ + 131, + 455, + 330, + 467 + ], + "type": "text", + "content": "3.2 Data Preprocessing and Statistics" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 130, + 474, + 481, + 605 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 474, + 481, + 605 + ], + "spans": [ + { + "bbox": [ + 130, + 474, + 481, + 605 + ], + "type": "text", + "content": "This section delves into the acquisition, processing, and production procedures of the original image data about the PDT dataset. We set up an isometric photo mode at an altitude of 150 meters to acquire 105 sets of raw image data and 3D point cloud data. Owing to the substantial size of the original images, they are unsuitable for training with the current mainstream object detection neural networks. We introduce a training approach that applies to the majority of neural network data. As illustrated in Fig. 3, we employ a methodology that encompasses \"data preprocessing - 'Human-in-the-loop' data labeling - manual correction\". The essence of this method is to leverage the preliminary analysis capabilities of artificial intelligence to aid human annotators, thereby enhancing the efficiency and accuracy of data annotation." + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 130, + 605, + 481, + 665 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 605, + 481, + 665 + ], + "spans": [ + { + "bbox": [ + 130, + 605, + 481, + 665 + ], + "type": "text", + "content": "Data Preprocessing. Fig. 3 (a) demonstrates the application of the sliding window technique on the original image data, which allows for the extraction of a usable, standard-sized image through regional cropping of the high-resolution image. To prevent information loss during the sliding window process, the window dimensions are set to " + }, + { + "bbox": [ + 130, + 605, + 481, + 665 + ], + "type": "inline_equation", + "content": "640 \\times 640" + }, + { + "bbox": [ + 130, + 605, + 481, + 665 + ], + "type": "text", + "content": " pixels, with a step size of " + }, + { + "bbox": [ + 130, + 605, + 481, + 665 + ], + "type": "inline_equation", + "content": "630 \\times 630" + }, + { + "bbox": [ + 130, + 605, + 481, + 665 + ], + "type": "text", + "content": " pixels." + } + ] + } + ], + "index": 14 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 391, + 91, + 447, + 100 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 391, + 91, + 447, + 100 + ], + "spans": [ + { + "bbox": [ + 391, + 91, + 447, + 100 + ], + "type": "text", + "content": "PDT Dataset" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 474, + 92, + 480, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 474, + 92, + 480, + 100 + ], + "spans": [ + { + "bbox": [ + 474, + 92, + 480, + 100 + ], + "type": "text", + "content": "5" + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 4 + }, + { + "para_blocks": [ + { + "type": "table", + "bbox": [ + 168, + 145, + 449, + 216 + ], + "blocks": [ + { + "bbox": [ + 201, + 114, + 412, + 126 + ], + "lines": [ + { + "bbox": [ + 201, + 114, + 412, + 126 + ], + "spans": [ + { + "bbox": [ + 201, + 114, + 412, + 126 + ], + "type": "text", + "content": "Table 1: Structure of PDT dataset and its sample." + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 168, + 145, + 449, + 216 + ], + "lines": [ + { + "bbox": [ + 168, + 145, + 449, + 216 + ], + "spans": [ + { + "bbox": [ + 168, + 145, + 449, + 216 + ], + "type": "table", + "html": "
EditionClassesStructureTargeted imagesUntargeted imagesImage sizeInstancesTarget Amount
S(Small)M(Medium)L(Large)
SampleunhealthyTrain811640×64025691896548179
Val1969152813825
LLunhealthyTrain316613709029070418163423530
Val395172640×6401252399262165432
Test3901771149489492095450
LHunhealthy-10505472×3648934749347400
", + "image_path": "f4064f094581d8868ecd333df2773089d9c97fa7abadbfb9794325f598a64bbd.jpg" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "table_body" + } + ], + "index": 3 + }, + { + "bbox": [ + 130, + 247, + 481, + 331 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 247, + 481, + 331 + ], + "spans": [ + { + "bbox": [ + 130, + 247, + 481, + 331 + ], + "type": "text", + "content": "Following the sliding window operation, 53 LL images were derived from a single LH image. Furthermore, due to the high similarity between the unhealthy pine tree depicted in Fig. 2 and the ground background, and to enhance the effectiveness of neural network training, we opt to retain images that do not contain targets. Upon generating an output from these data during training, a corresponding loss is produced, which aids the neural network in learning and reduces the rate of false positives." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 130, + 332, + 482, + 488 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 332, + 482, + 488 + ], + "spans": [ + { + "bbox": [ + 130, + 332, + 482, + 488 + ], + "type": "text", + "content": "Data Annotation. We proceed to annotate the preprocessed data. To ensure the data's validity, we utilize the widely-used Labelme " + }, + { + "bbox": [ + 130, + 332, + 482, + 488 + ], + "type": "inline_equation", + "content": "{}^{1}" + }, + { + "bbox": [ + 130, + 332, + 482, + 488 + ], + "type": "text", + "content": " data labeling software. For high-resolution datasets with dense, small targets, the sheer volume of targets per image makes manual annotation a tedious and time-intensive process. As illustrated in Fig. 3 (b), the \"Human-in-the-loop\" data annotation approach effectively addresses this challenge. Initially, we acquire the LL version samples of the PDT dataset through manual annotation. The structure of the sample dataset is depicted in Tab. 1, which comprises a training set and a validation set. Subsequently, the sample dataset is input into the YOLOv5s model for training over 300 epochs, with the optimal weights being saved. To ensure that no label data is missed, a confidence threshold of 0.25 and an NMS IOU (Intersection over Union) threshold of 0.45 were set to automatically label the LL data and obtain the original label set." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 130, + 488, + 482, + 583 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 488, + 482, + 583 + ], + "spans": [ + { + "bbox": [ + 130, + 488, + 482, + 583 + ], + "type": "text", + "content": "Manual Check. The labeled dataset underwent manual filtering to eliminate incorrect annotations, rectify erroneous ones, and add any omitted annotations. The LL datasets were generated in two widely recognized formats, YOLO txt and VOCXml, by automatically partitioning the data structure into an 8:1:1 ratio using scripts. For the LH data, we input the completed LL data into the YOLO-DP model for training to acquire the optimal weights. The \"Human-in-the-loop\" data annotation and subsequent manual verification were conducted on the LH data to obtain the final LH version dataset." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 130, + 584, + 482, + 645 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 584, + 482, + 645 + ], + "spans": [ + { + "bbox": [ + 130, + 584, + 482, + 645 + ], + "type": "text", + "content": "Data Statistics. The statistics of the PDT dataset are shown in Tab. 1. LL and LH versions consist of 5,670 and 105 images, respectively, with 114,307 and 93,474 samples, and feature a single class labeled as unhealthy. The training, validation, and test sets of the LL version include 1,370, 172, and 144 target-free images, respectively. On average, the LL and LH versions contain 29 and 890" + } + ] + } + ], + "index": 7 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "spans": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "type": "text", + "content": "6" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 166, + 91, + 226, + 101 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 166, + 91, + 226, + 101 + ], + "spans": [ + { + "bbox": [ + 166, + 91, + 226, + 101 + ], + "type": "text", + "content": "M. Zhou et al." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 124, + 653, + 292, + 665 + ], + "type": "page_footnote", + "angle": 0, + "lines": [ + { + "bbox": [ + 124, + 653, + 292, + 665 + ], + "spans": [ + { + "bbox": [ + 124, + 653, + 292, + 665 + ], + "type": "text", + "content": "1 https://github.com labelmeai/labelme" + } + ] + } + ], + "index": 8 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 5 + }, + { + "para_blocks": [ + { + "type": "table", + "bbox": [ + 169, + 145, + 447, + 209 + ], + "blocks": [ + { + "bbox": [ + 240, + 114, + 372, + 125 + ], + "lines": [ + { + "bbox": [ + 240, + 114, + 372, + 125 + ], + "spans": [ + { + "bbox": [ + 240, + 114, + 372, + 125 + ], + "type": "text", + "content": "Table 2: CWC dataset sources." + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 169, + 145, + 447, + 209 + ], + "lines": [ + { + "bbox": [ + 169, + 145, + 447, + 209 + ], + "spans": [ + { + "bbox": [ + 169, + 145, + 447, + 209 + ], + "type": "table", + "html": "
DatasetsCorn weed datasets [25]lettuce weed radish weed datasets [25] datasets [25]Fresh-weed-data [18]
Classesbluegrass, corn, sedge, chenopodium album, circium_setosumlettuceradishnightshadetomatocottonvelvet
Number2502002011151162438
Image Size800×600800×600800×600800×600800×600586×444643×500
", + "image_path": "2c20a19bc947c83bcb5b5bbe0004cf2654986182c548df647ba3afdeef82c8b3.jpg" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "table_body" + } + ], + "index": 3 + }, + { + "bbox": [ + 130, + 241, + 479, + 266 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 241, + 479, + 266 + ], + "spans": [ + { + "bbox": [ + 130, + 241, + 479, + 266 + ], + "type": "text", + "content": "samples per image, respectively, indicating that the PDT dataset is suitable for the task of high-density and high-precision inspection of plant protection drones." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 132, + 285, + 238, + 298 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 285, + 238, + 298 + ], + "spans": [ + { + "bbox": [ + 132, + 285, + 238, + 298 + ], + "type": "text", + "content": "4 CWC Dataset" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 130, + 311, + 482, + 443 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 311, + 482, + 443 + ], + "spans": [ + { + "bbox": [ + 130, + 311, + 482, + 443 + ], + "type": "text", + "content": "We compile the available data to construct the Common Weed and Crop datasets. In special cases, there may be crop crossovers (usually in small trial plots) where weeds and crop targets exhibit multi-class structural characteristics and similarities. The majority of existing public datasets are limited to single or double classes, which do not suffice for the training and classification capabilities required by models. Consequently, when training the neural network, it is essential to supply detailed texture data that differentiates between various objects, thereby equipping the model with robust classification capabilities. This enables the model to effectively sift through and eliminate highly similar negative samples during detection, thereby enhancing the model's detection accuracy. This section discusses the acquisition and processing of CWC original image data." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 132, + 461, + 228, + 472 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 461, + 228, + 472 + ], + "spans": [ + { + "bbox": [ + 132, + 461, + 228, + 472 + ], + "type": "text", + "content": "4.1 Data Sources" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 130, + 482, + 482, + 555 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 482, + 482, + 555 + ], + "spans": [ + { + "bbox": [ + 130, + 482, + 482, + 555 + ], + "type": "text", + "content": "In Tab. 2, we present the source, categories, initial quantities of the CWC raw data, as well as their respective resolutions. As depicted in Fig. 1 (b), the CWC original data comprises a total of 11 categories, characterized by similar size, texture, and color. Please note that we manually annotated the raw Corn weed datasets, Lettuce weed datasets, and Radish weed datasets, while the Fresh-weed-data was already pre-labeled and available in YOLO txt format." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 132, + 572, + 331, + 585 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 572, + 331, + 585 + ], + "spans": [ + { + "bbox": [ + 132, + 572, + 331, + 585 + ], + "type": "text", + "content": "4.2 Data Preprocessing and Statistics" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 130, + 594, + 482, + 666 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 594, + 482, + 666 + ], + "spans": [ + { + "bbox": [ + 130, + 594, + 482, + 666 + ], + "type": "text", + "content": "Data Preprocessing. Owing to the imbalance between positive and negative samples in the original CWC data, there is a risk that the model may become biased towards the majority class, leading to a diminished recognition capability for the minority class. We used the oversampling strategy, using the data augmentation method to equalize the number of samples. The methods include random rotation, random translation, random brightness change, random noise" + } + ] + } + ], + "index": 10 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 391, + 91, + 447, + 100 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 391, + 91, + 447, + 100 + ], + "spans": [ + { + "bbox": [ + 391, + 91, + 447, + 100 + ], + "type": "text", + "content": "PDT Dataset" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 474, + 91, + 480, + 99 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 474, + 91, + 480, + 99 + ], + "spans": [ + { + "bbox": [ + 474, + 91, + 480, + 99 + ], + "type": "text", + "content": "7" + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 6 + }, + { + "para_blocks": [ + { + "type": "table", + "bbox": [ + 151, + 145, + 465, + 222 + ], + "blocks": [ + { + "bbox": [ + 236, + 114, + 376, + 125 + ], + "lines": [ + { + "bbox": [ + 236, + 114, + 376, + 125 + ], + "spans": [ + { + "bbox": [ + 236, + 114, + 376, + 125 + ], + "type": "text", + "content": "Table 3: CWC dataset structure." + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 151, + 145, + 465, + 222 + ], + "lines": [ + { + "bbox": [ + 151, + 145, + 465, + 222 + ], + "spans": [ + { + "bbox": [ + 151, + 145, + 465, + 222 + ], + "type": "table", + "html": "
Classesbluegrasschenopodium-album-cirsium-setosumcornsedgelettuceradishnightshadetomatocottonvelvet
Targeted ImagesTrain200200200200200200200200200200200200
Val404040404040404040404040
Test101010101010101010101010
Targeted AmountS100500000003
M000900000000
L249250250236250444326250210268248
Image Size800×600800×600800×600800×600800×600800×600800×600800×600800×600586×444643×500
", + "image_path": "25c882fbb2aa0801f7b2a8179d116f69c7070878f2e2943ba7ee05d7608782a4.jpg" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "table_body" + } + ], + "index": 3 + }, + { + "bbox": [ + 130, + 254, + 479, + 289 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 254, + 479, + 289 + ], + "spans": [ + { + "bbox": [ + 130, + 254, + 479, + 289 + ], + "type": "text", + "content": "addition, flipping and cutout. This enhances the robustness of the dataset, simulating the illumination formula: " + }, + { + "bbox": [ + 130, + 254, + 479, + 289 + ], + "type": "inline_equation", + "content": "\\pmb{I}_{out} = \\pmb{I}_{in} \\times w + [255] \\times (1 - w)" + }, + { + "bbox": [ + 130, + 254, + 479, + 289 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 130, + 254, + 479, + 289 + ], + "type": "inline_equation", + "content": "w \\in [0.35, 1]" + }, + { + "bbox": [ + 130, + 254, + 479, + 289 + ], + "type": "text", + "content": ". Where " + }, + { + "bbox": [ + 130, + 254, + 479, + 289 + ], + "type": "inline_equation", + "content": "\\pmb{I}_{out}" + }, + { + "bbox": [ + 130, + 254, + 479, + 289 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 130, + 254, + 479, + 289 + ], + "type": "inline_equation", + "content": "\\pmb{I}_{in}" + }, + { + "bbox": [ + 130, + 254, + 479, + 289 + ], + "type": "text", + "content": " is the output and input images. " + }, + { + "bbox": [ + 130, + 254, + 479, + 289 + ], + "type": "inline_equation", + "content": "w" + }, + { + "bbox": [ + 130, + 254, + 479, + 289 + ], + "type": "text", + "content": " is the random weight." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 130, + 290, + 480, + 339 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 290, + 480, + 339 + ], + "spans": [ + { + "bbox": [ + 130, + 290, + 480, + 339 + ], + "type": "text", + "content": "Data Statistics. The statistics for the CWC dataset are presented in Tab. 3. The CWC dataset comprises 2,750 images and 2,599 data entries, with the majority featuring large-sized objects. Data enhancement makes the sample reach a balanced distribution, with an average image containing 1 to 2 objects." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 132, + 357, + 277, + 370 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 357, + 277, + 370 + ], + "spans": [ + { + "bbox": [ + 132, + 357, + 277, + 370 + ], + "type": "text", + "content": "5 Dataset Comparison" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 130, + 381, + 482, + 490 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 381, + 482, + 490 + ], + "spans": [ + { + "bbox": [ + 130, + 381, + 482, + 490 + ], + "type": "text", + "content": "Tab. 4 provides a thorough comparison of the datasets. The LH version of the PDT dataset is exceptionally clear, offering image quality that is 35 to 200 times superior to other publicly available datasets. The PDT dataset is characterized by density, small target and rich real environmental details, which accords with the typical environmental factors of high and low altitude UAV work, and is suitable for training the special UAV detection model for plant protection. Furthermore, the PDT dataset includes a 3D point cloud version. While not highlighted in this study, a 3D data version is planned for development and will be released to the public in the future." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 130, + 491, + 482, + 562 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 491, + 482, + 562 + ], + "spans": [ + { + "bbox": [ + 130, + 491, + 482, + 562 + ], + "type": "text", + "content": "On the other hand, the CWC dataset demonstrates outstanding classification performance, particularly in terms of category diversity, detailed texture information, and dataset authenticity. CWC dataset has a large number of categories, 2 to 5 times more than other publicly available datasets, which indicates that the CWC dataset is well suited for tasks involving high-precision classification of plant protection UAVs." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 132, + 580, + 261, + 593 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 580, + 261, + 593 + ], + "spans": [ + { + "bbox": [ + 132, + 580, + 261, + 593 + ], + "type": "text", + "content": "6 YOLO-DP Model" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 130, + 605, + 482, + 666 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 605, + 482, + 666 + ], + "spans": [ + { + "bbox": [ + 130, + 605, + 482, + 666 + ], + "type": "text", + "content": "To expedite research efforts in UAV detection models for intelligent agricultural plant protection, we have designed and proposed a model, YOLO-DP (Fig. 4), specifically dedicated to the detection of tree crop diseases and pests, building upon the YOLOv5s foundatin [16]. To ensure the scalability of the model, the Kmeans clustering algorithm is employed to determine the distribution of the" + } + ] + } + ], + "index": 10 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "spans": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "type": "text", + "content": "8" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 166, + 91, + 226, + 101 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 166, + 91, + 226, + 101 + ], + "spans": [ + { + "bbox": [ + 166, + 91, + 226, + 101 + ], + "type": "text", + "content": "M. Zhou et al." + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 7 + }, + { + "para_blocks": [ + { + "type": "table", + "bbox": [ + 151, + 145, + 465, + 319 + ], + "blocks": [ + { + "bbox": [ + 244, + 114, + 369, + 126 + ], + "lines": [ + { + "bbox": [ + 244, + 114, + 369, + 126 + ], + "spans": [ + { + "bbox": [ + 244, + 114, + 369, + 126 + ], + "type": "text", + "content": "Table 4: Dataset comparison." + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 151, + 145, + 465, + 319 + ], + "lines": [ + { + "bbox": [ + 151, + 145, + 465, + 319 + ], + "spans": [ + { + "bbox": [ + 151, + 145, + 465, + 319 + ], + "type": "table", + "html": "
DatasetResolutionClassesRich DetailsHigh DefinitionScaleDense TargetTarget ScaleUav CollectionNo Target Image3D Point Cloud DataAnnotation Quality
PDT dataset5472×36481-✓100+✓S✓✓✓“Human-in-the-loop”
640×6401--5K+✓S/M/L✓✓✓
CWC dataset800×60011✓-2K+-L---Manually
SugarBeet2017 [11]1296×9662--5K+-S/M--✓Manually
Multispectral Potato Plants Images (MPP) [5]750×7502--1K+✓L✓--Manually
Ribworth Dataset (RI) [4]320×2401✓-5K+-M/L---Manually
320×320
Corn weed Dataset [25]800×6005✓-6K+-L---No
lettuce weed Dataset [25]800×6002✓-700+-L---No
radish weed Dataset [25]800×6002✓-400+-L---No
Fresh -weed-data [29]800×6004✓-200+-L---Manually
586×444
643×500
crop and weed detection data (CAW) [10]512×5122✓-1K+-L---Manually
Weeds [7]480×4801--600+-S/M✓--Manually
", + "image_path": "ba80e14297bccb544bcbcc3c2542f2626cef5097b1203f37c6a1c003c7a09e47.jpg" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "table_body" + } + ], + "index": 3 + }, + { + "type": "image", + "bbox": [ + 151, + 342, + 463, + 424 + ], + "blocks": [ + { + "bbox": [ + 151, + 342, + 463, + 424 + ], + "lines": [ + { + "bbox": [ + 151, + 342, + 463, + 424 + ], + "spans": [ + { + "bbox": [ + 151, + 342, + 463, + 424 + ], + "type": "image", + "image_path": "7ef984eff2dea40bd293b288036c7be84dffa1767f08d087465ad7e40587b066.jpg" + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 130, + 434, + 482, + 469 + ], + "lines": [ + { + "bbox": [ + 130, + 434, + 482, + 469 + ], + "spans": [ + { + "bbox": [ + 130, + 434, + 482, + 469 + ], + "type": "text", + "content": "Fig. 4: YOLO-DP baseline model architecture. The FPN [21] + PAN [23] module consists of GhostConv [12], Upsample, Concat, and C3. C stands for Concat, S for Sigmoid, P for channel number dilation, " + }, + { + "bbox": [ + 130, + 434, + 482, + 469 + ], + "type": "inline_equation", + "content": "\\times" + }, + { + "bbox": [ + 130, + 434, + 482, + 469 + ], + "type": "text", + "content": " for matrix multiplication, and + for matrix addition." + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "image_caption" + } + ], + "index": 4 + }, + { + "bbox": [ + 130, + 495, + 482, + 639 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 495, + 482, + 639 + ], + "spans": [ + { + "bbox": [ + 130, + 495, + 482, + 639 + ], + "type": "text", + "content": "actual bounding box data, allowing for the dynamic adjustment of the anchor box sizes. In light of the UAV's large-scale detection capabilities, an adaptive Large Scale Selective Kernel is incorporated into the Backbone network to capture the location information of dense, small target pest-infested trees [20]. Taking into account the efficient detection mode of UAVs, GhostConv is utilized in the Neck network to reduce the model size and computational complexity. Concurrently, the receptive field is enlarged to capture more texture feature information, thereby enhancing the model's classification capabilities. A version with decoupled detection heads is provided to minimize the interference between classification and regression losses, allowing the model to focus on specific tasks, improving the accuracy and accuracy of the detection, and also improving the model's generalization ability." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 130, + 641, + 482, + 666 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 641, + 482, + 666 + ], + "spans": [ + { + "bbox": [ + 130, + 641, + 482, + 666 + ], + "type": "text", + "content": "Adaptive Large Scale Selective Kernel. To tackle the demanding detection task involving a broad range of dense, small targets, we gather pest and" + } + ] + } + ], + "index": 7 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 391, + 91, + 447, + 100 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 391, + 91, + 447, + 100 + ], + "spans": [ + { + "bbox": [ + 391, + 91, + 447, + 100 + ], + "type": "text", + "content": "PDT Dataset" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 474, + 91, + 481, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 474, + 91, + 481, + 100 + ], + "spans": [ + { + "bbox": [ + 474, + 91, + 481, + 100 + ], + "type": "text", + "content": "9" + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 8 + }, + { + "para_blocks": [ + { + "bbox": [ + 130, + 116, + 482, + 344 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 116, + 482, + 344 + ], + "spans": [ + { + "bbox": [ + 130, + 116, + 482, + 344 + ], + "type": "text", + "content": "disease target information by designing a Convolutional group with an extensive receptive field. Firstly, the matrix M obtains the shallow range information M1 by Conv with " + }, + { + "bbox": [ + 130, + 116, + 482, + 344 + ], + "type": "inline_equation", + "content": "\\text{Kernelsize}(k) = 5" + }, + { + "bbox": [ + 130, + 116, + 482, + 344 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 130, + 116, + 482, + 344 + ], + "type": "inline_equation", + "content": "\\text{Padding}(p) = 2" + }, + { + "bbox": [ + 130, + 116, + 482, + 344 + ], + "type": "text", + "content": ". The depth information is explored for M1, and the deep range information M2 is obtained using Conv with " + }, + { + "bbox": [ + 130, + 116, + 482, + 344 + ], + "type": "inline_equation", + "content": "k = 7" + }, + { + "bbox": [ + 130, + 116, + 482, + 344 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 130, + 116, + 482, + 344 + ], + "type": "inline_equation", + "content": "p = 9" + }, + { + "bbox": [ + 130, + 116, + 482, + 344 + ], + "type": "text", + "content": ". Feature selection was performed on shallow and depth range information, and M3 and M4 were obtained by grouping convolution on M1 and M2 using GhostConv with " + }, + { + "bbox": [ + 130, + 116, + 482, + 344 + ], + "type": "inline_equation", + "content": "k = 1" + }, + { + "bbox": [ + 130, + 116, + 482, + 344 + ], + "type": "text", + "content": ". The advantage of using GhostConv is that half of the range information is retained and the other half is processed to prevent excessive loss of information. Next, the spatial attention is calculated for M3 and M4 large-scale information. M3 and M4 are concatenated in the channel dimension to obtain M5, and M6 and M7 are obtained by averaging and maximizing the channel dimensions, respectively. M6 and M7 are concatenated in the channel dimension to obtain M8 with scale " + }, + { + "bbox": [ + 130, + 116, + 482, + 344 + ], + "type": "inline_equation", + "content": "b \\times 1 \\times w \\times h" + }, + { + "bbox": [ + 130, + 116, + 482, + 344 + ], + "type": "text", + "content": ", which is fed into Conv with " + }, + { + "bbox": [ + 130, + 116, + 482, + 344 + ], + "type": "inline_equation", + "content": "k = 7" + }, + { + "bbox": [ + 130, + 116, + 482, + 344 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 130, + 116, + 482, + 344 + ], + "type": "inline_equation", + "content": "p = 4" + }, + { + "bbox": [ + 130, + 116, + 482, + 344 + ], + "type": "text", + "content": " for range attention collection to obtain M9. After Sigmoid activation, the channel dimension is expanded to obtain M10 and M11. Matrix multiplication is performed with M3 and M4 to obtain M12 and M13, respectively. After bitwise addition, the final spatial attention matrix M14 is obtained after Conv (" + }, + { + "bbox": [ + 130, + 116, + 482, + 344 + ], + "type": "inline_equation", + "content": "k = 1" + }, + { + "bbox": [ + 130, + 116, + 482, + 344 + ], + "type": "text", + "content": "). Finally, the output is obtained by matrix multiplication with the input matrix." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 132, + 361, + 224, + 374 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 361, + 224, + 374 + ], + "spans": [ + { + "bbox": [ + 132, + 361, + 224, + 374 + ], + "type": "text", + "content": "7 Experiment" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 132, + 384, + 287, + 396 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 384, + 287, + 396 + ], + "spans": [ + { + "bbox": [ + 132, + 384, + 287, + 396 + ], + "type": "text", + "content": "7.1 Experimental Conditions" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 130, + 403, + 481, + 486 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 403, + 481, + 486 + ], + "spans": [ + { + "bbox": [ + 130, + 403, + 481, + 486 + ], + "type": "text", + "content": "The experimental environment of this paper is ubuntu 18.04, Python3.9, PyTorch 1.9.1 framework, CUDA 11.4 version, and cuDNN 8.0 version. This paper's model training and inference are performed on NVIDIA A100-SXM4 with 40GB GPU memory and 16 GB CPU memory on the experimental platform. The training metrics in this paper are: 300 epochs of iterative training, the early stopping mechanism is set to 100 rounds, the initial learning rate is 0.01, the weight decay of 0.001 and momentum of " + }, + { + "bbox": [ + 130, + 403, + 481, + 486 + ], + "type": "inline_equation", + "content": "0.9\\%" + }, + { + "bbox": [ + 130, + 403, + 481, + 486 + ], + "type": "text", + "content": " are used, the Batch-size is 16." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 130, + 487, + 481, + 606 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 487, + 481, + 606 + ], + "spans": [ + { + "bbox": [ + 130, + 487, + 481, + 606 + ], + "type": "text", + "content": "Experiments were carried out on the PDT dataset and CWC dataset presented in this paper, as well as on the public dataset SugarBeet2017. We test the performance of the model YOLO-DP according to the characteristics of the dataset. First, the PDT dataset is used to test the detection ability of the model for dense small targets in the UAV perspective. Secondly, the ability of the model to extract fine-grained texture information is measured on the CWC dataset to verify its classification ability. It is reflected in the P metrics. Finally, the detection accuracy was measured on the SugarBeet2017 dataset. It is reflected in the R metrics. In order not to lose the image information, the input sizes are " + }, + { + "bbox": [ + 130, + 487, + 481, + 606 + ], + "type": "inline_equation", + "content": "640 \\times 640" + }, + { + "bbox": [ + 130, + 487, + 481, + 606 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 130, + 487, + 481, + 606 + ], + "type": "inline_equation", + "content": "800 \\times 800" + }, + { + "bbox": [ + 130, + 487, + 481, + 606 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 130, + 487, + 481, + 606 + ], + "type": "inline_equation", + "content": "1296 \\times 1296" + }, + { + "bbox": [ + 130, + 487, + 481, + 606 + ], + "type": "text", + "content": " respectively." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 132, + 623, + 275, + 635 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 623, + 275, + 635 + ], + "spans": [ + { + "bbox": [ + 132, + 623, + 275, + 635 + ], + "type": "text", + "content": "7.2 Experimental Analysis" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 130, + 641, + 481, + 665 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 641, + 481, + 665 + ], + "spans": [ + { + "bbox": [ + 130, + 641, + 481, + 665 + ], + "type": "text", + "content": "Dataset Validation. We evaluated 6 detection models on seven datasets (2 of ours and 5 public datasets mentioned in the paper). Based on the dataset's" + } + ] + } + ], + "index": 8 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "spans": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "type": "text", + "content": "10" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 166, + 91, + 225, + 100 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 166, + 91, + 225, + 100 + ], + "spans": [ + { + "bbox": [ + 166, + 91, + 225, + 100 + ], + "type": "text", + "content": "M. Zhou et al." + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 9 + }, + { + "para_blocks": [ + { + "bbox": [ + 130, + 116, + 482, + 213 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 116, + 482, + 213 + ], + "spans": [ + { + "bbox": [ + 130, + 116, + 482, + 213 + ], + "type": "text", + "content": "characteristics, we choose different metrics for the ranking model (Rank (Metrics)). Specifically, dense target datasets collected in natural environments use the comprehensive F1 metrics, datasets with target size changes use the R metrics, and multiclass datasets use the P metrics. We sort models with the same metrics score again using Gflops. Our model has excellent performance on all 7 datasets, and the performance of the proposed 2 datasets differs from that of existing datasets, indicating that our dataset can provide unique data distribution to the research field (Tab. 5)." + } + ] + } + ], + "index": 2 + }, + { + "type": "table", + "bbox": [ + 187, + 255, + 430, + 306 + ], + "blocks": [ + { + "bbox": [ + 237, + 235, + 376, + 246 + ], + "lines": [ + { + "bbox": [ + 237, + 235, + 376, + 246 + ], + "spans": [ + { + "bbox": [ + 237, + 235, + 376, + 246 + ], + "type": "text", + "content": "Table 5: Datasets effect ranking." + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 187, + 255, + 430, + 306 + ], + "lines": [ + { + "bbox": [ + 187, + 255, + 430, + 306 + ], + "spans": [ + { + "bbox": [ + 187, + 255, + 430, + 306 + ], + "type": "table", + "html": "
MethodGflopsRank (F1)Rank (P)Rank (R)
PDT (LL) (our)MPP [5]CWC (our)Weeds [7]SugarBeet2017 [11]CAW [10]RI [4]
YOLO-DP11.71 (0.89)1 (0.42)2 (92.9%)2 (76.5%)1 (73.8%)2 (89.4%)1 (82.3%)
YOLOv3155.35 (0.88)5 (0.38)6 (86.6%)1 (83.0%)6 (46.2%)6 (73.1%)6 (74.0%)
YOLOv4s20.83 (0.88)2 (0.42)5 (87.3%)3 (75.6%)3 (60.3%)3 (86.5%)2 (81.7%)
YOLOv5s16.02 (0.89)3 (0.38)4 (88.6%)4 (75.3%)4 (58.3%)5 (82.5%)3 (81.5%)
YOLOv7105.16 (0.85)6 (0.24)1 (93.1%)5 (74.4%)5 (48.1%)4 (83.3%)4 (80.3%)
YOLOv8s28.64 (0.88)4 (0.38)3 (92.0%)6 (70.4%)2 (65.0%)1 (90.1%)5 (79.3%)
", + "image_path": "a2501b3721929a165d350bb084ca6fa6cf6d0fda1883aa19cd1a1cb89fb22b5f.jpg" + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "table_body" + } + ], + "index": 4 + }, + { + "bbox": [ + 130, + 329, + 482, + 400 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 329, + 482, + 400 + ], + "spans": [ + { + "bbox": [ + 130, + 329, + 482, + 400 + ], + "type": "text", + "content": "Comparative Experiment. We perform a comparative analysis of single-stage detectors including the full YOLO family, SSD, EfficientDet, and RetinaNet, the two-stage detector Fast-RCNN, anchor-free based CenterNet, and WeedNet-R, a specialized model for weed detection. Among them, some models use pre-trained weights. As some models do not provide the computation results for certain metrics, these are denoted by “-” in the table." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 130, + 401, + 482, + 593 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 401, + 482, + 593 + ], + "spans": [ + { + "bbox": [ + 130, + 401, + 482, + 593 + ], + "type": "text", + "content": "In Tab. 6, we present the comparative evaluations of the proposed YOLO-DP baseline model alongside other models across the three datasets. The YOLO-DP model does not need to load pre-trained weights, and the detection speed is significantly ahead of other models. Under this premise, YOLO-DP demonstrates the most outstanding overall performance on the PDT dataset, with its F1 composite metrics surpassing all other models. This shows that the YOLO-DP model can well adapt to the detection scenario of UAV with large-scale, high-precision, dense and small targets. On the CWC dataset, its P metrics is at the first-class level, which is " + }, + { + "bbox": [ + 130, + 401, + 482, + 593 + ], + "type": "inline_equation", + "content": "4.3\\%" + }, + { + "bbox": [ + 130, + 401, + 482, + 593 + ], + "type": "text", + "content": " higher than the YOLOv5s baseline model. This demonstrates that the YOLO-DP model possesses an effective capability for extracting fine-grained texture information and is capable of fulfilling the classification task requirements within this domain. On the authoritative dataset SugarBeet2017, while the overall performance does not match that of models with loaded and trained weights, the R metrics of YOLO-DP is comparable to the pre-trained model's performance. This substantiates that the YOLO-DP model possesses precise detection capabilities and can accurately pinpoint the detection targets." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 130, + 593, + 482, + 628 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 593, + 482, + 628 + ], + "spans": [ + { + "bbox": [ + 130, + 593, + 482, + 628 + ], + "type": "text", + "content": "The comparative experimental results show that the YOLO-DP model is suitable for the application of plant protection UAV in precision agriculture detection fields such as weeds, pests and diseases, and crops." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 130, + 629, + 482, + 665 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 629, + 482, + 665 + ], + "spans": [ + { + "bbox": [ + 130, + 629, + 482, + 665 + ], + "type": "text", + "content": "Ablation Experiment. We chose the YOLOv5s that performed well on the PDT dataset as the benchmark. We use the popular attention mechanism to ablate the adaptive large-scale selection kernel to verify its detection performance" + } + ] + } + ], + "index": 8 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 391, + 91, + 447, + 100 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 391, + 91, + 447, + 100 + ], + "spans": [ + { + "bbox": [ + 391, + 91, + 447, + 100 + ], + "type": "text", + "content": "PDT Dataset" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 470, + 91, + 480, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 470, + 91, + 480, + 100 + ], + "spans": [ + { + "bbox": [ + 470, + 91, + 480, + 100 + ], + "type": "text", + "content": "11" + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 10 + }, + { + "para_blocks": [ + { + "type": "table", + "bbox": [ + 177, + 145, + 438, + 407 + ], + "blocks": [ + { + "bbox": [ + 234, + 114, + 379, + 126 + ], + "lines": [ + { + "bbox": [ + 234, + 114, + 379, + 126 + ], + "spans": [ + { + "bbox": [ + 234, + 114, + 379, + 126 + ], + "type": "text", + "content": "Table 6: Comparative experiment." + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 177, + 145, + 438, + 407 + ], + "lines": [ + { + "bbox": [ + 177, + 145, + 438, + 407 + ], + "spans": [ + { + "bbox": [ + 177, + 145, + 438, + 407 + ], + "type": "table", + "html": "
DatasetsApproachPRmAP@.5mAP@.5::.95F1GflopsParametersFPS Pre-training
PDT dataset (LL)SSD [24]84.5%87.7%85.1%-0.86273.623.7M37✓
EfficientDet [28]92.6%73.4%72.3%-0.8211.56.7M12✓
RetinaNet [22]93.3%65.3%64.2%-0.79109.732.7M32✓
CenterNet [32]95.2%67.4%66.5%-0.79109.732.7M32✓
Faster-RCNN [27]57.8%70.5%61.7%-0.64401.7136.7M13✓
YOLOv3 [26]88.5%88.1%93.4%65.7%0.88155.361.5M41-
YOLOv4s [2]88.8%88.2%94.7%66.1%0.8820.89.1M51-
YOLOv5s_7_0 [16]88.9%88.5%94.2%67.0%0.8916.07.0M93-
YOLOv6s [19]--91.4%63.2%-44.117.2M43-
YOLOv7 [30]87.4%82.6%90.1%55.5%0.85105.137.2M32-
YOLOv8s [17]88.7%87.5%94.0%67.9%0.8828.611.1M60-
WeedNet-R [11]87.7%48.1%70.4%-0.6219.025.6M0.5-
YOLO-DP (our)90.2%88.0%94.5%67.5%0.8911.75.2M109-
CWC datasetSSD [24]97.7%77.6%85.7%-0.91426.923.7M29✓
EfficientDet [28]97.2%98.6%98.6%-0.9011.56.7M13✓
RetinaNet [22]95.1%98.3%98.0%-0.97261.336.4M24✓
CenterNet [32]96.6%73.8%73.3%-0.80171.432.7M27✓
YOLOv3 [26]86.8%89.4%93.2%82.3%0.88154.761.5M30-
YOLOv4s [2]87.3%87.9%91.9%81.5%0.8820.89.1M43-
YOLOv5s_7_0 [16]88.6%88.7%93.0%81.2%0.8916.07.0M65-
YOLOv6s [19]--92.7%84.3%-68.917.2M31-
YOLOv7 [30]93.1%76.4%88.1%75.6%0.84105.137.2M21-
YOLOv8s [17]92.0%89.1%94.0%86.2%0.9128.611.1M38-
WeedNet-R [11]86.1%51.8%71.6%-0.6519.025.6M0.5-
YOLO-DP (our)92.9%87.5%91.8%81.0%0.9011.75.2M72-
Sugar-Beet2017SSD [24]85.0%83.6%79.3%-0.85112023.7M19✓
EfficientDet [28]93.3%79.8%77.8%-0.8611.56.7M16✓
RetinaNet [22]91.7%78.8%76.6%-0.84256.436.3M23✓
CenterNet [32]97.9%51.2%51.0%-0.62117.432.7M41✓
Faster-RCNN [27]63.6%87.4%80.0%-0.73546.9136.7M25✓
YOLOv3 [26]34.8%46.2%39.4%25.6%0.40155.361.5M28-
YOLOv4s [2]28.1%60.3%41.1%26.4%0.3820.89.1M28-
YOLOv5s_7_0 [16]25.0%58.3%40.6%26.7%0.3516.07.0M50-
YOLOv6s [19]--24.6%15.0%-185.217.2M49-
YOLOv7 [30]34.2%48.1%38.6%24.9%0.40105.137.2M18-
YOLOv8s [17]23.9%65.0%39.1%26.1%0.3528.611.1M33-
WeedNet-R [11]90.1%68.4%84.8%-0.7819.025.6M0.5-
YOLO-DP (our)23.1%73.8%38.3%25.0%0.3511.75.2M62-
", + "image_path": "5ee361efae815e1cad7c68eb32e7b983cea315df510350bb2b1a09573339eab8.jpg" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "table_body" + } + ], + "index": 3 + }, + { + "bbox": [ + 130, + 437, + 482, + 547 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 437, + 482, + 547 + ], + "spans": [ + { + "bbox": [ + 130, + 437, + 482, + 547 + ], + "type": "text", + "content": "on PDT dataset. The attention modules are designed on Backbone's floors 3, 5, 7, and 9. Since the number of C3x and C3TR parameters is too large, we only replace them at the 9th layer in order to carry out the experiment. \"v5s\" means YOLOv5s_7.0. \"v5s__our\" stands for the use of Adaptive Large Scale Selective Kernel in the YOLOv5s_7.0. It can be seen from the results in Tab. 7 that the performance of YOLOv5s is greatly improved after Large Scale Selective Kernel is used. The mAP@.5, mAP@.5::.95 and F1 are the best metrics. This indicates that the Large Scale Selective Kernel used in this paper is suitable for dense object detection tasks." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 132, + 563, + 276, + 574 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 563, + 276, + 574 + ], + "spans": [ + { + "bbox": [ + 132, + 563, + 276, + 574 + ], + "type": "text", + "content": "7.3 Visualization Research" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 130, + 581, + 482, + 666 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 581, + 482, + 666 + ], + "spans": [ + { + "bbox": [ + 130, + 581, + 482, + 666 + ], + "type": "text", + "content": "In Fig. 5, we show the detection visualization results on the PDT dataset. We have selected the YOLOv5s and YOLOv8s models, which exhibited superior overall performance in the comparative experiment, to demonstrate and contrast with the YOLO-DP model. (a) and (e) illustrate the Ground Truth for the LL and LH versions, respectively. Observations indicate that the \"Human-in-the-loop\" data annotation approach is practical for weed and pest detection data annotation. The detection results in (b) and (c) suggest that the LL version" + } + ] + } + ], + "index": 6 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "spans": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "type": "text", + "content": "12" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 166, + 91, + 225, + 100 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 166, + 91, + 225, + 100 + ], + "spans": [ + { + "bbox": [ + 166, + 91, + 225, + 100 + ], + "type": "text", + "content": "M. Zhou et al." + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 11 + }, + { + "para_blocks": [ + { + "type": "table", + "bbox": [ + 203, + 145, + 413, + 240 + ], + "blocks": [ + { + "bbox": [ + 242, + 114, + 370, + 125 + ], + "lines": [ + { + "bbox": [ + 242, + 114, + 370, + 125 + ], + "spans": [ + { + "bbox": [ + 242, + 114, + 370, + 125 + ], + "type": "text", + "content": "Table 7: Ablation experiment." + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 203, + 145, + 413, + 240 + ], + "lines": [ + { + "bbox": [ + 203, + 145, + 413, + 240 + ], + "spans": [ + { + "bbox": [ + 203, + 145, + 413, + 240 + ], + "type": "table", + "html": "
DatasetsApproachPRmAP@.5mAP@.5:95F1GflopsParameters
PDT dataset (LL)v5s_C188.2%88.5%93.9%67.1%0.8825.310.0M
v5s_C288.8%88.4%94.1%67.0%0.8817.27.3M
v5s_C2f88.6%88.5%93.8%67.1%0.8817.47.5M
v5s_C388.9%88.5%94.2%67.0%0.8916.07.0M
v5s_C3x88.7%81.5%87.4%63.2%0.8514.56.5M
v5s_C3TR88.3%89.0%94.1%67.1%0.8915.77.0M
v5s_C3Ghost88.9%88.2%94.2%66.7%0.8812.55.9M
v5s_SE88.8%88.3%94.4%66.6%0.8810.65.1M
v5s_CBAM89.8%87.5%94.4%66.5%0.8910.95.6M
v5s_GAM89.2%87.7%94.0%67.1%0.8816.47.5M
v5s_ECA89.7%87.0%94.3%66.2%0.8810.55.1M
v5s_our89.1%88.5%94.4%67.2%0.8912.26.1M
", + "image_path": "3d3c9ec97bdbf9a28d94cb28f7c6ee7a97598eb8334b1e7a8ab40ae226403984.jpg" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "table_body" + } + ], + "index": 3 + }, + { + "type": "image", + "bbox": [ + 142, + 265, + 207, + 330 + ], + "blocks": [ + { + "bbox": [ + 142, + 265, + 207, + 330 + ], + "lines": [ + { + "bbox": [ + 142, + 265, + 207, + 330 + ], + "spans": [ + { + "bbox": [ + 142, + 265, + 207, + 330 + ], + "type": "image", + "image_path": "ceafb2fd488380bbfcc29dd894e742119ccdd0771e8ed2c3cf8fc56452afe56a.jpg" + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 147, + 331, + 202, + 338 + ], + "lines": [ + { + "bbox": [ + 147, + 331, + 202, + 338 + ], + "spans": [ + { + "bbox": [ + 147, + 331, + 202, + 338 + ], + "type": "text", + "content": "(a): Ground truth (LL)" + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "image_caption" + } + ], + "index": 4 + }, + { + "type": "image", + "bbox": [ + 209, + 265, + 297, + 330 + ], + "blocks": [ + { + "bbox": [ + 209, + 265, + 297, + 330 + ], + "lines": [ + { + "bbox": [ + 209, + 265, + 297, + 330 + ], + "spans": [ + { + "bbox": [ + 209, + 265, + 297, + 330 + ], + "type": "image", + "image_path": "a041a01751df4760b48b9e8f4ba36bca4d1a15ec659587807ca7715750dd4908.jpg" + } + ] + } + ], + "index": 6, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 248, + 331, + 282, + 338 + ], + "lines": [ + { + "bbox": [ + 248, + 331, + 282, + 338 + ], + "spans": [ + { + "bbox": [ + 248, + 331, + 282, + 338 + ], + "type": "text", + "content": "(b):YOLOv5s" + } + ] + } + ], + "index": 7, + "angle": 0, + "type": "image_caption" + } + ], + "index": 6 + }, + { + "type": "image", + "bbox": [ + 298, + 266, + 402, + 330 + ], + "blocks": [ + { + "bbox": [ + 298, + 266, + 402, + 330 + ], + "lines": [ + { + "bbox": [ + 298, + 266, + 402, + 330 + ], + "spans": [ + { + "bbox": [ + 298, + 266, + 402, + 330 + ], + "type": "image", + "image_path": "06fd8ca766369c50b1ce147af41b00eab3a5a3180ac70164c0b7da0b845fdaae.jpg" + } + ] + } + ], + "index": 8, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 335, + 331, + 370, + 338 + ], + "lines": [ + { + "bbox": [ + 335, + 331, + 370, + 338 + ], + "spans": [ + { + "bbox": [ + 335, + 331, + 370, + 338 + ], + "type": "text", + "content": "(c):YOLOv8s" + } + ] + } + ], + "index": 9, + "angle": 0, + "type": "image_caption" + } + ], + "index": 8 + }, + { + "type": "image", + "bbox": [ + 406, + 266, + 472, + 330 + ], + "blocks": [ + { + "bbox": [ + 406, + 266, + 472, + 330 + ], + "lines": [ + { + "bbox": [ + 406, + 266, + 472, + 330 + ], + "spans": [ + { + "bbox": [ + 406, + 266, + 472, + 330 + ], + "type": "image", + "image_path": "739026f48194d4c770fe7097cea99b6aced284cd23277737612851d892eb8385.jpg" + } + ] + } + ], + "index": 10, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 422, + 331, + 457, + 336 + ], + "lines": [ + { + "bbox": [ + 422, + 331, + 457, + 336 + ], + "spans": [ + { + "bbox": [ + 422, + 331, + 457, + 336 + ], + "type": "text", + "content": "(d):YOLO-DP" + } + ] + } + ], + "index": 11, + "angle": 0, + "type": "image_caption" + } + ], + "index": 10 + }, + { + "type": "image", + "bbox": [ + 142, + 339, + 306, + 448 + ], + "blocks": [ + { + "bbox": [ + 142, + 339, + 306, + 448 + ], + "lines": [ + { + "bbox": [ + 142, + 339, + 306, + 448 + ], + "spans": [ + { + "bbox": [ + 142, + 339, + 306, + 448 + ], + "type": "image", + "image_path": "3e3c91b210d3821aa0e1f7a0143f1686de5175e051793e82af5bc03f5822a233.jpg" + } + ] + } + ], + "index": 12, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 196, + 449, + 252, + 456 + ], + "lines": [ + { + "bbox": [ + 196, + 449, + 252, + 456 + ], + "spans": [ + { + "bbox": [ + 196, + 449, + 252, + 456 + ], + "type": "text", + "content": "(e): Ground truth (LH)" + } + ] + } + ], + "index": 13, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 208, + 466, + 403, + 477 + ], + "lines": [ + { + "bbox": [ + 208, + 466, + 403, + 477 + ], + "spans": [ + { + "bbox": [ + 208, + 466, + 403, + 477 + ], + "type": "text", + "content": "Fig. 5: Visualization of PDT dataset detection." + } + ] + } + ], + "index": 16, + "angle": 0, + "type": "image_caption" + } + ], + "index": 12 + }, + { + "type": "image", + "bbox": [ + 307, + 339, + 471, + 448 + ], + "blocks": [ + { + "bbox": [ + 307, + 339, + 471, + 448 + ], + "lines": [ + { + "bbox": [ + 307, + 339, + 471, + 448 + ], + "spans": [ + { + "bbox": [ + 307, + 339, + 471, + 448 + ], + "type": "image", + "image_path": "e33321c0f7850da74d1c51d2d9fb75d56c7599a9367a12a18332ac327c07fa7c.jpg" + } + ] + } + ], + "index": 14, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 372, + 449, + 407, + 456 + ], + "lines": [ + { + "bbox": [ + 372, + 449, + 407, + 456 + ], + "spans": [ + { + "bbox": [ + 372, + 449, + 407, + 456 + ], + "type": "text", + "content": "(f):YOLO-DP" + } + ] + } + ], + "index": 15, + "angle": 0, + "type": "image_caption" + } + ], + "index": 14 + }, + { + "bbox": [ + 130, + 506, + 482, + 614 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 506, + 482, + 614 + ], + "spans": [ + { + "bbox": [ + 130, + 506, + 482, + 614 + ], + "type": "text", + "content": "of PDT dataset is compatible with widely used detection models, confirming its utility. When compared to (d), it is evident that YOLO-DP delivers outstanding detection performance on LL, with no instances of missing, false, or duplicate detections. Examining the LH version's detection results in (f), we find no missed detections, and the critical issue of current detectors' inability to distinguish dead pine trees from the ground has been overcome. This leads us to conclude that the YOLO-DP model is well-suited for large-scale, high-precision, small-target UAV detection tasks. Moreover, the detection strategy depicted in Fig. 3 of this paper has been validated as effective." + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 130, + 617, + 482, + 666 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 617, + 482, + 666 + ], + "spans": [ + { + "bbox": [ + 130, + 617, + 482, + 666 + ], + "type": "text", + "content": "In Fig. 6, we presents a comparison between YOLO-DP and the Confusion Matrix of the high-performing classification models YOLOv7 and YOLOv8s on the CWC dataset. It is evident that YOLO-DP's classification performance surpasses that of YOLOv7. Furthermore, YOLO-DP's classification capability is" + } + ] + } + ], + "index": 18 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 391, + 91, + 447, + 100 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 391, + 91, + 447, + 100 + ], + "spans": [ + { + "bbox": [ + 391, + 91, + 447, + 100 + ], + "type": "text", + "content": "PDT Dataset" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 470, + 91, + 481, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 470, + 91, + 481, + 100 + ], + "spans": [ + { + "bbox": [ + 470, + 91, + 481, + 100 + ], + "type": "text", + "content": "13" + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 12 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 133, + 113, + 481, + 266 + ], + "blocks": [ + { + "bbox": [ + 133, + 113, + 481, + 266 + ], + "lines": [ + { + "bbox": [ + 133, + 113, + 481, + 266 + ], + "spans": [ + { + "bbox": [ + 133, + 113, + 481, + 266 + ], + "type": "image", + "image_path": "81072e389a695241e9a2bbcfb447b15e823afacfdc8252a4a78a3ceed113b550.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 132, + 277, + 480, + 310 + ], + "lines": [ + { + "bbox": [ + 132, + 277, + 480, + 310 + ], + "spans": [ + { + "bbox": [ + 132, + 277, + 480, + 310 + ], + "type": "text", + "content": "Fig. 6: Confusion matrix. The rows represent the true class, the columns represent the predicted class, and the confidence value is [0,1]. The format of the matrix is: YOLOv7/YOLOv8s/YOLO-DP." + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_caption" + } + ], + "index": 2 + }, + { + "bbox": [ + 132, + 335, + 480, + 357 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 335, + 480, + 357 + ], + "spans": [ + { + "bbox": [ + 132, + 335, + 480, + 357 + ], + "type": "text", + "content": "comparable to YOLOv8s, especially when its overall performance exceeds that of YOLOv8s." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 133, + 378, + 218, + 391 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 378, + 218, + 391 + ], + "spans": [ + { + "bbox": [ + 133, + 378, + 218, + 391 + ], + "type": "text", + "content": "8 Conclusion" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 132, + 403, + 481, + 583 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 403, + 481, + 583 + ], + "spans": [ + { + "bbox": [ + 132, + 403, + 481, + 583 + ], + "type": "text", + "content": "In this study, we introduce the PDT dataset to address the challenging task of UAV pest detection, with the goal of advancing research in UAV detection tasks within the realm of precision agriculture. We also present the CWC dataset for agricultural weed and crop classification, which compensates for the field's deficiency in training model classification capabilities. To tackle the scarcity of baseline model research in this domain, we propose the YOLO-DP model for dense, small-target UAV pest detection and validate its efficacy across three datasets. It is noteworthy that the PDT dataset features a dual-resolution version, and the CWC dataset boasts 11 detailed, texture-similar plant classes. Moreover, we provide a comprehensive evaluation of both datasets, objectively assessing their value. This work has its limitations: the 3D point cloud version of the PDT dataset is not discussed, and there is a size inconsistency issue within the CWC dataset. We plan to address these issues in future endeavors. Lastly, to ensure the continuity of this research, we make the datasets and associated code designed in this paper available on our website." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 133, + 603, + 239, + 616 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 603, + 239, + 616 + ], + "spans": [ + { + "bbox": [ + 133, + 603, + 239, + 616 + ], + "type": "text", + "content": "Acknowledgments" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 132, + 629, + 480, + 665 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 629, + 480, + 665 + ], + "spans": [ + { + "bbox": [ + 132, + 629, + 480, + 665 + ], + "type": "text", + "content": "This work was supported by Key R&D Program of Shandong Province, China (2023CXGC010112), the Taishan Scholars Program (NO. tsqn202103097, NO. tscy20221110)." + } + ] + } + ], + "index": 8 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "spans": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "type": "text", + "content": "14" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 166, + 91, + 225, + 100 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 166, + 91, + 225, + 100 + ], + "spans": [ + { + "bbox": [ + 166, + 91, + 225, + 100 + ], + "type": "text", + "content": "M. Zhou et al." + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 13 + }, + { + "para_blocks": [ + { + "bbox": [ + 133, + 114, + 197, + 126 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 114, + 197, + 126 + ], + "spans": [ + { + "bbox": [ + 133, + 114, + 197, + 126 + ], + "type": "text", + "content": "References" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 138, + 140, + 480, + 665 + ], + "type": "list", + "angle": 0, + "index": 20, + "blocks": [ + { + "bbox": [ + 138, + 140, + 480, + 172 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 140, + 480, + 172 + ], + "spans": [ + { + "bbox": [ + 138, + 140, + 480, + 172 + ], + "type": "text", + "content": "1. Agarwal, R., Hariharan, S., Rao, M.N., Agarwal, A.: Weed identification using k-means clustering with color spaces features in multi-spectral images taken by UAV. In: IGARSS. pp. 7047-7050. IEEE (2021)" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 138, + 173, + 480, + 194 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 173, + 480, + 194 + ], + "spans": [ + { + "bbox": [ + 138, + 173, + 480, + 194 + ], + "type": "text", + "content": "2. Bochkovskiy, A., Wang, C., Liao, H.M.: YOLOv4: Optimal speed and accuracy of object detection. CoRR abs/2004.10934 (2020)" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 138, + 195, + 480, + 238 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 195, + 480, + 238 + ], + "spans": [ + { + "bbox": [ + 138, + 195, + 480, + 238 + ], + "type": "text", + "content": "3. Boursianis, A.D., Papadopoulou, M.S., Diamantoulakis, P.D., Liopa-Tsakalidi, A., Barouchas, P., Salahas, G., Karagiannidis, G.K., Wan, S., Goudos, S.K.: Internet of things (IoT) and agricultural unmanned aerial vehicles (UAVs) in smart farming: A comprehensive review. Internet Things 18, 100187 (2022)" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 138, + 239, + 480, + 270 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 239, + 480, + 270 + ], + "spans": [ + { + "bbox": [ + 138, + 239, + 480, + 270 + ], + "type": "text", + "content": "4. university of burgandy: Ribworth dataset. https://universe.roboflow.com/university-of-burgandy-zowkw/ribworth (nov 2022), https://universe.roboflow.com/university-of-burgandy-zowkw/ribworth" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 138, + 271, + 480, + 304 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 271, + 480, + 304 + ], + "spans": [ + { + "bbox": [ + 138, + 271, + 480, + 304 + ], + "type": "text", + "content": "5. Butte, S., Vakanski, A., Duellman, K., Wang, H., Mirkouei, A.: Potato crop stress identification in aerial images using deep learning-based object detection. CoRR abs/2106.07770 (2021)" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 138, + 304, + 480, + 347 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 304, + 480, + 347 + ], + "spans": [ + { + "bbox": [ + 138, + 304, + 480, + 347 + ], + "type": "text", + "content": "6. Caras, T., Lati, R.N., Holland, D., Dubinin, V.M., Hatib, K., Shulner, I., Keiesar, O., Liddor, G., Paz-Kagan, T.: Monitoring the effects of weed management strategies on tree canopy structure and growth using uav-lidar in a young almond orchard. Comput. Electron. Agric. 216, 108467 (2024)" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 138, + 348, + 480, + 358 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 348, + 480, + 358 + ], + "spans": [ + { + "bbox": [ + 138, + 348, + 480, + 358 + ], + "type": "text", + "content": "7. Carboni: weeds pytorch (2023), https://github.com/carboni123/weeds-pytorch" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 138, + 360, + 480, + 391 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 360, + 480, + 391 + ], + "spans": [ + { + "bbox": [ + 138, + 360, + 480, + 391 + ], + "type": "text", + "content": "8. Chebrolu, N., Lottes, P., Schaefer, A., Winterhalter, W., Burgard, W., Stachniss, C.: Agricultural robot dataset for plant classification, localization and mapping on sugar beet fields. Int. J. Robotics Res. 36(10), 1045-1052 (2017)" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 138, + 392, + 480, + 435 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 392, + 480, + 435 + ], + "spans": [ + { + "bbox": [ + 138, + 392, + 480, + 435 + ], + "type": "text", + "content": "9. Costa, L.S., Sano, E.E., Ferreira, M.E., Munhoz, C.B.R., Costa, J.V.S., Junior, L.R.A., de Mello, T.R.B., da Cunha Bustamante, M.M.: Woody plant encroachment in a seasonal tropical savanna: Lessons about classifiers and accuracy from UAV images. Remote. Sens. 15(9), 2342 (2023)" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 138, + 436, + 480, + 456 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 436, + 480, + 456 + ], + "spans": [ + { + "bbox": [ + 138, + 436, + 480, + 456 + ], + "type": "text", + "content": "0. Dabhi: Crop and weed detection (2021), https://github.com/ravirajsinh45/Crop_and_weed_detection" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 138, + 457, + 480, + 490 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 457, + 480, + 490 + ], + "spans": [ + { + "bbox": [ + 138, + 457, + 480, + 490 + ], + "type": "text", + "content": "1. Guo, Z., Goh, H.H., Li, X., Zhang, M., Li, Y.: WeedNet-R: a sugar beet field weed detection algorithm based on enhanced retinanet and context semantic fusion. Frontiers in Plant Science 14, 1226329 (2023)" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 138, + 491, + 480, + 522 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 491, + 480, + 522 + ], + "spans": [ + { + "bbox": [ + 138, + 491, + 480, + 522 + ], + "type": "text", + "content": "2. Han, K., Wang, Y., Tian, Q., Guo, J., Xu, C., Xu, C.: GhostNet: more features from cheap operations. In: CVPR. pp. 1577-1586. Computer Vision Foundation / IEEE (2020)" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 138, + 523, + 480, + 555 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 523, + 480, + 555 + ], + "spans": [ + { + "bbox": [ + 138, + 523, + 480, + 555 + ], + "type": "text", + "content": "3. Huang, J., Luo, Y., Quan, Q., Wang, B., Xue, X., Zhang, Y.: An autonomous task assignment and decision-making method for coverage path planning of multiple pesticide spraying UAVs. Comput. Electron. Agric. 212, 108128 (2023)" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 138, + 556, + 480, + 589 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 556, + 480, + 589 + ], + "spans": [ + { + "bbox": [ + 138, + 556, + 480, + 589 + ], + "type": "text", + "content": "4. Innani, S., Dutande, P., Baheti, B., Talbar, S.N., Baid, U.: Fuse-PN: A novel architecture for anomaly pattern segmentation in aerial agricultural images. In: CVPR Workshops. pp. 2960-2968. Computer Vision Foundation / IEEE (2021)" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 138, + 590, + 480, + 621 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 590, + 480, + 621 + ], + "spans": [ + { + "bbox": [ + 138, + 590, + 480, + 621 + ], + "type": "text", + "content": "5. Jabir, B., Falih, N.: Deep learning-based decision support system for weeds detection in wheat fields. International Journal of Electrical and Computer Engineering 12(1), 816 (2022)" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 138, + 622, + 480, + 643 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 622, + 480, + 643 + ], + "spans": [ + { + "bbox": [ + 138, + 622, + 480, + 643 + ], + "type": "text", + "content": "6. Jocher, G.: YOLOv5 by Ultralytics (May 2020). https://doi.org/10.5281/ zenodo.3908559, https://github.com/ultralytics/yolov5" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 138, + 644, + 480, + 665 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 644, + 480, + 665 + ], + "spans": [ + { + "bbox": [ + 138, + 644, + 480, + 665 + ], + "type": "text", + "content": "7. Jocher, G., Chaurasia, A., Qiu, J.: Ultralytics YOLO (Jan 2023), https://github.com/ultralytics/ultralytics" + } + ] + } + ], + "index": 19 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 391, + 91, + 447, + 100 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 391, + 91, + 447, + 100 + ], + "spans": [ + { + "bbox": [ + 391, + 91, + 447, + 100 + ], + "type": "text", + "content": "PDT Dataset" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 470, + 92, + 481, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 470, + 92, + 481, + 100 + ], + "spans": [ + { + "bbox": [ + 470, + 92, + 481, + 100 + ], + "type": "text", + "content": "15" + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 14 + }, + { + "para_blocks": [ + { + "bbox": [ + 132, + 116, + 481, + 578 + ], + "type": "list", + "angle": 0, + "index": 17, + "blocks": [ + { + "bbox": [ + 133, + 116, + 481, + 160 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 116, + 481, + 160 + ], + "spans": [ + { + "bbox": [ + 133, + 116, + 481, + 160 + ], + "type": "text", + "content": "18. Lameski, P., Zdravevski, E., Trajkovik, V., Kulakov, A.: Weed detection dataset with RGB images taken under variable light conditions. In: ICT Innovations. Communications in Computer and Information Science, vol. 778, pp. 112-119. Springer (2017)" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 132, + 161, + 481, + 205 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 161, + 481, + 205 + ], + "spans": [ + { + "bbox": [ + 132, + 161, + 481, + 205 + ], + "type": "text", + "content": "19. Li, C., Li, L., Jiang, H., Weng, K., Geng, Y., Li, L., Ke, Z., Li, Q., Cheng, M., Nie, W., Li, Y., Zhang, B., Liang, Y., Zhou, L., Xu, X., Chu, X., Wei, X., Wei, X.: YOLOv6: A single-stage object detection framework for industrial applications. CoRR abs/2209.02976 (2022)" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 132, + 205, + 481, + 237 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 205, + 481, + 237 + ], + "spans": [ + { + "bbox": [ + 132, + 205, + 481, + 237 + ], + "type": "text", + "content": "20. Li, Y., Hou, Q., Zheng, Z., Cheng, M., Yang, J., Li, X.: Large selective kernel network for remote sensing object detection. In: ICCV. pp. 16748-16759. IEEE (2023)" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 132, + 238, + 481, + 270 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 238, + 481, + 270 + ], + "spans": [ + { + "bbox": [ + 132, + 238, + 481, + 270 + ], + "type": "text", + "content": "21. Lin, T., Dólar, P., Girshick, R.B., He, K., Hariharan, B., Belongie, S.J.: Feature pyramid networks for object detection. In: CVPR. pp. 936-944. IEEE Computer Society (2017)" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 132, + 271, + 481, + 293 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 271, + 481, + 293 + ], + "spans": [ + { + "bbox": [ + 132, + 271, + 481, + 293 + ], + "type": "text", + "content": "22. Lin, T., Goyal, P., Girshick, R.B., He, K., Dollar, P.: Focal loss for dense object detection. In: ICCV. pp. 2999-3007. IEEE Computer Society (2017)" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 132, + 293, + 481, + 326 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 293, + 481, + 326 + ], + "spans": [ + { + "bbox": [ + 132, + 293, + 481, + 326 + ], + "type": "text", + "content": "23. Liu, S., Qi, L., Qin, H., Shi, J., Jia, J.: Path aggregation network for instance segmentation. In: CVPR. pp. 8759-8768. Computer Vision Foundation / IEEE Computer Society (2018)" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 132, + 326, + 481, + 358 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 326, + 481, + 358 + ], + "spans": [ + { + "bbox": [ + 132, + 326, + 481, + 358 + ], + "type": "text", + "content": "24. Liu, W., Anguelov, D., Erhan, D., Szegedy, C., Reed, S.E., Fu, C., Berg, A.C.: SSD: single shot multibox detector. In: ECCV (1). Lecture Notes in Computer Science, vol. 9905, pp. 21-37. Springer (2016)" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 132, + 359, + 481, + 379 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 359, + 481, + 379 + ], + "spans": [ + { + "bbox": [ + 132, + 359, + 481, + 379 + ], + "type": "text", + "content": "25. Monster: weed dataset (2019), https://gitee.com/Monster7/weed-database/tree/master/" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 132, + 380, + 481, + 402 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 380, + 481, + 402 + ], + "spans": [ + { + "bbox": [ + 132, + 380, + 481, + 402 + ], + "type": "text", + "content": "26. Redmon, J., Farhadi, A.: YOLOv3: An incremental improvement. CoRR abs/1804.02767 (2018)" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 132, + 403, + 481, + 435 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 403, + 481, + 435 + ], + "spans": [ + { + "bbox": [ + 132, + 403, + 481, + 435 + ], + "type": "text", + "content": "27. Ren, S., He, K., Girshick, R.B., Sun, J.: Faster R-CNN: towards real-time object detection with region proposal networks. IEEE Trans. Pattern Anal. Mach. Intell. 39(6), 1137-1149 (2017)" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 132, + 435, + 481, + 456 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 435, + 481, + 456 + ], + "spans": [ + { + "bbox": [ + 132, + 435, + 481, + 456 + ], + "type": "text", + "content": "28. Tan, M., Pang, R., Le, Q.V.: EfficientDet: scalable and efficient object detection. In: CVPR. pp. 10778-10787. Computer Vision Foundation / IEEE (2020)" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 132, + 457, + 481, + 490 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 457, + 481, + 490 + ], + "spans": [ + { + "bbox": [ + 132, + 457, + 481, + 490 + ], + "type": "text", + "content": "29. Tang, J., Wang, D., Zhang, Z., He, L., Xin, J., Xu, Y.: Weed identification based on k-means feature learning combined with convolutional neural network. Comput. Electron. Agric. 135, 63-70 (2017)" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 132, + 491, + 481, + 522 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 491, + 481, + 522 + ], + "spans": [ + { + "bbox": [ + 132, + 491, + 481, + 522 + ], + "type": "text", + "content": "30. Wang, C., Bochkovskiy, A., Liao, H.M.: YOLOv7: Trainable bag-of-freebies sets new state-of-the-art for real-time object detectors. In: CVPR. pp. 7464-7475. IEEE (2023)" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 132, + 523, + 481, + 555 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 523, + 481, + 555 + ], + "spans": [ + { + "bbox": [ + 132, + 523, + 481, + 555 + ], + "type": "text", + "content": "31. Zhang, R., Wang, C., Hu, X., Liu, Y., Chen, S., et al.: Weed location and recognition based on UAV imaging and deep learning. International Journal of Precision Agricultural Aviation 3(1) (2020)" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 132, + 555, + 481, + 578 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 555, + 481, + 578 + ], + "spans": [ + { + "bbox": [ + 132, + 555, + 481, + 578 + ], + "type": "text", + "content": "32. Zhou, X., Wang, D., Krahenbuhl, P.: Objects as points. CoRR abs/1904.07850 (2019)" + } + ] + } + ], + "index": 16 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "spans": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "type": "text", + "content": "16" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 166, + 91, + 225, + 100 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 166, + 91, + 225, + 100 + ], + "spans": [ + { + "bbox": [ + 166, + 91, + 225, + 100 + ], + "type": "text", + "content": "M. Zhou et al." + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 15 + } + ], + "_backend": "vlm", + "_version_name": "2.6.4" +} \ No newline at end of file diff --git a/2024/PDiscoFormer_ Relaxing Part Discovery Constraints with Vision Transformers/aa33e6f8-0640-410e-98c0-983fb97aea81_content_list.json b/2024/PDiscoFormer_ Relaxing Part Discovery Constraints with Vision Transformers/aa33e6f8-0640-410e-98c0-983fb97aea81_content_list.json new file mode 100644 index 0000000000000000000000000000000000000000..8212c97fb2e43eba89de1108be4cb06e9c0f5712 --- /dev/null +++ b/2024/PDiscoFormer_ Relaxing Part Discovery Constraints with Vision Transformers/aa33e6f8-0640-410e-98c0-983fb97aea81_content_list.json @@ -0,0 +1,1469 @@ +[ + { + "type": "text", + "text": "PDiscoFormer: Relaxing Part Discovery Constraints with Vision Transformers", + "text_level": 1, + "bbox": [ + 272, + 141, + 728, + 184 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Ananthu Aniraj $^{1}$ , Cassio F. Dantas $^{2}$ , Dino Ienc $^{2}$ , and Diego Marcos $^{1}$", + "bbox": [ + 225, + 212, + 774, + 228 + ], + "page_idx": 0 + }, + { + "type": "list", + "sub_type": "text", + "list_items": [ + "1 Inria, Univ. Montpellier, LIRMM, UMR TETIS, Montpellier, France {ananthu.aniraj, diego.marcos}@inria.fr", + "2 Inria, Inrae, Univ. Montpellier, UMR TETIS, Montpellier, France {cassio.fraga-dantas, dino.ienco}@inrae.fr" + ], + "bbox": [ + 264, + 239, + 736, + 295 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Abstract. Computer vision methods that explicitly detect object parts and reason on them are a step towards inherently interpretable models. Existing approaches that perform part discovery driven by a fine-grained classification task make very restrictive assumptions on the geometric properties of the discovered parts; they should be small and compact. Although this prior is useful in some cases, in this paper we show that pre-trained transformer-based vision models, such as self-supervised DINOv2 ViT, enable the relaxation of these constraints. In particular, we find that a total variation (TV) prior, which allows for multiple connected components of any size, substantially outperforms previous work. We test our approach on three fine-grained classification benchmarks: CUB, PartImageNet and Oxford Flowers, and compare our results to previously published methods as well as a re-implementation of the state-of-the-art method PDiscoNet with a transformer-based backbone. We consistently obtain substantial improvements across the board, both on part discovery metrics and the downstream classification task, showing that the strong inductive biases in self-supervised ViT models require to rethink the geometric priors that can be used for unsupervised part discovery. Training code and pre-trained models are available at https://github.com/ananthu-aniraj/pdiscoformer.", + "bbox": [ + 259, + 333, + 743, + 611 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "1 Introduction", + "text_level": 1, + "bbox": [ + 215, + 637, + 374, + 652 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Deep neural networks, first in the form of convolutional neural networks (CNN) and, more recently, vision transformers (ViT), have revolutionized the field of", + "bbox": [ + 212, + 667, + 787, + 700 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/067bcb1bd9d7075196ae8e1b7e0ee8c8c055660bf65861a01f6d2b91f8b0eded.jpg", + "image_caption": [ + "Fig. 1: Concentration vs total variation as part priors." + ], + "image_footnote": [], + "bbox": [ + 217, + 724, + 785, + 811 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "computer vision. However, their black-box nature often limits their application in domains where interpretability is crucial. To ameliorate the understanding of the internal behaviour of deep learning models trained for computer vision tasks, post-hoc interpretability methods [25,27,31] are commonly used. These methods generate saliency maps indicating regions in the image that contributed most to a classification decision. However, these are only approximations to the actual internal process of the model and may be misleading [28].", + "bbox": [ + 212, + 146, + 787, + 253 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Recent efforts have been directed towards developing interpretable-by-design deep neural networks [4, 13, 14, 17, 21, 22]. These approaches aim to identify semantically interpretable features in an image and utilize this information for solving a downstream task. Primarily tailored for fine-grained image classification, these methods seek to emulate human reasoning by subdividing the final task, such as classifying flower species or dog breeds, into multiple explicit intermediate stages, such as identifying lower-level semantic features like body parts, which in turn are used to provide a classification score.", + "bbox": [ + 212, + 261, + 787, + 383 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "However, improving interpretability while maintaining performance remains a challenge. Some models, such as those proposed in [13, 14, 17, 18], rely on objective functions that impose priors on the shape and area of object parts such as compactness-inducing part shaping losses that, in practice, result in spatial Gaussian priors of fixed variance on the part maps. While effective for specific datasets, like the commonly used CUB [38] benchmark for bird species classification, where assumptions like parts occurring only once in an image and being compact are beneficial, these assumptions may not generalize well to other use cases. For instance, in plant species classification, the presence of multiple instances of object parts such as leaves or flowers, as well as irregularly-shaped parts like branches or stems, renders these assumptions inadequate.", + "bbox": [ + 212, + 392, + 787, + 559 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "As we show in this work, such strong part shape priors are required when using a CNN-based backbone pre-trained on ImageNet; in order to reach acceptable classification results, it is imperative to fine-tune the whole backbone. This means that the only inductive bias that is kept throughout the part discovery process is the one brought in by the convolutional nature of the CNN architecture. A strong part shape prior is required in order to obtain consistent part maps in this setting, leading to all recent methods for part discovery in the context of fine-grained classification to use a concentration loss that makes part maps small and compact [13, 14, 17]. Here, we propose a method that addresses this issue by refraining from assuming any priors on part shape or area. The only prior we assume is that the part maps should tend to be piece-wise constant and thus be formed of one or more spatially connected components. Our model is capable of localizing up to a fixed number of semantically consistent parts of an object without relying on part annotations, using only the classification labels as supervision signals. The lack of assumptions about part shape and size makes our approach suitable to various fine-grained image classification tasks, as demonstrated by our empirical results, generalizing beyond cases where compactness is a good part prior.", + "bbox": [ + 212, + 568, + 787, + 840 + ], + "page_idx": 1 + }, + { + "type": "page_number", + "text": "2", + "bbox": [ + 217, + 114, + 228, + 126 + ], + "page_idx": 1 + }, + { + "type": "header", + "text": "A. Aniraj et al.", + "bbox": [ + 271, + 114, + 377, + 128 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "The main contribution of this paper is to build upon [17] and relax its constraining concentration-based part prior by a more flexible total variation prior. We investigate which additional elements are required to compensate for the flexibility of the total variation loss and find that a combination of an entropy loss and Gumbel-Softmax on the part maps allows for consistent and high quality part maps. Our approach makes better use of the representation of self-supervised pre-trained ViT models and allows to apply the model to tasks with diverse types of parts.", + "bbox": [ + 212, + 146, + 787, + 268 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "2 Related Work", + "text_level": 1, + "bbox": [ + 215, + 291, + 388, + 308 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Unsupervised Part Discovery. Machine learning methods designed for unsupervised part discovery aim to identify a small number of semantically consistent parts, typically around $K = 8$ , shared among classes in a dataset. These methods visualize all discovered parts in an image through per-part saliency maps, facilitating easier interpretation. However, as this task is unsupervised, it often relies on assumptions about the discovered parts, such as their distribution and shape. For instance, Huang et al. [13] assume that the presence of each part across a mini-batch follows a beta distribution, while [17] assumes that discovered parts are present at least once in a mini-batch. Additionally, assumptions are often made about the shape of the parts. Methods such as [14, 17, 18, 40] impose a spatial Gaussian prior with a fixed variance on the part maps, resulting in compact parts appearing once in the image. Such assumptions on part shape can be highly restrictive and may not generalize well across different scenarios. To address this limitation, our proposed method refrains from assuming such priors on part shape and only imposes a spatial prior that discovered parts should form connected components.", + "bbox": [ + 212, + 325, + 787, + 566 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Part Prototype Networks. Chen et al. [4] introduced the Prototypical Part Network (ProtoPNet), an inherently interpretable method featuring a fixed number of learned prototypes per class. These prototypes represent visual patterns or features learned by the network during training. The model computes the similarity between each learned prototype and image patches and uses a linear combination of these similarity scores for classification. While explanations provided by ProtoPNet visualize all prototypes along with their weighted similarity scores, the sheer number of prototypes, often proportional to the number of classes, can lead to redundancy and interpretability challenges [10,21]. Subsequent works in this line of research have sought to address these challenges. Some approaches focus on reducing the number of prototypes shared per class [30,37], while others explore sharing prototypes between classes [20, 22]. Despite these efforts, recent works like [22] still utilize a large number ( $K > 500$ ) of prototypes, which can pose significant challenges for end-user interpretation.", + "bbox": [ + 212, + 566, + 787, + 779 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Regularization in Image Segmentation. Early works on image segmentation made direct use of low-level image features, chiefly color, along with geometric priors, in order to find coherent segments. The Mumford-Shah model [19], originally devised for piece-wise image reconstruction, was applied to find segments", + "bbox": [ + 212, + 779, + 787, + 840 + ], + "page_idx": 2 + }, + { + "type": "header", + "text": "PDiscoFormer: Part Discovery with Transformers", + "bbox": [ + 401, + 114, + 730, + 128 + ], + "page_idx": 2 + }, + { + "type": "page_number", + "text": "3", + "bbox": [ + 774, + 116, + 784, + 126 + ], + "page_idx": 2 + }, + { + "type": "image", + "img_path": "images/72f2233595f30e2531b43d5feca2bd0796ef12aabc0cbfd7ed18999f4312f4a9.jpg", + "image_caption": [ + "Fig. 2: The PDiscoFormer architecture for part discovery." + ], + "image_footnote": [], + "bbox": [ + 250, + 146, + 781, + 383 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "that are as homogeneous as possible while trying to minimize the length of their boundary as a regularization [3]. Total variation (TV) [29] has been shown to be a powerful and efficient regularization to approximate this boundary length prior, and has been extensively used for piece-wise signal and image reconstruction [7,35], as well as unsupervised image segmentation [8,36]. With the advent of deep learning, TV regularization has proven valuable for style transfer [16] and edge-preserving smoothing [41]. In this work we explore the use of TV regularization for weakly-supervised part discovery in order to alleviate the issues brought by previously used priors.", + "bbox": [ + 212, + 438, + 787, + 574 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Emergent grouping of ViT representations. Some recent research [26,39] has studied the emergence of object segmentation by supervising with image-text pairs. These methods learn to group similar patches in a ViT and assign them to textual concepts. They aim to identify the pixels in an input image corresponding to a predefined text prompt, focusing on segmenting whole objects rather than decomposing them into parts, as our model does with part discovery, and are limited to concepts that are represented in the text annotations.", + "bbox": [ + 212, + 574, + 787, + 681 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "3 Methodology", + "text_level": 1, + "bbox": [ + 215, + 702, + 380, + 720 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Here we introduce an end-to-end trainable method (see Fig. 2) capable of automatically identifying $K$ discriminative parts (semantically interpretable image regions) useful for solving a fine-grained image classification task. In contrast to earlier methods [13,17], our approach avoids any assumptions related to part shape, size or number of part instances. As in [13,17], our model is weakly supervised and relies solely on image-level class labels, thus uncovering discriminative parts without requiring any additional supervision. Our formulation is based", + "bbox": [ + 212, + 733, + 787, + 840 + ], + "page_idx": 3 + }, + { + "type": "page_number", + "text": "4", + "bbox": [ + 217, + 114, + 228, + 126 + ], + "page_idx": 3 + }, + { + "type": "header", + "text": "A. Aniraj et al.", + "bbox": [ + 271, + 114, + 377, + 128 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "on [17], from which we take take most elements of the architecture, except for the Gumbel-Softmax and the layer normalization-based part modulation. The total variation, entropy and background losses have not been used in previous part discovery works and are shown to be important in our results.", + "bbox": [ + 212, + 146, + 782, + 205 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "3.1 Model architecture", + "text_level": 1, + "bbox": [ + 215, + 233, + 419, + 247 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Let $\\mathbf{x} \\in \\mathbb{R}^{3 \\times M \\times N}$ be an image in the dataset, and let $y \\in \\{1, 2, \\dots, C\\}$ represent the corresponding classification label. Given a backbone model $h_{\\theta}(\\cdot)$ , we extract a feature tensor:", + "bbox": [ + 212, + 260, + 782, + 304 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\n\\mathbf {z} = h _ {\\theta} (\\mathbf {x}) \\in \\mathbb {R} ^ {D \\times H \\times W} \\tag {1}\n$$\n", + "text_format": "latex", + "bbox": [ + 421, + 306, + 784, + 325 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "We drop the class and register tokens of the ViT and reshape the patch tokens to obtain $\\mathbf{z}$ . Following prior research [13, 14, 17], we compute attention maps $\\mathbf{A} \\in [0,1]^{(K + 1) \\times H \\times W}$ , where the last channel in the first dimension is used to represent the background region in the image. These attention maps are computed using the negative squared Euclidean distance function between patch token features $\\mathbf{z}_{ij} \\in \\mathbb{R}^D$ , $i \\in \\{1, \\dots, H\\}$ , $j \\in \\{1, \\dots, W\\}$ and each of the learnable prototypes $\\mathbf{p}^k \\in \\mathbb{R}^D$ , with $k \\in \\{1, \\dots, K + 1\\}$ , followed by a Gumbel-Softmax [15] across the $K + 1$ channels:", + "bbox": [ + 212, + 334, + 784, + 454 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\na _ {i j} ^ {k} = \\frac {\\exp \\left(- \\| \\mathbf {z} _ {i j} - \\mathbf {p} ^ {k} \\| ^ {2} + \\gamma_ {k}\\right)}{\\sum_ {l = 1} ^ {K + 1} \\exp \\left(- \\| \\mathbf {z} _ {i j} - \\mathbf {p} ^ {l} \\| ^ {2} + \\gamma_ {l}\\right)} \\tag {2}\n$$\n", + "text_format": "latex", + "bbox": [ + 372, + 470, + 784, + 507 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "where $\\gamma_{k}$ and $\\gamma_{l}$ are independently drawn samples from the Gumbel(0,1) distribution, and $a_{ij}^{k}$ represents an element of $\\mathbf{A}$ .", + "bbox": [ + 212, + 513, + 782, + 544 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "The attention maps are used to compute part embedding vectors $\\mathbf{v}^k\\in \\mathbb{R}^D$ by using the attention values to calculate a weighted average over the feature tensor $\\mathbf{z}$ :", + "bbox": [ + 212, + 545, + 782, + 589 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\n\\mathbf {v} ^ {k} = \\frac {\\sum_ {i} \\sum_ {j} a _ {i j} ^ {k} \\mathbf {z} _ {i j}}{H W} \\tag {3}\n$$\n", + "text_format": "latex", + "bbox": [ + 433, + 606, + 784, + 638 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "These embedding vectors are then modulated using a layer normalization [2] operation, with normalization applied over the feature and prototype dimensions. This operation helps to make each embedding vector distinct for the classifier:", + "bbox": [ + 212, + 646, + 782, + 691 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\n\\mathbf {v} _ {m} ^ {k} = \\frac {\\left(\\mathbf {v} ^ {k} - \\operatorname {E} \\left[ \\mathbf {v} ^ {k} \\right]\\right)}{\\sqrt {\\operatorname {V a r} \\left[ \\mathbf {v} ^ {k} \\right] + \\epsilon}} \\odot \\mathbf {w} _ {m} ^ {k} + \\mathbf {b} _ {m} ^ {k} \\tag {4}\n$$\n", + "text_format": "latex", + "bbox": [ + 385, + 705, + 784, + 741 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "with $\\mathbf{w}_m^k, \\mathbf{b}_m^k \\in \\mathbb{R}^D$ being the modulation weights and biases per part. These modulated embedding vectors $\\mathbf{v}_m^k$ for $k \\in \\{1, \\dots, K\\}$ (we omit the background embedding) are then linearly projected to obtain a vector of class scores conditioned on the part embedding:", + "bbox": [ + 212, + 750, + 782, + 810 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\n\\mathbf {y} ^ {k} = W _ {c} \\cdot \\mathbf {v} _ {m} ^ {k} \\tag {5}\n$$\n", + "text_format": "latex", + "bbox": [ + 452, + 824, + 784, + 840 + ], + "page_idx": 4 + }, + { + "type": "header", + "text": "PDiscoFormer: Part Discovery with Transformers", + "bbox": [ + 401, + 114, + 732, + 128 + ], + "page_idx": 4 + }, + { + "type": "page_number", + "text": "5", + "bbox": [ + 774, + 116, + 784, + 126 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "The same linear classifier $W_{c} \\in \\mathbb{R}^{C \\times D}$ is applied to each of the modulated embedding vectors, i.e., for each of the $K$ discovered foreground parts. We then take the mean of the per-part class scores to obtain the classification scores:", + "bbox": [ + 212, + 145, + 787, + 191 + ], + "page_idx": 5 + }, + { + "type": "equation", + "text": "\n$$\n\\mathbf {y} = \\frac {1}{K} \\sum_ {k = 1} ^ {K} \\mathbf {y} ^ {k} \\tag {6}\n$$\n", + "text_format": "latex", + "bbox": [ + 450, + 204, + 785, + 244 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Finally, a softmax function is employed to obtain the classification probabilities. Part dropout. We adopt dropout [34] on the modulated embedding vectors $\\mathbf{v}_m^k$ , as proposed in [17], during model training. This drops entire part embedding vectors at a time and ensures that all discovered parts are discriminative, since discovering a single discriminative part would not be enough to solve the downstream classification task.", + "bbox": [ + 214, + 250, + 787, + 339 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "3.2 Loss Functions", + "text_level": 1, + "bbox": [ + 215, + 359, + 385, + 375 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Our main contribution consists in proposing a set of part-shaping loss functions that are able to make the most out of the inductive biases learned by self-supervised ViT models.", + "bbox": [ + 212, + 383, + 785, + 428 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Classification Loss $(\\mathcal{L}_{\\mathbf{c}})$ . Our model is mainly driven by the fine-grained classification objective during training. We use the standard negative log likelihood or multi-class cross-entropy for this purpose. While the classification loss helps ensure the discriminative nature of the discovered parts, it may not directly enforce the learned attention maps to represent salient parts of the object in the image. To address this, we incorporate additional objective functions to enforce priors on the discovered parts. These objectives effectively guide the learning process to focus on relevant regions of the image.", + "bbox": [ + 212, + 430, + 787, + 549 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Orthogonality Loss $(\\mathcal{L}_{\\perp})$ . We aim to discover parts that are semantically distinct. To achieve this, we apply the orthogonality loss, which encourages the learned part embedding vectors to be decorrelated from one another. This is accomplished by minimizing the cosine similarity between the embedding vectors. We apply this loss on the modulated embedding vectors $\\mathbf{v}_m^k$ .", + "bbox": [ + 212, + 550, + 787, + 628 + ], + "page_idx": 5 + }, + { + "type": "equation", + "text": "\n$$\n\\mathcal {L} _ {\\perp} = \\sum_ {k = 1} ^ {K + 1} \\sum_ {l \\neq k} \\frac {\\mathbf {v} _ {m} ^ {k} \\cdot \\mathbf {v} _ {m} ^ {l}}{\\| \\mathbf {v} _ {m} ^ {k} \\| \\cdot \\| \\mathbf {v} _ {m} ^ {l} \\|} \\tag {7}\n$$\n", + "text_format": "latex", + "bbox": [ + 401, + 638, + 785, + 678 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Equivalence Loss $(\\mathcal{L}_{\\mathrm{eq}})$ . We aim to detect the same parts even if the image is translated, rotated, or scaled. To achieve this, we formulate the equivariance loss, which encourages the learned part attention maps to be equivariant to rigid transformations. Specifically, given an input image, we use an affine transform $T$ with parameters randomly sampled from a pre-defined range and pass both the original image and the transformed image through our model. Then, we invert the transformation on the part attention maps of the transformed image. Finally, the equivariance loss is formulated as the cosine distance between the part attention maps of the original image and those of the corresponding transformed image.", + "bbox": [ + 212, + 688, + 787, + 840 + ], + "page_idx": 5 + }, + { + "type": "page_number", + "text": "6", + "bbox": [ + 217, + 114, + 228, + 126 + ], + "page_idx": 5 + }, + { + "type": "header", + "text": "A. Aniraj et al.", + "bbox": [ + 271, + 114, + 377, + 128 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Let $A^k(\\mathbf{x})$ be a function that returns the $k^{th}$ attention map given an input image $\\mathbf{x}$ . Then:", + "bbox": [ + 212, + 145, + 782, + 176 + ], + "page_idx": 6 + }, + { + "type": "equation", + "text": "\n$$\n\\mathcal {L} _ {\\mathrm {e q}} = 1 - \\frac {1}{K} \\sum_ {k} \\frac {\\left\\| A ^ {k} (\\mathbf {x}) \\cdot T ^ {- 1} \\left(A ^ {k} (T (\\mathbf {x})) \\right. \\right\\|}{\\| A ^ {k} (\\mathbf {x}) \\| \\cdot \\| A ^ {k} (T (\\mathbf {x})) \\|} \\tag {8}\n$$\n", + "text_format": "latex", + "bbox": [ + 346, + 186, + 785, + 227 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Presence Loss $(\\mathcal{L}_{\\mathbf{p}})$ . We assume that all discovered foreground parts are present in at least some images of the training dataset, while the background, the $(K + 1)^{th}$ part, is expected to be present in all images in the dataset. To enforce this, we formulate the presence loss, which consists of two parts. Firstly, we define a loss applied only to the discovered foreground parts, ensuring that each part is present at least once in a mini-batch. Given a batch $\\{\\mathbf{x}_1,\\dots ,\\mathbf{x}_B\\}$ , the presence loss is calculated as:", + "bbox": [ + 212, + 237, + 784, + 344 + ], + "page_idx": 6 + }, + { + "type": "equation", + "text": "\n$$\n\\mathcal {L} _ {\\mathrm {p} _ {1}} = 1 - \\frac {1}{K} \\sum_ {k} \\max _ {b, i, j} \\bar {a} _ {i j} ^ {k} \\left(\\mathbf {x} _ {b}\\right) \\tag {9}\n$$\n", + "text_format": "latex", + "bbox": [ + 398, + 354, + 785, + 390 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "where $\\bar{A}^k (\\mathbf{x}_b) = \\mathrm{avgpool}(A^k (\\mathbf{x}_b))$ is the output of a 2D average pooling operator with a small kernel size employed to prevent single-pixel attention maps.", + "bbox": [ + 212, + 401, + 782, + 433 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Secondly, to ensure the presence of the background, we introduce a stricter presence loss, which is satisfied if the background is present in all images of the mini-batch. This loss is formulated as follows:", + "bbox": [ + 212, + 434, + 782, + 477 + ], + "page_idx": 6 + }, + { + "type": "equation", + "text": "\n$$\n\\mathcal {L} _ {\\mathrm {p} _ {0}} = - \\frac {1}{B} \\sum_ {b} \\log \\left(\\max _ {i, j} m _ {i j} \\bar {a} _ {i j} ^ {K + 1} (\\mathbf {x} _ {b})\\right) \\tag {10}\n$$\n", + "text_format": "latex", + "bbox": [ + 357, + 487, + 785, + 523 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "where $M = [m_{ij}]^{H \\times W}$ is a soft mask with $m_{ij} \\in [0,1]$ that privileges entries placed farther from the image center (towards the borders):", + "bbox": [ + 212, + 535, + 782, + 568 + ], + "page_idx": 6 + }, + { + "type": "equation", + "text": "\n$$\nm _ {i j} = 2 \\left(\\frac {i - 1}{H - 1} - \\frac {1}{2}\\right) ^ {2} + 2 \\left(\\frac {j - 1}{W - 1} - \\frac {1}{2}\\right) ^ {2} \\tag {11}\n$$\n", + "text_format": "latex", + "bbox": [ + 344, + 578, + 785, + 614 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "This loss captures the notion that the background is at the same time expected in every image and is more likely to appear near the boundaries of the image.", + "bbox": [ + 212, + 623, + 782, + 667 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Entropy Loss $(\\mathcal{L}_{\\mathrm{ent}})$ . We aim to ensure that each patch token is assigned to a unique part. To achieve this, we formulate an entropy [32] loss applied to the part attention maps:", + "bbox": [ + 212, + 667, + 782, + 715 + ], + "page_idx": 6 + }, + { + "type": "equation", + "text": "\n$$\n\\mathcal {L} _ {\\text {e n t}} = \\frac {- 1}{K + 1} \\sum_ {k = 1} ^ {K + 1} \\sum_ {i j} a _ {i j} ^ {k} \\log \\left(a _ {i j} ^ {k}\\right) \\tag {12}\n$$\n", + "text_format": "latex", + "bbox": [ + 380, + 726, + 785, + 768 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Total Variation Loss $(\\mathcal{L}_{\\mathrm{tv}})$ . We do not want to make any assumptions about the part shape. However, we would still like the discovered parts to be composed of one or, at most, a few connected components. To achieve this, we make use of a total variation loss [29] applied to the part attention maps:", + "bbox": [ + 212, + 779, + 785, + 840 + ], + "page_idx": 6 + }, + { + "type": "header", + "text": "PDiscoFormer: Part Discovery with Transformers", + "bbox": [ + 401, + 114, + 732, + 128 + ], + "page_idx": 6 + }, + { + "type": "page_number", + "text": "7", + "bbox": [ + 774, + 114, + 785, + 126 + ], + "page_idx": 6 + }, + { + "type": "equation", + "text": "\n$$\n\\mathcal {L} _ {\\mathrm {t v}} = \\frac {1}{H W} \\sum_ {k = 1} ^ {K + 1} \\sum_ {i j} \\left| \\nabla a _ {i j} ^ {k} \\right| \\tag {13}\n$$\n", + "text_format": "latex", + "bbox": [ + 406, + 157, + 785, + 200 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "where $\\nabla a_{ij}^{k}$ is the spatial image gradient of part map $A^k$ in location $ij$ .", + "bbox": [ + 215, + 205, + 730, + 223 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "4 Experimental Setup", + "text_level": 1, + "bbox": [ + 215, + 241, + 447, + 258 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "We aim to perform part discovery using only image-level class labels as supervision signals, assuming that the discovered parts are shared among classes, which is common in fine-grained image classification settings. Moreover, we aim to compare our model with similar methods in the literature in scenarios where they have been empirically shown to work well, as well as in new and more challenging scenarios where underlying data characteristics may require more flexible geometric modeling.", + "bbox": [ + 212, + 270, + 787, + 375 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "All models were implemented in PyTorch. We used the DinoV2 [24] ViT [9] model with register tokens [6] as the starting weights for all experiments, utilizing the Base (ViT-B) variant. The training settings and compute requirements of our model are detailed in the supplementary material.", + "bbox": [ + 212, + 375, + 787, + 436 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "4.1 Datasets", + "text_level": 1, + "bbox": [ + 215, + 455, + 334, + 469 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "CUB Dataset. The CUB dataset [38] consists of images depicting 200 bird species, with 5,994 images in the training set and 5,794 images in the testing set. Keypoint annotations for 15 different bird body parts are available for each image. Bird parts are typically compact, although their presence may vary in images due to changes in pose and occlusion. Most images contain only a single instance of a bird. For this dataset, we evaluate the results using keypoint regression (Kp) using ground truth part keypoint annotations to compare them with centroids discovered by the model. Furthermore, we compute the Normalized Mutual Information (NMI) and Adjusted Rand Index (ARI) to quantify the clustering quality of the discovered parts, as in related literature [5].", + "bbox": [ + 212, + 477, + 787, + 628 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "PartImageNet Dataset. The PartImageNet dataset [11] comprises 158 classes distributed among 11 super-classes, such as airplanes, birds, and snakes. Each super-class is associated to between two and five possible parts, for a total of 41 part classes. Additionally, test images may contain multiple instances of objects, sometimes from different super-classes, while the classification label provides information about a single object. Notably, several super-classes, including snakes, airplanes, and reptiles, often exhibit irregular and non-compact part shapes. For our experiments, we use the PartImageNet 00D variant. Although originally intended for out-of-distribution detection, this variant is commonly used in related literature. We train our models on 14,865 training images and test on 1,658 images, respectively. For evaluation, we derive part centroids from ground truth part masks and calculate NMI and ARI scores to assess clustering quality. Additionally, we include an experiment on the newly released PartImageNet Seg variant in the supplementary material.", + "bbox": [ + 212, + 628, + 787, + 840 + ], + "page_idx": 7 + }, + { + "type": "page_number", + "text": "8", + "bbox": [ + 217, + 114, + 228, + 126 + ], + "page_idx": 7 + }, + { + "type": "header", + "text": "A. Aniraj et al.", + "bbox": [ + 271, + 114, + 377, + 128 + ], + "page_idx": 7 + }, + { + "type": "table", + "img_path": "images/e9634cda15f95bda50d3bfa2c751819b30ebf6e558509fdc2fff15c41abc96de.jpg", + "table_caption": [ + "Table 1: Quantitative comparison of the part discovery and classification capabilities of our model with related approaches from the literature. In this table, $K$ refers to the number of parts. The method with the label ** does not use any class supervision." + ], + "table_footnote": [], + "table_body": "
MethodCUB(%)PartImageNet OOD(%)Flowers(%)
KKp ↓NMI ↑ARI ↑Top-1 Acc. ↑KNMI ↑ARI ↑Top-1 Acc. ↑KFg. mIoU ↑Top-1 Acc. ↑
Dino [1]**4-31.1811.21-819.177.59-254.95-
8-47.2119.76-2531.4614.16-455.11-
16-50.5726.14-5037.8116.50-854.44-
Huang [13]411.5129.7414.0487.3085.881.5374.22229.9293.07
811.6035.7215.9086.05257.561.2573.56433.2293.14
1612.6043.9221.1085.935010.191.0573.20817.2692.86
PDiscoNet [17]49.1237.8215.2686.17827.138.7688.58219.0477.51
88.5250.0826.9686.722532.4110.6989.00434.7683.05
167.6756.8738.0587.495041.4914.1786.06849.1081.04
PdiscoNet + ViT-B [6]47.7052.5926.6688.61819.2834.7290.9524.9292.75
86.3465.0137.9086.952528.2350.3590.2941.9595.48
165.9568.6343.4184.045029.4827.8089.69813.1897.40
PDiscoFormer (Ours)47.4158.1325.1189.06829.0052.4089.75273.6299.61
85.9969.8743.4988.792544.7159.2790.77473.3299.54
165.7473.3855.8388.725046.2962.2191.01869.5999.64
", + "bbox": [ + 215, + 198, + 785, + 448 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "Oxford 102 Flower Dataset. The Oxford 102 Flower Dataset [23] comprises 8,189 images divided into 102 flower classes. The training and validation sets contain 10 images per class, totaling 1,020 images each, while the test set consists of 6,149 images, with a minimum of 20 images per class. Since no part annotations are provided for this dataset, we evaluate the quality of the discovered parts by combining them and comparing the result to the provided foreground/background masks. We evaluate model performance by calculating the mean Intersection over Union (mIoU) for the foreground mask. This dataset often includes multiple instances of parts in a single image, such as flowers, as well as parts with irregular shapes, such as flower petals.", + "bbox": [ + 212, + 474, + 787, + 627 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "5 Experiments", + "text_level": 1, + "bbox": [ + 215, + 648, + 375, + 666 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "5.1 Comparison with State-of-the-Art", + "text_level": 1, + "bbox": [ + 215, + 679, + 540, + 695 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "We conduct a quantitative analysis of our model's part discovery and classification capabilities and compare it with those of related methods from the literature. Additionally, we compare our model against a re-implementation of the state-of-the-art PDiCoNet [17], where we replaced the backbone with the variant used in this paper. The results of this comparison are summarized in Tab. 1. From the results, our method consistently outperforms existing approaches for both part discovery and image classification across all datasets. We attribute this to the synergistic effects of the strong inductive biases ingrained in the self-supervised ViT backbone and the enhanced flexibility of the proposed combina", + "bbox": [ + 212, + 703, + 785, + 840 + ], + "page_idx": 8 + }, + { + "type": "header", + "text": "PDiscoFormer: Part Discovery with Transformers", + "bbox": [ + 401, + 114, + 730, + 128 + ], + "page_idx": 8 + }, + { + "type": "page_number", + "text": "9", + "bbox": [ + 774, + 116, + 784, + 126 + ], + "page_idx": 8 + }, + { + "type": "image", + "img_path": "images/6d9d70a64ff3ec74c6be094189be276cf5689922e8bdbda6841cba801ae1ab09.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 217, + 143, + 795, + 329 + ], + "page_idx": 9 + }, + { + "type": "image", + "img_path": "images/90f23c12cb2713a453831b957e2251654f6cb2a561f37c1e2c5c5284fb43e2d8.jpg", + "image_caption": [ + "Fig. 3: Qualitative results on CUB for $K = 8$ .", + "Fig. 4: Qualitative results on PartImageNet for $K = 8$ ." + ], + "image_footnote": [], + "bbox": [ + 217, + 358, + 795, + 544 + ], + "page_idx": 9 + }, + { + "type": "text", + "text": "tion of geometric priors. Particularly noteworthy are the significant performance gains observed on the Flowers and PartImageNet datasets, which contain objects with irregular shapes occurring in multiple instances within the same image, underscoring the importance of our geometric part prior that favors piece-wise constant and spatially connected components. Even on CUB, where most parts fit well into the compactness prior, we see a consistent improvements of results with respect to PDiscoNet, even when using the same ViT backbone. With 16 parts, our model achieves $55.8\\%$ ARI and $73.4\\%$ NMI, compared to $43.4\\%$ ARI and $68.6\\%$ NMI of the best competing method, PDiscoNet with a ViT backbone. Despite competitive performance in classification tasks on CUB and PartImageNet, the PdiscoNet+ViT-B model tends to degrade in accuracy as part complexity increases, particularly in CUB, unlike our model, which maintains robust accuracy for any number of parts. In the Flowers dataset, all competing methods result in subpar performance, while our approach excels at both detecting the foreground and classification no matter the number of parts. In this", + "bbox": [ + 212, + 613, + 787, + 840 + ], + "page_idx": 9 + }, + { + "type": "page_number", + "text": "10", + "bbox": [ + 217, + 114, + 235, + 126 + ], + "page_idx": 9 + }, + { + "type": "header", + "text": "A. Aniraj et al.", + "bbox": [ + 271, + 114, + 377, + 128 + ], + "page_idx": 9 + }, + { + "type": "image", + "img_path": "images/7c2457fcb03457d8b8b5b415be86bd6763ac2661fd8dd858d2d2a975b9274320.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 217, + 143, + 795, + 329 + ], + "page_idx": 10 + }, + { + "type": "image", + "img_path": "images/9edc6bf3b7342da549ca128a391e19717491e20d60be69c4d33456616af0617b.jpg", + "image_caption": [ + "Fig. 5: Qualitative results on Flowers for $K = 2$ .", + "Fig. 6: Effect of $K$ on PDiscoFormer. The left three images correspond to Flowers (from top to bottom, $K = 2,4,8$ ). The three middle ones are CUB, with $K$ being 4, 8 and 16. The three on the right belong to PartImageNet, with $K = 8,25,50$ ." + ], + "image_footnote": [], + "bbox": [ + 217, + 357, + 795, + 542 + ], + "page_idx": 10 + }, + { + "type": "text", + "text": "dataset, PDiscoNet+ViT-B results in the second best classification accuracy, $97.4\\%$ , after the $99.6\\%$ obtained by our method, but completely fails to locate the foreground, with an mIoU of $13.2\\%$ , against $73.6\\%$ for PDiscoFormer. Additionally, we extend our comparisons by applying our proposed losses to a ResNet backbone and a fully frozen ViT architecture, showing that a pre-trained ViT is required to fully profit from the proposed set of losses and that fine-tuning the position, class and register tokens, while keeping the rest of the ViT frozen, results in better performances than leaving the whole ViT frozen. The results of these additional comparisons are presented in the supplementary material.", + "bbox": [ + 212, + 638, + 787, + 775 + ], + "page_idx": 10 + }, + { + "type": "text", + "text": "Qualitative Analysis. A qualitative analysis, as illustrated in Fig. 3, Fig. 4, Fig. 5 and Fig. 6 corroborates our quantitative findings, showcasing our model's ability to accurately discover irregularly shaped body parts such as bird wings (CUB), flower petals (Flowers) and the body of a snake (PartImageNet). At the", + "bbox": [ + 212, + 779, + 787, + 840 + ], + "page_idx": 10 + }, + { + "type": "header", + "text": "PDiscoFormer: Part Discovery with Transformers", + "bbox": [ + 401, + 114, + 730, + 128 + ], + "page_idx": 10 + }, + { + "type": "page_number", + "text": "11", + "bbox": [ + 767, + 116, + 782, + 126 + ], + "page_idx": 10 + }, + { + "type": "table", + "img_path": "images/96cf7c2a388be523d8e690ea8103240a91b629021a00c7373d13a6da53fae183.jpg", + "table_caption": [ + "Table 2: Results of the ablation study in PartImageNet and CUB." + ], + "table_footnote": [], + "table_body": "
PartImageNet OOD (K=25)CUB (K=16)
NMI ↑ARI ↑Top-1 Acc. ↑Kp ↓NMI ↑ARI ↑Top-1 Acc. ↑
Full Model44.7159.2790.775.7473.3855.8388.72
No Lp041.0549.0889.815.8568.6643.0983.64
No Lent39.1654.4691.195.5766.8845.6888.21
No Leq43.3854.1990.479.0556.8530.6383.71
No L⊥42.4458.8691.015.7970.5550.9088.63
No Lp133.9858.3191.077.1461.0434.2889.04
No Ltv33.7325.3590.476.0170.7851.8380.22
No Per-Part LayerNorm31.1440.7389.695.7470.7050.6269.73
No Part Dropout33.0843.9889.996.0870.2841.6987.14
No Gumbel32.9854.6890.716.4468.9243.6186.43
", + "bbox": [ + 261, + 170, + 743, + 369 + ], + "page_idx": 11 + }, + { + "type": "text", + "text": "same time, PDiscoNet [17] fails to discover these parts altogether, even when retrained on a ViT-B backbone. This indicates the usefulness of the flexible geometric priors enabled by our total variation loss. Compared to PDiscoNet [17], not only do we achieve a much tighter coverage of the object of interest, with our PDiscoFormer more closely following the object boundaries. This is particularly the case in parts that are relatively large, such as open bird wings and flower corollas. In addition to that, we can observe a substantial improvement in the consistency and semantic interpretability of parts. For instance, the parts discovered by our model for the four quadruped images displayed in Fig. 4 can immediately be interpreted (the green part corresponds to the snout while the brown one to the ears), while those from PDiscoNet are not clear. In Fig. 6 we vary the number of parts, $K$ . On Flowers, using two parts leads to a clear assignment of one part to the flower calyx and one to the corolla, while more parts tends to exhibit oversegmentation. CUB, on the other hand, is able to absorb a larger number of parts. With $K = 4$ , beak and legs are covered by the same part, while a fine-grained and consistent distinction of the different wing parts is achieved with $K = 16$ . Interestingly, in the case of PartImageNet, the effective number of parts being used by a specific super-class remains almost constant when increasing $K$ .", + "bbox": [ + 212, + 397, + 787, + 685 + ], + "page_idx": 11 + }, + { + "type": "text", + "text": "5.2 Ablation Study", + "text_level": 1, + "bbox": [ + 215, + 708, + 388, + 723 + ], + "page_idx": 11 + }, + { + "type": "text", + "text": "We conducted an ablation study by systematically removing one loss function or architectural element at a time and by replacing the Gumbel-Softmax with a regular Softmax. This was performed on the PartImageNet OOD dataset with $K = 25$ parts and the CUB dataset with $K = 16$ parts. The quantitative results of the ablation study are shown in Tab. 2 and some qualitative results can be seen in Fig. 7. From the results, we can observe the following: Total Variation $(\\mathcal{L}_{tv})$ prevents spurious part activations in the background, possibly reducing", + "bbox": [ + 212, + 733, + 787, + 842 + ], + "page_idx": 11 + }, + { + "type": "page_number", + "text": "12", + "bbox": [ + 217, + 114, + 235, + 126 + ], + "page_idx": 11 + }, + { + "type": "header", + "text": "A. Aniraj et al.", + "bbox": [ + 271, + 114, + 377, + 128 + ], + "page_idx": 11 + }, + { + "type": "image", + "img_path": "images/eaeb99af03887e3e9b1b7e4d6e8ffa81167451c6b916e49be4cf827a7f3db7ac.jpg", + "image_caption": [ + "Fig. 7: Qualitative results from the ablation study." + ], + "image_footnote": [], + "bbox": [ + 218, + 142, + 795, + 513 + ], + "page_idx": 12 + }, + { + "type": "text", + "text": "information leakage between foreground parts and the background. It has the largest impact on PartImageNet in terms of ARI. The absence of the Background Loss $(\\mathcal{L}_{\\mathbf{p}_0})$ results in the background part not being activated, causing the model to discover the background as a foreground part. Presence Loss $(\\mathcal{L}_{\\mathbf{p}})$ and Part Dropout promote the utilization of all parts, preventing domination by a few parts which can lead to multiple unused parts (Fig. 7). Removing the Per-Part Layer Norm leads to inconsistent part discovery (Fig. 7) indicating that it contributes to the creation of discriminative part-specific features driven by the classification loss. Equivalence Loss $(\\mathcal{L}_{eq})$ enhances robustness to affine geometric transformations, supporting stable part discovery across diverse inputs (Fig. 7). Gumbel-Softmax and Entropy Loss $(\\mathcal{L}_{\\mathrm{ent}})$ ensure unique part assignments for each patch token, further reducing information leakage, as shown in our supplementary material. As shown in Tab. 2, the complete model, incorporating all loss functions and architectural choices, generally performs best for both part discovery and classification, demonstrating the efficacy of our approach.", + "bbox": [ + 212, + 598, + 787, + 840 + ], + "page_idx": 12 + }, + { + "type": "header", + "text": "PDiscoFormer: Part Discovery with Transformers", + "bbox": [ + 401, + 114, + 730, + 128 + ], + "page_idx": 12 + }, + { + "type": "page_number", + "text": "13", + "bbox": [ + 767, + 114, + 785, + 126 + ], + "page_idx": 12 + }, + { + "type": "text", + "text": "6 Limitations and Future Work", + "text_level": 1, + "bbox": [ + 215, + 143, + 537, + 162 + ], + "page_idx": 13 + }, + { + "type": "text", + "text": "Our study focused on datasets with part annotations or FG-BG masks available for the test set. This allowed for quantitative comparison with state-of-the-art methods. However, exploring the training of our model on larger and more diverse datasets could provide valuable insights and further validate its performance in real-world scenarios. For example, larger datasets may introduce more variability in object appearances and poses, challenging the model to generalize better. Methods to enable the evaluation of part discovery in such settings would allow to take steps in this direction. The main hyper-parameter $K$ , the number of parts to discover, has a large impact on the granularity of the discovered parts (Fig. 6). Exploring ways of automatically estimating this value while preserving interpretability in the discovered semantic parts would be a valuable contribution. This could enhance the adaptability of the model to various datasets and streamline the training process by removing the need for manual selection of $K$ . Additionally, while our approach can detect parts of various shapes due to the total variation prior, it cannot ascertain if they belong to the same object. Integrating our method with unsupervised object discovery research [12,33] could enable part discovery within each detected object, improving capabilities in complex scenes. These methods aim to identify and localize individual objects without explicit supervision, complementing our approach for a more comprehensive scene understanding. These considerations suggest promising directions for extending and refining our method.", + "bbox": [ + 212, + 172, + 787, + 491 + ], + "page_idx": 13 + }, + { + "type": "text", + "text": "7 Conclusion", + "text_level": 1, + "bbox": [ + 215, + 508, + 359, + 525 + ], + "page_idx": 13 + }, + { + "type": "text", + "text": "In this paper, we presented PDiscoFormer, a new approach for unsupervised part discovery. Unlike previous methods that imposed strict geometric priors on discovered object parts, our model, based on self-supervised vision transformers, demonstrated the capability to relax these constraints. By leveraging a total variation (TV) prior, our model achieved substantial improvements in both part discovery and downstream classification accuracy when compared to the state-of-the-art across multiple datasets, specifically CUB, PartImageNet, and Oxford Flowers. Our findings highlight the importance of rethinking geometric priors in unsupervised part discovery tasks, especially in scenarios where objects exhibit complex shapes or occur multiple times within the same image. By refraining from assuming specific shape or size priors, our model demonstrated enhanced generalization capabilities, making it applicable to a broader range of fine-grained image classification tasks. Overall, our study shows that the strong inductive biases learned by vision transformers during the self-supervised pre-training stage enable the use of flexible shape priors such as total variation, leading to robust performance on part discovery and the downstream classification task on all of the tested datasets.", + "bbox": [ + 212, + 537, + 787, + 792 + ], + "page_idx": 13 + }, + { + "type": "text", + "text": "Acknowledgment. This work was supported in part by ANR project OBTEA ANR-22-CPJ1-0054-01 and, was granted access to the HPC resources of IDRIS under the allocation 2023-AD011014325 made by GENCI.", + "bbox": [ + 215, + 794, + 785, + 839 + ], + "page_idx": 13 + }, + { + "type": "page_number", + "text": "14", + "bbox": [ + 217, + 114, + 235, + 126 + ], + "page_idx": 13 + }, + { + "type": "header", + "text": "A. Aniraj et al.", + "bbox": [ + 271, + 114, + 377, + 128 + ], + "page_idx": 13 + }, + { + "type": "text", + "text": "References", + "text_level": 1, + "bbox": [ + 215, + 143, + 321, + 159 + ], + "page_idx": 14 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "1. Amir, S., Gandelsman, Y., Bagon, S., Dekel, T.: Deep vit features as dense visual descriptors. arXiv preprint arXiv:2112.05814 2(3), 4 (2021)", + "2. Ba, J.L., Kiros, J.R., Hinton, G.E.: Layer normalization. arXiv preprint arXiv:1607.06450 (2016)", + "3. Chan, T.F., Vese, L.A.: Active contours without edges. IEEE Transactions on image processing 10(2), 266-277 (2001)", + "4. Chen, C., Li, O., Tao, D., Barnett, A., Rudin, C., Su, J.K.: This looks like that: deep learning for interpretable image recognition. Advances in neural information processing systems 32 (2019)", + "5. Choudhury, S., Laina, I., Rupprecht, C., Vedaldi, A.: Unsupervised part discovery from contrastive reconstruction. Advances in Neural Information Processing Systems 34, 28104-28118 (2021)", + "6. Darcet, T., Oquab, M., Mairal, J., Bojanowski, P.: Vision transformers need registers. In: The Twelfth International Conference on Learning Representations (2023)", + "7. Dobson, D.C., Santosa, F.: Recovery of blocky images from noisy and blurred data. SIAM Journal on Applied Mathematics 56(4), 1181-1198 (1996)", + "8. Donoser, M., Urschler, M., Hirzer, M., Bischof, H.: Saliency driven total variation segmentation. In: ICCV. pp. 817-824. IEEE (2009)", + "9. Dosovitskiy, A., Beyer, L., Kolesnikov, A., Weissenborn, D., Zhai, X., Unterthiner, T., Dehghani, M., Minderer, M., Heigold, G., Gelly, S., et al.: An image is worth 16x16 words: Transformers for image recognition at scale. In: International Conference on Learning Representations (2020)", + "0. Gautam, S., Höhne, M.M.C., Hansen, S., Jenssen, R., Kampffmeyer, M.: This looks more like that: Enhancing self-explaining models by prototypical relevance propagation. Pattern Recognition 136, 109172 (2023)", + "1. He, J., Yang, S., Yang, S., Kortylewski, A., Yuan, X., Chen, J.N., Liu, S., Yang, C., Yu, Q., Yuille, A.: Partimagenet: A large, high-quality dataset of parts. In: European Conference on Computer Vision. pp. 128-145. Springer (2022)", + "2. Henaff, O.J., Koppula, S., Shelhamer, E., Zoran, D., Jaegle, A., Zisserman, A., Carreira, J., Arandjelovic, R.: Object discovery and representation networks. In: European Conference on Computer Vision. pp. 123-143. Springer (2022)", + "3. Huang, Z., Li, Y.: Interpretable and accurate fine-grained recognition via region grouping. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 8662-8672 (2020)", + "4. Hung, W.C., Jampani, V., Liu, S., Molchanov, P., Yang, M.H., Kautz, J.: SCOPS: Self-Supervised Co-Part Segmentation. In: IEEE Conf. on Computer Vision and Pattern Recognition (CVPR) (2019)", + "5. Jang, E., Gu, S., Poole, B.: Categorical reparameterization with gumbel-softmax. In: International Conference on Learning Representations (2017)", + "6. Johnson, J., Alahi, A., Fei-Fei, L.: Perceptual losses for real-time style transfer and super-resolution. In: Computer Vision-ECCV 2016: 14th European Conference, Amsterdam, The Netherlands, October 11-14, 2016, Proceedings, Part II 14. pp. 694-711. Springer (2016)", + "7. van der Klis, R., Alaniz, S., Mancini, M., Dantas, C.F., Ienco, D., Akata, Z., Marcos, D.: PDiscoNet: Semantically consistent part discovery for fine-grained recognition. In: Proceedings of the IEEE/CVF International Conference on Computer Vision. pp. 1866-1876 (2023)" + ], + "bbox": [ + 225, + 178, + 784, + 839 + ], + "page_idx": 14 + }, + { + "type": "header", + "text": "PDiscoFormer: Part Discovery with Transformers", + "bbox": [ + 401, + 114, + 730, + 128 + ], + "page_idx": 14 + }, + { + "type": "page_number", + "text": "15", + "bbox": [ + 767, + 116, + 784, + 126 + ], + "page_idx": 14 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "18. Marcos, D., Kierdorf, J., Cheeseman, T., Tuia, D., Roscher, R.: A whale's tail-finding the right whale in an uncertain world. In: International Workshop on Extending Explainable AI Beyond Deep Models and Classifiers. pp. 297-313. Springer (2020)", + "19. Mumford, D.B., Shah, J.: Optimal approximations by piecewise smooth functions and associated variational problems. Communications on pure and applied mathematics (1989)", + "20. Nauta, M., van Bree, R., Seifert, C.: Neural Prototype Trees for Interpretable Fine-Grained Image Recognition. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR). pp. 14933-14943 (6 2021)", + "21. Nauta, M., Jutte, A., Provoost, J., Seifert, C.: This looks like that, because... explaining prototypes for interpretable image recognition. In: Joint European Conference on Machine Learning and Knowledge Discovery in Databases. pp. 441-456. Springer (2021)", + "22. Nauta, M., Schlötterer, J., van Keulen, M., Seifert, C.: PIP-Net: Patch-Based Intuitive Prototypes for Interpretable Image Classification. Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (2023)", + "23. Nilsback, M.E., Zisserman, A.: Automated flower classification over a large number of classes. In: Indian Conference on Computer Vision, Graphics and Image Processing (Dec 2008)", + "24. Oquab, M., Darcet, T., Moutakanni, T., Vo, H., Szafraniec, M., Khalidov, V., Fernandez, P., Haziza, D., Massa, F., El-Nouby, A., Assran, M., Ballas, N., Galuba, W., Howes, R., Huang, P.Y., Li, S.W., Misra, I., Rabbat, M., Sharma, V., Synnaeve, G., Xu, H., Jegou, H., Mairal, J., Labatut, P., Joulin, A., Bojanowski, P.: DINOv2: Learning Robust Visual Features without Supervision. arXiv preprint arXiv:2304.07193 (2023), http://arxiv.org/abs/2304.07193", + "25. Ramaswamy, H.G., et al.: Ablation-cam: Visual explanations for deep convolutional network via gradient-free localization. In: proceedings of the IEEE/CVF winter conference on applications of computer vision. pp. 983-991 (2020)", + "26. Ranasinghe, K., McKinzie, B., Ravi, S., Yang, Y., Toshev, A., Shlens, J.: Perceptual grouping in contrastive vision-language models. In: Proceedings of the IEEE/CVF International Conference on Computer Vision. pp. 5571-5584 (2023)", + "27. Ribeiro, M.T., Singh, S., Guestrin, C.: \"Why Should I Trust You?\": Explaining the Predictions of Any Classifier. In: Proceedings of the 22nd ACM SIGKDD International Conference on Knowledge Discovery and Data Mining, San Francisco, CA, USA, August 13-17, 2016. pp. 1135-1144 (2016)", + "28. Rudin, C.: Stop explaining black box machine learning models for high stakes decisions and use interpretable models instead. Nature machine intelligence 1(5), 206-215 (2019)", + "29. Rudin, L.I., Osher, S., Fatemi, E.: Nonlinear total variation based noise removal algorithms. Physica D: nonlinear phenomena 60(1-4), 259-268 (1992)", + "30. Rymarczyk, D., Struski, Å., Tabor, J., Zielinski, B.: Protopshare: Prototypical parts sharing for similarity discovery in interpretable image classification. In: Proceedings of the 27th ACM SIGKDD Conference on Knowledge Discovery & Data Mining. pp. 1420-1430 (2021)", + "31. Selvaraju, R.R., Cogswell, M., Das, A., Vedantam, R., Parikh, D., Batra, D.: Gradcam: Visual explanations from deep networks via gradient-based localization. In: Proceedings of the IEEE international conference on computer vision. pp. 618-626 (2017)" + ], + "bbox": [ + 215, + 146, + 785, + 840 + ], + "page_idx": 15 + }, + { + "type": "page_number", + "text": "16", + "bbox": [ + 217, + 114, + 235, + 126 + ], + "page_idx": 15 + }, + { + "type": "header", + "text": "A. Aniraj et al.", + "bbox": [ + 271, + 114, + 375, + 128 + ], + "page_idx": 15 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "32. Shannon, C.E.: A Mathematical Theory of Communication. Bell System Technical Journal 27(3), 379-423 (1948). https://doi.org/10.1002/j.1538-7305.1948.tb01338.x", + "33. Simeoni, O., Puy, G., Vo, H.V., Roburin, S., Gidaris, S., Bursuc, A., Pérez, P., Marlet, R., Ponce, J.: Localizing objects with self-supervised transformers and no labels. arXiv preprint arXiv:2109.14279 (2021)", + "34. Srivastava, N., Hinton, G., Krizhevsky, A., Sutskever, I., Salakhutdinov, R.: Dropout: a simple way to prevent neural networks from overfitting. The journal of machine learning research 15(1), 1929-1958 (2014)", + "35. Strong, D., Chan, T.: Edge-preserving and scale-dependent properties of total variation regularization. Inverse problems 19(6), S165 (2003)", + "36. Unger, M., Pock, T., Trobin, W., Cremers, D., Bischof, H.: Tvseg-interactive total variation based image segmentation. In: BMVC. vol. 31, pp. 44-46. Citeseer (2008)", + "37. Wang, J., Liu, H., Wang, X., Jing, L.: Interpretable image recognition by constructing transparent embedding space. In: Proceedings of the IEEE/CVF International Conference on Computer Vision. pp. 895-904 (2021)", + "38. Welinder, P., Branson, S., Mita, T., Wah, C., Schroff, F., Belongie, S., Perona, P.: Caltech-UCSD Birds 200. Tech. Rep. CNS-TR-2010-001, California Institute of Technology (2010)", + "39. Xu, J., De Mello, S., Liu, S., Byeon, W., Breuel, T., Kautz, J., Wang, X.: Groupvit: Semantic segmentation emerges from text supervision. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 18134-18144 (2022)", + "40. Xu-Darme, R., Quenot, G., Chihani, Z., Rousset, M.C.: Particul: part identification with confidence measure using unsupervised learning. In: International Conference on Pattern Recognition. pp. 173-187. Springer (2022)", + "41. Yeh, R.A., Hu, Y.T., Ren, Z., Schwing, A.G.: Total variation optimization layers for computer vision. In: CVPR. pp. 711-721 (2022)" + ], + "bbox": [ + 212, + 146, + 787, + 535 + ], + "page_idx": 16 + }, + { + "type": "header", + "text": "PDiscoFormer: Part Discovery with Transformers", + "bbox": [ + 401, + 114, + 730, + 128 + ], + "page_idx": 16 + }, + { + "type": "page_number", + "text": "17", + "bbox": [ + 767, + 116, + 785, + 126 + ], + "page_idx": 16 + } +] \ No newline at end of file diff --git a/2024/PDiscoFormer_ Relaxing Part Discovery Constraints with Vision Transformers/aa33e6f8-0640-410e-98c0-983fb97aea81_model.json b/2024/PDiscoFormer_ Relaxing Part Discovery Constraints with Vision Transformers/aa33e6f8-0640-410e-98c0-983fb97aea81_model.json new file mode 100644 index 0000000000000000000000000000000000000000..0ceda122f16ac4779b1e54630469956fed485abe --- /dev/null +++ b/2024/PDiscoFormer_ Relaxing Part Discovery Constraints with Vision Transformers/aa33e6f8-0640-410e-98c0-983fb97aea81_model.json @@ -0,0 +1,1961 @@ +[ + [ + { + "type": "title", + "bbox": [ + 0.274, + 0.142, + 0.73, + 0.185 + ], + "angle": 0, + "content": "PDiscoFormer: Relaxing Part Discovery Constraints with Vision Transformers" + }, + { + "type": "text", + "bbox": [ + 0.227, + 0.213, + 0.776, + 0.229 + ], + "angle": 0, + "content": "Ananthu Aniraj\\(^{1}\\), Cassio F. Dantas\\(^{2}\\), Dino Ienc\\(^{2}\\), and Diego Marcos\\(^{1}\\)" + }, + { + "type": "text", + "bbox": [ + 0.265, + 0.24, + 0.737, + 0.268 + ], + "angle": 0, + "content": "1 Inria, Univ. Montpellier, LIRMM, UMR TETIS, Montpellier, France {ananthu.aniraj, diego.marcos}@inria.fr" + }, + { + "type": "text", + "bbox": [ + 0.274, + 0.269, + 0.727, + 0.296 + ], + "angle": 0, + "content": "2 Inria, Inrae, Univ. Montpellier, UMR TETIS, Montpellier, France {cassio.fraga-dantas, dino.ienco}@inrae.fr" + }, + { + "type": "list", + "bbox": [ + 0.265, + 0.24, + 0.737, + 0.296 + ], + "angle": 0, + "content": null + }, + { + "type": "text", + "bbox": [ + 0.261, + 0.334, + 0.744, + 0.612 + ], + "angle": 0, + "content": "Abstract. Computer vision methods that explicitly detect object parts and reason on them are a step towards inherently interpretable models. Existing approaches that perform part discovery driven by a fine-grained classification task make very restrictive assumptions on the geometric properties of the discovered parts; they should be small and compact. Although this prior is useful in some cases, in this paper we show that pre-trained transformer-based vision models, such as self-supervised DINOv2 ViT, enable the relaxation of these constraints. In particular, we find that a total variation (TV) prior, which allows for multiple connected components of any size, substantially outperforms previous work. We test our approach on three fine-grained classification benchmarks: CUB, PartImageNet and Oxford Flowers, and compare our results to previously published methods as well as a re-implementation of the state-of-the-art method PDiscoNet with a transformer-based backbone. We consistently obtain substantial improvements across the board, both on part discovery metrics and the downstream classification task, showing that the strong inductive biases in self-supervised ViT models require to rethink the geometric priors that can be used for unsupervised part discovery. Training code and pre-trained models are available at https://github.com/ananthu-aniraj/pdiscoformer." + }, + { + "type": "title", + "bbox": [ + 0.217, + 0.638, + 0.375, + 0.653 + ], + "angle": 0, + "content": "1 Introduction" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.669, + 0.789, + 0.701 + ], + "angle": 0, + "content": "Deep neural networks, first in the form of convolutional neural networks (CNN) and, more recently, vision transformers (ViT), have revolutionized the field of" + }, + { + "type": "image", + "bbox": [ + 0.218, + 0.726, + 0.787, + 0.813 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.316, + 0.822, + 0.684, + 0.837 + ], + "angle": 0, + "content": "Fig. 1: Concentration vs total variation as part priors." + } + ], + [ + { + "type": "page_number", + "bbox": [ + 0.218, + 0.116, + 0.23, + 0.127 + ], + "angle": 0, + "content": "2" + }, + { + "type": "header", + "bbox": [ + 0.272, + 0.115, + 0.378, + 0.129 + ], + "angle": 0, + "content": "A. Aniraj et al." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.147, + 0.788, + 0.254 + ], + "angle": 0, + "content": "computer vision. However, their black-box nature often limits their application in domains where interpretability is crucial. To ameliorate the understanding of the internal behaviour of deep learning models trained for computer vision tasks, post-hoc interpretability methods [25,27,31] are commonly used. These methods generate saliency maps indicating regions in the image that contributed most to a classification decision. However, these are only approximations to the actual internal process of the model and may be misleading [28]." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.262, + 0.788, + 0.385 + ], + "angle": 0, + "content": "Recent efforts have been directed towards developing interpretable-by-design deep neural networks [4, 13, 14, 17, 21, 22]. These approaches aim to identify semantically interpretable features in an image and utilize this information for solving a downstream task. Primarily tailored for fine-grained image classification, these methods seek to emulate human reasoning by subdividing the final task, such as classifying flower species or dog breeds, into multiple explicit intermediate stages, such as identifying lower-level semantic features like body parts, which in turn are used to provide a classification score." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.393, + 0.788, + 0.56 + ], + "angle": 0, + "content": "However, improving interpretability while maintaining performance remains a challenge. Some models, such as those proposed in [13, 14, 17, 18], rely on objective functions that impose priors on the shape and area of object parts such as compactness-inducing part shaping losses that, in practice, result in spatial Gaussian priors of fixed variance on the part maps. While effective for specific datasets, like the commonly used CUB [38] benchmark for bird species classification, where assumptions like parts occurring only once in an image and being compact are beneficial, these assumptions may not generalize well to other use cases. For instance, in plant species classification, the presence of multiple instances of object parts such as leaves or flowers, as well as irregularly-shaped parts like branches or stems, renders these assumptions inadequate." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.569, + 0.788, + 0.842 + ], + "angle": 0, + "content": "As we show in this work, such strong part shape priors are required when using a CNN-based backbone pre-trained on ImageNet; in order to reach acceptable classification results, it is imperative to fine-tune the whole backbone. This means that the only inductive bias that is kept throughout the part discovery process is the one brought in by the convolutional nature of the CNN architecture. A strong part shape prior is required in order to obtain consistent part maps in this setting, leading to all recent methods for part discovery in the context of fine-grained classification to use a concentration loss that makes part maps small and compact [13, 14, 17]. Here, we propose a method that addresses this issue by refraining from assuming any priors on part shape or area. The only prior we assume is that the part maps should tend to be piece-wise constant and thus be formed of one or more spatially connected components. Our model is capable of localizing up to a fixed number of semantically consistent parts of an object without relying on part annotations, using only the classification labels as supervision signals. The lack of assumptions about part shape and size makes our approach suitable to various fine-grained image classification tasks, as demonstrated by our empirical results, generalizing beyond cases where compactness is a good part prior." + } + ], + [ + { + "type": "header", + "bbox": [ + 0.402, + 0.115, + 0.731, + 0.129 + ], + "angle": 0, + "content": "PDiscoFormer: Part Discovery with Transformers" + }, + { + "type": "page_number", + "bbox": [ + 0.775, + 0.117, + 0.785, + 0.127 + ], + "angle": 0, + "content": "3" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.147, + 0.788, + 0.269 + ], + "angle": 0, + "content": "The main contribution of this paper is to build upon [17] and relax its constraining concentration-based part prior by a more flexible total variation prior. We investigate which additional elements are required to compensate for the flexibility of the total variation loss and find that a combination of an entropy loss and Gumbel-Softmax on the part maps allows for consistent and high quality part maps. Our approach makes better use of the representation of self-supervised pre-trained ViT models and allows to apply the model to tasks with diverse types of parts." + }, + { + "type": "title", + "bbox": [ + 0.216, + 0.292, + 0.389, + 0.309 + ], + "angle": 0, + "content": "2 Related Work" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.326, + 0.788, + 0.568 + ], + "angle": 0, + "content": "Unsupervised Part Discovery. Machine learning methods designed for unsupervised part discovery aim to identify a small number of semantically consistent parts, typically around \\( K = 8 \\), shared among classes in a dataset. These methods visualize all discovered parts in an image through per-part saliency maps, facilitating easier interpretation. However, as this task is unsupervised, it often relies on assumptions about the discovered parts, such as their distribution and shape. For instance, Huang et al. [13] assume that the presence of each part across a mini-batch follows a beta distribution, while [17] assumes that discovered parts are present at least once in a mini-batch. Additionally, assumptions are often made about the shape of the parts. Methods such as [14, 17, 18, 40] impose a spatial Gaussian prior with a fixed variance on the part maps, resulting in compact parts appearing once in the image. Such assumptions on part shape can be highly restrictive and may not generalize well across different scenarios. To address this limitation, our proposed method refrains from assuming such priors on part shape and only imposes a spatial prior that discovered parts should form connected components." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.568, + 0.788, + 0.78 + ], + "angle": 0, + "content": "Part Prototype Networks. Chen et al. [4] introduced the Prototypical Part Network (ProtoPNet), an inherently interpretable method featuring a fixed number of learned prototypes per class. These prototypes represent visual patterns or features learned by the network during training. The model computes the similarity between each learned prototype and image patches and uses a linear combination of these similarity scores for classification. While explanations provided by ProtoPNet visualize all prototypes along with their weighted similarity scores, the sheer number of prototypes, often proportional to the number of classes, can lead to redundancy and interpretability challenges [10,21]. Subsequent works in this line of research have sought to address these challenges. Some approaches focus on reducing the number of prototypes shared per class [30,37], while others explore sharing prototypes between classes [20, 22]. Despite these efforts, recent works like [22] still utilize a large number (\\(K > 500\\)) of prototypes, which can pose significant challenges for end-user interpretation." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.78, + 0.788, + 0.841 + ], + "angle": 0, + "content": "Regularization in Image Segmentation. Early works on image segmentation made direct use of low-level image features, chiefly color, along with geometric priors, in order to find coherent segments. The Mumford-Shah model [19], originally devised for piece-wise image reconstruction, was applied to find segments" + } + ], + [ + { + "type": "page_number", + "bbox": [ + 0.218, + 0.116, + 0.23, + 0.127 + ], + "angle": 0, + "content": "4" + }, + { + "type": "header", + "bbox": [ + 0.272, + 0.115, + 0.378, + 0.129 + ], + "angle": 0, + "content": "A. Aniraj et al." + }, + { + "type": "image", + "bbox": [ + 0.251, + 0.147, + 0.782, + 0.384 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.3, + 0.394, + 0.704, + 0.41 + ], + "angle": 0, + "content": "Fig. 2: The PDiscoFormer architecture for part discovery." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.439, + 0.788, + 0.575 + ], + "angle": 0, + "content": "that are as homogeneous as possible while trying to minimize the length of their boundary as a regularization [3]. Total variation (TV) [29] has been shown to be a powerful and efficient regularization to approximate this boundary length prior, and has been extensively used for piece-wise signal and image reconstruction [7,35], as well as unsupervised image segmentation [8,36]. With the advent of deep learning, TV regularization has proven valuable for style transfer [16] and edge-preserving smoothing [41]. In this work we explore the use of TV regularization for weakly-supervised part discovery in order to alleviate the issues brought by previously used priors." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.575, + 0.788, + 0.682 + ], + "angle": 0, + "content": "Emergent grouping of ViT representations. Some recent research [26,39] has studied the emergence of object segmentation by supervising with image-text pairs. These methods learn to group similar patches in a ViT and assign them to textual concepts. They aim to identify the pixels in an input image corresponding to a predefined text prompt, focusing on segmenting whole objects rather than decomposing them into parts, as our model does with part discovery, and are limited to concepts that are represented in the text annotations." + }, + { + "type": "title", + "bbox": [ + 0.216, + 0.703, + 0.381, + 0.722 + ], + "angle": 0, + "content": "3 Methodology" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.734, + 0.788, + 0.842 + ], + "angle": 0, + "content": "Here we introduce an end-to-end trainable method (see Fig. 2) capable of automatically identifying \\( K \\) discriminative parts (semantically interpretable image regions) useful for solving a fine-grained image classification task. In contrast to earlier methods [13,17], our approach avoids any assumptions related to part shape, size or number of part instances. As in [13,17], our model is weakly supervised and relies solely on image-level class labels, thus uncovering discriminative parts without requiring any additional supervision. Our formulation is based" + } + ], + [ + { + "type": "header", + "bbox": [ + 0.402, + 0.115, + 0.733, + 0.13 + ], + "angle": 0, + "content": "PDiscoFormer: Part Discovery with Transformers" + }, + { + "type": "page_number", + "bbox": [ + 0.775, + 0.117, + 0.785, + 0.127 + ], + "angle": 0, + "content": "5" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.147, + 0.784, + 0.207 + ], + "angle": 0, + "content": "on [17], from which we take take most elements of the architecture, except for the Gumbel-Softmax and the layer normalization-based part modulation. The total variation, entropy and background losses have not been used in previous part discovery works and are shown to be important in our results." + }, + { + "type": "title", + "bbox": [ + 0.216, + 0.234, + 0.421, + 0.248 + ], + "angle": 0, + "content": "3.1 Model architecture" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.261, + 0.784, + 0.305 + ], + "angle": 0, + "content": "Let \\(\\mathbf{x} \\in \\mathbb{R}^{3 \\times M \\times N}\\) be an image in the dataset, and let \\(y \\in \\{1, 2, \\dots, C\\}\\) represent the corresponding classification label. Given a backbone model \\(h_{\\theta}(\\cdot)\\), we extract a feature tensor:" + }, + { + "type": "equation", + "bbox": [ + 0.423, + 0.308, + 0.785, + 0.326 + ], + "angle": 0, + "content": "\\[\n\\mathbf {z} = h _ {\\theta} (\\mathbf {x}) \\in \\mathbb {R} ^ {D \\times H \\times W} \\tag {1}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.335, + 0.785, + 0.455 + ], + "angle": 0, + "content": "We drop the class and register tokens of the ViT and reshape the patch tokens to obtain \\(\\mathbf{z}\\). Following prior research [13, 14, 17], we compute attention maps \\(\\mathbf{A} \\in [0,1]^{(K + 1) \\times H \\times W}\\), where the last channel in the first dimension is used to represent the background region in the image. These attention maps are computed using the negative squared Euclidean distance function between patch token features \\(\\mathbf{z}_{ij} \\in \\mathbb{R}^D\\), \\(i \\in \\{1, \\dots, H\\}\\), \\(j \\in \\{1, \\dots, W\\}\\) and each of the learnable prototypes \\(\\mathbf{p}^k \\in \\mathbb{R}^D\\), with \\(k \\in \\{1, \\dots, K + 1\\}\\), followed by a Gumbel-Softmax [15] across the \\(K + 1\\) channels:" + }, + { + "type": "equation", + "bbox": [ + 0.373, + 0.471, + 0.785, + 0.508 + ], + "angle": 0, + "content": "\\[\na _ {i j} ^ {k} = \\frac {\\exp \\left(- \\| \\mathbf {z} _ {i j} - \\mathbf {p} ^ {k} \\| ^ {2} + \\gamma_ {k}\\right)}{\\sum_ {l = 1} ^ {K + 1} \\exp \\left(- \\| \\mathbf {z} _ {i j} - \\mathbf {p} ^ {l} \\| ^ {2} + \\gamma_ {l}\\right)} \\tag {2}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.514, + 0.784, + 0.545 + ], + "angle": 0, + "content": "where \\(\\gamma_{k}\\) and \\(\\gamma_{l}\\) are independently drawn samples from the Gumbel(0,1) distribution, and \\(a_{ij}^{k}\\) represents an element of \\(\\mathbf{A}\\)." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.546, + 0.784, + 0.59 + ], + "angle": 0, + "content": "The attention maps are used to compute part embedding vectors \\(\\mathbf{v}^k\\in \\mathbb{R}^D\\) by using the attention values to calculate a weighted average over the feature tensor \\(\\mathbf{z}\\):" + }, + { + "type": "equation", + "bbox": [ + 0.434, + 0.607, + 0.785, + 0.639 + ], + "angle": 0, + "content": "\\[\n\\mathbf {v} ^ {k} = \\frac {\\sum_ {i} \\sum_ {j} a _ {i j} ^ {k} \\mathbf {z} _ {i j}}{H W} \\tag {3}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.647, + 0.784, + 0.692 + ], + "angle": 0, + "content": "These embedding vectors are then modulated using a layer normalization [2] operation, with normalization applied over the feature and prototype dimensions. This operation helps to make each embedding vector distinct for the classifier:" + }, + { + "type": "equation", + "bbox": [ + 0.386, + 0.706, + 0.785, + 0.742 + ], + "angle": 0, + "content": "\\[\n\\mathbf {v} _ {m} ^ {k} = \\frac {\\left(\\mathbf {v} ^ {k} - \\operatorname {E} \\left[ \\mathbf {v} ^ {k} \\right]\\right)}{\\sqrt {\\operatorname {V a r} \\left[ \\mathbf {v} ^ {k} \\right] + \\epsilon}} \\odot \\mathbf {w} _ {m} ^ {k} + \\mathbf {b} _ {m} ^ {k} \\tag {4}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.75, + 0.784, + 0.811 + ], + "angle": 0, + "content": "with \\(\\mathbf{w}_m^k, \\mathbf{b}_m^k \\in \\mathbb{R}^D\\) being the modulation weights and biases per part. These modulated embedding vectors \\(\\mathbf{v}_m^k\\) for \\(k \\in \\{1, \\dots, K\\}\\) (we omit the background embedding) are then linearly projected to obtain a vector of class scores conditioned on the part embedding:" + }, + { + "type": "equation", + "bbox": [ + 0.453, + 0.825, + 0.785, + 0.842 + ], + "angle": 0, + "content": "\\[\n\\mathbf {y} ^ {k} = W _ {c} \\cdot \\mathbf {v} _ {m} ^ {k} \\tag {5}\n\\]" + } + ], + [ + { + "type": "page_number", + "bbox": [ + 0.218, + 0.116, + 0.23, + 0.127 + ], + "angle": 0, + "content": "6" + }, + { + "type": "header", + "bbox": [ + 0.272, + 0.115, + 0.378, + 0.129 + ], + "angle": 0, + "content": "A. Aniraj et al." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.146, + 0.788, + 0.193 + ], + "angle": 0, + "content": "The same linear classifier \\( W_{c} \\in \\mathbb{R}^{C \\times D} \\) is applied to each of the modulated embedding vectors, i.e., for each of the \\( K \\) discovered foreground parts. We then take the mean of the per-part class scores to obtain the classification scores:" + }, + { + "type": "equation", + "bbox": [ + 0.451, + 0.205, + 0.786, + 0.245 + ], + "angle": 0, + "content": "\\[\n\\mathbf {y} = \\frac {1}{K} \\sum_ {k = 1} ^ {K} \\mathbf {y} ^ {k} \\tag {6}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.215, + 0.25, + 0.788, + 0.34 + ], + "angle": 0, + "content": "Finally, a softmax function is employed to obtain the classification probabilities. Part dropout. We adopt dropout [34] on the modulated embedding vectors \\(\\mathbf{v}_m^k\\), as proposed in [17], during model training. This drops entire part embedding vectors at a time and ensures that all discovered parts are discriminative, since discovering a single discriminative part would not be enough to solve the downstream classification task." + }, + { + "type": "title", + "bbox": [ + 0.216, + 0.361, + 0.386, + 0.375 + ], + "angle": 0, + "content": "3.2 Loss Functions" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.385, + 0.787, + 0.429 + ], + "angle": 0, + "content": "Our main contribution consists in proposing a set of part-shaping loss functions that are able to make the most out of the inductive biases learned by self-supervised ViT models." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.431, + 0.788, + 0.55 + ], + "angle": 0, + "content": "Classification Loss \\((\\mathcal{L}_{\\mathbf{c}})\\). Our model is mainly driven by the fine-grained classification objective during training. We use the standard negative log likelihood or multi-class cross-entropy for this purpose. While the classification loss helps ensure the discriminative nature of the discovered parts, it may not directly enforce the learned attention maps to represent salient parts of the object in the image. To address this, we incorporate additional objective functions to enforce priors on the discovered parts. These objectives effectively guide the learning process to focus on relevant regions of the image." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.551, + 0.788, + 0.629 + ], + "angle": 0, + "content": "Orthogonality Loss \\((\\mathcal{L}_{\\perp})\\). We aim to discover parts that are semantically distinct. To achieve this, we apply the orthogonality loss, which encourages the learned part embedding vectors to be decorrelated from one another. This is accomplished by minimizing the cosine similarity between the embedding vectors. We apply this loss on the modulated embedding vectors \\(\\mathbf{v}_m^k\\)." + }, + { + "type": "equation", + "bbox": [ + 0.403, + 0.639, + 0.786, + 0.679 + ], + "angle": 0, + "content": "\\[\n\\mathcal {L} _ {\\perp} = \\sum_ {k = 1} ^ {K + 1} \\sum_ {l \\neq k} \\frac {\\mathbf {v} _ {m} ^ {k} \\cdot \\mathbf {v} _ {m} ^ {l}}{\\| \\mathbf {v} _ {m} ^ {k} \\| \\cdot \\| \\mathbf {v} _ {m} ^ {l} \\|} \\tag {7}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.689, + 0.789, + 0.841 + ], + "angle": 0, + "content": "Equivalence Loss \\((\\mathcal{L}_{\\mathrm{eq}})\\). We aim to detect the same parts even if the image is translated, rotated, or scaled. To achieve this, we formulate the equivariance loss, which encourages the learned part attention maps to be equivariant to rigid transformations. Specifically, given an input image, we use an affine transform \\(T\\) with parameters randomly sampled from a pre-defined range and pass both the original image and the transformed image through our model. Then, we invert the transformation on the part attention maps of the transformed image. Finally, the equivariance loss is formulated as the cosine distance between the part attention maps of the original image and those of the corresponding transformed image." + } + ], + [ + { + "type": "header", + "bbox": [ + 0.402, + 0.115, + 0.733, + 0.129 + ], + "angle": 0, + "content": "PDiscoFormer: Part Discovery with Transformers" + }, + { + "type": "page_number", + "bbox": [ + 0.775, + 0.116, + 0.787, + 0.127 + ], + "angle": 0, + "content": "7" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.146, + 0.784, + 0.178 + ], + "angle": 0, + "content": "Let \\( A^k(\\mathbf{x}) \\) be a function that returns the \\( k^{th} \\) attention map given an input image \\( \\mathbf{x} \\). Then:" + }, + { + "type": "equation", + "bbox": [ + 0.347, + 0.188, + 0.786, + 0.228 + ], + "angle": 0, + "content": "\\[\n\\mathcal {L} _ {\\mathrm {e q}} = 1 - \\frac {1}{K} \\sum_ {k} \\frac {\\left\\| A ^ {k} (\\mathbf {x}) \\cdot T ^ {- 1} \\left(A ^ {k} (T (\\mathbf {x})) \\right. \\right\\|}{\\| A ^ {k} (\\mathbf {x}) \\| \\cdot \\| A ^ {k} (T (\\mathbf {x})) \\|} \\tag {8}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.238, + 0.785, + 0.345 + ], + "angle": 0, + "content": "Presence Loss \\((\\mathcal{L}_{\\mathbf{p}})\\). We assume that all discovered foreground parts are present in at least some images of the training dataset, while the background, the \\((K + 1)^{th}\\) part, is expected to be present in all images in the dataset. To enforce this, we formulate the presence loss, which consists of two parts. Firstly, we define a loss applied only to the discovered foreground parts, ensuring that each part is present at least once in a mini-batch. Given a batch \\(\\{\\mathbf{x}_1,\\dots ,\\mathbf{x}_B\\}\\), the presence loss is calculated as:" + }, + { + "type": "equation", + "bbox": [ + 0.399, + 0.355, + 0.786, + 0.391 + ], + "angle": 0, + "content": "\\[\n\\mathcal {L} _ {\\mathrm {p} _ {1}} = 1 - \\frac {1}{K} \\sum_ {k} \\max _ {b, i, j} \\bar {a} _ {i j} ^ {k} \\left(\\mathbf {x} _ {b}\\right) \\tag {9}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.402, + 0.784, + 0.434 + ], + "angle": 0, + "content": "where \\(\\bar{A}^k (\\mathbf{x}_b) = \\mathrm{avgpool}(A^k (\\mathbf{x}_b))\\) is the output of a 2D average pooling operator with a small kernel size employed to prevent single-pixel attention maps." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.435, + 0.784, + 0.478 + ], + "angle": 0, + "content": "Secondly, to ensure the presence of the background, we introduce a stricter presence loss, which is satisfied if the background is present in all images of the mini-batch. This loss is formulated as follows:" + }, + { + "type": "equation", + "bbox": [ + 0.359, + 0.488, + 0.786, + 0.524 + ], + "angle": 0, + "content": "\\[\n\\mathcal {L} _ {\\mathrm {p} _ {0}} = - \\frac {1}{B} \\sum_ {b} \\log \\left(\\max _ {i, j} m _ {i j} \\bar {a} _ {i j} ^ {K + 1} (\\mathbf {x} _ {b})\\right) \\tag {10}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.536, + 0.784, + 0.569 + ], + "angle": 0, + "content": "where \\( M = [m_{ij}]^{H \\times W} \\) is a soft mask with \\( m_{ij} \\in [0,1] \\) that privileges entries placed farther from the image center (towards the borders):" + }, + { + "type": "equation", + "bbox": [ + 0.345, + 0.579, + 0.786, + 0.615 + ], + "angle": 0, + "content": "\\[\nm _ {i j} = 2 \\left(\\frac {i - 1}{H - 1} - \\frac {1}{2}\\right) ^ {2} + 2 \\left(\\frac {j - 1}{W - 1} - \\frac {1}{2}\\right) ^ {2} \\tag {11}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.624, + 0.784, + 0.668 + ], + "angle": 0, + "content": "This loss captures the notion that the background is at the same time expected in every image and is more likely to appear near the boundaries of the image." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.669, + 0.784, + 0.716 + ], + "angle": 0, + "content": "Entropy Loss \\((\\mathcal{L}_{\\mathrm{ent}})\\). We aim to ensure that each patch token is assigned to a unique part. To achieve this, we formulate an entropy [32] loss applied to the part attention maps:" + }, + { + "type": "equation", + "bbox": [ + 0.382, + 0.727, + 0.786, + 0.769 + ], + "angle": 0, + "content": "\\[\n\\mathcal {L} _ {\\text {e n t}} = \\frac {- 1}{K + 1} \\sum_ {k = 1} ^ {K + 1} \\sum_ {i j} a _ {i j} ^ {k} \\log \\left(a _ {i j} ^ {k}\\right) \\tag {12}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.78, + 0.786, + 0.842 + ], + "angle": 0, + "content": "Total Variation Loss \\((\\mathcal{L}_{\\mathrm{tv}})\\). We do not want to make any assumptions about the part shape. However, we would still like the discovered parts to be composed of one or, at most, a few connected components. To achieve this, we make use of a total variation loss [29] applied to the part attention maps:" + } + ], + [ + { + "type": "page_number", + "bbox": [ + 0.218, + 0.116, + 0.23, + 0.127 + ], + "angle": 0, + "content": "8" + }, + { + "type": "header", + "bbox": [ + 0.272, + 0.115, + 0.378, + 0.129 + ], + "angle": 0, + "content": "A. Aniraj et al." + }, + { + "type": "equation", + "bbox": [ + 0.407, + 0.159, + 0.786, + 0.201 + ], + "angle": 0, + "content": "\\[\n\\mathcal {L} _ {\\mathrm {t v}} = \\frac {1}{H W} \\sum_ {k = 1} ^ {K + 1} \\sum_ {i j} \\left| \\nabla a _ {i j} ^ {k} \\right| \\tag {13}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.216, + 0.206, + 0.731, + 0.224 + ], + "angle": 0, + "content": "where \\(\\nabla a_{ij}^{k}\\) is the spatial image gradient of part map \\(A^k\\) in location \\(ij\\)." + }, + { + "type": "title", + "bbox": [ + 0.216, + 0.242, + 0.449, + 0.26 + ], + "angle": 0, + "content": "4 Experimental Setup" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.271, + 0.788, + 0.375 + ], + "angle": 0, + "content": "We aim to perform part discovery using only image-level class labels as supervision signals, assuming that the discovered parts are shared among classes, which is common in fine-grained image classification settings. Moreover, we aim to compare our model with similar methods in the literature in scenarios where they have been empirically shown to work well, as well as in new and more challenging scenarios where underlying data characteristics may require more flexible geometric modeling." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.376, + 0.788, + 0.438 + ], + "angle": 0, + "content": "All models were implemented in PyTorch. We used the DinoV2 [24] ViT [9] model with register tokens [6] as the starting weights for all experiments, utilizing the Base (ViT-B) variant. The training settings and compute requirements of our model are detailed in the supplementary material." + }, + { + "type": "title", + "bbox": [ + 0.216, + 0.456, + 0.335, + 0.47 + ], + "angle": 0, + "content": "4.1 Datasets" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.478, + 0.788, + 0.629 + ], + "angle": 0, + "content": "CUB Dataset. The CUB dataset [38] consists of images depicting 200 bird species, with 5,994 images in the training set and 5,794 images in the testing set. Keypoint annotations for 15 different bird body parts are available for each image. Bird parts are typically compact, although their presence may vary in images due to changes in pose and occlusion. Most images contain only a single instance of a bird. For this dataset, we evaluate the results using keypoint regression (Kp) using ground truth part keypoint annotations to compare them with centroids discovered by the model. Furthermore, we compute the Normalized Mutual Information (NMI) and Adjusted Rand Index (ARI) to quantify the clustering quality of the discovered parts, as in related literature [5]." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.63, + 0.789, + 0.841 + ], + "angle": 0, + "content": "PartImageNet Dataset. The PartImageNet dataset [11] comprises 158 classes distributed among 11 super-classes, such as airplanes, birds, and snakes. Each super-class is associated to between two and five possible parts, for a total of 41 part classes. Additionally, test images may contain multiple instances of objects, sometimes from different super-classes, while the classification label provides information about a single object. Notably, several super-classes, including snakes, airplanes, and reptiles, often exhibit irregular and non-compact part shapes. For our experiments, we use the PartImageNet 00D variant. Although originally intended for out-of-distribution detection, this variant is commonly used in related literature. We train our models on 14,865 training images and test on 1,658 images, respectively. For evaluation, we derive part centroids from ground truth part masks and calculate NMI and ARI scores to assess clustering quality. Additionally, we include an experiment on the newly released PartImageNet Seg variant in the supplementary material." + } + ], + [ + { + "type": "header", + "bbox": [ + 0.402, + 0.115, + 0.731, + 0.129 + ], + "angle": 0, + "content": "PDiscoFormer: Part Discovery with Transformers" + }, + { + "type": "page_number", + "bbox": [ + 0.775, + 0.117, + 0.785, + 0.127 + ], + "angle": 0, + "content": "9" + }, + { + "type": "table_caption", + "bbox": [ + 0.217, + 0.145, + 0.788, + 0.187 + ], + "angle": 0, + "content": "Table 1: Quantitative comparison of the part discovery and classification capabilities of our model with related approaches from the literature. In this table, \\( K \\) refers to the number of parts. The method with the label ** does not use any class supervision." + }, + { + "type": "table", + "bbox": [ + 0.217, + 0.199, + 0.787, + 0.449 + ], + "angle": 0, + "content": "
MethodCUB(%)PartImageNet OOD(%)Flowers(%)
KKp ↓NMI ↑ARI ↑Top-1 Acc. ↑KNMI ↑ARI ↑Top-1 Acc. ↑KFg. mIoU ↑Top-1 Acc. ↑
Dino [1]**4-31.1811.21-819.177.59-254.95-
8-47.2119.76-2531.4614.16-455.11-
16-50.5726.14-5037.8116.50-854.44-
Huang [13]411.5129.7414.0487.3085.881.5374.22229.9293.07
811.6035.7215.9086.05257.561.2573.56433.2293.14
1612.6043.9221.1085.935010.191.0573.20817.2692.86
PDiscoNet [17]49.1237.8215.2686.17827.138.7688.58219.0477.51
88.5250.0826.9686.722532.4110.6989.00434.7683.05
167.6756.8738.0587.495041.4914.1786.06849.1081.04
PdiscoNet + ViT-B [6]47.7052.5926.6688.61819.2834.7290.9524.9292.75
86.3465.0137.9086.952528.2350.3590.2941.9595.48
165.9568.6343.4184.045029.4827.8089.69813.1897.40
PDiscoFormer (Ours)47.4158.1325.1189.06829.0052.4089.75273.6299.61
85.9969.8743.4988.792544.7159.2790.77473.3299.54
165.7473.3855.8388.725046.2962.2191.01869.5999.64
" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.476, + 0.788, + 0.628 + ], + "angle": 0, + "content": "Oxford 102 Flower Dataset. The Oxford 102 Flower Dataset [23] comprises 8,189 images divided into 102 flower classes. The training and validation sets contain 10 images per class, totaling 1,020 images each, while the test set consists of 6,149 images, with a minimum of 20 images per class. Since no part annotations are provided for this dataset, we evaluate the quality of the discovered parts by combining them and comparing the result to the provided foreground/background masks. We evaluate model performance by calculating the mean Intersection over Union (mIoU) for the foreground mask. This dataset often includes multiple instances of parts in a single image, such as flowers, as well as parts with irregular shapes, such as flower petals." + }, + { + "type": "title", + "bbox": [ + 0.216, + 0.65, + 0.377, + 0.667 + ], + "angle": 0, + "content": "5 Experiments" + }, + { + "type": "title", + "bbox": [ + 0.216, + 0.68, + 0.542, + 0.696 + ], + "angle": 0, + "content": "5.1 Comparison with State-of-the-Art" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.704, + 0.787, + 0.841 + ], + "angle": 0, + "content": "We conduct a quantitative analysis of our model's part discovery and classification capabilities and compare it with those of related methods from the literature. Additionally, we compare our model against a re-implementation of the state-of-the-art PDiCoNet [17], where we replaced the backbone with the variant used in this paper. The results of this comparison are summarized in Tab. 1. From the results, our method consistently outperforms existing approaches for both part discovery and image classification across all datasets. We attribute this to the synergistic effects of the strong inductive biases ingrained in the self-supervised ViT backbone and the enhanced flexibility of the proposed combina" + } + ], + [ + { + "type": "page_number", + "bbox": [ + 0.218, + 0.116, + 0.236, + 0.127 + ], + "angle": 0, + "content": "10" + }, + { + "type": "header", + "bbox": [ + 0.272, + 0.115, + 0.378, + 0.129 + ], + "angle": 0, + "content": "A. Aniraj et al." + }, + { + "type": "image", + "bbox": [ + 0.218, + 0.144, + 0.797, + 0.33 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.346, + 0.344, + 0.656, + 0.358 + ], + "angle": 0, + "content": "Fig. 3: Qualitative results on CUB for \\( K = 8 \\)." + }, + { + "type": "image", + "bbox": [ + 0.218, + 0.359, + 0.797, + 0.545 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.316, + 0.558, + 0.687, + 0.573 + ], + "angle": 0, + "content": "Fig. 4: Qualitative results on PartImageNet for \\( K = 8 \\)." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.614, + 0.789, + 0.842 + ], + "angle": 0, + "content": "tion of geometric priors. Particularly noteworthy are the significant performance gains observed on the Flowers and PartImageNet datasets, which contain objects with irregular shapes occurring in multiple instances within the same image, underscoring the importance of our geometric part prior that favors piece-wise constant and spatially connected components. Even on CUB, where most parts fit well into the compactness prior, we see a consistent improvements of results with respect to PDiscoNet, even when using the same ViT backbone. With 16 parts, our model achieves \\(55.8\\%\\) ARI and \\(73.4\\%\\) NMI, compared to \\(43.4\\%\\) ARI and \\(68.6\\%\\) NMI of the best competing method, PDiscoNet with a ViT backbone. Despite competitive performance in classification tasks on CUB and PartImageNet, the PdiscoNet+ViT-B model tends to degrade in accuracy as part complexity increases, particularly in CUB, unlike our model, which maintains robust accuracy for any number of parts. In the Flowers dataset, all competing methods result in subpar performance, while our approach excels at both detecting the foreground and classification no matter the number of parts. In this" + } + ], + [ + { + "type": "header", + "bbox": [ + 0.403, + 0.115, + 0.732, + 0.129 + ], + "angle": 0, + "content": "PDiscoFormer: Part Discovery with Transformers" + }, + { + "type": "page_number", + "bbox": [ + 0.769, + 0.117, + 0.784, + 0.127 + ], + "angle": 0, + "content": "11" + }, + { + "type": "image", + "bbox": [ + 0.218, + 0.144, + 0.797, + 0.33 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.338, + 0.344, + 0.663, + 0.357 + ], + "angle": 0, + "content": "Fig. 5: Qualitative results on Flowers for \\( K = 2 \\)." + }, + { + "type": "image", + "bbox": [ + 0.218, + 0.358, + 0.797, + 0.543 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.216, + 0.558, + 0.788, + 0.601 + ], + "angle": 0, + "content": "Fig. 6: Effect of \\( K \\) on PDiscoFormer. The left three images correspond to Flowers (from top to bottom, \\( K = 2,4,8 \\)). The three middle ones are CUB, with \\( K \\) being 4, 8 and 16. The three on the right belong to PartImageNet, with \\( K = 8,25,50 \\)." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.639, + 0.788, + 0.776 + ], + "angle": 0, + "content": "dataset, PDiscoNet+ViT-B results in the second best classification accuracy, \\(97.4\\%\\), after the \\(99.6\\%\\) obtained by our method, but completely fails to locate the foreground, with an mIoU of \\(13.2\\%\\), against \\(73.6\\%\\) for PDiscoFormer. Additionally, we extend our comparisons by applying our proposed losses to a ResNet backbone and a fully frozen ViT architecture, showing that a pre-trained ViT is required to fully profit from the proposed set of losses and that fine-tuning the position, class and register tokens, while keeping the rest of the ViT frozen, results in better performances than leaving the whole ViT frozen. The results of these additional comparisons are presented in the supplementary material." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.78, + 0.788, + 0.842 + ], + "angle": 0, + "content": "Qualitative Analysis. A qualitative analysis, as illustrated in Fig. 3, Fig. 4, Fig. 5 and Fig. 6 corroborates our quantitative findings, showcasing our model's ability to accurately discover irregularly shaped body parts such as bird wings (CUB), flower petals (Flowers) and the body of a snake (PartImageNet). At the" + } + ], + [ + { + "type": "page_number", + "bbox": [ + 0.218, + 0.116, + 0.236, + 0.127 + ], + "angle": 0, + "content": "12" + }, + { + "type": "header", + "bbox": [ + 0.272, + 0.115, + 0.378, + 0.129 + ], + "angle": 0, + "content": "A. Aniraj et al." + }, + { + "type": "table_caption", + "bbox": [ + 0.277, + 0.145, + 0.725, + 0.159 + ], + "angle": 0, + "content": "Table 2: Results of the ablation study in PartImageNet and CUB." + }, + { + "type": "table", + "bbox": [ + 0.262, + 0.171, + 0.744, + 0.371 + ], + "angle": 0, + "content": "
PartImageNet OOD (K=25)CUB (K=16)
NMI ↑ARI ↑Top-1 Acc. ↑Kp ↓NMI ↑ARI ↑Top-1 Acc. ↑
Full Model44.7159.2790.775.7473.3855.8388.72
No Lp041.0549.0889.815.8568.6643.0983.64
No Lent39.1654.4691.195.5766.8845.6888.21
No Leq43.3854.1990.479.0556.8530.6383.71
No L⊥42.4458.8691.015.7970.5550.9088.63
No Lp133.9858.3191.077.1461.0434.2889.04
No Ltv33.7325.3590.476.0170.7851.8380.22
No Per-Part LayerNorm31.1440.7389.695.7470.7050.6269.73
No Part Dropout33.0843.9889.996.0870.2841.6987.14
No Gumbel32.9854.6890.716.4468.9243.6186.43
" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.398, + 0.788, + 0.686 + ], + "angle": 0, + "content": "same time, PDiscoNet [17] fails to discover these parts altogether, even when retrained on a ViT-B backbone. This indicates the usefulness of the flexible geometric priors enabled by our total variation loss. Compared to PDiscoNet [17], not only do we achieve a much tighter coverage of the object of interest, with our PDiscoFormer more closely following the object boundaries. This is particularly the case in parts that are relatively large, such as open bird wings and flower corollas. In addition to that, we can observe a substantial improvement in the consistency and semantic interpretability of parts. For instance, the parts discovered by our model for the four quadruped images displayed in Fig. 4 can immediately be interpreted (the green part corresponds to the snout while the brown one to the ears), while those from PDiscoNet are not clear. In Fig. 6 we vary the number of parts, \\( K \\). On Flowers, using two parts leads to a clear assignment of one part to the flower calyx and one to the corolla, while more parts tends to exhibit oversegmentation. CUB, on the other hand, is able to absorb a larger number of parts. With \\( K = 4 \\), beak and legs are covered by the same part, while a fine-grained and consistent distinction of the different wing parts is achieved with \\( K = 16 \\). Interestingly, in the case of PartImageNet, the effective number of parts being used by a specific super-class remains almost constant when increasing \\( K \\)." + }, + { + "type": "title", + "bbox": [ + 0.216, + 0.709, + 0.39, + 0.724 + ], + "angle": 0, + "content": "5.2 Ablation Study" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.734, + 0.788, + 0.843 + ], + "angle": 0, + "content": "We conducted an ablation study by systematically removing one loss function or architectural element at a time and by replacing the Gumbel-Softmax with a regular Softmax. This was performed on the PartImageNet OOD dataset with \\( K = 25 \\) parts and the CUB dataset with \\( K = 16 \\) parts. The quantitative results of the ablation study are shown in Tab. 2 and some qualitative results can be seen in Fig. 7. From the results, we can observe the following: Total Variation \\( (\\mathcal{L}_{tv}) \\) prevents spurious part activations in the background, possibly reducing" + } + ], + [ + { + "type": "header", + "bbox": [ + 0.403, + 0.115, + 0.732, + 0.129 + ], + "angle": 0, + "content": "PDiscoFormer: Part Discovery with Transformers" + }, + { + "type": "page_number", + "bbox": [ + 0.769, + 0.116, + 0.786, + 0.127 + ], + "angle": 0, + "content": "13" + }, + { + "type": "image", + "bbox": [ + 0.22, + 0.143, + 0.797, + 0.515 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.33, + 0.529, + 0.673, + 0.544 + ], + "angle": 0, + "content": "Fig. 7: Qualitative results from the ablation study." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.599, + 0.788, + 0.841 + ], + "angle": 0, + "content": "information leakage between foreground parts and the background. It has the largest impact on PartImageNet in terms of ARI. The absence of the Background Loss \\((\\mathcal{L}_{\\mathbf{p}_0})\\) results in the background part not being activated, causing the model to discover the background as a foreground part. Presence Loss \\((\\mathcal{L}_{\\mathbf{p}})\\) and Part Dropout promote the utilization of all parts, preventing domination by a few parts which can lead to multiple unused parts (Fig. 7). Removing the Per-Part Layer Norm leads to inconsistent part discovery (Fig. 7) indicating that it contributes to the creation of discriminative part-specific features driven by the classification loss. Equivalence Loss \\((\\mathcal{L}_{eq})\\) enhances robustness to affine geometric transformations, supporting stable part discovery across diverse inputs (Fig. 7). Gumbel-Softmax and Entropy Loss \\((\\mathcal{L}_{\\mathrm{ent}})\\) ensure unique part assignments for each patch token, further reducing information leakage, as shown in our supplementary material. As shown in Tab. 2, the complete model, incorporating all loss functions and architectural choices, generally performs best for both part discovery and classification, demonstrating the efficacy of our approach." + } + ], + [ + { + "type": "page_number", + "bbox": [ + 0.218, + 0.116, + 0.236, + 0.127 + ], + "angle": 0, + "content": "14" + }, + { + "type": "header", + "bbox": [ + 0.272, + 0.115, + 0.378, + 0.129 + ], + "angle": 0, + "content": "A. Aniraj et al." + }, + { + "type": "title", + "bbox": [ + 0.216, + 0.145, + 0.538, + 0.163 + ], + "angle": 0, + "content": "6 Limitations and Future Work" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.174, + 0.789, + 0.492 + ], + "angle": 0, + "content": "Our study focused on datasets with part annotations or FG-BG masks available for the test set. This allowed for quantitative comparison with state-of-the-art methods. However, exploring the training of our model on larger and more diverse datasets could provide valuable insights and further validate its performance in real-world scenarios. For example, larger datasets may introduce more variability in object appearances and poses, challenging the model to generalize better. Methods to enable the evaluation of part discovery in such settings would allow to take steps in this direction. The main hyper-parameter \\( K \\), the number of parts to discover, has a large impact on the granularity of the discovered parts (Fig. 6). Exploring ways of automatically estimating this value while preserving interpretability in the discovered semantic parts would be a valuable contribution. This could enhance the adaptability of the model to various datasets and streamline the training process by removing the need for manual selection of \\( K \\). Additionally, while our approach can detect parts of various shapes due to the total variation prior, it cannot ascertain if they belong to the same object. Integrating our method with unsupervised object discovery research [12,33] could enable part discovery within each detected object, improving capabilities in complex scenes. These methods aim to identify and localize individual objects without explicit supervision, complementing our approach for a more comprehensive scene understanding. These considerations suggest promising directions for extending and refining our method." + }, + { + "type": "title", + "bbox": [ + 0.216, + 0.51, + 0.36, + 0.526 + ], + "angle": 0, + "content": "7 Conclusion" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.538, + 0.788, + 0.794 + ], + "angle": 0, + "content": "In this paper, we presented PDiscoFormer, a new approach for unsupervised part discovery. Unlike previous methods that imposed strict geometric priors on discovered object parts, our model, based on self-supervised vision transformers, demonstrated the capability to relax these constraints. By leveraging a total variation (TV) prior, our model achieved substantial improvements in both part discovery and downstream classification accuracy when compared to the state-of-the-art across multiple datasets, specifically CUB, PartImageNet, and Oxford Flowers. Our findings highlight the importance of rethinking geometric priors in unsupervised part discovery tasks, especially in scenarios where objects exhibit complex shapes or occur multiple times within the same image. By refraining from assuming specific shape or size priors, our model demonstrated enhanced generalization capabilities, making it applicable to a broader range of fine-grained image classification tasks. Overall, our study shows that the strong inductive biases learned by vision transformers during the self-supervised pre-training stage enable the use of flexible shape priors such as total variation, leading to robust performance on part discovery and the downstream classification task on all of the tested datasets." + }, + { + "type": "text", + "bbox": [ + 0.216, + 0.795, + 0.787, + 0.84 + ], + "angle": 0, + "content": "Acknowledgment. This work was supported in part by ANR project OBTEA ANR-22-CPJ1-0054-01 and, was granted access to the HPC resources of IDRIS under the allocation 2023-AD011014325 made by GENCI." + } + ], + [ + { + "type": "header", + "bbox": [ + 0.402, + 0.115, + 0.732, + 0.129 + ], + "angle": 0, + "content": "PDiscoFormer: Part Discovery with Transformers" + }, + { + "type": "page_number", + "bbox": [ + 0.769, + 0.117, + 0.785, + 0.127 + ], + "angle": 0, + "content": "15" + }, + { + "type": "title", + "bbox": [ + 0.217, + 0.145, + 0.323, + 0.16 + ], + "angle": 0, + "content": "References" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.179, + 0.785, + 0.208 + ], + "angle": 0, + "content": "1. Amir, S., Gandelsman, Y., Bagon, S., Dekel, T.: Deep vit features as dense visual descriptors. arXiv preprint arXiv:2112.05814 2(3), 4 (2021)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.209, + 0.785, + 0.236 + ], + "angle": 0, + "content": "2. Ba, J.L., Kiros, J.R., Hinton, G.E.: Layer normalization. arXiv preprint arXiv:1607.06450 (2016)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.237, + 0.785, + 0.264 + ], + "angle": 0, + "content": "3. Chan, T.F., Vese, L.A.: Active contours without edges. IEEE Transactions on image processing 10(2), 266-277 (2001)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.265, + 0.785, + 0.307 + ], + "angle": 0, + "content": "4. Chen, C., Li, O., Tao, D., Barnett, A., Rudin, C., Su, J.K.: This looks like that: deep learning for interpretable image recognition. Advances in neural information processing systems 32 (2019)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.308, + 0.785, + 0.349 + ], + "angle": 0, + "content": "5. Choudhury, S., Laina, I., Rupprecht, C., Vedaldi, A.: Unsupervised part discovery from contrastive reconstruction. Advances in Neural Information Processing Systems 34, 28104-28118 (2021)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.35, + 0.785, + 0.377 + ], + "angle": 0, + "content": "6. Darcet, T., Oquab, M., Mairal, J., Bojanowski, P.: Vision transformers need registers. In: The Twelfth International Conference on Learning Representations (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.378, + 0.785, + 0.405 + ], + "angle": 0, + "content": "7. Dobson, D.C., Santosa, F.: Recovery of blocky images from noisy and blurred data. SIAM Journal on Applied Mathematics 56(4), 1181-1198 (1996)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.406, + 0.785, + 0.433 + ], + "angle": 0, + "content": "8. Donoser, M., Urschler, M., Hirzer, M., Bischof, H.: Saliency driven total variation segmentation. In: ICCV. pp. 817-824. IEEE (2009)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.434, + 0.785, + 0.489 + ], + "angle": 0, + "content": "9. Dosovitskiy, A., Beyer, L., Kolesnikov, A., Weissenborn, D., Zhai, X., Unterthiner, T., Dehghani, M., Minderer, M., Heigold, G., Gelly, S., et al.: An image is worth 16x16 words: Transformers for image recognition at scale. In: International Conference on Learning Representations (2020)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.49, + 0.785, + 0.532 + ], + "angle": 0, + "content": "0. Gautam, S., Höhne, M.M.C., Hansen, S., Jenssen, R., Kampffmeyer, M.: This looks more like that: Enhancing self-explaining models by prototypical relevance propagation. Pattern Recognition 136, 109172 (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.533, + 0.785, + 0.574 + ], + "angle": 0, + "content": "1. He, J., Yang, S., Yang, S., Kortylewski, A., Yuan, X., Chen, J.N., Liu, S., Yang, C., Yu, Q., Yuille, A.: Partimagenet: A large, high-quality dataset of parts. In: European Conference on Computer Vision. pp. 128-145. Springer (2022)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.575, + 0.785, + 0.616 + ], + "angle": 0, + "content": "2. Henaff, O.J., Koppula, S., Shelhamer, E., Zoran, D., Jaegle, A., Zisserman, A., Carreira, J., Arandjelovic, R.: Object discovery and representation networks. In: European Conference on Computer Vision. pp. 123-143. Springer (2022)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.617, + 0.785, + 0.657 + ], + "angle": 0, + "content": "3. Huang, Z., Li, Y.: Interpretable and accurate fine-grained recognition via region grouping. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 8662-8672 (2020)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.659, + 0.785, + 0.7 + ], + "angle": 0, + "content": "4. Hung, W.C., Jampani, V., Liu, S., Molchanov, P., Yang, M.H., Kautz, J.: SCOPS: Self-Supervised Co-Part Segmentation. In: IEEE Conf. on Computer Vision and Pattern Recognition (CVPR) (2019)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.701, + 0.785, + 0.728 + ], + "angle": 0, + "content": "5. Jang, E., Gu, S., Poole, B.: Categorical reparameterization with gumbel-softmax. In: International Conference on Learning Representations (2017)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.729, + 0.785, + 0.784 + ], + "angle": 0, + "content": "6. Johnson, J., Alahi, A., Fei-Fei, L.: Perceptual losses for real-time style transfer and super-resolution. In: Computer Vision-ECCV 2016: 14th European Conference, Amsterdam, The Netherlands, October 11-14, 2016, Proceedings, Part II 14. pp. 694-711. Springer (2016)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.785, + 0.785, + 0.84 + ], + "angle": 0, + "content": "7. van der Klis, R., Alaniz, S., Mancini, M., Dantas, C.F., Ienco, D., Akata, Z., Marcos, D.: PDiscoNet: Semantically consistent part discovery for fine-grained recognition. In: Proceedings of the IEEE/CVF International Conference on Computer Vision. pp. 1866-1876 (2023)" + }, + { + "type": "list", + "bbox": [ + 0.226, + 0.179, + 0.785, + 0.84 + ], + "angle": 0, + "content": null + } + ], + [ + { + "type": "page_number", + "bbox": [ + 0.218, + 0.116, + 0.236, + 0.127 + ], + "angle": 0, + "content": "16" + }, + { + "type": "header", + "bbox": [ + 0.272, + 0.115, + 0.377, + 0.129 + ], + "angle": 0, + "content": "A. Aniraj et al." + }, + { + "type": "ref_text", + "bbox": [ + 0.218, + 0.147, + 0.787, + 0.205 + ], + "angle": 0, + "content": "18. Marcos, D., Kierdorf, J., Cheeseman, T., Tuia, D., Roscher, R.: A whale's tail-finding the right whale in an uncertain world. In: International Workshop on Extending Explainable AI Beyond Deep Models and Classifiers. pp. 297-313. Springer (2020)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.206, + 0.787, + 0.247 + ], + "angle": 0, + "content": "19. Mumford, D.B., Shah, J.: Optimal approximations by piecewise smooth functions and associated variational problems. Communications on pure and applied mathematics (1989)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.248, + 0.787, + 0.29 + ], + "angle": 0, + "content": "20. Nauta, M., van Bree, R., Seifert, C.: Neural Prototype Trees for Interpretable Fine-Grained Image Recognition. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR). pp. 14933-14943 (6 2021)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.291, + 0.787, + 0.346 + ], + "angle": 0, + "content": "21. Nauta, M., Jutte, A., Provoost, J., Seifert, C.: This looks like that, because... explaining prototypes for interpretable image recognition. In: Joint European Conference on Machine Learning and Knowledge Discovery in Databases. pp. 441-456. Springer (2021)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.346, + 0.787, + 0.389 + ], + "angle": 0, + "content": "22. Nauta, M., Schlötterer, J., van Keulen, M., Seifert, C.: PIP-Net: Patch-Based Intuitive Prototypes for Interpretable Image Classification. Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.39, + 0.787, + 0.432 + ], + "angle": 0, + "content": "23. Nilsback, M.E., Zisserman, A.: Automated flower classification over a large number of classes. In: Indian Conference on Computer Vision, Graphics and Image Processing (Dec 2008)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.432, + 0.787, + 0.515 + ], + "angle": 0, + "content": "24. Oquab, M., Darcet, T., Moutakanni, T., Vo, H., Szafraniec, M., Khalidov, V., Fernandez, P., Haziza, D., Massa, F., El-Nouby, A., Assran, M., Ballas, N., Galuba, W., Howes, R., Huang, P.Y., Li, S.W., Misra, I., Rabbat, M., Sharma, V., Synnaeve, G., Xu, H., Jegou, H., Mairal, J., Labatut, P., Joulin, A., Bojanowski, P.: DINOv2: Learning Robust Visual Features without Supervision. arXiv preprint arXiv:2304.07193 (2023), http://arxiv.org/abs/2304.07193" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.516, + 0.787, + 0.558 + ], + "angle": 0, + "content": "25. Ramaswamy, H.G., et al.: Ablation-cam: Visual explanations for deep convolutional network via gradient-free localization. In: proceedings of the IEEE/CVF winter conference on applications of computer vision. pp. 983-991 (2020)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.559, + 0.787, + 0.6 + ], + "angle": 0, + "content": "26. Ranasinghe, K., McKinzie, B., Ravi, S., Yang, Y., Toshev, A., Shlens, J.: Perceptual grouping in contrastive vision-language models. In: Proceedings of the IEEE/CVF International Conference on Computer Vision. pp. 5571-5584 (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.601, + 0.787, + 0.657 + ], + "angle": 0, + "content": "27. Ribeiro, M.T., Singh, S., Guestrin, C.: \"Why Should I Trust You?\": Explaining the Predictions of Any Classifier. In: Proceedings of the 22nd ACM SIGKDD International Conference on Knowledge Discovery and Data Mining, San Francisco, CA, USA, August 13-17, 2016. pp. 1135-1144 (2016)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.658, + 0.787, + 0.699 + ], + "angle": 0, + "content": "28. Rudin, C.: Stop explaining black box machine learning models for high stakes decisions and use interpretable models instead. Nature machine intelligence 1(5), 206-215 (2019)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.7, + 0.787, + 0.728 + ], + "angle": 0, + "content": "29. Rudin, L.I., Osher, S., Fatemi, E.: Nonlinear total variation based noise removal algorithms. Physica D: nonlinear phenomena 60(1-4), 259-268 (1992)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.729, + 0.787, + 0.784 + ], + "angle": 0, + "content": "30. Rymarczyk, D., Struski, Å., Tabor, J., Zielinski, B.: Protopshare: Prototypical parts sharing for similarity discovery in interpretable image classification. In: Proceedings of the 27th ACM SIGKDD Conference on Knowledge Discovery & Data Mining. pp. 1420-1430 (2021)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.785, + 0.787, + 0.841 + ], + "angle": 0, + "content": "31. Selvaraju, R.R., Cogswell, M., Das, A., Vedantam, R., Parikh, D., Batra, D.: Gradcam: Visual explanations from deep networks via gradient-based localization. In: Proceedings of the IEEE international conference on computer vision. pp. 618-626 (2017)" + }, + { + "type": "list", + "bbox": [ + 0.217, + 0.147, + 0.787, + 0.841 + ], + "angle": 0, + "content": null + } + ], + [ + { + "type": "header", + "bbox": [ + 0.402, + 0.115, + 0.732, + 0.129 + ], + "angle": 0, + "content": "PDiscoFormer: Part Discovery with Transformers" + }, + { + "type": "page_number", + "bbox": [ + 0.768, + 0.117, + 0.786, + 0.127 + ], + "angle": 0, + "content": "17" + }, + { + "type": "ref_text", + "bbox": [ + 0.214, + 0.147, + 0.788, + 0.188 + ], + "angle": 0, + "content": "32. Shannon, C.E.: A Mathematical Theory of Communication. Bell System Technical Journal 27(3), 379-423 (1948). https://doi.org/10.1002/j.1538-7305.1948.tb01338.x" + }, + { + "type": "ref_text", + "bbox": [ + 0.214, + 0.189, + 0.788, + 0.231 + ], + "angle": 0, + "content": "33. Simeoni, O., Puy, G., Vo, H.V., Roburin, S., Gidaris, S., Bursuc, A., Pérez, P., Marlet, R., Ponce, J.: Localizing objects with self-supervised transformers and no labels. arXiv preprint arXiv:2109.14279 (2021)" + }, + { + "type": "ref_text", + "bbox": [ + 0.214, + 0.231, + 0.788, + 0.272 + ], + "angle": 0, + "content": "34. Srivastava, N., Hinton, G., Krizhevsky, A., Sutskever, I., Salakhutdinov, R.: Dropout: a simple way to prevent neural networks from overfitting. The journal of machine learning research 15(1), 1929-1958 (2014)" + }, + { + "type": "ref_text", + "bbox": [ + 0.214, + 0.272, + 0.788, + 0.3 + ], + "angle": 0, + "content": "35. Strong, D., Chan, T.: Edge-preserving and scale-dependent properties of total variation regularization. Inverse problems 19(6), S165 (2003)" + }, + { + "type": "ref_text", + "bbox": [ + 0.214, + 0.3, + 0.788, + 0.328 + ], + "angle": 0, + "content": "36. Unger, M., Pock, T., Trobin, W., Cremers, D., Bischof, H.: Tvseg-interactive total variation based image segmentation. In: BMVC. vol. 31, pp. 44-46. Citeseer (2008)" + }, + { + "type": "ref_text", + "bbox": [ + 0.214, + 0.328, + 0.788, + 0.37 + ], + "angle": 0, + "content": "37. Wang, J., Liu, H., Wang, X., Jing, L.: Interpretable image recognition by constructing transparent embedding space. In: Proceedings of the IEEE/CVF International Conference on Computer Vision. pp. 895-904 (2021)" + }, + { + "type": "ref_text", + "bbox": [ + 0.214, + 0.37, + 0.788, + 0.412 + ], + "angle": 0, + "content": "38. Welinder, P., Branson, S., Mita, T., Wah, C., Schroff, F., Belongie, S., Perona, P.: Caltech-UCSD Birds 200. Tech. Rep. CNS-TR-2010-001, California Institute of Technology (2010)" + }, + { + "type": "ref_text", + "bbox": [ + 0.214, + 0.412, + 0.788, + 0.467 + ], + "angle": 0, + "content": "39. Xu, J., De Mello, S., Liu, S., Byeon, W., Breuel, T., Kautz, J., Wang, X.: Groupvit: Semantic segmentation emerges from text supervision. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 18134-18144 (2022)" + }, + { + "type": "ref_text", + "bbox": [ + 0.214, + 0.467, + 0.788, + 0.508 + ], + "angle": 0, + "content": "40. Xu-Darme, R., Quenot, G., Chihani, Z., Rousset, M.C.: Particul: part identification with confidence measure using unsupervised learning. In: International Conference on Pattern Recognition. pp. 173-187. Springer (2022)" + }, + { + "type": "ref_text", + "bbox": [ + 0.214, + 0.508, + 0.788, + 0.536 + ], + "angle": 0, + "content": "41. Yeh, R.A., Hu, Y.T., Ren, Z., Schwing, A.G.: Total variation optimization layers for computer vision. In: CVPR. pp. 711-721 (2022)" + }, + { + "type": "list", + "bbox": [ + 0.214, + 0.147, + 0.788, + 0.536 + ], + "angle": 0, + "content": null + } + ] +] \ No newline at end of file diff --git a/2024/PDiscoFormer_ Relaxing Part Discovery Constraints with Vision Transformers/aa33e6f8-0640-410e-98c0-983fb97aea81_origin.pdf b/2024/PDiscoFormer_ Relaxing Part Discovery Constraints with Vision Transformers/aa33e6f8-0640-410e-98c0-983fb97aea81_origin.pdf new file mode 100644 index 0000000000000000000000000000000000000000..e9d4b8b9d0401baab7ad2af5b13c3c213ab7fa4f --- /dev/null +++ b/2024/PDiscoFormer_ Relaxing Part Discovery Constraints with Vision Transformers/aa33e6f8-0640-410e-98c0-983fb97aea81_origin.pdf @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:426b05f77996ea271179fb5f01c8a8f0165e7f21dee1b3f7c2990c8634bb69bd +size 5755894 diff --git a/2024/PDiscoFormer_ Relaxing Part Discovery Constraints with Vision Transformers/full.md b/2024/PDiscoFormer_ Relaxing Part Discovery Constraints with Vision Transformers/full.md new file mode 100644 index 0000000000000000000000000000000000000000..245a344edae5f93f4e3d4ac1d23ffbda82a72783 --- /dev/null +++ b/2024/PDiscoFormer_ Relaxing Part Discovery Constraints with Vision Transformers/full.md @@ -0,0 +1,257 @@ +# PDiscoFormer: Relaxing Part Discovery Constraints with Vision Transformers + +Ananthu Aniraj $^{1}$ , Cassio F. Dantas $^{2}$ , Dino Ienc $^{2}$ , and Diego Marcos $^{1}$ + +1 Inria, Univ. Montpellier, LIRMM, UMR TETIS, Montpellier, France {ananthu.aniraj, diego.marcos}@inria.fr +2 Inria, Inrae, Univ. Montpellier, UMR TETIS, Montpellier, France {cassio.fraga-dantas, dino.ienco}@inrae.fr + +Abstract. Computer vision methods that explicitly detect object parts and reason on them are a step towards inherently interpretable models. Existing approaches that perform part discovery driven by a fine-grained classification task make very restrictive assumptions on the geometric properties of the discovered parts; they should be small and compact. Although this prior is useful in some cases, in this paper we show that pre-trained transformer-based vision models, such as self-supervised DINOv2 ViT, enable the relaxation of these constraints. In particular, we find that a total variation (TV) prior, which allows for multiple connected components of any size, substantially outperforms previous work. We test our approach on three fine-grained classification benchmarks: CUB, PartImageNet and Oxford Flowers, and compare our results to previously published methods as well as a re-implementation of the state-of-the-art method PDiscoNet with a transformer-based backbone. We consistently obtain substantial improvements across the board, both on part discovery metrics and the downstream classification task, showing that the strong inductive biases in self-supervised ViT models require to rethink the geometric priors that can be used for unsupervised part discovery. Training code and pre-trained models are available at https://github.com/ananthu-aniraj/pdiscoformer. + +# 1 Introduction + +Deep neural networks, first in the form of convolutional neural networks (CNN) and, more recently, vision transformers (ViT), have revolutionized the field of + +![](images/067bcb1bd9d7075196ae8e1b7e0ee8c8c055660bf65861a01f6d2b91f8b0eded.jpg) +Fig. 1: Concentration vs total variation as part priors. + +computer vision. However, their black-box nature often limits their application in domains where interpretability is crucial. To ameliorate the understanding of the internal behaviour of deep learning models trained for computer vision tasks, post-hoc interpretability methods [25,27,31] are commonly used. These methods generate saliency maps indicating regions in the image that contributed most to a classification decision. However, these are only approximations to the actual internal process of the model and may be misleading [28]. + +Recent efforts have been directed towards developing interpretable-by-design deep neural networks [4, 13, 14, 17, 21, 22]. These approaches aim to identify semantically interpretable features in an image and utilize this information for solving a downstream task. Primarily tailored for fine-grained image classification, these methods seek to emulate human reasoning by subdividing the final task, such as classifying flower species or dog breeds, into multiple explicit intermediate stages, such as identifying lower-level semantic features like body parts, which in turn are used to provide a classification score. + +However, improving interpretability while maintaining performance remains a challenge. Some models, such as those proposed in [13, 14, 17, 18], rely on objective functions that impose priors on the shape and area of object parts such as compactness-inducing part shaping losses that, in practice, result in spatial Gaussian priors of fixed variance on the part maps. While effective for specific datasets, like the commonly used CUB [38] benchmark for bird species classification, where assumptions like parts occurring only once in an image and being compact are beneficial, these assumptions may not generalize well to other use cases. For instance, in plant species classification, the presence of multiple instances of object parts such as leaves or flowers, as well as irregularly-shaped parts like branches or stems, renders these assumptions inadequate. + +As we show in this work, such strong part shape priors are required when using a CNN-based backbone pre-trained on ImageNet; in order to reach acceptable classification results, it is imperative to fine-tune the whole backbone. This means that the only inductive bias that is kept throughout the part discovery process is the one brought in by the convolutional nature of the CNN architecture. A strong part shape prior is required in order to obtain consistent part maps in this setting, leading to all recent methods for part discovery in the context of fine-grained classification to use a concentration loss that makes part maps small and compact [13, 14, 17]. Here, we propose a method that addresses this issue by refraining from assuming any priors on part shape or area. The only prior we assume is that the part maps should tend to be piece-wise constant and thus be formed of one or more spatially connected components. Our model is capable of localizing up to a fixed number of semantically consistent parts of an object without relying on part annotations, using only the classification labels as supervision signals. The lack of assumptions about part shape and size makes our approach suitable to various fine-grained image classification tasks, as demonstrated by our empirical results, generalizing beyond cases where compactness is a good part prior. + +The main contribution of this paper is to build upon [17] and relax its constraining concentration-based part prior by a more flexible total variation prior. We investigate which additional elements are required to compensate for the flexibility of the total variation loss and find that a combination of an entropy loss and Gumbel-Softmax on the part maps allows for consistent and high quality part maps. Our approach makes better use of the representation of self-supervised pre-trained ViT models and allows to apply the model to tasks with diverse types of parts. + +# 2 Related Work + +Unsupervised Part Discovery. Machine learning methods designed for unsupervised part discovery aim to identify a small number of semantically consistent parts, typically around $K = 8$ , shared among classes in a dataset. These methods visualize all discovered parts in an image through per-part saliency maps, facilitating easier interpretation. However, as this task is unsupervised, it often relies on assumptions about the discovered parts, such as their distribution and shape. For instance, Huang et al. [13] assume that the presence of each part across a mini-batch follows a beta distribution, while [17] assumes that discovered parts are present at least once in a mini-batch. Additionally, assumptions are often made about the shape of the parts. Methods such as [14, 17, 18, 40] impose a spatial Gaussian prior with a fixed variance on the part maps, resulting in compact parts appearing once in the image. Such assumptions on part shape can be highly restrictive and may not generalize well across different scenarios. To address this limitation, our proposed method refrains from assuming such priors on part shape and only imposes a spatial prior that discovered parts should form connected components. + +Part Prototype Networks. Chen et al. [4] introduced the Prototypical Part Network (ProtoPNet), an inherently interpretable method featuring a fixed number of learned prototypes per class. These prototypes represent visual patterns or features learned by the network during training. The model computes the similarity between each learned prototype and image patches and uses a linear combination of these similarity scores for classification. While explanations provided by ProtoPNet visualize all prototypes along with their weighted similarity scores, the sheer number of prototypes, often proportional to the number of classes, can lead to redundancy and interpretability challenges [10,21]. Subsequent works in this line of research have sought to address these challenges. Some approaches focus on reducing the number of prototypes shared per class [30,37], while others explore sharing prototypes between classes [20, 22]. Despite these efforts, recent works like [22] still utilize a large number ( $K > 500$ ) of prototypes, which can pose significant challenges for end-user interpretation. + +Regularization in Image Segmentation. Early works on image segmentation made direct use of low-level image features, chiefly color, along with geometric priors, in order to find coherent segments. The Mumford-Shah model [19], originally devised for piece-wise image reconstruction, was applied to find segments + +![](images/72f2233595f30e2531b43d5feca2bd0796ef12aabc0cbfd7ed18999f4312f4a9.jpg) +Fig. 2: The PDiscoFormer architecture for part discovery. + +that are as homogeneous as possible while trying to minimize the length of their boundary as a regularization [3]. Total variation (TV) [29] has been shown to be a powerful and efficient regularization to approximate this boundary length prior, and has been extensively used for piece-wise signal and image reconstruction [7,35], as well as unsupervised image segmentation [8,36]. With the advent of deep learning, TV regularization has proven valuable for style transfer [16] and edge-preserving smoothing [41]. In this work we explore the use of TV regularization for weakly-supervised part discovery in order to alleviate the issues brought by previously used priors. + +Emergent grouping of ViT representations. Some recent research [26,39] has studied the emergence of object segmentation by supervising with image-text pairs. These methods learn to group similar patches in a ViT and assign them to textual concepts. They aim to identify the pixels in an input image corresponding to a predefined text prompt, focusing on segmenting whole objects rather than decomposing them into parts, as our model does with part discovery, and are limited to concepts that are represented in the text annotations. + +# 3 Methodology + +Here we introduce an end-to-end trainable method (see Fig. 2) capable of automatically identifying $K$ discriminative parts (semantically interpretable image regions) useful for solving a fine-grained image classification task. In contrast to earlier methods [13,17], our approach avoids any assumptions related to part shape, size or number of part instances. As in [13,17], our model is weakly supervised and relies solely on image-level class labels, thus uncovering discriminative parts without requiring any additional supervision. Our formulation is based + +on [17], from which we take take most elements of the architecture, except for the Gumbel-Softmax and the layer normalization-based part modulation. The total variation, entropy and background losses have not been used in previous part discovery works and are shown to be important in our results. + +# 3.1 Model architecture + +Let $\mathbf{x} \in \mathbb{R}^{3 \times M \times N}$ be an image in the dataset, and let $y \in \{1, 2, \dots, C\}$ represent the corresponding classification label. Given a backbone model $h_{\theta}(\cdot)$ , we extract a feature tensor: + +$$ +\mathbf {z} = h _ {\theta} (\mathbf {x}) \in \mathbb {R} ^ {D \times H \times W} \tag {1} +$$ + +We drop the class and register tokens of the ViT and reshape the patch tokens to obtain $\mathbf{z}$ . Following prior research [13, 14, 17], we compute attention maps $\mathbf{A} \in [0,1]^{(K + 1) \times H \times W}$ , where the last channel in the first dimension is used to represent the background region in the image. These attention maps are computed using the negative squared Euclidean distance function between patch token features $\mathbf{z}_{ij} \in \mathbb{R}^D$ , $i \in \{1, \dots, H\}$ , $j \in \{1, \dots, W\}$ and each of the learnable prototypes $\mathbf{p}^k \in \mathbb{R}^D$ , with $k \in \{1, \dots, K + 1\}$ , followed by a Gumbel-Softmax [15] across the $K + 1$ channels: + +$$ +a _ {i j} ^ {k} = \frac {\exp \left(- \| \mathbf {z} _ {i j} - \mathbf {p} ^ {k} \| ^ {2} + \gamma_ {k}\right)}{\sum_ {l = 1} ^ {K + 1} \exp \left(- \| \mathbf {z} _ {i j} - \mathbf {p} ^ {l} \| ^ {2} + \gamma_ {l}\right)} \tag {2} +$$ + +where $\gamma_{k}$ and $\gamma_{l}$ are independently drawn samples from the Gumbel(0,1) distribution, and $a_{ij}^{k}$ represents an element of $\mathbf{A}$ . + +The attention maps are used to compute part embedding vectors $\mathbf{v}^k\in \mathbb{R}^D$ by using the attention values to calculate a weighted average over the feature tensor $\mathbf{z}$ : + +$$ +\mathbf {v} ^ {k} = \frac {\sum_ {i} \sum_ {j} a _ {i j} ^ {k} \mathbf {z} _ {i j}}{H W} \tag {3} +$$ + +These embedding vectors are then modulated using a layer normalization [2] operation, with normalization applied over the feature and prototype dimensions. This operation helps to make each embedding vector distinct for the classifier: + +$$ +\mathbf {v} _ {m} ^ {k} = \frac {\left(\mathbf {v} ^ {k} - \operatorname {E} \left[ \mathbf {v} ^ {k} \right]\right)}{\sqrt {\operatorname {V a r} \left[ \mathbf {v} ^ {k} \right] + \epsilon}} \odot \mathbf {w} _ {m} ^ {k} + \mathbf {b} _ {m} ^ {k} \tag {4} +$$ + +with $\mathbf{w}_m^k, \mathbf{b}_m^k \in \mathbb{R}^D$ being the modulation weights and biases per part. These modulated embedding vectors $\mathbf{v}_m^k$ for $k \in \{1, \dots, K\}$ (we omit the background embedding) are then linearly projected to obtain a vector of class scores conditioned on the part embedding: + +$$ +\mathbf {y} ^ {k} = W _ {c} \cdot \mathbf {v} _ {m} ^ {k} \tag {5} +$$ + +The same linear classifier $W_{c} \in \mathbb{R}^{C \times D}$ is applied to each of the modulated embedding vectors, i.e., for each of the $K$ discovered foreground parts. We then take the mean of the per-part class scores to obtain the classification scores: + +$$ +\mathbf {y} = \frac {1}{K} \sum_ {k = 1} ^ {K} \mathbf {y} ^ {k} \tag {6} +$$ + +Finally, a softmax function is employed to obtain the classification probabilities. Part dropout. We adopt dropout [34] on the modulated embedding vectors $\mathbf{v}_m^k$ , as proposed in [17], during model training. This drops entire part embedding vectors at a time and ensures that all discovered parts are discriminative, since discovering a single discriminative part would not be enough to solve the downstream classification task. + +# 3.2 Loss Functions + +Our main contribution consists in proposing a set of part-shaping loss functions that are able to make the most out of the inductive biases learned by self-supervised ViT models. + +Classification Loss $(\mathcal{L}_{\mathbf{c}})$ . Our model is mainly driven by the fine-grained classification objective during training. We use the standard negative log likelihood or multi-class cross-entropy for this purpose. While the classification loss helps ensure the discriminative nature of the discovered parts, it may not directly enforce the learned attention maps to represent salient parts of the object in the image. To address this, we incorporate additional objective functions to enforce priors on the discovered parts. These objectives effectively guide the learning process to focus on relevant regions of the image. + +Orthogonality Loss $(\mathcal{L}_{\perp})$ . We aim to discover parts that are semantically distinct. To achieve this, we apply the orthogonality loss, which encourages the learned part embedding vectors to be decorrelated from one another. This is accomplished by minimizing the cosine similarity between the embedding vectors. We apply this loss on the modulated embedding vectors $\mathbf{v}_m^k$ . + +$$ +\mathcal {L} _ {\perp} = \sum_ {k = 1} ^ {K + 1} \sum_ {l \neq k} \frac {\mathbf {v} _ {m} ^ {k} \cdot \mathbf {v} _ {m} ^ {l}}{\| \mathbf {v} _ {m} ^ {k} \| \cdot \| \mathbf {v} _ {m} ^ {l} \|} \tag {7} +$$ + +Equivalence Loss $(\mathcal{L}_{\mathrm{eq}})$ . We aim to detect the same parts even if the image is translated, rotated, or scaled. To achieve this, we formulate the equivariance loss, which encourages the learned part attention maps to be equivariant to rigid transformations. Specifically, given an input image, we use an affine transform $T$ with parameters randomly sampled from a pre-defined range and pass both the original image and the transformed image through our model. Then, we invert the transformation on the part attention maps of the transformed image. Finally, the equivariance loss is formulated as the cosine distance between the part attention maps of the original image and those of the corresponding transformed image. + +Let $A^k(\mathbf{x})$ be a function that returns the $k^{th}$ attention map given an input image $\mathbf{x}$ . Then: + +$$ +\mathcal {L} _ {\mathrm {e q}} = 1 - \frac {1}{K} \sum_ {k} \frac {\left\| A ^ {k} (\mathbf {x}) \cdot T ^ {- 1} \left(A ^ {k} (T (\mathbf {x})) \right. \right\|}{\| A ^ {k} (\mathbf {x}) \| \cdot \| A ^ {k} (T (\mathbf {x})) \|} \tag {8} +$$ + +Presence Loss $(\mathcal{L}_{\mathbf{p}})$ . We assume that all discovered foreground parts are present in at least some images of the training dataset, while the background, the $(K + 1)^{th}$ part, is expected to be present in all images in the dataset. To enforce this, we formulate the presence loss, which consists of two parts. Firstly, we define a loss applied only to the discovered foreground parts, ensuring that each part is present at least once in a mini-batch. Given a batch $\{\mathbf{x}_1,\dots ,\mathbf{x}_B\}$ , the presence loss is calculated as: + +$$ +\mathcal {L} _ {\mathrm {p} _ {1}} = 1 - \frac {1}{K} \sum_ {k} \max _ {b, i, j} \bar {a} _ {i j} ^ {k} \left(\mathbf {x} _ {b}\right) \tag {9} +$$ + +where $\bar{A}^k (\mathbf{x}_b) = \mathrm{avgpool}(A^k (\mathbf{x}_b))$ is the output of a 2D average pooling operator with a small kernel size employed to prevent single-pixel attention maps. + +Secondly, to ensure the presence of the background, we introduce a stricter presence loss, which is satisfied if the background is present in all images of the mini-batch. This loss is formulated as follows: + +$$ +\mathcal {L} _ {\mathrm {p} _ {0}} = - \frac {1}{B} \sum_ {b} \log \left(\max _ {i, j} m _ {i j} \bar {a} _ {i j} ^ {K + 1} (\mathbf {x} _ {b})\right) \tag {10} +$$ + +where $M = [m_{ij}]^{H \times W}$ is a soft mask with $m_{ij} \in [0,1]$ that privileges entries placed farther from the image center (towards the borders): + +$$ +m _ {i j} = 2 \left(\frac {i - 1}{H - 1} - \frac {1}{2}\right) ^ {2} + 2 \left(\frac {j - 1}{W - 1} - \frac {1}{2}\right) ^ {2} \tag {11} +$$ + +This loss captures the notion that the background is at the same time expected in every image and is more likely to appear near the boundaries of the image. + +Entropy Loss $(\mathcal{L}_{\mathrm{ent}})$ . We aim to ensure that each patch token is assigned to a unique part. To achieve this, we formulate an entropy [32] loss applied to the part attention maps: + +$$ +\mathcal {L} _ {\text {e n t}} = \frac {- 1}{K + 1} \sum_ {k = 1} ^ {K + 1} \sum_ {i j} a _ {i j} ^ {k} \log \left(a _ {i j} ^ {k}\right) \tag {12} +$$ + +Total Variation Loss $(\mathcal{L}_{\mathrm{tv}})$ . We do not want to make any assumptions about the part shape. However, we would still like the discovered parts to be composed of one or, at most, a few connected components. To achieve this, we make use of a total variation loss [29] applied to the part attention maps: + +$$ +\mathcal {L} _ {\mathrm {t v}} = \frac {1}{H W} \sum_ {k = 1} ^ {K + 1} \sum_ {i j} \left| \nabla a _ {i j} ^ {k} \right| \tag {13} +$$ + +where $\nabla a_{ij}^{k}$ is the spatial image gradient of part map $A^k$ in location $ij$ . + +# 4 Experimental Setup + +We aim to perform part discovery using only image-level class labels as supervision signals, assuming that the discovered parts are shared among classes, which is common in fine-grained image classification settings. Moreover, we aim to compare our model with similar methods in the literature in scenarios where they have been empirically shown to work well, as well as in new and more challenging scenarios where underlying data characteristics may require more flexible geometric modeling. + +All models were implemented in PyTorch. We used the DinoV2 [24] ViT [9] model with register tokens [6] as the starting weights for all experiments, utilizing the Base (ViT-B) variant. The training settings and compute requirements of our model are detailed in the supplementary material. + +# 4.1 Datasets + +CUB Dataset. The CUB dataset [38] consists of images depicting 200 bird species, with 5,994 images in the training set and 5,794 images in the testing set. Keypoint annotations for 15 different bird body parts are available for each image. Bird parts are typically compact, although their presence may vary in images due to changes in pose and occlusion. Most images contain only a single instance of a bird. For this dataset, we evaluate the results using keypoint regression (Kp) using ground truth part keypoint annotations to compare them with centroids discovered by the model. Furthermore, we compute the Normalized Mutual Information (NMI) and Adjusted Rand Index (ARI) to quantify the clustering quality of the discovered parts, as in related literature [5]. + +PartImageNet Dataset. The PartImageNet dataset [11] comprises 158 classes distributed among 11 super-classes, such as airplanes, birds, and snakes. Each super-class is associated to between two and five possible parts, for a total of 41 part classes. Additionally, test images may contain multiple instances of objects, sometimes from different super-classes, while the classification label provides information about a single object. Notably, several super-classes, including snakes, airplanes, and reptiles, often exhibit irregular and non-compact part shapes. For our experiments, we use the PartImageNet 00D variant. Although originally intended for out-of-distribution detection, this variant is commonly used in related literature. We train our models on 14,865 training images and test on 1,658 images, respectively. For evaluation, we derive part centroids from ground truth part masks and calculate NMI and ARI scores to assess clustering quality. Additionally, we include an experiment on the newly released PartImageNet Seg variant in the supplementary material. + +Table 1: Quantitative comparison of the part discovery and classification capabilities of our model with related approaches from the literature. In this table, $K$ refers to the number of parts. The method with the label ** does not use any class supervision. + +
MethodCUB(%)PartImageNet OOD(%)Flowers(%)
KKp ↓NMI ↑ARI ↑Top-1 Acc. ↑KNMI ↑ARI ↑Top-1 Acc. ↑KFg. mIoU ↑Top-1 Acc. ↑
Dino [1]**4-31.1811.21-819.177.59-254.95-
8-47.2119.76-2531.4614.16-455.11-
16-50.5726.14-5037.8116.50-854.44-
Huang [13]411.5129.7414.0487.3085.881.5374.22229.9293.07
811.6035.7215.9086.05257.561.2573.56433.2293.14
1612.6043.9221.1085.935010.191.0573.20817.2692.86
PDiscoNet [17]49.1237.8215.2686.17827.138.7688.58219.0477.51
88.5250.0826.9686.722532.4110.6989.00434.7683.05
167.6756.8738.0587.495041.4914.1786.06849.1081.04
PdiscoNet + ViT-B [6]47.7052.5926.6688.61819.2834.7290.9524.9292.75
86.3465.0137.9086.952528.2350.3590.2941.9595.48
165.9568.6343.4184.045029.4827.8089.69813.1897.40
PDiscoFormer (Ours)47.4158.1325.1189.06829.0052.4089.75273.6299.61
85.9969.8743.4988.792544.7159.2790.77473.3299.54
165.7473.3855.8388.725046.2962.2191.01869.5999.64
+ +Oxford 102 Flower Dataset. The Oxford 102 Flower Dataset [23] comprises 8,189 images divided into 102 flower classes. The training and validation sets contain 10 images per class, totaling 1,020 images each, while the test set consists of 6,149 images, with a minimum of 20 images per class. Since no part annotations are provided for this dataset, we evaluate the quality of the discovered parts by combining them and comparing the result to the provided foreground/background masks. We evaluate model performance by calculating the mean Intersection over Union (mIoU) for the foreground mask. This dataset often includes multiple instances of parts in a single image, such as flowers, as well as parts with irregular shapes, such as flower petals. + +# 5 Experiments + +# 5.1 Comparison with State-of-the-Art + +We conduct a quantitative analysis of our model's part discovery and classification capabilities and compare it with those of related methods from the literature. Additionally, we compare our model against a re-implementation of the state-of-the-art PDiCoNet [17], where we replaced the backbone with the variant used in this paper. The results of this comparison are summarized in Tab. 1. From the results, our method consistently outperforms existing approaches for both part discovery and image classification across all datasets. We attribute this to the synergistic effects of the strong inductive biases ingrained in the self-supervised ViT backbone and the enhanced flexibility of the proposed combina + +![](images/6d9d70a64ff3ec74c6be094189be276cf5689922e8bdbda6841cba801ae1ab09.jpg) + +![](images/90f23c12cb2713a453831b957e2251654f6cb2a561f37c1e2c5c5284fb43e2d8.jpg) +Fig. 3: Qualitative results on CUB for $K = 8$ . +Fig. 4: Qualitative results on PartImageNet for $K = 8$ . + +tion of geometric priors. Particularly noteworthy are the significant performance gains observed on the Flowers and PartImageNet datasets, which contain objects with irregular shapes occurring in multiple instances within the same image, underscoring the importance of our geometric part prior that favors piece-wise constant and spatially connected components. Even on CUB, where most parts fit well into the compactness prior, we see a consistent improvements of results with respect to PDiscoNet, even when using the same ViT backbone. With 16 parts, our model achieves $55.8\%$ ARI and $73.4\%$ NMI, compared to $43.4\%$ ARI and $68.6\%$ NMI of the best competing method, PDiscoNet with a ViT backbone. Despite competitive performance in classification tasks on CUB and PartImageNet, the PdiscoNet+ViT-B model tends to degrade in accuracy as part complexity increases, particularly in CUB, unlike our model, which maintains robust accuracy for any number of parts. In the Flowers dataset, all competing methods result in subpar performance, while our approach excels at both detecting the foreground and classification no matter the number of parts. In this + +![](images/7c2457fcb03457d8b8b5b415be86bd6763ac2661fd8dd858d2d2a975b9274320.jpg) + +![](images/9edc6bf3b7342da549ca128a391e19717491e20d60be69c4d33456616af0617b.jpg) +Fig. 5: Qualitative results on Flowers for $K = 2$ . +Fig. 6: Effect of $K$ on PDiscoFormer. The left three images correspond to Flowers (from top to bottom, $K = 2,4,8$ ). The three middle ones are CUB, with $K$ being 4, 8 and 16. The three on the right belong to PartImageNet, with $K = 8,25,50$ . + +dataset, PDiscoNet+ViT-B results in the second best classification accuracy, $97.4\%$ , after the $99.6\%$ obtained by our method, but completely fails to locate the foreground, with an mIoU of $13.2\%$ , against $73.6\%$ for PDiscoFormer. Additionally, we extend our comparisons by applying our proposed losses to a ResNet backbone and a fully frozen ViT architecture, showing that a pre-trained ViT is required to fully profit from the proposed set of losses and that fine-tuning the position, class and register tokens, while keeping the rest of the ViT frozen, results in better performances than leaving the whole ViT frozen. The results of these additional comparisons are presented in the supplementary material. + +Qualitative Analysis. A qualitative analysis, as illustrated in Fig. 3, Fig. 4, Fig. 5 and Fig. 6 corroborates our quantitative findings, showcasing our model's ability to accurately discover irregularly shaped body parts such as bird wings (CUB), flower petals (Flowers) and the body of a snake (PartImageNet). At the + +Table 2: Results of the ablation study in PartImageNet and CUB. + +
PartImageNet OOD (K=25)CUB (K=16)
NMI ↑ARI ↑Top-1 Acc. ↑Kp ↓NMI ↑ARI ↑Top-1 Acc. ↑
Full Model44.7159.2790.775.7473.3855.8388.72
No Lp041.0549.0889.815.8568.6643.0983.64
No Lent39.1654.4691.195.5766.8845.6888.21
No Leq43.3854.1990.479.0556.8530.6383.71
No L⊥42.4458.8691.015.7970.5550.9088.63
No Lp133.9858.3191.077.1461.0434.2889.04
No Ltv33.7325.3590.476.0170.7851.8380.22
No Per-Part LayerNorm31.1440.7389.695.7470.7050.6269.73
No Part Dropout33.0843.9889.996.0870.2841.6987.14
No Gumbel32.9854.6890.716.4468.9243.6186.43
+ +same time, PDiscoNet [17] fails to discover these parts altogether, even when retrained on a ViT-B backbone. This indicates the usefulness of the flexible geometric priors enabled by our total variation loss. Compared to PDiscoNet [17], not only do we achieve a much tighter coverage of the object of interest, with our PDiscoFormer more closely following the object boundaries. This is particularly the case in parts that are relatively large, such as open bird wings and flower corollas. In addition to that, we can observe a substantial improvement in the consistency and semantic interpretability of parts. For instance, the parts discovered by our model for the four quadruped images displayed in Fig. 4 can immediately be interpreted (the green part corresponds to the snout while the brown one to the ears), while those from PDiscoNet are not clear. In Fig. 6 we vary the number of parts, $K$ . On Flowers, using two parts leads to a clear assignment of one part to the flower calyx and one to the corolla, while more parts tends to exhibit oversegmentation. CUB, on the other hand, is able to absorb a larger number of parts. With $K = 4$ , beak and legs are covered by the same part, while a fine-grained and consistent distinction of the different wing parts is achieved with $K = 16$ . Interestingly, in the case of PartImageNet, the effective number of parts being used by a specific super-class remains almost constant when increasing $K$ . + +# 5.2 Ablation Study + +We conducted an ablation study by systematically removing one loss function or architectural element at a time and by replacing the Gumbel-Softmax with a regular Softmax. This was performed on the PartImageNet OOD dataset with $K = 25$ parts and the CUB dataset with $K = 16$ parts. The quantitative results of the ablation study are shown in Tab. 2 and some qualitative results can be seen in Fig. 7. From the results, we can observe the following: Total Variation $(\mathcal{L}_{tv})$ prevents spurious part activations in the background, possibly reducing + +![](images/eaeb99af03887e3e9b1b7e4d6e8ffa81167451c6b916e49be4cf827a7f3db7ac.jpg) +Fig. 7: Qualitative results from the ablation study. + +information leakage between foreground parts and the background. It has the largest impact on PartImageNet in terms of ARI. The absence of the Background Loss $(\mathcal{L}_{\mathbf{p}_0})$ results in the background part not being activated, causing the model to discover the background as a foreground part. Presence Loss $(\mathcal{L}_{\mathbf{p}})$ and Part Dropout promote the utilization of all parts, preventing domination by a few parts which can lead to multiple unused parts (Fig. 7). Removing the Per-Part Layer Norm leads to inconsistent part discovery (Fig. 7) indicating that it contributes to the creation of discriminative part-specific features driven by the classification loss. Equivalence Loss $(\mathcal{L}_{eq})$ enhances robustness to affine geometric transformations, supporting stable part discovery across diverse inputs (Fig. 7). Gumbel-Softmax and Entropy Loss $(\mathcal{L}_{\mathrm{ent}})$ ensure unique part assignments for each patch token, further reducing information leakage, as shown in our supplementary material. As shown in Tab. 2, the complete model, incorporating all loss functions and architectural choices, generally performs best for both part discovery and classification, demonstrating the efficacy of our approach. + +# 6 Limitations and Future Work + +Our study focused on datasets with part annotations or FG-BG masks available for the test set. This allowed for quantitative comparison with state-of-the-art methods. However, exploring the training of our model on larger and more diverse datasets could provide valuable insights and further validate its performance in real-world scenarios. For example, larger datasets may introduce more variability in object appearances and poses, challenging the model to generalize better. Methods to enable the evaluation of part discovery in such settings would allow to take steps in this direction. The main hyper-parameter $K$ , the number of parts to discover, has a large impact on the granularity of the discovered parts (Fig. 6). Exploring ways of automatically estimating this value while preserving interpretability in the discovered semantic parts would be a valuable contribution. This could enhance the adaptability of the model to various datasets and streamline the training process by removing the need for manual selection of $K$ . Additionally, while our approach can detect parts of various shapes due to the total variation prior, it cannot ascertain if they belong to the same object. Integrating our method with unsupervised object discovery research [12,33] could enable part discovery within each detected object, improving capabilities in complex scenes. These methods aim to identify and localize individual objects without explicit supervision, complementing our approach for a more comprehensive scene understanding. These considerations suggest promising directions for extending and refining our method. + +# 7 Conclusion + +In this paper, we presented PDiscoFormer, a new approach for unsupervised part discovery. Unlike previous methods that imposed strict geometric priors on discovered object parts, our model, based on self-supervised vision transformers, demonstrated the capability to relax these constraints. By leveraging a total variation (TV) prior, our model achieved substantial improvements in both part discovery and downstream classification accuracy when compared to the state-of-the-art across multiple datasets, specifically CUB, PartImageNet, and Oxford Flowers. Our findings highlight the importance of rethinking geometric priors in unsupervised part discovery tasks, especially in scenarios where objects exhibit complex shapes or occur multiple times within the same image. By refraining from assuming specific shape or size priors, our model demonstrated enhanced generalization capabilities, making it applicable to a broader range of fine-grained image classification tasks. Overall, our study shows that the strong inductive biases learned by vision transformers during the self-supervised pre-training stage enable the use of flexible shape priors such as total variation, leading to robust performance on part discovery and the downstream classification task on all of the tested datasets. + +Acknowledgment. This work was supported in part by ANR project OBTEA ANR-22-CPJ1-0054-01 and, was granted access to the HPC resources of IDRIS under the allocation 2023-AD011014325 made by GENCI. + +# References + +1. Amir, S., Gandelsman, Y., Bagon, S., Dekel, T.: Deep vit features as dense visual descriptors. arXiv preprint arXiv:2112.05814 2(3), 4 (2021) +2. Ba, J.L., Kiros, J.R., Hinton, G.E.: Layer normalization. arXiv preprint arXiv:1607.06450 (2016) +3. Chan, T.F., Vese, L.A.: Active contours without edges. IEEE Transactions on image processing 10(2), 266-277 (2001) +4. Chen, C., Li, O., Tao, D., Barnett, A., Rudin, C., Su, J.K.: This looks like that: deep learning for interpretable image recognition. Advances in neural information processing systems 32 (2019) +5. Choudhury, S., Laina, I., Rupprecht, C., Vedaldi, A.: Unsupervised part discovery from contrastive reconstruction. Advances in Neural Information Processing Systems 34, 28104-28118 (2021) +6. Darcet, T., Oquab, M., Mairal, J., Bojanowski, P.: Vision transformers need registers. In: The Twelfth International Conference on Learning Representations (2023) +7. Dobson, D.C., Santosa, F.: Recovery of blocky images from noisy and blurred data. SIAM Journal on Applied Mathematics 56(4), 1181-1198 (1996) +8. Donoser, M., Urschler, M., Hirzer, M., Bischof, H.: Saliency driven total variation segmentation. In: ICCV. pp. 817-824. IEEE (2009) +9. Dosovitskiy, A., Beyer, L., Kolesnikov, A., Weissenborn, D., Zhai, X., Unterthiner, T., Dehghani, M., Minderer, M., Heigold, G., Gelly, S., et al.: An image is worth 16x16 words: Transformers for image recognition at scale. In: International Conference on Learning Representations (2020) +0. Gautam, S., Höhne, M.M.C., Hansen, S., Jenssen, R., Kampffmeyer, M.: This looks more like that: Enhancing self-explaining models by prototypical relevance propagation. Pattern Recognition 136, 109172 (2023) +1. He, J., Yang, S., Yang, S., Kortylewski, A., Yuan, X., Chen, J.N., Liu, S., Yang, C., Yu, Q., Yuille, A.: Partimagenet: A large, high-quality dataset of parts. In: European Conference on Computer Vision. pp. 128-145. Springer (2022) +2. Henaff, O.J., Koppula, S., Shelhamer, E., Zoran, D., Jaegle, A., Zisserman, A., Carreira, J., Arandjelovic, R.: Object discovery and representation networks. In: European Conference on Computer Vision. pp. 123-143. Springer (2022) +3. Huang, Z., Li, Y.: Interpretable and accurate fine-grained recognition via region grouping. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 8662-8672 (2020) +4. Hung, W.C., Jampani, V., Liu, S., Molchanov, P., Yang, M.H., Kautz, J.: SCOPS: Self-Supervised Co-Part Segmentation. In: IEEE Conf. on Computer Vision and Pattern Recognition (CVPR) (2019) +5. Jang, E., Gu, S., Poole, B.: Categorical reparameterization with gumbel-softmax. In: International Conference on Learning Representations (2017) +6. Johnson, J., Alahi, A., Fei-Fei, L.: Perceptual losses for real-time style transfer and super-resolution. In: Computer Vision-ECCV 2016: 14th European Conference, Amsterdam, The Netherlands, October 11-14, 2016, Proceedings, Part II 14. pp. 694-711. Springer (2016) +7. van der Klis, R., Alaniz, S., Mancini, M., Dantas, C.F., Ienco, D., Akata, Z., Marcos, D.: PDiscoNet: Semantically consistent part discovery for fine-grained recognition. In: Proceedings of the IEEE/CVF International Conference on Computer Vision. pp. 1866-1876 (2023) + +18. Marcos, D., Kierdorf, J., Cheeseman, T., Tuia, D., Roscher, R.: A whale's tail-finding the right whale in an uncertain world. In: International Workshop on Extending Explainable AI Beyond Deep Models and Classifiers. pp. 297-313. Springer (2020) +19. Mumford, D.B., Shah, J.: Optimal approximations by piecewise smooth functions and associated variational problems. Communications on pure and applied mathematics (1989) +20. Nauta, M., van Bree, R., Seifert, C.: Neural Prototype Trees for Interpretable Fine-Grained Image Recognition. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR). pp. 14933-14943 (6 2021) +21. Nauta, M., Jutte, A., Provoost, J., Seifert, C.: This looks like that, because... explaining prototypes for interpretable image recognition. In: Joint European Conference on Machine Learning and Knowledge Discovery in Databases. pp. 441-456. Springer (2021) +22. Nauta, M., Schlötterer, J., van Keulen, M., Seifert, C.: PIP-Net: Patch-Based Intuitive Prototypes for Interpretable Image Classification. Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (2023) +23. Nilsback, M.E., Zisserman, A.: Automated flower classification over a large number of classes. In: Indian Conference on Computer Vision, Graphics and Image Processing (Dec 2008) +24. Oquab, M., Darcet, T., Moutakanni, T., Vo, H., Szafraniec, M., Khalidov, V., Fernandez, P., Haziza, D., Massa, F., El-Nouby, A., Assran, M., Ballas, N., Galuba, W., Howes, R., Huang, P.Y., Li, S.W., Misra, I., Rabbat, M., Sharma, V., Synnaeve, G., Xu, H., Jegou, H., Mairal, J., Labatut, P., Joulin, A., Bojanowski, P.: DINOv2: Learning Robust Visual Features without Supervision. arXiv preprint arXiv:2304.07193 (2023), http://arxiv.org/abs/2304.07193 +25. Ramaswamy, H.G., et al.: Ablation-cam: Visual explanations for deep convolutional network via gradient-free localization. In: proceedings of the IEEE/CVF winter conference on applications of computer vision. pp. 983-991 (2020) +26. Ranasinghe, K., McKinzie, B., Ravi, S., Yang, Y., Toshev, A., Shlens, J.: Perceptual grouping in contrastive vision-language models. In: Proceedings of the IEEE/CVF International Conference on Computer Vision. pp. 5571-5584 (2023) +27. Ribeiro, M.T., Singh, S., Guestrin, C.: "Why Should I Trust You?": Explaining the Predictions of Any Classifier. In: Proceedings of the 22nd ACM SIGKDD International Conference on Knowledge Discovery and Data Mining, San Francisco, CA, USA, August 13-17, 2016. pp. 1135-1144 (2016) +28. Rudin, C.: Stop explaining black box machine learning models for high stakes decisions and use interpretable models instead. Nature machine intelligence 1(5), 206-215 (2019) +29. Rudin, L.I., Osher, S., Fatemi, E.: Nonlinear total variation based noise removal algorithms. Physica D: nonlinear phenomena 60(1-4), 259-268 (1992) +30. Rymarczyk, D., Struski, Å., Tabor, J., Zielinski, B.: Protopshare: Prototypical parts sharing for similarity discovery in interpretable image classification. In: Proceedings of the 27th ACM SIGKDD Conference on Knowledge Discovery & Data Mining. pp. 1420-1430 (2021) +31. Selvaraju, R.R., Cogswell, M., Das, A., Vedantam, R., Parikh, D., Batra, D.: Gradcam: Visual explanations from deep networks via gradient-based localization. In: Proceedings of the IEEE international conference on computer vision. pp. 618-626 (2017) + +32. Shannon, C.E.: A Mathematical Theory of Communication. Bell System Technical Journal 27(3), 379-423 (1948). https://doi.org/10.1002/j.1538-7305.1948.tb01338.x +33. Simeoni, O., Puy, G., Vo, H.V., Roburin, S., Gidaris, S., Bursuc, A., Pérez, P., Marlet, R., Ponce, J.: Localizing objects with self-supervised transformers and no labels. arXiv preprint arXiv:2109.14279 (2021) +34. Srivastava, N., Hinton, G., Krizhevsky, A., Sutskever, I., Salakhutdinov, R.: Dropout: a simple way to prevent neural networks from overfitting. The journal of machine learning research 15(1), 1929-1958 (2014) +35. Strong, D., Chan, T.: Edge-preserving and scale-dependent properties of total variation regularization. Inverse problems 19(6), S165 (2003) +36. Unger, M., Pock, T., Trobin, W., Cremers, D., Bischof, H.: Tvseg-interactive total variation based image segmentation. In: BMVC. vol. 31, pp. 44-46. Citeseer (2008) +37. Wang, J., Liu, H., Wang, X., Jing, L.: Interpretable image recognition by constructing transparent embedding space. In: Proceedings of the IEEE/CVF International Conference on Computer Vision. pp. 895-904 (2021) +38. Welinder, P., Branson, S., Mita, T., Wah, C., Schroff, F., Belongie, S., Perona, P.: Caltech-UCSD Birds 200. Tech. Rep. CNS-TR-2010-001, California Institute of Technology (2010) +39. Xu, J., De Mello, S., Liu, S., Byeon, W., Breuel, T., Kautz, J., Wang, X.: Groupvit: Semantic segmentation emerges from text supervision. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 18134-18144 (2022) +40. Xu-Darme, R., Quenot, G., Chihani, Z., Rousset, M.C.: Particul: part identification with confidence measure using unsupervised learning. In: International Conference on Pattern Recognition. pp. 173-187. Springer (2022) +41. Yeh, R.A., Hu, Y.T., Ren, Z., Schwing, A.G.: Total variation optimization layers for computer vision. In: CVPR. pp. 711-721 (2022) \ No newline at end of file diff --git a/2024/PDiscoFormer_ Relaxing Part Discovery Constraints with Vision Transformers/images.zip b/2024/PDiscoFormer_ Relaxing Part Discovery Constraints with Vision Transformers/images.zip new file mode 100644 index 0000000000000000000000000000000000000000..4409accf5867d0f61fa48bb07d3f14a8ba654f0f --- /dev/null +++ b/2024/PDiscoFormer_ Relaxing Part Discovery Constraints with Vision Transformers/images.zip @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:eedd4cd81a593866191c81d9b8e917e174c7db43373c7b3c77b557ef7756b8a8 +size 842053 diff --git a/2024/PDiscoFormer_ Relaxing Part Discovery Constraints with Vision Transformers/layout.json b/2024/PDiscoFormer_ Relaxing Part Discovery Constraints with Vision Transformers/layout.json new file mode 100644 index 0000000000000000000000000000000000000000..c22b6c242cf2bfe8eacd433363aeab43fc6ec135 --- /dev/null +++ b/2024/PDiscoFormer_ Relaxing Part Discovery Constraints with Vision Transformers/layout.json @@ -0,0 +1,7564 @@ +{ + "pdf_info": [ + { + "para_blocks": [ + { + "bbox": [ + 167, + 112, + 446, + 146 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 167, + 112, + 446, + 146 + ], + "spans": [ + { + "bbox": [ + 167, + 112, + 446, + 146 + ], + "type": "text", + "content": "PDiscoFormer: Relaxing Part Discovery Constraints with Vision Transformers" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 138, + 168, + 474, + 181 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 168, + 474, + 181 + ], + "spans": [ + { + "bbox": [ + 138, + 168, + 474, + 181 + ], + "type": "text", + "content": "Ananthu Aniraj" + }, + { + "bbox": [ + 138, + 168, + 474, + 181 + ], + "type": "inline_equation", + "content": "^{1}" + }, + { + "bbox": [ + 138, + 168, + 474, + 181 + ], + "type": "text", + "content": ", Cassio F. Dantas" + }, + { + "bbox": [ + 138, + 168, + 474, + 181 + ], + "type": "inline_equation", + "content": "^{2}" + }, + { + "bbox": [ + 138, + 168, + 474, + 181 + ], + "type": "text", + "content": ", Dino Ienc" + }, + { + "bbox": [ + 138, + 168, + 474, + 181 + ], + "type": "inline_equation", + "content": "^{2}" + }, + { + "bbox": [ + 138, + 168, + 474, + 181 + ], + "type": "text", + "content": ", and Diego Marcos" + }, + { + "bbox": [ + 138, + 168, + 474, + 181 + ], + "type": "inline_equation", + "content": "^{1}" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 162, + 190, + 451, + 234 + ], + "type": "list", + "angle": 0, + "index": 4, + "blocks": [ + { + "bbox": [ + 162, + 190, + 451, + 212 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 162, + 190, + 451, + 212 + ], + "spans": [ + { + "bbox": [ + 162, + 190, + 451, + 212 + ], + "type": "text", + "content": "1 Inria, Univ. Montpellier, LIRMM, UMR TETIS, Montpellier, France {ananthu.aniraj, diego.marcos}@inria.fr" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 167, + 213, + 444, + 234 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 167, + 213, + 444, + 234 + ], + "spans": [ + { + "bbox": [ + 167, + 213, + 444, + 234 + ], + "type": "text", + "content": "2 Inria, Inrae, Univ. Montpellier, UMR TETIS, Montpellier, France {cassio.fraga-dantas, dino.ienco}@inrae.fr" + } + ] + } + ], + "index": 3 + } + ], + "sub_type": "text" + }, + { + "bbox": [ + 159, + 264, + 455, + 484 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 159, + 264, + 455, + 484 + ], + "spans": [ + { + "bbox": [ + 159, + 264, + 455, + 484 + ], + "type": "text", + "content": "Abstract. Computer vision methods that explicitly detect object parts and reason on them are a step towards inherently interpretable models. Existing approaches that perform part discovery driven by a fine-grained classification task make very restrictive assumptions on the geometric properties of the discovered parts; they should be small and compact. Although this prior is useful in some cases, in this paper we show that pre-trained transformer-based vision models, such as self-supervised DINOv2 ViT, enable the relaxation of these constraints. In particular, we find that a total variation (TV) prior, which allows for multiple connected components of any size, substantially outperforms previous work. We test our approach on three fine-grained classification benchmarks: CUB, PartImageNet and Oxford Flowers, and compare our results to previously published methods as well as a re-implementation of the state-of-the-art method PDiscoNet with a transformer-based backbone. We consistently obtain substantial improvements across the board, both on part discovery metrics and the downstream classification task, showing that the strong inductive biases in self-supervised ViT models require to rethink the geometric priors that can be used for unsupervised part discovery. Training code and pre-trained models are available at https://github.com/ananthu-aniraj/pdiscoformer." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 132, + 505, + 229, + 517 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 505, + 229, + 517 + ], + "spans": [ + { + "bbox": [ + 132, + 505, + 229, + 517 + ], + "type": "text", + "content": "1 Introduction" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 130, + 529, + 482, + 555 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 529, + 482, + 555 + ], + "spans": [ + { + "bbox": [ + 130, + 529, + 482, + 555 + ], + "type": "text", + "content": "Deep neural networks, first in the form of convolutional neural networks (CNN) and, more recently, vision transformers (ViT), have revolutionized the field of" + } + ] + } + ], + "index": 7 + }, + { + "type": "image", + "bbox": [ + 133, + 574, + 481, + 643 + ], + "blocks": [ + { + "bbox": [ + 133, + 574, + 481, + 643 + ], + "lines": [ + { + "bbox": [ + 133, + 574, + 481, + 643 + ], + "spans": [ + { + "bbox": [ + 133, + 574, + 481, + 643 + ], + "type": "image", + "image_path": "067bcb1bd9d7075196ae8e1b7e0ee8c8c055660bf65861a01f6d2b91f8b0eded.jpg" + } + ] + } + ], + "index": 8, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 193, + 651, + 418, + 662 + ], + "lines": [ + { + "bbox": [ + 193, + 651, + 418, + 662 + ], + "spans": [ + { + "bbox": [ + 193, + 651, + 418, + 662 + ], + "type": "text", + "content": "Fig. 1: Concentration vs total variation as part priors." + } + ] + } + ], + "index": 9, + "angle": 0, + "type": "image_caption" + } + ], + "index": 8 + } + ], + "discarded_blocks": [], + "page_size": [ + 612, + 792 + ], + "page_idx": 0 + }, + { + "para_blocks": [ + { + "bbox": [ + 130, + 116, + 482, + 201 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 116, + 482, + 201 + ], + "spans": [ + { + "bbox": [ + 130, + 116, + 482, + 201 + ], + "type": "text", + "content": "computer vision. However, their black-box nature often limits their application in domains where interpretability is crucial. To ameliorate the understanding of the internal behaviour of deep learning models trained for computer vision tasks, post-hoc interpretability methods [25,27,31] are commonly used. These methods generate saliency maps indicating regions in the image that contributed most to a classification decision. However, these are only approximations to the actual internal process of the model and may be misleading [28]." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 130, + 207, + 482, + 304 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 207, + 482, + 304 + ], + "spans": [ + { + "bbox": [ + 130, + 207, + 482, + 304 + ], + "type": "text", + "content": "Recent efforts have been directed towards developing interpretable-by-design deep neural networks [4, 13, 14, 17, 21, 22]. These approaches aim to identify semantically interpretable features in an image and utilize this information for solving a downstream task. Primarily tailored for fine-grained image classification, these methods seek to emulate human reasoning by subdividing the final task, such as classifying flower species or dog breeds, into multiple explicit intermediate stages, such as identifying lower-level semantic features like body parts, which in turn are used to provide a classification score." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 130, + 311, + 482, + 443 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 311, + 482, + 443 + ], + "spans": [ + { + "bbox": [ + 130, + 311, + 482, + 443 + ], + "type": "text", + "content": "However, improving interpretability while maintaining performance remains a challenge. Some models, such as those proposed in [13, 14, 17, 18], rely on objective functions that impose priors on the shape and area of object parts such as compactness-inducing part shaping losses that, in practice, result in spatial Gaussian priors of fixed variance on the part maps. While effective for specific datasets, like the commonly used CUB [38] benchmark for bird species classification, where assumptions like parts occurring only once in an image and being compact are beneficial, these assumptions may not generalize well to other use cases. For instance, in plant species classification, the presence of multiple instances of object parts such as leaves or flowers, as well as irregularly-shaped parts like branches or stems, renders these assumptions inadequate." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 130, + 450, + 482, + 666 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 450, + 482, + 666 + ], + "spans": [ + { + "bbox": [ + 130, + 450, + 482, + 666 + ], + "type": "text", + "content": "As we show in this work, such strong part shape priors are required when using a CNN-based backbone pre-trained on ImageNet; in order to reach acceptable classification results, it is imperative to fine-tune the whole backbone. This means that the only inductive bias that is kept throughout the part discovery process is the one brought in by the convolutional nature of the CNN architecture. A strong part shape prior is required in order to obtain consistent part maps in this setting, leading to all recent methods for part discovery in the context of fine-grained classification to use a concentration loss that makes part maps small and compact [13, 14, 17]. Here, we propose a method that addresses this issue by refraining from assuming any priors on part shape or area. The only prior we assume is that the part maps should tend to be piece-wise constant and thus be formed of one or more spatially connected components. Our model is capable of localizing up to a fixed number of semantically consistent parts of an object without relying on part annotations, using only the classification labels as supervision signals. The lack of assumptions about part shape and size makes our approach suitable to various fine-grained image classification tasks, as demonstrated by our empirical results, generalizing beyond cases where compactness is a good part prior." + } + ] + } + ], + "index": 5 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "spans": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "type": "text", + "content": "2" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 166, + 91, + 231, + 102 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 166, + 91, + 231, + 102 + ], + "spans": [ + { + "bbox": [ + 166, + 91, + 231, + 102 + ], + "type": "text", + "content": "A. Aniraj et al." + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 1 + }, + { + "para_blocks": [ + { + "bbox": [ + 130, + 116, + 482, + 213 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 116, + 482, + 213 + ], + "spans": [ + { + "bbox": [ + 130, + 116, + 482, + 213 + ], + "type": "text", + "content": "The main contribution of this paper is to build upon [17] and relax its constraining concentration-based part prior by a more flexible total variation prior. We investigate which additional elements are required to compensate for the flexibility of the total variation loss and find that a combination of an entropy loss and Gumbel-Softmax on the part maps allows for consistent and high quality part maps. Our approach makes better use of the representation of self-supervised pre-trained ViT models and allows to apply the model to tasks with diverse types of parts." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 132, + 231, + 238, + 244 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 231, + 238, + 244 + ], + "spans": [ + { + "bbox": [ + 132, + 231, + 238, + 244 + ], + "type": "text", + "content": "2 Related Work" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 130, + 258, + 482, + 449 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 258, + 482, + 449 + ], + "spans": [ + { + "bbox": [ + 130, + 258, + 482, + 449 + ], + "type": "text", + "content": "Unsupervised Part Discovery. Machine learning methods designed for unsupervised part discovery aim to identify a small number of semantically consistent parts, typically around " + }, + { + "bbox": [ + 130, + 258, + 482, + 449 + ], + "type": "inline_equation", + "content": "K = 8" + }, + { + "bbox": [ + 130, + 258, + 482, + 449 + ], + "type": "text", + "content": ", shared among classes in a dataset. These methods visualize all discovered parts in an image through per-part saliency maps, facilitating easier interpretation. However, as this task is unsupervised, it often relies on assumptions about the discovered parts, such as their distribution and shape. For instance, Huang et al. [13] assume that the presence of each part across a mini-batch follows a beta distribution, while [17] assumes that discovered parts are present at least once in a mini-batch. Additionally, assumptions are often made about the shape of the parts. Methods such as [14, 17, 18, 40] impose a spatial Gaussian prior with a fixed variance on the part maps, resulting in compact parts appearing once in the image. Such assumptions on part shape can be highly restrictive and may not generalize well across different scenarios. To address this limitation, our proposed method refrains from assuming such priors on part shape and only imposes a spatial prior that discovered parts should form connected components." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 130, + 449, + 482, + 617 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 449, + 482, + 617 + ], + "spans": [ + { + "bbox": [ + 130, + 449, + 482, + 617 + ], + "type": "text", + "content": "Part Prototype Networks. Chen et al. [4] introduced the Prototypical Part Network (ProtoPNet), an inherently interpretable method featuring a fixed number of learned prototypes per class. These prototypes represent visual patterns or features learned by the network during training. The model computes the similarity between each learned prototype and image patches and uses a linear combination of these similarity scores for classification. While explanations provided by ProtoPNet visualize all prototypes along with their weighted similarity scores, the sheer number of prototypes, often proportional to the number of classes, can lead to redundancy and interpretability challenges [10,21]. Subsequent works in this line of research have sought to address these challenges. Some approaches focus on reducing the number of prototypes shared per class [30,37], while others explore sharing prototypes between classes [20, 22]. Despite these efforts, recent works like [22] still utilize a large number (" + }, + { + "bbox": [ + 130, + 449, + 482, + 617 + ], + "type": "inline_equation", + "content": "K > 500" + }, + { + "bbox": [ + 130, + 449, + 482, + 617 + ], + "type": "text", + "content": ") of prototypes, which can pose significant challenges for end-user interpretation." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 130, + 617, + 482, + 666 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 617, + 482, + 666 + ], + "spans": [ + { + "bbox": [ + 130, + 617, + 482, + 666 + ], + "type": "text", + "content": "Regularization in Image Segmentation. Early works on image segmentation made direct use of low-level image features, chiefly color, along with geometric priors, in order to find coherent segments. The Mumford-Shah model [19], originally devised for piece-wise image reconstruction, was applied to find segments" + } + ] + } + ], + "index": 6 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 246, + 91, + 447, + 102 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 246, + 91, + 447, + 102 + ], + "spans": [ + { + "bbox": [ + 246, + 91, + 447, + 102 + ], + "type": "text", + "content": "PDiscoFormer: Part Discovery with Transformers" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 474, + 92, + 480, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 474, + 92, + 480, + 100 + ], + "spans": [ + { + "bbox": [ + 474, + 92, + 480, + 100 + ], + "type": "text", + "content": "3" + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 2 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 153, + 116, + 478, + 304 + ], + "blocks": [ + { + "bbox": [ + 153, + 116, + 478, + 304 + ], + "lines": [ + { + "bbox": [ + 153, + 116, + 478, + 304 + ], + "spans": [ + { + "bbox": [ + 153, + 116, + 478, + 304 + ], + "type": "image", + "image_path": "72f2233595f30e2531b43d5feca2bd0796ef12aabc0cbfd7ed18999f4312f4a9.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 183, + 312, + 430, + 324 + ], + "lines": [ + { + "bbox": [ + 183, + 312, + 430, + 324 + ], + "spans": [ + { + "bbox": [ + 183, + 312, + 430, + 324 + ], + "type": "text", + "content": "Fig. 2: The PDiscoFormer architecture for part discovery." + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_caption" + } + ], + "index": 2 + }, + { + "bbox": [ + 130, + 347, + 482, + 455 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 347, + 482, + 455 + ], + "spans": [ + { + "bbox": [ + 130, + 347, + 482, + 455 + ], + "type": "text", + "content": "that are as homogeneous as possible while trying to minimize the length of their boundary as a regularization [3]. Total variation (TV) [29] has been shown to be a powerful and efficient regularization to approximate this boundary length prior, and has been extensively used for piece-wise signal and image reconstruction [7,35], as well as unsupervised image segmentation [8,36]. With the advent of deep learning, TV regularization has proven valuable for style transfer [16] and edge-preserving smoothing [41]. In this work we explore the use of TV regularization for weakly-supervised part discovery in order to alleviate the issues brought by previously used priors." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 130, + 455, + 482, + 540 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 455, + 482, + 540 + ], + "spans": [ + { + "bbox": [ + 130, + 455, + 482, + 540 + ], + "type": "text", + "content": "Emergent grouping of ViT representations. Some recent research [26,39] has studied the emergence of object segmentation by supervising with image-text pairs. These methods learn to group similar patches in a ViT and assign them to textual concepts. They aim to identify the pixels in an input image corresponding to a predefined text prompt, focusing on segmenting whole objects rather than decomposing them into parts, as our model does with part discovery, and are limited to concepts that are represented in the text annotations." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 132, + 556, + 233, + 571 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 556, + 233, + 571 + ], + "spans": [ + { + "bbox": [ + 132, + 556, + 233, + 571 + ], + "type": "text", + "content": "3 Methodology" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 130, + 581, + 482, + 666 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 581, + 482, + 666 + ], + "spans": [ + { + "bbox": [ + 130, + 581, + 482, + 666 + ], + "type": "text", + "content": "Here we introduce an end-to-end trainable method (see Fig. 2) capable of automatically identifying " + }, + { + "bbox": [ + 130, + 581, + 482, + 666 + ], + "type": "inline_equation", + "content": "K" + }, + { + "bbox": [ + 130, + 581, + 482, + 666 + ], + "type": "text", + "content": " discriminative parts (semantically interpretable image regions) useful for solving a fine-grained image classification task. In contrast to earlier methods [13,17], our approach avoids any assumptions related to part shape, size or number of part instances. As in [13,17], our model is weakly supervised and relies solely on image-level class labels, thus uncovering discriminative parts without requiring any additional supervision. Our formulation is based" + } + ] + } + ], + "index": 7 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "spans": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "type": "text", + "content": "4" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 166, + 91, + 231, + 102 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 166, + 91, + 231, + 102 + ], + "spans": [ + { + "bbox": [ + 166, + 91, + 231, + 102 + ], + "type": "text", + "content": "A. Aniraj et al." + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 3 + }, + { + "para_blocks": [ + { + "bbox": [ + 130, + 116, + 479, + 163 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 116, + 479, + 163 + ], + "spans": [ + { + "bbox": [ + 130, + 116, + 479, + 163 + ], + "type": "text", + "content": "on [17], from which we take take most elements of the architecture, except for the Gumbel-Softmax and the layer normalization-based part modulation. The total variation, entropy and background losses have not been used in previous part discovery works and are shown to be important in our results." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 132, + 185, + 257, + 196 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 185, + 257, + 196 + ], + "spans": [ + { + "bbox": [ + 132, + 185, + 257, + 196 + ], + "type": "text", + "content": "3.1 Model architecture" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 130, + 206, + 479, + 241 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 206, + 479, + 241 + ], + "spans": [ + { + "bbox": [ + 130, + 206, + 479, + 241 + ], + "type": "text", + "content": "Let " + }, + { + "bbox": [ + 130, + 206, + 479, + 241 + ], + "type": "inline_equation", + "content": "\\mathbf{x} \\in \\mathbb{R}^{3 \\times M \\times N}" + }, + { + "bbox": [ + 130, + 206, + 479, + 241 + ], + "type": "text", + "content": " be an image in the dataset, and let " + }, + { + "bbox": [ + 130, + 206, + 479, + 241 + ], + "type": "inline_equation", + "content": "y \\in \\{1, 2, \\dots, C\\}" + }, + { + "bbox": [ + 130, + 206, + 479, + 241 + ], + "type": "text", + "content": " represent the corresponding classification label. Given a backbone model " + }, + { + "bbox": [ + 130, + 206, + 479, + 241 + ], + "type": "inline_equation", + "content": "h_{\\theta}(\\cdot)" + }, + { + "bbox": [ + 130, + 206, + 479, + 241 + ], + "type": "text", + "content": ", we extract a feature tensor:" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 258, + 243, + 480, + 258 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 258, + 243, + 480, + 258 + ], + "spans": [ + { + "bbox": [ + 258, + 243, + 480, + 258 + ], + "type": "interline_equation", + "content": "\\mathbf {z} = h _ {\\theta} (\\mathbf {x}) \\in \\mathbb {R} ^ {D \\times H \\times W} \\tag {1}", + "image_path": "8bd6147abc30a0a2bf5e795d6421e667914b9f5f12770ed79f7b7804bfac32fe.jpg" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 130, + 265, + 480, + 360 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 265, + 480, + 360 + ], + "spans": [ + { + "bbox": [ + 130, + 265, + 480, + 360 + ], + "type": "text", + "content": "We drop the class and register tokens of the ViT and reshape the patch tokens to obtain " + }, + { + "bbox": [ + 130, + 265, + 480, + 360 + ], + "type": "inline_equation", + "content": "\\mathbf{z}" + }, + { + "bbox": [ + 130, + 265, + 480, + 360 + ], + "type": "text", + "content": ". Following prior research [13, 14, 17], we compute attention maps " + }, + { + "bbox": [ + 130, + 265, + 480, + 360 + ], + "type": "inline_equation", + "content": "\\mathbf{A} \\in [0,1]^{(K + 1) \\times H \\times W}" + }, + { + "bbox": [ + 130, + 265, + 480, + 360 + ], + "type": "text", + "content": ", where the last channel in the first dimension is used to represent the background region in the image. These attention maps are computed using the negative squared Euclidean distance function between patch token features " + }, + { + "bbox": [ + 130, + 265, + 480, + 360 + ], + "type": "inline_equation", + "content": "\\mathbf{z}_{ij} \\in \\mathbb{R}^D" + }, + { + "bbox": [ + 130, + 265, + 480, + 360 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 130, + 265, + 480, + 360 + ], + "type": "inline_equation", + "content": "i \\in \\{1, \\dots, H\\}" + }, + { + "bbox": [ + 130, + 265, + 480, + 360 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 130, + 265, + 480, + 360 + ], + "type": "inline_equation", + "content": "j \\in \\{1, \\dots, W\\}" + }, + { + "bbox": [ + 130, + 265, + 480, + 360 + ], + "type": "text", + "content": " and each of the learnable prototypes " + }, + { + "bbox": [ + 130, + 265, + 480, + 360 + ], + "type": "inline_equation", + "content": "\\mathbf{p}^k \\in \\mathbb{R}^D" + }, + { + "bbox": [ + 130, + 265, + 480, + 360 + ], + "type": "text", + "content": ", with " + }, + { + "bbox": [ + 130, + 265, + 480, + 360 + ], + "type": "inline_equation", + "content": "k \\in \\{1, \\dots, K + 1\\}" + }, + { + "bbox": [ + 130, + 265, + 480, + 360 + ], + "type": "text", + "content": ", followed by a Gumbel-Softmax [15] across the " + }, + { + "bbox": [ + 130, + 265, + 480, + 360 + ], + "type": "inline_equation", + "content": "K + 1" + }, + { + "bbox": [ + 130, + 265, + 480, + 360 + ], + "type": "text", + "content": " channels:" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 228, + 373, + 480, + 402 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 228, + 373, + 480, + 402 + ], + "spans": [ + { + "bbox": [ + 228, + 373, + 480, + 402 + ], + "type": "interline_equation", + "content": "a _ {i j} ^ {k} = \\frac {\\exp \\left(- \\| \\mathbf {z} _ {i j} - \\mathbf {p} ^ {k} \\| ^ {2} + \\gamma_ {k}\\right)}{\\sum_ {l = 1} ^ {K + 1} \\exp \\left(- \\| \\mathbf {z} _ {i j} - \\mathbf {p} ^ {l} \\| ^ {2} + \\gamma_ {l}\\right)} \\tag {2}", + "image_path": "95552ee80858eb475662d36bf09e0c41a42fba45d9768d947d033feee2fe5bda.jpg" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 130, + 407, + 479, + 431 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 407, + 479, + 431 + ], + "spans": [ + { + "bbox": [ + 130, + 407, + 479, + 431 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 130, + 407, + 479, + 431 + ], + "type": "inline_equation", + "content": "\\gamma_{k}" + }, + { + "bbox": [ + 130, + 407, + 479, + 431 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 130, + 407, + 479, + 431 + ], + "type": "inline_equation", + "content": "\\gamma_{l}" + }, + { + "bbox": [ + 130, + 407, + 479, + 431 + ], + "type": "text", + "content": " are independently drawn samples from the Gumbel(0,1) distribution, and " + }, + { + "bbox": [ + 130, + 407, + 479, + 431 + ], + "type": "inline_equation", + "content": "a_{ij}^{k}" + }, + { + "bbox": [ + 130, + 407, + 479, + 431 + ], + "type": "text", + "content": " represents an element of " + }, + { + "bbox": [ + 130, + 407, + 479, + 431 + ], + "type": "inline_equation", + "content": "\\mathbf{A}" + }, + { + "bbox": [ + 130, + 407, + 479, + 431 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 130, + 432, + 479, + 467 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 432, + 479, + 467 + ], + "spans": [ + { + "bbox": [ + 130, + 432, + 479, + 467 + ], + "type": "text", + "content": "The attention maps are used to compute part embedding vectors " + }, + { + "bbox": [ + 130, + 432, + 479, + 467 + ], + "type": "inline_equation", + "content": "\\mathbf{v}^k\\in \\mathbb{R}^D" + }, + { + "bbox": [ + 130, + 432, + 479, + 467 + ], + "type": "text", + "content": " by using the attention values to calculate a weighted average over the feature tensor " + }, + { + "bbox": [ + 130, + 432, + 479, + 467 + ], + "type": "inline_equation", + "content": "\\mathbf{z}" + }, + { + "bbox": [ + 130, + 432, + 479, + 467 + ], + "type": "text", + "content": ":" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 265, + 480, + 480, + 506 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 265, + 480, + 480, + 506 + ], + "spans": [ + { + "bbox": [ + 265, + 480, + 480, + 506 + ], + "type": "interline_equation", + "content": "\\mathbf {v} ^ {k} = \\frac {\\sum_ {i} \\sum_ {j} a _ {i j} ^ {k} \\mathbf {z} _ {i j}}{H W} \\tag {3}", + "image_path": "e6d7c5d32698d0e1aaff436b6d10926a9b85bd88522884aa70b9bf2d20436342.jpg" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 130, + 512, + 479, + 548 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 512, + 479, + 548 + ], + "spans": [ + { + "bbox": [ + 130, + 512, + 479, + 548 + ], + "type": "text", + "content": "These embedding vectors are then modulated using a layer normalization [2] operation, with normalization applied over the feature and prototype dimensions. This operation helps to make each embedding vector distinct for the classifier:" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 236, + 559, + 480, + 587 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 236, + 559, + 480, + 587 + ], + "spans": [ + { + "bbox": [ + 236, + 559, + 480, + 587 + ], + "type": "interline_equation", + "content": "\\mathbf {v} _ {m} ^ {k} = \\frac {\\left(\\mathbf {v} ^ {k} - \\operatorname {E} \\left[ \\mathbf {v} ^ {k} \\right]\\right)}{\\sqrt {\\operatorname {V a r} \\left[ \\mathbf {v} ^ {k} \\right] + \\epsilon}} \\odot \\mathbf {w} _ {m} ^ {k} + \\mathbf {b} _ {m} ^ {k} \\tag {4}", + "image_path": "64aa99f21474c52c0242be3e9e2560f0225d607f8cb4fe9b1747e4de5ea5da13.jpg" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 130, + 594, + 479, + 642 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 594, + 479, + 642 + ], + "spans": [ + { + "bbox": [ + 130, + 594, + 479, + 642 + ], + "type": "text", + "content": "with " + }, + { + "bbox": [ + 130, + 594, + 479, + 642 + ], + "type": "inline_equation", + "content": "\\mathbf{w}_m^k, \\mathbf{b}_m^k \\in \\mathbb{R}^D" + }, + { + "bbox": [ + 130, + 594, + 479, + 642 + ], + "type": "text", + "content": " being the modulation weights and biases per part. These modulated embedding vectors " + }, + { + "bbox": [ + 130, + 594, + 479, + 642 + ], + "type": "inline_equation", + "content": "\\mathbf{v}_m^k" + }, + { + "bbox": [ + 130, + 594, + 479, + 642 + ], + "type": "text", + "content": " for " + }, + { + "bbox": [ + 130, + 594, + 479, + 642 + ], + "type": "inline_equation", + "content": "k \\in \\{1, \\dots, K\\}" + }, + { + "bbox": [ + 130, + 594, + 479, + 642 + ], + "type": "text", + "content": " (we omit the background embedding) are then linearly projected to obtain a vector of class scores conditioned on the part embedding:" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 277, + 653, + 480, + 666 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 277, + 653, + 480, + 666 + ], + "spans": [ + { + "bbox": [ + 277, + 653, + 480, + 666 + ], + "type": "interline_equation", + "content": "\\mathbf {y} ^ {k} = W _ {c} \\cdot \\mathbf {v} _ {m} ^ {k} \\tag {5}", + "image_path": "4e6c0ea1095c7093e2ab932ab79575082f71b2e37751ecbc54153f1bdd7edd4f.jpg" + } + ] + } + ], + "index": 14 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 246, + 91, + 448, + 102 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 246, + 91, + 448, + 102 + ], + "spans": [ + { + "bbox": [ + 246, + 91, + 448, + 102 + ], + "type": "text", + "content": "PDiscoFormer: Part Discovery with Transformers" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 474, + 92, + 480, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 474, + 92, + 480, + 100 + ], + "spans": [ + { + "bbox": [ + 474, + 92, + 480, + 100 + ], + "type": "text", + "content": "5" + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 4 + }, + { + "para_blocks": [ + { + "bbox": [ + 130, + 115, + 482, + 152 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 115, + 482, + 152 + ], + "spans": [ + { + "bbox": [ + 130, + 115, + 482, + 152 + ], + "type": "text", + "content": "The same linear classifier " + }, + { + "bbox": [ + 130, + 115, + 482, + 152 + ], + "type": "inline_equation", + "content": "W_{c} \\in \\mathbb{R}^{C \\times D}" + }, + { + "bbox": [ + 130, + 115, + 482, + 152 + ], + "type": "text", + "content": " is applied to each of the modulated embedding vectors, i.e., for each of the " + }, + { + "bbox": [ + 130, + 115, + 482, + 152 + ], + "type": "inline_equation", + "content": "K" + }, + { + "bbox": [ + 130, + 115, + 482, + 152 + ], + "type": "text", + "content": " discovered foreground parts. We then take the mean of the per-part class scores to obtain the classification scores:" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 276, + 162, + 481, + 194 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 276, + 162, + 481, + 194 + ], + "spans": [ + { + "bbox": [ + 276, + 162, + 481, + 194 + ], + "type": "interline_equation", + "content": "\\mathbf {y} = \\frac {1}{K} \\sum_ {k = 1} ^ {K} \\mathbf {y} ^ {k} \\tag {6}", + "image_path": "055be5f380210d014609cc84a26c47f9fc392d9cc404be30e012144ec85b1693.jpg" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 131, + 198, + 482, + 269 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 131, + 198, + 482, + 269 + ], + "spans": [ + { + "bbox": [ + 131, + 198, + 482, + 269 + ], + "type": "text", + "content": "Finally, a softmax function is employed to obtain the classification probabilities. Part dropout. We adopt dropout [34] on the modulated embedding vectors " + }, + { + "bbox": [ + 131, + 198, + 482, + 269 + ], + "type": "inline_equation", + "content": "\\mathbf{v}_m^k" + }, + { + "bbox": [ + 131, + 198, + 482, + 269 + ], + "type": "text", + "content": ", as proposed in [17], during model training. This drops entire part embedding vectors at a time and ensures that all discovered parts are discriminative, since discovering a single discriminative part would not be enough to solve the downstream classification task." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 132, + 285, + 236, + 297 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 285, + 236, + 297 + ], + "spans": [ + { + "bbox": [ + 132, + 285, + 236, + 297 + ], + "type": "text", + "content": "3.2 Loss Functions" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 130, + 304, + 481, + 339 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 304, + 481, + 339 + ], + "spans": [ + { + "bbox": [ + 130, + 304, + 481, + 339 + ], + "type": "text", + "content": "Our main contribution consists in proposing a set of part-shaping loss functions that are able to make the most out of the inductive biases learned by self-supervised ViT models." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 130, + 341, + 482, + 435 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 341, + 482, + 435 + ], + "spans": [ + { + "bbox": [ + 130, + 341, + 482, + 435 + ], + "type": "text", + "content": "Classification Loss " + }, + { + "bbox": [ + 130, + 341, + 482, + 435 + ], + "type": "inline_equation", + "content": "(\\mathcal{L}_{\\mathbf{c}})" + }, + { + "bbox": [ + 130, + 341, + 482, + 435 + ], + "type": "text", + "content": ". Our model is mainly driven by the fine-grained classification objective during training. We use the standard negative log likelihood or multi-class cross-entropy for this purpose. While the classification loss helps ensure the discriminative nature of the discovered parts, it may not directly enforce the learned attention maps to represent salient parts of the object in the image. To address this, we incorporate additional objective functions to enforce priors on the discovered parts. These objectives effectively guide the learning process to focus on relevant regions of the image." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 130, + 436, + 482, + 498 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 436, + 482, + 498 + ], + "spans": [ + { + "bbox": [ + 130, + 436, + 482, + 498 + ], + "type": "text", + "content": "Orthogonality Loss " + }, + { + "bbox": [ + 130, + 436, + 482, + 498 + ], + "type": "inline_equation", + "content": "(\\mathcal{L}_{\\perp})" + }, + { + "bbox": [ + 130, + 436, + 482, + 498 + ], + "type": "text", + "content": ". We aim to discover parts that are semantically distinct. To achieve this, we apply the orthogonality loss, which encourages the learned part embedding vectors to be decorrelated from one another. This is accomplished by minimizing the cosine similarity between the embedding vectors. We apply this loss on the modulated embedding vectors " + }, + { + "bbox": [ + 130, + 436, + 482, + 498 + ], + "type": "inline_equation", + "content": "\\mathbf{v}_m^k" + }, + { + "bbox": [ + 130, + 436, + 482, + 498 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 246, + 506, + 481, + 537 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 246, + 506, + 481, + 537 + ], + "spans": [ + { + "bbox": [ + 246, + 506, + 481, + 537 + ], + "type": "interline_equation", + "content": "\\mathcal {L} _ {\\perp} = \\sum_ {k = 1} ^ {K + 1} \\sum_ {l \\neq k} \\frac {\\mathbf {v} _ {m} ^ {k} \\cdot \\mathbf {v} _ {m} ^ {l}}{\\| \\mathbf {v} _ {m} ^ {k} \\| \\cdot \\| \\mathbf {v} _ {m} ^ {l} \\|} \\tag {7}", + "image_path": "ee361eef618e40a7759c5ede913daab306e991d7b058d6e55e31a919f174811c.jpg" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 130, + 545, + 482, + 666 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 545, + 482, + 666 + ], + "spans": [ + { + "bbox": [ + 130, + 545, + 482, + 666 + ], + "type": "text", + "content": "Equivalence Loss " + }, + { + "bbox": [ + 130, + 545, + 482, + 666 + ], + "type": "inline_equation", + "content": "(\\mathcal{L}_{\\mathrm{eq}})" + }, + { + "bbox": [ + 130, + 545, + 482, + 666 + ], + "type": "text", + "content": ". We aim to detect the same parts even if the image is translated, rotated, or scaled. To achieve this, we formulate the equivariance loss, which encourages the learned part attention maps to be equivariant to rigid transformations. Specifically, given an input image, we use an affine transform " + }, + { + "bbox": [ + 130, + 545, + 482, + 666 + ], + "type": "inline_equation", + "content": "T" + }, + { + "bbox": [ + 130, + 545, + 482, + 666 + ], + "type": "text", + "content": " with parameters randomly sampled from a pre-defined range and pass both the original image and the transformed image through our model. Then, we invert the transformation on the part attention maps of the transformed image. Finally, the equivariance loss is formulated as the cosine distance between the part attention maps of the original image and those of the corresponding transformed image." + } + ] + } + ], + "index": 10 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "spans": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "type": "text", + "content": "6" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 166, + 91, + 231, + 102 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 166, + 91, + 231, + 102 + ], + "spans": [ + { + "bbox": [ + 166, + 91, + 231, + 102 + ], + "type": "text", + "content": "A. Aniraj et al." + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 5 + }, + { + "para_blocks": [ + { + "bbox": [ + 130, + 115, + 479, + 140 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 115, + 479, + 140 + ], + "spans": [ + { + "bbox": [ + 130, + 115, + 479, + 140 + ], + "type": "text", + "content": "Let " + }, + { + "bbox": [ + 130, + 115, + 479, + 140 + ], + "type": "inline_equation", + "content": "A^k(\\mathbf{x})" + }, + { + "bbox": [ + 130, + 115, + 479, + 140 + ], + "type": "text", + "content": " be a function that returns the " + }, + { + "bbox": [ + 130, + 115, + 479, + 140 + ], + "type": "inline_equation", + "content": "k^{th}" + }, + { + "bbox": [ + 130, + 115, + 479, + 140 + ], + "type": "text", + "content": " attention map given an input image " + }, + { + "bbox": [ + 130, + 115, + 479, + 140 + ], + "type": "inline_equation", + "content": "\\mathbf{x}" + }, + { + "bbox": [ + 130, + 115, + 479, + 140 + ], + "type": "text", + "content": ". Then:" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 212, + 148, + 481, + 180 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 212, + 148, + 481, + 180 + ], + "spans": [ + { + "bbox": [ + 212, + 148, + 481, + 180 + ], + "type": "interline_equation", + "content": "\\mathcal {L} _ {\\mathrm {e q}} = 1 - \\frac {1}{K} \\sum_ {k} \\frac {\\left\\| A ^ {k} (\\mathbf {x}) \\cdot T ^ {- 1} \\left(A ^ {k} (T (\\mathbf {x})) \\right. \\right\\|}{\\| A ^ {k} (\\mathbf {x}) \\| \\cdot \\| A ^ {k} (T (\\mathbf {x})) \\|} \\tag {8}", + "image_path": "ff99f9411a111635fab17f00a15f7ceea19d6999e5357216f5bebd065604feaa.jpg" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 130, + 188, + 480, + 273 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 188, + 480, + 273 + ], + "spans": [ + { + "bbox": [ + 130, + 188, + 480, + 273 + ], + "type": "text", + "content": "Presence Loss " + }, + { + "bbox": [ + 130, + 188, + 480, + 273 + ], + "type": "inline_equation", + "content": "(\\mathcal{L}_{\\mathbf{p}})" + }, + { + "bbox": [ + 130, + 188, + 480, + 273 + ], + "type": "text", + "content": ". We assume that all discovered foreground parts are present in at least some images of the training dataset, while the background, the " + }, + { + "bbox": [ + 130, + 188, + 480, + 273 + ], + "type": "inline_equation", + "content": "(K + 1)^{th}" + }, + { + "bbox": [ + 130, + 188, + 480, + 273 + ], + "type": "text", + "content": " part, is expected to be present in all images in the dataset. To enforce this, we formulate the presence loss, which consists of two parts. Firstly, we define a loss applied only to the discovered foreground parts, ensuring that each part is present at least once in a mini-batch. Given a batch " + }, + { + "bbox": [ + 130, + 188, + 480, + 273 + ], + "type": "inline_equation", + "content": "\\{\\mathbf{x}_1,\\dots ,\\mathbf{x}_B\\}" + }, + { + "bbox": [ + 130, + 188, + 480, + 273 + ], + "type": "text", + "content": ", the presence loss is calculated as:" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 244, + 281, + 481, + 309 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 244, + 281, + 481, + 309 + ], + "spans": [ + { + "bbox": [ + 244, + 281, + 481, + 309 + ], + "type": "interline_equation", + "content": "\\mathcal {L} _ {\\mathrm {p} _ {1}} = 1 - \\frac {1}{K} \\sum_ {k} \\max _ {b, i, j} \\bar {a} _ {i j} ^ {k} \\left(\\mathbf {x} _ {b}\\right) \\tag {9}", + "image_path": "e5d074c49767049db21ce1f8613cf63cbc9605a6de20f2d054ee7b6e6ae07109.jpg" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 130, + 318, + 479, + 343 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 318, + 479, + 343 + ], + "spans": [ + { + "bbox": [ + 130, + 318, + 479, + 343 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 130, + 318, + 479, + 343 + ], + "type": "inline_equation", + "content": "\\bar{A}^k (\\mathbf{x}_b) = \\mathrm{avgpool}(A^k (\\mathbf{x}_b))" + }, + { + "bbox": [ + 130, + 318, + 479, + 343 + ], + "type": "text", + "content": " is the output of a 2D average pooling operator with a small kernel size employed to prevent single-pixel attention maps." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 130, + 344, + 479, + 378 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 344, + 479, + 378 + ], + "spans": [ + { + "bbox": [ + 130, + 344, + 479, + 378 + ], + "type": "text", + "content": "Secondly, to ensure the presence of the background, we introduce a stricter presence loss, which is satisfied if the background is present in all images of the mini-batch. This loss is formulated as follows:" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 219, + 386, + 481, + 415 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 219, + 386, + 481, + 415 + ], + "spans": [ + { + "bbox": [ + 219, + 386, + 481, + 415 + ], + "type": "interline_equation", + "content": "\\mathcal {L} _ {\\mathrm {p} _ {0}} = - \\frac {1}{B} \\sum_ {b} \\log \\left(\\max _ {i, j} m _ {i j} \\bar {a} _ {i j} ^ {K + 1} (\\mathbf {x} _ {b})\\right) \\tag {10}", + "image_path": "74cfe5584137db73386dd3f5b52c92cb34965fd6c5d630019526ecb7e9589bbf.jpg" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 130, + 424, + 479, + 450 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 424, + 479, + 450 + ], + "spans": [ + { + "bbox": [ + 130, + 424, + 479, + 450 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 130, + 424, + 479, + 450 + ], + "type": "inline_equation", + "content": "M = [m_{ij}]^{H \\times W}" + }, + { + "bbox": [ + 130, + 424, + 479, + 450 + ], + "type": "text", + "content": " is a soft mask with " + }, + { + "bbox": [ + 130, + 424, + 479, + 450 + ], + "type": "inline_equation", + "content": "m_{ij} \\in [0,1]" + }, + { + "bbox": [ + 130, + 424, + 479, + 450 + ], + "type": "text", + "content": " that privileges entries placed farther from the image center (towards the borders):" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 211, + 458, + 481, + 487 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 211, + 458, + 481, + 487 + ], + "spans": [ + { + "bbox": [ + 211, + 458, + 481, + 487 + ], + "type": "interline_equation", + "content": "m _ {i j} = 2 \\left(\\frac {i - 1}{H - 1} - \\frac {1}{2}\\right) ^ {2} + 2 \\left(\\frac {j - 1}{W - 1} - \\frac {1}{2}\\right) ^ {2} \\tag {11}", + "image_path": "b0ef0f775d6aafc62077d3b1b4512b1960bf7ef010952cfdb57590aaaedae85f.jpg" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 130, + 494, + 479, + 529 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 494, + 479, + 529 + ], + "spans": [ + { + "bbox": [ + 130, + 494, + 479, + 529 + ], + "type": "text", + "content": "This loss captures the notion that the background is at the same time expected in every image and is more likely to appear near the boundaries of the image." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 130, + 529, + 479, + 567 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 529, + 479, + 567 + ], + "spans": [ + { + "bbox": [ + 130, + 529, + 479, + 567 + ], + "type": "text", + "content": "Entropy Loss " + }, + { + "bbox": [ + 130, + 529, + 479, + 567 + ], + "type": "inline_equation", + "content": "(\\mathcal{L}_{\\mathrm{ent}})" + }, + { + "bbox": [ + 130, + 529, + 479, + 567 + ], + "type": "text", + "content": ". We aim to ensure that each patch token is assigned to a unique part. To achieve this, we formulate an entropy [32] loss applied to the part attention maps:" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 233, + 575, + 481, + 609 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 233, + 575, + 481, + 609 + ], + "spans": [ + { + "bbox": [ + 233, + 575, + 481, + 609 + ], + "type": "interline_equation", + "content": "\\mathcal {L} _ {\\text {e n t}} = \\frac {- 1}{K + 1} \\sum_ {k = 1} ^ {K + 1} \\sum_ {i j} a _ {i j} ^ {k} \\log \\left(a _ {i j} ^ {k}\\right) \\tag {12}", + "image_path": "02f4c8d0c242a730a8dfe11ef5d393a64fe008dfb842b9010e50325801fc7462.jpg" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 130, + 617, + 481, + 666 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 617, + 481, + 666 + ], + "spans": [ + { + "bbox": [ + 130, + 617, + 481, + 666 + ], + "type": "text", + "content": "Total Variation Loss " + }, + { + "bbox": [ + 130, + 617, + 481, + 666 + ], + "type": "inline_equation", + "content": "(\\mathcal{L}_{\\mathrm{tv}})" + }, + { + "bbox": [ + 130, + 617, + 481, + 666 + ], + "type": "text", + "content": ". We do not want to make any assumptions about the part shape. However, we would still like the discovered parts to be composed of one or, at most, a few connected components. To achieve this, we make use of a total variation loss [29] applied to the part attention maps:" + } + ] + } + ], + "index": 14 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 246, + 91, + 448, + 102 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 246, + 91, + 448, + 102 + ], + "spans": [ + { + "bbox": [ + 246, + 91, + 448, + 102 + ], + "type": "text", + "content": "PDiscoFormer: Part Discovery with Transformers" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 474, + 91, + 481, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 474, + 91, + 481, + 100 + ], + "spans": [ + { + "bbox": [ + 474, + 91, + 481, + 100 + ], + "type": "text", + "content": "7" + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 6 + }, + { + "para_blocks": [ + { + "bbox": [ + 249, + 125, + 481, + 159 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 249, + 125, + 481, + 159 + ], + "spans": [ + { + "bbox": [ + 249, + 125, + 481, + 159 + ], + "type": "interline_equation", + "content": "\\mathcal {L} _ {\\mathrm {t v}} = \\frac {1}{H W} \\sum_ {k = 1} ^ {K + 1} \\sum_ {i j} \\left| \\nabla a _ {i j} ^ {k} \\right| \\tag {13}", + "image_path": "837613dc7cd1bc8b6c83de503d0820a9d36f595a028a06f020c053155612d01e.jpg" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 132, + 163, + 447, + 177 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 163, + 447, + 177 + ], + "spans": [ + { + "bbox": [ + 132, + 163, + 447, + 177 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 132, + 163, + 447, + 177 + ], + "type": "inline_equation", + "content": "\\nabla a_{ij}^{k}" + }, + { + "bbox": [ + 132, + 163, + 447, + 177 + ], + "type": "text", + "content": " is the spatial image gradient of part map " + }, + { + "bbox": [ + 132, + 163, + 447, + 177 + ], + "type": "inline_equation", + "content": "A^k" + }, + { + "bbox": [ + 132, + 163, + 447, + 177 + ], + "type": "text", + "content": " in location " + }, + { + "bbox": [ + 132, + 163, + 447, + 177 + ], + "type": "inline_equation", + "content": "ij" + }, + { + "bbox": [ + 132, + 163, + 447, + 177 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 132, + 191, + 274, + 205 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 191, + 274, + 205 + ], + "spans": [ + { + "bbox": [ + 132, + 191, + 274, + 205 + ], + "type": "text", + "content": "4 Experimental Setup" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 130, + 214, + 482, + 297 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 214, + 482, + 297 + ], + "spans": [ + { + "bbox": [ + 130, + 214, + 482, + 297 + ], + "type": "text", + "content": "We aim to perform part discovery using only image-level class labels as supervision signals, assuming that the discovered parts are shared among classes, which is common in fine-grained image classification settings. Moreover, we aim to compare our model with similar methods in the literature in scenarios where they have been empirically shown to work well, as well as in new and more challenging scenarios where underlying data characteristics may require more flexible geometric modeling." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 130, + 297, + 482, + 346 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 297, + 482, + 346 + ], + "spans": [ + { + "bbox": [ + 130, + 297, + 482, + 346 + ], + "type": "text", + "content": "All models were implemented in PyTorch. We used the DinoV2 [24] ViT [9] model with register tokens [6] as the starting weights for all experiments, utilizing the Base (ViT-B) variant. The training settings and compute requirements of our model are detailed in the supplementary material." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 132, + 361, + 205, + 372 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 361, + 205, + 372 + ], + "spans": [ + { + "bbox": [ + 132, + 361, + 205, + 372 + ], + "type": "text", + "content": "4.1 Datasets" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 130, + 378, + 482, + 498 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 378, + 482, + 498 + ], + "spans": [ + { + "bbox": [ + 130, + 378, + 482, + 498 + ], + "type": "text", + "content": "CUB Dataset. The CUB dataset [38] consists of images depicting 200 bird species, with 5,994 images in the training set and 5,794 images in the testing set. Keypoint annotations for 15 different bird body parts are available for each image. Bird parts are typically compact, although their presence may vary in images due to changes in pose and occlusion. Most images contain only a single instance of a bird. For this dataset, we evaluate the results using keypoint regression (Kp) using ground truth part keypoint annotations to compare them with centroids discovered by the model. Furthermore, we compute the Normalized Mutual Information (NMI) and Adjusted Rand Index (ARI) to quantify the clustering quality of the discovered parts, as in related literature [5]." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 130, + 498, + 482, + 666 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 498, + 482, + 666 + ], + "spans": [ + { + "bbox": [ + 130, + 498, + 482, + 666 + ], + "type": "text", + "content": "PartImageNet Dataset. The PartImageNet dataset [11] comprises 158 classes distributed among 11 super-classes, such as airplanes, birds, and snakes. Each super-class is associated to between two and five possible parts, for a total of 41 part classes. Additionally, test images may contain multiple instances of objects, sometimes from different super-classes, while the classification label provides information about a single object. Notably, several super-classes, including snakes, airplanes, and reptiles, often exhibit irregular and non-compact part shapes. For our experiments, we use the PartImageNet 00D variant. Although originally intended for out-of-distribution detection, this variant is commonly used in related literature. We train our models on 14,865 training images and test on 1,658 images, respectively. For evaluation, we derive part centroids from ground truth part masks and calculate NMI and ARI scores to assess clustering quality. Additionally, we include an experiment on the newly released PartImageNet Seg variant in the supplementary material." + } + ] + } + ], + "index": 9 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "spans": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "type": "text", + "content": "8" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 166, + 91, + 231, + 102 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 166, + 91, + 231, + 102 + ], + "spans": [ + { + "bbox": [ + 166, + 91, + 231, + 102 + ], + "type": "text", + "content": "A. Aniraj et al." + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 7 + }, + { + "para_blocks": [ + { + "type": "table", + "bbox": [ + 132, + 157, + 481, + 355 + ], + "blocks": [ + { + "bbox": [ + 132, + 114, + 482, + 148 + ], + "lines": [ + { + "bbox": [ + 132, + 114, + 482, + 148 + ], + "spans": [ + { + "bbox": [ + 132, + 114, + 482, + 148 + ], + "type": "text", + "content": "Table 1: Quantitative comparison of the part discovery and classification capabilities of our model with related approaches from the literature. In this table, " + }, + { + "bbox": [ + 132, + 114, + 482, + 148 + ], + "type": "inline_equation", + "content": "K" + }, + { + "bbox": [ + 132, + 114, + 482, + 148 + ], + "type": "text", + "content": " refers to the number of parts. The method with the label ** does not use any class supervision." + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 132, + 157, + 481, + 355 + ], + "lines": [ + { + "bbox": [ + 132, + 157, + 481, + 355 + ], + "spans": [ + { + "bbox": [ + 132, + 157, + 481, + 355 + ], + "type": "table", + "html": "
MethodCUB(%)PartImageNet OOD(%)Flowers(%)
KKp ↓NMI ↑ARI ↑Top-1 Acc. ↑KNMI ↑ARI ↑Top-1 Acc. ↑KFg. mIoU ↑Top-1 Acc. ↑
Dino [1]**4-31.1811.21-819.177.59-254.95-
8-47.2119.76-2531.4614.16-455.11-
16-50.5726.14-5037.8116.50-854.44-
Huang [13]411.5129.7414.0487.3085.881.5374.22229.9293.07
811.6035.7215.9086.05257.561.2573.56433.2293.14
1612.6043.9221.1085.935010.191.0573.20817.2692.86
PDiscoNet [17]49.1237.8215.2686.17827.138.7688.58219.0477.51
88.5250.0826.9686.722532.4110.6989.00434.7683.05
167.6756.8738.0587.495041.4914.1786.06849.1081.04
PdiscoNet + ViT-B [6]47.7052.5926.6688.61819.2834.7290.9524.9292.75
86.3465.0137.9086.952528.2350.3590.2941.9595.48
165.9568.6343.4184.045029.4827.8089.69813.1897.40
PDiscoFormer (Ours)47.4158.1325.1189.06829.0052.4089.75273.6299.61
85.9969.8743.4988.792544.7159.2790.77473.3299.54
165.7473.3855.8388.725046.2962.2191.01869.5999.64
", + "image_path": "e9634cda15f95bda50d3bfa2c751819b30ebf6e558509fdc2fff15c41abc96de.jpg" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "table_body" + } + ], + "index": 3 + }, + { + "bbox": [ + 130, + 376, + 482, + 497 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 376, + 482, + 497 + ], + "spans": [ + { + "bbox": [ + 130, + 376, + 482, + 497 + ], + "type": "text", + "content": "Oxford 102 Flower Dataset. The Oxford 102 Flower Dataset [23] comprises 8,189 images divided into 102 flower classes. The training and validation sets contain 10 images per class, totaling 1,020 images each, while the test set consists of 6,149 images, with a minimum of 20 images per class. Since no part annotations are provided for this dataset, we evaluate the quality of the discovered parts by combining them and comparing the result to the provided foreground/background masks. We evaluate model performance by calculating the mean Intersection over Union (mIoU) for the foreground mask. This dataset often includes multiple instances of parts in a single image, such as flowers, as well as parts with irregular shapes, such as flower petals." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 132, + 514, + 230, + 528 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 514, + 230, + 528 + ], + "spans": [ + { + "bbox": [ + 132, + 514, + 230, + 528 + ], + "type": "text", + "content": "5 Experiments" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 132, + 538, + 331, + 551 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 538, + 331, + 551 + ], + "spans": [ + { + "bbox": [ + 132, + 538, + 331, + 551 + ], + "type": "text", + "content": "5.1 Comparison with State-of-the-Art" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 130, + 557, + 481, + 666 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 557, + 481, + 666 + ], + "spans": [ + { + "bbox": [ + 130, + 557, + 481, + 666 + ], + "type": "text", + "content": "We conduct a quantitative analysis of our model's part discovery and classification capabilities and compare it with those of related methods from the literature. Additionally, we compare our model against a re-implementation of the state-of-the-art PDiCoNet [17], where we replaced the backbone with the variant used in this paper. The results of this comparison are summarized in Tab. 1. From the results, our method consistently outperforms existing approaches for both part discovery and image classification across all datasets. We attribute this to the synergistic effects of the strong inductive biases ingrained in the self-supervised ViT backbone and the enhanced flexibility of the proposed combina" + } + ] + } + ], + "index": 7 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 246, + 91, + 447, + 102 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 246, + 91, + 447, + 102 + ], + "spans": [ + { + "bbox": [ + 246, + 91, + 447, + 102 + ], + "type": "text", + "content": "PDiscoFormer: Part Discovery with Transformers" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 474, + 92, + 480, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 474, + 92, + 480, + 100 + ], + "spans": [ + { + "bbox": [ + 474, + 92, + 480, + 100 + ], + "type": "text", + "content": "9" + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 8 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 133, + 114, + 487, + 261 + ], + "blocks": [ + { + "bbox": [ + 133, + 114, + 487, + 261 + ], + "lines": [ + { + "bbox": [ + 133, + 114, + 487, + 261 + ], + "spans": [ + { + "bbox": [ + 133, + 114, + 487, + 261 + ], + "type": "image", + "image_path": "6d9d70a64ff3ec74c6be094189be276cf5689922e8bdbda6841cba801ae1ab09.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_body" + } + ], + "index": 2 + }, + { + "type": "image", + "bbox": [ + 133, + 284, + 487, + 431 + ], + "blocks": [ + { + "bbox": [ + 211, + 272, + 401, + 283 + ], + "lines": [ + { + "bbox": [ + 211, + 272, + 401, + 283 + ], + "spans": [ + { + "bbox": [ + 211, + 272, + 401, + 283 + ], + "type": "text", + "content": "Fig. 3: Qualitative results on CUB for " + }, + { + "bbox": [ + 211, + 272, + 401, + 283 + ], + "type": "inline_equation", + "content": "K = 8" + }, + { + "bbox": [ + 211, + 272, + 401, + 283 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 133, + 284, + 487, + 431 + ], + "lines": [ + { + "bbox": [ + 133, + 284, + 487, + 431 + ], + "spans": [ + { + "bbox": [ + 133, + 284, + 487, + 431 + ], + "type": "image", + "image_path": "90f23c12cb2713a453831b957e2251654f6cb2a561f37c1e2c5c5284fb43e2d8.jpg" + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 193, + 441, + 420, + 453 + ], + "lines": [ + { + "bbox": [ + 193, + 441, + 420, + 453 + ], + "spans": [ + { + "bbox": [ + 193, + 441, + 420, + 453 + ], + "type": "text", + "content": "Fig. 4: Qualitative results on PartImageNet for " + }, + { + "bbox": [ + 193, + 441, + 420, + 453 + ], + "type": "inline_equation", + "content": "K = 8" + }, + { + "bbox": [ + 193, + 441, + 420, + 453 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "image_caption" + } + ], + "index": 4 + }, + { + "bbox": [ + 130, + 486, + 482, + 666 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 486, + 482, + 666 + ], + "spans": [ + { + "bbox": [ + 130, + 486, + 482, + 666 + ], + "type": "text", + "content": "tion of geometric priors. Particularly noteworthy are the significant performance gains observed on the Flowers and PartImageNet datasets, which contain objects with irregular shapes occurring in multiple instances within the same image, underscoring the importance of our geometric part prior that favors piece-wise constant and spatially connected components. Even on CUB, where most parts fit well into the compactness prior, we see a consistent improvements of results with respect to PDiscoNet, even when using the same ViT backbone. With 16 parts, our model achieves " + }, + { + "bbox": [ + 130, + 486, + 482, + 666 + ], + "type": "inline_equation", + "content": "55.8\\%" + }, + { + "bbox": [ + 130, + 486, + 482, + 666 + ], + "type": "text", + "content": " ARI and " + }, + { + "bbox": [ + 130, + 486, + 482, + 666 + ], + "type": "inline_equation", + "content": "73.4\\%" + }, + { + "bbox": [ + 130, + 486, + 482, + 666 + ], + "type": "text", + "content": " NMI, compared to " + }, + { + "bbox": [ + 130, + 486, + 482, + 666 + ], + "type": "inline_equation", + "content": "43.4\\%" + }, + { + "bbox": [ + 130, + 486, + 482, + 666 + ], + "type": "text", + "content": " ARI and " + }, + { + "bbox": [ + 130, + 486, + 482, + 666 + ], + "type": "inline_equation", + "content": "68.6\\%" + }, + { + "bbox": [ + 130, + 486, + 482, + 666 + ], + "type": "text", + "content": " NMI of the best competing method, PDiscoNet with a ViT backbone. Despite competitive performance in classification tasks on CUB and PartImageNet, the PdiscoNet+ViT-B model tends to degrade in accuracy as part complexity increases, particularly in CUB, unlike our model, which maintains robust accuracy for any number of parts. In the Flowers dataset, all competing methods result in subpar performance, while our approach excels at both detecting the foreground and classification no matter the number of parts. In this" + } + ] + } + ], + "index": 6 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "spans": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "type": "text", + "content": "10" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 166, + 91, + 231, + 102 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 166, + 91, + 231, + 102 + ], + "spans": [ + { + "bbox": [ + 166, + 91, + 231, + 102 + ], + "type": "text", + "content": "A. Aniraj et al." + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 9 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 133, + 114, + 487, + 261 + ], + "blocks": [ + { + "bbox": [ + 133, + 114, + 487, + 261 + ], + "lines": [ + { + "bbox": [ + 133, + 114, + 487, + 261 + ], + "spans": [ + { + "bbox": [ + 133, + 114, + 487, + 261 + ], + "type": "image", + "image_path": "7c2457fcb03457d8b8b5b415be86bd6763ac2661fd8dd858d2d2a975b9274320.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_body" + } + ], + "index": 2 + }, + { + "type": "image", + "bbox": [ + 133, + 283, + 487, + 430 + ], + "blocks": [ + { + "bbox": [ + 206, + 272, + 405, + 282 + ], + "lines": [ + { + "bbox": [ + 206, + 272, + 405, + 282 + ], + "spans": [ + { + "bbox": [ + 206, + 272, + 405, + 282 + ], + "type": "text", + "content": "Fig. 5: Qualitative results on Flowers for " + }, + { + "bbox": [ + 206, + 272, + 405, + 282 + ], + "type": "inline_equation", + "content": "K = 2" + }, + { + "bbox": [ + 206, + 272, + 405, + 282 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 133, + 283, + 487, + 430 + ], + "lines": [ + { + "bbox": [ + 133, + 283, + 487, + 430 + ], + "spans": [ + { + "bbox": [ + 133, + 283, + 487, + 430 + ], + "type": "image", + "image_path": "9edc6bf3b7342da549ca128a391e19717491e20d60be69c4d33456616af0617b.jpg" + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 132, + 441, + 482, + 475 + ], + "lines": [ + { + "bbox": [ + 132, + 441, + 482, + 475 + ], + "spans": [ + { + "bbox": [ + 132, + 441, + 482, + 475 + ], + "type": "text", + "content": "Fig. 6: Effect of " + }, + { + "bbox": [ + 132, + 441, + 482, + 475 + ], + "type": "inline_equation", + "content": "K" + }, + { + "bbox": [ + 132, + 441, + 482, + 475 + ], + "type": "text", + "content": " on PDiscoFormer. The left three images correspond to Flowers (from top to bottom, " + }, + { + "bbox": [ + 132, + 441, + 482, + 475 + ], + "type": "inline_equation", + "content": "K = 2,4,8" + }, + { + "bbox": [ + 132, + 441, + 482, + 475 + ], + "type": "text", + "content": "). The three middle ones are CUB, with " + }, + { + "bbox": [ + 132, + 441, + 482, + 475 + ], + "type": "inline_equation", + "content": "K" + }, + { + "bbox": [ + 132, + 441, + 482, + 475 + ], + "type": "text", + "content": " being 4, 8 and 16. The three on the right belong to PartImageNet, with " + }, + { + "bbox": [ + 132, + 441, + 482, + 475 + ], + "type": "inline_equation", + "content": "K = 8,25,50" + }, + { + "bbox": [ + 132, + 441, + 482, + 475 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "image_caption" + } + ], + "index": 4 + }, + { + "bbox": [ + 130, + 506, + 482, + 614 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 506, + 482, + 614 + ], + "spans": [ + { + "bbox": [ + 130, + 506, + 482, + 614 + ], + "type": "text", + "content": "dataset, PDiscoNet+ViT-B results in the second best classification accuracy, " + }, + { + "bbox": [ + 130, + 506, + 482, + 614 + ], + "type": "inline_equation", + "content": "97.4\\%" + }, + { + "bbox": [ + 130, + 506, + 482, + 614 + ], + "type": "text", + "content": ", after the " + }, + { + "bbox": [ + 130, + 506, + 482, + 614 + ], + "type": "inline_equation", + "content": "99.6\\%" + }, + { + "bbox": [ + 130, + 506, + 482, + 614 + ], + "type": "text", + "content": " obtained by our method, but completely fails to locate the foreground, with an mIoU of " + }, + { + "bbox": [ + 130, + 506, + 482, + 614 + ], + "type": "inline_equation", + "content": "13.2\\%" + }, + { + "bbox": [ + 130, + 506, + 482, + 614 + ], + "type": "text", + "content": ", against " + }, + { + "bbox": [ + 130, + 506, + 482, + 614 + ], + "type": "inline_equation", + "content": "73.6\\%" + }, + { + "bbox": [ + 130, + 506, + 482, + 614 + ], + "type": "text", + "content": " for PDiscoFormer. Additionally, we extend our comparisons by applying our proposed losses to a ResNet backbone and a fully frozen ViT architecture, showing that a pre-trained ViT is required to fully profit from the proposed set of losses and that fine-tuning the position, class and register tokens, while keeping the rest of the ViT frozen, results in better performances than leaving the whole ViT frozen. The results of these additional comparisons are presented in the supplementary material." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 130, + 617, + 482, + 666 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 617, + 482, + 666 + ], + "spans": [ + { + "bbox": [ + 130, + 617, + 482, + 666 + ], + "type": "text", + "content": "Qualitative Analysis. A qualitative analysis, as illustrated in Fig. 3, Fig. 4, Fig. 5 and Fig. 6 corroborates our quantitative findings, showcasing our model's ability to accurately discover irregularly shaped body parts such as bird wings (CUB), flower petals (Flowers) and the body of a snake (PartImageNet). At the" + } + ] + } + ], + "index": 7 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 246, + 91, + 447, + 102 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 246, + 91, + 447, + 102 + ], + "spans": [ + { + "bbox": [ + 246, + 91, + 447, + 102 + ], + "type": "text", + "content": "PDiscoFormer: Part Discovery with Transformers" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 470, + 92, + 479, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 470, + 92, + 479, + 100 + ], + "spans": [ + { + "bbox": [ + 470, + 92, + 479, + 100 + ], + "type": "text", + "content": "11" + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 10 + }, + { + "para_blocks": [ + { + "type": "table", + "bbox": [ + 160, + 135, + 455, + 293 + ], + "blocks": [ + { + "bbox": [ + 169, + 114, + 443, + 125 + ], + "lines": [ + { + "bbox": [ + 169, + 114, + 443, + 125 + ], + "spans": [ + { + "bbox": [ + 169, + 114, + 443, + 125 + ], + "type": "text", + "content": "Table 2: Results of the ablation study in PartImageNet and CUB." + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 160, + 135, + 455, + 293 + ], + "lines": [ + { + "bbox": [ + 160, + 135, + 455, + 293 + ], + "spans": [ + { + "bbox": [ + 160, + 135, + 455, + 293 + ], + "type": "table", + "html": "
PartImageNet OOD (K=25)CUB (K=16)
NMI ↑ARI ↑Top-1 Acc. ↑Kp ↓NMI ↑ARI ↑Top-1 Acc. ↑
Full Model44.7159.2790.775.7473.3855.8388.72
No Lp041.0549.0889.815.8568.6643.0983.64
No Lent39.1654.4691.195.5766.8845.6888.21
No Leq43.3854.1990.479.0556.8530.6383.71
No L⊥42.4458.8691.015.7970.5550.9088.63
No Lp133.9858.3191.077.1461.0434.2889.04
No Ltv33.7325.3590.476.0170.7851.8380.22
No Per-Part LayerNorm31.1440.7389.695.7470.7050.6269.73
No Part Dropout33.0843.9889.996.0870.2841.6987.14
No Gumbel32.9854.6890.716.4468.9243.6186.43
", + "image_path": "96cf7c2a388be523d8e690ea8103240a91b629021a00c7373d13a6da53fae183.jpg" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "table_body" + } + ], + "index": 3 + }, + { + "bbox": [ + 130, + 315, + 482, + 543 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 315, + 482, + 543 + ], + "spans": [ + { + "bbox": [ + 130, + 315, + 482, + 543 + ], + "type": "text", + "content": "same time, PDiscoNet [17] fails to discover these parts altogether, even when retrained on a ViT-B backbone. This indicates the usefulness of the flexible geometric priors enabled by our total variation loss. Compared to PDiscoNet [17], not only do we achieve a much tighter coverage of the object of interest, with our PDiscoFormer more closely following the object boundaries. This is particularly the case in parts that are relatively large, such as open bird wings and flower corollas. In addition to that, we can observe a substantial improvement in the consistency and semantic interpretability of parts. For instance, the parts discovered by our model for the four quadruped images displayed in Fig. 4 can immediately be interpreted (the green part corresponds to the snout while the brown one to the ears), while those from PDiscoNet are not clear. In Fig. 6 we vary the number of parts, " + }, + { + "bbox": [ + 130, + 315, + 482, + 543 + ], + "type": "inline_equation", + "content": "K" + }, + { + "bbox": [ + 130, + 315, + 482, + 543 + ], + "type": "text", + "content": ". On Flowers, using two parts leads to a clear assignment of one part to the flower calyx and one to the corolla, while more parts tends to exhibit oversegmentation. CUB, on the other hand, is able to absorb a larger number of parts. With " + }, + { + "bbox": [ + 130, + 315, + 482, + 543 + ], + "type": "inline_equation", + "content": "K = 4" + }, + { + "bbox": [ + 130, + 315, + 482, + 543 + ], + "type": "text", + "content": ", beak and legs are covered by the same part, while a fine-grained and consistent distinction of the different wing parts is achieved with " + }, + { + "bbox": [ + 130, + 315, + 482, + 543 + ], + "type": "inline_equation", + "content": "K = 16" + }, + { + "bbox": [ + 130, + 315, + 482, + 543 + ], + "type": "text", + "content": ". Interestingly, in the case of PartImageNet, the effective number of parts being used by a specific super-class remains almost constant when increasing " + }, + { + "bbox": [ + 130, + 315, + 482, + 543 + ], + "type": "inline_equation", + "content": "K" + }, + { + "bbox": [ + 130, + 315, + 482, + 543 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 132, + 561, + 238, + 573 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 561, + 238, + 573 + ], + "spans": [ + { + "bbox": [ + 132, + 561, + 238, + 573 + ], + "type": "text", + "content": "5.2 Ablation Study" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 130, + 581, + 482, + 667 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 581, + 482, + 667 + ], + "spans": [ + { + "bbox": [ + 130, + 581, + 482, + 667 + ], + "type": "text", + "content": "We conducted an ablation study by systematically removing one loss function or architectural element at a time and by replacing the Gumbel-Softmax with a regular Softmax. This was performed on the PartImageNet OOD dataset with " + }, + { + "bbox": [ + 130, + 581, + 482, + 667 + ], + "type": "inline_equation", + "content": "K = 25" + }, + { + "bbox": [ + 130, + 581, + 482, + 667 + ], + "type": "text", + "content": " parts and the CUB dataset with " + }, + { + "bbox": [ + 130, + 581, + 482, + 667 + ], + "type": "inline_equation", + "content": "K = 16" + }, + { + "bbox": [ + 130, + 581, + 482, + 667 + ], + "type": "text", + "content": " parts. The quantitative results of the ablation study are shown in Tab. 2 and some qualitative results can be seen in Fig. 7. From the results, we can observe the following: Total Variation " + }, + { + "bbox": [ + 130, + 581, + 482, + 667 + ], + "type": "inline_equation", + "content": "(\\mathcal{L}_{tv})" + }, + { + "bbox": [ + 130, + 581, + 482, + 667 + ], + "type": "text", + "content": " prevents spurious part activations in the background, possibly reducing" + } + ] + } + ], + "index": 6 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "spans": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "type": "text", + "content": "12" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 166, + 91, + 231, + 102 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 166, + 91, + 231, + 102 + ], + "spans": [ + { + "bbox": [ + 166, + 91, + 231, + 102 + ], + "type": "text", + "content": "A. Aniraj et al." + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 11 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 134, + 113, + 487, + 407 + ], + "blocks": [ + { + "bbox": [ + 134, + 113, + 487, + 407 + ], + "lines": [ + { + "bbox": [ + 134, + 113, + 487, + 407 + ], + "spans": [ + { + "bbox": [ + 134, + 113, + 487, + 407 + ], + "type": "image", + "image_path": "eaeb99af03887e3e9b1b7e4d6e8ffa81167451c6b916e49be4cf827a7f3db7ac.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 201, + 418, + 411, + 430 + ], + "lines": [ + { + "bbox": [ + 201, + 418, + 411, + 430 + ], + "spans": [ + { + "bbox": [ + 201, + 418, + 411, + 430 + ], + "type": "text", + "content": "Fig. 7: Qualitative results from the ablation study." + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_caption" + } + ], + "index": 2 + }, + { + "bbox": [ + 130, + 474, + 482, + 666 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 474, + 482, + 666 + ], + "spans": [ + { + "bbox": [ + 130, + 474, + 482, + 666 + ], + "type": "text", + "content": "information leakage between foreground parts and the background. It has the largest impact on PartImageNet in terms of ARI. The absence of the Background Loss " + }, + { + "bbox": [ + 130, + 474, + 482, + 666 + ], + "type": "inline_equation", + "content": "(\\mathcal{L}_{\\mathbf{p}_0})" + }, + { + "bbox": [ + 130, + 474, + 482, + 666 + ], + "type": "text", + "content": " results in the background part not being activated, causing the model to discover the background as a foreground part. Presence Loss " + }, + { + "bbox": [ + 130, + 474, + 482, + 666 + ], + "type": "inline_equation", + "content": "(\\mathcal{L}_{\\mathbf{p}})" + }, + { + "bbox": [ + 130, + 474, + 482, + 666 + ], + "type": "text", + "content": " and Part Dropout promote the utilization of all parts, preventing domination by a few parts which can lead to multiple unused parts (Fig. 7). Removing the Per-Part Layer Norm leads to inconsistent part discovery (Fig. 7) indicating that it contributes to the creation of discriminative part-specific features driven by the classification loss. Equivalence Loss " + }, + { + "bbox": [ + 130, + 474, + 482, + 666 + ], + "type": "inline_equation", + "content": "(\\mathcal{L}_{eq})" + }, + { + "bbox": [ + 130, + 474, + 482, + 666 + ], + "type": "text", + "content": " enhances robustness to affine geometric transformations, supporting stable part discovery across diverse inputs (Fig. 7). Gumbel-Softmax and Entropy Loss " + }, + { + "bbox": [ + 130, + 474, + 482, + 666 + ], + "type": "inline_equation", + "content": "(\\mathcal{L}_{\\mathrm{ent}})" + }, + { + "bbox": [ + 130, + 474, + 482, + 666 + ], + "type": "text", + "content": " ensure unique part assignments for each patch token, further reducing information leakage, as shown in our supplementary material. As shown in Tab. 2, the complete model, incorporating all loss functions and architectural choices, generally performs best for both part discovery and classification, demonstrating the efficacy of our approach." + } + ] + } + ], + "index": 4 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 246, + 91, + 447, + 102 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 246, + 91, + 447, + 102 + ], + "spans": [ + { + "bbox": [ + 246, + 91, + 447, + 102 + ], + "type": "text", + "content": "PDiscoFormer: Part Discovery with Transformers" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 470, + 91, + 481, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 470, + 91, + 481, + 100 + ], + "spans": [ + { + "bbox": [ + 470, + 91, + 481, + 100 + ], + "type": "text", + "content": "13" + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 12 + }, + { + "para_blocks": [ + { + "bbox": [ + 132, + 114, + 329, + 129 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 114, + 329, + 129 + ], + "spans": [ + { + "bbox": [ + 132, + 114, + 329, + 129 + ], + "type": "text", + "content": "6 Limitations and Future Work" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 130, + 137, + 482, + 389 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 137, + 482, + 389 + ], + "spans": [ + { + "bbox": [ + 130, + 137, + 482, + 389 + ], + "type": "text", + "content": "Our study focused on datasets with part annotations or FG-BG masks available for the test set. This allowed for quantitative comparison with state-of-the-art methods. However, exploring the training of our model on larger and more diverse datasets could provide valuable insights and further validate its performance in real-world scenarios. For example, larger datasets may introduce more variability in object appearances and poses, challenging the model to generalize better. Methods to enable the evaluation of part discovery in such settings would allow to take steps in this direction. The main hyper-parameter " + }, + { + "bbox": [ + 130, + 137, + 482, + 389 + ], + "type": "inline_equation", + "content": "K" + }, + { + "bbox": [ + 130, + 137, + 482, + 389 + ], + "type": "text", + "content": ", the number of parts to discover, has a large impact on the granularity of the discovered parts (Fig. 6). Exploring ways of automatically estimating this value while preserving interpretability in the discovered semantic parts would be a valuable contribution. This could enhance the adaptability of the model to various datasets and streamline the training process by removing the need for manual selection of " + }, + { + "bbox": [ + 130, + 137, + 482, + 389 + ], + "type": "inline_equation", + "content": "K" + }, + { + "bbox": [ + 130, + 137, + 482, + 389 + ], + "type": "text", + "content": ". Additionally, while our approach can detect parts of various shapes due to the total variation prior, it cannot ascertain if they belong to the same object. Integrating our method with unsupervised object discovery research [12,33] could enable part discovery within each detected object, improving capabilities in complex scenes. These methods aim to identify and localize individual objects without explicit supervision, complementing our approach for a more comprehensive scene understanding. These considerations suggest promising directions for extending and refining our method." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 132, + 403, + 220, + 416 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 403, + 220, + 416 + ], + "spans": [ + { + "bbox": [ + 132, + 403, + 220, + 416 + ], + "type": "text", + "content": "7 Conclusion" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 130, + 426, + 482, + 628 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 426, + 482, + 628 + ], + "spans": [ + { + "bbox": [ + 130, + 426, + 482, + 628 + ], + "type": "text", + "content": "In this paper, we presented PDiscoFormer, a new approach for unsupervised part discovery. Unlike previous methods that imposed strict geometric priors on discovered object parts, our model, based on self-supervised vision transformers, demonstrated the capability to relax these constraints. By leveraging a total variation (TV) prior, our model achieved substantial improvements in both part discovery and downstream classification accuracy when compared to the state-of-the-art across multiple datasets, specifically CUB, PartImageNet, and Oxford Flowers. Our findings highlight the importance of rethinking geometric priors in unsupervised part discovery tasks, especially in scenarios where objects exhibit complex shapes or occur multiple times within the same image. By refraining from assuming specific shape or size priors, our model demonstrated enhanced generalization capabilities, making it applicable to a broader range of fine-grained image classification tasks. Overall, our study shows that the strong inductive biases learned by vision transformers during the self-supervised pre-training stage enable the use of flexible shape priors such as total variation, leading to robust performance on part discovery and the downstream classification task on all of the tested datasets." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 132, + 629, + 481, + 665 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 629, + 481, + 665 + ], + "spans": [ + { + "bbox": [ + 132, + 629, + 481, + 665 + ], + "type": "text", + "content": "Acknowledgment. This work was supported in part by ANR project OBTEA ANR-22-CPJ1-0054-01 and, was granted access to the HPC resources of IDRIS under the allocation 2023-AD011014325 made by GENCI." + } + ] + } + ], + "index": 6 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "spans": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "type": "text", + "content": "14" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 166, + 91, + 231, + 102 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 166, + 91, + 231, + 102 + ], + "spans": [ + { + "bbox": [ + 166, + 91, + 231, + 102 + ], + "type": "text", + "content": "A. Aniraj et al." + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 13 + }, + { + "para_blocks": [ + { + "bbox": [ + 132, + 114, + 197, + 126 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 114, + 197, + 126 + ], + "spans": [ + { + "bbox": [ + 132, + 114, + 197, + 126 + ], + "type": "text", + "content": "References" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 138, + 141, + 480, + 665 + ], + "type": "list", + "angle": 0, + "index": 20, + "blocks": [ + { + "bbox": [ + 138, + 141, + 480, + 164 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 141, + 480, + 164 + ], + "spans": [ + { + "bbox": [ + 138, + 141, + 480, + 164 + ], + "type": "text", + "content": "1. Amir, S., Gandelsman, Y., Bagon, S., Dekel, T.: Deep vit features as dense visual descriptors. arXiv preprint arXiv:2112.05814 2(3), 4 (2021)" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 138, + 165, + 480, + 186 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 165, + 480, + 186 + ], + "spans": [ + { + "bbox": [ + 138, + 165, + 480, + 186 + ], + "type": "text", + "content": "2. Ba, J.L., Kiros, J.R., Hinton, G.E.: Layer normalization. arXiv preprint arXiv:1607.06450 (2016)" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 138, + 187, + 480, + 209 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 187, + 480, + 209 + ], + "spans": [ + { + "bbox": [ + 138, + 187, + 480, + 209 + ], + "type": "text", + "content": "3. Chan, T.F., Vese, L.A.: Active contours without edges. IEEE Transactions on image processing 10(2), 266-277 (2001)" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 138, + 209, + 480, + 243 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 209, + 480, + 243 + ], + "spans": [ + { + "bbox": [ + 138, + 209, + 480, + 243 + ], + "type": "text", + "content": "4. Chen, C., Li, O., Tao, D., Barnett, A., Rudin, C., Su, J.K.: This looks like that: deep learning for interpretable image recognition. Advances in neural information processing systems 32 (2019)" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 138, + 243, + 480, + 276 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 243, + 480, + 276 + ], + "spans": [ + { + "bbox": [ + 138, + 243, + 480, + 276 + ], + "type": "text", + "content": "5. Choudhury, S., Laina, I., Rupprecht, C., Vedaldi, A.: Unsupervised part discovery from contrastive reconstruction. Advances in Neural Information Processing Systems 34, 28104-28118 (2021)" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 138, + 277, + 480, + 298 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 277, + 480, + 298 + ], + "spans": [ + { + "bbox": [ + 138, + 277, + 480, + 298 + ], + "type": "text", + "content": "6. Darcet, T., Oquab, M., Mairal, J., Bojanowski, P.: Vision transformers need registers. In: The Twelfth International Conference on Learning Representations (2023)" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 138, + 299, + 480, + 320 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 299, + 480, + 320 + ], + "spans": [ + { + "bbox": [ + 138, + 299, + 480, + 320 + ], + "type": "text", + "content": "7. Dobson, D.C., Santosa, F.: Recovery of blocky images from noisy and blurred data. SIAM Journal on Applied Mathematics 56(4), 1181-1198 (1996)" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 138, + 321, + 480, + 342 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 321, + 480, + 342 + ], + "spans": [ + { + "bbox": [ + 138, + 321, + 480, + 342 + ], + "type": "text", + "content": "8. Donoser, M., Urschler, M., Hirzer, M., Bischof, H.: Saliency driven total variation segmentation. In: ICCV. pp. 817-824. IEEE (2009)" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 138, + 343, + 480, + 387 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 343, + 480, + 387 + ], + "spans": [ + { + "bbox": [ + 138, + 343, + 480, + 387 + ], + "type": "text", + "content": "9. Dosovitskiy, A., Beyer, L., Kolesnikov, A., Weissenborn, D., Zhai, X., Unterthiner, T., Dehghani, M., Minderer, M., Heigold, G., Gelly, S., et al.: An image is worth 16x16 words: Transformers for image recognition at scale. In: International Conference on Learning Representations (2020)" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 138, + 388, + 480, + 421 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 388, + 480, + 421 + ], + "spans": [ + { + "bbox": [ + 138, + 388, + 480, + 421 + ], + "type": "text", + "content": "0. Gautam, S., Höhne, M.M.C., Hansen, S., Jenssen, R., Kampffmeyer, M.: This looks more like that: Enhancing self-explaining models by prototypical relevance propagation. Pattern Recognition 136, 109172 (2023)" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 138, + 422, + 480, + 454 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 422, + 480, + 454 + ], + "spans": [ + { + "bbox": [ + 138, + 422, + 480, + 454 + ], + "type": "text", + "content": "1. He, J., Yang, S., Yang, S., Kortylewski, A., Yuan, X., Chen, J.N., Liu, S., Yang, C., Yu, Q., Yuille, A.: Partimagenet: A large, high-quality dataset of parts. In: European Conference on Computer Vision. pp. 128-145. Springer (2022)" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 138, + 455, + 480, + 487 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 455, + 480, + 487 + ], + "spans": [ + { + "bbox": [ + 138, + 455, + 480, + 487 + ], + "type": "text", + "content": "2. Henaff, O.J., Koppula, S., Shelhamer, E., Zoran, D., Jaegle, A., Zisserman, A., Carreira, J., Arandjelovic, R.: Object discovery and representation networks. In: European Conference on Computer Vision. pp. 123-143. Springer (2022)" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 138, + 488, + 480, + 520 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 488, + 480, + 520 + ], + "spans": [ + { + "bbox": [ + 138, + 488, + 480, + 520 + ], + "type": "text", + "content": "3. Huang, Z., Li, Y.: Interpretable and accurate fine-grained recognition via region grouping. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 8662-8672 (2020)" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 138, + 521, + 480, + 554 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 521, + 480, + 554 + ], + "spans": [ + { + "bbox": [ + 138, + 521, + 480, + 554 + ], + "type": "text", + "content": "4. Hung, W.C., Jampani, V., Liu, S., Molchanov, P., Yang, M.H., Kautz, J.: SCOPS: Self-Supervised Co-Part Segmentation. In: IEEE Conf. on Computer Vision and Pattern Recognition (CVPR) (2019)" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 138, + 555, + 480, + 576 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 555, + 480, + 576 + ], + "spans": [ + { + "bbox": [ + 138, + 555, + 480, + 576 + ], + "type": "text", + "content": "5. Jang, E., Gu, S., Poole, B.: Categorical reparameterization with gumbel-softmax. In: International Conference on Learning Representations (2017)" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 138, + 577, + 480, + 620 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 577, + 480, + 620 + ], + "spans": [ + { + "bbox": [ + 138, + 577, + 480, + 620 + ], + "type": "text", + "content": "6. Johnson, J., Alahi, A., Fei-Fei, L.: Perceptual losses for real-time style transfer and super-resolution. In: Computer Vision-ECCV 2016: 14th European Conference, Amsterdam, The Netherlands, October 11-14, 2016, Proceedings, Part II 14. pp. 694-711. Springer (2016)" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 138, + 621, + 480, + 665 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 621, + 480, + 665 + ], + "spans": [ + { + "bbox": [ + 138, + 621, + 480, + 665 + ], + "type": "text", + "content": "7. van der Klis, R., Alaniz, S., Mancini, M., Dantas, C.F., Ienco, D., Akata, Z., Marcos, D.: PDiscoNet: Semantically consistent part discovery for fine-grained recognition. In: Proceedings of the IEEE/CVF International Conference on Computer Vision. pp. 1866-1876 (2023)" + } + ] + } + ], + "index": 19 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 246, + 91, + 447, + 102 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 246, + 91, + 447, + 102 + ], + "spans": [ + { + "bbox": [ + 246, + 91, + 447, + 102 + ], + "type": "text", + "content": "PDiscoFormer: Part Discovery with Transformers" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 470, + 92, + 480, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 470, + 92, + 480, + 100 + ], + "spans": [ + { + "bbox": [ + 470, + 92, + 480, + 100 + ], + "type": "text", + "content": "15" + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 14 + }, + { + "para_blocks": [ + { + "bbox": [ + 132, + 116, + 481, + 666 + ], + "type": "list", + "angle": 0, + "index": 16, + "blocks": [ + { + "bbox": [ + 133, + 116, + 481, + 162 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 116, + 481, + 162 + ], + "spans": [ + { + "bbox": [ + 133, + 116, + 481, + 162 + ], + "type": "text", + "content": "18. Marcos, D., Kierdorf, J., Cheeseman, T., Tuia, D., Roscher, R.: A whale's tail-finding the right whale in an uncertain world. In: International Workshop on Extending Explainable AI Beyond Deep Models and Classifiers. pp. 297-313. Springer (2020)" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 132, + 163, + 481, + 195 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 163, + 481, + 195 + ], + "spans": [ + { + "bbox": [ + 132, + 163, + 481, + 195 + ], + "type": "text", + "content": "19. Mumford, D.B., Shah, J.: Optimal approximations by piecewise smooth functions and associated variational problems. Communications on pure and applied mathematics (1989)" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 132, + 196, + 481, + 229 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 196, + 481, + 229 + ], + "spans": [ + { + "bbox": [ + 132, + 196, + 481, + 229 + ], + "type": "text", + "content": "20. Nauta, M., van Bree, R., Seifert, C.: Neural Prototype Trees for Interpretable Fine-Grained Image Recognition. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR). pp. 14933-14943 (6 2021)" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 132, + 230, + 481, + 274 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 230, + 481, + 274 + ], + "spans": [ + { + "bbox": [ + 132, + 230, + 481, + 274 + ], + "type": "text", + "content": "21. Nauta, M., Jutte, A., Provoost, J., Seifert, C.: This looks like that, because... explaining prototypes for interpretable image recognition. In: Joint European Conference on Machine Learning and Knowledge Discovery in Databases. pp. 441-456. Springer (2021)" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 132, + 274, + 481, + 308 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 274, + 481, + 308 + ], + "spans": [ + { + "bbox": [ + 132, + 274, + 481, + 308 + ], + "type": "text", + "content": "22. Nauta, M., Schlötterer, J., van Keulen, M., Seifert, C.: PIP-Net: Patch-Based Intuitive Prototypes for Interpretable Image Classification. Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (2023)" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 132, + 308, + 481, + 342 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 308, + 481, + 342 + ], + "spans": [ + { + "bbox": [ + 132, + 308, + 481, + 342 + ], + "type": "text", + "content": "23. Nilsback, M.E., Zisserman, A.: Automated flower classification over a large number of classes. In: Indian Conference on Computer Vision, Graphics and Image Processing (Dec 2008)" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 132, + 342, + 481, + 407 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 342, + 481, + 407 + ], + "spans": [ + { + "bbox": [ + 132, + 342, + 481, + 407 + ], + "type": "text", + "content": "24. Oquab, M., Darcet, T., Moutakanni, T., Vo, H., Szafraniec, M., Khalidov, V., Fernandez, P., Haziza, D., Massa, F., El-Nouby, A., Assran, M., Ballas, N., Galuba, W., Howes, R., Huang, P.Y., Li, S.W., Misra, I., Rabbat, M., Sharma, V., Synnaeve, G., Xu, H., Jegou, H., Mairal, J., Labatut, P., Joulin, A., Bojanowski, P.: DINOv2: Learning Robust Visual Features without Supervision. arXiv preprint arXiv:2304.07193 (2023), http://arxiv.org/abs/2304.07193" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 132, + 408, + 481, + 441 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 408, + 481, + 441 + ], + "spans": [ + { + "bbox": [ + 132, + 408, + 481, + 441 + ], + "type": "text", + "content": "25. Ramaswamy, H.G., et al.: Ablation-cam: Visual explanations for deep convolutional network via gradient-free localization. In: proceedings of the IEEE/CVF winter conference on applications of computer vision. pp. 983-991 (2020)" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 132, + 442, + 481, + 475 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 442, + 481, + 475 + ], + "spans": [ + { + "bbox": [ + 132, + 442, + 481, + 475 + ], + "type": "text", + "content": "26. Ranasinghe, K., McKinzie, B., Ravi, S., Yang, Y., Toshev, A., Shlens, J.: Perceptual grouping in contrastive vision-language models. In: Proceedings of the IEEE/CVF International Conference on Computer Vision. pp. 5571-5584 (2023)" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 132, + 475, + 481, + 520 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 475, + 481, + 520 + ], + "spans": [ + { + "bbox": [ + 132, + 475, + 481, + 520 + ], + "type": "text", + "content": "27. Ribeiro, M.T., Singh, S., Guestrin, C.: \"Why Should I Trust You?\": Explaining the Predictions of Any Classifier. In: Proceedings of the 22nd ACM SIGKDD International Conference on Knowledge Discovery and Data Mining, San Francisco, CA, USA, August 13-17, 2016. pp. 1135-1144 (2016)" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 132, + 521, + 481, + 553 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 521, + 481, + 553 + ], + "spans": [ + { + "bbox": [ + 132, + 521, + 481, + 553 + ], + "type": "text", + "content": "28. Rudin, C.: Stop explaining black box machine learning models for high stakes decisions and use interpretable models instead. Nature machine intelligence 1(5), 206-215 (2019)" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 132, + 554, + 481, + 576 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 554, + 481, + 576 + ], + "spans": [ + { + "bbox": [ + 132, + 554, + 481, + 576 + ], + "type": "text", + "content": "29. Rudin, L.I., Osher, S., Fatemi, E.: Nonlinear total variation based noise removal algorithms. Physica D: nonlinear phenomena 60(1-4), 259-268 (1992)" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 132, + 577, + 481, + 620 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 577, + 481, + 620 + ], + "spans": [ + { + "bbox": [ + 132, + 577, + 481, + 620 + ], + "type": "text", + "content": "30. Rymarczyk, D., Struski, Å., Tabor, J., Zielinski, B.: Protopshare: Prototypical parts sharing for similarity discovery in interpretable image classification. In: Proceedings of the 27th ACM SIGKDD Conference on Knowledge Discovery & Data Mining. pp. 1420-1430 (2021)" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 132, + 621, + 481, + 666 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 621, + 481, + 666 + ], + "spans": [ + { + "bbox": [ + 132, + 621, + 481, + 666 + ], + "type": "text", + "content": "31. Selvaraju, R.R., Cogswell, M., Das, A., Vedantam, R., Parikh, D., Batra, D.: Gradcam: Visual explanations from deep networks via gradient-based localization. In: Proceedings of the IEEE international conference on computer vision. pp. 618-626 (2017)" + } + ] + } + ], + "index": 15 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "spans": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "type": "text", + "content": "16" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 166, + 91, + 230, + 102 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 166, + 91, + 230, + 102 + ], + "spans": [ + { + "bbox": [ + 166, + 91, + 230, + 102 + ], + "type": "text", + "content": "A. Aniraj et al." + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 15 + }, + { + "para_blocks": [ + { + "bbox": [ + 130, + 116, + 482, + 424 + ], + "type": "list", + "angle": 0, + "index": 12, + "blocks": [ + { + "bbox": [ + 130, + 116, + 482, + 148 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 116, + 482, + 148 + ], + "spans": [ + { + "bbox": [ + 130, + 116, + 482, + 148 + ], + "type": "text", + "content": "32. Shannon, C.E.: A Mathematical Theory of Communication. Bell System Technical Journal 27(3), 379-423 (1948). https://doi.org/10.1002/j.1538-7305.1948.tb01338.x" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 130, + 149, + 482, + 182 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 149, + 482, + 182 + ], + "spans": [ + { + "bbox": [ + 130, + 149, + 482, + 182 + ], + "type": "text", + "content": "33. Simeoni, O., Puy, G., Vo, H.V., Roburin, S., Gidaris, S., Bursuc, A., Pérez, P., Marlet, R., Ponce, J.: Localizing objects with self-supervised transformers and no labels. arXiv preprint arXiv:2109.14279 (2021)" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 130, + 182, + 482, + 215 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 182, + 482, + 215 + ], + "spans": [ + { + "bbox": [ + 130, + 182, + 482, + 215 + ], + "type": "text", + "content": "34. Srivastava, N., Hinton, G., Krizhevsky, A., Sutskever, I., Salakhutdinov, R.: Dropout: a simple way to prevent neural networks from overfitting. The journal of machine learning research 15(1), 1929-1958 (2014)" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 130, + 215, + 482, + 237 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 215, + 482, + 237 + ], + "spans": [ + { + "bbox": [ + 130, + 215, + 482, + 237 + ], + "type": "text", + "content": "35. Strong, D., Chan, T.: Edge-preserving and scale-dependent properties of total variation regularization. Inverse problems 19(6), S165 (2003)" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 130, + 237, + 482, + 259 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 237, + 482, + 259 + ], + "spans": [ + { + "bbox": [ + 130, + 237, + 482, + 259 + ], + "type": "text", + "content": "36. Unger, M., Pock, T., Trobin, W., Cremers, D., Bischof, H.: Tvseg-interactive total variation based image segmentation. In: BMVC. vol. 31, pp. 44-46. Citeseer (2008)" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 130, + 259, + 482, + 293 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 259, + 482, + 293 + ], + "spans": [ + { + "bbox": [ + 130, + 259, + 482, + 293 + ], + "type": "text", + "content": "37. Wang, J., Liu, H., Wang, X., Jing, L.: Interpretable image recognition by constructing transparent embedding space. In: Proceedings of the IEEE/CVF International Conference on Computer Vision. pp. 895-904 (2021)" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 130, + 293, + 482, + 326 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 293, + 482, + 326 + ], + "spans": [ + { + "bbox": [ + 130, + 293, + 482, + 326 + ], + "type": "text", + "content": "38. Welinder, P., Branson, S., Mita, T., Wah, C., Schroff, F., Belongie, S., Perona, P.: Caltech-UCSD Birds 200. Tech. Rep. CNS-TR-2010-001, California Institute of Technology (2010)" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 130, + 326, + 482, + 369 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 326, + 482, + 369 + ], + "spans": [ + { + "bbox": [ + 130, + 326, + 482, + 369 + ], + "type": "text", + "content": "39. Xu, J., De Mello, S., Liu, S., Byeon, W., Breuel, T., Kautz, J., Wang, X.: Groupvit: Semantic segmentation emerges from text supervision. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 18134-18144 (2022)" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 130, + 369, + 482, + 402 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 369, + 482, + 402 + ], + "spans": [ + { + "bbox": [ + 130, + 369, + 482, + 402 + ], + "type": "text", + "content": "40. Xu-Darme, R., Quenot, G., Chihani, Z., Rousset, M.C.: Particul: part identification with confidence measure using unsupervised learning. In: International Conference on Pattern Recognition. pp. 173-187. Springer (2022)" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 130, + 402, + 482, + 424 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 402, + 482, + 424 + ], + "spans": [ + { + "bbox": [ + 130, + 402, + 482, + 424 + ], + "type": "text", + "content": "41. Yeh, R.A., Hu, Y.T., Ren, Z., Schwing, A.G.: Total variation optimization layers for computer vision. In: CVPR. pp. 711-721 (2022)" + } + ] + } + ], + "index": 11 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 246, + 91, + 447, + 102 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 246, + 91, + 447, + 102 + ], + "spans": [ + { + "bbox": [ + 246, + 91, + 447, + 102 + ], + "type": "text", + "content": "PDiscoFormer: Part Discovery with Transformers" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 470, + 92, + 481, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 470, + 92, + 481, + 100 + ], + "spans": [ + { + "bbox": [ + 470, + 92, + 481, + 100 + ], + "type": "text", + "content": "17" + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 16 + } + ], + "_backend": "vlm", + "_version_name": "2.6.4" +} \ No newline at end of file diff --git a/2024/PEA-Diffusion_ Parameter-Efficient Adapter with Knowledge Distillation in non-English Text-to-Image Generation/e154c47c-6f75-47c7-a454-a9dcc00f12d4_content_list.json b/2024/PEA-Diffusion_ Parameter-Efficient Adapter with Knowledge Distillation in non-English Text-to-Image Generation/e154c47c-6f75-47c7-a454-a9dcc00f12d4_content_list.json new file mode 100644 index 0000000000000000000000000000000000000000..5bad73ff4e50c14006768900811bc104a833d788 --- /dev/null +++ b/2024/PEA-Diffusion_ Parameter-Efficient Adapter with Knowledge Distillation in non-English Text-to-Image Generation/e154c47c-6f75-47c7-a454-a9dcc00f12d4_content_list.json @@ -0,0 +1,1559 @@ +[ + { + "type": "text", + "text": "PEA-Diffusion: Parameter-Efficient Adapter with Knowledge Distillation in non-English Text-to-Image Generation", + "text_level": 1, + "bbox": [ + 220, + 140, + 782, + 209 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Jian Ma, Chen Chen†, Qingsong Xie, and Haonan Lu†", + "bbox": [ + 282, + 234, + 720, + 251 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "OPPO AI Center \n{majian2, chenchen4, xieqingsong1, luhaonan} @oppo.com", + "bbox": [ + 315, + 262, + 687, + 291 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Abstract. Text-to-image diffusion models are well known for their ability to generate realistic images based on textual prompts. However, the existing works have predominantly focused on English, lacking support for non-English text-to-image models. The most commonly used translation methods cannot solve the generation problem related to language culture, while training from scratch on a specific language dataset is prohibitively expensive. In this paper, we are inspired to propose a simple plug-and-play language transfer method based on knowledge distillation. All we need to do is train a lightweight MLP-like parameter-efficient adapter (PEA) with only 6M parameters under teacher knowledge distillation along with a small parallel data corpus. We are surprised to find that freezing the parameters of UNet can still achieve remarkable performance on the language-specific prompt evaluation set, demonstrating that PEA can stimulate the potential generation ability of the original UNet. Additionally, it closely approaches the performance of the English text-to-image model on a general prompt evaluation set. Furthermore, our adapter can be used as a plugin to achieve significant results in downstream tasks in cross-lingual text-to-image generation.", + "bbox": [ + 261, + 321, + 738, + 570 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Keywords: Diffusion models $\\cdot$ Parameter-efficient $\\cdot$ Plug-and-play", + "bbox": [ + 261, + 584, + 710, + 598 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "1 Introduction", + "text_level": 1, + "bbox": [ + 215, + 621, + 375, + 638 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Recent text-to-image (T2I) models, such as DALL-E2 [24], Imagen [28], and Stable Diffusion (SD) [26], have opened up a new era of AI painting due to their ability to generate realistic, photo-level, and creative images. Numerous downstream tasks have further driven the iteration and update of this technology, as well as the activity of the community, including image editing, controllable generation, etc. However, mainstream text-to-image models only support English prompts, which means non-English speakers usually rely on translations to use them. This may cause translation errors and information loss due to culture-specific concepts. For example, one famous Chinese dish, \"Hongshao shizi tou\", may be translated to \"Braised lion head\" in English. Additionally, the introduction of large translation networks leads to high computation consumption.", + "bbox": [ + 212, + 650, + 787, + 816 + ], + "page_idx": 0 + }, + { + "type": "page_footnote", + "text": "†Corresponding authors.", + "bbox": [ + 232, + 824, + 397, + 839 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "To address this issue, recent researchers have created T2I generative models that process native languages directly, bypassing any need for translation. Taiyi-Bilingual [35] incorporates a novel text encoder initialized with Chinese CLIP checkpoints, focusing only on training the text encoder. AltDiffusion [33] introduces a multi-stage training process for multilingual T2I generation, which is relatively costly. Some experts begin training T2I models from scratch using different language data. For example, ERNIE-ViLG 2.0 [6] is trained with native Chinese content and can generate high-quality images that reflect Chinese cultural elements. However, this approach is very expensive and often beyond the reach of many research facilities. Additionally, this method does not foster a community impact, necessitating the retraining of English models for various downstream or commonly used community tasks. GlueGen [22] introduces a GlueNet module to align the input features of the original image generator by adding a new encoder. Despite its low training cost, its effectiveness in generating culturally relevant images in multiple languages and handling broad prompts is comparatively limited.", + "bbox": [ + 212, + 146, + 787, + 388 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "As such, exploring a parameter-efficient plug-and-play method that adapts the publicly available, English-naive, pretrained T2I models to suit non-English and culture-specific concepts is becoming more and more urgent. Consequently, we concentrate on a critical research question (RQ): Is it possible to devise a low-cost method to acquire a non-English T2I model by adapting the existing pretrained T2I models efficiently? To this end, we take SD as the baseline, as it has shown surprisingly good results on T2I tasks. To finetune SD with minimal cost, we propose a novel parameter-efficient adapter (PEA) method with knowledge distillation (KD) to make full use of the pretrained SD model, referred to as PEA-Diffusion. To take advantage of the pre-trained SD model directly, the pre-trained SD is used as a teacher to guide student learning. Specifically, we propose the PEA module to align the representation space of the new condition encoder in the student with that of the teacher's image generator. PEA merely consists of a light-weight multi-layer perception (MLP), greatly simplifying the model construction process and preserving the image generation capabilities of the original SD model. Different from GlueNet [22] which forces text feature alignment of two text encoders at encoder space, we propose KD loss to compel representation alignment at feature maps and logits space from the UNet module between our PEA-Diffusion and pretrained English SD models. We argue that our method boosts the image generator to understand the information from the new text encoder. To avoid catastrophic forgetting and reduce trainable parameters, UNet parameters are kept frozen during the entire training procedure.", + "bbox": [ + 212, + 395, + 787, + 728 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Take Chinese SD as an example. We conduct a thorough experimental analysis on KD to help understand the effect of where to perform feature alignment between the teacher and the student. In addition, we provided intuitive analysis about what parameters should be updated, facilitating discovering the impact of the SD parameters for adapting SD into language-specific T2I models. It should be noted that the method could be applicable for any other non-English T2I generation model. Our core contributions are summarized as follows:", + "bbox": [ + 212, + 734, + 787, + 839 + ], + "page_idx": 1 + }, + { + "type": "page_number", + "text": "2", + "bbox": [ + 217, + 114, + 228, + 126 + ], + "page_idx": 1 + }, + { + "type": "header", + "text": "J. Ma et al.", + "bbox": [ + 271, + 114, + 351, + 127 + ], + "page_idx": 1 + }, + { + "type": "list", + "sub_type": "text", + "list_items": [ + "- We propose a novel PEA-Diffusion model to resolve the T2I model's English-native bias, supporting non-English prompts. The PEA aligns the feature and output spaces between the source and target UNet for bilingual prompts, leading to extremely low training costs.", + "- PEA module is plug-and-play, it can be directly applied to various T2I tasks, including LoRA, ControlNet, Inpainting, compressed SD models that have been pruned and models accelerated through various strategies.", + "- PEA-Diffusion can achieve remarkable image generation performance with culture-specific prompts compared with other cross-lingual or multilingual T2I models. It can also retain the general generation ability of the original English diffusion model with slight degradation on evaluation metrics." + ], + "bbox": [ + 223, + 146, + 785, + 314 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "2 Related Work.", + "text_level": 1, + "bbox": [ + 215, + 337, + 392, + 353 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "2.1 Multilingual Text-to-image Generation", + "text_level": 1, + "bbox": [ + 215, + 371, + 581, + 386 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Diffusion models have shown a remarkable capability of yielding photorealistic and diverse images given textual descriptions, such as GLIDE [20], Stable Diffusion [25], Imagen [28], DALL-E2 [24], and VQ-Diffusion [8]. Despite their powerful generation abilities, they still only understand English input, leading to limitations. To support non-English text prompt, Taiyi [36] finetunes the Chinese text encoder with a diffusion model. ERNIE-ViLG [34] trains the T2I models from scratch. AltDiffusion [33] proposes a multilingual text encoder to capture multiple language features, which serve as DM conditions. The model was trained in three stages: text encoder fine-tuning, $W_{K}, W_{V}$ fine-tuning, and full fine-tuning. GlueGen [22] has designed a GlueNet module that aligns the original conditional input features of the Image Generator by introducing a new encoder. ENSAD [17] proposes a translation-enhanced multilingual T2I generation model by ensembling multilingual T2I and neural machine translation. BDM [18] proposes a backbone-branch network architecture by integrating an English T2I backbone with a Chinese native semantics injection branch. Despite their success in the non-English I2I generation, these works are either constrained to deal with distinctive language concepts or isolated from English-native T2I communities. At the same time, it is not possible to balance the lower cost of training with the performance of evaluation sets. IAP [13] proposes to treat images as pivots to establish the connection between the embedding space of different text encoders, but it doesn't consider to enhance cultural-related generation ability and expand to downstream tasks.", + "bbox": [ + 212, + 397, + 787, + 731 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "2.2 Multilingual CLIP.", + "text_level": 1, + "bbox": [ + 215, + 753, + 418, + 768 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "CLIP uses contrastive language-image pretraining to connect text and image representations, based on massive English text-image sample pairs [23]. Recently, some works have applied teacher learning [2] or metric learning [1] to train multilingual CLIP models, expanding CLIP to other languages. AltCLIP [4] takes this", + "bbox": [ + 212, + 779, + 787, + 840 + ], + "page_idx": 2 + }, + { + "type": "header", + "text": "PEA-diffusion", + "bbox": [ + 635, + 114, + 730, + 127 + ], + "page_idx": 2 + }, + { + "type": "page_number", + "text": "3", + "bbox": [ + 774, + 116, + 785, + 126 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "one step further enhancing the language capabilities, which is implemented by teacher learning and contrastive learning by leveraging a pretrained multilingual text encoder, XLMR [5]. mCLIP [3] achieves multilingual CLIP by aligning the CLIP text model and a multilingual text encoder. To effectively transfer CLIP to language-specific scenarios, some researchers train language-specific CLIPs on large-scale native image-text pairs, e.g., Chinese CLIP [31]. In order to promote PEA-Diffusion to effectively understand language-specific concepts, we append the PEA module after the language-specific CLIP or Multilingual CLIP and train it to uncover non-English cultural concepts within the T2I. The goal is to establish a bridge between the new CLIP and T2I core structures, such as UNet using the PEA module.", + "bbox": [ + 212, + 146, + 787, + 313 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Knowledge Distillation. KD [9] transfers the knowledge in the output logits of a teacher model with a large size to a smaller student model without obvious performance degradation. The feature map in the hidden states and attention output space can also be used for KD to help students mimic teacher capability [11, 14]. Masked Generative Distillation [32] introduces a feature masking scheme into KD to improve students' representation power by forcing them to recover the masked features. Inspired by masked image modeling, the supervised masked KD model is proposed by inserting label information into self-distillation frameworks. However, to the best of our knowledge, KD has not been fully explored for training multilingual T2I generations. In this paper, we apply KD methods to language transfer in T2I and further harness the power of the PEA module through combination to achieve significant improvements.", + "bbox": [ + 212, + 338, + 789, + 521 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "3 Method", + "text_level": 1, + "bbox": [ + 215, + 546, + 330, + 561 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "A naive approach for generating images from non-English text is to substitute the CLIP text encoder in SD with a language-specific CLIP encoder. The text encoder is a crucial part of T2I models since exact and detailed text embedding is necessary for accurate generation. However, altering an existing text model to a different language-specific one is difficult. This difficulty arises because SD's text encoder and image generator are closely integrated, making it ineffective to just swap out the condition encoder. The primary issue is the mismatch between the new text encoder and the old image generator. Furthermore, our experiments show that joint fine-tuning depends heavily on high-quality text-image pairs and often fails due to catastrophic forgetting when updating well-trained parameters. In light of these insights, we keep the pretrained model's parameters fixed and develop a lightweight adapter with KD to synchronize the new text encoder and the image generator.", + "bbox": [ + 212, + 582, + 787, + 777 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Figure 1 illustrates the overall architecture of PEA-Diffusion for cross-lingual transfer in T2I generation. To inherit the knowledge of the pretrained diffusion model, we use the pretrained SD as a teacher to supervise PEA-Diffusion learning through KD. For language-specific T2I generation, we use the corresponding", + "bbox": [ + 212, + 779, + 787, + 840 + ], + "page_idx": 3 + }, + { + "type": "page_number", + "text": "4", + "bbox": [ + 217, + 114, + 228, + 126 + ], + "page_idx": 3 + }, + { + "type": "header", + "text": "J. Ma et al.", + "bbox": [ + 271, + 114, + 351, + 127 + ], + "page_idx": 3 + }, + { + "type": "image", + "img_path": "images/46c10031022c06625b26542269807a2a20bacc6b0c3e8b3b9f10c9a087c28369.jpg", + "image_caption": [ + "Fig. 1: An overview of the proposed PEA-Diffusion. Notice that only the lightweight adapter is trainable through the whole training process." + ], + "image_footnote": [], + "bbox": [ + 272, + 143, + 720, + 316 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "language-specific CLIP as a text encoder to capture text information. The generator in PEA-Diffusion is the same UNet as that used in SD. A light-weight MLP adapter is added on top of the CLIP text encoder in PEA-Diffusion, facilitating knowledge transfer from teacher to student.", + "bbox": [ + 212, + 387, + 782, + 446 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "3.1 Preliminary", + "text_level": 1, + "bbox": [ + 214, + 470, + 359, + 486 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "The diffusion model is composed of a forward diffusion process and a reverse denoising process. The forward process gradually adds random noise to clean data and diffuses it into pure Gaussian noise, while the reverse process reverses the diffusion process to create satisfactory samples from the Gaussian noise [10].", + "bbox": [ + 212, + 497, + 784, + 556 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Specifically, for an input image $x_0 \\in \\mathbb{R}^{H \\times W \\times 3}$ , the encoder $\\mathcal{E}$ of the autoencoder transforms it into a latent representation $z_0 \\in \\mathbb{R}^{h \\times w \\times c}$ , where $f = H / h = W / w$ is the downsampling factor and $c$ is the latent feature dimension. The diffusion process is then performed in the latent space, where a conditional UNet [27] denoiser $\\epsilon_{\\theta}$ is employed to predict noise $\\epsilon$ with current timestep $t$ , noisy latent $z_t$ and generation condition $C$ . In T2I scenarios, the condition $C = \\tau_{\\theta}(y)$ is produced by encoding the text prompts $y$ with a pretrained CLIP [23] text encoder $\\tau_{\\theta}$ . Therefore, the overall training objective of SD is defined as", + "bbox": [ + 212, + 556, + 785, + 679 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\n\\mathcal {L} _ {\\mathrm {S D}} = \\mathbb {E} _ {\\mathcal {E} (x _ {0}), C, \\epsilon \\sim \\mathcal {N} (0, 1), t} \\left[ \\| \\epsilon - \\epsilon_ {\\theta} (z _ {t}, t, C) \\| _ {2} ^ {2} \\right] \\tag {1}\n$$\n", + "text_format": "latex", + "bbox": [ + 339, + 689, + 785, + 718 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "3.2 Cross-Linguual Transfer", + "text_level": 1, + "bbox": [ + 214, + 750, + 449, + 763 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Objective. Given sets of parallel corpus $Y^{T} = \\{y_{i}^{t}\\}_{i = 1}^{N}$ and $Y^{S} = \\{y_{i}^{t}\\}_{i = 1}^{N}$ and $Y^{S^{\\prime}} = \\{y_{i}^{t}\\}_{i = 1}^{M}$ and the corresponding images $I = \\{img\\}_{i = 1}^{N}$ with $N$ denoting the total number of paired samples, $Y^{T}$ represents the English corpus that will be input into the teacher model, $Y^{S}$ represents the non-English corpus with", + "bbox": [ + 212, + 773, + 785, + 840 + ], + "page_idx": 4 + }, + { + "type": "header", + "text": "PEA-diffusion", + "bbox": [ + 635, + 114, + 730, + 126 + ], + "page_idx": 4 + }, + { + "type": "page_number", + "text": "5", + "bbox": [ + 774, + 116, + 784, + 126 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "culturally irrelevant content, and $Y^{S'}$ represents the non-English corpus with cultural relevance, they will be input into the student model. The teacher unet denoiser $\\epsilon^T$ projects pairs $(Y^T, I)$ into source middle feature maps and source logits output. Likewise, the student unet denoiser $\\epsilon^S$ projects pairs $(Y^S, I)$ into target middle feature maps and target logits output. Intuitively, there exists severe distribution mismatch between the source and target output, as well as between $\\epsilon^S$ and $\\epsilon^T$ due to the different languages. Therefore, our target is to make $\\epsilon^S$ match with $\\epsilon^T$ , so as to transfer teacher knowledge to the student, enabling cross-lingual T2I generation. For the student, we use a language-specific CLIP text encoder instead of the original English text encoder to enhance token- and sentence-level linguistic-native representations. For brevity, we omit the subscripts of $\\mathbb{E}_{\\mathcal{E}(x_0), \\tau_\\theta(y), \\epsilon \\sim \\mathcal{N}(0, 1), t}$ and $z_t, t$ conditions in the following notations.", + "bbox": [ + 212, + 145, + 787, + 330 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Knowledge Distillation for T2I generation. The original KD leverages Kullback-Leibler (KL) divergence to measure the difference between the categorical probability distribution of the teacher's logits output and the student's logits output [9]. However, in the task of T2I generation, since the middle feature maps are the latent value of pixels instead of categorical probability distribution, KL divergence is unable to measure the difference between student and teacher. Consequently, we replace KL divergence with the L2-norm distance between the logits output from student and teacher to extend raw KD for T2I generation, whose loss function is formulated as:", + "bbox": [ + 212, + 330, + 787, + 465 + ], + "page_idx": 5 + }, + { + "type": "equation", + "text": "\n$$\n\\mathcal {L} _ {L K D} = \\mathbb {E} \\left[ \\| \\epsilon^ {T} (y ^ {T}) - \\epsilon^ {S} (y ^ {S}) \\| _ {2} ^ {2} \\right] \\tag {2}\n$$\n", + "text_format": "latex", + "bbox": [ + 385, + 472, + 785, + 498 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "where $y^{T} \\in Y^{T}$ and $y^{S} \\in Y^{S}$ . $\\mathcal{L}_{LK D}$ merely enforces student distribution to match teacher distribution at logits output layer. To further reduce distribution shift of teacher and student, we boost feature alignment of the middle feature maps between $\\epsilon^{S}$ and $\\epsilon^{T}$ :", + "bbox": [ + 212, + 506, + 787, + 568 + ], + "page_idx": 5 + }, + { + "type": "equation", + "text": "\n$$\n\\mathcal {L} _ {F L K D} = \\mathbb {E} \\left[ \\sum_ {l} \\| f \\left(g _ {l} ^ {T} \\left(y ^ {T}\\right)\\right) - f \\left(g _ {l} ^ {S} \\left(y ^ {S}\\right)\\right) \\| _ {2} ^ {2} \\right] \\tag {3}\n$$\n", + "text_format": "latex", + "bbox": [ + 344, + 578, + 785, + 609 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "where $f$ represents the norm operation, $g_{l}^{T}$ and $g_{l}^{S}$ represent the feature maps of the $l$ -th layer in a predefined set of unet layers from the teacher and the student, respectively. With respective weights $\\lambda_{FL}$ and $\\lambda_{L}$ , the final KD objective is formalized as below:", + "bbox": [ + 212, + 619, + 787, + 681 + ], + "page_idx": 5 + }, + { + "type": "equation", + "text": "\n$$\n\\mathcal {L} _ {K D} = \\lambda_ {F L} \\mathcal {L} _ {F L K D} + \\lambda_ {L} \\mathcal {L} _ {L K D} \\tag {4}\n$$\n", + "text_format": "latex", + "bbox": [ + 390, + 694, + 785, + 709 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Adapter. Due to the dimension difference between English CLIP and non-English CLIP, we propose an adapter for simple feature transformation to align feature dimension. Moreover, the adapter is trained to learn language-specific information. The adapter is implemented by a MLP with only 6M parameters. To accelerate training and avoid destroying the generalization capability of the pretrained models, we freeze the CLIP encoder and image generator. As a result, the proposed PEA-Diffusion is both parameter- and computation-efficient because just the adapter is trained.", + "bbox": [ + 212, + 719, + 787, + 840 + ], + "page_idx": 5 + }, + { + "type": "page_number", + "text": "6", + "bbox": [ + 217, + 114, + 228, + 126 + ], + "page_idx": 5 + }, + { + "type": "header", + "text": "J. Ma et al.", + "bbox": [ + 271, + 114, + 351, + 127 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/c505fe5df1fc2348473de1214bea9d5e8b1ed6fdc267739f7d31aedb002d3f53.jpg", + "image_caption": [ + "Fig. 2: Image generation visualization of different models in Chinese specific language." + ], + "image_footnote": [], + "bbox": [ + 225, + 145, + 777, + 402 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "3.3 Training Strategy and Final Objective", + "text_level": 1, + "bbox": [ + 215, + 465, + 576, + 479 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "For our goal of training a multi-language T2I generation model, we only need to train on English data since both the new CLIP text encoder and the teacher model's CLIP text encoder can accept English inputs. In this case, we can use aligned KD loss $\\mathcal{L}_{KD}$ . For our goal of training a language-specific T2I generation model, we adopt a hybrid training strategy, for the target language's parallel corpus, we only need to use KD loss $\\mathcal{L}_{KD}$ to align the distribution. On the other hand, we also have language-specific culture-related data $Y^{S^{\\prime}}$ . Because there is a significant translation error when translating these data into English and the teacher model cannot provide high-quality knowledge guidance for this type of data, we only use $\\mathcal{L}_{\\mathrm{SD}}$ (1) for improving culture understanding ability. The final objective is formalized as follows:", + "bbox": [ + 212, + 489, + 787, + 659 + ], + "page_idx": 6 + }, + { + "type": "equation", + "text": "\n$$\n\\mathcal {L} = \\left\\{ \\begin{array}{l l} \\lambda_ {F L} \\mathbb {E} \\Big [ \\| \\epsilon^ {T} (y ^ {T}) - \\epsilon^ {S} (y) \\| _ {2} ^ {2} \\Big ] + \\lambda_ {L} \\mathbb {E} \\Big [ \\sum_ {l} \\| f (g _ {l} ^ {T} (y ^ {T})) - f (g _ {l} ^ {S} (y)) \\| _ {2} ^ {2} \\Big ], & y \\in Y ^ {S} \\\\ \\mathbb {E} \\Big [ \\| \\epsilon - \\epsilon^ {S} (y) \\| _ {2} ^ {2} \\Big ], & y \\in Y ^ {S ^ {\\prime}} \\end{array} \\right.\n$$\n", + "text_format": "latex", + "bbox": [ + 215, + 669, + 795, + 719 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "4 Experiments", + "text_level": 1, + "bbox": [ + 215, + 738, + 375, + 755 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "4.1 Data Preparation", + "text_level": 1, + "bbox": [ + 215, + 768, + 408, + 784 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "We download the Laion2B-multi [29] dataset along with English translations for image descriptions. Based on the open-source CLIP models available for different languages: Chinese CLIP, Russian CLIP, Japanese CLIP, Korean CLIP, and", + "bbox": [ + 212, + 794, + 785, + 840 + ], + "page_idx": 6 + }, + { + "type": "header", + "text": "PEA-diffusion", + "bbox": [ + 635, + 114, + 730, + 126 + ], + "page_idx": 6 + }, + { + "type": "page_number", + "text": "7", + "bbox": [ + 774, + 116, + 785, + 126 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Italian CLIP, we determine to conduct experiments on the above five languages. For Chinese, we also collect WuKong [7] data with Chinese cultural image-text pairs and Chinese-English parallel corpus translations of Laion2B-en. For more data, model, and code details, please refer to the Appendix.", + "bbox": [ + 212, + 146, + 787, + 207 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "4.2 Implementation Details and Evaluation", + "text_level": 1, + "bbox": [ + 215, + 239, + 584, + 255 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Implementation details. We conduct all experiments on Stable Diffusion (SD)- [25], and Stable Diffusion XL (SDXL)- [21] like backbones. Take SDXL base model as an example. PEA-Diffusion consists of VAE, UNet, CLIP text encoder, and the adapter module, comprising 3 billion parameters, out of which a mere 6 million parameters ( $\\approx 0.2\\%$ ) are trainable. We set $\\lambda_{FL}$ 0.1 and $\\lambda_{L}$ 1 respectively based on the loss performance. The entire model is trained on 64 A100 GPUs for 70000 steps with a batch size of 5 per GPU.", + "bbox": [ + 212, + 276, + 787, + 381 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Test benchmark. To evaluate the capability of PEA-Diffusion to generate images and capture culture-specific concepts of different languages, we introduce two datasets: Multilingual-General (MG) for generation quality evaluation and Multilingual-Cultural (MC) for culture-specific concepts evaluation. We collect 200 prompts for MG, which covered content such as people, animals, landscapes, and fantasy themes. The descriptions include diverse styles, such as cartoon, anime, illustration, watercolor, and cyberpunk. The data comes from various major art websites where users provide real prompts. On the other hand, we also construct MC data including 200 prompts. The collection is mainly based on the ChatGPT tool and consists of culturally relevant content for each specific language, including but not limited to local cuisine, architecture, clothing, and distinctive cultural features.", + "bbox": [ + 212, + 385, + 787, + 566 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Evaluation metrics. For the general evaluation data MG, the widely used FID metric can only measure the feature distance between generated images and specific images. It often fails to provide accurate assessment for high-quality generated images, as emphasized and experimentally validated by SDXL [21] and other studies. Therefore, in order to evaluate the quality of generated images more accurately, we introduce two human preference models as our evaluation metrics: ImageReward [30] and PickScore [15]. Additionally, we continue to utilize OpenCLIP ViT-bigG to measure the similarity between images and text. As for the language-specific evaluation data MC, constructed for each language, to ensure a fair comparison of content relevance, we only consider the CLIP model trained in the specific language to calculate the correlation between generated images and their corresponding text. Since the general CLIP model can only process non-English data after translation into English, this process can introduce significant performance degradation. Furthermore, our paper's focus is precisely on addressing the errors caused by translating non-English prompts into English to generate images. Therefore, other metrics that involve translation processes, such as the two models based on human preference mentioned above, will no longer be valid for evaluation.", + "bbox": [ + 212, + 568, + 787, + 839 + ], + "page_idx": 7 + }, + { + "type": "page_number", + "text": "8", + "bbox": [ + 217, + 114, + 228, + 126 + ], + "page_idx": 7 + }, + { + "type": "header", + "text": "J. Ma et al.", + "bbox": [ + 271, + 114, + 351, + 127 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Baseline methods. We generated image results in specific languages based on two T2I base models: SD1.5 and SDXL. We choose the following four categories of approaches as our baselines.", + "bbox": [ + 212, + 146, + 787, + 191 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "text", + "list_items": [ + "- Translation: we apply Google translation API.", + "- AltDiffusion [33]: we test two multilingual SD models based on the open-source AltDiffusion framework: AltDiffusion-m9 and AltDiffusion-18.", + "- GlueGen [22]: We evaluate our model on three languages that overlap with the checkpoint provided by GlueGen, which are Chinese, Italian, and Japanese. For the other two languages, Russian and Korean, we implement GlueGen based on its code and our data. In addition, to further compare the impact of different language-specific CLIPs on the results, we also retrain the GlueGen based on five languages' CLIP text encoders.", + "- Fine-tuning with LoRA [12] and $W_{K}, W_{V}$ (key and value projection matrices): As low-cost experiments for comparison, we also fine-tuning the model to transfer language capabilities." + ], + "bbox": [ + 225, + 200, + 792, + 382 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "4.3 Experimental Results", + "text_level": 1, + "bbox": [ + 215, + 404, + 441, + 420 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "Generation analysis. We validate our approach based on two different T2I base models. Firstly, to ensure a fair comparison with GlueGen and AltDiffusion, we train PEA-Diffusion based on the SD1.5 model. Secondly, to validate our low-cost experiment approach, we train a larger model based on SDXL. The compared methods are LoRA fine-tuning with rank 64 and $W_{K}$ , $W_{V}$ fine-tuning, both of which are parameter-efficient fine-tuning strategies. For the second approach, as proposed in [16], when updating the mapping from given text to image distribution, only updating $W_{K}^{(i)}$ , $W_{V}^{(i)}$ in each cross-attention block $i$ is sufficient since text features are the only input to the key and value projection matrix. Furthermore, to validate the wide application of the proposed approach with different text encoders, we train our model with a multi-language T5 encoder and a multi-language CLIP encoder. This type of training only requires English text data, which makes it even simpler and more efficient. Fig. 2 illustrates the comparisons between translation and our methods with different training strategies. It can be observed that our PEA-Diffusion can achieve Chinese culturally related image generation, but translation totally fails in these cases.", + "bbox": [ + 212, + 429, + 787, + 672 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "As shown in the comprehensive results of Table 1, when compared with the translation method, our PEA-Diffusion outperforms it for almost all languages (except Japanese) in terms of CLIPScore for MC datasets. It is obvious that the translation method cannot capture the essence of different languages' cultures, leading to a semantic gap between the original prompts and the translated English prompts. However, our approach can use additional culturally related image-text pairs to train the adapter, which can align the language understanding ability of the text encoder with the generation ability of the UNet. Even though the UNet is frozen during the whole training process, we argue that our PEA-Diffusion can still stimulate the potential generation ability for cultural-related images of the original UNet. Since SD1.5 and SDXL are all trained on", + "bbox": [ + 212, + 674, + 787, + 840 + ], + "page_idx": 8 + }, + { + "type": "header", + "text": "PEA-diffusion", + "bbox": [ + 635, + 114, + 730, + 127 + ], + "page_idx": 8 + }, + { + "type": "page_number", + "text": "9", + "bbox": [ + 774, + 114, + 785, + 126 + ], + "page_idx": 8 + }, + { + "type": "table", + "img_path": "images/30451c5e392fff7bfcac3cb0cb354f78b2376b3ef1c6ac706b4a285151b4e4c5.jpg", + "table_caption": [ + "Table 1: CLIPScore with different specific languages. For Chinese, we evaluated two general evaluation metrics with MG data. $\\dagger$ Indicates open source models." + ], + "table_footnote": [], + "table_body": "
Model BaseMethodszh-MCzh-MGru-MCit-MCko-MCja-MC
CLIPScore_zhImageRewardPickScoreCLIPScoreRuCLIPScore—itCLIPScore_koCLIPScore_ja
SD1.5Translate0.16860.24270.17990.49110.35560.19040.2885
Altdiffusion m9 [33]†0.2100-0.69660.12020.37150.31770.19060.2461
Altdiffusion m18 [33]†0.21720.33160.16480.47300.35020.19170.3008
GlueGen [22]†0.1700-0.51870.12240.41840.30650.17690.2506
GlueGen [22]0.2153-0.46850.12560.39870.29280.18430.2465
Ours0.25240.09850.17010.50070.37280.18960.2501
SDXLTranslate0.23831.36470.24020.48440.35910.20330.2958
LoRA0.23660.12170.15980.46560.34760.20340.1940
Wkand Wv0.24850.14130.16210.47850.34950.20880.2022
Ours w/ MuTI50.23010.82540.20010.45680.38050.19080.2996
Ours w/ Multilingual-CLIP0.24071.21220.23240.46730.38560.21570.3007
Ours0.26101.02450.22440.50120.41910.21940.2544
", + "bbox": [ + 218, + 185, + 785, + 306 + ], + "page_idx": 9 + }, + { + "type": "text", + "text": "billions of image-text pairs that contain wide image domains, our approach can successfully awaken their generative abilities in different cultural contexts. In terms of ImageReward and PickScore for the MG dataset, we must admit that PEA-Diffusion is slightly inferior to translation methods. This is because using other text encoders (e.g., Chinese CLIP and Multilingual CLIP) cannot ensure complete alignment between the text embedding and the UNet feature spaces.", + "bbox": [ + 212, + 337, + 782, + 428 + ], + "page_idx": 9 + }, + { + "type": "text", + "text": "Even though Altdiffusion retrans the multilingual CLIP text encoder and UNet on billions of image-text pairs, it still performs worse than our approach in terms of almost all metrics for all languages. GlueGen is the most related work on our concerned topic. The statistics presented in Table 1 reveal that both checkpoints provided by the authors or re-implemented by ourselves are significantly inferior to our PEA-Diffusion, especially for the ImageReward and PickScore metrics, resulting in very low generation quality since only aligning distributions of text embedding for different languages is very hard, even with its sophisticatedly designed GlenNet. In addition, it cannot generate culturally related images. Therefore, it's worth mentioning that our core contribution is that we align the model's generation ability through UNet distillation only by a MLP-like adapter between different languages. We don't forcibly align the text embedding spaces of different languages.", + "bbox": [ + 212, + 429, + 785, + 625 + ], + "page_idx": 9 + }, + { + "type": "text", + "text": "Additionally, we undertake various parameter-efficient fine-tuning strategies for comparison. The fine-tuning techniques, including LoRA and $W_{k}$ and $W_{v}$ , show subpar performance based on ImageReward, PickScore, and CLIPScore metrics. This is attributed to the fact that these fine-tuning methods depend significantly on high-quality image-text datasets, whereas our PEA-Diffusion requires merely some parallel multilingual texts and images and lacks stringent quality requirements. Further experiments can be found in the Appendix.", + "bbox": [ + 212, + 627, + 785, + 733 + ], + "page_idx": 9 + }, + { + "type": "text", + "text": "Analysis on different languages. The data in Table 1 indicate that the CLIPScores for Russian, Italian, and Korean either match or exceed those of other baseline methods by a notable margin. For Japanese, however, the CLIP-Score is slightly lower compared to translation and AltDiffusion. We ascribe this to the CLIP model's limited encoding ability for Japanese, which fails to align with the SD model via the adapter layer. Additionally, with only 1.6M training samples for Korean, our method still achieves effective language transfer, under-", + "bbox": [ + 212, + 734, + 785, + 839 + ], + "page_idx": 9 + }, + { + "type": "page_number", + "text": "10", + "bbox": [ + 217, + 114, + 235, + 126 + ], + "page_idx": 9 + }, + { + "type": "header", + "text": "J. Ma et al.", + "bbox": [ + 271, + 114, + 351, + 127 + ], + "page_idx": 9 + }, + { + "type": "text", + "text": "scoring that large-scale training datasets are not essential. Further visualization examples of culturally-related prompts can be found in the Appendix.", + "bbox": [ + 212, + 146, + 782, + 176 + ], + "page_idx": 10 + }, + { + "type": "text", + "text": "Analysis of training cost. Our PEADiffusion freezes the CLIP text encoder, the UNet, and VAE during the whole training process, only the MLP mapping adapter is trained, accounting for merely 6M training parameters, which is $0.2\\%$ of total parameters for SDXL-based models. We employed 64 A100 GPUs to train 70000 steps, resulting in 1600 GPU hours. As a comparison, AltDiffusion [33] trained the model for three stages with nearly 30000 GPU hours, which is $\\approx 20$ times our training cost. In addition, our method achieves higher performance in terms of almost all metrics and the five experimental languages. We also re-implemented and trained GlueGen [22] with the same dataset", + "bbox": [ + 212, + 178, + 537, + 419 + ], + "page_idx": 10 + }, + { + "type": "image", + "img_path": "images/ba27d4e8ef70fe6d6374aa4a2b3472a4c8d8498df69bea3a856dd84986fab0e4.jpg", + "image_caption": [ + "Fig.3: Training parameters(M) and cost comparison for different methods." + ], + "image_footnote": [], + "bbox": [ + 555, + 220, + 782, + 328 + ], + "page_idx": 10 + }, + { + "type": "text", + "text": "constructed by ourselves. It indeed achieves very low training costs with only 1 A100 GPU for 24 hours, but results in very poor performance no matter for image quality (-0.5 ImageReward score) or language transferring ability. Finally, we also conduct other parameter-efficient fine-tuning strategies (e.g., LoRA and tuning only $W_{k}$ and $W_{v}$ matrices of attention layers) and achieve much lower performance and higher training costs, which demonstrated the effectiveness and efficiency of our PEA-Diffusion. Figure 3 only shows the relationship between training parameters and GPU hours of different methods. For more information on model fitting and training data, as well as adapter module parameter variations, please refer to Appendix.", + "bbox": [ + 212, + 419, + 787, + 570 + ], + "page_idx": 10 + }, + { + "type": "text", + "text": "4.4 Ablation Studies", + "text_level": 1, + "bbox": [ + 215, + 597, + 401, + 611 + ], + "page_idx": 10 + }, + { + "type": "text", + "text": "Since Chinese data has certain advantages in terms of data quality and quantity, to eliminate errors caused by data, our ablation experiments are all based on Chinese data. The experiments include four parts: First, we conduct a baseline experiment where we only perform standard model fine-tuning without incorporating any KD strategies. Secondly, inspired by GlueGen [22], we employ text encoder KD to explore the impact of distillation at different positions in the SD framework. Next, we perform distillation based on the UNet model, including feature distillation between different UNet layers and logits distillation for the final output. Finally, we combine all of the distillation positions described above.", + "bbox": [ + 212, + 627, + 787, + 763 + ], + "page_idx": 10 + }, + { + "type": "text", + "text": "Effectiveness of KD. First, KD is crucial for guiding models onto the right learning trajectory for rapid convergence and is vital for the PEA-Diffusion method. As indicated in the first row of Table 2, the standalone fine-tuning method yields notably inferior performance in both image quality (ImageReward and PickScore) and image-text alignment (CLIPScores).", + "bbox": [ + 212, + 763, + 787, + 839 + ], + "page_idx": 10 + }, + { + "type": "header", + "text": "PEA-diffusion", + "bbox": [ + 635, + 114, + 730, + 126 + ], + "page_idx": 10 + }, + { + "type": "page_number", + "text": "11", + "bbox": [ + 767, + 114, + 782, + 126 + ], + "page_idx": 10 + }, + { + "type": "table", + "img_path": "images/07a47706fcd4e234d68d8f2d00b3acf594f20160f8561052edeb130a8bfff626.jpg", + "table_caption": [ + "Table 2: Ablation results for different knowledge strategies." + ], + "table_footnote": [], + "table_body": "
EvaluationImageRewardPickScoreCLIPScore_enCLIPScore_zh
Fine-tuning-1.31540.18540.32780.1901
Fine-tuning w/ TKD-0.98750.19970.35060.2287
Fine-tuning w/ ULKD-1.04470.19570.33310.1968
Fine-tuning w/ UFKD0.95560.21740.40850.2557
Fine-tuning w/ UTKD0.84900.21130.40110.2461
Ours(Fine-tuning w/ UKD)1.02450.22440.42240.2610
", + "bbox": [ + 277, + 172, + 728, + 260 + ], + "page_idx": 11 + }, + { + "type": "text", + "text": "Effectiveness of UNet Feature KD (UFKD). Our KD method consists of two parts: the first part involves distilling the logits results generated by the UNet (ULKD), we then introduce a distillation of intermediate layer features in the UNet, which serves as the primary knowledge guidance mechanism. As shown in the third and fourth lines of Table 2, we can observe that logit distillation can only improve marginally when compared with direct fine-tuning whereas feature distillation boosts the performance remarkably. We attribute this phenomenon to the generation ability of the large amount of UNet parameters, so that intermediate feature supervision is essential for knowledge transfer.", + "bbox": [ + 212, + 285, + 785, + 420 + ], + "page_idx": 11 + }, + { + "type": "text", + "text": "Effectiveness of UNet KD. Furthermore, different distillation positions have different effects. Our initial focus is on aligning the feature spaces of different text encoders, so a natural idea is to distill only the feature information of the teacher model's text encoder (TKD). However, as the results show in the second line of Table 2, although the overall performance exceeds the pure fine-tuning method, it is still far from our best performing approach. This illustrates that simply aligning the text encoder is insufficient and ineffective. This statistic also matches the experimental results of GlueGen in Table 1: simply aligning text embedding spaces cannot obtain satisfactory image generation results. Therefore, this further proves the necessity of distilling the features of the UNet itself.", + "bbox": [ + 212, + 421, + 785, + 571 + ], + "page_idx": 11 + }, + { + "type": "text", + "text": "Finally, we perform KD on both the UNet and text encoder (UTKD). However, the fifth line of Table 2 shows that its effect is worse than performing only UNet distillation. We suspect that the addition of text encoder KD actually weakens the learning of the SD core UNet structure. The main knowledge ability for generating images is mainly influenced by the UNet model, which is responsible for the adapter alignment feature in PEA-Diffusion. More ablation experiments can refer to Appendix.", + "bbox": [ + 212, + 571, + 785, + 678 + ], + "page_idx": 11 + }, + { + "type": "text", + "text": "4.5 Exploring the Domain Adaptation of PEA-Diffusion", + "text_level": 1, + "bbox": [ + 214, + 696, + 689, + 713 + ], + "page_idx": 11 + }, + { + "type": "text", + "text": "To demonstrate the domain adaptation of our model with minimal data guidance, we first constructed a dataset of 20,000 parallel image and text pairs in the domain of human. We then trained the models based on the models fine-tuned $W_{K}^{(i)}, W_{V}^{(i)}$ mentioned in section 4.3 with UKD, and evaluated on the MG and MC dataset. The results on Chinese are shown Table 3.", + "bbox": [ + 212, + 719, + 785, + 792 + ], + "page_idx": 11 + }, + { + "type": "text", + "text": "We observed significant improvements in the three metrics on the MG test data. This suggests that PEA can be guided by a small amount of data to perform out-of-domain knowledge transfer.", + "bbox": [ + 212, + 794, + 785, + 839 + ], + "page_idx": 11 + }, + { + "type": "page_number", + "text": "12", + "bbox": [ + 217, + 114, + 235, + 126 + ], + "page_idx": 11 + }, + { + "type": "header", + "text": "J. Ma et al.", + "bbox": [ + 271, + 114, + 351, + 127 + ], + "page_idx": 11 + }, + { + "type": "table", + "img_path": "images/c21e8083b1221e68b7df7a87d7f6fb2be7f4520bcc2fec796650d4451efca06d.jpg", + "table_caption": [ + "Table 3: Domain adaptation with 20,000 data in Chinese specific language." + ], + "table_footnote": [], + "table_body": "
EvaluationImageRewardPickScoreCLIPScore_enCLIPScore_zh
Wk and Wv0.14130.16210.36740.2485
Ours(Fine-tuning w/ UKD)0.83680.21450.40980.2501
", + "bbox": [ + 276, + 172, + 730, + 220 + ], + "page_idx": 12 + }, + { + "type": "text", + "text": "4.6 Adaptability Experiment of English Downstream Model", + "text_level": 1, + "bbox": [ + 215, + 247, + 723, + 263 + ], + "page_idx": 12 + }, + { + "type": "image", + "img_path": "images/653bc4464f408a5eae3cef29b77f50385c27afa055311bbcb173256cb65d62d3.jpg", + "image_caption": [ + "sal_scenery LoRA \n西湖，塔和瀑布，日出(West Lake, snow and waterfall, sunrise) \n江南，夏日湖畔的一村村庄(Jianxun, a village by the summer lake)" + ], + "image_footnote": [], + "bbox": [ + 217, + 299, + 501, + 356 + ], + "page_idx": 12 + }, + { + "type": "image", + "img_path": "images/e70f2a98840d160713a31812f6776e65287d45cc5c09fab109982deee3c7df53.jpg", + "image_caption": [ + "realistic/easy/animalism/3d-carton Fine-tuning \n一个漂亮的双面的调色板 \n一个白色的小球与white painting light) \n3D电影，4K，高度细致，男人坐在马桶上读报（3D movie, 4K, highly detailed, man sitting on toilet" + ], + "image_footnote": [], + "bbox": [ + 503, + 299, + 784, + 356 + ], + "page_idx": 12 + }, + { + "type": "image", + "img_path": "images/a8089f7a1a6d5ab98bdc9585dbecd91621a231660164c9a4e348c3f4485139f5.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 217, + 376, + 500, + 433 + ], + "page_idx": 12 + }, + { + "type": "image", + "img_path": "images/92eafbe56876f3ab5d6e6e24da6cbffc31a80436d09be4bf5ec0c423707e123e.jpg", + "image_caption": [ + "SD-XL, Inpainting 0.1 Painting \n2018年，由青岛双星玻璃有限公司（White porcelain plates on benches）一辆小型起重机放在了长椅上（A small crane was placed on the bench）" + ], + "image_footnote": [], + "bbox": [ + 500, + 377, + 784, + 433 + ], + "page_idx": 12 + }, + { + "type": "image", + "img_path": "images/b908fc72ce8f83e42fb39bf7c840e59a686cc5293d871901e0a81208305dabf5.jpg", + "image_caption": [ + "controlnet-canny ControlNet 鸳鸯双栖展双飞，满园存色惹人醉(Mandarin ducks and butterflies fly together, and the spring scenery in the garden is intoxicating); 北国风光，千里冰封，万里雪飘(Northern scenery, thousands of miles of ice, thousands of miles of snow drifting)", + "Icm-lora-adsl CDM 带着墨镜的猫咪手里拿着剑，在恶魔城怪里，仙鸽奇侠风格(The cat wearing sunglasses holds a sword in its hand, in the Demon Castle, and the Magic Knight) 红色灯笼，长城，秋天，高清，奇幻风格，史密德(Red Lutron, Great Wall, Autumn, HD, Fantasy Style, Epic Sense)", + "Fig. 4: Downstream task experiments of our proposed PEA. By only adding the trained adapter between the Chinese CLIP text encoder and the UNet, it can achieve language transfer for fine-tuning, LoRA, ControlNet, Inpainting, LCM and compressed models." + ], + "image_footnote": [], + "bbox": [ + 217, + 457, + 500, + 513 + ], + "page_idx": 12 + }, + { + "type": "image", + "img_path": "images/d901f8901fab1ae63dab5501bb89c8203b5b262fded9199baef354028fd4479f.jpg", + "image_caption": [ + "SSD-B I SSD \n一只超级可爱兔子穿着僧侣服装，肖像猫，皮克斯动画(A super cute rabbit wearing monk clothing, partition)。 \n中国汉奴女孩，阳光明媚，站在踏青花中；近景，摄影风景（Chinese Hanafi girl, sunny, standing in the midst of ruffled jeans, close-up, photography style）" + ], + "image_footnote": [], + "bbox": [ + 500, + 455, + 782, + 512 + ], + "page_idx": 12 + }, + { + "type": "text", + "text": "In this section, we would like to emphasize the plug-and-play feature of our PEA module. Currently, a wide range of community resources, such as fine-tuned models, LoRA, Dreambooth, ControlNet, compressed SD models, and accelerated sampling SD models, are derived from the English SD models, so that the English community is much more mature than other languages. Therefore, one key challenge we must overcome is bridging the compatibility gap between the new model and the English SD community after language transfer. Since our approach uses feature-level and logit-level KD to train only the PEA module while keeping the core UNet parameters unchanged, it naturally maintains compatibility with the English SD community and can seamlessly integrate with various English SD community plugins. Interestingly, the PEA module remains plug-and-play for models that have undergone changes in base SD UNet parameters in contrast to PEA-Diffusion but whose overall structure remains similar. This includes models such as fine-tuned SD and compressed SD trained using", + "bbox": [ + 212, + 628, + 787, + 840 + ], + "page_idx": 12 + }, + { + "type": "header", + "text": "PEA-diffusion", + "bbox": [ + 635, + 114, + 730, + 126 + ], + "page_idx": 12 + }, + { + "type": "page_number", + "text": "13", + "bbox": [ + 767, + 114, + 785, + 126 + ], + "page_idx": 12 + }, + { + "type": "text", + "text": "pruning and distillation techniques, among others. In the following, we will introduce several downstream image generation applications based on the SD or SDXL foundation model.", + "bbox": [ + 212, + 146, + 784, + 190 + ], + "page_idx": 13 + }, + { + "type": "text", + "text": "Fine-tuned checkpoint. We use two fine-tuned SDXL-based model checkpoints: \"xxmix9realistic\" and \"samaritan-3d-cartoon\". The \"xxmix9realistic\" model is trained on person, photorealistic, sexy, female, woman, girls, etc. The \"samaritan-3d-cartoon\" model is trained on anime, Disney, 3D, Pixar, etc.", + "bbox": [ + 212, + 191, + 795, + 251 + ], + "page_idx": 13 + }, + { + "type": "text", + "text": "LoRA. \"csal_scenery\" is a SDXL-based LoRA model that contains common ancient Chinese architectural forms such as pavilions, pagodas, towers, gardens, courtyards, bridges, temples, palaces, towns, etc.", + "bbox": [ + 212, + 252, + 784, + 296 + ], + "page_idx": 13 + }, + { + "type": "text", + "text": "ControlNet. \"controlnet-canny\" is trained on SDXL under canny condition.", + "bbox": [ + 238, + 297, + 784, + 311 + ], + "page_idx": 13 + }, + { + "type": "text", + "text": "Inpainting. SDXL Inpainting 0.1 is a latent T2I diffusion model capable of generating photorealistic images given any text input, with the extra capability of Inpainting the pictures by using a mask.", + "bbox": [ + 212, + 313, + 784, + 357 + ], + "page_idx": 13 + }, + { + "type": "text", + "text": "LCM. Latent Consistency Models [19] have achieved impressive performance in accelerating T2I generative tasks, producing high-quality images with minimal inference steps. \"lcm-lora-sdxl\" is an acceleration module based on SDXL.", + "bbox": [ + 212, + 358, + 784, + 402 + ], + "page_idx": 13 + }, + { + "type": "text", + "text": "SSD. The Segmind Stable Diffusion Model is a condensed version of the SDXL that is $50\\%$ smaller in size, delivering a $60\\%$ increase in speed while preserving high-quality text-to-image (T2I) generation capabilities.", + "bbox": [ + 212, + 402, + 784, + 448 + ], + "page_idx": 13 + }, + { + "type": "text", + "text": "In Figure 4, we have shown that the PEA module can be used as a plugin between the Chinese CLIP text encoder and the UNet, achieving the ability to transfer language from English to Chinese directly. For more details of the abovementioned checkpoints and generated images, please refer to our Appendix.", + "bbox": [ + 212, + 449, + 784, + 508 + ], + "page_idx": 13 + }, + { + "type": "text", + "text": "5 Conclusion and Limitation", + "text_level": 1, + "bbox": [ + 214, + 532, + 509, + 547 + ], + "page_idx": 13 + }, + { + "type": "text", + "text": "Conclusion. The high training expenses and limited training data available for multilingual T2I models have driven us to search for a more lightweight solution. In this paper, we introduce a straightforward and efficient method using KD and adapters. By focusing solely on training the adapter and keeping other parameters fixed, our approach harnesses the capability of T2I models for multilingual cultural comprehension and generation. Our method not only achieves significantly better results for culturally relevant data but also surpasses existing models while maintaining general synthesis capability. Comprehensive experiments demonstrate that our method is cost-effective, versatile, and easy to integrate.", + "bbox": [ + 212, + 563, + 785, + 714 + ], + "page_idx": 13 + }, + { + "type": "text", + "text": "**Limitation.** First, the effectiveness of the CLIP text encoder significantly influences both the quality of image generation and text relevance. Utilizing a weaker language-specific CLIP could impact the performance of PEA-Diffusion in broad open domains and may slightly compromise the model's generalization abilities. Second, the upper limit of our method is constrained by the capabilities of the English base model, indicating that our approach cannot exceed its inherent potential. Moving forward, we aim to conduct additional research to mitigate these limitations.", + "bbox": [ + 212, + 715, + 785, + 835 + ], + "page_idx": 13 + }, + { + "type": "page_number", + "text": "14", + "bbox": [ + 217, + 114, + 235, + 126 + ], + "page_idx": 13 + }, + { + "type": "header", + "text": "J. Ma et al.", + "bbox": [ + 271, + 114, + 351, + 127 + ], + "page_idx": 13 + }, + { + "type": "text", + "text": "References", + "text_level": 1, + "bbox": [ + 217, + 143, + 321, + 159 + ], + "page_idx": 14 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "1. Aggarwal, P., Kale, A.: Towards zero-shot cross-lingual image retrieval. arXiv preprint arXiv:2012.05107 (2020)", + "2. Carlsson, F., Eisen, P., Rekathati, F., Sahlgren, M.: Cross-lingual and multilingual clip. In: Proceedings of the Thirteenth Language Resources and Evaluation Conference. pp. 6848-6854 (2022)", + "3. Chen, G., Hou, L., Chen, Y., Dai, W., Shang, L., Jiang, X., Liu, Q., Pan, J., Wang, W.: mclip: Multilingual clip via cross-lingual transfer. In: Proceedings of the 61st Annual Meeting of the Association for Computational Linguistics (Volume 1: Long Papers). pp. 13028-13043 (2023)", + "4. Chen, Z., Liu, G., Zhang, B.W., Ye, F., Yang, Q., Wu, L.: Altclip: Altering the language encoder in clip for extended language capabilities (2022)", + "5. Conneau, A., Khandelwal, K., Goyal, N., Chaudhary, V., Wenzek, G., Guzmán, F., Grave, E., Ott, M., Zettlemoyer, L., Stoyanov, V.: Unsupervised cross-lingual representation learning at scale. arXiv preprint arXiv:1911.02116 (2019)", + "6. Ding, M., Yang, Z., Hong, W., Zheng, W., Zhou, C., Yin, D., Lin, J., Zou, X., Shao, Z., Yang, H., et al.: Cogview: Mastering text-to-image generation via transformers. Advances in Neural Information Processing Systems 34, 19822-19835 (2021)", + "7. Gu, J., Meng, X., Lu, G., Hou, L., Minzhe, N., Liang, X., Yao, L., Huang, R., Zhang, W., Jiang, X., et al.: Wukong: A 100 million large-scale chinese cross-modal pre-training benchmark. Advances in Neural Information Processing Systems 35, 26418-26431 (2022)", + "8. Gu, S., Chen, D., Bao, J., Wen, F., Zhang, B., Chen, D., Yuan, L., Guo, B.: Vector quantized diffusion model for text-to-image synthesis. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 10696-10706 (2022)", + "9. Hinton, G., Vinyals, O., Dean, J.: Distilling the knowledge in a neural network. arXiv preprint arXiv:1503.02531 (2015)", + "10. Ho, J., Jain, A., Abbeel, P.: Denoising diffusion probabilistic models. Advances in Neural Information Processing systems 33, 6840-6851 (2020)", + "1. Hou, L., Huang, Z., Shang, L., Jiang, X., Chen, X., Liu, Q.: Dynabert: Dynamic bert with adaptive width and depth. Advances in Neural Information Processing Systems 33, 9782-9793 (2020)", + "2. Hu, E.J., Shen, Y., Wallis, P., Allen-Zhu, Z., Li, Y., Wang, S., Wang, L., Chen, W.: Lora: Low-rank adaptation of large language models (2021)", + "3. Hu, J., Han, X., Yi, X., Chen, Y., Li, W., Liu, Z., Sun, M.: Efficient crosslingual transfer for chinese stable diffusion with images as pivots. arXiv preprint arXiv:2305.11540 (2023)", + "4. Jiao, X., Yin, Y., Shang, L., Jiang, X., Chen, X., Li, L., Wang, F., Liu, Q.: Tinybert: Distilling bert for natural language understanding. arXiv preprint arXiv:1909.10351 (2019)", + "5. Kirstain, Y., Polyak, A., Singer, U., Matiana, S., Penna, J., Levy, O.: Pick-a-pic: An open dataset of user preferences for text-to-image generation (2023)", + "6. Kumari, N., Zhang, B., Zhang, R., Shechtman, E., Zhu, J.Y.: Multi-concept customization of text-to-image diffusion. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 1931-1941 (2023)", + "7. Li, Y., Chang, C.Y., Rawls, S., Vulic, I., Korhonen, A.: Translation-enhanced multilingual text-to-image generation. arXiv preprint arXiv:2305.19216 (2023)" + ], + "bbox": [ + 225, + 178, + 784, + 839 + ], + "page_idx": 14 + }, + { + "type": "header", + "text": "PEA-diffusion", + "bbox": [ + 635, + 114, + 730, + 126 + ], + "page_idx": 14 + }, + { + "type": "page_number", + "text": "15", + "bbox": [ + 767, + 116, + 785, + 126 + ], + "page_idx": 14 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "18. Liu, S., Leng, D., Yin, Y.: Bridge diffusion model: bridge non-english language-native text-to-image diffusion model with english communities (2023)", + "19. Luo, S., Tan, Y., Huang, L., Li, J., Zhao, H.: Latent consistency models: Synthesizing high-resolution images with few-step inference (2023)", + "20. Nichol, A., Dhariwal, P., Ramesh, A., Shyam, P., Mishkin, P., McGrew, B., Sutskever, I., Chen, M.: Glide: Towards photorealistic image generation and editing with text-guided diffusion models. arXiv preprint arXiv:2112.10741 (2021)", + "21. Podell, D., English, Z., Lacey, K., Blattmann, A., Dockhorn, T., Müller, J., Penna, J., Rombach, R.: Sdxl: Improving latent diffusion models for high-resolution image synthesis (2023)", + "22. Qin, C., Yu, N., Xing, C., Zhang, S., Chen, Z., Ermon, S., Fu, Y., Xiong, C., Xu, R.: Gluegen: Plug and play multi-modal encoders for x-to-image generation. arXiv preprint arXiv:2303.10056 (2023)", + "23. Radford, A., Kim, J.W., Hallacy, C., Ramesh, A., Goh, G., Agarwal, S., Sastry, G., Askell, A., Mishkin, P., Clark, J., et al.: Learning transferable visual models from natural language supervision. In: International conference on machine learning. pp. 8748-8763. PMLR (2021)", + "24. Ramesh, A., Dhariwal, P., Nichol, A., Chu, C., Chen, M.: Hierarchical text-conditional image generation with clip latents (2022)", + "25. Rombach, R., Blattmann, A., Lorenz, D., Esser, P., Ommer, B.: High-resolution image synthesis with latent diffusion models. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 10684-10695 (2022)", + "26. Rombach, R., Blattmann, A., Lorenz, D., Esser, P., Ommer, B.: High-resolution image synthesis with latent diffusion models (2022)", + "27. Ronneberger, O., Fischer, P., Brox, T.: U-net: Convolutional networks for biomedical image segmentation. In: Medical Image Computing and Computer-Assisted Intervention-MICCAI 2015: 18th International Conference, Munich, Germany, October 5-9, 2015, Proceedings, Part III 18. pp. 234-241. Springer (2015)", + "28. Saharia, C., Chan, W., Saxena, S., Li, L., Whang, J., Denton, E., Ghasemipour, S.K.S., Ayan, B.K., Mahdavi, S.S., Lopes, R.G., Salimans, T., Ho, J., Fleet, D.J., Norouzi, M.: Photorealistic text-to-image diffusion models with deep language understanding (2022)", + "29. Schuhmann, C., Beaumont, R., Vencu, R., Gordon, C., Wightman, R., Cherti, M., Coombes, T., Katta, A., Mullis, C., Wortsman, M., Schramowski, P., Kundurthy, S., Crowson, K., Schmidt, L., Kaczmarczyk, R., Jitsev, J.: Laion-5b: An open large-scale dataset for training next generation image-text models (2022)", + "30. Xu, J., Liu, X., Wu, Y., Tong, Y., Li, Q., Ding, M., Tang, J., Dong, Y.: Imagereward: Learning and evaluating human preferences for text-to-image generation (2023)", + "31. Yang, A., Pan, J., Lin, J., Men, R., Zhang, Y., Zhou, J., Zhou, C.: Chinese clip: Contrastive vision-language pretraining in chinese. arXiv preprint arXiv:2211.01335 (2022)", + "32. Yang, Z., Li, Z., Shao, M., Shi, D., Yuan, Z., Yuan, C.: Masked generative distillation. In: European Conference on Computer Vision. pp. 53-69. Springer (2022)", + "33. Ye, F., Liu, G., Wu, X., Wu, L.: Altdiffusion: A multilingual text-to-image diffusion model (2023)", + "34. Zhang, H., Yin, W., Fang, Y., Li, L., Duan, B., Wu, Z., Sun, Y., Tian, H., Wu, H., Wang, H.: Ernie-vilg: Unified generative pre-training for bidirectional vision-language generation. arXiv preprint arXiv:2112.15283 (2021)" + ], + "bbox": [ + 215, + 147, + 784, + 839 + ], + "page_idx": 15 + }, + { + "type": "page_number", + "text": "16", + "bbox": [ + 217, + 114, + 235, + 126 + ], + "page_idx": 15 + }, + { + "type": "header", + "text": "J. Ma et al.", + "bbox": [ + 271, + 114, + 351, + 127 + ], + "page_idx": 15 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "35. Zhang, J., Gan, R., Wang, J., Zhang, Y., Zhang, L., Yang, P., Gao, X., Wu, Z., Dong, X., He, J., Zhuo, J., Yang, Q., Huang, Y., Li, X., Wu, Y., Lu, J., Zhu, X., Chen, W., Han, T., Pan, K., Wang, R., Wang, H., Wu, X., Zeng, Z., Chen, C.: Fengshenbang 1.0: Being the foundation of chinese cognitive intelligence. CoRR abs/2209.02970 (2022)", + "36. Zhang, J., Gan, R., Wang, J., Zhang, Y., Zhang, L., Yang, P., Gao, X., Wu, Z., Dong, X., He, J., et al.: Fengshenbang 1.0: Being the foundation of chinese cognitive intelligence. arXiv preprint arXiv:2209.02970 (2022)" + ], + "bbox": [ + 215, + 146, + 787, + 258 + ], + "page_idx": 16 + }, + { + "type": "header", + "text": "PEA-diffusion", + "bbox": [ + 635, + 114, + 730, + 126 + ], + "page_idx": 16 + }, + { + "type": "page_number", + "text": "17", + "bbox": [ + 767, + 116, + 785, + 126 + ], + "page_idx": 16 + } +] \ No newline at end of file diff --git a/2024/PEA-Diffusion_ Parameter-Efficient Adapter with Knowledge Distillation in non-English Text-to-Image Generation/e154c47c-6f75-47c7-a454-a9dcc00f12d4_model.json b/2024/PEA-Diffusion_ Parameter-Efficient Adapter with Knowledge Distillation in non-English Text-to-Image Generation/e154c47c-6f75-47c7-a454-a9dcc00f12d4_model.json new file mode 100644 index 0000000000000000000000000000000000000000..6480d90a85487c6eb8dd51e07290febc2affda93 --- /dev/null +++ b/2024/PEA-Diffusion_ Parameter-Efficient Adapter with Knowledge Distillation in non-English Text-to-Image Generation/e154c47c-6f75-47c7-a454-a9dcc00f12d4_model.json @@ -0,0 +1,2082 @@ +[ + [ + { + "type": "title", + "bbox": [ + 0.221, + 0.141, + 0.783, + 0.21 + ], + "angle": 0, + "content": "PEA-Diffusion: Parameter-Efficient Adapter with Knowledge Distillation in non-English Text-to-Image Generation" + }, + { + "type": "text", + "bbox": [ + 0.283, + 0.235, + 0.721, + 0.252 + ], + "angle": 0, + "content": "Jian Ma, Chen Chen†, Qingsong Xie, and Haonan Lu†" + }, + { + "type": "text", + "bbox": [ + 0.316, + 0.263, + 0.688, + 0.292 + ], + "angle": 0, + "content": "OPPO AI Center \n{majian2, chenchen4, xieqingsong1, luhaonan} @oppo.com" + }, + { + "type": "text", + "bbox": [ + 0.263, + 0.322, + 0.74, + 0.571 + ], + "angle": 0, + "content": "Abstract. Text-to-image diffusion models are well known for their ability to generate realistic images based on textual prompts. However, the existing works have predominantly focused on English, lacking support for non-English text-to-image models. The most commonly used translation methods cannot solve the generation problem related to language culture, while training from scratch on a specific language dataset is prohibitively expensive. In this paper, we are inspired to propose a simple plug-and-play language transfer method based on knowledge distillation. All we need to do is train a lightweight MLP-like parameter-efficient adapter (PEA) with only 6M parameters under teacher knowledge distillation along with a small parallel data corpus. We are surprised to find that freezing the parameters of UNet can still achieve remarkable performance on the language-specific prompt evaluation set, demonstrating that PEA can stimulate the potential generation ability of the original UNet. Additionally, it closely approaches the performance of the English text-to-image model on a general prompt evaluation set. Furthermore, our adapter can be used as a plugin to achieve significant results in downstream tasks in cross-lingual text-to-image generation." + }, + { + "type": "text", + "bbox": [ + 0.263, + 0.585, + 0.711, + 0.599 + ], + "angle": 0, + "content": "Keywords: Diffusion models \\(\\cdot\\) Parameter-efficient \\(\\cdot\\) Plug-and-play" + }, + { + "type": "title", + "bbox": [ + 0.217, + 0.622, + 0.377, + 0.639 + ], + "angle": 0, + "content": "1 Introduction" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.651, + 0.788, + 0.818 + ], + "angle": 0, + "content": "Recent text-to-image (T2I) models, such as DALL-E2 [24], Imagen [28], and Stable Diffusion (SD) [26], have opened up a new era of AI painting due to their ability to generate realistic, photo-level, and creative images. Numerous downstream tasks have further driven the iteration and update of this technology, as well as the activity of the community, including image editing, controllable generation, etc. However, mainstream text-to-image models only support English prompts, which means non-English speakers usually rely on translations to use them. This may cause translation errors and information loss due to culture-specific concepts. For example, one famous Chinese dish, \"Hongshao shizi tou\", may be translated to \"Braised lion head\" in English. Additionally, the introduction of large translation networks leads to high computation consumption." + }, + { + "type": "page_footnote", + "bbox": [ + 0.233, + 0.825, + 0.398, + 0.84 + ], + "angle": 0, + "content": "†Corresponding authors." + } + ], + [ + { + "type": "page_number", + "bbox": [ + 0.218, + 0.116, + 0.23, + 0.127 + ], + "angle": 0, + "content": "2" + }, + { + "type": "header", + "bbox": [ + 0.272, + 0.115, + 0.352, + 0.128 + ], + "angle": 0, + "content": "J. Ma et al." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.147, + 0.788, + 0.389 + ], + "angle": 0, + "content": "To address this issue, recent researchers have created T2I generative models that process native languages directly, bypassing any need for translation. Taiyi-Bilingual [35] incorporates a novel text encoder initialized with Chinese CLIP checkpoints, focusing only on training the text encoder. AltDiffusion [33] introduces a multi-stage training process for multilingual T2I generation, which is relatively costly. Some experts begin training T2I models from scratch using different language data. For example, ERNIE-ViLG 2.0 [6] is trained with native Chinese content and can generate high-quality images that reflect Chinese cultural elements. However, this approach is very expensive and often beyond the reach of many research facilities. Additionally, this method does not foster a community impact, necessitating the retraining of English models for various downstream or commonly used community tasks. GlueGen [22] introduces a GlueNet module to align the input features of the original image generator by adding a new encoder. Despite its low training cost, its effectiveness in generating culturally relevant images in multiple languages and handling broad prompts is comparatively limited." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.396, + 0.788, + 0.729 + ], + "angle": 0, + "content": "As such, exploring a parameter-efficient plug-and-play method that adapts the publicly available, English-naive, pretrained T2I models to suit non-English and culture-specific concepts is becoming more and more urgent. Consequently, we concentrate on a critical research question (RQ): Is it possible to devise a low-cost method to acquire a non-English T2I model by adapting the existing pretrained T2I models efficiently? To this end, we take SD as the baseline, as it has shown surprisingly good results on T2I tasks. To finetune SD with minimal cost, we propose a novel parameter-efficient adapter (PEA) method with knowledge distillation (KD) to make full use of the pretrained SD model, referred to as PEA-Diffusion. To take advantage of the pre-trained SD model directly, the pre-trained SD is used as a teacher to guide student learning. Specifically, we propose the PEA module to align the representation space of the new condition encoder in the student with that of the teacher's image generator. PEA merely consists of a light-weight multi-layer perception (MLP), greatly simplifying the model construction process and preserving the image generation capabilities of the original SD model. Different from GlueNet [22] which forces text feature alignment of two text encoders at encoder space, we propose KD loss to compel representation alignment at feature maps and logits space from the UNet module between our PEA-Diffusion and pretrained English SD models. We argue that our method boosts the image generator to understand the information from the new text encoder. To avoid catastrophic forgetting and reduce trainable parameters, UNet parameters are kept frozen during the entire training procedure." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.735, + 0.788, + 0.84 + ], + "angle": 0, + "content": "Take Chinese SD as an example. We conduct a thorough experimental analysis on KD to help understand the effect of where to perform feature alignment between the teacher and the student. In addition, we provided intuitive analysis about what parameters should be updated, facilitating discovering the impact of the SD parameters for adapting SD into language-specific T2I models. It should be noted that the method could be applicable for any other non-English T2I generation model. Our core contributions are summarized as follows:" + } + ], + [ + { + "type": "header", + "bbox": [ + 0.636, + 0.115, + 0.732, + 0.128 + ], + "angle": 0, + "content": "PEA-diffusion" + }, + { + "type": "page_number", + "bbox": [ + 0.775, + 0.117, + 0.787, + 0.127 + ], + "angle": 0, + "content": "3" + }, + { + "type": "text", + "bbox": [ + 0.225, + 0.147, + 0.786, + 0.207 + ], + "angle": 0, + "content": "- We propose a novel PEA-Diffusion model to resolve the T2I model's English-native bias, supporting non-English prompts. The PEA aligns the feature and output spaces between the source and target UNet for bilingual prompts, leading to extremely low training costs." + }, + { + "type": "text", + "bbox": [ + 0.226, + 0.208, + 0.786, + 0.254 + ], + "angle": 0, + "content": "- PEA module is plug-and-play, it can be directly applied to various T2I tasks, including LoRA, ControlNet, Inpainting, compressed SD models that have been pruned and models accelerated through various strategies." + }, + { + "type": "text", + "bbox": [ + 0.226, + 0.254, + 0.786, + 0.315 + ], + "angle": 0, + "content": "- PEA-Diffusion can achieve remarkable image generation performance with culture-specific prompts compared with other cross-lingual or multilingual T2I models. It can also retain the general generation ability of the original English diffusion model with slight degradation on evaluation metrics." + }, + { + "type": "list", + "bbox": [ + 0.225, + 0.147, + 0.786, + 0.315 + ], + "angle": 0, + "content": null + }, + { + "type": "title", + "bbox": [ + 0.216, + 0.338, + 0.393, + 0.354 + ], + "angle": 0, + "content": "2 Related Work." + }, + { + "type": "title", + "bbox": [ + 0.216, + 0.372, + 0.583, + 0.387 + ], + "angle": 0, + "content": "2.1 Multilingual Text-to-image Generation" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.398, + 0.789, + 0.732 + ], + "angle": 0, + "content": "Diffusion models have shown a remarkable capability of yielding photorealistic and diverse images given textual descriptions, such as GLIDE [20], Stable Diffusion [25], Imagen [28], DALL-E2 [24], and VQ-Diffusion [8]. Despite their powerful generation abilities, they still only understand English input, leading to limitations. To support non-English text prompt, Taiyi [36] finetunes the Chinese text encoder with a diffusion model. ERNIE-ViLG [34] trains the T2I models from scratch. AltDiffusion [33] proposes a multilingual text encoder to capture multiple language features, which serve as DM conditions. The model was trained in three stages: text encoder fine-tuning, \\( W_{K}, W_{V} \\) fine-tuning, and full fine-tuning. GlueGen [22] has designed a GlueNet module that aligns the original conditional input features of the Image Generator by introducing a new encoder. ENSAD [17] proposes a translation-enhanced multilingual T2I generation model by ensembling multilingual T2I and neural machine translation. BDM [18] proposes a backbone-branch network architecture by integrating an English T2I backbone with a Chinese native semantics injection branch. Despite their success in the non-English I2I generation, these works are either constrained to deal with distinctive language concepts or isolated from English-native T2I communities. At the same time, it is not possible to balance the lower cost of training with the performance of evaluation sets. IAP [13] proposes to treat images as pivots to establish the connection between the embedding space of different text encoders, but it doesn't consider to enhance cultural-related generation ability and expand to downstream tasks." + }, + { + "type": "title", + "bbox": [ + 0.216, + 0.754, + 0.419, + 0.769 + ], + "angle": 0, + "content": "2.2 Multilingual CLIP." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.78, + 0.788, + 0.842 + ], + "angle": 0, + "content": "CLIP uses contrastive language-image pretraining to connect text and image representations, based on massive English text-image sample pairs [23]. Recently, some works have applied teacher learning [2] or metric learning [1] to train multilingual CLIP models, expanding CLIP to other languages. AltCLIP [4] takes this" + } + ], + [ + { + "type": "page_number", + "bbox": [ + 0.218, + 0.116, + 0.23, + 0.127 + ], + "angle": 0, + "content": "4" + }, + { + "type": "header", + "bbox": [ + 0.272, + 0.115, + 0.352, + 0.128 + ], + "angle": 0, + "content": "J. Ma et al." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.147, + 0.788, + 0.314 + ], + "angle": 0, + "content": "one step further enhancing the language capabilities, which is implemented by teacher learning and contrastive learning by leveraging a pretrained multilingual text encoder, XLMR [5]. mCLIP [3] achieves multilingual CLIP by aligning the CLIP text model and a multilingual text encoder. To effectively transfer CLIP to language-specific scenarios, some researchers train language-specific CLIPs on large-scale native image-text pairs, e.g., Chinese CLIP [31]. In order to promote PEA-Diffusion to effectively understand language-specific concepts, we append the PEA module after the language-specific CLIP or Multilingual CLIP and train it to uncover non-English cultural concepts within the T2I. The goal is to establish a bridge between the new CLIP and T2I core structures, such as UNet using the PEA module." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.339, + 0.79, + 0.522 + ], + "angle": 0, + "content": "Knowledge Distillation. KD [9] transfers the knowledge in the output logits of a teacher model with a large size to a smaller student model without obvious performance degradation. The feature map in the hidden states and attention output space can also be used for KD to help students mimic teacher capability [11, 14]. Masked Generative Distillation [32] introduces a feature masking scheme into KD to improve students' representation power by forcing them to recover the masked features. Inspired by masked image modeling, the supervised masked KD model is proposed by inserting label information into self-distillation frameworks. However, to the best of our knowledge, KD has not been fully explored for training multilingual T2I generations. In this paper, we apply KD methods to language transfer in T2I and further harness the power of the PEA module through combination to achieve significant improvements." + }, + { + "type": "title", + "bbox": [ + 0.216, + 0.547, + 0.331, + 0.563 + ], + "angle": 0, + "content": "3 Method" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.583, + 0.788, + 0.779 + ], + "angle": 0, + "content": "A naive approach for generating images from non-English text is to substitute the CLIP text encoder in SD with a language-specific CLIP encoder. The text encoder is a crucial part of T2I models since exact and detailed text embedding is necessary for accurate generation. However, altering an existing text model to a different language-specific one is difficult. This difficulty arises because SD's text encoder and image generator are closely integrated, making it ineffective to just swap out the condition encoder. The primary issue is the mismatch between the new text encoder and the old image generator. Furthermore, our experiments show that joint fine-tuning depends heavily on high-quality text-image pairs and often fails due to catastrophic forgetting when updating well-trained parameters. In light of these insights, we keep the pretrained model's parameters fixed and develop a lightweight adapter with KD to synchronize the new text encoder and the image generator." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.78, + 0.788, + 0.842 + ], + "angle": 0, + "content": "Figure 1 illustrates the overall architecture of PEA-Diffusion for cross-lingual transfer in T2I generation. To inherit the knowledge of the pretrained diffusion model, we use the pretrained SD as a teacher to supervise PEA-Diffusion learning through KD. For language-specific T2I generation, we use the corresponding" + } + ], + [ + { + "type": "header", + "bbox": [ + 0.636, + 0.115, + 0.732, + 0.127 + ], + "angle": 0, + "content": "PEA-diffusion" + }, + { + "type": "page_number", + "bbox": [ + 0.775, + 0.117, + 0.785, + 0.127 + ], + "angle": 0, + "content": "5" + }, + { + "type": "image", + "bbox": [ + 0.274, + 0.144, + 0.722, + 0.317 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.215, + 0.329, + 0.785, + 0.358 + ], + "angle": 0, + "content": "Fig. 1: An overview of the proposed PEA-Diffusion. Notice that only the lightweight adapter is trainable through the whole training process." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.388, + 0.784, + 0.448 + ], + "angle": 0, + "content": "language-specific CLIP as a text encoder to capture text information. The generator in PEA-Diffusion is the same UNet as that used in SD. A light-weight MLP adapter is added on top of the CLIP text encoder in PEA-Diffusion, facilitating knowledge transfer from teacher to student." + }, + { + "type": "title", + "bbox": [ + 0.215, + 0.471, + 0.361, + 0.487 + ], + "angle": 0, + "content": "3.1 Preliminary" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.498, + 0.785, + 0.558 + ], + "angle": 0, + "content": "The diffusion model is composed of a forward diffusion process and a reverse denoising process. The forward process gradually adds random noise to clean data and diffuses it into pure Gaussian noise, while the reverse process reverses the diffusion process to create satisfactory samples from the Gaussian noise [10]." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.558, + 0.786, + 0.68 + ], + "angle": 0, + "content": "Specifically, for an input image \\( x_0 \\in \\mathbb{R}^{H \\times W \\times 3} \\), the encoder \\( \\mathcal{E} \\) of the autoencoder transforms it into a latent representation \\( z_0 \\in \\mathbb{R}^{h \\times w \\times c} \\), where \\( f = H / h = W / w \\) is the downsampling factor and \\( c \\) is the latent feature dimension. The diffusion process is then performed in the latent space, where a conditional UNet [27] denoiser \\( \\epsilon_{\\theta} \\) is employed to predict noise \\( \\epsilon \\) with current timestep \\( t \\), noisy latent \\( z_t \\) and generation condition \\( C \\). In T2I scenarios, the condition \\( C = \\tau_{\\theta}(y) \\) is produced by encoding the text prompts \\( y \\) with a pretrained CLIP [23] text encoder \\( \\tau_{\\theta} \\). Therefore, the overall training objective of SD is defined as" + }, + { + "type": "equation", + "bbox": [ + 0.34, + 0.69, + 0.787, + 0.719 + ], + "angle": 0, + "content": "\\[\n\\mathcal {L} _ {\\mathrm {S D}} = \\mathbb {E} _ {\\mathcal {E} (x _ {0}), C, \\epsilon \\sim \\mathcal {N} (0, 1), t} \\left[ \\| \\epsilon - \\epsilon_ {\\theta} (z _ {t}, t, C) \\| _ {2} ^ {2} \\right] \\tag {1}\n\\]" + }, + { + "type": "title", + "bbox": [ + 0.215, + 0.75, + 0.45, + 0.765 + ], + "angle": 0, + "content": "3.2 Cross-Linguual Transfer" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.775, + 0.787, + 0.841 + ], + "angle": 0, + "content": "Objective. Given sets of parallel corpus \\( Y^{T} = \\{y_{i}^{t}\\}_{i = 1}^{N} \\) and \\( Y^{S} = \\{y_{i}^{t}\\}_{i = 1}^{N} \\) and \\( Y^{S^{\\prime}} = \\{y_{i}^{t}\\}_{i = 1}^{M} \\) and the corresponding images \\( I = \\{img\\}_{i = 1}^{N} \\) with \\( N \\) denoting the total number of paired samples, \\( Y^{T} \\) represents the English corpus that will be input into the teacher model, \\( Y^{S} \\) represents the non-English corpus with" + } + ], + [ + { + "type": "page_number", + "bbox": [ + 0.218, + 0.116, + 0.23, + 0.127 + ], + "angle": 0, + "content": "6" + }, + { + "type": "header", + "bbox": [ + 0.272, + 0.115, + 0.352, + 0.128 + ], + "angle": 0, + "content": "J. Ma et al." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.146, + 0.788, + 0.331 + ], + "angle": 0, + "content": "culturally irrelevant content, and \\( Y^{S'} \\) represents the non-English corpus with cultural relevance, they will be input into the student model. The teacher unet denoiser \\( \\epsilon^T \\) projects pairs \\( (Y^T, I) \\) into source middle feature maps and source logits output. Likewise, the student unet denoiser \\( \\epsilon^S \\) projects pairs \\( (Y^S, I) \\) into target middle feature maps and target logits output. Intuitively, there exists severe distribution mismatch between the source and target output, as well as between \\( \\epsilon^S \\) and \\( \\epsilon^T \\) due to the different languages. Therefore, our target is to make \\( \\epsilon^S \\) match with \\( \\epsilon^T \\), so as to transfer teacher knowledge to the student, enabling cross-lingual T2I generation. For the student, we use a language-specific CLIP text encoder instead of the original English text encoder to enhance token- and sentence-level linguistic-native representations. For brevity, we omit the subscripts of \\( \\mathbb{E}_{\\mathcal{E}(x_0), \\tau_\\theta(y), \\epsilon \\sim \\mathcal{N}(0, 1), t} \\) and \\( z_t, t \\) conditions in the following notations." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.331, + 0.788, + 0.466 + ], + "angle": 0, + "content": "Knowledge Distillation for T2I generation. The original KD leverages Kullback-Leibler (KL) divergence to measure the difference between the categorical probability distribution of the teacher's logits output and the student's logits output [9]. However, in the task of T2I generation, since the middle feature maps are the latent value of pixels instead of categorical probability distribution, KL divergence is unable to measure the difference between student and teacher. Consequently, we replace KL divergence with the L2-norm distance between the logits output from student and teacher to extend raw KD for T2I generation, whose loss function is formulated as:" + }, + { + "type": "equation", + "bbox": [ + 0.387, + 0.473, + 0.786, + 0.499 + ], + "angle": 0, + "content": "\\[\n\\mathcal {L} _ {L K D} = \\mathbb {E} \\left[ \\| \\epsilon^ {T} (y ^ {T}) - \\epsilon^ {S} (y ^ {S}) \\| _ {2} ^ {2} \\right] \\tag {2}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.507, + 0.788, + 0.569 + ], + "angle": 0, + "content": "where \\( y^{T} \\in Y^{T} \\) and \\( y^{S} \\in Y^{S} \\). \\( \\mathcal{L}_{LK D} \\) merely enforces student distribution to match teacher distribution at logits output layer. To further reduce distribution shift of teacher and student, we boost feature alignment of the middle feature maps between \\( \\epsilon^{S} \\) and \\( \\epsilon^{T} \\):" + }, + { + "type": "equation", + "bbox": [ + 0.345, + 0.579, + 0.786, + 0.611 + ], + "angle": 0, + "content": "\\[\n\\mathcal {L} _ {F L K D} = \\mathbb {E} \\left[ \\sum_ {l} \\| f \\left(g _ {l} ^ {T} \\left(y ^ {T}\\right)\\right) - f \\left(g _ {l} ^ {S} \\left(y ^ {S}\\right)\\right) \\| _ {2} ^ {2} \\right] \\tag {3}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.621, + 0.788, + 0.682 + ], + "angle": 0, + "content": "where \\( f \\) represents the norm operation, \\( g_{l}^{T} \\) and \\( g_{l}^{S} \\) represent the feature maps of the \\( l \\)-th layer in a predefined set of unet layers from the teacher and the student, respectively. With respective weights \\( \\lambda_{FL} \\) and \\( \\lambda_{L} \\), the final KD objective is formalized as below:" + }, + { + "type": "equation", + "bbox": [ + 0.391, + 0.695, + 0.786, + 0.71 + ], + "angle": 0, + "content": "\\[\n\\mathcal {L} _ {K D} = \\lambda_ {F L} \\mathcal {L} _ {F L K D} + \\lambda_ {L} \\mathcal {L} _ {L K D} \\tag {4}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.72, + 0.788, + 0.841 + ], + "angle": 0, + "content": "Adapter. Due to the dimension difference between English CLIP and non-English CLIP, we propose an adapter for simple feature transformation to align feature dimension. Moreover, the adapter is trained to learn language-specific information. The adapter is implemented by a MLP with only 6M parameters. To accelerate training and avoid destroying the generalization capability of the pretrained models, we freeze the CLIP encoder and image generator. As a result, the proposed PEA-Diffusion is both parameter- and computation-efficient because just the adapter is trained." + } + ], + [ + { + "type": "header", + "bbox": [ + 0.636, + 0.115, + 0.732, + 0.127 + ], + "angle": 0, + "content": "PEA-diffusion" + }, + { + "type": "page_number", + "bbox": [ + 0.776, + 0.117, + 0.786, + 0.127 + ], + "angle": 0, + "content": "7" + }, + { + "type": "image", + "bbox": [ + 0.226, + 0.146, + 0.778, + 0.404 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.216, + 0.422, + 0.784, + 0.437 + ], + "angle": 0, + "content": "Fig. 2: Image generation visualization of different models in Chinese specific language." + }, + { + "type": "title", + "bbox": [ + 0.216, + 0.466, + 0.578, + 0.481 + ], + "angle": 0, + "content": "3.3 Training Strategy and Final Objective" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.491, + 0.788, + 0.66 + ], + "angle": 0, + "content": "For our goal of training a multi-language T2I generation model, we only need to train on English data since both the new CLIP text encoder and the teacher model's CLIP text encoder can accept English inputs. In this case, we can use aligned KD loss \\(\\mathcal{L}_{KD}\\). For our goal of training a language-specific T2I generation model, we adopt a hybrid training strategy, for the target language's parallel corpus, we only need to use KD loss \\(\\mathcal{L}_{KD}\\) to align the distribution. On the other hand, we also have language-specific culture-related data \\(Y^{S^{\\prime}}\\). Because there is a significant translation error when translating these data into English and the teacher model cannot provide high-quality knowledge guidance for this type of data, we only use \\(\\mathcal{L}_{\\mathrm{SD}}\\) (1) for improving culture understanding ability. The final objective is formalized as follows:" + }, + { + "type": "equation", + "bbox": [ + 0.216, + 0.67, + 0.796, + 0.72 + ], + "angle": 0, + "content": "\\[\n\\mathcal {L} = \\left\\{ \\begin{array}{l l} \\lambda_ {F L} \\mathbb {E} \\Big [ \\| \\epsilon^ {T} (y ^ {T}) - \\epsilon^ {S} (y) \\| _ {2} ^ {2} \\Big ] + \\lambda_ {L} \\mathbb {E} \\Big [ \\sum_ {l} \\| f (g _ {l} ^ {T} (y ^ {T})) - f (g _ {l} ^ {S} (y)) \\| _ {2} ^ {2} \\Big ], & y \\in Y ^ {S} \\\\ \\mathbb {E} \\Big [ \\| \\epsilon - \\epsilon^ {S} (y) \\| _ {2} ^ {2} \\Big ], & y \\in Y ^ {S ^ {\\prime}} \\end{array} \\right.\n\\]" + }, + { + "type": "title", + "bbox": [ + 0.216, + 0.739, + 0.376, + 0.756 + ], + "angle": 0, + "content": "4 Experiments" + }, + { + "type": "title", + "bbox": [ + 0.216, + 0.77, + 0.409, + 0.785 + ], + "angle": 0, + "content": "4.1 Data Preparation" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.795, + 0.787, + 0.842 + ], + "angle": 0, + "content": "We download the Laion2B-multi [29] dataset along with English translations for image descriptions. Based on the open-source CLIP models available for different languages: Chinese CLIP, Russian CLIP, Japanese CLIP, Korean CLIP, and" + } + ], + [ + { + "type": "page_number", + "bbox": [ + 0.218, + 0.116, + 0.23, + 0.127 + ], + "angle": 0, + "content": "8" + }, + { + "type": "header", + "bbox": [ + 0.272, + 0.115, + 0.352, + 0.128 + ], + "angle": 0, + "content": "J. Ma et al." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.147, + 0.788, + 0.208 + ], + "angle": 0, + "content": "Italian CLIP, we determine to conduct experiments on the above five languages. For Chinese, we also collect WuKong [7] data with Chinese cultural image-text pairs and Chinese-English parallel corpus translations of Laion2B-en. For more data, model, and code details, please refer to the Appendix." + }, + { + "type": "title", + "bbox": [ + 0.216, + 0.241, + 0.586, + 0.256 + ], + "angle": 0, + "content": "4.2 Implementation Details and Evaluation" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.277, + 0.788, + 0.382 + ], + "angle": 0, + "content": "Implementation details. We conduct all experiments on Stable Diffusion (SD)- [25], and Stable Diffusion XL (SDXL)- [21] like backbones. Take SDXL base model as an example. PEA-Diffusion consists of VAE, UNet, CLIP text encoder, and the adapter module, comprising 3 billion parameters, out of which a mere 6 million parameters (\\(\\approx 0.2\\%\\)) are trainable. We set \\(\\lambda_{FL}\\) 0.1 and \\(\\lambda_{L}\\) 1 respectively based on the loss performance. The entire model is trained on 64 A100 GPUs for 70000 steps with a batch size of 5 per GPU." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.386, + 0.788, + 0.567 + ], + "angle": 0, + "content": "Test benchmark. To evaluate the capability of PEA-Diffusion to generate images and capture culture-specific concepts of different languages, we introduce two datasets: Multilingual-General (MG) for generation quality evaluation and Multilingual-Cultural (MC) for culture-specific concepts evaluation. We collect 200 prompts for MG, which covered content such as people, animals, landscapes, and fantasy themes. The descriptions include diverse styles, such as cartoon, anime, illustration, watercolor, and cyberpunk. The data comes from various major art websites where users provide real prompts. On the other hand, we also construct MC data including 200 prompts. The collection is mainly based on the ChatGPT tool and consists of culturally relevant content for each specific language, including but not limited to local cuisine, architecture, clothing, and distinctive cultural features." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.569, + 0.788, + 0.84 + ], + "angle": 0, + "content": "Evaluation metrics. For the general evaluation data MG, the widely used FID metric can only measure the feature distance between generated images and specific images. It often fails to provide accurate assessment for high-quality generated images, as emphasized and experimentally validated by SDXL [21] and other studies. Therefore, in order to evaluate the quality of generated images more accurately, we introduce two human preference models as our evaluation metrics: ImageReward [30] and PickScore [15]. Additionally, we continue to utilize OpenCLIP ViT-bigG to measure the similarity between images and text. As for the language-specific evaluation data MC, constructed for each language, to ensure a fair comparison of content relevance, we only consider the CLIP model trained in the specific language to calculate the correlation between generated images and their corresponding text. Since the general CLIP model can only process non-English data after translation into English, this process can introduce significant performance degradation. Furthermore, our paper's focus is precisely on addressing the errors caused by translating non-English prompts into English to generate images. Therefore, other metrics that involve translation processes, such as the two models based on human preference mentioned above, will no longer be valid for evaluation." + } + ], + [ + { + "type": "header", + "bbox": [ + 0.636, + 0.115, + 0.732, + 0.128 + ], + "angle": 0, + "content": "PEA-diffusion" + }, + { + "type": "page_number", + "bbox": [ + 0.775, + 0.116, + 0.787, + 0.127 + ], + "angle": 0, + "content": "9" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.147, + 0.788, + 0.193 + ], + "angle": 0, + "content": "Baseline methods. We generated image results in specific languages based on two T2I base models: SD1.5 and SDXL. We choose the following four categories of approaches as our baselines." + }, + { + "type": "text", + "bbox": [ + 0.226, + 0.202, + 0.581, + 0.216 + ], + "angle": 0, + "content": "- Translation: we apply Google translation API." + }, + { + "type": "text", + "bbox": [ + 0.226, + 0.217, + 0.785, + 0.247 + ], + "angle": 0, + "content": "- AltDiffusion [33]: we test two multilingual SD models based on the open-source AltDiffusion framework: AltDiffusion-m9 and AltDiffusion-18." + }, + { + "type": "text", + "bbox": [ + 0.226, + 0.248, + 0.794, + 0.337 + ], + "angle": 0, + "content": "- GlueGen [22]: We evaluate our model on three languages that overlap with the checkpoint provided by GlueGen, which are Chinese, Italian, and Japanese. For the other two languages, Russian and Korean, we implement GlueGen based on its code and our data. In addition, to further compare the impact of different language-specific CLIPs on the results, we also retrain the GlueGen based on five languages' CLIP text encoders." + }, + { + "type": "text", + "bbox": [ + 0.226, + 0.338, + 0.785, + 0.383 + ], + "angle": 0, + "content": "- Fine-tuning with LoRA [12] and \\( W_{K}, W_{V} \\) (key and value projection matrices): As low-cost experiments for comparison, we also fine-tuning the model to transfer language capabilities." + }, + { + "type": "list", + "bbox": [ + 0.226, + 0.202, + 0.794, + 0.383 + ], + "angle": 0, + "content": null + }, + { + "type": "title", + "bbox": [ + 0.216, + 0.405, + 0.442, + 0.421 + ], + "angle": 0, + "content": "4.3 Experimental Results" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.43, + 0.789, + 0.674 + ], + "angle": 0, + "content": "Generation analysis. We validate our approach based on two different T2I base models. Firstly, to ensure a fair comparison with GlueGen and AltDiffusion, we train PEA-Diffusion based on the SD1.5 model. Secondly, to validate our low-cost experiment approach, we train a larger model based on SDXL. The compared methods are LoRA fine-tuning with rank 64 and \\( W_{K} \\), \\( W_{V} \\) fine-tuning, both of which are parameter-efficient fine-tuning strategies. For the second approach, as proposed in [16], when updating the mapping from given text to image distribution, only updating \\( W_{K}^{(i)} \\), \\( W_{V}^{(i)} \\) in each cross-attention block \\( i \\) is sufficient since text features are the only input to the key and value projection matrix. Furthermore, to validate the wide application of the proposed approach with different text encoders, we train our model with a multi-language T5 encoder and a multi-language CLIP encoder. This type of training only requires English text data, which makes it even simpler and more efficient. Fig. 2 illustrates the comparisons between translation and our methods with different training strategies. It can be observed that our PEA-Diffusion can achieve Chinese culturally related image generation, but translation totally fails in these cases." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.675, + 0.789, + 0.841 + ], + "angle": 0, + "content": "As shown in the comprehensive results of Table 1, when compared with the translation method, our PEA-Diffusion outperforms it for almost all languages (except Japanese) in terms of CLIPScore for MC datasets. It is obvious that the translation method cannot capture the essence of different languages' cultures, leading to a semantic gap between the original prompts and the translated English prompts. However, our approach can use additional culturally related image-text pairs to train the adapter, which can align the language understanding ability of the text encoder with the generation ability of the UNet. Even though the UNet is frozen during the whole training process, we argue that our PEA-Diffusion can still stimulate the potential generation ability for cultural-related images of the original UNet. Since SD1.5 and SDXL are all trained on" + } + ], + [ + { + "type": "page_number", + "bbox": [ + 0.218, + 0.116, + 0.236, + 0.127 + ], + "angle": 0, + "content": "10" + }, + { + "type": "header", + "bbox": [ + 0.272, + 0.115, + 0.352, + 0.128 + ], + "angle": 0, + "content": "J. Ma et al." + }, + { + "type": "table_caption", + "bbox": [ + 0.216, + 0.158, + 0.787, + 0.185 + ], + "angle": 0, + "content": "Table 1: CLIPScore with different specific languages. For Chinese, we evaluated two general evaluation metrics with MG data. \\(\\dagger\\) Indicates open source models." + }, + { + "type": "table", + "bbox": [ + 0.22, + 0.186, + 0.787, + 0.307 + ], + "angle": 0, + "content": "
Model BaseMethodszh-MCzh-MGru-MCit-MCko-MCja-MC
CLIPScore_zhImageRewardPickScoreCLIPScoreRuCLIPScore—itCLIPScore_koCLIPScore_ja
SD1.5Translate0.16860.24270.17990.49110.35560.19040.2885
Altdiffusion m9 [33]†0.2100-0.69660.12020.37150.31770.19060.2461
Altdiffusion m18 [33]†0.21720.33160.16480.47300.35020.19170.3008
GlueGen [22]†0.1700-0.51870.12240.41840.30650.17690.2506
GlueGen [22]0.2153-0.46850.12560.39870.29280.18430.2465
Ours0.25240.09850.17010.50070.37280.18960.2501
SDXLTranslate0.23831.36470.24020.48440.35910.20330.2958
LoRA0.23660.12170.15980.46560.34760.20340.1940
Wkand Wv0.24850.14130.16210.47850.34950.20880.2022
Ours w/ MuTI50.23010.82540.20010.45680.38050.19080.2996
Ours w/ Multilingual-CLIP0.24071.21220.23240.46730.38560.21570.3007
Ours0.26101.02450.22440.50120.41910.21940.2544
" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.338, + 0.784, + 0.429 + ], + "angle": 0, + "content": "billions of image-text pairs that contain wide image domains, our approach can successfully awaken their generative abilities in different cultural contexts. In terms of ImageReward and PickScore for the MG dataset, we must admit that PEA-Diffusion is slightly inferior to translation methods. This is because using other text encoders (e.g., Chinese CLIP and Multilingual CLIP) cannot ensure complete alignment between the text embedding and the UNet feature spaces." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.43, + 0.786, + 0.626 + ], + "angle": 0, + "content": "Even though Altdiffusion retrans the multilingual CLIP text encoder and UNet on billions of image-text pairs, it still performs worse than our approach in terms of almost all metrics for all languages. GlueGen is the most related work on our concerned topic. The statistics presented in Table 1 reveal that both checkpoints provided by the authors or re-implemented by ourselves are significantly inferior to our PEA-Diffusion, especially for the ImageReward and PickScore metrics, resulting in very low generation quality since only aligning distributions of text embedding for different languages is very hard, even with its sophisticatedly designed GlenNet. In addition, it cannot generate culturally related images. Therefore, it's worth mentioning that our core contribution is that we align the model's generation ability through UNet distillation only by a MLP-like adapter between different languages. We don't forcibly align the text embedding spaces of different languages." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.628, + 0.786, + 0.734 + ], + "angle": 0, + "content": "Additionally, we undertake various parameter-efficient fine-tuning strategies for comparison. The fine-tuning techniques, including LoRA and \\( W_{k} \\) and \\( W_{v} \\), show subpar performance based on ImageReward, PickScore, and CLIPScore metrics. This is attributed to the fact that these fine-tuning methods depend significantly on high-quality image-text datasets, whereas our PEA-Diffusion requires merely some parallel multilingual texts and images and lacks stringent quality requirements. Further experiments can be found in the Appendix." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.735, + 0.787, + 0.84 + ], + "angle": 0, + "content": "Analysis on different languages. The data in Table 1 indicate that the CLIPScores for Russian, Italian, and Korean either match or exceed those of other baseline methods by a notable margin. For Japanese, however, the CLIP-Score is slightly lower compared to translation and AltDiffusion. We ascribe this to the CLIP model's limited encoding ability for Japanese, which fails to align with the SD model via the adapter layer. Additionally, with only 1.6M training samples for Korean, our method still achieves effective language transfer, under-" + } + ], + [ + { + "type": "header", + "bbox": [ + 0.636, + 0.115, + 0.732, + 0.127 + ], + "angle": 0, + "content": "PEA-diffusion" + }, + { + "type": "page_number", + "bbox": [ + 0.769, + 0.116, + 0.784, + 0.127 + ], + "angle": 0, + "content": "11" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.147, + 0.784, + 0.178 + ], + "angle": 0, + "content": "scoring that large-scale training datasets are not essential. Further visualization examples of culturally-related prompts can be found in the Appendix." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.179, + 0.538, + 0.42 + ], + "angle": 0, + "content": "Analysis of training cost. Our PEADiffusion freezes the CLIP text encoder, the UNet, and VAE during the whole training process, only the MLP mapping adapter is trained, accounting for merely 6M training parameters, which is \\(0.2\\%\\) of total parameters for SDXL-based models. We employed 64 A100 GPUs to train 70000 steps, resulting in 1600 GPU hours. As a comparison, AltDiffusion [33] trained the model for three stages with nearly 30000 GPU hours, which is \\(\\approx 20\\) times our training cost. In addition, our method achieves higher performance in terms of almost all metrics and the five experimental languages. We also re-implemented and trained GlueGen [22] with the same dataset" + }, + { + "type": "image", + "bbox": [ + 0.557, + 0.221, + 0.784, + 0.329 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.549, + 0.338, + 0.785, + 0.379 + ], + "angle": 0, + "content": "Fig.3: Training parameters(M) and cost comparison for different methods." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.42, + 0.788, + 0.571 + ], + "angle": 0, + "content": "constructed by ourselves. It indeed achieves very low training costs with only 1 A100 GPU for 24 hours, but results in very poor performance no matter for image quality (-0.5 ImageReward score) or language transferring ability. Finally, we also conduct other parameter-efficient fine-tuning strategies (e.g., LoRA and tuning only \\( W_{k} \\) and \\( W_{v} \\) matrices of attention layers) and achieve much lower performance and higher training costs, which demonstrated the effectiveness and efficiency of our PEA-Diffusion. Figure 3 only shows the relationship between training parameters and GPU hours of different methods. For more information on model fitting and training data, as well as adapter module parameter variations, please refer to Appendix." + }, + { + "type": "title", + "bbox": [ + 0.216, + 0.598, + 0.402, + 0.612 + ], + "angle": 0, + "content": "4.4 Ablation Studies" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.628, + 0.788, + 0.764 + ], + "angle": 0, + "content": "Since Chinese data has certain advantages in terms of data quality and quantity, to eliminate errors caused by data, our ablation experiments are all based on Chinese data. The experiments include four parts: First, we conduct a baseline experiment where we only perform standard model fine-tuning without incorporating any KD strategies. Secondly, inspired by GlueGen [22], we employ text encoder KD to explore the impact of distillation at different positions in the SD framework. Next, we perform distillation based on the UNet model, including feature distillation between different UNet layers and logits distillation for the final output. Finally, we combine all of the distillation positions described above." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.765, + 0.788, + 0.84 + ], + "angle": 0, + "content": "Effectiveness of KD. First, KD is crucial for guiding models onto the right learning trajectory for rapid convergence and is vital for the PEA-Diffusion method. As indicated in the first row of Table 2, the standalone fine-tuning method yields notably inferior performance in both image quality (ImageReward and PickScore) and image-text alignment (CLIPScores)." + } + ], + [ + { + "type": "page_number", + "bbox": [ + 0.218, + 0.116, + 0.236, + 0.127 + ], + "angle": 0, + "content": "12" + }, + { + "type": "header", + "bbox": [ + 0.272, + 0.115, + 0.352, + 0.128 + ], + "angle": 0, + "content": "J. Ma et al." + }, + { + "type": "table_caption", + "bbox": [ + 0.3, + 0.158, + 0.701, + 0.171 + ], + "angle": 0, + "content": "Table 2: Ablation results for different knowledge strategies." + }, + { + "type": "table", + "bbox": [ + 0.278, + 0.173, + 0.73, + 0.261 + ], + "angle": 0, + "content": "
EvaluationImageRewardPickScoreCLIPScore_enCLIPScore_zh
Fine-tuning-1.31540.18540.32780.1901
Fine-tuning w/ TKD-0.98750.19970.35060.2287
Fine-tuning w/ ULKD-1.04470.19570.33310.1968
Fine-tuning w/ UFKD0.95560.21740.40850.2557
Fine-tuning w/ UTKD0.84900.21130.40110.2461
Ours(Fine-tuning w/ UKD)1.02450.22440.42240.2610
" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.286, + 0.787, + 0.421 + ], + "angle": 0, + "content": "Effectiveness of UNet Feature KD (UFKD). Our KD method consists of two parts: the first part involves distilling the logits results generated by the UNet (ULKD), we then introduce a distillation of intermediate layer features in the UNet, which serves as the primary knowledge guidance mechanism. As shown in the third and fourth lines of Table 2, we can observe that logit distillation can only improve marginally when compared with direct fine-tuning whereas feature distillation boosts the performance remarkably. We attribute this phenomenon to the generation ability of the large amount of UNet parameters, so that intermediate feature supervision is essential for knowledge transfer." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.422, + 0.787, + 0.572 + ], + "angle": 0, + "content": "Effectiveness of UNet KD. Furthermore, different distillation positions have different effects. Our initial focus is on aligning the feature spaces of different text encoders, so a natural idea is to distill only the feature information of the teacher model's text encoder (TKD). However, as the results show in the second line of Table 2, although the overall performance exceeds the pure fine-tuning method, it is still far from our best performing approach. This illustrates that simply aligning the text encoder is insufficient and ineffective. This statistic also matches the experimental results of GlueGen in Table 1: simply aligning text embedding spaces cannot obtain satisfactory image generation results. Therefore, this further proves the necessity of distilling the features of the UNet itself." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.573, + 0.787, + 0.679 + ], + "angle": 0, + "content": "Finally, we perform KD on both the UNet and text encoder (UTKD). However, the fifth line of Table 2 shows that its effect is worse than performing only UNet distillation. We suspect that the addition of text encoder KD actually weakens the learning of the SD core UNet structure. The main knowledge ability for generating images is mainly influenced by the UNet model, which is responsible for the adapter alignment feature in PEA-Diffusion. More ablation experiments can refer to Appendix." + }, + { + "type": "title", + "bbox": [ + 0.215, + 0.697, + 0.691, + 0.714 + ], + "angle": 0, + "content": "4.5 Exploring the Domain Adaptation of PEA-Diffusion" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.72, + 0.787, + 0.794 + ], + "angle": 0, + "content": "To demonstrate the domain adaptation of our model with minimal data guidance, we first constructed a dataset of 20,000 parallel image and text pairs in the domain of human. We then trained the models based on the models fine-tuned \\( W_{K}^{(i)}, W_{V}^{(i)} \\) mentioned in section 4.3 with UKD, and evaluated on the MG and MC dataset. The results on Chinese are shown Table 3." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.795, + 0.787, + 0.84 + ], + "angle": 0, + "content": "We observed significant improvements in the three metrics on the MG test data. This suggests that PEA can be guided by a small amount of data to perform out-of-domain knowledge transfer." + } + ], + [ + { + "type": "header", + "bbox": [ + 0.636, + 0.115, + 0.732, + 0.127 + ], + "angle": 0, + "content": "PEA-diffusion" + }, + { + "type": "page_number", + "bbox": [ + 0.769, + 0.116, + 0.786, + 0.127 + ], + "angle": 0, + "content": "13" + }, + { + "type": "table_caption", + "bbox": [ + 0.248, + 0.158, + 0.754, + 0.172 + ], + "angle": 0, + "content": "Table 3: Domain adaptation with 20,000 data in Chinese specific language." + }, + { + "type": "table", + "bbox": [ + 0.277, + 0.174, + 0.732, + 0.221 + ], + "angle": 0, + "content": "
EvaluationImageRewardPickScoreCLIPScore_enCLIPScore_zh
Wk and Wv0.14130.16210.36740.2485
Ours(Fine-tuning w/ UKD)0.83680.21450.40980.2501
" + }, + { + "type": "title", + "bbox": [ + 0.216, + 0.248, + 0.724, + 0.264 + ], + "angle": 0, + "content": "4.6 Adaptability Experiment of English Downstream Model" + }, + { + "type": "image", + "bbox": [ + 0.218, + 0.3, + 0.503, + 0.357 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.504, + 0.3, + 0.785, + 0.357 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.22, + 0.358, + 0.475, + 0.376 + ], + "angle": 0, + "content": "sal_scenery LoRA \n西湖，塔和瀑布，日出(West Lake, snow and waterfall, sunrise) \n江南，夏日湖畔的一村村庄(Jianxun, a village by the summer lake)" + }, + { + "type": "image_caption", + "bbox": [ + 0.522, + 0.357, + 0.767, + 0.375 + ], + "angle": 0, + "content": "realistic/easy/animalism/3d-carton Fine-tuning \n一个漂亮的双面的调色板 \n一个白色的小球与white painting light) \n3D电影，4K，高度细致，男人坐在马桶上读报（3D movie, 4K, highly detailed, man sitting on toilet" + }, + { + "type": "image", + "bbox": [ + 0.218, + 0.377, + 0.501, + 0.434 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.501, + 0.378, + 0.785, + 0.434 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.226, + 0.435, + 0.485, + 0.458 + ], + "angle": 0, + "content": "controlnet-canny ControlNet 鸳鸯双栖展双飞，满园存色惹人醉(Mandarin ducks and butterflies fly together, and the spring scenery in the garden is intoxicating); 北国风光，千里冰封，万里雪飘(Northern scenery, thousands of miles of ice, thousands of miles of snow drifting)" + }, + { + "type": "image_caption", + "bbox": [ + 0.521, + 0.436, + 0.761, + 0.454 + ], + "angle": 0, + "content": "SD-XL, Inpainting 0.1 Painting \n2018年，由青岛双星玻璃有限公司（White porcelain plates on benches）一辆小型起重机放在了长椅上（A small crane was placed on the bench）" + }, + { + "type": "image", + "bbox": [ + 0.218, + 0.458, + 0.501, + 0.514 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.501, + 0.457, + 0.784, + 0.513 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.232, + 0.515, + 0.49, + 0.544 + ], + "angle": 0, + "content": "Icm-lora-adsl CDM 带着墨镜的猫咪手里拿着剑，在恶魔城怪里，仙鸽奇侠风格(The cat wearing sunglasses holds a sword in its hand, in the Demon Castle, and the Magic Knight) 红色灯笼，长城，秋天，高清，奇幻风格，史密德(Red Lutron, Great Wall, Autumn, HD, Fantasy Style, Epic Sense)" + }, + { + "type": "image_caption", + "bbox": [ + 0.514, + 0.514, + 0.771, + 0.542 + ], + "angle": 0, + "content": "SSD-B I SSD \n一只超级可爱兔子穿着僧侣服装，肖像猫，皮克斯动画(A super cute rabbit wearing monk clothing, partition)。 \n中国汉奴女孩，阳光明媚，站在踏青花中；近景，摄影风景（Chinese Hanafi girl, sunny, standing in the midst of ruffled jeans, close-up, photography style）" + }, + { + "type": "image_caption", + "bbox": [ + 0.215, + 0.557, + 0.788, + 0.599 + ], + "angle": 0, + "content": "Fig. 4: Downstream task experiments of our proposed PEA. By only adding the trained adapter between the Chinese CLIP text encoder and the UNet, it can achieve language transfer for fine-tuning, LoRA, ControlNet, Inpainting, LCM and compressed models." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.629, + 0.789, + 0.841 + ], + "angle": 0, + "content": "In this section, we would like to emphasize the plug-and-play feature of our PEA module. Currently, a wide range of community resources, such as fine-tuned models, LoRA, Dreambooth, ControlNet, compressed SD models, and accelerated sampling SD models, are derived from the English SD models, so that the English community is much more mature than other languages. Therefore, one key challenge we must overcome is bridging the compatibility gap between the new model and the English SD community after language transfer. Since our approach uses feature-level and logit-level KD to train only the PEA module while keeping the core UNet parameters unchanged, it naturally maintains compatibility with the English SD community and can seamlessly integrate with various English SD community plugins. Interestingly, the PEA module remains plug-and-play for models that have undergone changes in base SD UNet parameters in contrast to PEA-Diffusion but whose overall structure remains similar. This includes models such as fine-tuned SD and compressed SD trained using" + } + ], + [ + { + "type": "page_number", + "bbox": [ + 0.218, + 0.116, + 0.236, + 0.127 + ], + "angle": 0, + "content": "14" + }, + { + "type": "header", + "bbox": [ + 0.272, + 0.115, + 0.352, + 0.128 + ], + "angle": 0, + "content": "J. Ma et al." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.147, + 0.785, + 0.191 + ], + "angle": 0, + "content": "pruning and distillation techniques, among others. In the following, we will introduce several downstream image generation applications based on the SD or SDXL foundation model." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.192, + 0.797, + 0.252 + ], + "angle": 0, + "content": "Fine-tuned checkpoint. We use two fine-tuned SDXL-based model checkpoints: \"xxmix9realistic\" and \"samaritan-3d-cartoon\". The \"xxmix9realistic\" model is trained on person, photorealistic, sexy, female, woman, girls, etc. The \"samaritan-3d-cartoon\" model is trained on anime, Disney, 3D, Pixar, etc." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.253, + 0.785, + 0.297 + ], + "angle": 0, + "content": "LoRA. \"csal_scenery\" is a SDXL-based LoRA model that contains common ancient Chinese architectural forms such as pavilions, pagodas, towers, gardens, courtyards, bridges, temples, palaces, towns, etc." + }, + { + "type": "text", + "bbox": [ + 0.24, + 0.298, + 0.785, + 0.313 + ], + "angle": 0, + "content": "ControlNet. \"controlnet-canny\" is trained on SDXL under canny condition." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.314, + 0.785, + 0.358 + ], + "angle": 0, + "content": "Inpainting. SDXL Inpainting 0.1 is a latent T2I diffusion model capable of generating photorealistic images given any text input, with the extra capability of Inpainting the pictures by using a mask." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.359, + 0.785, + 0.403 + ], + "angle": 0, + "content": "LCM. Latent Consistency Models [19] have achieved impressive performance in accelerating T2I generative tasks, producing high-quality images with minimal inference steps. \"lcm-lora-sdxl\" is an acceleration module based on SDXL." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.404, + 0.785, + 0.449 + ], + "angle": 0, + "content": "SSD. The Segmind Stable Diffusion Model is a condensed version of the SDXL that is \\(50\\%\\) smaller in size, delivering a \\(60\\%\\) increase in speed while preserving high-quality text-to-image (T2I) generation capabilities." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.45, + 0.785, + 0.509 + ], + "angle": 0, + "content": "In Figure 4, we have shown that the PEA module can be used as a plugin between the Chinese CLIP text encoder and the UNet, achieving the ability to transfer language from English to Chinese directly. For more details of the abovementioned checkpoints and generated images, please refer to our Appendix." + }, + { + "type": "title", + "bbox": [ + 0.215, + 0.533, + 0.51, + 0.549 + ], + "angle": 0, + "content": "5 Conclusion and Limitation" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.564, + 0.787, + 0.715 + ], + "angle": 0, + "content": "Conclusion. The high training expenses and limited training data available for multilingual T2I models have driven us to search for a more lightweight solution. In this paper, we introduce a straightforward and efficient method using KD and adapters. By focusing solely on training the adapter and keeping other parameters fixed, our approach harnesses the capability of T2I models for multilingual cultural comprehension and generation. Our method not only achieves significantly better results for culturally relevant data but also surpasses existing models while maintaining general synthesis capability. Comprehensive experiments demonstrate that our method is cost-effective, versatile, and easy to integrate." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.716, + 0.787, + 0.837 + ], + "angle": 0, + "content": "**Limitation.** First, the effectiveness of the CLIP text encoder significantly influences both the quality of image generation and text relevance. Utilizing a weaker language-specific CLIP could impact the performance of PEA-Diffusion in broad open domains and may slightly compromise the model's generalization abilities. Second, the upper limit of our method is constrained by the capabilities of the English base model, indicating that our approach cannot exceed its inherent potential. Moving forward, we aim to conduct additional research to mitigate these limitations." + } + ], + [ + { + "type": "header", + "bbox": [ + 0.636, + 0.115, + 0.732, + 0.127 + ], + "angle": 0, + "content": "PEA-diffusion" + }, + { + "type": "page_number", + "bbox": [ + 0.769, + 0.117, + 0.786, + 0.127 + ], + "angle": 0, + "content": "15" + }, + { + "type": "title", + "bbox": [ + 0.218, + 0.145, + 0.323, + 0.16 + ], + "angle": 0, + "content": "References" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.179, + 0.785, + 0.208 + ], + "angle": 0, + "content": "1. Aggarwal, P., Kale, A.: Towards zero-shot cross-lingual image retrieval. arXiv preprint arXiv:2012.05107 (2020)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.209, + 0.785, + 0.25 + ], + "angle": 0, + "content": "2. Carlsson, F., Eisen, P., Rekathati, F., Sahlgren, M.: Cross-lingual and multilingual clip. In: Proceedings of the Thirteenth Language Resources and Evaluation Conference. pp. 6848-6854 (2022)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.251, + 0.785, + 0.305 + ], + "angle": 0, + "content": "3. Chen, G., Hou, L., Chen, Y., Dai, W., Shang, L., Jiang, X., Liu, Q., Pan, J., Wang, W.: mclip: Multilingual clip via cross-lingual transfer. In: Proceedings of the 61st Annual Meeting of the Association for Computational Linguistics (Volume 1: Long Papers). pp. 13028-13043 (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.306, + 0.785, + 0.334 + ], + "angle": 0, + "content": "4. Chen, Z., Liu, G., Zhang, B.W., Ye, F., Yang, Q., Wu, L.: Altclip: Altering the language encoder in clip for extended language capabilities (2022)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.334, + 0.785, + 0.377 + ], + "angle": 0, + "content": "5. Conneau, A., Khandelwal, K., Goyal, N., Chaudhary, V., Wenzek, G., Guzmán, F., Grave, E., Ott, M., Zettlemoyer, L., Stoyanov, V.: Unsupervised cross-lingual representation learning at scale. arXiv preprint arXiv:1911.02116 (2019)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.378, + 0.785, + 0.418 + ], + "angle": 0, + "content": "6. Ding, M., Yang, Z., Hong, W., Zheng, W., Zhou, C., Yin, D., Lin, J., Zou, X., Shao, Z., Yang, H., et al.: Cogview: Mastering text-to-image generation via transformers. Advances in Neural Information Processing Systems 34, 19822-19835 (2021)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.42, + 0.785, + 0.473 + ], + "angle": 0, + "content": "7. Gu, J., Meng, X., Lu, G., Hou, L., Minzhe, N., Liang, X., Yao, L., Huang, R., Zhang, W., Jiang, X., et al.: Wukong: A 100 million large-scale chinese cross-modal pre-training benchmark. Advances in Neural Information Processing Systems 35, 26418-26431 (2022)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.475, + 0.785, + 0.53 + ], + "angle": 0, + "content": "8. Gu, S., Chen, D., Bao, J., Wen, F., Zhang, B., Chen, D., Yuan, L., Guo, B.: Vector quantized diffusion model for text-to-image synthesis. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 10696-10706 (2022)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.531, + 0.785, + 0.559 + ], + "angle": 0, + "content": "9. Hinton, G., Vinyals, O., Dean, J.: Distilling the knowledge in a neural network. arXiv preprint arXiv:1503.02531 (2015)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.56, + 0.785, + 0.587 + ], + "angle": 0, + "content": "10. Ho, J., Jain, A., Abbeel, P.: Denoising diffusion probabilistic models. Advances in Neural Information Processing systems 33, 6840-6851 (2020)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.588, + 0.785, + 0.628 + ], + "angle": 0, + "content": "1. Hou, L., Huang, Z., Shang, L., Jiang, X., Chen, X., Liu, Q.: Dynabert: Dynamic bert with adaptive width and depth. Advances in Neural Information Processing Systems 33, 9782-9793 (2020)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.63, + 0.785, + 0.657 + ], + "angle": 0, + "content": "2. Hu, E.J., Shen, Y., Wallis, P., Allen-Zhu, Z., Li, Y., Wang, S., Wang, L., Chen, W.: Lora: Low-rank adaptation of large language models (2021)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.659, + 0.785, + 0.699 + ], + "angle": 0, + "content": "3. Hu, J., Han, X., Yi, X., Chen, Y., Li, W., Liu, Z., Sun, M.: Efficient crosslingual transfer for chinese stable diffusion with images as pivots. arXiv preprint arXiv:2305.11540 (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.7, + 0.785, + 0.741 + ], + "angle": 0, + "content": "4. Jiao, X., Yin, Y., Shang, L., Jiang, X., Chen, X., Li, L., Wang, F., Liu, Q.: Tinybert: Distilling bert for natural language understanding. arXiv preprint arXiv:1909.10351 (2019)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.742, + 0.785, + 0.77 + ], + "angle": 0, + "content": "5. Kirstain, Y., Polyak, A., Singer, U., Matiana, S., Penna, J., Levy, O.: Pick-a-pic: An open dataset of user preferences for text-to-image generation (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.771, + 0.785, + 0.812 + ], + "angle": 0, + "content": "6. Kumari, N., Zhang, B., Zhang, R., Shechtman, E., Zhu, J.Y.: Multi-concept customization of text-to-image diffusion. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 1931-1941 (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.813, + 0.785, + 0.84 + ], + "angle": 0, + "content": "7. Li, Y., Chang, C.Y., Rawls, S., Vulic, I., Korhonen, A.: Translation-enhanced multilingual text-to-image generation. arXiv preprint arXiv:2305.19216 (2023)" + }, + { + "type": "list", + "bbox": [ + 0.226, + 0.179, + 0.785, + 0.84 + ], + "angle": 0, + "content": null + } + ], + [ + { + "type": "page_number", + "bbox": [ + 0.218, + 0.116, + 0.236, + 0.127 + ], + "angle": 0, + "content": "16" + }, + { + "type": "header", + "bbox": [ + 0.272, + 0.115, + 0.352, + 0.128 + ], + "angle": 0, + "content": "J. Ma et al." + }, + { + "type": "ref_text", + "bbox": [ + 0.218, + 0.148, + 0.785, + 0.175 + ], + "angle": 0, + "content": "18. Liu, S., Leng, D., Yin, Y.: Bridge diffusion model: bridge non-english language-native text-to-image diffusion model with english communities (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.218, + 0.177, + 0.785, + 0.204 + ], + "angle": 0, + "content": "19. Luo, S., Tan, Y., Huang, L., Li, J., Zhao, H.: Latent consistency models: Synthesizing high-resolution images with few-step inference (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.206, + 0.785, + 0.247 + ], + "angle": 0, + "content": "20. Nichol, A., Dhariwal, P., Ramesh, A., Shyam, P., Mishkin, P., McGrew, B., Sutskever, I., Chen, M.: Glide: Towards photorealistic image generation and editing with text-guided diffusion models. arXiv preprint arXiv:2112.10741 (2021)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.249, + 0.785, + 0.288 + ], + "angle": 0, + "content": "21. Podell, D., English, Z., Lacey, K., Blattmann, A., Dockhorn, T., Müller, J., Penna, J., Rombach, R.: Sdxl: Improving latent diffusion models for high-resolution image synthesis (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.291, + 0.785, + 0.331 + ], + "angle": 0, + "content": "22. Qin, C., Yu, N., Xing, C., Zhang, S., Chen, Z., Ermon, S., Fu, Y., Xiong, C., Xu, R.: Gluegen: Plug and play multi-modal encoders for x-to-image generation. arXiv preprint arXiv:2303.10056 (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.333, + 0.785, + 0.387 + ], + "angle": 0, + "content": "23. Radford, A., Kim, J.W., Hallacy, C., Ramesh, A., Goh, G., Agarwal, S., Sastry, G., Askell, A., Mishkin, P., Clark, J., et al.: Learning transferable visual models from natural language supervision. In: International conference on machine learning. pp. 8748-8763. PMLR (2021)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.389, + 0.785, + 0.416 + ], + "angle": 0, + "content": "24. Ramesh, A., Dhariwal, P., Nichol, A., Chu, C., Chen, M.: Hierarchical text-conditional image generation with clip latents (2022)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.418, + 0.785, + 0.459 + ], + "angle": 0, + "content": "25. Rombach, R., Blattmann, A., Lorenz, D., Esser, P., Ommer, B.: High-resolution image synthesis with latent diffusion models. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 10684-10695 (2022)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.461, + 0.785, + 0.487 + ], + "angle": 0, + "content": "26. Rombach, R., Blattmann, A., Lorenz, D., Esser, P., Ommer, B.: High-resolution image synthesis with latent diffusion models (2022)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.489, + 0.785, + 0.543 + ], + "angle": 0, + "content": "27. Ronneberger, O., Fischer, P., Brox, T.: U-net: Convolutional networks for biomedical image segmentation. In: Medical Image Computing and Computer-Assisted Intervention-MICCAI 2015: 18th International Conference, Munich, Germany, October 5-9, 2015, Proceedings, Part III 18. pp. 234-241. Springer (2015)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.546, + 0.785, + 0.599 + ], + "angle": 0, + "content": "28. Saharia, C., Chan, W., Saxena, S., Li, L., Whang, J., Denton, E., Ghasemipour, S.K.S., Ayan, B.K., Mahdavi, S.S., Lopes, R.G., Salimans, T., Ho, J., Fleet, D.J., Norouzi, M.: Photorealistic text-to-image diffusion models with deep language understanding (2022)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.601, + 0.785, + 0.655 + ], + "angle": 0, + "content": "29. Schuhmann, C., Beaumont, R., Vencu, R., Gordon, C., Wightman, R., Cherti, M., Coombes, T., Katta, A., Mullis, C., Wortsman, M., Schramowski, P., Kundurthy, S., Crowson, K., Schmidt, L., Kaczmarczyk, R., Jitsev, J.: Laion-5b: An open large-scale dataset for training next generation image-text models (2022)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.657, + 0.785, + 0.697 + ], + "angle": 0, + "content": "30. Xu, J., Liu, X., Wu, Y., Tong, Y., Li, Q., Ding, M., Tang, J., Dong, Y.: Imagereward: Learning and evaluating human preferences for text-to-image generation (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.7, + 0.785, + 0.74 + ], + "angle": 0, + "content": "31. Yang, A., Pan, J., Lin, J., Men, R., Zhang, Y., Zhou, J., Zhou, C.: Chinese clip: Contrastive vision-language pretraining in chinese. arXiv preprint arXiv:2211.01335 (2022)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.742, + 0.785, + 0.769 + ], + "angle": 0, + "content": "32. Yang, Z., Li, Z., Shao, M., Shi, D., Yuan, Z., Yuan, C.: Masked generative distillation. In: European Conference on Computer Vision. pp. 53-69. Springer (2022)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.771, + 0.785, + 0.797 + ], + "angle": 0, + "content": "33. Ye, F., Liu, G., Wu, X., Wu, L.: Altdiffusion: A multilingual text-to-image diffusion model (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.799, + 0.785, + 0.84 + ], + "angle": 0, + "content": "34. Zhang, H., Yin, W., Fang, Y., Li, L., Duan, B., Wu, Z., Sun, Y., Tian, H., Wu, H., Wang, H.: Ernie-vilg: Unified generative pre-training for bidirectional vision-language generation. arXiv preprint arXiv:2112.15283 (2021)" + }, + { + "type": "list", + "bbox": [ + 0.217, + 0.148, + 0.785, + 0.84 + ], + "angle": 0, + "content": null + } + ], + [ + { + "type": "header", + "bbox": [ + 0.636, + 0.115, + 0.732, + 0.127 + ], + "angle": 0, + "content": "PEA-diffusion" + }, + { + "type": "page_number", + "bbox": [ + 0.769, + 0.117, + 0.786, + 0.127 + ], + "angle": 0, + "content": "17" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.147, + 0.788, + 0.218 + ], + "angle": 0, + "content": "35. Zhang, J., Gan, R., Wang, J., Zhang, Y., Zhang, L., Yang, P., Gao, X., Wu, Z., Dong, X., He, J., Zhuo, J., Yang, Q., Huang, Y., Li, X., Wu, Y., Lu, J., Zhu, X., Chen, W., Han, T., Pan, K., Wang, R., Wang, H., Wu, X., Zeng, Z., Chen, C.: Fengshenbang 1.0: Being the foundation of chinese cognitive intelligence. CoRR abs/2209.02970 (2022)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.218, + 0.787, + 0.26 + ], + "angle": 0, + "content": "36. Zhang, J., Gan, R., Wang, J., Zhang, Y., Zhang, L., Yang, P., Gao, X., Wu, Z., Dong, X., He, J., et al.: Fengshenbang 1.0: Being the foundation of chinese cognitive intelligence. arXiv preprint arXiv:2209.02970 (2022)" + }, + { + "type": "list", + "bbox": [ + 0.217, + 0.147, + 0.788, + 0.26 + ], + "angle": 0, + "content": null + } + ] +] \ No newline at end of file diff --git a/2024/PEA-Diffusion_ Parameter-Efficient Adapter with Knowledge Distillation in non-English Text-to-Image Generation/e154c47c-6f75-47c7-a454-a9dcc00f12d4_origin.pdf b/2024/PEA-Diffusion_ Parameter-Efficient Adapter with Knowledge Distillation in non-English Text-to-Image Generation/e154c47c-6f75-47c7-a454-a9dcc00f12d4_origin.pdf new file mode 100644 index 0000000000000000000000000000000000000000..19b6c931b27f4c6c0cc08f8db274a879adbc8e7c --- /dev/null +++ b/2024/PEA-Diffusion_ Parameter-Efficient Adapter with Knowledge Distillation in non-English Text-to-Image Generation/e154c47c-6f75-47c7-a454-a9dcc00f12d4_origin.pdf @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:91ee605c745df797715e70a98e86fde47dc4c672710fd5afb867b2f8096b14f6 +size 8702178 diff --git a/2024/PEA-Diffusion_ Parameter-Efficient Adapter with Knowledge Distillation in non-English Text-to-Image Generation/full.md b/2024/PEA-Diffusion_ Parameter-Efficient Adapter with Knowledge Distillation in non-English Text-to-Image Generation/full.md new file mode 100644 index 0000000000000000000000000000000000000000..cbe0d5f9816b0587e88455fac00c12305ab641f2 --- /dev/null +++ b/2024/PEA-Diffusion_ Parameter-Efficient Adapter with Knowledge Distillation in non-English Text-to-Image Generation/full.md @@ -0,0 +1,266 @@ +# PEA-Diffusion: Parameter-Efficient Adapter with Knowledge Distillation in non-English Text-to-Image Generation + +Jian Ma, Chen Chen†, Qingsong Xie, and Haonan Lu† + +OPPO AI Center +{majian2, chenchen4, xieqingsong1, luhaonan} @oppo.com + +Abstract. Text-to-image diffusion models are well known for their ability to generate realistic images based on textual prompts. However, the existing works have predominantly focused on English, lacking support for non-English text-to-image models. The most commonly used translation methods cannot solve the generation problem related to language culture, while training from scratch on a specific language dataset is prohibitively expensive. In this paper, we are inspired to propose a simple plug-and-play language transfer method based on knowledge distillation. All we need to do is train a lightweight MLP-like parameter-efficient adapter (PEA) with only 6M parameters under teacher knowledge distillation along with a small parallel data corpus. We are surprised to find that freezing the parameters of UNet can still achieve remarkable performance on the language-specific prompt evaluation set, demonstrating that PEA can stimulate the potential generation ability of the original UNet. Additionally, it closely approaches the performance of the English text-to-image model on a general prompt evaluation set. Furthermore, our adapter can be used as a plugin to achieve significant results in downstream tasks in cross-lingual text-to-image generation. + +Keywords: Diffusion models $\cdot$ Parameter-efficient $\cdot$ Plug-and-play + +# 1 Introduction + +Recent text-to-image (T2I) models, such as DALL-E2 [24], Imagen [28], and Stable Diffusion (SD) [26], have opened up a new era of AI painting due to their ability to generate realistic, photo-level, and creative images. Numerous downstream tasks have further driven the iteration and update of this technology, as well as the activity of the community, including image editing, controllable generation, etc. However, mainstream text-to-image models only support English prompts, which means non-English speakers usually rely on translations to use them. This may cause translation errors and information loss due to culture-specific concepts. For example, one famous Chinese dish, "Hongshao shizi tou", may be translated to "Braised lion head" in English. Additionally, the introduction of large translation networks leads to high computation consumption. + +To address this issue, recent researchers have created T2I generative models that process native languages directly, bypassing any need for translation. Taiyi-Bilingual [35] incorporates a novel text encoder initialized with Chinese CLIP checkpoints, focusing only on training the text encoder. AltDiffusion [33] introduces a multi-stage training process for multilingual T2I generation, which is relatively costly. Some experts begin training T2I models from scratch using different language data. For example, ERNIE-ViLG 2.0 [6] is trained with native Chinese content and can generate high-quality images that reflect Chinese cultural elements. However, this approach is very expensive and often beyond the reach of many research facilities. Additionally, this method does not foster a community impact, necessitating the retraining of English models for various downstream or commonly used community tasks. GlueGen [22] introduces a GlueNet module to align the input features of the original image generator by adding a new encoder. Despite its low training cost, its effectiveness in generating culturally relevant images in multiple languages and handling broad prompts is comparatively limited. + +As such, exploring a parameter-efficient plug-and-play method that adapts the publicly available, English-naive, pretrained T2I models to suit non-English and culture-specific concepts is becoming more and more urgent. Consequently, we concentrate on a critical research question (RQ): Is it possible to devise a low-cost method to acquire a non-English T2I model by adapting the existing pretrained T2I models efficiently? To this end, we take SD as the baseline, as it has shown surprisingly good results on T2I tasks. To finetune SD with minimal cost, we propose a novel parameter-efficient adapter (PEA) method with knowledge distillation (KD) to make full use of the pretrained SD model, referred to as PEA-Diffusion. To take advantage of the pre-trained SD model directly, the pre-trained SD is used as a teacher to guide student learning. Specifically, we propose the PEA module to align the representation space of the new condition encoder in the student with that of the teacher's image generator. PEA merely consists of a light-weight multi-layer perception (MLP), greatly simplifying the model construction process and preserving the image generation capabilities of the original SD model. Different from GlueNet [22] which forces text feature alignment of two text encoders at encoder space, we propose KD loss to compel representation alignment at feature maps and logits space from the UNet module between our PEA-Diffusion and pretrained English SD models. We argue that our method boosts the image generator to understand the information from the new text encoder. To avoid catastrophic forgetting and reduce trainable parameters, UNet parameters are kept frozen during the entire training procedure. + +Take Chinese SD as an example. We conduct a thorough experimental analysis on KD to help understand the effect of where to perform feature alignment between the teacher and the student. In addition, we provided intuitive analysis about what parameters should be updated, facilitating discovering the impact of the SD parameters for adapting SD into language-specific T2I models. It should be noted that the method could be applicable for any other non-English T2I generation model. Our core contributions are summarized as follows: + +- We propose a novel PEA-Diffusion model to resolve the T2I model's English-native bias, supporting non-English prompts. The PEA aligns the feature and output spaces between the source and target UNet for bilingual prompts, leading to extremely low training costs. +- PEA module is plug-and-play, it can be directly applied to various T2I tasks, including LoRA, ControlNet, Inpainting, compressed SD models that have been pruned and models accelerated through various strategies. +- PEA-Diffusion can achieve remarkable image generation performance with culture-specific prompts compared with other cross-lingual or multilingual T2I models. It can also retain the general generation ability of the original English diffusion model with slight degradation on evaluation metrics. + +# 2 Related Work. + +# 2.1 Multilingual Text-to-image Generation + +Diffusion models have shown a remarkable capability of yielding photorealistic and diverse images given textual descriptions, such as GLIDE [20], Stable Diffusion [25], Imagen [28], DALL-E2 [24], and VQ-Diffusion [8]. Despite their powerful generation abilities, they still only understand English input, leading to limitations. To support non-English text prompt, Taiyi [36] finetunes the Chinese text encoder with a diffusion model. ERNIE-ViLG [34] trains the T2I models from scratch. AltDiffusion [33] proposes a multilingual text encoder to capture multiple language features, which serve as DM conditions. The model was trained in three stages: text encoder fine-tuning, $W_{K}, W_{V}$ fine-tuning, and full fine-tuning. GlueGen [22] has designed a GlueNet module that aligns the original conditional input features of the Image Generator by introducing a new encoder. ENSAD [17] proposes a translation-enhanced multilingual T2I generation model by ensembling multilingual T2I and neural machine translation. BDM [18] proposes a backbone-branch network architecture by integrating an English T2I backbone with a Chinese native semantics injection branch. Despite their success in the non-English I2I generation, these works are either constrained to deal with distinctive language concepts or isolated from English-native T2I communities. At the same time, it is not possible to balance the lower cost of training with the performance of evaluation sets. IAP [13] proposes to treat images as pivots to establish the connection between the embedding space of different text encoders, but it doesn't consider to enhance cultural-related generation ability and expand to downstream tasks. + +# 2.2 Multilingual CLIP. + +CLIP uses contrastive language-image pretraining to connect text and image representations, based on massive English text-image sample pairs [23]. Recently, some works have applied teacher learning [2] or metric learning [1] to train multilingual CLIP models, expanding CLIP to other languages. AltCLIP [4] takes this + +one step further enhancing the language capabilities, which is implemented by teacher learning and contrastive learning by leveraging a pretrained multilingual text encoder, XLMR [5]. mCLIP [3] achieves multilingual CLIP by aligning the CLIP text model and a multilingual text encoder. To effectively transfer CLIP to language-specific scenarios, some researchers train language-specific CLIPs on large-scale native image-text pairs, e.g., Chinese CLIP [31]. In order to promote PEA-Diffusion to effectively understand language-specific concepts, we append the PEA module after the language-specific CLIP or Multilingual CLIP and train it to uncover non-English cultural concepts within the T2I. The goal is to establish a bridge between the new CLIP and T2I core structures, such as UNet using the PEA module. + +Knowledge Distillation. KD [9] transfers the knowledge in the output logits of a teacher model with a large size to a smaller student model without obvious performance degradation. The feature map in the hidden states and attention output space can also be used for KD to help students mimic teacher capability [11, 14]. Masked Generative Distillation [32] introduces a feature masking scheme into KD to improve students' representation power by forcing them to recover the masked features. Inspired by masked image modeling, the supervised masked KD model is proposed by inserting label information into self-distillation frameworks. However, to the best of our knowledge, KD has not been fully explored for training multilingual T2I generations. In this paper, we apply KD methods to language transfer in T2I and further harness the power of the PEA module through combination to achieve significant improvements. + +# 3 Method + +A naive approach for generating images from non-English text is to substitute the CLIP text encoder in SD with a language-specific CLIP encoder. The text encoder is a crucial part of T2I models since exact and detailed text embedding is necessary for accurate generation. However, altering an existing text model to a different language-specific one is difficult. This difficulty arises because SD's text encoder and image generator are closely integrated, making it ineffective to just swap out the condition encoder. The primary issue is the mismatch between the new text encoder and the old image generator. Furthermore, our experiments show that joint fine-tuning depends heavily on high-quality text-image pairs and often fails due to catastrophic forgetting when updating well-trained parameters. In light of these insights, we keep the pretrained model's parameters fixed and develop a lightweight adapter with KD to synchronize the new text encoder and the image generator. + +Figure 1 illustrates the overall architecture of PEA-Diffusion for cross-lingual transfer in T2I generation. To inherit the knowledge of the pretrained diffusion model, we use the pretrained SD as a teacher to supervise PEA-Diffusion learning through KD. For language-specific T2I generation, we use the corresponding + +![](images/46c10031022c06625b26542269807a2a20bacc6b0c3e8b3b9f10c9a087c28369.jpg) +Fig. 1: An overview of the proposed PEA-Diffusion. Notice that only the lightweight adapter is trainable through the whole training process. + +language-specific CLIP as a text encoder to capture text information. The generator in PEA-Diffusion is the same UNet as that used in SD. A light-weight MLP adapter is added on top of the CLIP text encoder in PEA-Diffusion, facilitating knowledge transfer from teacher to student. + +# 3.1 Preliminary + +The diffusion model is composed of a forward diffusion process and a reverse denoising process. The forward process gradually adds random noise to clean data and diffuses it into pure Gaussian noise, while the reverse process reverses the diffusion process to create satisfactory samples from the Gaussian noise [10]. + +Specifically, for an input image $x_0 \in \mathbb{R}^{H \times W \times 3}$ , the encoder $\mathcal{E}$ of the autoencoder transforms it into a latent representation $z_0 \in \mathbb{R}^{h \times w \times c}$ , where $f = H / h = W / w$ is the downsampling factor and $c$ is the latent feature dimension. The diffusion process is then performed in the latent space, where a conditional UNet [27] denoiser $\epsilon_{\theta}$ is employed to predict noise $\epsilon$ with current timestep $t$ , noisy latent $z_t$ and generation condition $C$ . In T2I scenarios, the condition $C = \tau_{\theta}(y)$ is produced by encoding the text prompts $y$ with a pretrained CLIP [23] text encoder $\tau_{\theta}$ . Therefore, the overall training objective of SD is defined as + +$$ +\mathcal {L} _ {\mathrm {S D}} = \mathbb {E} _ {\mathcal {E} (x _ {0}), C, \epsilon \sim \mathcal {N} (0, 1), t} \left[ \| \epsilon - \epsilon_ {\theta} (z _ {t}, t, C) \| _ {2} ^ {2} \right] \tag {1} +$$ + +# 3.2 Cross-Linguual Transfer + +Objective. Given sets of parallel corpus $Y^{T} = \{y_{i}^{t}\}_{i = 1}^{N}$ and $Y^{S} = \{y_{i}^{t}\}_{i = 1}^{N}$ and $Y^{S^{\prime}} = \{y_{i}^{t}\}_{i = 1}^{M}$ and the corresponding images $I = \{img\}_{i = 1}^{N}$ with $N$ denoting the total number of paired samples, $Y^{T}$ represents the English corpus that will be input into the teacher model, $Y^{S}$ represents the non-English corpus with + +culturally irrelevant content, and $Y^{S'}$ represents the non-English corpus with cultural relevance, they will be input into the student model. The teacher unet denoiser $\epsilon^T$ projects pairs $(Y^T, I)$ into source middle feature maps and source logits output. Likewise, the student unet denoiser $\epsilon^S$ projects pairs $(Y^S, I)$ into target middle feature maps and target logits output. Intuitively, there exists severe distribution mismatch between the source and target output, as well as between $\epsilon^S$ and $\epsilon^T$ due to the different languages. Therefore, our target is to make $\epsilon^S$ match with $\epsilon^T$ , so as to transfer teacher knowledge to the student, enabling cross-lingual T2I generation. For the student, we use a language-specific CLIP text encoder instead of the original English text encoder to enhance token- and sentence-level linguistic-native representations. For brevity, we omit the subscripts of $\mathbb{E}_{\mathcal{E}(x_0), \tau_\theta(y), \epsilon \sim \mathcal{N}(0, 1), t}$ and $z_t, t$ conditions in the following notations. + +Knowledge Distillation for T2I generation. The original KD leverages Kullback-Leibler (KL) divergence to measure the difference between the categorical probability distribution of the teacher's logits output and the student's logits output [9]. However, in the task of T2I generation, since the middle feature maps are the latent value of pixels instead of categorical probability distribution, KL divergence is unable to measure the difference between student and teacher. Consequently, we replace KL divergence with the L2-norm distance between the logits output from student and teacher to extend raw KD for T2I generation, whose loss function is formulated as: + +$$ +\mathcal {L} _ {L K D} = \mathbb {E} \left[ \| \epsilon^ {T} (y ^ {T}) - \epsilon^ {S} (y ^ {S}) \| _ {2} ^ {2} \right] \tag {2} +$$ + +where $y^{T} \in Y^{T}$ and $y^{S} \in Y^{S}$ . $\mathcal{L}_{LK D}$ merely enforces student distribution to match teacher distribution at logits output layer. To further reduce distribution shift of teacher and student, we boost feature alignment of the middle feature maps between $\epsilon^{S}$ and $\epsilon^{T}$ : + +$$ +\mathcal {L} _ {F L K D} = \mathbb {E} \left[ \sum_ {l} \| f \left(g _ {l} ^ {T} \left(y ^ {T}\right)\right) - f \left(g _ {l} ^ {S} \left(y ^ {S}\right)\right) \| _ {2} ^ {2} \right] \tag {3} +$$ + +where $f$ represents the norm operation, $g_{l}^{T}$ and $g_{l}^{S}$ represent the feature maps of the $l$ -th layer in a predefined set of unet layers from the teacher and the student, respectively. With respective weights $\lambda_{FL}$ and $\lambda_{L}$ , the final KD objective is formalized as below: + +$$ +\mathcal {L} _ {K D} = \lambda_ {F L} \mathcal {L} _ {F L K D} + \lambda_ {L} \mathcal {L} _ {L K D} \tag {4} +$$ + +Adapter. Due to the dimension difference between English CLIP and non-English CLIP, we propose an adapter for simple feature transformation to align feature dimension. Moreover, the adapter is trained to learn language-specific information. The adapter is implemented by a MLP with only 6M parameters. To accelerate training and avoid destroying the generalization capability of the pretrained models, we freeze the CLIP encoder and image generator. As a result, the proposed PEA-Diffusion is both parameter- and computation-efficient because just the adapter is trained. + +![](images/c505fe5df1fc2348473de1214bea9d5e8b1ed6fdc267739f7d31aedb002d3f53.jpg) +Fig. 2: Image generation visualization of different models in Chinese specific language. + +# 3.3 Training Strategy and Final Objective + +For our goal of training a multi-language T2I generation model, we only need to train on English data since both the new CLIP text encoder and the teacher model's CLIP text encoder can accept English inputs. In this case, we can use aligned KD loss $\mathcal{L}_{KD}$ . For our goal of training a language-specific T2I generation model, we adopt a hybrid training strategy, for the target language's parallel corpus, we only need to use KD loss $\mathcal{L}_{KD}$ to align the distribution. On the other hand, we also have language-specific culture-related data $Y^{S^{\prime}}$ . Because there is a significant translation error when translating these data into English and the teacher model cannot provide high-quality knowledge guidance for this type of data, we only use $\mathcal{L}_{\mathrm{SD}}$ (1) for improving culture understanding ability. The final objective is formalized as follows: + +$$ +\mathcal {L} = \left\{ \begin{array}{l l} \lambda_ {F L} \mathbb {E} \Big [ \| \epsilon^ {T} (y ^ {T}) - \epsilon^ {S} (y) \| _ {2} ^ {2} \Big ] + \lambda_ {L} \mathbb {E} \Big [ \sum_ {l} \| f (g _ {l} ^ {T} (y ^ {T})) - f (g _ {l} ^ {S} (y)) \| _ {2} ^ {2} \Big ], & y \in Y ^ {S} \\ \mathbb {E} \Big [ \| \epsilon - \epsilon^ {S} (y) \| _ {2} ^ {2} \Big ], & y \in Y ^ {S ^ {\prime}} \end{array} \right. +$$ + +# 4 Experiments + +# 4.1 Data Preparation + +We download the Laion2B-multi [29] dataset along with English translations for image descriptions. Based on the open-source CLIP models available for different languages: Chinese CLIP, Russian CLIP, Japanese CLIP, Korean CLIP, and + +Italian CLIP, we determine to conduct experiments on the above five languages. For Chinese, we also collect WuKong [7] data with Chinese cultural image-text pairs and Chinese-English parallel corpus translations of Laion2B-en. For more data, model, and code details, please refer to the Appendix. + +# 4.2 Implementation Details and Evaluation + +Implementation details. We conduct all experiments on Stable Diffusion (SD)- [25], and Stable Diffusion XL (SDXL)- [21] like backbones. Take SDXL base model as an example. PEA-Diffusion consists of VAE, UNet, CLIP text encoder, and the adapter module, comprising 3 billion parameters, out of which a mere 6 million parameters ( $\approx 0.2\%$ ) are trainable. We set $\lambda_{FL}$ 0.1 and $\lambda_{L}$ 1 respectively based on the loss performance. The entire model is trained on 64 A100 GPUs for 70000 steps with a batch size of 5 per GPU. + +Test benchmark. To evaluate the capability of PEA-Diffusion to generate images and capture culture-specific concepts of different languages, we introduce two datasets: Multilingual-General (MG) for generation quality evaluation and Multilingual-Cultural (MC) for culture-specific concepts evaluation. We collect 200 prompts for MG, which covered content such as people, animals, landscapes, and fantasy themes. The descriptions include diverse styles, such as cartoon, anime, illustration, watercolor, and cyberpunk. The data comes from various major art websites where users provide real prompts. On the other hand, we also construct MC data including 200 prompts. The collection is mainly based on the ChatGPT tool and consists of culturally relevant content for each specific language, including but not limited to local cuisine, architecture, clothing, and distinctive cultural features. + +Evaluation metrics. For the general evaluation data MG, the widely used FID metric can only measure the feature distance between generated images and specific images. It often fails to provide accurate assessment for high-quality generated images, as emphasized and experimentally validated by SDXL [21] and other studies. Therefore, in order to evaluate the quality of generated images more accurately, we introduce two human preference models as our evaluation metrics: ImageReward [30] and PickScore [15]. Additionally, we continue to utilize OpenCLIP ViT-bigG to measure the similarity between images and text. As for the language-specific evaluation data MC, constructed for each language, to ensure a fair comparison of content relevance, we only consider the CLIP model trained in the specific language to calculate the correlation between generated images and their corresponding text. Since the general CLIP model can only process non-English data after translation into English, this process can introduce significant performance degradation. Furthermore, our paper's focus is precisely on addressing the errors caused by translating non-English prompts into English to generate images. Therefore, other metrics that involve translation processes, such as the two models based on human preference mentioned above, will no longer be valid for evaluation. + +Baseline methods. We generated image results in specific languages based on two T2I base models: SD1.5 and SDXL. We choose the following four categories of approaches as our baselines. + +- Translation: we apply Google translation API. +- AltDiffusion [33]: we test two multilingual SD models based on the open-source AltDiffusion framework: AltDiffusion-m9 and AltDiffusion-18. +- GlueGen [22]: We evaluate our model on three languages that overlap with the checkpoint provided by GlueGen, which are Chinese, Italian, and Japanese. For the other two languages, Russian and Korean, we implement GlueGen based on its code and our data. In addition, to further compare the impact of different language-specific CLIPs on the results, we also retrain the GlueGen based on five languages' CLIP text encoders. +- Fine-tuning with LoRA [12] and $W_{K}, W_{V}$ (key and value projection matrices): As low-cost experiments for comparison, we also fine-tuning the model to transfer language capabilities. + +# 4.3 Experimental Results + +Generation analysis. We validate our approach based on two different T2I base models. Firstly, to ensure a fair comparison with GlueGen and AltDiffusion, we train PEA-Diffusion based on the SD1.5 model. Secondly, to validate our low-cost experiment approach, we train a larger model based on SDXL. The compared methods are LoRA fine-tuning with rank 64 and $W_{K}$ , $W_{V}$ fine-tuning, both of which are parameter-efficient fine-tuning strategies. For the second approach, as proposed in [16], when updating the mapping from given text to image distribution, only updating $W_{K}^{(i)}$ , $W_{V}^{(i)}$ in each cross-attention block $i$ is sufficient since text features are the only input to the key and value projection matrix. Furthermore, to validate the wide application of the proposed approach with different text encoders, we train our model with a multi-language T5 encoder and a multi-language CLIP encoder. This type of training only requires English text data, which makes it even simpler and more efficient. Fig. 2 illustrates the comparisons between translation and our methods with different training strategies. It can be observed that our PEA-Diffusion can achieve Chinese culturally related image generation, but translation totally fails in these cases. + +As shown in the comprehensive results of Table 1, when compared with the translation method, our PEA-Diffusion outperforms it for almost all languages (except Japanese) in terms of CLIPScore for MC datasets. It is obvious that the translation method cannot capture the essence of different languages' cultures, leading to a semantic gap between the original prompts and the translated English prompts. However, our approach can use additional culturally related image-text pairs to train the adapter, which can align the language understanding ability of the text encoder with the generation ability of the UNet. Even though the UNet is frozen during the whole training process, we argue that our PEA-Diffusion can still stimulate the potential generation ability for cultural-related images of the original UNet. Since SD1.5 and SDXL are all trained on + +Table 1: CLIPScore with different specific languages. For Chinese, we evaluated two general evaluation metrics with MG data. $\dagger$ Indicates open source models. + +
Model BaseMethodszh-MCzh-MGru-MCit-MCko-MCja-MC
CLIPScore_zhImageRewardPickScoreCLIPScoreRuCLIPScore—itCLIPScore_koCLIPScore_ja
SD1.5Translate0.16860.24270.17990.49110.35560.19040.2885
Altdiffusion m9 [33]†0.2100-0.69660.12020.37150.31770.19060.2461
Altdiffusion m18 [33]†0.21720.33160.16480.47300.35020.19170.3008
GlueGen [22]†0.1700-0.51870.12240.41840.30650.17690.2506
GlueGen [22]0.2153-0.46850.12560.39870.29280.18430.2465
Ours0.25240.09850.17010.50070.37280.18960.2501
SDXLTranslate0.23831.36470.24020.48440.35910.20330.2958
LoRA0.23660.12170.15980.46560.34760.20340.1940
Wkand Wv0.24850.14130.16210.47850.34950.20880.2022
Ours w/ MuTI50.23010.82540.20010.45680.38050.19080.2996
Ours w/ Multilingual-CLIP0.24071.21220.23240.46730.38560.21570.3007
Ours0.26101.02450.22440.50120.41910.21940.2544
+ +billions of image-text pairs that contain wide image domains, our approach can successfully awaken their generative abilities in different cultural contexts. In terms of ImageReward and PickScore for the MG dataset, we must admit that PEA-Diffusion is slightly inferior to translation methods. This is because using other text encoders (e.g., Chinese CLIP and Multilingual CLIP) cannot ensure complete alignment between the text embedding and the UNet feature spaces. + +Even though Altdiffusion retrans the multilingual CLIP text encoder and UNet on billions of image-text pairs, it still performs worse than our approach in terms of almost all metrics for all languages. GlueGen is the most related work on our concerned topic. The statistics presented in Table 1 reveal that both checkpoints provided by the authors or re-implemented by ourselves are significantly inferior to our PEA-Diffusion, especially for the ImageReward and PickScore metrics, resulting in very low generation quality since only aligning distributions of text embedding for different languages is very hard, even with its sophisticatedly designed GlenNet. In addition, it cannot generate culturally related images. Therefore, it's worth mentioning that our core contribution is that we align the model's generation ability through UNet distillation only by a MLP-like adapter between different languages. We don't forcibly align the text embedding spaces of different languages. + +Additionally, we undertake various parameter-efficient fine-tuning strategies for comparison. The fine-tuning techniques, including LoRA and $W_{k}$ and $W_{v}$ , show subpar performance based on ImageReward, PickScore, and CLIPScore metrics. This is attributed to the fact that these fine-tuning methods depend significantly on high-quality image-text datasets, whereas our PEA-Diffusion requires merely some parallel multilingual texts and images and lacks stringent quality requirements. Further experiments can be found in the Appendix. + +Analysis on different languages. The data in Table 1 indicate that the CLIPScores for Russian, Italian, and Korean either match or exceed those of other baseline methods by a notable margin. For Japanese, however, the CLIP-Score is slightly lower compared to translation and AltDiffusion. We ascribe this to the CLIP model's limited encoding ability for Japanese, which fails to align with the SD model via the adapter layer. Additionally, with only 1.6M training samples for Korean, our method still achieves effective language transfer, under- + +scoring that large-scale training datasets are not essential. Further visualization examples of culturally-related prompts can be found in the Appendix. + +Analysis of training cost. Our PEADiffusion freezes the CLIP text encoder, the UNet, and VAE during the whole training process, only the MLP mapping adapter is trained, accounting for merely 6M training parameters, which is $0.2\%$ of total parameters for SDXL-based models. We employed 64 A100 GPUs to train 70000 steps, resulting in 1600 GPU hours. As a comparison, AltDiffusion [33] trained the model for three stages with nearly 30000 GPU hours, which is $\approx 20$ times our training cost. In addition, our method achieves higher performance in terms of almost all metrics and the five experimental languages. We also re-implemented and trained GlueGen [22] with the same dataset + +![](images/ba27d4e8ef70fe6d6374aa4a2b3472a4c8d8498df69bea3a856dd84986fab0e4.jpg) +Fig.3: Training parameters(M) and cost comparison for different methods. + +constructed by ourselves. It indeed achieves very low training costs with only 1 A100 GPU for 24 hours, but results in very poor performance no matter for image quality (-0.5 ImageReward score) or language transferring ability. Finally, we also conduct other parameter-efficient fine-tuning strategies (e.g., LoRA and tuning only $W_{k}$ and $W_{v}$ matrices of attention layers) and achieve much lower performance and higher training costs, which demonstrated the effectiveness and efficiency of our PEA-Diffusion. Figure 3 only shows the relationship between training parameters and GPU hours of different methods. For more information on model fitting and training data, as well as adapter module parameter variations, please refer to Appendix. + +# 4.4 Ablation Studies + +Since Chinese data has certain advantages in terms of data quality and quantity, to eliminate errors caused by data, our ablation experiments are all based on Chinese data. The experiments include four parts: First, we conduct a baseline experiment where we only perform standard model fine-tuning without incorporating any KD strategies. Secondly, inspired by GlueGen [22], we employ text encoder KD to explore the impact of distillation at different positions in the SD framework. Next, we perform distillation based on the UNet model, including feature distillation between different UNet layers and logits distillation for the final output. Finally, we combine all of the distillation positions described above. + +Effectiveness of KD. First, KD is crucial for guiding models onto the right learning trajectory for rapid convergence and is vital for the PEA-Diffusion method. As indicated in the first row of Table 2, the standalone fine-tuning method yields notably inferior performance in both image quality (ImageReward and PickScore) and image-text alignment (CLIPScores). + +Table 2: Ablation results for different knowledge strategies. + +
EvaluationImageRewardPickScoreCLIPScore_enCLIPScore_zh
Fine-tuning-1.31540.18540.32780.1901
Fine-tuning w/ TKD-0.98750.19970.35060.2287
Fine-tuning w/ ULKD-1.04470.19570.33310.1968
Fine-tuning w/ UFKD0.95560.21740.40850.2557
Fine-tuning w/ UTKD0.84900.21130.40110.2461
Ours(Fine-tuning w/ UKD)1.02450.22440.42240.2610
+ +Effectiveness of UNet Feature KD (UFKD). Our KD method consists of two parts: the first part involves distilling the logits results generated by the UNet (ULKD), we then introduce a distillation of intermediate layer features in the UNet, which serves as the primary knowledge guidance mechanism. As shown in the third and fourth lines of Table 2, we can observe that logit distillation can only improve marginally when compared with direct fine-tuning whereas feature distillation boosts the performance remarkably. We attribute this phenomenon to the generation ability of the large amount of UNet parameters, so that intermediate feature supervision is essential for knowledge transfer. + +Effectiveness of UNet KD. Furthermore, different distillation positions have different effects. Our initial focus is on aligning the feature spaces of different text encoders, so a natural idea is to distill only the feature information of the teacher model's text encoder (TKD). However, as the results show in the second line of Table 2, although the overall performance exceeds the pure fine-tuning method, it is still far from our best performing approach. This illustrates that simply aligning the text encoder is insufficient and ineffective. This statistic also matches the experimental results of GlueGen in Table 1: simply aligning text embedding spaces cannot obtain satisfactory image generation results. Therefore, this further proves the necessity of distilling the features of the UNet itself. + +Finally, we perform KD on both the UNet and text encoder (UTKD). However, the fifth line of Table 2 shows that its effect is worse than performing only UNet distillation. We suspect that the addition of text encoder KD actually weakens the learning of the SD core UNet structure. The main knowledge ability for generating images is mainly influenced by the UNet model, which is responsible for the adapter alignment feature in PEA-Diffusion. More ablation experiments can refer to Appendix. + +# 4.5 Exploring the Domain Adaptation of PEA-Diffusion + +To demonstrate the domain adaptation of our model with minimal data guidance, we first constructed a dataset of 20,000 parallel image and text pairs in the domain of human. We then trained the models based on the models fine-tuned $W_{K}^{(i)}, W_{V}^{(i)}$ mentioned in section 4.3 with UKD, and evaluated on the MG and MC dataset. The results on Chinese are shown Table 3. + +We observed significant improvements in the three metrics on the MG test data. This suggests that PEA can be guided by a small amount of data to perform out-of-domain knowledge transfer. + +Table 3: Domain adaptation with 20,000 data in Chinese specific language. + +
EvaluationImageRewardPickScoreCLIPScore_enCLIPScore_zh
Wk and Wv0.14130.16210.36740.2485
Ours(Fine-tuning w/ UKD)0.83680.21450.40980.2501
+ +# 4.6 Adaptability Experiment of English Downstream Model + +![](images/653bc4464f408a5eae3cef29b77f50385c27afa055311bbcb173256cb65d62d3.jpg) +sal_scenery LoRA +西湖，塔和瀑布，日出(West Lake, snow and waterfall, sunrise) +江南，夏日湖畔的一村村庄(Jianxun, a village by the summer lake) + +![](images/e70f2a98840d160713a31812f6776e65287d45cc5c09fab109982deee3c7df53.jpg) +realistic/easy/animalism/3d-carton Fine-tuning +一个漂亮的双面的调色板 +一个白色的小球与white painting light) +3D电影，4K，高度细致，男人坐在马桶上读报（3D movie, 4K, highly detailed, man sitting on toilet + +![](images/a8089f7a1a6d5ab98bdc9585dbecd91621a231660164c9a4e348c3f4485139f5.jpg) + +![](images/92eafbe56876f3ab5d6e6e24da6cbffc31a80436d09be4bf5ec0c423707e123e.jpg) +SD-XL, Inpainting 0.1 Painting +2018年，由青岛双星玻璃有限公司（White porcelain plates on benches）一辆小型起重机放在了长椅上（A small crane was placed on the bench） + +![](images/b908fc72ce8f83e42fb39bf7c840e59a686cc5293d871901e0a81208305dabf5.jpg) +controlnet-canny ControlNet 鸳鸯双栖展双飞，满园存色惹人醉(Mandarin ducks and butterflies fly together, and the spring scenery in the garden is intoxicating); 北国风光，千里冰封，万里雪飘(Northern scenery, thousands of miles of ice, thousands of miles of snow drifting) +Icm-lora-adsl CDM 带着墨镜的猫咪手里拿着剑，在恶魔城怪里，仙鸽奇侠风格(The cat wearing sunglasses holds a sword in its hand, in the Demon Castle, and the Magic Knight) 红色灯笼，长城，秋天，高清，奇幻风格，史密德(Red Lutron, Great Wall, Autumn, HD, Fantasy Style, Epic Sense) +Fig. 4: Downstream task experiments of our proposed PEA. By only adding the trained adapter between the Chinese CLIP text encoder and the UNet, it can achieve language transfer for fine-tuning, LoRA, ControlNet, Inpainting, LCM and compressed models. + +![](images/d901f8901fab1ae63dab5501bb89c8203b5b262fded9199baef354028fd4479f.jpg) +SSD-B I SSD +一只超级可爱兔子穿着僧侣服装，肖像猫，皮克斯动画(A super cute rabbit wearing monk clothing, partition)。 +中国汉奴女孩，阳光明媚，站在踏青花中；近景，摄影风景（Chinese Hanafi girl, sunny, standing in the midst of ruffled jeans, close-up, photography style） + +In this section, we would like to emphasize the plug-and-play feature of our PEA module. Currently, a wide range of community resources, such as fine-tuned models, LoRA, Dreambooth, ControlNet, compressed SD models, and accelerated sampling SD models, are derived from the English SD models, so that the English community is much more mature than other languages. Therefore, one key challenge we must overcome is bridging the compatibility gap between the new model and the English SD community after language transfer. Since our approach uses feature-level and logit-level KD to train only the PEA module while keeping the core UNet parameters unchanged, it naturally maintains compatibility with the English SD community and can seamlessly integrate with various English SD community plugins. Interestingly, the PEA module remains plug-and-play for models that have undergone changes in base SD UNet parameters in contrast to PEA-Diffusion but whose overall structure remains similar. This includes models such as fine-tuned SD and compressed SD trained using + +pruning and distillation techniques, among others. In the following, we will introduce several downstream image generation applications based on the SD or SDXL foundation model. + +Fine-tuned checkpoint. We use two fine-tuned SDXL-based model checkpoints: "xxmix9realistic" and "samaritan-3d-cartoon". The "xxmix9realistic" model is trained on person, photorealistic, sexy, female, woman, girls, etc. The "samaritan-3d-cartoon" model is trained on anime, Disney, 3D, Pixar, etc. + +LoRA. "csal_scenery" is a SDXL-based LoRA model that contains common ancient Chinese architectural forms such as pavilions, pagodas, towers, gardens, courtyards, bridges, temples, palaces, towns, etc. + +ControlNet. "controlnet-canny" is trained on SDXL under canny condition. + +Inpainting. SDXL Inpainting 0.1 is a latent T2I diffusion model capable of generating photorealistic images given any text input, with the extra capability of Inpainting the pictures by using a mask. + +LCM. Latent Consistency Models [19] have achieved impressive performance in accelerating T2I generative tasks, producing high-quality images with minimal inference steps. "lcm-lora-sdxl" is an acceleration module based on SDXL. + +SSD. The Segmind Stable Diffusion Model is a condensed version of the SDXL that is $50\%$ smaller in size, delivering a $60\%$ increase in speed while preserving high-quality text-to-image (T2I) generation capabilities. + +In Figure 4, we have shown that the PEA module can be used as a plugin between the Chinese CLIP text encoder and the UNet, achieving the ability to transfer language from English to Chinese directly. For more details of the abovementioned checkpoints and generated images, please refer to our Appendix. + +# 5 Conclusion and Limitation + +Conclusion. The high training expenses and limited training data available for multilingual T2I models have driven us to search for a more lightweight solution. In this paper, we introduce a straightforward and efficient method using KD and adapters. By focusing solely on training the adapter and keeping other parameters fixed, our approach harnesses the capability of T2I models for multilingual cultural comprehension and generation. Our method not only achieves significantly better results for culturally relevant data but also surpasses existing models while maintaining general synthesis capability. Comprehensive experiments demonstrate that our method is cost-effective, versatile, and easy to integrate. + +**Limitation.** First, the effectiveness of the CLIP text encoder significantly influences both the quality of image generation and text relevance. Utilizing a weaker language-specific CLIP could impact the performance of PEA-Diffusion in broad open domains and may slightly compromise the model's generalization abilities. Second, the upper limit of our method is constrained by the capabilities of the English base model, indicating that our approach cannot exceed its inherent potential. Moving forward, we aim to conduct additional research to mitigate these limitations. + +# References + +1. Aggarwal, P., Kale, A.: Towards zero-shot cross-lingual image retrieval. arXiv preprint arXiv:2012.05107 (2020) +2. Carlsson, F., Eisen, P., Rekathati, F., Sahlgren, M.: Cross-lingual and multilingual clip. In: Proceedings of the Thirteenth Language Resources and Evaluation Conference. pp. 6848-6854 (2022) +3. Chen, G., Hou, L., Chen, Y., Dai, W., Shang, L., Jiang, X., Liu, Q., Pan, J., Wang, W.: mclip: Multilingual clip via cross-lingual transfer. In: Proceedings of the 61st Annual Meeting of the Association for Computational Linguistics (Volume 1: Long Papers). pp. 13028-13043 (2023) +4. Chen, Z., Liu, G., Zhang, B.W., Ye, F., Yang, Q., Wu, L.: Altclip: Altering the language encoder in clip for extended language capabilities (2022) +5. Conneau, A., Khandelwal, K., Goyal, N., Chaudhary, V., Wenzek, G., Guzmán, F., Grave, E., Ott, M., Zettlemoyer, L., Stoyanov, V.: Unsupervised cross-lingual representation learning at scale. arXiv preprint arXiv:1911.02116 (2019) +6. Ding, M., Yang, Z., Hong, W., Zheng, W., Zhou, C., Yin, D., Lin, J., Zou, X., Shao, Z., Yang, H., et al.: Cogview: Mastering text-to-image generation via transformers. Advances in Neural Information Processing Systems 34, 19822-19835 (2021) +7. Gu, J., Meng, X., Lu, G., Hou, L., Minzhe, N., Liang, X., Yao, L., Huang, R., Zhang, W., Jiang, X., et al.: Wukong: A 100 million large-scale chinese cross-modal pre-training benchmark. Advances in Neural Information Processing Systems 35, 26418-26431 (2022) +8. Gu, S., Chen, D., Bao, J., Wen, F., Zhang, B., Chen, D., Yuan, L., Guo, B.: Vector quantized diffusion model for text-to-image synthesis. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 10696-10706 (2022) +9. Hinton, G., Vinyals, O., Dean, J.: Distilling the knowledge in a neural network. arXiv preprint arXiv:1503.02531 (2015) +10. Ho, J., Jain, A., Abbeel, P.: Denoising diffusion probabilistic models. Advances in Neural Information Processing systems 33, 6840-6851 (2020) +1. Hou, L., Huang, Z., Shang, L., Jiang, X., Chen, X., Liu, Q.: Dynabert: Dynamic bert with adaptive width and depth. Advances in Neural Information Processing Systems 33, 9782-9793 (2020) +2. Hu, E.J., Shen, Y., Wallis, P., Allen-Zhu, Z., Li, Y., Wang, S., Wang, L., Chen, W.: Lora: Low-rank adaptation of large language models (2021) +3. Hu, J., Han, X., Yi, X., Chen, Y., Li, W., Liu, Z., Sun, M.: Efficient crosslingual transfer for chinese stable diffusion with images as pivots. arXiv preprint arXiv:2305.11540 (2023) +4. Jiao, X., Yin, Y., Shang, L., Jiang, X., Chen, X., Li, L., Wang, F., Liu, Q.: Tinybert: Distilling bert for natural language understanding. arXiv preprint arXiv:1909.10351 (2019) +5. Kirstain, Y., Polyak, A., Singer, U., Matiana, S., Penna, J., Levy, O.: Pick-a-pic: An open dataset of user preferences for text-to-image generation (2023) +6. Kumari, N., Zhang, B., Zhang, R., Shechtman, E., Zhu, J.Y.: Multi-concept customization of text-to-image diffusion. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 1931-1941 (2023) +7. Li, Y., Chang, C.Y., Rawls, S., Vulic, I., Korhonen, A.: Translation-enhanced multilingual text-to-image generation. arXiv preprint arXiv:2305.19216 (2023) + +18. Liu, S., Leng, D., Yin, Y.: Bridge diffusion model: bridge non-english language-native text-to-image diffusion model with english communities (2023) +19. Luo, S., Tan, Y., Huang, L., Li, J., Zhao, H.: Latent consistency models: Synthesizing high-resolution images with few-step inference (2023) +20. Nichol, A., Dhariwal, P., Ramesh, A., Shyam, P., Mishkin, P., McGrew, B., Sutskever, I., Chen, M.: Glide: Towards photorealistic image generation and editing with text-guided diffusion models. arXiv preprint arXiv:2112.10741 (2021) +21. Podell, D., English, Z., Lacey, K., Blattmann, A., Dockhorn, T., Müller, J., Penna, J., Rombach, R.: Sdxl: Improving latent diffusion models for high-resolution image synthesis (2023) +22. Qin, C., Yu, N., Xing, C., Zhang, S., Chen, Z., Ermon, S., Fu, Y., Xiong, C., Xu, R.: Gluegen: Plug and play multi-modal encoders for x-to-image generation. arXiv preprint arXiv:2303.10056 (2023) +23. Radford, A., Kim, J.W., Hallacy, C., Ramesh, A., Goh, G., Agarwal, S., Sastry, G., Askell, A., Mishkin, P., Clark, J., et al.: Learning transferable visual models from natural language supervision. In: International conference on machine learning. pp. 8748-8763. PMLR (2021) +24. Ramesh, A., Dhariwal, P., Nichol, A., Chu, C., Chen, M.: Hierarchical text-conditional image generation with clip latents (2022) +25. Rombach, R., Blattmann, A., Lorenz, D., Esser, P., Ommer, B.: High-resolution image synthesis with latent diffusion models. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 10684-10695 (2022) +26. Rombach, R., Blattmann, A., Lorenz, D., Esser, P., Ommer, B.: High-resolution image synthesis with latent diffusion models (2022) +27. Ronneberger, O., Fischer, P., Brox, T.: U-net: Convolutional networks for biomedical image segmentation. In: Medical Image Computing and Computer-Assisted Intervention-MICCAI 2015: 18th International Conference, Munich, Germany, October 5-9, 2015, Proceedings, Part III 18. pp. 234-241. Springer (2015) +28. Saharia, C., Chan, W., Saxena, S., Li, L., Whang, J., Denton, E., Ghasemipour, S.K.S., Ayan, B.K., Mahdavi, S.S., Lopes, R.G., Salimans, T., Ho, J., Fleet, D.J., Norouzi, M.: Photorealistic text-to-image diffusion models with deep language understanding (2022) +29. Schuhmann, C., Beaumont, R., Vencu, R., Gordon, C., Wightman, R., Cherti, M., Coombes, T., Katta, A., Mullis, C., Wortsman, M., Schramowski, P., Kundurthy, S., Crowson, K., Schmidt, L., Kaczmarczyk, R., Jitsev, J.: Laion-5b: An open large-scale dataset for training next generation image-text models (2022) +30. Xu, J., Liu, X., Wu, Y., Tong, Y., Li, Q., Ding, M., Tang, J., Dong, Y.: Imagereward: Learning and evaluating human preferences for text-to-image generation (2023) +31. Yang, A., Pan, J., Lin, J., Men, R., Zhang, Y., Zhou, J., Zhou, C.: Chinese clip: Contrastive vision-language pretraining in chinese. arXiv preprint arXiv:2211.01335 (2022) +32. Yang, Z., Li, Z., Shao, M., Shi, D., Yuan, Z., Yuan, C.: Masked generative distillation. In: European Conference on Computer Vision. pp. 53-69. Springer (2022) +33. Ye, F., Liu, G., Wu, X., Wu, L.: Altdiffusion: A multilingual text-to-image diffusion model (2023) +34. Zhang, H., Yin, W., Fang, Y., Li, L., Duan, B., Wu, Z., Sun, Y., Tian, H., Wu, H., Wang, H.: Ernie-vilg: Unified generative pre-training for bidirectional vision-language generation. arXiv preprint arXiv:2112.15283 (2021) + +35. Zhang, J., Gan, R., Wang, J., Zhang, Y., Zhang, L., Yang, P., Gao, X., Wu, Z., Dong, X., He, J., Zhuo, J., Yang, Q., Huang, Y., Li, X., Wu, Y., Lu, J., Zhu, X., Chen, W., Han, T., Pan, K., Wang, R., Wang, H., Wu, X., Zeng, Z., Chen, C.: Fengshenbang 1.0: Being the foundation of chinese cognitive intelligence. CoRR abs/2209.02970 (2022) +36. Zhang, J., Gan, R., Wang, J., Zhang, Y., Zhang, L., Yang, P., Gao, X., Wu, Z., Dong, X., He, J., et al.: Fengshenbang 1.0: Being the foundation of chinese cognitive intelligence. arXiv preprint arXiv:2209.02970 (2022) \ No newline at end of file diff --git a/2024/PEA-Diffusion_ Parameter-Efficient Adapter with Knowledge Distillation in non-English Text-to-Image Generation/images.zip b/2024/PEA-Diffusion_ Parameter-Efficient Adapter with Knowledge Distillation in non-English Text-to-Image Generation/images.zip new file mode 100644 index 0000000000000000000000000000000000000000..2d0e8b5f555c0c817921f5666d5caa07f8ab3fa4 --- /dev/null +++ b/2024/PEA-Diffusion_ Parameter-Efficient Adapter with Knowledge Distillation in non-English Text-to-Image Generation/images.zip @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:dc67c90b05fbee4329d30e9cbe76515b75c124118a5a6c135a08e1cb0d255411 +size 373880 diff --git a/2024/PEA-Diffusion_ Parameter-Efficient Adapter with Knowledge Distillation in non-English Text-to-Image Generation/layout.json b/2024/PEA-Diffusion_ Parameter-Efficient Adapter with Knowledge Distillation in non-English Text-to-Image Generation/layout.json new file mode 100644 index 0000000000000000000000000000000000000000..9756d92db22228c9d1b6641135cebc4f08539fe7 --- /dev/null +++ b/2024/PEA-Diffusion_ Parameter-Efficient Adapter with Knowledge Distillation in non-English Text-to-Image Generation/layout.json @@ -0,0 +1,7748 @@ +{ + "pdf_info": [ + { + "para_blocks": [ + { + "bbox": [ + 135, + 111, + 479, + 166 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 135, + 111, + 479, + 166 + ], + "spans": [ + { + "bbox": [ + 135, + 111, + 479, + 166 + ], + "type": "text", + "content": "PEA-Diffusion: Parameter-Efficient Adapter with Knowledge Distillation in non-English Text-to-Image Generation" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 173, + 186, + 441, + 199 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 173, + 186, + 441, + 199 + ], + "spans": [ + { + "bbox": [ + 173, + 186, + 441, + 199 + ], + "type": "text", + "content": "Jian Ma, Chen Chen†, Qingsong Xie, and Haonan Lu†" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 193, + 208, + 421, + 231 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 193, + 208, + 421, + 231 + ], + "spans": [ + { + "bbox": [ + 193, + 208, + 421, + 231 + ], + "type": "text", + "content": "OPPO AI Center \n{majian2, chenchen4, xieqingsong1, luhaonan} @oppo.com" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 160, + 255, + 452, + 452 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 160, + 255, + 452, + 452 + ], + "spans": [ + { + "bbox": [ + 160, + 255, + 452, + 452 + ], + "type": "text", + "content": "Abstract. Text-to-image diffusion models are well known for their ability to generate realistic images based on textual prompts. However, the existing works have predominantly focused on English, lacking support for non-English text-to-image models. The most commonly used translation methods cannot solve the generation problem related to language culture, while training from scratch on a specific language dataset is prohibitively expensive. In this paper, we are inspired to propose a simple plug-and-play language transfer method based on knowledge distillation. All we need to do is train a lightweight MLP-like parameter-efficient adapter (PEA) with only 6M parameters under teacher knowledge distillation along with a small parallel data corpus. We are surprised to find that freezing the parameters of UNet can still achieve remarkable performance on the language-specific prompt evaluation set, demonstrating that PEA can stimulate the potential generation ability of the original UNet. Additionally, it closely approaches the performance of the English text-to-image model on a general prompt evaluation set. Furthermore, our adapter can be used as a plugin to achieve significant results in downstream tasks in cross-lingual text-to-image generation." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 160, + 463, + 435, + 474 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 160, + 463, + 435, + 474 + ], + "spans": [ + { + "bbox": [ + 160, + 463, + 435, + 474 + ], + "type": "text", + "content": "Keywords: Diffusion models " + }, + { + "bbox": [ + 160, + 463, + 435, + 474 + ], + "type": "inline_equation", + "content": "\\cdot" + }, + { + "bbox": [ + 160, + 463, + 435, + 474 + ], + "type": "text", + "content": " Parameter-efficient " + }, + { + "bbox": [ + 160, + 463, + 435, + 474 + ], + "type": "inline_equation", + "content": "\\cdot" + }, + { + "bbox": [ + 160, + 463, + 435, + 474 + ], + "type": "text", + "content": " Plug-and-play" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 132, + 492, + 230, + 506 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 492, + 230, + 506 + ], + "spans": [ + { + "bbox": [ + 132, + 492, + 230, + 506 + ], + "type": "text", + "content": "1 Introduction" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 130, + 515, + 482, + 647 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 515, + 482, + 647 + ], + "spans": [ + { + "bbox": [ + 130, + 515, + 482, + 647 + ], + "type": "text", + "content": "Recent text-to-image (T2I) models, such as DALL-E2 [24], Imagen [28], and Stable Diffusion (SD) [26], have opened up a new era of AI painting due to their ability to generate realistic, photo-level, and creative images. Numerous downstream tasks have further driven the iteration and update of this technology, as well as the activity of the community, including image editing, controllable generation, etc. However, mainstream text-to-image models only support English prompts, which means non-English speakers usually rely on translations to use them. This may cause translation errors and information loss due to culture-specific concepts. For example, one famous Chinese dish, \"Hongshao shizi tou\", may be translated to \"Braised lion head\" in English. Additionally, the introduction of large translation networks leads to high computation consumption." + } + ] + } + ], + "index": 6 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 142, + 653, + 243, + 665 + ], + "type": "page_footnote", + "angle": 0, + "lines": [ + { + "bbox": [ + 142, + 653, + 243, + 665 + ], + "spans": [ + { + "bbox": [ + 142, + 653, + 243, + 665 + ], + "type": "text", + "content": "†Corresponding authors." + } + ] + } + ], + "index": 7 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 0 + }, + { + "para_blocks": [ + { + "bbox": [ + 130, + 116, + 482, + 308 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 116, + 482, + 308 + ], + "spans": [ + { + "bbox": [ + 130, + 116, + 482, + 308 + ], + "type": "text", + "content": "To address this issue, recent researchers have created T2I generative models that process native languages directly, bypassing any need for translation. Taiyi-Bilingual [35] incorporates a novel text encoder initialized with Chinese CLIP checkpoints, focusing only on training the text encoder. AltDiffusion [33] introduces a multi-stage training process for multilingual T2I generation, which is relatively costly. Some experts begin training T2I models from scratch using different language data. For example, ERNIE-ViLG 2.0 [6] is trained with native Chinese content and can generate high-quality images that reflect Chinese cultural elements. However, this approach is very expensive and often beyond the reach of many research facilities. Additionally, this method does not foster a community impact, necessitating the retraining of English models for various downstream or commonly used community tasks. GlueGen [22] introduces a GlueNet module to align the input features of the original image generator by adding a new encoder. Despite its low training cost, its effectiveness in generating culturally relevant images in multiple languages and handling broad prompts is comparatively limited." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 130, + 313, + 482, + 577 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 313, + 482, + 577 + ], + "spans": [ + { + "bbox": [ + 130, + 313, + 482, + 577 + ], + "type": "text", + "content": "As such, exploring a parameter-efficient plug-and-play method that adapts the publicly available, English-naive, pretrained T2I models to suit non-English and culture-specific concepts is becoming more and more urgent. Consequently, we concentrate on a critical research question (RQ): Is it possible to devise a low-cost method to acquire a non-English T2I model by adapting the existing pretrained T2I models efficiently? To this end, we take SD as the baseline, as it has shown surprisingly good results on T2I tasks. To finetune SD with minimal cost, we propose a novel parameter-efficient adapter (PEA) method with knowledge distillation (KD) to make full use of the pretrained SD model, referred to as PEA-Diffusion. To take advantage of the pre-trained SD model directly, the pre-trained SD is used as a teacher to guide student learning. Specifically, we propose the PEA module to align the representation space of the new condition encoder in the student with that of the teacher's image generator. PEA merely consists of a light-weight multi-layer perception (MLP), greatly simplifying the model construction process and preserving the image generation capabilities of the original SD model. Different from GlueNet [22] which forces text feature alignment of two text encoders at encoder space, we propose KD loss to compel representation alignment at feature maps and logits space from the UNet module between our PEA-Diffusion and pretrained English SD models. We argue that our method boosts the image generator to understand the information from the new text encoder. To avoid catastrophic forgetting and reduce trainable parameters, UNet parameters are kept frozen during the entire training procedure." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 130, + 582, + 482, + 665 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 582, + 482, + 665 + ], + "spans": [ + { + "bbox": [ + 130, + 582, + 482, + 665 + ], + "type": "text", + "content": "Take Chinese SD as an example. We conduct a thorough experimental analysis on KD to help understand the effect of where to perform feature alignment between the teacher and the student. In addition, we provided intuitive analysis about what parameters should be updated, facilitating discovering the impact of the SD parameters for adapting SD into language-specific T2I models. It should be noted that the method could be applicable for any other non-English T2I generation model. Our core contributions are summarized as follows:" + } + ] + } + ], + "index": 4 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "spans": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "type": "text", + "content": "2" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 166, + 91, + 215, + 101 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 166, + 91, + 215, + 101 + ], + "spans": [ + { + "bbox": [ + 166, + 91, + 215, + 101 + ], + "type": "text", + "content": "J. Ma et al." + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 1 + }, + { + "para_blocks": [ + { + "bbox": [ + 137, + 116, + 481, + 249 + ], + "type": "list", + "angle": 0, + "index": 5, + "blocks": [ + { + "bbox": [ + 137, + 116, + 481, + 163 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 137, + 116, + 481, + 163 + ], + "spans": [ + { + "bbox": [ + 137, + 116, + 481, + 163 + ], + "type": "text", + "content": "- We propose a novel PEA-Diffusion model to resolve the T2I model's English-native bias, supporting non-English prompts. The PEA aligns the feature and output spaces between the source and target UNet for bilingual prompts, leading to extremely low training costs." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 138, + 164, + 481, + 201 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 164, + 481, + 201 + ], + "spans": [ + { + "bbox": [ + 138, + 164, + 481, + 201 + ], + "type": "text", + "content": "- PEA module is plug-and-play, it can be directly applied to various T2I tasks, including LoRA, ControlNet, Inpainting, compressed SD models that have been pruned and models accelerated through various strategies." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 138, + 201, + 481, + 249 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 201, + 481, + 249 + ], + "spans": [ + { + "bbox": [ + 138, + 201, + 481, + 249 + ], + "type": "text", + "content": "- PEA-Diffusion can achieve remarkable image generation performance with culture-specific prompts compared with other cross-lingual or multilingual T2I models. It can also retain the general generation ability of the original English diffusion model with slight degradation on evaluation metrics." + } + ] + } + ], + "index": 4 + } + ], + "sub_type": "text" + }, + { + "bbox": [ + 132, + 267, + 240, + 280 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 267, + 240, + 280 + ], + "spans": [ + { + "bbox": [ + 132, + 267, + 240, + 280 + ], + "type": "text", + "content": "2 Related Work." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 132, + 294, + 356, + 306 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 294, + 356, + 306 + ], + "spans": [ + { + "bbox": [ + 132, + 294, + 356, + 306 + ], + "type": "text", + "content": "2.1 Multilingual Text-to-image Generation" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 130, + 315, + 482, + 579 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 315, + 482, + 579 + ], + "spans": [ + { + "bbox": [ + 130, + 315, + 482, + 579 + ], + "type": "text", + "content": "Diffusion models have shown a remarkable capability of yielding photorealistic and diverse images given textual descriptions, such as GLIDE [20], Stable Diffusion [25], Imagen [28], DALL-E2 [24], and VQ-Diffusion [8]. Despite their powerful generation abilities, they still only understand English input, leading to limitations. To support non-English text prompt, Taiyi [36] finetunes the Chinese text encoder with a diffusion model. ERNIE-ViLG [34] trains the T2I models from scratch. AltDiffusion [33] proposes a multilingual text encoder to capture multiple language features, which serve as DM conditions. The model was trained in three stages: text encoder fine-tuning, " + }, + { + "bbox": [ + 130, + 315, + 482, + 579 + ], + "type": "inline_equation", + "content": "W_{K}, W_{V}" + }, + { + "bbox": [ + 130, + 315, + 482, + 579 + ], + "type": "text", + "content": " fine-tuning, and full fine-tuning. GlueGen [22] has designed a GlueNet module that aligns the original conditional input features of the Image Generator by introducing a new encoder. ENSAD [17] proposes a translation-enhanced multilingual T2I generation model by ensembling multilingual T2I and neural machine translation. BDM [18] proposes a backbone-branch network architecture by integrating an English T2I backbone with a Chinese native semantics injection branch. Despite their success in the non-English I2I generation, these works are either constrained to deal with distinctive language concepts or isolated from English-native T2I communities. At the same time, it is not possible to balance the lower cost of training with the performance of evaluation sets. IAP [13] proposes to treat images as pivots to establish the connection between the embedding space of different text encoders, but it doesn't consider to enhance cultural-related generation ability and expand to downstream tasks." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 132, + 597, + 256, + 609 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 597, + 256, + 609 + ], + "spans": [ + { + "bbox": [ + 132, + 597, + 256, + 609 + ], + "type": "text", + "content": "2.2 Multilingual CLIP." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 130, + 617, + 482, + 666 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 617, + 482, + 666 + ], + "spans": [ + { + "bbox": [ + 130, + 617, + 482, + 666 + ], + "type": "text", + "content": "CLIP uses contrastive language-image pretraining to connect text and image representations, based on massive English text-image sample pairs [23]. Recently, some works have applied teacher learning [2] or metric learning [1] to train multilingual CLIP models, expanding CLIP to other languages. AltCLIP [4] takes this" + } + ] + } + ], + "index": 10 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 389, + 91, + 447, + 101 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 389, + 91, + 447, + 101 + ], + "spans": [ + { + "bbox": [ + 389, + 91, + 447, + 101 + ], + "type": "text", + "content": "PEA-diffusion" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 474, + 92, + 481, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 474, + 92, + 481, + 100 + ], + "spans": [ + { + "bbox": [ + 474, + 92, + 481, + 100 + ], + "type": "text", + "content": "3" + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 2 + }, + { + "para_blocks": [ + { + "bbox": [ + 130, + 116, + 482, + 248 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 116, + 482, + 248 + ], + "spans": [ + { + "bbox": [ + 130, + 116, + 482, + 248 + ], + "type": "text", + "content": "one step further enhancing the language capabilities, which is implemented by teacher learning and contrastive learning by leveraging a pretrained multilingual text encoder, XLMR [5]. mCLIP [3] achieves multilingual CLIP by aligning the CLIP text model and a multilingual text encoder. To effectively transfer CLIP to language-specific scenarios, some researchers train language-specific CLIPs on large-scale native image-text pairs, e.g., Chinese CLIP [31]. In order to promote PEA-Diffusion to effectively understand language-specific concepts, we append the PEA module after the language-specific CLIP or Multilingual CLIP and train it to uncover non-English cultural concepts within the T2I. The goal is to establish a bridge between the new CLIP and T2I core structures, such as UNet using the PEA module." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 130, + 268, + 483, + 413 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 268, + 483, + 413 + ], + "spans": [ + { + "bbox": [ + 130, + 268, + 483, + 413 + ], + "type": "text", + "content": "Knowledge Distillation. KD [9] transfers the knowledge in the output logits of a teacher model with a large size to a smaller student model without obvious performance degradation. The feature map in the hidden states and attention output space can also be used for KD to help students mimic teacher capability [11, 14]. Masked Generative Distillation [32] introduces a feature masking scheme into KD to improve students' representation power by forcing them to recover the masked features. Inspired by masked image modeling, the supervised masked KD model is proposed by inserting label information into self-distillation frameworks. However, to the best of our knowledge, KD has not been fully explored for training multilingual T2I generations. In this paper, we apply KD methods to language transfer in T2I and further harness the power of the PEA module through combination to achieve significant improvements." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 132, + 433, + 202, + 445 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 433, + 202, + 445 + ], + "spans": [ + { + "bbox": [ + 132, + 433, + 202, + 445 + ], + "type": "text", + "content": "3 Method" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 130, + 461, + 482, + 616 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 461, + 482, + 616 + ], + "spans": [ + { + "bbox": [ + 130, + 461, + 482, + 616 + ], + "type": "text", + "content": "A naive approach for generating images from non-English text is to substitute the CLIP text encoder in SD with a language-specific CLIP encoder. The text encoder is a crucial part of T2I models since exact and detailed text embedding is necessary for accurate generation. However, altering an existing text model to a different language-specific one is difficult. This difficulty arises because SD's text encoder and image generator are closely integrated, making it ineffective to just swap out the condition encoder. The primary issue is the mismatch between the new text encoder and the old image generator. Furthermore, our experiments show that joint fine-tuning depends heavily on high-quality text-image pairs and often fails due to catastrophic forgetting when updating well-trained parameters. In light of these insights, we keep the pretrained model's parameters fixed and develop a lightweight adapter with KD to synchronize the new text encoder and the image generator." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 130, + 617, + 482, + 666 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 617, + 482, + 666 + ], + "spans": [ + { + "bbox": [ + 130, + 617, + 482, + 666 + ], + "type": "text", + "content": "Figure 1 illustrates the overall architecture of PEA-Diffusion for cross-lingual transfer in T2I generation. To inherit the knowledge of the pretrained diffusion model, we use the pretrained SD as a teacher to supervise PEA-Diffusion learning through KD. For language-specific T2I generation, we use the corresponding" + } + ] + } + ], + "index": 6 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "spans": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "type": "text", + "content": "4" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 166, + 91, + 215, + 101 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 166, + 91, + 215, + 101 + ], + "spans": [ + { + "bbox": [ + 166, + 91, + 215, + 101 + ], + "type": "text", + "content": "J. Ma et al." + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 3 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 167, + 114, + 441, + 251 + ], + "blocks": [ + { + "bbox": [ + 167, + 114, + 441, + 251 + ], + "lines": [ + { + "bbox": [ + 167, + 114, + 441, + 251 + ], + "spans": [ + { + "bbox": [ + 167, + 114, + 441, + 251 + ], + "type": "image", + "image_path": "46c10031022c06625b26542269807a2a20bacc6b0c3e8b3b9f10c9a087c28369.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 131, + 260, + 480, + 283 + ], + "lines": [ + { + "bbox": [ + 131, + 260, + 480, + 283 + ], + "spans": [ + { + "bbox": [ + 131, + 260, + 480, + 283 + ], + "type": "text", + "content": "Fig. 1: An overview of the proposed PEA-Diffusion. Notice that only the lightweight adapter is trainable through the whole training process." + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_caption" + } + ], + "index": 2 + }, + { + "bbox": [ + 130, + 307, + 479, + 354 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 307, + 479, + 354 + ], + "spans": [ + { + "bbox": [ + 130, + 307, + 479, + 354 + ], + "type": "text", + "content": "language-specific CLIP as a text encoder to capture text information. The generator in PEA-Diffusion is the same UNet as that used in SD. A light-weight MLP adapter is added on top of the CLIP text encoder in PEA-Diffusion, facilitating knowledge transfer from teacher to student." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 131, + 373, + 220, + 385 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 131, + 373, + 220, + 385 + ], + "spans": [ + { + "bbox": [ + 131, + 373, + 220, + 385 + ], + "type": "text", + "content": "3.1 Preliminary" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 130, + 394, + 480, + 441 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 394, + 480, + 441 + ], + "spans": [ + { + "bbox": [ + 130, + 394, + 480, + 441 + ], + "type": "text", + "content": "The diffusion model is composed of a forward diffusion process and a reverse denoising process. The forward process gradually adds random noise to clean data and diffuses it into pure Gaussian noise, while the reverse process reverses the diffusion process to create satisfactory samples from the Gaussian noise [10]." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 130, + 441, + 481, + 538 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 441, + 481, + 538 + ], + "spans": [ + { + "bbox": [ + 130, + 441, + 481, + 538 + ], + "type": "text", + "content": "Specifically, for an input image " + }, + { + "bbox": [ + 130, + 441, + 481, + 538 + ], + "type": "inline_equation", + "content": "x_0 \\in \\mathbb{R}^{H \\times W \\times 3}" + }, + { + "bbox": [ + 130, + 441, + 481, + 538 + ], + "type": "text", + "content": ", the encoder " + }, + { + "bbox": [ + 130, + 441, + 481, + 538 + ], + "type": "inline_equation", + "content": "\\mathcal{E}" + }, + { + "bbox": [ + 130, + 441, + 481, + 538 + ], + "type": "text", + "content": " of the autoencoder transforms it into a latent representation " + }, + { + "bbox": [ + 130, + 441, + 481, + 538 + ], + "type": "inline_equation", + "content": "z_0 \\in \\mathbb{R}^{h \\times w \\times c}" + }, + { + "bbox": [ + 130, + 441, + 481, + 538 + ], + "type": "text", + "content": ", where " + }, + { + "bbox": [ + 130, + 441, + 481, + 538 + ], + "type": "inline_equation", + "content": "f = H / h = W / w" + }, + { + "bbox": [ + 130, + 441, + 481, + 538 + ], + "type": "text", + "content": " is the downsampling factor and " + }, + { + "bbox": [ + 130, + 441, + 481, + 538 + ], + "type": "inline_equation", + "content": "c" + }, + { + "bbox": [ + 130, + 441, + 481, + 538 + ], + "type": "text", + "content": " is the latent feature dimension. The diffusion process is then performed in the latent space, where a conditional UNet [27] denoiser " + }, + { + "bbox": [ + 130, + 441, + 481, + 538 + ], + "type": "inline_equation", + "content": "\\epsilon_{\\theta}" + }, + { + "bbox": [ + 130, + 441, + 481, + 538 + ], + "type": "text", + "content": " is employed to predict noise " + }, + { + "bbox": [ + 130, + 441, + 481, + 538 + ], + "type": "inline_equation", + "content": "\\epsilon" + }, + { + "bbox": [ + 130, + 441, + 481, + 538 + ], + "type": "text", + "content": " with current timestep " + }, + { + "bbox": [ + 130, + 441, + 481, + 538 + ], + "type": "inline_equation", + "content": "t" + }, + { + "bbox": [ + 130, + 441, + 481, + 538 + ], + "type": "text", + "content": ", noisy latent " + }, + { + "bbox": [ + 130, + 441, + 481, + 538 + ], + "type": "inline_equation", + "content": "z_t" + }, + { + "bbox": [ + 130, + 441, + 481, + 538 + ], + "type": "text", + "content": " and generation condition " + }, + { + "bbox": [ + 130, + 441, + 481, + 538 + ], + "type": "inline_equation", + "content": "C" + }, + { + "bbox": [ + 130, + 441, + 481, + 538 + ], + "type": "text", + "content": ". In T2I scenarios, the condition " + }, + { + "bbox": [ + 130, + 441, + 481, + 538 + ], + "type": "inline_equation", + "content": "C = \\tau_{\\theta}(y)" + }, + { + "bbox": [ + 130, + 441, + 481, + 538 + ], + "type": "text", + "content": " is produced by encoding the text prompts " + }, + { + "bbox": [ + 130, + 441, + 481, + 538 + ], + "type": "inline_equation", + "content": "y" + }, + { + "bbox": [ + 130, + 441, + 481, + 538 + ], + "type": "text", + "content": " with a pretrained CLIP [23] text encoder " + }, + { + "bbox": [ + 130, + 441, + 481, + 538 + ], + "type": "inline_equation", + "content": "\\tau_{\\theta}" + }, + { + "bbox": [ + 130, + 441, + 481, + 538 + ], + "type": "text", + "content": ". Therefore, the overall training objective of SD is defined as" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 208, + 546, + 481, + 569 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 208, + 546, + 481, + 569 + ], + "spans": [ + { + "bbox": [ + 208, + 546, + 481, + 569 + ], + "type": "interline_equation", + "content": "\\mathcal {L} _ {\\mathrm {S D}} = \\mathbb {E} _ {\\mathcal {E} (x _ {0}), C, \\epsilon \\sim \\mathcal {N} (0, 1), t} \\left[ \\| \\epsilon - \\epsilon_ {\\theta} (z _ {t}, t, C) \\| _ {2} ^ {2} \\right] \\tag {1}", + "image_path": "16f331eb517c37e4e386de835d30a8db62c6f4d351df092e49cd803e5d8b992b.jpg" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 131, + 594, + 275, + 605 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 131, + 594, + 275, + 605 + ], + "spans": [ + { + "bbox": [ + 131, + 594, + 275, + 605 + ], + "type": "text", + "content": "3.2 Cross-Linguual Transfer" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 130, + 613, + 481, + 666 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 613, + 481, + 666 + ], + "spans": [ + { + "bbox": [ + 130, + 613, + 481, + 666 + ], + "type": "text", + "content": "Objective. Given sets of parallel corpus " + }, + { + "bbox": [ + 130, + 613, + 481, + 666 + ], + "type": "inline_equation", + "content": "Y^{T} = \\{y_{i}^{t}\\}_{i = 1}^{N}" + }, + { + "bbox": [ + 130, + 613, + 481, + 666 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 130, + 613, + 481, + 666 + ], + "type": "inline_equation", + "content": "Y^{S} = \\{y_{i}^{t}\\}_{i = 1}^{N}" + }, + { + "bbox": [ + 130, + 613, + 481, + 666 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 130, + 613, + 481, + 666 + ], + "type": "inline_equation", + "content": "Y^{S^{\\prime}} = \\{y_{i}^{t}\\}_{i = 1}^{M}" + }, + { + "bbox": [ + 130, + 613, + 481, + 666 + ], + "type": "text", + "content": " and the corresponding images " + }, + { + "bbox": [ + 130, + 613, + 481, + 666 + ], + "type": "inline_equation", + "content": "I = \\{img\\}_{i = 1}^{N}" + }, + { + "bbox": [ + 130, + 613, + 481, + 666 + ], + "type": "text", + "content": " with " + }, + { + "bbox": [ + 130, + 613, + 481, + 666 + ], + "type": "inline_equation", + "content": "N" + }, + { + "bbox": [ + 130, + 613, + 481, + 666 + ], + "type": "text", + "content": " denoting the total number of paired samples, " + }, + { + "bbox": [ + 130, + 613, + 481, + 666 + ], + "type": "inline_equation", + "content": "Y^{T}" + }, + { + "bbox": [ + 130, + 613, + 481, + 666 + ], + "type": "text", + "content": " represents the English corpus that will be input into the teacher model, " + }, + { + "bbox": [ + 130, + 613, + 481, + 666 + ], + "type": "inline_equation", + "content": "Y^{S}" + }, + { + "bbox": [ + 130, + 613, + 481, + 666 + ], + "type": "text", + "content": " represents the non-English corpus with" + } + ] + } + ], + "index": 10 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 389, + 91, + 447, + 100 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 389, + 91, + 447, + 100 + ], + "spans": [ + { + "bbox": [ + 389, + 91, + 447, + 100 + ], + "type": "text", + "content": "PEA-diffusion" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 474, + 92, + 480, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 474, + 92, + 480, + 100 + ], + "spans": [ + { + "bbox": [ + 474, + 92, + 480, + 100 + ], + "type": "text", + "content": "5" + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 4 + }, + { + "para_blocks": [ + { + "bbox": [ + 130, + 115, + 482, + 262 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 115, + 482, + 262 + ], + "spans": [ + { + "bbox": [ + 130, + 115, + 482, + 262 + ], + "type": "text", + "content": "culturally irrelevant content, and " + }, + { + "bbox": [ + 130, + 115, + 482, + 262 + ], + "type": "inline_equation", + "content": "Y^{S'}" + }, + { + "bbox": [ + 130, + 115, + 482, + 262 + ], + "type": "text", + "content": " represents the non-English corpus with cultural relevance, they will be input into the student model. The teacher unet denoiser " + }, + { + "bbox": [ + 130, + 115, + 482, + 262 + ], + "type": "inline_equation", + "content": "\\epsilon^T" + }, + { + "bbox": [ + 130, + 115, + 482, + 262 + ], + "type": "text", + "content": " projects pairs " + }, + { + "bbox": [ + 130, + 115, + 482, + 262 + ], + "type": "inline_equation", + "content": "(Y^T, I)" + }, + { + "bbox": [ + 130, + 115, + 482, + 262 + ], + "type": "text", + "content": " into source middle feature maps and source logits output. Likewise, the student unet denoiser " + }, + { + "bbox": [ + 130, + 115, + 482, + 262 + ], + "type": "inline_equation", + "content": "\\epsilon^S" + }, + { + "bbox": [ + 130, + 115, + 482, + 262 + ], + "type": "text", + "content": " projects pairs " + }, + { + "bbox": [ + 130, + 115, + 482, + 262 + ], + "type": "inline_equation", + "content": "(Y^S, I)" + }, + { + "bbox": [ + 130, + 115, + 482, + 262 + ], + "type": "text", + "content": " into target middle feature maps and target logits output. Intuitively, there exists severe distribution mismatch between the source and target output, as well as between " + }, + { + "bbox": [ + 130, + 115, + 482, + 262 + ], + "type": "inline_equation", + "content": "\\epsilon^S" + }, + { + "bbox": [ + 130, + 115, + 482, + 262 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 130, + 115, + 482, + 262 + ], + "type": "inline_equation", + "content": "\\epsilon^T" + }, + { + "bbox": [ + 130, + 115, + 482, + 262 + ], + "type": "text", + "content": " due to the different languages. Therefore, our target is to make " + }, + { + "bbox": [ + 130, + 115, + 482, + 262 + ], + "type": "inline_equation", + "content": "\\epsilon^S" + }, + { + "bbox": [ + 130, + 115, + 482, + 262 + ], + "type": "text", + "content": " match with " + }, + { + "bbox": [ + 130, + 115, + 482, + 262 + ], + "type": "inline_equation", + "content": "\\epsilon^T" + }, + { + "bbox": [ + 130, + 115, + 482, + 262 + ], + "type": "text", + "content": ", so as to transfer teacher knowledge to the student, enabling cross-lingual T2I generation. For the student, we use a language-specific CLIP text encoder instead of the original English text encoder to enhance token- and sentence-level linguistic-native representations. For brevity, we omit the subscripts of " + }, + { + "bbox": [ + 130, + 115, + 482, + 262 + ], + "type": "inline_equation", + "content": "\\mathbb{E}_{\\mathcal{E}(x_0), \\tau_\\theta(y), \\epsilon \\sim \\mathcal{N}(0, 1), t}" + }, + { + "bbox": [ + 130, + 115, + 482, + 262 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 130, + 115, + 482, + 262 + ], + "type": "inline_equation", + "content": "z_t, t" + }, + { + "bbox": [ + 130, + 115, + 482, + 262 + ], + "type": "text", + "content": " conditions in the following notations." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 130, + 262, + 482, + 369 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 262, + 482, + 369 + ], + "spans": [ + { + "bbox": [ + 130, + 262, + 482, + 369 + ], + "type": "text", + "content": "Knowledge Distillation for T2I generation. The original KD leverages Kullback-Leibler (KL) divergence to measure the difference between the categorical probability distribution of the teacher's logits output and the student's logits output [9]. However, in the task of T2I generation, since the middle feature maps are the latent value of pixels instead of categorical probability distribution, KL divergence is unable to measure the difference between student and teacher. Consequently, we replace KL divergence with the L2-norm distance between the logits output from student and teacher to extend raw KD for T2I generation, whose loss function is formulated as:" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 236, + 374, + 481, + 395 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 236, + 374, + 481, + 395 + ], + "spans": [ + { + "bbox": [ + 236, + 374, + 481, + 395 + ], + "type": "interline_equation", + "content": "\\mathcal {L} _ {L K D} = \\mathbb {E} \\left[ \\| \\epsilon^ {T} (y ^ {T}) - \\epsilon^ {S} (y ^ {S}) \\| _ {2} ^ {2} \\right] \\tag {2}", + "image_path": "4285fa667b075605e33e9732d159c5c7a1bfdef4584edfc616d2f102326d682b.jpg" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 130, + 401, + 482, + 450 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 401, + 482, + 450 + ], + "spans": [ + { + "bbox": [ + 130, + 401, + 482, + 450 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 130, + 401, + 482, + 450 + ], + "type": "inline_equation", + "content": "y^{T} \\in Y^{T}" + }, + { + "bbox": [ + 130, + 401, + 482, + 450 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 130, + 401, + 482, + 450 + ], + "type": "inline_equation", + "content": "y^{S} \\in Y^{S}" + }, + { + "bbox": [ + 130, + 401, + 482, + 450 + ], + "type": "text", + "content": ". " + }, + { + "bbox": [ + 130, + 401, + 482, + 450 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_{LK D}" + }, + { + "bbox": [ + 130, + 401, + 482, + 450 + ], + "type": "text", + "content": " merely enforces student distribution to match teacher distribution at logits output layer. To further reduce distribution shift of teacher and student, we boost feature alignment of the middle feature maps between " + }, + { + "bbox": [ + 130, + 401, + 482, + 450 + ], + "type": "inline_equation", + "content": "\\epsilon^{S}" + }, + { + "bbox": [ + 130, + 401, + 482, + 450 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 130, + 401, + 482, + 450 + ], + "type": "inline_equation", + "content": "\\epsilon^{T}" + }, + { + "bbox": [ + 130, + 401, + 482, + 450 + ], + "type": "text", + "content": ":" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 211, + 458, + 481, + 483 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 211, + 458, + 481, + 483 + ], + "spans": [ + { + "bbox": [ + 211, + 458, + 481, + 483 + ], + "type": "interline_equation", + "content": "\\mathcal {L} _ {F L K D} = \\mathbb {E} \\left[ \\sum_ {l} \\| f \\left(g _ {l} ^ {T} \\left(y ^ {T}\\right)\\right) - f \\left(g _ {l} ^ {S} \\left(y ^ {S}\\right)\\right) \\| _ {2} ^ {2} \\right] \\tag {3}", + "image_path": "e5ae6a14c125eb8075c74dd6bd40c2f2a6090fb03e1a272c07d5f6c5de4247f2.jpg" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 130, + 491, + 482, + 540 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 491, + 482, + 540 + ], + "spans": [ + { + "bbox": [ + 130, + 491, + 482, + 540 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 130, + 491, + 482, + 540 + ], + "type": "inline_equation", + "content": "f" + }, + { + "bbox": [ + 130, + 491, + 482, + 540 + ], + "type": "text", + "content": " represents the norm operation, " + }, + { + "bbox": [ + 130, + 491, + 482, + 540 + ], + "type": "inline_equation", + "content": "g_{l}^{T}" + }, + { + "bbox": [ + 130, + 491, + 482, + 540 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 130, + 491, + 482, + 540 + ], + "type": "inline_equation", + "content": "g_{l}^{S}" + }, + { + "bbox": [ + 130, + 491, + 482, + 540 + ], + "type": "text", + "content": " represent the feature maps of the " + }, + { + "bbox": [ + 130, + 491, + 482, + 540 + ], + "type": "inline_equation", + "content": "l" + }, + { + "bbox": [ + 130, + 491, + 482, + 540 + ], + "type": "text", + "content": "-th layer in a predefined set of unet layers from the teacher and the student, respectively. With respective weights " + }, + { + "bbox": [ + 130, + 491, + 482, + 540 + ], + "type": "inline_equation", + "content": "\\lambda_{FL}" + }, + { + "bbox": [ + 130, + 491, + 482, + 540 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 130, + 491, + 482, + 540 + ], + "type": "inline_equation", + "content": "\\lambda_{L}" + }, + { + "bbox": [ + 130, + 491, + 482, + 540 + ], + "type": "text", + "content": ", the final KD objective is formalized as below:" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 239, + 550, + 481, + 562 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 239, + 550, + 481, + 562 + ], + "spans": [ + { + "bbox": [ + 239, + 550, + 481, + 562 + ], + "type": "interline_equation", + "content": "\\mathcal {L} _ {K D} = \\lambda_ {F L} \\mathcal {L} _ {F L K D} + \\lambda_ {L} \\mathcal {L} _ {L K D} \\tag {4}", + "image_path": "51526af108d3bd64347d8e66ad48c6f7a41a7e54bdeed0862f3f2e0ae1661b13.jpg" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 130, + 570, + 482, + 666 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 570, + 482, + 666 + ], + "spans": [ + { + "bbox": [ + 130, + 570, + 482, + 666 + ], + "type": "text", + "content": "Adapter. Due to the dimension difference between English CLIP and non-English CLIP, we propose an adapter for simple feature transformation to align feature dimension. Moreover, the adapter is trained to learn language-specific information. The adapter is implemented by a MLP with only 6M parameters. To accelerate training and avoid destroying the generalization capability of the pretrained models, we freeze the CLIP encoder and image generator. As a result, the proposed PEA-Diffusion is both parameter- and computation-efficient because just the adapter is trained." + } + ] + } + ], + "index": 9 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "spans": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "type": "text", + "content": "6" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 166, + 91, + 215, + 101 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 166, + 91, + 215, + 101 + ], + "spans": [ + { + "bbox": [ + 166, + 91, + 215, + 101 + ], + "type": "text", + "content": "J. Ma et al." + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 5 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 138, + 115, + 476, + 319 + ], + "blocks": [ + { + "bbox": [ + 138, + 115, + 476, + 319 + ], + "lines": [ + { + "bbox": [ + 138, + 115, + 476, + 319 + ], + "spans": [ + { + "bbox": [ + 138, + 115, + 476, + 319 + ], + "type": "image", + "image_path": "c505fe5df1fc2348473de1214bea9d5e8b1ed6fdc267739f7d31aedb002d3f53.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 132, + 334, + 479, + 346 + ], + "lines": [ + { + "bbox": [ + 132, + 334, + 479, + 346 + ], + "spans": [ + { + "bbox": [ + 132, + 334, + 479, + 346 + ], + "type": "text", + "content": "Fig. 2: Image generation visualization of different models in Chinese specific language." + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_caption" + } + ], + "index": 2 + }, + { + "bbox": [ + 132, + 369, + 353, + 380 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 369, + 353, + 380 + ], + "spans": [ + { + "bbox": [ + 132, + 369, + 353, + 380 + ], + "type": "text", + "content": "3.3 Training Strategy and Final Objective" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 130, + 388, + 482, + 522 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 388, + 482, + 522 + ], + "spans": [ + { + "bbox": [ + 130, + 388, + 482, + 522 + ], + "type": "text", + "content": "For our goal of training a multi-language T2I generation model, we only need to train on English data since both the new CLIP text encoder and the teacher model's CLIP text encoder can accept English inputs. In this case, we can use aligned KD loss " + }, + { + "bbox": [ + 130, + 388, + 482, + 522 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_{KD}" + }, + { + "bbox": [ + 130, + 388, + 482, + 522 + ], + "type": "text", + "content": ". For our goal of training a language-specific T2I generation model, we adopt a hybrid training strategy, for the target language's parallel corpus, we only need to use KD loss " + }, + { + "bbox": [ + 130, + 388, + 482, + 522 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_{KD}" + }, + { + "bbox": [ + 130, + 388, + 482, + 522 + ], + "type": "text", + "content": " to align the distribution. On the other hand, we also have language-specific culture-related data " + }, + { + "bbox": [ + 130, + 388, + 482, + 522 + ], + "type": "inline_equation", + "content": "Y^{S^{\\prime}}" + }, + { + "bbox": [ + 130, + 388, + 482, + 522 + ], + "type": "text", + "content": ". Because there is a significant translation error when translating these data into English and the teacher model cannot provide high-quality knowledge guidance for this type of data, we only use " + }, + { + "bbox": [ + 130, + 388, + 482, + 522 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_{\\mathrm{SD}}" + }, + { + "bbox": [ + 130, + 388, + 482, + 522 + ], + "type": "text", + "content": " (1) for improving culture understanding ability. The final objective is formalized as follows:" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 132, + 530, + 487, + 570 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 530, + 487, + 570 + ], + "spans": [ + { + "bbox": [ + 132, + 530, + 487, + 570 + ], + "type": "interline_equation", + "content": "\\mathcal {L} = \\left\\{ \\begin{array}{l l} \\lambda_ {F L} \\mathbb {E} \\Big [ \\| \\epsilon^ {T} (y ^ {T}) - \\epsilon^ {S} (y) \\| _ {2} ^ {2} \\Big ] + \\lambda_ {L} \\mathbb {E} \\Big [ \\sum_ {l} \\| f (g _ {l} ^ {T} (y ^ {T})) - f (g _ {l} ^ {S} (y)) \\| _ {2} ^ {2} \\Big ], & y \\in Y ^ {S} \\\\ \\mathbb {E} \\Big [ \\| \\epsilon - \\epsilon^ {S} (y) \\| _ {2} ^ {2} \\Big ], & y \\in Y ^ {S ^ {\\prime}} \\end{array} \\right.", + "image_path": "63a2f82ae0847bb07b414015c749c292c83b122e04d8e6d92df0a8cbfba33778.jpg" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 132, + 585, + 230, + 598 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 585, + 230, + 598 + ], + "spans": [ + { + "bbox": [ + 132, + 585, + 230, + 598 + ], + "type": "text", + "content": "4 Experiments" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 132, + 609, + 250, + 621 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 609, + 250, + 621 + ], + "spans": [ + { + "bbox": [ + 132, + 609, + 250, + 621 + ], + "type": "text", + "content": "4.1 Data Preparation" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 130, + 629, + 481, + 666 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 629, + 481, + 666 + ], + "spans": [ + { + "bbox": [ + 130, + 629, + 481, + 666 + ], + "type": "text", + "content": "We download the Laion2B-multi [29] dataset along with English translations for image descriptions. Based on the open-source CLIP models available for different languages: Chinese CLIP, Russian CLIP, Japanese CLIP, Korean CLIP, and" + } + ] + } + ], + "index": 9 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 389, + 91, + 447, + 100 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 389, + 91, + 447, + 100 + ], + "spans": [ + { + "bbox": [ + 389, + 91, + 447, + 100 + ], + "type": "text", + "content": "PEA-diffusion" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 474, + 92, + 481, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 474, + 92, + 481, + 100 + ], + "spans": [ + { + "bbox": [ + 474, + 92, + 481, + 100 + ], + "type": "text", + "content": "7" + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 6 + }, + { + "para_blocks": [ + { + "bbox": [ + 130, + 116, + 482, + 164 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 116, + 482, + 164 + ], + "spans": [ + { + "bbox": [ + 130, + 116, + 482, + 164 + ], + "type": "text", + "content": "Italian CLIP, we determine to conduct experiments on the above five languages. For Chinese, we also collect WuKong [7] data with Chinese cultural image-text pairs and Chinese-English parallel corpus translations of Laion2B-en. For more data, model, and code details, please refer to the Appendix." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 132, + 190, + 358, + 202 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 190, + 358, + 202 + ], + "spans": [ + { + "bbox": [ + 132, + 190, + 358, + 202 + ], + "type": "text", + "content": "4.2 Implementation Details and Evaluation" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 130, + 219, + 482, + 302 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 219, + 482, + 302 + ], + "spans": [ + { + "bbox": [ + 130, + 219, + 482, + 302 + ], + "type": "text", + "content": "Implementation details. We conduct all experiments on Stable Diffusion (SD)- [25], and Stable Diffusion XL (SDXL)- [21] like backbones. Take SDXL base model as an example. PEA-Diffusion consists of VAE, UNet, CLIP text encoder, and the adapter module, comprising 3 billion parameters, out of which a mere 6 million parameters (" + }, + { + "bbox": [ + 130, + 219, + 482, + 302 + ], + "type": "inline_equation", + "content": "\\approx 0.2\\%" + }, + { + "bbox": [ + 130, + 219, + 482, + 302 + ], + "type": "text", + "content": ") are trainable. We set " + }, + { + "bbox": [ + 130, + 219, + 482, + 302 + ], + "type": "inline_equation", + "content": "\\lambda_{FL}" + }, + { + "bbox": [ + 130, + 219, + 482, + 302 + ], + "type": "text", + "content": " 0.1 and " + }, + { + "bbox": [ + 130, + 219, + 482, + 302 + ], + "type": "inline_equation", + "content": "\\lambda_{L}" + }, + { + "bbox": [ + 130, + 219, + 482, + 302 + ], + "type": "text", + "content": " 1 respectively based on the loss performance. The entire model is trained on 64 A100 GPUs for 70000 steps with a batch size of 5 per GPU." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 130, + 305, + 482, + 449 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 305, + 482, + 449 + ], + "spans": [ + { + "bbox": [ + 130, + 305, + 482, + 449 + ], + "type": "text", + "content": "Test benchmark. To evaluate the capability of PEA-Diffusion to generate images and capture culture-specific concepts of different languages, we introduce two datasets: Multilingual-General (MG) for generation quality evaluation and Multilingual-Cultural (MC) for culture-specific concepts evaluation. We collect 200 prompts for MG, which covered content such as people, animals, landscapes, and fantasy themes. The descriptions include diverse styles, such as cartoon, anime, illustration, watercolor, and cyberpunk. The data comes from various major art websites where users provide real prompts. On the other hand, we also construct MC data including 200 prompts. The collection is mainly based on the ChatGPT tool and consists of culturally relevant content for each specific language, including but not limited to local cuisine, architecture, clothing, and distinctive cultural features." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 130, + 450, + 482, + 665 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 450, + 482, + 665 + ], + "spans": [ + { + "bbox": [ + 130, + 450, + 482, + 665 + ], + "type": "text", + "content": "Evaluation metrics. For the general evaluation data MG, the widely used FID metric can only measure the feature distance between generated images and specific images. It often fails to provide accurate assessment for high-quality generated images, as emphasized and experimentally validated by SDXL [21] and other studies. Therefore, in order to evaluate the quality of generated images more accurately, we introduce two human preference models as our evaluation metrics: ImageReward [30] and PickScore [15]. Additionally, we continue to utilize OpenCLIP ViT-bigG to measure the similarity between images and text. As for the language-specific evaluation data MC, constructed for each language, to ensure a fair comparison of content relevance, we only consider the CLIP model trained in the specific language to calculate the correlation between generated images and their corresponding text. Since the general CLIP model can only process non-English data after translation into English, this process can introduce significant performance degradation. Furthermore, our paper's focus is precisely on addressing the errors caused by translating non-English prompts into English to generate images. Therefore, other metrics that involve translation processes, such as the two models based on human preference mentioned above, will no longer be valid for evaluation." + } + ] + } + ], + "index": 6 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "spans": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "type": "text", + "content": "8" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 166, + 91, + 215, + 101 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 166, + 91, + 215, + 101 + ], + "spans": [ + { + "bbox": [ + 166, + 91, + 215, + 101 + ], + "type": "text", + "content": "J. Ma et al." + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 7 + }, + { + "para_blocks": [ + { + "bbox": [ + 130, + 116, + 482, + 152 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 116, + 482, + 152 + ], + "spans": [ + { + "bbox": [ + 130, + 116, + 482, + 152 + ], + "type": "text", + "content": "Baseline methods. We generated image results in specific languages based on two T2I base models: SD1.5 and SDXL. We choose the following four categories of approaches as our baselines." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 138, + 159, + 485, + 303 + ], + "type": "list", + "angle": 0, + "index": 7, + "blocks": [ + { + "bbox": [ + 138, + 159, + 355, + 171 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 159, + 355, + 171 + ], + "spans": [ + { + "bbox": [ + 138, + 159, + 355, + 171 + ], + "type": "text", + "content": "- Translation: we apply Google translation API." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 138, + 171, + 480, + 195 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 171, + 480, + 195 + ], + "spans": [ + { + "bbox": [ + 138, + 171, + 480, + 195 + ], + "type": "text", + "content": "- AltDiffusion [33]: we test two multilingual SD models based on the open-source AltDiffusion framework: AltDiffusion-m9 and AltDiffusion-18." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 138, + 196, + 485, + 266 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 196, + 485, + 266 + ], + "spans": [ + { + "bbox": [ + 138, + 196, + 485, + 266 + ], + "type": "text", + "content": "- GlueGen [22]: We evaluate our model on three languages that overlap with the checkpoint provided by GlueGen, which are Chinese, Italian, and Japanese. For the other two languages, Russian and Korean, we implement GlueGen based on its code and our data. In addition, to further compare the impact of different language-specific CLIPs on the results, we also retrain the GlueGen based on five languages' CLIP text encoders." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 138, + 267, + 480, + 303 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 267, + 480, + 303 + ], + "spans": [ + { + "bbox": [ + 138, + 267, + 480, + 303 + ], + "type": "text", + "content": "- Fine-tuning with LoRA [12] and " + }, + { + "bbox": [ + 138, + 267, + 480, + 303 + ], + "type": "inline_equation", + "content": "W_{K}, W_{V}" + }, + { + "bbox": [ + 138, + 267, + 480, + 303 + ], + "type": "text", + "content": " (key and value projection matrices): As low-cost experiments for comparison, we also fine-tuning the model to transfer language capabilities." + } + ] + } + ], + "index": 6 + } + ], + "sub_type": "text" + }, + { + "bbox": [ + 132, + 320, + 270, + 333 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 320, + 270, + 333 + ], + "spans": [ + { + "bbox": [ + 132, + 320, + 270, + 333 + ], + "type": "text", + "content": "4.3 Experimental Results" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 130, + 340, + 482, + 533 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 340, + 482, + 533 + ], + "spans": [ + { + "bbox": [ + 130, + 340, + 482, + 533 + ], + "type": "text", + "content": "Generation analysis. We validate our approach based on two different T2I base models. Firstly, to ensure a fair comparison with GlueGen and AltDiffusion, we train PEA-Diffusion based on the SD1.5 model. Secondly, to validate our low-cost experiment approach, we train a larger model based on SDXL. The compared methods are LoRA fine-tuning with rank 64 and " + }, + { + "bbox": [ + 130, + 340, + 482, + 533 + ], + "type": "inline_equation", + "content": "W_{K}" + }, + { + "bbox": [ + 130, + 340, + 482, + 533 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 130, + 340, + 482, + 533 + ], + "type": "inline_equation", + "content": "W_{V}" + }, + { + "bbox": [ + 130, + 340, + 482, + 533 + ], + "type": "text", + "content": " fine-tuning, both of which are parameter-efficient fine-tuning strategies. For the second approach, as proposed in [16], when updating the mapping from given text to image distribution, only updating " + }, + { + "bbox": [ + 130, + 340, + 482, + 533 + ], + "type": "inline_equation", + "content": "W_{K}^{(i)}" + }, + { + "bbox": [ + 130, + 340, + 482, + 533 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 130, + 340, + 482, + 533 + ], + "type": "inline_equation", + "content": "W_{V}^{(i)}" + }, + { + "bbox": [ + 130, + 340, + 482, + 533 + ], + "type": "text", + "content": " in each cross-attention block " + }, + { + "bbox": [ + 130, + 340, + 482, + 533 + ], + "type": "inline_equation", + "content": "i" + }, + { + "bbox": [ + 130, + 340, + 482, + 533 + ], + "type": "text", + "content": " is sufficient since text features are the only input to the key and value projection matrix. Furthermore, to validate the wide application of the proposed approach with different text encoders, we train our model with a multi-language T5 encoder and a multi-language CLIP encoder. This type of training only requires English text data, which makes it even simpler and more efficient. Fig. 2 illustrates the comparisons between translation and our methods with different training strategies. It can be observed that our PEA-Diffusion can achieve Chinese culturally related image generation, but translation totally fails in these cases." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 130, + 534, + 482, + 666 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 534, + 482, + 666 + ], + "spans": [ + { + "bbox": [ + 130, + 534, + 482, + 666 + ], + "type": "text", + "content": "As shown in the comprehensive results of Table 1, when compared with the translation method, our PEA-Diffusion outperforms it for almost all languages (except Japanese) in terms of CLIPScore for MC datasets. It is obvious that the translation method cannot capture the essence of different languages' cultures, leading to a semantic gap between the original prompts and the translated English prompts. However, our approach can use additional culturally related image-text pairs to train the adapter, which can align the language understanding ability of the text encoder with the generation ability of the UNet. Even though the UNet is frozen during the whole training process, we argue that our PEA-Diffusion can still stimulate the potential generation ability for cultural-related images of the original UNet. Since SD1.5 and SDXL are all trained on" + } + ] + } + ], + "index": 10 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 389, + 91, + 447, + 101 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 389, + 91, + 447, + 101 + ], + "spans": [ + { + "bbox": [ + 389, + 91, + 447, + 101 + ], + "type": "text", + "content": "PEA-diffusion" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 474, + 91, + 481, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 474, + 91, + 481, + 100 + ], + "spans": [ + { + "bbox": [ + 474, + 91, + 481, + 100 + ], + "type": "text", + "content": "9" + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 8 + }, + { + "para_blocks": [ + { + "type": "table", + "bbox": [ + 134, + 147, + 481, + 243 + ], + "blocks": [ + { + "bbox": [ + 132, + 125, + 481, + 146 + ], + "lines": [ + { + "bbox": [ + 132, + 125, + 481, + 146 + ], + "spans": [ + { + "bbox": [ + 132, + 125, + 481, + 146 + ], + "type": "text", + "content": "Table 1: CLIPScore with different specific languages. For Chinese, we evaluated two general evaluation metrics with MG data. " + }, + { + "bbox": [ + 132, + 125, + 481, + 146 + ], + "type": "inline_equation", + "content": "\\dagger" + }, + { + "bbox": [ + 132, + 125, + 481, + 146 + ], + "type": "text", + "content": " Indicates open source models." + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 134, + 147, + 481, + 243 + ], + "lines": [ + { + "bbox": [ + 134, + 147, + 481, + 243 + ], + "spans": [ + { + "bbox": [ + 134, + 147, + 481, + 243 + ], + "type": "table", + "html": "
Model BaseMethodszh-MCzh-MGru-MCit-MCko-MCja-MC
CLIPScore_zhImageRewardPickScoreCLIPScoreRuCLIPScore—itCLIPScore_koCLIPScore_ja
SD1.5Translate0.16860.24270.17990.49110.35560.19040.2885
Altdiffusion m9 [33]†0.2100-0.69660.12020.37150.31770.19060.2461
Altdiffusion m18 [33]†0.21720.33160.16480.47300.35020.19170.3008
GlueGen [22]†0.1700-0.51870.12240.41840.30650.17690.2506
GlueGen [22]0.2153-0.46850.12560.39870.29280.18430.2465
Ours0.25240.09850.17010.50070.37280.18960.2501
SDXLTranslate0.23831.36470.24020.48440.35910.20330.2958
LoRA0.23660.12170.15980.46560.34760.20340.1940
Wkand Wv0.24850.14130.16210.47850.34950.20880.2022
Ours w/ MuTI50.23010.82540.20010.45680.38050.19080.2996
Ours w/ Multilingual-CLIP0.24071.21220.23240.46730.38560.21570.3007
Ours0.26101.02450.22440.50120.41910.21940.2544
", + "image_path": "30451c5e392fff7bfcac3cb0cb354f78b2376b3ef1c6ac706b4a285151b4e4c5.jpg" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "table_body" + } + ], + "index": 3 + }, + { + "bbox": [ + 130, + 267, + 479, + 339 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 267, + 479, + 339 + ], + "spans": [ + { + "bbox": [ + 130, + 267, + 479, + 339 + ], + "type": "text", + "content": "billions of image-text pairs that contain wide image domains, our approach can successfully awaken their generative abilities in different cultural contexts. In terms of ImageReward and PickScore for the MG dataset, we must admit that PEA-Diffusion is slightly inferior to translation methods. This is because using other text encoders (e.g., Chinese CLIP and Multilingual CLIP) cannot ensure complete alignment between the text embedding and the UNet feature spaces." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 130, + 340, + 481, + 495 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 340, + 481, + 495 + ], + "spans": [ + { + "bbox": [ + 130, + 340, + 481, + 495 + ], + "type": "text", + "content": "Even though Altdiffusion retrans the multilingual CLIP text encoder and UNet on billions of image-text pairs, it still performs worse than our approach in terms of almost all metrics for all languages. GlueGen is the most related work on our concerned topic. The statistics presented in Table 1 reveal that both checkpoints provided by the authors or re-implemented by ourselves are significantly inferior to our PEA-Diffusion, especially for the ImageReward and PickScore metrics, resulting in very low generation quality since only aligning distributions of text embedding for different languages is very hard, even with its sophisticatedly designed GlenNet. In addition, it cannot generate culturally related images. Therefore, it's worth mentioning that our core contribution is that we align the model's generation ability through UNet distillation only by a MLP-like adapter between different languages. We don't forcibly align the text embedding spaces of different languages." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 130, + 497, + 481, + 581 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 497, + 481, + 581 + ], + "spans": [ + { + "bbox": [ + 130, + 497, + 481, + 581 + ], + "type": "text", + "content": "Additionally, we undertake various parameter-efficient fine-tuning strategies for comparison. The fine-tuning techniques, including LoRA and " + }, + { + "bbox": [ + 130, + 497, + 481, + 581 + ], + "type": "inline_equation", + "content": "W_{k}" + }, + { + "bbox": [ + 130, + 497, + 481, + 581 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 130, + 497, + 481, + 581 + ], + "type": "inline_equation", + "content": "W_{v}" + }, + { + "bbox": [ + 130, + 497, + 481, + 581 + ], + "type": "text", + "content": ", show subpar performance based on ImageReward, PickScore, and CLIPScore metrics. This is attributed to the fact that these fine-tuning methods depend significantly on high-quality image-text datasets, whereas our PEA-Diffusion requires merely some parallel multilingual texts and images and lacks stringent quality requirements. Further experiments can be found in the Appendix." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 130, + 582, + 481, + 665 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 582, + 481, + 665 + ], + "spans": [ + { + "bbox": [ + 130, + 582, + 481, + 665 + ], + "type": "text", + "content": "Analysis on different languages. The data in Table 1 indicate that the CLIPScores for Russian, Italian, and Korean either match or exceed those of other baseline methods by a notable margin. For Japanese, however, the CLIP-Score is slightly lower compared to translation and AltDiffusion. We ascribe this to the CLIP model's limited encoding ability for Japanese, which fails to align with the SD model via the adapter layer. Additionally, with only 1.6M training samples for Korean, our method still achieves effective language transfer, under-" + } + ] + } + ], + "index": 7 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "spans": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "type": "text", + "content": "10" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 166, + 91, + 215, + 101 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 166, + 91, + 215, + 101 + ], + "spans": [ + { + "bbox": [ + 166, + 91, + 215, + 101 + ], + "type": "text", + "content": "J. Ma et al." + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 9 + }, + { + "para_blocks": [ + { + "bbox": [ + 130, + 116, + 479, + 140 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 116, + 479, + 140 + ], + "spans": [ + { + "bbox": [ + 130, + 116, + 479, + 140 + ], + "type": "text", + "content": "scoring that large-scale training datasets are not essential. Further visualization examples of culturally-related prompts can be found in the Appendix." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 130, + 141, + 329, + 332 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 141, + 329, + 332 + ], + "spans": [ + { + "bbox": [ + 130, + 141, + 329, + 332 + ], + "type": "text", + "content": "Analysis of training cost. Our PEADiffusion freezes the CLIP text encoder, the UNet, and VAE during the whole training process, only the MLP mapping adapter is trained, accounting for merely 6M training parameters, which is " + }, + { + "bbox": [ + 130, + 141, + 329, + 332 + ], + "type": "inline_equation", + "content": "0.2\\%" + }, + { + "bbox": [ + 130, + 141, + 329, + 332 + ], + "type": "text", + "content": " of total parameters for SDXL-based models. We employed 64 A100 GPUs to train 70000 steps, resulting in 1600 GPU hours. As a comparison, AltDiffusion [33] trained the model for three stages with nearly 30000 GPU hours, which is " + }, + { + "bbox": [ + 130, + 141, + 329, + 332 + ], + "type": "inline_equation", + "content": "\\approx 20" + }, + { + "bbox": [ + 130, + 141, + 329, + 332 + ], + "type": "text", + "content": " times our training cost. In addition, our method achieves higher performance in terms of almost all metrics and the five experimental languages. We also re-implemented and trained GlueGen [22] with the same dataset" + } + ] + } + ], + "index": 3 + }, + { + "type": "image", + "bbox": [ + 340, + 175, + 479, + 260 + ], + "blocks": [ + { + "bbox": [ + 340, + 175, + 479, + 260 + ], + "lines": [ + { + "bbox": [ + 340, + 175, + 479, + 260 + ], + "spans": [ + { + "bbox": [ + 340, + 175, + 479, + 260 + ], + "type": "image", + "image_path": "ba27d4e8ef70fe6d6374aa4a2b3472a4c8d8498df69bea3a856dd84986fab0e4.jpg" + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 335, + 267, + 480, + 300 + ], + "lines": [ + { + "bbox": [ + 335, + 267, + 480, + 300 + ], + "spans": [ + { + "bbox": [ + 335, + 267, + 480, + 300 + ], + "type": "text", + "content": "Fig.3: Training parameters(M) and cost comparison for different methods." + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "image_caption" + } + ], + "index": 4 + }, + { + "bbox": [ + 130, + 332, + 482, + 452 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 332, + 482, + 452 + ], + "spans": [ + { + "bbox": [ + 130, + 332, + 482, + 452 + ], + "type": "text", + "content": "constructed by ourselves. It indeed achieves very low training costs with only 1 A100 GPU for 24 hours, but results in very poor performance no matter for image quality (-0.5 ImageReward score) or language transferring ability. Finally, we also conduct other parameter-efficient fine-tuning strategies (e.g., LoRA and tuning only " + }, + { + "bbox": [ + 130, + 332, + 482, + 452 + ], + "type": "inline_equation", + "content": "W_{k}" + }, + { + "bbox": [ + 130, + 332, + 482, + 452 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 130, + 332, + 482, + 452 + ], + "type": "inline_equation", + "content": "W_{v}" + }, + { + "bbox": [ + 130, + 332, + 482, + 452 + ], + "type": "text", + "content": " matrices of attention layers) and achieve much lower performance and higher training costs, which demonstrated the effectiveness and efficiency of our PEA-Diffusion. Figure 3 only shows the relationship between training parameters and GPU hours of different methods. For more information on model fitting and training data, as well as adapter module parameter variations, please refer to Appendix." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 132, + 473, + 246, + 484 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 473, + 246, + 484 + ], + "spans": [ + { + "bbox": [ + 132, + 473, + 246, + 484 + ], + "type": "text", + "content": "4.4 Ablation Studies" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 130, + 497, + 482, + 605 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 497, + 482, + 605 + ], + "spans": [ + { + "bbox": [ + 130, + 497, + 482, + 605 + ], + "type": "text", + "content": "Since Chinese data has certain advantages in terms of data quality and quantity, to eliminate errors caused by data, our ablation experiments are all based on Chinese data. The experiments include four parts: First, we conduct a baseline experiment where we only perform standard model fine-tuning without incorporating any KD strategies. Secondly, inspired by GlueGen [22], we employ text encoder KD to explore the impact of distillation at different positions in the SD framework. Next, we perform distillation based on the UNet model, including feature distillation between different UNet layers and logits distillation for the final output. Finally, we combine all of the distillation positions described above." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 130, + 605, + 482, + 665 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 605, + 482, + 665 + ], + "spans": [ + { + "bbox": [ + 130, + 605, + 482, + 665 + ], + "type": "text", + "content": "Effectiveness of KD. First, KD is crucial for guiding models onto the right learning trajectory for rapid convergence and is vital for the PEA-Diffusion method. As indicated in the first row of Table 2, the standalone fine-tuning method yields notably inferior performance in both image quality (ImageReward and PickScore) and image-text alignment (CLIPScores)." + } + ] + } + ], + "index": 9 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 389, + 91, + 447, + 100 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 389, + 91, + 447, + 100 + ], + "spans": [ + { + "bbox": [ + 389, + 91, + 447, + 100 + ], + "type": "text", + "content": "PEA-diffusion" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 470, + 91, + 479, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 470, + 91, + 479, + 100 + ], + "spans": [ + { + "bbox": [ + 470, + 91, + 479, + 100 + ], + "type": "text", + "content": "11" + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 10 + }, + { + "para_blocks": [ + { + "type": "table", + "bbox": [ + 170, + 137, + 446, + 206 + ], + "blocks": [ + { + "bbox": [ + 183, + 125, + 429, + 135 + ], + "lines": [ + { + "bbox": [ + 183, + 125, + 429, + 135 + ], + "spans": [ + { + "bbox": [ + 183, + 125, + 429, + 135 + ], + "type": "text", + "content": "Table 2: Ablation results for different knowledge strategies." + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 170, + 137, + 446, + 206 + ], + "lines": [ + { + "bbox": [ + 170, + 137, + 446, + 206 + ], + "spans": [ + { + "bbox": [ + 170, + 137, + 446, + 206 + ], + "type": "table", + "html": "
EvaluationImageRewardPickScoreCLIPScore_enCLIPScore_zh
Fine-tuning-1.31540.18540.32780.1901
Fine-tuning w/ TKD-0.98750.19970.35060.2287
Fine-tuning w/ ULKD-1.04470.19570.33310.1968
Fine-tuning w/ UFKD0.95560.21740.40850.2557
Fine-tuning w/ UTKD0.84900.21130.40110.2461
Ours(Fine-tuning w/ UKD)1.02450.22440.42240.2610
", + "image_path": "07a47706fcd4e234d68d8f2d00b3acf594f20160f8561052edeb130a8bfff626.jpg" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "table_body" + } + ], + "index": 3 + }, + { + "bbox": [ + 130, + 226, + 481, + 333 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 226, + 481, + 333 + ], + "spans": [ + { + "bbox": [ + 130, + 226, + 481, + 333 + ], + "type": "text", + "content": "Effectiveness of UNet Feature KD (UFKD). Our KD method consists of two parts: the first part involves distilling the logits results generated by the UNet (ULKD), we then introduce a distillation of intermediate layer features in the UNet, which serves as the primary knowledge guidance mechanism. As shown in the third and fourth lines of Table 2, we can observe that logit distillation can only improve marginally when compared with direct fine-tuning whereas feature distillation boosts the performance remarkably. We attribute this phenomenon to the generation ability of the large amount of UNet parameters, so that intermediate feature supervision is essential for knowledge transfer." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 130, + 334, + 481, + 453 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 334, + 481, + 453 + ], + "spans": [ + { + "bbox": [ + 130, + 334, + 481, + 453 + ], + "type": "text", + "content": "Effectiveness of UNet KD. Furthermore, different distillation positions have different effects. Our initial focus is on aligning the feature spaces of different text encoders, so a natural idea is to distill only the feature information of the teacher model's text encoder (TKD). However, as the results show in the second line of Table 2, although the overall performance exceeds the pure fine-tuning method, it is still far from our best performing approach. This illustrates that simply aligning the text encoder is insufficient and ineffective. This statistic also matches the experimental results of GlueGen in Table 1: simply aligning text embedding spaces cannot obtain satisfactory image generation results. Therefore, this further proves the necessity of distilling the features of the UNet itself." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 130, + 453, + 481, + 537 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 453, + 481, + 537 + ], + "spans": [ + { + "bbox": [ + 130, + 453, + 481, + 537 + ], + "type": "text", + "content": "Finally, we perform KD on both the UNet and text encoder (UTKD). However, the fifth line of Table 2 shows that its effect is worse than performing only UNet distillation. We suspect that the addition of text encoder KD actually weakens the learning of the SD core UNet structure. The main knowledge ability for generating images is mainly influenced by the UNet model, which is responsible for the adapter alignment feature in PEA-Diffusion. More ablation experiments can refer to Appendix." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 131, + 552, + 422, + 565 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 131, + 552, + 422, + 565 + ], + "spans": [ + { + "bbox": [ + 131, + 552, + 422, + 565 + ], + "type": "text", + "content": "4.5 Exploring the Domain Adaptation of PEA-Diffusion" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 130, + 570, + 481, + 628 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 570, + 481, + 628 + ], + "spans": [ + { + "bbox": [ + 130, + 570, + 481, + 628 + ], + "type": "text", + "content": "To demonstrate the domain adaptation of our model with minimal data guidance, we first constructed a dataset of 20,000 parallel image and text pairs in the domain of human. We then trained the models based on the models fine-tuned " + }, + { + "bbox": [ + 130, + 570, + 481, + 628 + ], + "type": "inline_equation", + "content": "W_{K}^{(i)}, W_{V}^{(i)}" + }, + { + "bbox": [ + 130, + 570, + 481, + 628 + ], + "type": "text", + "content": " mentioned in section 4.3 with UKD, and evaluated on the MG and MC dataset. The results on Chinese are shown Table 3." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 130, + 629, + 481, + 665 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 629, + 481, + 665 + ], + "spans": [ + { + "bbox": [ + 130, + 629, + 481, + 665 + ], + "type": "text", + "content": "We observed significant improvements in the three metrics on the MG test data. This suggests that PEA can be guided by a small amount of data to perform out-of-domain knowledge transfer." + } + ] + } + ], + "index": 9 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "spans": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "type": "text", + "content": "12" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 166, + 91, + 215, + 101 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 166, + 91, + 215, + 101 + ], + "spans": [ + { + "bbox": [ + 166, + 91, + 215, + 101 + ], + "type": "text", + "content": "J. Ma et al." + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 11 + }, + { + "para_blocks": [ + { + "type": "table", + "bbox": [ + 169, + 137, + 447, + 175 + ], + "blocks": [ + { + "bbox": [ + 151, + 125, + 461, + 136 + ], + "lines": [ + { + "bbox": [ + 151, + 125, + 461, + 136 + ], + "spans": [ + { + "bbox": [ + 151, + 125, + 461, + 136 + ], + "type": "text", + "content": "Table 3: Domain adaptation with 20,000 data in Chinese specific language." + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 169, + 137, + 447, + 175 + ], + "lines": [ + { + "bbox": [ + 169, + 137, + 447, + 175 + ], + "spans": [ + { + "bbox": [ + 169, + 137, + 447, + 175 + ], + "type": "table", + "html": "
EvaluationImageRewardPickScoreCLIPScore_enCLIPScore_zh
Wk and Wv0.14130.16210.36740.2485
Ours(Fine-tuning w/ UKD)0.83680.21450.40980.2501
", + "image_path": "c21e8083b1221e68b7df7a87d7f6fb2be7f4520bcc2fec796650d4451efca06d.jpg" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "table_body" + } + ], + "index": 3 + }, + { + "bbox": [ + 132, + 196, + 443, + 209 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 196, + 443, + 209 + ], + "spans": [ + { + "bbox": [ + 132, + 196, + 443, + 209 + ], + "type": "text", + "content": "4.6 Adaptability Experiment of English Downstream Model" + } + ] + } + ], + "index": 4 + }, + { + "type": "image", + "bbox": [ + 133, + 237, + 307, + 282 + ], + "blocks": [ + { + "bbox": [ + 133, + 237, + 307, + 282 + ], + "lines": [ + { + "bbox": [ + 133, + 237, + 307, + 282 + ], + "spans": [ + { + "bbox": [ + 133, + 237, + 307, + 282 + ], + "type": "image", + "image_path": "653bc4464f408a5eae3cef29b77f50385c27afa055311bbcb173256cb65d62d3.jpg" + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 134, + 283, + 290, + 297 + ], + "lines": [ + { + "bbox": [ + 134, + 283, + 290, + 297 + ], + "spans": [ + { + "bbox": [ + 134, + 283, + 290, + 297 + ], + "type": "text", + "content": "sal_scenery LoRA \n西湖，塔和瀑布，日出(West Lake, snow and waterfall, sunrise) \n江南，夏日湖畔的一村村庄(Jianxun, a village by the summer lake)" + } + ] + } + ], + "index": 7, + "angle": 0, + "type": "image_caption" + } + ], + "index": 5 + }, + { + "type": "image", + "bbox": [ + 308, + 237, + 480, + 282 + ], + "blocks": [ + { + "bbox": [ + 308, + 237, + 480, + 282 + ], + "lines": [ + { + "bbox": [ + 308, + 237, + 480, + 282 + ], + "spans": [ + { + "bbox": [ + 308, + 237, + 480, + 282 + ], + "type": "image", + "image_path": "e70f2a98840d160713a31812f6776e65287d45cc5c09fab109982deee3c7df53.jpg" + } + ] + } + ], + "index": 6, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 319, + 282, + 469, + 297 + ], + "lines": [ + { + "bbox": [ + 319, + 282, + 469, + 297 + ], + "spans": [ + { + "bbox": [ + 319, + 282, + 469, + 297 + ], + "type": "text", + "content": "realistic/easy/animalism/3d-carton Fine-tuning \n一个漂亮的双面的调色板 \n一个白色的小球与white painting light) \n3D电影，4K，高度细致，男人坐在马桶上读报（3D movie, 4K, highly detailed, man sitting on toilet" + } + ] + } + ], + "index": 8, + "angle": 0, + "type": "image_caption" + } + ], + "index": 6 + }, + { + "type": "image", + "bbox": [ + 133, + 298, + 306, + 343 + ], + "blocks": [ + { + "bbox": [ + 133, + 298, + 306, + 343 + ], + "lines": [ + { + "bbox": [ + 133, + 298, + 306, + 343 + ], + "spans": [ + { + "bbox": [ + 133, + 298, + 306, + 343 + ], + "type": "image", + "image_path": "a8089f7a1a6d5ab98bdc9585dbecd91621a231660164c9a4e348c3f4485139f5.jpg" + } + ] + } + ], + "index": 9, + "angle": 0, + "type": "image_body" + } + ], + "index": 9 + }, + { + "type": "image", + "bbox": [ + 306, + 299, + 480, + 343 + ], + "blocks": [ + { + "bbox": [ + 306, + 299, + 480, + 343 + ], + "lines": [ + { + "bbox": [ + 306, + 299, + 480, + 343 + ], + "spans": [ + { + "bbox": [ + 306, + 299, + 480, + 343 + ], + "type": "image", + "image_path": "92eafbe56876f3ab5d6e6e24da6cbffc31a80436d09be4bf5ec0c423707e123e.jpg" + } + ] + } + ], + "index": 10, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 318, + 345, + 465, + 359 + ], + "lines": [ + { + "bbox": [ + 318, + 345, + 465, + 359 + ], + "spans": [ + { + "bbox": [ + 318, + 345, + 465, + 359 + ], + "type": "text", + "content": "SD-XL, Inpainting 0.1 Painting \n2018年，由青岛双星玻璃有限公司（White porcelain plates on benches）一辆小型起重机放在了长椅上（A small crane was placed on the bench）" + } + ] + } + ], + "index": 12, + "angle": 0, + "type": "image_caption" + } + ], + "index": 10 + }, + { + "type": "image", + "bbox": [ + 133, + 362, + 306, + 407 + ], + "blocks": [ + { + "bbox": [ + 138, + 344, + 296, + 362 + ], + "lines": [ + { + "bbox": [ + 138, + 344, + 296, + 362 + ], + "spans": [ + { + "bbox": [ + 138, + 344, + 296, + 362 + ], + "type": "text", + "content": "controlnet-canny ControlNet 鸳鸯双栖展双飞，满园存色惹人醉(Mandarin ducks and butterflies fly together, and the spring scenery in the garden is intoxicating); 北国风光，千里冰封，万里雪飘(Northern scenery, thousands of miles of ice, thousands of miles of snow drifting)" + } + ] + } + ], + "index": 11, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 133, + 362, + 306, + 407 + ], + "lines": [ + { + "bbox": [ + 133, + 362, + 306, + 407 + ], + "spans": [ + { + "bbox": [ + 133, + 362, + 306, + 407 + ], + "type": "image", + "image_path": "b908fc72ce8f83e42fb39bf7c840e59a686cc5293d871901e0a81208305dabf5.jpg" + } + ] + } + ], + "index": 13, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 141, + 407, + 299, + 430 + ], + "lines": [ + { + "bbox": [ + 141, + 407, + 299, + 430 + ], + "spans": [ + { + "bbox": [ + 141, + 407, + 299, + 430 + ], + "type": "text", + "content": "Icm-lora-adsl CDM 带着墨镜的猫咪手里拿着剑，在恶魔城怪里，仙鸽奇侠风格(The cat wearing sunglasses holds a sword in its hand, in the Demon Castle, and the Magic Knight) 红色灯笼，长城，秋天，高清，奇幻风格，史密德(Red Lutron, Great Wall, Autumn, HD, Fantasy Style, Epic Sense)" + } + ] + } + ], + "index": 15, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 131, + 441, + 482, + 474 + ], + "lines": [ + { + "bbox": [ + 131, + 441, + 482, + 474 + ], + "spans": [ + { + "bbox": [ + 131, + 441, + 482, + 474 + ], + "type": "text", + "content": "Fig. 4: Downstream task experiments of our proposed PEA. By only adding the trained adapter between the Chinese CLIP text encoder and the UNet, it can achieve language transfer for fine-tuning, LoRA, ControlNet, Inpainting, LCM and compressed models." + } + ] + } + ], + "index": 17, + "angle": 0, + "type": "image_caption" + } + ], + "index": 13 + }, + { + "type": "image", + "bbox": [ + 306, + 361, + 479, + 406 + ], + "blocks": [ + { + "bbox": [ + 306, + 361, + 479, + 406 + ], + "lines": [ + { + "bbox": [ + 306, + 361, + 479, + 406 + ], + "spans": [ + { + "bbox": [ + 306, + 361, + 479, + 406 + ], + "type": "image", + "image_path": "d901f8901fab1ae63dab5501bb89c8203b5b262fded9199baef354028fd4479f.jpg" + } + ] + } + ], + "index": 14, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 314, + 407, + 471, + 429 + ], + "lines": [ + { + "bbox": [ + 314, + 407, + 471, + 429 + ], + "spans": [ + { + "bbox": [ + 314, + 407, + 471, + 429 + ], + "type": "text", + "content": "SSD-B I SSD \n一只超级可爱兔子穿着僧侣服装，肖像猫，皮克斯动画(A super cute rabbit wearing monk clothing, partition)。 \n中国汉奴女孩，阳光明媚，站在踏青花中；近景，摄影风景（Chinese Hanafi girl, sunny, standing in the midst of ruffled jeans, close-up, photography style）" + } + ] + } + ], + "index": 16, + "angle": 0, + "type": "image_caption" + } + ], + "index": 14 + }, + { + "bbox": [ + 130, + 498, + 482, + 666 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 498, + 482, + 666 + ], + "spans": [ + { + "bbox": [ + 130, + 498, + 482, + 666 + ], + "type": "text", + "content": "In this section, we would like to emphasize the plug-and-play feature of our PEA module. Currently, a wide range of community resources, such as fine-tuned models, LoRA, Dreambooth, ControlNet, compressed SD models, and accelerated sampling SD models, are derived from the English SD models, so that the English community is much more mature than other languages. Therefore, one key challenge we must overcome is bridging the compatibility gap between the new model and the English SD community after language transfer. Since our approach uses feature-level and logit-level KD to train only the PEA module while keeping the core UNet parameters unchanged, it naturally maintains compatibility with the English SD community and can seamlessly integrate with various English SD community plugins. Interestingly, the PEA module remains plug-and-play for models that have undergone changes in base SD UNet parameters in contrast to PEA-Diffusion but whose overall structure remains similar. This includes models such as fine-tuned SD and compressed SD trained using" + } + ] + } + ], + "index": 18 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 389, + 91, + 447, + 100 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 389, + 91, + 447, + 100 + ], + "spans": [ + { + "bbox": [ + 389, + 91, + 447, + 100 + ], + "type": "text", + "content": "PEA-diffusion" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 470, + 91, + 481, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 470, + 91, + 481, + 100 + ], + "spans": [ + { + "bbox": [ + 470, + 91, + 481, + 100 + ], + "type": "text", + "content": "13" + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 12 + }, + { + "para_blocks": [ + { + "bbox": [ + 130, + 116, + 480, + 151 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 116, + 480, + 151 + ], + "spans": [ + { + "bbox": [ + 130, + 116, + 480, + 151 + ], + "type": "text", + "content": "pruning and distillation techniques, among others. In the following, we will introduce several downstream image generation applications based on the SD or SDXL foundation model." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 130, + 152, + 487, + 199 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 152, + 487, + 199 + ], + "spans": [ + { + "bbox": [ + 130, + 152, + 487, + 199 + ], + "type": "text", + "content": "Fine-tuned checkpoint. We use two fine-tuned SDXL-based model checkpoints: \"xxmix9realistic\" and \"samaritan-3d-cartoon\". The \"xxmix9realistic\" model is trained on person, photorealistic, sexy, female, woman, girls, etc. The \"samaritan-3d-cartoon\" model is trained on anime, Disney, 3D, Pixar, etc." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 130, + 200, + 480, + 235 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 200, + 480, + 235 + ], + "spans": [ + { + "bbox": [ + 130, + 200, + 480, + 235 + ], + "type": "text", + "content": "LoRA. \"csal_scenery\" is a SDXL-based LoRA model that contains common ancient Chinese architectural forms such as pavilions, pagodas, towers, gardens, courtyards, bridges, temples, palaces, towns, etc." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 146, + 236, + 480, + 247 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 146, + 236, + 480, + 247 + ], + "spans": [ + { + "bbox": [ + 146, + 236, + 480, + 247 + ], + "type": "text", + "content": "ControlNet. \"controlnet-canny\" is trained on SDXL under canny condition." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 130, + 248, + 480, + 283 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 248, + 480, + 283 + ], + "spans": [ + { + "bbox": [ + 130, + 248, + 480, + 283 + ], + "type": "text", + "content": "Inpainting. SDXL Inpainting 0.1 is a latent T2I diffusion model capable of generating photorealistic images given any text input, with the extra capability of Inpainting the pictures by using a mask." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 130, + 284, + 480, + 319 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 284, + 480, + 319 + ], + "spans": [ + { + "bbox": [ + 130, + 284, + 480, + 319 + ], + "type": "text", + "content": "LCM. Latent Consistency Models [19] have achieved impressive performance in accelerating T2I generative tasks, producing high-quality images with minimal inference steps. \"lcm-lora-sdxl\" is an acceleration module based on SDXL." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 130, + 319, + 480, + 355 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 319, + 480, + 355 + ], + "spans": [ + { + "bbox": [ + 130, + 319, + 480, + 355 + ], + "type": "text", + "content": "SSD. The Segmind Stable Diffusion Model is a condensed version of the SDXL that is " + }, + { + "bbox": [ + 130, + 319, + 480, + 355 + ], + "type": "inline_equation", + "content": "50\\%" + }, + { + "bbox": [ + 130, + 319, + 480, + 355 + ], + "type": "text", + "content": " smaller in size, delivering a " + }, + { + "bbox": [ + 130, + 319, + 480, + 355 + ], + "type": "inline_equation", + "content": "60\\%" + }, + { + "bbox": [ + 130, + 319, + 480, + 355 + ], + "type": "text", + "content": " increase in speed while preserving high-quality text-to-image (T2I) generation capabilities." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 130, + 356, + 480, + 403 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 356, + 480, + 403 + ], + "spans": [ + { + "bbox": [ + 130, + 356, + 480, + 403 + ], + "type": "text", + "content": "In Figure 4, we have shown that the PEA module can be used as a plugin between the Chinese CLIP text encoder and the UNet, achieving the ability to transfer language from English to Chinese directly. For more details of the abovementioned checkpoints and generated images, please refer to our Appendix." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 131, + 422, + 312, + 434 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 131, + 422, + 312, + 434 + ], + "spans": [ + { + "bbox": [ + 131, + 422, + 312, + 434 + ], + "type": "text", + "content": "5 Conclusion and Limitation" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 130, + 446, + 481, + 566 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 446, + 481, + 566 + ], + "spans": [ + { + "bbox": [ + 130, + 446, + 481, + 566 + ], + "type": "text", + "content": "Conclusion. The high training expenses and limited training data available for multilingual T2I models have driven us to search for a more lightweight solution. In this paper, we introduce a straightforward and efficient method using KD and adapters. By focusing solely on training the adapter and keeping other parameters fixed, our approach harnesses the capability of T2I models for multilingual cultural comprehension and generation. Our method not only achieves significantly better results for culturally relevant data but also surpasses existing models while maintaining general synthesis capability. Comprehensive experiments demonstrate that our method is cost-effective, versatile, and easy to integrate." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 130, + 567, + 481, + 662 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 567, + 481, + 662 + ], + "spans": [ + { + "bbox": [ + 130, + 567, + 481, + 662 + ], + "type": "text", + "content": "**Limitation.** First, the effectiveness of the CLIP text encoder significantly influences both the quality of image generation and text relevance. Utilizing a weaker language-specific CLIP could impact the performance of PEA-Diffusion in broad open domains and may slightly compromise the model's generalization abilities. Second, the upper limit of our method is constrained by the capabilities of the English base model, indicating that our approach cannot exceed its inherent potential. Moving forward, we aim to conduct additional research to mitigate these limitations." + } + ] + } + ], + "index": 12 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "spans": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "type": "text", + "content": "14" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 166, + 91, + 215, + 101 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 166, + 91, + 215, + 101 + ], + "spans": [ + { + "bbox": [ + 166, + 91, + 215, + 101 + ], + "type": "text", + "content": "J. Ma et al." + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 13 + }, + { + "para_blocks": [ + { + "bbox": [ + 133, + 114, + 197, + 126 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 114, + 197, + 126 + ], + "spans": [ + { + "bbox": [ + 133, + 114, + 197, + 126 + ], + "type": "text", + "content": "References" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 138, + 141, + 480, + 665 + ], + "type": "list", + "angle": 0, + "index": 20, + "blocks": [ + { + "bbox": [ + 138, + 141, + 480, + 164 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 141, + 480, + 164 + ], + "spans": [ + { + "bbox": [ + 138, + 141, + 480, + 164 + ], + "type": "text", + "content": "1. Aggarwal, P., Kale, A.: Towards zero-shot cross-lingual image retrieval. arXiv preprint arXiv:2012.05107 (2020)" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 138, + 165, + 480, + 198 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 165, + 480, + 198 + ], + "spans": [ + { + "bbox": [ + 138, + 165, + 480, + 198 + ], + "type": "text", + "content": "2. Carlsson, F., Eisen, P., Rekathati, F., Sahlgren, M.: Cross-lingual and multilingual clip. In: Proceedings of the Thirteenth Language Resources and Evaluation Conference. pp. 6848-6854 (2022)" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 138, + 198, + 480, + 241 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 198, + 480, + 241 + ], + "spans": [ + { + "bbox": [ + 138, + 198, + 480, + 241 + ], + "type": "text", + "content": "3. Chen, G., Hou, L., Chen, Y., Dai, W., Shang, L., Jiang, X., Liu, Q., Pan, J., Wang, W.: mclip: Multilingual clip via cross-lingual transfer. In: Proceedings of the 61st Annual Meeting of the Association for Computational Linguistics (Volume 1: Long Papers). pp. 13028-13043 (2023)" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 138, + 242, + 480, + 264 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 242, + 480, + 264 + ], + "spans": [ + { + "bbox": [ + 138, + 242, + 480, + 264 + ], + "type": "text", + "content": "4. Chen, Z., Liu, G., Zhang, B.W., Ye, F., Yang, Q., Wu, L.: Altclip: Altering the language encoder in clip for extended language capabilities (2022)" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 138, + 264, + 480, + 298 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 264, + 480, + 298 + ], + "spans": [ + { + "bbox": [ + 138, + 264, + 480, + 298 + ], + "type": "text", + "content": "5. Conneau, A., Khandelwal, K., Goyal, N., Chaudhary, V., Wenzek, G., Guzmán, F., Grave, E., Ott, M., Zettlemoyer, L., Stoyanov, V.: Unsupervised cross-lingual representation learning at scale. arXiv preprint arXiv:1911.02116 (2019)" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 138, + 299, + 480, + 331 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 299, + 480, + 331 + ], + "spans": [ + { + "bbox": [ + 138, + 299, + 480, + 331 + ], + "type": "text", + "content": "6. Ding, M., Yang, Z., Hong, W., Zheng, W., Zhou, C., Yin, D., Lin, J., Zou, X., Shao, Z., Yang, H., et al.: Cogview: Mastering text-to-image generation via transformers. Advances in Neural Information Processing Systems 34, 19822-19835 (2021)" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 138, + 332, + 480, + 374 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 332, + 480, + 374 + ], + "spans": [ + { + "bbox": [ + 138, + 332, + 480, + 374 + ], + "type": "text", + "content": "7. Gu, J., Meng, X., Lu, G., Hou, L., Minzhe, N., Liang, X., Yao, L., Huang, R., Zhang, W., Jiang, X., et al.: Wukong: A 100 million large-scale chinese cross-modal pre-training benchmark. Advances in Neural Information Processing Systems 35, 26418-26431 (2022)" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 138, + 376, + 480, + 419 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 376, + 480, + 419 + ], + "spans": [ + { + "bbox": [ + 138, + 376, + 480, + 419 + ], + "type": "text", + "content": "8. Gu, S., Chen, D., Bao, J., Wen, F., Zhang, B., Chen, D., Yuan, L., Guo, B.: Vector quantized diffusion model for text-to-image synthesis. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 10696-10706 (2022)" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 138, + 420, + 480, + 442 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 420, + 480, + 442 + ], + "spans": [ + { + "bbox": [ + 138, + 420, + 480, + 442 + ], + "type": "text", + "content": "9. Hinton, G., Vinyals, O., Dean, J.: Distilling the knowledge in a neural network. arXiv preprint arXiv:1503.02531 (2015)" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 138, + 443, + 480, + 464 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 443, + 480, + 464 + ], + "spans": [ + { + "bbox": [ + 138, + 443, + 480, + 464 + ], + "type": "text", + "content": "10. Ho, J., Jain, A., Abbeel, P.: Denoising diffusion probabilistic models. Advances in Neural Information Processing systems 33, 6840-6851 (2020)" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 138, + 465, + 480, + 497 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 465, + 480, + 497 + ], + "spans": [ + { + "bbox": [ + 138, + 465, + 480, + 497 + ], + "type": "text", + "content": "1. Hou, L., Huang, Z., Shang, L., Jiang, X., Chen, X., Liu, Q.: Dynabert: Dynamic bert with adaptive width and depth. Advances in Neural Information Processing Systems 33, 9782-9793 (2020)" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 138, + 498, + 480, + 520 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 498, + 480, + 520 + ], + "spans": [ + { + "bbox": [ + 138, + 498, + 480, + 520 + ], + "type": "text", + "content": "2. Hu, E.J., Shen, Y., Wallis, P., Allen-Zhu, Z., Li, Y., Wang, S., Wang, L., Chen, W.: Lora: Low-rank adaptation of large language models (2021)" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 138, + 521, + 480, + 553 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 521, + 480, + 553 + ], + "spans": [ + { + "bbox": [ + 138, + 521, + 480, + 553 + ], + "type": "text", + "content": "3. Hu, J., Han, X., Yi, X., Chen, Y., Li, W., Liu, Z., Sun, M.: Efficient crosslingual transfer for chinese stable diffusion with images as pivots. arXiv preprint arXiv:2305.11540 (2023)" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 138, + 554, + 480, + 586 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 554, + 480, + 586 + ], + "spans": [ + { + "bbox": [ + 138, + 554, + 480, + 586 + ], + "type": "text", + "content": "4. Jiao, X., Yin, Y., Shang, L., Jiang, X., Chen, X., Li, L., Wang, F., Liu, Q.: Tinybert: Distilling bert for natural language understanding. arXiv preprint arXiv:1909.10351 (2019)" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 138, + 587, + 480, + 609 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 587, + 480, + 609 + ], + "spans": [ + { + "bbox": [ + 138, + 587, + 480, + 609 + ], + "type": "text", + "content": "5. Kirstain, Y., Polyak, A., Singer, U., Matiana, S., Penna, J., Levy, O.: Pick-a-pic: An open dataset of user preferences for text-to-image generation (2023)" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 138, + 610, + 480, + 643 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 610, + 480, + 643 + ], + "spans": [ + { + "bbox": [ + 138, + 610, + 480, + 643 + ], + "type": "text", + "content": "6. Kumari, N., Zhang, B., Zhang, R., Shechtman, E., Zhu, J.Y.: Multi-concept customization of text-to-image diffusion. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 1931-1941 (2023)" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 138, + 643, + 480, + 665 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 643, + 480, + 665 + ], + "spans": [ + { + "bbox": [ + 138, + 643, + 480, + 665 + ], + "type": "text", + "content": "7. Li, Y., Chang, C.Y., Rawls, S., Vulic, I., Korhonen, A.: Translation-enhanced multilingual text-to-image generation. arXiv preprint arXiv:2305.19216 (2023)" + } + ] + } + ], + "index": 19 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 389, + 91, + 447, + 100 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 389, + 91, + 447, + 100 + ], + "spans": [ + { + "bbox": [ + 389, + 91, + 447, + 100 + ], + "type": "text", + "content": "PEA-diffusion" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 470, + 92, + 481, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 470, + 92, + 481, + 100 + ], + "spans": [ + { + "bbox": [ + 470, + 92, + 481, + 100 + ], + "type": "text", + "content": "15" + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 14 + }, + { + "para_blocks": [ + { + "bbox": [ + 132, + 117, + 480, + 665 + ], + "type": "list", + "angle": 0, + "index": 19, + "blocks": [ + { + "bbox": [ + 133, + 117, + 480, + 138 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 117, + 480, + 138 + ], + "spans": [ + { + "bbox": [ + 133, + 117, + 480, + 138 + ], + "type": "text", + "content": "18. Liu, S., Leng, D., Yin, Y.: Bridge diffusion model: bridge non-english language-native text-to-image diffusion model with english communities (2023)" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 133, + 140, + 480, + 161 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 140, + 480, + 161 + ], + "spans": [ + { + "bbox": [ + 133, + 140, + 480, + 161 + ], + "type": "text", + "content": "19. Luo, S., Tan, Y., Huang, L., Li, J., Zhao, H.: Latent consistency models: Synthesizing high-resolution images with few-step inference (2023)" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 132, + 163, + 480, + 195 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 163, + 480, + 195 + ], + "spans": [ + { + "bbox": [ + 132, + 163, + 480, + 195 + ], + "type": "text", + "content": "20. Nichol, A., Dhariwal, P., Ramesh, A., Shyam, P., Mishkin, P., McGrew, B., Sutskever, I., Chen, M.: Glide: Towards photorealistic image generation and editing with text-guided diffusion models. arXiv preprint arXiv:2112.10741 (2021)" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 132, + 197, + 480, + 228 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 197, + 480, + 228 + ], + "spans": [ + { + "bbox": [ + 132, + 197, + 480, + 228 + ], + "type": "text", + "content": "21. Podell, D., English, Z., Lacey, K., Blattmann, A., Dockhorn, T., Müller, J., Penna, J., Rombach, R.: Sdxl: Improving latent diffusion models for high-resolution image synthesis (2023)" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 132, + 230, + 480, + 262 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 230, + 480, + 262 + ], + "spans": [ + { + "bbox": [ + 132, + 230, + 480, + 262 + ], + "type": "text", + "content": "22. Qin, C., Yu, N., Xing, C., Zhang, S., Chen, Z., Ermon, S., Fu, Y., Xiong, C., Xu, R.: Gluegen: Plug and play multi-modal encoders for x-to-image generation. arXiv preprint arXiv:2303.10056 (2023)" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 132, + 263, + 480, + 306 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 263, + 480, + 306 + ], + "spans": [ + { + "bbox": [ + 132, + 263, + 480, + 306 + ], + "type": "text", + "content": "23. Radford, A., Kim, J.W., Hallacy, C., Ramesh, A., Goh, G., Agarwal, S., Sastry, G., Askell, A., Mishkin, P., Clark, J., et al.: Learning transferable visual models from natural language supervision. In: International conference on machine learning. pp. 8748-8763. PMLR (2021)" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 132, + 308, + 480, + 329 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 308, + 480, + 329 + ], + "spans": [ + { + "bbox": [ + 132, + 308, + 480, + 329 + ], + "type": "text", + "content": "24. Ramesh, A., Dhariwal, P., Nichol, A., Chu, C., Chen, M.: Hierarchical text-conditional image generation with clip latents (2022)" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 132, + 331, + 480, + 363 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 331, + 480, + 363 + ], + "spans": [ + { + "bbox": [ + 132, + 331, + 480, + 363 + ], + "type": "text", + "content": "25. Rombach, R., Blattmann, A., Lorenz, D., Esser, P., Ommer, B.: High-resolution image synthesis with latent diffusion models. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 10684-10695 (2022)" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 132, + 365, + 480, + 385 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 365, + 480, + 385 + ], + "spans": [ + { + "bbox": [ + 132, + 365, + 480, + 385 + ], + "type": "text", + "content": "26. Rombach, R., Blattmann, A., Lorenz, D., Esser, P., Ommer, B.: High-resolution image synthesis with latent diffusion models (2022)" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 132, + 387, + 480, + 430 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 387, + 480, + 430 + ], + "spans": [ + { + "bbox": [ + 132, + 387, + 480, + 430 + ], + "type": "text", + "content": "27. Ronneberger, O., Fischer, P., Brox, T.: U-net: Convolutional networks for biomedical image segmentation. In: Medical Image Computing and Computer-Assisted Intervention-MICCAI 2015: 18th International Conference, Munich, Germany, October 5-9, 2015, Proceedings, Part III 18. pp. 234-241. Springer (2015)" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 132, + 432, + 480, + 474 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 432, + 480, + 474 + ], + "spans": [ + { + "bbox": [ + 132, + 432, + 480, + 474 + ], + "type": "text", + "content": "28. Saharia, C., Chan, W., Saxena, S., Li, L., Whang, J., Denton, E., Ghasemipour, S.K.S., Ayan, B.K., Mahdavi, S.S., Lopes, R.G., Salimans, T., Ho, J., Fleet, D.J., Norouzi, M.: Photorealistic text-to-image diffusion models with deep language understanding (2022)" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 132, + 475, + 480, + 518 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 475, + 480, + 518 + ], + "spans": [ + { + "bbox": [ + 132, + 475, + 480, + 518 + ], + "type": "text", + "content": "29. Schuhmann, C., Beaumont, R., Vencu, R., Gordon, C., Wightman, R., Cherti, M., Coombes, T., Katta, A., Mullis, C., Wortsman, M., Schramowski, P., Kundurthy, S., Crowson, K., Schmidt, L., Kaczmarczyk, R., Jitsev, J.: Laion-5b: An open large-scale dataset for training next generation image-text models (2022)" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 132, + 520, + 480, + 552 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 520, + 480, + 552 + ], + "spans": [ + { + "bbox": [ + 132, + 520, + 480, + 552 + ], + "type": "text", + "content": "30. Xu, J., Liu, X., Wu, Y., Tong, Y., Li, Q., Ding, M., Tang, J., Dong, Y.: Imagereward: Learning and evaluating human preferences for text-to-image generation (2023)" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 132, + 554, + 480, + 586 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 554, + 480, + 586 + ], + "spans": [ + { + "bbox": [ + 132, + 554, + 480, + 586 + ], + "type": "text", + "content": "31. Yang, A., Pan, J., Lin, J., Men, R., Zhang, Y., Zhou, J., Zhou, C.: Chinese clip: Contrastive vision-language pretraining in chinese. arXiv preprint arXiv:2211.01335 (2022)" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 132, + 587, + 480, + 609 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 587, + 480, + 609 + ], + "spans": [ + { + "bbox": [ + 132, + 587, + 480, + 609 + ], + "type": "text", + "content": "32. Yang, Z., Li, Z., Shao, M., Shi, D., Yuan, Z., Yuan, C.: Masked generative distillation. In: European Conference on Computer Vision. pp. 53-69. Springer (2022)" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 132, + 610, + 480, + 631 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 610, + 480, + 631 + ], + "spans": [ + { + "bbox": [ + 132, + 610, + 480, + 631 + ], + "type": "text", + "content": "33. Ye, F., Liu, G., Wu, X., Wu, L.: Altdiffusion: A multilingual text-to-image diffusion model (2023)" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 132, + 632, + 480, + 665 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 632, + 480, + 665 + ], + "spans": [ + { + "bbox": [ + 132, + 632, + 480, + 665 + ], + "type": "text", + "content": "34. Zhang, H., Yin, W., Fang, Y., Li, L., Duan, B., Wu, Z., Sun, Y., Tian, H., Wu, H., Wang, H.: Ernie-vilg: Unified generative pre-training for bidirectional vision-language generation. arXiv preprint arXiv:2112.15283 (2021)" + } + ] + } + ], + "index": 18 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "spans": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "type": "text", + "content": "16" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 166, + 91, + 215, + 101 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 166, + 91, + 215, + 101 + ], + "spans": [ + { + "bbox": [ + 166, + 91, + 215, + 101 + ], + "type": "text", + "content": "J. Ma et al." + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 15 + }, + { + "para_blocks": [ + { + "bbox": [ + 132, + 116, + 482, + 205 + ], + "type": "list", + "angle": 0, + "index": 4, + "blocks": [ + { + "bbox": [ + 132, + 116, + 482, + 172 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 116, + 482, + 172 + ], + "spans": [ + { + "bbox": [ + 132, + 116, + 482, + 172 + ], + "type": "text", + "content": "35. Zhang, J., Gan, R., Wang, J., Zhang, Y., Zhang, L., Yang, P., Gao, X., Wu, Z., Dong, X., He, J., Zhuo, J., Yang, Q., Huang, Y., Li, X., Wu, Y., Lu, J., Zhu, X., Chen, W., Han, T., Pan, K., Wang, R., Wang, H., Wu, X., Zeng, Z., Chen, C.: Fengshenbang 1.0: Being the foundation of chinese cognitive intelligence. CoRR abs/2209.02970 (2022)" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 132, + 172, + 481, + 205 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 172, + 481, + 205 + ], + "spans": [ + { + "bbox": [ + 132, + 172, + 481, + 205 + ], + "type": "text", + "content": "36. Zhang, J., Gan, R., Wang, J., Zhang, Y., Zhang, L., Yang, P., Gao, X., Wu, Z., Dong, X., He, J., et al.: Fengshenbang 1.0: Being the foundation of chinese cognitive intelligence. arXiv preprint arXiv:2209.02970 (2022)" + } + ] + } + ], + "index": 3 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 389, + 91, + 447, + 100 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 389, + 91, + 447, + 100 + ], + "spans": [ + { + "bbox": [ + 389, + 91, + 447, + 100 + ], + "type": "text", + "content": "PEA-diffusion" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 470, + 92, + 481, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 470, + 92, + 481, + 100 + ], + "spans": [ + { + "bbox": [ + 470, + 92, + 481, + 100 + ], + "type": "text", + "content": "17" + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 16 + } + ], + "_backend": "vlm", + "_version_name": "2.6.4" +} \ No newline at end of file diff --git a/2024/PFGS_ High Fidelity Point Cloud Rendering via Feature Splatting/685f2bf6-434e-4901-92a2-92656824eb88_content_list.json b/2024/PFGS_ High Fidelity Point Cloud Rendering via Feature Splatting/685f2bf6-434e-4901-92a2-92656824eb88_content_list.json new file mode 100644 index 0000000000000000000000000000000000000000..4373580fbec4cf4b3f7fcc8fe20ec31b83b10713 --- /dev/null +++ b/2024/PFGS_ High Fidelity Point Cloud Rendering via Feature Splatting/685f2bf6-434e-4901-92a2-92656824eb88_content_list.json @@ -0,0 +1,2231 @@ +[ + { + "type": "text", + "text": "PFGS: High Fidelity Point Cloud Rendering via Feature Splatting", + "text_level": 1, + "bbox": [ + 225, + 140, + 779, + 186 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Jiaxu Wang $^{1\\dagger}$ , Ziyi Zhang $^{1\\dagger}$ , Junhao He $^{1}$ , and Renjing Xu $^{1\\boxplus}$", + "bbox": [ + 251, + 210, + 750, + 228 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "$^{1}$ Hong Kong University of Science and Technology (GZ) jwang457@connect.hkust-gz.edu.cn {ziyizhang, junhaohe, renjingxu}@hkust-gz.edu.cn", + "bbox": [ + 310, + 238, + 691, + 282 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Abstract. Rendering high-fidelity images from sparse point clouds is still challenging. Existing learning-based approaches suffer from either hole artifacts, missing details, or expensive computations. In this paper, we propose a novel framework to render high-quality images from sparse points. This method first attempts to bridge the 3D Gaussian Splatting and point cloud rendering, which includes several cascaded modules. We first use a regressor to estimate Gaussian properties in a point-wise manner, the estimated properties are used to rasterize neural feature descriptors into 2D planes which are extracted from a multiscale extractor. The projected feature volume is gradually decoded toward the final prediction via a multiscale and progressive decoder. The whole pipeline experiences a two-stage training and is driven by our well-designed progressive and multiscale reconstruction loss. Experiments on different benchmarks show the superiority of our method in terms of rendering qualities and the necessities of our main components. $^{1}$", + "bbox": [ + 259, + 309, + 743, + 517 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Keywords: Point cloud rendering $\\cdot$ 3D Gaussian Splatting", + "bbox": [ + 261, + 529, + 660, + 544 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "1 Introduction", + "text_level": 1, + "bbox": [ + 215, + 565, + 375, + 580 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Synthesizing photorealistic images from given colored point clouds and arbitrary camera views can be extensively applied to various fields including virtual/augmented reality [14], robotic navigation [11], automatic driving [35, 43], etc, which is academically named point cloud rendering. Traditional methods to render views from point clouds are based on graphics transformations. They directly project 3D point clouds into 2D planes by using the camera parameters and z-buffer rasterization. However, the sparse distribution of points leads to bleeding surfaces and hole artifacts. Even though the graphics renderer can treat points as graphical primitives and diminish these issues by color blending, rough and simple geometries of primitives entail blurred images and missing details.", + "bbox": [ + 212, + 593, + 787, + 743 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "In recent years, learning-based methods [9, 10, 24] have become prevailing to address the drawbacks of conventional graphics-based approaches. These methods project points and corresponding feature descriptors to 2D planes, obtaining sparse images. Then they use UNet-like networks to complete and restore", + "bbox": [ + 212, + 744, + 787, + 805 + ], + "page_idx": 0 + }, + { + "type": "page_footnote", + "text": "$\\star \\dagger$ Co-first authors; $\\boxtimes$ Corresponding authors", + "bbox": [ + 217, + 811, + 563, + 825 + ], + "page_idx": 0 + }, + { + "type": "page_footnote", + "text": "1 Project page: https://github.com/Mercerai/PPGS", + "bbox": [ + 220, + 825, + 575, + 839 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "the low-quality images [12, 24]. Nevertheless, these methods rely on 2D image restoration, which cannot guarantee 3D consistency and can result in the distinctive appearance of the same object from nearby different viewpoints. Moreover, their feature descriptors are associated with points that do not possess actual shapes and volumes in real space, thereby showing poor descriptions of local area structures. To obtain 3D-aware and consistent features, some approaches [8, 16] transform points to 3D volumes by trilinear interpolation and use 3D convolution to extract the features. However, they lead to a significant computational burden and sacrifice the efficient representation characteristics of point clouds. For example, voxels uniformly represent scenes, whereas points are densely distributed in complex areas and sparsely distributed or even absent in empty spaces.", + "bbox": [ + 212, + 146, + 787, + 313 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "To obtain efficient and consistent 3D representation, the Neural Radiance Field (NeRF) [20] has gradually been incorporated into point cloud renderers [9, 10]. These methods either consider points as anchors of NeRF or transform points into NeRF-like representations, then use the physical-based volume rendering to obtain 3D-aware feature maps which are decoded into corresponding images. With the assistance of NeRF, they achieve high-quality consistencies across nearby views and plausible images. However, discrete NeRF requires heavy memory consumption resulting from storing high-dimensional feature vectors. In addition, NeRF typically requires lots of sampling points and multiple accesses to the neural network, which significantly increases the rendering time and makes it challenging to meet real-time requirements.", + "bbox": [ + 212, + 315, + 787, + 482 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "More recently, 3D Gaussian Splatting (3DGS) [15] introduces a new representation that formulates points as 3D Gaussians with learnable parameters including 3D position, color, opacity, and anisotropic covariance. 3DGS achieves more photorealistic and faster rendering than NeRF by applying Gaussian rasterization and $\\alpha$ -blending. Conventional 3D Gaussian learns the primitive parameters associated with an initialized point cloud from multi-view images by backpropagation, it relies on per-subject parameter optimizations for several tens minutes. It is therefore impractical to be applied to the case in which only point cloud is available due to the need for multiview images and backpropagations.", + "bbox": [ + 212, + 484, + 787, + 621 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "In this work, we fill the gap between point cloud render and the 3DGS technique by introducing a novel multiscale, feature-augmented, 3DGS-based point renderer. Compared with previous learning-based point renderers, our method yields higher-quality and 3D consistent results. Compared with conventional 3DGS, our method does not necessarily require multiview images or fine-tuning procedures and can be directly employed to point modality.", + "bbox": [ + 212, + 625, + 787, + 715 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "We leverage an efficient multiscale feature extractor in the proposed framework to extract features from point cloud input. This extractor considers multiscale local geometries to extract discriminative feature descriptors. Then a regressor is used for estimating 3D Gaussian parameters in a point-wise manner. Next, we not only render images by these estimated Gaussian attributes in a multiscale way but also render the multiscale feature maps by the same Gaussian properties. Finally, the rendered multiscale images and feature maps are fused by a multiscale and recurrent UNet-like network to gradually transform", + "bbox": [ + 212, + 719, + 787, + 839 + ], + "page_idx": 1 + }, + { + "type": "page_number", + "text": "2", + "bbox": [ + 217, + 114, + 228, + 126 + ], + "page_idx": 1 + }, + { + "type": "header", + "text": "J. Wang et al.", + "bbox": [ + 271, + 114, + 367, + 128 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "into the final results. To the best knowledge of us, this is the first point cloud renderer combined with 3DGS.", + "bbox": [ + 212, + 146, + 782, + 175 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Our main contributions can be summarized in the following:", + "bbox": [ + 238, + 176, + 673, + 191 + ], + "page_idx": 2 + }, + { + "type": "list", + "sub_type": "text", + "list_items": [ + "1. We proposed a novel lightweight approach called PFGS (Point Feature Gaussian Splitting) that bridges point cloud rendering and 3DGS. It produces higher-fidelity images from sparse point clouds than currently prevailing approaches.", + "2. We proposed the point-based multiscale and feature-augmented 3D Gaussian framework, which consists of three main components including a multiscale feature extractor, a Gaussian regressor, and a multiscale and recurrent decoder, accompanied by the well-designed multiscale and progressive reconstruction losses.", + "3. Extensive experiments and ablation studies on three datasets show the effectiveness and advantages of the proposed method over current works in terms of reconstruction quality." + ], + "bbox": [ + 222, + 200, + 784, + 382 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "2 Related Work", + "text_level": 1, + "bbox": [ + 215, + 405, + 385, + 421 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "2.1 Point Cloud Render", + "text_level": 1, + "bbox": [ + 215, + 436, + 426, + 450 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Conventional point cloud rendering is implemented via graphics-based transformation [25,45] which simulates camera imaging processes to project points onto the 2D screens. Even though the rendering pipeline is fast and general for all scenes, it is heavily dependent on the density of point distributions. These methods produce empty holes and vacant areas when the distributions are sparse.", + "bbox": [ + 212, + 462, + 784, + 537 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Learning-based point renderers [3, 9, 10, 12, 24] to some extent handle the issues caused by traditional methods, which fill the holes by leveraging multiscale or multiview feature fusions. ME [3] implements a sparse Convolutional Neural Network to extract features from existing points and then calculates the features of arbitrary 3D points through ball querying. NPBG [12] enhances each point with a feature descriptor to encapsulate local geometry and appearance. Moreover, $\\mathrm{NPBG}++$ [24] introduces feature aggregation from multiview based on the NPBG to complement missing parts. NPCR [5] employs a 3DCNN to generate 3D volumes from point clouds and creates multiple depth layers to synthesize images. More recently, Point2pix [10] and Trivol [9] have incorporated NeRF into the point cloud rendering pipeline. They transform raw point clouds into NeRF-related representations and utilize volume rendering to access the representations to produce images. However, the above methods still cannot fill large vacant holes and only produce blurry rendering results, especially in local areas with fewer points. Even if incorporating NeRF representation is effective in generating 3D consistent feature encoding, they do not fully utilize the geometric information contained within the point clouds. In this work, we combine point rendering with a novel 3D representation, i.e. 3DGS, and implement multiscale and progressive feature decoding to obtain fine-grained and 3D-consistent predictions.", + "bbox": [ + 212, + 537, + 785, + 839 + ], + "page_idx": 2 + }, + { + "type": "header", + "text": "PFGS", + "bbox": [ + 687, + 114, + 730, + 126 + ], + "page_idx": 2 + }, + { + "type": "page_number", + "text": "3", + "bbox": [ + 774, + 116, + 785, + 126 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "2.2 Neural-based View Synthesis", + "text_level": 1, + "bbox": [ + 215, + 146, + 501, + 161 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Unlike traditional explicit representations [22, 23, 32, 33], implicit representations that encode scenes with the assistance of neural networks have been developed in recent years [18, 26, 38]. Among them, the Neural Radiance Field (NeRF) [20] attracts the most attention in both industries and research. NeRF is a continuous method that implicitly represents a 3D scene employing an MLP and synthesizing images of novel view by volume rendering. Recent studies have employed NeRF for various purposes, including dynamic reconstruction [1, 37, 40], physical reasoning [29], generative models [19, 28, 30], and fast rendering [6, 21, 34]. To achieve various goals, some studies attempted to combine the implicit NeRF with explicit primitives. For example, NeuMesh [41] regards mesh as a neural scaffold to store NeRF and achieve neural editing. PointNeRF [39] is the first to incorporate point clouds into NeRF reconstruction to obtain better rendering qualities and faster speed. Ref-NeuS [7] implements generalizable neural surface reconstruction. GPF [31] explicitly models the visibility by using the geometries in the point cloud. PointNeRF++ [27] solves point cloud sparsity by aggregating multiscale point clouds on sparse voxel grids at different resolutions.", + "bbox": [ + 212, + 167, + 787, + 409 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "It can be seen that many approaches we list above incorporate points into NeRF. However, NeRF-like volume rendering does not fully explore the rendering capability of the point cloud. In contrast to NeRF, 3DGS [15] proposed the 3D Gaussian Splatting to realize a remarkable differentiable rendering speed and yield high-quality 2D images. In this work, we first bridge point cloud rendering and 3DGS to generate higher-fidelity images.", + "bbox": [ + 212, + 409, + 787, + 501 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "3 Methodology", + "text_level": 1, + "bbox": [ + 215, + 518, + 380, + 537 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Given a colored point cloud $\\mathbf{P} = \\{\\mathbf{p}_k, \\mathbf{c}_k\\}$ , we aim to synthesize photorealistic images from any viewpoints defined by the camera intrinsic $K$ and poses $P$ are provided. The $\\mathbf{p}_k \\in R^3$ and $\\mathbf{c}_k \\in R^3$ are the point coordinates and colors respectively. The rendering process can be illustrated as $I_v = Re(\\mathbf{P}|K_v, P_v)$ where $Re$ is the rendering function that can be implemented by either graphics-based or learning-based methods as we introduced above. In this work, it is implemented by our proposed multiscale feature-based 3DGS rendering pipeline that contains a multiscale feature extractor, a Gaussian regressor, a feature-based Gaussian rendering module, and a multiscale recurrent decoder. All modules are implemented with small networks. In the following of this section, we sequentially introduce the aforementioned components.", + "bbox": [ + 212, + 547, + 787, + 715 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "3.1 Preliminaries", + "text_level": 1, + "bbox": [ + 215, + 732, + 372, + 747 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "3DGS is an explicit representation method for 3D scenes, which parameterizes the scene as a series of 3D Gaussian primitives. A 3D Gaussian is defined by a full 3D covariance matrix $\\Sigma$ , its center point $x$ , and the spherical harmonic $(SH)$ . The mean value of the Gaussian is defined as:", + "bbox": [ + 212, + 753, + 785, + 815 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\nG (x) = e ^ {- \\frac {1}{2} (x) ^ {T} \\Sigma^ {- 1} (x)} \\tag {1}\n$$\n", + "text_format": "latex", + "bbox": [ + 424, + 821, + 785, + 840 + ], + "page_idx": 3 + }, + { + "type": "page_number", + "text": "4", + "bbox": [ + 217, + 114, + 228, + 126 + ], + "page_idx": 3 + }, + { + "type": "header", + "text": "J. Wang et al.", + "bbox": [ + 271, + 114, + 367, + 128 + ], + "page_idx": 3 + }, + { + "type": "image", + "img_path": "images/50f7342bfdb9038ce4d05ff8d6c3ccdeebee297f41a9c9ee71b730e700512a80.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 217, + 143, + 787, + 301 + ], + "page_idx": 4 + }, + { + "type": "image", + "img_path": "images/36569bd060fabdb2b2ff0f2eab6037a3db14d1a5df81990b0b4da9e711d5ce22.jpg", + "image_caption": [ + "Fig. 1: The main pipeline of the proposed approach. The above panel includes the multiscale feature extractor and Gaussian regressor. Both Gaussian and features are splatted to the 2D plane and concatenated and fed to the progressive and multiscale feature decoding module described in the low panel." + ], + "image_footnote": [], + "bbox": [ + 228, + 301, + 784, + 463 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "To enable optimization via backpropagation, the covariance matrix could be decomposed into a rotation matrix $(R)$ and a scaling matrix $(S)$ :", + "bbox": [ + 212, + 539, + 787, + 570 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\n\\Sigma = R S S ^ {T} R ^ {T} \\tag {2}\n$$\n", + "text_format": "latex", + "bbox": [ + 446, + 578, + 785, + 597 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Given the camera trajectory, the projection of the 3D Gaussians to the 2D image plane can be characterized by a view transform matrix $(W)$ and the Jacobian of the affine approximation of the projective transformation $(J)$ , as in:", + "bbox": [ + 212, + 607, + 787, + 654 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\n\\Sigma^ {\\prime} = J W \\Sigma W ^ {T} J ^ {T} \\tag {3}\n$$\n", + "text_format": "latex", + "bbox": [ + 431, + 662, + 785, + 681 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "where the $\\Sigma^{\\prime}$ is the covariance matrix in 2D image planes. Thus, the $\\alpha$ -blend of $\\mathcal{N}$ ordered points overlapping a pixel is utilized to compute the final color $C$ of the pixel:", + "bbox": [ + 212, + 691, + 787, + 736 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\nC = \\sum_ {i \\in \\mathcal {N}} c _ {i} \\alpha_ {i} \\prod_ {j = 1} ^ {i - 1} (1 - \\alpha_ {j}) \\tag {4}\n$$\n", + "text_format": "latex", + "bbox": [ + 408, + 737, + 785, + 777 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "where $c_{i}$ and $\\alpha_{i}$ denote the color and density of the pixel with corresponding Gaussian parameters. Following [46], we projected the features similarly.", + "bbox": [ + 212, + 779, + 785, + 809 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "In summary, the above parameters can be characterized by the following attributes: $x$ denotes the position of a primitive where $x \\in R^3$ . The rotation", + "bbox": [ + 212, + 810, + 785, + 839 + ], + "page_idx": 4 + }, + { + "type": "header", + "text": "PFGS", + "bbox": [ + 687, + 114, + 730, + 126 + ], + "page_idx": 4 + }, + { + "type": "page_number", + "text": "5", + "bbox": [ + 774, + 116, + 785, + 126 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "matrix is characterized by a quaternion $q \\in R^4$ . The scale factor is defined by the anisotropy stretching $s \\in R^3$ . The opacity factor $\\alpha \\in R^1 \\in [0,1]$ . The 2D opacity $\\alpha \\in [0,1]$ is computed by $\\alpha_i(x) = o_i \\exp(-\\frac{1}{2} (x - \\mu_i)^T \\Sigma_i^T (x - \\mu_i))$ where the $\\mu$ and variance are the 2D-projected mean and variance of 3D Gaussians. And the color is represented by $SH$ .", + "bbox": [ + 212, + 146, + 787, + 223 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "3.2 Gaussian Feature Prediction with Multiscale Feature Extraction", + "text_level": 1, + "bbox": [ + 214, + 244, + 785, + 258 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "We first extract features for each point as their neural descriptors. We use a multi-input, single-output UNet-like architecture as the extraction network to encode points with different scales and capture features across different spatial distances. The architecture is mainly made of several PV Convolution modules [17], which is a highly efficient and commonly used point-based feature integration operation. We downsample the original point cloud with decreasing rates and concatenate the downsampled point clouds with different levels of feature maps as extra inputs.", + "bbox": [ + 212, + 268, + 787, + 390 + ], + "page_idx": 5 + }, + { + "type": "equation", + "text": "\n$$\nf _ {1: k} = \\mathcal {E} _ {\\theta} \\left(\\left(p _ {1: k}, c _ {1: k}\\right), \\left(p _ {1: k}, c _ {1: k}\\right) _ {\\downarrow r _ {1}}, \\left(p _ {1: k}, c _ {1: k}\\right) _ {\\downarrow r _ {2}}\\right) \\tag {5}\n$$\n", + "text_format": "latex", + "bbox": [ + 328, + 402, + 785, + 419 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "The $\\mathcal{E}_{\\theta}$ refers to the extractor network, $p, c$ denote point coordinates and colors. The down arrow represents a uniform downsampling operation and $r$ denotes the sampling rate where $r_1 > r_2$ . The output $f_{1:k}$ are point neural descriptors. This operation enables the extractor to recognize features of different scales and receptive fields. The detailed architecture and parameter setting are shown in our Appendix. The single output feature vector is considered as the descriptors for all points. Afterward, the neural descriptors, the coordinates, as well as colors are fed to another network to predict the Gaussian properties in a point-wise manner.", + "bbox": [ + 212, + 430, + 787, + 563 + ], + "page_idx": 5 + }, + { + "type": "equation", + "text": "\n$$\nF _ {k} = \\mathcal {R} _ {\\theta} \\left(f _ {k}, p _ {k}, c _ {k}\\right) \\tag {6}\n$$\n", + "text_format": "latex", + "bbox": [ + 325, + 564, + 785, + 588 + ], + "page_idx": 5 + }, + { + "type": "equation", + "text": "\n$$\nR, S, \\alpha = \\operatorname {n o r m} \\left(\\mathcal {H} _ {r} \\left(F _ {k}\\right)\\right), \\exp \\left(\\mathcal {H} _ {s} \\left(F _ {k}\\right)\\right), \\sigma \\left(\\mathcal {H} _ {o} \\left(F _ {k}\\right)\\right)\n$$\n", + "text_format": "latex", + "bbox": [ + 325, + 582, + 678, + 597 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "where $\\mathcal{R}_{\\theta}$ represents the main network of the regressor, it maps multiple inputs to a single unified feature vector $F_{k}$ . The $F_{k}$ is fed to three individual prediction heads $(\\mathcal{H}_r,\\mathcal{H}_s,\\mathcal{H}_o)$ , each of them is composed of two 1D Convolution layers with 1 kernel size, and activated by different functions to obtain corresponding Gaussian properties, namely rotation quaternion $(R)$ , scale factor $(S)$ and opacity $(\\alpha)$ . The color property $c$ remains that of the original points.", + "bbox": [ + 212, + 603, + 787, + 694 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "After the Gaussian attributes are obtained, one can render corresponding images from given camera viewpoints by Gaussian rasterization. We not only render the images by blending point colors on 2D planes but also render feature planes by replacing colors with the neural descriptors $(f_{k})$ .", + "bbox": [ + 212, + 694, + 787, + 755 + ], + "page_idx": 5 + }, + { + "type": "equation", + "text": "\n$$\nI _ {g s}, F _ {g s} = R e (R, S, \\alpha , p _ {1: k}, c _ {1: k} | K, P), R e (R, S, \\alpha , p _ {1: k}, f _ {1: k} | K, P) \\tag {7}\n$$\n", + "text_format": "latex", + "bbox": [ + 276, + 767, + 785, + 782 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "where $I_{gs}$ , $F_{gs}$ are corresponding image and feature map under the camera intrinsic $K$ and pose $P$ . The $I_{gs}$ can be considered a good result of point rendering when the number of points is sufficient.", + "bbox": [ + 212, + 794, + 787, + 839 + ], + "page_idx": 5 + }, + { + "type": "page_number", + "text": "6", + "bbox": [ + 217, + 114, + 228, + 126 + ], + "page_idx": 5 + }, + { + "type": "header", + "text": "J. Wang et al.", + "bbox": [ + 271, + 114, + 367, + 128 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "However, when the point sparsely distributes over the space, the representation capability of corresponding Gaussians is constrained, especially in complex local areas, resulting in poor-quality rendering results. On the contrary, the neural descriptors contain rich information to describe local structures with various scales. We therefore utilize the $F_{gs}$ to be gradually transformed to the final results in a multiscale and recurrent manner. The $F_{gs}$ is 3D-consistently rasterized with the same Gaussian parameters in Eq. 7, thus it precisely corresponds to the $I_{gs}$ in pixel-level. In other words, $F_{gs}$ can be regarded as a complement to $I_{gs}$ . Hence we concatenate $I_{gs}$ and $F_{gs}$ to be the input of the next module.", + "bbox": [ + 212, + 146, + 787, + 287 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "3.3 Multiscale and Progressive Feature Decoding", + "text_level": 1, + "bbox": [ + 214, + 310, + 635, + 327 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "After we render the feature volume $F_{gs}$ associated with the provided camera viewpoint, we concatenate the $I_{gs}$ and $F_{gs}$ as input of the decoder. The decoder is a multi-input and multi-output UNet-like architecture, which aims to gradually refine and decode the input features and transform them into the target view images. We downsample the input two times with the rate of $r_a$ and $r_b$ and fuse them with the next two low-resolution layers of the UNet. There are interlaced connections between layers of different levels in the network, as stated in the main Fig. 1.", + "bbox": [ + 212, + 340, + 787, + 464 + ], + "page_idx": 6 + }, + { + "type": "equation", + "text": "\n$$\n\\mathcal {F}, \\mathcal {F} _ {a}, \\mathcal {F} _ {b} = \\mathcal {D} _ {\\theta} \\left(\\left(F _ {g s} \\oplus I _ {g s}\\right), \\left(F _ {g s} \\oplus I _ {g s}\\right) _ {\\downarrow r _ {a}}, \\left(F _ {g s} \\oplus I _ {g s}\\right) _ {\\downarrow r _ {b}}\\right) \\tag {8}\n$$\n", + "text_format": "latex", + "bbox": [ + 297, + 477, + 785, + 494 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "In the above equation, $\\mathcal{D}_{\\theta}$ indicates the multiscale decoder. $\\downarrow r_{a}$ and $\\downarrow r_{b}$ are two downsampling operations. $\\mathcal{F},\\mathcal{F}_a,\\mathcal{F}_b$ refer to the feature output of different levels of layers in which $\\mathcal{F}$ denotes the full resolution feature output while $\\mathcal{F}$ with subscripts $a$ and $b$ represent the lower resolution feature outputs. We apply a prediction head (marked as $\\mathcal{H}_{\\eta}$ ) composed of two convolutional layers and a ReLU activation to transform these features into the target images with different scales.", + "bbox": [ + 212, + 507, + 787, + 611 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "More importantly, we recurrently run the decoder network twice to refine the feature maps and the predicted images. The following experiments show the positive influence of such recurrent refinement on restoring the missing details due to the sparse point distribution. The input for the second loop is related to the output of the first loop. The prediction after the first loop can be described as $I^{l}, I_{a}^{l}, I_{b}^{l} = \\mathcal{H}_{\\eta}(\\mathcal{F}), \\mathcal{H}_{\\eta}(\\mathcal{F}_{a}), \\mathcal{H}_{\\eta}(\\mathcal{F}_{b})$ where $l = 1$ refers to the loop number. Likewise, the second loop is accordingly explained as:", + "bbox": [ + 212, + 614, + 787, + 720 + ], + "page_idx": 6 + }, + { + "type": "equation", + "text": "\n$$\nI ^ {l = 2}, I _ {a} ^ {l = 2}, I _ {b} ^ {l = 2} = \\mathcal {H} _ {\\eta} \\star \\mathcal {D} _ {\\theta} ((\\mathcal {F} \\oplus I ^ {l = 1}), (\\mathcal {F} _ {a} \\oplus I _ {a} ^ {l = 1}), (\\mathcal {F} _ {b} \\oplus I _ {b} ^ {l = 1})) \\tag {9}\n$$\n", + "text_format": "latex", + "bbox": [ + 277, + 732, + 785, + 752 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "in which $\\mathcal{H}_{\\eta} \\star \\mathcal{D}_{\\theta}$ represents sequentially perform the two operations. We show the detailed architecture of the decoder in the Appendix. After we obtain the hierarchically multiscale predictions, we regard the full resolution result of the second loop $I^{l=2}$ as the final prediction. However, we leverage all predictions to optimize the full models, which we introduce in the next section.", + "bbox": [ + 212, + 763, + 787, + 840 + ], + "page_idx": 6 + }, + { + "type": "header", + "text": "PFGS", + "bbox": [ + 687, + 114, + 730, + 126 + ], + "page_idx": 6 + }, + { + "type": "page_number", + "text": "7", + "bbox": [ + 774, + 116, + 785, + 126 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "3.4 Training Strategies", + "text_level": 1, + "bbox": [ + 215, + 146, + 419, + 161 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "The proposed framework incorporates 3D-aware features into multiscale and progressive decoding by using 3D Gaussian Splatting. The integration of feature volumes under certain camera settings mainly relies on the Gaussian properties that are used for projections. Therefore, we first pretrain the feature extractor and Gaussian regressor in order to obtain a relatively reasonable rendering function with 8 epochs. In the first training stage, we employ the $L2$ rendering loss between the Gaussian predicted images and labels, i.e. $I_{gs}$ in Eq. 7.", + "bbox": [ + 212, + 179, + 787, + 287 + ], + "page_idx": 7 + }, + { + "type": "equation", + "text": "\n$$\n\\mathcal {L} _ {g s} = \\sum_ {\\mathbf {u} \\in \\Omega} \\frac {1}{\\Omega} \\left| \\left| I _ {g s} (\\mathbf {u}) - I _ {g t} (\\mathbf {u}) \\right| \\right| ^ {2} \\tag {10}\n$$\n", + "text_format": "latex", + "bbox": [ + 388, + 300, + 785, + 335 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "where $\\mathbf{u} = \\{u,v\\}$ and $\\Omega$ is the pixel domain. After the first pretraining, the regressor can yield plausible Gaussian rendering parameters, thus the rendered feature maps provide a certain and view-dependent description of the scene. Based on this, we combine the rest modules with these trained modules to carry out joint optimization. According to the above illustration, the final outputs of the framework include two levels, each comprising three images of different scales. We leverage all of the outputs to optimize the model by using progressively multiscale image loss.", + "bbox": [ + 212, + 349, + 787, + 470 + ], + "page_idx": 7 + }, + { + "type": "equation", + "text": "\n$$\n\\mathcal {L} _ {m i m} = \\sum_ {i = l} ^ {2} \\sum_ {s = r} ^ {3} \\sum_ {\\mathbf {u} \\in \\Omega} w (l) \\frac {1}{\\Omega} \\left| \\left| I _ {s} ^ {i} (\\mathbf {u}) - I _ {g t \\downarrow r _ {s}} (\\mathbf {u}) \\right| \\right| _ {1} ^ {1} \\tag {11}\n$$\n", + "text_format": "latex", + "bbox": [ + 333, + 487, + 785, + 527 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "in which $l$ refers to the loop number in the progressive decoder. $s$ denotes the scale number in each output of a loop and $\\downarrow r_s$ is $s$ ' corresponding downsampling operations. $w(l)$ is a simple weighting function related to $l$ . Here we set $w(l) = 0.75$ when $l = 1$ otherwise $w(l) = 1$ . In addition, we use the extra progressively multiscale frequency reconstruction loss to assist in restoring the high-frequency components, it is effective in reconstructing sharper edges and boundaries of images.", + "bbox": [ + 212, + 541, + 787, + 647 + ], + "page_idx": 7 + }, + { + "type": "equation", + "text": "\n$$\n\\mathcal {L} _ {m f r} = \\sum_ {i = l} ^ {2} \\sum_ {s = r} ^ {3} \\sum_ {\\mathbf {u} \\in \\Omega} w (l) \\frac {1}{\\Omega} | | F F T \\left(I _ {s} ^ {i} (\\mathbf {u})\\right) - F F T \\left(I _ {g t _ {\\downarrow r _ {s}}} (\\mathbf {u})\\right) | | _ {1} ^ {1} \\tag {12}\n$$\n", + "text_format": "latex", + "bbox": [ + 285, + 662, + 785, + 702 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "in which the $FFT$ represents the fast Fourier transform (FFT) that transfers the image signal to the frequency domain. The remaining symbols retain the same meanings as in the previous context. The total loss function can be defined as follows:", + "bbox": [ + 212, + 715, + 787, + 775 + ], + "page_idx": 7 + }, + { + "type": "equation", + "text": "\n$$\n\\mathcal {L} _ {\\text {t o t a l}} = \\gamma_ {1} \\mathcal {L} _ {g s} + \\gamma_ {2} \\mathcal {L} _ {\\text {m i m}} + \\gamma_ {3} \\mathcal {L} _ {\\text {m f r}} \\tag {13}\n$$\n", + "text_format": "latex", + "bbox": [ + 377, + 784, + 785, + 799 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "The $\\gamma_{1},\\gamma_{2}$ , and $\\gamma_{3}$ respectively control weights of these losses. In our implementation, we set $\\gamma_{1} = 0.75$ , $\\gamma_{2} = 1$ , and $\\gamma_{3} = 0.25$ respectively.", + "bbox": [ + 212, + 809, + 785, + 840 + ], + "page_idx": 7 + }, + { + "type": "page_number", + "text": "8", + "bbox": [ + 217, + 114, + 228, + 126 + ], + "page_idx": 7 + }, + { + "type": "header", + "text": "J. Wang et al.", + "bbox": [ + 271, + 114, + 367, + 128 + ], + "page_idx": 7 + }, + { + "type": "table", + "img_path": "images/a03e83b1e5756eed67a97c7dec4942e1bf6254f8a4ee2bdc78a96ce7b67eaefd.jpg", + "table_caption": [ + "Table 1: Quantitative Results of comparisons between our method and different baselines under three datasets." + ], + "table_footnote": [], + "table_body": "
ScanNetDTUTHuman2.0
PSNR↑SSIM↑LPIPS↓PSNR↑SSIM↑LPIPS↓PSNR↑SSIM↑LPIPS↓
Graphics13.620.5280.77912.150.5250.68220.260.9050.337
NPBG15.090.5920.62513.520.7030.51419.770.9150.112
NPBG++16.810.6710.58522.320.8330.32726.810.9520.062
Trivol18.560.7340.47319.250.5920.51825.970.9350.059
Ours19.860.7580.45225.440.9010.16434.740.9830.009
", + "bbox": [ + 215, + 179, + 797, + 303 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "4 Experiments", + "text_level": 1, + "bbox": [ + 215, + 321, + 374, + 339 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "4.1 Experiments setup", + "text_level": 1, + "bbox": [ + 215, + 351, + 415, + 367 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "Datasets. We conduct extensive experiments to evaluate the effectiveness of our approach on three different datasets, including the indoor scene dataset ScanNet [4], objects dataset DTU [13], and human body dataset THuman2.0 [42]. The ScanNet dataset includes more than 2.5 million images at different views and with different qualities in over 1500 scenes. The point clouds are constructed from RGBD images. Following the Trivol, [9] we select the first 1200 scenes as the training set and the rest of the $300+$ scenes as the test set. In addition, we use the preprocessed sparse point cloud in the dataset to test our model. For the DTU dataset, we transformed the point clouds of objects from their depth maps and images and we downsampled the point cloud by a factor of 0.5 to keep their point clouds distributing sparsely. In addition, we follow the MVSNeRF [2] to split the train and test groups of scans. The THuman2.0 dataset contains 500 high-quality human scans captured by a dense DLSR rig. For each scan, we render 36 views based on the provided 3D model and sparse sample points on the surface of the model. We keep the number of points for each scan at around 80,000 for their sparsity. We randomly select $75\\%$ as the training set and the remaining $25\\%$ as the test set.", + "bbox": [ + 215, + 375, + 785, + 628 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "Baselines and Metrics. We compare the proposed method with the currently prevailing point cloud rendering approaches to demonstrate the effectiveness of our framework. The baselines include a pure graphics-based renderer, learning-based renderers NPBG [12] and $\\mathrm{NPBG + + }$ [24], and the NeRF-based renderer Trivol [9]. As for the quality, we adopt PSNR, SSIM [36], and LPIPS [44]. These metrics overall determine the reconstruction accuracy and quality of these methods.", + "bbox": [ + 215, + 631, + 785, + 734 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "4.2 Comparisons With Other Point Cloud Renderers.", + "text_level": 1, + "bbox": [ + 215, + 756, + 669, + 771 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "Fig. 2, 3, and 4 show the results of qualitative comparisons on the ScanNet, DTU, and THuman2.0 datasets respectively. Table. 1 reports all the metrics on the three datasets of the baselines and ours. The best result is bold, and the second-best result is underlined in the Table.", + "bbox": [ + 215, + 779, + 785, + 838 + ], + "page_idx": 8 + }, + { + "type": "header", + "text": "PFGS", + "bbox": [ + 687, + 114, + 730, + 126 + ], + "page_idx": 8 + }, + { + "type": "page_number", + "text": "9", + "bbox": [ + 774, + 116, + 785, + 126 + ], + "page_idx": 8 + }, + { + "type": "image", + "img_path": "images/3a34bd6b3fd9041c6d8e1ed248ad047bc46b3ab79bf4cf9b2369f3e0eddcddbf.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 220, + 143, + 331, + 213 + ], + "page_idx": 9 + }, + { + "type": "image", + "img_path": "images/cc43a5f46defb72c8955dc67f03d48a9b06b620baa4aae84a5c5e1540f9f458e.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 334, + 145, + 444, + 214 + ], + "page_idx": 9 + }, + { + "type": "image", + "img_path": "images/0e433f96003de660997196971492f2bef15837eb533c1ff6ac7ff23d93edb4ed.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 446, + 145, + 557, + 214 + ], + "page_idx": 9 + }, + { + "type": "image", + "img_path": "images/2c0712f0f2820c8912d7cf09bcffbd8a5371e20d0466ccdbe0b444cf4ef1553a.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 558, + 145, + 669, + 214 + ], + "page_idx": 9 + }, + { + "type": "image", + "img_path": "images/3aa2436bbe65eeb2d38718659cd0200d0a186fa5db137bfa181349de7cc5cb7f.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 671, + 145, + 784, + 214 + ], + "page_idx": 9 + }, + { + "type": "image", + "img_path": "images/ce1adff760ab46ca04475fe5c6e4dcc75c63810fc9735f4e422f138c07f48382.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 220, + 215, + 331, + 284 + ], + "page_idx": 9 + }, + { + "type": "image", + "img_path": "images/3c2239014b8bb4b7c81584ecc721b4d4a47e7f66f8cc6f11197651a6db91a835.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 334, + 215, + 444, + 282 + ], + "page_idx": 9 + }, + { + "type": "image", + "img_path": "images/c003333cb0a57cb91b64ce8ea17f5bdb7a23d41c5907b55cf4ac76404157cb91.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 446, + 215, + 557, + 282 + ], + "page_idx": 9 + }, + { + "type": "image", + "img_path": "images/74dd1159eda6cf5c82ba31c9cec47f604215945d21a803177df86e7777b17116.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 558, + 215, + 669, + 282 + ], + "page_idx": 9 + }, + { + "type": "image", + "img_path": "images/e378b7e77f50b3b7e5701fd3be0429b08665b7bf869da15122cc949dd283e621.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 671, + 215, + 784, + 282 + ], + "page_idx": 9 + }, + { + "type": "image", + "img_path": "images/dde24abb49cd11fba5415acd11a6ba4813a54220966a8cfffe6c583f51842201.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 220, + 284, + 331, + 352 + ], + "page_idx": 9 + }, + { + "type": "image", + "img_path": "images/74f1c471e86b1611e32089b78b6dfb08a62e40c8decb61a38de026fcb886c80c.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 334, + 284, + 444, + 352 + ], + "page_idx": 9 + }, + { + "type": "image", + "img_path": "images/53bd43002cbe255690c7593b37cea9cd70e356bf85f2abdd804c4f0d217dec58.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 446, + 284, + 557, + 352 + ], + "page_idx": 9 + }, + { + "type": "image", + "img_path": "images/fbef94c1c6513a492a1ff2f3fe031eac8108a705efb18462d24c51f6f624d273.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 558, + 284, + 669, + 352 + ], + "page_idx": 9 + }, + { + "type": "image", + "img_path": "images/0e9f06e8b26b7a5c5b584938217a0ef49043b4e14a4030a5ad2cbb04975fa454.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 671, + 284, + 784, + 352 + ], + "page_idx": 9 + }, + { + "type": "image", + "img_path": "images/2791c9e2fa5b3582b72051a0ae31264bfacf7665f3c649964d4345b5dd250e62.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 220, + 353, + 331, + 422 + ], + "page_idx": 9 + }, + { + "type": "image", + "img_path": "images/3bc4518628076c2a0a12009731b8c4b3a7b0db19829f3a17d80631ce28fbd2da.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 334, + 353, + 444, + 422 + ], + "page_idx": 9 + }, + { + "type": "image", + "img_path": "images/0af877d64b84cca3384c704d9a80c62652bf21cdaef957fe547b5c04dc443f50.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 446, + 353, + 557, + 422 + ], + "page_idx": 9 + }, + { + "type": "image", + "img_path": "images/a90cbd7254f39a29aae73ddff4a65cb7783ec8e98c99a7eea112fe921ecab657.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 558, + 353, + 669, + 422 + ], + "page_idx": 9 + }, + { + "type": "image", + "img_path": "images/b3e1e21d50c15562a0184d1168f3f67cf88d4b9ae3a13dc701dc2b1578f06a91.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 671, + 353, + 784, + 422 + ], + "page_idx": 9 + }, + { + "type": "image", + "img_path": "images/a6f52cfabf32d04165f985a997907205baf85afdc61136f6c7de71b115e05663.jpg", + "image_caption": [ + "Graphics" + ], + "image_footnote": [], + "bbox": [ + 220, + 422, + 331, + 492 + ], + "page_idx": 9 + }, + { + "type": "image", + "img_path": "images/8417ae31182c63fdf8e40f9141b4d7f334c0b66583ef52ae86168691ca5e5459.jpg", + "image_caption": [ + "NPBG++" + ], + "image_footnote": [], + "bbox": [ + 334, + 422, + 444, + 492 + ], + "page_idx": 9 + }, + { + "type": "image", + "img_path": "images/72bbd56efc3a9d7feff2f75f0d3be39567f62bbef291cea7a9e513e7fa4577af.jpg", + "image_caption": [ + "Trivol" + ], + "image_footnote": [], + "bbox": [ + 446, + 422, + 557, + 492 + ], + "page_idx": 9 + }, + { + "type": "image", + "img_path": "images/c762a483576bf3e9b7f297c6fc4ed9c28f38fd78190ccc00942141bb8acd89ab.jpg", + "image_caption": [ + "Ours", + "Fig. 2: Qualitative comparisons between ours and other point cloud renderers on the ScanNet dataset." + ], + "image_footnote": [], + "bbox": [ + 558, + 422, + 669, + 492 + ], + "page_idx": 9 + }, + { + "type": "image", + "img_path": "images/a3e98e2c4eafe5396aec5e65042dbef3b3e0a281c61c143ae3dcf14347e9cbca.jpg", + "image_caption": [ + "Ground Truth" + ], + "image_footnote": [], + "bbox": [ + 671, + 422, + 784, + 492 + ], + "page_idx": 9 + }, + { + "type": "text", + "text": "ScanNet. The problem of uneven distribution and missing points exists in the real scene-level point cloud. As shown in the first column in Fig. 2, traditional graphical rendering leads to a large number of holes in the rendering result. When the point cloud distribution is uneven, the point cloud of nearby objects is relatively sparse, and the 2D-CNN of $\\mathrm{NPBG}++$ pays more attention to the features of distant objects, ignoring the 3D information in the point cloud, resulting that nearby objects are covered in the rendering result. Our method uses 3DGS for feature projection, which better integrates 3D context information. There are a large number of artifacts in the results of $\\mathrm{NPBG}++$ and Trivol. Our multi-scale feature extraction network excels at capturing features across various levels of detail, effectively decoding these projected features to enhance the rendering of lighting effects and details significantly.", + "bbox": [ + 212, + 551, + 787, + 733 + ], + "page_idx": 9 + }, + { + "type": "text", + "text": "DTU. As we stated in the Dataset description part, we maintain points distributed sparsely on the surface of the objects. Under this circumstance, the graphics-based transformation demonstrates more serious bleeding surfaces. Trivol always partially reconstructs the whole object because it depends on the fully implicit neural representation. This setting is more likely to be suited to class-specific point rendering, as they claimed in the original paper. In contrast, DTU includes many different classes that obstacle Trivol to learning a unified feature", + "bbox": [ + 212, + 734, + 792, + 839 + ], + "page_idx": 9 + }, + { + "type": "page_number", + "text": "10", + "bbox": [ + 217, + 114, + 235, + 126 + ], + "page_idx": 9 + }, + { + "type": "header", + "text": "J. Wang et al.", + "bbox": [ + 271, + 114, + 367, + 128 + ], + "page_idx": 9 + }, + { + "type": "image", + "img_path": "images/cd7e8e3c419b67b0a9e1c4a5e97afbcd2c38b6a0ff753cf76840af3a10b0d0b0.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 220, + 143, + 331, + 213 + ], + "page_idx": 10 + }, + { + "type": "image", + "img_path": "images/a5b0070e917b1518abd623da711ed569c7bd906498eff8a4ff350cffd37cae6f.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 220, + 214, + 331, + 282 + ], + "page_idx": 10 + }, + { + "type": "image", + "img_path": "images/9078a231cf16884ae347dabeb3421730963e94f45fa066e60dea7730cd2371ad.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 220, + 284, + 331, + 353 + ], + "page_idx": 10 + }, + { + "type": "image", + "img_path": "images/e496e1d75bbe501f37b082708972a92a1510f864ed5ed652968cfdddc8cbcb0b.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 220, + 354, + 331, + 422 + ], + "page_idx": 10 + }, + { + "type": "image", + "img_path": "images/534bbafc15cdef1376e5331cd5e573188d73b8868bff007e6f0f9ee7e149c22b.jpg", + "image_caption": [ + "Graphics" + ], + "image_footnote": [], + "bbox": [ + 220, + 424, + 331, + 493 + ], + "page_idx": 10 + }, + { + "type": "image", + "img_path": "images/fec8f76ef5e437d4a2e1d10d5cbc6d77d38b48501cc59de379a8a1f1da410e91.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 331, + 143, + 444, + 213 + ], + "page_idx": 10 + }, + { + "type": "image", + "img_path": "images/900cd61fe518779f43ae3c683195b156981eca68694aa8b0171f159da704c2a3.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 333, + 214, + 444, + 282 + ], + "page_idx": 10 + }, + { + "type": "image", + "img_path": "images/03a0b695b51f0a6522fc3190c07d4849d065a3716affb235fb5d1e142b88ee21.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 333, + 284, + 444, + 353 + ], + "page_idx": 10 + }, + { + "type": "image", + "img_path": "images/c678ab74a7fc9983411addf14e6825203587b972a1a7748c323af7cf23dc110f.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 333, + 354, + 444, + 422 + ], + "page_idx": 10 + }, + { + "type": "image", + "img_path": "images/7c58b6df8619f5572dc6e5ab98a387cffcebe6e0a702d2e9728cc22d866067ec.jpg", + "image_caption": [ + "NPBG++" + ], + "image_footnote": [], + "bbox": [ + 333, + 424, + 444, + 493 + ], + "page_idx": 10 + }, + { + "type": "image", + "img_path": "images/f07622ea5d480b41fbf1394413cbf4a79235acc77591d52ca856c5e66c7fe582.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 446, + 143, + 557, + 213 + ], + "page_idx": 10 + }, + { + "type": "image", + "img_path": "images/1c2ef355942df92686b8e69ea9f8f863b06ca9d7b9f7cf0d609e08e1571d091a.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 446, + 214, + 555, + 282 + ], + "page_idx": 10 + }, + { + "type": "image", + "img_path": "images/0204fb6946aa852a289a4fe93103322ed0ceaac017f75bd049c97e8b7690f482.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 446, + 284, + 555, + 353 + ], + "page_idx": 10 + }, + { + "type": "image", + "img_path": "images/90d3ea82c194a288caf340dcb500f09a0417d43124a13065efb9fc97bf1d4565.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 446, + 354, + 555, + 422 + ], + "page_idx": 10 + }, + { + "type": "image", + "img_path": "images/68d9e99e02a84480988c3fc4737ae8d08f7c01509e465e769cebf933dfd7f13d.jpg", + "image_caption": [ + "Trivol" + ], + "image_footnote": [], + "bbox": [ + 446, + 424, + 555, + 493 + ], + "page_idx": 10 + }, + { + "type": "image", + "img_path": "images/43628b3babe020c2f0e47f123d5734b533bea8b2fb7380fc3f5f342f2c863e7b.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 558, + 143, + 669, + 213 + ], + "page_idx": 10 + }, + { + "type": "image", + "img_path": "images/f834ef6cf179062b2d31581f7e0f23341eafb07eb4cccc4bf31be7b406829d4c.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 558, + 214, + 669, + 282 + ], + "page_idx": 10 + }, + { + "type": "image", + "img_path": "images/4ea53ddf98210f2921906d3ff6092d8a05414746b69fa7994c64b4ea498fc7fc.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 558, + 284, + 669, + 353 + ], + "page_idx": 10 + }, + { + "type": "image", + "img_path": "images/0549f9a2cb704ddb9ec2f9eb638ea6d4e224fb61265ef9ce5f0aafa5a1d1aedb.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 558, + 354, + 669, + 422 + ], + "page_idx": 10 + }, + { + "type": "image", + "img_path": "images/3802eeae13785f41156e779fd4da8862b9ee104e376e928a073b3464d5031a43.jpg", + "image_caption": [ + "Ours", + "Fig. 3: Qualitative comparisons between ours and other point cloud rendering methods on the DTU dataset." + ], + "image_footnote": [], + "bbox": [ + 558, + 424, + 669, + 493 + ], + "page_idx": 10 + }, + { + "type": "image", + "img_path": "images/652a412911889a2293185ae31179e01afdc677df9706e9a859728094f91176f4.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 671, + 143, + 782, + 213 + ], + "page_idx": 10 + }, + { + "type": "image", + "img_path": "images/b199285bf82babde402ff41797e7e2547fce0b6b80511b47771f5f718feec624.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 671, + 214, + 782, + 282 + ], + "page_idx": 10 + }, + { + "type": "image", + "img_path": "images/ddf8a049c33d52c67b1ce40c1eb424591d6dadc6d5924504b96fe166fac18340.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 671, + 284, + 782, + 353 + ], + "page_idx": 10 + }, + { + "type": "image", + "img_path": "images/99746c9dd055c8136156f052c2e9f07d0598659e05a73bf6824729b4e92a39e3.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 671, + 354, + 782, + 422 + ], + "page_idx": 10 + }, + { + "type": "image", + "img_path": "images/85d51e737b6dba932d5b080a49b3c3f32e2d75f82729291314bf02dc19963512.jpg", + "image_caption": [ + "Ground Truth" + ], + "image_footnote": [], + "bbox": [ + 671, + 424, + 782, + 493 + ], + "page_idx": 10 + }, + { + "type": "text", + "text": "transform underlying them. NPBG++ and our method deliver better results than Trivol because they leverage the original color and structure of point clouds without class-specific assumptions. Our approach yields the clearest predictions thanks to our more reasonable feature projection and decoding.", + "bbox": [ + 212, + 556, + 784, + 616 + ], + "page_idx": 10 + }, + { + "type": "text", + "text": "THuman2.0 The Graphics-based method reconstructs the main structures of the human body while inevitably seeing densely packed holes. The other two baselines generate similar results that are much better than those of the traditional methods. The difference is that Trivol fails to faithfully recover some cloth colors, which is because it is a fully implicit method and thereby cannot restore colors that are never seen in its train set. $\\mathrm{NPBG}++$ and ours directly utilize points' colors thus avoiding this domain generalization issue. Moreover, our method restores more details such as human facial expressions and clothes' folds.", + "bbox": [ + 212, + 618, + 785, + 752 + ], + "page_idx": 10 + }, + { + "type": "text", + "text": "4.3 Ablation studies.", + "text_level": 1, + "bbox": [ + 215, + 767, + 401, + 780 + ], + "page_idx": 10 + }, + { + "type": "text", + "text": "Ablations of main components. In this subsection, we evaluate the effectiveness of all main constituents proposed in our whole framework. We conduct experiments with different control variables set and report their performance on", + "bbox": [ + 212, + 794, + 784, + 839 + ], + "page_idx": 10 + }, + { + "type": "header", + "text": "PFGS", + "bbox": [ + 687, + 114, + 730, + 126 + ], + "page_idx": 10 + }, + { + "type": "page_number", + "text": "11", + "bbox": [ + 767, + 116, + 782, + 126 + ], + "page_idx": 10 + }, + { + "type": "image", + "img_path": "images/64a62b1e3fe8dd5a348e8144b5004164caaaaff1c6f88ba316caf5e832938f02.jpg", + "image_caption": [ + "Fig. 4: Qualitative comparisons between ours and other point cloud renderers on the THuman2.0 dataset." + ], + "image_footnote": [], + "bbox": [ + 225, + 148, + 784, + 558 + ], + "page_idx": 11 + }, + { + "type": "text", + "text": "the ScanNet dataset in Table 2. In the table, $w / o$ ProL refers to the version that removes all the progressive loss items, i.e. removes the first summation in Eq. 11 and Eq. 12. $w / o$ MSL denotes the version that removes all the multiscale loss items that are the second summation in Eq. 11 and 12. The recur1 illustrates that we discard the recurrent structure from the feature decoder and only utilize a single multi-input and multi-output UNet. Besides, the $w / o$ MSf describes a variant of the decoder which adopts a normal single scale UNet. In other words, this variant receives a single feature map as input and outputs a single prediction with full resolutions. Therefore, we only maintain the progressive losses and structures but discard all multiscale components for this version. Furthermore, we evaluate when the number of loops equals to 3, i.e. $l = 3$ in these equations. This is marked as recur3 in the Table. The results show that continuously increasing the number of iterations does not further improve the model's performance; instead, it may lead to a slight decline. Last, the $w / o$ Lmfr represents the version of removing the progressively multiscale frequency reconstruction loss (Eq. 12).", + "bbox": [ + 212, + 598, + 787, + 843 + ], + "page_idx": 11 + }, + { + "type": "page_number", + "text": "12", + "bbox": [ + 217, + 114, + 235, + 126 + ], + "page_idx": 11 + }, + { + "type": "header", + "text": "J. Wang et al.", + "bbox": [ + 271, + 114, + 367, + 128 + ], + "page_idx": 11 + }, + { + "type": "table", + "img_path": "images/c24eed72414f241be373810abc1da7708b39935022d41de1c300299bf4496768.jpg", + "table_caption": [ + "Table 2: Quantitative Results of ablation studies under various variants on the ScanNet dataset. The meaning of each abbreviation can be found in the text." + ], + "table_footnote": [], + "table_body": "
w/o ProLw/o MSLw/o MSfw/o Lmfrw/o 2trrecur1recur3Full
PSNR↑19.1919.3718.2719.4215.9219.1919.7719.86
SSIM↑0.7250.7330.7040.7400.6960.7290.7600.758
LPIPS↓0.4980.5090.6500.4780.6890.5160.4780.452
", + "bbox": [ + 218, + 178, + 781, + 253 + ], + "page_idx": 12 + }, + { + "type": "image", + "img_path": "images/79ba08dca64e76c1dbb5a5bc7094bf0f5047d6a196c66ebd6923737026dd9952.jpg", + "image_caption": [ + "Fig. 5: Qualitative Comparisons of the renderings produced by the Gaussian splatting after the first stage and the feature splatting after the whole framework." + ], + "image_footnote": [], + "bbox": [ + 218, + 266, + 782, + 345 + ], + "page_idx": 12 + }, + { + "type": "text", + "text": "It is observed that the full model delivers the most satisfactory rendering results across all metrics while the rest variants experience a decline to some extent. This indicates that both progressive and multi-scale features have a positive impact and contribute to the final promising performance of the model. Therefore, the entire design of this method is consistently characterized by progressive and multi-scale aspects. Moreover, the alignment of predictions and labels in the frequency domain progressively and in a multiscale manner makes prediction clearer and sharper.", + "bbox": [ + 212, + 393, + 784, + 513 + ], + "page_idx": 12 + }, + { + "type": "text", + "text": "Analysis of Gaussian and Feature Splatting. The proposed approach relies on the two-stage training scheme because the 3D-consistent feature should be projected and integrated into certain 2D planes given camera viewpoints. The projection of features is dependent on the 3D Gaussian rasterization and further reliant on the Gaussian parameters associated with each point which are estimated by Eq. 6. The significance of training the Gaussian regression network in the first training stage lies in laying a solid foundation for feature splatting. Even though the 3D Gaussian Splatting could also produce plausible images, namely the $I_{gs}$ in Eq. 7, the results are still severely influenced by points that are either locally or globally sparse, thereby missing details and showing holes. It can be seen in Fig. 5 that these Gaussian splatting results are indeed better than the pure graphics-based method due to the representation capabilities of 3DGS, but they still need to be improved further.", + "bbox": [ + 212, + 537, + 784, + 733 + ], + "page_idx": 12 + }, + { + "type": "text", + "text": "However, the results of feature splatting contain rich semantic information and local geometry cues, we consider them a good representation of restoring high-quality images and filling up blank areas. From Fig. 5 we can observe the capability of our method to generate and complete areas with voids and regions with sparse point clouds. We additionally evaluate an experiment in which we collaboratively train all modules in the framework rather than conducting the two-stage training and the results on the ScanNet dataset and we add this ad", + "bbox": [ + 212, + 734, + 785, + 839 + ], + "page_idx": 12 + }, + { + "type": "header", + "text": "PFGS", + "bbox": [ + 687, + 114, + 730, + 126 + ], + "page_idx": 12 + }, + { + "type": "page_number", + "text": "13", + "bbox": [ + 767, + 114, + 784, + 126 + ], + "page_idx": 12 + }, + { + "type": "table", + "img_path": "images/c100a01d98473941f776582841f8fe7b7a81221c1cbf56089df26157ee77cb7e.jpg", + "table_caption": [ + "Table 3: Ablation studies of the effect of different numbers of points on performance." + ], + "table_footnote": [], + "table_body": "
Train Points50k50k50k120k120k120k
Test Points50k80k120k50k80k120k
Trivol24.0525.4326.4824.7125.9626.86
Ours31.2932.8635.9531.6034.1739.18
", + "bbox": [ + 217, + 171, + 779, + 247 + ], + "page_idx": 13 + }, + { + "type": "text", + "text": "ditional result into Table. 2. $w / o$ 2tr means that we abandon the first training stage and jointly train all modules together. In this setting, we observe that the model completely collapses and sees a very fast degradation in performance.", + "bbox": [ + 212, + 258, + 784, + 305 + ], + "page_idx": 13 + }, + { + "type": "text", + "text": "Analysis of The Number of Points. We additionally analyze and evaluate how different training and test numbers of points affect the performance. On the THuman2.0 dataset, we set different training and test pairs to show the performance of our method and Trivol respectively, which includes 50,000 and 120,000 points used for training and 50,000, 80,000, and 120,000 points used for testing. The results are listed in Table 3. It is noted that Trivol is essentially unaffected by the number of points where our method sees significant fluctuations with changes in the number of points. However, our method exhibits very high performance when point clouds are abundant while the Trivol method has a relatively low upper limit, which is due to its fully implicit representation and lack of utilization of point cloud prior. More discussion can be seen in the Appendix.", + "bbox": [ + 212, + 323, + 787, + 491 + ], + "page_idx": 13 + }, + { + "type": "text", + "text": "5 Discussion", + "text_level": 1, + "bbox": [ + 215, + 508, + 354, + 525 + ], + "page_idx": 13 + }, + { + "type": "text", + "text": "Conclusion. This work proposes a novel point cloud rendering method called PFGS, that can render high-fidelity and photorealistic images from colored point clouds with arbitrary camera viewpoints. To address the shortcomings of current learning-based point cloud renderers such as slow, blurry, and inconsistent, we introduce the 3D Gaussian Splatting to point cloud rendering by several well-designed modules. We supervise the regressor to estimate Gaussian parameters in a point-wise way, and then we utilize the estimated Gaussian parameters to rasterize point-wise neural features that are extracted by a multiscale feature decoder. Furthermore, the projected feature volume would be gradually decoded to the final prediction by our progressive and multiscale decoder. The whole architecture is supervised by our finely-designed progressive and multiscale reconstruction losses. We carry out experiments on three datasets with different levels of objects including indoors, objects, and humans and the results show our method outperforms all baselines. The ablation studies also illustrate the significance of all main components in the approach.", + "bbox": [ + 212, + 537, + 787, + 763 + ], + "page_idx": 13 + }, + { + "type": "text", + "text": "Limitation. Even though the proposed method can produce photorealistic images from sparse and uneven point clouds, it still yields unreal details when the given point clouds are not precise enough. Future works can incorporate a large generative refiner trained on a large dataset to slightly refine the original point cloud to make points closer to the actual surface of objects.", + "bbox": [ + 212, + 765, + 787, + 839 + ], + "page_idx": 13 + }, + { + "type": "page_number", + "text": "14", + "bbox": [ + 217, + 114, + 235, + 126 + ], + "page_idx": 13 + }, + { + "type": "header", + "text": "J. Wang et al.", + "bbox": [ + 271, + 114, + 367, + 128 + ], + "page_idx": 13 + }, + { + "type": "text", + "text": "References", + "text_level": 1, + "bbox": [ + 217, + 143, + 321, + 159 + ], + "page_idx": 14 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "1. Cao, A., Johnson, J.: Hexplane: A fast representation for dynamic scenes. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 130-141 (2023)", + "2. Chen, A., Xu, Z., Zhao, F., Zhang, X., Xiang, F., Yu, J., Su, H.: Mvsnerf: Fast generalizable radiance field reconstruction from multi-view stereo. In: Proceedings of the IEEE/CVF International Conference on Computer Vision. pp. 14124-14133 (2021)", + "3. Choy, C., Gwak, J., Savarese, S.: 4d spatio-temporal convnets: Minkowski convolutional neural networks. In: Proceedings of the IEEE/CVF conference on computer vision and pattern recognition. pp. 3075-3084 (2019)", + "4. Dai, A., Chang, A.X., Savva, M., Halber, M., Funkhouser, T., Nießner, M.: Scannet: Richly-annotated 3d reconstructions of indoor scenes. In: Proceedings of the IEEE conference on computer vision and pattern recognition. pp. 5828-5839 (2017)", + "5. Dai, P., Zhang, Y., Li, Z., Liu, S., Zeng, B.: Neural point cloud rendering via multi-plane projection. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 7830-7839 (2020)", + "6. Fridovich-Keil, S., Yu, A., Tancik, M., Chen, Q., Recht, B., Kanazawa, A.: Plenoxels: Radiance fields without neural networks. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 5501-5510 (2022)", + "7. Ge, W., Hu, T., Zhao, H., Liu, S., Chen, Y.C.: Ref-neus: Ambiguity-reduced neural implicit surface learning for multi-view reconstruction with reflection. arXiv preprint arXiv:2303.10840 (2023)", + "8. Graham, B., Van der Maaten, L.: Submanifold sparse convolutional networks. arXiv preprint arXiv:1706.01307 (2017)", + "9. Hu, T., Xu, X., Chu, R., Jia, J.: Trivol: Point cloud rendering via triple volumes. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 20732-20741 (2023)", + "10. Hu, T., Xu, X., Liu, S., Jia, J.: Point2pix: Photo-realistic point cloud rendering via neural radiance fields. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 8349-8358 (2023)", + "1. Huang, C., Mees, O., Zeng, A., Burgard, W.: Visual language maps for robot navigation. In: 2023 IEEE International Conference on Robotics and Automation (ICRA). pp. 10608-10615. IEEE (2023)", + "2. Jena, S., Multon, F., Boukhayma, A.: Neural mesh-based graphics. In: European Conference on Computer Vision. pp. 739-757. Springer (2022)", + "3. Jensen, R., Dahl, A., Vogiatzis, G., Tola, E., Aanaes, H.: Large scale multi-view stereopsis evaluation. In: Proceedings of the IEEE conference on computer vision and pattern recognition. pp. 406-413 (2014)", + "4. Jiang, Y., Yu, C., Xie, T., Li, X., Feng, Y., Wang, H., Li, M., Lau, H., Gao, F., Yang, Y., et al.: Vr-gs: A physical dynamics-aware interactive gaussian splattering system in virtual reality. arXiv preprint arXiv:2401.16663 (2024)", + "5. Kerbl, B., Kopanas, G., Leimkuhler, T., Drettakis, G.: 3d gaussian splatting for real-time radiance field rendering. ACM Transactions on Graphics 42(4) (2023)", + "6. Lassner, C., Zollhofer, M.: Pulsar: Efficient sphere-based neural rendering. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 1440-1449 (2021)", + "7. Liu, Z., Tang, H., Lin, Y., Han, S.: Point-voxel cnn for efficient 3d deep learning. Advances in Neural Information Processing Systems 32 (2019)" + ], + "bbox": [ + 225, + 176, + 784, + 839 + ], + "page_idx": 14 + }, + { + "type": "header", + "text": "PFGS", + "bbox": [ + 687, + 114, + 730, + 126 + ], + "page_idx": 14 + }, + { + "type": "page_number", + "text": "15", + "bbox": [ + 767, + 116, + 785, + 126 + ], + "page_idx": 14 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "18. Lombardi, S., Simon, T., Saragih, J., Schwartz, G., Lehrmann, A., Sheikh, Y.: Neural volumes: Learning dynamic renderable volumes from images. arXiv preprint arXiv:1906.07751 (2019)", + "19. Metzer, G., Richardson, E., Patashnik, O., Giryes, R., Cohen-Or, D.: Latent-nerf for shape-guided generation of 3d shapes and textures. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 12663-12673 (2023)", + "20. Mildenhall, B., Srinivasan, P.P., Tancik, M., Barron, J.T., Ramamoorthi, R., Ng, R.: Nerf: Representing scenes as neural radiance fields for view synthesis. Communications of the ACM 65(1), 99-106 (2021)", + "21. Müller, T., Evans, A., Schied, C., Keller, A.: Instant neural graphics primitives with a multiresolution hash encoding. ACM Transactions on Graphics (ToG) 41(4), 1-15 (2022)", + "22. Qi, C.R., Su, H., Nießner, M., Dai, A., Yan, M., Guibas, L.J.: Volumetric and multi-view cnns for object classification on 3d data. In: Proceedings of the IEEE conference on computer vision and pattern recognition. pp. 5648-5656 (2016)", + "23. Qi, C.R., Yi, L., Su, H., Guibas, L.J.: Pointnet++: Deep hierarchical feature learning on point sets in a metric space. Advances in neural information processing systems 30 (2017)", + "24. Rakhimov, R., Ardelean, A.T., Lempitsky, V., Burnaev, E.: Npgb++: Accelerating neural point-based graphics. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 15969-15979 (2022)", + "25. Ravi, N., Reizenstein, J., Novotny, D., Gordon, T., Lo, W.Y., Johnson, J., Gkioxari, G.: Accelerating 3d deep learning with pytorch3d. arXiv preprint arXiv:2007.08501 (2020)", + "26. Sitzmann, V., Thies, J., Heide, F., Nießner, M., Wetzstein, G., Zollhofer, M.: Deepvoxels: Learning persistent 3d feature embeddings. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 2437-2446 (2019)", + "27. Sun, W., Trulls, E., Tseng, Y.C., Sambandam, S., Sharma, G., Tagliasacchi, A., Yi, K.M.: Pointner++: A multi-scale, point-based neural radiance field. arXiv preprint arXiv:2312.02362 (2023)", + "28. Wang, C., Chai, M., He, M., Chen, D., Liao, J.: Clip-nerf: Text-and-image driven manipulation of neural radiance fields. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 3835-3844 (2022)", + "29. Wang, J., He, J., Zhang, Z., Xu, R.: Physical priors augmented event-based 3d reconstruction. arXiv preprint arXiv:2401.17121 (2024)", + "30. Wang, J., Zhang, Z., Xu, R.: Learning to generate and manipulate 3d radiance field by a hierarchical diffusion framework with clip latent. In: Computer Graphics Forum. vol. 42, p. e14930. Wiley Online Library (2023)", + "31. Wang, J., Zhang, Z., Xu, R.: Learning robust generalizable radiance field with visibility and feature augmented point representation. arXiv preprint arXiv:2401.14354 (2024)", + "32. Wang, J., Sun, B., Lu, Y.: Mvpnet: Multi-view point regression networks for 3d object reconstruction from a single image. In: Proceedings of the AAAI Conference on Artificial Intelligence. vol. 33, pp. 8949-8956 (2019)", + "33. Wang, N., Zhang, Y., Li, Z., Fu, Y., Liu, W., Jiang, Y.G.: Pixel2mesh: Generating 3d mesh models from single rgb images. In: Proceedings of the European conference on computer vision (ECCV). pp. 52-67 (2018)" + ], + "bbox": [ + 215, + 146, + 785, + 839 + ], + "page_idx": 15 + }, + { + "type": "page_number", + "text": "16", + "bbox": [ + 217, + 114, + 235, + 126 + ], + "page_idx": 15 + }, + { + "type": "header", + "text": "J. Wang et al.", + "bbox": [ + 271, + 114, + 366, + 128 + ], + "page_idx": 15 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "34. Wang, P., Liu, Y., Chen, Z., Liu, L., Liu, Z., Komura, T., Theobalt, C., Wang, W.: F2-nerf: Fast neural radiance field training with free camera trajectories. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 4150-4159 (2023)", + "35. Wang, S., Liu, M.: Point cloud compression with range image-based entropy model for autonomous driving. In: European Conference on Computer Vision. pp. 323-340. Springer (2022)", + "36. Wang, Z., Bovik, A.C., Sheikh, H.R., Simoncelli, E.P.: Image quality assessment: from error visibility to structural similarity. IEEE transactions on image processing 13(4), 600-612 (2004)", + "37. Wu, G., Yi, T., Fang, J., Xie, L., Zhang, X., Wei, W., Liu, W., Tian, Q., Wang, X.: 4d gaussian splatting for real-time dynamic scene rendering. arXiv preprint arXiv:2310.08528 (2023)", + "38. Xiang, F., Xu, Z., Hasan, M., Hold-Geoffroy, Y., Sunkavalli, K., Su, H.: Neutex: Neural texture mapping for volumetric neural rendering. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 7119-7128 (2021)", + "39. Xu, Q., Xu, Z., Philip, J., Bi, S., Shu, Z., Sunkavalli, K., Neumann, U.: Point-nerf: Point-based neural radiance fields. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 5438-5448 (2022)", + "40. Yan, Z., Li, C., Lee, G.H.: Nerf-ds: Neural radiance fields for dynamic specular objects. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 8285-8295 (2023)", + "41. Yang, B., Bao, C., Zeng, J., Bao, H., Zhang, Y., Cui, Z., Zhang, G.: Neumesh: Learning disentangled neural mesh-based implicit field for geometry and texture editing. In: European Conference on Computer Vision. pp. 597-614. Springer (2022)", + "42. Yu, T., Zheng, Z., Guo, K., Liu, P., Dai, Q., Liu, Y.: Function4d: Real-time human volumetric capture from very sparse consumer rgbd sensors. In: IEEE Conference on Computer Vision and Pattern Recognition (CVPR2021) (June 2021)", + "43. Zhang, C., Guo, R., Zeng, W., Xiong, Y., Dai, B., Hu, R., Ren, M., Urtasun, R.: Rethinking closed-loop training for autonomous driving. In: European Conference on Computer Vision. pp. 264-282. Springer (2022)", + "44. Zhang, R., Isola, P., Efros, A.A., Shechtman, E., Wang, O.: The unreasonable effectiveness of deep features as a perceptual metric. In: Proceedings of the IEEE conference on computer vision and pattern recognition. pp. 586-595 (2018)", + "45. Zhou, Q.Y., Park, J., Koltun, V.: Open3d: A modern library for 3d data processing. arXiv preprint arXiv:1801.09847 (2018)", + "46. Zhou, S., Chang, H., Jiang, S., Fan, Z., Zhu, Z., Xu, D., Chari, P., You, S., Wang, Z., Kadambi, A.: Feature 3dgs: Supercharging 3d gaussian splatting to enable distilled feature fields. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 21676-21685 (2024)" + ], + "bbox": [ + 212, + 146, + 787, + 729 + ], + "page_idx": 16 + }, + { + "type": "header", + "text": "PFGS", + "bbox": [ + 687, + 114, + 730, + 126 + ], + "page_idx": 16 + }, + { + "type": "page_number", + "text": "17", + "bbox": [ + 767, + 114, + 785, + 126 + ], + "page_idx": 16 + } +] \ No newline at end of file diff --git a/2024/PFGS_ High Fidelity Point Cloud Rendering via Feature Splatting/685f2bf6-434e-4901-92a2-92656824eb88_model.json b/2024/PFGS_ High Fidelity Point Cloud Rendering via Feature Splatting/685f2bf6-434e-4901-92a2-92656824eb88_model.json new file mode 100644 index 0000000000000000000000000000000000000000..25031e40de185b2275c08d1c9e6a97353675ddfb --- /dev/null +++ b/2024/PFGS_ High Fidelity Point Cloud Rendering via Feature Splatting/685f2bf6-434e-4901-92a2-92656824eb88_model.json @@ -0,0 +1,2775 @@ +[ + [ + { + "type": "title", + "bbox": [ + 0.227, + 0.141, + 0.78, + 0.187 + ], + "angle": 0, + "content": "PFGS: High Fidelity Point Cloud Rendering via Feature Splatting" + }, + { + "type": "text", + "bbox": [ + 0.252, + 0.212, + 0.751, + 0.229 + ], + "angle": 0, + "content": "Jiaxu Wang\\(^{1\\dagger}\\), Ziyi Zhang\\(^{1\\dagger}\\), Junhao He\\(^{1}\\), and Renjing Xu\\(^{1\\boxplus}\\)" + }, + { + "type": "text", + "bbox": [ + 0.312, + 0.239, + 0.692, + 0.283 + ], + "angle": 0, + "content": "\\(^{1}\\) Hong Kong University of Science and Technology (GZ) jwang457@connect.hkust-gz.edu.cn {ziyizhang, junhaohe, renjingxu}@hkust-gz.edu.cn" + }, + { + "type": "text", + "bbox": [ + 0.261, + 0.31, + 0.744, + 0.518 + ], + "angle": 0, + "content": "Abstract. Rendering high-fidelity images from sparse point clouds is still challenging. Existing learning-based approaches suffer from either hole artifacts, missing details, or expensive computations. In this paper, we propose a novel framework to render high-quality images from sparse points. This method first attempts to bridge the 3D Gaussian Splatting and point cloud rendering, which includes several cascaded modules. We first use a regressor to estimate Gaussian properties in a point-wise manner, the estimated properties are used to rasterize neural feature descriptors into 2D planes which are extracted from a multiscale extractor. The projected feature volume is gradually decoded toward the final prediction via a multiscale and progressive decoder. The whole pipeline experiences a two-stage training and is driven by our well-designed progressive and multiscale reconstruction loss. Experiments on different benchmarks show the superiority of our method in terms of rendering qualities and the necessities of our main components.\\(^{1}\\)" + }, + { + "type": "text", + "bbox": [ + 0.262, + 0.53, + 0.661, + 0.545 + ], + "angle": 0, + "content": "Keywords: Point cloud rendering \\(\\cdot\\) 3D Gaussian Splatting" + }, + { + "type": "title", + "bbox": [ + 0.216, + 0.566, + 0.377, + 0.582 + ], + "angle": 0, + "content": "1 Introduction" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.594, + 0.788, + 0.744 + ], + "angle": 0, + "content": "Synthesizing photorealistic images from given colored point clouds and arbitrary camera views can be extensively applied to various fields including virtual/augmented reality [14], robotic navigation [11], automatic driving [35, 43], etc, which is academically named point cloud rendering. Traditional methods to render views from point clouds are based on graphics transformations. They directly project 3D point clouds into 2D planes by using the camera parameters and z-buffer rasterization. However, the sparse distribution of points leads to bleeding surfaces and hole artifacts. Even though the graphics renderer can treat points as graphical primitives and diminish these issues by color blending, rough and simple geometries of primitives entail blurred images and missing details." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.746, + 0.788, + 0.806 + ], + "angle": 0, + "content": "In recent years, learning-based methods [9, 10, 24] have become prevailing to address the drawbacks of conventional graphics-based approaches. These methods project points and corresponding feature descriptors to 2D planes, obtaining sparse images. Then they use UNet-like networks to complete and restore" + }, + { + "type": "page_footnote", + "bbox": [ + 0.218, + 0.812, + 0.565, + 0.826 + ], + "angle": 0, + "content": "\\(\\star \\dagger\\) Co-first authors; \\(\\boxtimes\\) Corresponding authors" + }, + { + "type": "page_footnote", + "bbox": [ + 0.222, + 0.827, + 0.576, + 0.84 + ], + "angle": 0, + "content": "1 Project page: https://github.com/Mercerai/PPGS" + }, + { + "type": "list", + "bbox": [ + 0.218, + 0.812, + 0.576, + 0.84 + ], + "angle": 0, + "content": null + } + ], + [ + { + "type": "page_number", + "bbox": [ + 0.218, + 0.116, + 0.23, + 0.127 + ], + "angle": 0, + "content": "2" + }, + { + "type": "header", + "bbox": [ + 0.272, + 0.115, + 0.368, + 0.129 + ], + "angle": 0, + "content": "J. Wang et al." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.147, + 0.788, + 0.314 + ], + "angle": 0, + "content": "the low-quality images [12, 24]. Nevertheless, these methods rely on 2D image restoration, which cannot guarantee 3D consistency and can result in the distinctive appearance of the same object from nearby different viewpoints. Moreover, their feature descriptors are associated with points that do not possess actual shapes and volumes in real space, thereby showing poor descriptions of local area structures. To obtain 3D-aware and consistent features, some approaches [8, 16] transform points to 3D volumes by trilinear interpolation and use 3D convolution to extract the features. However, they lead to a significant computational burden and sacrifice the efficient representation characteristics of point clouds. For example, voxels uniformly represent scenes, whereas points are densely distributed in complex areas and sparsely distributed or even absent in empty spaces." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.316, + 0.788, + 0.483 + ], + "angle": 0, + "content": "To obtain efficient and consistent 3D representation, the Neural Radiance Field (NeRF) [20] has gradually been incorporated into point cloud renderers [9, 10]. These methods either consider points as anchors of NeRF or transform points into NeRF-like representations, then use the physical-based volume rendering to obtain 3D-aware feature maps which are decoded into corresponding images. With the assistance of NeRF, they achieve high-quality consistencies across nearby views and plausible images. However, discrete NeRF requires heavy memory consumption resulting from storing high-dimensional feature vectors. In addition, NeRF typically requires lots of sampling points and multiple accesses to the neural network, which significantly increases the rendering time and makes it challenging to meet real-time requirements." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.486, + 0.788, + 0.622 + ], + "angle": 0, + "content": "More recently, 3D Gaussian Splatting (3DGS) [15] introduces a new representation that formulates points as 3D Gaussians with learnable parameters including 3D position, color, opacity, and anisotropic covariance. 3DGS achieves more photorealistic and faster rendering than NeRF by applying Gaussian rasterization and \\(\\alpha\\)-blending. Conventional 3D Gaussian learns the primitive parameters associated with an initialized point cloud from multi-view images by backpropagation, it relies on per-subject parameter optimizations for several tens minutes. It is therefore impractical to be applied to the case in which only point cloud is available due to the need for multiview images and backpropagations." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.626, + 0.788, + 0.716 + ], + "angle": 0, + "content": "In this work, we fill the gap between point cloud render and the 3DGS technique by introducing a novel multiscale, feature-augmented, 3DGS-based point renderer. Compared with previous learning-based point renderers, our method yields higher-quality and 3D consistent results. Compared with conventional 3DGS, our method does not necessarily require multiview images or fine-tuning procedures and can be directly employed to point modality." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.72, + 0.788, + 0.84 + ], + "angle": 0, + "content": "We leverage an efficient multiscale feature extractor in the proposed framework to extract features from point cloud input. This extractor considers multiscale local geometries to extract discriminative feature descriptors. Then a regressor is used for estimating 3D Gaussian parameters in a point-wise manner. Next, we not only render images by these estimated Gaussian attributes in a multiscale way but also render the multiscale feature maps by the same Gaussian properties. Finally, the rendered multiscale images and feature maps are fused by a multiscale and recurrent UNet-like network to gradually transform" + } + ], + [ + { + "type": "header", + "bbox": [ + 0.689, + 0.115, + 0.732, + 0.127 + ], + "angle": 0, + "content": "PFGS" + }, + { + "type": "page_number", + "bbox": [ + 0.776, + 0.117, + 0.786, + 0.127 + ], + "angle": 0, + "content": "3" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.147, + 0.784, + 0.176 + ], + "angle": 0, + "content": "into the final results. To the best knowledge of us, this is the first point cloud renderer combined with 3DGS." + }, + { + "type": "text", + "bbox": [ + 0.239, + 0.178, + 0.674, + 0.192 + ], + "angle": 0, + "content": "Our main contributions can be summarized in the following:" + }, + { + "type": "text", + "bbox": [ + 0.223, + 0.202, + 0.784, + 0.261 + ], + "angle": 0, + "content": "1. We proposed a novel lightweight approach called PFGS (Point Feature Gaussian Splitting) that bridges point cloud rendering and 3DGS. It produces higher-fidelity images from sparse point clouds than currently prevailing approaches." + }, + { + "type": "text", + "bbox": [ + 0.223, + 0.262, + 0.785, + 0.336 + ], + "angle": 0, + "content": "2. We proposed the point-based multiscale and feature-augmented 3D Gaussian framework, which consists of three main components including a multiscale feature extractor, a Gaussian regressor, and a multiscale and recurrent decoder, accompanied by the well-designed multiscale and progressive reconstruction losses." + }, + { + "type": "text", + "bbox": [ + 0.223, + 0.338, + 0.784, + 0.383 + ], + "angle": 0, + "content": "3. Extensive experiments and ablation studies on three datasets show the effectiveness and advantages of the proposed method over current works in terms of reconstruction quality." + }, + { + "type": "list", + "bbox": [ + 0.223, + 0.202, + 0.785, + 0.383 + ], + "angle": 0, + "content": null + }, + { + "type": "title", + "bbox": [ + 0.216, + 0.406, + 0.387, + 0.422 + ], + "angle": 0, + "content": "2 Related Work" + }, + { + "type": "title", + "bbox": [ + 0.216, + 0.438, + 0.428, + 0.452 + ], + "angle": 0, + "content": "2.1 Point Cloud Render" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.463, + 0.785, + 0.538 + ], + "angle": 0, + "content": "Conventional point cloud rendering is implemented via graphics-based transformation [25,45] which simulates camera imaging processes to project points onto the 2D screens. Even though the rendering pipeline is fast and general for all scenes, it is heavily dependent on the density of point distributions. These methods produce empty holes and vacant areas when the distributions are sparse." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.539, + 0.786, + 0.84 + ], + "angle": 0, + "content": "Learning-based point renderers [3, 9, 10, 12, 24] to some extent handle the issues caused by traditional methods, which fill the holes by leveraging multiscale or multiview feature fusions. ME [3] implements a sparse Convolutional Neural Network to extract features from existing points and then calculates the features of arbitrary 3D points through ball querying. NPBG [12] enhances each point with a feature descriptor to encapsulate local geometry and appearance. Moreover, \\(\\mathrm{NPBG}++\\) [24] introduces feature aggregation from multiview based on the NPBG to complement missing parts. NPCR [5] employs a 3DCNN to generate 3D volumes from point clouds and creates multiple depth layers to synthesize images. More recently, Point2pix [10] and Trivol [9] have incorporated NeRF into the point cloud rendering pipeline. They transform raw point clouds into NeRF-related representations and utilize volume rendering to access the representations to produce images. However, the above methods still cannot fill large vacant holes and only produce blurry rendering results, especially in local areas with fewer points. Even if incorporating NeRF representation is effective in generating 3D consistent feature encoding, they do not fully utilize the geometric information contained within the point clouds. In this work, we combine point rendering with a novel 3D representation, i.e. 3DGS, and implement multiscale and progressive feature decoding to obtain fine-grained and 3D-consistent predictions." + } + ], + [ + { + "type": "page_number", + "bbox": [ + 0.218, + 0.116, + 0.23, + 0.127 + ], + "angle": 0, + "content": "4" + }, + { + "type": "header", + "bbox": [ + 0.272, + 0.115, + 0.368, + 0.129 + ], + "angle": 0, + "content": "J. Wang et al." + }, + { + "type": "title", + "bbox": [ + 0.216, + 0.147, + 0.502, + 0.162 + ], + "angle": 0, + "content": "2.2 Neural-based View Synthesis" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.168, + 0.788, + 0.41 + ], + "angle": 0, + "content": "Unlike traditional explicit representations [22, 23, 32, 33], implicit representations that encode scenes with the assistance of neural networks have been developed in recent years [18, 26, 38]. Among them, the Neural Radiance Field (NeRF) [20] attracts the most attention in both industries and research. NeRF is a continuous method that implicitly represents a 3D scene employing an MLP and synthesizing images of novel view by volume rendering. Recent studies have employed NeRF for various purposes, including dynamic reconstruction [1, 37, 40], physical reasoning [29], generative models [19, 28, 30], and fast rendering [6, 21, 34]. To achieve various goals, some studies attempted to combine the implicit NeRF with explicit primitives. For example, NeuMesh [41] regards mesh as a neural scaffold to store NeRF and achieve neural editing. PointNeRF [39] is the first to incorporate point clouds into NeRF reconstruction to obtain better rendering qualities and faster speed. Ref-NeuS [7] implements generalizable neural surface reconstruction. GPF [31] explicitly models the visibility by using the geometries in the point cloud. PointNeRF++ [27] solves point cloud sparsity by aggregating multiscale point clouds on sparse voxel grids at different resolutions." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.41, + 0.788, + 0.502 + ], + "angle": 0, + "content": "It can be seen that many approaches we list above incorporate points into NeRF. However, NeRF-like volume rendering does not fully explore the rendering capability of the point cloud. In contrast to NeRF, 3DGS [15] proposed the 3D Gaussian Splatting to realize a remarkable differentiable rendering speed and yield high-quality 2D images. In this work, we first bridge point cloud rendering and 3DGS to generate higher-fidelity images." + }, + { + "type": "title", + "bbox": [ + 0.216, + 0.52, + 0.381, + 0.538 + ], + "angle": 0, + "content": "3 Methodology" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.548, + 0.788, + 0.716 + ], + "angle": 0, + "content": "Given a colored point cloud \\(\\mathbf{P} = \\{\\mathbf{p}_k, \\mathbf{c}_k\\}\\), we aim to synthesize photorealistic images from any viewpoints defined by the camera intrinsic \\(K\\) and poses \\(P\\) are provided. The \\(\\mathbf{p}_k \\in R^3\\) and \\(\\mathbf{c}_k \\in R^3\\) are the point coordinates and colors respectively. The rendering process can be illustrated as \\(I_v = Re(\\mathbf{P}|K_v, P_v)\\) where \\(Re\\) is the rendering function that can be implemented by either graphics-based or learning-based methods as we introduced above. In this work, it is implemented by our proposed multiscale feature-based 3DGS rendering pipeline that contains a multiscale feature extractor, a Gaussian regressor, a feature-based Gaussian rendering module, and a multiscale recurrent decoder. All modules are implemented with small networks. In the following of this section, we sequentially introduce the aforementioned components." + }, + { + "type": "title", + "bbox": [ + 0.216, + 0.733, + 0.373, + 0.748 + ], + "angle": 0, + "content": "3.1 Preliminaries" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.755, + 0.787, + 0.816 + ], + "angle": 0, + "content": "3DGS is an explicit representation method for 3D scenes, which parameterizes the scene as a series of 3D Gaussian primitives. A 3D Gaussian is defined by a full 3D covariance matrix \\(\\Sigma\\), its center point \\(x\\), and the spherical harmonic \\((SH)\\). The mean value of the Gaussian is defined as:" + }, + { + "type": "equation", + "bbox": [ + 0.425, + 0.822, + 0.786, + 0.842 + ], + "angle": 0, + "content": "\\[\nG (x) = e ^ {- \\frac {1}{2} (x) ^ {T} \\Sigma^ {- 1} (x)} \\tag {1}\n\\]" + } + ], + [ + { + "type": "header", + "bbox": [ + 0.689, + 0.115, + 0.732, + 0.127 + ], + "angle": 0, + "content": "PFGS" + }, + { + "type": "page_number", + "bbox": [ + 0.776, + 0.117, + 0.786, + 0.127 + ], + "angle": 0, + "content": "5" + }, + { + "type": "image", + "bbox": [ + 0.218, + 0.144, + 0.789, + 0.302 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.23, + 0.303, + 0.785, + 0.464 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.214, + 0.473, + 0.788, + 0.53 + ], + "angle": 0, + "content": "Fig. 1: The main pipeline of the proposed approach. The above panel includes the multiscale feature extractor and Gaussian regressor. Both Gaussian and features are splatted to the 2D plane and concatenated and fed to the progressive and multiscale feature decoding module described in the low panel." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.54, + 0.788, + 0.571 + ], + "angle": 0, + "content": "To enable optimization via backpropagation, the covariance matrix could be decomposed into a rotation matrix \\((R)\\) and a scaling matrix \\((S)\\):" + }, + { + "type": "equation", + "bbox": [ + 0.447, + 0.579, + 0.786, + 0.598 + ], + "angle": 0, + "content": "\\[\n\\Sigma = R S S ^ {T} R ^ {T} \\tag {2}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.608, + 0.788, + 0.655 + ], + "angle": 0, + "content": "Given the camera trajectory, the projection of the 3D Gaussians to the 2D image plane can be characterized by a view transform matrix \\((W)\\) and the Jacobian of the affine approximation of the projective transformation \\((J)\\), as in:" + }, + { + "type": "equation", + "bbox": [ + 0.433, + 0.663, + 0.786, + 0.682 + ], + "angle": 0, + "content": "\\[\n\\Sigma^ {\\prime} = J W \\Sigma W ^ {T} J ^ {T} \\tag {3}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.692, + 0.788, + 0.737 + ], + "angle": 0, + "content": "where the \\(\\Sigma^{\\prime}\\) is the covariance matrix in 2D image planes. Thus, the \\(\\alpha\\)-blend of \\(\\mathcal{N}\\) ordered points overlapping a pixel is utilized to compute the final color \\(C\\) of the pixel:" + }, + { + "type": "equation", + "bbox": [ + 0.41, + 0.738, + 0.786, + 0.779 + ], + "angle": 0, + "content": "\\[\nC = \\sum_ {i \\in \\mathcal {N}} c _ {i} \\alpha_ {i} \\prod_ {j = 1} ^ {i - 1} (1 - \\alpha_ {j}) \\tag {4}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.78, + 0.787, + 0.81 + ], + "angle": 0, + "content": "where \\( c_{i} \\) and \\( \\alpha_{i} \\) denote the color and density of the pixel with corresponding Gaussian parameters. Following [46], we projected the features similarly." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.811, + 0.787, + 0.84 + ], + "angle": 0, + "content": "In summary, the above parameters can be characterized by the following attributes: \\( x \\) denotes the position of a primitive where \\( x \\in R^3 \\). The rotation" + } + ], + [ + { + "type": "page_number", + "bbox": [ + 0.218, + 0.116, + 0.23, + 0.127 + ], + "angle": 0, + "content": "6" + }, + { + "type": "header", + "bbox": [ + 0.272, + 0.115, + 0.368, + 0.129 + ], + "angle": 0, + "content": "J. Wang et al." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.147, + 0.788, + 0.224 + ], + "angle": 0, + "content": "matrix is characterized by a quaternion \\( q \\in R^4 \\). The scale factor is defined by the anisotropy stretching \\( s \\in R^3 \\). The opacity factor \\( \\alpha \\in R^1 \\in [0,1] \\). The 2D opacity \\( \\alpha \\in [0,1] \\) is computed by \\( \\alpha_i(x) = o_i \\exp(-\\frac{1}{2} (x - \\mu_i)^T \\Sigma_i^T (x - \\mu_i)) \\) where the \\( \\mu \\) and variance are the 2D-projected mean and variance of 3D Gaussians. And the color is represented by \\( SH \\)." + }, + { + "type": "title", + "bbox": [ + 0.215, + 0.245, + 0.786, + 0.26 + ], + "angle": 0, + "content": "3.2 Gaussian Feature Prediction with Multiscale Feature Extraction" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.27, + 0.789, + 0.391 + ], + "angle": 0, + "content": "We first extract features for each point as their neural descriptors. We use a multi-input, single-output UNet-like architecture as the extraction network to encode points with different scales and capture features across different spatial distances. The architecture is mainly made of several PV Convolution modules [17], which is a highly efficient and commonly used point-based feature integration operation. We downsample the original point cloud with decreasing rates and concatenate the downsampled point clouds with different levels of feature maps as extra inputs." + }, + { + "type": "equation", + "bbox": [ + 0.329, + 0.403, + 0.786, + 0.42 + ], + "angle": 0, + "content": "\\[\nf _ {1: k} = \\mathcal {E} _ {\\theta} \\left(\\left(p _ {1: k}, c _ {1: k}\\right), \\left(p _ {1: k}, c _ {1: k}\\right) _ {\\downarrow r _ {1}}, \\left(p _ {1: k}, c _ {1: k}\\right) _ {\\downarrow r _ {2}}\\right) \\tag {5}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.431, + 0.788, + 0.564 + ], + "angle": 0, + "content": "The \\(\\mathcal{E}_{\\theta}\\) refers to the extractor network, \\(p, c\\) denote point coordinates and colors. The down arrow represents a uniform downsampling operation and \\(r\\) denotes the sampling rate where \\(r_1 > r_2\\). The output \\(f_{1:k}\\) are point neural descriptors. This operation enables the extractor to recognize features of different scales and receptive fields. The detailed architecture and parameter setting are shown in our Appendix. The single output feature vector is considered as the descriptors for all points. Afterward, the neural descriptors, the coordinates, as well as colors are fed to another network to predict the Gaussian properties in a point-wise manner." + }, + { + "type": "equation", + "bbox": [ + 0.326, + 0.565, + 0.786, + 0.589 + ], + "angle": 0, + "content": "\\[\nF _ {k} = \\mathcal {R} _ {\\theta} \\left(f _ {k}, p _ {k}, c _ {k}\\right) \\tag {6}\n\\]" + }, + { + "type": "equation", + "bbox": [ + 0.326, + 0.583, + 0.679, + 0.598 + ], + "angle": 0, + "content": "\\[\nR, S, \\alpha = \\operatorname {n o r m} \\left(\\mathcal {H} _ {r} \\left(F _ {k}\\right)\\right), \\exp \\left(\\mathcal {H} _ {s} \\left(F _ {k}\\right)\\right), \\sigma \\left(\\mathcal {H} _ {o} \\left(F _ {k}\\right)\\right)\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.604, + 0.788, + 0.695 + ], + "angle": 0, + "content": "where \\(\\mathcal{R}_{\\theta}\\) represents the main network of the regressor, it maps multiple inputs to a single unified feature vector \\(F_{k}\\). The \\(F_{k}\\) is fed to three individual prediction heads \\((\\mathcal{H}_r,\\mathcal{H}_s,\\mathcal{H}_o)\\), each of them is composed of two 1D Convolution layers with 1 kernel size, and activated by different functions to obtain corresponding Gaussian properties, namely rotation quaternion \\((R)\\), scale factor \\((S)\\) and opacity \\((\\alpha)\\). The color property \\(c\\) remains that of the original points." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.695, + 0.788, + 0.756 + ], + "angle": 0, + "content": "After the Gaussian attributes are obtained, one can render corresponding images from given camera viewpoints by Gaussian rasterization. We not only render the images by blending point colors on 2D planes but also render feature planes by replacing colors with the neural descriptors \\((f_{k})\\)." + }, + { + "type": "equation", + "bbox": [ + 0.277, + 0.768, + 0.786, + 0.784 + ], + "angle": 0, + "content": "\\[\nI _ {g s}, F _ {g s} = R e (R, S, \\alpha , p _ {1: k}, c _ {1: k} | K, P), R e (R, S, \\alpha , p _ {1: k}, f _ {1: k} | K, P) \\tag {7}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.795, + 0.788, + 0.84 + ], + "angle": 0, + "content": "where \\( I_{gs} \\), \\( F_{gs} \\) are corresponding image and feature map under the camera intrinsic \\( K \\) and pose \\( P \\). The \\( I_{gs} \\) can be considered a good result of point rendering when the number of points is sufficient." + } + ], + [ + { + "type": "header", + "bbox": [ + 0.689, + 0.115, + 0.732, + 0.127 + ], + "angle": 0, + "content": "PFGS" + }, + { + "type": "page_number", + "bbox": [ + 0.776, + 0.117, + 0.786, + 0.127 + ], + "angle": 0, + "content": "7" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.147, + 0.788, + 0.288 + ], + "angle": 0, + "content": "However, when the point sparsely distributes over the space, the representation capability of corresponding Gaussians is constrained, especially in complex local areas, resulting in poor-quality rendering results. On the contrary, the neural descriptors contain rich information to describe local structures with various scales. We therefore utilize the \\( F_{gs} \\) to be gradually transformed to the final results in a multiscale and recurrent manner. The \\( F_{gs} \\) is 3D-consistently rasterized with the same Gaussian parameters in Eq. 7, thus it precisely corresponds to the \\( I_{gs} \\) in pixel-level. In other words, \\( F_{gs} \\) can be regarded as a complement to \\( I_{gs} \\). Hence we concatenate \\( I_{gs} \\) and \\( F_{gs} \\) to be the input of the next module." + }, + { + "type": "title", + "bbox": [ + 0.215, + 0.311, + 0.636, + 0.328 + ], + "angle": 0, + "content": "3.3 Multiscale and Progressive Feature Decoding" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.342, + 0.788, + 0.465 + ], + "angle": 0, + "content": "After we render the feature volume \\( F_{gs} \\) associated with the provided camera viewpoint, we concatenate the \\( I_{gs} \\) and \\( F_{gs} \\) as input of the decoder. The decoder is a multi-input and multi-output UNet-like architecture, which aims to gradually refine and decode the input features and transform them into the target view images. We downsample the input two times with the rate of \\( r_a \\) and \\( r_b \\) and fuse them with the next two low-resolution layers of the UNet. There are interlaced connections between layers of different levels in the network, as stated in the main Fig. 1." + }, + { + "type": "equation", + "bbox": [ + 0.299, + 0.478, + 0.786, + 0.495 + ], + "angle": 0, + "content": "\\[\n\\mathcal {F}, \\mathcal {F} _ {a}, \\mathcal {F} _ {b} = \\mathcal {D} _ {\\theta} \\left(\\left(F _ {g s} \\oplus I _ {g s}\\right), \\left(F _ {g s} \\oplus I _ {g s}\\right) _ {\\downarrow r _ {a}}, \\left(F _ {g s} \\oplus I _ {g s}\\right) _ {\\downarrow r _ {b}}\\right) \\tag {8}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.508, + 0.788, + 0.612 + ], + "angle": 0, + "content": "In the above equation, \\(\\mathcal{D}_{\\theta}\\) indicates the multiscale decoder. \\(\\downarrow r_{a}\\) and \\(\\downarrow r_{b}\\) are two downsampling operations. \\(\\mathcal{F},\\mathcal{F}_a,\\mathcal{F}_b\\) refer to the feature output of different levels of layers in which \\(\\mathcal{F}\\) denotes the full resolution feature output while \\(\\mathcal{F}\\) with subscripts \\(a\\) and \\(b\\) represent the lower resolution feature outputs. We apply a prediction head (marked as \\(\\mathcal{H}_{\\eta}\\)) composed of two convolutional layers and a ReLU activation to transform these features into the target images with different scales." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.615, + 0.788, + 0.722 + ], + "angle": 0, + "content": "More importantly, we recurrently run the decoder network twice to refine the feature maps and the predicted images. The following experiments show the positive influence of such recurrent refinement on restoring the missing details due to the sparse point distribution. The input for the second loop is related to the output of the first loop. The prediction after the first loop can be described as \\( I^{l}, I_{a}^{l}, I_{b}^{l} = \\mathcal{H}_{\\eta}(\\mathcal{F}), \\mathcal{H}_{\\eta}(\\mathcal{F}_{a}), \\mathcal{H}_{\\eta}(\\mathcal{F}_{b}) \\) where \\( l = 1 \\) refers to the loop number. Likewise, the second loop is accordingly explained as:" + }, + { + "type": "equation", + "bbox": [ + 0.278, + 0.733, + 0.786, + 0.753 + ], + "angle": 0, + "content": "\\[\nI ^ {l = 2}, I _ {a} ^ {l = 2}, I _ {b} ^ {l = 2} = \\mathcal {H} _ {\\eta} \\star \\mathcal {D} _ {\\theta} ((\\mathcal {F} \\oplus I ^ {l = 1}), (\\mathcal {F} _ {a} \\oplus I _ {a} ^ {l = 1}), (\\mathcal {F} _ {b} \\oplus I _ {b} ^ {l = 1})) \\tag {9}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.765, + 0.788, + 0.842 + ], + "angle": 0, + "content": "in which \\(\\mathcal{H}_{\\eta} \\star \\mathcal{D}_{\\theta}\\) represents sequentially perform the two operations. We show the detailed architecture of the decoder in the Appendix. After we obtain the hierarchically multiscale predictions, we regard the full resolution result of the second loop \\(I^{l=2}\\) as the final prediction. However, we leverage all predictions to optimize the full models, which we introduce in the next section." + } + ], + [ + { + "type": "page_number", + "bbox": [ + 0.218, + 0.116, + 0.23, + 0.127 + ], + "angle": 0, + "content": "8" + }, + { + "type": "header", + "bbox": [ + 0.272, + 0.115, + 0.368, + 0.129 + ], + "angle": 0, + "content": "J. Wang et al." + }, + { + "type": "title", + "bbox": [ + 0.216, + 0.147, + 0.42, + 0.162 + ], + "angle": 0, + "content": "3.4 Training Strategies" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.18, + 0.788, + 0.288 + ], + "angle": 0, + "content": "The proposed framework incorporates 3D-aware features into multiscale and progressive decoding by using 3D Gaussian Splatting. The integration of feature volumes under certain camera settings mainly relies on the Gaussian properties that are used for projections. Therefore, we first pretrain the feature extractor and Gaussian regressor in order to obtain a relatively reasonable rendering function with 8 epochs. In the first training stage, we employ the \\(L2\\) rendering loss between the Gaussian predicted images and labels, i.e. \\(I_{gs}\\) in Eq. 7." + }, + { + "type": "equation", + "bbox": [ + 0.389, + 0.301, + 0.786, + 0.336 + ], + "angle": 0, + "content": "\\[\n\\mathcal {L} _ {g s} = \\sum_ {\\mathbf {u} \\in \\Omega} \\frac {1}{\\Omega} \\left| \\left| I _ {g s} (\\mathbf {u}) - I _ {g t} (\\mathbf {u}) \\right| \\right| ^ {2} \\tag {10}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.351, + 0.788, + 0.472 + ], + "angle": 0, + "content": "where \\(\\mathbf{u} = \\{u,v\\}\\) and \\(\\Omega\\) is the pixel domain. After the first pretraining, the regressor can yield plausible Gaussian rendering parameters, thus the rendered feature maps provide a certain and view-dependent description of the scene. Based on this, we combine the rest modules with these trained modules to carry out joint optimization. According to the above illustration, the final outputs of the framework include two levels, each comprising three images of different scales. We leverage all of the outputs to optimize the model by using progressively multiscale image loss." + }, + { + "type": "equation", + "bbox": [ + 0.334, + 0.488, + 0.787, + 0.528 + ], + "angle": 0, + "content": "\\[\n\\mathcal {L} _ {m i m} = \\sum_ {i = l} ^ {2} \\sum_ {s = r} ^ {3} \\sum_ {\\mathbf {u} \\in \\Omega} w (l) \\frac {1}{\\Omega} \\left| \\left| I _ {s} ^ {i} (\\mathbf {u}) - I _ {g t \\downarrow r _ {s}} (\\mathbf {u}) \\right| \\right| _ {1} ^ {1} \\tag {11}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.542, + 0.788, + 0.648 + ], + "angle": 0, + "content": "in which \\( l \\) refers to the loop number in the progressive decoder. \\( s \\) denotes the scale number in each output of a loop and \\( \\downarrow r_s \\) is \\( s \\)' corresponding downsampling operations. \\( w(l) \\) is a simple weighting function related to \\( l \\). Here we set \\( w(l) = 0.75 \\) when \\( l = 1 \\) otherwise \\( w(l) = 1 \\). In addition, we use the extra progressively multiscale frequency reconstruction loss to assist in restoring the high-frequency components, it is effective in reconstructing sharper edges and boundaries of images." + }, + { + "type": "equation", + "bbox": [ + 0.286, + 0.664, + 0.787, + 0.703 + ], + "angle": 0, + "content": "\\[\n\\mathcal {L} _ {m f r} = \\sum_ {i = l} ^ {2} \\sum_ {s = r} ^ {3} \\sum_ {\\mathbf {u} \\in \\Omega} w (l) \\frac {1}{\\Omega} | | F F T \\left(I _ {s} ^ {i} (\\mathbf {u})\\right) - F F T \\left(I _ {g t _ {\\downarrow r _ {s}}} (\\mathbf {u})\\right) | | _ {1} ^ {1} \\tag {12}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.717, + 0.788, + 0.776 + ], + "angle": 0, + "content": "in which the \\( FFT \\) represents the fast Fourier transform (FFT) that transfers the image signal to the frequency domain. The remaining symbols retain the same meanings as in the previous context. The total loss function can be defined as follows:" + }, + { + "type": "equation", + "bbox": [ + 0.378, + 0.785, + 0.786, + 0.8 + ], + "angle": 0, + "content": "\\[\n\\mathcal {L} _ {\\text {t o t a l}} = \\gamma_ {1} \\mathcal {L} _ {g s} + \\gamma_ {2} \\mathcal {L} _ {\\text {m i m}} + \\gamma_ {3} \\mathcal {L} _ {\\text {m f r}} \\tag {13}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.81, + 0.787, + 0.841 + ], + "angle": 0, + "content": "The \\(\\gamma_{1},\\gamma_{2}\\), and \\(\\gamma_{3}\\) respectively control weights of these losses. In our implementation, we set \\(\\gamma_{1} = 0.75\\), \\(\\gamma_{2} = 1\\), and \\(\\gamma_{3} = 0.25\\) respectively." + } + ], + [ + { + "type": "header", + "bbox": [ + 0.689, + 0.115, + 0.732, + 0.127 + ], + "angle": 0, + "content": "PFGS" + }, + { + "type": "page_number", + "bbox": [ + 0.776, + 0.117, + 0.786, + 0.127 + ], + "angle": 0, + "content": "9" + }, + { + "type": "table_caption", + "bbox": [ + 0.217, + 0.145, + 0.784, + 0.172 + ], + "angle": 0, + "content": "Table 1: Quantitative Results of comparisons between our method and different baselines under three datasets." + }, + { + "type": "table", + "bbox": [ + 0.217, + 0.18, + 0.798, + 0.304 + ], + "angle": 0, + "content": "
ScanNetDTUTHuman2.0
PSNR↑SSIM↑LPIPS↓PSNR↑SSIM↑LPIPS↓PSNR↑SSIM↑LPIPS↓
Graphics13.620.5280.77912.150.5250.68220.260.9050.337
NPBG15.090.5920.62513.520.7030.51419.770.9150.112
NPBG++16.810.6710.58522.320.8330.32726.810.9520.062
Trivol18.560.7340.47319.250.5920.51825.970.9350.059
Ours19.860.7580.45225.440.9010.16434.740.9830.009
" + }, + { + "type": "title", + "bbox": [ + 0.217, + 0.323, + 0.375, + 0.34 + ], + "angle": 0, + "content": "4 Experiments" + }, + { + "type": "title", + "bbox": [ + 0.217, + 0.352, + 0.416, + 0.368 + ], + "angle": 0, + "content": "4.1 Experiments setup" + }, + { + "type": "text", + "bbox": [ + 0.217, + 0.375, + 0.786, + 0.63 + ], + "angle": 0, + "content": "Datasets. We conduct extensive experiments to evaluate the effectiveness of our approach on three different datasets, including the indoor scene dataset ScanNet [4], objects dataset DTU [13], and human body dataset THuman2.0 [42]. The ScanNet dataset includes more than 2.5 million images at different views and with different qualities in over 1500 scenes. The point clouds are constructed from RGBD images. Following the Trivol, [9] we select the first 1200 scenes as the training set and the rest of the \\(300+\\) scenes as the test set. In addition, we use the preprocessed sparse point cloud in the dataset to test our model. For the DTU dataset, we transformed the point clouds of objects from their depth maps and images and we downsampled the point cloud by a factor of 0.5 to keep their point clouds distributing sparsely. In addition, we follow the MVSNeRF [2] to split the train and test groups of scans. The THuman2.0 dataset contains 500 high-quality human scans captured by a dense DLSR rig. For each scan, we render 36 views based on the provided 3D model and sparse sample points on the surface of the model. We keep the number of points for each scan at around 80,000 for their sparsity. We randomly select \\(75\\%\\) as the training set and the remaining \\(25\\%\\) as the test set." + }, + { + "type": "text", + "bbox": [ + 0.217, + 0.632, + 0.786, + 0.735 + ], + "angle": 0, + "content": "Baselines and Metrics. We compare the proposed method with the currently prevailing point cloud rendering approaches to demonstrate the effectiveness of our framework. The baselines include a pure graphics-based renderer, learning-based renderers NPBG [12] and \\(\\mathrm{NPBG + + }\\) [24], and the NeRF-based renderer Trivol [9]. As for the quality, we adopt PSNR, SSIM [36], and LPIPS [44]. These metrics overall determine the reconstruction accuracy and quality of these methods." + }, + { + "type": "title", + "bbox": [ + 0.217, + 0.757, + 0.67, + 0.772 + ], + "angle": 0, + "content": "4.2 Comparisons With Other Point Cloud Renderers." + }, + { + "type": "text", + "bbox": [ + 0.217, + 0.78, + 0.786, + 0.839 + ], + "angle": 0, + "content": "Fig. 2, 3, and 4 show the results of qualitative comparisons on the ScanNet, DTU, and THuman2.0 datasets respectively. Table. 1 reports all the metrics on the three datasets of the baselines and ours. The best result is bold, and the second-best result is underlined in the Table." + } + ], + [ + { + "type": "page_number", + "bbox": [ + 0.218, + 0.116, + 0.236, + 0.127 + ], + "angle": 0, + "content": "10" + }, + { + "type": "header", + "bbox": [ + 0.272, + 0.115, + 0.368, + 0.129 + ], + "angle": 0, + "content": "J. Wang et al." + }, + { + "type": "image", + "bbox": [ + 0.222, + 0.145, + 0.333, + 0.214 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.335, + 0.146, + 0.446, + 0.215 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.447, + 0.146, + 0.558, + 0.215 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.56, + 0.146, + 0.671, + 0.215 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.673, + 0.146, + 0.785, + 0.215 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.222, + 0.216, + 0.332, + 0.285 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.335, + 0.216, + 0.446, + 0.284 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.447, + 0.216, + 0.558, + 0.284 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.56, + 0.216, + 0.671, + 0.284 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.673, + 0.216, + 0.785, + 0.284 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.222, + 0.285, + 0.332, + 0.353 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.335, + 0.285, + 0.446, + 0.353 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.447, + 0.285, + 0.558, + 0.353 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.56, + 0.285, + 0.671, + 0.353 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.673, + 0.285, + 0.785, + 0.353 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.222, + 0.354, + 0.332, + 0.423 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.335, + 0.354, + 0.446, + 0.423 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.447, + 0.354, + 0.558, + 0.423 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.56, + 0.354, + 0.671, + 0.423 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.673, + 0.354, + 0.785, + 0.423 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.222, + 0.424, + 0.332, + 0.493 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.335, + 0.424, + 0.446, + 0.493 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.447, + 0.424, + 0.558, + 0.493 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.56, + 0.424, + 0.671, + 0.493 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.673, + 0.424, + 0.785, + 0.493 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.25, + 0.495, + 0.306, + 0.507 + ], + "angle": 0, + "content": "Graphics" + }, + { + "type": "image_caption", + "bbox": [ + 0.363, + 0.495, + 0.418, + 0.505 + ], + "angle": 0, + "content": "NPBG++" + }, + { + "type": "image_caption", + "bbox": [ + 0.486, + 0.495, + 0.522, + 0.505 + ], + "angle": 0, + "content": "Trivol" + }, + { + "type": "image_caption", + "bbox": [ + 0.6, + 0.495, + 0.633, + 0.506 + ], + "angle": 0, + "content": "Ours" + }, + { + "type": "image_caption", + "bbox": [ + 0.687, + 0.495, + 0.772, + 0.506 + ], + "angle": 0, + "content": "Ground Truth" + }, + { + "type": "image_caption", + "bbox": [ + 0.214, + 0.515, + 0.785, + 0.542 + ], + "angle": 0, + "content": "Fig. 2: Qualitative comparisons between ours and other point cloud renderers on the ScanNet dataset." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.552, + 0.788, + 0.734 + ], + "angle": 0, + "content": "ScanNet. The problem of uneven distribution and missing points exists in the real scene-level point cloud. As shown in the first column in Fig. 2, traditional graphical rendering leads to a large number of holes in the rendering result. When the point cloud distribution is uneven, the point cloud of nearby objects is relatively sparse, and the 2D-CNN of \\(\\mathrm{NPBG}++\\) pays more attention to the features of distant objects, ignoring the 3D information in the point cloud, resulting that nearby objects are covered in the rendering result. Our method uses 3DGS for feature projection, which better integrates 3D context information. There are a large number of artifacts in the results of \\(\\mathrm{NPBG}++\\) and Trivol. Our multi-scale feature extraction network excels at capturing features across various levels of detail, effectively decoding these projected features to enhance the rendering of lighting effects and details significantly." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.735, + 0.794, + 0.84 + ], + "angle": 0, + "content": "DTU. As we stated in the Dataset description part, we maintain points distributed sparsely on the surface of the objects. Under this circumstance, the graphics-based transformation demonstrates more serious bleeding surfaces. Trivol always partially reconstructs the whole object because it depends on the fully implicit neural representation. This setting is more likely to be suited to class-specific point rendering, as they claimed in the original paper. In contrast, DTU includes many different classes that obstacle Trivol to learning a unified feature" + } + ], + [ + { + "type": "header", + "bbox": [ + 0.689, + 0.115, + 0.732, + 0.127 + ], + "angle": 0, + "content": "PFGS" + }, + { + "type": "page_number", + "bbox": [ + 0.769, + 0.117, + 0.784, + 0.127 + ], + "angle": 0, + "content": "11" + }, + { + "type": "image", + "bbox": [ + 0.221, + 0.144, + 0.333, + 0.214 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.221, + 0.215, + 0.332, + 0.284 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.221, + 0.285, + 0.332, + 0.354 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.221, + 0.355, + 0.332, + 0.424 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.221, + 0.425, + 0.332, + 0.494 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.247, + 0.496, + 0.304, + 0.507 + ], + "angle": 0, + "content": "Graphics" + }, + { + "type": "image", + "bbox": [ + 0.333, + 0.145, + 0.445, + 0.214 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.334, + 0.215, + 0.445, + 0.284 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.334, + 0.285, + 0.445, + 0.354 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.334, + 0.355, + 0.445, + 0.424 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.334, + 0.425, + 0.445, + 0.494 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.365, + 0.496, + 0.42, + 0.506 + ], + "angle": 0, + "content": "NPBG++" + }, + { + "type": "image", + "bbox": [ + 0.447, + 0.145, + 0.558, + 0.214 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.447, + 0.215, + 0.557, + 0.284 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.447, + 0.285, + 0.557, + 0.354 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.447, + 0.355, + 0.557, + 0.424 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.447, + 0.425, + 0.557, + 0.494 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.486, + 0.496, + 0.524, + 0.506 + ], + "angle": 0, + "content": "Trivol" + }, + { + "type": "image", + "bbox": [ + 0.559, + 0.145, + 0.671, + 0.214 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.56, + 0.215, + 0.671, + 0.284 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.56, + 0.285, + 0.671, + 0.354 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.56, + 0.355, + 0.671, + 0.424 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.56, + 0.425, + 0.671, + 0.494 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.6, + 0.496, + 0.632, + 0.506 + ], + "angle": 0, + "content": "Ours" + }, + { + "type": "image", + "bbox": [ + 0.673, + 0.145, + 0.784, + 0.214 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.673, + 0.215, + 0.784, + 0.284 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.673, + 0.285, + 0.784, + 0.354 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.673, + 0.355, + 0.784, + 0.424 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.673, + 0.425, + 0.784, + 0.494 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.686, + 0.496, + 0.771, + 0.506 + ], + "angle": 0, + "content": "Ground Truth" + }, + { + "type": "image_caption", + "bbox": [ + 0.216, + 0.516, + 0.785, + 0.543 + ], + "angle": 0, + "content": "Fig. 3: Qualitative comparisons between ours and other point cloud rendering methods on the DTU dataset." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.557, + 0.785, + 0.617 + ], + "angle": 0, + "content": "transform underlying them. NPBG++ and our method deliver better results than Trivol because they leverage the original color and structure of point clouds without class-specific assumptions. Our approach yields the clearest predictions thanks to our more reasonable feature projection and decoding." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.619, + 0.787, + 0.753 + ], + "angle": 0, + "content": "THuman2.0 The Graphics-based method reconstructs the main structures of the human body while inevitably seeing densely packed holes. The other two baselines generate similar results that are much better than those of the traditional methods. The difference is that Trivol fails to faithfully recover some cloth colors, which is because it is a fully implicit method and thereby cannot restore colors that are never seen in its train set. \\(\\mathrm{NPBG}++\\) and ours directly utilize points' colors thus avoiding this domain generalization issue. Moreover, our method restores more details such as human facial expressions and clothes' folds." + }, + { + "type": "title", + "bbox": [ + 0.216, + 0.768, + 0.402, + 0.781 + ], + "angle": 0, + "content": "4.3 Ablation studies." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.795, + 0.785, + 0.84 + ], + "angle": 0, + "content": "Ablations of main components. In this subsection, we evaluate the effectiveness of all main constituents proposed in our whole framework. We conduct experiments with different control variables set and report their performance on" + } + ], + [ + { + "type": "page_number", + "bbox": [ + 0.218, + 0.116, + 0.236, + 0.127 + ], + "angle": 0, + "content": "12" + }, + { + "type": "header", + "bbox": [ + 0.272, + 0.115, + 0.368, + 0.129 + ], + "angle": 0, + "content": "J. Wang et al." + }, + { + "type": "image", + "bbox": [ + 0.226, + 0.149, + 0.785, + 0.559 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.216, + 0.568, + 0.788, + 0.595 + ], + "angle": 0, + "content": "Fig. 4: Qualitative comparisons between ours and other point cloud renderers on the THuman2.0 dataset." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.599, + 0.789, + 0.844 + ], + "angle": 0, + "content": "the ScanNet dataset in Table 2. In the table, \\( w / o \\) ProL refers to the version that removes all the progressive loss items, i.e. removes the first summation in Eq. 11 and Eq. 12. \\( w / o \\) MSL denotes the version that removes all the multiscale loss items that are the second summation in Eq. 11 and 12. The recur1 illustrates that we discard the recurrent structure from the feature decoder and only utilize a single multi-input and multi-output UNet. Besides, the \\( w / o \\) MSf describes a variant of the decoder which adopts a normal single scale UNet. In other words, this variant receives a single feature map as input and outputs a single prediction with full resolutions. Therefore, we only maintain the progressive losses and structures but discard all multiscale components for this version. Furthermore, we evaluate when the number of loops equals to 3, i.e. \\( l = 3 \\) in these equations. This is marked as recur3 in the Table. The results show that continuously increasing the number of iterations does not further improve the model's performance; instead, it may lead to a slight decline. Last, the \\( w / o \\) Lmfr represents the version of removing the progressively multiscale frequency reconstruction loss (Eq. 12)." + } + ], + [ + { + "type": "header", + "bbox": [ + 0.689, + 0.115, + 0.732, + 0.127 + ], + "angle": 0, + "content": "PFGS" + }, + { + "type": "page_number", + "bbox": [ + 0.769, + 0.116, + 0.785, + 0.127 + ], + "angle": 0, + "content": "13" + }, + { + "type": "table_caption", + "bbox": [ + 0.216, + 0.138, + 0.784, + 0.166 + ], + "angle": 0, + "content": "Table 2: Quantitative Results of ablation studies under various variants on the ScanNet dataset. The meaning of each abbreviation can be found in the text." + }, + { + "type": "table", + "bbox": [ + 0.22, + 0.179, + 0.782, + 0.254 + ], + "angle": 0, + "content": "
w/o ProLw/o MSLw/o MSfw/o Lmfrw/o 2trrecur1recur3Full
PSNR↑19.1919.3718.2719.4215.9219.1919.7719.86
SSIM↑0.7250.7330.7040.7400.6960.7290.7600.758
LPIPS↓0.4980.5090.6500.4780.6890.5160.4780.452
" + }, + { + "type": "image", + "bbox": [ + 0.22, + 0.267, + 0.784, + 0.347 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.216, + 0.351, + 0.784, + 0.378 + ], + "angle": 0, + "content": "Fig. 5: Qualitative Comparisons of the renderings produced by the Gaussian splatting after the first stage and the feature splatting after the whole framework." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.395, + 0.785, + 0.515 + ], + "angle": 0, + "content": "It is observed that the full model delivers the most satisfactory rendering results across all metrics while the rest variants experience a decline to some extent. This indicates that both progressive and multi-scale features have a positive impact and contribute to the final promising performance of the model. Therefore, the entire design of this method is consistently characterized by progressive and multi-scale aspects. Moreover, the alignment of predictions and labels in the frequency domain progressively and in a multiscale manner makes prediction clearer and sharper." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.538, + 0.785, + 0.734 + ], + "angle": 0, + "content": "Analysis of Gaussian and Feature Splatting. The proposed approach relies on the two-stage training scheme because the 3D-consistent feature should be projected and integrated into certain 2D planes given camera viewpoints. The projection of features is dependent on the 3D Gaussian rasterization and further reliant on the Gaussian parameters associated with each point which are estimated by Eq. 6. The significance of training the Gaussian regression network in the first training stage lies in laying a solid foundation for feature splatting. Even though the 3D Gaussian Splatting could also produce plausible images, namely the \\( I_{gs} \\) in Eq. 7, the results are still severely influenced by points that are either locally or globally sparse, thereby missing details and showing holes. It can be seen in Fig. 5 that these Gaussian splatting results are indeed better than the pure graphics-based method due to the representation capabilities of 3DGS, but they still need to be improved further." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.735, + 0.787, + 0.84 + ], + "angle": 0, + "content": "However, the results of feature splatting contain rich semantic information and local geometry cues, we consider them a good representation of restoring high-quality images and filling up blank areas. From Fig. 5 we can observe the capability of our method to generate and complete areas with voids and regions with sparse point clouds. We additionally evaluate an experiment in which we collaboratively train all modules in the framework rather than conducting the two-stage training and the results on the ScanNet dataset and we add this ad" + } + ], + [ + { + "type": "page_number", + "bbox": [ + 0.218, + 0.116, + 0.236, + 0.127 + ], + "angle": 0, + "content": "14" + }, + { + "type": "header", + "bbox": [ + 0.272, + 0.115, + 0.368, + 0.129 + ], + "angle": 0, + "content": "J. Wang et al." + }, + { + "type": "table_caption", + "bbox": [ + 0.217, + 0.145, + 0.785, + 0.16 + ], + "angle": 0, + "content": "Table 3: Ablation studies of the effect of different numbers of points on performance." + }, + { + "type": "table", + "bbox": [ + 0.218, + 0.172, + 0.781, + 0.248 + ], + "angle": 0, + "content": "
Train Points50k50k50k120k120k120k
Test Points50k80k120k50k80k120k
Trivol24.0525.4326.4824.7125.9626.86
Ours31.2932.8635.9531.6034.1739.18
" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.26, + 0.785, + 0.306 + ], + "angle": 0, + "content": "ditional result into Table. 2. \\( w / o \\) 2tr means that we abandon the first training stage and jointly train all modules together. In this setting, we observe that the model completely collapses and sees a very fast degradation in performance." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.324, + 0.788, + 0.492 + ], + "angle": 0, + "content": "Analysis of The Number of Points. We additionally analyze and evaluate how different training and test numbers of points affect the performance. On the THuman2.0 dataset, we set different training and test pairs to show the performance of our method and Trivol respectively, which includes 50,000 and 120,000 points used for training and 50,000, 80,000, and 120,000 points used for testing. The results are listed in Table 3. It is noted that Trivol is essentially unaffected by the number of points where our method sees significant fluctuations with changes in the number of points. However, our method exhibits very high performance when point clouds are abundant while the Trivol method has a relatively low upper limit, which is due to its fully implicit representation and lack of utilization of point cloud prior. More discussion can be seen in the Appendix." + }, + { + "type": "title", + "bbox": [ + 0.216, + 0.51, + 0.355, + 0.526 + ], + "angle": 0, + "content": "5 Discussion" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.538, + 0.788, + 0.765 + ], + "angle": 0, + "content": "Conclusion. This work proposes a novel point cloud rendering method called PFGS, that can render high-fidelity and photorealistic images from colored point clouds with arbitrary camera viewpoints. To address the shortcomings of current learning-based point cloud renderers such as slow, blurry, and inconsistent, we introduce the 3D Gaussian Splatting to point cloud rendering by several well-designed modules. We supervise the regressor to estimate Gaussian parameters in a point-wise way, and then we utilize the estimated Gaussian parameters to rasterize point-wise neural features that are extracted by a multiscale feature decoder. Furthermore, the projected feature volume would be gradually decoded to the final prediction by our progressive and multiscale decoder. The whole architecture is supervised by our finely-designed progressive and multiscale reconstruction losses. We carry out experiments on three datasets with different levels of objects including indoors, objects, and humans and the results show our method outperforms all baselines. The ablation studies also illustrate the significance of all main components in the approach." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.766, + 0.788, + 0.84 + ], + "angle": 0, + "content": "Limitation. Even though the proposed method can produce photorealistic images from sparse and uneven point clouds, it still yields unreal details when the given point clouds are not precise enough. Future works can incorporate a large generative refiner trained on a large dataset to slightly refine the original point cloud to make points closer to the actual surface of objects." + } + ], + [ + { + "type": "header", + "bbox": [ + 0.689, + 0.115, + 0.732, + 0.127 + ], + "angle": 0, + "content": "PFGS" + }, + { + "type": "page_number", + "bbox": [ + 0.769, + 0.117, + 0.786, + 0.127 + ], + "angle": 0, + "content": "15" + }, + { + "type": "title", + "bbox": [ + 0.218, + 0.145, + 0.323, + 0.16 + ], + "angle": 0, + "content": "References" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.177, + 0.785, + 0.218 + ], + "angle": 0, + "content": "1. Cao, A., Johnson, J.: Hexplane: A fast representation for dynamic scenes. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 130-141 (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.219, + 0.785, + 0.272 + ], + "angle": 0, + "content": "2. Chen, A., Xu, Z., Zhao, F., Zhang, X., Xiang, F., Yu, J., Su, H.: Mvsnerf: Fast generalizable radiance field reconstruction from multi-view stereo. In: Proceedings of the IEEE/CVF International Conference on Computer Vision. pp. 14124-14133 (2021)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.273, + 0.785, + 0.315 + ], + "angle": 0, + "content": "3. Choy, C., Gwak, J., Savarese, S.: 4d spatio-temporal convnets: Minkowski convolutional neural networks. In: Proceedings of the IEEE/CVF conference on computer vision and pattern recognition. pp. 3075-3084 (2019)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.316, + 0.785, + 0.357 + ], + "angle": 0, + "content": "4. Dai, A., Chang, A.X., Savva, M., Halber, M., Funkhouser, T., Nießner, M.: Scannet: Richly-annotated 3d reconstructions of indoor scenes. In: Proceedings of the IEEE conference on computer vision and pattern recognition. pp. 5828-5839 (2017)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.358, + 0.785, + 0.398 + ], + "angle": 0, + "content": "5. Dai, P., Zhang, Y., Li, Z., Liu, S., Zeng, B.: Neural point cloud rendering via multi-plane projection. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 7830-7839 (2020)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.399, + 0.785, + 0.44 + ], + "angle": 0, + "content": "6. Fridovich-Keil, S., Yu, A., Tancik, M., Chen, Q., Recht, B., Kanazawa, A.: Plenoxels: Radiance fields without neural networks. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 5501-5510 (2022)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.44, + 0.785, + 0.481 + ], + "angle": 0, + "content": "7. Ge, W., Hu, T., Zhao, H., Liu, S., Chen, Y.C.: Ref-neus: Ambiguity-reduced neural implicit surface learning for multi-view reconstruction with reflection. arXiv preprint arXiv:2303.10840 (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.482, + 0.785, + 0.509 + ], + "angle": 0, + "content": "8. Graham, B., Van der Maaten, L.: Submanifold sparse convolutional networks. arXiv preprint arXiv:1706.01307 (2017)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.509, + 0.785, + 0.55 + ], + "angle": 0, + "content": "9. Hu, T., Xu, X., Chu, R., Jia, J.: Trivol: Point cloud rendering via triple volumes. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 20732-20741 (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.55, + 0.785, + 0.592 + ], + "angle": 0, + "content": "10. Hu, T., Xu, X., Liu, S., Jia, J.: Point2pix: Photo-realistic point cloud rendering via neural radiance fields. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 8349-8358 (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.592, + 0.785, + 0.633 + ], + "angle": 0, + "content": "1. Huang, C., Mees, O., Zeng, A., Burgard, W.: Visual language maps for robot navigation. In: 2023 IEEE International Conference on Robotics and Automation (ICRA). pp. 10608-10615. IEEE (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.633, + 0.785, + 0.661 + ], + "angle": 0, + "content": "2. Jena, S., Multon, F., Boukhayma, A.: Neural mesh-based graphics. In: European Conference on Computer Vision. pp. 739-757. Springer (2022)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.661, + 0.785, + 0.702 + ], + "angle": 0, + "content": "3. Jensen, R., Dahl, A., Vogiatzis, G., Tola, E., Aanaes, H.: Large scale multi-view stereopsis evaluation. In: Proceedings of the IEEE conference on computer vision and pattern recognition. pp. 406-413 (2014)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.702, + 0.785, + 0.744 + ], + "angle": 0, + "content": "4. Jiang, Y., Yu, C., Xie, T., Li, X., Feng, Y., Wang, H., Li, M., Lau, H., Gao, F., Yang, Y., et al.: Vr-gs: A physical dynamics-aware interactive gaussian splattering system in virtual reality. arXiv preprint arXiv:2401.16663 (2024)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.744, + 0.785, + 0.772 + ], + "angle": 0, + "content": "5. Kerbl, B., Kopanas, G., Leimkuhler, T., Drettakis, G.: 3d gaussian splatting for real-time radiance field rendering. ACM Transactions on Graphics 42(4) (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.772, + 0.785, + 0.813 + ], + "angle": 0, + "content": "6. Lassner, C., Zollhofer, M.: Pulsar: Efficient sphere-based neural rendering. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 1440-1449 (2021)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.813, + 0.785, + 0.84 + ], + "angle": 0, + "content": "7. Liu, Z., Tang, H., Lin, Y., Han, S.: Point-voxel cnn for efficient 3d deep learning. Advances in Neural Information Processing Systems 32 (2019)" + }, + { + "type": "list", + "bbox": [ + 0.226, + 0.177, + 0.785, + 0.84 + ], + "angle": 0, + "content": null + } + ], + [ + { + "type": "page_number", + "bbox": [ + 0.218, + 0.116, + 0.236, + 0.127 + ], + "angle": 0, + "content": "16" + }, + { + "type": "header", + "bbox": [ + 0.272, + 0.115, + 0.367, + 0.129 + ], + "angle": 0, + "content": "J. Wang et al." + }, + { + "type": "ref_text", + "bbox": [ + 0.218, + 0.147, + 0.785, + 0.189 + ], + "angle": 0, + "content": "18. Lombardi, S., Simon, T., Saragih, J., Schwartz, G., Lehrmann, A., Sheikh, Y.: Neural volumes: Learning dynamic renderable volumes from images. arXiv preprint arXiv:1906.07751 (2019)" + }, + { + "type": "ref_text", + "bbox": [ + 0.218, + 0.19, + 0.787, + 0.245 + ], + "angle": 0, + "content": "19. Metzer, G., Richardson, E., Patashnik, O., Giryes, R., Cohen-Or, D.: Latent-nerf for shape-guided generation of 3d shapes and textures. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 12663-12673 (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.247, + 0.785, + 0.288 + ], + "angle": 0, + "content": "20. Mildenhall, B., Srinivasan, P.P., Tancik, M., Barron, J.T., Ramamoorthi, R., Ng, R.: Nerf: Representing scenes as neural radiance fields for view synthesis. Communications of the ACM 65(1), 99-106 (2021)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.289, + 0.785, + 0.33 + ], + "angle": 0, + "content": "21. Müller, T., Evans, A., Schied, C., Keller, A.: Instant neural graphics primitives with a multiresolution hash encoding. ACM Transactions on Graphics (ToG) 41(4), 1-15 (2022)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.332, + 0.785, + 0.373 + ], + "angle": 0, + "content": "22. Qi, C.R., Su, H., Nießner, M., Dai, A., Yan, M., Guibas, L.J.: Volumetric and multi-view cnns for object classification on 3d data. In: Proceedings of the IEEE conference on computer vision and pattern recognition. pp. 5648-5656 (2016)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.375, + 0.785, + 0.415 + ], + "angle": 0, + "content": "23. Qi, C.R., Yi, L., Su, H., Guibas, L.J.: Pointnet++: Deep hierarchical feature learning on point sets in a metric space. Advances in neural information processing systems 30 (2017)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.417, + 0.785, + 0.459 + ], + "angle": 0, + "content": "24. Rakhimov, R., Ardelean, A.T., Lempitsky, V., Burnaev, E.: Npgb++: Accelerating neural point-based graphics. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 15969-15979 (2022)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.46, + 0.785, + 0.5 + ], + "angle": 0, + "content": "25. Ravi, N., Reizenstein, J., Novotny, D., Gordon, T., Lo, W.Y., Johnson, J., Gkioxari, G.: Accelerating 3d deep learning with pytorch3d. arXiv preprint arXiv:2007.08501 (2020)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.502, + 0.785, + 0.557 + ], + "angle": 0, + "content": "26. Sitzmann, V., Thies, J., Heide, F., Nießner, M., Wetzstein, G., Zollhofer, M.: Deepvoxels: Learning persistent 3d feature embeddings. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 2437-2446 (2019)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.558, + 0.785, + 0.599 + ], + "angle": 0, + "content": "27. Sun, W., Trulls, E., Tseng, Y.C., Sambandam, S., Sharma, G., Tagliasacchi, A., Yi, K.M.: Pointner++: A multi-scale, point-based neural radiance field. arXiv preprint arXiv:2312.02362 (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.601, + 0.785, + 0.642 + ], + "angle": 0, + "content": "28. Wang, C., Chai, M., He, M., Chen, D., Liao, J.: Clip-nerf: Text-and-image driven manipulation of neural radiance fields. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 3835-3844 (2022)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.643, + 0.785, + 0.671 + ], + "angle": 0, + "content": "29. Wang, J., He, J., Zhang, Z., Xu, R.: Physical priors augmented event-based 3d reconstruction. arXiv preprint arXiv:2401.17121 (2024)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.672, + 0.785, + 0.713 + ], + "angle": 0, + "content": "30. Wang, J., Zhang, Z., Xu, R.: Learning to generate and manipulate 3d radiance field by a hierarchical diffusion framework with clip latent. In: Computer Graphics Forum. vol. 42, p. e14930. Wiley Online Library (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.714, + 0.785, + 0.755 + ], + "angle": 0, + "content": "31. Wang, J., Zhang, Z., Xu, R.: Learning robust generalizable radiance field with visibility and feature augmented point representation. arXiv preprint arXiv:2401.14354 (2024)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.757, + 0.785, + 0.798 + ], + "angle": 0, + "content": "32. Wang, J., Sun, B., Lu, Y.: Mvpnet: Multi-view point regression networks for 3d object reconstruction from a single image. In: Proceedings of the AAAI Conference on Artificial Intelligence. vol. 33, pp. 8949-8956 (2019)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.799, + 0.785, + 0.84 + ], + "angle": 0, + "content": "33. Wang, N., Zhang, Y., Li, Z., Fu, Y., Liu, W., Jiang, Y.G.: Pixel2mesh: Generating 3d mesh models from single rgb images. In: Proceedings of the European conference on computer vision (ECCV). pp. 52-67 (2018)" + }, + { + "type": "list", + "bbox": [ + 0.217, + 0.147, + 0.787, + 0.84 + ], + "angle": 0, + "content": null + } + ], + [ + { + "type": "header", + "bbox": [ + 0.689, + 0.115, + 0.732, + 0.127 + ], + "angle": 0, + "content": "PFGS" + }, + { + "type": "page_number", + "bbox": [ + 0.769, + 0.116, + 0.786, + 0.127 + ], + "angle": 0, + "content": "17" + }, + { + "type": "ref_text", + "bbox": [ + 0.214, + 0.147, + 0.788, + 0.204 + ], + "angle": 0, + "content": "34. Wang, P., Liu, Y., Chen, Z., Liu, L., Liu, Z., Komura, T., Theobalt, C., Wang, W.: F2-nerf: Fast neural radiance field training with free camera trajectories. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 4150-4159 (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.214, + 0.205, + 0.787, + 0.245 + ], + "angle": 0, + "content": "35. Wang, S., Liu, M.: Point cloud compression with range image-based entropy model for autonomous driving. In: European Conference on Computer Vision. pp. 323-340. Springer (2022)" + }, + { + "type": "ref_text", + "bbox": [ + 0.216, + 0.246, + 0.786, + 0.286 + ], + "angle": 0, + "content": "36. Wang, Z., Bovik, A.C., Sheikh, H.R., Simoncelli, E.P.: Image quality assessment: from error visibility to structural similarity. IEEE transactions on image processing 13(4), 600-612 (2004)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.287, + 0.787, + 0.327 + ], + "angle": 0, + "content": "37. Wu, G., Yi, T., Fang, J., Xie, L., Zhang, X., Wei, W., Liu, W., Tian, Q., Wang, X.: 4d gaussian splatting for real-time dynamic scene rendering. arXiv preprint arXiv:2310.08528 (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.328, + 0.787, + 0.383 + ], + "angle": 0, + "content": "38. Xiang, F., Xu, Z., Hasan, M., Hold-Geoffroy, Y., Sunkavalli, K., Su, H.: Neutex: Neural texture mapping for volumetric neural rendering. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 7119-7128 (2021)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.384, + 0.787, + 0.425 + ], + "angle": 0, + "content": "39. Xu, Q., Xu, Z., Philip, J., Bi, S., Shu, Z., Sunkavalli, K., Neumann, U.: Point-nerf: Point-based neural radiance fields. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 5438-5448 (2022)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.426, + 0.787, + 0.467 + ], + "angle": 0, + "content": "40. Yan, Z., Li, C., Lee, G.H.: Nerf-ds: Neural radiance fields for dynamic specular objects. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 8285-8295 (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.468, + 0.787, + 0.522 + ], + "angle": 0, + "content": "41. Yang, B., Bao, C., Zeng, J., Bao, H., Zhang, Y., Cui, Z., Zhang, G.: Neumesh: Learning disentangled neural mesh-based implicit field for geometry and texture editing. In: European Conference on Computer Vision. pp. 597-614. Springer (2022)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.523, + 0.787, + 0.563 + ], + "angle": 0, + "content": "42. Yu, T., Zheng, Z., Guo, K., Liu, P., Dai, Q., Liu, Y.: Function4d: Real-time human volumetric capture from very sparse consumer rgbd sensors. In: IEEE Conference on Computer Vision and Pattern Recognition (CVPR2021) (June 2021)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.565, + 0.787, + 0.605 + ], + "angle": 0, + "content": "43. Zhang, C., Guo, R., Zeng, W., Xiong, Y., Dai, B., Hu, R., Ren, M., Urtasun, R.: Rethinking closed-loop training for autonomous driving. In: European Conference on Computer Vision. pp. 264-282. Springer (2022)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.606, + 0.787, + 0.646 + ], + "angle": 0, + "content": "44. Zhang, R., Isola, P., Efros, A.A., Shechtman, E., Wang, O.: The unreasonable effectiveness of deep features as a perceptual metric. In: Proceedings of the IEEE conference on computer vision and pattern recognition. pp. 586-595 (2018)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.647, + 0.787, + 0.674 + ], + "angle": 0, + "content": "45. Zhou, Q.Y., Park, J., Koltun, V.: Open3d: A modern library for 3d data processing. arXiv preprint arXiv:1801.09847 (2018)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.675, + 0.787, + 0.73 + ], + "angle": 0, + "content": "46. Zhou, S., Chang, H., Jiang, S., Fan, Z., Zhu, Z., Xu, D., Chari, P., You, S., Wang, Z., Kadambi, A.: Feature 3dgs: Supercharging 3d gaussian splatting to enable distilled feature fields. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 21676-21685 (2024)" + }, + { + "type": "list", + "bbox": [ + 0.214, + 0.147, + 0.788, + 0.73 + ], + "angle": 0, + "content": null + } + ] +] \ No newline at end of file diff --git a/2024/PFGS_ High Fidelity Point Cloud Rendering via Feature Splatting/685f2bf6-434e-4901-92a2-92656824eb88_origin.pdf b/2024/PFGS_ High Fidelity Point Cloud Rendering via Feature Splatting/685f2bf6-434e-4901-92a2-92656824eb88_origin.pdf new file mode 100644 index 0000000000000000000000000000000000000000..08683818ba3750c33ad7a4349e3095956835385a --- /dev/null +++ b/2024/PFGS_ High Fidelity Point Cloud Rendering via Feature Splatting/685f2bf6-434e-4901-92a2-92656824eb88_origin.pdf @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:8cda94eea96970889da0b84addbdd2933f8c5f660d02eaf198e83db190f8062f +size 4777243 diff --git a/2024/PFGS_ High Fidelity Point Cloud Rendering via Feature Splatting/full.md b/2024/PFGS_ High Fidelity Point Cloud Rendering via Feature Splatting/full.md new file mode 100644 index 0000000000000000000000000000000000000000..1759b3a7901a2f7e03cf065320e92246391a42d7 --- /dev/null +++ b/2024/PFGS_ High Fidelity Point Cloud Rendering via Feature Splatting/full.md @@ -0,0 +1,387 @@ +# PFGS: High Fidelity Point Cloud Rendering via Feature Splatting + +Jiaxu Wang $^{1\dagger}$ , Ziyi Zhang $^{1\dagger}$ , Junhao He $^{1}$ , and Renjing Xu $^{1\boxplus}$ + +$^{1}$ Hong Kong University of Science and Technology (GZ) jwang457@connect.hkust-gz.edu.cn {ziyizhang, junhaohe, renjingxu}@hkust-gz.edu.cn + +Abstract. Rendering high-fidelity images from sparse point clouds is still challenging. Existing learning-based approaches suffer from either hole artifacts, missing details, or expensive computations. In this paper, we propose a novel framework to render high-quality images from sparse points. This method first attempts to bridge the 3D Gaussian Splatting and point cloud rendering, which includes several cascaded modules. We first use a regressor to estimate Gaussian properties in a point-wise manner, the estimated properties are used to rasterize neural feature descriptors into 2D planes which are extracted from a multiscale extractor. The projected feature volume is gradually decoded toward the final prediction via a multiscale and progressive decoder. The whole pipeline experiences a two-stage training and is driven by our well-designed progressive and multiscale reconstruction loss. Experiments on different benchmarks show the superiority of our method in terms of rendering qualities and the necessities of our main components. $^{1}$ + +Keywords: Point cloud rendering $\cdot$ 3D Gaussian Splatting + +# 1 Introduction + +Synthesizing photorealistic images from given colored point clouds and arbitrary camera views can be extensively applied to various fields including virtual/augmented reality [14], robotic navigation [11], automatic driving [35, 43], etc, which is academically named point cloud rendering. Traditional methods to render views from point clouds are based on graphics transformations. They directly project 3D point clouds into 2D planes by using the camera parameters and z-buffer rasterization. However, the sparse distribution of points leads to bleeding surfaces and hole artifacts. Even though the graphics renderer can treat points as graphical primitives and diminish these issues by color blending, rough and simple geometries of primitives entail blurred images and missing details. + +In recent years, learning-based methods [9, 10, 24] have become prevailing to address the drawbacks of conventional graphics-based approaches. These methods project points and corresponding feature descriptors to 2D planes, obtaining sparse images. Then they use UNet-like networks to complete and restore + +the low-quality images [12, 24]. Nevertheless, these methods rely on 2D image restoration, which cannot guarantee 3D consistency and can result in the distinctive appearance of the same object from nearby different viewpoints. Moreover, their feature descriptors are associated with points that do not possess actual shapes and volumes in real space, thereby showing poor descriptions of local area structures. To obtain 3D-aware and consistent features, some approaches [8, 16] transform points to 3D volumes by trilinear interpolation and use 3D convolution to extract the features. However, they lead to a significant computational burden and sacrifice the efficient representation characteristics of point clouds. For example, voxels uniformly represent scenes, whereas points are densely distributed in complex areas and sparsely distributed or even absent in empty spaces. + +To obtain efficient and consistent 3D representation, the Neural Radiance Field (NeRF) [20] has gradually been incorporated into point cloud renderers [9, 10]. These methods either consider points as anchors of NeRF or transform points into NeRF-like representations, then use the physical-based volume rendering to obtain 3D-aware feature maps which are decoded into corresponding images. With the assistance of NeRF, they achieve high-quality consistencies across nearby views and plausible images. However, discrete NeRF requires heavy memory consumption resulting from storing high-dimensional feature vectors. In addition, NeRF typically requires lots of sampling points and multiple accesses to the neural network, which significantly increases the rendering time and makes it challenging to meet real-time requirements. + +More recently, 3D Gaussian Splatting (3DGS) [15] introduces a new representation that formulates points as 3D Gaussians with learnable parameters including 3D position, color, opacity, and anisotropic covariance. 3DGS achieves more photorealistic and faster rendering than NeRF by applying Gaussian rasterization and $\alpha$ -blending. Conventional 3D Gaussian learns the primitive parameters associated with an initialized point cloud from multi-view images by backpropagation, it relies on per-subject parameter optimizations for several tens minutes. It is therefore impractical to be applied to the case in which only point cloud is available due to the need for multiview images and backpropagations. + +In this work, we fill the gap between point cloud render and the 3DGS technique by introducing a novel multiscale, feature-augmented, 3DGS-based point renderer. Compared with previous learning-based point renderers, our method yields higher-quality and 3D consistent results. Compared with conventional 3DGS, our method does not necessarily require multiview images or fine-tuning procedures and can be directly employed to point modality. + +We leverage an efficient multiscale feature extractor in the proposed framework to extract features from point cloud input. This extractor considers multiscale local geometries to extract discriminative feature descriptors. Then a regressor is used for estimating 3D Gaussian parameters in a point-wise manner. Next, we not only render images by these estimated Gaussian attributes in a multiscale way but also render the multiscale feature maps by the same Gaussian properties. Finally, the rendered multiscale images and feature maps are fused by a multiscale and recurrent UNet-like network to gradually transform + +into the final results. To the best knowledge of us, this is the first point cloud renderer combined with 3DGS. + +Our main contributions can be summarized in the following: + +1. We proposed a novel lightweight approach called PFGS (Point Feature Gaussian Splitting) that bridges point cloud rendering and 3DGS. It produces higher-fidelity images from sparse point clouds than currently prevailing approaches. +2. We proposed the point-based multiscale and feature-augmented 3D Gaussian framework, which consists of three main components including a multiscale feature extractor, a Gaussian regressor, and a multiscale and recurrent decoder, accompanied by the well-designed multiscale and progressive reconstruction losses. +3. Extensive experiments and ablation studies on three datasets show the effectiveness and advantages of the proposed method over current works in terms of reconstruction quality. + +# 2 Related Work + +# 2.1 Point Cloud Render + +Conventional point cloud rendering is implemented via graphics-based transformation [25,45] which simulates camera imaging processes to project points onto the 2D screens. Even though the rendering pipeline is fast and general for all scenes, it is heavily dependent on the density of point distributions. These methods produce empty holes and vacant areas when the distributions are sparse. + +Learning-based point renderers [3, 9, 10, 12, 24] to some extent handle the issues caused by traditional methods, which fill the holes by leveraging multiscale or multiview feature fusions. ME [3] implements a sparse Convolutional Neural Network to extract features from existing points and then calculates the features of arbitrary 3D points through ball querying. NPBG [12] enhances each point with a feature descriptor to encapsulate local geometry and appearance. Moreover, $\mathrm{NPBG}++$ [24] introduces feature aggregation from multiview based on the NPBG to complement missing parts. NPCR [5] employs a 3DCNN to generate 3D volumes from point clouds and creates multiple depth layers to synthesize images. More recently, Point2pix [10] and Trivol [9] have incorporated NeRF into the point cloud rendering pipeline. They transform raw point clouds into NeRF-related representations and utilize volume rendering to access the representations to produce images. However, the above methods still cannot fill large vacant holes and only produce blurry rendering results, especially in local areas with fewer points. Even if incorporating NeRF representation is effective in generating 3D consistent feature encoding, they do not fully utilize the geometric information contained within the point clouds. In this work, we combine point rendering with a novel 3D representation, i.e. 3DGS, and implement multiscale and progressive feature decoding to obtain fine-grained and 3D-consistent predictions. + +# 2.2 Neural-based View Synthesis + +Unlike traditional explicit representations [22, 23, 32, 33], implicit representations that encode scenes with the assistance of neural networks have been developed in recent years [18, 26, 38]. Among them, the Neural Radiance Field (NeRF) [20] attracts the most attention in both industries and research. NeRF is a continuous method that implicitly represents a 3D scene employing an MLP and synthesizing images of novel view by volume rendering. Recent studies have employed NeRF for various purposes, including dynamic reconstruction [1, 37, 40], physical reasoning [29], generative models [19, 28, 30], and fast rendering [6, 21, 34]. To achieve various goals, some studies attempted to combine the implicit NeRF with explicit primitives. For example, NeuMesh [41] regards mesh as a neural scaffold to store NeRF and achieve neural editing. PointNeRF [39] is the first to incorporate point clouds into NeRF reconstruction to obtain better rendering qualities and faster speed. Ref-NeuS [7] implements generalizable neural surface reconstruction. GPF [31] explicitly models the visibility by using the geometries in the point cloud. PointNeRF++ [27] solves point cloud sparsity by aggregating multiscale point clouds on sparse voxel grids at different resolutions. + +It can be seen that many approaches we list above incorporate points into NeRF. However, NeRF-like volume rendering does not fully explore the rendering capability of the point cloud. In contrast to NeRF, 3DGS [15] proposed the 3D Gaussian Splatting to realize a remarkable differentiable rendering speed and yield high-quality 2D images. In this work, we first bridge point cloud rendering and 3DGS to generate higher-fidelity images. + +# 3 Methodology + +Given a colored point cloud $\mathbf{P} = \{\mathbf{p}_k, \mathbf{c}_k\}$ , we aim to synthesize photorealistic images from any viewpoints defined by the camera intrinsic $K$ and poses $P$ are provided. The $\mathbf{p}_k \in R^3$ and $\mathbf{c}_k \in R^3$ are the point coordinates and colors respectively. The rendering process can be illustrated as $I_v = Re(\mathbf{P}|K_v, P_v)$ where $Re$ is the rendering function that can be implemented by either graphics-based or learning-based methods as we introduced above. In this work, it is implemented by our proposed multiscale feature-based 3DGS rendering pipeline that contains a multiscale feature extractor, a Gaussian regressor, a feature-based Gaussian rendering module, and a multiscale recurrent decoder. All modules are implemented with small networks. In the following of this section, we sequentially introduce the aforementioned components. + +# 3.1 Preliminaries + +3DGS is an explicit representation method for 3D scenes, which parameterizes the scene as a series of 3D Gaussian primitives. A 3D Gaussian is defined by a full 3D covariance matrix $\Sigma$ , its center point $x$ , and the spherical harmonic $(SH)$ . The mean value of the Gaussian is defined as: + +$$ +G (x) = e ^ {- \frac {1}{2} (x) ^ {T} \Sigma^ {- 1} (x)} \tag {1} +$$ + +![](images/50f7342bfdb9038ce4d05ff8d6c3ccdeebee297f41a9c9ee71b730e700512a80.jpg) + +![](images/36569bd060fabdb2b2ff0f2eab6037a3db14d1a5df81990b0b4da9e711d5ce22.jpg) +Fig. 1: The main pipeline of the proposed approach. The above panel includes the multiscale feature extractor and Gaussian regressor. Both Gaussian and features are splatted to the 2D plane and concatenated and fed to the progressive and multiscale feature decoding module described in the low panel. + +To enable optimization via backpropagation, the covariance matrix could be decomposed into a rotation matrix $(R)$ and a scaling matrix $(S)$ : + +$$ +\Sigma = R S S ^ {T} R ^ {T} \tag {2} +$$ + +Given the camera trajectory, the projection of the 3D Gaussians to the 2D image plane can be characterized by a view transform matrix $(W)$ and the Jacobian of the affine approximation of the projective transformation $(J)$ , as in: + +$$ +\Sigma^ {\prime} = J W \Sigma W ^ {T} J ^ {T} \tag {3} +$$ + +where the $\Sigma^{\prime}$ is the covariance matrix in 2D image planes. Thus, the $\alpha$ -blend of $\mathcal{N}$ ordered points overlapping a pixel is utilized to compute the final color $C$ of the pixel: + +$$ +C = \sum_ {i \in \mathcal {N}} c _ {i} \alpha_ {i} \prod_ {j = 1} ^ {i - 1} (1 - \alpha_ {j}) \tag {4} +$$ + +where $c_{i}$ and $\alpha_{i}$ denote the color and density of the pixel with corresponding Gaussian parameters. Following [46], we projected the features similarly. + +In summary, the above parameters can be characterized by the following attributes: $x$ denotes the position of a primitive where $x \in R^3$ . The rotation + +matrix is characterized by a quaternion $q \in R^4$ . The scale factor is defined by the anisotropy stretching $s \in R^3$ . The opacity factor $\alpha \in R^1 \in [0,1]$ . The 2D opacity $\alpha \in [0,1]$ is computed by $\alpha_i(x) = o_i \exp(-\frac{1}{2} (x - \mu_i)^T \Sigma_i^T (x - \mu_i))$ where the $\mu$ and variance are the 2D-projected mean and variance of 3D Gaussians. And the color is represented by $SH$ . + +# 3.2 Gaussian Feature Prediction with Multiscale Feature Extraction + +We first extract features for each point as their neural descriptors. We use a multi-input, single-output UNet-like architecture as the extraction network to encode points with different scales and capture features across different spatial distances. The architecture is mainly made of several PV Convolution modules [17], which is a highly efficient and commonly used point-based feature integration operation. We downsample the original point cloud with decreasing rates and concatenate the downsampled point clouds with different levels of feature maps as extra inputs. + +$$ +f _ {1: k} = \mathcal {E} _ {\theta} \left(\left(p _ {1: k}, c _ {1: k}\right), \left(p _ {1: k}, c _ {1: k}\right) _ {\downarrow r _ {1}}, \left(p _ {1: k}, c _ {1: k}\right) _ {\downarrow r _ {2}}\right) \tag {5} +$$ + +The $\mathcal{E}_{\theta}$ refers to the extractor network, $p, c$ denote point coordinates and colors. The down arrow represents a uniform downsampling operation and $r$ denotes the sampling rate where $r_1 > r_2$ . The output $f_{1:k}$ are point neural descriptors. This operation enables the extractor to recognize features of different scales and receptive fields. The detailed architecture and parameter setting are shown in our Appendix. The single output feature vector is considered as the descriptors for all points. Afterward, the neural descriptors, the coordinates, as well as colors are fed to another network to predict the Gaussian properties in a point-wise manner. + +$$ +F _ {k} = \mathcal {R} _ {\theta} \left(f _ {k}, p _ {k}, c _ {k}\right) \tag {6} +$$ + +$$ +R, S, \alpha = \operatorname {n o r m} \left(\mathcal {H} _ {r} \left(F _ {k}\right)\right), \exp \left(\mathcal {H} _ {s} \left(F _ {k}\right)\right), \sigma \left(\mathcal {H} _ {o} \left(F _ {k}\right)\right) +$$ + +where $\mathcal{R}_{\theta}$ represents the main network of the regressor, it maps multiple inputs to a single unified feature vector $F_{k}$ . The $F_{k}$ is fed to three individual prediction heads $(\mathcal{H}_r,\mathcal{H}_s,\mathcal{H}_o)$ , each of them is composed of two 1D Convolution layers with 1 kernel size, and activated by different functions to obtain corresponding Gaussian properties, namely rotation quaternion $(R)$ , scale factor $(S)$ and opacity $(\alpha)$ . The color property $c$ remains that of the original points. + +After the Gaussian attributes are obtained, one can render corresponding images from given camera viewpoints by Gaussian rasterization. We not only render the images by blending point colors on 2D planes but also render feature planes by replacing colors with the neural descriptors $(f_{k})$ . + +$$ +I _ {g s}, F _ {g s} = R e (R, S, \alpha , p _ {1: k}, c _ {1: k} | K, P), R e (R, S, \alpha , p _ {1: k}, f _ {1: k} | K, P) \tag {7} +$$ + +where $I_{gs}$ , $F_{gs}$ are corresponding image and feature map under the camera intrinsic $K$ and pose $P$ . The $I_{gs}$ can be considered a good result of point rendering when the number of points is sufficient. + +However, when the point sparsely distributes over the space, the representation capability of corresponding Gaussians is constrained, especially in complex local areas, resulting in poor-quality rendering results. On the contrary, the neural descriptors contain rich information to describe local structures with various scales. We therefore utilize the $F_{gs}$ to be gradually transformed to the final results in a multiscale and recurrent manner. The $F_{gs}$ is 3D-consistently rasterized with the same Gaussian parameters in Eq. 7, thus it precisely corresponds to the $I_{gs}$ in pixel-level. In other words, $F_{gs}$ can be regarded as a complement to $I_{gs}$ . Hence we concatenate $I_{gs}$ and $F_{gs}$ to be the input of the next module. + +# 3.3 Multiscale and Progressive Feature Decoding + +After we render the feature volume $F_{gs}$ associated with the provided camera viewpoint, we concatenate the $I_{gs}$ and $F_{gs}$ as input of the decoder. The decoder is a multi-input and multi-output UNet-like architecture, which aims to gradually refine and decode the input features and transform them into the target view images. We downsample the input two times with the rate of $r_a$ and $r_b$ and fuse them with the next two low-resolution layers of the UNet. There are interlaced connections between layers of different levels in the network, as stated in the main Fig. 1. + +$$ +\mathcal {F}, \mathcal {F} _ {a}, \mathcal {F} _ {b} = \mathcal {D} _ {\theta} \left(\left(F _ {g s} \oplus I _ {g s}\right), \left(F _ {g s} \oplus I _ {g s}\right) _ {\downarrow r _ {a}}, \left(F _ {g s} \oplus I _ {g s}\right) _ {\downarrow r _ {b}}\right) \tag {8} +$$ + +In the above equation, $\mathcal{D}_{\theta}$ indicates the multiscale decoder. $\downarrow r_{a}$ and $\downarrow r_{b}$ are two downsampling operations. $\mathcal{F},\mathcal{F}_a,\mathcal{F}_b$ refer to the feature output of different levels of layers in which $\mathcal{F}$ denotes the full resolution feature output while $\mathcal{F}$ with subscripts $a$ and $b$ represent the lower resolution feature outputs. We apply a prediction head (marked as $\mathcal{H}_{\eta}$ ) composed of two convolutional layers and a ReLU activation to transform these features into the target images with different scales. + +More importantly, we recurrently run the decoder network twice to refine the feature maps and the predicted images. The following experiments show the positive influence of such recurrent refinement on restoring the missing details due to the sparse point distribution. The input for the second loop is related to the output of the first loop. The prediction after the first loop can be described as $I^{l}, I_{a}^{l}, I_{b}^{l} = \mathcal{H}_{\eta}(\mathcal{F}), \mathcal{H}_{\eta}(\mathcal{F}_{a}), \mathcal{H}_{\eta}(\mathcal{F}_{b})$ where $l = 1$ refers to the loop number. Likewise, the second loop is accordingly explained as: + +$$ +I ^ {l = 2}, I _ {a} ^ {l = 2}, I _ {b} ^ {l = 2} = \mathcal {H} _ {\eta} \star \mathcal {D} _ {\theta} ((\mathcal {F} \oplus I ^ {l = 1}), (\mathcal {F} _ {a} \oplus I _ {a} ^ {l = 1}), (\mathcal {F} _ {b} \oplus I _ {b} ^ {l = 1})) \tag {9} +$$ + +in which $\mathcal{H}_{\eta} \star \mathcal{D}_{\theta}$ represents sequentially perform the two operations. We show the detailed architecture of the decoder in the Appendix. After we obtain the hierarchically multiscale predictions, we regard the full resolution result of the second loop $I^{l=2}$ as the final prediction. However, we leverage all predictions to optimize the full models, which we introduce in the next section. + +# 3.4 Training Strategies + +The proposed framework incorporates 3D-aware features into multiscale and progressive decoding by using 3D Gaussian Splatting. The integration of feature volumes under certain camera settings mainly relies on the Gaussian properties that are used for projections. Therefore, we first pretrain the feature extractor and Gaussian regressor in order to obtain a relatively reasonable rendering function with 8 epochs. In the first training stage, we employ the $L2$ rendering loss between the Gaussian predicted images and labels, i.e. $I_{gs}$ in Eq. 7. + +$$ +\mathcal {L} _ {g s} = \sum_ {\mathbf {u} \in \Omega} \frac {1}{\Omega} \left| \left| I _ {g s} (\mathbf {u}) - I _ {g t} (\mathbf {u}) \right| \right| ^ {2} \tag {10} +$$ + +where $\mathbf{u} = \{u,v\}$ and $\Omega$ is the pixel domain. After the first pretraining, the regressor can yield plausible Gaussian rendering parameters, thus the rendered feature maps provide a certain and view-dependent description of the scene. Based on this, we combine the rest modules with these trained modules to carry out joint optimization. According to the above illustration, the final outputs of the framework include two levels, each comprising three images of different scales. We leverage all of the outputs to optimize the model by using progressively multiscale image loss. + +$$ +\mathcal {L} _ {m i m} = \sum_ {i = l} ^ {2} \sum_ {s = r} ^ {3} \sum_ {\mathbf {u} \in \Omega} w (l) \frac {1}{\Omega} \left| \left| I _ {s} ^ {i} (\mathbf {u}) - I _ {g t \downarrow r _ {s}} (\mathbf {u}) \right| \right| _ {1} ^ {1} \tag {11} +$$ + +in which $l$ refers to the loop number in the progressive decoder. $s$ denotes the scale number in each output of a loop and $\downarrow r_s$ is $s$ ' corresponding downsampling operations. $w(l)$ is a simple weighting function related to $l$ . Here we set $w(l) = 0.75$ when $l = 1$ otherwise $w(l) = 1$ . In addition, we use the extra progressively multiscale frequency reconstruction loss to assist in restoring the high-frequency components, it is effective in reconstructing sharper edges and boundaries of images. + +$$ +\mathcal {L} _ {m f r} = \sum_ {i = l} ^ {2} \sum_ {s = r} ^ {3} \sum_ {\mathbf {u} \in \Omega} w (l) \frac {1}{\Omega} | | F F T \left(I _ {s} ^ {i} (\mathbf {u})\right) - F F T \left(I _ {g t _ {\downarrow r _ {s}}} (\mathbf {u})\right) | | _ {1} ^ {1} \tag {12} +$$ + +in which the $FFT$ represents the fast Fourier transform (FFT) that transfers the image signal to the frequency domain. The remaining symbols retain the same meanings as in the previous context. The total loss function can be defined as follows: + +$$ +\mathcal {L} _ {\text {t o t a l}} = \gamma_ {1} \mathcal {L} _ {g s} + \gamma_ {2} \mathcal {L} _ {\text {m i m}} + \gamma_ {3} \mathcal {L} _ {\text {m f r}} \tag {13} +$$ + +The $\gamma_{1},\gamma_{2}$ , and $\gamma_{3}$ respectively control weights of these losses. In our implementation, we set $\gamma_{1} = 0.75$ , $\gamma_{2} = 1$ , and $\gamma_{3} = 0.25$ respectively. + +Table 1: Quantitative Results of comparisons between our method and different baselines under three datasets. + +
ScanNetDTUTHuman2.0
PSNR↑SSIM↑LPIPS↓PSNR↑SSIM↑LPIPS↓PSNR↑SSIM↑LPIPS↓
Graphics13.620.5280.77912.150.5250.68220.260.9050.337
NPBG15.090.5920.62513.520.7030.51419.770.9150.112
NPBG++16.810.6710.58522.320.8330.32726.810.9520.062
Trivol18.560.7340.47319.250.5920.51825.970.9350.059
Ours19.860.7580.45225.440.9010.16434.740.9830.009
+ +# 4 Experiments + +# 4.1 Experiments setup + +Datasets. We conduct extensive experiments to evaluate the effectiveness of our approach on three different datasets, including the indoor scene dataset ScanNet [4], objects dataset DTU [13], and human body dataset THuman2.0 [42]. The ScanNet dataset includes more than 2.5 million images at different views and with different qualities in over 1500 scenes. The point clouds are constructed from RGBD images. Following the Trivol, [9] we select the first 1200 scenes as the training set and the rest of the $300+$ scenes as the test set. In addition, we use the preprocessed sparse point cloud in the dataset to test our model. For the DTU dataset, we transformed the point clouds of objects from their depth maps and images and we downsampled the point cloud by a factor of 0.5 to keep their point clouds distributing sparsely. In addition, we follow the MVSNeRF [2] to split the train and test groups of scans. The THuman2.0 dataset contains 500 high-quality human scans captured by a dense DLSR rig. For each scan, we render 36 views based on the provided 3D model and sparse sample points on the surface of the model. We keep the number of points for each scan at around 80,000 for their sparsity. We randomly select $75\%$ as the training set and the remaining $25\%$ as the test set. + +Baselines and Metrics. We compare the proposed method with the currently prevailing point cloud rendering approaches to demonstrate the effectiveness of our framework. The baselines include a pure graphics-based renderer, learning-based renderers NPBG [12] and $\mathrm{NPBG + + }$ [24], and the NeRF-based renderer Trivol [9]. As for the quality, we adopt PSNR, SSIM [36], and LPIPS [44]. These metrics overall determine the reconstruction accuracy and quality of these methods. + +# 4.2 Comparisons With Other Point Cloud Renderers. + +Fig. 2, 3, and 4 show the results of qualitative comparisons on the ScanNet, DTU, and THuman2.0 datasets respectively. Table. 1 reports all the metrics on the three datasets of the baselines and ours. The best result is bold, and the second-best result is underlined in the Table. + +![](images/3a34bd6b3fd9041c6d8e1ed248ad047bc46b3ab79bf4cf9b2369f3e0eddcddbf.jpg) + +![](images/cc43a5f46defb72c8955dc67f03d48a9b06b620baa4aae84a5c5e1540f9f458e.jpg) + +![](images/0e433f96003de660997196971492f2bef15837eb533c1ff6ac7ff23d93edb4ed.jpg) + +![](images/2c0712f0f2820c8912d7cf09bcffbd8a5371e20d0466ccdbe0b444cf4ef1553a.jpg) + +![](images/3aa2436bbe65eeb2d38718659cd0200d0a186fa5db137bfa181349de7cc5cb7f.jpg) + +![](images/ce1adff760ab46ca04475fe5c6e4dcc75c63810fc9735f4e422f138c07f48382.jpg) + +![](images/3c2239014b8bb4b7c81584ecc721b4d4a47e7f66f8cc6f11197651a6db91a835.jpg) + +![](images/c003333cb0a57cb91b64ce8ea17f5bdb7a23d41c5907b55cf4ac76404157cb91.jpg) + +![](images/74dd1159eda6cf5c82ba31c9cec47f604215945d21a803177df86e7777b17116.jpg) + +![](images/e378b7e77f50b3b7e5701fd3be0429b08665b7bf869da15122cc949dd283e621.jpg) + +![](images/dde24abb49cd11fba5415acd11a6ba4813a54220966a8cfffe6c583f51842201.jpg) + +![](images/74f1c471e86b1611e32089b78b6dfb08a62e40c8decb61a38de026fcb886c80c.jpg) + +![](images/53bd43002cbe255690c7593b37cea9cd70e356bf85f2abdd804c4f0d217dec58.jpg) + +![](images/fbef94c1c6513a492a1ff2f3fe031eac8108a705efb18462d24c51f6f624d273.jpg) + +![](images/0e9f06e8b26b7a5c5b584938217a0ef49043b4e14a4030a5ad2cbb04975fa454.jpg) + +![](images/2791c9e2fa5b3582b72051a0ae31264bfacf7665f3c649964d4345b5dd250e62.jpg) + +![](images/3bc4518628076c2a0a12009731b8c4b3a7b0db19829f3a17d80631ce28fbd2da.jpg) + +![](images/0af877d64b84cca3384c704d9a80c62652bf21cdaef957fe547b5c04dc443f50.jpg) + +![](images/a90cbd7254f39a29aae73ddff4a65cb7783ec8e98c99a7eea112fe921ecab657.jpg) + +![](images/b3e1e21d50c15562a0184d1168f3f67cf88d4b9ae3a13dc701dc2b1578f06a91.jpg) + +![](images/a6f52cfabf32d04165f985a997907205baf85afdc61136f6c7de71b115e05663.jpg) +Graphics + +![](images/8417ae31182c63fdf8e40f9141b4d7f334c0b66583ef52ae86168691ca5e5459.jpg) +NPBG++ + +![](images/72bbd56efc3a9d7feff2f75f0d3be39567f62bbef291cea7a9e513e7fa4577af.jpg) +Trivol + +![](images/c762a483576bf3e9b7f297c6fc4ed9c28f38fd78190ccc00942141bb8acd89ab.jpg) +Ours +Fig. 2: Qualitative comparisons between ours and other point cloud renderers on the ScanNet dataset. + +![](images/a3e98e2c4eafe5396aec5e65042dbef3b3e0a281c61c143ae3dcf14347e9cbca.jpg) +Ground Truth + +ScanNet. The problem of uneven distribution and missing points exists in the real scene-level point cloud. As shown in the first column in Fig. 2, traditional graphical rendering leads to a large number of holes in the rendering result. When the point cloud distribution is uneven, the point cloud of nearby objects is relatively sparse, and the 2D-CNN of $\mathrm{NPBG}++$ pays more attention to the features of distant objects, ignoring the 3D information in the point cloud, resulting that nearby objects are covered in the rendering result. Our method uses 3DGS for feature projection, which better integrates 3D context information. There are a large number of artifacts in the results of $\mathrm{NPBG}++$ and Trivol. Our multi-scale feature extraction network excels at capturing features across various levels of detail, effectively decoding these projected features to enhance the rendering of lighting effects and details significantly. + +DTU. As we stated in the Dataset description part, we maintain points distributed sparsely on the surface of the objects. Under this circumstance, the graphics-based transformation demonstrates more serious bleeding surfaces. Trivol always partially reconstructs the whole object because it depends on the fully implicit neural representation. This setting is more likely to be suited to class-specific point rendering, as they claimed in the original paper. In contrast, DTU includes many different classes that obstacle Trivol to learning a unified feature + +![](images/cd7e8e3c419b67b0a9e1c4a5e97afbcd2c38b6a0ff753cf76840af3a10b0d0b0.jpg) + +![](images/a5b0070e917b1518abd623da711ed569c7bd906498eff8a4ff350cffd37cae6f.jpg) + +![](images/9078a231cf16884ae347dabeb3421730963e94f45fa066e60dea7730cd2371ad.jpg) + +![](images/e496e1d75bbe501f37b082708972a92a1510f864ed5ed652968cfdddc8cbcb0b.jpg) + +![](images/534bbafc15cdef1376e5331cd5e573188d73b8868bff007e6f0f9ee7e149c22b.jpg) +Graphics + +![](images/fec8f76ef5e437d4a2e1d10d5cbc6d77d38b48501cc59de379a8a1f1da410e91.jpg) + +![](images/900cd61fe518779f43ae3c683195b156981eca68694aa8b0171f159da704c2a3.jpg) + +![](images/03a0b695b51f0a6522fc3190c07d4849d065a3716affb235fb5d1e142b88ee21.jpg) + +![](images/c678ab74a7fc9983411addf14e6825203587b972a1a7748c323af7cf23dc110f.jpg) + +![](images/7c58b6df8619f5572dc6e5ab98a387cffcebe6e0a702d2e9728cc22d866067ec.jpg) +NPBG++ + +![](images/f07622ea5d480b41fbf1394413cbf4a79235acc77591d52ca856c5e66c7fe582.jpg) + +![](images/1c2ef355942df92686b8e69ea9f8f863b06ca9d7b9f7cf0d609e08e1571d091a.jpg) + +![](images/0204fb6946aa852a289a4fe93103322ed0ceaac017f75bd049c97e8b7690f482.jpg) + +![](images/90d3ea82c194a288caf340dcb500f09a0417d43124a13065efb9fc97bf1d4565.jpg) + +![](images/68d9e99e02a84480988c3fc4737ae8d08f7c01509e465e769cebf933dfd7f13d.jpg) +Trivol + +![](images/43628b3babe020c2f0e47f123d5734b533bea8b2fb7380fc3f5f342f2c863e7b.jpg) + +![](images/f834ef6cf179062b2d31581f7e0f23341eafb07eb4cccc4bf31be7b406829d4c.jpg) + +![](images/4ea53ddf98210f2921906d3ff6092d8a05414746b69fa7994c64b4ea498fc7fc.jpg) + +![](images/0549f9a2cb704ddb9ec2f9eb638ea6d4e224fb61265ef9ce5f0aafa5a1d1aedb.jpg) + +![](images/3802eeae13785f41156e779fd4da8862b9ee104e376e928a073b3464d5031a43.jpg) +Ours +Fig. 3: Qualitative comparisons between ours and other point cloud rendering methods on the DTU dataset. + +![](images/652a412911889a2293185ae31179e01afdc677df9706e9a859728094f91176f4.jpg) + +![](images/b199285bf82babde402ff41797e7e2547fce0b6b80511b47771f5f718feec624.jpg) + +![](images/ddf8a049c33d52c67b1ce40c1eb424591d6dadc6d5924504b96fe166fac18340.jpg) + +![](images/99746c9dd055c8136156f052c2e9f07d0598659e05a73bf6824729b4e92a39e3.jpg) + +![](images/85d51e737b6dba932d5b080a49b3c3f32e2d75f82729291314bf02dc19963512.jpg) +Ground Truth + +transform underlying them. NPBG++ and our method deliver better results than Trivol because they leverage the original color and structure of point clouds without class-specific assumptions. Our approach yields the clearest predictions thanks to our more reasonable feature projection and decoding. + +THuman2.0 The Graphics-based method reconstructs the main structures of the human body while inevitably seeing densely packed holes. The other two baselines generate similar results that are much better than those of the traditional methods. The difference is that Trivol fails to faithfully recover some cloth colors, which is because it is a fully implicit method and thereby cannot restore colors that are never seen in its train set. $\mathrm{NPBG}++$ and ours directly utilize points' colors thus avoiding this domain generalization issue. Moreover, our method restores more details such as human facial expressions and clothes' folds. + +# 4.3 Ablation studies. + +Ablations of main components. In this subsection, we evaluate the effectiveness of all main constituents proposed in our whole framework. We conduct experiments with different control variables set and report their performance on + +![](images/64a62b1e3fe8dd5a348e8144b5004164caaaaff1c6f88ba316caf5e832938f02.jpg) +Fig. 4: Qualitative comparisons between ours and other point cloud renderers on the THuman2.0 dataset. + +the ScanNet dataset in Table 2. In the table, $w / o$ ProL refers to the version that removes all the progressive loss items, i.e. removes the first summation in Eq. 11 and Eq. 12. $w / o$ MSL denotes the version that removes all the multiscale loss items that are the second summation in Eq. 11 and 12. The recur1 illustrates that we discard the recurrent structure from the feature decoder and only utilize a single multi-input and multi-output UNet. Besides, the $w / o$ MSf describes a variant of the decoder which adopts a normal single scale UNet. In other words, this variant receives a single feature map as input and outputs a single prediction with full resolutions. Therefore, we only maintain the progressive losses and structures but discard all multiscale components for this version. Furthermore, we evaluate when the number of loops equals to 3, i.e. $l = 3$ in these equations. This is marked as recur3 in the Table. The results show that continuously increasing the number of iterations does not further improve the model's performance; instead, it may lead to a slight decline. Last, the $w / o$ Lmfr represents the version of removing the progressively multiscale frequency reconstruction loss (Eq. 12). + +Table 2: Quantitative Results of ablation studies under various variants on the ScanNet dataset. The meaning of each abbreviation can be found in the text. + +
w/o ProLw/o MSLw/o MSfw/o Lmfrw/o 2trrecur1recur3Full
PSNR↑19.1919.3718.2719.4215.9219.1919.7719.86
SSIM↑0.7250.7330.7040.7400.6960.7290.7600.758
LPIPS↓0.4980.5090.6500.4780.6890.5160.4780.452
+ +![](images/79ba08dca64e76c1dbb5a5bc7094bf0f5047d6a196c66ebd6923737026dd9952.jpg) +Fig. 5: Qualitative Comparisons of the renderings produced by the Gaussian splatting after the first stage and the feature splatting after the whole framework. + +It is observed that the full model delivers the most satisfactory rendering results across all metrics while the rest variants experience a decline to some extent. This indicates that both progressive and multi-scale features have a positive impact and contribute to the final promising performance of the model. Therefore, the entire design of this method is consistently characterized by progressive and multi-scale aspects. Moreover, the alignment of predictions and labels in the frequency domain progressively and in a multiscale manner makes prediction clearer and sharper. + +Analysis of Gaussian and Feature Splatting. The proposed approach relies on the two-stage training scheme because the 3D-consistent feature should be projected and integrated into certain 2D planes given camera viewpoints. The projection of features is dependent on the 3D Gaussian rasterization and further reliant on the Gaussian parameters associated with each point which are estimated by Eq. 6. The significance of training the Gaussian regression network in the first training stage lies in laying a solid foundation for feature splatting. Even though the 3D Gaussian Splatting could also produce plausible images, namely the $I_{gs}$ in Eq. 7, the results are still severely influenced by points that are either locally or globally sparse, thereby missing details and showing holes. It can be seen in Fig. 5 that these Gaussian splatting results are indeed better than the pure graphics-based method due to the representation capabilities of 3DGS, but they still need to be improved further. + +However, the results of feature splatting contain rich semantic information and local geometry cues, we consider them a good representation of restoring high-quality images and filling up blank areas. From Fig. 5 we can observe the capability of our method to generate and complete areas with voids and regions with sparse point clouds. We additionally evaluate an experiment in which we collaboratively train all modules in the framework rather than conducting the two-stage training and the results on the ScanNet dataset and we add this ad + +Table 3: Ablation studies of the effect of different numbers of points on performance. + +
Train Points50k50k50k120k120k120k
Test Points50k80k120k50k80k120k
Trivol24.0525.4326.4824.7125.9626.86
Ours31.2932.8635.9531.6034.1739.18
+ +ditional result into Table. 2. $w / o$ 2tr means that we abandon the first training stage and jointly train all modules together. In this setting, we observe that the model completely collapses and sees a very fast degradation in performance. + +Analysis of The Number of Points. We additionally analyze and evaluate how different training and test numbers of points affect the performance. On the THuman2.0 dataset, we set different training and test pairs to show the performance of our method and Trivol respectively, which includes 50,000 and 120,000 points used for training and 50,000, 80,000, and 120,000 points used for testing. The results are listed in Table 3. It is noted that Trivol is essentially unaffected by the number of points where our method sees significant fluctuations with changes in the number of points. However, our method exhibits very high performance when point clouds are abundant while the Trivol method has a relatively low upper limit, which is due to its fully implicit representation and lack of utilization of point cloud prior. More discussion can be seen in the Appendix. + +# 5 Discussion + +Conclusion. This work proposes a novel point cloud rendering method called PFGS, that can render high-fidelity and photorealistic images from colored point clouds with arbitrary camera viewpoints. To address the shortcomings of current learning-based point cloud renderers such as slow, blurry, and inconsistent, we introduce the 3D Gaussian Splatting to point cloud rendering by several well-designed modules. We supervise the regressor to estimate Gaussian parameters in a point-wise way, and then we utilize the estimated Gaussian parameters to rasterize point-wise neural features that are extracted by a multiscale feature decoder. Furthermore, the projected feature volume would be gradually decoded to the final prediction by our progressive and multiscale decoder. The whole architecture is supervised by our finely-designed progressive and multiscale reconstruction losses. We carry out experiments on three datasets with different levels of objects including indoors, objects, and humans and the results show our method outperforms all baselines. The ablation studies also illustrate the significance of all main components in the approach. + +Limitation. Even though the proposed method can produce photorealistic images from sparse and uneven point clouds, it still yields unreal details when the given point clouds are not precise enough. Future works can incorporate a large generative refiner trained on a large dataset to slightly refine the original point cloud to make points closer to the actual surface of objects. + +# References + +1. Cao, A., Johnson, J.: Hexplane: A fast representation for dynamic scenes. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 130-141 (2023) +2. Chen, A., Xu, Z., Zhao, F., Zhang, X., Xiang, F., Yu, J., Su, H.: Mvsnerf: Fast generalizable radiance field reconstruction from multi-view stereo. In: Proceedings of the IEEE/CVF International Conference on Computer Vision. pp. 14124-14133 (2021) +3. Choy, C., Gwak, J., Savarese, S.: 4d spatio-temporal convnets: Minkowski convolutional neural networks. In: Proceedings of the IEEE/CVF conference on computer vision and pattern recognition. pp. 3075-3084 (2019) +4. Dai, A., Chang, A.X., Savva, M., Halber, M., Funkhouser, T., Nießner, M.: Scannet: Richly-annotated 3d reconstructions of indoor scenes. In: Proceedings of the IEEE conference on computer vision and pattern recognition. pp. 5828-5839 (2017) +5. Dai, P., Zhang, Y., Li, Z., Liu, S., Zeng, B.: Neural point cloud rendering via multi-plane projection. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 7830-7839 (2020) +6. Fridovich-Keil, S., Yu, A., Tancik, M., Chen, Q., Recht, B., Kanazawa, A.: Plenoxels: Radiance fields without neural networks. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 5501-5510 (2022) +7. Ge, W., Hu, T., Zhao, H., Liu, S., Chen, Y.C.: Ref-neus: Ambiguity-reduced neural implicit surface learning for multi-view reconstruction with reflection. arXiv preprint arXiv:2303.10840 (2023) +8. Graham, B., Van der Maaten, L.: Submanifold sparse convolutional networks. arXiv preprint arXiv:1706.01307 (2017) +9. Hu, T., Xu, X., Chu, R., Jia, J.: Trivol: Point cloud rendering via triple volumes. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 20732-20741 (2023) +10. Hu, T., Xu, X., Liu, S., Jia, J.: Point2pix: Photo-realistic point cloud rendering via neural radiance fields. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 8349-8358 (2023) +1. Huang, C., Mees, O., Zeng, A., Burgard, W.: Visual language maps for robot navigation. In: 2023 IEEE International Conference on Robotics and Automation (ICRA). pp. 10608-10615. IEEE (2023) +2. Jena, S., Multon, F., Boukhayma, A.: Neural mesh-based graphics. In: European Conference on Computer Vision. pp. 739-757. Springer (2022) +3. Jensen, R., Dahl, A., Vogiatzis, G., Tola, E., Aanaes, H.: Large scale multi-view stereopsis evaluation. In: Proceedings of the IEEE conference on computer vision and pattern recognition. pp. 406-413 (2014) +4. Jiang, Y., Yu, C., Xie, T., Li, X., Feng, Y., Wang, H., Li, M., Lau, H., Gao, F., Yang, Y., et al.: Vr-gs: A physical dynamics-aware interactive gaussian splattering system in virtual reality. arXiv preprint arXiv:2401.16663 (2024) +5. Kerbl, B., Kopanas, G., Leimkuhler, T., Drettakis, G.: 3d gaussian splatting for real-time radiance field rendering. ACM Transactions on Graphics 42(4) (2023) +6. Lassner, C., Zollhofer, M.: Pulsar: Efficient sphere-based neural rendering. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 1440-1449 (2021) +7. Liu, Z., Tang, H., Lin, Y., Han, S.: Point-voxel cnn for efficient 3d deep learning. Advances in Neural Information Processing Systems 32 (2019) + +18. Lombardi, S., Simon, T., Saragih, J., Schwartz, G., Lehrmann, A., Sheikh, Y.: Neural volumes: Learning dynamic renderable volumes from images. arXiv preprint arXiv:1906.07751 (2019) +19. Metzer, G., Richardson, E., Patashnik, O., Giryes, R., Cohen-Or, D.: Latent-nerf for shape-guided generation of 3d shapes and textures. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 12663-12673 (2023) +20. Mildenhall, B., Srinivasan, P.P., Tancik, M., Barron, J.T., Ramamoorthi, R., Ng, R.: Nerf: Representing scenes as neural radiance fields for view synthesis. Communications of the ACM 65(1), 99-106 (2021) +21. Müller, T., Evans, A., Schied, C., Keller, A.: Instant neural graphics primitives with a multiresolution hash encoding. ACM Transactions on Graphics (ToG) 41(4), 1-15 (2022) +22. Qi, C.R., Su, H., Nießner, M., Dai, A., Yan, M., Guibas, L.J.: Volumetric and multi-view cnns for object classification on 3d data. In: Proceedings of the IEEE conference on computer vision and pattern recognition. pp. 5648-5656 (2016) +23. Qi, C.R., Yi, L., Su, H., Guibas, L.J.: Pointnet++: Deep hierarchical feature learning on point sets in a metric space. Advances in neural information processing systems 30 (2017) +24. Rakhimov, R., Ardelean, A.T., Lempitsky, V., Burnaev, E.: Npgb++: Accelerating neural point-based graphics. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 15969-15979 (2022) +25. Ravi, N., Reizenstein, J., Novotny, D., Gordon, T., Lo, W.Y., Johnson, J., Gkioxari, G.: Accelerating 3d deep learning with pytorch3d. arXiv preprint arXiv:2007.08501 (2020) +26. Sitzmann, V., Thies, J., Heide, F., Nießner, M., Wetzstein, G., Zollhofer, M.: Deepvoxels: Learning persistent 3d feature embeddings. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 2437-2446 (2019) +27. Sun, W., Trulls, E., Tseng, Y.C., Sambandam, S., Sharma, G., Tagliasacchi, A., Yi, K.M.: Pointner++: A multi-scale, point-based neural radiance field. arXiv preprint arXiv:2312.02362 (2023) +28. Wang, C., Chai, M., He, M., Chen, D., Liao, J.: Clip-nerf: Text-and-image driven manipulation of neural radiance fields. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 3835-3844 (2022) +29. Wang, J., He, J., Zhang, Z., Xu, R.: Physical priors augmented event-based 3d reconstruction. arXiv preprint arXiv:2401.17121 (2024) +30. Wang, J., Zhang, Z., Xu, R.: Learning to generate and manipulate 3d radiance field by a hierarchical diffusion framework with clip latent. In: Computer Graphics Forum. vol. 42, p. e14930. Wiley Online Library (2023) +31. Wang, J., Zhang, Z., Xu, R.: Learning robust generalizable radiance field with visibility and feature augmented point representation. arXiv preprint arXiv:2401.14354 (2024) +32. Wang, J., Sun, B., Lu, Y.: Mvpnet: Multi-view point regression networks for 3d object reconstruction from a single image. In: Proceedings of the AAAI Conference on Artificial Intelligence. vol. 33, pp. 8949-8956 (2019) +33. Wang, N., Zhang, Y., Li, Z., Fu, Y., Liu, W., Jiang, Y.G.: Pixel2mesh: Generating 3d mesh models from single rgb images. In: Proceedings of the European conference on computer vision (ECCV). pp. 52-67 (2018) + +34. Wang, P., Liu, Y., Chen, Z., Liu, L., Liu, Z., Komura, T., Theobalt, C., Wang, W.: F2-nerf: Fast neural radiance field training with free camera trajectories. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 4150-4159 (2023) +35. Wang, S., Liu, M.: Point cloud compression with range image-based entropy model for autonomous driving. In: European Conference on Computer Vision. pp. 323-340. Springer (2022) +36. Wang, Z., Bovik, A.C., Sheikh, H.R., Simoncelli, E.P.: Image quality assessment: from error visibility to structural similarity. IEEE transactions on image processing 13(4), 600-612 (2004) +37. Wu, G., Yi, T., Fang, J., Xie, L., Zhang, X., Wei, W., Liu, W., Tian, Q., Wang, X.: 4d gaussian splatting for real-time dynamic scene rendering. arXiv preprint arXiv:2310.08528 (2023) +38. Xiang, F., Xu, Z., Hasan, M., Hold-Geoffroy, Y., Sunkavalli, K., Su, H.: Neutex: Neural texture mapping for volumetric neural rendering. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 7119-7128 (2021) +39. Xu, Q., Xu, Z., Philip, J., Bi, S., Shu, Z., Sunkavalli, K., Neumann, U.: Point-nerf: Point-based neural radiance fields. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 5438-5448 (2022) +40. Yan, Z., Li, C., Lee, G.H.: Nerf-ds: Neural radiance fields for dynamic specular objects. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 8285-8295 (2023) +41. Yang, B., Bao, C., Zeng, J., Bao, H., Zhang, Y., Cui, Z., Zhang, G.: Neumesh: Learning disentangled neural mesh-based implicit field for geometry and texture editing. In: European Conference on Computer Vision. pp. 597-614. Springer (2022) +42. Yu, T., Zheng, Z., Guo, K., Liu, P., Dai, Q., Liu, Y.: Function4d: Real-time human volumetric capture from very sparse consumer rgbd sensors. In: IEEE Conference on Computer Vision and Pattern Recognition (CVPR2021) (June 2021) +43. Zhang, C., Guo, R., Zeng, W., Xiong, Y., Dai, B., Hu, R., Ren, M., Urtasun, R.: Rethinking closed-loop training for autonomous driving. In: European Conference on Computer Vision. pp. 264-282. Springer (2022) +44. Zhang, R., Isola, P., Efros, A.A., Shechtman, E., Wang, O.: The unreasonable effectiveness of deep features as a perceptual metric. In: Proceedings of the IEEE conference on computer vision and pattern recognition. pp. 586-595 (2018) +45. Zhou, Q.Y., Park, J., Koltun, V.: Open3d: A modern library for 3d data processing. arXiv preprint arXiv:1801.09847 (2018) +46. Zhou, S., Chang, H., Jiang, S., Fan, Z., Zhu, Z., Xu, D., Chari, P., You, S., Wang, Z., Kadambi, A.: Feature 3dgs: Supercharging 3d gaussian splatting to enable distilled feature fields. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 21676-21685 (2024) \ No newline at end of file diff --git a/2024/PFGS_ High Fidelity Point Cloud Rendering via Feature Splatting/images.zip b/2024/PFGS_ High Fidelity Point Cloud Rendering via Feature Splatting/images.zip new file mode 100644 index 0000000000000000000000000000000000000000..71cbd9555ddd67a06ed647641e88f72e9cea0198 --- /dev/null +++ b/2024/PFGS_ High Fidelity Point Cloud Rendering via Feature Splatting/images.zip @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:755a23aab68599b371173d33adaec183c23783df81280feee646e0bc2e2cde4d +size 666618 diff --git a/2024/PFGS_ High Fidelity Point Cloud Rendering via Feature Splatting/layout.json b/2024/PFGS_ High Fidelity Point Cloud Rendering via Feature Splatting/layout.json new file mode 100644 index 0000000000000000000000000000000000000000..17ce9e038993d4c6342fb5ce7714476e56d0ddb4 --- /dev/null +++ b/2024/PFGS_ High Fidelity Point Cloud Rendering via Feature Splatting/layout.json @@ -0,0 +1,11302 @@ +{ + "pdf_info": [ + { + "para_blocks": [ + { + "bbox": [ + 138, + 111, + 477, + 148 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 111, + 477, + 148 + ], + "spans": [ + { + "bbox": [ + 138, + 111, + 477, + 148 + ], + "type": "text", + "content": "PFGS: High Fidelity Point Cloud Rendering via Feature Splatting" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 154, + 167, + 459, + 181 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 154, + 167, + 459, + 181 + ], + "spans": [ + { + "bbox": [ + 154, + 167, + 459, + 181 + ], + "type": "text", + "content": "Jiaxu Wang" + }, + { + "bbox": [ + 154, + 167, + 459, + 181 + ], + "type": "inline_equation", + "content": "^{1\\dagger}" + }, + { + "bbox": [ + 154, + 167, + 459, + 181 + ], + "type": "text", + "content": ", Ziyi Zhang" + }, + { + "bbox": [ + 154, + 167, + 459, + 181 + ], + "type": "inline_equation", + "content": "^{1\\dagger}" + }, + { + "bbox": [ + 154, + 167, + 459, + 181 + ], + "type": "text", + "content": ", Junhao He" + }, + { + "bbox": [ + 154, + 167, + 459, + 181 + ], + "type": "inline_equation", + "content": "^{1}" + }, + { + "bbox": [ + 154, + 167, + 459, + 181 + ], + "type": "text", + "content": ", and Renjing Xu" + }, + { + "bbox": [ + 154, + 167, + 459, + 181 + ], + "type": "inline_equation", + "content": "^{1\\boxplus}" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 190, + 189, + 423, + 224 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 190, + 189, + 423, + 224 + ], + "spans": [ + { + "bbox": [ + 190, + 189, + 423, + 224 + ], + "type": "inline_equation", + "content": "^{1}" + }, + { + "bbox": [ + 190, + 189, + 423, + 224 + ], + "type": "text", + "content": " Hong Kong University of Science and Technology (GZ) jwang457@connect.hkust-gz.edu.cn {ziyizhang, junhaohe, renjingxu}@hkust-gz.edu.cn" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 159, + 245, + 455, + 410 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 159, + 245, + 455, + 410 + ], + "spans": [ + { + "bbox": [ + 159, + 245, + 455, + 410 + ], + "type": "text", + "content": "Abstract. Rendering high-fidelity images from sparse point clouds is still challenging. Existing learning-based approaches suffer from either hole artifacts, missing details, or expensive computations. In this paper, we propose a novel framework to render high-quality images from sparse points. This method first attempts to bridge the 3D Gaussian Splatting and point cloud rendering, which includes several cascaded modules. We first use a regressor to estimate Gaussian properties in a point-wise manner, the estimated properties are used to rasterize neural feature descriptors into 2D planes which are extracted from a multiscale extractor. The projected feature volume is gradually decoded toward the final prediction via a multiscale and progressive decoder. The whole pipeline experiences a two-stage training and is driven by our well-designed progressive and multiscale reconstruction loss. Experiments on different benchmarks show the superiority of our method in terms of rendering qualities and the necessities of our main components." + }, + { + "bbox": [ + 159, + 245, + 455, + 410 + ], + "type": "inline_equation", + "content": "^{1}" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 160, + 419, + 404, + 431 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 160, + 419, + 404, + 431 + ], + "spans": [ + { + "bbox": [ + 160, + 419, + 404, + 431 + ], + "type": "text", + "content": "Keywords: Point cloud rendering " + }, + { + "bbox": [ + 160, + 419, + 404, + 431 + ], + "type": "inline_equation", + "content": "\\cdot" + }, + { + "bbox": [ + 160, + 419, + 404, + 431 + ], + "type": "text", + "content": " 3D Gaussian Splatting" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 132, + 448, + 230, + 460 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 448, + 230, + 460 + ], + "spans": [ + { + "bbox": [ + 132, + 448, + 230, + 460 + ], + "type": "text", + "content": "1 Introduction" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 130, + 470, + 482, + 589 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 470, + 482, + 589 + ], + "spans": [ + { + "bbox": [ + 130, + 470, + 482, + 589 + ], + "type": "text", + "content": "Synthesizing photorealistic images from given colored point clouds and arbitrary camera views can be extensively applied to various fields including virtual/augmented reality [14], robotic navigation [11], automatic driving [35, 43], etc, which is academically named point cloud rendering. Traditional methods to render views from point clouds are based on graphics transformations. They directly project 3D point clouds into 2D planes by using the camera parameters and z-buffer rasterization. However, the sparse distribution of points leads to bleeding surfaces and hole artifacts. Even though the graphics renderer can treat points as graphical primitives and diminish these issues by color blending, rough and simple geometries of primitives entail blurred images and missing details." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 130, + 590, + 482, + 638 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 590, + 482, + 638 + ], + "spans": [ + { + "bbox": [ + 130, + 590, + 482, + 638 + ], + "type": "text", + "content": "In recent years, learning-based methods [9, 10, 24] have become prevailing to address the drawbacks of conventional graphics-based approaches. These methods project points and corresponding feature descriptors to 2D planes, obtaining sparse images. Then they use UNet-like networks to complete and restore" + } + ] + } + ], + "index": 7 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 133, + 643, + 345, + 654 + ], + "type": "page_footnote", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 643, + 345, + 654 + ], + "spans": [ + { + "bbox": [ + 133, + 643, + 345, + 654 + ], + "type": "inline_equation", + "content": "\\star \\dagger" + }, + { + "bbox": [ + 133, + 643, + 345, + 654 + ], + "type": "text", + "content": " Co-first authors; " + }, + { + "bbox": [ + 133, + 643, + 345, + 654 + ], + "type": "inline_equation", + "content": "\\boxtimes" + }, + { + "bbox": [ + 133, + 643, + 345, + 654 + ], + "type": "text", + "content": " Corresponding authors" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 135, + 654, + 352, + 665 + ], + "type": "page_footnote", + "angle": 0, + "lines": [ + { + "bbox": [ + 135, + 654, + 352, + 665 + ], + "spans": [ + { + "bbox": [ + 135, + 654, + 352, + 665 + ], + "type": "text", + "content": "1 Project page: https://github.com/Mercerai/PPGS" + } + ] + } + ], + "index": 9 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 0 + }, + { + "para_blocks": [ + { + "bbox": [ + 130, + 116, + 482, + 248 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 116, + 482, + 248 + ], + "spans": [ + { + "bbox": [ + 130, + 116, + 482, + 248 + ], + "type": "text", + "content": "the low-quality images [12, 24]. Nevertheless, these methods rely on 2D image restoration, which cannot guarantee 3D consistency and can result in the distinctive appearance of the same object from nearby different viewpoints. Moreover, their feature descriptors are associated with points that do not possess actual shapes and volumes in real space, thereby showing poor descriptions of local area structures. To obtain 3D-aware and consistent features, some approaches [8, 16] transform points to 3D volumes by trilinear interpolation and use 3D convolution to extract the features. However, they lead to a significant computational burden and sacrifice the efficient representation characteristics of point clouds. For example, voxels uniformly represent scenes, whereas points are densely distributed in complex areas and sparsely distributed or even absent in empty spaces." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 130, + 250, + 482, + 382 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 250, + 482, + 382 + ], + "spans": [ + { + "bbox": [ + 130, + 250, + 482, + 382 + ], + "type": "text", + "content": "To obtain efficient and consistent 3D representation, the Neural Radiance Field (NeRF) [20] has gradually been incorporated into point cloud renderers [9, 10]. These methods either consider points as anchors of NeRF or transform points into NeRF-like representations, then use the physical-based volume rendering to obtain 3D-aware feature maps which are decoded into corresponding images. With the assistance of NeRF, they achieve high-quality consistencies across nearby views and plausible images. However, discrete NeRF requires heavy memory consumption resulting from storing high-dimensional feature vectors. In addition, NeRF typically requires lots of sampling points and multiple accesses to the neural network, which significantly increases the rendering time and makes it challenging to meet real-time requirements." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 130, + 384, + 482, + 492 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 384, + 482, + 492 + ], + "spans": [ + { + "bbox": [ + 130, + 384, + 482, + 492 + ], + "type": "text", + "content": "More recently, 3D Gaussian Splatting (3DGS) [15] introduces a new representation that formulates points as 3D Gaussians with learnable parameters including 3D position, color, opacity, and anisotropic covariance. 3DGS achieves more photorealistic and faster rendering than NeRF by applying Gaussian rasterization and " + }, + { + "bbox": [ + 130, + 384, + 482, + 492 + ], + "type": "inline_equation", + "content": "\\alpha" + }, + { + "bbox": [ + 130, + 384, + 482, + 492 + ], + "type": "text", + "content": "-blending. Conventional 3D Gaussian learns the primitive parameters associated with an initialized point cloud from multi-view images by backpropagation, it relies on per-subject parameter optimizations for several tens minutes. It is therefore impractical to be applied to the case in which only point cloud is available due to the need for multiview images and backpropagations." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 130, + 495, + 482, + 567 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 495, + 482, + 567 + ], + "spans": [ + { + "bbox": [ + 130, + 495, + 482, + 567 + ], + "type": "text", + "content": "In this work, we fill the gap between point cloud render and the 3DGS technique by introducing a novel multiscale, feature-augmented, 3DGS-based point renderer. Compared with previous learning-based point renderers, our method yields higher-quality and 3D consistent results. Compared with conventional 3DGS, our method does not necessarily require multiview images or fine-tuning procedures and can be directly employed to point modality." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 130, + 570, + 482, + 665 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 570, + 482, + 665 + ], + "spans": [ + { + "bbox": [ + 130, + 570, + 482, + 665 + ], + "type": "text", + "content": "We leverage an efficient multiscale feature extractor in the proposed framework to extract features from point cloud input. This extractor considers multiscale local geometries to extract discriminative feature descriptors. Then a regressor is used for estimating 3D Gaussian parameters in a point-wise manner. Next, we not only render images by these estimated Gaussian attributes in a multiscale way but also render the multiscale feature maps by the same Gaussian properties. Finally, the rendered multiscale images and feature maps are fused by a multiscale and recurrent UNet-like network to gradually transform" + } + ] + } + ], + "index": 6 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "spans": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "type": "text", + "content": "2" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 166, + 91, + 225, + 102 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 166, + 91, + 225, + 102 + ], + "spans": [ + { + "bbox": [ + 166, + 91, + 225, + 102 + ], + "type": "text", + "content": "J. Wang et al." + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 1 + }, + { + "para_blocks": [ + { + "bbox": [ + 130, + 116, + 479, + 139 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 116, + 479, + 139 + ], + "spans": [ + { + "bbox": [ + 130, + 116, + 479, + 139 + ], + "type": "text", + "content": "into the final results. To the best knowledge of us, this is the first point cloud renderer combined with 3DGS." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 146, + 140, + 412, + 152 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 146, + 140, + 412, + 152 + ], + "spans": [ + { + "bbox": [ + 146, + 140, + 412, + 152 + ], + "type": "text", + "content": "Our main contributions can be summarized in the following:" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 136, + 159, + 480, + 303 + ], + "type": "list", + "angle": 0, + "index": 7, + "blocks": [ + { + "bbox": [ + 136, + 159, + 479, + 206 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 136, + 159, + 479, + 206 + ], + "spans": [ + { + "bbox": [ + 136, + 159, + 479, + 206 + ], + "type": "text", + "content": "1. We proposed a novel lightweight approach called PFGS (Point Feature Gaussian Splitting) that bridges point cloud rendering and 3DGS. It produces higher-fidelity images from sparse point clouds than currently prevailing approaches." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 136, + 207, + 480, + 266 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 136, + 207, + 480, + 266 + ], + "spans": [ + { + "bbox": [ + 136, + 207, + 480, + 266 + ], + "type": "text", + "content": "2. We proposed the point-based multiscale and feature-augmented 3D Gaussian framework, which consists of three main components including a multiscale feature extractor, a Gaussian regressor, and a multiscale and recurrent decoder, accompanied by the well-designed multiscale and progressive reconstruction losses." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 136, + 267, + 479, + 303 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 136, + 267, + 479, + 303 + ], + "spans": [ + { + "bbox": [ + 136, + 267, + 479, + 303 + ], + "type": "text", + "content": "3. Extensive experiments and ablation studies on three datasets show the effectiveness and advantages of the proposed method over current works in terms of reconstruction quality." + } + ] + } + ], + "index": 6 + } + ], + "sub_type": "text" + }, + { + "bbox": [ + 132, + 321, + 236, + 334 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 321, + 236, + 334 + ], + "spans": [ + { + "bbox": [ + 132, + 321, + 236, + 334 + ], + "type": "text", + "content": "2 Related Work" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 132, + 346, + 261, + 357 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 346, + 261, + 357 + ], + "spans": [ + { + "bbox": [ + 132, + 346, + 261, + 357 + ], + "type": "text", + "content": "2.1 Point Cloud Render" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 130, + 366, + 480, + 426 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 366, + 480, + 426 + ], + "spans": [ + { + "bbox": [ + 130, + 366, + 480, + 426 + ], + "type": "text", + "content": "Conventional point cloud rendering is implemented via graphics-based transformation [25,45] which simulates camera imaging processes to project points onto the 2D screens. Even though the rendering pipeline is fast and general for all scenes, it is heavily dependent on the density of point distributions. These methods produce empty holes and vacant areas when the distributions are sparse." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 130, + 426, + 481, + 665 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 426, + 481, + 665 + ], + "spans": [ + { + "bbox": [ + 130, + 426, + 481, + 665 + ], + "type": "text", + "content": "Learning-based point renderers [3, 9, 10, 12, 24] to some extent handle the issues caused by traditional methods, which fill the holes by leveraging multiscale or multiview feature fusions. ME [3] implements a sparse Convolutional Neural Network to extract features from existing points and then calculates the features of arbitrary 3D points through ball querying. NPBG [12] enhances each point with a feature descriptor to encapsulate local geometry and appearance. Moreover, " + }, + { + "bbox": [ + 130, + 426, + 481, + 665 + ], + "type": "inline_equation", + "content": "\\mathrm{NPBG}++" + }, + { + "bbox": [ + 130, + 426, + 481, + 665 + ], + "type": "text", + "content": " [24] introduces feature aggregation from multiview based on the NPBG to complement missing parts. NPCR [5] employs a 3DCNN to generate 3D volumes from point clouds and creates multiple depth layers to synthesize images. More recently, Point2pix [10] and Trivol [9] have incorporated NeRF into the point cloud rendering pipeline. They transform raw point clouds into NeRF-related representations and utilize volume rendering to access the representations to produce images. However, the above methods still cannot fill large vacant holes and only produce blurry rendering results, especially in local areas with fewer points. Even if incorporating NeRF representation is effective in generating 3D consistent feature encoding, they do not fully utilize the geometric information contained within the point clouds. In this work, we combine point rendering with a novel 3D representation, i.e. 3DGS, and implement multiscale and progressive feature decoding to obtain fine-grained and 3D-consistent predictions." + } + ] + } + ], + "index": 11 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 421, + 91, + 447, + 100 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 421, + 91, + 447, + 100 + ], + "spans": [ + { + "bbox": [ + 421, + 91, + 447, + 100 + ], + "type": "text", + "content": "PFGS" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 474, + 92, + 481, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 474, + 92, + 481, + 100 + ], + "spans": [ + { + "bbox": [ + 474, + 92, + 481, + 100 + ], + "type": "text", + "content": "3" + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 2 + }, + { + "para_blocks": [ + { + "bbox": [ + 132, + 116, + 307, + 128 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 116, + 307, + 128 + ], + "spans": [ + { + "bbox": [ + 132, + 116, + 307, + 128 + ], + "type": "text", + "content": "2.2 Neural-based View Synthesis" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 130, + 133, + 482, + 324 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 133, + 482, + 324 + ], + "spans": [ + { + "bbox": [ + 130, + 133, + 482, + 324 + ], + "type": "text", + "content": "Unlike traditional explicit representations [22, 23, 32, 33], implicit representations that encode scenes with the assistance of neural networks have been developed in recent years [18, 26, 38]. Among them, the Neural Radiance Field (NeRF) [20] attracts the most attention in both industries and research. NeRF is a continuous method that implicitly represents a 3D scene employing an MLP and synthesizing images of novel view by volume rendering. Recent studies have employed NeRF for various purposes, including dynamic reconstruction [1, 37, 40], physical reasoning [29], generative models [19, 28, 30], and fast rendering [6, 21, 34]. To achieve various goals, some studies attempted to combine the implicit NeRF with explicit primitives. For example, NeuMesh [41] regards mesh as a neural scaffold to store NeRF and achieve neural editing. PointNeRF [39] is the first to incorporate point clouds into NeRF reconstruction to obtain better rendering qualities and faster speed. Ref-NeuS [7] implements generalizable neural surface reconstruction. GPF [31] explicitly models the visibility by using the geometries in the point cloud. PointNeRF++ [27] solves point cloud sparsity by aggregating multiscale point clouds on sparse voxel grids at different resolutions." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 130, + 324, + 482, + 397 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 324, + 482, + 397 + ], + "spans": [ + { + "bbox": [ + 130, + 324, + 482, + 397 + ], + "type": "text", + "content": "It can be seen that many approaches we list above incorporate points into NeRF. However, NeRF-like volume rendering does not fully explore the rendering capability of the point cloud. In contrast to NeRF, 3DGS [15] proposed the 3D Gaussian Splatting to realize a remarkable differentiable rendering speed and yield high-quality 2D images. In this work, we first bridge point cloud rendering and 3DGS to generate higher-fidelity images." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 132, + 411, + 233, + 426 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 411, + 233, + 426 + ], + "spans": [ + { + "bbox": [ + 132, + 411, + 233, + 426 + ], + "type": "text", + "content": "3 Methodology" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 130, + 434, + 482, + 567 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 434, + 482, + 567 + ], + "spans": [ + { + "bbox": [ + 130, + 434, + 482, + 567 + ], + "type": "text", + "content": "Given a colored point cloud " + }, + { + "bbox": [ + 130, + 434, + 482, + 567 + ], + "type": "inline_equation", + "content": "\\mathbf{P} = \\{\\mathbf{p}_k, \\mathbf{c}_k\\}" + }, + { + "bbox": [ + 130, + 434, + 482, + 567 + ], + "type": "text", + "content": ", we aim to synthesize photorealistic images from any viewpoints defined by the camera intrinsic " + }, + { + "bbox": [ + 130, + 434, + 482, + 567 + ], + "type": "inline_equation", + "content": "K" + }, + { + "bbox": [ + 130, + 434, + 482, + 567 + ], + "type": "text", + "content": " and poses " + }, + { + "bbox": [ + 130, + 434, + 482, + 567 + ], + "type": "inline_equation", + "content": "P" + }, + { + "bbox": [ + 130, + 434, + 482, + 567 + ], + "type": "text", + "content": " are provided. The " + }, + { + "bbox": [ + 130, + 434, + 482, + 567 + ], + "type": "inline_equation", + "content": "\\mathbf{p}_k \\in R^3" + }, + { + "bbox": [ + 130, + 434, + 482, + 567 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 130, + 434, + 482, + 567 + ], + "type": "inline_equation", + "content": "\\mathbf{c}_k \\in R^3" + }, + { + "bbox": [ + 130, + 434, + 482, + 567 + ], + "type": "text", + "content": " are the point coordinates and colors respectively. The rendering process can be illustrated as " + }, + { + "bbox": [ + 130, + 434, + 482, + 567 + ], + "type": "inline_equation", + "content": "I_v = Re(\\mathbf{P}|K_v, P_v)" + }, + { + "bbox": [ + 130, + 434, + 482, + 567 + ], + "type": "text", + "content": " where " + }, + { + "bbox": [ + 130, + 434, + 482, + 567 + ], + "type": "inline_equation", + "content": "Re" + }, + { + "bbox": [ + 130, + 434, + 482, + 567 + ], + "type": "text", + "content": " is the rendering function that can be implemented by either graphics-based or learning-based methods as we introduced above. In this work, it is implemented by our proposed multiscale feature-based 3DGS rendering pipeline that contains a multiscale feature extractor, a Gaussian regressor, a feature-based Gaussian rendering module, and a multiscale recurrent decoder. All modules are implemented with small networks. In the following of this section, we sequentially introduce the aforementioned components." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 132, + 580, + 228, + 592 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 580, + 228, + 592 + ], + "spans": [ + { + "bbox": [ + 132, + 580, + 228, + 592 + ], + "type": "text", + "content": "3.1 Preliminaries" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 130, + 597, + 481, + 646 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 597, + 481, + 646 + ], + "spans": [ + { + "bbox": [ + 130, + 597, + 481, + 646 + ], + "type": "text", + "content": "3DGS is an explicit representation method for 3D scenes, which parameterizes the scene as a series of 3D Gaussian primitives. A 3D Gaussian is defined by a full 3D covariance matrix " + }, + { + "bbox": [ + 130, + 597, + 481, + 646 + ], + "type": "inline_equation", + "content": "\\Sigma" + }, + { + "bbox": [ + 130, + 597, + 481, + 646 + ], + "type": "text", + "content": ", its center point " + }, + { + "bbox": [ + 130, + 597, + 481, + 646 + ], + "type": "inline_equation", + "content": "x" + }, + { + "bbox": [ + 130, + 597, + 481, + 646 + ], + "type": "text", + "content": ", and the spherical harmonic " + }, + { + "bbox": [ + 130, + 597, + 481, + 646 + ], + "type": "inline_equation", + "content": "(SH)" + }, + { + "bbox": [ + 130, + 597, + 481, + 646 + ], + "type": "text", + "content": ". The mean value of the Gaussian is defined as:" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 260, + 651, + 481, + 666 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 260, + 651, + 481, + 666 + ], + "spans": [ + { + "bbox": [ + 260, + 651, + 481, + 666 + ], + "type": "interline_equation", + "content": "G (x) = e ^ {- \\frac {1}{2} (x) ^ {T} \\Sigma^ {- 1} (x)} \\tag {1}", + "image_path": "1bc9e704dd93083729c68aed4dd8bac513900ffa55506c9b6077463ee9c0d77d.jpg" + } + ] + } + ], + "index": 9 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "spans": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "type": "text", + "content": "4" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 166, + 91, + 225, + 102 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 166, + 91, + 225, + 102 + ], + "spans": [ + { + "bbox": [ + 166, + 91, + 225, + 102 + ], + "type": "text", + "content": "J. Wang et al." + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 3 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 133, + 114, + 482, + 239 + ], + "blocks": [ + { + "bbox": [ + 133, + 114, + 482, + 239 + ], + "lines": [ + { + "bbox": [ + 133, + 114, + 482, + 239 + ], + "spans": [ + { + "bbox": [ + 133, + 114, + 482, + 239 + ], + "type": "image", + "image_path": "50f7342bfdb9038ce4d05ff8d6c3ccdeebee297f41a9c9ee71b730e700512a80.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_body" + } + ], + "index": 2 + }, + { + "type": "image", + "bbox": [ + 140, + 239, + 480, + 367 + ], + "blocks": [ + { + "bbox": [ + 140, + 239, + 480, + 367 + ], + "lines": [ + { + "bbox": [ + 140, + 239, + 480, + 367 + ], + "spans": [ + { + "bbox": [ + 140, + 239, + 480, + 367 + ], + "type": "image", + "image_path": "36569bd060fabdb2b2ff0f2eab6037a3db14d1a5df81990b0b4da9e711d5ce22.jpg" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 130, + 374, + 482, + 419 + ], + "lines": [ + { + "bbox": [ + 130, + 374, + 482, + 419 + ], + "spans": [ + { + "bbox": [ + 130, + 374, + 482, + 419 + ], + "type": "text", + "content": "Fig. 1: The main pipeline of the proposed approach. The above panel includes the multiscale feature extractor and Gaussian regressor. Both Gaussian and features are splatted to the 2D plane and concatenated and fed to the progressive and multiscale feature decoding module described in the low panel." + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "image_caption" + } + ], + "index": 3 + }, + { + "bbox": [ + 130, + 427, + 482, + 452 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 427, + 482, + 452 + ], + "spans": [ + { + "bbox": [ + 130, + 427, + 482, + 452 + ], + "type": "text", + "content": "To enable optimization via backpropagation, the covariance matrix could be decomposed into a rotation matrix " + }, + { + "bbox": [ + 130, + 427, + 482, + 452 + ], + "type": "inline_equation", + "content": "(R)" + }, + { + "bbox": [ + 130, + 427, + 482, + 452 + ], + "type": "text", + "content": " and a scaling matrix " + }, + { + "bbox": [ + 130, + 427, + 482, + 452 + ], + "type": "inline_equation", + "content": "(S)" + }, + { + "bbox": [ + 130, + 427, + 482, + 452 + ], + "type": "text", + "content": ":" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 273, + 458, + 481, + 473 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 273, + 458, + 481, + 473 + ], + "spans": [ + { + "bbox": [ + 273, + 458, + 481, + 473 + ], + "type": "interline_equation", + "content": "\\Sigma = R S S ^ {T} R ^ {T} \\tag {2}", + "image_path": "f6dffbd9c5deeb754b8e881065b9c6500bfbd40cb149063db820b60f52fc086f.jpg" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 130, + 481, + 482, + 518 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 481, + 482, + 518 + ], + "spans": [ + { + "bbox": [ + 130, + 481, + 482, + 518 + ], + "type": "text", + "content": "Given the camera trajectory, the projection of the 3D Gaussians to the 2D image plane can be characterized by a view transform matrix " + }, + { + "bbox": [ + 130, + 481, + 482, + 518 + ], + "type": "inline_equation", + "content": "(W)" + }, + { + "bbox": [ + 130, + 481, + 482, + 518 + ], + "type": "text", + "content": " and the Jacobian of the affine approximation of the projective transformation " + }, + { + "bbox": [ + 130, + 481, + 482, + 518 + ], + "type": "inline_equation", + "content": "(J)" + }, + { + "bbox": [ + 130, + 481, + 482, + 518 + ], + "type": "text", + "content": ", as in:" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 264, + 525, + 481, + 540 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 264, + 525, + 481, + 540 + ], + "spans": [ + { + "bbox": [ + 264, + 525, + 481, + 540 + ], + "type": "interline_equation", + "content": "\\Sigma^ {\\prime} = J W \\Sigma W ^ {T} J ^ {T} \\tag {3}", + "image_path": "747366ce9fff4132b4466b2ade48bb9a2df72a454bb37c530e1b2095ce4197a2.jpg" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 130, + 548, + 482, + 583 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 548, + 482, + 583 + ], + "spans": [ + { + "bbox": [ + 130, + 548, + 482, + 583 + ], + "type": "text", + "content": "where the " + }, + { + "bbox": [ + 130, + 548, + 482, + 583 + ], + "type": "inline_equation", + "content": "\\Sigma^{\\prime}" + }, + { + "bbox": [ + 130, + 548, + 482, + 583 + ], + "type": "text", + "content": " is the covariance matrix in 2D image planes. Thus, the " + }, + { + "bbox": [ + 130, + 548, + 482, + 583 + ], + "type": "inline_equation", + "content": "\\alpha" + }, + { + "bbox": [ + 130, + 548, + 482, + 583 + ], + "type": "text", + "content": "-blend of " + }, + { + "bbox": [ + 130, + 548, + 482, + 583 + ], + "type": "inline_equation", + "content": "\\mathcal{N}" + }, + { + "bbox": [ + 130, + 548, + 482, + 583 + ], + "type": "text", + "content": " ordered points overlapping a pixel is utilized to compute the final color " + }, + { + "bbox": [ + 130, + 548, + 482, + 583 + ], + "type": "inline_equation", + "content": "C" + }, + { + "bbox": [ + 130, + 548, + 482, + 583 + ], + "type": "text", + "content": " of the pixel:" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 250, + 584, + 481, + 616 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 250, + 584, + 481, + 616 + ], + "spans": [ + { + "bbox": [ + 250, + 584, + 481, + 616 + ], + "type": "interline_equation", + "content": "C = \\sum_ {i \\in \\mathcal {N}} c _ {i} \\alpha_ {i} \\prod_ {j = 1} ^ {i - 1} (1 - \\alpha_ {j}) \\tag {4}", + "image_path": "f0c31903e22b66ac2b7545fc4ac00c70190ab582536cf501f702e31aaa808bcb.jpg" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 130, + 617, + 481, + 641 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 617, + 481, + 641 + ], + "spans": [ + { + "bbox": [ + 130, + 617, + 481, + 641 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 130, + 617, + 481, + 641 + ], + "type": "inline_equation", + "content": "c_{i}" + }, + { + "bbox": [ + 130, + 617, + 481, + 641 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 130, + 617, + 481, + 641 + ], + "type": "inline_equation", + "content": "\\alpha_{i}" + }, + { + "bbox": [ + 130, + 617, + 481, + 641 + ], + "type": "text", + "content": " denote the color and density of the pixel with corresponding Gaussian parameters. Following [46], we projected the features similarly." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 130, + 642, + 481, + 665 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 642, + 481, + 665 + ], + "spans": [ + { + "bbox": [ + 130, + 642, + 481, + 665 + ], + "type": "text", + "content": "In summary, the above parameters can be characterized by the following attributes: " + }, + { + "bbox": [ + 130, + 642, + 481, + 665 + ], + "type": "inline_equation", + "content": "x" + }, + { + "bbox": [ + 130, + 642, + 481, + 665 + ], + "type": "text", + "content": " denotes the position of a primitive where " + }, + { + "bbox": [ + 130, + 642, + 481, + 665 + ], + "type": "inline_equation", + "content": "x \\in R^3" + }, + { + "bbox": [ + 130, + 642, + 481, + 665 + ], + "type": "text", + "content": ". The rotation" + } + ] + } + ], + "index": 12 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 421, + 91, + 447, + 100 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 421, + 91, + 447, + 100 + ], + "spans": [ + { + "bbox": [ + 421, + 91, + 447, + 100 + ], + "type": "text", + "content": "PFGS" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 474, + 92, + 481, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 474, + 92, + 481, + 100 + ], + "spans": [ + { + "bbox": [ + 474, + 92, + 481, + 100 + ], + "type": "text", + "content": "5" + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 4 + }, + { + "para_blocks": [ + { + "bbox": [ + 130, + 116, + 482, + 177 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 116, + 482, + 177 + ], + "spans": [ + { + "bbox": [ + 130, + 116, + 482, + 177 + ], + "type": "text", + "content": "matrix is characterized by a quaternion " + }, + { + "bbox": [ + 130, + 116, + 482, + 177 + ], + "type": "inline_equation", + "content": "q \\in R^4" + }, + { + "bbox": [ + 130, + 116, + 482, + 177 + ], + "type": "text", + "content": ". The scale factor is defined by the anisotropy stretching " + }, + { + "bbox": [ + 130, + 116, + 482, + 177 + ], + "type": "inline_equation", + "content": "s \\in R^3" + }, + { + "bbox": [ + 130, + 116, + 482, + 177 + ], + "type": "text", + "content": ". The opacity factor " + }, + { + "bbox": [ + 130, + 116, + 482, + 177 + ], + "type": "inline_equation", + "content": "\\alpha \\in R^1 \\in [0,1]" + }, + { + "bbox": [ + 130, + 116, + 482, + 177 + ], + "type": "text", + "content": ". The 2D opacity " + }, + { + "bbox": [ + 130, + 116, + 482, + 177 + ], + "type": "inline_equation", + "content": "\\alpha \\in [0,1]" + }, + { + "bbox": [ + 130, + 116, + 482, + 177 + ], + "type": "text", + "content": " is computed by " + }, + { + "bbox": [ + 130, + 116, + 482, + 177 + ], + "type": "inline_equation", + "content": "\\alpha_i(x) = o_i \\exp(-\\frac{1}{2} (x - \\mu_i)^T \\Sigma_i^T (x - \\mu_i))" + }, + { + "bbox": [ + 130, + 116, + 482, + 177 + ], + "type": "text", + "content": " where the " + }, + { + "bbox": [ + 130, + 116, + 482, + 177 + ], + "type": "inline_equation", + "content": "\\mu" + }, + { + "bbox": [ + 130, + 116, + 482, + 177 + ], + "type": "text", + "content": " and variance are the 2D-projected mean and variance of 3D Gaussians. And the color is represented by " + }, + { + "bbox": [ + 130, + 116, + 482, + 177 + ], + "type": "inline_equation", + "content": "SH" + }, + { + "bbox": [ + 130, + 116, + 482, + 177 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 131, + 194, + 481, + 205 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 131, + 194, + 481, + 205 + ], + "spans": [ + { + "bbox": [ + 131, + 194, + 481, + 205 + ], + "type": "text", + "content": "3.2 Gaussian Feature Prediction with Multiscale Feature Extraction" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 130, + 213, + 482, + 309 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 213, + 482, + 309 + ], + "spans": [ + { + "bbox": [ + 130, + 213, + 482, + 309 + ], + "type": "text", + "content": "We first extract features for each point as their neural descriptors. We use a multi-input, single-output UNet-like architecture as the extraction network to encode points with different scales and capture features across different spatial distances. The architecture is mainly made of several PV Convolution modules [17], which is a highly efficient and commonly used point-based feature integration operation. We downsample the original point cloud with decreasing rates and concatenate the downsampled point clouds with different levels of feature maps as extra inputs." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 201, + 319, + 481, + 332 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 201, + 319, + 481, + 332 + ], + "spans": [ + { + "bbox": [ + 201, + 319, + 481, + 332 + ], + "type": "interline_equation", + "content": "f _ {1: k} = \\mathcal {E} _ {\\theta} \\left(\\left(p _ {1: k}, c _ {1: k}\\right), \\left(p _ {1: k}, c _ {1: k}\\right) _ {\\downarrow r _ {1}}, \\left(p _ {1: k}, c _ {1: k}\\right) _ {\\downarrow r _ {2}}\\right) \\tag {5}", + "image_path": "e4c5b83950f4511910bcfcac443a42c428e9ed7ed329d30faddf6c5bf6d5737a.jpg" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 130, + 341, + 482, + 446 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 341, + 482, + 446 + ], + "spans": [ + { + "bbox": [ + 130, + 341, + 482, + 446 + ], + "type": "text", + "content": "The " + }, + { + "bbox": [ + 130, + 341, + 482, + 446 + ], + "type": "inline_equation", + "content": "\\mathcal{E}_{\\theta}" + }, + { + "bbox": [ + 130, + 341, + 482, + 446 + ], + "type": "text", + "content": " refers to the extractor network, " + }, + { + "bbox": [ + 130, + 341, + 482, + 446 + ], + "type": "inline_equation", + "content": "p, c" + }, + { + "bbox": [ + 130, + 341, + 482, + 446 + ], + "type": "text", + "content": " denote point coordinates and colors. The down arrow represents a uniform downsampling operation and " + }, + { + "bbox": [ + 130, + 341, + 482, + 446 + ], + "type": "inline_equation", + "content": "r" + }, + { + "bbox": [ + 130, + 341, + 482, + 446 + ], + "type": "text", + "content": " denotes the sampling rate where " + }, + { + "bbox": [ + 130, + 341, + 482, + 446 + ], + "type": "inline_equation", + "content": "r_1 > r_2" + }, + { + "bbox": [ + 130, + 341, + 482, + 446 + ], + "type": "text", + "content": ". The output " + }, + { + "bbox": [ + 130, + 341, + 482, + 446 + ], + "type": "inline_equation", + "content": "f_{1:k}" + }, + { + "bbox": [ + 130, + 341, + 482, + 446 + ], + "type": "text", + "content": " are point neural descriptors. This operation enables the extractor to recognize features of different scales and receptive fields. The detailed architecture and parameter setting are shown in our Appendix. The single output feature vector is considered as the descriptors for all points. Afterward, the neural descriptors, the coordinates, as well as colors are fed to another network to predict the Gaussian properties in a point-wise manner." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 199, + 447, + 481, + 466 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 199, + 447, + 481, + 466 + ], + "spans": [ + { + "bbox": [ + 199, + 447, + 481, + 466 + ], + "type": "interline_equation", + "content": "F _ {k} = \\mathcal {R} _ {\\theta} \\left(f _ {k}, p _ {k}, c _ {k}\\right) \\tag {6}", + "image_path": "6588ec21921c1a61feaf51468ac8e9347de276206353cc2a80c64271ae349266.jpg" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 199, + 461, + 415, + 473 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 199, + 461, + 415, + 473 + ], + "spans": [ + { + "bbox": [ + 199, + 461, + 415, + 473 + ], + "type": "interline_equation", + "content": "R, S, \\alpha = \\operatorname {n o r m} \\left(\\mathcal {H} _ {r} \\left(F _ {k}\\right)\\right), \\exp \\left(\\mathcal {H} _ {s} \\left(F _ {k}\\right)\\right), \\sigma \\left(\\mathcal {H} _ {o} \\left(F _ {k}\\right)\\right)", + "image_path": "38a1fa69dbc9618292f36f19433b57d09936baec05610bf475f064efaaeedd4e.jpg" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 130, + 478, + 482, + 550 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 478, + 482, + 550 + ], + "spans": [ + { + "bbox": [ + 130, + 478, + 482, + 550 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 130, + 478, + 482, + 550 + ], + "type": "inline_equation", + "content": "\\mathcal{R}_{\\theta}" + }, + { + "bbox": [ + 130, + 478, + 482, + 550 + ], + "type": "text", + "content": " represents the main network of the regressor, it maps multiple inputs to a single unified feature vector " + }, + { + "bbox": [ + 130, + 478, + 482, + 550 + ], + "type": "inline_equation", + "content": "F_{k}" + }, + { + "bbox": [ + 130, + 478, + 482, + 550 + ], + "type": "text", + "content": ". The " + }, + { + "bbox": [ + 130, + 478, + 482, + 550 + ], + "type": "inline_equation", + "content": "F_{k}" + }, + { + "bbox": [ + 130, + 478, + 482, + 550 + ], + "type": "text", + "content": " is fed to three individual prediction heads " + }, + { + "bbox": [ + 130, + 478, + 482, + 550 + ], + "type": "inline_equation", + "content": "(\\mathcal{H}_r,\\mathcal{H}_s,\\mathcal{H}_o)" + }, + { + "bbox": [ + 130, + 478, + 482, + 550 + ], + "type": "text", + "content": ", each of them is composed of two 1D Convolution layers with 1 kernel size, and activated by different functions to obtain corresponding Gaussian properties, namely rotation quaternion " + }, + { + "bbox": [ + 130, + 478, + 482, + 550 + ], + "type": "inline_equation", + "content": "(R)" + }, + { + "bbox": [ + 130, + 478, + 482, + 550 + ], + "type": "text", + "content": ", scale factor " + }, + { + "bbox": [ + 130, + 478, + 482, + 550 + ], + "type": "inline_equation", + "content": "(S)" + }, + { + "bbox": [ + 130, + 478, + 482, + 550 + ], + "type": "text", + "content": " and opacity " + }, + { + "bbox": [ + 130, + 478, + 482, + 550 + ], + "type": "inline_equation", + "content": "(\\alpha)" + }, + { + "bbox": [ + 130, + 478, + 482, + 550 + ], + "type": "text", + "content": ". The color property " + }, + { + "bbox": [ + 130, + 478, + 482, + 550 + ], + "type": "inline_equation", + "content": "c" + }, + { + "bbox": [ + 130, + 478, + 482, + 550 + ], + "type": "text", + "content": " remains that of the original points." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 130, + 550, + 482, + 598 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 550, + 482, + 598 + ], + "spans": [ + { + "bbox": [ + 130, + 550, + 482, + 598 + ], + "type": "text", + "content": "After the Gaussian attributes are obtained, one can render corresponding images from given camera viewpoints by Gaussian rasterization. We not only render the images by blending point colors on 2D planes but also render feature planes by replacing colors with the neural descriptors " + }, + { + "bbox": [ + 130, + 550, + 482, + 598 + ], + "type": "inline_equation", + "content": "(f_{k})" + }, + { + "bbox": [ + 130, + 550, + 482, + 598 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 169, + 608, + 481, + 620 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 169, + 608, + 481, + 620 + ], + "spans": [ + { + "bbox": [ + 169, + 608, + 481, + 620 + ], + "type": "interline_equation", + "content": "I _ {g s}, F _ {g s} = R e (R, S, \\alpha , p _ {1: k}, c _ {1: k} | K, P), R e (R, S, \\alpha , p _ {1: k}, f _ {1: k} | K, P) \\tag {7}", + "image_path": "557995237b59157e537e07de79c93c80c4aa37b13c60a54a0d72a41cd00b8962.jpg" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 130, + 629, + 482, + 665 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 629, + 482, + 665 + ], + "spans": [ + { + "bbox": [ + 130, + 629, + 482, + 665 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 130, + 629, + 482, + 665 + ], + "type": "inline_equation", + "content": "I_{gs}" + }, + { + "bbox": [ + 130, + 629, + 482, + 665 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 130, + 629, + 482, + 665 + ], + "type": "inline_equation", + "content": "F_{gs}" + }, + { + "bbox": [ + 130, + 629, + 482, + 665 + ], + "type": "text", + "content": " are corresponding image and feature map under the camera intrinsic " + }, + { + "bbox": [ + 130, + 629, + 482, + 665 + ], + "type": "inline_equation", + "content": "K" + }, + { + "bbox": [ + 130, + 629, + 482, + 665 + ], + "type": "text", + "content": " and pose " + }, + { + "bbox": [ + 130, + 629, + 482, + 665 + ], + "type": "inline_equation", + "content": "P" + }, + { + "bbox": [ + 130, + 629, + 482, + 665 + ], + "type": "text", + "content": ". The " + }, + { + "bbox": [ + 130, + 629, + 482, + 665 + ], + "type": "inline_equation", + "content": "I_{gs}" + }, + { + "bbox": [ + 130, + 629, + 482, + 665 + ], + "type": "text", + "content": " can be considered a good result of point rendering when the number of points is sufficient." + } + ] + } + ], + "index": 12 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "spans": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "type": "text", + "content": "6" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 166, + 91, + 225, + 102 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 166, + 91, + 225, + 102 + ], + "spans": [ + { + "bbox": [ + 166, + 91, + 225, + 102 + ], + "type": "text", + "content": "J. Wang et al." + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 5 + }, + { + "para_blocks": [ + { + "bbox": [ + 130, + 116, + 482, + 228 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 116, + 482, + 228 + ], + "spans": [ + { + "bbox": [ + 130, + 116, + 482, + 228 + ], + "type": "text", + "content": "However, when the point sparsely distributes over the space, the representation capability of corresponding Gaussians is constrained, especially in complex local areas, resulting in poor-quality rendering results. On the contrary, the neural descriptors contain rich information to describe local structures with various scales. We therefore utilize the " + }, + { + "bbox": [ + 130, + 116, + 482, + 228 + ], + "type": "inline_equation", + "content": "F_{gs}" + }, + { + "bbox": [ + 130, + 116, + 482, + 228 + ], + "type": "text", + "content": " to be gradually transformed to the final results in a multiscale and recurrent manner. The " + }, + { + "bbox": [ + 130, + 116, + 482, + 228 + ], + "type": "inline_equation", + "content": "F_{gs}" + }, + { + "bbox": [ + 130, + 116, + 482, + 228 + ], + "type": "text", + "content": " is 3D-consistently rasterized with the same Gaussian parameters in Eq. 7, thus it precisely corresponds to the " + }, + { + "bbox": [ + 130, + 116, + 482, + 228 + ], + "type": "inline_equation", + "content": "I_{gs}" + }, + { + "bbox": [ + 130, + 116, + 482, + 228 + ], + "type": "text", + "content": " in pixel-level. In other words, " + }, + { + "bbox": [ + 130, + 116, + 482, + 228 + ], + "type": "inline_equation", + "content": "F_{gs}" + }, + { + "bbox": [ + 130, + 116, + 482, + 228 + ], + "type": "text", + "content": " can be regarded as a complement to " + }, + { + "bbox": [ + 130, + 116, + 482, + 228 + ], + "type": "inline_equation", + "content": "I_{gs}" + }, + { + "bbox": [ + 130, + 116, + 482, + 228 + ], + "type": "text", + "content": ". Hence we concatenate " + }, + { + "bbox": [ + 130, + 116, + 482, + 228 + ], + "type": "inline_equation", + "content": "I_{gs}" + }, + { + "bbox": [ + 130, + 116, + 482, + 228 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 130, + 116, + 482, + 228 + ], + "type": "inline_equation", + "content": "F_{gs}" + }, + { + "bbox": [ + 130, + 116, + 482, + 228 + ], + "type": "text", + "content": " to be the input of the next module." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 131, + 246, + 389, + 259 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 131, + 246, + 389, + 259 + ], + "spans": [ + { + "bbox": [ + 131, + 246, + 389, + 259 + ], + "type": "text", + "content": "3.3 Multiscale and Progressive Feature Decoding" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 130, + 270, + 482, + 368 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 270, + 482, + 368 + ], + "spans": [ + { + "bbox": [ + 130, + 270, + 482, + 368 + ], + "type": "text", + "content": "After we render the feature volume " + }, + { + "bbox": [ + 130, + 270, + 482, + 368 + ], + "type": "inline_equation", + "content": "F_{gs}" + }, + { + "bbox": [ + 130, + 270, + 482, + 368 + ], + "type": "text", + "content": " associated with the provided camera viewpoint, we concatenate the " + }, + { + "bbox": [ + 130, + 270, + 482, + 368 + ], + "type": "inline_equation", + "content": "I_{gs}" + }, + { + "bbox": [ + 130, + 270, + 482, + 368 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 130, + 270, + 482, + 368 + ], + "type": "inline_equation", + "content": "F_{gs}" + }, + { + "bbox": [ + 130, + 270, + 482, + 368 + ], + "type": "text", + "content": " as input of the decoder. The decoder is a multi-input and multi-output UNet-like architecture, which aims to gradually refine and decode the input features and transform them into the target view images. We downsample the input two times with the rate of " + }, + { + "bbox": [ + 130, + 270, + 482, + 368 + ], + "type": "inline_equation", + "content": "r_a" + }, + { + "bbox": [ + 130, + 270, + 482, + 368 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 130, + 270, + 482, + 368 + ], + "type": "inline_equation", + "content": "r_b" + }, + { + "bbox": [ + 130, + 270, + 482, + 368 + ], + "type": "text", + "content": " and fuse them with the next two low-resolution layers of the UNet. There are interlaced connections between layers of different levels in the network, as stated in the main Fig. 1." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 182, + 378, + 481, + 392 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 182, + 378, + 481, + 392 + ], + "spans": [ + { + "bbox": [ + 182, + 378, + 481, + 392 + ], + "type": "interline_equation", + "content": "\\mathcal {F}, \\mathcal {F} _ {a}, \\mathcal {F} _ {b} = \\mathcal {D} _ {\\theta} \\left(\\left(F _ {g s} \\oplus I _ {g s}\\right), \\left(F _ {g s} \\oplus I _ {g s}\\right) _ {\\downarrow r _ {a}}, \\left(F _ {g s} \\oplus I _ {g s}\\right) _ {\\downarrow r _ {b}}\\right) \\tag {8}", + "image_path": "1278e6e3d6367b841082a6f97062a711afeda92ed085258ac0300b40ccb22d0c.jpg" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 130, + 402, + 482, + 484 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 402, + 482, + 484 + ], + "spans": [ + { + "bbox": [ + 130, + 402, + 482, + 484 + ], + "type": "text", + "content": "In the above equation, " + }, + { + "bbox": [ + 130, + 402, + 482, + 484 + ], + "type": "inline_equation", + "content": "\\mathcal{D}_{\\theta}" + }, + { + "bbox": [ + 130, + 402, + 482, + 484 + ], + "type": "text", + "content": " indicates the multiscale decoder. " + }, + { + "bbox": [ + 130, + 402, + 482, + 484 + ], + "type": "inline_equation", + "content": "\\downarrow r_{a}" + }, + { + "bbox": [ + 130, + 402, + 482, + 484 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 130, + 402, + 482, + 484 + ], + "type": "inline_equation", + "content": "\\downarrow r_{b}" + }, + { + "bbox": [ + 130, + 402, + 482, + 484 + ], + "type": "text", + "content": " are two downsampling operations. " + }, + { + "bbox": [ + 130, + 402, + 482, + 484 + ], + "type": "inline_equation", + "content": "\\mathcal{F},\\mathcal{F}_a,\\mathcal{F}_b" + }, + { + "bbox": [ + 130, + 402, + 482, + 484 + ], + "type": "text", + "content": " refer to the feature output of different levels of layers in which " + }, + { + "bbox": [ + 130, + 402, + 482, + 484 + ], + "type": "inline_equation", + "content": "\\mathcal{F}" + }, + { + "bbox": [ + 130, + 402, + 482, + 484 + ], + "type": "text", + "content": " denotes the full resolution feature output while " + }, + { + "bbox": [ + 130, + 402, + 482, + 484 + ], + "type": "inline_equation", + "content": "\\mathcal{F}" + }, + { + "bbox": [ + 130, + 402, + 482, + 484 + ], + "type": "text", + "content": " with subscripts " + }, + { + "bbox": [ + 130, + 402, + 482, + 484 + ], + "type": "inline_equation", + "content": "a" + }, + { + "bbox": [ + 130, + 402, + 482, + 484 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 130, + 402, + 482, + 484 + ], + "type": "inline_equation", + "content": "b" + }, + { + "bbox": [ + 130, + 402, + 482, + 484 + ], + "type": "text", + "content": " represent the lower resolution feature outputs. We apply a prediction head (marked as " + }, + { + "bbox": [ + 130, + 402, + 482, + 484 + ], + "type": "inline_equation", + "content": "\\mathcal{H}_{\\eta}" + }, + { + "bbox": [ + 130, + 402, + 482, + 484 + ], + "type": "text", + "content": ") composed of two convolutional layers and a ReLU activation to transform these features into the target images with different scales." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 130, + 487, + 482, + 571 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 487, + 482, + 571 + ], + "spans": [ + { + "bbox": [ + 130, + 487, + 482, + 571 + ], + "type": "text", + "content": "More importantly, we recurrently run the decoder network twice to refine the feature maps and the predicted images. The following experiments show the positive influence of such recurrent refinement on restoring the missing details due to the sparse point distribution. The input for the second loop is related to the output of the first loop. The prediction after the first loop can be described as " + }, + { + "bbox": [ + 130, + 487, + 482, + 571 + ], + "type": "inline_equation", + "content": "I^{l}, I_{a}^{l}, I_{b}^{l} = \\mathcal{H}_{\\eta}(\\mathcal{F}), \\mathcal{H}_{\\eta}(\\mathcal{F}_{a}), \\mathcal{H}_{\\eta}(\\mathcal{F}_{b})" + }, + { + "bbox": [ + 130, + 487, + 482, + 571 + ], + "type": "text", + "content": " where " + }, + { + "bbox": [ + 130, + 487, + 482, + 571 + ], + "type": "inline_equation", + "content": "l = 1" + }, + { + "bbox": [ + 130, + 487, + 482, + 571 + ], + "type": "text", + "content": " refers to the loop number. Likewise, the second loop is accordingly explained as:" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 170, + 580, + 481, + 596 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 170, + 580, + 481, + 596 + ], + "spans": [ + { + "bbox": [ + 170, + 580, + 481, + 596 + ], + "type": "interline_equation", + "content": "I ^ {l = 2}, I _ {a} ^ {l = 2}, I _ {b} ^ {l = 2} = \\mathcal {H} _ {\\eta} \\star \\mathcal {D} _ {\\theta} ((\\mathcal {F} \\oplus I ^ {l = 1}), (\\mathcal {F} _ {a} \\oplus I _ {a} ^ {l = 1}), (\\mathcal {F} _ {b} \\oplus I _ {b} ^ {l = 1})) \\tag {9}", + "image_path": "0be48842fe60e910e3640d2288254f7f1992a651a2885e87ed8ef09c427ac6fd.jpg" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 130, + 605, + 482, + 666 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 605, + 482, + 666 + ], + "spans": [ + { + "bbox": [ + 130, + 605, + 482, + 666 + ], + "type": "text", + "content": "in which " + }, + { + "bbox": [ + 130, + 605, + 482, + 666 + ], + "type": "inline_equation", + "content": "\\mathcal{H}_{\\eta} \\star \\mathcal{D}_{\\theta}" + }, + { + "bbox": [ + 130, + 605, + 482, + 666 + ], + "type": "text", + "content": " represents sequentially perform the two operations. We show the detailed architecture of the decoder in the Appendix. After we obtain the hierarchically multiscale predictions, we regard the full resolution result of the second loop " + }, + { + "bbox": [ + 130, + 605, + 482, + 666 + ], + "type": "inline_equation", + "content": "I^{l=2}" + }, + { + "bbox": [ + 130, + 605, + 482, + 666 + ], + "type": "text", + "content": " as the final prediction. However, we leverage all predictions to optimize the full models, which we introduce in the next section." + } + ] + } + ], + "index": 9 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 421, + 91, + 447, + 100 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 421, + 91, + 447, + 100 + ], + "spans": [ + { + "bbox": [ + 421, + 91, + 447, + 100 + ], + "type": "text", + "content": "PFGS" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 474, + 92, + 481, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 474, + 92, + 481, + 100 + ], + "spans": [ + { + "bbox": [ + 474, + 92, + 481, + 100 + ], + "type": "text", + "content": "7" + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 6 + }, + { + "para_blocks": [ + { + "bbox": [ + 132, + 116, + 257, + 128 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 116, + 257, + 128 + ], + "spans": [ + { + "bbox": [ + 132, + 116, + 257, + 128 + ], + "type": "text", + "content": "3.4 Training Strategies" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 130, + 142, + 482, + 228 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 142, + 482, + 228 + ], + "spans": [ + { + "bbox": [ + 130, + 142, + 482, + 228 + ], + "type": "text", + "content": "The proposed framework incorporates 3D-aware features into multiscale and progressive decoding by using 3D Gaussian Splatting. The integration of feature volumes under certain camera settings mainly relies on the Gaussian properties that are used for projections. Therefore, we first pretrain the feature extractor and Gaussian regressor in order to obtain a relatively reasonable rendering function with 8 epochs. In the first training stage, we employ the " + }, + { + "bbox": [ + 130, + 142, + 482, + 228 + ], + "type": "inline_equation", + "content": "L2" + }, + { + "bbox": [ + 130, + 142, + 482, + 228 + ], + "type": "text", + "content": " rendering loss between the Gaussian predicted images and labels, i.e. " + }, + { + "bbox": [ + 130, + 142, + 482, + 228 + ], + "type": "inline_equation", + "content": "I_{gs}" + }, + { + "bbox": [ + 130, + 142, + 482, + 228 + ], + "type": "text", + "content": " in Eq. 7." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 238, + 238, + 481, + 266 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 238, + 238, + 481, + 266 + ], + "spans": [ + { + "bbox": [ + 238, + 238, + 481, + 266 + ], + "type": "interline_equation", + "content": "\\mathcal {L} _ {g s} = \\sum_ {\\mathbf {u} \\in \\Omega} \\frac {1}{\\Omega} \\left| \\left| I _ {g s} (\\mathbf {u}) - I _ {g t} (\\mathbf {u}) \\right| \\right| ^ {2} \\tag {10}", + "image_path": "07064b7a0e7a5f7f6ff10d64ef3426eea578e353010fa19aa973d91b9a924c49.jpg" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 130, + 277, + 482, + 373 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 277, + 482, + 373 + ], + "spans": [ + { + "bbox": [ + 130, + 277, + 482, + 373 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 130, + 277, + 482, + 373 + ], + "type": "inline_equation", + "content": "\\mathbf{u} = \\{u,v\\}" + }, + { + "bbox": [ + 130, + 277, + 482, + 373 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 130, + 277, + 482, + 373 + ], + "type": "inline_equation", + "content": "\\Omega" + }, + { + "bbox": [ + 130, + 277, + 482, + 373 + ], + "type": "text", + "content": " is the pixel domain. After the first pretraining, the regressor can yield plausible Gaussian rendering parameters, thus the rendered feature maps provide a certain and view-dependent description of the scene. Based on this, we combine the rest modules with these trained modules to carry out joint optimization. According to the above illustration, the final outputs of the framework include two levels, each comprising three images of different scales. We leverage all of the outputs to optimize the model by using progressively multiscale image loss." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 204, + 386, + 481, + 418 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 204, + 386, + 481, + 418 + ], + "spans": [ + { + "bbox": [ + 204, + 386, + 481, + 418 + ], + "type": "interline_equation", + "content": "\\mathcal {L} _ {m i m} = \\sum_ {i = l} ^ {2} \\sum_ {s = r} ^ {3} \\sum_ {\\mathbf {u} \\in \\Omega} w (l) \\frac {1}{\\Omega} \\left| \\left| I _ {s} ^ {i} (\\mathbf {u}) - I _ {g t \\downarrow r _ {s}} (\\mathbf {u}) \\right| \\right| _ {1} ^ {1} \\tag {11}", + "image_path": "d797556f30f75c543911de9fab0e4b0c3b3cae07e94e86587e47d274c8f0f9d4.jpg" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 130, + 429, + 482, + 513 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 429, + 482, + 513 + ], + "spans": [ + { + "bbox": [ + 130, + 429, + 482, + 513 + ], + "type": "text", + "content": "in which " + }, + { + "bbox": [ + 130, + 429, + 482, + 513 + ], + "type": "inline_equation", + "content": "l" + }, + { + "bbox": [ + 130, + 429, + 482, + 513 + ], + "type": "text", + "content": " refers to the loop number in the progressive decoder. " + }, + { + "bbox": [ + 130, + 429, + 482, + 513 + ], + "type": "inline_equation", + "content": "s" + }, + { + "bbox": [ + 130, + 429, + 482, + 513 + ], + "type": "text", + "content": " denotes the scale number in each output of a loop and " + }, + { + "bbox": [ + 130, + 429, + 482, + 513 + ], + "type": "inline_equation", + "content": "\\downarrow r_s" + }, + { + "bbox": [ + 130, + 429, + 482, + 513 + ], + "type": "text", + "content": " is " + }, + { + "bbox": [ + 130, + 429, + 482, + 513 + ], + "type": "inline_equation", + "content": "s" + }, + { + "bbox": [ + 130, + 429, + 482, + 513 + ], + "type": "text", + "content": "' corresponding downsampling operations. " + }, + { + "bbox": [ + 130, + 429, + 482, + 513 + ], + "type": "inline_equation", + "content": "w(l)" + }, + { + "bbox": [ + 130, + 429, + 482, + 513 + ], + "type": "text", + "content": " is a simple weighting function related to " + }, + { + "bbox": [ + 130, + 429, + 482, + 513 + ], + "type": "inline_equation", + "content": "l" + }, + { + "bbox": [ + 130, + 429, + 482, + 513 + ], + "type": "text", + "content": ". Here we set " + }, + { + "bbox": [ + 130, + 429, + 482, + 513 + ], + "type": "inline_equation", + "content": "w(l) = 0.75" + }, + { + "bbox": [ + 130, + 429, + 482, + 513 + ], + "type": "text", + "content": " when " + }, + { + "bbox": [ + 130, + 429, + 482, + 513 + ], + "type": "inline_equation", + "content": "l = 1" + }, + { + "bbox": [ + 130, + 429, + 482, + 513 + ], + "type": "text", + "content": " otherwise " + }, + { + "bbox": [ + 130, + 429, + 482, + 513 + ], + "type": "inline_equation", + "content": "w(l) = 1" + }, + { + "bbox": [ + 130, + 429, + 482, + 513 + ], + "type": "text", + "content": ". In addition, we use the extra progressively multiscale frequency reconstruction loss to assist in restoring the high-frequency components, it is effective in reconstructing sharper edges and boundaries of images." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 175, + 525, + 481, + 556 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 175, + 525, + 481, + 556 + ], + "spans": [ + { + "bbox": [ + 175, + 525, + 481, + 556 + ], + "type": "interline_equation", + "content": "\\mathcal {L} _ {m f r} = \\sum_ {i = l} ^ {2} \\sum_ {s = r} ^ {3} \\sum_ {\\mathbf {u} \\in \\Omega} w (l) \\frac {1}{\\Omega} | | F F T \\left(I _ {s} ^ {i} (\\mathbf {u})\\right) - F F T \\left(I _ {g t _ {\\downarrow r _ {s}}} (\\mathbf {u})\\right) | | _ {1} ^ {1} \\tag {12}", + "image_path": "62b9819c32415f5f3acd3d0b7d427398d2339a166f4b69c41b5c7ff86a13ee76.jpg" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 130, + 567, + 482, + 614 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 567, + 482, + 614 + ], + "spans": [ + { + "bbox": [ + 130, + 567, + 482, + 614 + ], + "type": "text", + "content": "in which the " + }, + { + "bbox": [ + 130, + 567, + 482, + 614 + ], + "type": "inline_equation", + "content": "FFT" + }, + { + "bbox": [ + 130, + 567, + 482, + 614 + ], + "type": "text", + "content": " represents the fast Fourier transform (FFT) that transfers the image signal to the frequency domain. The remaining symbols retain the same meanings as in the previous context. The total loss function can be defined as follows:" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 231, + 621, + 481, + 633 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 231, + 621, + 481, + 633 + ], + "spans": [ + { + "bbox": [ + 231, + 621, + 481, + 633 + ], + "type": "interline_equation", + "content": "\\mathcal {L} _ {\\text {t o t a l}} = \\gamma_ {1} \\mathcal {L} _ {g s} + \\gamma_ {2} \\mathcal {L} _ {\\text {m i m}} + \\gamma_ {3} \\mathcal {L} _ {\\text {m f r}} \\tag {13}", + "image_path": "08e2de4f7c639dea15a590f8c99874e777d4244b1a1dec2f3b02173540b2eae2.jpg" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 130, + 641, + 481, + 666 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 641, + 481, + 666 + ], + "spans": [ + { + "bbox": [ + 130, + 641, + 481, + 666 + ], + "type": "text", + "content": "The " + }, + { + "bbox": [ + 130, + 641, + 481, + 666 + ], + "type": "inline_equation", + "content": "\\gamma_{1},\\gamma_{2}" + }, + { + "bbox": [ + 130, + 641, + 481, + 666 + ], + "type": "text", + "content": ", and " + }, + { + "bbox": [ + 130, + 641, + 481, + 666 + ], + "type": "inline_equation", + "content": "\\gamma_{3}" + }, + { + "bbox": [ + 130, + 641, + 481, + 666 + ], + "type": "text", + "content": " respectively control weights of these losses. In our implementation, we set " + }, + { + "bbox": [ + 130, + 641, + 481, + 666 + ], + "type": "inline_equation", + "content": "\\gamma_{1} = 0.75" + }, + { + "bbox": [ + 130, + 641, + 481, + 666 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 130, + 641, + 481, + 666 + ], + "type": "inline_equation", + "content": "\\gamma_{2} = 1" + }, + { + "bbox": [ + 130, + 641, + 481, + 666 + ], + "type": "text", + "content": ", and " + }, + { + "bbox": [ + 130, + 641, + 481, + 666 + ], + "type": "inline_equation", + "content": "\\gamma_{3} = 0.25" + }, + { + "bbox": [ + 130, + 641, + 481, + 666 + ], + "type": "text", + "content": " respectively." + } + ] + } + ], + "index": 11 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "spans": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "type": "text", + "content": "8" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 166, + 91, + 225, + 102 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 166, + 91, + 225, + 102 + ], + "spans": [ + { + "bbox": [ + 166, + 91, + 225, + 102 + ], + "type": "text", + "content": "J. Wang et al." + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 7 + }, + { + "para_blocks": [ + { + "type": "table", + "bbox": [ + 132, + 142, + 488, + 240 + ], + "blocks": [ + { + "bbox": [ + 132, + 114, + 479, + 136 + ], + "lines": [ + { + "bbox": [ + 132, + 114, + 479, + 136 + ], + "spans": [ + { + "bbox": [ + 132, + 114, + 479, + 136 + ], + "type": "text", + "content": "Table 1: Quantitative Results of comparisons between our method and different baselines under three datasets." + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 132, + 142, + 488, + 240 + ], + "lines": [ + { + "bbox": [ + 132, + 142, + 488, + 240 + ], + "spans": [ + { + "bbox": [ + 132, + 142, + 488, + 240 + ], + "type": "table", + "html": "
ScanNetDTUTHuman2.0
PSNR↑SSIM↑LPIPS↓PSNR↑SSIM↑LPIPS↓PSNR↑SSIM↑LPIPS↓
Graphics13.620.5280.77912.150.5250.68220.260.9050.337
NPBG15.090.5920.62513.520.7030.51419.770.9150.112
NPBG++16.810.6710.58522.320.8330.32726.810.9520.062
Trivol18.560.7340.47319.250.5920.51825.970.9350.059
Ours19.860.7580.45225.440.9010.16434.740.9830.009
", + "image_path": "a03e83b1e5756eed67a97c7dec4942e1bf6254f8a4ee2bdc78a96ce7b67eaefd.jpg" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "table_body" + } + ], + "index": 3 + }, + { + "bbox": [ + 132, + 255, + 229, + 269 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 255, + 229, + 269 + ], + "spans": [ + { + "bbox": [ + 132, + 255, + 229, + 269 + ], + "type": "text", + "content": "4 Experiments" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 132, + 278, + 254, + 291 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 278, + 254, + 291 + ], + "spans": [ + { + "bbox": [ + 132, + 278, + 254, + 291 + ], + "type": "text", + "content": "4.1 Experiments setup" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 132, + 297, + 481, + 498 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 297, + 481, + 498 + ], + "spans": [ + { + "bbox": [ + 132, + 297, + 481, + 498 + ], + "type": "text", + "content": "Datasets. We conduct extensive experiments to evaluate the effectiveness of our approach on three different datasets, including the indoor scene dataset ScanNet [4], objects dataset DTU [13], and human body dataset THuman2.0 [42]. The ScanNet dataset includes more than 2.5 million images at different views and with different qualities in over 1500 scenes. The point clouds are constructed from RGBD images. Following the Trivol, [9] we select the first 1200 scenes as the training set and the rest of the " + }, + { + "bbox": [ + 132, + 297, + 481, + 498 + ], + "type": "inline_equation", + "content": "300+" + }, + { + "bbox": [ + 132, + 297, + 481, + 498 + ], + "type": "text", + "content": " scenes as the test set. In addition, we use the preprocessed sparse point cloud in the dataset to test our model. For the DTU dataset, we transformed the point clouds of objects from their depth maps and images and we downsampled the point cloud by a factor of 0.5 to keep their point clouds distributing sparsely. In addition, we follow the MVSNeRF [2] to split the train and test groups of scans. The THuman2.0 dataset contains 500 high-quality human scans captured by a dense DLSR rig. For each scan, we render 36 views based on the provided 3D model and sparse sample points on the surface of the model. We keep the number of points for each scan at around 80,000 for their sparsity. We randomly select " + }, + { + "bbox": [ + 132, + 297, + 481, + 498 + ], + "type": "inline_equation", + "content": "75\\%" + }, + { + "bbox": [ + 132, + 297, + 481, + 498 + ], + "type": "text", + "content": " as the training set and the remaining " + }, + { + "bbox": [ + 132, + 297, + 481, + 498 + ], + "type": "inline_equation", + "content": "25\\%" + }, + { + "bbox": [ + 132, + 297, + 481, + 498 + ], + "type": "text", + "content": " as the test set." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 132, + 500, + 481, + 582 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 500, + 481, + 582 + ], + "spans": [ + { + "bbox": [ + 132, + 500, + 481, + 582 + ], + "type": "text", + "content": "Baselines and Metrics. We compare the proposed method with the currently prevailing point cloud rendering approaches to demonstrate the effectiveness of our framework. The baselines include a pure graphics-based renderer, learning-based renderers NPBG [12] and " + }, + { + "bbox": [ + 132, + 500, + 481, + 582 + ], + "type": "inline_equation", + "content": "\\mathrm{NPBG + + }" + }, + { + "bbox": [ + 132, + 500, + 481, + 582 + ], + "type": "text", + "content": " [24], and the NeRF-based renderer Trivol [9]. As for the quality, we adopt PSNR, SSIM [36], and LPIPS [44]. These metrics overall determine the reconstruction accuracy and quality of these methods." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 132, + 599, + 410, + 611 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 599, + 410, + 611 + ], + "spans": [ + { + "bbox": [ + 132, + 599, + 410, + 611 + ], + "type": "text", + "content": "4.2 Comparisons With Other Point Cloud Renderers." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 132, + 617, + 481, + 664 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 617, + 481, + 664 + ], + "spans": [ + { + "bbox": [ + 132, + 617, + 481, + 664 + ], + "type": "text", + "content": "Fig. 2, 3, and 4 show the results of qualitative comparisons on the ScanNet, DTU, and THuman2.0 datasets respectively. Table. 1 reports all the metrics on the three datasets of the baselines and ours. The best result is bold, and the second-best result is underlined in the Table." + } + ] + } + ], + "index": 9 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 421, + 91, + 447, + 100 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 421, + 91, + 447, + 100 + ], + "spans": [ + { + "bbox": [ + 421, + 91, + 447, + 100 + ], + "type": "text", + "content": "PFGS" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 474, + 92, + 481, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 474, + 92, + 481, + 100 + ], + "spans": [ + { + "bbox": [ + 474, + 92, + 481, + 100 + ], + "type": "text", + "content": "9" + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 8 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 135, + 114, + 203, + 169 + ], + "blocks": [ + { + "bbox": [ + 135, + 114, + 203, + 169 + ], + "lines": [ + { + "bbox": [ + 135, + 114, + 203, + 169 + ], + "spans": [ + { + "bbox": [ + 135, + 114, + 203, + 169 + ], + "type": "image", + "image_path": "3a34bd6b3fd9041c6d8e1ed248ad047bc46b3ab79bf4cf9b2369f3e0eddcddbf.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_body" + } + ], + "index": 2 + }, + { + "type": "image", + "bbox": [ + 205, + 115, + 272, + 170 + ], + "blocks": [ + { + "bbox": [ + 205, + 115, + 272, + 170 + ], + "lines": [ + { + "bbox": [ + 205, + 115, + 272, + 170 + ], + "spans": [ + { + "bbox": [ + 205, + 115, + 272, + 170 + ], + "type": "image", + "image_path": "cc43a5f46defb72c8955dc67f03d48a9b06b620baa4aae84a5c5e1540f9f458e.jpg" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_body" + } + ], + "index": 3 + }, + { + "type": "image", + "bbox": [ + 273, + 115, + 341, + 170 + ], + "blocks": [ + { + "bbox": [ + 273, + 115, + 341, + 170 + ], + "lines": [ + { + "bbox": [ + 273, + 115, + 341, + 170 + ], + "spans": [ + { + "bbox": [ + 273, + 115, + 341, + 170 + ], + "type": "image", + "image_path": "0e433f96003de660997196971492f2bef15837eb533c1ff6ac7ff23d93edb4ed.jpg" + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "image_body" + } + ], + "index": 4 + }, + { + "type": "image", + "bbox": [ + 342, + 115, + 410, + 170 + ], + "blocks": [ + { + "bbox": [ + 342, + 115, + 410, + 170 + ], + "lines": [ + { + "bbox": [ + 342, + 115, + 410, + 170 + ], + "spans": [ + { + "bbox": [ + 342, + 115, + 410, + 170 + ], + "type": "image", + "image_path": "2c0712f0f2820c8912d7cf09bcffbd8a5371e20d0466ccdbe0b444cf4ef1553a.jpg" + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "image_body" + } + ], + "index": 5 + }, + { + "type": "image", + "bbox": [ + 411, + 115, + 480, + 170 + ], + "blocks": [ + { + "bbox": [ + 411, + 115, + 480, + 170 + ], + "lines": [ + { + "bbox": [ + 411, + 115, + 480, + 170 + ], + "spans": [ + { + "bbox": [ + 411, + 115, + 480, + 170 + ], + "type": "image", + "image_path": "3aa2436bbe65eeb2d38718659cd0200d0a186fa5db137bfa181349de7cc5cb7f.jpg" + } + ] + } + ], + "index": 6, + "angle": 0, + "type": "image_body" + } + ], + "index": 6 + }, + { + "type": "image", + "bbox": [ + 135, + 171, + 203, + 225 + ], + "blocks": [ + { + "bbox": [ + 135, + 171, + 203, + 225 + ], + "lines": [ + { + "bbox": [ + 135, + 171, + 203, + 225 + ], + "spans": [ + { + "bbox": [ + 135, + 171, + 203, + 225 + ], + "type": "image", + "image_path": "ce1adff760ab46ca04475fe5c6e4dcc75c63810fc9735f4e422f138c07f48382.jpg" + } + ] + } + ], + "index": 7, + "angle": 0, + "type": "image_body" + } + ], + "index": 7 + }, + { + "type": "image", + "bbox": [ + 205, + 171, + 272, + 224 + ], + "blocks": [ + { + "bbox": [ + 205, + 171, + 272, + 224 + ], + "lines": [ + { + "bbox": [ + 205, + 171, + 272, + 224 + ], + "spans": [ + { + "bbox": [ + 205, + 171, + 272, + 224 + ], + "type": "image", + "image_path": "3c2239014b8bb4b7c81584ecc721b4d4a47e7f66f8cc6f11197651a6db91a835.jpg" + } + ] + } + ], + "index": 8, + "angle": 0, + "type": "image_body" + } + ], + "index": 8 + }, + { + "type": "image", + "bbox": [ + 273, + 171, + 341, + 224 + ], + "blocks": [ + { + "bbox": [ + 273, + 171, + 341, + 224 + ], + "lines": [ + { + "bbox": [ + 273, + 171, + 341, + 224 + ], + "spans": [ + { + "bbox": [ + 273, + 171, + 341, + 224 + ], + "type": "image", + "image_path": "c003333cb0a57cb91b64ce8ea17f5bdb7a23d41c5907b55cf4ac76404157cb91.jpg" + } + ] + } + ], + "index": 9, + "angle": 0, + "type": "image_body" + } + ], + "index": 9 + }, + { + "type": "image", + "bbox": [ + 342, + 171, + 410, + 224 + ], + "blocks": [ + { + "bbox": [ + 342, + 171, + 410, + 224 + ], + "lines": [ + { + "bbox": [ + 342, + 171, + 410, + 224 + ], + "spans": [ + { + "bbox": [ + 342, + 171, + 410, + 224 + ], + "type": "image", + "image_path": "74dd1159eda6cf5c82ba31c9cec47f604215945d21a803177df86e7777b17116.jpg" + } + ] + } + ], + "index": 10, + "angle": 0, + "type": "image_body" + } + ], + "index": 10 + }, + { + "type": "image", + "bbox": [ + 411, + 171, + 480, + 224 + ], + "blocks": [ + { + "bbox": [ + 411, + 171, + 480, + 224 + ], + "lines": [ + { + "bbox": [ + 411, + 171, + 480, + 224 + ], + "spans": [ + { + "bbox": [ + 411, + 171, + 480, + 224 + ], + "type": "image", + "image_path": "e378b7e77f50b3b7e5701fd3be0429b08665b7bf869da15122cc949dd283e621.jpg" + } + ] + } + ], + "index": 11, + "angle": 0, + "type": "image_body" + } + ], + "index": 11 + }, + { + "type": "image", + "bbox": [ + 135, + 225, + 203, + 279 + ], + "blocks": [ + { + "bbox": [ + 135, + 225, + 203, + 279 + ], + "lines": [ + { + "bbox": [ + 135, + 225, + 203, + 279 + ], + "spans": [ + { + "bbox": [ + 135, + 225, + 203, + 279 + ], + "type": "image", + "image_path": "dde24abb49cd11fba5415acd11a6ba4813a54220966a8cfffe6c583f51842201.jpg" + } + ] + } + ], + "index": 12, + "angle": 0, + "type": "image_body" + } + ], + "index": 12 + }, + { + "type": "image", + "bbox": [ + 205, + 225, + 272, + 279 + ], + "blocks": [ + { + "bbox": [ + 205, + 225, + 272, + 279 + ], + "lines": [ + { + "bbox": [ + 205, + 225, + 272, + 279 + ], + "spans": [ + { + "bbox": [ + 205, + 225, + 272, + 279 + ], + "type": "image", + "image_path": "74f1c471e86b1611e32089b78b6dfb08a62e40c8decb61a38de026fcb886c80c.jpg" + } + ] + } + ], + "index": 13, + "angle": 0, + "type": "image_body" + } + ], + "index": 13 + }, + { + "type": "image", + "bbox": [ + 273, + 225, + 341, + 279 + ], + "blocks": [ + { + "bbox": [ + 273, + 225, + 341, + 279 + ], + "lines": [ + { + "bbox": [ + 273, + 225, + 341, + 279 + ], + "spans": [ + { + "bbox": [ + 273, + 225, + 341, + 279 + ], + "type": "image", + "image_path": "53bd43002cbe255690c7593b37cea9cd70e356bf85f2abdd804c4f0d217dec58.jpg" + } + ] + } + ], + "index": 14, + "angle": 0, + "type": "image_body" + } + ], + "index": 14 + }, + { + "type": "image", + "bbox": [ + 342, + 225, + 410, + 279 + ], + "blocks": [ + { + "bbox": [ + 342, + 225, + 410, + 279 + ], + "lines": [ + { + "bbox": [ + 342, + 225, + 410, + 279 + ], + "spans": [ + { + "bbox": [ + 342, + 225, + 410, + 279 + ], + "type": "image", + "image_path": "fbef94c1c6513a492a1ff2f3fe031eac8108a705efb18462d24c51f6f624d273.jpg" + } + ] + } + ], + "index": 15, + "angle": 0, + "type": "image_body" + } + ], + "index": 15 + }, + { + "type": "image", + "bbox": [ + 411, + 225, + 480, + 279 + ], + "blocks": [ + { + "bbox": [ + 411, + 225, + 480, + 279 + ], + "lines": [ + { + "bbox": [ + 411, + 225, + 480, + 279 + ], + "spans": [ + { + "bbox": [ + 411, + 225, + 480, + 279 + ], + "type": "image", + "image_path": "0e9f06e8b26b7a5c5b584938217a0ef49043b4e14a4030a5ad2cbb04975fa454.jpg" + } + ] + } + ], + "index": 16, + "angle": 0, + "type": "image_body" + } + ], + "index": 16 + }, + { + "type": "image", + "bbox": [ + 135, + 280, + 203, + 335 + ], + "blocks": [ + { + "bbox": [ + 135, + 280, + 203, + 335 + ], + "lines": [ + { + "bbox": [ + 135, + 280, + 203, + 335 + ], + "spans": [ + { + "bbox": [ + 135, + 280, + 203, + 335 + ], + "type": "image", + "image_path": "2791c9e2fa5b3582b72051a0ae31264bfacf7665f3c649964d4345b5dd250e62.jpg" + } + ] + } + ], + "index": 17, + "angle": 0, + "type": "image_body" + } + ], + "index": 17 + }, + { + "type": "image", + "bbox": [ + 205, + 280, + 272, + 335 + ], + "blocks": [ + { + "bbox": [ + 205, + 280, + 272, + 335 + ], + "lines": [ + { + "bbox": [ + 205, + 280, + 272, + 335 + ], + "spans": [ + { + "bbox": [ + 205, + 280, + 272, + 335 + ], + "type": "image", + "image_path": "3bc4518628076c2a0a12009731b8c4b3a7b0db19829f3a17d80631ce28fbd2da.jpg" + } + ] + } + ], + "index": 18, + "angle": 0, + "type": "image_body" + } + ], + "index": 18 + }, + { + "type": "image", + "bbox": [ + 273, + 280, + 341, + 335 + ], + "blocks": [ + { + "bbox": [ + 273, + 280, + 341, + 335 + ], + "lines": [ + { + "bbox": [ + 273, + 280, + 341, + 335 + ], + "spans": [ + { + "bbox": [ + 273, + 280, + 341, + 335 + ], + "type": "image", + "image_path": "0af877d64b84cca3384c704d9a80c62652bf21cdaef957fe547b5c04dc443f50.jpg" + } + ] + } + ], + "index": 19, + "angle": 0, + "type": "image_body" + } + ], + "index": 19 + }, + { + "type": "image", + "bbox": [ + 342, + 280, + 410, + 335 + ], + "blocks": [ + { + "bbox": [ + 342, + 280, + 410, + 335 + ], + "lines": [ + { + "bbox": [ + 342, + 280, + 410, + 335 + ], + "spans": [ + { + "bbox": [ + 342, + 280, + 410, + 335 + ], + "type": "image", + "image_path": "a90cbd7254f39a29aae73ddff4a65cb7783ec8e98c99a7eea112fe921ecab657.jpg" + } + ] + } + ], + "index": 20, + "angle": 0, + "type": "image_body" + } + ], + "index": 20 + }, + { + "type": "image", + "bbox": [ + 411, + 280, + 480, + 335 + ], + "blocks": [ + { + "bbox": [ + 411, + 280, + 480, + 335 + ], + "lines": [ + { + "bbox": [ + 411, + 280, + 480, + 335 + ], + "spans": [ + { + "bbox": [ + 411, + 280, + 480, + 335 + ], + "type": "image", + "image_path": "b3e1e21d50c15562a0184d1168f3f67cf88d4b9ae3a13dc701dc2b1578f06a91.jpg" + } + ] + } + ], + "index": 21, + "angle": 0, + "type": "image_body" + } + ], + "index": 21 + }, + { + "type": "image", + "bbox": [ + 135, + 335, + 203, + 390 + ], + "blocks": [ + { + "bbox": [ + 135, + 335, + 203, + 390 + ], + "lines": [ + { + "bbox": [ + 135, + 335, + 203, + 390 + ], + "spans": [ + { + "bbox": [ + 135, + 335, + 203, + 390 + ], + "type": "image", + "image_path": "a6f52cfabf32d04165f985a997907205baf85afdc61136f6c7de71b115e05663.jpg" + } + ] + } + ], + "index": 22, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 153, + 392, + 187, + 401 + ], + "lines": [ + { + "bbox": [ + 153, + 392, + 187, + 401 + ], + "spans": [ + { + "bbox": [ + 153, + 392, + 187, + 401 + ], + "type": "text", + "content": "Graphics" + } + ] + } + ], + "index": 27, + "angle": 0, + "type": "image_caption" + } + ], + "index": 22 + }, + { + "type": "image", + "bbox": [ + 205, + 335, + 272, + 390 + ], + "blocks": [ + { + "bbox": [ + 205, + 335, + 272, + 390 + ], + "lines": [ + { + "bbox": [ + 205, + 335, + 272, + 390 + ], + "spans": [ + { + "bbox": [ + 205, + 335, + 272, + 390 + ], + "type": "image", + "image_path": "8417ae31182c63fdf8e40f9141b4d7f334c0b66583ef52ae86168691ca5e5459.jpg" + } + ] + } + ], + "index": 23, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 222, + 392, + 255, + 399 + ], + "lines": [ + { + "bbox": [ + 222, + 392, + 255, + 399 + ], + "spans": [ + { + "bbox": [ + 222, + 392, + 255, + 399 + ], + "type": "text", + "content": "NPBG++" + } + ] + } + ], + "index": 28, + "angle": 0, + "type": "image_caption" + } + ], + "index": 23 + }, + { + "type": "image", + "bbox": [ + 273, + 335, + 341, + 390 + ], + "blocks": [ + { + "bbox": [ + 273, + 335, + 341, + 390 + ], + "lines": [ + { + "bbox": [ + 273, + 335, + 341, + 390 + ], + "spans": [ + { + "bbox": [ + 273, + 335, + 341, + 390 + ], + "type": "image", + "image_path": "72bbd56efc3a9d7feff2f75f0d3be39567f62bbef291cea7a9e513e7fa4577af.jpg" + } + ] + } + ], + "index": 24, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 297, + 392, + 319, + 399 + ], + "lines": [ + { + "bbox": [ + 297, + 392, + 319, + 399 + ], + "spans": [ + { + "bbox": [ + 297, + 392, + 319, + 399 + ], + "type": "text", + "content": "Trivol" + } + ] + } + ], + "index": 29, + "angle": 0, + "type": "image_caption" + } + ], + "index": 24 + }, + { + "type": "image", + "bbox": [ + 342, + 335, + 410, + 390 + ], + "blocks": [ + { + "bbox": [ + 342, + 335, + 410, + 390 + ], + "lines": [ + { + "bbox": [ + 342, + 335, + 410, + 390 + ], + "spans": [ + { + "bbox": [ + 342, + 335, + 410, + 390 + ], + "type": "image", + "image_path": "c762a483576bf3e9b7f297c6fc4ed9c28f38fd78190ccc00942141bb8acd89ab.jpg" + } + ] + } + ], + "index": 25, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 367, + 392, + 387, + 400 + ], + "lines": [ + { + "bbox": [ + 367, + 392, + 387, + 400 + ], + "spans": [ + { + "bbox": [ + 367, + 392, + 387, + 400 + ], + "type": "text", + "content": "Ours" + } + ] + } + ], + "index": 30, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 130, + 407, + 480, + 429 + ], + "lines": [ + { + "bbox": [ + 130, + 407, + 480, + 429 + ], + "spans": [ + { + "bbox": [ + 130, + 407, + 480, + 429 + ], + "type": "text", + "content": "Fig. 2: Qualitative comparisons between ours and other point cloud renderers on the ScanNet dataset." + } + ] + } + ], + "index": 32, + "angle": 0, + "type": "image_caption" + } + ], + "index": 25 + }, + { + "type": "image", + "bbox": [ + 411, + 335, + 480, + 390 + ], + "blocks": [ + { + "bbox": [ + 411, + 335, + 480, + 390 + ], + "lines": [ + { + "bbox": [ + 411, + 335, + 480, + 390 + ], + "spans": [ + { + "bbox": [ + 411, + 335, + 480, + 390 + ], + "type": "image", + "image_path": "a3e98e2c4eafe5396aec5e65042dbef3b3e0a281c61c143ae3dcf14347e9cbca.jpg" + } + ] + } + ], + "index": 26, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 420, + 392, + 472, + 400 + ], + "lines": [ + { + "bbox": [ + 420, + 392, + 472, + 400 + ], + "spans": [ + { + "bbox": [ + 420, + 392, + 472, + 400 + ], + "type": "text", + "content": "Ground Truth" + } + ] + } + ], + "index": 31, + "angle": 0, + "type": "image_caption" + } + ], + "index": 26 + }, + { + "bbox": [ + 130, + 437, + 482, + 581 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 437, + 482, + 581 + ], + "spans": [ + { + "bbox": [ + 130, + 437, + 482, + 581 + ], + "type": "text", + "content": "ScanNet. The problem of uneven distribution and missing points exists in the real scene-level point cloud. As shown in the first column in Fig. 2, traditional graphical rendering leads to a large number of holes in the rendering result. When the point cloud distribution is uneven, the point cloud of nearby objects is relatively sparse, and the 2D-CNN of " + }, + { + "bbox": [ + 130, + 437, + 482, + 581 + ], + "type": "inline_equation", + "content": "\\mathrm{NPBG}++" + }, + { + "bbox": [ + 130, + 437, + 482, + 581 + ], + "type": "text", + "content": " pays more attention to the features of distant objects, ignoring the 3D information in the point cloud, resulting that nearby objects are covered in the rendering result. Our method uses 3DGS for feature projection, which better integrates 3D context information. There are a large number of artifacts in the results of " + }, + { + "bbox": [ + 130, + 437, + 482, + 581 + ], + "type": "inline_equation", + "content": "\\mathrm{NPBG}++" + }, + { + "bbox": [ + 130, + 437, + 482, + 581 + ], + "type": "text", + "content": " and Trivol. Our multi-scale feature extraction network excels at capturing features across various levels of detail, effectively decoding these projected features to enhance the rendering of lighting effects and details significantly." + } + ] + } + ], + "index": 33 + }, + { + "bbox": [ + 130, + 582, + 485, + 665 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 582, + 485, + 665 + ], + "spans": [ + { + "bbox": [ + 130, + 582, + 485, + 665 + ], + "type": "text", + "content": "DTU. As we stated in the Dataset description part, we maintain points distributed sparsely on the surface of the objects. Under this circumstance, the graphics-based transformation demonstrates more serious bleeding surfaces. Trivol always partially reconstructs the whole object because it depends on the fully implicit neural representation. This setting is more likely to be suited to class-specific point rendering, as they claimed in the original paper. In contrast, DTU includes many different classes that obstacle Trivol to learning a unified feature" + } + ] + } + ], + "index": 34 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "spans": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "type": "text", + "content": "10" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 166, + 91, + 225, + 102 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 166, + 91, + 225, + 102 + ], + "spans": [ + { + "bbox": [ + 166, + 91, + 225, + 102 + ], + "type": "text", + "content": "J. Wang et al." + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 9 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 135, + 114, + 203, + 169 + ], + "blocks": [ + { + "bbox": [ + 135, + 114, + 203, + 169 + ], + "lines": [ + { + "bbox": [ + 135, + 114, + 203, + 169 + ], + "spans": [ + { + "bbox": [ + 135, + 114, + 203, + 169 + ], + "type": "image", + "image_path": "cd7e8e3c419b67b0a9e1c4a5e97afbcd2c38b6a0ff753cf76840af3a10b0d0b0.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_body" + } + ], + "index": 2 + }, + { + "type": "image", + "bbox": [ + 135, + 170, + 203, + 224 + ], + "blocks": [ + { + "bbox": [ + 135, + 170, + 203, + 224 + ], + "lines": [ + { + "bbox": [ + 135, + 170, + 203, + 224 + ], + "spans": [ + { + "bbox": [ + 135, + 170, + 203, + 224 + ], + "type": "image", + "image_path": "a5b0070e917b1518abd623da711ed569c7bd906498eff8a4ff350cffd37cae6f.jpg" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_body" + } + ], + "index": 3 + }, + { + "type": "image", + "bbox": [ + 135, + 225, + 203, + 280 + ], + "blocks": [ + { + "bbox": [ + 135, + 225, + 203, + 280 + ], + "lines": [ + { + "bbox": [ + 135, + 225, + 203, + 280 + ], + "spans": [ + { + "bbox": [ + 135, + 225, + 203, + 280 + ], + "type": "image", + "image_path": "9078a231cf16884ae347dabeb3421730963e94f45fa066e60dea7730cd2371ad.jpg" + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "image_body" + } + ], + "index": 4 + }, + { + "type": "image", + "bbox": [ + 135, + 281, + 203, + 335 + ], + "blocks": [ + { + "bbox": [ + 135, + 281, + 203, + 335 + ], + "lines": [ + { + "bbox": [ + 135, + 281, + 203, + 335 + ], + "spans": [ + { + "bbox": [ + 135, + 281, + 203, + 335 + ], + "type": "image", + "image_path": "e496e1d75bbe501f37b082708972a92a1510f864ed5ed652968cfdddc8cbcb0b.jpg" + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "image_body" + } + ], + "index": 5 + }, + { + "type": "image", + "bbox": [ + 135, + 336, + 203, + 391 + ], + "blocks": [ + { + "bbox": [ + 135, + 336, + 203, + 391 + ], + "lines": [ + { + "bbox": [ + 135, + 336, + 203, + 391 + ], + "spans": [ + { + "bbox": [ + 135, + 336, + 203, + 391 + ], + "type": "image", + "image_path": "534bbafc15cdef1376e5331cd5e573188d73b8868bff007e6f0f9ee7e149c22b.jpg" + } + ] + } + ], + "index": 6, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 151, + 392, + 186, + 401 + ], + "lines": [ + { + "bbox": [ + 151, + 392, + 186, + 401 + ], + "spans": [ + { + "bbox": [ + 151, + 392, + 186, + 401 + ], + "type": "text", + "content": "Graphics" + } + ] + } + ], + "index": 7, + "angle": 0, + "type": "image_caption" + } + ], + "index": 6 + }, + { + "type": "image", + "bbox": [ + 203, + 114, + 272, + 169 + ], + "blocks": [ + { + "bbox": [ + 203, + 114, + 272, + 169 + ], + "lines": [ + { + "bbox": [ + 203, + 114, + 272, + 169 + ], + "spans": [ + { + "bbox": [ + 203, + 114, + 272, + 169 + ], + "type": "image", + "image_path": "fec8f76ef5e437d4a2e1d10d5cbc6d77d38b48501cc59de379a8a1f1da410e91.jpg" + } + ] + } + ], + "index": 8, + "angle": 0, + "type": "image_body" + } + ], + "index": 8 + }, + { + "type": "image", + "bbox": [ + 204, + 170, + 272, + 224 + ], + "blocks": [ + { + "bbox": [ + 204, + 170, + 272, + 224 + ], + "lines": [ + { + "bbox": [ + 204, + 170, + 272, + 224 + ], + "spans": [ + { + "bbox": [ + 204, + 170, + 272, + 224 + ], + "type": "image", + "image_path": "900cd61fe518779f43ae3c683195b156981eca68694aa8b0171f159da704c2a3.jpg" + } + ] + } + ], + "index": 9, + "angle": 0, + "type": "image_body" + } + ], + "index": 9 + }, + { + "type": "image", + "bbox": [ + 204, + 225, + 272, + 280 + ], + "blocks": [ + { + "bbox": [ + 204, + 225, + 272, + 280 + ], + "lines": [ + { + "bbox": [ + 204, + 225, + 272, + 280 + ], + "spans": [ + { + "bbox": [ + 204, + 225, + 272, + 280 + ], + "type": "image", + "image_path": "03a0b695b51f0a6522fc3190c07d4849d065a3716affb235fb5d1e142b88ee21.jpg" + } + ] + } + ], + "index": 10, + "angle": 0, + "type": "image_body" + } + ], + "index": 10 + }, + { + "type": "image", + "bbox": [ + 204, + 281, + 272, + 335 + ], + "blocks": [ + { + "bbox": [ + 204, + 281, + 272, + 335 + ], + "lines": [ + { + "bbox": [ + 204, + 281, + 272, + 335 + ], + "spans": [ + { + "bbox": [ + 204, + 281, + 272, + 335 + ], + "type": "image", + "image_path": "c678ab74a7fc9983411addf14e6825203587b972a1a7748c323af7cf23dc110f.jpg" + } + ] + } + ], + "index": 11, + "angle": 0, + "type": "image_body" + } + ], + "index": 11 + }, + { + "type": "image", + "bbox": [ + 204, + 336, + 272, + 391 + ], + "blocks": [ + { + "bbox": [ + 204, + 336, + 272, + 391 + ], + "lines": [ + { + "bbox": [ + 204, + 336, + 272, + 391 + ], + "spans": [ + { + "bbox": [ + 204, + 336, + 272, + 391 + ], + "type": "image", + "image_path": "7c58b6df8619f5572dc6e5ab98a387cffcebe6e0a702d2e9728cc22d866067ec.jpg" + } + ] + } + ], + "index": 12, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 223, + 392, + 257, + 400 + ], + "lines": [ + { + "bbox": [ + 223, + 392, + 257, + 400 + ], + "spans": [ + { + "bbox": [ + 223, + 392, + 257, + 400 + ], + "type": "text", + "content": "NPBG++" + } + ] + } + ], + "index": 13, + "angle": 0, + "type": "image_caption" + } + ], + "index": 12 + }, + { + "type": "image", + "bbox": [ + 273, + 114, + 341, + 169 + ], + "blocks": [ + { + "bbox": [ + 273, + 114, + 341, + 169 + ], + "lines": [ + { + "bbox": [ + 273, + 114, + 341, + 169 + ], + "spans": [ + { + "bbox": [ + 273, + 114, + 341, + 169 + ], + "type": "image", + "image_path": "f07622ea5d480b41fbf1394413cbf4a79235acc77591d52ca856c5e66c7fe582.jpg" + } + ] + } + ], + "index": 14, + "angle": 0, + "type": "image_body" + } + ], + "index": 14 + }, + { + "type": "image", + "bbox": [ + 273, + 170, + 340, + 224 + ], + "blocks": [ + { + "bbox": [ + 273, + 170, + 340, + 224 + ], + "lines": [ + { + "bbox": [ + 273, + 170, + 340, + 224 + ], + "spans": [ + { + "bbox": [ + 273, + 170, + 340, + 224 + ], + "type": "image", + "image_path": "1c2ef355942df92686b8e69ea9f8f863b06ca9d7b9f7cf0d609e08e1571d091a.jpg" + } + ] + } + ], + "index": 15, + "angle": 0, + "type": "image_body" + } + ], + "index": 15 + }, + { + "type": "image", + "bbox": [ + 273, + 225, + 340, + 280 + ], + "blocks": [ + { + "bbox": [ + 273, + 225, + 340, + 280 + ], + "lines": [ + { + "bbox": [ + 273, + 225, + 340, + 280 + ], + "spans": [ + { + "bbox": [ + 273, + 225, + 340, + 280 + ], + "type": "image", + "image_path": "0204fb6946aa852a289a4fe93103322ed0ceaac017f75bd049c97e8b7690f482.jpg" + } + ] + } + ], + "index": 16, + "angle": 0, + "type": "image_body" + } + ], + "index": 16 + }, + { + "type": "image", + "bbox": [ + 273, + 281, + 340, + 335 + ], + "blocks": [ + { + "bbox": [ + 273, + 281, + 340, + 335 + ], + "lines": [ + { + "bbox": [ + 273, + 281, + 340, + 335 + ], + "spans": [ + { + "bbox": [ + 273, + 281, + 340, + 335 + ], + "type": "image", + "image_path": "90d3ea82c194a288caf340dcb500f09a0417d43124a13065efb9fc97bf1d4565.jpg" + } + ] + } + ], + "index": 17, + "angle": 0, + "type": "image_body" + } + ], + "index": 17 + }, + { + "type": "image", + "bbox": [ + 273, + 336, + 340, + 391 + ], + "blocks": [ + { + "bbox": [ + 273, + 336, + 340, + 391 + ], + "lines": [ + { + "bbox": [ + 273, + 336, + 340, + 391 + ], + "spans": [ + { + "bbox": [ + 273, + 336, + 340, + 391 + ], + "type": "image", + "image_path": "68d9e99e02a84480988c3fc4737ae8d08f7c01509e465e769cebf933dfd7f13d.jpg" + } + ] + } + ], + "index": 18, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 297, + 392, + 320, + 400 + ], + "lines": [ + { + "bbox": [ + 297, + 392, + 320, + 400 + ], + "spans": [ + { + "bbox": [ + 297, + 392, + 320, + 400 + ], + "type": "text", + "content": "Trivol" + } + ] + } + ], + "index": 19, + "angle": 0, + "type": "image_caption" + } + ], + "index": 18 + }, + { + "type": "image", + "bbox": [ + 342, + 114, + 410, + 169 + ], + "blocks": [ + { + "bbox": [ + 342, + 114, + 410, + 169 + ], + "lines": [ + { + "bbox": [ + 342, + 114, + 410, + 169 + ], + "spans": [ + { + "bbox": [ + 342, + 114, + 410, + 169 + ], + "type": "image", + "image_path": "43628b3babe020c2f0e47f123d5734b533bea8b2fb7380fc3f5f342f2c863e7b.jpg" + } + ] + } + ], + "index": 20, + "angle": 0, + "type": "image_body" + } + ], + "index": 20 + }, + { + "type": "image", + "bbox": [ + 342, + 170, + 410, + 224 + ], + "blocks": [ + { + "bbox": [ + 342, + 170, + 410, + 224 + ], + "lines": [ + { + "bbox": [ + 342, + 170, + 410, + 224 + ], + "spans": [ + { + "bbox": [ + 342, + 170, + 410, + 224 + ], + "type": "image", + "image_path": "f834ef6cf179062b2d31581f7e0f23341eafb07eb4cccc4bf31be7b406829d4c.jpg" + } + ] + } + ], + "index": 21, + "angle": 0, + "type": "image_body" + } + ], + "index": 21 + }, + { + "type": "image", + "bbox": [ + 342, + 225, + 410, + 280 + ], + "blocks": [ + { + "bbox": [ + 342, + 225, + 410, + 280 + ], + "lines": [ + { + "bbox": [ + 342, + 225, + 410, + 280 + ], + "spans": [ + { + "bbox": [ + 342, + 225, + 410, + 280 + ], + "type": "image", + "image_path": "4ea53ddf98210f2921906d3ff6092d8a05414746b69fa7994c64b4ea498fc7fc.jpg" + } + ] + } + ], + "index": 22, + "angle": 0, + "type": "image_body" + } + ], + "index": 22 + }, + { + "type": "image", + "bbox": [ + 342, + 281, + 410, + 335 + ], + "blocks": [ + { + "bbox": [ + 342, + 281, + 410, + 335 + ], + "lines": [ + { + "bbox": [ + 342, + 281, + 410, + 335 + ], + "spans": [ + { + "bbox": [ + 342, + 281, + 410, + 335 + ], + "type": "image", + "image_path": "0549f9a2cb704ddb9ec2f9eb638ea6d4e224fb61265ef9ce5f0aafa5a1d1aedb.jpg" + } + ] + } + ], + "index": 23, + "angle": 0, + "type": "image_body" + } + ], + "index": 23 + }, + { + "type": "image", + "bbox": [ + 342, + 336, + 410, + 391 + ], + "blocks": [ + { + "bbox": [ + 342, + 336, + 410, + 391 + ], + "lines": [ + { + "bbox": [ + 342, + 336, + 410, + 391 + ], + "spans": [ + { + "bbox": [ + 342, + 336, + 410, + 391 + ], + "type": "image", + "image_path": "3802eeae13785f41156e779fd4da8862b9ee104e376e928a073b3464d5031a43.jpg" + } + ] + } + ], + "index": 24, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 367, + 392, + 386, + 400 + ], + "lines": [ + { + "bbox": [ + 367, + 392, + 386, + 400 + ], + "spans": [ + { + "bbox": [ + 367, + 392, + 386, + 400 + ], + "type": "text", + "content": "Ours" + } + ] + } + ], + "index": 25, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 132, + 408, + 480, + 430 + ], + "lines": [ + { + "bbox": [ + 132, + 408, + 480, + 430 + ], + "spans": [ + { + "bbox": [ + 132, + 408, + 480, + 430 + ], + "type": "text", + "content": "Fig. 3: Qualitative comparisons between ours and other point cloud rendering methods on the DTU dataset." + } + ] + } + ], + "index": 32, + "angle": 0, + "type": "image_caption" + } + ], + "index": 24 + }, + { + "type": "image", + "bbox": [ + 411, + 114, + 479, + 169 + ], + "blocks": [ + { + "bbox": [ + 411, + 114, + 479, + 169 + ], + "lines": [ + { + "bbox": [ + 411, + 114, + 479, + 169 + ], + "spans": [ + { + "bbox": [ + 411, + 114, + 479, + 169 + ], + "type": "image", + "image_path": "652a412911889a2293185ae31179e01afdc677df9706e9a859728094f91176f4.jpg" + } + ] + } + ], + "index": 26, + "angle": 0, + "type": "image_body" + } + ], + "index": 26 + }, + { + "type": "image", + "bbox": [ + 411, + 170, + 479, + 224 + ], + "blocks": [ + { + "bbox": [ + 411, + 170, + 479, + 224 + ], + "lines": [ + { + "bbox": [ + 411, + 170, + 479, + 224 + ], + "spans": [ + { + "bbox": [ + 411, + 170, + 479, + 224 + ], + "type": "image", + "image_path": "b199285bf82babde402ff41797e7e2547fce0b6b80511b47771f5f718feec624.jpg" + } + ] + } + ], + "index": 27, + "angle": 0, + "type": "image_body" + } + ], + "index": 27 + }, + { + "type": "image", + "bbox": [ + 411, + 225, + 479, + 280 + ], + "blocks": [ + { + "bbox": [ + 411, + 225, + 479, + 280 + ], + "lines": [ + { + "bbox": [ + 411, + 225, + 479, + 280 + ], + "spans": [ + { + "bbox": [ + 411, + 225, + 479, + 280 + ], + "type": "image", + "image_path": "ddf8a049c33d52c67b1ce40c1eb424591d6dadc6d5924504b96fe166fac18340.jpg" + } + ] + } + ], + "index": 28, + "angle": 0, + "type": "image_body" + } + ], + "index": 28 + }, + { + "type": "image", + "bbox": [ + 411, + 281, + 479, + 335 + ], + "blocks": [ + { + "bbox": [ + 411, + 281, + 479, + 335 + ], + "lines": [ + { + "bbox": [ + 411, + 281, + 479, + 335 + ], + "spans": [ + { + "bbox": [ + 411, + 281, + 479, + 335 + ], + "type": "image", + "image_path": "99746c9dd055c8136156f052c2e9f07d0598659e05a73bf6824729b4e92a39e3.jpg" + } + ] + } + ], + "index": 29, + "angle": 0, + "type": "image_body" + } + ], + "index": 29 + }, + { + "type": "image", + "bbox": [ + 411, + 336, + 479, + 391 + ], + "blocks": [ + { + "bbox": [ + 411, + 336, + 479, + 391 + ], + "lines": [ + { + "bbox": [ + 411, + 336, + 479, + 391 + ], + "spans": [ + { + "bbox": [ + 411, + 336, + 479, + 391 + ], + "type": "image", + "image_path": "85d51e737b6dba932d5b080a49b3c3f32e2d75f82729291314bf02dc19963512.jpg" + } + ] + } + ], + "index": 30, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 419, + 392, + 471, + 400 + ], + "lines": [ + { + "bbox": [ + 419, + 392, + 471, + 400 + ], + "spans": [ + { + "bbox": [ + 419, + 392, + 471, + 400 + ], + "type": "text", + "content": "Ground Truth" + } + ] + } + ], + "index": 31, + "angle": 0, + "type": "image_caption" + } + ], + "index": 30 + }, + { + "bbox": [ + 130, + 441, + 480, + 488 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 441, + 480, + 488 + ], + "spans": [ + { + "bbox": [ + 130, + 441, + 480, + 488 + ], + "type": "text", + "content": "transform underlying them. NPBG++ and our method deliver better results than Trivol because they leverage the original color and structure of point clouds without class-specific assumptions. Our approach yields the clearest predictions thanks to our more reasonable feature projection and decoding." + } + ] + } + ], + "index": 33 + }, + { + "bbox": [ + 130, + 490, + 481, + 596 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 490, + 481, + 596 + ], + "spans": [ + { + "bbox": [ + 130, + 490, + 481, + 596 + ], + "type": "text", + "content": "THuman2.0 The Graphics-based method reconstructs the main structures of the human body while inevitably seeing densely packed holes. The other two baselines generate similar results that are much better than those of the traditional methods. The difference is that Trivol fails to faithfully recover some cloth colors, which is because it is a fully implicit method and thereby cannot restore colors that are never seen in its train set. " + }, + { + "bbox": [ + 130, + 490, + 481, + 596 + ], + "type": "inline_equation", + "content": "\\mathrm{NPBG}++" + }, + { + "bbox": [ + 130, + 490, + 481, + 596 + ], + "type": "text", + "content": " and ours directly utilize points' colors thus avoiding this domain generalization issue. Moreover, our method restores more details such as human facial expressions and clothes' folds." + } + ] + } + ], + "index": 34 + }, + { + "bbox": [ + 132, + 608, + 246, + 618 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 608, + 246, + 618 + ], + "spans": [ + { + "bbox": [ + 132, + 608, + 246, + 618 + ], + "type": "text", + "content": "4.3 Ablation studies." + } + ] + } + ], + "index": 35 + }, + { + "bbox": [ + 130, + 629, + 480, + 665 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 629, + 480, + 665 + ], + "spans": [ + { + "bbox": [ + 130, + 629, + 480, + 665 + ], + "type": "text", + "content": "Ablations of main components. In this subsection, we evaluate the effectiveness of all main constituents proposed in our whole framework. We conduct experiments with different control variables set and report their performance on" + } + ] + } + ], + "index": 36 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 421, + 91, + 447, + 100 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 421, + 91, + 447, + 100 + ], + "spans": [ + { + "bbox": [ + 421, + 91, + 447, + 100 + ], + "type": "text", + "content": "PFGS" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 470, + 92, + 479, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 470, + 92, + 479, + 100 + ], + "spans": [ + { + "bbox": [ + 470, + 92, + 479, + 100 + ], + "type": "text", + "content": "11" + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 10 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 138, + 118, + 480, + 442 + ], + "blocks": [ + { + "bbox": [ + 138, + 118, + 480, + 442 + ], + "lines": [ + { + "bbox": [ + 138, + 118, + 480, + 442 + ], + "spans": [ + { + "bbox": [ + 138, + 118, + 480, + 442 + ], + "type": "image", + "image_path": "64a62b1e3fe8dd5a348e8144b5004164caaaaff1c6f88ba316caf5e832938f02.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 132, + 449, + 482, + 471 + ], + "lines": [ + { + "bbox": [ + 132, + 449, + 482, + 471 + ], + "spans": [ + { + "bbox": [ + 132, + 449, + 482, + 471 + ], + "type": "text", + "content": "Fig. 4: Qualitative comparisons between ours and other point cloud renderers on the THuman2.0 dataset." + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_caption" + } + ], + "index": 2 + }, + { + "bbox": [ + 130, + 474, + 482, + 668 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 474, + 482, + 668 + ], + "spans": [ + { + "bbox": [ + 130, + 474, + 482, + 668 + ], + "type": "text", + "content": "the ScanNet dataset in Table 2. In the table, " + }, + { + "bbox": [ + 130, + 474, + 482, + 668 + ], + "type": "inline_equation", + "content": "w / o" + }, + { + "bbox": [ + 130, + 474, + 482, + 668 + ], + "type": "text", + "content": " ProL refers to the version that removes all the progressive loss items, i.e. removes the first summation in Eq. 11 and Eq. 12. " + }, + { + "bbox": [ + 130, + 474, + 482, + 668 + ], + "type": "inline_equation", + "content": "w / o" + }, + { + "bbox": [ + 130, + 474, + 482, + 668 + ], + "type": "text", + "content": " MSL denotes the version that removes all the multiscale loss items that are the second summation in Eq. 11 and 12. The recur1 illustrates that we discard the recurrent structure from the feature decoder and only utilize a single multi-input and multi-output UNet. Besides, the " + }, + { + "bbox": [ + 130, + 474, + 482, + 668 + ], + "type": "inline_equation", + "content": "w / o" + }, + { + "bbox": [ + 130, + 474, + 482, + 668 + ], + "type": "text", + "content": " MSf describes a variant of the decoder which adopts a normal single scale UNet. In other words, this variant receives a single feature map as input and outputs a single prediction with full resolutions. Therefore, we only maintain the progressive losses and structures but discard all multiscale components for this version. Furthermore, we evaluate when the number of loops equals to 3, i.e. " + }, + { + "bbox": [ + 130, + 474, + 482, + 668 + ], + "type": "inline_equation", + "content": "l = 3" + }, + { + "bbox": [ + 130, + 474, + 482, + 668 + ], + "type": "text", + "content": " in these equations. This is marked as recur3 in the Table. The results show that continuously increasing the number of iterations does not further improve the model's performance; instead, it may lead to a slight decline. Last, the " + }, + { + "bbox": [ + 130, + 474, + 482, + 668 + ], + "type": "inline_equation", + "content": "w / o" + }, + { + "bbox": [ + 130, + 474, + 482, + 668 + ], + "type": "text", + "content": " Lmfr represents the version of removing the progressively multiscale frequency reconstruction loss (Eq. 12)." + } + ] + } + ], + "index": 4 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "spans": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "type": "text", + "content": "12" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 166, + 91, + 225, + 102 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 166, + 91, + 225, + 102 + ], + "spans": [ + { + "bbox": [ + 166, + 91, + 225, + 102 + ], + "type": "text", + "content": "J. Wang et al." + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 11 + }, + { + "para_blocks": [ + { + "type": "table", + "bbox": [ + 134, + 141, + 478, + 201 + ], + "blocks": [ + { + "bbox": [ + 132, + 109, + 479, + 131 + ], + "lines": [ + { + "bbox": [ + 132, + 109, + 479, + 131 + ], + "spans": [ + { + "bbox": [ + 132, + 109, + 479, + 131 + ], + "type": "text", + "content": "Table 2: Quantitative Results of ablation studies under various variants on the ScanNet dataset. The meaning of each abbreviation can be found in the text." + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 134, + 141, + 478, + 201 + ], + "lines": [ + { + "bbox": [ + 134, + 141, + 478, + 201 + ], + "spans": [ + { + "bbox": [ + 134, + 141, + 478, + 201 + ], + "type": "table", + "html": "
w/o ProLw/o MSLw/o MSfw/o Lmfrw/o 2trrecur1recur3Full
PSNR↑19.1919.3718.2719.4215.9219.1919.7719.86
SSIM↑0.7250.7330.7040.7400.6960.7290.7600.758
LPIPS↓0.4980.5090.6500.4780.6890.5160.4780.452
", + "image_path": "c24eed72414f241be373810abc1da7708b39935022d41de1c300299bf4496768.jpg" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "table_body" + } + ], + "index": 3 + }, + { + "type": "image", + "bbox": [ + 134, + 211, + 479, + 274 + ], + "blocks": [ + { + "bbox": [ + 134, + 211, + 479, + 274 + ], + "lines": [ + { + "bbox": [ + 134, + 211, + 479, + 274 + ], + "spans": [ + { + "bbox": [ + 134, + 211, + 479, + 274 + ], + "type": "image", + "image_path": "79ba08dca64e76c1dbb5a5bc7094bf0f5047d6a196c66ebd6923737026dd9952.jpg" + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 132, + 277, + 479, + 299 + ], + "lines": [ + { + "bbox": [ + 132, + 277, + 479, + 299 + ], + "spans": [ + { + "bbox": [ + 132, + 277, + 479, + 299 + ], + "type": "text", + "content": "Fig. 5: Qualitative Comparisons of the renderings produced by the Gaussian splatting after the first stage and the feature splatting after the whole framework." + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "image_caption" + } + ], + "index": 4 + }, + { + "bbox": [ + 130, + 312, + 480, + 407 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 312, + 480, + 407 + ], + "spans": [ + { + "bbox": [ + 130, + 312, + 480, + 407 + ], + "type": "text", + "content": "It is observed that the full model delivers the most satisfactory rendering results across all metrics while the rest variants experience a decline to some extent. This indicates that both progressive and multi-scale features have a positive impact and contribute to the final promising performance of the model. Therefore, the entire design of this method is consistently characterized by progressive and multi-scale aspects. Moreover, the alignment of predictions and labels in the frequency domain progressively and in a multiscale manner makes prediction clearer and sharper." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 130, + 426, + 480, + 581 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 426, + 480, + 581 + ], + "spans": [ + { + "bbox": [ + 130, + 426, + 480, + 581 + ], + "type": "text", + "content": "Analysis of Gaussian and Feature Splatting. The proposed approach relies on the two-stage training scheme because the 3D-consistent feature should be projected and integrated into certain 2D planes given camera viewpoints. The projection of features is dependent on the 3D Gaussian rasterization and further reliant on the Gaussian parameters associated with each point which are estimated by Eq. 6. The significance of training the Gaussian regression network in the first training stage lies in laying a solid foundation for feature splatting. Even though the 3D Gaussian Splatting could also produce plausible images, namely the " + }, + { + "bbox": [ + 130, + 426, + 480, + 581 + ], + "type": "inline_equation", + "content": "I_{gs}" + }, + { + "bbox": [ + 130, + 426, + 480, + 581 + ], + "type": "text", + "content": " in Eq. 7, the results are still severely influenced by points that are either locally or globally sparse, thereby missing details and showing holes. It can be seen in Fig. 5 that these Gaussian splatting results are indeed better than the pure graphics-based method due to the representation capabilities of 3DGS, but they still need to be improved further." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 130, + 582, + 481, + 665 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 582, + 481, + 665 + ], + "spans": [ + { + "bbox": [ + 130, + 582, + 481, + 665 + ], + "type": "text", + "content": "However, the results of feature splatting contain rich semantic information and local geometry cues, we consider them a good representation of restoring high-quality images and filling up blank areas. From Fig. 5 we can observe the capability of our method to generate and complete areas with voids and regions with sparse point clouds. We additionally evaluate an experiment in which we collaboratively train all modules in the framework rather than conducting the two-stage training and the results on the ScanNet dataset and we add this ad" + } + ] + } + ], + "index": 8 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 421, + 91, + 447, + 100 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 421, + 91, + 447, + 100 + ], + "spans": [ + { + "bbox": [ + 421, + 91, + 447, + 100 + ], + "type": "text", + "content": "PFGS" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 470, + 91, + 480, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 470, + 91, + 480, + 100 + ], + "spans": [ + { + "bbox": [ + 470, + 91, + 480, + 100 + ], + "type": "text", + "content": "13" + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 12 + }, + { + "para_blocks": [ + { + "type": "table", + "bbox": [ + 133, + 136, + 477, + 196 + ], + "blocks": [ + { + "bbox": [ + 132, + 114, + 480, + 126 + ], + "lines": [ + { + "bbox": [ + 132, + 114, + 480, + 126 + ], + "spans": [ + { + "bbox": [ + 132, + 114, + 480, + 126 + ], + "type": "text", + "content": "Table 3: Ablation studies of the effect of different numbers of points on performance." + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 133, + 136, + 477, + 196 + ], + "lines": [ + { + "bbox": [ + 133, + 136, + 477, + 196 + ], + "spans": [ + { + "bbox": [ + 133, + 136, + 477, + 196 + ], + "type": "table", + "html": "
Train Points50k50k50k120k120k120k
Test Points50k80k120k50k80k120k
Trivol24.0525.4326.4824.7125.9626.86
Ours31.2932.8635.9531.6034.1739.18
", + "image_path": "c100a01d98473941f776582841f8fe7b7a81221c1cbf56089df26157ee77cb7e.jpg" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "table_body" + } + ], + "index": 3 + }, + { + "bbox": [ + 130, + 205, + 480, + 242 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 205, + 480, + 242 + ], + "spans": [ + { + "bbox": [ + 130, + 205, + 480, + 242 + ], + "type": "text", + "content": "ditional result into Table. 2. " + }, + { + "bbox": [ + 130, + 205, + 480, + 242 + ], + "type": "inline_equation", + "content": "w / o" + }, + { + "bbox": [ + 130, + 205, + 480, + 242 + ], + "type": "text", + "content": " 2tr means that we abandon the first training stage and jointly train all modules together. In this setting, we observe that the model completely collapses and sees a very fast degradation in performance." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 130, + 256, + 482, + 389 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 256, + 482, + 389 + ], + "spans": [ + { + "bbox": [ + 130, + 256, + 482, + 389 + ], + "type": "text", + "content": "Analysis of The Number of Points. We additionally analyze and evaluate how different training and test numbers of points affect the performance. On the THuman2.0 dataset, we set different training and test pairs to show the performance of our method and Trivol respectively, which includes 50,000 and 120,000 points used for training and 50,000, 80,000, and 120,000 points used for testing. The results are listed in Table 3. It is noted that Trivol is essentially unaffected by the number of points where our method sees significant fluctuations with changes in the number of points. However, our method exhibits very high performance when point clouds are abundant while the Trivol method has a relatively low upper limit, which is due to its fully implicit representation and lack of utilization of point cloud prior. More discussion can be seen in the Appendix." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 132, + 403, + 217, + 416 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 403, + 217, + 416 + ], + "spans": [ + { + "bbox": [ + 132, + 403, + 217, + 416 + ], + "type": "text", + "content": "5 Discussion" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 130, + 426, + 482, + 605 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 426, + 482, + 605 + ], + "spans": [ + { + "bbox": [ + 130, + 426, + 482, + 605 + ], + "type": "text", + "content": "Conclusion. This work proposes a novel point cloud rendering method called PFGS, that can render high-fidelity and photorealistic images from colored point clouds with arbitrary camera viewpoints. To address the shortcomings of current learning-based point cloud renderers such as slow, blurry, and inconsistent, we introduce the 3D Gaussian Splatting to point cloud rendering by several well-designed modules. We supervise the regressor to estimate Gaussian parameters in a point-wise way, and then we utilize the estimated Gaussian parameters to rasterize point-wise neural features that are extracted by a multiscale feature decoder. Furthermore, the projected feature volume would be gradually decoded to the final prediction by our progressive and multiscale decoder. The whole architecture is supervised by our finely-designed progressive and multiscale reconstruction losses. We carry out experiments on three datasets with different levels of objects including indoors, objects, and humans and the results show our method outperforms all baselines. The ablation studies also illustrate the significance of all main components in the approach." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 130, + 606, + 482, + 665 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 606, + 482, + 665 + ], + "spans": [ + { + "bbox": [ + 130, + 606, + 482, + 665 + ], + "type": "text", + "content": "Limitation. Even though the proposed method can produce photorealistic images from sparse and uneven point clouds, it still yields unreal details when the given point clouds are not precise enough. Future works can incorporate a large generative refiner trained on a large dataset to slightly refine the original point cloud to make points closer to the actual surface of objects." + } + ] + } + ], + "index": 8 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "spans": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "type": "text", + "content": "14" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 166, + 91, + 225, + 102 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 166, + 91, + 225, + 102 + ], + "spans": [ + { + "bbox": [ + 166, + 91, + 225, + 102 + ], + "type": "text", + "content": "J. Wang et al." + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 13 + }, + { + "para_blocks": [ + { + "bbox": [ + 133, + 114, + 197, + 126 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 114, + 197, + 126 + ], + "spans": [ + { + "bbox": [ + 133, + 114, + 197, + 126 + ], + "type": "text", + "content": "References" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 138, + 140, + 480, + 665 + ], + "type": "list", + "angle": 0, + "index": 20, + "blocks": [ + { + "bbox": [ + 138, + 140, + 480, + 172 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 140, + 480, + 172 + ], + "spans": [ + { + "bbox": [ + 138, + 140, + 480, + 172 + ], + "type": "text", + "content": "1. Cao, A., Johnson, J.: Hexplane: A fast representation for dynamic scenes. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 130-141 (2023)" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 138, + 173, + 480, + 215 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 173, + 480, + 215 + ], + "spans": [ + { + "bbox": [ + 138, + 173, + 480, + 215 + ], + "type": "text", + "content": "2. Chen, A., Xu, Z., Zhao, F., Zhang, X., Xiang, F., Yu, J., Su, H.: Mvsnerf: Fast generalizable radiance field reconstruction from multi-view stereo. In: Proceedings of the IEEE/CVF International Conference on Computer Vision. pp. 14124-14133 (2021)" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 138, + 216, + 480, + 249 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 216, + 480, + 249 + ], + "spans": [ + { + "bbox": [ + 138, + 216, + 480, + 249 + ], + "type": "text", + "content": "3. Choy, C., Gwak, J., Savarese, S.: 4d spatio-temporal convnets: Minkowski convolutional neural networks. In: Proceedings of the IEEE/CVF conference on computer vision and pattern recognition. pp. 3075-3084 (2019)" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 138, + 250, + 480, + 282 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 250, + 480, + 282 + ], + "spans": [ + { + "bbox": [ + 138, + 250, + 480, + 282 + ], + "type": "text", + "content": "4. Dai, A., Chang, A.X., Savva, M., Halber, M., Funkhouser, T., Nießner, M.: Scannet: Richly-annotated 3d reconstructions of indoor scenes. In: Proceedings of the IEEE conference on computer vision and pattern recognition. pp. 5828-5839 (2017)" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 138, + 283, + 480, + 315 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 283, + 480, + 315 + ], + "spans": [ + { + "bbox": [ + 138, + 283, + 480, + 315 + ], + "type": "text", + "content": "5. Dai, P., Zhang, Y., Li, Z., Liu, S., Zeng, B.: Neural point cloud rendering via multi-plane projection. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 7830-7839 (2020)" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 138, + 316, + 480, + 348 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 316, + 480, + 348 + ], + "spans": [ + { + "bbox": [ + 138, + 316, + 480, + 348 + ], + "type": "text", + "content": "6. Fridovich-Keil, S., Yu, A., Tancik, M., Chen, Q., Recht, B., Kanazawa, A.: Plenoxels: Radiance fields without neural networks. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 5501-5510 (2022)" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 138, + 348, + 480, + 380 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 348, + 480, + 380 + ], + "spans": [ + { + "bbox": [ + 138, + 348, + 480, + 380 + ], + "type": "text", + "content": "7. Ge, W., Hu, T., Zhao, H., Liu, S., Chen, Y.C.: Ref-neus: Ambiguity-reduced neural implicit surface learning for multi-view reconstruction with reflection. arXiv preprint arXiv:2303.10840 (2023)" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 138, + 381, + 480, + 403 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 381, + 480, + 403 + ], + "spans": [ + { + "bbox": [ + 138, + 381, + 480, + 403 + ], + "type": "text", + "content": "8. Graham, B., Van der Maaten, L.: Submanifold sparse convolutional networks. arXiv preprint arXiv:1706.01307 (2017)" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 138, + 403, + 480, + 435 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 403, + 480, + 435 + ], + "spans": [ + { + "bbox": [ + 138, + 403, + 480, + 435 + ], + "type": "text", + "content": "9. Hu, T., Xu, X., Chu, R., Jia, J.: Trivol: Point cloud rendering via triple volumes. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 20732-20741 (2023)" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 138, + 435, + 480, + 468 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 435, + 480, + 468 + ], + "spans": [ + { + "bbox": [ + 138, + 435, + 480, + 468 + ], + "type": "text", + "content": "10. Hu, T., Xu, X., Liu, S., Jia, J.: Point2pix: Photo-realistic point cloud rendering via neural radiance fields. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 8349-8358 (2023)" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 138, + 468, + 480, + 501 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 468, + 480, + 501 + ], + "spans": [ + { + "bbox": [ + 138, + 468, + 480, + 501 + ], + "type": "text", + "content": "1. Huang, C., Mees, O., Zeng, A., Burgard, W.: Visual language maps for robot navigation. In: 2023 IEEE International Conference on Robotics and Automation (ICRA). pp. 10608-10615. IEEE (2023)" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 138, + 501, + 480, + 523 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 501, + 480, + 523 + ], + "spans": [ + { + "bbox": [ + 138, + 501, + 480, + 523 + ], + "type": "text", + "content": "2. Jena, S., Multon, F., Boukhayma, A.: Neural mesh-based graphics. In: European Conference on Computer Vision. pp. 739-757. Springer (2022)" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 138, + 523, + 480, + 555 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 523, + 480, + 555 + ], + "spans": [ + { + "bbox": [ + 138, + 523, + 480, + 555 + ], + "type": "text", + "content": "3. Jensen, R., Dahl, A., Vogiatzis, G., Tola, E., Aanaes, H.: Large scale multi-view stereopsis evaluation. In: Proceedings of the IEEE conference on computer vision and pattern recognition. pp. 406-413 (2014)" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 138, + 555, + 480, + 589 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 555, + 480, + 589 + ], + "spans": [ + { + "bbox": [ + 138, + 555, + 480, + 589 + ], + "type": "text", + "content": "4. Jiang, Y., Yu, C., Xie, T., Li, X., Feng, Y., Wang, H., Li, M., Lau, H., Gao, F., Yang, Y., et al.: Vr-gs: A physical dynamics-aware interactive gaussian splattering system in virtual reality. arXiv preprint arXiv:2401.16663 (2024)" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 138, + 589, + 480, + 611 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 589, + 480, + 611 + ], + "spans": [ + { + "bbox": [ + 138, + 589, + 480, + 611 + ], + "type": "text", + "content": "5. Kerbl, B., Kopanas, G., Leimkuhler, T., Drettakis, G.: 3d gaussian splatting for real-time radiance field rendering. ACM Transactions on Graphics 42(4) (2023)" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 138, + 611, + 480, + 643 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 611, + 480, + 643 + ], + "spans": [ + { + "bbox": [ + 138, + 611, + 480, + 643 + ], + "type": "text", + "content": "6. Lassner, C., Zollhofer, M.: Pulsar: Efficient sphere-based neural rendering. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 1440-1449 (2021)" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 138, + 643, + 480, + 665 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 643, + 480, + 665 + ], + "spans": [ + { + "bbox": [ + 138, + 643, + 480, + 665 + ], + "type": "text", + "content": "7. Liu, Z., Tang, H., Lin, Y., Han, S.: Point-voxel cnn for efficient 3d deep learning. Advances in Neural Information Processing Systems 32 (2019)" + } + ] + } + ], + "index": 19 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 421, + 91, + 447, + 100 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 421, + 91, + 447, + 100 + ], + "spans": [ + { + "bbox": [ + 421, + 91, + 447, + 100 + ], + "type": "text", + "content": "PFGS" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 470, + 92, + 481, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 470, + 92, + 481, + 100 + ], + "spans": [ + { + "bbox": [ + 470, + 92, + 481, + 100 + ], + "type": "text", + "content": "15" + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 14 + }, + { + "para_blocks": [ + { + "bbox": [ + 132, + 116, + 481, + 665 + ], + "type": "list", + "angle": 0, + "index": 18, + "blocks": [ + { + "bbox": [ + 133, + 116, + 480, + 149 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 116, + 480, + 149 + ], + "spans": [ + { + "bbox": [ + 133, + 116, + 480, + 149 + ], + "type": "text", + "content": "18. Lombardi, S., Simon, T., Saragih, J., Schwartz, G., Lehrmann, A., Sheikh, Y.: Neural volumes: Learning dynamic renderable volumes from images. arXiv preprint arXiv:1906.07751 (2019)" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 133, + 150, + 481, + 194 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 150, + 481, + 194 + ], + "spans": [ + { + "bbox": [ + 133, + 150, + 481, + 194 + ], + "type": "text", + "content": "19. Metzer, G., Richardson, E., Patashnik, O., Giryes, R., Cohen-Or, D.: Latent-nerf for shape-guided generation of 3d shapes and textures. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 12663-12673 (2023)" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 132, + 195, + 480, + 228 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 195, + 480, + 228 + ], + "spans": [ + { + "bbox": [ + 132, + 195, + 480, + 228 + ], + "type": "text", + "content": "20. Mildenhall, B., Srinivasan, P.P., Tancik, M., Barron, J.T., Ramamoorthi, R., Ng, R.: Nerf: Representing scenes as neural radiance fields for view synthesis. Communications of the ACM 65(1), 99-106 (2021)" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 132, + 228, + 480, + 261 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 228, + 480, + 261 + ], + "spans": [ + { + "bbox": [ + 132, + 228, + 480, + 261 + ], + "type": "text", + "content": "21. Müller, T., Evans, A., Schied, C., Keller, A.: Instant neural graphics primitives with a multiresolution hash encoding. ACM Transactions on Graphics (ToG) 41(4), 1-15 (2022)" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 132, + 262, + 480, + 295 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 262, + 480, + 295 + ], + "spans": [ + { + "bbox": [ + 132, + 262, + 480, + 295 + ], + "type": "text", + "content": "22. Qi, C.R., Su, H., Nießner, M., Dai, A., Yan, M., Guibas, L.J.: Volumetric and multi-view cnns for object classification on 3d data. In: Proceedings of the IEEE conference on computer vision and pattern recognition. pp. 5648-5656 (2016)" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 132, + 297, + 480, + 328 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 297, + 480, + 328 + ], + "spans": [ + { + "bbox": [ + 132, + 297, + 480, + 328 + ], + "type": "text", + "content": "23. Qi, C.R., Yi, L., Su, H., Guibas, L.J.: Pointnet++: Deep hierarchical feature learning on point sets in a metric space. Advances in neural information processing systems 30 (2017)" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 132, + 330, + 480, + 363 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 330, + 480, + 363 + ], + "spans": [ + { + "bbox": [ + 132, + 330, + 480, + 363 + ], + "type": "text", + "content": "24. Rakhimov, R., Ardelean, A.T., Lempitsky, V., Burnaev, E.: Npgb++: Accelerating neural point-based graphics. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 15969-15979 (2022)" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 132, + 364, + 480, + 396 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 364, + 480, + 396 + ], + "spans": [ + { + "bbox": [ + 132, + 364, + 480, + 396 + ], + "type": "text", + "content": "25. Ravi, N., Reizenstein, J., Novotny, D., Gordon, T., Lo, W.Y., Johnson, J., Gkioxari, G.: Accelerating 3d deep learning with pytorch3d. arXiv preprint arXiv:2007.08501 (2020)" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 132, + 397, + 480, + 441 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 397, + 480, + 441 + ], + "spans": [ + { + "bbox": [ + 132, + 397, + 480, + 441 + ], + "type": "text", + "content": "26. Sitzmann, V., Thies, J., Heide, F., Nießner, M., Wetzstein, G., Zollhofer, M.: Deepvoxels: Learning persistent 3d feature embeddings. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 2437-2446 (2019)" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 132, + 441, + 480, + 474 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 441, + 480, + 474 + ], + "spans": [ + { + "bbox": [ + 132, + 441, + 480, + 474 + ], + "type": "text", + "content": "27. Sun, W., Trulls, E., Tseng, Y.C., Sambandam, S., Sharma, G., Tagliasacchi, A., Yi, K.M.: Pointner++: A multi-scale, point-based neural radiance field. arXiv preprint arXiv:2312.02362 (2023)" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 132, + 475, + 480, + 508 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 475, + 480, + 508 + ], + "spans": [ + { + "bbox": [ + 132, + 475, + 480, + 508 + ], + "type": "text", + "content": "28. Wang, C., Chai, M., He, M., Chen, D., Liao, J.: Clip-nerf: Text-and-image driven manipulation of neural radiance fields. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 3835-3844 (2022)" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 132, + 509, + 480, + 531 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 509, + 480, + 531 + ], + "spans": [ + { + "bbox": [ + 132, + 509, + 480, + 531 + ], + "type": "text", + "content": "29. Wang, J., He, J., Zhang, Z., Xu, R.: Physical priors augmented event-based 3d reconstruction. arXiv preprint arXiv:2401.17121 (2024)" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 132, + 532, + 480, + 564 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 532, + 480, + 564 + ], + "spans": [ + { + "bbox": [ + 132, + 532, + 480, + 564 + ], + "type": "text", + "content": "30. Wang, J., Zhang, Z., Xu, R.: Learning to generate and manipulate 3d radiance field by a hierarchical diffusion framework with clip latent. In: Computer Graphics Forum. vol. 42, p. e14930. Wiley Online Library (2023)" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 132, + 565, + 480, + 597 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 565, + 480, + 597 + ], + "spans": [ + { + "bbox": [ + 132, + 565, + 480, + 597 + ], + "type": "text", + "content": "31. Wang, J., Zhang, Z., Xu, R.: Learning robust generalizable radiance field with visibility and feature augmented point representation. arXiv preprint arXiv:2401.14354 (2024)" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 132, + 599, + 480, + 632 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 599, + 480, + 632 + ], + "spans": [ + { + "bbox": [ + 132, + 599, + 480, + 632 + ], + "type": "text", + "content": "32. Wang, J., Sun, B., Lu, Y.: Mvpnet: Multi-view point regression networks for 3d object reconstruction from a single image. In: Proceedings of the AAAI Conference on Artificial Intelligence. vol. 33, pp. 8949-8956 (2019)" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 132, + 632, + 480, + 665 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 632, + 480, + 665 + ], + "spans": [ + { + "bbox": [ + 132, + 632, + 480, + 665 + ], + "type": "text", + "content": "33. Wang, N., Zhang, Y., Li, Z., Fu, Y., Liu, W., Jiang, Y.G.: Pixel2mesh: Generating 3d mesh models from single rgb images. In: Proceedings of the European conference on computer vision (ECCV). pp. 52-67 (2018)" + } + ] + } + ], + "index": 17 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "spans": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "type": "text", + "content": "16" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 166, + 91, + 224, + 102 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 166, + 91, + 224, + 102 + ], + "spans": [ + { + "bbox": [ + 166, + 91, + 224, + 102 + ], + "type": "text", + "content": "J. Wang et al." + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 15 + }, + { + "para_blocks": [ + { + "bbox": [ + 130, + 116, + 482, + 578 + ], + "type": "list", + "angle": 0, + "index": 15, + "blocks": [ + { + "bbox": [ + 130, + 116, + 482, + 161 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 116, + 482, + 161 + ], + "spans": [ + { + "bbox": [ + 130, + 116, + 482, + 161 + ], + "type": "text", + "content": "34. Wang, P., Liu, Y., Chen, Z., Liu, L., Liu, Z., Komura, T., Theobalt, C., Wang, W.: F2-nerf: Fast neural radiance field training with free camera trajectories. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 4150-4159 (2023)" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 130, + 162, + 481, + 194 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 162, + 481, + 194 + ], + "spans": [ + { + "bbox": [ + 130, + 162, + 481, + 194 + ], + "type": "text", + "content": "35. Wang, S., Liu, M.: Point cloud compression with range image-based entropy model for autonomous driving. In: European Conference on Computer Vision. pp. 323-340. Springer (2022)" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 132, + 194, + 481, + 226 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 194, + 481, + 226 + ], + "spans": [ + { + "bbox": [ + 132, + 194, + 481, + 226 + ], + "type": "text", + "content": "36. Wang, Z., Bovik, A.C., Sheikh, H.R., Simoncelli, E.P.: Image quality assessment: from error visibility to structural similarity. IEEE transactions on image processing 13(4), 600-612 (2004)" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 132, + 227, + 481, + 258 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 227, + 481, + 258 + ], + "spans": [ + { + "bbox": [ + 132, + 227, + 481, + 258 + ], + "type": "text", + "content": "37. Wu, G., Yi, T., Fang, J., Xie, L., Zhang, X., Wei, W., Liu, W., Tian, Q., Wang, X.: 4d gaussian splatting for real-time dynamic scene rendering. arXiv preprint arXiv:2310.08528 (2023)" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 132, + 259, + 481, + 303 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 259, + 481, + 303 + ], + "spans": [ + { + "bbox": [ + 132, + 259, + 481, + 303 + ], + "type": "text", + "content": "38. Xiang, F., Xu, Z., Hasan, M., Hold-Geoffroy, Y., Sunkavalli, K., Su, H.: Neutex: Neural texture mapping for volumetric neural rendering. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 7119-7128 (2021)" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 132, + 304, + 481, + 336 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 304, + 481, + 336 + ], + "spans": [ + { + "bbox": [ + 132, + 304, + 481, + 336 + ], + "type": "text", + "content": "39. Xu, Q., Xu, Z., Philip, J., Bi, S., Shu, Z., Sunkavalli, K., Neumann, U.: Point-nerf: Point-based neural radiance fields. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 5438-5448 (2022)" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 132, + 337, + 481, + 369 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 337, + 481, + 369 + ], + "spans": [ + { + "bbox": [ + 132, + 337, + 481, + 369 + ], + "type": "text", + "content": "40. Yan, Z., Li, C., Lee, G.H.: Nerf-ds: Neural radiance fields for dynamic specular objects. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 8285-8295 (2023)" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 132, + 370, + 481, + 413 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 370, + 481, + 413 + ], + "spans": [ + { + "bbox": [ + 132, + 370, + 481, + 413 + ], + "type": "text", + "content": "41. Yang, B., Bao, C., Zeng, J., Bao, H., Zhang, Y., Cui, Z., Zhang, G.: Neumesh: Learning disentangled neural mesh-based implicit field for geometry and texture editing. In: European Conference on Computer Vision. pp. 597-614. Springer (2022)" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 132, + 414, + 481, + 445 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 414, + 481, + 445 + ], + "spans": [ + { + "bbox": [ + 132, + 414, + 481, + 445 + ], + "type": "text", + "content": "42. Yu, T., Zheng, Z., Guo, K., Liu, P., Dai, Q., Liu, Y.: Function4d: Real-time human volumetric capture from very sparse consumer rgbd sensors. In: IEEE Conference on Computer Vision and Pattern Recognition (CVPR2021) (June 2021)" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 132, + 447, + 481, + 479 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 447, + 481, + 479 + ], + "spans": [ + { + "bbox": [ + 132, + 447, + 481, + 479 + ], + "type": "text", + "content": "43. Zhang, C., Guo, R., Zeng, W., Xiong, Y., Dai, B., Hu, R., Ren, M., Urtasun, R.: Rethinking closed-loop training for autonomous driving. In: European Conference on Computer Vision. pp. 264-282. Springer (2022)" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 132, + 479, + 481, + 511 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 479, + 481, + 511 + ], + "spans": [ + { + "bbox": [ + 132, + 479, + 481, + 511 + ], + "type": "text", + "content": "44. Zhang, R., Isola, P., Efros, A.A., Shechtman, E., Wang, O.: The unreasonable effectiveness of deep features as a perceptual metric. In: Proceedings of the IEEE conference on computer vision and pattern recognition. pp. 586-595 (2018)" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 132, + 512, + 481, + 533 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 512, + 481, + 533 + ], + "spans": [ + { + "bbox": [ + 132, + 512, + 481, + 533 + ], + "type": "text", + "content": "45. Zhou, Q.Y., Park, J., Koltun, V.: Open3d: A modern library for 3d data processing. arXiv preprint arXiv:1801.09847 (2018)" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 132, + 534, + 481, + 578 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 534, + 481, + 578 + ], + "spans": [ + { + "bbox": [ + 132, + 534, + 481, + 578 + ], + "type": "text", + "content": "46. Zhou, S., Chang, H., Jiang, S., Fan, Z., Zhu, Z., Xu, D., Chari, P., You, S., Wang, Z., Kadambi, A.: Feature 3dgs: Supercharging 3d gaussian splatting to enable distilled feature fields. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 21676-21685 (2024)" + } + ] + } + ], + "index": 14 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 421, + 91, + 447, + 100 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 421, + 91, + 447, + 100 + ], + "spans": [ + { + "bbox": [ + 421, + 91, + 447, + 100 + ], + "type": "text", + "content": "PFGS" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 470, + 91, + 481, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 470, + 91, + 481, + 100 + ], + "spans": [ + { + "bbox": [ + 470, + 91, + 481, + 100 + ], + "type": "text", + "content": "17" + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 16 + } + ], + "_backend": "vlm", + "_version_name": "2.6.4" +} \ No newline at end of file diff --git a/2024/PFedEdit_ Personalized Federated Learning via Automated Model Editing/984c6b27-188f-4dc6-b40f-18bc786fe91d_content_list.json b/2024/PFedEdit_ Personalized Federated Learning via Automated Model Editing/984c6b27-188f-4dc6-b40f-18bc786fe91d_content_list.json new file mode 100644 index 0000000000000000000000000000000000000000..1ec9419b5fcc4e9241d9f1e198a3e00e3662d61d --- /dev/null +++ b/2024/PFedEdit_ Personalized Federated Learning via Automated Model Editing/984c6b27-188f-4dc6-b40f-18bc786fe91d_content_list.json @@ -0,0 +1,1544 @@ +[ + { + "type": "text", + "text": "PFEDEDIT: Personalized Federated Learning via Automated Model Editing", + "text_level": 1, + "bbox": [ + 225, + 140, + 779, + 186 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Haolin Yuan $^{1}$ , William Paul $^{2}$ , John Aucott $^{3}$ , Philippe Burlina $^{2}$ , and Yinzhi Cao $^{1}$", + "bbox": [ + 256, + 210, + 746, + 242 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "1 Johns Hopkins University {hyuan4, yinzhi.cao}@jhu.edu", + "bbox": [ + 295, + 253, + 704, + 268 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "2 Johns Hopkins Applied Physics Lab {william.paul,", + "bbox": [ + 316, + 268, + 683, + 284 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "philippe.burlina}@jhuapl.edu", + "bbox": [ + 390, + 284, + 611, + 297 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "3 Johns Hopkins University School of Medicine jaucott2@jhmi.edu", + "bbox": [ + 274, + 296, + 725, + 311 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Abstract. Federated learning (FL) allows clients to train a deep learning model collaboratively while maintaining their private data locally. One challenging problem facing FL is that the model utility drops significantly once the data distribution gets heterogeneous, or non-i.i.d, among clients. A promising solution is to personalize models for each client, e.g., keeping some layers locally without aggregation, which is thus called personalized FL. However, previous personalized FL often suffer from suboptimal utility because their choice of layer personalization is based on empirical knowledge and fixed for different datasets and distributions. In this work, we design PFEDEDIT, the first federated learning framework that leverages automated model editing to optimize the choice of personalization layers and improve model utility under a variety of data distributions including non-i.i.d. The high-level idea of PFEDEDIT is to assess the effectiveness of every global model layer in improving model utility on local data distribution once edited, and then to apply edits on the top- $k$ most effective layers. Our evaluation shows that PFEDEDIT outperforms six state-of-the-art approaches on three benchmark datasets by $6\\%$ on the model's performance on average, with the largest accuracy improvement being $26.6\\%$ . PFEDEDIT is open-source and available at this repository: https://github.com/Haolin-Yuan/PFedEdit", + "bbox": [ + 261, + 351, + 743, + 628 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "1 Introduction", + "text_level": 1, + "bbox": [ + 215, + 656, + 375, + 672 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Federated learning (FL) [13,15,20,31,33] has emerged as a widely used training algorithm of deep learning, especially in the image domain, due to its practical applications in real-world scenarios and its impressive performance. A federated learning training scenario entails multiple local clients and a global server. Each client trains their local model using their private image data. After performing local training, each client updates their local model weights to the global server for aggregation with other clients' model weights. Federated learning allows clients to train a collaborative model that performs well on multi-class tasks—even if each client only possesses a few classes of images—via only sharing models but not local training images.", + "bbox": [ + 212, + 688, + 789, + 840 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/771cb20bf21dae387b6f5b9a601b277237bea59b2b0ccd95e59bd8fa837624f5.jpg", + "image_caption": [ + "Fig. 1: High-level idea of PFEDEDIT (PFEDEDIT edits the global model returned from FL server to fit the local client's distribution.)" + ], + "image_footnote": [], + "bbox": [ + 218, + 143, + 787, + 236 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "As the goal of federated learning is to learn a global model for all devices cooperatively, if the data distributions among clients are similar, the global model is supposed to perform better than local models that are solely trained on local data. However, once the data distribution becomes skewed, or non-i.i.d., e.g., some clients may possess image samples of only one or two classes or some possess many more image samples from one class than others, the global model becomes ineffective to be employed as local models. The reason is that the local data distribution at one client could be dramatically different from others, thus causing local model weights to adversely affect each other during the model aggregation.", + "bbox": [ + 212, + 316, + 787, + 467 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Multiple works [11,14,16] have been proposed to improve the FL model accuracy under the non-i.i.d. distribution by optimizing local models. They either regularize local models to maximize the agreement between clients [14], or reduce clients' weights drift from the server end [11,16] for a better global model that generalizes on each local distribution. While such works improved model's accuracy, the model is still uniform across different clients, and cannot be tailored based on different clients' data distribution. Therefore, one natural step is a so-called personalized FL [2,4,18,36], which enables each user to have personalized layers that help customize local models to their data distributions. Those personalized layers are retained locally each communication round and are not updated from the global model, thereby preventing being aggregated with others and avoiding bias towards other clients' data distributions. The choices of those personalized layers are various but are all pre-defined before the training, e.g., it can be linear layer [2], batch normalization layer [4], or specially designed layer [36].", + "bbox": [ + 212, + 472, + 787, + 700 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "However, there are still two challenges that remain for personalized FLs despite the accuracy improvement. First, the personalized layers chosen in previous personalized FL works [2,4,18,36] are inspired by empirical findings and fixed, implying that those choices may not be optimal when given different data distributions or datasets. That is, while the weights of those layers might be personalized, the locations and types of those layers are not, e.g., either linear or batch normalization layers in certain locations of a learning model. Second, previous personalized works are limited in their application scenarios as they assume certain model structure [18], which may not be used for certain clients.", + "bbox": [ + 212, + 704, + 787, + 842 + ], + "page_idx": 1 + }, + { + "type": "page_number", + "text": "2", + "bbox": [ + 217, + 114, + 228, + 126 + ], + "page_idx": 1 + }, + { + "type": "header", + "text": "H. Yuan et al.", + "bbox": [ + 271, + 114, + 367, + 126 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "In this paper, to address these two challenges, we design a framework that automatically locates and chooses layers to be personalized. We propose the first personalized federated learning framework, called PFEDEDIT, which leverages automated model editing [3,5,21,23] to optimize the layer choosing and improve model utility under different settings including non-i.i.d. The high-level idea of PFEDEDIT, as shown in Figure 1, is to look for model layers that contain knowledge about local distribution that benefits clients the most, and by editing those layers, PFEDEDIT mitigates the negative influence brought by other clients, thus tailoring the global model weights towards the local model weights in each round. Specifically, PFEDEDIT assesses which layers of the global model are the most effective in maintaining the local distribution and then marks them as target layers. PFEDEDIT then edits the top- $k$ target layers to tailor the global model weights closer to local models thus improving the model utility. PFEDEDIT does not make any assumptions on model selection and thus is model-structure agnostics.", + "bbox": [ + 212, + 146, + 787, + 372 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "We performed extensive experiments on three benchmark datasets, namely CIFAR-10 [12], CIFAR-100 [12], and Lyme dataset [35], and we compared PFEDEDIT with FedAvg [20] and other six state-of-the-art personalized federated learning approaches. Our experimental results demonstrate that PFEDEDIT outperforms all SOTA personalized federated learning works on all the datasets, with an average $6\\%$ accuracy improvement.", + "bbox": [ + 212, + 373, + 813, + 463 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "In summary, our contributions are as follows:", + "bbox": [ + 238, + 464, + 566, + 478 + ], + "page_idx": 2 + }, + { + "type": "list", + "sub_type": "text", + "list_items": [ + "- We design PFEDEDIT, the first personalized federated learning framework that optimizes personalized layer selections for each client using model editing.", + "- PFEDEDIT is adaptable to all model structures because it does not impose any assumption on model selection.", + "- We conduct extensive experiments to show that PFEDEDIT outperforms other SOTA personalized federated works on different benchmark datasets." + ], + "bbox": [ + 212, + 481, + 784, + 570 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "2 Related Work", + "text_level": 1, + "bbox": [ + 215, + 595, + 387, + 612 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Federated Learning. Federated learning [13,15,20,31] (FL) is first proposed to predict user's text input on portable devices and it has been widely deployed in other industries, such as medical imaging [19], Internet of Things (IoT) [24], and block-chain [25]. It allows multiple clients to collaboratively train a model that can perform on samples from any clients' data distribution. Two common approaches in federated learning (FL) are FedSGD [28] and FedAvg [20]. FedSGD optimizes global weights by back-propagating each local gradient, while FedAvg allows for multiple local training epochs and client-side optimization. In FedAvg, the server's role is limited to averaging and distributing the aggregated weights. While FedAvg reduces communication costs compared to FedSGD, it lacks a convergence guarantee, and heterogeneous data distribution among clients can significantly reduce the global model's utility. Therefore, various works [2,4,14,18,32,36] have been proposed to solve such utility degradation caused by data heterogeneity. One direction [14] is to optimize local", + "bbox": [ + 212, + 627, + 787, + 843 + ], + "page_idx": 2 + }, + { + "type": "header", + "text": "PFEDEDIT", + "bbox": [ + 653, + 114, + 730, + 126 + ], + "page_idx": 2 + }, + { + "type": "page_number", + "text": "3", + "bbox": [ + 774, + 116, + 784, + 126 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "model either via personalized objective functions; yet, another, called personalized FL [2,4,18,36], allows clients to have personalized model layers that can be kept locally. As a comparison, PFEDEDIT is the first work that optimizes and personalizes the choice of model layers with further improved utilities of FL, especially under non-i.i.d. settings.", + "bbox": [ + 212, + 146, + 787, + 223 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Model Editing. Model editing 3.5.17.23.26 involves adding, removing, or adjusting layers and parameters of a model architecture to improve its performance or adapt it to new tasks. Specifically, let $h_i$ be the $i$ -th hidden states of any model, $\\{h_l | l \\in [1, L]\\}$ , where $L$ is the number of layers in model, model editing allows users to find each hidden state's contribution towards a specific prediction. It mainly contains three runs: i). clean run, in which the statistics of each state are recorded when fed a factual prompt $x$ ; ii). corrupted run, in which each hidden layer is added Gaussian noise such that $h_l \\coloneqq h_l + \\epsilon$ . Because the hidden layer is obfuscated by the added noise, the model will output incorrect predictions; iii). restoration run, in which the model gives predictions with noised hidden states recovered to clean one at a time to find out which layer is related to producing factual knowledge. Following all three steps, users can find out the causal importance of all model layers in the computation graph and edit weights of those knowledge-related layers to alter the final model output.", + "bbox": [ + 212, + 236, + 787, + 448 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "There are different applications of model editing. For example, model editing can be employed on large language models (LLMs) to facilitate the interpretation of LLMs as black boxes when generating particular content or to guide LLMs in producing desired outputs [22,30]. ROME [22] edits LLMs' linear layers to modify some learned factual knowledge and make LLMs output desired content such as \"Eiffel Tower is located in the city of Rome\". Meng et al. [23] utilizes Gaussian noise to perturb predictions to achieve mass editing in a transformer.", + "bbox": [ + 212, + 449, + 789, + 554 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "As a comparison, PFEDEDIT is different from prior model editing works in two aspects. First, PFEDEDIT is a layer-wise model editing approach, which replaces the entire model layer from the local model instead of editing specific neuron stats. The advantage is that PFEDEDIT is able to perform well against potential images from various data distributions. Second, PFEDEDIT considers the global model as corrupted instead of one with Gaussian noise to reduce computation overhead.", + "bbox": [ + 212, + 555, + 787, + 660 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "3 Overview", + "text_level": 1, + "bbox": [ + 215, + 684, + 344, + 699 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "We start by describing some backgrounds in defining personalized federated learning and then present a motivating example.", + "bbox": [ + 212, + 715, + 785, + 746 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "3.1 Background: Personalized Federated Learning", + "text_level": 1, + "bbox": [ + 212, + 768, + 640, + 785 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "We describe some backgrounds of personalized federated learning. Consider the initial weights for all local models as $\\{\\theta_1, \\theta_2, \\theta_3, \\dots, \\theta_n\\}$ . Let $\\mathcal{P}_i$ be the data distribution over domain $\\mathcal{X} \\times \\mathcal{Y}$ and $l_i$ be the loss function for client $i \\in \\{1, 2, 3.., n\\}$ .", + "bbox": [ + 212, + 794, + 785, + 842 + ], + "page_idx": 3 + }, + { + "type": "page_number", + "text": "4", + "bbox": [ + 217, + 114, + 228, + 126 + ], + "page_idx": 3 + }, + { + "type": "header", + "text": "H. Yuan et al.", + "bbox": [ + 271, + 114, + 367, + 127 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "The goal of personalized federated learning is to find the set of personalized weights $\\Theta$ for each client such that it minimizes empirical loss with respect to each $\\mathcal{P}_i$ as shown in Equation ①", + "bbox": [ + 212, + 146, + 782, + 193 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\n\\Theta = \\operatorname {a r g m i n} _ {\\theta_ {i}} \\frac {1}{n} \\sum_ {i = 1} ^ {n} \\mathbb {E} _ {x, y \\sim \\mathcal {P} _ {i}} \\left[ l _ {i} \\left(\\theta_ {i}, x, y\\right) \\right] \\tag {1}\n$$\n", + "text_format": "latex", + "bbox": [ + 359, + 204, + 785, + 243 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "3.2 A Motivating Example", + "text_level": 1, + "bbox": [ + 215, + 266, + 450, + 282 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "In this part, we run through an example to motivate PFEDEDIT. We start with a challenging question. Note that previous personalized FL work [24,18,36] has noted the advantages of keeping some layers locally under non-i.i.d. scenarios, e.g., all batch normalization layers for FedBN [18]. This observation raises an intriguing question: How do we choose personalized layers to improve FL's performance with heterogeneous data?", + "bbox": [ + 212, + 294, + 542, + 445 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Our motivating example provides insight into this question. Specifically, we consider a real-world federated learning scenario where there are 10 clients collaboratively training their models on the CIFAR-100 dataset [12]. Each client is assigned 20 classes of image samples following Yuan et al. [36] to form a non-i.i.d. scenario, and we choose ResNet18 [9] as their classifier. To inspect the correlation between the personalized layers and the model utility, we pick $\\{1,3,5,7,10,13,17\\}$ as the number of personalized layers, and choose three personalized algorithms for comparison.", + "bbox": [ + 212, + 446, + 542, + 642 + ], + "page_idx": 4 + }, + { + "type": "image", + "img_path": "images/e249f10e2e43a472744fb26833ae42c07a89cbd4003c06b20d19be9b37536e72.jpg", + "image_caption": [ + "Fig. 2: A motivating example: We consider three different ways to personalize layers of local models in a federated learning scenario: random choosing, fixed choosing, and optimized layerchoosing based on the loss value. For the fixed choosing algorithm, we allow clients to keep certain numbers of batch normalization layers locally each round, and this approach becomes equivalent to FedBN [18] when the number of personalized layers gets 17." + ], + "image_footnote": [], + "bbox": [ + 560, + 329, + 781, + 472 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "The first algorithm randomly chooses certain numbers of layers from local model and keeps those layers locally. The second algorithm follows FedBN [18] to personalize batch normalization layers, and we denoted it as fixed choosing. The third personalized algorithm optimizes layer-choosing based on loss values and it considers layers that, once kept locally, yield the least loss values personalized layer and keeps them locally each round.", + "bbox": [ + 212, + 643, + 787, + 734 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Figure 2 shows the results: The optimized-personalized algorithm clearly outperforms both fixed and random personalized algorithms in all cases. This motivates our approach of using model editing to choose the optimal layers for personalization. Note that the random-personalized algorithm fails to outperform FedAvg [20] in all cases, implying that random-personalized layers do not grant utility improvement as it comes with significant randomness. The fixed-personalized algorithm fails to outperform FedAvg [20] when with a small number", + "bbox": [ + 212, + 734, + 787, + 842 + ], + "page_idx": 4 + }, + { + "type": "header", + "text": "PFEDEDIT", + "bbox": [ + 653, + 114, + 730, + 126 + ], + "page_idx": 4 + }, + { + "type": "page_number", + "text": "5", + "bbox": [ + 774, + 114, + 784, + 125 + ], + "page_idx": 4 + }, + { + "type": "image", + "img_path": "images/729b0033d059f86faab17a92dfd9c92d05c756c54c0e5cc957457802375c19f3.jpg", + "image_caption": [ + "Fig. 3: Framework of PFEDEDIT. The federated learning still works as usual. That is, first, clients train their local models and upload their model weights to the global server. Then, the global server aggregates all model weights and returns the new global model weight to clients. Then, the main change is how PFEDEDIT merges global model with the local, which has two steps: i) PFEDEDIT generates a list of candidate model editorings, and ii) PFEDEDIT locates the top-k important model layers and outputs the edited model for the next-round local training." + ], + "image_footnote": [], + "bbox": [ + 223, + 143, + 781, + 268 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "of personalized layers, and its utility is quite similar to the random-choosing algorithm. Once the number of personalized layer gets 17, which means all batch normalization layers are locally personalized, the utility improvement increases significantly and it outperforms FedAvg [18] by $13\\%$ , which implies that fixed-personalized algorithm may require all certain layers to be involved to show any utility improvement.", + "bbox": [ + 212, + 407, + 787, + 498 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "4 Method", + "text_level": 1, + "bbox": [ + 215, + 520, + 330, + 536 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "In this section, we introduce our method, PFEDEDIT, which leverages model editing to improve model accuracy for heterogeneous data distributions. Figure 3 shows the entire workflow of PFEDEDIT. The training still follows a traditional FL, i.e., each client uploads their local model to the server for aggregation, and then the server returns the aggregated model. Then, when the client receives the global model, there are mainly two steps in PFEDEDIT, model editing and locating personalized layers, to update the client model based on the global model.", + "bbox": [ + 212, + 551, + 789, + 672 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "4.1 Step I: Generating Candidate Model Editings", + "text_level": 1, + "bbox": [ + 214, + 694, + 638, + 710 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "The purpose of this step is to generate a list of possible editorings of the local model based on the global model returned from the FL server. Specifically, PFEDEDIT considers the local model \"clean model\" as it is trained on local distribution. On the opposite, it considers the global model \"corrupted model\" as in non-i.i.d. scenarios, different models bias global model weights towards their local distributions, causing it not generalizing well on each local distribution. In the model editing step, as Figure 3 shows, PFEDEDIT hypothesizes all editorings that tailor the \"corrupted model\" towards local distribution.", + "bbox": [ + 212, + 719, + 787, + 840 + ], + "page_idx": 5 + }, + { + "type": "page_number", + "text": "6", + "bbox": [ + 217, + 114, + 228, + 126 + ], + "page_idx": 5 + }, + { + "type": "header", + "text": "H. Yuan et al.", + "bbox": [ + 271, + 114, + 367, + 127 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "We show the pseudocode of this step in Algorithm $\\mathbb{I}$ . Line 5-6. For each editing, PFEDEDIT replaces one layer $\\{h_t^g | t \\in (1, T)\\}$ , where $T$ is the total number of model layers, from the global model $f_{\\theta^g}$ with the corresponding layer $h_t^l$ from the local model $f_{\\theta^l}$ . Then the edited model is defined as $S(\\theta^*) = (S(\\theta^g) \\setminus \\{h_j^g\\}_{j=0}^k) \\cup \\{h_j^l\\}_{j=0}^k$ . As the global model is equipped with the local model layer, which is optimized on local distribution, the global model is tailored towards local distribution.", + "bbox": [ + 212, + 146, + 787, + 251 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "The first editing happens after the first round (Line 1-2), right after each client receives the aggregated global model weights (Line 4) and before they start next-round local training (Line 14). The reason is that the global model starts to bias towards other distributions starting from the second round as its weights is uniformly initialized in the first round (Line 1).", + "bbox": [ + 212, + 252, + 787, + 328 + ], + "page_idx": 6 + }, + { + "type": "code", + "sub_type": "algorithm", + "code_caption": [ + "Algorithm 1 PFEDEDIT" + ], + "code_body": "Input: client i's initialized global model $\\theta^g$ , communication round $R$ , number of model layers $T$ , number of layers to be edited $k$ , local training set $D_{train}$ , subset $D_p$ , $z = (x,y) \\in D_p$ , model layer set $S(\\cdot)$ , dictionary $W$ \nOutput: Edited global model weights $\\theta^*$ for $R$ -th round \n1: $\\theta^l \\gets$ train $\\theta^g$ on $D_{train}$ \n2: $\\theta^g \\gets \\frac{1}{n} \\sum^n \\theta_i^l$ \n3: for $r$ in range $(1,R+1)$ do \n4: for $h_t^g$ in $S(\\theta^g)$ do \n5: $S(\\theta^*) \\gets (S(\\theta^g) \\backslash h_t^g) \\cup h_t^l$ \n6: $g(\\theta_i^l, \\theta_i^*, z)$ , $(f_{\\theta^*}(x), y) \\gets f_{\\theta^*}(z)$ \n7: $\\varphi \\gets (g(\\theta_i^l, \\theta_i^*, z), (f_{\\theta^*}(x), y))$ \n8: $W[h_t^g] \\gets \\varphi$ \n9: end for \n10: $W \\gets$ Lexicographic_rank(W) \n11: for $j \\in$ range $(0,k)$ do \n12: $m \\gets W_j.\\text{key}()$ \n13: $S(\\theta^*) \\gets (S(\\theta^g) \\backslash h_m^g) \\cup h_m^l$ \n14: end for \n15: if $r == R$ then \n16: return $\\theta^*$ \n17: else \n18: $\\theta^l \\gets$ train $\\theta^*$ on $D_{train}$ \n19: $\\theta^g \\gets \\frac{1}{n} \\sum^n \\theta_l^l$ \n20: end if \n21: end for", + "bbox": [ + 215, + 373, + 787, + 642 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "4.2 Step II: Locating Personalized Layers by Ranking Editings", + "text_level": 1, + "bbox": [ + 212, + 693, + 746, + 709 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "The purpose of this step is to access all possible editorings from Step I and then locate a list of personalized layers. The detailed substeps are two-fold. First, PFEDEDIT evaluates model editing's effectiveness based on a Total Effectiveness (TE) Score (Step II-a). Second, PFEDEDIT generates prediction lists based on TE score and ground-truth mask to reflect changes in probability and predicted labels (Step II-b). Lastly, PFEDEDIT ranks all prediction lists following the lexicographic order and then selects the top-k best layer editorings as personalized layers for each client based on the ranking (Step II-c).", + "bbox": [ + 212, + 719, + 787, + 840 + ], + "page_idx": 6 + }, + { + "type": "header", + "text": "PFEDEDIT", + "bbox": [ + 653, + 114, + 730, + 126 + ], + "page_idx": 6 + }, + { + "type": "page_number", + "text": "7", + "bbox": [ + 774, + 114, + 784, + 125 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Step II-a: Evaluating model editing effectiveness. This substep is to calculate a so-called Total Effectiveness (TE) score to evaluate mode editing effectiveness. Specifically, PFEDEDIT queries local model $f(\\theta^l)$ using a representative subset of the training set, denoted as $D_p$ , which maintains the same distribution as the full training set $D_{train}$ and $|D_p| = p * |D_{train}|$ , where $p \\in (0,1)$ . PFEDEDIT considers that model weights $\\theta^l$ represents local distribution as it is trained on pure local data. PFEDEDIT then queries each edited model $f(\\theta^*)$ using the same subset to see the prediction of tailored weights on local distribution. Then by comparing probability outputs from both local and edited weights on the same image samples, PFEDEDIT measures the effectiveness of that editing as well as the capability of edited layer in improving model accuracy. Specifically, when querying the edited model $f(\\theta^*)$ , PFEDEDIT collects both TE scores and ground-truth masks, Line 7 in Algorithm 1. The Total Effectiveness (TE) score, as defined below, reflects the probability change on the ground-truth class while the ground truth mask, which is a boolean list indicating whether the predicted label matches the ground truth label, reflects the predicted label change.", + "bbox": [ + 212, + 146, + 787, + 388 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Now we describe the definition of TE score. Given model weight $\\theta$ and training dataset $D$ with $c$ classes, we denote by $P(y|\\theta ,x)$ the probability (softmax) of the ground-truth class $y$ . We define the total effectiveness (TE) score of editing the model weight $\\theta$ into the new weight $\\theta^{*}$ given dataset $D$ is defined as:", + "bbox": [ + 212, + 388, + 787, + 449 + ], + "page_idx": 7 + }, + { + "type": "equation", + "text": "\n$$\ng (\\theta , \\theta^ {*}, D) = \\frac {1}{N} \\sum_ {x, y \\in D} \\frac {P (y | \\theta^ {*} , x)}{P (y | \\theta , x)} - 1 \\tag {2}\n$$\n", + "text_format": "latex", + "bbox": [ + 372, + 455, + 785, + 491 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "A positive TE score represents an increased probability on ground-truth label while a negative TE score signifies a decrease.", + "bbox": [ + 214, + 500, + 785, + 529 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Step II-b: Generating prediction lists based on TE and ground-truth mask. This sub-task is to generate prediction lists based on TE and ground-truth mask. Specifically, the TE score, together with the ground-truth mask, makes up four combinations in total, and we call the list of all four cases a prediction list, denoted by $\\varphi$ . The prediction list then serves as a reference for evaluating the editing as well as the capability of such layer in improving model's accuracy. Below we show the prediction list that contains all four combinations of TE score and ground-truth mask in Equation 3", + "bbox": [ + 212, + 540, + 787, + 662 + ], + "page_idx": 7 + }, + { + "type": "equation", + "text": "\n$$\n\\varphi = \\left\\{ \\begin{array}{l} f _ {\\theta^ {*}} (x) = y, g \\left(\\theta , \\theta^ {*}, z\\right) > 0 \\\\ f _ {\\theta^ {*}} (x) = y, g \\left(\\theta , \\theta^ {*}, z\\right) < 0 \\\\ f _ {\\theta^ {*}} (x) \\neq y, g \\left(\\theta , \\theta^ {*}, z\\right) > 0 \\\\ f _ {\\theta^ {*}} (x) \\neq y, g \\left(\\theta , \\theta^ {*}, z\\right) < 0 \\end{array} \\right. \\tag {3}\n$$\n", + "text_format": "latex", + "bbox": [ + 380, + 672, + 785, + 741 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "The top case of the prediction list $\\varphi$ shows the best editing (predicted label same as ground-truth label, TE score greater than 0) and bottom case shows the worst editing effect (predicted label different to ground-truth label, TE score less than 0). PFEDEDIT collects values for such prediction list for each editing by splitting which case each image prediction belongs to, and then ranks all prediction lists using Lexicographic order defined below.", + "bbox": [ + 212, + 750, + 787, + 840 + ], + "page_idx": 7 + }, + { + "type": "page_number", + "text": "8", + "bbox": [ + 217, + 114, + 228, + 126 + ], + "page_idx": 7 + }, + { + "type": "header", + "text": "H. Yuan et al.", + "bbox": [ + 271, + 114, + 367, + 127 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/986db9a589683faa627a6a519f630ff92043d1593067de60cd0491f5fc85530b.jpg", + "image_caption": [ + "(a) Lyme" + ], + "image_footnote": [], + "bbox": [ + 223, + 154, + 380, + 250 + ], + "page_idx": 8 + }, + { + "type": "image", + "img_path": "images/6711746bb6761d3305c801b171f34a67ad5ecf9284bbe92b164002d91cf35d71.jpg", + "image_caption": [ + "(b) CIFAR-10", + "Fig. 4: Non-i.i.d. data distribution (2 classes per client and 20 classes for CIFAR-100) among clients in Section 5.2" + ], + "image_footnote": [], + "bbox": [ + 413, + 154, + 566, + 250 + ], + "page_idx": 8 + }, + { + "type": "image", + "img_path": "images/a24b23fbca0c61f39f108c57665b4cd1985b649e3dcd5248fcb9a2fe4689a3e3.jpg", + "image_caption": [ + "(c) CIFAR-100" + ], + "image_footnote": [], + "bbox": [ + 602, + 154, + 756, + 250 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "Step II-c: Locating top- $k$ personalized layers using lexicographic order. This sub-step is to rank all prediction lists $(\\varphi)$ using lexicographic order and select the top- $k$ as the personalized layers. Here is the definition of lexicographic order, which is used to rank $\\varphi$ . That is, given two prediction lists $\\varphi_1, \\varphi_2$ for two model editorings on layer $l_1, l_2$ , respectively, say $\\varphi_1 = a_1, \\ldots, a_4$ , $\\varphi_2 = b_1, \\ldots, b_4$ , where $a$ and $b$ are possible cases from prediction lists, the order of those two lists depends on the numeric order of the cases in the first place $i$ where the two case counts differ. That is, $\\varphi_1 < \\varphi_2$ if and only if $a_i < b_i$ in the natural order of numbers.", + "bbox": [ + 212, + 324, + 784, + 445 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "Note that the advantages of using such metric here are two-folds: i) We record $\\varphi_{j}(\\theta, \\theta^{*}, z)$ by counting the number of each case and compare any two prediction lists following the Lexicographic order instead of a summation manner, i.e., $\\sum_{|D|}^{i} \\varphi_{j}(\\theta, \\theta^{*}, z)$ , so that the impact of one case is not overridden by another, even if there are insufficient instances. ii) The case order listed in Equation 3 from top to the bottom, is intuitive and aligns with accuracy metric, thus implying that the editing with highest prediction list ranking comes with a largest accuracy improvement. We also discuss and compare our metrics with other potential ones, and evaluation details are shown in Section 5.4.", + "bbox": [ + 212, + 446, + 785, + 584 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "5 Evaluation", + "text_level": 1, + "bbox": [ + 215, + 607, + 354, + 622 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "In this section, we compare PFEDEDIT with baselines in terms of the accuracy and convergence. We then perform a study on parameters used in PFEDEDIT. The evaluation on computation overhead and complexity of PFEDEDIT can be found in Appendix A.4", + "bbox": [ + 212, + 638, + 784, + 700 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "5.1 Experimental Setup", + "text_level": 1, + "bbox": [ + 214, + 723, + 426, + 739 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "We describe four aspects, i.e., setup of federated learning, datasets, models, and then baselines, for our experimental setup.", + "bbox": [ + 212, + 750, + 784, + 777 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "FL Setup. Our default settings follow state-of-the-art works [2,6] in evaluating federated learning. That is, there are 10 clients to participate in a FL scenario with 100 communication rounds. The global sampling rate is set to be 1.0 in each communication round and each client has one local training epoch. The default", + "bbox": [ + 212, + 779, + 785, + 839 + ], + "page_idx": 8 + }, + { + "type": "header", + "text": "PFEDEDIT", + "bbox": [ + 653, + 114, + 730, + 126 + ], + "page_idx": 8 + }, + { + "type": "page_number", + "text": "9", + "bbox": [ + 774, + 114, + 784, + 126 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "value $(k)$ in selecting and editing top layers is $7\\%$ , and the default representative subset ratio $(p)$ is 0.1. Both values are selected based on our parameter study (Section 5.4). By default, personalized layers are chosen each round, and the comparison with fixed personalized layer setting can be found in Appendix A.2. Datasets. We evaluate PFEDEDIT on three benchmark datasets to show its performance on different domains. Specifically, we choose i). CIFAR-10 [12], a benchmark dataset that contains 60,000 $32\\times 32$ color images in 10 classes; ii). CIFAR-100 [12], a benchmark dataset that contains 60,000 $32\\times 32$ color images in 100 classes; and iii). Lyme [35] dataset, a skin disease dataset that contains 3,027 skin disease images in 4 classes that are confusers for Lyme. We resize CIFAR-10 image samples to $64\\times 64$ , CIFAR-100 samples to $224\\times 224$ following [36], and Lyme samples to $255\\times 255$ following [35].", + "bbox": [ + 212, + 145, + 787, + 328 + ], + "page_idx": 9 + }, + { + "type": "text", + "text": "To show that PFEDEDIT can perform on various non-i.i.d. scenario, we adopt the Distributional Heterogeneity level (DH-level) from Yuan et al. [36] to split data distribution into different levels. Specifically, we first assign certain numbers of classes to each client $i$ , and then for each local class $c$ , we sample $D_{i,c} \\in (0.4, 0.6)$ from $D_c$ . Each client is assigned $\\frac{D_{i,c}}{\\sum D_{n,c}}$ of samples in class $c$ . The DH-level ranges from $0\\%$ , which represents an i.i.d. scenario where each client gets all classes of data, all the way to $100\\%$ , which means that each client only gets one class of data. (10 classes for CIFAR-100). For Lyme [35] dataset, since it only possesses four classes, we only conduct evaluations on $50\\%$ (2 classes per clients) and $100\\%$ (4 classes per clients) DH-level distribution scenarios. Unless specified otherwise, the default data distribution among clients is 2 classes per client (20 classes per client for CIFAR-100) in our evaluation. We demonstrate the default heterogeneous data distribution for each dataset in Figure 4. We also consider feature space non-i.i.d. scenario, and the evaluation, in comparison with Cd2-pfed [27] and Partialfed-fix [29], can be found in Appendix A.3. Models. For model selection, we choose ResNet18 [9] for evaluation on CIFAR 10 and CIFAR 100, and ViT [8] for Lyme. We allow ResNet18 model to have pretrained weights and allow ViT to have b16 weights that is fine-tuned on Image1k. We set the learning rate for ViT to be 1e-3 and we follow FedPer [2] to set the learning rate as 1e-2 for ResNet18.", + "bbox": [ + 212, + 329, + 787, + 635 + ], + "page_idx": 9 + }, + { + "type": "text", + "text": "Baselines. We compare PFEDEDIT with FedAvg [20], MOON [14], and other five SOTA personalized FL works as baselines, namely, FedPer [2], FedRep [4], FedBN [18], APFL [6], and DisTrans [36]. We allow each baseline to have the same data distribution and model selection as PFEDEDIT. We consider all layer normalization layers personalized for FedBN [18] when employing ViT. For other parameters, we follow their default paper settings to assign values. The accuracy comparison with more baselines (Fedfa [38], FedALA [37], Fed-LAMB [10], FedRolex [1], and HeteroFL [7]) can be found in Appendix A.1.", + "bbox": [ + 212, + 635, + 787, + 760 + ], + "page_idx": 9 + }, + { + "type": "text", + "text": "5.2 Accuracy Comparison", + "text_level": 1, + "bbox": [ + 215, + 781, + 444, + 797 + ], + "page_idx": 9 + }, + { + "type": "text", + "text": "As Table 1 shows, PFEDEDIT outperforms other state-of-the-art works on all three datasets under different DH-level data distributions. Compared to the", + "bbox": [ + 212, + 808, + 787, + 840 + ], + "page_idx": 9 + }, + { + "type": "page_number", + "text": "10", + "bbox": [ + 217, + 114, + 235, + 126 + ], + "page_idx": 9 + }, + { + "type": "header", + "text": "H. Yuan et al.", + "bbox": [ + 271, + 114, + 367, + 126 + ], + "page_idx": 9 + }, + { + "type": "text", + "text": "baseline FedAvg, the accuracy improvement ranges from 0.007 to 0.266, with an average 0.09 accuracy improvement, and the largest accuracy improvement gets to 0.156 and average improvement gets to 0.05 when compared to other personalized FL works. The accuracy gap between PFEDEDIT and other STOA personalized works in CIFAR-10 is not as huge as in other two datasets, and the reason is that we allow all works to employ models with pretrained weights, which provides a greater lower bound for model accuracy and reduces the accuracy gap between different methods.", + "bbox": [ + 212, + 146, + 787, + 265 + ], + "page_idx": 10 + }, + { + "type": "text", + "text": "We then evaluate PFEDEDIT, together with other SOTA personalized FL methods under different DH levels. According to Table II, PFEDEDIT consistently achieves superior accuracy compared to other SOTA works under all data distributions, and we have two observations here. First, some personalized FL methods, e.g., FedPer, FedBN, and FedRep, fail to outperform FedAvg under i.i.d. scenarios, which is attributed to their natures as personalized FL works: they are strategically designed to align the global model with local distributions under non-i.i.d. scenarios. If the data distribution is i.i.d., where there is a close alignment between the global model distribution and each local one, the personalized layers may block favorable updates from other clients. On the other hand, PFEDEDIT outperforms FedAvg under i.i.d. scenarios, and the reason is that PFEDEDIT allow clients to update the weights of personalized layers so that they can favor other clients when local distributions are similar. Second, the accuracy improvement brought by PFEDEDIT compared to FedAvg increases as the DH-level increases. The reason is that as the distribution becomes more non-i.i.d., the gap between the global and local model distributions widens. Personalized layers help bridge this gap, tailoring the global model towards the local distribution and improving its generalization on local data.", + "bbox": [ + 212, + 267, + 787, + 539 + ], + "page_idx": 10 + }, + { + "type": "text", + "text": "It is noteworthy to mention that PFEDEDIT gets 0.989 accuracy on CIFAR-10 when DH-level gets to $100\\%$ . The reason is that, in an extreme non-i.i.d. scenario where clients only get one class of data samples, the local model is very likely to get 1.0 testing accuracy and the model editing effectively improves the model accuracy as the local model distribution is highly matched with the local data distribution.", + "bbox": [ + 212, + 541, + 787, + 630 + ], + "page_idx": 10 + }, + { + "type": "text", + "text": "5.3 Convergence", + "text_level": 1, + "bbox": [ + 214, + 659, + 369, + 674 + ], + "page_idx": 10 + }, + { + "type": "text", + "text": "In this section we evaluate the convergence rate of PFEDEDIT. Specifically, we compare PFEDEDIT with FedAvg and we demonstrate their best accuracy on each dataset and the numbers of rounds to achieve it. As Table 2 shows, PFEDEDIT converges much faster than FedAvg on all three datasets as it only requires 18 epochs to converge on CIFAR-100, 13 epochs on CIFAR-10, and 25 epochs on Lyme. On the opposite, FedAvg takes around three times numbers of rounds to get a suboptimal accuracy on both CIFAR-10 and CIFAR-100. For Lyme, it takes 87 epochs to get accuracy above 0.7, which only takes 4 epochs for PFEDEDIT. The reason is that FedAvg does not provide convergence guarantee as it aggregates all model weights each round without having any optimization,", + "bbox": [ + 212, + 688, + 787, + 840 + ], + "page_idx": 10 + }, + { + "type": "header", + "text": "PFEDEDIT", + "bbox": [ + 653, + 114, + 730, + 126 + ], + "page_idx": 10 + }, + { + "type": "page_number", + "text": "11", + "bbox": [ + 767, + 116, + 782, + 126 + ], + "page_idx": 10 + }, + { + "type": "table", + "img_path": "images/5c902251abfca3dc290541ab29acb7ad049912acb2a05cf75550409b9c0d8140.jpg", + "table_caption": [ + "Table 1: Evaluation of PFEDEDIT with other SOTA works on different datasets." + ], + "table_footnote": [], + "table_body": "
DatasetDH-levelFedAvg 20FedPer 2FedBN 18FedRep 4DisTrans 36MOON 14APFL 6PFEDEDIT
CIFAR-100 % (i.i.d)0.9420.8900.92720.8890.9470.9220.8910.955
40%0.9230.8730.9260.8840.9440.9310.8890.953
60%0.9390.8800.9400.9200.9600.9270.9100.965
80%0.8060.8870.9450.9330.9540.9380.9150.962
100%0.8300.9020.9610.9800.9720.9400.9270.989
CIFAR-1000 % (i.i.d)0.7830.7330.7460.7420.7880.7460.7440.790
40%0.7790.7530.7760.7470.7790.7620.7530.825
60%0.7720.7480.8020.7930.8300.7680.8020.855
80%0.7370.7690.8620.8590.8870.7700.8330.890
100%0.6640.7900.9040.8800.9030.7730.8590.930
Lyme0%0.7070.7000.7330.7300.7450.6990.7220.765
50 % (i.i.d)0.7160.8080.8210.8330.8500.7680.7810.912
", + "bbox": [ + 217, + 171, + 782, + 280 + ], + "page_idx": 11 + }, + { + "type": "table", + "img_path": "images/5597e1f04f9ea2fb5ce7553d4fefed3edcc29e8db8b8d92c7a2bd5772856f87b.jpg", + "table_caption": [ + "Table 2: Best Accuracy of PFEDEdit/FedAvg and number of rounds to achieve it. (- represents not achieved within 100 rounds)" + ], + "table_footnote": [], + "table_body": "
Dataset| PFEDEDIT/FedAvg# Rounds for target accuracy (PFEDEDIT/FedAvg)
>0.6>0.7>0.75>0.8>0.85>0.9
CIFAR-100.962/0.8064/56/87/139/3311/-13/-
CIFAR-1000.890/0.7376/98/2410/-12/-18/--/-
Lyme0.912/0.7163/324/875/-5/-9/-25/-
", + "bbox": [ + 218, + 332, + 781, + 411 + ], + "page_idx": 11 + }, + { + "type": "table", + "img_path": "images/d11e0e0fc645060808fe3c62ce5b98eefc608790579fbcc39406353dda745c9a.jpg", + "table_caption": [ + "Table 3: Accuracy of PFEDEDIT with different editing ratios (in comparison with FedAvg)." + ], + "table_footnote": [], + "table_body": "
Dataset| FedAvg| PFEDEDIT
1%3%5%7%9%
CIFAR-100.8060.9330.9470.9480.9620.952
CIFAR-1000.7480.8600.8780.8850.8900.874
Lyme0.7160.8680.8800.8890.9120.910
", + "bbox": [ + 218, + 463, + 781, + 536 + ], + "page_idx": 11 + }, + { + "type": "text", + "text": "and PFEDEDIT edits the global model and tailors it towards the local distribution, which makes it faster to generalize on local distributions.", + "bbox": [ + 215, + 564, + 785, + 595 + ], + "page_idx": 11 + }, + { + "type": "text", + "text": "5.4 Parameter Study", + "text_level": 1, + "bbox": [ + 215, + 617, + 405, + 633 + ], + "page_idx": 11 + }, + { + "type": "text", + "text": "Model Editing Ratio $(k)$ PFEDEDIT allows users to pre-define $k$ , the number of layers to be edited. It then edits the top- $k$ layers, which are considered to be the most effective in improving model accuracy. In this section we set up an experiment to show the correlation between the number of modified layers and the model accuracy. Intuitively, if the number of edited layers increases, the model accuracy gets higher as greater part of the model is tailored towards local distribution. However, such increase may not hold all the time as there are layers that do not possess any learnable parameters and the model may converge.", + "bbox": [ + 212, + 643, + 787, + 763 + ], + "page_idx": 11 + }, + { + "type": "text", + "text": "We choose the ratios of edited layers to be $1\\%$ , $3\\%$ , $5\\%$ , $7\\%$ , and $9\\%$ to the total model layers and we round up the number in each case. Shown in Table 3, as the editing ratio increases, the overall testing accuracy gets higher. The accuracy improvement is greater from improving the ratio from $1\\%$ to $3\\%$ and the increasing rate gets flatter when keep improving the ratio from $3\\%$ to", + "bbox": [ + 212, + 763, + 787, + 840 + ], + "page_idx": 11 + }, + { + "type": "page_number", + "text": "12", + "bbox": [ + 217, + 114, + 235, + 126 + ], + "page_idx": 11 + }, + { + "type": "header", + "text": "H. Yuan et al.", + "bbox": [ + 271, + 114, + 367, + 127 + ], + "page_idx": 11 + }, + { + "type": "text", + "text": "$5\\%$ , or $5\\%$ to $7\\%$ . At the end, when the ratio is set from $7\\%$ to $9\\%$ , the accuracy starts to drop a little bit, implying that the model may converge on that accuracy. Compared to the baseline, PFEDEDIT improves the model accuracy from 0.13, with editing $1\\%$ layers, to 0.16, with editing $7\\%$ layers, on three datasets on average.", + "bbox": [ + 212, + 145, + 787, + 220 + ], + "page_idx": 12 + }, + { + "type": "table", + "img_path": "images/1684719a56c2f35cf4b4b44b7dc17fe5d6fd237507db8786efa3033cefbc27fa.jpg", + "table_caption": [ + "Table 4: Evaluation of PFedEdit with representative subset ratio (i.e., $p$ )." + ], + "table_footnote": [], + "table_body": "
DatasetSubser ratio
p=0.01p=0.025p=0.05p=0.08p=0.1
CIFAR-100.9510.9530.9540.9540.962
CIFAR-100-0.8750.8840.8860.890
", + "bbox": [ + 215, + 262, + 511, + 319 + ], + "page_idx": 12 + }, + { + "type": "table", + "img_path": "images/4def843f70d054d4219c6ee4ca2affd09dc28d19721a01192c0ceb8c9c83b373.jpg", + "table_caption": [ + "Table 5: Evaluation of PFedEdit using different metrics in assessing model editing" + ], + "table_footnote": [], + "table_body": "
DatasetDifferent metrics
AccuracyLossTE scorePrediction list
CIFAR-100.5400.9050.8850.962
CIFAR-1000.7050.8860.8210.890
", + "bbox": [ + 527, + 262, + 784, + 320 + ], + "page_idx": 12 + }, + { + "type": "text", + "text": "Representative Subset Ratio $(p)$ As mentioned in Section 4, PFEDEDIT queries the model using a representative subset $D_{p}$ of the training set $D_{train}$ . PFEDEDIT employs a ratio $p$ to pre-define the subset size compared to the training set, $|D_{p}| = p * |D_{train}|$ . We want to ensure that the subset is representative enough for potential testing image samples while not introducing too much computation overhead. Therefore, we conduct a study on the size of the subset versus the performance of PFEDEDIT. Intuitively, a larger size covers a broader local data distribution and leads to a more accurate layer locating and thus a higher model accuracy. However, on the other hand, taken a larger subset from the original training set reduces the size of the remaining set. We want to find an optimal size of the subset that is large enough to give a complete representation of local data distribution but also leaves enough samples for the local training. We split the local training set into a subset that takes up of $p * |D_{train}|$ image samples, and a training set that takes the rest $|D_{train}^{*}| = |D_{train}| - |D_{p}|$ image samples. To control the variable, we fix the size of $D_{train}^{*}$ , and we choose different $p$ values from $\\{0.1, 0.08, 0.05, 0.025, 0.01\\}$ to adjust the subset size. We conduct the evaluation on CIFAR-10 and CIFAR-100 [12], with 10 clients and each client is assigned 2 classes of image samples. Each client possesses 4,500 image samples as training set, and the subset size varies from 50 image samples ( $p$ value 0.01) to 500 image samples ( $p$ value 0.1). Note that the $p$ value starts from 0.025 for CIFAR-100 because a subset that is 0.01 times the size of training set does not include image samples from all classes, i.e., 50 image samples do not cover 100 classes thus the subset is not representative.", + "bbox": [ + 212, + 340, + 787, + 688 + ], + "page_idx": 12 + }, + { + "type": "text", + "text": "As Table 4 shows, the overall trend of the testing accuracy keeps increasing as $p$ value gets larger from 0.01 times the training set to 0.1 times the training set for both datasets. However, the overall accuracy improvement is minimal compared to the subset size increase, demonstrating the capability of PFEDEDIT in improving model accuracy with a small size subset. For CIFAR-10, as $p$ value increases from 0.01 to 0.08, which is an 8 times increase, the model accuracy is improved from 0.951 to 0.954, which is a $0.3\\%$ improvement. As $p$ value increases from 0.08 to 0.1, the accuracy is further improved from 0.954 to 0.962, which is a $0.8\\%$ increase. For CIFAR-100, the accuracy improvement brought by increasing $p$ value is greater than that in CIFAR-10 as the model accuracy is improved from", + "bbox": [ + 212, + 688, + 787, + 840 + ], + "page_idx": 12 + }, + { + "type": "header", + "text": "PFEDEDIT", + "bbox": [ + 653, + 114, + 730, + 126 + ], + "page_idx": 12 + }, + { + "type": "page_number", + "text": "13", + "bbox": [ + 767, + 114, + 785, + 126 + ], + "page_idx": 12 + }, + { + "type": "text", + "text": "0.875 to 0.890 when $p$ value changes from 0.025 to 0.1. Therefore, PFEDEDIT is computation-friendly as it requires only 50 image samples to get a 0.951 model accuracy and 500 image samples to achieve a 0.962 accuracy on CIFAR-10. Similarly, PFEDEDIT requires only 125 image samples as the subset to achieve a 0.875 accuracy and 500 image samples to gets a 0.890 accuracy on CIFAR-100.", + "bbox": [ + 212, + 146, + 787, + 220 + ], + "page_idx": 13 + }, + { + "type": "text", + "text": "Metrics Selection In this section, we conduct experiments to explore possible metric selections for PFEDEDIT when assessing each model editing. Besides the prediction list that demonstrates all four combinations of TE score and ground-truth mask, we also choose model accuracy, loss value, and TE score. We select cross-entropy loss as our loss metric, which is defined as $H(P,Q) = -\\sum_{x \\in \\mathcal{X}} p(x) \\log q(x)$ , given two probability distribution $P$ and $Q$ , and the TE score is introduced in Section 4.", + "bbox": [ + 212, + 243, + 787, + 348 + ], + "page_idx": 13 + }, + { + "type": "text", + "text": "As Table 5 shows, the prediction list employed by PFEDEDIT outperforms other metrics and the accuracy gap ranges from 0.057 to 0.422. Among all possible metrics, using accuracy to assess model editorings leads to the lowest accuracy, 0.540 on CIFAR-10 and 0.705 on CIFAR-100, lower than the baseline FedAvg [20]. The reason is that the effects of some editorings are subtle and they only cause the probability change but not necessarily results in label change. As the way that accuracy is computed only counts on samples that are correctly classified, small probability change that do not alter the final prediction will be considered as no improvement at all. On the other hand, the TE score reflects the probability change on ground-truth label, but it does not reflects any label change. In other words, an increase in the probability of the ground-truth label suggests a positive editing, but a decrease in probability does not necessarily indicate a negative editing. From the accuracy perspective, the edited model performs identically to the original model despite these probability changes.", + "bbox": [ + 212, + 349, + 787, + 560 + ], + "page_idx": 13 + }, + { + "type": "text", + "text": "While employing loss functions as the metric provides insights into the probability of the ground-truth class, the averaged loss value may exhibit bias if outliers have a disproportionate impact on the probability of the ground-truth class. For example, If an editing improves the model accuracy in correctly classifying two samples with a slight increase in the ground-truth class and misclassifies one sample with a notable drop in the ground-truth class, the improvement introduced by the correct classifications might be nullified by the incorrect one, which leads to a biased averaged loss value with respect to each data sample.", + "bbox": [ + 212, + 560, + 787, + 681 + ], + "page_idx": 13 + }, + { + "type": "text", + "text": "6 Conclusion", + "text_level": 1, + "bbox": [ + 215, + 703, + 359, + 719 + ], + "page_idx": 13 + }, + { + "type": "text", + "text": "In this paper, we design the first personalized federated learning framework that utilizes automated model editing as a core strategy to improve the model's utility. Our approach leverages layer-specific editing, which automatically identifies specific layers to be edited based on local data distribution, allowing for more effective adaptation of the global model to heterogeneous data distributions across various clients. Our evaluation demonstrates that PFEDEDIT significantly improves the accuracy of the global model from non-i.i.d. to i.i.d. data distribution.", + "bbox": [ + 212, + 734, + 787, + 839 + ], + "page_idx": 13 + }, + { + "type": "page_number", + "text": "14", + "bbox": [ + 217, + 114, + 235, + 126 + ], + "page_idx": 13 + }, + { + "type": "header", + "text": "H. Yuan et al.", + "bbox": [ + 271, + 114, + 367, + 127 + ], + "page_idx": 13 + }, + { + "type": "text", + "text": "Acknowledgments", + "text_level": 1, + "bbox": [ + 217, + 143, + 392, + 162 + ], + "page_idx": 14 + }, + { + "type": "text", + "text": "This work was supported in part by National Science Foundation (NSF) under grants OAC-23-19742 and Johns Hopkins University Institute for Assured Autonomy (IAA) with grants 80052272 and 80052273. The views and conclusions contained herein are those of the authors and should not be interpreted as necessarily representing the official policies or endorsements, either expressed or implied, of NSF or JHU-IAA.", + "bbox": [ + 212, + 176, + 787, + 268 + ], + "page_idx": 14 + }, + { + "type": "text", + "text": "References", + "text_level": 1, + "bbox": [ + 217, + 292, + 323, + 308 + ], + "page_idx": 14 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "1. Alam, S., Liu, L., Yan, M., Zhang, M.: Fedrolex: Model-heterogeneous federated learning with rolling sub-model extraction (2023)", + "2. Arivazhagan, M.G., Aggarwal, V., Singh, A.K., Choudhary, S.: Federated learning with personalization layers. arXiv preprint arXiv:1912.00818 (2019)", + "3. Basu, S., Zhao, N., Morariu, V., Feizi, S., Manjunatha, V.: Localizing and editing knowledge in text-to-image generative models (2023)", + "4. Collins, L., Hassani, H., Mokhtari, A., Shakkottai, S.: Exploiting shared representations for personalized federated learning. In: Proceedings of the 38th International Conference on Machine Learning (2021)", + "5. Dai, D., Dong, L., Hao, Y., Sui, Z., Chang, B., Wei, F.: Knowledge neurons in pretrained transformers. arXiv preprint arXiv:2104.08696 (2021)", + "6. Deng, Y., Kamani, M.M., Mahdavi, M.: Adaptive personalized federated learning (2020)", + "7. Diao, E., Ding, J., Tarokh, V.: Heterofl: Computation and communication efficient federated learning for heterogeneous clients (2021)", + "8. Dosovitskiy, A., Beyer, L., Kolesnikov, A., Weissenborn, D., Zhai, X., Unterthiner, T., Dehghani, M., Minderer, M., Heigold, G., Gelly, S., Uszkoreit, J., Houlsby, N.: An image is worth 16x16 words: Transformers for image recognition at scale (2021)", + "9. He, K., Zhang, X., Ren, S., Sun, J.: Deep residual learning for image recognition. In: Proceedings of the IEEE conference on computer vision and pattern recognition. pp. 770-778 (2016)", + "0. Karimi, B., Li, P., Li, X.: Layer-wise and dimension-wise locally adaptive federated learning (2022)", + "1. Karimireddy, S.P., Kale, S., Mohri, M., Reddi, S., Stich, S., Suresh, A.T.: Scaffold: Stochastic controlled averaging for federated learning. In: International conference on machine learning (2020)", + "2. Krizhevsky, A.: Learning multiple layers of features from tiny images. Tech. rep. (2009)", + "3. Li, L., Fan, Y., Tse, M., Lin, K.Y.: A review of applications in federated learning. Computers & Industrial Engineering (2020)", + "4. Li, Q., He, B., Song, D.: Model-contrastive federated learning. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (2021)", + "5. Li, T., Sahu, A.K., Talwalkar, A., Smith, V.: Federated learning: Challenges, methods, and future directions. IEEE signal processing magazine (2020)", + "6. Li, T., Sahu, A.K., Zaheer, M., Sanjabi, M., Talwalkar, A., Smith, V.: Federated optimization in heterogeneous networks. Proceedings of Machine learning and systems 2, 429-450 (2020)" + ], + "bbox": [ + 225, + 324, + 784, + 839 + ], + "page_idx": 14 + }, + { + "type": "header", + "text": "PFEDEDIT", + "bbox": [ + 653, + 114, + 730, + 126 + ], + "page_idx": 14 + }, + { + "type": "page_number", + "text": "15", + "bbox": [ + 767, + 114, + 785, + 126 + ], + "page_idx": 14 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "17. Li, X., Li, S., Song, S., Yang, J., Ma, J., Yu, J.: Pmet: Precise model editing in a transformer. arXiv preprint arXiv:2308.08742 (2023)", + "18. Li, X., Jiang, M., Zhang, X., Kamp, M., Dou, Q.: Fed{bn}: Federated learning on non-{iid} features via local batch normalization. In: International Conference on Learning Representations (2021), https://openreview.net/pdf?id=6YEQUnOQICG", + "19. Liu, Q., Chen, C., Qin, J., Dou, Q., Heng, P.A.: Feddg: Federated domain generalization on medical image segmentation via episodic learning in continuous frequency space. The IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR) (2021)", + "20. McMahan, B., Moore, E., Ramage, D., Hampson, S., y Arcas, B.A.: Communication-efficient learning of deep networks from decentralized data. In: Artificial intelligence and statistics (2017)", + "21. Meng, K., Bau, D., Andonian, A., Belinkov, Y.: Locating and editing factual associations in GPT. Advances in Neural Information Processing Systems (2022)", + "22. Meng, K., Bau, D., Andonian, A., Belinkov, Y.: Locating and editing factual associations in GPT. Advances in Neural Information Processing Systems 35 (2022)", + "23. Meng, K., Sen Sharma, A., Andonian, A., Belinkov, Y., Bau, D.: Mass editing memory in a transformer. The Eleventh International Conference on Learning Representations (ICLR) (2023)", + "24. Nguyen, D.C., Ding, M., Pathirana, P.N., Seneviratne, A., Li, J., Vincent Poor, H.: Federated learning for internet of things: A comprehensive survey. IEEE Communications Surveys & Tutorials (2021)", + "25. Nguyen, D.C., Ding, M., Pham, Q.V., Pathirana, P.N., Le, L.B., Seneviratne, A., Li, J., Niyato, D., Poor, H.V.: Federated learning meets blockchain in edge computing: Opportunities and challenges. IEEE Internet of Things Journal (2021)", + "26. Nori, H., Jenkins, S., Koch, P., Caruana, R.: Interpretml: A unified framework for machine learning interpretability. arXiv preprint arXiv:1909.09223 (2019)", + "27. Shen, Y., Zhou, Y., Yu, L.: $\\mathrm{Cd^{2}}$ -pfed: Cyclic distillation-guided channel decoupling for model personalization in federated learning (2022)", + "28. Shokri, R., Shmatikov, V.: Privacy-preserving deep learning. In: Proceedings of the 22nd ACM SIGSAC conference on computer and communications security (2015)", + "29. Sun, B., Huo, H., YANG, Y., Bai, B.: Partialfed: Cross-domain personalized federated learning via partial initialization. In: Advances in Neural Information Processing Systems (2021)", + "30. Wang, P., Zhang, N., Xie, X., Yao, Y., Tian, B., Wang, M., Xi, Z., Cheng, S., Liu, K., Zheng, G., et al.: Easilyedit: An easy-to-use knowledge editing framework for large language models. arXiv preprint arXiv:2308.07269 (2023)", + "31. Yang, Q., Liu, Y., Chen, T., Tong, Y.: Federated machine learning: Concept and applications. ACM Transactions on Intelligent Systems and Technology (TIST) (2019)", + "32. Yang, Y., Hui, B., Yuan, H., Gong, N., Cao, Y.: PrivateFL: Accurate, differentially private federated learning via personalized data transformation. In: 32nd USENIX Security Symposium (USENIX Security 23). pp. 1595-1612 (2023)", + "33. Yang, Y., Yuan, H., Hui, B., Gong, N., Fendley, N., Burlina, P., Cao, Y.: Fortifying federated learning against membership inference attacks via client-level input perturbation. In: 2023 53rd Annual IEEE/IFIP International Conference on Dependable Systems and Networks (DSN) (2023)", + "34. Yang, Z., Zhang, Y., Zheng, Y., Tian, X., Peng, H., Liu, T., Han, B.: Fedfed: Feature distillation against data heterogeneity in federated learning. In: Thirty-seventh Conference on Neural Information Processing Systems (2023)" + ], + "bbox": [ + 215, + 147, + 785, + 839 + ], + "page_idx": 15 + }, + { + "type": "page_number", + "text": "16", + "bbox": [ + 217, + 114, + 235, + 126 + ], + "page_idx": 15 + }, + { + "type": "header", + "text": "H. Yuan et al.", + "bbox": [ + 271, + 114, + 367, + 127 + ], + "page_idx": 15 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "35. Yuan, H., Aucott, J., Hadzic, A., Paul, W., Villegas de Flores, M., Mathew, P., Burlina, P., Cao, Y.: Edgemixup: Embarrassingly simple data alteration to improve Lyme disease lesion segmentation and diagnosis fairness. In: Medical Image Computing and Computer Assisted Intervention - MICCAI 2023 (2023)", + "36. Yuan, H., Hui, B., Yang, Y., Burlina, P., Gong, N.Z., Cao, Y.: Addressing heterogeneity in federated learning via distributional transformation. In: Computer Vision - ECCV 2022: 17th European Conference, Tel Aviv, Israel, October 23-27, 2022, Proceedings, Part XXXVIII. p. 179-195 (2022)", + "37. Zhang, J., Hua, Y., Wang, H., Song, T., Xue, Z., Ma, R., Guan, H.: Fedala: Adaptive local aggregation for personalized federated learning. In: Proceedings of the AAAI Conference on Artificial Intelligence (2023)", + "38. Zhou, T., Konukoglu, E.: FedFA: Federated feature augmentation. In: The Eleventh International Conference on Learning Representations (ICLR) (2023)" + ], + "bbox": [ + 212, + 146, + 787, + 327 + ], + "page_idx": 16 + }, + { + "type": "header", + "text": "PFEDEDIT", + "bbox": [ + 653, + 114, + 730, + 126 + ], + "page_idx": 16 + }, + { + "type": "page_number", + "text": "17", + "bbox": [ + 767, + 114, + 785, + 126 + ], + "page_idx": 16 + } +] \ No newline at end of file diff --git a/2024/PFedEdit_ Personalized Federated Learning via Automated Model Editing/984c6b27-188f-4dc6-b40f-18bc786fe91d_model.json b/2024/PFedEdit_ Personalized Federated Learning via Automated Model Editing/984c6b27-188f-4dc6-b40f-18bc786fe91d_model.json new file mode 100644 index 0000000000000000000000000000000000000000..5406fc8d47ed63b4b21b9dd1f6a60b90552e86bb --- /dev/null +++ b/2024/PFedEdit_ Personalized Federated Learning via Automated Model Editing/984c6b27-188f-4dc6-b40f-18bc786fe91d_model.json @@ -0,0 +1,2049 @@ +[ + [ + { + "type": "title", + "bbox": [ + 0.227, + 0.141, + 0.78, + 0.187 + ], + "angle": 0, + "content": "PFEDEDIT: Personalized Federated Learning via Automated Model Editing" + }, + { + "type": "text", + "bbox": [ + 0.257, + 0.212, + 0.747, + 0.243 + ], + "angle": 0, + "content": "Haolin Yuan\\(^{1}\\), William Paul\\(^{2}\\), John Aucott\\(^{3}\\), Philippe Burlina\\(^{2}\\), and Yinzhi Cao\\(^{1}\\)" + }, + { + "type": "text", + "bbox": [ + 0.297, + 0.255, + 0.705, + 0.269 + ], + "angle": 0, + "content": "1 Johns Hopkins University {hyuan4, yinzhi.cao}@jhu.edu" + }, + { + "type": "text", + "bbox": [ + 0.318, + 0.27, + 0.684, + 0.285 + ], + "angle": 0, + "content": "2 Johns Hopkins Applied Physics Lab {william.paul," + }, + { + "type": "text", + "bbox": [ + 0.392, + 0.285, + 0.612, + 0.298 + ], + "angle": 0, + "content": "philippe.burlina}@jhuapl.edu" + }, + { + "type": "text", + "bbox": [ + 0.275, + 0.297, + 0.727, + 0.312 + ], + "angle": 0, + "content": "3 Johns Hopkins University School of Medicine jaucott2@jhmi.edu" + }, + { + "type": "text", + "bbox": [ + 0.262, + 0.352, + 0.744, + 0.63 + ], + "angle": 0, + "content": "Abstract. Federated learning (FL) allows clients to train a deep learning model collaboratively while maintaining their private data locally. One challenging problem facing FL is that the model utility drops significantly once the data distribution gets heterogeneous, or non-i.i.d, among clients. A promising solution is to personalize models for each client, e.g., keeping some layers locally without aggregation, which is thus called personalized FL. However, previous personalized FL often suffer from suboptimal utility because their choice of layer personalization is based on empirical knowledge and fixed for different datasets and distributions. In this work, we design PFEDEDIT, the first federated learning framework that leverages automated model editing to optimize the choice of personalization layers and improve model utility under a variety of data distributions including non-i.i.d. The high-level idea of PFEDEDIT is to assess the effectiveness of every global model layer in improving model utility on local data distribution once edited, and then to apply edits on the top-\\(k\\) most effective layers. Our evaluation shows that PFEDEDIT outperforms six state-of-the-art approaches on three benchmark datasets by \\(6\\%\\) on the model's performance on average, with the largest accuracy improvement being \\(26.6\\%\\). PFEDEDIT is open-source and available at this repository: https://github.com/Haolin-Yuan/PFedEdit" + }, + { + "type": "title", + "bbox": [ + 0.217, + 0.657, + 0.377, + 0.673 + ], + "angle": 0, + "content": "1 Introduction" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.689, + 0.79, + 0.841 + ], + "angle": 0, + "content": "Federated learning (FL) [13,15,20,31,33] has emerged as a widely used training algorithm of deep learning, especially in the image domain, due to its practical applications in real-world scenarios and its impressive performance. A federated learning training scenario entails multiple local clients and a global server. Each client trains their local model using their private image data. After performing local training, each client updates their local model weights to the global server for aggregation with other clients' model weights. Federated learning allows clients to train a collaborative model that performs well on multi-class tasks—even if each client only possesses a few classes of images—via only sharing models but not local training images." + } + ], + [ + { + "type": "page_number", + "bbox": [ + 0.218, + 0.116, + 0.23, + 0.127 + ], + "angle": 0, + "content": "2" + }, + { + "type": "header", + "bbox": [ + 0.272, + 0.115, + 0.368, + 0.127 + ], + "angle": 0, + "content": "H. Yuan et al." + }, + { + "type": "image", + "bbox": [ + 0.219, + 0.145, + 0.788, + 0.237 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.215, + 0.249, + 0.788, + 0.278 + ], + "angle": 0, + "content": "Fig. 1: High-level idea of PFEDEDIT (PFEDEDIT edits the global model returned from FL server to fit the local client's distribution.)" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.317, + 0.789, + 0.468 + ], + "angle": 0, + "content": "As the goal of federated learning is to learn a global model for all devices cooperatively, if the data distributions among clients are similar, the global model is supposed to perform better than local models that are solely trained on local data. However, once the data distribution becomes skewed, or non-i.i.d., e.g., some clients may possess image samples of only one or two classes or some possess many more image samples from one class than others, the global model becomes ineffective to be employed as local models. The reason is that the local data distribution at one client could be dramatically different from others, thus causing local model weights to adversely affect each other during the model aggregation." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.473, + 0.789, + 0.702 + ], + "angle": 0, + "content": "Multiple works [11,14,16] have been proposed to improve the FL model accuracy under the non-i.i.d. distribution by optimizing local models. They either regularize local models to maximize the agreement between clients [14], or reduce clients' weights drift from the server end [11,16] for a better global model that generalizes on each local distribution. While such works improved model's accuracy, the model is still uniform across different clients, and cannot be tailored based on different clients' data distribution. Therefore, one natural step is a so-called personalized FL [2,4,18,36], which enables each user to have personalized layers that help customize local models to their data distributions. Those personalized layers are retained locally each communication round and are not updated from the global model, thereby preventing being aggregated with others and avoiding bias towards other clients' data distributions. The choices of those personalized layers are various but are all pre-defined before the training, e.g., it can be linear layer [2], batch normalization layer [4], or specially designed layer [36]." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.705, + 0.789, + 0.843 + ], + "angle": 0, + "content": "However, there are still two challenges that remain for personalized FLs despite the accuracy improvement. First, the personalized layers chosen in previous personalized FL works [2,4,18,36] are inspired by empirical findings and fixed, implying that those choices may not be optimal when given different data distributions or datasets. That is, while the weights of those layers might be personalized, the locations and types of those layers are not, e.g., either linear or batch normalization layers in certain locations of a learning model. Second, previous personalized works are limited in their application scenarios as they assume certain model structure [18], which may not be used for certain clients." + } + ], + [ + { + "type": "header", + "bbox": [ + 0.655, + 0.115, + 0.732, + 0.127 + ], + "angle": 0, + "content": "PFEDEDIT" + }, + { + "type": "page_number", + "bbox": [ + 0.775, + 0.117, + 0.785, + 0.127 + ], + "angle": 0, + "content": "3" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.147, + 0.788, + 0.373 + ], + "angle": 0, + "content": "In this paper, to address these two challenges, we design a framework that automatically locates and chooses layers to be personalized. We propose the first personalized federated learning framework, called PFEDEDIT, which leverages automated model editing [3,5,21,23] to optimize the layer choosing and improve model utility under different settings including non-i.i.d. The high-level idea of PFEDEDIT, as shown in Figure 1, is to look for model layers that contain knowledge about local distribution that benefits clients the most, and by editing those layers, PFEDEDIT mitigates the negative influence brought by other clients, thus tailoring the global model weights towards the local model weights in each round. Specifically, PFEDEDIT assesses which layers of the global model are the most effective in maintaining the local distribution and then marks them as target layers. PFEDEDIT then edits the top-\\(k\\) target layers to tailor the global model weights closer to local models thus improving the model utility. PFEDEDIT does not make any assumptions on model selection and thus is model-structure agnostics." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.374, + 0.814, + 0.464 + ], + "angle": 0, + "content": "We performed extensive experiments on three benchmark datasets, namely CIFAR-10 [12], CIFAR-100 [12], and Lyme dataset [35], and we compared PFEDEDIT with FedAvg [20] and other six state-of-the-art personalized federated learning approaches. Our experimental results demonstrate that PFEDEDIT outperforms all SOTA personalized federated learning works on all the datasets, with an average \\(6\\%\\) accuracy improvement." + }, + { + "type": "text", + "bbox": [ + 0.239, + 0.465, + 0.568, + 0.479 + ], + "angle": 0, + "content": "In summary, our contributions are as follows:" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.482, + 0.785, + 0.511 + ], + "angle": 0, + "content": "- We design PFEDEDIT, the first personalized federated learning framework that optimizes personalized layer selections for each client using model editing." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.512, + 0.785, + 0.54 + ], + "angle": 0, + "content": "- PFEDEDIT is adaptable to all model structures because it does not impose any assumption on model selection." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.543, + 0.785, + 0.571 + ], + "angle": 0, + "content": "- We conduct extensive experiments to show that PFEDEDIT outperforms other SOTA personalized federated works on different benchmark datasets." + }, + { + "type": "list", + "bbox": [ + 0.214, + 0.482, + 0.785, + 0.571 + ], + "angle": 0, + "content": null + }, + { + "type": "title", + "bbox": [ + 0.216, + 0.597, + 0.388, + 0.613 + ], + "angle": 0, + "content": "2 Related Work" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.628, + 0.788, + 0.844 + ], + "angle": 0, + "content": "Federated Learning. Federated learning [13,15,20,31] (FL) is first proposed to predict user's text input on portable devices and it has been widely deployed in other industries, such as medical imaging [19], Internet of Things (IoT) [24], and block-chain [25]. It allows multiple clients to collaboratively train a model that can perform on samples from any clients' data distribution. Two common approaches in federated learning (FL) are FedSGD [28] and FedAvg [20]. FedSGD optimizes global weights by back-propagating each local gradient, while FedAvg allows for multiple local training epochs and client-side optimization. In FedAvg, the server's role is limited to averaging and distributing the aggregated weights. While FedAvg reduces communication costs compared to FedSGD, it lacks a convergence guarantee, and heterogeneous data distribution among clients can significantly reduce the global model's utility. Therefore, various works [2,4,14,18,32,36] have been proposed to solve such utility degradation caused by data heterogeneity. One direction [14] is to optimize local" + } + ], + [ + { + "type": "page_number", + "bbox": [ + 0.218, + 0.116, + 0.23, + 0.127 + ], + "angle": 0, + "content": "4" + }, + { + "type": "header", + "bbox": [ + 0.272, + 0.115, + 0.368, + 0.128 + ], + "angle": 0, + "content": "H. Yuan et al." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.147, + 0.788, + 0.224 + ], + "angle": 0, + "content": "model either via personalized objective functions; yet, another, called personalized FL [2,4,18,36], allows clients to have personalized model layers that can be kept locally. As a comparison, PFEDEDIT is the first work that optimizes and personalizes the choice of model layers with further improved utilities of FL, especially under non-i.i.d. settings." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.237, + 0.789, + 0.449 + ], + "angle": 0, + "content": "Model Editing. Model editing 3.5.17.23.26 involves adding, removing, or adjusting layers and parameters of a model architecture to improve its performance or adapt it to new tasks. Specifically, let \\( h_i \\) be the \\( i \\)-th hidden states of any model, \\( \\{h_l | l \\in [1, L]\\} \\), where \\( L \\) is the number of layers in model, model editing allows users to find each hidden state's contribution towards a specific prediction. It mainly contains three runs: i). clean run, in which the statistics of each state are recorded when fed a factual prompt \\( x \\); ii). corrupted run, in which each hidden layer is added Gaussian noise such that \\( h_l \\coloneqq h_l + \\epsilon \\). Because the hidden layer is obfuscated by the added noise, the model will output incorrect predictions; iii). restoration run, in which the model gives predictions with noised hidden states recovered to clean one at a time to find out which layer is related to producing factual knowledge. Following all three steps, users can find out the causal importance of all model layers in the computation graph and edit weights of those knowledge-related layers to alter the final model output." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.45, + 0.79, + 0.555 + ], + "angle": 0, + "content": "There are different applications of model editing. For example, model editing can be employed on large language models (LLMs) to facilitate the interpretation of LLMs as black boxes when generating particular content or to guide LLMs in producing desired outputs [22,30]. ROME [22] edits LLMs' linear layers to modify some learned factual knowledge and make LLMs output desired content such as \"Eiffel Tower is located in the city of Rome\". Meng et al. [23] utilizes Gaussian noise to perturb predictions to achieve mass editing in a transformer." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.556, + 0.789, + 0.661 + ], + "angle": 0, + "content": "As a comparison, PFEDEDIT is different from prior model editing works in two aspects. First, PFEDEDIT is a layer-wise model editing approach, which replaces the entire model layer from the local model instead of editing specific neuron stats. The advantage is that PFEDEDIT is able to perform well against potential images from various data distributions. Second, PFEDEDIT considers the global model as corrupted instead of one with Gaussian noise to reduce computation overhead." + }, + { + "type": "title", + "bbox": [ + 0.216, + 0.685, + 0.345, + 0.7 + ], + "angle": 0, + "content": "3 Overview" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.716, + 0.787, + 0.747 + ], + "angle": 0, + "content": "We start by describing some backgrounds in defining personalized federated learning and then present a motivating example." + }, + { + "type": "title", + "bbox": [ + 0.214, + 0.769, + 0.642, + 0.786 + ], + "angle": 0, + "content": "3.1 Background: Personalized Federated Learning" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.795, + 0.787, + 0.843 + ], + "angle": 0, + "content": "We describe some backgrounds of personalized federated learning. Consider the initial weights for all local models as \\(\\{\\theta_1, \\theta_2, \\theta_3, \\dots, \\theta_n\\}\\). Let \\(\\mathcal{P}_i\\) be the data distribution over domain \\(\\mathcal{X} \\times \\mathcal{Y}\\) and \\(l_i\\) be the loss function for client \\(i \\in \\{1, 2, 3.., n\\}\\)." + } + ], + [ + { + "type": "header", + "bbox": [ + 0.655, + 0.115, + 0.732, + 0.127 + ], + "angle": 0, + "content": "PFEDEDIT" + }, + { + "type": "page_number", + "bbox": [ + 0.775, + 0.116, + 0.785, + 0.126 + ], + "angle": 0, + "content": "5" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.147, + 0.784, + 0.194 + ], + "angle": 0, + "content": "The goal of personalized federated learning is to find the set of personalized weights \\(\\Theta\\) for each client such that it minimizes empirical loss with respect to each \\(\\mathcal{P}_i\\) as shown in Equation ①" + }, + { + "type": "equation", + "bbox": [ + 0.36, + 0.205, + 0.786, + 0.244 + ], + "angle": 0, + "content": "\\[\n\\Theta = \\operatorname {a r g m i n} _ {\\theta_ {i}} \\frac {1}{n} \\sum_ {i = 1} ^ {n} \\mathbb {E} _ {x, y \\sim \\mathcal {P} _ {i}} \\left[ l _ {i} \\left(\\theta_ {i}, x, y\\right) \\right] \\tag {1}\n\\]" + }, + { + "type": "title", + "bbox": [ + 0.216, + 0.267, + 0.452, + 0.284 + ], + "angle": 0, + "content": "3.2 A Motivating Example" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.295, + 0.544, + 0.446 + ], + "angle": 0, + "content": "In this part, we run through an example to motivate PFEDEDIT. We start with a challenging question. Note that previous personalized FL work [24,18,36] has noted the advantages of keeping some layers locally under non-i.i.d. scenarios, e.g., all batch normalization layers for FedBN [18]. This observation raises an intriguing question: How do we choose personalized layers to improve FL's performance with heterogeneous data?" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.447, + 0.544, + 0.643 + ], + "angle": 0, + "content": "Our motivating example provides insight into this question. Specifically, we consider a real-world federated learning scenario where there are 10 clients collaboratively training their models on the CIFAR-100 dataset [12]. Each client is assigned 20 classes of image samples following Yuan et al. [36] to form a non-i.i.d. scenario, and we choose ResNet18 [9] as their classifier. To inspect the correlation between the personalized layers and the model utility, we pick \\(\\{1,3,5,7,10,13,17\\}\\) as the number of personalized layers, and choose three personalized algorithms for comparison." + }, + { + "type": "image", + "bbox": [ + 0.561, + 0.33, + 0.782, + 0.473 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.554, + 0.486, + 0.788, + 0.61 + ], + "angle": 0, + "content": "Fig. 2: A motivating example: We consider three different ways to personalize layers of local models in a federated learning scenario: random choosing, fixed choosing, and optimized layerchoosing based on the loss value. For the fixed choosing algorithm, we allow clients to keep certain numbers of batch normalization layers locally each round, and this approach becomes equivalent to FedBN [18] when the number of personalized layers gets 17." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.644, + 0.788, + 0.735 + ], + "angle": 0, + "content": "The first algorithm randomly chooses certain numbers of layers from local model and keeps those layers locally. The second algorithm follows FedBN [18] to personalize batch normalization layers, and we denoted it as fixed choosing. The third personalized algorithm optimizes layer-choosing based on loss values and it considers layers that, once kept locally, yield the least loss values personalized layer and keeps them locally each round." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.735, + 0.788, + 0.843 + ], + "angle": 0, + "content": "Figure 2 shows the results: The optimized-personalized algorithm clearly outperforms both fixed and random personalized algorithms in all cases. This motivates our approach of using model editing to choose the optimal layers for personalization. Note that the random-personalized algorithm fails to outperform FedAvg [20] in all cases, implying that random-personalized layers do not grant utility improvement as it comes with significant randomness. The fixed-personalized algorithm fails to outperform FedAvg [20] when with a small number" + } + ], + [ + { + "type": "page_number", + "bbox": [ + 0.218, + 0.116, + 0.23, + 0.127 + ], + "angle": 0, + "content": "6" + }, + { + "type": "header", + "bbox": [ + 0.272, + 0.115, + 0.369, + 0.128 + ], + "angle": 0, + "content": "H. Yuan et al." + }, + { + "type": "image", + "bbox": [ + 0.224, + 0.144, + 0.782, + 0.27 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.214, + 0.281, + 0.789, + 0.379 + ], + "angle": 0, + "content": "Fig. 3: Framework of PFEDEDIT. The federated learning still works as usual. That is, first, clients train their local models and upload their model weights to the global server. Then, the global server aggregates all model weights and returns the new global model weight to clients. Then, the main change is how PFEDEDIT merges global model with the local, which has two steps: i) PFEDEDIT generates a list of candidate model editorings, and ii) PFEDEDIT locates the top-k important model layers and outputs the edited model for the next-round local training." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.408, + 0.789, + 0.499 + ], + "angle": 0, + "content": "of personalized layers, and its utility is quite similar to the random-choosing algorithm. Once the number of personalized layer gets 17, which means all batch normalization layers are locally personalized, the utility improvement increases significantly and it outperforms FedAvg [18] by \\(13\\%\\), which implies that fixed-personalized algorithm may require all certain layers to be involved to show any utility improvement." + }, + { + "type": "title", + "bbox": [ + 0.216, + 0.521, + 0.331, + 0.537 + ], + "angle": 0, + "content": "4 Method" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.552, + 0.79, + 0.673 + ], + "angle": 0, + "content": "In this section, we introduce our method, PFEDEDIT, which leverages model editing to improve model accuracy for heterogeneous data distributions. Figure 3 shows the entire workflow of PFEDEDIT. The training still follows a traditional FL, i.e., each client uploads their local model to the server for aggregation, and then the server returns the aggregated model. Then, when the client receives the global model, there are mainly two steps in PFEDEDIT, model editing and locating personalized layers, to update the client model based on the global model." + }, + { + "type": "title", + "bbox": [ + 0.215, + 0.695, + 0.64, + 0.712 + ], + "angle": 0, + "content": "4.1 Step I: Generating Candidate Model Editings" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.72, + 0.789, + 0.841 + ], + "angle": 0, + "content": "The purpose of this step is to generate a list of possible editorings of the local model based on the global model returned from the FL server. Specifically, PFEDEDIT considers the local model \"clean model\" as it is trained on local distribution. On the opposite, it considers the global model \"corrupted model\" as in non-i.i.d. scenarios, different models bias global model weights towards their local distributions, causing it not generalizing well on each local distribution. In the model editing step, as Figure 3 shows, PFEDEDIT hypothesizes all editorings that tailor the \"corrupted model\" towards local distribution." + } + ], + [ + { + "type": "header", + "bbox": [ + 0.655, + 0.115, + 0.732, + 0.127 + ], + "angle": 0, + "content": "PFEDEDIT" + }, + { + "type": "page_number", + "bbox": [ + 0.775, + 0.116, + 0.785, + 0.126 + ], + "angle": 0, + "content": "7" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.147, + 0.788, + 0.252 + ], + "angle": 0, + "content": "We show the pseudocode of this step in Algorithm \\(\\mathbb{I}\\). Line 5-6. For each editing, PFEDEDIT replaces one layer \\(\\{h_t^g | t \\in (1, T)\\}\\), where \\(T\\) is the total number of model layers, from the global model \\(f_{\\theta^g}\\) with the corresponding layer \\(h_t^l\\) from the local model \\(f_{\\theta^l}\\). Then the edited model is defined as \\(S(\\theta^*) = (S(\\theta^g) \\setminus \\{h_j^g\\}_{j=0}^k) \\cup \\{h_j^l\\}_{j=0}^k\\). As the global model is equipped with the local model layer, which is optimized on local distribution, the global model is tailored towards local distribution." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.253, + 0.788, + 0.329 + ], + "angle": 0, + "content": "The first editing happens after the first round (Line 1-2), right after each client receives the aggregated global model weights (Line 4) and before they start next-round local training (Line 14). The reason is that the global model starts to bias towards other distributions starting from the second round as its weights is uniformly initialized in the first round (Line 1)." + }, + { + "type": "code_caption", + "bbox": [ + 0.218, + 0.357, + 0.406, + 0.371 + ], + "angle": 0, + "content": "Algorithm 1 PFEDEDIT" + }, + { + "type": "algorithm", + "bbox": [ + 0.217, + 0.374, + 0.789, + 0.643 + ], + "angle": 0, + "content": "Input: client i's initialized global model \\(\\theta^g\\), communication round \\(R\\), number of model layers \\(T\\), number of layers to be edited \\(k\\), local training set \\(D_{train}\\), subset \\(D_p\\), \\(z = (x,y) \\in D_p\\), model layer set \\(S(\\cdot)\\), dictionary \\(W\\) \nOutput: Edited global model weights \\(\\theta^*\\) for \\(R\\)-th round \n1: \\(\\theta^l \\gets\\) train \\(\\theta^g\\) on \\(D_{train}\\) \n2: \\(\\theta^g \\gets \\frac{1}{n} \\sum^n \\theta_i^l\\) \n3: for \\(r\\) in range \\((1,R+1)\\) do \n4: for \\(h_t^g\\) in \\(S(\\theta^g)\\) do \n5: \\(S(\\theta^*) \\gets (S(\\theta^g) \\backslash h_t^g) \\cup h_t^l\\) \n6: \\(g(\\theta_i^l, \\theta_i^*, z)\\), \\((f_{\\theta^*}(x), y) \\gets f_{\\theta^*}(z)\\) \n7: \\(\\varphi \\gets (g(\\theta_i^l, \\theta_i^*, z), (f_{\\theta^*}(x), y))\\) \n8: \\(W[h_t^g] \\gets \\varphi\\) \n9: end for \n10: \\(W \\gets\\) Lexicographic_rank(W) \n11: for \\(j \\in\\) range \\((0,k)\\) do \n12: \\(m \\gets W_j.\\text{key}()\\) \n13: \\(S(\\theta^*) \\gets (S(\\theta^g) \\backslash h_m^g) \\cup h_m^l\\) \n14: end for \n15: if \\(r == R\\) then \n16: return \\(\\theta^*\\) \n17: else \n18: \\(\\theta^l \\gets\\) train \\(\\theta^*\\) on \\(D_{train}\\) \n19: \\(\\theta^g \\gets \\frac{1}{n} \\sum^n \\theta_l^l\\) \n20: end if \n21: end for" + }, + { + "type": "title", + "bbox": [ + 0.214, + 0.694, + 0.747, + 0.71 + ], + "angle": 0, + "content": "4.2 Step II: Locating Personalized Layers by Ranking Editings" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.72, + 0.788, + 0.841 + ], + "angle": 0, + "content": "The purpose of this step is to access all possible editorings from Step I and then locate a list of personalized layers. The detailed substeps are two-fold. First, PFEDEDIT evaluates model editing's effectiveness based on a Total Effectiveness (TE) Score (Step II-a). Second, PFEDEDIT generates prediction lists based on TE score and ground-truth mask to reflect changes in probability and predicted labels (Step II-b). Lastly, PFEDEDIT ranks all prediction lists following the lexicographic order and then selects the top-k best layer editorings as personalized layers for each client based on the ranking (Step II-c)." + } + ], + [ + { + "type": "page_number", + "bbox": [ + 0.218, + 0.116, + 0.23, + 0.127 + ], + "angle": 0, + "content": "8" + }, + { + "type": "header", + "bbox": [ + 0.272, + 0.115, + 0.368, + 0.128 + ], + "angle": 0, + "content": "H. Yuan et al." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.147, + 0.788, + 0.389 + ], + "angle": 0, + "content": "Step II-a: Evaluating model editing effectiveness. This substep is to calculate a so-called Total Effectiveness (TE) score to evaluate mode editing effectiveness. Specifically, PFEDEDIT queries local model \\( f(\\theta^l) \\) using a representative subset of the training set, denoted as \\( D_p \\), which maintains the same distribution as the full training set \\( D_{train} \\) and \\( |D_p| = p * |D_{train}| \\), where \\( p \\in (0,1) \\). PFEDEDIT considers that model weights \\( \\theta^l \\) represents local distribution as it is trained on pure local data. PFEDEDIT then queries each edited model \\( f(\\theta^*) \\) using the same subset to see the prediction of tailored weights on local distribution. Then by comparing probability outputs from both local and edited weights on the same image samples, PFEDEDIT measures the effectiveness of that editing as well as the capability of edited layer in improving model accuracy. Specifically, when querying the edited model \\( f(\\theta^*) \\), PFEDEDIT collects both TE scores and ground-truth masks, Line 7 in Algorithm 1. The Total Effectiveness (TE) score, as defined below, reflects the probability change on the ground-truth class while the ground truth mask, which is a boolean list indicating whether the predicted label matches the ground truth label, reflects the predicted label change." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.39, + 0.788, + 0.45 + ], + "angle": 0, + "content": "Now we describe the definition of TE score. Given model weight \\(\\theta\\) and training dataset \\(D\\) with \\(c\\) classes, we denote by \\(P(y|\\theta ,x)\\) the probability (softmax) of the ground-truth class \\(y\\). We define the total effectiveness (TE) score of editing the model weight \\(\\theta\\) into the new weight \\(\\theta^{*}\\) given dataset \\(D\\) is defined as:" + }, + { + "type": "equation", + "bbox": [ + 0.373, + 0.457, + 0.786, + 0.492 + ], + "angle": 0, + "content": "\\[\ng (\\theta , \\theta^ {*}, D) = \\frac {1}{N} \\sum_ {x, y \\in D} \\frac {P (y | \\theta^ {*} , x)}{P (y | \\theta , x)} - 1 \\tag {2}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.215, + 0.5, + 0.787, + 0.53 + ], + "angle": 0, + "content": "A positive TE score represents an increased probability on ground-truth label while a negative TE score signifies a decrease." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.541, + 0.788, + 0.663 + ], + "angle": 0, + "content": "Step II-b: Generating prediction lists based on TE and ground-truth mask. This sub-task is to generate prediction lists based on TE and ground-truth mask. Specifically, the TE score, together with the ground-truth mask, makes up four combinations in total, and we call the list of all four cases a prediction list, denoted by \\(\\varphi\\). The prediction list then serves as a reference for evaluating the editing as well as the capability of such layer in improving model's accuracy. Below we show the prediction list that contains all four combinations of TE score and ground-truth mask in Equation 3" + }, + { + "type": "equation", + "bbox": [ + 0.381, + 0.673, + 0.786, + 0.742 + ], + "angle": 0, + "content": "\\[\n\\varphi = \\left\\{ \\begin{array}{l} f _ {\\theta^ {*}} (x) = y, g \\left(\\theta , \\theta^ {*}, z\\right) > 0 \\\\ f _ {\\theta^ {*}} (x) = y, g \\left(\\theta , \\theta^ {*}, z\\right) < 0 \\\\ f _ {\\theta^ {*}} (x) \\neq y, g \\left(\\theta , \\theta^ {*}, z\\right) > 0 \\\\ f _ {\\theta^ {*}} (x) \\neq y, g \\left(\\theta , \\theta^ {*}, z\\right) < 0 \\end{array} \\right. \\tag {3}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.75, + 0.788, + 0.841 + ], + "angle": 0, + "content": "The top case of the prediction list \\(\\varphi\\) shows the best editing (predicted label same as ground-truth label, TE score greater than 0) and bottom case shows the worst editing effect (predicted label different to ground-truth label, TE score less than 0). PFEDEDIT collects values for such prediction list for each editing by splitting which case each image prediction belongs to, and then ranks all prediction lists using Lexicographic order defined below." + } + ], + [ + { + "type": "header", + "bbox": [ + 0.655, + 0.115, + 0.732, + 0.127 + ], + "angle": 0, + "content": "PFEDEDIT" + }, + { + "type": "page_number", + "bbox": [ + 0.775, + 0.116, + 0.785, + 0.127 + ], + "angle": 0, + "content": "9" + }, + { + "type": "image", + "bbox": [ + 0.225, + 0.155, + 0.382, + 0.25 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.281, + 0.253, + 0.338, + 0.264 + ], + "angle": 0, + "content": "(a) Lyme" + }, + { + "type": "image", + "bbox": [ + 0.414, + 0.155, + 0.568, + 0.251 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.458, + 0.253, + 0.542, + 0.265 + ], + "angle": 0, + "content": "(b) CIFAR-10" + }, + { + "type": "image", + "bbox": [ + 0.604, + 0.155, + 0.758, + 0.251 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.645, + 0.253, + 0.734, + 0.265 + ], + "angle": 0, + "content": "(c) CIFAR-100" + }, + { + "type": "image_caption", + "bbox": [ + 0.214, + 0.267, + 0.785, + 0.298 + ], + "angle": 0, + "content": "Fig. 4: Non-i.i.d. data distribution (2 classes per client and 20 classes for CIFAR-100) among clients in Section 5.2" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.325, + 0.785, + 0.446 + ], + "angle": 0, + "content": "Step II-c: Locating top-\\(k\\) personalized layers using lexicographic order. This sub-step is to rank all prediction lists \\((\\varphi)\\) using lexicographic order and select the top-\\(k\\) as the personalized layers. Here is the definition of lexicographic order, which is used to rank \\(\\varphi\\). That is, given two prediction lists \\(\\varphi_1, \\varphi_2\\) for two model editorings on layer \\(l_1, l_2\\), respectively, say \\(\\varphi_1 = a_1, \\ldots, a_4\\), \\(\\varphi_2 = b_1, \\ldots, b_4\\), where \\(a\\) and \\(b\\) are possible cases from prediction lists, the order of those two lists depends on the numeric order of the cases in the first place \\(i\\) where the two case counts differ. That is, \\(\\varphi_1 < \\varphi_2\\) if and only if \\(a_i < b_i\\) in the natural order of numbers." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.447, + 0.787, + 0.585 + ], + "angle": 0, + "content": "Note that the advantages of using such metric here are two-folds: i) We record \\(\\varphi_{j}(\\theta, \\theta^{*}, z)\\) by counting the number of each case and compare any two prediction lists following the Lexicographic order instead of a summation manner, i.e., \\(\\sum_{|D|}^{i} \\varphi_{j}(\\theta, \\theta^{*}, z)\\), so that the impact of one case is not overridden by another, even if there are insufficient instances. ii) The case order listed in Equation 3 from top to the bottom, is intuitive and aligns with accuracy metric, thus implying that the editing with highest prediction list ranking comes with a largest accuracy improvement. We also discuss and compare our metrics with other potential ones, and evaluation details are shown in Section 5.4." + }, + { + "type": "title", + "bbox": [ + 0.216, + 0.608, + 0.356, + 0.623 + ], + "angle": 0, + "content": "5 Evaluation" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.64, + 0.785, + 0.702 + ], + "angle": 0, + "content": "In this section, we compare PFEDEDIT with baselines in terms of the accuracy and convergence. We then perform a study on parameters used in PFEDEDIT. The evaluation on computation overhead and complexity of PFEDEDIT can be found in Appendix A.4" + }, + { + "type": "title", + "bbox": [ + 0.215, + 0.724, + 0.427, + 0.74 + ], + "angle": 0, + "content": "5.1 Experimental Setup" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.75, + 0.785, + 0.779 + ], + "angle": 0, + "content": "We describe four aspects, i.e., setup of federated learning, datasets, models, and then baselines, for our experimental setup." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.78, + 0.787, + 0.84 + ], + "angle": 0, + "content": "FL Setup. Our default settings follow state-of-the-art works [2,6] in evaluating federated learning. That is, there are 10 clients to participate in a FL scenario with 100 communication rounds. The global sampling rate is set to be 1.0 in each communication round and each client has one local training epoch. The default" + } + ], + [ + { + "type": "page_number", + "bbox": [ + 0.218, + 0.116, + 0.236, + 0.127 + ], + "angle": 0, + "content": "10" + }, + { + "type": "header", + "bbox": [ + 0.272, + 0.115, + 0.368, + 0.127 + ], + "angle": 0, + "content": "H. Yuan et al." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.146, + 0.788, + 0.329 + ], + "angle": 0, + "content": "value \\((k)\\) in selecting and editing top layers is \\(7\\%\\), and the default representative subset ratio \\((p)\\) is 0.1. Both values are selected based on our parameter study (Section 5.4). By default, personalized layers are chosen each round, and the comparison with fixed personalized layer setting can be found in Appendix A.2. Datasets. We evaluate PFEDEDIT on three benchmark datasets to show its performance on different domains. Specifically, we choose i). CIFAR-10 [12], a benchmark dataset that contains 60,000 \\(32\\times 32\\) color images in 10 classes; ii). CIFAR-100 [12], a benchmark dataset that contains 60,000 \\(32\\times 32\\) color images in 100 classes; and iii). Lyme [35] dataset, a skin disease dataset that contains 3,027 skin disease images in 4 classes that are confusers for Lyme. We resize CIFAR-10 image samples to \\(64\\times 64\\), CIFAR-100 samples to \\(224\\times 224\\) following [36], and Lyme samples to \\(255\\times 255\\) following [35]." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.33, + 0.789, + 0.636 + ], + "angle": 0, + "content": "To show that PFEDEDIT can perform on various non-i.i.d. scenario, we adopt the Distributional Heterogeneity level (DH-level) from Yuan et al. [36] to split data distribution into different levels. Specifically, we first assign certain numbers of classes to each client \\(i\\), and then for each local class \\(c\\), we sample \\(D_{i,c} \\in (0.4, 0.6)\\) from \\(D_c\\). Each client is assigned \\(\\frac{D_{i,c}}{\\sum D_{n,c}}\\) of samples in class \\(c\\). The DH-level ranges from \\(0\\%\\), which represents an i.i.d. scenario where each client gets all classes of data, all the way to \\(100\\%\\), which means that each client only gets one class of data. (10 classes for CIFAR-100). For Lyme [35] dataset, since it only possesses four classes, we only conduct evaluations on \\(50\\%\\) (2 classes per clients) and \\(100\\%\\) (4 classes per clients) DH-level distribution scenarios. Unless specified otherwise, the default data distribution among clients is 2 classes per client (20 classes per client for CIFAR-100) in our evaluation. We demonstrate the default heterogeneous data distribution for each dataset in Figure 4. We also consider feature space non-i.i.d. scenario, and the evaluation, in comparison with Cd2-pfed [27] and Partialfed-fix [29], can be found in Appendix A.3. Models. For model selection, we choose ResNet18 [9] for evaluation on CIFAR 10 and CIFAR 100, and ViT [8] for Lyme. We allow ResNet18 model to have pretrained weights and allow ViT to have b16 weights that is fine-tuned on Image1k. We set the learning rate for ViT to be 1e-3 and we follow FedPer [2] to set the learning rate as 1e-2 for ResNet18." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.636, + 0.789, + 0.761 + ], + "angle": 0, + "content": "Baselines. We compare PFEDEDIT with FedAvg [20], MOON [14], and other five SOTA personalized FL works as baselines, namely, FedPer [2], FedRep [4], FedBN [18], APFL [6], and DisTrans [36]. We allow each baseline to have the same data distribution and model selection as PFEDEDIT. We consider all layer normalization layers personalized for FedBN [18] when employing ViT. For other parameters, we follow their default paper settings to assign values. The accuracy comparison with more baselines (Fedfa [38], FedALA [37], Fed-LAMB [10], FedRolex [1], and HeteroFL [7]) can be found in Appendix A.1." + }, + { + "type": "title", + "bbox": [ + 0.216, + 0.782, + 0.445, + 0.798 + ], + "angle": 0, + "content": "5.2 Accuracy Comparison" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.809, + 0.789, + 0.842 + ], + "angle": 0, + "content": "As Table 1 shows, PFEDEDIT outperforms other state-of-the-art works on all three datasets under different DH-level data distributions. Compared to the" + } + ], + [ + { + "type": "header", + "bbox": [ + 0.655, + 0.115, + 0.732, + 0.127 + ], + "angle": 0, + "content": "PFEDEDIT" + }, + { + "type": "page_number", + "bbox": [ + 0.769, + 0.117, + 0.784, + 0.127 + ], + "angle": 0, + "content": "11" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.147, + 0.788, + 0.266 + ], + "angle": 0, + "content": "baseline FedAvg, the accuracy improvement ranges from 0.007 to 0.266, with an average 0.09 accuracy improvement, and the largest accuracy improvement gets to 0.156 and average improvement gets to 0.05 when compared to other personalized FL works. The accuracy gap between PFEDEDIT and other STOA personalized works in CIFAR-10 is not as huge as in other two datasets, and the reason is that we allow all works to employ models with pretrained weights, which provides a greater lower bound for model accuracy and reduces the accuracy gap between different methods." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.268, + 0.789, + 0.54 + ], + "angle": 0, + "content": "We then evaluate PFEDEDIT, together with other SOTA personalized FL methods under different DH levels. According to Table II, PFEDEDIT consistently achieves superior accuracy compared to other SOTA works under all data distributions, and we have two observations here. First, some personalized FL methods, e.g., FedPer, FedBN, and FedRep, fail to outperform FedAvg under i.i.d. scenarios, which is attributed to their natures as personalized FL works: they are strategically designed to align the global model with local distributions under non-i.i.d. scenarios. If the data distribution is i.i.d., where there is a close alignment between the global model distribution and each local one, the personalized layers may block favorable updates from other clients. On the other hand, PFEDEDIT outperforms FedAvg under i.i.d. scenarios, and the reason is that PFEDEDIT allow clients to update the weights of personalized layers so that they can favor other clients when local distributions are similar. Second, the accuracy improvement brought by PFEDEDIT compared to FedAvg increases as the DH-level increases. The reason is that as the distribution becomes more non-i.i.d., the gap between the global and local model distributions widens. Personalized layers help bridge this gap, tailoring the global model towards the local distribution and improving its generalization on local data." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.542, + 0.789, + 0.631 + ], + "angle": 0, + "content": "It is noteworthy to mention that PFEDEDIT gets 0.989 accuracy on CIFAR-10 when DH-level gets to \\(100\\%\\). The reason is that, in an extreme non-i.i.d. scenario where clients only get one class of data samples, the local model is very likely to get 1.0 testing accuracy and the model editing effectively improves the model accuracy as the local model distribution is highly matched with the local data distribution." + }, + { + "type": "title", + "bbox": [ + 0.215, + 0.66, + 0.37, + 0.675 + ], + "angle": 0, + "content": "5.3 Convergence" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.689, + 0.789, + 0.841 + ], + "angle": 0, + "content": "In this section we evaluate the convergence rate of PFEDEDIT. Specifically, we compare PFEDEDIT with FedAvg and we demonstrate their best accuracy on each dataset and the numbers of rounds to achieve it. As Table 2 shows, PFEDEDIT converges much faster than FedAvg on all three datasets as it only requires 18 epochs to converge on CIFAR-100, 13 epochs on CIFAR-10, and 25 epochs on Lyme. On the opposite, FedAvg takes around three times numbers of rounds to get a suboptimal accuracy on both CIFAR-10 and CIFAR-100. For Lyme, it takes 87 epochs to get accuracy above 0.7, which only takes 4 epochs for PFEDEDIT. The reason is that FedAvg does not provide convergence guarantee as it aggregates all model weights each round without having any optimization," + } + ], + [ + { + "type": "page_number", + "bbox": [ + 0.218, + 0.116, + 0.236, + 0.127 + ], + "angle": 0, + "content": "12" + }, + { + "type": "header", + "bbox": [ + 0.272, + 0.115, + 0.368, + 0.128 + ], + "angle": 0, + "content": "H. Yuan et al." + }, + { + "type": "table_caption", + "bbox": [ + 0.229, + 0.145, + 0.773, + 0.159 + ], + "angle": 0, + "content": "Table 1: Evaluation of PFEDEDIT with other SOTA works on different datasets." + }, + { + "type": "table", + "bbox": [ + 0.218, + 0.172, + 0.784, + 0.281 + ], + "angle": 0, + "content": "
DatasetDH-levelFedAvg 20FedPer 2FedBN 18FedRep 4DisTrans 36MOON 14APFL 6PFEDEDIT
CIFAR-100 % (i.i.d)0.9420.8900.92720.8890.9470.9220.8910.955
40%0.9230.8730.9260.8840.9440.9310.8890.953
60%0.9390.8800.9400.9200.9600.9270.9100.965
80%0.8060.8870.9450.9330.9540.9380.9150.962
100%0.8300.9020.9610.9800.9720.9400.9270.989
CIFAR-1000 % (i.i.d)0.7830.7330.7460.7420.7880.7460.7440.790
40%0.7790.7530.7760.7470.7790.7620.7530.825
60%0.7720.7480.8020.7930.8300.7680.8020.855
80%0.7370.7690.8620.8590.8870.7700.8330.890
100%0.6640.7900.9040.8800.9030.7730.8590.930
Lyme0%0.7070.7000.7330.7300.7450.6990.7220.765
50 % (i.i.d)0.7160.8080.8210.8330.8500.7680.7810.912
" + }, + { + "type": "table_caption", + "bbox": [ + 0.216, + 0.292, + 0.787, + 0.322 + ], + "angle": 0, + "content": "Table 2: Best Accuracy of PFEDEdit/FedAvg and number of rounds to achieve it. (- represents not achieved within 100 rounds)" + }, + { + "type": "table", + "bbox": [ + 0.22, + 0.333, + 0.782, + 0.412 + ], + "angle": 0, + "content": "
Dataset| PFEDEDIT/FedAvg# Rounds for target accuracy (PFEDEDIT/FedAvg)
>0.6>0.7>0.75>0.8>0.85>0.9
CIFAR-100.962/0.8064/56/87/139/3311/-13/-
CIFAR-1000.890/0.7376/98/2410/-12/-18/--/-
Lyme0.912/0.7163/324/875/-5/-9/-25/-
" + }, + { + "type": "table_caption", + "bbox": [ + 0.216, + 0.424, + 0.787, + 0.454 + ], + "angle": 0, + "content": "Table 3: Accuracy of PFEDEDIT with different editing ratios (in comparison with FedAvg)." + }, + { + "type": "table", + "bbox": [ + 0.22, + 0.464, + 0.782, + 0.537 + ], + "angle": 0, + "content": "
Dataset| FedAvg| PFEDEDIT
1%3%5%7%9%
CIFAR-100.8060.9330.9470.9480.9620.952
CIFAR-1000.7480.8600.8780.8850.8900.874
Lyme0.7160.8680.8800.8890.9120.910
" + }, + { + "type": "text", + "bbox": [ + 0.216, + 0.565, + 0.787, + 0.596 + ], + "angle": 0, + "content": "and PFEDEDIT edits the global model and tailors it towards the local distribution, which makes it faster to generalize on local distributions." + }, + { + "type": "title", + "bbox": [ + 0.216, + 0.618, + 0.406, + 0.634 + ], + "angle": 0, + "content": "5.4 Parameter Study" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.644, + 0.788, + 0.764 + ], + "angle": 0, + "content": "Model Editing Ratio \\((k)\\) PFEDEDIT allows users to pre-define \\(k\\), the number of layers to be edited. It then edits the top-\\(k\\) layers, which are considered to be the most effective in improving model accuracy. In this section we set up an experiment to show the correlation between the number of modified layers and the model accuracy. Intuitively, if the number of edited layers increases, the model accuracy gets higher as greater part of the model is tailored towards local distribution. However, such increase may not hold all the time as there are layers that do not possess any learnable parameters and the model may converge." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.765, + 0.788, + 0.841 + ], + "angle": 0, + "content": "We choose the ratios of edited layers to be \\(1\\%\\), \\(3\\%\\), \\(5\\%\\), \\(7\\%\\), and \\(9\\%\\) to the total model layers and we round up the number in each case. Shown in Table 3, as the editing ratio increases, the overall testing accuracy gets higher. The accuracy improvement is greater from improving the ratio from \\(1\\%\\) to \\(3\\%\\) and the increasing rate gets flatter when keep improving the ratio from \\(3\\%\\) to" + } + ], + [ + { + "type": "header", + "bbox": [ + 0.655, + 0.115, + 0.732, + 0.127 + ], + "angle": 0, + "content": "PFEDEDIT" + }, + { + "type": "page_number", + "bbox": [ + 0.769, + 0.116, + 0.786, + 0.127 + ], + "angle": 0, + "content": "13" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.146, + 0.788, + 0.222 + ], + "angle": 0, + "content": "\\(5\\%\\), or \\(5\\%\\) to \\(7\\%\\). At the end, when the ratio is set from \\(7\\%\\) to \\(9\\%\\), the accuracy starts to drop a little bit, implying that the model may converge on that accuracy. Compared to the baseline, PFEDEDIT improves the model accuracy from 0.13, with editing \\(1\\%\\) layers, to 0.16, with editing \\(7\\%\\) layers, on three datasets on average." + }, + { + "type": "table_caption", + "bbox": [ + 0.216, + 0.244, + 0.504, + 0.262 + ], + "angle": 0, + "content": "Table 4: Evaluation of PFedEdit with representative subset ratio (i.e., \\( p \\) )." + }, + { + "type": "table", + "bbox": [ + 0.217, + 0.263, + 0.513, + 0.32 + ], + "angle": 0, + "content": "
DatasetSubser ratio
p=0.01p=0.025p=0.05p=0.08p=0.1
CIFAR-100.9510.9530.9540.9540.962
CIFAR-100-0.8750.8840.8860.890
" + }, + { + "type": "table_caption", + "bbox": [ + 0.528, + 0.244, + 0.785, + 0.262 + ], + "angle": 0, + "content": "Table 5: Evaluation of PFedEdit using different metrics in assessing model editing" + }, + { + "type": "table", + "bbox": [ + 0.529, + 0.263, + 0.785, + 0.321 + ], + "angle": 0, + "content": "
DatasetDifferent metrics
AccuracyLossTE scorePrediction list
CIFAR-100.5400.9050.8850.962
CIFAR-1000.7050.8860.8210.890
" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.342, + 0.789, + 0.689 + ], + "angle": 0, + "content": "Representative Subset Ratio \\((p)\\) As mentioned in Section 4, PFEDEDIT queries the model using a representative subset \\(D_{p}\\) of the training set \\(D_{train}\\). PFEDEDIT employs a ratio \\(p\\) to pre-define the subset size compared to the training set, \\(|D_{p}| = p * |D_{train}|\\). We want to ensure that the subset is representative enough for potential testing image samples while not introducing too much computation overhead. Therefore, we conduct a study on the size of the subset versus the performance of PFEDEDIT. Intuitively, a larger size covers a broader local data distribution and leads to a more accurate layer locating and thus a higher model accuracy. However, on the other hand, taken a larger subset from the original training set reduces the size of the remaining set. We want to find an optimal size of the subset that is large enough to give a complete representation of local data distribution but also leaves enough samples for the local training. We split the local training set into a subset that takes up of \\(p * |D_{train}|\\) image samples, and a training set that takes the rest \\(|D_{train}^{*}| = |D_{train}| - |D_{p}|\\) image samples. To control the variable, we fix the size of \\(D_{train}^{*}\\), and we choose different \\(p\\) values from \\(\\{0.1, 0.08, 0.05, 0.025, 0.01\\}\\) to adjust the subset size. We conduct the evaluation on CIFAR-10 and CIFAR-100 [12], with 10 clients and each client is assigned 2 classes of image samples. Each client possesses 4,500 image samples as training set, and the subset size varies from 50 image samples (\\(p\\) value 0.01) to 500 image samples (\\(p\\) value 0.1). Note that the \\(p\\) value starts from 0.025 for CIFAR-100 because a subset that is 0.01 times the size of training set does not include image samples from all classes, i.e., 50 image samples do not cover 100 classes thus the subset is not representative." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.689, + 0.789, + 0.841 + ], + "angle": 0, + "content": "As Table 4 shows, the overall trend of the testing accuracy keeps increasing as \\( p \\) value gets larger from 0.01 times the training set to 0.1 times the training set for both datasets. However, the overall accuracy improvement is minimal compared to the subset size increase, demonstrating the capability of PFEDEDIT in improving model accuracy with a small size subset. For CIFAR-10, as \\( p \\) value increases from 0.01 to 0.08, which is an 8 times increase, the model accuracy is improved from 0.951 to 0.954, which is a \\( 0.3\\% \\) improvement. As \\( p \\) value increases from 0.08 to 0.1, the accuracy is further improved from 0.954 to 0.962, which is a \\( 0.8\\% \\) increase. For CIFAR-100, the accuracy improvement brought by increasing \\( p \\) value is greater than that in CIFAR-10 as the model accuracy is improved from" + } + ], + [ + { + "type": "page_number", + "bbox": [ + 0.218, + 0.116, + 0.236, + 0.127 + ], + "angle": 0, + "content": "14" + }, + { + "type": "header", + "bbox": [ + 0.272, + 0.115, + 0.368, + 0.128 + ], + "angle": 0, + "content": "H. Yuan et al." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.147, + 0.788, + 0.222 + ], + "angle": 0, + "content": "0.875 to 0.890 when \\( p \\) value changes from 0.025 to 0.1. Therefore, PFEDEDIT is computation-friendly as it requires only 50 image samples to get a 0.951 model accuracy and 500 image samples to achieve a 0.962 accuracy on CIFAR-10. Similarly, PFEDEDIT requires only 125 image samples as the subset to achieve a 0.875 accuracy and 500 image samples to gets a 0.890 accuracy on CIFAR-100." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.244, + 0.788, + 0.349 + ], + "angle": 0, + "content": "Metrics Selection In this section, we conduct experiments to explore possible metric selections for PFEDEDIT when assessing each model editing. Besides the prediction list that demonstrates all four combinations of TE score and ground-truth mask, we also choose model accuracy, loss value, and TE score. We select cross-entropy loss as our loss metric, which is defined as \\( H(P,Q) = -\\sum_{x \\in \\mathcal{X}} p(x) \\log q(x) \\), given two probability distribution \\( P \\) and \\( Q \\), and the TE score is introduced in Section 4." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.35, + 0.788, + 0.561 + ], + "angle": 0, + "content": "As Table 5 shows, the prediction list employed by PFEDEDIT outperforms other metrics and the accuracy gap ranges from 0.057 to 0.422. Among all possible metrics, using accuracy to assess model editorings leads to the lowest accuracy, 0.540 on CIFAR-10 and 0.705 on CIFAR-100, lower than the baseline FedAvg [20]. The reason is that the effects of some editorings are subtle and they only cause the probability change but not necessarily results in label change. As the way that accuracy is computed only counts on samples that are correctly classified, small probability change that do not alter the final prediction will be considered as no improvement at all. On the other hand, the TE score reflects the probability change on ground-truth label, but it does not reflects any label change. In other words, an increase in the probability of the ground-truth label suggests a positive editing, but a decrease in probability does not necessarily indicate a negative editing. From the accuracy perspective, the edited model performs identically to the original model despite these probability changes." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.561, + 0.788, + 0.682 + ], + "angle": 0, + "content": "While employing loss functions as the metric provides insights into the probability of the ground-truth class, the averaged loss value may exhibit bias if outliers have a disproportionate impact on the probability of the ground-truth class. For example, If an editing improves the model accuracy in correctly classifying two samples with a slight increase in the ground-truth class and misclassifies one sample with a notable drop in the ground-truth class, the improvement introduced by the correct classifications might be nullified by the incorrect one, which leads to a biased averaged loss value with respect to each data sample." + }, + { + "type": "title", + "bbox": [ + 0.216, + 0.704, + 0.36, + 0.72 + ], + "angle": 0, + "content": "6 Conclusion" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.735, + 0.788, + 0.84 + ], + "angle": 0, + "content": "In this paper, we design the first personalized federated learning framework that utilizes automated model editing as a core strategy to improve the model's utility. Our approach leverages layer-specific editing, which automatically identifies specific layers to be edited based on local data distribution, allowing for more effective adaptation of the global model to heterogeneous data distributions across various clients. Our evaluation demonstrates that PFEDEDIT significantly improves the accuracy of the global model from non-i.i.d. to i.i.d. data distribution." + } + ], + [ + { + "type": "header", + "bbox": [ + 0.655, + 0.115, + 0.732, + 0.127 + ], + "angle": 0, + "content": "PFEDEDIT" + }, + { + "type": "page_number", + "bbox": [ + 0.769, + 0.116, + 0.786, + 0.127 + ], + "angle": 0, + "content": "15" + }, + { + "type": "title", + "bbox": [ + 0.218, + 0.145, + 0.393, + 0.163 + ], + "angle": 0, + "content": "Acknowledgments" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.178, + 0.788, + 0.269 + ], + "angle": 0, + "content": "This work was supported in part by National Science Foundation (NSF) under grants OAC-23-19742 and Johns Hopkins University Institute for Assured Autonomy (IAA) with grants 80052272 and 80052273. The views and conclusions contained herein are those of the authors and should not be interpreted as necessarily representing the official policies or endorsements, either expressed or implied, of NSF or JHU-IAA." + }, + { + "type": "title", + "bbox": [ + 0.218, + 0.293, + 0.324, + 0.309 + ], + "angle": 0, + "content": "References" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.325, + 0.785, + 0.353 + ], + "angle": 0, + "content": "1. Alam, S., Liu, L., Yan, M., Zhang, M.: Fedrolex: Model-heterogeneous federated learning with rolling sub-model extraction (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.354, + 0.785, + 0.381 + ], + "angle": 0, + "content": "2. Arivazhagan, M.G., Aggarwal, V., Singh, A.K., Choudhary, S.: Federated learning with personalization layers. arXiv preprint arXiv:1912.00818 (2019)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.382, + 0.785, + 0.408 + ], + "angle": 0, + "content": "3. Basu, S., Zhao, N., Morariu, V., Feizi, S., Manjunatha, V.: Localizing and editing knowledge in text-to-image generative models (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.41, + 0.785, + 0.45 + ], + "angle": 0, + "content": "4. Collins, L., Hassani, H., Mokhtari, A., Shakkottai, S.: Exploiting shared representations for personalized federated learning. In: Proceedings of the 38th International Conference on Machine Learning (2021)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.451, + 0.785, + 0.478 + ], + "angle": 0, + "content": "5. Dai, D., Dong, L., Hao, Y., Sui, Z., Chang, B., Wei, F.: Knowledge neurons in pretrained transformers. arXiv preprint arXiv:2104.08696 (2021)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.479, + 0.785, + 0.506 + ], + "angle": 0, + "content": "6. Deng, Y., Kamani, M.M., Mahdavi, M.: Adaptive personalized federated learning (2020)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.507, + 0.785, + 0.534 + ], + "angle": 0, + "content": "7. Diao, E., Ding, J., Tarokh, V.: Heterofl: Computation and communication efficient federated learning for heterogeneous clients (2021)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.535, + 0.785, + 0.576 + ], + "angle": 0, + "content": "8. Dosovitskiy, A., Beyer, L., Kolesnikov, A., Weissenborn, D., Zhai, X., Unterthiner, T., Dehghani, M., Minderer, M., Heigold, G., Gelly, S., Uszkoreit, J., Houlsby, N.: An image is worth 16x16 words: Transformers for image recognition at scale (2021)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.577, + 0.785, + 0.618 + ], + "angle": 0, + "content": "9. He, K., Zhang, X., Ren, S., Sun, J.: Deep residual learning for image recognition. In: Proceedings of the IEEE conference on computer vision and pattern recognition. pp. 770-778 (2016)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.619, + 0.785, + 0.645 + ], + "angle": 0, + "content": "0. Karimi, B., Li, P., Li, X.: Layer-wise and dimension-wise locally adaptive federated learning (2022)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.646, + 0.785, + 0.687 + ], + "angle": 0, + "content": "1. Karimireddy, S.P., Kale, S., Mohri, M., Reddi, S., Stich, S., Suresh, A.T.: Scaffold: Stochastic controlled averaging for federated learning. In: International conference on machine learning (2020)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.688, + 0.785, + 0.715 + ], + "angle": 0, + "content": "2. Krizhevsky, A.: Learning multiple layers of features from tiny images. Tech. rep. (2009)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.716, + 0.785, + 0.743 + ], + "angle": 0, + "content": "3. Li, L., Fan, Y., Tse, M., Lin, K.Y.: A review of applications in federated learning. Computers & Industrial Engineering (2020)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.744, + 0.785, + 0.771 + ], + "angle": 0, + "content": "4. Li, Q., He, B., Song, D.: Model-contrastive federated learning. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (2021)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.772, + 0.785, + 0.799 + ], + "angle": 0, + "content": "5. Li, T., Sahu, A.K., Talwalkar, A., Smith, V.: Federated learning: Challenges, methods, and future directions. IEEE signal processing magazine (2020)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.799, + 0.785, + 0.84 + ], + "angle": 0, + "content": "6. Li, T., Sahu, A.K., Zaheer, M., Sanjabi, M., Talwalkar, A., Smith, V.: Federated optimization in heterogeneous networks. Proceedings of Machine learning and systems 2, 429-450 (2020)" + }, + { + "type": "list", + "bbox": [ + 0.226, + 0.325, + 0.785, + 0.84 + ], + "angle": 0, + "content": null + } + ], + [ + { + "type": "page_number", + "bbox": [ + 0.218, + 0.116, + 0.236, + 0.127 + ], + "angle": 0, + "content": "16" + }, + { + "type": "header", + "bbox": [ + 0.272, + 0.115, + 0.368, + 0.128 + ], + "angle": 0, + "content": "H. Yuan et al." + }, + { + "type": "ref_text", + "bbox": [ + 0.218, + 0.148, + 0.785, + 0.175 + ], + "angle": 0, + "content": "17. Li, X., Li, S., Song, S., Yang, J., Ma, J., Yu, J.: Pmet: Precise model editing in a transformer. arXiv preprint arXiv:2308.08742 (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.218, + 0.177, + 0.787, + 0.218 + ], + "angle": 0, + "content": "18. Li, X., Jiang, M., Zhang, X., Kamp, M., Dou, Q.: Fed{bn}: Federated learning on non-{iid} features via local batch normalization. In: International Conference on Learning Representations (2021), https://openreview.net/pdf?id=6YEQUnOQICG" + }, + { + "type": "ref_text", + "bbox": [ + 0.218, + 0.218, + 0.785, + 0.272 + ], + "angle": 0, + "content": "19. Liu, Q., Chen, C., Qin, J., Dou, Q., Heng, P.A.: Feddg: Federated domain generalization on medical image segmentation via episodic learning in continuous frequency space. The IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR) (2021)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.273, + 0.785, + 0.314 + ], + "angle": 0, + "content": "20. McMahan, B., Moore, E., Ramage, D., Hampson, S., y Arcas, B.A.: Communication-efficient learning of deep networks from decentralized data. In: Artificial intelligence and statistics (2017)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.315, + 0.785, + 0.342 + ], + "angle": 0, + "content": "21. Meng, K., Bau, D., Andonian, A., Belinkov, Y.: Locating and editing factual associations in GPT. Advances in Neural Information Processing Systems (2022)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.343, + 0.785, + 0.37 + ], + "angle": 0, + "content": "22. Meng, K., Bau, D., Andonian, A., Belinkov, Y.: Locating and editing factual associations in GPT. Advances in Neural Information Processing Systems 35 (2022)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.371, + 0.785, + 0.411 + ], + "angle": 0, + "content": "23. Meng, K., Sen Sharma, A., Andonian, A., Belinkov, Y., Bau, D.: Mass editing memory in a transformer. The Eleventh International Conference on Learning Representations (ICLR) (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.412, + 0.785, + 0.452 + ], + "angle": 0, + "content": "24. Nguyen, D.C., Ding, M., Pathirana, P.N., Seneviratne, A., Li, J., Vincent Poor, H.: Federated learning for internet of things: A comprehensive survey. IEEE Communications Surveys & Tutorials (2021)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.453, + 0.785, + 0.494 + ], + "angle": 0, + "content": "25. Nguyen, D.C., Ding, M., Pham, Q.V., Pathirana, P.N., Le, L.B., Seneviratne, A., Li, J., Niyato, D., Poor, H.V.: Federated learning meets blockchain in edge computing: Opportunities and challenges. IEEE Internet of Things Journal (2021)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.495, + 0.785, + 0.522 + ], + "angle": 0, + "content": "26. Nori, H., Jenkins, S., Koch, P., Caruana, R.: Interpretml: A unified framework for machine learning interpretability. arXiv preprint arXiv:1909.09223 (2019)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.523, + 0.785, + 0.549 + ], + "angle": 0, + "content": "27. Shen, Y., Zhou, Y., Yu, L.: \\(\\mathrm{Cd^{2}}\\)-pfed: Cyclic distillation-guided channel decoupling for model personalization in federated learning (2022)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.55, + 0.785, + 0.577 + ], + "angle": 0, + "content": "28. Shokri, R., Shmatikov, V.: Privacy-preserving deep learning. In: Proceedings of the 22nd ACM SIGSAC conference on computer and communications security (2015)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.578, + 0.785, + 0.618 + ], + "angle": 0, + "content": "29. Sun, B., Huo, H., YANG, Y., Bai, B.: Partialfed: Cross-domain personalized federated learning via partial initialization. In: Advances in Neural Information Processing Systems (2021)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.619, + 0.785, + 0.66 + ], + "angle": 0, + "content": "30. Wang, P., Zhang, N., Xie, X., Yao, Y., Tian, B., Wang, M., Xi, Z., Cheng, S., Liu, K., Zheng, G., et al.: Easilyedit: An easy-to-use knowledge editing framework for large language models. arXiv preprint arXiv:2308.07269 (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.661, + 0.785, + 0.701 + ], + "angle": 0, + "content": "31. Yang, Q., Liu, Y., Chen, T., Tong, Y.: Federated machine learning: Concept and applications. ACM Transactions on Intelligent Systems and Technology (TIST) (2019)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.702, + 0.785, + 0.742 + ], + "angle": 0, + "content": "32. Yang, Y., Hui, B., Yuan, H., Gong, N., Cao, Y.: PrivateFL: Accurate, differentially private federated learning via personalized data transformation. In: 32nd USENIX Security Symposium (USENIX Security 23). pp. 1595-1612 (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.743, + 0.785, + 0.799 + ], + "angle": 0, + "content": "33. Yang, Y., Yuan, H., Hui, B., Gong, N., Fendley, N., Burlina, P., Cao, Y.: Fortifying federated learning against membership inference attacks via client-level input perturbation. In: 2023 53rd Annual IEEE/IFIP International Conference on Dependable Systems and Networks (DSN) (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.8, + 0.785, + 0.84 + ], + "angle": 0, + "content": "34. Yang, Z., Zhang, Y., Zheng, Y., Tian, X., Peng, H., Liu, T., Han, B.: Fedfed: Feature distillation against data heterogeneity in federated learning. In: Thirty-seventh Conference on Neural Information Processing Systems (2023)" + }, + { + "type": "list", + "bbox": [ + 0.217, + 0.148, + 0.787, + 0.84 + ], + "angle": 0, + "content": null + } + ], + [ + { + "type": "header", + "bbox": [ + 0.655, + 0.115, + 0.732, + 0.127 + ], + "angle": 0, + "content": "PFEDEDIT" + }, + { + "type": "page_number", + "bbox": [ + 0.769, + 0.116, + 0.786, + 0.127 + ], + "angle": 0, + "content": "17" + }, + { + "type": "ref_text", + "bbox": [ + 0.214, + 0.147, + 0.788, + 0.203 + ], + "angle": 0, + "content": "35. Yuan, H., Aucott, J., Hadzic, A., Paul, W., Villegas de Flores, M., Mathew, P., Burlina, P., Cao, Y.: Edgemixup: Embarrassingly simple data alteration to improve Lyme disease lesion segmentation and diagnosis fairness. In: Medical Image Computing and Computer Assisted Intervention - MICCAI 2023 (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.214, + 0.204, + 0.788, + 0.259 + ], + "angle": 0, + "content": "36. Yuan, H., Hui, B., Yang, Y., Burlina, P., Gong, N.Z., Cao, Y.: Addressing heterogeneity in federated learning via distributional transformation. In: Computer Vision - ECCV 2022: 17th European Conference, Tel Aviv, Israel, October 23-27, 2022, Proceedings, Part XXXVIII. p. 179-195 (2022)" + }, + { + "type": "ref_text", + "bbox": [ + 0.215, + 0.259, + 0.787, + 0.3 + ], + "angle": 0, + "content": "37. Zhang, J., Hua, Y., Wang, H., Song, T., Xue, Z., Ma, R., Guan, H.: Fedala: Adaptive local aggregation for personalized federated learning. In: Proceedings of the AAAI Conference on Artificial Intelligence (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.215, + 0.3, + 0.787, + 0.328 + ], + "angle": 0, + "content": "38. Zhou, T., Konukoglu, E.: FedFA: Federated feature augmentation. In: The Eleventh International Conference on Learning Representations (ICLR) (2023)" + }, + { + "type": "list", + "bbox": [ + 0.214, + 0.147, + 0.788, + 0.328 + ], + "angle": 0, + "content": null + } + ] +] \ No newline at end of file diff --git a/2024/PFedEdit_ Personalized Federated Learning via Automated Model Editing/984c6b27-188f-4dc6-b40f-18bc786fe91d_origin.pdf b/2024/PFedEdit_ Personalized Federated Learning via Automated Model Editing/984c6b27-188f-4dc6-b40f-18bc786fe91d_origin.pdf new file mode 100644 index 0000000000000000000000000000000000000000..4d2562c956a7f81a815696b912f2cc113e21ab4f --- /dev/null +++ b/2024/PFedEdit_ Personalized Federated Learning via Automated Model Editing/984c6b27-188f-4dc6-b40f-18bc786fe91d_origin.pdf @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:f4ca3097979a054aa57b152405fd20a18f49017b9d1879ab333c4ed5923662c8 +size 1617809 diff --git a/2024/PFedEdit_ Personalized Federated Learning via Automated Model Editing/full.md b/2024/PFedEdit_ Personalized Federated Learning via Automated Model Editing/full.md new file mode 100644 index 0000000000000000000000000000000000000000..d8311b25366f9761a062b9b80330c241f96587ad --- /dev/null +++ b/2024/PFedEdit_ Personalized Federated Learning via Automated Model Editing/full.md @@ -0,0 +1,276 @@ +# PFEDEDIT: Personalized Federated Learning via Automated Model Editing + +Haolin Yuan $^{1}$ , William Paul $^{2}$ , John Aucott $^{3}$ , Philippe Burlina $^{2}$ , and Yinzhi Cao $^{1}$ + +1 Johns Hopkins University {hyuan4, yinzhi.cao}@jhu.edu + +2 Johns Hopkins Applied Physics Lab {william.paul, + +philippe.burlina}@jhuapl.edu + +3 Johns Hopkins University School of Medicine jaucott2@jhmi.edu + +Abstract. Federated learning (FL) allows clients to train a deep learning model collaboratively while maintaining their private data locally. One challenging problem facing FL is that the model utility drops significantly once the data distribution gets heterogeneous, or non-i.i.d, among clients. A promising solution is to personalize models for each client, e.g., keeping some layers locally without aggregation, which is thus called personalized FL. However, previous personalized FL often suffer from suboptimal utility because their choice of layer personalization is based on empirical knowledge and fixed for different datasets and distributions. In this work, we design PFEDEDIT, the first federated learning framework that leverages automated model editing to optimize the choice of personalization layers and improve model utility under a variety of data distributions including non-i.i.d. The high-level idea of PFEDEDIT is to assess the effectiveness of every global model layer in improving model utility on local data distribution once edited, and then to apply edits on the top- $k$ most effective layers. Our evaluation shows that PFEDEDIT outperforms six state-of-the-art approaches on three benchmark datasets by $6\%$ on the model's performance on average, with the largest accuracy improvement being $26.6\%$ . PFEDEDIT is open-source and available at this repository: https://github.com/Haolin-Yuan/PFedEdit + +# 1 Introduction + +Federated learning (FL) [13,15,20,31,33] has emerged as a widely used training algorithm of deep learning, especially in the image domain, due to its practical applications in real-world scenarios and its impressive performance. A federated learning training scenario entails multiple local clients and a global server. Each client trains their local model using their private image data. After performing local training, each client updates their local model weights to the global server for aggregation with other clients' model weights. Federated learning allows clients to train a collaborative model that performs well on multi-class tasks—even if each client only possesses a few classes of images—via only sharing models but not local training images. + +![](images/771cb20bf21dae387b6f5b9a601b277237bea59b2b0ccd95e59bd8fa837624f5.jpg) +Fig. 1: High-level idea of PFEDEDIT (PFEDEDIT edits the global model returned from FL server to fit the local client's distribution.) + +As the goal of federated learning is to learn a global model for all devices cooperatively, if the data distributions among clients are similar, the global model is supposed to perform better than local models that are solely trained on local data. However, once the data distribution becomes skewed, or non-i.i.d., e.g., some clients may possess image samples of only one or two classes or some possess many more image samples from one class than others, the global model becomes ineffective to be employed as local models. The reason is that the local data distribution at one client could be dramatically different from others, thus causing local model weights to adversely affect each other during the model aggregation. + +Multiple works [11,14,16] have been proposed to improve the FL model accuracy under the non-i.i.d. distribution by optimizing local models. They either regularize local models to maximize the agreement between clients [14], or reduce clients' weights drift from the server end [11,16] for a better global model that generalizes on each local distribution. While such works improved model's accuracy, the model is still uniform across different clients, and cannot be tailored based on different clients' data distribution. Therefore, one natural step is a so-called personalized FL [2,4,18,36], which enables each user to have personalized layers that help customize local models to their data distributions. Those personalized layers are retained locally each communication round and are not updated from the global model, thereby preventing being aggregated with others and avoiding bias towards other clients' data distributions. The choices of those personalized layers are various but are all pre-defined before the training, e.g., it can be linear layer [2], batch normalization layer [4], or specially designed layer [36]. + +However, there are still two challenges that remain for personalized FLs despite the accuracy improvement. First, the personalized layers chosen in previous personalized FL works [2,4,18,36] are inspired by empirical findings and fixed, implying that those choices may not be optimal when given different data distributions or datasets. That is, while the weights of those layers might be personalized, the locations and types of those layers are not, e.g., either linear or batch normalization layers in certain locations of a learning model. Second, previous personalized works are limited in their application scenarios as they assume certain model structure [18], which may not be used for certain clients. + +In this paper, to address these two challenges, we design a framework that automatically locates and chooses layers to be personalized. We propose the first personalized federated learning framework, called PFEDEDIT, which leverages automated model editing [3,5,21,23] to optimize the layer choosing and improve model utility under different settings including non-i.i.d. The high-level idea of PFEDEDIT, as shown in Figure 1, is to look for model layers that contain knowledge about local distribution that benefits clients the most, and by editing those layers, PFEDEDIT mitigates the negative influence brought by other clients, thus tailoring the global model weights towards the local model weights in each round. Specifically, PFEDEDIT assesses which layers of the global model are the most effective in maintaining the local distribution and then marks them as target layers. PFEDEDIT then edits the top- $k$ target layers to tailor the global model weights closer to local models thus improving the model utility. PFEDEDIT does not make any assumptions on model selection and thus is model-structure agnostics. + +We performed extensive experiments on three benchmark datasets, namely CIFAR-10 [12], CIFAR-100 [12], and Lyme dataset [35], and we compared PFEDEDIT with FedAvg [20] and other six state-of-the-art personalized federated learning approaches. Our experimental results demonstrate that PFEDEDIT outperforms all SOTA personalized federated learning works on all the datasets, with an average $6\%$ accuracy improvement. + +In summary, our contributions are as follows: + +- We design PFEDEDIT, the first personalized federated learning framework that optimizes personalized layer selections for each client using model editing. +- PFEDEDIT is adaptable to all model structures because it does not impose any assumption on model selection. +- We conduct extensive experiments to show that PFEDEDIT outperforms other SOTA personalized federated works on different benchmark datasets. + +# 2 Related Work + +Federated Learning. Federated learning [13,15,20,31] (FL) is first proposed to predict user's text input on portable devices and it has been widely deployed in other industries, such as medical imaging [19], Internet of Things (IoT) [24], and block-chain [25]. It allows multiple clients to collaboratively train a model that can perform on samples from any clients' data distribution. Two common approaches in federated learning (FL) are FedSGD [28] and FedAvg [20]. FedSGD optimizes global weights by back-propagating each local gradient, while FedAvg allows for multiple local training epochs and client-side optimization. In FedAvg, the server's role is limited to averaging and distributing the aggregated weights. While FedAvg reduces communication costs compared to FedSGD, it lacks a convergence guarantee, and heterogeneous data distribution among clients can significantly reduce the global model's utility. Therefore, various works [2,4,14,18,32,36] have been proposed to solve such utility degradation caused by data heterogeneity. One direction [14] is to optimize local + +model either via personalized objective functions; yet, another, called personalized FL [2,4,18,36], allows clients to have personalized model layers that can be kept locally. As a comparison, PFEDEDIT is the first work that optimizes and personalizes the choice of model layers with further improved utilities of FL, especially under non-i.i.d. settings. + +Model Editing. Model editing 3.5.17.23.26 involves adding, removing, or adjusting layers and parameters of a model architecture to improve its performance or adapt it to new tasks. Specifically, let $h_i$ be the $i$ -th hidden states of any model, $\{h_l | l \in [1, L]\}$ , where $L$ is the number of layers in model, model editing allows users to find each hidden state's contribution towards a specific prediction. It mainly contains three runs: i). clean run, in which the statistics of each state are recorded when fed a factual prompt $x$ ; ii). corrupted run, in which each hidden layer is added Gaussian noise such that $h_l \coloneqq h_l + \epsilon$ . Because the hidden layer is obfuscated by the added noise, the model will output incorrect predictions; iii). restoration run, in which the model gives predictions with noised hidden states recovered to clean one at a time to find out which layer is related to producing factual knowledge. Following all three steps, users can find out the causal importance of all model layers in the computation graph and edit weights of those knowledge-related layers to alter the final model output. + +There are different applications of model editing. For example, model editing can be employed on large language models (LLMs) to facilitate the interpretation of LLMs as black boxes when generating particular content or to guide LLMs in producing desired outputs [22,30]. ROME [22] edits LLMs' linear layers to modify some learned factual knowledge and make LLMs output desired content such as "Eiffel Tower is located in the city of Rome". Meng et al. [23] utilizes Gaussian noise to perturb predictions to achieve mass editing in a transformer. + +As a comparison, PFEDEDIT is different from prior model editing works in two aspects. First, PFEDEDIT is a layer-wise model editing approach, which replaces the entire model layer from the local model instead of editing specific neuron stats. The advantage is that PFEDEDIT is able to perform well against potential images from various data distributions. Second, PFEDEDIT considers the global model as corrupted instead of one with Gaussian noise to reduce computation overhead. + +# 3 Overview + +We start by describing some backgrounds in defining personalized federated learning and then present a motivating example. + +# 3.1 Background: Personalized Federated Learning + +We describe some backgrounds of personalized federated learning. Consider the initial weights for all local models as $\{\theta_1, \theta_2, \theta_3, \dots, \theta_n\}$ . Let $\mathcal{P}_i$ be the data distribution over domain $\mathcal{X} \times \mathcal{Y}$ and $l_i$ be the loss function for client $i \in \{1, 2, 3.., n\}$ . + +The goal of personalized federated learning is to find the set of personalized weights $\Theta$ for each client such that it minimizes empirical loss with respect to each $\mathcal{P}_i$ as shown in Equation ① + +$$ +\Theta = \operatorname {a r g m i n} _ {\theta_ {i}} \frac {1}{n} \sum_ {i = 1} ^ {n} \mathbb {E} _ {x, y \sim \mathcal {P} _ {i}} \left[ l _ {i} \left(\theta_ {i}, x, y\right) \right] \tag {1} +$$ + +# 3.2 A Motivating Example + +In this part, we run through an example to motivate PFEDEDIT. We start with a challenging question. Note that previous personalized FL work [24,18,36] has noted the advantages of keeping some layers locally under non-i.i.d. scenarios, e.g., all batch normalization layers for FedBN [18]. This observation raises an intriguing question: How do we choose personalized layers to improve FL's performance with heterogeneous data? + +Our motivating example provides insight into this question. Specifically, we consider a real-world federated learning scenario where there are 10 clients collaboratively training their models on the CIFAR-100 dataset [12]. Each client is assigned 20 classes of image samples following Yuan et al. [36] to form a non-i.i.d. scenario, and we choose ResNet18 [9] as their classifier. To inspect the correlation between the personalized layers and the model utility, we pick $\{1,3,5,7,10,13,17\}$ as the number of personalized layers, and choose three personalized algorithms for comparison. + +![](images/e249f10e2e43a472744fb26833ae42c07a89cbd4003c06b20d19be9b37536e72.jpg) +Fig. 2: A motivating example: We consider three different ways to personalize layers of local models in a federated learning scenario: random choosing, fixed choosing, and optimized layerchoosing based on the loss value. For the fixed choosing algorithm, we allow clients to keep certain numbers of batch normalization layers locally each round, and this approach becomes equivalent to FedBN [18] when the number of personalized layers gets 17. + +The first algorithm randomly chooses certain numbers of layers from local model and keeps those layers locally. The second algorithm follows FedBN [18] to personalize batch normalization layers, and we denoted it as fixed choosing. The third personalized algorithm optimizes layer-choosing based on loss values and it considers layers that, once kept locally, yield the least loss values personalized layer and keeps them locally each round. + +Figure 2 shows the results: The optimized-personalized algorithm clearly outperforms both fixed and random personalized algorithms in all cases. This motivates our approach of using model editing to choose the optimal layers for personalization. Note that the random-personalized algorithm fails to outperform FedAvg [20] in all cases, implying that random-personalized layers do not grant utility improvement as it comes with significant randomness. The fixed-personalized algorithm fails to outperform FedAvg [20] when with a small number + +![](images/729b0033d059f86faab17a92dfd9c92d05c756c54c0e5cc957457802375c19f3.jpg) +Fig. 3: Framework of PFEDEDIT. The federated learning still works as usual. That is, first, clients train their local models and upload their model weights to the global server. Then, the global server aggregates all model weights and returns the new global model weight to clients. Then, the main change is how PFEDEDIT merges global model with the local, which has two steps: i) PFEDEDIT generates a list of candidate model editorings, and ii) PFEDEDIT locates the top-k important model layers and outputs the edited model for the next-round local training. + +of personalized layers, and its utility is quite similar to the random-choosing algorithm. Once the number of personalized layer gets 17, which means all batch normalization layers are locally personalized, the utility improvement increases significantly and it outperforms FedAvg [18] by $13\%$ , which implies that fixed-personalized algorithm may require all certain layers to be involved to show any utility improvement. + +# 4 Method + +In this section, we introduce our method, PFEDEDIT, which leverages model editing to improve model accuracy for heterogeneous data distributions. Figure 3 shows the entire workflow of PFEDEDIT. The training still follows a traditional FL, i.e., each client uploads their local model to the server for aggregation, and then the server returns the aggregated model. Then, when the client receives the global model, there are mainly two steps in PFEDEDIT, model editing and locating personalized layers, to update the client model based on the global model. + +# 4.1 Step I: Generating Candidate Model Editings + +The purpose of this step is to generate a list of possible editorings of the local model based on the global model returned from the FL server. Specifically, PFEDEDIT considers the local model "clean model" as it is trained on local distribution. On the opposite, it considers the global model "corrupted model" as in non-i.i.d. scenarios, different models bias global model weights towards their local distributions, causing it not generalizing well on each local distribution. In the model editing step, as Figure 3 shows, PFEDEDIT hypothesizes all editorings that tailor the "corrupted model" towards local distribution. + +We show the pseudocode of this step in Algorithm $\mathbb{I}$ . Line 5-6. For each editing, PFEDEDIT replaces one layer $\{h_t^g | t \in (1, T)\}$ , where $T$ is the total number of model layers, from the global model $f_{\theta^g}$ with the corresponding layer $h_t^l$ from the local model $f_{\theta^l}$ . Then the edited model is defined as $S(\theta^*) = (S(\theta^g) \setminus \{h_j^g\}_{j=0}^k) \cup \{h_j^l\}_{j=0}^k$ . As the global model is equipped with the local model layer, which is optimized on local distribution, the global model is tailored towards local distribution. + +The first editing happens after the first round (Line 1-2), right after each client receives the aggregated global model weights (Line 4) and before they start next-round local training (Line 14). The reason is that the global model starts to bias towards other distributions starting from the second round as its weights is uniformly initialized in the first round (Line 1). + +Algorithm 1 PFEDEDIT +Input: client i's initialized global model $\theta^g$ , communication round $R$ , number of model layers $T$ , number of layers to be edited $k$ , local training set $D_{train}$ , subset $D_p$ , $z = (x,y) \in D_p$ , model layer set $S(\cdot)$ , dictionary $W$ +Output: Edited global model weights $\theta^*$ for $R$ -th round +1: $\theta^l \gets$ train $\theta^g$ on $D_{train}$ +2: $\theta^g \gets \frac{1}{n} \sum^n \theta_i^l$ +3: for $r$ in range $(1,R+1)$ do +4: for $h_t^g$ in $S(\theta^g)$ do +5: $S(\theta^*) \gets (S(\theta^g) \backslash h_t^g) \cup h_t^l$ +6: $g(\theta_i^l, \theta_i^*, z)$ , $(f_{\theta^*}(x), y) \gets f_{\theta^*}(z)$ +7: $\varphi \gets (g(\theta_i^l, \theta_i^*, z), (f_{\theta^*}(x), y))$ +8: $W[h_t^g] \gets \varphi$ +9: end for +10: $W \gets$ Lexicographic_rank(W) +11: for $j \in$ range $(0,k)$ do +12: $m \gets W_j.\text{key}()$ +13: $S(\theta^*) \gets (S(\theta^g) \backslash h_m^g) \cup h_m^l$ +14: end for +15: if $r == R$ then +16: return $\theta^*$ +17: else +18: $\theta^l \gets$ train $\theta^*$ on $D_{train}$ +19: $\theta^g \gets \frac{1}{n} \sum^n \theta_l^l$ +20: end if +21: end for + +# 4.2 Step II: Locating Personalized Layers by Ranking Editings + +The purpose of this step is to access all possible editorings from Step I and then locate a list of personalized layers. The detailed substeps are two-fold. First, PFEDEDIT evaluates model editing's effectiveness based on a Total Effectiveness (TE) Score (Step II-a). Second, PFEDEDIT generates prediction lists based on TE score and ground-truth mask to reflect changes in probability and predicted labels (Step II-b). Lastly, PFEDEDIT ranks all prediction lists following the lexicographic order and then selects the top-k best layer editorings as personalized layers for each client based on the ranking (Step II-c). + +Step II-a: Evaluating model editing effectiveness. This substep is to calculate a so-called Total Effectiveness (TE) score to evaluate mode editing effectiveness. Specifically, PFEDEDIT queries local model $f(\theta^l)$ using a representative subset of the training set, denoted as $D_p$ , which maintains the same distribution as the full training set $D_{train}$ and $|D_p| = p * |D_{train}|$ , where $p \in (0,1)$ . PFEDEDIT considers that model weights $\theta^l$ represents local distribution as it is trained on pure local data. PFEDEDIT then queries each edited model $f(\theta^*)$ using the same subset to see the prediction of tailored weights on local distribution. Then by comparing probability outputs from both local and edited weights on the same image samples, PFEDEDIT measures the effectiveness of that editing as well as the capability of edited layer in improving model accuracy. Specifically, when querying the edited model $f(\theta^*)$ , PFEDEDIT collects both TE scores and ground-truth masks, Line 7 in Algorithm 1. The Total Effectiveness (TE) score, as defined below, reflects the probability change on the ground-truth class while the ground truth mask, which is a boolean list indicating whether the predicted label matches the ground truth label, reflects the predicted label change. + +Now we describe the definition of TE score. Given model weight $\theta$ and training dataset $D$ with $c$ classes, we denote by $P(y|\theta ,x)$ the probability (softmax) of the ground-truth class $y$ . We define the total effectiveness (TE) score of editing the model weight $\theta$ into the new weight $\theta^{*}$ given dataset $D$ is defined as: + +$$ +g (\theta , \theta^ {*}, D) = \frac {1}{N} \sum_ {x, y \in D} \frac {P (y | \theta^ {*} , x)}{P (y | \theta , x)} - 1 \tag {2} +$$ + +A positive TE score represents an increased probability on ground-truth label while a negative TE score signifies a decrease. + +Step II-b: Generating prediction lists based on TE and ground-truth mask. This sub-task is to generate prediction lists based on TE and ground-truth mask. Specifically, the TE score, together with the ground-truth mask, makes up four combinations in total, and we call the list of all four cases a prediction list, denoted by $\varphi$ . The prediction list then serves as a reference for evaluating the editing as well as the capability of such layer in improving model's accuracy. Below we show the prediction list that contains all four combinations of TE score and ground-truth mask in Equation 3 + +$$ +\varphi = \left\{ \begin{array}{l} f _ {\theta^ {*}} (x) = y, g \left(\theta , \theta^ {*}, z\right) > 0 \\ f _ {\theta^ {*}} (x) = y, g \left(\theta , \theta^ {*}, z\right) < 0 \\ f _ {\theta^ {*}} (x) \neq y, g \left(\theta , \theta^ {*}, z\right) > 0 \\ f _ {\theta^ {*}} (x) \neq y, g \left(\theta , \theta^ {*}, z\right) < 0 \end{array} \right. \tag {3} +$$ + +The top case of the prediction list $\varphi$ shows the best editing (predicted label same as ground-truth label, TE score greater than 0) and bottom case shows the worst editing effect (predicted label different to ground-truth label, TE score less than 0). PFEDEDIT collects values for such prediction list for each editing by splitting which case each image prediction belongs to, and then ranks all prediction lists using Lexicographic order defined below. + +![](images/986db9a589683faa627a6a519f630ff92043d1593067de60cd0491f5fc85530b.jpg) +(a) Lyme + +![](images/6711746bb6761d3305c801b171f34a67ad5ecf9284bbe92b164002d91cf35d71.jpg) +(b) CIFAR-10 +Fig. 4: Non-i.i.d. data distribution (2 classes per client and 20 classes for CIFAR-100) among clients in Section 5.2 + +![](images/a24b23fbca0c61f39f108c57665b4cd1985b649e3dcd5248fcb9a2fe4689a3e3.jpg) +(c) CIFAR-100 + +Step II-c: Locating top- $k$ personalized layers using lexicographic order. This sub-step is to rank all prediction lists $(\varphi)$ using lexicographic order and select the top- $k$ as the personalized layers. Here is the definition of lexicographic order, which is used to rank $\varphi$ . That is, given two prediction lists $\varphi_1, \varphi_2$ for two model editorings on layer $l_1, l_2$ , respectively, say $\varphi_1 = a_1, \ldots, a_4$ , $\varphi_2 = b_1, \ldots, b_4$ , where $a$ and $b$ are possible cases from prediction lists, the order of those two lists depends on the numeric order of the cases in the first place $i$ where the two case counts differ. That is, $\varphi_1 < \varphi_2$ if and only if $a_i < b_i$ in the natural order of numbers. + +Note that the advantages of using such metric here are two-folds: i) We record $\varphi_{j}(\theta, \theta^{*}, z)$ by counting the number of each case and compare any two prediction lists following the Lexicographic order instead of a summation manner, i.e., $\sum_{|D|}^{i} \varphi_{j}(\theta, \theta^{*}, z)$ , so that the impact of one case is not overridden by another, even if there are insufficient instances. ii) The case order listed in Equation 3 from top to the bottom, is intuitive and aligns with accuracy metric, thus implying that the editing with highest prediction list ranking comes with a largest accuracy improvement. We also discuss and compare our metrics with other potential ones, and evaluation details are shown in Section 5.4. + +# 5 Evaluation + +In this section, we compare PFEDEDIT with baselines in terms of the accuracy and convergence. We then perform a study on parameters used in PFEDEDIT. The evaluation on computation overhead and complexity of PFEDEDIT can be found in Appendix A.4 + +# 5.1 Experimental Setup + +We describe four aspects, i.e., setup of federated learning, datasets, models, and then baselines, for our experimental setup. + +FL Setup. Our default settings follow state-of-the-art works [2,6] in evaluating federated learning. That is, there are 10 clients to participate in a FL scenario with 100 communication rounds. The global sampling rate is set to be 1.0 in each communication round and each client has one local training epoch. The default + +value $(k)$ in selecting and editing top layers is $7\%$ , and the default representative subset ratio $(p)$ is 0.1. Both values are selected based on our parameter study (Section 5.4). By default, personalized layers are chosen each round, and the comparison with fixed personalized layer setting can be found in Appendix A.2. Datasets. We evaluate PFEDEDIT on three benchmark datasets to show its performance on different domains. Specifically, we choose i). CIFAR-10 [12], a benchmark dataset that contains 60,000 $32\times 32$ color images in 10 classes; ii). CIFAR-100 [12], a benchmark dataset that contains 60,000 $32\times 32$ color images in 100 classes; and iii). Lyme [35] dataset, a skin disease dataset that contains 3,027 skin disease images in 4 classes that are confusers for Lyme. We resize CIFAR-10 image samples to $64\times 64$ , CIFAR-100 samples to $224\times 224$ following [36], and Lyme samples to $255\times 255$ following [35]. + +To show that PFEDEDIT can perform on various non-i.i.d. scenario, we adopt the Distributional Heterogeneity level (DH-level) from Yuan et al. [36] to split data distribution into different levels. Specifically, we first assign certain numbers of classes to each client $i$ , and then for each local class $c$ , we sample $D_{i,c} \in (0.4, 0.6)$ from $D_c$ . Each client is assigned $\frac{D_{i,c}}{\sum D_{n,c}}$ of samples in class $c$ . The DH-level ranges from $0\%$ , which represents an i.i.d. scenario where each client gets all classes of data, all the way to $100\%$ , which means that each client only gets one class of data. (10 classes for CIFAR-100). For Lyme [35] dataset, since it only possesses four classes, we only conduct evaluations on $50\%$ (2 classes per clients) and $100\%$ (4 classes per clients) DH-level distribution scenarios. Unless specified otherwise, the default data distribution among clients is 2 classes per client (20 classes per client for CIFAR-100) in our evaluation. We demonstrate the default heterogeneous data distribution for each dataset in Figure 4. We also consider feature space non-i.i.d. scenario, and the evaluation, in comparison with Cd2-pfed [27] and Partialfed-fix [29], can be found in Appendix A.3. Models. For model selection, we choose ResNet18 [9] for evaluation on CIFAR 10 and CIFAR 100, and ViT [8] for Lyme. We allow ResNet18 model to have pretrained weights and allow ViT to have b16 weights that is fine-tuned on Image1k. We set the learning rate for ViT to be 1e-3 and we follow FedPer [2] to set the learning rate as 1e-2 for ResNet18. + +Baselines. We compare PFEDEDIT with FedAvg [20], MOON [14], and other five SOTA personalized FL works as baselines, namely, FedPer [2], FedRep [4], FedBN [18], APFL [6], and DisTrans [36]. We allow each baseline to have the same data distribution and model selection as PFEDEDIT. We consider all layer normalization layers personalized for FedBN [18] when employing ViT. For other parameters, we follow their default paper settings to assign values. The accuracy comparison with more baselines (Fedfa [38], FedALA [37], Fed-LAMB [10], FedRolex [1], and HeteroFL [7]) can be found in Appendix A.1. + +# 5.2 Accuracy Comparison + +As Table 1 shows, PFEDEDIT outperforms other state-of-the-art works on all three datasets under different DH-level data distributions. Compared to the + +baseline FedAvg, the accuracy improvement ranges from 0.007 to 0.266, with an average 0.09 accuracy improvement, and the largest accuracy improvement gets to 0.156 and average improvement gets to 0.05 when compared to other personalized FL works. The accuracy gap between PFEDEDIT and other STOA personalized works in CIFAR-10 is not as huge as in other two datasets, and the reason is that we allow all works to employ models with pretrained weights, which provides a greater lower bound for model accuracy and reduces the accuracy gap between different methods. + +We then evaluate PFEDEDIT, together with other SOTA personalized FL methods under different DH levels. According to Table II, PFEDEDIT consistently achieves superior accuracy compared to other SOTA works under all data distributions, and we have two observations here. First, some personalized FL methods, e.g., FedPer, FedBN, and FedRep, fail to outperform FedAvg under i.i.d. scenarios, which is attributed to their natures as personalized FL works: they are strategically designed to align the global model with local distributions under non-i.i.d. scenarios. If the data distribution is i.i.d., where there is a close alignment between the global model distribution and each local one, the personalized layers may block favorable updates from other clients. On the other hand, PFEDEDIT outperforms FedAvg under i.i.d. scenarios, and the reason is that PFEDEDIT allow clients to update the weights of personalized layers so that they can favor other clients when local distributions are similar. Second, the accuracy improvement brought by PFEDEDIT compared to FedAvg increases as the DH-level increases. The reason is that as the distribution becomes more non-i.i.d., the gap between the global and local model distributions widens. Personalized layers help bridge this gap, tailoring the global model towards the local distribution and improving its generalization on local data. + +It is noteworthy to mention that PFEDEDIT gets 0.989 accuracy on CIFAR-10 when DH-level gets to $100\%$ . The reason is that, in an extreme non-i.i.d. scenario where clients only get one class of data samples, the local model is very likely to get 1.0 testing accuracy and the model editing effectively improves the model accuracy as the local model distribution is highly matched with the local data distribution. + +# 5.3 Convergence + +In this section we evaluate the convergence rate of PFEDEDIT. Specifically, we compare PFEDEDIT with FedAvg and we demonstrate their best accuracy on each dataset and the numbers of rounds to achieve it. As Table 2 shows, PFEDEDIT converges much faster than FedAvg on all three datasets as it only requires 18 epochs to converge on CIFAR-100, 13 epochs on CIFAR-10, and 25 epochs on Lyme. On the opposite, FedAvg takes around three times numbers of rounds to get a suboptimal accuracy on both CIFAR-10 and CIFAR-100. For Lyme, it takes 87 epochs to get accuracy above 0.7, which only takes 4 epochs for PFEDEDIT. The reason is that FedAvg does not provide convergence guarantee as it aggregates all model weights each round without having any optimization, + +Table 1: Evaluation of PFEDEDIT with other SOTA works on different datasets. + +
DatasetDH-levelFedAvg 20FedPer 2FedBN 18FedRep 4DisTrans 36MOON 14APFL 6PFEDEDIT
CIFAR-100 % (i.i.d)0.9420.8900.92720.8890.9470.9220.8910.955
40%0.9230.8730.9260.8840.9440.9310.8890.953
60%0.9390.8800.9400.9200.9600.9270.9100.965
80%0.8060.8870.9450.9330.9540.9380.9150.962
100%0.8300.9020.9610.9800.9720.9400.9270.989
CIFAR-1000 % (i.i.d)0.7830.7330.7460.7420.7880.7460.7440.790
40%0.7790.7530.7760.7470.7790.7620.7530.825
60%0.7720.7480.8020.7930.8300.7680.8020.855
80%0.7370.7690.8620.8590.8870.7700.8330.890
100%0.6640.7900.9040.8800.9030.7730.8590.930
Lyme0%0.7070.7000.7330.7300.7450.6990.7220.765
50 % (i.i.d)0.7160.8080.8210.8330.8500.7680.7810.912
+ +Table 2: Best Accuracy of PFEDEdit/FedAvg and number of rounds to achieve it. (- represents not achieved within 100 rounds) + +
Dataset| PFEDEDIT/FedAvg# Rounds for target accuracy (PFEDEDIT/FedAvg)
>0.6>0.7>0.75>0.8>0.85>0.9
CIFAR-100.962/0.8064/56/87/139/3311/-13/-
CIFAR-1000.890/0.7376/98/2410/-12/-18/--/-
Lyme0.912/0.7163/324/875/-5/-9/-25/-
+ +Table 3: Accuracy of PFEDEDIT with different editing ratios (in comparison with FedAvg). + +
Dataset| FedAvg| PFEDEDIT
1%3%5%7%9%
CIFAR-100.8060.9330.9470.9480.9620.952
CIFAR-1000.7480.8600.8780.8850.8900.874
Lyme0.7160.8680.8800.8890.9120.910
+ +and PFEDEDIT edits the global model and tailors it towards the local distribution, which makes it faster to generalize on local distributions. + +# 5.4 Parameter Study + +Model Editing Ratio $(k)$ PFEDEDIT allows users to pre-define $k$ , the number of layers to be edited. It then edits the top- $k$ layers, which are considered to be the most effective in improving model accuracy. In this section we set up an experiment to show the correlation between the number of modified layers and the model accuracy. Intuitively, if the number of edited layers increases, the model accuracy gets higher as greater part of the model is tailored towards local distribution. However, such increase may not hold all the time as there are layers that do not possess any learnable parameters and the model may converge. + +We choose the ratios of edited layers to be $1\%$ , $3\%$ , $5\%$ , $7\%$ , and $9\%$ to the total model layers and we round up the number in each case. Shown in Table 3, as the editing ratio increases, the overall testing accuracy gets higher. The accuracy improvement is greater from improving the ratio from $1\%$ to $3\%$ and the increasing rate gets flatter when keep improving the ratio from $3\%$ to + +$5\%$ , or $5\%$ to $7\%$ . At the end, when the ratio is set from $7\%$ to $9\%$ , the accuracy starts to drop a little bit, implying that the model may converge on that accuracy. Compared to the baseline, PFEDEDIT improves the model accuracy from 0.13, with editing $1\%$ layers, to 0.16, with editing $7\%$ layers, on three datasets on average. + +Table 4: Evaluation of PFedEdit with representative subset ratio (i.e., $p$ ). + +
DatasetSubser ratio
p=0.01p=0.025p=0.05p=0.08p=0.1
CIFAR-100.9510.9530.9540.9540.962
CIFAR-100-0.8750.8840.8860.890
+ +Table 5: Evaluation of PFedEdit using different metrics in assessing model editing + +
DatasetDifferent metrics
AccuracyLossTE scorePrediction list
CIFAR-100.5400.9050.8850.962
CIFAR-1000.7050.8860.8210.890
+ +Representative Subset Ratio $(p)$ As mentioned in Section 4, PFEDEDIT queries the model using a representative subset $D_{p}$ of the training set $D_{train}$ . PFEDEDIT employs a ratio $p$ to pre-define the subset size compared to the training set, $|D_{p}| = p * |D_{train}|$ . We want to ensure that the subset is representative enough for potential testing image samples while not introducing too much computation overhead. Therefore, we conduct a study on the size of the subset versus the performance of PFEDEDIT. Intuitively, a larger size covers a broader local data distribution and leads to a more accurate layer locating and thus a higher model accuracy. However, on the other hand, taken a larger subset from the original training set reduces the size of the remaining set. We want to find an optimal size of the subset that is large enough to give a complete representation of local data distribution but also leaves enough samples for the local training. We split the local training set into a subset that takes up of $p * |D_{train}|$ image samples, and a training set that takes the rest $|D_{train}^{*}| = |D_{train}| - |D_{p}|$ image samples. To control the variable, we fix the size of $D_{train}^{*}$ , and we choose different $p$ values from $\{0.1, 0.08, 0.05, 0.025, 0.01\}$ to adjust the subset size. We conduct the evaluation on CIFAR-10 and CIFAR-100 [12], with 10 clients and each client is assigned 2 classes of image samples. Each client possesses 4,500 image samples as training set, and the subset size varies from 50 image samples ( $p$ value 0.01) to 500 image samples ( $p$ value 0.1). Note that the $p$ value starts from 0.025 for CIFAR-100 because a subset that is 0.01 times the size of training set does not include image samples from all classes, i.e., 50 image samples do not cover 100 classes thus the subset is not representative. + +As Table 4 shows, the overall trend of the testing accuracy keeps increasing as $p$ value gets larger from 0.01 times the training set to 0.1 times the training set for both datasets. However, the overall accuracy improvement is minimal compared to the subset size increase, demonstrating the capability of PFEDEDIT in improving model accuracy with a small size subset. For CIFAR-10, as $p$ value increases from 0.01 to 0.08, which is an 8 times increase, the model accuracy is improved from 0.951 to 0.954, which is a $0.3\%$ improvement. As $p$ value increases from 0.08 to 0.1, the accuracy is further improved from 0.954 to 0.962, which is a $0.8\%$ increase. For CIFAR-100, the accuracy improvement brought by increasing $p$ value is greater than that in CIFAR-10 as the model accuracy is improved from + +0.875 to 0.890 when $p$ value changes from 0.025 to 0.1. Therefore, PFEDEDIT is computation-friendly as it requires only 50 image samples to get a 0.951 model accuracy and 500 image samples to achieve a 0.962 accuracy on CIFAR-10. Similarly, PFEDEDIT requires only 125 image samples as the subset to achieve a 0.875 accuracy and 500 image samples to gets a 0.890 accuracy on CIFAR-100. + +Metrics Selection In this section, we conduct experiments to explore possible metric selections for PFEDEDIT when assessing each model editing. Besides the prediction list that demonstrates all four combinations of TE score and ground-truth mask, we also choose model accuracy, loss value, and TE score. We select cross-entropy loss as our loss metric, which is defined as $H(P,Q) = -\sum_{x \in \mathcal{X}} p(x) \log q(x)$ , given two probability distribution $P$ and $Q$ , and the TE score is introduced in Section 4. + +As Table 5 shows, the prediction list employed by PFEDEDIT outperforms other metrics and the accuracy gap ranges from 0.057 to 0.422. Among all possible metrics, using accuracy to assess model editorings leads to the lowest accuracy, 0.540 on CIFAR-10 and 0.705 on CIFAR-100, lower than the baseline FedAvg [20]. The reason is that the effects of some editorings are subtle and they only cause the probability change but not necessarily results in label change. As the way that accuracy is computed only counts on samples that are correctly classified, small probability change that do not alter the final prediction will be considered as no improvement at all. On the other hand, the TE score reflects the probability change on ground-truth label, but it does not reflects any label change. In other words, an increase in the probability of the ground-truth label suggests a positive editing, but a decrease in probability does not necessarily indicate a negative editing. From the accuracy perspective, the edited model performs identically to the original model despite these probability changes. + +While employing loss functions as the metric provides insights into the probability of the ground-truth class, the averaged loss value may exhibit bias if outliers have a disproportionate impact on the probability of the ground-truth class. For example, If an editing improves the model accuracy in correctly classifying two samples with a slight increase in the ground-truth class and misclassifies one sample with a notable drop in the ground-truth class, the improvement introduced by the correct classifications might be nullified by the incorrect one, which leads to a biased averaged loss value with respect to each data sample. + +# 6 Conclusion + +In this paper, we design the first personalized federated learning framework that utilizes automated model editing as a core strategy to improve the model's utility. Our approach leverages layer-specific editing, which automatically identifies specific layers to be edited based on local data distribution, allowing for more effective adaptation of the global model to heterogeneous data distributions across various clients. Our evaluation demonstrates that PFEDEDIT significantly improves the accuracy of the global model from non-i.i.d. to i.i.d. data distribution. + +# Acknowledgments + +This work was supported in part by National Science Foundation (NSF) under grants OAC-23-19742 and Johns Hopkins University Institute for Assured Autonomy (IAA) with grants 80052272 and 80052273. The views and conclusions contained herein are those of the authors and should not be interpreted as necessarily representing the official policies or endorsements, either expressed or implied, of NSF or JHU-IAA. + +# References + +1. Alam, S., Liu, L., Yan, M., Zhang, M.: Fedrolex: Model-heterogeneous federated learning with rolling sub-model extraction (2023) +2. Arivazhagan, M.G., Aggarwal, V., Singh, A.K., Choudhary, S.: Federated learning with personalization layers. arXiv preprint arXiv:1912.00818 (2019) +3. Basu, S., Zhao, N., Morariu, V., Feizi, S., Manjunatha, V.: Localizing and editing knowledge in text-to-image generative models (2023) +4. Collins, L., Hassani, H., Mokhtari, A., Shakkottai, S.: Exploiting shared representations for personalized federated learning. In: Proceedings of the 38th International Conference on Machine Learning (2021) +5. Dai, D., Dong, L., Hao, Y., Sui, Z., Chang, B., Wei, F.: Knowledge neurons in pretrained transformers. arXiv preprint arXiv:2104.08696 (2021) +6. Deng, Y., Kamani, M.M., Mahdavi, M.: Adaptive personalized federated learning (2020) +7. Diao, E., Ding, J., Tarokh, V.: Heterofl: Computation and communication efficient federated learning for heterogeneous clients (2021) +8. Dosovitskiy, A., Beyer, L., Kolesnikov, A., Weissenborn, D., Zhai, X., Unterthiner, T., Dehghani, M., Minderer, M., Heigold, G., Gelly, S., Uszkoreit, J., Houlsby, N.: An image is worth 16x16 words: Transformers for image recognition at scale (2021) +9. He, K., Zhang, X., Ren, S., Sun, J.: Deep residual learning for image recognition. In: Proceedings of the IEEE conference on computer vision and pattern recognition. pp. 770-778 (2016) +0. Karimi, B., Li, P., Li, X.: Layer-wise and dimension-wise locally adaptive federated learning (2022) +1. Karimireddy, S.P., Kale, S., Mohri, M., Reddi, S., Stich, S., Suresh, A.T.: Scaffold: Stochastic controlled averaging for federated learning. In: International conference on machine learning (2020) +2. Krizhevsky, A.: Learning multiple layers of features from tiny images. Tech. rep. (2009) +3. Li, L., Fan, Y., Tse, M., Lin, K.Y.: A review of applications in federated learning. Computers & Industrial Engineering (2020) +4. Li, Q., He, B., Song, D.: Model-contrastive federated learning. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (2021) +5. Li, T., Sahu, A.K., Talwalkar, A., Smith, V.: Federated learning: Challenges, methods, and future directions. IEEE signal processing magazine (2020) +6. Li, T., Sahu, A.K., Zaheer, M., Sanjabi, M., Talwalkar, A., Smith, V.: Federated optimization in heterogeneous networks. Proceedings of Machine learning and systems 2, 429-450 (2020) + +17. Li, X., Li, S., Song, S., Yang, J., Ma, J., Yu, J.: Pmet: Precise model editing in a transformer. arXiv preprint arXiv:2308.08742 (2023) +18. Li, X., Jiang, M., Zhang, X., Kamp, M., Dou, Q.: Fed{bn}: Federated learning on non-{iid} features via local batch normalization. In: International Conference on Learning Representations (2021), https://openreview.net/pdf?id=6YEQUnOQICG +19. Liu, Q., Chen, C., Qin, J., Dou, Q., Heng, P.A.: Feddg: Federated domain generalization on medical image segmentation via episodic learning in continuous frequency space. The IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR) (2021) +20. McMahan, B., Moore, E., Ramage, D., Hampson, S., y Arcas, B.A.: Communication-efficient learning of deep networks from decentralized data. In: Artificial intelligence and statistics (2017) +21. Meng, K., Bau, D., Andonian, A., Belinkov, Y.: Locating and editing factual associations in GPT. Advances in Neural Information Processing Systems (2022) +22. Meng, K., Bau, D., Andonian, A., Belinkov, Y.: Locating and editing factual associations in GPT. Advances in Neural Information Processing Systems 35 (2022) +23. Meng, K., Sen Sharma, A., Andonian, A., Belinkov, Y., Bau, D.: Mass editing memory in a transformer. The Eleventh International Conference on Learning Representations (ICLR) (2023) +24. Nguyen, D.C., Ding, M., Pathirana, P.N., Seneviratne, A., Li, J., Vincent Poor, H.: Federated learning for internet of things: A comprehensive survey. IEEE Communications Surveys & Tutorials (2021) +25. Nguyen, D.C., Ding, M., Pham, Q.V., Pathirana, P.N., Le, L.B., Seneviratne, A., Li, J., Niyato, D., Poor, H.V.: Federated learning meets blockchain in edge computing: Opportunities and challenges. IEEE Internet of Things Journal (2021) +26. Nori, H., Jenkins, S., Koch, P., Caruana, R.: Interpretml: A unified framework for machine learning interpretability. arXiv preprint arXiv:1909.09223 (2019) +27. Shen, Y., Zhou, Y., Yu, L.: $\mathrm{Cd^{2}}$ -pfed: Cyclic distillation-guided channel decoupling for model personalization in federated learning (2022) +28. Shokri, R., Shmatikov, V.: Privacy-preserving deep learning. In: Proceedings of the 22nd ACM SIGSAC conference on computer and communications security (2015) +29. Sun, B., Huo, H., YANG, Y., Bai, B.: Partialfed: Cross-domain personalized federated learning via partial initialization. In: Advances in Neural Information Processing Systems (2021) +30. Wang, P., Zhang, N., Xie, X., Yao, Y., Tian, B., Wang, M., Xi, Z., Cheng, S., Liu, K., Zheng, G., et al.: Easilyedit: An easy-to-use knowledge editing framework for large language models. arXiv preprint arXiv:2308.07269 (2023) +31. Yang, Q., Liu, Y., Chen, T., Tong, Y.: Federated machine learning: Concept and applications. ACM Transactions on Intelligent Systems and Technology (TIST) (2019) +32. Yang, Y., Hui, B., Yuan, H., Gong, N., Cao, Y.: PrivateFL: Accurate, differentially private federated learning via personalized data transformation. In: 32nd USENIX Security Symposium (USENIX Security 23). pp. 1595-1612 (2023) +33. Yang, Y., Yuan, H., Hui, B., Gong, N., Fendley, N., Burlina, P., Cao, Y.: Fortifying federated learning against membership inference attacks via client-level input perturbation. In: 2023 53rd Annual IEEE/IFIP International Conference on Dependable Systems and Networks (DSN) (2023) +34. Yang, Z., Zhang, Y., Zheng, Y., Tian, X., Peng, H., Liu, T., Han, B.: Fedfed: Feature distillation against data heterogeneity in federated learning. In: Thirty-seventh Conference on Neural Information Processing Systems (2023) + +35. Yuan, H., Aucott, J., Hadzic, A., Paul, W., Villegas de Flores, M., Mathew, P., Burlina, P., Cao, Y.: Edgemixup: Embarrassingly simple data alteration to improve Lyme disease lesion segmentation and diagnosis fairness. In: Medical Image Computing and Computer Assisted Intervention - MICCAI 2023 (2023) +36. Yuan, H., Hui, B., Yang, Y., Burlina, P., Gong, N.Z., Cao, Y.: Addressing heterogeneity in federated learning via distributional transformation. In: Computer Vision - ECCV 2022: 17th European Conference, Tel Aviv, Israel, October 23-27, 2022, Proceedings, Part XXXVIII. p. 179-195 (2022) +37. Zhang, J., Hua, Y., Wang, H., Song, T., Xue, Z., Ma, R., Guan, H.: Fedala: Adaptive local aggregation for personalized federated learning. In: Proceedings of the AAAI Conference on Artificial Intelligence (2023) +38. Zhou, T., Konukoglu, E.: FedFA: Federated feature augmentation. In: The Eleventh International Conference on Learning Representations (ICLR) (2023) \ No newline at end of file diff --git a/2024/PFedEdit_ Personalized Federated Learning via Automated Model Editing/images.zip b/2024/PFedEdit_ Personalized Federated Learning via Automated Model Editing/images.zip new file mode 100644 index 0000000000000000000000000000000000000000..3476e2b38a35edf6be2a11b9c28423f6beb3b030 --- /dev/null +++ b/2024/PFedEdit_ Personalized Federated Learning via Automated Model Editing/images.zip @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:f69ab03ef0fc08442cca96cb8fd075fc5ebd6687e03e7fc45f718710ee296254 +size 262720 diff --git a/2024/PFedEdit_ Personalized Federated Learning via Automated Model Editing/layout.json b/2024/PFedEdit_ Personalized Federated Learning via Automated Model Editing/layout.json new file mode 100644 index 0000000000000000000000000000000000000000..267ff15b4ddac2a51f89a8eb81e9aafdad2a42c4 --- /dev/null +++ b/2024/PFedEdit_ Personalized Federated Learning via Automated Model Editing/layout.json @@ -0,0 +1,9528 @@ +{ + "pdf_info": [ + { + "para_blocks": [ + { + "bbox": [ + 138, + 111, + 477, + 148 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 111, + 477, + 148 + ], + "spans": [ + { + "bbox": [ + 138, + 111, + 477, + 148 + ], + "type": "text", + "content": "PFEDEDIT: Personalized Federated Learning via Automated Model Editing" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 157, + 167, + 457, + 192 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 157, + 167, + 457, + 192 + ], + "spans": [ + { + "bbox": [ + 157, + 167, + 457, + 192 + ], + "type": "text", + "content": "Haolin Yuan" + }, + { + "bbox": [ + 157, + 167, + 457, + 192 + ], + "type": "inline_equation", + "content": "^{1}" + }, + { + "bbox": [ + 157, + 167, + 457, + 192 + ], + "type": "text", + "content": ", William Paul" + }, + { + "bbox": [ + 157, + 167, + 457, + 192 + ], + "type": "inline_equation", + "content": "^{2}" + }, + { + "bbox": [ + 157, + 167, + 457, + 192 + ], + "type": "text", + "content": ", John Aucott" + }, + { + "bbox": [ + 157, + 167, + 457, + 192 + ], + "type": "inline_equation", + "content": "^{3}" + }, + { + "bbox": [ + 157, + 167, + 457, + 192 + ], + "type": "text", + "content": ", Philippe Burlina" + }, + { + "bbox": [ + 157, + 167, + 457, + 192 + ], + "type": "inline_equation", + "content": "^{2}" + }, + { + "bbox": [ + 157, + 167, + 457, + 192 + ], + "type": "text", + "content": ", and Yinzhi Cao" + }, + { + "bbox": [ + 157, + 167, + 457, + 192 + ], + "type": "inline_equation", + "content": "^{1}" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 181, + 201, + 431, + 213 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 181, + 201, + 431, + 213 + ], + "spans": [ + { + "bbox": [ + 181, + 201, + 431, + 213 + ], + "type": "text", + "content": "1 Johns Hopkins University {hyuan4, yinzhi.cao}@jhu.edu" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 194, + 213, + 418, + 225 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 194, + 213, + 418, + 225 + ], + "spans": [ + { + "bbox": [ + 194, + 213, + 418, + 225 + ], + "type": "text", + "content": "2 Johns Hopkins Applied Physics Lab {william.paul," + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 239, + 225, + 374, + 236 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 239, + 225, + 374, + 236 + ], + "spans": [ + { + "bbox": [ + 239, + 225, + 374, + 236 + ], + "type": "text", + "content": "philippe.burlina}@jhuapl.edu" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 168, + 235, + 444, + 247 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 168, + 235, + 444, + 247 + ], + "spans": [ + { + "bbox": [ + 168, + 235, + 444, + 247 + ], + "type": "text", + "content": "3 Johns Hopkins University School of Medicine jaucott2@jhmi.edu" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 160, + 278, + 455, + 498 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 160, + 278, + 455, + 498 + ], + "spans": [ + { + "bbox": [ + 160, + 278, + 455, + 498 + ], + "type": "text", + "content": "Abstract. Federated learning (FL) allows clients to train a deep learning model collaboratively while maintaining their private data locally. One challenging problem facing FL is that the model utility drops significantly once the data distribution gets heterogeneous, or non-i.i.d, among clients. A promising solution is to personalize models for each client, e.g., keeping some layers locally without aggregation, which is thus called personalized FL. However, previous personalized FL often suffer from suboptimal utility because their choice of layer personalization is based on empirical knowledge and fixed for different datasets and distributions. In this work, we design PFEDEDIT, the first federated learning framework that leverages automated model editing to optimize the choice of personalization layers and improve model utility under a variety of data distributions including non-i.i.d. The high-level idea of PFEDEDIT is to assess the effectiveness of every global model layer in improving model utility on local data distribution once edited, and then to apply edits on the top-" + }, + { + "bbox": [ + 160, + 278, + 455, + 498 + ], + "type": "inline_equation", + "content": "k" + }, + { + "bbox": [ + 160, + 278, + 455, + 498 + ], + "type": "text", + "content": " most effective layers. Our evaluation shows that PFEDEDIT outperforms six state-of-the-art approaches on three benchmark datasets by " + }, + { + "bbox": [ + 160, + 278, + 455, + 498 + ], + "type": "inline_equation", + "content": "6\\%" + }, + { + "bbox": [ + 160, + 278, + 455, + 498 + ], + "type": "text", + "content": " on the model's performance on average, with the largest accuracy improvement being " + }, + { + "bbox": [ + 160, + 278, + 455, + 498 + ], + "type": "inline_equation", + "content": "26.6\\%" + }, + { + "bbox": [ + 160, + 278, + 455, + 498 + ], + "type": "text", + "content": ". PFEDEDIT is open-source and available at this repository: https://github.com/Haolin-Yuan/PFedEdit" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 132, + 520, + 230, + 533 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 520, + 230, + 533 + ], + "spans": [ + { + "bbox": [ + 132, + 520, + 230, + 533 + ], + "type": "text", + "content": "1 Introduction" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 130, + 545, + 483, + 666 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 545, + 483, + 666 + ], + "spans": [ + { + "bbox": [ + 130, + 545, + 483, + 666 + ], + "type": "text", + "content": "Federated learning (FL) [13,15,20,31,33] has emerged as a widely used training algorithm of deep learning, especially in the image domain, due to its practical applications in real-world scenarios and its impressive performance. A federated learning training scenario entails multiple local clients and a global server. Each client trains their local model using their private image data. After performing local training, each client updates their local model weights to the global server for aggregation with other clients' model weights. Federated learning allows clients to train a collaborative model that performs well on multi-class tasks—even if each client only possesses a few classes of images—via only sharing models but not local training images." + } + ] + } + ], + "index": 8 + } + ], + "discarded_blocks": [], + "page_size": [ + 612, + 792 + ], + "page_idx": 0 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 134, + 114, + 482, + 187 + ], + "blocks": [ + { + "bbox": [ + 134, + 114, + 482, + 187 + ], + "lines": [ + { + "bbox": [ + 134, + 114, + 482, + 187 + ], + "spans": [ + { + "bbox": [ + 134, + 114, + 482, + 187 + ], + "type": "image", + "image_path": "771cb20bf21dae387b6f5b9a601b277237bea59b2b0ccd95e59bd8fa837624f5.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 131, + 197, + 482, + 220 + ], + "lines": [ + { + "bbox": [ + 131, + 197, + 482, + 220 + ], + "spans": [ + { + "bbox": [ + 131, + 197, + 482, + 220 + ], + "type": "text", + "content": "Fig. 1: High-level idea of PFEDEDIT (PFEDEDIT edits the global model returned from FL server to fit the local client's distribution.)" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_caption" + } + ], + "index": 2 + }, + { + "bbox": [ + 130, + 251, + 482, + 370 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 251, + 482, + 370 + ], + "spans": [ + { + "bbox": [ + 130, + 251, + 482, + 370 + ], + "type": "text", + "content": "As the goal of federated learning is to learn a global model for all devices cooperatively, if the data distributions among clients are similar, the global model is supposed to perform better than local models that are solely trained on local data. However, once the data distribution becomes skewed, or non-i.i.d., e.g., some clients may possess image samples of only one or two classes or some possess many more image samples from one class than others, the global model becomes ineffective to be employed as local models. The reason is that the local data distribution at one client could be dramatically different from others, thus causing local model weights to adversely affect each other during the model aggregation." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 130, + 374, + 482, + 555 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 374, + 482, + 555 + ], + "spans": [ + { + "bbox": [ + 130, + 374, + 482, + 555 + ], + "type": "text", + "content": "Multiple works [11,14,16] have been proposed to improve the FL model accuracy under the non-i.i.d. distribution by optimizing local models. They either regularize local models to maximize the agreement between clients [14], or reduce clients' weights drift from the server end [11,16] for a better global model that generalizes on each local distribution. While such works improved model's accuracy, the model is still uniform across different clients, and cannot be tailored based on different clients' data distribution. Therefore, one natural step is a so-called personalized FL [2,4,18,36], which enables each user to have personalized layers that help customize local models to their data distributions. Those personalized layers are retained locally each communication round and are not updated from the global model, thereby preventing being aggregated with others and avoiding bias towards other clients' data distributions. The choices of those personalized layers are various but are all pre-defined before the training, e.g., it can be linear layer [2], batch normalization layer [4], or specially designed layer [36]." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 130, + 558, + 482, + 667 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 558, + 482, + 667 + ], + "spans": [ + { + "bbox": [ + 130, + 558, + 482, + 667 + ], + "type": "text", + "content": "However, there are still two challenges that remain for personalized FLs despite the accuracy improvement. First, the personalized layers chosen in previous personalized FL works [2,4,18,36] are inspired by empirical findings and fixed, implying that those choices may not be optimal when given different data distributions or datasets. That is, while the weights of those layers might be personalized, the locations and types of those layers are not, e.g., either linear or batch normalization layers in certain locations of a learning model. Second, previous personalized works are limited in their application scenarios as they assume certain model structure [18], which may not be used for certain clients." + } + ] + } + ], + "index": 6 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "spans": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "type": "text", + "content": "2" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 166, + 91, + 225, + 100 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 166, + 91, + 225, + 100 + ], + "spans": [ + { + "bbox": [ + 166, + 91, + 225, + 100 + ], + "type": "text", + "content": "H. Yuan et al." + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 1 + }, + { + "para_blocks": [ + { + "bbox": [ + 130, + 116, + 482, + 295 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 116, + 482, + 295 + ], + "spans": [ + { + "bbox": [ + 130, + 116, + 482, + 295 + ], + "type": "text", + "content": "In this paper, to address these two challenges, we design a framework that automatically locates and chooses layers to be personalized. We propose the first personalized federated learning framework, called PFEDEDIT, which leverages automated model editing [3,5,21,23] to optimize the layer choosing and improve model utility under different settings including non-i.i.d. The high-level idea of PFEDEDIT, as shown in Figure 1, is to look for model layers that contain knowledge about local distribution that benefits clients the most, and by editing those layers, PFEDEDIT mitigates the negative influence brought by other clients, thus tailoring the global model weights towards the local model weights in each round. Specifically, PFEDEDIT assesses which layers of the global model are the most effective in maintaining the local distribution and then marks them as target layers. PFEDEDIT then edits the top-" + }, + { + "bbox": [ + 130, + 116, + 482, + 295 + ], + "type": "inline_equation", + "content": "k" + }, + { + "bbox": [ + 130, + 116, + 482, + 295 + ], + "type": "text", + "content": " target layers to tailor the global model weights closer to local models thus improving the model utility. PFEDEDIT does not make any assumptions on model selection and thus is model-structure agnostics." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 130, + 296, + 498, + 367 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 296, + 498, + 367 + ], + "spans": [ + { + "bbox": [ + 130, + 296, + 498, + 367 + ], + "type": "text", + "content": "We performed extensive experiments on three benchmark datasets, namely CIFAR-10 [12], CIFAR-100 [12], and Lyme dataset [35], and we compared PFEDEDIT with FedAvg [20] and other six state-of-the-art personalized federated learning approaches. Our experimental results demonstrate that PFEDEDIT outperforms all SOTA personalized federated learning works on all the datasets, with an average " + }, + { + "bbox": [ + 130, + 296, + 498, + 367 + ], + "type": "inline_equation", + "content": "6\\%" + }, + { + "bbox": [ + 130, + 296, + 498, + 367 + ], + "type": "text", + "content": " accuracy improvement." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 146, + 368, + 347, + 379 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 146, + 368, + 347, + 379 + ], + "spans": [ + { + "bbox": [ + 146, + 368, + 347, + 379 + ], + "type": "text", + "content": "In summary, our contributions are as follows:" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 130, + 381, + 480, + 452 + ], + "type": "list", + "angle": 0, + "index": 8, + "blocks": [ + { + "bbox": [ + 130, + 381, + 480, + 404 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 381, + 480, + 404 + ], + "spans": [ + { + "bbox": [ + 130, + 381, + 480, + 404 + ], + "type": "text", + "content": "- We design PFEDEDIT, the first personalized federated learning framework that optimizes personalized layer selections for each client using model editing." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 130, + 405, + 480, + 427 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 405, + 480, + 427 + ], + "spans": [ + { + "bbox": [ + 130, + 405, + 480, + 427 + ], + "type": "text", + "content": "- PFEDEDIT is adaptable to all model structures because it does not impose any assumption on model selection." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 130, + 430, + 480, + 452 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 430, + 480, + 452 + ], + "spans": [ + { + "bbox": [ + 130, + 430, + 480, + 452 + ], + "type": "text", + "content": "- We conduct extensive experiments to show that PFEDEDIT outperforms other SOTA personalized federated works on different benchmark datasets." + } + ] + } + ], + "index": 7 + } + ], + "sub_type": "text" + }, + { + "bbox": [ + 132, + 472, + 237, + 485 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 472, + 237, + 485 + ], + "spans": [ + { + "bbox": [ + 132, + 472, + 237, + 485 + ], + "type": "text", + "content": "2 Related Work" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 130, + 497, + 482, + 668 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 497, + 482, + 668 + ], + "spans": [ + { + "bbox": [ + 130, + 497, + 482, + 668 + ], + "type": "text", + "content": "Federated Learning. Federated learning [13,15,20,31] (FL) is first proposed to predict user's text input on portable devices and it has been widely deployed in other industries, such as medical imaging [19], Internet of Things (IoT) [24], and block-chain [25]. It allows multiple clients to collaboratively train a model that can perform on samples from any clients' data distribution. Two common approaches in federated learning (FL) are FedSGD [28] and FedAvg [20]. FedSGD optimizes global weights by back-propagating each local gradient, while FedAvg allows for multiple local training epochs and client-side optimization. In FedAvg, the server's role is limited to averaging and distributing the aggregated weights. While FedAvg reduces communication costs compared to FedSGD, it lacks a convergence guarantee, and heterogeneous data distribution among clients can significantly reduce the global model's utility. Therefore, various works [2,4,14,18,32,36] have been proposed to solve such utility degradation caused by data heterogeneity. One direction [14] is to optimize local" + } + ] + } + ], + "index": 10 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 400, + 91, + 447, + 100 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 400, + 91, + 447, + 100 + ], + "spans": [ + { + "bbox": [ + 400, + 91, + 447, + 100 + ], + "type": "text", + "content": "PFEDEDIT" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 474, + 92, + 480, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 474, + 92, + 480, + 100 + ], + "spans": [ + { + "bbox": [ + 474, + 92, + 480, + 100 + ], + "type": "text", + "content": "3" + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 2 + }, + { + "para_blocks": [ + { + "bbox": [ + 130, + 116, + 482, + 177 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 116, + 482, + 177 + ], + "spans": [ + { + "bbox": [ + 130, + 116, + 482, + 177 + ], + "type": "text", + "content": "model either via personalized objective functions; yet, another, called personalized FL [2,4,18,36], allows clients to have personalized model layers that can be kept locally. As a comparison, PFEDEDIT is the first work that optimizes and personalizes the choice of model layers with further improved utilities of FL, especially under non-i.i.d. settings." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 130, + 187, + 482, + 355 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 187, + 482, + 355 + ], + "spans": [ + { + "bbox": [ + 130, + 187, + 482, + 355 + ], + "type": "text", + "content": "Model Editing. Model editing 3.5.17.23.26 involves adding, removing, or adjusting layers and parameters of a model architecture to improve its performance or adapt it to new tasks. Specifically, let " + }, + { + "bbox": [ + 130, + 187, + 482, + 355 + ], + "type": "inline_equation", + "content": "h_i" + }, + { + "bbox": [ + 130, + 187, + 482, + 355 + ], + "type": "text", + "content": " be the " + }, + { + "bbox": [ + 130, + 187, + 482, + 355 + ], + "type": "inline_equation", + "content": "i" + }, + { + "bbox": [ + 130, + 187, + 482, + 355 + ], + "type": "text", + "content": "-th hidden states of any model, " + }, + { + "bbox": [ + 130, + 187, + 482, + 355 + ], + "type": "inline_equation", + "content": "\\{h_l | l \\in [1, L]\\}" + }, + { + "bbox": [ + 130, + 187, + 482, + 355 + ], + "type": "text", + "content": ", where " + }, + { + "bbox": [ + 130, + 187, + 482, + 355 + ], + "type": "inline_equation", + "content": "L" + }, + { + "bbox": [ + 130, + 187, + 482, + 355 + ], + "type": "text", + "content": " is the number of layers in model, model editing allows users to find each hidden state's contribution towards a specific prediction. It mainly contains three runs: i). clean run, in which the statistics of each state are recorded when fed a factual prompt " + }, + { + "bbox": [ + 130, + 187, + 482, + 355 + ], + "type": "inline_equation", + "content": "x" + }, + { + "bbox": [ + 130, + 187, + 482, + 355 + ], + "type": "text", + "content": "; ii). corrupted run, in which each hidden layer is added Gaussian noise such that " + }, + { + "bbox": [ + 130, + 187, + 482, + 355 + ], + "type": "inline_equation", + "content": "h_l \\coloneqq h_l + \\epsilon" + }, + { + "bbox": [ + 130, + 187, + 482, + 355 + ], + "type": "text", + "content": ". Because the hidden layer is obfuscated by the added noise, the model will output incorrect predictions; iii). restoration run, in which the model gives predictions with noised hidden states recovered to clean one at a time to find out which layer is related to producing factual knowledge. Following all three steps, users can find out the causal importance of all model layers in the computation graph and edit weights of those knowledge-related layers to alter the final model output." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 130, + 356, + 483, + 439 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 356, + 483, + 439 + ], + "spans": [ + { + "bbox": [ + 130, + 356, + 483, + 439 + ], + "type": "text", + "content": "There are different applications of model editing. For example, model editing can be employed on large language models (LLMs) to facilitate the interpretation of LLMs as black boxes when generating particular content or to guide LLMs in producing desired outputs [22,30]. ROME [22] edits LLMs' linear layers to modify some learned factual knowledge and make LLMs output desired content such as \"Eiffel Tower is located in the city of Rome\". Meng et al. [23] utilizes Gaussian noise to perturb predictions to achieve mass editing in a transformer." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 130, + 440, + 482, + 523 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 440, + 482, + 523 + ], + "spans": [ + { + "bbox": [ + 130, + 440, + 482, + 523 + ], + "type": "text", + "content": "As a comparison, PFEDEDIT is different from prior model editing works in two aspects. First, PFEDEDIT is a layer-wise model editing approach, which replaces the entire model layer from the local model instead of editing specific neuron stats. The advantage is that PFEDEDIT is able to perform well against potential images from various data distributions. Second, PFEDEDIT considers the global model as corrupted instead of one with Gaussian noise to reduce computation overhead." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 132, + 542, + 211, + 554 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 542, + 211, + 554 + ], + "spans": [ + { + "bbox": [ + 132, + 542, + 211, + 554 + ], + "type": "text", + "content": "3 Overview" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 130, + 567, + 481, + 591 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 567, + 481, + 591 + ], + "spans": [ + { + "bbox": [ + 130, + 567, + 481, + 591 + ], + "type": "text", + "content": "We start by describing some backgrounds in defining personalized federated learning and then present a motivating example." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 130, + 609, + 392, + 622 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 609, + 392, + 622 + ], + "spans": [ + { + "bbox": [ + 130, + 609, + 392, + 622 + ], + "type": "text", + "content": "3.1 Background: Personalized Federated Learning" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 130, + 629, + 481, + 667 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 629, + 481, + 667 + ], + "spans": [ + { + "bbox": [ + 130, + 629, + 481, + 667 + ], + "type": "text", + "content": "We describe some backgrounds of personalized federated learning. Consider the initial weights for all local models as " + }, + { + "bbox": [ + 130, + 629, + 481, + 667 + ], + "type": "inline_equation", + "content": "\\{\\theta_1, \\theta_2, \\theta_3, \\dots, \\theta_n\\}" + }, + { + "bbox": [ + 130, + 629, + 481, + 667 + ], + "type": "text", + "content": ". Let " + }, + { + "bbox": [ + 130, + 629, + 481, + 667 + ], + "type": "inline_equation", + "content": "\\mathcal{P}_i" + }, + { + "bbox": [ + 130, + 629, + 481, + 667 + ], + "type": "text", + "content": " be the data distribution over domain " + }, + { + "bbox": [ + 130, + 629, + 481, + 667 + ], + "type": "inline_equation", + "content": "\\mathcal{X} \\times \\mathcal{Y}" + }, + { + "bbox": [ + 130, + 629, + 481, + 667 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 130, + 629, + 481, + 667 + ], + "type": "inline_equation", + "content": "l_i" + }, + { + "bbox": [ + 130, + 629, + 481, + 667 + ], + "type": "text", + "content": " be the loss function for client " + }, + { + "bbox": [ + 130, + 629, + 481, + 667 + ], + "type": "inline_equation", + "content": "i \\in \\{1, 2, 3.., n\\}" + }, + { + "bbox": [ + 130, + 629, + 481, + 667 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 9 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "spans": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "type": "text", + "content": "4" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 166, + 91, + 225, + 101 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 166, + 91, + 225, + 101 + ], + "spans": [ + { + "bbox": [ + 166, + 91, + 225, + 101 + ], + "type": "text", + "content": "H. Yuan et al." + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 3 + }, + { + "para_blocks": [ + { + "bbox": [ + 130, + 116, + 479, + 153 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 116, + 479, + 153 + ], + "spans": [ + { + "bbox": [ + 130, + 116, + 479, + 153 + ], + "type": "text", + "content": "The goal of personalized federated learning is to find the set of personalized weights " + }, + { + "bbox": [ + 130, + 116, + 479, + 153 + ], + "type": "inline_equation", + "content": "\\Theta" + }, + { + "bbox": [ + 130, + 116, + 479, + 153 + ], + "type": "text", + "content": " for each client such that it minimizes empirical loss with respect to each " + }, + { + "bbox": [ + 130, + 116, + 479, + 153 + ], + "type": "inline_equation", + "content": "\\mathcal{P}_i" + }, + { + "bbox": [ + 130, + 116, + 479, + 153 + ], + "type": "text", + "content": " as shown in Equation ①" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 220, + 162, + 481, + 193 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 220, + 162, + 481, + 193 + ], + "spans": [ + { + "bbox": [ + 220, + 162, + 481, + 193 + ], + "type": "interline_equation", + "content": "\\Theta = \\operatorname {a r g m i n} _ {\\theta_ {i}} \\frac {1}{n} \\sum_ {i = 1} ^ {n} \\mathbb {E} _ {x, y \\sim \\mathcal {P} _ {i}} \\left[ l _ {i} \\left(\\theta_ {i}, x, y\\right) \\right] \\tag {1}", + "image_path": "52e15d67fc37b9db082c21313ded1ef729141e42e3ed55477b208c6a3a975251.jpg" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 132, + 211, + 276, + 224 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 211, + 276, + 224 + ], + "spans": [ + { + "bbox": [ + 132, + 211, + 276, + 224 + ], + "type": "text", + "content": "3.2 A Motivating Example" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 130, + 233, + 332, + 353 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 233, + 332, + 353 + ], + "spans": [ + { + "bbox": [ + 130, + 233, + 332, + 353 + ], + "type": "text", + "content": "In this part, we run through an example to motivate PFEDEDIT. We start with a challenging question. Note that previous personalized FL work [24,18,36] has noted the advantages of keeping some layers locally under non-i.i.d. scenarios, e.g., all batch normalization layers for FedBN [18]. This observation raises an intriguing question: How do we choose personalized layers to improve FL's performance with heterogeneous data?" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 130, + 354, + 332, + 509 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 354, + 332, + 509 + ], + "spans": [ + { + "bbox": [ + 130, + 354, + 332, + 509 + ], + "type": "text", + "content": "Our motivating example provides insight into this question. Specifically, we consider a real-world federated learning scenario where there are 10 clients collaboratively training their models on the CIFAR-100 dataset [12]. Each client is assigned 20 classes of image samples following Yuan et al. [36] to form a non-i.i.d. scenario, and we choose ResNet18 [9] as their classifier. To inspect the correlation between the personalized layers and the model utility, we pick " + }, + { + "bbox": [ + 130, + 354, + 332, + 509 + ], + "type": "inline_equation", + "content": "\\{1,3,5,7,10,13,17\\}" + }, + { + "bbox": [ + 130, + 354, + 332, + 509 + ], + "type": "text", + "content": " as the number of personalized layers, and choose three personalized algorithms for comparison." + } + ] + } + ], + "index": 6 + }, + { + "type": "image", + "bbox": [ + 343, + 261, + 478, + 374 + ], + "blocks": [ + { + "bbox": [ + 343, + 261, + 478, + 374 + ], + "lines": [ + { + "bbox": [ + 343, + 261, + 478, + 374 + ], + "spans": [ + { + "bbox": [ + 343, + 261, + 478, + 374 + ], + "type": "image", + "image_path": "e249f10e2e43a472744fb26833ae42c07a89cbd4003c06b20d19be9b37536e72.jpg" + } + ] + } + ], + "index": 7, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 339, + 384, + 482, + 483 + ], + "lines": [ + { + "bbox": [ + 339, + 384, + 482, + 483 + ], + "spans": [ + { + "bbox": [ + 339, + 384, + 482, + 483 + ], + "type": "text", + "content": "Fig. 2: A motivating example: We consider three different ways to personalize layers of local models in a federated learning scenario: random choosing, fixed choosing, and optimized layerchoosing based on the loss value. For the fixed choosing algorithm, we allow clients to keep certain numbers of batch normalization layers locally each round, and this approach becomes equivalent to FedBN [18] when the number of personalized layers gets 17." + } + ] + } + ], + "index": 8, + "angle": 0, + "type": "image_caption" + } + ], + "index": 7 + }, + { + "bbox": [ + 130, + 510, + 482, + 582 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 510, + 482, + 582 + ], + "spans": [ + { + "bbox": [ + 130, + 510, + 482, + 582 + ], + "type": "text", + "content": "The first algorithm randomly chooses certain numbers of layers from local model and keeps those layers locally. The second algorithm follows FedBN [18] to personalize batch normalization layers, and we denoted it as fixed choosing. The third personalized algorithm optimizes layer-choosing based on loss values and it considers layers that, once kept locally, yield the least loss values personalized layer and keeps them locally each round." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 130, + 582, + 482, + 667 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 582, + 482, + 667 + ], + "spans": [ + { + "bbox": [ + 130, + 582, + 482, + 667 + ], + "type": "text", + "content": "Figure 2 shows the results: The optimized-personalized algorithm clearly outperforms both fixed and random personalized algorithms in all cases. This motivates our approach of using model editing to choose the optimal layers for personalization. Note that the random-personalized algorithm fails to outperform FedAvg [20] in all cases, implying that random-personalized layers do not grant utility improvement as it comes with significant randomness. The fixed-personalized algorithm fails to outperform FedAvg [20] when with a small number" + } + ] + } + ], + "index": 10 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 400, + 91, + 447, + 100 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 400, + 91, + 447, + 100 + ], + "spans": [ + { + "bbox": [ + 400, + 91, + 447, + 100 + ], + "type": "text", + "content": "PFEDEDIT" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 474, + 91, + 480, + 99 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 474, + 91, + 480, + 99 + ], + "spans": [ + { + "bbox": [ + 474, + 91, + 480, + 99 + ], + "type": "text", + "content": "5" + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 4 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 137, + 114, + 478, + 213 + ], + "blocks": [ + { + "bbox": [ + 137, + 114, + 478, + 213 + ], + "lines": [ + { + "bbox": [ + 137, + 114, + 478, + 213 + ], + "spans": [ + { + "bbox": [ + 137, + 114, + 478, + 213 + ], + "type": "image", + "image_path": "729b0033d059f86faab17a92dfd9c92d05c756c54c0e5cc957457802375c19f3.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 130, + 222, + 482, + 300 + ], + "lines": [ + { + "bbox": [ + 130, + 222, + 482, + 300 + ], + "spans": [ + { + "bbox": [ + 130, + 222, + 482, + 300 + ], + "type": "text", + "content": "Fig. 3: Framework of PFEDEDIT. The federated learning still works as usual. That is, first, clients train their local models and upload their model weights to the global server. Then, the global server aggregates all model weights and returns the new global model weight to clients. Then, the main change is how PFEDEDIT merges global model with the local, which has two steps: i) PFEDEDIT generates a list of candidate model editorings, and ii) PFEDEDIT locates the top-k important model layers and outputs the edited model for the next-round local training." + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_caption" + } + ], + "index": 2 + }, + { + "bbox": [ + 130, + 323, + 482, + 395 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 323, + 482, + 395 + ], + "spans": [ + { + "bbox": [ + 130, + 323, + 482, + 395 + ], + "type": "text", + "content": "of personalized layers, and its utility is quite similar to the random-choosing algorithm. Once the number of personalized layer gets 17, which means all batch normalization layers are locally personalized, the utility improvement increases significantly and it outperforms FedAvg [18] by " + }, + { + "bbox": [ + 130, + 323, + 482, + 395 + ], + "type": "inline_equation", + "content": "13\\%" + }, + { + "bbox": [ + 130, + 323, + 482, + 395 + ], + "type": "text", + "content": ", which implies that fixed-personalized algorithm may require all certain layers to be involved to show any utility improvement." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 132, + 412, + 202, + 425 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 412, + 202, + 425 + ], + "spans": [ + { + "bbox": [ + 132, + 412, + 202, + 425 + ], + "type": "text", + "content": "4 Method" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 130, + 437, + 483, + 533 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 437, + 483, + 533 + ], + "spans": [ + { + "bbox": [ + 130, + 437, + 483, + 533 + ], + "type": "text", + "content": "In this section, we introduce our method, PFEDEDIT, which leverages model editing to improve model accuracy for heterogeneous data distributions. Figure 3 shows the entire workflow of PFEDEDIT. The training still follows a traditional FL, i.e., each client uploads their local model to the server for aggregation, and then the server returns the aggregated model. Then, when the client receives the global model, there are mainly two steps in PFEDEDIT, model editing and locating personalized layers, to update the client model based on the global model." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 131, + 550, + 391, + 563 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 131, + 550, + 391, + 563 + ], + "spans": [ + { + "bbox": [ + 131, + 550, + 391, + 563 + ], + "type": "text", + "content": "4.1 Step I: Generating Candidate Model Editings" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 130, + 570, + 482, + 666 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 570, + 482, + 666 + ], + "spans": [ + { + "bbox": [ + 130, + 570, + 482, + 666 + ], + "type": "text", + "content": "The purpose of this step is to generate a list of possible editorings of the local model based on the global model returned from the FL server. Specifically, PFEDEDIT considers the local model \"clean model\" as it is trained on local distribution. On the opposite, it considers the global model \"corrupted model\" as in non-i.i.d. scenarios, different models bias global model weights towards their local distributions, causing it not generalizing well on each local distribution. In the model editing step, as Figure 3 shows, PFEDEDIT hypothesizes all editorings that tailor the \"corrupted model\" towards local distribution." + } + ] + } + ], + "index": 8 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "spans": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "type": "text", + "content": "6" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 166, + 91, + 225, + 101 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 166, + 91, + 225, + 101 + ], + "spans": [ + { + "bbox": [ + 166, + 91, + 225, + 101 + ], + "type": "text", + "content": "H. Yuan et al." + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 5 + }, + { + "para_blocks": [ + { + "bbox": [ + 130, + 116, + 482, + 199 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 116, + 482, + 199 + ], + "spans": [ + { + "bbox": [ + 130, + 116, + 482, + 199 + ], + "type": "text", + "content": "We show the pseudocode of this step in Algorithm " + }, + { + "bbox": [ + 130, + 116, + 482, + 199 + ], + "type": "inline_equation", + "content": "\\mathbb{I}" + }, + { + "bbox": [ + 130, + 116, + 482, + 199 + ], + "type": "text", + "content": ". Line 5-6. For each editing, PFEDEDIT replaces one layer " + }, + { + "bbox": [ + 130, + 116, + 482, + 199 + ], + "type": "inline_equation", + "content": "\\{h_t^g | t \\in (1, T)\\}" + }, + { + "bbox": [ + 130, + 116, + 482, + 199 + ], + "type": "text", + "content": ", where " + }, + { + "bbox": [ + 130, + 116, + 482, + 199 + ], + "type": "inline_equation", + "content": "T" + }, + { + "bbox": [ + 130, + 116, + 482, + 199 + ], + "type": "text", + "content": " is the total number of model layers, from the global model " + }, + { + "bbox": [ + 130, + 116, + 482, + 199 + ], + "type": "inline_equation", + "content": "f_{\\theta^g}" + }, + { + "bbox": [ + 130, + 116, + 482, + 199 + ], + "type": "text", + "content": " with the corresponding layer " + }, + { + "bbox": [ + 130, + 116, + 482, + 199 + ], + "type": "inline_equation", + "content": "h_t^l" + }, + { + "bbox": [ + 130, + 116, + 482, + 199 + ], + "type": "text", + "content": " from the local model " + }, + { + "bbox": [ + 130, + 116, + 482, + 199 + ], + "type": "inline_equation", + "content": "f_{\\theta^l}" + }, + { + "bbox": [ + 130, + 116, + 482, + 199 + ], + "type": "text", + "content": ". Then the edited model is defined as " + }, + { + "bbox": [ + 130, + 116, + 482, + 199 + ], + "type": "inline_equation", + "content": "S(\\theta^*) = (S(\\theta^g) \\setminus \\{h_j^g\\}_{j=0}^k) \\cup \\{h_j^l\\}_{j=0}^k" + }, + { + "bbox": [ + 130, + 116, + 482, + 199 + ], + "type": "text", + "content": ". As the global model is equipped with the local model layer, which is optimized on local distribution, the global model is tailored towards local distribution." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 130, + 200, + 482, + 260 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 200, + 482, + 260 + ], + "spans": [ + { + "bbox": [ + 130, + 200, + 482, + 260 + ], + "type": "text", + "content": "The first editing happens after the first round (Line 1-2), right after each client receives the aggregated global model weights (Line 4) and before they start next-round local training (Line 14). The reason is that the global model starts to bias towards other distributions starting from the second round as its weights is uniformly initialized in the first round (Line 1)." + } + ] + } + ], + "index": 3 + }, + { + "type": "code", + "bbox": [ + 132, + 296, + 482, + 509 + ], + "blocks": [ + { + "bbox": [ + 133, + 282, + 248, + 293 + ], + "lines": [ + { + "bbox": [ + 133, + 282, + 248, + 293 + ], + "spans": [ + { + "bbox": [ + 133, + 282, + 248, + 293 + ], + "type": "text", + "content": "Algorithm 1 PFEDEDIT" + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "code_caption" + }, + { + "bbox": [ + 132, + 296, + 482, + 509 + ], + "lines": [ + { + "bbox": [ + 132, + 296, + 482, + 509 + ], + "spans": [ + { + "bbox": [ + 132, + 296, + 482, + 509 + ], + "type": "text", + "content": "Input: client i's initialized global model " + }, + { + "bbox": [ + 132, + 296, + 482, + 509 + ], + "type": "inline_equation", + "content": "\\theta^g" + }, + { + "bbox": [ + 132, + 296, + 482, + 509 + ], + "type": "text", + "content": ", communication round " + }, + { + "bbox": [ + 132, + 296, + 482, + 509 + ], + "type": "inline_equation", + "content": "R" + }, + { + "bbox": [ + 132, + 296, + 482, + 509 + ], + "type": "text", + "content": ", number of model layers " + }, + { + "bbox": [ + 132, + 296, + 482, + 509 + ], + "type": "inline_equation", + "content": "T" + }, + { + "bbox": [ + 132, + 296, + 482, + 509 + ], + "type": "text", + "content": ", number of layers to be edited " + }, + { + "bbox": [ + 132, + 296, + 482, + 509 + ], + "type": "inline_equation", + "content": "k" + }, + { + "bbox": [ + 132, + 296, + 482, + 509 + ], + "type": "text", + "content": ", local training set " + }, + { + "bbox": [ + 132, + 296, + 482, + 509 + ], + "type": "inline_equation", + "content": "D_{train}" + }, + { + "bbox": [ + 132, + 296, + 482, + 509 + ], + "type": "text", + "content": ", subset " + }, + { + "bbox": [ + 132, + 296, + 482, + 509 + ], + "type": "inline_equation", + "content": "D_p" + }, + { + "bbox": [ + 132, + 296, + 482, + 509 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 132, + 296, + 482, + 509 + ], + "type": "inline_equation", + "content": "z = (x,y) \\in D_p" + }, + { + "bbox": [ + 132, + 296, + 482, + 509 + ], + "type": "text", + "content": ", model layer set " + }, + { + "bbox": [ + 132, + 296, + 482, + 509 + ], + "type": "inline_equation", + "content": "S(\\cdot)" + }, + { + "bbox": [ + 132, + 296, + 482, + 509 + ], + "type": "text", + "content": ", dictionary " + }, + { + "bbox": [ + 132, + 296, + 482, + 509 + ], + "type": "inline_equation", + "content": "W" + }, + { + "bbox": [ + 132, + 296, + 482, + 509 + ], + "type": "text", + "content": " \nOutput: Edited global model weights " + }, + { + "bbox": [ + 132, + 296, + 482, + 509 + ], + "type": "inline_equation", + "content": "\\theta^*" + }, + { + "bbox": [ + 132, + 296, + 482, + 509 + ], + "type": "text", + "content": " for " + }, + { + "bbox": [ + 132, + 296, + 482, + 509 + ], + "type": "inline_equation", + "content": "R" + }, + { + "bbox": [ + 132, + 296, + 482, + 509 + ], + "type": "text", + "content": "-th round \n1: " + }, + { + "bbox": [ + 132, + 296, + 482, + 509 + ], + "type": "inline_equation", + "content": "\\theta^l \\gets" + }, + { + "bbox": [ + 132, + 296, + 482, + 509 + ], + "type": "text", + "content": " train " + }, + { + "bbox": [ + 132, + 296, + 482, + 509 + ], + "type": "inline_equation", + "content": "\\theta^g" + }, + { + "bbox": [ + 132, + 296, + 482, + 509 + ], + "type": "text", + "content": " on " + }, + { + "bbox": [ + 132, + 296, + 482, + 509 + ], + "type": "inline_equation", + "content": "D_{train}" + }, + { + "bbox": [ + 132, + 296, + 482, + 509 + ], + "type": "text", + "content": " \n2: " + }, + { + "bbox": [ + 132, + 296, + 482, + 509 + ], + "type": "inline_equation", + "content": "\\theta^g \\gets \\frac{1}{n} \\sum^n \\theta_i^l" + }, + { + "bbox": [ + 132, + 296, + 482, + 509 + ], + "type": "text", + "content": " \n3: for " + }, + { + "bbox": [ + 132, + 296, + 482, + 509 + ], + "type": "inline_equation", + "content": "r" + }, + { + "bbox": [ + 132, + 296, + 482, + 509 + ], + "type": "text", + "content": " in range " + }, + { + "bbox": [ + 132, + 296, + 482, + 509 + ], + "type": "inline_equation", + "content": "(1,R+1)" + }, + { + "bbox": [ + 132, + 296, + 482, + 509 + ], + "type": "text", + "content": " do \n4: for " + }, + { + "bbox": [ + 132, + 296, + 482, + 509 + ], + "type": "inline_equation", + "content": "h_t^g" + }, + { + "bbox": [ + 132, + 296, + 482, + 509 + ], + "type": "text", + "content": " in " + }, + { + "bbox": [ + 132, + 296, + 482, + 509 + ], + "type": "inline_equation", + "content": "S(\\theta^g)" + }, + { + "bbox": [ + 132, + 296, + 482, + 509 + ], + "type": "text", + "content": " do \n5: " + }, + { + "bbox": [ + 132, + 296, + 482, + 509 + ], + "type": "inline_equation", + "content": "S(\\theta^*) \\gets (S(\\theta^g) \\backslash h_t^g) \\cup h_t^l" + }, + { + "bbox": [ + 132, + 296, + 482, + 509 + ], + "type": "text", + "content": " \n6: " + }, + { + "bbox": [ + 132, + 296, + 482, + 509 + ], + "type": "inline_equation", + "content": "g(\\theta_i^l, \\theta_i^*, z)" + }, + { + "bbox": [ + 132, + 296, + 482, + 509 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 132, + 296, + 482, + 509 + ], + "type": "inline_equation", + "content": "(f_{\\theta^*}(x), y) \\gets f_{\\theta^*}(z)" + }, + { + "bbox": [ + 132, + 296, + 482, + 509 + ], + "type": "text", + "content": " \n7: " + }, + { + "bbox": [ + 132, + 296, + 482, + 509 + ], + "type": "inline_equation", + "content": "\\varphi \\gets (g(\\theta_i^l, \\theta_i^*, z), (f_{\\theta^*}(x), y))" + }, + { + "bbox": [ + 132, + 296, + 482, + 509 + ], + "type": "text", + "content": " \n8: " + }, + { + "bbox": [ + 132, + 296, + 482, + 509 + ], + "type": "inline_equation", + "content": "W[h_t^g] \\gets \\varphi" + }, + { + "bbox": [ + 132, + 296, + 482, + 509 + ], + "type": "text", + "content": " \n9: end for \n10: " + }, + { + "bbox": [ + 132, + 296, + 482, + 509 + ], + "type": "inline_equation", + "content": "W \\gets" + }, + { + "bbox": [ + 132, + 296, + 482, + 509 + ], + "type": "text", + "content": " Lexicographic_rank(W) \n11: for " + }, + { + "bbox": [ + 132, + 296, + 482, + 509 + ], + "type": "inline_equation", + "content": "j \\in" + }, + { + "bbox": [ + 132, + 296, + 482, + 509 + ], + "type": "text", + "content": " range " + }, + { + "bbox": [ + 132, + 296, + 482, + 509 + ], + "type": "inline_equation", + "content": "(0,k)" + }, + { + "bbox": [ + 132, + 296, + 482, + 509 + ], + "type": "text", + "content": " do \n12: " + }, + { + "bbox": [ + 132, + 296, + 482, + 509 + ], + "type": "inline_equation", + "content": "m \\gets W_j.\\text{key}()" + }, + { + "bbox": [ + 132, + 296, + 482, + 509 + ], + "type": "text", + "content": " \n13: " + }, + { + "bbox": [ + 132, + 296, + 482, + 509 + ], + "type": "inline_equation", + "content": "S(\\theta^*) \\gets (S(\\theta^g) \\backslash h_m^g) \\cup h_m^l" + }, + { + "bbox": [ + 132, + 296, + 482, + 509 + ], + "type": "text", + "content": " \n14: end for \n15: if " + }, + { + "bbox": [ + 132, + 296, + 482, + 509 + ], + "type": "inline_equation", + "content": "r == R" + }, + { + "bbox": [ + 132, + 296, + 482, + 509 + ], + "type": "text", + "content": " then \n16: return " + }, + { + "bbox": [ + 132, + 296, + 482, + 509 + ], + "type": "inline_equation", + "content": "\\theta^*" + }, + { + "bbox": [ + 132, + 296, + 482, + 509 + ], + "type": "text", + "content": " \n17: else \n18: " + }, + { + "bbox": [ + 132, + 296, + 482, + 509 + ], + "type": "inline_equation", + "content": "\\theta^l \\gets" + }, + { + "bbox": [ + 132, + 296, + 482, + 509 + ], + "type": "text", + "content": " train " + }, + { + "bbox": [ + 132, + 296, + 482, + 509 + ], + "type": "inline_equation", + "content": "\\theta^*" + }, + { + "bbox": [ + 132, + 296, + 482, + 509 + ], + "type": "text", + "content": " on " + }, + { + "bbox": [ + 132, + 296, + 482, + 509 + ], + "type": "inline_equation", + "content": "D_{train}" + }, + { + "bbox": [ + 132, + 296, + 482, + 509 + ], + "type": "text", + "content": " \n19: " + }, + { + "bbox": [ + 132, + 296, + 482, + 509 + ], + "type": "inline_equation", + "content": "\\theta^g \\gets \\frac{1}{n} \\sum^n \\theta_l^l" + }, + { + "bbox": [ + 132, + 296, + 482, + 509 + ], + "type": "text", + "content": " \n20: end if \n21: end for" + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "code_body" + } + ], + "index": 5, + "sub_type": "algorithm" + }, + { + "bbox": [ + 130, + 549, + 457, + 562 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 549, + 457, + 562 + ], + "spans": [ + { + "bbox": [ + 130, + 549, + 457, + 562 + ], + "type": "text", + "content": "4.2 Step II: Locating Personalized Layers by Ranking Editings" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 130, + 570, + 482, + 666 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 570, + 482, + 666 + ], + "spans": [ + { + "bbox": [ + 130, + 570, + 482, + 666 + ], + "type": "text", + "content": "The purpose of this step is to access all possible editorings from Step I and then locate a list of personalized layers. The detailed substeps are two-fold. First, PFEDEDIT evaluates model editing's effectiveness based on a Total Effectiveness (TE) Score (Step II-a). Second, PFEDEDIT generates prediction lists based on TE score and ground-truth mask to reflect changes in probability and predicted labels (Step II-b). Lastly, PFEDEDIT ranks all prediction lists following the lexicographic order and then selects the top-k best layer editorings as personalized layers for each client based on the ranking (Step II-c)." + } + ] + } + ], + "index": 7 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 400, + 91, + 447, + 100 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 400, + 91, + 447, + 100 + ], + "spans": [ + { + "bbox": [ + 400, + 91, + 447, + 100 + ], + "type": "text", + "content": "PFEDEDIT" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 474, + 91, + 480, + 99 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 474, + 91, + 480, + 99 + ], + "spans": [ + { + "bbox": [ + 474, + 91, + 480, + 99 + ], + "type": "text", + "content": "7" + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 6 + }, + { + "para_blocks": [ + { + "bbox": [ + 130, + 116, + 482, + 308 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 116, + 482, + 308 + ], + "spans": [ + { + "bbox": [ + 130, + 116, + 482, + 308 + ], + "type": "text", + "content": "Step II-a: Evaluating model editing effectiveness. This substep is to calculate a so-called Total Effectiveness (TE) score to evaluate mode editing effectiveness. Specifically, PFEDEDIT queries local model " + }, + { + "bbox": [ + 130, + 116, + 482, + 308 + ], + "type": "inline_equation", + "content": "f(\\theta^l)" + }, + { + "bbox": [ + 130, + 116, + 482, + 308 + ], + "type": "text", + "content": " using a representative subset of the training set, denoted as " + }, + { + "bbox": [ + 130, + 116, + 482, + 308 + ], + "type": "inline_equation", + "content": "D_p" + }, + { + "bbox": [ + 130, + 116, + 482, + 308 + ], + "type": "text", + "content": ", which maintains the same distribution as the full training set " + }, + { + "bbox": [ + 130, + 116, + 482, + 308 + ], + "type": "inline_equation", + "content": "D_{train}" + }, + { + "bbox": [ + 130, + 116, + 482, + 308 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 130, + 116, + 482, + 308 + ], + "type": "inline_equation", + "content": "|D_p| = p * |D_{train}|" + }, + { + "bbox": [ + 130, + 116, + 482, + 308 + ], + "type": "text", + "content": ", where " + }, + { + "bbox": [ + 130, + 116, + 482, + 308 + ], + "type": "inline_equation", + "content": "p \\in (0,1)" + }, + { + "bbox": [ + 130, + 116, + 482, + 308 + ], + "type": "text", + "content": ". PFEDEDIT considers that model weights " + }, + { + "bbox": [ + 130, + 116, + 482, + 308 + ], + "type": "inline_equation", + "content": "\\theta^l" + }, + { + "bbox": [ + 130, + 116, + 482, + 308 + ], + "type": "text", + "content": " represents local distribution as it is trained on pure local data. PFEDEDIT then queries each edited model " + }, + { + "bbox": [ + 130, + 116, + 482, + 308 + ], + "type": "inline_equation", + "content": "f(\\theta^*)" + }, + { + "bbox": [ + 130, + 116, + 482, + 308 + ], + "type": "text", + "content": " using the same subset to see the prediction of tailored weights on local distribution. Then by comparing probability outputs from both local and edited weights on the same image samples, PFEDEDIT measures the effectiveness of that editing as well as the capability of edited layer in improving model accuracy. Specifically, when querying the edited model " + }, + { + "bbox": [ + 130, + 116, + 482, + 308 + ], + "type": "inline_equation", + "content": "f(\\theta^*)" + }, + { + "bbox": [ + 130, + 116, + 482, + 308 + ], + "type": "text", + "content": ", PFEDEDIT collects both TE scores and ground-truth masks, Line 7 in Algorithm 1. The Total Effectiveness (TE) score, as defined below, reflects the probability change on the ground-truth class while the ground truth mask, which is a boolean list indicating whether the predicted label matches the ground truth label, reflects the predicted label change." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 130, + 308, + 482, + 356 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 308, + 482, + 356 + ], + "spans": [ + { + "bbox": [ + 130, + 308, + 482, + 356 + ], + "type": "text", + "content": "Now we describe the definition of TE score. Given model weight " + }, + { + "bbox": [ + 130, + 308, + 482, + 356 + ], + "type": "inline_equation", + "content": "\\theta" + }, + { + "bbox": [ + 130, + 308, + 482, + 356 + ], + "type": "text", + "content": " and training dataset " + }, + { + "bbox": [ + 130, + 308, + 482, + 356 + ], + "type": "inline_equation", + "content": "D" + }, + { + "bbox": [ + 130, + 308, + 482, + 356 + ], + "type": "text", + "content": " with " + }, + { + "bbox": [ + 130, + 308, + 482, + 356 + ], + "type": "inline_equation", + "content": "c" + }, + { + "bbox": [ + 130, + 308, + 482, + 356 + ], + "type": "text", + "content": " classes, we denote by " + }, + { + "bbox": [ + 130, + 308, + 482, + 356 + ], + "type": "inline_equation", + "content": "P(y|\\theta ,x)" + }, + { + "bbox": [ + 130, + 308, + 482, + 356 + ], + "type": "text", + "content": " the probability (softmax) of the ground-truth class " + }, + { + "bbox": [ + 130, + 308, + 482, + 356 + ], + "type": "inline_equation", + "content": "y" + }, + { + "bbox": [ + 130, + 308, + 482, + 356 + ], + "type": "text", + "content": ". We define the total effectiveness (TE) score of editing the model weight " + }, + { + "bbox": [ + 130, + 308, + 482, + 356 + ], + "type": "inline_equation", + "content": "\\theta" + }, + { + "bbox": [ + 130, + 308, + 482, + 356 + ], + "type": "text", + "content": " into the new weight " + }, + { + "bbox": [ + 130, + 308, + 482, + 356 + ], + "type": "inline_equation", + "content": "\\theta^{*}" + }, + { + "bbox": [ + 130, + 308, + 482, + 356 + ], + "type": "text", + "content": " given dataset " + }, + { + "bbox": [ + 130, + 308, + 482, + 356 + ], + "type": "inline_equation", + "content": "D" + }, + { + "bbox": [ + 130, + 308, + 482, + 356 + ], + "type": "text", + "content": " is defined as:" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 228, + 361, + 481, + 389 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 228, + 361, + 481, + 389 + ], + "spans": [ + { + "bbox": [ + 228, + 361, + 481, + 389 + ], + "type": "interline_equation", + "content": "g (\\theta , \\theta^ {*}, D) = \\frac {1}{N} \\sum_ {x, y \\in D} \\frac {P (y | \\theta^ {*} , x)}{P (y | \\theta , x)} - 1 \\tag {2}", + "image_path": "d42af7945f8d7e1b9e36bd41c3e5263a76bd427513923af30f80cfe202a358ec.jpg" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 131, + 396, + 481, + 419 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 131, + 396, + 481, + 419 + ], + "spans": [ + { + "bbox": [ + 131, + 396, + 481, + 419 + ], + "type": "text", + "content": "A positive TE score represents an increased probability on ground-truth label while a negative TE score signifies a decrease." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 130, + 428, + 482, + 525 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 428, + 482, + 525 + ], + "spans": [ + { + "bbox": [ + 130, + 428, + 482, + 525 + ], + "type": "text", + "content": "Step II-b: Generating prediction lists based on TE and ground-truth mask. This sub-task is to generate prediction lists based on TE and ground-truth mask. Specifically, the TE score, together with the ground-truth mask, makes up four combinations in total, and we call the list of all four cases a prediction list, denoted by " + }, + { + "bbox": [ + 130, + 428, + 482, + 525 + ], + "type": "inline_equation", + "content": "\\varphi" + }, + { + "bbox": [ + 130, + 428, + 482, + 525 + ], + "type": "text", + "content": ". The prediction list then serves as a reference for evaluating the editing as well as the capability of such layer in improving model's accuracy. Below we show the prediction list that contains all four combinations of TE score and ground-truth mask in Equation 3" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 233, + 533, + 481, + 587 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 233, + 533, + 481, + 587 + ], + "spans": [ + { + "bbox": [ + 233, + 533, + 481, + 587 + ], + "type": "interline_equation", + "content": "\\varphi = \\left\\{ \\begin{array}{l} f _ {\\theta^ {*}} (x) = y, g \\left(\\theta , \\theta^ {*}, z\\right) > 0 \\\\ f _ {\\theta^ {*}} (x) = y, g \\left(\\theta , \\theta^ {*}, z\\right) < 0 \\\\ f _ {\\theta^ {*}} (x) \\neq y, g \\left(\\theta , \\theta^ {*}, z\\right) > 0 \\\\ f _ {\\theta^ {*}} (x) \\neq y, g \\left(\\theta , \\theta^ {*}, z\\right) < 0 \\end{array} \\right. \\tag {3}", + "image_path": "60ba3d82c3ed9dc8f0608b5baabe64fcc24faec9597a9d40aff8810108f4e57a.jpg" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 130, + 594, + 482, + 666 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 594, + 482, + 666 + ], + "spans": [ + { + "bbox": [ + 130, + 594, + 482, + 666 + ], + "type": "text", + "content": "The top case of the prediction list " + }, + { + "bbox": [ + 130, + 594, + 482, + 666 + ], + "type": "inline_equation", + "content": "\\varphi" + }, + { + "bbox": [ + 130, + 594, + 482, + 666 + ], + "type": "text", + "content": " shows the best editing (predicted label same as ground-truth label, TE score greater than 0) and bottom case shows the worst editing effect (predicted label different to ground-truth label, TE score less than 0). PFEDEDIT collects values for such prediction list for each editing by splitting which case each image prediction belongs to, and then ranks all prediction lists using Lexicographic order defined below." + } + ] + } + ], + "index": 8 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "spans": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "type": "text", + "content": "8" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 166, + 91, + 225, + 101 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 166, + 91, + 225, + 101 + ], + "spans": [ + { + "bbox": [ + 166, + 91, + 225, + 101 + ], + "type": "text", + "content": "H. Yuan et al." + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 7 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 137, + 122, + 233, + 198 + ], + "blocks": [ + { + "bbox": [ + 137, + 122, + 233, + 198 + ], + "lines": [ + { + "bbox": [ + 137, + 122, + 233, + 198 + ], + "spans": [ + { + "bbox": [ + 137, + 122, + 233, + 198 + ], + "type": "image", + "image_path": "986db9a589683faa627a6a519f630ff92043d1593067de60cd0491f5fc85530b.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 171, + 200, + 206, + 209 + ], + "lines": [ + { + "bbox": [ + 171, + 200, + 206, + 209 + ], + "spans": [ + { + "bbox": [ + 171, + 200, + 206, + 209 + ], + "type": "text", + "content": "(a) Lyme" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_caption" + } + ], + "index": 2 + }, + { + "type": "image", + "bbox": [ + 253, + 122, + 347, + 198 + ], + "blocks": [ + { + "bbox": [ + 253, + 122, + 347, + 198 + ], + "lines": [ + { + "bbox": [ + 253, + 122, + 347, + 198 + ], + "spans": [ + { + "bbox": [ + 253, + 122, + 347, + 198 + ], + "type": "image", + "image_path": "6711746bb6761d3305c801b171f34a67ad5ecf9284bbe92b164002d91cf35d71.jpg" + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 280, + 200, + 331, + 209 + ], + "lines": [ + { + "bbox": [ + 280, + 200, + 331, + 209 + ], + "spans": [ + { + "bbox": [ + 280, + 200, + 331, + 209 + ], + "type": "text", + "content": "(b) CIFAR-10" + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 130, + 211, + 480, + 236 + ], + "lines": [ + { + "bbox": [ + 130, + 211, + 480, + 236 + ], + "spans": [ + { + "bbox": [ + 130, + 211, + 480, + 236 + ], + "type": "text", + "content": "Fig. 4: Non-i.i.d. data distribution (2 classes per client and 20 classes for CIFAR-100) among clients in Section 5.2" + } + ] + } + ], + "index": 8, + "angle": 0, + "type": "image_caption" + } + ], + "index": 4 + }, + { + "type": "image", + "bbox": [ + 369, + 122, + 463, + 198 + ], + "blocks": [ + { + "bbox": [ + 369, + 122, + 463, + 198 + ], + "lines": [ + { + "bbox": [ + 369, + 122, + 463, + 198 + ], + "spans": [ + { + "bbox": [ + 369, + 122, + 463, + 198 + ], + "type": "image", + "image_path": "a24b23fbca0c61f39f108c57665b4cd1985b649e3dcd5248fcb9a2fe4689a3e3.jpg" + } + ] + } + ], + "index": 6, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 394, + 200, + 449, + 209 + ], + "lines": [ + { + "bbox": [ + 394, + 200, + 449, + 209 + ], + "spans": [ + { + "bbox": [ + 394, + 200, + 449, + 209 + ], + "type": "text", + "content": "(c) CIFAR-100" + } + ] + } + ], + "index": 7, + "angle": 0, + "type": "image_caption" + } + ], + "index": 6 + }, + { + "bbox": [ + 130, + 257, + 480, + 353 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 257, + 480, + 353 + ], + "spans": [ + { + "bbox": [ + 130, + 257, + 480, + 353 + ], + "type": "text", + "content": "Step II-c: Locating top-" + }, + { + "bbox": [ + 130, + 257, + 480, + 353 + ], + "type": "inline_equation", + "content": "k" + }, + { + "bbox": [ + 130, + 257, + 480, + 353 + ], + "type": "text", + "content": " personalized layers using lexicographic order. This sub-step is to rank all prediction lists " + }, + { + "bbox": [ + 130, + 257, + 480, + 353 + ], + "type": "inline_equation", + "content": "(\\varphi)" + }, + { + "bbox": [ + 130, + 257, + 480, + 353 + ], + "type": "text", + "content": " using lexicographic order and select the top-" + }, + { + "bbox": [ + 130, + 257, + 480, + 353 + ], + "type": "inline_equation", + "content": "k" + }, + { + "bbox": [ + 130, + 257, + 480, + 353 + ], + "type": "text", + "content": " as the personalized layers. Here is the definition of lexicographic order, which is used to rank " + }, + { + "bbox": [ + 130, + 257, + 480, + 353 + ], + "type": "inline_equation", + "content": "\\varphi" + }, + { + "bbox": [ + 130, + 257, + 480, + 353 + ], + "type": "text", + "content": ". That is, given two prediction lists " + }, + { + "bbox": [ + 130, + 257, + 480, + 353 + ], + "type": "inline_equation", + "content": "\\varphi_1, \\varphi_2" + }, + { + "bbox": [ + 130, + 257, + 480, + 353 + ], + "type": "text", + "content": " for two model editorings on layer " + }, + { + "bbox": [ + 130, + 257, + 480, + 353 + ], + "type": "inline_equation", + "content": "l_1, l_2" + }, + { + "bbox": [ + 130, + 257, + 480, + 353 + ], + "type": "text", + "content": ", respectively, say " + }, + { + "bbox": [ + 130, + 257, + 480, + 353 + ], + "type": "inline_equation", + "content": "\\varphi_1 = a_1, \\ldots, a_4" + }, + { + "bbox": [ + 130, + 257, + 480, + 353 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 130, + 257, + 480, + 353 + ], + "type": "inline_equation", + "content": "\\varphi_2 = b_1, \\ldots, b_4" + }, + { + "bbox": [ + 130, + 257, + 480, + 353 + ], + "type": "text", + "content": ", where " + }, + { + "bbox": [ + 130, + 257, + 480, + 353 + ], + "type": "inline_equation", + "content": "a" + }, + { + "bbox": [ + 130, + 257, + 480, + 353 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 130, + 257, + 480, + 353 + ], + "type": "inline_equation", + "content": "b" + }, + { + "bbox": [ + 130, + 257, + 480, + 353 + ], + "type": "text", + "content": " are possible cases from prediction lists, the order of those two lists depends on the numeric order of the cases in the first place " + }, + { + "bbox": [ + 130, + 257, + 480, + 353 + ], + "type": "inline_equation", + "content": "i" + }, + { + "bbox": [ + 130, + 257, + 480, + 353 + ], + "type": "text", + "content": " where the two case counts differ. That is, " + }, + { + "bbox": [ + 130, + 257, + 480, + 353 + ], + "type": "inline_equation", + "content": "\\varphi_1 < \\varphi_2" + }, + { + "bbox": [ + 130, + 257, + 480, + 353 + ], + "type": "text", + "content": " if and only if " + }, + { + "bbox": [ + 130, + 257, + 480, + 353 + ], + "type": "inline_equation", + "content": "a_i < b_i" + }, + { + "bbox": [ + 130, + 257, + 480, + 353 + ], + "type": "text", + "content": " in the natural order of numbers." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 130, + 354, + 481, + 463 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 354, + 481, + 463 + ], + "spans": [ + { + "bbox": [ + 130, + 354, + 481, + 463 + ], + "type": "text", + "content": "Note that the advantages of using such metric here are two-folds: i) We record " + }, + { + "bbox": [ + 130, + 354, + 481, + 463 + ], + "type": "inline_equation", + "content": "\\varphi_{j}(\\theta, \\theta^{*}, z)" + }, + { + "bbox": [ + 130, + 354, + 481, + 463 + ], + "type": "text", + "content": " by counting the number of each case and compare any two prediction lists following the Lexicographic order instead of a summation manner, i.e., " + }, + { + "bbox": [ + 130, + 354, + 481, + 463 + ], + "type": "inline_equation", + "content": "\\sum_{|D|}^{i} \\varphi_{j}(\\theta, \\theta^{*}, z)" + }, + { + "bbox": [ + 130, + 354, + 481, + 463 + ], + "type": "text", + "content": ", so that the impact of one case is not overridden by another, even if there are insufficient instances. ii) The case order listed in Equation 3 from top to the bottom, is intuitive and aligns with accuracy metric, thus implying that the editing with highest prediction list ranking comes with a largest accuracy improvement. We also discuss and compare our metrics with other potential ones, and evaluation details are shown in Section 5.4." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 132, + 481, + 217, + 493 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 481, + 217, + 493 + ], + "spans": [ + { + "bbox": [ + 132, + 481, + 217, + 493 + ], + "type": "text", + "content": "5 Evaluation" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 130, + 506, + 480, + 555 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 506, + 480, + 555 + ], + "spans": [ + { + "bbox": [ + 130, + 506, + 480, + 555 + ], + "type": "text", + "content": "In this section, we compare PFEDEDIT with baselines in terms of the accuracy and convergence. We then perform a study on parameters used in PFEDEDIT. The evaluation on computation overhead and complexity of PFEDEDIT can be found in Appendix A.4" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 131, + 573, + 261, + 586 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 131, + 573, + 261, + 586 + ], + "spans": [ + { + "bbox": [ + 131, + 573, + 261, + 586 + ], + "type": "text", + "content": "5.1 Experimental Setup" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 130, + 594, + 480, + 616 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 594, + 480, + 616 + ], + "spans": [ + { + "bbox": [ + 130, + 594, + 480, + 616 + ], + "type": "text", + "content": "We describe four aspects, i.e., setup of federated learning, datasets, models, and then baselines, for our experimental setup." + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 130, + 617, + 481, + 665 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 617, + 481, + 665 + ], + "spans": [ + { + "bbox": [ + 130, + 617, + 481, + 665 + ], + "type": "text", + "content": "FL Setup. Our default settings follow state-of-the-art works [2,6] in evaluating federated learning. That is, there are 10 clients to participate in a FL scenario with 100 communication rounds. The global sampling rate is set to be 1.0 in each communication round and each client has one local training epoch. The default" + } + ] + } + ], + "index": 15 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 400, + 91, + 447, + 100 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 400, + 91, + 447, + 100 + ], + "spans": [ + { + "bbox": [ + 400, + 91, + 447, + 100 + ], + "type": "text", + "content": "PFEDEDIT" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 474, + 91, + 480, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 474, + 91, + 480, + 100 + ], + "spans": [ + { + "bbox": [ + 474, + 91, + 480, + 100 + ], + "type": "text", + "content": "9" + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 8 + }, + { + "para_blocks": [ + { + "bbox": [ + 130, + 115, + 482, + 260 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 115, + 482, + 260 + ], + "spans": [ + { + "bbox": [ + 130, + 115, + 482, + 260 + ], + "type": "text", + "content": "value " + }, + { + "bbox": [ + 130, + 115, + 482, + 260 + ], + "type": "inline_equation", + "content": "(k)" + }, + { + "bbox": [ + 130, + 115, + 482, + 260 + ], + "type": "text", + "content": " in selecting and editing top layers is " + }, + { + "bbox": [ + 130, + 115, + 482, + 260 + ], + "type": "inline_equation", + "content": "7\\%" + }, + { + "bbox": [ + 130, + 115, + 482, + 260 + ], + "type": "text", + "content": ", and the default representative subset ratio " + }, + { + "bbox": [ + 130, + 115, + 482, + 260 + ], + "type": "inline_equation", + "content": "(p)" + }, + { + "bbox": [ + 130, + 115, + 482, + 260 + ], + "type": "text", + "content": " is 0.1. Both values are selected based on our parameter study (Section 5.4). By default, personalized layers are chosen each round, and the comparison with fixed personalized layer setting can be found in Appendix A.2. Datasets. We evaluate PFEDEDIT on three benchmark datasets to show its performance on different domains. Specifically, we choose i). CIFAR-10 [12], a benchmark dataset that contains 60,000 " + }, + { + "bbox": [ + 130, + 115, + 482, + 260 + ], + "type": "inline_equation", + "content": "32\\times 32" + }, + { + "bbox": [ + 130, + 115, + 482, + 260 + ], + "type": "text", + "content": " color images in 10 classes; ii). CIFAR-100 [12], a benchmark dataset that contains 60,000 " + }, + { + "bbox": [ + 130, + 115, + 482, + 260 + ], + "type": "inline_equation", + "content": "32\\times 32" + }, + { + "bbox": [ + 130, + 115, + 482, + 260 + ], + "type": "text", + "content": " color images in 100 classes; and iii). Lyme [35] dataset, a skin disease dataset that contains 3,027 skin disease images in 4 classes that are confusers for Lyme. We resize CIFAR-10 image samples to " + }, + { + "bbox": [ + 130, + 115, + 482, + 260 + ], + "type": "inline_equation", + "content": "64\\times 64" + }, + { + "bbox": [ + 130, + 115, + 482, + 260 + ], + "type": "text", + "content": ", CIFAR-100 samples to " + }, + { + "bbox": [ + 130, + 115, + 482, + 260 + ], + "type": "inline_equation", + "content": "224\\times 224" + }, + { + "bbox": [ + 130, + 115, + 482, + 260 + ], + "type": "text", + "content": " following [36], and Lyme samples to " + }, + { + "bbox": [ + 130, + 115, + 482, + 260 + ], + "type": "inline_equation", + "content": "255\\times 255" + }, + { + "bbox": [ + 130, + 115, + 482, + 260 + ], + "type": "text", + "content": " following [35]." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 130, + 261, + 482, + 503 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 261, + 482, + 503 + ], + "spans": [ + { + "bbox": [ + 130, + 261, + 482, + 503 + ], + "type": "text", + "content": "To show that PFEDEDIT can perform on various non-i.i.d. scenario, we adopt the Distributional Heterogeneity level (DH-level) from Yuan et al. [36] to split data distribution into different levels. Specifically, we first assign certain numbers of classes to each client " + }, + { + "bbox": [ + 130, + 261, + 482, + 503 + ], + "type": "inline_equation", + "content": "i" + }, + { + "bbox": [ + 130, + 261, + 482, + 503 + ], + "type": "text", + "content": ", and then for each local class " + }, + { + "bbox": [ + 130, + 261, + 482, + 503 + ], + "type": "inline_equation", + "content": "c" + }, + { + "bbox": [ + 130, + 261, + 482, + 503 + ], + "type": "text", + "content": ", we sample " + }, + { + "bbox": [ + 130, + 261, + 482, + 503 + ], + "type": "inline_equation", + "content": "D_{i,c} \\in (0.4, 0.6)" + }, + { + "bbox": [ + 130, + 261, + 482, + 503 + ], + "type": "text", + "content": " from " + }, + { + "bbox": [ + 130, + 261, + 482, + 503 + ], + "type": "inline_equation", + "content": "D_c" + }, + { + "bbox": [ + 130, + 261, + 482, + 503 + ], + "type": "text", + "content": ". Each client is assigned " + }, + { + "bbox": [ + 130, + 261, + 482, + 503 + ], + "type": "inline_equation", + "content": "\\frac{D_{i,c}}{\\sum D_{n,c}}" + }, + { + "bbox": [ + 130, + 261, + 482, + 503 + ], + "type": "text", + "content": " of samples in class " + }, + { + "bbox": [ + 130, + 261, + 482, + 503 + ], + "type": "inline_equation", + "content": "c" + }, + { + "bbox": [ + 130, + 261, + 482, + 503 + ], + "type": "text", + "content": ". The DH-level ranges from " + }, + { + "bbox": [ + 130, + 261, + 482, + 503 + ], + "type": "inline_equation", + "content": "0\\%" + }, + { + "bbox": [ + 130, + 261, + 482, + 503 + ], + "type": "text", + "content": ", which represents an i.i.d. scenario where each client gets all classes of data, all the way to " + }, + { + "bbox": [ + 130, + 261, + 482, + 503 + ], + "type": "inline_equation", + "content": "100\\%" + }, + { + "bbox": [ + 130, + 261, + 482, + 503 + ], + "type": "text", + "content": ", which means that each client only gets one class of data. (10 classes for CIFAR-100). For Lyme [35] dataset, since it only possesses four classes, we only conduct evaluations on " + }, + { + "bbox": [ + 130, + 261, + 482, + 503 + ], + "type": "inline_equation", + "content": "50\\%" + }, + { + "bbox": [ + 130, + 261, + 482, + 503 + ], + "type": "text", + "content": " (2 classes per clients) and " + }, + { + "bbox": [ + 130, + 261, + 482, + 503 + ], + "type": "inline_equation", + "content": "100\\%" + }, + { + "bbox": [ + 130, + 261, + 482, + 503 + ], + "type": "text", + "content": " (4 classes per clients) DH-level distribution scenarios. Unless specified otherwise, the default data distribution among clients is 2 classes per client (20 classes per client for CIFAR-100) in our evaluation. We demonstrate the default heterogeneous data distribution for each dataset in Figure 4. We also consider feature space non-i.i.d. scenario, and the evaluation, in comparison with Cd2-pfed [27] and Partialfed-fix [29], can be found in Appendix A.3. Models. For model selection, we choose ResNet18 [9] for evaluation on CIFAR 10 and CIFAR 100, and ViT [8] for Lyme. We allow ResNet18 model to have pretrained weights and allow ViT to have b16 weights that is fine-tuned on Image1k. We set the learning rate for ViT to be 1e-3 and we follow FedPer [2] to set the learning rate as 1e-2 for ResNet18." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 130, + 503, + 482, + 602 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 503, + 482, + 602 + ], + "spans": [ + { + "bbox": [ + 130, + 503, + 482, + 602 + ], + "type": "text", + "content": "Baselines. We compare PFEDEDIT with FedAvg [20], MOON [14], and other five SOTA personalized FL works as baselines, namely, FedPer [2], FedRep [4], FedBN [18], APFL [6], and DisTrans [36]. We allow each baseline to have the same data distribution and model selection as PFEDEDIT. We consider all layer normalization layers personalized for FedBN [18] when employing ViT. For other parameters, we follow their default paper settings to assign values. The accuracy comparison with more baselines (Fedfa [38], FedALA [37], Fed-LAMB [10], FedRolex [1], and HeteroFL [7]) can be found in Appendix A.1." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 132, + 619, + 272, + 632 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 619, + 272, + 632 + ], + "spans": [ + { + "bbox": [ + 132, + 619, + 272, + 632 + ], + "type": "text", + "content": "5.2 Accuracy Comparison" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 130, + 640, + 482, + 666 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 640, + 482, + 666 + ], + "spans": [ + { + "bbox": [ + 130, + 640, + 482, + 666 + ], + "type": "text", + "content": "As Table 1 shows, PFEDEDIT outperforms other state-of-the-art works on all three datasets under different DH-level data distributions. Compared to the" + } + ] + } + ], + "index": 6 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "spans": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "type": "text", + "content": "10" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 166, + 91, + 225, + 100 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 166, + 91, + 225, + 100 + ], + "spans": [ + { + "bbox": [ + 166, + 91, + 225, + 100 + ], + "type": "text", + "content": "H. Yuan et al." + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 9 + }, + { + "para_blocks": [ + { + "bbox": [ + 130, + 116, + 482, + 210 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 116, + 482, + 210 + ], + "spans": [ + { + "bbox": [ + 130, + 116, + 482, + 210 + ], + "type": "text", + "content": "baseline FedAvg, the accuracy improvement ranges from 0.007 to 0.266, with an average 0.09 accuracy improvement, and the largest accuracy improvement gets to 0.156 and average improvement gets to 0.05 when compared to other personalized FL works. The accuracy gap between PFEDEDIT and other STOA personalized works in CIFAR-10 is not as huge as in other two datasets, and the reason is that we allow all works to employ models with pretrained weights, which provides a greater lower bound for model accuracy and reduces the accuracy gap between different methods." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 130, + 212, + 482, + 427 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 212, + 482, + 427 + ], + "spans": [ + { + "bbox": [ + 130, + 212, + 482, + 427 + ], + "type": "text", + "content": "We then evaluate PFEDEDIT, together with other SOTA personalized FL methods under different DH levels. According to Table II, PFEDEDIT consistently achieves superior accuracy compared to other SOTA works under all data distributions, and we have two observations here. First, some personalized FL methods, e.g., FedPer, FedBN, and FedRep, fail to outperform FedAvg under i.i.d. scenarios, which is attributed to their natures as personalized FL works: they are strategically designed to align the global model with local distributions under non-i.i.d. scenarios. If the data distribution is i.i.d., where there is a close alignment between the global model distribution and each local one, the personalized layers may block favorable updates from other clients. On the other hand, PFEDEDIT outperforms FedAvg under i.i.d. scenarios, and the reason is that PFEDEDIT allow clients to update the weights of personalized layers so that they can favor other clients when local distributions are similar. Second, the accuracy improvement brought by PFEDEDIT compared to FedAvg increases as the DH-level increases. The reason is that as the distribution becomes more non-i.i.d., the gap between the global and local model distributions widens. Personalized layers help bridge this gap, tailoring the global model towards the local distribution and improving its generalization on local data." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 130, + 429, + 482, + 499 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 429, + 482, + 499 + ], + "spans": [ + { + "bbox": [ + 130, + 429, + 482, + 499 + ], + "type": "text", + "content": "It is noteworthy to mention that PFEDEDIT gets 0.989 accuracy on CIFAR-10 when DH-level gets to " + }, + { + "bbox": [ + 130, + 429, + 482, + 499 + ], + "type": "inline_equation", + "content": "100\\%" + }, + { + "bbox": [ + 130, + 429, + 482, + 499 + ], + "type": "text", + "content": ". The reason is that, in an extreme non-i.i.d. scenario where clients only get one class of data samples, the local model is very likely to get 1.0 testing accuracy and the model editing effectively improves the model accuracy as the local model distribution is highly matched with the local data distribution." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 131, + 522, + 226, + 534 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 131, + 522, + 226, + 534 + ], + "spans": [ + { + "bbox": [ + 131, + 522, + 226, + 534 + ], + "type": "text", + "content": "5.3 Convergence" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 130, + 545, + 482, + 666 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 545, + 482, + 666 + ], + "spans": [ + { + "bbox": [ + 130, + 545, + 482, + 666 + ], + "type": "text", + "content": "In this section we evaluate the convergence rate of PFEDEDIT. Specifically, we compare PFEDEDIT with FedAvg and we demonstrate their best accuracy on each dataset and the numbers of rounds to achieve it. As Table 2 shows, PFEDEDIT converges much faster than FedAvg on all three datasets as it only requires 18 epochs to converge on CIFAR-100, 13 epochs on CIFAR-10, and 25 epochs on Lyme. On the opposite, FedAvg takes around three times numbers of rounds to get a suboptimal accuracy on both CIFAR-10 and CIFAR-100. For Lyme, it takes 87 epochs to get accuracy above 0.7, which only takes 4 epochs for PFEDEDIT. The reason is that FedAvg does not provide convergence guarantee as it aggregates all model weights each round without having any optimization," + } + ] + } + ], + "index": 6 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 400, + 91, + 447, + 100 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 400, + 91, + 447, + 100 + ], + "spans": [ + { + "bbox": [ + 400, + 91, + 447, + 100 + ], + "type": "text", + "content": "PFEDEDIT" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 470, + 92, + 479, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 470, + 92, + 479, + 100 + ], + "spans": [ + { + "bbox": [ + 470, + 92, + 479, + 100 + ], + "type": "text", + "content": "11" + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 10 + }, + { + "para_blocks": [ + { + "type": "table", + "bbox": [ + 133, + 136, + 479, + 222 + ], + "blocks": [ + { + "bbox": [ + 140, + 114, + 473, + 125 + ], + "lines": [ + { + "bbox": [ + 140, + 114, + 473, + 125 + ], + "spans": [ + { + "bbox": [ + 140, + 114, + 473, + 125 + ], + "type": "text", + "content": "Table 1: Evaluation of PFEDEDIT with other SOTA works on different datasets." + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 133, + 136, + 479, + 222 + ], + "lines": [ + { + "bbox": [ + 133, + 136, + 479, + 222 + ], + "spans": [ + { + "bbox": [ + 133, + 136, + 479, + 222 + ], + "type": "table", + "html": "
DatasetDH-levelFedAvg 20FedPer 2FedBN 18FedRep 4DisTrans 36MOON 14APFL 6PFEDEDIT
CIFAR-100 % (i.i.d)0.9420.8900.92720.8890.9470.9220.8910.955
40%0.9230.8730.9260.8840.9440.9310.8890.953
60%0.9390.8800.9400.9200.9600.9270.9100.965
80%0.8060.8870.9450.9330.9540.9380.9150.962
100%0.8300.9020.9610.9800.9720.9400.9270.989
CIFAR-1000 % (i.i.d)0.7830.7330.7460.7420.7880.7460.7440.790
40%0.7790.7530.7760.7470.7790.7620.7530.825
60%0.7720.7480.8020.7930.8300.7680.8020.855
80%0.7370.7690.8620.8590.8870.7700.8330.890
100%0.6640.7900.9040.8800.9030.7730.8590.930
Lyme0%0.7070.7000.7330.7300.7450.6990.7220.765
50 % (i.i.d)0.7160.8080.8210.8330.8500.7680.7810.912
", + "image_path": "5c902251abfca3dc290541ab29acb7ad049912acb2a05cf75550409b9c0d8140.jpg" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "table_body" + } + ], + "index": 3 + }, + { + "type": "table", + "bbox": [ + 134, + 263, + 478, + 326 + ], + "blocks": [ + { + "bbox": [ + 132, + 231, + 481, + 255 + ], + "lines": [ + { + "bbox": [ + 132, + 231, + 481, + 255 + ], + "spans": [ + { + "bbox": [ + 132, + 231, + 481, + 255 + ], + "type": "text", + "content": "Table 2: Best Accuracy of PFEDEdit/FedAvg and number of rounds to achieve it. (- represents not achieved within 100 rounds)" + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 134, + 263, + 478, + 326 + ], + "lines": [ + { + "bbox": [ + 134, + 263, + 478, + 326 + ], + "spans": [ + { + "bbox": [ + 134, + 263, + 478, + 326 + ], + "type": "table", + "html": "
Dataset| PFEDEDIT/FedAvg# Rounds for target accuracy (PFEDEDIT/FedAvg)
>0.6>0.7>0.75>0.8>0.85>0.9
CIFAR-100.962/0.8064/56/87/139/3311/-13/-
CIFAR-1000.890/0.7376/98/2410/-12/-18/--/-
Lyme0.912/0.7163/324/875/-5/-9/-25/-
", + "image_path": "5597e1f04f9ea2fb5ce7553d4fefed3edcc29e8db8b8d92c7a2bd5772856f87b.jpg" + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "table_body" + } + ], + "index": 5 + }, + { + "type": "table", + "bbox": [ + 134, + 367, + 478, + 425 + ], + "blocks": [ + { + "bbox": [ + 132, + 335, + 481, + 359 + ], + "lines": [ + { + "bbox": [ + 132, + 335, + 481, + 359 + ], + "spans": [ + { + "bbox": [ + 132, + 335, + 481, + 359 + ], + "type": "text", + "content": "Table 3: Accuracy of PFEDEDIT with different editing ratios (in comparison with FedAvg)." + } + ] + } + ], + "index": 6, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 134, + 367, + 478, + 425 + ], + "lines": [ + { + "bbox": [ + 134, + 367, + 478, + 425 + ], + "spans": [ + { + "bbox": [ + 134, + 367, + 478, + 425 + ], + "type": "table", + "html": "
Dataset| FedAvg| PFEDEDIT
1%3%5%7%9%
CIFAR-100.8060.9330.9470.9480.9620.952
CIFAR-1000.7480.8600.8780.8850.8900.874
Lyme0.7160.8680.8800.8890.9120.910
", + "image_path": "d11e0e0fc645060808fe3c62ce5b98eefc608790579fbcc39406353dda745c9a.jpg" + } + ] + } + ], + "index": 7, + "angle": 0, + "type": "table_body" + } + ], + "index": 7 + }, + { + "bbox": [ + 132, + 447, + 481, + 472 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 447, + 481, + 472 + ], + "spans": [ + { + "bbox": [ + 132, + 447, + 481, + 472 + ], + "type": "text", + "content": "and PFEDEDIT edits the global model and tailors it towards the local distribution, which makes it faster to generalize on local distributions." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 132, + 489, + 248, + 502 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 489, + 248, + 502 + ], + "spans": [ + { + "bbox": [ + 132, + 489, + 248, + 502 + ], + "type": "text", + "content": "5.4 Parameter Study" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 130, + 510, + 482, + 605 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 510, + 482, + 605 + ], + "spans": [ + { + "bbox": [ + 130, + 510, + 482, + 605 + ], + "type": "text", + "content": "Model Editing Ratio " + }, + { + "bbox": [ + 130, + 510, + 482, + 605 + ], + "type": "inline_equation", + "content": "(k)" + }, + { + "bbox": [ + 130, + 510, + 482, + 605 + ], + "type": "text", + "content": " PFEDEDIT allows users to pre-define " + }, + { + "bbox": [ + 130, + 510, + 482, + 605 + ], + "type": "inline_equation", + "content": "k" + }, + { + "bbox": [ + 130, + 510, + 482, + 605 + ], + "type": "text", + "content": ", the number of layers to be edited. It then edits the top-" + }, + { + "bbox": [ + 130, + 510, + 482, + 605 + ], + "type": "inline_equation", + "content": "k" + }, + { + "bbox": [ + 130, + 510, + 482, + 605 + ], + "type": "text", + "content": " layers, which are considered to be the most effective in improving model accuracy. In this section we set up an experiment to show the correlation between the number of modified layers and the model accuracy. Intuitively, if the number of edited layers increases, the model accuracy gets higher as greater part of the model is tailored towards local distribution. However, such increase may not hold all the time as there are layers that do not possess any learnable parameters and the model may converge." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 130, + 605, + 482, + 666 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 605, + 482, + 666 + ], + "spans": [ + { + "bbox": [ + 130, + 605, + 482, + 666 + ], + "type": "text", + "content": "We choose the ratios of edited layers to be " + }, + { + "bbox": [ + 130, + 605, + 482, + 666 + ], + "type": "inline_equation", + "content": "1\\%" + }, + { + "bbox": [ + 130, + 605, + 482, + 666 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 130, + 605, + 482, + 666 + ], + "type": "inline_equation", + "content": "3\\%" + }, + { + "bbox": [ + 130, + 605, + 482, + 666 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 130, + 605, + 482, + 666 + ], + "type": "inline_equation", + "content": "5\\%" + }, + { + "bbox": [ + 130, + 605, + 482, + 666 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 130, + 605, + 482, + 666 + ], + "type": "inline_equation", + "content": "7\\%" + }, + { + "bbox": [ + 130, + 605, + 482, + 666 + ], + "type": "text", + "content": ", and " + }, + { + "bbox": [ + 130, + 605, + 482, + 666 + ], + "type": "inline_equation", + "content": "9\\%" + }, + { + "bbox": [ + 130, + 605, + 482, + 666 + ], + "type": "text", + "content": " to the total model layers and we round up the number in each case. Shown in Table 3, as the editing ratio increases, the overall testing accuracy gets higher. The accuracy improvement is greater from improving the ratio from " + }, + { + "bbox": [ + 130, + 605, + 482, + 666 + ], + "type": "inline_equation", + "content": "1\\%" + }, + { + "bbox": [ + 130, + 605, + 482, + 666 + ], + "type": "text", + "content": " to " + }, + { + "bbox": [ + 130, + 605, + 482, + 666 + ], + "type": "inline_equation", + "content": "3\\%" + }, + { + "bbox": [ + 130, + 605, + 482, + 666 + ], + "type": "text", + "content": " and the increasing rate gets flatter when keep improving the ratio from " + }, + { + "bbox": [ + 130, + 605, + 482, + 666 + ], + "type": "inline_equation", + "content": "3\\%" + }, + { + "bbox": [ + 130, + 605, + 482, + 666 + ], + "type": "text", + "content": " to" + } + ] + } + ], + "index": 11 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "spans": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "type": "text", + "content": "12" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 166, + 91, + 225, + 101 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 166, + 91, + 225, + 101 + ], + "spans": [ + { + "bbox": [ + 166, + 91, + 225, + 101 + ], + "type": "text", + "content": "H. Yuan et al." + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 11 + }, + { + "para_blocks": [ + { + "bbox": [ + 130, + 115, + 482, + 175 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 115, + 482, + 175 + ], + "spans": [ + { + "bbox": [ + 130, + 115, + 482, + 175 + ], + "type": "inline_equation", + "content": "5\\%" + }, + { + "bbox": [ + 130, + 115, + 482, + 175 + ], + "type": "text", + "content": ", or " + }, + { + "bbox": [ + 130, + 115, + 482, + 175 + ], + "type": "inline_equation", + "content": "5\\%" + }, + { + "bbox": [ + 130, + 115, + 482, + 175 + ], + "type": "text", + "content": " to " + }, + { + "bbox": [ + 130, + 115, + 482, + 175 + ], + "type": "inline_equation", + "content": "7\\%" + }, + { + "bbox": [ + 130, + 115, + 482, + 175 + ], + "type": "text", + "content": ". At the end, when the ratio is set from " + }, + { + "bbox": [ + 130, + 115, + 482, + 175 + ], + "type": "inline_equation", + "content": "7\\%" + }, + { + "bbox": [ + 130, + 115, + 482, + 175 + ], + "type": "text", + "content": " to " + }, + { + "bbox": [ + 130, + 115, + 482, + 175 + ], + "type": "inline_equation", + "content": "9\\%" + }, + { + "bbox": [ + 130, + 115, + 482, + 175 + ], + "type": "text", + "content": ", the accuracy starts to drop a little bit, implying that the model may converge on that accuracy. Compared to the baseline, PFEDEDIT improves the model accuracy from 0.13, with editing " + }, + { + "bbox": [ + 130, + 115, + 482, + 175 + ], + "type": "inline_equation", + "content": "1\\%" + }, + { + "bbox": [ + 130, + 115, + 482, + 175 + ], + "type": "text", + "content": " layers, to 0.16, with editing " + }, + { + "bbox": [ + 130, + 115, + 482, + 175 + ], + "type": "inline_equation", + "content": "7\\%" + }, + { + "bbox": [ + 130, + 115, + 482, + 175 + ], + "type": "text", + "content": " layers, on three datasets on average." + } + ] + } + ], + "index": 2 + }, + { + "type": "table", + "bbox": [ + 132, + 208, + 313, + 253 + ], + "blocks": [ + { + "bbox": [ + 132, + 193, + 308, + 207 + ], + "lines": [ + { + "bbox": [ + 132, + 193, + 308, + 207 + ], + "spans": [ + { + "bbox": [ + 132, + 193, + 308, + 207 + ], + "type": "text", + "content": "Table 4: Evaluation of PFedEdit with representative subset ratio (i.e., " + }, + { + "bbox": [ + 132, + 193, + 308, + 207 + ], + "type": "inline_equation", + "content": "p" + }, + { + "bbox": [ + 132, + 193, + 308, + 207 + ], + "type": "text", + "content": " )." + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 132, + 208, + 313, + 253 + ], + "lines": [ + { + "bbox": [ + 132, + 208, + 313, + 253 + ], + "spans": [ + { + "bbox": [ + 132, + 208, + 313, + 253 + ], + "type": "table", + "html": "
DatasetSubser ratio
p=0.01p=0.025p=0.05p=0.08p=0.1
CIFAR-100.9510.9530.9540.9540.962
CIFAR-100-0.8750.8840.8860.890
", + "image_path": "1684719a56c2f35cf4b4b44b7dc17fe5d6fd237507db8786efa3033cefbc27fa.jpg" + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "table_body" + } + ], + "index": 4 + }, + { + "type": "table", + "bbox": [ + 323, + 208, + 480, + 254 + ], + "blocks": [ + { + "bbox": [ + 323, + 193, + 480, + 207 + ], + "lines": [ + { + "bbox": [ + 323, + 193, + 480, + 207 + ], + "spans": [ + { + "bbox": [ + 323, + 193, + 480, + 207 + ], + "type": "text", + "content": "Table 5: Evaluation of PFedEdit using different metrics in assessing model editing" + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 323, + 208, + 480, + 254 + ], + "lines": [ + { + "bbox": [ + 323, + 208, + 480, + 254 + ], + "spans": [ + { + "bbox": [ + 323, + 208, + 480, + 254 + ], + "type": "table", + "html": "
DatasetDifferent metrics
AccuracyLossTE scorePrediction list
CIFAR-100.5400.9050.8850.962
CIFAR-1000.7050.8860.8210.890
", + "image_path": "4def843f70d054d4219c6ee4ca2affd09dc28d19721a01192c0ceb8c9c83b373.jpg" + } + ] + } + ], + "index": 6, + "angle": 0, + "type": "table_body" + } + ], + "index": 6 + }, + { + "bbox": [ + 130, + 270, + 482, + 545 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 270, + 482, + 545 + ], + "spans": [ + { + "bbox": [ + 130, + 270, + 482, + 545 + ], + "type": "text", + "content": "Representative Subset Ratio " + }, + { + "bbox": [ + 130, + 270, + 482, + 545 + ], + "type": "inline_equation", + "content": "(p)" + }, + { + "bbox": [ + 130, + 270, + 482, + 545 + ], + "type": "text", + "content": " As mentioned in Section 4, PFEDEDIT queries the model using a representative subset " + }, + { + "bbox": [ + 130, + 270, + 482, + 545 + ], + "type": "inline_equation", + "content": "D_{p}" + }, + { + "bbox": [ + 130, + 270, + 482, + 545 + ], + "type": "text", + "content": " of the training set " + }, + { + "bbox": [ + 130, + 270, + 482, + 545 + ], + "type": "inline_equation", + "content": "D_{train}" + }, + { + "bbox": [ + 130, + 270, + 482, + 545 + ], + "type": "text", + "content": ". PFEDEDIT employs a ratio " + }, + { + "bbox": [ + 130, + 270, + 482, + 545 + ], + "type": "inline_equation", + "content": "p" + }, + { + "bbox": [ + 130, + 270, + 482, + 545 + ], + "type": "text", + "content": " to pre-define the subset size compared to the training set, " + }, + { + "bbox": [ + 130, + 270, + 482, + 545 + ], + "type": "inline_equation", + "content": "|D_{p}| = p * |D_{train}|" + }, + { + "bbox": [ + 130, + 270, + 482, + 545 + ], + "type": "text", + "content": ". We want to ensure that the subset is representative enough for potential testing image samples while not introducing too much computation overhead. Therefore, we conduct a study on the size of the subset versus the performance of PFEDEDIT. Intuitively, a larger size covers a broader local data distribution and leads to a more accurate layer locating and thus a higher model accuracy. However, on the other hand, taken a larger subset from the original training set reduces the size of the remaining set. We want to find an optimal size of the subset that is large enough to give a complete representation of local data distribution but also leaves enough samples for the local training. We split the local training set into a subset that takes up of " + }, + { + "bbox": [ + 130, + 270, + 482, + 545 + ], + "type": "inline_equation", + "content": "p * |D_{train}|" + }, + { + "bbox": [ + 130, + 270, + 482, + 545 + ], + "type": "text", + "content": " image samples, and a training set that takes the rest " + }, + { + "bbox": [ + 130, + 270, + 482, + 545 + ], + "type": "inline_equation", + "content": "|D_{train}^{*}| = |D_{train}| - |D_{p}|" + }, + { + "bbox": [ + 130, + 270, + 482, + 545 + ], + "type": "text", + "content": " image samples. To control the variable, we fix the size of " + }, + { + "bbox": [ + 130, + 270, + 482, + 545 + ], + "type": "inline_equation", + "content": "D_{train}^{*}" + }, + { + "bbox": [ + 130, + 270, + 482, + 545 + ], + "type": "text", + "content": ", and we choose different " + }, + { + "bbox": [ + 130, + 270, + 482, + 545 + ], + "type": "inline_equation", + "content": "p" + }, + { + "bbox": [ + 130, + 270, + 482, + 545 + ], + "type": "text", + "content": " values from " + }, + { + "bbox": [ + 130, + 270, + 482, + 545 + ], + "type": "inline_equation", + "content": "\\{0.1, 0.08, 0.05, 0.025, 0.01\\}" + }, + { + "bbox": [ + 130, + 270, + 482, + 545 + ], + "type": "text", + "content": " to adjust the subset size. We conduct the evaluation on CIFAR-10 and CIFAR-100 [12], with 10 clients and each client is assigned 2 classes of image samples. Each client possesses 4,500 image samples as training set, and the subset size varies from 50 image samples (" + }, + { + "bbox": [ + 130, + 270, + 482, + 545 + ], + "type": "inline_equation", + "content": "p" + }, + { + "bbox": [ + 130, + 270, + 482, + 545 + ], + "type": "text", + "content": " value 0.01) to 500 image samples (" + }, + { + "bbox": [ + 130, + 270, + 482, + 545 + ], + "type": "inline_equation", + "content": "p" + }, + { + "bbox": [ + 130, + 270, + 482, + 545 + ], + "type": "text", + "content": " value 0.1). Note that the " + }, + { + "bbox": [ + 130, + 270, + 482, + 545 + ], + "type": "inline_equation", + "content": "p" + }, + { + "bbox": [ + 130, + 270, + 482, + 545 + ], + "type": "text", + "content": " value starts from 0.025 for CIFAR-100 because a subset that is 0.01 times the size of training set does not include image samples from all classes, i.e., 50 image samples do not cover 100 classes thus the subset is not representative." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 130, + 545, + 482, + 666 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 545, + 482, + 666 + ], + "spans": [ + { + "bbox": [ + 130, + 545, + 482, + 666 + ], + "type": "text", + "content": "As Table 4 shows, the overall trend of the testing accuracy keeps increasing as " + }, + { + "bbox": [ + 130, + 545, + 482, + 666 + ], + "type": "inline_equation", + "content": "p" + }, + { + "bbox": [ + 130, + 545, + 482, + 666 + ], + "type": "text", + "content": " value gets larger from 0.01 times the training set to 0.1 times the training set for both datasets. However, the overall accuracy improvement is minimal compared to the subset size increase, demonstrating the capability of PFEDEDIT in improving model accuracy with a small size subset. For CIFAR-10, as " + }, + { + "bbox": [ + 130, + 545, + 482, + 666 + ], + "type": "inline_equation", + "content": "p" + }, + { + "bbox": [ + 130, + 545, + 482, + 666 + ], + "type": "text", + "content": " value increases from 0.01 to 0.08, which is an 8 times increase, the model accuracy is improved from 0.951 to 0.954, which is a " + }, + { + "bbox": [ + 130, + 545, + 482, + 666 + ], + "type": "inline_equation", + "content": "0.3\\%" + }, + { + "bbox": [ + 130, + 545, + 482, + 666 + ], + "type": "text", + "content": " improvement. As " + }, + { + "bbox": [ + 130, + 545, + 482, + 666 + ], + "type": "inline_equation", + "content": "p" + }, + { + "bbox": [ + 130, + 545, + 482, + 666 + ], + "type": "text", + "content": " value increases from 0.08 to 0.1, the accuracy is further improved from 0.954 to 0.962, which is a " + }, + { + "bbox": [ + 130, + 545, + 482, + 666 + ], + "type": "inline_equation", + "content": "0.8\\%" + }, + { + "bbox": [ + 130, + 545, + 482, + 666 + ], + "type": "text", + "content": " increase. For CIFAR-100, the accuracy improvement brought by increasing " + }, + { + "bbox": [ + 130, + 545, + 482, + 666 + ], + "type": "inline_equation", + "content": "p" + }, + { + "bbox": [ + 130, + 545, + 482, + 666 + ], + "type": "text", + "content": " value is greater than that in CIFAR-10 as the model accuracy is improved from" + } + ] + } + ], + "index": 8 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 400, + 91, + 447, + 100 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 400, + 91, + 447, + 100 + ], + "spans": [ + { + "bbox": [ + 400, + 91, + 447, + 100 + ], + "type": "text", + "content": "PFEDEDIT" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 470, + 91, + 481, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 470, + 91, + 481, + 100 + ], + "spans": [ + { + "bbox": [ + 470, + 91, + 481, + 100 + ], + "type": "text", + "content": "13" + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 12 + }, + { + "para_blocks": [ + { + "bbox": [ + 130, + 116, + 482, + 175 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 116, + 482, + 175 + ], + "spans": [ + { + "bbox": [ + 130, + 116, + 482, + 175 + ], + "type": "text", + "content": "0.875 to 0.890 when " + }, + { + "bbox": [ + 130, + 116, + 482, + 175 + ], + "type": "inline_equation", + "content": "p" + }, + { + "bbox": [ + 130, + 116, + 482, + 175 + ], + "type": "text", + "content": " value changes from 0.025 to 0.1. Therefore, PFEDEDIT is computation-friendly as it requires only 50 image samples to get a 0.951 model accuracy and 500 image samples to achieve a 0.962 accuracy on CIFAR-10. Similarly, PFEDEDIT requires only 125 image samples as the subset to achieve a 0.875 accuracy and 500 image samples to gets a 0.890 accuracy on CIFAR-100." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 130, + 193, + 482, + 276 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 193, + 482, + 276 + ], + "spans": [ + { + "bbox": [ + 130, + 193, + 482, + 276 + ], + "type": "text", + "content": "Metrics Selection In this section, we conduct experiments to explore possible metric selections for PFEDEDIT when assessing each model editing. Besides the prediction list that demonstrates all four combinations of TE score and ground-truth mask, we also choose model accuracy, loss value, and TE score. We select cross-entropy loss as our loss metric, which is defined as " + }, + { + "bbox": [ + 130, + 193, + 482, + 276 + ], + "type": "inline_equation", + "content": "H(P,Q) = -\\sum_{x \\in \\mathcal{X}} p(x) \\log q(x)" + }, + { + "bbox": [ + 130, + 193, + 482, + 276 + ], + "type": "text", + "content": ", given two probability distribution " + }, + { + "bbox": [ + 130, + 193, + 482, + 276 + ], + "type": "inline_equation", + "content": "P" + }, + { + "bbox": [ + 130, + 193, + 482, + 276 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 130, + 193, + 482, + 276 + ], + "type": "inline_equation", + "content": "Q" + }, + { + "bbox": [ + 130, + 193, + 482, + 276 + ], + "type": "text", + "content": ", and the TE score is introduced in Section 4." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 130, + 277, + 482, + 444 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 277, + 482, + 444 + ], + "spans": [ + { + "bbox": [ + 130, + 277, + 482, + 444 + ], + "type": "text", + "content": "As Table 5 shows, the prediction list employed by PFEDEDIT outperforms other metrics and the accuracy gap ranges from 0.057 to 0.422. Among all possible metrics, using accuracy to assess model editorings leads to the lowest accuracy, 0.540 on CIFAR-10 and 0.705 on CIFAR-100, lower than the baseline FedAvg [20]. The reason is that the effects of some editorings are subtle and they only cause the probability change but not necessarily results in label change. As the way that accuracy is computed only counts on samples that are correctly classified, small probability change that do not alter the final prediction will be considered as no improvement at all. On the other hand, the TE score reflects the probability change on ground-truth label, but it does not reflects any label change. In other words, an increase in the probability of the ground-truth label suggests a positive editing, but a decrease in probability does not necessarily indicate a negative editing. From the accuracy perspective, the edited model performs identically to the original model despite these probability changes." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 130, + 444, + 482, + 540 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 444, + 482, + 540 + ], + "spans": [ + { + "bbox": [ + 130, + 444, + 482, + 540 + ], + "type": "text", + "content": "While employing loss functions as the metric provides insights into the probability of the ground-truth class, the averaged loss value may exhibit bias if outliers have a disproportionate impact on the probability of the ground-truth class. For example, If an editing improves the model accuracy in correctly classifying two samples with a slight increase in the ground-truth class and misclassifies one sample with a notable drop in the ground-truth class, the improvement introduced by the correct classifications might be nullified by the incorrect one, which leads to a biased averaged loss value with respect to each data sample." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 132, + 557, + 220, + 570 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 557, + 220, + 570 + ], + "spans": [ + { + "bbox": [ + 132, + 557, + 220, + 570 + ], + "type": "text", + "content": "6 Conclusion" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 130, + 582, + 482, + 665 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 582, + 482, + 665 + ], + "spans": [ + { + "bbox": [ + 130, + 582, + 482, + 665 + ], + "type": "text", + "content": "In this paper, we design the first personalized federated learning framework that utilizes automated model editing as a core strategy to improve the model's utility. Our approach leverages layer-specific editing, which automatically identifies specific layers to be edited based on local data distribution, allowing for more effective adaptation of the global model to heterogeneous data distributions across various clients. Our evaluation demonstrates that PFEDEDIT significantly improves the accuracy of the global model from non-i.i.d. to i.i.d. data distribution." + } + ] + } + ], + "index": 7 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "spans": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "type": "text", + "content": "14" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 166, + 91, + 225, + 101 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 166, + 91, + 225, + 101 + ], + "spans": [ + { + "bbox": [ + 166, + 91, + 225, + 101 + ], + "type": "text", + "content": "H. Yuan et al." + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 13 + }, + { + "para_blocks": [ + { + "bbox": [ + 133, + 114, + 240, + 129 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 114, + 240, + 129 + ], + "spans": [ + { + "bbox": [ + 133, + 114, + 240, + 129 + ], + "type": "text", + "content": "Acknowledgments" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 130, + 140, + 482, + 213 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 140, + 482, + 213 + ], + "spans": [ + { + "bbox": [ + 130, + 140, + 482, + 213 + ], + "type": "text", + "content": "This work was supported in part by National Science Foundation (NSF) under grants OAC-23-19742 and Johns Hopkins University Institute for Assured Autonomy (IAA) with grants 80052272 and 80052273. The views and conclusions contained herein are those of the authors and should not be interpreted as necessarily representing the official policies or endorsements, either expressed or implied, of NSF or JHU-IAA." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 133, + 232, + 198, + 244 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 232, + 198, + 244 + ], + "spans": [ + { + "bbox": [ + 133, + 232, + 198, + 244 + ], + "type": "text", + "content": "References" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 138, + 257, + 480, + 665 + ], + "type": "list", + "angle": 0, + "index": 21, + "blocks": [ + { + "bbox": [ + 138, + 257, + 480, + 279 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 257, + 480, + 279 + ], + "spans": [ + { + "bbox": [ + 138, + 257, + 480, + 279 + ], + "type": "text", + "content": "1. Alam, S., Liu, L., Yan, M., Zhang, M.: Fedrolex: Model-heterogeneous federated learning with rolling sub-model extraction (2023)" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 138, + 280, + 480, + 301 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 280, + 480, + 301 + ], + "spans": [ + { + "bbox": [ + 138, + 280, + 480, + 301 + ], + "type": "text", + "content": "2. Arivazhagan, M.G., Aggarwal, V., Singh, A.K., Choudhary, S.: Federated learning with personalization layers. arXiv preprint arXiv:1912.00818 (2019)" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 138, + 302, + 480, + 323 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 302, + 480, + 323 + ], + "spans": [ + { + "bbox": [ + 138, + 302, + 480, + 323 + ], + "type": "text", + "content": "3. Basu, S., Zhao, N., Morariu, V., Feizi, S., Manjunatha, V.: Localizing and editing knowledge in text-to-image generative models (2023)" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 138, + 324, + 480, + 356 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 324, + 480, + 356 + ], + "spans": [ + { + "bbox": [ + 138, + 324, + 480, + 356 + ], + "type": "text", + "content": "4. Collins, L., Hassani, H., Mokhtari, A., Shakkottai, S.: Exploiting shared representations for personalized federated learning. In: Proceedings of the 38th International Conference on Machine Learning (2021)" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 138, + 357, + 480, + 378 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 357, + 480, + 378 + ], + "spans": [ + { + "bbox": [ + 138, + 357, + 480, + 378 + ], + "type": "text", + "content": "5. Dai, D., Dong, L., Hao, Y., Sui, Z., Chang, B., Wei, F.: Knowledge neurons in pretrained transformers. arXiv preprint arXiv:2104.08696 (2021)" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 138, + 379, + 480, + 400 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 379, + 480, + 400 + ], + "spans": [ + { + "bbox": [ + 138, + 379, + 480, + 400 + ], + "type": "text", + "content": "6. Deng, Y., Kamani, M.M., Mahdavi, M.: Adaptive personalized federated learning (2020)" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 138, + 401, + 480, + 422 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 401, + 480, + 422 + ], + "spans": [ + { + "bbox": [ + 138, + 401, + 480, + 422 + ], + "type": "text", + "content": "7. Diao, E., Ding, J., Tarokh, V.: Heterofl: Computation and communication efficient federated learning for heterogeneous clients (2021)" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 138, + 423, + 480, + 456 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 423, + 480, + 456 + ], + "spans": [ + { + "bbox": [ + 138, + 423, + 480, + 456 + ], + "type": "text", + "content": "8. Dosovitskiy, A., Beyer, L., Kolesnikov, A., Weissenborn, D., Zhai, X., Unterthiner, T., Dehghani, M., Minderer, M., Heigold, G., Gelly, S., Uszkoreit, J., Houlsby, N.: An image is worth 16x16 words: Transformers for image recognition at scale (2021)" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 138, + 456, + 480, + 489 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 456, + 480, + 489 + ], + "spans": [ + { + "bbox": [ + 138, + 456, + 480, + 489 + ], + "type": "text", + "content": "9. He, K., Zhang, X., Ren, S., Sun, J.: Deep residual learning for image recognition. In: Proceedings of the IEEE conference on computer vision and pattern recognition. pp. 770-778 (2016)" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 138, + 490, + 480, + 510 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 490, + 480, + 510 + ], + "spans": [ + { + "bbox": [ + 138, + 490, + 480, + 510 + ], + "type": "text", + "content": "0. Karimi, B., Li, P., Li, X.: Layer-wise and dimension-wise locally adaptive federated learning (2022)" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 138, + 511, + 480, + 544 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 511, + 480, + 544 + ], + "spans": [ + { + "bbox": [ + 138, + 511, + 480, + 544 + ], + "type": "text", + "content": "1. Karimireddy, S.P., Kale, S., Mohri, M., Reddi, S., Stich, S., Suresh, A.T.: Scaffold: Stochastic controlled averaging for federated learning. In: International conference on machine learning (2020)" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 138, + 544, + 480, + 566 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 544, + 480, + 566 + ], + "spans": [ + { + "bbox": [ + 138, + 544, + 480, + 566 + ], + "type": "text", + "content": "2. Krizhevsky, A.: Learning multiple layers of features from tiny images. Tech. rep. (2009)" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 138, + 567, + 480, + 588 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 567, + 480, + 588 + ], + "spans": [ + { + "bbox": [ + 138, + 567, + 480, + 588 + ], + "type": "text", + "content": "3. Li, L., Fan, Y., Tse, M., Lin, K.Y.: A review of applications in federated learning. Computers & Industrial Engineering (2020)" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 138, + 589, + 480, + 610 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 589, + 480, + 610 + ], + "spans": [ + { + "bbox": [ + 138, + 589, + 480, + 610 + ], + "type": "text", + "content": "4. Li, Q., He, B., Song, D.: Model-contrastive federated learning. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (2021)" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 138, + 611, + 480, + 632 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 611, + 480, + 632 + ], + "spans": [ + { + "bbox": [ + 138, + 611, + 480, + 632 + ], + "type": "text", + "content": "5. Li, T., Sahu, A.K., Talwalkar, A., Smith, V.: Federated learning: Challenges, methods, and future directions. IEEE signal processing magazine (2020)" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 138, + 632, + 480, + 665 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 632, + 480, + 665 + ], + "spans": [ + { + "bbox": [ + 138, + 632, + 480, + 665 + ], + "type": "text", + "content": "6. Li, T., Sahu, A.K., Zaheer, M., Sanjabi, M., Talwalkar, A., Smith, V.: Federated optimization in heterogeneous networks. Proceedings of Machine learning and systems 2, 429-450 (2020)" + } + ] + } + ], + "index": 20 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 400, + 91, + 447, + 100 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 400, + 91, + 447, + 100 + ], + "spans": [ + { + "bbox": [ + 400, + 91, + 447, + 100 + ], + "type": "text", + "content": "PFEDEDIT" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 470, + 91, + 481, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 470, + 91, + 481, + 100 + ], + "spans": [ + { + "bbox": [ + 470, + 91, + 481, + 100 + ], + "type": "text", + "content": "15" + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 14 + }, + { + "para_blocks": [ + { + "bbox": [ + 132, + 117, + 481, + 665 + ], + "type": "list", + "angle": 0, + "index": 20, + "blocks": [ + { + "bbox": [ + 133, + 117, + 480, + 138 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 117, + 480, + 138 + ], + "spans": [ + { + "bbox": [ + 133, + 117, + 480, + 138 + ], + "type": "text", + "content": "17. Li, X., Li, S., Song, S., Yang, J., Ma, J., Yu, J.: Pmet: Precise model editing in a transformer. arXiv preprint arXiv:2308.08742 (2023)" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 133, + 140, + 481, + 172 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 140, + 481, + 172 + ], + "spans": [ + { + "bbox": [ + 133, + 140, + 481, + 172 + ], + "type": "text", + "content": "18. Li, X., Jiang, M., Zhang, X., Kamp, M., Dou, Q.: Fed{bn}: Federated learning on non-{iid} features via local batch normalization. In: International Conference on Learning Representations (2021), https://openreview.net/pdf?id=6YEQUnOQICG" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 133, + 172, + 480, + 215 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 172, + 480, + 215 + ], + "spans": [ + { + "bbox": [ + 133, + 172, + 480, + 215 + ], + "type": "text", + "content": "19. Liu, Q., Chen, C., Qin, J., Dou, Q., Heng, P.A.: Feddg: Federated domain generalization on medical image segmentation via episodic learning in continuous frequency space. The IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR) (2021)" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 132, + 216, + 480, + 248 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 216, + 480, + 248 + ], + "spans": [ + { + "bbox": [ + 132, + 216, + 480, + 248 + ], + "type": "text", + "content": "20. McMahan, B., Moore, E., Ramage, D., Hampson, S., y Arcas, B.A.: Communication-efficient learning of deep networks from decentralized data. In: Artificial intelligence and statistics (2017)" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 132, + 249, + 480, + 270 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 249, + 480, + 270 + ], + "spans": [ + { + "bbox": [ + 132, + 249, + 480, + 270 + ], + "type": "text", + "content": "21. Meng, K., Bau, D., Andonian, A., Belinkov, Y.: Locating and editing factual associations in GPT. Advances in Neural Information Processing Systems (2022)" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 132, + 271, + 480, + 293 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 271, + 480, + 293 + ], + "spans": [ + { + "bbox": [ + 132, + 271, + 480, + 293 + ], + "type": "text", + "content": "22. Meng, K., Bau, D., Andonian, A., Belinkov, Y.: Locating and editing factual associations in GPT. Advances in Neural Information Processing Systems 35 (2022)" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 132, + 293, + 480, + 325 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 293, + 480, + 325 + ], + "spans": [ + { + "bbox": [ + 132, + 293, + 480, + 325 + ], + "type": "text", + "content": "23. Meng, K., Sen Sharma, A., Andonian, A., Belinkov, Y., Bau, D.: Mass editing memory in a transformer. The Eleventh International Conference on Learning Representations (ICLR) (2023)" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 132, + 326, + 480, + 357 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 326, + 480, + 357 + ], + "spans": [ + { + "bbox": [ + 132, + 326, + 480, + 357 + ], + "type": "text", + "content": "24. Nguyen, D.C., Ding, M., Pathirana, P.N., Seneviratne, A., Li, J., Vincent Poor, H.: Federated learning for internet of things: A comprehensive survey. IEEE Communications Surveys & Tutorials (2021)" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 132, + 358, + 480, + 391 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 358, + 480, + 391 + ], + "spans": [ + { + "bbox": [ + 132, + 358, + 480, + 391 + ], + "type": "text", + "content": "25. Nguyen, D.C., Ding, M., Pham, Q.V., Pathirana, P.N., Le, L.B., Seneviratne, A., Li, J., Niyato, D., Poor, H.V.: Federated learning meets blockchain in edge computing: Opportunities and challenges. IEEE Internet of Things Journal (2021)" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 132, + 392, + 480, + 413 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 392, + 480, + 413 + ], + "spans": [ + { + "bbox": [ + 132, + 392, + 480, + 413 + ], + "type": "text", + "content": "26. Nori, H., Jenkins, S., Koch, P., Caruana, R.: Interpretml: A unified framework for machine learning interpretability. arXiv preprint arXiv:1909.09223 (2019)" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 132, + 414, + 480, + 434 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 414, + 480, + 434 + ], + "spans": [ + { + "bbox": [ + 132, + 414, + 480, + 434 + ], + "type": "text", + "content": "27. Shen, Y., Zhou, Y., Yu, L.: " + }, + { + "bbox": [ + 132, + 414, + 480, + 434 + ], + "type": "inline_equation", + "content": "\\mathrm{Cd^{2}}" + }, + { + "bbox": [ + 132, + 414, + 480, + 434 + ], + "type": "text", + "content": "-pfed: Cyclic distillation-guided channel decoupling for model personalization in federated learning (2022)" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 132, + 435, + 480, + 456 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 435, + 480, + 456 + ], + "spans": [ + { + "bbox": [ + 132, + 435, + 480, + 456 + ], + "type": "text", + "content": "28. Shokri, R., Shmatikov, V.: Privacy-preserving deep learning. In: Proceedings of the 22nd ACM SIGSAC conference on computer and communications security (2015)" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 132, + 457, + 480, + 489 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 457, + 480, + 489 + ], + "spans": [ + { + "bbox": [ + 132, + 457, + 480, + 489 + ], + "type": "text", + "content": "29. Sun, B., Huo, H., YANG, Y., Bai, B.: Partialfed: Cross-domain personalized federated learning via partial initialization. In: Advances in Neural Information Processing Systems (2021)" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 132, + 490, + 480, + 522 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 490, + 480, + 522 + ], + "spans": [ + { + "bbox": [ + 132, + 490, + 480, + 522 + ], + "type": "text", + "content": "30. Wang, P., Zhang, N., Xie, X., Yao, Y., Tian, B., Wang, M., Xi, Z., Cheng, S., Liu, K., Zheng, G., et al.: Easilyedit: An easy-to-use knowledge editing framework for large language models. arXiv preprint arXiv:2308.07269 (2023)" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 132, + 523, + 480, + 555 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 523, + 480, + 555 + ], + "spans": [ + { + "bbox": [ + 132, + 523, + 480, + 555 + ], + "type": "text", + "content": "31. Yang, Q., Liu, Y., Chen, T., Tong, Y.: Federated machine learning: Concept and applications. ACM Transactions on Intelligent Systems and Technology (TIST) (2019)" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 132, + 555, + 480, + 587 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 555, + 480, + 587 + ], + "spans": [ + { + "bbox": [ + 132, + 555, + 480, + 587 + ], + "type": "text", + "content": "32. Yang, Y., Hui, B., Yuan, H., Gong, N., Cao, Y.: PrivateFL: Accurate, differentially private federated learning via personalized data transformation. In: 32nd USENIX Security Symposium (USENIX Security 23). pp. 1595-1612 (2023)" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 132, + 588, + 480, + 632 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 588, + 480, + 632 + ], + "spans": [ + { + "bbox": [ + 132, + 588, + 480, + 632 + ], + "type": "text", + "content": "33. Yang, Y., Yuan, H., Hui, B., Gong, N., Fendley, N., Burlina, P., Cao, Y.: Fortifying federated learning against membership inference attacks via client-level input perturbation. In: 2023 53rd Annual IEEE/IFIP International Conference on Dependable Systems and Networks (DSN) (2023)" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 132, + 633, + 480, + 665 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 633, + 480, + 665 + ], + "spans": [ + { + "bbox": [ + 132, + 633, + 480, + 665 + ], + "type": "text", + "content": "34. Yang, Z., Zhang, Y., Zheng, Y., Tian, X., Peng, H., Liu, T., Han, B.: Fedfed: Feature distillation against data heterogeneity in federated learning. In: Thirty-seventh Conference on Neural Information Processing Systems (2023)" + } + ] + } + ], + "index": 19 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "spans": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "type": "text", + "content": "16" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 166, + 91, + 225, + 101 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 166, + 91, + 225, + 101 + ], + "spans": [ + { + "bbox": [ + 166, + 91, + 225, + 101 + ], + "type": "text", + "content": "H. Yuan et al." + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 15 + }, + { + "para_blocks": [ + { + "bbox": [ + 130, + 116, + 482, + 259 + ], + "type": "list", + "angle": 0, + "index": 6, + "blocks": [ + { + "bbox": [ + 130, + 116, + 482, + 160 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 116, + 482, + 160 + ], + "spans": [ + { + "bbox": [ + 130, + 116, + 482, + 160 + ], + "type": "text", + "content": "35. Yuan, H., Aucott, J., Hadzic, A., Paul, W., Villegas de Flores, M., Mathew, P., Burlina, P., Cao, Y.: Edgemixup: Embarrassingly simple data alteration to improve Lyme disease lesion segmentation and diagnosis fairness. In: Medical Image Computing and Computer Assisted Intervention - MICCAI 2023 (2023)" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 130, + 161, + 482, + 205 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 161, + 482, + 205 + ], + "spans": [ + { + "bbox": [ + 130, + 161, + 482, + 205 + ], + "type": "text", + "content": "36. Yuan, H., Hui, B., Yang, Y., Burlina, P., Gong, N.Z., Cao, Y.: Addressing heterogeneity in federated learning via distributional transformation. In: Computer Vision - ECCV 2022: 17th European Conference, Tel Aviv, Israel, October 23-27, 2022, Proceedings, Part XXXVIII. p. 179-195 (2022)" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 131, + 205, + 481, + 237 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 131, + 205, + 481, + 237 + ], + "spans": [ + { + "bbox": [ + 131, + 205, + 481, + 237 + ], + "type": "text", + "content": "37. Zhang, J., Hua, Y., Wang, H., Song, T., Xue, Z., Ma, R., Guan, H.: Fedala: Adaptive local aggregation for personalized federated learning. In: Proceedings of the AAAI Conference on Artificial Intelligence (2023)" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 131, + 237, + 481, + 259 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 131, + 237, + 481, + 259 + ], + "spans": [ + { + "bbox": [ + 131, + 237, + 481, + 259 + ], + "type": "text", + "content": "38. Zhou, T., Konukoglu, E.: FedFA: Federated feature augmentation. In: The Eleventh International Conference on Learning Representations (ICLR) (2023)" + } + ] + } + ], + "index": 5 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 400, + 91, + 447, + 100 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 400, + 91, + 447, + 100 + ], + "spans": [ + { + "bbox": [ + 400, + 91, + 447, + 100 + ], + "type": "text", + "content": "PFEDEDIT" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 470, + 91, + 481, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 470, + 91, + 481, + 100 + ], + "spans": [ + { + "bbox": [ + 470, + 91, + 481, + 100 + ], + "type": "text", + "content": "17" + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 16 + } + ], + "_backend": "vlm", + "_version_name": "2.6.4" +} \ No newline at end of file diff --git a/2024/PILoRA_ Prototype Guided Incremental LoRA for Federated Class-Incremental Learning/b10fddc1-55da-4bf4-bfbe-8037f4bdbf91_content_list.json b/2024/PILoRA_ Prototype Guided Incremental LoRA for Federated Class-Incremental Learning/b10fddc1-55da-4bf4-bfbe-8037f4bdbf91_content_list.json new file mode 100644 index 0000000000000000000000000000000000000000..f28d8571a500be19302ca5122dad15edff4bfe67 --- /dev/null +++ b/2024/PILoRA_ Prototype Guided Incremental LoRA for Federated Class-Incremental Learning/b10fddc1-55da-4bf4-bfbe-8037f4bdbf91_content_list.json @@ -0,0 +1,1831 @@ +[ + { + "type": "text", + "text": "PILoRA: Prototype Guided Incremental LoRA for Federated Class-Incremental Learning", + "text_level": 1, + "bbox": [ + 233, + 140, + 772, + 186 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Haiyang Guo $^{1,2}$ , Fei Zhu $^{3}$ , Wenzhuo Liu $^{1,2}$ , Xu-Yao Zhang $^{1,2\\star}$ , and Cheng-Lin Liu $^{1,2}$", + "bbox": [ + 343, + 212, + 656, + 244 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "$^{1}$ MAIS, Institute of Automation, Chinese Academy of Sciences", + "bbox": [ + 285, + 253, + 712, + 268 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "$^{2}$ School of Artificial Intelligence, University of Chinese Academy of Sciences", + "bbox": [ + 243, + 268, + 756, + 282 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "3 Centre for Artificial Intelligence and Robotics, Hong Kong Institute of Science and Innovation, Chinese Academy of Sciences", + "bbox": [ + 217, + 282, + 782, + 310 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "{guohaiyang2023, zhufei2018}@ia.ac.cn, liuwenzhuo20@mails.ucas.ac.cn, xyz, liucl}@nlpr.ia.ac.cn", + "bbox": [ + 233, + 311, + 766, + 339 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Abstract. Existing federated learning methods have effectively dealt with decentralized learning in scenarios involving data privacy and non-IID data. However, in real-world situations, each client dynamically learns new classes, requiring the global model to classify all seen classes. To effectively mitigate catastrophic forgetting and data heterogeneity under low communication costs, we propose a simple and effective method named PILoRA. On the one hand, we adopt prototype learning to learn better feature representations and leverage the heuristic information between prototypes and class features to design a prototype re-weight module to solve the classifier bias caused by data heterogeneity without retraining the classifier. On the other hand, we view incremental learning as the process of learning distinct task vectors and encoding them within different LoRA parameters. Accordingly, we propose Incremental LoRA to mitigate catastrophic forgetting. Experimental results on standard datasets indicate that our method outperforms the state-of-the-art approaches significantly. More importantly, our method exhibits strong robustness and superiority in different settings and degrees of data heterogeneity. The code is available at https://github.com/GhyO501/PILORA.", + "bbox": [ + 259, + 381, + 743, + 632 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Keywords: Federated Learning $\\cdot$ Class Incremental Learning", + "bbox": [ + 261, + 643, + 679, + 660 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "1 Introduction", + "text_level": 1, + "bbox": [ + 215, + 688, + 375, + 704 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Federated learning (FL) [40] is a novel distributed machine learning paradigm that enables multiple data owners to collaboratively train a shared model while ensuring the privacy of their local data. In recent years, with the increasing emphasis on data privacy in society and the refinement of relevant regulations [52], FL has experienced rapid development and has been widely applied in various real-world scenarios [14,20,30,41,47].", + "bbox": [ + 212, + 722, + 787, + 815 + ], + "page_idx": 0 + }, + { + "type": "page_footnote", + "text": "* Corresponding author", + "bbox": [ + 217, + 824, + 382, + 840 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Existing FL methods [19, 28, 40] typically rely on the closed-world assumption [6, 13, 65], that is, the number of classes seen by the model remains the same and unchanged during both the training and testing stages. However, the real world is dynamic and constantly changing, the local clients often need to receive new data to update the global model continuously. Therefore, Federated Class Incremental Learning (FCIL) [10, 59] has been proposed to handle FL tasks in dynamic scenarios. Specifically, each local client can only update the model with new class data at each stage and upload the model parameters to the global server for aggregation, while the global model needs to maintain its discriminative ability for all seen classes. Moreover, the data distribution of local data follows the non-independent and identically distributed (non-IID) assumption [27].", + "bbox": [ + 212, + 146, + 787, + 329 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "FCIL presents a more realistic setting for real-world applications, but it also introduces greater challenges, as FCIL needs to simultaneously address the issue of catastrophic forgetting [23, 39] and data heterogeneity resulting from non-IID. In existing studies, one way is to store a subset of data from old classes and train them together when learning new tasks [9, 10], but the amount of old data stored is strictly limited due to privacy protection requirements. An alternative way is to utilize generative models to generate pseudo-samples of old data [43, 61], thus retaining the ability to classify the old classes. However, training a good generator will cause greater computational overhead, and the generator also suffers from catastrophic forgetting. Another direction is to use pre-trained models for fine-tuning [1, 33], these methods store different stages of knowledge by maintaining a pool of modules and inserting the corresponding modules at inference time based on the similarity between the input data and the modules. However, storing these modules takes up additional memory space. Although they provide different perspectives for solving FCIL, there is still a large unexplored space, especially the performance of the model under different non-IID settings and different degrees of data heterogeneity. So in order to better address this issue, we start by addressing a fundamental question, what is the panacea for dealing with CIL and FL?", + "bbox": [ + 212, + 334, + 787, + 622 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "In other words, if we can find common ground in dealing with CIL and FL, it will be of great help in solving the FCIL problem. Specifically, we observe that: 1) At the level of feature representation, both FL and CIL tasks require models to learn a intra-class compact and inter-class separable feature representation. On the one hand, for the CIL task, this kind of feature representation helps to reduce the overlap between features of the old and new classes in the deep feature space, thus alleviating the forgetting of old knowledge. For this reason, PASS [66] introduces self-supervised learning to assist the learning of feature representations. On the other hand, in the FL task, FedEx [51] imposes a constraint on the deep features of samples belonging to the same class. This constraint ensures that these features are proximate to the global prototypes of their respective classes. 2) At the classifier level, classifier drift is the common enemy of both. In CIL, when learning new classes, the decision boundaries learned from old classes can be severely disrupted, resulting in a significant bias", + "bbox": [ + 212, + 628, + 789, + 840 + ], + "page_idx": 1 + }, + { + "type": "page_number", + "text": "2", + "bbox": [ + 217, + 114, + 228, + 126 + ], + "page_idx": 1 + }, + { + "type": "header", + "text": "H. Guo et al.", + "bbox": [ + 271, + 114, + 359, + 127 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "in the classification layer [64, 66]. To address this issue, some methods [2, 3, 46] alleviate bias by directly retaining a portion of the data from previous categories and training the model with new data. Other methods [64, 66] address classifier bias by retaining pseudo features of old classes to assist in training the classification layer; In FL, CCVR [38] assesses the similarity among different layers within local models under data heterogeneity and reveals that classifiers exhibit the lowest similarity. This suggests that the classifier of each client models is seriously biased to the local data. Similarly to CIL, [38, 48] utilize the statistical information of local features to retrain the classifier on the global server to alleviate classifier bias.", + "bbox": [ + 212, + 146, + 787, + 297 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Motivated by the above findings, in this paper, we propose a Prototype Guided Incremental LoRA (PILoRA) model to tackle the FCIL problem. Specifically, we adopt prototype classifiers in our model to learn intraclass compact and inter-class separable feature representations, which is beneficial for both CIL and FL [28, 35, 56, 57]. We use the pre-trained Transformer model (ViT) as the backbone because the global interactions learned by Transformer are significantly more robust than the local patterns learned by CNN for FL tasks [44] (Fig. 1.), and also provide a good representation of the features. Considering the communication cost, we freeze the entire backbone network during training and use LoRA, a Parameter-Efficient Fine-Tuning (PEFT) method to train the model. To address the catastrophic forgetting problem in FCIL, we propose incremental LoRA, which constrains different stages of LoRA to be trained on", + "bbox": [ + 212, + 311, + 504, + 686 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "subspaces orthogonal to each other via orthogonal regularization, and acquires knowledge learned in the past by simple and efficient summation at inference time. To address the classifier bias caused by data heterogeneity, we design a prototype re-weight module for the global server, which aggregates local prototypes based on heuristic information between prototypes and corresponding class features. Compared to other existing FCIL methods, our model achieves superior results on standard datasets. Furthermore, we explore the model's performance under different non-IID settings and degrees of data heterogeneity. Experiments demonstrate the robustness of our approach, while other methods suffer from significant declines.", + "bbox": [ + 212, + 689, + 787, + 840 + ], + "page_idx": 2 + }, + { + "type": "image", + "img_path": "images/83ed409fef50ff66a6ccc5e13482f17907bf577f515fdbefe7df7ab37cbb14ac.jpg", + "image_caption": [ + "Fig.1: The local data of client 1 and client 2 are non-IID. Resnet50 (the first row) focuses more on local patterns, and the patterns learned are significantly different in the case of data heterogeneity, causing the average model to lose some important information (e.g., fish fins). However, ViT (the second row) is less affected by data heterogeneity, and the averaged model basically retains all the information learned by the local model." + ], + "image_footnote": [], + "bbox": [ + 521, + 347, + 782, + 492 + ], + "page_idx": 2 + }, + { + "type": "header", + "text": "PILORA", + "bbox": [ + 671, + 114, + 730, + 126 + ], + "page_idx": 2 + }, + { + "type": "page_number", + "text": "3", + "bbox": [ + 774, + 114, + 784, + 126 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "The main contributions of this paper can be summarized as:", + "bbox": [ + 217, + 146, + 648, + 161 + ], + "page_idx": 3 + }, + { + "type": "list", + "sub_type": "text", + "list_items": [ + "- We propose incremental LoRA, which performs incremental learning on orthogonal subspaces to mitigate catastrophic forgetting and aggregates previous knowledge through simple and efficient parameter summation.", + "- We propose a prototype re-weight module to form global prototypes using heuristic information between each class of local prototypes and the corresponding class of features. Our method effectively addresses classifier bias induced by data heterogeneity without retraining.", + "- We conduct extensive experiments on standard datasets and achieve state-of-the-art performance. Furthermore, in the extreme heterogeneous case, our method still maintains robustness, while all other methods suffer plummet." + ], + "bbox": [ + 223, + 169, + 784, + 318 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "2 Related Work", + "text_level": 1, + "bbox": [ + 215, + 340, + 387, + 357 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Federated Class-Incremental Learning: Research on FCIL has garnered considerable attention in recent years. Dong et al. [10] first introduces the concept of FCIL and proposes several loss functions on the local side and the global server side to alleviate local catastrophic forgetting and global catastrophic forgetting. However, their method uses a rehearsal buffer to store and retain old class data, and additionally design a proxy server to select the best model, resulting in large memory overhead and communication costs. LGA [9] extends the work of [10], but it still belongs to rehearsal-based FCIL. In the more challenging rehearsal-free FCIL problem, generate models are widely adopted to produce synthetic data that aim to mitigate the catastrophic forgetting on local side and global side [43,62], but its performance is highly dependent on the quality of synthetic data and will incur additional computational costs. Similar to our work, FedSpace [49] designs prototype-based loss to encourage feature vectors of the same class to be close together, while we achieve this goal by using prototype classifier. In contrast to the methods described above, some recent studies [1,33] combine pre-trained models with FCIL and achieve higher performance at a smaller communication cost. However, they both adopt a similarity-based selection strategy, which causes additional memory overheads in inference. Besides, they all use supervised pre-training weights, whereas we argue that this may pose privacy concerns because data from downstream tasks may overlap with pre-trained datasets.", + "bbox": [ + 212, + 371, + 787, + 686 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "PEFT for Pre-Trained Model: With the emergence of large-scale pre-trained models [4, 22, 45], how to effectively fine-tune these models to adapt the downstream tasks has been a focal point of attention. Recently, LoRA [17], Prompt [34], and Adapter [16] have emerged as standout techniques and have been widely used in CIL [12, 50, 54, 55] and FL [15, 58, 63] tasks. In FCIL, existing methods [1, 33] have attempted to combine Prompt and Adapter with pre-trained models. Specifically, they store the knowledge learned at each stage in the parameters of the Prompt or Adapter module and select the appropriate module for the current input to be embedded into the model by a specific similarity computation during the inference time, thus effectively mitigating catastrophic", + "bbox": [ + 212, + 688, + 787, + 840 + ], + "page_idx": 3 + }, + { + "type": "page_number", + "text": "4", + "bbox": [ + 217, + 114, + 228, + 126 + ], + "page_idx": 3 + }, + { + "type": "header", + "text": "H. Guo et al.", + "bbox": [ + 271, + 114, + 359, + 127 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "forgetting with a tiny communication costs. However, such similarity matching based approaches undoubtedly introduce inference delays since they require additional similarity computation modules. (In [33], they even train a separate CNN to compute the similarity.) Furthermore, they need a additional memory space in the global server to set up a parameter pool to store the module parameters.", + "bbox": [ + 212, + 146, + 787, + 223 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "3 Preliminaries", + "text_level": 1, + "bbox": [ + 215, + 244, + 382, + 262 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "In FCIL setting, each client has a local stream dataset $D_{k} = \\{D_{k}^{t}\\}_{t=1}^{T}$ , where $D_{k}^{t} = \\{\\mathbf{X}_{k}^{t}, \\mathbf{Y}_{k}^{t}\\} = \\{x_{k,i}^{t}, y_{k,i}^{t}\\}_{i=1}^{N_{t}}$ is the dataset of the $k$ -th client on task $t$ . Dataset $D_{k}^{t}$ contains $N_{k}^{t}$ training samples and their label $\\mathbf{Y}_{k}^{t} \\in C_{k}^{t}$ , where $C_{k}^{t}$ is the class set of the $k$ -th client in task $t$ . In particular, the distribution of different clients $k$ under the same task is non-IID and the class sets of different tasks $t$ are disjoint. For local client, the objective is to minimize a pre-defined loss function $\\mathcal{L}$ on current dataset $D_{k}^{t}$ , while avoiding interference with and possibly enhancing the knowledge acquired from previous learning stages:", + "bbox": [ + 214, + 277, + 789, + 402 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\n\\underset {\\boldsymbol {\\omega} _ {k} ^ {t}} {\\operatorname {a r g m i n}} \\mathcal {L} \\left(\\boldsymbol {\\omega} _ {k} ^ {t}; \\boldsymbol {\\omega} ^ {t - 1}, \\mathbf {X} _ {k} ^ {t}, \\mathbf {Y} _ {k} ^ {t}\\right), \\tag {1}\n$$\n", + "text_format": "latex", + "bbox": [ + 392, + 412, + 785, + 441 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "where $\\omega_{k}^{t}$ is the parameters of $k$ -th local model, and $\\omega^{t-1}$ is the global model at previous task. Then, the server updates the global model $\\omega^{t}$ by aggregating all uploaded parameters as follows:", + "bbox": [ + 214, + 453, + 787, + 500 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\n\\boldsymbol {\\omega} ^ {t} = \\sum_ {k = 1} ^ {K} \\gamma_ {k} \\boldsymbol {\\omega} _ {k} ^ {t}, \\text {w h e r e} \\gamma_ {k} = \\frac {N _ {k} ^ {t}}{\\sum_ {k ^ {\\prime}} N _ {k ^ {\\prime}} ^ {t}}. \\tag {2}\n$$\n", + "text_format": "latex", + "bbox": [ + 367, + 511, + 785, + 551 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "The global model aims to correctly classify the test samples of all seen classes and solve the problem of data heterogeneity at low communication cost.", + "bbox": [ + 214, + 561, + 787, + 593 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "4 Our method", + "text_level": 1, + "bbox": [ + 215, + 616, + 372, + 633 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "4.1 Incremental LoRA for Pre-Trained Model", + "text_level": 1, + "bbox": [ + 215, + 648, + 607, + 664 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "In FCIL, storing the knowledge of different stages in different modules through PEFT can effectively mitigate the catastrophic forgetting [1,33], but additional similarity computation units are required in the inference time in order to select the appropriate module to embed into the model based on the inputs, which leads to inference delay and additional memory overhead. In order to learn an end-to-end global model without taking up extra storage space, we intuitively believe that modules storing knowledge from different stages can be organically combined into one module that embeds knowledge from all stages. Accordingly, we propose Incremental LoRA to effectively address catastrophic forgetting in FCIL. On the one hand, LoRA has the natural advantages of low inference latency and more stable training [17], on the other hand, Inspired by [53, 55], we", + "bbox": [ + 212, + 672, + 787, + 843 + ], + "page_idx": 4 + }, + { + "type": "header", + "text": "PILORA", + "bbox": [ + 671, + 114, + 730, + 126 + ], + "page_idx": 4 + }, + { + "type": "page_number", + "text": "5", + "bbox": [ + 774, + 116, + 785, + 126 + ], + "page_idx": 4 + }, + { + "type": "image", + "img_path": "images/7e038bf45d14de35b69d245ef807388ab0f5880b01e46155eb88bf8431464b4a.jpg", + "image_caption": [ + "Fig. 2: Illustration of PILoRA for FCIL. The client fine-tunes LoRA and prototypes for each class using a local dataset. Upon upload, the global server aggregates the LoRA uploaded by different clients, and applies a prototype re-weight module before re-sending them to each client." + ], + "image_footnote": [], + "bbox": [ + 232, + 148, + 767, + 382 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "introduce the orthogonality loss to constrain LoRA to learn new knowledge in a subspace orthogonal to past tasks, thus better preserving the knowledge of the old classes.", + "bbox": [ + 212, + 484, + 784, + 527 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Specifically, we define the initialization parameters of the pre-training model as $\\mathbf{W} \\in \\mathbb{R}^{d \\times k}$ , and $\\Delta \\mathbf{W}_{\\mathbf{t}}$ represents the parameter to be updated for task $t$ . Then, the model's updates in different incremental learning stages can be expressed as:", + "bbox": [ + 212, + 529, + 784, + 575 + ], + "page_idx": 5 + }, + { + "type": "equation", + "text": "\n$$\n\\mathbf {W} + \\Delta \\mathbf {W} _ {\\mathbf {t}}. \\tag {3}\n$$\n", + "text_format": "latex", + "bbox": [ + 455, + 590, + 784, + 608 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "LoRA assumes that the weight changes of large-scale pre-trained models when adapting to downstream tasks occur in a low-rank space:", + "bbox": [ + 212, + 616, + 785, + 647 + ], + "page_idx": 5 + }, + { + "type": "equation", + "text": "\n$$\n\\mathbf {W} + \\Delta \\mathbf {W} _ {\\mathbf {t}} = \\mathbf {W} + \\mathbf {A} _ {t} \\mathbf {B} _ {t}, \\tag {4}\n$$\n", + "text_format": "latex", + "bbox": [ + 405, + 659, + 785, + 676 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "where $\\mathbf{A}_t \\in \\mathbb{R}^{d \\times r}$ , $\\mathbf{B}_t \\in \\mathbb{R}^{r \\times k}$ , and $r \\ll \\min\\{d, k\\}$ . At first, $\\mathbf{A}_t$ is initialized by a random Gaussian, while $\\mathbf{B}_t$ is initialized with zero. Therefore, $\\mathbf{B}_t$ can be regarded as the coefficient matrix of $\\mathbf{A}_t$ [53]. LoRA applies it as a bypass to query and value projection matrices in multi-head attention module, and during adaptation to downstream tasks, only the parameters of $\\mathbf{A}_t$ and $\\mathbf{B}_t$ are trainable.", + "bbox": [ + 212, + 686, + 785, + 763 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Inspired by [53], we argue that the parameters of LoRA can be viewed as containers that store different subspaces of task gradients. Thus, learning a series of incremental learning tasks can be viewed as learning a series of LoRA parameters. So how to use these LoRA modules to build a model that can classify all seen classes? An idea is to select appropriate modules to be embedded in the", + "bbox": [ + 212, + 765, + 785, + 839 + ], + "page_idx": 5 + }, + { + "type": "page_number", + "text": "6", + "bbox": [ + 217, + 114, + 228, + 126 + ], + "page_idx": 5 + }, + { + "type": "header", + "text": "H. Guo et al.", + "bbox": [ + 271, + 114, + 359, + 126 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "model based on the similarity between different modules and the corresponding inputs [1, 33]. However, computing similarity requires additional computational resources, which leads to delayed inference. Another idea is that since LoRA is essentially two weight matrices, it is possible to concat all LoRA parameters prior to stage $t$ into a new matrix to gain knowledge of all previous stages [53]. In particular, at stage $t$ , the knowledge learned in the previous $t - 1$ stages are stored in $\\mathbf{A}_{1:t - 1} = [\\mathbf{A}_1,\\dots ,\\mathbf{A}_{t - 1}]$ and $\\mathbf{B}_{1:t - 1} = [\\mathbf{B}_1,\\dots ,\\mathbf{B}_{t - 1}]$ . To obtain the weight matrix, we can sequentially concat these LoRA parameters into a bigger module:", + "bbox": [ + 212, + 146, + 782, + 280 + ], + "page_idx": 6 + }, + { + "type": "equation", + "text": "\n$$\n\\mathbf {W} + \\Delta \\mathbf {W} _ {\\mathbf {t}} = \\mathbf {W} + \\widetilde {\\mathbf {A}} _ {t} \\widetilde {\\mathbf {B}} _ {t}, \\tag {5}\n$$\n", + "text_format": "latex", + "bbox": [ + 408, + 281, + 784, + 297 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "where $\\widetilde{\\mathbf{A}}_t = \\mathrm{concat}([\\mathbf{A}_1,\\ldots ,\\mathbf{A}_t])$ , $\\widetilde{\\mathbf{B}}_t = \\mathrm{concat}([\\mathbf{B}_1,\\ldots ,\\mathbf{B}_t])$ . However, this will cause the parameters of LoRA in the global model to increase with the number of incremental learning tasks. Therefore, in order to ensure consistency between the local model and the global model, we are inspired by Task Arithmetic [7, 18, 42] in model edit and propose to integrate the LoRA parameters at different stages by summation:", + "bbox": [ + 212, + 305, + 784, + 398 + ], + "page_idx": 6 + }, + { + "type": "equation", + "text": "\n$$\n\\mathbf {W} + \\Delta \\mathbf {W} _ {\\mathbf {t}} = \\mathbf {W} + \\overline {{\\mathbf {A}}} _ {t} \\overline {{\\mathbf {B}}} _ {t}, \\text {w h e r e} \\overline {{\\mathbf {A}}} _ {t} = \\sum_ {i = 1} ^ {t} \\mathbf {A} _ {i}, \\overline {{\\mathbf {B}}} _ {t} = \\sum_ {i = 1} ^ {t} \\mathbf {B} _ {i}. \\tag {6}\n$$\n", + "text_format": "latex", + "bbox": [ + 287, + 411, + 784, + 450 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Specifically, distinct directions in the weight space correspond to various localized regions in the input space [42]. Consequently, a linear combination of these diverse directions encoded within the pre-trained weights enables the model to effectively discriminate between different inputs. In our method, The training data for different incremental tasks are disjoint from each other, so it is reasonable to directly sum the LoRA parameters corresponding to different tasks. Besides, Task Arithmetic also point out that orthogonality between different task parameter vectors helps to better integrate different task. So we propose a orthogonal regularization to achieve this by constraining the parameters of LoRA to be orthogonal to previous tasks:", + "bbox": [ + 212, + 455, + 784, + 607 + ], + "page_idx": 6 + }, + { + "type": "equation", + "text": "\n$$\nl _ {o r t} \\left(\\mathbf {A} _ {i}, \\mathbf {A} _ {t}\\right) = \\sum_ {i = 1} ^ {t - 1} \\left| \\mathbf {A} _ {i} ^ {T} \\cdot \\mathbf {A} _ {t} \\right|. \\tag {7}\n$$\n", + "text_format": "latex", + "bbox": [ + 401, + 619, + 784, + 657 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Meanwhile, the orthogonal regularization encourages the learning of distinct tasks along orthogonal directions, thereby effectively reducing the spatial overlap between them. This approach will be proved to be advantageous in mitigating catastrophic forgetting, as it helps preserve the knowledge learned from previous tasks while adapting to new ones.", + "bbox": [ + 212, + 670, + 782, + 746 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "4.2 Prototype learning with Prototype Re-weight", + "text_level": 1, + "bbox": [ + 214, + 768, + 635, + 784 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Prototype Learning: As mentioned before, a feature representation that is separable intra-class and compact inter-class is helpful for FCIL tasks. Therefore, the model not only needs to correctly classify known classes but also needs", + "bbox": [ + 212, + 794, + 782, + 840 + ], + "page_idx": 6 + }, + { + "type": "header", + "text": "PILORA", + "bbox": [ + 671, + 114, + 730, + 126 + ], + "page_idx": 6 + }, + { + "type": "page_number", + "text": "7", + "bbox": [ + 774, + 114, + 785, + 126 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "to model the distribution of known classes in the feature space. In open set recognition, CPN [57] designs a discriminative loss and generative loss for prototype learning to constrain the range of known classes in the feature space, thus reserving space for samples from unknown classes. Inspired by it, we introduce prototype learning in FCIL. In particular, we set a prototype $m = \\{m_i | i = 1, 2, \\dots, C\\}$ for each class, where $m_i \\in \\mathbb{R}^d$ , and the dimensions $d$ of each prototype are the same as the dimensions of the final deep feature space.", + "bbox": [ + 212, + 146, + 787, + 252 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "To shorten the distance between the class feature and the corresponding prototype, we apply the distance-based cross entropy (DCE) discriminative loss. Given a sample $(x,y)$ , DCE uses the distance between sample features $f_{\\theta}(x)$ and prototype $m_{i}$ to represent the probability of belonging to class $i$ . Considering the normalization of probability, DCE adopts the softmax operation:", + "bbox": [ + 212, + 252, + 787, + 329 + ], + "page_idx": 7 + }, + { + "type": "equation", + "text": "\n$$\np (x \\in m _ {i} | x) = \\frac {\\exp (- \\delta \\cdot \\| f _ {\\theta} (x) - m _ {i} \\| _ {2} ^ {2})}{\\sum_ {j = 1} ^ {C} \\exp (- \\delta \\cdot \\| f _ {\\theta} (x) - m _ {j} \\| _ {2} ^ {2})}, \\tag {8}\n$$\n", + "text_format": "latex", + "bbox": [ + 334, + 335, + 785, + 375 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "where $\\| f_{\\theta}(x) - m_i\\| _2^2$ is the Euclidean distance between the feature $f_{\\theta}(x)$ of the input sample and the prototype $m_{i}$ , and $\\delta$ is the temperature scalar controlling the hardness of the class distribution. Hence, the distance-based-cross-entropy loss can be defined as:", + "bbox": [ + 214, + 380, + 787, + 441 + ], + "page_idx": 7 + }, + { + "type": "equation", + "text": "\n$$\nl _ {d c e} ((x, y); \\theta , m) = - \\log (x \\in m _ {y} | x). \\tag {9}\n$$\n", + "text_format": "latex", + "bbox": [ + 372, + 450, + 785, + 468 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "By minimizing the DCE loss, the distance between the sample's feature and the correct prototype will be smaller than other incorrect prototypes. However, features learned only under the discriminative loss may not be compact, which may lead to an overlap in feature representations between new and old classes. To solve this problem, we introduce prototype learning (PL) loss [57]:", + "bbox": [ + 214, + 474, + 787, + 551 + ], + "page_idx": 7 + }, + { + "type": "equation", + "text": "\n$$\nl _ {p l} ((x, y); \\theta , m) = \\| f _ {\\theta} (x) - m _ {y} \\| _ {2} ^ {2}, \\tag {10}\n$$\n", + "text_format": "latex", + "bbox": [ + 383, + 558, + 785, + 575 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "PL loss decreases the distance between sample features and the corresponding correct prototype, making the model learn a more compact intra-class distribution. Essentially, PL loss is a maximum likelihood regularization of features $f_{\\theta}(x)$ under the Gaussian mixture density assumption [31, 32].", + "bbox": [ + 214, + 583, + 787, + 643 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Prototype Re-weight: Considering a specific class $c$ , under the non-IID setting, each client $k$ holds a portion of the training data $N_{c,k}$ and $\\sum_{k=1}^{K} N_{c,k} = N_{c}$ , where $N_{c}$ is the sum of training samples of class $c$ . In prototype learning, the prototype $m_{k,c}$ learned by each client can effectively reflect the distribution of class $c$ within its local data. That is, for clients that do not have training samples of class $c$ , the distance between their corresponding prototype and the features of class $c$ will be greater than for clients that do have training samples. If we give equal weight to these uploaded prototypes during parameter aggregation, classifier drift may occur, especially when the data heterogeneity is high, potentially resulting in the global model losing discriminative capability for that class.", + "bbox": [ + 212, + 643, + 787, + 794 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Therefore, we use the heuristic information contained in the distance between the prototype and the corresponding class average features to design the prototype re-weight module. Specifically, at stage $t$ , each client $k$ uploads prototypes", + "bbox": [ + 212, + 795, + 787, + 840 + ], + "page_idx": 7 + }, + { + "type": "page_number", + "text": "8", + "bbox": [ + 217, + 114, + 228, + 126 + ], + "page_idx": 7 + }, + { + "type": "header", + "text": "H. Guo et al.", + "bbox": [ + 271, + 114, + 359, + 127 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Algorithm 1 Prototype Re-weight of Task t.", + "bbox": [ + 217, + 146, + 545, + 161 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "Input: the number of classes $C_t$ , prototypes of each client $m_{t,k}$ , the set of mean features $\\mu_{t,k}$ of each client and the temperature coefficient $\\eta$", + "bbox": [ + 215, + 164, + 785, + 193 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "Output: global prototype $m_t$", + "bbox": [ + 215, + 193, + 419, + 205 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "text", + "list_items": [ + "1: for $c = 1 \\rightarrow C_t$ do", + "2: if $\\exists \\mu_{t,i,c} \\neq 0$ , where $i = 1 \\rightarrow K$ then", + "3: $\\mu_{t,c}^{*}\\gets \\mu_{t,c}$ retain all non-zero mean features", + "4: $d_{t,k,c} \\gets \\text{Distance} \\left( m_{t,k}, \\mu_{t,c}^* \\right)$ via Eq. (11)", + "5: $p_{t,k,c} \\gets$ take the reciprocal of $d_{t,k,c}$", + "6: $\\alpha_{t,k,c} \\gets \\text{Max-minNorm}(p_{t,k,c})$ via Eq. (12)", + "7: $\\omega_{t,k,c} \\gets \\text{Softmax}(\\alpha_{t,k,c}, \\eta)$ via Eq. (13)", + "8: else", + "9: $\\omega_{t,k,c} \\gets$ Assign with average value $1 / C_t$", + "10: end if", + "11: $\\omega_{t,c}\\gets \\mathrm{List}(\\omega_{t,1,c},\\dots ,\\omega_{t,K,c})$", + "12: $m_{t,c}\\gets Re\\text{-weight} (\\omega_{t,c},m_{t,k})$ via Eq. (14)", + "13: end for", + "14: Global prototype $m_t \\gets$ Concat all $m_{t,c}$" + ], + "bbox": [ + 217, + 207, + 588, + 400 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "$m_{t,k}$ and the set of average features for the class they have learned $\\mu_{t,k}$ (zero for classes without samples). On the global server, we first compute the sum of distances between the prototype $m_{t,k,c}$ and the average feature $\\mu_{t,i,c}$ uploaded by all clients:", + "bbox": [ + 212, + 431, + 785, + 492 + ], + "page_idx": 8 + }, + { + "type": "equation", + "text": "\n$$\nd _ {t, k, c} = \\sum_ {i = 1} ^ {K} \\| m _ {t, k, c} - \\mu_ {t, i, c} \\| _ {2} ^ {2}, \\tag {11}\n$$\n", + "text_format": "latex", + "bbox": [ + 393, + 492, + 785, + 532 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "the value of $d_{t,k,c}$ approximates the distance from the prototype $m_{t,k,c}$ to the overall features of class $c$ . A smaller value indicates that the prototype is closer to the features of that class uploaded by all clients and during aggregation, we aim to assign it a higher weight. So we perform max-min normalization on the set $D_{t,c} = \\{p_{t,k,c}|k = 1,2,\\dots ,K\\}$ , where $p_{t,k,c} = \\frac{1}{d_{t,k,c}}$ :", + "bbox": [ + 212, + 536, + 787, + 614 + ], + "page_idx": 8 + }, + { + "type": "equation", + "text": "\n$$\n\\alpha_ {t, k, c} = \\frac {p _ {t , k , c} - \\operatorname* {m i n} D _ {t , c}}{\\operatorname* {m a x} D _ {t , c} - \\operatorname* {m i n} D _ {t , c}}. \\tag {12}\n$$\n", + "text_format": "latex", + "bbox": [ + 393, + 623, + 785, + 656 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "To meet the normalization requirements of the weights, we perform softmax processing on them and obtain the weight coefficient:", + "bbox": [ + 212, + 662, + 785, + 693 + ], + "page_idx": 8 + }, + { + "type": "equation", + "text": "\n$$\n\\omega_ {t, k, c} = \\frac {\\exp \\left(\\eta \\cdot \\alpha_ {t , k , c}\\right)}{\\sum_ {i = 1} ^ {K} \\exp \\left(\\eta \\cdot \\alpha_ {t , i , c}\\right)}, \\tag {13}\n$$\n", + "text_format": "latex", + "bbox": [ + 395, + 700, + 785, + 738 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "where $\\eta$ is the temperature coefficient that controls the softness and hardness of weights. Finally, we re-weight all local prototypes corresponding to class $c$ according to the obtained weights to get the global prototype of class $c$ :", + "bbox": [ + 212, + 748, + 785, + 794 + ], + "page_idx": 8 + }, + { + "type": "equation", + "text": "\n$$\nm _ {t, c} = \\sum_ {i = 1} ^ {K} \\omega_ {t, i, c} \\cdot m _ {t, i, c}. \\tag {14}\n$$\n", + "text_format": "latex", + "bbox": [ + 413, + 801, + 785, + 844 + ], + "page_idx": 8 + }, + { + "type": "header", + "text": "PILORA", + "bbox": [ + 671, + 114, + 730, + 126 + ], + "page_idx": 8 + }, + { + "type": "page_number", + "text": "9", + "bbox": [ + 774, + 114, + 785, + 126 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "We argue that prototype re-weight module allows the global prototypes $m_{t,c}$ to differentially consider the data distribution information inherent in each local prototype, effectively mitigating the classifier bias issue. Compared with FedProto [51], our method further compresses the representation region of the same class in the feature space, which contributes to mitigate catastrophic forgetting. More importantly, our method does not require additional retraining of the classification layer, resulting in significant advantages in computational costs. Algorithm 1 presents the pseudo code of prototype re-weight module.", + "bbox": [ + 212, + 146, + 787, + 268 + ], + "page_idx": 9 + }, + { + "type": "text", + "text": "4.3 Integrated Objective of PILoRA", + "text_level": 1, + "bbox": [ + 214, + 290, + 532, + 306 + ], + "page_idx": 9 + }, + { + "type": "text", + "text": "The loss function of PILoRA can be defined as:", + "bbox": [ + 214, + 315, + 560, + 330 + ], + "page_idx": 9 + }, + { + "type": "equation", + "text": "\n$$\nl _ {t o t a l} = l _ {d c e} + \\lambda \\cdot l _ {p l} + \\gamma \\cdot l _ {o r t}, \\tag {15}\n$$\n", + "text_format": "latex", + "bbox": [ + 398, + 344, + 785, + 359 + ], + "page_idx": 9 + }, + { + "type": "text", + "text": "where $\\lambda$ and $\\gamma$ are two hyper-parameters, and the overall framework of PILoRA is shown in Fig. 2. In general, our method is simple and effective, showing strong performance in different non-IID settings and degrees of data heterogeneity.", + "bbox": [ + 212, + 371, + 787, + 417 + ], + "page_idx": 9 + }, + { + "type": "text", + "text": "5 Experiments", + "text_level": 1, + "bbox": [ + 214, + 440, + 375, + 458 + ], + "page_idx": 9 + }, + { + "type": "text", + "text": "5.1 Experimental Setups", + "text_level": 1, + "bbox": [ + 214, + 472, + 434, + 488 + ], + "page_idx": 9 + }, + { + "type": "text", + "text": "Benchmark: To evaluate the proposed PILoRA, we perform our experiments on two well-known datasets: CIFAR-100 [24] and TinyImageNet [25], We also test the performance of the model on large-scale datasets, specifically, we randomly select 200 classes from the ImageNet-1k [8] as a new dataset. Following the protocols proposed in [33], we split 10 incremental stages and only the data from the current stage is available. Besides, to challenge our method, the local dataset of each client is followed by two kinds of non-IID settings: quantity-based label imbalance and distribution-based label imbalance [27] and we denote the degree of heterogeneity of the two settings by $\\alpha$ and $\\beta$ . Details of both settings can be seen in Section A in the Supplementary Material.", + "bbox": [ + 212, + 497, + 787, + 648 + ], + "page_idx": 9 + }, + { + "type": "text", + "text": "Comparison: We compare our method with existing FCIL methods4: TARGET [61], GLFC [10], LGA [9]. we also adopt several CIL methods: EWC [23], LwF [29], iCaRL [46], L2P [55] and FL method FedNCM [26] in FCIL setting. In addition, we compare the use of cross-entropy loss for optimization during training and the use of class means as classifiers during inference, which we named FedCLM. We explore their performance in different non-IID settings and degrees of data heterogeneity. For a fair comparison, we tune all methods to the same pre-trained model as ours and fine-tune them using LoRA.", + "bbox": [ + 212, + 648, + 787, + 768 + ], + "page_idx": 9 + }, + { + "type": "text", + "text": "Implementation: Considering privacy issues, we evaluate the performance of our method on self-supervised pre-trained weights (Dino [5]) for ViT-B/16 [11], this setting is also widely used in CIL task [36,37,60]. Considering the trade-off", + "bbox": [ + 212, + 770, + 787, + 816 + ], + "page_idx": 9 + }, + { + "type": "page_number", + "text": "10", + "bbox": [ + 217, + 114, + 235, + 126 + ], + "page_idx": 9 + }, + { + "type": "header", + "text": "H. Guo et al.", + "bbox": [ + 271, + 114, + 359, + 127 + ], + "page_idx": 9 + }, + { + "type": "page_footnote", + "text": "4 GLFC and LGA only perform experiments on quantity-based label imbalance.", + "bbox": [ + 217, + 824, + 751, + 840 + ], + "page_idx": 9 + }, + { + "type": "table", + "img_path": "images/e402a7b0ebb20fce09c17ba1dc4be8770dd38242ab2d48b9672be3014c56c1da.jpg", + "table_caption": [ + "Table 1: Results (%) on CIFAR-100. Results are included for 10 tasks (10 classes per task) and under different degree of two non-IID setting." + ], + "table_footnote": [], + "table_body": "
Non-IIDQuantity-based label imbalanceDistribution-based label imbalance
Partitionα = 6α = 4α = 2β = 0.5β = 0.1β = 0.05
MethodsANAvg.ANAvg.ANAvg.ANAvg.ANAvg.ANAvg.
Joint88.6-84.3-79.8-90.1-87.8-85.9-
EWC+FL57.969.155.966.842.252.765.577.857.873.243.559.2
LwF+FL57.468.855.166.740.852.964.777.554.663.345.764.5
iCaRL+FL35.856.537.158.943.455.351.367.750.165.944.663.0
L2P+FL63.465.159.058.22.65.653.951.662.971.438.732.2
FedCLM58.969.457.667.644.357.966.577.461.071.848.863.5
FedNCM65.674.461.971.149.659.866.877.962.172.450.965.9
TARGET60.971.358.869.545.256.566.177.860.571.151.865.3
GLFC58.270.453.765.913.137.7------
LGA64.573.661.170.521.640.9------
Ours69.578.665.174.454.962.668.578.163.473.754.867.1
Ours+HT69.678.565.874.856.564.370.278.663.673.857.969.2
", + "bbox": [ + 233, + 184, + 772, + 386 + ], + "page_idx": 10 + }, + { + "type": "text", + "text": "between performance and number of parameters, we insert the LoRA module into only the first block (See Section B.1) and set $r = 4$ . We train our models using Adam [21] with a batch size of 64 and followed by [60], we adopt different learning rates for prototypes layer of $2e^{-3}$ and LoRA parameters of $1e^{-5}$ on CIFAR-100, and $5e^{-3} / 5e^{-6}$ on TinyImageNet. Moreover, cosine annealing is also used in training processes. We set $\\delta = 1$ , $\\lambda = 0.001$ , $\\gamma = 0.5$ and $\\eta = 0.2$ . We initialize 10 local clients to train and upload the parameters at each communication round. The local training epoch is 5 and the communication round is 30.", + "bbox": [ + 212, + 412, + 787, + 537 + ], + "page_idx": 10 + }, + { + "type": "text", + "text": "5.2 Comparative Results", + "text_level": 1, + "bbox": [ + 215, + 556, + 434, + 573 + ], + "page_idx": 10 + }, + { + "type": "text", + "text": "We report $A_{N}(\\uparrow)$ and $\\mathrm{Avg}(\\uparrow)$ to evaluate the performance of the methods, where $A_{N}$ is the accuracy of all seen classes in the final task and Avg. is calculated as the average accuracy of all tasks. Results are shown in Tab. 1 and Tab. 2, we can obverse that our method outperforms other comparison methods and demonstrates strong robustness across different non-IID settings. Among FCIL methods, TARGET achieves relatively better performance under different data heterogeneity, while LGA and GLFC show a steep drop in performance and lack of robustness under extreme data heterogeneity. Notably, FedNCM in the FCIL setting even outperforms TARGET in some cases, which suggests that existing FCIL methods lack the handling of data heterogeneity, whereas our approach dramatically improves the performance of the model under different data heterogeneity through a lightweight prototype re-weight module.", + "bbox": [ + 212, + 583, + 787, + 763 + ], + "page_idx": 10 + }, + { + "type": "text", + "text": "The CIL methods also suffer from data heterogeneity in the FCIL setting, with L2P+FL being the most obvious. We believe that there are two main reasons for this, firstly the performance of L2P relies on supervised pre-trained weights (e.g., ImageNet-1k [8]) and its performance degrades significantly when using self-supervised pre-trained weights [60]. Secondly, L2P needs to choose the", + "bbox": [ + 212, + 765, + 787, + 840 + ], + "page_idx": 10 + }, + { + "type": "header", + "text": "PILORA", + "bbox": [ + 671, + 114, + 730, + 126 + ], + "page_idx": 10 + }, + { + "type": "page_number", + "text": "11", + "bbox": [ + 767, + 114, + 784, + 126 + ], + "page_idx": 10 + }, + { + "type": "table", + "img_path": "images/115cf2bab88c315fa029c5f4c196ce03bec8502f5c98f8229d93f98170441386.jpg", + "table_caption": [ + "Table 2: Results (\\%) on TinyImageNet. Results are included for 10 tasks (20 classes per task) and under different degree of two non-IID setting." + ], + "table_footnote": [], + "table_body": "
Non-IIDQuantity-based label imbalanceDistribution-based label imbalance
Partitionα = 12α = 8α = 4β = 0.5β = 0.1β = 0.05
MethodsANAvg.ANAvg.ANAvg.ANAvg.ANAvg.ANAvg.
Joint83.6-82.9-80.2-84.3-83.3-82.8-
L2P+FL61.658.049.439.38.210.264.266.956.352.551.943.2
FedCLM61.672.451.860.345.856.966.577.460.471.046.757.8
FedNCM71.681.669.579.457.264.773.781.670.880.468.478.0
TARGET72.681.670.379.663.873.571.680.971.080.169.379.1
GLFC69.177.961.373.525.139.4------
LGA71.379.465.875.336.748.8------
Ours74.481.074.380.970.177.874.580.974.381.073.680.2
Ours+HT75.081.374.981.267.975.674.781.074.380.973.480.1
", + "bbox": [ + 228, + 184, + 777, + 345 + ], + "page_idx": 11 + }, + { + "type": "image", + "img_path": "images/6c9ab0bf0ffc2e44905efff2593d0aeb5534d72e962d2d5bd3c8839ecf6c8885.jpg", + "image_caption": [ + "(a)" + ], + "image_footnote": [], + "bbox": [ + 228, + 364, + 380, + 460 + ], + "page_idx": 11 + }, + { + "type": "image", + "img_path": "images/ccf7e08d3b18cfc5db266ea983653c1b9f6c324715f101ce31d773aa9c456466.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 383, + 364, + 537, + 460 + ], + "page_idx": 11 + }, + { + "type": "image", + "img_path": "images/50c43133f5895e13f792703c80e5366ab30e0bb8b19af7e025358fffc3f14dcc.jpg", + "image_caption": [ + "(b)", + "Fig. 3: (a) Comparison of real proportions of the same class among different clients and the weights obtained from Prototype Re-weighting during a local training process. (b) The $L2$ distance between classes feature extracted by different clients and global model with corresponding prototypes." + ], + "image_footnote": [], + "bbox": [ + 565, + 369, + 774, + 462 + ], + "page_idx": 11 + }, + { + "type": "text", + "text": "appropriate prompt for embedding in the model based on the similarity calculation. However, in cases of extreme heterogeneity, there exists a significant disparity in what each client learns, even at the same stage. This discrepancy prevents the selection of a uniform prompt for reasoning. Instead, our method bypasses redundant similarity calculations or knowledge distillation [29] and effectively mitigates catastrophic forgetting directly through the summation of LoRA parameters. In addition, we combine the HeadTune in FedNCM with our prototype classifier, and the performance of our model is further improved by the better initialization provided by HeadTune. This suggests that there is still a lot of space for improvement in the existing methods in solving the problem of data heterogeneity in FCIL.", + "bbox": [ + 212, + 580, + 787, + 747 + ], + "page_idx": 11 + }, + { + "type": "text", + "text": "Analysis of Prototype Re-weight: The core idea of our proposed prototype re-weight module is to aggregate prototypes in a manner that closely resembles the true data distribution among clients without disclosing local data information. In Fig. 3a, we show the real proportions of the same class of data across clients and we can observe that the weights calculated through re-weight module align with the distribution of data among different clients. Hence, the global", + "bbox": [ + 212, + 750, + 787, + 840 + ], + "page_idx": 11 + }, + { + "type": "page_number", + "text": "12", + "bbox": [ + 217, + 114, + 235, + 126 + ], + "page_idx": 11 + }, + { + "type": "header", + "text": "H. Guo et al.", + "bbox": [ + 271, + 114, + 359, + 126 + ], + "page_idx": 11 + }, + { + "type": "table", + "img_path": "images/503cb48cbe8f52b4b2c62e3e5b8cb585a7364c393e0eae1bafddacb4228e2d4b.jpg", + "table_caption": [ + "Table 3: Ablation Results (\\%) on CIFAR-100. Experimental results based on self-supervised pre-trained weights. $F_N(\\downarrow)$ is the average forgetting and the calculation process can be referred to [66]." + ], + "table_footnote": [], + "table_body": "
Partitionα = 6β = 0.5
MethodsANAvg.FNANAvg.FN
Ours-w/o PR26.830.95.541.448.89.3
Ours-w/o lort68.777.910.367.177.011.1
Ours-w/o LoRA (fintune)11.329.57.313.329.98.9
Ours-w/o LoRA (frozen)68.677.710.067.877.211.6
Ours69.578.69.568.578.110.6
", + "bbox": [ + 318, + 198, + 689, + 296 + ], + "page_idx": 12 + }, + { + "type": "text", + "text": "prototypes effectively retain all learned prototype information. In contrast, other methods typically use direct averaging for the classification layer, leading to the fusion of a considerable amount of irrelevant information in the classifier when the data heterogeneity is high, consequently causing classifier bias. In Fig. 3b, we compare the distances between the prototype calculated by both prototype re-weight and average for the global model and the deep features of the test data. It can be seen that our proposed prototype re-weight method effectively 'takes the best' of all client uploaded prototypes, so that the global prototype is better adapted to the features of the corresponding class, whereas on average the correct prototype is far away from the features of the corresponding class, especially when the data is very heterogeneous.", + "bbox": [ + 212, + 324, + 787, + 491 + ], + "page_idx": 12 + }, + { + "type": "text", + "text": "5.3 Ablation Study", + "text_level": 1, + "bbox": [ + 214, + 512, + 388, + 527 + ], + "page_idx": 12 + }, + { + "type": "text", + "text": "To evaluate the effect of each component in PILoRA, we perform the ablation study and show the results of 10 phases set in CIFAR-100 in Tab. 3. We can observe that without prototype re-weight (PR), The classification accuracy $A_{N}$ and Avg. of the model decrease significantly, which is due to the fact that the global prototype obtained by simply averaging at this point does not represent the information of each class well, which leads to classifier bias. Instead, our proposed method significantly improves the performance of the model by heuristically reweight the local prototypes. When there is no orthogonal regularization $(l_{ort})$ , all the metrics of the model show a certain decrease, which indicates that there is a partial overlap between the parameter spaces of different incremental tasks. By imposing orthogonal regularization the parameter spaces corresponding to different input spaces can be made to train the model in the direction of orthogonality to each other, which further improves the model performance. We visualize the cosine similarity between LoRA in Section B.2.", + "bbox": [ + 212, + 537, + 787, + 748 + ], + "page_idx": 12 + }, + { + "type": "text", + "text": "To better demonstrate the contribution of LoRA, we compare the effects of fine-tuning the entire backbone and freezing the entire backbone under self-supervised pre-trained weights when $\\alpha = 6$ and $\\beta = 0.5$ . As can be seen in Tab. 3, when fine-tuning the entire backbone, the performance of the model is severely degraded, we believe that this is due to the fact that the model parameters of ViT are too large to achieve a balance between the ability to remember old", + "bbox": [ + 212, + 750, + 787, + 840 + ], + "page_idx": 12 + }, + { + "type": "header", + "text": "PILORA", + "bbox": [ + 671, + 114, + 730, + 126 + ], + "page_idx": 12 + }, + { + "type": "page_number", + "text": "13", + "bbox": [ + 767, + 114, + 785, + 126 + ], + "page_idx": 12 + }, + { + "type": "text", + "text": "classes and the ability to discriminate new classes under the constraint of knowledge distillation. Freezing the entire backbone network can retain discriminative ability for old classes, but at this point, the model has too few trainable parameters, with only prototypes used for classification, thereby limiting the expressive capacity of the model. Therefore, to achieve a trade-off between communication cost and model performance, we fine-tune and achieve the best performance with a tiny number of parameters through Incremental LoRA.", + "bbox": [ + 212, + 146, + 787, + 252 + ], + "page_idx": 13 + }, + { + "type": "text", + "text": "5.4 Further Analysis", + "text_level": 1, + "bbox": [ + 215, + 271, + 401, + 286 + ], + "page_idx": 13 + }, + { + "type": "text", + "text": "Memory usage analysis. For our PILoRA, in addition to the model for each client, we store the LoRA parameters $A_{q}^{1:t - 1}$ and $A_{v}^{1:t - 1}$ up to the current stage $t$ to compute the orthogonal regularization. Whereas TARGET requires an additional generator to generate old samples, as well as additional memory space to store the generated images, LGA and GLFC similarly take up additional space to store old class samples. Compared to them, our method takes up very little memory space, storing on average only $0.04\\%$ of the parameters equivalent to the ViT-base.", + "bbox": [ + 212, + 292, + 486, + 531 + ], + "page_idx": 13 + }, + { + "type": "image", + "img_path": "images/eba9aef916c37020ca0159111514aacd6192e007e387183f7e255efe8119c05e.jpg", + "image_caption": [ + "Fig.4: The impact of different number of $K$ on CIFAR-100, where we consider $\\alpha = 6$ and $\\beta = 0.5$ ." + ], + "image_footnote": [], + "bbox": [ + 498, + 325, + 638, + 431 + ], + "page_idx": 13 + }, + { + "type": "image", + "img_path": "images/ae690ecd148e5ca54d747d9d26469f771e9b33fc02aa59931e30013d2a665802.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 643, + 325, + 782, + 431 + ], + "page_idx": 13 + }, + { + "type": "text", + "text": "Increase the number of local clients $(K)$ . As depicted in Fig. 4, we investigate the performance of the model as the number of local clients increase, by respectively setting $K = \\{10, 15, 20\\}$ . From the results shown in Fig. 4, we observe that the model's performance slightly declines as the number of clients increases. We believe that under the same non-IID setting, enlarging the number of clients further exacerbates the heterogeneity among clients, consequently impacting the model's performance. More results can be seen in Section B.3.", + "bbox": [ + 212, + 534, + 787, + 642 + ], + "page_idx": 13 + }, + { + "type": "text", + "text": "6 Conclusion", + "text_level": 1, + "bbox": [ + 215, + 660, + 359, + 676 + ], + "page_idx": 13 + }, + { + "type": "text", + "text": "In this paper, we propose a simple and effective method of $PILoRA$ for FCIL. PILoRA is based on pre-trained ViT models and fine-tunes with a tiny number of parameters using LoRA. To address the catastrophic forgetting problem in FCIL, we propose incremental LoRA, which can efficiently combine different incremental tasks by summing the orthogonal LoRA parameter space; To deal with the classifier bias caused by data heterogeneity, we adopt prototype learning and propose prototype re-weight, which utilize heuristic information between prototypes and features to perform weighted aggregation of global prototypes. Experimental results show that our method achieves state-of-the-art results on standard datasets and maintains robustness under extreme data heterogeneity.", + "bbox": [ + 212, + 688, + 787, + 840 + ], + "page_idx": 13 + }, + { + "type": "page_number", + "text": "14", + "bbox": [ + 217, + 114, + 235, + 126 + ], + "page_idx": 13 + }, + { + "type": "header", + "text": "H. Guo et al.", + "bbox": [ + 271, + 114, + 359, + 127 + ], + "page_idx": 13 + }, + { + "type": "text", + "text": "Acknowledgements", + "text_level": 1, + "bbox": [ + 217, + 143, + 401, + 162 + ], + "page_idx": 14 + }, + { + "type": "text", + "text": "This work has been supported by the National Science and Technology Major Project (2022ZD0116500), National Natural Science Foundation of China (U20A20223, 62222609, 62076236), CAS Project for Young Scientists in Basic Research (YSBR-083), and Key Research Program of Frontier Sciences of CAS (ZDBS-LY-7004) and the InnoHK program.", + "bbox": [ + 212, + 176, + 787, + 253 + ], + "page_idx": 14 + }, + { + "type": "text", + "text": "References", + "text_level": 1, + "bbox": [ + 215, + 277, + 321, + 292 + ], + "page_idx": 14 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "1. Bagwe, G., Yuan, X., Pan, M., Zhang, L.: Fed-cprompt: Contrastive prompt for rehearsal-free federated continual learning. In: Federated Learning and Analytics in Practice: Algorithms, Systems, Applications, and Opportunities (2023)", + "2. Belouadah, E., Popescu, A.: Deesil: Deep-shallow incremental learning. In: Proceedings of the European Conference on Computer Vision (ECCV) Workshops. pp. 0-0 (2018)", + "3. Belouadah, E., Popescu, A.: II2m: Class incremental learning with dual memory. In: Proceedings of the IEEE/CVF international conference on computer vision. pp. 583-592 (2019)", + "4. Brown, T., Mann, B., Ryder, N., Subbiah, M., Kaplan, J.D., Dhariwal, P., Neelakantan, A., Shyam, P., Sastry, G., Askell, A., et al.: Language models are few-shot learners. Advances in neural information processing systems 33, 1877-1901 (2020)", + "5. Caron, M., Touvron, H., Misra, I., Jégou, H., Mairal, J., Bojanowski, P., Joulin, A.: Emerging properties in self-supervised vision transformers. In: Proceedings of the International Conference on Computer Vision (ICCV) (2021)", + "6. Cheng, Z., Zhang, X.Y., Liu, C.L.: Unified classification and rejection: A one-versus-all framework. arXiv preprint arXiv:2311.13355 (2023)", + "7. Chitale, R., Vaidya, A., Kane, A., Ghotkar, A.: Task arithmetic with lora for continual learning. arXiv preprint arXiv:2311.02428 (2023)", + "8. Deng, J., Dong, W., Socher, R., Li, L.J., Li, K., Fei-Fei, L.: Imagenet: A large-scale hierarchical image database. In: 2009 IEEE conference on computer vision and pattern recognition. pp. 248-255. IEEE (2009)", + "9. Dong, J., Cong, Y., Sun, G., Zhang, Y., Schiele, B., Dai, D.: No one left behind: Real-world federated class-incremental learning. arXiv preprint arXiv:2302.00903 (2023)", + "0. Dong, J., Wang, L., Fang, Z., Sun, G., Xu, S., Wang, X., Zhu, Q.: Federated class incremental learning. In: Proceedings of the IEEE/CVF conference on computer vision and pattern recognition. pp. 10164-10173 (2022)", + "11. Dosovitskiy, A., Beyer, L., Kolesnikov, A., Weissenborn, D., Zhai, X., Unterthiner, T., Dehghani, M., Minderer, M., Heigold, G., Gelly, S., et al.: An image is worth 16x16 words: Transformers for image recognition at scale. arXiv preprint arXiv:2010.11929 (2020)", + "12. Gao, Q., Zhao, C., Sun, Y., Xi, T., Zhang, G., Ghanem, B., Zhang, J.: A unified continual learning framework with general parameter-efficient tuning. arXiv preprint arXiv:2303.10070 (2023)", + "13. Geng, C., Huang, S.j., Chen, S.: Recent advances in open set recognition: A survey. IEEE transactions on pattern analysis and machine intelligence 43(10), 3614-3631 (2020)" + ], + "bbox": [ + 218, + 310, + 784, + 839 + ], + "page_idx": 14 + }, + { + "type": "header", + "text": "PILORA", + "bbox": [ + 671, + 114, + 730, + 126 + ], + "page_idx": 14 + }, + { + "type": "page_number", + "text": "15", + "bbox": [ + 767, + 114, + 785, + 126 + ], + "page_idx": 14 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "14. Giuseppi, A., Manfredi, S., Pietrabissa, A.: A weighted average consensus approach for decentralized federated learning. Machine Intelligence Research 19(4), 319-330 (2022). https://doi.org/10.1007/s11633-022-1338-z, https://www.mi-research.net/en/article/doi/10.1007/s11633-022-1338-z", + "15. Guo, T., Guo, S., Wang, J., Tang, X., Xu, W.: Promptfl: Let federated participants cooperatively learn prompts instead of models-federated learning in age of foundation model. IEEE Transactions on Mobile Computing (2023)", + "16. Houlsby, N., Giurgiu, A., Jastrzebski, S., Morrone, B., De Laroussilhe, Q., Gesmundo, A., Attariyan, M., Gelly, S.: Parameter-efficient transfer learning for nlp. In: International Conference on Machine Learning. pp. 2790-2799. PMLR (2019)", + "17. Hu, E.J., Shen, Y., Wallis, P., Allen-Zhu, Z., Li, Y., Wang, S., Wang, L., Chen, W.: Lora: Low-rank adaptation of large language models. arXiv preprint arXiv:2106.09685 (2021)", + "18. Ilharco, G., Ribeiro, M.T., Wortsman, M., Gururangan, S., Schmidt, L., Hajishirzi, H., Farhadi, A.: Editing models with task arithmetic. arXiv preprint arXiv:2212.04089 (2022)", + "19. Karimireddy, S.P., Kale, S., Mohri, M., Reddi, S., Stich, S., Suresh, A.T.: Scaffold: Stochastic controlled averaging for federated learning. In: International conference on machine learning. pp. 5132-5143. PMLR (2020)", + "20. Khan, L.U., Saad, W., Han, Z., Hossain, E., Hong, C.S.: Federated learning for internet of things: Recent advances, taxonomy, and open challenges. IEEE Communications Surveys & Tutorials 23(3), 1759-1799 (2021)", + "21. Kingma, D.P., Ba, J.: Adam: A method for stochastic optimization. arXiv preprint arXiv:1412.6980 (2014)", + "22. Kirillov, A., Mintun, E., Ravi, N., Mao, H., Rolland, C., Gustafson, L., Xiao, T., Whitehead, S., Berg, A.C., Lo, W.Y., et al.: Segment anything. arXiv preprint arXiv:2304.02643 (2023)", + "23. Kirkpatrick, J., Pascanu, R., Rabinowitz, N., Veness, J., Desjardins, G., Rusu, A.A., Milan, K., Quan, J., Ramalho, T., Grabska-Barwinska, A., et al.: Overcoming catastrophic forgetting in neural networks. Proceedings of the national academy of sciences 114(13), 3521-3526 (2017)", + "24. Krizhevsky, A., Hinton, G., et al.: Learning multiple layers of features from tiny images (2009)", + "25. Le, Y., Yang, X.: Tiny imagenet visual recognition challenge. CS 231N 7(7), 3 (2015)", + "26. Legate, G., Bernier, N., Page-Caccia, L., Oyallon, E., Belilovsky, E.: Guiding the last layer in federated learning with pre-trained models. Advances in Neural Information Processing Systems 36 (2024)", + "27. Li, Q., Diao, Y., Chen, Q., He, B.: Federated learning on non-iid data silos: An experimental study. In: 2022 IEEE 38th International Conference on Data Engineering (ICDE). pp. 965-978. IEEE (2022)", + "28. Li, T., Sahu, A.K., Zaheer, M., Sanjabi, M., Talwalkar, A., Smith, V.: Federated optimization in heterogeneous networks. Proceedings of Machine learning and systems 2, 429-450 (2020)", + "29. Li, Z., Hoiem, D.: Learning without forgetting. IEEE transactions on pattern analysis and machine intelligence 40(12), 2935-2947 (2017)", + "30. Lin, Y.M., Gao, Y., Gong, M.G., Zhang, S.J., Zhang, Y.Q., Li, Z.Y.: Federated learning on multimodal data: A comprehensive survey. Machine Intelligence Research 20(4), 539-553 (2023). https://doi.org/10.1007/s11633-022-1398-0, https://www.mi-research.net/en/article/doi/10.1007/s11633-022-1398-0" + ], + "bbox": [ + 215, + 147, + 785, + 839 + ], + "page_idx": 15 + }, + { + "type": "page_number", + "text": "16", + "bbox": [ + 217, + 114, + 235, + 126 + ], + "page_idx": 15 + }, + { + "type": "header", + "text": "H. Guo et al.", + "bbox": [ + 271, + 114, + 359, + 126 + ], + "page_idx": 15 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "31. Liu, C.L., Sako, H., Fujisawa, H.: Discriminative learning quadratic discriminant function for handwriting recognition. IEEE Transactions on Neural Networks 15(2), 430-444 (2004)", + "32. Liu, C.L., Sako, H., Fujisawa, H.: Effects of classifier structures and training regimes on integrated segmentation and recognition of handwritten numeral strings. IEEE transactions on pattern analysis and machine intelligence 26(11), 1395-1407 (2004)", + "33. Liu, C., Qu, X., Wang, J., Xiao, J.: Fedet: A communication-efficient federated class-incremental learning framework based on enhanced transformer. arXiv preprint arXiv:2306.15347 (2023)", + "34. Liu, P., Yuan, W., Fu, J., Jiang, Z., Hayashi, H., Neubig, G.: Pre-train, prompt, and predict: A systematic survey of prompting methods in natural language processing. ACM Computing Surveys 55(9), 1-35 (2023)", + "35. Liu, W., Wu, X., Zhu, F., Yu, M., Wang, C., Liu, C.L.: Class incremental learning with self-supervised pre-training and prototype learning. arXiv preprint arXiv:2308.02346 (2023)", + "36. Liu, W., Zhu, F., Liu, C.L.: Branch-tuning: Balancing stability and plasticity for continual self-supervised learning. arXiv preprint arXiv:2403.18266 (2024)", + "37. Liu, W., Zhu, F., Liu, C.L.: Towards non-exemplar semi-supervised class-incremental learning. arXiv preprint arXiv:2403.18291 (2024)", + "38. Luo, M., Chen, F., Hu, D., Zhang, Y., Liang, J., Feng, J.: No fear of heterogeneity: Classifier calibration for federated learning with non-iid data. Advances in Neural Information Processing Systems 34, 5972-5984 (2021)", + "39. McCloskey, M., Cohen, N.J.: Catastrophic interference in connectionist networks: The sequential learning problem. In: Psychology of learning and motivation, vol. 24, pp. 109-165. Elsevier (1989)", + "40. McMahan, B., Moore, E., Ramage, D., Hampson, S., y Arcas, B.A.: Communication-efficient learning of deep networks from decentralized data. In: Artificial intelligence and statistics. pp. 1273-1282. PMLR (2017)", + "41. Niknam, S., Dhillon, H.S., Reed, J.H.: Federated learning for wireless communications: Motivation, opportunities, and challenges. IEEE Communications Magazine 58(6), 46-51 (2020)", + "42. Ortiz-Jimenez, G., Favero, A., Frossard, P.: Task arithmetic in the tangent space: Improved editing of pre-trained models. Advances in Neural Information Processing Systems 36 (2024)", + "43. Qi, D., Zhao, H., Li, S.: Better generative replay for continual federated learning. In: The Eleventh International Conference on Learning Representations (2022)", + "44. Qu, L., Zhou, Y., Liang, P.P., Xia, Y., Wang, F., Adeli, E., Fei-Fei, L., Rubin, D.: Rethinking architecture design for tackling data heterogeneity in federated learning. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 10061-10071 (2022)", + "45. Radford, A., Kim, J.W., Hallacy, C., Ramesh, A., Goh, G., Agarwal, S., Sastry, G., Askell, A., Mishkin, P., Clark, J., et al.: Learning transferable visual models from natural language supervision. In: International conference on machine learning. pp. 8748-8763. PMLR (2021)", + "46. Rebuffi, S.A., Kolesnikov, A., Sperl, G., Lampert, C.H.: icarl: Incremental classifier and representation learning. In: Proceedings of the IEEE conference on Computer Vision and Pattern Recognition. pp. 2001-2010 (2017)", + "47. Rieke, N., Hancox, J., Li, W., Miletari, F., Roth, H.R., Albarqouni, S., Bakas, S., Galtier, M.N., Landman, B.A., Maier-Hein, K., et al.: The future of digital health with federated learning. NPJ digital medicine 3(1), 119 (2020)" + ], + "bbox": [ + 212, + 146, + 787, + 840 + ], + "page_idx": 16 + }, + { + "type": "header", + "text": "PILoRA", + "bbox": [ + 673, + 114, + 730, + 126 + ], + "page_idx": 16 + }, + { + "type": "page_number", + "text": "17", + "bbox": [ + 767, + 114, + 785, + 126 + ], + "page_idx": 16 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "48. Shang, X., Lu, Y., Huang, G., Wang, H.: Federated learning on heterogeneous and long-tailed data via classifier re-training with federated features. arXiv preprint arXiv:2204.13399 (2022)", + "49. Shenaj, D., Toldo, M., Rigon, A., Zanuttigh, P.: Asynchronous federated continual learning. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 5054-5062 (2023)", + "50. Smith, J.S., Karlinsky, L., Gutta, V., Cascante-Bonilla, P., Kim, D., Arbelle, A., Panda, R., Feris, R., Kira, Z.: Coda-prompt: Continual decomposed attention-based prompting for rehearsal-free continual learning. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 11909-11919 (2023)", + "51. Tan, Y., Long, G., Liu, L., Zhou, T., Lu, Q., Jiang, J., Zhang, C.: Fedproto: Federated prototype learning across heterogeneous clients. In: Proceedings of the AAAI Conference on Artificial Intelligence. vol. 36, pp. 8432-8440 (2022)", + "52. Voigt, P., Von dem Bussche, A.: The eu general data protection regulation (gdpr). A Practical Guide, 1st Ed., Cham: Springer International Publishing 10(3152676), 10-5555 (2017)", + "53. Wang, X., Chen, T., Ge, Q., Xia, H., Bao, R., Zheng, R., Zhang, Q., Gui, T., Huang, X.: Orthogonal subspace learning for language model continual learning. arXiv preprint arXiv:2310.14152 (2023)", + "54. Wang, Z., Zhang, Z., Ebrahimi, S., Sun, R., Zhang, H., Lee, C.Y., Ren, X., Su, G., Perot, V., Dy, J., et al.: Dualprompt: Complementary prompting for rehearsal-free continual learning. In: European Conference on Computer Vision. pp. 631-648. Springer (2022)", + "55. Wang, Z., Zhang, Z., Lee, C.Y., Zhang, H., Sun, R., Ren, X., Su, G., Perot, V., Dy, J., Pfister, T.: Learning to prompt for continual learning. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 139-149 (2022)", + "56. Yang, H.M., Zhang, X.Y., Yin, F., Liu, C.L.: Robust classification with convolutional prototype learning. In: Proceedings of the IEEE conference on computer vision and pattern recognition. pp. 3474-3482 (2018)", + "57. Yang, H.M., Zhang, X.Y., Yin, F., Yang, Q., Liu, C.L.: Convolutional prototype network for open set recognition. IEEE Transactions on Pattern Analysis and Machine Intelligence 44(5), 2358-2370 (2020)", + "58. Yi, L., Yu, H., Wang, G., Liu, X.: Fedlora: Model-heterogeneous personalized federated learning with lora tuning. arXiv preprint arXiv:2310.13283 (2023)", + "59. Yoon, J., Jeong, W., Lee, G., Yang, E., Hwang, S.J.: Federated continual learning with weighted inter-client transfer. In: International Conference on Machine Learning. pp. 12073-12086. PMLR (2021)", + "60. Zhang, G., Wang, L., Kang, G., Chen, L., Wei, Y.: Slow learner with classifier alignment for continual learning on a pre-trained model. arXiv preprint arXiv:2303.05118 (2023)", + "61. Zhang, J., Chen, C., Zhuang, W., Lv, L.: Addressing catastrophic forgetting in federated class-continual learning. arXiv preprint arXiv:2303.06937 (2023)", + "62. Zhang, J., Chen, C., Zhuang, W., Lyu, L.: Target: Federated class-continual learning via exemplar-free distillation. In: Proceedings of the IEEE/CVF International Conference on Computer Vision. pp. 4782-4793 (2023)", + "63. Zhao, H., Du, W., Li, F., Li, P., Liu, G.: Reduce communication costs and preserve privacy: Prompt tuning method in federated learning. arXiv preprint arXiv:2208.12268 (2022)" + ], + "bbox": [ + 215, + 147, + 784, + 839 + ], + "page_idx": 17 + }, + { + "type": "page_number", + "text": "18", + "bbox": [ + 217, + 114, + 235, + 126 + ], + "page_idx": 17 + }, + { + "type": "header", + "text": "H. Guo et al.", + "bbox": [ + 271, + 114, + 359, + 126 + ], + "page_idx": 17 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "64. Zhu, F., Cheng, Z., Zhang, X.y., Liu, C.l.: Class-incremental learning via dual augmentation. Advances in Neural Information Processing Systems 34, 14306-14318 (2021)", + "65. Zhu, F., Ma, S., Cheng, Z., Zhang, X.Y., Zhang, Z., Liu, C.L.: Open-world machine learning: A review and new outlooks. arXiv preprint arXiv:2403.01759 (2024)", + "66. Zhu, F., Zhang, X.Y., Wang, C., Yin, F., Liu, C.L.: Prototype augmentation and self-supervision for incremental learning. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 5871-5880 (2021)" + ], + "bbox": [ + 212, + 146, + 787, + 258 + ], + "page_idx": 18 + }, + { + "type": "header", + "text": "PILORA", + "bbox": [ + 671, + 114, + 730, + 126 + ], + "page_idx": 18 + }, + { + "type": "page_number", + "text": "19", + "bbox": [ + 767, + 114, + 785, + 126 + ], + "page_idx": 18 + } +] \ No newline at end of file diff --git a/2024/PILoRA_ Prototype Guided Incremental LoRA for Federated Class-Incremental Learning/b10fddc1-55da-4bf4-bfbe-8037f4bdbf91_model.json b/2024/PILoRA_ Prototype Guided Incremental LoRA for Federated Class-Incremental Learning/b10fddc1-55da-4bf4-bfbe-8037f4bdbf91_model.json new file mode 100644 index 0000000000000000000000000000000000000000..f9b2401cc275efa4ae7948fd46ebb5a97ef7f3b0 --- /dev/null +++ b/2024/PILoRA_ Prototype Guided Incremental LoRA for Federated Class-Incremental Learning/b10fddc1-55da-4bf4-bfbe-8037f4bdbf91_model.json @@ -0,0 +1,2713 @@ +[ + [ + { + "type": "title", + "bbox": [ + 0.234, + 0.141, + 0.773, + 0.187 + ], + "angle": 0, + "content": "PILoRA: Prototype Guided Incremental LoRA for Federated Class-Incremental Learning" + }, + { + "type": "text", + "bbox": [ + 0.344, + 0.213, + 0.658, + 0.245 + ], + "angle": 0, + "content": "Haiyang Guo\\(^{1,2}\\), Fei Zhu\\(^{3}\\), Wenzhuo Liu\\(^{1,2}\\), Xu-Yao Zhang\\(^{1,2\\star}\\), and Cheng-Lin Liu\\(^{1,2}\\)" + }, + { + "type": "text", + "bbox": [ + 0.287, + 0.255, + 0.714, + 0.27 + ], + "angle": 0, + "content": "\\(^{1}\\) MAIS, Institute of Automation, Chinese Academy of Sciences" + }, + { + "type": "text", + "bbox": [ + 0.245, + 0.27, + 0.757, + 0.283 + ], + "angle": 0, + "content": "\\(^{2}\\) School of Artificial Intelligence, University of Chinese Academy of Sciences" + }, + { + "type": "text", + "bbox": [ + 0.218, + 0.283, + 0.783, + 0.311 + ], + "angle": 0, + "content": "3 Centre for Artificial Intelligence and Robotics, Hong Kong Institute of Science and Innovation, Chinese Academy of Sciences" + }, + { + "type": "text", + "bbox": [ + 0.235, + 0.312, + 0.767, + 0.34 + ], + "angle": 0, + "content": "{guohaiyang2023, zhufei2018}@ia.ac.cn, liuwenzhuo20@mails.ucas.ac.cn, xyz, liucl}@nlpr.ia.ac.cn" + }, + { + "type": "text", + "bbox": [ + 0.261, + 0.382, + 0.744, + 0.633 + ], + "angle": 0, + "content": "Abstract. Existing federated learning methods have effectively dealt with decentralized learning in scenarios involving data privacy and non-IID data. However, in real-world situations, each client dynamically learns new classes, requiring the global model to classify all seen classes. To effectively mitigate catastrophic forgetting and data heterogeneity under low communication costs, we propose a simple and effective method named PILoRA. On the one hand, we adopt prototype learning to learn better feature representations and leverage the heuristic information between prototypes and class features to design a prototype re-weight module to solve the classifier bias caused by data heterogeneity without retraining the classifier. On the other hand, we view incremental learning as the process of learning distinct task vectors and encoding them within different LoRA parameters. Accordingly, we propose Incremental LoRA to mitigate catastrophic forgetting. Experimental results on standard datasets indicate that our method outperforms the state-of-the-art approaches significantly. More importantly, our method exhibits strong robustness and superiority in different settings and degrees of data heterogeneity. The code is available at https://github.com/GhyO501/PILORA." + }, + { + "type": "text", + "bbox": [ + 0.262, + 0.645, + 0.68, + 0.661 + ], + "angle": 0, + "content": "Keywords: Federated Learning \\(\\cdot\\) Class Incremental Learning" + }, + { + "type": "title", + "bbox": [ + 0.216, + 0.689, + 0.377, + 0.705 + ], + "angle": 0, + "content": "1 Introduction" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.723, + 0.788, + 0.816 + ], + "angle": 0, + "content": "Federated learning (FL) [40] is a novel distributed machine learning paradigm that enables multiple data owners to collaboratively train a shared model while ensuring the privacy of their local data. In recent years, with the increasing emphasis on data privacy in society and the refinement of relevant regulations [52], FL has experienced rapid development and has been widely applied in various real-world scenarios [14,20,30,41,47]." + }, + { + "type": "page_footnote", + "bbox": [ + 0.218, + 0.825, + 0.383, + 0.841 + ], + "angle": 0, + "content": "* Corresponding author" + } + ], + [ + { + "type": "page_number", + "bbox": [ + 0.218, + 0.116, + 0.23, + 0.127 + ], + "angle": 0, + "content": "2" + }, + { + "type": "header", + "bbox": [ + 0.272, + 0.115, + 0.361, + 0.128 + ], + "angle": 0, + "content": "H. Guo et al." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.147, + 0.788, + 0.33 + ], + "angle": 0, + "content": "Existing FL methods [19, 28, 40] typically rely on the closed-world assumption [6, 13, 65], that is, the number of classes seen by the model remains the same and unchanged during both the training and testing stages. However, the real world is dynamic and constantly changing, the local clients often need to receive new data to update the global model continuously. Therefore, Federated Class Incremental Learning (FCIL) [10, 59] has been proposed to handle FL tasks in dynamic scenarios. Specifically, each local client can only update the model with new class data at each stage and upload the model parameters to the global server for aggregation, while the global model needs to maintain its discriminative ability for all seen classes. Moreover, the data distribution of local data follows the non-independent and identically distributed (non-IID) assumption [27]." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.335, + 0.789, + 0.623 + ], + "angle": 0, + "content": "FCIL presents a more realistic setting for real-world applications, but it also introduces greater challenges, as FCIL needs to simultaneously address the issue of catastrophic forgetting [23, 39] and data heterogeneity resulting from non-IID. In existing studies, one way is to store a subset of data from old classes and train them together when learning new tasks [9, 10], but the amount of old data stored is strictly limited due to privacy protection requirements. An alternative way is to utilize generative models to generate pseudo-samples of old data [43, 61], thus retaining the ability to classify the old classes. However, training a good generator will cause greater computational overhead, and the generator also suffers from catastrophic forgetting. Another direction is to use pre-trained models for fine-tuning [1, 33], these methods store different stages of knowledge by maintaining a pool of modules and inserting the corresponding modules at inference time based on the similarity between the input data and the modules. However, storing these modules takes up additional memory space. Although they provide different perspectives for solving FCIL, there is still a large unexplored space, especially the performance of the model under different non-IID settings and different degrees of data heterogeneity. So in order to better address this issue, we start by addressing a fundamental question, what is the panacea for dealing with CIL and FL?" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.629, + 0.79, + 0.842 + ], + "angle": 0, + "content": "In other words, if we can find common ground in dealing with CIL and FL, it will be of great help in solving the FCIL problem. Specifically, we observe that: 1) At the level of feature representation, both FL and CIL tasks require models to learn a intra-class compact and inter-class separable feature representation. On the one hand, for the CIL task, this kind of feature representation helps to reduce the overlap between features of the old and new classes in the deep feature space, thus alleviating the forgetting of old knowledge. For this reason, PASS [66] introduces self-supervised learning to assist the learning of feature representations. On the other hand, in the FL task, FedEx [51] imposes a constraint on the deep features of samples belonging to the same class. This constraint ensures that these features are proximate to the global prototypes of their respective classes. 2) At the classifier level, classifier drift is the common enemy of both. In CIL, when learning new classes, the decision boundaries learned from old classes can be severely disrupted, resulting in a significant bias" + } + ], + [ + { + "type": "header", + "bbox": [ + 0.673, + 0.115, + 0.732, + 0.127 + ], + "angle": 0, + "content": "PILORA" + }, + { + "type": "page_number", + "bbox": [ + 0.775, + 0.116, + 0.785, + 0.127 + ], + "angle": 0, + "content": "3" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.147, + 0.788, + 0.298 + ], + "angle": 0, + "content": "in the classification layer [64, 66]. To address this issue, some methods [2, 3, 46] alleviate bias by directly retaining a portion of the data from previous categories and training the model with new data. Other methods [64, 66] address classifier bias by retaining pseudo features of old classes to assist in training the classification layer; In FL, CCVR [38] assesses the similarity among different layers within local models under data heterogeneity and reveals that classifiers exhibit the lowest similarity. This suggests that the classifier of each client models is seriously biased to the local data. Similarly to CIL, [38, 48] utilize the statistical information of local features to retrain the classifier on the global server to alleviate classifier bias." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.312, + 0.506, + 0.688 + ], + "angle": 0, + "content": "Motivated by the above findings, in this paper, we propose a Prototype Guided Incremental LoRA (PILoRA) model to tackle the FCIL problem. Specifically, we adopt prototype classifiers in our model to learn intraclass compact and inter-class separable feature representations, which is beneficial for both CIL and FL [28, 35, 56, 57]. We use the pre-trained Transformer model (ViT) as the backbone because the global interactions learned by Transformer are significantly more robust than the local patterns learned by CNN for FL tasks [44] (Fig. 1.), and also provide a good representation of the features. Considering the communication cost, we freeze the entire backbone network during training and use LoRA, a Parameter-Efficient Fine-Tuning (PEFT) method to train the model. To address the catastrophic forgetting problem in FCIL, we propose incremental LoRA, which constrains different stages of LoRA to be trained on" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.69, + 0.788, + 0.841 + ], + "angle": 0, + "content": "subspaces orthogonal to each other via orthogonal regularization, and acquires knowledge learned in the past by simple and efficient summation at inference time. To address the classifier bias caused by data heterogeneity, we design a prototype re-weight module for the global server, which aggregates local prototypes based on heuristic information between prototypes and corresponding class features. Compared to other existing FCIL methods, our model achieves superior results on standard datasets. Furthermore, we explore the model's performance under different non-IID settings and degrees of data heterogeneity. Experiments demonstrate the robustness of our approach, while other methods suffer from significant declines." + }, + { + "type": "image", + "bbox": [ + 0.522, + 0.348, + 0.783, + 0.493 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.514, + 0.509, + 0.789, + 0.661 + ], + "angle": 0, + "content": "Fig.1: The local data of client 1 and client 2 are non-IID. Resnet50 (the first row) focuses more on local patterns, and the patterns learned are significantly different in the case of data heterogeneity, causing the average model to lose some important information (e.g., fish fins). However, ViT (the second row) is less affected by data heterogeneity, and the averaged model basically retains all the information learned by the local model." + } + ], + [ + { + "type": "page_number", + "bbox": [ + 0.218, + 0.116, + 0.23, + 0.127 + ], + "angle": 0, + "content": "4" + }, + { + "type": "header", + "bbox": [ + 0.272, + 0.115, + 0.361, + 0.128 + ], + "angle": 0, + "content": "H. Guo et al." + }, + { + "type": "text", + "bbox": [ + 0.218, + 0.147, + 0.649, + 0.162 + ], + "angle": 0, + "content": "The main contributions of this paper can be summarized as:" + }, + { + "type": "text", + "bbox": [ + 0.225, + 0.17, + 0.784, + 0.214 + ], + "angle": 0, + "content": "- We propose incremental LoRA, which performs incremental learning on orthogonal subspaces to mitigate catastrophic forgetting and aggregates previous knowledge through simple and efficient parameter summation." + }, + { + "type": "text", + "bbox": [ + 0.226, + 0.216, + 0.785, + 0.274 + ], + "angle": 0, + "content": "- We propose a prototype re-weight module to form global prototypes using heuristic information between each class of local prototypes and the corresponding class of features. Our method effectively addresses classifier bias induced by data heterogeneity without retraining." + }, + { + "type": "text", + "bbox": [ + 0.226, + 0.276, + 0.784, + 0.319 + ], + "angle": 0, + "content": "- We conduct extensive experiments on standard datasets and achieve state-of-the-art performance. Furthermore, in the extreme heterogeneous case, our method still maintains robustness, while all other methods suffer plummet." + }, + { + "type": "list", + "bbox": [ + 0.225, + 0.17, + 0.785, + 0.319 + ], + "angle": 0, + "content": null + }, + { + "type": "title", + "bbox": [ + 0.217, + 0.342, + 0.388, + 0.358 + ], + "angle": 0, + "content": "2 Related Work" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.372, + 0.788, + 0.688 + ], + "angle": 0, + "content": "Federated Class-Incremental Learning: Research on FCIL has garnered considerable attention in recent years. Dong et al. [10] first introduces the concept of FCIL and proposes several loss functions on the local side and the global server side to alleviate local catastrophic forgetting and global catastrophic forgetting. However, their method uses a rehearsal buffer to store and retain old class data, and additionally design a proxy server to select the best model, resulting in large memory overhead and communication costs. LGA [9] extends the work of [10], but it still belongs to rehearsal-based FCIL. In the more challenging rehearsal-free FCIL problem, generate models are widely adopted to produce synthetic data that aim to mitigate the catastrophic forgetting on local side and global side [43,62], but its performance is highly dependent on the quality of synthetic data and will incur additional computational costs. Similar to our work, FedSpace [49] designs prototype-based loss to encourage feature vectors of the same class to be close together, while we achieve this goal by using prototype classifier. In contrast to the methods described above, some recent studies [1,33] combine pre-trained models with FCIL and achieve higher performance at a smaller communication cost. However, they both adopt a similarity-based selection strategy, which causes additional memory overheads in inference. Besides, they all use supervised pre-training weights, whereas we argue that this may pose privacy concerns because data from downstream tasks may overlap with pre-trained datasets." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.689, + 0.788, + 0.841 + ], + "angle": 0, + "content": "PEFT for Pre-Trained Model: With the emergence of large-scale pre-trained models [4, 22, 45], how to effectively fine-tune these models to adapt the downstream tasks has been a focal point of attention. Recently, LoRA [17], Prompt [34], and Adapter [16] have emerged as standout techniques and have been widely used in CIL [12, 50, 54, 55] and FL [15, 58, 63] tasks. In FCIL, existing methods [1, 33] have attempted to combine Prompt and Adapter with pre-trained models. Specifically, they store the knowledge learned at each stage in the parameters of the Prompt or Adapter module and select the appropriate module for the current input to be embedded into the model by a specific similarity computation during the inference time, thus effectively mitigating catastrophic" + } + ], + [ + { + "type": "header", + "bbox": [ + 0.673, + 0.115, + 0.732, + 0.127 + ], + "angle": 0, + "content": "PILORA" + }, + { + "type": "page_number", + "bbox": [ + 0.776, + 0.117, + 0.786, + 0.127 + ], + "angle": 0, + "content": "5" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.147, + 0.788, + 0.224 + ], + "angle": 0, + "content": "forgetting with a tiny communication costs. However, such similarity matching based approaches undoubtedly introduce inference delays since they require additional similarity computation modules. (In [33], they even train a separate CNN to compute the similarity.) Furthermore, they need a additional memory space in the global server to set up a parameter pool to store the module parameters." + }, + { + "type": "title", + "bbox": [ + 0.216, + 0.246, + 0.383, + 0.263 + ], + "angle": 0, + "content": "3 Preliminaries" + }, + { + "type": "text", + "bbox": [ + 0.215, + 0.278, + 0.79, + 0.404 + ], + "angle": 0, + "content": "In FCIL setting, each client has a local stream dataset \\( D_{k} = \\{D_{k}^{t}\\}_{t=1}^{T} \\), where \\( D_{k}^{t} = \\{\\mathbf{X}_{k}^{t}, \\mathbf{Y}_{k}^{t}\\} = \\{x_{k,i}^{t}, y_{k,i}^{t}\\}_{i=1}^{N_{t}} \\) is the dataset of the \\( k \\)-th client on task \\( t \\). Dataset \\( D_{k}^{t} \\) contains \\( N_{k}^{t} \\) training samples and their label \\( \\mathbf{Y}_{k}^{t} \\in C_{k}^{t} \\), where \\( C_{k}^{t} \\) is the class set of the \\( k \\)-th client in task \\( t \\). In particular, the distribution of different clients \\( k \\) under the same task is non-IID and the class sets of different tasks \\( t \\) are disjoint. For local client, the objective is to minimize a pre-defined loss function \\( \\mathcal{L} \\) on current dataset \\( D_{k}^{t} \\), while avoiding interference with and possibly enhancing the knowledge acquired from previous learning stages:" + }, + { + "type": "equation", + "bbox": [ + 0.393, + 0.413, + 0.786, + 0.442 + ], + "angle": 0, + "content": "\\[\n\\underset {\\boldsymbol {\\omega} _ {k} ^ {t}} {\\operatorname {a r g m i n}} \\mathcal {L} \\left(\\boldsymbol {\\omega} _ {k} ^ {t}; \\boldsymbol {\\omega} ^ {t - 1}, \\mathbf {X} _ {k} ^ {t}, \\mathbf {Y} _ {k} ^ {t}\\right), \\tag {1}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.215, + 0.454, + 0.788, + 0.501 + ], + "angle": 0, + "content": "where \\(\\omega_{k}^{t}\\) is the parameters of \\(k\\)-th local model, and \\(\\omega^{t-1}\\) is the global model at previous task. Then, the server updates the global model \\(\\omega^{t}\\) by aggregating all uploaded parameters as follows:" + }, + { + "type": "equation", + "bbox": [ + 0.369, + 0.512, + 0.786, + 0.553 + ], + "angle": 0, + "content": "\\[\n\\boldsymbol {\\omega} ^ {t} = \\sum_ {k = 1} ^ {K} \\gamma_ {k} \\boldsymbol {\\omega} _ {k} ^ {t}, \\text {w h e r e} \\gamma_ {k} = \\frac {N _ {k} ^ {t}}{\\sum_ {k ^ {\\prime}} N _ {k ^ {\\prime}} ^ {t}}. \\tag {2}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.215, + 0.563, + 0.788, + 0.594 + ], + "angle": 0, + "content": "The global model aims to correctly classify the test samples of all seen classes and solve the problem of data heterogeneity at low communication cost." + }, + { + "type": "title", + "bbox": [ + 0.216, + 0.617, + 0.374, + 0.634 + ], + "angle": 0, + "content": "4 Our method" + }, + { + "type": "title", + "bbox": [ + 0.216, + 0.649, + 0.609, + 0.665 + ], + "angle": 0, + "content": "4.1 Incremental LoRA for Pre-Trained Model" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.674, + 0.789, + 0.844 + ], + "angle": 0, + "content": "In FCIL, storing the knowledge of different stages in different modules through PEFT can effectively mitigate the catastrophic forgetting [1,33], but additional similarity computation units are required in the inference time in order to select the appropriate module to embed into the model based on the inputs, which leads to inference delay and additional memory overhead. In order to learn an end-to-end global model without taking up extra storage space, we intuitively believe that modules storing knowledge from different stages can be organically combined into one module that embeds knowledge from all stages. Accordingly, we propose Incremental LoRA to effectively address catastrophic forgetting in FCIL. On the one hand, LoRA has the natural advantages of low inference latency and more stable training [17], on the other hand, Inspired by [53, 55], we" + } + ], + [ + { + "type": "page_number", + "bbox": [ + 0.218, + 0.116, + 0.23, + 0.127 + ], + "angle": 0, + "content": "6" + }, + { + "type": "header", + "bbox": [ + 0.272, + 0.115, + 0.361, + 0.127 + ], + "angle": 0, + "content": "H. Guo et al." + }, + { + "type": "image", + "bbox": [ + 0.233, + 0.149, + 0.769, + 0.383 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.214, + 0.397, + 0.788, + 0.454 + ], + "angle": 0, + "content": "Fig. 2: Illustration of PILoRA for FCIL. The client fine-tunes LoRA and prototypes for each class using a local dataset. Upon upload, the global server aggregates the LoRA uploaded by different clients, and applies a prototype re-weight module before re-sending them to each client." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.485, + 0.785, + 0.529 + ], + "angle": 0, + "content": "introduce the orthogonality loss to constrain LoRA to learn new knowledge in a subspace orthogonal to past tasks, thus better preserving the knowledge of the old classes." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.53, + 0.785, + 0.577 + ], + "angle": 0, + "content": "Specifically, we define the initialization parameters of the pre-training model as \\(\\mathbf{W} \\in \\mathbb{R}^{d \\times k}\\), and \\(\\Delta \\mathbf{W}_{\\mathbf{t}}\\) represents the parameter to be updated for task \\(t\\). Then, the model's updates in different incremental learning stages can be expressed as:" + }, + { + "type": "equation", + "bbox": [ + 0.457, + 0.592, + 0.785, + 0.609 + ], + "angle": 0, + "content": "\\[\n\\mathbf {W} + \\Delta \\mathbf {W} _ {\\mathbf {t}}. \\tag {3}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.617, + 0.787, + 0.648 + ], + "angle": 0, + "content": "LoRA assumes that the weight changes of large-scale pre-trained models when adapting to downstream tasks occur in a low-rank space:" + }, + { + "type": "equation", + "bbox": [ + 0.406, + 0.66, + 0.786, + 0.677 + ], + "angle": 0, + "content": "\\[\n\\mathbf {W} + \\Delta \\mathbf {W} _ {\\mathbf {t}} = \\mathbf {W} + \\mathbf {A} _ {t} \\mathbf {B} _ {t}, \\tag {4}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.688, + 0.787, + 0.764 + ], + "angle": 0, + "content": "where \\(\\mathbf{A}_t \\in \\mathbb{R}^{d \\times r}\\), \\(\\mathbf{B}_t \\in \\mathbb{R}^{r \\times k}\\), and \\(r \\ll \\min\\{d, k\\}\\). At first, \\(\\mathbf{A}_t\\) is initialized by a random Gaussian, while \\(\\mathbf{B}_t\\) is initialized with zero. Therefore, \\(\\mathbf{B}_t\\) can be regarded as the coefficient matrix of \\(\\mathbf{A}_t\\) [53]. LoRA applies it as a bypass to query and value projection matrices in multi-head attention module, and during adaptation to downstream tasks, only the parameters of \\(\\mathbf{A}_t\\) and \\(\\mathbf{B}_t\\) are trainable." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.766, + 0.787, + 0.84 + ], + "angle": 0, + "content": "Inspired by [53], we argue that the parameters of LoRA can be viewed as containers that store different subspaces of task gradients. Thus, learning a series of incremental learning tasks can be viewed as learning a series of LoRA parameters. So how to use these LoRA modules to build a model that can classify all seen classes? An idea is to select appropriate modules to be embedded in the" + } + ], + [ + { + "type": "header", + "bbox": [ + 0.673, + 0.115, + 0.732, + 0.127 + ], + "angle": 0, + "content": "PILORA" + }, + { + "type": "page_number", + "bbox": [ + 0.776, + 0.116, + 0.786, + 0.127 + ], + "angle": 0, + "content": "7" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.147, + 0.784, + 0.281 + ], + "angle": 0, + "content": "model based on the similarity between different modules and the corresponding inputs [1, 33]. However, computing similarity requires additional computational resources, which leads to delayed inference. Another idea is that since LoRA is essentially two weight matrices, it is possible to concat all LoRA parameters prior to stage \\( t \\) into a new matrix to gain knowledge of all previous stages [53]. In particular, at stage \\( t \\), the knowledge learned in the previous \\( t - 1 \\) stages are stored in \\( \\mathbf{A}_{1:t - 1} = [\\mathbf{A}_1,\\dots ,\\mathbf{A}_{t - 1}] \\) and \\( \\mathbf{B}_{1:t - 1} = [\\mathbf{B}_1,\\dots ,\\mathbf{B}_{t - 1}] \\). To obtain the weight matrix, we can sequentially concat these LoRA parameters into a bigger module:" + }, + { + "type": "equation", + "bbox": [ + 0.41, + 0.282, + 0.785, + 0.299 + ], + "angle": 0, + "content": "\\[\n\\mathbf {W} + \\Delta \\mathbf {W} _ {\\mathbf {t}} = \\mathbf {W} + \\widetilde {\\mathbf {A}} _ {t} \\widetilde {\\mathbf {B}} _ {t}, \\tag {5}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.306, + 0.785, + 0.399 + ], + "angle": 0, + "content": "where \\(\\widetilde{\\mathbf{A}}_t = \\mathrm{concat}([\\mathbf{A}_1,\\ldots ,\\mathbf{A}_t])\\), \\(\\widetilde{\\mathbf{B}}_t = \\mathrm{concat}([\\mathbf{B}_1,\\ldots ,\\mathbf{B}_t])\\). However, this will cause the parameters of LoRA in the global model to increase with the number of incremental learning tasks. Therefore, in order to ensure consistency between the local model and the global model, we are inspired by Task Arithmetic [7, 18, 42] in model edit and propose to integrate the LoRA parameters at different stages by summation:" + }, + { + "type": "equation", + "bbox": [ + 0.289, + 0.412, + 0.785, + 0.451 + ], + "angle": 0, + "content": "\\[\n\\mathbf {W} + \\Delta \\mathbf {W} _ {\\mathbf {t}} = \\mathbf {W} + \\overline {{\\mathbf {A}}} _ {t} \\overline {{\\mathbf {B}}} _ {t}, \\text {w h e r e} \\overline {{\\mathbf {A}}} _ {t} = \\sum_ {i = 1} ^ {t} \\mathbf {A} _ {i}, \\overline {{\\mathbf {B}}} _ {t} = \\sum_ {i = 1} ^ {t} \\mathbf {B} _ {i}. \\tag {6}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.457, + 0.785, + 0.608 + ], + "angle": 0, + "content": "Specifically, distinct directions in the weight space correspond to various localized regions in the input space [42]. Consequently, a linear combination of these diverse directions encoded within the pre-trained weights enables the model to effectively discriminate between different inputs. In our method, The training data for different incremental tasks are disjoint from each other, so it is reasonable to directly sum the LoRA parameters corresponding to different tasks. Besides, Task Arithmetic also point out that orthogonality between different task parameter vectors helps to better integrate different task. So we propose a orthogonal regularization to achieve this by constraining the parameters of LoRA to be orthogonal to previous tasks:" + }, + { + "type": "equation", + "bbox": [ + 0.402, + 0.621, + 0.785, + 0.658 + ], + "angle": 0, + "content": "\\[\nl _ {o r t} \\left(\\mathbf {A} _ {i}, \\mathbf {A} _ {t}\\right) = \\sum_ {i = 1} ^ {t - 1} \\left| \\mathbf {A} _ {i} ^ {T} \\cdot \\mathbf {A} _ {t} \\right|. \\tag {7}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.671, + 0.784, + 0.747 + ], + "angle": 0, + "content": "Meanwhile, the orthogonal regularization encourages the learning of distinct tasks along orthogonal directions, thereby effectively reducing the spatial overlap between them. This approach will be proved to be advantageous in mitigating catastrophic forgetting, as it helps preserve the knowledge learned from previous tasks while adapting to new ones." + }, + { + "type": "title", + "bbox": [ + 0.215, + 0.769, + 0.637, + 0.785 + ], + "angle": 0, + "content": "4.2 Prototype learning with Prototype Re-weight" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.795, + 0.784, + 0.841 + ], + "angle": 0, + "content": "Prototype Learning: As mentioned before, a feature representation that is separable intra-class and compact inter-class is helpful for FCIL tasks. Therefore, the model not only needs to correctly classify known classes but also needs" + } + ], + [ + { + "type": "page_number", + "bbox": [ + 0.218, + 0.116, + 0.23, + 0.127 + ], + "angle": 0, + "content": "8" + }, + { + "type": "header", + "bbox": [ + 0.272, + 0.115, + 0.361, + 0.128 + ], + "angle": 0, + "content": "H. Guo et al." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.147, + 0.788, + 0.253 + ], + "angle": 0, + "content": "to model the distribution of known classes in the feature space. In open set recognition, CPN [57] designs a discriminative loss and generative loss for prototype learning to constrain the range of known classes in the feature space, thus reserving space for samples from unknown classes. Inspired by it, we introduce prototype learning in FCIL. In particular, we set a prototype \\( m = \\{m_i | i = 1, 2, \\dots, C\\} \\) for each class, where \\( m_i \\in \\mathbb{R}^d \\), and the dimensions \\( d \\) of each prototype are the same as the dimensions of the final deep feature space." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.253, + 0.788, + 0.33 + ], + "angle": 0, + "content": "To shorten the distance between the class feature and the corresponding prototype, we apply the distance-based cross entropy (DCE) discriminative loss. Given a sample \\((x,y)\\), DCE uses the distance between sample features \\(f_{\\theta}(x)\\) and prototype \\(m_{i}\\) to represent the probability of belonging to class \\(i\\). Considering the normalization of probability, DCE adopts the softmax operation:" + }, + { + "type": "equation", + "bbox": [ + 0.335, + 0.336, + 0.786, + 0.375 + ], + "angle": 0, + "content": "\\[\np (x \\in m _ {i} | x) = \\frac {\\exp (- \\delta \\cdot \\| f _ {\\theta} (x) - m _ {i} \\| _ {2} ^ {2})}{\\sum_ {j = 1} ^ {C} \\exp (- \\delta \\cdot \\| f _ {\\theta} (x) - m _ {j} \\| _ {2} ^ {2})}, \\tag {8}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.215, + 0.381, + 0.789, + 0.442 + ], + "angle": 0, + "content": "where \\( \\| f_{\\theta}(x) - m_i\\| _2^2 \\) is the Euclidean distance between the feature \\( f_{\\theta}(x) \\) of the input sample and the prototype \\( m_{i} \\), and \\( \\delta \\) is the temperature scalar controlling the hardness of the class distribution. Hence, the distance-based-cross-entropy loss can be defined as:" + }, + { + "type": "equation", + "bbox": [ + 0.374, + 0.452, + 0.786, + 0.469 + ], + "angle": 0, + "content": "\\[\nl _ {d c e} ((x, y); \\theta , m) = - \\log (x \\in m _ {y} | x). \\tag {9}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.215, + 0.475, + 0.788, + 0.552 + ], + "angle": 0, + "content": "By minimizing the DCE loss, the distance between the sample's feature and the correct prototype will be smaller than other incorrect prototypes. However, features learned only under the discriminative loss may not be compact, which may lead to an overlap in feature representations between new and old classes. To solve this problem, we introduce prototype learning (PL) loss [57]:" + }, + { + "type": "equation", + "bbox": [ + 0.384, + 0.559, + 0.786, + 0.577 + ], + "angle": 0, + "content": "\\[\nl _ {p l} ((x, y); \\theta , m) = \\| f _ {\\theta} (x) - m _ {y} \\| _ {2} ^ {2}, \\tag {10}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.215, + 0.584, + 0.788, + 0.645 + ], + "angle": 0, + "content": "PL loss decreases the distance between sample features and the corresponding correct prototype, making the model learn a more compact intra-class distribution. Essentially, PL loss is a maximum likelihood regularization of features \\( f_{\\theta}(x) \\) under the Gaussian mixture density assumption [31, 32]." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.645, + 0.788, + 0.795 + ], + "angle": 0, + "content": "Prototype Re-weight: Considering a specific class \\( c \\), under the non-IID setting, each client \\( k \\) holds a portion of the training data \\( N_{c,k} \\) and \\( \\sum_{k=1}^{K} N_{c,k} = N_{c} \\), where \\( N_{c} \\) is the sum of training samples of class \\( c \\). In prototype learning, the prototype \\( m_{k,c} \\) learned by each client can effectively reflect the distribution of class \\( c \\) within its local data. That is, for clients that do not have training samples of class \\( c \\), the distance between their corresponding prototype and the features of class \\( c \\) will be greater than for clients that do have training samples. If we give equal weight to these uploaded prototypes during parameter aggregation, classifier drift may occur, especially when the data heterogeneity is high, potentially resulting in the global model losing discriminative capability for that class." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.796, + 0.788, + 0.842 + ], + "angle": 0, + "content": "Therefore, we use the heuristic information contained in the distance between the prototype and the corresponding class average features to design the prototype re-weight module. Specifically, at stage \\( t \\), each client \\( k \\) uploads prototypes" + } + ], + [ + { + "type": "header", + "bbox": [ + 0.673, + 0.115, + 0.732, + 0.127 + ], + "angle": 0, + "content": "PILORA" + }, + { + "type": "page_number", + "bbox": [ + 0.776, + 0.116, + 0.786, + 0.127 + ], + "angle": 0, + "content": "9" + }, + { + "type": "code_caption", + "bbox": [ + 0.218, + 0.147, + 0.547, + 0.162 + ], + "angle": 0, + "content": "Algorithm 1 Prototype Re-weight of Task t." + }, + { + "type": "text", + "bbox": [ + 0.217, + 0.165, + 0.787, + 0.194 + ], + "angle": 0, + "content": "Input: the number of classes \\( C_t \\), prototypes of each client \\( m_{t,k} \\), the set of mean features \\( \\mu_{t,k} \\) of each client and the temperature coefficient \\( \\eta \\)" + }, + { + "type": "text", + "bbox": [ + 0.217, + 0.194, + 0.421, + 0.207 + ], + "angle": 0, + "content": "Output: global prototype \\( m_t \\)" + }, + { + "type": "text", + "bbox": [ + 0.223, + 0.208, + 0.376, + 0.221 + ], + "angle": 0, + "content": "1: for \\( c = 1 \\rightarrow C_t \\) do" + }, + { + "type": "text", + "bbox": [ + 0.223, + 0.222, + 0.524, + 0.235 + ], + "angle": 0, + "content": "2: if \\(\\exists \\mu_{t,i,c} \\neq 0\\), where \\(i = 1 \\rightarrow K\\) then" + }, + { + "type": "text", + "bbox": [ + 0.223, + 0.236, + 0.589, + 0.249 + ], + "angle": 0, + "content": "3: \\(\\mu_{t,c}^{*}\\gets \\mu_{t,c}\\) retain all non-zero mean features" + }, + { + "type": "text", + "bbox": [ + 0.223, + 0.25, + 0.575, + 0.263 + ], + "angle": 0, + "content": "4: \\(d_{t,k,c} \\gets \\text{Distance} \\left( m_{t,k}, \\mu_{t,c}^* \\right) \\) via Eq. (11)" + }, + { + "type": "text", + "bbox": [ + 0.223, + 0.264, + 0.528, + 0.277 + ], + "angle": 0, + "content": "5: \\( p_{t,k,c} \\gets \\) take the reciprocal of \\( d_{t,k,c} \\)" + }, + { + "type": "text", + "bbox": [ + 0.223, + 0.277, + 0.585, + 0.291 + ], + "angle": 0, + "content": "6: \\(\\alpha_{t,k,c} \\gets \\text{Max-minNorm}(p_{t,k,c})\\) via Eq. (12)" + }, + { + "type": "text", + "bbox": [ + 0.223, + 0.292, + 0.563, + 0.305 + ], + "angle": 0, + "content": "7: \\(\\omega_{t,k,c} \\gets \\text{Softmax}(\\alpha_{t,k,c}, \\eta)\\) via Eq. (13)" + }, + { + "type": "text", + "bbox": [ + 0.223, + 0.306, + 0.297, + 0.316 + ], + "angle": 0, + "content": "8: else" + }, + { + "type": "text", + "bbox": [ + 0.223, + 0.318, + 0.561, + 0.332 + ], + "angle": 0, + "content": "9: \\(\\omega_{t,k,c} \\gets\\) Assign with average value \\(1 / C_t\\)" + }, + { + "type": "text", + "bbox": [ + 0.218, + 0.333, + 0.315, + 0.344 + ], + "angle": 0, + "content": "10: end if" + }, + { + "type": "text", + "bbox": [ + 0.218, + 0.346, + 0.469, + 0.36 + ], + "angle": 0, + "content": "11: \\(\\omega_{t,c}\\gets \\mathrm{List}(\\omega_{t,1,c},\\dots ,\\omega_{t,K,c})\\)" + }, + { + "type": "text", + "bbox": [ + 0.218, + 0.361, + 0.556, + 0.374 + ], + "angle": 0, + "content": "12: \\(m_{t,c}\\gets Re\\text{-weight} (\\omega_{t,c},m_{t,k})\\) via Eq. (14)" + }, + { + "type": "text", + "bbox": [ + 0.218, + 0.375, + 0.302, + 0.386 + ], + "angle": 0, + "content": "13: end for" + }, + { + "type": "text", + "bbox": [ + 0.218, + 0.388, + 0.513, + 0.401 + ], + "angle": 0, + "content": "14: Global prototype \\( m_t \\gets \\) Concat all \\( m_{t,c} \\)" + }, + { + "type": "list", + "bbox": [ + 0.218, + 0.208, + 0.589, + 0.401 + ], + "angle": 0, + "content": null + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.433, + 0.787, + 0.493 + ], + "angle": 0, + "content": "\\(m_{t,k}\\) and the set of average features for the class they have learned \\(\\mu_{t,k}\\) (zero for classes without samples). On the global server, we first compute the sum of distances between the prototype \\(m_{t,k,c}\\) and the average feature \\(\\mu_{t,i,c}\\) uploaded by all clients:" + }, + { + "type": "equation", + "bbox": [ + 0.395, + 0.493, + 0.786, + 0.533 + ], + "angle": 0, + "content": "\\[\nd _ {t, k, c} = \\sum_ {i = 1} ^ {K} \\| m _ {t, k, c} - \\mu_ {t, i, c} \\| _ {2} ^ {2}, \\tag {11}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.537, + 0.788, + 0.616 + ], + "angle": 0, + "content": "the value of \\( d_{t,k,c} \\) approximates the distance from the prototype \\( m_{t,k,c} \\) to the overall features of class \\( c \\). A smaller value indicates that the prototype is closer to the features of that class uploaded by all clients and during aggregation, we aim to assign it a higher weight. So we perform max-min normalization on the set \\( D_{t,c} = \\{p_{t,k,c}|k = 1,2,\\dots ,K\\} \\), where \\( p_{t,k,c} = \\frac{1}{d_{t,k,c}} \\):" + }, + { + "type": "equation", + "bbox": [ + 0.394, + 0.624, + 0.786, + 0.657 + ], + "angle": 0, + "content": "\\[\n\\alpha_ {t, k, c} = \\frac {p _ {t , k , c} - \\operatorname* {m i n} D _ {t , c}}{\\operatorname* {m a x} D _ {t , c} - \\operatorname* {m i n} D _ {t , c}}. \\tag {12}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.663, + 0.787, + 0.694 + ], + "angle": 0, + "content": "To meet the normalization requirements of the weights, we perform softmax processing on them and obtain the weight coefficient:" + }, + { + "type": "equation", + "bbox": [ + 0.397, + 0.702, + 0.786, + 0.739 + ], + "angle": 0, + "content": "\\[\n\\omega_ {t, k, c} = \\frac {\\exp \\left(\\eta \\cdot \\alpha_ {t , k , c}\\right)}{\\sum_ {i = 1} ^ {K} \\exp \\left(\\eta \\cdot \\alpha_ {t , i , c}\\right)}, \\tag {13}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.749, + 0.787, + 0.795 + ], + "angle": 0, + "content": "where \\(\\eta\\) is the temperature coefficient that controls the softness and hardness of weights. Finally, we re-weight all local prototypes corresponding to class \\(c\\) according to the obtained weights to get the global prototype of class \\(c\\):" + }, + { + "type": "equation", + "bbox": [ + 0.414, + 0.802, + 0.786, + 0.845 + ], + "angle": 0, + "content": "\\[\nm _ {t, c} = \\sum_ {i = 1} ^ {K} \\omega_ {t, i, c} \\cdot m _ {t, i, c}. \\tag {14}\n\\]" + } + ], + [ + { + "type": "page_number", + "bbox": [ + 0.218, + 0.116, + 0.236, + 0.127 + ], + "angle": 0, + "content": "10" + }, + { + "type": "header", + "bbox": [ + 0.272, + 0.115, + 0.361, + 0.128 + ], + "angle": 0, + "content": "H. Guo et al." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.147, + 0.788, + 0.269 + ], + "angle": 0, + "content": "We argue that prototype re-weight module allows the global prototypes \\( m_{t,c} \\) to differentially consider the data distribution information inherent in each local prototype, effectively mitigating the classifier bias issue. Compared with FedProto [51], our method further compresses the representation region of the same class in the feature space, which contributes to mitigate catastrophic forgetting. More importantly, our method does not require additional retraining of the classification layer, resulting in significant advantages in computational costs. Algorithm 1 presents the pseudo code of prototype re-weight module." + }, + { + "type": "title", + "bbox": [ + 0.215, + 0.291, + 0.533, + 0.307 + ], + "angle": 0, + "content": "4.3 Integrated Objective of PILoRA" + }, + { + "type": "text", + "bbox": [ + 0.215, + 0.316, + 0.562, + 0.332 + ], + "angle": 0, + "content": "The loss function of PILoRA can be defined as:" + }, + { + "type": "equation", + "bbox": [ + 0.399, + 0.345, + 0.786, + 0.361 + ], + "angle": 0, + "content": "\\[\nl _ {t o t a l} = l _ {d c e} + \\lambda \\cdot l _ {p l} + \\gamma \\cdot l _ {o r t}, \\tag {15}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.372, + 0.788, + 0.418 + ], + "angle": 0, + "content": "where \\(\\lambda\\) and \\(\\gamma\\) are two hyper-parameters, and the overall framework of PILoRA is shown in Fig. 2. In general, our method is simple and effective, showing strong performance in different non-IID settings and degrees of data heterogeneity." + }, + { + "type": "title", + "bbox": [ + 0.215, + 0.441, + 0.376, + 0.459 + ], + "angle": 0, + "content": "5 Experiments" + }, + { + "type": "title", + "bbox": [ + 0.215, + 0.473, + 0.436, + 0.489 + ], + "angle": 0, + "content": "5.1 Experimental Setups" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.498, + 0.789, + 0.649 + ], + "angle": 0, + "content": "Benchmark: To evaluate the proposed PILoRA, we perform our experiments on two well-known datasets: CIFAR-100 [24] and TinyImageNet [25], We also test the performance of the model on large-scale datasets, specifically, we randomly select 200 classes from the ImageNet-1k [8] as a new dataset. Following the protocols proposed in [33], we split 10 incremental stages and only the data from the current stage is available. Besides, to challenge our method, the local dataset of each client is followed by two kinds of non-IID settings: quantity-based label imbalance and distribution-based label imbalance [27] and we denote the degree of heterogeneity of the two settings by \\(\\alpha\\) and \\(\\beta\\). Details of both settings can be seen in Section A in the Supplementary Material." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.65, + 0.788, + 0.77 + ], + "angle": 0, + "content": "Comparison: We compare our method with existing FCIL methods4: TARGET [61], GLFC [10], LGA [9]. we also adopt several CIL methods: EWC [23], LwF [29], iCaRL [46], L2P [55] and FL method FedNCM [26] in FCIL setting. In addition, we compare the use of cross-entropy loss for optimization during training and the use of class means as classifiers during inference, which we named FedCLM. We explore their performance in different non-IID settings and degrees of data heterogeneity. For a fair comparison, we tune all methods to the same pre-trained model as ours and fine-tune them using LoRA." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.771, + 0.788, + 0.817 + ], + "angle": 0, + "content": "Implementation: Considering privacy issues, we evaluate the performance of our method on self-supervised pre-trained weights (Dino [5]) for ViT-B/16 [11], this setting is also widely used in CIL task [36,37,60]. Considering the trade-off" + }, + { + "type": "page_footnote", + "bbox": [ + 0.218, + 0.825, + 0.753, + 0.841 + ], + "angle": 0, + "content": "4 GLFC and LGA only perform experiments on quantity-based label imbalance." + } + ], + [ + { + "type": "header", + "bbox": [ + 0.673, + 0.115, + 0.732, + 0.127 + ], + "angle": 0, + "content": "PILORA" + }, + { + "type": "page_number", + "bbox": [ + 0.769, + 0.116, + 0.785, + 0.127 + ], + "angle": 0, + "content": "11" + }, + { + "type": "table_caption", + "bbox": [ + 0.216, + 0.145, + 0.785, + 0.174 + ], + "angle": 0, + "content": "Table 1: Results (%) on CIFAR-100. Results are included for 10 tasks (10 classes per task) and under different degree of two non-IID setting." + }, + { + "type": "table", + "bbox": [ + 0.234, + 0.185, + 0.773, + 0.387 + ], + "angle": 0, + "content": "
Non-IIDQuantity-based label imbalanceDistribution-based label imbalance
Partitionα = 6α = 4α = 2β = 0.5β = 0.1β = 0.05
MethodsANAvg.ANAvg.ANAvg.ANAvg.ANAvg.ANAvg.
Joint88.6-84.3-79.8-90.1-87.8-85.9-
EWC+FL57.969.155.966.842.252.765.577.857.873.243.559.2
LwF+FL57.468.855.166.740.852.964.777.554.663.345.764.5
iCaRL+FL35.856.537.158.943.455.351.367.750.165.944.663.0
L2P+FL63.465.159.058.22.65.653.951.662.971.438.732.2
FedCLM58.969.457.667.644.357.966.577.461.071.848.863.5
FedNCM65.674.461.971.149.659.866.877.962.172.450.965.9
TARGET60.971.358.869.545.256.566.177.860.571.151.865.3
GLFC58.270.453.765.913.137.7------
LGA64.573.661.170.521.640.9------
Ours69.578.665.174.454.962.668.578.163.473.754.867.1
Ours+HT69.678.565.874.856.564.370.278.663.673.857.969.2
" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.414, + 0.788, + 0.538 + ], + "angle": 0, + "content": "between performance and number of parameters, we insert the LoRA module into only the first block (See Section B.1) and set \\( r = 4 \\). We train our models using Adam [21] with a batch size of 64 and followed by [60], we adopt different learning rates for prototypes layer of \\( 2e^{-3} \\) and LoRA parameters of \\( 1e^{-5} \\) on CIFAR-100, and \\( 5e^{-3} / 5e^{-6} \\) on TinyImageNet. Moreover, cosine annealing is also used in training processes. We set \\( \\delta = 1 \\), \\( \\lambda = 0.001 \\), \\( \\gamma = 0.5 \\) and \\( \\eta = 0.2 \\). We initialize 10 local clients to train and upload the parameters at each communication round. The local training epoch is 5 and the communication round is 30." + }, + { + "type": "title", + "bbox": [ + 0.216, + 0.558, + 0.436, + 0.574 + ], + "angle": 0, + "content": "5.2 Comparative Results" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.584, + 0.788, + 0.765 + ], + "angle": 0, + "content": "We report \\( A_{N}(\\uparrow) \\) and \\( \\mathrm{Avg}(\\uparrow) \\) to evaluate the performance of the methods, where \\( A_{N} \\) is the accuracy of all seen classes in the final task and Avg. is calculated as the average accuracy of all tasks. Results are shown in Tab. 1 and Tab. 2, we can obverse that our method outperforms other comparison methods and demonstrates strong robustness across different non-IID settings. Among FCIL methods, TARGET achieves relatively better performance under different data heterogeneity, while LGA and GLFC show a steep drop in performance and lack of robustness under extreme data heterogeneity. Notably, FedNCM in the FCIL setting even outperforms TARGET in some cases, which suggests that existing FCIL methods lack the handling of data heterogeneity, whereas our approach dramatically improves the performance of the model under different data heterogeneity through a lightweight prototype re-weight module." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.766, + 0.788, + 0.842 + ], + "angle": 0, + "content": "The CIL methods also suffer from data heterogeneity in the FCIL setting, with L2P+FL being the most obvious. We believe that there are two main reasons for this, firstly the performance of L2P relies on supervised pre-trained weights (e.g., ImageNet-1k [8]) and its performance degrades significantly when using self-supervised pre-trained weights [60]. Secondly, L2P needs to choose the" + } + ], + [ + { + "type": "page_number", + "bbox": [ + 0.218, + 0.116, + 0.236, + 0.127 + ], + "angle": 0, + "content": "12" + }, + { + "type": "header", + "bbox": [ + 0.272, + 0.115, + 0.361, + 0.127 + ], + "angle": 0, + "content": "H. Guo et al." + }, + { + "type": "table_caption", + "bbox": [ + 0.216, + 0.145, + 0.788, + 0.174 + ], + "angle": 0, + "content": "Table 2: Results (\\%) on TinyImageNet. Results are included for 10 tasks (20 classes per task) and under different degree of two non-IID setting." + }, + { + "type": "table", + "bbox": [ + 0.23, + 0.185, + 0.779, + 0.347 + ], + "angle": 0, + "content": "
Non-IIDQuantity-based label imbalanceDistribution-based label imbalance
Partitionα = 12α = 8α = 4β = 0.5β = 0.1β = 0.05
MethodsANAvg.ANAvg.ANAvg.ANAvg.ANAvg.ANAvg.
Joint83.6-82.9-80.2-84.3-83.3-82.8-
L2P+FL61.658.049.439.38.210.264.266.956.352.551.943.2
FedCLM61.672.451.860.345.856.966.577.460.471.046.757.8
FedNCM71.681.669.579.457.264.773.781.670.880.468.478.0
TARGET72.681.670.379.663.873.571.680.971.080.169.379.1
GLFC69.177.961.373.525.139.4------
LGA71.379.465.875.336.748.8------
Ours74.481.074.380.970.177.874.580.974.381.073.680.2
Ours+HT75.081.374.981.267.975.674.781.074.380.973.480.1
" + }, + { + "type": "image", + "bbox": [ + 0.229, + 0.365, + 0.382, + 0.462 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.384, + 0.365, + 0.538, + 0.462 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.374, + 0.468, + 0.394, + 0.48 + ], + "angle": 0, + "content": "(a)" + }, + { + "type": "image", + "bbox": [ + 0.566, + 0.37, + 0.776, + 0.463 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.661, + 0.468, + 0.682, + 0.48 + ], + "angle": 0, + "content": "(b)" + }, + { + "type": "image_caption", + "bbox": [ + 0.214, + 0.491, + 0.789, + 0.548 + ], + "angle": 0, + "content": "Fig. 3: (a) Comparison of real proportions of the same class among different clients and the weights obtained from Prototype Re-weighting during a local training process. (b) The \\( L2 \\) distance between classes feature extracted by different clients and global model with corresponding prototypes." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.581, + 0.789, + 0.748 + ], + "angle": 0, + "content": "appropriate prompt for embedding in the model based on the similarity calculation. However, in cases of extreme heterogeneity, there exists a significant disparity in what each client learns, even at the same stage. This discrepancy prevents the selection of a uniform prompt for reasoning. Instead, our method bypasses redundant similarity calculations or knowledge distillation [29] and effectively mitigates catastrophic forgetting directly through the summation of LoRA parameters. In addition, we combine the HeadTune in FedNCM with our prototype classifier, and the performance of our model is further improved by the better initialization provided by HeadTune. This suggests that there is still a lot of space for improvement in the existing methods in solving the problem of data heterogeneity in FCIL." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.75, + 0.789, + 0.841 + ], + "angle": 0, + "content": "Analysis of Prototype Re-weight: The core idea of our proposed prototype re-weight module is to aggregate prototypes in a manner that closely resembles the true data distribution among clients without disclosing local data information. In Fig. 3a, we show the real proportions of the same class of data across clients and we can observe that the weights calculated through re-weight module align with the distribution of data among different clients. Hence, the global" + } + ], + [ + { + "type": "header", + "bbox": [ + 0.673, + 0.115, + 0.732, + 0.127 + ], + "angle": 0, + "content": "PILORA" + }, + { + "type": "page_number", + "bbox": [ + 0.769, + 0.116, + 0.786, + 0.127 + ], + "angle": 0, + "content": "13" + }, + { + "type": "table_caption", + "bbox": [ + 0.214, + 0.145, + 0.788, + 0.189 + ], + "angle": 0, + "content": "Table 3: Ablation Results (\\%) on CIFAR-100. Experimental results based on self-supervised pre-trained weights. \\( F_N(\\downarrow) \\) is the average forgetting and the calculation process can be referred to [66]." + }, + { + "type": "table", + "bbox": [ + 0.32, + 0.199, + 0.691, + 0.297 + ], + "angle": 0, + "content": "
Partitionα = 6β = 0.5
MethodsANAvg.FNANAvg.FN
Ours-w/o PR26.830.95.541.448.89.3
Ours-w/o lort68.777.910.367.177.011.1
Ours-w/o LoRA (fintune)11.329.57.313.329.98.9
Ours-w/o LoRA (frozen)68.677.710.067.877.211.6
Ours69.578.69.568.578.110.6
" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.325, + 0.788, + 0.492 + ], + "angle": 0, + "content": "prototypes effectively retain all learned prototype information. In contrast, other methods typically use direct averaging for the classification layer, leading to the fusion of a considerable amount of irrelevant information in the classifier when the data heterogeneity is high, consequently causing classifier bias. In Fig. 3b, we compare the distances between the prototype calculated by both prototype re-weight and average for the global model and the deep features of the test data. It can be seen that our proposed prototype re-weight method effectively 'takes the best' of all client uploaded prototypes, so that the global prototype is better adapted to the features of the corresponding class, whereas on average the correct prototype is far away from the features of the corresponding class, especially when the data is very heterogeneous." + }, + { + "type": "title", + "bbox": [ + 0.215, + 0.513, + 0.39, + 0.529 + ], + "angle": 0, + "content": "5.3 Ablation Study" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.538, + 0.788, + 0.749 + ], + "angle": 0, + "content": "To evaluate the effect of each component in PILoRA, we perform the ablation study and show the results of 10 phases set in CIFAR-100 in Tab. 3. We can observe that without prototype re-weight (PR), The classification accuracy \\(A_{N}\\) and Avg. of the model decrease significantly, which is due to the fact that the global prototype obtained by simply averaging at this point does not represent the information of each class well, which leads to classifier bias. Instead, our proposed method significantly improves the performance of the model by heuristically reweight the local prototypes. When there is no orthogonal regularization \\((l_{ort})\\), all the metrics of the model show a certain decrease, which indicates that there is a partial overlap between the parameter spaces of different incremental tasks. By imposing orthogonal regularization the parameter spaces corresponding to different input spaces can be made to train the model in the direction of orthogonality to each other, which further improves the model performance. We visualize the cosine similarity between LoRA in Section B.2." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.75, + 0.788, + 0.841 + ], + "angle": 0, + "content": "To better demonstrate the contribution of LoRA, we compare the effects of fine-tuning the entire backbone and freezing the entire backbone under self-supervised pre-trained weights when \\(\\alpha = 6\\) and \\(\\beta = 0.5\\). As can be seen in Tab. 3, when fine-tuning the entire backbone, the performance of the model is severely degraded, we believe that this is due to the fact that the model parameters of ViT are too large to achieve a balance between the ability to remember old" + } + ], + [ + { + "type": "page_number", + "bbox": [ + 0.218, + 0.116, + 0.236, + 0.127 + ], + "angle": 0, + "content": "14" + }, + { + "type": "header", + "bbox": [ + 0.272, + 0.115, + 0.361, + 0.128 + ], + "angle": 0, + "content": "H. Guo et al." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.147, + 0.788, + 0.253 + ], + "angle": 0, + "content": "classes and the ability to discriminate new classes under the constraint of knowledge distillation. Freezing the entire backbone network can retain discriminative ability for old classes, but at this point, the model has too few trainable parameters, with only prototypes used for classification, thereby limiting the expressive capacity of the model. Therefore, to achieve a trade-off between communication cost and model performance, we fine-tune and achieve the best performance with a tiny number of parameters through Incremental LoRA." + }, + { + "type": "title", + "bbox": [ + 0.216, + 0.272, + 0.402, + 0.287 + ], + "angle": 0, + "content": "5.4 Further Analysis" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.294, + 0.487, + 0.532 + ], + "angle": 0, + "content": "Memory usage analysis. For our PILoRA, in addition to the model for each client, we store the LoRA parameters \\( A_{q}^{1:t - 1} \\) and \\( A_{v}^{1:t - 1} \\) up to the current stage \\( t \\) to compute the orthogonal regularization. Whereas TARGET requires an additional generator to generate old samples, as well as additional memory space to store the generated images, LGA and GLFC similarly take up additional space to store old class samples. Compared to them, our method takes up very little memory space, storing on average only \\( 0.04\\% \\) of the parameters equivalent to the ViT-base." + }, + { + "type": "image", + "bbox": [ + 0.499, + 0.327, + 0.639, + 0.432 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.645, + 0.327, + 0.783, + 0.432 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.497, + 0.447, + 0.788, + 0.489 + ], + "angle": 0, + "content": "Fig.4: The impact of different number of \\(K\\) on CIFAR-100, where we consider \\(\\alpha = 6\\) and \\(\\beta = 0.5\\)." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.535, + 0.788, + 0.643 + ], + "angle": 0, + "content": "Increase the number of local clients \\((K)\\). As depicted in Fig. 4, we investigate the performance of the model as the number of local clients increase, by respectively setting \\( K = \\{10, 15, 20\\} \\). From the results shown in Fig. 4, we observe that the model's performance slightly declines as the number of clients increases. We believe that under the same non-IID setting, enlarging the number of clients further exacerbates the heterogeneity among clients, consequently impacting the model's performance. More results can be seen in Section B.3." + }, + { + "type": "title", + "bbox": [ + 0.216, + 0.661, + 0.36, + 0.677 + ], + "angle": 0, + "content": "6 Conclusion" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.689, + 0.788, + 0.841 + ], + "angle": 0, + "content": "In this paper, we propose a simple and effective method of \\( PILoRA \\) for FCIL. PILoRA is based on pre-trained ViT models and fine-tunes with a tiny number of parameters using LoRA. To address the catastrophic forgetting problem in FCIL, we propose incremental LoRA, which can efficiently combine different incremental tasks by summing the orthogonal LoRA parameter space; To deal with the classifier bias caused by data heterogeneity, we adopt prototype learning and propose prototype re-weight, which utilize heuristic information between prototypes and features to perform weighted aggregation of global prototypes. Experimental results show that our method achieves state-of-the-art results on standard datasets and maintains robustness under extreme data heterogeneity." + } + ], + [ + { + "type": "header", + "bbox": [ + 0.673, + 0.115, + 0.732, + 0.127 + ], + "angle": 0, + "content": "PILORA" + }, + { + "type": "page_number", + "bbox": [ + 0.769, + 0.116, + 0.786, + 0.127 + ], + "angle": 0, + "content": "15" + }, + { + "type": "title", + "bbox": [ + 0.218, + 0.145, + 0.403, + 0.163 + ], + "angle": 0, + "content": "Acknowledgements" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.178, + 0.788, + 0.254 + ], + "angle": 0, + "content": "This work has been supported by the National Science and Technology Major Project (2022ZD0116500), National Natural Science Foundation of China (U20A20223, 62222609, 62076236), CAS Project for Young Scientists in Basic Research (YSBR-083), and Key Research Program of Frontier Sciences of CAS (ZDBS-LY-7004) and the InnoHK program." + }, + { + "type": "title", + "bbox": [ + 0.217, + 0.279, + 0.323, + 0.294 + ], + "angle": 0, + "content": "References" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.311, + 0.785, + 0.354 + ], + "angle": 0, + "content": "1. Bagwe, G., Yuan, X., Pan, M., Zhang, L.: Fed-cprompt: Contrastive prompt for rehearsal-free federated continual learning. In: Federated Learning and Analytics in Practice: Algorithms, Systems, Applications, and Opportunities (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.355, + 0.785, + 0.395 + ], + "angle": 0, + "content": "2. Belouadah, E., Popescu, A.: Deesil: Deep-shallow incremental learning. In: Proceedings of the European Conference on Computer Vision (ECCV) Workshops. pp. 0-0 (2018)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.396, + 0.785, + 0.436 + ], + "angle": 0, + "content": "3. Belouadah, E., Popescu, A.: II2m: Class incremental learning with dual memory. In: Proceedings of the IEEE/CVF international conference on computer vision. pp. 583-592 (2019)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.437, + 0.785, + 0.478 + ], + "angle": 0, + "content": "4. Brown, T., Mann, B., Ryder, N., Subbiah, M., Kaplan, J.D., Dhariwal, P., Neelakantan, A., Shyam, P., Sastry, G., Askell, A., et al.: Language models are few-shot learners. Advances in neural information processing systems 33, 1877-1901 (2020)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.479, + 0.785, + 0.52 + ], + "angle": 0, + "content": "5. Caron, M., Touvron, H., Misra, I., Jégou, H., Mairal, J., Bojanowski, P., Joulin, A.: Emerging properties in self-supervised vision transformers. In: Proceedings of the International Conference on Computer Vision (ICCV) (2021)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.521, + 0.785, + 0.548 + ], + "angle": 0, + "content": "6. Cheng, Z., Zhang, X.Y., Liu, C.L.: Unified classification and rejection: A one-versus-all framework. arXiv preprint arXiv:2311.13355 (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.549, + 0.785, + 0.576 + ], + "angle": 0, + "content": "7. Chitale, R., Vaidya, A., Kane, A., Ghotkar, A.: Task arithmetic with lora for continual learning. arXiv preprint arXiv:2311.02428 (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.577, + 0.785, + 0.618 + ], + "angle": 0, + "content": "8. Deng, J., Dong, W., Socher, R., Li, L.J., Li, K., Fei-Fei, L.: Imagenet: A large-scale hierarchical image database. In: 2009 IEEE conference on computer vision and pattern recognition. pp. 248-255. IEEE (2009)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.619, + 0.785, + 0.659 + ], + "angle": 0, + "content": "9. Dong, J., Cong, Y., Sun, G., Zhang, Y., Schiele, B., Dai, D.: No one left behind: Real-world federated class-incremental learning. arXiv preprint arXiv:2302.00903 (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.66, + 0.785, + 0.701 + ], + "angle": 0, + "content": "0. Dong, J., Wang, L., Fang, Z., Sun, G., Xu, S., Wang, X., Zhu, Q.: Federated class incremental learning. In: Proceedings of the IEEE/CVF conference on computer vision and pattern recognition. pp. 10164-10173 (2022)" + }, + { + "type": "ref_text", + "bbox": [ + 0.22, + 0.702, + 0.785, + 0.757 + ], + "angle": 0, + "content": "11. Dosovitskiy, A., Beyer, L., Kolesnikov, A., Weissenborn, D., Zhai, X., Unterthiner, T., Dehghani, M., Minderer, M., Heigold, G., Gelly, S., et al.: An image is worth 16x16 words: Transformers for image recognition at scale. arXiv preprint arXiv:2010.11929 (2020)" + }, + { + "type": "ref_text", + "bbox": [ + 0.22, + 0.758, + 0.785, + 0.799 + ], + "angle": 0, + "content": "12. Gao, Q., Zhao, C., Sun, Y., Xi, T., Zhang, G., Ghanem, B., Zhang, J.: A unified continual learning framework with general parameter-efficient tuning. arXiv preprint arXiv:2303.10070 (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.22, + 0.8, + 0.785, + 0.84 + ], + "angle": 0, + "content": "13. Geng, C., Huang, S.j., Chen, S.: Recent advances in open set recognition: A survey. IEEE transactions on pattern analysis and machine intelligence 43(10), 3614-3631 (2020)" + }, + { + "type": "list", + "bbox": [ + 0.22, + 0.311, + 0.785, + 0.84 + ], + "angle": 0, + "content": null + } + ], + [ + { + "type": "page_number", + "bbox": [ + 0.218, + 0.116, + 0.236, + 0.127 + ], + "angle": 0, + "content": "16" + }, + { + "type": "header", + "bbox": [ + 0.272, + 0.115, + 0.361, + 0.127 + ], + "angle": 0, + "content": "H. Guo et al." + }, + { + "type": "ref_text", + "bbox": [ + 0.218, + 0.148, + 0.787, + 0.203 + ], + "angle": 0, + "content": "14. Giuseppi, A., Manfredi, S., Pietrabissa, A.: A weighted average consensus approach for decentralized federated learning. Machine Intelligence Research 19(4), 319-330 (2022). https://doi.org/10.1007/s11633-022-1338-z, https://www.mi-research.net/en/article/doi/10.1007/s11633-022-1338-z" + }, + { + "type": "ref_text", + "bbox": [ + 0.218, + 0.204, + 0.787, + 0.245 + ], + "angle": 0, + "content": "15. Guo, T., Guo, S., Wang, J., Tang, X., Xu, W.: Promptfl: Let federated participants cooperatively learn prompts instead of models-federated learning in age of foundation model. IEEE Transactions on Mobile Computing (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.218, + 0.246, + 0.786, + 0.286 + ], + "angle": 0, + "content": "16. Houlsby, N., Giurgiu, A., Jastrzebski, S., Morrone, B., De Laroussilhe, Q., Gesmundo, A., Attariyan, M., Gelly, S.: Parameter-efficient transfer learning for nlp. In: International Conference on Machine Learning. pp. 2790-2799. PMLR (2019)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.287, + 0.787, + 0.327 + ], + "angle": 0, + "content": "17. Hu, E.J., Shen, Y., Wallis, P., Allen-Zhu, Z., Li, Y., Wang, S., Wang, L., Chen, W.: Lora: Low-rank adaptation of large language models. arXiv preprint arXiv:2106.09685 (2021)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.328, + 0.787, + 0.37 + ], + "angle": 0, + "content": "18. Ilharco, G., Ribeiro, M.T., Wortsman, M., Gururangan, S., Schmidt, L., Hajishirzi, H., Farhadi, A.: Editing models with task arithmetic. arXiv preprint arXiv:2212.04089 (2022)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.371, + 0.787, + 0.412 + ], + "angle": 0, + "content": "19. Karimireddy, S.P., Kale, S., Mohri, M., Reddi, S., Stich, S., Suresh, A.T.: Scaffold: Stochastic controlled averaging for federated learning. In: International conference on machine learning. pp. 5132-5143. PMLR (2020)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.412, + 0.787, + 0.453 + ], + "angle": 0, + "content": "20. Khan, L.U., Saad, W., Han, Z., Hossain, E., Hong, C.S.: Federated learning for internet of things: Recent advances, taxonomy, and open challenges. IEEE Communications Surveys & Tutorials 23(3), 1759-1799 (2021)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.453, + 0.787, + 0.479 + ], + "angle": 0, + "content": "21. Kingma, D.P., Ba, J.: Adam: A method for stochastic optimization. arXiv preprint arXiv:1412.6980 (2014)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.48, + 0.787, + 0.522 + ], + "angle": 0, + "content": "22. Kirillov, A., Mintun, E., Ravi, N., Mao, H., Rolland, C., Gustafson, L., Xiao, T., Whitehead, S., Berg, A.C., Lo, W.Y., et al.: Segment anything. arXiv preprint arXiv:2304.02643 (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.523, + 0.787, + 0.577 + ], + "angle": 0, + "content": "23. Kirkpatrick, J., Pascanu, R., Rabinowitz, N., Veness, J., Desjardins, G., Rusu, A.A., Milan, K., Quan, J., Ramalho, T., Grabska-Barwinska, A., et al.: Overcoming catastrophic forgetting in neural networks. Proceedings of the national academy of sciences 114(13), 3521-3526 (2017)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.578, + 0.787, + 0.604 + ], + "angle": 0, + "content": "24. Krizhevsky, A., Hinton, G., et al.: Learning multiple layers of features from tiny images (2009)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.605, + 0.787, + 0.633 + ], + "angle": 0, + "content": "25. Le, Y., Yang, X.: Tiny imagenet visual recognition challenge. CS 231N 7(7), 3 (2015)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.634, + 0.787, + 0.674 + ], + "angle": 0, + "content": "26. Legate, G., Bernier, N., Page-Caccia, L., Oyallon, E., Belilovsky, E.: Guiding the last layer in federated learning with pre-trained models. Advances in Neural Information Processing Systems 36 (2024)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.675, + 0.787, + 0.715 + ], + "angle": 0, + "content": "27. Li, Q., Diao, Y., Chen, Q., He, B.: Federated learning on non-iid data silos: An experimental study. In: 2022 IEEE 38th International Conference on Data Engineering (ICDE). pp. 965-978. IEEE (2022)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.716, + 0.787, + 0.757 + ], + "angle": 0, + "content": "28. Li, T., Sahu, A.K., Zaheer, M., Sanjabi, M., Talwalkar, A., Smith, V.: Federated optimization in heterogeneous networks. Proceedings of Machine learning and systems 2, 429-450 (2020)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.758, + 0.787, + 0.785 + ], + "angle": 0, + "content": "29. Li, Z., Hoiem, D.: Learning without forgetting. IEEE transactions on pattern analysis and machine intelligence 40(12), 2935-2947 (2017)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.786, + 0.787, + 0.84 + ], + "angle": 0, + "content": "30. Lin, Y.M., Gao, Y., Gong, M.G., Zhang, S.J., Zhang, Y.Q., Li, Z.Y.: Federated learning on multimodal data: A comprehensive survey. Machine Intelligence Research 20(4), 539-553 (2023). https://doi.org/10.1007/s11633-022-1398-0, https://www.mi-research.net/en/article/doi/10.1007/s11633-022-1398-0" + }, + { + "type": "list", + "bbox": [ + 0.217, + 0.148, + 0.787, + 0.84 + ], + "angle": 0, + "content": null + } + ], + [ + { + "type": "header", + "bbox": [ + 0.674, + 0.115, + 0.732, + 0.127 + ], + "angle": 0, + "content": "PILoRA" + }, + { + "type": "page_number", + "bbox": [ + 0.769, + 0.116, + 0.786, + 0.127 + ], + "angle": 0, + "content": "17" + }, + { + "type": "ref_text", + "bbox": [ + 0.214, + 0.147, + 0.788, + 0.189 + ], + "angle": 0, + "content": "31. Liu, C.L., Sako, H., Fujisawa, H.: Discriminative learning quadratic discriminant function for handwriting recognition. IEEE Transactions on Neural Networks 15(2), 430-444 (2004)" + }, + { + "type": "ref_text", + "bbox": [ + 0.214, + 0.19, + 0.787, + 0.232 + ], + "angle": 0, + "content": "32. Liu, C.L., Sako, H., Fujisawa, H.: Effects of classifier structures and training regimes on integrated segmentation and recognition of handwritten numeral strings. IEEE transactions on pattern analysis and machine intelligence 26(11), 1395-1407 (2004)" + }, + { + "type": "ref_text", + "bbox": [ + 0.214, + 0.232, + 0.787, + 0.272 + ], + "angle": 0, + "content": "33. Liu, C., Qu, X., Wang, J., Xiao, J.: Fedet: A communication-efficient federated class-incremental learning framework based on enhanced transformer. arXiv preprint arXiv:2306.15347 (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.214, + 0.273, + 0.787, + 0.314 + ], + "angle": 0, + "content": "34. Liu, P., Yuan, W., Fu, J., Jiang, Z., Hayashi, H., Neubig, G.: Pre-train, prompt, and predict: A systematic survey of prompting methods in natural language processing. ACM Computing Surveys 55(9), 1-35 (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.214, + 0.315, + 0.787, + 0.355 + ], + "angle": 0, + "content": "35. Liu, W., Wu, X., Zhu, F., Yu, M., Wang, C., Liu, C.L.: Class incremental learning with self-supervised pre-training and prototype learning. arXiv preprint arXiv:2308.02346 (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.214, + 0.356, + 0.787, + 0.384 + ], + "angle": 0, + "content": "36. Liu, W., Zhu, F., Liu, C.L.: Branch-tuning: Balancing stability and plasticity for continual self-supervised learning. arXiv preprint arXiv:2403.18266 (2024)" + }, + { + "type": "ref_text", + "bbox": [ + 0.214, + 0.384, + 0.787, + 0.412 + ], + "angle": 0, + "content": "37. Liu, W., Zhu, F., Liu, C.L.: Towards non-exemplar semi-supervised class-incremental learning. arXiv preprint arXiv:2403.18291 (2024)" + }, + { + "type": "ref_text", + "bbox": [ + 0.214, + 0.412, + 0.787, + 0.453 + ], + "angle": 0, + "content": "38. Luo, M., Chen, F., Hu, D., Zhang, Y., Liang, J., Feng, J.: No fear of heterogeneity: Classifier calibration for federated learning with non-iid data. Advances in Neural Information Processing Systems 34, 5972-5984 (2021)" + }, + { + "type": "ref_text", + "bbox": [ + 0.214, + 0.453, + 0.787, + 0.494 + ], + "angle": 0, + "content": "39. McCloskey, M., Cohen, N.J.: Catastrophic interference in connectionist networks: The sequential learning problem. In: Psychology of learning and motivation, vol. 24, pp. 109-165. Elsevier (1989)" + }, + { + "type": "ref_text", + "bbox": [ + 0.214, + 0.495, + 0.787, + 0.536 + ], + "angle": 0, + "content": "40. McMahan, B., Moore, E., Ramage, D., Hampson, S., y Arcas, B.A.: Communication-efficient learning of deep networks from decentralized data. In: Artificial intelligence and statistics. pp. 1273-1282. PMLR (2017)" + }, + { + "type": "ref_text", + "bbox": [ + 0.214, + 0.536, + 0.787, + 0.577 + ], + "angle": 0, + "content": "41. Niknam, S., Dhillon, H.S., Reed, J.H.: Federated learning for wireless communications: Motivation, opportunities, and challenges. IEEE Communications Magazine 58(6), 46-51 (2020)" + }, + { + "type": "ref_text", + "bbox": [ + 0.214, + 0.578, + 0.787, + 0.619 + ], + "angle": 0, + "content": "42. Ortiz-Jimenez, G., Favero, A., Frossard, P.: Task arithmetic in the tangent space: Improved editing of pre-trained models. Advances in Neural Information Processing Systems 36 (2024)" + }, + { + "type": "ref_text", + "bbox": [ + 0.214, + 0.62, + 0.787, + 0.646 + ], + "angle": 0, + "content": "43. Qi, D., Zhao, H., Li, S.: Better generative replay for continual federated learning. In: The Eleventh International Conference on Learning Representations (2022)" + }, + { + "type": "ref_text", + "bbox": [ + 0.214, + 0.647, + 0.787, + 0.702 + ], + "angle": 0, + "content": "44. Qu, L., Zhou, Y., Liang, P.P., Xia, Y., Wang, F., Adeli, E., Fei-Fei, L., Rubin, D.: Rethinking architecture design for tackling data heterogeneity in federated learning. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 10061-10071 (2022)" + }, + { + "type": "ref_text", + "bbox": [ + 0.214, + 0.702, + 0.787, + 0.757 + ], + "angle": 0, + "content": "45. Radford, A., Kim, J.W., Hallacy, C., Ramesh, A., Goh, G., Agarwal, S., Sastry, G., Askell, A., Mishkin, P., Clark, J., et al.: Learning transferable visual models from natural language supervision. In: International conference on machine learning. pp. 8748-8763. PMLR (2021)" + }, + { + "type": "ref_text", + "bbox": [ + 0.214, + 0.758, + 0.787, + 0.799 + ], + "angle": 0, + "content": "46. Rebuffi, S.A., Kolesnikov, A., Sperl, G., Lampert, C.H.: icarl: Incremental classifier and representation learning. In: Proceedings of the IEEE conference on Computer Vision and Pattern Recognition. pp. 2001-2010 (2017)" + }, + { + "type": "ref_text", + "bbox": [ + 0.214, + 0.799, + 0.787, + 0.841 + ], + "angle": 0, + "content": "47. Rieke, N., Hancox, J., Li, W., Miletari, F., Roth, H.R., Albarqouni, S., Bakas, S., Galtier, M.N., Landman, B.A., Maier-Hein, K., et al.: The future of digital health with federated learning. NPJ digital medicine 3(1), 119 (2020)" + }, + { + "type": "list", + "bbox": [ + 0.214, + 0.147, + 0.788, + 0.841 + ], + "angle": 0, + "content": null + } + ], + [ + { + "type": "page_number", + "bbox": [ + 0.218, + 0.116, + 0.236, + 0.127 + ], + "angle": 0, + "content": "18" + }, + { + "type": "header", + "bbox": [ + 0.272, + 0.115, + 0.361, + 0.127 + ], + "angle": 0, + "content": "H. Guo et al." + }, + { + "type": "ref_text", + "bbox": [ + 0.218, + 0.148, + 0.785, + 0.189 + ], + "angle": 0, + "content": "48. Shang, X., Lu, Y., Huang, G., Wang, H.: Federated learning on heterogeneous and long-tailed data via classifier re-training with federated features. arXiv preprint arXiv:2204.13399 (2022)" + }, + { + "type": "ref_text", + "bbox": [ + 0.218, + 0.19, + 0.785, + 0.232 + ], + "angle": 0, + "content": "49. Shenaj, D., Toldo, M., Rigon, A., Zanuttigh, P.: Asynchronous federated continual learning. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 5054-5062 (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.232, + 0.785, + 0.3 + ], + "angle": 0, + "content": "50. Smith, J.S., Karlinsky, L., Gutta, V., Cascante-Bonilla, P., Kim, D., Arbelle, A., Panda, R., Feris, R., Kira, Z.: Coda-prompt: Continual decomposed attention-based prompting for rehearsal-free continual learning. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 11909-11919 (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.3, + 0.785, + 0.342 + ], + "angle": 0, + "content": "51. Tan, Y., Long, G., Liu, L., Zhou, T., Lu, Q., Jiang, J., Zhang, C.: Fedproto: Federated prototype learning across heterogeneous clients. In: Proceedings of the AAAI Conference on Artificial Intelligence. vol. 36, pp. 8432-8440 (2022)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.342, + 0.785, + 0.383 + ], + "angle": 0, + "content": "52. Voigt, P., Von dem Bussche, A.: The eu general data protection regulation (gdpr). A Practical Guide, 1st Ed., Cham: Springer International Publishing 10(3152676), 10-5555 (2017)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.383, + 0.785, + 0.425 + ], + "angle": 0, + "content": "53. Wang, X., Chen, T., Ge, Q., Xia, H., Bao, R., Zheng, R., Zhang, Q., Gui, T., Huang, X.: Orthogonal subspace learning for language model continual learning. arXiv preprint arXiv:2310.14152 (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.425, + 0.785, + 0.48 + ], + "angle": 0, + "content": "54. Wang, Z., Zhang, Z., Ebrahimi, S., Sun, R., Zhang, H., Lee, C.Y., Ren, X., Su, G., Perot, V., Dy, J., et al.: Dualprompt: Complementary prompting for rehearsal-free continual learning. In: European Conference on Computer Vision. pp. 631-648. Springer (2022)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.481, + 0.785, + 0.535 + ], + "angle": 0, + "content": "55. Wang, Z., Zhang, Z., Lee, C.Y., Zhang, H., Sun, R., Ren, X., Su, G., Perot, V., Dy, J., Pfister, T.: Learning to prompt for continual learning. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 139-149 (2022)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.535, + 0.785, + 0.577 + ], + "angle": 0, + "content": "56. Yang, H.M., Zhang, X.Y., Yin, F., Liu, C.L.: Robust classification with convolutional prototype learning. In: Proceedings of the IEEE conference on computer vision and pattern recognition. pp. 3474-3482 (2018)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.577, + 0.785, + 0.619 + ], + "angle": 0, + "content": "57. Yang, H.M., Zhang, X.Y., Yin, F., Yang, Q., Liu, C.L.: Convolutional prototype network for open set recognition. IEEE Transactions on Pattern Analysis and Machine Intelligence 44(5), 2358-2370 (2020)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.619, + 0.785, + 0.647 + ], + "angle": 0, + "content": "58. Yi, L., Yu, H., Wang, G., Liu, X.: Fedlora: Model-heterogeneous personalized federated learning with lora tuning. arXiv preprint arXiv:2310.13283 (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.647, + 0.785, + 0.688 + ], + "angle": 0, + "content": "59. Yoon, J., Jeong, W., Lee, G., Yang, E., Hwang, S.J.: Federated continual learning with weighted inter-client transfer. In: International Conference on Machine Learning. pp. 12073-12086. PMLR (2021)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.688, + 0.785, + 0.73 + ], + "angle": 0, + "content": "60. Zhang, G., Wang, L., Kang, G., Chen, L., Wei, Y.: Slow learner with classifier alignment for continual learning on a pre-trained model. arXiv preprint arXiv:2303.05118 (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.73, + 0.785, + 0.758 + ], + "angle": 0, + "content": "61. Zhang, J., Chen, C., Zhuang, W., Lv, L.: Addressing catastrophic forgetting in federated class-continual learning. arXiv preprint arXiv:2303.06937 (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.758, + 0.785, + 0.799 + ], + "angle": 0, + "content": "62. Zhang, J., Chen, C., Zhuang, W., Lyu, L.: Target: Federated class-continual learning via exemplar-free distillation. In: Proceedings of the IEEE/CVF International Conference on Computer Vision. pp. 4782-4793 (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.799, + 0.785, + 0.84 + ], + "angle": 0, + "content": "63. Zhao, H., Du, W., Li, F., Li, P., Liu, G.: Reduce communication costs and preserve privacy: Prompt tuning method in federated learning. arXiv preprint arXiv:2208.12268 (2022)" + }, + { + "type": "list", + "bbox": [ + 0.217, + 0.148, + 0.785, + 0.84 + ], + "angle": 0, + "content": null + } + ], + [ + { + "type": "header", + "bbox": [ + 0.673, + 0.115, + 0.732, + 0.127 + ], + "angle": 0, + "content": "PILORA" + }, + { + "type": "page_number", + "bbox": [ + 0.769, + 0.116, + 0.786, + 0.127 + ], + "angle": 0, + "content": "19" + }, + { + "type": "ref_text", + "bbox": [ + 0.214, + 0.147, + 0.788, + 0.189 + ], + "angle": 0, + "content": "64. Zhu, F., Cheng, Z., Zhang, X.y., Liu, C.l.: Class-incremental learning via dual augmentation. Advances in Neural Information Processing Systems 34, 14306-14318 (2021)" + }, + { + "type": "ref_text", + "bbox": [ + 0.214, + 0.19, + 0.787, + 0.218 + ], + "angle": 0, + "content": "65. Zhu, F., Ma, S., Cheng, Z., Zhang, X.Y., Zhang, Z., Liu, C.L.: Open-world machine learning: A review and new outlooks. arXiv preprint arXiv:2403.01759 (2024)" + }, + { + "type": "ref_text", + "bbox": [ + 0.216, + 0.218, + 0.787, + 0.259 + ], + "angle": 0, + "content": "66. Zhu, F., Zhang, X.Y., Wang, C., Yin, F., Liu, C.L.: Prototype augmentation and self-supervision for incremental learning. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 5871-5880 (2021)" + }, + { + "type": "list", + "bbox": [ + 0.214, + 0.147, + 0.788, + 0.259 + ], + "angle": 0, + "content": null + } + ] +] \ No newline at end of file diff --git a/2024/PILoRA_ Prototype Guided Incremental LoRA for Federated Class-Incremental Learning/b10fddc1-55da-4bf4-bfbe-8037f4bdbf91_origin.pdf b/2024/PILoRA_ Prototype Guided Incremental LoRA for Federated Class-Incremental Learning/b10fddc1-55da-4bf4-bfbe-8037f4bdbf91_origin.pdf new file mode 100644 index 0000000000000000000000000000000000000000..7a17e1693db16044debfad573b6a9f6cab79c2f7 --- /dev/null +++ b/2024/PILoRA_ Prototype Guided Incremental LoRA for Federated Class-Incremental Learning/b10fddc1-55da-4bf4-bfbe-8037f4bdbf91_origin.pdf @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:e715bc800f4e2530cff767c0900f32a26774f24ae8905188758baa91cfee259d +size 3345042 diff --git a/2024/PILoRA_ Prototype Guided Incremental LoRA for Federated Class-Incremental Learning/full.md b/2024/PILoRA_ Prototype Guided Incremental LoRA for Federated Class-Incremental Learning/full.md new file mode 100644 index 0000000000000000000000000000000000000000..e565dea2b212abfc0cd8b0e642f9ce7849e4e366 --- /dev/null +++ b/2024/PILoRA_ Prototype Guided Incremental LoRA for Federated Class-Incremental Learning/full.md @@ -0,0 +1,343 @@ +# PILoRA: Prototype Guided Incremental LoRA for Federated Class-Incremental Learning + +Haiyang Guo $^{1,2}$ , Fei Zhu $^{3}$ , Wenzhuo Liu $^{1,2}$ , Xu-Yao Zhang $^{1,2\star}$ , and Cheng-Lin Liu $^{1,2}$ + +$^{1}$ MAIS, Institute of Automation, Chinese Academy of Sciences + +$^{2}$ School of Artificial Intelligence, University of Chinese Academy of Sciences + +3 Centre for Artificial Intelligence and Robotics, Hong Kong Institute of Science and Innovation, Chinese Academy of Sciences + +{guohaiyang2023, zhufei2018}@ia.ac.cn, liuwenzhuo20@mails.ucas.ac.cn, xyz, liucl}@nlpr.ia.ac.cn + +Abstract. Existing federated learning methods have effectively dealt with decentralized learning in scenarios involving data privacy and non-IID data. However, in real-world situations, each client dynamically learns new classes, requiring the global model to classify all seen classes. To effectively mitigate catastrophic forgetting and data heterogeneity under low communication costs, we propose a simple and effective method named PILoRA. On the one hand, we adopt prototype learning to learn better feature representations and leverage the heuristic information between prototypes and class features to design a prototype re-weight module to solve the classifier bias caused by data heterogeneity without retraining the classifier. On the other hand, we view incremental learning as the process of learning distinct task vectors and encoding them within different LoRA parameters. Accordingly, we propose Incremental LoRA to mitigate catastrophic forgetting. Experimental results on standard datasets indicate that our method outperforms the state-of-the-art approaches significantly. More importantly, our method exhibits strong robustness and superiority in different settings and degrees of data heterogeneity. The code is available at https://github.com/GhyO501/PILORA. + +Keywords: Federated Learning $\cdot$ Class Incremental Learning + +# 1 Introduction + +Federated learning (FL) [40] is a novel distributed machine learning paradigm that enables multiple data owners to collaboratively train a shared model while ensuring the privacy of their local data. In recent years, with the increasing emphasis on data privacy in society and the refinement of relevant regulations [52], FL has experienced rapid development and has been widely applied in various real-world scenarios [14,20,30,41,47]. + +Existing FL methods [19, 28, 40] typically rely on the closed-world assumption [6, 13, 65], that is, the number of classes seen by the model remains the same and unchanged during both the training and testing stages. However, the real world is dynamic and constantly changing, the local clients often need to receive new data to update the global model continuously. Therefore, Federated Class Incremental Learning (FCIL) [10, 59] has been proposed to handle FL tasks in dynamic scenarios. Specifically, each local client can only update the model with new class data at each stage and upload the model parameters to the global server for aggregation, while the global model needs to maintain its discriminative ability for all seen classes. Moreover, the data distribution of local data follows the non-independent and identically distributed (non-IID) assumption [27]. + +FCIL presents a more realistic setting for real-world applications, but it also introduces greater challenges, as FCIL needs to simultaneously address the issue of catastrophic forgetting [23, 39] and data heterogeneity resulting from non-IID. In existing studies, one way is to store a subset of data from old classes and train them together when learning new tasks [9, 10], but the amount of old data stored is strictly limited due to privacy protection requirements. An alternative way is to utilize generative models to generate pseudo-samples of old data [43, 61], thus retaining the ability to classify the old classes. However, training a good generator will cause greater computational overhead, and the generator also suffers from catastrophic forgetting. Another direction is to use pre-trained models for fine-tuning [1, 33], these methods store different stages of knowledge by maintaining a pool of modules and inserting the corresponding modules at inference time based on the similarity between the input data and the modules. However, storing these modules takes up additional memory space. Although they provide different perspectives for solving FCIL, there is still a large unexplored space, especially the performance of the model under different non-IID settings and different degrees of data heterogeneity. So in order to better address this issue, we start by addressing a fundamental question, what is the panacea for dealing with CIL and FL? + +In other words, if we can find common ground in dealing with CIL and FL, it will be of great help in solving the FCIL problem. Specifically, we observe that: 1) At the level of feature representation, both FL and CIL tasks require models to learn a intra-class compact and inter-class separable feature representation. On the one hand, for the CIL task, this kind of feature representation helps to reduce the overlap between features of the old and new classes in the deep feature space, thus alleviating the forgetting of old knowledge. For this reason, PASS [66] introduces self-supervised learning to assist the learning of feature representations. On the other hand, in the FL task, FedEx [51] imposes a constraint on the deep features of samples belonging to the same class. This constraint ensures that these features are proximate to the global prototypes of their respective classes. 2) At the classifier level, classifier drift is the common enemy of both. In CIL, when learning new classes, the decision boundaries learned from old classes can be severely disrupted, resulting in a significant bias + +in the classification layer [64, 66]. To address this issue, some methods [2, 3, 46] alleviate bias by directly retaining a portion of the data from previous categories and training the model with new data. Other methods [64, 66] address classifier bias by retaining pseudo features of old classes to assist in training the classification layer; In FL, CCVR [38] assesses the similarity among different layers within local models under data heterogeneity and reveals that classifiers exhibit the lowest similarity. This suggests that the classifier of each client models is seriously biased to the local data. Similarly to CIL, [38, 48] utilize the statistical information of local features to retrain the classifier on the global server to alleviate classifier bias. + +Motivated by the above findings, in this paper, we propose a Prototype Guided Incremental LoRA (PILoRA) model to tackle the FCIL problem. Specifically, we adopt prototype classifiers in our model to learn intraclass compact and inter-class separable feature representations, which is beneficial for both CIL and FL [28, 35, 56, 57]. We use the pre-trained Transformer model (ViT) as the backbone because the global interactions learned by Transformer are significantly more robust than the local patterns learned by CNN for FL tasks [44] (Fig. 1.), and also provide a good representation of the features. Considering the communication cost, we freeze the entire backbone network during training and use LoRA, a Parameter-Efficient Fine-Tuning (PEFT) method to train the model. To address the catastrophic forgetting problem in FCIL, we propose incremental LoRA, which constrains different stages of LoRA to be trained on + +subspaces orthogonal to each other via orthogonal regularization, and acquires knowledge learned in the past by simple and efficient summation at inference time. To address the classifier bias caused by data heterogeneity, we design a prototype re-weight module for the global server, which aggregates local prototypes based on heuristic information between prototypes and corresponding class features. Compared to other existing FCIL methods, our model achieves superior results on standard datasets. Furthermore, we explore the model's performance under different non-IID settings and degrees of data heterogeneity. Experiments demonstrate the robustness of our approach, while other methods suffer from significant declines. + +![](images/83ed409fef50ff66a6ccc5e13482f17907bf577f515fdbefe7df7ab37cbb14ac.jpg) +Fig.1: The local data of client 1 and client 2 are non-IID. Resnet50 (the first row) focuses more on local patterns, and the patterns learned are significantly different in the case of data heterogeneity, causing the average model to lose some important information (e.g., fish fins). However, ViT (the second row) is less affected by data heterogeneity, and the averaged model basically retains all the information learned by the local model. + +The main contributions of this paper can be summarized as: + +- We propose incremental LoRA, which performs incremental learning on orthogonal subspaces to mitigate catastrophic forgetting and aggregates previous knowledge through simple and efficient parameter summation. +- We propose a prototype re-weight module to form global prototypes using heuristic information between each class of local prototypes and the corresponding class of features. Our method effectively addresses classifier bias induced by data heterogeneity without retraining. +- We conduct extensive experiments on standard datasets and achieve state-of-the-art performance. Furthermore, in the extreme heterogeneous case, our method still maintains robustness, while all other methods suffer plummet. + +# 2 Related Work + +Federated Class-Incremental Learning: Research on FCIL has garnered considerable attention in recent years. Dong et al. [10] first introduces the concept of FCIL and proposes several loss functions on the local side and the global server side to alleviate local catastrophic forgetting and global catastrophic forgetting. However, their method uses a rehearsal buffer to store and retain old class data, and additionally design a proxy server to select the best model, resulting in large memory overhead and communication costs. LGA [9] extends the work of [10], but it still belongs to rehearsal-based FCIL. In the more challenging rehearsal-free FCIL problem, generate models are widely adopted to produce synthetic data that aim to mitigate the catastrophic forgetting on local side and global side [43,62], but its performance is highly dependent on the quality of synthetic data and will incur additional computational costs. Similar to our work, FedSpace [49] designs prototype-based loss to encourage feature vectors of the same class to be close together, while we achieve this goal by using prototype classifier. In contrast to the methods described above, some recent studies [1,33] combine pre-trained models with FCIL and achieve higher performance at a smaller communication cost. However, they both adopt a similarity-based selection strategy, which causes additional memory overheads in inference. Besides, they all use supervised pre-training weights, whereas we argue that this may pose privacy concerns because data from downstream tasks may overlap with pre-trained datasets. + +PEFT for Pre-Trained Model: With the emergence of large-scale pre-trained models [4, 22, 45], how to effectively fine-tune these models to adapt the downstream tasks has been a focal point of attention. Recently, LoRA [17], Prompt [34], and Adapter [16] have emerged as standout techniques and have been widely used in CIL [12, 50, 54, 55] and FL [15, 58, 63] tasks. In FCIL, existing methods [1, 33] have attempted to combine Prompt and Adapter with pre-trained models. Specifically, they store the knowledge learned at each stage in the parameters of the Prompt or Adapter module and select the appropriate module for the current input to be embedded into the model by a specific similarity computation during the inference time, thus effectively mitigating catastrophic + +forgetting with a tiny communication costs. However, such similarity matching based approaches undoubtedly introduce inference delays since they require additional similarity computation modules. (In [33], they even train a separate CNN to compute the similarity.) Furthermore, they need a additional memory space in the global server to set up a parameter pool to store the module parameters. + +# 3 Preliminaries + +In FCIL setting, each client has a local stream dataset $D_{k} = \{D_{k}^{t}\}_{t=1}^{T}$ , where $D_{k}^{t} = \{\mathbf{X}_{k}^{t}, \mathbf{Y}_{k}^{t}\} = \{x_{k,i}^{t}, y_{k,i}^{t}\}_{i=1}^{N_{t}}$ is the dataset of the $k$ -th client on task $t$ . Dataset $D_{k}^{t}$ contains $N_{k}^{t}$ training samples and their label $\mathbf{Y}_{k}^{t} \in C_{k}^{t}$ , where $C_{k}^{t}$ is the class set of the $k$ -th client in task $t$ . In particular, the distribution of different clients $k$ under the same task is non-IID and the class sets of different tasks $t$ are disjoint. For local client, the objective is to minimize a pre-defined loss function $\mathcal{L}$ on current dataset $D_{k}^{t}$ , while avoiding interference with and possibly enhancing the knowledge acquired from previous learning stages: + +$$ +\underset {\boldsymbol {\omega} _ {k} ^ {t}} {\operatorname {a r g m i n}} \mathcal {L} \left(\boldsymbol {\omega} _ {k} ^ {t}; \boldsymbol {\omega} ^ {t - 1}, \mathbf {X} _ {k} ^ {t}, \mathbf {Y} _ {k} ^ {t}\right), \tag {1} +$$ + +where $\omega_{k}^{t}$ is the parameters of $k$ -th local model, and $\omega^{t-1}$ is the global model at previous task. Then, the server updates the global model $\omega^{t}$ by aggregating all uploaded parameters as follows: + +$$ +\boldsymbol {\omega} ^ {t} = \sum_ {k = 1} ^ {K} \gamma_ {k} \boldsymbol {\omega} _ {k} ^ {t}, \text {w h e r e} \gamma_ {k} = \frac {N _ {k} ^ {t}}{\sum_ {k ^ {\prime}} N _ {k ^ {\prime}} ^ {t}}. \tag {2} +$$ + +The global model aims to correctly classify the test samples of all seen classes and solve the problem of data heterogeneity at low communication cost. + +# 4 Our method + +# 4.1 Incremental LoRA for Pre-Trained Model + +In FCIL, storing the knowledge of different stages in different modules through PEFT can effectively mitigate the catastrophic forgetting [1,33], but additional similarity computation units are required in the inference time in order to select the appropriate module to embed into the model based on the inputs, which leads to inference delay and additional memory overhead. In order to learn an end-to-end global model without taking up extra storage space, we intuitively believe that modules storing knowledge from different stages can be organically combined into one module that embeds knowledge from all stages. Accordingly, we propose Incremental LoRA to effectively address catastrophic forgetting in FCIL. On the one hand, LoRA has the natural advantages of low inference latency and more stable training [17], on the other hand, Inspired by [53, 55], we + +![](images/7e038bf45d14de35b69d245ef807388ab0f5880b01e46155eb88bf8431464b4a.jpg) +Fig. 2: Illustration of PILoRA for FCIL. The client fine-tunes LoRA and prototypes for each class using a local dataset. Upon upload, the global server aggregates the LoRA uploaded by different clients, and applies a prototype re-weight module before re-sending them to each client. + +introduce the orthogonality loss to constrain LoRA to learn new knowledge in a subspace orthogonal to past tasks, thus better preserving the knowledge of the old classes. + +Specifically, we define the initialization parameters of the pre-training model as $\mathbf{W} \in \mathbb{R}^{d \times k}$ , and $\Delta \mathbf{W}_{\mathbf{t}}$ represents the parameter to be updated for task $t$ . Then, the model's updates in different incremental learning stages can be expressed as: + +$$ +\mathbf {W} + \Delta \mathbf {W} _ {\mathbf {t}}. \tag {3} +$$ + +LoRA assumes that the weight changes of large-scale pre-trained models when adapting to downstream tasks occur in a low-rank space: + +$$ +\mathbf {W} + \Delta \mathbf {W} _ {\mathbf {t}} = \mathbf {W} + \mathbf {A} _ {t} \mathbf {B} _ {t}, \tag {4} +$$ + +where $\mathbf{A}_t \in \mathbb{R}^{d \times r}$ , $\mathbf{B}_t \in \mathbb{R}^{r \times k}$ , and $r \ll \min\{d, k\}$ . At first, $\mathbf{A}_t$ is initialized by a random Gaussian, while $\mathbf{B}_t$ is initialized with zero. Therefore, $\mathbf{B}_t$ can be regarded as the coefficient matrix of $\mathbf{A}_t$ [53]. LoRA applies it as a bypass to query and value projection matrices in multi-head attention module, and during adaptation to downstream tasks, only the parameters of $\mathbf{A}_t$ and $\mathbf{B}_t$ are trainable. + +Inspired by [53], we argue that the parameters of LoRA can be viewed as containers that store different subspaces of task gradients. Thus, learning a series of incremental learning tasks can be viewed as learning a series of LoRA parameters. So how to use these LoRA modules to build a model that can classify all seen classes? An idea is to select appropriate modules to be embedded in the + +model based on the similarity between different modules and the corresponding inputs [1, 33]. However, computing similarity requires additional computational resources, which leads to delayed inference. Another idea is that since LoRA is essentially two weight matrices, it is possible to concat all LoRA parameters prior to stage $t$ into a new matrix to gain knowledge of all previous stages [53]. In particular, at stage $t$ , the knowledge learned in the previous $t - 1$ stages are stored in $\mathbf{A}_{1:t - 1} = [\mathbf{A}_1,\dots ,\mathbf{A}_{t - 1}]$ and $\mathbf{B}_{1:t - 1} = [\mathbf{B}_1,\dots ,\mathbf{B}_{t - 1}]$ . To obtain the weight matrix, we can sequentially concat these LoRA parameters into a bigger module: + +$$ +\mathbf {W} + \Delta \mathbf {W} _ {\mathbf {t}} = \mathbf {W} + \widetilde {\mathbf {A}} _ {t} \widetilde {\mathbf {B}} _ {t}, \tag {5} +$$ + +where $\widetilde{\mathbf{A}}_t = \mathrm{concat}([\mathbf{A}_1,\ldots ,\mathbf{A}_t])$ , $\widetilde{\mathbf{B}}_t = \mathrm{concat}([\mathbf{B}_1,\ldots ,\mathbf{B}_t])$ . However, this will cause the parameters of LoRA in the global model to increase with the number of incremental learning tasks. Therefore, in order to ensure consistency between the local model and the global model, we are inspired by Task Arithmetic [7, 18, 42] in model edit and propose to integrate the LoRA parameters at different stages by summation: + +$$ +\mathbf {W} + \Delta \mathbf {W} _ {\mathbf {t}} = \mathbf {W} + \overline {{\mathbf {A}}} _ {t} \overline {{\mathbf {B}}} _ {t}, \text {w h e r e} \overline {{\mathbf {A}}} _ {t} = \sum_ {i = 1} ^ {t} \mathbf {A} _ {i}, \overline {{\mathbf {B}}} _ {t} = \sum_ {i = 1} ^ {t} \mathbf {B} _ {i}. \tag {6} +$$ + +Specifically, distinct directions in the weight space correspond to various localized regions in the input space [42]. Consequently, a linear combination of these diverse directions encoded within the pre-trained weights enables the model to effectively discriminate between different inputs. In our method, The training data for different incremental tasks are disjoint from each other, so it is reasonable to directly sum the LoRA parameters corresponding to different tasks. Besides, Task Arithmetic also point out that orthogonality between different task parameter vectors helps to better integrate different task. So we propose a orthogonal regularization to achieve this by constraining the parameters of LoRA to be orthogonal to previous tasks: + +$$ +l _ {o r t} \left(\mathbf {A} _ {i}, \mathbf {A} _ {t}\right) = \sum_ {i = 1} ^ {t - 1} \left| \mathbf {A} _ {i} ^ {T} \cdot \mathbf {A} _ {t} \right|. \tag {7} +$$ + +Meanwhile, the orthogonal regularization encourages the learning of distinct tasks along orthogonal directions, thereby effectively reducing the spatial overlap between them. This approach will be proved to be advantageous in mitigating catastrophic forgetting, as it helps preserve the knowledge learned from previous tasks while adapting to new ones. + +# 4.2 Prototype learning with Prototype Re-weight + +Prototype Learning: As mentioned before, a feature representation that is separable intra-class and compact inter-class is helpful for FCIL tasks. Therefore, the model not only needs to correctly classify known classes but also needs + +to model the distribution of known classes in the feature space. In open set recognition, CPN [57] designs a discriminative loss and generative loss for prototype learning to constrain the range of known classes in the feature space, thus reserving space for samples from unknown classes. Inspired by it, we introduce prototype learning in FCIL. In particular, we set a prototype $m = \{m_i | i = 1, 2, \dots, C\}$ for each class, where $m_i \in \mathbb{R}^d$ , and the dimensions $d$ of each prototype are the same as the dimensions of the final deep feature space. + +To shorten the distance between the class feature and the corresponding prototype, we apply the distance-based cross entropy (DCE) discriminative loss. Given a sample $(x,y)$ , DCE uses the distance between sample features $f_{\theta}(x)$ and prototype $m_{i}$ to represent the probability of belonging to class $i$ . Considering the normalization of probability, DCE adopts the softmax operation: + +$$ +p (x \in m _ {i} | x) = \frac {\exp (- \delta \cdot \| f _ {\theta} (x) - m _ {i} \| _ {2} ^ {2})}{\sum_ {j = 1} ^ {C} \exp (- \delta \cdot \| f _ {\theta} (x) - m _ {j} \| _ {2} ^ {2})}, \tag {8} +$$ + +where $\| f_{\theta}(x) - m_i\| _2^2$ is the Euclidean distance between the feature $f_{\theta}(x)$ of the input sample and the prototype $m_{i}$ , and $\delta$ is the temperature scalar controlling the hardness of the class distribution. Hence, the distance-based-cross-entropy loss can be defined as: + +$$ +l _ {d c e} ((x, y); \theta , m) = - \log (x \in m _ {y} | x). \tag {9} +$$ + +By minimizing the DCE loss, the distance between the sample's feature and the correct prototype will be smaller than other incorrect prototypes. However, features learned only under the discriminative loss may not be compact, which may lead to an overlap in feature representations between new and old classes. To solve this problem, we introduce prototype learning (PL) loss [57]: + +$$ +l _ {p l} ((x, y); \theta , m) = \| f _ {\theta} (x) - m _ {y} \| _ {2} ^ {2}, \tag {10} +$$ + +PL loss decreases the distance between sample features and the corresponding correct prototype, making the model learn a more compact intra-class distribution. Essentially, PL loss is a maximum likelihood regularization of features $f_{\theta}(x)$ under the Gaussian mixture density assumption [31, 32]. + +Prototype Re-weight: Considering a specific class $c$ , under the non-IID setting, each client $k$ holds a portion of the training data $N_{c,k}$ and $\sum_{k=1}^{K} N_{c,k} = N_{c}$ , where $N_{c}$ is the sum of training samples of class $c$ . In prototype learning, the prototype $m_{k,c}$ learned by each client can effectively reflect the distribution of class $c$ within its local data. That is, for clients that do not have training samples of class $c$ , the distance between their corresponding prototype and the features of class $c$ will be greater than for clients that do have training samples. If we give equal weight to these uploaded prototypes during parameter aggregation, classifier drift may occur, especially when the data heterogeneity is high, potentially resulting in the global model losing discriminative capability for that class. + +Therefore, we use the heuristic information contained in the distance between the prototype and the corresponding class average features to design the prototype re-weight module. Specifically, at stage $t$ , each client $k$ uploads prototypes + +Algorithm 1 Prototype Re-weight of Task t. + +Input: the number of classes $C_t$ , prototypes of each client $m_{t,k}$ , the set of mean features $\mu_{t,k}$ of each client and the temperature coefficient $\eta$ + +Output: global prototype $m_t$ + +1: for $c = 1 \rightarrow C_t$ do +2: if $\exists \mu_{t,i,c} \neq 0$ , where $i = 1 \rightarrow K$ then +3: $\mu_{t,c}^{*}\gets \mu_{t,c}$ retain all non-zero mean features +4: $d_{t,k,c} \gets \text{Distance} \left( m_{t,k}, \mu_{t,c}^* \right)$ via Eq. (11) +5: $p_{t,k,c} \gets$ take the reciprocal of $d_{t,k,c}$ +6: $\alpha_{t,k,c} \gets \text{Max-minNorm}(p_{t,k,c})$ via Eq. (12) +7: $\omega_{t,k,c} \gets \text{Softmax}(\alpha_{t,k,c}, \eta)$ via Eq. (13) +8: else +9: $\omega_{t,k,c} \gets$ Assign with average value $1 / C_t$ +10: end if +11: $\omega_{t,c}\gets \mathrm{List}(\omega_{t,1,c},\dots ,\omega_{t,K,c})$ +12: $m_{t,c}\gets Re\text{-weight} (\omega_{t,c},m_{t,k})$ via Eq. (14) +13: end for +14: Global prototype $m_t \gets$ Concat all $m_{t,c}$ + +$m_{t,k}$ and the set of average features for the class they have learned $\mu_{t,k}$ (zero for classes without samples). On the global server, we first compute the sum of distances between the prototype $m_{t,k,c}$ and the average feature $\mu_{t,i,c}$ uploaded by all clients: + +$$ +d _ {t, k, c} = \sum_ {i = 1} ^ {K} \| m _ {t, k, c} - \mu_ {t, i, c} \| _ {2} ^ {2}, \tag {11} +$$ + +the value of $d_{t,k,c}$ approximates the distance from the prototype $m_{t,k,c}$ to the overall features of class $c$ . A smaller value indicates that the prototype is closer to the features of that class uploaded by all clients and during aggregation, we aim to assign it a higher weight. So we perform max-min normalization on the set $D_{t,c} = \{p_{t,k,c}|k = 1,2,\dots ,K\}$ , where $p_{t,k,c} = \frac{1}{d_{t,k,c}}$ : + +$$ +\alpha_ {t, k, c} = \frac {p _ {t , k , c} - \operatorname* {m i n} D _ {t , c}}{\operatorname* {m a x} D _ {t , c} - \operatorname* {m i n} D _ {t , c}}. \tag {12} +$$ + +To meet the normalization requirements of the weights, we perform softmax processing on them and obtain the weight coefficient: + +$$ +\omega_ {t, k, c} = \frac {\exp \left(\eta \cdot \alpha_ {t , k , c}\right)}{\sum_ {i = 1} ^ {K} \exp \left(\eta \cdot \alpha_ {t , i , c}\right)}, \tag {13} +$$ + +where $\eta$ is the temperature coefficient that controls the softness and hardness of weights. Finally, we re-weight all local prototypes corresponding to class $c$ according to the obtained weights to get the global prototype of class $c$ : + +$$ +m _ {t, c} = \sum_ {i = 1} ^ {K} \omega_ {t, i, c} \cdot m _ {t, i, c}. \tag {14} +$$ + +We argue that prototype re-weight module allows the global prototypes $m_{t,c}$ to differentially consider the data distribution information inherent in each local prototype, effectively mitigating the classifier bias issue. Compared with FedProto [51], our method further compresses the representation region of the same class in the feature space, which contributes to mitigate catastrophic forgetting. More importantly, our method does not require additional retraining of the classification layer, resulting in significant advantages in computational costs. Algorithm 1 presents the pseudo code of prototype re-weight module. + +# 4.3 Integrated Objective of PILoRA + +The loss function of PILoRA can be defined as: + +$$ +l _ {t o t a l} = l _ {d c e} + \lambda \cdot l _ {p l} + \gamma \cdot l _ {o r t}, \tag {15} +$$ + +where $\lambda$ and $\gamma$ are two hyper-parameters, and the overall framework of PILoRA is shown in Fig. 2. In general, our method is simple and effective, showing strong performance in different non-IID settings and degrees of data heterogeneity. + +# 5 Experiments + +# 5.1 Experimental Setups + +Benchmark: To evaluate the proposed PILoRA, we perform our experiments on two well-known datasets: CIFAR-100 [24] and TinyImageNet [25], We also test the performance of the model on large-scale datasets, specifically, we randomly select 200 classes from the ImageNet-1k [8] as a new dataset. Following the protocols proposed in [33], we split 10 incremental stages and only the data from the current stage is available. Besides, to challenge our method, the local dataset of each client is followed by two kinds of non-IID settings: quantity-based label imbalance and distribution-based label imbalance [27] and we denote the degree of heterogeneity of the two settings by $\alpha$ and $\beta$ . Details of both settings can be seen in Section A in the Supplementary Material. + +Comparison: We compare our method with existing FCIL methods4: TARGET [61], GLFC [10], LGA [9]. we also adopt several CIL methods: EWC [23], LwF [29], iCaRL [46], L2P [55] and FL method FedNCM [26] in FCIL setting. In addition, we compare the use of cross-entropy loss for optimization during training and the use of class means as classifiers during inference, which we named FedCLM. We explore their performance in different non-IID settings and degrees of data heterogeneity. For a fair comparison, we tune all methods to the same pre-trained model as ours and fine-tune them using LoRA. + +Implementation: Considering privacy issues, we evaluate the performance of our method on self-supervised pre-trained weights (Dino [5]) for ViT-B/16 [11], this setting is also widely used in CIL task [36,37,60]. Considering the trade-off + +Table 1: Results (%) on CIFAR-100. Results are included for 10 tasks (10 classes per task) and under different degree of two non-IID setting. + +
Non-IIDQuantity-based label imbalanceDistribution-based label imbalance
Partitionα = 6α = 4α = 2β = 0.5β = 0.1β = 0.05
MethodsANAvg.ANAvg.ANAvg.ANAvg.ANAvg.ANAvg.
Joint88.6-84.3-79.8-90.1-87.8-85.9-
EWC+FL57.969.155.966.842.252.765.577.857.873.243.559.2
LwF+FL57.468.855.166.740.852.964.777.554.663.345.764.5
iCaRL+FL35.856.537.158.943.455.351.367.750.165.944.663.0
L2P+FL63.465.159.058.22.65.653.951.662.971.438.732.2
FedCLM58.969.457.667.644.357.966.577.461.071.848.863.5
FedNCM65.674.461.971.149.659.866.877.962.172.450.965.9
TARGET60.971.358.869.545.256.566.177.860.571.151.865.3
GLFC58.270.453.765.913.137.7------
LGA64.573.661.170.521.640.9------
Ours69.578.665.174.454.962.668.578.163.473.754.867.1
Ours+HT69.678.565.874.856.564.370.278.663.673.857.969.2
+ +between performance and number of parameters, we insert the LoRA module into only the first block (See Section B.1) and set $r = 4$ . We train our models using Adam [21] with a batch size of 64 and followed by [60], we adopt different learning rates for prototypes layer of $2e^{-3}$ and LoRA parameters of $1e^{-5}$ on CIFAR-100, and $5e^{-3} / 5e^{-6}$ on TinyImageNet. Moreover, cosine annealing is also used in training processes. We set $\delta = 1$ , $\lambda = 0.001$ , $\gamma = 0.5$ and $\eta = 0.2$ . We initialize 10 local clients to train and upload the parameters at each communication round. The local training epoch is 5 and the communication round is 30. + +# 5.2 Comparative Results + +We report $A_{N}(\uparrow)$ and $\mathrm{Avg}(\uparrow)$ to evaluate the performance of the methods, where $A_{N}$ is the accuracy of all seen classes in the final task and Avg. is calculated as the average accuracy of all tasks. Results are shown in Tab. 1 and Tab. 2, we can obverse that our method outperforms other comparison methods and demonstrates strong robustness across different non-IID settings. Among FCIL methods, TARGET achieves relatively better performance under different data heterogeneity, while LGA and GLFC show a steep drop in performance and lack of robustness under extreme data heterogeneity. Notably, FedNCM in the FCIL setting even outperforms TARGET in some cases, which suggests that existing FCIL methods lack the handling of data heterogeneity, whereas our approach dramatically improves the performance of the model under different data heterogeneity through a lightweight prototype re-weight module. + +The CIL methods also suffer from data heterogeneity in the FCIL setting, with L2P+FL being the most obvious. We believe that there are two main reasons for this, firstly the performance of L2P relies on supervised pre-trained weights (e.g., ImageNet-1k [8]) and its performance degrades significantly when using self-supervised pre-trained weights [60]. Secondly, L2P needs to choose the + +Table 2: Results (\%) on TinyImageNet. Results are included for 10 tasks (20 classes per task) and under different degree of two non-IID setting. + +
Non-IIDQuantity-based label imbalanceDistribution-based label imbalance
Partitionα = 12α = 8α = 4β = 0.5β = 0.1β = 0.05
MethodsANAvg.ANAvg.ANAvg.ANAvg.ANAvg.ANAvg.
Joint83.6-82.9-80.2-84.3-83.3-82.8-
L2P+FL61.658.049.439.38.210.264.266.956.352.551.943.2
FedCLM61.672.451.860.345.856.966.577.460.471.046.757.8
FedNCM71.681.669.579.457.264.773.781.670.880.468.478.0
TARGET72.681.670.379.663.873.571.680.971.080.169.379.1
GLFC69.177.961.373.525.139.4------
LGA71.379.465.875.336.748.8------
Ours74.481.074.380.970.177.874.580.974.381.073.680.2
Ours+HT75.081.374.981.267.975.674.781.074.380.973.480.1
+ +![](images/6c9ab0bf0ffc2e44905efff2593d0aeb5534d72e962d2d5bd3c8839ecf6c8885.jpg) +(a) + +![](images/ccf7e08d3b18cfc5db266ea983653c1b9f6c324715f101ce31d773aa9c456466.jpg) + +![](images/50c43133f5895e13f792703c80e5366ab30e0bb8b19af7e025358fffc3f14dcc.jpg) +(b) +Fig. 3: (a) Comparison of real proportions of the same class among different clients and the weights obtained from Prototype Re-weighting during a local training process. (b) The $L2$ distance between classes feature extracted by different clients and global model with corresponding prototypes. + +appropriate prompt for embedding in the model based on the similarity calculation. However, in cases of extreme heterogeneity, there exists a significant disparity in what each client learns, even at the same stage. This discrepancy prevents the selection of a uniform prompt for reasoning. Instead, our method bypasses redundant similarity calculations or knowledge distillation [29] and effectively mitigates catastrophic forgetting directly through the summation of LoRA parameters. In addition, we combine the HeadTune in FedNCM with our prototype classifier, and the performance of our model is further improved by the better initialization provided by HeadTune. This suggests that there is still a lot of space for improvement in the existing methods in solving the problem of data heterogeneity in FCIL. + +Analysis of Prototype Re-weight: The core idea of our proposed prototype re-weight module is to aggregate prototypes in a manner that closely resembles the true data distribution among clients without disclosing local data information. In Fig. 3a, we show the real proportions of the same class of data across clients and we can observe that the weights calculated through re-weight module align with the distribution of data among different clients. Hence, the global + +Table 3: Ablation Results (\%) on CIFAR-100. Experimental results based on self-supervised pre-trained weights. $F_N(\downarrow)$ is the average forgetting and the calculation process can be referred to [66]. + +
Partitionα = 6β = 0.5
MethodsANAvg.FNANAvg.FN
Ours-w/o PR26.830.95.541.448.89.3
Ours-w/o lort68.777.910.367.177.011.1
Ours-w/o LoRA (fintune)11.329.57.313.329.98.9
Ours-w/o LoRA (frozen)68.677.710.067.877.211.6
Ours69.578.69.568.578.110.6
+ +prototypes effectively retain all learned prototype information. In contrast, other methods typically use direct averaging for the classification layer, leading to the fusion of a considerable amount of irrelevant information in the classifier when the data heterogeneity is high, consequently causing classifier bias. In Fig. 3b, we compare the distances between the prototype calculated by both prototype re-weight and average for the global model and the deep features of the test data. It can be seen that our proposed prototype re-weight method effectively 'takes the best' of all client uploaded prototypes, so that the global prototype is better adapted to the features of the corresponding class, whereas on average the correct prototype is far away from the features of the corresponding class, especially when the data is very heterogeneous. + +# 5.3 Ablation Study + +To evaluate the effect of each component in PILoRA, we perform the ablation study and show the results of 10 phases set in CIFAR-100 in Tab. 3. We can observe that without prototype re-weight (PR), The classification accuracy $A_{N}$ and Avg. of the model decrease significantly, which is due to the fact that the global prototype obtained by simply averaging at this point does not represent the information of each class well, which leads to classifier bias. Instead, our proposed method significantly improves the performance of the model by heuristically reweight the local prototypes. When there is no orthogonal regularization $(l_{ort})$ , all the metrics of the model show a certain decrease, which indicates that there is a partial overlap between the parameter spaces of different incremental tasks. By imposing orthogonal regularization the parameter spaces corresponding to different input spaces can be made to train the model in the direction of orthogonality to each other, which further improves the model performance. We visualize the cosine similarity between LoRA in Section B.2. + +To better demonstrate the contribution of LoRA, we compare the effects of fine-tuning the entire backbone and freezing the entire backbone under self-supervised pre-trained weights when $\alpha = 6$ and $\beta = 0.5$ . As can be seen in Tab. 3, when fine-tuning the entire backbone, the performance of the model is severely degraded, we believe that this is due to the fact that the model parameters of ViT are too large to achieve a balance between the ability to remember old + +classes and the ability to discriminate new classes under the constraint of knowledge distillation. Freezing the entire backbone network can retain discriminative ability for old classes, but at this point, the model has too few trainable parameters, with only prototypes used for classification, thereby limiting the expressive capacity of the model. Therefore, to achieve a trade-off between communication cost and model performance, we fine-tune and achieve the best performance with a tiny number of parameters through Incremental LoRA. + +# 5.4 Further Analysis + +Memory usage analysis. For our PILoRA, in addition to the model for each client, we store the LoRA parameters $A_{q}^{1:t - 1}$ and $A_{v}^{1:t - 1}$ up to the current stage $t$ to compute the orthogonal regularization. Whereas TARGET requires an additional generator to generate old samples, as well as additional memory space to store the generated images, LGA and GLFC similarly take up additional space to store old class samples. Compared to them, our method takes up very little memory space, storing on average only $0.04\%$ of the parameters equivalent to the ViT-base. + +![](images/eba9aef916c37020ca0159111514aacd6192e007e387183f7e255efe8119c05e.jpg) +Fig.4: The impact of different number of $K$ on CIFAR-100, where we consider $\alpha = 6$ and $\beta = 0.5$ . + +![](images/ae690ecd148e5ca54d747d9d26469f771e9b33fc02aa59931e30013d2a665802.jpg) + +Increase the number of local clients $(K)$ . As depicted in Fig. 4, we investigate the performance of the model as the number of local clients increase, by respectively setting $K = \{10, 15, 20\}$ . From the results shown in Fig. 4, we observe that the model's performance slightly declines as the number of clients increases. We believe that under the same non-IID setting, enlarging the number of clients further exacerbates the heterogeneity among clients, consequently impacting the model's performance. More results can be seen in Section B.3. + +# 6 Conclusion + +In this paper, we propose a simple and effective method of $PILoRA$ for FCIL. PILoRA is based on pre-trained ViT models and fine-tunes with a tiny number of parameters using LoRA. To address the catastrophic forgetting problem in FCIL, we propose incremental LoRA, which can efficiently combine different incremental tasks by summing the orthogonal LoRA parameter space; To deal with the classifier bias caused by data heterogeneity, we adopt prototype learning and propose prototype re-weight, which utilize heuristic information between prototypes and features to perform weighted aggregation of global prototypes. Experimental results show that our method achieves state-of-the-art results on standard datasets and maintains robustness under extreme data heterogeneity. + +# Acknowledgements + +This work has been supported by the National Science and Technology Major Project (2022ZD0116500), National Natural Science Foundation of China (U20A20223, 62222609, 62076236), CAS Project for Young Scientists in Basic Research (YSBR-083), and Key Research Program of Frontier Sciences of CAS (ZDBS-LY-7004) and the InnoHK program. + +# References + +1. Bagwe, G., Yuan, X., Pan, M., Zhang, L.: Fed-cprompt: Contrastive prompt for rehearsal-free federated continual learning. In: Federated Learning and Analytics in Practice: Algorithms, Systems, Applications, and Opportunities (2023) +2. Belouadah, E., Popescu, A.: Deesil: Deep-shallow incremental learning. In: Proceedings of the European Conference on Computer Vision (ECCV) Workshops. pp. 0-0 (2018) +3. Belouadah, E., Popescu, A.: II2m: Class incremental learning with dual memory. In: Proceedings of the IEEE/CVF international conference on computer vision. pp. 583-592 (2019) +4. Brown, T., Mann, B., Ryder, N., Subbiah, M., Kaplan, J.D., Dhariwal, P., Neelakantan, A., Shyam, P., Sastry, G., Askell, A., et al.: Language models are few-shot learners. Advances in neural information processing systems 33, 1877-1901 (2020) +5. Caron, M., Touvron, H., Misra, I., Jégou, H., Mairal, J., Bojanowski, P., Joulin, A.: Emerging properties in self-supervised vision transformers. In: Proceedings of the International Conference on Computer Vision (ICCV) (2021) +6. Cheng, Z., Zhang, X.Y., Liu, C.L.: Unified classification and rejection: A one-versus-all framework. arXiv preprint arXiv:2311.13355 (2023) +7. Chitale, R., Vaidya, A., Kane, A., Ghotkar, A.: Task arithmetic with lora for continual learning. arXiv preprint arXiv:2311.02428 (2023) +8. Deng, J., Dong, W., Socher, R., Li, L.J., Li, K., Fei-Fei, L.: Imagenet: A large-scale hierarchical image database. In: 2009 IEEE conference on computer vision and pattern recognition. pp. 248-255. IEEE (2009) +9. Dong, J., Cong, Y., Sun, G., Zhang, Y., Schiele, B., Dai, D.: No one left behind: Real-world federated class-incremental learning. arXiv preprint arXiv:2302.00903 (2023) +0. Dong, J., Wang, L., Fang, Z., Sun, G., Xu, S., Wang, X., Zhu, Q.: Federated class incremental learning. In: Proceedings of the IEEE/CVF conference on computer vision and pattern recognition. pp. 10164-10173 (2022) +11. Dosovitskiy, A., Beyer, L., Kolesnikov, A., Weissenborn, D., Zhai, X., Unterthiner, T., Dehghani, M., Minderer, M., Heigold, G., Gelly, S., et al.: An image is worth 16x16 words: Transformers for image recognition at scale. arXiv preprint arXiv:2010.11929 (2020) +12. Gao, Q., Zhao, C., Sun, Y., Xi, T., Zhang, G., Ghanem, B., Zhang, J.: A unified continual learning framework with general parameter-efficient tuning. arXiv preprint arXiv:2303.10070 (2023) +13. Geng, C., Huang, S.j., Chen, S.: Recent advances in open set recognition: A survey. IEEE transactions on pattern analysis and machine intelligence 43(10), 3614-3631 (2020) + +14. Giuseppi, A., Manfredi, S., Pietrabissa, A.: A weighted average consensus approach for decentralized federated learning. Machine Intelligence Research 19(4), 319-330 (2022). https://doi.org/10.1007/s11633-022-1338-z, https://www.mi-research.net/en/article/doi/10.1007/s11633-022-1338-z +15. Guo, T., Guo, S., Wang, J., Tang, X., Xu, W.: Promptfl: Let federated participants cooperatively learn prompts instead of models-federated learning in age of foundation model. IEEE Transactions on Mobile Computing (2023) +16. Houlsby, N., Giurgiu, A., Jastrzebski, S., Morrone, B., De Laroussilhe, Q., Gesmundo, A., Attariyan, M., Gelly, S.: Parameter-efficient transfer learning for nlp. In: International Conference on Machine Learning. pp. 2790-2799. PMLR (2019) +17. Hu, E.J., Shen, Y., Wallis, P., Allen-Zhu, Z., Li, Y., Wang, S., Wang, L., Chen, W.: Lora: Low-rank adaptation of large language models. arXiv preprint arXiv:2106.09685 (2021) +18. Ilharco, G., Ribeiro, M.T., Wortsman, M., Gururangan, S., Schmidt, L., Hajishirzi, H., Farhadi, A.: Editing models with task arithmetic. arXiv preprint arXiv:2212.04089 (2022) +19. Karimireddy, S.P., Kale, S., Mohri, M., Reddi, S., Stich, S., Suresh, A.T.: Scaffold: Stochastic controlled averaging for federated learning. In: International conference on machine learning. pp. 5132-5143. PMLR (2020) +20. Khan, L.U., Saad, W., Han, Z., Hossain, E., Hong, C.S.: Federated learning for internet of things: Recent advances, taxonomy, and open challenges. IEEE Communications Surveys & Tutorials 23(3), 1759-1799 (2021) +21. Kingma, D.P., Ba, J.: Adam: A method for stochastic optimization. arXiv preprint arXiv:1412.6980 (2014) +22. Kirillov, A., Mintun, E., Ravi, N., Mao, H., Rolland, C., Gustafson, L., Xiao, T., Whitehead, S., Berg, A.C., Lo, W.Y., et al.: Segment anything. arXiv preprint arXiv:2304.02643 (2023) +23. Kirkpatrick, J., Pascanu, R., Rabinowitz, N., Veness, J., Desjardins, G., Rusu, A.A., Milan, K., Quan, J., Ramalho, T., Grabska-Barwinska, A., et al.: Overcoming catastrophic forgetting in neural networks. Proceedings of the national academy of sciences 114(13), 3521-3526 (2017) +24. Krizhevsky, A., Hinton, G., et al.: Learning multiple layers of features from tiny images (2009) +25. Le, Y., Yang, X.: Tiny imagenet visual recognition challenge. CS 231N 7(7), 3 (2015) +26. Legate, G., Bernier, N., Page-Caccia, L., Oyallon, E., Belilovsky, E.: Guiding the last layer in federated learning with pre-trained models. Advances in Neural Information Processing Systems 36 (2024) +27. Li, Q., Diao, Y., Chen, Q., He, B.: Federated learning on non-iid data silos: An experimental study. In: 2022 IEEE 38th International Conference on Data Engineering (ICDE). pp. 965-978. IEEE (2022) +28. Li, T., Sahu, A.K., Zaheer, M., Sanjabi, M., Talwalkar, A., Smith, V.: Federated optimization in heterogeneous networks. Proceedings of Machine learning and systems 2, 429-450 (2020) +29. Li, Z., Hoiem, D.: Learning without forgetting. IEEE transactions on pattern analysis and machine intelligence 40(12), 2935-2947 (2017) +30. Lin, Y.M., Gao, Y., Gong, M.G., Zhang, S.J., Zhang, Y.Q., Li, Z.Y.: Federated learning on multimodal data: A comprehensive survey. Machine Intelligence Research 20(4), 539-553 (2023). https://doi.org/10.1007/s11633-022-1398-0, https://www.mi-research.net/en/article/doi/10.1007/s11633-022-1398-0 + +31. Liu, C.L., Sako, H., Fujisawa, H.: Discriminative learning quadratic discriminant function for handwriting recognition. IEEE Transactions on Neural Networks 15(2), 430-444 (2004) +32. Liu, C.L., Sako, H., Fujisawa, H.: Effects of classifier structures and training regimes on integrated segmentation and recognition of handwritten numeral strings. IEEE transactions on pattern analysis and machine intelligence 26(11), 1395-1407 (2004) +33. Liu, C., Qu, X., Wang, J., Xiao, J.: Fedet: A communication-efficient federated class-incremental learning framework based on enhanced transformer. arXiv preprint arXiv:2306.15347 (2023) +34. Liu, P., Yuan, W., Fu, J., Jiang, Z., Hayashi, H., Neubig, G.: Pre-train, prompt, and predict: A systematic survey of prompting methods in natural language processing. ACM Computing Surveys 55(9), 1-35 (2023) +35. Liu, W., Wu, X., Zhu, F., Yu, M., Wang, C., Liu, C.L.: Class incremental learning with self-supervised pre-training and prototype learning. arXiv preprint arXiv:2308.02346 (2023) +36. Liu, W., Zhu, F., Liu, C.L.: Branch-tuning: Balancing stability and plasticity for continual self-supervised learning. arXiv preprint arXiv:2403.18266 (2024) +37. Liu, W., Zhu, F., Liu, C.L.: Towards non-exemplar semi-supervised class-incremental learning. arXiv preprint arXiv:2403.18291 (2024) +38. Luo, M., Chen, F., Hu, D., Zhang, Y., Liang, J., Feng, J.: No fear of heterogeneity: Classifier calibration for federated learning with non-iid data. Advances in Neural Information Processing Systems 34, 5972-5984 (2021) +39. McCloskey, M., Cohen, N.J.: Catastrophic interference in connectionist networks: The sequential learning problem. In: Psychology of learning and motivation, vol. 24, pp. 109-165. Elsevier (1989) +40. McMahan, B., Moore, E., Ramage, D., Hampson, S., y Arcas, B.A.: Communication-efficient learning of deep networks from decentralized data. In: Artificial intelligence and statistics. pp. 1273-1282. PMLR (2017) +41. Niknam, S., Dhillon, H.S., Reed, J.H.: Federated learning for wireless communications: Motivation, opportunities, and challenges. IEEE Communications Magazine 58(6), 46-51 (2020) +42. Ortiz-Jimenez, G., Favero, A., Frossard, P.: Task arithmetic in the tangent space: Improved editing of pre-trained models. Advances in Neural Information Processing Systems 36 (2024) +43. Qi, D., Zhao, H., Li, S.: Better generative replay for continual federated learning. In: The Eleventh International Conference on Learning Representations (2022) +44. Qu, L., Zhou, Y., Liang, P.P., Xia, Y., Wang, F., Adeli, E., Fei-Fei, L., Rubin, D.: Rethinking architecture design for tackling data heterogeneity in federated learning. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 10061-10071 (2022) +45. Radford, A., Kim, J.W., Hallacy, C., Ramesh, A., Goh, G., Agarwal, S., Sastry, G., Askell, A., Mishkin, P., Clark, J., et al.: Learning transferable visual models from natural language supervision. In: International conference on machine learning. pp. 8748-8763. PMLR (2021) +46. Rebuffi, S.A., Kolesnikov, A., Sperl, G., Lampert, C.H.: icarl: Incremental classifier and representation learning. In: Proceedings of the IEEE conference on Computer Vision and Pattern Recognition. pp. 2001-2010 (2017) +47. Rieke, N., Hancox, J., Li, W., Miletari, F., Roth, H.R., Albarqouni, S., Bakas, S., Galtier, M.N., Landman, B.A., Maier-Hein, K., et al.: The future of digital health with federated learning. NPJ digital medicine 3(1), 119 (2020) + +48. Shang, X., Lu, Y., Huang, G., Wang, H.: Federated learning on heterogeneous and long-tailed data via classifier re-training with federated features. arXiv preprint arXiv:2204.13399 (2022) +49. Shenaj, D., Toldo, M., Rigon, A., Zanuttigh, P.: Asynchronous federated continual learning. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 5054-5062 (2023) +50. Smith, J.S., Karlinsky, L., Gutta, V., Cascante-Bonilla, P., Kim, D., Arbelle, A., Panda, R., Feris, R., Kira, Z.: Coda-prompt: Continual decomposed attention-based prompting for rehearsal-free continual learning. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 11909-11919 (2023) +51. Tan, Y., Long, G., Liu, L., Zhou, T., Lu, Q., Jiang, J., Zhang, C.: Fedproto: Federated prototype learning across heterogeneous clients. In: Proceedings of the AAAI Conference on Artificial Intelligence. vol. 36, pp. 8432-8440 (2022) +52. Voigt, P., Von dem Bussche, A.: The eu general data protection regulation (gdpr). A Practical Guide, 1st Ed., Cham: Springer International Publishing 10(3152676), 10-5555 (2017) +53. Wang, X., Chen, T., Ge, Q., Xia, H., Bao, R., Zheng, R., Zhang, Q., Gui, T., Huang, X.: Orthogonal subspace learning for language model continual learning. arXiv preprint arXiv:2310.14152 (2023) +54. Wang, Z., Zhang, Z., Ebrahimi, S., Sun, R., Zhang, H., Lee, C.Y., Ren, X., Su, G., Perot, V., Dy, J., et al.: Dualprompt: Complementary prompting for rehearsal-free continual learning. In: European Conference on Computer Vision. pp. 631-648. Springer (2022) +55. Wang, Z., Zhang, Z., Lee, C.Y., Zhang, H., Sun, R., Ren, X., Su, G., Perot, V., Dy, J., Pfister, T.: Learning to prompt for continual learning. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 139-149 (2022) +56. Yang, H.M., Zhang, X.Y., Yin, F., Liu, C.L.: Robust classification with convolutional prototype learning. In: Proceedings of the IEEE conference on computer vision and pattern recognition. pp. 3474-3482 (2018) +57. Yang, H.M., Zhang, X.Y., Yin, F., Yang, Q., Liu, C.L.: Convolutional prototype network for open set recognition. IEEE Transactions on Pattern Analysis and Machine Intelligence 44(5), 2358-2370 (2020) +58. Yi, L., Yu, H., Wang, G., Liu, X.: Fedlora: Model-heterogeneous personalized federated learning with lora tuning. arXiv preprint arXiv:2310.13283 (2023) +59. Yoon, J., Jeong, W., Lee, G., Yang, E., Hwang, S.J.: Federated continual learning with weighted inter-client transfer. In: International Conference on Machine Learning. pp. 12073-12086. PMLR (2021) +60. Zhang, G., Wang, L., Kang, G., Chen, L., Wei, Y.: Slow learner with classifier alignment for continual learning on a pre-trained model. arXiv preprint arXiv:2303.05118 (2023) +61. Zhang, J., Chen, C., Zhuang, W., Lv, L.: Addressing catastrophic forgetting in federated class-continual learning. arXiv preprint arXiv:2303.06937 (2023) +62. Zhang, J., Chen, C., Zhuang, W., Lyu, L.: Target: Federated class-continual learning via exemplar-free distillation. In: Proceedings of the IEEE/CVF International Conference on Computer Vision. pp. 4782-4793 (2023) +63. Zhao, H., Du, W., Li, F., Li, P., Liu, G.: Reduce communication costs and preserve privacy: Prompt tuning method in federated learning. arXiv preprint arXiv:2208.12268 (2022) + +64. Zhu, F., Cheng, Z., Zhang, X.y., Liu, C.l.: Class-incremental learning via dual augmentation. Advances in Neural Information Processing Systems 34, 14306-14318 (2021) +65. Zhu, F., Ma, S., Cheng, Z., Zhang, X.Y., Zhang, Z., Liu, C.L.: Open-world machine learning: A review and new outlooks. arXiv preprint arXiv:2403.01759 (2024) +66. Zhu, F., Zhang, X.Y., Wang, C., Yin, F., Liu, C.L.: Prototype augmentation and self-supervision for incremental learning. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 5871-5880 (2021) \ No newline at end of file diff --git a/2024/PILoRA_ Prototype Guided Incremental LoRA for Federated Class-Incremental Learning/images.zip b/2024/PILoRA_ Prototype Guided Incremental LoRA for Federated Class-Incremental Learning/images.zip new file mode 100644 index 0000000000000000000000000000000000000000..2a5b191f15340d9882d2f0a4fd6dee76068b02db --- /dev/null +++ b/2024/PILoRA_ Prototype Guided Incremental LoRA for Federated Class-Incremental Learning/images.zip @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:561479152977fb62aa7ff43119b88bc96671ffd9c19ae5edc72001e6403e26c4 +size 419829 diff --git a/2024/PILoRA_ Prototype Guided Incremental LoRA for Federated Class-Incremental Learning/layout.json b/2024/PILoRA_ Prototype Guided Incremental LoRA for Federated Class-Incremental Learning/layout.json new file mode 100644 index 0000000000000000000000000000000000000000..17181cd7ab31c9133bf3ab82bd758453a8dae13b --- /dev/null +++ b/2024/PILoRA_ Prototype Guided Incremental LoRA for Federated Class-Incremental Learning/layout.json @@ -0,0 +1,10799 @@ +{ + "pdf_info": [ + { + "para_blocks": [ + { + "bbox": [ + 143, + 111, + 473, + 148 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 143, + 111, + 473, + 148 + ], + "spans": [ + { + "bbox": [ + 143, + 111, + 473, + 148 + ], + "type": "text", + "content": "PILoRA: Prototype Guided Incremental LoRA for Federated Class-Incremental Learning" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 210, + 168, + 402, + 194 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 210, + 168, + 402, + 194 + ], + "spans": [ + { + "bbox": [ + 210, + 168, + 402, + 194 + ], + "type": "text", + "content": "Haiyang Guo" + }, + { + "bbox": [ + 210, + 168, + 402, + 194 + ], + "type": "inline_equation", + "content": "^{1,2}" + }, + { + "bbox": [ + 210, + 168, + 402, + 194 + ], + "type": "text", + "content": ", Fei Zhu" + }, + { + "bbox": [ + 210, + 168, + 402, + 194 + ], + "type": "inline_equation", + "content": "^{3}" + }, + { + "bbox": [ + 210, + 168, + 402, + 194 + ], + "type": "text", + "content": ", Wenzhuo Liu" + }, + { + "bbox": [ + 210, + 168, + 402, + 194 + ], + "type": "inline_equation", + "content": "^{1,2}" + }, + { + "bbox": [ + 210, + 168, + 402, + 194 + ], + "type": "text", + "content": ", Xu-Yao Zhang" + }, + { + "bbox": [ + 210, + 168, + 402, + 194 + ], + "type": "inline_equation", + "content": "^{1,2\\star}" + }, + { + "bbox": [ + 210, + 168, + 402, + 194 + ], + "type": "text", + "content": ", and Cheng-Lin Liu" + }, + { + "bbox": [ + 210, + 168, + 402, + 194 + ], + "type": "inline_equation", + "content": "^{1,2}" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 175, + 201, + 436, + 213 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 175, + 201, + 436, + 213 + ], + "spans": [ + { + "bbox": [ + 175, + 201, + 436, + 213 + ], + "type": "inline_equation", + "content": "^{1}" + }, + { + "bbox": [ + 175, + 201, + 436, + 213 + ], + "type": "text", + "content": " MAIS, Institute of Automation, Chinese Academy of Sciences" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 149, + 213, + 463, + 224 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 149, + 213, + 463, + 224 + ], + "spans": [ + { + "bbox": [ + 149, + 213, + 463, + 224 + ], + "type": "inline_equation", + "content": "^{2}" + }, + { + "bbox": [ + 149, + 213, + 463, + 224 + ], + "type": "text", + "content": " School of Artificial Intelligence, University of Chinese Academy of Sciences" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 133, + 224, + 479, + 246 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 224, + 479, + 246 + ], + "spans": [ + { + "bbox": [ + 133, + 224, + 479, + 246 + ], + "type": "text", + "content": "3 Centre for Artificial Intelligence and Robotics, Hong Kong Institute of Science and Innovation, Chinese Academy of Sciences" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 143, + 247, + 469, + 269 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 143, + 247, + 469, + 269 + ], + "spans": [ + { + "bbox": [ + 143, + 247, + 469, + 269 + ], + "type": "text", + "content": "{guohaiyang2023, zhufei2018}@ia.ac.cn, liuwenzhuo20@mails.ucas.ac.cn, xyz, liucl}@nlpr.ia.ac.cn" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 159, + 302, + 455, + 501 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 159, + 302, + 455, + 501 + ], + "spans": [ + { + "bbox": [ + 159, + 302, + 455, + 501 + ], + "type": "text", + "content": "Abstract. Existing federated learning methods have effectively dealt with decentralized learning in scenarios involving data privacy and non-IID data. However, in real-world situations, each client dynamically learns new classes, requiring the global model to classify all seen classes. To effectively mitigate catastrophic forgetting and data heterogeneity under low communication costs, we propose a simple and effective method named PILoRA. On the one hand, we adopt prototype learning to learn better feature representations and leverage the heuristic information between prototypes and class features to design a prototype re-weight module to solve the classifier bias caused by data heterogeneity without retraining the classifier. On the other hand, we view incremental learning as the process of learning distinct task vectors and encoding them within different LoRA parameters. Accordingly, we propose Incremental LoRA to mitigate catastrophic forgetting. Experimental results on standard datasets indicate that our method outperforms the state-of-the-art approaches significantly. More importantly, our method exhibits strong robustness and superiority in different settings and degrees of data heterogeneity. The code is available at https://github.com/GhyO501/PILORA." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 160, + 510, + 416, + 523 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 160, + 510, + 416, + 523 + ], + "spans": [ + { + "bbox": [ + 160, + 510, + 416, + 523 + ], + "type": "text", + "content": "Keywords: Federated Learning " + }, + { + "bbox": [ + 160, + 510, + 416, + 523 + ], + "type": "inline_equation", + "content": "\\cdot" + }, + { + "bbox": [ + 160, + 510, + 416, + 523 + ], + "type": "text", + "content": " Class Incremental Learning" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 132, + 545, + 230, + 558 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 545, + 230, + 558 + ], + "spans": [ + { + "bbox": [ + 132, + 545, + 230, + 558 + ], + "type": "text", + "content": "1 Introduction" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 130, + 572, + 482, + 646 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 572, + 482, + 646 + ], + "spans": [ + { + "bbox": [ + 130, + 572, + 482, + 646 + ], + "type": "text", + "content": "Federated learning (FL) [40] is a novel distributed machine learning paradigm that enables multiple data owners to collaboratively train a shared model while ensuring the privacy of their local data. In recent years, with the increasing emphasis on data privacy in society and the refinement of relevant regulations [52], FL has experienced rapid development and has been widely applied in various real-world scenarios [14,20,30,41,47]." + } + ] + } + ], + "index": 9 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 133, + 653, + 234, + 666 + ], + "type": "page_footnote", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 653, + 234, + 666 + ], + "spans": [ + { + "bbox": [ + 133, + 653, + 234, + 666 + ], + "type": "text", + "content": "* Corresponding author" + } + ] + } + ], + "index": 10 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 0 + }, + { + "para_blocks": [ + { + "bbox": [ + 130, + 116, + 482, + 261 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 116, + 482, + 261 + ], + "spans": [ + { + "bbox": [ + 130, + 116, + 482, + 261 + ], + "type": "text", + "content": "Existing FL methods [19, 28, 40] typically rely on the closed-world assumption [6, 13, 65], that is, the number of classes seen by the model remains the same and unchanged during both the training and testing stages. However, the real world is dynamic and constantly changing, the local clients often need to receive new data to update the global model continuously. Therefore, Federated Class Incremental Learning (FCIL) [10, 59] has been proposed to handle FL tasks in dynamic scenarios. Specifically, each local client can only update the model with new class data at each stage and upload the model parameters to the global server for aggregation, while the global model needs to maintain its discriminative ability for all seen classes. Moreover, the data distribution of local data follows the non-independent and identically distributed (non-IID) assumption [27]." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 130, + 265, + 482, + 493 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 265, + 482, + 493 + ], + "spans": [ + { + "bbox": [ + 130, + 265, + 482, + 493 + ], + "type": "text", + "content": "FCIL presents a more realistic setting for real-world applications, but it also introduces greater challenges, as FCIL needs to simultaneously address the issue of catastrophic forgetting [23, 39] and data heterogeneity resulting from non-IID. In existing studies, one way is to store a subset of data from old classes and train them together when learning new tasks [9, 10], but the amount of old data stored is strictly limited due to privacy protection requirements. An alternative way is to utilize generative models to generate pseudo-samples of old data [43, 61], thus retaining the ability to classify the old classes. However, training a good generator will cause greater computational overhead, and the generator also suffers from catastrophic forgetting. Another direction is to use pre-trained models for fine-tuning [1, 33], these methods store different stages of knowledge by maintaining a pool of modules and inserting the corresponding modules at inference time based on the similarity between the input data and the modules. However, storing these modules takes up additional memory space. Although they provide different perspectives for solving FCIL, there is still a large unexplored space, especially the performance of the model under different non-IID settings and different degrees of data heterogeneity. So in order to better address this issue, we start by addressing a fundamental question, what is the panacea for dealing with CIL and FL?" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 130, + 498, + 483, + 666 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 498, + 483, + 666 + ], + "spans": [ + { + "bbox": [ + 130, + 498, + 483, + 666 + ], + "type": "text", + "content": "In other words, if we can find common ground in dealing with CIL and FL, it will be of great help in solving the FCIL problem. Specifically, we observe that: 1) At the level of feature representation, both FL and CIL tasks require models to learn a intra-class compact and inter-class separable feature representation. On the one hand, for the CIL task, this kind of feature representation helps to reduce the overlap between features of the old and new classes in the deep feature space, thus alleviating the forgetting of old knowledge. For this reason, PASS [66] introduces self-supervised learning to assist the learning of feature representations. On the other hand, in the FL task, FedEx [51] imposes a constraint on the deep features of samples belonging to the same class. This constraint ensures that these features are proximate to the global prototypes of their respective classes. 2) At the classifier level, classifier drift is the common enemy of both. In CIL, when learning new classes, the decision boundaries learned from old classes can be severely disrupted, resulting in a significant bias" + } + ] + } + ], + "index": 4 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "spans": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "type": "text", + "content": "2" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 166, + 91, + 220, + 101 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 166, + 91, + 220, + 101 + ], + "spans": [ + { + "bbox": [ + 166, + 91, + 220, + 101 + ], + "type": "text", + "content": "H. Guo et al." + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 1 + }, + { + "para_blocks": [ + { + "bbox": [ + 130, + 116, + 482, + 236 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 116, + 482, + 236 + ], + "spans": [ + { + "bbox": [ + 130, + 116, + 482, + 236 + ], + "type": "text", + "content": "in the classification layer [64, 66]. To address this issue, some methods [2, 3, 46] alleviate bias by directly retaining a portion of the data from previous categories and training the model with new data. Other methods [64, 66] address classifier bias by retaining pseudo features of old classes to assist in training the classification layer; In FL, CCVR [38] assesses the similarity among different layers within local models under data heterogeneity and reveals that classifiers exhibit the lowest similarity. This suggests that the classifier of each client models is seriously biased to the local data. Similarly to CIL, [38, 48] utilize the statistical information of local features to retrain the classifier on the global server to alleviate classifier bias." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 130, + 247, + 309, + 544 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 247, + 309, + 544 + ], + "spans": [ + { + "bbox": [ + 130, + 247, + 309, + 544 + ], + "type": "text", + "content": "Motivated by the above findings, in this paper, we propose a Prototype Guided Incremental LoRA (PILoRA) model to tackle the FCIL problem. Specifically, we adopt prototype classifiers in our model to learn intraclass compact and inter-class separable feature representations, which is beneficial for both CIL and FL [28, 35, 56, 57]. We use the pre-trained Transformer model (ViT) as the backbone because the global interactions learned by Transformer are significantly more robust than the local patterns learned by CNN for FL tasks [44] (Fig. 1.), and also provide a good representation of the features. Considering the communication cost, we freeze the entire backbone network during training and use LoRA, a Parameter-Efficient Fine-Tuning (PEFT) method to train the model. To address the catastrophic forgetting problem in FCIL, we propose incremental LoRA, which constrains different stages of LoRA to be trained on" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 130, + 546, + 482, + 666 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 546, + 482, + 666 + ], + "spans": [ + { + "bbox": [ + 130, + 546, + 482, + 666 + ], + "type": "text", + "content": "subspaces orthogonal to each other via orthogonal regularization, and acquires knowledge learned in the past by simple and efficient summation at inference time. To address the classifier bias caused by data heterogeneity, we design a prototype re-weight module for the global server, which aggregates local prototypes based on heuristic information between prototypes and corresponding class features. Compared to other existing FCIL methods, our model achieves superior results on standard datasets. Furthermore, we explore the model's performance under different non-IID settings and degrees of data heterogeneity. Experiments demonstrate the robustness of our approach, while other methods suffer from significant declines." + } + ] + } + ], + "index": 4 + }, + { + "type": "image", + "bbox": [ + 319, + 275, + 479, + 390 + ], + "blocks": [ + { + "bbox": [ + 319, + 275, + 479, + 390 + ], + "lines": [ + { + "bbox": [ + 319, + 275, + 479, + 390 + ], + "spans": [ + { + "bbox": [ + 319, + 275, + 479, + 390 + ], + "type": "image", + "image_path": "83ed409fef50ff66a6ccc5e13482f17907bf577f515fdbefe7df7ab37cbb14ac.jpg" + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 314, + 403, + 482, + 523 + ], + "lines": [ + { + "bbox": [ + 314, + 403, + 482, + 523 + ], + "spans": [ + { + "bbox": [ + 314, + 403, + 482, + 523 + ], + "type": "text", + "content": "Fig.1: The local data of client 1 and client 2 are non-IID. Resnet50 (the first row) focuses more on local patterns, and the patterns learned are significantly different in the case of data heterogeneity, causing the average model to lose some important information (e.g., fish fins). However, ViT (the second row) is less affected by data heterogeneity, and the averaged model basically retains all the information learned by the local model." + } + ] + } + ], + "index": 6, + "angle": 0, + "type": "image_caption" + } + ], + "index": 5 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 411, + 91, + 447, + 100 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 411, + 91, + 447, + 100 + ], + "spans": [ + { + "bbox": [ + 411, + 91, + 447, + 100 + ], + "type": "text", + "content": "PILORA" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 474, + 91, + 480, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 474, + 91, + 480, + 100 + ], + "spans": [ + { + "bbox": [ + 474, + 91, + 480, + 100 + ], + "type": "text", + "content": "3" + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 2 + }, + { + "para_blocks": [ + { + "bbox": [ + 133, + 116, + 397, + 128 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 116, + 397, + 128 + ], + "spans": [ + { + "bbox": [ + 133, + 116, + 397, + 128 + ], + "type": "text", + "content": "The main contributions of this paper can be summarized as:" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 137, + 134, + 480, + 252 + ], + "type": "list", + "angle": 0, + "index": 6, + "blocks": [ + { + "bbox": [ + 137, + 134, + 479, + 169 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 137, + 134, + 479, + 169 + ], + "spans": [ + { + "bbox": [ + 137, + 134, + 479, + 169 + ], + "type": "text", + "content": "- We propose incremental LoRA, which performs incremental learning on orthogonal subspaces to mitigate catastrophic forgetting and aggregates previous knowledge through simple and efficient parameter summation." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 138, + 171, + 480, + 217 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 171, + 480, + 217 + ], + "spans": [ + { + "bbox": [ + 138, + 171, + 480, + 217 + ], + "type": "text", + "content": "- We propose a prototype re-weight module to form global prototypes using heuristic information between each class of local prototypes and the corresponding class of features. Our method effectively addresses classifier bias induced by data heterogeneity without retraining." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 138, + 218, + 479, + 252 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 218, + 479, + 252 + ], + "spans": [ + { + "bbox": [ + 138, + 218, + 479, + 252 + ], + "type": "text", + "content": "- We conduct extensive experiments on standard datasets and achieve state-of-the-art performance. Furthermore, in the extreme heterogeneous case, our method still maintains robustness, while all other methods suffer plummet." + } + ] + } + ], + "index": 5 + } + ], + "sub_type": "text" + }, + { + "bbox": [ + 132, + 270, + 237, + 283 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 270, + 237, + 283 + ], + "spans": [ + { + "bbox": [ + 132, + 270, + 237, + 283 + ], + "type": "text", + "content": "2 Related Work" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 130, + 294, + 482, + 544 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 294, + 482, + 544 + ], + "spans": [ + { + "bbox": [ + 130, + 294, + 482, + 544 + ], + "type": "text", + "content": "Federated Class-Incremental Learning: Research on FCIL has garnered considerable attention in recent years. Dong et al. [10] first introduces the concept of FCIL and proposes several loss functions on the local side and the global server side to alleviate local catastrophic forgetting and global catastrophic forgetting. However, their method uses a rehearsal buffer to store and retain old class data, and additionally design a proxy server to select the best model, resulting in large memory overhead and communication costs. LGA [9] extends the work of [10], but it still belongs to rehearsal-based FCIL. In the more challenging rehearsal-free FCIL problem, generate models are widely adopted to produce synthetic data that aim to mitigate the catastrophic forgetting on local side and global side [43,62], but its performance is highly dependent on the quality of synthetic data and will incur additional computational costs. Similar to our work, FedSpace [49] designs prototype-based loss to encourage feature vectors of the same class to be close together, while we achieve this goal by using prototype classifier. In contrast to the methods described above, some recent studies [1,33] combine pre-trained models with FCIL and achieve higher performance at a smaller communication cost. However, they both adopt a similarity-based selection strategy, which causes additional memory overheads in inference. Besides, they all use supervised pre-training weights, whereas we argue that this may pose privacy concerns because data from downstream tasks may overlap with pre-trained datasets." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 130, + 545, + 482, + 666 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 545, + 482, + 666 + ], + "spans": [ + { + "bbox": [ + 130, + 545, + 482, + 666 + ], + "type": "text", + "content": "PEFT for Pre-Trained Model: With the emergence of large-scale pre-trained models [4, 22, 45], how to effectively fine-tune these models to adapt the downstream tasks has been a focal point of attention. Recently, LoRA [17], Prompt [34], and Adapter [16] have emerged as standout techniques and have been widely used in CIL [12, 50, 54, 55] and FL [15, 58, 63] tasks. In FCIL, existing methods [1, 33] have attempted to combine Prompt and Adapter with pre-trained models. Specifically, they store the knowledge learned at each stage in the parameters of the Prompt or Adapter module and select the appropriate module for the current input to be embedded into the model by a specific similarity computation during the inference time, thus effectively mitigating catastrophic" + } + ] + } + ], + "index": 9 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "spans": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "type": "text", + "content": "4" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 166, + 91, + 220, + 101 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 166, + 91, + 220, + 101 + ], + "spans": [ + { + "bbox": [ + 166, + 91, + 220, + 101 + ], + "type": "text", + "content": "H. Guo et al." + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 3 + }, + { + "para_blocks": [ + { + "bbox": [ + 130, + 116, + 482, + 177 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 116, + 482, + 177 + ], + "spans": [ + { + "bbox": [ + 130, + 116, + 482, + 177 + ], + "type": "text", + "content": "forgetting with a tiny communication costs. However, such similarity matching based approaches undoubtedly introduce inference delays since they require additional similarity computation modules. (In [33], they even train a separate CNN to compute the similarity.) Furthermore, they need a additional memory space in the global server to set up a parameter pool to store the module parameters." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 132, + 194, + 234, + 208 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 194, + 234, + 208 + ], + "spans": [ + { + "bbox": [ + 132, + 194, + 234, + 208 + ], + "type": "text", + "content": "3 Preliminaries" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 131, + 220, + 483, + 319 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 131, + 220, + 483, + 319 + ], + "spans": [ + { + "bbox": [ + 131, + 220, + 483, + 319 + ], + "type": "text", + "content": "In FCIL setting, each client has a local stream dataset " + }, + { + "bbox": [ + 131, + 220, + 483, + 319 + ], + "type": "inline_equation", + "content": "D_{k} = \\{D_{k}^{t}\\}_{t=1}^{T}" + }, + { + "bbox": [ + 131, + 220, + 483, + 319 + ], + "type": "text", + "content": ", where " + }, + { + "bbox": [ + 131, + 220, + 483, + 319 + ], + "type": "inline_equation", + "content": "D_{k}^{t} = \\{\\mathbf{X}_{k}^{t}, \\mathbf{Y}_{k}^{t}\\} = \\{x_{k,i}^{t}, y_{k,i}^{t}\\}_{i=1}^{N_{t}}" + }, + { + "bbox": [ + 131, + 220, + 483, + 319 + ], + "type": "text", + "content": " is the dataset of the " + }, + { + "bbox": [ + 131, + 220, + 483, + 319 + ], + "type": "inline_equation", + "content": "k" + }, + { + "bbox": [ + 131, + 220, + 483, + 319 + ], + "type": "text", + "content": "-th client on task " + }, + { + "bbox": [ + 131, + 220, + 483, + 319 + ], + "type": "inline_equation", + "content": "t" + }, + { + "bbox": [ + 131, + 220, + 483, + 319 + ], + "type": "text", + "content": ". Dataset " + }, + { + "bbox": [ + 131, + 220, + 483, + 319 + ], + "type": "inline_equation", + "content": "D_{k}^{t}" + }, + { + "bbox": [ + 131, + 220, + 483, + 319 + ], + "type": "text", + "content": " contains " + }, + { + "bbox": [ + 131, + 220, + 483, + 319 + ], + "type": "inline_equation", + "content": "N_{k}^{t}" + }, + { + "bbox": [ + 131, + 220, + 483, + 319 + ], + "type": "text", + "content": " training samples and their label " + }, + { + "bbox": [ + 131, + 220, + 483, + 319 + ], + "type": "inline_equation", + "content": "\\mathbf{Y}_{k}^{t} \\in C_{k}^{t}" + }, + { + "bbox": [ + 131, + 220, + 483, + 319 + ], + "type": "text", + "content": ", where " + }, + { + "bbox": [ + 131, + 220, + 483, + 319 + ], + "type": "inline_equation", + "content": "C_{k}^{t}" + }, + { + "bbox": [ + 131, + 220, + 483, + 319 + ], + "type": "text", + "content": " is the class set of the " + }, + { + "bbox": [ + 131, + 220, + 483, + 319 + ], + "type": "inline_equation", + "content": "k" + }, + { + "bbox": [ + 131, + 220, + 483, + 319 + ], + "type": "text", + "content": "-th client in task " + }, + { + "bbox": [ + 131, + 220, + 483, + 319 + ], + "type": "inline_equation", + "content": "t" + }, + { + "bbox": [ + 131, + 220, + 483, + 319 + ], + "type": "text", + "content": ". In particular, the distribution of different clients " + }, + { + "bbox": [ + 131, + 220, + 483, + 319 + ], + "type": "inline_equation", + "content": "k" + }, + { + "bbox": [ + 131, + 220, + 483, + 319 + ], + "type": "text", + "content": " under the same task is non-IID and the class sets of different tasks " + }, + { + "bbox": [ + 131, + 220, + 483, + 319 + ], + "type": "inline_equation", + "content": "t" + }, + { + "bbox": [ + 131, + 220, + 483, + 319 + ], + "type": "text", + "content": " are disjoint. For local client, the objective is to minimize a pre-defined loss function " + }, + { + "bbox": [ + 131, + 220, + 483, + 319 + ], + "type": "inline_equation", + "content": "\\mathcal{L}" + }, + { + "bbox": [ + 131, + 220, + 483, + 319 + ], + "type": "text", + "content": " on current dataset " + }, + { + "bbox": [ + 131, + 220, + 483, + 319 + ], + "type": "inline_equation", + "content": "D_{k}^{t}" + }, + { + "bbox": [ + 131, + 220, + 483, + 319 + ], + "type": "text", + "content": ", while avoiding interference with and possibly enhancing the knowledge acquired from previous learning stages:" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 240, + 327, + 481, + 350 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 240, + 327, + 481, + 350 + ], + "spans": [ + { + "bbox": [ + 240, + 327, + 481, + 350 + ], + "type": "interline_equation", + "content": "\\underset {\\boldsymbol {\\omega} _ {k} ^ {t}} {\\operatorname {a r g m i n}} \\mathcal {L} \\left(\\boldsymbol {\\omega} _ {k} ^ {t}; \\boldsymbol {\\omega} ^ {t - 1}, \\mathbf {X} _ {k} ^ {t}, \\mathbf {Y} _ {k} ^ {t}\\right), \\tag {1}", + "image_path": "dc45422c9f6d7251d28bc472252ede81f59a739bf6a3604d441bbc2ff432eaad.jpg" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 131, + 359, + 482, + 396 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 131, + 359, + 482, + 396 + ], + "spans": [ + { + "bbox": [ + 131, + 359, + 482, + 396 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 131, + 359, + 482, + 396 + ], + "type": "inline_equation", + "content": "\\omega_{k}^{t}" + }, + { + "bbox": [ + 131, + 359, + 482, + 396 + ], + "type": "text", + "content": " is the parameters of " + }, + { + "bbox": [ + 131, + 359, + 482, + 396 + ], + "type": "inline_equation", + "content": "k" + }, + { + "bbox": [ + 131, + 359, + 482, + 396 + ], + "type": "text", + "content": "-th local model, and " + }, + { + "bbox": [ + 131, + 359, + 482, + 396 + ], + "type": "inline_equation", + "content": "\\omega^{t-1}" + }, + { + "bbox": [ + 131, + 359, + 482, + 396 + ], + "type": "text", + "content": " is the global model at previous task. Then, the server updates the global model " + }, + { + "bbox": [ + 131, + 359, + 482, + 396 + ], + "type": "inline_equation", + "content": "\\omega^{t}" + }, + { + "bbox": [ + 131, + 359, + 482, + 396 + ], + "type": "text", + "content": " by aggregating all uploaded parameters as follows:" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 225, + 405, + 481, + 437 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 225, + 405, + 481, + 437 + ], + "spans": [ + { + "bbox": [ + 225, + 405, + 481, + 437 + ], + "type": "interline_equation", + "content": "\\boldsymbol {\\omega} ^ {t} = \\sum_ {k = 1} ^ {K} \\gamma_ {k} \\boldsymbol {\\omega} _ {k} ^ {t}, \\text {w h e r e} \\gamma_ {k} = \\frac {N _ {k} ^ {t}}{\\sum_ {k ^ {\\prime}} N _ {k ^ {\\prime}} ^ {t}}. \\tag {2}", + "image_path": "1584901ea7a5b35307a2df9c39ba83cb22fb77914427e699072180c0231e6f80.jpg" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 131, + 445, + 482, + 470 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 131, + 445, + 482, + 470 + ], + "spans": [ + { + "bbox": [ + 131, + 445, + 482, + 470 + ], + "type": "text", + "content": "The global model aims to correctly classify the test samples of all seen classes and solve the problem of data heterogeneity at low communication cost." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 132, + 488, + 228, + 502 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 488, + 228, + 502 + ], + "spans": [ + { + "bbox": [ + 132, + 488, + 228, + 502 + ], + "type": "text", + "content": "4 Our method" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 132, + 514, + 372, + 526 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 514, + 372, + 526 + ], + "spans": [ + { + "bbox": [ + 132, + 514, + 372, + 526 + ], + "type": "text", + "content": "4.1 Incremental LoRA for Pre-Trained Model" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 130, + 533, + 482, + 668 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 533, + 482, + 668 + ], + "spans": [ + { + "bbox": [ + 130, + 533, + 482, + 668 + ], + "type": "text", + "content": "In FCIL, storing the knowledge of different stages in different modules through PEFT can effectively mitigate the catastrophic forgetting [1,33], but additional similarity computation units are required in the inference time in order to select the appropriate module to embed into the model based on the inputs, which leads to inference delay and additional memory overhead. In order to learn an end-to-end global model without taking up extra storage space, we intuitively believe that modules storing knowledge from different stages can be organically combined into one module that embeds knowledge from all stages. Accordingly, we propose Incremental LoRA to effectively address catastrophic forgetting in FCIL. On the one hand, LoRA has the natural advantages of low inference latency and more stable training [17], on the other hand, Inspired by [53, 55], we" + } + ] + } + ], + "index": 11 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 411, + 91, + 447, + 100 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 411, + 91, + 447, + 100 + ], + "spans": [ + { + "bbox": [ + 411, + 91, + 447, + 100 + ], + "type": "text", + "content": "PILORA" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 474, + 92, + 481, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 474, + 92, + 481, + 100 + ], + "spans": [ + { + "bbox": [ + 474, + 92, + 481, + 100 + ], + "type": "text", + "content": "5" + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 4 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 142, + 118, + 470, + 303 + ], + "blocks": [ + { + "bbox": [ + 142, + 118, + 470, + 303 + ], + "lines": [ + { + "bbox": [ + 142, + 118, + 470, + 303 + ], + "spans": [ + { + "bbox": [ + 142, + 118, + 470, + 303 + ], + "type": "image", + "image_path": "7e038bf45d14de35b69d245ef807388ab0f5880b01e46155eb88bf8431464b4a.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 130, + 314, + 482, + 359 + ], + "lines": [ + { + "bbox": [ + 130, + 314, + 482, + 359 + ], + "spans": [ + { + "bbox": [ + 130, + 314, + 482, + 359 + ], + "type": "text", + "content": "Fig. 2: Illustration of PILoRA for FCIL. The client fine-tunes LoRA and prototypes for each class using a local dataset. Upon upload, the global server aggregates the LoRA uploaded by different clients, and applies a prototype re-weight module before re-sending them to each client." + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_caption" + } + ], + "index": 2 + }, + { + "bbox": [ + 130, + 384, + 480, + 418 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 384, + 480, + 418 + ], + "spans": [ + { + "bbox": [ + 130, + 384, + 480, + 418 + ], + "type": "text", + "content": "introduce the orthogonality loss to constrain LoRA to learn new knowledge in a subspace orthogonal to past tasks, thus better preserving the knowledge of the old classes." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 130, + 419, + 480, + 456 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 419, + 480, + 456 + ], + "spans": [ + { + "bbox": [ + 130, + 419, + 480, + 456 + ], + "type": "text", + "content": "Specifically, we define the initialization parameters of the pre-training model as " + }, + { + "bbox": [ + 130, + 419, + 480, + 456 + ], + "type": "inline_equation", + "content": "\\mathbf{W} \\in \\mathbb{R}^{d \\times k}" + }, + { + "bbox": [ + 130, + 419, + 480, + 456 + ], + "type": "text", + "content": ", and " + }, + { + "bbox": [ + 130, + 419, + 480, + 456 + ], + "type": "inline_equation", + "content": "\\Delta \\mathbf{W}_{\\mathbf{t}}" + }, + { + "bbox": [ + 130, + 419, + 480, + 456 + ], + "type": "text", + "content": " represents the parameter to be updated for task " + }, + { + "bbox": [ + 130, + 419, + 480, + 456 + ], + "type": "inline_equation", + "content": "t" + }, + { + "bbox": [ + 130, + 419, + 480, + 456 + ], + "type": "text", + "content": ". Then, the model's updates in different incremental learning stages can be expressed as:" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 279, + 468, + 480, + 482 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 279, + 468, + 480, + 482 + ], + "spans": [ + { + "bbox": [ + 279, + 468, + 480, + 482 + ], + "type": "interline_equation", + "content": "\\mathbf {W} + \\Delta \\mathbf {W} _ {\\mathbf {t}}. \\tag {3}", + "image_path": "88792d7332ee62e62916b98bf4f8e8fa8495518229d18e4bcf7d5546be26a6c1.jpg" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 130, + 488, + 481, + 513 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 488, + 481, + 513 + ], + "spans": [ + { + "bbox": [ + 130, + 488, + 481, + 513 + ], + "type": "text", + "content": "LoRA assumes that the weight changes of large-scale pre-trained models when adapting to downstream tasks occur in a low-rank space:" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 248, + 522, + 481, + 536 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 248, + 522, + 481, + 536 + ], + "spans": [ + { + "bbox": [ + 248, + 522, + 481, + 536 + ], + "type": "interline_equation", + "content": "\\mathbf {W} + \\Delta \\mathbf {W} _ {\\mathbf {t}} = \\mathbf {W} + \\mathbf {A} _ {t} \\mathbf {B} _ {t}, \\tag {4}", + "image_path": "e31a99ea61a48d364eefb67089d6191f9c3362736ec555acc2ee1ca63004a8e6.jpg" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 130, + 544, + 481, + 605 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 544, + 481, + 605 + ], + "spans": [ + { + "bbox": [ + 130, + 544, + 481, + 605 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 130, + 544, + 481, + 605 + ], + "type": "inline_equation", + "content": "\\mathbf{A}_t \\in \\mathbb{R}^{d \\times r}" + }, + { + "bbox": [ + 130, + 544, + 481, + 605 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 130, + 544, + 481, + 605 + ], + "type": "inline_equation", + "content": "\\mathbf{B}_t \\in \\mathbb{R}^{r \\times k}" + }, + { + "bbox": [ + 130, + 544, + 481, + 605 + ], + "type": "text", + "content": ", and " + }, + { + "bbox": [ + 130, + 544, + 481, + 605 + ], + "type": "inline_equation", + "content": "r \\ll \\min\\{d, k\\}" + }, + { + "bbox": [ + 130, + 544, + 481, + 605 + ], + "type": "text", + "content": ". At first, " + }, + { + "bbox": [ + 130, + 544, + 481, + 605 + ], + "type": "inline_equation", + "content": "\\mathbf{A}_t" + }, + { + "bbox": [ + 130, + 544, + 481, + 605 + ], + "type": "text", + "content": " is initialized by a random Gaussian, while " + }, + { + "bbox": [ + 130, + 544, + 481, + 605 + ], + "type": "inline_equation", + "content": "\\mathbf{B}_t" + }, + { + "bbox": [ + 130, + 544, + 481, + 605 + ], + "type": "text", + "content": " is initialized with zero. Therefore, " + }, + { + "bbox": [ + 130, + 544, + 481, + 605 + ], + "type": "inline_equation", + "content": "\\mathbf{B}_t" + }, + { + "bbox": [ + 130, + 544, + 481, + 605 + ], + "type": "text", + "content": " can be regarded as the coefficient matrix of " + }, + { + "bbox": [ + 130, + 544, + 481, + 605 + ], + "type": "inline_equation", + "content": "\\mathbf{A}_t" + }, + { + "bbox": [ + 130, + 544, + 481, + 605 + ], + "type": "text", + "content": " [53]. LoRA applies it as a bypass to query and value projection matrices in multi-head attention module, and during adaptation to downstream tasks, only the parameters of " + }, + { + "bbox": [ + 130, + 544, + 481, + 605 + ], + "type": "inline_equation", + "content": "\\mathbf{A}_t" + }, + { + "bbox": [ + 130, + 544, + 481, + 605 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 130, + 544, + 481, + 605 + ], + "type": "inline_equation", + "content": "\\mathbf{B}_t" + }, + { + "bbox": [ + 130, + 544, + 481, + 605 + ], + "type": "text", + "content": " are trainable." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 130, + 606, + 481, + 665 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 606, + 481, + 665 + ], + "spans": [ + { + "bbox": [ + 130, + 606, + 481, + 665 + ], + "type": "text", + "content": "Inspired by [53], we argue that the parameters of LoRA can be viewed as containers that store different subspaces of task gradients. Thus, learning a series of incremental learning tasks can be viewed as learning a series of LoRA parameters. So how to use these LoRA modules to build a model that can classify all seen classes? An idea is to select appropriate modules to be embedded in the" + } + ] + } + ], + "index": 10 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "spans": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "type": "text", + "content": "6" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 166, + 91, + 220, + 100 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 166, + 91, + 220, + 100 + ], + "spans": [ + { + "bbox": [ + 166, + 91, + 220, + 100 + ], + "type": "text", + "content": "H. Guo et al." + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 5 + }, + { + "para_blocks": [ + { + "bbox": [ + 130, + 116, + 479, + 222 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 116, + 479, + 222 + ], + "spans": [ + { + "bbox": [ + 130, + 116, + 479, + 222 + ], + "type": "text", + "content": "model based on the similarity between different modules and the corresponding inputs [1, 33]. However, computing similarity requires additional computational resources, which leads to delayed inference. Another idea is that since LoRA is essentially two weight matrices, it is possible to concat all LoRA parameters prior to stage " + }, + { + "bbox": [ + 130, + 116, + 479, + 222 + ], + "type": "inline_equation", + "content": "t" + }, + { + "bbox": [ + 130, + 116, + 479, + 222 + ], + "type": "text", + "content": " into a new matrix to gain knowledge of all previous stages [53]. In particular, at stage " + }, + { + "bbox": [ + 130, + 116, + 479, + 222 + ], + "type": "inline_equation", + "content": "t" + }, + { + "bbox": [ + 130, + 116, + 479, + 222 + ], + "type": "text", + "content": ", the knowledge learned in the previous " + }, + { + "bbox": [ + 130, + 116, + 479, + 222 + ], + "type": "inline_equation", + "content": "t - 1" + }, + { + "bbox": [ + 130, + 116, + 479, + 222 + ], + "type": "text", + "content": " stages are stored in " + }, + { + "bbox": [ + 130, + 116, + 479, + 222 + ], + "type": "inline_equation", + "content": "\\mathbf{A}_{1:t - 1} = [\\mathbf{A}_1,\\dots ,\\mathbf{A}_{t - 1}]" + }, + { + "bbox": [ + 130, + 116, + 479, + 222 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 130, + 116, + 479, + 222 + ], + "type": "inline_equation", + "content": "\\mathbf{B}_{1:t - 1} = [\\mathbf{B}_1,\\dots ,\\mathbf{B}_{t - 1}]" + }, + { + "bbox": [ + 130, + 116, + 479, + 222 + ], + "type": "text", + "content": ". To obtain the weight matrix, we can sequentially concat these LoRA parameters into a bigger module:" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 250, + 223, + 480, + 236 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 250, + 223, + 480, + 236 + ], + "spans": [ + { + "bbox": [ + 250, + 223, + 480, + 236 + ], + "type": "interline_equation", + "content": "\\mathbf {W} + \\Delta \\mathbf {W} _ {\\mathbf {t}} = \\mathbf {W} + \\widetilde {\\mathbf {A}} _ {t} \\widetilde {\\mathbf {B}} _ {t}, \\tag {5}", + "image_path": "db2b73bfab9360c3ab26391216d90b2ffdfc7843c0549718cd0344094f00ab24.jpg" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 130, + 242, + 480, + 316 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 242, + 480, + 316 + ], + "spans": [ + { + "bbox": [ + 130, + 242, + 480, + 316 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 130, + 242, + 480, + 316 + ], + "type": "inline_equation", + "content": "\\widetilde{\\mathbf{A}}_t = \\mathrm{concat}([\\mathbf{A}_1,\\ldots ,\\mathbf{A}_t])" + }, + { + "bbox": [ + 130, + 242, + 480, + 316 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 130, + 242, + 480, + 316 + ], + "type": "inline_equation", + "content": "\\widetilde{\\mathbf{B}}_t = \\mathrm{concat}([\\mathbf{B}_1,\\ldots ,\\mathbf{B}_t])" + }, + { + "bbox": [ + 130, + 242, + 480, + 316 + ], + "type": "text", + "content": ". However, this will cause the parameters of LoRA in the global model to increase with the number of incremental learning tasks. Therefore, in order to ensure consistency between the local model and the global model, we are inspired by Task Arithmetic [7, 18, 42] in model edit and propose to integrate the LoRA parameters at different stages by summation:" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 176, + 326, + 480, + 357 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 176, + 326, + 480, + 357 + ], + "spans": [ + { + "bbox": [ + 176, + 326, + 480, + 357 + ], + "type": "interline_equation", + "content": "\\mathbf {W} + \\Delta \\mathbf {W} _ {\\mathbf {t}} = \\mathbf {W} + \\overline {{\\mathbf {A}}} _ {t} \\overline {{\\mathbf {B}}} _ {t}, \\text {w h e r e} \\overline {{\\mathbf {A}}} _ {t} = \\sum_ {i = 1} ^ {t} \\mathbf {A} _ {i}, \\overline {{\\mathbf {B}}} _ {t} = \\sum_ {i = 1} ^ {t} \\mathbf {B} _ {i}. \\tag {6}", + "image_path": "fe795501cd463b0b8af435452f432d1748fd64d253ef9a5981221ee2ebc74dd3.jpg" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 130, + 361, + 480, + 481 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 361, + 480, + 481 + ], + "spans": [ + { + "bbox": [ + 130, + 361, + 480, + 481 + ], + "type": "text", + "content": "Specifically, distinct directions in the weight space correspond to various localized regions in the input space [42]. Consequently, a linear combination of these diverse directions encoded within the pre-trained weights enables the model to effectively discriminate between different inputs. In our method, The training data for different incremental tasks are disjoint from each other, so it is reasonable to directly sum the LoRA parameters corresponding to different tasks. Besides, Task Arithmetic also point out that orthogonality between different task parameter vectors helps to better integrate different task. So we propose a orthogonal regularization to achieve this by constraining the parameters of LoRA to be orthogonal to previous tasks:" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 246, + 491, + 480, + 521 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 246, + 491, + 480, + 521 + ], + "spans": [ + { + "bbox": [ + 246, + 491, + 480, + 521 + ], + "type": "interline_equation", + "content": "l _ {o r t} \\left(\\mathbf {A} _ {i}, \\mathbf {A} _ {t}\\right) = \\sum_ {i = 1} ^ {t - 1} \\left| \\mathbf {A} _ {i} ^ {T} \\cdot \\mathbf {A} _ {t} \\right|. \\tag {7}", + "image_path": "753f13cc561071e1d95efc909b60cfc7b605dcccae1c60c3958cccf39c68406b.jpg" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 130, + 531, + 479, + 591 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 531, + 479, + 591 + ], + "spans": [ + { + "bbox": [ + 130, + 531, + 479, + 591 + ], + "type": "text", + "content": "Meanwhile, the orthogonal regularization encourages the learning of distinct tasks along orthogonal directions, thereby effectively reducing the spatial overlap between them. This approach will be proved to be advantageous in mitigating catastrophic forgetting, as it helps preserve the knowledge learned from previous tasks while adapting to new ones." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 131, + 609, + 389, + 621 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 131, + 609, + 389, + 621 + ], + "spans": [ + { + "bbox": [ + 131, + 609, + 389, + 621 + ], + "type": "text", + "content": "4.2 Prototype learning with Prototype Re-weight" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 130, + 629, + 479, + 666 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 629, + 479, + 666 + ], + "spans": [ + { + "bbox": [ + 130, + 629, + 479, + 666 + ], + "type": "text", + "content": "Prototype Learning: As mentioned before, a feature representation that is separable intra-class and compact inter-class is helpful for FCIL tasks. Therefore, the model not only needs to correctly classify known classes but also needs" + } + ] + } + ], + "index": 10 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 411, + 91, + 447, + 100 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 411, + 91, + 447, + 100 + ], + "spans": [ + { + "bbox": [ + 411, + 91, + 447, + 100 + ], + "type": "text", + "content": "PILORA" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 474, + 91, + 481, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 474, + 91, + 481, + 100 + ], + "spans": [ + { + "bbox": [ + 474, + 91, + 481, + 100 + ], + "type": "text", + "content": "7" + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 6 + }, + { + "para_blocks": [ + { + "bbox": [ + 130, + 116, + 482, + 200 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 116, + 482, + 200 + ], + "spans": [ + { + "bbox": [ + 130, + 116, + 482, + 200 + ], + "type": "text", + "content": "to model the distribution of known classes in the feature space. In open set recognition, CPN [57] designs a discriminative loss and generative loss for prototype learning to constrain the range of known classes in the feature space, thus reserving space for samples from unknown classes. Inspired by it, we introduce prototype learning in FCIL. In particular, we set a prototype " + }, + { + "bbox": [ + 130, + 116, + 482, + 200 + ], + "type": "inline_equation", + "content": "m = \\{m_i | i = 1, 2, \\dots, C\\}" + }, + { + "bbox": [ + 130, + 116, + 482, + 200 + ], + "type": "text", + "content": " for each class, where " + }, + { + "bbox": [ + 130, + 116, + 482, + 200 + ], + "type": "inline_equation", + "content": "m_i \\in \\mathbb{R}^d" + }, + { + "bbox": [ + 130, + 116, + 482, + 200 + ], + "type": "text", + "content": ", and the dimensions " + }, + { + "bbox": [ + 130, + 116, + 482, + 200 + ], + "type": "inline_equation", + "content": "d" + }, + { + "bbox": [ + 130, + 116, + 482, + 200 + ], + "type": "text", + "content": " of each prototype are the same as the dimensions of the final deep feature space." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 130, + 200, + 482, + 261 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 200, + 482, + 261 + ], + "spans": [ + { + "bbox": [ + 130, + 200, + 482, + 261 + ], + "type": "text", + "content": "To shorten the distance between the class feature and the corresponding prototype, we apply the distance-based cross entropy (DCE) discriminative loss. Given a sample " + }, + { + "bbox": [ + 130, + 200, + 482, + 261 + ], + "type": "inline_equation", + "content": "(x,y)" + }, + { + "bbox": [ + 130, + 200, + 482, + 261 + ], + "type": "text", + "content": ", DCE uses the distance between sample features " + }, + { + "bbox": [ + 130, + 200, + 482, + 261 + ], + "type": "inline_equation", + "content": "f_{\\theta}(x)" + }, + { + "bbox": [ + 130, + 200, + 482, + 261 + ], + "type": "text", + "content": " and prototype " + }, + { + "bbox": [ + 130, + 200, + 482, + 261 + ], + "type": "inline_equation", + "content": "m_{i}" + }, + { + "bbox": [ + 130, + 200, + 482, + 261 + ], + "type": "text", + "content": " to represent the probability of belonging to class " + }, + { + "bbox": [ + 130, + 200, + 482, + 261 + ], + "type": "inline_equation", + "content": "i" + }, + { + "bbox": [ + 130, + 200, + 482, + 261 + ], + "type": "text", + "content": ". Considering the normalization of probability, DCE adopts the softmax operation:" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 205, + 266, + 481, + 297 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 205, + 266, + 481, + 297 + ], + "spans": [ + { + "bbox": [ + 205, + 266, + 481, + 297 + ], + "type": "interline_equation", + "content": "p (x \\in m _ {i} | x) = \\frac {\\exp (- \\delta \\cdot \\| f _ {\\theta} (x) - m _ {i} \\| _ {2} ^ {2})}{\\sum_ {j = 1} ^ {C} \\exp (- \\delta \\cdot \\| f _ {\\theta} (x) - m _ {j} \\| _ {2} ^ {2})}, \\tag {8}", + "image_path": "725e3c25d4342e43db4864067253ccc2f4d460c3adc6ee2b80a5dcd162630999.jpg" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 131, + 301, + 482, + 350 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 131, + 301, + 482, + 350 + ], + "spans": [ + { + "bbox": [ + 131, + 301, + 482, + 350 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 131, + 301, + 482, + 350 + ], + "type": "inline_equation", + "content": "\\| f_{\\theta}(x) - m_i\\| _2^2" + }, + { + "bbox": [ + 131, + 301, + 482, + 350 + ], + "type": "text", + "content": " is the Euclidean distance between the feature " + }, + { + "bbox": [ + 131, + 301, + 482, + 350 + ], + "type": "inline_equation", + "content": "f_{\\theta}(x)" + }, + { + "bbox": [ + 131, + 301, + 482, + 350 + ], + "type": "text", + "content": " of the input sample and the prototype " + }, + { + "bbox": [ + 131, + 301, + 482, + 350 + ], + "type": "inline_equation", + "content": "m_{i}" + }, + { + "bbox": [ + 131, + 301, + 482, + 350 + ], + "type": "text", + "content": ", and " + }, + { + "bbox": [ + 131, + 301, + 482, + 350 + ], + "type": "inline_equation", + "content": "\\delta" + }, + { + "bbox": [ + 131, + 301, + 482, + 350 + ], + "type": "text", + "content": " is the temperature scalar controlling the hardness of the class distribution. Hence, the distance-based-cross-entropy loss can be defined as:" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 228, + 357, + 481, + 371 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 228, + 357, + 481, + 371 + ], + "spans": [ + { + "bbox": [ + 228, + 357, + 481, + 371 + ], + "type": "interline_equation", + "content": "l _ {d c e} ((x, y); \\theta , m) = - \\log (x \\in m _ {y} | x). \\tag {9}", + "image_path": "d4f270561df412ec5e9aabaff2bb377b64650c3378bbc7c235011cba3b37ee25.jpg" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 131, + 376, + 482, + 437 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 131, + 376, + 482, + 437 + ], + "spans": [ + { + "bbox": [ + 131, + 376, + 482, + 437 + ], + "type": "text", + "content": "By minimizing the DCE loss, the distance between the sample's feature and the correct prototype will be smaller than other incorrect prototypes. However, features learned only under the discriminative loss may not be compact, which may lead to an overlap in feature representations between new and old classes. To solve this problem, we introduce prototype learning (PL) loss [57]:" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 235, + 442, + 481, + 456 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 235, + 442, + 481, + 456 + ], + "spans": [ + { + "bbox": [ + 235, + 442, + 481, + 456 + ], + "type": "interline_equation", + "content": "l _ {p l} ((x, y); \\theta , m) = \\| f _ {\\theta} (x) - m _ {y} \\| _ {2} ^ {2}, \\tag {10}", + "image_path": "3ccca292c145c917ccd2a7049ba8646073eaa04ee841440a7c6c1d9c8438df6a.jpg" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 131, + 462, + 482, + 510 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 131, + 462, + 482, + 510 + ], + "spans": [ + { + "bbox": [ + 131, + 462, + 482, + 510 + ], + "type": "text", + "content": "PL loss decreases the distance between sample features and the corresponding correct prototype, making the model learn a more compact intra-class distribution. Essentially, PL loss is a maximum likelihood regularization of features " + }, + { + "bbox": [ + 131, + 462, + 482, + 510 + ], + "type": "inline_equation", + "content": "f_{\\theta}(x)" + }, + { + "bbox": [ + 131, + 462, + 482, + 510 + ], + "type": "text", + "content": " under the Gaussian mixture density assumption [31, 32]." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 130, + 510, + 482, + 629 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 510, + 482, + 629 + ], + "spans": [ + { + "bbox": [ + 130, + 510, + 482, + 629 + ], + "type": "text", + "content": "Prototype Re-weight: Considering a specific class " + }, + { + "bbox": [ + 130, + 510, + 482, + 629 + ], + "type": "inline_equation", + "content": "c" + }, + { + "bbox": [ + 130, + 510, + 482, + 629 + ], + "type": "text", + "content": ", under the non-IID setting, each client " + }, + { + "bbox": [ + 130, + 510, + 482, + 629 + ], + "type": "inline_equation", + "content": "k" + }, + { + "bbox": [ + 130, + 510, + 482, + 629 + ], + "type": "text", + "content": " holds a portion of the training data " + }, + { + "bbox": [ + 130, + 510, + 482, + 629 + ], + "type": "inline_equation", + "content": "N_{c,k}" + }, + { + "bbox": [ + 130, + 510, + 482, + 629 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 130, + 510, + 482, + 629 + ], + "type": "inline_equation", + "content": "\\sum_{k=1}^{K} N_{c,k} = N_{c}" + }, + { + "bbox": [ + 130, + 510, + 482, + 629 + ], + "type": "text", + "content": ", where " + }, + { + "bbox": [ + 130, + 510, + 482, + 629 + ], + "type": "inline_equation", + "content": "N_{c}" + }, + { + "bbox": [ + 130, + 510, + 482, + 629 + ], + "type": "text", + "content": " is the sum of training samples of class " + }, + { + "bbox": [ + 130, + 510, + 482, + 629 + ], + "type": "inline_equation", + "content": "c" + }, + { + "bbox": [ + 130, + 510, + 482, + 629 + ], + "type": "text", + "content": ". In prototype learning, the prototype " + }, + { + "bbox": [ + 130, + 510, + 482, + 629 + ], + "type": "inline_equation", + "content": "m_{k,c}" + }, + { + "bbox": [ + 130, + 510, + 482, + 629 + ], + "type": "text", + "content": " learned by each client can effectively reflect the distribution of class " + }, + { + "bbox": [ + 130, + 510, + 482, + 629 + ], + "type": "inline_equation", + "content": "c" + }, + { + "bbox": [ + 130, + 510, + 482, + 629 + ], + "type": "text", + "content": " within its local data. That is, for clients that do not have training samples of class " + }, + { + "bbox": [ + 130, + 510, + 482, + 629 + ], + "type": "inline_equation", + "content": "c" + }, + { + "bbox": [ + 130, + 510, + 482, + 629 + ], + "type": "text", + "content": ", the distance between their corresponding prototype and the features of class " + }, + { + "bbox": [ + 130, + 510, + 482, + 629 + ], + "type": "inline_equation", + "content": "c" + }, + { + "bbox": [ + 130, + 510, + 482, + 629 + ], + "type": "text", + "content": " will be greater than for clients that do have training samples. If we give equal weight to these uploaded prototypes during parameter aggregation, classifier drift may occur, especially when the data heterogeneity is high, potentially resulting in the global model losing discriminative capability for that class." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 130, + 630, + 482, + 666 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 630, + 482, + 666 + ], + "spans": [ + { + "bbox": [ + 130, + 630, + 482, + 666 + ], + "type": "text", + "content": "Therefore, we use the heuristic information contained in the distance between the prototype and the corresponding class average features to design the prototype re-weight module. Specifically, at stage " + }, + { + "bbox": [ + 130, + 630, + 482, + 666 + ], + "type": "inline_equation", + "content": "t" + }, + { + "bbox": [ + 130, + 630, + 482, + 666 + ], + "type": "text", + "content": ", each client " + }, + { + "bbox": [ + 130, + 630, + 482, + 666 + ], + "type": "inline_equation", + "content": "k" + }, + { + "bbox": [ + 130, + 630, + 482, + 666 + ], + "type": "text", + "content": " uploads prototypes" + } + ] + } + ], + "index": 11 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "spans": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "type": "text", + "content": "8" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 166, + 91, + 220, + 101 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 166, + 91, + 220, + 101 + ], + "spans": [ + { + "bbox": [ + 166, + 91, + 220, + 101 + ], + "type": "text", + "content": "H. Guo et al." + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 7 + }, + { + "para_blocks": [ + { + "bbox": [ + 133, + 116, + 334, + 128 + ], + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 116, + 334, + 128 + ], + "spans": [ + { + "bbox": [ + 133, + 116, + 334, + 128 + ], + "type": "text", + "content": "Algorithm 1 Prototype Re-weight of Task t." + } + ] + } + ], + "index": 2, + "type": "text" + }, + { + "bbox": [ + 132, + 130, + 481, + 153 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 130, + 481, + 153 + ], + "spans": [ + { + "bbox": [ + 132, + 130, + 481, + 153 + ], + "type": "text", + "content": "Input: the number of classes " + }, + { + "bbox": [ + 132, + 130, + 481, + 153 + ], + "type": "inline_equation", + "content": "C_t" + }, + { + "bbox": [ + 132, + 130, + 481, + 153 + ], + "type": "text", + "content": ", prototypes of each client " + }, + { + "bbox": [ + 132, + 130, + 481, + 153 + ], + "type": "inline_equation", + "content": "m_{t,k}" + }, + { + "bbox": [ + 132, + 130, + 481, + 153 + ], + "type": "text", + "content": ", the set of mean features " + }, + { + "bbox": [ + 132, + 130, + 481, + 153 + ], + "type": "inline_equation", + "content": "\\mu_{t,k}" + }, + { + "bbox": [ + 132, + 130, + 481, + 153 + ], + "type": "text", + "content": " of each client and the temperature coefficient " + }, + { + "bbox": [ + 132, + 130, + 481, + 153 + ], + "type": "inline_equation", + "content": "\\eta" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 132, + 153, + 257, + 163 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 153, + 257, + 163 + ], + "spans": [ + { + "bbox": [ + 132, + 153, + 257, + 163 + ], + "type": "text", + "content": "Output: global prototype " + }, + { + "bbox": [ + 132, + 153, + 257, + 163 + ], + "type": "inline_equation", + "content": "m_t" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 133, + 164, + 360, + 317 + ], + "type": "list", + "angle": 0, + "index": 19, + "blocks": [ + { + "bbox": [ + 136, + 164, + 230, + 175 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 136, + 164, + 230, + 175 + ], + "spans": [ + { + "bbox": [ + 136, + 164, + 230, + 175 + ], + "type": "text", + "content": "1: for " + }, + { + "bbox": [ + 136, + 164, + 230, + 175 + ], + "type": "inline_equation", + "content": "c = 1 \\rightarrow C_t" + }, + { + "bbox": [ + 136, + 164, + 230, + 175 + ], + "type": "text", + "content": " do" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 136, + 175, + 320, + 186 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 136, + 175, + 320, + 186 + ], + "spans": [ + { + "bbox": [ + 136, + 175, + 320, + 186 + ], + "type": "text", + "content": "2: if " + }, + { + "bbox": [ + 136, + 175, + 320, + 186 + ], + "type": "inline_equation", + "content": "\\exists \\mu_{t,i,c} \\neq 0" + }, + { + "bbox": [ + 136, + 175, + 320, + 186 + ], + "type": "text", + "content": ", where " + }, + { + "bbox": [ + 136, + 175, + 320, + 186 + ], + "type": "inline_equation", + "content": "i = 1 \\rightarrow K" + }, + { + "bbox": [ + 136, + 175, + 320, + 186 + ], + "type": "text", + "content": " then" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 136, + 186, + 360, + 197 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 136, + 186, + 360, + 197 + ], + "spans": [ + { + "bbox": [ + 136, + 186, + 360, + 197 + ], + "type": "text", + "content": "3: " + }, + { + "bbox": [ + 136, + 186, + 360, + 197 + ], + "type": "inline_equation", + "content": "\\mu_{t,c}^{*}\\gets \\mu_{t,c}" + }, + { + "bbox": [ + 136, + 186, + 360, + 197 + ], + "type": "text", + "content": " retain all non-zero mean features" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 136, + 198, + 351, + 208 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 136, + 198, + 351, + 208 + ], + "spans": [ + { + "bbox": [ + 136, + 198, + 351, + 208 + ], + "type": "text", + "content": "4: " + }, + { + "bbox": [ + 136, + 198, + 351, + 208 + ], + "type": "inline_equation", + "content": "d_{t,k,c} \\gets \\text{Distance} \\left( m_{t,k}, \\mu_{t,c}^* \\right)" + }, + { + "bbox": [ + 136, + 198, + 351, + 208 + ], + "type": "text", + "content": " via Eq. (11)" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 136, + 209, + 323, + 219 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 136, + 209, + 323, + 219 + ], + "spans": [ + { + "bbox": [ + 136, + 209, + 323, + 219 + ], + "type": "text", + "content": "5: " + }, + { + "bbox": [ + 136, + 209, + 323, + 219 + ], + "type": "inline_equation", + "content": "p_{t,k,c} \\gets" + }, + { + "bbox": [ + 136, + 209, + 323, + 219 + ], + "type": "text", + "content": " take the reciprocal of " + }, + { + "bbox": [ + 136, + 209, + 323, + 219 + ], + "type": "inline_equation", + "content": "d_{t,k,c}" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 136, + 219, + 358, + 230 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 136, + 219, + 358, + 230 + ], + "spans": [ + { + "bbox": [ + 136, + 219, + 358, + 230 + ], + "type": "text", + "content": "6: " + }, + { + "bbox": [ + 136, + 219, + 358, + 230 + ], + "type": "inline_equation", + "content": "\\alpha_{t,k,c} \\gets \\text{Max-minNorm}(p_{t,k,c})" + }, + { + "bbox": [ + 136, + 219, + 358, + 230 + ], + "type": "text", + "content": " via Eq. (12)" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 136, + 231, + 344, + 241 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 136, + 231, + 344, + 241 + ], + "spans": [ + { + "bbox": [ + 136, + 231, + 344, + 241 + ], + "type": "text", + "content": "7: " + }, + { + "bbox": [ + 136, + 231, + 344, + 241 + ], + "type": "inline_equation", + "content": "\\omega_{t,k,c} \\gets \\text{Softmax}(\\alpha_{t,k,c}, \\eta)" + }, + { + "bbox": [ + 136, + 231, + 344, + 241 + ], + "type": "text", + "content": " via Eq. (13)" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 136, + 242, + 181, + 250 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 136, + 242, + 181, + 250 + ], + "spans": [ + { + "bbox": [ + 136, + 242, + 181, + 250 + ], + "type": "text", + "content": "8: else" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 136, + 251, + 343, + 262 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 136, + 251, + 343, + 262 + ], + "spans": [ + { + "bbox": [ + 136, + 251, + 343, + 262 + ], + "type": "text", + "content": "9: " + }, + { + "bbox": [ + 136, + 251, + 343, + 262 + ], + "type": "inline_equation", + "content": "\\omega_{t,k,c} \\gets" + }, + { + "bbox": [ + 136, + 251, + 343, + 262 + ], + "type": "text", + "content": " Assign with average value " + }, + { + "bbox": [ + 136, + 251, + 343, + 262 + ], + "type": "inline_equation", + "content": "1 / C_t" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 133, + 263, + 192, + 272 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 263, + 192, + 272 + ], + "spans": [ + { + "bbox": [ + 133, + 263, + 192, + 272 + ], + "type": "text", + "content": "10: end if" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 133, + 274, + 287, + 285 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 274, + 287, + 285 + ], + "spans": [ + { + "bbox": [ + 133, + 274, + 287, + 285 + ], + "type": "text", + "content": "11: " + }, + { + "bbox": [ + 133, + 274, + 287, + 285 + ], + "type": "inline_equation", + "content": "\\omega_{t,c}\\gets \\mathrm{List}(\\omega_{t,1,c},\\dots ,\\omega_{t,K,c})" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 133, + 285, + 340, + 296 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 285, + 340, + 296 + ], + "spans": [ + { + "bbox": [ + 133, + 285, + 340, + 296 + ], + "type": "text", + "content": "12: " + }, + { + "bbox": [ + 133, + 285, + 340, + 296 + ], + "type": "inline_equation", + "content": "m_{t,c}\\gets Re\\text{-weight} (\\omega_{t,c},m_{t,k})" + }, + { + "bbox": [ + 133, + 285, + 340, + 296 + ], + "type": "text", + "content": " via Eq. (14)" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 133, + 297, + 184, + 305 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 297, + 184, + 305 + ], + "spans": [ + { + "bbox": [ + 133, + 297, + 184, + 305 + ], + "type": "text", + "content": "13: end for" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 133, + 307, + 313, + 317 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 307, + 313, + 317 + ], + "spans": [ + { + "bbox": [ + 133, + 307, + 313, + 317 + ], + "type": "text", + "content": "14: Global prototype " + }, + { + "bbox": [ + 133, + 307, + 313, + 317 + ], + "type": "inline_equation", + "content": "m_t \\gets" + }, + { + "bbox": [ + 133, + 307, + 313, + 317 + ], + "type": "text", + "content": " Concat all " + }, + { + "bbox": [ + 133, + 307, + 313, + 317 + ], + "type": "inline_equation", + "content": "m_{t,c}" + } + ] + } + ], + "index": 18 + } + ], + "sub_type": "text" + }, + { + "bbox": [ + 130, + 342, + 481, + 390 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 342, + 481, + 390 + ], + "spans": [ + { + "bbox": [ + 130, + 342, + 481, + 390 + ], + "type": "inline_equation", + "content": "m_{t,k}" + }, + { + "bbox": [ + 130, + 342, + 481, + 390 + ], + "type": "text", + "content": " and the set of average features for the class they have learned " + }, + { + "bbox": [ + 130, + 342, + 481, + 390 + ], + "type": "inline_equation", + "content": "\\mu_{t,k}" + }, + { + "bbox": [ + 130, + 342, + 481, + 390 + ], + "type": "text", + "content": " (zero for classes without samples). On the global server, we first compute the sum of distances between the prototype " + }, + { + "bbox": [ + 130, + 342, + 481, + 390 + ], + "type": "inline_equation", + "content": "m_{t,k,c}" + }, + { + "bbox": [ + 130, + 342, + 481, + 390 + ], + "type": "text", + "content": " and the average feature " + }, + { + "bbox": [ + 130, + 342, + 481, + 390 + ], + "type": "inline_equation", + "content": "\\mu_{t,i,c}" + }, + { + "bbox": [ + 130, + 342, + 481, + 390 + ], + "type": "text", + "content": " uploaded by all clients:" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 241, + 390, + 481, + 422 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 241, + 390, + 481, + 422 + ], + "spans": [ + { + "bbox": [ + 241, + 390, + 481, + 422 + ], + "type": "interline_equation", + "content": "d _ {t, k, c} = \\sum_ {i = 1} ^ {K} \\| m _ {t, k, c} - \\mu_ {t, i, c} \\| _ {2} ^ {2}, \\tag {11}", + "image_path": "95411fdda007590476ab78105d51658782b20086f2335c3b6909ac93d747f638.jpg" + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 130, + 425, + 482, + 487 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 425, + 482, + 487 + ], + "spans": [ + { + "bbox": [ + 130, + 425, + 482, + 487 + ], + "type": "text", + "content": "the value of " + }, + { + "bbox": [ + 130, + 425, + 482, + 487 + ], + "type": "inline_equation", + "content": "d_{t,k,c}" + }, + { + "bbox": [ + 130, + 425, + 482, + 487 + ], + "type": "text", + "content": " approximates the distance from the prototype " + }, + { + "bbox": [ + 130, + 425, + 482, + 487 + ], + "type": "inline_equation", + "content": "m_{t,k,c}" + }, + { + "bbox": [ + 130, + 425, + 482, + 487 + ], + "type": "text", + "content": " to the overall features of class " + }, + { + "bbox": [ + 130, + 425, + 482, + 487 + ], + "type": "inline_equation", + "content": "c" + }, + { + "bbox": [ + 130, + 425, + 482, + 487 + ], + "type": "text", + "content": ". A smaller value indicates that the prototype is closer to the features of that class uploaded by all clients and during aggregation, we aim to assign it a higher weight. So we perform max-min normalization on the set " + }, + { + "bbox": [ + 130, + 425, + 482, + 487 + ], + "type": "inline_equation", + "content": "D_{t,c} = \\{p_{t,k,c}|k = 1,2,\\dots ,K\\}" + }, + { + "bbox": [ + 130, + 425, + 482, + 487 + ], + "type": "text", + "content": ", where " + }, + { + "bbox": [ + 130, + 425, + 482, + 487 + ], + "type": "inline_equation", + "content": "p_{t,k,c} = \\frac{1}{d_{t,k,c}}" + }, + { + "bbox": [ + 130, + 425, + 482, + 487 + ], + "type": "text", + "content": ":" + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 241, + 494, + 481, + 520 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 241, + 494, + 481, + 520 + ], + "spans": [ + { + "bbox": [ + 241, + 494, + 481, + 520 + ], + "type": "interline_equation", + "content": "\\alpha_ {t, k, c} = \\frac {p _ {t , k , c} - \\operatorname* {m i n} D _ {t , c}}{\\operatorname* {m a x} D _ {t , c} - \\operatorname* {m i n} D _ {t , c}}. \\tag {12}", + "image_path": "92da4f3a05dcab4c3a4f65822645cb2473c7029fb3bbc8955999931fe732fab5.jpg" + } + ] + } + ], + "index": 23 + }, + { + "bbox": [ + 130, + 525, + 481, + 549 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 525, + 481, + 549 + ], + "spans": [ + { + "bbox": [ + 130, + 525, + 481, + 549 + ], + "type": "text", + "content": "To meet the normalization requirements of the weights, we perform softmax processing on them and obtain the weight coefficient:" + } + ] + } + ], + "index": 24 + }, + { + "bbox": [ + 242, + 555, + 481, + 585 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 242, + 555, + 481, + 585 + ], + "spans": [ + { + "bbox": [ + 242, + 555, + 481, + 585 + ], + "type": "interline_equation", + "content": "\\omega_ {t, k, c} = \\frac {\\exp \\left(\\eta \\cdot \\alpha_ {t , k , c}\\right)}{\\sum_ {i = 1} ^ {K} \\exp \\left(\\eta \\cdot \\alpha_ {t , i , c}\\right)}, \\tag {13}", + "image_path": "223eb4674e53aac2f4611bca8537c2afc415b514c6f8efd372d154d36d67aa9d.jpg" + } + ] + } + ], + "index": 25 + }, + { + "bbox": [ + 130, + 593, + 481, + 629 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 593, + 481, + 629 + ], + "spans": [ + { + "bbox": [ + 130, + 593, + 481, + 629 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 130, + 593, + 481, + 629 + ], + "type": "inline_equation", + "content": "\\eta" + }, + { + "bbox": [ + 130, + 593, + 481, + 629 + ], + "type": "text", + "content": " is the temperature coefficient that controls the softness and hardness of weights. Finally, we re-weight all local prototypes corresponding to class " + }, + { + "bbox": [ + 130, + 593, + 481, + 629 + ], + "type": "inline_equation", + "content": "c" + }, + { + "bbox": [ + 130, + 593, + 481, + 629 + ], + "type": "text", + "content": " according to the obtained weights to get the global prototype of class " + }, + { + "bbox": [ + 130, + 593, + 481, + 629 + ], + "type": "inline_equation", + "content": "c" + }, + { + "bbox": [ + 130, + 593, + 481, + 629 + ], + "type": "text", + "content": ":" + } + ] + } + ], + "index": 26 + }, + { + "bbox": [ + 253, + 635, + 481, + 669 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 253, + 635, + 481, + 669 + ], + "spans": [ + { + "bbox": [ + 253, + 635, + 481, + 669 + ], + "type": "interline_equation", + "content": "m _ {t, c} = \\sum_ {i = 1} ^ {K} \\omega_ {t, i, c} \\cdot m _ {t, i, c}. \\tag {14}", + "image_path": "3cd20105875430e688c3760ae0e5b3f30263fa0e526465b78335e7d2f6a31a43.jpg" + } + ] + } + ], + "index": 27 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 411, + 91, + 447, + 100 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 411, + 91, + 447, + 100 + ], + "spans": [ + { + "bbox": [ + 411, + 91, + 447, + 100 + ], + "type": "text", + "content": "PILORA" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 474, + 91, + 481, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 474, + 91, + 481, + 100 + ], + "spans": [ + { + "bbox": [ + 474, + 91, + 481, + 100 + ], + "type": "text", + "content": "9" + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 8 + }, + { + "para_blocks": [ + { + "bbox": [ + 130, + 116, + 482, + 213 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 116, + 482, + 213 + ], + "spans": [ + { + "bbox": [ + 130, + 116, + 482, + 213 + ], + "type": "text", + "content": "We argue that prototype re-weight module allows the global prototypes " + }, + { + "bbox": [ + 130, + 116, + 482, + 213 + ], + "type": "inline_equation", + "content": "m_{t,c}" + }, + { + "bbox": [ + 130, + 116, + 482, + 213 + ], + "type": "text", + "content": " to differentially consider the data distribution information inherent in each local prototype, effectively mitigating the classifier bias issue. Compared with FedProto [51], our method further compresses the representation region of the same class in the feature space, which contributes to mitigate catastrophic forgetting. More importantly, our method does not require additional retraining of the classification layer, resulting in significant advantages in computational costs. Algorithm 1 presents the pseudo code of prototype re-weight module." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 131, + 230, + 326, + 243 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 131, + 230, + 326, + 243 + ], + "spans": [ + { + "bbox": [ + 131, + 230, + 326, + 243 + ], + "type": "text", + "content": "4.3 Integrated Objective of PILoRA" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 131, + 250, + 343, + 262 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 131, + 250, + 343, + 262 + ], + "spans": [ + { + "bbox": [ + 131, + 250, + 343, + 262 + ], + "type": "text", + "content": "The loss function of PILoRA can be defined as:" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 244, + 273, + 481, + 285 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 244, + 273, + 481, + 285 + ], + "spans": [ + { + "bbox": [ + 244, + 273, + 481, + 285 + ], + "type": "interline_equation", + "content": "l _ {t o t a l} = l _ {d c e} + \\lambda \\cdot l _ {p l} + \\gamma \\cdot l _ {o r t}, \\tag {15}", + "image_path": "959078a5890596ea10120324e6419f082cfdac94cdc9a79abf7f2e8befa91165.jpg" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 130, + 294, + 482, + 331 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 294, + 482, + 331 + ], + "spans": [ + { + "bbox": [ + 130, + 294, + 482, + 331 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 130, + 294, + 482, + 331 + ], + "type": "inline_equation", + "content": "\\lambda" + }, + { + "bbox": [ + 130, + 294, + 482, + 331 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 130, + 294, + 482, + 331 + ], + "type": "inline_equation", + "content": "\\gamma" + }, + { + "bbox": [ + 130, + 294, + 482, + 331 + ], + "type": "text", + "content": " are two hyper-parameters, and the overall framework of PILoRA is shown in Fig. 2. In general, our method is simple and effective, showing strong performance in different non-IID settings and degrees of data heterogeneity." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 131, + 349, + 230, + 363 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 131, + 349, + 230, + 363 + ], + "spans": [ + { + "bbox": [ + 131, + 349, + 230, + 363 + ], + "type": "text", + "content": "5 Experiments" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 131, + 374, + 266, + 387 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 131, + 374, + 266, + 387 + ], + "spans": [ + { + "bbox": [ + 131, + 374, + 266, + 387 + ], + "type": "text", + "content": "5.1 Experimental Setups" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 130, + 394, + 482, + 514 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 394, + 482, + 514 + ], + "spans": [ + { + "bbox": [ + 130, + 394, + 482, + 514 + ], + "type": "text", + "content": "Benchmark: To evaluate the proposed PILoRA, we perform our experiments on two well-known datasets: CIFAR-100 [24] and TinyImageNet [25], We also test the performance of the model on large-scale datasets, specifically, we randomly select 200 classes from the ImageNet-1k [8] as a new dataset. Following the protocols proposed in [33], we split 10 incremental stages and only the data from the current stage is available. Besides, to challenge our method, the local dataset of each client is followed by two kinds of non-IID settings: quantity-based label imbalance and distribution-based label imbalance [27] and we denote the degree of heterogeneity of the two settings by " + }, + { + "bbox": [ + 130, + 394, + 482, + 514 + ], + "type": "inline_equation", + "content": "\\alpha" + }, + { + "bbox": [ + 130, + 394, + 482, + 514 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 130, + 394, + 482, + 514 + ], + "type": "inline_equation", + "content": "\\beta" + }, + { + "bbox": [ + 130, + 394, + 482, + 514 + ], + "type": "text", + "content": ". Details of both settings can be seen in Section A in the Supplementary Material." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 130, + 514, + 482, + 609 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 514, + 482, + 609 + ], + "spans": [ + { + "bbox": [ + 130, + 514, + 482, + 609 + ], + "type": "text", + "content": "Comparison: We compare our method with existing FCIL methods4: TARGET [61], GLFC [10], LGA [9]. we also adopt several CIL methods: EWC [23], LwF [29], iCaRL [46], L2P [55] and FL method FedNCM [26] in FCIL setting. In addition, we compare the use of cross-entropy loss for optimization during training and the use of class means as classifiers during inference, which we named FedCLM. We explore their performance in different non-IID settings and degrees of data heterogeneity. For a fair comparison, we tune all methods to the same pre-trained model as ours and fine-tune them using LoRA." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 130, + 610, + 482, + 647 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 610, + 482, + 647 + ], + "spans": [ + { + "bbox": [ + 130, + 610, + 482, + 647 + ], + "type": "text", + "content": "Implementation: Considering privacy issues, we evaluate the performance of our method on self-supervised pre-trained weights (Dino [5]) for ViT-B/16 [11], this setting is also widely used in CIL task [36,37,60]. Considering the trade-off" + } + ] + } + ], + "index": 11 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "spans": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "type": "text", + "content": "10" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 166, + 91, + 220, + 101 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 166, + 91, + 220, + 101 + ], + "spans": [ + { + "bbox": [ + 166, + 91, + 220, + 101 + ], + "type": "text", + "content": "H. Guo et al." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 133, + 653, + 460, + 666 + ], + "type": "page_footnote", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 653, + 460, + 666 + ], + "spans": [ + { + "bbox": [ + 133, + 653, + 460, + 666 + ], + "type": "text", + "content": "4 GLFC and LGA only perform experiments on quantity-based label imbalance." + } + ] + } + ], + "index": 12 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 9 + }, + { + "para_blocks": [ + { + "type": "table", + "bbox": [ + 143, + 146, + 473, + 306 + ], + "blocks": [ + { + "bbox": [ + 132, + 114, + 480, + 137 + ], + "lines": [ + { + "bbox": [ + 132, + 114, + 480, + 137 + ], + "spans": [ + { + "bbox": [ + 132, + 114, + 480, + 137 + ], + "type": "text", + "content": "Table 1: Results (%) on CIFAR-100. Results are included for 10 tasks (10 classes per task) and under different degree of two non-IID setting." + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 143, + 146, + 473, + 306 + ], + "lines": [ + { + "bbox": [ + 143, + 146, + 473, + 306 + ], + "spans": [ + { + "bbox": [ + 143, + 146, + 473, + 306 + ], + "type": "table", + "html": "
Non-IIDQuantity-based label imbalanceDistribution-based label imbalance
Partitionα = 6α = 4α = 2β = 0.5β = 0.1β = 0.05
MethodsANAvg.ANAvg.ANAvg.ANAvg.ANAvg.ANAvg.
Joint88.6-84.3-79.8-90.1-87.8-85.9-
EWC+FL57.969.155.966.842.252.765.577.857.873.243.559.2
LwF+FL57.468.855.166.740.852.964.777.554.663.345.764.5
iCaRL+FL35.856.537.158.943.455.351.367.750.165.944.663.0
L2P+FL63.465.159.058.22.65.653.951.662.971.438.732.2
FedCLM58.969.457.667.644.357.966.577.461.071.848.863.5
FedNCM65.674.461.971.149.659.866.877.962.172.450.965.9
TARGET60.971.358.869.545.256.566.177.860.571.151.865.3
GLFC58.270.453.765.913.137.7------
LGA64.573.661.170.521.640.9------
Ours69.578.665.174.454.962.668.578.163.473.754.867.1
Ours+HT69.678.565.874.856.564.370.278.663.673.857.969.2
", + "image_path": "e402a7b0ebb20fce09c17ba1dc4be8770dd38242ab2d48b9672be3014c56c1da.jpg" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "table_body" + } + ], + "index": 3 + }, + { + "bbox": [ + 130, + 327, + 482, + 426 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 327, + 482, + 426 + ], + "spans": [ + { + "bbox": [ + 130, + 327, + 482, + 426 + ], + "type": "text", + "content": "between performance and number of parameters, we insert the LoRA module into only the first block (See Section B.1) and set " + }, + { + "bbox": [ + 130, + 327, + 482, + 426 + ], + "type": "inline_equation", + "content": "r = 4" + }, + { + "bbox": [ + 130, + 327, + 482, + 426 + ], + "type": "text", + "content": ". We train our models using Adam [21] with a batch size of 64 and followed by [60], we adopt different learning rates for prototypes layer of " + }, + { + "bbox": [ + 130, + 327, + 482, + 426 + ], + "type": "inline_equation", + "content": "2e^{-3}" + }, + { + "bbox": [ + 130, + 327, + 482, + 426 + ], + "type": "text", + "content": " and LoRA parameters of " + }, + { + "bbox": [ + 130, + 327, + 482, + 426 + ], + "type": "inline_equation", + "content": "1e^{-5}" + }, + { + "bbox": [ + 130, + 327, + 482, + 426 + ], + "type": "text", + "content": " on CIFAR-100, and " + }, + { + "bbox": [ + 130, + 327, + 482, + 426 + ], + "type": "inline_equation", + "content": "5e^{-3} / 5e^{-6}" + }, + { + "bbox": [ + 130, + 327, + 482, + 426 + ], + "type": "text", + "content": " on TinyImageNet. Moreover, cosine annealing is also used in training processes. We set " + }, + { + "bbox": [ + 130, + 327, + 482, + 426 + ], + "type": "inline_equation", + "content": "\\delta = 1" + }, + { + "bbox": [ + 130, + 327, + 482, + 426 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 130, + 327, + 482, + 426 + ], + "type": "inline_equation", + "content": "\\lambda = 0.001" + }, + { + "bbox": [ + 130, + 327, + 482, + 426 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 130, + 327, + 482, + 426 + ], + "type": "inline_equation", + "content": "\\gamma = 0.5" + }, + { + "bbox": [ + 130, + 327, + 482, + 426 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 130, + 327, + 482, + 426 + ], + "type": "inline_equation", + "content": "\\eta = 0.2" + }, + { + "bbox": [ + 130, + 327, + 482, + 426 + ], + "type": "text", + "content": ". We initialize 10 local clients to train and upload the parameters at each communication round. The local training epoch is 5 and the communication round is 30." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 132, + 441, + 266, + 454 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 441, + 266, + 454 + ], + "spans": [ + { + "bbox": [ + 132, + 441, + 266, + 454 + ], + "type": "text", + "content": "5.2 Comparative Results" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 130, + 462, + 482, + 605 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 462, + 482, + 605 + ], + "spans": [ + { + "bbox": [ + 130, + 462, + 482, + 605 + ], + "type": "text", + "content": "We report " + }, + { + "bbox": [ + 130, + 462, + 482, + 605 + ], + "type": "inline_equation", + "content": "A_{N}(\\uparrow)" + }, + { + "bbox": [ + 130, + 462, + 482, + 605 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 130, + 462, + 482, + 605 + ], + "type": "inline_equation", + "content": "\\mathrm{Avg}(\\uparrow)" + }, + { + "bbox": [ + 130, + 462, + 482, + 605 + ], + "type": "text", + "content": " to evaluate the performance of the methods, where " + }, + { + "bbox": [ + 130, + 462, + 482, + 605 + ], + "type": "inline_equation", + "content": "A_{N}" + }, + { + "bbox": [ + 130, + 462, + 482, + 605 + ], + "type": "text", + "content": " is the accuracy of all seen classes in the final task and Avg. is calculated as the average accuracy of all tasks. Results are shown in Tab. 1 and Tab. 2, we can obverse that our method outperforms other comparison methods and demonstrates strong robustness across different non-IID settings. Among FCIL methods, TARGET achieves relatively better performance under different data heterogeneity, while LGA and GLFC show a steep drop in performance and lack of robustness under extreme data heterogeneity. Notably, FedNCM in the FCIL setting even outperforms TARGET in some cases, which suggests that existing FCIL methods lack the handling of data heterogeneity, whereas our approach dramatically improves the performance of the model under different data heterogeneity through a lightweight prototype re-weight module." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 130, + 606, + 482, + 666 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 606, + 482, + 666 + ], + "spans": [ + { + "bbox": [ + 130, + 606, + 482, + 666 + ], + "type": "text", + "content": "The CIL methods also suffer from data heterogeneity in the FCIL setting, with L2P+FL being the most obvious. We believe that there are two main reasons for this, firstly the performance of L2P relies on supervised pre-trained weights (e.g., ImageNet-1k [8]) and its performance degrades significantly when using self-supervised pre-trained weights [60]. Secondly, L2P needs to choose the" + } + ] + } + ], + "index": 7 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 411, + 91, + 447, + 100 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 411, + 91, + 447, + 100 + ], + "spans": [ + { + "bbox": [ + 411, + 91, + 447, + 100 + ], + "type": "text", + "content": "PILORA" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 470, + 91, + 480, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 470, + 91, + 480, + 100 + ], + "spans": [ + { + "bbox": [ + 470, + 91, + 480, + 100 + ], + "type": "text", + "content": "11" + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 10 + }, + { + "para_blocks": [ + { + "type": "table", + "bbox": [ + 140, + 146, + 476, + 274 + ], + "blocks": [ + { + "bbox": [ + 132, + 114, + 482, + 137 + ], + "lines": [ + { + "bbox": [ + 132, + 114, + 482, + 137 + ], + "spans": [ + { + "bbox": [ + 132, + 114, + 482, + 137 + ], + "type": "text", + "content": "Table 2: Results (\\%) on TinyImageNet. Results are included for 10 tasks (20 classes per task) and under different degree of two non-IID setting." + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 140, + 146, + 476, + 274 + ], + "lines": [ + { + "bbox": [ + 140, + 146, + 476, + 274 + ], + "spans": [ + { + "bbox": [ + 140, + 146, + 476, + 274 + ], + "type": "table", + "html": "
Non-IIDQuantity-based label imbalanceDistribution-based label imbalance
Partitionα = 12α = 8α = 4β = 0.5β = 0.1β = 0.05
MethodsANAvg.ANAvg.ANAvg.ANAvg.ANAvg.ANAvg.
Joint83.6-82.9-80.2-84.3-83.3-82.8-
L2P+FL61.658.049.439.38.210.264.266.956.352.551.943.2
FedCLM61.672.451.860.345.856.966.577.460.471.046.757.8
FedNCM71.681.669.579.457.264.773.781.670.880.468.478.0
TARGET72.681.670.379.663.873.571.680.971.080.169.379.1
GLFC69.177.961.373.525.139.4------
LGA71.379.465.875.336.748.8------
Ours74.481.074.380.970.177.874.580.974.381.073.680.2
Ours+HT75.081.374.981.267.975.674.781.074.380.973.480.1
", + "image_path": "115cf2bab88c315fa029c5f4c196ce03bec8502f5c98f8229d93f98170441386.jpg" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "table_body" + } + ], + "index": 3 + }, + { + "type": "image", + "bbox": [ + 140, + 289, + 233, + 365 + ], + "blocks": [ + { + "bbox": [ + 140, + 289, + 233, + 365 + ], + "lines": [ + { + "bbox": [ + 140, + 289, + 233, + 365 + ], + "spans": [ + { + "bbox": [ + 140, + 289, + 233, + 365 + ], + "type": "image", + "image_path": "6c9ab0bf0ffc2e44905efff2593d0aeb5534d72e962d2d5bd3c8839ecf6c8885.jpg" + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 228, + 370, + 241, + 380 + ], + "lines": [ + { + "bbox": [ + 228, + 370, + 241, + 380 + ], + "spans": [ + { + "bbox": [ + 228, + 370, + 241, + 380 + ], + "type": "text", + "content": "(a)" + } + ] + } + ], + "index": 6, + "angle": 0, + "type": "image_caption" + } + ], + "index": 4 + }, + { + "type": "image", + "bbox": [ + 235, + 289, + 329, + 365 + ], + "blocks": [ + { + "bbox": [ + 235, + 289, + 329, + 365 + ], + "lines": [ + { + "bbox": [ + 235, + 289, + 329, + 365 + ], + "spans": [ + { + "bbox": [ + 235, + 289, + 329, + 365 + ], + "type": "image", + "image_path": "ccf7e08d3b18cfc5db266ea983653c1b9f6c324715f101ce31d773aa9c456466.jpg" + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "image_body" + } + ], + "index": 5 + }, + { + "type": "image", + "bbox": [ + 346, + 293, + 474, + 366 + ], + "blocks": [ + { + "bbox": [ + 346, + 293, + 474, + 366 + ], + "lines": [ + { + "bbox": [ + 346, + 293, + 474, + 366 + ], + "spans": [ + { + "bbox": [ + 346, + 293, + 474, + 366 + ], + "type": "image", + "image_path": "50c43133f5895e13f792703c80e5366ab30e0bb8b19af7e025358fffc3f14dcc.jpg" + } + ] + } + ], + "index": 7, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 404, + 370, + 417, + 380 + ], + "lines": [ + { + "bbox": [ + 404, + 370, + 417, + 380 + ], + "spans": [ + { + "bbox": [ + 404, + 370, + 417, + 380 + ], + "type": "text", + "content": "(b)" + } + ] + } + ], + "index": 8, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 130, + 388, + 482, + 434 + ], + "lines": [ + { + "bbox": [ + 130, + 388, + 482, + 434 + ], + "spans": [ + { + "bbox": [ + 130, + 388, + 482, + 434 + ], + "type": "text", + "content": "Fig. 3: (a) Comparison of real proportions of the same class among different clients and the weights obtained from Prototype Re-weighting during a local training process. (b) The " + }, + { + "bbox": [ + 130, + 388, + 482, + 434 + ], + "type": "inline_equation", + "content": "L2" + }, + { + "bbox": [ + 130, + 388, + 482, + 434 + ], + "type": "text", + "content": " distance between classes feature extracted by different clients and global model with corresponding prototypes." + } + ] + } + ], + "index": 9, + "angle": 0, + "type": "image_caption" + } + ], + "index": 7 + }, + { + "bbox": [ + 130, + 460, + 482, + 592 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 460, + 482, + 592 + ], + "spans": [ + { + "bbox": [ + 130, + 460, + 482, + 592 + ], + "type": "text", + "content": "appropriate prompt for embedding in the model based on the similarity calculation. However, in cases of extreme heterogeneity, there exists a significant disparity in what each client learns, even at the same stage. This discrepancy prevents the selection of a uniform prompt for reasoning. Instead, our method bypasses redundant similarity calculations or knowledge distillation [29] and effectively mitigates catastrophic forgetting directly through the summation of LoRA parameters. In addition, we combine the HeadTune in FedNCM with our prototype classifier, and the performance of our model is further improved by the better initialization provided by HeadTune. This suggests that there is still a lot of space for improvement in the existing methods in solving the problem of data heterogeneity in FCIL." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 130, + 594, + 482, + 666 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 594, + 482, + 666 + ], + "spans": [ + { + "bbox": [ + 130, + 594, + 482, + 666 + ], + "type": "text", + "content": "Analysis of Prototype Re-weight: The core idea of our proposed prototype re-weight module is to aggregate prototypes in a manner that closely resembles the true data distribution among clients without disclosing local data information. In Fig. 3a, we show the real proportions of the same class of data across clients and we can observe that the weights calculated through re-weight module align with the distribution of data among different clients. Hence, the global" + } + ] + } + ], + "index": 11 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "spans": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "type": "text", + "content": "12" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 166, + 91, + 220, + 100 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 166, + 91, + 220, + 100 + ], + "spans": [ + { + "bbox": [ + 166, + 91, + 220, + 100 + ], + "type": "text", + "content": "H. Guo et al." + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 11 + }, + { + "para_blocks": [ + { + "type": "table", + "bbox": [ + 195, + 157, + 422, + 235 + ], + "blocks": [ + { + "bbox": [ + 130, + 114, + 482, + 149 + ], + "lines": [ + { + "bbox": [ + 130, + 114, + 482, + 149 + ], + "spans": [ + { + "bbox": [ + 130, + 114, + 482, + 149 + ], + "type": "text", + "content": "Table 3: Ablation Results (\\%) on CIFAR-100. Experimental results based on self-supervised pre-trained weights. " + }, + { + "bbox": [ + 130, + 114, + 482, + 149 + ], + "type": "inline_equation", + "content": "F_N(\\downarrow)" + }, + { + "bbox": [ + 130, + 114, + 482, + 149 + ], + "type": "text", + "content": " is the average forgetting and the calculation process can be referred to [66]." + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 195, + 157, + 422, + 235 + ], + "lines": [ + { + "bbox": [ + 195, + 157, + 422, + 235 + ], + "spans": [ + { + "bbox": [ + 195, + 157, + 422, + 235 + ], + "type": "table", + "html": "
Partitionα = 6β = 0.5
MethodsANAvg.FNANAvg.FN
Ours-w/o PR26.830.95.541.448.89.3
Ours-w/o lort68.777.910.367.177.011.1
Ours-w/o LoRA (fintune)11.329.57.313.329.98.9
Ours-w/o LoRA (frozen)68.677.710.067.877.211.6
Ours69.578.69.568.578.110.6
", + "image_path": "503cb48cbe8f52b4b2c62e3e5b8cb585a7364c393e0eae1bafddacb4228e2d4b.jpg" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "table_body" + } + ], + "index": 3 + }, + { + "bbox": [ + 130, + 257, + 482, + 389 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 257, + 482, + 389 + ], + "spans": [ + { + "bbox": [ + 130, + 257, + 482, + 389 + ], + "type": "text", + "content": "prototypes effectively retain all learned prototype information. In contrast, other methods typically use direct averaging for the classification layer, leading to the fusion of a considerable amount of irrelevant information in the classifier when the data heterogeneity is high, consequently causing classifier bias. In Fig. 3b, we compare the distances between the prototype calculated by both prototype re-weight and average for the global model and the deep features of the test data. It can be seen that our proposed prototype re-weight method effectively 'takes the best' of all client uploaded prototypes, so that the global prototype is better adapted to the features of the corresponding class, whereas on average the correct prototype is far away from the features of the corresponding class, especially when the data is very heterogeneous." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 131, + 406, + 238, + 418 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 131, + 406, + 238, + 418 + ], + "spans": [ + { + "bbox": [ + 131, + 406, + 238, + 418 + ], + "type": "text", + "content": "5.3 Ablation Study" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 130, + 426, + 482, + 593 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 426, + 482, + 593 + ], + "spans": [ + { + "bbox": [ + 130, + 426, + 482, + 593 + ], + "type": "text", + "content": "To evaluate the effect of each component in PILoRA, we perform the ablation study and show the results of 10 phases set in CIFAR-100 in Tab. 3. We can observe that without prototype re-weight (PR), The classification accuracy " + }, + { + "bbox": [ + 130, + 426, + 482, + 593 + ], + "type": "inline_equation", + "content": "A_{N}" + }, + { + "bbox": [ + 130, + 426, + 482, + 593 + ], + "type": "text", + "content": " and Avg. of the model decrease significantly, which is due to the fact that the global prototype obtained by simply averaging at this point does not represent the information of each class well, which leads to classifier bias. Instead, our proposed method significantly improves the performance of the model by heuristically reweight the local prototypes. When there is no orthogonal regularization " + }, + { + "bbox": [ + 130, + 426, + 482, + 593 + ], + "type": "inline_equation", + "content": "(l_{ort})" + }, + { + "bbox": [ + 130, + 426, + 482, + 593 + ], + "type": "text", + "content": ", all the metrics of the model show a certain decrease, which indicates that there is a partial overlap between the parameter spaces of different incremental tasks. By imposing orthogonal regularization the parameter spaces corresponding to different input spaces can be made to train the model in the direction of orthogonality to each other, which further improves the model performance. We visualize the cosine similarity between LoRA in Section B.2." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 130, + 594, + 482, + 666 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 594, + 482, + 666 + ], + "spans": [ + { + "bbox": [ + 130, + 594, + 482, + 666 + ], + "type": "text", + "content": "To better demonstrate the contribution of LoRA, we compare the effects of fine-tuning the entire backbone and freezing the entire backbone under self-supervised pre-trained weights when " + }, + { + "bbox": [ + 130, + 594, + 482, + 666 + ], + "type": "inline_equation", + "content": "\\alpha = 6" + }, + { + "bbox": [ + 130, + 594, + 482, + 666 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 130, + 594, + 482, + 666 + ], + "type": "inline_equation", + "content": "\\beta = 0.5" + }, + { + "bbox": [ + 130, + 594, + 482, + 666 + ], + "type": "text", + "content": ". As can be seen in Tab. 3, when fine-tuning the entire backbone, the performance of the model is severely degraded, we believe that this is due to the fact that the model parameters of ViT are too large to achieve a balance between the ability to remember old" + } + ] + } + ], + "index": 7 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 411, + 91, + 447, + 100 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 411, + 91, + 447, + 100 + ], + "spans": [ + { + "bbox": [ + 411, + 91, + 447, + 100 + ], + "type": "text", + "content": "PILORA" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 470, + 91, + 481, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 470, + 91, + 481, + 100 + ], + "spans": [ + { + "bbox": [ + 470, + 91, + 481, + 100 + ], + "type": "text", + "content": "13" + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 12 + }, + { + "para_blocks": [ + { + "bbox": [ + 130, + 116, + 482, + 200 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 116, + 482, + 200 + ], + "spans": [ + { + "bbox": [ + 130, + 116, + 482, + 200 + ], + "type": "text", + "content": "classes and the ability to discriminate new classes under the constraint of knowledge distillation. Freezing the entire backbone network can retain discriminative ability for old classes, but at this point, the model has too few trainable parameters, with only prototypes used for classification, thereby limiting the expressive capacity of the model. Therefore, to achieve a trade-off between communication cost and model performance, we fine-tune and achieve the best performance with a tiny number of parameters through Incremental LoRA." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 132, + 215, + 246, + 227 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 215, + 246, + 227 + ], + "spans": [ + { + "bbox": [ + 132, + 215, + 246, + 227 + ], + "type": "text", + "content": "5.4 Further Analysis" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 130, + 232, + 298, + 421 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 232, + 298, + 421 + ], + "spans": [ + { + "bbox": [ + 130, + 232, + 298, + 421 + ], + "type": "text", + "content": "Memory usage analysis. For our PILoRA, in addition to the model for each client, we store the LoRA parameters " + }, + { + "bbox": [ + 130, + 232, + 298, + 421 + ], + "type": "inline_equation", + "content": "A_{q}^{1:t - 1}" + }, + { + "bbox": [ + 130, + 232, + 298, + 421 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 130, + 232, + 298, + 421 + ], + "type": "inline_equation", + "content": "A_{v}^{1:t - 1}" + }, + { + "bbox": [ + 130, + 232, + 298, + 421 + ], + "type": "text", + "content": " up to the current stage " + }, + { + "bbox": [ + 130, + 232, + 298, + 421 + ], + "type": "inline_equation", + "content": "t" + }, + { + "bbox": [ + 130, + 232, + 298, + 421 + ], + "type": "text", + "content": " to compute the orthogonal regularization. Whereas TARGET requires an additional generator to generate old samples, as well as additional memory space to store the generated images, LGA and GLFC similarly take up additional space to store old class samples. Compared to them, our method takes up very little memory space, storing on average only " + }, + { + "bbox": [ + 130, + 232, + 298, + 421 + ], + "type": "inline_equation", + "content": "0.04\\%" + }, + { + "bbox": [ + 130, + 232, + 298, + 421 + ], + "type": "text", + "content": " of the parameters equivalent to the ViT-base." + } + ] + } + ], + "index": 4 + }, + { + "type": "image", + "bbox": [ + 305, + 258, + 391, + 342 + ], + "blocks": [ + { + "bbox": [ + 305, + 258, + 391, + 342 + ], + "lines": [ + { + "bbox": [ + 305, + 258, + 391, + 342 + ], + "spans": [ + { + "bbox": [ + 305, + 258, + 391, + 342 + ], + "type": "image", + "image_path": "eba9aef916c37020ca0159111514aacd6192e007e387183f7e255efe8119c05e.jpg" + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 304, + 354, + 482, + 387 + ], + "lines": [ + { + "bbox": [ + 304, + 354, + 482, + 387 + ], + "spans": [ + { + "bbox": [ + 304, + 354, + 482, + 387 + ], + "type": "text", + "content": "Fig.4: The impact of different number of " + }, + { + "bbox": [ + 304, + 354, + 482, + 387 + ], + "type": "inline_equation", + "content": "K" + }, + { + "bbox": [ + 304, + 354, + 482, + 387 + ], + "type": "text", + "content": " on CIFAR-100, where we consider " + }, + { + "bbox": [ + 304, + 354, + 482, + 387 + ], + "type": "inline_equation", + "content": "\\alpha = 6" + }, + { + "bbox": [ + 304, + 354, + 482, + 387 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 304, + 354, + 482, + 387 + ], + "type": "inline_equation", + "content": "\\beta = 0.5" + }, + { + "bbox": [ + 304, + 354, + 482, + 387 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 7, + "angle": 0, + "type": "image_caption" + } + ], + "index": 5 + }, + { + "type": "image", + "bbox": [ + 394, + 258, + 479, + 342 + ], + "blocks": [ + { + "bbox": [ + 394, + 258, + 479, + 342 + ], + "lines": [ + { + "bbox": [ + 394, + 258, + 479, + 342 + ], + "spans": [ + { + "bbox": [ + 394, + 258, + 479, + 342 + ], + "type": "image", + "image_path": "ae690ecd148e5ca54d747d9d26469f771e9b33fc02aa59931e30013d2a665802.jpg" + } + ] + } + ], + "index": 6, + "angle": 0, + "type": "image_body" + } + ], + "index": 6 + }, + { + "bbox": [ + 130, + 423, + 482, + 509 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 423, + 482, + 509 + ], + "spans": [ + { + "bbox": [ + 130, + 423, + 482, + 509 + ], + "type": "text", + "content": "Increase the number of local clients " + }, + { + "bbox": [ + 130, + 423, + 482, + 509 + ], + "type": "inline_equation", + "content": "(K)" + }, + { + "bbox": [ + 130, + 423, + 482, + 509 + ], + "type": "text", + "content": ". As depicted in Fig. 4, we investigate the performance of the model as the number of local clients increase, by respectively setting " + }, + { + "bbox": [ + 130, + 423, + 482, + 509 + ], + "type": "inline_equation", + "content": "K = \\{10, 15, 20\\}" + }, + { + "bbox": [ + 130, + 423, + 482, + 509 + ], + "type": "text", + "content": ". From the results shown in Fig. 4, we observe that the model's performance slightly declines as the number of clients increases. We believe that under the same non-IID setting, enlarging the number of clients further exacerbates the heterogeneity among clients, consequently impacting the model's performance. More results can be seen in Section B.3." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 132, + 523, + 220, + 536 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 523, + 220, + 536 + ], + "spans": [ + { + "bbox": [ + 132, + 523, + 220, + 536 + ], + "type": "text", + "content": "6 Conclusion" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 130, + 545, + 482, + 666 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 545, + 482, + 666 + ], + "spans": [ + { + "bbox": [ + 130, + 545, + 482, + 666 + ], + "type": "text", + "content": "In this paper, we propose a simple and effective method of " + }, + { + "bbox": [ + 130, + 545, + 482, + 666 + ], + "type": "inline_equation", + "content": "PILoRA" + }, + { + "bbox": [ + 130, + 545, + 482, + 666 + ], + "type": "text", + "content": " for FCIL. PILoRA is based on pre-trained ViT models and fine-tunes with a tiny number of parameters using LoRA. To address the catastrophic forgetting problem in FCIL, we propose incremental LoRA, which can efficiently combine different incremental tasks by summing the orthogonal LoRA parameter space; To deal with the classifier bias caused by data heterogeneity, we adopt prototype learning and propose prototype re-weight, which utilize heuristic information between prototypes and features to perform weighted aggregation of global prototypes. Experimental results show that our method achieves state-of-the-art results on standard datasets and maintains robustness under extreme data heterogeneity." + } + ] + } + ], + "index": 10 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "spans": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "type": "text", + "content": "14" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 166, + 91, + 220, + 101 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 166, + 91, + 220, + 101 + ], + "spans": [ + { + "bbox": [ + 166, + 91, + 220, + 101 + ], + "type": "text", + "content": "H. Guo et al." + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 13 + }, + { + "para_blocks": [ + { + "bbox": [ + 133, + 114, + 246, + 129 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 114, + 246, + 129 + ], + "spans": [ + { + "bbox": [ + 133, + 114, + 246, + 129 + ], + "type": "text", + "content": "Acknowledgements" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 130, + 140, + 482, + 201 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 140, + 482, + 201 + ], + "spans": [ + { + "bbox": [ + 130, + 140, + 482, + 201 + ], + "type": "text", + "content": "This work has been supported by the National Science and Technology Major Project (2022ZD0116500), National Natural Science Foundation of China (U20A20223, 62222609, 62076236), CAS Project for Young Scientists in Basic Research (YSBR-083), and Key Research Program of Frontier Sciences of CAS (ZDBS-LY-7004) and the InnoHK program." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 132, + 220, + 197, + 232 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 220, + 197, + 232 + ], + "spans": [ + { + "bbox": [ + 132, + 220, + 197, + 232 + ], + "type": "text", + "content": "References" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 134, + 246, + 480, + 665 + ], + "type": "list", + "angle": 0, + "index": 18, + "blocks": [ + { + "bbox": [ + 138, + 246, + 480, + 280 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 246, + 480, + 280 + ], + "spans": [ + { + "bbox": [ + 138, + 246, + 480, + 280 + ], + "type": "text", + "content": "1. Bagwe, G., Yuan, X., Pan, M., Zhang, L.: Fed-cprompt: Contrastive prompt for rehearsal-free federated continual learning. In: Federated Learning and Analytics in Practice: Algorithms, Systems, Applications, and Opportunities (2023)" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 138, + 281, + 480, + 312 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 281, + 480, + 312 + ], + "spans": [ + { + "bbox": [ + 138, + 281, + 480, + 312 + ], + "type": "text", + "content": "2. Belouadah, E., Popescu, A.: Deesil: Deep-shallow incremental learning. In: Proceedings of the European Conference on Computer Vision (ECCV) Workshops. pp. 0-0 (2018)" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 138, + 313, + 480, + 345 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 313, + 480, + 345 + ], + "spans": [ + { + "bbox": [ + 138, + 313, + 480, + 345 + ], + "type": "text", + "content": "3. Belouadah, E., Popescu, A.: II2m: Class incremental learning with dual memory. In: Proceedings of the IEEE/CVF international conference on computer vision. pp. 583-592 (2019)" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 138, + 346, + 480, + 378 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 346, + 480, + 378 + ], + "spans": [ + { + "bbox": [ + 138, + 346, + 480, + 378 + ], + "type": "text", + "content": "4. Brown, T., Mann, B., Ryder, N., Subbiah, M., Kaplan, J.D., Dhariwal, P., Neelakantan, A., Shyam, P., Sastry, G., Askell, A., et al.: Language models are few-shot learners. Advances in neural information processing systems 33, 1877-1901 (2020)" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 138, + 379, + 480, + 411 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 379, + 480, + 411 + ], + "spans": [ + { + "bbox": [ + 138, + 379, + 480, + 411 + ], + "type": "text", + "content": "5. Caron, M., Touvron, H., Misra, I., Jégou, H., Mairal, J., Bojanowski, P., Joulin, A.: Emerging properties in self-supervised vision transformers. In: Proceedings of the International Conference on Computer Vision (ICCV) (2021)" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 138, + 412, + 480, + 434 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 412, + 480, + 434 + ], + "spans": [ + { + "bbox": [ + 138, + 412, + 480, + 434 + ], + "type": "text", + "content": "6. Cheng, Z., Zhang, X.Y., Liu, C.L.: Unified classification and rejection: A one-versus-all framework. arXiv preprint arXiv:2311.13355 (2023)" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 138, + 434, + 480, + 456 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 434, + 480, + 456 + ], + "spans": [ + { + "bbox": [ + 138, + 434, + 480, + 456 + ], + "type": "text", + "content": "7. Chitale, R., Vaidya, A., Kane, A., Ghotkar, A.: Task arithmetic with lora for continual learning. arXiv preprint arXiv:2311.02428 (2023)" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 138, + 456, + 480, + 489 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 456, + 480, + 489 + ], + "spans": [ + { + "bbox": [ + 138, + 456, + 480, + 489 + ], + "type": "text", + "content": "8. Deng, J., Dong, W., Socher, R., Li, L.J., Li, K., Fei-Fei, L.: Imagenet: A large-scale hierarchical image database. In: 2009 IEEE conference on computer vision and pattern recognition. pp. 248-255. IEEE (2009)" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 138, + 490, + 480, + 521 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 490, + 480, + 521 + ], + "spans": [ + { + "bbox": [ + 138, + 490, + 480, + 521 + ], + "type": "text", + "content": "9. Dong, J., Cong, Y., Sun, G., Zhang, Y., Schiele, B., Dai, D.: No one left behind: Real-world federated class-incremental learning. arXiv preprint arXiv:2302.00903 (2023)" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 138, + 522, + 480, + 555 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 522, + 480, + 555 + ], + "spans": [ + { + "bbox": [ + 138, + 522, + 480, + 555 + ], + "type": "text", + "content": "0. Dong, J., Wang, L., Fang, Z., Sun, G., Xu, S., Wang, X., Zhu, Q.: Federated class incremental learning. In: Proceedings of the IEEE/CVF conference on computer vision and pattern recognition. pp. 10164-10173 (2022)" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 134, + 555, + 480, + 599 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 134, + 555, + 480, + 599 + ], + "spans": [ + { + "bbox": [ + 134, + 555, + 480, + 599 + ], + "type": "text", + "content": "11. Dosovitskiy, A., Beyer, L., Kolesnikov, A., Weissenborn, D., Zhai, X., Unterthiner, T., Dehghani, M., Minderer, M., Heigold, G., Gelly, S., et al.: An image is worth 16x16 words: Transformers for image recognition at scale. arXiv preprint arXiv:2010.11929 (2020)" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 134, + 600, + 480, + 632 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 134, + 600, + 480, + 632 + ], + "spans": [ + { + "bbox": [ + 134, + 600, + 480, + 632 + ], + "type": "text", + "content": "12. Gao, Q., Zhao, C., Sun, Y., Xi, T., Zhang, G., Ghanem, B., Zhang, J.: A unified continual learning framework with general parameter-efficient tuning. arXiv preprint arXiv:2303.10070 (2023)" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 134, + 633, + 480, + 665 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 134, + 633, + 480, + 665 + ], + "spans": [ + { + "bbox": [ + 134, + 633, + 480, + 665 + ], + "type": "text", + "content": "13. Geng, C., Huang, S.j., Chen, S.: Recent advances in open set recognition: A survey. IEEE transactions on pattern analysis and machine intelligence 43(10), 3614-3631 (2020)" + } + ] + } + ], + "index": 17 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 411, + 91, + 447, + 100 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 411, + 91, + 447, + 100 + ], + "spans": [ + { + "bbox": [ + 411, + 91, + 447, + 100 + ], + "type": "text", + "content": "PILORA" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 470, + 91, + 481, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 470, + 91, + 481, + 100 + ], + "spans": [ + { + "bbox": [ + 470, + 91, + 481, + 100 + ], + "type": "text", + "content": "15" + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 14 + }, + { + "para_blocks": [ + { + "bbox": [ + 132, + 117, + 481, + 665 + ], + "type": "list", + "angle": 0, + "index": 19, + "blocks": [ + { + "bbox": [ + 133, + 117, + 481, + 160 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 117, + 481, + 160 + ], + "spans": [ + { + "bbox": [ + 133, + 117, + 481, + 160 + ], + "type": "text", + "content": "14. Giuseppi, A., Manfredi, S., Pietrabissa, A.: A weighted average consensus approach for decentralized federated learning. Machine Intelligence Research 19(4), 319-330 (2022). https://doi.org/10.1007/s11633-022-1338-z, https://www.mi-research.net/en/article/doi/10.1007/s11633-022-1338-z" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 133, + 161, + 481, + 194 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 161, + 481, + 194 + ], + "spans": [ + { + "bbox": [ + 133, + 161, + 481, + 194 + ], + "type": "text", + "content": "15. Guo, T., Guo, S., Wang, J., Tang, X., Xu, W.: Promptfl: Let federated participants cooperatively learn prompts instead of models-federated learning in age of foundation model. IEEE Transactions on Mobile Computing (2023)" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 133, + 194, + 481, + 226 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 194, + 481, + 226 + ], + "spans": [ + { + "bbox": [ + 133, + 194, + 481, + 226 + ], + "type": "text", + "content": "16. Houlsby, N., Giurgiu, A., Jastrzebski, S., Morrone, B., De Laroussilhe, Q., Gesmundo, A., Attariyan, M., Gelly, S.: Parameter-efficient transfer learning for nlp. In: International Conference on Machine Learning. pp. 2790-2799. PMLR (2019)" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 132, + 227, + 481, + 258 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 227, + 481, + 258 + ], + "spans": [ + { + "bbox": [ + 132, + 227, + 481, + 258 + ], + "type": "text", + "content": "17. Hu, E.J., Shen, Y., Wallis, P., Allen-Zhu, Z., Li, Y., Wang, S., Wang, L., Chen, W.: Lora: Low-rank adaptation of large language models. arXiv preprint arXiv:2106.09685 (2021)" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 132, + 259, + 481, + 293 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 259, + 481, + 293 + ], + "spans": [ + { + "bbox": [ + 132, + 259, + 481, + 293 + ], + "type": "text", + "content": "18. Ilharco, G., Ribeiro, M.T., Wortsman, M., Gururangan, S., Schmidt, L., Hajishirzi, H., Farhadi, A.: Editing models with task arithmetic. arXiv preprint arXiv:2212.04089 (2022)" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 132, + 293, + 481, + 326 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 293, + 481, + 326 + ], + "spans": [ + { + "bbox": [ + 132, + 293, + 481, + 326 + ], + "type": "text", + "content": "19. Karimireddy, S.P., Kale, S., Mohri, M., Reddi, S., Stich, S., Suresh, A.T.: Scaffold: Stochastic controlled averaging for federated learning. In: International conference on machine learning. pp. 5132-5143. PMLR (2020)" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 132, + 326, + 481, + 358 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 326, + 481, + 358 + ], + "spans": [ + { + "bbox": [ + 132, + 326, + 481, + 358 + ], + "type": "text", + "content": "20. Khan, L.U., Saad, W., Han, Z., Hossain, E., Hong, C.S.: Federated learning for internet of things: Recent advances, taxonomy, and open challenges. IEEE Communications Surveys & Tutorials 23(3), 1759-1799 (2021)" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 132, + 358, + 481, + 379 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 358, + 481, + 379 + ], + "spans": [ + { + "bbox": [ + 132, + 358, + 481, + 379 + ], + "type": "text", + "content": "21. Kingma, D.P., Ba, J.: Adam: A method for stochastic optimization. arXiv preprint arXiv:1412.6980 (2014)" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 132, + 380, + 481, + 413 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 380, + 481, + 413 + ], + "spans": [ + { + "bbox": [ + 132, + 380, + 481, + 413 + ], + "type": "text", + "content": "22. Kirillov, A., Mintun, E., Ravi, N., Mao, H., Rolland, C., Gustafson, L., Xiao, T., Whitehead, S., Berg, A.C., Lo, W.Y., et al.: Segment anything. arXiv preprint arXiv:2304.02643 (2023)" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 132, + 414, + 481, + 456 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 414, + 481, + 456 + ], + "spans": [ + { + "bbox": [ + 132, + 414, + 481, + 456 + ], + "type": "text", + "content": "23. Kirkpatrick, J., Pascanu, R., Rabinowitz, N., Veness, J., Desjardins, G., Rusu, A.A., Milan, K., Quan, J., Ramalho, T., Grabska-Barwinska, A., et al.: Overcoming catastrophic forgetting in neural networks. Proceedings of the national academy of sciences 114(13), 3521-3526 (2017)" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 132, + 457, + 481, + 478 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 457, + 481, + 478 + ], + "spans": [ + { + "bbox": [ + 132, + 457, + 481, + 478 + ], + "type": "text", + "content": "24. Krizhevsky, A., Hinton, G., et al.: Learning multiple layers of features from tiny images (2009)" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 132, + 479, + 481, + 501 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 479, + 481, + 501 + ], + "spans": [ + { + "bbox": [ + 132, + 479, + 481, + 501 + ], + "type": "text", + "content": "25. Le, Y., Yang, X.: Tiny imagenet visual recognition challenge. CS 231N 7(7), 3 (2015)" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 132, + 502, + 481, + 533 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 502, + 481, + 533 + ], + "spans": [ + { + "bbox": [ + 132, + 502, + 481, + 533 + ], + "type": "text", + "content": "26. Legate, G., Bernier, N., Page-Caccia, L., Oyallon, E., Belilovsky, E.: Guiding the last layer in federated learning with pre-trained models. Advances in Neural Information Processing Systems 36 (2024)" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 132, + 534, + 481, + 566 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 534, + 481, + 566 + ], + "spans": [ + { + "bbox": [ + 132, + 534, + 481, + 566 + ], + "type": "text", + "content": "27. Li, Q., Diao, Y., Chen, Q., He, B.: Federated learning on non-iid data silos: An experimental study. In: 2022 IEEE 38th International Conference on Data Engineering (ICDE). pp. 965-978. IEEE (2022)" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 132, + 567, + 481, + 599 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 567, + 481, + 599 + ], + "spans": [ + { + "bbox": [ + 132, + 567, + 481, + 599 + ], + "type": "text", + "content": "28. Li, T., Sahu, A.K., Zaheer, M., Sanjabi, M., Talwalkar, A., Smith, V.: Federated optimization in heterogeneous networks. Proceedings of Machine learning and systems 2, 429-450 (2020)" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 132, + 600, + 481, + 621 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 600, + 481, + 621 + ], + "spans": [ + { + "bbox": [ + 132, + 600, + 481, + 621 + ], + "type": "text", + "content": "29. Li, Z., Hoiem, D.: Learning without forgetting. IEEE transactions on pattern analysis and machine intelligence 40(12), 2935-2947 (2017)" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 132, + 622, + 481, + 665 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 622, + 481, + 665 + ], + "spans": [ + { + "bbox": [ + 132, + 622, + 481, + 665 + ], + "type": "text", + "content": "30. Lin, Y.M., Gao, Y., Gong, M.G., Zhang, S.J., Zhang, Y.Q., Li, Z.Y.: Federated learning on multimodal data: A comprehensive survey. Machine Intelligence Research 20(4), 539-553 (2023). https://doi.org/10.1007/s11633-022-1398-0, https://www.mi-research.net/en/article/doi/10.1007/s11633-022-1398-0" + } + ] + } + ], + "index": 18 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "spans": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "type": "text", + "content": "16" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 166, + 91, + 220, + 100 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 166, + 91, + 220, + 100 + ], + "spans": [ + { + "bbox": [ + 166, + 91, + 220, + 100 + ], + "type": "text", + "content": "H. Guo et al." + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 15 + }, + { + "para_blocks": [ + { + "bbox": [ + 130, + 116, + 482, + 666 + ], + "type": "list", + "angle": 0, + "index": 19, + "blocks": [ + { + "bbox": [ + 130, + 116, + 482, + 149 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 116, + 482, + 149 + ], + "spans": [ + { + "bbox": [ + 130, + 116, + 482, + 149 + ], + "type": "text", + "content": "31. Liu, C.L., Sako, H., Fujisawa, H.: Discriminative learning quadratic discriminant function for handwriting recognition. IEEE Transactions on Neural Networks 15(2), 430-444 (2004)" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 130, + 150, + 481, + 183 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 150, + 481, + 183 + ], + "spans": [ + { + "bbox": [ + 130, + 150, + 481, + 183 + ], + "type": "text", + "content": "32. Liu, C.L., Sako, H., Fujisawa, H.: Effects of classifier structures and training regimes on integrated segmentation and recognition of handwritten numeral strings. IEEE transactions on pattern analysis and machine intelligence 26(11), 1395-1407 (2004)" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 130, + 183, + 481, + 215 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 183, + 481, + 215 + ], + "spans": [ + { + "bbox": [ + 130, + 183, + 481, + 215 + ], + "type": "text", + "content": "33. Liu, C., Qu, X., Wang, J., Xiao, J.: Fedet: A communication-efficient federated class-incremental learning framework based on enhanced transformer. arXiv preprint arXiv:2306.15347 (2023)" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 130, + 216, + 481, + 248 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 216, + 481, + 248 + ], + "spans": [ + { + "bbox": [ + 130, + 216, + 481, + 248 + ], + "type": "text", + "content": "34. Liu, P., Yuan, W., Fu, J., Jiang, Z., Hayashi, H., Neubig, G.: Pre-train, prompt, and predict: A systematic survey of prompting methods in natural language processing. ACM Computing Surveys 55(9), 1-35 (2023)" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 130, + 249, + 481, + 281 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 249, + 481, + 281 + ], + "spans": [ + { + "bbox": [ + 130, + 249, + 481, + 281 + ], + "type": "text", + "content": "35. Liu, W., Wu, X., Zhu, F., Yu, M., Wang, C., Liu, C.L.: Class incremental learning with self-supervised pre-training and prototype learning. arXiv preprint arXiv:2308.02346 (2023)" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 130, + 281, + 481, + 304 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 281, + 481, + 304 + ], + "spans": [ + { + "bbox": [ + 130, + 281, + 481, + 304 + ], + "type": "text", + "content": "36. Liu, W., Zhu, F., Liu, C.L.: Branch-tuning: Balancing stability and plasticity for continual self-supervised learning. arXiv preprint arXiv:2403.18266 (2024)" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 130, + 304, + 481, + 326 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 304, + 481, + 326 + ], + "spans": [ + { + "bbox": [ + 130, + 304, + 481, + 326 + ], + "type": "text", + "content": "37. Liu, W., Zhu, F., Liu, C.L.: Towards non-exemplar semi-supervised class-incremental learning. arXiv preprint arXiv:2403.18291 (2024)" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 130, + 326, + 481, + 358 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 326, + 481, + 358 + ], + "spans": [ + { + "bbox": [ + 130, + 326, + 481, + 358 + ], + "type": "text", + "content": "38. Luo, M., Chen, F., Hu, D., Zhang, Y., Liang, J., Feng, J.: No fear of heterogeneity: Classifier calibration for federated learning with non-iid data. Advances in Neural Information Processing Systems 34, 5972-5984 (2021)" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 130, + 358, + 481, + 391 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 358, + 481, + 391 + ], + "spans": [ + { + "bbox": [ + 130, + 358, + 481, + 391 + ], + "type": "text", + "content": "39. McCloskey, M., Cohen, N.J.: Catastrophic interference in connectionist networks: The sequential learning problem. In: Psychology of learning and motivation, vol. 24, pp. 109-165. Elsevier (1989)" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 130, + 392, + 481, + 424 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 392, + 481, + 424 + ], + "spans": [ + { + "bbox": [ + 130, + 392, + 481, + 424 + ], + "type": "text", + "content": "40. McMahan, B., Moore, E., Ramage, D., Hampson, S., y Arcas, B.A.: Communication-efficient learning of deep networks from decentralized data. In: Artificial intelligence and statistics. pp. 1273-1282. PMLR (2017)" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 130, + 424, + 481, + 456 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 424, + 481, + 456 + ], + "spans": [ + { + "bbox": [ + 130, + 424, + 481, + 456 + ], + "type": "text", + "content": "41. Niknam, S., Dhillon, H.S., Reed, J.H.: Federated learning for wireless communications: Motivation, opportunities, and challenges. IEEE Communications Magazine 58(6), 46-51 (2020)" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 130, + 457, + 481, + 490 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 457, + 481, + 490 + ], + "spans": [ + { + "bbox": [ + 130, + 457, + 481, + 490 + ], + "type": "text", + "content": "42. Ortiz-Jimenez, G., Favero, A., Frossard, P.: Task arithmetic in the tangent space: Improved editing of pre-trained models. Advances in Neural Information Processing Systems 36 (2024)" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 130, + 491, + 481, + 511 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 491, + 481, + 511 + ], + "spans": [ + { + "bbox": [ + 130, + 491, + 481, + 511 + ], + "type": "text", + "content": "43. Qi, D., Zhao, H., Li, S.: Better generative replay for continual federated learning. In: The Eleventh International Conference on Learning Representations (2022)" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 130, + 512, + 481, + 555 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 512, + 481, + 555 + ], + "spans": [ + { + "bbox": [ + 130, + 512, + 481, + 555 + ], + "type": "text", + "content": "44. Qu, L., Zhou, Y., Liang, P.P., Xia, Y., Wang, F., Adeli, E., Fei-Fei, L., Rubin, D.: Rethinking architecture design for tackling data heterogeneity in federated learning. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 10061-10071 (2022)" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 130, + 555, + 481, + 599 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 555, + 481, + 599 + ], + "spans": [ + { + "bbox": [ + 130, + 555, + 481, + 599 + ], + "type": "text", + "content": "45. Radford, A., Kim, J.W., Hallacy, C., Ramesh, A., Goh, G., Agarwal, S., Sastry, G., Askell, A., Mishkin, P., Clark, J., et al.: Learning transferable visual models from natural language supervision. In: International conference on machine learning. pp. 8748-8763. PMLR (2021)" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 130, + 600, + 481, + 632 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 600, + 481, + 632 + ], + "spans": [ + { + "bbox": [ + 130, + 600, + 481, + 632 + ], + "type": "text", + "content": "46. Rebuffi, S.A., Kolesnikov, A., Sperl, G., Lampert, C.H.: icarl: Incremental classifier and representation learning. In: Proceedings of the IEEE conference on Computer Vision and Pattern Recognition. pp. 2001-2010 (2017)" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 130, + 632, + 481, + 666 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 632, + 481, + 666 + ], + "spans": [ + { + "bbox": [ + 130, + 632, + 481, + 666 + ], + "type": "text", + "content": "47. Rieke, N., Hancox, J., Li, W., Miletari, F., Roth, H.R., Albarqouni, S., Bakas, S., Galtier, M.N., Landman, B.A., Maier-Hein, K., et al.: The future of digital health with federated learning. NPJ digital medicine 3(1), 119 (2020)" + } + ] + } + ], + "index": 18 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 412, + 91, + 447, + 100 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 412, + 91, + 447, + 100 + ], + "spans": [ + { + "bbox": [ + 412, + 91, + 447, + 100 + ], + "type": "text", + "content": "PILoRA" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 470, + 91, + 481, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 470, + 91, + 481, + 100 + ], + "spans": [ + { + "bbox": [ + 470, + 91, + 481, + 100 + ], + "type": "text", + "content": "17" + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 16 + }, + { + "para_blocks": [ + { + "bbox": [ + 132, + 117, + 480, + 665 + ], + "type": "list", + "angle": 0, + "index": 18, + "blocks": [ + { + "bbox": [ + 133, + 117, + 480, + 149 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 117, + 480, + 149 + ], + "spans": [ + { + "bbox": [ + 133, + 117, + 480, + 149 + ], + "type": "text", + "content": "48. Shang, X., Lu, Y., Huang, G., Wang, H.: Federated learning on heterogeneous and long-tailed data via classifier re-training with federated features. arXiv preprint arXiv:2204.13399 (2022)" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 133, + 150, + 480, + 183 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 150, + 480, + 183 + ], + "spans": [ + { + "bbox": [ + 133, + 150, + 480, + 183 + ], + "type": "text", + "content": "49. Shenaj, D., Toldo, M., Rigon, A., Zanuttigh, P.: Asynchronous federated continual learning. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 5054-5062 (2023)" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 132, + 183, + 480, + 237 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 183, + 480, + 237 + ], + "spans": [ + { + "bbox": [ + 132, + 183, + 480, + 237 + ], + "type": "text", + "content": "50. Smith, J.S., Karlinsky, L., Gutta, V., Cascante-Bonilla, P., Kim, D., Arbelle, A., Panda, R., Feris, R., Kira, Z.: Coda-prompt: Continual decomposed attention-based prompting for rehearsal-free continual learning. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 11909-11919 (2023)" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 132, + 237, + 480, + 270 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 237, + 480, + 270 + ], + "spans": [ + { + "bbox": [ + 132, + 237, + 480, + 270 + ], + "type": "text", + "content": "51. Tan, Y., Long, G., Liu, L., Zhou, T., Lu, Q., Jiang, J., Zhang, C.: Fedproto: Federated prototype learning across heterogeneous clients. In: Proceedings of the AAAI Conference on Artificial Intelligence. vol. 36, pp. 8432-8440 (2022)" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 132, + 270, + 480, + 303 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 270, + 480, + 303 + ], + "spans": [ + { + "bbox": [ + 132, + 270, + 480, + 303 + ], + "type": "text", + "content": "52. Voigt, P., Von dem Bussche, A.: The eu general data protection regulation (gdpr). A Practical Guide, 1st Ed., Cham: Springer International Publishing 10(3152676), 10-5555 (2017)" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 132, + 303, + 480, + 336 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 303, + 480, + 336 + ], + "spans": [ + { + "bbox": [ + 132, + 303, + 480, + 336 + ], + "type": "text", + "content": "53. Wang, X., Chen, T., Ge, Q., Xia, H., Bao, R., Zheng, R., Zhang, Q., Gui, T., Huang, X.: Orthogonal subspace learning for language model continual learning. arXiv preprint arXiv:2310.14152 (2023)" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 132, + 336, + 480, + 380 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 336, + 480, + 380 + ], + "spans": [ + { + "bbox": [ + 132, + 336, + 480, + 380 + ], + "type": "text", + "content": "54. Wang, Z., Zhang, Z., Ebrahimi, S., Sun, R., Zhang, H., Lee, C.Y., Ren, X., Su, G., Perot, V., Dy, J., et al.: Dualprompt: Complementary prompting for rehearsal-free continual learning. In: European Conference on Computer Vision. pp. 631-648. Springer (2022)" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 132, + 380, + 480, + 423 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 380, + 480, + 423 + ], + "spans": [ + { + "bbox": [ + 132, + 380, + 480, + 423 + ], + "type": "text", + "content": "55. Wang, Z., Zhang, Z., Lee, C.Y., Zhang, H., Sun, R., Ren, X., Su, G., Perot, V., Dy, J., Pfister, T.: Learning to prompt for continual learning. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 139-149 (2022)" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 132, + 423, + 480, + 456 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 423, + 480, + 456 + ], + "spans": [ + { + "bbox": [ + 132, + 423, + 480, + 456 + ], + "type": "text", + "content": "56. Yang, H.M., Zhang, X.Y., Yin, F., Liu, C.L.: Robust classification with convolutional prototype learning. In: Proceedings of the IEEE conference on computer vision and pattern recognition. pp. 3474-3482 (2018)" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 132, + 456, + 480, + 490 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 456, + 480, + 490 + ], + "spans": [ + { + "bbox": [ + 132, + 456, + 480, + 490 + ], + "type": "text", + "content": "57. Yang, H.M., Zhang, X.Y., Yin, F., Yang, Q., Liu, C.L.: Convolutional prototype network for open set recognition. IEEE Transactions on Pattern Analysis and Machine Intelligence 44(5), 2358-2370 (2020)" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 132, + 490, + 480, + 512 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 490, + 480, + 512 + ], + "spans": [ + { + "bbox": [ + 132, + 490, + 480, + 512 + ], + "type": "text", + "content": "58. Yi, L., Yu, H., Wang, G., Liu, X.: Fedlora: Model-heterogeneous personalized federated learning with lora tuning. arXiv preprint arXiv:2310.13283 (2023)" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 132, + 512, + 480, + 544 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 512, + 480, + 544 + ], + "spans": [ + { + "bbox": [ + 132, + 512, + 480, + 544 + ], + "type": "text", + "content": "59. Yoon, J., Jeong, W., Lee, G., Yang, E., Hwang, S.J.: Federated continual learning with weighted inter-client transfer. In: International Conference on Machine Learning. pp. 12073-12086. PMLR (2021)" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 132, + 544, + 480, + 578 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 544, + 480, + 578 + ], + "spans": [ + { + "bbox": [ + 132, + 544, + 480, + 578 + ], + "type": "text", + "content": "60. Zhang, G., Wang, L., Kang, G., Chen, L., Wei, Y.: Slow learner with classifier alignment for continual learning on a pre-trained model. arXiv preprint arXiv:2303.05118 (2023)" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 132, + 578, + 480, + 600 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 578, + 480, + 600 + ], + "spans": [ + { + "bbox": [ + 132, + 578, + 480, + 600 + ], + "type": "text", + "content": "61. Zhang, J., Chen, C., Zhuang, W., Lv, L.: Addressing catastrophic forgetting in federated class-continual learning. arXiv preprint arXiv:2303.06937 (2023)" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 132, + 600, + 480, + 632 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 600, + 480, + 632 + ], + "spans": [ + { + "bbox": [ + 132, + 600, + 480, + 632 + ], + "type": "text", + "content": "62. Zhang, J., Chen, C., Zhuang, W., Lyu, L.: Target: Federated class-continual learning via exemplar-free distillation. In: Proceedings of the IEEE/CVF International Conference on Computer Vision. pp. 4782-4793 (2023)" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 132, + 632, + 480, + 665 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 632, + 480, + 665 + ], + "spans": [ + { + "bbox": [ + 132, + 632, + 480, + 665 + ], + "type": "text", + "content": "63. Zhao, H., Du, W., Li, F., Li, P., Liu, G.: Reduce communication costs and preserve privacy: Prompt tuning method in federated learning. arXiv preprint arXiv:2208.12268 (2022)" + } + ] + } + ], + "index": 17 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "spans": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "type": "text", + "content": "18" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 166, + 91, + 220, + 100 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 166, + 91, + 220, + 100 + ], + "spans": [ + { + "bbox": [ + 166, + 91, + 220, + 100 + ], + "type": "text", + "content": "H. Guo et al." + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 17 + }, + { + "para_blocks": [ + { + "bbox": [ + 130, + 116, + 482, + 205 + ], + "type": "list", + "angle": 0, + "index": 5, + "blocks": [ + { + "bbox": [ + 130, + 116, + 482, + 149 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 116, + 482, + 149 + ], + "spans": [ + { + "bbox": [ + 130, + 116, + 482, + 149 + ], + "type": "text", + "content": "64. Zhu, F., Cheng, Z., Zhang, X.y., Liu, C.l.: Class-incremental learning via dual augmentation. Advances in Neural Information Processing Systems 34, 14306-14318 (2021)" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 130, + 150, + 481, + 172 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 150, + 481, + 172 + ], + "spans": [ + { + "bbox": [ + 130, + 150, + 481, + 172 + ], + "type": "text", + "content": "65. Zhu, F., Ma, S., Cheng, Z., Zhang, X.Y., Zhang, Z., Liu, C.L.: Open-world machine learning: A review and new outlooks. arXiv preprint arXiv:2403.01759 (2024)" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 132, + 172, + 481, + 205 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 172, + 481, + 205 + ], + "spans": [ + { + "bbox": [ + 132, + 172, + 481, + 205 + ], + "type": "text", + "content": "66. Zhu, F., Zhang, X.Y., Wang, C., Yin, F., Liu, C.L.: Prototype augmentation and self-supervision for incremental learning. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 5871-5880 (2021)" + } + ] + } + ], + "index": 4 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 411, + 91, + 447, + 100 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 411, + 91, + 447, + 100 + ], + "spans": [ + { + "bbox": [ + 411, + 91, + 447, + 100 + ], + "type": "text", + "content": "PILORA" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 470, + 91, + 481, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 470, + 91, + 481, + 100 + ], + "spans": [ + { + "bbox": [ + 470, + 91, + 481, + 100 + ], + "type": "text", + "content": "19" + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 18 + } + ], + "_backend": "vlm", + "_version_name": "2.6.4" +} \ No newline at end of file diff --git a/2024/PISR_ Polarimetric Neural Implicit Surface Reconstruction for Textureless and Specular Objects/949f096c-d4bf-4988-bb1a-9febc3a46def_content_list.json b/2024/PISR_ Polarimetric Neural Implicit Surface Reconstruction for Textureless and Specular Objects/949f096c-d4bf-4988-bb1a-9febc3a46def_content_list.json new file mode 100644 index 0000000000000000000000000000000000000000..03132c36d8dc3d0edc73ba85f31912133f3e8f52 --- /dev/null +++ b/2024/PISR_ Polarimetric Neural Implicit Surface Reconstruction for Textureless and Specular Objects/949f096c-d4bf-4988-bb1a-9febc3a46def_content_list.json @@ -0,0 +1,2174 @@ +[ + { + "type": "text", + "text": "PISR: Polarimetric Neural Implicit Surface Reconstruction for Textureless and Specular Objects", + "text_level": 1, + "bbox": [ + 277, + 140, + 725, + 209 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Guangcheng Chen $^{1,2}$ , Yicheng He $^{1,2}$ , Li He $^{1,2}$ , and Hong Zhang $^{\\star 1,2}$", + "bbox": [ + 236, + 233, + 764, + 251 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "$^{1}$ Southern University of Science and Technology, Shenzhen, China $^{2}$ Shenzhen Key Laboratory of Robotics and Computer Vision {chengc2023,12332136}@mail.sustech.edu.cn, {hel, hzhang}@sustech.edu.cn", + "bbox": [ + 228, + 261, + 772, + 305 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Abstract. Neural implicit surface reconstruction has achieved remarkable progress recently. Despite resorting to complex radiance modeling, state-of-the-art methods still struggle with textureless and specular surfaces. Different from RGB images, polarization images can provide direct constraints on the azimuth angles of the surface normals. In this paper, we present PISR, a novel method that utilizes a geometrically accurate polarimetric loss to refine shape independently of appearance. In addition, PISR smooths surface normals in image space to eliminate severe shape distortions and leverages the hash-grid-based neural signed distance function to accelerate the reconstruction. Experimental results demonstrate that PISR achieves higher accuracy and robustness, with an L1 Chamfer distance of $0.5\\mathrm{mm}$ and an F-score of $99.5\\%$ at $1\\mathrm{mm}$ , while converging $4 \\sim 30\\times$ faster than previous polarimetric surface reconstruction methods.", + "bbox": [ + 259, + 340, + 743, + 537 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Keywords: Polarization $\\cdot$ Surface reconstruction $\\cdot$ Multi-view reconstruction $\\cdot$ Neural implicit surface $\\cdot$ Polarimetric constraint", + "bbox": [ + 259, + 549, + 740, + 578 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "1 Introduction", + "text_level": 1, + "bbox": [ + 215, + 604, + 375, + 619 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Surface reconstruction from multi-view images is a fundamental problem in computer vision, with various applications in graphics, robotics, and more. Recently, neural implicit surface has shown powerful flexibility for representing 3D shape and appearance. Combined with differentiable volume rendering [23], neural implicit surface reconstruction is able to estimate 3D shapes from multi-view images accurately [42, 47]. However, these methods struggle with textureless and specular objects due to the shape-radiance ambiguity [52]. Recent methods resort to diffuse and specular separation modeling [40], specular reflection detection [9] and inverse rendering techniques [18, 51]. While these methods show promising results, they may still encounter difficulties since their optimizations rely on image reconstruction losses that tightly couple shape and appearance.", + "bbox": [ + 212, + 636, + 787, + 803 + ], + "page_idx": 0 + }, + { + "type": "page_footnote", + "text": "* Corresponding author. The source code is available at https://github.com/GCChen97/PISR", + "bbox": [ + 217, + 810, + 691, + 839 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/6a469094d20c368d430fe3b7fa6a5133e6f628dce87676295cae92f606d652dc.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 228, + 142, + 385, + 215 + ], + "page_idx": 1 + }, + { + "type": "image", + "img_path": "images/a9a1be8c9f3aacf37a7af9a94811b3fe6741c52a3798b0a0f0ec1683286ad4ee.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 388, + 143, + 571, + 215 + ], + "page_idx": 1 + }, + { + "type": "image", + "img_path": "images/b43769659760d7d04a02cb3d57e87d681c848a8afa1b14128ddd2de94d077a7f.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 576, + 143, + 756, + 215 + ], + "page_idx": 1 + }, + { + "type": "image", + "img_path": "images/03958391f55e6d912b84c5d3ff6c3c770da3390fad1a49b0bdd521592166caca.jpg", + "image_caption": [ + "Fig. 1: Normal maps of reconstructed surfaces. By leveraging polarization information, our method is able to reconstruct textureless and specular surfaces with fine-grained details." + ], + "image_footnote": [], + "bbox": [ + 228, + 217, + 385, + 305 + ], + "page_idx": 1 + }, + { + "type": "image", + "img_path": "images/79fe8c59d1ab29dcf49fabd5da5546765bf44b592933c83e5e5395b5265f2e31.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 388, + 224, + 571, + 306 + ], + "page_idx": 1 + }, + { + "type": "image", + "img_path": "images/bbe473ac767a597842053e424acc7a6b6c52d8b5791f226911cd83b0038db1ee.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 576, + 224, + 756, + 305 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Polarization, as another property of light, provides direct cues on object shape. The angle of polarization of light from object surfaces is independent of light intensity and mainly determined by the surface shapes. From the perspective of a camera, the angle of polarization of a pixel is equal to the azimuth of the surface normal, i.e., the 2D projection direction of the normal onto the image plane, up to a $\\pi/2$ -ambiguity and a $\\pi$ -ambiguity [5]. This geometric property has been extensively explored in shape from polarization [1, 2, 25, 49], photo-polarimetric stereo [7, 8, 27, 38], polarimetric multi-view 3D reconstruction [5, 54, 55] and dense mapping [34, 46].", + "bbox": [ + 212, + 402, + 784, + 541 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "In this work, we introduce PISR (polarimetric neural implicit surface reconstruction), a novel method that combines the merits of polarimetric 3D reconstruction and neural implicit surface reconstruction for textureless and specular objects, such as ceramics and plastics. Different from recent works [6, 16] that utilize the polarization cues through polarimetric volume rendering, we propose a novel polarimetric loss to refine an object shape directly. Thanks to the integration of the perspective polarimetric constraint [4] in the proposed loss, PISR eliminates artifacts on the reconstructed surfaces, resulting in a $30\\%$ reduction in L1 Chamfer distance. Besides, instead of treating all possible normals equally [54], our loss adaptively adjusts according to the degree of polarization to address specular-reflection-dominant surfaces, such as ceramics with dark color.", + "bbox": [ + 212, + 546, + 787, + 712 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "To accelerate the optimization process, PISR uses a multi-resolution hash grid and multi-layer perceptrons (MLP) to represent an object, achieving $4 \\sim 30 \\times$ speed up as shown in Tab. 3. This representation is also crucial for accurately reconstructing highly reflective surfaces, enabling PISR to rectify the shape topology during optimization, a processing step not available for the mesh-based method, PMVIR [54]. However, the discrete nature of a hash grid also reduces the smoothness of the representation and can lead to the reconstruction", + "bbox": [ + 212, + 734, + 787, + 839 + ], + "page_idx": 1 + }, + { + "type": "page_number", + "text": "2", + "bbox": [ + 217, + 114, + 228, + 126 + ], + "page_idx": 1 + }, + { + "type": "header", + "text": "G. Chen et al.", + "bbox": [ + 271, + 114, + 367, + 126 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "results being trapped in local minima. To mitigate the problem, PISR smooths rendered normals for better robustness against poor shape initialization.", + "bbox": [ + 212, + 146, + 782, + 176 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Since there is no publicly available dataset for quantitative evaluation of polarimetric 3D reconstruction, we collect a dataset containing textureless and specular objects with a color polarization camera [36]. Using the dataset, we show that PISR is able to improve the surface reconstruction accuracy over the SOTA methods, with an L1 Chamfer distance of $0.5\\mathrm{mm}$ and an F-score of $99.5\\%$ at $1\\mathrm{mm}$ . Shown in Fig. 1 are recovered shapes by different methods and our result shows an accurate shape and fine-grained details.", + "bbox": [ + 212, + 178, + 787, + 282 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "In summary, our contributions are as follows:", + "bbox": [ + 238, + 285, + 566, + 299 + ], + "page_idx": 2 + }, + { + "type": "list", + "sub_type": "text", + "list_items": [ + "- PISR, a novel method that leverages polarization cues and hash-grid-based neural SDF for robust and efficient reconstruction of textureless and specular objects.", + "- A novel polarimetric loss that utilizes an accurate formulation of the polarimetric constraint for higher surface reconstruction accuracy.", + "- A polarization image dataset of real textureless and specular objects with ground-truth meshes for quantitative evaluation." + ], + "bbox": [ + 223, + 314, + 782, + 421 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "2 Related Works", + "text_level": 1, + "bbox": [ + 215, + 452, + 395, + 468 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Neural Implicit Surface Reconstruction: The success of neural radiance field reconstruction [23] in novel view synthesis greatly inspires the development of neural implicit surface [28, 30, 48]. Combining with differentiable volume rendering, neural implicit surface reconstruction [29, 42, 47] is able to reconstruct implicit surfaces from dense multi-view images while showing excellent performance in reconstructing fine-grained details. Since then, several works have improved it in reconstruction quality [50], sparse-view performance [19] and optimization speed [43].", + "bbox": [ + 212, + 489, + 787, + 611 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Yet textureless and specular surfaces are still challenging for neural implicit surface reconstruction. To address the problem, NeRO [18] incorporates inverse rendering to alleviate the shape-radiance ambiguity [52] induced by specular reflection. Ref-NeuS [9] borrows the idea of anomaly detection to decrease the weights of specular pixels during optimization and conditions the radiance on reflection direction in reflective scenes as in Ref-NeRF [39]. UniSDF [40] models diffuse and specular radiances in two MLPs separately.", + "bbox": [ + 212, + 612, + 787, + 717 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Estimating a 3D radiance field and an SDF from 2D images is a highly ill-posed problem. During optimization, these methods may distort shapes to fit the reflection colors, causing severe reconstruction errors, or, as is commonly referred to, the shape-radiance ambiguity [52]. In contrast, polarization images provide linear constraints on surface normal by the angle of polarization. Therefore, to avoid the shape-radiance ambiguity, our method utilizes the geometry cues from polarization to regularize object shapes independently of appearance, keeping the shape from distortions induced by the visually distorted appearance.", + "bbox": [ + 212, + 719, + 787, + 839 + ], + "page_idx": 2 + }, + { + "type": "header", + "text": "PISR for Textureless and Specular Objects", + "bbox": [ + 444, + 114, + 732, + 128 + ], + "page_idx": 2 + }, + { + "type": "page_number", + "text": "3", + "bbox": [ + 774, + 116, + 784, + 126 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Polarimetric 3D Reconstruction: Shape from polarization (SfP) [1,25,31,49] and photo-polarimetric stereo [7,8,27,38] are single-view model-based methods that utilize polarization images for normal and depth estimation. These methods rely on assumptions of either spotlighting or distant lighting to explicitly model the relationship between the degree of polarization and the zenith angle of a surface normal, making them difficult to work under natural lighting conditions. Several deep learning-based methods [2,21,35] leverage data priors to enhance the robustness to more complex indoor lighting conditions. SfPW [15] and DPS-Net [37] learn to estimate scene-level normal maps and depth maps. Yet their generalization is still limited by the relatively small amount of training datasets of polarization images.", + "bbox": [ + 212, + 145, + 787, + 311 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Polarimetric multi-view 3D reconstruction leverages multi-view polarimetric constraints to adapt to natural lighting conditions. The polarimetric constraint is established on the fundamental fact that the azimuth angle of a surface normal is equal to the angle of polarization up to a $\\pi/2$ -ambiguity and a $\\pi$ -ambiguity under natural lighting [5, 54], which is relatively independent of lighting [5]. Miyazaki et al. [24] uses multi-view polarimetric constraints to reconstruct black specular objects. PolMVS [5] proposes to propagate the depths of sparse points and refine the depth maps with this constraint. Inspired by PolMVS, PDMS [46] and PMDM [34] integrate this constraint into SLAM for dense mapping. PMVIR [54] proposes a framework to simultaneously refine coarse meshes and estimate illuminations and surface albedos. PolarPMS [55] combines this constraint with photo consistency to estimate depth maps from multi-view polarization images.", + "bbox": [ + 212, + 313, + 787, + 494 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "More closely related to our approach is PMVIR [54]. PMVIR represents shapes using meshes and thereby has limitations in handling incorrect shape topology. In contrast, by leveraging neural SDF, our method is able to rectify incorrect shape topology during the optimization. More importantly, our method utilizes the perspective polarimetric constraint [4] to consider the camera ray directions in the polarimetric loss and thus achieves further performance improvement.", + "bbox": [ + 212, + 494, + 787, + 602 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Polarimetric Inverse Rendering: The development of polarimetric bidirectional reflectance distribution function (pBRDF) [3, 11, 12, 14] has enabled polarimetric inverse rendering. Based on mesh representation, sparse ellipsometry [11] leverages polarimetric inverse rendering to simultaneously estimate object meshes and materials using polarization images captured under flashlight illumination.", + "bbox": [ + 212, + 627, + 787, + 717 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "With the emergence of neural fields in 3D reconstruction, a few recent works [6, 16] have integrated polarization into neural implicit surfaces through polarization image rendering. Concurrent work NeISF [16] introduces the incident Stokes field to represent multi-path polarized light and reparameterizes pBRDF using MLPs, achieving highly accurate geometry reconstruction. PANDORA [6] aims to decompose diffuse and specular reflections using polarization as auxiliary cues, relying on the Fresnel equations to calculate the diffuse and specular components of radiance.", + "bbox": [ + 212, + 719, + 787, + 839 + ], + "page_idx": 3 + }, + { + "type": "page_number", + "text": "4", + "bbox": [ + 217, + 114, + 228, + 126 + ], + "page_idx": 3 + }, + { + "type": "header", + "text": "G. Chen et al.", + "bbox": [ + 271, + 114, + 367, + 127 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Although polarimetric inverse rendering excels at accurate material and illumination estimation, it requires proper geometry initialization to decouple shape and appearance. As a result, it may still suffer from the shape-radiance ambiguity, especially when dealing with textureless and specular surfaces. To avoid this problem, our method utilizes the polarimetric constraint to regularize shape directly, independent of appearance. The surface reconstruction results of our method can be used to initialize polarimetric inverse rendering.", + "bbox": [ + 212, + 146, + 782, + 252 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "3 Preliminary", + "text_level": 1, + "bbox": [ + 215, + 280, + 366, + 297 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Polarimetric Constraint: Assuming four synchronized and aligned images captured at polarizer angles of 0, 45, 90, 135 degrees [36], denoted as $\\mathbf{I}_0$ , $\\mathbf{I}_{45}$ , $\\mathbf{I}_{90}$ and $\\mathbf{I}_{135}$ . For each pixel, the angle of polarization (AoP) $\\varphi \\in [-\\pi /2,\\pi /2]$ and degree of polarization (DoP) $\\rho \\in [0,1]$ are calculated as:", + "bbox": [ + 212, + 316, + 782, + 377 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\n\\varphi = \\frac {1}{2} \\arctan 2 \\left(s _ {2}, s _ {1}\\right), \\quad \\rho = \\frac {\\sqrt {s _ {1} ^ {2} + s _ {2} ^ {2}}}{s _ {0}} \\tag {1}\n$$\n", + "text_format": "latex", + "bbox": [ + 362, + 391, + 784, + 424 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "where $\\mathbf{s} = [s_0, s_1, s_2]^T = \\left[\\frac{1}{2}\\sum I_i, I_0 - I_{90}, I_{45} - I_{135}\\right]^T$ is the Stokes vector for representing the polarization state of light. Given the estimated normal $\\hat{\\mathbf{n}}$ in the camera frame and the AoP $\\varphi$ of a pixel, the orthographic polarimetric constraint on the azimuth angle of the normal is calculated as:", + "bbox": [ + 212, + 436, + 782, + 497 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\n[ \\sin (\\varphi + \\Delta), \\cos (\\varphi + \\Delta), 0 ] \\cdot \\hat {\\mathbf {n}} = 0, \\tag {2}\n$$\n", + "text_format": "latex", + "bbox": [ + 375, + 513, + 784, + 527 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "where the disambiguation scalar $\\varDelta\\in\\{0,\\pi/2\\}$ is used for solving the $\\pi /2$ ambiguity [5]. While this constraint is commonly used in previous works [5,34, 38,46,54], there is a perspective polarimetric constraint which is geometrically more accurate than it by considering the perspective effect of the lens [4]:", + "bbox": [ + 212, + 542, + 782, + 603 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\n\\left[ v _ {z} \\sin \\varphi^ {\\prime}, v _ {z} \\cos \\varphi^ {\\prime}, - \\left(v _ {y} \\cos \\varphi^ {\\prime} + v _ {x} \\sin \\varphi^ {\\prime}\\right) \\right] \\cdot \\hat {\\mathbf {n}} = 0, \\quad \\varphi^ {\\prime} = \\varphi + \\Delta \\tag {3}\n$$\n", + "text_format": "latex", + "bbox": [ + 269, + 617, + 784, + 633 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "where $\\mathbf{v} = [v_x, v_y, v_z]^T$ is a normalized camera ray in the camera frame.", + "bbox": [ + 212, + 645, + 728, + 662 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "$\\pi /2$ ambiguity: The $\\pi /2$ -ambiguity arises from the differences between the AoPs of specular reflection and diffuse reflection [5]. According to the Fresnel equations, reflected and transmitted light have distinct energy distributions in the parallel and perpendicular directions to the plane of incident, resulting in a $\\pi /2$ difference between their AoPs, i.e., $\\Delta = 0$ for diffuse reflection and $\\Delta = \\pi /2$ for specular reflection. Under natural lighting conditions, the diffuse reflections and specular reflections are usually mixed, making it difficult to distinguish the dominant component, and thus resulting in the $\\pi /2$ -ambiguity. Therefore, the $\\pi /2$ -ambiguity poses a challenge in terms of applying the polarimetric constraints.", + "bbox": [ + 212, + 688, + 784, + 838 + ], + "page_idx": 4 + }, + { + "type": "header", + "text": "PISR for Textureless and Specular Objects", + "bbox": [ + 444, + 114, + 732, + 128 + ], + "page_idx": 4 + }, + { + "type": "page_number", + "text": "5", + "bbox": [ + 774, + 116, + 784, + 126 + ], + "page_idx": 4 + }, + { + "type": "image", + "img_path": "images/5a2e64f767d0edd25388219e1196d3019d957322d8aa6b5e2a8c01a6422f6e29.jpg", + "image_caption": [ + "Fig. 2: Overview of PISR. Pixels for the optimization are sampled in criss-cross patterns at the middle stage of the optimization. The polarimetric loss $\\mathcal{L}_{\\mathrm{pol}}^{\\mathrm{p}}$ and the normal loss $\\mathcal{L}_{\\mathrm{normal}}$ are used for regularizing the shape independently of appearance." + ], + "image_footnote": [], + "bbox": [ + 227, + 142, + 777, + 300 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "4 Method", + "text_level": 1, + "bbox": [ + 215, + 378, + 330, + 393 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "PISR aims to recover object shapes from multi-view polarization images and camera poses. Fig. 2 shows an overview of our method. Object shape is represented as a hash-grid-based neural SDF and rendered to images for computing losses (Sec. 4.1). During the reconstruction, the coarse shape of an object is estimated with the photometric loss $\\mathcal{L}_{\\mathrm{color}}$ . Then the polarimetric loss $\\mathcal{L}_{\\mathrm{pol}}^{\\mathrm{p}}$ (Sec. 4.2) and the normal loss $\\mathcal{L}_{\\mathrm{normal}}$ with criss-cross pattern sampling (Sec. 4.3) are jointly used to rectify the shape from distortions. Finally, the shape is refined with $\\mathcal{L}_{\\mathrm{color}}$ and $\\mathcal{L}_{\\mathrm{pol}}^{\\mathrm{p}}$ . The optimization scheme is introduced in Sec. 4.4.", + "bbox": [ + 212, + 407, + 787, + 532 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "4.1 Surface Representation and Image Rendering", + "text_level": 1, + "bbox": [ + 212, + 546, + 637, + 561 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Since Instant-NGP [26], combining a multi-resolution hash grid and MLPs has been commonly used for faster convergence in neural field reconstruction. Following recent works [17, 41, 43], we represent the shape and appearance of an object as an $L$ -level multi-resolution hash grid $\\mathcal{G} = \\{\\mathcal{G}_l\\}_{l=1}^L$ with two MLPs $\\varPhi_s$ and $\\varPhi_{\\mathbf{c}}$ for decoding SDF values $s$ and color $\\mathbf{c}$ respectively. For each 3D space point $\\mathbf{x}$ , the SDF value $s$ and color $\\mathbf{c}$ are computed as", + "bbox": [ + 212, + 568, + 787, + 659 + ], + "page_idx": 5 + }, + { + "type": "equation", + "text": "\n$$\n(s, \\mathbf {f}) = \\varPhi_ {s} (\\mathcal {G} (\\mathbf {x})), \\quad \\mathbf {c} = \\varPhi_ {\\mathbf {c}} (s, \\mathbf {f}, \\mathbf {v}, \\mathbf {n}) \\tag {4}\n$$\n", + "text_format": "latex", + "bbox": [ + 366, + 666, + 784, + 683 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "with the camera ray $\\mathbf{v}$ and surface normal $\\mathbf{n}$ . The normal $\\mathbf{n}$ can be computed by normalizing the gradient of SDF w.r.t. to $\\mathbf{x}$ or by finite difference [17]. Besides the foreground object, we use Instant-NGP [26] to represent the background.", + "bbox": [ + 212, + 690, + 787, + 734 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "To optimize the shape with images, we use differentiable volume rendering [22] to render images given the surface representation. The pixel color $\\hat{\\mathbf{c}}$ and normal $\\hat{\\mathbf{n}}$ are computed as the weighted sum of $M$ samples $\\{\\mathbf{x}_i = \\mathbf{o} + t_i\\mathbf{v}\\}_{i=1}^M$ on a cast-off pixel ray with the camera position $\\mathbf{o}$ and different depth $t_i$ :", + "bbox": [ + 212, + 734, + 787, + 796 + ], + "page_idx": 5 + }, + { + "type": "equation", + "text": "\n$$\n\\hat {\\mathbf {c}} = \\sum_ {i = 1} ^ {M} w _ {i} \\mathbf {c} _ {i}, \\quad \\hat {\\mathbf {n}} = \\sum_ {i = 1} ^ {M} w _ {i} \\mathbf {n} _ {i}, \\tag {5}\n$$\n", + "text_format": "latex", + "bbox": [ + 393, + 801, + 785, + 843 + ], + "page_idx": 5 + }, + { + "type": "page_number", + "text": "6", + "bbox": [ + 217, + 114, + 228, + 126 + ], + "page_idx": 5 + }, + { + "type": "header", + "text": "G. Chen et al.", + "bbox": [ + 271, + 114, + 367, + 126 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "where $w_{i}$ is the blending weight converted from SDF values using the unbiased weight function of NeuS [42], and $\\mathbf{c}_i$ and $\\mathbf{n}_i$ are the color and normal of the sample $\\mathbf{x}_i$ .", + "bbox": [ + 212, + 146, + 782, + 191 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "4.2 Polarimetric and Photometric Losses", + "text_level": 1, + "bbox": [ + 215, + 215, + 565, + 229 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Polarimetric Loss: Previous works [5,34,38,46,54] leverage polarization cues with the following loss function $h^{\\mathrm{o}}$ that is derived from the orthographic polarimetric constraint Eq. (2):", + "bbox": [ + 212, + 243, + 782, + 289 + ], + "page_idx": 6 + }, + { + "type": "equation", + "text": "\n$$\nh ^ {\\mathrm {o}} (\\varphi , \\Delta , \\mathbf {n}) = \\left(\\mathbf {a} ^ {\\mathrm {o}} (\\varphi , \\Delta) \\cdot \\mathbf {n}\\right) ^ {2}, \\mathbf {a} ^ {\\mathrm {o}} (\\varphi , \\Delta) = [ \\sin (\\varphi + \\Delta), \\cos (\\varphi + \\Delta), 0 ], \\tag {6}\n$$\n", + "text_format": "latex", + "bbox": [ + 233, + 301, + 784, + 329 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "where the coefficient vector $\\mathbf{a}^{\\mathrm{o}}$ is calculated with the AoP $\\varphi$ and the disambiguation scalar $\\Delta$ . However, as discussed in [4], Eq. (2) neglects the perspective effect of the lens and may introduce distortions in the reconstruction results.", + "bbox": [ + 212, + 339, + 782, + 383 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Therefore, we instead adopt the perspective polarimetric constraint Eq. (3) [4] and propose a novel loss function $h^{\\mathrm{p}}$ :", + "bbox": [ + 214, + 385, + 782, + 416 + ], + "page_idx": 6 + }, + { + "type": "equation", + "text": "\n$$\nh ^ {\\mathrm {p}} (\\varphi , \\Delta , \\mathbf {v}, \\mathbf {n}) = \\left(\\frac {\\mathbf {a} ^ {\\mathrm {p}} (\\varphi , \\Delta , \\mathbf {v}) \\cdot \\mathbf {n}}{\\| \\mathbf {a} ^ {\\mathrm {p}} (\\varphi , \\Delta , \\mathbf {v}) \\|}\\right) ^ {2} \\tag {7}\n$$\n", + "text_format": "latex", + "bbox": [ + 375, + 428, + 784, + 460 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "with $\\mathbf{a}^{\\mathrm{p}}$ calculated as:", + "bbox": [ + 214, + 470, + 375, + 484 + ], + "page_idx": 6 + }, + { + "type": "equation", + "text": "\n$$\n\\mathbf {a} ^ {\\mathrm {p}} (\\varphi , \\Delta , \\mathbf {v}) = \\left[ v _ {z} \\sin \\varphi^ {\\prime}, v _ {z} \\cos \\varphi^ {\\prime}, - \\left(v _ {y} \\cos \\varphi^ {\\prime} + v _ {x} \\sin \\varphi^ {\\prime}\\right) \\right], \\varphi^ {\\prime} = \\varphi + \\Delta . \\tag {8}\n$$\n", + "text_format": "latex", + "bbox": [ + 236, + 500, + 784, + 515 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "The loss function $h^{\\mathrm{p}}$ additionally considers the camera ray direction $\\mathbf{v}$ to model the perspective effect. In Sec. 5.3, we will show that $h^{\\mathrm{p}}$ effectively improves the reconstruction accuracy.", + "bbox": [ + 212, + 527, + 782, + 571 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Besides, the $\\pi/2$ -ambiguity must be addressed when applying the polarimetric constraint. We combine the merits of PolMVS [5] and PMVIR [54], proposing the segmented loss function $f^{\\mathrm{p}}$ as follows:", + "bbox": [ + 212, + 573, + 782, + 618 + ], + "page_idx": 6 + }, + { + "type": "equation", + "text": "\n$$\nf ^ {\\mathrm {p}} (\\varphi , \\mathbf {v}, \\mathbf {n}, \\rho) = \\left\\{ \\begin{array}{l l} h ^ {\\mathrm {p}} (\\varphi , 0, \\mathbf {v}, \\mathbf {n}) \\cdot h ^ {\\mathrm {p}} (\\varphi , \\pi / 2, \\mathbf {v}, \\mathbf {n}) & \\text {i f} \\rho < \\theta , \\\\ h ^ {\\mathrm {p}} (\\varphi , \\pi / 2, \\mathbf {v}, \\mathbf {n}) & \\text {i f} \\rho \\geq \\theta , \\end{array} \\right. \\tag {9}\n$$\n", + "text_format": "latex", + "bbox": [ + 294, + 631, + 784, + 678 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "where $\\rho$ is the DoP and $\\theta$ is a preset threshold. This design is based on two observations: 1) the reflected light of an object is usually a mixture of the diffuse and specular components, and thus the DoP is relatively low [5], 2) only the DoP of specular reflection of dielectric objects can exceed 0.4 according to Fresnel equations. We use multiplication instead of addition when $\\rho < \\theta$ because only one of the two needs to be minimized.", + "bbox": [ + 212, + 688, + 784, + 777 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Finally, we calculate the polarimetric loss for a batch of sampled pixels $S$ as follows:", + "bbox": [ + 212, + 780, + 782, + 806 + ], + "page_idx": 6 + }, + { + "type": "equation", + "text": "\n$$\n\\mathcal {L} _ {\\mathrm {p o l}} ^ {\\mathrm {p}} = \\frac {1}{| S |} \\sum_ {u \\in \\mathcal {S}} f ^ {\\mathrm {p}} \\left(\\varphi_ {u}, \\mathbf {v} _ {u}, \\hat {\\mathbf {n}} _ {u}, \\rho_ {u}\\right). \\tag {10}\n$$\n", + "text_format": "latex", + "bbox": [ + 383, + 808, + 784, + 842 + ], + "page_idx": 6 + }, + { + "type": "header", + "text": "PISR for Textureless and Specular Objects", + "bbox": [ + 444, + 114, + 732, + 128 + ], + "page_idx": 6 + }, + { + "type": "page_number", + "text": "7", + "bbox": [ + 774, + 116, + 784, + 126 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Note that the camera ray $\\mathbf{v}_u$ and normal $\\hat{\\mathbf{n}}_u$ are transformed into the camera frame according to the camera pose. In Sec. 5.3, the experiments will demonstrate that calculating the loss $\\mathcal{L}_{\\mathrm{pol}}^{\\mathrm{p}}$ with $h^{\\mathrm{p}}$ instead of $h^{\\mathrm{o}}$ can reduce the L1 Chamfer distance by $30\\%$ .", + "bbox": [ + 212, + 146, + 782, + 205 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Photometric Loss: Since the polarimetric loss alone is insufficient for recovering object shapes due to the $\\pi /2$ ambiguity and image noises, we used the photometric loss for initializing and maintaining the basic shape:", + "bbox": [ + 212, + 210, + 782, + 255 + ], + "page_idx": 7 + }, + { + "type": "equation", + "text": "\n$$\n\\mathcal {L} _ {\\text {c o l o r}} = \\frac {1}{| \\mathcal {S} |} \\sum_ {u \\in \\mathcal {S}} \\left| \\hat {\\mathbf {c}} _ {u} - \\mathbf {c} _ {u} \\right|. \\tag {11}\n$$\n", + "text_format": "latex", + "bbox": [ + 408, + 263, + 785, + 297 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "4.3 Normal Regularization", + "text_level": 1, + "bbox": [ + 215, + 316, + 450, + 333 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "The discrete nature of the hash grid makes the estimated SDF less smooth than those purely MLP-based, which can introduce reconstruction artifacts such as holes or cracks. Directly smoothing the normals in 3D space as in [17, 53] can lead to a bumpy surface because the sampled normals are solely constrained by the smoothness loss. Inspired by the belief propagation used in traditional MVS methods [33, 45, 56], we smooth the rendered normals so that all the samples are also constrained by both the photometric loss and the polarimetric loss. We therefore utilize the following normal loss:", + "bbox": [ + 212, + 339, + 784, + 460 + ], + "page_idx": 7 + }, + { + "type": "equation", + "text": "\n$$\n\\mathcal {L} _ {\\text {n o r m a l}} = - \\frac {1}{| \\mathcal {S} _ {\\mathrm {c}} |} \\sum_ {u \\in \\mathcal {S} _ {\\mathrm {c}}} \\frac {1}{| \\mathcal {N} _ {u} |} \\sum_ {\\hat {\\mathbf {n}} _ {j} \\in \\mathcal {N} _ {u}} \\operatorname {S G} \\left(\\hat {\\mathbf {n}} _ {u}\\right) ^ {T} \\hat {\\mathbf {n}} _ {j} \\tag {12}\n$$\n", + "text_format": "latex", + "bbox": [ + 344, + 468, + 785, + 505 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "where $S_{c}$ is a set of pixels sampled among pixels of all the images, $\\mathcal{N}_u$ is a set of pixel normals sampled around $\\hat{\\mathbf{n}}_u$ in the image space. The function SG (\"Stop Gradient\") is used to stop propagating the gradients of the losses w.r.t. $\\hat{\\mathbf{n}}_u$ during backpropagation, i.e., fix the values of $\\hat{\\mathbf{n}}_u$ so that it can be propagated to neighbor pixels.", + "bbox": [ + 212, + 512, + 782, + 587 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Now the pixel sample set $S$ of each iteration during the optimization is the union of the set of center pixels $S_{c}$ and the sets of neighbor pixels $\\mathcal{N}_u$ :", + "bbox": [ + 212, + 588, + 782, + 619 + ], + "page_idx": 7 + }, + { + "type": "equation", + "text": "\n$$\n\\mathcal {S} = \\mathcal {S} _ {c} \\cup \\left\\{\\bigcup_ {u \\in S _ {c}} \\mathcal {N} _ {u} \\right\\} \\tag {13}\n$$\n", + "text_format": "latex", + "bbox": [ + 431, + 628, + 785, + 659 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Empirically, larger $\\mathcal{S}_c$ can lead to better results because it helps capture global structures, which are important for optimizing a neural field [44], while larger $\\mathcal{N}_u$ is beneficial to smooth the surface. Therefore, we sample $\\mathcal{N}_u$ in a criss-cross pattern instead of a whole patch around $\\hat{\\mathbf{n}}_u$ to balance $|\\mathcal{S}_c|$ and $|\\mathcal{N}_u|$ with fixed $|S|$ . In principle, $\\mathcal{N}_u$ can be sampled in other relatively sparse patterns. Here we use the criss-cross one for its simple implementation in code.", + "bbox": [ + 212, + 666, + 784, + 757 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "4.4 Optimization Strategy", + "text_level": 1, + "bbox": [ + 215, + 777, + 446, + 792 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "The overall loss is", + "bbox": [ + 215, + 800, + 349, + 813 + ], + "page_idx": 7 + }, + { + "type": "equation", + "text": "\n$$\n\\mathcal {L} _ {\\text {a l l}} = \\mathcal {L} _ {\\text {c o l o r}} + \\lambda_ {\\mathrm {p}} \\mathcal {L} _ {\\text {p o l}} ^ {\\mathrm {p}} + \\lambda_ {\\mathrm {n}} \\mathcal {L} _ {\\text {n o r m a l}} + \\lambda_ {\\mathrm {e}} \\mathcal {L} _ {\\text {e i k o n a l}} \\tag {14}\n$$\n", + "text_format": "latex", + "bbox": [ + 336, + 825, + 785, + 842 + ], + "page_idx": 7 + }, + { + "type": "page_number", + "text": "8", + "bbox": [ + 217, + 114, + 228, + 126 + ], + "page_idx": 7 + }, + { + "type": "header", + "text": "G. Chen et al.", + "bbox": [ + 271, + 114, + 367, + 127 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/ad2a90825e6a009c1f1cda070f0a1e0a427a491413a6cc3cfbecf8222c46ead7.jpg", + "image_caption": [ + "(a)" + ], + "image_footnote": [], + "bbox": [ + 222, + 143, + 439, + 271 + ], + "page_idx": 8 + }, + { + "type": "image", + "img_path": "images/4482a83617f06baf723fd609e86a07738f14f38a91835588448fc51771ec4e25.jpg", + "image_caption": [ + "(b)", + "Fig. 3: Dataset and collection. (a) Objects with ground-truth meshes. (b) Objects without ground-truth meshes. (c) Capture setup." + ], + "image_footnote": [], + "bbox": [ + 450, + 143, + 550, + 271 + ], + "page_idx": 8 + }, + { + "type": "image", + "img_path": "images/a60ac6b24b2d8417635b8e40757e67e9d17fe15162457302ca6a5593ba9a3cb3.jpg", + "image_caption": [ + "(c)" + ], + "image_footnote": [], + "bbox": [ + 560, + 143, + 781, + 271 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "The extra Eikonal loss $\\mathcal{L}_{\\mathrm{eikonal}} = \\frac{1}{M|S|}\\sum_{k=1}^{M|S|}(\\|\\nabla\\hat{s}(\\mathbf{x}_k)\\| - 1)^2$ is used for encouraging the neural field to approximate a signed distance field [10]. During the optimization, we apply progressive training [17] on the multi-resolution hash grid. To initialize a coarse shape, we set $\\lambda_{\\mathrm{p}} = \\lambda_{\\mathrm{n}} = 0$ and use the photometric and eikonal losses at the early stage. Then we gradually increase $\\lambda_{\\mathrm{p}}$ and $\\lambda_{\\mathrm{n}}$ to rectify the distorted shape. Finally, we gradually decrease $\\lambda_{\\mathrm{n}}$ to 0 to avoid over smoothing the details.", + "bbox": [ + 212, + 353, + 787, + 464 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "5 Experiments", + "text_level": 1, + "bbox": [ + 215, + 489, + 375, + 507 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "5.1 Experimental Setup", + "text_level": 1, + "bbox": [ + 214, + 523, + 426, + 540 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "Dataset: Since there is no public polarization image dataset of textureless and specular objects with ground-truth meshes, we collect a real-world dataset, including four objects (Black Dragon, Red Dragon, Standing Rabbit, Lying Rabbit, Figure and Car) as shown in Fig. 3. The reasons for capturing real object images, rather than synthesizing using a rendering engine, such as [13], are two-fold: 1) the natural ambient lighting conditions are complex and therefore difficult to synthesize, and 2) the shadows and occlusions of the camera can be realistically generated.", + "bbox": [ + 212, + 551, + 787, + 672 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "Each object's polarization images are captured indoors under natural lighting conditions from 40 (60 for Figure for its relatively complex shape) different viewpoints around the object with a color polarization camera [36]. Color images are obtained by first decomposing polarization images into images of four polarizer angles and then averaging and demosaicing. Camera poses are estimated from color images using SuperGlue [32] and COLMAP [33] given pre-calibrated camera intrinsics. We use a checkerboard to help pose estimation. To obtain ground-truth meshes, the objects are coated by 3D scanning spray to eliminate specular reflection and scanned by a structure light-based scanner except for Figure and Car. Please refer to our supplementary material for more details.", + "bbox": [ + 212, + 672, + 787, + 825 + ], + "page_idx": 8 + }, + { + "type": "header", + "text": "PISR for Textureless and Specular Objects", + "bbox": [ + 444, + 114, + 732, + 128 + ], + "page_idx": 8 + }, + { + "type": "page_number", + "text": "9", + "bbox": [ + 774, + 116, + 784, + 126 + ], + "page_idx": 8 + }, + { + "type": "image", + "img_path": "images/af839f7fac945c5590b7ba519169fc518c49a16245ce1ac542e8e1f20729b32e.jpg", + "image_caption": [ + "Baaee" + ], + "image_footnote": [], + "bbox": [ + 250, + 146, + 320, + 200 + ], + "page_idx": 9 + }, + { + "type": "image", + "img_path": "images/402c09e0b7c2f735a2509080170bd6be041b5d78bf0c57c7930944245cccd929.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 336, + 147, + 406, + 200 + ], + "page_idx": 9 + }, + { + "type": "image", + "img_path": "images/afb9c62b8b4d84827e8c7916c0d6d406c7bcb31dad5ebdffb9eb467c5e433a11.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 421, + 146, + 491, + 200 + ], + "page_idx": 9 + }, + { + "type": "image", + "img_path": "images/390a22fd15d12af1da682339f55c73b3fb93e8ea777d4f5b65b6bc39e78a516d.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 508, + 147, + 576, + 199 + ], + "page_idx": 9 + }, + { + "type": "image", + "img_path": "images/7062568af730f6c1253dc2085c3eca3eb5c9e2b11e0598e8202691cfffd82a76.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 593, + 147, + 663, + 199 + ], + "page_idx": 9 + }, + { + "type": "image", + "img_path": "images/61c5d10507c487387c2ef71eeda335f322d42a005aaae000c4ad9f94c32865d0.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 679, + 147, + 750, + 199 + ], + "page_idx": 9 + }, + { + "type": "image", + "img_path": "images/aaa03de075311943c9d8c87a64bae4ce9266bfef6f8e9bda6dcfc4484474a6d0.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 754, + 157, + 784, + 383 + ], + "page_idx": 9 + }, + { + "type": "image", + "img_path": "images/9f30589a6477021164cc8b6be51d64d03893dc6a7f622d12385b259030df037d.jpg", + "image_caption": [ + "Rrnnnne" + ], + "image_footnote": [], + "bbox": [ + 250, + 207, + 318, + 258 + ], + "page_idx": 9 + }, + { + "type": "image", + "img_path": "images/305e0fcc06b9c08958713a3231d6b7fd32065ec2214600a4ca8b473e9497ea41.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 334, + 207, + 406, + 258 + ], + "page_idx": 9 + }, + { + "type": "image", + "img_path": "images/f265945b3935956bd26ef4c63b2293b7ae64cf42f97de268e9eecd7198645d21.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 421, + 207, + 491, + 258 + ], + "page_idx": 9 + }, + { + "type": "image", + "img_path": "images/9606c7417819930d655e83645eb54dc0bd03c34e8818f2154a5f1c59d30bd4cc.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 506, + 207, + 576, + 258 + ], + "page_idx": 9 + }, + { + "type": "image", + "img_path": "images/76b2f19e9fce38ef428ac2d2d9cce2cedbb3d90478328e8f3373fac83c19a65b.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 593, + 207, + 661, + 258 + ], + "page_idx": 9 + }, + { + "type": "image", + "img_path": "images/ff231fc2dc87098a84b70215659f9173c94fa48f3224b4bb3c6b9e63d1d7fa05.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 679, + 207, + 750, + 258 + ], + "page_idx": 9 + }, + { + "type": "image", + "img_path": "images/47cd50d7ccdb63edd1679924aa24fd59372459fbe1f1fa200df06b218b51e0ee.jpg", + "image_caption": [ + "Rabitt S." + ], + "image_footnote": [], + "bbox": [ + 251, + 262, + 315, + 321 + ], + "page_idx": 9 + }, + { + "type": "image", + "img_path": "images/b74f8ea4b5a4492c45304007ba65d9435825d1c380392928732b05fce8d683cd.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 334, + 262, + 401, + 321 + ], + "page_idx": 9 + }, + { + "type": "image", + "img_path": "images/86ce32ee94a377de92169710dda7344dc8412f3c5432cd142b4bd5bf32a0354a.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 421, + 262, + 486, + 321 + ], + "page_idx": 9 + }, + { + "type": "image", + "img_path": "images/bd7ac0a4ff2779891e7487a0635ee61474cd59a66448ca4e96233f5b8db64c5e.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 509, + 263, + 573, + 321 + ], + "page_idx": 9 + }, + { + "type": "image", + "img_path": "images/4439d24aad9e33560e33cd0609d3f4ba9f5709d06677687d8f1e32841fec8096.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 593, + 263, + 660, + 321 + ], + "page_idx": 9 + }, + { + "type": "image", + "img_path": "images/d5bf77d25e8ed23cc261fca96f7e2f0997e9c2c1e0dde4b70937ad091ecff674.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 679, + 263, + 745, + 321 + ], + "page_idx": 9 + }, + { + "type": "image", + "img_path": "images/16097e7320adf1e3539591de277398634af04835b033a7fc4d41b4715c4b8aea.jpg", + "image_caption": [ + "Rabibit L" + ], + "image_footnote": [], + "bbox": [ + 251, + 323, + 315, + 376 + ], + "page_idx": 9 + }, + { + "type": "image", + "img_path": "images/68d8cfb4ca6944f5204923ceeb568d26aa322fb1c02321043fb8ead27430e6a2.jpg", + "image_caption": [ + "Fig. 4: Signed error maps. Red and blue mean surface swelling and shrinking respectively. Regions with higher color saturation indicate longer distances between the reconstruction result and the ground truth. Errors are truncated to within $\\pm 2\\mathrm{mm}$ for better visualization." + ], + "image_footnote": [], + "bbox": [ + 336, + 323, + 401, + 376 + ], + "page_idx": 9 + }, + { + "type": "image", + "img_path": "images/064fbcc29c2aa257066d05b6fe93579dcc289a6747e2726699c43cc8fd673364.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 424, + 323, + 486, + 376 + ], + "page_idx": 9 + }, + { + "type": "image", + "img_path": "images/b38268a9062d69a8ba55f4189b733e82426ac06681ffc84860aa545ea0c7ff00.jpg", + "image_caption": [ + "PANDORA" + ], + "image_footnote": [], + "bbox": [ + 509, + 323, + 573, + 377 + ], + "page_idx": 9 + }, + { + "type": "image", + "img_path": "images/04cc31e212237a2f738d991801c3ac69647ad7ac6414ca84d647ee4142dcf504.jpg", + "image_caption": [ + "PMVIR" + ], + "image_footnote": [], + "bbox": [ + 594, + 323, + 658, + 377 + ], + "page_idx": 9 + }, + { + "type": "image", + "img_path": "images/847a5e8d15f4b341aeb9f51669bb0703530c86b954db100484761f3fdb161633.jpg", + "image_caption": [ + "PISR (ours)" + ], + "image_footnote": [], + "bbox": [ + 679, + 323, + 745, + 377 + ], + "page_idx": 9 + }, + { + "type": "table", + "img_path": "images/7a85cb613c7375159a74e0c94a83c6484789eaf2b256b61eb004c7a13c807f03.jpg", + "table_caption": [ + "Table 1: Summary of Methods in Experiments" + ], + "table_footnote": [], + "table_body": "
MethodsImageShape Repre.Appearance Repre.
Ref-NeuS [9]RGBNeural SDFNeural radiance
NeRO [18]RGBNeural SDFNeural BRDF
NeuS [42]RGBNeural SDFNeural radiance
PANDORA [6]Pol. RGBNeural SDFNeural BRDF
PMVIR [54]Pol. RGBMeshSpherical harmonics
PISR (Ours)Pol. RGBGrids+Neural SDFNeural radiance
", + "bbox": [ + 261, + 506, + 740, + 621 + ], + "page_idx": 9 + }, + { + "type": "text", + "text": "Performance Metrics: Since we focus on geometry reconstruction, we only evaluate the surface reconstruction quality using L1 Chamfer distance and F-score. We remove the background meshes before calculating the metrics.", + "bbox": [ + 212, + 655, + 787, + 702 + ], + "page_idx": 9 + }, + { + "type": "text", + "text": "Implementation Details: The SDF MLP $\\varPhi_s$ has one hidden layer with a size of 64, and the color MLP $\\varPhi_{\\mathbf{c}}$ has two hidden layers with sizes of 64. The multi-resolution grid contains 16 levels of resolution, and the coarsest and finest resolutions are 32 and 2700, respectively. The hash table size of each level is $2^{19}$ . The optimization of the grid starts at the fourth-coarsest resolutions and progressively extends one level every 1000 iterations. Throughout the whole optimization process, the threshold of DoP $\\theta$ is set to 0.3 and $\\lambda_{\\mathrm{eikonal}}$ is set to 0.1. After the first 2.5k iterations, $\\lambda_{\\mathrm{pol}}$ , $\\lambda_{\\mathrm{normal}}$ and $|\\mathcal{N}_p|$ are linearly increased from", + "bbox": [ + 212, + 718, + 787, + 842 + ], + "page_idx": 9 + }, + { + "type": "page_number", + "text": "10", + "bbox": [ + 217, + 114, + 235, + 126 + ], + "page_idx": 9 + }, + { + "type": "header", + "text": "G. Chen et al.", + "bbox": [ + 271, + 114, + 367, + 126 + ], + "page_idx": 9 + }, + { + "type": "table", + "img_path": "images/a2a510f2bee08f6e4e9b0cb345e02ba4786afdcfcec80128f9dfa015fe214eeb.jpg", + "table_caption": [ + "Table 2: Reconstruction accuracy in Chamfer distance (CD↓, mm) and F-score (FS↑, %). FS is calculated at a threshold of 1 mm. Note that the values in the table are rounded to one decimal place. The CD of PMVIR and PISR on Rabbit L. are 0.584 mm and 0.578 mm respectively. Bold, underline, and double underline mean the best, second-best and third-best performance, respectively." + ], + "table_footnote": [], + "table_body": "
MethodsDragon B.Dragon R.Rabbit S.Rabbit L.Avg.
CDFSCDFSCDFSCDFSCDFS
RGBRef-NeuS2.165.32.166.11.284.51.090.51.676.6
NeRO2.459.21.866.31.380.51.385.01.772.8
NeuS1.967.11.286.01.186.80.993.01.383.2
RGBPANDORA2.355.31.864.41.478.31.577.11.868.8
PMVIR1.876.51.090.50.697.50.697.31.090.5
PISR (Ours)0.599.90.699.10.599.70.699.40.599.5
", + "bbox": [ + 238, + 224, + 766, + 363 + ], + "page_idx": 10 + }, + { + "type": "text", + "text": "0 to 2.0, 1.0 and 28 in $2.5\\mathrm{k}$ iterations, respectively. Then $\\lambda_{\\mathrm{normal}}$ and $|\\mathcal{N}_p|$ are linearly decreased to 0 in another $2.5\\mathrm{k}$ iterations. The maximum number of iterations is $20\\mathrm{k}$ . The final SDF is converted to meshes using Marching Cubes [20].", + "bbox": [ + 212, + 387, + 787, + 435 + ], + "page_idx": 10 + }, + { + "type": "text", + "text": "Baselines: We compare our method with five methods, including PMVIR [54], PANDORA [6], NeRO [18], Ref-NeuS [9] and NeuS [42]. Tab. 1 summarizes the type of input images, geometry representation and color representation of all the six methods. Since PMVIR is a refinement method based on an initial shape, we provide the coarse initial meshes reconstructed by PISR-C (see the ablation study in Sec. 5.3 for the details of PISR-C).", + "bbox": [ + 212, + 446, + 787, + 540 + ], + "page_idx": 10 + }, + { + "type": "text", + "text": "5.2 Results", + "text_level": 1, + "bbox": [ + 214, + 559, + 326, + 573 + ], + "page_idx": 10 + }, + { + "type": "text", + "text": "Quantitative Results: The Chamfer distances and F-scores of the reconstruction results are reported in Tab. 2. Our method achieves the best reconstruction quality on all four objects shown in Fig. 3a and significantly outperforms the other five methods. The average Chamfer distance and F-score of our method are $0.5 \\, \\mathrm{mm}$ and $99.5\\%$ , respectively. As the second-best method in the comparison, PMVIR achieves twice the Chamfer distance and a $9\\%$ lower F-score compared to ours. Both PMVIR and our method utilize polarimetric constraints, and the results show that the geometry cue of polarization is effective for shape estimation. Although PANDORA also uses polarization images as input, it implicitly utilizes polarization cues through polarimetric rendering, i.e., optimizing shapes using an image reconstruction loss, which could suffer from the shape-radiance ambiguity as RGB-based methods. Among RGB-based methods, NeuS achieves better results than Ref-NeuS and NeRO which are designed for specular objects.", + "bbox": [ + 212, + 583, + 787, + 780 + ], + "page_idx": 10 + }, + { + "type": "text", + "text": "We visualize the error distributions in Fig. 4. For methods that rely solely on image reconstruction losses (Ref-NeuS, NeRO, NeuS and PANDORA), their reconstruction errors in non-convex regions are notably higher compared to", + "bbox": [ + 212, + 794, + 787, + 840 + ], + "page_idx": 10 + }, + { + "type": "header", + "text": "PISR for Textureless and Specular Objects", + "bbox": [ + 444, + 114, + 730, + 128 + ], + "page_idx": 10 + }, + { + "type": "page_number", + "text": "11", + "bbox": [ + 767, + 114, + 782, + 126 + ], + "page_idx": 10 + }, + { + "type": "image", + "img_path": "images/4ff13a0f845138e44cc2454e391498a61ed2575212ca31a0f43b8e964a7357a8.jpg", + "image_caption": [ + "Initialization" + ], + "image_footnote": [], + "bbox": [ + 218, + 142, + 405, + 275 + ], + "page_idx": 11 + }, + { + "type": "image", + "img_path": "images/0b11c7fca60e3e8be9ed524bbd37ca08b6d52b836ed3a90e0dd9c8e87cd4071c.jpg", + "image_caption": [ + "PMVIR" + ], + "image_footnote": [], + "bbox": [ + 410, + 142, + 596, + 273 + ], + "page_idx": 11 + }, + { + "type": "image", + "img_path": "images/cd9c71569bc759d4b33bef45b1e71bf08f5bf8134105d35b2bf99a2aed20cd9b.jpg", + "image_caption": [ + "PISR", + "Fig. 5: Normal maps of Dragon B. before and after optimizations. Using neural SDFs to represent shapes allows changes in the shape topology during the optimization." + ], + "image_footnote": [], + "bbox": [ + 601, + 142, + 787, + 273 + ], + "page_idx": 11 + }, + { + "type": "table", + "img_path": "images/b3a9476d2ff4a0e394e00ccd1e1d06962d30135887cfd0bdd4af910454970a06.jpg", + "table_caption": [ + "Table 3: Average optimization time of different methods." + ], + "table_footnote": [], + "table_body": "
MethodsRef-NeuSNeRONeuSPANDORAPMVIRPISR (Ours)
Iterations200k200k200k100k420k
Time (h)20.020.510.815.02.00.5
", + "bbox": [ + 246, + 364, + 754, + 422 + ], + "page_idx": 11 + }, + { + "type": "text", + "text": "PMVIR and our method. While the inherent continuity of MLPs ensures the watertightness of their reconstructed shapes, their accuracy cannot be guaranteed due to the shape-radiance ambiguity introduced by specular reflections and the lack of textures. Additionally, for NeRO and PANDORA, the lighting shadows cast by the camera further exacerbate the difficulty of estimating surface material parameters. The real HDR ambient lighting also poses challenges for PANDORA and Ref-NeuS to accurately estimate object shapes. In contrast, PMVIR and our method can recover non-convex regions using the polarimetric constraints. Despite the polarization cues, representing the shapes as neural SDF helps our method better reconstruct regions that are poorly initialized. As shown in Fig. 5, neural SDF allows continuous changes of the shape topology during optimization, which is crucial for evolving surfaces from holes.", + "bbox": [ + 212, + 449, + 787, + 630 + ], + "page_idx": 11 + }, + { + "type": "text", + "text": "**Optimization Time:** We report the average optimization time in Tab. 3. The experiments are conducted on a machine equipped with an Intel Core i9-12900K and an NVIDIA RTX 3090 GPU. We set the maximum number of iterations to $200\\mathrm{k}$ . If the shape converges, we stop the optimization earlier. Our method represents shape as a multi-resolution hash grid and two shallow MLPs, significantly reducing the time complexity for rendering a pixel. Therefore, our method converges much faster, achieving up to $4 \\sim 41 \\times$ convergence speed while maintaining the highest reconstruction accuracy.", + "bbox": [ + 212, + 645, + 787, + 767 + ], + "page_idx": 11 + }, + { + "type": "text", + "text": "5.3 Ablation Study", + "text_level": 1, + "bbox": [ + 214, + 786, + 388, + 801 + ], + "page_idx": 11 + }, + { + "type": "text", + "text": "We evaluate the effectiveness of the polarimetric loss and normal regularization, as well as the influence of the perspective effect of the lens. We compare PISR(-", + "bbox": [ + 212, + 809, + 785, + 839 + ], + "page_idx": 11 + }, + { + "type": "page_number", + "text": "12", + "bbox": [ + 217, + 114, + 235, + 126 + ], + "page_idx": 11 + }, + { + "type": "header", + "text": "G. Chen et al.", + "bbox": [ + 271, + 114, + 367, + 126 + ], + "page_idx": 11 + }, + { + "type": "table", + "img_path": "images/2db590b15ae31261eee482fe89e4975ebdc7fc381d50dd96f428db69490000da.jpg", + "table_caption": [ + "Table 4: Results of the ablation study on different combinations of losses. The L1 CD↓ (mm) and FSs↑ (%) at 1.0 mm and 0.5 are the averages over four objects in Fig.3a. The ranks are the averages over the three metrics. $\\mathcal{L}_{\\mathrm{pol}}^{\\mathrm{o}}$ differs from $\\mathcal{L}_{\\mathrm{pol}}^{\\mathrm{p}}$ by its use of the orthographic constraint Eq. 2 instead of the perspective one Eq. 3. Note that $\\mathcal{L}_{\\mathrm{eikonal}}$ is used in all experiments and is omitted it in this table." + ], + "table_footnote": [], + "table_body": "
VariantsUsed LossesAvg.Rank
LcolorLpoiLpolLnormalCDFS(1.0)FS(0.5)
PISR-C✓1.1088.1070.306
PISR-N✓✓1.0088.2071.105
PISR-O✓✓0.8096.6075.543
PISR-ON✓✓✓0.7796.1075.504
PISR-P✓✓0.5199.7692.791
PISR(-PN)✓✓✓0.5499.5491.752
", + "bbox": [ + 251, + 224, + 751, + 362 + ], + "page_idx": 12 + }, + { + "type": "text", + "text": "PN) with its five variants, referred to as $PISR-P$ , $PISR-ON$ , $PISR-O$ , $PISR-N$ and $PISR-C$ , based on different combinations of losses as shown in Tab. 4. The quantitative experiment is conducted on objects shown in Fig. 3a and the qualitative experiment is conducted on objects shown in Fig. 3b.", + "bbox": [ + 215, + 390, + 785, + 450 + ], + "page_idx": 12 + }, + { + "type": "text", + "text": "As shown in Tab. 4, the top four performing variants are all incorporated with the polarimetric losses, and fine-grained details can be recovered without complicated appearance modeling. As for the normal regularization, although the quantitative comparison between PISR-PN and PISR-P shows that it slightly reduces the accuracy, the qualitative results in Fig. 6 show its effectiveness for recovering surfaces from severe distortions.", + "bbox": [ + 215, + 450, + 785, + 541 + ], + "page_idx": 12 + }, + { + "type": "text", + "text": "Among the top four performing variants, PISR-PN and PISR-P achieve higher accuracy by considering the perspective effect of the lens using Eq. (3) instead of Eq. (2). As the example error maps of Rabbit L. shown in Fig. 7, the perspective effect leads to surface swelling and shrinking, causing severe cracks on the surface. This phenomenon does not appear in the results of PMVIR and PANDORA. We speculate the reason to be that the hash-grid-based neural SDF we use is too flexible in topology compared to the mesh-based one used in PMVIR, and lacks the smoothness of purely MLP-based one used in PANDORA. Besides, the results in Tab. 4 show that PISR-PN outperforms PISR-ON by a $30\\%$ reduction in L1 Chamfer distance and a $17\\%$ improvement in F-score at $0.5 \\mathrm{~mm}$ , which suggests the necessity of considering the perspective effect to improve the accuracy of polarimetric 3D reconstruction.", + "bbox": [ + 215, + 541, + 785, + 723 + ], + "page_idx": 12 + }, + { + "type": "text", + "text": "6 Conclusion", + "text_level": 1, + "bbox": [ + 217, + 746, + 357, + 762 + ], + "page_idx": 12 + }, + { + "type": "text", + "text": "We have proposed PISR, an accurate and efficient polarimetric neural implicit surface reconstruction method for textureless and specular objects. The key idea of PISR is to explicitly regularize the SDF using the polarimetric constraints to avoid the shape-radiance ambiguity. More importantly, PISR achieves high accu", + "bbox": [ + 215, + 779, + 785, + 839 + ], + "page_idx": 12 + }, + { + "type": "header", + "text": "PISR for Textureless and Specular Objects", + "bbox": [ + 444, + 114, + 730, + 128 + ], + "page_idx": 12 + }, + { + "type": "page_number", + "text": "13", + "bbox": [ + 767, + 114, + 784, + 126 + ], + "page_idx": 12 + }, + { + "type": "image", + "img_path": "images/98032b4b80934bcbd775b5d12cd8772aa34459ecad177d68ece041a695612cf2.jpg", + "image_caption": [ + "Fig. 6: Normal maps of Figure and Car. The polarimetric loss and normal loss jointly recover the shape from distortions induced by strong reflections." + ], + "image_footnote": [], + "bbox": [ + 215, + 142, + 784, + 373 + ], + "page_idx": 13 + }, + { + "type": "image", + "img_path": "images/d5d3b40d27e58a5369d5df98e9c3be7a9f49ba008cd2cf5197d0147d817945ac.jpg", + "image_caption": [ + "Fig. 7: Signed error maps of Rabbit L. PISR-PN leverages the perspective polarimetric constraint to eliminate surface cracks encountered by PISR-ON. Additional error maps can be found in our supplementary material." + ], + "image_footnote": [], + "bbox": [ + 276, + 444, + 715, + 575 + ], + "page_idx": 13 + }, + { + "type": "text", + "text": "racy by leveraging the perspective polarimetric constraint, eliminating the surface cracks caused by the traditional orthographic one. Besides, PISR regularizes surface normals in image space for robustness and integrates the multi-resolution hash grid for efficiency.", + "bbox": [ + 212, + 678, + 787, + 739 + ], + "page_idx": 13 + }, + { + "type": "text", + "text": "Limitations: Despite the high reconstruction accuracy in the experiments, PISR's performance degenerates with noisy AoP maps, limiting its application in reconstructing rough, well-textured or mirror-like objects. Recent works [16,18] have shown solutions to mitigate this problem. In addition, PISR also requires dense input views, adding shape constraints derived from normal and depth estimation models would be a promising solution.", + "bbox": [ + 212, + 750, + 787, + 840 + ], + "page_idx": 13 + }, + { + "type": "page_number", + "text": "14", + "bbox": [ + 217, + 114, + 235, + 126 + ], + "page_idx": 13 + }, + { + "type": "header", + "text": "G. Chen et al.", + "bbox": [ + 271, + 114, + 367, + 126 + ], + "page_idx": 13 + }, + { + "type": "text", + "text": "References", + "text_level": 1, + "bbox": [ + 217, + 143, + 321, + 159 + ], + "page_idx": 14 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "1. Atkinson, G.A., Hancock, E.R.: Recovery of surface orientation from diffuse polarization. IEEE transactions on image processing 15(6), 1653-1664 (2006)", + "2. Ba, Y., Gilbert, A., Wang, F., Yang, J., Chen, R., Wang, Y., Yan, L., Shi, B., Kadambi, A.: Deep shape from polarization. In: European Conference on Computer Vision. pp. 554-571. Springer (2020)", + "3. Baek, S.H., Zeltner, T., Ku, H., Hwang, I., Tong, X., Jakob, W., Kim, M.H.: Image-based acquisition and modeling of polarimetric reflectance. ACM Trans. Graph. 39(4), 139 (2020)", + "4. Chen, G., He, L., Guan, Y., Zhang, H.: Perspective phase angle model for polarimetric 3d reconstruction. In: European Conference on Computer Vision. pp. 398-414. Springer (2022)", + "5. Cui, Z., Gu, J., Shi, B., Tan, P., Kautz, J.: Polarimetric multi-view stereo. In: Proceedings of the IEEE conference on computer vision and pattern recognition. pp. 1558-1567 (2017)", + "6. Dave, A., Zhao, Y., Veeraraghavan, A.: Pandora: Polarization-aided neural decomposition of radiance. In: European Conference on Computer Vision. pp. 538-556. Springer (2022)", + "7. Ding, Y., Ji, Y., Ye, J.: Polar-photometric stereo under natural illumination. In: 2022 International Conference on 3D Vision (3DV). pp. 690–699. IEEE (2022)", + "8. Ding, Y., Ji, Y., Zhou, M., Kang, S.B., Ye, J.: Polarimetric helmholtz stereopsis. In: Proceedings of the IEEE/CVF International Conference on Computer Vision. pp. 5037-5046 (2021)", + "9. Ge, W., Hu, T., Zhao, H., Liu, S., Chen, Y.C.: Ref-neus: Ambiguity-reduced neural implicit surface learning for multi-view reconstruction with reflection. arXiv preprint arXiv:2303.10840 (2023)", + "0. Gropp, A., Yariv, L., Haim, N., Atzmon, M., Lipman, Y.: Implicit geometric regularization for learning shapes. In: Proceedings of the 37th International Conference on Machine Learning. pp. 3789-3799 (2020)", + "1. Hwang, I., Jeon, D.S., Munoz, A., Gutierrez, D., Tong, X., Kim, M.H.: Sparse ellipsometry: portable acquisition of polarimetric svbrdf and shape with unstructured flash photography. ACM Transactions on Graphics (TOG) 41(4), 1-14 (2022)", + "2. Ichikawa, T., Fukao, Y., Nobuhara, S., Nishino, K.: Fresnel microfacet brdf: Unification of polari-radiometric surface-body reflection. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 16489-16497 (2023)", + "3. Jakob, W., Speierer, S., Roussel, N., Nimier-David, M., Vicini, D., Zeltner, T., Nicolet, B., Crespo, M., Leroy, V., Zhang, Z.: Mitsuba 3 renderer (2022), https://mitsuba-renderer.org", + "4. Kondo, Y., Ono, T., Sun, L., Hirasawa, Y., Murayama, J.: Accurate polarimetric brdf for real polarization scene rendering. In: Computer Vision-ECCV 2020: 16th European Conference, Glasgow, UK, August 23-28, 2020, Proceedings, Part XIX 16. pp. 220-236. Springer (2020)", + "5. Lei, C., Qi, C., Xie, J., Fan, N., Koltun, V., Chen, Q.: Shape from polarization for complex scenes in the wild. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR). pp. 12632-12641 (June 2022)", + "6. Li, C., Ono, T., Uemori, T., Mihara, H., Gatto, A., Nagahara, H., Moriuchi, Y.: Neisf: Neural incident stokes field for geometry and material estimation. arXiv preprint arXiv:2311.13187 (2023)" + ], + "bbox": [ + 225, + 175, + 784, + 839 + ], + "page_idx": 14 + }, + { + "type": "header", + "text": "PISR for Textureless and Specular Objects", + "bbox": [ + 444, + 114, + 730, + 128 + ], + "page_idx": 14 + }, + { + "type": "page_number", + "text": "15", + "bbox": [ + 767, + 116, + 784, + 126 + ], + "page_idx": 14 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "17. Li, Z., Müller, T., Evans, A., Taylor, R.H., Unberath, M., Liu, M.Y., Lin, C.H.: Neuralangelo: High-fidelity neural surface reconstruction. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 8456-8465 (2023)", + "18. Liu, Y., Wang, P., Lin, C., Long, X., Wang, J., Liu, L., Komura, T., Wang, W.: Nano: Neural geometry and brdf reconstruction of reflective objects from multiview images. In: SIGGRAPH (2023)", + "19. Long, X., Lin, C., Wang, P., Komura, T., Wang, W.: Sparseneus: Fast generalizable neural surface reconstruction from sparse views. In: European Conference on Computer Vision. pp. 210-227. Springer (2022)", + "20. Lorensen, W.E., Cline, H.E.: Marching cubes: A high resolution 3d surface construction algorithm. ACM SIGGRAPH Computer Graphics 21(4), 163-169 (1987)", + "21. Lyu, Y., Zhao, L., Li, S., Shi, B.: Shape from polarization with distant lighting estimation. IEEE Transactions on Pattern Analysis and Machine Intelligence 45(11), 13991-14004 (2023). https://doi.org/10.1109/TPAMI.2023.3298376", + "22. Martin-Brualla, R., Radwan, N., Sajjadi, M.S., Barron, J.T., Dosovitskiy, A., Duckworth, D.: Nerf in the wild: Neural radiance fields for unconstrained photo collections. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 7210–7219 (2021)", + "23. Mildenhall, B., Srinivasan, P.P., Tancik, M., Barron, J.T., Ramamoorthi, R., Ng, R.: Nerf: Representing scenes as neural radiance fields for view synthesis. Communications of the ACM 65(1), 99-106 (2021)", + "24. Miyazaki, D., Shigetomi, T., Baba, M., Furukawa, R., Hiura, S., Asada, N.: Polarization-based surface normal estimation of black specular objects from multiple viewpoints. In: 2012 Second International Conference on 3D Imaging, Modeling, Processing, Visualization & Transmission. pp. 104-111. IEEE (2012)", + "25. Miyazaki, D., Tan, R.T., Hara, K., Ikeuchi, K.: Polarization-based inverse rendering from a single view. In: Computer Vision, IEEE International Conference on. vol. 3, pp. 982-982. IEEE Computer Society (2003)", + "26. Müller, T., Evans, A., Schied, C., Keller, A.: Instant neural graphics primitives with a multiresolution hash encoding. ACM Transactions on Graphics (ToG) 41(4), 1-15 (2022)", + "27. Ngo Thanh, T., Nagahara, H., Taniguchi, R.i.: Shape and light directions from shading and polarization. In: Proceedings of the IEEE conference on computer vision and pattern recognition. pp. 2310-2318. IEEE (2015)", + "28. Niemeyer, M., Mescheder, L., Oechsle, M., Geiger, A.: Differentiable volumetric rendering: Learning implicit 3d representations without 3d supervision. In: Proc. IEEE Conf. on Computer Vision and Pattern Recognition (CVPR) (2020)", + "29. Oechsle, M., Peng, S., Geiger, A.: Unisurf: Unifying neural implicit surfaces and radiance fields for multi-view reconstruction. In: International Conference on Computer Vision (ICCV) (2021)", + "30. Park, J.J., Florence, P., Straub, J., Newcombe, R., Lovegrove, S.: Deepsdf: Learning continuous signed distance functions for shape representation. In: Proceedings of the IEEE/CVF conference on computer vision and pattern recognition. pp. 165-174 (2019)", + "31. Rahmann, S., Canterakis, N.: Reconstruction of specular surfaces using polarization imaging. In: Proceedings of the 2001 IEEE Computer Society Conference on Computer Vision and Pattern Recognition. CVPR 2001. vol. 1, pp. I-I. IEEE (2001)", + "32. Sarlin, P.E., DeTone, D., Malisiewicz, T., Rabinovich, A.: SuperGlue: Learning feature matching with graph neural networks. In: CVPR (2020)" + ], + "bbox": [ + 215, + 146, + 785, + 840 + ], + "page_idx": 15 + }, + { + "type": "page_number", + "text": "16", + "bbox": [ + 217, + 114, + 235, + 126 + ], + "page_idx": 15 + }, + { + "type": "header", + "text": "G. Chen et al.", + "bbox": [ + 271, + 114, + 367, + 126 + ], + "page_idx": 15 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "33. Schonberger, J.L., Frahm, J.M.: Structure-from-motion revisited. In: Conference on Computer Vision and Pattern Recognition (CVPR) (2016)", + "34. Shakeri, M., Loo, Y., Zhang, H.: Polarimetric monocular dense mapping using relative deep depth prior [j]. 2021. Shakeri M Loo Y Zhang H Polarimetric Monocular Dense Mapping Using Relative Deep Depth Prior (2021)", + "35. Shao, M., Xia, C., Yang, Z., Huang, J., Wang, X.: Transparent shape from a single view polarization image. In: Proceedings of the IEEE/CVF International Conference on Computer Vision (ICCV). pp. 9277-9286 (October 2023)", + "36. Sony: Polarization image sensor with four-directional on-chip polarizer and global shutter function, https://www.sony-semicon.co.jp/e/products/IS/industry/product/polarization.html, accessed 2024-02-04", + "37. Tian, C., Pan, W., Wang, Z., Mao, M., Zhang, G., Bao, H., Tan, P., Cui, Z.: Dpsnet: Deep polarimetric stereo depth estimation. In: The IEEE/CVF International Conference on Computer Vision (ICCV) (2023)", + "38. Tozza, S., Zhu, D., Smith, W., Ramamoorthi, R., Hancock, E.: Uncalibrated, two source photo-polarimetric stereo. IEEE Transactions on Pattern Analysis and Machine Intelligence (2021)", + "39. Verbin, D., Hedman, P., Mildenhall, B., Zickler, T., Barron, J.T., Srinivasan, P.P.: Ref-nerf: Structured view-dependent appearance for neural radiance fields. In: 2022 IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR). pp. 5481-5490. IEEE (2022)", + "40. Wang, F., Rakotosaona, M.J., Niemeyer, M., Szeliski, R., Pollefeys, M., Tombari, F.: Unisdf: Unifying neural representations for high-fidelity 3d reconstruction of complex scenes with reflections. In: arXiv (2023)", + "41. Wang, H., Wang, J., Agapito, L.: Co-slam: Joint coordinate and sparse parametric encodings for neural real-time slam. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 13293-13302 (2023)", + "42. Wang, P., Liu, L., Liu, Y., Theobalt, C., Komura, T., Wang, W.: Neus: Learning neural implicit surfaces by volume rendering for multi-view reconstruction. Advances in Neural Information Processing Systems 34, 27171-27183 (2021)", + "43. Wang, Y., Han, Q., Habermann, M., Daniilidis, K., Theobalt, C., Liu, L.: Neus2: Fast learning of neural implicit surfaces for multi-view reconstruction. In: Proceedings of the IEEE/CVF International Conference on Computer Vision. pp. 3295-3306 (2023)", + "44. Xie, Z., Yang, X., Yang, Y., Sun, Q., Jiang, Y., Wang, H., Cai, Y., Sun, M.: S3im: Stochastic structural similarity and its unreasonable effectiveness for neural fields. In: Proceedings of the IEEE/CVF International Conference on Computer Vision. pp. 18024-18034 (2023)", + "45. Xu, Q., Tao, W.: Multi-scale geometric consistency guided multi-view stereo. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR) (June 2019)", + "46. Yang, L., Tan, F., Li, A., Cui, Z., Furukawa, Y., Tan, P.: Polarimetric dense monocular slam. In: Proceedings of the IEEE conference on computer vision and pattern recognition. pp. 3857-3866 (2018)", + "47. Yariv, L., Gu, J., Kasten, Y., Lipman, Y.: Volume rendering of neural implicit surfaces. In: Thirty-Fifth Conference on Neural Information Processing Systems (2021)", + "48. Yariv, L., Kasten, Y., Moran, D., Galun, M., Atzmon, M., Ronen, B., Lipman, Y.: Multiview neural surface reconstruction by disentangling geometry and appearance. Advances in Neural Information Processing Systems 33 (2020)" + ], + "bbox": [ + 215, + 147, + 784, + 839 + ], + "page_idx": 16 + }, + { + "type": "header", + "text": "PISR for Textureless and Specular Objects", + "bbox": [ + 444, + 114, + 730, + 128 + ], + "page_idx": 16 + }, + { + "type": "page_number", + "text": "17", + "bbox": [ + 767, + 116, + 784, + 126 + ], + "page_idx": 16 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "49. Yu, Y., Zhu, D., Smith, W.A.: Shape-from-polarisation: a nonlinear least squares approach. In: Proceedings of the IEEE International Conference on Computer Vision Workshops. pp. 2969-2976 (2017)", + "50. Yu, Z., Peng, S., Niemeyer, M., Sattler, T., Geiger, A.: Monosdf: Exploring monocular geometric cues for neural implicit surface reconstruction. Advances in neural information processing systems 35, 25018-25032 (2022)", + "51. Zhang, K., Luan, F., Wang, Q., Bala, K., Snavely, N.: PhySG: Inverse rendering with spherical gaussians for physics-based material editing and relighting. In: The IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR) (2021)", + "52. Zhang, K., Riegler, G., Snavely, N., Koltun, V.: Nerf++: Analyzing and improving neural radiance fields. arXiv:2010.07492 (2020)", + "53. Zhang, X., Srinivasan, P.P., Deng, B., Debevec, P., Freeman, W.T., Barron, J.T.: Nerfactor: Neural factorization of shape and reflectance under an unknown illumination. ACM Transactions on Graphics (ToG) 40(6), 1-18 (2021)", + "54. Zhao, J., Monno, Y., Okutomi, M.: Polarimetric multi-view inverse rendering. IEEE Transactions on Pattern Analysis and Machine Intelligence (2022)", + "55. Zhao, J., Oishi, J., Monno, Y., Okutomi, M.: Polarimetric patchmatch multi-view stereo. In: Proceedings of the IEEE/CVF Winter Conference on Applications of Computer Vision. pp. 3476-3484 (2024)", + "56. Zheng, E., Dunn, E., Jojic, V., Frahm, J.M.: Patchmatch based joint view selection and depthmap estimation. In: Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition (CVPR) (June 2014)" + ], + "bbox": [ + 215, + 146, + 785, + 467 + ], + "page_idx": 17 + }, + { + "type": "page_number", + "text": "18", + "bbox": [ + 217, + 114, + 235, + 126 + ], + "page_idx": 17 + }, + { + "type": "header", + "text": "G. Chen et al.", + "bbox": [ + 271, + 114, + 367, + 127 + ], + "page_idx": 17 + } +] \ No newline at end of file diff --git a/2024/PISR_ Polarimetric Neural Implicit Surface Reconstruction for Textureless and Specular Objects/949f096c-d4bf-4988-bb1a-9febc3a46def_model.json b/2024/PISR_ Polarimetric Neural Implicit Surface Reconstruction for Textureless and Specular Objects/949f096c-d4bf-4988-bb1a-9febc3a46def_model.json new file mode 100644 index 0000000000000000000000000000000000000000..d490dddf5a21edaa9dd78bff7369649ff22cdfd1 --- /dev/null +++ b/2024/PISR_ Polarimetric Neural Implicit Surface Reconstruction for Textureless and Specular Objects/949f096c-d4bf-4988-bb1a-9febc3a46def_model.json @@ -0,0 +1,2887 @@ +[ + [ + { + "type": "title", + "bbox": [ + 0.278, + 0.141, + 0.726, + 0.21 + ], + "angle": 0, + "content": "PISR: Polarimetric Neural Implicit Surface Reconstruction for Textureless and Specular Objects" + }, + { + "type": "text", + "bbox": [ + 0.237, + 0.234, + 0.766, + 0.252 + ], + "angle": 0, + "content": "Guangcheng Chen\\(^{1,2}\\), Yicheng He\\(^{1,2}\\), Li He\\(^{1,2}\\), and Hong Zhang\\(^{\\star 1,2}\\)" + }, + { + "type": "text", + "bbox": [ + 0.23, + 0.262, + 0.774, + 0.306 + ], + "angle": 0, + "content": "\\(^{1}\\) Southern University of Science and Technology, Shenzhen, China \\(^{2}\\) Shenzhen Key Laboratory of Robotics and Computer Vision {chengc2023,12332136}@mail.sustech.edu.cn, {hel, hzhang}@sustech.edu.cn" + }, + { + "type": "text", + "bbox": [ + 0.261, + 0.342, + 0.744, + 0.538 + ], + "angle": 0, + "content": "Abstract. Neural implicit surface reconstruction has achieved remarkable progress recently. Despite resorting to complex radiance modeling, state-of-the-art methods still struggle with textureless and specular surfaces. Different from RGB images, polarization images can provide direct constraints on the azimuth angles of the surface normals. In this paper, we present PISR, a novel method that utilizes a geometrically accurate polarimetric loss to refine shape independently of appearance. In addition, PISR smooths surface normals in image space to eliminate severe shape distortions and leverages the hash-grid-based neural signed distance function to accelerate the reconstruction. Experimental results demonstrate that PISR achieves higher accuracy and robustness, with an L1 Chamfer distance of \\(0.5\\mathrm{mm}\\) and an F-score of \\(99.5\\%\\) at \\(1\\mathrm{mm}\\), while converging \\(4 \\sim 30\\times\\) faster than previous polarimetric surface reconstruction methods." + }, + { + "type": "text", + "bbox": [ + 0.261, + 0.55, + 0.741, + 0.579 + ], + "angle": 0, + "content": "Keywords: Polarization \\(\\cdot\\) Surface reconstruction \\(\\cdot\\) Multi-view reconstruction \\(\\cdot\\) Neural implicit surface \\(\\cdot\\) Polarimetric constraint" + }, + { + "type": "title", + "bbox": [ + 0.216, + 0.605, + 0.377, + 0.621 + ], + "angle": 0, + "content": "1 Introduction" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.637, + 0.788, + 0.804 + ], + "angle": 0, + "content": "Surface reconstruction from multi-view images is a fundamental problem in computer vision, with various applications in graphics, robotics, and more. Recently, neural implicit surface has shown powerful flexibility for representing 3D shape and appearance. Combined with differentiable volume rendering [23], neural implicit surface reconstruction is able to estimate 3D shapes from multi-view images accurately [42, 47]. However, these methods struggle with textureless and specular objects due to the shape-radiance ambiguity [52]. Recent methods resort to diffuse and specular separation modeling [40], specular reflection detection [9] and inverse rendering techniques [18, 51]. While these methods show promising results, they may still encounter difficulties since their optimizations rely on image reconstruction losses that tightly couple shape and appearance." + }, + { + "type": "page_footnote", + "bbox": [ + 0.218, + 0.811, + 0.692, + 0.84 + ], + "angle": 0, + "content": "* Corresponding author. The source code is available at https://github.com/GCChen97/PISR" + } + ], + [ + { + "type": "page_number", + "bbox": [ + 0.218, + 0.116, + 0.23, + 0.127 + ], + "angle": 0, + "content": "2" + }, + { + "type": "header", + "bbox": [ + 0.272, + 0.115, + 0.369, + 0.127 + ], + "angle": 0, + "content": "G. Chen et al." + }, + { + "type": "image", + "bbox": [ + 0.23, + 0.143, + 0.387, + 0.216 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.39, + 0.144, + 0.573, + 0.217 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.577, + 0.144, + 0.758, + 0.217 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.23, + 0.218, + 0.386, + 0.306 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.39, + 0.225, + 0.573, + 0.307 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.577, + 0.225, + 0.758, + 0.306 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.214, + 0.318, + 0.785, + 0.36 + ], + "angle": 0, + "content": "Fig. 1: Normal maps of reconstructed surfaces. By leveraging polarization information, our method is able to reconstruct textureless and specular surfaces with fine-grained details." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.404, + 0.785, + 0.542 + ], + "angle": 0, + "content": "Polarization, as another property of light, provides direct cues on object shape. The angle of polarization of light from object surfaces is independent of light intensity and mainly determined by the surface shapes. From the perspective of a camera, the angle of polarization of a pixel is equal to the azimuth of the surface normal, i.e., the 2D projection direction of the normal onto the image plane, up to a \\(\\pi/2\\)-ambiguity and a \\(\\pi\\)-ambiguity [5]. This geometric property has been extensively explored in shape from polarization [1, 2, 25, 49], photo-polarimetric stereo [7, 8, 27, 38], polarimetric multi-view 3D reconstruction [5, 54, 55] and dense mapping [34, 46]." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.547, + 0.788, + 0.713 + ], + "angle": 0, + "content": "In this work, we introduce PISR (polarimetric neural implicit surface reconstruction), a novel method that combines the merits of polarimetric 3D reconstruction and neural implicit surface reconstruction for textureless and specular objects, such as ceramics and plastics. Different from recent works [6, 16] that utilize the polarization cues through polarimetric volume rendering, we propose a novel polarimetric loss to refine an object shape directly. Thanks to the integration of the perspective polarimetric constraint [4] in the proposed loss, PISR eliminates artifacts on the reconstructed surfaces, resulting in a \\(30\\%\\) reduction in L1 Chamfer distance. Besides, instead of treating all possible normals equally [54], our loss adaptively adjusts according to the degree of polarization to address specular-reflection-dominant surfaces, such as ceramics with dark color." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.735, + 0.788, + 0.84 + ], + "angle": 0, + "content": "To accelerate the optimization process, PISR uses a multi-resolution hash grid and multi-layer perceptrons (MLP) to represent an object, achieving \\(4 \\sim 30 \\times\\) speed up as shown in Tab. 3. This representation is also crucial for accurately reconstructing highly reflective surfaces, enabling PISR to rectify the shape topology during optimization, a processing step not available for the mesh-based method, PMVIR [54]. However, the discrete nature of a hash grid also reduces the smoothness of the representation and can lead to the reconstruction" + } + ], + [ + { + "type": "header", + "bbox": [ + 0.445, + 0.115, + 0.733, + 0.13 + ], + "angle": 0, + "content": "PISR for Textureless and Specular Objects" + }, + { + "type": "page_number", + "bbox": [ + 0.775, + 0.117, + 0.785, + 0.127 + ], + "angle": 0, + "content": "3" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.147, + 0.784, + 0.177 + ], + "angle": 0, + "content": "results being trapped in local minima. To mitigate the problem, PISR smooths rendered normals for better robustness against poor shape initialization." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.179, + 0.788, + 0.284 + ], + "angle": 0, + "content": "Since there is no publicly available dataset for quantitative evaluation of polarimetric 3D reconstruction, we collect a dataset containing textureless and specular objects with a color polarization camera [36]. Using the dataset, we show that PISR is able to improve the surface reconstruction accuracy over the SOTA methods, with an L1 Chamfer distance of \\(0.5\\mathrm{mm}\\) and an F-score of \\(99.5\\%\\) at \\(1\\mathrm{mm}\\). Shown in Fig. 1 are recovered shapes by different methods and our result shows an accurate shape and fine-grained details." + }, + { + "type": "text", + "bbox": [ + 0.24, + 0.286, + 0.568, + 0.3 + ], + "angle": 0, + "content": "In summary, our contributions are as follows:" + }, + { + "type": "text", + "bbox": [ + 0.226, + 0.315, + 0.784, + 0.359 + ], + "angle": 0, + "content": "- PISR, a novel method that leverages polarization cues and hash-grid-based neural SDF for robust and efficient reconstruction of textureless and specular objects." + }, + { + "type": "text", + "bbox": [ + 0.225, + 0.362, + 0.784, + 0.391 + ], + "angle": 0, + "content": "- A novel polarimetric loss that utilizes an accurate formulation of the polarimetric constraint for higher surface reconstruction accuracy." + }, + { + "type": "text", + "bbox": [ + 0.225, + 0.393, + 0.784, + 0.422 + ], + "angle": 0, + "content": "- A polarization image dataset of real textureless and specular objects with ground-truth meshes for quantitative evaluation." + }, + { + "type": "list", + "bbox": [ + 0.225, + 0.315, + 0.784, + 0.422 + ], + "angle": 0, + "content": null + }, + { + "type": "title", + "bbox": [ + 0.216, + 0.453, + 0.396, + 0.469 + ], + "angle": 0, + "content": "2 Related Works" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.491, + 0.788, + 0.612 + ], + "angle": 0, + "content": "Neural Implicit Surface Reconstruction: The success of neural radiance field reconstruction [23] in novel view synthesis greatly inspires the development of neural implicit surface [28, 30, 48]. Combining with differentiable volume rendering, neural implicit surface reconstruction [29, 42, 47] is able to reconstruct implicit surfaces from dense multi-view images while showing excellent performance in reconstructing fine-grained details. Since then, several works have improved it in reconstruction quality [50], sparse-view performance [19] and optimization speed [43]." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.613, + 0.788, + 0.718 + ], + "angle": 0, + "content": "Yet textureless and specular surfaces are still challenging for neural implicit surface reconstruction. To address the problem, NeRO [18] incorporates inverse rendering to alleviate the shape-radiance ambiguity [52] induced by specular reflection. Ref-NeuS [9] borrows the idea of anomaly detection to decrease the weights of specular pixels during optimization and conditions the radiance on reflection direction in reflective scenes as in Ref-NeRF [39]. UniSDF [40] models diffuse and specular radiances in two MLPs separately." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.72, + 0.788, + 0.84 + ], + "angle": 0, + "content": "Estimating a 3D radiance field and an SDF from 2D images is a highly ill-posed problem. During optimization, these methods may distort shapes to fit the reflection colors, causing severe reconstruction errors, or, as is commonly referred to, the shape-radiance ambiguity [52]. In contrast, polarization images provide linear constraints on surface normal by the angle of polarization. Therefore, to avoid the shape-radiance ambiguity, our method utilizes the geometry cues from polarization to regularize object shapes independently of appearance, keeping the shape from distortions induced by the visually distorted appearance." + } + ], + [ + { + "type": "page_number", + "bbox": [ + 0.218, + 0.116, + 0.23, + 0.127 + ], + "angle": 0, + "content": "4" + }, + { + "type": "header", + "bbox": [ + 0.272, + 0.115, + 0.369, + 0.128 + ], + "angle": 0, + "content": "G. Chen et al." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.146, + 0.788, + 0.313 + ], + "angle": 0, + "content": "Polarimetric 3D Reconstruction: Shape from polarization (SfP) [1,25,31,49] and photo-polarimetric stereo [7,8,27,38] are single-view model-based methods that utilize polarization images for normal and depth estimation. These methods rely on assumptions of either spotlighting or distant lighting to explicitly model the relationship between the degree of polarization and the zenith angle of a surface normal, making them difficult to work under natural lighting conditions. Several deep learning-based methods [2,21,35] leverage data priors to enhance the robustness to more complex indoor lighting conditions. SfPW [15] and DPS-Net [37] learn to estimate scene-level normal maps and depth maps. Yet their generalization is still limited by the relatively small amount of training datasets of polarization images." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.314, + 0.788, + 0.495 + ], + "angle": 0, + "content": "Polarimetric multi-view 3D reconstruction leverages multi-view polarimetric constraints to adapt to natural lighting conditions. The polarimetric constraint is established on the fundamental fact that the azimuth angle of a surface normal is equal to the angle of polarization up to a \\(\\pi/2\\)-ambiguity and a \\(\\pi\\)-ambiguity under natural lighting [5, 54], which is relatively independent of lighting [5]. Miyazaki et al. [24] uses multi-view polarimetric constraints to reconstruct black specular objects. PolMVS [5] proposes to propagate the depths of sparse points and refine the depth maps with this constraint. Inspired by PolMVS, PDMS [46] and PMDM [34] integrate this constraint into SLAM for dense mapping. PMVIR [54] proposes a framework to simultaneously refine coarse meshes and estimate illuminations and surface albedos. PolarPMS [55] combines this constraint with photo consistency to estimate depth maps from multi-view polarization images." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.496, + 0.788, + 0.603 + ], + "angle": 0, + "content": "More closely related to our approach is PMVIR [54]. PMVIR represents shapes using meshes and thereby has limitations in handling incorrect shape topology. In contrast, by leveraging neural SDF, our method is able to rectify incorrect shape topology during the optimization. More importantly, our method utilizes the perspective polarimetric constraint [4] to consider the camera ray directions in the polarimetric loss and thus achieves further performance improvement." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.628, + 0.788, + 0.718 + ], + "angle": 0, + "content": "Polarimetric Inverse Rendering: The development of polarimetric bidirectional reflectance distribution function (pBRDF) [3, 11, 12, 14] has enabled polarimetric inverse rendering. Based on mesh representation, sparse ellipsometry [11] leverages polarimetric inverse rendering to simultaneously estimate object meshes and materials using polarization images captured under flashlight illumination." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.72, + 0.788, + 0.84 + ], + "angle": 0, + "content": "With the emergence of neural fields in 3D reconstruction, a few recent works [6, 16] have integrated polarization into neural implicit surfaces through polarization image rendering. Concurrent work NeISF [16] introduces the incident Stokes field to represent multi-path polarized light and reparameterizes pBRDF using MLPs, achieving highly accurate geometry reconstruction. PANDORA [6] aims to decompose diffuse and specular reflections using polarization as auxiliary cues, relying on the Fresnel equations to calculate the diffuse and specular components of radiance." + } + ], + [ + { + "type": "header", + "bbox": [ + 0.445, + 0.115, + 0.733, + 0.13 + ], + "angle": 0, + "content": "PISR for Textureless and Specular Objects" + }, + { + "type": "page_number", + "bbox": [ + 0.775, + 0.117, + 0.785, + 0.127 + ], + "angle": 0, + "content": "5" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.147, + 0.784, + 0.253 + ], + "angle": 0, + "content": "Although polarimetric inverse rendering excels at accurate material and illumination estimation, it requires proper geometry initialization to decouple shape and appearance. As a result, it may still suffer from the shape-radiance ambiguity, especially when dealing with textureless and specular surfaces. To avoid this problem, our method utilizes the polarimetric constraint to regularize shape directly, independent of appearance. The surface reconstruction results of our method can be used to initialize polarimetric inverse rendering." + }, + { + "type": "title", + "bbox": [ + 0.216, + 0.281, + 0.367, + 0.298 + ], + "angle": 0, + "content": "3 Preliminary" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.317, + 0.784, + 0.378 + ], + "angle": 0, + "content": "Polarimetric Constraint: Assuming four synchronized and aligned images captured at polarizer angles of 0, 45, 90, 135 degrees [36], denoted as \\(\\mathbf{I}_0\\), \\(\\mathbf{I}_{45}\\), \\(\\mathbf{I}_{90}\\) and \\(\\mathbf{I}_{135}\\). For each pixel, the angle of polarization (AoP) \\(\\varphi \\in [-\\pi /2,\\pi /2]\\) and degree of polarization (DoP) \\(\\rho \\in [0,1]\\) are calculated as:" + }, + { + "type": "equation", + "bbox": [ + 0.363, + 0.392, + 0.785, + 0.425 + ], + "angle": 0, + "content": "\\[\n\\varphi = \\frac {1}{2} \\arctan 2 \\left(s _ {2}, s _ {1}\\right), \\quad \\rho = \\frac {\\sqrt {s _ {1} ^ {2} + s _ {2} ^ {2}}}{s _ {0}} \\tag {1}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.438, + 0.784, + 0.498 + ], + "angle": 0, + "content": "where \\(\\mathbf{s} = [s_0, s_1, s_2]^T = \\left[\\frac{1}{2}\\sum I_i, I_0 - I_{90}, I_{45} - I_{135}\\right]^T\\) is the Stokes vector for representing the polarization state of light. Given the estimated normal \\(\\hat{\\mathbf{n}}\\) in the camera frame and the AoP \\(\\varphi\\) of a pixel, the orthographic polarimetric constraint on the azimuth angle of the normal is calculated as:" + }, + { + "type": "equation", + "bbox": [ + 0.376, + 0.514, + 0.785, + 0.529 + ], + "angle": 0, + "content": "\\[\n[ \\sin (\\varphi + \\Delta), \\cos (\\varphi + \\Delta), 0 ] \\cdot \\hat {\\mathbf {n}} = 0, \\tag {2}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.543, + 0.784, + 0.604 + ], + "angle": 0, + "content": "where the disambiguation scalar \\(\\varDelta\\in\\{0,\\pi/2\\}\\) is used for solving the \\(\\pi /2\\) ambiguity [5]. While this constraint is commonly used in previous works [5,34, 38,46,54], there is a perspective polarimetric constraint which is geometrically more accurate than it by considering the perspective effect of the lens [4]:" + }, + { + "type": "equation", + "bbox": [ + 0.27, + 0.618, + 0.785, + 0.634 + ], + "angle": 0, + "content": "\\[\n\\left[ v _ {z} \\sin \\varphi^ {\\prime}, v _ {z} \\cos \\varphi^ {\\prime}, - \\left(v _ {y} \\cos \\varphi^ {\\prime} + v _ {x} \\sin \\varphi^ {\\prime}\\right) \\right] \\cdot \\hat {\\mathbf {n}} = 0, \\quad \\varphi^ {\\prime} = \\varphi + \\Delta \\tag {3}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.646, + 0.73, + 0.664 + ], + "angle": 0, + "content": "where \\(\\mathbf{v} = [v_x, v_y, v_z]^T\\) is a normalized camera ray in the camera frame." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.689, + 0.785, + 0.839 + ], + "angle": 0, + "content": "\\(\\pi /2\\) ambiguity: The \\(\\pi /2\\)-ambiguity arises from the differences between the AoPs of specular reflection and diffuse reflection [5]. According to the Fresnel equations, reflected and transmitted light have distinct energy distributions in the parallel and perpendicular directions to the plane of incident, resulting in a \\(\\pi /2\\) difference between their AoPs, i.e., \\(\\Delta = 0\\) for diffuse reflection and \\(\\Delta = \\pi /2\\) for specular reflection. Under natural lighting conditions, the diffuse reflections and specular reflections are usually mixed, making it difficult to distinguish the dominant component, and thus resulting in the \\(\\pi /2\\)-ambiguity. Therefore, the \\(\\pi /2\\)-ambiguity poses a challenge in terms of applying the polarimetric constraints." + } + ], + [ + { + "type": "page_number", + "bbox": [ + 0.218, + 0.116, + 0.23, + 0.127 + ], + "angle": 0, + "content": "6" + }, + { + "type": "header", + "bbox": [ + 0.272, + 0.115, + 0.369, + 0.127 + ], + "angle": 0, + "content": "G. Chen et al." + }, + { + "type": "image", + "bbox": [ + 0.228, + 0.143, + 0.778, + 0.301 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.214, + 0.313, + 0.785, + 0.356 + ], + "angle": 0, + "content": "Fig. 2: Overview of PISR. Pixels for the optimization are sampled in criss-cross patterns at the middle stage of the optimization. The polarimetric loss \\(\\mathcal{L}_{\\mathrm{pol}}^{\\mathrm{p}}\\) and the normal loss \\(\\mathcal{L}_{\\mathrm{normal}}\\) are used for regularizing the shape independently of appearance." + }, + { + "type": "title", + "bbox": [ + 0.216, + 0.38, + 0.331, + 0.395 + ], + "angle": 0, + "content": "4 Method" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.408, + 0.788, + 0.533 + ], + "angle": 0, + "content": "PISR aims to recover object shapes from multi-view polarization images and camera poses. Fig. 2 shows an overview of our method. Object shape is represented as a hash-grid-based neural SDF and rendered to images for computing losses (Sec. 4.1). During the reconstruction, the coarse shape of an object is estimated with the photometric loss \\(\\mathcal{L}_{\\mathrm{color}}\\). Then the polarimetric loss \\(\\mathcal{L}_{\\mathrm{pol}}^{\\mathrm{p}}\\) (Sec. 4.2) and the normal loss \\(\\mathcal{L}_{\\mathrm{normal}}\\) with criss-cross pattern sampling (Sec. 4.3) are jointly used to rectify the shape from distortions. Finally, the shape is refined with \\(\\mathcal{L}_{\\mathrm{color}}\\) and \\(\\mathcal{L}_{\\mathrm{pol}}^{\\mathrm{p}}\\). The optimization scheme is introduced in Sec. 4.4." + }, + { + "type": "title", + "bbox": [ + 0.214, + 0.547, + 0.638, + 0.563 + ], + "angle": 0, + "content": "4.1 Surface Representation and Image Rendering" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.569, + 0.788, + 0.66 + ], + "angle": 0, + "content": "Since Instant-NGP [26], combining a multi-resolution hash grid and MLPs has been commonly used for faster convergence in neural field reconstruction. Following recent works [17, 41, 43], we represent the shape and appearance of an object as an \\(L\\)-level multi-resolution hash grid \\(\\mathcal{G} = \\{\\mathcal{G}_l\\}_{l=1}^L\\) with two MLPs \\(\\varPhi_s\\) and \\(\\varPhi_{\\mathbf{c}}\\) for decoding SDF values \\(s\\) and color \\(\\mathbf{c}\\) respectively. For each 3D space point \\(\\mathbf{x}\\), the SDF value \\(s\\) and color \\(\\mathbf{c}\\) are computed as" + }, + { + "type": "equation", + "bbox": [ + 0.367, + 0.667, + 0.785, + 0.684 + ], + "angle": 0, + "content": "\\[\n(s, \\mathbf {f}) = \\varPhi_ {s} (\\mathcal {G} (\\mathbf {x})), \\quad \\mathbf {c} = \\varPhi_ {\\mathbf {c}} (s, \\mathbf {f}, \\mathbf {v}, \\mathbf {n}) \\tag {4}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.691, + 0.788, + 0.735 + ], + "angle": 0, + "content": "with the camera ray \\(\\mathbf{v}\\) and surface normal \\(\\mathbf{n}\\). The normal \\(\\mathbf{n}\\) can be computed by normalizing the gradient of SDF w.r.t. to \\(\\mathbf{x}\\) or by finite difference [17]. Besides the foreground object, we use Instant-NGP [26] to represent the background." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.736, + 0.788, + 0.797 + ], + "angle": 0, + "content": "To optimize the shape with images, we use differentiable volume rendering [22] to render images given the surface representation. The pixel color \\(\\hat{\\mathbf{c}}\\) and normal \\(\\hat{\\mathbf{n}}\\) are computed as the weighted sum of \\(M\\) samples \\(\\{\\mathbf{x}_i = \\mathbf{o} + t_i\\mathbf{v}\\}_{i=1}^M\\) on a cast-off pixel ray with the camera position \\(\\mathbf{o}\\) and different depth \\(t_i\\):" + }, + { + "type": "equation", + "bbox": [ + 0.394, + 0.802, + 0.787, + 0.844 + ], + "angle": 0, + "content": "\\[\n\\hat {\\mathbf {c}} = \\sum_ {i = 1} ^ {M} w _ {i} \\mathbf {c} _ {i}, \\quad \\hat {\\mathbf {n}} = \\sum_ {i = 1} ^ {M} w _ {i} \\mathbf {n} _ {i}, \\tag {5}\n\\]" + } + ], + [ + { + "type": "header", + "bbox": [ + 0.445, + 0.115, + 0.733, + 0.13 + ], + "angle": 0, + "content": "PISR for Textureless and Specular Objects" + }, + { + "type": "page_number", + "bbox": [ + 0.775, + 0.117, + 0.785, + 0.127 + ], + "angle": 0, + "content": "7" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.147, + 0.784, + 0.193 + ], + "angle": 0, + "content": "where \\( w_{i} \\) is the blending weight converted from SDF values using the unbiased weight function of NeuS [42], and \\( \\mathbf{c}_i \\) and \\( \\mathbf{n}_i \\) are the color and normal of the sample \\( \\mathbf{x}_i \\)." + }, + { + "type": "title", + "bbox": [ + 0.216, + 0.217, + 0.566, + 0.23 + ], + "angle": 0, + "content": "4.2 Polarimetric and Photometric Losses" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.244, + 0.784, + 0.29 + ], + "angle": 0, + "content": "Polarimetric Loss: Previous works [5,34,38,46,54] leverage polarization cues with the following loss function \\( h^{\\mathrm{o}} \\) that is derived from the orthographic polarimetric constraint Eq. (2):" + }, + { + "type": "equation", + "bbox": [ + 0.235, + 0.302, + 0.785, + 0.33 + ], + "angle": 0, + "content": "\\[\nh ^ {\\mathrm {o}} (\\varphi , \\Delta , \\mathbf {n}) = \\left(\\mathbf {a} ^ {\\mathrm {o}} (\\varphi , \\Delta) \\cdot \\mathbf {n}\\right) ^ {2}, \\mathbf {a} ^ {\\mathrm {o}} (\\varphi , \\Delta) = [ \\sin (\\varphi + \\Delta), \\cos (\\varphi + \\Delta), 0 ], \\tag {6}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.34, + 0.784, + 0.384 + ], + "angle": 0, + "content": "where the coefficient vector \\(\\mathbf{a}^{\\mathrm{o}}\\) is calculated with the AoP \\(\\varphi\\) and the disambiguation scalar \\(\\Delta\\). However, as discussed in [4], Eq. (2) neglects the perspective effect of the lens and may introduce distortions in the reconstruction results." + }, + { + "type": "text", + "bbox": [ + 0.215, + 0.386, + 0.784, + 0.417 + ], + "angle": 0, + "content": "Therefore, we instead adopt the perspective polarimetric constraint Eq. (3) [4] and propose a novel loss function \\( h^{\\mathrm{p}} \\):" + }, + { + "type": "equation", + "bbox": [ + 0.377, + 0.429, + 0.785, + 0.462 + ], + "angle": 0, + "content": "\\[\nh ^ {\\mathrm {p}} (\\varphi , \\Delta , \\mathbf {v}, \\mathbf {n}) = \\left(\\frac {\\mathbf {a} ^ {\\mathrm {p}} (\\varphi , \\Delta , \\mathbf {v}) \\cdot \\mathbf {n}}{\\| \\mathbf {a} ^ {\\mathrm {p}} (\\varphi , \\Delta , \\mathbf {v}) \\|}\\right) ^ {2} \\tag {7}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.215, + 0.472, + 0.377, + 0.485 + ], + "angle": 0, + "content": "with \\(\\mathbf{a}^{\\mathrm{p}}\\) calculated as:" + }, + { + "type": "equation", + "bbox": [ + 0.238, + 0.5, + 0.785, + 0.516 + ], + "angle": 0, + "content": "\\[\n\\mathbf {a} ^ {\\mathrm {p}} (\\varphi , \\Delta , \\mathbf {v}) = \\left[ v _ {z} \\sin \\varphi^ {\\prime}, v _ {z} \\cos \\varphi^ {\\prime}, - \\left(v _ {y} \\cos \\varphi^ {\\prime} + v _ {x} \\sin \\varphi^ {\\prime}\\right) \\right], \\varphi^ {\\prime} = \\varphi + \\Delta . \\tag {8}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.528, + 0.784, + 0.572 + ], + "angle": 0, + "content": "The loss function \\( h^{\\mathrm{p}} \\) additionally considers the camera ray direction \\( \\mathbf{v} \\) to model the perspective effect. In Sec. 5.3, we will show that \\( h^{\\mathrm{p}} \\) effectively improves the reconstruction accuracy." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.574, + 0.784, + 0.619 + ], + "angle": 0, + "content": "Besides, the \\(\\pi/2\\)-ambiguity must be addressed when applying the polarimetric constraint. We combine the merits of PolMVS [5] and PMVIR [54], proposing the segmented loss function \\(f^{\\mathrm{p}}\\) as follows:" + }, + { + "type": "equation", + "bbox": [ + 0.295, + 0.632, + 0.785, + 0.679 + ], + "angle": 0, + "content": "\\[\nf ^ {\\mathrm {p}} (\\varphi , \\mathbf {v}, \\mathbf {n}, \\rho) = \\left\\{ \\begin{array}{l l} h ^ {\\mathrm {p}} (\\varphi , 0, \\mathbf {v}, \\mathbf {n}) \\cdot h ^ {\\mathrm {p}} (\\varphi , \\pi / 2, \\mathbf {v}, \\mathbf {n}) & \\text {i f} \\rho < \\theta , \\\\ h ^ {\\mathrm {p}} (\\varphi , \\pi / 2, \\mathbf {v}, \\mathbf {n}) & \\text {i f} \\rho \\geq \\theta , \\end{array} \\right. \\tag {9}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.689, + 0.785, + 0.779 + ], + "angle": 0, + "content": "where \\(\\rho\\) is the DoP and \\(\\theta\\) is a preset threshold. This design is based on two observations: 1) the reflected light of an object is usually a mixture of the diffuse and specular components, and thus the DoP is relatively low [5], 2) only the DoP of specular reflection of dielectric objects can exceed 0.4 according to Fresnel equations. We use multiplication instead of addition when \\(\\rho < \\theta\\) because only one of the two needs to be minimized." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.781, + 0.784, + 0.808 + ], + "angle": 0, + "content": "Finally, we calculate the polarimetric loss for a batch of sampled pixels \\( S \\) as follows:" + }, + { + "type": "equation", + "bbox": [ + 0.384, + 0.809, + 0.785, + 0.843 + ], + "angle": 0, + "content": "\\[\n\\mathcal {L} _ {\\mathrm {p o l}} ^ {\\mathrm {p}} = \\frac {1}{| S |} \\sum_ {u \\in \\mathcal {S}} f ^ {\\mathrm {p}} \\left(\\varphi_ {u}, \\mathbf {v} _ {u}, \\hat {\\mathbf {n}} _ {u}, \\rho_ {u}\\right). \\tag {10}\n\\]" + } + ], + [ + { + "type": "page_number", + "bbox": [ + 0.218, + 0.116, + 0.23, + 0.127 + ], + "angle": 0, + "content": "8" + }, + { + "type": "header", + "bbox": [ + 0.272, + 0.115, + 0.369, + 0.128 + ], + "angle": 0, + "content": "G. Chen et al." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.147, + 0.784, + 0.207 + ], + "angle": 0, + "content": "Note that the camera ray \\(\\mathbf{v}_u\\) and normal \\(\\hat{\\mathbf{n}}_u\\) are transformed into the camera frame according to the camera pose. In Sec. 5.3, the experiments will demonstrate that calculating the loss \\(\\mathcal{L}_{\\mathrm{pol}}^{\\mathrm{p}}\\) with \\(h^{\\mathrm{p}}\\) instead of \\(h^{\\mathrm{o}}\\) can reduce the L1 Chamfer distance by \\(30\\%\\)." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.211, + 0.784, + 0.256 + ], + "angle": 0, + "content": "Photometric Loss: Since the polarimetric loss alone is insufficient for recovering object shapes due to the \\(\\pi /2\\) ambiguity and image noises, we used the photometric loss for initializing and maintaining the basic shape:" + }, + { + "type": "equation", + "bbox": [ + 0.41, + 0.264, + 0.786, + 0.299 + ], + "angle": 0, + "content": "\\[\n\\mathcal {L} _ {\\text {c o l o r}} = \\frac {1}{| \\mathcal {S} |} \\sum_ {u \\in \\mathcal {S}} \\left| \\hat {\\mathbf {c}} _ {u} - \\mathbf {c} _ {u} \\right|. \\tag {11}\n\\]" + }, + { + "type": "title", + "bbox": [ + 0.216, + 0.317, + 0.452, + 0.334 + ], + "angle": 0, + "content": "4.3 Normal Regularization" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.34, + 0.785, + 0.461 + ], + "angle": 0, + "content": "The discrete nature of the hash grid makes the estimated SDF less smooth than those purely MLP-based, which can introduce reconstruction artifacts such as holes or cracks. Directly smoothing the normals in 3D space as in [17, 53] can lead to a bumpy surface because the sampled normals are solely constrained by the smoothness loss. Inspired by the belief propagation used in traditional MVS methods [33, 45, 56], we smooth the rendered normals so that all the samples are also constrained by both the photometric loss and the polarimetric loss. We therefore utilize the following normal loss:" + }, + { + "type": "equation", + "bbox": [ + 0.345, + 0.469, + 0.786, + 0.506 + ], + "angle": 0, + "content": "\\[\n\\mathcal {L} _ {\\text {n o r m a l}} = - \\frac {1}{| \\mathcal {S} _ {\\mathrm {c}} |} \\sum_ {u \\in \\mathcal {S} _ {\\mathrm {c}}} \\frac {1}{| \\mathcal {N} _ {u} |} \\sum_ {\\hat {\\mathbf {n}} _ {j} \\in \\mathcal {N} _ {u}} \\operatorname {S G} \\left(\\hat {\\mathbf {n}} _ {u}\\right) ^ {T} \\hat {\\mathbf {n}} _ {j} \\tag {12}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.513, + 0.784, + 0.588 + ], + "angle": 0, + "content": "where \\( S_{c} \\) is a set of pixels sampled among pixels of all the images, \\( \\mathcal{N}_u \\) is a set of pixel normals sampled around \\( \\hat{\\mathbf{n}}_u \\) in the image space. The function SG (\"Stop Gradient\") is used to stop propagating the gradients of the losses w.r.t. \\( \\hat{\\mathbf{n}}_u \\) during backpropagation, i.e., fix the values of \\( \\hat{\\mathbf{n}}_u \\) so that it can be propagated to neighbor pixels." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.589, + 0.784, + 0.62 + ], + "angle": 0, + "content": "Now the pixel sample set \\( S \\) of each iteration during the optimization is the union of the set of center pixels \\( S_{c} \\) and the sets of neighbor pixels \\( \\mathcal{N}_u \\):" + }, + { + "type": "equation", + "bbox": [ + 0.432, + 0.629, + 0.786, + 0.66 + ], + "angle": 0, + "content": "\\[\n\\mathcal {S} = \\mathcal {S} _ {c} \\cup \\left\\{\\bigcup_ {u \\in S _ {c}} \\mathcal {N} _ {u} \\right\\} \\tag {13}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.667, + 0.785, + 0.758 + ], + "angle": 0, + "content": "Empirically, larger \\(\\mathcal{S}_c\\) can lead to better results because it helps capture global structures, which are important for optimizing a neural field [44], while larger \\(\\mathcal{N}_u\\) is beneficial to smooth the surface. Therefore, we sample \\(\\mathcal{N}_u\\) in a criss-cross pattern instead of a whole patch around \\(\\hat{\\mathbf{n}}_u\\) to balance \\(|\\mathcal{S}_c|\\) and \\(|\\mathcal{N}_u|\\) with fixed \\(|S|\\). In principle, \\(\\mathcal{N}_u\\) can be sampled in other relatively sparse patterns. Here we use the criss-cross one for its simple implementation in code." + }, + { + "type": "title", + "bbox": [ + 0.216, + 0.778, + 0.447, + 0.794 + ], + "angle": 0, + "content": "4.4 Optimization Strategy" + }, + { + "type": "text", + "bbox": [ + 0.216, + 0.801, + 0.351, + 0.814 + ], + "angle": 0, + "content": "The overall loss is" + }, + { + "type": "equation", + "bbox": [ + 0.337, + 0.826, + 0.786, + 0.843 + ], + "angle": 0, + "content": "\\[\n\\mathcal {L} _ {\\text {a l l}} = \\mathcal {L} _ {\\text {c o l o r}} + \\lambda_ {\\mathrm {p}} \\mathcal {L} _ {\\text {p o l}} ^ {\\mathrm {p}} + \\lambda_ {\\mathrm {n}} \\mathcal {L} _ {\\text {n o r m a l}} + \\lambda_ {\\mathrm {e}} \\mathcal {L} _ {\\text {e i k o n a l}} \\tag {14}\n\\]" + } + ], + [ + { + "type": "header", + "bbox": [ + 0.445, + 0.115, + 0.733, + 0.13 + ], + "angle": 0, + "content": "PISR for Textureless and Specular Objects" + }, + { + "type": "page_number", + "bbox": [ + 0.775, + 0.117, + 0.785, + 0.127 + ], + "angle": 0, + "content": "9" + }, + { + "type": "image", + "bbox": [ + 0.223, + 0.144, + 0.441, + 0.272 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.32, + 0.273, + 0.341, + 0.285 + ], + "angle": 0, + "content": "(a)" + }, + { + "type": "image", + "bbox": [ + 0.452, + 0.144, + 0.551, + 0.272 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.489, + 0.274, + 0.511, + 0.285 + ], + "angle": 0, + "content": "(b)" + }, + { + "type": "image", + "bbox": [ + 0.561, + 0.144, + 0.782, + 0.272 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.66, + 0.274, + 0.679, + 0.285 + ], + "angle": 0, + "content": "(c)" + }, + { + "type": "image_caption", + "bbox": [ + 0.215, + 0.296, + 0.785, + 0.325 + ], + "angle": 0, + "content": "Fig. 3: Dataset and collection. (a) Objects with ground-truth meshes. (b) Objects without ground-truth meshes. (c) Capture setup." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.354, + 0.788, + 0.465 + ], + "angle": 0, + "content": "The extra Eikonal loss \\(\\mathcal{L}_{\\mathrm{eikonal}} = \\frac{1}{M|S|}\\sum_{k=1}^{M|S|}(\\|\\nabla\\hat{s}(\\mathbf{x}_k)\\| - 1)^2\\) is used for encouraging the neural field to approximate a signed distance field [10]. During the optimization, we apply progressive training [17] on the multi-resolution hash grid. To initialize a coarse shape, we set \\(\\lambda_{\\mathrm{p}} = \\lambda_{\\mathrm{n}} = 0\\) and use the photometric and eikonal losses at the early stage. Then we gradually increase \\(\\lambda_{\\mathrm{p}}\\) and \\(\\lambda_{\\mathrm{n}}\\) to rectify the distorted shape. Finally, we gradually decrease \\(\\lambda_{\\mathrm{n}}\\) to 0 to avoid over smoothing the details." + }, + { + "type": "title", + "bbox": [ + 0.216, + 0.491, + 0.376, + 0.508 + ], + "angle": 0, + "content": "5 Experiments" + }, + { + "type": "title", + "bbox": [ + 0.215, + 0.525, + 0.428, + 0.541 + ], + "angle": 0, + "content": "5.1 Experimental Setup" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.553, + 0.788, + 0.673 + ], + "angle": 0, + "content": "Dataset: Since there is no public polarization image dataset of textureless and specular objects with ground-truth meshes, we collect a real-world dataset, including four objects (Black Dragon, Red Dragon, Standing Rabbit, Lying Rabbit, Figure and Car) as shown in Fig. 3. The reasons for capturing real object images, rather than synthesizing using a rendering engine, such as [13], are two-fold: 1) the natural ambient lighting conditions are complex and therefore difficult to synthesize, and 2) the shadows and occlusions of the camera can be realistically generated." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.674, + 0.789, + 0.827 + ], + "angle": 0, + "content": "Each object's polarization images are captured indoors under natural lighting conditions from 40 (60 for Figure for its relatively complex shape) different viewpoints around the object with a color polarization camera [36]. Color images are obtained by first decomposing polarization images into images of four polarizer angles and then averaging and demosaicing. Camera poses are estimated from color images using SuperGlue [32] and COLMAP [33] given pre-calibrated camera intrinsics. We use a checkerboard to help pose estimation. To obtain ground-truth meshes, the objects are coated by 3D scanning spray to eliminate specular reflection and scanned by a structure light-based scanner except for Figure and Car. Please refer to our supplementary material for more details." + } + ], + [ + { + "type": "page_number", + "bbox": [ + 0.218, + 0.116, + 0.236, + 0.127 + ], + "angle": 0, + "content": "10" + }, + { + "type": "header", + "bbox": [ + 0.272, + 0.115, + 0.369, + 0.127 + ], + "angle": 0, + "content": "G. Chen et al." + }, + { + "type": "image_caption", + "bbox": [ + 0.227, + 0.151, + 0.244, + 0.196 + ], + "angle": 0, + "content": "Baaee" + }, + { + "type": "image", + "bbox": [ + 0.25, + 0.147, + 0.321, + 0.202 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.338, + 0.148, + 0.407, + 0.201 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.423, + 0.147, + 0.493, + 0.201 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.509, + 0.148, + 0.578, + 0.2 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.594, + 0.148, + 0.664, + 0.2 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.681, + 0.148, + 0.75, + 0.2 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.755, + 0.159, + 0.785, + 0.384 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.228, + 0.216, + 0.244, + 0.258 + ], + "angle": 0, + "content": "Rrnnnne" + }, + { + "type": "image", + "bbox": [ + 0.251, + 0.208, + 0.32, + 0.259 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.336, + 0.208, + 0.407, + 0.259 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.422, + 0.208, + 0.492, + 0.259 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.507, + 0.208, + 0.577, + 0.26 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.594, + 0.208, + 0.663, + 0.26 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.68, + 0.208, + 0.75, + 0.26 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.228, + 0.277, + 0.244, + 0.316 + ], + "angle": 0, + "content": "Rabitt S." + }, + { + "type": "image", + "bbox": [ + 0.252, + 0.263, + 0.316, + 0.322 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.336, + 0.263, + 0.403, + 0.322 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.423, + 0.263, + 0.488, + 0.323 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.51, + 0.264, + 0.574, + 0.323 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.594, + 0.264, + 0.661, + 0.322 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.68, + 0.264, + 0.746, + 0.322 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.228, + 0.335, + 0.244, + 0.375 + ], + "angle": 0, + "content": "Rabibit L" + }, + { + "type": "image", + "bbox": [ + 0.253, + 0.324, + 0.316, + 0.377 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.338, + 0.324, + 0.403, + 0.377 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.425, + 0.324, + 0.488, + 0.377 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.51, + 0.324, + 0.574, + 0.378 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.508, + 0.38, + 0.572, + 0.391 + ], + "angle": 0, + "content": "PANDORA" + }, + { + "type": "image", + "bbox": [ + 0.595, + 0.324, + 0.66, + 0.378 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.604, + 0.38, + 0.647, + 0.391 + ], + "angle": 0, + "content": "PMVIR" + }, + { + "type": "image", + "bbox": [ + 0.681, + 0.324, + 0.746, + 0.378 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.68, + 0.38, + 0.744, + 0.391 + ], + "angle": 0, + "content": "PISR (ours)" + }, + { + "type": "image_caption", + "bbox": [ + 0.214, + 0.404, + 0.789, + 0.46 + ], + "angle": 0, + "content": "Fig. 4: Signed error maps. Red and blue mean surface swelling and shrinking respectively. Regions with higher color saturation indicate longer distances between the reconstruction result and the ground truth. Errors are truncated to within \\(\\pm 2\\mathrm{mm}\\) for better visualization." + }, + { + "type": "table_caption", + "bbox": [ + 0.342, + 0.481, + 0.662, + 0.496 + ], + "angle": 0, + "content": "Table 1: Summary of Methods in Experiments" + }, + { + "type": "table", + "bbox": [ + 0.263, + 0.507, + 0.741, + 0.622 + ], + "angle": 0, + "content": "
MethodsImageShape Repre.Appearance Repre.
Ref-NeuS [9]RGBNeural SDFNeural radiance
NeRO [18]RGBNeural SDFNeural BRDF
NeuS [42]RGBNeural SDFNeural radiance
PANDORA [6]Pol. RGBNeural SDFNeural BRDF
PMVIR [54]Pol. RGBMeshSpherical harmonics
PISR (Ours)Pol. RGBGrids+Neural SDFNeural radiance
" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.656, + 0.788, + 0.703 + ], + "angle": 0, + "content": "Performance Metrics: Since we focus on geometry reconstruction, we only evaluate the surface reconstruction quality using L1 Chamfer distance and F-score. We remove the background meshes before calculating the metrics." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.719, + 0.789, + 0.843 + ], + "angle": 0, + "content": "Implementation Details: The SDF MLP \\(\\varPhi_s\\) has one hidden layer with a size of 64, and the color MLP \\(\\varPhi_{\\mathbf{c}}\\) has two hidden layers with sizes of 64. The multi-resolution grid contains 16 levels of resolution, and the coarsest and finest resolutions are 32 and 2700, respectively. The hash table size of each level is \\(2^{19}\\). The optimization of the grid starts at the fourth-coarsest resolutions and progressively extends one level every 1000 iterations. Throughout the whole optimization process, the threshold of DoP \\(\\theta\\) is set to 0.3 and \\(\\lambda_{\\mathrm{eikonal}}\\) is set to 0.1. After the first 2.5k iterations, \\(\\lambda_{\\mathrm{pol}}\\), \\(\\lambda_{\\mathrm{normal}}\\) and \\(|\\mathcal{N}_p|\\) are linearly increased from" + } + ], + [ + { + "type": "header", + "bbox": [ + 0.445, + 0.115, + 0.732, + 0.13 + ], + "angle": 0, + "content": "PISR for Textureless and Specular Objects" + }, + { + "type": "page_number", + "bbox": [ + 0.769, + 0.116, + 0.784, + 0.127 + ], + "angle": 0, + "content": "11" + }, + { + "type": "table_caption", + "bbox": [ + 0.214, + 0.145, + 0.788, + 0.215 + ], + "angle": 0, + "content": "Table 2: Reconstruction accuracy in Chamfer distance (CD↓, mm) and F-score (FS↑, %). FS is calculated at a threshold of 1 mm. Note that the values in the table are rounded to one decimal place. The CD of PMVIR and PISR on Rabbit L. are 0.584 mm and 0.578 mm respectively. Bold, underline, and double underline mean the best, second-best and third-best performance, respectively." + }, + { + "type": "table", + "bbox": [ + 0.239, + 0.226, + 0.767, + 0.364 + ], + "angle": 0, + "content": "
MethodsDragon B.Dragon R.Rabbit S.Rabbit L.Avg.
CDFSCDFSCDFSCDFSCDFS
RGBRef-NeuS2.165.32.166.11.284.51.090.51.676.6
NeRO2.459.21.866.31.380.51.385.01.772.8
NeuS1.967.11.286.01.186.80.993.01.383.2
RGBPANDORA2.355.31.864.41.478.31.577.11.868.8
PMVIR1.876.51.090.50.697.50.697.31.090.5
PISR (Ours)0.599.90.699.10.599.70.699.40.599.5
" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.388, + 0.788, + 0.436 + ], + "angle": 0, + "content": "0 to 2.0, 1.0 and 28 in \\(2.5\\mathrm{k}\\) iterations, respectively. Then \\(\\lambda_{\\mathrm{normal}}\\) and \\(|\\mathcal{N}_p|\\) are linearly decreased to 0 in another \\(2.5\\mathrm{k}\\) iterations. The maximum number of iterations is \\(20\\mathrm{k}\\). The final SDF is converted to meshes using Marching Cubes [20]." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.448, + 0.788, + 0.541 + ], + "angle": 0, + "content": "Baselines: We compare our method with five methods, including PMVIR [54], PANDORA [6], NeRO [18], Ref-NeuS [9] and NeuS [42]. Tab. 1 summarizes the type of input images, geometry representation and color representation of all the six methods. Since PMVIR is a refinement method based on an initial shape, we provide the coarse initial meshes reconstructed by PISR-C (see the ablation study in Sec. 5.3 for the details of PISR-C)." + }, + { + "type": "title", + "bbox": [ + 0.215, + 0.56, + 0.327, + 0.574 + ], + "angle": 0, + "content": "5.2 Results" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.584, + 0.788, + 0.781 + ], + "angle": 0, + "content": "Quantitative Results: The Chamfer distances and F-scores of the reconstruction results are reported in Tab. 2. Our method achieves the best reconstruction quality on all four objects shown in Fig. 3a and significantly outperforms the other five methods. The average Chamfer distance and F-score of our method are \\(0.5 \\, \\mathrm{mm}\\) and \\(99.5\\%\\), respectively. As the second-best method in the comparison, PMVIR achieves twice the Chamfer distance and a \\(9\\%\\) lower F-score compared to ours. Both PMVIR and our method utilize polarimetric constraints, and the results show that the geometry cue of polarization is effective for shape estimation. Although PANDORA also uses polarization images as input, it implicitly utilizes polarization cues through polarimetric rendering, i.e., optimizing shapes using an image reconstruction loss, which could suffer from the shape-radiance ambiguity as RGB-based methods. Among RGB-based methods, NeuS achieves better results than Ref-NeuS and NeRO which are designed for specular objects." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.795, + 0.788, + 0.842 + ], + "angle": 0, + "content": "We visualize the error distributions in Fig. 4. For methods that rely solely on image reconstruction losses (Ref-NeuS, NeRO, NeuS and PANDORA), their reconstruction errors in non-convex regions are notably higher compared to" + } + ], + [ + { + "type": "page_number", + "bbox": [ + 0.218, + 0.116, + 0.236, + 0.127 + ], + "angle": 0, + "content": "12" + }, + { + "type": "header", + "bbox": [ + 0.272, + 0.115, + 0.369, + 0.127 + ], + "angle": 0, + "content": "G. Chen et al." + }, + { + "type": "image", + "bbox": [ + 0.219, + 0.143, + 0.406, + 0.276 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.271, + 0.277, + 0.341, + 0.286 + ], + "angle": 0, + "content": "Initialization" + }, + { + "type": "image", + "bbox": [ + 0.411, + 0.143, + 0.597, + 0.275 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.48, + 0.276, + 0.525, + 0.286 + ], + "angle": 0, + "content": "PMVIR" + }, + { + "type": "image", + "bbox": [ + 0.602, + 0.143, + 0.788, + 0.275 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.674, + 0.276, + 0.705, + 0.286 + ], + "angle": 0, + "content": "PISR" + }, + { + "type": "image_caption", + "bbox": [ + 0.215, + 0.296, + 0.785, + 0.325 + ], + "angle": 0, + "content": "Fig. 5: Normal maps of Dragon B. before and after optimizations. Using neural SDFs to represent shapes allows changes in the shape topology during the optimization." + }, + { + "type": "table_caption", + "bbox": [ + 0.308, + 0.338, + 0.694, + 0.352 + ], + "angle": 0, + "content": "Table 3: Average optimization time of different methods." + }, + { + "type": "table", + "bbox": [ + 0.248, + 0.365, + 0.756, + 0.424 + ], + "angle": 0, + "content": "
MethodsRef-NeuSNeRONeuSPANDORAPMVIRPISR (Ours)
Iterations200k200k200k100k420k
Time (h)20.020.510.815.02.00.5
" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.45, + 0.788, + 0.631 + ], + "angle": 0, + "content": "PMVIR and our method. While the inherent continuity of MLPs ensures the watertightness of their reconstructed shapes, their accuracy cannot be guaranteed due to the shape-radiance ambiguity introduced by specular reflections and the lack of textures. Additionally, for NeRO and PANDORA, the lighting shadows cast by the camera further exacerbate the difficulty of estimating surface material parameters. The real HDR ambient lighting also poses challenges for PANDORA and Ref-NeuS to accurately estimate object shapes. In contrast, PMVIR and our method can recover non-convex regions using the polarimetric constraints. Despite the polarization cues, representing the shapes as neural SDF helps our method better reconstruct regions that are poorly initialized. As shown in Fig. 5, neural SDF allows continuous changes of the shape topology during optimization, which is crucial for evolving surfaces from holes." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.646, + 0.788, + 0.768 + ], + "angle": 0, + "content": "**Optimization Time:** We report the average optimization time in Tab. 3. The experiments are conducted on a machine equipped with an Intel Core i9-12900K and an NVIDIA RTX 3090 GPU. We set the maximum number of iterations to \\(200\\mathrm{k}\\). If the shape converges, we stop the optimization earlier. Our method represents shape as a multi-resolution hash grid and two shallow MLPs, significantly reducing the time complexity for rendering a pixel. Therefore, our method converges much faster, achieving up to \\(4 \\sim 41 \\times\\) convergence speed while maintaining the highest reconstruction accuracy." + }, + { + "type": "title", + "bbox": [ + 0.215, + 0.787, + 0.389, + 0.803 + ], + "angle": 0, + "content": "5.3 Ablation Study" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.81, + 0.787, + 0.84 + ], + "angle": 0, + "content": "We evaluate the effectiveness of the polarimetric loss and normal regularization, as well as the influence of the perspective effect of the lens. We compare PISR(-" + } + ], + [ + { + "type": "header", + "bbox": [ + 0.445, + 0.115, + 0.732, + 0.13 + ], + "angle": 0, + "content": "PISR for Textureless and Specular Objects" + }, + { + "type": "page_number", + "bbox": [ + 0.769, + 0.116, + 0.785, + 0.127 + ], + "angle": 0, + "content": "13" + }, + { + "type": "table_caption", + "bbox": [ + 0.217, + 0.145, + 0.788, + 0.215 + ], + "angle": 0, + "content": "Table 4: Results of the ablation study on different combinations of losses. The L1 CD↓ (mm) and FSs↑ (%) at 1.0 mm and 0.5 are the averages over four objects in Fig.3a. The ranks are the averages over the three metrics. \\(\\mathcal{L}_{\\mathrm{pol}}^{\\mathrm{o}}\\) differs from \\(\\mathcal{L}_{\\mathrm{pol}}^{\\mathrm{p}}\\) by its use of the orthographic constraint Eq. 2 instead of the perspective one Eq. 3. Note that \\(\\mathcal{L}_{\\mathrm{eikonal}}\\) is used in all experiments and is omitted it in this table." + }, + { + "type": "table", + "bbox": [ + 0.252, + 0.226, + 0.753, + 0.363 + ], + "angle": 0, + "content": "
VariantsUsed LossesAvg.Rank
LcolorLpoiLpolLnormalCDFS(1.0)FS(0.5)
PISR-C✓1.1088.1070.306
PISR-N✓✓1.0088.2071.105
PISR-O✓✓0.8096.6075.543
PISR-ON✓✓✓0.7796.1075.504
PISR-P✓✓0.5199.7692.791
PISR(-PN)✓✓✓0.5499.5491.752
" + }, + { + "type": "text", + "bbox": [ + 0.217, + 0.391, + 0.787, + 0.451 + ], + "angle": 0, + "content": "PN) with its five variants, referred to as \\( PISR-P \\), \\( PISR-ON \\), \\( PISR-O \\), \\( PISR-N \\) and \\( PISR-C \\), based on different combinations of losses as shown in Tab. 4. The quantitative experiment is conducted on objects shown in Fig. 3a and the qualitative experiment is conducted on objects shown in Fig. 3b." + }, + { + "type": "text", + "bbox": [ + 0.217, + 0.452, + 0.787, + 0.542 + ], + "angle": 0, + "content": "As shown in Tab. 4, the top four performing variants are all incorporated with the polarimetric losses, and fine-grained details can be recovered without complicated appearance modeling. As for the normal regularization, although the quantitative comparison between PISR-PN and PISR-P shows that it slightly reduces the accuracy, the qualitative results in Fig. 6 show its effectiveness for recovering surfaces from severe distortions." + }, + { + "type": "text", + "bbox": [ + 0.217, + 0.542, + 0.787, + 0.724 + ], + "angle": 0, + "content": "Among the top four performing variants, PISR-PN and PISR-P achieve higher accuracy by considering the perspective effect of the lens using Eq. (3) instead of Eq. (2). As the example error maps of Rabbit L. shown in Fig. 7, the perspective effect leads to surface swelling and shrinking, causing severe cracks on the surface. This phenomenon does not appear in the results of PMVIR and PANDORA. We speculate the reason to be that the hash-grid-based neural SDF we use is too flexible in topology compared to the mesh-based one used in PMVIR, and lacks the smoothness of purely MLP-based one used in PANDORA. Besides, the results in Tab. 4 show that PISR-PN outperforms PISR-ON by a \\(30\\%\\) reduction in L1 Chamfer distance and a \\(17\\%\\) improvement in F-score at \\(0.5 \\mathrm{~mm}\\), which suggests the necessity of considering the perspective effect to improve the accuracy of polarimetric 3D reconstruction." + }, + { + "type": "title", + "bbox": [ + 0.218, + 0.747, + 0.358, + 0.763 + ], + "angle": 0, + "content": "6 Conclusion" + }, + { + "type": "text", + "bbox": [ + 0.217, + 0.78, + 0.787, + 0.84 + ], + "angle": 0, + "content": "We have proposed PISR, an accurate and efficient polarimetric neural implicit surface reconstruction method for textureless and specular objects. The key idea of PISR is to explicitly regularize the SDF using the polarimetric constraints to avoid the shape-radiance ambiguity. More importantly, PISR achieves high accu" + } + ], + [ + { + "type": "page_number", + "bbox": [ + 0.218, + 0.116, + 0.236, + 0.127 + ], + "angle": 0, + "content": "14" + }, + { + "type": "header", + "bbox": [ + 0.272, + 0.115, + 0.369, + 0.127 + ], + "angle": 0, + "content": "G. Chen et al." + }, + { + "type": "image", + "bbox": [ + 0.216, + 0.143, + 0.785, + 0.374 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.214, + 0.383, + 0.788, + 0.412 + ], + "angle": 0, + "content": "Fig. 6: Normal maps of Figure and Car. The polarimetric loss and normal loss jointly recover the shape from distortions induced by strong reflections." + }, + { + "type": "image", + "bbox": [ + 0.277, + 0.445, + 0.716, + 0.576 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.214, + 0.585, + 0.788, + 0.629 + ], + "angle": 0, + "content": "Fig. 7: Signed error maps of Rabbit L. PISR-PN leverages the perspective polarimetric constraint to eliminate surface cracks encountered by PISR-ON. Additional error maps can be found in our supplementary material." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.679, + 0.788, + 0.741 + ], + "angle": 0, + "content": "racy by leveraging the perspective polarimetric constraint, eliminating the surface cracks caused by the traditional orthographic one. Besides, PISR regularizes surface normals in image space for robustness and integrates the multi-resolution hash grid for efficiency." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.75, + 0.788, + 0.841 + ], + "angle": 0, + "content": "Limitations: Despite the high reconstruction accuracy in the experiments, PISR's performance degenerates with noisy AoP maps, limiting its application in reconstructing rough, well-textured or mirror-like objects. Recent works [16,18] have shown solutions to mitigate this problem. In addition, PISR also requires dense input views, adding shape constraints derived from normal and depth estimation models would be a promising solution." + } + ], + [ + { + "type": "header", + "bbox": [ + 0.445, + 0.115, + 0.732, + 0.129 + ], + "angle": 0, + "content": "PISR for Textureless and Specular Objects" + }, + { + "type": "page_number", + "bbox": [ + 0.769, + 0.117, + 0.785, + 0.127 + ], + "angle": 0, + "content": "15" + }, + { + "type": "title", + "bbox": [ + 0.218, + 0.145, + 0.323, + 0.16 + ], + "angle": 0, + "content": "References" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.176, + 0.785, + 0.204 + ], + "angle": 0, + "content": "1. Atkinson, G.A., Hancock, E.R.: Recovery of surface orientation from diffuse polarization. IEEE transactions on image processing 15(6), 1653-1664 (2006)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.205, + 0.785, + 0.245 + ], + "angle": 0, + "content": "2. Ba, Y., Gilbert, A., Wang, F., Yang, J., Chen, R., Wang, Y., Yan, L., Shi, B., Kadambi, A.: Deep shape from polarization. In: European Conference on Computer Vision. pp. 554-571. Springer (2020)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.247, + 0.785, + 0.286 + ], + "angle": 0, + "content": "3. Baek, S.H., Zeltner, T., Ku, H., Hwang, I., Tong, X., Jakob, W., Kim, M.H.: Image-based acquisition and modeling of polarimetric reflectance. ACM Trans. Graph. 39(4), 139 (2020)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.288, + 0.785, + 0.328 + ], + "angle": 0, + "content": "4. Chen, G., He, L., Guan, Y., Zhang, H.: Perspective phase angle model for polarimetric 3d reconstruction. In: European Conference on Computer Vision. pp. 398-414. Springer (2022)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.33, + 0.785, + 0.37 + ], + "angle": 0, + "content": "5. Cui, Z., Gu, J., Shi, B., Tan, P., Kautz, J.: Polarimetric multi-view stereo. In: Proceedings of the IEEE conference on computer vision and pattern recognition. pp. 1558-1567 (2017)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.371, + 0.785, + 0.411 + ], + "angle": 0, + "content": "6. Dave, A., Zhao, Y., Veeraraghavan, A.: Pandora: Polarization-aided neural decomposition of radiance. In: European Conference on Computer Vision. pp. 538-556. Springer (2022)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.412, + 0.785, + 0.439 + ], + "angle": 0, + "content": "7. Ding, Y., Ji, Y., Ye, J.: Polar-photometric stereo under natural illumination. In: 2022 International Conference on 3D Vision (3DV). pp. 690–699. IEEE (2022)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.44, + 0.785, + 0.481 + ], + "angle": 0, + "content": "8. Ding, Y., Ji, Y., Zhou, M., Kang, S.B., Ye, J.: Polarimetric helmholtz stereopsis. In: Proceedings of the IEEE/CVF International Conference on Computer Vision. pp. 5037-5046 (2021)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.482, + 0.785, + 0.522 + ], + "angle": 0, + "content": "9. Ge, W., Hu, T., Zhao, H., Liu, S., Chen, Y.C.: Ref-neus: Ambiguity-reduced neural implicit surface learning for multi-view reconstruction with reflection. arXiv preprint arXiv:2303.10840 (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.523, + 0.785, + 0.563 + ], + "angle": 0, + "content": "0. Gropp, A., Yariv, L., Haim, N., Atzmon, M., Lipman, Y.: Implicit geometric regularization for learning shapes. In: Proceedings of the 37th International Conference on Machine Learning. pp. 3789-3799 (2020)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.565, + 0.785, + 0.605 + ], + "angle": 0, + "content": "1. Hwang, I., Jeon, D.S., Munoz, A., Gutierrez, D., Tong, X., Kim, M.H.: Sparse ellipsometry: portable acquisition of polarimetric svbrdf and shape with unstructured flash photography. ACM Transactions on Graphics (TOG) 41(4), 1-14 (2022)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.606, + 0.785, + 0.66 + ], + "angle": 0, + "content": "2. Ichikawa, T., Fukao, Y., Nobuhara, S., Nishino, K.: Fresnel microfacet brdf: Unification of polari-radiometric surface-body reflection. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 16489-16497 (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.661, + 0.785, + 0.702 + ], + "angle": 0, + "content": "3. Jakob, W., Speierer, S., Roussel, N., Nimier-David, M., Vicini, D., Zeltner, T., Nicolet, B., Crespo, M., Leroy, V., Zhang, Z.: Mitsuba 3 renderer (2022), https://mitsuba-renderer.org" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.703, + 0.785, + 0.757 + ], + "angle": 0, + "content": "4. Kondo, Y., Ono, T., Sun, L., Hirasawa, Y., Murayama, J.: Accurate polarimetric brdf for real polarization scene rendering. In: Computer Vision-ECCV 2020: 16th European Conference, Glasgow, UK, August 23-28, 2020, Proceedings, Part XIX 16. pp. 220-236. Springer (2020)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.758, + 0.785, + 0.799 + ], + "angle": 0, + "content": "5. Lei, C., Qi, C., Xie, J., Fan, N., Koltun, V., Chen, Q.: Shape from polarization for complex scenes in the wild. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR). pp. 12632-12641 (June 2022)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.8, + 0.785, + 0.84 + ], + "angle": 0, + "content": "6. Li, C., Ono, T., Uemori, T., Mihara, H., Gatto, A., Nagahara, H., Moriuchi, Y.: Neisf: Neural incident stokes field for geometry and material estimation. arXiv preprint arXiv:2311.13187 (2023)" + }, + { + "type": "list", + "bbox": [ + 0.226, + 0.176, + 0.785, + 0.84 + ], + "angle": 0, + "content": null + } + ], + [ + { + "type": "page_number", + "bbox": [ + 0.218, + 0.116, + 0.236, + 0.127 + ], + "angle": 0, + "content": "16" + }, + { + "type": "header", + "bbox": [ + 0.272, + 0.115, + 0.368, + 0.127 + ], + "angle": 0, + "content": "G. Chen et al." + }, + { + "type": "ref_text", + "bbox": [ + 0.218, + 0.147, + 0.787, + 0.203 + ], + "angle": 0, + "content": "17. Li, Z., Müller, T., Evans, A., Taylor, R.H., Unberath, M., Liu, M.Y., Lin, C.H.: Neuralangelo: High-fidelity neural surface reconstruction. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 8456-8465 (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.218, + 0.203, + 0.787, + 0.244 + ], + "angle": 0, + "content": "18. Liu, Y., Wang, P., Lin, C., Long, X., Wang, J., Liu, L., Komura, T., Wang, W.: Nano: Neural geometry and brdf reconstruction of reflective objects from multiview images. In: SIGGRAPH (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.218, + 0.244, + 0.787, + 0.285 + ], + "angle": 0, + "content": "19. Long, X., Lin, C., Wang, P., Komura, T., Wang, W.: Sparseneus: Fast generalizable neural surface reconstruction from sparse views. In: European Conference on Computer Vision. pp. 210-227. Springer (2022)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.285, + 0.787, + 0.311 + ], + "angle": 0, + "content": "20. Lorensen, W.E., Cline, H.E.: Marching cubes: A high resolution 3d surface construction algorithm. ACM SIGGRAPH Computer Graphics 21(4), 163-169 (1987)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.311, + 0.787, + 0.352 + ], + "angle": 0, + "content": "21. Lyu, Y., Zhao, L., Li, S., Shi, B.: Shape from polarization with distant lighting estimation. IEEE Transactions on Pattern Analysis and Machine Intelligence 45(11), 13991-14004 (2023). https://doi.org/10.1109/TPAMI.2023.3298376" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.352, + 0.787, + 0.406 + ], + "angle": 0, + "content": "22. Martin-Brualla, R., Radwan, N., Sajjadi, M.S., Barron, J.T., Dosovitskiy, A., Duckworth, D.: Nerf in the wild: Neural radiance fields for unconstrained photo collections. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 7210–7219 (2021)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.406, + 0.787, + 0.447 + ], + "angle": 0, + "content": "23. Mildenhall, B., Srinivasan, P.P., Tancik, M., Barron, J.T., Ramamoorthi, R., Ng, R.: Nerf: Representing scenes as neural radiance fields for view synthesis. Communications of the ACM 65(1), 99-106 (2021)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.447, + 0.787, + 0.502 + ], + "angle": 0, + "content": "24. Miyazaki, D., Shigetomi, T., Baba, M., Furukawa, R., Hiura, S., Asada, N.: Polarization-based surface normal estimation of black specular objects from multiple viewpoints. In: 2012 Second International Conference on 3D Imaging, Modeling, Processing, Visualization & Transmission. pp. 104-111. IEEE (2012)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.502, + 0.787, + 0.543 + ], + "angle": 0, + "content": "25. Miyazaki, D., Tan, R.T., Hara, K., Ikeuchi, K.: Polarization-based inverse rendering from a single view. In: Computer Vision, IEEE International Conference on. vol. 3, pp. 982-982. IEEE Computer Society (2003)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.543, + 0.787, + 0.582 + ], + "angle": 0, + "content": "26. Müller, T., Evans, A., Schied, C., Keller, A.: Instant neural graphics primitives with a multiresolution hash encoding. ACM Transactions on Graphics (ToG) 41(4), 1-15 (2022)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.582, + 0.787, + 0.624 + ], + "angle": 0, + "content": "27. Ngo Thanh, T., Nagahara, H., Taniguchi, R.i.: Shape and light directions from shading and polarization. In: Proceedings of the IEEE conference on computer vision and pattern recognition. pp. 2310-2318. IEEE (2015)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.624, + 0.787, + 0.664 + ], + "angle": 0, + "content": "28. Niemeyer, M., Mescheder, L., Oechsle, M., Geiger, A.: Differentiable volumetric rendering: Learning implicit 3d representations without 3d supervision. In: Proc. IEEE Conf. on Computer Vision and Pattern Recognition (CVPR) (2020)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.664, + 0.787, + 0.704 + ], + "angle": 0, + "content": "29. Oechsle, M., Peng, S., Geiger, A.: Unisurf: Unifying neural implicit surfaces and radiance fields for multi-view reconstruction. In: International Conference on Computer Vision (ICCV) (2021)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.704, + 0.787, + 0.759 + ], + "angle": 0, + "content": "30. Park, J.J., Florence, P., Straub, J., Newcombe, R., Lovegrove, S.: Deepsdf: Learning continuous signed distance functions for shape representation. In: Proceedings of the IEEE/CVF conference on computer vision and pattern recognition. pp. 165-174 (2019)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.759, + 0.787, + 0.813 + ], + "angle": 0, + "content": "31. Rahmann, S., Canterakis, N.: Reconstruction of specular surfaces using polarization imaging. In: Proceedings of the 2001 IEEE Computer Society Conference on Computer Vision and Pattern Recognition. CVPR 2001. vol. 1, pp. I-I. IEEE (2001)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.813, + 0.787, + 0.841 + ], + "angle": 0, + "content": "32. Sarlin, P.E., DeTone, D., Malisiewicz, T., Rabinovich, A.: SuperGlue: Learning feature matching with graph neural networks. In: CVPR (2020)" + }, + { + "type": "list", + "bbox": [ + 0.217, + 0.147, + 0.787, + 0.841 + ], + "angle": 0, + "content": null + } + ], + [ + { + "type": "header", + "bbox": [ + 0.445, + 0.115, + 0.732, + 0.129 + ], + "angle": 0, + "content": "PISR for Textureless and Specular Objects" + }, + { + "type": "page_number", + "bbox": [ + 0.769, + 0.117, + 0.785, + 0.127 + ], + "angle": 0, + "content": "17" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.148, + 0.785, + 0.175 + ], + "angle": 0, + "content": "33. Schonberger, J.L., Frahm, J.M.: Structure-from-motion revisited. In: Conference on Computer Vision and Pattern Recognition (CVPR) (2016)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.177, + 0.785, + 0.217 + ], + "angle": 0, + "content": "34. Shakeri, M., Loo, Y., Zhang, H.: Polarimetric monocular dense mapping using relative deep depth prior [j]. 2021. Shakeri M Loo Y Zhang H Polarimetric Monocular Dense Mapping Using Relative Deep Depth Prior (2021)" + }, + { + "type": "ref_text", + "bbox": [ + 0.218, + 0.218, + 0.785, + 0.259 + ], + "angle": 0, + "content": "35. Shao, M., Xia, C., Yang, Z., Huang, J., Wang, X.: Transparent shape from a single view polarization image. In: Proceedings of the IEEE/CVF International Conference on Computer Vision (ICCV). pp. 9277-9286 (October 2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.26, + 0.785, + 0.3 + ], + "angle": 0, + "content": "36. Sony: Polarization image sensor with four-directional on-chip polarizer and global shutter function, https://www.sony-semicon.co.jp/e/products/IS/industry/product/polarization.html, accessed 2024-02-04" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.301, + 0.785, + 0.342 + ], + "angle": 0, + "content": "37. Tian, C., Pan, W., Wang, Z., Mao, M., Zhang, G., Bao, H., Tan, P., Cui, Z.: Dpsnet: Deep polarimetric stereo depth estimation. In: The IEEE/CVF International Conference on Computer Vision (ICCV) (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.343, + 0.785, + 0.383 + ], + "angle": 0, + "content": "38. Tozza, S., Zhu, D., Smith, W., Ramamoorthi, R., Hancock, E.: Uncalibrated, two source photo-polarimetric stereo. IEEE Transactions on Pattern Analysis and Machine Intelligence (2021)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.384, + 0.785, + 0.438 + ], + "angle": 0, + "content": "39. Verbin, D., Hedman, P., Mildenhall, B., Zickler, T., Barron, J.T., Srinivasan, P.P.: Ref-nerf: Structured view-dependent appearance for neural radiance fields. In: 2022 IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR). pp. 5481-5490. IEEE (2022)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.439, + 0.785, + 0.481 + ], + "angle": 0, + "content": "40. Wang, F., Rakotosaona, M.J., Niemeyer, M., Szeliski, R., Pollefeys, M., Tombari, F.: Unisdf: Unifying neural representations for high-fidelity 3d reconstruction of complex scenes with reflections. In: arXiv (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.481, + 0.785, + 0.522 + ], + "angle": 0, + "content": "41. Wang, H., Wang, J., Agapito, L.: Co-slam: Joint coordinate and sparse parametric encodings for neural real-time slam. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 13293-13302 (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.523, + 0.785, + 0.563 + ], + "angle": 0, + "content": "42. Wang, P., Liu, L., Liu, Y., Theobalt, C., Komura, T., Wang, W.: Neus: Learning neural implicit surfaces by volume rendering for multi-view reconstruction. Advances in Neural Information Processing Systems 34, 27171-27183 (2021)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.565, + 0.785, + 0.618 + ], + "angle": 0, + "content": "43. Wang, Y., Han, Q., Habermann, M., Daniilidis, K., Theobalt, C., Liu, L.: Neus2: Fast learning of neural implicit surfaces for multi-view reconstruction. In: Proceedings of the IEEE/CVF International Conference on Computer Vision. pp. 3295-3306 (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.619, + 0.785, + 0.674 + ], + "angle": 0, + "content": "44. Xie, Z., Yang, X., Yang, Y., Sun, Q., Jiang, Y., Wang, H., Cai, Y., Sun, M.: S3im: Stochastic structural similarity and its unreasonable effectiveness for neural fields. In: Proceedings of the IEEE/CVF International Conference on Computer Vision. pp. 18024-18034 (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.675, + 0.785, + 0.715 + ], + "angle": 0, + "content": "45. Xu, Q., Tao, W.: Multi-scale geometric consistency guided multi-view stereo. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR) (June 2019)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.716, + 0.785, + 0.757 + ], + "angle": 0, + "content": "46. Yang, L., Tan, F., Li, A., Cui, Z., Furukawa, Y., Tan, P.: Polarimetric dense monocular slam. In: Proceedings of the IEEE conference on computer vision and pattern recognition. pp. 3857-3866 (2018)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.758, + 0.785, + 0.799 + ], + "angle": 0, + "content": "47. Yariv, L., Gu, J., Kasten, Y., Lipman, Y.: Volume rendering of neural implicit surfaces. In: Thirty-Fifth Conference on Neural Information Processing Systems (2021)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.799, + 0.785, + 0.84 + ], + "angle": 0, + "content": "48. Yariv, L., Kasten, Y., Moran, D., Galun, M., Atzmon, M., Ronen, B., Lipman, Y.: Multiview neural surface reconstruction by disentangling geometry and appearance. Advances in Neural Information Processing Systems 33 (2020)" + }, + { + "type": "list", + "bbox": [ + 0.217, + 0.148, + 0.785, + 0.84 + ], + "angle": 0, + "content": null + } + ], + [ + { + "type": "page_number", + "bbox": [ + 0.218, + 0.116, + 0.236, + 0.127 + ], + "angle": 0, + "content": "18" + }, + { + "type": "header", + "bbox": [ + 0.272, + 0.115, + 0.369, + 0.128 + ], + "angle": 0, + "content": "G. Chen et al." + }, + { + "type": "ref_text", + "bbox": [ + 0.218, + 0.147, + 0.787, + 0.189 + ], + "angle": 0, + "content": "49. Yu, Y., Zhu, D., Smith, W.A.: Shape-from-polarisation: a nonlinear least squares approach. In: Proceedings of the IEEE International Conference on Computer Vision Workshops. pp. 2969-2976 (2017)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.19, + 0.787, + 0.232 + ], + "angle": 0, + "content": "50. Yu, Z., Peng, S., Niemeyer, M., Sattler, T., Geiger, A.: Monosdf: Exploring monocular geometric cues for neural implicit surface reconstruction. Advances in neural information processing systems 35, 25018-25032 (2022)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.232, + 0.787, + 0.286 + ], + "angle": 0, + "content": "51. Zhang, K., Luan, F., Wang, Q., Bala, K., Snavely, N.: PhySG: Inverse rendering with spherical gaussians for physics-based material editing and relighting. In: The IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR) (2021)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.287, + 0.787, + 0.314 + ], + "angle": 0, + "content": "52. Zhang, K., Riegler, G., Snavely, N., Koltun, V.: Nerf++: Analyzing and improving neural radiance fields. arXiv:2010.07492 (2020)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.315, + 0.787, + 0.357 + ], + "angle": 0, + "content": "53. Zhang, X., Srinivasan, P.P., Deng, B., Debevec, P., Freeman, W.T., Barron, J.T.: Nerfactor: Neural factorization of shape and reflectance under an unknown illumination. ACM Transactions on Graphics (ToG) 40(6), 1-18 (2021)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.357, + 0.787, + 0.384 + ], + "angle": 0, + "content": "54. Zhao, J., Monno, Y., Okutomi, M.: Polarimetric multi-view inverse rendering. IEEE Transactions on Pattern Analysis and Machine Intelligence (2022)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.384, + 0.787, + 0.425 + ], + "angle": 0, + "content": "55. Zhao, J., Oishi, J., Monno, Y., Okutomi, M.: Polarimetric patchmatch multi-view stereo. In: Proceedings of the IEEE/CVF Winter Conference on Applications of Computer Vision. pp. 3476-3484 (2024)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.425, + 0.787, + 0.468 + ], + "angle": 0, + "content": "56. Zheng, E., Dunn, E., Jojic, V., Frahm, J.M.: Patchmatch based joint view selection and depthmap estimation. In: Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition (CVPR) (June 2014)" + }, + { + "type": "list", + "bbox": [ + 0.217, + 0.147, + 0.787, + 0.468 + ], + "angle": 0, + "content": null + } + ] +] \ No newline at end of file diff --git a/2024/PISR_ Polarimetric Neural Implicit Surface Reconstruction for Textureless and Specular Objects/949f096c-d4bf-4988-bb1a-9febc3a46def_origin.pdf b/2024/PISR_ Polarimetric Neural Implicit Surface Reconstruction for Textureless and Specular Objects/949f096c-d4bf-4988-bb1a-9febc3a46def_origin.pdf new file mode 100644 index 0000000000000000000000000000000000000000..fc297de06948371b272e5dea8730e01ca8f12126 --- /dev/null +++ b/2024/PISR_ Polarimetric Neural Implicit Surface Reconstruction for Textureless and Specular Objects/949f096c-d4bf-4988-bb1a-9febc3a46def_origin.pdf @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:5769a45cb5b3969c62c04190a86765f16533cb361f696fa928454fe66003d9a7 +size 9545137 diff --git a/2024/PISR_ Polarimetric Neural Implicit Surface Reconstruction for Textureless and Specular Objects/full.md b/2024/PISR_ Polarimetric Neural Implicit Surface Reconstruction for Textureless and Specular Objects/full.md new file mode 100644 index 0000000000000000000000000000000000000000..f8b4892b7ef321afd6749994abe4ac4d12f0b49a --- /dev/null +++ b/2024/PISR_ Polarimetric Neural Implicit Surface Reconstruction for Textureless and Specular Objects/full.md @@ -0,0 +1,391 @@ +# PISR: Polarimetric Neural Implicit Surface Reconstruction for Textureless and Specular Objects + +Guangcheng Chen $^{1,2}$ , Yicheng He $^{1,2}$ , Li He $^{1,2}$ , and Hong Zhang $^{\star 1,2}$ + +$^{1}$ Southern University of Science and Technology, Shenzhen, China $^{2}$ Shenzhen Key Laboratory of Robotics and Computer Vision {chengc2023,12332136}@mail.sustech.edu.cn, {hel, hzhang}@sustech.edu.cn + +Abstract. Neural implicit surface reconstruction has achieved remarkable progress recently. Despite resorting to complex radiance modeling, state-of-the-art methods still struggle with textureless and specular surfaces. Different from RGB images, polarization images can provide direct constraints on the azimuth angles of the surface normals. In this paper, we present PISR, a novel method that utilizes a geometrically accurate polarimetric loss to refine shape independently of appearance. In addition, PISR smooths surface normals in image space to eliminate severe shape distortions and leverages the hash-grid-based neural signed distance function to accelerate the reconstruction. Experimental results demonstrate that PISR achieves higher accuracy and robustness, with an L1 Chamfer distance of $0.5\mathrm{mm}$ and an F-score of $99.5\%$ at $1\mathrm{mm}$ , while converging $4 \sim 30\times$ faster than previous polarimetric surface reconstruction methods. + +Keywords: Polarization $\cdot$ Surface reconstruction $\cdot$ Multi-view reconstruction $\cdot$ Neural implicit surface $\cdot$ Polarimetric constraint + +# 1 Introduction + +Surface reconstruction from multi-view images is a fundamental problem in computer vision, with various applications in graphics, robotics, and more. Recently, neural implicit surface has shown powerful flexibility for representing 3D shape and appearance. Combined with differentiable volume rendering [23], neural implicit surface reconstruction is able to estimate 3D shapes from multi-view images accurately [42, 47]. However, these methods struggle with textureless and specular objects due to the shape-radiance ambiguity [52]. Recent methods resort to diffuse and specular separation modeling [40], specular reflection detection [9] and inverse rendering techniques [18, 51]. While these methods show promising results, they may still encounter difficulties since their optimizations rely on image reconstruction losses that tightly couple shape and appearance. + +![](images/6a469094d20c368d430fe3b7fa6a5133e6f628dce87676295cae92f606d652dc.jpg) + +![](images/a9a1be8c9f3aacf37a7af9a94811b3fe6741c52a3798b0a0f0ec1683286ad4ee.jpg) + +![](images/b43769659760d7d04a02cb3d57e87d681c848a8afa1b14128ddd2de94d077a7f.jpg) + +![](images/03958391f55e6d912b84c5d3ff6c3c770da3390fad1a49b0bdd521592166caca.jpg) +Fig. 1: Normal maps of reconstructed surfaces. By leveraging polarization information, our method is able to reconstruct textureless and specular surfaces with fine-grained details. + +![](images/79fe8c59d1ab29dcf49fabd5da5546765bf44b592933c83e5e5395b5265f2e31.jpg) + +![](images/bbe473ac767a597842053e424acc7a6b6c52d8b5791f226911cd83b0038db1ee.jpg) + +Polarization, as another property of light, provides direct cues on object shape. The angle of polarization of light from object surfaces is independent of light intensity and mainly determined by the surface shapes. From the perspective of a camera, the angle of polarization of a pixel is equal to the azimuth of the surface normal, i.e., the 2D projection direction of the normal onto the image plane, up to a $\pi/2$ -ambiguity and a $\pi$ -ambiguity [5]. This geometric property has been extensively explored in shape from polarization [1, 2, 25, 49], photo-polarimetric stereo [7, 8, 27, 38], polarimetric multi-view 3D reconstruction [5, 54, 55] and dense mapping [34, 46]. + +In this work, we introduce PISR (polarimetric neural implicit surface reconstruction), a novel method that combines the merits of polarimetric 3D reconstruction and neural implicit surface reconstruction for textureless and specular objects, such as ceramics and plastics. Different from recent works [6, 16] that utilize the polarization cues through polarimetric volume rendering, we propose a novel polarimetric loss to refine an object shape directly. Thanks to the integration of the perspective polarimetric constraint [4] in the proposed loss, PISR eliminates artifacts on the reconstructed surfaces, resulting in a $30\%$ reduction in L1 Chamfer distance. Besides, instead of treating all possible normals equally [54], our loss adaptively adjusts according to the degree of polarization to address specular-reflection-dominant surfaces, such as ceramics with dark color. + +To accelerate the optimization process, PISR uses a multi-resolution hash grid and multi-layer perceptrons (MLP) to represent an object, achieving $4 \sim 30 \times$ speed up as shown in Tab. 3. This representation is also crucial for accurately reconstructing highly reflective surfaces, enabling PISR to rectify the shape topology during optimization, a processing step not available for the mesh-based method, PMVIR [54]. However, the discrete nature of a hash grid also reduces the smoothness of the representation and can lead to the reconstruction + +results being trapped in local minima. To mitigate the problem, PISR smooths rendered normals for better robustness against poor shape initialization. + +Since there is no publicly available dataset for quantitative evaluation of polarimetric 3D reconstruction, we collect a dataset containing textureless and specular objects with a color polarization camera [36]. Using the dataset, we show that PISR is able to improve the surface reconstruction accuracy over the SOTA methods, with an L1 Chamfer distance of $0.5\mathrm{mm}$ and an F-score of $99.5\%$ at $1\mathrm{mm}$ . Shown in Fig. 1 are recovered shapes by different methods and our result shows an accurate shape and fine-grained details. + +In summary, our contributions are as follows: + +- PISR, a novel method that leverages polarization cues and hash-grid-based neural SDF for robust and efficient reconstruction of textureless and specular objects. +- A novel polarimetric loss that utilizes an accurate formulation of the polarimetric constraint for higher surface reconstruction accuracy. +- A polarization image dataset of real textureless and specular objects with ground-truth meshes for quantitative evaluation. + +# 2 Related Works + +Neural Implicit Surface Reconstruction: The success of neural radiance field reconstruction [23] in novel view synthesis greatly inspires the development of neural implicit surface [28, 30, 48]. Combining with differentiable volume rendering, neural implicit surface reconstruction [29, 42, 47] is able to reconstruct implicit surfaces from dense multi-view images while showing excellent performance in reconstructing fine-grained details. Since then, several works have improved it in reconstruction quality [50], sparse-view performance [19] and optimization speed [43]. + +Yet textureless and specular surfaces are still challenging for neural implicit surface reconstruction. To address the problem, NeRO [18] incorporates inverse rendering to alleviate the shape-radiance ambiguity [52] induced by specular reflection. Ref-NeuS [9] borrows the idea of anomaly detection to decrease the weights of specular pixels during optimization and conditions the radiance on reflection direction in reflective scenes as in Ref-NeRF [39]. UniSDF [40] models diffuse and specular radiances in two MLPs separately. + +Estimating a 3D radiance field and an SDF from 2D images is a highly ill-posed problem. During optimization, these methods may distort shapes to fit the reflection colors, causing severe reconstruction errors, or, as is commonly referred to, the shape-radiance ambiguity [52]. In contrast, polarization images provide linear constraints on surface normal by the angle of polarization. Therefore, to avoid the shape-radiance ambiguity, our method utilizes the geometry cues from polarization to regularize object shapes independently of appearance, keeping the shape from distortions induced by the visually distorted appearance. + +Polarimetric 3D Reconstruction: Shape from polarization (SfP) [1,25,31,49] and photo-polarimetric stereo [7,8,27,38] are single-view model-based methods that utilize polarization images for normal and depth estimation. These methods rely on assumptions of either spotlighting or distant lighting to explicitly model the relationship between the degree of polarization and the zenith angle of a surface normal, making them difficult to work under natural lighting conditions. Several deep learning-based methods [2,21,35] leverage data priors to enhance the robustness to more complex indoor lighting conditions. SfPW [15] and DPS-Net [37] learn to estimate scene-level normal maps and depth maps. Yet their generalization is still limited by the relatively small amount of training datasets of polarization images. + +Polarimetric multi-view 3D reconstruction leverages multi-view polarimetric constraints to adapt to natural lighting conditions. The polarimetric constraint is established on the fundamental fact that the azimuth angle of a surface normal is equal to the angle of polarization up to a $\pi/2$ -ambiguity and a $\pi$ -ambiguity under natural lighting [5, 54], which is relatively independent of lighting [5]. Miyazaki et al. [24] uses multi-view polarimetric constraints to reconstruct black specular objects. PolMVS [5] proposes to propagate the depths of sparse points and refine the depth maps with this constraint. Inspired by PolMVS, PDMS [46] and PMDM [34] integrate this constraint into SLAM for dense mapping. PMVIR [54] proposes a framework to simultaneously refine coarse meshes and estimate illuminations and surface albedos. PolarPMS [55] combines this constraint with photo consistency to estimate depth maps from multi-view polarization images. + +More closely related to our approach is PMVIR [54]. PMVIR represents shapes using meshes and thereby has limitations in handling incorrect shape topology. In contrast, by leveraging neural SDF, our method is able to rectify incorrect shape topology during the optimization. More importantly, our method utilizes the perspective polarimetric constraint [4] to consider the camera ray directions in the polarimetric loss and thus achieves further performance improvement. + +Polarimetric Inverse Rendering: The development of polarimetric bidirectional reflectance distribution function (pBRDF) [3, 11, 12, 14] has enabled polarimetric inverse rendering. Based on mesh representation, sparse ellipsometry [11] leverages polarimetric inverse rendering to simultaneously estimate object meshes and materials using polarization images captured under flashlight illumination. + +With the emergence of neural fields in 3D reconstruction, a few recent works [6, 16] have integrated polarization into neural implicit surfaces through polarization image rendering. Concurrent work NeISF [16] introduces the incident Stokes field to represent multi-path polarized light and reparameterizes pBRDF using MLPs, achieving highly accurate geometry reconstruction. PANDORA [6] aims to decompose diffuse and specular reflections using polarization as auxiliary cues, relying on the Fresnel equations to calculate the diffuse and specular components of radiance. + +Although polarimetric inverse rendering excels at accurate material and illumination estimation, it requires proper geometry initialization to decouple shape and appearance. As a result, it may still suffer from the shape-radiance ambiguity, especially when dealing with textureless and specular surfaces. To avoid this problem, our method utilizes the polarimetric constraint to regularize shape directly, independent of appearance. The surface reconstruction results of our method can be used to initialize polarimetric inverse rendering. + +# 3 Preliminary + +Polarimetric Constraint: Assuming four synchronized and aligned images captured at polarizer angles of 0, 45, 90, 135 degrees [36], denoted as $\mathbf{I}_0$ , $\mathbf{I}_{45}$ , $\mathbf{I}_{90}$ and $\mathbf{I}_{135}$ . For each pixel, the angle of polarization (AoP) $\varphi \in [-\pi /2,\pi /2]$ and degree of polarization (DoP) $\rho \in [0,1]$ are calculated as: + +$$ +\varphi = \frac {1}{2} \arctan 2 \left(s _ {2}, s _ {1}\right), \quad \rho = \frac {\sqrt {s _ {1} ^ {2} + s _ {2} ^ {2}}}{s _ {0}} \tag {1} +$$ + +where $\mathbf{s} = [s_0, s_1, s_2]^T = \left[\frac{1}{2}\sum I_i, I_0 - I_{90}, I_{45} - I_{135}\right]^T$ is the Stokes vector for representing the polarization state of light. Given the estimated normal $\hat{\mathbf{n}}$ in the camera frame and the AoP $\varphi$ of a pixel, the orthographic polarimetric constraint on the azimuth angle of the normal is calculated as: + +$$ +[ \sin (\varphi + \Delta), \cos (\varphi + \Delta), 0 ] \cdot \hat {\mathbf {n}} = 0, \tag {2} +$$ + +where the disambiguation scalar $\varDelta\in\{0,\pi/2\}$ is used for solving the $\pi /2$ ambiguity [5]. While this constraint is commonly used in previous works [5,34, 38,46,54], there is a perspective polarimetric constraint which is geometrically more accurate than it by considering the perspective effect of the lens [4]: + +$$ +\left[ v _ {z} \sin \varphi^ {\prime}, v _ {z} \cos \varphi^ {\prime}, - \left(v _ {y} \cos \varphi^ {\prime} + v _ {x} \sin \varphi^ {\prime}\right) \right] \cdot \hat {\mathbf {n}} = 0, \quad \varphi^ {\prime} = \varphi + \Delta \tag {3} +$$ + +where $\mathbf{v} = [v_x, v_y, v_z]^T$ is a normalized camera ray in the camera frame. + +$\pi /2$ ambiguity: The $\pi /2$ -ambiguity arises from the differences between the AoPs of specular reflection and diffuse reflection [5]. According to the Fresnel equations, reflected and transmitted light have distinct energy distributions in the parallel and perpendicular directions to the plane of incident, resulting in a $\pi /2$ difference between their AoPs, i.e., $\Delta = 0$ for diffuse reflection and $\Delta = \pi /2$ for specular reflection. Under natural lighting conditions, the diffuse reflections and specular reflections are usually mixed, making it difficult to distinguish the dominant component, and thus resulting in the $\pi /2$ -ambiguity. Therefore, the $\pi /2$ -ambiguity poses a challenge in terms of applying the polarimetric constraints. + +![](images/5a2e64f767d0edd25388219e1196d3019d957322d8aa6b5e2a8c01a6422f6e29.jpg) +Fig. 2: Overview of PISR. Pixels for the optimization are sampled in criss-cross patterns at the middle stage of the optimization. The polarimetric loss $\mathcal{L}_{\mathrm{pol}}^{\mathrm{p}}$ and the normal loss $\mathcal{L}_{\mathrm{normal}}$ are used for regularizing the shape independently of appearance. + +# 4 Method + +PISR aims to recover object shapes from multi-view polarization images and camera poses. Fig. 2 shows an overview of our method. Object shape is represented as a hash-grid-based neural SDF and rendered to images for computing losses (Sec. 4.1). During the reconstruction, the coarse shape of an object is estimated with the photometric loss $\mathcal{L}_{\mathrm{color}}$ . Then the polarimetric loss $\mathcal{L}_{\mathrm{pol}}^{\mathrm{p}}$ (Sec. 4.2) and the normal loss $\mathcal{L}_{\mathrm{normal}}$ with criss-cross pattern sampling (Sec. 4.3) are jointly used to rectify the shape from distortions. Finally, the shape is refined with $\mathcal{L}_{\mathrm{color}}$ and $\mathcal{L}_{\mathrm{pol}}^{\mathrm{p}}$ . The optimization scheme is introduced in Sec. 4.4. + +# 4.1 Surface Representation and Image Rendering + +Since Instant-NGP [26], combining a multi-resolution hash grid and MLPs has been commonly used for faster convergence in neural field reconstruction. Following recent works [17, 41, 43], we represent the shape and appearance of an object as an $L$ -level multi-resolution hash grid $\mathcal{G} = \{\mathcal{G}_l\}_{l=1}^L$ with two MLPs $\varPhi_s$ and $\varPhi_{\mathbf{c}}$ for decoding SDF values $s$ and color $\mathbf{c}$ respectively. For each 3D space point $\mathbf{x}$ , the SDF value $s$ and color $\mathbf{c}$ are computed as + +$$ +(s, \mathbf {f}) = \varPhi_ {s} (\mathcal {G} (\mathbf {x})), \quad \mathbf {c} = \varPhi_ {\mathbf {c}} (s, \mathbf {f}, \mathbf {v}, \mathbf {n}) \tag {4} +$$ + +with the camera ray $\mathbf{v}$ and surface normal $\mathbf{n}$ . The normal $\mathbf{n}$ can be computed by normalizing the gradient of SDF w.r.t. to $\mathbf{x}$ or by finite difference [17]. Besides the foreground object, we use Instant-NGP [26] to represent the background. + +To optimize the shape with images, we use differentiable volume rendering [22] to render images given the surface representation. The pixel color $\hat{\mathbf{c}}$ and normal $\hat{\mathbf{n}}$ are computed as the weighted sum of $M$ samples $\{\mathbf{x}_i = \mathbf{o} + t_i\mathbf{v}\}_{i=1}^M$ on a cast-off pixel ray with the camera position $\mathbf{o}$ and different depth $t_i$ : + +$$ +\hat {\mathbf {c}} = \sum_ {i = 1} ^ {M} w _ {i} \mathbf {c} _ {i}, \quad \hat {\mathbf {n}} = \sum_ {i = 1} ^ {M} w _ {i} \mathbf {n} _ {i}, \tag {5} +$$ + +where $w_{i}$ is the blending weight converted from SDF values using the unbiased weight function of NeuS [42], and $\mathbf{c}_i$ and $\mathbf{n}_i$ are the color and normal of the sample $\mathbf{x}_i$ . + +# 4.2 Polarimetric and Photometric Losses + +Polarimetric Loss: Previous works [5,34,38,46,54] leverage polarization cues with the following loss function $h^{\mathrm{o}}$ that is derived from the orthographic polarimetric constraint Eq. (2): + +$$ +h ^ {\mathrm {o}} (\varphi , \Delta , \mathbf {n}) = \left(\mathbf {a} ^ {\mathrm {o}} (\varphi , \Delta) \cdot \mathbf {n}\right) ^ {2}, \mathbf {a} ^ {\mathrm {o}} (\varphi , \Delta) = [ \sin (\varphi + \Delta), \cos (\varphi + \Delta), 0 ], \tag {6} +$$ + +where the coefficient vector $\mathbf{a}^{\mathrm{o}}$ is calculated with the AoP $\varphi$ and the disambiguation scalar $\Delta$ . However, as discussed in [4], Eq. (2) neglects the perspective effect of the lens and may introduce distortions in the reconstruction results. + +Therefore, we instead adopt the perspective polarimetric constraint Eq. (3) [4] and propose a novel loss function $h^{\mathrm{p}}$ : + +$$ +h ^ {\mathrm {p}} (\varphi , \Delta , \mathbf {v}, \mathbf {n}) = \left(\frac {\mathbf {a} ^ {\mathrm {p}} (\varphi , \Delta , \mathbf {v}) \cdot \mathbf {n}}{\| \mathbf {a} ^ {\mathrm {p}} (\varphi , \Delta , \mathbf {v}) \|}\right) ^ {2} \tag {7} +$$ + +with $\mathbf{a}^{\mathrm{p}}$ calculated as: + +$$ +\mathbf {a} ^ {\mathrm {p}} (\varphi , \Delta , \mathbf {v}) = \left[ v _ {z} \sin \varphi^ {\prime}, v _ {z} \cos \varphi^ {\prime}, - \left(v _ {y} \cos \varphi^ {\prime} + v _ {x} \sin \varphi^ {\prime}\right) \right], \varphi^ {\prime} = \varphi + \Delta . \tag {8} +$$ + +The loss function $h^{\mathrm{p}}$ additionally considers the camera ray direction $\mathbf{v}$ to model the perspective effect. In Sec. 5.3, we will show that $h^{\mathrm{p}}$ effectively improves the reconstruction accuracy. + +Besides, the $\pi/2$ -ambiguity must be addressed when applying the polarimetric constraint. We combine the merits of PolMVS [5] and PMVIR [54], proposing the segmented loss function $f^{\mathrm{p}}$ as follows: + +$$ +f ^ {\mathrm {p}} (\varphi , \mathbf {v}, \mathbf {n}, \rho) = \left\{ \begin{array}{l l} h ^ {\mathrm {p}} (\varphi , 0, \mathbf {v}, \mathbf {n}) \cdot h ^ {\mathrm {p}} (\varphi , \pi / 2, \mathbf {v}, \mathbf {n}) & \text {i f} \rho < \theta , \\ h ^ {\mathrm {p}} (\varphi , \pi / 2, \mathbf {v}, \mathbf {n}) & \text {i f} \rho \geq \theta , \end{array} \right. \tag {9} +$$ + +where $\rho$ is the DoP and $\theta$ is a preset threshold. This design is based on two observations: 1) the reflected light of an object is usually a mixture of the diffuse and specular components, and thus the DoP is relatively low [5], 2) only the DoP of specular reflection of dielectric objects can exceed 0.4 according to Fresnel equations. We use multiplication instead of addition when $\rho < \theta$ because only one of the two needs to be minimized. + +Finally, we calculate the polarimetric loss for a batch of sampled pixels $S$ as follows: + +$$ +\mathcal {L} _ {\mathrm {p o l}} ^ {\mathrm {p}} = \frac {1}{| S |} \sum_ {u \in \mathcal {S}} f ^ {\mathrm {p}} \left(\varphi_ {u}, \mathbf {v} _ {u}, \hat {\mathbf {n}} _ {u}, \rho_ {u}\right). \tag {10} +$$ + +Note that the camera ray $\mathbf{v}_u$ and normal $\hat{\mathbf{n}}_u$ are transformed into the camera frame according to the camera pose. In Sec. 5.3, the experiments will demonstrate that calculating the loss $\mathcal{L}_{\mathrm{pol}}^{\mathrm{p}}$ with $h^{\mathrm{p}}$ instead of $h^{\mathrm{o}}$ can reduce the L1 Chamfer distance by $30\%$ . + +Photometric Loss: Since the polarimetric loss alone is insufficient for recovering object shapes due to the $\pi /2$ ambiguity and image noises, we used the photometric loss for initializing and maintaining the basic shape: + +$$ +\mathcal {L} _ {\text {c o l o r}} = \frac {1}{| \mathcal {S} |} \sum_ {u \in \mathcal {S}} \left| \hat {\mathbf {c}} _ {u} - \mathbf {c} _ {u} \right|. \tag {11} +$$ + +# 4.3 Normal Regularization + +The discrete nature of the hash grid makes the estimated SDF less smooth than those purely MLP-based, which can introduce reconstruction artifacts such as holes or cracks. Directly smoothing the normals in 3D space as in [17, 53] can lead to a bumpy surface because the sampled normals are solely constrained by the smoothness loss. Inspired by the belief propagation used in traditional MVS methods [33, 45, 56], we smooth the rendered normals so that all the samples are also constrained by both the photometric loss and the polarimetric loss. We therefore utilize the following normal loss: + +$$ +\mathcal {L} _ {\text {n o r m a l}} = - \frac {1}{| \mathcal {S} _ {\mathrm {c}} |} \sum_ {u \in \mathcal {S} _ {\mathrm {c}}} \frac {1}{| \mathcal {N} _ {u} |} \sum_ {\hat {\mathbf {n}} _ {j} \in \mathcal {N} _ {u}} \operatorname {S G} \left(\hat {\mathbf {n}} _ {u}\right) ^ {T} \hat {\mathbf {n}} _ {j} \tag {12} +$$ + +where $S_{c}$ is a set of pixels sampled among pixels of all the images, $\mathcal{N}_u$ is a set of pixel normals sampled around $\hat{\mathbf{n}}_u$ in the image space. The function SG ("Stop Gradient") is used to stop propagating the gradients of the losses w.r.t. $\hat{\mathbf{n}}_u$ during backpropagation, i.e., fix the values of $\hat{\mathbf{n}}_u$ so that it can be propagated to neighbor pixels. + +Now the pixel sample set $S$ of each iteration during the optimization is the union of the set of center pixels $S_{c}$ and the sets of neighbor pixels $\mathcal{N}_u$ : + +$$ +\mathcal {S} = \mathcal {S} _ {c} \cup \left\{\bigcup_ {u \in S _ {c}} \mathcal {N} _ {u} \right\} \tag {13} +$$ + +Empirically, larger $\mathcal{S}_c$ can lead to better results because it helps capture global structures, which are important for optimizing a neural field [44], while larger $\mathcal{N}_u$ is beneficial to smooth the surface. Therefore, we sample $\mathcal{N}_u$ in a criss-cross pattern instead of a whole patch around $\hat{\mathbf{n}}_u$ to balance $|\mathcal{S}_c|$ and $|\mathcal{N}_u|$ with fixed $|S|$ . In principle, $\mathcal{N}_u$ can be sampled in other relatively sparse patterns. Here we use the criss-cross one for its simple implementation in code. + +# 4.4 Optimization Strategy + +The overall loss is + +$$ +\mathcal {L} _ {\text {a l l}} = \mathcal {L} _ {\text {c o l o r}} + \lambda_ {\mathrm {p}} \mathcal {L} _ {\text {p o l}} ^ {\mathrm {p}} + \lambda_ {\mathrm {n}} \mathcal {L} _ {\text {n o r m a l}} + \lambda_ {\mathrm {e}} \mathcal {L} _ {\text {e i k o n a l}} \tag {14} +$$ + +![](images/ad2a90825e6a009c1f1cda070f0a1e0a427a491413a6cc3cfbecf8222c46ead7.jpg) +(a) + +![](images/4482a83617f06baf723fd609e86a07738f14f38a91835588448fc51771ec4e25.jpg) +(b) +Fig. 3: Dataset and collection. (a) Objects with ground-truth meshes. (b) Objects without ground-truth meshes. (c) Capture setup. + +![](images/a60ac6b24b2d8417635b8e40757e67e9d17fe15162457302ca6a5593ba9a3cb3.jpg) +(c) + +The extra Eikonal loss $\mathcal{L}_{\mathrm{eikonal}} = \frac{1}{M|S|}\sum_{k=1}^{M|S|}(\|\nabla\hat{s}(\mathbf{x}_k)\| - 1)^2$ is used for encouraging the neural field to approximate a signed distance field [10]. During the optimization, we apply progressive training [17] on the multi-resolution hash grid. To initialize a coarse shape, we set $\lambda_{\mathrm{p}} = \lambda_{\mathrm{n}} = 0$ and use the photometric and eikonal losses at the early stage. Then we gradually increase $\lambda_{\mathrm{p}}$ and $\lambda_{\mathrm{n}}$ to rectify the distorted shape. Finally, we gradually decrease $\lambda_{\mathrm{n}}$ to 0 to avoid over smoothing the details. + +# 5 Experiments + +# 5.1 Experimental Setup + +Dataset: Since there is no public polarization image dataset of textureless and specular objects with ground-truth meshes, we collect a real-world dataset, including four objects (Black Dragon, Red Dragon, Standing Rabbit, Lying Rabbit, Figure and Car) as shown in Fig. 3. The reasons for capturing real object images, rather than synthesizing using a rendering engine, such as [13], are two-fold: 1) the natural ambient lighting conditions are complex and therefore difficult to synthesize, and 2) the shadows and occlusions of the camera can be realistically generated. + +Each object's polarization images are captured indoors under natural lighting conditions from 40 (60 for Figure for its relatively complex shape) different viewpoints around the object with a color polarization camera [36]. Color images are obtained by first decomposing polarization images into images of four polarizer angles and then averaging and demosaicing. Camera poses are estimated from color images using SuperGlue [32] and COLMAP [33] given pre-calibrated camera intrinsics. We use a checkerboard to help pose estimation. To obtain ground-truth meshes, the objects are coated by 3D scanning spray to eliminate specular reflection and scanned by a structure light-based scanner except for Figure and Car. Please refer to our supplementary material for more details. + +![](images/af839f7fac945c5590b7ba519169fc518c49a16245ce1ac542e8e1f20729b32e.jpg) +Baaee + +![](images/402c09e0b7c2f735a2509080170bd6be041b5d78bf0c57c7930944245cccd929.jpg) + +![](images/afb9c62b8b4d84827e8c7916c0d6d406c7bcb31dad5ebdffb9eb467c5e433a11.jpg) + +![](images/390a22fd15d12af1da682339f55c73b3fb93e8ea777d4f5b65b6bc39e78a516d.jpg) + +![](images/7062568af730f6c1253dc2085c3eca3eb5c9e2b11e0598e8202691cfffd82a76.jpg) + +![](images/61c5d10507c487387c2ef71eeda335f322d42a005aaae000c4ad9f94c32865d0.jpg) + +![](images/aaa03de075311943c9d8c87a64bae4ce9266bfef6f8e9bda6dcfc4484474a6d0.jpg) + +![](images/9f30589a6477021164cc8b6be51d64d03893dc6a7f622d12385b259030df037d.jpg) +Rrnnnne + +![](images/305e0fcc06b9c08958713a3231d6b7fd32065ec2214600a4ca8b473e9497ea41.jpg) + +![](images/f265945b3935956bd26ef4c63b2293b7ae64cf42f97de268e9eecd7198645d21.jpg) + +![](images/9606c7417819930d655e83645eb54dc0bd03c34e8818f2154a5f1c59d30bd4cc.jpg) + +![](images/76b2f19e9fce38ef428ac2d2d9cce2cedbb3d90478328e8f3373fac83c19a65b.jpg) + +![](images/ff231fc2dc87098a84b70215659f9173c94fa48f3224b4bb3c6b9e63d1d7fa05.jpg) + +![](images/47cd50d7ccdb63edd1679924aa24fd59372459fbe1f1fa200df06b218b51e0ee.jpg) +Rabitt S. + +![](images/b74f8ea4b5a4492c45304007ba65d9435825d1c380392928732b05fce8d683cd.jpg) + +![](images/86ce32ee94a377de92169710dda7344dc8412f3c5432cd142b4bd5bf32a0354a.jpg) + +![](images/bd7ac0a4ff2779891e7487a0635ee61474cd59a66448ca4e96233f5b8db64c5e.jpg) + +![](images/4439d24aad9e33560e33cd0609d3f4ba9f5709d06677687d8f1e32841fec8096.jpg) + +![](images/d5bf77d25e8ed23cc261fca96f7e2f0997e9c2c1e0dde4b70937ad091ecff674.jpg) + +![](images/16097e7320adf1e3539591de277398634af04835b033a7fc4d41b4715c4b8aea.jpg) +Rabibit L + +![](images/68d8cfb4ca6944f5204923ceeb568d26aa322fb1c02321043fb8ead27430e6a2.jpg) +Fig. 4: Signed error maps. Red and blue mean surface swelling and shrinking respectively. Regions with higher color saturation indicate longer distances between the reconstruction result and the ground truth. Errors are truncated to within $\pm 2\mathrm{mm}$ for better visualization. + +![](images/064fbcc29c2aa257066d05b6fe93579dcc289a6747e2726699c43cc8fd673364.jpg) + +![](images/b38268a9062d69a8ba55f4189b733e82426ac06681ffc84860aa545ea0c7ff00.jpg) +PANDORA + +![](images/04cc31e212237a2f738d991801c3ac69647ad7ac6414ca84d647ee4142dcf504.jpg) +PMVIR + +![](images/847a5e8d15f4b341aeb9f51669bb0703530c86b954db100484761f3fdb161633.jpg) +PISR (ours) + +Table 1: Summary of Methods in Experiments + +
MethodsImageShape Repre.Appearance Repre.
Ref-NeuS [9]RGBNeural SDFNeural radiance
NeRO [18]RGBNeural SDFNeural BRDF
NeuS [42]RGBNeural SDFNeural radiance
PANDORA [6]Pol. RGBNeural SDFNeural BRDF
PMVIR [54]Pol. RGBMeshSpherical harmonics
PISR (Ours)Pol. RGBGrids+Neural SDFNeural radiance
+ +Performance Metrics: Since we focus on geometry reconstruction, we only evaluate the surface reconstruction quality using L1 Chamfer distance and F-score. We remove the background meshes before calculating the metrics. + +Implementation Details: The SDF MLP $\varPhi_s$ has one hidden layer with a size of 64, and the color MLP $\varPhi_{\mathbf{c}}$ has two hidden layers with sizes of 64. The multi-resolution grid contains 16 levels of resolution, and the coarsest and finest resolutions are 32 and 2700, respectively. The hash table size of each level is $2^{19}$ . The optimization of the grid starts at the fourth-coarsest resolutions and progressively extends one level every 1000 iterations. Throughout the whole optimization process, the threshold of DoP $\theta$ is set to 0.3 and $\lambda_{\mathrm{eikonal}}$ is set to 0.1. After the first 2.5k iterations, $\lambda_{\mathrm{pol}}$ , $\lambda_{\mathrm{normal}}$ and $|\mathcal{N}_p|$ are linearly increased from + +Table 2: Reconstruction accuracy in Chamfer distance (CD↓, mm) and F-score (FS↑, %). FS is calculated at a threshold of 1 mm. Note that the values in the table are rounded to one decimal place. The CD of PMVIR and PISR on Rabbit L. are 0.584 mm and 0.578 mm respectively. Bold, underline, and double underline mean the best, second-best and third-best performance, respectively. + +
MethodsDragon B.Dragon R.Rabbit S.Rabbit L.Avg.
CDFSCDFSCDFSCDFSCDFS
RGBRef-NeuS2.165.32.166.11.284.51.090.51.676.6
NeRO2.459.21.866.31.380.51.385.01.772.8
NeuS1.967.11.286.01.186.80.993.01.383.2
RGBPANDORA2.355.31.864.41.478.31.577.11.868.8
PMVIR1.876.51.090.50.697.50.697.31.090.5
PISR (Ours)0.599.90.699.10.599.70.699.40.599.5
+ +0 to 2.0, 1.0 and 28 in $2.5\mathrm{k}$ iterations, respectively. Then $\lambda_{\mathrm{normal}}$ and $|\mathcal{N}_p|$ are linearly decreased to 0 in another $2.5\mathrm{k}$ iterations. The maximum number of iterations is $20\mathrm{k}$ . The final SDF is converted to meshes using Marching Cubes [20]. + +Baselines: We compare our method with five methods, including PMVIR [54], PANDORA [6], NeRO [18], Ref-NeuS [9] and NeuS [42]. Tab. 1 summarizes the type of input images, geometry representation and color representation of all the six methods. Since PMVIR is a refinement method based on an initial shape, we provide the coarse initial meshes reconstructed by PISR-C (see the ablation study in Sec. 5.3 for the details of PISR-C). + +# 5.2 Results + +Quantitative Results: The Chamfer distances and F-scores of the reconstruction results are reported in Tab. 2. Our method achieves the best reconstruction quality on all four objects shown in Fig. 3a and significantly outperforms the other five methods. The average Chamfer distance and F-score of our method are $0.5 \, \mathrm{mm}$ and $99.5\%$ , respectively. As the second-best method in the comparison, PMVIR achieves twice the Chamfer distance and a $9\%$ lower F-score compared to ours. Both PMVIR and our method utilize polarimetric constraints, and the results show that the geometry cue of polarization is effective for shape estimation. Although PANDORA also uses polarization images as input, it implicitly utilizes polarization cues through polarimetric rendering, i.e., optimizing shapes using an image reconstruction loss, which could suffer from the shape-radiance ambiguity as RGB-based methods. Among RGB-based methods, NeuS achieves better results than Ref-NeuS and NeRO which are designed for specular objects. + +We visualize the error distributions in Fig. 4. For methods that rely solely on image reconstruction losses (Ref-NeuS, NeRO, NeuS and PANDORA), their reconstruction errors in non-convex regions are notably higher compared to + +![](images/4ff13a0f845138e44cc2454e391498a61ed2575212ca31a0f43b8e964a7357a8.jpg) +Initialization + +![](images/0b11c7fca60e3e8be9ed524bbd37ca08b6d52b836ed3a90e0dd9c8e87cd4071c.jpg) +PMVIR + +![](images/cd9c71569bc759d4b33bef45b1e71bf08f5bf8134105d35b2bf99a2aed20cd9b.jpg) +PISR +Fig. 5: Normal maps of Dragon B. before and after optimizations. Using neural SDFs to represent shapes allows changes in the shape topology during the optimization. + +Table 3: Average optimization time of different methods. + +
MethodsRef-NeuSNeRONeuSPANDORAPMVIRPISR (Ours)
Iterations200k200k200k100k420k
Time (h)20.020.510.815.02.00.5
+ +PMVIR and our method. While the inherent continuity of MLPs ensures the watertightness of their reconstructed shapes, their accuracy cannot be guaranteed due to the shape-radiance ambiguity introduced by specular reflections and the lack of textures. Additionally, for NeRO and PANDORA, the lighting shadows cast by the camera further exacerbate the difficulty of estimating surface material parameters. The real HDR ambient lighting also poses challenges for PANDORA and Ref-NeuS to accurately estimate object shapes. In contrast, PMVIR and our method can recover non-convex regions using the polarimetric constraints. Despite the polarization cues, representing the shapes as neural SDF helps our method better reconstruct regions that are poorly initialized. As shown in Fig. 5, neural SDF allows continuous changes of the shape topology during optimization, which is crucial for evolving surfaces from holes. + +**Optimization Time:** We report the average optimization time in Tab. 3. The experiments are conducted on a machine equipped with an Intel Core i9-12900K and an NVIDIA RTX 3090 GPU. We set the maximum number of iterations to $200\mathrm{k}$ . If the shape converges, we stop the optimization earlier. Our method represents shape as a multi-resolution hash grid and two shallow MLPs, significantly reducing the time complexity for rendering a pixel. Therefore, our method converges much faster, achieving up to $4 \sim 41 \times$ convergence speed while maintaining the highest reconstruction accuracy. + +# 5.3 Ablation Study + +We evaluate the effectiveness of the polarimetric loss and normal regularization, as well as the influence of the perspective effect of the lens. We compare PISR(- + +Table 4: Results of the ablation study on different combinations of losses. The L1 CD↓ (mm) and FSs↑ (%) at 1.0 mm and 0.5 are the averages over four objects in Fig.3a. The ranks are the averages over the three metrics. $\mathcal{L}_{\mathrm{pol}}^{\mathrm{o}}$ differs from $\mathcal{L}_{\mathrm{pol}}^{\mathrm{p}}$ by its use of the orthographic constraint Eq. 2 instead of the perspective one Eq. 3. Note that $\mathcal{L}_{\mathrm{eikonal}}$ is used in all experiments and is omitted it in this table. + +
VariantsUsed LossesAvg.Rank
LcolorLpoiLpolLnormalCDFS(1.0)FS(0.5)
PISR-C✓1.1088.1070.306
PISR-N✓✓1.0088.2071.105
PISR-O✓✓0.8096.6075.543
PISR-ON✓✓✓0.7796.1075.504
PISR-P✓✓0.5199.7692.791
PISR(-PN)✓✓✓0.5499.5491.752
+ +PN) with its five variants, referred to as $PISR-P$ , $PISR-ON$ , $PISR-O$ , $PISR-N$ and $PISR-C$ , based on different combinations of losses as shown in Tab. 4. The quantitative experiment is conducted on objects shown in Fig. 3a and the qualitative experiment is conducted on objects shown in Fig. 3b. + +As shown in Tab. 4, the top four performing variants are all incorporated with the polarimetric losses, and fine-grained details can be recovered without complicated appearance modeling. As for the normal regularization, although the quantitative comparison between PISR-PN and PISR-P shows that it slightly reduces the accuracy, the qualitative results in Fig. 6 show its effectiveness for recovering surfaces from severe distortions. + +Among the top four performing variants, PISR-PN and PISR-P achieve higher accuracy by considering the perspective effect of the lens using Eq. (3) instead of Eq. (2). As the example error maps of Rabbit L. shown in Fig. 7, the perspective effect leads to surface swelling and shrinking, causing severe cracks on the surface. This phenomenon does not appear in the results of PMVIR and PANDORA. We speculate the reason to be that the hash-grid-based neural SDF we use is too flexible in topology compared to the mesh-based one used in PMVIR, and lacks the smoothness of purely MLP-based one used in PANDORA. Besides, the results in Tab. 4 show that PISR-PN outperforms PISR-ON by a $30\%$ reduction in L1 Chamfer distance and a $17\%$ improvement in F-score at $0.5 \mathrm{~mm}$ , which suggests the necessity of considering the perspective effect to improve the accuracy of polarimetric 3D reconstruction. + +# 6 Conclusion + +We have proposed PISR, an accurate and efficient polarimetric neural implicit surface reconstruction method for textureless and specular objects. The key idea of PISR is to explicitly regularize the SDF using the polarimetric constraints to avoid the shape-radiance ambiguity. More importantly, PISR achieves high accu + +![](images/98032b4b80934bcbd775b5d12cd8772aa34459ecad177d68ece041a695612cf2.jpg) +Fig. 6: Normal maps of Figure and Car. The polarimetric loss and normal loss jointly recover the shape from distortions induced by strong reflections. + +![](images/d5d3b40d27e58a5369d5df98e9c3be7a9f49ba008cd2cf5197d0147d817945ac.jpg) +Fig. 7: Signed error maps of Rabbit L. PISR-PN leverages the perspective polarimetric constraint to eliminate surface cracks encountered by PISR-ON. Additional error maps can be found in our supplementary material. + +racy by leveraging the perspective polarimetric constraint, eliminating the surface cracks caused by the traditional orthographic one. Besides, PISR regularizes surface normals in image space for robustness and integrates the multi-resolution hash grid for efficiency. + +Limitations: Despite the high reconstruction accuracy in the experiments, PISR's performance degenerates with noisy AoP maps, limiting its application in reconstructing rough, well-textured or mirror-like objects. Recent works [16,18] have shown solutions to mitigate this problem. In addition, PISR also requires dense input views, adding shape constraints derived from normal and depth estimation models would be a promising solution. + +# References + +1. Atkinson, G.A., Hancock, E.R.: Recovery of surface orientation from diffuse polarization. IEEE transactions on image processing 15(6), 1653-1664 (2006) +2. Ba, Y., Gilbert, A., Wang, F., Yang, J., Chen, R., Wang, Y., Yan, L., Shi, B., Kadambi, A.: Deep shape from polarization. In: European Conference on Computer Vision. pp. 554-571. Springer (2020) +3. Baek, S.H., Zeltner, T., Ku, H., Hwang, I., Tong, X., Jakob, W., Kim, M.H.: Image-based acquisition and modeling of polarimetric reflectance. ACM Trans. Graph. 39(4), 139 (2020) +4. Chen, G., He, L., Guan, Y., Zhang, H.: Perspective phase angle model for polarimetric 3d reconstruction. In: European Conference on Computer Vision. pp. 398-414. Springer (2022) +5. Cui, Z., Gu, J., Shi, B., Tan, P., Kautz, J.: Polarimetric multi-view stereo. In: Proceedings of the IEEE conference on computer vision and pattern recognition. pp. 1558-1567 (2017) +6. Dave, A., Zhao, Y., Veeraraghavan, A.: Pandora: Polarization-aided neural decomposition of radiance. In: European Conference on Computer Vision. pp. 538-556. Springer (2022) +7. Ding, Y., Ji, Y., Ye, J.: Polar-photometric stereo under natural illumination. In: 2022 International Conference on 3D Vision (3DV). pp. 690–699. IEEE (2022) +8. Ding, Y., Ji, Y., Zhou, M., Kang, S.B., Ye, J.: Polarimetric helmholtz stereopsis. In: Proceedings of the IEEE/CVF International Conference on Computer Vision. pp. 5037-5046 (2021) +9. Ge, W., Hu, T., Zhao, H., Liu, S., Chen, Y.C.: Ref-neus: Ambiguity-reduced neural implicit surface learning for multi-view reconstruction with reflection. arXiv preprint arXiv:2303.10840 (2023) +0. Gropp, A., Yariv, L., Haim, N., Atzmon, M., Lipman, Y.: Implicit geometric regularization for learning shapes. In: Proceedings of the 37th International Conference on Machine Learning. pp. 3789-3799 (2020) +1. Hwang, I., Jeon, D.S., Munoz, A., Gutierrez, D., Tong, X., Kim, M.H.: Sparse ellipsometry: portable acquisition of polarimetric svbrdf and shape with unstructured flash photography. ACM Transactions on Graphics (TOG) 41(4), 1-14 (2022) +2. Ichikawa, T., Fukao, Y., Nobuhara, S., Nishino, K.: Fresnel microfacet brdf: Unification of polari-radiometric surface-body reflection. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 16489-16497 (2023) +3. Jakob, W., Speierer, S., Roussel, N., Nimier-David, M., Vicini, D., Zeltner, T., Nicolet, B., Crespo, M., Leroy, V., Zhang, Z.: Mitsuba 3 renderer (2022), https://mitsuba-renderer.org +4. Kondo, Y., Ono, T., Sun, L., Hirasawa, Y., Murayama, J.: Accurate polarimetric brdf for real polarization scene rendering. In: Computer Vision-ECCV 2020: 16th European Conference, Glasgow, UK, August 23-28, 2020, Proceedings, Part XIX 16. pp. 220-236. Springer (2020) +5. Lei, C., Qi, C., Xie, J., Fan, N., Koltun, V., Chen, Q.: Shape from polarization for complex scenes in the wild. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR). pp. 12632-12641 (June 2022) +6. Li, C., Ono, T., Uemori, T., Mihara, H., Gatto, A., Nagahara, H., Moriuchi, Y.: Neisf: Neural incident stokes field for geometry and material estimation. arXiv preprint arXiv:2311.13187 (2023) + +17. Li, Z., Müller, T., Evans, A., Taylor, R.H., Unberath, M., Liu, M.Y., Lin, C.H.: Neuralangelo: High-fidelity neural surface reconstruction. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 8456-8465 (2023) +18. Liu, Y., Wang, P., Lin, C., Long, X., Wang, J., Liu, L., Komura, T., Wang, W.: Nano: Neural geometry and brdf reconstruction of reflective objects from multiview images. In: SIGGRAPH (2023) +19. Long, X., Lin, C., Wang, P., Komura, T., Wang, W.: Sparseneus: Fast generalizable neural surface reconstruction from sparse views. In: European Conference on Computer Vision. pp. 210-227. Springer (2022) +20. Lorensen, W.E., Cline, H.E.: Marching cubes: A high resolution 3d surface construction algorithm. ACM SIGGRAPH Computer Graphics 21(4), 163-169 (1987) +21. Lyu, Y., Zhao, L., Li, S., Shi, B.: Shape from polarization with distant lighting estimation. IEEE Transactions on Pattern Analysis and Machine Intelligence 45(11), 13991-14004 (2023). https://doi.org/10.1109/TPAMI.2023.3298376 +22. Martin-Brualla, R., Radwan, N., Sajjadi, M.S., Barron, J.T., Dosovitskiy, A., Duckworth, D.: Nerf in the wild: Neural radiance fields for unconstrained photo collections. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 7210–7219 (2021) +23. Mildenhall, B., Srinivasan, P.P., Tancik, M., Barron, J.T., Ramamoorthi, R., Ng, R.: Nerf: Representing scenes as neural radiance fields for view synthesis. Communications of the ACM 65(1), 99-106 (2021) +24. Miyazaki, D., Shigetomi, T., Baba, M., Furukawa, R., Hiura, S., Asada, N.: Polarization-based surface normal estimation of black specular objects from multiple viewpoints. In: 2012 Second International Conference on 3D Imaging, Modeling, Processing, Visualization & Transmission. pp. 104-111. IEEE (2012) +25. Miyazaki, D., Tan, R.T., Hara, K., Ikeuchi, K.: Polarization-based inverse rendering from a single view. In: Computer Vision, IEEE International Conference on. vol. 3, pp. 982-982. IEEE Computer Society (2003) +26. Müller, T., Evans, A., Schied, C., Keller, A.: Instant neural graphics primitives with a multiresolution hash encoding. ACM Transactions on Graphics (ToG) 41(4), 1-15 (2022) +27. Ngo Thanh, T., Nagahara, H., Taniguchi, R.i.: Shape and light directions from shading and polarization. In: Proceedings of the IEEE conference on computer vision and pattern recognition. pp. 2310-2318. IEEE (2015) +28. Niemeyer, M., Mescheder, L., Oechsle, M., Geiger, A.: Differentiable volumetric rendering: Learning implicit 3d representations without 3d supervision. In: Proc. IEEE Conf. on Computer Vision and Pattern Recognition (CVPR) (2020) +29. Oechsle, M., Peng, S., Geiger, A.: Unisurf: Unifying neural implicit surfaces and radiance fields for multi-view reconstruction. In: International Conference on Computer Vision (ICCV) (2021) +30. Park, J.J., Florence, P., Straub, J., Newcombe, R., Lovegrove, S.: Deepsdf: Learning continuous signed distance functions for shape representation. In: Proceedings of the IEEE/CVF conference on computer vision and pattern recognition. pp. 165-174 (2019) +31. Rahmann, S., Canterakis, N.: Reconstruction of specular surfaces using polarization imaging. In: Proceedings of the 2001 IEEE Computer Society Conference on Computer Vision and Pattern Recognition. CVPR 2001. vol. 1, pp. I-I. IEEE (2001) +32. Sarlin, P.E., DeTone, D., Malisiewicz, T., Rabinovich, A.: SuperGlue: Learning feature matching with graph neural networks. In: CVPR (2020) + +33. Schonberger, J.L., Frahm, J.M.: Structure-from-motion revisited. In: Conference on Computer Vision and Pattern Recognition (CVPR) (2016) +34. Shakeri, M., Loo, Y., Zhang, H.: Polarimetric monocular dense mapping using relative deep depth prior [j]. 2021. Shakeri M Loo Y Zhang H Polarimetric Monocular Dense Mapping Using Relative Deep Depth Prior (2021) +35. Shao, M., Xia, C., Yang, Z., Huang, J., Wang, X.: Transparent shape from a single view polarization image. In: Proceedings of the IEEE/CVF International Conference on Computer Vision (ICCV). pp. 9277-9286 (October 2023) +36. Sony: Polarization image sensor with four-directional on-chip polarizer and global shutter function, https://www.sony-semicon.co.jp/e/products/IS/industry/product/polarization.html, accessed 2024-02-04 +37. Tian, C., Pan, W., Wang, Z., Mao, M., Zhang, G., Bao, H., Tan, P., Cui, Z.: Dpsnet: Deep polarimetric stereo depth estimation. In: The IEEE/CVF International Conference on Computer Vision (ICCV) (2023) +38. Tozza, S., Zhu, D., Smith, W., Ramamoorthi, R., Hancock, E.: Uncalibrated, two source photo-polarimetric stereo. IEEE Transactions on Pattern Analysis and Machine Intelligence (2021) +39. Verbin, D., Hedman, P., Mildenhall, B., Zickler, T., Barron, J.T., Srinivasan, P.P.: Ref-nerf: Structured view-dependent appearance for neural radiance fields. In: 2022 IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR). pp. 5481-5490. IEEE (2022) +40. Wang, F., Rakotosaona, M.J., Niemeyer, M., Szeliski, R., Pollefeys, M., Tombari, F.: Unisdf: Unifying neural representations for high-fidelity 3d reconstruction of complex scenes with reflections. In: arXiv (2023) +41. Wang, H., Wang, J., Agapito, L.: Co-slam: Joint coordinate and sparse parametric encodings for neural real-time slam. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 13293-13302 (2023) +42. Wang, P., Liu, L., Liu, Y., Theobalt, C., Komura, T., Wang, W.: Neus: Learning neural implicit surfaces by volume rendering for multi-view reconstruction. Advances in Neural Information Processing Systems 34, 27171-27183 (2021) +43. Wang, Y., Han, Q., Habermann, M., Daniilidis, K., Theobalt, C., Liu, L.: Neus2: Fast learning of neural implicit surfaces for multi-view reconstruction. In: Proceedings of the IEEE/CVF International Conference on Computer Vision. pp. 3295-3306 (2023) +44. Xie, Z., Yang, X., Yang, Y., Sun, Q., Jiang, Y., Wang, H., Cai, Y., Sun, M.: S3im: Stochastic structural similarity and its unreasonable effectiveness for neural fields. In: Proceedings of the IEEE/CVF International Conference on Computer Vision. pp. 18024-18034 (2023) +45. Xu, Q., Tao, W.: Multi-scale geometric consistency guided multi-view stereo. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR) (June 2019) +46. Yang, L., Tan, F., Li, A., Cui, Z., Furukawa, Y., Tan, P.: Polarimetric dense monocular slam. In: Proceedings of the IEEE conference on computer vision and pattern recognition. pp. 3857-3866 (2018) +47. Yariv, L., Gu, J., Kasten, Y., Lipman, Y.: Volume rendering of neural implicit surfaces. In: Thirty-Fifth Conference on Neural Information Processing Systems (2021) +48. Yariv, L., Kasten, Y., Moran, D., Galun, M., Atzmon, M., Ronen, B., Lipman, Y.: Multiview neural surface reconstruction by disentangling geometry and appearance. Advances in Neural Information Processing Systems 33 (2020) + +49. Yu, Y., Zhu, D., Smith, W.A.: Shape-from-polarisation: a nonlinear least squares approach. In: Proceedings of the IEEE International Conference on Computer Vision Workshops. pp. 2969-2976 (2017) +50. Yu, Z., Peng, S., Niemeyer, M., Sattler, T., Geiger, A.: Monosdf: Exploring monocular geometric cues for neural implicit surface reconstruction. Advances in neural information processing systems 35, 25018-25032 (2022) +51. Zhang, K., Luan, F., Wang, Q., Bala, K., Snavely, N.: PhySG: Inverse rendering with spherical gaussians for physics-based material editing and relighting. In: The IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR) (2021) +52. Zhang, K., Riegler, G., Snavely, N., Koltun, V.: Nerf++: Analyzing and improving neural radiance fields. arXiv:2010.07492 (2020) +53. Zhang, X., Srinivasan, P.P., Deng, B., Debevec, P., Freeman, W.T., Barron, J.T.: Nerfactor: Neural factorization of shape and reflectance under an unknown illumination. ACM Transactions on Graphics (ToG) 40(6), 1-18 (2021) +54. Zhao, J., Monno, Y., Okutomi, M.: Polarimetric multi-view inverse rendering. IEEE Transactions on Pattern Analysis and Machine Intelligence (2022) +55. Zhao, J., Oishi, J., Monno, Y., Okutomi, M.: Polarimetric patchmatch multi-view stereo. In: Proceedings of the IEEE/CVF Winter Conference on Applications of Computer Vision. pp. 3476-3484 (2024) +56. Zheng, E., Dunn, E., Jojic, V., Frahm, J.M.: Patchmatch based joint view selection and depthmap estimation. In: Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition (CVPR) (June 2014) \ No newline at end of file diff --git a/2024/PISR_ Polarimetric Neural Implicit Surface Reconstruction for Textureless and Specular Objects/images.zip b/2024/PISR_ Polarimetric Neural Implicit Surface Reconstruction for Textureless and Specular Objects/images.zip new file mode 100644 index 0000000000000000000000000000000000000000..66aa3d02c8bbb5b068dd5085731cfac00e1f9ba6 --- /dev/null +++ b/2024/PISR_ Polarimetric Neural Implicit Surface Reconstruction for Textureless and Specular Objects/images.zip @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:0a9689cdb071eb10b3d89717b20140096eb75eb544a15adc73471df475f9356a +size 671408 diff --git a/2024/PISR_ Polarimetric Neural Implicit Surface Reconstruction for Textureless and Specular Objects/layout.json b/2024/PISR_ Polarimetric Neural Implicit Surface Reconstruction for Textureless and Specular Objects/layout.json new file mode 100644 index 0000000000000000000000000000000000000000..8f0a3d546e7c1ec964a17e914da1ea9ac22ddd6b --- /dev/null +++ b/2024/PISR_ Polarimetric Neural Implicit Surface Reconstruction for Textureless and Specular Objects/layout.json @@ -0,0 +1,12072 @@ +{ + "pdf_info": [ + { + "para_blocks": [ + { + "bbox": [ + 170, + 111, + 444, + 166 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 170, + 111, + 444, + 166 + ], + "spans": [ + { + "bbox": [ + 170, + 111, + 444, + 166 + ], + "type": "text", + "content": "PISR: Polarimetric Neural Implicit Surface Reconstruction for Textureless and Specular Objects" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 145, + 185, + 468, + 199 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 145, + 185, + 468, + 199 + ], + "spans": [ + { + "bbox": [ + 145, + 185, + 468, + 199 + ], + "type": "text", + "content": "Guangcheng Chen" + }, + { + "bbox": [ + 145, + 185, + 468, + 199 + ], + "type": "inline_equation", + "content": "^{1,2}" + }, + { + "bbox": [ + 145, + 185, + 468, + 199 + ], + "type": "text", + "content": ", Yicheng He" + }, + { + "bbox": [ + 145, + 185, + 468, + 199 + ], + "type": "inline_equation", + "content": "^{1,2}" + }, + { + "bbox": [ + 145, + 185, + 468, + 199 + ], + "type": "text", + "content": ", Li He" + }, + { + "bbox": [ + 145, + 185, + 468, + 199 + ], + "type": "inline_equation", + "content": "^{1,2}" + }, + { + "bbox": [ + 145, + 185, + 468, + 199 + ], + "type": "text", + "content": ", and Hong Zhang" + }, + { + "bbox": [ + 145, + 185, + 468, + 199 + ], + "type": "inline_equation", + "content": "^{\\star 1,2}" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 140, + 207, + 473, + 242 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 140, + 207, + 473, + 242 + ], + "spans": [ + { + "bbox": [ + 140, + 207, + 473, + 242 + ], + "type": "inline_equation", + "content": "^{1}" + }, + { + "bbox": [ + 140, + 207, + 473, + 242 + ], + "type": "text", + "content": " Southern University of Science and Technology, Shenzhen, China " + }, + { + "bbox": [ + 140, + 207, + 473, + 242 + ], + "type": "inline_equation", + "content": "^{2}" + }, + { + "bbox": [ + 140, + 207, + 473, + 242 + ], + "type": "text", + "content": " Shenzhen Key Laboratory of Robotics and Computer Vision {chengc2023,12332136}@mail.sustech.edu.cn, {hel, hzhang}@sustech.edu.cn" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 159, + 270, + 455, + 426 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 159, + 270, + 455, + 426 + ], + "spans": [ + { + "bbox": [ + 159, + 270, + 455, + 426 + ], + "type": "text", + "content": "Abstract. Neural implicit surface reconstruction has achieved remarkable progress recently. Despite resorting to complex radiance modeling, state-of-the-art methods still struggle with textureless and specular surfaces. Different from RGB images, polarization images can provide direct constraints on the azimuth angles of the surface normals. In this paper, we present PISR, a novel method that utilizes a geometrically accurate polarimetric loss to refine shape independently of appearance. In addition, PISR smooths surface normals in image space to eliminate severe shape distortions and leverages the hash-grid-based neural signed distance function to accelerate the reconstruction. Experimental results demonstrate that PISR achieves higher accuracy and robustness, with an L1 Chamfer distance of " + }, + { + "bbox": [ + 159, + 270, + 455, + 426 + ], + "type": "inline_equation", + "content": "0.5\\mathrm{mm}" + }, + { + "bbox": [ + 159, + 270, + 455, + 426 + ], + "type": "text", + "content": " and an F-score of " + }, + { + "bbox": [ + 159, + 270, + 455, + 426 + ], + "type": "inline_equation", + "content": "99.5\\%" + }, + { + "bbox": [ + 159, + 270, + 455, + 426 + ], + "type": "text", + "content": " at " + }, + { + "bbox": [ + 159, + 270, + 455, + 426 + ], + "type": "inline_equation", + "content": "1\\mathrm{mm}" + }, + { + "bbox": [ + 159, + 270, + 455, + 426 + ], + "type": "text", + "content": ", while converging " + }, + { + "bbox": [ + 159, + 270, + 455, + 426 + ], + "type": "inline_equation", + "content": "4 \\sim 30\\times" + }, + { + "bbox": [ + 159, + 270, + 455, + 426 + ], + "type": "text", + "content": " faster than previous polarimetric surface reconstruction methods." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 159, + 435, + 453, + 458 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 159, + 435, + 453, + 458 + ], + "spans": [ + { + "bbox": [ + 159, + 435, + 453, + 458 + ], + "type": "text", + "content": "Keywords: Polarization " + }, + { + "bbox": [ + 159, + 435, + 453, + 458 + ], + "type": "inline_equation", + "content": "\\cdot" + }, + { + "bbox": [ + 159, + 435, + 453, + 458 + ], + "type": "text", + "content": " Surface reconstruction " + }, + { + "bbox": [ + 159, + 435, + 453, + 458 + ], + "type": "inline_equation", + "content": "\\cdot" + }, + { + "bbox": [ + 159, + 435, + 453, + 458 + ], + "type": "text", + "content": " Multi-view reconstruction " + }, + { + "bbox": [ + 159, + 435, + 453, + 458 + ], + "type": "inline_equation", + "content": "\\cdot" + }, + { + "bbox": [ + 159, + 435, + 453, + 458 + ], + "type": "text", + "content": " Neural implicit surface " + }, + { + "bbox": [ + 159, + 435, + 453, + 458 + ], + "type": "inline_equation", + "content": "\\cdot" + }, + { + "bbox": [ + 159, + 435, + 453, + 458 + ], + "type": "text", + "content": " Polarimetric constraint" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 132, + 479, + 230, + 491 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 479, + 230, + 491 + ], + "spans": [ + { + "bbox": [ + 132, + 479, + 230, + 491 + ], + "type": "text", + "content": "1 Introduction" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 130, + 504, + 482, + 636 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 504, + 482, + 636 + ], + "spans": [ + { + "bbox": [ + 130, + 504, + 482, + 636 + ], + "type": "text", + "content": "Surface reconstruction from multi-view images is a fundamental problem in computer vision, with various applications in graphics, robotics, and more. Recently, neural implicit surface has shown powerful flexibility for representing 3D shape and appearance. Combined with differentiable volume rendering [23], neural implicit surface reconstruction is able to estimate 3D shapes from multi-view images accurately [42, 47]. However, these methods struggle with textureless and specular objects due to the shape-radiance ambiguity [52]. Recent methods resort to diffuse and specular separation modeling [40], specular reflection detection [9] and inverse rendering techniques [18, 51]. While these methods show promising results, they may still encounter difficulties since their optimizations rely on image reconstruction losses that tightly couple shape and appearance." + } + ] + } + ], + "index": 6 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 133, + 642, + 423, + 665 + ], + "type": "page_footnote", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 642, + 423, + 665 + ], + "spans": [ + { + "bbox": [ + 133, + 642, + 423, + 665 + ], + "type": "text", + "content": "* Corresponding author. The source code is available at https://github.com/GCChen97/PISR" + } + ] + } + ], + "index": 7 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 0 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 140, + 113, + 236, + 171 + ], + "blocks": [ + { + "bbox": [ + 140, + 113, + 236, + 171 + ], + "lines": [ + { + "bbox": [ + 140, + 113, + 236, + 171 + ], + "spans": [ + { + "bbox": [ + 140, + 113, + 236, + 171 + ], + "type": "image", + "image_path": "6a469094d20c368d430fe3b7fa6a5133e6f628dce87676295cae92f606d652dc.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_body" + } + ], + "index": 2 + }, + { + "type": "image", + "bbox": [ + 238, + 114, + 350, + 171 + ], + "blocks": [ + { + "bbox": [ + 238, + 114, + 350, + 171 + ], + "lines": [ + { + "bbox": [ + 238, + 114, + 350, + 171 + ], + "spans": [ + { + "bbox": [ + 238, + 114, + 350, + 171 + ], + "type": "image", + "image_path": "a9a1be8c9f3aacf37a7af9a94811b3fe6741c52a3798b0a0f0ec1683286ad4ee.jpg" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_body" + } + ], + "index": 3 + }, + { + "type": "image", + "bbox": [ + 353, + 114, + 463, + 171 + ], + "blocks": [ + { + "bbox": [ + 353, + 114, + 463, + 171 + ], + "lines": [ + { + "bbox": [ + 353, + 114, + 463, + 171 + ], + "spans": [ + { + "bbox": [ + 353, + 114, + 463, + 171 + ], + "type": "image", + "image_path": "b43769659760d7d04a02cb3d57e87d681c848a8afa1b14128ddd2de94d077a7f.jpg" + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "image_body" + } + ], + "index": 4 + }, + { + "type": "image", + "bbox": [ + 140, + 172, + 236, + 242 + ], + "blocks": [ + { + "bbox": [ + 140, + 172, + 236, + 242 + ], + "lines": [ + { + "bbox": [ + 140, + 172, + 236, + 242 + ], + "spans": [ + { + "bbox": [ + 140, + 172, + 236, + 242 + ], + "type": "image", + "image_path": "03958391f55e6d912b84c5d3ff6c3c770da3390fad1a49b0bdd521592166caca.jpg" + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 130, + 251, + 480, + 285 + ], + "lines": [ + { + "bbox": [ + 130, + 251, + 480, + 285 + ], + "spans": [ + { + "bbox": [ + 130, + 251, + 480, + 285 + ], + "type": "text", + "content": "Fig. 1: Normal maps of reconstructed surfaces. By leveraging polarization information, our method is able to reconstruct textureless and specular surfaces with fine-grained details." + } + ] + } + ], + "index": 8, + "angle": 0, + "type": "image_caption" + } + ], + "index": 5 + }, + { + "type": "image", + "bbox": [ + 238, + 178, + 350, + 243 + ], + "blocks": [ + { + "bbox": [ + 238, + 178, + 350, + 243 + ], + "lines": [ + { + "bbox": [ + 238, + 178, + 350, + 243 + ], + "spans": [ + { + "bbox": [ + 238, + 178, + 350, + 243 + ], + "type": "image", + "image_path": "79fe8c59d1ab29dcf49fabd5da5546765bf44b592933c83e5e5395b5265f2e31.jpg" + } + ] + } + ], + "index": 6, + "angle": 0, + "type": "image_body" + } + ], + "index": 6 + }, + { + "type": "image", + "bbox": [ + 353, + 178, + 463, + 242 + ], + "blocks": [ + { + "bbox": [ + 353, + 178, + 463, + 242 + ], + "lines": [ + { + "bbox": [ + 353, + 178, + 463, + 242 + ], + "spans": [ + { + "bbox": [ + 353, + 178, + 463, + 242 + ], + "type": "image", + "image_path": "bbe473ac767a597842053e424acc7a6b6c52d8b5791f226911cd83b0038db1ee.jpg" + } + ] + } + ], + "index": 7, + "angle": 0, + "type": "image_body" + } + ], + "index": 7 + }, + { + "bbox": [ + 130, + 319, + 480, + 429 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 319, + 480, + 429 + ], + "spans": [ + { + "bbox": [ + 130, + 319, + 480, + 429 + ], + "type": "text", + "content": "Polarization, as another property of light, provides direct cues on object shape. The angle of polarization of light from object surfaces is independent of light intensity and mainly determined by the surface shapes. From the perspective of a camera, the angle of polarization of a pixel is equal to the azimuth of the surface normal, i.e., the 2D projection direction of the normal onto the image plane, up to a " + }, + { + "bbox": [ + 130, + 319, + 480, + 429 + ], + "type": "inline_equation", + "content": "\\pi/2" + }, + { + "bbox": [ + 130, + 319, + 480, + 429 + ], + "type": "text", + "content": "-ambiguity and a " + }, + { + "bbox": [ + 130, + 319, + 480, + 429 + ], + "type": "inline_equation", + "content": "\\pi" + }, + { + "bbox": [ + 130, + 319, + 480, + 429 + ], + "type": "text", + "content": "-ambiguity [5]. This geometric property has been extensively explored in shape from polarization [1, 2, 25, 49], photo-polarimetric stereo [7, 8, 27, 38], polarimetric multi-view 3D reconstruction [5, 54, 55] and dense mapping [34, 46]." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 130, + 433, + 482, + 564 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 433, + 482, + 564 + ], + "spans": [ + { + "bbox": [ + 130, + 433, + 482, + 564 + ], + "type": "text", + "content": "In this work, we introduce PISR (polarimetric neural implicit surface reconstruction), a novel method that combines the merits of polarimetric 3D reconstruction and neural implicit surface reconstruction for textureless and specular objects, such as ceramics and plastics. Different from recent works [6, 16] that utilize the polarization cues through polarimetric volume rendering, we propose a novel polarimetric loss to refine an object shape directly. Thanks to the integration of the perspective polarimetric constraint [4] in the proposed loss, PISR eliminates artifacts on the reconstructed surfaces, resulting in a " + }, + { + "bbox": [ + 130, + 433, + 482, + 564 + ], + "type": "inline_equation", + "content": "30\\%" + }, + { + "bbox": [ + 130, + 433, + 482, + 564 + ], + "type": "text", + "content": " reduction in L1 Chamfer distance. Besides, instead of treating all possible normals equally [54], our loss adaptively adjusts according to the degree of polarization to address specular-reflection-dominant surfaces, such as ceramics with dark color." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 130, + 582, + 482, + 665 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 582, + 482, + 665 + ], + "spans": [ + { + "bbox": [ + 130, + 582, + 482, + 665 + ], + "type": "text", + "content": "To accelerate the optimization process, PISR uses a multi-resolution hash grid and multi-layer perceptrons (MLP) to represent an object, achieving " + }, + { + "bbox": [ + 130, + 582, + 482, + 665 + ], + "type": "inline_equation", + "content": "4 \\sim 30 \\times" + }, + { + "bbox": [ + 130, + 582, + 482, + 665 + ], + "type": "text", + "content": " speed up as shown in Tab. 3. This representation is also crucial for accurately reconstructing highly reflective surfaces, enabling PISR to rectify the shape topology during optimization, a processing step not available for the mesh-based method, PMVIR [54]. However, the discrete nature of a hash grid also reduces the smoothness of the representation and can lead to the reconstruction" + } + ] + } + ], + "index": 11 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "spans": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "type": "text", + "content": "2" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 166, + 91, + 225, + 100 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 166, + 91, + 225, + 100 + ], + "spans": [ + { + "bbox": [ + 166, + 91, + 225, + 100 + ], + "type": "text", + "content": "G. Chen et al." + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 1 + }, + { + "para_blocks": [ + { + "bbox": [ + 130, + 116, + 479, + 140 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 116, + 479, + 140 + ], + "spans": [ + { + "bbox": [ + 130, + 116, + 479, + 140 + ], + "type": "text", + "content": "results being trapped in local minima. To mitigate the problem, PISR smooths rendered normals for better robustness against poor shape initialization." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 130, + 141, + 482, + 224 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 141, + 482, + 224 + ], + "spans": [ + { + "bbox": [ + 130, + 141, + 482, + 224 + ], + "type": "text", + "content": "Since there is no publicly available dataset for quantitative evaluation of polarimetric 3D reconstruction, we collect a dataset containing textureless and specular objects with a color polarization camera [36]. Using the dataset, we show that PISR is able to improve the surface reconstruction accuracy over the SOTA methods, with an L1 Chamfer distance of " + }, + { + "bbox": [ + 130, + 141, + 482, + 224 + ], + "type": "inline_equation", + "content": "0.5\\mathrm{mm}" + }, + { + "bbox": [ + 130, + 141, + 482, + 224 + ], + "type": "text", + "content": " and an F-score of " + }, + { + "bbox": [ + 130, + 141, + 482, + 224 + ], + "type": "inline_equation", + "content": "99.5\\%" + }, + { + "bbox": [ + 130, + 141, + 482, + 224 + ], + "type": "text", + "content": " at " + }, + { + "bbox": [ + 130, + 141, + 482, + 224 + ], + "type": "inline_equation", + "content": "1\\mathrm{mm}" + }, + { + "bbox": [ + 130, + 141, + 482, + 224 + ], + "type": "text", + "content": ". Shown in Fig. 1 are recovered shapes by different methods and our result shows an accurate shape and fine-grained details." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 146, + 226, + 347, + 237 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 146, + 226, + 347, + 237 + ], + "spans": [ + { + "bbox": [ + 146, + 226, + 347, + 237 + ], + "type": "text", + "content": "In summary, our contributions are as follows:" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 137, + 249, + 479, + 334 + ], + "type": "list", + "angle": 0, + "index": 8, + "blocks": [ + { + "bbox": [ + 138, + 249, + 479, + 284 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 249, + 479, + 284 + ], + "spans": [ + { + "bbox": [ + 138, + 249, + 479, + 284 + ], + "type": "text", + "content": "- PISR, a novel method that leverages polarization cues and hash-grid-based neural SDF for robust and efficient reconstruction of textureless and specular objects." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 137, + 286, + 479, + 309 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 137, + 286, + 479, + 309 + ], + "spans": [ + { + "bbox": [ + 137, + 286, + 479, + 309 + ], + "type": "text", + "content": "- A novel polarimetric loss that utilizes an accurate formulation of the polarimetric constraint for higher surface reconstruction accuracy." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 137, + 311, + 479, + 334 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 137, + 311, + 479, + 334 + ], + "spans": [ + { + "bbox": [ + 137, + 311, + 479, + 334 + ], + "type": "text", + "content": "- A polarization image dataset of real textureless and specular objects with ground-truth meshes for quantitative evaluation." + } + ] + } + ], + "index": 7 + } + ], + "sub_type": "text" + }, + { + "bbox": [ + 132, + 358, + 242, + 371 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 358, + 242, + 371 + ], + "spans": [ + { + "bbox": [ + 132, + 358, + 242, + 371 + ], + "type": "text", + "content": "2 Related Works" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 130, + 388, + 482, + 484 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 388, + 482, + 484 + ], + "spans": [ + { + "bbox": [ + 130, + 388, + 482, + 484 + ], + "type": "text", + "content": "Neural Implicit Surface Reconstruction: The success of neural radiance field reconstruction [23] in novel view synthesis greatly inspires the development of neural implicit surface [28, 30, 48]. Combining with differentiable volume rendering, neural implicit surface reconstruction [29, 42, 47] is able to reconstruct implicit surfaces from dense multi-view images while showing excellent performance in reconstructing fine-grained details. Since then, several works have improved it in reconstruction quality [50], sparse-view performance [19] and optimization speed [43]." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 130, + 485, + 482, + 568 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 485, + 482, + 568 + ], + "spans": [ + { + "bbox": [ + 130, + 485, + 482, + 568 + ], + "type": "text", + "content": "Yet textureless and specular surfaces are still challenging for neural implicit surface reconstruction. To address the problem, NeRO [18] incorporates inverse rendering to alleviate the shape-radiance ambiguity [52] induced by specular reflection. Ref-NeuS [9] borrows the idea of anomaly detection to decrease the weights of specular pixels during optimization and conditions the radiance on reflection direction in reflective scenes as in Ref-NeRF [39]. UniSDF [40] models diffuse and specular radiances in two MLPs separately." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 130, + 570, + 482, + 665 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 570, + 482, + 665 + ], + "spans": [ + { + "bbox": [ + 130, + 570, + 482, + 665 + ], + "type": "text", + "content": "Estimating a 3D radiance field and an SDF from 2D images is a highly ill-posed problem. During optimization, these methods may distort shapes to fit the reflection colors, causing severe reconstruction errors, or, as is commonly referred to, the shape-radiance ambiguity [52]. In contrast, polarization images provide linear constraints on surface normal by the angle of polarization. Therefore, to avoid the shape-radiance ambiguity, our method utilizes the geometry cues from polarization to regularize object shapes independently of appearance, keeping the shape from distortions induced by the visually distorted appearance." + } + ] + } + ], + "index": 12 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 272, + 91, + 448, + 102 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 272, + 91, + 448, + 102 + ], + "spans": [ + { + "bbox": [ + 272, + 91, + 448, + 102 + ], + "type": "text", + "content": "PISR for Textureless and Specular Objects" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 474, + 92, + 480, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 474, + 92, + 480, + 100 + ], + "spans": [ + { + "bbox": [ + 474, + 92, + 480, + 100 + ], + "type": "text", + "content": "3" + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 2 + }, + { + "para_blocks": [ + { + "bbox": [ + 130, + 115, + 482, + 247 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 115, + 482, + 247 + ], + "spans": [ + { + "bbox": [ + 130, + 115, + 482, + 247 + ], + "type": "text", + "content": "Polarimetric 3D Reconstruction: Shape from polarization (SfP) [1,25,31,49] and photo-polarimetric stereo [7,8,27,38] are single-view model-based methods that utilize polarization images for normal and depth estimation. These methods rely on assumptions of either spotlighting or distant lighting to explicitly model the relationship between the degree of polarization and the zenith angle of a surface normal, making them difficult to work under natural lighting conditions. Several deep learning-based methods [2,21,35] leverage data priors to enhance the robustness to more complex indoor lighting conditions. SfPW [15] and DPS-Net [37] learn to estimate scene-level normal maps and depth maps. Yet their generalization is still limited by the relatively small amount of training datasets of polarization images." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 130, + 248, + 482, + 392 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 248, + 482, + 392 + ], + "spans": [ + { + "bbox": [ + 130, + 248, + 482, + 392 + ], + "type": "text", + "content": "Polarimetric multi-view 3D reconstruction leverages multi-view polarimetric constraints to adapt to natural lighting conditions. The polarimetric constraint is established on the fundamental fact that the azimuth angle of a surface normal is equal to the angle of polarization up to a " + }, + { + "bbox": [ + 130, + 248, + 482, + 392 + ], + "type": "inline_equation", + "content": "\\pi/2" + }, + { + "bbox": [ + 130, + 248, + 482, + 392 + ], + "type": "text", + "content": "-ambiguity and a " + }, + { + "bbox": [ + 130, + 248, + 482, + 392 + ], + "type": "inline_equation", + "content": "\\pi" + }, + { + "bbox": [ + 130, + 248, + 482, + 392 + ], + "type": "text", + "content": "-ambiguity under natural lighting [5, 54], which is relatively independent of lighting [5]. Miyazaki et al. [24] uses multi-view polarimetric constraints to reconstruct black specular objects. PolMVS [5] proposes to propagate the depths of sparse points and refine the depth maps with this constraint. Inspired by PolMVS, PDMS [46] and PMDM [34] integrate this constraint into SLAM for dense mapping. PMVIR [54] proposes a framework to simultaneously refine coarse meshes and estimate illuminations and surface albedos. PolarPMS [55] combines this constraint with photo consistency to estimate depth maps from multi-view polarization images." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 130, + 392, + 482, + 477 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 392, + 482, + 477 + ], + "spans": [ + { + "bbox": [ + 130, + 392, + 482, + 477 + ], + "type": "text", + "content": "More closely related to our approach is PMVIR [54]. PMVIR represents shapes using meshes and thereby has limitations in handling incorrect shape topology. In contrast, by leveraging neural SDF, our method is able to rectify incorrect shape topology during the optimization. More importantly, our method utilizes the perspective polarimetric constraint [4] to consider the camera ray directions in the polarimetric loss and thus achieves further performance improvement." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 130, + 497, + 482, + 568 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 497, + 482, + 568 + ], + "spans": [ + { + "bbox": [ + 130, + 497, + 482, + 568 + ], + "type": "text", + "content": "Polarimetric Inverse Rendering: The development of polarimetric bidirectional reflectance distribution function (pBRDF) [3, 11, 12, 14] has enabled polarimetric inverse rendering. Based on mesh representation, sparse ellipsometry [11] leverages polarimetric inverse rendering to simultaneously estimate object meshes and materials using polarization images captured under flashlight illumination." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 130, + 570, + 482, + 665 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 570, + 482, + 665 + ], + "spans": [ + { + "bbox": [ + 130, + 570, + 482, + 665 + ], + "type": "text", + "content": "With the emergence of neural fields in 3D reconstruction, a few recent works [6, 16] have integrated polarization into neural implicit surfaces through polarization image rendering. Concurrent work NeISF [16] introduces the incident Stokes field to represent multi-path polarized light and reparameterizes pBRDF using MLPs, achieving highly accurate geometry reconstruction. PANDORA [6] aims to decompose diffuse and specular reflections using polarization as auxiliary cues, relying on the Fresnel equations to calculate the diffuse and specular components of radiance." + } + ] + } + ], + "index": 6 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "spans": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "type": "text", + "content": "4" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 166, + 91, + 225, + 101 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 166, + 91, + 225, + 101 + ], + "spans": [ + { + "bbox": [ + 166, + 91, + 225, + 101 + ], + "type": "text", + "content": "G. Chen et al." + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 3 + }, + { + "para_blocks": [ + { + "bbox": [ + 130, + 116, + 479, + 200 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 116, + 479, + 200 + ], + "spans": [ + { + "bbox": [ + 130, + 116, + 479, + 200 + ], + "type": "text", + "content": "Although polarimetric inverse rendering excels at accurate material and illumination estimation, it requires proper geometry initialization to decouple shape and appearance. As a result, it may still suffer from the shape-radiance ambiguity, especially when dealing with textureless and specular surfaces. To avoid this problem, our method utilizes the polarimetric constraint to regularize shape directly, independent of appearance. The surface reconstruction results of our method can be used to initialize polarimetric inverse rendering." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 132, + 222, + 224, + 236 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 222, + 224, + 236 + ], + "spans": [ + { + "bbox": [ + 132, + 222, + 224, + 236 + ], + "type": "text", + "content": "3 Preliminary" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 130, + 251, + 479, + 299 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 251, + 479, + 299 + ], + "spans": [ + { + "bbox": [ + 130, + 251, + 479, + 299 + ], + "type": "text", + "content": "Polarimetric Constraint: Assuming four synchronized and aligned images captured at polarizer angles of 0, 45, 90, 135 degrees [36], denoted as " + }, + { + "bbox": [ + 130, + 251, + 479, + 299 + ], + "type": "inline_equation", + "content": "\\mathbf{I}_0" + }, + { + "bbox": [ + 130, + 251, + 479, + 299 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 130, + 251, + 479, + 299 + ], + "type": "inline_equation", + "content": "\\mathbf{I}_{45}" + }, + { + "bbox": [ + 130, + 251, + 479, + 299 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 130, + 251, + 479, + 299 + ], + "type": "inline_equation", + "content": "\\mathbf{I}_{90}" + }, + { + "bbox": [ + 130, + 251, + 479, + 299 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 130, + 251, + 479, + 299 + ], + "type": "inline_equation", + "content": "\\mathbf{I}_{135}" + }, + { + "bbox": [ + 130, + 251, + 479, + 299 + ], + "type": "text", + "content": ". For each pixel, the angle of polarization (AoP) " + }, + { + "bbox": [ + 130, + 251, + 479, + 299 + ], + "type": "inline_equation", + "content": "\\varphi \\in [-\\pi /2,\\pi /2]" + }, + { + "bbox": [ + 130, + 251, + 479, + 299 + ], + "type": "text", + "content": " and degree of polarization (DoP) " + }, + { + "bbox": [ + 130, + 251, + 479, + 299 + ], + "type": "inline_equation", + "content": "\\rho \\in [0,1]" + }, + { + "bbox": [ + 130, + 251, + 479, + 299 + ], + "type": "text", + "content": " are calculated as:" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 222, + 310, + 480, + 336 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 222, + 310, + 480, + 336 + ], + "spans": [ + { + "bbox": [ + 222, + 310, + 480, + 336 + ], + "type": "interline_equation", + "content": "\\varphi = \\frac {1}{2} \\arctan 2 \\left(s _ {2}, s _ {1}\\right), \\quad \\rho = \\frac {\\sqrt {s _ {1} ^ {2} + s _ {2} ^ {2}}}{s _ {0}} \\tag {1}", + "image_path": "1de4175c3ce27e2e490b3f5116ceb4e6c38e39c60e6861e4f3c75bd4722d9b77.jpg" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 130, + 346, + 479, + 394 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 346, + 479, + 394 + ], + "spans": [ + { + "bbox": [ + 130, + 346, + 479, + 394 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 130, + 346, + 479, + 394 + ], + "type": "inline_equation", + "content": "\\mathbf{s} = [s_0, s_1, s_2]^T = \\left[\\frac{1}{2}\\sum I_i, I_0 - I_{90}, I_{45} - I_{135}\\right]^T" + }, + { + "bbox": [ + 130, + 346, + 479, + 394 + ], + "type": "text", + "content": " is the Stokes vector for representing the polarization state of light. Given the estimated normal " + }, + { + "bbox": [ + 130, + 346, + 479, + 394 + ], + "type": "inline_equation", + "content": "\\hat{\\mathbf{n}}" + }, + { + "bbox": [ + 130, + 346, + 479, + 394 + ], + "type": "text", + "content": " in the camera frame and the AoP " + }, + { + "bbox": [ + 130, + 346, + 479, + 394 + ], + "type": "inline_equation", + "content": "\\varphi" + }, + { + "bbox": [ + 130, + 346, + 479, + 394 + ], + "type": "text", + "content": " of a pixel, the orthographic polarimetric constraint on the azimuth angle of the normal is calculated as:" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 230, + 407, + 480, + 418 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 230, + 407, + 480, + 418 + ], + "spans": [ + { + "bbox": [ + 230, + 407, + 480, + 418 + ], + "type": "interline_equation", + "content": "[ \\sin (\\varphi + \\Delta), \\cos (\\varphi + \\Delta), 0 ] \\cdot \\hat {\\mathbf {n}} = 0, \\tag {2}", + "image_path": "629651d577aa0f9bdee682259421518d5a99b2286559ea41b9331683ac438ced.jpg" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 130, + 430, + 479, + 478 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 430, + 479, + 478 + ], + "spans": [ + { + "bbox": [ + 130, + 430, + 479, + 478 + ], + "type": "text", + "content": "where the disambiguation scalar " + }, + { + "bbox": [ + 130, + 430, + 479, + 478 + ], + "type": "inline_equation", + "content": "\\varDelta\\in\\{0,\\pi/2\\}" + }, + { + "bbox": [ + 130, + 430, + 479, + 478 + ], + "type": "text", + "content": " is used for solving the " + }, + { + "bbox": [ + 130, + 430, + 479, + 478 + ], + "type": "inline_equation", + "content": "\\pi /2" + }, + { + "bbox": [ + 130, + 430, + 479, + 478 + ], + "type": "text", + "content": " ambiguity [5]. While this constraint is commonly used in previous works [5,34, 38,46,54], there is a perspective polarimetric constraint which is geometrically more accurate than it by considering the perspective effect of the lens [4]:" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 165, + 489, + 480, + 502 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 165, + 489, + 480, + 502 + ], + "spans": [ + { + "bbox": [ + 165, + 489, + 480, + 502 + ], + "type": "interline_equation", + "content": "\\left[ v _ {z} \\sin \\varphi^ {\\prime}, v _ {z} \\cos \\varphi^ {\\prime}, - \\left(v _ {y} \\cos \\varphi^ {\\prime} + v _ {x} \\sin \\varphi^ {\\prime}\\right) \\right] \\cdot \\hat {\\mathbf {n}} = 0, \\quad \\varphi^ {\\prime} = \\varphi + \\Delta \\tag {3}", + "image_path": "4fc2b5bfaeac4f3dd81d1e28f3ab42d596742e86966defd710cb5f9c34e75217.jpg" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 130, + 511, + 446, + 525 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 511, + 446, + 525 + ], + "spans": [ + { + "bbox": [ + 130, + 511, + 446, + 525 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 130, + 511, + 446, + 525 + ], + "type": "inline_equation", + "content": "\\mathbf{v} = [v_x, v_y, v_z]^T" + }, + { + "bbox": [ + 130, + 511, + 446, + 525 + ], + "type": "text", + "content": " is a normalized camera ray in the camera frame." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 130, + 545, + 480, + 664 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 545, + 480, + 664 + ], + "spans": [ + { + "bbox": [ + 130, + 545, + 480, + 664 + ], + "type": "inline_equation", + "content": "\\pi /2" + }, + { + "bbox": [ + 130, + 545, + 480, + 664 + ], + "type": "text", + "content": " ambiguity: The " + }, + { + "bbox": [ + 130, + 545, + 480, + 664 + ], + "type": "inline_equation", + "content": "\\pi /2" + }, + { + "bbox": [ + 130, + 545, + 480, + 664 + ], + "type": "text", + "content": "-ambiguity arises from the differences between the AoPs of specular reflection and diffuse reflection [5]. According to the Fresnel equations, reflected and transmitted light have distinct energy distributions in the parallel and perpendicular directions to the plane of incident, resulting in a " + }, + { + "bbox": [ + 130, + 545, + 480, + 664 + ], + "type": "inline_equation", + "content": "\\pi /2" + }, + { + "bbox": [ + 130, + 545, + 480, + 664 + ], + "type": "text", + "content": " difference between their AoPs, i.e., " + }, + { + "bbox": [ + 130, + 545, + 480, + 664 + ], + "type": "inline_equation", + "content": "\\Delta = 0" + }, + { + "bbox": [ + 130, + 545, + 480, + 664 + ], + "type": "text", + "content": " for diffuse reflection and " + }, + { + "bbox": [ + 130, + 545, + 480, + 664 + ], + "type": "inline_equation", + "content": "\\Delta = \\pi /2" + }, + { + "bbox": [ + 130, + 545, + 480, + 664 + ], + "type": "text", + "content": " for specular reflection. Under natural lighting conditions, the diffuse reflections and specular reflections are usually mixed, making it difficult to distinguish the dominant component, and thus resulting in the " + }, + { + "bbox": [ + 130, + 545, + 480, + 664 + ], + "type": "inline_equation", + "content": "\\pi /2" + }, + { + "bbox": [ + 130, + 545, + 480, + 664 + ], + "type": "text", + "content": "-ambiguity. Therefore, the " + }, + { + "bbox": [ + 130, + 545, + 480, + 664 + ], + "type": "inline_equation", + "content": "\\pi /2" + }, + { + "bbox": [ + 130, + 545, + 480, + 664 + ], + "type": "text", + "content": "-ambiguity poses a challenge in terms of applying the polarimetric constraints." + } + ] + } + ], + "index": 11 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 272, + 91, + 448, + 102 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 272, + 91, + 448, + 102 + ], + "spans": [ + { + "bbox": [ + 272, + 91, + 448, + 102 + ], + "type": "text", + "content": "PISR for Textureless and Specular Objects" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 474, + 92, + 480, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 474, + 92, + 480, + 100 + ], + "spans": [ + { + "bbox": [ + 474, + 92, + 480, + 100 + ], + "type": "text", + "content": "5" + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 4 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 139, + 113, + 476, + 238 + ], + "blocks": [ + { + "bbox": [ + 139, + 113, + 476, + 238 + ], + "lines": [ + { + "bbox": [ + 139, + 113, + 476, + 238 + ], + "spans": [ + { + "bbox": [ + 139, + 113, + 476, + 238 + ], + "type": "image", + "image_path": "5a2e64f767d0edd25388219e1196d3019d957322d8aa6b5e2a8c01a6422f6e29.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 130, + 247, + 480, + 281 + ], + "lines": [ + { + "bbox": [ + 130, + 247, + 480, + 281 + ], + "spans": [ + { + "bbox": [ + 130, + 247, + 480, + 281 + ], + "type": "text", + "content": "Fig. 2: Overview of PISR. Pixels for the optimization are sampled in criss-cross patterns at the middle stage of the optimization. The polarimetric loss " + }, + { + "bbox": [ + 130, + 247, + 480, + 281 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_{\\mathrm{pol}}^{\\mathrm{p}}" + }, + { + "bbox": [ + 130, + 247, + 480, + 281 + ], + "type": "text", + "content": " and the normal loss " + }, + { + "bbox": [ + 130, + 247, + 480, + 281 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_{\\mathrm{normal}}" + }, + { + "bbox": [ + 130, + 247, + 480, + 281 + ], + "type": "text", + "content": " are used for regularizing the shape independently of appearance." + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_caption" + } + ], + "index": 2 + }, + { + "bbox": [ + 132, + 300, + 202, + 312 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 300, + 202, + 312 + ], + "spans": [ + { + "bbox": [ + 132, + 300, + 202, + 312 + ], + "type": "text", + "content": "4 Method" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 130, + 323, + 482, + 422 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 323, + 482, + 422 + ], + "spans": [ + { + "bbox": [ + 130, + 323, + 482, + 422 + ], + "type": "text", + "content": "PISR aims to recover object shapes from multi-view polarization images and camera poses. Fig. 2 shows an overview of our method. Object shape is represented as a hash-grid-based neural SDF and rendered to images for computing losses (Sec. 4.1). During the reconstruction, the coarse shape of an object is estimated with the photometric loss " + }, + { + "bbox": [ + 130, + 323, + 482, + 422 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_{\\mathrm{color}}" + }, + { + "bbox": [ + 130, + 323, + 482, + 422 + ], + "type": "text", + "content": ". Then the polarimetric loss " + }, + { + "bbox": [ + 130, + 323, + 482, + 422 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_{\\mathrm{pol}}^{\\mathrm{p}}" + }, + { + "bbox": [ + 130, + 323, + 482, + 422 + ], + "type": "text", + "content": " (Sec. 4.2) and the normal loss " + }, + { + "bbox": [ + 130, + 323, + 482, + 422 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_{\\mathrm{normal}}" + }, + { + "bbox": [ + 130, + 323, + 482, + 422 + ], + "type": "text", + "content": " with criss-cross pattern sampling (Sec. 4.3) are jointly used to rectify the shape from distortions. Finally, the shape is refined with " + }, + { + "bbox": [ + 130, + 323, + 482, + 422 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_{\\mathrm{color}}" + }, + { + "bbox": [ + 130, + 323, + 482, + 422 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 130, + 323, + 482, + 422 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_{\\mathrm{pol}}^{\\mathrm{p}}" + }, + { + "bbox": [ + 130, + 323, + 482, + 422 + ], + "type": "text", + "content": ". The optimization scheme is introduced in Sec. 4.4." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 130, + 433, + 390, + 445 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 433, + 390, + 445 + ], + "spans": [ + { + "bbox": [ + 130, + 433, + 390, + 445 + ], + "type": "text", + "content": "4.1 Surface Representation and Image Rendering" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 130, + 450, + 482, + 522 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 450, + 482, + 522 + ], + "spans": [ + { + "bbox": [ + 130, + 450, + 482, + 522 + ], + "type": "text", + "content": "Since Instant-NGP [26], combining a multi-resolution hash grid and MLPs has been commonly used for faster convergence in neural field reconstruction. Following recent works [17, 41, 43], we represent the shape and appearance of an object as an " + }, + { + "bbox": [ + 130, + 450, + 482, + 522 + ], + "type": "inline_equation", + "content": "L" + }, + { + "bbox": [ + 130, + 450, + 482, + 522 + ], + "type": "text", + "content": "-level multi-resolution hash grid " + }, + { + "bbox": [ + 130, + 450, + 482, + 522 + ], + "type": "inline_equation", + "content": "\\mathcal{G} = \\{\\mathcal{G}_l\\}_{l=1}^L" + }, + { + "bbox": [ + 130, + 450, + 482, + 522 + ], + "type": "text", + "content": " with two MLPs " + }, + { + "bbox": [ + 130, + 450, + 482, + 522 + ], + "type": "inline_equation", + "content": "\\varPhi_s" + }, + { + "bbox": [ + 130, + 450, + 482, + 522 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 130, + 450, + 482, + 522 + ], + "type": "inline_equation", + "content": "\\varPhi_{\\mathbf{c}}" + }, + { + "bbox": [ + 130, + 450, + 482, + 522 + ], + "type": "text", + "content": " for decoding SDF values " + }, + { + "bbox": [ + 130, + 450, + 482, + 522 + ], + "type": "inline_equation", + "content": "s" + }, + { + "bbox": [ + 130, + 450, + 482, + 522 + ], + "type": "text", + "content": " and color " + }, + { + "bbox": [ + 130, + 450, + 482, + 522 + ], + "type": "inline_equation", + "content": "\\mathbf{c}" + }, + { + "bbox": [ + 130, + 450, + 482, + 522 + ], + "type": "text", + "content": " respectively. For each 3D space point " + }, + { + "bbox": [ + 130, + 450, + 482, + 522 + ], + "type": "inline_equation", + "content": "\\mathbf{x}" + }, + { + "bbox": [ + 130, + 450, + 482, + 522 + ], + "type": "text", + "content": ", the SDF value " + }, + { + "bbox": [ + 130, + 450, + 482, + 522 + ], + "type": "inline_equation", + "content": "s" + }, + { + "bbox": [ + 130, + 450, + 482, + 522 + ], + "type": "text", + "content": " and color " + }, + { + "bbox": [ + 130, + 450, + 482, + 522 + ], + "type": "inline_equation", + "content": "\\mathbf{c}" + }, + { + "bbox": [ + 130, + 450, + 482, + 522 + ], + "type": "text", + "content": " are computed as" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 224, + 528, + 480, + 541 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 224, + 528, + 480, + 541 + ], + "spans": [ + { + "bbox": [ + 224, + 528, + 480, + 541 + ], + "type": "interline_equation", + "content": "(s, \\mathbf {f}) = \\varPhi_ {s} (\\mathcal {G} (\\mathbf {x})), \\quad \\mathbf {c} = \\varPhi_ {\\mathbf {c}} (s, \\mathbf {f}, \\mathbf {v}, \\mathbf {n}) \\tag {4}", + "image_path": "f5a20245cefe165096e7b9344c8bcc83593db2065b1ea4c99aab649a01c5e4d9.jpg" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 130, + 547, + 482, + 582 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 547, + 482, + 582 + ], + "spans": [ + { + "bbox": [ + 130, + 547, + 482, + 582 + ], + "type": "text", + "content": "with the camera ray " + }, + { + "bbox": [ + 130, + 547, + 482, + 582 + ], + "type": "inline_equation", + "content": "\\mathbf{v}" + }, + { + "bbox": [ + 130, + 547, + 482, + 582 + ], + "type": "text", + "content": " and surface normal " + }, + { + "bbox": [ + 130, + 547, + 482, + 582 + ], + "type": "inline_equation", + "content": "\\mathbf{n}" + }, + { + "bbox": [ + 130, + 547, + 482, + 582 + ], + "type": "text", + "content": ". The normal " + }, + { + "bbox": [ + 130, + 547, + 482, + 582 + ], + "type": "inline_equation", + "content": "\\mathbf{n}" + }, + { + "bbox": [ + 130, + 547, + 482, + 582 + ], + "type": "text", + "content": " can be computed by normalizing the gradient of SDF w.r.t. to " + }, + { + "bbox": [ + 130, + 547, + 482, + 582 + ], + "type": "inline_equation", + "content": "\\mathbf{x}" + }, + { + "bbox": [ + 130, + 547, + 482, + 582 + ], + "type": "text", + "content": " or by finite difference [17]. Besides the foreground object, we use Instant-NGP [26] to represent the background." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 130, + 582, + 482, + 631 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 582, + 482, + 631 + ], + "spans": [ + { + "bbox": [ + 130, + 582, + 482, + 631 + ], + "type": "text", + "content": "To optimize the shape with images, we use differentiable volume rendering [22] to render images given the surface representation. The pixel color " + }, + { + "bbox": [ + 130, + 582, + 482, + 631 + ], + "type": "inline_equation", + "content": "\\hat{\\mathbf{c}}" + }, + { + "bbox": [ + 130, + 582, + 482, + 631 + ], + "type": "text", + "content": " and normal " + }, + { + "bbox": [ + 130, + 582, + 482, + 631 + ], + "type": "inline_equation", + "content": "\\hat{\\mathbf{n}}" + }, + { + "bbox": [ + 130, + 582, + 482, + 631 + ], + "type": "text", + "content": " are computed as the weighted sum of " + }, + { + "bbox": [ + 130, + 582, + 482, + 631 + ], + "type": "inline_equation", + "content": "M" + }, + { + "bbox": [ + 130, + 582, + 482, + 631 + ], + "type": "text", + "content": " samples " + }, + { + "bbox": [ + 130, + 582, + 482, + 631 + ], + "type": "inline_equation", + "content": "\\{\\mathbf{x}_i = \\mathbf{o} + t_i\\mathbf{v}\\}_{i=1}^M" + }, + { + "bbox": [ + 130, + 582, + 482, + 631 + ], + "type": "text", + "content": " on a cast-off pixel ray with the camera position " + }, + { + "bbox": [ + 130, + 582, + 482, + 631 + ], + "type": "inline_equation", + "content": "\\mathbf{o}" + }, + { + "bbox": [ + 130, + 582, + 482, + 631 + ], + "type": "text", + "content": " and different depth " + }, + { + "bbox": [ + 130, + 582, + 482, + 631 + ], + "type": "inline_equation", + "content": "t_i" + }, + { + "bbox": [ + 130, + 582, + 482, + 631 + ], + "type": "text", + "content": ":" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 241, + 635, + 481, + 668 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 241, + 635, + 481, + 668 + ], + "spans": [ + { + "bbox": [ + 241, + 635, + 481, + 668 + ], + "type": "interline_equation", + "content": "\\hat {\\mathbf {c}} = \\sum_ {i = 1} ^ {M} w _ {i} \\mathbf {c} _ {i}, \\quad \\hat {\\mathbf {n}} = \\sum_ {i = 1} ^ {M} w _ {i} \\mathbf {n} _ {i}, \\tag {5}", + "image_path": "dd6819cc694f373d746c2c29911985b1c7cae5618b2d8d5cd37fe6493df0a7d1.jpg" + } + ] + } + ], + "index": 11 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "spans": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "type": "text", + "content": "6" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 166, + 91, + 225, + 100 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 166, + 91, + 225, + 100 + ], + "spans": [ + { + "bbox": [ + 166, + 91, + 225, + 100 + ], + "type": "text", + "content": "G. Chen et al." + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 5 + }, + { + "para_blocks": [ + { + "bbox": [ + 130, + 116, + 479, + 152 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 116, + 479, + 152 + ], + "spans": [ + { + "bbox": [ + 130, + 116, + 479, + 152 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 130, + 116, + 479, + 152 + ], + "type": "inline_equation", + "content": "w_{i}" + }, + { + "bbox": [ + 130, + 116, + 479, + 152 + ], + "type": "text", + "content": " is the blending weight converted from SDF values using the unbiased weight function of NeuS [42], and " + }, + { + "bbox": [ + 130, + 116, + 479, + 152 + ], + "type": "inline_equation", + "content": "\\mathbf{c}_i" + }, + { + "bbox": [ + 130, + 116, + 479, + 152 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 130, + 116, + 479, + 152 + ], + "type": "inline_equation", + "content": "\\mathbf{n}_i" + }, + { + "bbox": [ + 130, + 116, + 479, + 152 + ], + "type": "text", + "content": " are the color and normal of the sample " + }, + { + "bbox": [ + 130, + 116, + 479, + 152 + ], + "type": "inline_equation", + "content": "\\mathbf{x}_i" + }, + { + "bbox": [ + 130, + 116, + 479, + 152 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 132, + 171, + 346, + 182 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 171, + 346, + 182 + ], + "spans": [ + { + "bbox": [ + 132, + 171, + 346, + 182 + ], + "type": "text", + "content": "4.2 Polarimetric and Photometric Losses" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 130, + 193, + 479, + 229 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 193, + 479, + 229 + ], + "spans": [ + { + "bbox": [ + 130, + 193, + 479, + 229 + ], + "type": "text", + "content": "Polarimetric Loss: Previous works [5,34,38,46,54] leverage polarization cues with the following loss function " + }, + { + "bbox": [ + 130, + 193, + 479, + 229 + ], + "type": "inline_equation", + "content": "h^{\\mathrm{o}}" + }, + { + "bbox": [ + 130, + 193, + 479, + 229 + ], + "type": "text", + "content": " that is derived from the orthographic polarimetric constraint Eq. (2):" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 143, + 239, + 480, + 261 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 143, + 239, + 480, + 261 + ], + "spans": [ + { + "bbox": [ + 143, + 239, + 480, + 261 + ], + "type": "interline_equation", + "content": "h ^ {\\mathrm {o}} (\\varphi , \\Delta , \\mathbf {n}) = \\left(\\mathbf {a} ^ {\\mathrm {o}} (\\varphi , \\Delta) \\cdot \\mathbf {n}\\right) ^ {2}, \\mathbf {a} ^ {\\mathrm {o}} (\\varphi , \\Delta) = [ \\sin (\\varphi + \\Delta), \\cos (\\varphi + \\Delta), 0 ], \\tag {6}", + "image_path": "7fb0fe87dceb271b5489d9ea794f0bf3a565c6cd637d74a86635fb74a614442f.jpg" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 130, + 269, + 479, + 304 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 269, + 479, + 304 + ], + "spans": [ + { + "bbox": [ + 130, + 269, + 479, + 304 + ], + "type": "text", + "content": "where the coefficient vector " + }, + { + "bbox": [ + 130, + 269, + 479, + 304 + ], + "type": "inline_equation", + "content": "\\mathbf{a}^{\\mathrm{o}}" + }, + { + "bbox": [ + 130, + 269, + 479, + 304 + ], + "type": "text", + "content": " is calculated with the AoP " + }, + { + "bbox": [ + 130, + 269, + 479, + 304 + ], + "type": "inline_equation", + "content": "\\varphi" + }, + { + "bbox": [ + 130, + 269, + 479, + 304 + ], + "type": "text", + "content": " and the disambiguation scalar " + }, + { + "bbox": [ + 130, + 269, + 479, + 304 + ], + "type": "inline_equation", + "content": "\\Delta" + }, + { + "bbox": [ + 130, + 269, + 479, + 304 + ], + "type": "text", + "content": ". However, as discussed in [4], Eq. (2) neglects the perspective effect of the lens and may introduce distortions in the reconstruction results." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 131, + 305, + 479, + 330 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 131, + 305, + 479, + 330 + ], + "spans": [ + { + "bbox": [ + 131, + 305, + 479, + 330 + ], + "type": "text", + "content": "Therefore, we instead adopt the perspective polarimetric constraint Eq. (3) [4] and propose a novel loss function " + }, + { + "bbox": [ + 131, + 305, + 479, + 330 + ], + "type": "inline_equation", + "content": "h^{\\mathrm{p}}" + }, + { + "bbox": [ + 131, + 305, + 479, + 330 + ], + "type": "text", + "content": ":" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 230, + 339, + 480, + 365 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 230, + 339, + 480, + 365 + ], + "spans": [ + { + "bbox": [ + 230, + 339, + 480, + 365 + ], + "type": "interline_equation", + "content": "h ^ {\\mathrm {p}} (\\varphi , \\Delta , \\mathbf {v}, \\mathbf {n}) = \\left(\\frac {\\mathbf {a} ^ {\\mathrm {p}} (\\varphi , \\Delta , \\mathbf {v}) \\cdot \\mathbf {n}}{\\| \\mathbf {a} ^ {\\mathrm {p}} (\\varphi , \\Delta , \\mathbf {v}) \\|}\\right) ^ {2} \\tag {7}", + "image_path": "93419d6d4df70a3bf8a7afe6e823203aaa4e29fd65383de6e35d5f12a317b120.jpg" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 131, + 373, + 230, + 384 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 131, + 373, + 230, + 384 + ], + "spans": [ + { + "bbox": [ + 131, + 373, + 230, + 384 + ], + "type": "text", + "content": "with " + }, + { + "bbox": [ + 131, + 373, + 230, + 384 + ], + "type": "inline_equation", + "content": "\\mathbf{a}^{\\mathrm{p}}" + }, + { + "bbox": [ + 131, + 373, + 230, + 384 + ], + "type": "text", + "content": " calculated as:" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 145, + 396, + 480, + 408 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 145, + 396, + 480, + 408 + ], + "spans": [ + { + "bbox": [ + 145, + 396, + 480, + 408 + ], + "type": "interline_equation", + "content": "\\mathbf {a} ^ {\\mathrm {p}} (\\varphi , \\Delta , \\mathbf {v}) = \\left[ v _ {z} \\sin \\varphi^ {\\prime}, v _ {z} \\cos \\varphi^ {\\prime}, - \\left(v _ {y} \\cos \\varphi^ {\\prime} + v _ {x} \\sin \\varphi^ {\\prime}\\right) \\right], \\varphi^ {\\prime} = \\varphi + \\Delta . \\tag {8}", + "image_path": "ae37bd5f60e9c4800e2107cd476ac008de9b7c30c1d7dfb36dc02a8db147c4c8.jpg" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 130, + 418, + 479, + 453 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 418, + 479, + 453 + ], + "spans": [ + { + "bbox": [ + 130, + 418, + 479, + 453 + ], + "type": "text", + "content": "The loss function " + }, + { + "bbox": [ + 130, + 418, + 479, + 453 + ], + "type": "inline_equation", + "content": "h^{\\mathrm{p}}" + }, + { + "bbox": [ + 130, + 418, + 479, + 453 + ], + "type": "text", + "content": " additionally considers the camera ray direction " + }, + { + "bbox": [ + 130, + 418, + 479, + 453 + ], + "type": "inline_equation", + "content": "\\mathbf{v}" + }, + { + "bbox": [ + 130, + 418, + 479, + 453 + ], + "type": "text", + "content": " to model the perspective effect. In Sec. 5.3, we will show that " + }, + { + "bbox": [ + 130, + 418, + 479, + 453 + ], + "type": "inline_equation", + "content": "h^{\\mathrm{p}}" + }, + { + "bbox": [ + 130, + 418, + 479, + 453 + ], + "type": "text", + "content": " effectively improves the reconstruction accuracy." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 130, + 454, + 479, + 490 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 454, + 479, + 490 + ], + "spans": [ + { + "bbox": [ + 130, + 454, + 479, + 490 + ], + "type": "text", + "content": "Besides, the " + }, + { + "bbox": [ + 130, + 454, + 479, + 490 + ], + "type": "inline_equation", + "content": "\\pi/2" + }, + { + "bbox": [ + 130, + 454, + 479, + 490 + ], + "type": "text", + "content": "-ambiguity must be addressed when applying the polarimetric constraint. We combine the merits of PolMVS [5] and PMVIR [54], proposing the segmented loss function " + }, + { + "bbox": [ + 130, + 454, + 479, + 490 + ], + "type": "inline_equation", + "content": "f^{\\mathrm{p}}" + }, + { + "bbox": [ + 130, + 454, + 479, + 490 + ], + "type": "text", + "content": " as follows:" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 180, + 500, + 480, + 537 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 180, + 500, + 480, + 537 + ], + "spans": [ + { + "bbox": [ + 180, + 500, + 480, + 537 + ], + "type": "interline_equation", + "content": "f ^ {\\mathrm {p}} (\\varphi , \\mathbf {v}, \\mathbf {n}, \\rho) = \\left\\{ \\begin{array}{l l} h ^ {\\mathrm {p}} (\\varphi , 0, \\mathbf {v}, \\mathbf {n}) \\cdot h ^ {\\mathrm {p}} (\\varphi , \\pi / 2, \\mathbf {v}, \\mathbf {n}) & \\text {i f} \\rho < \\theta , \\\\ h ^ {\\mathrm {p}} (\\varphi , \\pi / 2, \\mathbf {v}, \\mathbf {n}) & \\text {i f} \\rho \\geq \\theta , \\end{array} \\right. \\tag {9}", + "image_path": "b5ed5f389635061c3a1e8e503a9aa02939877f3c484d4bad5cdb9e7f8961c512.jpg" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 130, + 545, + 480, + 616 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 545, + 480, + 616 + ], + "spans": [ + { + "bbox": [ + 130, + 545, + 480, + 616 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 130, + 545, + 480, + 616 + ], + "type": "inline_equation", + "content": "\\rho" + }, + { + "bbox": [ + 130, + 545, + 480, + 616 + ], + "type": "text", + "content": " is the DoP and " + }, + { + "bbox": [ + 130, + 545, + 480, + 616 + ], + "type": "inline_equation", + "content": "\\theta" + }, + { + "bbox": [ + 130, + 545, + 480, + 616 + ], + "type": "text", + "content": " is a preset threshold. This design is based on two observations: 1) the reflected light of an object is usually a mixture of the diffuse and specular components, and thus the DoP is relatively low [5], 2) only the DoP of specular reflection of dielectric objects can exceed 0.4 according to Fresnel equations. We use multiplication instead of addition when " + }, + { + "bbox": [ + 130, + 545, + 480, + 616 + ], + "type": "inline_equation", + "content": "\\rho < \\theta" + }, + { + "bbox": [ + 130, + 545, + 480, + 616 + ], + "type": "text", + "content": " because only one of the two needs to be minimized." + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 130, + 618, + 479, + 639 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 618, + 479, + 639 + ], + "spans": [ + { + "bbox": [ + 130, + 618, + 479, + 639 + ], + "type": "text", + "content": "Finally, we calculate the polarimetric loss for a batch of sampled pixels " + }, + { + "bbox": [ + 130, + 618, + 479, + 639 + ], + "type": "inline_equation", + "content": "S" + }, + { + "bbox": [ + 130, + 618, + 479, + 639 + ], + "type": "text", + "content": " as follows:" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 235, + 640, + 480, + 667 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 235, + 640, + 480, + 667 + ], + "spans": [ + { + "bbox": [ + 235, + 640, + 480, + 667 + ], + "type": "interline_equation", + "content": "\\mathcal {L} _ {\\mathrm {p o l}} ^ {\\mathrm {p}} = \\frac {1}{| S |} \\sum_ {u \\in \\mathcal {S}} f ^ {\\mathrm {p}} \\left(\\varphi_ {u}, \\mathbf {v} _ {u}, \\hat {\\mathbf {n}} _ {u}, \\rho_ {u}\\right). \\tag {10}", + "image_path": "3fe0c0729812898602c048e7b08716ceae6937b38cd601b4941159d98822a849.jpg" + } + ] + } + ], + "index": 16 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 272, + 91, + 448, + 102 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 272, + 91, + 448, + 102 + ], + "spans": [ + { + "bbox": [ + 272, + 91, + 448, + 102 + ], + "type": "text", + "content": "PISR for Textureless and Specular Objects" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 474, + 92, + 480, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 474, + 92, + 480, + 100 + ], + "spans": [ + { + "bbox": [ + 474, + 92, + 480, + 100 + ], + "type": "text", + "content": "7" + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 6 + }, + { + "para_blocks": [ + { + "bbox": [ + 130, + 116, + 479, + 163 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 116, + 479, + 163 + ], + "spans": [ + { + "bbox": [ + 130, + 116, + 479, + 163 + ], + "type": "text", + "content": "Note that the camera ray " + }, + { + "bbox": [ + 130, + 116, + 479, + 163 + ], + "type": "inline_equation", + "content": "\\mathbf{v}_u" + }, + { + "bbox": [ + 130, + 116, + 479, + 163 + ], + "type": "text", + "content": " and normal " + }, + { + "bbox": [ + 130, + 116, + 479, + 163 + ], + "type": "inline_equation", + "content": "\\hat{\\mathbf{n}}_u" + }, + { + "bbox": [ + 130, + 116, + 479, + 163 + ], + "type": "text", + "content": " are transformed into the camera frame according to the camera pose. In Sec. 5.3, the experiments will demonstrate that calculating the loss " + }, + { + "bbox": [ + 130, + 116, + 479, + 163 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_{\\mathrm{pol}}^{\\mathrm{p}}" + }, + { + "bbox": [ + 130, + 116, + 479, + 163 + ], + "type": "text", + "content": " with " + }, + { + "bbox": [ + 130, + 116, + 479, + 163 + ], + "type": "inline_equation", + "content": "h^{\\mathrm{p}}" + }, + { + "bbox": [ + 130, + 116, + 479, + 163 + ], + "type": "text", + "content": " instead of " + }, + { + "bbox": [ + 130, + 116, + 479, + 163 + ], + "type": "inline_equation", + "content": "h^{\\mathrm{o}}" + }, + { + "bbox": [ + 130, + 116, + 479, + 163 + ], + "type": "text", + "content": " can reduce the L1 Chamfer distance by " + }, + { + "bbox": [ + 130, + 116, + 479, + 163 + ], + "type": "inline_equation", + "content": "30\\%" + }, + { + "bbox": [ + 130, + 116, + 479, + 163 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 130, + 167, + 479, + 202 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 167, + 479, + 202 + ], + "spans": [ + { + "bbox": [ + 130, + 167, + 479, + 202 + ], + "type": "text", + "content": "Photometric Loss: Since the polarimetric loss alone is insufficient for recovering object shapes due to the " + }, + { + "bbox": [ + 130, + 167, + 479, + 202 + ], + "type": "inline_equation", + "content": "\\pi /2" + }, + { + "bbox": [ + 130, + 167, + 479, + 202 + ], + "type": "text", + "content": " ambiguity and image noises, we used the photometric loss for initializing and maintaining the basic shape:" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 250, + 209, + 481, + 236 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 250, + 209, + 481, + 236 + ], + "spans": [ + { + "bbox": [ + 250, + 209, + 481, + 236 + ], + "type": "interline_equation", + "content": "\\mathcal {L} _ {\\text {c o l o r}} = \\frac {1}{| \\mathcal {S} |} \\sum_ {u \\in \\mathcal {S}} \\left| \\hat {\\mathbf {c}} _ {u} - \\mathbf {c} _ {u} \\right|. \\tag {11}", + "image_path": "3096cf80a1b971d59ea09d46b5516b223d08860e69da52f12223eb82639d4361.jpg" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 132, + 251, + 276, + 264 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 251, + 276, + 264 + ], + "spans": [ + { + "bbox": [ + 132, + 251, + 276, + 264 + ], + "type": "text", + "content": "4.3 Normal Regularization" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 130, + 269, + 480, + 365 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 269, + 480, + 365 + ], + "spans": [ + { + "bbox": [ + 130, + 269, + 480, + 365 + ], + "type": "text", + "content": "The discrete nature of the hash grid makes the estimated SDF less smooth than those purely MLP-based, which can introduce reconstruction artifacts such as holes or cracks. Directly smoothing the normals in 3D space as in [17, 53] can lead to a bumpy surface because the sampled normals are solely constrained by the smoothness loss. Inspired by the belief propagation used in traditional MVS methods [33, 45, 56], we smooth the rendered normals so that all the samples are also constrained by both the photometric loss and the polarimetric loss. We therefore utilize the following normal loss:" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 211, + 371, + 481, + 400 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 211, + 371, + 481, + 400 + ], + "spans": [ + { + "bbox": [ + 211, + 371, + 481, + 400 + ], + "type": "interline_equation", + "content": "\\mathcal {L} _ {\\text {n o r m a l}} = - \\frac {1}{| \\mathcal {S} _ {\\mathrm {c}} |} \\sum_ {u \\in \\mathcal {S} _ {\\mathrm {c}}} \\frac {1}{| \\mathcal {N} _ {u} |} \\sum_ {\\hat {\\mathbf {n}} _ {j} \\in \\mathcal {N} _ {u}} \\operatorname {S G} \\left(\\hat {\\mathbf {n}} _ {u}\\right) ^ {T} \\hat {\\mathbf {n}} _ {j} \\tag {12}", + "image_path": "ddb07c4056f90140813aa460306806582386953cac249e528de18158f4de6dbb.jpg" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 130, + 406, + 479, + 465 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 406, + 479, + 465 + ], + "spans": [ + { + "bbox": [ + 130, + 406, + 479, + 465 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 130, + 406, + 479, + 465 + ], + "type": "inline_equation", + "content": "S_{c}" + }, + { + "bbox": [ + 130, + 406, + 479, + 465 + ], + "type": "text", + "content": " is a set of pixels sampled among pixels of all the images, " + }, + { + "bbox": [ + 130, + 406, + 479, + 465 + ], + "type": "inline_equation", + "content": "\\mathcal{N}_u" + }, + { + "bbox": [ + 130, + 406, + 479, + 465 + ], + "type": "text", + "content": " is a set of pixel normals sampled around " + }, + { + "bbox": [ + 130, + 406, + 479, + 465 + ], + "type": "inline_equation", + "content": "\\hat{\\mathbf{n}}_u" + }, + { + "bbox": [ + 130, + 406, + 479, + 465 + ], + "type": "text", + "content": " in the image space. The function SG (\"Stop Gradient\") is used to stop propagating the gradients of the losses w.r.t. " + }, + { + "bbox": [ + 130, + 406, + 479, + 465 + ], + "type": "inline_equation", + "content": "\\hat{\\mathbf{n}}_u" + }, + { + "bbox": [ + 130, + 406, + 479, + 465 + ], + "type": "text", + "content": " during backpropagation, i.e., fix the values of " + }, + { + "bbox": [ + 130, + 406, + 479, + 465 + ], + "type": "inline_equation", + "content": "\\hat{\\mathbf{n}}_u" + }, + { + "bbox": [ + 130, + 406, + 479, + 465 + ], + "type": "text", + "content": " so that it can be propagated to neighbor pixels." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 130, + 466, + 479, + 491 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 466, + 479, + 491 + ], + "spans": [ + { + "bbox": [ + 130, + 466, + 479, + 491 + ], + "type": "text", + "content": "Now the pixel sample set " + }, + { + "bbox": [ + 130, + 466, + 479, + 491 + ], + "type": "inline_equation", + "content": "S" + }, + { + "bbox": [ + 130, + 466, + 479, + 491 + ], + "type": "text", + "content": " of each iteration during the optimization is the union of the set of center pixels " + }, + { + "bbox": [ + 130, + 466, + 479, + 491 + ], + "type": "inline_equation", + "content": "S_{c}" + }, + { + "bbox": [ + 130, + 466, + 479, + 491 + ], + "type": "text", + "content": " and the sets of neighbor pixels " + }, + { + "bbox": [ + 130, + 466, + 479, + 491 + ], + "type": "inline_equation", + "content": "\\mathcal{N}_u" + }, + { + "bbox": [ + 130, + 466, + 479, + 491 + ], + "type": "text", + "content": ":" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 264, + 498, + 481, + 522 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 264, + 498, + 481, + 522 + ], + "spans": [ + { + "bbox": [ + 264, + 498, + 481, + 522 + ], + "type": "interline_equation", + "content": "\\mathcal {S} = \\mathcal {S} _ {c} \\cup \\left\\{\\bigcup_ {u \\in S _ {c}} \\mathcal {N} _ {u} \\right\\} \\tag {13}", + "image_path": "5dddeacd46cbfc8e22e46be64af19fdb0364c47f444600fd1f0f4c412c9ea638.jpg" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 130, + 528, + 480, + 600 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 528, + 480, + 600 + ], + "spans": [ + { + "bbox": [ + 130, + 528, + 480, + 600 + ], + "type": "text", + "content": "Empirically, larger " + }, + { + "bbox": [ + 130, + 528, + 480, + 600 + ], + "type": "inline_equation", + "content": "\\mathcal{S}_c" + }, + { + "bbox": [ + 130, + 528, + 480, + 600 + ], + "type": "text", + "content": " can lead to better results because it helps capture global structures, which are important for optimizing a neural field [44], while larger " + }, + { + "bbox": [ + 130, + 528, + 480, + 600 + ], + "type": "inline_equation", + "content": "\\mathcal{N}_u" + }, + { + "bbox": [ + 130, + 528, + 480, + 600 + ], + "type": "text", + "content": " is beneficial to smooth the surface. Therefore, we sample " + }, + { + "bbox": [ + 130, + 528, + 480, + 600 + ], + "type": "inline_equation", + "content": "\\mathcal{N}_u" + }, + { + "bbox": [ + 130, + 528, + 480, + 600 + ], + "type": "text", + "content": " in a criss-cross pattern instead of a whole patch around " + }, + { + "bbox": [ + 130, + 528, + 480, + 600 + ], + "type": "inline_equation", + "content": "\\hat{\\mathbf{n}}_u" + }, + { + "bbox": [ + 130, + 528, + 480, + 600 + ], + "type": "text", + "content": " to balance " + }, + { + "bbox": [ + 130, + 528, + 480, + 600 + ], + "type": "inline_equation", + "content": "|\\mathcal{S}_c|" + }, + { + "bbox": [ + 130, + 528, + 480, + 600 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 130, + 528, + 480, + 600 + ], + "type": "inline_equation", + "content": "|\\mathcal{N}_u|" + }, + { + "bbox": [ + 130, + 528, + 480, + 600 + ], + "type": "text", + "content": " with fixed " + }, + { + "bbox": [ + 130, + 528, + 480, + 600 + ], + "type": "inline_equation", + "content": "|S|" + }, + { + "bbox": [ + 130, + 528, + 480, + 600 + ], + "type": "text", + "content": ". In principle, " + }, + { + "bbox": [ + 130, + 528, + 480, + 600 + ], + "type": "inline_equation", + "content": "\\mathcal{N}_u" + }, + { + "bbox": [ + 130, + 528, + 480, + 600 + ], + "type": "text", + "content": " can be sampled in other relatively sparse patterns. Here we use the criss-cross one for its simple implementation in code." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 132, + 616, + 273, + 628 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 616, + 273, + 628 + ], + "spans": [ + { + "bbox": [ + 132, + 616, + 273, + 628 + ], + "type": "text", + "content": "4.4 Optimization Strategy" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 132, + 634, + 214, + 644 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 634, + 214, + 644 + ], + "spans": [ + { + "bbox": [ + 132, + 634, + 214, + 644 + ], + "type": "text", + "content": "The overall loss is" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 206, + 654, + 481, + 667 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 206, + 654, + 481, + 667 + ], + "spans": [ + { + "bbox": [ + 206, + 654, + 481, + 667 + ], + "type": "interline_equation", + "content": "\\mathcal {L} _ {\\text {a l l}} = \\mathcal {L} _ {\\text {c o l o r}} + \\lambda_ {\\mathrm {p}} \\mathcal {L} _ {\\text {p o l}} ^ {\\mathrm {p}} + \\lambda_ {\\mathrm {n}} \\mathcal {L} _ {\\text {n o r m a l}} + \\lambda_ {\\mathrm {e}} \\mathcal {L} _ {\\text {e i k o n a l}} \\tag {14}", + "image_path": "5ea05532572a58548e4475396766ed6ed43c9bf1714aab2f31258a80b3aea6d8.jpg" + } + ] + } + ], + "index": 14 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "spans": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "type": "text", + "content": "8" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 166, + 91, + 225, + 101 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 166, + 91, + 225, + 101 + ], + "spans": [ + { + "bbox": [ + 166, + 91, + 225, + 101 + ], + "type": "text", + "content": "G. Chen et al." + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 7 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 136, + 114, + 269, + 215 + ], + "blocks": [ + { + "bbox": [ + 136, + 114, + 269, + 215 + ], + "lines": [ + { + "bbox": [ + 136, + 114, + 269, + 215 + ], + "spans": [ + { + "bbox": [ + 136, + 114, + 269, + 215 + ], + "type": "image", + "image_path": "ad2a90825e6a009c1f1cda070f0a1e0a427a491413a6cc3cfbecf8222c46ead7.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 195, + 216, + 208, + 225 + ], + "lines": [ + { + "bbox": [ + 195, + 216, + 208, + 225 + ], + "spans": [ + { + "bbox": [ + 195, + 216, + 208, + 225 + ], + "type": "text", + "content": "(a)" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_caption" + } + ], + "index": 2 + }, + { + "type": "image", + "bbox": [ + 276, + 114, + 337, + 215 + ], + "blocks": [ + { + "bbox": [ + 276, + 114, + 337, + 215 + ], + "lines": [ + { + "bbox": [ + 276, + 114, + 337, + 215 + ], + "spans": [ + { + "bbox": [ + 276, + 114, + 337, + 215 + ], + "type": "image", + "image_path": "4482a83617f06baf723fd609e86a07738f14f38a91835588448fc51771ec4e25.jpg" + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 299, + 217, + 312, + 225 + ], + "lines": [ + { + "bbox": [ + 299, + 217, + 312, + 225 + ], + "spans": [ + { + "bbox": [ + 299, + 217, + 312, + 225 + ], + "type": "text", + "content": "(b)" + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 131, + 234, + 480, + 257 + ], + "lines": [ + { + "bbox": [ + 131, + 234, + 480, + 257 + ], + "spans": [ + { + "bbox": [ + 131, + 234, + 480, + 257 + ], + "type": "text", + "content": "Fig. 3: Dataset and collection. (a) Objects with ground-truth meshes. (b) Objects without ground-truth meshes. (c) Capture setup." + } + ] + } + ], + "index": 8, + "angle": 0, + "type": "image_caption" + } + ], + "index": 4 + }, + { + "type": "image", + "bbox": [ + 343, + 114, + 478, + 215 + ], + "blocks": [ + { + "bbox": [ + 343, + 114, + 478, + 215 + ], + "lines": [ + { + "bbox": [ + 343, + 114, + 478, + 215 + ], + "spans": [ + { + "bbox": [ + 343, + 114, + 478, + 215 + ], + "type": "image", + "image_path": "a60ac6b24b2d8417635b8e40757e67e9d17fe15162457302ca6a5593ba9a3cb3.jpg" + } + ] + } + ], + "index": 6, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 403, + 217, + 415, + 225 + ], + "lines": [ + { + "bbox": [ + 403, + 217, + 415, + 225 + ], + "spans": [ + { + "bbox": [ + 403, + 217, + 415, + 225 + ], + "type": "text", + "content": "(c)" + } + ] + } + ], + "index": 7, + "angle": 0, + "type": "image_caption" + } + ], + "index": 6 + }, + { + "bbox": [ + 130, + 280, + 482, + 368 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 280, + 482, + 368 + ], + "spans": [ + { + "bbox": [ + 130, + 280, + 482, + 368 + ], + "type": "text", + "content": "The extra Eikonal loss " + }, + { + "bbox": [ + 130, + 280, + 482, + 368 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_{\\mathrm{eikonal}} = \\frac{1}{M|S|}\\sum_{k=1}^{M|S|}(\\|\\nabla\\hat{s}(\\mathbf{x}_k)\\| - 1)^2" + }, + { + "bbox": [ + 130, + 280, + 482, + 368 + ], + "type": "text", + "content": " is used for encouraging the neural field to approximate a signed distance field [10]. During the optimization, we apply progressive training [17] on the multi-resolution hash grid. To initialize a coarse shape, we set " + }, + { + "bbox": [ + 130, + 280, + 482, + 368 + ], + "type": "inline_equation", + "content": "\\lambda_{\\mathrm{p}} = \\lambda_{\\mathrm{n}} = 0" + }, + { + "bbox": [ + 130, + 280, + 482, + 368 + ], + "type": "text", + "content": " and use the photometric and eikonal losses at the early stage. Then we gradually increase " + }, + { + "bbox": [ + 130, + 280, + 482, + 368 + ], + "type": "inline_equation", + "content": "\\lambda_{\\mathrm{p}}" + }, + { + "bbox": [ + 130, + 280, + 482, + 368 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 130, + 280, + 482, + 368 + ], + "type": "inline_equation", + "content": "\\lambda_{\\mathrm{n}}" + }, + { + "bbox": [ + 130, + 280, + 482, + 368 + ], + "type": "text", + "content": " to rectify the distorted shape. Finally, we gradually decrease " + }, + { + "bbox": [ + 130, + 280, + 482, + 368 + ], + "type": "inline_equation", + "content": "\\lambda_{\\mathrm{n}}" + }, + { + "bbox": [ + 130, + 280, + 482, + 368 + ], + "type": "text", + "content": " to 0 to avoid over smoothing the details." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 132, + 388, + 230, + 402 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 388, + 230, + 402 + ], + "spans": [ + { + "bbox": [ + 132, + 388, + 230, + 402 + ], + "type": "text", + "content": "5 Experiments" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 131, + 415, + 261, + 428 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 131, + 415, + 261, + 428 + ], + "spans": [ + { + "bbox": [ + 131, + 415, + 261, + 428 + ], + "type": "text", + "content": "5.1 Experimental Setup" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 130, + 437, + 482, + 533 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 437, + 482, + 533 + ], + "spans": [ + { + "bbox": [ + 130, + 437, + 482, + 533 + ], + "type": "text", + "content": "Dataset: Since there is no public polarization image dataset of textureless and specular objects with ground-truth meshes, we collect a real-world dataset, including four objects (Black Dragon, Red Dragon, Standing Rabbit, Lying Rabbit, Figure and Car) as shown in Fig. 3. The reasons for capturing real object images, rather than synthesizing using a rendering engine, such as [13], are two-fold: 1) the natural ambient lighting conditions are complex and therefore difficult to synthesize, and 2) the shadows and occlusions of the camera can be realistically generated." + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 130, + 533, + 482, + 654 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 533, + 482, + 654 + ], + "spans": [ + { + "bbox": [ + 130, + 533, + 482, + 654 + ], + "type": "text", + "content": "Each object's polarization images are captured indoors under natural lighting conditions from 40 (60 for Figure for its relatively complex shape) different viewpoints around the object with a color polarization camera [36]. Color images are obtained by first decomposing polarization images into images of four polarizer angles and then averaging and demosaicing. Camera poses are estimated from color images using SuperGlue [32] and COLMAP [33] given pre-calibrated camera intrinsics. We use a checkerboard to help pose estimation. To obtain ground-truth meshes, the objects are coated by 3D scanning spray to eliminate specular reflection and scanned by a structure light-based scanner except for Figure and Car. Please refer to our supplementary material for more details." + } + ] + } + ], + "index": 13 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 272, + 91, + 448, + 102 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 272, + 91, + 448, + 102 + ], + "spans": [ + { + "bbox": [ + 272, + 91, + 448, + 102 + ], + "type": "text", + "content": "PISR for Textureless and Specular Objects" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 474, + 92, + 480, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 474, + 92, + 480, + 100 + ], + "spans": [ + { + "bbox": [ + 474, + 92, + 480, + 100 + ], + "type": "text", + "content": "9" + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 8 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 153, + 116, + 196, + 159 + ], + "blocks": [ + { + "bbox": [ + 138, + 119, + 149, + 155 + ], + "lines": [ + { + "bbox": [ + 138, + 119, + 149, + 155 + ], + "spans": [ + { + "bbox": [ + 138, + 119, + 149, + 155 + ], + "type": "text", + "content": "Baaee" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 153, + 116, + 196, + 159 + ], + "lines": [ + { + "bbox": [ + 153, + 116, + 196, + 159 + ], + "spans": [ + { + "bbox": [ + 153, + 116, + 196, + 159 + ], + "type": "image", + "image_path": "af839f7fac945c5590b7ba519169fc518c49a16245ce1ac542e8e1f20729b32e.jpg" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_body" + } + ], + "index": 3 + }, + { + "type": "image", + "bbox": [ + 206, + 117, + 249, + 159 + ], + "blocks": [ + { + "bbox": [ + 206, + 117, + 249, + 159 + ], + "lines": [ + { + "bbox": [ + 206, + 117, + 249, + 159 + ], + "spans": [ + { + "bbox": [ + 206, + 117, + 249, + 159 + ], + "type": "image", + "image_path": "402c09e0b7c2f735a2509080170bd6be041b5d78bf0c57c7930944245cccd929.jpg" + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "image_body" + } + ], + "index": 4 + }, + { + "type": "image", + "bbox": [ + 258, + 116, + 301, + 159 + ], + "blocks": [ + { + "bbox": [ + 258, + 116, + 301, + 159 + ], + "lines": [ + { + "bbox": [ + 258, + 116, + 301, + 159 + ], + "spans": [ + { + "bbox": [ + 258, + 116, + 301, + 159 + ], + "type": "image", + "image_path": "afb9c62b8b4d84827e8c7916c0d6d406c7bcb31dad5ebdffb9eb467c5e433a11.jpg" + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "image_body" + } + ], + "index": 5 + }, + { + "type": "image", + "bbox": [ + 311, + 117, + 353, + 158 + ], + "blocks": [ + { + "bbox": [ + 311, + 117, + 353, + 158 + ], + "lines": [ + { + "bbox": [ + 311, + 117, + 353, + 158 + ], + "spans": [ + { + "bbox": [ + 311, + 117, + 353, + 158 + ], + "type": "image", + "image_path": "390a22fd15d12af1da682339f55c73b3fb93e8ea777d4f5b65b6bc39e78a516d.jpg" + } + ] + } + ], + "index": 6, + "angle": 0, + "type": "image_body" + } + ], + "index": 6 + }, + { + "type": "image", + "bbox": [ + 363, + 117, + 406, + 158 + ], + "blocks": [ + { + "bbox": [ + 363, + 117, + 406, + 158 + ], + "lines": [ + { + "bbox": [ + 363, + 117, + 406, + 158 + ], + "spans": [ + { + "bbox": [ + 363, + 117, + 406, + 158 + ], + "type": "image", + "image_path": "7062568af730f6c1253dc2085c3eca3eb5c9e2b11e0598e8202691cfffd82a76.jpg" + } + ] + } + ], + "index": 7, + "angle": 0, + "type": "image_body" + } + ], + "index": 7 + }, + { + "type": "image", + "bbox": [ + 416, + 117, + 459, + 158 + ], + "blocks": [ + { + "bbox": [ + 416, + 117, + 459, + 158 + ], + "lines": [ + { + "bbox": [ + 416, + 117, + 459, + 158 + ], + "spans": [ + { + "bbox": [ + 416, + 117, + 459, + 158 + ], + "type": "image", + "image_path": "61c5d10507c487387c2ef71eeda335f322d42a005aaae000c4ad9f94c32865d0.jpg" + } + ] + } + ], + "index": 8, + "angle": 0, + "type": "image_body" + } + ], + "index": 8 + }, + { + "type": "image", + "bbox": [ + 462, + 125, + 480, + 304 + ], + "blocks": [ + { + "bbox": [ + 462, + 125, + 480, + 304 + ], + "lines": [ + { + "bbox": [ + 462, + 125, + 480, + 304 + ], + "spans": [ + { + "bbox": [ + 462, + 125, + 480, + 304 + ], + "type": "image", + "image_path": "aaa03de075311943c9d8c87a64bae4ce9266bfef6f8e9bda6dcfc4484474a6d0.jpg" + } + ] + } + ], + "index": 9, + "angle": 0, + "type": "image_body" + } + ], + "index": 9 + }, + { + "type": "image", + "bbox": [ + 153, + 164, + 195, + 205 + ], + "blocks": [ + { + "bbox": [ + 139, + 171, + 149, + 204 + ], + "lines": [ + { + "bbox": [ + 139, + 171, + 149, + 204 + ], + "spans": [ + { + "bbox": [ + 139, + 171, + 149, + 204 + ], + "type": "text", + "content": "Rrnnnne" + } + ] + } + ], + "index": 10, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 153, + 164, + 195, + 205 + ], + "lines": [ + { + "bbox": [ + 153, + 164, + 195, + 205 + ], + "spans": [ + { + "bbox": [ + 153, + 164, + 195, + 205 + ], + "type": "image", + "image_path": "9f30589a6477021164cc8b6be51d64d03893dc6a7f622d12385b259030df037d.jpg" + } + ] + } + ], + "index": 11, + "angle": 0, + "type": "image_body" + } + ], + "index": 11 + }, + { + "type": "image", + "bbox": [ + 205, + 164, + 249, + 205 + ], + "blocks": [ + { + "bbox": [ + 205, + 164, + 249, + 205 + ], + "lines": [ + { + "bbox": [ + 205, + 164, + 249, + 205 + ], + "spans": [ + { + "bbox": [ + 205, + 164, + 249, + 205 + ], + "type": "image", + "image_path": "305e0fcc06b9c08958713a3231d6b7fd32065ec2214600a4ca8b473e9497ea41.jpg" + } + ] + } + ], + "index": 12, + "angle": 0, + "type": "image_body" + } + ], + "index": 12 + }, + { + "type": "image", + "bbox": [ + 258, + 164, + 301, + 205 + ], + "blocks": [ + { + "bbox": [ + 258, + 164, + 301, + 205 + ], + "lines": [ + { + "bbox": [ + 258, + 164, + 301, + 205 + ], + "spans": [ + { + "bbox": [ + 258, + 164, + 301, + 205 + ], + "type": "image", + "image_path": "f265945b3935956bd26ef4c63b2293b7ae64cf42f97de268e9eecd7198645d21.jpg" + } + ] + } + ], + "index": 13, + "angle": 0, + "type": "image_body" + } + ], + "index": 13 + }, + { + "type": "image", + "bbox": [ + 310, + 164, + 353, + 205 + ], + "blocks": [ + { + "bbox": [ + 310, + 164, + 353, + 205 + ], + "lines": [ + { + "bbox": [ + 310, + 164, + 353, + 205 + ], + "spans": [ + { + "bbox": [ + 310, + 164, + 353, + 205 + ], + "type": "image", + "image_path": "9606c7417819930d655e83645eb54dc0bd03c34e8818f2154a5f1c59d30bd4cc.jpg" + } + ] + } + ], + "index": 14, + "angle": 0, + "type": "image_body" + } + ], + "index": 14 + }, + { + "type": "image", + "bbox": [ + 363, + 164, + 405, + 205 + ], + "blocks": [ + { + "bbox": [ + 363, + 164, + 405, + 205 + ], + "lines": [ + { + "bbox": [ + 363, + 164, + 405, + 205 + ], + "spans": [ + { + "bbox": [ + 363, + 164, + 405, + 205 + ], + "type": "image", + "image_path": "76b2f19e9fce38ef428ac2d2d9cce2cedbb3d90478328e8f3373fac83c19a65b.jpg" + } + ] + } + ], + "index": 15, + "angle": 0, + "type": "image_body" + } + ], + "index": 15 + }, + { + "type": "image", + "bbox": [ + 416, + 164, + 459, + 205 + ], + "blocks": [ + { + "bbox": [ + 416, + 164, + 459, + 205 + ], + "lines": [ + { + "bbox": [ + 416, + 164, + 459, + 205 + ], + "spans": [ + { + "bbox": [ + 416, + 164, + 459, + 205 + ], + "type": "image", + "image_path": "ff231fc2dc87098a84b70215659f9173c94fa48f3224b4bb3c6b9e63d1d7fa05.jpg" + } + ] + } + ], + "index": 16, + "angle": 0, + "type": "image_body" + } + ], + "index": 16 + }, + { + "type": "image", + "bbox": [ + 154, + 208, + 193, + 255 + ], + "blocks": [ + { + "bbox": [ + 139, + 219, + 149, + 250 + ], + "lines": [ + { + "bbox": [ + 139, + 219, + 149, + 250 + ], + "spans": [ + { + "bbox": [ + 139, + 219, + 149, + 250 + ], + "type": "text", + "content": "Rabitt S." + } + ] + } + ], + "index": 17, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 154, + 208, + 193, + 255 + ], + "lines": [ + { + "bbox": [ + 154, + 208, + 193, + 255 + ], + "spans": [ + { + "bbox": [ + 154, + 208, + 193, + 255 + ], + "type": "image", + "image_path": "47cd50d7ccdb63edd1679924aa24fd59372459fbe1f1fa200df06b218b51e0ee.jpg" + } + ] + } + ], + "index": 18, + "angle": 0, + "type": "image_body" + } + ], + "index": 18 + }, + { + "type": "image", + "bbox": [ + 205, + 208, + 246, + 255 + ], + "blocks": [ + { + "bbox": [ + 205, + 208, + 246, + 255 + ], + "lines": [ + { + "bbox": [ + 205, + 208, + 246, + 255 + ], + "spans": [ + { + "bbox": [ + 205, + 208, + 246, + 255 + ], + "type": "image", + "image_path": "b74f8ea4b5a4492c45304007ba65d9435825d1c380392928732b05fce8d683cd.jpg" + } + ] + } + ], + "index": 19, + "angle": 0, + "type": "image_body" + } + ], + "index": 19 + }, + { + "type": "image", + "bbox": [ + 258, + 208, + 298, + 255 + ], + "blocks": [ + { + "bbox": [ + 258, + 208, + 298, + 255 + ], + "lines": [ + { + "bbox": [ + 258, + 208, + 298, + 255 + ], + "spans": [ + { + "bbox": [ + 258, + 208, + 298, + 255 + ], + "type": "image", + "image_path": "86ce32ee94a377de92169710dda7344dc8412f3c5432cd142b4bd5bf32a0354a.jpg" + } + ] + } + ], + "index": 20, + "angle": 0, + "type": "image_body" + } + ], + "index": 20 + }, + { + "type": "image", + "bbox": [ + 312, + 209, + 351, + 255 + ], + "blocks": [ + { + "bbox": [ + 312, + 209, + 351, + 255 + ], + "lines": [ + { + "bbox": [ + 312, + 209, + 351, + 255 + ], + "spans": [ + { + "bbox": [ + 312, + 209, + 351, + 255 + ], + "type": "image", + "image_path": "bd7ac0a4ff2779891e7487a0635ee61474cd59a66448ca4e96233f5b8db64c5e.jpg" + } + ] + } + ], + "index": 21, + "angle": 0, + "type": "image_body" + } + ], + "index": 21 + }, + { + "type": "image", + "bbox": [ + 363, + 209, + 404, + 255 + ], + "blocks": [ + { + "bbox": [ + 363, + 209, + 404, + 255 + ], + "lines": [ + { + "bbox": [ + 363, + 209, + 404, + 255 + ], + "spans": [ + { + "bbox": [ + 363, + 209, + 404, + 255 + ], + "type": "image", + "image_path": "4439d24aad9e33560e33cd0609d3f4ba9f5709d06677687d8f1e32841fec8096.jpg" + } + ] + } + ], + "index": 22, + "angle": 0, + "type": "image_body" + } + ], + "index": 22 + }, + { + "type": "image", + "bbox": [ + 416, + 209, + 456, + 255 + ], + "blocks": [ + { + "bbox": [ + 416, + 209, + 456, + 255 + ], + "lines": [ + { + "bbox": [ + 416, + 209, + 456, + 255 + ], + "spans": [ + { + "bbox": [ + 416, + 209, + 456, + 255 + ], + "type": "image", + "image_path": "d5bf77d25e8ed23cc261fca96f7e2f0997e9c2c1e0dde4b70937ad091ecff674.jpg" + } + ] + } + ], + "index": 23, + "angle": 0, + "type": "image_body" + } + ], + "index": 23 + }, + { + "type": "image", + "bbox": [ + 154, + 256, + 193, + 298 + ], + "blocks": [ + { + "bbox": [ + 139, + 265, + 149, + 297 + ], + "lines": [ + { + "bbox": [ + 139, + 265, + 149, + 297 + ], + "spans": [ + { + "bbox": [ + 139, + 265, + 149, + 297 + ], + "type": "text", + "content": "Rabibit L" + } + ] + } + ], + "index": 24, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 154, + 256, + 193, + 298 + ], + "lines": [ + { + "bbox": [ + 154, + 256, + 193, + 298 + ], + "spans": [ + { + "bbox": [ + 154, + 256, + 193, + 298 + ], + "type": "image", + "image_path": "16097e7320adf1e3539591de277398634af04835b033a7fc4d41b4715c4b8aea.jpg" + } + ] + } + ], + "index": 25, + "angle": 0, + "type": "image_body" + } + ], + "index": 25 + }, + { + "type": "image", + "bbox": [ + 206, + 256, + 246, + 298 + ], + "blocks": [ + { + "bbox": [ + 206, + 256, + 246, + 298 + ], + "lines": [ + { + "bbox": [ + 206, + 256, + 246, + 298 + ], + "spans": [ + { + "bbox": [ + 206, + 256, + 246, + 298 + ], + "type": "image", + "image_path": "68d8cfb4ca6944f5204923ceeb568d26aa322fb1c02321043fb8ead27430e6a2.jpg" + } + ] + } + ], + "index": 26, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 130, + 319, + 482, + 364 + ], + "lines": [ + { + "bbox": [ + 130, + 319, + 482, + 364 + ], + "spans": [ + { + "bbox": [ + 130, + 319, + 482, + 364 + ], + "type": "text", + "content": "Fig. 4: Signed error maps. Red and blue mean surface swelling and shrinking respectively. Regions with higher color saturation indicate longer distances between the reconstruction result and the ground truth. Errors are truncated to within " + }, + { + "bbox": [ + 130, + 319, + 482, + 364 + ], + "type": "inline_equation", + "content": "\\pm 2\\mathrm{mm}" + }, + { + "bbox": [ + 130, + 319, + 482, + 364 + ], + "type": "text", + "content": " for better visualization." + } + ] + } + ], + "index": 34, + "angle": 0, + "type": "image_caption" + } + ], + "index": 26 + }, + { + "type": "image", + "bbox": [ + 260, + 256, + 298, + 298 + ], + "blocks": [ + { + "bbox": [ + 260, + 256, + 298, + 298 + ], + "lines": [ + { + "bbox": [ + 260, + 256, + 298, + 298 + ], + "spans": [ + { + "bbox": [ + 260, + 256, + 298, + 298 + ], + "type": "image", + "image_path": "064fbcc29c2aa257066d05b6fe93579dcc289a6747e2726699c43cc8fd673364.jpg" + } + ] + } + ], + "index": 27, + "angle": 0, + "type": "image_body" + } + ], + "index": 27 + }, + { + "type": "image", + "bbox": [ + 312, + 256, + 351, + 299 + ], + "blocks": [ + { + "bbox": [ + 312, + 256, + 351, + 299 + ], + "lines": [ + { + "bbox": [ + 312, + 256, + 351, + 299 + ], + "spans": [ + { + "bbox": [ + 312, + 256, + 351, + 299 + ], + "type": "image", + "image_path": "b38268a9062d69a8ba55f4189b733e82426ac06681ffc84860aa545ea0c7ff00.jpg" + } + ] + } + ], + "index": 28, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 310, + 300, + 350, + 309 + ], + "lines": [ + { + "bbox": [ + 310, + 300, + 350, + 309 + ], + "spans": [ + { + "bbox": [ + 310, + 300, + 350, + 309 + ], + "type": "text", + "content": "PANDORA" + } + ] + } + ], + "index": 29, + "angle": 0, + "type": "image_caption" + } + ], + "index": 28 + }, + { + "type": "image", + "bbox": [ + 364, + 256, + 403, + 299 + ], + "blocks": [ + { + "bbox": [ + 364, + 256, + 403, + 299 + ], + "lines": [ + { + "bbox": [ + 364, + 256, + 403, + 299 + ], + "spans": [ + { + "bbox": [ + 364, + 256, + 403, + 299 + ], + "type": "image", + "image_path": "04cc31e212237a2f738d991801c3ac69647ad7ac6414ca84d647ee4142dcf504.jpg" + } + ] + } + ], + "index": 30, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 369, + 300, + 395, + 309 + ], + "lines": [ + { + "bbox": [ + 369, + 300, + 395, + 309 + ], + "spans": [ + { + "bbox": [ + 369, + 300, + 395, + 309 + ], + "type": "text", + "content": "PMVIR" + } + ] + } + ], + "index": 31, + "angle": 0, + "type": "image_caption" + } + ], + "index": 30 + }, + { + "type": "image", + "bbox": [ + 416, + 256, + 456, + 299 + ], + "blocks": [ + { + "bbox": [ + 416, + 256, + 456, + 299 + ], + "lines": [ + { + "bbox": [ + 416, + 256, + 456, + 299 + ], + "spans": [ + { + "bbox": [ + 416, + 256, + 456, + 299 + ], + "type": "image", + "image_path": "847a5e8d15f4b341aeb9f51669bb0703530c86b954db100484761f3fdb161633.jpg" + } + ] + } + ], + "index": 32, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 416, + 300, + 455, + 309 + ], + "lines": [ + { + "bbox": [ + 416, + 300, + 455, + 309 + ], + "spans": [ + { + "bbox": [ + 416, + 300, + 455, + 309 + ], + "type": "text", + "content": "PISR (ours)" + } + ] + } + ], + "index": 33, + "angle": 0, + "type": "image_caption" + } + ], + "index": 32 + }, + { + "type": "table", + "bbox": [ + 160, + 401, + 453, + 492 + ], + "blocks": [ + { + "bbox": [ + 209, + 380, + 405, + 392 + ], + "lines": [ + { + "bbox": [ + 209, + 380, + 405, + 392 + ], + "spans": [ + { + "bbox": [ + 209, + 380, + 405, + 392 + ], + "type": "text", + "content": "Table 1: Summary of Methods in Experiments" + } + ] + } + ], + "index": 35, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 160, + 401, + 453, + 492 + ], + "lines": [ + { + "bbox": [ + 160, + 401, + 453, + 492 + ], + "spans": [ + { + "bbox": [ + 160, + 401, + 453, + 492 + ], + "type": "table", + "html": "
MethodsImageShape Repre.Appearance Repre.
Ref-NeuS [9]RGBNeural SDFNeural radiance
NeRO [18]RGBNeural SDFNeural BRDF
NeuS [42]RGBNeural SDFNeural radiance
PANDORA [6]Pol. RGBNeural SDFNeural BRDF
PMVIR [54]Pol. RGBMeshSpherical harmonics
PISR (Ours)Pol. RGBGrids+Neural SDFNeural radiance
", + "image_path": "7a85cb613c7375159a74e0c94a83c6484789eaf2b256b61eb004c7a13c807f03.jpg" + } + ] + } + ], + "index": 36, + "angle": 0, + "type": "table_body" + } + ], + "index": 36 + }, + { + "bbox": [ + 130, + 519, + 482, + 556 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 519, + 482, + 556 + ], + "spans": [ + { + "bbox": [ + 130, + 519, + 482, + 556 + ], + "type": "text", + "content": "Performance Metrics: Since we focus on geometry reconstruction, we only evaluate the surface reconstruction quality using L1 Chamfer distance and F-score. We remove the background meshes before calculating the metrics." + } + ] + } + ], + "index": 37 + }, + { + "bbox": [ + 130, + 569, + 482, + 667 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 569, + 482, + 667 + ], + "spans": [ + { + "bbox": [ + 130, + 569, + 482, + 667 + ], + "type": "text", + "content": "Implementation Details: The SDF MLP " + }, + { + "bbox": [ + 130, + 569, + 482, + 667 + ], + "type": "inline_equation", + "content": "\\varPhi_s" + }, + { + "bbox": [ + 130, + 569, + 482, + 667 + ], + "type": "text", + "content": " has one hidden layer with a size of 64, and the color MLP " + }, + { + "bbox": [ + 130, + 569, + 482, + 667 + ], + "type": "inline_equation", + "content": "\\varPhi_{\\mathbf{c}}" + }, + { + "bbox": [ + 130, + 569, + 482, + 667 + ], + "type": "text", + "content": " has two hidden layers with sizes of 64. The multi-resolution grid contains 16 levels of resolution, and the coarsest and finest resolutions are 32 and 2700, respectively. The hash table size of each level is " + }, + { + "bbox": [ + 130, + 569, + 482, + 667 + ], + "type": "inline_equation", + "content": "2^{19}" + }, + { + "bbox": [ + 130, + 569, + 482, + 667 + ], + "type": "text", + "content": ". The optimization of the grid starts at the fourth-coarsest resolutions and progressively extends one level every 1000 iterations. Throughout the whole optimization process, the threshold of DoP " + }, + { + "bbox": [ + 130, + 569, + 482, + 667 + ], + "type": "inline_equation", + "content": "\\theta" + }, + { + "bbox": [ + 130, + 569, + 482, + 667 + ], + "type": "text", + "content": " is set to 0.3 and " + }, + { + "bbox": [ + 130, + 569, + 482, + 667 + ], + "type": "inline_equation", + "content": "\\lambda_{\\mathrm{eikonal}}" + }, + { + "bbox": [ + 130, + 569, + 482, + 667 + ], + "type": "text", + "content": " is set to 0.1. After the first 2.5k iterations, " + }, + { + "bbox": [ + 130, + 569, + 482, + 667 + ], + "type": "inline_equation", + "content": "\\lambda_{\\mathrm{pol}}" + }, + { + "bbox": [ + 130, + 569, + 482, + 667 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 130, + 569, + 482, + 667 + ], + "type": "inline_equation", + "content": "\\lambda_{\\mathrm{normal}}" + }, + { + "bbox": [ + 130, + 569, + 482, + 667 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 130, + 569, + 482, + 667 + ], + "type": "inline_equation", + "content": "|\\mathcal{N}_p|" + }, + { + "bbox": [ + 130, + 569, + 482, + 667 + ], + "type": "text", + "content": " are linearly increased from" + } + ] + } + ], + "index": 38 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "spans": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "type": "text", + "content": "10" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 166, + 91, + 225, + 100 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 166, + 91, + 225, + 100 + ], + "spans": [ + { + "bbox": [ + 166, + 91, + 225, + 100 + ], + "type": "text", + "content": "G. Chen et al." + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 9 + }, + { + "para_blocks": [ + { + "type": "table", + "bbox": [ + 146, + 178, + 469, + 288 + ], + "blocks": [ + { + "bbox": [ + 130, + 114, + 482, + 170 + ], + "lines": [ + { + "bbox": [ + 130, + 114, + 482, + 170 + ], + "spans": [ + { + "bbox": [ + 130, + 114, + 482, + 170 + ], + "type": "text", + "content": "Table 2: Reconstruction accuracy in Chamfer distance (CD↓, mm) and F-score (FS↑, %). FS is calculated at a threshold of 1 mm. Note that the values in the table are rounded to one decimal place. The CD of PMVIR and PISR on Rabbit L. are 0.584 mm and 0.578 mm respectively. Bold, underline, and double underline mean the best, second-best and third-best performance, respectively." + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 146, + 178, + 469, + 288 + ], + "lines": [ + { + "bbox": [ + 146, + 178, + 469, + 288 + ], + "spans": [ + { + "bbox": [ + 146, + 178, + 469, + 288 + ], + "type": "table", + "html": "
MethodsDragon B.Dragon R.Rabbit S.Rabbit L.Avg.
CDFSCDFSCDFSCDFSCDFS
RGBRef-NeuS2.165.32.166.11.284.51.090.51.676.6
NeRO2.459.21.866.31.380.51.385.01.772.8
NeuS1.967.11.286.01.186.80.993.01.383.2
RGBPANDORA2.355.31.864.41.478.31.577.11.868.8
PMVIR1.876.51.090.50.697.50.697.31.090.5
PISR (Ours)0.599.90.699.10.599.70.699.40.599.5
", + "image_path": "a2a510f2bee08f6e4e9b0cb345e02ba4786afdcfcec80128f9dfa015fe214eeb.jpg" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "table_body" + } + ], + "index": 3 + }, + { + "bbox": [ + 130, + 307, + 482, + 345 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 307, + 482, + 345 + ], + "spans": [ + { + "bbox": [ + 130, + 307, + 482, + 345 + ], + "type": "text", + "content": "0 to 2.0, 1.0 and 28 in " + }, + { + "bbox": [ + 130, + 307, + 482, + 345 + ], + "type": "inline_equation", + "content": "2.5\\mathrm{k}" + }, + { + "bbox": [ + 130, + 307, + 482, + 345 + ], + "type": "text", + "content": " iterations, respectively. Then " + }, + { + "bbox": [ + 130, + 307, + 482, + 345 + ], + "type": "inline_equation", + "content": "\\lambda_{\\mathrm{normal}}" + }, + { + "bbox": [ + 130, + 307, + 482, + 345 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 130, + 307, + 482, + 345 + ], + "type": "inline_equation", + "content": "|\\mathcal{N}_p|" + }, + { + "bbox": [ + 130, + 307, + 482, + 345 + ], + "type": "text", + "content": " are linearly decreased to 0 in another " + }, + { + "bbox": [ + 130, + 307, + 482, + 345 + ], + "type": "inline_equation", + "content": "2.5\\mathrm{k}" + }, + { + "bbox": [ + 130, + 307, + 482, + 345 + ], + "type": "text", + "content": " iterations. The maximum number of iterations is " + }, + { + "bbox": [ + 130, + 307, + 482, + 345 + ], + "type": "inline_equation", + "content": "20\\mathrm{k}" + }, + { + "bbox": [ + 130, + 307, + 482, + 345 + ], + "type": "text", + "content": ". The final SDF is converted to meshes using Marching Cubes [20]." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 130, + 354, + 482, + 428 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 354, + 482, + 428 + ], + "spans": [ + { + "bbox": [ + 130, + 354, + 482, + 428 + ], + "type": "text", + "content": "Baselines: We compare our method with five methods, including PMVIR [54], PANDORA [6], NeRO [18], Ref-NeuS [9] and NeuS [42]. Tab. 1 summarizes the type of input images, geometry representation and color representation of all the six methods. Since PMVIR is a refinement method based on an initial shape, we provide the coarse initial meshes reconstructed by PISR-C (see the ablation study in Sec. 5.3 for the details of PISR-C)." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 131, + 443, + 200, + 454 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 131, + 443, + 200, + 454 + ], + "spans": [ + { + "bbox": [ + 131, + 443, + 200, + 454 + ], + "type": "text", + "content": "5.2 Results" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 130, + 462, + 482, + 618 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 462, + 482, + 618 + ], + "spans": [ + { + "bbox": [ + 130, + 462, + 482, + 618 + ], + "type": "text", + "content": "Quantitative Results: The Chamfer distances and F-scores of the reconstruction results are reported in Tab. 2. Our method achieves the best reconstruction quality on all four objects shown in Fig. 3a and significantly outperforms the other five methods. The average Chamfer distance and F-score of our method are " + }, + { + "bbox": [ + 130, + 462, + 482, + 618 + ], + "type": "inline_equation", + "content": "0.5 \\, \\mathrm{mm}" + }, + { + "bbox": [ + 130, + 462, + 482, + 618 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 130, + 462, + 482, + 618 + ], + "type": "inline_equation", + "content": "99.5\\%" + }, + { + "bbox": [ + 130, + 462, + 482, + 618 + ], + "type": "text", + "content": ", respectively. As the second-best method in the comparison, PMVIR achieves twice the Chamfer distance and a " + }, + { + "bbox": [ + 130, + 462, + 482, + 618 + ], + "type": "inline_equation", + "content": "9\\%" + }, + { + "bbox": [ + 130, + 462, + 482, + 618 + ], + "type": "text", + "content": " lower F-score compared to ours. Both PMVIR and our method utilize polarimetric constraints, and the results show that the geometry cue of polarization is effective for shape estimation. Although PANDORA also uses polarization images as input, it implicitly utilizes polarization cues through polarimetric rendering, i.e., optimizing shapes using an image reconstruction loss, which could suffer from the shape-radiance ambiguity as RGB-based methods. Among RGB-based methods, NeuS achieves better results than Ref-NeuS and NeRO which are designed for specular objects." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 130, + 629, + 482, + 666 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 629, + 482, + 666 + ], + "spans": [ + { + "bbox": [ + 130, + 629, + 482, + 666 + ], + "type": "text", + "content": "We visualize the error distributions in Fig. 4. For methods that rely solely on image reconstruction losses (Ref-NeuS, NeRO, NeuS and PANDORA), their reconstruction errors in non-convex regions are notably higher compared to" + } + ] + } + ], + "index": 8 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 272, + 91, + 447, + 102 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 272, + 91, + 447, + 102 + ], + "spans": [ + { + "bbox": [ + 272, + 91, + 447, + 102 + ], + "type": "text", + "content": "PISR for Textureless and Specular Objects" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 470, + 91, + 479, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 470, + 91, + 479, + 100 + ], + "spans": [ + { + "bbox": [ + 470, + 91, + 479, + 100 + ], + "type": "text", + "content": "11" + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 10 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 134, + 113, + 248, + 218 + ], + "blocks": [ + { + "bbox": [ + 134, + 113, + 248, + 218 + ], + "lines": [ + { + "bbox": [ + 134, + 113, + 248, + 218 + ], + "spans": [ + { + "bbox": [ + 134, + 113, + 248, + 218 + ], + "type": "image", + "image_path": "4ff13a0f845138e44cc2454e391498a61ed2575212ca31a0f43b8e964a7357a8.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 165, + 219, + 208, + 226 + ], + "lines": [ + { + "bbox": [ + 165, + 219, + 208, + 226 + ], + "spans": [ + { + "bbox": [ + 165, + 219, + 208, + 226 + ], + "type": "text", + "content": "Initialization" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_caption" + } + ], + "index": 2 + }, + { + "type": "image", + "bbox": [ + 251, + 113, + 365, + 217 + ], + "blocks": [ + { + "bbox": [ + 251, + 113, + 365, + 217 + ], + "lines": [ + { + "bbox": [ + 251, + 113, + 365, + 217 + ], + "spans": [ + { + "bbox": [ + 251, + 113, + 365, + 217 + ], + "type": "image", + "image_path": "0b11c7fca60e3e8be9ed524bbd37ca08b6d52b836ed3a90e0dd9c8e87cd4071c.jpg" + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 293, + 218, + 321, + 226 + ], + "lines": [ + { + "bbox": [ + 293, + 218, + 321, + 226 + ], + "spans": [ + { + "bbox": [ + 293, + 218, + 321, + 226 + ], + "type": "text", + "content": "PMVIR" + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "image_caption" + } + ], + "index": 4 + }, + { + "type": "image", + "bbox": [ + 368, + 113, + 482, + 217 + ], + "blocks": [ + { + "bbox": [ + 368, + 113, + 482, + 217 + ], + "lines": [ + { + "bbox": [ + 368, + 113, + 482, + 217 + ], + "spans": [ + { + "bbox": [ + 368, + 113, + 482, + 217 + ], + "type": "image", + "image_path": "cd9c71569bc759d4b33bef45b1e71bf08f5bf8134105d35b2bf99a2aed20cd9b.jpg" + } + ] + } + ], + "index": 6, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 412, + 218, + 431, + 226 + ], + "lines": [ + { + "bbox": [ + 412, + 218, + 431, + 226 + ], + "spans": [ + { + "bbox": [ + 412, + 218, + 431, + 226 + ], + "type": "text", + "content": "PISR" + } + ] + } + ], + "index": 7, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 131, + 234, + 480, + 257 + ], + "lines": [ + { + "bbox": [ + 131, + 234, + 480, + 257 + ], + "spans": [ + { + "bbox": [ + 131, + 234, + 480, + 257 + ], + "type": "text", + "content": "Fig. 5: Normal maps of Dragon B. before and after optimizations. Using neural SDFs to represent shapes allows changes in the shape topology during the optimization." + } + ] + } + ], + "index": 8, + "angle": 0, + "type": "image_caption" + } + ], + "index": 6 + }, + { + "type": "table", + "bbox": [ + 151, + 289, + 462, + 335 + ], + "blocks": [ + { + "bbox": [ + 188, + 267, + 424, + 278 + ], + "lines": [ + { + "bbox": [ + 188, + 267, + 424, + 278 + ], + "spans": [ + { + "bbox": [ + 188, + 267, + 424, + 278 + ], + "type": "text", + "content": "Table 3: Average optimization time of different methods." + } + ] + } + ], + "index": 9, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 151, + 289, + 462, + 335 + ], + "lines": [ + { + "bbox": [ + 151, + 289, + 462, + 335 + ], + "spans": [ + { + "bbox": [ + 151, + 289, + 462, + 335 + ], + "type": "table", + "html": "
MethodsRef-NeuSNeRONeuSPANDORAPMVIRPISR (Ours)
Iterations200k200k200k100k420k
Time (h)20.020.510.815.02.00.5
", + "image_path": "b3a9476d2ff4a0e394e00ccd1e1d06962d30135887cfd0bdd4af910454970a06.jpg" + } + ] + } + ], + "index": 10, + "angle": 0, + "type": "table_body" + } + ], + "index": 10 + }, + { + "bbox": [ + 130, + 356, + 482, + 499 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 356, + 482, + 499 + ], + "spans": [ + { + "bbox": [ + 130, + 356, + 482, + 499 + ], + "type": "text", + "content": "PMVIR and our method. While the inherent continuity of MLPs ensures the watertightness of their reconstructed shapes, their accuracy cannot be guaranteed due to the shape-radiance ambiguity introduced by specular reflections and the lack of textures. Additionally, for NeRO and PANDORA, the lighting shadows cast by the camera further exacerbate the difficulty of estimating surface material parameters. The real HDR ambient lighting also poses challenges for PANDORA and Ref-NeuS to accurately estimate object shapes. In contrast, PMVIR and our method can recover non-convex regions using the polarimetric constraints. Despite the polarization cues, representing the shapes as neural SDF helps our method better reconstruct regions that are poorly initialized. As shown in Fig. 5, neural SDF allows continuous changes of the shape topology during optimization, which is crucial for evolving surfaces from holes." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 130, + 511, + 482, + 608 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 511, + 482, + 608 + ], + "spans": [ + { + "bbox": [ + 130, + 511, + 482, + 608 + ], + "type": "text", + "content": "**Optimization Time:** We report the average optimization time in Tab. 3. The experiments are conducted on a machine equipped with an Intel Core i9-12900K and an NVIDIA RTX 3090 GPU. We set the maximum number of iterations to " + }, + { + "bbox": [ + 130, + 511, + 482, + 608 + ], + "type": "inline_equation", + "content": "200\\mathrm{k}" + }, + { + "bbox": [ + 130, + 511, + 482, + 608 + ], + "type": "text", + "content": ". If the shape converges, we stop the optimization earlier. Our method represents shape as a multi-resolution hash grid and two shallow MLPs, significantly reducing the time complexity for rendering a pixel. Therefore, our method converges much faster, achieving up to " + }, + { + "bbox": [ + 130, + 511, + 482, + 608 + ], + "type": "inline_equation", + "content": "4 \\sim 41 \\times" + }, + { + "bbox": [ + 130, + 511, + 482, + 608 + ], + "type": "text", + "content": " convergence speed while maintaining the highest reconstruction accuracy." + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 131, + 623, + 238, + 635 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 131, + 623, + 238, + 635 + ], + "spans": [ + { + "bbox": [ + 131, + 623, + 238, + 635 + ], + "type": "text", + "content": "5.3 Ablation Study" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 130, + 641, + 481, + 665 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 641, + 481, + 665 + ], + "spans": [ + { + "bbox": [ + 130, + 641, + 481, + 665 + ], + "type": "text", + "content": "We evaluate the effectiveness of the polarimetric loss and normal regularization, as well as the influence of the perspective effect of the lens. We compare PISR(-" + } + ] + } + ], + "index": 14 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "spans": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "type": "text", + "content": "12" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 166, + 91, + 225, + 100 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 166, + 91, + 225, + 100 + ], + "spans": [ + { + "bbox": [ + 166, + 91, + 225, + 100 + ], + "type": "text", + "content": "G. Chen et al." + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 11 + }, + { + "para_blocks": [ + { + "type": "table", + "bbox": [ + 154, + 178, + 460, + 287 + ], + "blocks": [ + { + "bbox": [ + 132, + 114, + 482, + 170 + ], + "lines": [ + { + "bbox": [ + 132, + 114, + 482, + 170 + ], + "spans": [ + { + "bbox": [ + 132, + 114, + 482, + 170 + ], + "type": "text", + "content": "Table 4: Results of the ablation study on different combinations of losses. The L1 CD↓ (mm) and FSs↑ (%) at 1.0 mm and 0.5 are the averages over four objects in Fig.3a. The ranks are the averages over the three metrics. " + }, + { + "bbox": [ + 132, + 114, + 482, + 170 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_{\\mathrm{pol}}^{\\mathrm{o}}" + }, + { + "bbox": [ + 132, + 114, + 482, + 170 + ], + "type": "text", + "content": " differs from " + }, + { + "bbox": [ + 132, + 114, + 482, + 170 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_{\\mathrm{pol}}^{\\mathrm{p}}" + }, + { + "bbox": [ + 132, + 114, + 482, + 170 + ], + "type": "text", + "content": " by its use of the orthographic constraint Eq. 2 instead of the perspective one Eq. 3. Note that " + }, + { + "bbox": [ + 132, + 114, + 482, + 170 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_{\\mathrm{eikonal}}" + }, + { + "bbox": [ + 132, + 114, + 482, + 170 + ], + "type": "text", + "content": " is used in all experiments and is omitted it in this table." + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 154, + 178, + 460, + 287 + ], + "lines": [ + { + "bbox": [ + 154, + 178, + 460, + 287 + ], + "spans": [ + { + "bbox": [ + 154, + 178, + 460, + 287 + ], + "type": "table", + "html": "
VariantsUsed LossesAvg.Rank
LcolorLpoiLpolLnormalCDFS(1.0)FS(0.5)
PISR-C✓1.1088.1070.306
PISR-N✓✓1.0088.2071.105
PISR-O✓✓0.8096.6075.543
PISR-ON✓✓✓0.7796.1075.504
PISR-P✓✓0.5199.7692.791
PISR(-PN)✓✓✓0.5499.5491.752
", + "image_path": "2db590b15ae31261eee482fe89e4975ebdc7fc381d50dd96f428db69490000da.jpg" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "table_body" + } + ], + "index": 3 + }, + { + "bbox": [ + 132, + 309, + 481, + 357 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 309, + 481, + 357 + ], + "spans": [ + { + "bbox": [ + 132, + 309, + 481, + 357 + ], + "type": "text", + "content": "PN) with its five variants, referred to as " + }, + { + "bbox": [ + 132, + 309, + 481, + 357 + ], + "type": "inline_equation", + "content": "PISR-P" + }, + { + "bbox": [ + 132, + 309, + 481, + 357 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 132, + 309, + 481, + 357 + ], + "type": "inline_equation", + "content": "PISR-ON" + }, + { + "bbox": [ + 132, + 309, + 481, + 357 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 132, + 309, + 481, + 357 + ], + "type": "inline_equation", + "content": "PISR-O" + }, + { + "bbox": [ + 132, + 309, + 481, + 357 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 132, + 309, + 481, + 357 + ], + "type": "inline_equation", + "content": "PISR-N" + }, + { + "bbox": [ + 132, + 309, + 481, + 357 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 132, + 309, + 481, + 357 + ], + "type": "inline_equation", + "content": "PISR-C" + }, + { + "bbox": [ + 132, + 309, + 481, + 357 + ], + "type": "text", + "content": ", based on different combinations of losses as shown in Tab. 4. The quantitative experiment is conducted on objects shown in Fig. 3a and the qualitative experiment is conducted on objects shown in Fig. 3b." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 132, + 357, + 481, + 429 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 357, + 481, + 429 + ], + "spans": [ + { + "bbox": [ + 132, + 357, + 481, + 429 + ], + "type": "text", + "content": "As shown in Tab. 4, the top four performing variants are all incorporated with the polarimetric losses, and fine-grained details can be recovered without complicated appearance modeling. As for the normal regularization, although the quantitative comparison between PISR-PN and PISR-P shows that it slightly reduces the accuracy, the qualitative results in Fig. 6 show its effectiveness for recovering surfaces from severe distortions." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 132, + 429, + 481, + 573 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 429, + 481, + 573 + ], + "spans": [ + { + "bbox": [ + 132, + 429, + 481, + 573 + ], + "type": "text", + "content": "Among the top four performing variants, PISR-PN and PISR-P achieve higher accuracy by considering the perspective effect of the lens using Eq. (3) instead of Eq. (2). As the example error maps of Rabbit L. shown in Fig. 7, the perspective effect leads to surface swelling and shrinking, causing severe cracks on the surface. This phenomenon does not appear in the results of PMVIR and PANDORA. We speculate the reason to be that the hash-grid-based neural SDF we use is too flexible in topology compared to the mesh-based one used in PMVIR, and lacks the smoothness of purely MLP-based one used in PANDORA. Besides, the results in Tab. 4 show that PISR-PN outperforms PISR-ON by a " + }, + { + "bbox": [ + 132, + 429, + 481, + 573 + ], + "type": "inline_equation", + "content": "30\\%" + }, + { + "bbox": [ + 132, + 429, + 481, + 573 + ], + "type": "text", + "content": " reduction in L1 Chamfer distance and a " + }, + { + "bbox": [ + 132, + 429, + 481, + 573 + ], + "type": "inline_equation", + "content": "17\\%" + }, + { + "bbox": [ + 132, + 429, + 481, + 573 + ], + "type": "text", + "content": " improvement in F-score at " + }, + { + "bbox": [ + 132, + 429, + 481, + 573 + ], + "type": "inline_equation", + "content": "0.5 \\mathrm{~mm}" + }, + { + "bbox": [ + 132, + 429, + 481, + 573 + ], + "type": "text", + "content": ", which suggests the necessity of considering the perspective effect to improve the accuracy of polarimetric 3D reconstruction." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 133, + 591, + 219, + 604 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 591, + 219, + 604 + ], + "spans": [ + { + "bbox": [ + 133, + 591, + 219, + 604 + ], + "type": "text", + "content": "6 Conclusion" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 132, + 617, + 481, + 665 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 617, + 481, + 665 + ], + "spans": [ + { + "bbox": [ + 132, + 617, + 481, + 665 + ], + "type": "text", + "content": "We have proposed PISR, an accurate and efficient polarimetric neural implicit surface reconstruction method for textureless and specular objects. The key idea of PISR is to explicitly regularize the SDF using the polarimetric constraints to avoid the shape-radiance ambiguity. More importantly, PISR achieves high accu" + } + ] + } + ], + "index": 8 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 272, + 91, + 447, + 102 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 272, + 91, + 447, + 102 + ], + "spans": [ + { + "bbox": [ + 272, + 91, + 447, + 102 + ], + "type": "text", + "content": "PISR for Textureless and Specular Objects" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 470, + 91, + 480, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 470, + 91, + 480, + 100 + ], + "spans": [ + { + "bbox": [ + 470, + 91, + 480, + 100 + ], + "type": "text", + "content": "13" + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 12 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 132, + 113, + 480, + 296 + ], + "blocks": [ + { + "bbox": [ + 132, + 113, + 480, + 296 + ], + "lines": [ + { + "bbox": [ + 132, + 113, + 480, + 296 + ], + "spans": [ + { + "bbox": [ + 132, + 113, + 480, + 296 + ], + "type": "image", + "image_path": "98032b4b80934bcbd775b5d12cd8772aa34459ecad177d68ece041a695612cf2.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 130, + 303, + 482, + 326 + ], + "lines": [ + { + "bbox": [ + 130, + 303, + 482, + 326 + ], + "spans": [ + { + "bbox": [ + 130, + 303, + 482, + 326 + ], + "type": "text", + "content": "Fig. 6: Normal maps of Figure and Car. The polarimetric loss and normal loss jointly recover the shape from distortions induced by strong reflections." + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_caption" + } + ], + "index": 2 + }, + { + "type": "image", + "bbox": [ + 169, + 352, + 438, + 456 + ], + "blocks": [ + { + "bbox": [ + 169, + 352, + 438, + 456 + ], + "lines": [ + { + "bbox": [ + 169, + 352, + 438, + 456 + ], + "spans": [ + { + "bbox": [ + 169, + 352, + 438, + 456 + ], + "type": "image", + "image_path": "d5d3b40d27e58a5369d5df98e9c3be7a9f49ba008cd2cf5197d0147d817945ac.jpg" + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 130, + 463, + 482, + 498 + ], + "lines": [ + { + "bbox": [ + 130, + 463, + 482, + 498 + ], + "spans": [ + { + "bbox": [ + 130, + 463, + 482, + 498 + ], + "type": "text", + "content": "Fig. 7: Signed error maps of Rabbit L. PISR-PN leverages the perspective polarimetric constraint to eliminate surface cracks encountered by PISR-ON. Additional error maps can be found in our supplementary material." + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "image_caption" + } + ], + "index": 4 + }, + { + "bbox": [ + 130, + 537, + 482, + 586 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 537, + 482, + 586 + ], + "spans": [ + { + "bbox": [ + 130, + 537, + 482, + 586 + ], + "type": "text", + "content": "racy by leveraging the perspective polarimetric constraint, eliminating the surface cracks caused by the traditional orthographic one. Besides, PISR regularizes surface normals in image space for robustness and integrates the multi-resolution hash grid for efficiency." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 130, + 594, + 482, + 666 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 594, + 482, + 666 + ], + "spans": [ + { + "bbox": [ + 130, + 594, + 482, + 666 + ], + "type": "text", + "content": "Limitations: Despite the high reconstruction accuracy in the experiments, PISR's performance degenerates with noisy AoP maps, limiting its application in reconstructing rough, well-textured or mirror-like objects. Recent works [16,18] have shown solutions to mitigate this problem. In addition, PISR also requires dense input views, adding shape constraints derived from normal and depth estimation models would be a promising solution." + } + ] + } + ], + "index": 7 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "spans": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "type": "text", + "content": "14" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 166, + 91, + 225, + 100 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 166, + 91, + 225, + 100 + ], + "spans": [ + { + "bbox": [ + 166, + 91, + 225, + 100 + ], + "type": "text", + "content": "G. Chen et al." + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 13 + }, + { + "para_blocks": [ + { + "bbox": [ + 133, + 114, + 197, + 126 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 114, + 197, + 126 + ], + "spans": [ + { + "bbox": [ + 133, + 114, + 197, + 126 + ], + "type": "text", + "content": "References" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 138, + 139, + 480, + 665 + ], + "type": "list", + "angle": 0, + "index": 19, + "blocks": [ + { + "bbox": [ + 138, + 139, + 480, + 161 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 139, + 480, + 161 + ], + "spans": [ + { + "bbox": [ + 138, + 139, + 480, + 161 + ], + "type": "text", + "content": "1. Atkinson, G.A., Hancock, E.R.: Recovery of surface orientation from diffuse polarization. IEEE transactions on image processing 15(6), 1653-1664 (2006)" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 138, + 162, + 480, + 194 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 162, + 480, + 194 + ], + "spans": [ + { + "bbox": [ + 138, + 162, + 480, + 194 + ], + "type": "text", + "content": "2. Ba, Y., Gilbert, A., Wang, F., Yang, J., Chen, R., Wang, Y., Yan, L., Shi, B., Kadambi, A.: Deep shape from polarization. In: European Conference on Computer Vision. pp. 554-571. Springer (2020)" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 138, + 195, + 480, + 226 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 195, + 480, + 226 + ], + "spans": [ + { + "bbox": [ + 138, + 195, + 480, + 226 + ], + "type": "text", + "content": "3. Baek, S.H., Zeltner, T., Ku, H., Hwang, I., Tong, X., Jakob, W., Kim, M.H.: Image-based acquisition and modeling of polarimetric reflectance. ACM Trans. Graph. 39(4), 139 (2020)" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 138, + 228, + 480, + 259 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 228, + 480, + 259 + ], + "spans": [ + { + "bbox": [ + 138, + 228, + 480, + 259 + ], + "type": "text", + "content": "4. Chen, G., He, L., Guan, Y., Zhang, H.: Perspective phase angle model for polarimetric 3d reconstruction. In: European Conference on Computer Vision. pp. 398-414. Springer (2022)" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 138, + 261, + 480, + 293 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 261, + 480, + 293 + ], + "spans": [ + { + "bbox": [ + 138, + 261, + 480, + 293 + ], + "type": "text", + "content": "5. Cui, Z., Gu, J., Shi, B., Tan, P., Kautz, J.: Polarimetric multi-view stereo. In: Proceedings of the IEEE conference on computer vision and pattern recognition. pp. 1558-1567 (2017)" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 138, + 293, + 480, + 325 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 293, + 480, + 325 + ], + "spans": [ + { + "bbox": [ + 138, + 293, + 480, + 325 + ], + "type": "text", + "content": "6. Dave, A., Zhao, Y., Veeraraghavan, A.: Pandora: Polarization-aided neural decomposition of radiance. In: European Conference on Computer Vision. pp. 538-556. Springer (2022)" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 138, + 326, + 480, + 347 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 326, + 480, + 347 + ], + "spans": [ + { + "bbox": [ + 138, + 326, + 480, + 347 + ], + "type": "text", + "content": "7. Ding, Y., Ji, Y., Ye, J.: Polar-photometric stereo under natural illumination. In: 2022 International Conference on 3D Vision (3DV). pp. 690–699. IEEE (2022)" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 138, + 348, + 480, + 380 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 348, + 480, + 380 + ], + "spans": [ + { + "bbox": [ + 138, + 348, + 480, + 380 + ], + "type": "text", + "content": "8. Ding, Y., Ji, Y., Zhou, M., Kang, S.B., Ye, J.: Polarimetric helmholtz stereopsis. In: Proceedings of the IEEE/CVF International Conference on Computer Vision. pp. 5037-5046 (2021)" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 138, + 381, + 480, + 413 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 381, + 480, + 413 + ], + "spans": [ + { + "bbox": [ + 138, + 381, + 480, + 413 + ], + "type": "text", + "content": "9. Ge, W., Hu, T., Zhao, H., Liu, S., Chen, Y.C.: Ref-neus: Ambiguity-reduced neural implicit surface learning for multi-view reconstruction with reflection. arXiv preprint arXiv:2303.10840 (2023)" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 138, + 414, + 480, + 445 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 414, + 480, + 445 + ], + "spans": [ + { + "bbox": [ + 138, + 414, + 480, + 445 + ], + "type": "text", + "content": "0. Gropp, A., Yariv, L., Haim, N., Atzmon, M., Lipman, Y.: Implicit geometric regularization for learning shapes. In: Proceedings of the 37th International Conference on Machine Learning. pp. 3789-3799 (2020)" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 138, + 447, + 480, + 479 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 447, + 480, + 479 + ], + "spans": [ + { + "bbox": [ + 138, + 447, + 480, + 479 + ], + "type": "text", + "content": "1. Hwang, I., Jeon, D.S., Munoz, A., Gutierrez, D., Tong, X., Kim, M.H.: Sparse ellipsometry: portable acquisition of polarimetric svbrdf and shape with unstructured flash photography. ACM Transactions on Graphics (TOG) 41(4), 1-14 (2022)" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 138, + 479, + 480, + 522 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 479, + 480, + 522 + ], + "spans": [ + { + "bbox": [ + 138, + 479, + 480, + 522 + ], + "type": "text", + "content": "2. Ichikawa, T., Fukao, Y., Nobuhara, S., Nishino, K.: Fresnel microfacet brdf: Unification of polari-radiometric surface-body reflection. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 16489-16497 (2023)" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 138, + 523, + 480, + 555 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 523, + 480, + 555 + ], + "spans": [ + { + "bbox": [ + 138, + 523, + 480, + 555 + ], + "type": "text", + "content": "3. Jakob, W., Speierer, S., Roussel, N., Nimier-David, M., Vicini, D., Zeltner, T., Nicolet, B., Crespo, M., Leroy, V., Zhang, Z.: Mitsuba 3 renderer (2022), https://mitsuba-renderer.org" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 138, + 556, + 480, + 599 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 556, + 480, + 599 + ], + "spans": [ + { + "bbox": [ + 138, + 556, + 480, + 599 + ], + "type": "text", + "content": "4. Kondo, Y., Ono, T., Sun, L., Hirasawa, Y., Murayama, J.: Accurate polarimetric brdf for real polarization scene rendering. In: Computer Vision-ECCV 2020: 16th European Conference, Glasgow, UK, August 23-28, 2020, Proceedings, Part XIX 16. pp. 220-236. Springer (2020)" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 138, + 600, + 480, + 632 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 600, + 480, + 632 + ], + "spans": [ + { + "bbox": [ + 138, + 600, + 480, + 632 + ], + "type": "text", + "content": "5. Lei, C., Qi, C., Xie, J., Fan, N., Koltun, V., Chen, Q.: Shape from polarization for complex scenes in the wild. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR). pp. 12632-12641 (June 2022)" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 138, + 633, + 480, + 665 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 633, + 480, + 665 + ], + "spans": [ + { + "bbox": [ + 138, + 633, + 480, + 665 + ], + "type": "text", + "content": "6. Li, C., Ono, T., Uemori, T., Mihara, H., Gatto, A., Nagahara, H., Moriuchi, Y.: Neisf: Neural incident stokes field for geometry and material estimation. arXiv preprint arXiv:2311.13187 (2023)" + } + ] + } + ], + "index": 18 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 272, + 91, + 447, + 102 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 272, + 91, + 447, + 102 + ], + "spans": [ + { + "bbox": [ + 272, + 91, + 447, + 102 + ], + "type": "text", + "content": "PISR for Textureless and Specular Objects" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 470, + 92, + 480, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 470, + 92, + 480, + 100 + ], + "spans": [ + { + "bbox": [ + 470, + 92, + 480, + 100 + ], + "type": "text", + "content": "15" + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 14 + }, + { + "para_blocks": [ + { + "bbox": [ + 132, + 116, + 481, + 666 + ], + "type": "list", + "angle": 0, + "index": 18, + "blocks": [ + { + "bbox": [ + 133, + 116, + 481, + 160 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 116, + 481, + 160 + ], + "spans": [ + { + "bbox": [ + 133, + 116, + 481, + 160 + ], + "type": "text", + "content": "17. Li, Z., Müller, T., Evans, A., Taylor, R.H., Unberath, M., Liu, M.Y., Lin, C.H.: Neuralangelo: High-fidelity neural surface reconstruction. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 8456-8465 (2023)" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 133, + 160, + 481, + 193 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 160, + 481, + 193 + ], + "spans": [ + { + "bbox": [ + 133, + 160, + 481, + 193 + ], + "type": "text", + "content": "18. Liu, Y., Wang, P., Lin, C., Long, X., Wang, J., Liu, L., Komura, T., Wang, W.: Nano: Neural geometry and brdf reconstruction of reflective objects from multiview images. In: SIGGRAPH (2023)" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 133, + 193, + 481, + 225 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 193, + 481, + 225 + ], + "spans": [ + { + "bbox": [ + 133, + 193, + 481, + 225 + ], + "type": "text", + "content": "19. Long, X., Lin, C., Wang, P., Komura, T., Wang, W.: Sparseneus: Fast generalizable neural surface reconstruction from sparse views. In: European Conference on Computer Vision. pp. 210-227. Springer (2022)" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 132, + 225, + 481, + 246 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 225, + 481, + 246 + ], + "spans": [ + { + "bbox": [ + 132, + 225, + 481, + 246 + ], + "type": "text", + "content": "20. Lorensen, W.E., Cline, H.E.: Marching cubes: A high resolution 3d surface construction algorithm. ACM SIGGRAPH Computer Graphics 21(4), 163-169 (1987)" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 132, + 246, + 481, + 278 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 246, + 481, + 278 + ], + "spans": [ + { + "bbox": [ + 132, + 246, + 481, + 278 + ], + "type": "text", + "content": "21. Lyu, Y., Zhao, L., Li, S., Shi, B.: Shape from polarization with distant lighting estimation. IEEE Transactions on Pattern Analysis and Machine Intelligence 45(11), 13991-14004 (2023). https://doi.org/10.1109/TPAMI.2023.3298376" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 132, + 278, + 481, + 321 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 278, + 481, + 321 + ], + "spans": [ + { + "bbox": [ + 132, + 278, + 481, + 321 + ], + "type": "text", + "content": "22. Martin-Brualla, R., Radwan, N., Sajjadi, M.S., Barron, J.T., Dosovitskiy, A., Duckworth, D.: Nerf in the wild: Neural radiance fields for unconstrained photo collections. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 7210–7219 (2021)" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 132, + 321, + 481, + 354 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 321, + 481, + 354 + ], + "spans": [ + { + "bbox": [ + 132, + 321, + 481, + 354 + ], + "type": "text", + "content": "23. Mildenhall, B., Srinivasan, P.P., Tancik, M., Barron, J.T., Ramamoorthi, R., Ng, R.: Nerf: Representing scenes as neural radiance fields for view synthesis. Communications of the ACM 65(1), 99-106 (2021)" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 132, + 354, + 481, + 397 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 354, + 481, + 397 + ], + "spans": [ + { + "bbox": [ + 132, + 354, + 481, + 397 + ], + "type": "text", + "content": "24. Miyazaki, D., Shigetomi, T., Baba, M., Furukawa, R., Hiura, S., Asada, N.: Polarization-based surface normal estimation of black specular objects from multiple viewpoints. In: 2012 Second International Conference on 3D Imaging, Modeling, Processing, Visualization & Transmission. pp. 104-111. IEEE (2012)" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 132, + 397, + 481, + 430 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 397, + 481, + 430 + ], + "spans": [ + { + "bbox": [ + 132, + 397, + 481, + 430 + ], + "type": "text", + "content": "25. Miyazaki, D., Tan, R.T., Hara, K., Ikeuchi, K.: Polarization-based inverse rendering from a single view. In: Computer Vision, IEEE International Conference on. vol. 3, pp. 982-982. IEEE Computer Society (2003)" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 132, + 430, + 481, + 460 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 430, + 481, + 460 + ], + "spans": [ + { + "bbox": [ + 132, + 430, + 481, + 460 + ], + "type": "text", + "content": "26. Müller, T., Evans, A., Schied, C., Keller, A.: Instant neural graphics primitives with a multiresolution hash encoding. ACM Transactions on Graphics (ToG) 41(4), 1-15 (2022)" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 132, + 460, + 481, + 494 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 460, + 481, + 494 + ], + "spans": [ + { + "bbox": [ + 132, + 460, + 481, + 494 + ], + "type": "text", + "content": "27. Ngo Thanh, T., Nagahara, H., Taniguchi, R.i.: Shape and light directions from shading and polarization. In: Proceedings of the IEEE conference on computer vision and pattern recognition. pp. 2310-2318. IEEE (2015)" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 132, + 494, + 481, + 525 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 494, + 481, + 525 + ], + "spans": [ + { + "bbox": [ + 132, + 494, + 481, + 525 + ], + "type": "text", + "content": "28. Niemeyer, M., Mescheder, L., Oechsle, M., Geiger, A.: Differentiable volumetric rendering: Learning implicit 3d representations without 3d supervision. In: Proc. IEEE Conf. on Computer Vision and Pattern Recognition (CVPR) (2020)" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 132, + 525, + 481, + 557 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 525, + 481, + 557 + ], + "spans": [ + { + "bbox": [ + 132, + 525, + 481, + 557 + ], + "type": "text", + "content": "29. Oechsle, M., Peng, S., Geiger, A.: Unisurf: Unifying neural implicit surfaces and radiance fields for multi-view reconstruction. In: International Conference on Computer Vision (ICCV) (2021)" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 132, + 557, + 481, + 601 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 557, + 481, + 601 + ], + "spans": [ + { + "bbox": [ + 132, + 557, + 481, + 601 + ], + "type": "text", + "content": "30. Park, J.J., Florence, P., Straub, J., Newcombe, R., Lovegrove, S.: Deepsdf: Learning continuous signed distance functions for shape representation. In: Proceedings of the IEEE/CVF conference on computer vision and pattern recognition. pp. 165-174 (2019)" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 132, + 601, + 481, + 643 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 601, + 481, + 643 + ], + "spans": [ + { + "bbox": [ + 132, + 601, + 481, + 643 + ], + "type": "text", + "content": "31. Rahmann, S., Canterakis, N.: Reconstruction of specular surfaces using polarization imaging. In: Proceedings of the 2001 IEEE Computer Society Conference on Computer Vision and Pattern Recognition. CVPR 2001. vol. 1, pp. I-I. IEEE (2001)" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 132, + 643, + 481, + 666 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 643, + 481, + 666 + ], + "spans": [ + { + "bbox": [ + 132, + 643, + 481, + 666 + ], + "type": "text", + "content": "32. Sarlin, P.E., DeTone, D., Malisiewicz, T., Rabinovich, A.: SuperGlue: Learning feature matching with graph neural networks. In: CVPR (2020)" + } + ] + } + ], + "index": 17 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "spans": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "type": "text", + "content": "16" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 166, + 91, + 225, + 100 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 166, + 91, + 225, + 100 + ], + "spans": [ + { + "bbox": [ + 166, + 91, + 225, + 100 + ], + "type": "text", + "content": "G. Chen et al." + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 15 + }, + { + "para_blocks": [ + { + "bbox": [ + 132, + 117, + 480, + 665 + ], + "type": "list", + "angle": 0, + "index": 18, + "blocks": [ + { + "bbox": [ + 132, + 117, + 480, + 138 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 117, + 480, + 138 + ], + "spans": [ + { + "bbox": [ + 132, + 117, + 480, + 138 + ], + "type": "text", + "content": "33. Schonberger, J.L., Frahm, J.M.: Structure-from-motion revisited. In: Conference on Computer Vision and Pattern Recognition (CVPR) (2016)" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 132, + 140, + 480, + 171 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 140, + 480, + 171 + ], + "spans": [ + { + "bbox": [ + 132, + 140, + 480, + 171 + ], + "type": "text", + "content": "34. Shakeri, M., Loo, Y., Zhang, H.: Polarimetric monocular dense mapping using relative deep depth prior [j]. 2021. Shakeri M Loo Y Zhang H Polarimetric Monocular Dense Mapping Using Relative Deep Depth Prior (2021)" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 133, + 172, + 480, + 205 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 172, + 480, + 205 + ], + "spans": [ + { + "bbox": [ + 133, + 172, + 480, + 205 + ], + "type": "text", + "content": "35. Shao, M., Xia, C., Yang, Z., Huang, J., Wang, X.: Transparent shape from a single view polarization image. In: Proceedings of the IEEE/CVF International Conference on Computer Vision (ICCV). pp. 9277-9286 (October 2023)" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 132, + 205, + 480, + 237 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 205, + 480, + 237 + ], + "spans": [ + { + "bbox": [ + 132, + 205, + 480, + 237 + ], + "type": "text", + "content": "36. Sony: Polarization image sensor with four-directional on-chip polarizer and global shutter function, https://www.sony-semicon.co.jp/e/products/IS/industry/product/polarization.html, accessed 2024-02-04" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 132, + 238, + 480, + 270 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 238, + 480, + 270 + ], + "spans": [ + { + "bbox": [ + 132, + 238, + 480, + 270 + ], + "type": "text", + "content": "37. Tian, C., Pan, W., Wang, Z., Mao, M., Zhang, G., Bao, H., Tan, P., Cui, Z.: Dpsnet: Deep polarimetric stereo depth estimation. In: The IEEE/CVF International Conference on Computer Vision (ICCV) (2023)" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 132, + 271, + 480, + 303 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 271, + 480, + 303 + ], + "spans": [ + { + "bbox": [ + 132, + 271, + 480, + 303 + ], + "type": "text", + "content": "38. Tozza, S., Zhu, D., Smith, W., Ramamoorthi, R., Hancock, E.: Uncalibrated, two source photo-polarimetric stereo. IEEE Transactions on Pattern Analysis and Machine Intelligence (2021)" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 132, + 304, + 480, + 346 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 304, + 480, + 346 + ], + "spans": [ + { + "bbox": [ + 132, + 304, + 480, + 346 + ], + "type": "text", + "content": "39. Verbin, D., Hedman, P., Mildenhall, B., Zickler, T., Barron, J.T., Srinivasan, P.P.: Ref-nerf: Structured view-dependent appearance for neural radiance fields. In: 2022 IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR). pp. 5481-5490. IEEE (2022)" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 132, + 347, + 480, + 380 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 347, + 480, + 380 + ], + "spans": [ + { + "bbox": [ + 132, + 347, + 480, + 380 + ], + "type": "text", + "content": "40. Wang, F., Rakotosaona, M.J., Niemeyer, M., Szeliski, R., Pollefeys, M., Tombari, F.: Unisdf: Unifying neural representations for high-fidelity 3d reconstruction of complex scenes with reflections. In: arXiv (2023)" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 132, + 380, + 480, + 413 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 380, + 480, + 413 + ], + "spans": [ + { + "bbox": [ + 132, + 380, + 480, + 413 + ], + "type": "text", + "content": "41. Wang, H., Wang, J., Agapito, L.: Co-slam: Joint coordinate and sparse parametric encodings for neural real-time slam. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 13293-13302 (2023)" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 132, + 414, + 480, + 445 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 414, + 480, + 445 + ], + "spans": [ + { + "bbox": [ + 132, + 414, + 480, + 445 + ], + "type": "text", + "content": "42. Wang, P., Liu, L., Liu, Y., Theobalt, C., Komura, T., Wang, W.: Neus: Learning neural implicit surfaces by volume rendering for multi-view reconstruction. Advances in Neural Information Processing Systems 34, 27171-27183 (2021)" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 132, + 447, + 480, + 489 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 447, + 480, + 489 + ], + "spans": [ + { + "bbox": [ + 132, + 447, + 480, + 489 + ], + "type": "text", + "content": "43. Wang, Y., Han, Q., Habermann, M., Daniilidis, K., Theobalt, C., Liu, L.: Neus2: Fast learning of neural implicit surfaces for multi-view reconstruction. In: Proceedings of the IEEE/CVF International Conference on Computer Vision. pp. 3295-3306 (2023)" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 132, + 490, + 480, + 533 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 490, + 480, + 533 + ], + "spans": [ + { + "bbox": [ + 132, + 490, + 480, + 533 + ], + "type": "text", + "content": "44. Xie, Z., Yang, X., Yang, Y., Sun, Q., Jiang, Y., Wang, H., Cai, Y., Sun, M.: S3im: Stochastic structural similarity and its unreasonable effectiveness for neural fields. In: Proceedings of the IEEE/CVF International Conference on Computer Vision. pp. 18024-18034 (2023)" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 132, + 534, + 480, + 566 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 534, + 480, + 566 + ], + "spans": [ + { + "bbox": [ + 132, + 534, + 480, + 566 + ], + "type": "text", + "content": "45. Xu, Q., Tao, W.: Multi-scale geometric consistency guided multi-view stereo. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR) (June 2019)" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 132, + 567, + 480, + 599 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 567, + 480, + 599 + ], + "spans": [ + { + "bbox": [ + 132, + 567, + 480, + 599 + ], + "type": "text", + "content": "46. Yang, L., Tan, F., Li, A., Cui, Z., Furukawa, Y., Tan, P.: Polarimetric dense monocular slam. In: Proceedings of the IEEE conference on computer vision and pattern recognition. pp. 3857-3866 (2018)" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 132, + 600, + 480, + 632 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 600, + 480, + 632 + ], + "spans": [ + { + "bbox": [ + 132, + 600, + 480, + 632 + ], + "type": "text", + "content": "47. Yariv, L., Gu, J., Kasten, Y., Lipman, Y.: Volume rendering of neural implicit surfaces. In: Thirty-Fifth Conference on Neural Information Processing Systems (2021)" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 132, + 632, + 480, + 665 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 632, + 480, + 665 + ], + "spans": [ + { + "bbox": [ + 132, + 632, + 480, + 665 + ], + "type": "text", + "content": "48. Yariv, L., Kasten, Y., Moran, D., Galun, M., Atzmon, M., Ronen, B., Lipman, Y.: Multiview neural surface reconstruction by disentangling geometry and appearance. Advances in Neural Information Processing Systems 33 (2020)" + } + ] + } + ], + "index": 17 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 272, + 91, + 447, + 102 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 272, + 91, + 447, + 102 + ], + "spans": [ + { + "bbox": [ + 272, + 91, + 447, + 102 + ], + "type": "text", + "content": "PISR for Textureless and Specular Objects" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 470, + 92, + 480, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 470, + 92, + 480, + 100 + ], + "spans": [ + { + "bbox": [ + 470, + 92, + 480, + 100 + ], + "type": "text", + "content": "17" + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 16 + }, + { + "para_blocks": [ + { + "bbox": [ + 132, + 116, + 481, + 370 + ], + "type": "list", + "angle": 0, + "index": 10, + "blocks": [ + { + "bbox": [ + 133, + 116, + 481, + 149 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 116, + 481, + 149 + ], + "spans": [ + { + "bbox": [ + 133, + 116, + 481, + 149 + ], + "type": "text", + "content": "49. Yu, Y., Zhu, D., Smith, W.A.: Shape-from-polarisation: a nonlinear least squares approach. In: Proceedings of the IEEE International Conference on Computer Vision Workshops. pp. 2969-2976 (2017)" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 132, + 150, + 481, + 183 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 150, + 481, + 183 + ], + "spans": [ + { + "bbox": [ + 132, + 150, + 481, + 183 + ], + "type": "text", + "content": "50. Yu, Z., Peng, S., Niemeyer, M., Sattler, T., Geiger, A.: Monosdf: Exploring monocular geometric cues for neural implicit surface reconstruction. Advances in neural information processing systems 35, 25018-25032 (2022)" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 132, + 183, + 481, + 226 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 183, + 481, + 226 + ], + "spans": [ + { + "bbox": [ + 132, + 183, + 481, + 226 + ], + "type": "text", + "content": "51. Zhang, K., Luan, F., Wang, Q., Bala, K., Snavely, N.: PhySG: Inverse rendering with spherical gaussians for physics-based material editing and relighting. In: The IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR) (2021)" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 132, + 227, + 481, + 248 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 227, + 481, + 248 + ], + "spans": [ + { + "bbox": [ + 132, + 227, + 481, + 248 + ], + "type": "text", + "content": "52. Zhang, K., Riegler, G., Snavely, N., Koltun, V.: Nerf++: Analyzing and improving neural radiance fields. arXiv:2010.07492 (2020)" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 132, + 249, + 481, + 282 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 249, + 481, + 282 + ], + "spans": [ + { + "bbox": [ + 132, + 249, + 481, + 282 + ], + "type": "text", + "content": "53. Zhang, X., Srinivasan, P.P., Deng, B., Debevec, P., Freeman, W.T., Barron, J.T.: Nerfactor: Neural factorization of shape and reflectance under an unknown illumination. ACM Transactions on Graphics (ToG) 40(6), 1-18 (2021)" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 132, + 282, + 481, + 304 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 282, + 481, + 304 + ], + "spans": [ + { + "bbox": [ + 132, + 282, + 481, + 304 + ], + "type": "text", + "content": "54. Zhao, J., Monno, Y., Okutomi, M.: Polarimetric multi-view inverse rendering. IEEE Transactions on Pattern Analysis and Machine Intelligence (2022)" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 132, + 304, + 481, + 336 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 304, + 481, + 336 + ], + "spans": [ + { + "bbox": [ + 132, + 304, + 481, + 336 + ], + "type": "text", + "content": "55. Zhao, J., Oishi, J., Monno, Y., Okutomi, M.: Polarimetric patchmatch multi-view stereo. In: Proceedings of the IEEE/CVF Winter Conference on Applications of Computer Vision. pp. 3476-3484 (2024)" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 132, + 336, + 481, + 370 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 336, + 481, + 370 + ], + "spans": [ + { + "bbox": [ + 132, + 336, + 481, + 370 + ], + "type": "text", + "content": "56. Zheng, E., Dunn, E., Jojic, V., Frahm, J.M.: Patchmatch based joint view selection and depthmap estimation. In: Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition (CVPR) (June 2014)" + } + ] + } + ], + "index": 9 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "spans": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "type": "text", + "content": "18" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 166, + 91, + 225, + 101 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 166, + 91, + 225, + 101 + ], + "spans": [ + { + "bbox": [ + 166, + 91, + 225, + 101 + ], + "type": "text", + "content": "G. Chen et al." + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 17 + } + ], + "_backend": "vlm", + "_version_name": "2.6.4" +} \ No newline at end of file diff --git a/2024/PLOT_ Text-based Person Search with Part Slot Attention for Corresponding Part Discovery/aafb2e8f-8788-410d-8a77-952604d06d27_content_list.json b/2024/PLOT_ Text-based Person Search with Part Slot Attention for Corresponding Part Discovery/aafb2e8f-8788-410d-8a77-952604d06d27_content_list.json new file mode 100644 index 0000000000000000000000000000000000000000..0a98d71b5bac8a1eec06f0dcb4f6582629900c48 --- /dev/null +++ b/2024/PLOT_ Text-based Person Search with Part Slot Attention for Corresponding Part Discovery/aafb2e8f-8788-410d-8a77-952604d06d27_content_list.json @@ -0,0 +1,1741 @@ +[ + { + "type": "text", + "text": "PLOT: Text-based Person Search with Part Slot Attention for Corresponding Part Discovery", + "text_level": 1, + "bbox": [ + 218, + 140, + 782, + 186 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Jicheol Park D, Dongwon Kim D, Boseung Jeong D, and Suha Kwak D", + "bbox": [ + 245, + 213, + 754, + 228 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Pohang University of Science and Technology (POSTECH), South Korea {jicheol, kdwon,boseung01, suha.kwak}@postech.ac.kr https://cvlab.postech.ac.kr/research/PLOT", + "bbox": [ + 256, + 239, + 743, + 282 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Abstract. Text-based person search, employing free-form text queries to identify individuals within a vast image collection, presents a unique challenge in aligning visual and textual representations, particularly at the human part level. Existing methods often struggle with part feature extraction and alignment due to the lack of direct part-level supervision and reliance on heuristic features. We propose a novel framework that leverages a part discovery module based on slot attention to autonomously identify and align distinctive parts across modalities, enhancing interpretability and retrieval accuracy without explicit part-level correspondence supervision. Additionally, text-based dynamic part attention adjusts the importance of each part, further improving retrieval outcomes. Our method is evaluated on three public benchmarks, significantly outperforming existing methods.", + "bbox": [ + 261, + 325, + 743, + 505 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Keywords: Text-Based Person Search $\\cdot$ Multi-Modal Retrieval", + "bbox": [ + 261, + 518, + 697, + 534 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "1 Introduction", + "text_level": 1, + "bbox": [ + 215, + 561, + 375, + 578 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Text-based person search is the task of identifying the target person from the vast collection of images with a free-form text query. This task demands extracting identifiable features, such as human parts, from both textual and visual modalities to capture subtle differences between individuals. Hence, establishing correspondence between the extracted human part features across image and text modalities is essential for accurate text-based person search. However, it is not straightforward to extract these part features and establish their correspondences between the two modalities without part-level supervision.", + "bbox": [ + 212, + 597, + 787, + 718 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "To address this challenge, previous work [2, 22] relies on the heuristic part features obtained by equi-horizontally cropping the entire image; such features are then matched to the free-form text queries for person search. However, the heuristic part features used in this approach are susceptible to deformation caused by occlusion and pose variation. Meanwhile, earlier work [8, 20, 24] proposed a learning-based approach for part feature extraction. Nevertheless, these methods tend to generate redundant part features that lack disentanglement [20], or they demand access to additional part-level supervisions [8, 24].", + "bbox": [ + 212, + 719, + 787, + 840 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/8a390316999a60ab9111737922f170dec25ac14a89cbf5747f3a9421ce3c96f6.jpg", + "image_caption": [ + "Fig. 1: The overall architecture of PLOT." + ], + "image_footnote": [], + "bbox": [ + 218, + 147, + 787, + 363 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "To tackle the above issues, we introduce a new framework that discovers distinctive parts in both modalities and matches them between the two modalities without any correspondence supervision; its overall architecture is illustrated in Fig. 1. To discover distinctive parts and extract their features from both modalities, we propose PLOT, a Part discovery module based on the sLOT attention mechanism [15]. Slot attention is an attention mechanism designed for object-centric learning, which segments input data into a set of slots representing individual entities without requiring object-level supervision. In PLOT, we first define a set of learnable embedding vectors, termed part slots, that contain primitive information related to human body parts shared between the two modalities. Then, these part slots undergo refinement through several iterative attention processes, where they compete amongst themselves to bind with the input data; ultimately, the part slots are transformed into part features, termed part embeddings, that represent distinctive parts in the input data.", + "bbox": [ + 212, + 409, + 787, + 621 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "To ensure correspondence between part embeddings from different modalities, the part slots are shared between the visual and textual modalities. Part embeddings from the same part slots are then learned to represent the identical part although they are computed from different modalities. This mechanism enables PLOT to match the discovered parts from the two modalities without supervision for the correspondence as well as capturing part-level fine-grained appearance features from the both modalities. Hence, PLOT improves performance of text-based person search through the rich and fine-grained part features, and at the same time, it guarantees interpretable retrieval by providing part-level correspondences between query text and retrieved images.", + "bbox": [ + 212, + 625, + 787, + 776 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "In addition, PLOT introduces a new similarity aggregation method for part embeddings called text-based dynamic part attention (TDPA) pooling, which dynamically adjusts the weights of part embeddings based on the text query. Given a text query as input, TDPA predicts the importance weight of each", + "bbox": [ + 212, + 779, + 789, + 840 + ], + "page_idx": 1 + }, + { + "type": "page_number", + "text": "2", + "bbox": [ + 217, + 114, + 228, + 126 + ], + "page_idx": 1 + }, + { + "type": "header", + "text": "J. Park et al.", + "bbox": [ + 271, + 114, + 359, + 126 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "slot for retrieval and applies the predicted importance weights to aggregate the similarities of part embeddings between the two modalities. TDPA allows the entire retrieval system to perform optimized retrieval for each query, leading to improved performance.", + "bbox": [ + 212, + 146, + 784, + 205 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Our method was evaluated and compared with prior work on three public benchmarks [4, 11, 32], where it clearly outperformed all existing methods thanks to the rich representation based on part embeddings. The main contribution of our work is four-fold:", + "bbox": [ + 212, + 207, + 785, + 266 + ], + "page_idx": 2 + }, + { + "type": "list", + "sub_type": "text", + "list_items": [ + "- We introduce PLOT, a new framework for text-based person search that discovers distinguished human body parts, extracts their embeddings, and establishes their correspondences between the two modalities with no human intervention.", + "- The part embeddings provided by PLOT enables an interpretable text-based person search thanks to the part-level correspondences it provides.", + "- We introduce a novel similarity aggregation method that adaptively determines importance of each discovered part based on each text query and consequently enables retrieval optimized per query.", + "- Our model with PLOT achieved the best on all the public benchmarks for text-based person search." + ], + "bbox": [ + 223, + 277, + 784, + 440 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "2 Related Work", + "text_level": 1, + "bbox": [ + 215, + 465, + 387, + 481 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "2.1 Text-Based Person Search", + "text_level": 1, + "bbox": [ + 214, + 497, + 477, + 511 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "In recent years, the task of text-based person search has gained significant attention in the computer vision community. Li et al., [11] proposed a gated neural attention-based recurrent neural network (GNA-RNN) for learning the affinity between text descriptions and images, along with providing a benchmark dataset CUHK-PEDES for model evaluation. Zhang et al., [30] proposed cross-modal projection matching and classification (CMPM+CMPC) loss, for learning deep discriminative image-text global embeddings. However, these methods primarily focus on global representations of input data and are thus not capable of capturing distinctive human part details, leading to limited performance in the text-based person search.", + "bbox": [ + 212, + 522, + 787, + 672 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "To address the above problem, a line of studies focuses on extracting fine-grained representations. One of the prominent examples of exploiting fine-grained information is to cut human images horizontally and use them as human parts [2, 4, 5, 16, 22, 23]. Chen et al., [2] extracts image part embeddings through equi-horizontal cropping of the entire image and aligns additional network to transform the textual global into corresponding parts. However, these heuristic part features inevitably include no-informative information, such as background elements. To avoid the above limitations, Suo et al., [22] proposed the simple and robust correlation filtering method to extract foreground features on the heuristic part features. Yet, the heuristically divided part features fundamentally fall short of capturing the complex human parts, underscoring a critical limitation in their", + "bbox": [ + 212, + 674, + 787, + 840 + ], + "page_idx": 2 + }, + { + "type": "header", + "text": "PLOT: TBPS with Part Slot Attention for Corresponding Part Discovery", + "bbox": [ + 241, + 114, + 730, + 128 + ], + "page_idx": 2 + }, + { + "type": "page_number", + "text": "3", + "bbox": [ + 774, + 116, + 784, + 126 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "expressive capability. To move beyond such heuristic part structures, Shao et al., [2] proposes a learning-based approach for part feature extraction. However, this method struggles to extract distinctive part embeddings due to the extraction of redundant part information, which lacks sufficient disentanglement. To extract exquisite human parts, several studies tried to utilize useful information (e.g., human attributes and human keypoints) via external tools [1,8,24]. Wang et al., [24] introduced an auxiliary attribute segmentation to align the visual part features with the textual attributes parsed from text description. Jing et al., [8] proposed a new multi-granularity attention network to learn the part feature alignment between visual and textual with human pose estimation. However, these approaches have inevitable limitations of high computational cost and dependence on the performance of external tools for local feature extraction.", + "bbox": [ + 212, + 146, + 787, + 325 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Most recent work, Jiang et al., [7] utilizes pre-trained CLIP [18] model, capitalizing on the rich knowledge of models trained on extensive data for text-to-image matching, to excel in text-based person search. Yet, this methods primarily focus on global features, not specifically designed for extracting human part features. Not only do we leverage pre-trained knowledge, but we transcend heuristic human parts methods to extract sophisticated human parts without external tools or part-level supervision.", + "bbox": [ + 212, + 328, + 787, + 434 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "2.2 Slot Attention", + "text_level": 1, + "bbox": [ + 215, + 460, + 380, + 474 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Slot attention [15] is a recently proposed attention mechanism for object-centric learning, a problem focusing on discovering constituent visual entities within an image. The unique property of slot attention is that it can represent input images as a set of slots, where the slots are representations corresponding to individual visual entities, without any object-level supervision during training. Within slot attention, slots iteratively compete for aggregating input data, ensuring distinct representations focusing on individual visual entities. By incorporating slot attention, the proposed framework facilitates the unsupervised identification of a structure underlying image and text queries, enabling the model to discern and represent individual human parts without explicit supervision. This capability is particularly valuable in person search datasets, where recognizing subtle differences and understanding the correspondence of human parts across different modalities are crucial.", + "bbox": [ + 212, + 489, + 787, + 685 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "3 Proposed Method", + "text_level": 1, + "bbox": [ + 215, + 713, + 426, + 729 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "The following subsection offers details of global and part embeddings extraction for each modality with its backbone (Sec 3.1). Subsequently, we present our novel framework, PLOT, which includes the part discovery module (Sec. 3.2) and similarity aggregation between part embeddings (Sec. 3.3), concluding with a discussion on the learning objective designed to optimize our proposed framework (Sec. 3.4) and inference of our framework (Sec. 3.5).", + "bbox": [ + 212, + 750, + 787, + 840 + ], + "page_idx": 3 + }, + { + "type": "page_number", + "text": "4", + "bbox": [ + 217, + 114, + 228, + 126 + ], + "page_idx": 3 + }, + { + "type": "header", + "text": "J. Park et al.", + "bbox": [ + 271, + 114, + 359, + 126 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "3.1 Global and Part Embeddings", + "text_level": 1, + "bbox": [ + 215, + 146, + 501, + 161 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "In our framework, a single input data is described by two different types of representations: a global embedding and multiple part embeddings. The global embedding is used to represent the input data holistically, while each part embedding describes the appearance of distinctive human parts (e.g., arm, leg, torso, etc.). We below provide details of global and part embeddings computation for visual and textual modalities. Following the previous work [7], a pre-trained CLIP [18] is used as the backbone networks for the visual and textual modalities. Visual Modality: We utilize the vision transformer (ViT) from the CLIP-B/16 [18] architecture as a visual backbone network. Initially, an input image of a person is split into $N$ distinct, non-overlapping patches, which are subsequently transformed into patch tokens through linear projection. The patch tokens and an extra [c1s] token are then fed into the visual backbone network. The token sequence is processed throughout multiple self-attention blocks, and the [c1s] token of the last block is used as a global embedding $\\mathbf{g}^{\\mathcal{V}} \\in \\mathbb{R}^{D}$ . To obtain part embeddings $\\mathbf{P}^{\\mathcal{V}} \\in \\mathbb{R}^{K \\times D}$ , remaining patch tokens of last block $\\mathbf{x}^{\\mathcal{V}} \\in \\mathbb{R}^{N \\times D}$ is passed to the part discovery module. The part discovery module aggregates the patch features describing coherent human parts into the same part embedding. We provide more detailed information about the part discovery module in Sec. 3.2. Textual Modality: For the textual backbone network, we utilize a transformer architecture from the CLIP-Xformer [18] text encoder. This encoder operates on text input transformed into byte pair encoding (BPE) sequences. Initially, the text query undergoes tokenization via BPE, followed by the enclosing with [S0S] and [E0S] tokens. The resulting sequence of tokens is then inputted into the textual backbone network. Here, the [E0S] token from the final block serves as the global embedding, denoted as $\\mathbf{g}^{\\mathcal{T}} \\in \\mathbb{R}^{D}$ . Analogous to the approach for the visual modality, we process the remainder of the text tokens, $\\mathbf{x}^{\\mathcal{T}} \\in \\mathbb{R}^{L \\times D}$ , into part embeddings, $\\mathbf{P}^{\\mathcal{T}} \\in \\mathbb{R}^{K \\times D}$ , utilizing the part discovery module.", + "bbox": [ + 212, + 167, + 787, + 575 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "3.2 Part Discovery Module", + "text_level": 1, + "bbox": [ + 215, + 594, + 454, + 609 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "For extracting part embeddings in each modality, part discovery module aggregates patch tokens $\\mathbf{x}^{\\mathcal{V}}$ and text tokens $\\mathbf{x}^{\\mathcal{T}}$ into the visual part embeddings $\\mathbf{P}^{\\mathcal{V}}$ and textual part embeddings $\\mathbf{P}^{\\mathcal{T}}$ , respectively. It is worth noting that the part discovery module for each modality has identical model architecture and functions equivalently. Therefore, we will only explain the part discovery module on the visual modality for brevity.", + "bbox": [ + 212, + 616, + 785, + 705 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Part discovery module consists of initial part slots $\\mathbf{S}^0\\in \\mathbb{R}^{K\\times D}$ and $T$ multiple iteration of the part slot attention block (PSA block). To extract part embeddings, we first initialize a set of learnable embeddings part slots $\\mathbf{S}^0\\in \\mathbb{R}^{K\\times D}$ , where $K$ indicates the number of part slots. Then, through a series of $T$ iterations of our PSA block, the initial $\\mathbf{S}^0$ evolves into refined $\\mathbf{S}^T$ , where each slot captures distinct parts within input data. The refined part slots $\\mathbf{S}^T$ are used as visual part embeddings $\\mathbf{P}^{\\mathcal{V}}$ . The part discovery module can be formulated as follows:", + "bbox": [ + 212, + 705, + 787, + 811 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\n\\mathbf {P} ^ {\\mathcal {V}} := \\mathbf {S} ^ {T}, \\text {w h e r e} \\quad \\mathbf {S} ^ {t} = \\operatorname {P S A} _ {-} \\operatorname {B l o c k} ^ {\\mathcal {V}} \\left(\\mathbf {x} ^ {\\mathcal {V}}; \\mathbf {S} ^ {t - 1}\\right). \\tag {1}\n$$\n", + "text_format": "latex", + "bbox": [ + 338, + 818, + 785, + 835 + ], + "page_idx": 4 + }, + { + "type": "header", + "text": "PLOT: TBPS with Part Slot Attention for Corresponding Part Discovery", + "bbox": [ + 241, + 114, + 730, + 128 + ], + "page_idx": 4 + }, + { + "type": "page_number", + "text": "5", + "bbox": [ + 774, + 116, + 784, + 126 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "The PSA block first transforms the inputs $\\mathbf{S}^{t - 1}$ and $\\mathbf{x}^{\\mathcal{V}}$ with layer normalization and linear projection layers $q(\\cdot), k(\\cdot)$ and $\\nu (\\cdot)$ to obtain embeddings of $D_{h}$ dimension. Then the attention map $A\\in \\mathbb{R}^{N\\times K}$ between $\\mathbf{S}^{t - 1}$ and $\\mathbf{x}^{\\mathcal{V}}$ is computed by", + "bbox": [ + 212, + 145, + 787, + 205 + ], + "page_idx": 5 + }, + { + "type": "equation", + "text": "\n$$\nA _ {n, k} = \\frac {e ^ {M _ {n , k}}}{\\sum_ {i = 1} ^ {K} e ^ {M _ {n , i}}}, \\mathrm {w h e r e} M = \\frac {k (\\mathbf {x} ^ {\\mathcal {V}}) q (\\mathbf {S} ^ {t - 1}) ^ {\\top}}{\\sqrt {D _ {h}}}. (2)\n$$\n", + "text_format": "latex", + "bbox": [ + 334, + 205, + 787, + 242 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "The attention map is obtained through normalization across part slots; this normalization encourages competition among themselves to bind distinct sets of patch tokens to each slot. We update part slots with weighted mean of patch tokens obtained by attention map $A_{n,k}$ and then feed it to a gated recurrent unit (GRU) using $\\mathbf{S}^{t-1}$ as hidden state as follows:", + "bbox": [ + 212, + 244, + 787, + 321 + ], + "page_idx": 5 + }, + { + "type": "equation", + "text": "\n$$\n\\bar {A} _ {n, k} = \\frac {A _ {n , k}}{\\sum_ {i = 1} ^ {N} A _ {i , k}}, \\bar {\\mathbf {S}} ^ {t} = \\mathbb {G R U} \\left(\\mathbf {S} ^ {t - 1}, \\bar {A} ^ {\\top} v \\left(\\mathbf {x} ^ {\\mathcal {V}}\\right)\\right). \\tag {3}\n$$\n", + "text_format": "latex", + "bbox": [ + 348, + 332, + 787, + 367 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Then, we obtain $t$ -th part slot by feeding $\\overline{\\mathbf{S}}^{t-1}$ into a multi-layer perceptron (MLP) with layer normalization, ReLU activation, and residual connection:", + "bbox": [ + 212, + 376, + 785, + 407 + ], + "page_idx": 5 + }, + { + "type": "equation", + "text": "\n$$\n\\mathbf {S} ^ {t} = \\operatorname {P S A} _ {\\text {B l o c k}} ^ {\\mathcal {V}} (\\mathbf {x} ^ {\\mathcal {V}}; \\mathbf {S} ^ {t - 1}) = \\operatorname {M L P} \\left(\\bar {\\mathbf {S}} ^ {t - 1}\\right) + \\bar {\\mathbf {S}} ^ {t - 1}. \\tag {4}\n$$\n", + "text_format": "latex", + "bbox": [ + 333, + 417, + 787, + 435 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Finally, we can obtain the visual part embeddings $\\mathbf{P}^{\\mathcal{V}} = \\{\\mathbf{p}_k^{\\mathcal{V}}\\}_{k=1}^K$ which is the output of $T$ -th iteration of PSA block: $\\mathbf{P}^{\\mathcal{V}} = \\mathbf{S}_{\\mathcal{V}}^T \\in \\mathbb{R}^{K \\times D}$ .", + "bbox": [ + 212, + 443, + 785, + 477 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Part Correspondence through Slot Sharing: Additionally, we share the learnable part slots $\\mathbf{S}^0$ between two part discovery modules to establish correspondences between part embeddings extracted from each modality, considering that part embeddings extracted from the same part slot are corresponding part across modalities, thereby contributing to a clearer comparison between modalities.", + "bbox": [ + 212, + 477, + 787, + 553 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "3.3 Measuring Similarity between Embeddings", + "text_level": 1, + "bbox": [ + 214, + 574, + 614, + 589 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "In the context of training and applying our retrieval model, selecting an appropriate similarity function between embeddings is crucial. The challenge lies in dealing with two distinct types of embeddings: global embeddings and part embeddings. For global embeddings, cosine similarity offers a straightforward and effective means of measuring similarity.", + "bbox": [ + 212, + 598, + 785, + 672 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "However, the situation becomes more complex when considering part embeddings. A direct method for addressing this complexity involves calculating the average cosine similarity across all pairs of part embeddings. This approach, while straightforward, has its drawbacks, primarily because it treats all part-wise similarities as equally significant. In reality, the relevance of specific part embeddings to the actual similarity between data instances can significantly vary, influenced by the context of a text query. For instance, if a text query focuses exclusively on particular features of a human figure, the similarity contributions from other unrelated part embeddings should be less relevant. This challenge highlights the necessity for an approach that can dynamically assess and prioritize the relevance of part embeddings based on the context provided by the query.", + "bbox": [ + 212, + 674, + 787, + 840 + ], + "page_idx": 5 + }, + { + "type": "page_number", + "text": "6", + "bbox": [ + 217, + 114, + 228, + 126 + ], + "page_idx": 5 + }, + { + "type": "header", + "text": "J. Park et al.", + "bbox": [ + 271, + 114, + 359, + 127 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "To address this challenge, we introduce text-based dynamic part attention (TDPA) to aggregate the similarities between part embeddings. Firstly, textual global embedding $\\mathbf{g}^{\\mathcal{T}}$ is transformed to the TDPA $\\mathbf{a} \\in \\mathbb{R}^{K}$ , using MLP and a softmax function. Using TDPA, the aggregated similarity between part embeddings are computed as follows:", + "bbox": [ + 212, + 146, + 789, + 220 + ], + "page_idx": 6 + }, + { + "type": "equation", + "text": "\n$$\n\\mathbf {a} = \\sigma (\\operatorname {M L P} \\left(\\mathbf {g} ^ {\\mathcal {T}}\\right)) \\in \\mathbb {R} ^ {K}, \\tag {5}\n$$\n", + "text_format": "latex", + "bbox": [ + 364, + 229, + 787, + 247 + ], + "page_idx": 6 + }, + { + "type": "equation", + "text": "\n$$\nc _ {\\mathrm {a g g}} \\left(\\mathbf {P} ^ {\\mathcal {V}}, \\mathbf {P} ^ {\\mathcal {T}}; \\mathbf {g} ^ {\\mathcal {T}}\\right) := \\sum_ {k = 1} ^ {K} \\mathrm {a} _ {k} \\cdot c \\left(\\mathbf {p} _ {k} ^ {\\mathcal {V}}, \\mathbf {p} _ {k} ^ {\\mathcal {T}}\\right), \\tag {6}\n$$\n", + "text_format": "latex", + "bbox": [ + 364, + 252, + 787, + 291 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "where $c(\\cdot, \\cdot)$ denotes cosine similarity between two embeddings, $a_k$ is the $k$ -th value of $\\mathbf{a}$ , and $\\sigma(\\cdot)$ is a softmax function. If given textual global embedding $\\mathbf{g}^{\\mathcal{T}}$ , TDPA is computed by MLP which is learned in an end-to-end manner by minimizing our partNCE loss, as will be introduced in Sec 3.4. This dynamic attention enables us to adaptively assign importance to each part embedding, which reduces the impact of non-informative part embedding and promotes a more informative similarity measurement that reflects the significant semantic similarity of each text query.", + "bbox": [ + 212, + 297, + 787, + 420 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "3.4 Learning Objective", + "text_level": 1, + "bbox": [ + 214, + 439, + 419, + 455 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Our model is trained through the establishment of cross-modal alignments, considering both global and part levels. The alignment at the global-level is accomplished by aligning global embeddings $\\mathbf{g}^{\\mathcal{V}}$ and $\\mathbf{g}^{\\mathcal{T}}$ that capture the comprehensive information of each modality. For part-level alignment, we leverage part embeddings $\\mathbf{P}^{\\mathcal{V}}$ and $\\mathbf{P}^{\\mathcal{T}}$ that have locally exclusive features within each modality due to the slot attention. Furthermore, by dynamically adjusting weights of specific part slots based on the text global embedding $\\mathbf{g}^{\\mathcal{T}}$ , we facilitate the learning of more informative alignments.", + "bbox": [ + 212, + 460, + 787, + 582 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Global Alignment Loss: To align global embeddings $\\mathbf{g}^{\\mathcal{V}}$ and $\\mathbf{g}^{\\mathcal{T}}$ extracted from each modality backbone, we first define a batch of global embeddings $\\mathcal{B}_{\\mathrm{global}} = \\{(\\mathbf{g}_i^{\\mathcal{V}},\\mathbf{g}_i^{\\mathcal{T}})\\}_{i = 1}^B$ , where $B$ is batch size. Then we adopt the InfoNCE loss [17] which is a contrastive learning objective that maximizes the similarity between embeddings of positive pairs while minimizing the similarity between negative pairs in the batch. Consequently, our global NCE loss with cosine similarity function $c(\\cdot ,\\cdot)$ is formulated as follows:", + "bbox": [ + 214, + 582, + 787, + 688 + ], + "page_idx": 6 + }, + { + "type": "equation", + "text": "\n$$\n\\mathcal {L} _ {\\mathrm {N C E}} = - \\sum_ {i = 1} ^ {B} \\left(\\log \\frac {e ^ {c \\left(\\mathbf {g} _ {i} ^ {\\mathcal {V}} , \\mathbf {g} _ {i} ^ {\\mathcal {T}}\\right) / \\tau}}{\\sum_ {j = 1} ^ {B} e ^ {c \\left(\\mathbf {g} _ {i} ^ {\\mathcal {V}} , \\mathbf {g} _ {j} ^ {\\mathcal {T}}\\right) / \\tau}} + \\log \\frac {e ^ {c \\left(\\mathbf {g} _ {i} ^ {\\mathcal {V}} , \\mathbf {g} _ {i} ^ {\\mathcal {T}}\\right) / \\tau}}{\\sum_ {j = 1} ^ {B} e ^ {c \\left(\\mathbf {g} _ {j} ^ {\\mathcal {V}} , \\mathbf {g} _ {i} ^ {\\mathcal {T}}\\right) / \\tau}}\\right), \\tag {7}\n$$\n", + "text_format": "latex", + "bbox": [ + 295, + 696, + 787, + 744 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "where the $\\tau$ is temperature term. Additionally, we employ an identity classification loss $\\mathcal{L}_{\\mathrm{ID}}$ to ensure that the embeddings extracted from the same identity become similar. The $\\mathcal{L}_{\\mathrm{ID}}$ is denoted by", + "bbox": [ + 212, + 750, + 787, + 796 + ], + "page_idx": 6 + }, + { + "type": "equation", + "text": "\n$$\n\\mathcal {L} _ {\\mathrm {I D}} = - \\sum_ {i = 1} ^ {B} \\left(\\boldsymbol {y} _ {i} \\log \\sigma \\left(\\mathbf {g} _ {i} ^ {\\mathcal {V}} \\mathbf {W} _ {\\mathrm {I D}}\\right) + \\boldsymbol {y} _ {i} \\log \\sigma \\left(\\mathbf {g} _ {i} ^ {\\mathcal {T}} \\mathbf {W} _ {\\mathrm {I D}}\\right)\\right), \\tag {8}\n$$\n", + "text_format": "latex", + "bbox": [ + 316, + 804, + 787, + 842 + ], + "page_idx": 6 + }, + { + "type": "header", + "text": "PLOT: TBPS with Part Slot Attention for Corresponding Part Discovery", + "bbox": [ + 240, + 114, + 732, + 130 + ], + "page_idx": 6 + }, + { + "type": "page_number", + "text": "7", + "bbox": [ + 774, + 114, + 785, + 127 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "where $\\mathbf{y}_i\\in \\mathbb{R}^C$ is the identity ground truth of corresponding global embedding $\\mathbf{g}_i$ represented by a one-hot vector and $C$ is the number of identities, $\\mathbf{W}_{\\mathrm{ID}}\\in \\mathbb{R}^{D\\times C}$ is a classifier shared between the two modalities. For cross-modal alignments, the commonly used CMPM loss [7,30] is additionally adopted. Finally, we describe global alignment loss as below:", + "bbox": [ + 212, + 146, + 787, + 223 + ], + "page_idx": 7 + }, + { + "type": "equation", + "text": "\n$$\n\\mathcal {L} _ {\\text {G l o b a l}} = \\mathcal {L} _ {\\text {N C E}} + \\mathcal {L} _ {\\text {I D}} + \\mathcal {L} _ {\\text {C M P M}}. \\tag {9}\n$$\n", + "text_format": "latex", + "bbox": [ + 379, + 234, + 787, + 250 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Part Alignment Loss: Similar to the alignment of global embeddings, we first define a batch of part embeddings $\\mathcal{B}_{\\mathrm{part}} = \\{(\\mathbf{P}_i^{\\mathcal{V}},\\mathbf{P}_i^{\\mathcal{T}})\\}_{i = 1}^B$ with batch size $B$ , and we adopt the InfoNCE loss to align part embeddings extracted from two modalities. For learning text query-based informative alignment of part embeddings between modalities, we employ text query-based similarity aggregation function $c_{\\mathrm{agg}}$ in Eq. 6 to compute InfoNCE loss. We termed this InfoNCE with $c_{\\mathrm{agg}}$ as PartNCE loss, and it is formulated by", + "bbox": [ + 214, + 261, + 789, + 367 + ], + "page_idx": 7 + }, + { + "type": "equation", + "text": "\n$$\n\\mathcal {L} _ {\\mathrm {P a r t N C E}} = - \\sum_ {i = 1} ^ {B} \\left(\\log \\frac {e ^ {c _ {\\mathrm {a g g}} (\\mathbf {P} _ {i} ^ {\\mathcal {V}} , \\mathbf {P} _ {i} ^ {\\mathcal {T}} ; \\mathbf {g} _ {i} ^ {\\mathcal {T}}) / \\tau}}{\\sum_ {j = 1} ^ {B} e ^ {c _ {\\mathrm {a g g}} (\\mathbf {P} _ {i} ^ {\\mathcal {V}} , \\mathbf {P} _ {j} ^ {\\mathcal {T}} ; \\mathbf {g} _ {j} ^ {\\mathcal {T}}) / \\tau}} + \\log \\frac {e ^ {c _ {\\mathrm {a g g}} (\\mathbf {P} _ {i} ^ {\\mathcal {V}} , \\mathbf {P} _ {i} ^ {\\mathcal {T}} ; \\mathbf {g} _ {i} ^ {\\mathcal {T}}) / \\tau}}{\\sum_ {j = 1} ^ {B} e ^ {c _ {\\mathrm {a g g}} \\left(\\mathbf {P} _ {j} ^ {\\mathcal {V}}, \\mathbf {P} _ {i} ^ {\\mathcal {T}}; \\mathbf {g} _ {i} ^ {\\mathcal {T}}\\right) / \\tau}\\right), \\tag {10}\n$$\n", + "text_format": "latex", + "bbox": [ + 228, + 377, + 787, + 428 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Similar to global alignment loss, we adopt the identity loss that shares classifier weights between the two modalities for part alignment; however, the difference is that the part embeddings are concatenated along the embedding dimension. The identity loss for part alignment is formulated by", + "bbox": [ + 214, + 435, + 787, + 497 + ], + "page_idx": 7 + }, + { + "type": "equation", + "text": "\n$$\n\\mathcal {L} _ {\\text {P a r t I D}} = - \\sum_ {i = 1} ^ {B} \\left(\\boldsymbol {y} _ {i} \\log \\sigma \\left(\\left[ \\boldsymbol {P} _ {i} ^ {\\mathcal {V}} \\right] \\boldsymbol {W} _ {\\text {P a r t I D}}\\right) + \\boldsymbol {y} _ {i} \\log \\sigma \\left(\\left[ \\boldsymbol {P} _ {i} ^ {\\mathcal {T}} \\right] \\boldsymbol {W} _ {\\text {P a r t I D}}\\right)\\right), \\tag {11}\n$$\n", + "text_format": "latex", + "bbox": [ + 256, + 507, + 787, + 547 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "where $[\\mathbf{P}] \\in \\mathbb{R}^{KD}$ is the concatenation of part embeddings $\\mathbf{P}$ along the embedding dimension, and $\\mathbf{W}_{\\mathrm{PartID}} \\in \\mathbb{R}^{KD \\times C}$ is a classifier shared between both modalities. Finally, the part alignment loss is computed by", + "bbox": [ + 214, + 559, + 789, + 604 + ], + "page_idx": 7 + }, + { + "type": "equation", + "text": "\n$$\n\\mathcal {L} _ {\\text {P a r t}} = \\mathcal {L} _ {\\text {P a r t N C E}} + \\mathcal {L} _ {\\text {P a r t I D}}. \\tag {12}\n$$\n", + "text_format": "latex", + "bbox": [ + 395, + 617, + 787, + 633 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Cross-Modal Masked Language Modeling Loss: Following the practices of employing Transformer-based backbones in existing methods [7,21], we adopt an auxiliary loss, the cross-modal masked language modeling (CMLM) loss, to facilitate the learning of interactions between modali", + "bbox": [ + 214, + 657, + 517, + 763 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/d6bcbcab86d9b112f25ee0477689121305b220c886e6338817255b1b29ca810c.jpg", + "image_caption": [ + "Cross-Modal Masked Language Model (CMLM)", + "Fig. 2: Illustration of CMLM." + ], + "image_footnote": [], + "bbox": [ + 532, + 650, + 777, + 734 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "ties. Similar to BERT [3], given a text description, we randomly select text tokens with a $15\\%$ probability and replace them with the learnable [MASK] token. The masked text description is then processed through the text backbone to obtain the masked textual tokens. After obtaining the masked textual tokens, we concatenate them with the visual tokens extracted via the image backbone.", + "bbox": [ + 212, + 763, + 787, + 840 + ], + "page_idx": 7 + }, + { + "type": "page_number", + "text": "8", + "bbox": [ + 217, + 114, + 228, + 126 + ], + "page_idx": 7 + }, + { + "type": "header", + "text": "J. Park et al.", + "bbox": [ + 271, + 114, + 359, + 127 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "The concatenated tokens are then fed into a transformer to acquire cross-modal fused tokens. Among the fused tokens, those corresponding to the indices of textual tokens are denoted as $\\mathbf{F}$ composed as $\\{\\mathbf{f}_1,\\dots \\mathbf{f}_L\\}$ , where a fused token $\\mathbf{f}\\in \\mathbb{R}^{1\\times D}$ and $L$ indicates the max length of input text descriptions. Ultimately, these tokens are fed into a CMLM classifier $\\mathbf{W}_{\\mathrm{CMLM}}$ to predict the probability of vocabulary IDs. The overall procedure of CMLM is illustrated in Fig. 2 and through this procedure we compute the $\\mathcal{L}_{\\mathrm{CMLM}}$ loss as follows:", + "bbox": [ + 212, + 146, + 787, + 253 + ], + "page_idx": 8 + }, + { + "type": "equation", + "text": "\n$$\n\\mathcal {L} _ {\\mathrm {C M L M}} = - \\frac {1}{L} \\sum_ {l = 1} ^ {L} \\mathbf {y} _ {l} \\log (\\sigma (\\mathbf {f} _ {l} \\mathbf {W} _ {\\mathrm {C M L M}})), \\tag {13}\n$$\n", + "text_format": "latex", + "bbox": [ + 362, + 263, + 787, + 303 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "where $\\mathbf{W}_{\\mathrm{CMLM}} \\in \\mathbb{R}^{D \\times V}$ and $\\pmb{y}_l \\in \\mathbb{R}^V$ is the vocabulary ground truth of $l$ -th text tokens, represented by a one-hot vector and $V$ is the size of vocabulary. Finally, our overall objective function for training is denoted by", + "bbox": [ + 212, + 314, + 789, + 361 + ], + "page_idx": 8 + }, + { + "type": "equation", + "text": "\n$$\n\\mathcal {L} = \\mathcal {L} _ {\\text {G l o b a l}} + \\mathcal {L} _ {\\text {P a r t}} + \\mathcal {L} _ {\\text {C M L M}}. \\tag {14}\n$$\n", + "text_format": "latex", + "bbox": [ + 392, + 375, + 787, + 388 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "3.5 Inference", + "text_level": 1, + "bbox": [ + 215, + 411, + 339, + 425 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "During testing, the global and part embeddings of each modality input are fully exploited to calculate the similarity between the image-text pair. In particular, the similarities between visual part embeddings and their corresponding textual part embeddings are linearly combined with the attention weights $\\mathbf{a}$ to aggregate them. The image-text pair similarity is defined as the sum of the similarity between the global embeddings of the image-text pair and the similarity between the part embeddings of it, which can be computed by $c(\\mathbf{g}^{\\mathcal{V}},\\mathbf{g}^{\\mathcal{T}}) + c_{\\mathrm{agg}}(\\mathbf{P}^{\\mathcal{V}},\\mathbf{P}^{\\mathcal{T}};\\mathbf{g}^{\\mathcal{T}})$ . Finally, given the text query, the images in the gallery are ranked according to similarity scores between the images and the text for inference.", + "bbox": [ + 212, + 435, + 789, + 571 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "4 Experiments", + "text_level": 1, + "bbox": [ + 215, + 595, + 375, + 612 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "In this section, we provide a detailed account of our experimental setup (Sec. 4.1), evaluate our method, and compare it with state of the arts on three benchmark datasets for text-based person search (Sec. 4.2). Furthermore, we qualitatively present retrieval results and analyze the effectiveness of the part discovery module and TDPA with visualization results (Sec. 4.3). We also conduct ablation studies on the losses employed in model training, the methodologies of part discovery, and the strategies for part similarity aggregation (Sec. 4.4).", + "bbox": [ + 212, + 626, + 787, + 733 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "4.1 Experimental Setup", + "text_level": 1, + "bbox": [ + 215, + 753, + 428, + 771 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "Datasets: On three benchmark datasets, CUHK-PEDES [11], ICFG-PEDES [4], and RSTPReid [32], we evaluate and compare the performance of our method against previous methods. In CUHK-PEDES collected from five existing person re-identification datasets [6,12,13,28,31], it contains 40,206 images corresponding", + "bbox": [ + 212, + 779, + 787, + 842 + ], + "page_idx": 8 + }, + { + "type": "header", + "text": "PLOT: TBPS with Part Slot Attention for Corresponding Part Discovery", + "bbox": [ + 241, + 114, + 732, + 130 + ], + "page_idx": 8 + }, + { + "type": "page_number", + "text": "9", + "bbox": [ + 774, + 116, + 784, + 126 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "to 13,003 individual IDs, with each image being approximately matched with two annotated text descriptions. We follow the data split of [11] with 34,054 images from 11,003 person IDs and 68,126 text descriptions for training, 3,078 images from 1,000 IDs and 6,158 text descriptions for validation, and 3,074 images from 1,000 IDs and 6,156 text descriptions for testing. The remaining two datasets are collected from MSMT17 [26]. ICFG-PEDES consists of 54,522 image-text pairs from 4,102 individual IDs, which are split into 34,674 and 19,848 for training and testing, respectively. RSTPReid contains 20,505 images of 4,101 individual IDs, with each ID having 5 images and each image associated with the corresponding two annotated text descriptions. We follow the data split of [32] with 18,505 images from 3,701 IDs and 37,010 text descriptions for training, 1,000 images from 200 IDs and 2,000 text descriptions for validation, and 1,000 images from 200 IDs and 2,000 text descriptions for testing, respectively.", + "bbox": [ + 212, + 146, + 787, + 343 + ], + "page_idx": 9 + }, + { + "type": "text", + "text": "Evaluation Protocol: We employ the standard metric of rank at K (R@K=1,5, 10) for all retrieval experiments. Specifically, given a query text, images are sorted based on their similarity to the query text. The search is considered correct if at least one relevant image appears in the top K positions of the ranking.", + "bbox": [ + 212, + 343, + 787, + 404 + ], + "page_idx": 9 + }, + { + "type": "text", + "text": "Network Architecture: We adopt the pre-trained CLIP models from OpenAI [18] for both image and text encoders, where the size of the image encoder is ViT-B/16. The input images are resized to $384 \\times 128$ . Random horizontal flipping, random cropping, and random erasing are applied for the data augmentation in training time. The maximum text length is set to 77.", + "bbox": [ + 212, + 404, + 787, + 478 + ], + "page_idx": 9 + }, + { + "type": "text", + "text": "Network Optimization: Our model is trained using the Adam optimizer for 60 epochs with a batch size of 128 for all experiments. For the CLIP encoders, the initial learning rate is set to $5\\mathrm{e} - 6$ , using a cosine schedule with the warm-up strategy at the first five epochs; we use a high learning rate for the remaining parameters by scaling 20 times.", + "bbox": [ + 212, + 479, + 787, + 554 + ], + "page_idx": 9 + }, + { + "type": "text", + "text": "Hyperparameters: We set the number of part slots to 8. The number of iterations in the part slot attention block is set to 5. The temperature parameter $\\tau$ is set to 0.015.", + "bbox": [ + 212, + 554, + 787, + 598 + ], + "page_idx": 9 + }, + { + "type": "text", + "text": "4.2 Quantitative Results", + "text_level": 1, + "bbox": [ + 214, + 619, + 433, + 635 + ], + "page_idx": 9 + }, + { + "type": "text", + "text": "We compare our method with previous text-based person search methods on CUHK-PEDES [11], ICFG-PEDES [4], and RSTPReid [32]. The performance comparison and the backbones for each modality employed by each method are shown in Table 1. Specifically, our method achieves an outstanding $\\mathbb{R}@\\mathbb{1}$ metric of $75.28\\%$ , $65.76\\%$ , and $61.80\\%$ on the CUHK-PEDES, ICFG-PEDES, and RSTPReid datasets, respectively, outperforming previous methods utilizing different backbones from ours. Moreover, our method improves the previous state of the art, IRRA [7], on $\\mathbb{R}@\\mathbb{1}$ by a large margin as $1.9\\% \\mathbb{p}$ , $1.7\\% \\mathbb{p}$ , and $1.6\\% \\mathbb{p}$ , respectively. Since IRRA only focuses on aligning the global embeddings of each modality, it is hard to capture the fine-grained differences. In contrast, our method not only takes account of the global embeddings but also aims to discover and align discriminative part embeddings from each modality; it allows the model to effectively find the target person.", + "bbox": [ + 212, + 643, + 787, + 840 + ], + "page_idx": 9 + }, + { + "type": "page_number", + "text": "10", + "bbox": [ + 217, + 114, + 235, + 126 + ], + "page_idx": 9 + }, + { + "type": "header", + "text": "J. Park et al.", + "bbox": [ + 271, + 114, + 359, + 126 + ], + "page_idx": 9 + }, + { + "type": "table", + "img_path": "images/f8e7e1bca7d6fde188ba5cee8e5e2996f7c250bda42d91c40189a9b69af832ea.jpg", + "table_caption": [ + "Table 1: Performance of text-based person search methods on the three datasets. Bold and underline denote the best and the second best." + ], + "table_footnote": [], + "table_body": "
MethodsBackboneCUHK-PEDESICFG-PEDESRSTPReid
ImageTextR@1R@5R@10R@1R@5R@10R@1R@5R@10
GNA-RNN [11]RN50LSTM19.05-53.64------
CMPM/C [30]RN50LSTM49.3771.6979.2743.5165.4474.26---
PMA [8]RN50BERT53.8173.5481.23------
TIMAM [19]RN101BERT54.5177.5684.78------
SCAN [9]RN50BERT55.8675.9783.6950.0569.6577.21---
ViTAA [24]RN50LSTM55.9775.8483.5250.9868.7975.78---
NAFS [5]RN50BERT59.9479.8686.70------
DSSL [32]RN50BERT59.9880.4187.56---32.4355.0863.19
MGEL [23]RN50LSTM60.2780.0186.74------
SSAN [4]RN50LSTM61.3780.1586.7354.2372.6379.5343.5067.8077.15
LapsCore [27]RN50BERT63.40-87.80------
SRCF [22]RN50BERT64.0482.9988.8157.1875.0181.49---
LGUR [20]RN50BERT64.2181.9487.9357.4274.9781.45---
TIPCB [2]RN50BERT64.2683.1989.10------
CAIBC [25]RN50BERT64.4382.8788.37---47.3569.5579.00
SAF [10]ViT-B/16BERT64.1382.6288.40------
IVT [21]ViT-B/16BERT65.5983.1189.2156.0473.6080.2246.7070.0078.80
CFine [29]CLIP-ViT-B/16BERT69.5785.9391.1560.8375.5582.4250.5572.5081.60
IRRA [7]CLIP-ViT-B/16CLIP-Xformer73.3889.9393.7163.4680.2485.8260.2081.3088.20
OursCLIP-ViT-B/16CLIP-Xformer75.2890.4294.1265.7681.3986.7361.8082.8589.45
", + "bbox": [ + 215, + 184, + 782, + 439 + ], + "page_idx": 10 + }, + { + "type": "text", + "text": "4.3 Qualitative Results", + "text_level": 1, + "bbox": [ + 215, + 468, + 419, + 483 + ], + "page_idx": 10 + }, + { + "type": "text", + "text": "Retrieval Results: Top-5 retrieval results of our method on the CUHK-PEDES dataset are illustrated in Fig. 3. Above all, it shows the overall satisfactory retrieval results. In particular, we can observe our model retrieves targets well, even with distinctive human parts that are small or located in various positions. For instance, the small distinctive human parts like \"ponytails\" and \"high skirts\" in Fig. 3(a), \"black shoes\" in Fig. 3(b), and \"blue plaid shorts\" in Fig. 3(c), as well as human parts that could appear in various viewpoints such as \"yellow shoulder bags\" in Fig. 3(a), \"floaty dresses\" in Fig. 3(b), and \"red backpacks\" in Fig. 3(d). The CUHK-PEDES dataset typically contains three target images on average in the search space, thereby most of the false matches in the figure are observed by a lack of additional targets. Despite the false matching due to the limitation of the dataset, the retrieval results are reasonable in that the retrieved false matching contains distinctive human parts described by the query description.", + "bbox": [ + 212, + 492, + 787, + 688 + ], + "page_idx": 10 + }, + { + "type": "text", + "text": "Visualization of Attention Map $\\bar{A}_k$ in PSABlock: To demonstrate the effectiveness of our part discovery module, we visualize the attention map $\\bar{A}_k$ in $T$ -th iteration of the PSA block for both visual and textual modalities (in Eq. (2)). The visualization results are illustrated in Fig. 4. It not only demonstrates that the part embeddings extracted by our part discovery module capture distinctive human parts but also shows that the part embeddings extracted from the same part slot attend to the semantically identical human parts regardless of modality. For example, the 1st part slot typically focuses on bottom clothes, the 4th on footwear, the 5th on objects being held, the 7th on top clothes, and the 8th on the person's head. Furthermore, in Fig. 4 compared (a) and (b) of visual modality,", + "bbox": [ + 212, + 688, + 789, + 840 + ], + "page_idx": 10 + }, + { + "type": "header", + "text": "PLOT: TBPS with Part Slot Attention for Corresponding Part Discovery", + "bbox": [ + 241, + 114, + 730, + 128 + ], + "page_idx": 10 + }, + { + "type": "page_number", + "text": "11", + "bbox": [ + 767, + 116, + 782, + 126 + ], + "page_idx": 10 + }, + { + "type": "text", + "text": "Query: a girl with a ponytail wearing a red top with blue denim thigh high skirt and carrying a yellow shoulder bag looking at a phone.", + "bbox": [ + 241, + 143, + 477, + 179 + ], + "page_idx": 11 + }, + { + "type": "image", + "img_path": "images/cbfd0667d02d7bd6323ed76a4faf2d80e3e8f2baa6e044985e825293d7c7ee61.jpg", + "image_caption": [ + "(a)" + ], + "image_footnote": [], + "bbox": [ + 235, + 181, + 486, + 250 + ], + "page_idx": 11 + }, + { + "type": "text", + "text": "Query: the little boy is wearing his hat backwards. he has on a blue, short sleeved shirt and grey plaid shorts.", + "bbox": [ + 241, + 265, + 483, + 299 + ], + "page_idx": 11 + }, + { + "type": "text", + "text": "Query: the woman is wearing a white cardigan over a long floaty dress and black shoes. she is carrying a back carry-on bag.", + "bbox": [ + 516, + 143, + 750, + 179 + ], + "page_idx": 11 + }, + { + "type": "image", + "img_path": "images/67b8f416eb5048dfab7e4fd5e2cfb1ca67048742e3b933f260eeeb49137f2040.jpg", + "image_caption": [ + "(b)" + ], + "image_footnote": [], + "bbox": [ + 511, + 181, + 764, + 250 + ], + "page_idx": 11 + }, + { + "type": "text", + "text": "Query: person wearing a dark brown sweater, denim cropped pants that go just below the knees, and white sneakers. they are wearing a red backpack over both shoulders.", + "bbox": [ + 516, + 265, + 750, + 309 + ], + "page_idx": 11 + }, + { + "type": "image", + "img_path": "images/6e4c87227f1095c68d6de0a8a4b79ed07c5e5422f0256aec1c9ea1762bed1401.jpg", + "image_caption": [ + "(c)" + ], + "image_footnote": [], + "bbox": [ + 235, + 311, + 486, + 382 + ], + "page_idx": 11 + }, + { + "type": "image", + "img_path": "images/31787daf2f547edde4d967cfa6876118dc51700a5992bd0461d3cee9ca9e9460.jpg", + "image_caption": [ + "(d)", + "Fig. 3: Top-5 retrieval results of our method on the CUHK-PEDES dataset. Images are sorted from left to right according to their ranks below each text query. Green and red boxes indicate true and false matches, respectively." + ], + "image_footnote": [], + "bbox": [ + 511, + 311, + 763, + 382 + ], + "page_idx": 11 + }, + { + "type": "text", + "text": "our part discovery module is capable of capturing the distinctive human parts, while it is robust against pose variations and viewpoint changes.", + "bbox": [ + 212, + 459, + 787, + 489 + ], + "page_idx": 11 + }, + { + "type": "text", + "text": "Visualization of TDPA: We visualize TDPA weights $\\mathbf{a}$ in Eq. (6) to demonstrate the effectiveness of similarity aggregation between part embeddings of two modalities with TDPA. The visualization results are presented in the bottom left in Fig. 4(a, b), respectively. Comparing examples (a) and (b), the 5th part slot typically focuses on human parts associated with held objects, like bags, leading to a predicted lower TDPA weight for this part slot in (b) due to the absence of such distinctive information in its query description. However, the presence of the \"hat\" in the query description of (b) leads to having a high TDPA weight for the 8th part slot that typically focuses on the human head part. In contrast, since this related human head part information is not provided in the query description of (a), the TDPA weight for the 8th part slot in (a) is predicted to be a low value. These observations highlight the capability of TDPA, which adaptively improves part-based retrieval depending on the contents of the text query.", + "bbox": [ + 212, + 489, + 787, + 700 + ], + "page_idx": 11 + }, + { + "type": "text", + "text": "4.4 Ablation Studies", + "text_level": 1, + "bbox": [ + 215, + 724, + 401, + 738 + ], + "page_idx": 11 + }, + { + "type": "text", + "text": "In our ablation studies conducted on the CUHK-PEDES datasets, we evaluate the effectiveness of proposed components and their combinations in improving text-based person search performance. In Table 2, we compare several configurations: The baseline method employs only global embeddings $\\mathbf{g}$ trained by InfoNCE $(\\mathcal{L}_{\\mathrm{NCE}})$ loss. When incorporating the cross-modal masked language modeling $(\\mathcal{L}_{\\mathrm{CMLM}})$ , there is a slight improvement across all metrics. The addition", + "bbox": [ + 212, + 750, + 787, + 840 + ], + "page_idx": 11 + }, + { + "type": "page_number", + "text": "12", + "bbox": [ + 217, + 114, + 235, + 126 + ], + "page_idx": 11 + }, + { + "type": "header", + "text": "J. Park et al.", + "bbox": [ + 271, + 114, + 359, + 126 + ], + "page_idx": 11 + }, + { + "type": "text", + "text": "Text: the woman is wearing a white cardigan over a long floaty dress and black shoes. she is carrying a back carry-on bag.", + "bbox": [ + 251, + 146, + 746, + 174 + ], + "page_idx": 12 + }, + { + "type": "image", + "img_path": "images/a54dd864b9e7b3c752ebe3dc49e32c81f39056b1f34bd7903632f0d11e86113a.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 256, + 175, + 772, + 421 + ], + "page_idx": 12 + }, + { + "type": "text", + "text": "Text: the little boy is wearing his hat backwards. he has on a blue, short sleeved shirt and grey plaid shorts.", + "bbox": [ + 253, + 426, + 751, + 455 + ], + "page_idx": 12 + }, + { + "type": "image", + "img_path": "images/2cb3e51c334281e25db0501e542484791a9ea7dc23a36fbe744dded371a898f7.jpg", + "image_caption": [ + "Fig. 4: Visualization of each modality's attention map $\\bar{A}_k$ in $T$ -th iteration of PSA block and TDPA weights a on CUHK-PEDES dataset." + ], + "image_footnote": [], + "bbox": [ + 251, + 459, + 772, + 714 + ], + "page_idx": 12 + }, + { + "type": "text", + "text": "of part embeddings $\\mathbf{P}$ that without $\\mathcal{L}_{\\mathrm{PartID}}$ also shows significant enhancement in R@1 and R@5, where we observe an increase of $3.46\\% \\mathrm{p}$ and $2.64\\% \\mathrm{p}$ , respectively. Notably, the full configuration achieves the best performance with $\\mathcal{L}_{\\mathrm{PartID}}$ , underscoring the importance of part embeddings for accurate text-based person search.", + "bbox": [ + 212, + 779, + 789, + 840 + ], + "page_idx": 12 + }, + { + "type": "header", + "text": "PLOT: TBPS with Part Slot Attention for Corresponding Part Discovery", + "bbox": [ + 241, + 114, + 730, + 128 + ], + "page_idx": 12 + }, + { + "type": "page_number", + "text": "13", + "bbox": [ + 767, + 114, + 784, + 126 + ], + "page_idx": 12 + }, + { + "type": "table", + "img_path": "images/d2b3e36dac03003270bb7594c1125c67cd3f9111eeaba7167825cd1d030e2c71.jpg", + "table_caption": [ + "Table 2: Ablation studies on the CUHK-PEDES datasets." + ], + "table_footnote": [], + "table_body": "
MethodLossCUHK-PEDES
LNCELIDLCMLMLPartNCELPartIDR@1R@5R@10
Global Only✓XXXX71.39 (+0.0)87.65 (+0.0)92.74 (+0.0)
✓✓XXX71.83 (+0.44)88.06 (+0.41)92.58 (-0.16)
+ CMLM✓✓✓XX72.65 (+1.26)88.58 (+0.93)92.93 (+0.19)
+ Part Embeddings✓✓✓✓X74.85 (+3.46)90.29 (+2.64)94.10 (+1.36)
✓✓✓✓✓75.28 (+3.89)90.42 (+2.77)94.12 (+1.38)
", + "bbox": [ + 218, + 170, + 784, + 276 + ], + "page_idx": 13 + }, + { + "type": "text", + "text": "This suggests that our part embeddings significantly contribute to discerning fine-grained details critical for text-based person search.", + "bbox": [ + 212, + 304, + 782, + 333 + ], + "page_idx": 13 + }, + { + "type": "text", + "text": "Ablation Study on Part Discovery Methods: To validate the effectiveness of our proposed part discovery module, we conduct experiments by replacing our part discovery module with other methods such as TIPCB [2] and PAT [14]. TIPCB is a simple yet effective heuristic part discovery method, which extracts", + "bbox": [ + 212, + 334, + 501, + 454 + ], + "page_idx": 13 + }, + { + "type": "table", + "img_path": "images/7396b1de6649c605a9f245be02236adbc6b6f692bd1416fd2a2671fbbda94395.jpg", + "table_caption": [ + "Table 3: Ablation study of different part discovery methods on the CUHK-PEDES." + ], + "table_footnote": [], + "table_body": "
MethodsR@1R@5R@10
Ours + TIPCB [2]73.2389.1094.04
Ours + PAT [14]72.7689.2393.42
Ours75.2890.4294.12
", + "bbox": [ + 511, + 366, + 789, + 444 + ], + "page_idx": 13 + }, + { + "type": "text", + "text": "image part embeddings through equi-horizontal cropping of the entire image and transforms the global embedding of the text modality into corresponding parts by using additional learnable MLPs for each image part. PAT performs part discovery with a querying transformer approach by leveraging learnable queries and conventional cross-attention. A key difference of our method from PAT is that ours explicitly encourages the discovered parts to be spatially separated since our part slots compete with each other to aggregate input data. In contrast, parts found by PAT, based on conventional cross-attention, often over-capture salient regions and are likely to miss fine details. To demonstrate this limitation, we show the visualization of part attentions with PAT applied to ours in the supplementary materials. As indicated in Table 3, our part discovery method outperforms these heuristic and conventional cross-attention based part discovery methods across all metrics, with the most significant difference observed in the R@1 metric, which requires a precise discernment of human parts.", + "bbox": [ + 212, + 455, + 787, + 667 + ], + "page_idx": 13 + }, + { + "type": "text", + "text": "5 Conclusion", + "text_level": 1, + "bbox": [ + 215, + 688, + 359, + 704 + ], + "page_idx": 13 + }, + { + "type": "text", + "text": "We proposed a novel framework that extracts distinctive human parts corresponding across visual and textual modalities through part discovery module without part-level supervision. The introduced TDPA further refines the retrieval process by adjusting the importance of each part slot based on the contents of the text query, leading to more precise and relevant retrieval.", + "bbox": [ + 212, + 719, + 787, + 794 + ], + "page_idx": 13 + }, + { + "type": "text", + "text": "Limitation: Since slots learn to occupy the entire image and text, some of them may indicate irrelevant part of input. TDPA addresses this issue to some extent, but a more explicit solution would further enhance our system.", + "bbox": [ + 212, + 795, + 787, + 839 + ], + "page_idx": 13 + }, + { + "type": "page_number", + "text": "14", + "bbox": [ + 217, + 114, + 235, + 126 + ], + "page_idx": 13 + }, + { + "type": "header", + "text": "J. Park et al.", + "bbox": [ + 271, + 114, + 359, + 126 + ], + "page_idx": 13 + }, + { + "type": "text", + "text": "Acknowledgements", + "text_level": 1, + "bbox": [ + 217, + 145, + 401, + 162 + ], + "page_idx": 14 + }, + { + "type": "text", + "text": "This work was supported by the IITP grants and the NRF grants funded by Ministry of Science and ICT, Korea (RS-2019-II191906; RS-2022-II220926; NRF-2018R1A5A1060031; NRF-2021R1A2C3012728)", + "bbox": [ + 212, + 176, + 792, + 220 + ], + "page_idx": 14 + }, + { + "type": "text", + "text": "References", + "text_level": 1, + "bbox": [ + 217, + 243, + 321, + 258 + ], + "page_idx": 14 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "1. Aggarwal, S., Radhakrishnan, V.B., Chakraborty, A.: Text-based person search via attribute-aided matching. In: Proc. Winter Conference on Applications of Computer Vision (WACV) (2020)", + "2. Chen, Y., Zhang, G., Lu, Y., Wang, Z., Zheng, Y.: Tipcb: A simple but effective part-based convolutional baseline for text-based person search. Neurocomputing 494 (2022)", + "3. Devlin, J., Chang, M.W., Lee, K., Toutanova, K.: Bert: Pre-training of deep bidirectional transformers for language understanding. In: Annual Conference of the North American Chapter of the Association for Computational Linguistics (2019)", + "4. Ding, Z., Ding, C., Shao, Z., Tao, D.: Semantically self-aligned network for text-to-image part-aware person re-identification. arXiv preprint arXiv:2107.12666 (2021)", + "5. Gao, C., Cai, G., Jiang, X., Zheng, F., Zhang, J., Gong, Y., Peng, P., Guo, X., Sun, X.: Contextual non-local alignment over full-scale representation for text-based person search. arXiv preprint arXiv:2101.03036 (2021)", + "6. Gray, D., Brennan, S., Tao, H.: Evaluating appearance models for recognition, reacquisition, and tracking. In: Proc. IEEE international workshop on performance evaluation for tracking and surveillance (PETS) (2007)", + "7. Jiang, D., Ye, M.: Cross-modal implicit relation reasoning and aligning for text-to-image person retrieval. In: IEEE International Conference on Computer Vision and Pattern Recognition (CVPR) (2023)", + "8. Jing, Y., Si, C., Wang, J., Wang, W., Wang, L., Tan, T.: Pose-guided multi-granularity attention network for text-based person search. In: Proc. AAAI Conference on Artificial Intelligence (AAAI) (2020)", + "9. Lee, K.H., Chen, X., Hua, G., Hu, H., He, X.: Stacked cross attention for image-text matching. In: Proc. European Conference on Computer Vision (ECCV) (2018)", + "10. Li, S., Cao, M., Zhang, M.: Learning semantic-aligned feature representation for text-based person search. In: Proc. IEEE International Conference on Acoustics, Speech and Signal Processing (ICASSP). IEEE (2022)", + "11. Li, S., Xiao, T., Li, H., Zhou, B., Yue, D., Wang, X.: Person search with natural language description. In: Proc. IEEE Conference on Computer Vision and Pattern Recognition (CVPR) (2017)", + "12. Li, W., Zhao, R., Wang, X.: Human reidentification with transferred metric learning. In: Computer Vision-ACCV 2012: 11th Asian Conference on Computer Vision, Daejeon, Korea, November 5-9, 2012, Revised Selected Papers, Part I 11. Springer (2013)", + "13. Li, W., Zhao, R., Xiao, T., Wang, X.: Deepreid: Deep filter pairing neural network for person re-identification. In: Proc. IEEE Conference on Computer Vision and Pattern Recognition (CVPR) (2014)", + "14. Li, Y., He, J., Zhang, T., Liu, X., Zhang, Y., Wu, F.: Diverse part discovery: Occluded person re-identification with part-aware transformer. In: Proc. IEEE Conference on Computer Vision and Pattern Recognition (CVPR) (2021)" + ], + "bbox": [ + 218, + 273, + 785, + 839 + ], + "page_idx": 14 + }, + { + "type": "header", + "text": "PLOT: TBPS with Part Slot Attention for Corresponding Part Discovery", + "bbox": [ + 241, + 114, + 730, + 128 + ], + "page_idx": 14 + }, + { + "type": "page_number", + "text": "15", + "bbox": [ + 767, + 116, + 784, + 126 + ], + "page_idx": 14 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "15. Locatello, F., Weissenborn, D., Unterthiner, T., Mahendran, A., Heigold, G., Uszkoreit, J., Dosovitskiy, A., Kipf, T.: Object-centric learning with slot attention (2020)", + "16. Niu, K., Huang, Y., Ouyang, W., Wang, L.: Improving description-based person re-identification by multi-granularity image-text alignments. IEEE Transactions on Image Processing 29, 5542-5556 (2020)", + "17. Oord, A.v.d., Li, Y., Vinyals, O.: Representation learning with contrastive predictive coding. arXiv preprint arXiv:1807.03748 (2018)", + "18. Radford, A., Kim, J.W., Hallacy, C., Ramesh, A., Goh, G., Agarwal, S., Sastry, G., Askell, A., Mishkin, P., Clark, J., et al.: Learning transferable visual models from natural language supervision. In: Proc. International Conference on Machine Learning (ICML) (2021)", + "19. Sarafianos, N., Xu, X., Kakadiaris, I.A.: Adversarial representation learning for text-to-image matching. In: Proc. IEEE International Conference on Computer Vision (ICCV) (2019)", + "20. Shao, Z., Zhang, X., Fang, M., Lin, Z., Wang, J., Ding, C.: Learning granularity-unified representations for text-to-image person re-identification. In: Proc. ACM Multimedia Conference (ACMMM) (2022)", + "21. Shu, X., Wen, W., Wu, H., Chen, K., Song, Y., Qiao, R., Ren, B., Wang, X.: See finer, see more: Implicit modality alignment for text-based person retrieval. In: Proc. European Conference on Computer Vision Workshop on Real-World Surveillance, (ECCVW) (2022)", + "22. Suo, W., Sun, M., Niu, K., Gao, Y., Wang, P., Zhang, Y., Wu, Q.: A simple and robust correlation filtering method for text-based person search. In: Proc. European Conference on Computer Vision (ECCV) (2022)", + "23. Wang, C., Luo, Z., Lin, Y., Li, S.: Text-based person search via multi-granularity embedding learning. In: Proc. International Joint Conferences on Artificial Intelligence (IJCAI) (2021)", + "24. Wang, Z., Fang, Z., Wang, J., Yang, Y.: Vitaa: Visual-textual attributes alignment in person search by natural language. In: Proc. European Conference on Computer Vision (ECCV) (2020)", + "25. Wang, Z., Zhu, A., Xue, J., Wan, X., Liu, C., Wang, T., Li, Y.: Caibc: Capturing all-round information beyond color for text-based person retrieval. In: Proc. ACM Multimedia Conference (ACMMM) (2022)", + "26. Wei, L., Zhang, S., Gao, W., Tian, Q.: Person transfer gan to bridge domain gap for person re-identification. In: Proc. IEEE Conference on Computer Vision and Pattern Recognition (CVPR) (2018)", + "27. Wu, Y., Yan, Z., Han, X., Li, G., Zou, C., Cui, S.: Lipscore: language-guided person search via color reasoning. In: Proc. IEEE International Conference on Computer Vision (ICCV) (2021)", + "28. Xiao, T., Li, S., Wang, B., Lin, L., Wang, X.: End-to-end deep learning for person search. arXiv preprint arXiv:1604.01850 (2016)", + "29. Yan, S., Dong, N., Zhang, L., Tang, J.: Clip-driven fine-grained text-image person re-identification. IEEE Transactions on Image Processing (2023)", + "30. Zhang, Y., Lu, H.: Deep cross-modal projection learning for image-text matching. In: Proc. European Conference on Computer Vision (ECCV) (2018)", + "31. Zheng, L., Shen, L., Tian, L., Wang, S., Wang, J., Tian, Q.: Scalable person re-identification: A benchmark. In: Proc. IEEE International Conference on Computer Vision (ICCV) (2015)", + "32. Zhu, A., Wang, Z., Li, Y., Wan, X., Jin, J., Wang, T., Hu, F., Hua, G.: Dssl: deep surroundings-person separation learning for text-based person retrieval. In:" + ], + "bbox": [ + 215, + 147, + 785, + 839 + ], + "page_idx": 15 + }, + { + "type": "page_number", + "text": "16", + "bbox": [ + 217, + 114, + 235, + 126 + ], + "page_idx": 15 + }, + { + "type": "header", + "text": "J. Park et al.", + "bbox": [ + 271, + 114, + 359, + 126 + ], + "page_idx": 15 + }, + { + "type": "text", + "text": "Proceedings of the 29th ACM International Conference on Multimedia. pp. 209-217 (2021)", + "bbox": [ + 241, + 147, + 787, + 174 + ], + "page_idx": 16 + }, + { + "type": "header", + "text": "PLOT: TBPS with Part Slot Attention for Corresponding Part Discovery", + "bbox": [ + 241, + 114, + 730, + 128 + ], + "page_idx": 16 + }, + { + "type": "page_number", + "text": "17", + "bbox": [ + 767, + 116, + 784, + 126 + ], + "page_idx": 16 + } +] \ No newline at end of file diff --git a/2024/PLOT_ Text-based Person Search with Part Slot Attention for Corresponding Part Discovery/aafb2e8f-8788-410d-8a77-952604d06d27_model.json b/2024/PLOT_ Text-based Person Search with Part Slot Attention for Corresponding Part Discovery/aafb2e8f-8788-410d-8a77-952604d06d27_model.json new file mode 100644 index 0000000000000000000000000000000000000000..94b01b5d87e6d5e8b986c7bfd16ef9165adfead0 --- /dev/null +++ b/2024/PLOT_ Text-based Person Search with Part Slot Attention for Corresponding Part Discovery/aafb2e8f-8788-410d-8a77-952604d06d27_model.json @@ -0,0 +1,2181 @@ +[ + [ + { + "type": "title", + "bbox": [ + 0.22, + 0.141, + 0.783, + 0.187 + ], + "angle": 0, + "content": "PLOT: Text-based Person Search with Part Slot Attention for Corresponding Part Discovery" + }, + { + "type": "text", + "bbox": [ + 0.246, + 0.214, + 0.756, + 0.229 + ], + "angle": 0, + "content": "Jicheol Park D, Dongwon Kim D, Boseung Jeong D, and Suha Kwak D" + }, + { + "type": "text", + "bbox": [ + 0.258, + 0.241, + 0.745, + 0.283 + ], + "angle": 0, + "content": "Pohang University of Science and Technology (POSTECH), South Korea {jicheol, kdwon,boseung01, suha.kwak}@postech.ac.kr https://cvlab.postech.ac.kr/research/PLOT" + }, + { + "type": "text", + "bbox": [ + 0.262, + 0.326, + 0.744, + 0.506 + ], + "angle": 0, + "content": "Abstract. Text-based person search, employing free-form text queries to identify individuals within a vast image collection, presents a unique challenge in aligning visual and textual representations, particularly at the human part level. Existing methods often struggle with part feature extraction and alignment due to the lack of direct part-level supervision and reliance on heuristic features. We propose a novel framework that leverages a part discovery module based on slot attention to autonomously identify and align distinctive parts across modalities, enhancing interpretability and retrieval accuracy without explicit part-level correspondence supervision. Additionally, text-based dynamic part attention adjusts the importance of each part, further improving retrieval outcomes. Our method is evaluated on three public benchmarks, significantly outperforming existing methods." + }, + { + "type": "text", + "bbox": [ + 0.262, + 0.52, + 0.698, + 0.535 + ], + "angle": 0, + "content": "Keywords: Text-Based Person Search \\(\\cdot\\) Multi-Modal Retrieval" + }, + { + "type": "title", + "bbox": [ + 0.216, + 0.563, + 0.377, + 0.579 + ], + "angle": 0, + "content": "1 Introduction" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.598, + 0.788, + 0.719 + ], + "angle": 0, + "content": "Text-based person search is the task of identifying the target person from the vast collection of images with a free-form text query. This task demands extracting identifiable features, such as human parts, from both textual and visual modalities to capture subtle differences between individuals. Hence, establishing correspondence between the extracted human part features across image and text modalities is essential for accurate text-based person search. However, it is not straightforward to extract these part features and establish their correspondences between the two modalities without part-level supervision." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.72, + 0.788, + 0.841 + ], + "angle": 0, + "content": "To address this challenge, previous work [2, 22] relies on the heuristic part features obtained by equi-horizontally cropping the entire image; such features are then matched to the free-form text queries for person search. However, the heuristic part features used in this approach are susceptible to deformation caused by occlusion and pose variation. Meanwhile, earlier work [8, 20, 24] proposed a learning-based approach for part feature extraction. Nevertheless, these methods tend to generate redundant part features that lack disentanglement [20], or they demand access to additional part-level supervisions [8, 24]." + } + ], + [ + { + "type": "page_number", + "bbox": [ + 0.218, + 0.116, + 0.23, + 0.127 + ], + "angle": 0, + "content": "2" + }, + { + "type": "header", + "bbox": [ + 0.272, + 0.115, + 0.361, + 0.127 + ], + "angle": 0, + "content": "J. Park et al." + }, + { + "type": "image", + "bbox": [ + 0.219, + 0.148, + 0.788, + 0.364 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.36, + 0.377, + 0.642, + 0.391 + ], + "angle": 0, + "content": "Fig. 1: The overall architecture of PLOT." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.41, + 0.789, + 0.622 + ], + "angle": 0, + "content": "To tackle the above issues, we introduce a new framework that discovers distinctive parts in both modalities and matches them between the two modalities without any correspondence supervision; its overall architecture is illustrated in Fig. 1. To discover distinctive parts and extract their features from both modalities, we propose PLOT, a Part discovery module based on the sLOT attention mechanism [15]. Slot attention is an attention mechanism designed for object-centric learning, which segments input data into a set of slots representing individual entities without requiring object-level supervision. In PLOT, we first define a set of learnable embedding vectors, termed part slots, that contain primitive information related to human body parts shared between the two modalities. Then, these part slots undergo refinement through several iterative attention processes, where they compete amongst themselves to bind with the input data; ultimately, the part slots are transformed into part features, termed part embeddings, that represent distinctive parts in the input data." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.625, + 0.789, + 0.777 + ], + "angle": 0, + "content": "To ensure correspondence between part embeddings from different modalities, the part slots are shared between the visual and textual modalities. Part embeddings from the same part slots are then learned to represent the identical part although they are computed from different modalities. This mechanism enables PLOT to match the discovered parts from the two modalities without supervision for the correspondence as well as capturing part-level fine-grained appearance features from the both modalities. Hence, PLOT improves performance of text-based person search through the rich and fine-grained part features, and at the same time, it guarantees interpretable retrieval by providing part-level correspondences between query text and retrieved images." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.78, + 0.79, + 0.842 + ], + "angle": 0, + "content": "In addition, PLOT introduces a new similarity aggregation method for part embeddings called text-based dynamic part attention (TDPA) pooling, which dynamically adjusts the weights of part embeddings based on the text query. Given a text query as input, TDPA predicts the importance weight of each" + } + ], + [ + { + "type": "header", + "bbox": [ + 0.242, + 0.115, + 0.732, + 0.129 + ], + "angle": 0, + "content": "PLOT: TBPS with Part Slot Attention for Corresponding Part Discovery" + }, + { + "type": "page_number", + "bbox": [ + 0.775, + 0.117, + 0.785, + 0.127 + ], + "angle": 0, + "content": "3" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.147, + 0.785, + 0.207 + ], + "angle": 0, + "content": "slot for retrieval and applies the predicted importance weights to aggregate the similarities of part embeddings between the two modalities. TDPA allows the entire retrieval system to perform optimized retrieval for each query, leading to improved performance." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.208, + 0.787, + 0.267 + ], + "angle": 0, + "content": "Our method was evaluated and compared with prior work on three public benchmarks [4, 11, 32], where it clearly outperformed all existing methods thanks to the rich representation based on part embeddings. The main contribution of our work is four-fold:" + }, + { + "type": "text", + "bbox": [ + 0.226, + 0.278, + 0.784, + 0.335 + ], + "angle": 0, + "content": "- We introduce PLOT, a new framework for text-based person search that discovers distinguished human body parts, extracts their embeddings, and establishes their correspondences between the two modalities with no human intervention." + }, + { + "type": "text", + "bbox": [ + 0.225, + 0.338, + 0.784, + 0.366 + ], + "angle": 0, + "content": "- The part embeddings provided by PLOT enables an interpretable text-based person search thanks to the part-level correspondences it provides." + }, + { + "type": "text", + "bbox": [ + 0.225, + 0.369, + 0.785, + 0.411 + ], + "angle": 0, + "content": "- We introduce a novel similarity aggregation method that adaptively determines importance of each discovered part based on each text query and consequently enables retrieval optimized per query." + }, + { + "type": "text", + "bbox": [ + 0.225, + 0.413, + 0.784, + 0.441 + ], + "angle": 0, + "content": "- Our model with PLOT achieved the best on all the public benchmarks for text-based person search." + }, + { + "type": "list", + "bbox": [ + 0.225, + 0.278, + 0.785, + 0.441 + ], + "angle": 0, + "content": null + }, + { + "type": "title", + "bbox": [ + 0.216, + 0.467, + 0.388, + 0.482 + ], + "angle": 0, + "content": "2 Related Work" + }, + { + "type": "title", + "bbox": [ + 0.215, + 0.498, + 0.478, + 0.512 + ], + "angle": 0, + "content": "2.1 Text-Based Person Search" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.523, + 0.788, + 0.674 + ], + "angle": 0, + "content": "In recent years, the task of text-based person search has gained significant attention in the computer vision community. Li et al., [11] proposed a gated neural attention-based recurrent neural network (GNA-RNN) for learning the affinity between text descriptions and images, along with providing a benchmark dataset CUHK-PEDES for model evaluation. Zhang et al., [30] proposed cross-modal projection matching and classification (CMPM+CMPC) loss, for learning deep discriminative image-text global embeddings. However, these methods primarily focus on global representations of input data and are thus not capable of capturing distinctive human part details, leading to limited performance in the text-based person search." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.675, + 0.788, + 0.841 + ], + "angle": 0, + "content": "To address the above problem, a line of studies focuses on extracting fine-grained representations. One of the prominent examples of exploiting fine-grained information is to cut human images horizontally and use them as human parts [2, 4, 5, 16, 22, 23]. Chen et al., [2] extracts image part embeddings through equi-horizontal cropping of the entire image and aligns additional network to transform the textual global into corresponding parts. However, these heuristic part features inevitably include no-informative information, such as background elements. To avoid the above limitations, Suo et al., [22] proposed the simple and robust correlation filtering method to extract foreground features on the heuristic part features. Yet, the heuristically divided part features fundamentally fall short of capturing the complex human parts, underscoring a critical limitation in their" + } + ], + [ + { + "type": "page_number", + "bbox": [ + 0.218, + 0.116, + 0.23, + 0.127 + ], + "angle": 0, + "content": "4" + }, + { + "type": "header", + "bbox": [ + 0.272, + 0.115, + 0.361, + 0.127 + ], + "angle": 0, + "content": "J. Park et al." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.147, + 0.788, + 0.327 + ], + "angle": 0, + "content": "expressive capability. To move beyond such heuristic part structures, Shao et al., [2] proposes a learning-based approach for part feature extraction. However, this method struggles to extract distinctive part embeddings due to the extraction of redundant part information, which lacks sufficient disentanglement. To extract exquisite human parts, several studies tried to utilize useful information (e.g., human attributes and human keypoints) via external tools [1,8,24]. Wang et al., [24] introduced an auxiliary attribute segmentation to align the visual part features with the textual attributes parsed from text description. Jing et al., [8] proposed a new multi-granularity attention network to learn the part feature alignment between visual and textual with human pose estimation. However, these approaches have inevitable limitations of high computational cost and dependence on the performance of external tools for local feature extraction." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.329, + 0.788, + 0.435 + ], + "angle": 0, + "content": "Most recent work, Jiang et al., [7] utilizes pre-trained CLIP [18] model, capitalizing on the rich knowledge of models trained on extensive data for text-to-image matching, to excel in text-based person search. Yet, this methods primarily focus on global features, not specifically designed for extracting human part features. Not only do we leverage pre-trained knowledge, but we transcend heuristic human parts methods to extract sophisticated human parts without external tools or part-level supervision." + }, + { + "type": "title", + "bbox": [ + 0.216, + 0.461, + 0.382, + 0.475 + ], + "angle": 0, + "content": "2.2 Slot Attention" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.49, + 0.788, + 0.686 + ], + "angle": 0, + "content": "Slot attention [15] is a recently proposed attention mechanism for object-centric learning, a problem focusing on discovering constituent visual entities within an image. The unique property of slot attention is that it can represent input images as a set of slots, where the slots are representations corresponding to individual visual entities, without any object-level supervision during training. Within slot attention, slots iteratively compete for aggregating input data, ensuring distinct representations focusing on individual visual entities. By incorporating slot attention, the proposed framework facilitates the unsupervised identification of a structure underlying image and text queries, enabling the model to discern and represent individual human parts without explicit supervision. This capability is particularly valuable in person search datasets, where recognizing subtle differences and understanding the correspondence of human parts across different modalities are crucial." + }, + { + "type": "title", + "bbox": [ + 0.216, + 0.714, + 0.427, + 0.731 + ], + "angle": 0, + "content": "3 Proposed Method" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.75, + 0.788, + 0.841 + ], + "angle": 0, + "content": "The following subsection offers details of global and part embeddings extraction for each modality with its backbone (Sec 3.1). Subsequently, we present our novel framework, PLOT, which includes the part discovery module (Sec. 3.2) and similarity aggregation between part embeddings (Sec. 3.3), concluding with a discussion on the learning objective designed to optimize our proposed framework (Sec. 3.4) and inference of our framework (Sec. 3.5)." + } + ], + [ + { + "type": "header", + "bbox": [ + 0.242, + 0.115, + 0.732, + 0.129 + ], + "angle": 0, + "content": "PLOT: TBPS with Part Slot Attention for Corresponding Part Discovery" + }, + { + "type": "page_number", + "bbox": [ + 0.775, + 0.117, + 0.785, + 0.127 + ], + "angle": 0, + "content": "5" + }, + { + "type": "title", + "bbox": [ + 0.216, + 0.147, + 0.503, + 0.162 + ], + "angle": 0, + "content": "3.1 Global and Part Embeddings" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.169, + 0.788, + 0.577 + ], + "angle": 0, + "content": "In our framework, a single input data is described by two different types of representations: a global embedding and multiple part embeddings. The global embedding is used to represent the input data holistically, while each part embedding describes the appearance of distinctive human parts (e.g., arm, leg, torso, etc.). We below provide details of global and part embeddings computation for visual and textual modalities. Following the previous work [7], a pre-trained CLIP [18] is used as the backbone networks for the visual and textual modalities. Visual Modality: We utilize the vision transformer (ViT) from the CLIP-B/16 [18] architecture as a visual backbone network. Initially, an input image of a person is split into \\(N\\) distinct, non-overlapping patches, which are subsequently transformed into patch tokens through linear projection. The patch tokens and an extra [c1s] token are then fed into the visual backbone network. The token sequence is processed throughout multiple self-attention blocks, and the [c1s] token of the last block is used as a global embedding \\(\\mathbf{g}^{\\mathcal{V}} \\in \\mathbb{R}^{D}\\). To obtain part embeddings \\(\\mathbf{P}^{\\mathcal{V}} \\in \\mathbb{R}^{K \\times D}\\), remaining patch tokens of last block \\(\\mathbf{x}^{\\mathcal{V}} \\in \\mathbb{R}^{N \\times D}\\) is passed to the part discovery module. The part discovery module aggregates the patch features describing coherent human parts into the same part embedding. We provide more detailed information about the part discovery module in Sec. 3.2. Textual Modality: For the textual backbone network, we utilize a transformer architecture from the CLIP-Xformer [18] text encoder. This encoder operates on text input transformed into byte pair encoding (BPE) sequences. Initially, the text query undergoes tokenization via BPE, followed by the enclosing with [S0S] and [E0S] tokens. The resulting sequence of tokens is then inputted into the textual backbone network. Here, the [E0S] token from the final block serves as the global embedding, denoted as \\(\\mathbf{g}^{\\mathcal{T}} \\in \\mathbb{R}^{D}\\). Analogous to the approach for the visual modality, we process the remainder of the text tokens, \\(\\mathbf{x}^{\\mathcal{T}} \\in \\mathbb{R}^{L \\times D}\\), into part embeddings, \\(\\mathbf{P}^{\\mathcal{T}} \\in \\mathbb{R}^{K \\times D}\\), utilizing the part discovery module." + }, + { + "type": "title", + "bbox": [ + 0.216, + 0.595, + 0.455, + 0.61 + ], + "angle": 0, + "content": "3.2 Part Discovery Module" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.617, + 0.787, + 0.706 + ], + "angle": 0, + "content": "For extracting part embeddings in each modality, part discovery module aggregates patch tokens \\(\\mathbf{x}^{\\mathcal{V}}\\) and text tokens \\(\\mathbf{x}^{\\mathcal{T}}\\) into the visual part embeddings \\(\\mathbf{P}^{\\mathcal{V}}\\) and textual part embeddings \\(\\mathbf{P}^{\\mathcal{T}}\\), respectively. It is worth noting that the part discovery module for each modality has identical model architecture and functions equivalently. Therefore, we will only explain the part discovery module on the visual modality for brevity." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.707, + 0.788, + 0.812 + ], + "angle": 0, + "content": "Part discovery module consists of initial part slots \\(\\mathbf{S}^0\\in \\mathbb{R}^{K\\times D}\\) and \\(T\\) multiple iteration of the part slot attention block (PSA block). To extract part embeddings, we first initialize a set of learnable embeddings part slots \\(\\mathbf{S}^0\\in \\mathbb{R}^{K\\times D}\\), where \\(K\\) indicates the number of part slots. Then, through a series of \\(T\\) iterations of our PSA block, the initial \\(\\mathbf{S}^0\\) evolves into refined \\(\\mathbf{S}^T\\), where each slot captures distinct parts within input data. The refined part slots \\(\\mathbf{S}^T\\) are used as visual part embeddings \\(\\mathbf{P}^{\\mathcal{V}}\\). The part discovery module can be formulated as follows:" + }, + { + "type": "equation", + "bbox": [ + 0.339, + 0.819, + 0.787, + 0.836 + ], + "angle": 0, + "content": "\\[\n\\mathbf {P} ^ {\\mathcal {V}} := \\mathbf {S} ^ {T}, \\text {w h e r e} \\quad \\mathbf {S} ^ {t} = \\operatorname {P S A} _ {-} \\operatorname {B l o c k} ^ {\\mathcal {V}} \\left(\\mathbf {x} ^ {\\mathcal {V}}; \\mathbf {S} ^ {t - 1}\\right). \\tag {1}\n\\]" + } + ], + [ + { + "type": "page_number", + "bbox": [ + 0.218, + 0.116, + 0.23, + 0.127 + ], + "angle": 0, + "content": "6" + }, + { + "type": "header", + "bbox": [ + 0.272, + 0.115, + 0.361, + 0.128 + ], + "angle": 0, + "content": "J. Park et al." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.146, + 0.788, + 0.206 + ], + "angle": 0, + "content": "The PSA block first transforms the inputs \\(\\mathbf{S}^{t - 1}\\) and \\(\\mathbf{x}^{\\mathcal{V}}\\) with layer normalization and linear projection layers \\(q(\\cdot), k(\\cdot)\\) and \\(\\nu (\\cdot)\\) to obtain embeddings of \\(D_{h}\\) dimension. Then the attention map \\(A\\in \\mathbb{R}^{N\\times K}\\) between \\(\\mathbf{S}^{t - 1}\\) and \\(\\mathbf{x}^{\\mathcal{V}}\\) is computed by" + }, + { + "type": "equation", + "bbox": [ + 0.335, + 0.207, + 0.788, + 0.243 + ], + "angle": 0, + "content": "\\[\nA _ {n, k} = \\frac {e ^ {M _ {n , k}}}{\\sum_ {i = 1} ^ {K} e ^ {M _ {n , i}}}, \\mathrm {w h e r e} M = \\frac {k (\\mathbf {x} ^ {\\mathcal {V}}) q (\\mathbf {S} ^ {t - 1}) ^ {\\top}}{\\sqrt {D _ {h}}}. (2)\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.246, + 0.788, + 0.322 + ], + "angle": 0, + "content": "The attention map is obtained through normalization across part slots; this normalization encourages competition among themselves to bind distinct sets of patch tokens to each slot. We update part slots with weighted mean of patch tokens obtained by attention map \\( A_{n,k} \\) and then feed it to a gated recurrent unit (GRU) using \\( \\mathbf{S}^{t-1} \\) as hidden state as follows:" + }, + { + "type": "equation", + "bbox": [ + 0.349, + 0.333, + 0.788, + 0.368 + ], + "angle": 0, + "content": "\\[\n\\bar {A} _ {n, k} = \\frac {A _ {n , k}}{\\sum_ {i = 1} ^ {N} A _ {i , k}}, \\bar {\\mathbf {S}} ^ {t} = \\mathbb {G R U} \\left(\\mathbf {S} ^ {t - 1}, \\bar {A} ^ {\\top} v \\left(\\mathbf {x} ^ {\\mathcal {V}}\\right)\\right). \\tag {3}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.377, + 0.787, + 0.408 + ], + "angle": 0, + "content": "Then, we obtain \\( t \\)-th part slot by feeding \\( \\overline{\\mathbf{S}}^{t-1} \\) into a multi-layer perceptron (MLP) with layer normalization, ReLU activation, and residual connection:" + }, + { + "type": "equation", + "bbox": [ + 0.334, + 0.418, + 0.788, + 0.436 + ], + "angle": 0, + "content": "\\[\n\\mathbf {S} ^ {t} = \\operatorname {P S A} _ {\\text {B l o c k}} ^ {\\mathcal {V}} (\\mathbf {x} ^ {\\mathcal {V}}; \\mathbf {S} ^ {t - 1}) = \\operatorname {M L P} \\left(\\bar {\\mathbf {S}} ^ {t - 1}\\right) + \\bar {\\mathbf {S}} ^ {t - 1}. \\tag {4}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.444, + 0.787, + 0.478 + ], + "angle": 0, + "content": "Finally, we can obtain the visual part embeddings \\(\\mathbf{P}^{\\mathcal{V}} = \\{\\mathbf{p}_k^{\\mathcal{V}}\\}_{k=1}^K\\) which is the output of \\(T\\)-th iteration of PSA block: \\(\\mathbf{P}^{\\mathcal{V}} = \\mathbf{S}_{\\mathcal{V}}^T \\in \\mathbb{R}^{K \\times D}\\)." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.478, + 0.788, + 0.554 + ], + "angle": 0, + "content": "Part Correspondence through Slot Sharing: Additionally, we share the learnable part slots \\(\\mathbf{S}^0\\) between two part discovery modules to establish correspondences between part embeddings extracted from each modality, considering that part embeddings extracted from the same part slot are corresponding part across modalities, thereby contributing to a clearer comparison between modalities." + }, + { + "type": "title", + "bbox": [ + 0.215, + 0.575, + 0.616, + 0.59 + ], + "angle": 0, + "content": "3.3 Measuring Similarity between Embeddings" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.599, + 0.787, + 0.674 + ], + "angle": 0, + "content": "In the context of training and applying our retrieval model, selecting an appropriate similarity function between embeddings is crucial. The challenge lies in dealing with two distinct types of embeddings: global embeddings and part embeddings. For global embeddings, cosine similarity offers a straightforward and effective means of measuring similarity." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.675, + 0.789, + 0.841 + ], + "angle": 0, + "content": "However, the situation becomes more complex when considering part embeddings. A direct method for addressing this complexity involves calculating the average cosine similarity across all pairs of part embeddings. This approach, while straightforward, has its drawbacks, primarily because it treats all part-wise similarities as equally significant. In reality, the relevance of specific part embeddings to the actual similarity between data instances can significantly vary, influenced by the context of a text query. For instance, if a text query focuses exclusively on particular features of a human figure, the similarity contributions from other unrelated part embeddings should be less relevant. This challenge highlights the necessity for an approach that can dynamically assess and prioritize the relevance of part embeddings based on the context provided by the query." + } + ], + [ + { + "type": "header", + "bbox": [ + 0.241, + 0.115, + 0.733, + 0.131 + ], + "angle": 0, + "content": "PLOT: TBPS with Part Slot Attention for Corresponding Part Discovery" + }, + { + "type": "page_number", + "bbox": [ + 0.775, + 0.116, + 0.787, + 0.128 + ], + "angle": 0, + "content": "7" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.147, + 0.79, + 0.222 + ], + "angle": 0, + "content": "To address this challenge, we introduce text-based dynamic part attention (TDPA) to aggregate the similarities between part embeddings. Firstly, textual global embedding \\(\\mathbf{g}^{\\mathcal{T}}\\) is transformed to the TDPA \\(\\mathbf{a} \\in \\mathbb{R}^{K}\\), using MLP and a softmax function. Using TDPA, the aggregated similarity between part embeddings are computed as follows:" + }, + { + "type": "equation", + "bbox": [ + 0.365, + 0.231, + 0.788, + 0.248 + ], + "angle": 0, + "content": "\\[\n\\mathbf {a} = \\sigma (\\operatorname {M L P} \\left(\\mathbf {g} ^ {\\mathcal {T}}\\right)) \\in \\mathbb {R} ^ {K}, \\tag {5}\n\\]" + }, + { + "type": "equation", + "bbox": [ + 0.365, + 0.253, + 0.788, + 0.292 + ], + "angle": 0, + "content": "\\[\nc _ {\\mathrm {a g g}} \\left(\\mathbf {P} ^ {\\mathcal {V}}, \\mathbf {P} ^ {\\mathcal {T}}; \\mathbf {g} ^ {\\mathcal {T}}\\right) := \\sum_ {k = 1} ^ {K} \\mathrm {a} _ {k} \\cdot c \\left(\\mathbf {p} _ {k} ^ {\\mathcal {V}}, \\mathbf {p} _ {k} ^ {\\mathcal {T}}\\right), \\tag {6}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.299, + 0.789, + 0.421 + ], + "angle": 0, + "content": "where \\(c(\\cdot, \\cdot)\\) denotes cosine similarity between two embeddings, \\(a_k\\) is the \\(k\\)-th value of \\(\\mathbf{a}\\), and \\(\\sigma(\\cdot)\\) is a softmax function. If given textual global embedding \\(\\mathbf{g}^{\\mathcal{T}}\\), TDPA is computed by MLP which is learned in an end-to-end manner by minimizing our partNCE loss, as will be introduced in Sec 3.4. This dynamic attention enables us to adaptively assign importance to each part embedding, which reduces the impact of non-informative part embedding and promotes a more informative similarity measurement that reflects the significant semantic similarity of each text query." + }, + { + "type": "title", + "bbox": [ + 0.215, + 0.44, + 0.421, + 0.456 + ], + "angle": 0, + "content": "3.4 Learning Objective" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.462, + 0.789, + 0.583 + ], + "angle": 0, + "content": "Our model is trained through the establishment of cross-modal alignments, considering both global and part levels. The alignment at the global-level is accomplished by aligning global embeddings \\(\\mathbf{g}^{\\mathcal{V}}\\) and \\(\\mathbf{g}^{\\mathcal{T}}\\) that capture the comprehensive information of each modality. For part-level alignment, we leverage part embeddings \\(\\mathbf{P}^{\\mathcal{V}}\\) and \\(\\mathbf{P}^{\\mathcal{T}}\\) that have locally exclusive features within each modality due to the slot attention. Furthermore, by dynamically adjusting weights of specific part slots based on the text global embedding \\(\\mathbf{g}^{\\mathcal{T}}\\), we facilitate the learning of more informative alignments." + }, + { + "type": "text", + "bbox": [ + 0.215, + 0.583, + 0.789, + 0.689 + ], + "angle": 0, + "content": "Global Alignment Loss: To align global embeddings \\(\\mathbf{g}^{\\mathcal{V}}\\) and \\(\\mathbf{g}^{\\mathcal{T}}\\) extracted from each modality backbone, we first define a batch of global embeddings \\(\\mathcal{B}_{\\mathrm{global}} = \\{(\\mathbf{g}_i^{\\mathcal{V}},\\mathbf{g}_i^{\\mathcal{T}})\\}_{i = 1}^B\\), where \\(B\\) is batch size. Then we adopt the InfoNCE loss [17] which is a contrastive learning objective that maximizes the similarity between embeddings of positive pairs while minimizing the similarity between negative pairs in the batch. Consequently, our global NCE loss with cosine similarity function \\(c(\\cdot ,\\cdot)\\) is formulated as follows:" + }, + { + "type": "equation", + "bbox": [ + 0.297, + 0.698, + 0.788, + 0.745 + ], + "angle": 0, + "content": "\\[\n\\mathcal {L} _ {\\mathrm {N C E}} = - \\sum_ {i = 1} ^ {B} \\left(\\log \\frac {e ^ {c \\left(\\mathbf {g} _ {i} ^ {\\mathcal {V}} , \\mathbf {g} _ {i} ^ {\\mathcal {T}}\\right) / \\tau}}{\\sum_ {j = 1} ^ {B} e ^ {c \\left(\\mathbf {g} _ {i} ^ {\\mathcal {V}} , \\mathbf {g} _ {j} ^ {\\mathcal {T}}\\right) / \\tau}} + \\log \\frac {e ^ {c \\left(\\mathbf {g} _ {i} ^ {\\mathcal {V}} , \\mathbf {g} _ {i} ^ {\\mathcal {T}}\\right) / \\tau}}{\\sum_ {j = 1} ^ {B} e ^ {c \\left(\\mathbf {g} _ {j} ^ {\\mathcal {V}} , \\mathbf {g} _ {i} ^ {\\mathcal {T}}\\right) / \\tau}}\\right), \\tag {7}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.75, + 0.788, + 0.797 + ], + "angle": 0, + "content": "where the \\(\\tau\\) is temperature term. Additionally, we employ an identity classification loss \\(\\mathcal{L}_{\\mathrm{ID}}\\) to ensure that the embeddings extracted from the same identity become similar. The \\(\\mathcal{L}_{\\mathrm{ID}}\\) is denoted by" + }, + { + "type": "equation", + "bbox": [ + 0.317, + 0.805, + 0.788, + 0.843 + ], + "angle": 0, + "content": "\\[\n\\mathcal {L} _ {\\mathrm {I D}} = - \\sum_ {i = 1} ^ {B} \\left(\\boldsymbol {y} _ {i} \\log \\sigma \\left(\\mathbf {g} _ {i} ^ {\\mathcal {V}} \\mathbf {W} _ {\\mathrm {I D}}\\right) + \\boldsymbol {y} _ {i} \\log \\sigma \\left(\\mathbf {g} _ {i} ^ {\\mathcal {T}} \\mathbf {W} _ {\\mathrm {I D}}\\right)\\right), \\tag {8}\n\\]" + } + ], + [ + { + "type": "page_number", + "bbox": [ + 0.218, + 0.116, + 0.23, + 0.127 + ], + "angle": 0, + "content": "8" + }, + { + "type": "header", + "bbox": [ + 0.272, + 0.115, + 0.361, + 0.128 + ], + "angle": 0, + "content": "J. Park et al." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.147, + 0.788, + 0.224 + ], + "angle": 0, + "content": "where \\(\\mathbf{y}_i\\in \\mathbb{R}^C\\) is the identity ground truth of corresponding global embedding \\(\\mathbf{g}_i\\) represented by a one-hot vector and \\(C\\) is the number of identities, \\(\\mathbf{W}_{\\mathrm{ID}}\\in \\mathbb{R}^{D\\times C}\\) is a classifier shared between the two modalities. For cross-modal alignments, the commonly used CMPM loss [7,30] is additionally adopted. Finally, we describe global alignment loss as below:" + }, + { + "type": "equation", + "bbox": [ + 0.38, + 0.235, + 0.788, + 0.251 + ], + "angle": 0, + "content": "\\[\n\\mathcal {L} _ {\\text {G l o b a l}} = \\mathcal {L} _ {\\text {N C E}} + \\mathcal {L} _ {\\text {I D}} + \\mathcal {L} _ {\\text {C M P M}}. \\tag {9}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.215, + 0.262, + 0.79, + 0.368 + ], + "angle": 0, + "content": "Part Alignment Loss: Similar to the alignment of global embeddings, we first define a batch of part embeddings \\(\\mathcal{B}_{\\mathrm{part}} = \\{(\\mathbf{P}_i^{\\mathcal{V}},\\mathbf{P}_i^{\\mathcal{T}})\\}_{i = 1}^B\\) with batch size \\(B\\), and we adopt the InfoNCE loss to align part embeddings extracted from two modalities. For learning text query-based informative alignment of part embeddings between modalities, we employ text query-based similarity aggregation function \\(c_{\\mathrm{agg}}\\) in Eq. 6 to compute InfoNCE loss. We termed this InfoNCE with \\(c_{\\mathrm{agg}}\\) as PartNCE loss, and it is formulated by" + }, + { + "type": "equation", + "bbox": [ + 0.23, + 0.378, + 0.788, + 0.429 + ], + "angle": 0, + "content": "\\[\n\\mathcal {L} _ {\\mathrm {P a r t N C E}} = - \\sum_ {i = 1} ^ {B} \\left(\\log \\frac {e ^ {c _ {\\mathrm {a g g}} (\\mathbf {P} _ {i} ^ {\\mathcal {V}} , \\mathbf {P} _ {i} ^ {\\mathcal {T}} ; \\mathbf {g} _ {i} ^ {\\mathcal {T}}) / \\tau}}{\\sum_ {j = 1} ^ {B} e ^ {c _ {\\mathrm {a g g}} (\\mathbf {P} _ {i} ^ {\\mathcal {V}} , \\mathbf {P} _ {j} ^ {\\mathcal {T}} ; \\mathbf {g} _ {j} ^ {\\mathcal {T}}) / \\tau}} + \\log \\frac {e ^ {c _ {\\mathrm {a g g}} (\\mathbf {P} _ {i} ^ {\\mathcal {V}} , \\mathbf {P} _ {i} ^ {\\mathcal {T}} ; \\mathbf {g} _ {i} ^ {\\mathcal {T}}) / \\tau}}{\\sum_ {j = 1} ^ {B} e ^ {c _ {\\mathrm {a g g}} \\left(\\mathbf {P} _ {j} ^ {\\mathcal {V}}, \\mathbf {P} _ {i} ^ {\\mathcal {T}}; \\mathbf {g} _ {i} ^ {\\mathcal {T}}\\right) / \\tau}\\right), \\tag {10}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.215, + 0.436, + 0.788, + 0.498 + ], + "angle": 0, + "content": "Similar to global alignment loss, we adopt the identity loss that shares classifier weights between the two modalities for part alignment; however, the difference is that the part embeddings are concatenated along the embedding dimension. The identity loss for part alignment is formulated by" + }, + { + "type": "equation", + "bbox": [ + 0.257, + 0.508, + 0.788, + 0.548 + ], + "angle": 0, + "content": "\\[\n\\mathcal {L} _ {\\text {P a r t I D}} = - \\sum_ {i = 1} ^ {B} \\left(\\boldsymbol {y} _ {i} \\log \\sigma \\left(\\left[ \\boldsymbol {P} _ {i} ^ {\\mathcal {V}} \\right] \\boldsymbol {W} _ {\\text {P a r t I D}}\\right) + \\boldsymbol {y} _ {i} \\log \\sigma \\left(\\left[ \\boldsymbol {P} _ {i} ^ {\\mathcal {T}} \\right] \\boldsymbol {W} _ {\\text {P a r t I D}}\\right)\\right), \\tag {11}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.215, + 0.56, + 0.79, + 0.606 + ], + "angle": 0, + "content": "where \\([\\mathbf{P}] \\in \\mathbb{R}^{KD}\\) is the concatenation of part embeddings \\(\\mathbf{P}\\) along the embedding dimension, and \\(\\mathbf{W}_{\\mathrm{PartID}} \\in \\mathbb{R}^{KD \\times C}\\) is a classifier shared between both modalities. Finally, the part alignment loss is computed by" + }, + { + "type": "equation", + "bbox": [ + 0.397, + 0.618, + 0.788, + 0.634 + ], + "angle": 0, + "content": "\\[\n\\mathcal {L} _ {\\text {P a r t}} = \\mathcal {L} _ {\\text {P a r t N C E}} + \\mathcal {L} _ {\\text {P a r t I D}}. \\tag {12}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.215, + 0.659, + 0.519, + 0.764 + ], + "angle": 0, + "content": "Cross-Modal Masked Language Modeling Loss: Following the practices of employing Transformer-based backbones in existing methods [7,21], we adopt an auxiliary loss, the cross-modal masked language modeling (CMLM) loss, to facilitate the learning of interactions between modali" + }, + { + "type": "image", + "bbox": [ + 0.534, + 0.651, + 0.778, + 0.735 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.539, + 0.737, + 0.77, + 0.748 + ], + "angle": 0, + "content": "Cross-Modal Masked Language Model (CMLM)" + }, + { + "type": "image_caption", + "bbox": [ + 0.555, + 0.749, + 0.758, + 0.762 + ], + "angle": 0, + "content": "Fig. 2: Illustration of CMLM." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.765, + 0.788, + 0.841 + ], + "angle": 0, + "content": "ties. Similar to BERT [3], given a text description, we randomly select text tokens with a \\(15\\%\\) probability and replace them with the learnable [MASK] token. The masked text description is then processed through the text backbone to obtain the masked textual tokens. After obtaining the masked textual tokens, we concatenate them with the visual tokens extracted via the image backbone." + } + ], + [ + { + "type": "header", + "bbox": [ + 0.242, + 0.115, + 0.733, + 0.131 + ], + "angle": 0, + "content": "PLOT: TBPS with Part Slot Attention for Corresponding Part Discovery" + }, + { + "type": "page_number", + "bbox": [ + 0.775, + 0.117, + 0.785, + 0.127 + ], + "angle": 0, + "content": "9" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.147, + 0.788, + 0.254 + ], + "angle": 0, + "content": "The concatenated tokens are then fed into a transformer to acquire cross-modal fused tokens. Among the fused tokens, those corresponding to the indices of textual tokens are denoted as \\(\\mathbf{F}\\) composed as \\(\\{\\mathbf{f}_1,\\dots \\mathbf{f}_L\\}\\), where a fused token \\(\\mathbf{f}\\in \\mathbb{R}^{1\\times D}\\) and \\(L\\) indicates the max length of input text descriptions. Ultimately, these tokens are fed into a CMLM classifier \\(\\mathbf{W}_{\\mathrm{CMLM}}\\) to predict the probability of vocabulary IDs. The overall procedure of CMLM is illustrated in Fig. 2 and through this procedure we compute the \\(\\mathcal{L}_{\\mathrm{CMLM}}\\) loss as follows:" + }, + { + "type": "equation", + "bbox": [ + 0.363, + 0.265, + 0.788, + 0.304 + ], + "angle": 0, + "content": "\\[\n\\mathcal {L} _ {\\mathrm {C M L M}} = - \\frac {1}{L} \\sum_ {l = 1} ^ {L} \\mathbf {y} _ {l} \\log (\\sigma (\\mathbf {f} _ {l} \\mathbf {W} _ {\\mathrm {C M L M}})), \\tag {13}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.315, + 0.79, + 0.362 + ], + "angle": 0, + "content": "where \\(\\mathbf{W}_{\\mathrm{CMLM}} \\in \\mathbb{R}^{D \\times V}\\) and \\(\\pmb{y}_l \\in \\mathbb{R}^V\\) is the vocabulary ground truth of \\(l\\)-th text tokens, represented by a one-hot vector and \\(V\\) is the size of vocabulary. Finally, our overall objective function for training is denoted by" + }, + { + "type": "equation", + "bbox": [ + 0.393, + 0.376, + 0.788, + 0.39 + ], + "angle": 0, + "content": "\\[\n\\mathcal {L} = \\mathcal {L} _ {\\text {G l o b a l}} + \\mathcal {L} _ {\\text {P a r t}} + \\mathcal {L} _ {\\text {C M L M}}. \\tag {14}\n\\]" + }, + { + "type": "title", + "bbox": [ + 0.216, + 0.412, + 0.341, + 0.426 + ], + "angle": 0, + "content": "3.5 Inference" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.436, + 0.79, + 0.573 + ], + "angle": 0, + "content": "During testing, the global and part embeddings of each modality input are fully exploited to calculate the similarity between the image-text pair. In particular, the similarities between visual part embeddings and their corresponding textual part embeddings are linearly combined with the attention weights \\(\\mathbf{a}\\) to aggregate them. The image-text pair similarity is defined as the sum of the similarity between the global embeddings of the image-text pair and the similarity between the part embeddings of it, which can be computed by \\(c(\\mathbf{g}^{\\mathcal{V}},\\mathbf{g}^{\\mathcal{T}}) + c_{\\mathrm{agg}}(\\mathbf{P}^{\\mathcal{V}},\\mathbf{P}^{\\mathcal{T}};\\mathbf{g}^{\\mathcal{T}})\\). Finally, given the text query, the images in the gallery are ranked according to similarity scores between the images and the text for inference." + }, + { + "type": "title", + "bbox": [ + 0.216, + 0.596, + 0.377, + 0.613 + ], + "angle": 0, + "content": "4 Experiments" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.627, + 0.788, + 0.734 + ], + "angle": 0, + "content": "In this section, we provide a detailed account of our experimental setup (Sec. 4.1), evaluate our method, and compare it with state of the arts on three benchmark datasets for text-based person search (Sec. 4.2). Furthermore, we qualitatively present retrieval results and analyze the effectiveness of the part discovery module and TDPA with visualization results (Sec. 4.3). We also conduct ablation studies on the losses employed in model training, the methodologies of part discovery, and the strategies for part similarity aggregation (Sec. 4.4)." + }, + { + "type": "title", + "bbox": [ + 0.216, + 0.755, + 0.429, + 0.772 + ], + "angle": 0, + "content": "4.1 Experimental Setup" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.78, + 0.788, + 0.843 + ], + "angle": 0, + "content": "Datasets: On three benchmark datasets, CUHK-PEDES [11], ICFG-PEDES [4], and RSTPReid [32], we evaluate and compare the performance of our method against previous methods. In CUHK-PEDES collected from five existing person re-identification datasets [6,12,13,28,31], it contains 40,206 images corresponding" + } + ], + [ + { + "type": "page_number", + "bbox": [ + 0.218, + 0.116, + 0.236, + 0.127 + ], + "angle": 0, + "content": "10" + }, + { + "type": "header", + "bbox": [ + 0.272, + 0.115, + 0.361, + 0.127 + ], + "angle": 0, + "content": "J. Park et al." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.147, + 0.788, + 0.344 + ], + "angle": 0, + "content": "to 13,003 individual IDs, with each image being approximately matched with two annotated text descriptions. We follow the data split of [11] with 34,054 images from 11,003 person IDs and 68,126 text descriptions for training, 3,078 images from 1,000 IDs and 6,158 text descriptions for validation, and 3,074 images from 1,000 IDs and 6,156 text descriptions for testing. The remaining two datasets are collected from MSMT17 [26]. ICFG-PEDES consists of 54,522 image-text pairs from 4,102 individual IDs, which are split into 34,674 and 19,848 for training and testing, respectively. RSTPReid contains 20,505 images of 4,101 individual IDs, with each ID having 5 images and each image associated with the corresponding two annotated text descriptions. We follow the data split of [32] with 18,505 images from 3,701 IDs and 37,010 text descriptions for training, 1,000 images from 200 IDs and 2,000 text descriptions for validation, and 1,000 images from 200 IDs and 2,000 text descriptions for testing, respectively." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.344, + 0.788, + 0.405 + ], + "angle": 0, + "content": "Evaluation Protocol: We employ the standard metric of rank at K (R@K=1,5, 10) for all retrieval experiments. Specifically, given a query text, images are sorted based on their similarity to the query text. The search is considered correct if at least one relevant image appears in the top K positions of the ranking." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.405, + 0.788, + 0.479 + ], + "angle": 0, + "content": "Network Architecture: We adopt the pre-trained CLIP models from OpenAI [18] for both image and text encoders, where the size of the image encoder is ViT-B/16. The input images are resized to \\(384 \\times 128\\). Random horizontal flipping, random cropping, and random erasing are applied for the data augmentation in training time. The maximum text length is set to 77." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.48, + 0.788, + 0.555 + ], + "angle": 0, + "content": "Network Optimization: Our model is trained using the Adam optimizer for 60 epochs with a batch size of 128 for all experiments. For the CLIP encoders, the initial learning rate is set to \\(5\\mathrm{e} - 6\\), using a cosine schedule with the warm-up strategy at the first five epochs; we use a high learning rate for the remaining parameters by scaling 20 times." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.555, + 0.788, + 0.599 + ], + "angle": 0, + "content": "Hyperparameters: We set the number of part slots to 8. The number of iterations in the part slot attention block is set to 5. The temperature parameter \\(\\tau\\) is set to 0.015." + }, + { + "type": "title", + "bbox": [ + 0.215, + 0.621, + 0.434, + 0.636 + ], + "angle": 0, + "content": "4.2 Quantitative Results" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.644, + 0.788, + 0.841 + ], + "angle": 0, + "content": "We compare our method with previous text-based person search methods on CUHK-PEDES [11], ICFG-PEDES [4], and RSTPReid [32]. The performance comparison and the backbones for each modality employed by each method are shown in Table 1. Specifically, our method achieves an outstanding \\( \\mathbb{R}@\\mathbb{1} \\) metric of \\( 75.28\\% \\), \\( 65.76\\% \\), and \\( 61.80\\% \\) on the CUHK-PEDES, ICFG-PEDES, and RSTPReid datasets, respectively, outperforming previous methods utilizing different backbones from ours. Moreover, our method improves the previous state of the art, IRRA [7], on \\( \\mathbb{R}@\\mathbb{1} \\) by a large margin as \\( 1.9\\% \\mathbb{p} \\), \\( 1.7\\% \\mathbb{p} \\), and \\( 1.6\\% \\mathbb{p} \\), respectively. Since IRRA only focuses on aligning the global embeddings of each modality, it is hard to capture the fine-grained differences. In contrast, our method not only takes account of the global embeddings but also aims to discover and align discriminative part embeddings from each modality; it allows the model to effectively find the target person." + } + ], + [ + { + "type": "header", + "bbox": [ + 0.242, + 0.115, + 0.732, + 0.129 + ], + "angle": 0, + "content": "PLOT: TBPS with Part Slot Attention for Corresponding Part Discovery" + }, + { + "type": "page_number", + "bbox": [ + 0.769, + 0.117, + 0.784, + 0.127 + ], + "angle": 0, + "content": "11" + }, + { + "type": "table_caption", + "bbox": [ + 0.215, + 0.145, + 0.788, + 0.173 + ], + "angle": 0, + "content": "Table 1: Performance of text-based person search methods on the three datasets. Bold and underline denote the best and the second best." + }, + { + "type": "table", + "bbox": [ + 0.216, + 0.185, + 0.784, + 0.44 + ], + "angle": 0, + "content": "
MethodsBackboneCUHK-PEDESICFG-PEDESRSTPReid
ImageTextR@1R@5R@10R@1R@5R@10R@1R@5R@10
GNA-RNN [11]RN50LSTM19.05-53.64------
CMPM/C [30]RN50LSTM49.3771.6979.2743.5165.4474.26---
PMA [8]RN50BERT53.8173.5481.23------
TIMAM [19]RN101BERT54.5177.5684.78------
SCAN [9]RN50BERT55.8675.9783.6950.0569.6577.21---
ViTAA [24]RN50LSTM55.9775.8483.5250.9868.7975.78---
NAFS [5]RN50BERT59.9479.8686.70------
DSSL [32]RN50BERT59.9880.4187.56---32.4355.0863.19
MGEL [23]RN50LSTM60.2780.0186.74------
SSAN [4]RN50LSTM61.3780.1586.7354.2372.6379.5343.5067.8077.15
LapsCore [27]RN50BERT63.40-87.80------
SRCF [22]RN50BERT64.0482.9988.8157.1875.0181.49---
LGUR [20]RN50BERT64.2181.9487.9357.4274.9781.45---
TIPCB [2]RN50BERT64.2683.1989.10------
CAIBC [25]RN50BERT64.4382.8788.37---47.3569.5579.00
SAF [10]ViT-B/16BERT64.1382.6288.40------
IVT [21]ViT-B/16BERT65.5983.1189.2156.0473.6080.2246.7070.0078.80
CFine [29]CLIP-ViT-B/16BERT69.5785.9391.1560.8375.5582.4250.5572.5081.60
IRRA [7]CLIP-ViT-B/16CLIP-Xformer73.3889.9393.7163.4680.2485.8260.2081.3088.20
OursCLIP-ViT-B/16CLIP-Xformer75.2890.4294.1265.7681.3986.7361.8082.8589.45
" + }, + { + "type": "title", + "bbox": [ + 0.216, + 0.469, + 0.421, + 0.484 + ], + "angle": 0, + "content": "4.3 Qualitative Results" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.493, + 0.789, + 0.689 + ], + "angle": 0, + "content": "Retrieval Results: Top-5 retrieval results of our method on the CUHK-PEDES dataset are illustrated in Fig. 3. Above all, it shows the overall satisfactory retrieval results. In particular, we can observe our model retrieves targets well, even with distinctive human parts that are small or located in various positions. For instance, the small distinctive human parts like \"ponytails\" and \"high skirts\" in Fig. 3(a), \"black shoes\" in Fig. 3(b), and \"blue plaid shorts\" in Fig. 3(c), as well as human parts that could appear in various viewpoints such as \"yellow shoulder bags\" in Fig. 3(a), \"floaty dresses\" in Fig. 3(b), and \"red backpacks\" in Fig. 3(d). The CUHK-PEDES dataset typically contains three target images on average in the search space, thereby most of the false matches in the figure are observed by a lack of additional targets. Despite the false matching due to the limitation of the dataset, the retrieval results are reasonable in that the retrieved false matching contains distinctive human parts described by the query description." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.689, + 0.79, + 0.841 + ], + "angle": 0, + "content": "Visualization of Attention Map \\(\\bar{A}_k\\) in PSABlock: To demonstrate the effectiveness of our part discovery module, we visualize the attention map \\(\\bar{A}_k\\) in \\(T\\)-th iteration of the PSA block for both visual and textual modalities (in Eq. (2)). The visualization results are illustrated in Fig. 4. It not only demonstrates that the part embeddings extracted by our part discovery module capture distinctive human parts but also shows that the part embeddings extracted from the same part slot attend to the semantically identical human parts regardless of modality. For example, the 1st part slot typically focuses on bottom clothes, the 4th on footwear, the 5th on objects being held, the 7th on top clothes, and the 8th on the person's head. Furthermore, in Fig. 4 compared (a) and (b) of visual modality," + } + ], + [ + { + "type": "page_number", + "bbox": [ + 0.218, + 0.116, + 0.236, + 0.127 + ], + "angle": 0, + "content": "12" + }, + { + "type": "header", + "bbox": [ + 0.272, + 0.115, + 0.361, + 0.127 + ], + "angle": 0, + "content": "J. Park et al." + }, + { + "type": "text", + "bbox": [ + 0.242, + 0.145, + 0.478, + 0.18 + ], + "angle": 0, + "content": "Query: a girl with a ponytail wearing a red top with blue denim thigh high skirt and carrying a yellow shoulder bag looking at a phone." + }, + { + "type": "image", + "bbox": [ + 0.236, + 0.182, + 0.487, + 0.251 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.352, + 0.254, + 0.369, + 0.264 + ], + "angle": 0, + "content": "(a)" + }, + { + "type": "text", + "bbox": [ + 0.243, + 0.266, + 0.484, + 0.3 + ], + "angle": 0, + "content": "Query: the little boy is wearing his hat backwards. he has on a blue, short sleeved shirt and grey plaid shorts." + }, + { + "type": "text", + "bbox": [ + 0.517, + 0.145, + 0.751, + 0.18 + ], + "angle": 0, + "content": "Query: the woman is wearing a white cardigan over a long floaty dress and black shoes. she is carrying a back carry-on bag." + }, + { + "type": "image", + "bbox": [ + 0.513, + 0.182, + 0.765, + 0.25 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.631, + 0.253, + 0.647, + 0.264 + ], + "angle": 0, + "content": "(b)" + }, + { + "type": "text", + "bbox": [ + 0.517, + 0.266, + 0.75, + 0.31 + ], + "angle": 0, + "content": "Query: person wearing a dark brown sweater, denim cropped pants that go just below the knees, and white sneakers. they are wearing a red backpack over both shoulders." + }, + { + "type": "image", + "bbox": [ + 0.236, + 0.313, + 0.487, + 0.383 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.353, + 0.387, + 0.369, + 0.397 + ], + "angle": 0, + "content": "(c)" + }, + { + "type": "image", + "bbox": [ + 0.513, + 0.313, + 0.764, + 0.383 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.631, + 0.387, + 0.648, + 0.397 + ], + "angle": 0, + "content": "(d)" + }, + { + "type": "image_caption", + "bbox": [ + 0.214, + 0.397, + 0.787, + 0.439 + ], + "angle": 0, + "content": "Fig. 3: Top-5 retrieval results of our method on the CUHK-PEDES dataset. Images are sorted from left to right according to their ranks below each text query. Green and red boxes indicate true and false matches, respectively." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.46, + 0.788, + 0.49 + ], + "angle": 0, + "content": "our part discovery module is capable of capturing the distinctive human parts, while it is robust against pose variations and viewpoint changes." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.491, + 0.789, + 0.702 + ], + "angle": 0, + "content": "Visualization of TDPA: We visualize TDPA weights \\( \\mathbf{a} \\) in Eq. (6) to demonstrate the effectiveness of similarity aggregation between part embeddings of two modalities with TDPA. The visualization results are presented in the bottom left in Fig. 4(a, b), respectively. Comparing examples (a) and (b), the 5th part slot typically focuses on human parts associated with held objects, like bags, leading to a predicted lower TDPA weight for this part slot in (b) due to the absence of such distinctive information in its query description. However, the presence of the \"hat\" in the query description of (b) leads to having a high TDPA weight for the 8th part slot that typically focuses on the human head part. In contrast, since this related human head part information is not provided in the query description of (a), the TDPA weight for the 8th part slot in (a) is predicted to be a low value. These observations highlight the capability of TDPA, which adaptively improves part-based retrieval depending on the contents of the text query." + }, + { + "type": "title", + "bbox": [ + 0.216, + 0.725, + 0.402, + 0.739 + ], + "angle": 0, + "content": "4.4 Ablation Studies" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.75, + 0.788, + 0.841 + ], + "angle": 0, + "content": "In our ablation studies conducted on the CUHK-PEDES datasets, we evaluate the effectiveness of proposed components and their combinations in improving text-based person search performance. In Table 2, we compare several configurations: The baseline method employs only global embeddings \\(\\mathbf{g}\\) trained by InfoNCE \\((\\mathcal{L}_{\\mathrm{NCE}})\\) loss. When incorporating the cross-modal masked language modeling \\((\\mathcal{L}_{\\mathrm{CMLM}})\\), there is a slight improvement across all metrics. The addition" + } + ], + [ + { + "type": "header", + "bbox": [ + 0.242, + 0.115, + 0.732, + 0.129 + ], + "angle": 0, + "content": "PLOT: TBPS with Part Slot Attention for Corresponding Part Discovery" + }, + { + "type": "page_number", + "bbox": [ + 0.769, + 0.116, + 0.785, + 0.127 + ], + "angle": 0, + "content": "13" + }, + { + "type": "text", + "bbox": [ + 0.253, + 0.147, + 0.747, + 0.175 + ], + "angle": 0, + "content": "Text: the woman is wearing a white cardigan over a long floaty dress and black shoes. she is carrying a back carry-on bag." + }, + { + "type": "image", + "bbox": [ + 0.257, + 0.176, + 0.774, + 0.422 + ], + "angle": 0, + "content": null + }, + { + "type": "text", + "bbox": [ + 0.254, + 0.428, + 0.753, + 0.457 + ], + "angle": 0, + "content": "Text: the little boy is wearing his hat backwards. he has on a blue, short sleeved shirt and grey plaid shorts." + }, + { + "type": "image", + "bbox": [ + 0.253, + 0.46, + 0.774, + 0.715 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.214, + 0.716, + 0.788, + 0.744 + ], + "angle": 0, + "content": "Fig. 4: Visualization of each modality's attention map \\(\\bar{A}_k\\) in \\(T\\)-th iteration of PSA block and TDPA weights a on CUHK-PEDES dataset." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.78, + 0.79, + 0.842 + ], + "angle": 0, + "content": "of part embeddings \\(\\mathbf{P}\\) that without \\(\\mathcal{L}_{\\mathrm{PartID}}\\) also shows significant enhancement in R@1 and R@5, where we observe an increase of \\(3.46\\% \\mathrm{p}\\) and \\(2.64\\% \\mathrm{p}\\), respectively. Notably, the full configuration achieves the best performance with \\(\\mathcal{L}_{\\mathrm{PartID}}\\), underscoring the importance of part embeddings for accurate text-based person search." + } + ], + [ + { + "type": "page_number", + "bbox": [ + 0.218, + 0.116, + 0.236, + 0.127 + ], + "angle": 0, + "content": "14" + }, + { + "type": "header", + "bbox": [ + 0.272, + 0.115, + 0.361, + 0.127 + ], + "angle": 0, + "content": "J. Park et al." + }, + { + "type": "table_caption", + "bbox": [ + 0.303, + 0.145, + 0.698, + 0.159 + ], + "angle": 0, + "content": "Table 2: Ablation studies on the CUHK-PEDES datasets." + }, + { + "type": "table", + "bbox": [ + 0.22, + 0.171, + 0.785, + 0.277 + ], + "angle": 0, + "content": "
MethodLossCUHK-PEDES
LNCELIDLCMLMLPartNCELPartIDR@1R@5R@10
Global Only✓XXXX71.39 (+0.0)87.65 (+0.0)92.74 (+0.0)
✓✓XXX71.83 (+0.44)88.06 (+0.41)92.58 (-0.16)
+ CMLM✓✓✓XX72.65 (+1.26)88.58 (+0.93)92.93 (+0.19)
+ Part Embeddings✓✓✓✓X74.85 (+3.46)90.29 (+2.64)94.10 (+1.36)
✓✓✓✓✓75.28 (+3.89)90.42 (+2.77)94.12 (+1.38)
" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.305, + 0.784, + 0.334 + ], + "angle": 0, + "content": "This suggests that our part embeddings significantly contribute to discerning fine-grained details critical for text-based person search." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.335, + 0.503, + 0.455 + ], + "angle": 0, + "content": "Ablation Study on Part Discovery Methods: To validate the effectiveness of our proposed part discovery module, we conduct experiments by replacing our part discovery module with other methods such as TIPCB [2] and PAT [14]. TIPCB is a simple yet effective heuristic part discovery method, which extracts" + }, + { + "type": "table_caption", + "bbox": [ + 0.513, + 0.338, + 0.788, + 0.365 + ], + "angle": 0, + "content": "Table 3: Ablation study of different part discovery methods on the CUHK-PEDES." + }, + { + "type": "table", + "bbox": [ + 0.513, + 0.367, + 0.79, + 0.445 + ], + "angle": 0, + "content": "
MethodsR@1R@5R@10
Ours + TIPCB [2]73.2389.1094.04
Ours + PAT [14]72.7689.2393.42
Ours75.2890.4294.12
" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.456, + 0.788, + 0.668 + ], + "angle": 0, + "content": "image part embeddings through equi-horizontal cropping of the entire image and transforms the global embedding of the text modality into corresponding parts by using additional learnable MLPs for each image part. PAT performs part discovery with a querying transformer approach by leveraging learnable queries and conventional cross-attention. A key difference of our method from PAT is that ours explicitly encourages the discovered parts to be spatially separated since our part slots compete with each other to aggregate input data. In contrast, parts found by PAT, based on conventional cross-attention, often over-capture salient regions and are likely to miss fine details. To demonstrate this limitation, we show the visualization of part attentions with PAT applied to ours in the supplementary materials. As indicated in Table 3, our part discovery method outperforms these heuristic and conventional cross-attention based part discovery methods across all metrics, with the most significant difference observed in the R@1 metric, which requires a precise discernment of human parts." + }, + { + "type": "title", + "bbox": [ + 0.216, + 0.689, + 0.36, + 0.705 + ], + "angle": 0, + "content": "5 Conclusion" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.72, + 0.788, + 0.795 + ], + "angle": 0, + "content": "We proposed a novel framework that extracts distinctive human parts corresponding across visual and textual modalities through part discovery module without part-level supervision. The introduced TDPA further refines the retrieval process by adjusting the importance of each part slot based on the contents of the text query, leading to more precise and relevant retrieval." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.796, + 0.788, + 0.84 + ], + "angle": 0, + "content": "Limitation: Since slots learn to occupy the entire image and text, some of them may indicate irrelevant part of input. TDPA addresses this issue to some extent, but a more explicit solution would further enhance our system." + } + ], + [ + { + "type": "header", + "bbox": [ + 0.242, + 0.115, + 0.732, + 0.129 + ], + "angle": 0, + "content": "PLOT: TBPS with Part Slot Attention for Corresponding Part Discovery" + }, + { + "type": "page_number", + "bbox": [ + 0.769, + 0.117, + 0.785, + 0.127 + ], + "angle": 0, + "content": "15" + }, + { + "type": "title", + "bbox": [ + 0.218, + 0.146, + 0.403, + 0.163 + ], + "angle": 0, + "content": "Acknowledgements" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.177, + 0.793, + 0.222 + ], + "angle": 0, + "content": "This work was supported by the IITP grants and the NRF grants funded by Ministry of Science and ICT, Korea (RS-2019-II191906; RS-2022-II220926; NRF-2018R1A5A1060031; NRF-2021R1A2C3012728)" + }, + { + "type": "title", + "bbox": [ + 0.218, + 0.244, + 0.323, + 0.26 + ], + "angle": 0, + "content": "References" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.274, + 0.787, + 0.317 + ], + "angle": 0, + "content": "1. Aggarwal, S., Radhakrishnan, V.B., Chakraborty, A.: Text-based person search via attribute-aided matching. In: Proc. Winter Conference on Applications of Computer Vision (WACV) (2020)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.318, + 0.787, + 0.357 + ], + "angle": 0, + "content": "2. Chen, Y., Zhang, G., Lu, Y., Wang, Z., Zheng, Y.: Tipcb: A simple but effective part-based convolutional baseline for text-based person search. Neurocomputing 494 (2022)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.359, + 0.787, + 0.399 + ], + "angle": 0, + "content": "3. Devlin, J., Chang, M.W., Lee, K., Toutanova, K.: Bert: Pre-training of deep bidirectional transformers for language understanding. In: Annual Conference of the North American Chapter of the Association for Computational Linguistics (2019)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.401, + 0.787, + 0.427 + ], + "angle": 0, + "content": "4. Ding, Z., Ding, C., Shao, Z., Tao, D.: Semantically self-aligned network for text-to-image part-aware person re-identification. arXiv preprint arXiv:2107.12666 (2021)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.428, + 0.787, + 0.468 + ], + "angle": 0, + "content": "5. Gao, C., Cai, G., Jiang, X., Zheng, F., Zhang, J., Gong, Y., Peng, P., Guo, X., Sun, X.: Contextual non-local alignment over full-scale representation for text-based person search. arXiv preprint arXiv:2101.03036 (2021)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.47, + 0.787, + 0.509 + ], + "angle": 0, + "content": "6. Gray, D., Brennan, S., Tao, H.: Evaluating appearance models for recognition, reacquisition, and tracking. In: Proc. IEEE international workshop on performance evaluation for tracking and surveillance (PETS) (2007)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.511, + 0.787, + 0.55 + ], + "angle": 0, + "content": "7. Jiang, D., Ye, M.: Cross-modal implicit relation reasoning and aligning for text-to-image person retrieval. In: IEEE International Conference on Computer Vision and Pattern Recognition (CVPR) (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.552, + 0.787, + 0.592 + ], + "angle": 0, + "content": "8. Jing, Y., Si, C., Wang, J., Wang, W., Wang, L., Tan, T.: Pose-guided multi-granularity attention network for text-based person search. In: Proc. AAAI Conference on Artificial Intelligence (AAAI) (2020)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.594, + 0.787, + 0.62 + ], + "angle": 0, + "content": "9. Lee, K.H., Chen, X., Hua, G., Hu, H., He, X.: Stacked cross attention for image-text matching. In: Proc. European Conference on Computer Vision (ECCV) (2018)" + }, + { + "type": "ref_text", + "bbox": [ + 0.22, + 0.621, + 0.787, + 0.661 + ], + "angle": 0, + "content": "10. Li, S., Cao, M., Zhang, M.: Learning semantic-aligned feature representation for text-based person search. In: Proc. IEEE International Conference on Acoustics, Speech and Signal Processing (ICASSP). IEEE (2022)" + }, + { + "type": "ref_text", + "bbox": [ + 0.22, + 0.662, + 0.787, + 0.702 + ], + "angle": 0, + "content": "11. Li, S., Xiao, T., Li, H., Zhou, B., Yue, D., Wang, X.: Person search with natural language description. In: Proc. IEEE Conference on Computer Vision and Pattern Recognition (CVPR) (2017)" + }, + { + "type": "ref_text", + "bbox": [ + 0.22, + 0.703, + 0.787, + 0.757 + ], + "angle": 0, + "content": "12. Li, W., Zhao, R., Wang, X.: Human reidentification with transferred metric learning. In: Computer Vision-ACCV 2012: 11th Asian Conference on Computer Vision, Daejeon, Korea, November 5-9, 2012, Revised Selected Papers, Part I 11. Springer (2013)" + }, + { + "type": "ref_text", + "bbox": [ + 0.22, + 0.758, + 0.787, + 0.799 + ], + "angle": 0, + "content": "13. Li, W., Zhao, R., Xiao, T., Wang, X.: Deepreid: Deep filter pairing neural network for person re-identification. In: Proc. IEEE Conference on Computer Vision and Pattern Recognition (CVPR) (2014)" + }, + { + "type": "ref_text", + "bbox": [ + 0.22, + 0.8, + 0.787, + 0.84 + ], + "angle": 0, + "content": "14. Li, Y., He, J., Zhang, T., Liu, X., Zhang, Y., Wu, F.: Diverse part discovery: Occluded person re-identification with part-aware transformer. In: Proc. IEEE Conference on Computer Vision and Pattern Recognition (CVPR) (2021)" + }, + { + "type": "list", + "bbox": [ + 0.22, + 0.274, + 0.787, + 0.84 + ], + "angle": 0, + "content": null + } + ], + [ + { + "type": "page_number", + "bbox": [ + 0.218, + 0.116, + 0.236, + 0.127 + ], + "angle": 0, + "content": "16" + }, + { + "type": "header", + "bbox": [ + 0.272, + 0.115, + 0.361, + 0.127 + ], + "angle": 0, + "content": "J. Park et al." + }, + { + "type": "ref_text", + "bbox": [ + 0.218, + 0.148, + 0.787, + 0.175 + ], + "angle": 0, + "content": "15. Locatello, F., Weissenborn, D., Unterthiner, T., Mahendran, A., Heigold, G., Uszkoreit, J., Dosovitskiy, A., Kipf, T.: Object-centric learning with slot attention (2020)" + }, + { + "type": "ref_text", + "bbox": [ + 0.218, + 0.177, + 0.787, + 0.217 + ], + "angle": 0, + "content": "16. Niu, K., Huang, Y., Ouyang, W., Wang, L.: Improving description-based person re-identification by multi-granularity image-text alignments. IEEE Transactions on Image Processing 29, 5542-5556 (2020)" + }, + { + "type": "ref_text", + "bbox": [ + 0.218, + 0.218, + 0.787, + 0.244 + ], + "angle": 0, + "content": "17. Oord, A.v.d., Li, Y., Vinyals, O.: Representation learning with contrastive predictive coding. arXiv preprint arXiv:1807.03748 (2018)" + }, + { + "type": "ref_text", + "bbox": [ + 0.218, + 0.245, + 0.787, + 0.3 + ], + "angle": 0, + "content": "18. Radford, A., Kim, J.W., Hallacy, C., Ramesh, A., Goh, G., Agarwal, S., Sastry, G., Askell, A., Mishkin, P., Clark, J., et al.: Learning transferable visual models from natural language supervision. In: Proc. International Conference on Machine Learning (ICML) (2021)" + }, + { + "type": "ref_text", + "bbox": [ + 0.218, + 0.3, + 0.787, + 0.342 + ], + "angle": 0, + "content": "19. Sarafianos, N., Xu, X., Kakadiaris, I.A.: Adversarial representation learning for text-to-image matching. In: Proc. IEEE International Conference on Computer Vision (ICCV) (2019)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.342, + 0.787, + 0.383 + ], + "angle": 0, + "content": "20. Shao, Z., Zhang, X., Fang, M., Lin, Z., Wang, J., Ding, C.: Learning granularity-unified representations for text-to-image person re-identification. In: Proc. ACM Multimedia Conference (ACMMM) (2022)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.383, + 0.787, + 0.438 + ], + "angle": 0, + "content": "21. Shu, X., Wen, W., Wu, H., Chen, K., Song, Y., Qiao, R., Ren, B., Wang, X.: See finer, see more: Implicit modality alignment for text-based person retrieval. In: Proc. European Conference on Computer Vision Workshop on Real-World Surveillance, (ECCVW) (2022)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.439, + 0.787, + 0.481 + ], + "angle": 0, + "content": "22. Suo, W., Sun, M., Niu, K., Gao, Y., Wang, P., Zhang, Y., Wu, Q.: A simple and robust correlation filtering method for text-based person search. In: Proc. European Conference on Computer Vision (ECCV) (2022)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.481, + 0.787, + 0.522 + ], + "angle": 0, + "content": "23. Wang, C., Luo, Z., Lin, Y., Li, S.: Text-based person search via multi-granularity embedding learning. In: Proc. International Joint Conferences on Artificial Intelligence (IJCAI) (2021)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.522, + 0.787, + 0.563 + ], + "angle": 0, + "content": "24. Wang, Z., Fang, Z., Wang, J., Yang, Y.: Vitaa: Visual-textual attributes alignment in person search by natural language. In: Proc. European Conference on Computer Vision (ECCV) (2020)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.564, + 0.787, + 0.605 + ], + "angle": 0, + "content": "25. Wang, Z., Zhu, A., Xue, J., Wan, X., Liu, C., Wang, T., Li, Y.: Caibc: Capturing all-round information beyond color for text-based person retrieval. In: Proc. ACM Multimedia Conference (ACMMM) (2022)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.606, + 0.787, + 0.646 + ], + "angle": 0, + "content": "26. Wei, L., Zhang, S., Gao, W., Tian, Q.: Person transfer gan to bridge domain gap for person re-identification. In: Proc. IEEE Conference on Computer Vision and Pattern Recognition (CVPR) (2018)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.647, + 0.787, + 0.688 + ], + "angle": 0, + "content": "27. Wu, Y., Yan, Z., Han, X., Li, G., Zou, C., Cui, S.: Lipscore: language-guided person search via color reasoning. In: Proc. IEEE International Conference on Computer Vision (ICCV) (2021)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.689, + 0.787, + 0.716 + ], + "angle": 0, + "content": "28. Xiao, T., Li, S., Wang, B., Lin, L., Wang, X.: End-to-end deep learning for person search. arXiv preprint arXiv:1604.01850 (2016)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.717, + 0.787, + 0.743 + ], + "angle": 0, + "content": "29. Yan, S., Dong, N., Zhang, L., Tang, J.: Clip-driven fine-grained text-image person re-identification. IEEE Transactions on Image Processing (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.744, + 0.787, + 0.771 + ], + "angle": 0, + "content": "30. Zhang, Y., Lu, H.: Deep cross-modal projection learning for image-text matching. In: Proc. European Conference on Computer Vision (ECCV) (2018)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.772, + 0.787, + 0.812 + ], + "angle": 0, + "content": "31. Zheng, L., Shen, L., Tian, L., Wang, S., Wang, J., Tian, Q.: Scalable person re-identification: A benchmark. In: Proc. IEEE International Conference on Computer Vision (ICCV) (2015)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.813, + 0.787, + 0.84 + ], + "angle": 0, + "content": "32. Zhu, A., Wang, Z., Li, Y., Wan, X., Jin, J., Wang, T., Hu, F., Hua, G.: Dssl: deep surroundings-person separation learning for text-based person retrieval. In:" + }, + { + "type": "list", + "bbox": [ + 0.217, + 0.148, + 0.787, + 0.84 + ], + "angle": 0, + "content": null + } + ], + [ + { + "type": "header", + "bbox": [ + 0.242, + 0.115, + 0.732, + 0.129 + ], + "angle": 0, + "content": "PLOT: TBPS with Part Slot Attention for Corresponding Part Discovery" + }, + { + "type": "page_number", + "bbox": [ + 0.768, + 0.117, + 0.785, + 0.127 + ], + "angle": 0, + "content": "17" + }, + { + "type": "text", + "bbox": [ + 0.242, + 0.148, + 0.788, + 0.175 + ], + "angle": 0, + "content": "Proceedings of the 29th ACM International Conference on Multimedia. pp. 209-217 (2021)" + } + ] +] \ No newline at end of file diff --git a/2024/PLOT_ Text-based Person Search with Part Slot Attention for Corresponding Part Discovery/aafb2e8f-8788-410d-8a77-952604d06d27_origin.pdf b/2024/PLOT_ Text-based Person Search with Part Slot Attention for Corresponding Part Discovery/aafb2e8f-8788-410d-8a77-952604d06d27_origin.pdf new file mode 100644 index 0000000000000000000000000000000000000000..30de91c9ddeb2a16d29ae3d1760cd7e3b9813c1d --- /dev/null +++ b/2024/PLOT_ Text-based Person Search with Part Slot Attention for Corresponding Part Discovery/aafb2e8f-8788-410d-8a77-952604d06d27_origin.pdf @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:2dd80f084b7d3c556baa902b429996042e5921b83ea1243839f7559749636934 +size 1598886 diff --git a/2024/PLOT_ Text-based Person Search with Part Slot Attention for Corresponding Part Discovery/full.md b/2024/PLOT_ Text-based Person Search with Part Slot Attention for Corresponding Part Discovery/full.md new file mode 100644 index 0000000000000000000000000000000000000000..d86834d17e23be20f60e1cb4e6db358239c46533 --- /dev/null +++ b/2024/PLOT_ Text-based Person Search with Part Slot Attention for Corresponding Part Discovery/full.md @@ -0,0 +1,305 @@ +# PLOT: Text-based Person Search with Part Slot Attention for Corresponding Part Discovery + +Jicheol Park D, Dongwon Kim D, Boseung Jeong D, and Suha Kwak D + +Pohang University of Science and Technology (POSTECH), South Korea {jicheol, kdwon,boseung01, suha.kwak}@postech.ac.kr https://cvlab.postech.ac.kr/research/PLOT + +Abstract. Text-based person search, employing free-form text queries to identify individuals within a vast image collection, presents a unique challenge in aligning visual and textual representations, particularly at the human part level. Existing methods often struggle with part feature extraction and alignment due to the lack of direct part-level supervision and reliance on heuristic features. We propose a novel framework that leverages a part discovery module based on slot attention to autonomously identify and align distinctive parts across modalities, enhancing interpretability and retrieval accuracy without explicit part-level correspondence supervision. Additionally, text-based dynamic part attention adjusts the importance of each part, further improving retrieval outcomes. Our method is evaluated on three public benchmarks, significantly outperforming existing methods. + +Keywords: Text-Based Person Search $\cdot$ Multi-Modal Retrieval + +# 1 Introduction + +Text-based person search is the task of identifying the target person from the vast collection of images with a free-form text query. This task demands extracting identifiable features, such as human parts, from both textual and visual modalities to capture subtle differences between individuals. Hence, establishing correspondence between the extracted human part features across image and text modalities is essential for accurate text-based person search. However, it is not straightforward to extract these part features and establish their correspondences between the two modalities without part-level supervision. + +To address this challenge, previous work [2, 22] relies on the heuristic part features obtained by equi-horizontally cropping the entire image; such features are then matched to the free-form text queries for person search. However, the heuristic part features used in this approach are susceptible to deformation caused by occlusion and pose variation. Meanwhile, earlier work [8, 20, 24] proposed a learning-based approach for part feature extraction. Nevertheless, these methods tend to generate redundant part features that lack disentanglement [20], or they demand access to additional part-level supervisions [8, 24]. + +![](images/8a390316999a60ab9111737922f170dec25ac14a89cbf5747f3a9421ce3c96f6.jpg) +Fig. 1: The overall architecture of PLOT. + +To tackle the above issues, we introduce a new framework that discovers distinctive parts in both modalities and matches them between the two modalities without any correspondence supervision; its overall architecture is illustrated in Fig. 1. To discover distinctive parts and extract their features from both modalities, we propose PLOT, a Part discovery module based on the sLOT attention mechanism [15]. Slot attention is an attention mechanism designed for object-centric learning, which segments input data into a set of slots representing individual entities without requiring object-level supervision. In PLOT, we first define a set of learnable embedding vectors, termed part slots, that contain primitive information related to human body parts shared between the two modalities. Then, these part slots undergo refinement through several iterative attention processes, where they compete amongst themselves to bind with the input data; ultimately, the part slots are transformed into part features, termed part embeddings, that represent distinctive parts in the input data. + +To ensure correspondence between part embeddings from different modalities, the part slots are shared between the visual and textual modalities. Part embeddings from the same part slots are then learned to represent the identical part although they are computed from different modalities. This mechanism enables PLOT to match the discovered parts from the two modalities without supervision for the correspondence as well as capturing part-level fine-grained appearance features from the both modalities. Hence, PLOT improves performance of text-based person search through the rich and fine-grained part features, and at the same time, it guarantees interpretable retrieval by providing part-level correspondences between query text and retrieved images. + +In addition, PLOT introduces a new similarity aggregation method for part embeddings called text-based dynamic part attention (TDPA) pooling, which dynamically adjusts the weights of part embeddings based on the text query. Given a text query as input, TDPA predicts the importance weight of each + +slot for retrieval and applies the predicted importance weights to aggregate the similarities of part embeddings between the two modalities. TDPA allows the entire retrieval system to perform optimized retrieval for each query, leading to improved performance. + +Our method was evaluated and compared with prior work on three public benchmarks [4, 11, 32], where it clearly outperformed all existing methods thanks to the rich representation based on part embeddings. The main contribution of our work is four-fold: + +- We introduce PLOT, a new framework for text-based person search that discovers distinguished human body parts, extracts their embeddings, and establishes their correspondences between the two modalities with no human intervention. +- The part embeddings provided by PLOT enables an interpretable text-based person search thanks to the part-level correspondences it provides. +- We introduce a novel similarity aggregation method that adaptively determines importance of each discovered part based on each text query and consequently enables retrieval optimized per query. +- Our model with PLOT achieved the best on all the public benchmarks for text-based person search. + +# 2 Related Work + +# 2.1 Text-Based Person Search + +In recent years, the task of text-based person search has gained significant attention in the computer vision community. Li et al., [11] proposed a gated neural attention-based recurrent neural network (GNA-RNN) for learning the affinity between text descriptions and images, along with providing a benchmark dataset CUHK-PEDES for model evaluation. Zhang et al., [30] proposed cross-modal projection matching and classification (CMPM+CMPC) loss, for learning deep discriminative image-text global embeddings. However, these methods primarily focus on global representations of input data and are thus not capable of capturing distinctive human part details, leading to limited performance in the text-based person search. + +To address the above problem, a line of studies focuses on extracting fine-grained representations. One of the prominent examples of exploiting fine-grained information is to cut human images horizontally and use them as human parts [2, 4, 5, 16, 22, 23]. Chen et al., [2] extracts image part embeddings through equi-horizontal cropping of the entire image and aligns additional network to transform the textual global into corresponding parts. However, these heuristic part features inevitably include no-informative information, such as background elements. To avoid the above limitations, Suo et al., [22] proposed the simple and robust correlation filtering method to extract foreground features on the heuristic part features. Yet, the heuristically divided part features fundamentally fall short of capturing the complex human parts, underscoring a critical limitation in their + +expressive capability. To move beyond such heuristic part structures, Shao et al., [2] proposes a learning-based approach for part feature extraction. However, this method struggles to extract distinctive part embeddings due to the extraction of redundant part information, which lacks sufficient disentanglement. To extract exquisite human parts, several studies tried to utilize useful information (e.g., human attributes and human keypoints) via external tools [1,8,24]. Wang et al., [24] introduced an auxiliary attribute segmentation to align the visual part features with the textual attributes parsed from text description. Jing et al., [8] proposed a new multi-granularity attention network to learn the part feature alignment between visual and textual with human pose estimation. However, these approaches have inevitable limitations of high computational cost and dependence on the performance of external tools for local feature extraction. + +Most recent work, Jiang et al., [7] utilizes pre-trained CLIP [18] model, capitalizing on the rich knowledge of models trained on extensive data for text-to-image matching, to excel in text-based person search. Yet, this methods primarily focus on global features, not specifically designed for extracting human part features. Not only do we leverage pre-trained knowledge, but we transcend heuristic human parts methods to extract sophisticated human parts without external tools or part-level supervision. + +# 2.2 Slot Attention + +Slot attention [15] is a recently proposed attention mechanism for object-centric learning, a problem focusing on discovering constituent visual entities within an image. The unique property of slot attention is that it can represent input images as a set of slots, where the slots are representations corresponding to individual visual entities, without any object-level supervision during training. Within slot attention, slots iteratively compete for aggregating input data, ensuring distinct representations focusing on individual visual entities. By incorporating slot attention, the proposed framework facilitates the unsupervised identification of a structure underlying image and text queries, enabling the model to discern and represent individual human parts without explicit supervision. This capability is particularly valuable in person search datasets, where recognizing subtle differences and understanding the correspondence of human parts across different modalities are crucial. + +# 3 Proposed Method + +The following subsection offers details of global and part embeddings extraction for each modality with its backbone (Sec 3.1). Subsequently, we present our novel framework, PLOT, which includes the part discovery module (Sec. 3.2) and similarity aggregation between part embeddings (Sec. 3.3), concluding with a discussion on the learning objective designed to optimize our proposed framework (Sec. 3.4) and inference of our framework (Sec. 3.5). + +# 3.1 Global and Part Embeddings + +In our framework, a single input data is described by two different types of representations: a global embedding and multiple part embeddings. The global embedding is used to represent the input data holistically, while each part embedding describes the appearance of distinctive human parts (e.g., arm, leg, torso, etc.). We below provide details of global and part embeddings computation for visual and textual modalities. Following the previous work [7], a pre-trained CLIP [18] is used as the backbone networks for the visual and textual modalities. Visual Modality: We utilize the vision transformer (ViT) from the CLIP-B/16 [18] architecture as a visual backbone network. Initially, an input image of a person is split into $N$ distinct, non-overlapping patches, which are subsequently transformed into patch tokens through linear projection. The patch tokens and an extra [c1s] token are then fed into the visual backbone network. The token sequence is processed throughout multiple self-attention blocks, and the [c1s] token of the last block is used as a global embedding $\mathbf{g}^{\mathcal{V}} \in \mathbb{R}^{D}$ . To obtain part embeddings $\mathbf{P}^{\mathcal{V}} \in \mathbb{R}^{K \times D}$ , remaining patch tokens of last block $\mathbf{x}^{\mathcal{V}} \in \mathbb{R}^{N \times D}$ is passed to the part discovery module. The part discovery module aggregates the patch features describing coherent human parts into the same part embedding. We provide more detailed information about the part discovery module in Sec. 3.2. Textual Modality: For the textual backbone network, we utilize a transformer architecture from the CLIP-Xformer [18] text encoder. This encoder operates on text input transformed into byte pair encoding (BPE) sequences. Initially, the text query undergoes tokenization via BPE, followed by the enclosing with [S0S] and [E0S] tokens. The resulting sequence of tokens is then inputted into the textual backbone network. Here, the [E0S] token from the final block serves as the global embedding, denoted as $\mathbf{g}^{\mathcal{T}} \in \mathbb{R}^{D}$ . Analogous to the approach for the visual modality, we process the remainder of the text tokens, $\mathbf{x}^{\mathcal{T}} \in \mathbb{R}^{L \times D}$ , into part embeddings, $\mathbf{P}^{\mathcal{T}} \in \mathbb{R}^{K \times D}$ , utilizing the part discovery module. + +# 3.2 Part Discovery Module + +For extracting part embeddings in each modality, part discovery module aggregates patch tokens $\mathbf{x}^{\mathcal{V}}$ and text tokens $\mathbf{x}^{\mathcal{T}}$ into the visual part embeddings $\mathbf{P}^{\mathcal{V}}$ and textual part embeddings $\mathbf{P}^{\mathcal{T}}$ , respectively. It is worth noting that the part discovery module for each modality has identical model architecture and functions equivalently. Therefore, we will only explain the part discovery module on the visual modality for brevity. + +Part discovery module consists of initial part slots $\mathbf{S}^0\in \mathbb{R}^{K\times D}$ and $T$ multiple iteration of the part slot attention block (PSA block). To extract part embeddings, we first initialize a set of learnable embeddings part slots $\mathbf{S}^0\in \mathbb{R}^{K\times D}$ , where $K$ indicates the number of part slots. Then, through a series of $T$ iterations of our PSA block, the initial $\mathbf{S}^0$ evolves into refined $\mathbf{S}^T$ , where each slot captures distinct parts within input data. The refined part slots $\mathbf{S}^T$ are used as visual part embeddings $\mathbf{P}^{\mathcal{V}}$ . The part discovery module can be formulated as follows: + +$$ +\mathbf {P} ^ {\mathcal {V}} := \mathbf {S} ^ {T}, \text {w h e r e} \quad \mathbf {S} ^ {t} = \operatorname {P S A} _ {-} \operatorname {B l o c k} ^ {\mathcal {V}} \left(\mathbf {x} ^ {\mathcal {V}}; \mathbf {S} ^ {t - 1}\right). \tag {1} +$$ + +The PSA block first transforms the inputs $\mathbf{S}^{t - 1}$ and $\mathbf{x}^{\mathcal{V}}$ with layer normalization and linear projection layers $q(\cdot), k(\cdot)$ and $\nu (\cdot)$ to obtain embeddings of $D_{h}$ dimension. Then the attention map $A\in \mathbb{R}^{N\times K}$ between $\mathbf{S}^{t - 1}$ and $\mathbf{x}^{\mathcal{V}}$ is computed by + +$$ +A _ {n, k} = \frac {e ^ {M _ {n , k}}}{\sum_ {i = 1} ^ {K} e ^ {M _ {n , i}}}, \mathrm {w h e r e} M = \frac {k (\mathbf {x} ^ {\mathcal {V}}) q (\mathbf {S} ^ {t - 1}) ^ {\top}}{\sqrt {D _ {h}}}. (2) +$$ + +The attention map is obtained through normalization across part slots; this normalization encourages competition among themselves to bind distinct sets of patch tokens to each slot. We update part slots with weighted mean of patch tokens obtained by attention map $A_{n,k}$ and then feed it to a gated recurrent unit (GRU) using $\mathbf{S}^{t-1}$ as hidden state as follows: + +$$ +\bar {A} _ {n, k} = \frac {A _ {n , k}}{\sum_ {i = 1} ^ {N} A _ {i , k}}, \bar {\mathbf {S}} ^ {t} = \mathbb {G R U} \left(\mathbf {S} ^ {t - 1}, \bar {A} ^ {\top} v \left(\mathbf {x} ^ {\mathcal {V}}\right)\right). \tag {3} +$$ + +Then, we obtain $t$ -th part slot by feeding $\overline{\mathbf{S}}^{t-1}$ into a multi-layer perceptron (MLP) with layer normalization, ReLU activation, and residual connection: + +$$ +\mathbf {S} ^ {t} = \operatorname {P S A} _ {\text {B l o c k}} ^ {\mathcal {V}} (\mathbf {x} ^ {\mathcal {V}}; \mathbf {S} ^ {t - 1}) = \operatorname {M L P} \left(\bar {\mathbf {S}} ^ {t - 1}\right) + \bar {\mathbf {S}} ^ {t - 1}. \tag {4} +$$ + +Finally, we can obtain the visual part embeddings $\mathbf{P}^{\mathcal{V}} = \{\mathbf{p}_k^{\mathcal{V}}\}_{k=1}^K$ which is the output of $T$ -th iteration of PSA block: $\mathbf{P}^{\mathcal{V}} = \mathbf{S}_{\mathcal{V}}^T \in \mathbb{R}^{K \times D}$ . + +Part Correspondence through Slot Sharing: Additionally, we share the learnable part slots $\mathbf{S}^0$ between two part discovery modules to establish correspondences between part embeddings extracted from each modality, considering that part embeddings extracted from the same part slot are corresponding part across modalities, thereby contributing to a clearer comparison between modalities. + +# 3.3 Measuring Similarity between Embeddings + +In the context of training and applying our retrieval model, selecting an appropriate similarity function between embeddings is crucial. The challenge lies in dealing with two distinct types of embeddings: global embeddings and part embeddings. For global embeddings, cosine similarity offers a straightforward and effective means of measuring similarity. + +However, the situation becomes more complex when considering part embeddings. A direct method for addressing this complexity involves calculating the average cosine similarity across all pairs of part embeddings. This approach, while straightforward, has its drawbacks, primarily because it treats all part-wise similarities as equally significant. In reality, the relevance of specific part embeddings to the actual similarity between data instances can significantly vary, influenced by the context of a text query. For instance, if a text query focuses exclusively on particular features of a human figure, the similarity contributions from other unrelated part embeddings should be less relevant. This challenge highlights the necessity for an approach that can dynamically assess and prioritize the relevance of part embeddings based on the context provided by the query. + +To address this challenge, we introduce text-based dynamic part attention (TDPA) to aggregate the similarities between part embeddings. Firstly, textual global embedding $\mathbf{g}^{\mathcal{T}}$ is transformed to the TDPA $\mathbf{a} \in \mathbb{R}^{K}$ , using MLP and a softmax function. Using TDPA, the aggregated similarity between part embeddings are computed as follows: + +$$ +\mathbf {a} = \sigma (\operatorname {M L P} \left(\mathbf {g} ^ {\mathcal {T}}\right)) \in \mathbb {R} ^ {K}, \tag {5} +$$ + +$$ +c _ {\mathrm {a g g}} \left(\mathbf {P} ^ {\mathcal {V}}, \mathbf {P} ^ {\mathcal {T}}; \mathbf {g} ^ {\mathcal {T}}\right) := \sum_ {k = 1} ^ {K} \mathrm {a} _ {k} \cdot c \left(\mathbf {p} _ {k} ^ {\mathcal {V}}, \mathbf {p} _ {k} ^ {\mathcal {T}}\right), \tag {6} +$$ + +where $c(\cdot, \cdot)$ denotes cosine similarity between two embeddings, $a_k$ is the $k$ -th value of $\mathbf{a}$ , and $\sigma(\cdot)$ is a softmax function. If given textual global embedding $\mathbf{g}^{\mathcal{T}}$ , TDPA is computed by MLP which is learned in an end-to-end manner by minimizing our partNCE loss, as will be introduced in Sec 3.4. This dynamic attention enables us to adaptively assign importance to each part embedding, which reduces the impact of non-informative part embedding and promotes a more informative similarity measurement that reflects the significant semantic similarity of each text query. + +# 3.4 Learning Objective + +Our model is trained through the establishment of cross-modal alignments, considering both global and part levels. The alignment at the global-level is accomplished by aligning global embeddings $\mathbf{g}^{\mathcal{V}}$ and $\mathbf{g}^{\mathcal{T}}$ that capture the comprehensive information of each modality. For part-level alignment, we leverage part embeddings $\mathbf{P}^{\mathcal{V}}$ and $\mathbf{P}^{\mathcal{T}}$ that have locally exclusive features within each modality due to the slot attention. Furthermore, by dynamically adjusting weights of specific part slots based on the text global embedding $\mathbf{g}^{\mathcal{T}}$ , we facilitate the learning of more informative alignments. + +Global Alignment Loss: To align global embeddings $\mathbf{g}^{\mathcal{V}}$ and $\mathbf{g}^{\mathcal{T}}$ extracted from each modality backbone, we first define a batch of global embeddings $\mathcal{B}_{\mathrm{global}} = \{(\mathbf{g}_i^{\mathcal{V}},\mathbf{g}_i^{\mathcal{T}})\}_{i = 1}^B$ , where $B$ is batch size. Then we adopt the InfoNCE loss [17] which is a contrastive learning objective that maximizes the similarity between embeddings of positive pairs while minimizing the similarity between negative pairs in the batch. Consequently, our global NCE loss with cosine similarity function $c(\cdot ,\cdot)$ is formulated as follows: + +$$ +\mathcal {L} _ {\mathrm {N C E}} = - \sum_ {i = 1} ^ {B} \left(\log \frac {e ^ {c \left(\mathbf {g} _ {i} ^ {\mathcal {V}} , \mathbf {g} _ {i} ^ {\mathcal {T}}\right) / \tau}}{\sum_ {j = 1} ^ {B} e ^ {c \left(\mathbf {g} _ {i} ^ {\mathcal {V}} , \mathbf {g} _ {j} ^ {\mathcal {T}}\right) / \tau}} + \log \frac {e ^ {c \left(\mathbf {g} _ {i} ^ {\mathcal {V}} , \mathbf {g} _ {i} ^ {\mathcal {T}}\right) / \tau}}{\sum_ {j = 1} ^ {B} e ^ {c \left(\mathbf {g} _ {j} ^ {\mathcal {V}} , \mathbf {g} _ {i} ^ {\mathcal {T}}\right) / \tau}}\right), \tag {7} +$$ + +where the $\tau$ is temperature term. Additionally, we employ an identity classification loss $\mathcal{L}_{\mathrm{ID}}$ to ensure that the embeddings extracted from the same identity become similar. The $\mathcal{L}_{\mathrm{ID}}$ is denoted by + +$$ +\mathcal {L} _ {\mathrm {I D}} = - \sum_ {i = 1} ^ {B} \left(\boldsymbol {y} _ {i} \log \sigma \left(\mathbf {g} _ {i} ^ {\mathcal {V}} \mathbf {W} _ {\mathrm {I D}}\right) + \boldsymbol {y} _ {i} \log \sigma \left(\mathbf {g} _ {i} ^ {\mathcal {T}} \mathbf {W} _ {\mathrm {I D}}\right)\right), \tag {8} +$$ + +where $\mathbf{y}_i\in \mathbb{R}^C$ is the identity ground truth of corresponding global embedding $\mathbf{g}_i$ represented by a one-hot vector and $C$ is the number of identities, $\mathbf{W}_{\mathrm{ID}}\in \mathbb{R}^{D\times C}$ is a classifier shared between the two modalities. For cross-modal alignments, the commonly used CMPM loss [7,30] is additionally adopted. Finally, we describe global alignment loss as below: + +$$ +\mathcal {L} _ {\text {G l o b a l}} = \mathcal {L} _ {\text {N C E}} + \mathcal {L} _ {\text {I D}} + \mathcal {L} _ {\text {C M P M}}. \tag {9} +$$ + +Part Alignment Loss: Similar to the alignment of global embeddings, we first define a batch of part embeddings $\mathcal{B}_{\mathrm{part}} = \{(\mathbf{P}_i^{\mathcal{V}},\mathbf{P}_i^{\mathcal{T}})\}_{i = 1}^B$ with batch size $B$ , and we adopt the InfoNCE loss to align part embeddings extracted from two modalities. For learning text query-based informative alignment of part embeddings between modalities, we employ text query-based similarity aggregation function $c_{\mathrm{agg}}$ in Eq. 6 to compute InfoNCE loss. We termed this InfoNCE with $c_{\mathrm{agg}}$ as PartNCE loss, and it is formulated by + +$$ +\mathcal {L} _ {\mathrm {P a r t N C E}} = - \sum_ {i = 1} ^ {B} \left(\log \frac {e ^ {c _ {\mathrm {a g g}} (\mathbf {P} _ {i} ^ {\mathcal {V}} , \mathbf {P} _ {i} ^ {\mathcal {T}} ; \mathbf {g} _ {i} ^ {\mathcal {T}}) / \tau}}{\sum_ {j = 1} ^ {B} e ^ {c _ {\mathrm {a g g}} (\mathbf {P} _ {i} ^ {\mathcal {V}} , \mathbf {P} _ {j} ^ {\mathcal {T}} ; \mathbf {g} _ {j} ^ {\mathcal {T}}) / \tau}} + \log \frac {e ^ {c _ {\mathrm {a g g}} (\mathbf {P} _ {i} ^ {\mathcal {V}} , \mathbf {P} _ {i} ^ {\mathcal {T}} ; \mathbf {g} _ {i} ^ {\mathcal {T}}) / \tau}}{\sum_ {j = 1} ^ {B} e ^ {c _ {\mathrm {a g g}} \left(\mathbf {P} _ {j} ^ {\mathcal {V}}, \mathbf {P} _ {i} ^ {\mathcal {T}}; \mathbf {g} _ {i} ^ {\mathcal {T}}\right) / \tau}\right), \tag {10} +$$ + +Similar to global alignment loss, we adopt the identity loss that shares classifier weights between the two modalities for part alignment; however, the difference is that the part embeddings are concatenated along the embedding dimension. The identity loss for part alignment is formulated by + +$$ +\mathcal {L} _ {\text {P a r t I D}} = - \sum_ {i = 1} ^ {B} \left(\boldsymbol {y} _ {i} \log \sigma \left(\left[ \boldsymbol {P} _ {i} ^ {\mathcal {V}} \right] \boldsymbol {W} _ {\text {P a r t I D}}\right) + \boldsymbol {y} _ {i} \log \sigma \left(\left[ \boldsymbol {P} _ {i} ^ {\mathcal {T}} \right] \boldsymbol {W} _ {\text {P a r t I D}}\right)\right), \tag {11} +$$ + +where $[\mathbf{P}] \in \mathbb{R}^{KD}$ is the concatenation of part embeddings $\mathbf{P}$ along the embedding dimension, and $\mathbf{W}_{\mathrm{PartID}} \in \mathbb{R}^{KD \times C}$ is a classifier shared between both modalities. Finally, the part alignment loss is computed by + +$$ +\mathcal {L} _ {\text {P a r t}} = \mathcal {L} _ {\text {P a r t N C E}} + \mathcal {L} _ {\text {P a r t I D}}. \tag {12} +$$ + +Cross-Modal Masked Language Modeling Loss: Following the practices of employing Transformer-based backbones in existing methods [7,21], we adopt an auxiliary loss, the cross-modal masked language modeling (CMLM) loss, to facilitate the learning of interactions between modali + +![](images/d6bcbcab86d9b112f25ee0477689121305b220c886e6338817255b1b29ca810c.jpg) +Cross-Modal Masked Language Model (CMLM) +Fig. 2: Illustration of CMLM. + +ties. Similar to BERT [3], given a text description, we randomly select text tokens with a $15\%$ probability and replace them with the learnable [MASK] token. The masked text description is then processed through the text backbone to obtain the masked textual tokens. After obtaining the masked textual tokens, we concatenate them with the visual tokens extracted via the image backbone. + +The concatenated tokens are then fed into a transformer to acquire cross-modal fused tokens. Among the fused tokens, those corresponding to the indices of textual tokens are denoted as $\mathbf{F}$ composed as $\{\mathbf{f}_1,\dots \mathbf{f}_L\}$ , where a fused token $\mathbf{f}\in \mathbb{R}^{1\times D}$ and $L$ indicates the max length of input text descriptions. Ultimately, these tokens are fed into a CMLM classifier $\mathbf{W}_{\mathrm{CMLM}}$ to predict the probability of vocabulary IDs. The overall procedure of CMLM is illustrated in Fig. 2 and through this procedure we compute the $\mathcal{L}_{\mathrm{CMLM}}$ loss as follows: + +$$ +\mathcal {L} _ {\mathrm {C M L M}} = - \frac {1}{L} \sum_ {l = 1} ^ {L} \mathbf {y} _ {l} \log (\sigma (\mathbf {f} _ {l} \mathbf {W} _ {\mathrm {C M L M}})), \tag {13} +$$ + +where $\mathbf{W}_{\mathrm{CMLM}} \in \mathbb{R}^{D \times V}$ and $\pmb{y}_l \in \mathbb{R}^V$ is the vocabulary ground truth of $l$ -th text tokens, represented by a one-hot vector and $V$ is the size of vocabulary. Finally, our overall objective function for training is denoted by + +$$ +\mathcal {L} = \mathcal {L} _ {\text {G l o b a l}} + \mathcal {L} _ {\text {P a r t}} + \mathcal {L} _ {\text {C M L M}}. \tag {14} +$$ + +# 3.5 Inference + +During testing, the global and part embeddings of each modality input are fully exploited to calculate the similarity between the image-text pair. In particular, the similarities between visual part embeddings and their corresponding textual part embeddings are linearly combined with the attention weights $\mathbf{a}$ to aggregate them. The image-text pair similarity is defined as the sum of the similarity between the global embeddings of the image-text pair and the similarity between the part embeddings of it, which can be computed by $c(\mathbf{g}^{\mathcal{V}},\mathbf{g}^{\mathcal{T}}) + c_{\mathrm{agg}}(\mathbf{P}^{\mathcal{V}},\mathbf{P}^{\mathcal{T}};\mathbf{g}^{\mathcal{T}})$ . Finally, given the text query, the images in the gallery are ranked according to similarity scores between the images and the text for inference. + +# 4 Experiments + +In this section, we provide a detailed account of our experimental setup (Sec. 4.1), evaluate our method, and compare it with state of the arts on three benchmark datasets for text-based person search (Sec. 4.2). Furthermore, we qualitatively present retrieval results and analyze the effectiveness of the part discovery module and TDPA with visualization results (Sec. 4.3). We also conduct ablation studies on the losses employed in model training, the methodologies of part discovery, and the strategies for part similarity aggregation (Sec. 4.4). + +# 4.1 Experimental Setup + +Datasets: On three benchmark datasets, CUHK-PEDES [11], ICFG-PEDES [4], and RSTPReid [32], we evaluate and compare the performance of our method against previous methods. In CUHK-PEDES collected from five existing person re-identification datasets [6,12,13,28,31], it contains 40,206 images corresponding + +to 13,003 individual IDs, with each image being approximately matched with two annotated text descriptions. We follow the data split of [11] with 34,054 images from 11,003 person IDs and 68,126 text descriptions for training, 3,078 images from 1,000 IDs and 6,158 text descriptions for validation, and 3,074 images from 1,000 IDs and 6,156 text descriptions for testing. The remaining two datasets are collected from MSMT17 [26]. ICFG-PEDES consists of 54,522 image-text pairs from 4,102 individual IDs, which are split into 34,674 and 19,848 for training and testing, respectively. RSTPReid contains 20,505 images of 4,101 individual IDs, with each ID having 5 images and each image associated with the corresponding two annotated text descriptions. We follow the data split of [32] with 18,505 images from 3,701 IDs and 37,010 text descriptions for training, 1,000 images from 200 IDs and 2,000 text descriptions for validation, and 1,000 images from 200 IDs and 2,000 text descriptions for testing, respectively. + +Evaluation Protocol: We employ the standard metric of rank at K (R@K=1,5, 10) for all retrieval experiments. Specifically, given a query text, images are sorted based on their similarity to the query text. The search is considered correct if at least one relevant image appears in the top K positions of the ranking. + +Network Architecture: We adopt the pre-trained CLIP models from OpenAI [18] for both image and text encoders, where the size of the image encoder is ViT-B/16. The input images are resized to $384 \times 128$ . Random horizontal flipping, random cropping, and random erasing are applied for the data augmentation in training time. The maximum text length is set to 77. + +Network Optimization: Our model is trained using the Adam optimizer for 60 epochs with a batch size of 128 for all experiments. For the CLIP encoders, the initial learning rate is set to $5\mathrm{e} - 6$ , using a cosine schedule with the warm-up strategy at the first five epochs; we use a high learning rate for the remaining parameters by scaling 20 times. + +Hyperparameters: We set the number of part slots to 8. The number of iterations in the part slot attention block is set to 5. The temperature parameter $\tau$ is set to 0.015. + +# 4.2 Quantitative Results + +We compare our method with previous text-based person search methods on CUHK-PEDES [11], ICFG-PEDES [4], and RSTPReid [32]. The performance comparison and the backbones for each modality employed by each method are shown in Table 1. Specifically, our method achieves an outstanding $\mathbb{R}@\mathbb{1}$ metric of $75.28\%$ , $65.76\%$ , and $61.80\%$ on the CUHK-PEDES, ICFG-PEDES, and RSTPReid datasets, respectively, outperforming previous methods utilizing different backbones from ours. Moreover, our method improves the previous state of the art, IRRA [7], on $\mathbb{R}@\mathbb{1}$ by a large margin as $1.9\% \mathbb{p}$ , $1.7\% \mathbb{p}$ , and $1.6\% \mathbb{p}$ , respectively. Since IRRA only focuses on aligning the global embeddings of each modality, it is hard to capture the fine-grained differences. In contrast, our method not only takes account of the global embeddings but also aims to discover and align discriminative part embeddings from each modality; it allows the model to effectively find the target person. + +Table 1: Performance of text-based person search methods on the three datasets. Bold and underline denote the best and the second best. + +
MethodsBackboneCUHK-PEDESICFG-PEDESRSTPReid
ImageTextR@1R@5R@10R@1R@5R@10R@1R@5R@10
GNA-RNN [11]RN50LSTM19.05-53.64------
CMPM/C [30]RN50LSTM49.3771.6979.2743.5165.4474.26---
PMA [8]RN50BERT53.8173.5481.23------
TIMAM [19]RN101BERT54.5177.5684.78------
SCAN [9]RN50BERT55.8675.9783.6950.0569.6577.21---
ViTAA [24]RN50LSTM55.9775.8483.5250.9868.7975.78---
NAFS [5]RN50BERT59.9479.8686.70------
DSSL [32]RN50BERT59.9880.4187.56---32.4355.0863.19
MGEL [23]RN50LSTM60.2780.0186.74------
SSAN [4]RN50LSTM61.3780.1586.7354.2372.6379.5343.5067.8077.15
LapsCore [27]RN50BERT63.40-87.80------
SRCF [22]RN50BERT64.0482.9988.8157.1875.0181.49---
LGUR [20]RN50BERT64.2181.9487.9357.4274.9781.45---
TIPCB [2]RN50BERT64.2683.1989.10------
CAIBC [25]RN50BERT64.4382.8788.37---47.3569.5579.00
SAF [10]ViT-B/16BERT64.1382.6288.40------
IVT [21]ViT-B/16BERT65.5983.1189.2156.0473.6080.2246.7070.0078.80
CFine [29]CLIP-ViT-B/16BERT69.5785.9391.1560.8375.5582.4250.5572.5081.60
IRRA [7]CLIP-ViT-B/16CLIP-Xformer73.3889.9393.7163.4680.2485.8260.2081.3088.20
OursCLIP-ViT-B/16CLIP-Xformer75.2890.4294.1265.7681.3986.7361.8082.8589.45
+ +# 4.3 Qualitative Results + +Retrieval Results: Top-5 retrieval results of our method on the CUHK-PEDES dataset are illustrated in Fig. 3. Above all, it shows the overall satisfactory retrieval results. In particular, we can observe our model retrieves targets well, even with distinctive human parts that are small or located in various positions. For instance, the small distinctive human parts like "ponytails" and "high skirts" in Fig. 3(a), "black shoes" in Fig. 3(b), and "blue plaid shorts" in Fig. 3(c), as well as human parts that could appear in various viewpoints such as "yellow shoulder bags" in Fig. 3(a), "floaty dresses" in Fig. 3(b), and "red backpacks" in Fig. 3(d). The CUHK-PEDES dataset typically contains three target images on average in the search space, thereby most of the false matches in the figure are observed by a lack of additional targets. Despite the false matching due to the limitation of the dataset, the retrieval results are reasonable in that the retrieved false matching contains distinctive human parts described by the query description. + +Visualization of Attention Map $\bar{A}_k$ in PSABlock: To demonstrate the effectiveness of our part discovery module, we visualize the attention map $\bar{A}_k$ in $T$ -th iteration of the PSA block for both visual and textual modalities (in Eq. (2)). The visualization results are illustrated in Fig. 4. It not only demonstrates that the part embeddings extracted by our part discovery module capture distinctive human parts but also shows that the part embeddings extracted from the same part slot attend to the semantically identical human parts regardless of modality. For example, the 1st part slot typically focuses on bottom clothes, the 4th on footwear, the 5th on objects being held, the 7th on top clothes, and the 8th on the person's head. Furthermore, in Fig. 4 compared (a) and (b) of visual modality, + +Query: a girl with a ponytail wearing a red top with blue denim thigh high skirt and carrying a yellow shoulder bag looking at a phone. + +![](images/cbfd0667d02d7bd6323ed76a4faf2d80e3e8f2baa6e044985e825293d7c7ee61.jpg) +(a) + +Query: the little boy is wearing his hat backwards. he has on a blue, short sleeved shirt and grey plaid shorts. + +Query: the woman is wearing a white cardigan over a long floaty dress and black shoes. she is carrying a back carry-on bag. + +![](images/67b8f416eb5048dfab7e4fd5e2cfb1ca67048742e3b933f260eeeb49137f2040.jpg) +(b) + +Query: person wearing a dark brown sweater, denim cropped pants that go just below the knees, and white sneakers. they are wearing a red backpack over both shoulders. + +![](images/6e4c87227f1095c68d6de0a8a4b79ed07c5e5422f0256aec1c9ea1762bed1401.jpg) +(c) + +![](images/31787daf2f547edde4d967cfa6876118dc51700a5992bd0461d3cee9ca9e9460.jpg) +(d) +Fig. 3: Top-5 retrieval results of our method on the CUHK-PEDES dataset. Images are sorted from left to right according to their ranks below each text query. Green and red boxes indicate true and false matches, respectively. + +our part discovery module is capable of capturing the distinctive human parts, while it is robust against pose variations and viewpoint changes. + +Visualization of TDPA: We visualize TDPA weights $\mathbf{a}$ in Eq. (6) to demonstrate the effectiveness of similarity aggregation between part embeddings of two modalities with TDPA. The visualization results are presented in the bottom left in Fig. 4(a, b), respectively. Comparing examples (a) and (b), the 5th part slot typically focuses on human parts associated with held objects, like bags, leading to a predicted lower TDPA weight for this part slot in (b) due to the absence of such distinctive information in its query description. However, the presence of the "hat" in the query description of (b) leads to having a high TDPA weight for the 8th part slot that typically focuses on the human head part. In contrast, since this related human head part information is not provided in the query description of (a), the TDPA weight for the 8th part slot in (a) is predicted to be a low value. These observations highlight the capability of TDPA, which adaptively improves part-based retrieval depending on the contents of the text query. + +# 4.4 Ablation Studies + +In our ablation studies conducted on the CUHK-PEDES datasets, we evaluate the effectiveness of proposed components and their combinations in improving text-based person search performance. In Table 2, we compare several configurations: The baseline method employs only global embeddings $\mathbf{g}$ trained by InfoNCE $(\mathcal{L}_{\mathrm{NCE}})$ loss. When incorporating the cross-modal masked language modeling $(\mathcal{L}_{\mathrm{CMLM}})$ , there is a slight improvement across all metrics. The addition + +Text: the woman is wearing a white cardigan over a long floaty dress and black shoes. she is carrying a back carry-on bag. + +![](images/a54dd864b9e7b3c752ebe3dc49e32c81f39056b1f34bd7903632f0d11e86113a.jpg) + +Text: the little boy is wearing his hat backwards. he has on a blue, short sleeved shirt and grey plaid shorts. + +![](images/2cb3e51c334281e25db0501e542484791a9ea7dc23a36fbe744dded371a898f7.jpg) +Fig. 4: Visualization of each modality's attention map $\bar{A}_k$ in $T$ -th iteration of PSA block and TDPA weights a on CUHK-PEDES dataset. + +of part embeddings $\mathbf{P}$ that without $\mathcal{L}_{\mathrm{PartID}}$ also shows significant enhancement in R@1 and R@5, where we observe an increase of $3.46\% \mathrm{p}$ and $2.64\% \mathrm{p}$ , respectively. Notably, the full configuration achieves the best performance with $\mathcal{L}_{\mathrm{PartID}}$ , underscoring the importance of part embeddings for accurate text-based person search. + +Table 2: Ablation studies on the CUHK-PEDES datasets. + +
MethodLossCUHK-PEDES
LNCELIDLCMLMLPartNCELPartIDR@1R@5R@10
Global Only✓XXXX71.39 (+0.0)87.65 (+0.0)92.74 (+0.0)
✓✓XXX71.83 (+0.44)88.06 (+0.41)92.58 (-0.16)
+ CMLM✓✓✓XX72.65 (+1.26)88.58 (+0.93)92.93 (+0.19)
+ Part Embeddings✓✓✓✓X74.85 (+3.46)90.29 (+2.64)94.10 (+1.36)
✓✓✓✓✓75.28 (+3.89)90.42 (+2.77)94.12 (+1.38)
+ +This suggests that our part embeddings significantly contribute to discerning fine-grained details critical for text-based person search. + +Ablation Study on Part Discovery Methods: To validate the effectiveness of our proposed part discovery module, we conduct experiments by replacing our part discovery module with other methods such as TIPCB [2] and PAT [14]. TIPCB is a simple yet effective heuristic part discovery method, which extracts + +Table 3: Ablation study of different part discovery methods on the CUHK-PEDES. + +
MethodsR@1R@5R@10
Ours + TIPCB [2]73.2389.1094.04
Ours + PAT [14]72.7689.2393.42
Ours75.2890.4294.12
+ +image part embeddings through equi-horizontal cropping of the entire image and transforms the global embedding of the text modality into corresponding parts by using additional learnable MLPs for each image part. PAT performs part discovery with a querying transformer approach by leveraging learnable queries and conventional cross-attention. A key difference of our method from PAT is that ours explicitly encourages the discovered parts to be spatially separated since our part slots compete with each other to aggregate input data. In contrast, parts found by PAT, based on conventional cross-attention, often over-capture salient regions and are likely to miss fine details. To demonstrate this limitation, we show the visualization of part attentions with PAT applied to ours in the supplementary materials. As indicated in Table 3, our part discovery method outperforms these heuristic and conventional cross-attention based part discovery methods across all metrics, with the most significant difference observed in the R@1 metric, which requires a precise discernment of human parts. + +# 5 Conclusion + +We proposed a novel framework that extracts distinctive human parts corresponding across visual and textual modalities through part discovery module without part-level supervision. The introduced TDPA further refines the retrieval process by adjusting the importance of each part slot based on the contents of the text query, leading to more precise and relevant retrieval. + +Limitation: Since slots learn to occupy the entire image and text, some of them may indicate irrelevant part of input. TDPA addresses this issue to some extent, but a more explicit solution would further enhance our system. + +# Acknowledgements + +This work was supported by the IITP grants and the NRF grants funded by Ministry of Science and ICT, Korea (RS-2019-II191906; RS-2022-II220926; NRF-2018R1A5A1060031; NRF-2021R1A2C3012728) + +# References + +1. Aggarwal, S., Radhakrishnan, V.B., Chakraborty, A.: Text-based person search via attribute-aided matching. In: Proc. Winter Conference on Applications of Computer Vision (WACV) (2020) +2. Chen, Y., Zhang, G., Lu, Y., Wang, Z., Zheng, Y.: Tipcb: A simple but effective part-based convolutional baseline for text-based person search. Neurocomputing 494 (2022) +3. Devlin, J., Chang, M.W., Lee, K., Toutanova, K.: Bert: Pre-training of deep bidirectional transformers for language understanding. In: Annual Conference of the North American Chapter of the Association for Computational Linguistics (2019) +4. Ding, Z., Ding, C., Shao, Z., Tao, D.: Semantically self-aligned network for text-to-image part-aware person re-identification. arXiv preprint arXiv:2107.12666 (2021) +5. Gao, C., Cai, G., Jiang, X., Zheng, F., Zhang, J., Gong, Y., Peng, P., Guo, X., Sun, X.: Contextual non-local alignment over full-scale representation for text-based person search. arXiv preprint arXiv:2101.03036 (2021) +6. Gray, D., Brennan, S., Tao, H.: Evaluating appearance models for recognition, reacquisition, and tracking. In: Proc. IEEE international workshop on performance evaluation for tracking and surveillance (PETS) (2007) +7. Jiang, D., Ye, M.: Cross-modal implicit relation reasoning and aligning for text-to-image person retrieval. In: IEEE International Conference on Computer Vision and Pattern Recognition (CVPR) (2023) +8. Jing, Y., Si, C., Wang, J., Wang, W., Wang, L., Tan, T.: Pose-guided multi-granularity attention network for text-based person search. In: Proc. AAAI Conference on Artificial Intelligence (AAAI) (2020) +9. Lee, K.H., Chen, X., Hua, G., Hu, H., He, X.: Stacked cross attention for image-text matching. In: Proc. European Conference on Computer Vision (ECCV) (2018) +10. Li, S., Cao, M., Zhang, M.: Learning semantic-aligned feature representation for text-based person search. In: Proc. IEEE International Conference on Acoustics, Speech and Signal Processing (ICASSP). IEEE (2022) +11. Li, S., Xiao, T., Li, H., Zhou, B., Yue, D., Wang, X.: Person search with natural language description. In: Proc. IEEE Conference on Computer Vision and Pattern Recognition (CVPR) (2017) +12. Li, W., Zhao, R., Wang, X.: Human reidentification with transferred metric learning. In: Computer Vision-ACCV 2012: 11th Asian Conference on Computer Vision, Daejeon, Korea, November 5-9, 2012, Revised Selected Papers, Part I 11. Springer (2013) +13. Li, W., Zhao, R., Xiao, T., Wang, X.: Deepreid: Deep filter pairing neural network for person re-identification. In: Proc. IEEE Conference on Computer Vision and Pattern Recognition (CVPR) (2014) +14. Li, Y., He, J., Zhang, T., Liu, X., Zhang, Y., Wu, F.: Diverse part discovery: Occluded person re-identification with part-aware transformer. In: Proc. IEEE Conference on Computer Vision and Pattern Recognition (CVPR) (2021) + +15. Locatello, F., Weissenborn, D., Unterthiner, T., Mahendran, A., Heigold, G., Uszkoreit, J., Dosovitskiy, A., Kipf, T.: Object-centric learning with slot attention (2020) +16. Niu, K., Huang, Y., Ouyang, W., Wang, L.: Improving description-based person re-identification by multi-granularity image-text alignments. IEEE Transactions on Image Processing 29, 5542-5556 (2020) +17. Oord, A.v.d., Li, Y., Vinyals, O.: Representation learning with contrastive predictive coding. arXiv preprint arXiv:1807.03748 (2018) +18. Radford, A., Kim, J.W., Hallacy, C., Ramesh, A., Goh, G., Agarwal, S., Sastry, G., Askell, A., Mishkin, P., Clark, J., et al.: Learning transferable visual models from natural language supervision. In: Proc. International Conference on Machine Learning (ICML) (2021) +19. Sarafianos, N., Xu, X., Kakadiaris, I.A.: Adversarial representation learning for text-to-image matching. In: Proc. IEEE International Conference on Computer Vision (ICCV) (2019) +20. Shao, Z., Zhang, X., Fang, M., Lin, Z., Wang, J., Ding, C.: Learning granularity-unified representations for text-to-image person re-identification. In: Proc. ACM Multimedia Conference (ACMMM) (2022) +21. Shu, X., Wen, W., Wu, H., Chen, K., Song, Y., Qiao, R., Ren, B., Wang, X.: See finer, see more: Implicit modality alignment for text-based person retrieval. In: Proc. European Conference on Computer Vision Workshop on Real-World Surveillance, (ECCVW) (2022) +22. Suo, W., Sun, M., Niu, K., Gao, Y., Wang, P., Zhang, Y., Wu, Q.: A simple and robust correlation filtering method for text-based person search. In: Proc. European Conference on Computer Vision (ECCV) (2022) +23. Wang, C., Luo, Z., Lin, Y., Li, S.: Text-based person search via multi-granularity embedding learning. In: Proc. International Joint Conferences on Artificial Intelligence (IJCAI) (2021) +24. Wang, Z., Fang, Z., Wang, J., Yang, Y.: Vitaa: Visual-textual attributes alignment in person search by natural language. In: Proc. European Conference on Computer Vision (ECCV) (2020) +25. Wang, Z., Zhu, A., Xue, J., Wan, X., Liu, C., Wang, T., Li, Y.: Caibc: Capturing all-round information beyond color for text-based person retrieval. In: Proc. ACM Multimedia Conference (ACMMM) (2022) +26. Wei, L., Zhang, S., Gao, W., Tian, Q.: Person transfer gan to bridge domain gap for person re-identification. In: Proc. IEEE Conference on Computer Vision and Pattern Recognition (CVPR) (2018) +27. Wu, Y., Yan, Z., Han, X., Li, G., Zou, C., Cui, S.: Lipscore: language-guided person search via color reasoning. In: Proc. IEEE International Conference on Computer Vision (ICCV) (2021) +28. Xiao, T., Li, S., Wang, B., Lin, L., Wang, X.: End-to-end deep learning for person search. arXiv preprint arXiv:1604.01850 (2016) +29. Yan, S., Dong, N., Zhang, L., Tang, J.: Clip-driven fine-grained text-image person re-identification. IEEE Transactions on Image Processing (2023) +30. Zhang, Y., Lu, H.: Deep cross-modal projection learning for image-text matching. In: Proc. European Conference on Computer Vision (ECCV) (2018) +31. Zheng, L., Shen, L., Tian, L., Wang, S., Wang, J., Tian, Q.: Scalable person re-identification: A benchmark. In: Proc. IEEE International Conference on Computer Vision (ICCV) (2015) +32. Zhu, A., Wang, Z., Li, Y., Wan, X., Jin, J., Wang, T., Hu, F., Hua, G.: Dssl: deep surroundings-person separation learning for text-based person retrieval. In: + +Proceedings of the 29th ACM International Conference on Multimedia. pp. 209-217 (2021) \ No newline at end of file diff --git a/2024/PLOT_ Text-based Person Search with Part Slot Attention for Corresponding Part Discovery/images.zip b/2024/PLOT_ Text-based Person Search with Part Slot Attention for Corresponding Part Discovery/images.zip new file mode 100644 index 0000000000000000000000000000000000000000..00e199cbca05be96461f753a2c629e589849dba1 --- /dev/null +++ b/2024/PLOT_ Text-based Person Search with Part Slot Attention for Corresponding Part Discovery/images.zip @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:7ec304a318d2b79c437fa997b41813b668aadd0c23b0cc1c313353d46ec425ab +size 547860 diff --git a/2024/PLOT_ Text-based Person Search with Part Slot Attention for Corresponding Part Discovery/layout.json b/2024/PLOT_ Text-based Person Search with Part Slot Attention for Corresponding Part Discovery/layout.json new file mode 100644 index 0000000000000000000000000000000000000000..052dc63c0d8b4687d4e2356370f88ab27a049274 --- /dev/null +++ b/2024/PLOT_ Text-based Person Search with Part Slot Attention for Corresponding Part Discovery/layout.json @@ -0,0 +1,8859 @@ +{ + "pdf_info": [ + { + "para_blocks": [ + { + "bbox": [ + 134, + 111, + 479, + 148 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 134, + 111, + 479, + 148 + ], + "spans": [ + { + "bbox": [ + 134, + 111, + 479, + 148 + ], + "type": "text", + "content": "PLOT: Text-based Person Search with Part Slot Attention for Corresponding Part Discovery" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 150, + 169, + 462, + 181 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 150, + 169, + 462, + 181 + ], + "spans": [ + { + "bbox": [ + 150, + 169, + 462, + 181 + ], + "type": "text", + "content": "Jicheol Park D, Dongwon Kim D, Boseung Jeong D, and Suha Kwak D" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 157, + 190, + 455, + 224 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 157, + 190, + 455, + 224 + ], + "spans": [ + { + "bbox": [ + 157, + 190, + 455, + 224 + ], + "type": "text", + "content": "Pohang University of Science and Technology (POSTECH), South Korea {jicheol, kdwon,boseung01, suha.kwak}@postech.ac.kr https://cvlab.postech.ac.kr/research/PLOT" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 160, + 258, + 455, + 400 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 160, + 258, + 455, + 400 + ], + "spans": [ + { + "bbox": [ + 160, + 258, + 455, + 400 + ], + "type": "text", + "content": "Abstract. Text-based person search, employing free-form text queries to identify individuals within a vast image collection, presents a unique challenge in aligning visual and textual representations, particularly at the human part level. Existing methods often struggle with part feature extraction and alignment due to the lack of direct part-level supervision and reliance on heuristic features. We propose a novel framework that leverages a part discovery module based on slot attention to autonomously identify and align distinctive parts across modalities, enhancing interpretability and retrieval accuracy without explicit part-level correspondence supervision. Additionally, text-based dynamic part attention adjusts the importance of each part, further improving retrieval outcomes. Our method is evaluated on three public benchmarks, significantly outperforming existing methods." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 160, + 411, + 427, + 423 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 160, + 411, + 427, + 423 + ], + "spans": [ + { + "bbox": [ + 160, + 411, + 427, + 423 + ], + "type": "text", + "content": "Keywords: Text-Based Person Search " + }, + { + "bbox": [ + 160, + 411, + 427, + 423 + ], + "type": "inline_equation", + "content": "\\cdot" + }, + { + "bbox": [ + 160, + 411, + 427, + 423 + ], + "type": "text", + "content": " Multi-Modal Retrieval" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 132, + 445, + 230, + 458 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 445, + 230, + 458 + ], + "spans": [ + { + "bbox": [ + 132, + 445, + 230, + 458 + ], + "type": "text", + "content": "1 Introduction" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 130, + 473, + 482, + 569 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 473, + 482, + 569 + ], + "spans": [ + { + "bbox": [ + 130, + 473, + 482, + 569 + ], + "type": "text", + "content": "Text-based person search is the task of identifying the target person from the vast collection of images with a free-form text query. This task demands extracting identifiable features, such as human parts, from both textual and visual modalities to capture subtle differences between individuals. Hence, establishing correspondence between the extracted human part features across image and text modalities is essential for accurate text-based person search. However, it is not straightforward to extract these part features and establish their correspondences between the two modalities without part-level supervision." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 130, + 570, + 482, + 666 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 570, + 482, + 666 + ], + "spans": [ + { + "bbox": [ + 130, + 570, + 482, + 666 + ], + "type": "text", + "content": "To address this challenge, previous work [2, 22] relies on the heuristic part features obtained by equi-horizontally cropping the entire image; such features are then matched to the free-form text queries for person search. However, the heuristic part features used in this approach are susceptible to deformation caused by occlusion and pose variation. Meanwhile, earlier work [8, 20, 24] proposed a learning-based approach for part feature extraction. Nevertheless, these methods tend to generate redundant part features that lack disentanglement [20], or they demand access to additional part-level supervisions [8, 24]." + } + ] + } + ], + "index": 7 + } + ], + "discarded_blocks": [], + "page_size": [ + 612, + 792 + ], + "page_idx": 0 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 134, + 117, + 482, + 288 + ], + "blocks": [ + { + "bbox": [ + 134, + 117, + 482, + 288 + ], + "lines": [ + { + "bbox": [ + 134, + 117, + 482, + 288 + ], + "spans": [ + { + "bbox": [ + 134, + 117, + 482, + 288 + ], + "type": "image", + "image_path": "8a390316999a60ab9111737922f170dec25ac14a89cbf5747f3a9421ce3c96f6.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 220, + 298, + 392, + 309 + ], + "lines": [ + { + "bbox": [ + 220, + 298, + 392, + 309 + ], + "spans": [ + { + "bbox": [ + 220, + 298, + 392, + 309 + ], + "type": "text", + "content": "Fig. 1: The overall architecture of PLOT." + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_caption" + } + ], + "index": 2 + }, + { + "bbox": [ + 130, + 324, + 482, + 492 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 324, + 482, + 492 + ], + "spans": [ + { + "bbox": [ + 130, + 324, + 482, + 492 + ], + "type": "text", + "content": "To tackle the above issues, we introduce a new framework that discovers distinctive parts in both modalities and matches them between the two modalities without any correspondence supervision; its overall architecture is illustrated in Fig. 1. To discover distinctive parts and extract their features from both modalities, we propose PLOT, a Part discovery module based on the sLOT attention mechanism [15]. Slot attention is an attention mechanism designed for object-centric learning, which segments input data into a set of slots representing individual entities without requiring object-level supervision. In PLOT, we first define a set of learnable embedding vectors, termed part slots, that contain primitive information related to human body parts shared between the two modalities. Then, these part slots undergo refinement through several iterative attention processes, where they compete amongst themselves to bind with the input data; ultimately, the part slots are transformed into part features, termed part embeddings, that represent distinctive parts in the input data." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 130, + 495, + 482, + 615 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 495, + 482, + 615 + ], + "spans": [ + { + "bbox": [ + 130, + 495, + 482, + 615 + ], + "type": "text", + "content": "To ensure correspondence between part embeddings from different modalities, the part slots are shared between the visual and textual modalities. Part embeddings from the same part slots are then learned to represent the identical part although they are computed from different modalities. This mechanism enables PLOT to match the discovered parts from the two modalities without supervision for the correspondence as well as capturing part-level fine-grained appearance features from the both modalities. Hence, PLOT improves performance of text-based person search through the rich and fine-grained part features, and at the same time, it guarantees interpretable retrieval by providing part-level correspondences between query text and retrieved images." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 130, + 617, + 483, + 666 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 617, + 483, + 666 + ], + "spans": [ + { + "bbox": [ + 130, + 617, + 483, + 666 + ], + "type": "text", + "content": "In addition, PLOT introduces a new similarity aggregation method for part embeddings called text-based dynamic part attention (TDPA) pooling, which dynamically adjusts the weights of part embeddings based on the text query. Given a text query as input, TDPA predicts the importance weight of each" + } + ] + } + ], + "index": 6 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "spans": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "type": "text", + "content": "2" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 166, + 91, + 220, + 100 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 166, + 91, + 220, + 100 + ], + "spans": [ + { + "bbox": [ + 166, + 91, + 220, + 100 + ], + "type": "text", + "content": "J. Park et al." + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 1 + }, + { + "para_blocks": [ + { + "bbox": [ + 130, + 116, + 480, + 163 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 116, + 480, + 163 + ], + "spans": [ + { + "bbox": [ + 130, + 116, + 480, + 163 + ], + "type": "text", + "content": "slot for retrieval and applies the predicted importance weights to aggregate the similarities of part embeddings between the two modalities. TDPA allows the entire retrieval system to perform optimized retrieval for each query, leading to improved performance." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 130, + 164, + 481, + 211 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 164, + 481, + 211 + ], + "spans": [ + { + "bbox": [ + 130, + 164, + 481, + 211 + ], + "type": "text", + "content": "Our method was evaluated and compared with prior work on three public benchmarks [4, 11, 32], where it clearly outperformed all existing methods thanks to the rich representation based on part embeddings. The main contribution of our work is four-fold:" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 137, + 220, + 480, + 349 + ], + "type": "list", + "angle": 0, + "index": 8, + "blocks": [ + { + "bbox": [ + 138, + 220, + 479, + 265 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 220, + 479, + 265 + ], + "spans": [ + { + "bbox": [ + 138, + 220, + 479, + 265 + ], + "type": "text", + "content": "- We introduce PLOT, a new framework for text-based person search that discovers distinguished human body parts, extracts their embeddings, and establishes their correspondences between the two modalities with no human intervention." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 137, + 267, + 479, + 289 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 137, + 267, + 479, + 289 + ], + "spans": [ + { + "bbox": [ + 137, + 267, + 479, + 289 + ], + "type": "text", + "content": "- The part embeddings provided by PLOT enables an interpretable text-based person search thanks to the part-level correspondences it provides." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 137, + 292, + 480, + 325 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 137, + 292, + 480, + 325 + ], + "spans": [ + { + "bbox": [ + 137, + 292, + 480, + 325 + ], + "type": "text", + "content": "- We introduce a novel similarity aggregation method that adaptively determines importance of each discovered part based on each text query and consequently enables retrieval optimized per query." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 137, + 327, + 479, + 349 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 137, + 327, + 479, + 349 + ], + "spans": [ + { + "bbox": [ + 137, + 327, + 479, + 349 + ], + "type": "text", + "content": "- Our model with PLOT achieved the best on all the public benchmarks for text-based person search." + } + ] + } + ], + "index": 7 + } + ], + "sub_type": "text" + }, + { + "bbox": [ + 132, + 369, + 237, + 381 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 369, + 237, + 381 + ], + "spans": [ + { + "bbox": [ + 132, + 369, + 237, + 381 + ], + "type": "text", + "content": "2 Related Work" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 131, + 394, + 292, + 405 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 131, + 394, + 292, + 405 + ], + "spans": [ + { + "bbox": [ + 131, + 394, + 292, + 405 + ], + "type": "text", + "content": "2.1 Text-Based Person Search" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 130, + 414, + 482, + 533 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 414, + 482, + 533 + ], + "spans": [ + { + "bbox": [ + 130, + 414, + 482, + 533 + ], + "type": "text", + "content": "In recent years, the task of text-based person search has gained significant attention in the computer vision community. Li et al., [11] proposed a gated neural attention-based recurrent neural network (GNA-RNN) for learning the affinity between text descriptions and images, along with providing a benchmark dataset CUHK-PEDES for model evaluation. Zhang et al., [30] proposed cross-modal projection matching and classification (CMPM+CMPC) loss, for learning deep discriminative image-text global embeddings. However, these methods primarily focus on global representations of input data and are thus not capable of capturing distinctive human part details, leading to limited performance in the text-based person search." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 130, + 534, + 482, + 666 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 534, + 482, + 666 + ], + "spans": [ + { + "bbox": [ + 130, + 534, + 482, + 666 + ], + "type": "text", + "content": "To address the above problem, a line of studies focuses on extracting fine-grained representations. One of the prominent examples of exploiting fine-grained information is to cut human images horizontally and use them as human parts [2, 4, 5, 16, 22, 23]. Chen et al., [2] extracts image part embeddings through equi-horizontal cropping of the entire image and aligns additional network to transform the textual global into corresponding parts. However, these heuristic part features inevitably include no-informative information, such as background elements. To avoid the above limitations, Suo et al., [22] proposed the simple and robust correlation filtering method to extract foreground features on the heuristic part features. Yet, the heuristically divided part features fundamentally fall short of capturing the complex human parts, underscoring a critical limitation in their" + } + ] + } + ], + "index": 12 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 148, + 91, + 447, + 102 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 148, + 91, + 447, + 102 + ], + "spans": [ + { + "bbox": [ + 148, + 91, + 447, + 102 + ], + "type": "text", + "content": "PLOT: TBPS with Part Slot Attention for Corresponding Part Discovery" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 474, + 92, + 480, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 474, + 92, + 480, + 100 + ], + "spans": [ + { + "bbox": [ + 474, + 92, + 480, + 100 + ], + "type": "text", + "content": "3" + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 2 + }, + { + "para_blocks": [ + { + "bbox": [ + 130, + 116, + 482, + 258 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 116, + 482, + 258 + ], + "spans": [ + { + "bbox": [ + 130, + 116, + 482, + 258 + ], + "type": "text", + "content": "expressive capability. To move beyond such heuristic part structures, Shao et al., [2] proposes a learning-based approach for part feature extraction. However, this method struggles to extract distinctive part embeddings due to the extraction of redundant part information, which lacks sufficient disentanglement. To extract exquisite human parts, several studies tried to utilize useful information (e.g., human attributes and human keypoints) via external tools [1,8,24]. Wang et al., [24] introduced an auxiliary attribute segmentation to align the visual part features with the textual attributes parsed from text description. Jing et al., [8] proposed a new multi-granularity attention network to learn the part feature alignment between visual and textual with human pose estimation. However, these approaches have inevitable limitations of high computational cost and dependence on the performance of external tools for local feature extraction." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 130, + 260, + 482, + 344 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 260, + 482, + 344 + ], + "spans": [ + { + "bbox": [ + 130, + 260, + 482, + 344 + ], + "type": "text", + "content": "Most recent work, Jiang et al., [7] utilizes pre-trained CLIP [18] model, capitalizing on the rich knowledge of models trained on extensive data for text-to-image matching, to excel in text-based person search. Yet, this methods primarily focus on global features, not specifically designed for extracting human part features. Not only do we leverage pre-trained knowledge, but we transcend heuristic human parts methods to extract sophisticated human parts without external tools or part-level supervision." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 132, + 365, + 233, + 376 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 365, + 233, + 376 + ], + "spans": [ + { + "bbox": [ + 132, + 365, + 233, + 376 + ], + "type": "text", + "content": "2.2 Slot Attention" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 130, + 388, + 482, + 543 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 388, + 482, + 543 + ], + "spans": [ + { + "bbox": [ + 130, + 388, + 482, + 543 + ], + "type": "text", + "content": "Slot attention [15] is a recently proposed attention mechanism for object-centric learning, a problem focusing on discovering constituent visual entities within an image. The unique property of slot attention is that it can represent input images as a set of slots, where the slots are representations corresponding to individual visual entities, without any object-level supervision during training. Within slot attention, slots iteratively compete for aggregating input data, ensuring distinct representations focusing on individual visual entities. By incorporating slot attention, the proposed framework facilitates the unsupervised identification of a structure underlying image and text queries, enabling the model to discern and represent individual human parts without explicit supervision. This capability is particularly valuable in person search datasets, where recognizing subtle differences and understanding the correspondence of human parts across different modalities are crucial." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 132, + 565, + 261, + 578 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 565, + 261, + 578 + ], + "spans": [ + { + "bbox": [ + 132, + 565, + 261, + 578 + ], + "type": "text", + "content": "3 Proposed Method" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 130, + 594, + 482, + 666 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 594, + 482, + 666 + ], + "spans": [ + { + "bbox": [ + 130, + 594, + 482, + 666 + ], + "type": "text", + "content": "The following subsection offers details of global and part embeddings extraction for each modality with its backbone (Sec 3.1). Subsequently, we present our novel framework, PLOT, which includes the part discovery module (Sec. 3.2) and similarity aggregation between part embeddings (Sec. 3.3), concluding with a discussion on the learning objective designed to optimize our proposed framework (Sec. 3.4) and inference of our framework (Sec. 3.5)." + } + ] + } + ], + "index": 7 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "spans": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "type": "text", + "content": "4" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 166, + 91, + 220, + 100 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 166, + 91, + 220, + 100 + ], + "spans": [ + { + "bbox": [ + 166, + 91, + 220, + 100 + ], + "type": "text", + "content": "J. Park et al." + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 3 + }, + { + "para_blocks": [ + { + "bbox": [ + 132, + 116, + 307, + 128 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 116, + 307, + 128 + ], + "spans": [ + { + "bbox": [ + 132, + 116, + 307, + 128 + ], + "type": "text", + "content": "3.1 Global and Part Embeddings" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 130, + 133, + 482, + 456 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 133, + 482, + 456 + ], + "spans": [ + { + "bbox": [ + 130, + 133, + 482, + 456 + ], + "type": "text", + "content": "In our framework, a single input data is described by two different types of representations: a global embedding and multiple part embeddings. The global embedding is used to represent the input data holistically, while each part embedding describes the appearance of distinctive human parts (e.g., arm, leg, torso, etc.). We below provide details of global and part embeddings computation for visual and textual modalities. Following the previous work [7], a pre-trained CLIP [18] is used as the backbone networks for the visual and textual modalities. Visual Modality: We utilize the vision transformer (ViT) from the CLIP-B/16 [18] architecture as a visual backbone network. Initially, an input image of a person is split into " + }, + { + "bbox": [ + 130, + 133, + 482, + 456 + ], + "type": "inline_equation", + "content": "N" + }, + { + "bbox": [ + 130, + 133, + 482, + 456 + ], + "type": "text", + "content": " distinct, non-overlapping patches, which are subsequently transformed into patch tokens through linear projection. The patch tokens and an extra [c1s] token are then fed into the visual backbone network. The token sequence is processed throughout multiple self-attention blocks, and the [c1s] token of the last block is used as a global embedding " + }, + { + "bbox": [ + 130, + 133, + 482, + 456 + ], + "type": "inline_equation", + "content": "\\mathbf{g}^{\\mathcal{V}} \\in \\mathbb{R}^{D}" + }, + { + "bbox": [ + 130, + 133, + 482, + 456 + ], + "type": "text", + "content": ". To obtain part embeddings " + }, + { + "bbox": [ + 130, + 133, + 482, + 456 + ], + "type": "inline_equation", + "content": "\\mathbf{P}^{\\mathcal{V}} \\in \\mathbb{R}^{K \\times D}" + }, + { + "bbox": [ + 130, + 133, + 482, + 456 + ], + "type": "text", + "content": ", remaining patch tokens of last block " + }, + { + "bbox": [ + 130, + 133, + 482, + 456 + ], + "type": "inline_equation", + "content": "\\mathbf{x}^{\\mathcal{V}} \\in \\mathbb{R}^{N \\times D}" + }, + { + "bbox": [ + 130, + 133, + 482, + 456 + ], + "type": "text", + "content": " is passed to the part discovery module. The part discovery module aggregates the patch features describing coherent human parts into the same part embedding. We provide more detailed information about the part discovery module in Sec. 3.2. Textual Modality: For the textual backbone network, we utilize a transformer architecture from the CLIP-Xformer [18] text encoder. This encoder operates on text input transformed into byte pair encoding (BPE) sequences. Initially, the text query undergoes tokenization via BPE, followed by the enclosing with [S0S] and [E0S] tokens. The resulting sequence of tokens is then inputted into the textual backbone network. Here, the [E0S] token from the final block serves as the global embedding, denoted as " + }, + { + "bbox": [ + 130, + 133, + 482, + 456 + ], + "type": "inline_equation", + "content": "\\mathbf{g}^{\\mathcal{T}} \\in \\mathbb{R}^{D}" + }, + { + "bbox": [ + 130, + 133, + 482, + 456 + ], + "type": "text", + "content": ". Analogous to the approach for the visual modality, we process the remainder of the text tokens, " + }, + { + "bbox": [ + 130, + 133, + 482, + 456 + ], + "type": "inline_equation", + "content": "\\mathbf{x}^{\\mathcal{T}} \\in \\mathbb{R}^{L \\times D}" + }, + { + "bbox": [ + 130, + 133, + 482, + 456 + ], + "type": "text", + "content": ", into part embeddings, " + }, + { + "bbox": [ + 130, + 133, + 482, + 456 + ], + "type": "inline_equation", + "content": "\\mathbf{P}^{\\mathcal{T}} \\in \\mathbb{R}^{K \\times D}" + }, + { + "bbox": [ + 130, + 133, + 482, + 456 + ], + "type": "text", + "content": ", utilizing the part discovery module." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 132, + 471, + 278, + 483 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 471, + 278, + 483 + ], + "spans": [ + { + "bbox": [ + 132, + 471, + 278, + 483 + ], + "type": "text", + "content": "3.2 Part Discovery Module" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 130, + 488, + 481, + 559 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 488, + 481, + 559 + ], + "spans": [ + { + "bbox": [ + 130, + 488, + 481, + 559 + ], + "type": "text", + "content": "For extracting part embeddings in each modality, part discovery module aggregates patch tokens " + }, + { + "bbox": [ + 130, + 488, + 481, + 559 + ], + "type": "inline_equation", + "content": "\\mathbf{x}^{\\mathcal{V}}" + }, + { + "bbox": [ + 130, + 488, + 481, + 559 + ], + "type": "text", + "content": " and text tokens " + }, + { + "bbox": [ + 130, + 488, + 481, + 559 + ], + "type": "inline_equation", + "content": "\\mathbf{x}^{\\mathcal{T}}" + }, + { + "bbox": [ + 130, + 488, + 481, + 559 + ], + "type": "text", + "content": " into the visual part embeddings " + }, + { + "bbox": [ + 130, + 488, + 481, + 559 + ], + "type": "inline_equation", + "content": "\\mathbf{P}^{\\mathcal{V}}" + }, + { + "bbox": [ + 130, + 488, + 481, + 559 + ], + "type": "text", + "content": " and textual part embeddings " + }, + { + "bbox": [ + 130, + 488, + 481, + 559 + ], + "type": "inline_equation", + "content": "\\mathbf{P}^{\\mathcal{T}}" + }, + { + "bbox": [ + 130, + 488, + 481, + 559 + ], + "type": "text", + "content": ", respectively. It is worth noting that the part discovery module for each modality has identical model architecture and functions equivalently. Therefore, we will only explain the part discovery module on the visual modality for brevity." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 130, + 559, + 482, + 643 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 559, + 482, + 643 + ], + "spans": [ + { + "bbox": [ + 130, + 559, + 482, + 643 + ], + "type": "text", + "content": "Part discovery module consists of initial part slots " + }, + { + "bbox": [ + 130, + 559, + 482, + 643 + ], + "type": "inline_equation", + "content": "\\mathbf{S}^0\\in \\mathbb{R}^{K\\times D}" + }, + { + "bbox": [ + 130, + 559, + 482, + 643 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 130, + 559, + 482, + 643 + ], + "type": "inline_equation", + "content": "T" + }, + { + "bbox": [ + 130, + 559, + 482, + 643 + ], + "type": "text", + "content": " multiple iteration of the part slot attention block (PSA block). To extract part embeddings, we first initialize a set of learnable embeddings part slots " + }, + { + "bbox": [ + 130, + 559, + 482, + 643 + ], + "type": "inline_equation", + "content": "\\mathbf{S}^0\\in \\mathbb{R}^{K\\times D}" + }, + { + "bbox": [ + 130, + 559, + 482, + 643 + ], + "type": "text", + "content": ", where " + }, + { + "bbox": [ + 130, + 559, + 482, + 643 + ], + "type": "inline_equation", + "content": "K" + }, + { + "bbox": [ + 130, + 559, + 482, + 643 + ], + "type": "text", + "content": " indicates the number of part slots. Then, through a series of " + }, + { + "bbox": [ + 130, + 559, + 482, + 643 + ], + "type": "inline_equation", + "content": "T" + }, + { + "bbox": [ + 130, + 559, + 482, + 643 + ], + "type": "text", + "content": " iterations of our PSA block, the initial " + }, + { + "bbox": [ + 130, + 559, + 482, + 643 + ], + "type": "inline_equation", + "content": "\\mathbf{S}^0" + }, + { + "bbox": [ + 130, + 559, + 482, + 643 + ], + "type": "text", + "content": " evolves into refined " + }, + { + "bbox": [ + 130, + 559, + 482, + 643 + ], + "type": "inline_equation", + "content": "\\mathbf{S}^T" + }, + { + "bbox": [ + 130, + 559, + 482, + 643 + ], + "type": "text", + "content": ", where each slot captures distinct parts within input data. The refined part slots " + }, + { + "bbox": [ + 130, + 559, + 482, + 643 + ], + "type": "inline_equation", + "content": "\\mathbf{S}^T" + }, + { + "bbox": [ + 130, + 559, + 482, + 643 + ], + "type": "text", + "content": " are used as visual part embeddings " + }, + { + "bbox": [ + 130, + 559, + 482, + 643 + ], + "type": "inline_equation", + "content": "\\mathbf{P}^{\\mathcal{V}}" + }, + { + "bbox": [ + 130, + 559, + 482, + 643 + ], + "type": "text", + "content": ". The part discovery module can be formulated as follows:" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 207, + 648, + 481, + 662 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 207, + 648, + 481, + 662 + ], + "spans": [ + { + "bbox": [ + 207, + 648, + 481, + 662 + ], + "type": "interline_equation", + "content": "\\mathbf {P} ^ {\\mathcal {V}} := \\mathbf {S} ^ {T}, \\text {w h e r e} \\quad \\mathbf {S} ^ {t} = \\operatorname {P S A} _ {-} \\operatorname {B l o c k} ^ {\\mathcal {V}} \\left(\\mathbf {x} ^ {\\mathcal {V}}; \\mathbf {S} ^ {t - 1}\\right). \\tag {1}", + "image_path": "672145e6a655765892c2bf0967aa8acde326cee9654f849b0d73e0cf63cc9638.jpg" + } + ] + } + ], + "index": 7 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 148, + 91, + 447, + 102 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 148, + 91, + 447, + 102 + ], + "spans": [ + { + "bbox": [ + 148, + 91, + 447, + 102 + ], + "type": "text", + "content": "PLOT: TBPS with Part Slot Attention for Corresponding Part Discovery" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 474, + 92, + 480, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 474, + 92, + 480, + 100 + ], + "spans": [ + { + "bbox": [ + 474, + 92, + 480, + 100 + ], + "type": "text", + "content": "5" + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 4 + }, + { + "para_blocks": [ + { + "bbox": [ + 130, + 115, + 482, + 163 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 115, + 482, + 163 + ], + "spans": [ + { + "bbox": [ + 130, + 115, + 482, + 163 + ], + "type": "text", + "content": "The PSA block first transforms the inputs " + }, + { + "bbox": [ + 130, + 115, + 482, + 163 + ], + "type": "inline_equation", + "content": "\\mathbf{S}^{t - 1}" + }, + { + "bbox": [ + 130, + 115, + 482, + 163 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 130, + 115, + 482, + 163 + ], + "type": "inline_equation", + "content": "\\mathbf{x}^{\\mathcal{V}}" + }, + { + "bbox": [ + 130, + 115, + 482, + 163 + ], + "type": "text", + "content": " with layer normalization and linear projection layers " + }, + { + "bbox": [ + 130, + 115, + 482, + 163 + ], + "type": "inline_equation", + "content": "q(\\cdot), k(\\cdot)" + }, + { + "bbox": [ + 130, + 115, + 482, + 163 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 130, + 115, + 482, + 163 + ], + "type": "inline_equation", + "content": "\\nu (\\cdot)" + }, + { + "bbox": [ + 130, + 115, + 482, + 163 + ], + "type": "text", + "content": " to obtain embeddings of " + }, + { + "bbox": [ + 130, + 115, + 482, + 163 + ], + "type": "inline_equation", + "content": "D_{h}" + }, + { + "bbox": [ + 130, + 115, + 482, + 163 + ], + "type": "text", + "content": " dimension. Then the attention map " + }, + { + "bbox": [ + 130, + 115, + 482, + 163 + ], + "type": "inline_equation", + "content": "A\\in \\mathbb{R}^{N\\times K}" + }, + { + "bbox": [ + 130, + 115, + 482, + 163 + ], + "type": "text", + "content": " between " + }, + { + "bbox": [ + 130, + 115, + 482, + 163 + ], + "type": "inline_equation", + "content": "\\mathbf{S}^{t - 1}" + }, + { + "bbox": [ + 130, + 115, + 482, + 163 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 130, + 115, + 482, + 163 + ], + "type": "inline_equation", + "content": "\\mathbf{x}^{\\mathcal{V}}" + }, + { + "bbox": [ + 130, + 115, + 482, + 163 + ], + "type": "text", + "content": " is computed by" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 205, + 163, + 482, + 192 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 205, + 163, + 482, + 192 + ], + "spans": [ + { + "bbox": [ + 205, + 163, + 482, + 192 + ], + "type": "interline_equation", + "content": "A _ {n, k} = \\frac {e ^ {M _ {n , k}}}{\\sum_ {i = 1} ^ {K} e ^ {M _ {n , i}}}, \\mathrm {w h e r e} M = \\frac {k (\\mathbf {x} ^ {\\mathcal {V}}) q (\\mathbf {S} ^ {t - 1}) ^ {\\top}}{\\sqrt {D _ {h}}}. (2)", + "image_path": "b114fa15414949553cbb353aa2fe9149fd1d9aa99e9858ff1e6d08419e355008.jpg" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 130, + 194, + 482, + 255 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 194, + 482, + 255 + ], + "spans": [ + { + "bbox": [ + 130, + 194, + 482, + 255 + ], + "type": "text", + "content": "The attention map is obtained through normalization across part slots; this normalization encourages competition among themselves to bind distinct sets of patch tokens to each slot. We update part slots with weighted mean of patch tokens obtained by attention map " + }, + { + "bbox": [ + 130, + 194, + 482, + 255 + ], + "type": "inline_equation", + "content": "A_{n,k}" + }, + { + "bbox": [ + 130, + 194, + 482, + 255 + ], + "type": "text", + "content": " and then feed it to a gated recurrent unit (GRU) using " + }, + { + "bbox": [ + 130, + 194, + 482, + 255 + ], + "type": "inline_equation", + "content": "\\mathbf{S}^{t-1}" + }, + { + "bbox": [ + 130, + 194, + 482, + 255 + ], + "type": "text", + "content": " as hidden state as follows:" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 213, + 263, + 482, + 291 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 213, + 263, + 482, + 291 + ], + "spans": [ + { + "bbox": [ + 213, + 263, + 482, + 291 + ], + "type": "interline_equation", + "content": "\\bar {A} _ {n, k} = \\frac {A _ {n , k}}{\\sum_ {i = 1} ^ {N} A _ {i , k}}, \\bar {\\mathbf {S}} ^ {t} = \\mathbb {G R U} \\left(\\mathbf {S} ^ {t - 1}, \\bar {A} ^ {\\top} v \\left(\\mathbf {x} ^ {\\mathcal {V}}\\right)\\right). \\tag {3}", + "image_path": "0bb68cba01dbf1b60349a4a262573bcc4bbae10f4e14da212bf832cc326a1324.jpg" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 130, + 298, + 481, + 323 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 298, + 481, + 323 + ], + "spans": [ + { + "bbox": [ + 130, + 298, + 481, + 323 + ], + "type": "text", + "content": "Then, we obtain " + }, + { + "bbox": [ + 130, + 298, + 481, + 323 + ], + "type": "inline_equation", + "content": "t" + }, + { + "bbox": [ + 130, + 298, + 481, + 323 + ], + "type": "text", + "content": "-th part slot by feeding " + }, + { + "bbox": [ + 130, + 298, + 481, + 323 + ], + "type": "inline_equation", + "content": "\\overline{\\mathbf{S}}^{t-1}" + }, + { + "bbox": [ + 130, + 298, + 481, + 323 + ], + "type": "text", + "content": " into a multi-layer perceptron (MLP) with layer normalization, ReLU activation, and residual connection:" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 204, + 331, + 482, + 345 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 204, + 331, + 482, + 345 + ], + "spans": [ + { + "bbox": [ + 204, + 331, + 482, + 345 + ], + "type": "interline_equation", + "content": "\\mathbf {S} ^ {t} = \\operatorname {P S A} _ {\\text {B l o c k}} ^ {\\mathcal {V}} (\\mathbf {x} ^ {\\mathcal {V}}; \\mathbf {S} ^ {t - 1}) = \\operatorname {M L P} \\left(\\bar {\\mathbf {S}} ^ {t - 1}\\right) + \\bar {\\mathbf {S}} ^ {t - 1}. \\tag {4}", + "image_path": "c5cba7bda8bb7ae7df875beef5a3c3f9deb295e4a9c42bfc7287d0879664986a.jpg" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 130, + 351, + 481, + 378 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 351, + 481, + 378 + ], + "spans": [ + { + "bbox": [ + 130, + 351, + 481, + 378 + ], + "type": "text", + "content": "Finally, we can obtain the visual part embeddings " + }, + { + "bbox": [ + 130, + 351, + 481, + 378 + ], + "type": "inline_equation", + "content": "\\mathbf{P}^{\\mathcal{V}} = \\{\\mathbf{p}_k^{\\mathcal{V}}\\}_{k=1}^K" + }, + { + "bbox": [ + 130, + 351, + 481, + 378 + ], + "type": "text", + "content": " which is the output of " + }, + { + "bbox": [ + 130, + 351, + 481, + 378 + ], + "type": "inline_equation", + "content": "T" + }, + { + "bbox": [ + 130, + 351, + 481, + 378 + ], + "type": "text", + "content": "-th iteration of PSA block: " + }, + { + "bbox": [ + 130, + 351, + 481, + 378 + ], + "type": "inline_equation", + "content": "\\mathbf{P}^{\\mathcal{V}} = \\mathbf{S}_{\\mathcal{V}}^T \\in \\mathbb{R}^{K \\times D}" + }, + { + "bbox": [ + 130, + 351, + 481, + 378 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 130, + 378, + 482, + 438 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 378, + 482, + 438 + ], + "spans": [ + { + "bbox": [ + 130, + 378, + 482, + 438 + ], + "type": "text", + "content": "Part Correspondence through Slot Sharing: Additionally, we share the learnable part slots " + }, + { + "bbox": [ + 130, + 378, + 482, + 438 + ], + "type": "inline_equation", + "content": "\\mathbf{S}^0" + }, + { + "bbox": [ + 130, + 378, + 482, + 438 + ], + "type": "text", + "content": " between two part discovery modules to establish correspondences between part embeddings extracted from each modality, considering that part embeddings extracted from the same part slot are corresponding part across modalities, thereby contributing to a clearer comparison between modalities." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 131, + 455, + 376, + 467 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 131, + 455, + 376, + 467 + ], + "spans": [ + { + "bbox": [ + 131, + 455, + 376, + 467 + ], + "type": "text", + "content": "3.3 Measuring Similarity between Embeddings" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 130, + 474, + 481, + 533 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 474, + 481, + 533 + ], + "spans": [ + { + "bbox": [ + 130, + 474, + 481, + 533 + ], + "type": "text", + "content": "In the context of training and applying our retrieval model, selecting an appropriate similarity function between embeddings is crucial. The challenge lies in dealing with two distinct types of embeddings: global embeddings and part embeddings. For global embeddings, cosine similarity offers a straightforward and effective means of measuring similarity." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 130, + 534, + 482, + 666 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 534, + 482, + 666 + ], + "spans": [ + { + "bbox": [ + 130, + 534, + 482, + 666 + ], + "type": "text", + "content": "However, the situation becomes more complex when considering part embeddings. A direct method for addressing this complexity involves calculating the average cosine similarity across all pairs of part embeddings. This approach, while straightforward, has its drawbacks, primarily because it treats all part-wise similarities as equally significant. In reality, the relevance of specific part embeddings to the actual similarity between data instances can significantly vary, influenced by the context of a text query. For instance, if a text query focuses exclusively on particular features of a human figure, the similarity contributions from other unrelated part embeddings should be less relevant. This challenge highlights the necessity for an approach that can dynamically assess and prioritize the relevance of part embeddings based on the context provided by the query." + } + ] + } + ], + "index": 12 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "spans": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "type": "text", + "content": "6" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 166, + 91, + 220, + 101 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 166, + 91, + 220, + 101 + ], + "spans": [ + { + "bbox": [ + 166, + 91, + 220, + 101 + ], + "type": "text", + "content": "J. Park et al." + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 5 + }, + { + "para_blocks": [ + { + "bbox": [ + 130, + 116, + 483, + 175 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 116, + 483, + 175 + ], + "spans": [ + { + "bbox": [ + 130, + 116, + 483, + 175 + ], + "type": "text", + "content": "To address this challenge, we introduce text-based dynamic part attention (TDPA) to aggregate the similarities between part embeddings. Firstly, textual global embedding " + }, + { + "bbox": [ + 130, + 116, + 483, + 175 + ], + "type": "inline_equation", + "content": "\\mathbf{g}^{\\mathcal{T}}" + }, + { + "bbox": [ + 130, + 116, + 483, + 175 + ], + "type": "text", + "content": " is transformed to the TDPA " + }, + { + "bbox": [ + 130, + 116, + 483, + 175 + ], + "type": "inline_equation", + "content": "\\mathbf{a} \\in \\mathbb{R}^{K}" + }, + { + "bbox": [ + 130, + 116, + 483, + 175 + ], + "type": "text", + "content": ", using MLP and a softmax function. Using TDPA, the aggregated similarity between part embeddings are computed as follows:" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 223, + 182, + 482, + 196 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 223, + 182, + 482, + 196 + ], + "spans": [ + { + "bbox": [ + 223, + 182, + 482, + 196 + ], + "type": "interline_equation", + "content": "\\mathbf {a} = \\sigma (\\operatorname {M L P} \\left(\\mathbf {g} ^ {\\mathcal {T}}\\right)) \\in \\mathbb {R} ^ {K}, \\tag {5}", + "image_path": "750f449ca03671d64956566eea128c7a8b3e14e6ecf690cbbd310ecffc8de877.jpg" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 223, + 200, + 482, + 231 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 223, + 200, + 482, + 231 + ], + "spans": [ + { + "bbox": [ + 223, + 200, + 482, + 231 + ], + "type": "interline_equation", + "content": "c _ {\\mathrm {a g g}} \\left(\\mathbf {P} ^ {\\mathcal {V}}, \\mathbf {P} ^ {\\mathcal {T}}; \\mathbf {g} ^ {\\mathcal {T}}\\right) := \\sum_ {k = 1} ^ {K} \\mathrm {a} _ {k} \\cdot c \\left(\\mathbf {p} _ {k} ^ {\\mathcal {V}}, \\mathbf {p} _ {k} ^ {\\mathcal {T}}\\right), \\tag {6}", + "image_path": "cf017370f88232500ebae64b0ed8a9d00dd269b7c89e4895c7f56372c4c172ef.jpg" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 130, + 236, + 482, + 333 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 236, + 482, + 333 + ], + "spans": [ + { + "bbox": [ + 130, + 236, + 482, + 333 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 130, + 236, + 482, + 333 + ], + "type": "inline_equation", + "content": "c(\\cdot, \\cdot)" + }, + { + "bbox": [ + 130, + 236, + 482, + 333 + ], + "type": "text", + "content": " denotes cosine similarity between two embeddings, " + }, + { + "bbox": [ + 130, + 236, + 482, + 333 + ], + "type": "inline_equation", + "content": "a_k" + }, + { + "bbox": [ + 130, + 236, + 482, + 333 + ], + "type": "text", + "content": " is the " + }, + { + "bbox": [ + 130, + 236, + 482, + 333 + ], + "type": "inline_equation", + "content": "k" + }, + { + "bbox": [ + 130, + 236, + 482, + 333 + ], + "type": "text", + "content": "-th value of " + }, + { + "bbox": [ + 130, + 236, + 482, + 333 + ], + "type": "inline_equation", + "content": "\\mathbf{a}" + }, + { + "bbox": [ + 130, + 236, + 482, + 333 + ], + "type": "text", + "content": ", and " + }, + { + "bbox": [ + 130, + 236, + 482, + 333 + ], + "type": "inline_equation", + "content": "\\sigma(\\cdot)" + }, + { + "bbox": [ + 130, + 236, + 482, + 333 + ], + "type": "text", + "content": " is a softmax function. If given textual global embedding " + }, + { + "bbox": [ + 130, + 236, + 482, + 333 + ], + "type": "inline_equation", + "content": "\\mathbf{g}^{\\mathcal{T}}" + }, + { + "bbox": [ + 130, + 236, + 482, + 333 + ], + "type": "text", + "content": ", TDPA is computed by MLP which is learned in an end-to-end manner by minimizing our partNCE loss, as will be introduced in Sec 3.4. This dynamic attention enables us to adaptively assign importance to each part embedding, which reduces the impact of non-informative part embedding and promotes a more informative similarity measurement that reflects the significant semantic similarity of each text query." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 131, + 348, + 257, + 361 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 131, + 348, + 257, + 361 + ], + "spans": [ + { + "bbox": [ + 131, + 348, + 257, + 361 + ], + "type": "text", + "content": "3.4 Learning Objective" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 130, + 365, + 482, + 461 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 365, + 482, + 461 + ], + "spans": [ + { + "bbox": [ + 130, + 365, + 482, + 461 + ], + "type": "text", + "content": "Our model is trained through the establishment of cross-modal alignments, considering both global and part levels. The alignment at the global-level is accomplished by aligning global embeddings " + }, + { + "bbox": [ + 130, + 365, + 482, + 461 + ], + "type": "inline_equation", + "content": "\\mathbf{g}^{\\mathcal{V}}" + }, + { + "bbox": [ + 130, + 365, + 482, + 461 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 130, + 365, + 482, + 461 + ], + "type": "inline_equation", + "content": "\\mathbf{g}^{\\mathcal{T}}" + }, + { + "bbox": [ + 130, + 365, + 482, + 461 + ], + "type": "text", + "content": " that capture the comprehensive information of each modality. For part-level alignment, we leverage part embeddings " + }, + { + "bbox": [ + 130, + 365, + 482, + 461 + ], + "type": "inline_equation", + "content": "\\mathbf{P}^{\\mathcal{V}}" + }, + { + "bbox": [ + 130, + 365, + 482, + 461 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 130, + 365, + 482, + 461 + ], + "type": "inline_equation", + "content": "\\mathbf{P}^{\\mathcal{T}}" + }, + { + "bbox": [ + 130, + 365, + 482, + 461 + ], + "type": "text", + "content": " that have locally exclusive features within each modality due to the slot attention. Furthermore, by dynamically adjusting weights of specific part slots based on the text global embedding " + }, + { + "bbox": [ + 130, + 365, + 482, + 461 + ], + "type": "inline_equation", + "content": "\\mathbf{g}^{\\mathcal{T}}" + }, + { + "bbox": [ + 130, + 365, + 482, + 461 + ], + "type": "text", + "content": ", we facilitate the learning of more informative alignments." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 131, + 461, + 482, + 545 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 131, + 461, + 482, + 545 + ], + "spans": [ + { + "bbox": [ + 131, + 461, + 482, + 545 + ], + "type": "text", + "content": "Global Alignment Loss: To align global embeddings " + }, + { + "bbox": [ + 131, + 461, + 482, + 545 + ], + "type": "inline_equation", + "content": "\\mathbf{g}^{\\mathcal{V}}" + }, + { + "bbox": [ + 131, + 461, + 482, + 545 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 131, + 461, + 482, + 545 + ], + "type": "inline_equation", + "content": "\\mathbf{g}^{\\mathcal{T}}" + }, + { + "bbox": [ + 131, + 461, + 482, + 545 + ], + "type": "text", + "content": " extracted from each modality backbone, we first define a batch of global embeddings " + }, + { + "bbox": [ + 131, + 461, + 482, + 545 + ], + "type": "inline_equation", + "content": "\\mathcal{B}_{\\mathrm{global}} = \\{(\\mathbf{g}_i^{\\mathcal{V}},\\mathbf{g}_i^{\\mathcal{T}})\\}_{i = 1}^B" + }, + { + "bbox": [ + 131, + 461, + 482, + 545 + ], + "type": "text", + "content": ", where " + }, + { + "bbox": [ + 131, + 461, + 482, + 545 + ], + "type": "inline_equation", + "content": "B" + }, + { + "bbox": [ + 131, + 461, + 482, + 545 + ], + "type": "text", + "content": " is batch size. Then we adopt the InfoNCE loss [17] which is a contrastive learning objective that maximizes the similarity between embeddings of positive pairs while minimizing the similarity between negative pairs in the batch. Consequently, our global NCE loss with cosine similarity function " + }, + { + "bbox": [ + 131, + 461, + 482, + 545 + ], + "type": "inline_equation", + "content": "c(\\cdot ,\\cdot)" + }, + { + "bbox": [ + 131, + 461, + 482, + 545 + ], + "type": "text", + "content": " is formulated as follows:" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 181, + 552, + 482, + 590 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 181, + 552, + 482, + 590 + ], + "spans": [ + { + "bbox": [ + 181, + 552, + 482, + 590 + ], + "type": "interline_equation", + "content": "\\mathcal {L} _ {\\mathrm {N C E}} = - \\sum_ {i = 1} ^ {B} \\left(\\log \\frac {e ^ {c \\left(\\mathbf {g} _ {i} ^ {\\mathcal {V}} , \\mathbf {g} _ {i} ^ {\\mathcal {T}}\\right) / \\tau}}{\\sum_ {j = 1} ^ {B} e ^ {c \\left(\\mathbf {g} _ {i} ^ {\\mathcal {V}} , \\mathbf {g} _ {j} ^ {\\mathcal {T}}\\right) / \\tau}} + \\log \\frac {e ^ {c \\left(\\mathbf {g} _ {i} ^ {\\mathcal {V}} , \\mathbf {g} _ {i} ^ {\\mathcal {T}}\\right) / \\tau}}{\\sum_ {j = 1} ^ {B} e ^ {c \\left(\\mathbf {g} _ {j} ^ {\\mathcal {V}} , \\mathbf {g} _ {i} ^ {\\mathcal {T}}\\right) / \\tau}}\\right), \\tag {7}", + "image_path": "84810b92869ca528459155167d3db84a222c49d7d6518cc8d30bb8c3ccb508ea.jpg" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 130, + 594, + 482, + 631 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 594, + 482, + 631 + ], + "spans": [ + { + "bbox": [ + 130, + 594, + 482, + 631 + ], + "type": "text", + "content": "where the " + }, + { + "bbox": [ + 130, + 594, + 482, + 631 + ], + "type": "inline_equation", + "content": "\\tau" + }, + { + "bbox": [ + 130, + 594, + 482, + 631 + ], + "type": "text", + "content": " is temperature term. Additionally, we employ an identity classification loss " + }, + { + "bbox": [ + 130, + 594, + 482, + 631 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_{\\mathrm{ID}}" + }, + { + "bbox": [ + 130, + 594, + 482, + 631 + ], + "type": "text", + "content": " to ensure that the embeddings extracted from the same identity become similar. The " + }, + { + "bbox": [ + 130, + 594, + 482, + 631 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_{\\mathrm{ID}}" + }, + { + "bbox": [ + 130, + 594, + 482, + 631 + ], + "type": "text", + "content": " is denoted by" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 194, + 637, + 482, + 667 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 194, + 637, + 482, + 667 + ], + "spans": [ + { + "bbox": [ + 194, + 637, + 482, + 667 + ], + "type": "interline_equation", + "content": "\\mathcal {L} _ {\\mathrm {I D}} = - \\sum_ {i = 1} ^ {B} \\left(\\boldsymbol {y} _ {i} \\log \\sigma \\left(\\mathbf {g} _ {i} ^ {\\mathcal {V}} \\mathbf {W} _ {\\mathrm {I D}}\\right) + \\boldsymbol {y} _ {i} \\log \\sigma \\left(\\mathbf {g} _ {i} ^ {\\mathcal {T}} \\mathbf {W} _ {\\mathrm {I D}}\\right)\\right), \\tag {8}", + "image_path": "dd3691ab529b58bd41b8e63d914693d8089194494b8d5d70bf7c8d821a6b89a1.jpg" + } + ] + } + ], + "index": 11 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 147, + 91, + 448, + 103 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 147, + 91, + 448, + 103 + ], + "spans": [ + { + "bbox": [ + 147, + 91, + 448, + 103 + ], + "type": "text", + "content": "PLOT: TBPS with Part Slot Attention for Corresponding Part Discovery" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 474, + 91, + 481, + 101 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 474, + 91, + 481, + 101 + ], + "spans": [ + { + "bbox": [ + 474, + 91, + 481, + 101 + ], + "type": "text", + "content": "7" + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 6 + }, + { + "para_blocks": [ + { + "bbox": [ + 130, + 116, + 482, + 177 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 116, + 482, + 177 + ], + "spans": [ + { + "bbox": [ + 130, + 116, + 482, + 177 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 130, + 116, + 482, + 177 + ], + "type": "inline_equation", + "content": "\\mathbf{y}_i\\in \\mathbb{R}^C" + }, + { + "bbox": [ + 130, + 116, + 482, + 177 + ], + "type": "text", + "content": " is the identity ground truth of corresponding global embedding " + }, + { + "bbox": [ + 130, + 116, + 482, + 177 + ], + "type": "inline_equation", + "content": "\\mathbf{g}_i" + }, + { + "bbox": [ + 130, + 116, + 482, + 177 + ], + "type": "text", + "content": " represented by a one-hot vector and " + }, + { + "bbox": [ + 130, + 116, + 482, + 177 + ], + "type": "inline_equation", + "content": "C" + }, + { + "bbox": [ + 130, + 116, + 482, + 177 + ], + "type": "text", + "content": " is the number of identities, " + }, + { + "bbox": [ + 130, + 116, + 482, + 177 + ], + "type": "inline_equation", + "content": "\\mathbf{W}_{\\mathrm{ID}}\\in \\mathbb{R}^{D\\times C}" + }, + { + "bbox": [ + 130, + 116, + 482, + 177 + ], + "type": "text", + "content": " is a classifier shared between the two modalities. For cross-modal alignments, the commonly used CMPM loss [7,30] is additionally adopted. Finally, we describe global alignment loss as below:" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 232, + 186, + 482, + 198 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 232, + 186, + 482, + 198 + ], + "spans": [ + { + "bbox": [ + 232, + 186, + 482, + 198 + ], + "type": "interline_equation", + "content": "\\mathcal {L} _ {\\text {G l o b a l}} = \\mathcal {L} _ {\\text {N C E}} + \\mathcal {L} _ {\\text {I D}} + \\mathcal {L} _ {\\text {C M P M}}. \\tag {9}", + "image_path": "f2990bbc0805c6ebab5e48d3d0a94effd8b1b19e03ff1239923c54fdebaaa06b.jpg" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 131, + 207, + 483, + 291 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 131, + 207, + 483, + 291 + ], + "spans": [ + { + "bbox": [ + 131, + 207, + 483, + 291 + ], + "type": "text", + "content": "Part Alignment Loss: Similar to the alignment of global embeddings, we first define a batch of part embeddings " + }, + { + "bbox": [ + 131, + 207, + 483, + 291 + ], + "type": "inline_equation", + "content": "\\mathcal{B}_{\\mathrm{part}} = \\{(\\mathbf{P}_i^{\\mathcal{V}},\\mathbf{P}_i^{\\mathcal{T}})\\}_{i = 1}^B" + }, + { + "bbox": [ + 131, + 207, + 483, + 291 + ], + "type": "text", + "content": " with batch size " + }, + { + "bbox": [ + 131, + 207, + 483, + 291 + ], + "type": "inline_equation", + "content": "B" + }, + { + "bbox": [ + 131, + 207, + 483, + 291 + ], + "type": "text", + "content": ", and we adopt the InfoNCE loss to align part embeddings extracted from two modalities. For learning text query-based informative alignment of part embeddings between modalities, we employ text query-based similarity aggregation function " + }, + { + "bbox": [ + 131, + 207, + 483, + 291 + ], + "type": "inline_equation", + "content": "c_{\\mathrm{agg}}" + }, + { + "bbox": [ + 131, + 207, + 483, + 291 + ], + "type": "text", + "content": " in Eq. 6 to compute InfoNCE loss. We termed this InfoNCE with " + }, + { + "bbox": [ + 131, + 207, + 483, + 291 + ], + "type": "inline_equation", + "content": "c_{\\mathrm{agg}}" + }, + { + "bbox": [ + 131, + 207, + 483, + 291 + ], + "type": "text", + "content": " as PartNCE loss, and it is formulated by" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 140, + 299, + 482, + 339 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 140, + 299, + 482, + 339 + ], + "spans": [ + { + "bbox": [ + 140, + 299, + 482, + 339 + ], + "type": "interline_equation", + "content": "\\mathcal {L} _ {\\mathrm {P a r t N C E}} = - \\sum_ {i = 1} ^ {B} \\left(\\log \\frac {e ^ {c _ {\\mathrm {a g g}} (\\mathbf {P} _ {i} ^ {\\mathcal {V}} , \\mathbf {P} _ {i} ^ {\\mathcal {T}} ; \\mathbf {g} _ {i} ^ {\\mathcal {T}}) / \\tau}}{\\sum_ {j = 1} ^ {B} e ^ {c _ {\\mathrm {a g g}} (\\mathbf {P} _ {i} ^ {\\mathcal {V}} , \\mathbf {P} _ {j} ^ {\\mathcal {T}} ; \\mathbf {g} _ {j} ^ {\\mathcal {T}}) / \\tau}} + \\log \\frac {e ^ {c _ {\\mathrm {a g g}} (\\mathbf {P} _ {i} ^ {\\mathcal {V}} , \\mathbf {P} _ {i} ^ {\\mathcal {T}} ; \\mathbf {g} _ {i} ^ {\\mathcal {T}}) / \\tau}}{\\sum_ {j = 1} ^ {B} e ^ {c _ {\\mathrm {a g g}} \\left(\\mathbf {P} _ {j} ^ {\\mathcal {V}}, \\mathbf {P} _ {i} ^ {\\mathcal {T}}; \\mathbf {g} _ {i} ^ {\\mathcal {T}}\\right) / \\tau}\\right), \\tag {10}", + "image_path": "1419123bb57d707fc626c78717923a44601743c91c162cf9c2c1240ddbe51901.jpg" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 131, + 345, + 482, + 394 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 131, + 345, + 482, + 394 + ], + "spans": [ + { + "bbox": [ + 131, + 345, + 482, + 394 + ], + "type": "text", + "content": "Similar to global alignment loss, we adopt the identity loss that shares classifier weights between the two modalities for part alignment; however, the difference is that the part embeddings are concatenated along the embedding dimension. The identity loss for part alignment is formulated by" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 157, + 402, + 482, + 434 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 157, + 402, + 482, + 434 + ], + "spans": [ + { + "bbox": [ + 157, + 402, + 482, + 434 + ], + "type": "interline_equation", + "content": "\\mathcal {L} _ {\\text {P a r t I D}} = - \\sum_ {i = 1} ^ {B} \\left(\\boldsymbol {y} _ {i} \\log \\sigma \\left(\\left[ \\boldsymbol {P} _ {i} ^ {\\mathcal {V}} \\right] \\boldsymbol {W} _ {\\text {P a r t I D}}\\right) + \\boldsymbol {y} _ {i} \\log \\sigma \\left(\\left[ \\boldsymbol {P} _ {i} ^ {\\mathcal {T}} \\right] \\boldsymbol {W} _ {\\text {P a r t I D}}\\right)\\right), \\tag {11}", + "image_path": "bcb3f04276dabf91c6b63df6b653a545d41a2acbc13ebfbfc050d3bf4908c8dc.jpg" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 131, + 443, + 483, + 479 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 131, + 443, + 483, + 479 + ], + "spans": [ + { + "bbox": [ + 131, + 443, + 483, + 479 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 131, + 443, + 483, + 479 + ], + "type": "inline_equation", + "content": "[\\mathbf{P}] \\in \\mathbb{R}^{KD}" + }, + { + "bbox": [ + 131, + 443, + 483, + 479 + ], + "type": "text", + "content": " is the concatenation of part embeddings " + }, + { + "bbox": [ + 131, + 443, + 483, + 479 + ], + "type": "inline_equation", + "content": "\\mathbf{P}" + }, + { + "bbox": [ + 131, + 443, + 483, + 479 + ], + "type": "text", + "content": " along the embedding dimension, and " + }, + { + "bbox": [ + 131, + 443, + 483, + 479 + ], + "type": "inline_equation", + "content": "\\mathbf{W}_{\\mathrm{PartID}} \\in \\mathbb{R}^{KD \\times C}" + }, + { + "bbox": [ + 131, + 443, + 483, + 479 + ], + "type": "text", + "content": " is a classifier shared between both modalities. Finally, the part alignment loss is computed by" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 242, + 489, + 482, + 502 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 242, + 489, + 482, + 502 + ], + "spans": [ + { + "bbox": [ + 242, + 489, + 482, + 502 + ], + "type": "interline_equation", + "content": "\\mathcal {L} _ {\\text {P a r t}} = \\mathcal {L} _ {\\text {P a r t N C E}} + \\mathcal {L} _ {\\text {P a r t I D}}. \\tag {12}", + "image_path": "e14aa32e20cf4ba99bebae3729ce6bae9079e00880499140a0eebba88081282d.jpg" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 131, + 521, + 317, + 605 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 131, + 521, + 317, + 605 + ], + "spans": [ + { + "bbox": [ + 131, + 521, + 317, + 605 + ], + "type": "text", + "content": "Cross-Modal Masked Language Modeling Loss: Following the practices of employing Transformer-based backbones in existing methods [7,21], we adopt an auxiliary loss, the cross-modal masked language modeling (CMLM) loss, to facilitate the learning of interactions between modali" + } + ] + } + ], + "index": 10 + }, + { + "type": "image", + "bbox": [ + 326, + 515, + 476, + 582 + ], + "blocks": [ + { + "bbox": [ + 326, + 515, + 476, + 582 + ], + "lines": [ + { + "bbox": [ + 326, + 515, + 476, + 582 + ], + "spans": [ + { + "bbox": [ + 326, + 515, + 476, + 582 + ], + "type": "image", + "image_path": "d6bcbcab86d9b112f25ee0477689121305b220c886e6338817255b1b29ca810c.jpg" + } + ] + } + ], + "index": 11, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 329, + 583, + 471, + 592 + ], + "lines": [ + { + "bbox": [ + 329, + 583, + 471, + 592 + ], + "spans": [ + { + "bbox": [ + 329, + 583, + 471, + 592 + ], + "type": "text", + "content": "Cross-Modal Masked Language Model (CMLM)" + } + ] + } + ], + "index": 12, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 339, + 593, + 463, + 603 + ], + "lines": [ + { + "bbox": [ + 339, + 593, + 463, + 603 + ], + "spans": [ + { + "bbox": [ + 339, + 593, + 463, + 603 + ], + "type": "text", + "content": "Fig. 2: Illustration of CMLM." + } + ] + } + ], + "index": 13, + "angle": 0, + "type": "image_caption" + } + ], + "index": 11 + }, + { + "bbox": [ + 130, + 605, + 482, + 666 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 605, + 482, + 666 + ], + "spans": [ + { + "bbox": [ + 130, + 605, + 482, + 666 + ], + "type": "text", + "content": "ties. Similar to BERT [3], given a text description, we randomly select text tokens with a " + }, + { + "bbox": [ + 130, + 605, + 482, + 666 + ], + "type": "inline_equation", + "content": "15\\%" + }, + { + "bbox": [ + 130, + 605, + 482, + 666 + ], + "type": "text", + "content": " probability and replace them with the learnable [MASK] token. The masked text description is then processed through the text backbone to obtain the masked textual tokens. After obtaining the masked textual tokens, we concatenate them with the visual tokens extracted via the image backbone." + } + ] + } + ], + "index": 14 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "spans": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "type": "text", + "content": "8" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 166, + 91, + 220, + 101 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 166, + 91, + 220, + 101 + ], + "spans": [ + { + "bbox": [ + 166, + 91, + 220, + 101 + ], + "type": "text", + "content": "J. Park et al." + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 7 + }, + { + "para_blocks": [ + { + "bbox": [ + 130, + 116, + 482, + 201 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 116, + 482, + 201 + ], + "spans": [ + { + "bbox": [ + 130, + 116, + 482, + 201 + ], + "type": "text", + "content": "The concatenated tokens are then fed into a transformer to acquire cross-modal fused tokens. Among the fused tokens, those corresponding to the indices of textual tokens are denoted as " + }, + { + "bbox": [ + 130, + 116, + 482, + 201 + ], + "type": "inline_equation", + "content": "\\mathbf{F}" + }, + { + "bbox": [ + 130, + 116, + 482, + 201 + ], + "type": "text", + "content": " composed as " + }, + { + "bbox": [ + 130, + 116, + 482, + 201 + ], + "type": "inline_equation", + "content": "\\{\\mathbf{f}_1,\\dots \\mathbf{f}_L\\}" + }, + { + "bbox": [ + 130, + 116, + 482, + 201 + ], + "type": "text", + "content": ", where a fused token " + }, + { + "bbox": [ + 130, + 116, + 482, + 201 + ], + "type": "inline_equation", + "content": "\\mathbf{f}\\in \\mathbb{R}^{1\\times D}" + }, + { + "bbox": [ + 130, + 116, + 482, + 201 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 130, + 116, + 482, + 201 + ], + "type": "inline_equation", + "content": "L" + }, + { + "bbox": [ + 130, + 116, + 482, + 201 + ], + "type": "text", + "content": " indicates the max length of input text descriptions. Ultimately, these tokens are fed into a CMLM classifier " + }, + { + "bbox": [ + 130, + 116, + 482, + 201 + ], + "type": "inline_equation", + "content": "\\mathbf{W}_{\\mathrm{CMLM}}" + }, + { + "bbox": [ + 130, + 116, + 482, + 201 + ], + "type": "text", + "content": " to predict the probability of vocabulary IDs. The overall procedure of CMLM is illustrated in Fig. 2 and through this procedure we compute the " + }, + { + "bbox": [ + 130, + 116, + 482, + 201 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_{\\mathrm{CMLM}}" + }, + { + "bbox": [ + 130, + 116, + 482, + 201 + ], + "type": "text", + "content": " loss as follows:" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 222, + 209, + 482, + 240 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 222, + 209, + 482, + 240 + ], + "spans": [ + { + "bbox": [ + 222, + 209, + 482, + 240 + ], + "type": "interline_equation", + "content": "\\mathcal {L} _ {\\mathrm {C M L M}} = - \\frac {1}{L} \\sum_ {l = 1} ^ {L} \\mathbf {y} _ {l} \\log (\\sigma (\\mathbf {f} _ {l} \\mathbf {W} _ {\\mathrm {C M L M}})), \\tag {13}", + "image_path": "e3c4d41bf01bdf8e5ab0f68346c0f4b517eacf02b07297bdabd1ff813420ed56.jpg" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 130, + 249, + 483, + 286 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 249, + 483, + 286 + ], + "spans": [ + { + "bbox": [ + 130, + 249, + 483, + 286 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 130, + 249, + 483, + 286 + ], + "type": "inline_equation", + "content": "\\mathbf{W}_{\\mathrm{CMLM}} \\in \\mathbb{R}^{D \\times V}" + }, + { + "bbox": [ + 130, + 249, + 483, + 286 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 130, + 249, + 483, + 286 + ], + "type": "inline_equation", + "content": "\\pmb{y}_l \\in \\mathbb{R}^V" + }, + { + "bbox": [ + 130, + 249, + 483, + 286 + ], + "type": "text", + "content": " is the vocabulary ground truth of " + }, + { + "bbox": [ + 130, + 249, + 483, + 286 + ], + "type": "inline_equation", + "content": "l" + }, + { + "bbox": [ + 130, + 249, + 483, + 286 + ], + "type": "text", + "content": "-th text tokens, represented by a one-hot vector and " + }, + { + "bbox": [ + 130, + 249, + 483, + 286 + ], + "type": "inline_equation", + "content": "V" + }, + { + "bbox": [ + 130, + 249, + 483, + 286 + ], + "type": "text", + "content": " is the size of vocabulary. Finally, our overall objective function for training is denoted by" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 240, + 297, + 482, + 308 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 240, + 297, + 482, + 308 + ], + "spans": [ + { + "bbox": [ + 240, + 297, + 482, + 308 + ], + "type": "interline_equation", + "content": "\\mathcal {L} = \\mathcal {L} _ {\\text {G l o b a l}} + \\mathcal {L} _ {\\text {P a r t}} + \\mathcal {L} _ {\\text {C M L M}}. \\tag {14}", + "image_path": "7db23b092451e68593c7ed8d7fcb495b9de94ae6f537dd727c56816e489ed530.jpg" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 132, + 326, + 208, + 337 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 326, + 208, + 337 + ], + "spans": [ + { + "bbox": [ + 132, + 326, + 208, + 337 + ], + "type": "text", + "content": "3.5 Inference" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 130, + 345, + 483, + 453 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 345, + 483, + 453 + ], + "spans": [ + { + "bbox": [ + 130, + 345, + 483, + 453 + ], + "type": "text", + "content": "During testing, the global and part embeddings of each modality input are fully exploited to calculate the similarity between the image-text pair. In particular, the similarities between visual part embeddings and their corresponding textual part embeddings are linearly combined with the attention weights " + }, + { + "bbox": [ + 130, + 345, + 483, + 453 + ], + "type": "inline_equation", + "content": "\\mathbf{a}" + }, + { + "bbox": [ + 130, + 345, + 483, + 453 + ], + "type": "text", + "content": " to aggregate them. The image-text pair similarity is defined as the sum of the similarity between the global embeddings of the image-text pair and the similarity between the part embeddings of it, which can be computed by " + }, + { + "bbox": [ + 130, + 345, + 483, + 453 + ], + "type": "inline_equation", + "content": "c(\\mathbf{g}^{\\mathcal{V}},\\mathbf{g}^{\\mathcal{T}}) + c_{\\mathrm{agg}}(\\mathbf{P}^{\\mathcal{V}},\\mathbf{P}^{\\mathcal{T}};\\mathbf{g}^{\\mathcal{T}})" + }, + { + "bbox": [ + 130, + 345, + 483, + 453 + ], + "type": "text", + "content": ". Finally, given the text query, the images in the gallery are ranked according to similarity scores between the images and the text for inference." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 132, + 472, + 230, + 485 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 472, + 230, + 485 + ], + "spans": [ + { + "bbox": [ + 132, + 472, + 230, + 485 + ], + "type": "text", + "content": "4 Experiments" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 130, + 496, + 482, + 581 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 496, + 482, + 581 + ], + "spans": [ + { + "bbox": [ + 130, + 496, + 482, + 581 + ], + "type": "text", + "content": "In this section, we provide a detailed account of our experimental setup (Sec. 4.1), evaluate our method, and compare it with state of the arts on three benchmark datasets for text-based person search (Sec. 4.2). Furthermore, we qualitatively present retrieval results and analyze the effectiveness of the part discovery module and TDPA with visualization results (Sec. 4.3). We also conduct ablation studies on the losses employed in model training, the methodologies of part discovery, and the strategies for part similarity aggregation (Sec. 4.4)." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 132, + 597, + 262, + 611 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 597, + 262, + 611 + ], + "spans": [ + { + "bbox": [ + 132, + 597, + 262, + 611 + ], + "type": "text", + "content": "4.1 Experimental Setup" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 130, + 617, + 482, + 667 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 617, + 482, + 667 + ], + "spans": [ + { + "bbox": [ + 130, + 617, + 482, + 667 + ], + "type": "text", + "content": "Datasets: On three benchmark datasets, CUHK-PEDES [11], ICFG-PEDES [4], and RSTPReid [32], we evaluate and compare the performance of our method against previous methods. In CUHK-PEDES collected from five existing person re-identification datasets [6,12,13,28,31], it contains 40,206 images corresponding" + } + ] + } + ], + "index": 11 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 148, + 91, + 448, + 103 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 148, + 91, + 448, + 103 + ], + "spans": [ + { + "bbox": [ + 148, + 91, + 448, + 103 + ], + "type": "text", + "content": "PLOT: TBPS with Part Slot Attention for Corresponding Part Discovery" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 474, + 92, + 480, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 474, + 92, + 480, + 100 + ], + "spans": [ + { + "bbox": [ + 474, + 92, + 480, + 100 + ], + "type": "text", + "content": "9" + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 8 + }, + { + "para_blocks": [ + { + "bbox": [ + 130, + 116, + 482, + 272 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 116, + 482, + 272 + ], + "spans": [ + { + "bbox": [ + 130, + 116, + 482, + 272 + ], + "type": "text", + "content": "to 13,003 individual IDs, with each image being approximately matched with two annotated text descriptions. We follow the data split of [11] with 34,054 images from 11,003 person IDs and 68,126 text descriptions for training, 3,078 images from 1,000 IDs and 6,158 text descriptions for validation, and 3,074 images from 1,000 IDs and 6,156 text descriptions for testing. The remaining two datasets are collected from MSMT17 [26]. ICFG-PEDES consists of 54,522 image-text pairs from 4,102 individual IDs, which are split into 34,674 and 19,848 for training and testing, respectively. RSTPReid contains 20,505 images of 4,101 individual IDs, with each ID having 5 images and each image associated with the corresponding two annotated text descriptions. We follow the data split of [32] with 18,505 images from 3,701 IDs and 37,010 text descriptions for training, 1,000 images from 200 IDs and 2,000 text descriptions for validation, and 1,000 images from 200 IDs and 2,000 text descriptions for testing, respectively." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 130, + 272, + 482, + 320 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 272, + 482, + 320 + ], + "spans": [ + { + "bbox": [ + 130, + 272, + 482, + 320 + ], + "type": "text", + "content": "Evaluation Protocol: We employ the standard metric of rank at K (R@K=1,5, 10) for all retrieval experiments. Specifically, given a query text, images are sorted based on their similarity to the query text. The search is considered correct if at least one relevant image appears in the top K positions of the ranking." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 130, + 320, + 482, + 379 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 320, + 482, + 379 + ], + "spans": [ + { + "bbox": [ + 130, + 320, + 482, + 379 + ], + "type": "text", + "content": "Network Architecture: We adopt the pre-trained CLIP models from OpenAI [18] for both image and text encoders, where the size of the image encoder is ViT-B/16. The input images are resized to " + }, + { + "bbox": [ + 130, + 320, + 482, + 379 + ], + "type": "inline_equation", + "content": "384 \\times 128" + }, + { + "bbox": [ + 130, + 320, + 482, + 379 + ], + "type": "text", + "content": ". Random horizontal flipping, random cropping, and random erasing are applied for the data augmentation in training time. The maximum text length is set to 77." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 130, + 380, + 482, + 439 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 380, + 482, + 439 + ], + "spans": [ + { + "bbox": [ + 130, + 380, + 482, + 439 + ], + "type": "text", + "content": "Network Optimization: Our model is trained using the Adam optimizer for 60 epochs with a batch size of 128 for all experiments. For the CLIP encoders, the initial learning rate is set to " + }, + { + "bbox": [ + 130, + 380, + 482, + 439 + ], + "type": "inline_equation", + "content": "5\\mathrm{e} - 6" + }, + { + "bbox": [ + 130, + 380, + 482, + 439 + ], + "type": "text", + "content": ", using a cosine schedule with the warm-up strategy at the first five epochs; we use a high learning rate for the remaining parameters by scaling 20 times." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 130, + 439, + 482, + 474 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 439, + 482, + 474 + ], + "spans": [ + { + "bbox": [ + 130, + 439, + 482, + 474 + ], + "type": "text", + "content": "Hyperparameters: We set the number of part slots to 8. The number of iterations in the part slot attention block is set to 5. The temperature parameter " + }, + { + "bbox": [ + 130, + 439, + 482, + 474 + ], + "type": "inline_equation", + "content": "\\tau" + }, + { + "bbox": [ + 130, + 439, + 482, + 474 + ], + "type": "text", + "content": " is set to 0.015." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 131, + 491, + 265, + 503 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 131, + 491, + 265, + 503 + ], + "spans": [ + { + "bbox": [ + 131, + 491, + 265, + 503 + ], + "type": "text", + "content": "4.2 Quantitative Results" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 130, + 510, + 482, + 666 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 510, + 482, + 666 + ], + "spans": [ + { + "bbox": [ + 130, + 510, + 482, + 666 + ], + "type": "text", + "content": "We compare our method with previous text-based person search methods on CUHK-PEDES [11], ICFG-PEDES [4], and RSTPReid [32]. The performance comparison and the backbones for each modality employed by each method are shown in Table 1. Specifically, our method achieves an outstanding " + }, + { + "bbox": [ + 130, + 510, + 482, + 666 + ], + "type": "inline_equation", + "content": "\\mathbb{R}@\\mathbb{1}" + }, + { + "bbox": [ + 130, + 510, + 482, + 666 + ], + "type": "text", + "content": " metric of " + }, + { + "bbox": [ + 130, + 510, + 482, + 666 + ], + "type": "inline_equation", + "content": "75.28\\%" + }, + { + "bbox": [ + 130, + 510, + 482, + 666 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 130, + 510, + 482, + 666 + ], + "type": "inline_equation", + "content": "65.76\\%" + }, + { + "bbox": [ + 130, + 510, + 482, + 666 + ], + "type": "text", + "content": ", and " + }, + { + "bbox": [ + 130, + 510, + 482, + 666 + ], + "type": "inline_equation", + "content": "61.80\\%" + }, + { + "bbox": [ + 130, + 510, + 482, + 666 + ], + "type": "text", + "content": " on the CUHK-PEDES, ICFG-PEDES, and RSTPReid datasets, respectively, outperforming previous methods utilizing different backbones from ours. Moreover, our method improves the previous state of the art, IRRA [7], on " + }, + { + "bbox": [ + 130, + 510, + 482, + 666 + ], + "type": "inline_equation", + "content": "\\mathbb{R}@\\mathbb{1}" + }, + { + "bbox": [ + 130, + 510, + 482, + 666 + ], + "type": "text", + "content": " by a large margin as " + }, + { + "bbox": [ + 130, + 510, + 482, + 666 + ], + "type": "inline_equation", + "content": "1.9\\% \\mathbb{p}" + }, + { + "bbox": [ + 130, + 510, + 482, + 666 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 130, + 510, + 482, + 666 + ], + "type": "inline_equation", + "content": "1.7\\% \\mathbb{p}" + }, + { + "bbox": [ + 130, + 510, + 482, + 666 + ], + "type": "text", + "content": ", and " + }, + { + "bbox": [ + 130, + 510, + 482, + 666 + ], + "type": "inline_equation", + "content": "1.6\\% \\mathbb{p}" + }, + { + "bbox": [ + 130, + 510, + 482, + 666 + ], + "type": "text", + "content": ", respectively. Since IRRA only focuses on aligning the global embeddings of each modality, it is hard to capture the fine-grained differences. In contrast, our method not only takes account of the global embeddings but also aims to discover and align discriminative part embeddings from each modality; it allows the model to effectively find the target person." + } + ] + } + ], + "index": 8 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "spans": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "type": "text", + "content": "10" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 166, + 91, + 220, + 100 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 166, + 91, + 220, + 100 + ], + "spans": [ + { + "bbox": [ + 166, + 91, + 220, + 100 + ], + "type": "text", + "content": "J. Park et al." + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 9 + }, + { + "para_blocks": [ + { + "type": "table", + "bbox": [ + 132, + 146, + 479, + 348 + ], + "blocks": [ + { + "bbox": [ + 131, + 114, + 482, + 137 + ], + "lines": [ + { + "bbox": [ + 131, + 114, + 482, + 137 + ], + "spans": [ + { + "bbox": [ + 131, + 114, + 482, + 137 + ], + "type": "text", + "content": "Table 1: Performance of text-based person search methods on the three datasets. Bold and underline denote the best and the second best." + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 132, + 146, + 479, + 348 + ], + "lines": [ + { + "bbox": [ + 132, + 146, + 479, + 348 + ], + "spans": [ + { + "bbox": [ + 132, + 146, + 479, + 348 + ], + "type": "table", + "html": "
MethodsBackboneCUHK-PEDESICFG-PEDESRSTPReid
ImageTextR@1R@5R@10R@1R@5R@10R@1R@5R@10
GNA-RNN [11]RN50LSTM19.05-53.64------
CMPM/C [30]RN50LSTM49.3771.6979.2743.5165.4474.26---
PMA [8]RN50BERT53.8173.5481.23------
TIMAM [19]RN101BERT54.5177.5684.78------
SCAN [9]RN50BERT55.8675.9783.6950.0569.6577.21---
ViTAA [24]RN50LSTM55.9775.8483.5250.9868.7975.78---
NAFS [5]RN50BERT59.9479.8686.70------
DSSL [32]RN50BERT59.9880.4187.56---32.4355.0863.19
MGEL [23]RN50LSTM60.2780.0186.74------
SSAN [4]RN50LSTM61.3780.1586.7354.2372.6379.5343.5067.8077.15
LapsCore [27]RN50BERT63.40-87.80------
SRCF [22]RN50BERT64.0482.9988.8157.1875.0181.49---
LGUR [20]RN50BERT64.2181.9487.9357.4274.9781.45---
TIPCB [2]RN50BERT64.2683.1989.10------
CAIBC [25]RN50BERT64.4382.8788.37---47.3569.5579.00
SAF [10]ViT-B/16BERT64.1382.6288.40------
IVT [21]ViT-B/16BERT65.5983.1189.2156.0473.6080.2246.7070.0078.80
CFine [29]CLIP-ViT-B/16BERT69.5785.9391.1560.8375.5582.4250.5572.5081.60
IRRA [7]CLIP-ViT-B/16CLIP-Xformer73.3889.9393.7163.4680.2485.8260.2081.3088.20
OursCLIP-ViT-B/16CLIP-Xformer75.2890.4294.1265.7681.3986.7361.8082.8589.45
", + "image_path": "f8e7e1bca7d6fde188ba5cee8e5e2996f7c250bda42d91c40189a9b69af832ea.jpg" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "table_body" + } + ], + "index": 3 + }, + { + "bbox": [ + 132, + 371, + 257, + 383 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 371, + 257, + 383 + ], + "spans": [ + { + "bbox": [ + 132, + 371, + 257, + 383 + ], + "type": "text", + "content": "4.3 Qualitative Results" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 130, + 390, + 482, + 545 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 390, + 482, + 545 + ], + "spans": [ + { + "bbox": [ + 130, + 390, + 482, + 545 + ], + "type": "text", + "content": "Retrieval Results: Top-5 retrieval results of our method on the CUHK-PEDES dataset are illustrated in Fig. 3. Above all, it shows the overall satisfactory retrieval results. In particular, we can observe our model retrieves targets well, even with distinctive human parts that are small or located in various positions. For instance, the small distinctive human parts like \"ponytails\" and \"high skirts\" in Fig. 3(a), \"black shoes\" in Fig. 3(b), and \"blue plaid shorts\" in Fig. 3(c), as well as human parts that could appear in various viewpoints such as \"yellow shoulder bags\" in Fig. 3(a), \"floaty dresses\" in Fig. 3(b), and \"red backpacks\" in Fig. 3(d). The CUHK-PEDES dataset typically contains three target images on average in the search space, thereby most of the false matches in the figure are observed by a lack of additional targets. Despite the false matching due to the limitation of the dataset, the retrieval results are reasonable in that the retrieved false matching contains distinctive human parts described by the query description." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 130, + 545, + 483, + 666 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 545, + 483, + 666 + ], + "spans": [ + { + "bbox": [ + 130, + 545, + 483, + 666 + ], + "type": "text", + "content": "Visualization of Attention Map " + }, + { + "bbox": [ + 130, + 545, + 483, + 666 + ], + "type": "inline_equation", + "content": "\\bar{A}_k" + }, + { + "bbox": [ + 130, + 545, + 483, + 666 + ], + "type": "text", + "content": " in PSABlock: To demonstrate the effectiveness of our part discovery module, we visualize the attention map " + }, + { + "bbox": [ + 130, + 545, + 483, + 666 + ], + "type": "inline_equation", + "content": "\\bar{A}_k" + }, + { + "bbox": [ + 130, + 545, + 483, + 666 + ], + "type": "text", + "content": " in " + }, + { + "bbox": [ + 130, + 545, + 483, + 666 + ], + "type": "inline_equation", + "content": "T" + }, + { + "bbox": [ + 130, + 545, + 483, + 666 + ], + "type": "text", + "content": "-th iteration of the PSA block for both visual and textual modalities (in Eq. (2)). The visualization results are illustrated in Fig. 4. It not only demonstrates that the part embeddings extracted by our part discovery module capture distinctive human parts but also shows that the part embeddings extracted from the same part slot attend to the semantically identical human parts regardless of modality. For example, the 1st part slot typically focuses on bottom clothes, the 4th on footwear, the 5th on objects being held, the 7th on top clothes, and the 8th on the person's head. Furthermore, in Fig. 4 compared (a) and (b) of visual modality," + } + ] + } + ], + "index": 6 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 148, + 91, + 447, + 102 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 148, + 91, + 447, + 102 + ], + "spans": [ + { + "bbox": [ + 148, + 91, + 447, + 102 + ], + "type": "text", + "content": "PLOT: TBPS with Part Slot Attention for Corresponding Part Discovery" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 470, + 92, + 479, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 470, + 92, + 479, + 100 + ], + "spans": [ + { + "bbox": [ + 470, + 92, + 479, + 100 + ], + "type": "text", + "content": "11" + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 10 + }, + { + "para_blocks": [ + { + "bbox": [ + 148, + 114, + 292, + 142 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 148, + 114, + 292, + 142 + ], + "spans": [ + { + "bbox": [ + 148, + 114, + 292, + 142 + ], + "type": "text", + "content": "Query: a girl with a ponytail wearing a red top with blue denim thigh high skirt and carrying a yellow shoulder bag looking at a phone." + } + ] + } + ], + "index": 2 + }, + { + "type": "image", + "bbox": [ + 144, + 144, + 298, + 198 + ], + "blocks": [ + { + "bbox": [ + 144, + 144, + 298, + 198 + ], + "lines": [ + { + "bbox": [ + 144, + 144, + 298, + 198 + ], + "spans": [ + { + "bbox": [ + 144, + 144, + 298, + 198 + ], + "type": "image", + "image_path": "cbfd0667d02d7bd6323ed76a4faf2d80e3e8f2baa6e044985e825293d7c7ee61.jpg" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 215, + 201, + 225, + 209 + ], + "lines": [ + { + "bbox": [ + 215, + 201, + 225, + 209 + ], + "spans": [ + { + "bbox": [ + 215, + 201, + 225, + 209 + ], + "type": "text", + "content": "(a)" + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "image_caption" + } + ], + "index": 3 + }, + { + "bbox": [ + 148, + 210, + 296, + 237 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 148, + 210, + 296, + 237 + ], + "spans": [ + { + "bbox": [ + 148, + 210, + 296, + 237 + ], + "type": "text", + "content": "Query: the little boy is wearing his hat backwards. he has on a blue, short sleeved shirt and grey plaid shorts." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 316, + 114, + 459, + 142 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 114, + 459, + 142 + ], + "spans": [ + { + "bbox": [ + 316, + 114, + 459, + 142 + ], + "type": "text", + "content": "Query: the woman is wearing a white cardigan over a long floaty dress and black shoes. she is carrying a back carry-on bag." + } + ] + } + ], + "index": 6 + }, + { + "type": "image", + "bbox": [ + 313, + 144, + 468, + 198 + ], + "blocks": [ + { + "bbox": [ + 313, + 144, + 468, + 198 + ], + "lines": [ + { + "bbox": [ + 313, + 144, + 468, + 198 + ], + "spans": [ + { + "bbox": [ + 313, + 144, + 468, + 198 + ], + "type": "image", + "image_path": "67b8f416eb5048dfab7e4fd5e2cfb1ca67048742e3b933f260eeeb49137f2040.jpg" + } + ] + } + ], + "index": 7, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 386, + 200, + 395, + 209 + ], + "lines": [ + { + "bbox": [ + 386, + 200, + 395, + 209 + ], + "spans": [ + { + "bbox": [ + 386, + 200, + 395, + 209 + ], + "type": "text", + "content": "(b)" + } + ] + } + ], + "index": 8, + "angle": 0, + "type": "image_caption" + } + ], + "index": 7 + }, + { + "bbox": [ + 316, + 210, + 459, + 245 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 210, + 459, + 245 + ], + "spans": [ + { + "bbox": [ + 316, + 210, + 459, + 245 + ], + "type": "text", + "content": "Query: person wearing a dark brown sweater, denim cropped pants that go just below the knees, and white sneakers. they are wearing a red backpack over both shoulders." + } + ] + } + ], + "index": 9 + }, + { + "type": "image", + "bbox": [ + 144, + 247, + 298, + 303 + ], + "blocks": [ + { + "bbox": [ + 144, + 247, + 298, + 303 + ], + "lines": [ + { + "bbox": [ + 144, + 247, + 298, + 303 + ], + "spans": [ + { + "bbox": [ + 144, + 247, + 298, + 303 + ], + "type": "image", + "image_path": "6e4c87227f1095c68d6de0a8a4b79ed07c5e5422f0256aec1c9ea1762bed1401.jpg" + } + ] + } + ], + "index": 10, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 216, + 306, + 225, + 314 + ], + "lines": [ + { + "bbox": [ + 216, + 306, + 225, + 314 + ], + "spans": [ + { + "bbox": [ + 216, + 306, + 225, + 314 + ], + "type": "text", + "content": "(c)" + } + ] + } + ], + "index": 11, + "angle": 0, + "type": "image_caption" + } + ], + "index": 10 + }, + { + "type": "image", + "bbox": [ + 313, + 247, + 467, + 303 + ], + "blocks": [ + { + "bbox": [ + 313, + 247, + 467, + 303 + ], + "lines": [ + { + "bbox": [ + 313, + 247, + 467, + 303 + ], + "spans": [ + { + "bbox": [ + 313, + 247, + 467, + 303 + ], + "type": "image", + "image_path": "31787daf2f547edde4d967cfa6876118dc51700a5992bd0461d3cee9ca9e9460.jpg" + } + ] + } + ], + "index": 12, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 386, + 306, + 396, + 314 + ], + "lines": [ + { + "bbox": [ + 386, + 306, + 396, + 314 + ], + "spans": [ + { + "bbox": [ + 386, + 306, + 396, + 314 + ], + "type": "text", + "content": "(d)" + } + ] + } + ], + "index": 13, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 130, + 314, + 481, + 347 + ], + "lines": [ + { + "bbox": [ + 130, + 314, + 481, + 347 + ], + "spans": [ + { + "bbox": [ + 130, + 314, + 481, + 347 + ], + "type": "text", + "content": "Fig. 3: Top-5 retrieval results of our method on the CUHK-PEDES dataset. Images are sorted from left to right according to their ranks below each text query. Green and red boxes indicate true and false matches, respectively." + } + ] + } + ], + "index": 14, + "angle": 0, + "type": "image_caption" + } + ], + "index": 12 + }, + { + "bbox": [ + 130, + 364, + 482, + 388 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 364, + 482, + 388 + ], + "spans": [ + { + "bbox": [ + 130, + 364, + 482, + 388 + ], + "type": "text", + "content": "our part discovery module is capable of capturing the distinctive human parts, while it is robust against pose variations and viewpoint changes." + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 130, + 388, + 482, + 555 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 388, + 482, + 555 + ], + "spans": [ + { + "bbox": [ + 130, + 388, + 482, + 555 + ], + "type": "text", + "content": "Visualization of TDPA: We visualize TDPA weights " + }, + { + "bbox": [ + 130, + 388, + 482, + 555 + ], + "type": "inline_equation", + "content": "\\mathbf{a}" + }, + { + "bbox": [ + 130, + 388, + 482, + 555 + ], + "type": "text", + "content": " in Eq. (6) to demonstrate the effectiveness of similarity aggregation between part embeddings of two modalities with TDPA. The visualization results are presented in the bottom left in Fig. 4(a, b), respectively. Comparing examples (a) and (b), the 5th part slot typically focuses on human parts associated with held objects, like bags, leading to a predicted lower TDPA weight for this part slot in (b) due to the absence of such distinctive information in its query description. However, the presence of the \"hat\" in the query description of (b) leads to having a high TDPA weight for the 8th part slot that typically focuses on the human head part. In contrast, since this related human head part information is not provided in the query description of (a), the TDPA weight for the 8th part slot in (a) is predicted to be a low value. These observations highlight the capability of TDPA, which adaptively improves part-based retrieval depending on the contents of the text query." + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 132, + 574, + 246, + 585 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 574, + 246, + 585 + ], + "spans": [ + { + "bbox": [ + 132, + 574, + 246, + 585 + ], + "type": "text", + "content": "4.4 Ablation Studies" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 130, + 594, + 482, + 666 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 594, + 482, + 666 + ], + "spans": [ + { + "bbox": [ + 130, + 594, + 482, + 666 + ], + "type": "text", + "content": "In our ablation studies conducted on the CUHK-PEDES datasets, we evaluate the effectiveness of proposed components and their combinations in improving text-based person search performance. In Table 2, we compare several configurations: The baseline method employs only global embeddings " + }, + { + "bbox": [ + 130, + 594, + 482, + 666 + ], + "type": "inline_equation", + "content": "\\mathbf{g}" + }, + { + "bbox": [ + 130, + 594, + 482, + 666 + ], + "type": "text", + "content": " trained by InfoNCE " + }, + { + "bbox": [ + 130, + 594, + 482, + 666 + ], + "type": "inline_equation", + "content": "(\\mathcal{L}_{\\mathrm{NCE}})" + }, + { + "bbox": [ + 130, + 594, + 482, + 666 + ], + "type": "text", + "content": " loss. When incorporating the cross-modal masked language modeling " + }, + { + "bbox": [ + 130, + 594, + 482, + 666 + ], + "type": "inline_equation", + "content": "(\\mathcal{L}_{\\mathrm{CMLM}})" + }, + { + "bbox": [ + 130, + 594, + 482, + 666 + ], + "type": "text", + "content": ", there is a slight improvement across all metrics. The addition" + } + ] + } + ], + "index": 18 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "spans": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "type": "text", + "content": "12" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 166, + 91, + 220, + 100 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 166, + 91, + 220, + 100 + ], + "spans": [ + { + "bbox": [ + 166, + 91, + 220, + 100 + ], + "type": "text", + "content": "J. Park et al." + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 11 + }, + { + "para_blocks": [ + { + "bbox": [ + 154, + 116, + 457, + 138 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 154, + 116, + 457, + 138 + ], + "spans": [ + { + "bbox": [ + 154, + 116, + 457, + 138 + ], + "type": "text", + "content": "Text: the woman is wearing a white cardigan over a long floaty dress and black shoes. she is carrying a back carry-on bag." + } + ] + } + ], + "index": 2 + }, + { + "type": "image", + "bbox": [ + 157, + 139, + 473, + 334 + ], + "blocks": [ + { + "bbox": [ + 157, + 139, + 473, + 334 + ], + "lines": [ + { + "bbox": [ + 157, + 139, + 473, + 334 + ], + "spans": [ + { + "bbox": [ + 157, + 139, + 473, + 334 + ], + "type": "image", + "image_path": "a54dd864b9e7b3c752ebe3dc49e32c81f39056b1f34bd7903632f0d11e86113a.jpg" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_body" + } + ], + "index": 3 + }, + { + "bbox": [ + 155, + 338, + 460, + 361 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 155, + 338, + 460, + 361 + ], + "spans": [ + { + "bbox": [ + 155, + 338, + 460, + 361 + ], + "type": "text", + "content": "Text: the little boy is wearing his hat backwards. he has on a blue, short sleeved shirt and grey plaid shorts." + } + ] + } + ], + "index": 4 + }, + { + "type": "image", + "bbox": [ + 154, + 364, + 473, + 566 + ], + "blocks": [ + { + "bbox": [ + 154, + 364, + 473, + 566 + ], + "lines": [ + { + "bbox": [ + 154, + 364, + 473, + 566 + ], + "spans": [ + { + "bbox": [ + 154, + 364, + 473, + 566 + ], + "type": "image", + "image_path": "2cb3e51c334281e25db0501e542484791a9ea7dc23a36fbe744dded371a898f7.jpg" + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 130, + 567, + 482, + 589 + ], + "lines": [ + { + "bbox": [ + 130, + 567, + 482, + 589 + ], + "spans": [ + { + "bbox": [ + 130, + 567, + 482, + 589 + ], + "type": "text", + "content": "Fig. 4: Visualization of each modality's attention map " + }, + { + "bbox": [ + 130, + 567, + 482, + 589 + ], + "type": "inline_equation", + "content": "\\bar{A}_k" + }, + { + "bbox": [ + 130, + 567, + 482, + 589 + ], + "type": "text", + "content": " in " + }, + { + "bbox": [ + 130, + 567, + 482, + 589 + ], + "type": "inline_equation", + "content": "T" + }, + { + "bbox": [ + 130, + 567, + 482, + 589 + ], + "type": "text", + "content": "-th iteration of PSA block and TDPA weights a on CUHK-PEDES dataset." + } + ] + } + ], + "index": 6, + "angle": 0, + "type": "image_caption" + } + ], + "index": 5 + }, + { + "bbox": [ + 130, + 617, + 483, + 666 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 617, + 483, + 666 + ], + "spans": [ + { + "bbox": [ + 130, + 617, + 483, + 666 + ], + "type": "text", + "content": "of part embeddings " + }, + { + "bbox": [ + 130, + 617, + 483, + 666 + ], + "type": "inline_equation", + "content": "\\mathbf{P}" + }, + { + "bbox": [ + 130, + 617, + 483, + 666 + ], + "type": "text", + "content": " that without " + }, + { + "bbox": [ + 130, + 617, + 483, + 666 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_{\\mathrm{PartID}}" + }, + { + "bbox": [ + 130, + 617, + 483, + 666 + ], + "type": "text", + "content": " also shows significant enhancement in R@1 and R@5, where we observe an increase of " + }, + { + "bbox": [ + 130, + 617, + 483, + 666 + ], + "type": "inline_equation", + "content": "3.46\\% \\mathrm{p}" + }, + { + "bbox": [ + 130, + 617, + 483, + 666 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 130, + 617, + 483, + 666 + ], + "type": "inline_equation", + "content": "2.64\\% \\mathrm{p}" + }, + { + "bbox": [ + 130, + 617, + 483, + 666 + ], + "type": "text", + "content": ", respectively. Notably, the full configuration achieves the best performance with " + }, + { + "bbox": [ + 130, + 617, + 483, + 666 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_{\\mathrm{PartID}}" + }, + { + "bbox": [ + 130, + 617, + 483, + 666 + ], + "type": "text", + "content": ", underscoring the importance of part embeddings for accurate text-based person search." + } + ] + } + ], + "index": 7 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 148, + 91, + 447, + 102 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 148, + 91, + 447, + 102 + ], + "spans": [ + { + "bbox": [ + 148, + 91, + 447, + 102 + ], + "type": "text", + "content": "PLOT: TBPS with Part Slot Attention for Corresponding Part Discovery" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 470, + 91, + 480, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 470, + 91, + 480, + 100 + ], + "spans": [ + { + "bbox": [ + 470, + 91, + 480, + 100 + ], + "type": "text", + "content": "13" + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 12 + }, + { + "para_blocks": [ + { + "type": "table", + "bbox": [ + 134, + 135, + 480, + 219 + ], + "blocks": [ + { + "bbox": [ + 185, + 114, + 427, + 125 + ], + "lines": [ + { + "bbox": [ + 185, + 114, + 427, + 125 + ], + "spans": [ + { + "bbox": [ + 185, + 114, + 427, + 125 + ], + "type": "text", + "content": "Table 2: Ablation studies on the CUHK-PEDES datasets." + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 134, + 135, + 480, + 219 + ], + "lines": [ + { + "bbox": [ + 134, + 135, + 480, + 219 + ], + "spans": [ + { + "bbox": [ + 134, + 135, + 480, + 219 + ], + "type": "table", + "html": "
MethodLossCUHK-PEDES
LNCELIDLCMLMLPartNCELPartIDR@1R@5R@10
Global Only✓XXXX71.39 (+0.0)87.65 (+0.0)92.74 (+0.0)
✓✓XXX71.83 (+0.44)88.06 (+0.41)92.58 (-0.16)
+ CMLM✓✓✓XX72.65 (+1.26)88.58 (+0.93)92.93 (+0.19)
+ Part Embeddings✓✓✓✓X74.85 (+3.46)90.29 (+2.64)94.10 (+1.36)
✓✓✓✓✓75.28 (+3.89)90.42 (+2.77)94.12 (+1.38)
", + "image_path": "d2b3e36dac03003270bb7594c1125c67cd3f9111eeaba7167825cd1d030e2c71.jpg" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "table_body" + } + ], + "index": 3 + }, + { + "bbox": [ + 130, + 241, + 479, + 264 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 241, + 479, + 264 + ], + "spans": [ + { + "bbox": [ + 130, + 241, + 479, + 264 + ], + "type": "text", + "content": "This suggests that our part embeddings significantly contribute to discerning fine-grained details critical for text-based person search." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 130, + 265, + 307, + 360 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 265, + 307, + 360 + ], + "spans": [ + { + "bbox": [ + 130, + 265, + 307, + 360 + ], + "type": "text", + "content": "Ablation Study on Part Discovery Methods: To validate the effectiveness of our proposed part discovery module, we conduct experiments by replacing our part discovery module with other methods such as TIPCB [2] and PAT [14]. TIPCB is a simple yet effective heuristic part discovery method, which extracts" + } + ] + } + ], + "index": 5 + }, + { + "type": "table", + "bbox": [ + 313, + 290, + 483, + 352 + ], + "blocks": [ + { + "bbox": [ + 313, + 267, + 482, + 289 + ], + "lines": [ + { + "bbox": [ + 313, + 267, + 482, + 289 + ], + "spans": [ + { + "bbox": [ + 313, + 267, + 482, + 289 + ], + "type": "text", + "content": "Table 3: Ablation study of different part discovery methods on the CUHK-PEDES." + } + ] + } + ], + "index": 6, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 313, + 290, + 483, + 352 + ], + "lines": [ + { + "bbox": [ + 313, + 290, + 483, + 352 + ], + "spans": [ + { + "bbox": [ + 313, + 290, + 483, + 352 + ], + "type": "table", + "html": "
MethodsR@1R@5R@10
Ours + TIPCB [2]73.2389.1094.04
Ours + PAT [14]72.7689.2393.42
Ours75.2890.4294.12
", + "image_path": "7396b1de6649c605a9f245be02236adbc6b6f692bd1416fd2a2671fbbda94395.jpg" + } + ] + } + ], + "index": 7, + "angle": 0, + "type": "table_body" + } + ], + "index": 7 + }, + { + "bbox": [ + 130, + 361, + 482, + 529 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 361, + 482, + 529 + ], + "spans": [ + { + "bbox": [ + 130, + 361, + 482, + 529 + ], + "type": "text", + "content": "image part embeddings through equi-horizontal cropping of the entire image and transforms the global embedding of the text modality into corresponding parts by using additional learnable MLPs for each image part. PAT performs part discovery with a querying transformer approach by leveraging learnable queries and conventional cross-attention. A key difference of our method from PAT is that ours explicitly encourages the discovered parts to be spatially separated since our part slots compete with each other to aggregate input data. In contrast, parts found by PAT, based on conventional cross-attention, often over-capture salient regions and are likely to miss fine details. To demonstrate this limitation, we show the visualization of part attentions with PAT applied to ours in the supplementary materials. As indicated in Table 3, our part discovery method outperforms these heuristic and conventional cross-attention based part discovery methods across all metrics, with the most significant difference observed in the R@1 metric, which requires a precise discernment of human parts." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 132, + 545, + 220, + 558 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 545, + 220, + 558 + ], + "spans": [ + { + "bbox": [ + 132, + 545, + 220, + 558 + ], + "type": "text", + "content": "5 Conclusion" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 130, + 570, + 482, + 629 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 570, + 482, + 629 + ], + "spans": [ + { + "bbox": [ + 130, + 570, + 482, + 629 + ], + "type": "text", + "content": "We proposed a novel framework that extracts distinctive human parts corresponding across visual and textual modalities through part discovery module without part-level supervision. The introduced TDPA further refines the retrieval process by adjusting the importance of each part slot based on the contents of the text query, leading to more precise and relevant retrieval." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 130, + 630, + 482, + 665 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 630, + 482, + 665 + ], + "spans": [ + { + "bbox": [ + 130, + 630, + 482, + 665 + ], + "type": "text", + "content": "Limitation: Since slots learn to occupy the entire image and text, some of them may indicate irrelevant part of input. TDPA addresses this issue to some extent, but a more explicit solution would further enhance our system." + } + ] + } + ], + "index": 11 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "spans": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "type": "text", + "content": "14" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 166, + 91, + 220, + 100 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 166, + 91, + 220, + 100 + ], + "spans": [ + { + "bbox": [ + 166, + 91, + 220, + 100 + ], + "type": "text", + "content": "J. Park et al." + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 13 + }, + { + "para_blocks": [ + { + "bbox": [ + 133, + 115, + 246, + 129 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 115, + 246, + 129 + ], + "spans": [ + { + "bbox": [ + 133, + 115, + 246, + 129 + ], + "type": "text", + "content": "Acknowledgements" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 130, + 140, + 485, + 175 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 140, + 485, + 175 + ], + "spans": [ + { + "bbox": [ + 130, + 140, + 485, + 175 + ], + "type": "text", + "content": "This work was supported by the IITP grants and the NRF grants funded by Ministry of Science and ICT, Korea (RS-2019-II191906; RS-2022-II220926; NRF-2018R1A5A1060031; NRF-2021R1A2C3012728)" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 133, + 193, + 197, + 205 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 193, + 197, + 205 + ], + "spans": [ + { + "bbox": [ + 133, + 193, + 197, + 205 + ], + "type": "text", + "content": "References" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 134, + 217, + 481, + 665 + ], + "type": "list", + "angle": 0, + "index": 19, + "blocks": [ + { + "bbox": [ + 138, + 217, + 481, + 251 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 217, + 481, + 251 + ], + "spans": [ + { + "bbox": [ + 138, + 217, + 481, + 251 + ], + "type": "text", + "content": "1. Aggarwal, S., Radhakrishnan, V.B., Chakraborty, A.: Text-based person search via attribute-aided matching. In: Proc. Winter Conference on Applications of Computer Vision (WACV) (2020)" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 138, + 251, + 481, + 282 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 251, + 481, + 282 + ], + "spans": [ + { + "bbox": [ + 138, + 251, + 481, + 282 + ], + "type": "text", + "content": "2. Chen, Y., Zhang, G., Lu, Y., Wang, Z., Zheng, Y.: Tipcb: A simple but effective part-based convolutional baseline for text-based person search. Neurocomputing 494 (2022)" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 138, + 284, + 481, + 316 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 284, + 481, + 316 + ], + "spans": [ + { + "bbox": [ + 138, + 284, + 481, + 316 + ], + "type": "text", + "content": "3. Devlin, J., Chang, M.W., Lee, K., Toutanova, K.: Bert: Pre-training of deep bidirectional transformers for language understanding. In: Annual Conference of the North American Chapter of the Association for Computational Linguistics (2019)" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 138, + 317, + 481, + 338 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 317, + 481, + 338 + ], + "spans": [ + { + "bbox": [ + 138, + 317, + 481, + 338 + ], + "type": "text", + "content": "4. Ding, Z., Ding, C., Shao, Z., Tao, D.: Semantically self-aligned network for text-to-image part-aware person re-identification. arXiv preprint arXiv:2107.12666 (2021)" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 138, + 338, + 481, + 370 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 338, + 481, + 370 + ], + "spans": [ + { + "bbox": [ + 138, + 338, + 481, + 370 + ], + "type": "text", + "content": "5. Gao, C., Cai, G., Jiang, X., Zheng, F., Zhang, J., Gong, Y., Peng, P., Guo, X., Sun, X.: Contextual non-local alignment over full-scale representation for text-based person search. arXiv preprint arXiv:2101.03036 (2021)" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 138, + 372, + 481, + 403 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 372, + 481, + 403 + ], + "spans": [ + { + "bbox": [ + 138, + 372, + 481, + 403 + ], + "type": "text", + "content": "6. Gray, D., Brennan, S., Tao, H.: Evaluating appearance models for recognition, reacquisition, and tracking. In: Proc. IEEE international workshop on performance evaluation for tracking and surveillance (PETS) (2007)" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 138, + 404, + 481, + 435 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 404, + 481, + 435 + ], + "spans": [ + { + "bbox": [ + 138, + 404, + 481, + 435 + ], + "type": "text", + "content": "7. Jiang, D., Ye, M.: Cross-modal implicit relation reasoning and aligning for text-to-image person retrieval. In: IEEE International Conference on Computer Vision and Pattern Recognition (CVPR) (2023)" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 138, + 437, + 481, + 468 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 437, + 481, + 468 + ], + "spans": [ + { + "bbox": [ + 138, + 437, + 481, + 468 + ], + "type": "text", + "content": "8. Jing, Y., Si, C., Wang, J., Wang, W., Wang, L., Tan, T.: Pose-guided multi-granularity attention network for text-based person search. In: Proc. AAAI Conference on Artificial Intelligence (AAAI) (2020)" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 138, + 470, + 481, + 491 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 470, + 481, + 491 + ], + "spans": [ + { + "bbox": [ + 138, + 470, + 481, + 491 + ], + "type": "text", + "content": "9. Lee, K.H., Chen, X., Hua, G., Hu, H., He, X.: Stacked cross attention for image-text matching. In: Proc. European Conference on Computer Vision (ECCV) (2018)" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 134, + 491, + 481, + 523 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 134, + 491, + 481, + 523 + ], + "spans": [ + { + "bbox": [ + 134, + 491, + 481, + 523 + ], + "type": "text", + "content": "10. Li, S., Cao, M., Zhang, M.: Learning semantic-aligned feature representation for text-based person search. In: Proc. IEEE International Conference on Acoustics, Speech and Signal Processing (ICASSP). IEEE (2022)" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 134, + 524, + 481, + 555 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 134, + 524, + 481, + 555 + ], + "spans": [ + { + "bbox": [ + 134, + 524, + 481, + 555 + ], + "type": "text", + "content": "11. Li, S., Xiao, T., Li, H., Zhou, B., Yue, D., Wang, X.: Person search with natural language description. In: Proc. IEEE Conference on Computer Vision and Pattern Recognition (CVPR) (2017)" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 134, + 556, + 481, + 599 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 134, + 556, + 481, + 599 + ], + "spans": [ + { + "bbox": [ + 134, + 556, + 481, + 599 + ], + "type": "text", + "content": "12. Li, W., Zhao, R., Wang, X.: Human reidentification with transferred metric learning. In: Computer Vision-ACCV 2012: 11th Asian Conference on Computer Vision, Daejeon, Korea, November 5-9, 2012, Revised Selected Papers, Part I 11. Springer (2013)" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 134, + 600, + 481, + 632 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 134, + 600, + 481, + 632 + ], + "spans": [ + { + "bbox": [ + 134, + 600, + 481, + 632 + ], + "type": "text", + "content": "13. Li, W., Zhao, R., Xiao, T., Wang, X.: Deepreid: Deep filter pairing neural network for person re-identification. In: Proc. IEEE Conference on Computer Vision and Pattern Recognition (CVPR) (2014)" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 134, + 633, + 481, + 665 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 134, + 633, + 481, + 665 + ], + "spans": [ + { + "bbox": [ + 134, + 633, + 481, + 665 + ], + "type": "text", + "content": "14. Li, Y., He, J., Zhang, T., Liu, X., Zhang, Y., Wu, F.: Diverse part discovery: Occluded person re-identification with part-aware transformer. In: Proc. IEEE Conference on Computer Vision and Pattern Recognition (CVPR) (2021)" + } + ] + } + ], + "index": 18 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 148, + 91, + 447, + 102 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 148, + 91, + 447, + 102 + ], + "spans": [ + { + "bbox": [ + 148, + 91, + 447, + 102 + ], + "type": "text", + "content": "PLOT: TBPS with Part Slot Attention for Corresponding Part Discovery" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 470, + 92, + 480, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 470, + 92, + 480, + 100 + ], + "spans": [ + { + "bbox": [ + 470, + 92, + 480, + 100 + ], + "type": "text", + "content": "15" + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 14 + }, + { + "para_blocks": [ + { + "bbox": [ + 132, + 117, + 481, + 665 + ], + "type": "list", + "angle": 0, + "index": 20, + "blocks": [ + { + "bbox": [ + 133, + 117, + 481, + 138 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 117, + 481, + 138 + ], + "spans": [ + { + "bbox": [ + 133, + 117, + 481, + 138 + ], + "type": "text", + "content": "15. Locatello, F., Weissenborn, D., Unterthiner, T., Mahendran, A., Heigold, G., Uszkoreit, J., Dosovitskiy, A., Kipf, T.: Object-centric learning with slot attention (2020)" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 133, + 140, + 481, + 171 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 140, + 481, + 171 + ], + "spans": [ + { + "bbox": [ + 133, + 140, + 481, + 171 + ], + "type": "text", + "content": "16. Niu, K., Huang, Y., Ouyang, W., Wang, L.: Improving description-based person re-identification by multi-granularity image-text alignments. IEEE Transactions on Image Processing 29, 5542-5556 (2020)" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 133, + 172, + 481, + 193 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 172, + 481, + 193 + ], + "spans": [ + { + "bbox": [ + 133, + 172, + 481, + 193 + ], + "type": "text", + "content": "17. Oord, A.v.d., Li, Y., Vinyals, O.: Representation learning with contrastive predictive coding. arXiv preprint arXiv:1807.03748 (2018)" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 133, + 194, + 481, + 237 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 194, + 481, + 237 + ], + "spans": [ + { + "bbox": [ + 133, + 194, + 481, + 237 + ], + "type": "text", + "content": "18. Radford, A., Kim, J.W., Hallacy, C., Ramesh, A., Goh, G., Agarwal, S., Sastry, G., Askell, A., Mishkin, P., Clark, J., et al.: Learning transferable visual models from natural language supervision. In: Proc. International Conference on Machine Learning (ICML) (2021)" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 133, + 237, + 481, + 270 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 237, + 481, + 270 + ], + "spans": [ + { + "bbox": [ + 133, + 237, + 481, + 270 + ], + "type": "text", + "content": "19. Sarafianos, N., Xu, X., Kakadiaris, I.A.: Adversarial representation learning for text-to-image matching. In: Proc. IEEE International Conference on Computer Vision (ICCV) (2019)" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 132, + 270, + 481, + 303 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 270, + 481, + 303 + ], + "spans": [ + { + "bbox": [ + 132, + 270, + 481, + 303 + ], + "type": "text", + "content": "20. Shao, Z., Zhang, X., Fang, M., Lin, Z., Wang, J., Ding, C.: Learning granularity-unified representations for text-to-image person re-identification. In: Proc. ACM Multimedia Conference (ACMMM) (2022)" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 132, + 303, + 481, + 346 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 303, + 481, + 346 + ], + "spans": [ + { + "bbox": [ + 132, + 303, + 481, + 346 + ], + "type": "text", + "content": "21. Shu, X., Wen, W., Wu, H., Chen, K., Song, Y., Qiao, R., Ren, B., Wang, X.: See finer, see more: Implicit modality alignment for text-based person retrieval. In: Proc. European Conference on Computer Vision Workshop on Real-World Surveillance, (ECCVW) (2022)" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 132, + 347, + 481, + 380 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 347, + 481, + 380 + ], + "spans": [ + { + "bbox": [ + 132, + 347, + 481, + 380 + ], + "type": "text", + "content": "22. Suo, W., Sun, M., Niu, K., Gao, Y., Wang, P., Zhang, Y., Wu, Q.: A simple and robust correlation filtering method for text-based person search. In: Proc. European Conference on Computer Vision (ECCV) (2022)" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 132, + 380, + 481, + 413 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 380, + 481, + 413 + ], + "spans": [ + { + "bbox": [ + 132, + 380, + 481, + 413 + ], + "type": "text", + "content": "23. Wang, C., Luo, Z., Lin, Y., Li, S.: Text-based person search via multi-granularity embedding learning. In: Proc. International Joint Conferences on Artificial Intelligence (IJCAI) (2021)" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 132, + 413, + 481, + 445 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 413, + 481, + 445 + ], + "spans": [ + { + "bbox": [ + 132, + 413, + 481, + 445 + ], + "type": "text", + "content": "24. Wang, Z., Fang, Z., Wang, J., Yang, Y.: Vitaa: Visual-textual attributes alignment in person search by natural language. In: Proc. European Conference on Computer Vision (ECCV) (2020)" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 132, + 446, + 481, + 479 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 446, + 481, + 479 + ], + "spans": [ + { + "bbox": [ + 132, + 446, + 481, + 479 + ], + "type": "text", + "content": "25. Wang, Z., Zhu, A., Xue, J., Wan, X., Liu, C., Wang, T., Li, Y.: Caibc: Capturing all-round information beyond color for text-based person retrieval. In: Proc. ACM Multimedia Conference (ACMMM) (2022)" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 132, + 479, + 481, + 511 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 479, + 481, + 511 + ], + "spans": [ + { + "bbox": [ + 132, + 479, + 481, + 511 + ], + "type": "text", + "content": "26. Wei, L., Zhang, S., Gao, W., Tian, Q.: Person transfer gan to bridge domain gap for person re-identification. In: Proc. IEEE Conference on Computer Vision and Pattern Recognition (CVPR) (2018)" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 132, + 512, + 481, + 544 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 512, + 481, + 544 + ], + "spans": [ + { + "bbox": [ + 132, + 512, + 481, + 544 + ], + "type": "text", + "content": "27. Wu, Y., Yan, Z., Han, X., Li, G., Zou, C., Cui, S.: Lipscore: language-guided person search via color reasoning. In: Proc. IEEE International Conference on Computer Vision (ICCV) (2021)" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 132, + 545, + 481, + 567 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 545, + 481, + 567 + ], + "spans": [ + { + "bbox": [ + 132, + 545, + 481, + 567 + ], + "type": "text", + "content": "28. Xiao, T., Li, S., Wang, B., Lin, L., Wang, X.: End-to-end deep learning for person search. arXiv preprint arXiv:1604.01850 (2016)" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 132, + 567, + 481, + 588 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 567, + 481, + 588 + ], + "spans": [ + { + "bbox": [ + 132, + 567, + 481, + 588 + ], + "type": "text", + "content": "29. Yan, S., Dong, N., Zhang, L., Tang, J.: Clip-driven fine-grained text-image person re-identification. IEEE Transactions on Image Processing (2023)" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 132, + 589, + 481, + 610 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 589, + 481, + 610 + ], + "spans": [ + { + "bbox": [ + 132, + 589, + 481, + 610 + ], + "type": "text", + "content": "30. Zhang, Y., Lu, H.: Deep cross-modal projection learning for image-text matching. In: Proc. European Conference on Computer Vision (ECCV) (2018)" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 132, + 611, + 481, + 643 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 611, + 481, + 643 + ], + "spans": [ + { + "bbox": [ + 132, + 611, + 481, + 643 + ], + "type": "text", + "content": "31. Zheng, L., Shen, L., Tian, L., Wang, S., Wang, J., Tian, Q.: Scalable person re-identification: A benchmark. In: Proc. IEEE International Conference on Computer Vision (ICCV) (2015)" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 132, + 643, + 481, + 665 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 643, + 481, + 665 + ], + "spans": [ + { + "bbox": [ + 132, + 643, + 481, + 665 + ], + "type": "text", + "content": "32. Zhu, A., Wang, Z., Li, Y., Wan, X., Jin, J., Wang, T., Hu, F., Hua, G.: Dssl: deep surroundings-person separation learning for text-based person retrieval. In:" + } + ] + } + ], + "index": 19 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "spans": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "type": "text", + "content": "16" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 166, + 91, + 220, + 100 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 166, + 91, + 220, + 100 + ], + "spans": [ + { + "bbox": [ + 166, + 91, + 220, + 100 + ], + "type": "text", + "content": "J. Park et al." + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 15 + }, + { + "para_blocks": [ + { + "bbox": [ + 148, + 117, + 482, + 138 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 148, + 117, + 482, + 138 + ], + "spans": [ + { + "bbox": [ + 148, + 117, + 482, + 138 + ], + "type": "text", + "content": "Proceedings of the 29th ACM International Conference on Multimedia. pp. 209-217 (2021)" + } + ] + } + ], + "index": 2 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 148, + 91, + 447, + 102 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 148, + 91, + 447, + 102 + ], + "spans": [ + { + "bbox": [ + 148, + 91, + 447, + 102 + ], + "type": "text", + "content": "PLOT: TBPS with Part Slot Attention for Corresponding Part Discovery" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 470, + 92, + 480, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 470, + 92, + 480, + 100 + ], + "spans": [ + { + "bbox": [ + 470, + 92, + 480, + 100 + ], + "type": "text", + "content": "17" + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 16 + } + ], + "_backend": "vlm", + "_version_name": "2.6.4" +} \ No newline at end of file diff --git a/2024/PMT_ Progressive Mean Teacher via Exploring Temporal Consistency for Semi-Supervised Medical Image Segmentation/049d7a98-0307-4aa9-ae68-729601a51f9f_content_list.json b/2024/PMT_ Progressive Mean Teacher via Exploring Temporal Consistency for Semi-Supervised Medical Image Segmentation/049d7a98-0307-4aa9-ae68-729601a51f9f_content_list.json new file mode 100644 index 0000000000000000000000000000000000000000..0bdefa16802ca6118523ca6582c1ffeb2b753e37 --- /dev/null +++ b/2024/PMT_ Progressive Mean Teacher via Exploring Temporal Consistency for Semi-Supervised Medical Image Segmentation/049d7a98-0307-4aa9-ae68-729601a51f9f_content_list.json @@ -0,0 +1,1611 @@ +[ + { + "type": "text", + "text": "PMT: Progressive Mean Teacher via Exploring Temporal Consistency for Semi-Supervised Medical Image Segmentation", + "text_level": 1, + "bbox": [ + 233, + 140, + 772, + 209 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Ning Gao $^{\\text{D}}$ , Sanping Zhou $^{*}$ , Le Wang $^{\\text{D}}$ , and Nanning Zheng $^{\\text{D}}$", + "bbox": [ + 269, + 234, + 730, + 252 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "National Key Laboratory of Human-Machine Hybrid Augmented Intelligence, National Engineering Research Center for Visual Information and Applications, Institute of Artificial Intelligence and Robotics, Xi'an Jiaotong University", + "bbox": [ + 235, + 262, + 764, + 305 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Abstract. Semi-supervised learning has emerged as a widely adopted technique in the field of medical image segmentation. The existing works either focuses on the construction of consistency constraints or the generation of pseudo labels to provide high-quality supervisory signals, whose main challenge mainly comes from how to keep the continuous improvement of model capabilities. In this paper, we propose a simple yet effective semi-supervised learning framework, termed Progressive Mean Teachers (PMT), for medical image segmentation, whose goal is to generate high-fidelity pseudo labels by learning robust and diverse features in the training process. Specifically, our PMT employs a standard mean teacher to penalize the consistency of the current state and utilizes two sets of MT architectures for co-training. The two sets of MT architectures are individually updated for prolonged periods to maintain stable model diversity established through performance gaps generated by iteration differences. Additionally, a difference-driven alignment regularizer is employed to expedite the alignment of lagging models with the representation capabilities of leading models. Furthermore, a simple yet effective pseudo-label filtering algorithm is employed for facile evaluation of models and selection of high-fidelity pseudo-labels outputted when models are operating at high performance for co-training purposes. Experimental results on two datasets with different modalities, i.e., CT and MRI, demonstrate that our method outperforms the state-of-the-art medical image segmentation approaches across various dimensions. The code is available at https://github.com/Axi404/PMT.", + "bbox": [ + 259, + 340, + 743, + 674 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Keywords: Semi-supervised learning $\\cdot$ Medical image segmentation $\\cdot$ Temporal consistency regularization", + "bbox": [ + 259, + 686, + 740, + 715 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "1 Introduction", + "text_level": 1, + "bbox": [ + 215, + 739, + 375, + 757 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Semi-supervised learning is an important field in deep learning, which offers an effective way to tackle problems with limited labeled data [16, 25, 27, 32, 33]. With the continuous emergence of large-scale data, semi-supervised learning has", + "bbox": [ + 212, + 771, + 787, + 816 + ], + "page_idx": 0 + }, + { + "type": "page_footnote", + "text": "* Corresponding author.", + "bbox": [ + 217, + 824, + 385, + 840 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/062538e6c64842cb76ba27d2a32d430a0131c159211272c07e580bf7119e48bf.jpg", + "image_caption": [ + "Fig. 1: Motivation of our PMT. In particular, the standard MT helps to learn robust features by keeping the consistency between teacher and student networks at the current iteration, while our PMT further helps to learn diverse features by maintaining the difference between student networks at different iterations. As a result, more and more high-fidelity pseudo labels will be generated for semi-supervised medical image segmentation." + ], + "image_footnote": [], + "bbox": [ + 320, + 146, + 681, + 327 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "become a hot topic in both machine learning and computer vision communities. It has gained particular attention in the medical image segmentation [21], due to the need for expert annotation in determining the accurate boundaries of targets. As a result, the labeled data is very lacking in medical image segmentation, which makes the semi-supervised learning very popular in this domain.", + "bbox": [ + 212, + 459, + 787, + 535 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Most of the semi-supervised learning methods involve seeking a guidance mechanism using limited annotated data, so as to maximize the usage of unlabeled data to train the models on different downstream tasks. In general, two technologies, i.e., consistency regularization [9, 10] and pseudo label generation [10, 19], are commonly used in semi-supervised medical segmentation. The former ones aim at enhancing the representation capability of networks by exploring the consistency based on different prior assumptions [5], while the latter ones focus on improving the performance on downstream tasks by generating the pseudo labels based on current model capability. For example, the well-known Mean Teacher (MT) [19] utilizes an ingenious Exponential Moving Average (EMA) that is equivalent to data augmentation and enforces consistency regularization between the outputs of teacher and student models. Besides, the representative Noisy Student [26] designs a teacher model to generate pseudo labels, so as to train a larger student model.", + "bbox": [ + 212, + 536, + 787, + 747 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Even though the above methods have achieved significant improvements in semi-supervised medical image segmentation, we still argue that they are weak in exploring high-quality supervisory signals to consistently enhance the model's capability. To address this issue, an iterative unified optimization framework has been introduced to semi-supervised medical image segmentation [20, 23, 24], in which the consistency regularization strategy is taken to enhance the model's", + "bbox": [ + 212, + 750, + 787, + 840 + ], + "page_idx": 1 + }, + { + "type": "page_number", + "text": "2", + "bbox": [ + 217, + 114, + 228, + 126 + ], + "page_idx": 1 + }, + { + "type": "header", + "text": "N. Gao et al.", + "bbox": [ + 271, + 114, + 359, + 126 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "representation capability and the pseudo label generation algorithm is applied to generate high-fidelity pseudo labels. For example, the recent MCF [23] takes VNet and 3D ResNet for representation learning, and heterogeneous networks for dynamic pseudo label generation. The challenges to these methods lie on how to continuously generate diverse pseudo labels in the forward propagation, and enhance model's capability in the backward propagation.", + "bbox": [ + 212, + 146, + 782, + 236 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "In this paper, we design a novel semi-supervised learning framework, termed Progressive Mean Teachers (PMT), for medical image segmentation, whose main idea focuses on how to obtain a diverse set of accurate pseudo labels. Inspired by the positive relationship between iteration and performance, we try to maintain the diversity of networks by exploring their states at different training epochs. Specifically, the standard MT is first taken as basic architecture to learn the parameters of network at each iteration, which can enhance the network's representation capability by using the EMA data augmentation. Our PMT further explores network diversity during training, alternating between two homogeneous MT architectures trained on the same dataset, a process we term progressive design. These models exhibit significant iteration leads due to alternating continuous individual updates, establishing performance gaps between networks at different epochs, as shown in Fig. 1. Consequently, the student network can acquire robust yet diverse features for medical image segmentation. Then, the Discrepancy Driven Alignment (DDA) regularizer is further designed to examine disparities between predictions obtained by the student network at different training epochs, facilitating rapid alignment to high-fidelity generated images. Finally, we design a simple Pseudo Label Filtering (PLF) algorithm to refine the basic interaction process, enabling the retention of high-fidelity pseudo-labels for training by comparing student network performance across different training epochs. As a result, more and more high-fidelity pseudo labels can be fed to train the other student networks, which will in turn generate more accurate predictions for pseudo label generation.", + "bbox": [ + 212, + 238, + 787, + 584 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "In summary, the main contributions of this work are as follows: (1) We design a novel Progressive Mean Teacher framework for semi-supervised medical image segmentation. (2) We design a novel Discrepancy Driven Alignment regularizer to rapidly align the representational capacity gap between lagging and leading networks. (3) We design a simple yet effective Pseudo Label Filtering algorithm to select high-fidelity pseudo labels. Extensive experiments in both Left Atrial [28] and Pancreas-NIH [17] datasets show that our PMT can achieve the state-of-the-art results in semi-supervised medical image segmentation.", + "bbox": [ + 212, + 585, + 787, + 705 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "2 Related Work", + "text_level": 1, + "bbox": [ + 214, + 732, + 387, + 748 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "2.1 Consistency Regularization", + "text_level": 1, + "bbox": [ + 214, + 767, + 486, + 781 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Consistency regularization is often employed in semi-supervised learning, so as to enhance the stable representational capacity of model. The behind idea is to preserve the invariance of predictions made by the same model when facing", + "bbox": [ + 212, + 794, + 785, + 839 + ], + "page_idx": 2 + }, + { + "type": "header", + "text": "Progressive Mean Teacher", + "bbox": [ + 555, + 114, + 730, + 128 + ], + "page_idx": 2 + }, + { + "type": "page_number", + "text": "3", + "bbox": [ + 774, + 116, + 784, + 126 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "perturbations applied in different regions, by constraining the consistency between output results under different perturbations. For example, $\\varPi$ model [9] introduces image-level regularization by adding perturbations to images. The well-known MT [19] introduces parameter-level regularization by using EMA to constrain the outputs between teacher and student models. Thanks to its simplicity and effectiveness, more and more works are paying attention to improve the generalization ability by formulating different consistency regularizers. For example, SASSnet [11] focuses on the regularity of geometric shapes for target object classes within consistency regularization. Besides, CPCL [29] establishes regularization between supervised and unsupervised training within a cyclic framework. Furthermore, some later works have made progress in regularization at different task and model levels. For example, DTC [12] introduces task-level regularization, presenting a novel dual-task consistency semi-supervised framework. Besides, MCF [23] introduces model-level regularization by using heterogeneous models to constrain output consistency. Compared to previous work, our PMT seeks the cross-temporal regularization between different training periods, which can help learn diverse yet robust features for subsequent pseudo label generation.", + "bbox": [ + 212, + 146, + 787, + 402 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "2.2 Pseudo Label Generation", + "text_level": 1, + "bbox": [ + 215, + 424, + 472, + 436 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Pseudo label generation is often employed in semi-supervised learning, so as to enhance the discriminative ability of model. The behind idea is to train a prior model with the labeled data and then apply it to generate the pseudo labels for unlabeled data. It is widely accepted to categorize pseudo-label generation into two methods [7], direct generation focuses on selecting pseudo labels with higher confidence, while indirect generation explores methods to generate high-fidelity pseudo labels. For direct generation, work [10] uses a fixed threshold to choose high-confidence pseudo labels. SsaNet [22] employs a trust module to reevaluate pseudo labels. UA-MT [30] introduces uncertainty estimation to filter out unreliable pseudo labels. Co-BioNet [15] introduces a feedback network to measure the uncertainty and choose high-confidence predictions of different models. For indirect generation, Tri-Net [6] is proposed to use two subnetworks to generate pseudo labels for a third subnetwork. Work [1] improves pseudo label generation using Simple Linear Iterative Clustering (SLIC) algorithm. MCF [23] dynamically generates pseudo labels using a heterogeneous network and effectively addresses cognitive bias, while DeSCO [3] focuses on the spatial correlation of medical images and generates pseudo labels using orthogonal slices. Compared to previous work, our PMT emphasizes the enhancement of pseudo label quality in the temporal domain and generates diverse and robust pseudo labels across different training periods, significantly improving performance.", + "bbox": [ + 212, + 446, + 787, + 750 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "2.3 Multi-Model Framework", + "text_level": 1, + "bbox": [ + 215, + 770, + 465, + 785 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "The multi-model framework is often employed in semi-supervised learning, so as to enhance the model representation by acquiring multiple views or diversity. Its development aligns closely with consistency regularization. Here, we", + "bbox": [ + 212, + 794, + 785, + 839 + ], + "page_idx": 3 + }, + { + "type": "page_number", + "text": "4", + "bbox": [ + 217, + 114, + 228, + 126 + ], + "page_idx": 3 + }, + { + "type": "header", + "text": "N. Gao et al.", + "bbox": [ + 271, + 114, + 359, + 127 + ], + "page_idx": 3 + }, + { + "type": "image", + "img_path": "images/fdeffe67ee6516f4a7f3439e0d10a005ee9fc08aa83cf44a01c443edf12a99a2.jpg", + "image_caption": [ + "Fig. 2: An overview of PMT. We employed a progressive design and utilized the architecture of MT. The PMT framework maintains a data buffer of length $\\mathcal{B}$ for cros-temporal training. The total loss function $\\mathrm{L}_{\\mathrm{total}}$ for each network includes supervised losses $\\mathrm{L_{CE}}, \\mathrm{L_{DICE}}, \\mathrm{L_{aln}}$ , and unsupervised loss $\\mathrm{L_U}, \\mathrm{L_T}$ ." + ], + "image_footnote": [], + "bbox": [ + 222, + 146, + 782, + 349 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "focus on the evolution of model structures. For example, MT [19] introduces a Teacher model with significantly improved representation capabilities at a lower cost using EMA, establishing the MT architecture. Besides, CPC [8] utilizes the confidence vector of multi-model outputs as pseudo-labels for co-training, while CPS [4] employs one-hot labels, both considered as starting points for co-training. In recent years, efforts have been made to combine these approaches. For example, UCMT [18] employs two student models for co-training, simultaneously updating the same Mean Teacher using EMA, while Dual Teacher [14] alternates updates between two Teacher models using EMA with a single student model, thereby fostering diversity among Teacher models. Compared to previous work, our PMT utilizes two sets of Mean Teachers in a progressive training framework, and employ two student models to update teacher models independently. This enables our model to rapidly establish robust performance disparities across iteration gap and maintain stable diversity among models.", + "bbox": [ + 212, + 445, + 787, + 657 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "3 Method", + "text_level": 1, + "bbox": [ + 215, + 679, + 328, + 694 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "3.1 The Overall Process of PMT", + "text_level": 1, + "bbox": [ + 214, + 709, + 501, + 724 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "The training process of PMT is illustrated in Fig. 2. As previously mentioned, PMT can consist of multiple networks. For ease of explanation, the number of networks is assumed to be two by default in our work. In our approach, these networks share the same structure and can be denoted as $f_{i}(\\cdot) \\in \\mathcal{F}$ , where $\\mathcal{F}$ stands for the function space within which these networks reside. In a semi-supervised process, the training data includes a small number of labeled data denoted as $\\mathbf{D}_L = \\{(x_i^L, y_i^L)\\}_{i=1}^N$ , and a large amount of unlabeled data denoted", + "bbox": [ + 212, + 733, + 787, + 842 + ], + "page_idx": 4 + }, + { + "type": "header", + "text": "Progressive Mean Teacher", + "bbox": [ + 555, + 114, + 730, + 128 + ], + "page_idx": 4 + }, + { + "type": "page_number", + "text": "5", + "bbox": [ + 774, + 116, + 784, + 126 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "as $\\mathbf{D}_U = \\{(x_i^U)\\}_{i = N + 1}^{N + M}$ , where $N\\ll M$ , $x_{i}\\in \\mathbb{R}^{H\\times W\\times D}$ represents medical volumes, and $y_{i}\\in \\{0,1\\}^{H\\times W\\times D}$ represents ground truth labels. Batches of input data $\\mathbf{X}$ consist of an equal proportion of labeled data $(\\mathbf{X}^L,\\mathbf{Y}^L)$ and unlabeled data $\\mathbf{X}^U$ . These volumes are fed into $f_{i}(\\cdot)$ and $f_{i + 1}(\\cdot)$ :", + "bbox": [ + 212, + 143, + 787, + 209 + ], + "page_idx": 5 + }, + { + "type": "equation", + "text": "\n$$\n\\hat {\\mathbf {Y}} _ {i} = f _ {i} (\\mathbf {X}), \\quad \\hat {\\mathbf {Y}} _ {i + 1} = f _ {i + 1} (\\mathbf {X}). \\tag {1}\n$$\n", + "text_format": "latex", + "bbox": [ + 388, + 220, + 785, + 239 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "The output consists of volume predictions for both labeled and unlabeled data: $\\hat{\\mathbf{Y}} = \\hat{\\mathbf{Y}}^L\\cup \\hat{\\mathbf{Y}}^U$ . For the sake of simplicity, the network index subscripts are omitted here.", + "bbox": [ + 212, + 243, + 785, + 287 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "We have ingeniously reintroduced MT architecture into the framework of semi-supervised learning. The introduction serves the purpose that, while providing high-fidelity pseudo labels to the model, the model's architecture needs to support the stable improvement of its representational capacity, aiming for better performance, while maintaining stable diversity. The teacher network is structurally identical to the student network but does not actively participate in the training process. Instead, it updates all of its parameters through EMA and applies consistency regularization to the student model through its output. The parameter is updated as follows:", + "bbox": [ + 212, + 289, + 787, + 426 + ], + "page_idx": 5 + }, + { + "type": "equation", + "text": "\n$$\n\\theta_ {\\text {t e a c h e r}} = \\alpha \\theta_ {\\text {t e a c h e r}} + (1 - \\alpha) \\theta_ {\\text {s t u d e n t}}, \\tag {2}\n$$\n", + "text_format": "latex", + "bbox": [ + 372, + 441, + 784, + 455 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "where $\\theta_{\\mathrm{teacher}}$ represents the parameters of teacher network, $\\theta_{\\mathrm{student}}$ represents the parameters of student network, and $\\alpha$ is the EMA decay rate.", + "bbox": [ + 212, + 460, + 785, + 491 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "In practice, we propose a cross-temporal training approach, which involves introducing a phase shift among different models during iterations and training the most lagging model in terms of iterations. To ensure that each model can train across temporal, the number of iterations for the model recorded in the sequence of iterations is denoted as $\\mathcal{I}_i$ . Our model maintains a static iteration gap through the data buffer length $\\mathcal{B}$ :", + "bbox": [ + 212, + 491, + 787, + 580 + ], + "page_idx": 5 + }, + { + "type": "equation", + "text": "\n$$\n\\forall i \\in [ 1, n - 1 ], \\operatorname {L a r} (\\mathcal {I}, i) - \\operatorname {L a r} (\\mathcal {I}, i + 1) = \\frac {\\mathcal {B}}{(n - 1)}, \\tag {3}\n$$\n", + "text_format": "latex", + "bbox": [ + 323, + 593, + 785, + 625 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "where, $\\operatorname{Lar}(\\mathcal{I}, i)$ represents the $i$ -th largest number in $\\mathcal{I}$ .", + "bbox": [ + 212, + 628, + 622, + 643 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "During each training iteration, the model with the least advanced iteration progress, referred to as the Current Progressive Model (CPM), sequentially utilizes data from the buffer for training until it has advanced ahead of the model with the most advanced iteration progress $\\mathcal{B} / n$ times. Throughout the training process, the PLF screens out pseudo-labels generated by models whose performance is inferior to that of the CPM by measuring the performance gap between the CPM and other models, and learns from the remaining pseudo-labels. Simultaneously, the model computes the DDA with respect to a given input $\\mathbf{X}$ and, along with all other models, examines regions of inconsistent predictions among different networks. When the performance of the CPM lags behind that of other models, the DDA corrects the CPM, enabling it to quickly align its performance with models that outperform it. It is worth noting that if the CPM is the best-performing model for that iteration, it will not guide other models in reverse. It", + "bbox": [ + 212, + 643, + 787, + 840 + ], + "page_idx": 5 + }, + { + "type": "page_number", + "text": "6", + "bbox": [ + 217, + 114, + 228, + 126 + ], + "page_idx": 5 + }, + { + "type": "header", + "text": "N. Gao et al.", + "bbox": [ + 271, + 114, + 359, + 127 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/71cf79dd32c6620cdbcc9651461e85d067aaecdc4ff856c4fd342be0857c2784.jpg", + "image_caption": [ + "(a)" + ], + "image_footnote": [], + "bbox": [ + 336, + 150, + 496, + 280 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/137070e520425265a242a5cb371ba0abccefb748d358ce452af54722233ebb6e.jpg", + "image_caption": [ + "(b)", + "Fig. 3: Illustration of PLF and DDA. PLF is utilized for comparing the representational capacity of models, while DDA aligns model outputs by examining differences." + ], + "image_footnote": [], + "bbox": [ + 501, + 150, + 663, + 280 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "receives guidance during the training process and only guides others once they have completed training.", + "bbox": [ + 212, + 367, + 782, + 398 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "3.2 Pseudo Label Filtering", + "text_level": 1, + "bbox": [ + 214, + 425, + 452, + 441 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "To guide the learning of the current progressive model effectively, we introduce PLF, as shown in Fig. 3 (a). PLF involves a naive approximation where we consider the representational capacity of models during the supervised learning phase as a criterion for judging the representational capacity during the unsupervised learning phase to some extent. During the supervised learning phase, PLF entails inference by all models, and through combining labels, calculates the representational capacity of each model in the current iteration. The metrics for evaluating representational capacity vary with tasks; in this task, the Dice loss is regarded as a statistical measure of model performance. We utilize the representational capacity of the CPM as a threshold and filter out pseudo-labels generated by models with representational capacities below that of the CPM. The pseudo label's passage through the PLF is recorded as $\\mathcal{P}(f_i,f_{i + 1})$ , where a value of 1 denotes passage and 0 denotes exclusion. This aids in the model's better learning of pseudo-labels with high fidelity. Furthermore, to ensure that the pseudo-labels are accurate and well-defined, we apply a sharpening function to high-fidelity pseudo-labels:", + "bbox": [ + 212, + 455, + 787, + 698 + ], + "page_idx": 6 + }, + { + "type": "equation", + "text": "\n$$\n\\mathbf {Y} ^ {U} = \\frac {\\mathbf {P _ {s}} ^ {1 / T}}{\\mathbf {P _ {s}} ^ {1 / T} + (1 - \\mathbf {P _ {s}}) ^ {1 / T}}, \\tag {4}\n$$\n", + "text_format": "latex", + "bbox": [ + 395, + 712, + 785, + 751 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "where $\\mathbf{P_s}$ represents the output of the good student, and $T$ is a hyperparameter.", + "bbox": [ + 214, + 758, + 782, + 773 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "The resulting loss function $\\mathrm{L_U}$ is computed as MSE between the pseudo labels $\\mathbf{Y}^U$ and the model's predictions of CPM $\\hat{\\mathbf{Y}}_c^U$ :", + "bbox": [ + 214, + 773, + 785, + 805 + ], + "page_idx": 6 + }, + { + "type": "equation", + "text": "\n$$\n\\mathrm {L} _ {\\mathrm {U}} = \\operatorname {M S E} \\left(\\hat {\\mathbf {Y}} _ {c} ^ {U}, \\mathbf {Y} ^ {U}\\right). \\tag {5}\n$$\n", + "text_format": "latex", + "bbox": [ + 419, + 821, + 785, + 840 + ], + "page_idx": 6 + }, + { + "type": "header", + "text": "Progressive Mean Teacher", + "bbox": [ + 555, + 114, + 730, + 128 + ], + "page_idx": 6 + }, + { + "type": "page_number", + "text": "7", + "bbox": [ + 774, + 114, + 784, + 126 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "3.3 Discrepancy Driven Alignment", + "text_level": 1, + "bbox": [ + 215, + 146, + 517, + 162 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "We propose DDA to compare the outputs of multiple models and achieve rapid alignment between models, as shown in Fig. 3 (b). During the alternating process of progressive Learning, the performance of two groups of models alternately improves, providing stable pseudo-labels to each other. However, concerns often arise due to the disappearance of the performance gap between models caused by random factors. While the expectation is for alternating performance leadership between models, in practice, it may turn into one model consistently leading. Therefore, a fast and high-confidence alignment module is needed. In traditional consistency regularization, contrastive algorithms enforce consistency among the outputs of different models. However, this approach often raises concerns because we are uncertain whether this consistency regularization guides the models in the right direction. DDA focuses on the different parts of predictions among models, which often reflect the differences in representational capabilities between models. We aim to enhance the consistency regularization specifically on these parts to achieve alignment of representational capabilities between models. It is worth noting that we expect the models aligned by DDA to have representation capabilities surpassing CPM. Therefore, DDA will only take effect when PLF allows other models to provide pseudo-labels. The mask of different parts of predictions among different models can be obtained by taking the difference between the union and intersection of the binary outputs of softmax outputs $\\hat{\\mathbf{Y}}^L$ of multiple networks:", + "bbox": [ + 212, + 172, + 789, + 492 + ], + "page_idx": 7 + }, + { + "type": "equation", + "text": "\n$$\n\\mathcal {M} _ {\\text {d i f f}} = \\bigcup_ {i = 1} ^ {n} \\operatorname {B I N A} \\left(\\hat {\\mathbf {Y}} _ {i} ^ {L}\\right) - \\bigcap_ {i = 1} ^ {N} \\operatorname {B I N A} \\left(\\hat {\\mathbf {Y}} _ {i} ^ {L}\\right). \\tag {6}\n$$\n", + "text_format": "latex", + "bbox": [ + 356, + 503, + 787, + 546 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Subsequently, MSE of the model's predictions is computed and denoted as $\\mathcal{M}_{\\mathrm{dist}}$ :", + "bbox": [ + 215, + 551, + 789, + 584 + ], + "page_idx": 7 + }, + { + "type": "equation", + "text": "\n$$\n\\mathcal {M} _ {\\mathrm {d i s t}} = \\operatorname {M S E} \\left(\\hat {\\mathbf {Y}} _ {i} ^ {L}, \\mathbf {Y} ^ {L}\\right). \\tag {7}\n$$\n", + "text_format": "latex", + "bbox": [ + 410, + 595, + 785, + 621 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Next, the total count of elements in $\\mathcal{M}_{\\mathrm{diff}}$ is calculated, and $\\mathcal{M}_{\\mathrm{diff}}$ is used to mask $\\mathcal{M}_{\\mathrm{dist}}$ . The ratio of the sum of the masked $\\mathcal{M}_{\\mathrm{dist}}$ to the total count of elements in $\\mathcal{M}_{\\mathrm{diff}}$ is denoted as $\\mathrm{L_{aln}}$ , which characterizes the gap to the ground truth within the prediction differences between two models:", + "bbox": [ + 212, + 626, + 787, + 686 + ], + "page_idx": 7 + }, + { + "type": "equation", + "text": "\n$$\n\\mathrm {L} _ {\\mathrm {a l n}} = \\frac {\\sum \\left(\\mathcal {M} _ {\\mathrm {d i f f}} \\cdot \\mathcal {M} _ {\\mathrm {d i s t}}\\right)}{\\sum \\left(\\mathcal {M} _ {\\mathrm {d i f f}}\\right)}. \\tag {8}\n$$\n", + "text_format": "latex", + "bbox": [ + 408, + 699, + 785, + 734 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "3.4 Loss Function", + "text_level": 1, + "bbox": [ + 215, + 753, + 377, + 768 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "In general, during an iteration cycle, CPM conducts the process of backward propagation concerning its loss function, which is defined as follow:", + "bbox": [ + 212, + 781, + 787, + 811 + ], + "page_idx": 7 + }, + { + "type": "equation", + "text": "\n$$\n\\mathrm {L} _ {\\text {t o t a l}} = \\mathrm {L} _ {\\mathrm {s}} + \\lambda_ {1} \\mathcal {P} (f _ {i}, f _ {i + 1}) \\mathrm {L} _ {\\mathrm {U}} + \\lambda_ {2} \\mathrm {L} _ {\\mathrm {T}}, \\tag {9}\n$$\n", + "text_format": "latex", + "bbox": [ + 366, + 825, + 785, + 842 + ], + "page_idx": 7 + }, + { + "type": "page_number", + "text": "8", + "bbox": [ + 217, + 114, + 228, + 126 + ], + "page_idx": 7 + }, + { + "type": "header", + "text": "N. Gao et al.", + "bbox": [ + 271, + 114, + 359, + 127 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "where $\\lambda_{1}$ and $\\lambda_{2}$ increase as the number of iterations grows, up to a point where they stop growing after a fixed iteration.", + "bbox": [ + 214, + 146, + 782, + 176 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "We employed two independent Gaussian warm-up function to control the loss function weights, $\\lambda_{1}$ and $\\lambda_{2}$ , using different parameters:", + "bbox": [ + 214, + 176, + 782, + 208 + ], + "page_idx": 8 + }, + { + "type": "equation", + "text": "\n$$\n\\lambda_ {1} (t) = \\left\\{ \\begin{array}{l l} \\hat {\\lambda} _ {1} \\cdot e ^ {- 5 \\left(1 - \\frac {2 t}{t _ {\\operatorname* {m a x}}}\\right) ^ {2}}, & t < \\frac {t _ {\\operatorname* {m a x}}}{2} \\\\ \\hat {\\lambda} _ {1}, & t \\geq \\frac {t _ {\\operatorname* {m a x}}}{2} \\end{array} \\right. \\tag {10}\n$$\n", + "text_format": "latex", + "bbox": [ + 367, + 222, + 784, + 268 + ], + "page_idx": 8 + }, + { + "type": "equation", + "text": "\n$$\n\\lambda_ {2} (t) = \\left\\{ \\begin{array}{l l} \\hat {\\lambda} _ {2} \\cdot e ^ {- 5 \\left(1 - \\frac {2 t}{t _ {\\max}}\\right) ^ {2}}, & t < \\frac {t _ {\\max}}{2} \\\\ \\hat {\\lambda} _ {2}, & t \\geq \\frac {t _ {\\max}}{2} \\end{array} \\right.\n$$\n", + "text_format": "latex", + "bbox": [ + 367, + 267, + 633, + 305 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "where $t$ represents the current iteration number, and $t_{\\mathrm{max}}$ represents the total number of training iterations. The hyperparameter $\\hat{\\lambda}_1$ and $\\hat{\\lambda}_2$ were empirically set to 20.0 and 10.0.", + "bbox": [ + 214, + 313, + 782, + 356 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "Loss function $\\mathrm{L_s}$ represents the loss during supervised learning and consists of the following components:", + "bbox": [ + 214, + 359, + 782, + 388 + ], + "page_idx": 8 + }, + { + "type": "equation", + "text": "\n$$\n\\mathrm {L} _ {\\mathrm {s}} = \\mathrm {C E} + \\mathrm {D I C E} + \\beta \\mathcal {P} \\left(f _ {i}, f _ {i + 1}\\right) \\mathrm {L} _ {\\mathrm {a l n}} \\left(\\mathcal {M} _ {\\text {d i f f}}, \\mathcal {M} _ {\\text {d i s t}}\\right), \\tag {11}\n$$\n", + "text_format": "latex", + "bbox": [ + 323, + 407, + 782, + 422 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "where $\\beta$ , a constant set to 0.5, is used to balance alignment loss and other losses.", + "bbox": [ + 214, + 431, + 782, + 445 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "Loss function $\\mathrm{L}_{\\mathrm{T}}$ is the consistency loss generated by the MT, which is derived by the teacher through the sharpen function to convert the model outputs into pseudo-labels, and subsequently generated by computing the MSE with the student model:", + "bbox": [ + 214, + 446, + 782, + 506 + ], + "page_idx": 8 + }, + { + "type": "equation", + "text": "\n$$\n\\mathrm {L} _ {\\mathrm {T}} = \\operatorname {M S E} \\left(f _ {i} \\left(\\mathbf {X} ^ {U}\\right), \\frac {\\mathbf {P} _ {\\mathbf {t}} ^ {1 / T}}{\\mathbf {P} _ {\\mathbf {t}} ^ {1 / T} + (1 - \\mathbf {P} _ {\\mathbf {t}}) ^ {1 / T}}\\right). \\tag {12}\n$$\n", + "text_format": "latex", + "bbox": [ + 336, + 522, + 782, + 561 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "4 Experiments", + "text_level": 1, + "bbox": [ + 215, + 594, + 374, + 612 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "4.1 Implementation Details", + "text_level": 1, + "bbox": [ + 215, + 628, + 455, + 643 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "We selected the VNet [13] model as a baseline network, which performs well in conditions with limited data and is essentially a 3D convolutional version of UNet. During inference, we use the average of the outputs from two networks as the final prediction. Specifically, the SGD optimizer was used to update the network parameters with weight decay of 0.0001 and a momentum of 0.9. The initial learning rate was set to 0.01, divided by 10 every 2500 iterations, for a total of 6000 iterations.", + "bbox": [ + 214, + 657, + 782, + 762 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "Following the practice in comparative literature [11, 12, 23, 30], our methods are trained for a fixed number of 6,000 iterations to obtain the final model. Additionally, our models all use a batch size of 4, with a labeled data quantity of 2. We tested the performance of the models, and all experiments were conducted on NVIDIA® GeForce A40 48GB running Ubuntu 20.04 and PyTorch 1.11.0.", + "bbox": [ + 214, + 763, + 782, + 839 + ], + "page_idx": 8 + }, + { + "type": "header", + "text": "Progressive Mean Teacher", + "bbox": [ + 555, + 114, + 730, + 128 + ], + "page_idx": 8 + }, + { + "type": "page_number", + "text": "9", + "bbox": [ + 774, + 114, + 785, + 126 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "4.2 Datasets and Metrics", + "text_level": 1, + "bbox": [ + 215, + 146, + 437, + 159 + ], + "page_idx": 9 + }, + { + "type": "text", + "text": "In the experiment, we selected two datasets with different modalities and utilized four distinct metrics to assess the performance of the model. For each dataset, $80\\%$ of the data was used as the training set and $20\\%$ as the test set. The proportion of supervised data was determined based on the training set.", + "bbox": [ + 212, + 176, + 782, + 234 + ], + "page_idx": 9 + }, + { + "type": "text", + "text": "LA Dataset. It [28] includes 100 3D gadolinium-enhanced MR imaging volumes of left atrial with an isotropic resolution of $0.625 \\times 0.625 \\times 0.625mm^3$ and the corresponding ground truth labels. For pre-processing, we first normalize all volumes to zero mean and unit variance, then crop each 3D MRI volume with enlarged margins according to the targets. During training, the training volumes are randomly cropped to $112 \\times 112 \\times 80$ as the model input. During inference, a sliding window of the same size is used to obtain segmentation results with a stride of $18 \\times 18 \\times 4$ .", + "bbox": [ + 212, + 237, + 784, + 356 + ], + "page_idx": 9 + }, + { + "type": "text", + "text": "Pancreas-NIH Dataset. It [17] provides 82 contrast-enhanced abdominal 3D CT volumes of pancreas with manual annotation. The size of each CT volume is $512 \\times 512 \\times D$ , where $D \\in [181, 466]$ . In pre-processing, we use the soft tissue CT window of $[-120, 240]$ HU, and we crop the CT scans centering at the pancreas region, and enlarge margins with 25 voxels. The training volumes are randomly cropped to $96 \\times 96 \\times 96$ as the model input. During inference, a sliding window of the same size is used to obtain segmentation results with a stride of $16 \\times 16 \\times 16$ .", + "bbox": [ + 212, + 359, + 784, + 463 + ], + "page_idx": 9 + }, + { + "type": "text", + "text": "Metrics. Following [2, 11, 12, 23, 29, 30], we use four metrics to evaluate model performance, including regional sensitive metrics: Dice similarity coefficient (Dice) [30], Jaccard similarity coefficient (Jaccard) [12], and edge sensitive metrics: $95\\%$ Hausdorff Distance (95HD) [29] and Average Surface Distance (ASD) [2].", + "bbox": [ + 212, + 465, + 784, + 542 + ], + "page_idx": 9 + }, + { + "type": "text", + "text": "4.3 Ablation Study", + "text_level": 1, + "bbox": [ + 215, + 566, + 387, + 583 + ], + "page_idx": 9 + }, + { + "type": "text", + "text": "For simplicity, we conducted ablation experiments on LA dataset to evaluate our design choices for each component. For a reliable assessment, all other parts were kept consistent except for the component under investigation.", + "bbox": [ + 212, + 597, + 782, + 642 + ], + "page_idx": 9 + }, + { + "type": "text", + "text": "Analysis of PMT Framework. The PMT architecture under progressive design is at the core of our work, wherein progressive design ensures the continuous generation of diverse pseudo-labels during the forward propagation process. Within the PMT architecture, PLF and DDA respectively denote the pseudolabel filtering and model alignment methods proposed in our paper, while MT ensures stable enhancement of model capability. We employ a non-progressive design co-training framework consisting of two VNet components as the baseline model. When adding PLF and DDA methods to the baseline model, the progressive design is simultaneously incorporated into the baseline model. Finally, we evaluate the impact of the MT architecture on model representation capability, with results presented in Table 1. Overall, compared to the baseline model, which already possesses good representation capability, the PMT framework greatly enhances the model's representation capability.", + "bbox": [ + 212, + 643, + 784, + 839 + ], + "page_idx": 9 + }, + { + "type": "page_number", + "text": "10", + "bbox": [ + 217, + 114, + 235, + 126 + ], + "page_idx": 9 + }, + { + "type": "header", + "text": "N. Gao et al.", + "bbox": [ + 271, + 114, + 359, + 126 + ], + "page_idx": 9 + }, + { + "type": "table", + "img_path": "images/f9ee592f83af732a3750272dcb1ec1c02f09820ba2f1f65e88da46d391600aba.jpg", + "table_caption": [ + "Table 1: Ablation results about PMT framework on LA dataset" + ], + "table_footnote": [], + "table_body": "
MethodLabeledMetrics
PLFDDAMTDice↑Jaccard↑95HD↓ASD↓
---8(10%)88.2879.317.592.34
✓--8(10%)89.9081.736.141.72
-✓-8(10%)90.4382.605.981.66
✓✓-8(10%)90.4382.605.491.49
✓✓✓8(10%)90.8183.235.611.50
", + "bbox": [ + 236, + 170, + 761, + 292 + ], + "page_idx": 10 + }, + { + "type": "text", + "text": "Analysis of Regularization Strength. The hyperparameters $\\hat{\\lambda}_1$ and $\\hat{\\lambda}_2$ respectively characterize the regularization strength of the progressive methods and MT architecture on the model. We use a tenfold scale and scale $\\hat{\\lambda}_1$ and $\\hat{\\lambda}_2$ based on the hyperparameter specifications we use, as shown in Table 2. The results indicate that variations in the two parameters within a certain range do not significantly affect the model's performance, demonstrating a certain level of robustness of our model to parameter variations. Within a certain range, on the LA dataset with $10\\%$ labeled data, our model achieves the best performance when $\\hat{\\lambda}_1$ and $\\hat{\\lambda}_2$ are 20.0 and 10.0 respectively.", + "bbox": [ + 212, + 315, + 787, + 454 + ], + "page_idx": 10 + }, + { + "type": "table", + "img_path": "images/09e9aabbc4b197991aefacd36a7babddd50585b8b8a60a93d9b2d32ec6d08db7.jpg", + "table_caption": [ + "Table 2: Ablation results about regularization strength on LA dataset" + ], + "table_footnote": [], + "table_body": "
λLabeledMetrics
λ1maxλ2maxDice↑Jaccard↑95HD↓ASD↓
2.01.08(10%)89.6681.336.351.84
2.010.08(10%)90.4782.665.581.64
2.0100.08(10%)87.7478.289.332.84
20.01.08(10%)90.6782.986.311.77
20.010.08(10%)90.8183.235.611.50
20.0100.08(10%)88.0078.739.052.78
200.01.08(10%)89.0880.4211.403.07
200.010.08(10%)89.7581.477.082.22
200.0100.08(10%)83.8072.5516.445.11
", + "bbox": [ + 241, + 503, + 754, + 680 + ], + "page_idx": 10 + }, + { + "type": "text", + "text": "4.4 Comparison with Other Methods", + "text_level": 1, + "bbox": [ + 215, + 726, + 535, + 742 + ], + "page_idx": 10 + }, + { + "type": "text", + "text": "We compared our approach with previous state-of-the-art methods on LA dataset and Pancreas-NIH dataset.", + "bbox": [ + 212, + 750, + 782, + 777 + ], + "page_idx": 10 + }, + { + "type": "text", + "text": "We chose VNet as baseline models for comparison. For the selected alternative models, we opted for UA-MT [30] with uncertainty estimation, SASSNet [11] focusing on the regularity of geometric shapes, DTC [12] with task-level regularization, BCP [2] using bidirectional CutMix [31], and MCF [23] with model-level", + "bbox": [ + 212, + 779, + 784, + 840 + ], + "page_idx": 10 + }, + { + "type": "header", + "text": "Progressive Mean Teacher", + "bbox": [ + 555, + 114, + 730, + 128 + ], + "page_idx": 10 + }, + { + "type": "page_number", + "text": "11", + "bbox": [ + 767, + 114, + 782, + 126 + ], + "page_idx": 10 + }, + { + "type": "text", + "text": "regularization, with BCP and MCF being state-of-the-art results. Noting that, for BCP, we follow its parameter settings for pre-training 2,000 times and self-training 15,000 times.", + "bbox": [ + 212, + 146, + 782, + 191 + ], + "page_idx": 11 + }, + { + "type": "text", + "text": "Comparison on LA Dataset. We conducted a cross-model comparison on the classic LA dataset. We tested with $5\\%$ and $10\\%$ of labeled data. Results of the experiments are presented in Table 3. To provide a more intuitive demonstration of the performance of various models on the LA dataset, we have selected some representative results for visualization, as illustrated in Fig. 4. Areas with inaccurate segmentation have been annotated accordingly.", + "bbox": [ + 212, + 191, + 785, + 282 + ], + "page_idx": 11 + }, + { + "type": "table", + "img_path": "images/3e4e4610f48e2fd2241950049299ca5be2dffcf4a882f6228ce47ba87981d950.jpg", + "table_caption": [ + "Table 3: Comparison results on LA dataset with $5\\%$ and $10\\%$ labeled data" + ], + "table_footnote": [], + "table_body": "
MethodLabeledMetrics
Dice↑Jaccard↑95HD↓ASD↓
VNet4(5%)52.5539.6047.059.87
UA-MT4(5%)82.2670.9813.713.82
SASSNet4(5%)81.6069.6316.163.58
DTC4(5%)81.2569.3314.903.99
MCFSOTA4(5%)----
BCPSOTA4(5%)88.0278.727.902.15
PMT(Ours)4(5%)89.4781.046.451.86
VNet8(10%)82.7471.7213.353.26
UA-MT8(10%)86.2876.1118.714.63
SASSNet8(10%)85.2275.0911.182.89
MCFSOTA8(10%)88.7180.416.321.90
BCPSOTA8(10%)89.6281.316.811.76
PMT(Ours)8(10%)90.8183.235.611.50
", + "bbox": [ + 236, + 333, + 761, + 571 + ], + "page_idx": 11 + }, + { + "type": "text", + "text": "Our model outperformed previous models on all four metrics across $5\\%$ and $10\\%$ labeled data. At $5\\%$ of the data, compared to the best results from previous work, our PMT showed a $1.45\\%$ improvement in Dice, a $2.32\\%$ improvement in Jaccard, a reduction of 1.45 in 95HD, and a reduction of 0.29 in ASD. Our PMT also showed a $1.19\\%$ improvement in Dice, a $1.92\\%$ improvement in Jaccard, a reduction of 1.20 in 95HD, and a reduction of 0.26 in ASD at $10\\%$ of the data. It is noteworthy that our model maintains good performance even with a small amount of data. In comparison with state-of-the-art results, our model achieves a leading performance in the majority of metrics using only half of their data. For instance, we surpass MCF's performance at $10\\%$ labeled data using only $5\\%$ labeled data. The above indicates that our model achieves superior results compared to existing methods on differently proportioned annotated data in the task of segmenting the left atrium.", + "bbox": [ + 212, + 598, + 787, + 792 + ], + "page_idx": 11 + }, + { + "type": "text", + "text": "Comparison on Pancreas-NIH Dataset. We conducted a cross-model comparison on the classic Pancreas dataset. Detailed results of the experiments are presented in Table 4. We tested with $10\\%$ and $20\\%$ of labeled data. To", + "bbox": [ + 212, + 795, + 785, + 839 + ], + "page_idx": 11 + }, + { + "type": "page_number", + "text": "12", + "bbox": [ + 217, + 114, + 235, + 126 + ], + "page_idx": 11 + }, + { + "type": "header", + "text": "N. Gao et al.", + "bbox": [ + 271, + 114, + 359, + 126 + ], + "page_idx": 11 + }, + { + "type": "image", + "img_path": "images/6f78183ef03b2653e9009c0a2726c7e3b72f0700e96b436b845a88d012677779.jpg", + "image_caption": [ + "Fig. 4: 2D & 3D segmentation visualization of different semi-supervised methods under $10\\%$ labeled on LA (upper) and pancreas (bottom) dataset." + ], + "image_footnote": [], + "bbox": [ + 215, + 143, + 785, + 448 + ], + "page_idx": 12 + }, + { + "type": "text", + "text": "provide a more intuitive demonstration of the performance of various models on the Pancreas-NIH dataset, we have selected some representative results for visualization, as illustrated in Fig. 4. Areas with inaccurate segmentation have been annotated accordingly. It is worth noting that the results in the table clearly indicate that Pancreas-NIH dataset is significantly more challenging than LA dataset. Therefore, we increased $\\hat{\\lambda}_{1}$ by a factor of two, resulting in improved performance.", + "bbox": [ + 212, + 517, + 784, + 622 + ], + "page_idx": 12 + }, + { + "type": "text", + "text": "Our model outperformed previous models on all four metrics across $10\\%$ and $20\\%$ labeled data. At $10\\%$ of the data, compared to the best results from previous work, our PMT showed a $7.17\\%$ improvement in Dice, a $9.09\\%$ improvement in Jaccard, a reduction of 6.35 in 95HD, and a reduction of 2.10 in ASD. Our PMT also showed a $0.31\\%$ improvement in Dice, a $0.55\\%$ improvement in Jaccard, and a reduction of 0.36 in ASD at $20\\%$ of the data. The above indicates that our model achieves superior results compared to existing methods on differently proportioned annotated data in the task of segmenting the pancreas.", + "bbox": [ + 212, + 625, + 785, + 744 + ], + "page_idx": 12 + }, + { + "type": "text", + "text": "5 Conclusion", + "text_level": 1, + "bbox": [ + 215, + 773, + 357, + 787 + ], + "page_idx": 12 + }, + { + "type": "text", + "text": "In this paper, we propose a semi-supervised medical image segmentation framework named Progress Mean Teacher (PMT). PMT adopts a progressive", + "bbox": [ + 212, + 809, + 784, + 839 + ], + "page_idx": 12 + }, + { + "type": "header", + "text": "Progressive Mean Teacher", + "bbox": [ + 555, + 114, + 730, + 127 + ], + "page_idx": 12 + }, + { + "type": "page_number", + "text": "13", + "bbox": [ + 767, + 114, + 784, + 126 + ], + "page_idx": 12 + }, + { + "type": "table", + "img_path": "images/387f66b0ca3ff35902e2ef3becfb5089c9f1576e246792d4326b458a2fe7526a.jpg", + "table_caption": [ + "Table 4: Comparison results on Pancreas-NIH dataset with ${10}\\%$ and ${20}\\%$ labeled data" + ], + "table_footnote": [], + "table_body": "
MethodLabeledMetrics
Dice↑Jaccard↑95HD↓ASD↓
VNet6(10%)55.6041.7445.3318.63
UA-MT6(10%)66.3453.2117.214.57
SASSNet6(10%)68.7853.8619.026.26
DTC6(10%)69.2154.0617.215.95
BCPSOTA6(10%)73.8359.2412.713.72
MCFSOTA6(10%)----
PMT(Ours)6(10%)81.0068.336.361.62
VNet12(20%)72.3858.2619.355.89
UA-MT12(20%)76.1062.6210.842.43
SASSNet12(20%)77.6664.0810.933.05
DTC12(20%)78.2764.758.362.25
BCPSOTA12(20%)82.9170.976.432.25
MCFSOTA12(20%)75.0061.2711.593.27
PMT(Ours)12(20%)83.2271.527.601.89
", + "bbox": [ + 236, + 171, + 761, + 422 + ], + "page_idx": 13 + }, + { + "type": "text", + "text": "design training process, establishing temporal-level model alignment and pseudolabel filtering while leveraging a network architecture based on MT. The core idea of this model framework is to generate diverse pseudo-labels consistently during the backpropagation process by establishing representation capability differences caused by iteration gaps, thereby stabilizing and enhancing the model's representation capability. In addition to the progressive architecture, PMT employs two simple yet effective methods, Pseudo Label Filtering (PLF) and Discrepancy Driven Alignment (DDA). PLF utilizes the representation capability of the model to filter pseudo-labels, discarding low-fidelity ones detrimental to constraining, while DDA aligns the differences in model predictions, allowing lagging models to catch up with leading models rapidly. Results from ablation experiments demonstrate that each component of the PMT framework significantly enhances the model's performance. In comparative experiments with other methods, PMT achieves state-of-the-art performance in terms of accuracy, surpassing previous methods significantly, and maintains this advantage compared to other methods in situations with more limited data and more challenging tasks.", + "bbox": [ + 217, + 488, + 785, + 729 + ], + "page_idx": 13 + }, + { + "type": "text", + "text": "Limitation and Future Work. Despite the outstanding performance of PMT, there is still room for further exploration of semi-supervised training architectures established through temporal consistency. Investigating more advanced and stable strategies under a progressive design paradigm, and assessing whether they can further enhance performance, are topics worthy of further research in the future.", + "bbox": [ + 217, + 748, + 785, + 839 + ], + "page_idx": 13 + }, + { + "type": "page_number", + "text": "14", + "bbox": [ + 217, + 114, + 235, + 126 + ], + "page_idx": 13 + }, + { + "type": "header", + "text": "N. Gao et al.", + "bbox": [ + 271, + 114, + 359, + 126 + ], + "page_idx": 13 + }, + { + "type": "text", + "text": "Acknowledgments", + "text_level": 1, + "bbox": [ + 217, + 143, + 392, + 162 + ], + "page_idx": 14 + }, + { + "type": "text", + "text": "This work was supported in part by National Natural Science Foundation of China under Grants 62088102, 12326608 and 62106192, Natural Science Foundation of Shaanxi Province under Grant 2022JC-41, and Fundamental Research Funds for the Central Universities under Grant XTR042021005.", + "bbox": [ + 212, + 172, + 787, + 233 + ], + "page_idx": 14 + }, + { + "type": "text", + "text": "References", + "text_level": 1, + "bbox": [ + 217, + 253, + 321, + 268 + ], + "page_idx": 14 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "1. Achanta, R., Shaji, A., Smith, K., Lucchi, A., Fua, P., Süsstrunk, S.: Slic superpixels compared to state-of-the-art superpixel methods. IEEE transactions on pattern analysis and machine intelligence 34(11), 2274-2282 (2012) 4", + "2. Bai, Y., Chen, D., Li, Q., Shen, W., Wang, Y.: Bidirectional copy-paste for semi-supervised medical image segmentation. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 11514-11524 (2023) 10, 11", + "3. Cai, H., Li, S., Qi, L., Yu, Q., Shi, Y., Gao, Y.: Orthogonal annotation benefits barely-supervised medical image segmentation. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 3302-3311 (2023) 4", + "4. Chen, X., Yuan, Y., Zeng, G., Wang, J.: Semi-supervised semantic segmentation with cross pseudo supervision. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 2613-2622 (2021) 5", + "5. Chen, X., He, K.: Exploring simple siamese representation learning. In: Proceedings of the IEEE/CVF conference on computer vision and pattern recognition. pp. 15750-15758 (2021) 2", + "6. Dong-DongChen, W., WeiGao, Z.: Tri-net for semi-supervised deep learning. In: Proceedings of twenty-seventh international joint conference on artificial intelligence. pp. 2014-2020 (2018) 4", + "7. Jiao, R., Zhang, Y., Ding, L., Cai, R., Zhang, J.: Learning with limited annotations: a survey on deep semi-supervised learning for medical image segmentation. arXiv preprint arXiv:2207.14191 (2022) 4", + "8. Ke, Z., Qiu, D., Li, K., Yan, Q., Lau, R.W.: Guided collaborative training for pixelwise semi-supervised learning. In: Computer Vision-ECCV 2020: 16th European Conference, Glasgow, UK, August 23-28, 2020, Proceedings, Part XIII 16. pp. 429-445. Springer (2020) 5", + "9. Laine, S., Aila, T.: Temporal ensembling for semi-supervised learning. arXiv preprint arXiv:1610.02242 (2016) 2, 4", + "10. Lee, D.H., et al.: Pseudo-label: The simple and efficient semi-supervised learning method for deep neural networks. In: Workshop on challenges in representation learning, ICML. vol. 3, p. 896. Atlanta (2013) 2, 4", + "11. Li, S., Zhang, C., He, X.: Shape-aware semi-supervised 3d semantic segmentation for medical images. In: Medical Image Computing and Computer Assisted Intervention-MICCAI 2020: 23rd International Conference, Lima, Peru, October 4-8, 2020, Proceedings, Part I 23. pp. 552-561. Springer (2020) 4, 9, 10, 11", + "12. Luo, X., Chen, J., Song, T., Wang, G.: Semi-supervised medical image segmentation through dual-task consistency. In: Proceedings of the AAAI conference on artificial intelligence. vol. 35, pp. 8801-8809 (2021) 4, 9, 10, 11", + "13. Miletari, F., Navab, N., Ahmadi, S.A.: V-net: Fully convolutional neural networks for volumetric medical image segmentation. In: 2016 fourth international conference on 3D vision (3DV). pp. 565-571. IEEE (2016) 9" + ], + "bbox": [ + 218, + 281, + 784, + 839 + ], + "page_idx": 14 + }, + { + "type": "header", + "text": "Progressive Mean Teacher", + "bbox": [ + 555, + 114, + 730, + 128 + ], + "page_idx": 14 + }, + { + "type": "page_number", + "text": "15", + "bbox": [ + 767, + 116, + 784, + 126 + ], + "page_idx": 14 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "14. Na, J., Ha, J.W., Chang, H.J., Han, D., Hwang, W.: Switching temporary teachers for semi-supervised semantic segmentation. Advances in Neural Information Processing Systems 36 (2024) 5", + "15. Peiris, H., Hayat, M., Chen, Z., Egan, G., Harandi, M.: Uncertainty-guided dual-views for semi-supervised volumetric medical image segmentation. Nature Machine Intelligence 5(7), 724-738 (2023) 4", + "16. Rizve, M.N., Kardan, N., Shah, M.: Towards realistic semi-supervised learning. In: European Conference on Computer Vision. pp. 437-455. Springer (2022) 1", + "17. Roth, H.R., Lu, L., Farag, A., Shin, H.C., Liu, J., Turkbey, E.B., Summers, R.M.: Deeporgan: Multi-level deep convolutional networks for automated pancreas segmentation. In: Medical Image Computing and Computer-Assisted Intervention-MICCAI 2015: 18th International Conference, Munich, Germany, October 5-9, 2015, Proceedings, Part I 18. pp. 556-564. Springer (2015) 3, 10", + "18. Shen, Z., Cao, P., Yang, H., Liu, X., Yang, J., Zaiane, O.R.: Co-training with high-confidence pseudo labels for semi-supervised medical image segmentation. arXiv preprint arXiv:2301.04465 (2023) 5", + "19. Tarvainen, A., Valpola, H.: Mean teachers are better role models: Weight-averaged consistency targets improve semi-supervised deep learning results. Advances in neural information processing systems 30 (2017) 2, 4, 5", + "20. Wang, H., Li, X.: Dhc: Dual-debiased heterogeneous co-training framework for class-imbalanced semi-supervised medical image segmentation. In: International Conference on Medical Image Computing and Computer-Assisted Intervention. pp. 582-591. Springer (2023) 2", + "21. Wang, J., Zhou, S., Fang, C., Wang, L., Wang, J.: Meta corrupted pixels mining for medical image segmentation. In: Medical Image Computing and Computer Assisted Intervention-MICCAI 2020: 23rd International Conference, Lima, Peru, October 4-8, 2020, Proceedings, Part I 23. pp. 335-345. Springer (2020) 2", + "22. Wang, X., Yuan, Y., Guo, D., Huang, X., Cui, Y., Xia, M., Wang, Z., Bai, C., Chen, S.: Ssa-net: Spatial self-attention network for COVID-19 pneumonia infection segmentation with semi-supervised few-shot learning. Medical Image Analysis 79, 102459 (2022) 4", + "23. Wang, Y., Xiao, B., Bi, X., Li, W., Gao, X.: Mcf: Mutual correction framework for semi-supervised medical image segmentation. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 15651-15660 (2023) 2, 3, 4, 9, 10, 11", + "24. Wu, H., Wang, Z., Song, Y., Yang, L., Qin, J.: Cross-patch dense contrastive learning for semi-supervised segmentation of cellular nuclei in histopathologic images. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 11666-11675 (2022) 2", + "25. Xia, K., Wang, L., Zhou, S., Hua, G., Tang, W.: Learning from noisy pseudo labels for semi-supervised temporal action localization. In: Proceedings of the IEEE/CVF International Conference on Computer Vision. pp. 10160-10169 (2023) 1", + "26. Xie, Q., Luong, M.T., Hovy, E., Le, Q.V.: Self-training with noisy student improves imagenet classification. In: Proceedings of the IEEE/CVF conference on computer vision and pattern recognition. pp. 10687-10698 (2020) 2", + "27. Xin, X., Wang, J., Xie, R., Zhou, S., Huang, W., Zheng, N.: Semi-supervised person re-identification using multi-view clustering. Pattern Recognition 88, 285–297 (2019) 1", + "28. Xiong, Z., Xia, Q., Hu, Z., Huang, N., Bian, C., Zheng, Y., Vesal, S., Ravikumar, N., Maier, A., Yang, X., et al.: A global benchmark of algorithms for segmenting" + ], + "bbox": [ + 215, + 146, + 785, + 840 + ], + "page_idx": 15 + }, + { + "type": "page_number", + "text": "16", + "bbox": [ + 217, + 114, + 235, + 126 + ], + "page_idx": 15 + }, + { + "type": "header", + "text": "N. Gao et al.", + "bbox": [ + 271, + 114, + 359, + 126 + ], + "page_idx": 15 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "the left atrium from late gadolinium-enhanced cardiac magnetic resonance imaging. Medical image analysis 67, 101832 (2021) 3, 10", + "29. Xu, Z., Wang, Y., Lu, D., Yu, L., Yan, J., Luo, J., Ma, K., Zheng, Y., Tong, R.K.y.: All-around real label supervision: Cyclic prototype consistency learning for semi-supervised medical image segmentation. IEEE Journal of Biomedical and Health Informatics 26(7), 3174-3184 (2022) 4, 10", + "30. Yu, L., Wang, S., Li, X., Fu, C.W., Heng, P.A.: Uncertainty-aware self-ensembling model for semi-supervised 3d left atrium segmentation. In: Medical Image Computing and Computer Assisted Intervention-MICCAI 2019: 22nd International Conference, Shenzhen, China, October 13–17, 2019, Proceedings, Part II 22. pp. 605–613. Springer (2019) 4, 9, 10, 11", + "31. Yun, S., Han, D., Oh, S.J., Chun, S., Choe, J., Yoo, Y.: Cutmix: Regularization strategy to train strong classifiers with localizable features. In: Proceedings of the IEEE/CVF international conference on computer vision. pp. 6023-6032 (2019) 11", + "32. Zheng, M., You, S., Huang, L., Wang, F., Qian, C., Xu, C.: Simmatch: Semi-supervised learning with similarity matching. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 14471-14481 (2022) 1", + "33. Zhou, S., Wang, J., Shu, J., Meng, D., Wang, L., Zheng, N.: Multinetwork collaborative feature learning for semisupervised person reidentification. IEEE Transactions on Neural Networks and Learning Systems 33(9), 4826-4839 (2021) 1" + ], + "bbox": [ + 215, + 146, + 785, + 439 + ], + "page_idx": 16 + }, + { + "type": "header", + "text": "Progressive Mean Teacher", + "bbox": [ + 555, + 114, + 730, + 128 + ], + "page_idx": 16 + }, + { + "type": "page_number", + "text": "17", + "bbox": [ + 767, + 114, + 785, + 126 + ], + "page_idx": 16 + } +] \ No newline at end of file diff --git a/2024/PMT_ Progressive Mean Teacher via Exploring Temporal Consistency for Semi-Supervised Medical Image Segmentation/049d7a98-0307-4aa9-ae68-729601a51f9f_model.json b/2024/PMT_ Progressive Mean Teacher via Exploring Temporal Consistency for Semi-Supervised Medical Image Segmentation/049d7a98-0307-4aa9-ae68-729601a51f9f_model.json new file mode 100644 index 0000000000000000000000000000000000000000..13aa2a3da9530a2f89702b5dc0fdf73760b08a87 --- /dev/null +++ b/2024/PMT_ Progressive Mean Teacher via Exploring Temporal Consistency for Semi-Supervised Medical Image Segmentation/049d7a98-0307-4aa9-ae68-729601a51f9f_model.json @@ -0,0 +1,2016 @@ +[ + [ + { + "type": "title", + "bbox": [ + 0.235, + 0.141, + 0.773, + 0.21 + ], + "angle": 0, + "content": "PMT: Progressive Mean Teacher via Exploring Temporal Consistency for Semi-Supervised Medical Image Segmentation" + }, + { + "type": "text", + "bbox": [ + 0.271, + 0.236, + 0.731, + 0.253 + ], + "angle": 0, + "content": "Ning Gao\\(^{\\text{D}}\\), Sanping Zhou\\(^{*}\\), Le Wang\\(^{\\text{D}}\\), and Nanning Zheng\\(^{\\text{D}}\\)" + }, + { + "type": "text", + "bbox": [ + 0.236, + 0.263, + 0.766, + 0.306 + ], + "angle": 0, + "content": "National Key Laboratory of Human-Machine Hybrid Augmented Intelligence, National Engineering Research Center for Visual Information and Applications, Institute of Artificial Intelligence and Robotics, Xi'an Jiaotong University" + }, + { + "type": "text", + "bbox": [ + 0.261, + 0.341, + 0.744, + 0.675 + ], + "angle": 0, + "content": "Abstract. Semi-supervised learning has emerged as a widely adopted technique in the field of medical image segmentation. The existing works either focuses on the construction of consistency constraints or the generation of pseudo labels to provide high-quality supervisory signals, whose main challenge mainly comes from how to keep the continuous improvement of model capabilities. In this paper, we propose a simple yet effective semi-supervised learning framework, termed Progressive Mean Teachers (PMT), for medical image segmentation, whose goal is to generate high-fidelity pseudo labels by learning robust and diverse features in the training process. Specifically, our PMT employs a standard mean teacher to penalize the consistency of the current state and utilizes two sets of MT architectures for co-training. The two sets of MT architectures are individually updated for prolonged periods to maintain stable model diversity established through performance gaps generated by iteration differences. Additionally, a difference-driven alignment regularizer is employed to expedite the alignment of lagging models with the representation capabilities of leading models. Furthermore, a simple yet effective pseudo-label filtering algorithm is employed for facile evaluation of models and selection of high-fidelity pseudo-labels outputted when models are operating at high performance for co-training purposes. Experimental results on two datasets with different modalities, i.e., CT and MRI, demonstrate that our method outperforms the state-of-the-art medical image segmentation approaches across various dimensions. The code is available at https://github.com/Axi404/PMT." + }, + { + "type": "text", + "bbox": [ + 0.261, + 0.687, + 0.741, + 0.716 + ], + "angle": 0, + "content": "Keywords: Semi-supervised learning \\(\\cdot\\) Medical image segmentation \\(\\cdot\\) Temporal consistency regularization" + }, + { + "type": "title", + "bbox": [ + 0.216, + 0.741, + 0.377, + 0.758 + ], + "angle": 0, + "content": "1 Introduction" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.772, + 0.788, + 0.818 + ], + "angle": 0, + "content": "Semi-supervised learning is an important field in deep learning, which offers an effective way to tackle problems with limited labeled data [16, 25, 27, 32, 33]. With the continuous emergence of large-scale data, semi-supervised learning has" + }, + { + "type": "page_footnote", + "bbox": [ + 0.218, + 0.825, + 0.386, + 0.841 + ], + "angle": 0, + "content": "* Corresponding author." + } + ], + [ + { + "type": "page_number", + "bbox": [ + 0.218, + 0.116, + 0.23, + 0.127 + ], + "angle": 0, + "content": "2" + }, + { + "type": "header", + "bbox": [ + 0.272, + 0.115, + 0.361, + 0.127 + ], + "angle": 0, + "content": "N. Gao et al." + }, + { + "type": "image", + "bbox": [ + 0.321, + 0.147, + 0.682, + 0.328 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.214, + 0.343, + 0.788, + 0.428 + ], + "angle": 0, + "content": "Fig. 1: Motivation of our PMT. In particular, the standard MT helps to learn robust features by keeping the consistency between teacher and student networks at the current iteration, while our PMT further helps to learn diverse features by maintaining the difference between student networks at different iterations. As a result, more and more high-fidelity pseudo labels will be generated for semi-supervised medical image segmentation." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.46, + 0.788, + 0.536 + ], + "angle": 0, + "content": "become a hot topic in both machine learning and computer vision communities. It has gained particular attention in the medical image segmentation [21], due to the need for expert annotation in determining the accurate boundaries of targets. As a result, the labeled data is very lacking in medical image segmentation, which makes the semi-supervised learning very popular in this domain." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.537, + 0.788, + 0.748 + ], + "angle": 0, + "content": "Most of the semi-supervised learning methods involve seeking a guidance mechanism using limited annotated data, so as to maximize the usage of unlabeled data to train the models on different downstream tasks. In general, two technologies, i.e., consistency regularization [9, 10] and pseudo label generation [10, 19], are commonly used in semi-supervised medical segmentation. The former ones aim at enhancing the representation capability of networks by exploring the consistency based on different prior assumptions [5], while the latter ones focus on improving the performance on downstream tasks by generating the pseudo labels based on current model capability. For example, the well-known Mean Teacher (MT) [19] utilizes an ingenious Exponential Moving Average (EMA) that is equivalent to data augmentation and enforces consistency regularization between the outputs of teacher and student models. Besides, the representative Noisy Student [26] designs a teacher model to generate pseudo labels, so as to train a larger student model." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.75, + 0.788, + 0.841 + ], + "angle": 0, + "content": "Even though the above methods have achieved significant improvements in semi-supervised medical image segmentation, we still argue that they are weak in exploring high-quality supervisory signals to consistently enhance the model's capability. To address this issue, an iterative unified optimization framework has been introduced to semi-supervised medical image segmentation [20, 23, 24], in which the consistency regularization strategy is taken to enhance the model's" + } + ], + [ + { + "type": "header", + "bbox": [ + 0.557, + 0.115, + 0.732, + 0.129 + ], + "angle": 0, + "content": "Progressive Mean Teacher" + }, + { + "type": "page_number", + "bbox": [ + 0.775, + 0.117, + 0.785, + 0.127 + ], + "angle": 0, + "content": "3" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.147, + 0.784, + 0.237 + ], + "angle": 0, + "content": "representation capability and the pseudo label generation algorithm is applied to generate high-fidelity pseudo labels. For example, the recent MCF [23] takes VNet and 3D ResNet for representation learning, and heterogeneous networks for dynamic pseudo label generation. The challenges to these methods lie on how to continuously generate diverse pseudo labels in the forward propagation, and enhance model's capability in the backward propagation." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.239, + 0.788, + 0.585 + ], + "angle": 0, + "content": "In this paper, we design a novel semi-supervised learning framework, termed Progressive Mean Teachers (PMT), for medical image segmentation, whose main idea focuses on how to obtain a diverse set of accurate pseudo labels. Inspired by the positive relationship between iteration and performance, we try to maintain the diversity of networks by exploring their states at different training epochs. Specifically, the standard MT is first taken as basic architecture to learn the parameters of network at each iteration, which can enhance the network's representation capability by using the EMA data augmentation. Our PMT further explores network diversity during training, alternating between two homogeneous MT architectures trained on the same dataset, a process we term progressive design. These models exhibit significant iteration leads due to alternating continuous individual updates, establishing performance gaps between networks at different epochs, as shown in Fig. 1. Consequently, the student network can acquire robust yet diverse features for medical image segmentation. Then, the Discrepancy Driven Alignment (DDA) regularizer is further designed to examine disparities between predictions obtained by the student network at different training epochs, facilitating rapid alignment to high-fidelity generated images. Finally, we design a simple Pseudo Label Filtering (PLF) algorithm to refine the basic interaction process, enabling the retention of high-fidelity pseudo-labels for training by comparing student network performance across different training epochs. As a result, more and more high-fidelity pseudo labels can be fed to train the other student networks, which will in turn generate more accurate predictions for pseudo label generation." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.587, + 0.788, + 0.707 + ], + "angle": 0, + "content": "In summary, the main contributions of this work are as follows: (1) We design a novel Progressive Mean Teacher framework for semi-supervised medical image segmentation. (2) We design a novel Discrepancy Driven Alignment regularizer to rapidly align the representational capacity gap between lagging and leading networks. (3) We design a simple yet effective Pseudo Label Filtering algorithm to select high-fidelity pseudo labels. Extensive experiments in both Left Atrial [28] and Pancreas-NIH [17] datasets show that our PMT can achieve the state-of-the-art results in semi-supervised medical image segmentation." + }, + { + "type": "title", + "bbox": [ + 0.215, + 0.733, + 0.388, + 0.749 + ], + "angle": 0, + "content": "2 Related Work" + }, + { + "type": "title", + "bbox": [ + 0.215, + 0.768, + 0.487, + 0.782 + ], + "angle": 0, + "content": "2.1 Consistency Regularization" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.795, + 0.787, + 0.84 + ], + "angle": 0, + "content": "Consistency regularization is often employed in semi-supervised learning, so as to enhance the stable representational capacity of model. The behind idea is to preserve the invariance of predictions made by the same model when facing" + } + ], + [ + { + "type": "page_number", + "bbox": [ + 0.218, + 0.116, + 0.23, + 0.127 + ], + "angle": 0, + "content": "4" + }, + { + "type": "header", + "bbox": [ + 0.272, + 0.115, + 0.361, + 0.128 + ], + "angle": 0, + "content": "N. Gao et al." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.147, + 0.788, + 0.404 + ], + "angle": 0, + "content": "perturbations applied in different regions, by constraining the consistency between output results under different perturbations. For example, \\(\\varPi\\) model [9] introduces image-level regularization by adding perturbations to images. The well-known MT [19] introduces parameter-level regularization by using EMA to constrain the outputs between teacher and student models. Thanks to its simplicity and effectiveness, more and more works are paying attention to improve the generalization ability by formulating different consistency regularizers. For example, SASSnet [11] focuses on the regularity of geometric shapes for target object classes within consistency regularization. Besides, CPCL [29] establishes regularization between supervised and unsupervised training within a cyclic framework. Furthermore, some later works have made progress in regularization at different task and model levels. For example, DTC [12] introduces task-level regularization, presenting a novel dual-task consistency semi-supervised framework. Besides, MCF [23] introduces model-level regularization by using heterogeneous models to constrain output consistency. Compared to previous work, our PMT seeks the cross-temporal regularization between different training periods, which can help learn diverse yet robust features for subsequent pseudo label generation." + }, + { + "type": "title", + "bbox": [ + 0.216, + 0.425, + 0.473, + 0.438 + ], + "angle": 0, + "content": "2.2 Pseudo Label Generation" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.448, + 0.789, + 0.751 + ], + "angle": 0, + "content": "Pseudo label generation is often employed in semi-supervised learning, so as to enhance the discriminative ability of model. The behind idea is to train a prior model with the labeled data and then apply it to generate the pseudo labels for unlabeled data. It is widely accepted to categorize pseudo-label generation into two methods [7], direct generation focuses on selecting pseudo labels with higher confidence, while indirect generation explores methods to generate high-fidelity pseudo labels. For direct generation, work [10] uses a fixed threshold to choose high-confidence pseudo labels. SsaNet [22] employs a trust module to reevaluate pseudo labels. UA-MT [30] introduces uncertainty estimation to filter out unreliable pseudo labels. Co-BioNet [15] introduces a feedback network to measure the uncertainty and choose high-confidence predictions of different models. For indirect generation, Tri-Net [6] is proposed to use two subnetworks to generate pseudo labels for a third subnetwork. Work [1] improves pseudo label generation using Simple Linear Iterative Clustering (SLIC) algorithm. MCF [23] dynamically generates pseudo labels using a heterogeneous network and effectively addresses cognitive bias, while DeSCO [3] focuses on the spatial correlation of medical images and generates pseudo labels using orthogonal slices. Compared to previous work, our PMT emphasizes the enhancement of pseudo label quality in the temporal domain and generates diverse and robust pseudo labels across different training periods, significantly improving performance." + }, + { + "type": "title", + "bbox": [ + 0.216, + 0.771, + 0.466, + 0.786 + ], + "angle": 0, + "content": "2.3 Multi-Model Framework" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.795, + 0.787, + 0.84 + ], + "angle": 0, + "content": "The multi-model framework is often employed in semi-supervised learning, so as to enhance the model representation by acquiring multiple views or diversity. Its development aligns closely with consistency regularization. Here, we" + } + ], + [ + { + "type": "header", + "bbox": [ + 0.557, + 0.115, + 0.732, + 0.129 + ], + "angle": 0, + "content": "Progressive Mean Teacher" + }, + { + "type": "page_number", + "bbox": [ + 0.775, + 0.117, + 0.785, + 0.127 + ], + "angle": 0, + "content": "5" + }, + { + "type": "image", + "bbox": [ + 0.223, + 0.147, + 0.784, + 0.351 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.214, + 0.361, + 0.788, + 0.417 + ], + "angle": 0, + "content": "Fig. 2: An overview of PMT. We employed a progressive design and utilized the architecture of MT. The PMT framework maintains a data buffer of length \\(\\mathcal{B}\\) for cros-temporal training. The total loss function \\(\\mathrm{L}_{\\mathrm{total}}\\) for each network includes supervised losses \\(\\mathrm{L_{CE}}, \\mathrm{L_{DICE}}, \\mathrm{L_{aln}}\\), and unsupervised loss \\(\\mathrm{L_U}, \\mathrm{L_T}\\)." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.446, + 0.789, + 0.658 + ], + "angle": 0, + "content": "focus on the evolution of model structures. For example, MT [19] introduces a Teacher model with significantly improved representation capabilities at a lower cost using EMA, establishing the MT architecture. Besides, CPC [8] utilizes the confidence vector of multi-model outputs as pseudo-labels for co-training, while CPS [4] employs one-hot labels, both considered as starting points for co-training. In recent years, efforts have been made to combine these approaches. For example, UCMT [18] employs two student models for co-training, simultaneously updating the same Mean Teacher using EMA, while Dual Teacher [14] alternates updates between two Teacher models using EMA with a single student model, thereby fostering diversity among Teacher models. Compared to previous work, our PMT utilizes two sets of Mean Teachers in a progressive training framework, and employ two student models to update teacher models independently. This enables our model to rapidly establish robust performance disparities across iteration gap and maintain stable diversity among models." + }, + { + "type": "title", + "bbox": [ + 0.216, + 0.68, + 0.33, + 0.695 + ], + "angle": 0, + "content": "3 Method" + }, + { + "type": "title", + "bbox": [ + 0.215, + 0.71, + 0.502, + 0.725 + ], + "angle": 0, + "content": "3.1 The Overall Process of PMT" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.734, + 0.788, + 0.843 + ], + "angle": 0, + "content": "The training process of PMT is illustrated in Fig. 2. As previously mentioned, PMT can consist of multiple networks. For ease of explanation, the number of networks is assumed to be two by default in our work. In our approach, these networks share the same structure and can be denoted as \\( f_{i}(\\cdot) \\in \\mathcal{F} \\), where \\( \\mathcal{F} \\) stands for the function space within which these networks reside. In a semi-supervised process, the training data includes a small number of labeled data denoted as \\( \\mathbf{D}_L = \\{(x_i^L, y_i^L)\\}_{i=1}^N \\), and a large amount of unlabeled data denoted" + } + ], + [ + { + "type": "page_number", + "bbox": [ + 0.218, + 0.116, + 0.23, + 0.127 + ], + "angle": 0, + "content": "6" + }, + { + "type": "header", + "bbox": [ + 0.272, + 0.115, + 0.361, + 0.128 + ], + "angle": 0, + "content": "N. Gao et al." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.145, + 0.788, + 0.21 + ], + "angle": 0, + "content": "as \\(\\mathbf{D}_U = \\{(x_i^U)\\}_{i = N + 1}^{N + M}\\), where \\(N\\ll M\\), \\(x_{i}\\in \\mathbb{R}^{H\\times W\\times D}\\) represents medical volumes, and \\(y_{i}\\in \\{0,1\\}^{H\\times W\\times D}\\) represents ground truth labels. Batches of input data \\(\\mathbf{X}\\) consist of an equal proportion of labeled data \\((\\mathbf{X}^L,\\mathbf{Y}^L)\\) and unlabeled data \\(\\mathbf{X}^U\\). These volumes are fed into \\(f_{i}(\\cdot)\\) and \\(f_{i + 1}(\\cdot)\\):" + }, + { + "type": "equation", + "bbox": [ + 0.389, + 0.222, + 0.786, + 0.24 + ], + "angle": 0, + "content": "\\[\n\\hat {\\mathbf {Y}} _ {i} = f _ {i} (\\mathbf {X}), \\quad \\hat {\\mathbf {Y}} _ {i + 1} = f _ {i + 1} (\\mathbf {X}). \\tag {1}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.244, + 0.786, + 0.289 + ], + "angle": 0, + "content": "The output consists of volume predictions for both labeled and unlabeled data: \\(\\hat{\\mathbf{Y}} = \\hat{\\mathbf{Y}}^L\\cup \\hat{\\mathbf{Y}}^U\\). For the sake of simplicity, the network index subscripts are omitted here." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.29, + 0.789, + 0.427 + ], + "angle": 0, + "content": "We have ingeniously reintroduced MT architecture into the framework of semi-supervised learning. The introduction serves the purpose that, while providing high-fidelity pseudo labels to the model, the model's architecture needs to support the stable improvement of its representational capacity, aiming for better performance, while maintaining stable diversity. The teacher network is structurally identical to the student network but does not actively participate in the training process. Instead, it updates all of its parameters through EMA and applies consistency regularization to the student model through its output. The parameter is updated as follows:" + }, + { + "type": "equation", + "bbox": [ + 0.373, + 0.442, + 0.785, + 0.457 + ], + "angle": 0, + "content": "\\[\n\\theta_ {\\text {t e a c h e r}} = \\alpha \\theta_ {\\text {t e a c h e r}} + (1 - \\alpha) \\theta_ {\\text {s t u d e n t}}, \\tag {2}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.462, + 0.786, + 0.492 + ], + "angle": 0, + "content": "where \\(\\theta_{\\mathrm{teacher}}\\) represents the parameters of teacher network, \\(\\theta_{\\mathrm{student}}\\) represents the parameters of student network, and \\(\\alpha\\) is the EMA decay rate." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.492, + 0.788, + 0.582 + ], + "angle": 0, + "content": "In practice, we propose a cross-temporal training approach, which involves introducing a phase shift among different models during iterations and training the most lagging model in terms of iterations. To ensure that each model can train across temporal, the number of iterations for the model recorded in the sequence of iterations is denoted as \\(\\mathcal{I}_i\\). Our model maintains a static iteration gap through the data buffer length \\(\\mathcal{B}\\):" + }, + { + "type": "equation", + "bbox": [ + 0.325, + 0.594, + 0.786, + 0.625 + ], + "angle": 0, + "content": "\\[\n\\forall i \\in [ 1, n - 1 ], \\operatorname {L a r} (\\mathcal {I}, i) - \\operatorname {L a r} (\\mathcal {I}, i + 1) = \\frac {\\mathcal {B}}{(n - 1)}, \\tag {3}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.629, + 0.623, + 0.644 + ], + "angle": 0, + "content": "where, \\(\\operatorname{Lar}(\\mathcal{I}, i)\\) represents the \\(i\\)-th largest number in \\(\\mathcal{I}\\)." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.645, + 0.789, + 0.841 + ], + "angle": 0, + "content": "During each training iteration, the model with the least advanced iteration progress, referred to as the Current Progressive Model (CPM), sequentially utilizes data from the buffer for training until it has advanced ahead of the model with the most advanced iteration progress \\( \\mathcal{B} / n \\) times. Throughout the training process, the PLF screens out pseudo-labels generated by models whose performance is inferior to that of the CPM by measuring the performance gap between the CPM and other models, and learns from the remaining pseudo-labels. Simultaneously, the model computes the DDA with respect to a given input \\( \\mathbf{X} \\) and, along with all other models, examines regions of inconsistent predictions among different networks. When the performance of the CPM lags behind that of other models, the DDA corrects the CPM, enabling it to quickly align its performance with models that outperform it. It is worth noting that if the CPM is the best-performing model for that iteration, it will not guide other models in reverse. It" + } + ], + [ + { + "type": "header", + "bbox": [ + 0.557, + 0.115, + 0.732, + 0.129 + ], + "angle": 0, + "content": "Progressive Mean Teacher" + }, + { + "type": "page_number", + "bbox": [ + 0.775, + 0.116, + 0.785, + 0.127 + ], + "angle": 0, + "content": "7" + }, + { + "type": "image", + "bbox": [ + 0.337, + 0.151, + 0.497, + 0.281 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.408, + 0.282, + 0.428, + 0.295 + ], + "angle": 0, + "content": "(a)" + }, + { + "type": "image", + "bbox": [ + 0.502, + 0.151, + 0.664, + 0.281 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.574, + 0.282, + 0.593, + 0.295 + ], + "angle": 0, + "content": "(b)" + }, + { + "type": "image_caption", + "bbox": [ + 0.215, + 0.307, + 0.785, + 0.336 + ], + "angle": 0, + "content": "Fig. 3: Illustration of PLF and DDA. PLF is utilized for comparing the representational capacity of models, while DDA aligns model outputs by examining differences." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.368, + 0.784, + 0.399 + ], + "angle": 0, + "content": "receives guidance during the training process and only guides others once they have completed training." + }, + { + "type": "title", + "bbox": [ + 0.215, + 0.426, + 0.453, + 0.442 + ], + "angle": 0, + "content": "3.2 Pseudo Label Filtering" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.457, + 0.788, + 0.699 + ], + "angle": 0, + "content": "To guide the learning of the current progressive model effectively, we introduce PLF, as shown in Fig. 3 (a). PLF involves a naive approximation where we consider the representational capacity of models during the supervised learning phase as a criterion for judging the representational capacity during the unsupervised learning phase to some extent. During the supervised learning phase, PLF entails inference by all models, and through combining labels, calculates the representational capacity of each model in the current iteration. The metrics for evaluating representational capacity vary with tasks; in this task, the Dice loss is regarded as a statistical measure of model performance. We utilize the representational capacity of the CPM as a threshold and filter out pseudo-labels generated by models with representational capacities below that of the CPM. The pseudo label's passage through the PLF is recorded as \\(\\mathcal{P}(f_i,f_{i + 1})\\), where a value of 1 denotes passage and 0 denotes exclusion. This aids in the model's better learning of pseudo-labels with high fidelity. Furthermore, to ensure that the pseudo-labels are accurate and well-defined, we apply a sharpening function to high-fidelity pseudo-labels:" + }, + { + "type": "equation", + "bbox": [ + 0.397, + 0.713, + 0.787, + 0.752 + ], + "angle": 0, + "content": "\\[\n\\mathbf {Y} ^ {U} = \\frac {\\mathbf {P _ {s}} ^ {1 / T}}{\\mathbf {P _ {s}} ^ {1 / T} + (1 - \\mathbf {P _ {s}}) ^ {1 / T}}, \\tag {4}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.215, + 0.759, + 0.784, + 0.774 + ], + "angle": 0, + "content": "where \\(\\mathbf{P_s}\\) represents the output of the good student, and \\(T\\) is a hyperparameter." + }, + { + "type": "text", + "bbox": [ + 0.215, + 0.775, + 0.786, + 0.806 + ], + "angle": 0, + "content": "The resulting loss function \\(\\mathrm{L_U}\\) is computed as MSE between the pseudo labels \\(\\mathbf{Y}^U\\) and the model's predictions of CPM \\(\\hat{\\mathbf{Y}}_c^U\\):" + }, + { + "type": "equation", + "bbox": [ + 0.421, + 0.823, + 0.786, + 0.842 + ], + "angle": 0, + "content": "\\[\n\\mathrm {L} _ {\\mathrm {U}} = \\operatorname {M S E} \\left(\\hat {\\mathbf {Y}} _ {c} ^ {U}, \\mathbf {Y} ^ {U}\\right). \\tag {5}\n\\]" + } + ], + [ + { + "type": "page_number", + "bbox": [ + 0.218, + 0.116, + 0.23, + 0.127 + ], + "angle": 0, + "content": "8" + }, + { + "type": "header", + "bbox": [ + 0.272, + 0.115, + 0.361, + 0.128 + ], + "angle": 0, + "content": "N. Gao et al." + }, + { + "type": "title", + "bbox": [ + 0.216, + 0.147, + 0.519, + 0.163 + ], + "angle": 0, + "content": "3.3 Discrepancy Driven Alignment" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.174, + 0.79, + 0.493 + ], + "angle": 0, + "content": "We propose DDA to compare the outputs of multiple models and achieve rapid alignment between models, as shown in Fig. 3 (b). During the alternating process of progressive Learning, the performance of two groups of models alternately improves, providing stable pseudo-labels to each other. However, concerns often arise due to the disappearance of the performance gap between models caused by random factors. While the expectation is for alternating performance leadership between models, in practice, it may turn into one model consistently leading. Therefore, a fast and high-confidence alignment module is needed. In traditional consistency regularization, contrastive algorithms enforce consistency among the outputs of different models. However, this approach often raises concerns because we are uncertain whether this consistency regularization guides the models in the right direction. DDA focuses on the different parts of predictions among models, which often reflect the differences in representational capabilities between models. We aim to enhance the consistency regularization specifically on these parts to achieve alignment of representational capabilities between models. It is worth noting that we expect the models aligned by DDA to have representation capabilities surpassing CPM. Therefore, DDA will only take effect when PLF allows other models to provide pseudo-labels. The mask of different parts of predictions among different models can be obtained by taking the difference between the union and intersection of the binary outputs of softmax outputs \\(\\hat{\\mathbf{Y}}^L\\) of multiple networks:" + }, + { + "type": "equation", + "bbox": [ + 0.357, + 0.505, + 0.788, + 0.547 + ], + "angle": 0, + "content": "\\[\n\\mathcal {M} _ {\\text {d i f f}} = \\bigcup_ {i = 1} ^ {n} \\operatorname {B I N A} \\left(\\hat {\\mathbf {Y}} _ {i} ^ {L}\\right) - \\bigcap_ {i = 1} ^ {N} \\operatorname {B I N A} \\left(\\hat {\\mathbf {Y}} _ {i} ^ {L}\\right). \\tag {6}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.216, + 0.552, + 0.79, + 0.585 + ], + "angle": 0, + "content": "Subsequently, MSE of the model's predictions is computed and denoted as \\(\\mathcal{M}_{\\mathrm{dist}}\\):" + }, + { + "type": "equation", + "bbox": [ + 0.411, + 0.596, + 0.787, + 0.622 + ], + "angle": 0, + "content": "\\[\n\\mathcal {M} _ {\\mathrm {d i s t}} = \\operatorname {M S E} \\left(\\hat {\\mathbf {Y}} _ {i} ^ {L}, \\mathbf {Y} ^ {L}\\right). \\tag {7}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.627, + 0.789, + 0.687 + ], + "angle": 0, + "content": "Next, the total count of elements in \\(\\mathcal{M}_{\\mathrm{diff}}\\) is calculated, and \\(\\mathcal{M}_{\\mathrm{diff}}\\) is used to mask \\(\\mathcal{M}_{\\mathrm{dist}}\\). The ratio of the sum of the masked \\(\\mathcal{M}_{\\mathrm{dist}}\\) to the total count of elements in \\(\\mathcal{M}_{\\mathrm{diff}}\\) is denoted as \\(\\mathrm{L_{aln}}\\), which characterizes the gap to the ground truth within the prediction differences between two models:" + }, + { + "type": "equation", + "bbox": [ + 0.41, + 0.7, + 0.787, + 0.735 + ], + "angle": 0, + "content": "\\[\n\\mathrm {L} _ {\\mathrm {a l n}} = \\frac {\\sum \\left(\\mathcal {M} _ {\\mathrm {d i f f}} \\cdot \\mathcal {M} _ {\\mathrm {d i s t}}\\right)}{\\sum \\left(\\mathcal {M} _ {\\mathrm {d i f f}}\\right)}. \\tag {8}\n\\]" + }, + { + "type": "title", + "bbox": [ + 0.216, + 0.755, + 0.379, + 0.769 + ], + "angle": 0, + "content": "3.4 Loss Function" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.782, + 0.788, + 0.813 + ], + "angle": 0, + "content": "In general, during an iteration cycle, CPM conducts the process of backward propagation concerning its loss function, which is defined as follow:" + }, + { + "type": "equation", + "bbox": [ + 0.367, + 0.826, + 0.787, + 0.843 + ], + "angle": 0, + "content": "\\[\n\\mathrm {L} _ {\\text {t o t a l}} = \\mathrm {L} _ {\\mathrm {s}} + \\lambda_ {1} \\mathcal {P} (f _ {i}, f _ {i + 1}) \\mathrm {L} _ {\\mathrm {U}} + \\lambda_ {2} \\mathrm {L} _ {\\mathrm {T}}, \\tag {9}\n\\]" + } + ], + [ + { + "type": "header", + "bbox": [ + 0.556, + 0.115, + 0.732, + 0.129 + ], + "angle": 0, + "content": "Progressive Mean Teacher" + }, + { + "type": "page_number", + "bbox": [ + 0.775, + 0.116, + 0.787, + 0.127 + ], + "angle": 0, + "content": "9" + }, + { + "type": "text", + "bbox": [ + 0.215, + 0.147, + 0.784, + 0.177 + ], + "angle": 0, + "content": "where \\(\\lambda_{1}\\) and \\(\\lambda_{2}\\) increase as the number of iterations grows, up to a point where they stop growing after a fixed iteration." + }, + { + "type": "text", + "bbox": [ + 0.215, + 0.178, + 0.784, + 0.209 + ], + "angle": 0, + "content": "We employed two independent Gaussian warm-up function to control the loss function weights, \\(\\lambda_{1}\\) and \\(\\lambda_{2}\\), using different parameters:" + }, + { + "type": "equation", + "bbox": [ + 0.368, + 0.223, + 0.785, + 0.27 + ], + "angle": 0, + "content": "\\[\n\\lambda_ {1} (t) = \\left\\{ \\begin{array}{l l} \\hat {\\lambda} _ {1} \\cdot e ^ {- 5 \\left(1 - \\frac {2 t}{t _ {\\operatorname* {m a x}}}\\right) ^ {2}}, & t < \\frac {t _ {\\operatorname* {m a x}}}{2} \\\\ \\hat {\\lambda} _ {1}, & t \\geq \\frac {t _ {\\operatorname* {m a x}}}{2} \\end{array} \\right. \\tag {10}\n\\]" + }, + { + "type": "equation", + "bbox": [ + 0.368, + 0.268, + 0.635, + 0.306 + ], + "angle": 0, + "content": "\\[\n\\lambda_ {2} (t) = \\left\\{ \\begin{array}{l l} \\hat {\\lambda} _ {2} \\cdot e ^ {- 5 \\left(1 - \\frac {2 t}{t _ {\\max}}\\right) ^ {2}}, & t < \\frac {t _ {\\max}}{2} \\\\ \\hat {\\lambda} _ {2}, & t \\geq \\frac {t _ {\\max}}{2} \\end{array} \\right.\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.215, + 0.314, + 0.784, + 0.357 + ], + "angle": 0, + "content": "where \\( t \\) represents the current iteration number, and \\( t_{\\mathrm{max}} \\) represents the total number of training iterations. The hyperparameter \\( \\hat{\\lambda}_1 \\) and \\( \\hat{\\lambda}_2 \\) were empirically set to 20.0 and 10.0." + }, + { + "type": "text", + "bbox": [ + 0.215, + 0.36, + 0.784, + 0.39 + ], + "angle": 0, + "content": "Loss function \\( \\mathrm{L_s} \\) represents the loss during supervised learning and consists of the following components:" + }, + { + "type": "equation", + "bbox": [ + 0.324, + 0.409, + 0.784, + 0.423 + ], + "angle": 0, + "content": "\\[\n\\mathrm {L} _ {\\mathrm {s}} = \\mathrm {C E} + \\mathrm {D I C E} + \\beta \\mathcal {P} \\left(f _ {i}, f _ {i + 1}\\right) \\mathrm {L} _ {\\mathrm {a l n}} \\left(\\mathcal {M} _ {\\text {d i f f}}, \\mathcal {M} _ {\\text {d i s t}}\\right), \\tag {11}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.215, + 0.432, + 0.784, + 0.446 + ], + "angle": 0, + "content": "where \\(\\beta\\), a constant set to 0.5, is used to balance alignment loss and other losses." + }, + { + "type": "text", + "bbox": [ + 0.215, + 0.448, + 0.784, + 0.507 + ], + "angle": 0, + "content": "Loss function \\( \\mathrm{L}_{\\mathrm{T}} \\) is the consistency loss generated by the MT, which is derived by the teacher through the sharpen function to convert the model outputs into pseudo-labels, and subsequently generated by computing the MSE with the student model:" + }, + { + "type": "equation", + "bbox": [ + 0.337, + 0.523, + 0.784, + 0.562 + ], + "angle": 0, + "content": "\\[\n\\mathrm {L} _ {\\mathrm {T}} = \\operatorname {M S E} \\left(f _ {i} \\left(\\mathbf {X} ^ {U}\\right), \\frac {\\mathbf {P} _ {\\mathbf {t}} ^ {1 / T}}{\\mathbf {P} _ {\\mathbf {t}} ^ {1 / T} + (1 - \\mathbf {P} _ {\\mathbf {t}}) ^ {1 / T}}\\right). \\tag {12}\n\\]" + }, + { + "type": "title", + "bbox": [ + 0.216, + 0.595, + 0.375, + 0.613 + ], + "angle": 0, + "content": "4 Experiments" + }, + { + "type": "title", + "bbox": [ + 0.216, + 0.63, + 0.457, + 0.645 + ], + "angle": 0, + "content": "4.1 Implementation Details" + }, + { + "type": "text", + "bbox": [ + 0.215, + 0.658, + 0.784, + 0.763 + ], + "angle": 0, + "content": "We selected the VNet [13] model as a baseline network, which performs well in conditions with limited data and is essentially a 3D convolutional version of UNet. During inference, we use the average of the outputs from two networks as the final prediction. Specifically, the SGD optimizer was used to update the network parameters with weight decay of 0.0001 and a momentum of 0.9. The initial learning rate was set to 0.01, divided by 10 every 2500 iterations, for a total of 6000 iterations." + }, + { + "type": "text", + "bbox": [ + 0.215, + 0.765, + 0.784, + 0.84 + ], + "angle": 0, + "content": "Following the practice in comparative literature [11, 12, 23, 30], our methods are trained for a fixed number of 6,000 iterations to obtain the final model. Additionally, our models all use a batch size of 4, with a labeled data quantity of 2. We tested the performance of the models, and all experiments were conducted on NVIDIA® GeForce A40 48GB running Ubuntu 20.04 and PyTorch 1.11.0." + } + ], + [ + { + "type": "page_number", + "bbox": [ + 0.218, + 0.116, + 0.236, + 0.127 + ], + "angle": 0, + "content": "10" + }, + { + "type": "header", + "bbox": [ + 0.272, + 0.115, + 0.361, + 0.127 + ], + "angle": 0, + "content": "N. Gao et al." + }, + { + "type": "title", + "bbox": [ + 0.216, + 0.147, + 0.439, + 0.16 + ], + "angle": 0, + "content": "4.2 Datasets and Metrics" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.177, + 0.784, + 0.236 + ], + "angle": 0, + "content": "In the experiment, we selected two datasets with different modalities and utilized four distinct metrics to assess the performance of the model. For each dataset, \\(80\\%\\) of the data was used as the training set and \\(20\\%\\) as the test set. The proportion of supervised data was determined based on the training set." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.238, + 0.785, + 0.357 + ], + "angle": 0, + "content": "LA Dataset. It [28] includes 100 3D gadolinium-enhanced MR imaging volumes of left atrial with an isotropic resolution of \\(0.625 \\times 0.625 \\times 0.625mm^3\\) and the corresponding ground truth labels. For pre-processing, we first normalize all volumes to zero mean and unit variance, then crop each 3D MRI volume with enlarged margins according to the targets. During training, the training volumes are randomly cropped to \\(112 \\times 112 \\times 80\\) as the model input. During inference, a sliding window of the same size is used to obtain segmentation results with a stride of \\(18 \\times 18 \\times 4\\)." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.36, + 0.785, + 0.464 + ], + "angle": 0, + "content": "Pancreas-NIH Dataset. It [17] provides 82 contrast-enhanced abdominal 3D CT volumes of pancreas with manual annotation. The size of each CT volume is \\(512 \\times 512 \\times D\\), where \\(D \\in [181, 466]\\). In pre-processing, we use the soft tissue CT window of \\([-120, 240]\\) HU, and we crop the CT scans centering at the pancreas region, and enlarge margins with 25 voxels. The training volumes are randomly cropped to \\(96 \\times 96 \\times 96\\) as the model input. During inference, a sliding window of the same size is used to obtain segmentation results with a stride of \\(16 \\times 16 \\times 16\\)." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.466, + 0.785, + 0.543 + ], + "angle": 0, + "content": "Metrics. Following [2, 11, 12, 23, 29, 30], we use four metrics to evaluate model performance, including regional sensitive metrics: Dice similarity coefficient (Dice) [30], Jaccard similarity coefficient (Jaccard) [12], and edge sensitive metrics: \\(95\\%\\) Hausdorff Distance (95HD) [29] and Average Surface Distance (ASD) [2]." + }, + { + "type": "title", + "bbox": [ + 0.216, + 0.568, + 0.388, + 0.584 + ], + "angle": 0, + "content": "4.3 Ablation Study" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.598, + 0.784, + 0.643 + ], + "angle": 0, + "content": "For simplicity, we conducted ablation experiments on LA dataset to evaluate our design choices for each component. For a reliable assessment, all other parts were kept consistent except for the component under investigation." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.645, + 0.785, + 0.84 + ], + "angle": 0, + "content": "Analysis of PMT Framework. The PMT architecture under progressive design is at the core of our work, wherein progressive design ensures the continuous generation of diverse pseudo-labels during the forward propagation process. Within the PMT architecture, PLF and DDA respectively denote the pseudolabel filtering and model alignment methods proposed in our paper, while MT ensures stable enhancement of model capability. We employ a non-progressive design co-training framework consisting of two VNet components as the baseline model. When adding PLF and DDA methods to the baseline model, the progressive design is simultaneously incorporated into the baseline model. Finally, we evaluate the impact of the MT architecture on model representation capability, with results presented in Table 1. Overall, compared to the baseline model, which already possesses good representation capability, the PMT framework greatly enhances the model's representation capability." + } + ], + [ + { + "type": "header", + "bbox": [ + 0.557, + 0.115, + 0.732, + 0.129 + ], + "angle": 0, + "content": "Progressive Mean Teacher" + }, + { + "type": "page_number", + "bbox": [ + 0.769, + 0.116, + 0.784, + 0.127 + ], + "angle": 0, + "content": "11" + }, + { + "type": "table_caption", + "bbox": [ + 0.276, + 0.145, + 0.726, + 0.159 + ], + "angle": 0, + "content": "Table 1: Ablation results about PMT framework on LA dataset" + }, + { + "type": "table", + "bbox": [ + 0.237, + 0.171, + 0.763, + 0.293 + ], + "angle": 0, + "content": "
MethodLabeledMetrics
PLFDDAMTDice↑Jaccard↑95HD↓ASD↓
---8(10%)88.2879.317.592.34
✓--8(10%)89.9081.736.141.72
-✓-8(10%)90.4382.605.981.66
✓✓-8(10%)90.4382.605.491.49
✓✓✓8(10%)90.8183.235.611.50
" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.316, + 0.788, + 0.455 + ], + "angle": 0, + "content": "Analysis of Regularization Strength. The hyperparameters \\(\\hat{\\lambda}_1\\) and \\(\\hat{\\lambda}_2\\) respectively characterize the regularization strength of the progressive methods and MT architecture on the model. We use a tenfold scale and scale \\(\\hat{\\lambda}_1\\) and \\(\\hat{\\lambda}_2\\) based on the hyperparameter specifications we use, as shown in Table 2. The results indicate that variations in the two parameters within a certain range do not significantly affect the model's performance, demonstrating a certain level of robustness of our model to parameter variations. Within a certain range, on the LA dataset with \\(10\\%\\) labeled data, our model achieves the best performance when \\(\\hat{\\lambda}_1\\) and \\(\\hat{\\lambda}_2\\) are 20.0 and 10.0 respectively." + }, + { + "type": "table_caption", + "bbox": [ + 0.251, + 0.478, + 0.751, + 0.494 + ], + "angle": 0, + "content": "Table 2: Ablation results about regularization strength on LA dataset" + }, + { + "type": "table", + "bbox": [ + 0.243, + 0.505, + 0.756, + 0.681 + ], + "angle": 0, + "content": "
λLabeledMetrics
λ1maxλ2maxDice↑Jaccard↑95HD↓ASD↓
2.01.08(10%)89.6681.336.351.84
2.010.08(10%)90.4782.665.581.64
2.0100.08(10%)87.7478.289.332.84
20.01.08(10%)90.6782.986.311.77
20.010.08(10%)90.8183.235.611.50
20.0100.08(10%)88.0078.739.052.78
200.01.08(10%)89.0880.4211.403.07
200.010.08(10%)89.7581.477.082.22
200.0100.08(10%)83.8072.5516.445.11
" + }, + { + "type": "title", + "bbox": [ + 0.216, + 0.727, + 0.536, + 0.743 + ], + "angle": 0, + "content": "4.4 Comparison with Other Methods" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.75, + 0.784, + 0.779 + ], + "angle": 0, + "content": "We compared our approach with previous state-of-the-art methods on LA dataset and Pancreas-NIH dataset." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.78, + 0.785, + 0.841 + ], + "angle": 0, + "content": "We chose VNet as baseline models for comparison. For the selected alternative models, we opted for UA-MT [30] with uncertainty estimation, SASSNet [11] focusing on the regularity of geometric shapes, DTC [12] with task-level regularization, BCP [2] using bidirectional CutMix [31], and MCF [23] with model-level" + } + ], + [ + { + "type": "page_number", + "bbox": [ + 0.218, + 0.116, + 0.236, + 0.127 + ], + "angle": 0, + "content": "12" + }, + { + "type": "header", + "bbox": [ + 0.272, + 0.115, + 0.361, + 0.127 + ], + "angle": 0, + "content": "N. Gao et al." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.147, + 0.783, + 0.192 + ], + "angle": 0, + "content": "regularization, with BCP and MCF being state-of-the-art results. Noting that, for BCP, we follow its parameter settings for pre-training 2,000 times and self-training 15,000 times." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.193, + 0.787, + 0.284 + ], + "angle": 0, + "content": "Comparison on LA Dataset. We conducted a cross-model comparison on the classic LA dataset. We tested with \\(5\\%\\) and \\(10\\%\\) of labeled data. Results of the experiments are presented in Table 3. To provide a more intuitive demonstration of the performance of various models on the LA dataset, we have selected some representative results for visualization, as illustrated in Fig. 4. Areas with inaccurate segmentation have been annotated accordingly." + }, + { + "type": "table_caption", + "bbox": [ + 0.248, + 0.307, + 0.755, + 0.322 + ], + "angle": 0, + "content": "Table 3: Comparison results on LA dataset with \\(5\\%\\) and \\(10\\%\\) labeled data" + }, + { + "type": "table", + "bbox": [ + 0.237, + 0.334, + 0.763, + 0.573 + ], + "angle": 0, + "content": "
MethodLabeledMetrics
Dice↑Jaccard↑95HD↓ASD↓
VNet4(5%)52.5539.6047.059.87
UA-MT4(5%)82.2670.9813.713.82
SASSNet4(5%)81.6069.6316.163.58
DTC4(5%)81.2569.3314.903.99
MCFSOTA4(5%)----
BCPSOTA4(5%)88.0278.727.902.15
PMT(Ours)4(5%)89.4781.046.451.86
VNet8(10%)82.7471.7213.353.26
UA-MT8(10%)86.2876.1118.714.63
SASSNet8(10%)85.2275.0911.182.89
MCFSOTA8(10%)88.7180.416.321.90
BCPSOTA8(10%)89.6281.316.811.76
PMT(Ours)8(10%)90.8183.235.611.50
" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.599, + 0.788, + 0.794 + ], + "angle": 0, + "content": "Our model outperformed previous models on all four metrics across \\(5\\%\\) and \\(10\\%\\) labeled data. At \\(5\\%\\) of the data, compared to the best results from previous work, our PMT showed a \\(1.45\\%\\) improvement in Dice, a \\(2.32\\%\\) improvement in Jaccard, a reduction of 1.45 in 95HD, and a reduction of 0.29 in ASD. Our PMT also showed a \\(1.19\\%\\) improvement in Dice, a \\(1.92\\%\\) improvement in Jaccard, a reduction of 1.20 in 95HD, and a reduction of 0.26 in ASD at \\(10\\%\\) of the data. It is noteworthy that our model maintains good performance even with a small amount of data. In comparison with state-of-the-art results, our model achieves a leading performance in the majority of metrics using only half of their data. For instance, we surpass MCF's performance at \\(10\\%\\) labeled data using only \\(5\\%\\) labeled data. The above indicates that our model achieves superior results compared to existing methods on differently proportioned annotated data in the task of segmenting the left atrium." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.796, + 0.787, + 0.84 + ], + "angle": 0, + "content": "Comparison on Pancreas-NIH Dataset. We conducted a cross-model comparison on the classic Pancreas dataset. Detailed results of the experiments are presented in Table 4. We tested with \\(10\\%\\) and \\(20\\%\\) of labeled data. To" + } + ], + [ + { + "type": "header", + "bbox": [ + 0.557, + 0.115, + 0.732, + 0.128 + ], + "angle": 0, + "content": "Progressive Mean Teacher" + }, + { + "type": "page_number", + "bbox": [ + 0.769, + 0.116, + 0.785, + 0.127 + ], + "angle": 0, + "content": "13" + }, + { + "type": "image", + "bbox": [ + 0.217, + 0.144, + 0.787, + 0.449 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.216, + 0.458, + 0.785, + 0.486 + ], + "angle": 0, + "content": "Fig. 4: 2D & 3D segmentation visualization of different semi-supervised methods under \\(10\\%\\) labeled on LA (upper) and pancreas (bottom) dataset." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.518, + 0.785, + 0.623 + ], + "angle": 0, + "content": "provide a more intuitive demonstration of the performance of various models on the Pancreas-NIH dataset, we have selected some representative results for visualization, as illustrated in Fig. 4. Areas with inaccurate segmentation have been annotated accordingly. It is worth noting that the results in the table clearly indicate that Pancreas-NIH dataset is significantly more challenging than LA dataset. Therefore, we increased \\(\\hat{\\lambda}_{1}\\) by a factor of two, resulting in improved performance." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.625, + 0.787, + 0.746 + ], + "angle": 0, + "content": "Our model outperformed previous models on all four metrics across \\(10\\%\\) and \\(20\\%\\) labeled data. At \\(10\\%\\) of the data, compared to the best results from previous work, our PMT showed a \\(7.17\\%\\) improvement in Dice, a \\(9.09\\%\\) improvement in Jaccard, a reduction of 6.35 in 95HD, and a reduction of 2.10 in ASD. Our PMT also showed a \\(0.31\\%\\) improvement in Dice, a \\(0.55\\%\\) improvement in Jaccard, and a reduction of 0.36 in ASD at \\(20\\%\\) of the data. The above indicates that our model achieves superior results compared to existing methods on differently proportioned annotated data in the task of segmenting the pancreas." + }, + { + "type": "title", + "bbox": [ + 0.216, + 0.774, + 0.359, + 0.789 + ], + "angle": 0, + "content": "5 Conclusion" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.81, + 0.785, + 0.84 + ], + "angle": 0, + "content": "In this paper, we propose a semi-supervised medical image segmentation framework named Progress Mean Teacher (PMT). PMT adopts a progressive" + } + ], + [ + { + "type": "page_number", + "bbox": [ + 0.218, + 0.116, + 0.236, + 0.127 + ], + "angle": 0, + "content": "14" + }, + { + "type": "header", + "bbox": [ + 0.272, + 0.115, + 0.361, + 0.127 + ], + "angle": 0, + "content": "N. Gao et al." + }, + { + "type": "table_caption", + "bbox": [ + 0.218, + 0.145, + 0.787, + 0.159 + ], + "angle": 0, + "content": "Table 4: Comparison results on Pancreas-NIH dataset with \\( {10}\\% \\) and \\( {20}\\% \\) labeled data" + }, + { + "type": "table", + "bbox": [ + 0.237, + 0.172, + 0.763, + 0.424 + ], + "angle": 0, + "content": "
MethodLabeledMetrics
Dice↑Jaccard↑95HD↓ASD↓
VNet6(10%)55.6041.7445.3318.63
UA-MT6(10%)66.3453.2117.214.57
SASSNet6(10%)68.7853.8619.026.26
DTC6(10%)69.2154.0617.215.95
BCPSOTA6(10%)73.8359.2412.713.72
MCFSOTA6(10%)----
PMT(Ours)6(10%)81.0068.336.361.62
VNet12(20%)72.3858.2619.355.89
UA-MT12(20%)76.1062.6210.842.43
SASSNet12(20%)77.6664.0810.933.05
DTC12(20%)78.2764.758.362.25
BCPSOTA12(20%)82.9170.976.432.25
MCFSOTA12(20%)75.0061.2711.593.27
PMT(Ours)12(20%)83.2271.527.601.89
" + }, + { + "type": "text", + "bbox": [ + 0.218, + 0.489, + 0.787, + 0.731 + ], + "angle": 0, + "content": "design training process, establishing temporal-level model alignment and pseudolabel filtering while leveraging a network architecture based on MT. The core idea of this model framework is to generate diverse pseudo-labels consistently during the backpropagation process by establishing representation capability differences caused by iteration gaps, thereby stabilizing and enhancing the model's representation capability. In addition to the progressive architecture, PMT employs two simple yet effective methods, Pseudo Label Filtering (PLF) and Discrepancy Driven Alignment (DDA). PLF utilizes the representation capability of the model to filter pseudo-labels, discarding low-fidelity ones detrimental to constraining, while DDA aligns the differences in model predictions, allowing lagging models to catch up with leading models rapidly. Results from ablation experiments demonstrate that each component of the PMT framework significantly enhances the model's performance. In comparative experiments with other methods, PMT achieves state-of-the-art performance in terms of accuracy, surpassing previous methods significantly, and maintains this advantage compared to other methods in situations with more limited data and more challenging tasks." + }, + { + "type": "text", + "bbox": [ + 0.218, + 0.749, + 0.787, + 0.84 + ], + "angle": 0, + "content": "Limitation and Future Work. Despite the outstanding performance of PMT, there is still room for further exploration of semi-supervised training architectures established through temporal consistency. Investigating more advanced and stable strategies under a progressive design paradigm, and assessing whether they can further enhance performance, are topics worthy of further research in the future." + } + ], + [ + { + "type": "header", + "bbox": [ + 0.557, + 0.115, + 0.732, + 0.129 + ], + "angle": 0, + "content": "Progressive Mean Teacher" + }, + { + "type": "page_number", + "bbox": [ + 0.769, + 0.117, + 0.785, + 0.127 + ], + "angle": 0, + "content": "15" + }, + { + "type": "title", + "bbox": [ + 0.218, + 0.145, + 0.393, + 0.163 + ], + "angle": 0, + "content": "Acknowledgments" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.174, + 0.788, + 0.234 + ], + "angle": 0, + "content": "This work was supported in part by National Natural Science Foundation of China under Grants 62088102, 12326608 and 62106192, Natural Science Foundation of Shaanxi Province under Grant 2022JC-41, and Fundamental Research Funds for the Central Universities under Grant XTR042021005." + }, + { + "type": "title", + "bbox": [ + 0.218, + 0.254, + 0.323, + 0.27 + ], + "angle": 0, + "content": "References" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.282, + 0.785, + 0.323 + ], + "angle": 0, + "content": "1. Achanta, R., Shaji, A., Smith, K., Lucchi, A., Fua, P., Süsstrunk, S.: Slic superpixels compared to state-of-the-art superpixel methods. IEEE transactions on pattern analysis and machine intelligence 34(11), 2274-2282 (2012) 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.324, + 0.785, + 0.377 + ], + "angle": 0, + "content": "2. Bai, Y., Chen, D., Li, Q., Shen, W., Wang, Y.: Bidirectional copy-paste for semi-supervised medical image segmentation. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 11514-11524 (2023) 10, 11" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.378, + 0.785, + 0.418 + ], + "angle": 0, + "content": "3. Cai, H., Li, S., Qi, L., Yu, Q., Shi, Y., Gao, Y.: Orthogonal annotation benefits barely-supervised medical image segmentation. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 3302-3311 (2023) 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.42, + 0.785, + 0.459 + ], + "angle": 0, + "content": "4. Chen, X., Yuan, Y., Zeng, G., Wang, J.: Semi-supervised semantic segmentation with cross pseudo supervision. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 2613-2622 (2021) 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.461, + 0.785, + 0.499 + ], + "angle": 0, + "content": "5. Chen, X., He, K.: Exploring simple siamese representation learning. In: Proceedings of the IEEE/CVF conference on computer vision and pattern recognition. pp. 15750-15758 (2021) 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.501, + 0.785, + 0.54 + ], + "angle": 0, + "content": "6. Dong-DongChen, W., WeiGao, Z.: Tri-net for semi-supervised deep learning. In: Proceedings of twenty-seventh international joint conference on artificial intelligence. pp. 2014-2020 (2018) 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.542, + 0.785, + 0.581 + ], + "angle": 0, + "content": "7. Jiao, R., Zhang, Y., Ding, L., Cai, R., Zhang, J.: Learning with limited annotations: a survey on deep semi-supervised learning for medical image segmentation. arXiv preprint arXiv:2207.14191 (2022) 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.582, + 0.785, + 0.636 + ], + "angle": 0, + "content": "8. Ke, Z., Qiu, D., Li, K., Yan, Q., Lau, R.W.: Guided collaborative training for pixelwise semi-supervised learning. In: Computer Vision-ECCV 2020: 16th European Conference, Glasgow, UK, August 23-28, 2020, Proceedings, Part XIII 16. pp. 429-445. Springer (2020) 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.637, + 0.785, + 0.663 + ], + "angle": 0, + "content": "9. Laine, S., Aila, T.: Temporal ensembling for semi-supervised learning. arXiv preprint arXiv:1610.02242 (2016) 2, 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.22, + 0.664, + 0.785, + 0.703 + ], + "angle": 0, + "content": "10. Lee, D.H., et al.: Pseudo-label: The simple and efficient semi-supervised learning method for deep neural networks. In: Workshop on challenges in representation learning, ICML. vol. 3, p. 896. Atlanta (2013) 2, 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.22, + 0.704, + 0.785, + 0.758 + ], + "angle": 0, + "content": "11. Li, S., Zhang, C., He, X.: Shape-aware semi-supervised 3d semantic segmentation for medical images. In: Medical Image Computing and Computer Assisted Intervention-MICCAI 2020: 23rd International Conference, Lima, Peru, October 4-8, 2020, Proceedings, Part I 23. pp. 552-561. Springer (2020) 4, 9, 10, 11" + }, + { + "type": "ref_text", + "bbox": [ + 0.22, + 0.759, + 0.785, + 0.799 + ], + "angle": 0, + "content": "12. Luo, X., Chen, J., Song, T., Wang, G.: Semi-supervised medical image segmentation through dual-task consistency. In: Proceedings of the AAAI conference on artificial intelligence. vol. 35, pp. 8801-8809 (2021) 4, 9, 10, 11" + }, + { + "type": "ref_text", + "bbox": [ + 0.22, + 0.8, + 0.785, + 0.84 + ], + "angle": 0, + "content": "13. Miletari, F., Navab, N., Ahmadi, S.A.: V-net: Fully convolutional neural networks for volumetric medical image segmentation. In: 2016 fourth international conference on 3D vision (3DV). pp. 565-571. IEEE (2016) 9" + }, + { + "type": "list", + "bbox": [ + 0.22, + 0.282, + 0.785, + 0.84 + ], + "angle": 0, + "content": null + } + ], + [ + { + "type": "page_number", + "bbox": [ + 0.218, + 0.116, + 0.236, + 0.127 + ], + "angle": 0, + "content": "16" + }, + { + "type": "header", + "bbox": [ + 0.272, + 0.115, + 0.361, + 0.127 + ], + "angle": 0, + "content": "N. Gao et al." + }, + { + "type": "ref_text", + "bbox": [ + 0.218, + 0.147, + 0.787, + 0.189 + ], + "angle": 0, + "content": "14. Na, J., Ha, J.W., Chang, H.J., Han, D., Hwang, W.: Switching temporary teachers for semi-supervised semantic segmentation. Advances in Neural Information Processing Systems 36 (2024) 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.218, + 0.19, + 0.787, + 0.232 + ], + "angle": 0, + "content": "15. Peiris, H., Hayat, M., Chen, Z., Egan, G., Harandi, M.: Uncertainty-guided dual-views for semi-supervised volumetric medical image segmentation. Nature Machine Intelligence 5(7), 724-738 (2023) 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.232, + 0.787, + 0.259 + ], + "angle": 0, + "content": "16. Rizve, M.N., Kardan, N., Shah, M.: Towards realistic semi-supervised learning. In: European Conference on Computer Vision. pp. 437-455. Springer (2022) 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.259, + 0.787, + 0.328 + ], + "angle": 0, + "content": "17. Roth, H.R., Lu, L., Farag, A., Shin, H.C., Liu, J., Turkbey, E.B., Summers, R.M.: Deeporgan: Multi-level deep convolutional networks for automated pancreas segmentation. In: Medical Image Computing and Computer-Assisted Intervention-MICCAI 2015: 18th International Conference, Munich, Germany, October 5-9, 2015, Proceedings, Part I 18. pp. 556-564. Springer (2015) 3, 10" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.328, + 0.787, + 0.37 + ], + "angle": 0, + "content": "18. Shen, Z., Cao, P., Yang, H., Liu, X., Yang, J., Zaiane, O.R.: Co-training with high-confidence pseudo labels for semi-supervised medical image segmentation. arXiv preprint arXiv:2301.04465 (2023) 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.371, + 0.787, + 0.412 + ], + "angle": 0, + "content": "19. Tarvainen, A., Valpola, H.: Mean teachers are better role models: Weight-averaged consistency targets improve semi-supervised deep learning results. Advances in neural information processing systems 30 (2017) 2, 4, 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.412, + 0.787, + 0.467 + ], + "angle": 0, + "content": "20. Wang, H., Li, X.: Dhc: Dual-debiased heterogeneous co-training framework for class-imbalanced semi-supervised medical image segmentation. In: International Conference on Medical Image Computing and Computer-Assisted Intervention. pp. 582-591. Springer (2023) 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.467, + 0.787, + 0.522 + ], + "angle": 0, + "content": "21. Wang, J., Zhou, S., Fang, C., Wang, L., Wang, J.: Meta corrupted pixels mining for medical image segmentation. In: Medical Image Computing and Computer Assisted Intervention-MICCAI 2020: 23rd International Conference, Lima, Peru, October 4-8, 2020, Proceedings, Part I 23. pp. 335-345. Springer (2020) 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.522, + 0.787, + 0.577 + ], + "angle": 0, + "content": "22. Wang, X., Yuan, Y., Guo, D., Huang, X., Cui, Y., Xia, M., Wang, Z., Bai, C., Chen, S.: Ssa-net: Spatial self-attention network for COVID-19 pneumonia infection segmentation with semi-supervised few-shot learning. Medical Image Analysis 79, 102459 (2022) 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.578, + 0.787, + 0.632 + ], + "angle": 0, + "content": "23. Wang, Y., Xiao, B., Bi, X., Li, W., Gao, X.: Mcf: Mutual correction framework for semi-supervised medical image segmentation. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 15651-15660 (2023) 2, 3, 4, 9, 10, 11" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.633, + 0.787, + 0.688 + ], + "angle": 0, + "content": "24. Wu, H., Wang, Z., Song, Y., Yang, L., Qin, J.: Cross-patch dense contrastive learning for semi-supervised segmentation of cellular nuclei in histopathologic images. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 11666-11675 (2022) 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.689, + 0.787, + 0.73 + ], + "angle": 0, + "content": "25. Xia, K., Wang, L., Zhou, S., Hua, G., Tang, W.: Learning from noisy pseudo labels for semi-supervised temporal action localization. In: Proceedings of the IEEE/CVF International Conference on Computer Vision. pp. 10160-10169 (2023) 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.73, + 0.787, + 0.771 + ], + "angle": 0, + "content": "26. Xie, Q., Luong, M.T., Hovy, E., Le, Q.V.: Self-training with noisy student improves imagenet classification. In: Proceedings of the IEEE/CVF conference on computer vision and pattern recognition. pp. 10687-10698 (2020) 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.772, + 0.787, + 0.813 + ], + "angle": 0, + "content": "27. Xin, X., Wang, J., Xie, R., Zhou, S., Huang, W., Zheng, N.: Semi-supervised person re-identification using multi-view clustering. Pattern Recognition 88, 285–297 (2019) 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.813, + 0.787, + 0.841 + ], + "angle": 0, + "content": "28. Xiong, Z., Xia, Q., Hu, Z., Huang, N., Bian, C., Zheng, Y., Vesal, S., Ravikumar, N., Maier, A., Yang, X., et al.: A global benchmark of algorithms for segmenting" + }, + { + "type": "list", + "bbox": [ + 0.217, + 0.147, + 0.787, + 0.841 + ], + "angle": 0, + "content": null + } + ], + [ + { + "type": "header", + "bbox": [ + 0.557, + 0.115, + 0.732, + 0.129 + ], + "angle": 0, + "content": "Progressive Mean Teacher" + }, + { + "type": "page_number", + "bbox": [ + 0.769, + 0.116, + 0.786, + 0.127 + ], + "angle": 0, + "content": "17" + }, + { + "type": "ref_text", + "bbox": [ + 0.245, + 0.147, + 0.785, + 0.175 + ], + "angle": 0, + "content": "the left atrium from late gadolinium-enhanced cardiac magnetic resonance imaging. Medical image analysis 67, 101832 (2021) 3, 10" + }, + { + "type": "ref_text", + "bbox": [ + 0.218, + 0.176, + 0.787, + 0.231 + ], + "angle": 0, + "content": "29. Xu, Z., Wang, Y., Lu, D., Yu, L., Yan, J., Luo, J., Ma, K., Zheng, Y., Tong, R.K.y.: All-around real label supervision: Cyclic prototype consistency learning for semi-supervised medical image segmentation. IEEE Journal of Biomedical and Health Informatics 26(7), 3174-3184 (2022) 4, 10" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.232, + 0.786, + 0.3 + ], + "angle": 0, + "content": "30. Yu, L., Wang, S., Li, X., Fu, C.W., Heng, P.A.: Uncertainty-aware self-ensembling model for semi-supervised 3d left atrium segmentation. In: Medical Image Computing and Computer Assisted Intervention-MICCAI 2019: 22nd International Conference, Shenzhen, China, October 13–17, 2019, Proceedings, Part II 22. pp. 605–613. Springer (2019) 4, 9, 10, 11" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.3, + 0.786, + 0.342 + ], + "angle": 0, + "content": "31. Yun, S., Han, D., Oh, S.J., Chun, S., Choe, J., Yoo, Y.: Cutmix: Regularization strategy to train strong classifiers with localizable features. In: Proceedings of the IEEE/CVF international conference on computer vision. pp. 6023-6032 (2019) 11" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.342, + 0.786, + 0.395 + ], + "angle": 0, + "content": "32. Zheng, M., You, S., Huang, L., Wang, F., Qian, C., Xu, C.: Simmatch: Semi-supervised learning with similarity matching. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 14471-14481 (2022) 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.397, + 0.786, + 0.44 + ], + "angle": 0, + "content": "33. Zhou, S., Wang, J., Shu, J., Meng, D., Wang, L., Zheng, N.: Multinetwork collaborative feature learning for semisupervised person reidentification. IEEE Transactions on Neural Networks and Learning Systems 33(9), 4826-4839 (2021) 1" + }, + { + "type": "list", + "bbox": [ + 0.217, + 0.147, + 0.787, + 0.44 + ], + "angle": 0, + "content": null + } + ] +] \ No newline at end of file diff --git a/2024/PMT_ Progressive Mean Teacher via Exploring Temporal Consistency for Semi-Supervised Medical Image Segmentation/049d7a98-0307-4aa9-ae68-729601a51f9f_origin.pdf b/2024/PMT_ Progressive Mean Teacher via Exploring Temporal Consistency for Semi-Supervised Medical Image Segmentation/049d7a98-0307-4aa9-ae68-729601a51f9f_origin.pdf new file mode 100644 index 0000000000000000000000000000000000000000..2f7e1ca89f0e5cf24f0d47e9635e08b7108be787 --- /dev/null +++ b/2024/PMT_ Progressive Mean Teacher via Exploring Temporal Consistency for Semi-Supervised Medical Image Segmentation/049d7a98-0307-4aa9-ae68-729601a51f9f_origin.pdf @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:4678d198b05670634e59a69b0d97d068a3835ac000cee9762c01180702c13a98 +size 1630997 diff --git a/2024/PMT_ Progressive Mean Teacher via Exploring Temporal Consistency for Semi-Supervised Medical Image Segmentation/full.md b/2024/PMT_ Progressive Mean Teacher via Exploring Temporal Consistency for Semi-Supervised Medical Image Segmentation/full.md new file mode 100644 index 0000000000000000000000000000000000000000..094c75bb06384984c36f0745d1780f8f2332c3a1 --- /dev/null +++ b/2024/PMT_ Progressive Mean Teacher via Exploring Temporal Consistency for Semi-Supervised Medical Image Segmentation/full.md @@ -0,0 +1,276 @@ +# PMT: Progressive Mean Teacher via Exploring Temporal Consistency for Semi-Supervised Medical Image Segmentation + +Ning Gao $^{\text{D}}$ , Sanping Zhou $^{*}$ , Le Wang $^{\text{D}}$ , and Nanning Zheng $^{\text{D}}$ + +National Key Laboratory of Human-Machine Hybrid Augmented Intelligence, National Engineering Research Center for Visual Information and Applications, Institute of Artificial Intelligence and Robotics, Xi'an Jiaotong University + +Abstract. Semi-supervised learning has emerged as a widely adopted technique in the field of medical image segmentation. The existing works either focuses on the construction of consistency constraints or the generation of pseudo labels to provide high-quality supervisory signals, whose main challenge mainly comes from how to keep the continuous improvement of model capabilities. In this paper, we propose a simple yet effective semi-supervised learning framework, termed Progressive Mean Teachers (PMT), for medical image segmentation, whose goal is to generate high-fidelity pseudo labels by learning robust and diverse features in the training process. Specifically, our PMT employs a standard mean teacher to penalize the consistency of the current state and utilizes two sets of MT architectures for co-training. The two sets of MT architectures are individually updated for prolonged periods to maintain stable model diversity established through performance gaps generated by iteration differences. Additionally, a difference-driven alignment regularizer is employed to expedite the alignment of lagging models with the representation capabilities of leading models. Furthermore, a simple yet effective pseudo-label filtering algorithm is employed for facile evaluation of models and selection of high-fidelity pseudo-labels outputted when models are operating at high performance for co-training purposes. Experimental results on two datasets with different modalities, i.e., CT and MRI, demonstrate that our method outperforms the state-of-the-art medical image segmentation approaches across various dimensions. The code is available at https://github.com/Axi404/PMT. + +Keywords: Semi-supervised learning $\cdot$ Medical image segmentation $\cdot$ Temporal consistency regularization + +# 1 Introduction + +Semi-supervised learning is an important field in deep learning, which offers an effective way to tackle problems with limited labeled data [16, 25, 27, 32, 33]. With the continuous emergence of large-scale data, semi-supervised learning has + +![](images/062538e6c64842cb76ba27d2a32d430a0131c159211272c07e580bf7119e48bf.jpg) +Fig. 1: Motivation of our PMT. In particular, the standard MT helps to learn robust features by keeping the consistency between teacher and student networks at the current iteration, while our PMT further helps to learn diverse features by maintaining the difference between student networks at different iterations. As a result, more and more high-fidelity pseudo labels will be generated for semi-supervised medical image segmentation. + +become a hot topic in both machine learning and computer vision communities. It has gained particular attention in the medical image segmentation [21], due to the need for expert annotation in determining the accurate boundaries of targets. As a result, the labeled data is very lacking in medical image segmentation, which makes the semi-supervised learning very popular in this domain. + +Most of the semi-supervised learning methods involve seeking a guidance mechanism using limited annotated data, so as to maximize the usage of unlabeled data to train the models on different downstream tasks. In general, two technologies, i.e., consistency regularization [9, 10] and pseudo label generation [10, 19], are commonly used in semi-supervised medical segmentation. The former ones aim at enhancing the representation capability of networks by exploring the consistency based on different prior assumptions [5], while the latter ones focus on improving the performance on downstream tasks by generating the pseudo labels based on current model capability. For example, the well-known Mean Teacher (MT) [19] utilizes an ingenious Exponential Moving Average (EMA) that is equivalent to data augmentation and enforces consistency regularization between the outputs of teacher and student models. Besides, the representative Noisy Student [26] designs a teacher model to generate pseudo labels, so as to train a larger student model. + +Even though the above methods have achieved significant improvements in semi-supervised medical image segmentation, we still argue that they are weak in exploring high-quality supervisory signals to consistently enhance the model's capability. To address this issue, an iterative unified optimization framework has been introduced to semi-supervised medical image segmentation [20, 23, 24], in which the consistency regularization strategy is taken to enhance the model's + +representation capability and the pseudo label generation algorithm is applied to generate high-fidelity pseudo labels. For example, the recent MCF [23] takes VNet and 3D ResNet for representation learning, and heterogeneous networks for dynamic pseudo label generation. The challenges to these methods lie on how to continuously generate diverse pseudo labels in the forward propagation, and enhance model's capability in the backward propagation. + +In this paper, we design a novel semi-supervised learning framework, termed Progressive Mean Teachers (PMT), for medical image segmentation, whose main idea focuses on how to obtain a diverse set of accurate pseudo labels. Inspired by the positive relationship between iteration and performance, we try to maintain the diversity of networks by exploring their states at different training epochs. Specifically, the standard MT is first taken as basic architecture to learn the parameters of network at each iteration, which can enhance the network's representation capability by using the EMA data augmentation. Our PMT further explores network diversity during training, alternating between two homogeneous MT architectures trained on the same dataset, a process we term progressive design. These models exhibit significant iteration leads due to alternating continuous individual updates, establishing performance gaps between networks at different epochs, as shown in Fig. 1. Consequently, the student network can acquire robust yet diverse features for medical image segmentation. Then, the Discrepancy Driven Alignment (DDA) regularizer is further designed to examine disparities between predictions obtained by the student network at different training epochs, facilitating rapid alignment to high-fidelity generated images. Finally, we design a simple Pseudo Label Filtering (PLF) algorithm to refine the basic interaction process, enabling the retention of high-fidelity pseudo-labels for training by comparing student network performance across different training epochs. As a result, more and more high-fidelity pseudo labels can be fed to train the other student networks, which will in turn generate more accurate predictions for pseudo label generation. + +In summary, the main contributions of this work are as follows: (1) We design a novel Progressive Mean Teacher framework for semi-supervised medical image segmentation. (2) We design a novel Discrepancy Driven Alignment regularizer to rapidly align the representational capacity gap between lagging and leading networks. (3) We design a simple yet effective Pseudo Label Filtering algorithm to select high-fidelity pseudo labels. Extensive experiments in both Left Atrial [28] and Pancreas-NIH [17] datasets show that our PMT can achieve the state-of-the-art results in semi-supervised medical image segmentation. + +# 2 Related Work + +# 2.1 Consistency Regularization + +Consistency regularization is often employed in semi-supervised learning, so as to enhance the stable representational capacity of model. The behind idea is to preserve the invariance of predictions made by the same model when facing + +perturbations applied in different regions, by constraining the consistency between output results under different perturbations. For example, $\varPi$ model [9] introduces image-level regularization by adding perturbations to images. The well-known MT [19] introduces parameter-level regularization by using EMA to constrain the outputs between teacher and student models. Thanks to its simplicity and effectiveness, more and more works are paying attention to improve the generalization ability by formulating different consistency regularizers. For example, SASSnet [11] focuses on the regularity of geometric shapes for target object classes within consistency regularization. Besides, CPCL [29] establishes regularization between supervised and unsupervised training within a cyclic framework. Furthermore, some later works have made progress in regularization at different task and model levels. For example, DTC [12] introduces task-level regularization, presenting a novel dual-task consistency semi-supervised framework. Besides, MCF [23] introduces model-level regularization by using heterogeneous models to constrain output consistency. Compared to previous work, our PMT seeks the cross-temporal regularization between different training periods, which can help learn diverse yet robust features for subsequent pseudo label generation. + +# 2.2 Pseudo Label Generation + +Pseudo label generation is often employed in semi-supervised learning, so as to enhance the discriminative ability of model. The behind idea is to train a prior model with the labeled data and then apply it to generate the pseudo labels for unlabeled data. It is widely accepted to categorize pseudo-label generation into two methods [7], direct generation focuses on selecting pseudo labels with higher confidence, while indirect generation explores methods to generate high-fidelity pseudo labels. For direct generation, work [10] uses a fixed threshold to choose high-confidence pseudo labels. SsaNet [22] employs a trust module to reevaluate pseudo labels. UA-MT [30] introduces uncertainty estimation to filter out unreliable pseudo labels. Co-BioNet [15] introduces a feedback network to measure the uncertainty and choose high-confidence predictions of different models. For indirect generation, Tri-Net [6] is proposed to use two subnetworks to generate pseudo labels for a third subnetwork. Work [1] improves pseudo label generation using Simple Linear Iterative Clustering (SLIC) algorithm. MCF [23] dynamically generates pseudo labels using a heterogeneous network and effectively addresses cognitive bias, while DeSCO [3] focuses on the spatial correlation of medical images and generates pseudo labels using orthogonal slices. Compared to previous work, our PMT emphasizes the enhancement of pseudo label quality in the temporal domain and generates diverse and robust pseudo labels across different training periods, significantly improving performance. + +# 2.3 Multi-Model Framework + +The multi-model framework is often employed in semi-supervised learning, so as to enhance the model representation by acquiring multiple views or diversity. Its development aligns closely with consistency regularization. Here, we + +![](images/fdeffe67ee6516f4a7f3439e0d10a005ee9fc08aa83cf44a01c443edf12a99a2.jpg) +Fig. 2: An overview of PMT. We employed a progressive design and utilized the architecture of MT. The PMT framework maintains a data buffer of length $\mathcal{B}$ for cros-temporal training. The total loss function $\mathrm{L}_{\mathrm{total}}$ for each network includes supervised losses $\mathrm{L_{CE}}, \mathrm{L_{DICE}}, \mathrm{L_{aln}}$ , and unsupervised loss $\mathrm{L_U}, \mathrm{L_T}$ . + +focus on the evolution of model structures. For example, MT [19] introduces a Teacher model with significantly improved representation capabilities at a lower cost using EMA, establishing the MT architecture. Besides, CPC [8] utilizes the confidence vector of multi-model outputs as pseudo-labels for co-training, while CPS [4] employs one-hot labels, both considered as starting points for co-training. In recent years, efforts have been made to combine these approaches. For example, UCMT [18] employs two student models for co-training, simultaneously updating the same Mean Teacher using EMA, while Dual Teacher [14] alternates updates between two Teacher models using EMA with a single student model, thereby fostering diversity among Teacher models. Compared to previous work, our PMT utilizes two sets of Mean Teachers in a progressive training framework, and employ two student models to update teacher models independently. This enables our model to rapidly establish robust performance disparities across iteration gap and maintain stable diversity among models. + +# 3 Method + +# 3.1 The Overall Process of PMT + +The training process of PMT is illustrated in Fig. 2. As previously mentioned, PMT can consist of multiple networks. For ease of explanation, the number of networks is assumed to be two by default in our work. In our approach, these networks share the same structure and can be denoted as $f_{i}(\cdot) \in \mathcal{F}$ , where $\mathcal{F}$ stands for the function space within which these networks reside. In a semi-supervised process, the training data includes a small number of labeled data denoted as $\mathbf{D}_L = \{(x_i^L, y_i^L)\}_{i=1}^N$ , and a large amount of unlabeled data denoted + +as $\mathbf{D}_U = \{(x_i^U)\}_{i = N + 1}^{N + M}$ , where $N\ll M$ , $x_{i}\in \mathbb{R}^{H\times W\times D}$ represents medical volumes, and $y_{i}\in \{0,1\}^{H\times W\times D}$ represents ground truth labels. Batches of input data $\mathbf{X}$ consist of an equal proportion of labeled data $(\mathbf{X}^L,\mathbf{Y}^L)$ and unlabeled data $\mathbf{X}^U$ . These volumes are fed into $f_{i}(\cdot)$ and $f_{i + 1}(\cdot)$ : + +$$ +\hat {\mathbf {Y}} _ {i} = f _ {i} (\mathbf {X}), \quad \hat {\mathbf {Y}} _ {i + 1} = f _ {i + 1} (\mathbf {X}). \tag {1} +$$ + +The output consists of volume predictions for both labeled and unlabeled data: $\hat{\mathbf{Y}} = \hat{\mathbf{Y}}^L\cup \hat{\mathbf{Y}}^U$ . For the sake of simplicity, the network index subscripts are omitted here. + +We have ingeniously reintroduced MT architecture into the framework of semi-supervised learning. The introduction serves the purpose that, while providing high-fidelity pseudo labels to the model, the model's architecture needs to support the stable improvement of its representational capacity, aiming for better performance, while maintaining stable diversity. The teacher network is structurally identical to the student network but does not actively participate in the training process. Instead, it updates all of its parameters through EMA and applies consistency regularization to the student model through its output. The parameter is updated as follows: + +$$ +\theta_ {\text {t e a c h e r}} = \alpha \theta_ {\text {t e a c h e r}} + (1 - \alpha) \theta_ {\text {s t u d e n t}}, \tag {2} +$$ + +where $\theta_{\mathrm{teacher}}$ represents the parameters of teacher network, $\theta_{\mathrm{student}}$ represents the parameters of student network, and $\alpha$ is the EMA decay rate. + +In practice, we propose a cross-temporal training approach, which involves introducing a phase shift among different models during iterations and training the most lagging model in terms of iterations. To ensure that each model can train across temporal, the number of iterations for the model recorded in the sequence of iterations is denoted as $\mathcal{I}_i$ . Our model maintains a static iteration gap through the data buffer length $\mathcal{B}$ : + +$$ +\forall i \in [ 1, n - 1 ], \operatorname {L a r} (\mathcal {I}, i) - \operatorname {L a r} (\mathcal {I}, i + 1) = \frac {\mathcal {B}}{(n - 1)}, \tag {3} +$$ + +where, $\operatorname{Lar}(\mathcal{I}, i)$ represents the $i$ -th largest number in $\mathcal{I}$ . + +During each training iteration, the model with the least advanced iteration progress, referred to as the Current Progressive Model (CPM), sequentially utilizes data from the buffer for training until it has advanced ahead of the model with the most advanced iteration progress $\mathcal{B} / n$ times. Throughout the training process, the PLF screens out pseudo-labels generated by models whose performance is inferior to that of the CPM by measuring the performance gap between the CPM and other models, and learns from the remaining pseudo-labels. Simultaneously, the model computes the DDA with respect to a given input $\mathbf{X}$ and, along with all other models, examines regions of inconsistent predictions among different networks. When the performance of the CPM lags behind that of other models, the DDA corrects the CPM, enabling it to quickly align its performance with models that outperform it. It is worth noting that if the CPM is the best-performing model for that iteration, it will not guide other models in reverse. It + +![](images/71cf79dd32c6620cdbcc9651461e85d067aaecdc4ff856c4fd342be0857c2784.jpg) +(a) + +![](images/137070e520425265a242a5cb371ba0abccefb748d358ce452af54722233ebb6e.jpg) +(b) +Fig. 3: Illustration of PLF and DDA. PLF is utilized for comparing the representational capacity of models, while DDA aligns model outputs by examining differences. + +receives guidance during the training process and only guides others once they have completed training. + +# 3.2 Pseudo Label Filtering + +To guide the learning of the current progressive model effectively, we introduce PLF, as shown in Fig. 3 (a). PLF involves a naive approximation where we consider the representational capacity of models during the supervised learning phase as a criterion for judging the representational capacity during the unsupervised learning phase to some extent. During the supervised learning phase, PLF entails inference by all models, and through combining labels, calculates the representational capacity of each model in the current iteration. The metrics for evaluating representational capacity vary with tasks; in this task, the Dice loss is regarded as a statistical measure of model performance. We utilize the representational capacity of the CPM as a threshold and filter out pseudo-labels generated by models with representational capacities below that of the CPM. The pseudo label's passage through the PLF is recorded as $\mathcal{P}(f_i,f_{i + 1})$ , where a value of 1 denotes passage and 0 denotes exclusion. This aids in the model's better learning of pseudo-labels with high fidelity. Furthermore, to ensure that the pseudo-labels are accurate and well-defined, we apply a sharpening function to high-fidelity pseudo-labels: + +$$ +\mathbf {Y} ^ {U} = \frac {\mathbf {P _ {s}} ^ {1 / T}}{\mathbf {P _ {s}} ^ {1 / T} + (1 - \mathbf {P _ {s}}) ^ {1 / T}}, \tag {4} +$$ + +where $\mathbf{P_s}$ represents the output of the good student, and $T$ is a hyperparameter. + +The resulting loss function $\mathrm{L_U}$ is computed as MSE between the pseudo labels $\mathbf{Y}^U$ and the model's predictions of CPM $\hat{\mathbf{Y}}_c^U$ : + +$$ +\mathrm {L} _ {\mathrm {U}} = \operatorname {M S E} \left(\hat {\mathbf {Y}} _ {c} ^ {U}, \mathbf {Y} ^ {U}\right). \tag {5} +$$ + +# 3.3 Discrepancy Driven Alignment + +We propose DDA to compare the outputs of multiple models and achieve rapid alignment between models, as shown in Fig. 3 (b). During the alternating process of progressive Learning, the performance of two groups of models alternately improves, providing stable pseudo-labels to each other. However, concerns often arise due to the disappearance of the performance gap between models caused by random factors. While the expectation is for alternating performance leadership between models, in practice, it may turn into one model consistently leading. Therefore, a fast and high-confidence alignment module is needed. In traditional consistency regularization, contrastive algorithms enforce consistency among the outputs of different models. However, this approach often raises concerns because we are uncertain whether this consistency regularization guides the models in the right direction. DDA focuses on the different parts of predictions among models, which often reflect the differences in representational capabilities between models. We aim to enhance the consistency regularization specifically on these parts to achieve alignment of representational capabilities between models. It is worth noting that we expect the models aligned by DDA to have representation capabilities surpassing CPM. Therefore, DDA will only take effect when PLF allows other models to provide pseudo-labels. The mask of different parts of predictions among different models can be obtained by taking the difference between the union and intersection of the binary outputs of softmax outputs $\hat{\mathbf{Y}}^L$ of multiple networks: + +$$ +\mathcal {M} _ {\text {d i f f}} = \bigcup_ {i = 1} ^ {n} \operatorname {B I N A} \left(\hat {\mathbf {Y}} _ {i} ^ {L}\right) - \bigcap_ {i = 1} ^ {N} \operatorname {B I N A} \left(\hat {\mathbf {Y}} _ {i} ^ {L}\right). \tag {6} +$$ + +Subsequently, MSE of the model's predictions is computed and denoted as $\mathcal{M}_{\mathrm{dist}}$ : + +$$ +\mathcal {M} _ {\mathrm {d i s t}} = \operatorname {M S E} \left(\hat {\mathbf {Y}} _ {i} ^ {L}, \mathbf {Y} ^ {L}\right). \tag {7} +$$ + +Next, the total count of elements in $\mathcal{M}_{\mathrm{diff}}$ is calculated, and $\mathcal{M}_{\mathrm{diff}}$ is used to mask $\mathcal{M}_{\mathrm{dist}}$ . The ratio of the sum of the masked $\mathcal{M}_{\mathrm{dist}}$ to the total count of elements in $\mathcal{M}_{\mathrm{diff}}$ is denoted as $\mathrm{L_{aln}}$ , which characterizes the gap to the ground truth within the prediction differences between two models: + +$$ +\mathrm {L} _ {\mathrm {a l n}} = \frac {\sum \left(\mathcal {M} _ {\mathrm {d i f f}} \cdot \mathcal {M} _ {\mathrm {d i s t}}\right)}{\sum \left(\mathcal {M} _ {\mathrm {d i f f}}\right)}. \tag {8} +$$ + +# 3.4 Loss Function + +In general, during an iteration cycle, CPM conducts the process of backward propagation concerning its loss function, which is defined as follow: + +$$ +\mathrm {L} _ {\text {t o t a l}} = \mathrm {L} _ {\mathrm {s}} + \lambda_ {1} \mathcal {P} (f _ {i}, f _ {i + 1}) \mathrm {L} _ {\mathrm {U}} + \lambda_ {2} \mathrm {L} _ {\mathrm {T}}, \tag {9} +$$ + +where $\lambda_{1}$ and $\lambda_{2}$ increase as the number of iterations grows, up to a point where they stop growing after a fixed iteration. + +We employed two independent Gaussian warm-up function to control the loss function weights, $\lambda_{1}$ and $\lambda_{2}$ , using different parameters: + +$$ +\lambda_ {1} (t) = \left\{ \begin{array}{l l} \hat {\lambda} _ {1} \cdot e ^ {- 5 \left(1 - \frac {2 t}{t _ {\operatorname* {m a x}}}\right) ^ {2}}, & t < \frac {t _ {\operatorname* {m a x}}}{2} \\ \hat {\lambda} _ {1}, & t \geq \frac {t _ {\operatorname* {m a x}}}{2} \end{array} \right. \tag {10} +$$ + +$$ +\lambda_ {2} (t) = \left\{ \begin{array}{l l} \hat {\lambda} _ {2} \cdot e ^ {- 5 \left(1 - \frac {2 t}{t _ {\max}}\right) ^ {2}}, & t < \frac {t _ {\max}}{2} \\ \hat {\lambda} _ {2}, & t \geq \frac {t _ {\max}}{2} \end{array} \right. +$$ + +where $t$ represents the current iteration number, and $t_{\mathrm{max}}$ represents the total number of training iterations. The hyperparameter $\hat{\lambda}_1$ and $\hat{\lambda}_2$ were empirically set to 20.0 and 10.0. + +Loss function $\mathrm{L_s}$ represents the loss during supervised learning and consists of the following components: + +$$ +\mathrm {L} _ {\mathrm {s}} = \mathrm {C E} + \mathrm {D I C E} + \beta \mathcal {P} \left(f _ {i}, f _ {i + 1}\right) \mathrm {L} _ {\mathrm {a l n}} \left(\mathcal {M} _ {\text {d i f f}}, \mathcal {M} _ {\text {d i s t}}\right), \tag {11} +$$ + +where $\beta$ , a constant set to 0.5, is used to balance alignment loss and other losses. + +Loss function $\mathrm{L}_{\mathrm{T}}$ is the consistency loss generated by the MT, which is derived by the teacher through the sharpen function to convert the model outputs into pseudo-labels, and subsequently generated by computing the MSE with the student model: + +$$ +\mathrm {L} _ {\mathrm {T}} = \operatorname {M S E} \left(f _ {i} \left(\mathbf {X} ^ {U}\right), \frac {\mathbf {P} _ {\mathbf {t}} ^ {1 / T}}{\mathbf {P} _ {\mathbf {t}} ^ {1 / T} + (1 - \mathbf {P} _ {\mathbf {t}}) ^ {1 / T}}\right). \tag {12} +$$ + +# 4 Experiments + +# 4.1 Implementation Details + +We selected the VNet [13] model as a baseline network, which performs well in conditions with limited data and is essentially a 3D convolutional version of UNet. During inference, we use the average of the outputs from two networks as the final prediction. Specifically, the SGD optimizer was used to update the network parameters with weight decay of 0.0001 and a momentum of 0.9. The initial learning rate was set to 0.01, divided by 10 every 2500 iterations, for a total of 6000 iterations. + +Following the practice in comparative literature [11, 12, 23, 30], our methods are trained for a fixed number of 6,000 iterations to obtain the final model. Additionally, our models all use a batch size of 4, with a labeled data quantity of 2. We tested the performance of the models, and all experiments were conducted on NVIDIA® GeForce A40 48GB running Ubuntu 20.04 and PyTorch 1.11.0. + +# 4.2 Datasets and Metrics + +In the experiment, we selected two datasets with different modalities and utilized four distinct metrics to assess the performance of the model. For each dataset, $80\%$ of the data was used as the training set and $20\%$ as the test set. The proportion of supervised data was determined based on the training set. + +LA Dataset. It [28] includes 100 3D gadolinium-enhanced MR imaging volumes of left atrial with an isotropic resolution of $0.625 \times 0.625 \times 0.625mm^3$ and the corresponding ground truth labels. For pre-processing, we first normalize all volumes to zero mean and unit variance, then crop each 3D MRI volume with enlarged margins according to the targets. During training, the training volumes are randomly cropped to $112 \times 112 \times 80$ as the model input. During inference, a sliding window of the same size is used to obtain segmentation results with a stride of $18 \times 18 \times 4$ . + +Pancreas-NIH Dataset. It [17] provides 82 contrast-enhanced abdominal 3D CT volumes of pancreas with manual annotation. The size of each CT volume is $512 \times 512 \times D$ , where $D \in [181, 466]$ . In pre-processing, we use the soft tissue CT window of $[-120, 240]$ HU, and we crop the CT scans centering at the pancreas region, and enlarge margins with 25 voxels. The training volumes are randomly cropped to $96 \times 96 \times 96$ as the model input. During inference, a sliding window of the same size is used to obtain segmentation results with a stride of $16 \times 16 \times 16$ . + +Metrics. Following [2, 11, 12, 23, 29, 30], we use four metrics to evaluate model performance, including regional sensitive metrics: Dice similarity coefficient (Dice) [30], Jaccard similarity coefficient (Jaccard) [12], and edge sensitive metrics: $95\%$ Hausdorff Distance (95HD) [29] and Average Surface Distance (ASD) [2]. + +# 4.3 Ablation Study + +For simplicity, we conducted ablation experiments on LA dataset to evaluate our design choices for each component. For a reliable assessment, all other parts were kept consistent except for the component under investigation. + +Analysis of PMT Framework. The PMT architecture under progressive design is at the core of our work, wherein progressive design ensures the continuous generation of diverse pseudo-labels during the forward propagation process. Within the PMT architecture, PLF and DDA respectively denote the pseudolabel filtering and model alignment methods proposed in our paper, while MT ensures stable enhancement of model capability. We employ a non-progressive design co-training framework consisting of two VNet components as the baseline model. When adding PLF and DDA methods to the baseline model, the progressive design is simultaneously incorporated into the baseline model. Finally, we evaluate the impact of the MT architecture on model representation capability, with results presented in Table 1. Overall, compared to the baseline model, which already possesses good representation capability, the PMT framework greatly enhances the model's representation capability. + +Table 1: Ablation results about PMT framework on LA dataset + +
MethodLabeledMetrics
PLFDDAMTDice↑Jaccard↑95HD↓ASD↓
---8(10%)88.2879.317.592.34
✓--8(10%)89.9081.736.141.72
-✓-8(10%)90.4382.605.981.66
✓✓-8(10%)90.4382.605.491.49
✓✓✓8(10%)90.8183.235.611.50
+ +Analysis of Regularization Strength. The hyperparameters $\hat{\lambda}_1$ and $\hat{\lambda}_2$ respectively characterize the regularization strength of the progressive methods and MT architecture on the model. We use a tenfold scale and scale $\hat{\lambda}_1$ and $\hat{\lambda}_2$ based on the hyperparameter specifications we use, as shown in Table 2. The results indicate that variations in the two parameters within a certain range do not significantly affect the model's performance, demonstrating a certain level of robustness of our model to parameter variations. Within a certain range, on the LA dataset with $10\%$ labeled data, our model achieves the best performance when $\hat{\lambda}_1$ and $\hat{\lambda}_2$ are 20.0 and 10.0 respectively. + +Table 2: Ablation results about regularization strength on LA dataset + +
λLabeledMetrics
λ1maxλ2maxDice↑Jaccard↑95HD↓ASD↓
2.01.08(10%)89.6681.336.351.84
2.010.08(10%)90.4782.665.581.64
2.0100.08(10%)87.7478.289.332.84
20.01.08(10%)90.6782.986.311.77
20.010.08(10%)90.8183.235.611.50
20.0100.08(10%)88.0078.739.052.78
200.01.08(10%)89.0880.4211.403.07
200.010.08(10%)89.7581.477.082.22
200.0100.08(10%)83.8072.5516.445.11
+ +# 4.4 Comparison with Other Methods + +We compared our approach with previous state-of-the-art methods on LA dataset and Pancreas-NIH dataset. + +We chose VNet as baseline models for comparison. For the selected alternative models, we opted for UA-MT [30] with uncertainty estimation, SASSNet [11] focusing on the regularity of geometric shapes, DTC [12] with task-level regularization, BCP [2] using bidirectional CutMix [31], and MCF [23] with model-level + +regularization, with BCP and MCF being state-of-the-art results. Noting that, for BCP, we follow its parameter settings for pre-training 2,000 times and self-training 15,000 times. + +Comparison on LA Dataset. We conducted a cross-model comparison on the classic LA dataset. We tested with $5\%$ and $10\%$ of labeled data. Results of the experiments are presented in Table 3. To provide a more intuitive demonstration of the performance of various models on the LA dataset, we have selected some representative results for visualization, as illustrated in Fig. 4. Areas with inaccurate segmentation have been annotated accordingly. + +Table 3: Comparison results on LA dataset with $5\%$ and $10\%$ labeled data + +
MethodLabeledMetrics
Dice↑Jaccard↑95HD↓ASD↓
VNet4(5%)52.5539.6047.059.87
UA-MT4(5%)82.2670.9813.713.82
SASSNet4(5%)81.6069.6316.163.58
DTC4(5%)81.2569.3314.903.99
MCFSOTA4(5%)----
BCPSOTA4(5%)88.0278.727.902.15
PMT(Ours)4(5%)89.4781.046.451.86
VNet8(10%)82.7471.7213.353.26
UA-MT8(10%)86.2876.1118.714.63
SASSNet8(10%)85.2275.0911.182.89
MCFSOTA8(10%)88.7180.416.321.90
BCPSOTA8(10%)89.6281.316.811.76
PMT(Ours)8(10%)90.8183.235.611.50
+ +Our model outperformed previous models on all four metrics across $5\%$ and $10\%$ labeled data. At $5\%$ of the data, compared to the best results from previous work, our PMT showed a $1.45\%$ improvement in Dice, a $2.32\%$ improvement in Jaccard, a reduction of 1.45 in 95HD, and a reduction of 0.29 in ASD. Our PMT also showed a $1.19\%$ improvement in Dice, a $1.92\%$ improvement in Jaccard, a reduction of 1.20 in 95HD, and a reduction of 0.26 in ASD at $10\%$ of the data. It is noteworthy that our model maintains good performance even with a small amount of data. In comparison with state-of-the-art results, our model achieves a leading performance in the majority of metrics using only half of their data. For instance, we surpass MCF's performance at $10\%$ labeled data using only $5\%$ labeled data. The above indicates that our model achieves superior results compared to existing methods on differently proportioned annotated data in the task of segmenting the left atrium. + +Comparison on Pancreas-NIH Dataset. We conducted a cross-model comparison on the classic Pancreas dataset. Detailed results of the experiments are presented in Table 4. We tested with $10\%$ and $20\%$ of labeled data. To + +![](images/6f78183ef03b2653e9009c0a2726c7e3b72f0700e96b436b845a88d012677779.jpg) +Fig. 4: 2D & 3D segmentation visualization of different semi-supervised methods under $10\%$ labeled on LA (upper) and pancreas (bottom) dataset. + +provide a more intuitive demonstration of the performance of various models on the Pancreas-NIH dataset, we have selected some representative results for visualization, as illustrated in Fig. 4. Areas with inaccurate segmentation have been annotated accordingly. It is worth noting that the results in the table clearly indicate that Pancreas-NIH dataset is significantly more challenging than LA dataset. Therefore, we increased $\hat{\lambda}_{1}$ by a factor of two, resulting in improved performance. + +Our model outperformed previous models on all four metrics across $10\%$ and $20\%$ labeled data. At $10\%$ of the data, compared to the best results from previous work, our PMT showed a $7.17\%$ improvement in Dice, a $9.09\%$ improvement in Jaccard, a reduction of 6.35 in 95HD, and a reduction of 2.10 in ASD. Our PMT also showed a $0.31\%$ improvement in Dice, a $0.55\%$ improvement in Jaccard, and a reduction of 0.36 in ASD at $20\%$ of the data. The above indicates that our model achieves superior results compared to existing methods on differently proportioned annotated data in the task of segmenting the pancreas. + +# 5 Conclusion + +In this paper, we propose a semi-supervised medical image segmentation framework named Progress Mean Teacher (PMT). PMT adopts a progressive + +Table 4: Comparison results on Pancreas-NIH dataset with ${10}\%$ and ${20}\%$ labeled data + +
MethodLabeledMetrics
Dice↑Jaccard↑95HD↓ASD↓
VNet6(10%)55.6041.7445.3318.63
UA-MT6(10%)66.3453.2117.214.57
SASSNet6(10%)68.7853.8619.026.26
DTC6(10%)69.2154.0617.215.95
BCPSOTA6(10%)73.8359.2412.713.72
MCFSOTA6(10%)----
PMT(Ours)6(10%)81.0068.336.361.62
VNet12(20%)72.3858.2619.355.89
UA-MT12(20%)76.1062.6210.842.43
SASSNet12(20%)77.6664.0810.933.05
DTC12(20%)78.2764.758.362.25
BCPSOTA12(20%)82.9170.976.432.25
MCFSOTA12(20%)75.0061.2711.593.27
PMT(Ours)12(20%)83.2271.527.601.89
+ +design training process, establishing temporal-level model alignment and pseudolabel filtering while leveraging a network architecture based on MT. The core idea of this model framework is to generate diverse pseudo-labels consistently during the backpropagation process by establishing representation capability differences caused by iteration gaps, thereby stabilizing and enhancing the model's representation capability. In addition to the progressive architecture, PMT employs two simple yet effective methods, Pseudo Label Filtering (PLF) and Discrepancy Driven Alignment (DDA). PLF utilizes the representation capability of the model to filter pseudo-labels, discarding low-fidelity ones detrimental to constraining, while DDA aligns the differences in model predictions, allowing lagging models to catch up with leading models rapidly. Results from ablation experiments demonstrate that each component of the PMT framework significantly enhances the model's performance. In comparative experiments with other methods, PMT achieves state-of-the-art performance in terms of accuracy, surpassing previous methods significantly, and maintains this advantage compared to other methods in situations with more limited data and more challenging tasks. + +Limitation and Future Work. Despite the outstanding performance of PMT, there is still room for further exploration of semi-supervised training architectures established through temporal consistency. Investigating more advanced and stable strategies under a progressive design paradigm, and assessing whether they can further enhance performance, are topics worthy of further research in the future. + +# Acknowledgments + +This work was supported in part by National Natural Science Foundation of China under Grants 62088102, 12326608 and 62106192, Natural Science Foundation of Shaanxi Province under Grant 2022JC-41, and Fundamental Research Funds for the Central Universities under Grant XTR042021005. + +# References + +1. Achanta, R., Shaji, A., Smith, K., Lucchi, A., Fua, P., Süsstrunk, S.: Slic superpixels compared to state-of-the-art superpixel methods. IEEE transactions on pattern analysis and machine intelligence 34(11), 2274-2282 (2012) 4 +2. Bai, Y., Chen, D., Li, Q., Shen, W., Wang, Y.: Bidirectional copy-paste for semi-supervised medical image segmentation. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 11514-11524 (2023) 10, 11 +3. Cai, H., Li, S., Qi, L., Yu, Q., Shi, Y., Gao, Y.: Orthogonal annotation benefits barely-supervised medical image segmentation. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 3302-3311 (2023) 4 +4. Chen, X., Yuan, Y., Zeng, G., Wang, J.: Semi-supervised semantic segmentation with cross pseudo supervision. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 2613-2622 (2021) 5 +5. Chen, X., He, K.: Exploring simple siamese representation learning. In: Proceedings of the IEEE/CVF conference on computer vision and pattern recognition. pp. 15750-15758 (2021) 2 +6. Dong-DongChen, W., WeiGao, Z.: Tri-net for semi-supervised deep learning. In: Proceedings of twenty-seventh international joint conference on artificial intelligence. pp. 2014-2020 (2018) 4 +7. Jiao, R., Zhang, Y., Ding, L., Cai, R., Zhang, J.: Learning with limited annotations: a survey on deep semi-supervised learning for medical image segmentation. arXiv preprint arXiv:2207.14191 (2022) 4 +8. Ke, Z., Qiu, D., Li, K., Yan, Q., Lau, R.W.: Guided collaborative training for pixelwise semi-supervised learning. In: Computer Vision-ECCV 2020: 16th European Conference, Glasgow, UK, August 23-28, 2020, Proceedings, Part XIII 16. pp. 429-445. Springer (2020) 5 +9. Laine, S., Aila, T.: Temporal ensembling for semi-supervised learning. arXiv preprint arXiv:1610.02242 (2016) 2, 4 +10. Lee, D.H., et al.: Pseudo-label: The simple and efficient semi-supervised learning method for deep neural networks. In: Workshop on challenges in representation learning, ICML. vol. 3, p. 896. Atlanta (2013) 2, 4 +11. Li, S., Zhang, C., He, X.: Shape-aware semi-supervised 3d semantic segmentation for medical images. In: Medical Image Computing and Computer Assisted Intervention-MICCAI 2020: 23rd International Conference, Lima, Peru, October 4-8, 2020, Proceedings, Part I 23. pp. 552-561. Springer (2020) 4, 9, 10, 11 +12. Luo, X., Chen, J., Song, T., Wang, G.: Semi-supervised medical image segmentation through dual-task consistency. In: Proceedings of the AAAI conference on artificial intelligence. vol. 35, pp. 8801-8809 (2021) 4, 9, 10, 11 +13. Miletari, F., Navab, N., Ahmadi, S.A.: V-net: Fully convolutional neural networks for volumetric medical image segmentation. In: 2016 fourth international conference on 3D vision (3DV). pp. 565-571. IEEE (2016) 9 + +14. Na, J., Ha, J.W., Chang, H.J., Han, D., Hwang, W.: Switching temporary teachers for semi-supervised semantic segmentation. Advances in Neural Information Processing Systems 36 (2024) 5 +15. Peiris, H., Hayat, M., Chen, Z., Egan, G., Harandi, M.: Uncertainty-guided dual-views for semi-supervised volumetric medical image segmentation. Nature Machine Intelligence 5(7), 724-738 (2023) 4 +16. Rizve, M.N., Kardan, N., Shah, M.: Towards realistic semi-supervised learning. In: European Conference on Computer Vision. pp. 437-455. Springer (2022) 1 +17. Roth, H.R., Lu, L., Farag, A., Shin, H.C., Liu, J., Turkbey, E.B., Summers, R.M.: Deeporgan: Multi-level deep convolutional networks for automated pancreas segmentation. In: Medical Image Computing and Computer-Assisted Intervention-MICCAI 2015: 18th International Conference, Munich, Germany, October 5-9, 2015, Proceedings, Part I 18. pp. 556-564. Springer (2015) 3, 10 +18. Shen, Z., Cao, P., Yang, H., Liu, X., Yang, J., Zaiane, O.R.: Co-training with high-confidence pseudo labels for semi-supervised medical image segmentation. arXiv preprint arXiv:2301.04465 (2023) 5 +19. Tarvainen, A., Valpola, H.: Mean teachers are better role models: Weight-averaged consistency targets improve semi-supervised deep learning results. Advances in neural information processing systems 30 (2017) 2, 4, 5 +20. Wang, H., Li, X.: Dhc: Dual-debiased heterogeneous co-training framework for class-imbalanced semi-supervised medical image segmentation. In: International Conference on Medical Image Computing and Computer-Assisted Intervention. pp. 582-591. Springer (2023) 2 +21. Wang, J., Zhou, S., Fang, C., Wang, L., Wang, J.: Meta corrupted pixels mining for medical image segmentation. In: Medical Image Computing and Computer Assisted Intervention-MICCAI 2020: 23rd International Conference, Lima, Peru, October 4-8, 2020, Proceedings, Part I 23. pp. 335-345. Springer (2020) 2 +22. Wang, X., Yuan, Y., Guo, D., Huang, X., Cui, Y., Xia, M., Wang, Z., Bai, C., Chen, S.: Ssa-net: Spatial self-attention network for COVID-19 pneumonia infection segmentation with semi-supervised few-shot learning. Medical Image Analysis 79, 102459 (2022) 4 +23. Wang, Y., Xiao, B., Bi, X., Li, W., Gao, X.: Mcf: Mutual correction framework for semi-supervised medical image segmentation. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 15651-15660 (2023) 2, 3, 4, 9, 10, 11 +24. Wu, H., Wang, Z., Song, Y., Yang, L., Qin, J.: Cross-patch dense contrastive learning for semi-supervised segmentation of cellular nuclei in histopathologic images. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 11666-11675 (2022) 2 +25. Xia, K., Wang, L., Zhou, S., Hua, G., Tang, W.: Learning from noisy pseudo labels for semi-supervised temporal action localization. In: Proceedings of the IEEE/CVF International Conference on Computer Vision. pp. 10160-10169 (2023) 1 +26. Xie, Q., Luong, M.T., Hovy, E., Le, Q.V.: Self-training with noisy student improves imagenet classification. In: Proceedings of the IEEE/CVF conference on computer vision and pattern recognition. pp. 10687-10698 (2020) 2 +27. Xin, X., Wang, J., Xie, R., Zhou, S., Huang, W., Zheng, N.: Semi-supervised person re-identification using multi-view clustering. Pattern Recognition 88, 285–297 (2019) 1 +28. Xiong, Z., Xia, Q., Hu, Z., Huang, N., Bian, C., Zheng, Y., Vesal, S., Ravikumar, N., Maier, A., Yang, X., et al.: A global benchmark of algorithms for segmenting + +the left atrium from late gadolinium-enhanced cardiac magnetic resonance imaging. Medical image analysis 67, 101832 (2021) 3, 10 +29. Xu, Z., Wang, Y., Lu, D., Yu, L., Yan, J., Luo, J., Ma, K., Zheng, Y., Tong, R.K.y.: All-around real label supervision: Cyclic prototype consistency learning for semi-supervised medical image segmentation. IEEE Journal of Biomedical and Health Informatics 26(7), 3174-3184 (2022) 4, 10 +30. Yu, L., Wang, S., Li, X., Fu, C.W., Heng, P.A.: Uncertainty-aware self-ensembling model for semi-supervised 3d left atrium segmentation. In: Medical Image Computing and Computer Assisted Intervention-MICCAI 2019: 22nd International Conference, Shenzhen, China, October 13–17, 2019, Proceedings, Part II 22. pp. 605–613. Springer (2019) 4, 9, 10, 11 +31. Yun, S., Han, D., Oh, S.J., Chun, S., Choe, J., Yoo, Y.: Cutmix: Regularization strategy to train strong classifiers with localizable features. In: Proceedings of the IEEE/CVF international conference on computer vision. pp. 6023-6032 (2019) 11 +32. Zheng, M., You, S., Huang, L., Wang, F., Qian, C., Xu, C.: Simmatch: Semi-supervised learning with similarity matching. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 14471-14481 (2022) 1 +33. Zhou, S., Wang, J., Shu, J., Meng, D., Wang, L., Zheng, N.: Multinetwork collaborative feature learning for semisupervised person reidentification. IEEE Transactions on Neural Networks and Learning Systems 33(9), 4826-4839 (2021) 1 \ No newline at end of file diff --git a/2024/PMT_ Progressive Mean Teacher via Exploring Temporal Consistency for Semi-Supervised Medical Image Segmentation/images.zip b/2024/PMT_ Progressive Mean Teacher via Exploring Temporal Consistency for Semi-Supervised Medical Image Segmentation/images.zip new file mode 100644 index 0000000000000000000000000000000000000000..2879dece959a49f9040094af07d7d219498240a9 --- /dev/null +++ b/2024/PMT_ Progressive Mean Teacher via Exploring Temporal Consistency for Semi-Supervised Medical Image Segmentation/images.zip @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:ead29650306ae0d2a95b841f44fa8666fb81bdf1250f02ef91ecd23f7463f7a3 +size 532814 diff --git a/2024/PMT_ Progressive Mean Teacher via Exploring Temporal Consistency for Semi-Supervised Medical Image Segmentation/layout.json b/2024/PMT_ Progressive Mean Teacher via Exploring Temporal Consistency for Semi-Supervised Medical Image Segmentation/layout.json new file mode 100644 index 0000000000000000000000000000000000000000..2a3cfe273476ded8d301dd4e7ffb9e550a476489 --- /dev/null +++ b/2024/PMT_ Progressive Mean Teacher via Exploring Temporal Consistency for Semi-Supervised Medical Image Segmentation/layout.json @@ -0,0 +1,8291 @@ +{ + "pdf_info": [ + { + "para_blocks": [ + { + "bbox": [ + 143, + 111, + 473, + 166 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 143, + 111, + 473, + 166 + ], + "spans": [ + { + "bbox": [ + 143, + 111, + 473, + 166 + ], + "type": "text", + "content": "PMT: Progressive Mean Teacher via Exploring Temporal Consistency for Semi-Supervised Medical Image Segmentation" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 165, + 186, + 447, + 200 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 165, + 186, + 447, + 200 + ], + "spans": [ + { + "bbox": [ + 165, + 186, + 447, + 200 + ], + "type": "text", + "content": "Ning Gao" + }, + { + "bbox": [ + 165, + 186, + 447, + 200 + ], + "type": "inline_equation", + "content": "^{\\text{D}}" + }, + { + "bbox": [ + 165, + 186, + 447, + 200 + ], + "type": "text", + "content": ", Sanping Zhou" + }, + { + "bbox": [ + 165, + 186, + 447, + 200 + ], + "type": "inline_equation", + "content": "^{*}" + }, + { + "bbox": [ + 165, + 186, + 447, + 200 + ], + "type": "text", + "content": ", Le Wang" + }, + { + "bbox": [ + 165, + 186, + 447, + 200 + ], + "type": "inline_equation", + "content": "^{\\text{D}}" + }, + { + "bbox": [ + 165, + 186, + 447, + 200 + ], + "type": "text", + "content": ", and Nanning Zheng" + }, + { + "bbox": [ + 165, + 186, + 447, + 200 + ], + "type": "inline_equation", + "content": "^{\\text{D}}" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 144, + 208, + 468, + 242 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 144, + 208, + 468, + 242 + ], + "spans": [ + { + "bbox": [ + 144, + 208, + 468, + 242 + ], + "type": "text", + "content": "National Key Laboratory of Human-Machine Hybrid Augmented Intelligence, National Engineering Research Center for Visual Information and Applications, Institute of Artificial Intelligence and Robotics, Xi'an Jiaotong University" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 159, + 270, + 455, + 534 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 159, + 270, + 455, + 534 + ], + "spans": [ + { + "bbox": [ + 159, + 270, + 455, + 534 + ], + "type": "text", + "content": "Abstract. Semi-supervised learning has emerged as a widely adopted technique in the field of medical image segmentation. The existing works either focuses on the construction of consistency constraints or the generation of pseudo labels to provide high-quality supervisory signals, whose main challenge mainly comes from how to keep the continuous improvement of model capabilities. In this paper, we propose a simple yet effective semi-supervised learning framework, termed Progressive Mean Teachers (PMT), for medical image segmentation, whose goal is to generate high-fidelity pseudo labels by learning robust and diverse features in the training process. Specifically, our PMT employs a standard mean teacher to penalize the consistency of the current state and utilizes two sets of MT architectures for co-training. The two sets of MT architectures are individually updated for prolonged periods to maintain stable model diversity established through performance gaps generated by iteration differences. Additionally, a difference-driven alignment regularizer is employed to expedite the alignment of lagging models with the representation capabilities of leading models. Furthermore, a simple yet effective pseudo-label filtering algorithm is employed for facile evaluation of models and selection of high-fidelity pseudo-labels outputted when models are operating at high performance for co-training purposes. Experimental results on two datasets with different modalities, i.e., CT and MRI, demonstrate that our method outperforms the state-of-the-art medical image segmentation approaches across various dimensions. The code is available at https://github.com/Axi404/PMT." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 159, + 544, + 453, + 567 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 159, + 544, + 453, + 567 + ], + "spans": [ + { + "bbox": [ + 159, + 544, + 453, + 567 + ], + "type": "text", + "content": "Keywords: Semi-supervised learning " + }, + { + "bbox": [ + 159, + 544, + 453, + 567 + ], + "type": "inline_equation", + "content": "\\cdot" + }, + { + "bbox": [ + 159, + 544, + 453, + 567 + ], + "type": "text", + "content": " Medical image segmentation " + }, + { + "bbox": [ + 159, + 544, + 453, + 567 + ], + "type": "inline_equation", + "content": "\\cdot" + }, + { + "bbox": [ + 159, + 544, + 453, + 567 + ], + "type": "text", + "content": " Temporal consistency regularization" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 132, + 586, + 230, + 600 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 586, + 230, + 600 + ], + "spans": [ + { + "bbox": [ + 132, + 586, + 230, + 600 + ], + "type": "text", + "content": "1 Introduction" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 130, + 611, + 482, + 647 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 611, + 482, + 647 + ], + "spans": [ + { + "bbox": [ + 130, + 611, + 482, + 647 + ], + "type": "text", + "content": "Semi-supervised learning is an important field in deep learning, which offers an effective way to tackle problems with limited labeled data [16, 25, 27, 32, 33]. With the continuous emergence of large-scale data, semi-supervised learning has" + } + ] + } + ], + "index": 6 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 133, + 653, + 236, + 666 + ], + "type": "page_footnote", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 653, + 236, + 666 + ], + "spans": [ + { + "bbox": [ + 133, + 653, + 236, + 666 + ], + "type": "text", + "content": "* Corresponding author." + } + ] + } + ], + "index": 7 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 0 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 196, + 116, + 417, + 259 + ], + "blocks": [ + { + "bbox": [ + 196, + 116, + 417, + 259 + ], + "lines": [ + { + "bbox": [ + 196, + 116, + 417, + 259 + ], + "spans": [ + { + "bbox": [ + 196, + 116, + 417, + 259 + ], + "type": "image", + "image_path": "062538e6c64842cb76ba27d2a32d430a0131c159211272c07e580bf7119e48bf.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 130, + 271, + 482, + 338 + ], + "lines": [ + { + "bbox": [ + 130, + 271, + 482, + 338 + ], + "spans": [ + { + "bbox": [ + 130, + 271, + 482, + 338 + ], + "type": "text", + "content": "Fig. 1: Motivation of our PMT. In particular, the standard MT helps to learn robust features by keeping the consistency between teacher and student networks at the current iteration, while our PMT further helps to learn diverse features by maintaining the difference between student networks at different iterations. As a result, more and more high-fidelity pseudo labels will be generated for semi-supervised medical image segmentation." + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_caption" + } + ], + "index": 2 + }, + { + "bbox": [ + 130, + 364, + 482, + 424 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 364, + 482, + 424 + ], + "spans": [ + { + "bbox": [ + 130, + 364, + 482, + 424 + ], + "type": "text", + "content": "become a hot topic in both machine learning and computer vision communities. It has gained particular attention in the medical image segmentation [21], due to the need for expert annotation in determining the accurate boundaries of targets. As a result, the labeled data is very lacking in medical image segmentation, which makes the semi-supervised learning very popular in this domain." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 130, + 425, + 482, + 592 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 425, + 482, + 592 + ], + "spans": [ + { + "bbox": [ + 130, + 425, + 482, + 592 + ], + "type": "text", + "content": "Most of the semi-supervised learning methods involve seeking a guidance mechanism using limited annotated data, so as to maximize the usage of unlabeled data to train the models on different downstream tasks. In general, two technologies, i.e., consistency regularization [9, 10] and pseudo label generation [10, 19], are commonly used in semi-supervised medical segmentation. The former ones aim at enhancing the representation capability of networks by exploring the consistency based on different prior assumptions [5], while the latter ones focus on improving the performance on downstream tasks by generating the pseudo labels based on current model capability. For example, the well-known Mean Teacher (MT) [19] utilizes an ingenious Exponential Moving Average (EMA) that is equivalent to data augmentation and enforces consistency regularization between the outputs of teacher and student models. Besides, the representative Noisy Student [26] designs a teacher model to generate pseudo labels, so as to train a larger student model." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 130, + 594, + 482, + 666 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 594, + 482, + 666 + ], + "spans": [ + { + "bbox": [ + 130, + 594, + 482, + 666 + ], + "type": "text", + "content": "Even though the above methods have achieved significant improvements in semi-supervised medical image segmentation, we still argue that they are weak in exploring high-quality supervisory signals to consistently enhance the model's capability. To address this issue, an iterative unified optimization framework has been introduced to semi-supervised medical image segmentation [20, 23, 24], in which the consistency regularization strategy is taken to enhance the model's" + } + ] + } + ], + "index": 6 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "spans": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "type": "text", + "content": "2" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 166, + 91, + 220, + 100 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 166, + 91, + 220, + 100 + ], + "spans": [ + { + "bbox": [ + 166, + 91, + 220, + 100 + ], + "type": "text", + "content": "N. Gao et al." + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 1 + }, + { + "para_blocks": [ + { + "bbox": [ + 130, + 116, + 479, + 187 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 116, + 479, + 187 + ], + "spans": [ + { + "bbox": [ + 130, + 116, + 479, + 187 + ], + "type": "text", + "content": "representation capability and the pseudo label generation algorithm is applied to generate high-fidelity pseudo labels. For example, the recent MCF [23] takes VNet and 3D ResNet for representation learning, and heterogeneous networks for dynamic pseudo label generation. The challenges to these methods lie on how to continuously generate diverse pseudo labels in the forward propagation, and enhance model's capability in the backward propagation." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 130, + 189, + 482, + 463 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 189, + 482, + 463 + ], + "spans": [ + { + "bbox": [ + 130, + 189, + 482, + 463 + ], + "type": "text", + "content": "In this paper, we design a novel semi-supervised learning framework, termed Progressive Mean Teachers (PMT), for medical image segmentation, whose main idea focuses on how to obtain a diverse set of accurate pseudo labels. Inspired by the positive relationship between iteration and performance, we try to maintain the diversity of networks by exploring their states at different training epochs. Specifically, the standard MT is first taken as basic architecture to learn the parameters of network at each iteration, which can enhance the network's representation capability by using the EMA data augmentation. Our PMT further explores network diversity during training, alternating between two homogeneous MT architectures trained on the same dataset, a process we term progressive design. These models exhibit significant iteration leads due to alternating continuous individual updates, establishing performance gaps between networks at different epochs, as shown in Fig. 1. Consequently, the student network can acquire robust yet diverse features for medical image segmentation. Then, the Discrepancy Driven Alignment (DDA) regularizer is further designed to examine disparities between predictions obtained by the student network at different training epochs, facilitating rapid alignment to high-fidelity generated images. Finally, we design a simple Pseudo Label Filtering (PLF) algorithm to refine the basic interaction process, enabling the retention of high-fidelity pseudo-labels for training by comparing student network performance across different training epochs. As a result, more and more high-fidelity pseudo labels can be fed to train the other student networks, which will in turn generate more accurate predictions for pseudo label generation." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 130, + 464, + 482, + 559 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 464, + 482, + 559 + ], + "spans": [ + { + "bbox": [ + 130, + 464, + 482, + 559 + ], + "type": "text", + "content": "In summary, the main contributions of this work are as follows: (1) We design a novel Progressive Mean Teacher framework for semi-supervised medical image segmentation. (2) We design a novel Discrepancy Driven Alignment regularizer to rapidly align the representational capacity gap between lagging and leading networks. (3) We design a simple yet effective Pseudo Label Filtering algorithm to select high-fidelity pseudo labels. Extensive experiments in both Left Atrial [28] and Pancreas-NIH [17] datasets show that our PMT can achieve the state-of-the-art results in semi-supervised medical image segmentation." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 131, + 580, + 237, + 593 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 131, + 580, + 237, + 593 + ], + "spans": [ + { + "bbox": [ + 131, + 580, + 237, + 593 + ], + "type": "text", + "content": "2 Related Work" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 131, + 608, + 298, + 619 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 131, + 608, + 298, + 619 + ], + "spans": [ + { + "bbox": [ + 131, + 608, + 298, + 619 + ], + "type": "text", + "content": "2.1 Consistency Regularization" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 130, + 629, + 481, + 665 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 629, + 481, + 665 + ], + "spans": [ + { + "bbox": [ + 130, + 629, + 481, + 665 + ], + "type": "text", + "content": "Consistency regularization is often employed in semi-supervised learning, so as to enhance the stable representational capacity of model. The behind idea is to preserve the invariance of predictions made by the same model when facing" + } + ] + } + ], + "index": 7 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 340, + 91, + 447, + 102 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 340, + 91, + 447, + 102 + ], + "spans": [ + { + "bbox": [ + 340, + 91, + 447, + 102 + ], + "type": "text", + "content": "Progressive Mean Teacher" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 474, + 92, + 480, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 474, + 92, + 480, + 100 + ], + "spans": [ + { + "bbox": [ + 474, + 92, + 480, + 100 + ], + "type": "text", + "content": "3" + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 2 + }, + { + "para_blocks": [ + { + "bbox": [ + 130, + 116, + 482, + 319 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 116, + 482, + 319 + ], + "spans": [ + { + "bbox": [ + 130, + 116, + 482, + 319 + ], + "type": "text", + "content": "perturbations applied in different regions, by constraining the consistency between output results under different perturbations. For example, " + }, + { + "bbox": [ + 130, + 116, + 482, + 319 + ], + "type": "inline_equation", + "content": "\\varPi" + }, + { + "bbox": [ + 130, + 116, + 482, + 319 + ], + "type": "text", + "content": " model [9] introduces image-level regularization by adding perturbations to images. The well-known MT [19] introduces parameter-level regularization by using EMA to constrain the outputs between teacher and student models. Thanks to its simplicity and effectiveness, more and more works are paying attention to improve the generalization ability by formulating different consistency regularizers. For example, SASSnet [11] focuses on the regularity of geometric shapes for target object classes within consistency regularization. Besides, CPCL [29] establishes regularization between supervised and unsupervised training within a cyclic framework. Furthermore, some later works have made progress in regularization at different task and model levels. For example, DTC [12] introduces task-level regularization, presenting a novel dual-task consistency semi-supervised framework. Besides, MCF [23] introduces model-level regularization by using heterogeneous models to constrain output consistency. Compared to previous work, our PMT seeks the cross-temporal regularization between different training periods, which can help learn diverse yet robust features for subsequent pseudo label generation." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 132, + 336, + 289, + 346 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 336, + 289, + 346 + ], + "spans": [ + { + "bbox": [ + 132, + 336, + 289, + 346 + ], + "type": "text", + "content": "2.2 Pseudo Label Generation" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 130, + 354, + 482, + 594 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 354, + 482, + 594 + ], + "spans": [ + { + "bbox": [ + 130, + 354, + 482, + 594 + ], + "type": "text", + "content": "Pseudo label generation is often employed in semi-supervised learning, so as to enhance the discriminative ability of model. The behind idea is to train a prior model with the labeled data and then apply it to generate the pseudo labels for unlabeled data. It is widely accepted to categorize pseudo-label generation into two methods [7], direct generation focuses on selecting pseudo labels with higher confidence, while indirect generation explores methods to generate high-fidelity pseudo labels. For direct generation, work [10] uses a fixed threshold to choose high-confidence pseudo labels. SsaNet [22] employs a trust module to reevaluate pseudo labels. UA-MT [30] introduces uncertainty estimation to filter out unreliable pseudo labels. Co-BioNet [15] introduces a feedback network to measure the uncertainty and choose high-confidence predictions of different models. For indirect generation, Tri-Net [6] is proposed to use two subnetworks to generate pseudo labels for a third subnetwork. Work [1] improves pseudo label generation using Simple Linear Iterative Clustering (SLIC) algorithm. MCF [23] dynamically generates pseudo labels using a heterogeneous network and effectively addresses cognitive bias, while DeSCO [3] focuses on the spatial correlation of medical images and generates pseudo labels using orthogonal slices. Compared to previous work, our PMT emphasizes the enhancement of pseudo label quality in the temporal domain and generates diverse and robust pseudo labels across different training periods, significantly improving performance." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 132, + 610, + 285, + 622 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 610, + 285, + 622 + ], + "spans": [ + { + "bbox": [ + 132, + 610, + 285, + 622 + ], + "type": "text", + "content": "2.3 Multi-Model Framework" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 130, + 629, + 481, + 665 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 629, + 481, + 665 + ], + "spans": [ + { + "bbox": [ + 130, + 629, + 481, + 665 + ], + "type": "text", + "content": "The multi-model framework is often employed in semi-supervised learning, so as to enhance the model representation by acquiring multiple views or diversity. Its development aligns closely with consistency regularization. Here, we" + } + ] + } + ], + "index": 6 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "spans": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "type": "text", + "content": "4" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 166, + 91, + 220, + 101 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 166, + 91, + 220, + 101 + ], + "spans": [ + { + "bbox": [ + 166, + 91, + 220, + 101 + ], + "type": "text", + "content": "N. Gao et al." + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 3 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 136, + 116, + 479, + 277 + ], + "blocks": [ + { + "bbox": [ + 136, + 116, + 479, + 277 + ], + "lines": [ + { + "bbox": [ + 136, + 116, + 479, + 277 + ], + "spans": [ + { + "bbox": [ + 136, + 116, + 479, + 277 + ], + "type": "image", + "image_path": "fdeffe67ee6516f4a7f3439e0d10a005ee9fc08aa83cf44a01c443edf12a99a2.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 130, + 285, + 482, + 330 + ], + "lines": [ + { + "bbox": [ + 130, + 285, + 482, + 330 + ], + "spans": [ + { + "bbox": [ + 130, + 285, + 482, + 330 + ], + "type": "text", + "content": "Fig. 2: An overview of PMT. We employed a progressive design and utilized the architecture of MT. The PMT framework maintains a data buffer of length " + }, + { + "bbox": [ + 130, + 285, + 482, + 330 + ], + "type": "inline_equation", + "content": "\\mathcal{B}" + }, + { + "bbox": [ + 130, + 285, + 482, + 330 + ], + "type": "text", + "content": " for cros-temporal training. The total loss function " + }, + { + "bbox": [ + 130, + 285, + 482, + 330 + ], + "type": "inline_equation", + "content": "\\mathrm{L}_{\\mathrm{total}}" + }, + { + "bbox": [ + 130, + 285, + 482, + 330 + ], + "type": "text", + "content": " for each network includes supervised losses " + }, + { + "bbox": [ + 130, + 285, + 482, + 330 + ], + "type": "inline_equation", + "content": "\\mathrm{L_{CE}}, \\mathrm{L_{DICE}}, \\mathrm{L_{aln}}" + }, + { + "bbox": [ + 130, + 285, + 482, + 330 + ], + "type": "text", + "content": ", and unsupervised loss " + }, + { + "bbox": [ + 130, + 285, + 482, + 330 + ], + "type": "inline_equation", + "content": "\\mathrm{L_U}, \\mathrm{L_T}" + }, + { + "bbox": [ + 130, + 285, + 482, + 330 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_caption" + } + ], + "index": 2 + }, + { + "bbox": [ + 130, + 353, + 482, + 521 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 353, + 482, + 521 + ], + "spans": [ + { + "bbox": [ + 130, + 353, + 482, + 521 + ], + "type": "text", + "content": "focus on the evolution of model structures. For example, MT [19] introduces a Teacher model with significantly improved representation capabilities at a lower cost using EMA, establishing the MT architecture. Besides, CPC [8] utilizes the confidence vector of multi-model outputs as pseudo-labels for co-training, while CPS [4] employs one-hot labels, both considered as starting points for co-training. In recent years, efforts have been made to combine these approaches. For example, UCMT [18] employs two student models for co-training, simultaneously updating the same Mean Teacher using EMA, while Dual Teacher [14] alternates updates between two Teacher models using EMA with a single student model, thereby fostering diversity among Teacher models. Compared to previous work, our PMT utilizes two sets of Mean Teachers in a progressive training framework, and employ two student models to update teacher models independently. This enables our model to rapidly establish robust performance disparities across iteration gap and maintain stable diversity among models." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 132, + 538, + 201, + 550 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 538, + 201, + 550 + ], + "spans": [ + { + "bbox": [ + 132, + 538, + 201, + 550 + ], + "type": "text", + "content": "3 Method" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 131, + 562, + 307, + 574 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 131, + 562, + 307, + 574 + ], + "spans": [ + { + "bbox": [ + 131, + 562, + 307, + 574 + ], + "type": "text", + "content": "3.1 The Overall Process of PMT" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 130, + 581, + 482, + 667 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 581, + 482, + 667 + ], + "spans": [ + { + "bbox": [ + 130, + 581, + 482, + 667 + ], + "type": "text", + "content": "The training process of PMT is illustrated in Fig. 2. As previously mentioned, PMT can consist of multiple networks. For ease of explanation, the number of networks is assumed to be two by default in our work. In our approach, these networks share the same structure and can be denoted as " + }, + { + "bbox": [ + 130, + 581, + 482, + 667 + ], + "type": "inline_equation", + "content": "f_{i}(\\cdot) \\in \\mathcal{F}" + }, + { + "bbox": [ + 130, + 581, + 482, + 667 + ], + "type": "text", + "content": ", where " + }, + { + "bbox": [ + 130, + 581, + 482, + 667 + ], + "type": "inline_equation", + "content": "\\mathcal{F}" + }, + { + "bbox": [ + 130, + 581, + 482, + 667 + ], + "type": "text", + "content": " stands for the function space within which these networks reside. In a semi-supervised process, the training data includes a small number of labeled data denoted as " + }, + { + "bbox": [ + 130, + 581, + 482, + 667 + ], + "type": "inline_equation", + "content": "\\mathbf{D}_L = \\{(x_i^L, y_i^L)\\}_{i=1}^N" + }, + { + "bbox": [ + 130, + 581, + 482, + 667 + ], + "type": "text", + "content": ", and a large amount of unlabeled data denoted" + } + ] + } + ], + "index": 7 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 340, + 91, + 447, + 102 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 340, + 91, + 447, + 102 + ], + "spans": [ + { + "bbox": [ + 340, + 91, + 447, + 102 + ], + "type": "text", + "content": "Progressive Mean Teacher" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 474, + 92, + 480, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 474, + 92, + 480, + 100 + ], + "spans": [ + { + "bbox": [ + 474, + 92, + 480, + 100 + ], + "type": "text", + "content": "5" + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 4 + }, + { + "para_blocks": [ + { + "bbox": [ + 130, + 114, + 482, + 166 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 114, + 482, + 166 + ], + "spans": [ + { + "bbox": [ + 130, + 114, + 482, + 166 + ], + "type": "text", + "content": "as " + }, + { + "bbox": [ + 130, + 114, + 482, + 166 + ], + "type": "inline_equation", + "content": "\\mathbf{D}_U = \\{(x_i^U)\\}_{i = N + 1}^{N + M}" + }, + { + "bbox": [ + 130, + 114, + 482, + 166 + ], + "type": "text", + "content": ", where " + }, + { + "bbox": [ + 130, + 114, + 482, + 166 + ], + "type": "inline_equation", + "content": "N\\ll M" + }, + { + "bbox": [ + 130, + 114, + 482, + 166 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 130, + 114, + 482, + 166 + ], + "type": "inline_equation", + "content": "x_{i}\\in \\mathbb{R}^{H\\times W\\times D}" + }, + { + "bbox": [ + 130, + 114, + 482, + 166 + ], + "type": "text", + "content": " represents medical volumes, and " + }, + { + "bbox": [ + 130, + 114, + 482, + 166 + ], + "type": "inline_equation", + "content": "y_{i}\\in \\{0,1\\}^{H\\times W\\times D}" + }, + { + "bbox": [ + 130, + 114, + 482, + 166 + ], + "type": "text", + "content": " represents ground truth labels. Batches of input data " + }, + { + "bbox": [ + 130, + 114, + 482, + 166 + ], + "type": "inline_equation", + "content": "\\mathbf{X}" + }, + { + "bbox": [ + 130, + 114, + 482, + 166 + ], + "type": "text", + "content": " consist of an equal proportion of labeled data " + }, + { + "bbox": [ + 130, + 114, + 482, + 166 + ], + "type": "inline_equation", + "content": "(\\mathbf{X}^L,\\mathbf{Y}^L)" + }, + { + "bbox": [ + 130, + 114, + 482, + 166 + ], + "type": "text", + "content": " and unlabeled data " + }, + { + "bbox": [ + 130, + 114, + 482, + 166 + ], + "type": "inline_equation", + "content": "\\mathbf{X}^U" + }, + { + "bbox": [ + 130, + 114, + 482, + 166 + ], + "type": "text", + "content": ". These volumes are fed into " + }, + { + "bbox": [ + 130, + 114, + 482, + 166 + ], + "type": "inline_equation", + "content": "f_{i}(\\cdot)" + }, + { + "bbox": [ + 130, + 114, + 482, + 166 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 130, + 114, + 482, + 166 + ], + "type": "inline_equation", + "content": "f_{i + 1}(\\cdot)" + }, + { + "bbox": [ + 130, + 114, + 482, + 166 + ], + "type": "text", + "content": ":" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 238, + 175, + 481, + 190 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 238, + 175, + 481, + 190 + ], + "spans": [ + { + "bbox": [ + 238, + 175, + 481, + 190 + ], + "type": "interline_equation", + "content": "\\hat {\\mathbf {Y}} _ {i} = f _ {i} (\\mathbf {X}), \\quad \\hat {\\mathbf {Y}} _ {i + 1} = f _ {i + 1} (\\mathbf {X}). \\tag {1}", + "image_path": "1eac0df4e49900a0e0baf82d1cd4841da949963fb238f8c23510e514ecdf6bff.jpg" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 130, + 193, + 481, + 228 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 193, + 481, + 228 + ], + "spans": [ + { + "bbox": [ + 130, + 193, + 481, + 228 + ], + "type": "text", + "content": "The output consists of volume predictions for both labeled and unlabeled data: " + }, + { + "bbox": [ + 130, + 193, + 481, + 228 + ], + "type": "inline_equation", + "content": "\\hat{\\mathbf{Y}} = \\hat{\\mathbf{Y}}^L\\cup \\hat{\\mathbf{Y}}^U" + }, + { + "bbox": [ + 130, + 193, + 481, + 228 + ], + "type": "text", + "content": ". For the sake of simplicity, the network index subscripts are omitted here." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 130, + 229, + 482, + 338 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 229, + 482, + 338 + ], + "spans": [ + { + "bbox": [ + 130, + 229, + 482, + 338 + ], + "type": "text", + "content": "We have ingeniously reintroduced MT architecture into the framework of semi-supervised learning. The introduction serves the purpose that, while providing high-fidelity pseudo labels to the model, the model's architecture needs to support the stable improvement of its representational capacity, aiming for better performance, while maintaining stable diversity. The teacher network is structurally identical to the student network but does not actively participate in the training process. Instead, it updates all of its parameters through EMA and applies consistency regularization to the student model through its output. The parameter is updated as follows:" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 228, + 350, + 480, + 361 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 228, + 350, + 480, + 361 + ], + "spans": [ + { + "bbox": [ + 228, + 350, + 480, + 361 + ], + "type": "interline_equation", + "content": "\\theta_ {\\text {t e a c h e r}} = \\alpha \\theta_ {\\text {t e a c h e r}} + (1 - \\alpha) \\theta_ {\\text {s t u d e n t}}, \\tag {2}", + "image_path": "47212c3d23089ca2fdb769e8892710375b2c972e24e7edc8e220ec772eaf1a77.jpg" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 130, + 365, + 481, + 389 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 365, + 481, + 389 + ], + "spans": [ + { + "bbox": [ + 130, + 365, + 481, + 389 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 130, + 365, + 481, + 389 + ], + "type": "inline_equation", + "content": "\\theta_{\\mathrm{teacher}}" + }, + { + "bbox": [ + 130, + 365, + 481, + 389 + ], + "type": "text", + "content": " represents the parameters of teacher network, " + }, + { + "bbox": [ + 130, + 365, + 481, + 389 + ], + "type": "inline_equation", + "content": "\\theta_{\\mathrm{student}}" + }, + { + "bbox": [ + 130, + 365, + 481, + 389 + ], + "type": "text", + "content": " represents the parameters of student network, and " + }, + { + "bbox": [ + 130, + 365, + 481, + 389 + ], + "type": "inline_equation", + "content": "\\alpha" + }, + { + "bbox": [ + 130, + 365, + 481, + 389 + ], + "type": "text", + "content": " is the EMA decay rate." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 130, + 389, + 482, + 460 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 389, + 482, + 460 + ], + "spans": [ + { + "bbox": [ + 130, + 389, + 482, + 460 + ], + "type": "text", + "content": "In practice, we propose a cross-temporal training approach, which involves introducing a phase shift among different models during iterations and training the most lagging model in terms of iterations. To ensure that each model can train across temporal, the number of iterations for the model recorded in the sequence of iterations is denoted as " + }, + { + "bbox": [ + 130, + 389, + 482, + 460 + ], + "type": "inline_equation", + "content": "\\mathcal{I}_i" + }, + { + "bbox": [ + 130, + 389, + 482, + 460 + ], + "type": "text", + "content": ". Our model maintains a static iteration gap through the data buffer length " + }, + { + "bbox": [ + 130, + 389, + 482, + 460 + ], + "type": "inline_equation", + "content": "\\mathcal{B}" + }, + { + "bbox": [ + 130, + 389, + 482, + 460 + ], + "type": "text", + "content": ":" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 198, + 470, + 481, + 495 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 198, + 470, + 481, + 495 + ], + "spans": [ + { + "bbox": [ + 198, + 470, + 481, + 495 + ], + "type": "interline_equation", + "content": "\\forall i \\in [ 1, n - 1 ], \\operatorname {L a r} (\\mathcal {I}, i) - \\operatorname {L a r} (\\mathcal {I}, i + 1) = \\frac {\\mathcal {B}}{(n - 1)}, \\tag {3}", + "image_path": "e631699f12c4a2d0a89a6e866a02c695400112678285708648297f5a07af4ed4.jpg" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 130, + 498, + 381, + 510 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 498, + 381, + 510 + ], + "spans": [ + { + "bbox": [ + 130, + 498, + 381, + 510 + ], + "type": "text", + "content": "where, " + }, + { + "bbox": [ + 130, + 498, + 381, + 510 + ], + "type": "inline_equation", + "content": "\\operatorname{Lar}(\\mathcal{I}, i)" + }, + { + "bbox": [ + 130, + 498, + 381, + 510 + ], + "type": "text", + "content": " represents the " + }, + { + "bbox": [ + 130, + 498, + 381, + 510 + ], + "type": "inline_equation", + "content": "i" + }, + { + "bbox": [ + 130, + 498, + 381, + 510 + ], + "type": "text", + "content": "-th largest number in " + }, + { + "bbox": [ + 130, + 498, + 381, + 510 + ], + "type": "inline_equation", + "content": "\\mathcal{I}" + }, + { + "bbox": [ + 130, + 498, + 381, + 510 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 130, + 510, + 482, + 666 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 510, + 482, + 666 + ], + "spans": [ + { + "bbox": [ + 130, + 510, + 482, + 666 + ], + "type": "text", + "content": "During each training iteration, the model with the least advanced iteration progress, referred to as the Current Progressive Model (CPM), sequentially utilizes data from the buffer for training until it has advanced ahead of the model with the most advanced iteration progress " + }, + { + "bbox": [ + 130, + 510, + 482, + 666 + ], + "type": "inline_equation", + "content": "\\mathcal{B} / n" + }, + { + "bbox": [ + 130, + 510, + 482, + 666 + ], + "type": "text", + "content": " times. Throughout the training process, the PLF screens out pseudo-labels generated by models whose performance is inferior to that of the CPM by measuring the performance gap between the CPM and other models, and learns from the remaining pseudo-labels. Simultaneously, the model computes the DDA with respect to a given input " + }, + { + "bbox": [ + 130, + 510, + 482, + 666 + ], + "type": "inline_equation", + "content": "\\mathbf{X}" + }, + { + "bbox": [ + 130, + 510, + 482, + 666 + ], + "type": "text", + "content": " and, along with all other models, examines regions of inconsistent predictions among different networks. When the performance of the CPM lags behind that of other models, the DDA corrects the CPM, enabling it to quickly align its performance with models that outperform it. It is worth noting that if the CPM is the best-performing model for that iteration, it will not guide other models in reverse. It" + } + ] + } + ], + "index": 11 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "spans": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "type": "text", + "content": "6" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 166, + 91, + 220, + 101 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 166, + 91, + 220, + 101 + ], + "spans": [ + { + "bbox": [ + 166, + 91, + 220, + 101 + ], + "type": "text", + "content": "N. Gao et al." + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 5 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 206, + 119, + 304, + 222 + ], + "blocks": [ + { + "bbox": [ + 206, + 119, + 304, + 222 + ], + "lines": [ + { + "bbox": [ + 206, + 119, + 304, + 222 + ], + "spans": [ + { + "bbox": [ + 206, + 119, + 304, + 222 + ], + "type": "image", + "image_path": "71cf79dd32c6620cdbcc9651461e85d067aaecdc4ff856c4fd342be0857c2784.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 249, + 223, + 261, + 233 + ], + "lines": [ + { + "bbox": [ + 249, + 223, + 261, + 233 + ], + "spans": [ + { + "bbox": [ + 249, + 223, + 261, + 233 + ], + "type": "text", + "content": "(a)" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_caption" + } + ], + "index": 2 + }, + { + "type": "image", + "bbox": [ + 307, + 119, + 406, + 222 + ], + "blocks": [ + { + "bbox": [ + 307, + 119, + 406, + 222 + ], + "lines": [ + { + "bbox": [ + 307, + 119, + 406, + 222 + ], + "spans": [ + { + "bbox": [ + 307, + 119, + 406, + 222 + ], + "type": "image", + "image_path": "137070e520425265a242a5cb371ba0abccefb748d358ce452af54722233ebb6e.jpg" + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 351, + 223, + 362, + 233 + ], + "lines": [ + { + "bbox": [ + 351, + 223, + 362, + 233 + ], + "spans": [ + { + "bbox": [ + 351, + 223, + 362, + 233 + ], + "type": "text", + "content": "(b)" + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 131, + 243, + 480, + 266 + ], + "lines": [ + { + "bbox": [ + 131, + 243, + 480, + 266 + ], + "spans": [ + { + "bbox": [ + 131, + 243, + 480, + 266 + ], + "type": "text", + "content": "Fig. 3: Illustration of PLF and DDA. PLF is utilized for comparing the representational capacity of models, while DDA aligns model outputs by examining differences." + } + ] + } + ], + "index": 6, + "angle": 0, + "type": "image_caption" + } + ], + "index": 4 + }, + { + "bbox": [ + 130, + 291, + 479, + 316 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 291, + 479, + 316 + ], + "spans": [ + { + "bbox": [ + 130, + 291, + 479, + 316 + ], + "type": "text", + "content": "receives guidance during the training process and only guides others once they have completed training." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 131, + 337, + 277, + 350 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 131, + 337, + 277, + 350 + ], + "spans": [ + { + "bbox": [ + 131, + 337, + 277, + 350 + ], + "type": "text", + "content": "3.2 Pseudo Label Filtering" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 130, + 361, + 482, + 553 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 361, + 482, + 553 + ], + "spans": [ + { + "bbox": [ + 130, + 361, + 482, + 553 + ], + "type": "text", + "content": "To guide the learning of the current progressive model effectively, we introduce PLF, as shown in Fig. 3 (a). PLF involves a naive approximation where we consider the representational capacity of models during the supervised learning phase as a criterion for judging the representational capacity during the unsupervised learning phase to some extent. During the supervised learning phase, PLF entails inference by all models, and through combining labels, calculates the representational capacity of each model in the current iteration. The metrics for evaluating representational capacity vary with tasks; in this task, the Dice loss is regarded as a statistical measure of model performance. We utilize the representational capacity of the CPM as a threshold and filter out pseudo-labels generated by models with representational capacities below that of the CPM. The pseudo label's passage through the PLF is recorded as " + }, + { + "bbox": [ + 130, + 361, + 482, + 553 + ], + "type": "inline_equation", + "content": "\\mathcal{P}(f_i,f_{i + 1})" + }, + { + "bbox": [ + 130, + 361, + 482, + 553 + ], + "type": "text", + "content": ", where a value of 1 denotes passage and 0 denotes exclusion. This aids in the model's better learning of pseudo-labels with high fidelity. Furthermore, to ensure that the pseudo-labels are accurate and well-defined, we apply a sharpening function to high-fidelity pseudo-labels:" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 242, + 564, + 481, + 595 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 242, + 564, + 481, + 595 + ], + "spans": [ + { + "bbox": [ + 242, + 564, + 481, + 595 + ], + "type": "interline_equation", + "content": "\\mathbf {Y} ^ {U} = \\frac {\\mathbf {P _ {s}} ^ {1 / T}}{\\mathbf {P _ {s}} ^ {1 / T} + (1 - \\mathbf {P _ {s}}) ^ {1 / T}}, \\tag {4}", + "image_path": "91bf23a1fbbe6a52ce31f781ea68a9ee776d70236ad98837372eb5edd97dbce4.jpg" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 131, + 601, + 479, + 613 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 131, + 601, + 479, + 613 + ], + "spans": [ + { + "bbox": [ + 131, + 601, + 479, + 613 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 131, + 601, + 479, + 613 + ], + "type": "inline_equation", + "content": "\\mathbf{P_s}" + }, + { + "bbox": [ + 131, + 601, + 479, + 613 + ], + "type": "text", + "content": " represents the output of the good student, and " + }, + { + "bbox": [ + 131, + 601, + 479, + 613 + ], + "type": "inline_equation", + "content": "T" + }, + { + "bbox": [ + 131, + 601, + 479, + 613 + ], + "type": "text", + "content": " is a hyperparameter." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 131, + 613, + 481, + 638 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 131, + 613, + 481, + 638 + ], + "spans": [ + { + "bbox": [ + 131, + 613, + 481, + 638 + ], + "type": "text", + "content": "The resulting loss function " + }, + { + "bbox": [ + 131, + 613, + 481, + 638 + ], + "type": "inline_equation", + "content": "\\mathrm{L_U}" + }, + { + "bbox": [ + 131, + 613, + 481, + 638 + ], + "type": "text", + "content": " is computed as MSE between the pseudo labels " + }, + { + "bbox": [ + 131, + 613, + 481, + 638 + ], + "type": "inline_equation", + "content": "\\mathbf{Y}^U" + }, + { + "bbox": [ + 131, + 613, + 481, + 638 + ], + "type": "text", + "content": " and the model's predictions of CPM " + }, + { + "bbox": [ + 131, + 613, + 481, + 638 + ], + "type": "inline_equation", + "content": "\\hat{\\mathbf{Y}}_c^U" + }, + { + "bbox": [ + 131, + 613, + 481, + 638 + ], + "type": "text", + "content": ":" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 257, + 651, + 481, + 666 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 257, + 651, + 481, + 666 + ], + "spans": [ + { + "bbox": [ + 257, + 651, + 481, + 666 + ], + "type": "interline_equation", + "content": "\\mathrm {L} _ {\\mathrm {U}} = \\operatorname {M S E} \\left(\\hat {\\mathbf {Y}} _ {c} ^ {U}, \\mathbf {Y} ^ {U}\\right). \\tag {5}", + "image_path": "aab3792543c55cb128f1aea00eb0d1b7e3f79f05317844f96c315030885ec75c.jpg" + } + ] + } + ], + "index": 13 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 340, + 91, + 447, + 102 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 340, + 91, + 447, + 102 + ], + "spans": [ + { + "bbox": [ + 340, + 91, + 447, + 102 + ], + "type": "text", + "content": "Progressive Mean Teacher" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 474, + 91, + 480, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 474, + 91, + 480, + 100 + ], + "spans": [ + { + "bbox": [ + 474, + 91, + 480, + 100 + ], + "type": "text", + "content": "7" + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 6 + }, + { + "para_blocks": [ + { + "bbox": [ + 132, + 116, + 317, + 129 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 116, + 317, + 129 + ], + "spans": [ + { + "bbox": [ + 132, + 116, + 317, + 129 + ], + "type": "text", + "content": "3.3 Discrepancy Driven Alignment" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 130, + 137, + 483, + 390 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 137, + 483, + 390 + ], + "spans": [ + { + "bbox": [ + 130, + 137, + 483, + 390 + ], + "type": "text", + "content": "We propose DDA to compare the outputs of multiple models and achieve rapid alignment between models, as shown in Fig. 3 (b). During the alternating process of progressive Learning, the performance of two groups of models alternately improves, providing stable pseudo-labels to each other. However, concerns often arise due to the disappearance of the performance gap between models caused by random factors. While the expectation is for alternating performance leadership between models, in practice, it may turn into one model consistently leading. Therefore, a fast and high-confidence alignment module is needed. In traditional consistency regularization, contrastive algorithms enforce consistency among the outputs of different models. However, this approach often raises concerns because we are uncertain whether this consistency regularization guides the models in the right direction. DDA focuses on the different parts of predictions among models, which often reflect the differences in representational capabilities between models. We aim to enhance the consistency regularization specifically on these parts to achieve alignment of representational capabilities between models. It is worth noting that we expect the models aligned by DDA to have representation capabilities surpassing CPM. Therefore, DDA will only take effect when PLF allows other models to provide pseudo-labels. The mask of different parts of predictions among different models can be obtained by taking the difference between the union and intersection of the binary outputs of softmax outputs " + }, + { + "bbox": [ + 130, + 137, + 483, + 390 + ], + "type": "inline_equation", + "content": "\\hat{\\mathbf{Y}}^L" + }, + { + "bbox": [ + 130, + 137, + 483, + 390 + ], + "type": "text", + "content": " of multiple networks:" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 218, + 399, + 482, + 433 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 218, + 399, + 482, + 433 + ], + "spans": [ + { + "bbox": [ + 218, + 399, + 482, + 433 + ], + "type": "interline_equation", + "content": "\\mathcal {M} _ {\\text {d i f f}} = \\bigcup_ {i = 1} ^ {n} \\operatorname {B I N A} \\left(\\hat {\\mathbf {Y}} _ {i} ^ {L}\\right) - \\bigcap_ {i = 1} ^ {N} \\operatorname {B I N A} \\left(\\hat {\\mathbf {Y}} _ {i} ^ {L}\\right). \\tag {6}", + "image_path": "d9a3cceae7cca7b875a1160ed63e3b4dbbee0b58700f6cebd16588e80919fb88.jpg" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 132, + 437, + 483, + 463 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 437, + 483, + 463 + ], + "spans": [ + { + "bbox": [ + 132, + 437, + 483, + 463 + ], + "type": "text", + "content": "Subsequently, MSE of the model's predictions is computed and denoted as " + }, + { + "bbox": [ + 132, + 437, + 483, + 463 + ], + "type": "inline_equation", + "content": "\\mathcal{M}_{\\mathrm{dist}}" + }, + { + "bbox": [ + 132, + 437, + 483, + 463 + ], + "type": "text", + "content": ":" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 251, + 472, + 481, + 492 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 251, + 472, + 481, + 492 + ], + "spans": [ + { + "bbox": [ + 251, + 472, + 481, + 492 + ], + "type": "interline_equation", + "content": "\\mathcal {M} _ {\\mathrm {d i s t}} = \\operatorname {M S E} \\left(\\hat {\\mathbf {Y}} _ {i} ^ {L}, \\mathbf {Y} ^ {L}\\right). \\tag {7}", + "image_path": "9fd4ac1cb0cc3e0207dd0f25aad002ebdbe9748386a4753a5f884d8561b08824.jpg" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 130, + 496, + 482, + 544 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 496, + 482, + 544 + ], + "spans": [ + { + "bbox": [ + 130, + 496, + 482, + 544 + ], + "type": "text", + "content": "Next, the total count of elements in " + }, + { + "bbox": [ + 130, + 496, + 482, + 544 + ], + "type": "inline_equation", + "content": "\\mathcal{M}_{\\mathrm{diff}}" + }, + { + "bbox": [ + 130, + 496, + 482, + 544 + ], + "type": "text", + "content": " is calculated, and " + }, + { + "bbox": [ + 130, + 496, + 482, + 544 + ], + "type": "inline_equation", + "content": "\\mathcal{M}_{\\mathrm{diff}}" + }, + { + "bbox": [ + 130, + 496, + 482, + 544 + ], + "type": "text", + "content": " is used to mask " + }, + { + "bbox": [ + 130, + 496, + 482, + 544 + ], + "type": "inline_equation", + "content": "\\mathcal{M}_{\\mathrm{dist}}" + }, + { + "bbox": [ + 130, + 496, + 482, + 544 + ], + "type": "text", + "content": ". The ratio of the sum of the masked " + }, + { + "bbox": [ + 130, + 496, + 482, + 544 + ], + "type": "inline_equation", + "content": "\\mathcal{M}_{\\mathrm{dist}}" + }, + { + "bbox": [ + 130, + 496, + 482, + 544 + ], + "type": "text", + "content": " to the total count of elements in " + }, + { + "bbox": [ + 130, + 496, + 482, + 544 + ], + "type": "inline_equation", + "content": "\\mathcal{M}_{\\mathrm{diff}}" + }, + { + "bbox": [ + 130, + 496, + 482, + 544 + ], + "type": "text", + "content": " is denoted as " + }, + { + "bbox": [ + 130, + 496, + 482, + 544 + ], + "type": "inline_equation", + "content": "\\mathrm{L_{aln}}" + }, + { + "bbox": [ + 130, + 496, + 482, + 544 + ], + "type": "text", + "content": ", which characterizes the gap to the ground truth within the prediction differences between two models:" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 250, + 554, + 481, + 582 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 250, + 554, + 481, + 582 + ], + "spans": [ + { + "bbox": [ + 250, + 554, + 481, + 582 + ], + "type": "interline_equation", + "content": "\\mathrm {L} _ {\\mathrm {a l n}} = \\frac {\\sum \\left(\\mathcal {M} _ {\\mathrm {d i f f}} \\cdot \\mathcal {M} _ {\\mathrm {d i s t}}\\right)}{\\sum \\left(\\mathcal {M} _ {\\mathrm {d i f f}}\\right)}. \\tag {8}", + "image_path": "397dfe475261fcec2a3c29c13faa0ed12f5de2102d1264a4bd60c7b8d9287539.jpg" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 132, + 597, + 231, + 609 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 597, + 231, + 609 + ], + "spans": [ + { + "bbox": [ + 132, + 597, + 231, + 609 + ], + "type": "text", + "content": "3.4 Loss Function" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 130, + 619, + 482, + 643 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 619, + 482, + 643 + ], + "spans": [ + { + "bbox": [ + 130, + 619, + 482, + 643 + ], + "type": "text", + "content": "In general, during an iteration cycle, CPM conducts the process of backward propagation concerning its loss function, which is defined as follow:" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 224, + 654, + 481, + 667 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 224, + 654, + 481, + 667 + ], + "spans": [ + { + "bbox": [ + 224, + 654, + 481, + 667 + ], + "type": "interline_equation", + "content": "\\mathrm {L} _ {\\text {t o t a l}} = \\mathrm {L} _ {\\mathrm {s}} + \\lambda_ {1} \\mathcal {P} (f _ {i}, f _ {i + 1}) \\mathrm {L} _ {\\mathrm {U}} + \\lambda_ {2} \\mathrm {L} _ {\\mathrm {T}}, \\tag {9}", + "image_path": "5fbb20a3f2199c4cb7b49905a7e12c7eb03332005b58654ba7c8d7791ecad196.jpg" + } + ] + } + ], + "index": 11 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "spans": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "type": "text", + "content": "8" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 166, + 91, + 220, + 101 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 166, + 91, + 220, + 101 + ], + "spans": [ + { + "bbox": [ + 166, + 91, + 220, + 101 + ], + "type": "text", + "content": "N. Gao et al." + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 7 + }, + { + "para_blocks": [ + { + "bbox": [ + 131, + 116, + 479, + 140 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 131, + 116, + 479, + 140 + ], + "spans": [ + { + "bbox": [ + 131, + 116, + 479, + 140 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 131, + 116, + 479, + 140 + ], + "type": "inline_equation", + "content": "\\lambda_{1}" + }, + { + "bbox": [ + 131, + 116, + 479, + 140 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 131, + 116, + 479, + 140 + ], + "type": "inline_equation", + "content": "\\lambda_{2}" + }, + { + "bbox": [ + 131, + 116, + 479, + 140 + ], + "type": "text", + "content": " increase as the number of iterations grows, up to a point where they stop growing after a fixed iteration." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 131, + 140, + 479, + 165 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 131, + 140, + 479, + 165 + ], + "spans": [ + { + "bbox": [ + 131, + 140, + 479, + 165 + ], + "type": "text", + "content": "We employed two independent Gaussian warm-up function to control the loss function weights, " + }, + { + "bbox": [ + 131, + 140, + 479, + 165 + ], + "type": "inline_equation", + "content": "\\lambda_{1}" + }, + { + "bbox": [ + 131, + 140, + 479, + 165 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 131, + 140, + 479, + 165 + ], + "type": "inline_equation", + "content": "\\lambda_{2}" + }, + { + "bbox": [ + 131, + 140, + 479, + 165 + ], + "type": "text", + "content": ", using different parameters:" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 225, + 176, + 480, + 213 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 225, + 176, + 480, + 213 + ], + "spans": [ + { + "bbox": [ + 225, + 176, + 480, + 213 + ], + "type": "interline_equation", + "content": "\\lambda_ {1} (t) = \\left\\{ \\begin{array}{l l} \\hat {\\lambda} _ {1} \\cdot e ^ {- 5 \\left(1 - \\frac {2 t}{t _ {\\operatorname* {m a x}}}\\right) ^ {2}}, & t < \\frac {t _ {\\operatorname* {m a x}}}{2} \\\\ \\hat {\\lambda} _ {1}, & t \\geq \\frac {t _ {\\operatorname* {m a x}}}{2} \\end{array} \\right. \\tag {10}", + "image_path": "826d0d40e267e0165752096e3af2c12f1c0c4c0da785f61543b70a077b6634c9.jpg" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 225, + 212, + 388, + 242 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 225, + 212, + 388, + 242 + ], + "spans": [ + { + "bbox": [ + 225, + 212, + 388, + 242 + ], + "type": "interline_equation", + "content": "\\lambda_ {2} (t) = \\left\\{ \\begin{array}{l l} \\hat {\\lambda} _ {2} \\cdot e ^ {- 5 \\left(1 - \\frac {2 t}{t _ {\\max}}\\right) ^ {2}}, & t < \\frac {t _ {\\max}}{2} \\\\ \\hat {\\lambda} _ {2}, & t \\geq \\frac {t _ {\\max}}{2} \\end{array} \\right.", + "image_path": "631794b5a76525ba90c6b936004038b696ace87b6b49f9e1facae7bc11e24dda.jpg" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 131, + 248, + 479, + 282 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 131, + 248, + 479, + 282 + ], + "spans": [ + { + "bbox": [ + 131, + 248, + 479, + 282 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 131, + 248, + 479, + 282 + ], + "type": "inline_equation", + "content": "t" + }, + { + "bbox": [ + 131, + 248, + 479, + 282 + ], + "type": "text", + "content": " represents the current iteration number, and " + }, + { + "bbox": [ + 131, + 248, + 479, + 282 + ], + "type": "inline_equation", + "content": "t_{\\mathrm{max}}" + }, + { + "bbox": [ + 131, + 248, + 479, + 282 + ], + "type": "text", + "content": " represents the total number of training iterations. The hyperparameter " + }, + { + "bbox": [ + 131, + 248, + 479, + 282 + ], + "type": "inline_equation", + "content": "\\hat{\\lambda}_1" + }, + { + "bbox": [ + 131, + 248, + 479, + 282 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 131, + 248, + 479, + 282 + ], + "type": "inline_equation", + "content": "\\hat{\\lambda}_2" + }, + { + "bbox": [ + 131, + 248, + 479, + 282 + ], + "type": "text", + "content": " were empirically set to 20.0 and 10.0." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 131, + 285, + 479, + 308 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 131, + 285, + 479, + 308 + ], + "spans": [ + { + "bbox": [ + 131, + 285, + 479, + 308 + ], + "type": "text", + "content": "Loss function " + }, + { + "bbox": [ + 131, + 285, + 479, + 308 + ], + "type": "inline_equation", + "content": "\\mathrm{L_s}" + }, + { + "bbox": [ + 131, + 285, + 479, + 308 + ], + "type": "text", + "content": " represents the loss during supervised learning and consists of the following components:" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 198, + 323, + 479, + 335 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 198, + 323, + 479, + 335 + ], + "spans": [ + { + "bbox": [ + 198, + 323, + 479, + 335 + ], + "type": "interline_equation", + "content": "\\mathrm {L} _ {\\mathrm {s}} = \\mathrm {C E} + \\mathrm {D I C E} + \\beta \\mathcal {P} \\left(f _ {i}, f _ {i + 1}\\right) \\mathrm {L} _ {\\mathrm {a l n}} \\left(\\mathcal {M} _ {\\text {d i f f}}, \\mathcal {M} _ {\\text {d i s t}}\\right), \\tag {11}", + "image_path": "a2825bd3d4bd036633dfd6ca8c26ed469d2680abacf9b423b356f335568eb7aa.jpg" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 131, + 342, + 479, + 353 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 131, + 342, + 479, + 353 + ], + "spans": [ + { + "bbox": [ + 131, + 342, + 479, + 353 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 131, + 342, + 479, + 353 + ], + "type": "inline_equation", + "content": "\\beta" + }, + { + "bbox": [ + 131, + 342, + 479, + 353 + ], + "type": "text", + "content": ", a constant set to 0.5, is used to balance alignment loss and other losses." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 131, + 354, + 479, + 401 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 131, + 354, + 479, + 401 + ], + "spans": [ + { + "bbox": [ + 131, + 354, + 479, + 401 + ], + "type": "text", + "content": "Loss function " + }, + { + "bbox": [ + 131, + 354, + 479, + 401 + ], + "type": "inline_equation", + "content": "\\mathrm{L}_{\\mathrm{T}}" + }, + { + "bbox": [ + 131, + 354, + 479, + 401 + ], + "type": "text", + "content": " is the consistency loss generated by the MT, which is derived by the teacher through the sharpen function to convert the model outputs into pseudo-labels, and subsequently generated by computing the MSE with the student model:" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 206, + 414, + 479, + 445 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 206, + 414, + 479, + 445 + ], + "spans": [ + { + "bbox": [ + 206, + 414, + 479, + 445 + ], + "type": "interline_equation", + "content": "\\mathrm {L} _ {\\mathrm {T}} = \\operatorname {M S E} \\left(f _ {i} \\left(\\mathbf {X} ^ {U}\\right), \\frac {\\mathbf {P} _ {\\mathbf {t}} ^ {1 / T}}{\\mathbf {P} _ {\\mathbf {t}} ^ {1 / T} + (1 - \\mathbf {P} _ {\\mathbf {t}}) ^ {1 / T}}\\right). \\tag {12}", + "image_path": "fef0208d54175a048710f335f5762767008bd942906858d1ba9712026cd01a91.jpg" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 132, + 471, + 229, + 485 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 471, + 229, + 485 + ], + "spans": [ + { + "bbox": [ + 132, + 471, + 229, + 485 + ], + "type": "text", + "content": "4 Experiments" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 132, + 498, + 279, + 510 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 498, + 279, + 510 + ], + "spans": [ + { + "bbox": [ + 132, + 498, + 279, + 510 + ], + "type": "text", + "content": "4.1 Implementation Details" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 131, + 521, + 479, + 604 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 131, + 521, + 479, + 604 + ], + "spans": [ + { + "bbox": [ + 131, + 521, + 479, + 604 + ], + "type": "text", + "content": "We selected the VNet [13] model as a baseline network, which performs well in conditions with limited data and is essentially a 3D convolutional version of UNet. During inference, we use the average of the outputs from two networks as the final prediction. Specifically, the SGD optimizer was used to update the network parameters with weight decay of 0.0001 and a momentum of 0.9. The initial learning rate was set to 0.01, divided by 10 every 2500 iterations, for a total of 6000 iterations." + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 131, + 605, + 479, + 665 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 131, + 605, + 479, + 665 + ], + "spans": [ + { + "bbox": [ + 131, + 605, + 479, + 665 + ], + "type": "text", + "content": "Following the practice in comparative literature [11, 12, 23, 30], our methods are trained for a fixed number of 6,000 iterations to obtain the final model. Additionally, our models all use a batch size of 4, with a labeled data quantity of 2. We tested the performance of the models, and all experiments were conducted on NVIDIA® GeForce A40 48GB running Ubuntu 20.04 and PyTorch 1.11.0." + } + ] + } + ], + "index": 15 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 340, + 91, + 447, + 102 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 340, + 91, + 447, + 102 + ], + "spans": [ + { + "bbox": [ + 340, + 91, + 447, + 102 + ], + "type": "text", + "content": "Progressive Mean Teacher" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 474, + 91, + 481, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 474, + 91, + 481, + 100 + ], + "spans": [ + { + "bbox": [ + 474, + 91, + 481, + 100 + ], + "type": "text", + "content": "9" + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 8 + }, + { + "para_blocks": [ + { + "bbox": [ + 132, + 116, + 268, + 126 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 116, + 268, + 126 + ], + "spans": [ + { + "bbox": [ + 132, + 116, + 268, + 126 + ], + "type": "text", + "content": "4.2 Datasets and Metrics" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 130, + 140, + 479, + 186 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 140, + 479, + 186 + ], + "spans": [ + { + "bbox": [ + 130, + 140, + 479, + 186 + ], + "type": "text", + "content": "In the experiment, we selected two datasets with different modalities and utilized four distinct metrics to assess the performance of the model. For each dataset, " + }, + { + "bbox": [ + 130, + 140, + 479, + 186 + ], + "type": "inline_equation", + "content": "80\\%" + }, + { + "bbox": [ + 130, + 140, + 479, + 186 + ], + "type": "text", + "content": " of the data was used as the training set and " + }, + { + "bbox": [ + 130, + 140, + 479, + 186 + ], + "type": "inline_equation", + "content": "20\\%" + }, + { + "bbox": [ + 130, + 140, + 479, + 186 + ], + "type": "text", + "content": " as the test set. The proportion of supervised data was determined based on the training set." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 130, + 188, + 480, + 282 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 188, + 480, + 282 + ], + "spans": [ + { + "bbox": [ + 130, + 188, + 480, + 282 + ], + "type": "text", + "content": "LA Dataset. It [28] includes 100 3D gadolinium-enhanced MR imaging volumes of left atrial with an isotropic resolution of " + }, + { + "bbox": [ + 130, + 188, + 480, + 282 + ], + "type": "inline_equation", + "content": "0.625 \\times 0.625 \\times 0.625mm^3" + }, + { + "bbox": [ + 130, + 188, + 480, + 282 + ], + "type": "text", + "content": " and the corresponding ground truth labels. For pre-processing, we first normalize all volumes to zero mean and unit variance, then crop each 3D MRI volume with enlarged margins according to the targets. During training, the training volumes are randomly cropped to " + }, + { + "bbox": [ + 130, + 188, + 480, + 282 + ], + "type": "inline_equation", + "content": "112 \\times 112 \\times 80" + }, + { + "bbox": [ + 130, + 188, + 480, + 282 + ], + "type": "text", + "content": " as the model input. During inference, a sliding window of the same size is used to obtain segmentation results with a stride of " + }, + { + "bbox": [ + 130, + 188, + 480, + 282 + ], + "type": "inline_equation", + "content": "18 \\times 18 \\times 4" + }, + { + "bbox": [ + 130, + 188, + 480, + 282 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 130, + 285, + 480, + 367 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 285, + 480, + 367 + ], + "spans": [ + { + "bbox": [ + 130, + 285, + 480, + 367 + ], + "type": "text", + "content": "Pancreas-NIH Dataset. It [17] provides 82 contrast-enhanced abdominal 3D CT volumes of pancreas with manual annotation. The size of each CT volume is " + }, + { + "bbox": [ + 130, + 285, + 480, + 367 + ], + "type": "inline_equation", + "content": "512 \\times 512 \\times D" + }, + { + "bbox": [ + 130, + 285, + 480, + 367 + ], + "type": "text", + "content": ", where " + }, + { + "bbox": [ + 130, + 285, + 480, + 367 + ], + "type": "inline_equation", + "content": "D \\in [181, 466]" + }, + { + "bbox": [ + 130, + 285, + 480, + 367 + ], + "type": "text", + "content": ". In pre-processing, we use the soft tissue CT window of " + }, + { + "bbox": [ + 130, + 285, + 480, + 367 + ], + "type": "inline_equation", + "content": "[-120, 240]" + }, + { + "bbox": [ + 130, + 285, + 480, + 367 + ], + "type": "text", + "content": " HU, and we crop the CT scans centering at the pancreas region, and enlarge margins with 25 voxels. The training volumes are randomly cropped to " + }, + { + "bbox": [ + 130, + 285, + 480, + 367 + ], + "type": "inline_equation", + "content": "96 \\times 96 \\times 96" + }, + { + "bbox": [ + 130, + 285, + 480, + 367 + ], + "type": "text", + "content": " as the model input. During inference, a sliding window of the same size is used to obtain segmentation results with a stride of " + }, + { + "bbox": [ + 130, + 285, + 480, + 367 + ], + "type": "inline_equation", + "content": "16 \\times 16 \\times 16" + }, + { + "bbox": [ + 130, + 285, + 480, + 367 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 130, + 369, + 480, + 430 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 369, + 480, + 430 + ], + "spans": [ + { + "bbox": [ + 130, + 369, + 480, + 430 + ], + "type": "text", + "content": "Metrics. Following [2, 11, 12, 23, 29, 30], we use four metrics to evaluate model performance, including regional sensitive metrics: Dice similarity coefficient (Dice) [30], Jaccard similarity coefficient (Jaccard) [12], and edge sensitive metrics: " + }, + { + "bbox": [ + 130, + 369, + 480, + 430 + ], + "type": "inline_equation", + "content": "95\\%" + }, + { + "bbox": [ + 130, + 369, + 480, + 430 + ], + "type": "text", + "content": " Hausdorff Distance (95HD) [29] and Average Surface Distance (ASD) [2]." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 132, + 449, + 237, + 462 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 449, + 237, + 462 + ], + "spans": [ + { + "bbox": [ + 132, + 449, + 237, + 462 + ], + "type": "text", + "content": "4.3 Ablation Study" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 130, + 473, + 479, + 509 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 473, + 479, + 509 + ], + "spans": [ + { + "bbox": [ + 130, + 473, + 479, + 509 + ], + "type": "text", + "content": "For simplicity, we conducted ablation experiments on LA dataset to evaluate our design choices for each component. For a reliable assessment, all other parts were kept consistent except for the component under investigation." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 130, + 510, + 480, + 665 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 510, + 480, + 665 + ], + "spans": [ + { + "bbox": [ + 130, + 510, + 480, + 665 + ], + "type": "text", + "content": "Analysis of PMT Framework. The PMT architecture under progressive design is at the core of our work, wherein progressive design ensures the continuous generation of diverse pseudo-labels during the forward propagation process. Within the PMT architecture, PLF and DDA respectively denote the pseudolabel filtering and model alignment methods proposed in our paper, while MT ensures stable enhancement of model capability. We employ a non-progressive design co-training framework consisting of two VNet components as the baseline model. When adding PLF and DDA methods to the baseline model, the progressive design is simultaneously incorporated into the baseline model. Finally, we evaluate the impact of the MT architecture on model representation capability, with results presented in Table 1. Overall, compared to the baseline model, which already possesses good representation capability, the PMT framework greatly enhances the model's representation capability." + } + ] + } + ], + "index": 9 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "spans": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "type": "text", + "content": "10" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 166, + 91, + 220, + 100 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 166, + 91, + 220, + 100 + ], + "spans": [ + { + "bbox": [ + 166, + 91, + 220, + 100 + ], + "type": "text", + "content": "N. Gao et al." + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 9 + }, + { + "para_blocks": [ + { + "type": "table", + "bbox": [ + 145, + 135, + 466, + 232 + ], + "blocks": [ + { + "bbox": [ + 168, + 114, + 444, + 125 + ], + "lines": [ + { + "bbox": [ + 168, + 114, + 444, + 125 + ], + "spans": [ + { + "bbox": [ + 168, + 114, + 444, + 125 + ], + "type": "text", + "content": "Table 1: Ablation results about PMT framework on LA dataset" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 145, + 135, + 466, + 232 + ], + "lines": [ + { + "bbox": [ + 145, + 135, + 466, + 232 + ], + "spans": [ + { + "bbox": [ + 145, + 135, + 466, + 232 + ], + "type": "table", + "html": "
MethodLabeledMetrics
PLFDDAMTDice↑Jaccard↑95HD↓ASD↓
---8(10%)88.2879.317.592.34
✓--8(10%)89.9081.736.141.72
-✓-8(10%)90.4382.605.981.66
✓✓-8(10%)90.4382.605.491.49
✓✓✓8(10%)90.8183.235.611.50
", + "image_path": "f9ee592f83af732a3750272dcb1ec1c02f09820ba2f1f65e88da46d391600aba.jpg" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "table_body" + } + ], + "index": 3 + }, + { + "bbox": [ + 130, + 250, + 482, + 360 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 250, + 482, + 360 + ], + "spans": [ + { + "bbox": [ + 130, + 250, + 482, + 360 + ], + "type": "text", + "content": "Analysis of Regularization Strength. The hyperparameters " + }, + { + "bbox": [ + 130, + 250, + 482, + 360 + ], + "type": "inline_equation", + "content": "\\hat{\\lambda}_1" + }, + { + "bbox": [ + 130, + 250, + 482, + 360 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 130, + 250, + 482, + 360 + ], + "type": "inline_equation", + "content": "\\hat{\\lambda}_2" + }, + { + "bbox": [ + 130, + 250, + 482, + 360 + ], + "type": "text", + "content": " respectively characterize the regularization strength of the progressive methods and MT architecture on the model. We use a tenfold scale and scale " + }, + { + "bbox": [ + 130, + 250, + 482, + 360 + ], + "type": "inline_equation", + "content": "\\hat{\\lambda}_1" + }, + { + "bbox": [ + 130, + 250, + 482, + 360 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 130, + 250, + 482, + 360 + ], + "type": "inline_equation", + "content": "\\hat{\\lambda}_2" + }, + { + "bbox": [ + 130, + 250, + 482, + 360 + ], + "type": "text", + "content": " based on the hyperparameter specifications we use, as shown in Table 2. The results indicate that variations in the two parameters within a certain range do not significantly affect the model's performance, demonstrating a certain level of robustness of our model to parameter variations. Within a certain range, on the LA dataset with " + }, + { + "bbox": [ + 130, + 250, + 482, + 360 + ], + "type": "inline_equation", + "content": "10\\%" + }, + { + "bbox": [ + 130, + 250, + 482, + 360 + ], + "type": "text", + "content": " labeled data, our model achieves the best performance when " + }, + { + "bbox": [ + 130, + 250, + 482, + 360 + ], + "type": "inline_equation", + "content": "\\hat{\\lambda}_1" + }, + { + "bbox": [ + 130, + 250, + 482, + 360 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 130, + 250, + 482, + 360 + ], + "type": "inline_equation", + "content": "\\hat{\\lambda}_2" + }, + { + "bbox": [ + 130, + 250, + 482, + 360 + ], + "type": "text", + "content": " are 20.0 and 10.0 respectively." + } + ] + } + ], + "index": 4 + }, + { + "type": "table", + "bbox": [ + 148, + 399, + 462, + 539 + ], + "blocks": [ + { + "bbox": [ + 153, + 378, + 459, + 391 + ], + "lines": [ + { + "bbox": [ + 153, + 378, + 459, + 391 + ], + "spans": [ + { + "bbox": [ + 153, + 378, + 459, + 391 + ], + "type": "text", + "content": "Table 2: Ablation results about regularization strength on LA dataset" + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 148, + 399, + 462, + 539 + ], + "lines": [ + { + "bbox": [ + 148, + 399, + 462, + 539 + ], + "spans": [ + { + "bbox": [ + 148, + 399, + 462, + 539 + ], + "type": "table", + "html": "
λLabeledMetrics
λ1maxλ2maxDice↑Jaccard↑95HD↓ASD↓
2.01.08(10%)89.6681.336.351.84
2.010.08(10%)90.4782.665.581.64
2.0100.08(10%)87.7478.289.332.84
20.01.08(10%)90.6782.986.311.77
20.010.08(10%)90.8183.235.611.50
20.0100.08(10%)88.0078.739.052.78
200.01.08(10%)89.0880.4211.403.07
200.010.08(10%)89.7581.477.082.22
200.0100.08(10%)83.8072.5516.445.11
", + "image_path": "09e9aabbc4b197991aefacd36a7babddd50585b8b8a60a93d9b2d32ec6d08db7.jpg" + } + ] + } + ], + "index": 6, + "angle": 0, + "type": "table_body" + } + ], + "index": 6 + }, + { + "bbox": [ + 132, + 575, + 328, + 588 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 575, + 328, + 588 + ], + "spans": [ + { + "bbox": [ + 132, + 575, + 328, + 588 + ], + "type": "text", + "content": "4.4 Comparison with Other Methods" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 130, + 594, + 479, + 616 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 594, + 479, + 616 + ], + "spans": [ + { + "bbox": [ + 130, + 594, + 479, + 616 + ], + "type": "text", + "content": "We compared our approach with previous state-of-the-art methods on LA dataset and Pancreas-NIH dataset." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 130, + 617, + 480, + 666 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 617, + 480, + 666 + ], + "spans": [ + { + "bbox": [ + 130, + 617, + 480, + 666 + ], + "type": "text", + "content": "We chose VNet as baseline models for comparison. For the selected alternative models, we opted for UA-MT [30] with uncertainty estimation, SASSNet [11] focusing on the regularity of geometric shapes, DTC [12] with task-level regularization, BCP [2] using bidirectional CutMix [31], and MCF [23] with model-level" + } + ] + } + ], + "index": 9 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 340, + 91, + 447, + 102 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 340, + 91, + 447, + 102 + ], + "spans": [ + { + "bbox": [ + 340, + 91, + 447, + 102 + ], + "type": "text", + "content": "Progressive Mean Teacher" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 470, + 91, + 479, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 470, + 91, + 479, + 100 + ], + "spans": [ + { + "bbox": [ + 470, + 91, + 479, + 100 + ], + "type": "text", + "content": "11" + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 10 + }, + { + "para_blocks": [ + { + "bbox": [ + 130, + 116, + 479, + 152 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 116, + 479, + 152 + ], + "spans": [ + { + "bbox": [ + 130, + 116, + 479, + 152 + ], + "type": "text", + "content": "regularization, with BCP and MCF being state-of-the-art results. Noting that, for BCP, we follow its parameter settings for pre-training 2,000 times and self-training 15,000 times." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 130, + 152, + 481, + 224 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 152, + 481, + 224 + ], + "spans": [ + { + "bbox": [ + 130, + 152, + 481, + 224 + ], + "type": "text", + "content": "Comparison on LA Dataset. We conducted a cross-model comparison on the classic LA dataset. We tested with " + }, + { + "bbox": [ + 130, + 152, + 481, + 224 + ], + "type": "inline_equation", + "content": "5\\%" + }, + { + "bbox": [ + 130, + 152, + 481, + 224 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 130, + 152, + 481, + 224 + ], + "type": "inline_equation", + "content": "10\\%" + }, + { + "bbox": [ + 130, + 152, + 481, + 224 + ], + "type": "text", + "content": " of labeled data. Results of the experiments are presented in Table 3. To provide a more intuitive demonstration of the performance of various models on the LA dataset, we have selected some representative results for visualization, as illustrated in Fig. 4. Areas with inaccurate segmentation have been annotated accordingly." + } + ] + } + ], + "index": 3 + }, + { + "type": "table", + "bbox": [ + 145, + 264, + 466, + 453 + ], + "blocks": [ + { + "bbox": [ + 151, + 243, + 462, + 255 + ], + "lines": [ + { + "bbox": [ + 151, + 243, + 462, + 255 + ], + "spans": [ + { + "bbox": [ + 151, + 243, + 462, + 255 + ], + "type": "text", + "content": "Table 3: Comparison results on LA dataset with " + }, + { + "bbox": [ + 151, + 243, + 462, + 255 + ], + "type": "inline_equation", + "content": "5\\%" + }, + { + "bbox": [ + 151, + 243, + 462, + 255 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 151, + 243, + 462, + 255 + ], + "type": "inline_equation", + "content": "10\\%" + }, + { + "bbox": [ + 151, + 243, + 462, + 255 + ], + "type": "text", + "content": " labeled data" + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 145, + 264, + 466, + 453 + ], + "lines": [ + { + "bbox": [ + 145, + 264, + 466, + 453 + ], + "spans": [ + { + "bbox": [ + 145, + 264, + 466, + 453 + ], + "type": "table", + "html": "
MethodLabeledMetrics
Dice↑Jaccard↑95HD↓ASD↓
VNet4(5%)52.5539.6047.059.87
UA-MT4(5%)82.2670.9813.713.82
SASSNet4(5%)81.6069.6316.163.58
DTC4(5%)81.2569.3314.903.99
MCFSOTA4(5%)----
BCPSOTA4(5%)88.0278.727.902.15
PMT(Ours)4(5%)89.4781.046.451.86
VNet8(10%)82.7471.7213.353.26
UA-MT8(10%)86.2876.1118.714.63
SASSNet8(10%)85.2275.0911.182.89
MCFSOTA8(10%)88.7180.416.321.90
BCPSOTA8(10%)89.6281.316.811.76
PMT(Ours)8(10%)90.8183.235.611.50
", + "image_path": "3e4e4610f48e2fd2241950049299ca5be2dffcf4a882f6228ce47ba87981d950.jpg" + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "table_body" + } + ], + "index": 5 + }, + { + "bbox": [ + 130, + 474, + 482, + 628 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 474, + 482, + 628 + ], + "spans": [ + { + "bbox": [ + 130, + 474, + 482, + 628 + ], + "type": "text", + "content": "Our model outperformed previous models on all four metrics across " + }, + { + "bbox": [ + 130, + 474, + 482, + 628 + ], + "type": "inline_equation", + "content": "5\\%" + }, + { + "bbox": [ + 130, + 474, + 482, + 628 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 130, + 474, + 482, + 628 + ], + "type": "inline_equation", + "content": "10\\%" + }, + { + "bbox": [ + 130, + 474, + 482, + 628 + ], + "type": "text", + "content": " labeled data. At " + }, + { + "bbox": [ + 130, + 474, + 482, + 628 + ], + "type": "inline_equation", + "content": "5\\%" + }, + { + "bbox": [ + 130, + 474, + 482, + 628 + ], + "type": "text", + "content": " of the data, compared to the best results from previous work, our PMT showed a " + }, + { + "bbox": [ + 130, + 474, + 482, + 628 + ], + "type": "inline_equation", + "content": "1.45\\%" + }, + { + "bbox": [ + 130, + 474, + 482, + 628 + ], + "type": "text", + "content": " improvement in Dice, a " + }, + { + "bbox": [ + 130, + 474, + 482, + 628 + ], + "type": "inline_equation", + "content": "2.32\\%" + }, + { + "bbox": [ + 130, + 474, + 482, + 628 + ], + "type": "text", + "content": " improvement in Jaccard, a reduction of 1.45 in 95HD, and a reduction of 0.29 in ASD. Our PMT also showed a " + }, + { + "bbox": [ + 130, + 474, + 482, + 628 + ], + "type": "inline_equation", + "content": "1.19\\%" + }, + { + "bbox": [ + 130, + 474, + 482, + 628 + ], + "type": "text", + "content": " improvement in Dice, a " + }, + { + "bbox": [ + 130, + 474, + 482, + 628 + ], + "type": "inline_equation", + "content": "1.92\\%" + }, + { + "bbox": [ + 130, + 474, + 482, + 628 + ], + "type": "text", + "content": " improvement in Jaccard, a reduction of 1.20 in 95HD, and a reduction of 0.26 in ASD at " + }, + { + "bbox": [ + 130, + 474, + 482, + 628 + ], + "type": "inline_equation", + "content": "10\\%" + }, + { + "bbox": [ + 130, + 474, + 482, + 628 + ], + "type": "text", + "content": " of the data. It is noteworthy that our model maintains good performance even with a small amount of data. In comparison with state-of-the-art results, our model achieves a leading performance in the majority of metrics using only half of their data. For instance, we surpass MCF's performance at " + }, + { + "bbox": [ + 130, + 474, + 482, + 628 + ], + "type": "inline_equation", + "content": "10\\%" + }, + { + "bbox": [ + 130, + 474, + 482, + 628 + ], + "type": "text", + "content": " labeled data using only " + }, + { + "bbox": [ + 130, + 474, + 482, + 628 + ], + "type": "inline_equation", + "content": "5\\%" + }, + { + "bbox": [ + 130, + 474, + 482, + 628 + ], + "type": "text", + "content": " labeled data. The above indicates that our model achieves superior results compared to existing methods on differently proportioned annotated data in the task of segmenting the left atrium." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 130, + 630, + 481, + 665 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 630, + 481, + 665 + ], + "spans": [ + { + "bbox": [ + 130, + 630, + 481, + 665 + ], + "type": "text", + "content": "Comparison on Pancreas-NIH Dataset. We conducted a cross-model comparison on the classic Pancreas dataset. Detailed results of the experiments are presented in Table 4. We tested with " + }, + { + "bbox": [ + 130, + 630, + 481, + 665 + ], + "type": "inline_equation", + "content": "10\\%" + }, + { + "bbox": [ + 130, + 630, + 481, + 665 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 130, + 630, + 481, + 665 + ], + "type": "inline_equation", + "content": "20\\%" + }, + { + "bbox": [ + 130, + 630, + 481, + 665 + ], + "type": "text", + "content": " of labeled data. To" + } + ] + } + ], + "index": 7 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "spans": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "type": "text", + "content": "12" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 166, + 91, + 220, + 100 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 166, + 91, + 220, + 100 + ], + "spans": [ + { + "bbox": [ + 166, + 91, + 220, + 100 + ], + "type": "text", + "content": "N. Gao et al." + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 11 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 132, + 114, + 481, + 355 + ], + "blocks": [ + { + "bbox": [ + 132, + 114, + 481, + 355 + ], + "lines": [ + { + "bbox": [ + 132, + 114, + 481, + 355 + ], + "spans": [ + { + "bbox": [ + 132, + 114, + 481, + 355 + ], + "type": "image", + "image_path": "6f78183ef03b2653e9009c0a2726c7e3b72f0700e96b436b845a88d012677779.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 132, + 362, + 480, + 384 + ], + "lines": [ + { + "bbox": [ + 132, + 362, + 480, + 384 + ], + "spans": [ + { + "bbox": [ + 132, + 362, + 480, + 384 + ], + "type": "text", + "content": "Fig. 4: 2D & 3D segmentation visualization of different semi-supervised methods under " + }, + { + "bbox": [ + 132, + 362, + 480, + 384 + ], + "type": "inline_equation", + "content": "10\\%" + }, + { + "bbox": [ + 132, + 362, + 480, + 384 + ], + "type": "text", + "content": " labeled on LA (upper) and pancreas (bottom) dataset." + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_caption" + } + ], + "index": 2 + }, + { + "bbox": [ + 130, + 410, + 480, + 493 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 410, + 480, + 493 + ], + "spans": [ + { + "bbox": [ + 130, + 410, + 480, + 493 + ], + "type": "text", + "content": "provide a more intuitive demonstration of the performance of various models on the Pancreas-NIH dataset, we have selected some representative results for visualization, as illustrated in Fig. 4. Areas with inaccurate segmentation have been annotated accordingly. It is worth noting that the results in the table clearly indicate that Pancreas-NIH dataset is significantly more challenging than LA dataset. Therefore, we increased " + }, + { + "bbox": [ + 130, + 410, + 480, + 493 + ], + "type": "inline_equation", + "content": "\\hat{\\lambda}_{1}" + }, + { + "bbox": [ + 130, + 410, + 480, + 493 + ], + "type": "text", + "content": " by a factor of two, resulting in improved performance." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 130, + 495, + 481, + 590 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 495, + 481, + 590 + ], + "spans": [ + { + "bbox": [ + 130, + 495, + 481, + 590 + ], + "type": "text", + "content": "Our model outperformed previous models on all four metrics across " + }, + { + "bbox": [ + 130, + 495, + 481, + 590 + ], + "type": "inline_equation", + "content": "10\\%" + }, + { + "bbox": [ + 130, + 495, + 481, + 590 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 130, + 495, + 481, + 590 + ], + "type": "inline_equation", + "content": "20\\%" + }, + { + "bbox": [ + 130, + 495, + 481, + 590 + ], + "type": "text", + "content": " labeled data. At " + }, + { + "bbox": [ + 130, + 495, + 481, + 590 + ], + "type": "inline_equation", + "content": "10\\%" + }, + { + "bbox": [ + 130, + 495, + 481, + 590 + ], + "type": "text", + "content": " of the data, compared to the best results from previous work, our PMT showed a " + }, + { + "bbox": [ + 130, + 495, + 481, + 590 + ], + "type": "inline_equation", + "content": "7.17\\%" + }, + { + "bbox": [ + 130, + 495, + 481, + 590 + ], + "type": "text", + "content": " improvement in Dice, a " + }, + { + "bbox": [ + 130, + 495, + 481, + 590 + ], + "type": "inline_equation", + "content": "9.09\\%" + }, + { + "bbox": [ + 130, + 495, + 481, + 590 + ], + "type": "text", + "content": " improvement in Jaccard, a reduction of 6.35 in 95HD, and a reduction of 2.10 in ASD. Our PMT also showed a " + }, + { + "bbox": [ + 130, + 495, + 481, + 590 + ], + "type": "inline_equation", + "content": "0.31\\%" + }, + { + "bbox": [ + 130, + 495, + 481, + 590 + ], + "type": "text", + "content": " improvement in Dice, a " + }, + { + "bbox": [ + 130, + 495, + 481, + 590 + ], + "type": "inline_equation", + "content": "0.55\\%" + }, + { + "bbox": [ + 130, + 495, + 481, + 590 + ], + "type": "text", + "content": " improvement in Jaccard, and a reduction of 0.36 in ASD at " + }, + { + "bbox": [ + 130, + 495, + 481, + 590 + ], + "type": "inline_equation", + "content": "20\\%" + }, + { + "bbox": [ + 130, + 495, + 481, + 590 + ], + "type": "text", + "content": " of the data. The above indicates that our model achieves superior results compared to existing methods on differently proportioned annotated data in the task of segmenting the pancreas." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 132, + 613, + 219, + 624 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 613, + 219, + 624 + ], + "spans": [ + { + "bbox": [ + 132, + 613, + 219, + 624 + ], + "type": "text", + "content": "5 Conclusion" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 130, + 641, + 480, + 665 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 641, + 480, + 665 + ], + "spans": [ + { + "bbox": [ + 130, + 641, + 480, + 665 + ], + "type": "text", + "content": "In this paper, we propose a semi-supervised medical image segmentation framework named Progress Mean Teacher (PMT). PMT adopts a progressive" + } + ] + } + ], + "index": 7 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 340, + 91, + 447, + 101 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 340, + 91, + 447, + 101 + ], + "spans": [ + { + "bbox": [ + 340, + 91, + 447, + 101 + ], + "type": "text", + "content": "Progressive Mean Teacher" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 470, + 91, + 480, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 470, + 91, + 480, + 100 + ], + "spans": [ + { + "bbox": [ + 470, + 91, + 480, + 100 + ], + "type": "text", + "content": "13" + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 12 + }, + { + "para_blocks": [ + { + "type": "table", + "bbox": [ + 145, + 136, + 466, + 335 + ], + "blocks": [ + { + "bbox": [ + 133, + 114, + 481, + 125 + ], + "lines": [ + { + "bbox": [ + 133, + 114, + 481, + 125 + ], + "spans": [ + { + "bbox": [ + 133, + 114, + 481, + 125 + ], + "type": "text", + "content": "Table 4: Comparison results on Pancreas-NIH dataset with " + }, + { + "bbox": [ + 133, + 114, + 481, + 125 + ], + "type": "inline_equation", + "content": "{10}\\%" + }, + { + "bbox": [ + 133, + 114, + 481, + 125 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 133, + 114, + 481, + 125 + ], + "type": "inline_equation", + "content": "{20}\\%" + }, + { + "bbox": [ + 133, + 114, + 481, + 125 + ], + "type": "text", + "content": " labeled data" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 145, + 136, + 466, + 335 + ], + "lines": [ + { + "bbox": [ + 145, + 136, + 466, + 335 + ], + "spans": [ + { + "bbox": [ + 145, + 136, + 466, + 335 + ], + "type": "table", + "html": "
MethodLabeledMetrics
Dice↑Jaccard↑95HD↓ASD↓
VNet6(10%)55.6041.7445.3318.63
UA-MT6(10%)66.3453.2117.214.57
SASSNet6(10%)68.7853.8619.026.26
DTC6(10%)69.2154.0617.215.95
BCPSOTA6(10%)73.8359.2412.713.72
MCFSOTA6(10%)----
PMT(Ours)6(10%)81.0068.336.361.62
VNet12(20%)72.3858.2619.355.89
UA-MT12(20%)76.1062.6210.842.43
SASSNet12(20%)77.6664.0810.933.05
DTC12(20%)78.2764.758.362.25
BCPSOTA12(20%)82.9170.976.432.25
MCFSOTA12(20%)75.0061.2711.593.27
PMT(Ours)12(20%)83.2271.527.601.89
", + "image_path": "387f66b0ca3ff35902e2ef3becfb5089c9f1576e246792d4326b458a2fe7526a.jpg" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "table_body" + } + ], + "index": 3 + }, + { + "bbox": [ + 133, + 387, + 481, + 578 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 387, + 481, + 578 + ], + "spans": [ + { + "bbox": [ + 133, + 387, + 481, + 578 + ], + "type": "text", + "content": "design training process, establishing temporal-level model alignment and pseudolabel filtering while leveraging a network architecture based on MT. The core idea of this model framework is to generate diverse pseudo-labels consistently during the backpropagation process by establishing representation capability differences caused by iteration gaps, thereby stabilizing and enhancing the model's representation capability. In addition to the progressive architecture, PMT employs two simple yet effective methods, Pseudo Label Filtering (PLF) and Discrepancy Driven Alignment (DDA). PLF utilizes the representation capability of the model to filter pseudo-labels, discarding low-fidelity ones detrimental to constraining, while DDA aligns the differences in model predictions, allowing lagging models to catch up with leading models rapidly. Results from ablation experiments demonstrate that each component of the PMT framework significantly enhances the model's performance. In comparative experiments with other methods, PMT achieves state-of-the-art performance in terms of accuracy, surpassing previous methods significantly, and maintains this advantage compared to other methods in situations with more limited data and more challenging tasks." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 133, + 593, + 481, + 665 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 593, + 481, + 665 + ], + "spans": [ + { + "bbox": [ + 133, + 593, + 481, + 665 + ], + "type": "text", + "content": "Limitation and Future Work. Despite the outstanding performance of PMT, there is still room for further exploration of semi-supervised training architectures established through temporal consistency. Investigating more advanced and stable strategies under a progressive design paradigm, and assessing whether they can further enhance performance, are topics worthy of further research in the future." + } + ] + } + ], + "index": 5 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "spans": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "type": "text", + "content": "14" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 166, + 91, + 220, + 100 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 166, + 91, + 220, + 100 + ], + "spans": [ + { + "bbox": [ + 166, + 91, + 220, + 100 + ], + "type": "text", + "content": "N. Gao et al." + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 13 + }, + { + "para_blocks": [ + { + "bbox": [ + 133, + 114, + 240, + 129 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 114, + 240, + 129 + ], + "spans": [ + { + "bbox": [ + 133, + 114, + 240, + 129 + ], + "type": "text", + "content": "Acknowledgments" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 130, + 137, + 482, + 185 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 137, + 482, + 185 + ], + "spans": [ + { + "bbox": [ + 130, + 137, + 482, + 185 + ], + "type": "text", + "content": "This work was supported in part by National Natural Science Foundation of China under Grants 62088102, 12326608 and 62106192, Natural Science Foundation of Shaanxi Province under Grant 2022JC-41, and Fundamental Research Funds for the Central Universities under Grant XTR042021005." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 133, + 201, + 197, + 213 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 201, + 197, + 213 + ], + "spans": [ + { + "bbox": [ + 133, + 201, + 197, + 213 + ], + "type": "text", + "content": "References" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 134, + 223, + 480, + 665 + ], + "type": "list", + "angle": 0, + "index": 18, + "blocks": [ + { + "bbox": [ + 138, + 223, + 480, + 255 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 223, + 480, + 255 + ], + "spans": [ + { + "bbox": [ + 138, + 223, + 480, + 255 + ], + "type": "text", + "content": "1. Achanta, R., Shaji, A., Smith, K., Lucchi, A., Fua, P., Süsstrunk, S.: Slic superpixels compared to state-of-the-art superpixel methods. IEEE transactions on pattern analysis and machine intelligence 34(11), 2274-2282 (2012) 4" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 138, + 256, + 480, + 298 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 256, + 480, + 298 + ], + "spans": [ + { + "bbox": [ + 138, + 256, + 480, + 298 + ], + "type": "text", + "content": "2. Bai, Y., Chen, D., Li, Q., Shen, W., Wang, Y.: Bidirectional copy-paste for semi-supervised medical image segmentation. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 11514-11524 (2023) 10, 11" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 138, + 299, + 480, + 331 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 299, + 480, + 331 + ], + "spans": [ + { + "bbox": [ + 138, + 299, + 480, + 331 + ], + "type": "text", + "content": "3. Cai, H., Li, S., Qi, L., Yu, Q., Shi, Y., Gao, Y.: Orthogonal annotation benefits barely-supervised medical image segmentation. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 3302-3311 (2023) 4" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 138, + 332, + 480, + 363 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 332, + 480, + 363 + ], + "spans": [ + { + "bbox": [ + 138, + 332, + 480, + 363 + ], + "type": "text", + "content": "4. Chen, X., Yuan, Y., Zeng, G., Wang, J.: Semi-supervised semantic segmentation with cross pseudo supervision. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 2613-2622 (2021) 5" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 138, + 365, + 480, + 395 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 365, + 480, + 395 + ], + "spans": [ + { + "bbox": [ + 138, + 365, + 480, + 395 + ], + "type": "text", + "content": "5. Chen, X., He, K.: Exploring simple siamese representation learning. In: Proceedings of the IEEE/CVF conference on computer vision and pattern recognition. pp. 15750-15758 (2021) 2" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 138, + 396, + 480, + 427 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 396, + 480, + 427 + ], + "spans": [ + { + "bbox": [ + 138, + 396, + 480, + 427 + ], + "type": "text", + "content": "6. Dong-DongChen, W., WeiGao, Z.: Tri-net for semi-supervised deep learning. In: Proceedings of twenty-seventh international joint conference on artificial intelligence. pp. 2014-2020 (2018) 4" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 138, + 429, + 480, + 460 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 429, + 480, + 460 + ], + "spans": [ + { + "bbox": [ + 138, + 429, + 480, + 460 + ], + "type": "text", + "content": "7. Jiao, R., Zhang, Y., Ding, L., Cai, R., Zhang, J.: Learning with limited annotations: a survey on deep semi-supervised learning for medical image segmentation. arXiv preprint arXiv:2207.14191 (2022) 4" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 138, + 460, + 480, + 503 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 460, + 480, + 503 + ], + "spans": [ + { + "bbox": [ + 138, + 460, + 480, + 503 + ], + "type": "text", + "content": "8. Ke, Z., Qiu, D., Li, K., Yan, Q., Lau, R.W.: Guided collaborative training for pixelwise semi-supervised learning. In: Computer Vision-ECCV 2020: 16th European Conference, Glasgow, UK, August 23-28, 2020, Proceedings, Part XIII 16. pp. 429-445. Springer (2020) 5" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 138, + 504, + 480, + 525 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 504, + 480, + 525 + ], + "spans": [ + { + "bbox": [ + 138, + 504, + 480, + 525 + ], + "type": "text", + "content": "9. Laine, S., Aila, T.: Temporal ensembling for semi-supervised learning. arXiv preprint arXiv:1610.02242 (2016) 2, 4" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 134, + 525, + 480, + 556 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 134, + 525, + 480, + 556 + ], + "spans": [ + { + "bbox": [ + 134, + 525, + 480, + 556 + ], + "type": "text", + "content": "10. Lee, D.H., et al.: Pseudo-label: The simple and efficient semi-supervised learning method for deep neural networks. In: Workshop on challenges in representation learning, ICML. vol. 3, p. 896. Atlanta (2013) 2, 4" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 134, + 557, + 480, + 600 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 134, + 557, + 480, + 600 + ], + "spans": [ + { + "bbox": [ + 134, + 557, + 480, + 600 + ], + "type": "text", + "content": "11. Li, S., Zhang, C., He, X.: Shape-aware semi-supervised 3d semantic segmentation for medical images. In: Medical Image Computing and Computer Assisted Intervention-MICCAI 2020: 23rd International Conference, Lima, Peru, October 4-8, 2020, Proceedings, Part I 23. pp. 552-561. Springer (2020) 4, 9, 10, 11" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 134, + 601, + 480, + 632 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 134, + 601, + 480, + 632 + ], + "spans": [ + { + "bbox": [ + 134, + 601, + 480, + 632 + ], + "type": "text", + "content": "12. Luo, X., Chen, J., Song, T., Wang, G.: Semi-supervised medical image segmentation through dual-task consistency. In: Proceedings of the AAAI conference on artificial intelligence. vol. 35, pp. 8801-8809 (2021) 4, 9, 10, 11" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 134, + 633, + 480, + 665 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 134, + 633, + 480, + 665 + ], + "spans": [ + { + "bbox": [ + 134, + 633, + 480, + 665 + ], + "type": "text", + "content": "13. Miletari, F., Navab, N., Ahmadi, S.A.: V-net: Fully convolutional neural networks for volumetric medical image segmentation. In: 2016 fourth international conference on 3D vision (3DV). pp. 565-571. IEEE (2016) 9" + } + ] + } + ], + "index": 17 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 340, + 91, + 447, + 102 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 340, + 91, + 447, + 102 + ], + "spans": [ + { + "bbox": [ + 340, + 91, + 447, + 102 + ], + "type": "text", + "content": "Progressive Mean Teacher" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 470, + 92, + 480, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 470, + 92, + 480, + 100 + ], + "spans": [ + { + "bbox": [ + 470, + 92, + 480, + 100 + ], + "type": "text", + "content": "15" + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 14 + }, + { + "para_blocks": [ + { + "bbox": [ + 132, + 116, + 481, + 666 + ], + "type": "list", + "angle": 0, + "index": 17, + "blocks": [ + { + "bbox": [ + 133, + 116, + 481, + 149 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 116, + 481, + 149 + ], + "spans": [ + { + "bbox": [ + 133, + 116, + 481, + 149 + ], + "type": "text", + "content": "14. Na, J., Ha, J.W., Chang, H.J., Han, D., Hwang, W.: Switching temporary teachers for semi-supervised semantic segmentation. Advances in Neural Information Processing Systems 36 (2024) 5" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 133, + 150, + 481, + 183 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 150, + 481, + 183 + ], + "spans": [ + { + "bbox": [ + 133, + 150, + 481, + 183 + ], + "type": "text", + "content": "15. Peiris, H., Hayat, M., Chen, Z., Egan, G., Harandi, M.: Uncertainty-guided dual-views for semi-supervised volumetric medical image segmentation. Nature Machine Intelligence 5(7), 724-738 (2023) 4" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 132, + 183, + 481, + 205 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 183, + 481, + 205 + ], + "spans": [ + { + "bbox": [ + 132, + 183, + 481, + 205 + ], + "type": "text", + "content": "16. Rizve, M.N., Kardan, N., Shah, M.: Towards realistic semi-supervised learning. In: European Conference on Computer Vision. pp. 437-455. Springer (2022) 1" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 132, + 205, + 481, + 259 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 205, + 481, + 259 + ], + "spans": [ + { + "bbox": [ + 132, + 205, + 481, + 259 + ], + "type": "text", + "content": "17. Roth, H.R., Lu, L., Farag, A., Shin, H.C., Liu, J., Turkbey, E.B., Summers, R.M.: Deeporgan: Multi-level deep convolutional networks for automated pancreas segmentation. In: Medical Image Computing and Computer-Assisted Intervention-MICCAI 2015: 18th International Conference, Munich, Germany, October 5-9, 2015, Proceedings, Part I 18. pp. 556-564. Springer (2015) 3, 10" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 132, + 259, + 481, + 293 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 259, + 481, + 293 + ], + "spans": [ + { + "bbox": [ + 132, + 259, + 481, + 293 + ], + "type": "text", + "content": "18. Shen, Z., Cao, P., Yang, H., Liu, X., Yang, J., Zaiane, O.R.: Co-training with high-confidence pseudo labels for semi-supervised medical image segmentation. arXiv preprint arXiv:2301.04465 (2023) 5" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 132, + 293, + 481, + 326 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 293, + 481, + 326 + ], + "spans": [ + { + "bbox": [ + 132, + 293, + 481, + 326 + ], + "type": "text", + "content": "19. Tarvainen, A., Valpola, H.: Mean teachers are better role models: Weight-averaged consistency targets improve semi-supervised deep learning results. Advances in neural information processing systems 30 (2017) 2, 4, 5" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 132, + 326, + 481, + 369 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 326, + 481, + 369 + ], + "spans": [ + { + "bbox": [ + 132, + 326, + 481, + 369 + ], + "type": "text", + "content": "20. Wang, H., Li, X.: Dhc: Dual-debiased heterogeneous co-training framework for class-imbalanced semi-supervised medical image segmentation. In: International Conference on Medical Image Computing and Computer-Assisted Intervention. pp. 582-591. Springer (2023) 2" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 132, + 369, + 481, + 413 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 369, + 481, + 413 + ], + "spans": [ + { + "bbox": [ + 132, + 369, + 481, + 413 + ], + "type": "text", + "content": "21. Wang, J., Zhou, S., Fang, C., Wang, L., Wang, J.: Meta corrupted pixels mining for medical image segmentation. In: Medical Image Computing and Computer Assisted Intervention-MICCAI 2020: 23rd International Conference, Lima, Peru, October 4-8, 2020, Proceedings, Part I 23. pp. 335-345. Springer (2020) 2" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 132, + 413, + 481, + 456 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 413, + 481, + 456 + ], + "spans": [ + { + "bbox": [ + 132, + 413, + 481, + 456 + ], + "type": "text", + "content": "22. Wang, X., Yuan, Y., Guo, D., Huang, X., Cui, Y., Xia, M., Wang, Z., Bai, C., Chen, S.: Ssa-net: Spatial self-attention network for COVID-19 pneumonia infection segmentation with semi-supervised few-shot learning. Medical Image Analysis 79, 102459 (2022) 4" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 132, + 457, + 481, + 500 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 457, + 481, + 500 + ], + "spans": [ + { + "bbox": [ + 132, + 457, + 481, + 500 + ], + "type": "text", + "content": "23. Wang, Y., Xiao, B., Bi, X., Li, W., Gao, X.: Mcf: Mutual correction framework for semi-supervised medical image segmentation. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 15651-15660 (2023) 2, 3, 4, 9, 10, 11" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 132, + 501, + 481, + 544 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 501, + 481, + 544 + ], + "spans": [ + { + "bbox": [ + 132, + 501, + 481, + 544 + ], + "type": "text", + "content": "24. Wu, H., Wang, Z., Song, Y., Yang, L., Qin, J.: Cross-patch dense contrastive learning for semi-supervised segmentation of cellular nuclei in histopathologic images. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 11666-11675 (2022) 2" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 132, + 545, + 481, + 578 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 545, + 481, + 578 + ], + "spans": [ + { + "bbox": [ + 132, + 545, + 481, + 578 + ], + "type": "text", + "content": "25. Xia, K., Wang, L., Zhou, S., Hua, G., Tang, W.: Learning from noisy pseudo labels for semi-supervised temporal action localization. In: Proceedings of the IEEE/CVF International Conference on Computer Vision. pp. 10160-10169 (2023) 1" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 132, + 578, + 481, + 610 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 578, + 481, + 610 + ], + "spans": [ + { + "bbox": [ + 132, + 578, + 481, + 610 + ], + "type": "text", + "content": "26. Xie, Q., Luong, M.T., Hovy, E., Le, Q.V.: Self-training with noisy student improves imagenet classification. In: Proceedings of the IEEE/CVF conference on computer vision and pattern recognition. pp. 10687-10698 (2020) 2" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 132, + 611, + 481, + 643 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 611, + 481, + 643 + ], + "spans": [ + { + "bbox": [ + 132, + 611, + 481, + 643 + ], + "type": "text", + "content": "27. Xin, X., Wang, J., Xie, R., Zhou, S., Huang, W., Zheng, N.: Semi-supervised person re-identification using multi-view clustering. Pattern Recognition 88, 285–297 (2019) 1" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 132, + 643, + 481, + 666 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 643, + 481, + 666 + ], + "spans": [ + { + "bbox": [ + 132, + 643, + 481, + 666 + ], + "type": "text", + "content": "28. Xiong, Z., Xia, Q., Hu, Z., Huang, N., Bian, C., Zheng, Y., Vesal, S., Ravikumar, N., Maier, A., Yang, X., et al.: A global benchmark of algorithms for segmenting" + } + ] + } + ], + "index": 16 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "spans": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "type": "text", + "content": "16" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 166, + 91, + 220, + 100 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 166, + 91, + 220, + 100 + ], + "spans": [ + { + "bbox": [ + 166, + 91, + 220, + 100 + ], + "type": "text", + "content": "N. Gao et al." + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 15 + }, + { + "para_blocks": [ + { + "bbox": [ + 132, + 116, + 481, + 348 + ], + "type": "list", + "angle": 0, + "index": 8, + "blocks": [ + { + "bbox": [ + 149, + 116, + 480, + 138 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 149, + 116, + 480, + 138 + ], + "spans": [ + { + "bbox": [ + 149, + 116, + 480, + 138 + ], + "type": "text", + "content": "the left atrium from late gadolinium-enhanced cardiac magnetic resonance imaging. Medical image analysis 67, 101832 (2021) 3, 10" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 133, + 139, + 481, + 182 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 139, + 481, + 182 + ], + "spans": [ + { + "bbox": [ + 133, + 139, + 481, + 182 + ], + "type": "text", + "content": "29. Xu, Z., Wang, Y., Lu, D., Yu, L., Yan, J., Luo, J., Ma, K., Zheng, Y., Tong, R.K.y.: All-around real label supervision: Cyclic prototype consistency learning for semi-supervised medical image segmentation. IEEE Journal of Biomedical and Health Informatics 26(7), 3174-3184 (2022) 4, 10" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 132, + 183, + 481, + 237 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 183, + 481, + 237 + ], + "spans": [ + { + "bbox": [ + 132, + 183, + 481, + 237 + ], + "type": "text", + "content": "30. Yu, L., Wang, S., Li, X., Fu, C.W., Heng, P.A.: Uncertainty-aware self-ensembling model for semi-supervised 3d left atrium segmentation. In: Medical Image Computing and Computer Assisted Intervention-MICCAI 2019: 22nd International Conference, Shenzhen, China, October 13–17, 2019, Proceedings, Part II 22. pp. 605–613. Springer (2019) 4, 9, 10, 11" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 132, + 237, + 481, + 270 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 237, + 481, + 270 + ], + "spans": [ + { + "bbox": [ + 132, + 237, + 481, + 270 + ], + "type": "text", + "content": "31. Yun, S., Han, D., Oh, S.J., Chun, S., Choe, J., Yoo, Y.: Cutmix: Regularization strategy to train strong classifiers with localizable features. In: Proceedings of the IEEE/CVF international conference on computer vision. pp. 6023-6032 (2019) 11" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 132, + 270, + 481, + 312 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 270, + 481, + 312 + ], + "spans": [ + { + "bbox": [ + 132, + 270, + 481, + 312 + ], + "type": "text", + "content": "32. Zheng, M., You, S., Huang, L., Wang, F., Qian, C., Xu, C.: Simmatch: Semi-supervised learning with similarity matching. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 14471-14481 (2022) 1" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 132, + 314, + 481, + 348 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 314, + 481, + 348 + ], + "spans": [ + { + "bbox": [ + 132, + 314, + 481, + 348 + ], + "type": "text", + "content": "33. Zhou, S., Wang, J., Shu, J., Meng, D., Wang, L., Zheng, N.: Multinetwork collaborative feature learning for semisupervised person reidentification. IEEE Transactions on Neural Networks and Learning Systems 33(9), 4826-4839 (2021) 1" + } + ] + } + ], + "index": 7 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 340, + 91, + 447, + 102 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 340, + 91, + 447, + 102 + ], + "spans": [ + { + "bbox": [ + 340, + 91, + 447, + 102 + ], + "type": "text", + "content": "Progressive Mean Teacher" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 470, + 91, + 481, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 470, + 91, + 481, + 100 + ], + "spans": [ + { + "bbox": [ + 470, + 91, + 481, + 100 + ], + "type": "text", + "content": "17" + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 16 + } + ], + "_backend": "vlm", + "_version_name": "2.6.4" +} \ No newline at end of file diff --git a/2024/POA_ Pre-training Once for Models of All Sizes/9b5de88d-0728-41f9-a801-73fd67e5b8a0_content_list.json b/2024/POA_ Pre-training Once for Models of All Sizes/9b5de88d-0728-41f9-a801-73fd67e5b8a0_content_list.json new file mode 100644 index 0000000000000000000000000000000000000000..eab01a775cbf15db0db4951f92e77ba8c11cda8b --- /dev/null +++ b/2024/POA_ Pre-training Once for Models of All Sizes/9b5de88d-0728-41f9-a801-73fd67e5b8a0_content_list.json @@ -0,0 +1,1589 @@ +[ + { + "type": "text", + "text": "POA: Pre-training Once for Models of All Sizes", + "text_level": 1, + "bbox": [ + 228, + 141, + 774, + 162 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Yingying Zhang, Xin Guo, Jiangwei Lao, Lei Yu, Lixiang Ru, Jian Wang, Guo Ye, Huimei He, Jingdong Chen, and Ming Yang", + "bbox": [ + 230, + 189, + 771, + 220 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Ant Group.", + "bbox": [ + 460, + 233, + 539, + 246 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "qichu.zyy@antgroup.com", + "bbox": [ + 416, + 247, + 589, + 260 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Abstract. Large-scale self-supervised pre-training has paved the way for one foundation model to handle many different vision tasks. Most pre-training methodologies train a single model of a certain size at one time. Nevertheless, various computation or storage constraints in real-world scenarios require substantial efforts to develop a series of models with different sizes to deploy. Thus, in this study, we propose a novel tri-branch self-supervised training framework, termed as POA (Pre-training Once for All), to tackle this aforementioned issue. Our approach introduces an innovative elastic student branch into a modern self-distillation paradigm. At each pre-training step, we randomly sample a sub-network from the original student to form the elastic student and train all branches in a self-distilling fashion. Once pre-trained, POA allows the extraction of pre-trained models of diverse sizes for downstream tasks. Remarkably, the elastic student facilitates the simultaneous pre-training of multiple models with different sizes, which also acts as an additional ensemble of models of various sizes to enhance representation learning. Extensive experiments, including k-nearest neighbors, linear probing evaluation and assessments on multiple downstream tasks demonstrate the effectiveness and advantages of our POA. It achieves state-of-the-art performance using ViT, Swin Transformer and ResNet backbones, producing around a hundred models with different sizes through a single pre-training session. The code is available at: https://github.com/Qichuzyy/POA.", + "bbox": [ + 261, + 297, + 738, + 602 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Keywords: Self-supervised Learning $\\cdot$ Pre-training Once for All", + "bbox": [ + 261, + 616, + 694, + 630 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "1 Introduction", + "text_level": 1, + "bbox": [ + 217, + 657, + 374, + 672 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Learning generalizable visual representations in a large model by self-supervised learning has delivered superior performance across a wide range of visual tasks [10, 13, 20, 46, 49, 53] in recent years. Nevertheless, when deployed to real-world applications, a large model has to be adapted to diverse resource limitations in terms of computation, storage, or power consumption, etc. For example, a well-engineered AI product typically comprises a suite of models tailored for varying scenarios, such as Gemini Nano, Pro and Ultra [37]. Given a large pre-trained model, common solutions to deploy it to multiple application scenarios with different resource constraints include additional weight pruning [28, 42, 51], knowledge distillation [22, 24], or even re-training a small network from scratch, which", + "bbox": [ + 212, + 688, + 787, + 840 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "all require substantial development efforts. Consequently, this issue prompts a critical question: is it possible to pre-train once to produce multiple models with different sizes simultaneously, each delivering sufficiently good representations?", + "bbox": [ + 212, + 146, + 787, + 194 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "To address this challenge, we introduce a new paradigm of self-supervised learning, called POA (Pre-training Once for All). POA is built upon the prevalent teacher-student self-distillation framework [5,33,55], with an additional innovative elastic student branch. The elastic student branch embeds a series of sub-networks through parameter sharing, upon the observation that smaller-sized models are sub-networks of a larger-sized one for modern network structures [14,21,30]. Moreover, the parameters of this branch are shared with the original, or intact student. At each pre-training step, we randomly sample a subset of parameters from the intact student to form the corresponding elastic student. Both the original intact student and the elastic student are trained to emulate the output of the teacher network. The teacher itself is continually refined via an exponential moving average (EMA) of the student's parameters, including the sampled elastic student. The elastic student facilitates effective and efficient pre-training on different subsets of parameters, leading to the successful extraction of high-performance sub-networks from the pre-trained teacher for subsequent downstream scenarios. It also acts as a form of training regularization by enforcing the outputs to match between the teacher and various sub-networks, contributing to a stable training process. Additionally, by serving as an ensemble of different sub-networks across different pre-training steps, the elastic student improves the representation learning [44].", + "bbox": [ + 212, + 205, + 789, + 510 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "To the best of our knowledge, POA represents the first self-supervised learning method capable of training multiple-sized models concurrently, each obtaining high-quality representations for different resource constraints without further pre-training. Figure 1 displays the k-nearest neighbors (k-NN) evaluation results for 143 sub-networks extracted from a ViT-L model [14] pre-trained by POA. By choosing different elastic widths and depths, which will be explained in Sec. 3.1, the pre-trained teacher model can generate a sufficient number of candidate sub-networks for the selection of the suitable model tailored for downstream applications according to the", + "bbox": [ + 212, + 522, + 495, + 794 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "computational resources available. Notably, each sub-network is well-trained and exhibits superior performance, thanks to our meticulous design on same-view distillation, as validated in Sec. 4.4. In particular, the ViT-S, ViT-B and ViT-", + "bbox": [ + 212, + 795, + 787, + 840 + ], + "page_idx": 1 + }, + { + "type": "image", + "img_path": "images/58642ae34be5e2e6be6ff4bb4893376192d54c89b6fa430015f321e75fff488a.jpg", + "image_caption": [ + "Fig.1: The k-NN evaluation accuracy of 143 elastic ViTs derived from the ViT-L/16 teacher model pre-trained with POA." + ], + "image_footnote": [], + "bbox": [ + 504, + 527, + 790, + 728 + ], + "page_idx": 1 + }, + { + "type": "page_number", + "text": "2", + "bbox": [ + 217, + 114, + 228, + 126 + ], + "page_idx": 1 + }, + { + "type": "header", + "text": "Y. Zhang et al.", + "bbox": [ + 271, + 114, + 374, + 128 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "L models set new benchmarks, achieving the state-of-the-art (SOTA) results compared with those pre-trained by existing methods [33, 34, 55].", + "bbox": [ + 212, + 146, + 782, + 176 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "To rigorously evaluate the efficacy of our approach, we conduct extensive experiments using three widely-used backbone architectures, i.e., ViT [14], Swin Transformer [30], and ResNet [21]. Each backbone is pre-trained on ImageNet-1K dataset and assessed using k-NN and linear probing classification evaluation, as well as downstream dense prediction tasks such as object detection and semantic segmentation. Our method achieves SOTA accuracy across various model sizes with a single pre-training session. The technical contributions of this work are summarized as follows:", + "bbox": [ + 212, + 176, + 787, + 297 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "- To the best of our knowledge, POA is the first pre-training paradigm that integrates unsupervised representation learning and once-for-all model generation within a single pre-training session. It tackles the Pre-training-Once-for-All challenge which has been seldomly explored by the community but is of great practical significance for real-world deployment that usually requires a suite of models.", + "bbox": [ + 225, + 306, + 782, + 393 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "- We propose a novel and elegant component called Elastic Student, featuring a range of elastic operators that enable POA to be compatible with popular backbone structures including ViT, Swin Transformer and ResNet. It provides the capability to generate models of diverse sizes. Furthermore, it serves as a model ensemble to smooth training process and improve learned representations.", + "bbox": [ + 225, + 398, + 782, + 484 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "- Through thorough assessments using k-NN, linear probing and downstream dense task evaluation, our approach exhibits superior performance over existing state-of-the-art pre-training methods across multiple metrics. Moreover, we compare our POA with Self-Supervised Distillation (SEED) [15], a knowledge distillation method designed especially for self-supervised learning, further validating POA's effectiveness.", + "bbox": [ + 225, + 488, + 782, + 575 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "2 Related Work", + "text_level": 1, + "bbox": [ + 215, + 601, + 387, + 617 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "2.1 Self-supervised Learning", + "text_level": 1, + "bbox": [ + 214, + 633, + 465, + 648 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Self-supervised learning (SSL) frameworks commonly fall into two categories, generative and discriminative SSL. Most generative SSL approaches [7,9,18,23, 25,26,38,48,52] focus on learning image representations directly in pixel space. On the other hand, discriminative SSL methods [4,6,19,27,36,41,43,47] aim to learn representations by pulling those of different views of the same image closer while separating the representations of views from different images.", + "bbox": [ + 212, + 657, + 782, + 747 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Contrastive learning (CL) with the InfoNCE loss [32] has emerged as a popular approach for discriminative SSL methods, attracting significant research interests in recent years. Although CL methods prevent the collapse of network representations through the use of negative samples, they still suffer from the dimensional collapse, where representations tend to collapse into a low-dimensional manifold. Grill et al. introduced a distillation-based asymmetric", + "bbox": [ + 212, + 750, + 784, + 838 + ], + "page_idx": 2 + }, + { + "type": "header", + "text": "POA", + "bbox": [ + 694, + 114, + 730, + 126 + ], + "page_idx": 2 + }, + { + "type": "page_number", + "text": "3", + "bbox": [ + 774, + 114, + 785, + 126 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "framework known as BYOL [17], which circumvents collapse without self-labeling or contrastive loss relying on negative samples. Following this work, distillation-based frameworks have become a prevailing trend in self-supervised learning. These frameworks [8, 16, 47] often merge with others to enhance overall performance. DINO [5] presented a simple self-distillation framework and has demonstrated impressive results in ViT pre-training. Subsequent works [33, 55] further improved the pre-training performance via masked token or the novel clustering design. Given the substantial benefits of distillation-based methods over other SSL techniques, we have developed our POA SSL framework upon these successful methodologies.", + "bbox": [ + 212, + 146, + 787, + 297 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "2.2 Dynamic Architecture.", + "text_level": 1, + "bbox": [ + 215, + 320, + 452, + 335 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Recently, Chen et al. proposed AutoFormer [45], which trained a supernet to support the effective search of optimal sub-network under some specific parameter number constraints. On the basis of [45], MaskTAS [50] introduced a self-supervised transformer architecture search method. Cai et al. [2] trained a network that accommodates various architectural configurations to reduce the training expense. Their methodologies enable the extraction of a specialized sub-network from the trained supernet. The design of the elastic student in our POA SSL is inspired by the weight-sharing strategy employed in these neural architecture search (NAS) methods. However, our implementation differs significantly due to the distinct purpose from NAS. Specifically, NAS aims to discover the optimal architecture within certain parameter constraints, which typically involves a huge number of sub-networks (more than $10^{16}$ ) in the search space. Given the limitations on the number of training iterations and the network parameter capacity, it is challenging to ensure high performance across all sub-networks. After training the supernet, NAS requires a subsequent phase of searching and re-training to finalize the output model. In contrast, our POA SSL defines a sufficiently large yet compact set of sub-networks with elastic depths and widths for the purpose of pre-training models of various sizes via a single training session. Additional design on the same-view distillation guarantees that all elastic sub-networks within our framework are adequately and efficiently trained. As a result, our POA can readily extract a range of sub-networks from the teacher model without the need for extra pre-training, facilitating an easy selection of a suitable sub-network for different computational contexts. The design of elastic student is somewhat akin to the supervised learning method, Cosub [40]. The main difference is that Cosub simply skips blocks, making only the depth elastic.", + "bbox": [ + 212, + 345, + 787, + 724 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "3 POA Self-supervised Learning Framework", + "text_level": 1, + "bbox": [ + 215, + 747, + 660, + 765 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Our primary goal is to pre-train models of multiple sizes via a single self-supervised pre-training session. To this end, we propose a novel SSL framework named POA, inspired by the latest progress in self-distillation techniques [5, 33, 55]. The architecture of POA is illustrated in Figure 2, encompassing a", + "bbox": [ + 212, + 779, + 787, + 840 + ], + "page_idx": 3 + }, + { + "type": "page_number", + "text": "4", + "bbox": [ + 217, + 114, + 228, + 126 + ], + "page_idx": 3 + }, + { + "type": "header", + "text": "Y. Zhang et al.", + "bbox": [ + 271, + 114, + 372, + 128 + ], + "page_idx": 3 + }, + { + "type": "image", + "img_path": "images/f899a486424e9769904603e459e5736d7358d72eec370a022540471bbfad631c.jpg", + "image_caption": [ + "Fig. 2: Overview of the POA SSL: Given an image $x$ , two augmented views $x_{a}$ and $x_{b}$ are generated. These views are input into three branches: a teacher, an intact student, and an elastic student, the latter being derived from the intact student. POA optimizes distillation losses in a twofold manner: the intact and the elastic students are distilled from the teacher using the cross-view data respectively, and additionally, the elastic student is distilled from the intact student using the same-view data." + ], + "image_footnote": [], + "bbox": [ + 220, + 143, + 782, + 295 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "teacher, an intact student, an elastic student and their corresponding heads. The teacher is updated with an EMA of the students. The elastic student is a derivative of the intact one, with both the backbones and heads' parameters shared. We leverage distillation in two aspects: both the intact and elastic students are distilled from the teacher using different views of the same image, while the elastic student additionally learns from the intact student using the identical views. The cross-view distillation works as a form of representation learning, as introduced in [5, 33, 55]. Notably, in addition to the regular EMA update with only the intact student as [5, 33, 55], the elastic student provides a randomly-sampled sub-network at each pre-training step to participate in the teacher's EMA refinement. This procedure actually simulates an ensemble of multiple sub-networks, which is also proven to be beneficial in the realm of supervised learning [44]. The same-view distillation is a standard knowledge distillation between the intact and elastic students, promoting the quality of the elastic one. Experiments in Sec. 4.4 validate our design in details.", + "bbox": [ + 212, + 396, + 787, + 623 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "3.1 Design of Elastic Student", + "text_level": 1, + "bbox": [ + 215, + 646, + 472, + 662 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "The elastic student is a sub-network whose parameters are extracted from the intact student. In the context of a transformer backbone, the width refers to the dimensionality of the tokens, whereas for a convolutional backbone, the width indicates the number of channels. We denote the depth as the number of basic blocks in either a transformer or a convolutional network. Given the value of the width and depth, it yields a certain network structure. For simplicity, we focus on detailing the elastic design of the ViT in this section. For the similar elastic design of the Swin Transformer and ResNet, please refer to Appendix A.", + "bbox": [ + 212, + 672, + 787, + 792 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "A basic block of ViT mainly consists of a multi-head self-attention (MSA) module and a multi-layer perceptron (MLP) module. Layer Normalization (LN) [1] is applied before each module, with residual connections after each module.", + "bbox": [ + 212, + 794, + 785, + 840 + ], + "page_idx": 4 + }, + { + "type": "header", + "text": "POA", + "bbox": [ + 694, + 114, + 730, + 126 + ], + "page_idx": 4 + }, + { + "type": "page_number", + "text": "5", + "bbox": [ + 774, + 114, + 784, + 126 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "As shown in the left part of Figure 3, an elastic block refers to a stack of elastic MSA, MLP, and LN after adjusting the width in the original basic block in ViT. In our approach, the elastic student branch is constructed by assembling a specific number of these elastic blocks at each training iteration.", + "bbox": [ + 212, + 146, + 787, + 205 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/998f376e1f2880a4ce99e0e8de331a39cf8d97fc432940a0dceff96ec234665c.jpg", + "image_caption": [ + "Fig. 3: Illustration of the elastic MSA in an elastic ViT block. To be concise, we simply exclude the projection layers that correspond to $K$ and $V$ in each head." + ], + "image_footnote": [], + "bbox": [ + 256, + 208, + 743, + 348 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "**Elastic MSA** An original, or intact MSA module consists of three major components, i.e., input projection layers, an operator contains attention & concatenation, and an output projection layer. We define a projection layer as $(w^{*}, b^{*})$ , where $w^{*}$ is the linear transformation weight, $b^{*}$ denotes the corresponding bias, and * indicates the name of the layer. As shown in the right part of Figure 3, given a token dimension of $D_{max} = N_h \\cdot D_h$ , where $N_h$ is the number of attention heads and $D_h$ is the head dimension, an input sequence $z \\in \\mathbb{R}^{T \\times D_{max}}$ with length $T$ is initially projected to form the queries $Q \\in \\mathbb{R}^{T \\times D_h}$ , keys $K \\in \\mathbb{R}^{T \\times D_h}$ , and values $V \\in \\mathbb{R}^{T \\times D_h}$ by the intact MSA. To generate elastic MSA, we define $\\mathrm{M} + 1$ elastic widths, including $D_{max}$ , spaced at intervals of $D_h$ as follows:", + "bbox": [ + 212, + 407, + 787, + 559 + ], + "page_idx": 5 + }, + { + "type": "equation", + "text": "\n$$\nD _ {i} = \\left(N _ {h} - i\\right) \\cdot D _ {h}, \\quad \\forall i \\in \\{0, 1, \\dots , M \\}, \\quad M < N _ {h}. \\tag {1}\n$$\n", + "text_format": "latex", + "bbox": [ + 313, + 571, + 785, + 587 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "For each elastic width $D_{i}$ , the weight $w_{i}^{a1} \\in \\mathbb{R}^{D_{h} \\times D_{i}}$ and bias $b_{i}^{a1} \\in \\mathbb{R}^{D_{h}}$ that generate $Q$ , $K$ , and $V$ of each head are extracted from the corresponding input projection layer ( $w^{a1}$ , $b^{a1}$ ) in the intact MSA, as $w_{i}^{a1} = w^{a1}[:,: D_{i}] \\cdot \\alpha_{i}$ and $b_{i}^{a1} = b^{a1}$ . Here, $\\alpha_{i}$ represents the scaling factor to cope with the reduction of the input dimension, calculated as $\\alpha_{i} = D_{max} / D_{i}$ . As the reduction of width, the number of attention heads in the elastic MAS is naturally reduced to $N_{h} - i$ . Similarly, for the output projection layer ( $w^{a2}$ , $b^{a2}$ ), the weight $w_{i}^{a2} \\in \\mathbb{R}^{D_{i} \\times D_{i}}$ and bias $b_{i}^{a2} \\in \\mathbb{R}^{D_{i}}$ are extracted as:", + "bbox": [ + 212, + 598, + 787, + 720 + ], + "page_idx": 5 + }, + { + "type": "equation", + "text": "\n$$\nw _ {i} ^ {a 2} = w ^ {a 2} [: D _ {i},: D _ {i} ] \\cdot \\alpha_ {i} \\quad b _ {i} ^ {a 2} = b ^ {a 2} [: D _ {i} ]. \\tag {2}\n$$\n", + "text_format": "latex", + "bbox": [ + 343, + 729, + 785, + 750 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "**Elastic MLP** An original, or intact MLP module in the ViT block contains two projection layers. The first layer ( $w^{m1}, b^{m1}$ ) expands the dimension of embedding by a factor of $s$ , which is generally set to 4 in the ViT structure. The second layer ( $w^{m2}, b^{m2}$ ) then projects it back to the original dimension. The parameters for both layers of the elastic MLP are extracted in a manner analogous to that", + "bbox": [ + 212, + 763, + 787, + 840 + ], + "page_idx": 5 + }, + { + "type": "page_number", + "text": "6", + "bbox": [ + 217, + 114, + 228, + 126 + ], + "page_idx": 5 + }, + { + "type": "header", + "text": "Y. Zhang et al.", + "bbox": [ + 271, + 114, + 374, + 128 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "described in Equation 2 as follows:", + "bbox": [ + 215, + 146, + 467, + 161 + ], + "page_idx": 6 + }, + { + "type": "equation", + "text": "\n$$\nw _ {i} ^ {m 1} = w ^ {m 1} \\left[ \\colon D _ {i} \\cdot s, \\colon D _ {i} \\right] \\cdot \\alpha_ {i} \\quad b _ {i} ^ {m 1} = b ^ {m 1} \\left[ \\colon D _ {i} \\cdot s \\right]\n$$\n", + "text_format": "latex", + "bbox": [ + 321, + 172, + 681, + 189 + ], + "page_idx": 6 + }, + { + "type": "equation", + "text": "\n$$\nw _ {i} ^ {m 2} = w ^ {m 2} [: D _ {i},: D _ {i} \\cdot s ] \\cdot \\alpha_ {i} \\quad b _ {i} ^ {m 2} = b ^ {m 2} [: D _ {i} ]. \\tag {3}\n$$\n", + "text_format": "latex", + "bbox": [ + 321, + 185, + 784, + 210 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "**Elastic LN** For elastic LN, we directly use the first $D_{i}$ elements of the parameter inside the original LN, akin to the bias extraction in Equation 2.", + "bbox": [ + 215, + 222, + 782, + 252 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "To create a sub-network with $L_{i}$ elastic blocks from the intact ViT comprising $L_{max}$ blocks, we introduce a set of $N + 1$ elastic depths, defined as $L_{i} = L_{max} - i$ , $\\forall i \\in \\{0,1,\\dots,N\\}$ , $N < L_{max}$ . For a specific depth $L_{i}$ , we select the corresponding blocks based on their block IDs at equal intervals. Each block ID $BID_{j}^{L_{i}}$ that is activated for depth $L_{i}$ can be formulated as:", + "bbox": [ + 215, + 252, + 784, + 330 + ], + "page_idx": 6 + }, + { + "type": "equation", + "text": "\n$$\nB I D _ {j} ^ {L _ {i}} = \\left\\lfloor \\frac {\\left(L _ {\\max } - 1\\right) \\cdot j}{L _ {i} - 1} \\right\\rfloor , \\quad \\forall j \\in \\{0, 1, \\dots , L _ {i} - 1 \\}. \\tag {4}\n$$\n", + "text_format": "latex", + "bbox": [ + 316, + 340, + 784, + 372 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Consequently, by combining elastic widths and depths, we can generate a total of $(N + 1) \\cdot (M + 1)$ distinct sub-networks. For instance, by setting the elastic width to 384 and the elastic depth to 12, we can directly extract a ViT-S from the intact network such as ViT-L. During each iteration of the pre-training, we randomly select one of these sub-networks to serve as the elastic student branch.", + "bbox": [ + 214, + 377, + 782, + 452 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "3.2 Distillation between Views", + "text_level": 1, + "bbox": [ + 215, + 474, + 482, + 488 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "POA performs distillation across its three branches accordingly. Given a pair of globally augmented views of an input image $x$ , denoted as $x_{a}$ and $x_{b}$ , the teacher encoder $E_{T}$ extracts the feature $Z_{a} = E_{T}(x_{a})$ using $x_{a}$ as input. Simultaneously, $x_{b}$ is fed into both the intact student encoder $E_{IS}$ and the elastic student encoder $E_{ES}$ , producing the features $Z_{b1} = E_{IS}(x_{b})$ and $Z_{b2} = E_{ES}(x_{b})$ , respectively. The feature output from the teacher encoder, $Z_{a}$ , is then processed by the teacher head $H_{T}$ , followed by centering with the Sinkhorn-Knopp (SK) [11] algorithm and normalization using a temperature-scaled softmax to generate the probability $p_{a}$ as follows:", + "bbox": [ + 214, + 500, + 784, + 635 + ], + "page_idx": 6 + }, + { + "type": "equation", + "text": "\n$$\nl _ {a} = S K \\left(H _ {T} \\left(Z _ {a}\\right)\\right), l _ {a} \\in \\mathbb {R} ^ {P} \\quad p _ {a} ^ {i} = \\frac {\\exp \\left(l _ {a} ^ {i} / \\tau\\right)}{\\sum_ {k = 0} ^ {P - 1} \\exp \\left(l _ {a} ^ {k} / \\tau\\right)}, \\forall i \\in \\{0, \\dots , P - 1 \\}, \\tag {5}\n$$\n", + "text_format": "latex", + "bbox": [ + 227, + 659, + 784, + 696 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "where $P$ is the number of prototypes and $\\tau > 0$ is a temperature parameter. In a similar fashion, we compute the probabilities $p_{b1}^{i}$ and $p_{b2}^{i}$ for the intact and elastic student encoders, respectively, by processing the outputs through the student heads $H_{IS}$ and $H_{ES}$ . These outputs are then passed through a temperature-scaled softmax function using a temperature parameter $\\tau'$ tailored for the student. It should be noted that $H_{IS}$ and $H_{ES}$ share the identical parameters, except that the first projection layer of $H_{ES}$ is adapted similarly as in Equation 2 to conform to align with the corresponding dimensionality. For simplicity, we omit the explicit expressions for $p_{b1}^{i}$ and $p_{b2}^{i}$ , as they follow a similar", + "bbox": [ + 214, + 704, + 785, + 840 + ], + "page_idx": 6 + }, + { + "type": "header", + "text": "POA", + "bbox": [ + 694, + 114, + 730, + 126 + ], + "page_idx": 6 + }, + { + "type": "page_number", + "text": "7", + "bbox": [ + 774, + 114, + 785, + 126 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "calculation to Equation 5. For the intact student branch, we perform distillation from the teacher using the cross-view data as follows:", + "bbox": [ + 212, + 146, + 782, + 176 + ], + "page_idx": 7 + }, + { + "type": "equation", + "text": "\n$$\n\\mathcal {L} _ {I S} ^ {g} = - p _ {a} \\log \\left(p _ {b 1}\\right). \\tag {6}\n$$\n", + "text_format": "latex", + "bbox": [ + 431, + 191, + 785, + 209 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "The elastic student branch plays a pivotal role in our POA framework. To ensure adequate training of this branch, we employ a dual distillation from the teacher and intact student branch. The first distillation involves the teacher model, which utilizes the cross-view data to guide the learning of representations. The second is a distillation process with the intact student model, which uses the same-view data. This same-view distillation is responsible for transferring the representations learned by the intact student to the elastic student branch. The loss functions for this dual distillation process are formulated as follows:", + "bbox": [ + 212, + 213, + 784, + 333 + ], + "page_idx": 7 + }, + { + "type": "equation", + "text": "\n$$\n\\mathcal {L} _ {E S 1} ^ {g} = - p _ {a} \\log (p _ {b 2}), \\quad \\mathcal {L} _ {E S 2} ^ {g} = - p _ {b 1} \\log (p _ {b 2}). \\qquad (7)\n$$\n", + "text_format": "latex", + "bbox": [ + 339, + 348, + 785, + 367 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Note that in both loss functions, we sum over all prototypes to compute the cross-entropy loss between the respective probability distributions.", + "bbox": [ + 212, + 369, + 782, + 401 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "3.3 Overall Loss of POA", + "text_level": 1, + "bbox": [ + 215, + 421, + 431, + 435 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Following the SSL methods such as [5,33,55], we employ a multi-crop strategy [4] to create various distorted views from a single image. Apart from the two global views previously mentioned, we also generate $v$ local views with lower resolution $x_{l_1}, x_{l_2}, \\ldots, x_{l_v}$ . These local views are processed by both students to promote local-to-global correspondence. The local distillation losses for the intact and elastic students are computed as follows:", + "bbox": [ + 212, + 444, + 782, + 534 + ], + "page_idx": 7 + }, + { + "type": "equation", + "text": "\n$$\n\\mathcal {L} _ {I S} ^ {l} = - \\frac {1}{v} \\sum_ {i = 1} ^ {v} p _ {a} \\log \\left(p _ {l _ {i} 1}\\right), \\tag {8}\n$$\n", + "text_format": "latex", + "bbox": [ + 408, + 547, + 785, + 585 + ], + "page_idx": 7 + }, + { + "type": "equation", + "text": "\n$$\n\\mathcal {L} _ {E S 1} ^ {l} = - \\frac {1}{v} \\sum_ {i = 1} ^ {v} p _ {a} \\log \\left(p _ {l _ {i} 2}\\right), \\quad \\mathcal {L} _ {E S 2} ^ {l} = - \\frac {1}{v} \\sum_ {i = 1} ^ {v} p _ {l _ {i} 1} \\log \\left(p _ {l _ {i} 2}\\right), \\tag {9}\n$$\n", + "text_format": "latex", + "bbox": [ + 292, + 593, + 785, + 631 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "where $p_{l_{i1}}$ and $p_{l_{i2}}$ are the probability produced by the intact and elastic student branches for the local view $l_i$ , respectively. The total distillation loss of the intact and elastic student is calculated by summing them with the factor $\\lambda$ :", + "bbox": [ + 212, + 635, + 782, + 680 + ], + "page_idx": 7 + }, + { + "type": "equation", + "text": "\n$$\n\\begin{array}{l} \\mathcal {L} _ {\\mathcal {S}} = \\lambda \\left(\\mathcal {L} _ {I S} ^ {g} + \\mathcal {L} _ {I S} ^ {l}\\right) + (1 - \\lambda) \\left(\\left(\\mathcal {L} _ {E S 1} ^ {g} + \\mathcal {L} _ {E S 1} ^ {l}\\right) + \\left(\\mathcal {L} _ {E S 2} ^ {g} + \\mathcal {L} _ {E S 2} ^ {l}\\right)\\right) \\tag {10} \\\\ = \\lambda \\mathcal {L} _ {I S} + (1 - \\lambda) \\left(\\mathcal {L} _ {E S 1} + \\mathcal {L} _ {E S 2}\\right). \\\\ \\end{array}\n$$\n", + "text_format": "latex", + "bbox": [ + 259, + 689, + 785, + 724 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "To ensure sufficient training of each sub-network from the elastic student, we introduce multiple projection heads (MPH) positioned after the backbone network. Each projection head has exactly the same structure, except for a different number of prototypes. For each head, the distillation loss $\\mathcal{L}_{\\mathcal{S}_i}$ for both the intact and elastic student is calculated with Equation. 10. Finally, the overall loss function in the POA framework with $H$ projection heads is formulated as: $\\mathcal{L} = \\frac{1}{H}\\sum_{i = 1}^{H}\\mathcal{L}_{\\mathcal{S}_i}$ .", + "bbox": [ + 212, + 734, + 784, + 842 + ], + "page_idx": 7 + }, + { + "type": "page_number", + "text": "8", + "bbox": [ + 217, + 114, + 228, + 126 + ], + "page_idx": 7 + }, + { + "type": "header", + "text": "Y. Zhang et al.", + "bbox": [ + 271, + 114, + 374, + 128 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "4 Experiments", + "text_level": 1, + "bbox": [ + 215, + 145, + 374, + 162 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "4.1 Implementation Details", + "text_level": 1, + "bbox": [ + 215, + 176, + 455, + 191 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "Backbones. We have trained our POA using ViT, Swin Transformer and ResNet backbones, respectively. For the ViT, we configure the patch size to 16 and the dimension of each head in the MSA to 64. This aligns with the configurations typically used in existing SSL methods [5,34,35,55]. For the smallest and largest elastic networks of ViT, we choose the ViT-S and ViT-L, respectively. This leads to 11 elastic widths and 13 elastic depths, yielding a total of 143 ViT subnetworks. In the case of the Swin Transformer, we set the smallest and largest elastic networks as Swin-T and Swin-B, respectively. This configuration yields a total of 39 Swin sub-networks by multiplying the number of widths and depths as $3 \\times 13$ . For the ResNet, we establish the smallest and largest elastic network configurations as ResNet-50 and ResNet-152. Consequently, this setup accounts for a total of 465 ResNet sub-networks, which is the product of $3 \\times 155$ .", + "bbox": [ + 218, + 200, + 787, + 383 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "Pre-Training Setup. We pre-train all models on the ImageNet-1K dataset [12] without the labels. The process employs the AdamW optimizer [31] with a batch size of 1600, which is distributed across 32 A100 GPUs when employing a ViT backbone. We adopt a learning rate schedule that begins with a linear warm-up during the first 10 epochs, reaching a base value that is scaled proportionally to the total batch size: $\\mathrm{lr} = 0.004 \\times \\sqrt{\\text{batch size} / 1024}$ , in line with [33]. Following this warm-up period, the learning rate is decayed with a cosine schedule. Similarly, the weight decay follows a cosine schedule, starting at 0.04 and increasing to 0.4. The student network's temperature $\\tau'$ is fixed at 0.1, whereas the teacher temperature $\\tau$ changes with a linear warm-up from 0.04 to 0.07 over the first 30 epochs. For further details, please refer to Appendix C.", + "bbox": [ + 218, + 398, + 787, + 565 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "4.2 Experiments on ImageNet-1K", + "text_level": 1, + "bbox": [ + 215, + 587, + 508, + 603 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "After unsupervised pre-training, we assess the model's performance using two widely-recognized evaluation protocols in SSL domain on ImageNet-1K dataset, i.e., k-NN and linear probing. To ensure a fair comparison between SSL methods that employ different numbers of crop views for data augmentation, Zhou et al. [55] introduced the effective training epoch as a measure to quantify the extent of a method's pre-training. We report the effective training epochs of the SSL methods for comparison. For additional evaluations including fine-tuning, semi-supervised and unsupervised learning, please refer to Appendix D.", + "bbox": [ + 218, + 613, + 787, + 733 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "k-NN and Linear Probing. To assess the quality of pre-trained features, we employ a k-NN classifier and a linear classifier on the frozen representations. For both the k-NN and linear probing (LP) evaluation, we follow the evaluation protocols established in [5, 33, 55]. The performance of our method when being trained using ResNet, Swin Transformer and ViT backbones is reported in Table 1. Our POA SSL achieves the SOTA k-NN accuracy of $82.3\\%$ and LP accuracy", + "bbox": [ + 215, + 750, + 787, + 839 + ], + "page_idx": 8 + }, + { + "type": "header", + "text": "POA", + "bbox": [ + 694, + 114, + 730, + 126 + ], + "page_idx": 8 + }, + { + "type": "page_number", + "text": "9", + "bbox": [ + 774, + 116, + 785, + 126 + ], + "page_idx": 8 + }, + { + "type": "table", + "img_path": "images/9340ef9b865f2b883823d586bede6d6f2ab06e59800dcc838d868f142bb49f50.jpg", + "table_caption": [ + "Table 1: Comparison results of k-NN and linear probing classification accuracy (\\%) on the ImageNet-1K dataset. \"Param.\" refers to the number of parameters. \"Epo.\" represents the number of effective training epochs following [55]. \"/16\" denotes patch size of 16. \"/W7\" means the window size of 7. \"**\" indicates our implementation based on official codebase. \"+\" denotes reproduced results using the released code." + ], + "table_footnote": [ + "(a) ResNet backbone.", + "(b) Swin backbone.", + "(c) ViT backbone." + ], + "table_body": "
MethodEpo.k-NNLPMethodEpo.k-NNLPMethodEpo.k-NNLP
ResNet-50(Param. 23M)Swin-T/W7(Param. 28M)ViT-S/16(Param. 21M)
VICReg2000-73.2SMoG1200-77.7DINO320074.577.0
SwAV240065.775.3iBOT120075.378.6iBOT320075.277.9
DINO320067.575.3DINOv2*120075.478.0ENT320075.377.7
UniGrad2400-75.5EsViT120075.778.1POA076.877.6
SCFS320068.575.7POA077.578.9ViT-B/16(Param. 85M)
ReLICv24000-77.1Swin-S/W7(Param. 49M)DINO160076.178.2
POA073.476.9DINOv2*120076.179.8ENT160077.179.1
ResNet-101(Param. 41M)EsViT120077.779.5iBOT160077.179.5
ReLICv24000-78.7POA079.381.3POA080.981.7
POA075.779.1Swin-B/W7(Param. 87M)ViT-L/16(Param. 307M)
ResNet-152(Param. 56M)DINOv2*120076.980.9iBOT120078.081.0
ReLICv24000-79.3EsViT120078.980.4DINOv2†120082.083.3
POA240076.479.9POA120079.682.0POA120082.383.6
", + "bbox": [ + 215, + 224, + 777, + 463 + ], + "page_idx": 9 + }, + { + "type": "text", + "text": "of $83.6\\%$ when using the ViT-L/16 backbone. By employing the sub-network extraction approach outlined in Section 3.1, we derive the sub-networks ViT-S/16 and ViT-B/16 from the teacher ViT-L/16 without any additional pre-training. Thus, the number of effective training epochs of them is 0. Notably, the extracted ViT-S/16 achieves the SOTA k-NN accuracy of $76.8\\%$ and LP accuracy on par with the previous SOTA of $77.9\\%$ reported by iBOT [55]. Our derived ViT-B/16 model sets new benchmarks for both k-NN and LP, with accuracy of $80.9\\%$ and $81.7\\%$ , respectively. For the models using Swin Transformer and ResNet backbones, POA also reaches SOTA performance in k-NN and LP accuracy. The only exception is the LP accuracy of ResNet-50, which is competitive with that of ReLICv2 [39], despite the latter being trained over a much longer period (4000 epochs). The superior performance achieved across a range of backbone architectures confirms the effectiveness and versatility of our method. For additional detailed comparison with 27 existing methods, please see Appendix D.1.", + "bbox": [ + 212, + 493, + 787, + 705 + ], + "page_idx": 9 + }, + { + "type": "text", + "text": "4.3 Evaluation on Downstream Visual Tasks", + "text_level": 1, + "bbox": [ + 215, + 724, + 594, + 739 + ], + "page_idx": 9 + }, + { + "type": "text", + "text": "Object Detection and Instance Segmentation on COCO Dataset. For a fair comparison, we utilize the Cascade Mask R-CNN framework [3,20], which generates both bounding boxes and instance masks, in line with previous approach [55], on the COCO dataset [29]. We benchmark our results against existing SSL methods that generate pre-trained ViT-S/16 and ViT-B/16 backbones. As shown in Table 2, POA boosts the bounding box average precision $(\\mathrm{AP}^b)$ for", + "bbox": [ + 212, + 748, + 787, + 840 + ], + "page_idx": 9 + }, + { + "type": "page_number", + "text": "10", + "bbox": [ + 217, + 114, + 235, + 126 + ], + "page_idx": 9 + }, + { + "type": "header", + "text": "Y. Zhang et al.", + "bbox": [ + 271, + 114, + 374, + 128 + ], + "page_idx": 9 + }, + { + "type": "text", + "text": "ViT-S/16 from 49.4 to 50.6 and the mask average precision $(\\mathrm{AP}^m)$ from 42.6 to 43.8. When applied to ViT-B/16, POA attains an $\\mathrm{AP}^b$ of 52.4 and an $\\mathrm{AP}^m$ of 45.4, which represents a remarkable step forward over previous SOTA.", + "bbox": [ + 212, + 145, + 787, + 191 + ], + "page_idx": 10 + }, + { + "type": "text", + "text": "Semantic Segmentation on ADE20K Dataset. For the semantic segmentation task, we primarily focus on two settings on the ADE20K dataset [54], following [55]. First, akin to the linear evaluation protocol in classification, we evaluate the quality of representations by keeping the patch features fixed and only fine-tuning one linear layer. This approach offers a clear comparison of representation quality. Second, we employ the UPerNet [46] as the task head and fine-tune the entire network. As depicted in Table 2, our POA significantly outperforms the supervised baseline using the ViT-S/16 backbone, achieving a substantial increase of 2.2 in mean Intersection over Union (mIoU), and surpassing iBOT by 1.3 mIoU. With the ViT-B/16 backbone, POA exceeds the previously best-performing method, iBOT, by 0.4 mIoU when utilizing UPerNet. Furthermore, under the assessment using solely a linear head, POA obtains an impressive improvement of 2.0 mIoU over iBOT as the performance is largely determined by the quality of the pre-trained representation.", + "bbox": [ + 212, + 203, + 787, + 414 + ], + "page_idx": 10 + }, + { + "type": "table", + "img_path": "images/f34a8722e9d380d7b21c8742b47700e8a5528ff6ffd3f9c2f7824a92c636f093.jpg", + "table_caption": [ + "Table 2: Evaluation results on downstream detection and segmentation tasks. $\\mathrm{Seg.}^{\\dagger}$ indicates the use of a linear head for semantic segmentation." + ], + "table_footnote": [], + "table_body": "
MethodArch.Det. APbISeg. APmSeg. mIoUArch.Det. APbISeg. APmSeg.† mIoUSeg. mIoU
Sup.ViT-S/1646.240.144.5ViT-B/1649.843.235.446.6
BEiTViT-S/16---ViT-B/1650.143.527.445.8
DINOViT-S/16---ViT-B/1650.143.434.546.8
iBOTViT-S/1649.442.645.4ViT-B/1651.244.238.350.0
POAViT-S/1650.643.846.7ViT-B/1652.445.440.350.4
", + "bbox": [ + 241, + 457, + 761, + 571 + ], + "page_idx": 10 + }, + { + "type": "text", + "text": "4.4 Ablations and Discussions", + "text_level": 1, + "bbox": [ + 214, + 594, + 478, + 608 + ], + "page_idx": 10 + }, + { + "type": "text", + "text": "In this section, we conduct an empirical analysis of POA using ViT as backbone. Our investigation includes the impact of the loss functions $\\mathcal{L}_{ES1}$ and $\\mathcal{L}_{ES2}$ , in addition with the effectiveness of multiple projection heads. Moreover, we compare our POA with knowledge distillation techniques for self-supervised learning to demonstrate the advantages of combining once-for-all model generation with self-distillation in a unified framework. For more ablation studies, we direct readers to the Appendix D.8. In addition, we further contrast our POA with three variants tailored for elastic pre-training to showcase POA's superiority. Finally, we discuss how the elastic student facilitates the pre-training.", + "bbox": [ + 212, + 614, + 787, + 752 + ], + "page_idx": 10 + }, + { + "type": "text", + "text": "Importance of Each Component. We evaluate the contributions of the components on a ViT backbone. Table 3 shows the performance of different component combinations. First, we note that employing multiple projection heads (MPH) enhances the learned representations for each elastic sub-network, particularly for smaller ones. The design consideration behind MPH is that for each", + "bbox": [ + 212, + 763, + 787, + 839 + ], + "page_idx": 10 + }, + { + "type": "header", + "text": "POA", + "bbox": [ + 694, + 114, + 730, + 126 + ], + "page_idx": 10 + }, + { + "type": "page_number", + "text": "11", + "bbox": [ + 767, + 114, + 784, + 126 + ], + "page_idx": 10 + }, + { + "type": "table", + "img_path": "images/73bb1f1a90420509d108ec5fc9f688229ef3a5b604e65f9ca9a223841c39366c.jpg", + "table_caption": [ + "Table 3: Contributions of each component in POA framework. We conduct assessments with the k-NN and linear probing (LP) evaluations. MPH denotes the multiple projection heads with different numbers of prototypes." + ], + "table_footnote": [], + "table_body": "
MPHLES1LES2k-NNLP
ViT-S/16ViT-B/16ViT-L/16ViT-S/16ViT-B/16ViT-L/16
✓✓✓76.880.982.377.681.783.6
✓✓76.280.782.277.381.683.4
✓75.180.282.275.881.083.4
✓72.879.182.175.380.883.3
", + "bbox": [ + 225, + 198, + 777, + 297 + ], + "page_idx": 11 + }, + { + "type": "text", + "text": "pre-training iteration, the sub-network is chosen randomly, leading to a relatively insufficient optimization. MPH introduces different sets of prototypes, which act as multiple semantic spaces for representation learning, enabling the teacher to distill various aspects of learned knowledge into the sub-network. Furthermore, we ascertain that the same-view distillation loss $\\mathcal{L}_{ES2}$ is crucial for the representation quality of elastic sub-networks. Omitting $\\mathcal{L}_{ES2}$ causes a significant drop in k-NN accuracy, by $3.4\\%$ for ViT-S/16 and $1.6\\%$ for ViT-B/16. In addition, $\\mathcal{L}_{ES2}$ is more important than cross-view distillation $\\mathcal{L}_{ES1}$ in terms of sub-networks' representation quality. The underlying reason is that the cross-view distillation is for the unsupervised representation learning, while the same-view distillation improves the sub-networks by distilling previously learned representations from the intact student. Table 4 confirms that distillation from already-good representations is more effective than representation learning, especially for smaller networks. It also explains why employing three different views in our POA is unnecessary.", + "bbox": [ + 212, + 303, + 787, + 529 + ], + "page_idx": 11 + }, + { + "type": "text", + "text": "Table 4: Comparison with knowledge distillation for self-supervised learning. The teacher name \"ViT-L/16-600\" denotes a teacher model (ViT-L/16) that has been pretrained with 600 effective epochs. The student name \"ViT-S/16-600\" refers to a student model (ViT-S/16) that has undergone distillation from the pre-trained teacher with 600 effective epochs. The number of the left side of \" $\\rightarrow$ \" indicates the performance of the teacher model when pre-trained individually. The number of the right side denotes the performance achieved after distillation using the SEED [15].", + "bbox": [ + 212, + 531, + 789, + 630 + ], + "page_idx": 11 + }, + { + "type": "table", + "img_path": "images/82af6ed40a3c842b8077784a3655cdd874ce1e88cb85454ba3f2a516af52d216.jpg", + "table_caption": [], + "table_footnote": [], + "table_body": "
MethodTeacherStudentTotal Epochsk-NNLP
DINOv2ViT-S/16120072.273.1
DINOv2+SEEDViT-L/16-600ViT-S/16-600120081.3 → 74.082.4 → 75.2
DINOv2+SEEDViT-L/16-1200ViT-S/16-1200240082.0 → 75.583.3 → 76.2
POAViT-S/16120076.877.6
DINOv2ViT-B/16120077.478.5
DINOv2+SEEDViT-L/16-600ViT-B/16-600120081.3 → 78.882.4 → 80.0
DINOv2+SEEDViT-L/16-1200ViT-B/16-1200240082.0 → 79.783.3 → 80.9
POAViT-B/16120080.981.7
", + "bbox": [ + 215, + 641, + 789, + 790 + ], + "page_idx": 11 + }, + { + "type": "text", + "text": "Comparison with Knowledge Distillation. Knowledge distillation (KD) is a proven strategy for enhancing small networks by leveraging the knowledge of", + "bbox": [ + 215, + 808, + 787, + 840 + ], + "page_idx": 11 + }, + { + "type": "page_number", + "text": "12", + "bbox": [ + 217, + 114, + 235, + 126 + ], + "page_idx": 11 + }, + { + "type": "header", + "text": "Y. Zhang et al.", + "bbox": [ + 271, + 114, + 374, + 128 + ], + "page_idx": 11 + }, + { + "type": "text", + "text": "a well-trained, larger network. We compare our method with a self-supervised KD method SEED [15] that employs a pre-trained network from the previous SOTA DINOv2 as the teacher for distilling knowledge into smaller networks. We observe that SEED significantly boosts the performance of ViT-S/16 and ViT-B/16 when a pre-trained ViT-L/16 serves as the teacher. Specifically, with the same number of training epochs, SEED yields k-NN accuracy improvements of $1.8\\%$ and $1.4\\%$ for ViT-S/16 and ViT-B/16 respectively, compared with a learnfrom-scratch setup. Our POA further outperforms SEED, delivering superior k-NN accuracy gains of $2.8\\%$ (76.8% vs. 74.0%) and $2.1\\%$ (80.9% vs. 78.8%) on ViT-S/16 and ViT-B/16. Moreover, the ViT-S/16 and ViT-B/16 derived directly from the pre-trained teacher from our POA, without additional pre-training, perform better than those enhanced by SEED, despite the latter undergoing two times of training epochs. These results demonstrate the effectiveness of our unified design, which integrates pre-training with once-for-all model generation. Additional comparison on training cost can be found in Appendix D.11.", + "bbox": [ + 212, + 146, + 787, + 373 + ], + "page_idx": 12 + }, + { + "type": "text", + "text": "Alternative Designs to Elastic Pre-training. Upon the proposed POA, we investigate three alternatives for concurrently pre-training multi-sized models within a single pre-training session. As illustrated in Figure 4, the first variant, POA-V1, eliminates the intact student and modifies the teacher to be elastic which aligns with the architecture of the elastic student. The second variant, POA-V2, only removes the intact student. However, in this model, the intact teacher is directly updated via an EMA of the elastic students. These two variants bear resemblance to current teacher-student self-distillation paradigm, with the primary differences in the structure of the teacher or student network. For the POA-V3, we introduce an additional elastic teacher branch that mirrors the architecture of the elastic student, enabling further cross-view distillation. For more details of the three variants, please refer to the Appendix D.9.", + "bbox": [ + 212, + 391, + 787, + 571 + ], + "page_idx": 12 + }, + { + "type": "image", + "img_path": "images/63f78a649a4a00db6dcb1fd57c087ad50e27c7a5cd5e7f1d95f95aebf15b26ae.jpg", + "image_caption": [ + "Fig. 4: Illustration of different variants of POA." + ], + "image_footnote": [], + "bbox": [ + 222, + 574, + 789, + 688 + ], + "page_idx": 12 + }, + { + "type": "text", + "text": "We compare the k-NN and linear probing accuracy of each variant in Table 5. The results indicate that the performance drops dramatically for the variants with two branches. This decline can be attributed to two reasons: 1) in each training iteration, only a subset of the student's parameters are updated, leading to relatively insufficient overall pre-training; and 2) each elastic network undergoes training exclusively with the cross-view distillation, missing out the standard distillation guidance from a larger network, which our ablation study suggests is crucial for the efficacy of elastic networks. Although the POA-V3,", + "bbox": [ + 212, + 719, + 787, + 840 + ], + "page_idx": 12 + }, + { + "type": "header", + "text": "POA", + "bbox": [ + 694, + 114, + 730, + 126 + ], + "page_idx": 12 + }, + { + "type": "page_number", + "text": "13", + "bbox": [ + 767, + 114, + 785, + 126 + ], + "page_idx": 12 + }, + { + "type": "text", + "text": "featuring an additional elastic teacher branch, performs slightly better than our POA, it does so at the expense of increased model complexity and computational costs, making it a less appealing scheme than POA.", + "bbox": [ + 212, + 146, + 782, + 191 + ], + "page_idx": 13 + }, + { + "type": "table", + "img_path": "images/5283dc82d277162ae9426518cd5c6f50f63888feeba0b7f79d4fd593724d6140.jpg", + "table_caption": [ + "Table 5: k-NN and linear probing (LP) comparison of different variants of POA." + ], + "table_footnote": [], + "table_body": "
Methodk-NNLP
ViT-S/16ViT-B/16ViT-L/16ViT-S/16ViT-B/16ViT-L/16
POA76.880.982.377.681.783.6
POA-V173.778.579.274.579.480.9
POA-V274.379.180.975.180.082.3
POA-V376.981.082.377.881.883.6
", + "bbox": [ + 223, + 220, + 779, + 321 + ], + "page_idx": 13 + }, + { + "type": "text", + "text": "How Does Elastic Student Facilitate Pre-training? The elastic student branch enables the derivation of diverse-sized models directly from a pre-trained teacher, while simultaneously enhancing the learned representation. The elastic branch plays a dual role. First, it acts as a training regularization to stabilize the training progress. In our preliminary experiments, we observe that pre-training loss of a ResNet backbone easily diverges to NaN without the elastic branch. Conversely, the inclusion of the elastic student yields a highly stable pre-training. Furthermore, the elastic student provides hundreds of sub-network candidates that will be assembled to the teacher during pre-training. Unlike existing self-distillation methods, the teacher in the POA SSL integrates a series of sub-networks through an EMA update. Wortsman et al. [44] have shown that averaging the weights of multiple models typically improves accuracy and robustness. Similarly, the integration of different sub-networks for the teacher is effective in improving the representation quality. At the same time, an improved teacher promotes the learning of students in return, fostering a positive feedback loop for pre-training. For the visualization of representations, please refer to Appendix E.2.", + "bbox": [ + 212, + 340, + 787, + 599 + ], + "page_idx": 13 + }, + { + "type": "text", + "text": "5 Conclusion", + "text_level": 1, + "bbox": [ + 215, + 625, + 359, + 641 + ], + "page_idx": 13 + }, + { + "type": "text", + "text": "In this study, we tackle the challenge of efficiently and effectively pre-training models of various sizes within a single self-supervised learning session, facilitating model deployment given different resource constraints. We propose a novel self-supervised learning paradigm, termed POA, which features an integrated design combining self-distillation and once-for-all model generation. It allows for the simultaneous pre-training of models of multiple sizes through an elastic branch design. POA enables the direct generation of varied-sized models from a pretrained teacher, which are ready for downstream tasks without additional pretraining. This advantage significantly improves the deployment flexibility and facilitates our pre-trained model to achieve SOTA results across various vision tasks. Looking forward, we plan to extend POA to Multimodal Large Language Models, tapping into its vast potential for real-world AI product deployment.", + "bbox": [ + 212, + 657, + 787, + 840 + ], + "page_idx": 13 + }, + { + "type": "page_number", + "text": "14", + "bbox": [ + 217, + 114, + 235, + 126 + ], + "page_idx": 13 + }, + { + "type": "header", + "text": "Y. Zhang et al.", + "bbox": [ + 271, + 114, + 372, + 128 + ], + "page_idx": 13 + }, + { + "type": "text", + "text": "References", + "text_level": 1, + "bbox": [ + 215, + 143, + 323, + 160 + ], + "page_idx": 14 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "1. Ba, J., Kiros, J.R., Hinton, G.E.: Layer normalization. ArXiv abs/1607.06450 (2016)", + "2. Cai, H., Gan, C., Wang, T., Zhang, Z., Han, S.: Once for all: Train one network and specialize it for efficient deployment. In: International Conference on Learning Representations (2020)", + "3. Cai, Z., Vasconcelos, N.: Cascade r-cnn: High quality object detection and instance segmentation. IEEE Transactions on Pattern Analysis and Machine Intelligence 43, 1483-1498 (2019)", + "4. Caron, M., Misra, I., Mairal, J., Goyal, P., Bojanowski, P., Joulin, A.: Unsupervised learning of visual features by contrasting cluster assignments. In: Larochelle, H., Ranzato, M., Hadsell, R., Balcan, M., Lin, H. (eds.) Advances in Neural Information Processing Systems. vol. 33, pp. 9912-9924. Curran Associates, Inc. (2020)", + "5. Caron, M., Touvron, H., Misra, I., J'egou, H., Mairal, J., Bojanowski, P., Joulin, A.: Emerging properties in self-supervised vision transformers. 2021 IEEE/CVF International Conference on Computer Vision (ICCV) pp. 9630-9640 (2021)", + "6. Chen, T., Kornblith, S., Norouzi, M., Hinton, G.E.: A simple framework for contrastive learning of visual representations. International Conference on Machine Learning (2021)", + "7. Chen, T., Zhai, X., Ritter, M., Lucic, M., Houlsby, N.: Self-supervised gans via auxiliary rotation loss. 2019 IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR) pp. 12146-12155 (2018)", + "8. Chen, X., Fan, H., Girshick, R., He, K.: Improved baselines with momentum contrastive learning. arXiv preprint arXiv:2003.04297 (2020)", + "9. Chen, Y., Liu, Y., Jiang, D., Zhang, X., Dai, W., Xiong, H., Tian, Q.: Sdae: Self-distillated masked autoencoder. In: European Conference on Computer Vision (2022)", + "10. Cheng, X., Chen, J., Wang, R.: Modified dual attention triplet-supervised hashing network for image retrieval. Signal, Image and Video Processing 18, 1939-1948 (2023)", + "1. Cuturi, M.: Sinkhorn distances: Lightspeed computation of optimal transport. In: Neural Information Processing Systems (2013)", + "2. Deng, J., Dong, W., Socher, R., Li, L.J., Li, K., Fei-Fei, L.: ImageNet: A LargeScale Hierarchical Image Database. In: CVPR 2009 (2009)", + "3. Dong, C., Loy, C.C., He, K., Tang, X.: Image super-resolution using deep convolutional networks. IEEE Transactions on Pattern Analysis and Machine Intelligence 38, 295-307 (2014)", + "4. Dosovitskiy, A., Beyer, L., Kolesnikov, A., Weissenborn, D., Zhai, X., Unterthiner, T., Dehghani, M., Minderer, M., Heigold, G., Gelly, S., Uszkoreit, J., Houlsby, N.: An image is worth 16x16 words: Transformers for image recognition at scale. International Conference on Learning Representations (ICLR) (2021)", + "5. Fang, Z., Wang, J., Wang, L., Zhang, L., Yang, Y., Liu, Z.: Seed: Self-supervised distillation for visual representation. International Conference on Learning Representations (2021)", + "6. Gao, Y., Zhuang, J.X., Lin, S., Cheng, H., Sun, X., Li, K., Shen, C.: Disco: Remedying self-supervised learning on lightweight models with distilled contrastive learning. In: European Conference on Computer Vision (ECCV) (2022)", + "7. Grill, J.B., Strub, F., Altché, F., Tallec, C., Richemond, P., Buchatskaya, E., Doersch, C., Avila Pires, B., Guo, Z., Gheshlaghi Azar, M., et al.: Bootstrap your own" + ], + "bbox": [ + 225, + 176, + 784, + 839 + ], + "page_idx": 14 + }, + { + "type": "header", + "text": "POA", + "bbox": [ + 694, + 114, + 730, + 126 + ], + "page_idx": 14 + }, + { + "type": "page_number", + "text": "15", + "bbox": [ + 767, + 114, + 785, + 126 + ], + "page_idx": 14 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "latent-a new approach to self-supervised learning. Advances in neural information processing systems 33, 21271-21284 (2020)", + "18. He, K., Chen, X., Xie, S., Li, Y., Doll'ar, P., Girshick, R.B.: Masked autoencoders are scalable vision learners. 2022 IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR) pp. 15979-15988 (2021)", + "19. He, K., Fan, H., Wu, Y., Xie, S., Girshick, R.B.: Momentum contrast for unsupervised visual representation learning. 2020 IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR) pp. 9726-9735 (2019)", + "20. He, K., Gkioxari, G., Dolkar, P., Girshick, R.B.: Mask r-cnn. 2017 IEEE International Conference on Computer Vision (ICCV) pp. 2980-2988 (2017)", + "21. He, K., Zhang, X., Ren, S., Sun, J.: Deep residual learning for image recognition. 2016 IEEE Conference on Computer Vision and Pattern Recognition (CVPR) pp. 770-778 (2015)", + "22. Hinton, G.E., Vinyals, O., Dean, J.: Distilling the knowledge in a neural network. ArXiv abs/1503.02531 (2015)", + "23. Hou, L., Shen, H., Cao, Q., Cheng, X.: Self-supervised gans with label augmentation. ArXiv abs/2106.08601 (2021)", + "24. Hu, C., Li, X., Liu, D., Wu, H., Chen, X., Wang, J., Liu, X.: Teacher-student architecture for knowledge distillation: A survey. ArXiv abs/2308.04268 (2023)", + "25. Huang, Z., Jin, X., Lu, C., Hou, Q., Cheng, M.M., Fu, D., Shen, X., Feng, J.: Contrastive masked autoencoders are stronger vision learners. IEEE transactions on pattern analysis and machine intelligence PP (2022)", + "26. Li, J., Wang, Y., ZHANG, X., Chen, Y., Jiang, D., Dai, W., Li, C., Xiong, H., Tian, Q.: Progressively compressed auto-encoder for self-supervised representation learning. In: The Eleventh International Conference on Learning Representations (2023)", + "27. Li, Z., Zhu, Y., Yang, F., Li, W., Zhao, C., Chen, Y., Chen, Z., Xie, J., Wu, L., Zhao, R., Tang, M., Wang, J.: Univip: A unified framework for self-supervised visual pre-training. 2022 IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR) pp. 14607-14616 (2022)", + "28. Lin, J., Rao, Y., Lu, J., Zhou, J.: Runtime neural pruning. In: Guyon, I., Luxburg, U.V., Bengio, S., Wallach, H., Fergus, R., Vishwanathan, S., Garnett, R. (eds.) Advances in Neural Information Processing Systems. vol. 30. Curran Associates, Inc. (2017)", + "29. Lin, T.Y., Maire, M., Belongie, S.J., Hays, J., Perona, P., Ramanan, D., Dóllar, P., Zitnick, C.L.: Microsoft coco: Common objects in context. In: European Conference on Computer Vision (2014)", + "30. Liu, Z., Lin, Y., Cao, Y., Hu, H., Wei, Y., Zhang, Z., Lin, S., Guo, B.: Swin transformer: Hierarchical vision transformer using shifted windows. 2021 IEEE/CVF International Conference on Computer Vision (ICCV) pp. 9992-10002 (2021)", + "31. Loshchilov, I., Hutter, F.: Fixing weight decay regularization in adam. ArXiv abs/1711.05101 (2017)", + "32. van den Oord, A., Li, Y., Vinyals, O.: Representation learning with contrastive predictive coding. ArXiv abs/1807.03748 (2018)", + "33. Oquab, M., Darcet, T., Moutakanni, T., Vo, H.V., Szafraniec, M., Khalidov, V., Fernandez, P., HAZIZA, D., Massa, F., El-Nouby, A., Assran, M., Ballas, N., Galuba, W., Howes, R., Huang, P.Y., Li, S.W., Misra, I., Rabbat, M., Sharma, V., Synnaeve, G., Xu, H., Jegou, H., Mairal, J., Labatut, P., Joulin, A., Bojanowski, P.: DINOv2: Learning robust visual features without supervision. Transactions on Machine Learning Research (2024)" + ], + "bbox": [ + 215, + 147, + 785, + 839 + ], + "page_idx": 15 + }, + { + "type": "page_number", + "text": "16", + "bbox": [ + 217, + 114, + 235, + 126 + ], + "page_idx": 15 + }, + { + "type": "header", + "text": "Y. Zhang et al.", + "bbox": [ + 271, + 114, + 374, + 128 + ], + "page_idx": 15 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "34. Ruan, Y., Singh, S., Morningstar, W.R., Alemi, A.A., Ioffe, S., Fischer, I., Dillon, J.V.: Weighted ensemble self-supervised learning. In: The Eleventh International Conference on Learning Representations (2023)", + "35. Song, K., Xie, J., Zhang, S., Luo, Z.: Multi-mode online knowledge distillation for self-supervised visual representation learning. 2023 IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR) pp. 11848-11857 (2023)", + "36. Tao, C., Wang, H., Zhu, X., Dong, J., Song, S., Huang, G., Dai, J.: Exploring the equivalence of siamese self-supervised learning via a unified gradient framework. 2022 IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR) pp. 14411-14420 (2021)", + "37. Team, G.G.: Introducing gemini: our largest and most capable ai model. Google (2023)", + "38. Tian, K., Jiang, Y., qishuai diao, Lin, C., Wang, L., Yuan, Z.: Designing BERT for convolutional networks: Sparse and hierarchical masked modeling. In: The Eleventh International Conference on Learning Representations (2023)", + "39. Tomasev, N., Bica, I., McWilliams, B., Buesing, L., Pascanu, R., Blundell, C., Mitrovic, J.: Pushing the limits of self-supervised resnets: Can we outperform supervised learning without labels onImagenet? arXiv preprint arXiv:2201.05119 (2022)", + "40. Touvron, H., Cord, M., Oquab, M., Bojanowski, P., Verbeek, J., Jegou, H.: Cotraining 2l submodels for visual recognition. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR). pp. 11701-11710 (June 2023)", + "41. Wang, X., Huang, Y., Zeng, D., Qi, G.J.: Caco: Both positive and negative samples are directly learnable via cooperative-adversarial contrastive learning. IEEE Transactions on Pattern Analysis and Machine Intelligence 45, 10718-10730 (2022)", + "42. Wang, X., Yu, F., Dou, Z.Y., Gonzalez, J.: Skipnet: Learning dynamic routing in convolutional networks. ArXiv abs/1711.09485 (2017)", + "43. Wang, X., Zhang, R., Shen, C., Kong, T., Li, L.: Dense contrastive learning for self-supervised visual pre-training. 2021 IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR) pp. 3023-3032 (2020)", + "44. Wortsman, M., Ilharco, G., Gadre, S.Y., Roelofs, R., Gontijo-Lopes, R., Morcos, A.S., Namkoong, H., Farhadi, A., Carmon, Y., Kornblith, S., Schmidt, L.: Model soups: averaging weights of multiple fine-tuned models improves accuracy without increasing inference time. In: International Conference on Machine Learning (ICML) (2022)", + "45. Wu, H., Xu, J., Wang, J., Long, M.: Autoformer: Decomposition transformers with auto-correlation for long-term series forecasting. In: Neural Information Processing Systems (2021)", + "46. Xiao, T., Liu, Y., Zhou, B., Jiang, Y., Sun, J.: Unified perceptual parsing for scene understanding. ArXiv abs/1807.10221 (2018)", + "47. Xie, Z., Lin, Y., Yao, Z., Zhang, Z., Dai, Q., Cao, Y., Hu, H.: Self-supervised learning with swin transformers. ArXiv abs/2105.04553 (2021)", + "48. Xie, Z., Zhang, Z., Cao, Y., Lin, Y., Bao, J., Yao, Z., Dai, Q., Hu, H.: Simmim: a simple framework for masked image modeling. 2022 IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR) pp. 9643-9653 (2021)", + "49. Yan, B., Jiang, Y., Sun, P., Wang, D., Yuan, Z., Luo, P., Lu, H.: Towards grand unification of object tracking. In: European Conference on Computer Vision (2022)", + "50. Yan, C., Chang, X., Li, Z., Yao, L., Luo, M., Zheng, Q.: Masked distillation advances self-supervised transformer architecture search. In: The Twelfth Interna" + ], + "bbox": [ + 212, + 146, + 787, + 839 + ], + "page_idx": 16 + }, + { + "type": "header", + "text": "POA", + "bbox": [ + 694, + 114, + 730, + 126 + ], + "page_idx": 16 + }, + { + "type": "page_number", + "text": "17", + "bbox": [ + 767, + 114, + 785, + 126 + ], + "page_idx": 16 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "tional Conference on Learning Representations (2024), https://openreview.net/ forum?id=LUpC8KTvdV", + "51. Yu, F., Huang, K., Wang, M., Cheng, Y., Chu, W., Cui, L.: Width & depth pruning for vision transformers. In: AAAI Conference on Artificial Intelligence (2022)", + "52. Yu, J., Li, X., Koh, J.Y., Zhang, H., Pang, R., Qin, J., Ku, A., Xu, Y., Baldridge, J., Wu, Y.: Vector-quantized image modeling with improved VQGAN. In: International Conference on Learning Representations (2022)", + "53. Zhang, Y., Zhong, Q., Ma, L., Xie, D., Pu, S.: Learning incremental triplet margin for person re-identification. AAAI Conference on Artificial Intelligence (2019)", + "54. Zhou, B., Zhao, H., Puig, X., Fidler, S., Barriuso, A., Torralba, A.: Scene parsing through ade20k dataset. 2017 IEEE Conference on Computer Vision and Pattern Recognition (CVPR) pp. 5122-5130 (2017)", + "55. Zhou, J., Wei, C., Wang, H., Shen, W., Xie, C., Yuille, A., Kong, T.: ibot: Image bert pre-training with online tokenizer. International Conference on Learning Representations (ICLR) (2022)" + ], + "bbox": [ + 215, + 147, + 785, + 354 + ], + "page_idx": 17 + }, + { + "type": "page_number", + "text": "18", + "bbox": [ + 217, + 114, + 235, + 126 + ], + "page_idx": 17 + }, + { + "type": "header", + "text": "Y. Zhang et al.", + "bbox": [ + 271, + 114, + 374, + 128 + ], + "page_idx": 17 + } +] \ No newline at end of file diff --git a/2024/POA_ Pre-training Once for Models of All Sizes/9b5de88d-0728-41f9-a801-73fd67e5b8a0_model.json b/2024/POA_ Pre-training Once for Models of All Sizes/9b5de88d-0728-41f9-a801-73fd67e5b8a0_model.json new file mode 100644 index 0000000000000000000000000000000000000000..dec230e381c76438166debe335b3f8bc00024b3c --- /dev/null +++ b/2024/POA_ Pre-training Once for Models of All Sizes/9b5de88d-0728-41f9-a801-73fd67e5b8a0_model.json @@ -0,0 +1,2238 @@ +[ + [ + { + "type": "title", + "bbox": [ + 0.23, + 0.142, + 0.776, + 0.163 + ], + "angle": 0, + "content": "POA: Pre-training Once for Models of All Sizes" + }, + { + "type": "text", + "bbox": [ + 0.231, + 0.19, + 0.772, + 0.221 + ], + "angle": 0, + "content": "Yingying Zhang, Xin Guo, Jiangwei Lao, Lei Yu, Lixiang Ru, Jian Wang, Guo Ye, Huimei He, Jingdong Chen, and Ming Yang" + }, + { + "type": "text", + "bbox": [ + 0.462, + 0.234, + 0.54, + 0.247 + ], + "angle": 0, + "content": "Ant Group." + }, + { + "type": "text", + "bbox": [ + 0.418, + 0.248, + 0.59, + 0.261 + ], + "angle": 0, + "content": "qichu.zyy@antgroup.com" + }, + { + "type": "text", + "bbox": [ + 0.263, + 0.299, + 0.74, + 0.603 + ], + "angle": 0, + "content": "Abstract. Large-scale self-supervised pre-training has paved the way for one foundation model to handle many different vision tasks. Most pre-training methodologies train a single model of a certain size at one time. Nevertheless, various computation or storage constraints in real-world scenarios require substantial efforts to develop a series of models with different sizes to deploy. Thus, in this study, we propose a novel tri-branch self-supervised training framework, termed as POA (Pre-training Once for All), to tackle this aforementioned issue. Our approach introduces an innovative elastic student branch into a modern self-distillation paradigm. At each pre-training step, we randomly sample a sub-network from the original student to form the elastic student and train all branches in a self-distilling fashion. Once pre-trained, POA allows the extraction of pre-trained models of diverse sizes for downstream tasks. Remarkably, the elastic student facilitates the simultaneous pre-training of multiple models with different sizes, which also acts as an additional ensemble of models of various sizes to enhance representation learning. Extensive experiments, including k-nearest neighbors, linear probing evaluation and assessments on multiple downstream tasks demonstrate the effectiveness and advantages of our POA. It achieves state-of-the-art performance using ViT, Swin Transformer and ResNet backbones, producing around a hundred models with different sizes through a single pre-training session. The code is available at: https://github.com/Qichuzyy/POA." + }, + { + "type": "text", + "bbox": [ + 0.263, + 0.617, + 0.695, + 0.631 + ], + "angle": 0, + "content": "Keywords: Self-supervised Learning \\(\\cdot\\) Pre-training Once for All" + }, + { + "type": "title", + "bbox": [ + 0.218, + 0.658, + 0.375, + 0.673 + ], + "angle": 0, + "content": "1 Introduction" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.689, + 0.788, + 0.841 + ], + "angle": 0, + "content": "Learning generalizable visual representations in a large model by self-supervised learning has delivered superior performance across a wide range of visual tasks [10, 13, 20, 46, 49, 53] in recent years. Nevertheless, when deployed to real-world applications, a large model has to be adapted to diverse resource limitations in terms of computation, storage, or power consumption, etc. For example, a well-engineered AI product typically comprises a suite of models tailored for varying scenarios, such as Gemini Nano, Pro and Ultra [37]. Given a large pre-trained model, common solutions to deploy it to multiple application scenarios with different resource constraints include additional weight pruning [28, 42, 51], knowledge distillation [22, 24], or even re-training a small network from scratch, which" + } + ], + [ + { + "type": "page_number", + "bbox": [ + 0.218, + 0.116, + 0.23, + 0.127 + ], + "angle": 0, + "content": "2" + }, + { + "type": "header", + "bbox": [ + 0.272, + 0.115, + 0.375, + 0.129 + ], + "angle": 0, + "content": "Y. Zhang et al." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.147, + 0.788, + 0.195 + ], + "angle": 0, + "content": "all require substantial development efforts. Consequently, this issue prompts a critical question: is it possible to pre-train once to produce multiple models with different sizes simultaneously, each delivering sufficiently good representations?" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.207, + 0.79, + 0.511 + ], + "angle": 0, + "content": "To address this challenge, we introduce a new paradigm of self-supervised learning, called POA (Pre-training Once for All). POA is built upon the prevalent teacher-student self-distillation framework [5,33,55], with an additional innovative elastic student branch. The elastic student branch embeds a series of sub-networks through parameter sharing, upon the observation that smaller-sized models are sub-networks of a larger-sized one for modern network structures [14,21,30]. Moreover, the parameters of this branch are shared with the original, or intact student. At each pre-training step, we randomly sample a subset of parameters from the intact student to form the corresponding elastic student. Both the original intact student and the elastic student are trained to emulate the output of the teacher network. The teacher itself is continually refined via an exponential moving average (EMA) of the student's parameters, including the sampled elastic student. The elastic student facilitates effective and efficient pre-training on different subsets of parameters, leading to the successful extraction of high-performance sub-networks from the pre-trained teacher for subsequent downstream scenarios. It also acts as a form of training regularization by enforcing the outputs to match between the teacher and various sub-networks, contributing to a stable training process. Additionally, by serving as an ensemble of different sub-networks across different pre-training steps, the elastic student improves the representation learning [44]." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.523, + 0.496, + 0.795 + ], + "angle": 0, + "content": "To the best of our knowledge, POA represents the first self-supervised learning method capable of training multiple-sized models concurrently, each obtaining high-quality representations for different resource constraints without further pre-training. Figure 1 displays the k-nearest neighbors (k-NN) evaluation results for 143 sub-networks extracted from a ViT-L model [14] pre-trained by POA. By choosing different elastic widths and depths, which will be explained in Sec. 3.1, the pre-trained teacher model can generate a sufficient number of candidate sub-networks for the selection of the suitable model tailored for downstream applications according to the" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.796, + 0.788, + 0.842 + ], + "angle": 0, + "content": "computational resources available. Notably, each sub-network is well-trained and exhibits superior performance, thanks to our meticulous design on same-view distillation, as validated in Sec. 4.4. In particular, the ViT-S, ViT-B and ViT-" + }, + { + "type": "image", + "bbox": [ + 0.505, + 0.528, + 0.792, + 0.729 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.503, + 0.741, + 0.79, + 0.784 + ], + "angle": 0, + "content": "Fig.1: The k-NN evaluation accuracy of 143 elastic ViTs derived from the ViT-L/16 teacher model pre-trained with POA." + } + ], + [ + { + "type": "header", + "bbox": [ + 0.695, + 0.115, + 0.732, + 0.127 + ], + "angle": 0, + "content": "POA" + }, + { + "type": "page_number", + "bbox": [ + 0.775, + 0.116, + 0.787, + 0.127 + ], + "angle": 0, + "content": "3" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.147, + 0.784, + 0.177 + ], + "angle": 0, + "content": "L models set new benchmarks, achieving the state-of-the-art (SOTA) results compared with those pre-trained by existing methods [33, 34, 55]." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.178, + 0.788, + 0.298 + ], + "angle": 0, + "content": "To rigorously evaluate the efficacy of our approach, we conduct extensive experiments using three widely-used backbone architectures, i.e., ViT [14], Swin Transformer [30], and ResNet [21]. Each backbone is pre-trained on ImageNet-1K dataset and assessed using k-NN and linear probing classification evaluation, as well as downstream dense prediction tasks such as object detection and semantic segmentation. Our method achieves SOTA accuracy across various model sizes with a single pre-training session. The technical contributions of this work are summarized as follows:" + }, + { + "type": "text", + "bbox": [ + 0.226, + 0.308, + 0.784, + 0.395 + ], + "angle": 0, + "content": "- To the best of our knowledge, POA is the first pre-training paradigm that integrates unsupervised representation learning and once-for-all model generation within a single pre-training session. It tackles the Pre-training-Once-for-All challenge which has been seldomly explored by the community but is of great practical significance for real-world deployment that usually requires a suite of models." + }, + { + "type": "text", + "bbox": [ + 0.226, + 0.399, + 0.784, + 0.486 + ], + "angle": 0, + "content": "- We propose a novel and elegant component called Elastic Student, featuring a range of elastic operators that enable POA to be compatible with popular backbone structures including ViT, Swin Transformer and ResNet. It provides the capability to generate models of diverse sizes. Furthermore, it serves as a model ensemble to smooth training process and improve learned representations." + }, + { + "type": "text", + "bbox": [ + 0.226, + 0.489, + 0.784, + 0.576 + ], + "angle": 0, + "content": "- Through thorough assessments using k-NN, linear probing and downstream dense task evaluation, our approach exhibits superior performance over existing state-of-the-art pre-training methods across multiple metrics. Moreover, we compare our POA with Self-Supervised Distillation (SEED) [15], a knowledge distillation method designed especially for self-supervised learning, further validating POA's effectiveness." + }, + { + "type": "title", + "bbox": [ + 0.216, + 0.602, + 0.388, + 0.618 + ], + "angle": 0, + "content": "2 Related Work" + }, + { + "type": "title", + "bbox": [ + 0.215, + 0.634, + 0.466, + 0.649 + ], + "angle": 0, + "content": "2.1 Self-supervised Learning" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.659, + 0.784, + 0.748 + ], + "angle": 0, + "content": "Self-supervised learning (SSL) frameworks commonly fall into two categories, generative and discriminative SSL. Most generative SSL approaches [7,9,18,23, 25,26,38,48,52] focus on learning image representations directly in pixel space. On the other hand, discriminative SSL methods [4,6,19,27,36,41,43,47] aim to learn representations by pulling those of different views of the same image closer while separating the representations of views from different images." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.75, + 0.785, + 0.839 + ], + "angle": 0, + "content": "Contrastive learning (CL) with the InfoNCE loss [32] has emerged as a popular approach for discriminative SSL methods, attracting significant research interests in recent years. Although CL methods prevent the collapse of network representations through the use of negative samples, they still suffer from the dimensional collapse, where representations tend to collapse into a low-dimensional manifold. Grill et al. introduced a distillation-based asymmetric" + } + ], + [ + { + "type": "page_number", + "bbox": [ + 0.218, + 0.116, + 0.23, + 0.127 + ], + "angle": 0, + "content": "4" + }, + { + "type": "header", + "bbox": [ + 0.272, + 0.115, + 0.374, + 0.129 + ], + "angle": 0, + "content": "Y. Zhang et al." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.147, + 0.788, + 0.298 + ], + "angle": 0, + "content": "framework known as BYOL [17], which circumvents collapse without self-labeling or contrastive loss relying on negative samples. Following this work, distillation-based frameworks have become a prevailing trend in self-supervised learning. These frameworks [8, 16, 47] often merge with others to enhance overall performance. DINO [5] presented a simple self-distillation framework and has demonstrated impressive results in ViT pre-training. Subsequent works [33, 55] further improved the pre-training performance via masked token or the novel clustering design. Given the substantial benefits of distillation-based methods over other SSL techniques, we have developed our POA SSL framework upon these successful methodologies." + }, + { + "type": "title", + "bbox": [ + 0.216, + 0.321, + 0.453, + 0.337 + ], + "angle": 0, + "content": "2.2 Dynamic Architecture." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.346, + 0.789, + 0.725 + ], + "angle": 0, + "content": "Recently, Chen et al. proposed AutoFormer [45], which trained a supernet to support the effective search of optimal sub-network under some specific parameter number constraints. On the basis of [45], MaskTAS [50] introduced a self-supervised transformer architecture search method. Cai et al. [2] trained a network that accommodates various architectural configurations to reduce the training expense. Their methodologies enable the extraction of a specialized sub-network from the trained supernet. The design of the elastic student in our POA SSL is inspired by the weight-sharing strategy employed in these neural architecture search (NAS) methods. However, our implementation differs significantly due to the distinct purpose from NAS. Specifically, NAS aims to discover the optimal architecture within certain parameter constraints, which typically involves a huge number of sub-networks (more than \\(10^{16}\\)) in the search space. Given the limitations on the number of training iterations and the network parameter capacity, it is challenging to ensure high performance across all sub-networks. After training the supernet, NAS requires a subsequent phase of searching and re-training to finalize the output model. In contrast, our POA SSL defines a sufficiently large yet compact set of sub-networks with elastic depths and widths for the purpose of pre-training models of various sizes via a single training session. Additional design on the same-view distillation guarantees that all elastic sub-networks within our framework are adequately and efficiently trained. As a result, our POA can readily extract a range of sub-networks from the teacher model without the need for extra pre-training, facilitating an easy selection of a suitable sub-network for different computational contexts. The design of elastic student is somewhat akin to the supervised learning method, Cosub [40]. The main difference is that Cosub simply skips blocks, making only the depth elastic." + }, + { + "type": "title", + "bbox": [ + 0.216, + 0.748, + 0.661, + 0.766 + ], + "angle": 0, + "content": "3 POA Self-supervised Learning Framework" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.78, + 0.788, + 0.842 + ], + "angle": 0, + "content": "Our primary goal is to pre-train models of multiple sizes via a single self-supervised pre-training session. To this end, we propose a novel SSL framework named POA, inspired by the latest progress in self-distillation techniques [5, 33, 55]. The architecture of POA is illustrated in Figure 2, encompassing a" + } + ], + [ + { + "type": "header", + "bbox": [ + 0.695, + 0.115, + 0.732, + 0.127 + ], + "angle": 0, + "content": "POA" + }, + { + "type": "page_number", + "bbox": [ + 0.775, + 0.116, + 0.785, + 0.127 + ], + "angle": 0, + "content": "5" + }, + { + "type": "image", + "bbox": [ + 0.222, + 0.144, + 0.784, + 0.296 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.214, + 0.308, + 0.785, + 0.391 + ], + "angle": 0, + "content": "Fig. 2: Overview of the POA SSL: Given an image \\( x \\), two augmented views \\( x_{a} \\) and \\( x_{b} \\) are generated. These views are input into three branches: a teacher, an intact student, and an elastic student, the latter being derived from the intact student. POA optimizes distillation losses in a twofold manner: the intact and the elastic students are distilled from the teacher using the cross-view data respectively, and additionally, the elastic student is distilled from the intact student using the same-view data." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.397, + 0.788, + 0.624 + ], + "angle": 0, + "content": "teacher, an intact student, an elastic student and their corresponding heads. The teacher is updated with an EMA of the students. The elastic student is a derivative of the intact one, with both the backbones and heads' parameters shared. We leverage distillation in two aspects: both the intact and elastic students are distilled from the teacher using different views of the same image, while the elastic student additionally learns from the intact student using the identical views. The cross-view distillation works as a form of representation learning, as introduced in [5, 33, 55]. Notably, in addition to the regular EMA update with only the intact student as [5, 33, 55], the elastic student provides a randomly-sampled sub-network at each pre-training step to participate in the teacher's EMA refinement. This procedure actually simulates an ensemble of multiple sub-networks, which is also proven to be beneficial in the realm of supervised learning [44]. The same-view distillation is a standard knowledge distillation between the intact and elastic students, promoting the quality of the elastic one. Experiments in Sec. 4.4 validate our design in details." + }, + { + "type": "title", + "bbox": [ + 0.216, + 0.647, + 0.473, + 0.663 + ], + "angle": 0, + "content": "3.1 Design of Elastic Student" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.674, + 0.788, + 0.794 + ], + "angle": 0, + "content": "The elastic student is a sub-network whose parameters are extracted from the intact student. In the context of a transformer backbone, the width refers to the dimensionality of the tokens, whereas for a convolutional backbone, the width indicates the number of channels. We denote the depth as the number of basic blocks in either a transformer or a convolutional network. Given the value of the width and depth, it yields a certain network structure. For simplicity, we focus on detailing the elastic design of the ViT in this section. For the similar elastic design of the Swin Transformer and ResNet, please refer to Appendix A." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.795, + 0.787, + 0.842 + ], + "angle": 0, + "content": "A basic block of ViT mainly consists of a multi-head self-attention (MSA) module and a multi-layer perceptron (MLP) module. Layer Normalization (LN) [1] is applied before each module, with residual connections after each module." + } + ], + [ + { + "type": "page_number", + "bbox": [ + 0.218, + 0.116, + 0.23, + 0.127 + ], + "angle": 0, + "content": "6" + }, + { + "type": "header", + "bbox": [ + 0.272, + 0.115, + 0.375, + 0.129 + ], + "angle": 0, + "content": "Y. Zhang et al." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.147, + 0.788, + 0.207 + ], + "angle": 0, + "content": "As shown in the left part of Figure 3, an elastic block refers to a stack of elastic MSA, MLP, and LN after adjusting the width in the original basic block in ViT. In our approach, the elastic student branch is constructed by assembling a specific number of these elastic blocks at each training iteration." + }, + { + "type": "image", + "bbox": [ + 0.257, + 0.209, + 0.744, + 0.349 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.214, + 0.358, + 0.786, + 0.389 + ], + "angle": 0, + "content": "Fig. 3: Illustration of the elastic MSA in an elastic ViT block. To be concise, we simply exclude the projection layers that correspond to \\( K \\) and \\( V \\) in each head." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.408, + 0.789, + 0.56 + ], + "angle": 0, + "content": "**Elastic MSA** An original, or intact MSA module consists of three major components, i.e., input projection layers, an operator contains attention & concatenation, and an output projection layer. We define a projection layer as \\((w^{*}, b^{*})\\), where \\(w^{*}\\) is the linear transformation weight, \\(b^{*}\\) denotes the corresponding bias, and * indicates the name of the layer. As shown in the right part of Figure 3, given a token dimension of \\(D_{max} = N_h \\cdot D_h\\), where \\(N_h\\) is the number of attention heads and \\(D_h\\) is the head dimension, an input sequence \\(z \\in \\mathbb{R}^{T \\times D_{max}}\\) with length \\(T\\) is initially projected to form the queries \\(Q \\in \\mathbb{R}^{T \\times D_h}\\), keys \\(K \\in \\mathbb{R}^{T \\times D_h}\\), and values \\(V \\in \\mathbb{R}^{T \\times D_h}\\) by the intact MSA. To generate elastic MSA, we define \\(\\mathrm{M} + 1\\) elastic widths, including \\(D_{max}\\), spaced at intervals of \\(D_h\\) as follows:" + }, + { + "type": "equation", + "bbox": [ + 0.315, + 0.572, + 0.787, + 0.588 + ], + "angle": 0, + "content": "\\[\nD _ {i} = \\left(N _ {h} - i\\right) \\cdot D _ {h}, \\quad \\forall i \\in \\{0, 1, \\dots , M \\}, \\quad M < N _ {h}. \\tag {1}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.599, + 0.788, + 0.721 + ], + "angle": 0, + "content": "For each elastic width \\(D_{i}\\), the weight \\(w_{i}^{a1} \\in \\mathbb{R}^{D_{h} \\times D_{i}}\\) and bias \\(b_{i}^{a1} \\in \\mathbb{R}^{D_{h}}\\) that generate \\(Q\\), \\(K\\), and \\(V\\) of each head are extracted from the corresponding input projection layer (\\(w^{a1}\\), \\(b^{a1}\\)) in the intact MSA, as \\(w_{i}^{a1} = w^{a1}[:,: D_{i}] \\cdot \\alpha_{i}\\) and \\(b_{i}^{a1} = b^{a1}\\). Here, \\(\\alpha_{i}\\) represents the scaling factor to cope with the reduction of the input dimension, calculated as \\(\\alpha_{i} = D_{max} / D_{i}\\). As the reduction of width, the number of attention heads in the elastic MAS is naturally reduced to \\(N_{h} - i\\). Similarly, for the output projection layer (\\(w^{a2}\\), \\(b^{a2}\\)), the weight \\(w_{i}^{a2} \\in \\mathbb{R}^{D_{i} \\times D_{i}}\\) and bias \\(b_{i}^{a2} \\in \\mathbb{R}^{D_{i}}\\) are extracted as:" + }, + { + "type": "equation", + "bbox": [ + 0.344, + 0.731, + 0.787, + 0.751 + ], + "angle": 0, + "content": "\\[\nw _ {i} ^ {a 2} = w ^ {a 2} [: D _ {i},: D _ {i} ] \\cdot \\alpha_ {i} \\quad b _ {i} ^ {a 2} = b ^ {a 2} [: D _ {i} ]. \\tag {2}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.765, + 0.788, + 0.842 + ], + "angle": 0, + "content": "**Elastic MLP** An original, or intact MLP module in the ViT block contains two projection layers. The first layer (\\(w^{m1}, b^{m1}\\)) expands the dimension of embedding by a factor of \\(s\\), which is generally set to 4 in the ViT structure. The second layer (\\(w^{m2}, b^{m2}\\)) then projects it back to the original dimension. The parameters for both layers of the elastic MLP are extracted in a manner analogous to that" + } + ], + [ + { + "type": "header", + "bbox": [ + 0.695, + 0.115, + 0.732, + 0.127 + ], + "angle": 0, + "content": "POA" + }, + { + "type": "page_number", + "bbox": [ + 0.776, + 0.116, + 0.786, + 0.127 + ], + "angle": 0, + "content": "7" + }, + { + "type": "text", + "bbox": [ + 0.216, + 0.147, + 0.468, + 0.162 + ], + "angle": 0, + "content": "described in Equation 2 as follows:" + }, + { + "type": "equation", + "bbox": [ + 0.323, + 0.174, + 0.682, + 0.19 + ], + "angle": 0, + "content": "\\[\nw _ {i} ^ {m 1} = w ^ {m 1} \\left[ \\colon D _ {i} \\cdot s, \\colon D _ {i} \\right] \\cdot \\alpha_ {i} \\quad b _ {i} ^ {m 1} = b ^ {m 1} \\left[ \\colon D _ {i} \\cdot s \\right]\n\\]" + }, + { + "type": "equation", + "bbox": [ + 0.322, + 0.186, + 0.785, + 0.211 + ], + "angle": 0, + "content": "\\[\nw _ {i} ^ {m 2} = w ^ {m 2} [: D _ {i},: D _ {i} \\cdot s ] \\cdot \\alpha_ {i} \\quad b _ {i} ^ {m 2} = b ^ {m 2} [: D _ {i} ]. \\tag {3}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.216, + 0.223, + 0.784, + 0.253 + ], + "angle": 0, + "content": "**Elastic LN** For elastic LN, we directly use the first \\( D_{i} \\) elements of the parameter inside the original LN, akin to the bias extraction in Equation 2." + }, + { + "type": "text", + "bbox": [ + 0.216, + 0.253, + 0.785, + 0.331 + ], + "angle": 0, + "content": "To create a sub-network with \\( L_{i} \\) elastic blocks from the intact ViT comprising \\( L_{max} \\) blocks, we introduce a set of \\( N + 1 \\) elastic depths, defined as \\( L_{i} = L_{max} - i \\), \\( \\forall i \\in \\{0,1,\\dots,N\\} \\), \\( N < L_{max} \\). For a specific depth \\( L_{i} \\), we select the corresponding blocks based on their block IDs at equal intervals. Each block ID \\( BID_{j}^{L_{i}} \\) that is activated for depth \\( L_{i} \\) can be formulated as:" + }, + { + "type": "equation", + "bbox": [ + 0.317, + 0.341, + 0.785, + 0.373 + ], + "angle": 0, + "content": "\\[\nB I D _ {j} ^ {L _ {i}} = \\left\\lfloor \\frac {\\left(L _ {\\max } - 1\\right) \\cdot j}{L _ {i} - 1} \\right\\rfloor , \\quad \\forall j \\in \\{0, 1, \\dots , L _ {i} - 1 \\}. \\tag {4}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.215, + 0.378, + 0.784, + 0.453 + ], + "angle": 0, + "content": "Consequently, by combining elastic widths and depths, we can generate a total of \\((N + 1) \\cdot (M + 1)\\) distinct sub-networks. For instance, by setting the elastic width to 384 and the elastic depth to 12, we can directly extract a ViT-S from the intact network such as ViT-L. During each iteration of the pre-training, we randomly select one of these sub-networks to serve as the elastic student branch." + }, + { + "type": "title", + "bbox": [ + 0.216, + 0.475, + 0.483, + 0.489 + ], + "angle": 0, + "content": "3.2 Distillation between Views" + }, + { + "type": "text", + "bbox": [ + 0.215, + 0.5, + 0.785, + 0.636 + ], + "angle": 0, + "content": "POA performs distillation across its three branches accordingly. Given a pair of globally augmented views of an input image \\( x \\), denoted as \\( x_{a} \\) and \\( x_{b} \\), the teacher encoder \\( E_{T} \\) extracts the feature \\( Z_{a} = E_{T}(x_{a}) \\) using \\( x_{a} \\) as input. Simultaneously, \\( x_{b} \\) is fed into both the intact student encoder \\( E_{IS} \\) and the elastic student encoder \\( E_{ES} \\), producing the features \\( Z_{b1} = E_{IS}(x_{b}) \\) and \\( Z_{b2} = E_{ES}(x_{b}) \\), respectively. The feature output from the teacher encoder, \\( Z_{a} \\), is then processed by the teacher head \\( H_{T} \\), followed by centering with the Sinkhorn-Knopp (SK) [11] algorithm and normalization using a temperature-scaled softmax to generate the probability \\( p_{a} \\) as follows:" + }, + { + "type": "equation", + "bbox": [ + 0.228, + 0.66, + 0.785, + 0.697 + ], + "angle": 0, + "content": "\\[\nl _ {a} = S K \\left(H _ {T} \\left(Z _ {a}\\right)\\right), l _ {a} \\in \\mathbb {R} ^ {P} \\quad p _ {a} ^ {i} = \\frac {\\exp \\left(l _ {a} ^ {i} / \\tau\\right)}{\\sum_ {k = 0} ^ {P - 1} \\exp \\left(l _ {a} ^ {k} / \\tau\\right)}, \\forall i \\in \\{0, \\dots , P - 1 \\}, \\tag {5}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.215, + 0.705, + 0.786, + 0.842 + ], + "angle": 0, + "content": "where \\( P \\) is the number of prototypes and \\( \\tau > 0 \\) is a temperature parameter. In a similar fashion, we compute the probabilities \\( p_{b1}^{i} \\) and \\( p_{b2}^{i} \\) for the intact and elastic student encoders, respectively, by processing the outputs through the student heads \\( H_{IS} \\) and \\( H_{ES} \\). These outputs are then passed through a temperature-scaled softmax function using a temperature parameter \\( \\tau' \\) tailored for the student. It should be noted that \\( H_{IS} \\) and \\( H_{ES} \\) share the identical parameters, except that the first projection layer of \\( H_{ES} \\) is adapted similarly as in Equation 2 to conform to align with the corresponding dimensionality. For simplicity, we omit the explicit expressions for \\( p_{b1}^{i} \\) and \\( p_{b2}^{i} \\), as they follow a similar" + } + ], + [ + { + "type": "page_number", + "bbox": [ + 0.218, + 0.116, + 0.23, + 0.127 + ], + "angle": 0, + "content": "8" + }, + { + "type": "header", + "bbox": [ + 0.272, + 0.115, + 0.375, + 0.129 + ], + "angle": 0, + "content": "Y. Zhang et al." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.147, + 0.784, + 0.178 + ], + "angle": 0, + "content": "calculation to Equation 5. For the intact student branch, we perform distillation from the teacher using the cross-view data as follows:" + }, + { + "type": "equation", + "bbox": [ + 0.433, + 0.193, + 0.786, + 0.21 + ], + "angle": 0, + "content": "\\[\n\\mathcal {L} _ {I S} ^ {g} = - p _ {a} \\log \\left(p _ {b 1}\\right). \\tag {6}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.214, + 0.785, + 0.334 + ], + "angle": 0, + "content": "The elastic student branch plays a pivotal role in our POA framework. To ensure adequate training of this branch, we employ a dual distillation from the teacher and intact student branch. The first distillation involves the teacher model, which utilizes the cross-view data to guide the learning of representations. The second is a distillation process with the intact student model, which uses the same-view data. This same-view distillation is responsible for transferring the representations learned by the intact student to the elastic student branch. The loss functions for this dual distillation process are formulated as follows:" + }, + { + "type": "equation", + "bbox": [ + 0.34, + 0.349, + 0.786, + 0.368 + ], + "angle": 0, + "content": "\\[\n\\mathcal {L} _ {E S 1} ^ {g} = - p _ {a} \\log (p _ {b 2}), \\quad \\mathcal {L} _ {E S 2} ^ {g} = - p _ {b 1} \\log (p _ {b 2}). \\qquad (7)\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.371, + 0.784, + 0.402 + ], + "angle": 0, + "content": "Note that in both loss functions, we sum over all prototypes to compute the cross-entropy loss between the respective probability distributions." + }, + { + "type": "title", + "bbox": [ + 0.216, + 0.422, + 0.432, + 0.436 + ], + "angle": 0, + "content": "3.3 Overall Loss of POA" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.445, + 0.784, + 0.535 + ], + "angle": 0, + "content": "Following the SSL methods such as [5,33,55], we employ a multi-crop strategy [4] to create various distorted views from a single image. Apart from the two global views previously mentioned, we also generate \\( v \\) local views with lower resolution \\( x_{l_1}, x_{l_2}, \\ldots, x_{l_v} \\). These local views are processed by both students to promote local-to-global correspondence. The local distillation losses for the intact and elastic students are computed as follows:" + }, + { + "type": "equation", + "bbox": [ + 0.41, + 0.548, + 0.786, + 0.587 + ], + "angle": 0, + "content": "\\[\n\\mathcal {L} _ {I S} ^ {l} = - \\frac {1}{v} \\sum_ {i = 1} ^ {v} p _ {a} \\log \\left(p _ {l _ {i} 1}\\right), \\tag {8}\n\\]" + }, + { + "type": "equation", + "bbox": [ + 0.293, + 0.594, + 0.786, + 0.632 + ], + "angle": 0, + "content": "\\[\n\\mathcal {L} _ {E S 1} ^ {l} = - \\frac {1}{v} \\sum_ {i = 1} ^ {v} p _ {a} \\log \\left(p _ {l _ {i} 2}\\right), \\quad \\mathcal {L} _ {E S 2} ^ {l} = - \\frac {1}{v} \\sum_ {i = 1} ^ {v} p _ {l _ {i} 1} \\log \\left(p _ {l _ {i} 2}\\right), \\tag {9}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.636, + 0.784, + 0.681 + ], + "angle": 0, + "content": "where \\( p_{l_{i1}} \\) and \\( p_{l_{i2}} \\) are the probability produced by the intact and elastic student branches for the local view \\( l_i \\), respectively. The total distillation loss of the intact and elastic student is calculated by summing them with the factor \\( \\lambda \\):" + }, + { + "type": "equation", + "bbox": [ + 0.26, + 0.69, + 0.786, + 0.726 + ], + "angle": 0, + "content": "\\[\n\\begin{array}{l} \\mathcal {L} _ {\\mathcal {S}} = \\lambda \\left(\\mathcal {L} _ {I S} ^ {g} + \\mathcal {L} _ {I S} ^ {l}\\right) + (1 - \\lambda) \\left(\\left(\\mathcal {L} _ {E S 1} ^ {g} + \\mathcal {L} _ {E S 1} ^ {l}\\right) + \\left(\\mathcal {L} _ {E S 2} ^ {g} + \\mathcal {L} _ {E S 2} ^ {l}\\right)\\right) \\tag {10} \\\\ = \\lambda \\mathcal {L} _ {I S} + (1 - \\lambda) \\left(\\mathcal {L} _ {E S 1} + \\mathcal {L} _ {E S 2}\\right). \\\\ \\end{array}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.735, + 0.785, + 0.843 + ], + "angle": 0, + "content": "To ensure sufficient training of each sub-network from the elastic student, we introduce multiple projection heads (MPH) positioned after the backbone network. Each projection head has exactly the same structure, except for a different number of prototypes. For each head, the distillation loss \\(\\mathcal{L}_{\\mathcal{S}_i}\\) for both the intact and elastic student is calculated with Equation. 10. Finally, the overall loss function in the POA framework with \\(H\\) projection heads is formulated as: \\(\\mathcal{L} = \\frac{1}{H}\\sum_{i = 1}^{H}\\mathcal{L}_{\\mathcal{S}_i}\\)." + } + ], + [ + { + "type": "header", + "bbox": [ + 0.695, + 0.115, + 0.732, + 0.127 + ], + "angle": 0, + "content": "POA" + }, + { + "type": "page_number", + "bbox": [ + 0.776, + 0.117, + 0.786, + 0.127 + ], + "angle": 0, + "content": "9" + }, + { + "type": "title", + "bbox": [ + 0.217, + 0.146, + 0.375, + 0.163 + ], + "angle": 0, + "content": "4 Experiments" + }, + { + "type": "title", + "bbox": [ + 0.217, + 0.177, + 0.456, + 0.193 + ], + "angle": 0, + "content": "4.1 Implementation Details" + }, + { + "type": "text", + "bbox": [ + 0.219, + 0.202, + 0.788, + 0.384 + ], + "angle": 0, + "content": "Backbones. We have trained our POA using ViT, Swin Transformer and ResNet backbones, respectively. For the ViT, we configure the patch size to 16 and the dimension of each head in the MSA to 64. This aligns with the configurations typically used in existing SSL methods [5,34,35,55]. For the smallest and largest elastic networks of ViT, we choose the ViT-S and ViT-L, respectively. This leads to 11 elastic widths and 13 elastic depths, yielding a total of 143 ViT subnetworks. In the case of the Swin Transformer, we set the smallest and largest elastic networks as Swin-T and Swin-B, respectively. This configuration yields a total of 39 Swin sub-networks by multiplying the number of widths and depths as \\(3 \\times 13\\). For the ResNet, we establish the smallest and largest elastic network configurations as ResNet-50 and ResNet-152. Consequently, this setup accounts for a total of 465 ResNet sub-networks, which is the product of \\(3 \\times 155\\)." + }, + { + "type": "text", + "bbox": [ + 0.219, + 0.399, + 0.788, + 0.566 + ], + "angle": 0, + "content": "Pre-Training Setup. We pre-train all models on the ImageNet-1K dataset [12] without the labels. The process employs the AdamW optimizer [31] with a batch size of 1600, which is distributed across 32 A100 GPUs when employing a ViT backbone. We adopt a learning rate schedule that begins with a linear warm-up during the first 10 epochs, reaching a base value that is scaled proportionally to the total batch size: \\(\\mathrm{lr} = 0.004 \\times \\sqrt{\\text{batch size} / 1024}\\), in line with [33]. Following this warm-up period, the learning rate is decayed with a cosine schedule. Similarly, the weight decay follows a cosine schedule, starting at 0.04 and increasing to 0.4. The student network's temperature \\(\\tau'\\) is fixed at 0.1, whereas the teacher temperature \\(\\tau\\) changes with a linear warm-up from 0.04 to 0.07 over the first 30 epochs. For further details, please refer to Appendix C." + }, + { + "type": "title", + "bbox": [ + 0.217, + 0.588, + 0.509, + 0.604 + ], + "angle": 0, + "content": "4.2 Experiments on ImageNet-1K" + }, + { + "type": "text", + "bbox": [ + 0.219, + 0.614, + 0.788, + 0.734 + ], + "angle": 0, + "content": "After unsupervised pre-training, we assess the model's performance using two widely-recognized evaluation protocols in SSL domain on ImageNet-1K dataset, i.e., k-NN and linear probing. To ensure a fair comparison between SSL methods that employ different numbers of crop views for data augmentation, Zhou et al. [55] introduced the effective training epoch as a measure to quantify the extent of a method's pre-training. We report the effective training epochs of the SSL methods for comparison. For additional evaluations including fine-tuning, semi-supervised and unsupervised learning, please refer to Appendix D." + }, + { + "type": "text", + "bbox": [ + 0.217, + 0.75, + 0.788, + 0.84 + ], + "angle": 0, + "content": "k-NN and Linear Probing. To assess the quality of pre-trained features, we employ a k-NN classifier and a linear classifier on the frozen representations. For both the k-NN and linear probing (LP) evaluation, we follow the evaluation protocols established in [5, 33, 55]. The performance of our method when being trained using ResNet, Swin Transformer and ViT backbones is reported in Table 1. Our POA SSL achieves the SOTA k-NN accuracy of \\(82.3\\%\\) and LP accuracy" + } + ], + [ + { + "type": "page_number", + "bbox": [ + 0.218, + 0.116, + 0.236, + 0.127 + ], + "angle": 0, + "content": "10" + }, + { + "type": "header", + "bbox": [ + 0.272, + 0.115, + 0.375, + 0.129 + ], + "angle": 0, + "content": "Y. Zhang et al." + }, + { + "type": "table_caption", + "bbox": [ + 0.214, + 0.145, + 0.788, + 0.215 + ], + "angle": 0, + "content": "Table 1: Comparison results of k-NN and linear probing classification accuracy (\\%) on the ImageNet-1K dataset. \"Param.\" refers to the number of parameters. \"Epo.\" represents the number of effective training epochs following [55]. \"/16\" denotes patch size of 16. \"/W7\" means the window size of 7. \"**\" indicates our implementation based on official codebase. \"+\" denotes reproduced results using the released code." + }, + { + "type": "table", + "bbox": [ + 0.216, + 0.226, + 0.778, + 0.464 + ], + "angle": 0, + "content": "
MethodEpo.k-NNLPMethodEpo.k-NNLPMethodEpo.k-NNLP
ResNet-50(Param. 23M)Swin-T/W7(Param. 28M)ViT-S/16(Param. 21M)
VICReg2000-73.2SMoG1200-77.7DINO320074.577.0
SwAV240065.775.3iBOT120075.378.6iBOT320075.277.9
DINO320067.575.3DINOv2*120075.478.0ENT320075.377.7
UniGrad2400-75.5EsViT120075.778.1POA076.877.6
SCFS320068.575.7POA077.578.9ViT-B/16(Param. 85M)
ReLICv24000-77.1Swin-S/W7(Param. 49M)DINO160076.178.2
POA073.476.9DINOv2*120076.179.8ENT160077.179.1
ResNet-101(Param. 41M)EsViT120077.779.5iBOT160077.179.5
ReLICv24000-78.7POA079.381.3POA080.981.7
POA075.779.1Swin-B/W7(Param. 87M)ViT-L/16(Param. 307M)
ResNet-152(Param. 56M)DINOv2*120076.980.9iBOT120078.081.0
ReLICv24000-79.3EsViT120078.980.4DINOv2†120082.083.3
POA240076.479.9POA120079.682.0POA120082.383.6
" + }, + { + "type": "table_footnote", + "bbox": [ + 0.245, + 0.465, + 0.373, + 0.478 + ], + "angle": 0, + "content": "(a) ResNet backbone." + }, + { + "type": "table_footnote", + "bbox": [ + 0.438, + 0.465, + 0.553, + 0.477 + ], + "angle": 0, + "content": "(b) Swin backbone." + }, + { + "type": "table_footnote", + "bbox": [ + 0.627, + 0.465, + 0.736, + 0.477 + ], + "angle": 0, + "content": "(c) ViT backbone." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.494, + 0.789, + 0.707 + ], + "angle": 0, + "content": "of \\(83.6\\%\\) when using the ViT-L/16 backbone. By employing the sub-network extraction approach outlined in Section 3.1, we derive the sub-networks ViT-S/16 and ViT-B/16 from the teacher ViT-L/16 without any additional pre-training. Thus, the number of effective training epochs of them is 0. Notably, the extracted ViT-S/16 achieves the SOTA k-NN accuracy of \\(76.8\\%\\) and LP accuracy on par with the previous SOTA of \\(77.9\\%\\) reported by iBOT [55]. Our derived ViT-B/16 model sets new benchmarks for both k-NN and LP, with accuracy of \\(80.9\\%\\) and \\(81.7\\%\\), respectively. For the models using Swin Transformer and ResNet backbones, POA also reaches SOTA performance in k-NN and LP accuracy. The only exception is the LP accuracy of ResNet-50, which is competitive with that of ReLICv2 [39], despite the latter being trained over a much longer period (4000 epochs). The superior performance achieved across a range of backbone architectures confirms the effectiveness and versatility of our method. For additional detailed comparison with 27 existing methods, please see Appendix D.1." + }, + { + "type": "title", + "bbox": [ + 0.216, + 0.726, + 0.595, + 0.74 + ], + "angle": 0, + "content": "4.3 Evaluation on Downstream Visual Tasks" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.749, + 0.788, + 0.841 + ], + "angle": 0, + "content": "Object Detection and Instance Segmentation on COCO Dataset. For a fair comparison, we utilize the Cascade Mask R-CNN framework [3,20], which generates both bounding boxes and instance masks, in line with previous approach [55], on the COCO dataset [29]. We benchmark our results against existing SSL methods that generate pre-trained ViT-S/16 and ViT-B/16 backbones. As shown in Table 2, POA boosts the bounding box average precision \\((\\mathrm{AP}^b)\\) for" + } + ], + [ + { + "type": "header", + "bbox": [ + 0.695, + 0.115, + 0.732, + 0.127 + ], + "angle": 0, + "content": "POA" + }, + { + "type": "page_number", + "bbox": [ + 0.769, + 0.116, + 0.785, + 0.127 + ], + "angle": 0, + "content": "11" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.146, + 0.788, + 0.193 + ], + "angle": 0, + "content": "ViT-S/16 from 49.4 to 50.6 and the mask average precision \\((\\mathrm{AP}^m)\\) from 42.6 to 43.8. When applied to ViT-B/16, POA attains an \\(\\mathrm{AP}^b\\) of 52.4 and an \\(\\mathrm{AP}^m\\) of 45.4, which represents a remarkable step forward over previous SOTA." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.204, + 0.789, + 0.415 + ], + "angle": 0, + "content": "Semantic Segmentation on ADE20K Dataset. For the semantic segmentation task, we primarily focus on two settings on the ADE20K dataset [54], following [55]. First, akin to the linear evaluation protocol in classification, we evaluate the quality of representations by keeping the patch features fixed and only fine-tuning one linear layer. This approach offers a clear comparison of representation quality. Second, we employ the UPerNet [46] as the task head and fine-tune the entire network. As depicted in Table 2, our POA significantly outperforms the supervised baseline using the ViT-S/16 backbone, achieving a substantial increase of 2.2 in mean Intersection over Union (mIoU), and surpassing iBOT by 1.3 mIoU. With the ViT-B/16 backbone, POA exceeds the previously best-performing method, iBOT, by 0.4 mIoU when utilizing UPerNet. Furthermore, under the assessment using solely a linear head, POA obtains an impressive improvement of 2.0 mIoU over iBOT as the performance is largely determined by the quality of the pre-trained representation." + }, + { + "type": "table_caption", + "bbox": [ + 0.214, + 0.416, + 0.787, + 0.446 + ], + "angle": 0, + "content": "Table 2: Evaluation results on downstream detection and segmentation tasks. \\(\\mathrm{Seg.}^{\\dagger}\\) indicates the use of a linear head for semantic segmentation." + }, + { + "type": "table", + "bbox": [ + 0.243, + 0.458, + 0.763, + 0.572 + ], + "angle": 0, + "content": "
MethodArch.Det. APbISeg. APmSeg. mIoUArch.Det. APbISeg. APmSeg.† mIoUSeg. mIoU
Sup.ViT-S/1646.240.144.5ViT-B/1649.843.235.446.6
BEiTViT-S/16---ViT-B/1650.143.527.445.8
DINOViT-S/16---ViT-B/1650.143.434.546.8
iBOTViT-S/1649.442.645.4ViT-B/1651.244.238.350.0
POAViT-S/1650.643.846.7ViT-B/1652.445.440.350.4
" + }, + { + "type": "title", + "bbox": [ + 0.215, + 0.595, + 0.479, + 0.609 + ], + "angle": 0, + "content": "4.4 Ablations and Discussions" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.616, + 0.788, + 0.753 + ], + "angle": 0, + "content": "In this section, we conduct an empirical analysis of POA using ViT as backbone. Our investigation includes the impact of the loss functions \\(\\mathcal{L}_{ES1}\\) and \\(\\mathcal{L}_{ES2}\\), in addition with the effectiveness of multiple projection heads. Moreover, we compare our POA with knowledge distillation techniques for self-supervised learning to demonstrate the advantages of combining once-for-all model generation with self-distillation in a unified framework. For more ablation studies, we direct readers to the Appendix D.8. In addition, we further contrast our POA with three variants tailored for elastic pre-training to showcase POA's superiority. Finally, we discuss how the elastic student facilitates the pre-training." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.765, + 0.788, + 0.84 + ], + "angle": 0, + "content": "Importance of Each Component. We evaluate the contributions of the components on a ViT backbone. Table 3 shows the performance of different component combinations. First, we note that employing multiple projection heads (MPH) enhances the learned representations for each elastic sub-network, particularly for smaller ones. The design consideration behind MPH is that for each" + } + ], + [ + { + "type": "page_number", + "bbox": [ + 0.218, + 0.116, + 0.236, + 0.127 + ], + "angle": 0, + "content": "12" + }, + { + "type": "header", + "bbox": [ + 0.272, + 0.115, + 0.375, + 0.129 + ], + "angle": 0, + "content": "Y. Zhang et al." + }, + { + "type": "table_caption", + "bbox": [ + 0.214, + 0.145, + 0.788, + 0.189 + ], + "angle": 0, + "content": "Table 3: Contributions of each component in POA framework. We conduct assessments with the k-NN and linear probing (LP) evaluations. MPH denotes the multiple projection heads with different numbers of prototypes." + }, + { + "type": "table", + "bbox": [ + 0.227, + 0.199, + 0.779, + 0.299 + ], + "angle": 0, + "content": "
MPHLES1LES2k-NNLP
ViT-S/16ViT-B/16ViT-L/16ViT-S/16ViT-B/16ViT-L/16
✓✓✓76.880.982.377.681.783.6
✓✓76.280.782.277.381.683.4
✓75.180.282.275.881.083.4
✓72.879.182.175.380.883.3
" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.304, + 0.789, + 0.53 + ], + "angle": 0, + "content": "pre-training iteration, the sub-network is chosen randomly, leading to a relatively insufficient optimization. MPH introduces different sets of prototypes, which act as multiple semantic spaces for representation learning, enabling the teacher to distill various aspects of learned knowledge into the sub-network. Furthermore, we ascertain that the same-view distillation loss \\(\\mathcal{L}_{ES2}\\) is crucial for the representation quality of elastic sub-networks. Omitting \\(\\mathcal{L}_{ES2}\\) causes a significant drop in k-NN accuracy, by \\(3.4\\%\\) for ViT-S/16 and \\(1.6\\%\\) for ViT-B/16. In addition, \\(\\mathcal{L}_{ES2}\\) is more important than cross-view distillation \\(\\mathcal{L}_{ES1}\\) in terms of sub-networks' representation quality. The underlying reason is that the cross-view distillation is for the unsupervised representation learning, while the same-view distillation improves the sub-networks by distilling previously learned representations from the intact student. Table 4 confirms that distillation from already-good representations is more effective than representation learning, especially for smaller networks. It also explains why employing three different views in our POA is unnecessary." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.532, + 0.79, + 0.631 + ], + "angle": 0, + "content": "Table 4: Comparison with knowledge distillation for self-supervised learning. The teacher name \"ViT-L/16-600\" denotes a teacher model (ViT-L/16) that has been pretrained with 600 effective epochs. The student name \"ViT-S/16-600\" refers to a student model (ViT-S/16) that has undergone distillation from the pre-trained teacher with 600 effective epochs. The number of the left side of \"\\(\\rightarrow\\)\" indicates the performance of the teacher model when pre-trained individually. The number of the right side denotes the performance achieved after distillation using the SEED [15]." + }, + { + "type": "table", + "bbox": [ + 0.216, + 0.642, + 0.79, + 0.791 + ], + "angle": 0, + "content": "
MethodTeacherStudentTotal Epochsk-NNLP
DINOv2ViT-S/16120072.273.1
DINOv2+SEEDViT-L/16-600ViT-S/16-600120081.3 → 74.082.4 → 75.2
DINOv2+SEEDViT-L/16-1200ViT-S/16-1200240082.0 → 75.583.3 → 76.2
POAViT-S/16120076.877.6
DINOv2ViT-B/16120077.478.5
DINOv2+SEEDViT-L/16-600ViT-B/16-600120081.3 → 78.882.4 → 80.0
DINOv2+SEEDViT-L/16-1200ViT-B/16-1200240082.0 → 79.783.3 → 80.9
POAViT-B/16120080.981.7
" + }, + { + "type": "text", + "bbox": [ + 0.216, + 0.809, + 0.789, + 0.842 + ], + "angle": 0, + "content": "Comparison with Knowledge Distillation. Knowledge distillation (KD) is a proven strategy for enhancing small networks by leveraging the knowledge of" + } + ], + [ + { + "type": "header", + "bbox": [ + 0.695, + 0.115, + 0.732, + 0.127 + ], + "angle": 0, + "content": "POA" + }, + { + "type": "page_number", + "bbox": [ + 0.769, + 0.116, + 0.786, + 0.127 + ], + "angle": 0, + "content": "13" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.147, + 0.788, + 0.374 + ], + "angle": 0, + "content": "a well-trained, larger network. We compare our method with a self-supervised KD method SEED [15] that employs a pre-trained network from the previous SOTA DINOv2 as the teacher for distilling knowledge into smaller networks. We observe that SEED significantly boosts the performance of ViT-S/16 and ViT-B/16 when a pre-trained ViT-L/16 serves as the teacher. Specifically, with the same number of training epochs, SEED yields k-NN accuracy improvements of \\(1.8\\%\\) and \\(1.4\\%\\) for ViT-S/16 and ViT-B/16 respectively, compared with a learnfrom-scratch setup. Our POA further outperforms SEED, delivering superior k-NN accuracy gains of \\(2.8\\%\\) (76.8% vs. 74.0%) and \\(2.1\\%\\) (80.9% vs. 78.8%) on ViT-S/16 and ViT-B/16. Moreover, the ViT-S/16 and ViT-B/16 derived directly from the pre-trained teacher from our POA, without additional pre-training, perform better than those enhanced by SEED, despite the latter undergoing two times of training epochs. These results demonstrate the effectiveness of our unified design, which integrates pre-training with once-for-all model generation. Additional comparison on training cost can be found in Appendix D.11." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.392, + 0.788, + 0.573 + ], + "angle": 0, + "content": "Alternative Designs to Elastic Pre-training. Upon the proposed POA, we investigate three alternatives for concurrently pre-training multi-sized models within a single pre-training session. As illustrated in Figure 4, the first variant, POA-V1, eliminates the intact student and modifies the teacher to be elastic which aligns with the architecture of the elastic student. The second variant, POA-V2, only removes the intact student. However, in this model, the intact teacher is directly updated via an EMA of the elastic students. These two variants bear resemblance to current teacher-student self-distillation paradigm, with the primary differences in the structure of the teacher or student network. For the POA-V3, we introduce an additional elastic teacher branch that mirrors the architecture of the elastic student, enabling further cross-view distillation. For more details of the three variants, please refer to the Appendix D.9." + }, + { + "type": "image", + "bbox": [ + 0.223, + 0.575, + 0.79, + 0.689 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.339, + 0.699, + 0.663, + 0.713 + ], + "angle": 0, + "content": "Fig. 4: Illustration of different variants of POA." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.72, + 0.788, + 0.841 + ], + "angle": 0, + "content": "We compare the k-NN and linear probing accuracy of each variant in Table 5. The results indicate that the performance drops dramatically for the variants with two branches. This decline can be attributed to two reasons: 1) in each training iteration, only a subset of the student's parameters are updated, leading to relatively insufficient overall pre-training; and 2) each elastic network undergoes training exclusively with the cross-view distillation, missing out the standard distillation guidance from a larger network, which our ablation study suggests is crucial for the efficacy of elastic networks. Although the POA-V3," + } + ], + [ + { + "type": "page_number", + "bbox": [ + 0.218, + 0.116, + 0.236, + 0.127 + ], + "angle": 0, + "content": "14" + }, + { + "type": "header", + "bbox": [ + 0.272, + 0.115, + 0.374, + 0.129 + ], + "angle": 0, + "content": "Y. Zhang et al." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.147, + 0.783, + 0.192 + ], + "angle": 0, + "content": "featuring an additional elastic teacher branch, performs slightly better than our POA, it does so at the expense of increased model complexity and computational costs, making it a less appealing scheme than POA." + }, + { + "type": "table_caption", + "bbox": [ + 0.23, + 0.194, + 0.772, + 0.209 + ], + "angle": 0, + "content": "Table 5: k-NN and linear probing (LP) comparison of different variants of POA." + }, + { + "type": "table", + "bbox": [ + 0.224, + 0.221, + 0.78, + 0.322 + ], + "angle": 0, + "content": "
Methodk-NNLP
ViT-S/16ViT-B/16ViT-L/16ViT-S/16ViT-B/16ViT-L/16
POA76.880.982.377.681.783.6
POA-V173.778.579.274.579.480.9
POA-V274.379.180.975.180.082.3
POA-V376.981.082.377.881.883.6
" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.342, + 0.788, + 0.6 + ], + "angle": 0, + "content": "How Does Elastic Student Facilitate Pre-training? The elastic student branch enables the derivation of diverse-sized models directly from a pre-trained teacher, while simultaneously enhancing the learned representation. The elastic branch plays a dual role. First, it acts as a training regularization to stabilize the training progress. In our preliminary experiments, we observe that pre-training loss of a ResNet backbone easily diverges to NaN without the elastic branch. Conversely, the inclusion of the elastic student yields a highly stable pre-training. Furthermore, the elastic student provides hundreds of sub-network candidates that will be assembled to the teacher during pre-training. Unlike existing self-distillation methods, the teacher in the POA SSL integrates a series of sub-networks through an EMA update. Wortsman et al. [44] have shown that averaging the weights of multiple models typically improves accuracy and robustness. Similarly, the integration of different sub-networks for the teacher is effective in improving the representation quality. At the same time, an improved teacher promotes the learning of students in return, fostering a positive feedback loop for pre-training. For the visualization of representations, please refer to Appendix E.2." + }, + { + "type": "title", + "bbox": [ + 0.216, + 0.625, + 0.36, + 0.642 + ], + "angle": 0, + "content": "5 Conclusion" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.659, + 0.788, + 0.841 + ], + "angle": 0, + "content": "In this study, we tackle the challenge of efficiently and effectively pre-training models of various sizes within a single self-supervised learning session, facilitating model deployment given different resource constraints. We propose a novel self-supervised learning paradigm, termed POA, which features an integrated design combining self-distillation and once-for-all model generation. It allows for the simultaneous pre-training of models of multiple sizes through an elastic branch design. POA enables the direct generation of varied-sized models from a pretrained teacher, which are ready for downstream tasks without additional pretraining. This advantage significantly improves the deployment flexibility and facilitates our pre-trained model to achieve SOTA results across various vision tasks. Looking forward, we plan to extend POA to Multimodal Large Language Models, tapping into its vast potential for real-world AI product deployment." + } + ], + [ + { + "type": "header", + "bbox": [ + 0.695, + 0.115, + 0.732, + 0.127 + ], + "angle": 0, + "content": "POA" + }, + { + "type": "page_number", + "bbox": [ + 0.769, + 0.116, + 0.786, + 0.127 + ], + "angle": 0, + "content": "15" + }, + { + "type": "title", + "bbox": [ + 0.217, + 0.145, + 0.324, + 0.161 + ], + "angle": 0, + "content": "References" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.177, + 0.785, + 0.204 + ], + "angle": 0, + "content": "1. Ba, J., Kiros, J.R., Hinton, G.E.: Layer normalization. ArXiv abs/1607.06450 (2016)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.205, + 0.785, + 0.245 + ], + "angle": 0, + "content": "2. Cai, H., Gan, C., Wang, T., Zhang, Z., Han, S.: Once for all: Train one network and specialize it for efficient deployment. In: International Conference on Learning Representations (2020)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.247, + 0.785, + 0.286 + ], + "angle": 0, + "content": "3. Cai, Z., Vasconcelos, N.: Cascade r-cnn: High quality object detection and instance segmentation. IEEE Transactions on Pattern Analysis and Machine Intelligence 43, 1483-1498 (2019)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.288, + 0.785, + 0.342 + ], + "angle": 0, + "content": "4. Caron, M., Misra, I., Mairal, J., Goyal, P., Bojanowski, P., Joulin, A.: Unsupervised learning of visual features by contrasting cluster assignments. In: Larochelle, H., Ranzato, M., Hadsell, R., Balcan, M., Lin, H. (eds.) Advances in Neural Information Processing Systems. vol. 33, pp. 9912-9924. Curran Associates, Inc. (2020)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.343, + 0.785, + 0.384 + ], + "angle": 0, + "content": "5. Caron, M., Touvron, H., Misra, I., J'egou, H., Mairal, J., Bojanowski, P., Joulin, A.: Emerging properties in self-supervised vision transformers. 2021 IEEE/CVF International Conference on Computer Vision (ICCV) pp. 9630-9640 (2021)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.385, + 0.785, + 0.425 + ], + "angle": 0, + "content": "6. Chen, T., Kornblith, S., Norouzi, M., Hinton, G.E.: A simple framework for contrastive learning of visual representations. International Conference on Machine Learning (2021)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.426, + 0.785, + 0.467 + ], + "angle": 0, + "content": "7. Chen, T., Zhai, X., Ritter, M., Lucic, M., Houlsby, N.: Self-supervised gans via auxiliary rotation loss. 2019 IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR) pp. 12146-12155 (2018)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.468, + 0.785, + 0.494 + ], + "angle": 0, + "content": "8. Chen, X., Fan, H., Girshick, R., He, K.: Improved baselines with momentum contrastive learning. arXiv preprint arXiv:2003.04297 (2020)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.496, + 0.785, + 0.535 + ], + "angle": 0, + "content": "9. Chen, Y., Liu, Y., Jiang, D., Zhang, X., Dai, W., Xiong, H., Tian, Q.: Sdae: Self-distillated masked autoencoder. In: European Conference on Computer Vision (2022)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.537, + 0.785, + 0.577 + ], + "angle": 0, + "content": "10. Cheng, X., Chen, J., Wang, R.: Modified dual attention triplet-supervised hashing network for image retrieval. Signal, Image and Video Processing 18, 1939-1948 (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.579, + 0.785, + 0.605 + ], + "angle": 0, + "content": "1. Cuturi, M.: Sinkhorn distances: Lightspeed computation of optimal transport. In: Neural Information Processing Systems (2013)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.607, + 0.785, + 0.633 + ], + "angle": 0, + "content": "2. Deng, J., Dong, W., Socher, R., Li, L.J., Li, K., Fei-Fei, L.: ImageNet: A LargeScale Hierarchical Image Database. In: CVPR 2009 (2009)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.634, + 0.785, + 0.674 + ], + "angle": 0, + "content": "3. Dong, C., Loy, C.C., He, K., Tang, X.: Image super-resolution using deep convolutional networks. IEEE Transactions on Pattern Analysis and Machine Intelligence 38, 295-307 (2014)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.675, + 0.785, + 0.729 + ], + "angle": 0, + "content": "4. Dosovitskiy, A., Beyer, L., Kolesnikov, A., Weissenborn, D., Zhai, X., Unterthiner, T., Dehghani, M., Minderer, M., Heigold, G., Gelly, S., Uszkoreit, J., Houlsby, N.: An image is worth 16x16 words: Transformers for image recognition at scale. International Conference on Learning Representations (ICLR) (2021)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.73, + 0.785, + 0.77 + ], + "angle": 0, + "content": "5. Fang, Z., Wang, J., Wang, L., Zhang, L., Yang, Y., Liu, Z.: Seed: Self-supervised distillation for visual representation. International Conference on Learning Representations (2021)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.772, + 0.785, + 0.812 + ], + "angle": 0, + "content": "6. Gao, Y., Zhuang, J.X., Lin, S., Cheng, H., Sun, X., Li, K., Shen, C.: Disco: Remedying self-supervised learning on lightweight models with distilled contrastive learning. In: European Conference on Computer Vision (ECCV) (2022)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.813, + 0.785, + 0.84 + ], + "angle": 0, + "content": "7. Grill, J.B., Strub, F., Altché, F., Tallec, C., Richemond, P., Buchatskaya, E., Doersch, C., Avila Pires, B., Guo, Z., Gheshlaghi Azar, M., et al.: Bootstrap your own" + }, + { + "type": "list", + "bbox": [ + 0.226, + 0.177, + 0.785, + 0.84 + ], + "angle": 0, + "content": null + } + ], + [ + { + "type": "page_number", + "bbox": [ + 0.218, + 0.116, + 0.236, + 0.127 + ], + "angle": 0, + "content": "16" + }, + { + "type": "header", + "bbox": [ + 0.272, + 0.115, + 0.375, + 0.129 + ], + "angle": 0, + "content": "Y. Zhang et al." + }, + { + "type": "ref_text", + "bbox": [ + 0.241, + 0.148, + 0.786, + 0.175 + ], + "angle": 0, + "content": "latent-a new approach to self-supervised learning. Advances in neural information processing systems 33, 21271-21284 (2020)" + }, + { + "type": "ref_text", + "bbox": [ + 0.218, + 0.176, + 0.787, + 0.218 + ], + "angle": 0, + "content": "18. He, K., Chen, X., Xie, S., Li, Y., Doll'ar, P., Girshick, R.B.: Masked autoencoders are scalable vision learners. 2022 IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR) pp. 15979-15988 (2021)" + }, + { + "type": "ref_text", + "bbox": [ + 0.218, + 0.218, + 0.787, + 0.259 + ], + "angle": 0, + "content": "19. He, K., Fan, H., Wu, Y., Xie, S., Girshick, R.B.: Momentum contrast for unsupervised visual representation learning. 2020 IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR) pp. 9726-9735 (2019)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.259, + 0.786, + 0.286 + ], + "angle": 0, + "content": "20. He, K., Gkioxari, G., Dolkar, P., Girshick, R.B.: Mask r-cnn. 2017 IEEE International Conference on Computer Vision (ICCV) pp. 2980-2988 (2017)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.286, + 0.786, + 0.327 + ], + "angle": 0, + "content": "21. He, K., Zhang, X., Ren, S., Sun, J.: Deep residual learning for image recognition. 2016 IEEE Conference on Computer Vision and Pattern Recognition (CVPR) pp. 770-778 (2015)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.328, + 0.786, + 0.355 + ], + "angle": 0, + "content": "22. Hinton, G.E., Vinyals, O., Dean, J.: Distilling the knowledge in a neural network. ArXiv abs/1503.02531 (2015)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.356, + 0.786, + 0.383 + ], + "angle": 0, + "content": "23. Hou, L., Shen, H., Cao, Q., Cheng, X.: Self-supervised gans with label augmentation. ArXiv abs/2106.08601 (2021)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.383, + 0.786, + 0.411 + ], + "angle": 0, + "content": "24. Hu, C., Li, X., Liu, D., Wu, H., Chen, X., Wang, J., Liu, X.: Teacher-student architecture for knowledge distillation: A survey. ArXiv abs/2308.04268 (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.411, + 0.786, + 0.452 + ], + "angle": 0, + "content": "25. Huang, Z., Jin, X., Lu, C., Hou, Q., Cheng, M.M., Fu, D., Shen, X., Feng, J.: Contrastive masked autoencoders are stronger vision learners. IEEE transactions on pattern analysis and machine intelligence PP (2022)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.452, + 0.786, + 0.508 + ], + "angle": 0, + "content": "26. Li, J., Wang, Y., ZHANG, X., Chen, Y., Jiang, D., Dai, W., Li, C., Xiong, H., Tian, Q.: Progressively compressed auto-encoder for self-supervised representation learning. In: The Eleventh International Conference on Learning Representations (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.508, + 0.786, + 0.563 + ], + "angle": 0, + "content": "27. Li, Z., Zhu, Y., Yang, F., Li, W., Zhao, C., Chen, Y., Chen, Z., Xie, J., Wu, L., Zhao, R., Tang, M., Wang, J.: Univip: A unified framework for self-supervised visual pre-training. 2022 IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR) pp. 14607-14616 (2022)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.564, + 0.786, + 0.618 + ], + "angle": 0, + "content": "28. Lin, J., Rao, Y., Lu, J., Zhou, J.: Runtime neural pruning. In: Guyon, I., Luxburg, U.V., Bengio, S., Wallach, H., Fergus, R., Vishwanathan, S., Garnett, R. (eds.) Advances in Neural Information Processing Systems. vol. 30. Curran Associates, Inc. (2017)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.619, + 0.786, + 0.66 + ], + "angle": 0, + "content": "29. Lin, T.Y., Maire, M., Belongie, S.J., Hays, J., Perona, P., Ramanan, D., Dóllar, P., Zitnick, C.L.: Microsoft coco: Common objects in context. In: European Conference on Computer Vision (2014)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.66, + 0.786, + 0.702 + ], + "angle": 0, + "content": "30. Liu, Z., Lin, Y., Cao, Y., Hu, H., Wei, Y., Zhang, Z., Lin, S., Guo, B.: Swin transformer: Hierarchical vision transformer using shifted windows. 2021 IEEE/CVF International Conference on Computer Vision (ICCV) pp. 9992-10002 (2021)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.702, + 0.786, + 0.73 + ], + "angle": 0, + "content": "31. Loshchilov, I., Hutter, F.: Fixing weight decay regularization in adam. ArXiv abs/1711.05101 (2017)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.73, + 0.786, + 0.757 + ], + "angle": 0, + "content": "32. van den Oord, A., Li, Y., Vinyals, O.: Representation learning with contrastive predictive coding. ArXiv abs/1807.03748 (2018)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.757, + 0.786, + 0.84 + ], + "angle": 0, + "content": "33. Oquab, M., Darcet, T., Moutakanni, T., Vo, H.V., Szafraniec, M., Khalidov, V., Fernandez, P., HAZIZA, D., Massa, F., El-Nouby, A., Assran, M., Ballas, N., Galuba, W., Howes, R., Huang, P.Y., Li, S.W., Misra, I., Rabbat, M., Sharma, V., Synnaeve, G., Xu, H., Jegou, H., Mairal, J., Labatut, P., Joulin, A., Bojanowski, P.: DINOv2: Learning robust visual features without supervision. Transactions on Machine Learning Research (2024)" + }, + { + "type": "list", + "bbox": [ + 0.217, + 0.148, + 0.787, + 0.84 + ], + "angle": 0, + "content": null + } + ], + [ + { + "type": "header", + "bbox": [ + 0.695, + 0.115, + 0.732, + 0.127 + ], + "angle": 0, + "content": "POA" + }, + { + "type": "page_number", + "bbox": [ + 0.769, + 0.116, + 0.786, + 0.127 + ], + "angle": 0, + "content": "17" + }, + { + "type": "ref_text", + "bbox": [ + 0.214, + 0.147, + 0.788, + 0.189 + ], + "angle": 0, + "content": "34. Ruan, Y., Singh, S., Morningstar, W.R., Alemi, A.A., Ioffe, S., Fischer, I., Dillon, J.V.: Weighted ensemble self-supervised learning. In: The Eleventh International Conference on Learning Representations (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.214, + 0.19, + 0.787, + 0.232 + ], + "angle": 0, + "content": "35. Song, K., Xie, J., Zhang, S., Luo, Z.: Multi-mode online knowledge distillation for self-supervised visual representation learning. 2023 IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR) pp. 11848-11857 (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.214, + 0.232, + 0.787, + 0.286 + ], + "angle": 0, + "content": "36. Tao, C., Wang, H., Zhu, X., Dong, J., Song, S., Huang, G., Dai, J.: Exploring the equivalence of siamese self-supervised learning via a unified gradient framework. 2022 IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR) pp. 14411-14420 (2021)" + }, + { + "type": "ref_text", + "bbox": [ + 0.214, + 0.287, + 0.787, + 0.314 + ], + "angle": 0, + "content": "37. Team, G.G.: Introducing gemini: our largest and most capable ai model. Google (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.214, + 0.315, + 0.787, + 0.357 + ], + "angle": 0, + "content": "38. Tian, K., Jiang, Y., qishuai diao, Lin, C., Wang, L., Yuan, Z.: Designing BERT for convolutional networks: Sparse and hierarchical masked modeling. In: The Eleventh International Conference on Learning Representations (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.214, + 0.357, + 0.787, + 0.412 + ], + "angle": 0, + "content": "39. Tomasev, N., Bica, I., McWilliams, B., Buesing, L., Pascanu, R., Blundell, C., Mitrovic, J.: Pushing the limits of self-supervised resnets: Can we outperform supervised learning without labels onImagenet? arXiv preprint arXiv:2201.05119 (2022)" + }, + { + "type": "ref_text", + "bbox": [ + 0.214, + 0.412, + 0.787, + 0.467 + ], + "angle": 0, + "content": "40. Touvron, H., Cord, M., Oquab, M., Bojanowski, P., Verbeek, J., Jegou, H.: Cotraining 2l submodels for visual recognition. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR). pp. 11701-11710 (June 2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.214, + 0.467, + 0.787, + 0.509 + ], + "angle": 0, + "content": "41. Wang, X., Huang, Y., Zeng, D., Qi, G.J.: Caco: Both positive and negative samples are directly learnable via cooperative-adversarial contrastive learning. IEEE Transactions on Pattern Analysis and Machine Intelligence 45, 10718-10730 (2022)" + }, + { + "type": "ref_text", + "bbox": [ + 0.214, + 0.509, + 0.787, + 0.536 + ], + "angle": 0, + "content": "42. Wang, X., Yu, F., Dou, Z.Y., Gonzalez, J.: Skipnet: Learning dynamic routing in convolutional networks. ArXiv abs/1711.09485 (2017)" + }, + { + "type": "ref_text", + "bbox": [ + 0.214, + 0.536, + 0.787, + 0.578 + ], + "angle": 0, + "content": "43. Wang, X., Zhang, R., Shen, C., Kong, T., Li, L.: Dense contrastive learning for self-supervised visual pre-training. 2021 IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR) pp. 3023-3032 (2020)" + }, + { + "type": "ref_text", + "bbox": [ + 0.214, + 0.578, + 0.787, + 0.647 + ], + "angle": 0, + "content": "44. Wortsman, M., Ilharco, G., Gadre, S.Y., Roelofs, R., Gontijo-Lopes, R., Morcos, A.S., Namkoong, H., Farhadi, A., Carmon, Y., Kornblith, S., Schmidt, L.: Model soups: averaging weights of multiple fine-tuned models improves accuracy without increasing inference time. In: International Conference on Machine Learning (ICML) (2022)" + }, + { + "type": "ref_text", + "bbox": [ + 0.214, + 0.647, + 0.787, + 0.688 + ], + "angle": 0, + "content": "45. Wu, H., Xu, J., Wang, J., Long, M.: Autoformer: Decomposition transformers with auto-correlation for long-term series forecasting. In: Neural Information Processing Systems (2021)" + }, + { + "type": "ref_text", + "bbox": [ + 0.214, + 0.689, + 0.787, + 0.716 + ], + "angle": 0, + "content": "46. Xiao, T., Liu, Y., Zhou, B., Jiang, Y., Sun, J.: Unified perceptual parsing for scene understanding. ArXiv abs/1807.10221 (2018)" + }, + { + "type": "ref_text", + "bbox": [ + 0.214, + 0.716, + 0.787, + 0.744 + ], + "angle": 0, + "content": "47. Xie, Z., Lin, Y., Yao, Z., Zhang, Z., Dai, Q., Cao, Y., Hu, H.: Self-supervised learning with swin transformers. ArXiv abs/2105.04553 (2021)" + }, + { + "type": "ref_text", + "bbox": [ + 0.214, + 0.744, + 0.787, + 0.786 + ], + "angle": 0, + "content": "48. Xie, Z., Zhang, Z., Cao, Y., Lin, Y., Bao, J., Yao, Z., Dai, Q., Hu, H.: Simmim: a simple framework for masked image modeling. 2022 IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR) pp. 9643-9653 (2021)" + }, + { + "type": "ref_text", + "bbox": [ + 0.214, + 0.786, + 0.787, + 0.813 + ], + "angle": 0, + "content": "49. Yan, B., Jiang, Y., Sun, P., Wang, D., Yuan, Z., Luo, P., Lu, H.: Towards grand unification of object tracking. In: European Conference on Computer Vision (2022)" + }, + { + "type": "ref_text", + "bbox": [ + 0.214, + 0.813, + 0.787, + 0.84 + ], + "angle": 0, + "content": "50. Yan, C., Chang, X., Li, Z., Yao, L., Luo, M., Zheng, Q.: Masked distillation advances self-supervised transformer architecture search. In: The Twelfth Interna" + }, + { + "type": "list", + "bbox": [ + 0.214, + 0.147, + 0.788, + 0.84 + ], + "angle": 0, + "content": null + } + ], + [ + { + "type": "page_number", + "bbox": [ + 0.218, + 0.116, + 0.236, + 0.127 + ], + "angle": 0, + "content": "18" + }, + { + "type": "header", + "bbox": [ + 0.272, + 0.115, + 0.375, + 0.129 + ], + "angle": 0, + "content": "Y. Zhang et al." + }, + { + "type": "ref_text", + "bbox": [ + 0.241, + 0.148, + 0.786, + 0.175 + ], + "angle": 0, + "content": "tional Conference on Learning Representations (2024), https://openreview.net/ forum?id=LUpC8KTvdV" + }, + { + "type": "ref_text", + "bbox": [ + 0.218, + 0.176, + 0.786, + 0.203 + ], + "angle": 0, + "content": "51. Yu, F., Huang, K., Wang, M., Cheng, Y., Chu, W., Cui, L.: Width & depth pruning for vision transformers. In: AAAI Conference on Artificial Intelligence (2022)" + }, + { + "type": "ref_text", + "bbox": [ + 0.218, + 0.204, + 0.786, + 0.245 + ], + "angle": 0, + "content": "52. Yu, J., Li, X., Koh, J.Y., Zhang, H., Pang, R., Qin, J., Ku, A., Xu, Y., Baldridge, J., Wu, Y.: Vector-quantized image modeling with improved VQGAN. In: International Conference on Learning Representations (2022)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.246, + 0.786, + 0.273 + ], + "angle": 0, + "content": "53. Zhang, Y., Zhong, Q., Ma, L., Xie, D., Pu, S.: Learning incremental triplet margin for person re-identification. AAAI Conference on Artificial Intelligence (2019)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.273, + 0.786, + 0.314 + ], + "angle": 0, + "content": "54. Zhou, B., Zhao, H., Puig, X., Fidler, S., Barriuso, A., Torralba, A.: Scene parsing through ade20k dataset. 2017 IEEE Conference on Computer Vision and Pattern Recognition (CVPR) pp. 5122-5130 (2017)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.315, + 0.786, + 0.356 + ], + "angle": 0, + "content": "55. Zhou, J., Wei, C., Wang, H., Shen, W., Xie, C., Yuille, A., Kong, T.: ibot: Image bert pre-training with online tokenizer. International Conference on Learning Representations (ICLR) (2022)" + }, + { + "type": "list", + "bbox": [ + 0.217, + 0.148, + 0.786, + 0.356 + ], + "angle": 0, + "content": null + } + ] +] \ No newline at end of file diff --git a/2024/POA_ Pre-training Once for Models of All Sizes/9b5de88d-0728-41f9-a801-73fd67e5b8a0_origin.pdf b/2024/POA_ Pre-training Once for Models of All Sizes/9b5de88d-0728-41f9-a801-73fd67e5b8a0_origin.pdf new file mode 100644 index 0000000000000000000000000000000000000000..d1bac1171641dd536ecc19cd57b6f719b3d9cc86 --- /dev/null +++ b/2024/POA_ Pre-training Once for Models of All Sizes/9b5de88d-0728-41f9-a801-73fd67e5b8a0_origin.pdf @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:cd6c4d631892bbef76854b7b21a04755ae33243374b0de1de5d91614721f3709 +size 829634 diff --git a/2024/POA_ Pre-training Once for Models of All Sizes/full.md b/2024/POA_ Pre-training Once for Models of All Sizes/full.md new file mode 100644 index 0000000000000000000000000000000000000000..b0927e85088feaa285b50b99684e9dadcbc7f6ce --- /dev/null +++ b/2024/POA_ Pre-training Once for Models of All Sizes/full.md @@ -0,0 +1,286 @@ +# POA: Pre-training Once for Models of All Sizes + +Yingying Zhang, Xin Guo, Jiangwei Lao, Lei Yu, Lixiang Ru, Jian Wang, Guo Ye, Huimei He, Jingdong Chen, and Ming Yang + +Ant Group. + +qichu.zyy@antgroup.com + +Abstract. Large-scale self-supervised pre-training has paved the way for one foundation model to handle many different vision tasks. Most pre-training methodologies train a single model of a certain size at one time. Nevertheless, various computation or storage constraints in real-world scenarios require substantial efforts to develop a series of models with different sizes to deploy. Thus, in this study, we propose a novel tri-branch self-supervised training framework, termed as POA (Pre-training Once for All), to tackle this aforementioned issue. Our approach introduces an innovative elastic student branch into a modern self-distillation paradigm. At each pre-training step, we randomly sample a sub-network from the original student to form the elastic student and train all branches in a self-distilling fashion. Once pre-trained, POA allows the extraction of pre-trained models of diverse sizes for downstream tasks. Remarkably, the elastic student facilitates the simultaneous pre-training of multiple models with different sizes, which also acts as an additional ensemble of models of various sizes to enhance representation learning. Extensive experiments, including k-nearest neighbors, linear probing evaluation and assessments on multiple downstream tasks demonstrate the effectiveness and advantages of our POA. It achieves state-of-the-art performance using ViT, Swin Transformer and ResNet backbones, producing around a hundred models with different sizes through a single pre-training session. The code is available at: https://github.com/Qichuzyy/POA. + +Keywords: Self-supervised Learning $\cdot$ Pre-training Once for All + +# 1 Introduction + +Learning generalizable visual representations in a large model by self-supervised learning has delivered superior performance across a wide range of visual tasks [10, 13, 20, 46, 49, 53] in recent years. Nevertheless, when deployed to real-world applications, a large model has to be adapted to diverse resource limitations in terms of computation, storage, or power consumption, etc. For example, a well-engineered AI product typically comprises a suite of models tailored for varying scenarios, such as Gemini Nano, Pro and Ultra [37]. Given a large pre-trained model, common solutions to deploy it to multiple application scenarios with different resource constraints include additional weight pruning [28, 42, 51], knowledge distillation [22, 24], or even re-training a small network from scratch, which + +all require substantial development efforts. Consequently, this issue prompts a critical question: is it possible to pre-train once to produce multiple models with different sizes simultaneously, each delivering sufficiently good representations? + +To address this challenge, we introduce a new paradigm of self-supervised learning, called POA (Pre-training Once for All). POA is built upon the prevalent teacher-student self-distillation framework [5,33,55], with an additional innovative elastic student branch. The elastic student branch embeds a series of sub-networks through parameter sharing, upon the observation that smaller-sized models are sub-networks of a larger-sized one for modern network structures [14,21,30]. Moreover, the parameters of this branch are shared with the original, or intact student. At each pre-training step, we randomly sample a subset of parameters from the intact student to form the corresponding elastic student. Both the original intact student and the elastic student are trained to emulate the output of the teacher network. The teacher itself is continually refined via an exponential moving average (EMA) of the student's parameters, including the sampled elastic student. The elastic student facilitates effective and efficient pre-training on different subsets of parameters, leading to the successful extraction of high-performance sub-networks from the pre-trained teacher for subsequent downstream scenarios. It also acts as a form of training regularization by enforcing the outputs to match between the teacher and various sub-networks, contributing to a stable training process. Additionally, by serving as an ensemble of different sub-networks across different pre-training steps, the elastic student improves the representation learning [44]. + +To the best of our knowledge, POA represents the first self-supervised learning method capable of training multiple-sized models concurrently, each obtaining high-quality representations for different resource constraints without further pre-training. Figure 1 displays the k-nearest neighbors (k-NN) evaluation results for 143 sub-networks extracted from a ViT-L model [14] pre-trained by POA. By choosing different elastic widths and depths, which will be explained in Sec. 3.1, the pre-trained teacher model can generate a sufficient number of candidate sub-networks for the selection of the suitable model tailored for downstream applications according to the + +computational resources available. Notably, each sub-network is well-trained and exhibits superior performance, thanks to our meticulous design on same-view distillation, as validated in Sec. 4.4. In particular, the ViT-S, ViT-B and ViT- + +![](images/58642ae34be5e2e6be6ff4bb4893376192d54c89b6fa430015f321e75fff488a.jpg) +Fig.1: The k-NN evaluation accuracy of 143 elastic ViTs derived from the ViT-L/16 teacher model pre-trained with POA. + +L models set new benchmarks, achieving the state-of-the-art (SOTA) results compared with those pre-trained by existing methods [33, 34, 55]. + +To rigorously evaluate the efficacy of our approach, we conduct extensive experiments using three widely-used backbone architectures, i.e., ViT [14], Swin Transformer [30], and ResNet [21]. Each backbone is pre-trained on ImageNet-1K dataset and assessed using k-NN and linear probing classification evaluation, as well as downstream dense prediction tasks such as object detection and semantic segmentation. Our method achieves SOTA accuracy across various model sizes with a single pre-training session. The technical contributions of this work are summarized as follows: + +- To the best of our knowledge, POA is the first pre-training paradigm that integrates unsupervised representation learning and once-for-all model generation within a single pre-training session. It tackles the Pre-training-Once-for-All challenge which has been seldomly explored by the community but is of great practical significance for real-world deployment that usually requires a suite of models. + +- We propose a novel and elegant component called Elastic Student, featuring a range of elastic operators that enable POA to be compatible with popular backbone structures including ViT, Swin Transformer and ResNet. It provides the capability to generate models of diverse sizes. Furthermore, it serves as a model ensemble to smooth training process and improve learned representations. + +- Through thorough assessments using k-NN, linear probing and downstream dense task evaluation, our approach exhibits superior performance over existing state-of-the-art pre-training methods across multiple metrics. Moreover, we compare our POA with Self-Supervised Distillation (SEED) [15], a knowledge distillation method designed especially for self-supervised learning, further validating POA's effectiveness. + +# 2 Related Work + +# 2.1 Self-supervised Learning + +Self-supervised learning (SSL) frameworks commonly fall into two categories, generative and discriminative SSL. Most generative SSL approaches [7,9,18,23, 25,26,38,48,52] focus on learning image representations directly in pixel space. On the other hand, discriminative SSL methods [4,6,19,27,36,41,43,47] aim to learn representations by pulling those of different views of the same image closer while separating the representations of views from different images. + +Contrastive learning (CL) with the InfoNCE loss [32] has emerged as a popular approach for discriminative SSL methods, attracting significant research interests in recent years. Although CL methods prevent the collapse of network representations through the use of negative samples, they still suffer from the dimensional collapse, where representations tend to collapse into a low-dimensional manifold. Grill et al. introduced a distillation-based asymmetric + +framework known as BYOL [17], which circumvents collapse without self-labeling or contrastive loss relying on negative samples. Following this work, distillation-based frameworks have become a prevailing trend in self-supervised learning. These frameworks [8, 16, 47] often merge with others to enhance overall performance. DINO [5] presented a simple self-distillation framework and has demonstrated impressive results in ViT pre-training. Subsequent works [33, 55] further improved the pre-training performance via masked token or the novel clustering design. Given the substantial benefits of distillation-based methods over other SSL techniques, we have developed our POA SSL framework upon these successful methodologies. + +# 2.2 Dynamic Architecture. + +Recently, Chen et al. proposed AutoFormer [45], which trained a supernet to support the effective search of optimal sub-network under some specific parameter number constraints. On the basis of [45], MaskTAS [50] introduced a self-supervised transformer architecture search method. Cai et al. [2] trained a network that accommodates various architectural configurations to reduce the training expense. Their methodologies enable the extraction of a specialized sub-network from the trained supernet. The design of the elastic student in our POA SSL is inspired by the weight-sharing strategy employed in these neural architecture search (NAS) methods. However, our implementation differs significantly due to the distinct purpose from NAS. Specifically, NAS aims to discover the optimal architecture within certain parameter constraints, which typically involves a huge number of sub-networks (more than $10^{16}$ ) in the search space. Given the limitations on the number of training iterations and the network parameter capacity, it is challenging to ensure high performance across all sub-networks. After training the supernet, NAS requires a subsequent phase of searching and re-training to finalize the output model. In contrast, our POA SSL defines a sufficiently large yet compact set of sub-networks with elastic depths and widths for the purpose of pre-training models of various sizes via a single training session. Additional design on the same-view distillation guarantees that all elastic sub-networks within our framework are adequately and efficiently trained. As a result, our POA can readily extract a range of sub-networks from the teacher model without the need for extra pre-training, facilitating an easy selection of a suitable sub-network for different computational contexts. The design of elastic student is somewhat akin to the supervised learning method, Cosub [40]. The main difference is that Cosub simply skips blocks, making only the depth elastic. + +# 3 POA Self-supervised Learning Framework + +Our primary goal is to pre-train models of multiple sizes via a single self-supervised pre-training session. To this end, we propose a novel SSL framework named POA, inspired by the latest progress in self-distillation techniques [5, 33, 55]. The architecture of POA is illustrated in Figure 2, encompassing a + +![](images/f899a486424e9769904603e459e5736d7358d72eec370a022540471bbfad631c.jpg) +Fig. 2: Overview of the POA SSL: Given an image $x$ , two augmented views $x_{a}$ and $x_{b}$ are generated. These views are input into three branches: a teacher, an intact student, and an elastic student, the latter being derived from the intact student. POA optimizes distillation losses in a twofold manner: the intact and the elastic students are distilled from the teacher using the cross-view data respectively, and additionally, the elastic student is distilled from the intact student using the same-view data. + +teacher, an intact student, an elastic student and their corresponding heads. The teacher is updated with an EMA of the students. The elastic student is a derivative of the intact one, with both the backbones and heads' parameters shared. We leverage distillation in two aspects: both the intact and elastic students are distilled from the teacher using different views of the same image, while the elastic student additionally learns from the intact student using the identical views. The cross-view distillation works as a form of representation learning, as introduced in [5, 33, 55]. Notably, in addition to the regular EMA update with only the intact student as [5, 33, 55], the elastic student provides a randomly-sampled sub-network at each pre-training step to participate in the teacher's EMA refinement. This procedure actually simulates an ensemble of multiple sub-networks, which is also proven to be beneficial in the realm of supervised learning [44]. The same-view distillation is a standard knowledge distillation between the intact and elastic students, promoting the quality of the elastic one. Experiments in Sec. 4.4 validate our design in details. + +# 3.1 Design of Elastic Student + +The elastic student is a sub-network whose parameters are extracted from the intact student. In the context of a transformer backbone, the width refers to the dimensionality of the tokens, whereas for a convolutional backbone, the width indicates the number of channels. We denote the depth as the number of basic blocks in either a transformer or a convolutional network. Given the value of the width and depth, it yields a certain network structure. For simplicity, we focus on detailing the elastic design of the ViT in this section. For the similar elastic design of the Swin Transformer and ResNet, please refer to Appendix A. + +A basic block of ViT mainly consists of a multi-head self-attention (MSA) module and a multi-layer perceptron (MLP) module. Layer Normalization (LN) [1] is applied before each module, with residual connections after each module. + +As shown in the left part of Figure 3, an elastic block refers to a stack of elastic MSA, MLP, and LN after adjusting the width in the original basic block in ViT. In our approach, the elastic student branch is constructed by assembling a specific number of these elastic blocks at each training iteration. + +![](images/998f376e1f2880a4ce99e0e8de331a39cf8d97fc432940a0dceff96ec234665c.jpg) +Fig. 3: Illustration of the elastic MSA in an elastic ViT block. To be concise, we simply exclude the projection layers that correspond to $K$ and $V$ in each head. + +**Elastic MSA** An original, or intact MSA module consists of three major components, i.e., input projection layers, an operator contains attention & concatenation, and an output projection layer. We define a projection layer as $(w^{*}, b^{*})$ , where $w^{*}$ is the linear transformation weight, $b^{*}$ denotes the corresponding bias, and * indicates the name of the layer. As shown in the right part of Figure 3, given a token dimension of $D_{max} = N_h \cdot D_h$ , where $N_h$ is the number of attention heads and $D_h$ is the head dimension, an input sequence $z \in \mathbb{R}^{T \times D_{max}}$ with length $T$ is initially projected to form the queries $Q \in \mathbb{R}^{T \times D_h}$ , keys $K \in \mathbb{R}^{T \times D_h}$ , and values $V \in \mathbb{R}^{T \times D_h}$ by the intact MSA. To generate elastic MSA, we define $\mathrm{M} + 1$ elastic widths, including $D_{max}$ , spaced at intervals of $D_h$ as follows: + +$$ +D _ {i} = \left(N _ {h} - i\right) \cdot D _ {h}, \quad \forall i \in \{0, 1, \dots , M \}, \quad M < N _ {h}. \tag {1} +$$ + +For each elastic width $D_{i}$ , the weight $w_{i}^{a1} \in \mathbb{R}^{D_{h} \times D_{i}}$ and bias $b_{i}^{a1} \in \mathbb{R}^{D_{h}}$ that generate $Q$ , $K$ , and $V$ of each head are extracted from the corresponding input projection layer ( $w^{a1}$ , $b^{a1}$ ) in the intact MSA, as $w_{i}^{a1} = w^{a1}[:,: D_{i}] \cdot \alpha_{i}$ and $b_{i}^{a1} = b^{a1}$ . Here, $\alpha_{i}$ represents the scaling factor to cope with the reduction of the input dimension, calculated as $\alpha_{i} = D_{max} / D_{i}$ . As the reduction of width, the number of attention heads in the elastic MAS is naturally reduced to $N_{h} - i$ . Similarly, for the output projection layer ( $w^{a2}$ , $b^{a2}$ ), the weight $w_{i}^{a2} \in \mathbb{R}^{D_{i} \times D_{i}}$ and bias $b_{i}^{a2} \in \mathbb{R}^{D_{i}}$ are extracted as: + +$$ +w _ {i} ^ {a 2} = w ^ {a 2} [: D _ {i},: D _ {i} ] \cdot \alpha_ {i} \quad b _ {i} ^ {a 2} = b ^ {a 2} [: D _ {i} ]. \tag {2} +$$ + +**Elastic MLP** An original, or intact MLP module in the ViT block contains two projection layers. The first layer ( $w^{m1}, b^{m1}$ ) expands the dimension of embedding by a factor of $s$ , which is generally set to 4 in the ViT structure. The second layer ( $w^{m2}, b^{m2}$ ) then projects it back to the original dimension. The parameters for both layers of the elastic MLP are extracted in a manner analogous to that + +described in Equation 2 as follows: + +$$ +w _ {i} ^ {m 1} = w ^ {m 1} \left[ \colon D _ {i} \cdot s, \colon D _ {i} \right] \cdot \alpha_ {i} \quad b _ {i} ^ {m 1} = b ^ {m 1} \left[ \colon D _ {i} \cdot s \right] +$$ + +$$ +w _ {i} ^ {m 2} = w ^ {m 2} [: D _ {i},: D _ {i} \cdot s ] \cdot \alpha_ {i} \quad b _ {i} ^ {m 2} = b ^ {m 2} [: D _ {i} ]. \tag {3} +$$ + +**Elastic LN** For elastic LN, we directly use the first $D_{i}$ elements of the parameter inside the original LN, akin to the bias extraction in Equation 2. + +To create a sub-network with $L_{i}$ elastic blocks from the intact ViT comprising $L_{max}$ blocks, we introduce a set of $N + 1$ elastic depths, defined as $L_{i} = L_{max} - i$ , $\forall i \in \{0,1,\dots,N\}$ , $N < L_{max}$ . For a specific depth $L_{i}$ , we select the corresponding blocks based on their block IDs at equal intervals. Each block ID $BID_{j}^{L_{i}}$ that is activated for depth $L_{i}$ can be formulated as: + +$$ +B I D _ {j} ^ {L _ {i}} = \left\lfloor \frac {\left(L _ {\max } - 1\right) \cdot j}{L _ {i} - 1} \right\rfloor , \quad \forall j \in \{0, 1, \dots , L _ {i} - 1 \}. \tag {4} +$$ + +Consequently, by combining elastic widths and depths, we can generate a total of $(N + 1) \cdot (M + 1)$ distinct sub-networks. For instance, by setting the elastic width to 384 and the elastic depth to 12, we can directly extract a ViT-S from the intact network such as ViT-L. During each iteration of the pre-training, we randomly select one of these sub-networks to serve as the elastic student branch. + +# 3.2 Distillation between Views + +POA performs distillation across its three branches accordingly. Given a pair of globally augmented views of an input image $x$ , denoted as $x_{a}$ and $x_{b}$ , the teacher encoder $E_{T}$ extracts the feature $Z_{a} = E_{T}(x_{a})$ using $x_{a}$ as input. Simultaneously, $x_{b}$ is fed into both the intact student encoder $E_{IS}$ and the elastic student encoder $E_{ES}$ , producing the features $Z_{b1} = E_{IS}(x_{b})$ and $Z_{b2} = E_{ES}(x_{b})$ , respectively. The feature output from the teacher encoder, $Z_{a}$ , is then processed by the teacher head $H_{T}$ , followed by centering with the Sinkhorn-Knopp (SK) [11] algorithm and normalization using a temperature-scaled softmax to generate the probability $p_{a}$ as follows: + +$$ +l _ {a} = S K \left(H _ {T} \left(Z _ {a}\right)\right), l _ {a} \in \mathbb {R} ^ {P} \quad p _ {a} ^ {i} = \frac {\exp \left(l _ {a} ^ {i} / \tau\right)}{\sum_ {k = 0} ^ {P - 1} \exp \left(l _ {a} ^ {k} / \tau\right)}, \forall i \in \{0, \dots , P - 1 \}, \tag {5} +$$ + +where $P$ is the number of prototypes and $\tau > 0$ is a temperature parameter. In a similar fashion, we compute the probabilities $p_{b1}^{i}$ and $p_{b2}^{i}$ for the intact and elastic student encoders, respectively, by processing the outputs through the student heads $H_{IS}$ and $H_{ES}$ . These outputs are then passed through a temperature-scaled softmax function using a temperature parameter $\tau'$ tailored for the student. It should be noted that $H_{IS}$ and $H_{ES}$ share the identical parameters, except that the first projection layer of $H_{ES}$ is adapted similarly as in Equation 2 to conform to align with the corresponding dimensionality. For simplicity, we omit the explicit expressions for $p_{b1}^{i}$ and $p_{b2}^{i}$ , as they follow a similar + +calculation to Equation 5. For the intact student branch, we perform distillation from the teacher using the cross-view data as follows: + +$$ +\mathcal {L} _ {I S} ^ {g} = - p _ {a} \log \left(p _ {b 1}\right). \tag {6} +$$ + +The elastic student branch plays a pivotal role in our POA framework. To ensure adequate training of this branch, we employ a dual distillation from the teacher and intact student branch. The first distillation involves the teacher model, which utilizes the cross-view data to guide the learning of representations. The second is a distillation process with the intact student model, which uses the same-view data. This same-view distillation is responsible for transferring the representations learned by the intact student to the elastic student branch. The loss functions for this dual distillation process are formulated as follows: + +$$ +\mathcal {L} _ {E S 1} ^ {g} = - p _ {a} \log (p _ {b 2}), \quad \mathcal {L} _ {E S 2} ^ {g} = - p _ {b 1} \log (p _ {b 2}). \qquad (7) +$$ + +Note that in both loss functions, we sum over all prototypes to compute the cross-entropy loss between the respective probability distributions. + +# 3.3 Overall Loss of POA + +Following the SSL methods such as [5,33,55], we employ a multi-crop strategy [4] to create various distorted views from a single image. Apart from the two global views previously mentioned, we also generate $v$ local views with lower resolution $x_{l_1}, x_{l_2}, \ldots, x_{l_v}$ . These local views are processed by both students to promote local-to-global correspondence. The local distillation losses for the intact and elastic students are computed as follows: + +$$ +\mathcal {L} _ {I S} ^ {l} = - \frac {1}{v} \sum_ {i = 1} ^ {v} p _ {a} \log \left(p _ {l _ {i} 1}\right), \tag {8} +$$ + +$$ +\mathcal {L} _ {E S 1} ^ {l} = - \frac {1}{v} \sum_ {i = 1} ^ {v} p _ {a} \log \left(p _ {l _ {i} 2}\right), \quad \mathcal {L} _ {E S 2} ^ {l} = - \frac {1}{v} \sum_ {i = 1} ^ {v} p _ {l _ {i} 1} \log \left(p _ {l _ {i} 2}\right), \tag {9} +$$ + +where $p_{l_{i1}}$ and $p_{l_{i2}}$ are the probability produced by the intact and elastic student branches for the local view $l_i$ , respectively. The total distillation loss of the intact and elastic student is calculated by summing them with the factor $\lambda$ : + +$$ +\begin{array}{l} \mathcal {L} _ {\mathcal {S}} = \lambda \left(\mathcal {L} _ {I S} ^ {g} + \mathcal {L} _ {I S} ^ {l}\right) + (1 - \lambda) \left(\left(\mathcal {L} _ {E S 1} ^ {g} + \mathcal {L} _ {E S 1} ^ {l}\right) + \left(\mathcal {L} _ {E S 2} ^ {g} + \mathcal {L} _ {E S 2} ^ {l}\right)\right) \tag {10} \\ = \lambda \mathcal {L} _ {I S} + (1 - \lambda) \left(\mathcal {L} _ {E S 1} + \mathcal {L} _ {E S 2}\right). \\ \end{array} +$$ + +To ensure sufficient training of each sub-network from the elastic student, we introduce multiple projection heads (MPH) positioned after the backbone network. Each projection head has exactly the same structure, except for a different number of prototypes. For each head, the distillation loss $\mathcal{L}_{\mathcal{S}_i}$ for both the intact and elastic student is calculated with Equation. 10. Finally, the overall loss function in the POA framework with $H$ projection heads is formulated as: $\mathcal{L} = \frac{1}{H}\sum_{i = 1}^{H}\mathcal{L}_{\mathcal{S}_i}$ . + +# 4 Experiments + +# 4.1 Implementation Details + +Backbones. We have trained our POA using ViT, Swin Transformer and ResNet backbones, respectively. For the ViT, we configure the patch size to 16 and the dimension of each head in the MSA to 64. This aligns with the configurations typically used in existing SSL methods [5,34,35,55]. For the smallest and largest elastic networks of ViT, we choose the ViT-S and ViT-L, respectively. This leads to 11 elastic widths and 13 elastic depths, yielding a total of 143 ViT subnetworks. In the case of the Swin Transformer, we set the smallest and largest elastic networks as Swin-T and Swin-B, respectively. This configuration yields a total of 39 Swin sub-networks by multiplying the number of widths and depths as $3 \times 13$ . For the ResNet, we establish the smallest and largest elastic network configurations as ResNet-50 and ResNet-152. Consequently, this setup accounts for a total of 465 ResNet sub-networks, which is the product of $3 \times 155$ . + +Pre-Training Setup. We pre-train all models on the ImageNet-1K dataset [12] without the labels. The process employs the AdamW optimizer [31] with a batch size of 1600, which is distributed across 32 A100 GPUs when employing a ViT backbone. We adopt a learning rate schedule that begins with a linear warm-up during the first 10 epochs, reaching a base value that is scaled proportionally to the total batch size: $\mathrm{lr} = 0.004 \times \sqrt{\text{batch size} / 1024}$ , in line with [33]. Following this warm-up period, the learning rate is decayed with a cosine schedule. Similarly, the weight decay follows a cosine schedule, starting at 0.04 and increasing to 0.4. The student network's temperature $\tau'$ is fixed at 0.1, whereas the teacher temperature $\tau$ changes with a linear warm-up from 0.04 to 0.07 over the first 30 epochs. For further details, please refer to Appendix C. + +# 4.2 Experiments on ImageNet-1K + +After unsupervised pre-training, we assess the model's performance using two widely-recognized evaluation protocols in SSL domain on ImageNet-1K dataset, i.e., k-NN and linear probing. To ensure a fair comparison between SSL methods that employ different numbers of crop views for data augmentation, Zhou et al. [55] introduced the effective training epoch as a measure to quantify the extent of a method's pre-training. We report the effective training epochs of the SSL methods for comparison. For additional evaluations including fine-tuning, semi-supervised and unsupervised learning, please refer to Appendix D. + +k-NN and Linear Probing. To assess the quality of pre-trained features, we employ a k-NN classifier and a linear classifier on the frozen representations. For both the k-NN and linear probing (LP) evaluation, we follow the evaluation protocols established in [5, 33, 55]. The performance of our method when being trained using ResNet, Swin Transformer and ViT backbones is reported in Table 1. Our POA SSL achieves the SOTA k-NN accuracy of $82.3\%$ and LP accuracy + +Table 1: Comparison results of k-NN and linear probing classification accuracy (\%) on the ImageNet-1K dataset. "Param." refers to the number of parameters. "Epo." represents the number of effective training epochs following [55]. "/16" denotes patch size of 16. "/W7" means the window size of 7. "**" indicates our implementation based on official codebase. "+" denotes reproduced results using the released code. + +
MethodEpo.k-NNLPMethodEpo.k-NNLPMethodEpo.k-NNLP
ResNet-50(Param. 23M)Swin-T/W7(Param. 28M)ViT-S/16(Param. 21M)
VICReg2000-73.2SMoG1200-77.7DINO320074.577.0
SwAV240065.775.3iBOT120075.378.6iBOT320075.277.9
DINO320067.575.3DINOv2*120075.478.0ENT320075.377.7
UniGrad2400-75.5EsViT120075.778.1POA076.877.6
SCFS320068.575.7POA077.578.9ViT-B/16(Param. 85M)
ReLICv24000-77.1Swin-S/W7(Param. 49M)DINO160076.178.2
POA073.476.9DINOv2*120076.179.8ENT160077.179.1
ResNet-101(Param. 41M)EsViT120077.779.5iBOT160077.179.5
ReLICv24000-78.7POA079.381.3POA080.981.7
POA075.779.1Swin-B/W7(Param. 87M)ViT-L/16(Param. 307M)
ResNet-152(Param. 56M)DINOv2*120076.980.9iBOT120078.081.0
ReLICv24000-79.3EsViT120078.980.4DINOv2†120082.083.3
POA240076.479.9POA120079.682.0POA120082.383.6
+ +(a) ResNet backbone. +(b) Swin backbone. +(c) ViT backbone. + +of $83.6\%$ when using the ViT-L/16 backbone. By employing the sub-network extraction approach outlined in Section 3.1, we derive the sub-networks ViT-S/16 and ViT-B/16 from the teacher ViT-L/16 without any additional pre-training. Thus, the number of effective training epochs of them is 0. Notably, the extracted ViT-S/16 achieves the SOTA k-NN accuracy of $76.8\%$ and LP accuracy on par with the previous SOTA of $77.9\%$ reported by iBOT [55]. Our derived ViT-B/16 model sets new benchmarks for both k-NN and LP, with accuracy of $80.9\%$ and $81.7\%$ , respectively. For the models using Swin Transformer and ResNet backbones, POA also reaches SOTA performance in k-NN and LP accuracy. The only exception is the LP accuracy of ResNet-50, which is competitive with that of ReLICv2 [39], despite the latter being trained over a much longer period (4000 epochs). The superior performance achieved across a range of backbone architectures confirms the effectiveness and versatility of our method. For additional detailed comparison with 27 existing methods, please see Appendix D.1. + +# 4.3 Evaluation on Downstream Visual Tasks + +Object Detection and Instance Segmentation on COCO Dataset. For a fair comparison, we utilize the Cascade Mask R-CNN framework [3,20], which generates both bounding boxes and instance masks, in line with previous approach [55], on the COCO dataset [29]. We benchmark our results against existing SSL methods that generate pre-trained ViT-S/16 and ViT-B/16 backbones. As shown in Table 2, POA boosts the bounding box average precision $(\mathrm{AP}^b)$ for + +ViT-S/16 from 49.4 to 50.6 and the mask average precision $(\mathrm{AP}^m)$ from 42.6 to 43.8. When applied to ViT-B/16, POA attains an $\mathrm{AP}^b$ of 52.4 and an $\mathrm{AP}^m$ of 45.4, which represents a remarkable step forward over previous SOTA. + +Semantic Segmentation on ADE20K Dataset. For the semantic segmentation task, we primarily focus on two settings on the ADE20K dataset [54], following [55]. First, akin to the linear evaluation protocol in classification, we evaluate the quality of representations by keeping the patch features fixed and only fine-tuning one linear layer. This approach offers a clear comparison of representation quality. Second, we employ the UPerNet [46] as the task head and fine-tune the entire network. As depicted in Table 2, our POA significantly outperforms the supervised baseline using the ViT-S/16 backbone, achieving a substantial increase of 2.2 in mean Intersection over Union (mIoU), and surpassing iBOT by 1.3 mIoU. With the ViT-B/16 backbone, POA exceeds the previously best-performing method, iBOT, by 0.4 mIoU when utilizing UPerNet. Furthermore, under the assessment using solely a linear head, POA obtains an impressive improvement of 2.0 mIoU over iBOT as the performance is largely determined by the quality of the pre-trained representation. + +Table 2: Evaluation results on downstream detection and segmentation tasks. $\mathrm{Seg.}^{\dagger}$ indicates the use of a linear head for semantic segmentation. + +
MethodArch.Det. APbISeg. APmSeg. mIoUArch.Det. APbISeg. APmSeg.† mIoUSeg. mIoU
Sup.ViT-S/1646.240.144.5ViT-B/1649.843.235.446.6
BEiTViT-S/16---ViT-B/1650.143.527.445.8
DINOViT-S/16---ViT-B/1650.143.434.546.8
iBOTViT-S/1649.442.645.4ViT-B/1651.244.238.350.0
POAViT-S/1650.643.846.7ViT-B/1652.445.440.350.4
+ +# 4.4 Ablations and Discussions + +In this section, we conduct an empirical analysis of POA using ViT as backbone. Our investigation includes the impact of the loss functions $\mathcal{L}_{ES1}$ and $\mathcal{L}_{ES2}$ , in addition with the effectiveness of multiple projection heads. Moreover, we compare our POA with knowledge distillation techniques for self-supervised learning to demonstrate the advantages of combining once-for-all model generation with self-distillation in a unified framework. For more ablation studies, we direct readers to the Appendix D.8. In addition, we further contrast our POA with three variants tailored for elastic pre-training to showcase POA's superiority. Finally, we discuss how the elastic student facilitates the pre-training. + +Importance of Each Component. We evaluate the contributions of the components on a ViT backbone. Table 3 shows the performance of different component combinations. First, we note that employing multiple projection heads (MPH) enhances the learned representations for each elastic sub-network, particularly for smaller ones. The design consideration behind MPH is that for each + +Table 3: Contributions of each component in POA framework. We conduct assessments with the k-NN and linear probing (LP) evaluations. MPH denotes the multiple projection heads with different numbers of prototypes. + +
MPHLES1LES2k-NNLP
ViT-S/16ViT-B/16ViT-L/16ViT-S/16ViT-B/16ViT-L/16
✓✓✓76.880.982.377.681.783.6
✓✓76.280.782.277.381.683.4
✓75.180.282.275.881.083.4
✓72.879.182.175.380.883.3
+ +pre-training iteration, the sub-network is chosen randomly, leading to a relatively insufficient optimization. MPH introduces different sets of prototypes, which act as multiple semantic spaces for representation learning, enabling the teacher to distill various aspects of learned knowledge into the sub-network. Furthermore, we ascertain that the same-view distillation loss $\mathcal{L}_{ES2}$ is crucial for the representation quality of elastic sub-networks. Omitting $\mathcal{L}_{ES2}$ causes a significant drop in k-NN accuracy, by $3.4\%$ for ViT-S/16 and $1.6\%$ for ViT-B/16. In addition, $\mathcal{L}_{ES2}$ is more important than cross-view distillation $\mathcal{L}_{ES1}$ in terms of sub-networks' representation quality. The underlying reason is that the cross-view distillation is for the unsupervised representation learning, while the same-view distillation improves the sub-networks by distilling previously learned representations from the intact student. Table 4 confirms that distillation from already-good representations is more effective than representation learning, especially for smaller networks. It also explains why employing three different views in our POA is unnecessary. + +Table 4: Comparison with knowledge distillation for self-supervised learning. The teacher name "ViT-L/16-600" denotes a teacher model (ViT-L/16) that has been pretrained with 600 effective epochs. The student name "ViT-S/16-600" refers to a student model (ViT-S/16) that has undergone distillation from the pre-trained teacher with 600 effective epochs. The number of the left side of " $\rightarrow$ " indicates the performance of the teacher model when pre-trained individually. The number of the right side denotes the performance achieved after distillation using the SEED [15]. + +
MethodTeacherStudentTotal Epochsk-NNLP
DINOv2ViT-S/16120072.273.1
DINOv2+SEEDViT-L/16-600ViT-S/16-600120081.3 → 74.082.4 → 75.2
DINOv2+SEEDViT-L/16-1200ViT-S/16-1200240082.0 → 75.583.3 → 76.2
POAViT-S/16120076.877.6
DINOv2ViT-B/16120077.478.5
DINOv2+SEEDViT-L/16-600ViT-B/16-600120081.3 → 78.882.4 → 80.0
DINOv2+SEEDViT-L/16-1200ViT-B/16-1200240082.0 → 79.783.3 → 80.9
POAViT-B/16120080.981.7
+ +Comparison with Knowledge Distillation. Knowledge distillation (KD) is a proven strategy for enhancing small networks by leveraging the knowledge of + +a well-trained, larger network. We compare our method with a self-supervised KD method SEED [15] that employs a pre-trained network from the previous SOTA DINOv2 as the teacher for distilling knowledge into smaller networks. We observe that SEED significantly boosts the performance of ViT-S/16 and ViT-B/16 when a pre-trained ViT-L/16 serves as the teacher. Specifically, with the same number of training epochs, SEED yields k-NN accuracy improvements of $1.8\%$ and $1.4\%$ for ViT-S/16 and ViT-B/16 respectively, compared with a learnfrom-scratch setup. Our POA further outperforms SEED, delivering superior k-NN accuracy gains of $2.8\%$ (76.8% vs. 74.0%) and $2.1\%$ (80.9% vs. 78.8%) on ViT-S/16 and ViT-B/16. Moreover, the ViT-S/16 and ViT-B/16 derived directly from the pre-trained teacher from our POA, without additional pre-training, perform better than those enhanced by SEED, despite the latter undergoing two times of training epochs. These results demonstrate the effectiveness of our unified design, which integrates pre-training with once-for-all model generation. Additional comparison on training cost can be found in Appendix D.11. + +Alternative Designs to Elastic Pre-training. Upon the proposed POA, we investigate three alternatives for concurrently pre-training multi-sized models within a single pre-training session. As illustrated in Figure 4, the first variant, POA-V1, eliminates the intact student and modifies the teacher to be elastic which aligns with the architecture of the elastic student. The second variant, POA-V2, only removes the intact student. However, in this model, the intact teacher is directly updated via an EMA of the elastic students. These two variants bear resemblance to current teacher-student self-distillation paradigm, with the primary differences in the structure of the teacher or student network. For the POA-V3, we introduce an additional elastic teacher branch that mirrors the architecture of the elastic student, enabling further cross-view distillation. For more details of the three variants, please refer to the Appendix D.9. + +![](images/63f78a649a4a00db6dcb1fd57c087ad50e27c7a5cd5e7f1d95f95aebf15b26ae.jpg) +Fig. 4: Illustration of different variants of POA. + +We compare the k-NN and linear probing accuracy of each variant in Table 5. The results indicate that the performance drops dramatically for the variants with two branches. This decline can be attributed to two reasons: 1) in each training iteration, only a subset of the student's parameters are updated, leading to relatively insufficient overall pre-training; and 2) each elastic network undergoes training exclusively with the cross-view distillation, missing out the standard distillation guidance from a larger network, which our ablation study suggests is crucial for the efficacy of elastic networks. Although the POA-V3, + +featuring an additional elastic teacher branch, performs slightly better than our POA, it does so at the expense of increased model complexity and computational costs, making it a less appealing scheme than POA. + +Table 5: k-NN and linear probing (LP) comparison of different variants of POA. + +
Methodk-NNLP
ViT-S/16ViT-B/16ViT-L/16ViT-S/16ViT-B/16ViT-L/16
POA76.880.982.377.681.783.6
POA-V173.778.579.274.579.480.9
POA-V274.379.180.975.180.082.3
POA-V376.981.082.377.881.883.6
+ +How Does Elastic Student Facilitate Pre-training? The elastic student branch enables the derivation of diverse-sized models directly from a pre-trained teacher, while simultaneously enhancing the learned representation. The elastic branch plays a dual role. First, it acts as a training regularization to stabilize the training progress. In our preliminary experiments, we observe that pre-training loss of a ResNet backbone easily diverges to NaN without the elastic branch. Conversely, the inclusion of the elastic student yields a highly stable pre-training. Furthermore, the elastic student provides hundreds of sub-network candidates that will be assembled to the teacher during pre-training. Unlike existing self-distillation methods, the teacher in the POA SSL integrates a series of sub-networks through an EMA update. Wortsman et al. [44] have shown that averaging the weights of multiple models typically improves accuracy and robustness. Similarly, the integration of different sub-networks for the teacher is effective in improving the representation quality. At the same time, an improved teacher promotes the learning of students in return, fostering a positive feedback loop for pre-training. For the visualization of representations, please refer to Appendix E.2. + +# 5 Conclusion + +In this study, we tackle the challenge of efficiently and effectively pre-training models of various sizes within a single self-supervised learning session, facilitating model deployment given different resource constraints. We propose a novel self-supervised learning paradigm, termed POA, which features an integrated design combining self-distillation and once-for-all model generation. It allows for the simultaneous pre-training of models of multiple sizes through an elastic branch design. POA enables the direct generation of varied-sized models from a pretrained teacher, which are ready for downstream tasks without additional pretraining. This advantage significantly improves the deployment flexibility and facilitates our pre-trained model to achieve SOTA results across various vision tasks. Looking forward, we plan to extend POA to Multimodal Large Language Models, tapping into its vast potential for real-world AI product deployment. + +# References + +1. Ba, J., Kiros, J.R., Hinton, G.E.: Layer normalization. ArXiv abs/1607.06450 (2016) +2. Cai, H., Gan, C., Wang, T., Zhang, Z., Han, S.: Once for all: Train one network and specialize it for efficient deployment. In: International Conference on Learning Representations (2020) +3. Cai, Z., Vasconcelos, N.: Cascade r-cnn: High quality object detection and instance segmentation. IEEE Transactions on Pattern Analysis and Machine Intelligence 43, 1483-1498 (2019) +4. Caron, M., Misra, I., Mairal, J., Goyal, P., Bojanowski, P., Joulin, A.: Unsupervised learning of visual features by contrasting cluster assignments. In: Larochelle, H., Ranzato, M., Hadsell, R., Balcan, M., Lin, H. (eds.) Advances in Neural Information Processing Systems. vol. 33, pp. 9912-9924. Curran Associates, Inc. (2020) +5. Caron, M., Touvron, H., Misra, I., J'egou, H., Mairal, J., Bojanowski, P., Joulin, A.: Emerging properties in self-supervised vision transformers. 2021 IEEE/CVF International Conference on Computer Vision (ICCV) pp. 9630-9640 (2021) +6. Chen, T., Kornblith, S., Norouzi, M., Hinton, G.E.: A simple framework for contrastive learning of visual representations. International Conference on Machine Learning (2021) +7. Chen, T., Zhai, X., Ritter, M., Lucic, M., Houlsby, N.: Self-supervised gans via auxiliary rotation loss. 2019 IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR) pp. 12146-12155 (2018) +8. Chen, X., Fan, H., Girshick, R., He, K.: Improved baselines with momentum contrastive learning. arXiv preprint arXiv:2003.04297 (2020) +9. Chen, Y., Liu, Y., Jiang, D., Zhang, X., Dai, W., Xiong, H., Tian, Q.: Sdae: Self-distillated masked autoencoder. In: European Conference on Computer Vision (2022) +10. Cheng, X., Chen, J., Wang, R.: Modified dual attention triplet-supervised hashing network for image retrieval. Signal, Image and Video Processing 18, 1939-1948 (2023) +1. Cuturi, M.: Sinkhorn distances: Lightspeed computation of optimal transport. In: Neural Information Processing Systems (2013) +2. Deng, J., Dong, W., Socher, R., Li, L.J., Li, K., Fei-Fei, L.: ImageNet: A LargeScale Hierarchical Image Database. In: CVPR 2009 (2009) +3. Dong, C., Loy, C.C., He, K., Tang, X.: Image super-resolution using deep convolutional networks. IEEE Transactions on Pattern Analysis and Machine Intelligence 38, 295-307 (2014) +4. Dosovitskiy, A., Beyer, L., Kolesnikov, A., Weissenborn, D., Zhai, X., Unterthiner, T., Dehghani, M., Minderer, M., Heigold, G., Gelly, S., Uszkoreit, J., Houlsby, N.: An image is worth 16x16 words: Transformers for image recognition at scale. International Conference on Learning Representations (ICLR) (2021) +5. Fang, Z., Wang, J., Wang, L., Zhang, L., Yang, Y., Liu, Z.: Seed: Self-supervised distillation for visual representation. International Conference on Learning Representations (2021) +6. Gao, Y., Zhuang, J.X., Lin, S., Cheng, H., Sun, X., Li, K., Shen, C.: Disco: Remedying self-supervised learning on lightweight models with distilled contrastive learning. In: European Conference on Computer Vision (ECCV) (2022) +7. Grill, J.B., Strub, F., Altché, F., Tallec, C., Richemond, P., Buchatskaya, E., Doersch, C., Avila Pires, B., Guo, Z., Gheshlaghi Azar, M., et al.: Bootstrap your own + +latent-a new approach to self-supervised learning. Advances in neural information processing systems 33, 21271-21284 (2020) +18. He, K., Chen, X., Xie, S., Li, Y., Doll'ar, P., Girshick, R.B.: Masked autoencoders are scalable vision learners. 2022 IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR) pp. 15979-15988 (2021) +19. He, K., Fan, H., Wu, Y., Xie, S., Girshick, R.B.: Momentum contrast for unsupervised visual representation learning. 2020 IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR) pp. 9726-9735 (2019) +20. He, K., Gkioxari, G., Dolkar, P., Girshick, R.B.: Mask r-cnn. 2017 IEEE International Conference on Computer Vision (ICCV) pp. 2980-2988 (2017) +21. He, K., Zhang, X., Ren, S., Sun, J.: Deep residual learning for image recognition. 2016 IEEE Conference on Computer Vision and Pattern Recognition (CVPR) pp. 770-778 (2015) +22. Hinton, G.E., Vinyals, O., Dean, J.: Distilling the knowledge in a neural network. ArXiv abs/1503.02531 (2015) +23. Hou, L., Shen, H., Cao, Q., Cheng, X.: Self-supervised gans with label augmentation. ArXiv abs/2106.08601 (2021) +24. Hu, C., Li, X., Liu, D., Wu, H., Chen, X., Wang, J., Liu, X.: Teacher-student architecture for knowledge distillation: A survey. ArXiv abs/2308.04268 (2023) +25. Huang, Z., Jin, X., Lu, C., Hou, Q., Cheng, M.M., Fu, D., Shen, X., Feng, J.: Contrastive masked autoencoders are stronger vision learners. IEEE transactions on pattern analysis and machine intelligence PP (2022) +26. Li, J., Wang, Y., ZHANG, X., Chen, Y., Jiang, D., Dai, W., Li, C., Xiong, H., Tian, Q.: Progressively compressed auto-encoder for self-supervised representation learning. In: The Eleventh International Conference on Learning Representations (2023) +27. Li, Z., Zhu, Y., Yang, F., Li, W., Zhao, C., Chen, Y., Chen, Z., Xie, J., Wu, L., Zhao, R., Tang, M., Wang, J.: Univip: A unified framework for self-supervised visual pre-training. 2022 IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR) pp. 14607-14616 (2022) +28. Lin, J., Rao, Y., Lu, J., Zhou, J.: Runtime neural pruning. In: Guyon, I., Luxburg, U.V., Bengio, S., Wallach, H., Fergus, R., Vishwanathan, S., Garnett, R. (eds.) Advances in Neural Information Processing Systems. vol. 30. Curran Associates, Inc. (2017) +29. Lin, T.Y., Maire, M., Belongie, S.J., Hays, J., Perona, P., Ramanan, D., Dóllar, P., Zitnick, C.L.: Microsoft coco: Common objects in context. In: European Conference on Computer Vision (2014) +30. Liu, Z., Lin, Y., Cao, Y., Hu, H., Wei, Y., Zhang, Z., Lin, S., Guo, B.: Swin transformer: Hierarchical vision transformer using shifted windows. 2021 IEEE/CVF International Conference on Computer Vision (ICCV) pp. 9992-10002 (2021) +31. Loshchilov, I., Hutter, F.: Fixing weight decay regularization in adam. ArXiv abs/1711.05101 (2017) +32. van den Oord, A., Li, Y., Vinyals, O.: Representation learning with contrastive predictive coding. ArXiv abs/1807.03748 (2018) +33. Oquab, M., Darcet, T., Moutakanni, T., Vo, H.V., Szafraniec, M., Khalidov, V., Fernandez, P., HAZIZA, D., Massa, F., El-Nouby, A., Assran, M., Ballas, N., Galuba, W., Howes, R., Huang, P.Y., Li, S.W., Misra, I., Rabbat, M., Sharma, V., Synnaeve, G., Xu, H., Jegou, H., Mairal, J., Labatut, P., Joulin, A., Bojanowski, P.: DINOv2: Learning robust visual features without supervision. Transactions on Machine Learning Research (2024) + +34. Ruan, Y., Singh, S., Morningstar, W.R., Alemi, A.A., Ioffe, S., Fischer, I., Dillon, J.V.: Weighted ensemble self-supervised learning. In: The Eleventh International Conference on Learning Representations (2023) +35. Song, K., Xie, J., Zhang, S., Luo, Z.: Multi-mode online knowledge distillation for self-supervised visual representation learning. 2023 IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR) pp. 11848-11857 (2023) +36. Tao, C., Wang, H., Zhu, X., Dong, J., Song, S., Huang, G., Dai, J.: Exploring the equivalence of siamese self-supervised learning via a unified gradient framework. 2022 IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR) pp. 14411-14420 (2021) +37. Team, G.G.: Introducing gemini: our largest and most capable ai model. Google (2023) +38. Tian, K., Jiang, Y., qishuai diao, Lin, C., Wang, L., Yuan, Z.: Designing BERT for convolutional networks: Sparse and hierarchical masked modeling. In: The Eleventh International Conference on Learning Representations (2023) +39. Tomasev, N., Bica, I., McWilliams, B., Buesing, L., Pascanu, R., Blundell, C., Mitrovic, J.: Pushing the limits of self-supervised resnets: Can we outperform supervised learning without labels onImagenet? arXiv preprint arXiv:2201.05119 (2022) +40. Touvron, H., Cord, M., Oquab, M., Bojanowski, P., Verbeek, J., Jegou, H.: Cotraining 2l submodels for visual recognition. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR). pp. 11701-11710 (June 2023) +41. Wang, X., Huang, Y., Zeng, D., Qi, G.J.: Caco: Both positive and negative samples are directly learnable via cooperative-adversarial contrastive learning. IEEE Transactions on Pattern Analysis and Machine Intelligence 45, 10718-10730 (2022) +42. Wang, X., Yu, F., Dou, Z.Y., Gonzalez, J.: Skipnet: Learning dynamic routing in convolutional networks. ArXiv abs/1711.09485 (2017) +43. Wang, X., Zhang, R., Shen, C., Kong, T., Li, L.: Dense contrastive learning for self-supervised visual pre-training. 2021 IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR) pp. 3023-3032 (2020) +44. Wortsman, M., Ilharco, G., Gadre, S.Y., Roelofs, R., Gontijo-Lopes, R., Morcos, A.S., Namkoong, H., Farhadi, A., Carmon, Y., Kornblith, S., Schmidt, L.: Model soups: averaging weights of multiple fine-tuned models improves accuracy without increasing inference time. In: International Conference on Machine Learning (ICML) (2022) +45. Wu, H., Xu, J., Wang, J., Long, M.: Autoformer: Decomposition transformers with auto-correlation for long-term series forecasting. In: Neural Information Processing Systems (2021) +46. Xiao, T., Liu, Y., Zhou, B., Jiang, Y., Sun, J.: Unified perceptual parsing for scene understanding. ArXiv abs/1807.10221 (2018) +47. Xie, Z., Lin, Y., Yao, Z., Zhang, Z., Dai, Q., Cao, Y., Hu, H.: Self-supervised learning with swin transformers. ArXiv abs/2105.04553 (2021) +48. Xie, Z., Zhang, Z., Cao, Y., Lin, Y., Bao, J., Yao, Z., Dai, Q., Hu, H.: Simmim: a simple framework for masked image modeling. 2022 IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR) pp. 9643-9653 (2021) +49. Yan, B., Jiang, Y., Sun, P., Wang, D., Yuan, Z., Luo, P., Lu, H.: Towards grand unification of object tracking. In: European Conference on Computer Vision (2022) +50. Yan, C., Chang, X., Li, Z., Yao, L., Luo, M., Zheng, Q.: Masked distillation advances self-supervised transformer architecture search. In: The Twelfth Interna + +tional Conference on Learning Representations (2024), https://openreview.net/ forum?id=LUpC8KTvdV +51. Yu, F., Huang, K., Wang, M., Cheng, Y., Chu, W., Cui, L.: Width & depth pruning for vision transformers. In: AAAI Conference on Artificial Intelligence (2022) +52. Yu, J., Li, X., Koh, J.Y., Zhang, H., Pang, R., Qin, J., Ku, A., Xu, Y., Baldridge, J., Wu, Y.: Vector-quantized image modeling with improved VQGAN. In: International Conference on Learning Representations (2022) +53. Zhang, Y., Zhong, Q., Ma, L., Xie, D., Pu, S.: Learning incremental triplet margin for person re-identification. AAAI Conference on Artificial Intelligence (2019) +54. Zhou, B., Zhao, H., Puig, X., Fidler, S., Barriuso, A., Torralba, A.: Scene parsing through ade20k dataset. 2017 IEEE Conference on Computer Vision and Pattern Recognition (CVPR) pp. 5122-5130 (2017) +55. Zhou, J., Wei, C., Wang, H., Shen, W., Xie, C., Yuille, A., Kong, T.: ibot: Image bert pre-training with online tokenizer. International Conference on Learning Representations (ICLR) (2022) \ No newline at end of file diff --git a/2024/POA_ Pre-training Once for Models of All Sizes/images.zip b/2024/POA_ Pre-training Once for Models of All Sizes/images.zip new file mode 100644 index 0000000000000000000000000000000000000000..cacd09677cdd4f0bdea36c1d3ad2d656368cc2d1 --- /dev/null +++ b/2024/POA_ Pre-training Once for Models of All Sizes/images.zip @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:d954c210fe4800bd6ec6041b2a892ce3df4543e173c45b878ebf0a92f4cf71f0 +size 494184 diff --git a/2024/POA_ Pre-training Once for Models of All Sizes/layout.json b/2024/POA_ Pre-training Once for Models of All Sizes/layout.json new file mode 100644 index 0000000000000000000000000000000000000000..00965a39ff9b61621d0eb8348adf5f5c32f96439 --- /dev/null +++ b/2024/POA_ Pre-training Once for Models of All Sizes/layout.json @@ -0,0 +1,9171 @@ +{ + "pdf_info": [ + { + "para_blocks": [ + { + "bbox": [ + 140, + 112, + 474, + 129 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 140, + 112, + 474, + 129 + ], + "spans": [ + { + "bbox": [ + 140, + 112, + 474, + 129 + ], + "type": "text", + "content": "POA: Pre-training Once for Models of All Sizes" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 141, + 150, + 472, + 175 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 141, + 150, + 472, + 175 + ], + "spans": [ + { + "bbox": [ + 141, + 150, + 472, + 175 + ], + "type": "text", + "content": "Yingying Zhang, Xin Guo, Jiangwei Lao, Lei Yu, Lixiang Ru, Jian Wang, Guo Ye, Huimei He, Jingdong Chen, and Ming Yang" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 282, + 185, + 330, + 195 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 282, + 185, + 330, + 195 + ], + "spans": [ + { + "bbox": [ + 282, + 185, + 330, + 195 + ], + "type": "text", + "content": "Ant Group." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 255, + 196, + 361, + 206 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 255, + 196, + 361, + 206 + ], + "spans": [ + { + "bbox": [ + 255, + 196, + 361, + 206 + ], + "type": "text", + "content": "qichu.zyy@antgroup.com" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 160, + 236, + 452, + 477 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 160, + 236, + 452, + 477 + ], + "spans": [ + { + "bbox": [ + 160, + 236, + 452, + 477 + ], + "type": "text", + "content": "Abstract. Large-scale self-supervised pre-training has paved the way for one foundation model to handle many different vision tasks. Most pre-training methodologies train a single model of a certain size at one time. Nevertheless, various computation or storage constraints in real-world scenarios require substantial efforts to develop a series of models with different sizes to deploy. Thus, in this study, we propose a novel tri-branch self-supervised training framework, termed as POA (Pre-training Once for All), to tackle this aforementioned issue. Our approach introduces an innovative elastic student branch into a modern self-distillation paradigm. At each pre-training step, we randomly sample a sub-network from the original student to form the elastic student and train all branches in a self-distilling fashion. Once pre-trained, POA allows the extraction of pre-trained models of diverse sizes for downstream tasks. Remarkably, the elastic student facilitates the simultaneous pre-training of multiple models with different sizes, which also acts as an additional ensemble of models of various sizes to enhance representation learning. Extensive experiments, including k-nearest neighbors, linear probing evaluation and assessments on multiple downstream tasks demonstrate the effectiveness and advantages of our POA. It achieves state-of-the-art performance using ViT, Swin Transformer and ResNet backbones, producing around a hundred models with different sizes through a single pre-training session. The code is available at: https://github.com/Qichuzyy/POA." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 160, + 488, + 425, + 499 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 160, + 488, + 425, + 499 + ], + "spans": [ + { + "bbox": [ + 160, + 488, + 425, + 499 + ], + "type": "text", + "content": "Keywords: Self-supervised Learning " + }, + { + "bbox": [ + 160, + 488, + 425, + 499 + ], + "type": "inline_equation", + "content": "\\cdot" + }, + { + "bbox": [ + 160, + 488, + 425, + 499 + ], + "type": "text", + "content": " Pre-training Once for All" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 133, + 521, + 229, + 533 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 521, + 229, + 533 + ], + "spans": [ + { + "bbox": [ + 133, + 521, + 229, + 533 + ], + "type": "text", + "content": "1 Introduction" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 130, + 545, + 482, + 666 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 545, + 482, + 666 + ], + "spans": [ + { + "bbox": [ + 130, + 545, + 482, + 666 + ], + "type": "text", + "content": "Learning generalizable visual representations in a large model by self-supervised learning has delivered superior performance across a wide range of visual tasks [10, 13, 20, 46, 49, 53] in recent years. Nevertheless, when deployed to real-world applications, a large model has to be adapted to diverse resource limitations in terms of computation, storage, or power consumption, etc. For example, a well-engineered AI product typically comprises a suite of models tailored for varying scenarios, such as Gemini Nano, Pro and Ultra [37]. Given a large pre-trained model, common solutions to deploy it to multiple application scenarios with different resource constraints include additional weight pruning [28, 42, 51], knowledge distillation [22, 24], or even re-training a small network from scratch, which" + } + ] + } + ], + "index": 7 + } + ], + "discarded_blocks": [], + "page_size": [ + 612, + 792 + ], + "page_idx": 0 + }, + { + "para_blocks": [ + { + "bbox": [ + 130, + 116, + 482, + 154 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 116, + 482, + 154 + ], + "spans": [ + { + "bbox": [ + 130, + 116, + 482, + 154 + ], + "type": "text", + "content": "all require substantial development efforts. Consequently, this issue prompts a critical question: is it possible to pre-train once to produce multiple models with different sizes simultaneously, each delivering sufficiently good representations?" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 130, + 163, + 483, + 404 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 163, + 483, + 404 + ], + "spans": [ + { + "bbox": [ + 130, + 163, + 483, + 404 + ], + "type": "text", + "content": "To address this challenge, we introduce a new paradigm of self-supervised learning, called POA (Pre-training Once for All). POA is built upon the prevalent teacher-student self-distillation framework [5,33,55], with an additional innovative elastic student branch. The elastic student branch embeds a series of sub-networks through parameter sharing, upon the observation that smaller-sized models are sub-networks of a larger-sized one for modern network structures [14,21,30]. Moreover, the parameters of this branch are shared with the original, or intact student. At each pre-training step, we randomly sample a subset of parameters from the intact student to form the corresponding elastic student. Both the original intact student and the elastic student are trained to emulate the output of the teacher network. The teacher itself is continually refined via an exponential moving average (EMA) of the student's parameters, including the sampled elastic student. The elastic student facilitates effective and efficient pre-training on different subsets of parameters, leading to the successful extraction of high-performance sub-networks from the pre-trained teacher for subsequent downstream scenarios. It also acts as a form of training regularization by enforcing the outputs to match between the teacher and various sub-networks, contributing to a stable training process. Additionally, by serving as an ensemble of different sub-networks across different pre-training steps, the elastic student improves the representation learning [44]." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 130, + 414, + 303, + 629 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 414, + 303, + 629 + ], + "spans": [ + { + "bbox": [ + 130, + 414, + 303, + 629 + ], + "type": "text", + "content": "To the best of our knowledge, POA represents the first self-supervised learning method capable of training multiple-sized models concurrently, each obtaining high-quality representations for different resource constraints without further pre-training. Figure 1 displays the k-nearest neighbors (k-NN) evaluation results for 143 sub-networks extracted from a ViT-L model [14] pre-trained by POA. By choosing different elastic widths and depths, which will be explained in Sec. 3.1, the pre-trained teacher model can generate a sufficient number of candidate sub-networks for the selection of the suitable model tailored for downstream applications according to the" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 130, + 630, + 482, + 666 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 630, + 482, + 666 + ], + "spans": [ + { + "bbox": [ + 130, + 630, + 482, + 666 + ], + "type": "text", + "content": "computational resources available. Notably, each sub-network is well-trained and exhibits superior performance, thanks to our meticulous design on same-view distillation, as validated in Sec. 4.4. In particular, the ViT-S, ViT-B and ViT-" + } + ] + } + ], + "index": 5 + }, + { + "type": "image", + "bbox": [ + 309, + 418, + 484, + 577 + ], + "blocks": [ + { + "bbox": [ + 309, + 418, + 484, + 577 + ], + "lines": [ + { + "bbox": [ + 309, + 418, + 484, + 577 + ], + "spans": [ + { + "bbox": [ + 309, + 418, + 484, + 577 + ], + "type": "image", + "image_path": "58642ae34be5e2e6be6ff4bb4893376192d54c89b6fa430015f321e75fff488a.jpg" + } + ] + } + ], + "index": 6, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 307, + 586, + 483, + 620 + ], + "lines": [ + { + "bbox": [ + 307, + 586, + 483, + 620 + ], + "spans": [ + { + "bbox": [ + 307, + 586, + 483, + 620 + ], + "type": "text", + "content": "Fig.1: The k-NN evaluation accuracy of 143 elastic ViTs derived from the ViT-L/16 teacher model pre-trained with POA." + } + ] + } + ], + "index": 7, + "angle": 0, + "type": "image_caption" + } + ], + "index": 6 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "spans": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "type": "text", + "content": "2" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 166, + 91, + 229, + 102 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 166, + 91, + 229, + 102 + ], + "spans": [ + { + "bbox": [ + 166, + 91, + 229, + 102 + ], + "type": "text", + "content": "Y. Zhang et al." + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 1 + }, + { + "para_blocks": [ + { + "bbox": [ + 130, + 116, + 479, + 140 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 116, + 479, + 140 + ], + "spans": [ + { + "bbox": [ + 130, + 116, + 479, + 140 + ], + "type": "text", + "content": "L models set new benchmarks, achieving the state-of-the-art (SOTA) results compared with those pre-trained by existing methods [33, 34, 55]." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 130, + 140, + 482, + 236 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 140, + 482, + 236 + ], + "spans": [ + { + "bbox": [ + 130, + 140, + 482, + 236 + ], + "type": "text", + "content": "To rigorously evaluate the efficacy of our approach, we conduct extensive experiments using three widely-used backbone architectures, i.e., ViT [14], Swin Transformer [30], and ResNet [21]. Each backbone is pre-trained on ImageNet-1K dataset and assessed using k-NN and linear probing classification evaluation, as well as downstream dense prediction tasks such as object detection and semantic segmentation. Our method achieves SOTA accuracy across various model sizes with a single pre-training session. The technical contributions of this work are summarized as follows:" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 138, + 243, + 479, + 312 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 243, + 479, + 312 + ], + "spans": [ + { + "bbox": [ + 138, + 243, + 479, + 312 + ], + "type": "text", + "content": "- To the best of our knowledge, POA is the first pre-training paradigm that integrates unsupervised representation learning and once-for-all model generation within a single pre-training session. It tackles the Pre-training-Once-for-All challenge which has been seldomly explored by the community but is of great practical significance for real-world deployment that usually requires a suite of models." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 138, + 316, + 479, + 384 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 316, + 479, + 384 + ], + "spans": [ + { + "bbox": [ + 138, + 316, + 479, + 384 + ], + "type": "text", + "content": "- We propose a novel and elegant component called Elastic Student, featuring a range of elastic operators that enable POA to be compatible with popular backbone structures including ViT, Swin Transformer and ResNet. It provides the capability to generate models of diverse sizes. Furthermore, it serves as a model ensemble to smooth training process and improve learned representations." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 138, + 387, + 479, + 456 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 387, + 479, + 456 + ], + "spans": [ + { + "bbox": [ + 138, + 387, + 479, + 456 + ], + "type": "text", + "content": "- Through thorough assessments using k-NN, linear probing and downstream dense task evaluation, our approach exhibits superior performance over existing state-of-the-art pre-training methods across multiple metrics. Moreover, we compare our POA with Self-Supervised Distillation (SEED) [15], a knowledge distillation method designed especially for self-supervised learning, further validating POA's effectiveness." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 132, + 476, + 237, + 489 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 476, + 237, + 489 + ], + "spans": [ + { + "bbox": [ + 132, + 476, + 237, + 489 + ], + "type": "text", + "content": "2 Related Work" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 131, + 502, + 285, + 514 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 131, + 502, + 285, + 514 + ], + "spans": [ + { + "bbox": [ + 131, + 502, + 285, + 514 + ], + "type": "text", + "content": "2.1 Self-supervised Learning" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 130, + 521, + 479, + 592 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 521, + 479, + 592 + ], + "spans": [ + { + "bbox": [ + 130, + 521, + 479, + 592 + ], + "type": "text", + "content": "Self-supervised learning (SSL) frameworks commonly fall into two categories, generative and discriminative SSL. Most generative SSL approaches [7,9,18,23, 25,26,38,48,52] focus on learning image representations directly in pixel space. On the other hand, discriminative SSL methods [4,6,19,27,36,41,43,47] aim to learn representations by pulling those of different views of the same image closer while separating the representations of views from different images." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 130, + 594, + 480, + 664 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 594, + 480, + 664 + ], + "spans": [ + { + "bbox": [ + 130, + 594, + 480, + 664 + ], + "type": "text", + "content": "Contrastive learning (CL) with the InfoNCE loss [32] has emerged as a popular approach for discriminative SSL methods, attracting significant research interests in recent years. Although CL methods prevent the collapse of network representations through the use of negative samples, they still suffer from the dimensional collapse, where representations tend to collapse into a low-dimensional manifold. Grill et al. introduced a distillation-based asymmetric" + } + ] + } + ], + "index": 10 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 425, + 91, + 447, + 100 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 425, + 91, + 447, + 100 + ], + "spans": [ + { + "bbox": [ + 425, + 91, + 447, + 100 + ], + "type": "text", + "content": "POA" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 474, + 91, + 481, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 474, + 91, + 481, + 100 + ], + "spans": [ + { + "bbox": [ + 474, + 91, + 481, + 100 + ], + "type": "text", + "content": "3" + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 2 + }, + { + "para_blocks": [ + { + "bbox": [ + 130, + 116, + 482, + 236 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 116, + 482, + 236 + ], + "spans": [ + { + "bbox": [ + 130, + 116, + 482, + 236 + ], + "type": "text", + "content": "framework known as BYOL [17], which circumvents collapse without self-labeling or contrastive loss relying on negative samples. Following this work, distillation-based frameworks have become a prevailing trend in self-supervised learning. These frameworks [8, 16, 47] often merge with others to enhance overall performance. DINO [5] presented a simple self-distillation framework and has demonstrated impressive results in ViT pre-training. Subsequent works [33, 55] further improved the pre-training performance via masked token or the novel clustering design. Given the substantial benefits of distillation-based methods over other SSL techniques, we have developed our POA SSL framework upon these successful methodologies." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 132, + 254, + 277, + 266 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 254, + 277, + 266 + ], + "spans": [ + { + "bbox": [ + 132, + 254, + 277, + 266 + ], + "type": "text", + "content": "2.2 Dynamic Architecture." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 130, + 274, + 482, + 574 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 274, + 482, + 574 + ], + "spans": [ + { + "bbox": [ + 130, + 274, + 482, + 574 + ], + "type": "text", + "content": "Recently, Chen et al. proposed AutoFormer [45], which trained a supernet to support the effective search of optimal sub-network under some specific parameter number constraints. On the basis of [45], MaskTAS [50] introduced a self-supervised transformer architecture search method. Cai et al. [2] trained a network that accommodates various architectural configurations to reduce the training expense. Their methodologies enable the extraction of a specialized sub-network from the trained supernet. The design of the elastic student in our POA SSL is inspired by the weight-sharing strategy employed in these neural architecture search (NAS) methods. However, our implementation differs significantly due to the distinct purpose from NAS. Specifically, NAS aims to discover the optimal architecture within certain parameter constraints, which typically involves a huge number of sub-networks (more than " + }, + { + "bbox": [ + 130, + 274, + 482, + 574 + ], + "type": "inline_equation", + "content": "10^{16}" + }, + { + "bbox": [ + 130, + 274, + 482, + 574 + ], + "type": "text", + "content": ") in the search space. Given the limitations on the number of training iterations and the network parameter capacity, it is challenging to ensure high performance across all sub-networks. After training the supernet, NAS requires a subsequent phase of searching and re-training to finalize the output model. In contrast, our POA SSL defines a sufficiently large yet compact set of sub-networks with elastic depths and widths for the purpose of pre-training models of various sizes via a single training session. Additional design on the same-view distillation guarantees that all elastic sub-networks within our framework are adequately and efficiently trained. As a result, our POA can readily extract a range of sub-networks from the teacher model without the need for extra pre-training, facilitating an easy selection of a suitable sub-network for different computational contexts. The design of elastic student is somewhat akin to the supervised learning method, Cosub [40]. The main difference is that Cosub simply skips blocks, making only the depth elastic." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 132, + 592, + 404, + 606 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 592, + 404, + 606 + ], + "spans": [ + { + "bbox": [ + 132, + 592, + 404, + 606 + ], + "type": "text", + "content": "3 POA Self-supervised Learning Framework" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 130, + 617, + 482, + 666 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 617, + 482, + 666 + ], + "spans": [ + { + "bbox": [ + 130, + 617, + 482, + 666 + ], + "type": "text", + "content": "Our primary goal is to pre-train models of multiple sizes via a single self-supervised pre-training session. To this end, we propose a novel SSL framework named POA, inspired by the latest progress in self-distillation techniques [5, 33, 55]. The architecture of POA is illustrated in Figure 2, encompassing a" + } + ] + } + ], + "index": 6 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "spans": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "type": "text", + "content": "4" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 166, + 91, + 228, + 102 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 166, + 91, + 228, + 102 + ], + "spans": [ + { + "bbox": [ + 166, + 91, + 228, + 102 + ], + "type": "text", + "content": "Y. Zhang et al." + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 3 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 135, + 114, + 479, + 234 + ], + "blocks": [ + { + "bbox": [ + 135, + 114, + 479, + 234 + ], + "lines": [ + { + "bbox": [ + 135, + 114, + 479, + 234 + ], + "spans": [ + { + "bbox": [ + 135, + 114, + 479, + 234 + ], + "type": "image", + "image_path": "f899a486424e9769904603e459e5736d7358d72eec370a022540471bbfad631c.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 130, + 243, + 480, + 309 + ], + "lines": [ + { + "bbox": [ + 130, + 243, + 480, + 309 + ], + "spans": [ + { + "bbox": [ + 130, + 243, + 480, + 309 + ], + "type": "text", + "content": "Fig. 2: Overview of the POA SSL: Given an image " + }, + { + "bbox": [ + 130, + 243, + 480, + 309 + ], + "type": "inline_equation", + "content": "x" + }, + { + "bbox": [ + 130, + 243, + 480, + 309 + ], + "type": "text", + "content": ", two augmented views " + }, + { + "bbox": [ + 130, + 243, + 480, + 309 + ], + "type": "inline_equation", + "content": "x_{a}" + }, + { + "bbox": [ + 130, + 243, + 480, + 309 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 130, + 243, + 480, + 309 + ], + "type": "inline_equation", + "content": "x_{b}" + }, + { + "bbox": [ + 130, + 243, + 480, + 309 + ], + "type": "text", + "content": " are generated. These views are input into three branches: a teacher, an intact student, and an elastic student, the latter being derived from the intact student. POA optimizes distillation losses in a twofold manner: the intact and the elastic students are distilled from the teacher using the cross-view data respectively, and additionally, the elastic student is distilled from the intact student using the same-view data." + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_caption" + } + ], + "index": 2 + }, + { + "bbox": [ + 130, + 314, + 482, + 494 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 314, + 482, + 494 + ], + "spans": [ + { + "bbox": [ + 130, + 314, + 482, + 494 + ], + "type": "text", + "content": "teacher, an intact student, an elastic student and their corresponding heads. The teacher is updated with an EMA of the students. The elastic student is a derivative of the intact one, with both the backbones and heads' parameters shared. We leverage distillation in two aspects: both the intact and elastic students are distilled from the teacher using different views of the same image, while the elastic student additionally learns from the intact student using the identical views. The cross-view distillation works as a form of representation learning, as introduced in [5, 33, 55]. Notably, in addition to the regular EMA update with only the intact student as [5, 33, 55], the elastic student provides a randomly-sampled sub-network at each pre-training step to participate in the teacher's EMA refinement. This procedure actually simulates an ensemble of multiple sub-networks, which is also proven to be beneficial in the realm of supervised learning [44]. The same-view distillation is a standard knowledge distillation between the intact and elastic students, promoting the quality of the elastic one. Experiments in Sec. 4.4 validate our design in details." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 132, + 512, + 289, + 525 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 512, + 289, + 525 + ], + "spans": [ + { + "bbox": [ + 132, + 512, + 289, + 525 + ], + "type": "text", + "content": "3.1 Design of Elastic Student" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 130, + 533, + 482, + 628 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 533, + 482, + 628 + ], + "spans": [ + { + "bbox": [ + 130, + 533, + 482, + 628 + ], + "type": "text", + "content": "The elastic student is a sub-network whose parameters are extracted from the intact student. In the context of a transformer backbone, the width refers to the dimensionality of the tokens, whereas for a convolutional backbone, the width indicates the number of channels. We denote the depth as the number of basic blocks in either a transformer or a convolutional network. Given the value of the width and depth, it yields a certain network structure. For simplicity, we focus on detailing the elastic design of the ViT in this section. For the similar elastic design of the Swin Transformer and ResNet, please refer to Appendix A." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 130, + 629, + 481, + 666 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 629, + 481, + 666 + ], + "spans": [ + { + "bbox": [ + 130, + 629, + 481, + 666 + ], + "type": "text", + "content": "A basic block of ViT mainly consists of a multi-head self-attention (MSA) module and a multi-layer perceptron (MLP) module. Layer Normalization (LN) [1] is applied before each module, with residual connections after each module." + } + ] + } + ], + "index": 7 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 425, + 91, + 447, + 100 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 425, + 91, + 447, + 100 + ], + "spans": [ + { + "bbox": [ + 425, + 91, + 447, + 100 + ], + "type": "text", + "content": "POA" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 474, + 91, + 480, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 474, + 91, + 480, + 100 + ], + "spans": [ + { + "bbox": [ + 474, + 91, + 480, + 100 + ], + "type": "text", + "content": "5" + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 4 + }, + { + "para_blocks": [ + { + "bbox": [ + 130, + 116, + 482, + 163 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 116, + 482, + 163 + ], + "spans": [ + { + "bbox": [ + 130, + 116, + 482, + 163 + ], + "type": "text", + "content": "As shown in the left part of Figure 3, an elastic block refers to a stack of elastic MSA, MLP, and LN after adjusting the width in the original basic block in ViT. In our approach, the elastic student branch is constructed by assembling a specific number of these elastic blocks at each training iteration." + } + ] + } + ], + "index": 2 + }, + { + "type": "image", + "bbox": [ + 157, + 165, + 455, + 276 + ], + "blocks": [ + { + "bbox": [ + 157, + 165, + 455, + 276 + ], + "lines": [ + { + "bbox": [ + 157, + 165, + 455, + 276 + ], + "spans": [ + { + "bbox": [ + 157, + 165, + 455, + 276 + ], + "type": "image", + "image_path": "998f376e1f2880a4ce99e0e8de331a39cf8d97fc432940a0dceff96ec234665c.jpg" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 130, + 283, + 481, + 308 + ], + "lines": [ + { + "bbox": [ + 130, + 283, + 481, + 308 + ], + "spans": [ + { + "bbox": [ + 130, + 283, + 481, + 308 + ], + "type": "text", + "content": "Fig. 3: Illustration of the elastic MSA in an elastic ViT block. To be concise, we simply exclude the projection layers that correspond to " + }, + { + "bbox": [ + 130, + 283, + 481, + 308 + ], + "type": "inline_equation", + "content": "K" + }, + { + "bbox": [ + 130, + 283, + 481, + 308 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 130, + 283, + 481, + 308 + ], + "type": "inline_equation", + "content": "V" + }, + { + "bbox": [ + 130, + 283, + 481, + 308 + ], + "type": "text", + "content": " in each head." + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "image_caption" + } + ], + "index": 3 + }, + { + "bbox": [ + 130, + 323, + 482, + 443 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 323, + 482, + 443 + ], + "spans": [ + { + "bbox": [ + 130, + 323, + 482, + 443 + ], + "type": "text", + "content": "**Elastic MSA** An original, or intact MSA module consists of three major components, i.e., input projection layers, an operator contains attention & concatenation, and an output projection layer. We define a projection layer as " + }, + { + "bbox": [ + 130, + 323, + 482, + 443 + ], + "type": "inline_equation", + "content": "(w^{*}, b^{*})" + }, + { + "bbox": [ + 130, + 323, + 482, + 443 + ], + "type": "text", + "content": ", where " + }, + { + "bbox": [ + 130, + 323, + 482, + 443 + ], + "type": "inline_equation", + "content": "w^{*}" + }, + { + "bbox": [ + 130, + 323, + 482, + 443 + ], + "type": "text", + "content": " is the linear transformation weight, " + }, + { + "bbox": [ + 130, + 323, + 482, + 443 + ], + "type": "inline_equation", + "content": "b^{*}" + }, + { + "bbox": [ + 130, + 323, + 482, + 443 + ], + "type": "text", + "content": " denotes the corresponding bias, and * indicates the name of the layer. As shown in the right part of Figure 3, given a token dimension of " + }, + { + "bbox": [ + 130, + 323, + 482, + 443 + ], + "type": "inline_equation", + "content": "D_{max} = N_h \\cdot D_h" + }, + { + "bbox": [ + 130, + 323, + 482, + 443 + ], + "type": "text", + "content": ", where " + }, + { + "bbox": [ + 130, + 323, + 482, + 443 + ], + "type": "inline_equation", + "content": "N_h" + }, + { + "bbox": [ + 130, + 323, + 482, + 443 + ], + "type": "text", + "content": " is the number of attention heads and " + }, + { + "bbox": [ + 130, + 323, + 482, + 443 + ], + "type": "inline_equation", + "content": "D_h" + }, + { + "bbox": [ + 130, + 323, + 482, + 443 + ], + "type": "text", + "content": " is the head dimension, an input sequence " + }, + { + "bbox": [ + 130, + 323, + 482, + 443 + ], + "type": "inline_equation", + "content": "z \\in \\mathbb{R}^{T \\times D_{max}}" + }, + { + "bbox": [ + 130, + 323, + 482, + 443 + ], + "type": "text", + "content": " with length " + }, + { + "bbox": [ + 130, + 323, + 482, + 443 + ], + "type": "inline_equation", + "content": "T" + }, + { + "bbox": [ + 130, + 323, + 482, + 443 + ], + "type": "text", + "content": " is initially projected to form the queries " + }, + { + "bbox": [ + 130, + 323, + 482, + 443 + ], + "type": "inline_equation", + "content": "Q \\in \\mathbb{R}^{T \\times D_h}" + }, + { + "bbox": [ + 130, + 323, + 482, + 443 + ], + "type": "text", + "content": ", keys " + }, + { + "bbox": [ + 130, + 323, + 482, + 443 + ], + "type": "inline_equation", + "content": "K \\in \\mathbb{R}^{T \\times D_h}" + }, + { + "bbox": [ + 130, + 323, + 482, + 443 + ], + "type": "text", + "content": ", and values " + }, + { + "bbox": [ + 130, + 323, + 482, + 443 + ], + "type": "inline_equation", + "content": "V \\in \\mathbb{R}^{T \\times D_h}" + }, + { + "bbox": [ + 130, + 323, + 482, + 443 + ], + "type": "text", + "content": " by the intact MSA. To generate elastic MSA, we define " + }, + { + "bbox": [ + 130, + 323, + 482, + 443 + ], + "type": "inline_equation", + "content": "\\mathrm{M} + 1" + }, + { + "bbox": [ + 130, + 323, + 482, + 443 + ], + "type": "text", + "content": " elastic widths, including " + }, + { + "bbox": [ + 130, + 323, + 482, + 443 + ], + "type": "inline_equation", + "content": "D_{max}" + }, + { + "bbox": [ + 130, + 323, + 482, + 443 + ], + "type": "text", + "content": ", spaced at intervals of " + }, + { + "bbox": [ + 130, + 323, + 482, + 443 + ], + "type": "inline_equation", + "content": "D_h" + }, + { + "bbox": [ + 130, + 323, + 482, + 443 + ], + "type": "text", + "content": " as follows:" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 192, + 453, + 481, + 465 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 192, + 453, + 481, + 465 + ], + "spans": [ + { + "bbox": [ + 192, + 453, + 481, + 465 + ], + "type": "interline_equation", + "content": "D _ {i} = \\left(N _ {h} - i\\right) \\cdot D _ {h}, \\quad \\forall i \\in \\{0, 1, \\dots , M \\}, \\quad M < N _ {h}. \\tag {1}", + "image_path": "02135369c05a5e4cb0c8392a19ed46fb69b1db04f76d7a8c778e4fe1f52c3054.jpg" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 130, + 474, + 482, + 571 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 474, + 482, + 571 + ], + "spans": [ + { + "bbox": [ + 130, + 474, + 482, + 571 + ], + "type": "text", + "content": "For each elastic width " + }, + { + "bbox": [ + 130, + 474, + 482, + 571 + ], + "type": "inline_equation", + "content": "D_{i}" + }, + { + "bbox": [ + 130, + 474, + 482, + 571 + ], + "type": "text", + "content": ", the weight " + }, + { + "bbox": [ + 130, + 474, + 482, + 571 + ], + "type": "inline_equation", + "content": "w_{i}^{a1} \\in \\mathbb{R}^{D_{h} \\times D_{i}}" + }, + { + "bbox": [ + 130, + 474, + 482, + 571 + ], + "type": "text", + "content": " and bias " + }, + { + "bbox": [ + 130, + 474, + 482, + 571 + ], + "type": "inline_equation", + "content": "b_{i}^{a1} \\in \\mathbb{R}^{D_{h}}" + }, + { + "bbox": [ + 130, + 474, + 482, + 571 + ], + "type": "text", + "content": " that generate " + }, + { + "bbox": [ + 130, + 474, + 482, + 571 + ], + "type": "inline_equation", + "content": "Q" + }, + { + "bbox": [ + 130, + 474, + 482, + 571 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 130, + 474, + 482, + 571 + ], + "type": "inline_equation", + "content": "K" + }, + { + "bbox": [ + 130, + 474, + 482, + 571 + ], + "type": "text", + "content": ", and " + }, + { + "bbox": [ + 130, + 474, + 482, + 571 + ], + "type": "inline_equation", + "content": "V" + }, + { + "bbox": [ + 130, + 474, + 482, + 571 + ], + "type": "text", + "content": " of each head are extracted from the corresponding input projection layer (" + }, + { + "bbox": [ + 130, + 474, + 482, + 571 + ], + "type": "inline_equation", + "content": "w^{a1}" + }, + { + "bbox": [ + 130, + 474, + 482, + 571 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 130, + 474, + 482, + 571 + ], + "type": "inline_equation", + "content": "b^{a1}" + }, + { + "bbox": [ + 130, + 474, + 482, + 571 + ], + "type": "text", + "content": ") in the intact MSA, as " + }, + { + "bbox": [ + 130, + 474, + 482, + 571 + ], + "type": "inline_equation", + "content": "w_{i}^{a1} = w^{a1}[:,: D_{i}] \\cdot \\alpha_{i}" + }, + { + "bbox": [ + 130, + 474, + 482, + 571 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 130, + 474, + 482, + 571 + ], + "type": "inline_equation", + "content": "b_{i}^{a1} = b^{a1}" + }, + { + "bbox": [ + 130, + 474, + 482, + 571 + ], + "type": "text", + "content": ". Here, " + }, + { + "bbox": [ + 130, + 474, + 482, + 571 + ], + "type": "inline_equation", + "content": "\\alpha_{i}" + }, + { + "bbox": [ + 130, + 474, + 482, + 571 + ], + "type": "text", + "content": " represents the scaling factor to cope with the reduction of the input dimension, calculated as " + }, + { + "bbox": [ + 130, + 474, + 482, + 571 + ], + "type": "inline_equation", + "content": "\\alpha_{i} = D_{max} / D_{i}" + }, + { + "bbox": [ + 130, + 474, + 482, + 571 + ], + "type": "text", + "content": ". As the reduction of width, the number of attention heads in the elastic MAS is naturally reduced to " + }, + { + "bbox": [ + 130, + 474, + 482, + 571 + ], + "type": "inline_equation", + "content": "N_{h} - i" + }, + { + "bbox": [ + 130, + 474, + 482, + 571 + ], + "type": "text", + "content": ". Similarly, for the output projection layer (" + }, + { + "bbox": [ + 130, + 474, + 482, + 571 + ], + "type": "inline_equation", + "content": "w^{a2}" + }, + { + "bbox": [ + 130, + 474, + 482, + 571 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 130, + 474, + 482, + 571 + ], + "type": "inline_equation", + "content": "b^{a2}" + }, + { + "bbox": [ + 130, + 474, + 482, + 571 + ], + "type": "text", + "content": "), the weight " + }, + { + "bbox": [ + 130, + 474, + 482, + 571 + ], + "type": "inline_equation", + "content": "w_{i}^{a2} \\in \\mathbb{R}^{D_{i} \\times D_{i}}" + }, + { + "bbox": [ + 130, + 474, + 482, + 571 + ], + "type": "text", + "content": " and bias " + }, + { + "bbox": [ + 130, + 474, + 482, + 571 + ], + "type": "inline_equation", + "content": "b_{i}^{a2} \\in \\mathbb{R}^{D_{i}}" + }, + { + "bbox": [ + 130, + 474, + 482, + 571 + ], + "type": "text", + "content": " are extracted as:" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 210, + 578, + 481, + 594 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 210, + 578, + 481, + 594 + ], + "spans": [ + { + "bbox": [ + 210, + 578, + 481, + 594 + ], + "type": "interline_equation", + "content": "w _ {i} ^ {a 2} = w ^ {a 2} [: D _ {i},: D _ {i} ] \\cdot \\alpha_ {i} \\quad b _ {i} ^ {a 2} = b ^ {a 2} [: D _ {i} ]. \\tag {2}", + "image_path": "5b74fef333f81da37bcb89c411b1c8ee9ebbd3be699a62264a8f1ad1ab15b425.jpg" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 130, + 605, + 482, + 666 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 605, + 482, + 666 + ], + "spans": [ + { + "bbox": [ + 130, + 605, + 482, + 666 + ], + "type": "text", + "content": "**Elastic MLP** An original, or intact MLP module in the ViT block contains two projection layers. The first layer (" + }, + { + "bbox": [ + 130, + 605, + 482, + 666 + ], + "type": "inline_equation", + "content": "w^{m1}, b^{m1}" + }, + { + "bbox": [ + 130, + 605, + 482, + 666 + ], + "type": "text", + "content": ") expands the dimension of embedding by a factor of " + }, + { + "bbox": [ + 130, + 605, + 482, + 666 + ], + "type": "inline_equation", + "content": "s" + }, + { + "bbox": [ + 130, + 605, + 482, + 666 + ], + "type": "text", + "content": ", which is generally set to 4 in the ViT structure. The second layer (" + }, + { + "bbox": [ + 130, + 605, + 482, + 666 + ], + "type": "inline_equation", + "content": "w^{m2}, b^{m2}" + }, + { + "bbox": [ + 130, + 605, + 482, + 666 + ], + "type": "text", + "content": ") then projects it back to the original dimension. The parameters for both layers of the elastic MLP are extracted in a manner analogous to that" + } + ] + } + ], + "index": 9 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "spans": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "type": "text", + "content": "6" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 166, + 91, + 229, + 102 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 166, + 91, + 229, + 102 + ], + "spans": [ + { + "bbox": [ + 166, + 91, + 229, + 102 + ], + "type": "text", + "content": "Y. Zhang et al." + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 5 + }, + { + "para_blocks": [ + { + "bbox": [ + 132, + 116, + 286, + 128 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 116, + 286, + 128 + ], + "spans": [ + { + "bbox": [ + 132, + 116, + 286, + 128 + ], + "type": "text", + "content": "described in Equation 2 as follows:" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 197, + 137, + 417, + 150 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 197, + 137, + 417, + 150 + ], + "spans": [ + { + "bbox": [ + 197, + 137, + 417, + 150 + ], + "type": "interline_equation", + "content": "w _ {i} ^ {m 1} = w ^ {m 1} \\left[ \\colon D _ {i} \\cdot s, \\colon D _ {i} \\right] \\cdot \\alpha_ {i} \\quad b _ {i} ^ {m 1} = b ^ {m 1} \\left[ \\colon D _ {i} \\cdot s \\right]", + "image_path": "adf183c824ebc011939e9c184658042bcf5e4f9753e2241f4c149da753c5e83c.jpg" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 197, + 147, + 480, + 167 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 197, + 147, + 480, + 167 + ], + "spans": [ + { + "bbox": [ + 197, + 147, + 480, + 167 + ], + "type": "interline_equation", + "content": "w _ {i} ^ {m 2} = w ^ {m 2} [: D _ {i},: D _ {i} \\cdot s ] \\cdot \\alpha_ {i} \\quad b _ {i} ^ {m 2} = b ^ {m 2} [: D _ {i} ]. \\tag {3}", + "image_path": "493540742ae65bdde1b4f4c88e214a47384cc985ddb37d761ef87237771609b4.jpg" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 132, + 176, + 479, + 200 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 176, + 479, + 200 + ], + "spans": [ + { + "bbox": [ + 132, + 176, + 479, + 200 + ], + "type": "text", + "content": "**Elastic LN** For elastic LN, we directly use the first " + }, + { + "bbox": [ + 132, + 176, + 479, + 200 + ], + "type": "inline_equation", + "content": "D_{i}" + }, + { + "bbox": [ + 132, + 176, + 479, + 200 + ], + "type": "text", + "content": " elements of the parameter inside the original LN, akin to the bias extraction in Equation 2." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 132, + 200, + 480, + 262 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 200, + 480, + 262 + ], + "spans": [ + { + "bbox": [ + 132, + 200, + 480, + 262 + ], + "type": "text", + "content": "To create a sub-network with " + }, + { + "bbox": [ + 132, + 200, + 480, + 262 + ], + "type": "inline_equation", + "content": "L_{i}" + }, + { + "bbox": [ + 132, + 200, + 480, + 262 + ], + "type": "text", + "content": " elastic blocks from the intact ViT comprising " + }, + { + "bbox": [ + 132, + 200, + 480, + 262 + ], + "type": "inline_equation", + "content": "L_{max}" + }, + { + "bbox": [ + 132, + 200, + 480, + 262 + ], + "type": "text", + "content": " blocks, we introduce a set of " + }, + { + "bbox": [ + 132, + 200, + 480, + 262 + ], + "type": "inline_equation", + "content": "N + 1" + }, + { + "bbox": [ + 132, + 200, + 480, + 262 + ], + "type": "text", + "content": " elastic depths, defined as " + }, + { + "bbox": [ + 132, + 200, + 480, + 262 + ], + "type": "inline_equation", + "content": "L_{i} = L_{max} - i" + }, + { + "bbox": [ + 132, + 200, + 480, + 262 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 132, + 200, + 480, + 262 + ], + "type": "inline_equation", + "content": "\\forall i \\in \\{0,1,\\dots,N\\}" + }, + { + "bbox": [ + 132, + 200, + 480, + 262 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 132, + 200, + 480, + 262 + ], + "type": "inline_equation", + "content": "N < L_{max}" + }, + { + "bbox": [ + 132, + 200, + 480, + 262 + ], + "type": "text", + "content": ". For a specific depth " + }, + { + "bbox": [ + 132, + 200, + 480, + 262 + ], + "type": "inline_equation", + "content": "L_{i}" + }, + { + "bbox": [ + 132, + 200, + 480, + 262 + ], + "type": "text", + "content": ", we select the corresponding blocks based on their block IDs at equal intervals. Each block ID " + }, + { + "bbox": [ + 132, + 200, + 480, + 262 + ], + "type": "inline_equation", + "content": "BID_{j}^{L_{i}}" + }, + { + "bbox": [ + 132, + 200, + 480, + 262 + ], + "type": "text", + "content": " that is activated for depth " + }, + { + "bbox": [ + 132, + 200, + 480, + 262 + ], + "type": "inline_equation", + "content": "L_{i}" + }, + { + "bbox": [ + 132, + 200, + 480, + 262 + ], + "type": "text", + "content": " can be formulated as:" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 194, + 270, + 480, + 295 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 194, + 270, + 480, + 295 + ], + "spans": [ + { + "bbox": [ + 194, + 270, + 480, + 295 + ], + "type": "interline_equation", + "content": "B I D _ {j} ^ {L _ {i}} = \\left\\lfloor \\frac {\\left(L _ {\\max } - 1\\right) \\cdot j}{L _ {i} - 1} \\right\\rfloor , \\quad \\forall j \\in \\{0, 1, \\dots , L _ {i} - 1 \\}. \\tag {4}", + "image_path": "4fc3515e49c97e979b28586a4b2e1fa9db7e823f74e785372df87ba5d3765e78.jpg" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 131, + 299, + 479, + 358 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 131, + 299, + 479, + 358 + ], + "spans": [ + { + "bbox": [ + 131, + 299, + 479, + 358 + ], + "type": "text", + "content": "Consequently, by combining elastic widths and depths, we can generate a total of " + }, + { + "bbox": [ + 131, + 299, + 479, + 358 + ], + "type": "inline_equation", + "content": "(N + 1) \\cdot (M + 1)" + }, + { + "bbox": [ + 131, + 299, + 479, + 358 + ], + "type": "text", + "content": " distinct sub-networks. For instance, by setting the elastic width to 384 and the elastic depth to 12, we can directly extract a ViT-S from the intact network such as ViT-L. During each iteration of the pre-training, we randomly select one of these sub-networks to serve as the elastic student branch." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 132, + 376, + 295, + 387 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 376, + 295, + 387 + ], + "spans": [ + { + "bbox": [ + 132, + 376, + 295, + 387 + ], + "type": "text", + "content": "3.2 Distillation between Views" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 131, + 396, + 480, + 503 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 131, + 396, + 480, + 503 + ], + "spans": [ + { + "bbox": [ + 131, + 396, + 480, + 503 + ], + "type": "text", + "content": "POA performs distillation across its three branches accordingly. Given a pair of globally augmented views of an input image " + }, + { + "bbox": [ + 131, + 396, + 480, + 503 + ], + "type": "inline_equation", + "content": "x" + }, + { + "bbox": [ + 131, + 396, + 480, + 503 + ], + "type": "text", + "content": ", denoted as " + }, + { + "bbox": [ + 131, + 396, + 480, + 503 + ], + "type": "inline_equation", + "content": "x_{a}" + }, + { + "bbox": [ + 131, + 396, + 480, + 503 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 131, + 396, + 480, + 503 + ], + "type": "inline_equation", + "content": "x_{b}" + }, + { + "bbox": [ + 131, + 396, + 480, + 503 + ], + "type": "text", + "content": ", the teacher encoder " + }, + { + "bbox": [ + 131, + 396, + 480, + 503 + ], + "type": "inline_equation", + "content": "E_{T}" + }, + { + "bbox": [ + 131, + 396, + 480, + 503 + ], + "type": "text", + "content": " extracts the feature " + }, + { + "bbox": [ + 131, + 396, + 480, + 503 + ], + "type": "inline_equation", + "content": "Z_{a} = E_{T}(x_{a})" + }, + { + "bbox": [ + 131, + 396, + 480, + 503 + ], + "type": "text", + "content": " using " + }, + { + "bbox": [ + 131, + 396, + 480, + 503 + ], + "type": "inline_equation", + "content": "x_{a}" + }, + { + "bbox": [ + 131, + 396, + 480, + 503 + ], + "type": "text", + "content": " as input. Simultaneously, " + }, + { + "bbox": [ + 131, + 396, + 480, + 503 + ], + "type": "inline_equation", + "content": "x_{b}" + }, + { + "bbox": [ + 131, + 396, + 480, + 503 + ], + "type": "text", + "content": " is fed into both the intact student encoder " + }, + { + "bbox": [ + 131, + 396, + 480, + 503 + ], + "type": "inline_equation", + "content": "E_{IS}" + }, + { + "bbox": [ + 131, + 396, + 480, + 503 + ], + "type": "text", + "content": " and the elastic student encoder " + }, + { + "bbox": [ + 131, + 396, + 480, + 503 + ], + "type": "inline_equation", + "content": "E_{ES}" + }, + { + "bbox": [ + 131, + 396, + 480, + 503 + ], + "type": "text", + "content": ", producing the features " + }, + { + "bbox": [ + 131, + 396, + 480, + 503 + ], + "type": "inline_equation", + "content": "Z_{b1} = E_{IS}(x_{b})" + }, + { + "bbox": [ + 131, + 396, + 480, + 503 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 131, + 396, + 480, + 503 + ], + "type": "inline_equation", + "content": "Z_{b2} = E_{ES}(x_{b})" + }, + { + "bbox": [ + 131, + 396, + 480, + 503 + ], + "type": "text", + "content": ", respectively. The feature output from the teacher encoder, " + }, + { + "bbox": [ + 131, + 396, + 480, + 503 + ], + "type": "inline_equation", + "content": "Z_{a}" + }, + { + "bbox": [ + 131, + 396, + 480, + 503 + ], + "type": "text", + "content": ", is then processed by the teacher head " + }, + { + "bbox": [ + 131, + 396, + 480, + 503 + ], + "type": "inline_equation", + "content": "H_{T}" + }, + { + "bbox": [ + 131, + 396, + 480, + 503 + ], + "type": "text", + "content": ", followed by centering with the Sinkhorn-Knopp (SK) [11] algorithm and normalization using a temperature-scaled softmax to generate the probability " + }, + { + "bbox": [ + 131, + 396, + 480, + 503 + ], + "type": "inline_equation", + "content": "p_{a}" + }, + { + "bbox": [ + 131, + 396, + 480, + 503 + ], + "type": "text", + "content": " as follows:" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 139, + 522, + 480, + 552 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 139, + 522, + 480, + 552 + ], + "spans": [ + { + "bbox": [ + 139, + 522, + 480, + 552 + ], + "type": "interline_equation", + "content": "l _ {a} = S K \\left(H _ {T} \\left(Z _ {a}\\right)\\right), l _ {a} \\in \\mathbb {R} ^ {P} \\quad p _ {a} ^ {i} = \\frac {\\exp \\left(l _ {a} ^ {i} / \\tau\\right)}{\\sum_ {k = 0} ^ {P - 1} \\exp \\left(l _ {a} ^ {k} / \\tau\\right)}, \\forall i \\in \\{0, \\dots , P - 1 \\}, \\tag {5}", + "image_path": "b8bf71f25a9f6a6f340ac00de3aef78e284364d8bb27c1f26b3f94f1153a12b8.jpg" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 131, + 558, + 481, + 666 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 131, + 558, + 481, + 666 + ], + "spans": [ + { + "bbox": [ + 131, + 558, + 481, + 666 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 131, + 558, + 481, + 666 + ], + "type": "inline_equation", + "content": "P" + }, + { + "bbox": [ + 131, + 558, + 481, + 666 + ], + "type": "text", + "content": " is the number of prototypes and " + }, + { + "bbox": [ + 131, + 558, + 481, + 666 + ], + "type": "inline_equation", + "content": "\\tau > 0" + }, + { + "bbox": [ + 131, + 558, + 481, + 666 + ], + "type": "text", + "content": " is a temperature parameter. In a similar fashion, we compute the probabilities " + }, + { + "bbox": [ + 131, + 558, + 481, + 666 + ], + "type": "inline_equation", + "content": "p_{b1}^{i}" + }, + { + "bbox": [ + 131, + 558, + 481, + 666 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 131, + 558, + 481, + 666 + ], + "type": "inline_equation", + "content": "p_{b2}^{i}" + }, + { + "bbox": [ + 131, + 558, + 481, + 666 + ], + "type": "text", + "content": " for the intact and elastic student encoders, respectively, by processing the outputs through the student heads " + }, + { + "bbox": [ + 131, + 558, + 481, + 666 + ], + "type": "inline_equation", + "content": "H_{IS}" + }, + { + "bbox": [ + 131, + 558, + 481, + 666 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 131, + 558, + 481, + 666 + ], + "type": "inline_equation", + "content": "H_{ES}" + }, + { + "bbox": [ + 131, + 558, + 481, + 666 + ], + "type": "text", + "content": ". These outputs are then passed through a temperature-scaled softmax function using a temperature parameter " + }, + { + "bbox": [ + 131, + 558, + 481, + 666 + ], + "type": "inline_equation", + "content": "\\tau'" + }, + { + "bbox": [ + 131, + 558, + 481, + 666 + ], + "type": "text", + "content": " tailored for the student. It should be noted that " + }, + { + "bbox": [ + 131, + 558, + 481, + 666 + ], + "type": "inline_equation", + "content": "H_{IS}" + }, + { + "bbox": [ + 131, + 558, + 481, + 666 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 131, + 558, + 481, + 666 + ], + "type": "inline_equation", + "content": "H_{ES}" + }, + { + "bbox": [ + 131, + 558, + 481, + 666 + ], + "type": "text", + "content": " share the identical parameters, except that the first projection layer of " + }, + { + "bbox": [ + 131, + 558, + 481, + 666 + ], + "type": "inline_equation", + "content": "H_{ES}" + }, + { + "bbox": [ + 131, + 558, + 481, + 666 + ], + "type": "text", + "content": " is adapted similarly as in Equation 2 to conform to align with the corresponding dimensionality. For simplicity, we omit the explicit expressions for " + }, + { + "bbox": [ + 131, + 558, + 481, + 666 + ], + "type": "inline_equation", + "content": "p_{b1}^{i}" + }, + { + "bbox": [ + 131, + 558, + 481, + 666 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 131, + 558, + 481, + 666 + ], + "type": "inline_equation", + "content": "p_{b2}^{i}" + }, + { + "bbox": [ + 131, + 558, + 481, + 666 + ], + "type": "text", + "content": ", as they follow a similar" + } + ] + } + ], + "index": 12 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 425, + 91, + 447, + 100 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 425, + 91, + 447, + 100 + ], + "spans": [ + { + "bbox": [ + 425, + 91, + 447, + 100 + ], + "type": "text", + "content": "POA" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 474, + 91, + 481, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 474, + 91, + 481, + 100 + ], + "spans": [ + { + "bbox": [ + 474, + 91, + 481, + 100 + ], + "type": "text", + "content": "7" + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 6 + }, + { + "para_blocks": [ + { + "bbox": [ + 130, + 116, + 479, + 140 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 116, + 479, + 140 + ], + "spans": [ + { + "bbox": [ + 130, + 116, + 479, + 140 + ], + "type": "text", + "content": "calculation to Equation 5. For the intact student branch, we perform distillation from the teacher using the cross-view data as follows:" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 264, + 152, + 481, + 166 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 264, + 152, + 481, + 166 + ], + "spans": [ + { + "bbox": [ + 264, + 152, + 481, + 166 + ], + "type": "interline_equation", + "content": "\\mathcal {L} _ {I S} ^ {g} = - p _ {a} \\log \\left(p _ {b 1}\\right). \\tag {6}", + "image_path": "7af346efc887897d67a31c4ce82621dd2f41ba5bca64bc4ad44a4e352ed2abde.jpg" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 130, + 169, + 480, + 264 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 169, + 480, + 264 + ], + "spans": [ + { + "bbox": [ + 130, + 169, + 480, + 264 + ], + "type": "text", + "content": "The elastic student branch plays a pivotal role in our POA framework. To ensure adequate training of this branch, we employ a dual distillation from the teacher and intact student branch. The first distillation involves the teacher model, which utilizes the cross-view data to guide the learning of representations. The second is a distillation process with the intact student model, which uses the same-view data. This same-view distillation is responsible for transferring the representations learned by the intact student to the elastic student branch. The loss functions for this dual distillation process are formulated as follows:" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 208, + 276, + 481, + 291 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 208, + 276, + 481, + 291 + ], + "spans": [ + { + "bbox": [ + 208, + 276, + 481, + 291 + ], + "type": "interline_equation", + "content": "\\mathcal {L} _ {E S 1} ^ {g} = - p _ {a} \\log (p _ {b 2}), \\quad \\mathcal {L} _ {E S 2} ^ {g} = - p _ {b 1} \\log (p _ {b 2}). \\qquad (7)", + "image_path": "2c13a4d35dfea18169965f196f6f37ad0a1c9b03b1a4b4d2a5d221640d623fc3.jpg" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 130, + 293, + 479, + 318 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 293, + 479, + 318 + ], + "spans": [ + { + "bbox": [ + 130, + 293, + 479, + 318 + ], + "type": "text", + "content": "Note that in both loss functions, we sum over all prototypes to compute the cross-entropy loss between the respective probability distributions." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 132, + 334, + 264, + 345 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 334, + 264, + 345 + ], + "spans": [ + { + "bbox": [ + 132, + 334, + 264, + 345 + ], + "type": "text", + "content": "3.3 Overall Loss of POA" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 130, + 352, + 479, + 423 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 352, + 479, + 423 + ], + "spans": [ + { + "bbox": [ + 130, + 352, + 479, + 423 + ], + "type": "text", + "content": "Following the SSL methods such as [5,33,55], we employ a multi-crop strategy [4] to create various distorted views from a single image. Apart from the two global views previously mentioned, we also generate " + }, + { + "bbox": [ + 130, + 352, + 479, + 423 + ], + "type": "inline_equation", + "content": "v" + }, + { + "bbox": [ + 130, + 352, + 479, + 423 + ], + "type": "text", + "content": " local views with lower resolution " + }, + { + "bbox": [ + 130, + 352, + 479, + 423 + ], + "type": "inline_equation", + "content": "x_{l_1}, x_{l_2}, \\ldots, x_{l_v}" + }, + { + "bbox": [ + 130, + 352, + 479, + 423 + ], + "type": "text", + "content": ". These local views are processed by both students to promote local-to-global correspondence. The local distillation losses for the intact and elastic students are computed as follows:" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 250, + 434, + 481, + 464 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 250, + 434, + 481, + 464 + ], + "spans": [ + { + "bbox": [ + 250, + 434, + 481, + 464 + ], + "type": "interline_equation", + "content": "\\mathcal {L} _ {I S} ^ {l} = - \\frac {1}{v} \\sum_ {i = 1} ^ {v} p _ {a} \\log \\left(p _ {l _ {i} 1}\\right), \\tag {8}", + "image_path": "cd67e6c9b1b452e33b1a443630b56fea5507937e98948e64c1fcd5eee91f49ea.jpg" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 179, + 470, + 481, + 500 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 179, + 470, + 481, + 500 + ], + "spans": [ + { + "bbox": [ + 179, + 470, + 481, + 500 + ], + "type": "interline_equation", + "content": "\\mathcal {L} _ {E S 1} ^ {l} = - \\frac {1}{v} \\sum_ {i = 1} ^ {v} p _ {a} \\log \\left(p _ {l _ {i} 2}\\right), \\quad \\mathcal {L} _ {E S 2} ^ {l} = - \\frac {1}{v} \\sum_ {i = 1} ^ {v} p _ {l _ {i} 1} \\log \\left(p _ {l _ {i} 2}\\right), \\tag {9}", + "image_path": "9f0bed9876892b74a3c790b2222eb1a6ecf002154d9830114e602cd4fc2ceecc.jpg" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 130, + 503, + 479, + 539 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 503, + 479, + 539 + ], + "spans": [ + { + "bbox": [ + 130, + 503, + 479, + 539 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 130, + 503, + 479, + 539 + ], + "type": "inline_equation", + "content": "p_{l_{i1}}" + }, + { + "bbox": [ + 130, + 503, + 479, + 539 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 130, + 503, + 479, + 539 + ], + "type": "inline_equation", + "content": "p_{l_{i2}}" + }, + { + "bbox": [ + 130, + 503, + 479, + 539 + ], + "type": "text", + "content": " are the probability produced by the intact and elastic student branches for the local view " + }, + { + "bbox": [ + 130, + 503, + 479, + 539 + ], + "type": "inline_equation", + "content": "l_i" + }, + { + "bbox": [ + 130, + 503, + 479, + 539 + ], + "type": "text", + "content": ", respectively. The total distillation loss of the intact and elastic student is calculated by summing them with the factor " + }, + { + "bbox": [ + 130, + 503, + 479, + 539 + ], + "type": "inline_equation", + "content": "\\lambda" + }, + { + "bbox": [ + 130, + 503, + 479, + 539 + ], + "type": "text", + "content": ":" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 159, + 546, + 481, + 574 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 159, + 546, + 481, + 574 + ], + "spans": [ + { + "bbox": [ + 159, + 546, + 481, + 574 + ], + "type": "interline_equation", + "content": "\\begin{array}{l} \\mathcal {L} _ {\\mathcal {S}} = \\lambda \\left(\\mathcal {L} _ {I S} ^ {g} + \\mathcal {L} _ {I S} ^ {l}\\right) + (1 - \\lambda) \\left(\\left(\\mathcal {L} _ {E S 1} ^ {g} + \\mathcal {L} _ {E S 1} ^ {l}\\right) + \\left(\\mathcal {L} _ {E S 2} ^ {g} + \\mathcal {L} _ {E S 2} ^ {l}\\right)\\right) \\tag {10} \\\\ = \\lambda \\mathcal {L} _ {I S} + (1 - \\lambda) \\left(\\mathcal {L} _ {E S 1} + \\mathcal {L} _ {E S 2}\\right). \\\\ \\end{array}", + "image_path": "86cc6f239f8925eb3248e36e91ed50c8974cdc01f2973f7c4b4335d7fcd78c31.jpg" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 130, + 582, + 480, + 667 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 582, + 480, + 667 + ], + "spans": [ + { + "bbox": [ + 130, + 582, + 480, + 667 + ], + "type": "text", + "content": "To ensure sufficient training of each sub-network from the elastic student, we introduce multiple projection heads (MPH) positioned after the backbone network. Each projection head has exactly the same structure, except for a different number of prototypes. For each head, the distillation loss " + }, + { + "bbox": [ + 130, + 582, + 480, + 667 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_{\\mathcal{S}_i}" + }, + { + "bbox": [ + 130, + 582, + 480, + 667 + ], + "type": "text", + "content": " for both the intact and elastic student is calculated with Equation. 10. Finally, the overall loss function in the POA framework with " + }, + { + "bbox": [ + 130, + 582, + 480, + 667 + ], + "type": "inline_equation", + "content": "H" + }, + { + "bbox": [ + 130, + 582, + 480, + 667 + ], + "type": "text", + "content": " projection heads is formulated as: " + }, + { + "bbox": [ + 130, + 582, + 480, + 667 + ], + "type": "inline_equation", + "content": "\\mathcal{L} = \\frac{1}{H}\\sum_{i = 1}^{H}\\mathcal{L}_{\\mathcal{S}_i}" + }, + { + "bbox": [ + 130, + 582, + 480, + 667 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 13 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "spans": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "type": "text", + "content": "8" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 166, + 91, + 229, + 102 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 166, + 91, + 229, + 102 + ], + "spans": [ + { + "bbox": [ + 166, + 91, + 229, + 102 + ], + "type": "text", + "content": "Y. Zhang et al." + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 7 + }, + { + "para_blocks": [ + { + "bbox": [ + 132, + 115, + 229, + 129 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 115, + 229, + 129 + ], + "spans": [ + { + "bbox": [ + 132, + 115, + 229, + 129 + ], + "type": "text", + "content": "4 Experiments" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 132, + 140, + 279, + 152 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 140, + 279, + 152 + ], + "spans": [ + { + "bbox": [ + 132, + 140, + 279, + 152 + ], + "type": "text", + "content": "4.1 Implementation Details" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 134, + 159, + 482, + 304 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 134, + 159, + 482, + 304 + ], + "spans": [ + { + "bbox": [ + 134, + 159, + 482, + 304 + ], + "type": "text", + "content": "Backbones. We have trained our POA using ViT, Swin Transformer and ResNet backbones, respectively. For the ViT, we configure the patch size to 16 and the dimension of each head in the MSA to 64. This aligns with the configurations typically used in existing SSL methods [5,34,35,55]. For the smallest and largest elastic networks of ViT, we choose the ViT-S and ViT-L, respectively. This leads to 11 elastic widths and 13 elastic depths, yielding a total of 143 ViT subnetworks. In the case of the Swin Transformer, we set the smallest and largest elastic networks as Swin-T and Swin-B, respectively. This configuration yields a total of 39 Swin sub-networks by multiplying the number of widths and depths as " + }, + { + "bbox": [ + 134, + 159, + 482, + 304 + ], + "type": "inline_equation", + "content": "3 \\times 13" + }, + { + "bbox": [ + 134, + 159, + 482, + 304 + ], + "type": "text", + "content": ". For the ResNet, we establish the smallest and largest elastic network configurations as ResNet-50 and ResNet-152. Consequently, this setup accounts for a total of 465 ResNet sub-networks, which is the product of " + }, + { + "bbox": [ + 134, + 159, + 482, + 304 + ], + "type": "inline_equation", + "content": "3 \\times 155" + }, + { + "bbox": [ + 134, + 159, + 482, + 304 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 134, + 316, + 482, + 448 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 134, + 316, + 482, + 448 + ], + "spans": [ + { + "bbox": [ + 134, + 316, + 482, + 448 + ], + "type": "text", + "content": "Pre-Training Setup. We pre-train all models on the ImageNet-1K dataset [12] without the labels. The process employs the AdamW optimizer [31] with a batch size of 1600, which is distributed across 32 A100 GPUs when employing a ViT backbone. We adopt a learning rate schedule that begins with a linear warm-up during the first 10 epochs, reaching a base value that is scaled proportionally to the total batch size: " + }, + { + "bbox": [ + 134, + 316, + 482, + 448 + ], + "type": "inline_equation", + "content": "\\mathrm{lr} = 0.004 \\times \\sqrt{\\text{batch size} / 1024}" + }, + { + "bbox": [ + 134, + 316, + 482, + 448 + ], + "type": "text", + "content": ", in line with [33]. Following this warm-up period, the learning rate is decayed with a cosine schedule. Similarly, the weight decay follows a cosine schedule, starting at 0.04 and increasing to 0.4. The student network's temperature " + }, + { + "bbox": [ + 134, + 316, + 482, + 448 + ], + "type": "inline_equation", + "content": "\\tau'" + }, + { + "bbox": [ + 134, + 316, + 482, + 448 + ], + "type": "text", + "content": " is fixed at 0.1, whereas the teacher temperature " + }, + { + "bbox": [ + 134, + 316, + 482, + 448 + ], + "type": "inline_equation", + "content": "\\tau" + }, + { + "bbox": [ + 134, + 316, + 482, + 448 + ], + "type": "text", + "content": " changes with a linear warm-up from 0.04 to 0.07 over the first 30 epochs. For further details, please refer to Appendix C." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 132, + 465, + 311, + 478 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 465, + 311, + 478 + ], + "spans": [ + { + "bbox": [ + 132, + 465, + 311, + 478 + ], + "type": "text", + "content": "4.2 Experiments on ImageNet-1K" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 134, + 486, + 482, + 581 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 134, + 486, + 482, + 581 + ], + "spans": [ + { + "bbox": [ + 134, + 486, + 482, + 581 + ], + "type": "text", + "content": "After unsupervised pre-training, we assess the model's performance using two widely-recognized evaluation protocols in SSL domain on ImageNet-1K dataset, i.e., k-NN and linear probing. To ensure a fair comparison between SSL methods that employ different numbers of crop views for data augmentation, Zhou et al. [55] introduced the effective training epoch as a measure to quantify the extent of a method's pre-training. We report the effective training epochs of the SSL methods for comparison. For additional evaluations including fine-tuning, semi-supervised and unsupervised learning, please refer to Appendix D." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 132, + 594, + 482, + 665 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 594, + 482, + 665 + ], + "spans": [ + { + "bbox": [ + 132, + 594, + 482, + 665 + ], + "type": "text", + "content": "k-NN and Linear Probing. To assess the quality of pre-trained features, we employ a k-NN classifier and a linear classifier on the frozen representations. For both the k-NN and linear probing (LP) evaluation, we follow the evaluation protocols established in [5, 33, 55]. The performance of our method when being trained using ResNet, Swin Transformer and ViT backbones is reported in Table 1. Our POA SSL achieves the SOTA k-NN accuracy of " + }, + { + "bbox": [ + 132, + 594, + 482, + 665 + ], + "type": "inline_equation", + "content": "82.3\\%" + }, + { + "bbox": [ + 132, + 594, + 482, + 665 + ], + "type": "text", + "content": " and LP accuracy" + } + ] + } + ], + "index": 8 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 425, + 91, + 447, + 100 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 425, + 91, + 447, + 100 + ], + "spans": [ + { + "bbox": [ + 425, + 91, + 447, + 100 + ], + "type": "text", + "content": "POA" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 474, + 92, + 481, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 474, + 92, + 481, + 100 + ], + "spans": [ + { + "bbox": [ + 474, + 92, + 481, + 100 + ], + "type": "text", + "content": "9" + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 8 + }, + { + "para_blocks": [ + { + "type": "table", + "bbox": [ + 132, + 178, + 476, + 367 + ], + "blocks": [ + { + "bbox": [ + 130, + 114, + 482, + 170 + ], + "lines": [ + { + "bbox": [ + 130, + 114, + 482, + 170 + ], + "spans": [ + { + "bbox": [ + 130, + 114, + 482, + 170 + ], + "type": "text", + "content": "Table 1: Comparison results of k-NN and linear probing classification accuracy (\\%) on the ImageNet-1K dataset. \"Param.\" refers to the number of parameters. \"Epo.\" represents the number of effective training epochs following [55]. \"/16\" denotes patch size of 16. \"/W7\" means the window size of 7. \"**\" indicates our implementation based on official codebase. \"+\" denotes reproduced results using the released code." + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 132, + 178, + 476, + 367 + ], + "lines": [ + { + "bbox": [ + 132, + 178, + 476, + 367 + ], + "spans": [ + { + "bbox": [ + 132, + 178, + 476, + 367 + ], + "type": "table", + "html": "
MethodEpo.k-NNLPMethodEpo.k-NNLPMethodEpo.k-NNLP
ResNet-50(Param. 23M)Swin-T/W7(Param. 28M)ViT-S/16(Param. 21M)
VICReg2000-73.2SMoG1200-77.7DINO320074.577.0
SwAV240065.775.3iBOT120075.378.6iBOT320075.277.9
DINO320067.575.3DINOv2*120075.478.0ENT320075.377.7
UniGrad2400-75.5EsViT120075.778.1POA076.877.6
SCFS320068.575.7POA077.578.9ViT-B/16(Param. 85M)
ReLICv24000-77.1Swin-S/W7(Param. 49M)DINO160076.178.2
POA073.476.9DINOv2*120076.179.8ENT160077.179.1
ResNet-101(Param. 41M)EsViT120077.779.5iBOT160077.179.5
ReLICv24000-78.7POA079.381.3POA080.981.7
POA075.779.1Swin-B/W7(Param. 87M)ViT-L/16(Param. 307M)
ResNet-152(Param. 56M)DINOv2*120076.980.9iBOT120078.081.0
ReLICv24000-79.3EsViT120078.980.4DINOv2†120082.083.3
POA240076.479.9POA120079.682.0POA120082.383.6
", + "image_path": "9340ef9b865f2b883823d586bede6d6f2ab06e59800dcc838d868f142bb49f50.jpg" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "table_body" + }, + { + "bbox": [ + 149, + 368, + 228, + 378 + ], + "lines": [ + { + "bbox": [ + 149, + 368, + 228, + 378 + ], + "spans": [ + { + "bbox": [ + 149, + 368, + 228, + 378 + ], + "type": "text", + "content": "(a) ResNet backbone." + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "table_footnote" + }, + { + "bbox": [ + 268, + 368, + 338, + 377 + ], + "lines": [ + { + "bbox": [ + 268, + 368, + 338, + 377 + ], + "spans": [ + { + "bbox": [ + 268, + 368, + 338, + 377 + ], + "type": "text", + "content": "(b) Swin backbone." + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "table_footnote" + }, + { + "bbox": [ + 383, + 368, + 450, + 377 + ], + "lines": [ + { + "bbox": [ + 383, + 368, + 450, + 377 + ], + "spans": [ + { + "bbox": [ + 383, + 368, + 450, + 377 + ], + "type": "text", + "content": "(c) ViT backbone." + } + ] + } + ], + "index": 6, + "angle": 0, + "type": "table_footnote" + } + ], + "index": 3 + }, + { + "bbox": [ + 130, + 391, + 482, + 559 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 391, + 482, + 559 + ], + "spans": [ + { + "bbox": [ + 130, + 391, + 482, + 559 + ], + "type": "text", + "content": "of " + }, + { + "bbox": [ + 130, + 391, + 482, + 559 + ], + "type": "inline_equation", + "content": "83.6\\%" + }, + { + "bbox": [ + 130, + 391, + 482, + 559 + ], + "type": "text", + "content": " when using the ViT-L/16 backbone. By employing the sub-network extraction approach outlined in Section 3.1, we derive the sub-networks ViT-S/16 and ViT-B/16 from the teacher ViT-L/16 without any additional pre-training. Thus, the number of effective training epochs of them is 0. Notably, the extracted ViT-S/16 achieves the SOTA k-NN accuracy of " + }, + { + "bbox": [ + 130, + 391, + 482, + 559 + ], + "type": "inline_equation", + "content": "76.8\\%" + }, + { + "bbox": [ + 130, + 391, + 482, + 559 + ], + "type": "text", + "content": " and LP accuracy on par with the previous SOTA of " + }, + { + "bbox": [ + 130, + 391, + 482, + 559 + ], + "type": "inline_equation", + "content": "77.9\\%" + }, + { + "bbox": [ + 130, + 391, + 482, + 559 + ], + "type": "text", + "content": " reported by iBOT [55]. Our derived ViT-B/16 model sets new benchmarks for both k-NN and LP, with accuracy of " + }, + { + "bbox": [ + 130, + 391, + 482, + 559 + ], + "type": "inline_equation", + "content": "80.9\\%" + }, + { + "bbox": [ + 130, + 391, + 482, + 559 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 130, + 391, + 482, + 559 + ], + "type": "inline_equation", + "content": "81.7\\%" + }, + { + "bbox": [ + 130, + 391, + 482, + 559 + ], + "type": "text", + "content": ", respectively. For the models using Swin Transformer and ResNet backbones, POA also reaches SOTA performance in k-NN and LP accuracy. The only exception is the LP accuracy of ResNet-50, which is competitive with that of ReLICv2 [39], despite the latter being trained over a much longer period (4000 epochs). The superior performance achieved across a range of backbone architectures confirms the effectiveness and versatility of our method. For additional detailed comparison with 27 existing methods, please see Appendix D.1." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 132, + 574, + 364, + 586 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 574, + 364, + 586 + ], + "spans": [ + { + "bbox": [ + 132, + 574, + 364, + 586 + ], + "type": "text", + "content": "4.3 Evaluation on Downstream Visual Tasks" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 130, + 593, + 482, + 666 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 593, + 482, + 666 + ], + "spans": [ + { + "bbox": [ + 130, + 593, + 482, + 666 + ], + "type": "text", + "content": "Object Detection and Instance Segmentation on COCO Dataset. For a fair comparison, we utilize the Cascade Mask R-CNN framework [3,20], which generates both bounding boxes and instance masks, in line with previous approach [55], on the COCO dataset [29]. We benchmark our results against existing SSL methods that generate pre-trained ViT-S/16 and ViT-B/16 backbones. As shown in Table 2, POA boosts the bounding box average precision " + }, + { + "bbox": [ + 130, + 593, + 482, + 666 + ], + "type": "inline_equation", + "content": "(\\mathrm{AP}^b)" + }, + { + "bbox": [ + 130, + 593, + 482, + 666 + ], + "type": "text", + "content": " for" + } + ] + } + ], + "index": 9 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "spans": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "type": "text", + "content": "10" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 166, + 91, + 229, + 102 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 166, + 91, + 229, + 102 + ], + "spans": [ + { + "bbox": [ + 166, + 91, + 229, + 102 + ], + "type": "text", + "content": "Y. Zhang et al." + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 9 + }, + { + "para_blocks": [ + { + "bbox": [ + 130, + 115, + 482, + 152 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 115, + 482, + 152 + ], + "spans": [ + { + "bbox": [ + 130, + 115, + 482, + 152 + ], + "type": "text", + "content": "ViT-S/16 from 49.4 to 50.6 and the mask average precision " + }, + { + "bbox": [ + 130, + 115, + 482, + 152 + ], + "type": "inline_equation", + "content": "(\\mathrm{AP}^m)" + }, + { + "bbox": [ + 130, + 115, + 482, + 152 + ], + "type": "text", + "content": " from 42.6 to 43.8. When applied to ViT-B/16, POA attains an " + }, + { + "bbox": [ + 130, + 115, + 482, + 152 + ], + "type": "inline_equation", + "content": "\\mathrm{AP}^b" + }, + { + "bbox": [ + 130, + 115, + 482, + 152 + ], + "type": "text", + "content": " of 52.4 and an " + }, + { + "bbox": [ + 130, + 115, + 482, + 152 + ], + "type": "inline_equation", + "content": "\\mathrm{AP}^m" + }, + { + "bbox": [ + 130, + 115, + 482, + 152 + ], + "type": "text", + "content": " of 45.4, which represents a remarkable step forward over previous SOTA." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 130, + 161, + 482, + 328 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 161, + 482, + 328 + ], + "spans": [ + { + "bbox": [ + 130, + 161, + 482, + 328 + ], + "type": "text", + "content": "Semantic Segmentation on ADE20K Dataset. For the semantic segmentation task, we primarily focus on two settings on the ADE20K dataset [54], following [55]. First, akin to the linear evaluation protocol in classification, we evaluate the quality of representations by keeping the patch features fixed and only fine-tuning one linear layer. This approach offers a clear comparison of representation quality. Second, we employ the UPerNet [46] as the task head and fine-tune the entire network. As depicted in Table 2, our POA significantly outperforms the supervised baseline using the ViT-S/16 backbone, achieving a substantial increase of 2.2 in mean Intersection over Union (mIoU), and surpassing iBOT by 1.3 mIoU. With the ViT-B/16 backbone, POA exceeds the previously best-performing method, iBOT, by 0.4 mIoU when utilizing UPerNet. Furthermore, under the assessment using solely a linear head, POA obtains an impressive improvement of 2.0 mIoU over iBOT as the performance is largely determined by the quality of the pre-trained representation." + } + ] + } + ], + "index": 3 + }, + { + "type": "table", + "bbox": [ + 148, + 362, + 466, + 453 + ], + "blocks": [ + { + "bbox": [ + 130, + 329, + 481, + 353 + ], + "lines": [ + { + "bbox": [ + 130, + 329, + 481, + 353 + ], + "spans": [ + { + "bbox": [ + 130, + 329, + 481, + 353 + ], + "type": "text", + "content": "Table 2: Evaluation results on downstream detection and segmentation tasks. " + }, + { + "bbox": [ + 130, + 329, + 481, + 353 + ], + "type": "inline_equation", + "content": "\\mathrm{Seg.}^{\\dagger}" + }, + { + "bbox": [ + 130, + 329, + 481, + 353 + ], + "type": "text", + "content": " indicates the use of a linear head for semantic segmentation." + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 148, + 362, + 466, + 453 + ], + "lines": [ + { + "bbox": [ + 148, + 362, + 466, + 453 + ], + "spans": [ + { + "bbox": [ + 148, + 362, + 466, + 453 + ], + "type": "table", + "html": "
MethodArch.Det. APbISeg. APmSeg. mIoUArch.Det. APbISeg. APmSeg.† mIoUSeg. mIoU
Sup.ViT-S/1646.240.144.5ViT-B/1649.843.235.446.6
BEiTViT-S/16---ViT-B/1650.143.527.445.8
DINOViT-S/16---ViT-B/1650.143.434.546.8
iBOTViT-S/1649.442.645.4ViT-B/1651.244.238.350.0
POAViT-S/1650.643.846.7ViT-B/1652.445.440.350.4
", + "image_path": "f34a8722e9d380d7b21c8742b47700e8a5528ff6ffd3f9c2f7824a92c636f093.jpg" + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "table_body" + } + ], + "index": 5 + }, + { + "bbox": [ + 131, + 471, + 293, + 482 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 131, + 471, + 293, + 482 + ], + "spans": [ + { + "bbox": [ + 131, + 471, + 293, + 482 + ], + "type": "text", + "content": "4.4 Ablations and Discussions" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 130, + 487, + 482, + 596 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 487, + 482, + 596 + ], + "spans": [ + { + "bbox": [ + 130, + 487, + 482, + 596 + ], + "type": "text", + "content": "In this section, we conduct an empirical analysis of POA using ViT as backbone. Our investigation includes the impact of the loss functions " + }, + { + "bbox": [ + 130, + 487, + 482, + 596 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_{ES1}" + }, + { + "bbox": [ + 130, + 487, + 482, + 596 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 130, + 487, + 482, + 596 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_{ES2}" + }, + { + "bbox": [ + 130, + 487, + 482, + 596 + ], + "type": "text", + "content": ", in addition with the effectiveness of multiple projection heads. Moreover, we compare our POA with knowledge distillation techniques for self-supervised learning to demonstrate the advantages of combining once-for-all model generation with self-distillation in a unified framework. For more ablation studies, we direct readers to the Appendix D.8. In addition, we further contrast our POA with three variants tailored for elastic pre-training to showcase POA's superiority. Finally, we discuss how the elastic student facilitates the pre-training." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 130, + 605, + 482, + 665 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 605, + 482, + 665 + ], + "spans": [ + { + "bbox": [ + 130, + 605, + 482, + 665 + ], + "type": "text", + "content": "Importance of Each Component. We evaluate the contributions of the components on a ViT backbone. Table 3 shows the performance of different component combinations. First, we note that employing multiple projection heads (MPH) enhances the learned representations for each elastic sub-network, particularly for smaller ones. The design consideration behind MPH is that for each" + } + ] + } + ], + "index": 8 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 425, + 91, + 447, + 100 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 425, + 91, + 447, + 100 + ], + "spans": [ + { + "bbox": [ + 425, + 91, + 447, + 100 + ], + "type": "text", + "content": "POA" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 470, + 91, + 480, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 470, + 91, + 480, + 100 + ], + "spans": [ + { + "bbox": [ + 470, + 91, + 480, + 100 + ], + "type": "text", + "content": "11" + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 10 + }, + { + "para_blocks": [ + { + "type": "table", + "bbox": [ + 138, + 157, + 476, + 236 + ], + "blocks": [ + { + "bbox": [ + 130, + 114, + 482, + 149 + ], + "lines": [ + { + "bbox": [ + 130, + 114, + 482, + 149 + ], + "spans": [ + { + "bbox": [ + 130, + 114, + 482, + 149 + ], + "type": "text", + "content": "Table 3: Contributions of each component in POA framework. We conduct assessments with the k-NN and linear probing (LP) evaluations. MPH denotes the multiple projection heads with different numbers of prototypes." + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 138, + 157, + 476, + 236 + ], + "lines": [ + { + "bbox": [ + 138, + 157, + 476, + 236 + ], + "spans": [ + { + "bbox": [ + 138, + 157, + 476, + 236 + ], + "type": "table", + "html": "
MPHLES1LES2k-NNLP
ViT-S/16ViT-B/16ViT-L/16ViT-S/16ViT-B/16ViT-L/16
✓✓✓76.880.982.377.681.783.6
✓✓76.280.782.277.381.683.4
✓75.180.282.275.881.083.4
✓72.879.182.175.380.883.3
", + "image_path": "73bb1f1a90420509d108ec5fc9f688229ef3a5b604e65f9ca9a223841c39366c.jpg" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "table_body" + } + ], + "index": 3 + }, + { + "bbox": [ + 130, + 240, + 482, + 419 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 240, + 482, + 419 + ], + "spans": [ + { + "bbox": [ + 130, + 240, + 482, + 419 + ], + "type": "text", + "content": "pre-training iteration, the sub-network is chosen randomly, leading to a relatively insufficient optimization. MPH introduces different sets of prototypes, which act as multiple semantic spaces for representation learning, enabling the teacher to distill various aspects of learned knowledge into the sub-network. Furthermore, we ascertain that the same-view distillation loss " + }, + { + "bbox": [ + 130, + 240, + 482, + 419 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_{ES2}" + }, + { + "bbox": [ + 130, + 240, + 482, + 419 + ], + "type": "text", + "content": " is crucial for the representation quality of elastic sub-networks. Omitting " + }, + { + "bbox": [ + 130, + 240, + 482, + 419 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_{ES2}" + }, + { + "bbox": [ + 130, + 240, + 482, + 419 + ], + "type": "text", + "content": " causes a significant drop in k-NN accuracy, by " + }, + { + "bbox": [ + 130, + 240, + 482, + 419 + ], + "type": "inline_equation", + "content": "3.4\\%" + }, + { + "bbox": [ + 130, + 240, + 482, + 419 + ], + "type": "text", + "content": " for ViT-S/16 and " + }, + { + "bbox": [ + 130, + 240, + 482, + 419 + ], + "type": "inline_equation", + "content": "1.6\\%" + }, + { + "bbox": [ + 130, + 240, + 482, + 419 + ], + "type": "text", + "content": " for ViT-B/16. In addition, " + }, + { + "bbox": [ + 130, + 240, + 482, + 419 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_{ES2}" + }, + { + "bbox": [ + 130, + 240, + 482, + 419 + ], + "type": "text", + "content": " is more important than cross-view distillation " + }, + { + "bbox": [ + 130, + 240, + 482, + 419 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_{ES1}" + }, + { + "bbox": [ + 130, + 240, + 482, + 419 + ], + "type": "text", + "content": " in terms of sub-networks' representation quality. The underlying reason is that the cross-view distillation is for the unsupervised representation learning, while the same-view distillation improves the sub-networks by distilling previously learned representations from the intact student. Table 4 confirms that distillation from already-good representations is more effective than representation learning, especially for smaller networks. It also explains why employing three different views in our POA is unnecessary." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 130, + 421, + 483, + 499 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 421, + 483, + 499 + ], + "spans": [ + { + "bbox": [ + 130, + 421, + 483, + 499 + ], + "type": "text", + "content": "Table 4: Comparison with knowledge distillation for self-supervised learning. The teacher name \"ViT-L/16-600\" denotes a teacher model (ViT-L/16) that has been pretrained with 600 effective epochs. The student name \"ViT-S/16-600\" refers to a student model (ViT-S/16) that has undergone distillation from the pre-trained teacher with 600 effective epochs. The number of the left side of \"" + }, + { + "bbox": [ + 130, + 421, + 483, + 499 + ], + "type": "inline_equation", + "content": "\\rightarrow" + }, + { + "bbox": [ + 130, + 421, + 483, + 499 + ], + "type": "text", + "content": "\" indicates the performance of the teacher model when pre-trained individually. The number of the right side denotes the performance achieved after distillation using the SEED [15]." + } + ] + } + ], + "index": 5 + }, + { + "type": "table", + "bbox": [ + 132, + 508, + 483, + 626 + ], + "blocks": [ + { + "bbox": [ + 132, + 508, + 483, + 626 + ], + "lines": [ + { + "bbox": [ + 132, + 508, + 483, + 626 + ], + "spans": [ + { + "bbox": [ + 132, + 508, + 483, + 626 + ], + "type": "table", + "html": "
MethodTeacherStudentTotal Epochsk-NNLP
DINOv2ViT-S/16120072.273.1
DINOv2+SEEDViT-L/16-600ViT-S/16-600120081.3 → 74.082.4 → 75.2
DINOv2+SEEDViT-L/16-1200ViT-S/16-1200240082.0 → 75.583.3 → 76.2
POAViT-S/16120076.877.6
DINOv2ViT-B/16120077.478.5
DINOv2+SEEDViT-L/16-600ViT-B/16-600120081.3 → 78.882.4 → 80.0
DINOv2+SEEDViT-L/16-1200ViT-B/16-1200240082.0 → 79.783.3 → 80.9
POAViT-B/16120080.981.7
", + "image_path": "82af6ed40a3c842b8077784a3655cdd874ce1e88cb85454ba3f2a516af52d216.jpg" + } + ] + } + ], + "index": 6, + "angle": 0, + "type": "table_body" + } + ], + "index": 6 + }, + { + "bbox": [ + 132, + 640, + 482, + 666 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 640, + 482, + 666 + ], + "spans": [ + { + "bbox": [ + 132, + 640, + 482, + 666 + ], + "type": "text", + "content": "Comparison with Knowledge Distillation. Knowledge distillation (KD) is a proven strategy for enhancing small networks by leveraging the knowledge of" + } + ] + } + ], + "index": 7 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "spans": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "type": "text", + "content": "12" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 166, + 91, + 229, + 102 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 166, + 91, + 229, + 102 + ], + "spans": [ + { + "bbox": [ + 166, + 91, + 229, + 102 + ], + "type": "text", + "content": "Y. Zhang et al." + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 11 + }, + { + "para_blocks": [ + { + "bbox": [ + 130, + 116, + 482, + 296 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 116, + 482, + 296 + ], + "spans": [ + { + "bbox": [ + 130, + 116, + 482, + 296 + ], + "type": "text", + "content": "a well-trained, larger network. We compare our method with a self-supervised KD method SEED [15] that employs a pre-trained network from the previous SOTA DINOv2 as the teacher for distilling knowledge into smaller networks. We observe that SEED significantly boosts the performance of ViT-S/16 and ViT-B/16 when a pre-trained ViT-L/16 serves as the teacher. Specifically, with the same number of training epochs, SEED yields k-NN accuracy improvements of " + }, + { + "bbox": [ + 130, + 116, + 482, + 296 + ], + "type": "inline_equation", + "content": "1.8\\%" + }, + { + "bbox": [ + 130, + 116, + 482, + 296 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 130, + 116, + 482, + 296 + ], + "type": "inline_equation", + "content": "1.4\\%" + }, + { + "bbox": [ + 130, + 116, + 482, + 296 + ], + "type": "text", + "content": " for ViT-S/16 and ViT-B/16 respectively, compared with a learnfrom-scratch setup. Our POA further outperforms SEED, delivering superior k-NN accuracy gains of " + }, + { + "bbox": [ + 130, + 116, + 482, + 296 + ], + "type": "inline_equation", + "content": "2.8\\%" + }, + { + "bbox": [ + 130, + 116, + 482, + 296 + ], + "type": "text", + "content": " (76.8% vs. 74.0%) and " + }, + { + "bbox": [ + 130, + 116, + 482, + 296 + ], + "type": "inline_equation", + "content": "2.1\\%" + }, + { + "bbox": [ + 130, + 116, + 482, + 296 + ], + "type": "text", + "content": " (80.9% vs. 78.8%) on ViT-S/16 and ViT-B/16. Moreover, the ViT-S/16 and ViT-B/16 derived directly from the pre-trained teacher from our POA, without additional pre-training, perform better than those enhanced by SEED, despite the latter undergoing two times of training epochs. These results demonstrate the effectiveness of our unified design, which integrates pre-training with once-for-all model generation. Additional comparison on training cost can be found in Appendix D.11." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 130, + 310, + 482, + 453 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 310, + 482, + 453 + ], + "spans": [ + { + "bbox": [ + 130, + 310, + 482, + 453 + ], + "type": "text", + "content": "Alternative Designs to Elastic Pre-training. Upon the proposed POA, we investigate three alternatives for concurrently pre-training multi-sized models within a single pre-training session. As illustrated in Figure 4, the first variant, POA-V1, eliminates the intact student and modifies the teacher to be elastic which aligns with the architecture of the elastic student. The second variant, POA-V2, only removes the intact student. However, in this model, the intact teacher is directly updated via an EMA of the elastic students. These two variants bear resemblance to current teacher-student self-distillation paradigm, with the primary differences in the structure of the teacher or student network. For the POA-V3, we introduce an additional elastic teacher branch that mirrors the architecture of the elastic student, enabling further cross-view distillation. For more details of the three variants, please refer to the Appendix D.9." + } + ] + } + ], + "index": 3 + }, + { + "type": "image", + "bbox": [ + 136, + 455, + 483, + 545 + ], + "blocks": [ + { + "bbox": [ + 136, + 455, + 483, + 545 + ], + "lines": [ + { + "bbox": [ + 136, + 455, + 483, + 545 + ], + "spans": [ + { + "bbox": [ + 136, + 455, + 483, + 545 + ], + "type": "image", + "image_path": "63f78a649a4a00db6dcb1fd57c087ad50e27c7a5cd5e7f1d95f95aebf15b26ae.jpg" + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 207, + 553, + 405, + 564 + ], + "lines": [ + { + "bbox": [ + 207, + 553, + 405, + 564 + ], + "spans": [ + { + "bbox": [ + 207, + 553, + 405, + 564 + ], + "type": "text", + "content": "Fig. 4: Illustration of different variants of POA." + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "image_caption" + } + ], + "index": 4 + }, + { + "bbox": [ + 130, + 570, + 482, + 666 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 570, + 482, + 666 + ], + "spans": [ + { + "bbox": [ + 130, + 570, + 482, + 666 + ], + "type": "text", + "content": "We compare the k-NN and linear probing accuracy of each variant in Table 5. The results indicate that the performance drops dramatically for the variants with two branches. This decline can be attributed to two reasons: 1) in each training iteration, only a subset of the student's parameters are updated, leading to relatively insufficient overall pre-training; and 2) each elastic network undergoes training exclusively with the cross-view distillation, missing out the standard distillation guidance from a larger network, which our ablation study suggests is crucial for the efficacy of elastic networks. Although the POA-V3," + } + ] + } + ], + "index": 6 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 425, + 91, + 447, + 100 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 425, + 91, + 447, + 100 + ], + "spans": [ + { + "bbox": [ + 425, + 91, + 447, + 100 + ], + "type": "text", + "content": "POA" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 470, + 91, + 481, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 470, + 91, + 481, + 100 + ], + "spans": [ + { + "bbox": [ + 470, + 91, + 481, + 100 + ], + "type": "text", + "content": "13" + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 12 + }, + { + "para_blocks": [ + { + "bbox": [ + 130, + 116, + 479, + 152 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 116, + 479, + 152 + ], + "spans": [ + { + "bbox": [ + 130, + 116, + 479, + 152 + ], + "type": "text", + "content": "featuring an additional elastic teacher branch, performs slightly better than our POA, it does so at the expense of increased model complexity and computational costs, making it a less appealing scheme than POA." + } + ] + } + ], + "index": 2 + }, + { + "type": "table", + "bbox": [ + 137, + 175, + 477, + 255 + ], + "blocks": [ + { + "bbox": [ + 140, + 153, + 472, + 165 + ], + "lines": [ + { + "bbox": [ + 140, + 153, + 472, + 165 + ], + "spans": [ + { + "bbox": [ + 140, + 153, + 472, + 165 + ], + "type": "text", + "content": "Table 5: k-NN and linear probing (LP) comparison of different variants of POA." + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 137, + 175, + 477, + 255 + ], + "lines": [ + { + "bbox": [ + 137, + 175, + 477, + 255 + ], + "spans": [ + { + "bbox": [ + 137, + 175, + 477, + 255 + ], + "type": "table", + "html": "
Methodk-NNLP
ViT-S/16ViT-B/16ViT-L/16ViT-S/16ViT-B/16ViT-L/16
POA76.880.982.377.681.783.6
POA-V173.778.579.274.579.480.9
POA-V274.379.180.975.180.082.3
POA-V376.981.082.377.881.883.6
", + "image_path": "5283dc82d277162ae9426518cd5c6f50f63888feeba0b7f79d4fd593724d6140.jpg" + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "table_body" + } + ], + "index": 4 + }, + { + "bbox": [ + 130, + 270, + 482, + 475 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 270, + 482, + 475 + ], + "spans": [ + { + "bbox": [ + 130, + 270, + 482, + 475 + ], + "type": "text", + "content": "How Does Elastic Student Facilitate Pre-training? The elastic student branch enables the derivation of diverse-sized models directly from a pre-trained teacher, while simultaneously enhancing the learned representation. The elastic branch plays a dual role. First, it acts as a training regularization to stabilize the training progress. In our preliminary experiments, we observe that pre-training loss of a ResNet backbone easily diverges to NaN without the elastic branch. Conversely, the inclusion of the elastic student yields a highly stable pre-training. Furthermore, the elastic student provides hundreds of sub-network candidates that will be assembled to the teacher during pre-training. Unlike existing self-distillation methods, the teacher in the POA SSL integrates a series of sub-networks through an EMA update. Wortsman et al. [44] have shown that averaging the weights of multiple models typically improves accuracy and robustness. Similarly, the integration of different sub-networks for the teacher is effective in improving the representation quality. At the same time, an improved teacher promotes the learning of students in return, fostering a positive feedback loop for pre-training. For the visualization of representations, please refer to Appendix E.2." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 132, + 495, + 220, + 508 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 495, + 220, + 508 + ], + "spans": [ + { + "bbox": [ + 132, + 495, + 220, + 508 + ], + "type": "text", + "content": "5 Conclusion" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 130, + 521, + 482, + 666 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 521, + 482, + 666 + ], + "spans": [ + { + "bbox": [ + 130, + 521, + 482, + 666 + ], + "type": "text", + "content": "In this study, we tackle the challenge of efficiently and effectively pre-training models of various sizes within a single self-supervised learning session, facilitating model deployment given different resource constraints. We propose a novel self-supervised learning paradigm, termed POA, which features an integrated design combining self-distillation and once-for-all model generation. It allows for the simultaneous pre-training of models of multiple sizes through an elastic branch design. POA enables the direct generation of varied-sized models from a pretrained teacher, which are ready for downstream tasks without additional pretraining. This advantage significantly improves the deployment flexibility and facilitates our pre-trained model to achieve SOTA results across various vision tasks. Looking forward, we plan to extend POA to Multimodal Large Language Models, tapping into its vast potential for real-world AI product deployment." + } + ] + } + ], + "index": 7 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "spans": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "type": "text", + "content": "14" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 166, + 91, + 228, + 102 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 166, + 91, + 228, + 102 + ], + "spans": [ + { + "bbox": [ + 166, + 91, + 228, + 102 + ], + "type": "text", + "content": "Y. Zhang et al." + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 13 + }, + { + "para_blocks": [ + { + "bbox": [ + 132, + 114, + 198, + 127 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 114, + 198, + 127 + ], + "spans": [ + { + "bbox": [ + 132, + 114, + 198, + 127 + ], + "type": "text", + "content": "References" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 138, + 140, + 480, + 665 + ], + "type": "list", + "angle": 0, + "index": 20, + "blocks": [ + { + "bbox": [ + 138, + 140, + 480, + 161 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 140, + 480, + 161 + ], + "spans": [ + { + "bbox": [ + 138, + 140, + 480, + 161 + ], + "type": "text", + "content": "1. Ba, J., Kiros, J.R., Hinton, G.E.: Layer normalization. ArXiv abs/1607.06450 (2016)" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 138, + 162, + 480, + 194 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 162, + 480, + 194 + ], + "spans": [ + { + "bbox": [ + 138, + 162, + 480, + 194 + ], + "type": "text", + "content": "2. Cai, H., Gan, C., Wang, T., Zhang, Z., Han, S.: Once for all: Train one network and specialize it for efficient deployment. In: International Conference on Learning Representations (2020)" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 138, + 195, + 480, + 226 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 195, + 480, + 226 + ], + "spans": [ + { + "bbox": [ + 138, + 195, + 480, + 226 + ], + "type": "text", + "content": "3. Cai, Z., Vasconcelos, N.: Cascade r-cnn: High quality object detection and instance segmentation. IEEE Transactions on Pattern Analysis and Machine Intelligence 43, 1483-1498 (2019)" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 138, + 228, + 480, + 270 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 228, + 480, + 270 + ], + "spans": [ + { + "bbox": [ + 138, + 228, + 480, + 270 + ], + "type": "text", + "content": "4. Caron, M., Misra, I., Mairal, J., Goyal, P., Bojanowski, P., Joulin, A.: Unsupervised learning of visual features by contrasting cluster assignments. In: Larochelle, H., Ranzato, M., Hadsell, R., Balcan, M., Lin, H. (eds.) Advances in Neural Information Processing Systems. vol. 33, pp. 9912-9924. Curran Associates, Inc. (2020)" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 138, + 271, + 480, + 304 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 271, + 480, + 304 + ], + "spans": [ + { + "bbox": [ + 138, + 271, + 480, + 304 + ], + "type": "text", + "content": "5. Caron, M., Touvron, H., Misra, I., J'egou, H., Mairal, J., Bojanowski, P., Joulin, A.: Emerging properties in self-supervised vision transformers. 2021 IEEE/CVF International Conference on Computer Vision (ICCV) pp. 9630-9640 (2021)" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 138, + 304, + 480, + 336 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 304, + 480, + 336 + ], + "spans": [ + { + "bbox": [ + 138, + 304, + 480, + 336 + ], + "type": "text", + "content": "6. Chen, T., Kornblith, S., Norouzi, M., Hinton, G.E.: A simple framework for contrastive learning of visual representations. International Conference on Machine Learning (2021)" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 138, + 337, + 480, + 369 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 337, + 480, + 369 + ], + "spans": [ + { + "bbox": [ + 138, + 337, + 480, + 369 + ], + "type": "text", + "content": "7. Chen, T., Zhai, X., Ritter, M., Lucic, M., Houlsby, N.: Self-supervised gans via auxiliary rotation loss. 2019 IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR) pp. 12146-12155 (2018)" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 138, + 370, + 480, + 391 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 370, + 480, + 391 + ], + "spans": [ + { + "bbox": [ + 138, + 370, + 480, + 391 + ], + "type": "text", + "content": "8. Chen, X., Fan, H., Girshick, R., He, K.: Improved baselines with momentum contrastive learning. arXiv preprint arXiv:2003.04297 (2020)" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 138, + 392, + 480, + 423 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 392, + 480, + 423 + ], + "spans": [ + { + "bbox": [ + 138, + 392, + 480, + 423 + ], + "type": "text", + "content": "9. Chen, Y., Liu, Y., Jiang, D., Zhang, X., Dai, W., Xiong, H., Tian, Q.: Sdae: Self-distillated masked autoencoder. In: European Conference on Computer Vision (2022)" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 138, + 425, + 480, + 456 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 425, + 480, + 456 + ], + "spans": [ + { + "bbox": [ + 138, + 425, + 480, + 456 + ], + "type": "text", + "content": "10. Cheng, X., Chen, J., Wang, R.: Modified dual attention triplet-supervised hashing network for image retrieval. Signal, Image and Video Processing 18, 1939-1948 (2023)" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 138, + 458, + 480, + 479 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 458, + 480, + 479 + ], + "spans": [ + { + "bbox": [ + 138, + 458, + 480, + 479 + ], + "type": "text", + "content": "1. Cuturi, M.: Sinkhorn distances: Lightspeed computation of optimal transport. In: Neural Information Processing Systems (2013)" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 138, + 480, + 480, + 501 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 480, + 480, + 501 + ], + "spans": [ + { + "bbox": [ + 138, + 480, + 480, + 501 + ], + "type": "text", + "content": "2. Deng, J., Dong, W., Socher, R., Li, L.J., Li, K., Fei-Fei, L.: ImageNet: A LargeScale Hierarchical Image Database. In: CVPR 2009 (2009)" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 138, + 502, + 480, + 533 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 502, + 480, + 533 + ], + "spans": [ + { + "bbox": [ + 138, + 502, + 480, + 533 + ], + "type": "text", + "content": "3. Dong, C., Loy, C.C., He, K., Tang, X.: Image super-resolution using deep convolutional networks. IEEE Transactions on Pattern Analysis and Machine Intelligence 38, 295-307 (2014)" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 138, + 534, + 480, + 577 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 534, + 480, + 577 + ], + "spans": [ + { + "bbox": [ + 138, + 534, + 480, + 577 + ], + "type": "text", + "content": "4. Dosovitskiy, A., Beyer, L., Kolesnikov, A., Weissenborn, D., Zhai, X., Unterthiner, T., Dehghani, M., Minderer, M., Heigold, G., Gelly, S., Uszkoreit, J., Houlsby, N.: An image is worth 16x16 words: Transformers for image recognition at scale. International Conference on Learning Representations (ICLR) (2021)" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 138, + 578, + 480, + 609 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 578, + 480, + 609 + ], + "spans": [ + { + "bbox": [ + 138, + 578, + 480, + 609 + ], + "type": "text", + "content": "5. Fang, Z., Wang, J., Wang, L., Zhang, L., Yang, Y., Liu, Z.: Seed: Self-supervised distillation for visual representation. International Conference on Learning Representations (2021)" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 138, + 611, + 480, + 643 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 611, + 480, + 643 + ], + "spans": [ + { + "bbox": [ + 138, + 611, + 480, + 643 + ], + "type": "text", + "content": "6. Gao, Y., Zhuang, J.X., Lin, S., Cheng, H., Sun, X., Li, K., Shen, C.: Disco: Remedying self-supervised learning on lightweight models with distilled contrastive learning. In: European Conference on Computer Vision (ECCV) (2022)" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 138, + 643, + 480, + 665 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 643, + 480, + 665 + ], + "spans": [ + { + "bbox": [ + 138, + 643, + 480, + 665 + ], + "type": "text", + "content": "7. Grill, J.B., Strub, F., Altché, F., Tallec, C., Richemond, P., Buchatskaya, E., Doersch, C., Avila Pires, B., Guo, Z., Gheshlaghi Azar, M., et al.: Bootstrap your own" + } + ] + } + ], + "index": 19 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 425, + 91, + 447, + 100 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 425, + 91, + 447, + 100 + ], + "spans": [ + { + "bbox": [ + 425, + 91, + 447, + 100 + ], + "type": "text", + "content": "POA" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 470, + 91, + 481, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 470, + 91, + 481, + 100 + ], + "spans": [ + { + "bbox": [ + 470, + 91, + 481, + 100 + ], + "type": "text", + "content": "15" + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 14 + }, + { + "para_blocks": [ + { + "bbox": [ + 132, + 117, + 481, + 665 + ], + "type": "list", + "angle": 0, + "index": 19, + "blocks": [ + { + "bbox": [ + 147, + 117, + 481, + 138 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 147, + 117, + 481, + 138 + ], + "spans": [ + { + "bbox": [ + 147, + 117, + 481, + 138 + ], + "type": "text", + "content": "latent-a new approach to self-supervised learning. Advances in neural information processing systems 33, 21271-21284 (2020)" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 133, + 139, + 481, + 172 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 139, + 481, + 172 + ], + "spans": [ + { + "bbox": [ + 133, + 139, + 481, + 172 + ], + "type": "text", + "content": "18. He, K., Chen, X., Xie, S., Li, Y., Doll'ar, P., Girshick, R.B.: Masked autoencoders are scalable vision learners. 2022 IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR) pp. 15979-15988 (2021)" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 133, + 172, + 481, + 205 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 172, + 481, + 205 + ], + "spans": [ + { + "bbox": [ + 133, + 172, + 481, + 205 + ], + "type": "text", + "content": "19. He, K., Fan, H., Wu, Y., Xie, S., Girshick, R.B.: Momentum contrast for unsupervised visual representation learning. 2020 IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR) pp. 9726-9735 (2019)" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 132, + 205, + 481, + 226 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 205, + 481, + 226 + ], + "spans": [ + { + "bbox": [ + 132, + 205, + 481, + 226 + ], + "type": "text", + "content": "20. He, K., Gkioxari, G., Dolkar, P., Girshick, R.B.: Mask r-cnn. 2017 IEEE International Conference on Computer Vision (ICCV) pp. 2980-2988 (2017)" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 132, + 226, + 481, + 258 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 226, + 481, + 258 + ], + "spans": [ + { + "bbox": [ + 132, + 226, + 481, + 258 + ], + "type": "text", + "content": "21. He, K., Zhang, X., Ren, S., Sun, J.: Deep residual learning for image recognition. 2016 IEEE Conference on Computer Vision and Pattern Recognition (CVPR) pp. 770-778 (2015)" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 132, + 259, + 481, + 281 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 259, + 481, + 281 + ], + "spans": [ + { + "bbox": [ + 132, + 259, + 481, + 281 + ], + "type": "text", + "content": "22. Hinton, G.E., Vinyals, O., Dean, J.: Distilling the knowledge in a neural network. ArXiv abs/1503.02531 (2015)" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 132, + 281, + 481, + 303 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 281, + 481, + 303 + ], + "spans": [ + { + "bbox": [ + 132, + 281, + 481, + 303 + ], + "type": "text", + "content": "23. Hou, L., Shen, H., Cao, Q., Cheng, X.: Self-supervised gans with label augmentation. ArXiv abs/2106.08601 (2021)" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 132, + 303, + 481, + 325 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 303, + 481, + 325 + ], + "spans": [ + { + "bbox": [ + 132, + 303, + 481, + 325 + ], + "type": "text", + "content": "24. Hu, C., Li, X., Liu, D., Wu, H., Chen, X., Wang, J., Liu, X.: Teacher-student architecture for knowledge distillation: A survey. ArXiv abs/2308.04268 (2023)" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 132, + 325, + 481, + 357 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 325, + 481, + 357 + ], + "spans": [ + { + "bbox": [ + 132, + 325, + 481, + 357 + ], + "type": "text", + "content": "25. Huang, Z., Jin, X., Lu, C., Hou, Q., Cheng, M.M., Fu, D., Shen, X., Feng, J.: Contrastive masked autoencoders are stronger vision learners. IEEE transactions on pattern analysis and machine intelligence PP (2022)" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 132, + 357, + 481, + 402 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 357, + 481, + 402 + ], + "spans": [ + { + "bbox": [ + 132, + 357, + 481, + 402 + ], + "type": "text", + "content": "26. Li, J., Wang, Y., ZHANG, X., Chen, Y., Jiang, D., Dai, W., Li, C., Xiong, H., Tian, Q.: Progressively compressed auto-encoder for self-supervised representation learning. In: The Eleventh International Conference on Learning Representations (2023)" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 132, + 402, + 481, + 445 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 402, + 481, + 445 + ], + "spans": [ + { + "bbox": [ + 132, + 402, + 481, + 445 + ], + "type": "text", + "content": "27. Li, Z., Zhu, Y., Yang, F., Li, W., Zhao, C., Chen, Y., Chen, Z., Xie, J., Wu, L., Zhao, R., Tang, M., Wang, J.: Univip: A unified framework for self-supervised visual pre-training. 2022 IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR) pp. 14607-14616 (2022)" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 132, + 446, + 481, + 489 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 446, + 481, + 489 + ], + "spans": [ + { + "bbox": [ + 132, + 446, + 481, + 489 + ], + "type": "text", + "content": "28. Lin, J., Rao, Y., Lu, J., Zhou, J.: Runtime neural pruning. In: Guyon, I., Luxburg, U.V., Bengio, S., Wallach, H., Fergus, R., Vishwanathan, S., Garnett, R. (eds.) Advances in Neural Information Processing Systems. vol. 30. Curran Associates, Inc. (2017)" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 132, + 490, + 481, + 522 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 490, + 481, + 522 + ], + "spans": [ + { + "bbox": [ + 132, + 490, + 481, + 522 + ], + "type": "text", + "content": "29. Lin, T.Y., Maire, M., Belongie, S.J., Hays, J., Perona, P., Ramanan, D., Dóllar, P., Zitnick, C.L.: Microsoft coco: Common objects in context. In: European Conference on Computer Vision (2014)" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 132, + 522, + 481, + 555 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 522, + 481, + 555 + ], + "spans": [ + { + "bbox": [ + 132, + 522, + 481, + 555 + ], + "type": "text", + "content": "30. Liu, Z., Lin, Y., Cao, Y., Hu, H., Wei, Y., Zhang, Z., Lin, S., Guo, B.: Swin transformer: Hierarchical vision transformer using shifted windows. 2021 IEEE/CVF International Conference on Computer Vision (ICCV) pp. 9992-10002 (2021)" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 132, + 555, + 481, + 578 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 555, + 481, + 578 + ], + "spans": [ + { + "bbox": [ + 132, + 555, + 481, + 578 + ], + "type": "text", + "content": "31. Loshchilov, I., Hutter, F.: Fixing weight decay regularization in adam. ArXiv abs/1711.05101 (2017)" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 132, + 578, + 481, + 599 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 578, + 481, + 599 + ], + "spans": [ + { + "bbox": [ + 132, + 578, + 481, + 599 + ], + "type": "text", + "content": "32. van den Oord, A., Li, Y., Vinyals, O.: Representation learning with contrastive predictive coding. ArXiv abs/1807.03748 (2018)" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 132, + 599, + 481, + 665 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 599, + 481, + 665 + ], + "spans": [ + { + "bbox": [ + 132, + 599, + 481, + 665 + ], + "type": "text", + "content": "33. Oquab, M., Darcet, T., Moutakanni, T., Vo, H.V., Szafraniec, M., Khalidov, V., Fernandez, P., HAZIZA, D., Massa, F., El-Nouby, A., Assran, M., Ballas, N., Galuba, W., Howes, R., Huang, P.Y., Li, S.W., Misra, I., Rabbat, M., Sharma, V., Synnaeve, G., Xu, H., Jegou, H., Mairal, J., Labatut, P., Joulin, A., Bojanowski, P.: DINOv2: Learning robust visual features without supervision. Transactions on Machine Learning Research (2024)" + } + ] + } + ], + "index": 18 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "spans": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "type": "text", + "content": "16" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 166, + 91, + 229, + 102 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 166, + 91, + 229, + 102 + ], + "spans": [ + { + "bbox": [ + 166, + 91, + 229, + 102 + ], + "type": "text", + "content": "Y. Zhang et al." + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 15 + }, + { + "para_blocks": [ + { + "bbox": [ + 130, + 116, + 482, + 665 + ], + "type": "list", + "angle": 0, + "index": 19, + "blocks": [ + { + "bbox": [ + 130, + 116, + 482, + 149 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 116, + 482, + 149 + ], + "spans": [ + { + "bbox": [ + 130, + 116, + 482, + 149 + ], + "type": "text", + "content": "34. Ruan, Y., Singh, S., Morningstar, W.R., Alemi, A.A., Ioffe, S., Fischer, I., Dillon, J.V.: Weighted ensemble self-supervised learning. In: The Eleventh International Conference on Learning Representations (2023)" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 130, + 150, + 481, + 183 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 150, + 481, + 183 + ], + "spans": [ + { + "bbox": [ + 130, + 150, + 481, + 183 + ], + "type": "text", + "content": "35. Song, K., Xie, J., Zhang, S., Luo, Z.: Multi-mode online knowledge distillation for self-supervised visual representation learning. 2023 IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR) pp. 11848-11857 (2023)" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 130, + 183, + 481, + 226 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 183, + 481, + 226 + ], + "spans": [ + { + "bbox": [ + 130, + 183, + 481, + 226 + ], + "type": "text", + "content": "36. Tao, C., Wang, H., Zhu, X., Dong, J., Song, S., Huang, G., Dai, J.: Exploring the equivalence of siamese self-supervised learning via a unified gradient framework. 2022 IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR) pp. 14411-14420 (2021)" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 130, + 227, + 481, + 248 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 227, + 481, + 248 + ], + "spans": [ + { + "bbox": [ + 130, + 227, + 481, + 248 + ], + "type": "text", + "content": "37. Team, G.G.: Introducing gemini: our largest and most capable ai model. Google (2023)" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 130, + 249, + 481, + 282 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 249, + 481, + 282 + ], + "spans": [ + { + "bbox": [ + 130, + 249, + 481, + 282 + ], + "type": "text", + "content": "38. Tian, K., Jiang, Y., qishuai diao, Lin, C., Wang, L., Yuan, Z.: Designing BERT for convolutional networks: Sparse and hierarchical masked modeling. In: The Eleventh International Conference on Learning Representations (2023)" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 130, + 282, + 481, + 326 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 282, + 481, + 326 + ], + "spans": [ + { + "bbox": [ + 130, + 282, + 481, + 326 + ], + "type": "text", + "content": "39. Tomasev, N., Bica, I., McWilliams, B., Buesing, L., Pascanu, R., Blundell, C., Mitrovic, J.: Pushing the limits of self-supervised resnets: Can we outperform supervised learning without labels onImagenet? arXiv preprint arXiv:2201.05119 (2022)" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 130, + 326, + 481, + 369 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 326, + 481, + 369 + ], + "spans": [ + { + "bbox": [ + 130, + 326, + 481, + 369 + ], + "type": "text", + "content": "40. Touvron, H., Cord, M., Oquab, M., Bojanowski, P., Verbeek, J., Jegou, H.: Cotraining 2l submodels for visual recognition. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR). pp. 11701-11710 (June 2023)" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 130, + 369, + 481, + 403 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 369, + 481, + 403 + ], + "spans": [ + { + "bbox": [ + 130, + 369, + 481, + 403 + ], + "type": "text", + "content": "41. Wang, X., Huang, Y., Zeng, D., Qi, G.J.: Caco: Both positive and negative samples are directly learnable via cooperative-adversarial contrastive learning. IEEE Transactions on Pattern Analysis and Machine Intelligence 45, 10718-10730 (2022)" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 130, + 403, + 481, + 424 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 403, + 481, + 424 + ], + "spans": [ + { + "bbox": [ + 130, + 403, + 481, + 424 + ], + "type": "text", + "content": "42. Wang, X., Yu, F., Dou, Z.Y., Gonzalez, J.: Skipnet: Learning dynamic routing in convolutional networks. ArXiv abs/1711.09485 (2017)" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 130, + 424, + 481, + 457 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 424, + 481, + 457 + ], + "spans": [ + { + "bbox": [ + 130, + 424, + 481, + 457 + ], + "type": "text", + "content": "43. Wang, X., Zhang, R., Shen, C., Kong, T., Li, L.: Dense contrastive learning for self-supervised visual pre-training. 2021 IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR) pp. 3023-3032 (2020)" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 130, + 457, + 481, + 512 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 457, + 481, + 512 + ], + "spans": [ + { + "bbox": [ + 130, + 457, + 481, + 512 + ], + "type": "text", + "content": "44. Wortsman, M., Ilharco, G., Gadre, S.Y., Roelofs, R., Gontijo-Lopes, R., Morcos, A.S., Namkoong, H., Farhadi, A., Carmon, Y., Kornblith, S., Schmidt, L.: Model soups: averaging weights of multiple fine-tuned models improves accuracy without increasing inference time. In: International Conference on Machine Learning (ICML) (2022)" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 130, + 512, + 481, + 544 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 512, + 481, + 544 + ], + "spans": [ + { + "bbox": [ + 130, + 512, + 481, + 544 + ], + "type": "text", + "content": "45. Wu, H., Xu, J., Wang, J., Long, M.: Autoformer: Decomposition transformers with auto-correlation for long-term series forecasting. In: Neural Information Processing Systems (2021)" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 130, + 545, + 481, + 567 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 545, + 481, + 567 + ], + "spans": [ + { + "bbox": [ + 130, + 545, + 481, + 567 + ], + "type": "text", + "content": "46. Xiao, T., Liu, Y., Zhou, B., Jiang, Y., Sun, J.: Unified perceptual parsing for scene understanding. ArXiv abs/1807.10221 (2018)" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 130, + 567, + 481, + 589 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 567, + 481, + 589 + ], + "spans": [ + { + "bbox": [ + 130, + 567, + 481, + 589 + ], + "type": "text", + "content": "47. Xie, Z., Lin, Y., Yao, Z., Zhang, Z., Dai, Q., Cao, Y., Hu, H.: Self-supervised learning with swin transformers. ArXiv abs/2105.04553 (2021)" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 130, + 589, + 481, + 622 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 589, + 481, + 622 + ], + "spans": [ + { + "bbox": [ + 130, + 589, + 481, + 622 + ], + "type": "text", + "content": "48. Xie, Z., Zhang, Z., Cao, Y., Lin, Y., Bao, J., Yao, Z., Dai, Q., Hu, H.: Simmim: a simple framework for masked image modeling. 2022 IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR) pp. 9643-9653 (2021)" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 130, + 622, + 481, + 643 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 622, + 481, + 643 + ], + "spans": [ + { + "bbox": [ + 130, + 622, + 481, + 643 + ], + "type": "text", + "content": "49. Yan, B., Jiang, Y., Sun, P., Wang, D., Yuan, Z., Luo, P., Lu, H.: Towards grand unification of object tracking. In: European Conference on Computer Vision (2022)" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 130, + 643, + 481, + 665 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 643, + 481, + 665 + ], + "spans": [ + { + "bbox": [ + 130, + 643, + 481, + 665 + ], + "type": "text", + "content": "50. Yan, C., Chang, X., Li, Z., Yao, L., Luo, M., Zheng, Q.: Masked distillation advances self-supervised transformer architecture search. In: The Twelfth Interna" + } + ] + } + ], + "index": 18 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 425, + 91, + 447, + 100 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 425, + 91, + 447, + 100 + ], + "spans": [ + { + "bbox": [ + 425, + 91, + 447, + 100 + ], + "type": "text", + "content": "POA" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 470, + 91, + 481, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 470, + 91, + 481, + 100 + ], + "spans": [ + { + "bbox": [ + 470, + 91, + 481, + 100 + ], + "type": "text", + "content": "17" + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 16 + }, + { + "para_blocks": [ + { + "bbox": [ + 132, + 117, + 481, + 281 + ], + "type": "list", + "angle": 0, + "index": 8, + "blocks": [ + { + "bbox": [ + 147, + 117, + 481, + 138 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 147, + 117, + 481, + 138 + ], + "spans": [ + { + "bbox": [ + 147, + 117, + 481, + 138 + ], + "type": "text", + "content": "tional Conference on Learning Representations (2024), https://openreview.net/ forum?id=LUpC8KTvdV" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 133, + 139, + 481, + 160 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 139, + 481, + 160 + ], + "spans": [ + { + "bbox": [ + 133, + 139, + 481, + 160 + ], + "type": "text", + "content": "51. Yu, F., Huang, K., Wang, M., Cheng, Y., Chu, W., Cui, L.: Width & depth pruning for vision transformers. In: AAAI Conference on Artificial Intelligence (2022)" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 133, + 161, + 481, + 194 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 161, + 481, + 194 + ], + "spans": [ + { + "bbox": [ + 133, + 161, + 481, + 194 + ], + "type": "text", + "content": "52. Yu, J., Li, X., Koh, J.Y., Zhang, H., Pang, R., Qin, J., Ku, A., Xu, Y., Baldridge, J., Wu, Y.: Vector-quantized image modeling with improved VQGAN. In: International Conference on Learning Representations (2022)" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 132, + 194, + 481, + 216 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 194, + 481, + 216 + ], + "spans": [ + { + "bbox": [ + 132, + 194, + 481, + 216 + ], + "type": "text", + "content": "53. Zhang, Y., Zhong, Q., Ma, L., Xie, D., Pu, S.: Learning incremental triplet margin for person re-identification. AAAI Conference on Artificial Intelligence (2019)" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 132, + 216, + 481, + 248 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 216, + 481, + 248 + ], + "spans": [ + { + "bbox": [ + 132, + 216, + 481, + 248 + ], + "type": "text", + "content": "54. Zhou, B., Zhao, H., Puig, X., Fidler, S., Barriuso, A., Torralba, A.: Scene parsing through ade20k dataset. 2017 IEEE Conference on Computer Vision and Pattern Recognition (CVPR) pp. 5122-5130 (2017)" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 132, + 249, + 481, + 281 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 249, + 481, + 281 + ], + "spans": [ + { + "bbox": [ + 132, + 249, + 481, + 281 + ], + "type": "text", + "content": "55. Zhou, J., Wei, C., Wang, H., Shen, W., Xie, C., Yuille, A., Kong, T.: ibot: Image bert pre-training with online tokenizer. International Conference on Learning Representations (ICLR) (2022)" + } + ] + } + ], + "index": 7 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "spans": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "type": "text", + "content": "18" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 166, + 91, + 229, + 102 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 166, + 91, + 229, + 102 + ], + "spans": [ + { + "bbox": [ + 166, + 91, + 229, + 102 + ], + "type": "text", + "content": "Y. Zhang et al." + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 17 + } + ], + "_backend": "vlm", + "_version_name": "2.6.4" +} \ No newline at end of file diff --git a/2024/POCA_ Post-training Quantization with Temporal Alignment for Codec Avatars/f5688d31-6662-4aa8-b308-c286c399e498_content_list.json b/2024/POCA_ Post-training Quantization with Temporal Alignment for Codec Avatars/f5688d31-6662-4aa8-b308-c286c399e498_content_list.json new file mode 100644 index 0000000000000000000000000000000000000000..3d19a917f8edb602d86fb97c4b281f668fb85f10 --- /dev/null +++ b/2024/POCA_ Post-training Quantization with Temporal Alignment for Codec Avatars/f5688d31-6662-4aa8-b308-c286c399e498_content_list.json @@ -0,0 +1,1672 @@ +[ + { + "type": "text", + "text": "POCA: Post-training Quantization with Temporal Alignment for Codec Avatars", + "text_level": 1, + "bbox": [ + 274, + 140, + 728, + 186 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Jian Meng $^{1,2}$ , Yuecheng Li $^{1}$ , Chenghui (Leo) Li $^{1}$ , Syed Shakib Sarwar $^{1}$ , Dilin Wang $^{1}$ , and Jae-sun Seo $^{2}$", + "bbox": [ + 223, + 210, + 779, + 243 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "1 Meta Codec Avatars Lab, Pittsburgh, USA", + "bbox": [ + 349, + 253, + 651, + 268 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "2 Cornell Tech, New York City, USA", + "bbox": [ + 375, + 268, + 624, + 282 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Abstract. Real-time decoding generates high-quality assets for rendering photorealistic Codec Avatars for immersive social telepresence with AR/VR. However, high-quality avatar decoding incurs expensive computation and memory consumption, which necessitates the design of a decoder compression algorithm (e.g., quantization). Although quantization has been widely studied, the quantization of avatar decoders is an urgent yet under-explored need. Furthermore, the requirement of fast \"User-Avatar\" deployment prioritizes the post-training quantization (PTQ) over the time-consuming quantization-aware training (QAT). As the first work in this area, we reveal the sensitivity of the avatar decoding quality under low precision. In particular, the state-of-the-art (SoTA) QAT and PTQ algorithms introduce massive amounts of temporal noise to the rendered avatars, even with the well-established 8-bit precision. To resolve these issues, a novel PTQ algorithm is proposed for quantizing the avatar decoder with low-precision weights and activation (8-bit and 6-bit), without introducing temporal noise to the rendered avatar. Furthermore, the proposed method only needs $10\\%$ of the activations of each layer to calibrate quantization parameters without any distribution manipulations or extensive boundary search. The proposed method is evaluated on various face avatars with different facial characteristics. The proposed method compresses the decoder model by $5.3\\times$ while recovering the quality on par with the full precision baseline. In addition to the avatar rendering tasks, POCA is also applicable to image resolution enhancement tasks, achieving new SoTA image quality. https://mengjian0502.github.io/poca.github.io/", + "bbox": [ + 261, + 316, + 743, + 664 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "1 Introduction", + "text_level": 1, + "bbox": [ + 215, + 688, + 375, + 704 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Photorealistic telepresence [14, 15, 23, 27] has emerged as the cornerstone technology in augmented or virtual reality (AR/VR), enabling virtual interactions across different devices. Specifically, the ultimate goal of photorealistic rendering is animating the high-fidelity Codec Avatars with multiple partial facial images captured by the head-mounted cameras (HMC) [29]. The variational autoencoder-based model [14, 15] has been proved and deployed to VR device (e.g., Quest Pro) as a feasible solution. During inference, the encoder model projects the HMC-captured facial motion into the latent space, and the decoder", + "bbox": [ + 212, + 718, + 787, + 840 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/4f5bb62931d49a77b0787612bcf87fe07dc0cbda4605eb3139e2927f11d35465.jpg", + "image_caption": [ + "Fig. 1: Decoding and rendering with low-precision avatar decoder. The SoTA 8-bit PTQ algorithm distorts the avatar with massive and random temporal noise, while the proposed method compresses the decoder with the high-quality avatar without noise." + ], + "image_footnote": [ + "Noise is visible in the supplementary video: \"supplementary/ID002643814/poca_all_002643814_w8a8_400029.mp4\"" + ], + "bbox": [ + 263, + 145, + 738, + 362 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "reconstructs the latent code into view dependent high-quality texture and mesh, which are then rendered into the 3-D Codec AvATAR on screen. In practice, the critical ingredient of photorealistic telepresence is transmitting the encoded latent information from one VR to another and rendering the avatar in real-time. In practice, the VAE-based Codec AvATAR model has been employed for generative Pixel Codec AvATAR [15] or 3D Gaussian-based AvATAR [22]. Naturally, the quality and latency of telepresence mainly depends on the efficiency of the decoder, which impacts the real-time visual quality.", + "bbox": [ + 212, + 460, + 787, + 582 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Despite the success of the deep appearance model [14], the high computational cost and memory consumption of avatar decoding hinder the fast on-device telepresence. Naïvely deploying the full-precision decoder to the headset leads to a low frame rate, causing degraded user experience. On the other hand, relying on cloud-based computation adds additional \"headset-cloud\" communication latency and increases the risk of avatar data breaching.", + "bbox": [ + 212, + 582, + 787, + 672 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "The conflict between high-fidelity rendering and the urgent needs of on-device computing necessitates the high-performance compression algorithm designed for the avatar decoder. In particular, quantization [2, 18, 28, 30] compresses the model with a unified data compression ratio on weights and activation, leading to easy hardware deployment [24]. Unfortunately, the SoTA quantization algorithms mainly focus on classification [11, 18], object detection [26], and natural language processing tasks [30]. Recent works have investigated high-resolution static image rendering with the low-precision model [25]. However, the high-quality quantization algorithm for avatar decoding remains unexplored. In fact, our research shows that the current efforts on quantization are highly incompatible to the Avatar decoder with the following challenges:", + "bbox": [ + 212, + 672, + 787, + 840 + ], + "page_idx": 1 + }, + { + "type": "page_number", + "text": "2", + "bbox": [ + 217, + 114, + 228, + 126 + ], + "page_idx": 1 + }, + { + "type": "header", + "text": "J. Meng et al.", + "bbox": [ + 271, + 114, + 364, + 127 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Massive temporal noises introduced by INT8 quantization. Although the popular 8-bit quantization has been widely investigated with quantization-aware training [20] and post-training quantization [2,18,24], our discovery shows that the well-established INT8 precision introduces a massive amount of temporal noises to the rendered avatar, even with the SoTA methods. As shown in Figure 1, the facial motion generated by the 8-bit decoder is filled with temporally flickering noises between two consecutive frames, which immerses the original facial expressions and degrades the quality of the rendered avatar. Most importantly, the temporal noise is not visible in the static single frame, distorting the avatar in a continuous manner. As the first work in this area, this paper reveals the root cause of the temporal avatar noise, as described in Section 3.", + "bbox": [ + 212, + 146, + 787, + 313 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Difficulty of characterizing temporal noise. Common quantization noises in conventional classification and object detection tasks can be modeled by a standard statistical model [12]. However, the temporal avatar noises introduced by quantized decoder exhibit inconsistent intensities and granularities in different facial regions. As a result, it is almost impossible to model the temporal noises globally without additional recurrent information. Furthermore, applying an additional filtering strategy after quantization introduces additional energy and latency consumption to the AR/VR devices, which deteriorates the efficiency of Codec Avatar telecommunication. Therefore, the noise suppression process should be achieved all at once during post training quantization.", + "bbox": [ + 212, + 325, + 787, + 474 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "To resolve the aforementioned problems, we propose POCA, a novel Post Training Quantization (PTQ) algorithm designed for high quality and temporally noise-free Codec Avatar decoding with outstanding simplicity and efficiency. The major contributions of this work are listed as follows:", + "bbox": [ + 212, + 476, + 787, + 536 + ], + "page_idx": 2 + }, + { + "type": "list", + "sub_type": "text", + "list_items": [ + "- As the first work that investigates the low-precision avatar decoding, we reveal the root cause of the temporal noises induced by the temporally dynamic and long-tailed distribution.", + "- POCA eliminates the temporal facial noise with 8-bit and 6-bit precision, outperforming the SoTA method with fully recovered avatar quality [16].", + "- POCA only requires $\\leq 10\\%$ of layer-wise feature map data and $\\leq 5\\%$ of the total training data for PTQ.", + "- POCA is evaluated on a variety of avatar identities with a full range of viewing directions and dynamic expressions, exhibiting a high degree of robustness and consistent noise-free avatar animation across different genders, age groups, hairstyles, skin colors, and view directions.", + "- POCA is also applicable to image super resolution task [8,25], achieving the SoTA image quality with $80\\%$ calibration sparsity." + ], + "bbox": [ + 225, + 542, + 784, + 736 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "2 Related Work", + "text_level": 1, + "bbox": [ + 215, + 757, + 387, + 773 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "2.1 Codec Avatar", + "text_level": 1, + "bbox": [ + 215, + 787, + 375, + 801 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Photorealistic human face modeling and animation are challenging computer vision and computer graphics tasks in the sense of achieving minimum visual", + "bbox": [ + 212, + 809, + 785, + 839 + ], + "page_idx": 2 + }, + { + "type": "header", + "text": "POCA", + "bbox": [ + 683, + 114, + 730, + 126 + ], + "page_idx": 2 + }, + { + "type": "page_number", + "text": "3", + "bbox": [ + 774, + 116, + 784, + 126 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "distortion for reconstructed 3-D human faces. Early works mainly focus on accurate 3-D reconstruction with blend shape models [9], which are computationally expensive and impossible for mobile on-device telepresence. Empowered by the deep learning models, recent research [14,15] considers human face reconstruction as data-driven generative task with the variational auto-encoder neural network. In particular, the variational encoder [6] is trained for encoding the multi-view HMC images into the latent information $\\mathbf{z}$ , which describes the state of the face with rich expression and motion. Subsequently, the decoder network reconstructs the 3-D faces based on the latent code and the given view direction $\\mathbf{v}$ , where the assets for rendering 3D face avatar are decoded separately by the dedicated mesh decoder and texture decoder.", + "bbox": [ + 212, + 146, + 787, + 311 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "The compact latent code is transmitted between VR devices with low communication bandwidth requirements. However, the desire for real-time, on-device 3-D face avatar animation is hindered by the expensive computation cost and high memory consumption of the decoder. Various efforts have been established to improve the computation efficiency of the VAE model [5, 14]. Pixel Codec Avatar [15] proposes a novel architecture to allow lightweight decoder by creatively leveraging rendering pipeline features, skipping computations for invisible pixels while maintaining the high-fidelity for visible regions. Recent research optimizes the encoder architecture via neural architecture search (NAS) [3], leading to reduced computational costs. Different from such prior works, our work proposes a decoder-dedicated quantization algorithm that enables temporal alignments and noise-free avatar decoding with low-precision model, which further enhances instant and high quality on-device deployment of Codec Avatars for resource-constrained mobile AR/VR headsets.", + "bbox": [ + 212, + 311, + 787, + 523 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "2.2 Quantization", + "text_level": 1, + "bbox": [ + 215, + 545, + 370, + 559 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Quantization has been widely investigated together with the evolution of deep learning. Starting from the well-established 8-bit precision [4,24] all the way to the aggressive binary neural network [32], the quantized model largely improves the computation and memory efficiency, making the complex DNN models deployable to the resource-constrained edge devices (e.g., AR/VR headset). The core idea of quantization is constructing the accurate matching between the high-precision data (e.g., FP32) and the low-precision representation. Normally, the quantization process can be generalized into the \"quantize-and-dequantize\" workflow connected by the shared scaling factor in between. Mathematically, given the full precision weights $\\mathcal{W}^l$ and input activation $\\mathcal{X}^l$ of a layer $l$ , we have:", + "bbox": [ + 212, + 569, + 787, + 720 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\n\\mathcal {W} _ {Q} ^ {l} = \\left\\lfloor \\frac {\\mathcal {W} ^ {l}}{S _ {w} ^ {l}} \\right] \\in \\operatorname {I N T} _ {\\mathbf {n}}, \\quad \\mathcal {X} _ {Q} ^ {l} = \\left\\lfloor \\frac {\\mathcal {X} ^ {l}}{S _ {x} ^ {l}} \\right] \\in \\operatorname {I N T} _ {\\mathbf {n}} \\tag {1}\n$$\n", + "text_format": "latex", + "bbox": [ + 336, + 731, + 785, + 765 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Where $S$ is the scaling factor determined by data precision $\\mathbf{n}$ and the numerical range. Specifically, $S$ characterizes the \"interval\" between adjacent levels in the floating point dequantized values, bounded by the range [lb, ub]:", + "bbox": [ + 212, + 768, + 787, + 814 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\nS = | \\mathrm {u b} - \\mathrm {l b} | / 2 ^ {\\mathbf {n}} - 1 \\tag {2}\n$$\n", + "text_format": "latex", + "bbox": [ + 426, + 825, + 785, + 840 + ], + "page_idx": 3 + }, + { + "type": "page_number", + "text": "4", + "bbox": [ + 217, + 114, + 228, + 126 + ], + "page_idx": 3 + }, + { + "type": "header", + "text": "J. Meng et al.", + "bbox": [ + 271, + 114, + 364, + 128 + ], + "page_idx": 3 + }, + { + "type": "image", + "img_path": "images/fb71659a89c7e4a15b921ebd294a5c266747bcdaa27d6bf348d6432f958aecf8.jpg", + "image_caption": [ + "Fig. 2: Top: The rendered avatar exhibits minimal quality (PSNR) reduction in static frames and large distortion with animation. Bottom: SoTA 8-bit quantization [2,18] amplifies the pixel shifting between frames in the rendered avatar across RGB channels." + ], + "image_footnote": [], + "bbox": [ + 261, + 143, + 738, + 412 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Since the number of levels is directly determined by the precision $\\mathbf{n}$ , the essence of quantization accurately approximates the floating-point distribution with the dequantized tensors while minimizing the model-level error. To that end, a variety of quantization algorithms have been proposed to locate the optimal scaling factors with QAT [13] or PTQ [11, 18, 28], including boundary learning [1, 10], scaling factor learning [2], and quantization error modeling [12], etc. Recent efforts introduce additional scaling to avoid the error caused by the extreme distribution [30]. Promoted by the fast compression process, PTQ exhibits high compatibility for quantizing the avatar VAE model compared to QAT. Ideally, compressing the Codec Avatar models with PTQ would lead to instant \"user-avatar\" conversion for AR/VR devices, compared to time-consuming QAT. Recent VAE-based quantization focuses on the video compression task with quantized latent code [7, 21], while the quantization of the codec-avatar decoder model has not been explored despite the urgent needs on edge device and high rendering quality.", + "bbox": [ + 212, + 497, + 787, + 726 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "3 Root Cause of Temporal Avatar Noise", + "text_level": 1, + "bbox": [ + 214, + 750, + 622, + 767 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "3.1 Temporal Noise is Dynamic Pixel Color Oscillation", + "text_level": 1, + "bbox": [ + 214, + 782, + 683, + 797 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Recent 3D rendering tasks (e.g., NeRF) [17] uses peak signal-to-noise ratio (PSNR) to evaluate the quality between the rendered output and the ground truth im", + "bbox": [ + 212, + 809, + 795, + 840 + ], + "page_idx": 4 + }, + { + "type": "header", + "text": "POCA", + "bbox": [ + 683, + 114, + 730, + 126 + ], + "page_idx": 4 + }, + { + "type": "page_number", + "text": "5", + "bbox": [ + 774, + 116, + 784, + 126 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "age. However, the temporal distortion of Codec Avatar cannot be effectively quantified with the same metrics.", + "bbox": [ + 212, + 146, + 782, + 175 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "With the low-precision decoding, the temporal noise shows minimal impact on the static frame quality, but largely distorts the dynamic motion of the avatar, as shown in Figure 2 and supplementary material. Prior works only focus on the static image quality [25] with low-precision models, but the dynamic quality reduction has not been investigated well.", + "bbox": [ + 212, + 176, + 784, + 252 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "In addition to the FovVideoVDP score [16], the temporal pixel oscillation can be characterized by the mean pixel value differences between two consecutive frames $F_{t}$ and $F_{t + 1}$ of the rendered avatar:", + "bbox": [ + 212, + 253, + 784, + 297 + ], + "page_idx": 5 + }, + { + "type": "equation", + "text": "\n$$\n\\Delta_ {F} ^ {c} = \\frac {1}{\\mathcal {H W}} \\times | F _ {t + 1} - F _ {t} | \\tag {3}\n$$\n", + "text_format": "latex", + "bbox": [ + 411, + 311, + 784, + 339 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Where $\\mathcal{H}$ and $\\mathcal{W}$ represent the height and width of the rendered frame. We evaluate the pixel value differences separately for RGB channels such that $c \\in \\{R, G, B\\}$ . The trend of the pixel changes characterizes the basic transition of the facial expression (e.g., Tongue in and out), as shown in Figure 2 (Bottom).", + "bbox": [ + 214, + 348, + 785, + 409 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "With the 8-bit VAE decoder, the frame-wise pixel dynamics exhibit a similar trend as the full-precision decoder baseline. However, the quantized decoder adds an additional \"offset\" to the transition of the facial motion. In other words, the additional distortion is introduced to the avatar on top of the natural motion, which leads to random texture fluctuation in the rendered Codec Avatar. Figure 2 (Bottom) shows that the temporal noises uniformly exist across the entire facial motion with low intensity of $\\pm 0.44$ , $\\pm 0.47$ , and $\\pm 0.25$ average pixel shifting on the R, G, B channels, respectively. We are fully aware of the existence of other motion capture matrices (e.g., optical flow), but the temporal noises are significantly fine-grained, which requires detailed pixel-level analysis.", + "bbox": [ + 212, + 410, + 785, + 560 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Formally, such random pixel flickering can be captured by the FovVideoVDP [16] score between the rendered avatar and ground truth image. As shown in Figure 2, the INT8 avatar decoder largely degrades the animation quality of the avatar, even with the SoTA 8-bit quantization method [2, 18]. The high density of the yellow region indicates the strongly mismatched animation caused by the temporal noise. The continuous pixel oscillation and temporal noises worsen the fidelity of the rendered Codec Avatar in motion.", + "bbox": [ + 212, + 561, + 785, + 667 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "3.2 Root Cause of Noisy Avatars", + "text_level": 1, + "bbox": [ + 214, + 691, + 501, + 707 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Since the temporal noises are induced by the pixel oscillation of the avatar in motion, an intuitive understanding of the noises is the \"unstable rounding\" during quantization [12, 19]. However, the rounding instability is still the phenomenon of quantization instead of the root cause of the noise. In this work, we unravel the fundamental reason of the temporal noise induced by low-precision avatar decoding.", + "bbox": [ + 212, + 718, + 784, + 808 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Dynamically long-tailed distribution of human face. Human facial features naturally display a diverse range of distributions characterized by dif", + "bbox": [ + 212, + 809, + 785, + 839 + ], + "page_idx": 5 + }, + { + "type": "page_number", + "text": "6", + "bbox": [ + 217, + 114, + 228, + 126 + ], + "page_idx": 5 + }, + { + "type": "header", + "text": "J. Meng et al.", + "bbox": [ + 271, + 114, + 364, + 128 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/7edb81847558169fa6b09c1dff05d88de636549afab9f57ace19fd39adf3339a.jpg", + "image_caption": [ + "(a)" + ], + "image_footnote": [], + "bbox": [ + 282, + 145, + 478, + 277 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/f0c263706cb0acc9a3dd87cc5cec5cf3c3d880209167bea05464059504959cde.jpg", + "image_caption": [ + "(b)", + "Fig. 3: (a) Row-channel plane of the texture decoder activation and (b) Regional standard deviation of the row-channel planes with respect to different facial expressions." + ], + "image_footnote": [], + "bbox": [ + 480, + 152, + 728, + 277 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "ferent facial regions (for instance, hair compared to cheeks). The broad spectrum of variability in human facial characteristics formulates various statistical properties. For the avatar decoding, the texture decoder of the VAE model [14] decodes face avatar texture from the latent space via transposed convolutional layers, which consecutively up-scales the activation feature map from $4 \\times 4$ to $1024 \\times 1024$ (height and width). Therefore, the statistical variation of the facial texture is also embedded into the activation of the earlier layer.", + "bbox": [ + 212, + 364, + 784, + 470 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Proof of facial dynamics. We prove the dynamic range variations of the texture by analyzing the regional distribution of the latent layer activation. Assume $\\mathcal{C}$ , $\\mathcal{H}$ , and $\\mathcal{W}$ represent the channel depth, height, and width of the activation tensor. We analyze the statistics by decomposing the entire face texture feature map into multiple \"row-channel planes\". Each plane contains the regional statistics from hair to jaw, with the total number of planes = $\\mathcal{H}$ .", + "bbox": [ + 212, + 474, + 785, + 566 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Figure 3(a) shows the row-channel decomposition based on the intermediate activation feature map of the texture decoder with the shape of $16 \\times 256 \\times 256$ . On top of that, we compute the standard deviation of each plane, which characterizes the local tailedness and dynamic range in each region of the face, with respect to different facial expressions. As depicted in Figure 3(b), smiling and eye blinking lead to the long-tailed dynamics in the mouth and eye regions, respectively. As a result, the distribution of the entire face is dynamically long-tailed, which is harmful for quantization. In other words:", + "bbox": [ + 212, + 571, + 785, + 691 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "The distribution of a decoder feature map contains both constrained and long-tailed distribution, while the tailedness varies between different facial expressions.", + "bbox": [ + 214, + 698, + 785, + 729 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "In other words, the root cause of the temporal noises is the conflict between the long-tailed, highly dynamic distribution of activations and the global error minimization during PTQ. To minimize the global quantization error, the long-tailed distribution stretches the quantization interval (scaling factor) in order to cover larger range, which in turn, elevates the local rounding instability and leads to unstable numerical representation of the pixels. We prove this hypothesis in Section 4.3 with the proposed method.", + "bbox": [ + 212, + 734, + 785, + 839 + ], + "page_idx": 6 + }, + { + "type": "header", + "text": "POCA", + "bbox": [ + 683, + 114, + 730, + 126 + ], + "page_idx": 6 + }, + { + "type": "page_number", + "text": "7", + "bbox": [ + 774, + 114, + 784, + 126 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "4 Proposed Method: POCA", + "text_level": 1, + "bbox": [ + 215, + 143, + 503, + 162 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "To resolve the temporal noises caused by low-precision avatar decoding, this section presents POCA, a novel post-training quantization algorithm that successfully compresses the deep appearance decoder and eliminates the temporal noises from the rendered Codec Avatar. In particular, POCA is empowered by two novel techniques: a) View-aware PTQ and b) PTQ with shape-wise filtering.", + "bbox": [ + 212, + 179, + 787, + 257 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "4.1 View-aware PTQ", + "text_level": 1, + "bbox": [ + 215, + 281, + 405, + 297 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "As pointed out in Section 3, the face texture contains both region-dependent dispersion and facial expression-driven dynamics in the entire face. As a result, the sample space of the training data and layer-wise activation contains high variation with redundant information. In particular, the texture decoder is trained by the whole face texture map with the size of $1024 \\times 1024$ . However, the final 3D avatar is rendered based on the visibility information from the given view direction. Therefore, quantizing the decoder without considering the visibility for inner layers leads to the incorrectly learned quantization parameter, which further generates the mismatched quantization decision between PTQ and actual rendering. Motivated by that, we propose the view-aware PTQ for the avatar texture decoder.", + "bbox": [ + 212, + 310, + 787, + 474 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Basic Setup. Given the pre-trained full-precision VAE decoder (texture and mesh), we first randomly select $m$ training samples from the full-sized dataset with size $\\mathcal{D}$ ( $m << \\mathcal{D}$ ). Each training sample contains the input and pre-computed output of the decoder ( $[z_t, v_t], [\\hat{\\mathbf{T}}_t, \\hat{\\mathbf{M}}_t]$ ), where $z_t, v_t, \\hat{\\mathbf{T}}_t$ , and $\\hat{\\mathbf{M}}_t$ represents the encoded latent code, view direction, decoded texture, and decoded mesh for sample $t$ , respectively. Following the standard PTQ strategy, we extract input $\\mathcal{X}^l$ and output activation $\\mathcal{Y}^l$ of each layer as the calibration data. We adopted AdaRound [18] and LSQ [2] as the weight and activation quantizers, respectively. The learnable quantization parameters ( $\\alpha, S^l$ ) are optimized by minimizing the mean square error (MSE) between the quantized layer output $\\hat{\\mathcal{Y}}^l$ and full precision output $\\mathcal{Y}^l$ :", + "bbox": [ + 212, + 477, + 787, + 645 + ], + "page_idx": 7 + }, + { + "type": "equation", + "text": "\n$$\n\\underset {\\alpha , S ^ {l}} {\\arg \\min } \\frac {1}{\\mathcal {H W C}} \\sum \\left(\\mathcal {Y} ^ {l} - \\hat {\\mathcal {Y}} ^ {l}\\right) ^ {2} \\tag {4}\n$$\n", + "text_format": "latex", + "bbox": [ + 400, + 657, + 785, + 689 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "View-dependent visibility mask. The computation of visibility is inherited from the rendering pipeline utilizing the pre-trained VAE model. Specifically, the decoded mesh vertices $\\hat{\\mathbf{M}}_t$ are rasterized to formulate the barycentric coordinates, which are leveraged to interpolate the UV coordinates $\\varPhi$ across the entire avatar. Following this, the visibility is determined by sampling the view-specific pixels from the 2-D texture map based on the interpolated UV indexes, $\\varPhi = [\\phi_x,\\phi_y]$ $(0\\leq \\phi_{x},\\phi_{y}\\leq 1)$ .", + "bbox": [ + 212, + 702, + 787, + 809 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "The value of $\\phi_x$ and $\\phi_y$ represents the normalized coordinates of the 2-D texture map with the shape of $\\mathcal{H} \\times \\mathcal{W}$ . Different from the default pipeline, the", + "bbox": [ + 212, + 809, + 787, + 840 + ], + "page_idx": 7 + }, + { + "type": "page_number", + "text": "8", + "bbox": [ + 217, + 114, + 228, + 126 + ], + "page_idx": 7 + }, + { + "type": "header", + "text": "J. Meng et al.", + "bbox": [ + 271, + 114, + 364, + 128 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/5f091620bbd48526a320aaa752ccb47cd4a8c80de29fcecb15789ba85aea0f78.jpg", + "image_caption": [ + "Fig. 4: Overall PTQ process of POCA based on layer-wise calibration (Layer Calib). POCA fully quantizes all the layers in the VAE decoder, where the proposed view-dependent $+$ SWD filter is applied to the decoder with transpose convolution layers." + ], + "image_footnote": [], + "bbox": [ + 246, + 145, + 754, + 348 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "We proposed POCA algorithm generates the binary \"visibility mask\" $\\Gamma$ to select the pixels in the 2-D texture map $\\hat{\\mathbf{T}}_t$ . Formally, we have:", + "bbox": [ + 212, + 428, + 782, + 458 + ], + "page_idx": 8 + }, + { + "type": "equation", + "text": "\n$$\n\\Phi = \\text {I n t e r p o l a t e} (\\text {R a s t e r i z e} (\\mathbf {M} _ {t}, \\dots), u v) \\tag {5}\n$$\n", + "text_format": "latex", + "bbox": [ + 321, + 465, + 785, + 483 + ], + "page_idx": 8 + }, + { + "type": "equation", + "text": "\n$$\n\\Gamma = \\mathbb {1} \\cdot \\left\\{\\left(\\gamma_ {x}, \\gamma_ {y}\\right) \\mid \\gamma_ {x} \\in \\left\\{\\phi_ {x} \\times \\mathcal {H} \\right\\}, \\gamma_ {y} \\in \\left\\{\\phi_ {y} \\times \\mathcal {W} \\right\\} \\right\\} \\tag {6}\n$$\n", + "text_format": "latex", + "bbox": [ + 323, + 484, + 784, + 501 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "Where $\\mathbb{1}$ represents the indicator function that signifies whether the pixel at $(\\gamma_x,\\gamma_y)$ should be active at the given visibility. $\\{\\phi_x\\times \\mathcal{H}\\}$ and $\\{\\phi_y\\times \\mathcal{W}\\}$ are the visibility coordinates (matrix indexes) scaled based on the interpolated UV coordinates $\\varPhi$ with the dimensionality of the texture map.", + "bbox": [ + 214, + 507, + 782, + 566 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "The visibility masks are generated during the data-fetching process. Since the model is already pre-trained, the visibility masks $\\varGamma$ are accessible with each full-precision texture $\\hat{\\mathbf{T}}_t$ and triangle index $uv$ at the output layer of the texture decoder. The final layer $\\varGamma$ ( $1024 \\times 1024$ ) will be max-pooled to formulate the visibility mask for the earlier layer. For each layer of the texture decoder, POCA performs PTQ with the visible pixel only:", + "bbox": [ + 212, + 568, + 784, + 659 + ], + "page_idx": 8 + }, + { + "type": "equation", + "text": "\n$$\n\\underset {\\alpha^ {l}, S ^ {l}} {\\arg \\min } \\frac {1}{\\mathcal {H W C}} \\sum \\left(\\mathcal {Y} _ {\\mathbf {T}} ^ {l} \\odot \\Gamma^ {l} - \\hat {\\mathcal {Y}} _ {\\mathbf {T}} ^ {l} \\odot \\Gamma^ {l}\\right) ^ {2} \\tag {7}\n$$\n", + "text_format": "latex", + "bbox": [ + 354, + 664, + 785, + 698 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "Where $\\odot$ represents the element-wise product between tensors. Subsequently, the learnable quantization parameters $\\alpha^l$ and $S^l$ are trained based on the gradient that resulted from the view-dependent visibility $\\Gamma$ . As a result, POCA optimizes the quantization decision based on the most relevant information of the training sample, instead of being diversified by the widely dispersed texture space with long-tailed and dynamic distribution. Figure 4 depicts the proposed view-dependent PTQ. Although the quantizer is calibrated based on the visible pixels, POCA only runs PTQ once, and the resultant model will be deployed and rendered for all views.", + "bbox": [ + 212, + 704, + 787, + 839 + ], + "page_idx": 8 + }, + { + "type": "header", + "text": "POCA", + "bbox": [ + 683, + 114, + 730, + 126 + ], + "page_idx": 8 + }, + { + "type": "page_number", + "text": "9", + "bbox": [ + 774, + 116, + 785, + 126 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "4.2 PTQ with Shape-wise Dynamic Filtering", + "text_level": 1, + "bbox": [ + 215, + 146, + 599, + 162 + ], + "page_idx": 9 + }, + { + "type": "text", + "text": "Besides the PTQ with the view-dependent visibility masking, we further suppress the impact of the dynamically long-tailed distribution by eliminating the pixels with a high channel-wise dynamic range, or \"Shape-Wise Dynamics\" (SWD). Given the full-precision output $\\mathcal{V}_{\\mathbf{T}}^{l}$ with the shape of $\\mathcal{C} \\times \\mathcal{H} \\times \\mathcal{W}$ , the SWD is characterized by the standard deviation score of each $\\mathcal{C} \\times 1 \\times 1$ vector, which formulates the overall texture SWD map with the shape of $\\mathcal{H} \\times \\mathcal{W}$ .", + "bbox": [ + 212, + 172, + 787, + 263 + ], + "page_idx": 9 + }, + { + "type": "text", + "text": "Subsequently, POCA generates the SWD mask $\\Theta$ by filtering out the pixels in the SWD map with the highest tailedness. The sparsity of the SWD mask is controlled by a tunable hyperparameter $\\tau$ . E.g., $\\tau = 0.7$ means masking out the top $70\\%$ pixels with the highest tailedness (standard deviation). The resultant SWD mask $\\Theta$ will be multiplied with the view-dependent mask $\\Gamma$ and generate the final learning objective of the layer-wise post-training quantization of POCA:", + "bbox": [ + 212, + 263, + 787, + 354 + ], + "page_idx": 9 + }, + { + "type": "equation", + "text": "\n$$\n\\underset {\\alpha^ {l}, S ^ {l}} {\\arg \\min } \\frac {1}{\\mathcal {H W C}} \\sum \\left(\\mathcal {Y} _ {\\mathbf {T}} ^ {l} \\odot \\Gamma^ {l} \\odot \\Theta^ {l} - \\hat {\\mathcal {Y}} _ {\\mathbf {T}} ^ {l} \\odot \\Gamma^ {l} \\odot \\Theta^ {l}\\right) ^ {2} \\tag {8}\n$$\n", + "text_format": "latex", + "bbox": [ + 321, + 367, + 785, + 398 + ], + "page_idx": 9 + }, + { + "type": "text", + "text": "As shown in Eq. (8), POCA optimizes the quantizers by learning the most relevant information (view-dependent) and least diversified pixels (SWD masking) of the avatar, which suppresses the drawbacks of the long-tailedness and high dynamics of the avatar.", + "bbox": [ + 212, + 411, + 785, + 470 + ], + "page_idx": 9 + }, + { + "type": "text", + "text": "The masking strategy of POCA introduces a high degree of sparsity in layer-wise PTQ, controlled by visibility and the SWD threshold $\\tau$ . Our experimental results in Section 5 show that the optimal performance is achieved with $\\tau = 0.9$ . As a result, POCA calibrates the low-precision layer based on highly sparse data $[\\mathcal{X}^l,\\mathcal{Y}_{\\mathbf{T}}^l]$ with $\\geq 90\\%$ sparsity. In other words, POCA only requires $\\leq 10\\%$ of the activation for layer-wise PTQ, showing low memory overhead for the efficient PTQ calibration process.", + "bbox": [ + 212, + 472, + 787, + 579 + ], + "page_idx": 9 + }, + { + "type": "text", + "text": "4.3 Rationale behind POCA", + "text_level": 1, + "bbox": [ + 215, + 601, + 465, + 616 + ], + "page_idx": 9 + }, + { + "type": "text", + "text": "POCA alleviates the impact of rounding instability. With the proposed view-aware PTQ and SWD masking, POCA aims to optimize the layerwise quantizers with the most relevant and quantization-friendly features. The POCA-trained activation quantizer (LSQ) has a generally smaller scaling factor $S^l$ compared to the PTQ without POCA, as shown in Figure 5. The reduced scaling factors formulate the low-precision grid with small intervals (scaling factor) between adjacent levels, alleviating the numerical differences between adjacent levels in the dequantized high-precision domain. As a result, the impact of the aforementioned rounding instability is minimized and the temporal noises become invisible, which proves the aforementioned hypothesis in Section 3.2.", + "bbox": [ + 212, + 627, + 787, + 779 + ], + "page_idx": 9 + }, + { + "type": "text", + "text": "Sparse masking $\\neq$ biased learning. Although two types of sparse masking are introduced into PTQ, the POCA-trained quantizers will not lead to a biased representation in the quantized model. As shown in Figure 4, the sparsified calibration mask $(\\varGamma^l\\odot\\varTheta^l)$ contains sparse but uniformly sampled information", + "bbox": [ + 212, + 780, + 787, + 840 + ], + "page_idx": 9 + }, + { + "type": "page_number", + "text": "10", + "bbox": [ + 217, + 114, + 235, + 126 + ], + "page_idx": 9 + }, + { + "type": "header", + "text": "J. Meng et al.", + "bbox": [ + 271, + 114, + 364, + 128 + ], + "page_idx": 9 + }, + { + "type": "image", + "img_path": "images/3cc20d1972f613dc437500148f413b5b2e4e3394fd117e438146aaeb86142734.jpg", + "image_caption": [ + "Fig. 5: Scaling factor $S^l$ of the texture decoder trained by POCA." + ], + "image_footnote": [], + "bbox": [ + 313, + 143, + 687, + 271 + ], + "page_idx": 10 + }, + { + "type": "text", + "text": "across all the regions of the avatar texture. As a result, the learnable quantizers are uniformly trained with all the relevant facial information.", + "bbox": [ + 212, + 306, + 782, + 334 + ], + "page_idx": 10 + }, + { + "type": "text", + "text": "Highly-correlated data and high fidelity sparse learning. Unlike classification tasks, the training samples of the VAE model are highly correlated with respect to the single identity. As a result, the sparsified information of each sample can be compensated from other training data with different viewing perspectives, leading to sparse but robust PTQ with the quantized decoder.", + "bbox": [ + 212, + 335, + 784, + 411 + ], + "page_idx": 10 + }, + { + "type": "text", + "text": "5 Experimental Results", + "text_level": 1, + "bbox": [ + 215, + 431, + 460, + 450 + ], + "page_idx": 10 + }, + { + "type": "text", + "text": "We evaluated the proposed POCA algorithm based on Multiface [29], a public avatar dataset captured by the multi-camera system. POCA is applied to eight avatars among three viewing directions with all 65 facial expressions of the dataset. We want to highlight that the eight avatars comprehensively cover different genders, age groups, hairstyles, skin types, and facial hairs.", + "bbox": [ + 212, + 460, + 784, + 537 + ], + "page_idx": 10 + }, + { + "type": "text", + "text": "For each identity, we use the corresponding pre-trained deep appearance VAE model as the starting point of POCA. Following the standard PTQ strategy, we quantize all layers of the decoders (texture and mesh) iteratively, where each layer is calibrated by 100 epochs with the sampled data. POCA only acquires 512 training samples from each identity, which is $\\leq 5\\%$ samples of the dataset. For each layer, the quantizers are optimized by Adam with 1e-4 learning rate and eight samples per batch.", + "bbox": [ + 212, + 537, + 785, + 643 + ], + "page_idx": 10 + }, + { + "type": "text", + "text": "Prior Works for Comparison. As the first work in the low-precision avatar decoding, we select three prior quantization works that reported the SoTA INT8 performance with the classification tasks: (1) Adaptive Rounding (AdaRound) [18], (2) learned step size quantization (LSQ) [2], which are also incorporated for hardware deployment [24], and (3) W8A8 naive QAT [20] in PyTorch. These selected prior works achieved superior INT8 performance [2, 18] on the ImageNet-1K dataset, even outperforming the full-precision baseline.", + "bbox": [ + 212, + 643, + 785, + 748 + ], + "page_idx": 10 + }, + { + "type": "text", + "text": "Evaluation Metric. We evaluate the proposed POCA algorithm based on the FovVideoVDP [16] between the rendered avatar and the ground truth video to quantify the temporal noise of the dynamic Codec Avatar. For each avatar identity, we compute the averaged FovVideoVDP score among all the 65 expressions in the Multiface dataset [29]. The performance of POCA is demonstrated with the rendered avatar videos in the supplementary material.", + "bbox": [ + 212, + 750, + 785, + 839 + ], + "page_idx": 10 + }, + { + "type": "header", + "text": "POCA", + "bbox": [ + 683, + 114, + 730, + 126 + ], + "page_idx": 10 + }, + { + "type": "page_number", + "text": "11", + "bbox": [ + 767, + 114, + 784, + 126 + ], + "page_idx": 10 + }, + { + "type": "table", + "img_path": "images/d3782da200063a75a37548e56ea4c2713a8ed3892a3b85507a4b2de9b21a6c57.jpg", + "table_caption": [ + "Table 1: FovVideoVDP score [16] between the rendered avatar and the ground truth based on eight avatars across all the expressions from three viewing perspectives. The proposed POCA algorithm achieves the highest quality compared to the SoTA PTQ and QAT methods with both 8-bit and 6-bit precision." + ], + "table_footnote": [], + "table_body": "
Avatar IDPrecision26438142183941537202126453108870559667444329145896795937
Male, GoateeBlack, FemaleMohawk, MaleAsian, FemaleOlder WomanMale, Mid HairBlonde GirlMale, Short Hair
FrontTopFrontTopFrontTopFrontTopFrontTopFrontTopFrontTopFrontTop
Baseline [29]FP326.214.945.306.544.995.456.144.875.406.325.025.596.054.955.136.02
SoTA PTQ [2,18]W8A85.814.434.875.914.405.005.804.525.195.874.565.105.434.444.685.52
W8A8 QAT [20]5.924.554.996.024.575.125.884.565.235.974.665.165.574.504.795.60
POCA6.194.935.266.314.805.386.094.835.336.214.955.425.954.825.085.99
SoTA PTQ [2,18]W6A64.503.664.064.963.814.124.554.104.475.223.914.454.293.733.964.25
W6A6 QAT5.083.754.395.544.014.234.944.194.495.364.074.664.363.854.094.46
POCA6.064.555.026.294.425.136.014.615.166.024.665.215.454.494.995.66
", + "bbox": [ + 215, + 212, + 785, + 319 + ], + "page_idx": 11 + }, + { + "type": "image", + "img_path": "images/dfc31f7dfbe95a7a21ecbb8634d9eeb7138c6e23f09312d14c0d85952b626fe9.jpg", + "image_caption": [ + "Yellow region $=$ Mismatched pixel motion", + "FovVideoVDP [14] between the rendered avatar and ground truth (Max = 10.0)", + "Fig. 6: Rendered avatars and temporal pixel differences between two consecutive frames of the avatar video. The lightened yellow region of POCA signifies the removed temporal noise as the full precision baseline." + ], + "image_footnote": [], + "bbox": [ + 222, + 330, + 782, + 489 + ], + "page_idx": 11 + }, + { + "type": "text", + "text": "5.1 Temporal Noise Reduction with High Fidelity", + "text_level": 1, + "bbox": [ + 214, + 582, + 640, + 598 + ], + "page_idx": 11 + }, + { + "type": "text", + "text": "We evaluate the proposed POCA algorithm with eight avatars with different genders, hairstyles, and skin colors. Table 1 summarizes the FovVideoVDP score [16] of the proposed POCA algorithm with 8-bit and 6-bit avatar decoders. The SoTA INT8 PTQ quantization algorithm failed to preserve the avatar quality as the baseline full precision model, exhibiting the degraded FovVideoVDP score. Even with the expensive quantization-aware training (QAT), the quality of the rendered avatar cannot be restored.", + "bbox": [ + 212, + 612, + 787, + 717 + ], + "page_idx": 11 + }, + { + "type": "text", + "text": "Compared to the previous methods with 8-bit precision, POCA eliminates the temporal noises in the rendered avatar. POCA successfully compresses the decoder model while achieving on-par quality as the full precision decoder baseline. Figure 6 shows the high avatar quality and restored FovVideoVDP score (heat map) of the proposed POCA algorithm (closer to the FP32 baseline = better quantization). Regarding the efficiency of PTQ, POCA only requires $\\leq 10\\%$ of the activations for layer-wise PTQ, showing low memory overhead for the efficient PTQ calibration process.", + "bbox": [ + 212, + 719, + 787, + 840 + ], + "page_idx": 11 + }, + { + "type": "page_number", + "text": "12", + "bbox": [ + 217, + 114, + 235, + 126 + ], + "page_idx": 11 + }, + { + "type": "header", + "text": "J. Meng et al.", + "bbox": [ + 271, + 114, + 364, + 128 + ], + "page_idx": 11 + }, + { + "type": "image", + "img_path": "images/07e345d143445ab9dc251b3dfb8a8c907ed40d885528d5bf819d149708673aa7.jpg", + "image_caption": [ + "SSIM=0.741 \nFovVideo $\\mathrm{DVP}=6.213$" + ], + "image_footnote": [], + "bbox": [ + 274, + 143, + 361, + 255 + ], + "page_idx": 12 + }, + { + "type": "image", + "img_path": "images/4f149d2510c6863f4b7bf29d333575109256d8111bcaca2469a9a0109e6e185d.jpg", + "image_caption": [ + "SSIM=0.741 \nFovVideoDVP=6.119" + ], + "image_footnote": [], + "bbox": [ + 362, + 143, + 450, + 255 + ], + "page_idx": 12 + }, + { + "type": "image", + "img_path": "images/b09af302b5c3cae5805f4af93690fbc0b60e4a438927578e8106ae8f58a30cd0.jpg", + "image_caption": [ + "SSIM=0.741 \nFovVideoVDP=1.685", + "Fig. 7: The impact of the SWD threshold. Extreme filtering leads to distorted and missing texture in the resultant avatar." + ], + "image_footnote": [], + "bbox": [ + 452, + 143, + 544, + 255 + ], + "page_idx": 12 + }, + { + "type": "image", + "img_path": "images/67276d87212373da8dc64a0adf798dfbc307d6ef36804bdbfda6b1889cec60bf.jpg", + "image_caption": [ + "SSIM=0.740 \nFovVideoVDP=6.121" + ], + "image_footnote": [], + "bbox": [ + 545, + 143, + 633, + 255 + ], + "page_idx": 12 + }, + { + "type": "image", + "img_path": "images/be98ce8fa9a901c66f1f0e6b976be335103a3f511e8e23d0f714b5a20bd6fab4.jpg", + "image_caption": [ + "SSIM=0.739 \nFoxVideoVDP=6.106" + ], + "image_footnote": [], + "bbox": [ + 635, + 143, + 728, + 255 + ], + "page_idx": 12 + }, + { + "type": "table", + "img_path": "images/b61a4cc34dd6833349360b27eed229d29c2e0699529ee9d2b7b12c21cf361a8c.jpg", + "table_caption": [ + "Table 2: Training time comparison between POCA and QAT." + ], + "table_footnote": [], + "table_body": "
MethodQuantizationMethodW/ATimeGPUDecoder Size (MB)
FP32 Baseline-32/32---32.78
Native W8A8 [20]QAT8/8>10 hrs2 × A60008.19 (↓ 4×)
POCA (This work)PTQ8/8~13 min1 × A60008.19 (↓ 4×)
POCA (This work)PTQ6/6~13 min1 × A60006.15 (↓ 5.3×)
", + "bbox": [ + 269, + 333, + 727, + 407 + ], + "page_idx": 12 + }, + { + "type": "text", + "text": "POCA generates clean avatars without introducing any graphic filtering techniques [31]. From the perspective of quantization, POCA keeps the vanilla quantization strategy in Eq. (1) in both training and inference, compressing the model without introducing distribution manipulations [28] on the additional scaling [30] and bias [12]. Compared to traditional QAT, POCA quantizes the entire model in only 13 minutes (Table 2). The fast deployment workflow with the actual VR device (e.g., Quest Pro) can further accelerate the process of \"user-to-meta verse\".", + "bbox": [ + 212, + 436, + 787, + 556 + ], + "page_idx": 12 + }, + { + "type": "text", + "text": "Future Applicability: POCA is not limited to the deep appearance model [14]. The proposed POCA algorithm is applicable to all mesh-based 3D reconstruction tasks with VAE models. For example, the proposed view-dependent visibility mask and shape-wise filtering can be easily transferred to the recent Pixel Codec Avatar (PiCA)[15], which has the similar autoencoder-based architecture.", + "bbox": [ + 212, + 559, + 805, + 635 + ], + "page_idx": 12 + }, + { + "type": "text", + "text": "5.2 Impact of the SWD Filtering Threshold", + "text_level": 1, + "bbox": [ + 215, + 661, + 589, + 678 + ], + "page_idx": 12 + }, + { + "type": "text", + "text": "POCA introduced the tunable parameter $\\tau$ in Eq. (8) to control the amount of activation pixels for layer-wise PTQ. Figure 7 demonstrates the impact of the SWD threshold $\\tau$ with respect to the avatar quality (SSIM). POCA achieves the best tradeoff between filter intensity and avatar quality with $\\tau^{*} = 0.9$ . Quantizing the decoder with $\\tau > \\tau^{*}$ leads to missing texture pixels and distorted rendering with no quality improvement. The selected SWD threshold exhibits low sensitivity between different avatars and viewing perspectives, as demonstrated by the high-quality rendering in Figure 6.", + "bbox": [ + 212, + 691, + 787, + 813 + ], + "page_idx": 12 + }, + { + "type": "header", + "text": "POCA", + "bbox": [ + 683, + 114, + 730, + 126 + ], + "page_idx": 12 + }, + { + "type": "page_number", + "text": "13", + "bbox": [ + 767, + 114, + 785, + 126 + ], + "page_idx": 12 + }, + { + "type": "page_footnote", + "text": "3 PiCA results are available at https://mengjian0502.github.io/poca.github.io/", + "bbox": [ + 217, + 824, + 784, + 839 + ], + "page_idx": 12 + }, + { + "type": "text", + "text": "5.3 Applicability to Other Quality Enhancement Tasks", + "text_level": 1, + "bbox": [ + 215, + 146, + 679, + 161 + ], + "page_idx": 13 + }, + { + "type": "text", + "text": "We extend the experiment by evaluating the proposed method on super-resolution tasks. The proposed \"shape-wise-filtering\" (SWD) exhibits strong performance in image quality enhancement [25] with the low-precision SRResNet model [8].", + "bbox": [ + 212, + 167, + 787, + 214 + ], + "page_idx": 13 + }, + { + "type": "table", + "img_path": "images/6ac2ec29a1f4c966cb6daa522cb79e6ff95f1a3531b883bbc567e46cb14f5737.jpg", + "table_caption": [ + "Table 3: SoTA image quality (PSNR/SSIM) achieved by the proposed POCA-SWD algorithm with 8-bit and 6-bit model precision, outperforms the recent SR-PTQ [25]." + ], + "table_footnote": [], + "table_body": "
MethodPrecisionSparsitySet5 (× 4)Set14 (× 4)BSD100 (× 4)Urban100 (× 4)
SRResNet Baseline [25]FP32(Baseline)N/A32.234 / 0.89628.656 / 0.78427.63 / 0.73826.229 / 0.791
SRResNet [8]32.05 / 0.90128.49 / 0.81827.58 / 0.762Not Reported
SRResNet (This work)31.927 / 0.90128.588 / 0.79927.628 / 0.75626.223 / 0.800
AdaRound [18] + LSQ [2]W8A8Dense31.870 / 0.89928.550 / 0.79827.533 / 0.75426.078 / 0.797
SR-PTQ [25]Dense32.207 / 0.89528.619 / 0.78327.618 / 0.73826.191 / 0.790
POCA80%31.898 / 0.90028.566 / 0.79827.620 / 0.75626.195 / 0.799
Quality Drop [2,18]W8A80.057 / 0.0020.038 / 0.0010.095 / 0.0020.145 / 0.003
Quality Drop [25]0.027 / 0.0010.037 / 0.0010.012 / 0.0000.038 / 0.001
Quality Drop (POCA)0.029 / 0.0010.022 / 0.0010.008 / 0.0000.028 / 0.001
AdaRound [18] + LSQ [2]W6A6Dense31.765 / 0.89728.449 / 0.79527.470 / 0.75225.988 / 0.794
SR-PTQ [25]Dense32.089 / 0.89228.504 / 0.77927.561 / 0.73326.011 / 0.783
POCA80%31.789 / 0.89928.499 / 0.79627.565 / 0.75326.110 / 0.795
Quality Drop [2,18]W6A60.162 / 0.0040.139 / 0.0030.158 / 0.0040.235 / 0.006
Quality Drop [25]0.145 / 0.0040.152 / 0.0050.069 / 0.0050.218 / 0.008
Quality Drop (POCA)0.138 / 0.0020.089 / 0.0040.061 / 0.0030.113 / 0.005
", + "bbox": [ + 228, + 253, + 764, + 438 + ], + "page_idx": 13 + }, + { + "type": "text", + "text": "With 8-bit and 6-bit precision, the proposed method outperforms [25] in BSD100 and Urban100 datasets, as shown in Table 3. For a fair comparison, we listed all the baseline full-precision models. Our method outperforms [25] on all datasets with the new SoTA image quality. Compared to [18] and [2], the improved image quality implies the effectiveness of the proposed SWD. More importantly, our method enables the sparse PTQ by using only $20\\%$ pixels of the selected samples, whereas all prior works [2,18,25] require dense PTQ.", + "bbox": [ + 212, + 449, + 787, + 555 + ], + "page_idx": 13 + }, + { + "type": "text", + "text": "6 Conclusion and Discussion", + "text_level": 1, + "bbox": [ + 215, + 574, + 508, + 589 + ], + "page_idx": 13 + }, + { + "type": "text", + "text": "In this paper, we propose a novel post-training quantization algorithm designed for high-quality Codec Avatar rendering. As the first work in this area, we discover the challenge of temporal noise with low-precision avatar decoding. Through thorough investigation, we reveal the root cause of the temporal noise that originated from the high dynamics and long-tailedness of the texture distribution. To address that, we propose POCA, which consists of view-aware PTQ and shape-wise filtering (SWD). Compared to the previous SoTA methods, POCA eliminates the temporal noises even with 6-bit precision. Furthermore, the proposed SWD algorithm is also applicable to super-resolution tasks, achieving the new SoTA image quality with highly efficient sparse calibration.", + "bbox": [ + 212, + 602, + 785, + 753 + ], + "page_idx": 13 + }, + { + "type": "text", + "text": "Limitations. The proposed POCA algorithm mainly focuses on the VAE-based Codec Avatar generation, which has been proven as a successful strategy with the actual commercial-level VR product deployment. Recently, 3D Gaussian-based avatar rendering also emerged as a feasible method for generative Avatars, while low-precision quantization of such models will remain as future work.", + "bbox": [ + 212, + 763, + 785, + 839 + ], + "page_idx": 13 + }, + { + "type": "page_number", + "text": "14", + "bbox": [ + 217, + 114, + 235, + 126 + ], + "page_idx": 13 + }, + { + "type": "header", + "text": "J. Meng et al.", + "bbox": [ + 271, + 114, + 364, + 128 + ], + "page_idx": 13 + }, + { + "type": "text", + "text": "References", + "text_level": 1, + "bbox": [ + 215, + 143, + 323, + 159 + ], + "page_idx": 14 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "1. Choi, J., Venkataramani, S., Srinivasan, V.V., Gopalakrishnan, K., Wang, Z., Chuang, P.: Accurate and efficient 2-bit quantized neural networks. Proceedings of Machine Learning and Systems 1, 348-359 (2019) 5", + "2. Esser, S.K., McKinstry, J.L., Bablani, D., Appuswamy, R., Modha, D.S.: Learned step size quantization. In: International Conference on Learning Representations (ICLR) (2020) 2, 3, 5, 6, 8, 11, 12, 14", + "3. Fu, Y., Li, Y., Li, C., Saragih, J., Zhang, P., Dai, X., Lin, Y.C.: Auto-card: Efficient and robust codec avatar driving for real-time mobile telepresence. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR) (2023) 4", + "4. Gorbachev, Y., Fedorov, M., Slavutin, I., Tugarev, A., Fatekhov, M., Tarkan, Y.: Openvino deep learning workbench: Comprehensive analysis and tuning of neural networks inference. In: Proceedings of the IEEE/CVF International Conference on Computer Vision Workshops (CVPR-W) (2019) 4", + "5. Han, Y., Yang, S., Wang, W., Liu, J.: From design draft to real attire: Unaligned fashion image translation. In: Proceedings of the 28th ACM International Conference on Multimedia. p. 1533-1541 (2020) 4", + "6. Kingma, D.P., Welling, M.: Auto-encoding variational bayes. arXiv preprint arXiv:1312.6114 (2013) 4", + "7. Le, H., Zhang, L., Said, A., Sautiere, G., Yang, Y., Shrestha, P., Yin, F., Pourreza, R., Wiggers, A.: Mobilecodec: neural inter-frame video compression on mobile devices. In: Proceedings of the 13th ACM Multimedia Systems Conference. pp. 324-330 (2022) 5", + "8. Ledig, C., Theis, L., Huszár, F., Caballero, J., Cunningham, A., Acosta, A., Aitken, A., Tejani, A., Totz, J., Wang, Z., et al.: Photo-realistic single image superresolution using a generative adversarial network. In: Proceedings of the IEEE conference on computer vision and pattern recognition. pp. 4681-4690 (2017) 3, 14", + "9. Lewis, J.P., Anjyo, K., Rhee, T., Zhang, M., Pighin, F.H., Deng, Z.: Practice and theory of blendshape facial models. Eurographics (State of the Art Reports) 1(8), 2 (2014) 4", + "0. Li, Y., Dong, X., Wang, W.: Additive powers-of-two quantization: An efficient non-uniform discretization for neural networks. In: International Conference on Learning Representations (ICLR) (2020) 5", + "1. Li, Y., Gong, R., Tan, X., Yang, Y., Hu, P., Zhang, Q., Yu, F., Wang, W., Gu, S.: {BRECQ}: Pushing the limit of post-training quantization by block reconstruction. In: International Conference on Learning Representations (ICLR) (2021) 2, 5", + "2. Liu, Y., Yang, H., Dong, Z., Keutzer, K., Du, L., Zhang, S.: Noisyquant: Noisy bias-enhanced post-training activation quantization for vision transformers. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR) (2023) 3, 5, 6, 13", + "3. Liu, Z., Cheng, K.T., Huang, D., Xing, E.P., Shen, Z.: Nonuniform-to-uniform quantization: Towards accurate quantization via generalized straight-through estimation. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR) (2022) 5", + "4. Lombardi, S., Saragih, J., Simon, T., Sheikh, Y.: Deep appearance models for face rendering. ACM Transactions on Graphics (ToG) (2018) 1, 2, 4, 7, 13" + ], + "bbox": [ + 225, + 179, + 785, + 839 + ], + "page_idx": 14 + }, + { + "type": "header", + "text": "POCA", + "bbox": [ + 683, + 114, + 730, + 126 + ], + "page_idx": 14 + }, + { + "type": "page_number", + "text": "15", + "bbox": [ + 767, + 114, + 785, + 126 + ], + "page_idx": 14 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "15. Ma, S., Simon, T., Saragih, J., Wang, D., Li, Y., De La Torre, F., Sheikh, Y.: Pixel codec avatars. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR) (2021) 1, 2, 4, 13", + "16. Mantiuk, R.K., Denes, G., Shapiro, A., Kaplanyan, A., Rufo, G., Bachy, R., Lian, T., Patney, A.: FovVideoVDP: A visible difference predictor for wide field-of-view video. In: ACM Transactions on Graphics (TOG). ACM New York, NY, USA (2021) 3, 6, 11, 12", + "17. Mildenhall, B., Srinivasan, P.P., Tancik, M., Barron, J.T., Ramamoorthi, R., Ng, R.: NeRF: Representing scenes as neural radiance fields for view synthesis. Communications of the ACM 65(1), 99-106 (2021) 5", + "18. Nagel, M., Amjad, R.A., Van Baalen, M., Louizos, C., Blankevoort, T.: Up or down? adaptive rounding for post-training quantization. In: International Conference on Machine Learning (ICML) (2020) 2, 3, 5, 6, 8, 11, 12, 14", + "19. Park, E., Yoo, S.: Profit: A novel training method for sub-4-bit mobilenet models. In: European Conference on Computer Vision (ECCV) (2020) 6", + "20. Paszke, A., Gross, S., Massa, F., Lerer, A., Bradbury, J., Chanan, G., Killeen, T., Lin, Z., Gimelshein, N., Antiga, L., Desmaison, A., Kopf, A., Yang, E., DeVito, Z., Raison, M., Tejani, A., Chilamkurthy, S., Steiner, B., Fang, L., Bai, J., Chintala, S.: Pytorch: An imperative style, high-performance deep learning library. In: Advances in Neural Information Processing Systems (NeurIPS) (2019) 3, 11, 12, 13", + "21. van Rozendaal, T., Singhal, T., Le, H., Sautiere, G., Said, A., Buska, K., Raha, A., Kalatzis, D., Mehta, H., Mayer, F., et al.: Mobilenvc: Real-time 1080p neural video compression on a mobile device. In: Proceedings of the IEEE/CVF Winter Conference on Applications of Computer Vision. pp. 4323-4333 (2024) 5", + "22. Saito, S., Schwartz, G., Simon, T., Li, J., Nam, G.: Relightable gaussian codec avatars. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 130-141 (2024) 2", + "23. Schwartz, G., Wei, S.E., Wang, T.L., Lombardi, S., Simon, T., Saragih, J., Sheikh, Y.: The eyes have it: An integrated eye and face model for photorealistic facial animation. ACM Transactions on Graphics (TOG) (2020) 1", + "24. Siddegowda, S., Fournarakis, M., Nagel, M., Blankevoort, T., Patel, C., Khobare, A.: Neural network quantization with AI model efficiency toolkit (AIMET). arXiv preprint arXiv:2201.08442 (2022) 2, 3, 4, 11", + "25. Tu, Z., Hu, J., Chen, H., Wang, Y.: Toward accurate post-training quantization for image super resolution. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR) (2023) 2, 3, 6, 14", + "26. Wang, Z., Wu, Z., Lu, J., Zhou, J.: Bidet: An efficient binarized object detector. In: Proceedings of the IEEE/CVF conference on computer vision and pattern recognition (CVPR) (2020) 2", + "27. Wei, S.E., Saragih, J., Simon, T., Harley, A.W., Lombardi, S., Perdoch, M., Hypes, A., Wang, D., Badino, H., Sheikh, Y.: VR facial animation via multiview image translation. ACM Transactions on Graphics (TOG) (2019) 1", + "28. Wei, X., Gong, R., Li, Y., Liu, X., Yu, F.: QDrop: Randomly dropping quantization for extremely low-bit post-training quantization. In: International Conference on Learning Representations (ICLR) (2022) 2, 5, 13", + "29. Wuu, C.h., Zheng, N., Ardisson, S., Bali, R., Belko, D., Brockmeyer, E., Evans, L., Godisart, T., Ha, H., Huang, X., Hypes, A., Koska, T., Krenn, S., Lombardi, S., Luo, X., McPhail, K., Millerschoen, L., Perdoch, M., Pitts, M., Richard, A., Saragih, J., Saragih, J., Shiratori, T., Simon, T., Stewart, M., Trimble, A., Weng, X., Whitewolf, D., Wu, C., Yu, S.I., Sheikh, Y.: Multiface: A dataset for neural" + ], + "bbox": [ + 215, + 146, + 785, + 839 + ], + "page_idx": 15 + }, + { + "type": "page_number", + "text": "16", + "bbox": [ + 217, + 114, + 235, + 126 + ], + "page_idx": 15 + }, + { + "type": "header", + "text": "J. Meng et al.", + "bbox": [ + 271, + 114, + 364, + 128 + ], + "page_idx": 15 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "face rendering. In: arXiv (2022). https://doi.org/10.48550/ARXIV.2207.11243, https://arxiv.org/abs/2207.11243 1, 11, 12", + "30. Xiao, G., Lin, J., Seznek, M., Wu, H., Demouth, J., Han, S.: Smoothquant: Accurate and efficient post-training quantization for large language models. In: International Conference on Machine Learning (ICML) (2023) 2, 5, 13", + "31. Yin, H., Gong, Y., Qiu, G.: Side window filtering. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR) (2019) 13", + "32. Zhang, Y., Zhang, Z., Lew, L.: Pokebnn: A binary pursuit of lightweight accuracy. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR) (2022) 4" + ], + "bbox": [ + 215, + 146, + 785, + 285 + ], + "page_idx": 16 + }, + { + "type": "header", + "text": "POCA", + "bbox": [ + 683, + 114, + 730, + 126 + ], + "page_idx": 16 + }, + { + "type": "page_number", + "text": "17", + "bbox": [ + 767, + 114, + 785, + 126 + ], + "page_idx": 16 + } +] \ No newline at end of file diff --git a/2024/POCA_ Post-training Quantization with Temporal Alignment for Codec Avatars/f5688d31-6662-4aa8-b308-c286c399e498_model.json b/2024/POCA_ Post-training Quantization with Temporal Alignment for Codec Avatars/f5688d31-6662-4aa8-b308-c286c399e498_model.json new file mode 100644 index 0000000000000000000000000000000000000000..8f0b64ecac506141933cbaeba1ed663d82adfc29 --- /dev/null +++ b/2024/POCA_ Post-training Quantization with Temporal Alignment for Codec Avatars/f5688d31-6662-4aa8-b308-c286c399e498_model.json @@ -0,0 +1,2203 @@ +[ + [ + { + "type": "title", + "bbox": [ + 0.275, + 0.141, + 0.73, + 0.187 + ], + "angle": 0, + "content": "POCA: Post-training Quantization with Temporal Alignment for Codec Avatars" + }, + { + "type": "text", + "bbox": [ + 0.225, + 0.212, + 0.78, + 0.244 + ], + "angle": 0, + "content": "Jian Meng\\(^{1,2}\\), Yuecheng Li\\(^{1}\\), Chenghui (Leo) Li\\(^{1}\\), Syed Shakib Sarwar\\(^{1}\\), Dilin Wang\\(^{1}\\), and Jae-sun Seo\\(^{2}\\)" + }, + { + "type": "text", + "bbox": [ + 0.35, + 0.255, + 0.652, + 0.269 + ], + "angle": 0, + "content": "1 Meta Codec Avatars Lab, Pittsburgh, USA" + }, + { + "type": "text", + "bbox": [ + 0.377, + 0.269, + 0.625, + 0.283 + ], + "angle": 0, + "content": "2 Cornell Tech, New York City, USA" + }, + { + "type": "text", + "bbox": [ + 0.262, + 0.318, + 0.744, + 0.665 + ], + "angle": 0, + "content": "Abstract. Real-time decoding generates high-quality assets for rendering photorealistic Codec Avatars for immersive social telepresence with AR/VR. However, high-quality avatar decoding incurs expensive computation and memory consumption, which necessitates the design of a decoder compression algorithm (e.g., quantization). Although quantization has been widely studied, the quantization of avatar decoders is an urgent yet under-explored need. Furthermore, the requirement of fast \"User-Avatar\" deployment prioritizes the post-training quantization (PTQ) over the time-consuming quantization-aware training (QAT). As the first work in this area, we reveal the sensitivity of the avatar decoding quality under low precision. In particular, the state-of-the-art (SoTA) QAT and PTQ algorithms introduce massive amounts of temporal noise to the rendered avatars, even with the well-established 8-bit precision. To resolve these issues, a novel PTQ algorithm is proposed for quantizing the avatar decoder with low-precision weights and activation (8-bit and 6-bit), without introducing temporal noise to the rendered avatar. Furthermore, the proposed method only needs \\(10\\%\\) of the activations of each layer to calibrate quantization parameters without any distribution manipulations or extensive boundary search. The proposed method is evaluated on various face avatars with different facial characteristics. The proposed method compresses the decoder model by \\(5.3\\times\\) while recovering the quality on par with the full precision baseline. In addition to the avatar rendering tasks, POCA is also applicable to image resolution enhancement tasks, achieving new SoTA image quality. https://mengjian0502.github.io/poca.github.io/" + }, + { + "type": "title", + "bbox": [ + 0.217, + 0.689, + 0.377, + 0.705 + ], + "angle": 0, + "content": "1 Introduction" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.719, + 0.788, + 0.841 + ], + "angle": 0, + "content": "Photorealistic telepresence [14, 15, 23, 27] has emerged as the cornerstone technology in augmented or virtual reality (AR/VR), enabling virtual interactions across different devices. Specifically, the ultimate goal of photorealistic rendering is animating the high-fidelity Codec Avatars with multiple partial facial images captured by the head-mounted cameras (HMC) [29]. The variational autoencoder-based model [14, 15] has been proved and deployed to VR device (e.g., Quest Pro) as a feasible solution. During inference, the encoder model projects the HMC-captured facial motion into the latent space, and the decoder" + } + ], + [ + { + "type": "page_number", + "bbox": [ + 0.218, + 0.116, + 0.23, + 0.127 + ], + "angle": 0, + "content": "2" + }, + { + "type": "header", + "bbox": [ + 0.272, + 0.115, + 0.366, + 0.128 + ], + "angle": 0, + "content": "J. Meng et al." + }, + { + "type": "image", + "bbox": [ + 0.264, + 0.146, + 0.74, + 0.363 + ], + "angle": 0, + "content": null + }, + { + "type": "image_footnote", + "bbox": [ + 0.26, + 0.365, + 0.743, + 0.375 + ], + "angle": 0, + "content": "Noise is visible in the supplementary video: \"supplementary/ID002643814/poca_all_002643814_w8a8_400029.mp4\"" + }, + { + "type": "image_caption", + "bbox": [ + 0.215, + 0.388, + 0.789, + 0.431 + ], + "angle": 0, + "content": "Fig. 1: Decoding and rendering with low-precision avatar decoder. The SoTA 8-bit PTQ algorithm distorts the avatar with massive and random temporal noise, while the proposed method compresses the decoder with the high-quality avatar without noise." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.462, + 0.788, + 0.583 + ], + "angle": 0, + "content": "reconstructs the latent code into view dependent high-quality texture and mesh, which are then rendered into the 3-D Codec AvATAR on screen. In practice, the critical ingredient of photorealistic telepresence is transmitting the encoded latent information from one VR to another and rendering the avatar in real-time. In practice, the VAE-based Codec AvATAR model has been employed for generative Pixel Codec AvATAR [15] or 3D Gaussian-based AvATAR [22]. Naturally, the quality and latency of telepresence mainly depends on the efficiency of the decoder, which impacts the real-time visual quality." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.583, + 0.788, + 0.674 + ], + "angle": 0, + "content": "Despite the success of the deep appearance model [14], the high computational cost and memory consumption of avatar decoding hinder the fast on-device telepresence. Naïvely deploying the full-precision decoder to the headset leads to a low frame rate, causing degraded user experience. On the other hand, relying on cloud-based computation adds additional \"headset-cloud\" communication latency and increases the risk of avatar data breaching." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.674, + 0.789, + 0.841 + ], + "angle": 0, + "content": "The conflict between high-fidelity rendering and the urgent needs of on-device computing necessitates the high-performance compression algorithm designed for the avatar decoder. In particular, quantization [2, 18, 28, 30] compresses the model with a unified data compression ratio on weights and activation, leading to easy hardware deployment [24]. Unfortunately, the SoTA quantization algorithms mainly focus on classification [11, 18], object detection [26], and natural language processing tasks [30]. Recent works have investigated high-resolution static image rendering with the low-precision model [25]. However, the high-quality quantization algorithm for avatar decoding remains unexplored. In fact, our research shows that the current efforts on quantization are highly incompatible to the Avatar decoder with the following challenges:" + } + ], + [ + { + "type": "header", + "bbox": [ + 0.684, + 0.115, + 0.732, + 0.127 + ], + "angle": 0, + "content": "POCA" + }, + { + "type": "page_number", + "bbox": [ + 0.775, + 0.117, + 0.785, + 0.127 + ], + "angle": 0, + "content": "3" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.147, + 0.788, + 0.314 + ], + "angle": 0, + "content": "Massive temporal noises introduced by INT8 quantization. Although the popular 8-bit quantization has been widely investigated with quantization-aware training [20] and post-training quantization [2,18,24], our discovery shows that the well-established INT8 precision introduces a massive amount of temporal noises to the rendered avatar, even with the SoTA methods. As shown in Figure 1, the facial motion generated by the 8-bit decoder is filled with temporally flickering noises between two consecutive frames, which immerses the original facial expressions and degrades the quality of the rendered avatar. Most importantly, the temporal noise is not visible in the static single frame, distorting the avatar in a continuous manner. As the first work in this area, this paper reveals the root cause of the temporal avatar noise, as described in Section 3." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.326, + 0.788, + 0.476 + ], + "angle": 0, + "content": "Difficulty of characterizing temporal noise. Common quantization noises in conventional classification and object detection tasks can be modeled by a standard statistical model [12]. However, the temporal avatar noises introduced by quantized decoder exhibit inconsistent intensities and granularities in different facial regions. As a result, it is almost impossible to model the temporal noises globally without additional recurrent information. Furthermore, applying an additional filtering strategy after quantization introduces additional energy and latency consumption to the AR/VR devices, which deteriorates the efficiency of Codec Avatar telecommunication. Therefore, the noise suppression process should be achieved all at once during post training quantization." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.477, + 0.788, + 0.537 + ], + "angle": 0, + "content": "To resolve the aforementioned problems, we propose POCA, a novel Post Training Quantization (PTQ) algorithm designed for high quality and temporally noise-free Codec Avatar decoding with outstanding simplicity and efficiency. The major contributions of this work are listed as follows:" + }, + { + "type": "text", + "bbox": [ + 0.226, + 0.544, + 0.785, + 0.588 + ], + "angle": 0, + "content": "- As the first work that investigates the low-precision avatar decoding, we reveal the root cause of the temporal noises induced by the temporally dynamic and long-tailed distribution." + }, + { + "type": "text", + "bbox": [ + 0.226, + 0.589, + 0.785, + 0.619 + ], + "angle": 0, + "content": "- POCA eliminates the temporal facial noise with 8-bit and 6-bit precision, outperforming the SoTA method with fully recovered avatar quality [16]." + }, + { + "type": "text", + "bbox": [ + 0.226, + 0.619, + 0.785, + 0.648 + ], + "angle": 0, + "content": "- POCA only requires \\(\\leq 10\\%\\) of layer-wise feature map data and \\(\\leq 5\\%\\) of the total training data for PTQ." + }, + { + "type": "text", + "bbox": [ + 0.226, + 0.648, + 0.785, + 0.707 + ], + "angle": 0, + "content": "- POCA is evaluated on a variety of avatar identities with a full range of viewing directions and dynamic expressions, exhibiting a high degree of robustness and consistent noise-free avatar animation across different genders, age groups, hairstyles, skin colors, and view directions." + }, + { + "type": "text", + "bbox": [ + 0.226, + 0.708, + 0.785, + 0.737 + ], + "angle": 0, + "content": "- POCA is also applicable to image super resolution task [8,25], achieving the SoTA image quality with \\(80\\%\\) calibration sparsity." + }, + { + "type": "list", + "bbox": [ + 0.226, + 0.544, + 0.785, + 0.737 + ], + "angle": 0, + "content": null + }, + { + "type": "title", + "bbox": [ + 0.216, + 0.758, + 0.388, + 0.774 + ], + "angle": 0, + "content": "2 Related Work" + }, + { + "type": "title", + "bbox": [ + 0.216, + 0.788, + 0.377, + 0.802 + ], + "angle": 0, + "content": "2.1 Codec Avatar" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.81, + 0.787, + 0.84 + ], + "angle": 0, + "content": "Photorealistic human face modeling and animation are challenging computer vision and computer graphics tasks in the sense of achieving minimum visual" + } + ], + [ + { + "type": "page_number", + "bbox": [ + 0.218, + 0.116, + 0.23, + 0.127 + ], + "angle": 0, + "content": "4" + }, + { + "type": "header", + "bbox": [ + 0.272, + 0.115, + 0.366, + 0.129 + ], + "angle": 0, + "content": "J. Meng et al." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.147, + 0.788, + 0.312 + ], + "angle": 0, + "content": "distortion for reconstructed 3-D human faces. Early works mainly focus on accurate 3-D reconstruction with blend shape models [9], which are computationally expensive and impossible for mobile on-device telepresence. Empowered by the deep learning models, recent research [14,15] considers human face reconstruction as data-driven generative task with the variational auto-encoder neural network. In particular, the variational encoder [6] is trained for encoding the multi-view HMC images into the latent information \\(\\mathbf{z}\\), which describes the state of the face with rich expression and motion. Subsequently, the decoder network reconstructs the 3-D faces based on the latent code and the given view direction \\(\\mathbf{v}\\), where the assets for rendering 3D face avatar are decoded separately by the dedicated mesh decoder and texture decoder." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.313, + 0.789, + 0.525 + ], + "angle": 0, + "content": "The compact latent code is transmitted between VR devices with low communication bandwidth requirements. However, the desire for real-time, on-device 3-D face avatar animation is hindered by the expensive computation cost and high memory consumption of the decoder. Various efforts have been established to improve the computation efficiency of the VAE model [5, 14]. Pixel Codec Avatar [15] proposes a novel architecture to allow lightweight decoder by creatively leveraging rendering pipeline features, skipping computations for invisible pixels while maintaining the high-fidelity for visible regions. Recent research optimizes the encoder architecture via neural architecture search (NAS) [3], leading to reduced computational costs. Different from such prior works, our work proposes a decoder-dedicated quantization algorithm that enables temporal alignments and noise-free avatar decoding with low-precision model, which further enhances instant and high quality on-device deployment of Codec Avatars for resource-constrained mobile AR/VR headsets." + }, + { + "type": "title", + "bbox": [ + 0.216, + 0.546, + 0.371, + 0.56 + ], + "angle": 0, + "content": "2.2 Quantization" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.57, + 0.788, + 0.722 + ], + "angle": 0, + "content": "Quantization has been widely investigated together with the evolution of deep learning. Starting from the well-established 8-bit precision [4,24] all the way to the aggressive binary neural network [32], the quantized model largely improves the computation and memory efficiency, making the complex DNN models deployable to the resource-constrained edge devices (e.g., AR/VR headset). The core idea of quantization is constructing the accurate matching between the high-precision data (e.g., FP32) and the low-precision representation. Normally, the quantization process can be generalized into the \"quantize-and-dequantize\" workflow connected by the shared scaling factor in between. Mathematically, given the full precision weights \\(\\mathcal{W}^l\\) and input activation \\(\\mathcal{X}^l\\) of a layer \\(l\\), we have:" + }, + { + "type": "equation", + "bbox": [ + 0.338, + 0.732, + 0.786, + 0.766 + ], + "angle": 0, + "content": "\\[\n\\mathcal {W} _ {Q} ^ {l} = \\left\\lfloor \\frac {\\mathcal {W} ^ {l}}{S _ {w} ^ {l}} \\right] \\in \\operatorname {I N T} _ {\\mathbf {n}}, \\quad \\mathcal {X} _ {Q} ^ {l} = \\left\\lfloor \\frac {\\mathcal {X} ^ {l}}{S _ {x} ^ {l}} \\right] \\in \\operatorname {I N T} _ {\\mathbf {n}} \\tag {1}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.769, + 0.788, + 0.815 + ], + "angle": 0, + "content": "Where \\( S \\) is the scaling factor determined by data precision \\( \\mathbf{n} \\) and the numerical range. Specifically, \\( S \\) characterizes the \"interval\" between adjacent levels in the floating point dequantized values, bounded by the range [lb, ub]:" + }, + { + "type": "equation", + "bbox": [ + 0.428, + 0.826, + 0.786, + 0.842 + ], + "angle": 0, + "content": "\\[\nS = | \\mathrm {u b} - \\mathrm {l b} | / 2 ^ {\\mathbf {n}} - 1 \\tag {2}\n\\]" + } + ], + [ + { + "type": "header", + "bbox": [ + 0.684, + 0.115, + 0.732, + 0.127 + ], + "angle": 0, + "content": "POCA" + }, + { + "type": "page_number", + "bbox": [ + 0.775, + 0.117, + 0.785, + 0.127 + ], + "angle": 0, + "content": "5" + }, + { + "type": "image", + "bbox": [ + 0.262, + 0.145, + 0.74, + 0.414 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.214, + 0.425, + 0.788, + 0.468 + ], + "angle": 0, + "content": "Fig. 2: Top: The rendered avatar exhibits minimal quality (PSNR) reduction in static frames and large distortion with animation. Bottom: SoTA 8-bit quantization [2,18] amplifies the pixel shifting between frames in the rendered avatar across RGB channels." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.498, + 0.789, + 0.727 + ], + "angle": 0, + "content": "Since the number of levels is directly determined by the precision \\(\\mathbf{n}\\), the essence of quantization accurately approximates the floating-point distribution with the dequantized tensors while minimizing the model-level error. To that end, a variety of quantization algorithms have been proposed to locate the optimal scaling factors with QAT [13] or PTQ [11, 18, 28], including boundary learning [1, 10], scaling factor learning [2], and quantization error modeling [12], etc. Recent efforts introduce additional scaling to avoid the error caused by the extreme distribution [30]. Promoted by the fast compression process, PTQ exhibits high compatibility for quantizing the avatar VAE model compared to QAT. Ideally, compressing the Codec Avatar models with PTQ would lead to instant \"user-avatar\" conversion for AR/VR devices, compared to time-consuming QAT. Recent VAE-based quantization focuses on the video compression task with quantized latent code [7, 21], while the quantization of the codec-avatar decoder model has not been explored despite the urgent needs on edge device and high rendering quality." + }, + { + "type": "title", + "bbox": [ + 0.215, + 0.75, + 0.624, + 0.768 + ], + "angle": 0, + "content": "3 Root Cause of Temporal Avatar Noise" + }, + { + "type": "title", + "bbox": [ + 0.215, + 0.783, + 0.684, + 0.799 + ], + "angle": 0, + "content": "3.1 Temporal Noise is Dynamic Pixel Color Oscillation" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.81, + 0.796, + 0.842 + ], + "angle": 0, + "content": "Recent 3D rendering tasks (e.g., NeRF) [17] uses peak signal-to-noise ratio (PSNR) to evaluate the quality between the rendered output and the ground truth im" + } + ], + [ + { + "type": "page_number", + "bbox": [ + 0.218, + 0.116, + 0.23, + 0.127 + ], + "angle": 0, + "content": "6" + }, + { + "type": "header", + "bbox": [ + 0.272, + 0.115, + 0.366, + 0.129 + ], + "angle": 0, + "content": "J. Meng et al." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.147, + 0.784, + 0.176 + ], + "angle": 0, + "content": "age. However, the temporal distortion of Codec Avatar cannot be effectively quantified with the same metrics." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.178, + 0.785, + 0.253 + ], + "angle": 0, + "content": "With the low-precision decoding, the temporal noise shows minimal impact on the static frame quality, but largely distorts the dynamic motion of the avatar, as shown in Figure 2 and supplementary material. Prior works only focus on the static image quality [25] with low-precision models, but the dynamic quality reduction has not been investigated well." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.254, + 0.785, + 0.299 + ], + "angle": 0, + "content": "In addition to the FovVideoVDP score [16], the temporal pixel oscillation can be characterized by the mean pixel value differences between two consecutive frames \\(F_{t}\\) and \\(F_{t + 1}\\) of the rendered avatar:" + }, + { + "type": "equation", + "bbox": [ + 0.413, + 0.312, + 0.785, + 0.34 + ], + "angle": 0, + "content": "\\[\n\\Delta_ {F} ^ {c} = \\frac {1}{\\mathcal {H W}} \\times | F _ {t + 1} - F _ {t} | \\tag {3}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.215, + 0.349, + 0.787, + 0.41 + ], + "angle": 0, + "content": "Where \\(\\mathcal{H}\\) and \\(\\mathcal{W}\\) represent the height and width of the rendered frame. We evaluate the pixel value differences separately for RGB channels such that \\(c \\in \\{R, G, B\\}\\). The trend of the pixel changes characterizes the basic transition of the facial expression (e.g., Tongue in and out), as shown in Figure 2 (Bottom)." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.411, + 0.786, + 0.561 + ], + "angle": 0, + "content": "With the 8-bit VAE decoder, the frame-wise pixel dynamics exhibit a similar trend as the full-precision decoder baseline. However, the quantized decoder adds an additional \"offset\" to the transition of the facial motion. In other words, the additional distortion is introduced to the avatar on top of the natural motion, which leads to random texture fluctuation in the rendered Codec Avatar. Figure 2 (Bottom) shows that the temporal noises uniformly exist across the entire facial motion with low intensity of \\(\\pm 0.44\\), \\(\\pm 0.47\\), and \\(\\pm 0.25\\) average pixel shifting on the R, G, B channels, respectively. We are fully aware of the existence of other motion capture matrices (e.g., optical flow), but the temporal noises are significantly fine-grained, which requires detailed pixel-level analysis." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.562, + 0.787, + 0.668 + ], + "angle": 0, + "content": "Formally, such random pixel flickering can be captured by the FovVideoVDP [16] score between the rendered avatar and ground truth image. As shown in Figure 2, the INT8 avatar decoder largely degrades the animation quality of the avatar, even with the SoTA 8-bit quantization method [2, 18]. The high density of the yellow region indicates the strongly mismatched animation caused by the temporal noise. The continuous pixel oscillation and temporal noises worsen the fidelity of the rendered Codec Avatar in motion." + }, + { + "type": "title", + "bbox": [ + 0.215, + 0.692, + 0.502, + 0.708 + ], + "angle": 0, + "content": "3.2 Root Cause of Noisy Avatars" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.719, + 0.785, + 0.809 + ], + "angle": 0, + "content": "Since the temporal noises are induced by the pixel oscillation of the avatar in motion, an intuitive understanding of the noises is the \"unstable rounding\" during quantization [12, 19]. However, the rounding instability is still the phenomenon of quantization instead of the root cause of the noise. In this work, we unravel the fundamental reason of the temporal noise induced by low-precision avatar decoding." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.81, + 0.787, + 0.84 + ], + "angle": 0, + "content": "Dynamically long-tailed distribution of human face. Human facial features naturally display a diverse range of distributions characterized by dif" + } + ], + [ + { + "type": "header", + "bbox": [ + 0.684, + 0.115, + 0.732, + 0.127 + ], + "angle": 0, + "content": "POCA" + }, + { + "type": "page_number", + "bbox": [ + 0.775, + 0.116, + 0.785, + 0.127 + ], + "angle": 0, + "content": "7" + }, + { + "type": "image", + "bbox": [ + 0.284, + 0.146, + 0.48, + 0.279 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.32, + 0.282, + 0.342, + 0.297 + ], + "angle": 0, + "content": "(a)" + }, + { + "type": "image", + "bbox": [ + 0.482, + 0.153, + 0.729, + 0.279 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.588, + 0.282, + 0.61, + 0.297 + ], + "angle": 0, + "content": "(b)" + }, + { + "type": "image_caption", + "bbox": [ + 0.215, + 0.308, + 0.785, + 0.336 + ], + "angle": 0, + "content": "Fig. 3: (a) Row-channel plane of the texture decoder activation and (b) Regional standard deviation of the row-channel planes with respect to different facial expressions." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.365, + 0.785, + 0.471 + ], + "angle": 0, + "content": "ferent facial regions (for instance, hair compared to cheeks). The broad spectrum of variability in human facial characteristics formulates various statistical properties. For the avatar decoding, the texture decoder of the VAE model [14] decodes face avatar texture from the latent space via transposed convolutional layers, which consecutively up-scales the activation feature map from \\( 4 \\times 4 \\) to \\( 1024 \\times 1024 \\) (height and width). Therefore, the statistical variation of the facial texture is also embedded into the activation of the earlier layer." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.476, + 0.787, + 0.568 + ], + "angle": 0, + "content": "Proof of facial dynamics. We prove the dynamic range variations of the texture by analyzing the regional distribution of the latent layer activation. Assume \\(\\mathcal{C}\\), \\(\\mathcal{H}\\), and \\(\\mathcal{W}\\) represent the channel depth, height, and width of the activation tensor. We analyze the statistics by decomposing the entire face texture feature map into multiple \"row-channel planes\". Each plane contains the regional statistics from hair to jaw, with the total number of planes = \\(\\mathcal{H}\\)." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.572, + 0.787, + 0.693 + ], + "angle": 0, + "content": "Figure 3(a) shows the row-channel decomposition based on the intermediate activation feature map of the texture decoder with the shape of \\(16 \\times 256 \\times 256\\). On top of that, we compute the standard deviation of each plane, which characterizes the local tailedness and dynamic range in each region of the face, with respect to different facial expressions. As depicted in Figure 3(b), smiling and eye blinking lead to the long-tailed dynamics in the mouth and eye regions, respectively. As a result, the distribution of the entire face is dynamically long-tailed, which is harmful for quantization. In other words:" + }, + { + "type": "text", + "bbox": [ + 0.215, + 0.699, + 0.787, + 0.731 + ], + "angle": 0, + "content": "The distribution of a decoder feature map contains both constrained and long-tailed distribution, while the tailedness varies between different facial expressions." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.735, + 0.787, + 0.84 + ], + "angle": 0, + "content": "In other words, the root cause of the temporal noises is the conflict between the long-tailed, highly dynamic distribution of activations and the global error minimization during PTQ. To minimize the global quantization error, the long-tailed distribution stretches the quantization interval (scaling factor) in order to cover larger range, which in turn, elevates the local rounding instability and leads to unstable numerical representation of the pixels. We prove this hypothesis in Section 4.3 with the proposed method." + } + ], + [ + { + "type": "page_number", + "bbox": [ + 0.218, + 0.116, + 0.23, + 0.127 + ], + "angle": 0, + "content": "8" + }, + { + "type": "header", + "bbox": [ + 0.272, + 0.115, + 0.366, + 0.129 + ], + "angle": 0, + "content": "J. Meng et al." + }, + { + "type": "title", + "bbox": [ + 0.216, + 0.145, + 0.504, + 0.163 + ], + "angle": 0, + "content": "4 Proposed Method: POCA" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.18, + 0.788, + 0.258 + ], + "angle": 0, + "content": "To resolve the temporal noises caused by low-precision avatar decoding, this section presents POCA, a novel post-training quantization algorithm that successfully compresses the deep appearance decoder and eliminates the temporal noises from the rendered Codec Avatar. In particular, POCA is empowered by two novel techniques: a) View-aware PTQ and b) PTQ with shape-wise filtering." + }, + { + "type": "title", + "bbox": [ + 0.216, + 0.282, + 0.406, + 0.298 + ], + "angle": 0, + "content": "4.1 View-aware PTQ" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.311, + 0.789, + 0.476 + ], + "angle": 0, + "content": "As pointed out in Section 3, the face texture contains both region-dependent dispersion and facial expression-driven dynamics in the entire face. As a result, the sample space of the training data and layer-wise activation contains high variation with redundant information. In particular, the texture decoder is trained by the whole face texture map with the size of \\(1024 \\times 1024\\). However, the final 3D avatar is rendered based on the visibility information from the given view direction. Therefore, quantizing the decoder without considering the visibility for inner layers leads to the incorrectly learned quantization parameter, which further generates the mismatched quantization decision between PTQ and actual rendering. Motivated by that, we propose the view-aware PTQ for the avatar texture decoder." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.478, + 0.789, + 0.646 + ], + "angle": 0, + "content": "Basic Setup. Given the pre-trained full-precision VAE decoder (texture and mesh), we first randomly select \\( m \\) training samples from the full-sized dataset with size \\( \\mathcal{D} \\) (\\( m << \\mathcal{D} \\)). Each training sample contains the input and pre-computed output of the decoder (\\([z_t, v_t], [\\hat{\\mathbf{T}}_t, \\hat{\\mathbf{M}}_t]\\)), where \\( z_t, v_t, \\hat{\\mathbf{T}}_t \\), and \\( \\hat{\\mathbf{M}}_t \\) represents the encoded latent code, view direction, decoded texture, and decoded mesh for sample \\( t \\), respectively. Following the standard PTQ strategy, we extract input \\( \\mathcal{X}^l \\) and output activation \\( \\mathcal{Y}^l \\) of each layer as the calibration data. We adopted AdaRound [18] and LSQ [2] as the weight and activation quantizers, respectively. The learnable quantization parameters (\\( \\alpha, S^l \\)) are optimized by minimizing the mean square error (MSE) between the quantized layer output \\( \\hat{\\mathcal{Y}}^l \\) and full precision output \\( \\mathcal{Y}^l \\):" + }, + { + "type": "equation", + "bbox": [ + 0.401, + 0.658, + 0.787, + 0.69 + ], + "angle": 0, + "content": "\\[\n\\underset {\\alpha , S ^ {l}} {\\arg \\min } \\frac {1}{\\mathcal {H W C}} \\sum \\left(\\mathcal {Y} ^ {l} - \\hat {\\mathcal {Y}} ^ {l}\\right) ^ {2} \\tag {4}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.703, + 0.788, + 0.81 + ], + "angle": 0, + "content": "View-dependent visibility mask. The computation of visibility is inherited from the rendering pipeline utilizing the pre-trained VAE model. Specifically, the decoded mesh vertices \\(\\hat{\\mathbf{M}}_t\\) are rasterized to formulate the barycentric coordinates, which are leveraged to interpolate the UV coordinates \\(\\varPhi\\) across the entire avatar. Following this, the visibility is determined by sampling the view-specific pixels from the 2-D texture map based on the interpolated UV indexes, \\(\\varPhi = [\\phi_x,\\phi_y]\\) \\((0\\leq \\phi_{x},\\phi_{y}\\leq 1)\\)." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.81, + 0.788, + 0.842 + ], + "angle": 0, + "content": "The value of \\(\\phi_x\\) and \\(\\phi_y\\) represents the normalized coordinates of the 2-D texture map with the shape of \\(\\mathcal{H} \\times \\mathcal{W}\\). Different from the default pipeline, the" + } + ], + [ + { + "type": "header", + "bbox": [ + 0.684, + 0.115, + 0.732, + 0.127 + ], + "angle": 0, + "content": "POCA" + }, + { + "type": "page_number", + "bbox": [ + 0.776, + 0.117, + 0.786, + 0.127 + ], + "angle": 0, + "content": "9" + }, + { + "type": "image", + "bbox": [ + 0.248, + 0.146, + 0.756, + 0.349 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.215, + 0.36, + 0.785, + 0.403 + ], + "angle": 0, + "content": "Fig. 4: Overall PTQ process of POCA based on layer-wise calibration (Layer Calib). POCA fully quantizes all the layers in the VAE decoder, where the proposed view-dependent \\(+\\) SWD filter is applied to the decoder with transpose convolution layers." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.429, + 0.784, + 0.459 + ], + "angle": 0, + "content": "We proposed POCA algorithm generates the binary \"visibility mask\" \\(\\Gamma\\) to select the pixels in the 2-D texture map \\(\\hat{\\mathbf{T}}_t\\). Formally, we have:" + }, + { + "type": "equation", + "bbox": [ + 0.322, + 0.467, + 0.786, + 0.484 + ], + "angle": 0, + "content": "\\[\n\\Phi = \\text {I n t e r p o l a t e} (\\text {R a s t e r i z e} (\\mathbf {M} _ {t}, \\dots), u v) \\tag {5}\n\\]" + }, + { + "type": "equation", + "bbox": [ + 0.324, + 0.485, + 0.785, + 0.502 + ], + "angle": 0, + "content": "\\[\n\\Gamma = \\mathbb {1} \\cdot \\left\\{\\left(\\gamma_ {x}, \\gamma_ {y}\\right) \\mid \\gamma_ {x} \\in \\left\\{\\phi_ {x} \\times \\mathcal {H} \\right\\}, \\gamma_ {y} \\in \\left\\{\\phi_ {y} \\times \\mathcal {W} \\right\\} \\right\\} \\tag {6}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.215, + 0.508, + 0.784, + 0.568 + ], + "angle": 0, + "content": "Where \\(\\mathbb{1}\\) represents the indicator function that signifies whether the pixel at \\((\\gamma_x,\\gamma_y)\\) should be active at the given visibility. \\(\\{\\phi_x\\times \\mathcal{H}\\}\\) and \\(\\{\\phi_y\\times \\mathcal{W}\\}\\) are the visibility coordinates (matrix indexes) scaled based on the interpolated UV coordinates \\(\\varPhi\\) with the dimensionality of the texture map." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.569, + 0.785, + 0.66 + ], + "angle": 0, + "content": "The visibility masks are generated during the data-fetching process. Since the model is already pre-trained, the visibility masks \\(\\varGamma\\) are accessible with each full-precision texture \\(\\hat{\\mathbf{T}}_t\\) and triangle index \\(uv\\) at the output layer of the texture decoder. The final layer \\(\\varGamma\\) (\\(1024 \\times 1024\\)) will be max-pooled to formulate the visibility mask for the earlier layer. For each layer of the texture decoder, POCA performs PTQ with the visible pixel only:" + }, + { + "type": "equation", + "bbox": [ + 0.355, + 0.665, + 0.786, + 0.699 + ], + "angle": 0, + "content": "\\[\n\\underset {\\alpha^ {l}, S ^ {l}} {\\arg \\min } \\frac {1}{\\mathcal {H W C}} \\sum \\left(\\mathcal {Y} _ {\\mathbf {T}} ^ {l} \\odot \\Gamma^ {l} - \\hat {\\mathcal {Y}} _ {\\mathbf {T}} ^ {l} \\odot \\Gamma^ {l}\\right) ^ {2} \\tag {7}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.705, + 0.788, + 0.84 + ], + "angle": 0, + "content": "Where \\(\\odot\\) represents the element-wise product between tensors. Subsequently, the learnable quantization parameters \\(\\alpha^l\\) and \\(S^l\\) are trained based on the gradient that resulted from the view-dependent visibility \\(\\Gamma\\). As a result, POCA optimizes the quantization decision based on the most relevant information of the training sample, instead of being diversified by the widely dispersed texture space with long-tailed and dynamic distribution. Figure 4 depicts the proposed view-dependent PTQ. Although the quantizer is calibrated based on the visible pixels, POCA only runs PTQ once, and the resultant model will be deployed and rendered for all views." + } + ], + [ + { + "type": "page_number", + "bbox": [ + 0.218, + 0.116, + 0.236, + 0.127 + ], + "angle": 0, + "content": "10" + }, + { + "type": "header", + "bbox": [ + 0.272, + 0.115, + 0.366, + 0.129 + ], + "angle": 0, + "content": "J. Meng et al." + }, + { + "type": "title", + "bbox": [ + 0.216, + 0.147, + 0.6, + 0.163 + ], + "angle": 0, + "content": "4.2 PTQ with Shape-wise Dynamic Filtering" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.174, + 0.788, + 0.264 + ], + "angle": 0, + "content": "Besides the PTQ with the view-dependent visibility masking, we further suppress the impact of the dynamically long-tailed distribution by eliminating the pixels with a high channel-wise dynamic range, or \"Shape-Wise Dynamics\" (SWD). Given the full-precision output \\(\\mathcal{V}_{\\mathbf{T}}^{l}\\) with the shape of \\(\\mathcal{C} \\times \\mathcal{H} \\times \\mathcal{W}\\), the SWD is characterized by the standard deviation score of each \\(\\mathcal{C} \\times 1 \\times 1\\) vector, which formulates the overall texture SWD map with the shape of \\(\\mathcal{H} \\times \\mathcal{W}\\)." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.265, + 0.788, + 0.356 + ], + "angle": 0, + "content": "Subsequently, POCA generates the SWD mask \\(\\Theta\\) by filtering out the pixels in the SWD map with the highest tailedness. The sparsity of the SWD mask is controlled by a tunable hyperparameter \\(\\tau\\). E.g., \\(\\tau = 0.7\\) means masking out the top \\(70\\%\\) pixels with the highest tailedness (standard deviation). The resultant SWD mask \\(\\Theta\\) will be multiplied with the view-dependent mask \\(\\Gamma\\) and generate the final learning objective of the layer-wise post-training quantization of POCA:" + }, + { + "type": "equation", + "bbox": [ + 0.322, + 0.368, + 0.786, + 0.4 + ], + "angle": 0, + "content": "\\[\n\\underset {\\alpha^ {l}, S ^ {l}} {\\arg \\min } \\frac {1}{\\mathcal {H W C}} \\sum \\left(\\mathcal {Y} _ {\\mathbf {T}} ^ {l} \\odot \\Gamma^ {l} \\odot \\Theta^ {l} - \\hat {\\mathcal {Y}} _ {\\mathbf {T}} ^ {l} \\odot \\Gamma^ {l} \\odot \\Theta^ {l}\\right) ^ {2} \\tag {8}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.412, + 0.787, + 0.472 + ], + "angle": 0, + "content": "As shown in Eq. (8), POCA optimizes the quantizers by learning the most relevant information (view-dependent) and least diversified pixels (SWD masking) of the avatar, which suppresses the drawbacks of the long-tailedness and high dynamics of the avatar." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.473, + 0.788, + 0.58 + ], + "angle": 0, + "content": "The masking strategy of POCA introduces a high degree of sparsity in layer-wise PTQ, controlled by visibility and the SWD threshold \\(\\tau\\). Our experimental results in Section 5 show that the optimal performance is achieved with \\(\\tau = 0.9\\). As a result, POCA calibrates the low-precision layer based on highly sparse data \\([\\mathcal{X}^l,\\mathcal{Y}_{\\mathbf{T}}^l]\\) with \\(\\geq 90\\%\\) sparsity. In other words, POCA only requires \\(\\leq 10\\%\\) of the activation for layer-wise PTQ, showing low memory overhead for the efficient PTQ calibration process." + }, + { + "type": "title", + "bbox": [ + 0.216, + 0.602, + 0.466, + 0.617 + ], + "angle": 0, + "content": "4.3 Rationale behind POCA" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.628, + 0.788, + 0.78 + ], + "angle": 0, + "content": "POCA alleviates the impact of rounding instability. With the proposed view-aware PTQ and SWD masking, POCA aims to optimize the layerwise quantizers with the most relevant and quantization-friendly features. The POCA-trained activation quantizer (LSQ) has a generally smaller scaling factor \\( S^l \\) compared to the PTQ without POCA, as shown in Figure 5. The reduced scaling factors formulate the low-precision grid with small intervals (scaling factor) between adjacent levels, alleviating the numerical differences between adjacent levels in the dequantized high-precision domain. As a result, the impact of the aforementioned rounding instability is minimized and the temporal noises become invisible, which proves the aforementioned hypothesis in Section 3.2." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.781, + 0.788, + 0.842 + ], + "angle": 0, + "content": "Sparse masking \\(\\neq\\) biased learning. Although two types of sparse masking are introduced into PTQ, the POCA-trained quantizers will not lead to a biased representation in the quantized model. As shown in Figure 4, the sparsified calibration mask \\((\\varGamma^l\\odot\\varTheta^l)\\) contains sparse but uniformly sampled information" + } + ], + [ + { + "type": "header", + "bbox": [ + 0.684, + 0.115, + 0.732, + 0.127 + ], + "angle": 0, + "content": "POCA" + }, + { + "type": "page_number", + "bbox": [ + 0.769, + 0.116, + 0.785, + 0.127 + ], + "angle": 0, + "content": "11" + }, + { + "type": "image", + "bbox": [ + 0.315, + 0.144, + 0.689, + 0.272 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.28, + 0.274, + 0.723, + 0.289 + ], + "angle": 0, + "content": "Fig. 5: Scaling factor \\( S^l \\) of the texture decoder trained by POCA." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.307, + 0.784, + 0.335 + ], + "angle": 0, + "content": "across all the regions of the avatar texture. As a result, the learnable quantizers are uniformly trained with all the relevant facial information." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.336, + 0.785, + 0.412 + ], + "angle": 0, + "content": "Highly-correlated data and high fidelity sparse learning. Unlike classification tasks, the training samples of the VAE model are highly correlated with respect to the single identity. As a result, the sparsified information of each sample can be compensated from other training data with different viewing perspectives, leading to sparse but robust PTQ with the quantized decoder." + }, + { + "type": "title", + "bbox": [ + 0.216, + 0.433, + 0.461, + 0.451 + ], + "angle": 0, + "content": "5 Experimental Results" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.462, + 0.785, + 0.538 + ], + "angle": 0, + "content": "We evaluated the proposed POCA algorithm based on Multiface [29], a public avatar dataset captured by the multi-camera system. POCA is applied to eight avatars among three viewing directions with all 65 facial expressions of the dataset. We want to highlight that the eight avatars comprehensively cover different genders, age groups, hairstyles, skin types, and facial hairs." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.539, + 0.787, + 0.644 + ], + "angle": 0, + "content": "For each identity, we use the corresponding pre-trained deep appearance VAE model as the starting point of POCA. Following the standard PTQ strategy, we quantize all layers of the decoders (texture and mesh) iteratively, where each layer is calibrated by 100 epochs with the sampled data. POCA only acquires 512 training samples from each identity, which is \\(\\leq 5\\%\\) samples of the dataset. For each layer, the quantizers are optimized by Adam with 1e-4 learning rate and eight samples per batch." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.645, + 0.787, + 0.749 + ], + "angle": 0, + "content": "Prior Works for Comparison. As the first work in the low-precision avatar decoding, we select three prior quantization works that reported the SoTA INT8 performance with the classification tasks: (1) Adaptive Rounding (AdaRound) [18], (2) learned step size quantization (LSQ) [2], which are also incorporated for hardware deployment [24], and (3) W8A8 naive QAT [20] in PyTorch. These selected prior works achieved superior INT8 performance [2, 18] on the ImageNet-1K dataset, even outperforming the full-precision baseline." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.75, + 0.787, + 0.84 + ], + "angle": 0, + "content": "Evaluation Metric. We evaluate the proposed POCA algorithm based on the FovVideoVDP [16] between the rendered avatar and the ground truth video to quantify the temporal noise of the dynamic Codec Avatar. For each avatar identity, we compute the averaged FovVideoVDP score among all the 65 expressions in the Multiface dataset [29]. The performance of POCA is demonstrated with the rendered avatar videos in the supplementary material." + } + ], + [ + { + "type": "page_number", + "bbox": [ + 0.218, + 0.116, + 0.236, + 0.127 + ], + "angle": 0, + "content": "12" + }, + { + "type": "header", + "bbox": [ + 0.272, + 0.115, + 0.366, + 0.129 + ], + "angle": 0, + "content": "J. Meng et al." + }, + { + "type": "table_caption", + "bbox": [ + 0.215, + 0.145, + 0.788, + 0.201 + ], + "angle": 0, + "content": "Table 1: FovVideoVDP score [16] between the rendered avatar and the ground truth based on eight avatars across all the expressions from three viewing perspectives. The proposed POCA algorithm achieves the highest quality compared to the SoTA PTQ and QAT methods with both 8-bit and 6-bit precision." + }, + { + "type": "table", + "bbox": [ + 0.217, + 0.213, + 0.787, + 0.32 + ], + "angle": 0, + "content": "
Avatar IDPrecision26438142183941537202126453108870559667444329145896795937
Male, GoateeBlack, FemaleMohawk, MaleAsian, FemaleOlder WomanMale, Mid HairBlonde GirlMale, Short Hair
FrontTopFrontTopFrontTopFrontTopFrontTopFrontTopFrontTopFrontTop
Baseline [29]FP326.214.945.306.544.995.456.144.875.406.325.025.596.054.955.136.02
SoTA PTQ [2,18]W8A85.814.434.875.914.405.005.804.525.195.874.565.105.434.444.685.52
W8A8 QAT [20]5.924.554.996.024.575.125.884.565.235.974.665.165.574.504.795.60
POCA6.194.935.266.314.805.386.094.835.336.214.955.425.954.825.085.99
SoTA PTQ [2,18]W6A64.503.664.064.963.814.124.554.104.475.223.914.454.293.733.964.25
W6A6 QAT5.083.754.395.544.014.234.944.194.495.364.074.664.363.854.094.46
POCA6.064.555.026.294.425.136.014.615.166.024.665.215.454.494.995.66
" + }, + { + "type": "image_caption", + "bbox": [ + 0.277, + 0.321, + 0.45, + 0.331 + ], + "angle": 0, + "content": "Yellow region \\(=\\) Mismatched pixel motion" + }, + { + "type": "image", + "bbox": [ + 0.223, + 0.331, + 0.784, + 0.491 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.336, + 0.492, + 0.696, + 0.503 + ], + "angle": 0, + "content": "FovVideoVDP [14] between the rendered avatar and ground truth (Max = 10.0)" + }, + { + "type": "image_caption", + "bbox": [ + 0.215, + 0.51, + 0.788, + 0.552 + ], + "angle": 0, + "content": "Fig. 6: Rendered avatars and temporal pixel differences between two consecutive frames of the avatar video. The lightened yellow region of POCA signifies the removed temporal noise as the full precision baseline." + }, + { + "type": "title", + "bbox": [ + 0.215, + 0.583, + 0.642, + 0.599 + ], + "angle": 0, + "content": "5.1 Temporal Noise Reduction with High Fidelity" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.613, + 0.788, + 0.718 + ], + "angle": 0, + "content": "We evaluate the proposed POCA algorithm with eight avatars with different genders, hairstyles, and skin colors. Table 1 summarizes the FovVideoVDP score [16] of the proposed POCA algorithm with 8-bit and 6-bit avatar decoders. The SoTA INT8 PTQ quantization algorithm failed to preserve the avatar quality as the baseline full precision model, exhibiting the degraded FovVideoVDP score. Even with the expensive quantization-aware training (QAT), the quality of the rendered avatar cannot be restored." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.72, + 0.789, + 0.841 + ], + "angle": 0, + "content": "Compared to the previous methods with 8-bit precision, POCA eliminates the temporal noises in the rendered avatar. POCA successfully compresses the decoder model while achieving on-par quality as the full precision decoder baseline. Figure 6 shows the high avatar quality and restored FovVideoVDP score (heat map) of the proposed POCA algorithm (closer to the FP32 baseline = better quantization). Regarding the efficiency of PTQ, POCA only requires \\(\\leq 10\\%\\) of the activations for layer-wise PTQ, showing low memory overhead for the efficient PTQ calibration process." + } + ], + [ + { + "type": "header", + "bbox": [ + 0.684, + 0.115, + 0.732, + 0.127 + ], + "angle": 0, + "content": "POCA" + }, + { + "type": "page_number", + "bbox": [ + 0.769, + 0.116, + 0.786, + 0.127 + ], + "angle": 0, + "content": "13" + }, + { + "type": "image", + "bbox": [ + 0.276, + 0.145, + 0.362, + 0.256 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.281, + 0.257, + 0.363, + 0.275 + ], + "angle": 0, + "content": "SSIM=0.741 \nFovVideo\\(\\mathrm{DVP}=6.213\\)" + }, + { + "type": "image", + "bbox": [ + 0.364, + 0.145, + 0.451, + 0.256 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.368, + 0.257, + 0.451, + 0.275 + ], + "angle": 0, + "content": "SSIM=0.741 \nFovVideoDVP=6.119" + }, + { + "type": "image", + "bbox": [ + 0.453, + 0.145, + 0.545, + 0.256 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.455, + 0.257, + 0.539, + 0.275 + ], + "angle": 0, + "content": "SSIM=0.741 \nFovVideoVDP=1.685" + }, + { + "type": "image", + "bbox": [ + 0.546, + 0.145, + 0.634, + 0.256 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.551, + 0.257, + 0.633, + 0.275 + ], + "angle": 0, + "content": "SSIM=0.740 \nFovVideoVDP=6.121" + }, + { + "type": "image", + "bbox": [ + 0.636, + 0.145, + 0.729, + 0.256 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.639, + 0.257, + 0.722, + 0.275 + ], + "angle": 0, + "content": "SSIM=0.739 \nFoxVideoVDP=6.106" + }, + { + "type": "image_caption", + "bbox": [ + 0.216, + 0.285, + 0.785, + 0.314 + ], + "angle": 0, + "content": "Fig. 7: The impact of the SWD threshold. Extreme filtering leads to distorted and missing texture in the resultant avatar." + }, + { + "type": "table_caption", + "bbox": [ + 0.291, + 0.318, + 0.71, + 0.332 + ], + "angle": 0, + "content": "Table 2: Training time comparison between POCA and QAT." + }, + { + "type": "table", + "bbox": [ + 0.271, + 0.334, + 0.728, + 0.409 + ], + "angle": 0, + "content": "
MethodQuantizationMethodW/ATimeGPUDecoder Size (MB)
FP32 Baseline-32/32---32.78
Native W8A8 [20]QAT8/8>10 hrs2 × A60008.19 (↓ 4×)
POCA (This work)PTQ8/8~13 min1 × A60008.19 (↓ 4×)
POCA (This work)PTQ6/6~13 min1 × A60006.15 (↓ 5.3×)
" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.438, + 0.788, + 0.558 + ], + "angle": 0, + "content": "POCA generates clean avatars without introducing any graphic filtering techniques [31]. From the perspective of quantization, POCA keeps the vanilla quantization strategy in Eq. (1) in both training and inference, compressing the model without introducing distribution manipulations [28] on the additional scaling [30] and bias [12]. Compared to traditional QAT, POCA quantizes the entire model in only 13 minutes (Table 2). The fast deployment workflow with the actual VR device (e.g., Quest Pro) can further accelerate the process of \"user-to-meta verse\"." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.56, + 0.807, + 0.636 + ], + "angle": 0, + "content": "Future Applicability: POCA is not limited to the deep appearance model [14]. The proposed POCA algorithm is applicable to all mesh-based 3D reconstruction tasks with VAE models. For example, the proposed view-dependent visibility mask and shape-wise filtering can be easily transferred to the recent Pixel Codec Avatar (PiCA)[15], which has the similar autoencoder-based architecture." + }, + { + "type": "title", + "bbox": [ + 0.216, + 0.662, + 0.591, + 0.679 + ], + "angle": 0, + "content": "5.2 Impact of the SWD Filtering Threshold" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.692, + 0.788, + 0.814 + ], + "angle": 0, + "content": "POCA introduced the tunable parameter \\(\\tau\\) in Eq. (8) to control the amount of activation pixels for layer-wise PTQ. Figure 7 demonstrates the impact of the SWD threshold \\(\\tau\\) with respect to the avatar quality (SSIM). POCA achieves the best tradeoff between filter intensity and avatar quality with \\(\\tau^{*} = 0.9\\). Quantizing the decoder with \\(\\tau > \\tau^{*}\\) leads to missing texture pixels and distorted rendering with no quality improvement. The selected SWD threshold exhibits low sensitivity between different avatars and viewing perspectives, as demonstrated by the high-quality rendering in Figure 6." + }, + { + "type": "page_footnote", + "bbox": [ + 0.218, + 0.825, + 0.785, + 0.84 + ], + "angle": 0, + "content": "3 PiCA results are available at https://mengjian0502.github.io/poca.github.io/" + } + ], + [ + { + "type": "page_number", + "bbox": [ + 0.218, + 0.116, + 0.236, + 0.127 + ], + "angle": 0, + "content": "14" + }, + { + "type": "header", + "bbox": [ + 0.272, + 0.115, + 0.366, + 0.129 + ], + "angle": 0, + "content": "J. Meng et al." + }, + { + "type": "title", + "bbox": [ + 0.216, + 0.147, + 0.681, + 0.162 + ], + "angle": 0, + "content": "5.3 Applicability to Other Quality Enhancement Tasks" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.168, + 0.788, + 0.215 + ], + "angle": 0, + "content": "We extend the experiment by evaluating the proposed method on super-resolution tasks. The proposed \"shape-wise-filtering\" (SWD) exhibits strong performance in image quality enhancement [25] with the low-precision SRResNet model [8]." + }, + { + "type": "table_caption", + "bbox": [ + 0.216, + 0.223, + 0.786, + 0.252 + ], + "angle": 0, + "content": "Table 3: SoTA image quality (PSNR/SSIM) achieved by the proposed POCA-SWD algorithm with 8-bit and 6-bit model precision, outperforms the recent SR-PTQ [25]." + }, + { + "type": "table", + "bbox": [ + 0.229, + 0.255, + 0.765, + 0.439 + ], + "angle": 0, + "content": "
MethodPrecisionSparsitySet5 (× 4)Set14 (× 4)BSD100 (× 4)Urban100 (× 4)
SRResNet Baseline [25]FP32(Baseline)N/A32.234 / 0.89628.656 / 0.78427.63 / 0.73826.229 / 0.791
SRResNet [8]32.05 / 0.90128.49 / 0.81827.58 / 0.762Not Reported
SRResNet (This work)31.927 / 0.90128.588 / 0.79927.628 / 0.75626.223 / 0.800
AdaRound [18] + LSQ [2]W8A8Dense31.870 / 0.89928.550 / 0.79827.533 / 0.75426.078 / 0.797
SR-PTQ [25]Dense32.207 / 0.89528.619 / 0.78327.618 / 0.73826.191 / 0.790
POCA80%31.898 / 0.90028.566 / 0.79827.620 / 0.75626.195 / 0.799
Quality Drop [2,18]W8A80.057 / 0.0020.038 / 0.0010.095 / 0.0020.145 / 0.003
Quality Drop [25]0.027 / 0.0010.037 / 0.0010.012 / 0.0000.038 / 0.001
Quality Drop (POCA)0.029 / 0.0010.022 / 0.0010.008 / 0.0000.028 / 0.001
AdaRound [18] + LSQ [2]W6A6Dense31.765 / 0.89728.449 / 0.79527.470 / 0.75225.988 / 0.794
SR-PTQ [25]Dense32.089 / 0.89228.504 / 0.77927.561 / 0.73326.011 / 0.783
POCA80%31.789 / 0.89928.499 / 0.79627.565 / 0.75326.110 / 0.795
Quality Drop [2,18]W6A60.162 / 0.0040.139 / 0.0030.158 / 0.0040.235 / 0.006
Quality Drop [25]0.145 / 0.0040.152 / 0.0050.069 / 0.0050.218 / 0.008
Quality Drop (POCA)0.138 / 0.0020.089 / 0.0040.061 / 0.0030.113 / 0.005
" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.45, + 0.788, + 0.556 + ], + "angle": 0, + "content": "With 8-bit and 6-bit precision, the proposed method outperforms [25] in BSD100 and Urban100 datasets, as shown in Table 3. For a fair comparison, we listed all the baseline full-precision models. Our method outperforms [25] on all datasets with the new SoTA image quality. Compared to [18] and [2], the improved image quality implies the effectiveness of the proposed SWD. More importantly, our method enables the sparse PTQ by using only \\(20\\%\\) pixels of the selected samples, whereas all prior works [2,18,25] require dense PTQ." + }, + { + "type": "title", + "bbox": [ + 0.216, + 0.575, + 0.509, + 0.59 + ], + "angle": 0, + "content": "6 Conclusion and Discussion" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.603, + 0.787, + 0.754 + ], + "angle": 0, + "content": "In this paper, we propose a novel post-training quantization algorithm designed for high-quality Codec Avatar rendering. As the first work in this area, we discover the challenge of temporal noise with low-precision avatar decoding. Through thorough investigation, we reveal the root cause of the temporal noise that originated from the high dynamics and long-tailedness of the texture distribution. To address that, we propose POCA, which consists of view-aware PTQ and shape-wise filtering (SWD). Compared to the previous SoTA methods, POCA eliminates the temporal noises even with 6-bit precision. Furthermore, the proposed SWD algorithm is also applicable to super-resolution tasks, achieving the new SoTA image quality with highly efficient sparse calibration." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.765, + 0.787, + 0.84 + ], + "angle": 0, + "content": "Limitations. The proposed POCA algorithm mainly focuses on the VAE-based Codec Avatar generation, which has been proven as a successful strategy with the actual commercial-level VR product deployment. Recently, 3D Gaussian-based avatar rendering also emerged as a feasible method for generative Avatars, while low-precision quantization of such models will remain as future work." + } + ], + [ + { + "type": "header", + "bbox": [ + 0.684, + 0.115, + 0.732, + 0.127 + ], + "angle": 0, + "content": "POCA" + }, + { + "type": "page_number", + "bbox": [ + 0.769, + 0.116, + 0.786, + 0.127 + ], + "angle": 0, + "content": "15" + }, + { + "type": "title", + "bbox": [ + 0.217, + 0.145, + 0.324, + 0.16 + ], + "angle": 0, + "content": "References" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.18, + 0.787, + 0.223 + ], + "angle": 0, + "content": "1. Choi, J., Venkataramani, S., Srinivasan, V.V., Gopalakrishnan, K., Wang, Z., Chuang, P.: Accurate and efficient 2-bit quantized neural networks. Proceedings of Machine Learning and Systems 1, 348-359 (2019) 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.223, + 0.787, + 0.264 + ], + "angle": 0, + "content": "2. Esser, S.K., McKinstry, J.L., Bablani, D., Appuswamy, R., Modha, D.S.: Learned step size quantization. In: International Conference on Learning Representations (ICLR) (2020) 2, 3, 5, 6, 8, 11, 12, 14" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.265, + 0.787, + 0.32 + ], + "angle": 0, + "content": "3. Fu, Y., Li, Y., Li, C., Saragih, J., Zhang, P., Dai, X., Lin, Y.C.: Auto-card: Efficient and robust codec avatar driving for real-time mobile telepresence. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR) (2023) 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.321, + 0.787, + 0.377 + ], + "angle": 0, + "content": "4. Gorbachev, Y., Fedorov, M., Slavutin, I., Tugarev, A., Fatekhov, M., Tarkan, Y.: Openvino deep learning workbench: Comprehensive analysis and tuning of neural networks inference. In: Proceedings of the IEEE/CVF International Conference on Computer Vision Workshops (CVPR-W) (2019) 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.378, + 0.787, + 0.418 + ], + "angle": 0, + "content": "5. Han, Y., Yang, S., Wang, W., Liu, J.: From design draft to real attire: Unaligned fashion image translation. In: Proceedings of the 28th ACM International Conference on Multimedia. p. 1533-1541 (2020) 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.42, + 0.787, + 0.447 + ], + "angle": 0, + "content": "6. Kingma, D.P., Welling, M.: Auto-encoding variational bayes. arXiv preprint arXiv:1312.6114 (2013) 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.448, + 0.787, + 0.504 + ], + "angle": 0, + "content": "7. Le, H., Zhang, L., Said, A., Sautiere, G., Yang, Y., Shrestha, P., Yin, F., Pourreza, R., Wiggers, A.: Mobilecodec: neural inter-frame video compression on mobile devices. In: Proceedings of the 13th ACM Multimedia Systems Conference. pp. 324-330 (2022) 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.505, + 0.787, + 0.572 + ], + "angle": 0, + "content": "8. Ledig, C., Theis, L., Huszár, F., Caballero, J., Cunningham, A., Acosta, A., Aitken, A., Tejani, A., Totz, J., Wang, Z., et al.: Photo-realistic single image superresolution using a generative adversarial network. In: Proceedings of the IEEE conference on computer vision and pattern recognition. pp. 4681-4690 (2017) 3, 14" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.574, + 0.787, + 0.615 + ], + "angle": 0, + "content": "9. Lewis, J.P., Anjyo, K., Rhee, T., Zhang, M., Pighin, F.H., Deng, Z.: Practice and theory of blendshape facial models. Eurographics (State of the Art Reports) 1(8), 2 (2014) 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.616, + 0.787, + 0.657 + ], + "angle": 0, + "content": "0. Li, Y., Dong, X., Wang, W.: Additive powers-of-two quantization: An efficient non-uniform discretization for neural networks. In: International Conference on Learning Representations (ICLR) (2020) 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.659, + 0.787, + 0.699 + ], + "angle": 0, + "content": "1. Li, Y., Gong, R., Tan, X., Yang, Y., Hu, P., Zhang, Q., Yu, F., Wang, W., Gu, S.: {BRECQ}: Pushing the limit of post-training quantization by block reconstruction. In: International Conference on Learning Representations (ICLR) (2021) 2, 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.7, + 0.787, + 0.756 + ], + "angle": 0, + "content": "2. Liu, Y., Yang, H., Dong, Z., Keutzer, K., Du, L., Zhang, S.: Noisyquant: Noisy bias-enhanced post-training activation quantization for vision transformers. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR) (2023) 3, 5, 6, 13" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.757, + 0.787, + 0.812 + ], + "angle": 0, + "content": "3. Liu, Z., Cheng, K.T., Huang, D., Xing, E.P., Shen, Z.: Nonuniform-to-uniform quantization: Towards accurate quantization via generalized straight-through estimation. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR) (2022) 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.813, + 0.787, + 0.84 + ], + "angle": 0, + "content": "4. Lombardi, S., Saragih, J., Simon, T., Sheikh, Y.: Deep appearance models for face rendering. ACM Transactions on Graphics (ToG) (2018) 1, 2, 4, 7, 13" + }, + { + "type": "list", + "bbox": [ + 0.226, + 0.18, + 0.787, + 0.84 + ], + "angle": 0, + "content": null + } + ], + [ + { + "type": "page_number", + "bbox": [ + 0.218, + 0.116, + 0.236, + 0.127 + ], + "angle": 0, + "content": "16" + }, + { + "type": "header", + "bbox": [ + 0.272, + 0.115, + 0.366, + 0.129 + ], + "angle": 0, + "content": "J. Meng et al." + }, + { + "type": "ref_text", + "bbox": [ + 0.218, + 0.147, + 0.787, + 0.189 + ], + "angle": 0, + "content": "15. Ma, S., Simon, T., Saragih, J., Wang, D., Li, Y., De La Torre, F., Sheikh, Y.: Pixel codec avatars. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR) (2021) 1, 2, 4, 13" + }, + { + "type": "ref_text", + "bbox": [ + 0.218, + 0.19, + 0.787, + 0.245 + ], + "angle": 0, + "content": "16. Mantiuk, R.K., Denes, G., Shapiro, A., Kaplanyan, A., Rufo, G., Bachy, R., Lian, T., Patney, A.: FovVideoVDP: A visible difference predictor for wide field-of-view video. In: ACM Transactions on Graphics (TOG). ACM New York, NY, USA (2021) 3, 6, 11, 12" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.246, + 0.787, + 0.286 + ], + "angle": 0, + "content": "17. Mildenhall, B., Srinivasan, P.P., Tancik, M., Barron, J.T., Ramamoorthi, R., Ng, R.: NeRF: Representing scenes as neural radiance fields for view synthesis. Communications of the ACM 65(1), 99-106 (2021) 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.287, + 0.787, + 0.328 + ], + "angle": 0, + "content": "18. Nagel, M., Amjad, R.A., Van Baalen, M., Louizos, C., Blankevoort, T.: Up or down? adaptive rounding for post-training quantization. In: International Conference on Machine Learning (ICML) (2020) 2, 3, 5, 6, 8, 11, 12, 14" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.329, + 0.787, + 0.357 + ], + "angle": 0, + "content": "19. Park, E., Yoo, S.: Profit: A novel training method for sub-4-bit mobilenet models. In: European Conference on Computer Vision (ECCV) (2020) 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.357, + 0.787, + 0.424 + ], + "angle": 0, + "content": "20. Paszke, A., Gross, S., Massa, F., Lerer, A., Bradbury, J., Chanan, G., Killeen, T., Lin, Z., Gimelshein, N., Antiga, L., Desmaison, A., Kopf, A., Yang, E., DeVito, Z., Raison, M., Tejani, A., Chilamkurthy, S., Steiner, B., Fang, L., Bai, J., Chintala, S.: Pytorch: An imperative style, high-performance deep learning library. In: Advances in Neural Information Processing Systems (NeurIPS) (2019) 3, 11, 12, 13" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.425, + 0.787, + 0.481 + ], + "angle": 0, + "content": "21. van Rozendaal, T., Singhal, T., Le, H., Sautiere, G., Said, A., Buska, K., Raha, A., Kalatzis, D., Mehta, H., Mayer, F., et al.: Mobilenvc: Real-time 1080p neural video compression on a mobile device. In: Proceedings of the IEEE/CVF Winter Conference on Applications of Computer Vision. pp. 4323-4333 (2024) 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.481, + 0.787, + 0.522 + ], + "angle": 0, + "content": "22. Saito, S., Schwartz, G., Simon, T., Li, J., Nam, G.: Relightable gaussian codec avatars. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 130-141 (2024) 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.522, + 0.787, + 0.564 + ], + "angle": 0, + "content": "23. Schwartz, G., Wei, S.E., Wang, T.L., Lombardi, S., Simon, T., Saragih, J., Sheikh, Y.: The eyes have it: An integrated eye and face model for photorealistic facial animation. ACM Transactions on Graphics (TOG) (2020) 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.565, + 0.787, + 0.605 + ], + "angle": 0, + "content": "24. Siddegowda, S., Fournarakis, M., Nagel, M., Blankevoort, T., Patel, C., Khobare, A.: Neural network quantization with AI model efficiency toolkit (AIMET). arXiv preprint arXiv:2201.08442 (2022) 2, 3, 4, 11" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.606, + 0.787, + 0.647 + ], + "angle": 0, + "content": "25. Tu, Z., Hu, J., Chen, H., Wang, Y.: Toward accurate post-training quantization for image super resolution. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR) (2023) 2, 3, 6, 14" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.647, + 0.787, + 0.688 + ], + "angle": 0, + "content": "26. Wang, Z., Wu, Z., Lu, J., Zhou, J.: Bidet: An efficient binarized object detector. In: Proceedings of the IEEE/CVF conference on computer vision and pattern recognition (CVPR) (2020) 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.689, + 0.787, + 0.73 + ], + "angle": 0, + "content": "27. Wei, S.E., Saragih, J., Simon, T., Harley, A.W., Lombardi, S., Perdoch, M., Hypes, A., Wang, D., Badino, H., Sheikh, Y.: VR facial animation via multiview image translation. ACM Transactions on Graphics (TOG) (2019) 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.73, + 0.787, + 0.771 + ], + "angle": 0, + "content": "28. Wei, X., Gong, R., Li, Y., Liu, X., Yu, F.: QDrop: Randomly dropping quantization for extremely low-bit post-training quantization. In: International Conference on Learning Representations (ICLR) (2022) 2, 5, 13" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.772, + 0.787, + 0.84 + ], + "angle": 0, + "content": "29. Wuu, C.h., Zheng, N., Ardisson, S., Bali, R., Belko, D., Brockmeyer, E., Evans, L., Godisart, T., Ha, H., Huang, X., Hypes, A., Koska, T., Krenn, S., Lombardi, S., Luo, X., McPhail, K., Millerschoen, L., Perdoch, M., Pitts, M., Richard, A., Saragih, J., Saragih, J., Shiratori, T., Simon, T., Stewart, M., Trimble, A., Weng, X., Whitewolf, D., Wu, C., Yu, S.I., Sheikh, Y.: Multiface: A dataset for neural" + }, + { + "type": "list", + "bbox": [ + 0.217, + 0.147, + 0.787, + 0.84 + ], + "angle": 0, + "content": null + } + ], + [ + { + "type": "header", + "bbox": [ + 0.684, + 0.115, + 0.732, + 0.127 + ], + "angle": 0, + "content": "POCA" + }, + { + "type": "page_number", + "bbox": [ + 0.769, + 0.116, + 0.786, + 0.127 + ], + "angle": 0, + "content": "17" + }, + { + "type": "ref_text", + "bbox": [ + 0.241, + 0.147, + 0.786, + 0.175 + ], + "angle": 0, + "content": "face rendering. In: arXiv (2022). https://doi.org/10.48550/ARXIV.2207.11243, https://arxiv.org/abs/2207.11243 1, 11, 12" + }, + { + "type": "ref_text", + "bbox": [ + 0.218, + 0.177, + 0.787, + 0.217 + ], + "angle": 0, + "content": "30. Xiao, G., Lin, J., Seznek, M., Wu, H., Demouth, J., Han, S.: Smoothquant: Accurate and efficient post-training quantization for large language models. In: International Conference on Machine Learning (ICML) (2023) 2, 5, 13" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.218, + 0.786, + 0.244 + ], + "angle": 0, + "content": "31. Yin, H., Gong, Y., Qiu, G.: Side window filtering. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR) (2019) 13" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.246, + 0.786, + 0.286 + ], + "angle": 0, + "content": "32. Zhang, Y., Zhang, Z., Lew, L.: Pokebnn: A binary pursuit of lightweight accuracy. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR) (2022) 4" + }, + { + "type": "list", + "bbox": [ + 0.217, + 0.147, + 0.787, + 0.286 + ], + "angle": 0, + "content": null + } + ] +] \ No newline at end of file diff --git a/2024/POCA_ Post-training Quantization with Temporal Alignment for Codec Avatars/f5688d31-6662-4aa8-b308-c286c399e498_origin.pdf b/2024/POCA_ Post-training Quantization with Temporal Alignment for Codec Avatars/f5688d31-6662-4aa8-b308-c286c399e498_origin.pdf new file mode 100644 index 0000000000000000000000000000000000000000..9880d57164c8e9263e9507bf004daef542e013ea --- /dev/null +++ b/2024/POCA_ Post-training Quantization with Temporal Alignment for Codec Avatars/f5688d31-6662-4aa8-b308-c286c399e498_origin.pdf @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:7188d7955dcb3e44a896ee1e08aae377e462f57335a85dcbd23cc1f62031791c +size 5685471 diff --git a/2024/POCA_ Post-training Quantization with Temporal Alignment for Codec Avatars/full.md b/2024/POCA_ Post-training Quantization with Temporal Alignment for Codec Avatars/full.md new file mode 100644 index 0000000000000000000000000000000000000000..070069c8cc625ac5a486fdec9d53feb1335bd7f2 --- /dev/null +++ b/2024/POCA_ Post-training Quantization with Temporal Alignment for Codec Avatars/full.md @@ -0,0 +1,289 @@ +# POCA: Post-training Quantization with Temporal Alignment for Codec Avatars + +Jian Meng $^{1,2}$ , Yuecheng Li $^{1}$ , Chenghui (Leo) Li $^{1}$ , Syed Shakib Sarwar $^{1}$ , Dilin Wang $^{1}$ , and Jae-sun Seo $^{2}$ + +1 Meta Codec Avatars Lab, Pittsburgh, USA + +2 Cornell Tech, New York City, USA + +Abstract. Real-time decoding generates high-quality assets for rendering photorealistic Codec Avatars for immersive social telepresence with AR/VR. However, high-quality avatar decoding incurs expensive computation and memory consumption, which necessitates the design of a decoder compression algorithm (e.g., quantization). Although quantization has been widely studied, the quantization of avatar decoders is an urgent yet under-explored need. Furthermore, the requirement of fast "User-Avatar" deployment prioritizes the post-training quantization (PTQ) over the time-consuming quantization-aware training (QAT). As the first work in this area, we reveal the sensitivity of the avatar decoding quality under low precision. In particular, the state-of-the-art (SoTA) QAT and PTQ algorithms introduce massive amounts of temporal noise to the rendered avatars, even with the well-established 8-bit precision. To resolve these issues, a novel PTQ algorithm is proposed for quantizing the avatar decoder with low-precision weights and activation (8-bit and 6-bit), without introducing temporal noise to the rendered avatar. Furthermore, the proposed method only needs $10\%$ of the activations of each layer to calibrate quantization parameters without any distribution manipulations or extensive boundary search. The proposed method is evaluated on various face avatars with different facial characteristics. The proposed method compresses the decoder model by $5.3\times$ while recovering the quality on par with the full precision baseline. In addition to the avatar rendering tasks, POCA is also applicable to image resolution enhancement tasks, achieving new SoTA image quality. https://mengjian0502.github.io/poca.github.io/ + +# 1 Introduction + +Photorealistic telepresence [14, 15, 23, 27] has emerged as the cornerstone technology in augmented or virtual reality (AR/VR), enabling virtual interactions across different devices. Specifically, the ultimate goal of photorealistic rendering is animating the high-fidelity Codec Avatars with multiple partial facial images captured by the head-mounted cameras (HMC) [29]. The variational autoencoder-based model [14, 15] has been proved and deployed to VR device (e.g., Quest Pro) as a feasible solution. During inference, the encoder model projects the HMC-captured facial motion into the latent space, and the decoder + +Fig. 1: Decoding and rendering with low-precision avatar decoder. The SoTA 8-bit PTQ algorithm distorts the avatar with massive and random temporal noise, while the proposed method compresses the decoder with the high-quality avatar without noise. +![](images/4f5bb62931d49a77b0787612bcf87fe07dc0cbda4605eb3139e2927f11d35465.jpg) +Noise is visible in the supplementary video: "supplementary/ID002643814/poca_all_002643814_w8a8_400029.mp4" + +reconstructs the latent code into view dependent high-quality texture and mesh, which are then rendered into the 3-D Codec AvATAR on screen. In practice, the critical ingredient of photorealistic telepresence is transmitting the encoded latent information from one VR to another and rendering the avatar in real-time. In practice, the VAE-based Codec AvATAR model has been employed for generative Pixel Codec AvATAR [15] or 3D Gaussian-based AvATAR [22]. Naturally, the quality and latency of telepresence mainly depends on the efficiency of the decoder, which impacts the real-time visual quality. + +Despite the success of the deep appearance model [14], the high computational cost and memory consumption of avatar decoding hinder the fast on-device telepresence. Naïvely deploying the full-precision decoder to the headset leads to a low frame rate, causing degraded user experience. On the other hand, relying on cloud-based computation adds additional "headset-cloud" communication latency and increases the risk of avatar data breaching. + +The conflict between high-fidelity rendering and the urgent needs of on-device computing necessitates the high-performance compression algorithm designed for the avatar decoder. In particular, quantization [2, 18, 28, 30] compresses the model with a unified data compression ratio on weights and activation, leading to easy hardware deployment [24]. Unfortunately, the SoTA quantization algorithms mainly focus on classification [11, 18], object detection [26], and natural language processing tasks [30]. Recent works have investigated high-resolution static image rendering with the low-precision model [25]. However, the high-quality quantization algorithm for avatar decoding remains unexplored. In fact, our research shows that the current efforts on quantization are highly incompatible to the Avatar decoder with the following challenges: + +Massive temporal noises introduced by INT8 quantization. Although the popular 8-bit quantization has been widely investigated with quantization-aware training [20] and post-training quantization [2,18,24], our discovery shows that the well-established INT8 precision introduces a massive amount of temporal noises to the rendered avatar, even with the SoTA methods. As shown in Figure 1, the facial motion generated by the 8-bit decoder is filled with temporally flickering noises between two consecutive frames, which immerses the original facial expressions and degrades the quality of the rendered avatar. Most importantly, the temporal noise is not visible in the static single frame, distorting the avatar in a continuous manner. As the first work in this area, this paper reveals the root cause of the temporal avatar noise, as described in Section 3. + +Difficulty of characterizing temporal noise. Common quantization noises in conventional classification and object detection tasks can be modeled by a standard statistical model [12]. However, the temporal avatar noises introduced by quantized decoder exhibit inconsistent intensities and granularities in different facial regions. As a result, it is almost impossible to model the temporal noises globally without additional recurrent information. Furthermore, applying an additional filtering strategy after quantization introduces additional energy and latency consumption to the AR/VR devices, which deteriorates the efficiency of Codec Avatar telecommunication. Therefore, the noise suppression process should be achieved all at once during post training quantization. + +To resolve the aforementioned problems, we propose POCA, a novel Post Training Quantization (PTQ) algorithm designed for high quality and temporally noise-free Codec Avatar decoding with outstanding simplicity and efficiency. The major contributions of this work are listed as follows: + +- As the first work that investigates the low-precision avatar decoding, we reveal the root cause of the temporal noises induced by the temporally dynamic and long-tailed distribution. +- POCA eliminates the temporal facial noise with 8-bit and 6-bit precision, outperforming the SoTA method with fully recovered avatar quality [16]. +- POCA only requires $\leq 10\%$ of layer-wise feature map data and $\leq 5\%$ of the total training data for PTQ. +- POCA is evaluated on a variety of avatar identities with a full range of viewing directions and dynamic expressions, exhibiting a high degree of robustness and consistent noise-free avatar animation across different genders, age groups, hairstyles, skin colors, and view directions. +- POCA is also applicable to image super resolution task [8,25], achieving the SoTA image quality with $80\%$ calibration sparsity. + +# 2 Related Work + +# 2.1 Codec Avatar + +Photorealistic human face modeling and animation are challenging computer vision and computer graphics tasks in the sense of achieving minimum visual + +distortion for reconstructed 3-D human faces. Early works mainly focus on accurate 3-D reconstruction with blend shape models [9], which are computationally expensive and impossible for mobile on-device telepresence. Empowered by the deep learning models, recent research [14,15] considers human face reconstruction as data-driven generative task with the variational auto-encoder neural network. In particular, the variational encoder [6] is trained for encoding the multi-view HMC images into the latent information $\mathbf{z}$ , which describes the state of the face with rich expression and motion. Subsequently, the decoder network reconstructs the 3-D faces based on the latent code and the given view direction $\mathbf{v}$ , where the assets for rendering 3D face avatar are decoded separately by the dedicated mesh decoder and texture decoder. + +The compact latent code is transmitted between VR devices with low communication bandwidth requirements. However, the desire for real-time, on-device 3-D face avatar animation is hindered by the expensive computation cost and high memory consumption of the decoder. Various efforts have been established to improve the computation efficiency of the VAE model [5, 14]. Pixel Codec Avatar [15] proposes a novel architecture to allow lightweight decoder by creatively leveraging rendering pipeline features, skipping computations for invisible pixels while maintaining the high-fidelity for visible regions. Recent research optimizes the encoder architecture via neural architecture search (NAS) [3], leading to reduced computational costs. Different from such prior works, our work proposes a decoder-dedicated quantization algorithm that enables temporal alignments and noise-free avatar decoding with low-precision model, which further enhances instant and high quality on-device deployment of Codec Avatars for resource-constrained mobile AR/VR headsets. + +# 2.2 Quantization + +Quantization has been widely investigated together with the evolution of deep learning. Starting from the well-established 8-bit precision [4,24] all the way to the aggressive binary neural network [32], the quantized model largely improves the computation and memory efficiency, making the complex DNN models deployable to the resource-constrained edge devices (e.g., AR/VR headset). The core idea of quantization is constructing the accurate matching between the high-precision data (e.g., FP32) and the low-precision representation. Normally, the quantization process can be generalized into the "quantize-and-dequantize" workflow connected by the shared scaling factor in between. Mathematically, given the full precision weights $\mathcal{W}^l$ and input activation $\mathcal{X}^l$ of a layer $l$ , we have: + +$$ +\mathcal {W} _ {Q} ^ {l} = \left\lfloor \frac {\mathcal {W} ^ {l}}{S _ {w} ^ {l}} \right] \in \operatorname {I N T} _ {\mathbf {n}}, \quad \mathcal {X} _ {Q} ^ {l} = \left\lfloor \frac {\mathcal {X} ^ {l}}{S _ {x} ^ {l}} \right] \in \operatorname {I N T} _ {\mathbf {n}} \tag {1} +$$ + +Where $S$ is the scaling factor determined by data precision $\mathbf{n}$ and the numerical range. Specifically, $S$ characterizes the "interval" between adjacent levels in the floating point dequantized values, bounded by the range [lb, ub]: + +$$ +S = | \mathrm {u b} - \mathrm {l b} | / 2 ^ {\mathbf {n}} - 1 \tag {2} +$$ + +![](images/fb71659a89c7e4a15b921ebd294a5c266747bcdaa27d6bf348d6432f958aecf8.jpg) +Fig. 2: Top: The rendered avatar exhibits minimal quality (PSNR) reduction in static frames and large distortion with animation. Bottom: SoTA 8-bit quantization [2,18] amplifies the pixel shifting between frames in the rendered avatar across RGB channels. + +Since the number of levels is directly determined by the precision $\mathbf{n}$ , the essence of quantization accurately approximates the floating-point distribution with the dequantized tensors while minimizing the model-level error. To that end, a variety of quantization algorithms have been proposed to locate the optimal scaling factors with QAT [13] or PTQ [11, 18, 28], including boundary learning [1, 10], scaling factor learning [2], and quantization error modeling [12], etc. Recent efforts introduce additional scaling to avoid the error caused by the extreme distribution [30]. Promoted by the fast compression process, PTQ exhibits high compatibility for quantizing the avatar VAE model compared to QAT. Ideally, compressing the Codec Avatar models with PTQ would lead to instant "user-avatar" conversion for AR/VR devices, compared to time-consuming QAT. Recent VAE-based quantization focuses on the video compression task with quantized latent code [7, 21], while the quantization of the codec-avatar decoder model has not been explored despite the urgent needs on edge device and high rendering quality. + +# 3 Root Cause of Temporal Avatar Noise + +# 3.1 Temporal Noise is Dynamic Pixel Color Oscillation + +Recent 3D rendering tasks (e.g., NeRF) [17] uses peak signal-to-noise ratio (PSNR) to evaluate the quality between the rendered output and the ground truth im + +age. However, the temporal distortion of Codec Avatar cannot be effectively quantified with the same metrics. + +With the low-precision decoding, the temporal noise shows minimal impact on the static frame quality, but largely distorts the dynamic motion of the avatar, as shown in Figure 2 and supplementary material. Prior works only focus on the static image quality [25] with low-precision models, but the dynamic quality reduction has not been investigated well. + +In addition to the FovVideoVDP score [16], the temporal pixel oscillation can be characterized by the mean pixel value differences between two consecutive frames $F_{t}$ and $F_{t + 1}$ of the rendered avatar: + +$$ +\Delta_ {F} ^ {c} = \frac {1}{\mathcal {H W}} \times | F _ {t + 1} - F _ {t} | \tag {3} +$$ + +Where $\mathcal{H}$ and $\mathcal{W}$ represent the height and width of the rendered frame. We evaluate the pixel value differences separately for RGB channels such that $c \in \{R, G, B\}$ . The trend of the pixel changes characterizes the basic transition of the facial expression (e.g., Tongue in and out), as shown in Figure 2 (Bottom). + +With the 8-bit VAE decoder, the frame-wise pixel dynamics exhibit a similar trend as the full-precision decoder baseline. However, the quantized decoder adds an additional "offset" to the transition of the facial motion. In other words, the additional distortion is introduced to the avatar on top of the natural motion, which leads to random texture fluctuation in the rendered Codec Avatar. Figure 2 (Bottom) shows that the temporal noises uniformly exist across the entire facial motion with low intensity of $\pm 0.44$ , $\pm 0.47$ , and $\pm 0.25$ average pixel shifting on the R, G, B channels, respectively. We are fully aware of the existence of other motion capture matrices (e.g., optical flow), but the temporal noises are significantly fine-grained, which requires detailed pixel-level analysis. + +Formally, such random pixel flickering can be captured by the FovVideoVDP [16] score between the rendered avatar and ground truth image. As shown in Figure 2, the INT8 avatar decoder largely degrades the animation quality of the avatar, even with the SoTA 8-bit quantization method [2, 18]. The high density of the yellow region indicates the strongly mismatched animation caused by the temporal noise. The continuous pixel oscillation and temporal noises worsen the fidelity of the rendered Codec Avatar in motion. + +# 3.2 Root Cause of Noisy Avatars + +Since the temporal noises are induced by the pixel oscillation of the avatar in motion, an intuitive understanding of the noises is the "unstable rounding" during quantization [12, 19]. However, the rounding instability is still the phenomenon of quantization instead of the root cause of the noise. In this work, we unravel the fundamental reason of the temporal noise induced by low-precision avatar decoding. + +Dynamically long-tailed distribution of human face. Human facial features naturally display a diverse range of distributions characterized by dif + +![](images/7edb81847558169fa6b09c1dff05d88de636549afab9f57ace19fd39adf3339a.jpg) +(a) + +![](images/f0c263706cb0acc9a3dd87cc5cec5cf3c3d880209167bea05464059504959cde.jpg) +(b) +Fig. 3: (a) Row-channel plane of the texture decoder activation and (b) Regional standard deviation of the row-channel planes with respect to different facial expressions. + +ferent facial regions (for instance, hair compared to cheeks). The broad spectrum of variability in human facial characteristics formulates various statistical properties. For the avatar decoding, the texture decoder of the VAE model [14] decodes face avatar texture from the latent space via transposed convolutional layers, which consecutively up-scales the activation feature map from $4 \times 4$ to $1024 \times 1024$ (height and width). Therefore, the statistical variation of the facial texture is also embedded into the activation of the earlier layer. + +Proof of facial dynamics. We prove the dynamic range variations of the texture by analyzing the regional distribution of the latent layer activation. Assume $\mathcal{C}$ , $\mathcal{H}$ , and $\mathcal{W}$ represent the channel depth, height, and width of the activation tensor. We analyze the statistics by decomposing the entire face texture feature map into multiple "row-channel planes". Each plane contains the regional statistics from hair to jaw, with the total number of planes = $\mathcal{H}$ . + +Figure 3(a) shows the row-channel decomposition based on the intermediate activation feature map of the texture decoder with the shape of $16 \times 256 \times 256$ . On top of that, we compute the standard deviation of each plane, which characterizes the local tailedness and dynamic range in each region of the face, with respect to different facial expressions. As depicted in Figure 3(b), smiling and eye blinking lead to the long-tailed dynamics in the mouth and eye regions, respectively. As a result, the distribution of the entire face is dynamically long-tailed, which is harmful for quantization. In other words: + +The distribution of a decoder feature map contains both constrained and long-tailed distribution, while the tailedness varies between different facial expressions. + +In other words, the root cause of the temporal noises is the conflict between the long-tailed, highly dynamic distribution of activations and the global error minimization during PTQ. To minimize the global quantization error, the long-tailed distribution stretches the quantization interval (scaling factor) in order to cover larger range, which in turn, elevates the local rounding instability and leads to unstable numerical representation of the pixels. We prove this hypothesis in Section 4.3 with the proposed method. + +# 4 Proposed Method: POCA + +To resolve the temporal noises caused by low-precision avatar decoding, this section presents POCA, a novel post-training quantization algorithm that successfully compresses the deep appearance decoder and eliminates the temporal noises from the rendered Codec Avatar. In particular, POCA is empowered by two novel techniques: a) View-aware PTQ and b) PTQ with shape-wise filtering. + +# 4.1 View-aware PTQ + +As pointed out in Section 3, the face texture contains both region-dependent dispersion and facial expression-driven dynamics in the entire face. As a result, the sample space of the training data and layer-wise activation contains high variation with redundant information. In particular, the texture decoder is trained by the whole face texture map with the size of $1024 \times 1024$ . However, the final 3D avatar is rendered based on the visibility information from the given view direction. Therefore, quantizing the decoder without considering the visibility for inner layers leads to the incorrectly learned quantization parameter, which further generates the mismatched quantization decision between PTQ and actual rendering. Motivated by that, we propose the view-aware PTQ for the avatar texture decoder. + +Basic Setup. Given the pre-trained full-precision VAE decoder (texture and mesh), we first randomly select $m$ training samples from the full-sized dataset with size $\mathcal{D}$ ( $m << \mathcal{D}$ ). Each training sample contains the input and pre-computed output of the decoder ( $[z_t, v_t], [\hat{\mathbf{T}}_t, \hat{\mathbf{M}}_t]$ ), where $z_t, v_t, \hat{\mathbf{T}}_t$ , and $\hat{\mathbf{M}}_t$ represents the encoded latent code, view direction, decoded texture, and decoded mesh for sample $t$ , respectively. Following the standard PTQ strategy, we extract input $\mathcal{X}^l$ and output activation $\mathcal{Y}^l$ of each layer as the calibration data. We adopted AdaRound [18] and LSQ [2] as the weight and activation quantizers, respectively. The learnable quantization parameters ( $\alpha, S^l$ ) are optimized by minimizing the mean square error (MSE) between the quantized layer output $\hat{\mathcal{Y}}^l$ and full precision output $\mathcal{Y}^l$ : + +$$ +\underset {\alpha , S ^ {l}} {\arg \min } \frac {1}{\mathcal {H W C}} \sum \left(\mathcal {Y} ^ {l} - \hat {\mathcal {Y}} ^ {l}\right) ^ {2} \tag {4} +$$ + +View-dependent visibility mask. The computation of visibility is inherited from the rendering pipeline utilizing the pre-trained VAE model. Specifically, the decoded mesh vertices $\hat{\mathbf{M}}_t$ are rasterized to formulate the barycentric coordinates, which are leveraged to interpolate the UV coordinates $\varPhi$ across the entire avatar. Following this, the visibility is determined by sampling the view-specific pixels from the 2-D texture map based on the interpolated UV indexes, $\varPhi = [\phi_x,\phi_y]$ $(0\leq \phi_{x},\phi_{y}\leq 1)$ . + +The value of $\phi_x$ and $\phi_y$ represents the normalized coordinates of the 2-D texture map with the shape of $\mathcal{H} \times \mathcal{W}$ . Different from the default pipeline, the + +![](images/5f091620bbd48526a320aaa752ccb47cd4a8c80de29fcecb15789ba85aea0f78.jpg) +Fig. 4: Overall PTQ process of POCA based on layer-wise calibration (Layer Calib). POCA fully quantizes all the layers in the VAE decoder, where the proposed view-dependent $+$ SWD filter is applied to the decoder with transpose convolution layers. + +We proposed POCA algorithm generates the binary "visibility mask" $\Gamma$ to select the pixels in the 2-D texture map $\hat{\mathbf{T}}_t$ . Formally, we have: + +$$ +\Phi = \text {I n t e r p o l a t e} (\text {R a s t e r i z e} (\mathbf {M} _ {t}, \dots), u v) \tag {5} +$$ + +$$ +\Gamma = \mathbb {1} \cdot \left\{\left(\gamma_ {x}, \gamma_ {y}\right) \mid \gamma_ {x} \in \left\{\phi_ {x} \times \mathcal {H} \right\}, \gamma_ {y} \in \left\{\phi_ {y} \times \mathcal {W} \right\} \right\} \tag {6} +$$ + +Where $\mathbb{1}$ represents the indicator function that signifies whether the pixel at $(\gamma_x,\gamma_y)$ should be active at the given visibility. $\{\phi_x\times \mathcal{H}\}$ and $\{\phi_y\times \mathcal{W}\}$ are the visibility coordinates (matrix indexes) scaled based on the interpolated UV coordinates $\varPhi$ with the dimensionality of the texture map. + +The visibility masks are generated during the data-fetching process. Since the model is already pre-trained, the visibility masks $\varGamma$ are accessible with each full-precision texture $\hat{\mathbf{T}}_t$ and triangle index $uv$ at the output layer of the texture decoder. The final layer $\varGamma$ ( $1024 \times 1024$ ) will be max-pooled to formulate the visibility mask for the earlier layer. For each layer of the texture decoder, POCA performs PTQ with the visible pixel only: + +$$ +\underset {\alpha^ {l}, S ^ {l}} {\arg \min } \frac {1}{\mathcal {H W C}} \sum \left(\mathcal {Y} _ {\mathbf {T}} ^ {l} \odot \Gamma^ {l} - \hat {\mathcal {Y}} _ {\mathbf {T}} ^ {l} \odot \Gamma^ {l}\right) ^ {2} \tag {7} +$$ + +Where $\odot$ represents the element-wise product between tensors. Subsequently, the learnable quantization parameters $\alpha^l$ and $S^l$ are trained based on the gradient that resulted from the view-dependent visibility $\Gamma$ . As a result, POCA optimizes the quantization decision based on the most relevant information of the training sample, instead of being diversified by the widely dispersed texture space with long-tailed and dynamic distribution. Figure 4 depicts the proposed view-dependent PTQ. Although the quantizer is calibrated based on the visible pixels, POCA only runs PTQ once, and the resultant model will be deployed and rendered for all views. + +# 4.2 PTQ with Shape-wise Dynamic Filtering + +Besides the PTQ with the view-dependent visibility masking, we further suppress the impact of the dynamically long-tailed distribution by eliminating the pixels with a high channel-wise dynamic range, or "Shape-Wise Dynamics" (SWD). Given the full-precision output $\mathcal{V}_{\mathbf{T}}^{l}$ with the shape of $\mathcal{C} \times \mathcal{H} \times \mathcal{W}$ , the SWD is characterized by the standard deviation score of each $\mathcal{C} \times 1 \times 1$ vector, which formulates the overall texture SWD map with the shape of $\mathcal{H} \times \mathcal{W}$ . + +Subsequently, POCA generates the SWD mask $\Theta$ by filtering out the pixels in the SWD map with the highest tailedness. The sparsity of the SWD mask is controlled by a tunable hyperparameter $\tau$ . E.g., $\tau = 0.7$ means masking out the top $70\%$ pixels with the highest tailedness (standard deviation). The resultant SWD mask $\Theta$ will be multiplied with the view-dependent mask $\Gamma$ and generate the final learning objective of the layer-wise post-training quantization of POCA: + +$$ +\underset {\alpha^ {l}, S ^ {l}} {\arg \min } \frac {1}{\mathcal {H W C}} \sum \left(\mathcal {Y} _ {\mathbf {T}} ^ {l} \odot \Gamma^ {l} \odot \Theta^ {l} - \hat {\mathcal {Y}} _ {\mathbf {T}} ^ {l} \odot \Gamma^ {l} \odot \Theta^ {l}\right) ^ {2} \tag {8} +$$ + +As shown in Eq. (8), POCA optimizes the quantizers by learning the most relevant information (view-dependent) and least diversified pixels (SWD masking) of the avatar, which suppresses the drawbacks of the long-tailedness and high dynamics of the avatar. + +The masking strategy of POCA introduces a high degree of sparsity in layer-wise PTQ, controlled by visibility and the SWD threshold $\tau$ . Our experimental results in Section 5 show that the optimal performance is achieved with $\tau = 0.9$ . As a result, POCA calibrates the low-precision layer based on highly sparse data $[\mathcal{X}^l,\mathcal{Y}_{\mathbf{T}}^l]$ with $\geq 90\%$ sparsity. In other words, POCA only requires $\leq 10\%$ of the activation for layer-wise PTQ, showing low memory overhead for the efficient PTQ calibration process. + +# 4.3 Rationale behind POCA + +POCA alleviates the impact of rounding instability. With the proposed view-aware PTQ and SWD masking, POCA aims to optimize the layerwise quantizers with the most relevant and quantization-friendly features. The POCA-trained activation quantizer (LSQ) has a generally smaller scaling factor $S^l$ compared to the PTQ without POCA, as shown in Figure 5. The reduced scaling factors formulate the low-precision grid with small intervals (scaling factor) between adjacent levels, alleviating the numerical differences between adjacent levels in the dequantized high-precision domain. As a result, the impact of the aforementioned rounding instability is minimized and the temporal noises become invisible, which proves the aforementioned hypothesis in Section 3.2. + +Sparse masking $\neq$ biased learning. Although two types of sparse masking are introduced into PTQ, the POCA-trained quantizers will not lead to a biased representation in the quantized model. As shown in Figure 4, the sparsified calibration mask $(\varGamma^l\odot\varTheta^l)$ contains sparse but uniformly sampled information + +![](images/3cc20d1972f613dc437500148f413b5b2e4e3394fd117e438146aaeb86142734.jpg) +Fig. 5: Scaling factor $S^l$ of the texture decoder trained by POCA. + +across all the regions of the avatar texture. As a result, the learnable quantizers are uniformly trained with all the relevant facial information. + +Highly-correlated data and high fidelity sparse learning. Unlike classification tasks, the training samples of the VAE model are highly correlated with respect to the single identity. As a result, the sparsified information of each sample can be compensated from other training data with different viewing perspectives, leading to sparse but robust PTQ with the quantized decoder. + +# 5 Experimental Results + +We evaluated the proposed POCA algorithm based on Multiface [29], a public avatar dataset captured by the multi-camera system. POCA is applied to eight avatars among three viewing directions with all 65 facial expressions of the dataset. We want to highlight that the eight avatars comprehensively cover different genders, age groups, hairstyles, skin types, and facial hairs. + +For each identity, we use the corresponding pre-trained deep appearance VAE model as the starting point of POCA. Following the standard PTQ strategy, we quantize all layers of the decoders (texture and mesh) iteratively, where each layer is calibrated by 100 epochs with the sampled data. POCA only acquires 512 training samples from each identity, which is $\leq 5\%$ samples of the dataset. For each layer, the quantizers are optimized by Adam with 1e-4 learning rate and eight samples per batch. + +Prior Works for Comparison. As the first work in the low-precision avatar decoding, we select three prior quantization works that reported the SoTA INT8 performance with the classification tasks: (1) Adaptive Rounding (AdaRound) [18], (2) learned step size quantization (LSQ) [2], which are also incorporated for hardware deployment [24], and (3) W8A8 naive QAT [20] in PyTorch. These selected prior works achieved superior INT8 performance [2, 18] on the ImageNet-1K dataset, even outperforming the full-precision baseline. + +Evaluation Metric. We evaluate the proposed POCA algorithm based on the FovVideoVDP [16] between the rendered avatar and the ground truth video to quantify the temporal noise of the dynamic Codec Avatar. For each avatar identity, we compute the averaged FovVideoVDP score among all the 65 expressions in the Multiface dataset [29]. The performance of POCA is demonstrated with the rendered avatar videos in the supplementary material. + +Table 1: FovVideoVDP score [16] between the rendered avatar and the ground truth based on eight avatars across all the expressions from three viewing perspectives. The proposed POCA algorithm achieves the highest quality compared to the SoTA PTQ and QAT methods with both 8-bit and 6-bit precision. + +
Avatar IDPrecision26438142183941537202126453108870559667444329145896795937
Male, GoateeBlack, FemaleMohawk, MaleAsian, FemaleOlder WomanMale, Mid HairBlonde GirlMale, Short Hair
FrontTopFrontTopFrontTopFrontTopFrontTopFrontTopFrontTopFrontTop
Baseline [29]FP326.214.945.306.544.995.456.144.875.406.325.025.596.054.955.136.02
SoTA PTQ [2,18]W8A85.814.434.875.914.405.005.804.525.195.874.565.105.434.444.685.52
W8A8 QAT [20]5.924.554.996.024.575.125.884.565.235.974.665.165.574.504.795.60
POCA6.194.935.266.314.805.386.094.835.336.214.955.425.954.825.085.99
SoTA PTQ [2,18]W6A64.503.664.064.963.814.124.554.104.475.223.914.454.293.733.964.25
W6A6 QAT5.083.754.395.544.014.234.944.194.495.364.074.664.363.854.094.46
POCA6.064.555.026.294.425.136.014.615.166.024.665.215.454.494.995.66
+ +![](images/dfc31f7dfbe95a7a21ecbb8634d9eeb7138c6e23f09312d14c0d85952b626fe9.jpg) +Yellow region $=$ Mismatched pixel motion +FovVideoVDP [14] between the rendered avatar and ground truth (Max = 10.0) +Fig. 6: Rendered avatars and temporal pixel differences between two consecutive frames of the avatar video. The lightened yellow region of POCA signifies the removed temporal noise as the full precision baseline. + +# 5.1 Temporal Noise Reduction with High Fidelity + +We evaluate the proposed POCA algorithm with eight avatars with different genders, hairstyles, and skin colors. Table 1 summarizes the FovVideoVDP score [16] of the proposed POCA algorithm with 8-bit and 6-bit avatar decoders. The SoTA INT8 PTQ quantization algorithm failed to preserve the avatar quality as the baseline full precision model, exhibiting the degraded FovVideoVDP score. Even with the expensive quantization-aware training (QAT), the quality of the rendered avatar cannot be restored. + +Compared to the previous methods with 8-bit precision, POCA eliminates the temporal noises in the rendered avatar. POCA successfully compresses the decoder model while achieving on-par quality as the full precision decoder baseline. Figure 6 shows the high avatar quality and restored FovVideoVDP score (heat map) of the proposed POCA algorithm (closer to the FP32 baseline = better quantization). Regarding the efficiency of PTQ, POCA only requires $\leq 10\%$ of the activations for layer-wise PTQ, showing low memory overhead for the efficient PTQ calibration process. + +![](images/07e345d143445ab9dc251b3dfb8a8c907ed40d885528d5bf819d149708673aa7.jpg) +SSIM=0.741 +FovVideo $\mathrm{DVP}=6.213$ + +![](images/4f149d2510c6863f4b7bf29d333575109256d8111bcaca2469a9a0109e6e185d.jpg) +SSIM=0.741 +FovVideoDVP=6.119 + +![](images/b09af302b5c3cae5805f4af93690fbc0b60e4a438927578e8106ae8f58a30cd0.jpg) +SSIM=0.741 +FovVideoVDP=1.685 +Fig. 7: The impact of the SWD threshold. Extreme filtering leads to distorted and missing texture in the resultant avatar. + +![](images/67276d87212373da8dc64a0adf798dfbc307d6ef36804bdbfda6b1889cec60bf.jpg) +SSIM=0.740 +FovVideoVDP=6.121 + +![](images/be98ce8fa9a901c66f1f0e6b976be335103a3f511e8e23d0f714b5a20bd6fab4.jpg) +SSIM=0.739 +FoxVideoVDP=6.106 + +Table 2: Training time comparison between POCA and QAT. + +
MethodQuantizationMethodW/ATimeGPUDecoder Size (MB)
FP32 Baseline-32/32---32.78
Native W8A8 [20]QAT8/8>10 hrs2 × A60008.19 (↓ 4×)
POCA (This work)PTQ8/8~13 min1 × A60008.19 (↓ 4×)
POCA (This work)PTQ6/6~13 min1 × A60006.15 (↓ 5.3×)
+ +POCA generates clean avatars without introducing any graphic filtering techniques [31]. From the perspective of quantization, POCA keeps the vanilla quantization strategy in Eq. (1) in both training and inference, compressing the model without introducing distribution manipulations [28] on the additional scaling [30] and bias [12]. Compared to traditional QAT, POCA quantizes the entire model in only 13 minutes (Table 2). The fast deployment workflow with the actual VR device (e.g., Quest Pro) can further accelerate the process of "user-to-meta verse". + +Future Applicability: POCA is not limited to the deep appearance model [14]. The proposed POCA algorithm is applicable to all mesh-based 3D reconstruction tasks with VAE models. For example, the proposed view-dependent visibility mask and shape-wise filtering can be easily transferred to the recent Pixel Codec Avatar (PiCA)[15], which has the similar autoencoder-based architecture. + +# 5.2 Impact of the SWD Filtering Threshold + +POCA introduced the tunable parameter $\tau$ in Eq. (8) to control the amount of activation pixels for layer-wise PTQ. Figure 7 demonstrates the impact of the SWD threshold $\tau$ with respect to the avatar quality (SSIM). POCA achieves the best tradeoff between filter intensity and avatar quality with $\tau^{*} = 0.9$ . Quantizing the decoder with $\tau > \tau^{*}$ leads to missing texture pixels and distorted rendering with no quality improvement. The selected SWD threshold exhibits low sensitivity between different avatars and viewing perspectives, as demonstrated by the high-quality rendering in Figure 6. + +# 5.3 Applicability to Other Quality Enhancement Tasks + +We extend the experiment by evaluating the proposed method on super-resolution tasks. The proposed "shape-wise-filtering" (SWD) exhibits strong performance in image quality enhancement [25] with the low-precision SRResNet model [8]. + +Table 3: SoTA image quality (PSNR/SSIM) achieved by the proposed POCA-SWD algorithm with 8-bit and 6-bit model precision, outperforms the recent SR-PTQ [25]. + +
MethodPrecisionSparsitySet5 (× 4)Set14 (× 4)BSD100 (× 4)Urban100 (× 4)
SRResNet Baseline [25]FP32(Baseline)N/A32.234 / 0.89628.656 / 0.78427.63 / 0.73826.229 / 0.791
SRResNet [8]32.05 / 0.90128.49 / 0.81827.58 / 0.762Not Reported
SRResNet (This work)31.927 / 0.90128.588 / 0.79927.628 / 0.75626.223 / 0.800
AdaRound [18] + LSQ [2]W8A8Dense31.870 / 0.89928.550 / 0.79827.533 / 0.75426.078 / 0.797
SR-PTQ [25]Dense32.207 / 0.89528.619 / 0.78327.618 / 0.73826.191 / 0.790
POCA80%31.898 / 0.90028.566 / 0.79827.620 / 0.75626.195 / 0.799
Quality Drop [2,18]W8A80.057 / 0.0020.038 / 0.0010.095 / 0.0020.145 / 0.003
Quality Drop [25]0.027 / 0.0010.037 / 0.0010.012 / 0.0000.038 / 0.001
Quality Drop (POCA)0.029 / 0.0010.022 / 0.0010.008 / 0.0000.028 / 0.001
AdaRound [18] + LSQ [2]W6A6Dense31.765 / 0.89728.449 / 0.79527.470 / 0.75225.988 / 0.794
SR-PTQ [25]Dense32.089 / 0.89228.504 / 0.77927.561 / 0.73326.011 / 0.783
POCA80%31.789 / 0.89928.499 / 0.79627.565 / 0.75326.110 / 0.795
Quality Drop [2,18]W6A60.162 / 0.0040.139 / 0.0030.158 / 0.0040.235 / 0.006
Quality Drop [25]0.145 / 0.0040.152 / 0.0050.069 / 0.0050.218 / 0.008
Quality Drop (POCA)0.138 / 0.0020.089 / 0.0040.061 / 0.0030.113 / 0.005
+ +With 8-bit and 6-bit precision, the proposed method outperforms [25] in BSD100 and Urban100 datasets, as shown in Table 3. For a fair comparison, we listed all the baseline full-precision models. Our method outperforms [25] on all datasets with the new SoTA image quality. Compared to [18] and [2], the improved image quality implies the effectiveness of the proposed SWD. More importantly, our method enables the sparse PTQ by using only $20\%$ pixels of the selected samples, whereas all prior works [2,18,25] require dense PTQ. + +# 6 Conclusion and Discussion + +In this paper, we propose a novel post-training quantization algorithm designed for high-quality Codec Avatar rendering. As the first work in this area, we discover the challenge of temporal noise with low-precision avatar decoding. Through thorough investigation, we reveal the root cause of the temporal noise that originated from the high dynamics and long-tailedness of the texture distribution. To address that, we propose POCA, which consists of view-aware PTQ and shape-wise filtering (SWD). Compared to the previous SoTA methods, POCA eliminates the temporal noises even with 6-bit precision. Furthermore, the proposed SWD algorithm is also applicable to super-resolution tasks, achieving the new SoTA image quality with highly efficient sparse calibration. + +Limitations. The proposed POCA algorithm mainly focuses on the VAE-based Codec Avatar generation, which has been proven as a successful strategy with the actual commercial-level VR product deployment. Recently, 3D Gaussian-based avatar rendering also emerged as a feasible method for generative Avatars, while low-precision quantization of such models will remain as future work. + +# References + +1. Choi, J., Venkataramani, S., Srinivasan, V.V., Gopalakrishnan, K., Wang, Z., Chuang, P.: Accurate and efficient 2-bit quantized neural networks. Proceedings of Machine Learning and Systems 1, 348-359 (2019) 5 +2. Esser, S.K., McKinstry, J.L., Bablani, D., Appuswamy, R., Modha, D.S.: Learned step size quantization. In: International Conference on Learning Representations (ICLR) (2020) 2, 3, 5, 6, 8, 11, 12, 14 +3. Fu, Y., Li, Y., Li, C., Saragih, J., Zhang, P., Dai, X., Lin, Y.C.: Auto-card: Efficient and robust codec avatar driving for real-time mobile telepresence. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR) (2023) 4 +4. Gorbachev, Y., Fedorov, M., Slavutin, I., Tugarev, A., Fatekhov, M., Tarkan, Y.: Openvino deep learning workbench: Comprehensive analysis and tuning of neural networks inference. In: Proceedings of the IEEE/CVF International Conference on Computer Vision Workshops (CVPR-W) (2019) 4 +5. Han, Y., Yang, S., Wang, W., Liu, J.: From design draft to real attire: Unaligned fashion image translation. In: Proceedings of the 28th ACM International Conference on Multimedia. p. 1533-1541 (2020) 4 +6. Kingma, D.P., Welling, M.: Auto-encoding variational bayes. arXiv preprint arXiv:1312.6114 (2013) 4 +7. Le, H., Zhang, L., Said, A., Sautiere, G., Yang, Y., Shrestha, P., Yin, F., Pourreza, R., Wiggers, A.: Mobilecodec: neural inter-frame video compression on mobile devices. In: Proceedings of the 13th ACM Multimedia Systems Conference. pp. 324-330 (2022) 5 +8. Ledig, C., Theis, L., Huszár, F., Caballero, J., Cunningham, A., Acosta, A., Aitken, A., Tejani, A., Totz, J., Wang, Z., et al.: Photo-realistic single image superresolution using a generative adversarial network. In: Proceedings of the IEEE conference on computer vision and pattern recognition. pp. 4681-4690 (2017) 3, 14 +9. Lewis, J.P., Anjyo, K., Rhee, T., Zhang, M., Pighin, F.H., Deng, Z.: Practice and theory of blendshape facial models. Eurographics (State of the Art Reports) 1(8), 2 (2014) 4 +0. Li, Y., Dong, X., Wang, W.: Additive powers-of-two quantization: An efficient non-uniform discretization for neural networks. In: International Conference on Learning Representations (ICLR) (2020) 5 +1. Li, Y., Gong, R., Tan, X., Yang, Y., Hu, P., Zhang, Q., Yu, F., Wang, W., Gu, S.: {BRECQ}: Pushing the limit of post-training quantization by block reconstruction. In: International Conference on Learning Representations (ICLR) (2021) 2, 5 +2. Liu, Y., Yang, H., Dong, Z., Keutzer, K., Du, L., Zhang, S.: Noisyquant: Noisy bias-enhanced post-training activation quantization for vision transformers. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR) (2023) 3, 5, 6, 13 +3. Liu, Z., Cheng, K.T., Huang, D., Xing, E.P., Shen, Z.: Nonuniform-to-uniform quantization: Towards accurate quantization via generalized straight-through estimation. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR) (2022) 5 +4. Lombardi, S., Saragih, J., Simon, T., Sheikh, Y.: Deep appearance models for face rendering. ACM Transactions on Graphics (ToG) (2018) 1, 2, 4, 7, 13 + +15. Ma, S., Simon, T., Saragih, J., Wang, D., Li, Y., De La Torre, F., Sheikh, Y.: Pixel codec avatars. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR) (2021) 1, 2, 4, 13 +16. Mantiuk, R.K., Denes, G., Shapiro, A., Kaplanyan, A., Rufo, G., Bachy, R., Lian, T., Patney, A.: FovVideoVDP: A visible difference predictor for wide field-of-view video. In: ACM Transactions on Graphics (TOG). ACM New York, NY, USA (2021) 3, 6, 11, 12 +17. Mildenhall, B., Srinivasan, P.P., Tancik, M., Barron, J.T., Ramamoorthi, R., Ng, R.: NeRF: Representing scenes as neural radiance fields for view synthesis. Communications of the ACM 65(1), 99-106 (2021) 5 +18. Nagel, M., Amjad, R.A., Van Baalen, M., Louizos, C., Blankevoort, T.: Up or down? adaptive rounding for post-training quantization. In: International Conference on Machine Learning (ICML) (2020) 2, 3, 5, 6, 8, 11, 12, 14 +19. Park, E., Yoo, S.: Profit: A novel training method for sub-4-bit mobilenet models. In: European Conference on Computer Vision (ECCV) (2020) 6 +20. Paszke, A., Gross, S., Massa, F., Lerer, A., Bradbury, J., Chanan, G., Killeen, T., Lin, Z., Gimelshein, N., Antiga, L., Desmaison, A., Kopf, A., Yang, E., DeVito, Z., Raison, M., Tejani, A., Chilamkurthy, S., Steiner, B., Fang, L., Bai, J., Chintala, S.: Pytorch: An imperative style, high-performance deep learning library. In: Advances in Neural Information Processing Systems (NeurIPS) (2019) 3, 11, 12, 13 +21. van Rozendaal, T., Singhal, T., Le, H., Sautiere, G., Said, A., Buska, K., Raha, A., Kalatzis, D., Mehta, H., Mayer, F., et al.: Mobilenvc: Real-time 1080p neural video compression on a mobile device. In: Proceedings of the IEEE/CVF Winter Conference on Applications of Computer Vision. pp. 4323-4333 (2024) 5 +22. Saito, S., Schwartz, G., Simon, T., Li, J., Nam, G.: Relightable gaussian codec avatars. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 130-141 (2024) 2 +23. Schwartz, G., Wei, S.E., Wang, T.L., Lombardi, S., Simon, T., Saragih, J., Sheikh, Y.: The eyes have it: An integrated eye and face model for photorealistic facial animation. ACM Transactions on Graphics (TOG) (2020) 1 +24. Siddegowda, S., Fournarakis, M., Nagel, M., Blankevoort, T., Patel, C., Khobare, A.: Neural network quantization with AI model efficiency toolkit (AIMET). arXiv preprint arXiv:2201.08442 (2022) 2, 3, 4, 11 +25. Tu, Z., Hu, J., Chen, H., Wang, Y.: Toward accurate post-training quantization for image super resolution. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR) (2023) 2, 3, 6, 14 +26. Wang, Z., Wu, Z., Lu, J., Zhou, J.: Bidet: An efficient binarized object detector. In: Proceedings of the IEEE/CVF conference on computer vision and pattern recognition (CVPR) (2020) 2 +27. Wei, S.E., Saragih, J., Simon, T., Harley, A.W., Lombardi, S., Perdoch, M., Hypes, A., Wang, D., Badino, H., Sheikh, Y.: VR facial animation via multiview image translation. ACM Transactions on Graphics (TOG) (2019) 1 +28. Wei, X., Gong, R., Li, Y., Liu, X., Yu, F.: QDrop: Randomly dropping quantization for extremely low-bit post-training quantization. In: International Conference on Learning Representations (ICLR) (2022) 2, 5, 13 +29. Wuu, C.h., Zheng, N., Ardisson, S., Bali, R., Belko, D., Brockmeyer, E., Evans, L., Godisart, T., Ha, H., Huang, X., Hypes, A., Koska, T., Krenn, S., Lombardi, S., Luo, X., McPhail, K., Millerschoen, L., Perdoch, M., Pitts, M., Richard, A., Saragih, J., Saragih, J., Shiratori, T., Simon, T., Stewart, M., Trimble, A., Weng, X., Whitewolf, D., Wu, C., Yu, S.I., Sheikh, Y.: Multiface: A dataset for neural + +face rendering. In: arXiv (2022). https://doi.org/10.48550/ARXIV.2207.11243, https://arxiv.org/abs/2207.11243 1, 11, 12 +30. Xiao, G., Lin, J., Seznek, M., Wu, H., Demouth, J., Han, S.: Smoothquant: Accurate and efficient post-training quantization for large language models. In: International Conference on Machine Learning (ICML) (2023) 2, 5, 13 +31. Yin, H., Gong, Y., Qiu, G.: Side window filtering. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR) (2019) 13 +32. Zhang, Y., Zhang, Z., Lew, L.: Pokebnn: A binary pursuit of lightweight accuracy. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR) (2022) 4 \ No newline at end of file diff --git a/2024/POCA_ Post-training Quantization with Temporal Alignment for Codec Avatars/images.zip b/2024/POCA_ Post-training Quantization with Temporal Alignment for Codec Avatars/images.zip new file mode 100644 index 0000000000000000000000000000000000000000..96024e7778e81e38440586bf946b5f0bd9b658a2 --- /dev/null +++ b/2024/POCA_ Post-training Quantization with Temporal Alignment for Codec Avatars/images.zip @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:d20bed86b963ad5c432c56495ac1ef3935d9535e91fca55f4623dc722a12b0c0 +size 642592 diff --git a/2024/POCA_ Post-training Quantization with Temporal Alignment for Codec Avatars/layout.json b/2024/POCA_ Post-training Quantization with Temporal Alignment for Codec Avatars/layout.json new file mode 100644 index 0000000000000000000000000000000000000000..cf63b9944eaee792fe2d138674871958a0e02380 --- /dev/null +++ b/2024/POCA_ Post-training Quantization with Temporal Alignment for Codec Avatars/layout.json @@ -0,0 +1,8778 @@ +{ + "pdf_info": [ + { + "para_blocks": [ + { + "bbox": [ + 168, + 111, + 446, + 148 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 168, + 111, + 446, + 148 + ], + "spans": [ + { + "bbox": [ + 168, + 111, + 446, + 148 + ], + "type": "text", + "content": "POCA: Post-training Quantization with Temporal Alignment for Codec Avatars" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 137, + 167, + 477, + 193 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 137, + 167, + 477, + 193 + ], + "spans": [ + { + "bbox": [ + 137, + 167, + 477, + 193 + ], + "type": "text", + "content": "Jian Meng" + }, + { + "bbox": [ + 137, + 167, + 477, + 193 + ], + "type": "inline_equation", + "content": "^{1,2}" + }, + { + "bbox": [ + 137, + 167, + 477, + 193 + ], + "type": "text", + "content": ", Yuecheng Li" + }, + { + "bbox": [ + 137, + 167, + 477, + 193 + ], + "type": "inline_equation", + "content": "^{1}" + }, + { + "bbox": [ + 137, + 167, + 477, + 193 + ], + "type": "text", + "content": ", Chenghui (Leo) Li" + }, + { + "bbox": [ + 137, + 167, + 477, + 193 + ], + "type": "inline_equation", + "content": "^{1}" + }, + { + "bbox": [ + 137, + 167, + 477, + 193 + ], + "type": "text", + "content": ", Syed Shakib Sarwar" + }, + { + "bbox": [ + 137, + 167, + 477, + 193 + ], + "type": "inline_equation", + "content": "^{1}" + }, + { + "bbox": [ + 137, + 167, + 477, + 193 + ], + "type": "text", + "content": ", Dilin Wang" + }, + { + "bbox": [ + 137, + 167, + 477, + 193 + ], + "type": "inline_equation", + "content": "^{1}" + }, + { + "bbox": [ + 137, + 167, + 477, + 193 + ], + "type": "text", + "content": ", and Jae-sun Seo" + }, + { + "bbox": [ + 137, + 167, + 477, + 193 + ], + "type": "inline_equation", + "content": "^{2}" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 214, + 201, + 399, + 213 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 214, + 201, + 399, + 213 + ], + "spans": [ + { + "bbox": [ + 214, + 201, + 399, + 213 + ], + "type": "text", + "content": "1 Meta Codec Avatars Lab, Pittsburgh, USA" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 230, + 213, + 382, + 224 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 230, + 213, + 382, + 224 + ], + "spans": [ + { + "bbox": [ + 230, + 213, + 382, + 224 + ], + "type": "text", + "content": "2 Cornell Tech, New York City, USA" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 160, + 251, + 455, + 526 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 160, + 251, + 455, + 526 + ], + "spans": [ + { + "bbox": [ + 160, + 251, + 455, + 526 + ], + "type": "text", + "content": "Abstract. Real-time decoding generates high-quality assets for rendering photorealistic Codec Avatars for immersive social telepresence with AR/VR. However, high-quality avatar decoding incurs expensive computation and memory consumption, which necessitates the design of a decoder compression algorithm (e.g., quantization). Although quantization has been widely studied, the quantization of avatar decoders is an urgent yet under-explored need. Furthermore, the requirement of fast \"User-Avatar\" deployment prioritizes the post-training quantization (PTQ) over the time-consuming quantization-aware training (QAT). As the first work in this area, we reveal the sensitivity of the avatar decoding quality under low precision. In particular, the state-of-the-art (SoTA) QAT and PTQ algorithms introduce massive amounts of temporal noise to the rendered avatars, even with the well-established 8-bit precision. To resolve these issues, a novel PTQ algorithm is proposed for quantizing the avatar decoder with low-precision weights and activation (8-bit and 6-bit), without introducing temporal noise to the rendered avatar. Furthermore, the proposed method only needs " + }, + { + "bbox": [ + 160, + 251, + 455, + 526 + ], + "type": "inline_equation", + "content": "10\\%" + }, + { + "bbox": [ + 160, + 251, + 455, + 526 + ], + "type": "text", + "content": " of the activations of each layer to calibrate quantization parameters without any distribution manipulations or extensive boundary search. The proposed method is evaluated on various face avatars with different facial characteristics. The proposed method compresses the decoder model by " + }, + { + "bbox": [ + 160, + 251, + 455, + 526 + ], + "type": "inline_equation", + "content": "5.3\\times" + }, + { + "bbox": [ + 160, + 251, + 455, + 526 + ], + "type": "text", + "content": " while recovering the quality on par with the full precision baseline. In addition to the avatar rendering tasks, POCA is also applicable to image resolution enhancement tasks, achieving new SoTA image quality. https://mengjian0502.github.io/poca.github.io/" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 132, + 545, + 230, + 558 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 545, + 230, + 558 + ], + "spans": [ + { + "bbox": [ + 132, + 545, + 230, + 558 + ], + "type": "text", + "content": "1 Introduction" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 130, + 569, + 482, + 666 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 569, + 482, + 666 + ], + "spans": [ + { + "bbox": [ + 130, + 569, + 482, + 666 + ], + "type": "text", + "content": "Photorealistic telepresence [14, 15, 23, 27] has emerged as the cornerstone technology in augmented or virtual reality (AR/VR), enabling virtual interactions across different devices. Specifically, the ultimate goal of photorealistic rendering is animating the high-fidelity Codec Avatars with multiple partial facial images captured by the head-mounted cameras (HMC) [29]. The variational autoencoder-based model [14, 15] has been proved and deployed to VR device (e.g., Quest Pro) as a feasible solution. During inference, the encoder model projects the HMC-captured facial motion into the latent space, and the decoder" + } + ] + } + ], + "index": 6 + } + ], + "discarded_blocks": [], + "page_size": [ + 612, + 792 + ], + "page_idx": 0 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 161, + 115, + 452, + 287 + ], + "blocks": [ + { + "bbox": [ + 161, + 115, + 452, + 287 + ], + "lines": [ + { + "bbox": [ + 161, + 115, + 452, + 287 + ], + "spans": [ + { + "bbox": [ + 161, + 115, + 452, + 287 + ], + "type": "image", + "image_path": "4f5bb62931d49a77b0787612bcf87fe07dc0cbda4605eb3139e2927f11d35465.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 159, + 289, + 454, + 297 + ], + "lines": [ + { + "bbox": [ + 159, + 289, + 454, + 297 + ], + "spans": [ + { + "bbox": [ + 159, + 289, + 454, + 297 + ], + "type": "text", + "content": "Noise is visible in the supplementary video: \"supplementary/ID002643814/poca_all_002643814_w8a8_400029.mp4\"" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_footnote" + }, + { + "bbox": [ + 131, + 307, + 482, + 341 + ], + "lines": [ + { + "bbox": [ + 131, + 307, + 482, + 341 + ], + "spans": [ + { + "bbox": [ + 131, + 307, + 482, + 341 + ], + "type": "text", + "content": "Fig. 1: Decoding and rendering with low-precision avatar decoder. The SoTA 8-bit PTQ algorithm distorts the avatar with massive and random temporal noise, while the proposed method compresses the decoder with the high-quality avatar without noise." + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "image_caption" + } + ], + "index": 2 + }, + { + "bbox": [ + 130, + 365, + 482, + 461 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 365, + 482, + 461 + ], + "spans": [ + { + "bbox": [ + 130, + 365, + 482, + 461 + ], + "type": "text", + "content": "reconstructs the latent code into view dependent high-quality texture and mesh, which are then rendered into the 3-D Codec AvATAR on screen. In practice, the critical ingredient of photorealistic telepresence is transmitting the encoded latent information from one VR to another and rendering the avatar in real-time. In practice, the VAE-based Codec AvATAR model has been employed for generative Pixel Codec AvATAR [15] or 3D Gaussian-based AvATAR [22]. Naturally, the quality and latency of telepresence mainly depends on the efficiency of the decoder, which impacts the real-time visual quality." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 130, + 461, + 482, + 533 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 461, + 482, + 533 + ], + "spans": [ + { + "bbox": [ + 130, + 461, + 482, + 533 + ], + "type": "text", + "content": "Despite the success of the deep appearance model [14], the high computational cost and memory consumption of avatar decoding hinder the fast on-device telepresence. Naïvely deploying the full-precision decoder to the headset leads to a low frame rate, causing degraded user experience. On the other hand, relying on cloud-based computation adds additional \"headset-cloud\" communication latency and increases the risk of avatar data breaching." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 130, + 533, + 482, + 666 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 533, + 482, + 666 + ], + "spans": [ + { + "bbox": [ + 130, + 533, + 482, + 666 + ], + "type": "text", + "content": "The conflict between high-fidelity rendering and the urgent needs of on-device computing necessitates the high-performance compression algorithm designed for the avatar decoder. In particular, quantization [2, 18, 28, 30] compresses the model with a unified data compression ratio on weights and activation, leading to easy hardware deployment [24]. Unfortunately, the SoTA quantization algorithms mainly focus on classification [11, 18], object detection [26], and natural language processing tasks [30]. Recent works have investigated high-resolution static image rendering with the low-precision model [25]. However, the high-quality quantization algorithm for avatar decoding remains unexplored. In fact, our research shows that the current efforts on quantization are highly incompatible to the Avatar decoder with the following challenges:" + } + ] + } + ], + "index": 7 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "spans": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "type": "text", + "content": "2" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 166, + 91, + 223, + 101 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 166, + 91, + 223, + 101 + ], + "spans": [ + { + "bbox": [ + 166, + 91, + 223, + 101 + ], + "type": "text", + "content": "J. Meng et al." + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 1 + }, + { + "para_blocks": [ + { + "bbox": [ + 130, + 116, + 482, + 248 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 116, + 482, + 248 + ], + "spans": [ + { + "bbox": [ + 130, + 116, + 482, + 248 + ], + "type": "text", + "content": "Massive temporal noises introduced by INT8 quantization. Although the popular 8-bit quantization has been widely investigated with quantization-aware training [20] and post-training quantization [2,18,24], our discovery shows that the well-established INT8 precision introduces a massive amount of temporal noises to the rendered avatar, even with the SoTA methods. As shown in Figure 1, the facial motion generated by the 8-bit decoder is filled with temporally flickering noises between two consecutive frames, which immerses the original facial expressions and degrades the quality of the rendered avatar. Most importantly, the temporal noise is not visible in the static single frame, distorting the avatar in a continuous manner. As the first work in this area, this paper reveals the root cause of the temporal avatar noise, as described in Section 3." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 130, + 258, + 482, + 376 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 258, + 482, + 376 + ], + "spans": [ + { + "bbox": [ + 130, + 258, + 482, + 376 + ], + "type": "text", + "content": "Difficulty of characterizing temporal noise. Common quantization noises in conventional classification and object detection tasks can be modeled by a standard statistical model [12]. However, the temporal avatar noises introduced by quantized decoder exhibit inconsistent intensities and granularities in different facial regions. As a result, it is almost impossible to model the temporal noises globally without additional recurrent information. Furthermore, applying an additional filtering strategy after quantization introduces additional energy and latency consumption to the AR/VR devices, which deteriorates the efficiency of Codec Avatar telecommunication. Therefore, the noise suppression process should be achieved all at once during post training quantization." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 130, + 377, + 482, + 425 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 377, + 482, + 425 + ], + "spans": [ + { + "bbox": [ + 130, + 377, + 482, + 425 + ], + "type": "text", + "content": "To resolve the aforementioned problems, we propose POCA, a novel Post Training Quantization (PTQ) algorithm designed for high quality and temporally noise-free Codec Avatar decoding with outstanding simplicity and efficiency. The major contributions of this work are listed as follows:" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 138, + 430, + 480, + 583 + ], + "type": "list", + "angle": 0, + "index": 10, + "blocks": [ + { + "bbox": [ + 138, + 430, + 480, + 465 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 430, + 480, + 465 + ], + "spans": [ + { + "bbox": [ + 138, + 430, + 480, + 465 + ], + "type": "text", + "content": "- As the first work that investigates the low-precision avatar decoding, we reveal the root cause of the temporal noises induced by the temporally dynamic and long-tailed distribution." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 138, + 466, + 480, + 490 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 466, + 480, + 490 + ], + "spans": [ + { + "bbox": [ + 138, + 466, + 480, + 490 + ], + "type": "text", + "content": "- POCA eliminates the temporal facial noise with 8-bit and 6-bit precision, outperforming the SoTA method with fully recovered avatar quality [16]." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 138, + 490, + 480, + 513 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 490, + 480, + 513 + ], + "spans": [ + { + "bbox": [ + 138, + 490, + 480, + 513 + ], + "type": "text", + "content": "- POCA only requires " + }, + { + "bbox": [ + 138, + 490, + 480, + 513 + ], + "type": "inline_equation", + "content": "\\leq 10\\%" + }, + { + "bbox": [ + 138, + 490, + 480, + 513 + ], + "type": "text", + "content": " of layer-wise feature map data and " + }, + { + "bbox": [ + 138, + 490, + 480, + 513 + ], + "type": "inline_equation", + "content": "\\leq 5\\%" + }, + { + "bbox": [ + 138, + 490, + 480, + 513 + ], + "type": "text", + "content": " of the total training data for PTQ." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 138, + 513, + 480, + 559 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 513, + 480, + 559 + ], + "spans": [ + { + "bbox": [ + 138, + 513, + 480, + 559 + ], + "type": "text", + "content": "- POCA is evaluated on a variety of avatar identities with a full range of viewing directions and dynamic expressions, exhibiting a high degree of robustness and consistent noise-free avatar animation across different genders, age groups, hairstyles, skin colors, and view directions." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 138, + 560, + 480, + 583 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 560, + 480, + 583 + ], + "spans": [ + { + "bbox": [ + 138, + 560, + 480, + 583 + ], + "type": "text", + "content": "- POCA is also applicable to image super resolution task [8,25], achieving the SoTA image quality with " + }, + { + "bbox": [ + 138, + 560, + 480, + 583 + ], + "type": "inline_equation", + "content": "80\\%" + }, + { + "bbox": [ + 138, + 560, + 480, + 583 + ], + "type": "text", + "content": " calibration sparsity." + } + ] + } + ], + "index": 9 + } + ], + "sub_type": "text" + }, + { + "bbox": [ + 132, + 600, + 237, + 613 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 600, + 237, + 613 + ], + "spans": [ + { + "bbox": [ + 132, + 600, + 237, + 613 + ], + "type": "text", + "content": "2 Related Work" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 132, + 624, + 230, + 635 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 624, + 230, + 635 + ], + "spans": [ + { + "bbox": [ + 132, + 624, + 230, + 635 + ], + "type": "text", + "content": "2.1 Codec Avatar" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 130, + 641, + 481, + 665 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 641, + 481, + 665 + ], + "spans": [ + { + "bbox": [ + 130, + 641, + 481, + 665 + ], + "type": "text", + "content": "Photorealistic human face modeling and animation are challenging computer vision and computer graphics tasks in the sense of achieving minimum visual" + } + ] + } + ], + "index": 13 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 418, + 91, + 447, + 100 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 418, + 91, + 447, + 100 + ], + "spans": [ + { + "bbox": [ + 418, + 91, + 447, + 100 + ], + "type": "text", + "content": "POCA" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 474, + 92, + 480, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 474, + 92, + 480, + 100 + ], + "spans": [ + { + "bbox": [ + 474, + 92, + 480, + 100 + ], + "type": "text", + "content": "3" + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 2 + }, + { + "para_blocks": [ + { + "bbox": [ + 130, + 116, + 482, + 247 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 116, + 482, + 247 + ], + "spans": [ + { + "bbox": [ + 130, + 116, + 482, + 247 + ], + "type": "text", + "content": "distortion for reconstructed 3-D human faces. Early works mainly focus on accurate 3-D reconstruction with blend shape models [9], which are computationally expensive and impossible for mobile on-device telepresence. Empowered by the deep learning models, recent research [14,15] considers human face reconstruction as data-driven generative task with the variational auto-encoder neural network. In particular, the variational encoder [6] is trained for encoding the multi-view HMC images into the latent information " + }, + { + "bbox": [ + 130, + 116, + 482, + 247 + ], + "type": "inline_equation", + "content": "\\mathbf{z}" + }, + { + "bbox": [ + 130, + 116, + 482, + 247 + ], + "type": "text", + "content": ", which describes the state of the face with rich expression and motion. Subsequently, the decoder network reconstructs the 3-D faces based on the latent code and the given view direction " + }, + { + "bbox": [ + 130, + 116, + 482, + 247 + ], + "type": "inline_equation", + "content": "\\mathbf{v}" + }, + { + "bbox": [ + 130, + 116, + 482, + 247 + ], + "type": "text", + "content": ", where the assets for rendering 3D face avatar are decoded separately by the dedicated mesh decoder and texture decoder." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 130, + 247, + 482, + 415 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 247, + 482, + 415 + ], + "spans": [ + { + "bbox": [ + 130, + 247, + 482, + 415 + ], + "type": "text", + "content": "The compact latent code is transmitted between VR devices with low communication bandwidth requirements. However, the desire for real-time, on-device 3-D face avatar animation is hindered by the expensive computation cost and high memory consumption of the decoder. Various efforts have been established to improve the computation efficiency of the VAE model [5, 14]. Pixel Codec Avatar [15] proposes a novel architecture to allow lightweight decoder by creatively leveraging rendering pipeline features, skipping computations for invisible pixels while maintaining the high-fidelity for visible regions. Recent research optimizes the encoder architecture via neural architecture search (NAS) [3], leading to reduced computational costs. Different from such prior works, our work proposes a decoder-dedicated quantization algorithm that enables temporal alignments and noise-free avatar decoding with low-precision model, which further enhances instant and high quality on-device deployment of Codec Avatars for resource-constrained mobile AR/VR headsets." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 132, + 432, + 227, + 443 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 432, + 227, + 443 + ], + "spans": [ + { + "bbox": [ + 132, + 432, + 227, + 443 + ], + "type": "text", + "content": "2.2 Quantization" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 130, + 451, + 482, + 571 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 451, + 482, + 571 + ], + "spans": [ + { + "bbox": [ + 130, + 451, + 482, + 571 + ], + "type": "text", + "content": "Quantization has been widely investigated together with the evolution of deep learning. Starting from the well-established 8-bit precision [4,24] all the way to the aggressive binary neural network [32], the quantized model largely improves the computation and memory efficiency, making the complex DNN models deployable to the resource-constrained edge devices (e.g., AR/VR headset). The core idea of quantization is constructing the accurate matching between the high-precision data (e.g., FP32) and the low-precision representation. Normally, the quantization process can be generalized into the \"quantize-and-dequantize\" workflow connected by the shared scaling factor in between. Mathematically, given the full precision weights " + }, + { + "bbox": [ + 130, + 451, + 482, + 571 + ], + "type": "inline_equation", + "content": "\\mathcal{W}^l" + }, + { + "bbox": [ + 130, + 451, + 482, + 571 + ], + "type": "text", + "content": " and input activation " + }, + { + "bbox": [ + 130, + 451, + 482, + 571 + ], + "type": "inline_equation", + "content": "\\mathcal{X}^l" + }, + { + "bbox": [ + 130, + 451, + 482, + 571 + ], + "type": "text", + "content": " of a layer " + }, + { + "bbox": [ + 130, + 451, + 482, + 571 + ], + "type": "inline_equation", + "content": "l" + }, + { + "bbox": [ + 130, + 451, + 482, + 571 + ], + "type": "text", + "content": ", we have:" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 206, + 579, + 481, + 606 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 206, + 579, + 481, + 606 + ], + "spans": [ + { + "bbox": [ + 206, + 579, + 481, + 606 + ], + "type": "interline_equation", + "content": "\\mathcal {W} _ {Q} ^ {l} = \\left\\lfloor \\frac {\\mathcal {W} ^ {l}}{S _ {w} ^ {l}} \\right] \\in \\operatorname {I N T} _ {\\mathbf {n}}, \\quad \\mathcal {X} _ {Q} ^ {l} = \\left\\lfloor \\frac {\\mathcal {X} ^ {l}}{S _ {x} ^ {l}} \\right] \\in \\operatorname {I N T} _ {\\mathbf {n}} \\tag {1}", + "image_path": "1f6b9aa351ed117c3a1135d7cd12bff1ba4438bda4ab4fc025d77dd5657f0904.jpg" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 130, + 609, + 482, + 645 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 609, + 482, + 645 + ], + "spans": [ + { + "bbox": [ + 130, + 609, + 482, + 645 + ], + "type": "text", + "content": "Where " + }, + { + "bbox": [ + 130, + 609, + 482, + 645 + ], + "type": "inline_equation", + "content": "S" + }, + { + "bbox": [ + 130, + 609, + 482, + 645 + ], + "type": "text", + "content": " is the scaling factor determined by data precision " + }, + { + "bbox": [ + 130, + 609, + 482, + 645 + ], + "type": "inline_equation", + "content": "\\mathbf{n}" + }, + { + "bbox": [ + 130, + 609, + 482, + 645 + ], + "type": "text", + "content": " and the numerical range. Specifically, " + }, + { + "bbox": [ + 130, + 609, + 482, + 645 + ], + "type": "inline_equation", + "content": "S" + }, + { + "bbox": [ + 130, + 609, + 482, + 645 + ], + "type": "text", + "content": " characterizes the \"interval\" between adjacent levels in the floating point dequantized values, bounded by the range [lb, ub]:" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 261, + 654, + 481, + 666 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 261, + 654, + 481, + 666 + ], + "spans": [ + { + "bbox": [ + 261, + 654, + 481, + 666 + ], + "type": "interline_equation", + "content": "S = | \\mathrm {u b} - \\mathrm {l b} | / 2 ^ {\\mathbf {n}} - 1 \\tag {2}", + "image_path": "a966121298f99932f76507eca8e0793a0d736b26f2d9e24ba2cc8b746c613c56.jpg" + } + ] + } + ], + "index": 8 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "spans": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "type": "text", + "content": "4" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 166, + 91, + 223, + 102 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 166, + 91, + 223, + 102 + ], + "spans": [ + { + "bbox": [ + 166, + 91, + 223, + 102 + ], + "type": "text", + "content": "J. Meng et al." + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 3 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 160, + 114, + 452, + 327 + ], + "blocks": [ + { + "bbox": [ + 160, + 114, + 452, + 327 + ], + "lines": [ + { + "bbox": [ + 160, + 114, + 452, + 327 + ], + "spans": [ + { + "bbox": [ + 160, + 114, + 452, + 327 + ], + "type": "image", + "image_path": "fb71659a89c7e4a15b921ebd294a5c266747bcdaa27d6bf348d6432f958aecf8.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 130, + 336, + 482, + 370 + ], + "lines": [ + { + "bbox": [ + 130, + 336, + 482, + 370 + ], + "spans": [ + { + "bbox": [ + 130, + 336, + 482, + 370 + ], + "type": "text", + "content": "Fig. 2: Top: The rendered avatar exhibits minimal quality (PSNR) reduction in static frames and large distortion with animation. Bottom: SoTA 8-bit quantization [2,18] amplifies the pixel shifting between frames in the rendered avatar across RGB channels." + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_caption" + } + ], + "index": 2 + }, + { + "bbox": [ + 130, + 394, + 482, + 575 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 394, + 482, + 575 + ], + "spans": [ + { + "bbox": [ + 130, + 394, + 482, + 575 + ], + "type": "text", + "content": "Since the number of levels is directly determined by the precision " + }, + { + "bbox": [ + 130, + 394, + 482, + 575 + ], + "type": "inline_equation", + "content": "\\mathbf{n}" + }, + { + "bbox": [ + 130, + 394, + 482, + 575 + ], + "type": "text", + "content": ", the essence of quantization accurately approximates the floating-point distribution with the dequantized tensors while minimizing the model-level error. To that end, a variety of quantization algorithms have been proposed to locate the optimal scaling factors with QAT [13] or PTQ [11, 18, 28], including boundary learning [1, 10], scaling factor learning [2], and quantization error modeling [12], etc. Recent efforts introduce additional scaling to avoid the error caused by the extreme distribution [30]. Promoted by the fast compression process, PTQ exhibits high compatibility for quantizing the avatar VAE model compared to QAT. Ideally, compressing the Codec Avatar models with PTQ would lead to instant \"user-avatar\" conversion for AR/VR devices, compared to time-consuming QAT. Recent VAE-based quantization focuses on the video compression task with quantized latent code [7, 21], while the quantization of the codec-avatar decoder model has not been explored despite the urgent needs on edge device and high rendering quality." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 131, + 594, + 381, + 608 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 131, + 594, + 381, + 608 + ], + "spans": [ + { + "bbox": [ + 131, + 594, + 381, + 608 + ], + "type": "text", + "content": "3 Root Cause of Temporal Avatar Noise" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 131, + 620, + 418, + 632 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 131, + 620, + 418, + 632 + ], + "spans": [ + { + "bbox": [ + 131, + 620, + 418, + 632 + ], + "type": "text", + "content": "3.1 Temporal Noise is Dynamic Pixel Color Oscillation" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 130, + 641, + 487, + 666 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 641, + 487, + 666 + ], + "spans": [ + { + "bbox": [ + 130, + 641, + 487, + 666 + ], + "type": "text", + "content": "Recent 3D rendering tasks (e.g., NeRF) [17] uses peak signal-to-noise ratio (PSNR) to evaluate the quality between the rendered output and the ground truth im" + } + ] + } + ], + "index": 7 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 418, + 91, + 447, + 100 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 418, + 91, + 447, + 100 + ], + "spans": [ + { + "bbox": [ + 418, + 91, + 447, + 100 + ], + "type": "text", + "content": "POCA" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 474, + 92, + 480, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 474, + 92, + 480, + 100 + ], + "spans": [ + { + "bbox": [ + 474, + 92, + 480, + 100 + ], + "type": "text", + "content": "5" + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 4 + }, + { + "para_blocks": [ + { + "bbox": [ + 130, + 116, + 479, + 139 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 116, + 479, + 139 + ], + "spans": [ + { + "bbox": [ + 130, + 116, + 479, + 139 + ], + "type": "text", + "content": "age. However, the temporal distortion of Codec Avatar cannot be effectively quantified with the same metrics." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 130, + 140, + 480, + 200 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 140, + 480, + 200 + ], + "spans": [ + { + "bbox": [ + 130, + 140, + 480, + 200 + ], + "type": "text", + "content": "With the low-precision decoding, the temporal noise shows minimal impact on the static frame quality, but largely distorts the dynamic motion of the avatar, as shown in Figure 2 and supplementary material. Prior works only focus on the static image quality [25] with low-precision models, but the dynamic quality reduction has not been investigated well." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 130, + 201, + 480, + 236 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 201, + 480, + 236 + ], + "spans": [ + { + "bbox": [ + 130, + 201, + 480, + 236 + ], + "type": "text", + "content": "In addition to the FovVideoVDP score [16], the temporal pixel oscillation can be characterized by the mean pixel value differences between two consecutive frames " + }, + { + "bbox": [ + 130, + 201, + 480, + 236 + ], + "type": "inline_equation", + "content": "F_{t}" + }, + { + "bbox": [ + 130, + 201, + 480, + 236 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 130, + 201, + 480, + 236 + ], + "type": "inline_equation", + "content": "F_{t + 1}" + }, + { + "bbox": [ + 130, + 201, + 480, + 236 + ], + "type": "text", + "content": " of the rendered avatar:" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 252, + 247, + 480, + 269 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 252, + 247, + 480, + 269 + ], + "spans": [ + { + "bbox": [ + 252, + 247, + 480, + 269 + ], + "type": "interline_equation", + "content": "\\Delta_ {F} ^ {c} = \\frac {1}{\\mathcal {H W}} \\times | F _ {t + 1} - F _ {t} | \\tag {3}", + "image_path": "3365e7991dc6780d13a291cfa7badd5f76dc7b414acbd7056729334546111124.jpg" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 131, + 276, + 481, + 324 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 131, + 276, + 481, + 324 + ], + "spans": [ + { + "bbox": [ + 131, + 276, + 481, + 324 + ], + "type": "text", + "content": "Where " + }, + { + "bbox": [ + 131, + 276, + 481, + 324 + ], + "type": "inline_equation", + "content": "\\mathcal{H}" + }, + { + "bbox": [ + 131, + 276, + 481, + 324 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 131, + 276, + 481, + 324 + ], + "type": "inline_equation", + "content": "\\mathcal{W}" + }, + { + "bbox": [ + 131, + 276, + 481, + 324 + ], + "type": "text", + "content": " represent the height and width of the rendered frame. We evaluate the pixel value differences separately for RGB channels such that " + }, + { + "bbox": [ + 131, + 276, + 481, + 324 + ], + "type": "inline_equation", + "content": "c \\in \\{R, G, B\\}" + }, + { + "bbox": [ + 131, + 276, + 481, + 324 + ], + "type": "text", + "content": ". The trend of the pixel changes characterizes the basic transition of the facial expression (e.g., Tongue in and out), as shown in Figure 2 (Bottom)." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 130, + 325, + 481, + 444 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 325, + 481, + 444 + ], + "spans": [ + { + "bbox": [ + 130, + 325, + 481, + 444 + ], + "type": "text", + "content": "With the 8-bit VAE decoder, the frame-wise pixel dynamics exhibit a similar trend as the full-precision decoder baseline. However, the quantized decoder adds an additional \"offset\" to the transition of the facial motion. In other words, the additional distortion is introduced to the avatar on top of the natural motion, which leads to random texture fluctuation in the rendered Codec Avatar. Figure 2 (Bottom) shows that the temporal noises uniformly exist across the entire facial motion with low intensity of " + }, + { + "bbox": [ + 130, + 325, + 481, + 444 + ], + "type": "inline_equation", + "content": "\\pm 0.44" + }, + { + "bbox": [ + 130, + 325, + 481, + 444 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 130, + 325, + 481, + 444 + ], + "type": "inline_equation", + "content": "\\pm 0.47" + }, + { + "bbox": [ + 130, + 325, + 481, + 444 + ], + "type": "text", + "content": ", and " + }, + { + "bbox": [ + 130, + 325, + 481, + 444 + ], + "type": "inline_equation", + "content": "\\pm 0.25" + }, + { + "bbox": [ + 130, + 325, + 481, + 444 + ], + "type": "text", + "content": " average pixel shifting on the R, G, B channels, respectively. We are fully aware of the existence of other motion capture matrices (e.g., optical flow), but the temporal noises are significantly fine-grained, which requires detailed pixel-level analysis." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 130, + 445, + 481, + 529 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 445, + 481, + 529 + ], + "spans": [ + { + "bbox": [ + 130, + 445, + 481, + 529 + ], + "type": "text", + "content": "Formally, such random pixel flickering can be captured by the FovVideoVDP [16] score between the rendered avatar and ground truth image. As shown in Figure 2, the INT8 avatar decoder largely degrades the animation quality of the avatar, even with the SoTA 8-bit quantization method [2, 18]. The high density of the yellow region indicates the strongly mismatched animation caused by the temporal noise. The continuous pixel oscillation and temporal noises worsen the fidelity of the rendered Codec Avatar in motion." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 131, + 548, + 307, + 560 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 131, + 548, + 307, + 560 + ], + "spans": [ + { + "bbox": [ + 131, + 548, + 307, + 560 + ], + "type": "text", + "content": "3.2 Root Cause of Noisy Avatars" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 130, + 569, + 480, + 640 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 569, + 480, + 640 + ], + "spans": [ + { + "bbox": [ + 130, + 569, + 480, + 640 + ], + "type": "text", + "content": "Since the temporal noises are induced by the pixel oscillation of the avatar in motion, an intuitive understanding of the noises is the \"unstable rounding\" during quantization [12, 19]. However, the rounding instability is still the phenomenon of quantization instead of the root cause of the noise. In this work, we unravel the fundamental reason of the temporal noise induced by low-precision avatar decoding." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 130, + 641, + 481, + 665 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 641, + 481, + 665 + ], + "spans": [ + { + "bbox": [ + 130, + 641, + 481, + 665 + ], + "type": "text", + "content": "Dynamically long-tailed distribution of human face. Human facial features naturally display a diverse range of distributions characterized by dif" + } + ] + } + ], + "index": 11 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "spans": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "type": "text", + "content": "6" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 166, + 91, + 223, + 102 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 166, + 91, + 223, + 102 + ], + "spans": [ + { + "bbox": [ + 166, + 91, + 223, + 102 + ], + "type": "text", + "content": "J. Meng et al." + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 5 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 173, + 115, + 293, + 220 + ], + "blocks": [ + { + "bbox": [ + 173, + 115, + 293, + 220 + ], + "lines": [ + { + "bbox": [ + 173, + 115, + 293, + 220 + ], + "spans": [ + { + "bbox": [ + 173, + 115, + 293, + 220 + ], + "type": "image", + "image_path": "7edb81847558169fa6b09c1dff05d88de636549afab9f57ace19fd39adf3339a.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 195, + 223, + 209, + 235 + ], + "lines": [ + { + "bbox": [ + 195, + 223, + 209, + 235 + ], + "spans": [ + { + "bbox": [ + 195, + 223, + 209, + 235 + ], + "type": "text", + "content": "(a)" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_caption" + } + ], + "index": 2 + }, + { + "type": "image", + "bbox": [ + 294, + 121, + 446, + 220 + ], + "blocks": [ + { + "bbox": [ + 294, + 121, + 446, + 220 + ], + "lines": [ + { + "bbox": [ + 294, + 121, + 446, + 220 + ], + "spans": [ + { + "bbox": [ + 294, + 121, + 446, + 220 + ], + "type": "image", + "image_path": "f0c263706cb0acc9a3dd87cc5cec5cf3c3d880209167bea05464059504959cde.jpg" + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 359, + 223, + 373, + 235 + ], + "lines": [ + { + "bbox": [ + 359, + 223, + 373, + 235 + ], + "spans": [ + { + "bbox": [ + 359, + 223, + 373, + 235 + ], + "type": "text", + "content": "(b)" + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 131, + 243, + 480, + 266 + ], + "lines": [ + { + "bbox": [ + 131, + 243, + 480, + 266 + ], + "spans": [ + { + "bbox": [ + 131, + 243, + 480, + 266 + ], + "type": "text", + "content": "Fig. 3: (a) Row-channel plane of the texture decoder activation and (b) Regional standard deviation of the row-channel planes with respect to different facial expressions." + } + ] + } + ], + "index": 6, + "angle": 0, + "type": "image_caption" + } + ], + "index": 4 + }, + { + "bbox": [ + 130, + 289, + 480, + 373 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 289, + 480, + 373 + ], + "spans": [ + { + "bbox": [ + 130, + 289, + 480, + 373 + ], + "type": "text", + "content": "ferent facial regions (for instance, hair compared to cheeks). The broad spectrum of variability in human facial characteristics formulates various statistical properties. For the avatar decoding, the texture decoder of the VAE model [14] decodes face avatar texture from the latent space via transposed convolutional layers, which consecutively up-scales the activation feature map from " + }, + { + "bbox": [ + 130, + 289, + 480, + 373 + ], + "type": "inline_equation", + "content": "4 \\times 4" + }, + { + "bbox": [ + 130, + 289, + 480, + 373 + ], + "type": "text", + "content": " to " + }, + { + "bbox": [ + 130, + 289, + 480, + 373 + ], + "type": "inline_equation", + "content": "1024 \\times 1024" + }, + { + "bbox": [ + 130, + 289, + 480, + 373 + ], + "type": "text", + "content": " (height and width). Therefore, the statistical variation of the facial texture is also embedded into the activation of the earlier layer." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 130, + 376, + 481, + 449 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 376, + 481, + 449 + ], + "spans": [ + { + "bbox": [ + 130, + 376, + 481, + 449 + ], + "type": "text", + "content": "Proof of facial dynamics. We prove the dynamic range variations of the texture by analyzing the regional distribution of the latent layer activation. Assume " + }, + { + "bbox": [ + 130, + 376, + 481, + 449 + ], + "type": "inline_equation", + "content": "\\mathcal{C}" + }, + { + "bbox": [ + 130, + 376, + 481, + 449 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 130, + 376, + 481, + 449 + ], + "type": "inline_equation", + "content": "\\mathcal{H}" + }, + { + "bbox": [ + 130, + 376, + 481, + 449 + ], + "type": "text", + "content": ", and " + }, + { + "bbox": [ + 130, + 376, + 481, + 449 + ], + "type": "inline_equation", + "content": "\\mathcal{W}" + }, + { + "bbox": [ + 130, + 376, + 481, + 449 + ], + "type": "text", + "content": " represent the channel depth, height, and width of the activation tensor. We analyze the statistics by decomposing the entire face texture feature map into multiple \"row-channel planes\". Each plane contains the regional statistics from hair to jaw, with the total number of planes = " + }, + { + "bbox": [ + 130, + 376, + 481, + 449 + ], + "type": "inline_equation", + "content": "\\mathcal{H}" + }, + { + "bbox": [ + 130, + 376, + 481, + 449 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 130, + 453, + 481, + 548 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 453, + 481, + 548 + ], + "spans": [ + { + "bbox": [ + 130, + 453, + 481, + 548 + ], + "type": "text", + "content": "Figure 3(a) shows the row-channel decomposition based on the intermediate activation feature map of the texture decoder with the shape of " + }, + { + "bbox": [ + 130, + 453, + 481, + 548 + ], + "type": "inline_equation", + "content": "16 \\times 256 \\times 256" + }, + { + "bbox": [ + 130, + 453, + 481, + 548 + ], + "type": "text", + "content": ". On top of that, we compute the standard deviation of each plane, which characterizes the local tailedness and dynamic range in each region of the face, with respect to different facial expressions. As depicted in Figure 3(b), smiling and eye blinking lead to the long-tailed dynamics in the mouth and eye regions, respectively. As a result, the distribution of the entire face is dynamically long-tailed, which is harmful for quantization. In other words:" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 131, + 553, + 481, + 578 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 131, + 553, + 481, + 578 + ], + "spans": [ + { + "bbox": [ + 131, + 553, + 481, + 578 + ], + "type": "text", + "content": "The distribution of a decoder feature map contains both constrained and long-tailed distribution, while the tailedness varies between different facial expressions." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 130, + 582, + 481, + 665 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 582, + 481, + 665 + ], + "spans": [ + { + "bbox": [ + 130, + 582, + 481, + 665 + ], + "type": "text", + "content": "In other words, the root cause of the temporal noises is the conflict between the long-tailed, highly dynamic distribution of activations and the global error minimization during PTQ. To minimize the global quantization error, the long-tailed distribution stretches the quantization interval (scaling factor) in order to cover larger range, which in turn, elevates the local rounding instability and leads to unstable numerical representation of the pixels. We prove this hypothesis in Section 4.3 with the proposed method." + } + ] + } + ], + "index": 11 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 418, + 91, + 447, + 100 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 418, + 91, + 447, + 100 + ], + "spans": [ + { + "bbox": [ + 418, + 91, + 447, + 100 + ], + "type": "text", + "content": "POCA" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 474, + 91, + 480, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 474, + 91, + 480, + 100 + ], + "spans": [ + { + "bbox": [ + 474, + 91, + 480, + 100 + ], + "type": "text", + "content": "7" + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 6 + }, + { + "para_blocks": [ + { + "bbox": [ + 132, + 114, + 308, + 129 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 114, + 308, + 129 + ], + "spans": [ + { + "bbox": [ + 132, + 114, + 308, + 129 + ], + "type": "text", + "content": "4 Proposed Method: POCA" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 130, + 142, + 482, + 204 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 142, + 482, + 204 + ], + "spans": [ + { + "bbox": [ + 130, + 142, + 482, + 204 + ], + "type": "text", + "content": "To resolve the temporal noises caused by low-precision avatar decoding, this section presents POCA, a novel post-training quantization algorithm that successfully compresses the deep appearance decoder and eliminates the temporal noises from the rendered Codec Avatar. In particular, POCA is empowered by two novel techniques: a) View-aware PTQ and b) PTQ with shape-wise filtering." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 132, + 223, + 248, + 236 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 223, + 248, + 236 + ], + "spans": [ + { + "bbox": [ + 132, + 223, + 248, + 236 + ], + "type": "text", + "content": "4.1 View-aware PTQ" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 130, + 246, + 482, + 376 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 246, + 482, + 376 + ], + "spans": [ + { + "bbox": [ + 130, + 246, + 482, + 376 + ], + "type": "text", + "content": "As pointed out in Section 3, the face texture contains both region-dependent dispersion and facial expression-driven dynamics in the entire face. As a result, the sample space of the training data and layer-wise activation contains high variation with redundant information. In particular, the texture decoder is trained by the whole face texture map with the size of " + }, + { + "bbox": [ + 130, + 246, + 482, + 376 + ], + "type": "inline_equation", + "content": "1024 \\times 1024" + }, + { + "bbox": [ + 130, + 246, + 482, + 376 + ], + "type": "text", + "content": ". However, the final 3D avatar is rendered based on the visibility information from the given view direction. Therefore, quantizing the decoder without considering the visibility for inner layers leads to the incorrectly learned quantization parameter, which further generates the mismatched quantization decision between PTQ and actual rendering. Motivated by that, we propose the view-aware PTQ for the avatar texture decoder." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 130, + 378, + 482, + 511 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 378, + 482, + 511 + ], + "spans": [ + { + "bbox": [ + 130, + 378, + 482, + 511 + ], + "type": "text", + "content": "Basic Setup. Given the pre-trained full-precision VAE decoder (texture and mesh), we first randomly select " + }, + { + "bbox": [ + 130, + 378, + 482, + 511 + ], + "type": "inline_equation", + "content": "m" + }, + { + "bbox": [ + 130, + 378, + 482, + 511 + ], + "type": "text", + "content": " training samples from the full-sized dataset with size " + }, + { + "bbox": [ + 130, + 378, + 482, + 511 + ], + "type": "inline_equation", + "content": "\\mathcal{D}" + }, + { + "bbox": [ + 130, + 378, + 482, + 511 + ], + "type": "text", + "content": " (" + }, + { + "bbox": [ + 130, + 378, + 482, + 511 + ], + "type": "inline_equation", + "content": "m << \\mathcal{D}" + }, + { + "bbox": [ + 130, + 378, + 482, + 511 + ], + "type": "text", + "content": "). Each training sample contains the input and pre-computed output of the decoder (" + }, + { + "bbox": [ + 130, + 378, + 482, + 511 + ], + "type": "inline_equation", + "content": "[z_t, v_t], [\\hat{\\mathbf{T}}_t, \\hat{\\mathbf{M}}_t]" + }, + { + "bbox": [ + 130, + 378, + 482, + 511 + ], + "type": "text", + "content": "), where " + }, + { + "bbox": [ + 130, + 378, + 482, + 511 + ], + "type": "inline_equation", + "content": "z_t, v_t, \\hat{\\mathbf{T}}_t" + }, + { + "bbox": [ + 130, + 378, + 482, + 511 + ], + "type": "text", + "content": ", and " + }, + { + "bbox": [ + 130, + 378, + 482, + 511 + ], + "type": "inline_equation", + "content": "\\hat{\\mathbf{M}}_t" + }, + { + "bbox": [ + 130, + 378, + 482, + 511 + ], + "type": "text", + "content": " represents the encoded latent code, view direction, decoded texture, and decoded mesh for sample " + }, + { + "bbox": [ + 130, + 378, + 482, + 511 + ], + "type": "inline_equation", + "content": "t" + }, + { + "bbox": [ + 130, + 378, + 482, + 511 + ], + "type": "text", + "content": ", respectively. Following the standard PTQ strategy, we extract input " + }, + { + "bbox": [ + 130, + 378, + 482, + 511 + ], + "type": "inline_equation", + "content": "\\mathcal{X}^l" + }, + { + "bbox": [ + 130, + 378, + 482, + 511 + ], + "type": "text", + "content": " and output activation " + }, + { + "bbox": [ + 130, + 378, + 482, + 511 + ], + "type": "inline_equation", + "content": "\\mathcal{Y}^l" + }, + { + "bbox": [ + 130, + 378, + 482, + 511 + ], + "type": "text", + "content": " of each layer as the calibration data. We adopted AdaRound [18] and LSQ [2] as the weight and activation quantizers, respectively. The learnable quantization parameters (" + }, + { + "bbox": [ + 130, + 378, + 482, + 511 + ], + "type": "inline_equation", + "content": "\\alpha, S^l" + }, + { + "bbox": [ + 130, + 378, + 482, + 511 + ], + "type": "text", + "content": ") are optimized by minimizing the mean square error (MSE) between the quantized layer output " + }, + { + "bbox": [ + 130, + 378, + 482, + 511 + ], + "type": "inline_equation", + "content": "\\hat{\\mathcal{Y}}^l" + }, + { + "bbox": [ + 130, + 378, + 482, + 511 + ], + "type": "text", + "content": " and full precision output " + }, + { + "bbox": [ + 130, + 378, + 482, + 511 + ], + "type": "inline_equation", + "content": "\\mathcal{Y}^l" + }, + { + "bbox": [ + 130, + 378, + 482, + 511 + ], + "type": "text", + "content": ":" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 245, + 521, + 481, + 546 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 245, + 521, + 481, + 546 + ], + "spans": [ + { + "bbox": [ + 245, + 521, + 481, + 546 + ], + "type": "interline_equation", + "content": "\\underset {\\alpha , S ^ {l}} {\\arg \\min } \\frac {1}{\\mathcal {H W C}} \\sum \\left(\\mathcal {Y} ^ {l} - \\hat {\\mathcal {Y}} ^ {l}\\right) ^ {2} \\tag {4}", + "image_path": "34e421daf304567c3426d2235ae6b23927ea5af2e13b347569ba714e899f6ea3.jpg" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 130, + 556, + 482, + 641 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 556, + 482, + 641 + ], + "spans": [ + { + "bbox": [ + 130, + 556, + 482, + 641 + ], + "type": "text", + "content": "View-dependent visibility mask. The computation of visibility is inherited from the rendering pipeline utilizing the pre-trained VAE model. Specifically, the decoded mesh vertices " + }, + { + "bbox": [ + 130, + 556, + 482, + 641 + ], + "type": "inline_equation", + "content": "\\hat{\\mathbf{M}}_t" + }, + { + "bbox": [ + 130, + 556, + 482, + 641 + ], + "type": "text", + "content": " are rasterized to formulate the barycentric coordinates, which are leveraged to interpolate the UV coordinates " + }, + { + "bbox": [ + 130, + 556, + 482, + 641 + ], + "type": "inline_equation", + "content": "\\varPhi" + }, + { + "bbox": [ + 130, + 556, + 482, + 641 + ], + "type": "text", + "content": " across the entire avatar. Following this, the visibility is determined by sampling the view-specific pixels from the 2-D texture map based on the interpolated UV indexes, " + }, + { + "bbox": [ + 130, + 556, + 482, + 641 + ], + "type": "inline_equation", + "content": "\\varPhi = [\\phi_x,\\phi_y]" + }, + { + "bbox": [ + 130, + 556, + 482, + 641 + ], + "type": "inline_equation", + "content": "(0\\leq \\phi_{x},\\phi_{y}\\leq 1)" + }, + { + "bbox": [ + 130, + 556, + 482, + 641 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 130, + 641, + 482, + 666 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 641, + 482, + 666 + ], + "spans": [ + { + "bbox": [ + 130, + 641, + 482, + 666 + ], + "type": "text", + "content": "The value of " + }, + { + "bbox": [ + 130, + 641, + 482, + 666 + ], + "type": "inline_equation", + "content": "\\phi_x" + }, + { + "bbox": [ + 130, + 641, + 482, + 666 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 130, + 641, + 482, + 666 + ], + "type": "inline_equation", + "content": "\\phi_y" + }, + { + "bbox": [ + 130, + 641, + 482, + 666 + ], + "type": "text", + "content": " represents the normalized coordinates of the 2-D texture map with the shape of " + }, + { + "bbox": [ + 130, + 641, + 482, + 666 + ], + "type": "inline_equation", + "content": "\\mathcal{H} \\times \\mathcal{W}" + }, + { + "bbox": [ + 130, + 641, + 482, + 666 + ], + "type": "text", + "content": ". Different from the default pipeline, the" + } + ] + } + ], + "index": 9 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "spans": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "type": "text", + "content": "8" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 166, + 91, + 223, + 102 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 166, + 91, + 223, + 102 + ], + "spans": [ + { + "bbox": [ + 166, + 91, + 223, + 102 + ], + "type": "text", + "content": "J. Meng et al." + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 7 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 151, + 115, + 462, + 276 + ], + "blocks": [ + { + "bbox": [ + 151, + 115, + 462, + 276 + ], + "lines": [ + { + "bbox": [ + 151, + 115, + 462, + 276 + ], + "spans": [ + { + "bbox": [ + 151, + 115, + 462, + 276 + ], + "type": "image", + "image_path": "5f091620bbd48526a320aaa752ccb47cd4a8c80de29fcecb15789ba85aea0f78.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 131, + 285, + 480, + 319 + ], + "lines": [ + { + "bbox": [ + 131, + 285, + 480, + 319 + ], + "spans": [ + { + "bbox": [ + 131, + 285, + 480, + 319 + ], + "type": "text", + "content": "Fig. 4: Overall PTQ process of POCA based on layer-wise calibration (Layer Calib). POCA fully quantizes all the layers in the VAE decoder, where the proposed view-dependent " + }, + { + "bbox": [ + 131, + 285, + 480, + 319 + ], + "type": "inline_equation", + "content": "+" + }, + { + "bbox": [ + 131, + 285, + 480, + 319 + ], + "type": "text", + "content": " SWD filter is applied to the decoder with transpose convolution layers." + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_caption" + } + ], + "index": 2 + }, + { + "bbox": [ + 130, + 339, + 479, + 363 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 339, + 479, + 363 + ], + "spans": [ + { + "bbox": [ + 130, + 339, + 479, + 363 + ], + "type": "text", + "content": "We proposed POCA algorithm generates the binary \"visibility mask\" " + }, + { + "bbox": [ + 130, + 339, + 479, + 363 + ], + "type": "inline_equation", + "content": "\\Gamma" + }, + { + "bbox": [ + 130, + 339, + 479, + 363 + ], + "type": "text", + "content": " to select the pixels in the 2-D texture map " + }, + { + "bbox": [ + 130, + 339, + 479, + 363 + ], + "type": "inline_equation", + "content": "\\hat{\\mathbf{T}}_t" + }, + { + "bbox": [ + 130, + 339, + 479, + 363 + ], + "type": "text", + "content": ". Formally, we have:" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 197, + 369, + 481, + 383 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 197, + 369, + 481, + 383 + ], + "spans": [ + { + "bbox": [ + 197, + 369, + 481, + 383 + ], + "type": "interline_equation", + "content": "\\Phi = \\text {I n t e r p o l a t e} (\\text {R a s t e r i z e} (\\mathbf {M} _ {t}, \\dots), u v) \\tag {5}", + "image_path": "003db8c8b5afc045050666de4ed8eca324ee4823df9c7704148ca49597db67b8.jpg" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 198, + 384, + 480, + 397 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 198, + 384, + 480, + 397 + ], + "spans": [ + { + "bbox": [ + 198, + 384, + 480, + 397 + ], + "type": "interline_equation", + "content": "\\Gamma = \\mathbb {1} \\cdot \\left\\{\\left(\\gamma_ {x}, \\gamma_ {y}\\right) \\mid \\gamma_ {x} \\in \\left\\{\\phi_ {x} \\times \\mathcal {H} \\right\\}, \\gamma_ {y} \\in \\left\\{\\phi_ {y} \\times \\mathcal {W} \\right\\} \\right\\} \\tag {6}", + "image_path": "24cf56dfe7bb304c9a7cb312698ddfd54f28666315018ecc37b639cca4cd7f00.jpg" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 131, + 402, + 479, + 449 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 131, + 402, + 479, + 449 + ], + "spans": [ + { + "bbox": [ + 131, + 402, + 479, + 449 + ], + "type": "text", + "content": "Where " + }, + { + "bbox": [ + 131, + 402, + 479, + 449 + ], + "type": "inline_equation", + "content": "\\mathbb{1}" + }, + { + "bbox": [ + 131, + 402, + 479, + 449 + ], + "type": "text", + "content": " represents the indicator function that signifies whether the pixel at " + }, + { + "bbox": [ + 131, + 402, + 479, + 449 + ], + "type": "inline_equation", + "content": "(\\gamma_x,\\gamma_y)" + }, + { + "bbox": [ + 131, + 402, + 479, + 449 + ], + "type": "text", + "content": " should be active at the given visibility. " + }, + { + "bbox": [ + 131, + 402, + 479, + 449 + ], + "type": "inline_equation", + "content": "\\{\\phi_x\\times \\mathcal{H}\\}" + }, + { + "bbox": [ + 131, + 402, + 479, + 449 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 131, + 402, + 479, + 449 + ], + "type": "inline_equation", + "content": "\\{\\phi_y\\times \\mathcal{W}\\}" + }, + { + "bbox": [ + 131, + 402, + 479, + 449 + ], + "type": "text", + "content": " are the visibility coordinates (matrix indexes) scaled based on the interpolated UV coordinates " + }, + { + "bbox": [ + 131, + 402, + 479, + 449 + ], + "type": "inline_equation", + "content": "\\varPhi" + }, + { + "bbox": [ + 131, + 402, + 479, + 449 + ], + "type": "text", + "content": " with the dimensionality of the texture map." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 130, + 450, + 480, + 522 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 450, + 480, + 522 + ], + "spans": [ + { + "bbox": [ + 130, + 450, + 480, + 522 + ], + "type": "text", + "content": "The visibility masks are generated during the data-fetching process. Since the model is already pre-trained, the visibility masks " + }, + { + "bbox": [ + 130, + 450, + 480, + 522 + ], + "type": "inline_equation", + "content": "\\varGamma" + }, + { + "bbox": [ + 130, + 450, + 480, + 522 + ], + "type": "text", + "content": " are accessible with each full-precision texture " + }, + { + "bbox": [ + 130, + 450, + 480, + 522 + ], + "type": "inline_equation", + "content": "\\hat{\\mathbf{T}}_t" + }, + { + "bbox": [ + 130, + 450, + 480, + 522 + ], + "type": "text", + "content": " and triangle index " + }, + { + "bbox": [ + 130, + 450, + 480, + 522 + ], + "type": "inline_equation", + "content": "uv" + }, + { + "bbox": [ + 130, + 450, + 480, + 522 + ], + "type": "text", + "content": " at the output layer of the texture decoder. The final layer " + }, + { + "bbox": [ + 130, + 450, + 480, + 522 + ], + "type": "inline_equation", + "content": "\\varGamma" + }, + { + "bbox": [ + 130, + 450, + 480, + 522 + ], + "type": "text", + "content": " (" + }, + { + "bbox": [ + 130, + 450, + 480, + 522 + ], + "type": "inline_equation", + "content": "1024 \\times 1024" + }, + { + "bbox": [ + 130, + 450, + 480, + 522 + ], + "type": "text", + "content": ") will be max-pooled to formulate the visibility mask for the earlier layer. For each layer of the texture decoder, POCA performs PTQ with the visible pixel only:" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 217, + 526, + 481, + 553 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 217, + 526, + 481, + 553 + ], + "spans": [ + { + "bbox": [ + 217, + 526, + 481, + 553 + ], + "type": "interline_equation", + "content": "\\underset {\\alpha^ {l}, S ^ {l}} {\\arg \\min } \\frac {1}{\\mathcal {H W C}} \\sum \\left(\\mathcal {Y} _ {\\mathbf {T}} ^ {l} \\odot \\Gamma^ {l} - \\hat {\\mathcal {Y}} _ {\\mathbf {T}} ^ {l} \\odot \\Gamma^ {l}\\right) ^ {2} \\tag {7}", + "image_path": "e825149b977d7d0ed7c395019a5a9e90d468902eb5acc40d2bd8945437e1d461.jpg" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 130, + 558, + 482, + 665 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 558, + 482, + 665 + ], + "spans": [ + { + "bbox": [ + 130, + 558, + 482, + 665 + ], + "type": "text", + "content": "Where " + }, + { + "bbox": [ + 130, + 558, + 482, + 665 + ], + "type": "inline_equation", + "content": "\\odot" + }, + { + "bbox": [ + 130, + 558, + 482, + 665 + ], + "type": "text", + "content": " represents the element-wise product between tensors. Subsequently, the learnable quantization parameters " + }, + { + "bbox": [ + 130, + 558, + 482, + 665 + ], + "type": "inline_equation", + "content": "\\alpha^l" + }, + { + "bbox": [ + 130, + 558, + 482, + 665 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 130, + 558, + 482, + 665 + ], + "type": "inline_equation", + "content": "S^l" + }, + { + "bbox": [ + 130, + 558, + 482, + 665 + ], + "type": "text", + "content": " are trained based on the gradient that resulted from the view-dependent visibility " + }, + { + "bbox": [ + 130, + 558, + 482, + 665 + ], + "type": "inline_equation", + "content": "\\Gamma" + }, + { + "bbox": [ + 130, + 558, + 482, + 665 + ], + "type": "text", + "content": ". As a result, POCA optimizes the quantization decision based on the most relevant information of the training sample, instead of being diversified by the widely dispersed texture space with long-tailed and dynamic distribution. Figure 4 depicts the proposed view-dependent PTQ. Although the quantizer is calibrated based on the visible pixels, POCA only runs PTQ once, and the resultant model will be deployed and rendered for all views." + } + ] + } + ], + "index": 10 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 418, + 91, + 447, + 100 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 418, + 91, + 447, + 100 + ], + "spans": [ + { + "bbox": [ + 418, + 91, + 447, + 100 + ], + "type": "text", + "content": "POCA" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 474, + 92, + 481, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 474, + 92, + 481, + 100 + ], + "spans": [ + { + "bbox": [ + 474, + 92, + 481, + 100 + ], + "type": "text", + "content": "9" + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 8 + }, + { + "para_blocks": [ + { + "bbox": [ + 132, + 116, + 367, + 129 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 116, + 367, + 129 + ], + "spans": [ + { + "bbox": [ + 132, + 116, + 367, + 129 + ], + "type": "text", + "content": "4.2 PTQ with Shape-wise Dynamic Filtering" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 130, + 137, + 482, + 209 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 137, + 482, + 209 + ], + "spans": [ + { + "bbox": [ + 130, + 137, + 482, + 209 + ], + "type": "text", + "content": "Besides the PTQ with the view-dependent visibility masking, we further suppress the impact of the dynamically long-tailed distribution by eliminating the pixels with a high channel-wise dynamic range, or \"Shape-Wise Dynamics\" (SWD). Given the full-precision output " + }, + { + "bbox": [ + 130, + 137, + 482, + 209 + ], + "type": "inline_equation", + "content": "\\mathcal{V}_{\\mathbf{T}}^{l}" + }, + { + "bbox": [ + 130, + 137, + 482, + 209 + ], + "type": "text", + "content": " with the shape of " + }, + { + "bbox": [ + 130, + 137, + 482, + 209 + ], + "type": "inline_equation", + "content": "\\mathcal{C} \\times \\mathcal{H} \\times \\mathcal{W}" + }, + { + "bbox": [ + 130, + 137, + 482, + 209 + ], + "type": "text", + "content": ", the SWD is characterized by the standard deviation score of each " + }, + { + "bbox": [ + 130, + 137, + 482, + 209 + ], + "type": "inline_equation", + "content": "\\mathcal{C} \\times 1 \\times 1" + }, + { + "bbox": [ + 130, + 137, + 482, + 209 + ], + "type": "text", + "content": " vector, which formulates the overall texture SWD map with the shape of " + }, + { + "bbox": [ + 130, + 137, + 482, + 209 + ], + "type": "inline_equation", + "content": "\\mathcal{H} \\times \\mathcal{W}" + }, + { + "bbox": [ + 130, + 137, + 482, + 209 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 130, + 209, + 482, + 281 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 209, + 482, + 281 + ], + "spans": [ + { + "bbox": [ + 130, + 209, + 482, + 281 + ], + "type": "text", + "content": "Subsequently, POCA generates the SWD mask " + }, + { + "bbox": [ + 130, + 209, + 482, + 281 + ], + "type": "inline_equation", + "content": "\\Theta" + }, + { + "bbox": [ + 130, + 209, + 482, + 281 + ], + "type": "text", + "content": " by filtering out the pixels in the SWD map with the highest tailedness. The sparsity of the SWD mask is controlled by a tunable hyperparameter " + }, + { + "bbox": [ + 130, + 209, + 482, + 281 + ], + "type": "inline_equation", + "content": "\\tau" + }, + { + "bbox": [ + 130, + 209, + 482, + 281 + ], + "type": "text", + "content": ". E.g., " + }, + { + "bbox": [ + 130, + 209, + 482, + 281 + ], + "type": "inline_equation", + "content": "\\tau = 0.7" + }, + { + "bbox": [ + 130, + 209, + 482, + 281 + ], + "type": "text", + "content": " means masking out the top " + }, + { + "bbox": [ + 130, + 209, + 482, + 281 + ], + "type": "inline_equation", + "content": "70\\%" + }, + { + "bbox": [ + 130, + 209, + 482, + 281 + ], + "type": "text", + "content": " pixels with the highest tailedness (standard deviation). The resultant SWD mask " + }, + { + "bbox": [ + 130, + 209, + 482, + 281 + ], + "type": "inline_equation", + "content": "\\Theta" + }, + { + "bbox": [ + 130, + 209, + 482, + 281 + ], + "type": "text", + "content": " will be multiplied with the view-dependent mask " + }, + { + "bbox": [ + 130, + 209, + 482, + 281 + ], + "type": "inline_equation", + "content": "\\Gamma" + }, + { + "bbox": [ + 130, + 209, + 482, + 281 + ], + "type": "text", + "content": " and generate the final learning objective of the layer-wise post-training quantization of POCA:" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 197, + 291, + 481, + 316 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 197, + 291, + 481, + 316 + ], + "spans": [ + { + "bbox": [ + 197, + 291, + 481, + 316 + ], + "type": "interline_equation", + "content": "\\underset {\\alpha^ {l}, S ^ {l}} {\\arg \\min } \\frac {1}{\\mathcal {H W C}} \\sum \\left(\\mathcal {Y} _ {\\mathbf {T}} ^ {l} \\odot \\Gamma^ {l} \\odot \\Theta^ {l} - \\hat {\\mathcal {Y}} _ {\\mathbf {T}} ^ {l} \\odot \\Gamma^ {l} \\odot \\Theta^ {l}\\right) ^ {2} \\tag {8}", + "image_path": "ae6b912749f65658aacd780c1de8e49ddcae793b9ca50936dfa414e313ab6a19.jpg" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 130, + 326, + 481, + 373 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 326, + 481, + 373 + ], + "spans": [ + { + "bbox": [ + 130, + 326, + 481, + 373 + ], + "type": "text", + "content": "As shown in Eq. (8), POCA optimizes the quantizers by learning the most relevant information (view-dependent) and least diversified pixels (SWD masking) of the avatar, which suppresses the drawbacks of the long-tailedness and high dynamics of the avatar." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 130, + 374, + 482, + 459 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 374, + 482, + 459 + ], + "spans": [ + { + "bbox": [ + 130, + 374, + 482, + 459 + ], + "type": "text", + "content": "The masking strategy of POCA introduces a high degree of sparsity in layer-wise PTQ, controlled by visibility and the SWD threshold " + }, + { + "bbox": [ + 130, + 374, + 482, + 459 + ], + "type": "inline_equation", + "content": "\\tau" + }, + { + "bbox": [ + 130, + 374, + 482, + 459 + ], + "type": "text", + "content": ". Our experimental results in Section 5 show that the optimal performance is achieved with " + }, + { + "bbox": [ + 130, + 374, + 482, + 459 + ], + "type": "inline_equation", + "content": "\\tau = 0.9" + }, + { + "bbox": [ + 130, + 374, + 482, + 459 + ], + "type": "text", + "content": ". As a result, POCA calibrates the low-precision layer based on highly sparse data " + }, + { + "bbox": [ + 130, + 374, + 482, + 459 + ], + "type": "inline_equation", + "content": "[\\mathcal{X}^l,\\mathcal{Y}_{\\mathbf{T}}^l]" + }, + { + "bbox": [ + 130, + 374, + 482, + 459 + ], + "type": "text", + "content": " with " + }, + { + "bbox": [ + 130, + 374, + 482, + 459 + ], + "type": "inline_equation", + "content": "\\geq 90\\%" + }, + { + "bbox": [ + 130, + 374, + 482, + 459 + ], + "type": "text", + "content": " sparsity. In other words, POCA only requires " + }, + { + "bbox": [ + 130, + 374, + 482, + 459 + ], + "type": "inline_equation", + "content": "\\leq 10\\%" + }, + { + "bbox": [ + 130, + 374, + 482, + 459 + ], + "type": "text", + "content": " of the activation for layer-wise PTQ, showing low memory overhead for the efficient PTQ calibration process." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 132, + 476, + 285, + 488 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 476, + 285, + 488 + ], + "spans": [ + { + "bbox": [ + 132, + 476, + 285, + 488 + ], + "type": "text", + "content": "4.3 Rationale behind POCA" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 130, + 497, + 482, + 617 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 497, + 482, + 617 + ], + "spans": [ + { + "bbox": [ + 130, + 497, + 482, + 617 + ], + "type": "text", + "content": "POCA alleviates the impact of rounding instability. With the proposed view-aware PTQ and SWD masking, POCA aims to optimize the layerwise quantizers with the most relevant and quantization-friendly features. The POCA-trained activation quantizer (LSQ) has a generally smaller scaling factor " + }, + { + "bbox": [ + 130, + 497, + 482, + 617 + ], + "type": "inline_equation", + "content": "S^l" + }, + { + "bbox": [ + 130, + 497, + 482, + 617 + ], + "type": "text", + "content": " compared to the PTQ without POCA, as shown in Figure 5. The reduced scaling factors formulate the low-precision grid with small intervals (scaling factor) between adjacent levels, alleviating the numerical differences between adjacent levels in the dequantized high-precision domain. As a result, the impact of the aforementioned rounding instability is minimized and the temporal noises become invisible, which proves the aforementioned hypothesis in Section 3.2." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 130, + 618, + 482, + 666 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 618, + 482, + 666 + ], + "spans": [ + { + "bbox": [ + 130, + 618, + 482, + 666 + ], + "type": "text", + "content": "Sparse masking " + }, + { + "bbox": [ + 130, + 618, + 482, + 666 + ], + "type": "inline_equation", + "content": "\\neq" + }, + { + "bbox": [ + 130, + 618, + 482, + 666 + ], + "type": "text", + "content": " biased learning. Although two types of sparse masking are introduced into PTQ, the POCA-trained quantizers will not lead to a biased representation in the quantized model. As shown in Figure 4, the sparsified calibration mask " + }, + { + "bbox": [ + 130, + 618, + 482, + 666 + ], + "type": "inline_equation", + "content": "(\\varGamma^l\\odot\\varTheta^l)" + }, + { + "bbox": [ + 130, + 618, + 482, + 666 + ], + "type": "text", + "content": " contains sparse but uniformly sampled information" + } + ] + } + ], + "index": 10 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "spans": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "type": "text", + "content": "10" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 166, + 91, + 223, + 102 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 166, + 91, + 223, + 102 + ], + "spans": [ + { + "bbox": [ + 166, + 91, + 223, + 102 + ], + "type": "text", + "content": "J. Meng et al." + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 9 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 192, + 114, + 421, + 215 + ], + "blocks": [ + { + "bbox": [ + 192, + 114, + 421, + 215 + ], + "lines": [ + { + "bbox": [ + 192, + 114, + 421, + 215 + ], + "spans": [ + { + "bbox": [ + 192, + 114, + 421, + 215 + ], + "type": "image", + "image_path": "3cc20d1972f613dc437500148f413b5b2e4e3394fd117e438146aaeb86142734.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 171, + 217, + 442, + 228 + ], + "lines": [ + { + "bbox": [ + 171, + 217, + 442, + 228 + ], + "spans": [ + { + "bbox": [ + 171, + 217, + 442, + 228 + ], + "type": "text", + "content": "Fig. 5: Scaling factor " + }, + { + "bbox": [ + 171, + 217, + 442, + 228 + ], + "type": "inline_equation", + "content": "S^l" + }, + { + "bbox": [ + 171, + 217, + 442, + 228 + ], + "type": "text", + "content": " of the texture decoder trained by POCA." + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_caption" + } + ], + "index": 2 + }, + { + "bbox": [ + 130, + 243, + 479, + 265 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 243, + 479, + 265 + ], + "spans": [ + { + "bbox": [ + 130, + 243, + 479, + 265 + ], + "type": "text", + "content": "across all the regions of the avatar texture. As a result, the learnable quantizers are uniformly trained with all the relevant facial information." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 130, + 266, + 480, + 326 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 266, + 480, + 326 + ], + "spans": [ + { + "bbox": [ + 130, + 266, + 480, + 326 + ], + "type": "text", + "content": "Highly-correlated data and high fidelity sparse learning. Unlike classification tasks, the training samples of the VAE model are highly correlated with respect to the single identity. As a result, the sparsified information of each sample can be compensated from other training data with different viewing perspectives, leading to sparse but robust PTQ with the quantized decoder." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 132, + 342, + 282, + 357 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 342, + 282, + 357 + ], + "spans": [ + { + "bbox": [ + 132, + 342, + 282, + 357 + ], + "type": "text", + "content": "5 Experimental Results" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 130, + 365, + 480, + 426 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 365, + 480, + 426 + ], + "spans": [ + { + "bbox": [ + 130, + 365, + 480, + 426 + ], + "type": "text", + "content": "We evaluated the proposed POCA algorithm based on Multiface [29], a public avatar dataset captured by the multi-camera system. POCA is applied to eight avatars among three viewing directions with all 65 facial expressions of the dataset. We want to highlight that the eight avatars comprehensively cover different genders, age groups, hairstyles, skin types, and facial hairs." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 130, + 426, + 481, + 510 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 426, + 481, + 510 + ], + "spans": [ + { + "bbox": [ + 130, + 426, + 481, + 510 + ], + "type": "text", + "content": "For each identity, we use the corresponding pre-trained deep appearance VAE model as the starting point of POCA. Following the standard PTQ strategy, we quantize all layers of the decoders (texture and mesh) iteratively, where each layer is calibrated by 100 epochs with the sampled data. POCA only acquires 512 training samples from each identity, which is " + }, + { + "bbox": [ + 130, + 426, + 481, + 510 + ], + "type": "inline_equation", + "content": "\\leq 5\\%" + }, + { + "bbox": [ + 130, + 426, + 481, + 510 + ], + "type": "text", + "content": " samples of the dataset. For each layer, the quantizers are optimized by Adam with 1e-4 learning rate and eight samples per batch." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 130, + 510, + 481, + 593 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 510, + 481, + 593 + ], + "spans": [ + { + "bbox": [ + 130, + 510, + 481, + 593 + ], + "type": "text", + "content": "Prior Works for Comparison. As the first work in the low-precision avatar decoding, we select three prior quantization works that reported the SoTA INT8 performance with the classification tasks: (1) Adaptive Rounding (AdaRound) [18], (2) learned step size quantization (LSQ) [2], which are also incorporated for hardware deployment [24], and (3) W8A8 naive QAT [20] in PyTorch. These selected prior works achieved superior INT8 performance [2, 18] on the ImageNet-1K dataset, even outperforming the full-precision baseline." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 130, + 594, + 481, + 665 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 594, + 481, + 665 + ], + "spans": [ + { + "bbox": [ + 130, + 594, + 481, + 665 + ], + "type": "text", + "content": "Evaluation Metric. We evaluate the proposed POCA algorithm based on the FovVideoVDP [16] between the rendered avatar and the ground truth video to quantify the temporal noise of the dynamic Codec Avatar. For each avatar identity, we compute the averaged FovVideoVDP score among all the 65 expressions in the Multiface dataset [29]. The performance of POCA is demonstrated with the rendered avatar videos in the supplementary material." + } + ] + } + ], + "index": 10 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 418, + 91, + 447, + 100 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 418, + 91, + 447, + 100 + ], + "spans": [ + { + "bbox": [ + 418, + 91, + 447, + 100 + ], + "type": "text", + "content": "POCA" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 470, + 91, + 480, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 470, + 91, + 480, + 100 + ], + "spans": [ + { + "bbox": [ + 470, + 91, + 480, + 100 + ], + "type": "text", + "content": "11" + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 10 + }, + { + "para_blocks": [ + { + "type": "table", + "bbox": [ + 132, + 168, + 481, + 253 + ], + "blocks": [ + { + "bbox": [ + 131, + 114, + 482, + 159 + ], + "lines": [ + { + "bbox": [ + 131, + 114, + 482, + 159 + ], + "spans": [ + { + "bbox": [ + 131, + 114, + 482, + 159 + ], + "type": "text", + "content": "Table 1: FovVideoVDP score [16] between the rendered avatar and the ground truth based on eight avatars across all the expressions from three viewing perspectives. The proposed POCA algorithm achieves the highest quality compared to the SoTA PTQ and QAT methods with both 8-bit and 6-bit precision." + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 132, + 168, + 481, + 253 + ], + "lines": [ + { + "bbox": [ + 132, + 168, + 481, + 253 + ], + "spans": [ + { + "bbox": [ + 132, + 168, + 481, + 253 + ], + "type": "table", + "html": "
Avatar IDPrecision26438142183941537202126453108870559667444329145896795937
Male, GoateeBlack, FemaleMohawk, MaleAsian, FemaleOlder WomanMale, Mid HairBlonde GirlMale, Short Hair
FrontTopFrontTopFrontTopFrontTopFrontTopFrontTopFrontTopFrontTop
Baseline [29]FP326.214.945.306.544.995.456.144.875.406.325.025.596.054.955.136.02
SoTA PTQ [2,18]W8A85.814.434.875.914.405.005.804.525.195.874.565.105.434.444.685.52
W8A8 QAT [20]5.924.554.996.024.575.125.884.565.235.974.665.165.574.504.795.60
POCA6.194.935.266.314.805.386.094.835.336.214.955.425.954.825.085.99
SoTA PTQ [2,18]W6A64.503.664.064.963.814.124.554.104.475.223.914.454.293.733.964.25
W6A6 QAT5.083.754.395.544.014.234.944.194.495.364.074.664.363.854.094.46
POCA6.064.555.026.294.425.136.014.615.166.024.665.215.454.494.995.66
", + "image_path": "d3782da200063a75a37548e56ea4c2713a8ed3892a3b85507a4b2de9b21a6c57.jpg" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "table_body" + } + ], + "index": 3 + }, + { + "type": "image", + "bbox": [ + 136, + 262, + 479, + 388 + ], + "blocks": [ + { + "bbox": [ + 169, + 254, + 275, + 262 + ], + "lines": [ + { + "bbox": [ + 169, + 254, + 275, + 262 + ], + "spans": [ + { + "bbox": [ + 169, + 254, + 275, + 262 + ], + "type": "text", + "content": "Yellow region " + }, + { + "bbox": [ + 169, + 254, + 275, + 262 + ], + "type": "inline_equation", + "content": "=" + }, + { + "bbox": [ + 169, + 254, + 275, + 262 + ], + "type": "text", + "content": " Mismatched pixel motion" + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 136, + 262, + 479, + 388 + ], + "lines": [ + { + "bbox": [ + 136, + 262, + 479, + 388 + ], + "spans": [ + { + "bbox": [ + 136, + 262, + 479, + 388 + ], + "type": "image", + "image_path": "dfc31f7dfbe95a7a21ecbb8634d9eeb7138c6e23f09312d14c0d85952b626fe9.jpg" + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 205, + 389, + 425, + 398 + ], + "lines": [ + { + "bbox": [ + 205, + 389, + 425, + 398 + ], + "spans": [ + { + "bbox": [ + 205, + 389, + 425, + 398 + ], + "type": "text", + "content": "FovVideoVDP [14] between the rendered avatar and ground truth (Max = 10.0)" + } + ] + } + ], + "index": 6, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 131, + 403, + 482, + 437 + ], + "lines": [ + { + "bbox": [ + 131, + 403, + 482, + 437 + ], + "spans": [ + { + "bbox": [ + 131, + 403, + 482, + 437 + ], + "type": "text", + "content": "Fig. 6: Rendered avatars and temporal pixel differences between two consecutive frames of the avatar video. The lightened yellow region of POCA signifies the removed temporal noise as the full precision baseline." + } + ] + } + ], + "index": 7, + "angle": 0, + "type": "image_caption" + } + ], + "index": 5 + }, + { + "bbox": [ + 131, + 461, + 392, + 474 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 131, + 461, + 392, + 474 + ], + "spans": [ + { + "bbox": [ + 131, + 461, + 392, + 474 + ], + "type": "text", + "content": "5.1 Temporal Noise Reduction with High Fidelity" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 130, + 485, + 482, + 568 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 485, + 482, + 568 + ], + "spans": [ + { + "bbox": [ + 130, + 485, + 482, + 568 + ], + "type": "text", + "content": "We evaluate the proposed POCA algorithm with eight avatars with different genders, hairstyles, and skin colors. Table 1 summarizes the FovVideoVDP score [16] of the proposed POCA algorithm with 8-bit and 6-bit avatar decoders. The SoTA INT8 PTQ quantization algorithm failed to preserve the avatar quality as the baseline full precision model, exhibiting the degraded FovVideoVDP score. Even with the expensive quantization-aware training (QAT), the quality of the rendered avatar cannot be restored." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 130, + 570, + 482, + 666 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 570, + 482, + 666 + ], + "spans": [ + { + "bbox": [ + 130, + 570, + 482, + 666 + ], + "type": "text", + "content": "Compared to the previous methods with 8-bit precision, POCA eliminates the temporal noises in the rendered avatar. POCA successfully compresses the decoder model while achieving on-par quality as the full precision decoder baseline. Figure 6 shows the high avatar quality and restored FovVideoVDP score (heat map) of the proposed POCA algorithm (closer to the FP32 baseline = better quantization). Regarding the efficiency of PTQ, POCA only requires " + }, + { + "bbox": [ + 130, + 570, + 482, + 666 + ], + "type": "inline_equation", + "content": "\\leq 10\\%" + }, + { + "bbox": [ + 130, + 570, + 482, + 666 + ], + "type": "text", + "content": " of the activations for layer-wise PTQ, showing low memory overhead for the efficient PTQ calibration process." + } + ] + } + ], + "index": 10 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "spans": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "type": "text", + "content": "12" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 166, + 91, + 223, + 102 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 166, + 91, + 223, + 102 + ], + "spans": [ + { + "bbox": [ + 166, + 91, + 223, + 102 + ], + "type": "text", + "content": "J. Meng et al." + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 11 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 168, + 114, + 221, + 202 + ], + "blocks": [ + { + "bbox": [ + 168, + 114, + 221, + 202 + ], + "lines": [ + { + "bbox": [ + 168, + 114, + 221, + 202 + ], + "spans": [ + { + "bbox": [ + 168, + 114, + 221, + 202 + ], + "type": "image", + "image_path": "07e345d143445ab9dc251b3dfb8a8c907ed40d885528d5bf819d149708673aa7.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 171, + 203, + 222, + 217 + ], + "lines": [ + { + "bbox": [ + 171, + 203, + 222, + 217 + ], + "spans": [ + { + "bbox": [ + 171, + 203, + 222, + 217 + ], + "type": "text", + "content": "SSIM=0.741 \nFovVideo" + }, + { + "bbox": [ + 171, + 203, + 222, + 217 + ], + "type": "inline_equation", + "content": "\\mathrm{DVP}=6.213" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_caption" + } + ], + "index": 2 + }, + { + "type": "image", + "bbox": [ + 222, + 114, + 276, + 202 + ], + "blocks": [ + { + "bbox": [ + 222, + 114, + 276, + 202 + ], + "lines": [ + { + "bbox": [ + 222, + 114, + 276, + 202 + ], + "spans": [ + { + "bbox": [ + 222, + 114, + 276, + 202 + ], + "type": "image", + "image_path": "4f149d2510c6863f4b7bf29d333575109256d8111bcaca2469a9a0109e6e185d.jpg" + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 225, + 203, + 276, + 217 + ], + "lines": [ + { + "bbox": [ + 225, + 203, + 276, + 217 + ], + "spans": [ + { + "bbox": [ + 225, + 203, + 276, + 217 + ], + "type": "text", + "content": "SSIM=0.741 \nFovVideoDVP=6.119" + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "image_caption" + } + ], + "index": 4 + }, + { + "type": "image", + "bbox": [ + 277, + 114, + 333, + 202 + ], + "blocks": [ + { + "bbox": [ + 277, + 114, + 333, + 202 + ], + "lines": [ + { + "bbox": [ + 277, + 114, + 333, + 202 + ], + "spans": [ + { + "bbox": [ + 277, + 114, + 333, + 202 + ], + "type": "image", + "image_path": "b09af302b5c3cae5805f4af93690fbc0b60e4a438927578e8106ae8f58a30cd0.jpg" + } + ] + } + ], + "index": 6, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 278, + 203, + 329, + 217 + ], + "lines": [ + { + "bbox": [ + 278, + 203, + 329, + 217 + ], + "spans": [ + { + "bbox": [ + 278, + 203, + 329, + 217 + ], + "type": "text", + "content": "SSIM=0.741 \nFovVideoVDP=1.685" + } + ] + } + ], + "index": 7, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 132, + 225, + 480, + 248 + ], + "lines": [ + { + "bbox": [ + 132, + 225, + 480, + 248 + ], + "spans": [ + { + "bbox": [ + 132, + 225, + 480, + 248 + ], + "type": "text", + "content": "Fig. 7: The impact of the SWD threshold. Extreme filtering leads to distorted and missing texture in the resultant avatar." + } + ] + } + ], + "index": 12, + "angle": 0, + "type": "image_caption" + } + ], + "index": 6 + }, + { + "type": "image", + "bbox": [ + 334, + 114, + 388, + 202 + ], + "blocks": [ + { + "bbox": [ + 334, + 114, + 388, + 202 + ], + "lines": [ + { + "bbox": [ + 334, + 114, + 388, + 202 + ], + "spans": [ + { + "bbox": [ + 334, + 114, + 388, + 202 + ], + "type": "image", + "image_path": "67276d87212373da8dc64a0adf798dfbc307d6ef36804bdbfda6b1889cec60bf.jpg" + } + ] + } + ], + "index": 8, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 337, + 203, + 387, + 217 + ], + "lines": [ + { + "bbox": [ + 337, + 203, + 387, + 217 + ], + "spans": [ + { + "bbox": [ + 337, + 203, + 387, + 217 + ], + "type": "text", + "content": "SSIM=0.740 \nFovVideoVDP=6.121" + } + ] + } + ], + "index": 9, + "angle": 0, + "type": "image_caption" + } + ], + "index": 8 + }, + { + "type": "image", + "bbox": [ + 389, + 114, + 446, + 202 + ], + "blocks": [ + { + "bbox": [ + 389, + 114, + 446, + 202 + ], + "lines": [ + { + "bbox": [ + 389, + 114, + 446, + 202 + ], + "spans": [ + { + "bbox": [ + 389, + 114, + 446, + 202 + ], + "type": "image", + "image_path": "be98ce8fa9a901c66f1f0e6b976be335103a3f511e8e23d0f714b5a20bd6fab4.jpg" + } + ] + } + ], + "index": 10, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 391, + 203, + 441, + 217 + ], + "lines": [ + { + "bbox": [ + 391, + 203, + 441, + 217 + ], + "spans": [ + { + "bbox": [ + 391, + 203, + 441, + 217 + ], + "type": "text", + "content": "SSIM=0.739 \nFoxVideoVDP=6.106" + } + ] + } + ], + "index": 11, + "angle": 0, + "type": "image_caption" + } + ], + "index": 10 + }, + { + "type": "table", + "bbox": [ + 165, + 264, + 445, + 323 + ], + "blocks": [ + { + "bbox": [ + 178, + 251, + 434, + 262 + ], + "lines": [ + { + "bbox": [ + 178, + 251, + 434, + 262 + ], + "spans": [ + { + "bbox": [ + 178, + 251, + 434, + 262 + ], + "type": "text", + "content": "Table 2: Training time comparison between POCA and QAT." + } + ] + } + ], + "index": 13, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 165, + 264, + 445, + 323 + ], + "lines": [ + { + "bbox": [ + 165, + 264, + 445, + 323 + ], + "spans": [ + { + "bbox": [ + 165, + 264, + 445, + 323 + ], + "type": "table", + "html": "
MethodQuantizationMethodW/ATimeGPUDecoder Size (MB)
FP32 Baseline-32/32---32.78
Native W8A8 [20]QAT8/8>10 hrs2 × A60008.19 (↓ 4×)
POCA (This work)PTQ8/8~13 min1 × A60008.19 (↓ 4×)
POCA (This work)PTQ6/6~13 min1 × A60006.15 (↓ 5.3×)
", + "image_path": "b61a4cc34dd6833349360b27eed229d29c2e0699529ee9d2b7b12c21cf361a8c.jpg" + } + ] + } + ], + "index": 14, + "angle": 0, + "type": "table_body" + } + ], + "index": 14 + }, + { + "bbox": [ + 130, + 346, + 482, + 441 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 346, + 482, + 441 + ], + "spans": [ + { + "bbox": [ + 130, + 346, + 482, + 441 + ], + "type": "text", + "content": "POCA generates clean avatars without introducing any graphic filtering techniques [31]. From the perspective of quantization, POCA keeps the vanilla quantization strategy in Eq. (1) in both training and inference, compressing the model without introducing distribution manipulations [28] on the additional scaling [30] and bias [12]. Compared to traditional QAT, POCA quantizes the entire model in only 13 minutes (Table 2). The fast deployment workflow with the actual VR device (e.g., Quest Pro) can further accelerate the process of \"user-to-meta verse\"." + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 130, + 443, + 493, + 503 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 443, + 493, + 503 + ], + "spans": [ + { + "bbox": [ + 130, + 443, + 493, + 503 + ], + "type": "text", + "content": "Future Applicability: POCA is not limited to the deep appearance model [14]. The proposed POCA algorithm is applicable to all mesh-based 3D reconstruction tasks with VAE models. For example, the proposed view-dependent visibility mask and shape-wise filtering can be easily transferred to the recent Pixel Codec Avatar (PiCA)[15], which has the similar autoencoder-based architecture." + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 132, + 524, + 361, + 537 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 524, + 361, + 537 + ], + "spans": [ + { + "bbox": [ + 132, + 524, + 361, + 537 + ], + "type": "text", + "content": "5.2 Impact of the SWD Filtering Threshold" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 130, + 548, + 482, + 644 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 548, + 482, + 644 + ], + "spans": [ + { + "bbox": [ + 130, + 548, + 482, + 644 + ], + "type": "text", + "content": "POCA introduced the tunable parameter " + }, + { + "bbox": [ + 130, + 548, + 482, + 644 + ], + "type": "inline_equation", + "content": "\\tau" + }, + { + "bbox": [ + 130, + 548, + 482, + 644 + ], + "type": "text", + "content": " in Eq. (8) to control the amount of activation pixels for layer-wise PTQ. Figure 7 demonstrates the impact of the SWD threshold " + }, + { + "bbox": [ + 130, + 548, + 482, + 644 + ], + "type": "inline_equation", + "content": "\\tau" + }, + { + "bbox": [ + 130, + 548, + 482, + 644 + ], + "type": "text", + "content": " with respect to the avatar quality (SSIM). POCA achieves the best tradeoff between filter intensity and avatar quality with " + }, + { + "bbox": [ + 130, + 548, + 482, + 644 + ], + "type": "inline_equation", + "content": "\\tau^{*} = 0.9" + }, + { + "bbox": [ + 130, + 548, + 482, + 644 + ], + "type": "text", + "content": ". Quantizing the decoder with " + }, + { + "bbox": [ + 130, + 548, + 482, + 644 + ], + "type": "inline_equation", + "content": "\\tau > \\tau^{*}" + }, + { + "bbox": [ + 130, + 548, + 482, + 644 + ], + "type": "text", + "content": " leads to missing texture pixels and distorted rendering with no quality improvement. The selected SWD threshold exhibits low sensitivity between different avatars and viewing perspectives, as demonstrated by the high-quality rendering in Figure 6." + } + ] + } + ], + "index": 18 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 418, + 91, + 447, + 100 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 418, + 91, + 447, + 100 + ], + "spans": [ + { + "bbox": [ + 418, + 91, + 447, + 100 + ], + "type": "text", + "content": "POCA" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 470, + 91, + 481, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 470, + 91, + 481, + 100 + ], + "spans": [ + { + "bbox": [ + 470, + 91, + 481, + 100 + ], + "type": "text", + "content": "13" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 133, + 653, + 480, + 665 + ], + "type": "page_footnote", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 653, + 480, + 665 + ], + "spans": [ + { + "bbox": [ + 133, + 653, + 480, + 665 + ], + "type": "text", + "content": "3 PiCA results are available at https://mengjian0502.github.io/poca.github.io/" + } + ] + } + ], + "index": 19 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 12 + }, + { + "para_blocks": [ + { + "bbox": [ + 132, + 116, + 416, + 128 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 116, + 416, + 128 + ], + "spans": [ + { + "bbox": [ + 132, + 116, + 416, + 128 + ], + "type": "text", + "content": "5.3 Applicability to Other Quality Enhancement Tasks" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 130, + 133, + 482, + 170 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 133, + 482, + 170 + ], + "spans": [ + { + "bbox": [ + 130, + 133, + 482, + 170 + ], + "type": "text", + "content": "We extend the experiment by evaluating the proposed method on super-resolution tasks. The proposed \"shape-wise-filtering\" (SWD) exhibits strong performance in image quality enhancement [25] with the low-precision SRResNet model [8]." + } + ] + } + ], + "index": 3 + }, + { + "type": "table", + "bbox": [ + 140, + 201, + 468, + 347 + ], + "blocks": [ + { + "bbox": [ + 132, + 176, + 481, + 199 + ], + "lines": [ + { + "bbox": [ + 132, + 176, + 481, + 199 + ], + "spans": [ + { + "bbox": [ + 132, + 176, + 481, + 199 + ], + "type": "text", + "content": "Table 3: SoTA image quality (PSNR/SSIM) achieved by the proposed POCA-SWD algorithm with 8-bit and 6-bit model precision, outperforms the recent SR-PTQ [25]." + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 140, + 201, + 468, + 347 + ], + "lines": [ + { + "bbox": [ + 140, + 201, + 468, + 347 + ], + "spans": [ + { + "bbox": [ + 140, + 201, + 468, + 347 + ], + "type": "table", + "html": "
MethodPrecisionSparsitySet5 (× 4)Set14 (× 4)BSD100 (× 4)Urban100 (× 4)
SRResNet Baseline [25]FP32(Baseline)N/A32.234 / 0.89628.656 / 0.78427.63 / 0.73826.229 / 0.791
SRResNet [8]32.05 / 0.90128.49 / 0.81827.58 / 0.762Not Reported
SRResNet (This work)31.927 / 0.90128.588 / 0.79927.628 / 0.75626.223 / 0.800
AdaRound [18] + LSQ [2]W8A8Dense31.870 / 0.89928.550 / 0.79827.533 / 0.75426.078 / 0.797
SR-PTQ [25]Dense32.207 / 0.89528.619 / 0.78327.618 / 0.73826.191 / 0.790
POCA80%31.898 / 0.90028.566 / 0.79827.620 / 0.75626.195 / 0.799
Quality Drop [2,18]W8A80.057 / 0.0020.038 / 0.0010.095 / 0.0020.145 / 0.003
Quality Drop [25]0.027 / 0.0010.037 / 0.0010.012 / 0.0000.038 / 0.001
Quality Drop (POCA)0.029 / 0.0010.022 / 0.0010.008 / 0.0000.028 / 0.001
AdaRound [18] + LSQ [2]W6A6Dense31.765 / 0.89728.449 / 0.79527.470 / 0.75225.988 / 0.794
SR-PTQ [25]Dense32.089 / 0.89228.504 / 0.77927.561 / 0.73326.011 / 0.783
POCA80%31.789 / 0.89928.499 / 0.79627.565 / 0.75326.110 / 0.795
Quality Drop [2,18]W6A60.162 / 0.0040.139 / 0.0030.158 / 0.0040.235 / 0.006
Quality Drop [25]0.145 / 0.0040.152 / 0.0050.069 / 0.0050.218 / 0.008
Quality Drop (POCA)0.138 / 0.0020.089 / 0.0040.061 / 0.0030.113 / 0.005
", + "image_path": "6ac2ec29a1f4c966cb6daa522cb79e6ff95f1a3531b883bbc567e46cb14f5737.jpg" + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "table_body" + } + ], + "index": 5 + }, + { + "bbox": [ + 130, + 356, + 482, + 440 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 356, + 482, + 440 + ], + "spans": [ + { + "bbox": [ + 130, + 356, + 482, + 440 + ], + "type": "text", + "content": "With 8-bit and 6-bit precision, the proposed method outperforms [25] in BSD100 and Urban100 datasets, as shown in Table 3. For a fair comparison, we listed all the baseline full-precision models. Our method outperforms [25] on all datasets with the new SoTA image quality. Compared to [18] and [2], the improved image quality implies the effectiveness of the proposed SWD. More importantly, our method enables the sparse PTQ by using only " + }, + { + "bbox": [ + 130, + 356, + 482, + 440 + ], + "type": "inline_equation", + "content": "20\\%" + }, + { + "bbox": [ + 130, + 356, + 482, + 440 + ], + "type": "text", + "content": " pixels of the selected samples, whereas all prior works [2,18,25] require dense PTQ." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 132, + 455, + 311, + 467 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 455, + 311, + 467 + ], + "spans": [ + { + "bbox": [ + 132, + 455, + 311, + 467 + ], + "type": "text", + "content": "6 Conclusion and Discussion" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 130, + 477, + 481, + 597 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 477, + 481, + 597 + ], + "spans": [ + { + "bbox": [ + 130, + 477, + 481, + 597 + ], + "type": "text", + "content": "In this paper, we propose a novel post-training quantization algorithm designed for high-quality Codec Avatar rendering. As the first work in this area, we discover the challenge of temporal noise with low-precision avatar decoding. Through thorough investigation, we reveal the root cause of the temporal noise that originated from the high dynamics and long-tailedness of the texture distribution. To address that, we propose POCA, which consists of view-aware PTQ and shape-wise filtering (SWD). Compared to the previous SoTA methods, POCA eliminates the temporal noises even with 6-bit precision. Furthermore, the proposed SWD algorithm is also applicable to super-resolution tasks, achieving the new SoTA image quality with highly efficient sparse calibration." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 130, + 605, + 481, + 665 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 605, + 481, + 665 + ], + "spans": [ + { + "bbox": [ + 130, + 605, + 481, + 665 + ], + "type": "text", + "content": "Limitations. The proposed POCA algorithm mainly focuses on the VAE-based Codec Avatar generation, which has been proven as a successful strategy with the actual commercial-level VR product deployment. Recently, 3D Gaussian-based avatar rendering also emerged as a feasible method for generative Avatars, while low-precision quantization of such models will remain as future work." + } + ] + } + ], + "index": 9 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "spans": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "type": "text", + "content": "14" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 166, + 91, + 223, + 102 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 166, + 91, + 223, + 102 + ], + "spans": [ + { + "bbox": [ + 166, + 91, + 223, + 102 + ], + "type": "text", + "content": "J. Meng et al." + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 13 + }, + { + "para_blocks": [ + { + "bbox": [ + 132, + 114, + 198, + 126 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 114, + 198, + 126 + ], + "spans": [ + { + "bbox": [ + 132, + 114, + 198, + 126 + ], + "type": "text", + "content": "References" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 138, + 142, + 481, + 665 + ], + "type": "list", + "angle": 0, + "index": 17, + "blocks": [ + { + "bbox": [ + 138, + 142, + 481, + 176 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 142, + 481, + 176 + ], + "spans": [ + { + "bbox": [ + 138, + 142, + 481, + 176 + ], + "type": "text", + "content": "1. Choi, J., Venkataramani, S., Srinivasan, V.V., Gopalakrishnan, K., Wang, Z., Chuang, P.: Accurate and efficient 2-bit quantized neural networks. Proceedings of Machine Learning and Systems 1, 348-359 (2019) 5" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 138, + 176, + 481, + 209 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 176, + 481, + 209 + ], + "spans": [ + { + "bbox": [ + 138, + 176, + 481, + 209 + ], + "type": "text", + "content": "2. Esser, S.K., McKinstry, J.L., Bablani, D., Appuswamy, R., Modha, D.S.: Learned step size quantization. In: International Conference on Learning Representations (ICLR) (2020) 2, 3, 5, 6, 8, 11, 12, 14" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 138, + 209, + 481, + 253 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 209, + 481, + 253 + ], + "spans": [ + { + "bbox": [ + 138, + 209, + 481, + 253 + ], + "type": "text", + "content": "3. Fu, Y., Li, Y., Li, C., Saragih, J., Zhang, P., Dai, X., Lin, Y.C.: Auto-card: Efficient and robust codec avatar driving for real-time mobile telepresence. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR) (2023) 4" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 138, + 254, + 481, + 298 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 254, + 481, + 298 + ], + "spans": [ + { + "bbox": [ + 138, + 254, + 481, + 298 + ], + "type": "text", + "content": "4. Gorbachev, Y., Fedorov, M., Slavutin, I., Tugarev, A., Fatekhov, M., Tarkan, Y.: Openvino deep learning workbench: Comprehensive analysis and tuning of neural networks inference. In: Proceedings of the IEEE/CVF International Conference on Computer Vision Workshops (CVPR-W) (2019) 4" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 138, + 299, + 481, + 331 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 299, + 481, + 331 + ], + "spans": [ + { + "bbox": [ + 138, + 299, + 481, + 331 + ], + "type": "text", + "content": "5. Han, Y., Yang, S., Wang, W., Liu, J.: From design draft to real attire: Unaligned fashion image translation. In: Proceedings of the 28th ACM International Conference on Multimedia. p. 1533-1541 (2020) 4" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 138, + 332, + 481, + 354 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 332, + 481, + 354 + ], + "spans": [ + { + "bbox": [ + 138, + 332, + 481, + 354 + ], + "type": "text", + "content": "6. Kingma, D.P., Welling, M.: Auto-encoding variational bayes. arXiv preprint arXiv:1312.6114 (2013) 4" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 138, + 354, + 481, + 399 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 354, + 481, + 399 + ], + "spans": [ + { + "bbox": [ + 138, + 354, + 481, + 399 + ], + "type": "text", + "content": "7. Le, H., Zhang, L., Said, A., Sautiere, G., Yang, Y., Shrestha, P., Yin, F., Pourreza, R., Wiggers, A.: Mobilecodec: neural inter-frame video compression on mobile devices. In: Proceedings of the 13th ACM Multimedia Systems Conference. pp. 324-330 (2022) 5" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 138, + 399, + 481, + 453 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 399, + 481, + 453 + ], + "spans": [ + { + "bbox": [ + 138, + 399, + 481, + 453 + ], + "type": "text", + "content": "8. Ledig, C., Theis, L., Huszár, F., Caballero, J., Cunningham, A., Acosta, A., Aitken, A., Tejani, A., Totz, J., Wang, Z., et al.: Photo-realistic single image superresolution using a generative adversarial network. In: Proceedings of the IEEE conference on computer vision and pattern recognition. pp. 4681-4690 (2017) 3, 14" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 138, + 454, + 481, + 487 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 454, + 481, + 487 + ], + "spans": [ + { + "bbox": [ + 138, + 454, + 481, + 487 + ], + "type": "text", + "content": "9. Lewis, J.P., Anjyo, K., Rhee, T., Zhang, M., Pighin, F.H., Deng, Z.: Practice and theory of blendshape facial models. Eurographics (State of the Art Reports) 1(8), 2 (2014) 4" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 138, + 487, + 481, + 520 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 487, + 481, + 520 + ], + "spans": [ + { + "bbox": [ + 138, + 487, + 481, + 520 + ], + "type": "text", + "content": "0. Li, Y., Dong, X., Wang, W.: Additive powers-of-two quantization: An efficient non-uniform discretization for neural networks. In: International Conference on Learning Representations (ICLR) (2020) 5" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 138, + 521, + 481, + 553 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 521, + 481, + 553 + ], + "spans": [ + { + "bbox": [ + 138, + 521, + 481, + 553 + ], + "type": "text", + "content": "1. Li, Y., Gong, R., Tan, X., Yang, Y., Hu, P., Zhang, Q., Yu, F., Wang, W., Gu, S.: {BRECQ}: Pushing the limit of post-training quantization by block reconstruction. In: International Conference on Learning Representations (ICLR) (2021) 2, 5" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 138, + 554, + 481, + 598 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 554, + 481, + 598 + ], + "spans": [ + { + "bbox": [ + 138, + 554, + 481, + 598 + ], + "type": "text", + "content": "2. Liu, Y., Yang, H., Dong, Z., Keutzer, K., Du, L., Zhang, S.: Noisyquant: Noisy bias-enhanced post-training activation quantization for vision transformers. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR) (2023) 3, 5, 6, 13" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 138, + 599, + 481, + 643 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 599, + 481, + 643 + ], + "spans": [ + { + "bbox": [ + 138, + 599, + 481, + 643 + ], + "type": "text", + "content": "3. Liu, Z., Cheng, K.T., Huang, D., Xing, E.P., Shen, Z.: Nonuniform-to-uniform quantization: Towards accurate quantization via generalized straight-through estimation. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR) (2022) 5" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 138, + 643, + 481, + 665 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 643, + 481, + 665 + ], + "spans": [ + { + "bbox": [ + 138, + 643, + 481, + 665 + ], + "type": "text", + "content": "4. Lombardi, S., Saragih, J., Simon, T., Sheikh, Y.: Deep appearance models for face rendering. ACM Transactions on Graphics (ToG) (2018) 1, 2, 4, 7, 13" + } + ] + } + ], + "index": 16 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 418, + 91, + 447, + 100 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 418, + 91, + 447, + 100 + ], + "spans": [ + { + "bbox": [ + 418, + 91, + 447, + 100 + ], + "type": "text", + "content": "POCA" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 470, + 91, + 481, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 470, + 91, + 481, + 100 + ], + "spans": [ + { + "bbox": [ + 470, + 91, + 481, + 100 + ], + "type": "text", + "content": "15" + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 14 + }, + { + "para_blocks": [ + { + "bbox": [ + 132, + 116, + 481, + 665 + ], + "type": "list", + "angle": 0, + "index": 17, + "blocks": [ + { + "bbox": [ + 133, + 116, + 481, + 149 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 116, + 481, + 149 + ], + "spans": [ + { + "bbox": [ + 133, + 116, + 481, + 149 + ], + "type": "text", + "content": "15. Ma, S., Simon, T., Saragih, J., Wang, D., Li, Y., De La Torre, F., Sheikh, Y.: Pixel codec avatars. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR) (2021) 1, 2, 4, 13" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 133, + 150, + 481, + 194 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 150, + 481, + 194 + ], + "spans": [ + { + "bbox": [ + 133, + 150, + 481, + 194 + ], + "type": "text", + "content": "16. Mantiuk, R.K., Denes, G., Shapiro, A., Kaplanyan, A., Rufo, G., Bachy, R., Lian, T., Patney, A.: FovVideoVDP: A visible difference predictor for wide field-of-view video. In: ACM Transactions on Graphics (TOG). ACM New York, NY, USA (2021) 3, 6, 11, 12" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 132, + 194, + 481, + 226 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 194, + 481, + 226 + ], + "spans": [ + { + "bbox": [ + 132, + 194, + 481, + 226 + ], + "type": "text", + "content": "17. Mildenhall, B., Srinivasan, P.P., Tancik, M., Barron, J.T., Ramamoorthi, R., Ng, R.: NeRF: Representing scenes as neural radiance fields for view synthesis. Communications of the ACM 65(1), 99-106 (2021) 5" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 132, + 227, + 481, + 259 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 227, + 481, + 259 + ], + "spans": [ + { + "bbox": [ + 132, + 227, + 481, + 259 + ], + "type": "text", + "content": "18. Nagel, M., Amjad, R.A., Van Baalen, M., Louizos, C., Blankevoort, T.: Up or down? adaptive rounding for post-training quantization. In: International Conference on Machine Learning (ICML) (2020) 2, 3, 5, 6, 8, 11, 12, 14" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 132, + 260, + 481, + 282 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 260, + 481, + 282 + ], + "spans": [ + { + "bbox": [ + 132, + 260, + 481, + 282 + ], + "type": "text", + "content": "19. Park, E., Yoo, S.: Profit: A novel training method for sub-4-bit mobilenet models. In: European Conference on Computer Vision (ECCV) (2020) 6" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 132, + 282, + 481, + 335 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 282, + 481, + 335 + ], + "spans": [ + { + "bbox": [ + 132, + 282, + 481, + 335 + ], + "type": "text", + "content": "20. Paszke, A., Gross, S., Massa, F., Lerer, A., Bradbury, J., Chanan, G., Killeen, T., Lin, Z., Gimelshein, N., Antiga, L., Desmaison, A., Kopf, A., Yang, E., DeVito, Z., Raison, M., Tejani, A., Chilamkurthy, S., Steiner, B., Fang, L., Bai, J., Chintala, S.: Pytorch: An imperative style, high-performance deep learning library. In: Advances in Neural Information Processing Systems (NeurIPS) (2019) 3, 11, 12, 13" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 132, + 336, + 481, + 380 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 336, + 481, + 380 + ], + "spans": [ + { + "bbox": [ + 132, + 336, + 481, + 380 + ], + "type": "text", + "content": "21. van Rozendaal, T., Singhal, T., Le, H., Sautiere, G., Said, A., Buska, K., Raha, A., Kalatzis, D., Mehta, H., Mayer, F., et al.: Mobilenvc: Real-time 1080p neural video compression on a mobile device. In: Proceedings of the IEEE/CVF Winter Conference on Applications of Computer Vision. pp. 4323-4333 (2024) 5" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 132, + 380, + 481, + 413 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 380, + 481, + 413 + ], + "spans": [ + { + "bbox": [ + 132, + 380, + 481, + 413 + ], + "type": "text", + "content": "22. Saito, S., Schwartz, G., Simon, T., Li, J., Nam, G.: Relightable gaussian codec avatars. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 130-141 (2024) 2" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 132, + 413, + 481, + 446 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 413, + 481, + 446 + ], + "spans": [ + { + "bbox": [ + 132, + 413, + 481, + 446 + ], + "type": "text", + "content": "23. Schwartz, G., Wei, S.E., Wang, T.L., Lombardi, S., Simon, T., Saragih, J., Sheikh, Y.: The eyes have it: An integrated eye and face model for photorealistic facial animation. ACM Transactions on Graphics (TOG) (2020) 1" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 132, + 447, + 481, + 479 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 447, + 481, + 479 + ], + "spans": [ + { + "bbox": [ + 132, + 447, + 481, + 479 + ], + "type": "text", + "content": "24. Siddegowda, S., Fournarakis, M., Nagel, M., Blankevoort, T., Patel, C., Khobare, A.: Neural network quantization with AI model efficiency toolkit (AIMET). arXiv preprint arXiv:2201.08442 (2022) 2, 3, 4, 11" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 132, + 479, + 481, + 512 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 479, + 481, + 512 + ], + "spans": [ + { + "bbox": [ + 132, + 479, + 481, + 512 + ], + "type": "text", + "content": "25. Tu, Z., Hu, J., Chen, H., Wang, Y.: Toward accurate post-training quantization for image super resolution. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR) (2023) 2, 3, 6, 14" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 132, + 512, + 481, + 544 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 512, + 481, + 544 + ], + "spans": [ + { + "bbox": [ + 132, + 512, + 481, + 544 + ], + "type": "text", + "content": "26. Wang, Z., Wu, Z., Lu, J., Zhou, J.: Bidet: An efficient binarized object detector. In: Proceedings of the IEEE/CVF conference on computer vision and pattern recognition (CVPR) (2020) 2" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 132, + 545, + 481, + 578 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 545, + 481, + 578 + ], + "spans": [ + { + "bbox": [ + 132, + 545, + 481, + 578 + ], + "type": "text", + "content": "27. Wei, S.E., Saragih, J., Simon, T., Harley, A.W., Lombardi, S., Perdoch, M., Hypes, A., Wang, D., Badino, H., Sheikh, Y.: VR facial animation via multiview image translation. ACM Transactions on Graphics (TOG) (2019) 1" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 132, + 578, + 481, + 610 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 578, + 481, + 610 + ], + "spans": [ + { + "bbox": [ + 132, + 578, + 481, + 610 + ], + "type": "text", + "content": "28. Wei, X., Gong, R., Li, Y., Liu, X., Yu, F.: QDrop: Randomly dropping quantization for extremely low-bit post-training quantization. In: International Conference on Learning Representations (ICLR) (2022) 2, 5, 13" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 132, + 611, + 481, + 665 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 611, + 481, + 665 + ], + "spans": [ + { + "bbox": [ + 132, + 611, + 481, + 665 + ], + "type": "text", + "content": "29. Wuu, C.h., Zheng, N., Ardisson, S., Bali, R., Belko, D., Brockmeyer, E., Evans, L., Godisart, T., Ha, H., Huang, X., Hypes, A., Koska, T., Krenn, S., Lombardi, S., Luo, X., McPhail, K., Millerschoen, L., Perdoch, M., Pitts, M., Richard, A., Saragih, J., Saragih, J., Shiratori, T., Simon, T., Stewart, M., Trimble, A., Weng, X., Whitewolf, D., Wu, C., Yu, S.I., Sheikh, Y.: Multiface: A dataset for neural" + } + ] + } + ], + "index": 16 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "spans": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "type": "text", + "content": "16" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 166, + 91, + 223, + 102 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 166, + 91, + 223, + 102 + ], + "spans": [ + { + "bbox": [ + 166, + 91, + 223, + 102 + ], + "type": "text", + "content": "J. Meng et al." + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 15 + }, + { + "para_blocks": [ + { + "bbox": [ + 132, + 116, + 481, + 226 + ], + "type": "list", + "angle": 0, + "index": 6, + "blocks": [ + { + "bbox": [ + 147, + 116, + 481, + 138 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 147, + 116, + 481, + 138 + ], + "spans": [ + { + "bbox": [ + 147, + 116, + 481, + 138 + ], + "type": "text", + "content": "face rendering. In: arXiv (2022). https://doi.org/10.48550/ARXIV.2207.11243, https://arxiv.org/abs/2207.11243 1, 11, 12" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 133, + 140, + 481, + 171 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 140, + 481, + 171 + ], + "spans": [ + { + "bbox": [ + 133, + 140, + 481, + 171 + ], + "type": "text", + "content": "30. Xiao, G., Lin, J., Seznek, M., Wu, H., Demouth, J., Han, S.: Smoothquant: Accurate and efficient post-training quantization for large language models. In: International Conference on Machine Learning (ICML) (2023) 2, 5, 13" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 132, + 172, + 481, + 193 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 172, + 481, + 193 + ], + "spans": [ + { + "bbox": [ + 132, + 172, + 481, + 193 + ], + "type": "text", + "content": "31. Yin, H., Gong, Y., Qiu, G.: Side window filtering. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR) (2019) 13" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 132, + 194, + 481, + 226 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 194, + 481, + 226 + ], + "spans": [ + { + "bbox": [ + 132, + 194, + 481, + 226 + ], + "type": "text", + "content": "32. Zhang, Y., Zhang, Z., Lew, L.: Pokebnn: A binary pursuit of lightweight accuracy. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR) (2022) 4" + } + ] + } + ], + "index": 5 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 418, + 91, + 447, + 100 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 418, + 91, + 447, + 100 + ], + "spans": [ + { + "bbox": [ + 418, + 91, + 447, + 100 + ], + "type": "text", + "content": "POCA" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 470, + 91, + 481, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 470, + 91, + 481, + 100 + ], + "spans": [ + { + "bbox": [ + 470, + 91, + 481, + 100 + ], + "type": "text", + "content": "17" + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 16 + } + ], + "_backend": "vlm", + "_version_name": "2.6.4" +} \ No newline at end of file diff --git a/2024/POET_ Prompt Offset Tuning for Continual Human Action Adaptation/9942db11-78b9-4c6d-815f-c44748e2ea7c_content_list.json b/2024/POET_ Prompt Offset Tuning for Continual Human Action Adaptation/9942db11-78b9-4c6d-815f-c44748e2ea7c_content_list.json new file mode 100644 index 0000000000000000000000000000000000000000..2b12561ed4747be990698599e2b45c51430af42d --- /dev/null +++ b/2024/POET_ Prompt Offset Tuning for Continual Human Action Adaptation/9942db11-78b9-4c6d-815f-c44748e2ea7c_content_list.json @@ -0,0 +1,1749 @@ +[ + { + "type": "text", + "text": "POET: Prompt Offset Tuning for Continual Human Action Adaptation", + "text_level": 1, + "bbox": [ + 250, + 140, + 754, + 186 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Prachi Garg $^{1}$ , K J Joseph $^{3}$ , Vineeth N Balasubramanian $^{3}$ , Necati Cihan Camgoz $^{2}$ , Chengde Wan $^{2}$ , Kenrick Kin $^{2}$ , Weiguang Si $^{2}$ , Shugao Ma $^{2}$ , and Fernando De La Torre $^{1}$", + "bbox": [ + 236, + 212, + 766, + 258 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "1 Carnegie Mellon University, USA", + "bbox": [ + 380, + 268, + 620, + 285 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Meta Reality Labs", + "bbox": [ + 431, + 285, + 571, + 297 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "3 Indian Institute of Technology, Hyderabad", + "bbox": [ + 352, + 297, + 650, + 311 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Abstract. As extended reality (XR) is redefining how users interact with computing devices, research in human action recognition is gaining prominence. Typically, models deployed on immersive computing devices are static and limited to their default set of classes. The goal of our research is to provide users and developers with the capability to personalize their experience by adding new action classes to their device models continually. Importantly, a user should be able to add new classes in a low-shot and efficient manner, while this process should not require storing or replaying any of user's sensitive training data. We formalize this problem as privacy-aware few-shot continual action recognition. Towards this end, we propose POET: Prompt-offset Tuning. While existing prompt tuning approaches have shown great promise for continual learning of image, text, and video modalities; they demand access to extensively pretrained transformers. Breaking away from this assumption, POET demonstrates the efficacy of prompt tuning a significantly lightweight backbone, pretrained exclusively on the base class data. We propose a novel spatio-temporal learnable prompt offset tuning approach, and are the first to apply such prompt tuning to Graph Neural Networks. We contribute two new benchmarks for our new problem setting in human action recognition: (i) NTU RGB+D dataset for activity recognition, and (ii) SHREC-2017 dataset for hand gesture recognition. We find that POET consistently outperforms comprehensive benchmarks.", + "bbox": [ + 259, + 356, + 743, + 661 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Keywords: 3D Skeleton Activity Recognition $\\cdot$ Extended Reality (XR) $\\cdot$ Continual Learning $\\cdot$ Prompt Tuning.", + "bbox": [ + 261, + 674, + 740, + 703 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "1 Introduction", + "text_level": 1, + "bbox": [ + 215, + 733, + 375, + 750 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "A key input modality to virtual, augmented and mixed reality (often together termed as extended reality, XR) devices today is through recognizing human", + "bbox": [ + 215, + 768, + 787, + 799 + ], + "page_idx": 0 + }, + { + "type": "page_footnote", + "text": "Source Code at https://github.com/humansensinglab/POET-continual-action-recognition", + "bbox": [ + 217, + 810, + 785, + 839 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/30acfb9b2daf18ec6aef71a76a4b9c3a515194d19662e3e8544a9a71a9e5b0cf.jpg", + "image_caption": [ + "Fig. 1: Proposed POET method continually adapts skeleton-based human action recognition models pretrained on a pre-defined set of categories to new user categories with few training examples. Users can thus expand the capabilities of XR systems with novel action classes by providing a few examples of each new class. We discard the user-sensitive data as soon as the model is updated on the new categories." + ], + "image_footnote": [], + "bbox": [ + 222, + 143, + 782, + 277 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "activity and hand gestures based on body and hand pose estimates. Recognizing human actions5 facilitates seamless user interactions in head-mounted XR devices such as the Meta Quest 3 and Apple Vision Pro. If the provided action recognition models are static, then developers and users are limited to a predefined set of action categories. With the growing use of such devices in new contexts and the increasing demand for personalized technology delivery, there is an impending need to enable the action recognition models in such systems to adapt and learn new user actions over time. Defining their own action categories allows users to customize their experience and expand the functionality of their XR devices. Addressing this need is the primary objective of this work.", + "bbox": [ + 212, + 368, + 787, + 518 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Adapting human action models to new user categories over time faces a few challenges. Firstly, the model must be capable of learning new actions with minimal amount of training data so users can add new classes by providing just a few training examples per class. Secondly, due to the increasing use of XR devices for personal assistance, there is a need for privacy preservation in user action recognition-based pipelines [2, 14]. Hence, the adaptation of such action recognition models to new user categories must also be 'data-free', i.e., it cannot store and replay previously seen user training data in subsequent continual sessions. Considering these requirements, we leverage the recent success of 'data-free' prompt-based learning [49] and propose a new spatio-temporal prompt offset tuning approach to efficiently adapt the default model without finetuning.", + "bbox": [ + 212, + 520, + 787, + 685 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Human action recognition systems are moving to skeleton-based approaches, especially in applications that require low-shot action recognition capabilities such as medical action recognition [27, 56]. Skeletons offer a robust and compact alternative to videos in such low-shot regimes, due to their relatively low dimensionality and lesser variance under background conditions. While there have been a wide variety of efforts in skeleton-based human action recognition over the years [36, 52, 53], there have been fewer efforts on adapting such models to newer user categories. Efforts like [1, 22] attempted to continually learn new user", + "bbox": [ + 212, + 686, + 787, + 806 + ], + "page_idx": 1 + }, + { + "type": "page_number", + "text": "2", + "bbox": [ + 217, + 114, + 228, + 126 + ], + "page_idx": 1 + }, + { + "type": "header", + "text": "P. Garg et al.", + "bbox": [ + 271, + 114, + 362, + 128 + ], + "page_idx": 1 + }, + { + "type": "page_footnote", + "text": "5 We use human action as an umbrella term for both hand gesture and body activity in this work for ease of presentation.", + "bbox": [ + 217, + 815, + 787, + 839 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "categories over time in skeleton-based human action recognition, but relied on fully-supervised data for the new classes. On the other hand, few-shot learning works [27, 46, 56] adapt a pretrained skeleton-based action recognition model to new data, but without explicitly retaining past categories. In this work, we seek to learn new user categories in trained human action models with very few labeled samples for the new classes, while being data-free (not storing samples from previously trained categories). Fig 1 summarizes our overall objective. One could view our setting as privacy-aware few-shot continual learning for skeleton-based action recognition.", + "bbox": [ + 212, + 146, + 787, + 282 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "To this end, we propose a prompt offset tuning methodology that can be integrated with existing backbone architectures for skeleton-based human action recognition. Our learnable (soft) prompts are selected from a shared knowledge pool of prompts based on an input instance dependent attention mechanism. In particular, we propose prompt selection using an ordered query-key matching that enables a temporal prompt frame order selection consistent with the input instance. We show that such an approach allows us to learn new user categories without having to store data from past classes, without overwriting the pre-existing categories. To the best of our knowledge, this is the first effort on leveraging prompt tuning for skeleton-based models, as well as on spatiotemporal prompt selection and tuning.", + "bbox": [ + 212, + 282, + 787, + 449 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Our key contributions are summarized below:", + "bbox": [ + 214, + 452, + 545, + 467 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "- We formalize a novel problem setting which continually adapts human action models to new user categories over time, in a privacy aware manner.", + "bbox": [ + 212, + 473, + 782, + 500 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "- To address this problem, we propose a novel spatio-temporal Prompt OffsEt Tuning methodology (POET). In particular, it is designed to seamlessly plug-and-play with a pretrained model's input embedding, without any significant architectural changes.", + "bbox": [ + 212, + 503, + 782, + 559 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "- Our comprehensive experimental evaluation on two benchmark datasets brings out the efficacy of our proposed approach.", + "bbox": [ + 212, + 561, + 782, + 590 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "2 Related Works", + "text_level": 1, + "bbox": [ + 214, + 614, + 397, + 631 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "2.1 Prompt Tuning", + "text_level": 1, + "bbox": [ + 214, + 648, + 392, + 664 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "The idea of prompting, as it originated from Large Language Models (LLMs), is to include additional information, known as a text prompt, to condition the model's input for generating an output relevant to the prompt. Instead of applying a discrete, pre-defined 'hard' language prompt token, prompt and prefix tuning [20,23] formalized the concept of applying 'soft prompts' to the input. A set of learnable parameters are prepended (concatenated) to the input text and trained along with the classifier while keeping the backbone parameters frozen. Similar to prompt tuning of LLMs, recent works have popularized prompt tuning of ViTs [16] as an effective way of adapting large pretrained models to downstream tasks [49,57]. However, it remains unexplored and undefined (to the best of our knowledge) for non-transformer architectures such as GNNs.", + "bbox": [ + 212, + 672, + 787, + 840 + ], + "page_idx": 2 + }, + { + "type": "header", + "text": "POET: Prompt Offset Tuning for Continual Human Action Adaptation", + "bbox": [ + 254, + 114, + 730, + 128 + ], + "page_idx": 2 + }, + { + "type": "page_number", + "text": "3", + "bbox": [ + 774, + 116, + 785, + 126 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "2.2 Prompt Tuning for Continual Learning", + "text_level": 1, + "bbox": [ + 215, + 146, + 583, + 162 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Prompt tuning provides a simple and cost-effective way of learning task-specific signal condensed into 'soft prompts'. For continual learning, training a set of prompts for each sequential task provides a natural alternative to storing privacy violating exemplars and replaying them. Training task-specific prompts for each sequential task is straightforward when authors assume access to task identity at both train and inference time, like in Progressive Prompts [34]. However, if task identity is unavailable at inference, the model will not know which task's prompts or classifier to use for evaluating a test sample. In this respect, S-prompts [47] and A-la-carte prompt tuning (APT) [4] learn an independent set of prompts for each domain/task and employ a KNN-based search for domain/task identity at test time. Since these methods learn stand-alone prompts for every task, the prompt feature space is task-specific, and there is no forgetting of old knowledge when learning new tasks (by design). At the same time however, these 'no forgetting' prompts cannot share knowledge across tasks.", + "bbox": [ + 212, + 176, + 787, + 387 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "This leads to another ideology for continual prompt tuning, i.e., treat each prompt unit as being a part of a larger shared (knowledge) pool of prompts. Then the desired number of prompt units can be selected from the pool, conditioned on the input instance itself [41, 48, 49]. Given the scarcity of new data in our setting, we hypothesize that sharing of knowledge will benefit new tasks and draw inspiration from this line of works. Most recently, Adaptive Prompt Generator (APG) [42] challenges the intensive ImageNet21K pre-training assumption as it prompts a ViT pretrained only on the continual benchmark's base class data (similar to us). However, they use replay and knowledge distillation-style 'antiforgetting learning', in addition to using prompts. Even though our backbone is trained only on the base classes, we propose a simple prompt tuning-only strategy to counter forgetting. This implies that a prompt strategy is all we need to continually add new action semantics in a few-shot manner.", + "bbox": [ + 212, + 388, + 787, + 584 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "2.3 Few-Shot Class Incremental Learning", + "text_level": 1, + "bbox": [ + 215, + 609, + 571, + 626 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "FSCIL is a challenging continual learning setting where a model overfits to new classes, with the simultaneous heightened (often complete) forgetting of old knowledge as soon as the base model is fine-tuned on few-shot data [10,43]. Since the backbone feature extractor is the only source of previously seen knowledge, if it is updated, knowledge is lost forever. Typically, existing works decouple the learning of (backbone) feature representations from the classifier by learning the model only on the base data and relying on non-parametric class-mean classifiers for classification in subsequent steps [13,31,55]. This leads to a feature-classifier misalignment issue [32, 51] because new class prototypes are extracted from a backbone representation trained only on the base classes. We hypothesize that optimizing input prompt vectors along with a dynamically expanding parametric classifier on top of a frozen backbone can alleviate this misalignment issue. Our work not only provides a fresh perspective into FSCIL, but to our best knowledge is also the only work not designed for and evaluated on image benchmarks.", + "bbox": [ + 212, + 640, + 787, + 852 + ], + "page_idx": 3 + }, + { + "type": "page_number", + "text": "4", + "bbox": [ + 217, + 114, + 228, + 126 + ], + "page_idx": 3 + }, + { + "type": "header", + "text": "P. Garg et al.", + "bbox": [ + 271, + 114, + 362, + 128 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "3 Preliminaries", + "text_level": 1, + "bbox": [ + 215, + 143, + 382, + 160 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Skeleton Action Recognition Using Graph Representations. Our input $\\mathbf{X} \\in \\mathbb{R}^{T \\times J \\times 3}$ is a video sequence of $T$ frames, each frame containing $J$ joints of the human body (25 joints) or hand skeleton (22 joints) in 3D Cartesian coordinate system. Such a skeleton action sequence is naturally represented as a graph topology $G = \\{\\mathcal{V}, \\mathcal{E}\\}$ with $\\mathcal{V}$ vertices and $\\mathcal{E}$ edges. Graphs are modeled using Graph Neural Networks (GNNs) [12], which can either be sparse graph convolutional networks (GCN) or fully connected graph transformers (GT). Our main model (a GNN) is defined as $f(\\mathbf{X}) = f_c \\circ f_g \\circ f_e(\\mathbf{X})$ (as also shown in Fig. 2). Input $\\mathbf{X}$ is first passed through an input embedding layer $f_e$ to get an embedding of human joints $\\mathbf{X_e} = f_e(\\mathbf{X}), \\mathbf{X_e} \\in \\mathbb{R}^{T \\times J \\times C_e}$ , with feature dimension $C_e$ . $\\mathbf{X_e}$ is further passed to a graph feature extractor $f_g$ composed of a stack of convolutional layers (in GCNs) or attention layers (in GTs), and finally a classifier $f_c$ which predicts the action class label $\\mathbf{y}$ . In POET, we propose to attach learnable parameters $\\mathbf{P_T}$ (called prompt offsets) to the embedding $\\mathbf{X_e}$ .", + "bbox": [ + 212, + 172, + 787, + 385 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Problem Definition. Given a default (pre)trained model deployed on a user's device, we would like to extend this model to new action classes over $T$ subsequent user sessions (also called tasks) $\\{\\mathcal{US}^{(1)},\\dots,\\mathcal{US}^{(T)}\\}^6$ . In each user session $\\mathcal{US}^{(t)}$ , the model learns a dataset $\\mathcal{D}^{(t)} = (\\mathbf{X}_i^t,\\mathbf{y}_i^t)_{i = 1}^{|D^{(t)}|}$ of skeleton action sequence and label pairs provided by the user, $\\mathbf{X}_i^t\\in \\mathbb{R}^{T\\times J\\times 3}$ , $y_i^t\\in \\mathbb{R}^{y^{(t)}}$ . In each session, the user typically provides a few training instances $F$ (e.g. $F\\leq 5$ ) for each of the $N$ new classes being added, such that $|\\mathcal{D}^{(t)}| = NF$ . The base (default) model's session $\\mathcal{UB}^{(0)}$ is assumed to have a large number of default action classes $\\mathcal{V}^{(0)}$ trained on sufficient data $\\mathcal{D}^{(0)}$ , which is most often proprietary and cannot be accessed in later user sessions. In each session, the user adds new action classes such that, $\\mathcal{V}^{(t)}\\cap \\mathcal{V}^{(t')} = \\emptyset ,\\forall t\\neq t'^7$ . Due to the aforementioned privacy constraints, in any training session $\\mathcal{US}^{(t)}$ , the model has access to only $\\mathcal{D}^{(t)}$ ; after training this data is made inaccessible for use in subsequent sessions (no exemplar or prototypes stored). After training on every new session $\\mathcal{US}^{(t)}$ , the model is evaluated on the test set of all classes seen so far $\\cup_{i = 0}^{t}\\mathcal{V}^{(i)}$ . The challenge is to alleviate forgetting of old classes while not overfitting to the user-provided new class samples. One could view our setting as privacy-aware few-shot continual action recognition, a problem of practical relevance in human action recognition - which has not received adequate attention.", + "bbox": [ + 212, + 388, + 787, + 686 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "4 Methodology: Prompt Offset Tuning (POET)", + "text_level": 1, + "bbox": [ + 215, + 703, + 692, + 722 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Overview. We propose to prompt tune a base GNN model $f(\\cdot)$ by prompts $\\mathbf{P_T}$ to address our overall objective. As shown in Fig. 3, for each input instance $\\mathbf{X}$ , corresponding prompts $\\mathbf{P_T}$ are selected from a pool of prompt parameters, using an input-dependent query and key attention mechanism. The selected prompts", + "bbox": [ + 212, + 729, + 787, + 792 + ], + "page_idx": 4 + }, + { + "type": "header", + "text": "POET: Prompt Offset Tuning for Continual Human Action Adaptation", + "bbox": [ + 253, + 113, + 732, + 130 + ], + "page_idx": 4 + }, + { + "type": "page_number", + "text": "5", + "bbox": [ + 774, + 114, + 785, + 127 + ], + "page_idx": 4 + }, + { + "type": "page_footnote", + "text": "6 User sessions may be spaced at arbitrary time intervals.", + "bbox": [ + 217, + 803, + 555, + 815 + ], + "page_idx": 4 + }, + { + "type": "page_footnote", + "text": "We make this assumption considering this is a first of such efforts; allowing for overlapping action classes and users to 'update' older classes would be interesting extensions of our proposed work.", + "bbox": [ + 220, + 816, + 785, + 839 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "are added to the input feature embedding (and hence the term 'prompt offsets'), before forwarding to the feature extractor and classifier (shown in Fig. 2).", + "bbox": [ + 212, + 146, + 782, + 176 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "To this end, our method, POET uses the same number of prompts as the number of temporal frames in the input, to maintain temporal consistency between the prompt and the input. Focusing solely on prompt offsets allows us to adapt the model to subsequent user sessions without having to update the input embeddings or the feature extraction backbone. Our prompt selection mechanism is learnable and trained along with the classifier to make this method simple and efficient.", + "bbox": [ + 212, + 176, + 573, + 329 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "What are Prompt Offsets? Learnable (or soft [20]) prompts are parameter vectors in a continuous space which are optimized to adapt the pretrained frozen backbone $f_{g}$ to each continual task. We define our spatio-temporal prompt offsets $\\mathbf{P_T}$ as a set of $T$ prompts (same in number as skeletal frames in input), each prompt $P_{i}$ having length equal to the number of joints in a frame $J$ and feature dimension same as input feature embedding $\\mathbf{X_e}$ , i.e., $P_{i} \\in \\mathbb{R}^{J \\times C_{e}}$ .", + "bbox": [ + 212, + 333, + 573, + 484 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Existing prompt tuning efforts, for example in image classification, focus on concatenating learnable prompts to the input token sequence in transformer architectures [16, 20]. Even though trans", + "bbox": [ + 212, + 484, + 573, + 545 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "formers can be generalized to graphs [7, 11, 29], it is non-trivial to attach prompts to a GNN. This is because transformers can be viewed as treating sentences or images as fully connected graphs where any word (or image patch) can attend to any other word in the sentence [12]. However, our input is a spatio-temporal graph skeleton of the human joint-bone structure with its own edge connectivity. Concatenating prompts along spatial or temporal dimensions would affect the graph semantics, and also affect standard training strategies such as a forward pass or backpropagation (especially in GCNs). Hence, we attach the selected prompts $\\mathbf{P_T}$ to the corresponding input feature embedding $\\mathbf{X_e}$ via a prompt attachment operator $f_{p}(.)$ . The class logit distribution $\\mathbf{y}$ is thus obtained as:", + "bbox": [ + 212, + 545, + 787, + 696 + ], + "page_idx": 5 + }, + { + "type": "equation", + "text": "\n$$\n\\mathbf {y} = f (\\mathbf {X}, \\mathbf {P} _ {\\mathbf {T}}) = f _ {c} \\circ f _ {g} \\circ f _ {p} \\left(f _ {e} (\\mathbf {X}), \\mathbf {P} _ {\\mathbf {T}}\\right) \\tag {1}\n$$\n", + "text_format": "latex", + "bbox": [ + 354, + 702, + 785, + 720 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "In every user session $t > 0$ , the classifier output dimension expands by $N$ to accommodate the new action classes. Unlike most existing continual prompt tuning works, our feature extractor backbone $f_{g}$ is trained only on the base class data $\\mathcal{D}^{(0)}$ and is never fine-tuned on classes from new user sessions $\\mathcal{U}\\mathcal{S}^{(t)}, t > 0$ . After the base session training, parameters of $f_{g}, f_{e}$ are frozen.", + "bbox": [ + 212, + 726, + 787, + 805 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Prompt Pool Design. As stated in Sec. 2.2, to encourage knowledge sharing across user sessions, we choose to construct a single prompt pool $\\mathbf{P}$ which encodes", + "bbox": [ + 212, + 809, + 787, + 840 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/a2ff12afd4f1ec19e137cbe1431a41753ef0ef8f4958409d3b5efa49301b4fb8.jpg", + "image_caption": [ + "Fig.2: POET: Prompt-offset Tuning proposes to offset the input feature embedding $\\mathbf{X}_{\\mathrm{e}}$ of the main model by learnable prompt parameters $\\mathbf{P_T}$ for privacy-aware few-shot continual action recognition. We explain prompt selection mechanism in Fig. 3." + ], + "image_footnote": [], + "bbox": [ + 588, + 195, + 781, + 411 + ], + "page_idx": 5 + }, + { + "type": "page_number", + "text": "6", + "bbox": [ + 217, + 114, + 228, + 126 + ], + "page_idx": 5 + }, + { + "type": "header", + "text": "P. Garg et al.", + "bbox": [ + 271, + 114, + 362, + 128 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/1d64d8218068d4851448335f14f345fe791b511cc7f3c49d714e715015bbf95a.jpg", + "image_caption": [ + "Fig. 3: Selection of our prompts $\\mathbf{P_T}$ : Input-dependent query $q$ is matched with keys $K$ using sorted cosine similarity to get an ordered index sequence $(s_i)_{i=1}^T$ of the top $T$ keys. This ordered index sequence is used to select the corresponding ordered prompt sequence $\\mathbf{P_T}$ from prompt pool $\\mathbf{P}$ . We add $\\mathbf{P_T}$ to $\\mathbf{X_e}$ , thereby adding an offset to it. Our experimental evaluation confirms that such an additive spatio-temporal prompt offset can balance the plasticity to learn new classes from a few action samples, while maintaining stability on previously learned classes." + ], + "image_footnote": [], + "bbox": [ + 230, + 143, + 767, + 280 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "knowledge across the sessions:", + "bbox": [ + 215, + 359, + 434, + 375 + ], + "page_idx": 6 + }, + { + "type": "equation", + "text": "\n$$\n\\mathbf {P} = \\left\\{P _ {1},.. P _ {i},..., P _ {M} \\right\\}, \\qquad P _ {i} \\in \\mathbb {R} ^ {J \\times C _ {e}}; M = \\# \\text {p r o m p t s a t t i m e t} \\tag {2}\n$$\n", + "text_format": "latex", + "bbox": [ + 259, + 378, + 785, + 398 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "For selecting prompts from this pool (Fig. 3), we construct a bijective key-value codebook, treating prompts in the pool $\\mathbf{P}$ as values and defining learnable key vectors $\\pmb{K} = \\{\\pmb{k}_1, \\dots, \\pmb{k}_i, \\dots, \\pmb{k}_M\\}, \\pmb{k}_i \\in \\mathbb{R}^{C_e}$ . A cosine similarity matching $\\gamma(.)$ between the query $\\pmb{q}$ and keys $\\pmb{K}$ is used to find indices of the $T$ closest keys $\\mathbb{Z}$ , which in turn are used to select prompts from the pool:", + "bbox": [ + 214, + 404, + 782, + 479 + ], + "page_idx": 6 + }, + { + "type": "equation", + "text": "\n$$\n\\mathbb {Z} = \\underset {T} {\\operatorname {a r g m a x}} \\gamma \\left(f _ {q} (\\mathbf {X}), \\mathbf {K}\\right) \\tag {3}\n$$\n", + "text_format": "latex", + "bbox": [ + 408, + 484, + 785, + 510 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "This quantization process is enabled by a query function $f_{q}(.)$ , which is a pretrained encoder that maps an input instance $\\mathbf{X}$ to a query $\\pmb{q}$ as:", + "bbox": [ + 215, + 517, + 782, + 547 + ], + "page_idx": 6 + }, + { + "type": "equation", + "text": "\n$$\n\\boldsymbol {q} = f _ {q} (\\mathbf {X}) = f _ {Q A} \\circ f _ {g} ^ {\\prime} \\circ f _ {e} ^ {\\prime} (\\mathbf {X}), \\quad f _ {q}: \\mathbb {R} ^ {T \\times J \\times 3} \\longrightarrow \\mathbb {R} ^ {C _ {e}} \\tag {4}\n$$\n", + "text_format": "latex", + "bbox": [ + 297, + 551, + 785, + 571 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "where the query adaptor $f_{QA}$ is a fully connected layer mapping the $f_{g}^{\\prime}$ output dimension to the desired prompt embedding dimension $C_e$ .", + "bbox": [ + 214, + 578, + 782, + 608 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Coupled Optimization in User Sessions $t > 0$ . Typically, the argmax operator in Eq. 3 decouples the optimization of keys from the prompt pool and main model as it prevents backpropagation of gradients to the keys (also seen in earlier works such as [48, 49]). However, this approach does not work for our setting, even more so since we assume no large-scale pre-training of our base model. Due to the lack of off-the-shelf availability of large-scale models for skeletal action data, our query function $f_{q}$ is pretrained only on base class data $\\mathcal{D}^{(0)}$ . Hence, it becomes important that $f_{q}$ is updated as the model learns new classes. As shown in red boxes in Figs. 2 and 3, we propose to couple this optimization process such that the overall cross-entropy loss for new tasks updates: (i) the classifier $f_{c}$ , (ii) selected prompts in $\\mathbf{P}$ , (iii) selected keys in $K$ , as well as (iv) query adaptor $f_{QA}$ . We achieve this by approximating the gradient for $K$ and $f_{QA}$ by the straight-through estimator reparameterization trick as in [3, 44]. We freeze the query feature extractor layers $f_{g}^{\\prime}, f_{e}^{\\prime}$ in $t > 0$ to prevent catastrophic forgetting of base knowledge in $f_{q}$ . Our cross-entropy loss is hence given by:", + "bbox": [ + 214, + 613, + 787, + 843 + ], + "page_idx": 6 + }, + { + "type": "header", + "text": "POET: Prompt Offset Tuning for Continual Human Action Adaptation", + "bbox": [ + 254, + 114, + 732, + 128 + ], + "page_idx": 6 + }, + { + "type": "page_number", + "text": "7", + "bbox": [ + 774, + 116, + 785, + 126 + ], + "page_idx": 6 + }, + { + "type": "equation", + "text": "\n$$\n\\min _ {\\theta_ {f _ {Q A}}, \\theta_ {K}, \\theta_ {\\mathbf {P}}, \\theta_ {f _ {c}}} \\mathcal {L} (f (\\mathbf {X}, \\mathbf {P} _ {\\mathbf {T}}), \\mathbf {y}) \\tag {5}\n$$\n", + "text_format": "latex", + "bbox": [ + 395, + 146, + 784, + 171 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "To move queries closer to their aligned $T$ keys during training, we use a vector quantization clustering loss inspired from VQ-VAE [44] as:", + "bbox": [ + 212, + 176, + 782, + 207 + ], + "page_idx": 7 + }, + { + "type": "equation", + "text": "\n$$\n\\max _ {\\theta_ {f _ {Q A}}, \\theta_ {K}} \\lambda \\sum_ {i \\in \\mathbb {Z}} \\gamma \\left(f _ {q} (\\mathbf {X}), K _ {i}\\right) \\tag {6}\n$$\n", + "text_format": "latex", + "bbox": [ + 403, + 210, + 785, + 242 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "where $\\lambda$ is the clustering loss coefficient. Our end-to-end optimization thus establishes a prompt optimization framework which is amenable to prompt tuning when extensive pre-training is not possible. This sets the foundation for our spatio-temporal prompt selection module, described next.", + "bbox": [ + 212, + 247, + 782, + 306 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Spatio-Temporal Prompt Selection. In order to ensure that our learned prompts respect temporal information in the input video sequence, we choose the number of selected prompts to be equal to the number of frames in the input", + "bbox": [ + 212, + 311, + 782, + 358 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "video $T$ . After coupling the prompt pool and keys, we observed in our initial experiments with pool size $M > T$ that the same set of prompts get selected across training iterations and user sessions (Fig. 4A). More concretely, as the vector quantization loss (Eqn 6) brings the query close to the selected keys, the same set of active prompts get selected and optimized in each iteration, not using other prompts at all. This is similar to the well-known issue of 'codebook col", + "bbox": [ + 212, + 359, + 486, + 551 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/655464329a10d8067e2e10803916ebc135458dab373e9d280dcd48d26a17afed.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 500, + 359, + 785, + 429 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/1183da4b6d6d075e6477a447af1a7ac4f6075e36005e75afa5cc5623a79d6a0d.jpg", + "image_caption": [ + "Fig.4: $M > T$ Case: Prompt Pool Collapse. (Top) Certain prompt indices remain unused across user sessions. (Bottom) Our POET pool expansion strategy alleviates pool collapse." + ], + "image_footnote": [], + "bbox": [ + 500, + 435, + 784, + 503 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "lapse' in VQ-VAE [9, 50, 54]. Based on this observation, we design two prompt pool update mechanisms in user sessions $t > 0$ as below:", + "bbox": [ + 212, + 554, + 782, + 583 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "1. Case 1, $M = T \\forall t$ : No pool expansion, Algorithm 1. All prompts are selected in all tasks. But the order of their selection $(s_i)_{i=1}^T$ varies with each input instance as we replace Eq. 3 by sorting the cosine similarity before selecting the top $T$ indices as follows:", + "bbox": [ + 214, + 585, + 784, + 646 + ], + "page_idx": 7 + }, + { + "type": "equation", + "text": "\n$$\n\\mathbb {Z} = \\underset {\\left(s _ {i}\\right) _ {i = 1} ^ {T}} {\\operatorname {a r g s o r t}} \\gamma \\left(f _ {q} (\\mathbf {X}), \\mathbf {K}\\right) \\tag {7}\n$$\n", + "text_format": "latex", + "bbox": [ + 419, + 648, + 784, + 676 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "In Fig. 5, we visualize the positions occupied by indices in this (sorted) ordered key index sequence $(s_i)_{i=1}^T$ . Entropy increase across tasks $t = 1$ to $t = 4$ (bottom row of figure) shows that our selection mechanism learns to select a unique temporal code for all inputs.", + "bbox": [ + 233, + 680, + 785, + 739 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "2. Case 2, $M = T + (R * t), t > 0$ . Expand pool with $R$ prompts. We also propose an order-aware prompt pool expansion strategy (Appendix B) that selects prompts from an expanded pool in a temporally coherent manner, for $t > 0$ . This alleviates prompt pool collapse as shown in Fig. 4B.", + "bbox": [ + 214, + 739, + 782, + 801 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Prompt Offset Attachment. Since concatenation is not meaningful for graph data, we use addition as our choice for the prompt attachment operator as:", + "bbox": [ + 212, + 809, + 782, + 839 + ], + "page_idx": 7 + }, + { + "type": "page_number", + "text": "8", + "bbox": [ + 217, + 114, + 228, + 126 + ], + "page_idx": 7 + }, + { + "type": "header", + "text": "P. Garg et al.", + "bbox": [ + 271, + 114, + 362, + 128 + ], + "page_idx": 7 + }, + { + "type": "equation", + "text": "\n$$\nf _ {p} \\left(\\mathbf {X} _ {\\mathbf {e}}, \\mathbf {P} _ {\\mathbf {T}}\\right) = \\mathbf {X} _ {\\mathbf {e}} + \\mathbf {P} _ {\\mathbf {T}} \\tag {8}\n$$\n", + "text_format": "latex", + "bbox": [ + 413, + 145, + 785, + 162 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "Hence, we call our approach as prompt offset tuning. We also study this empirically through experiments that support this choice in Sec. 6.", + "bbox": [ + 212, + 167, + 782, + 198 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "Interpreting Prompt Offset Tuning of GNNs. Our additive prompt offsets are open to interpretation, as shown in Fig. 5. (i) Adding our selected prompts $\\mathbf{P_T}$ to input feature embedding $\\mathbf{X_e}$ acts like an input-dependent transformation", + "bbox": [ + 212, + 200, + 784, + 247 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "for spatio-temporal joints.", + "bbox": [ + 212, + 247, + 400, + 262 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "(ii) As our prompts have same size as $\\mathbf{X}_{\\mathbf{e}}$ it can also be thought of as a learned prompt encoding, bearing similarity with learnable position encoding works [12, 24, 26]. Our purpose is different however as prompt offsets seek to dynamically condition the input for adapting the backbone continually,", + "bbox": [ + 212, + 263, + 388, + 474 + ], + "page_idx": 8 + }, + { + "type": "image", + "img_path": "images/8b6936c5caba33ff9e421a5b926c64d9a3beb54d70507904b14af0dd6f349945.jpg", + "image_caption": [ + "Fig. 5: Here we visualize the order $(s_i)_{i=1}^T$ in which the $M = 64$ prompts in the pool are selected at train time, across 4 user sessions $\\mathcal{U}\\mathcal{S}^{(t)}$ . X-axis: prompt index, Y-axis: index position in selected sequence. Top: The no sorting case uses the default sequence (hence diagonal matrices), giving equal importance to all prompts. Bottom (Our Method): Even though the same 64 prompts are selected and updated, the ordering is temporally unique and consistent with input." + ], + "image_footnote": [], + "bbox": [ + 388, + 250, + 787, + 383 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "instead of learning positions. (iii) POET also bears similarity with auto-decoders like DeepSDF [30] which learn latent codes for each style or shape and use relevant codes along with a frozen decoder at inference. (iv) Prompt tuning can also be thought of as a parameter isolation technique for continual learning [28,34,35,38]. POET's ordered prompt selection as seen in Fig. 5 learns to isolate the relevant sequence of prompts for each input action sequence.", + "bbox": [ + 212, + 474, + 787, + 566 + ], + "page_idx": 8 + }, + { + "type": "code", + "sub_type": "algorithm", + "code_caption": [ + "Algorithm 1 POET at Train Time, $t > 0$ (Case 1 $M = T$ , No pool expansion)" + ], + "code_body": "Input: Query function $f_{q}$ keys $\\pmb {K} = \\{\\pmb {k}_j\\}_{j = 1}^T$ prompt pool $\\mathbf{P} = \\{\\pmb {P}_j\\}_{j = 1}^T$ main model $f_{e},f_{g},f_{c}$ Initialize: $\\mathbf{P},\\mathbf{K}$ from $t - 1$ ; Expand $f_{c}$ by $N$ new classes. Initialize $f_{c}$ as: (i) copy $f_{c}^{old}$ weights, (ii) $f_{c}^{new}\\gets Mean(f_{c}^{old})$ Freeze: query layers $f_{g}^{\\prime},f_{e}^{\\prime}$ ; main model layers $f_{e},f_{g}$ for epochs and batch $(\\mathbf{X}_i^t,\\mathbf{y}_i^t)_{i = 1}^{NK}$ do 1. Get query feature $\\pmb{q}$ (Eq.4); Compute $\\gamma (\\cdot)$ b/w query $\\pmb{q}$ and keys $\\pmb{K}$ 2.Sort $\\gamma (\\cdot)$ ; Get ordered key index sequence $(s_i)_i = 1$ Eq.7 3. Read pool memory $\\mathbf{P}$ in order $(s_i)^T_{i = 1}\\rightarrow$ Get prompt offsets $\\mathbf{P_T}$ 4.Get $\\mathbf{X_e};$ Add $\\mathbf{P_T}$ to it (Eq.8); get prediction y from prompted input (Eq. 1) 5.Use cross entropy loss (Equation 5) to update $f_{QA},K,\\mathbf{P},f_{c}$ 6.Use clustering loss (Equation 6) to update $f_{QA}$ and $\\pmb{K}$ end // See $t = 0$ training protocol in Algorithm 2 in Appendix", + "bbox": [ + 215, + 593, + 787, + 732 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "5 Experiments and Results", + "text_level": 1, + "bbox": [ + 215, + 751, + 496, + 768 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "Datasets. We evaluated our method on well-known action recognition datasets: (i) activity recognition on the NTU RGB+D dataset [39]; and (ii) hand gesture", + "bbox": [ + 212, + 772, + 787, + 805 + ], + "page_idx": 8 + }, + { + "type": "header", + "text": "POET: Prompt Offset Tuning for Continual Human Action Adaptation", + "bbox": [ + 253, + 113, + 732, + 130 + ], + "page_idx": 8 + }, + { + "type": "page_number", + "text": "9", + "bbox": [ + 774, + 114, + 785, + 126 + ], + "page_idx": 8 + }, + { + "type": "page_footnote", + "text": "8 The datasets used in this work were accessed and processed at and by CMU. They were not accessed, processed, stored, or maintained at Meta.", + "bbox": [ + 217, + 810, + 785, + 839 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "recognition on the SHREC-2017 dataset [40]. As we introduce a new problem setting in human action recognition, we contribute two new benchmarks to the community for this setting, on the NTU RGB+D and SHREC-2017 datasets.", + "bbox": [ + 212, + 146, + 784, + 191 + ], + "page_idx": 9 + }, + { + "type": "text", + "text": "For the NTU RGB+D dataset, we divide the 60 daily action categories into 40 base classes, learning the remaining 20 classes in subsequent user sessions. In few-shot learning parlance, our protocol is 4-task 5-way 5-shot, i.e. 5 novel classes using 5 user training instances in 4 user sessions. Each input 3D skeleton sequence has 64 temporal frames, each consisting of 25 body keypoints, such that $x \\in \\mathcal{R}^{64 \\times 25 \\times 3}$ . We use the spatio-temporal GCN, CTR-GCN [8], as the architecture for NTU RGB+D, where we choose the joint input modality for better interpretability of prompt tuning.", + "bbox": [ + 212, + 195, + 787, + 316 + ], + "page_idx": 9 + }, + { + "type": "text", + "text": "For SHREC-2017, we divide the 14 fine-grained hand gesture classes into 8 base classes and 6 classes learned in subsequent user sessions. This is done in a 3-task 2-way 5-shot protocol, i.e. 2 novel classes using 5 user training instances in 3 user sessions. For each input instance of SHREC-2017, we use 8 temporal frames each having 22 hand keypoints, such that input $x \\in \\mathcal{R}^{8 \\times 22 \\times 3}$ . We use a fully-connected graph transformer backbone, DG-STA [7] for SHREC-2017. We select DG-STA due to easily reproducible code and to validate if our method POET works equally well across graph convolutional networks and graph transformers.", + "bbox": [ + 212, + 319, + 787, + 441 + ], + "page_idx": 9 + }, + { + "type": "text", + "text": "Evaluation Metrics. Following earlier work in similar settings [31], we report: (i) Average accuracy 'Avg' of all classes seen so far, and (ii) Harmonic Mean $A_{HM}$ between 'accuracy only on Old classes' and 'accuracy only on New classes' after learning each new user session. Note that the average accuracy tends to be biased towards the base session $\\mathcal{T}^{(0)}$ performance due to more number of base classes. A higher $A_{HM}$ implies better stability-plasticity trade-off between new task performance and old tasks' retention. Unlike many earlier CIL efforts, we report accuracy for both Old and New classes in each user session for transparency.", + "bbox": [ + 214, + 443, + 787, + 566 + ], + "page_idx": 9 + }, + { + "type": "text", + "text": "Implementation Details. We observe that a key source of forgetting in our setting is from the classifier as the logits tend to become heavily biased towards the few-shot samples of new classes. We use a cosine classifier for activity recognition experiments on CTR-GCN. For gesture recognition on the lightweight DG-STA, we use a standard fully-connected layer as classifier, but freeze old class parameters in the classifier by zeroing their gradients. We attach prompts after the 1st layer of DG-STA and 1st CTR-GC block of CTR-GCN. For both datasets, we have equal or higher learning rates in user sessions when compared to the base model's training in order to accommodate new knowledge in the model (for better plasticity). For exact implementation details (including learning rates, epochs, hyperparameter analysis, and backward forgetting metric), see Appendix A. In earlier efforts that more generally tune prompts for class-incremental learning [41, 45, 47-49], it is common to rely on an ImageNet21K pretrained ViT [37] or CLIP [33] as the backbone. However, such backbones do not exist for skeleton-based human action recognition. Our base feature extractor is hence trained on the base session dataset itself without any pretraining, making this one of the first efforts of prompt tuning without extensive pretraining (scale of data 3-5 times lower order of magnitude).", + "bbox": [ + 214, + 568, + 787, + 840 + ], + "page_idx": 9 + }, + { + "type": "page_number", + "text": "10", + "bbox": [ + 217, + 114, + 235, + 126 + ], + "page_idx": 9 + }, + { + "type": "header", + "text": "P. Garg et al.", + "bbox": [ + 271, + 114, + 362, + 128 + ], + "page_idx": 9 + }, + { + "type": "text", + "text": "Results. Since there are no existing baselines for our proposed setting in skeletal action recognition, we compare our method by adapting continual learning (CL) baselines to skeletal data in Sec 5.1, Tables 1, 2. We first compare POET with prompt tuning based class-incremental learning (CIL) approaches originally designed for images (L2P [49], CODA-P [41], APT [4]) and find that it has very low performance on new classes as they do not update their query function. We find any fine-tuning or knowledge distillation based approaches (LWF [25], EWC [17], LUCIR [15]) lead to rapid forgetting of base knowledge as the model overfits to user's few-shots. We also compare with multiple variants of Feature Extraction (FE) to check if prompts truly have merit (POET=FE+Prompts) and provide upper bound baselines. In Sec 6, we first show the importance of prompts in POET by removing the prompts. We discuss the value of our coupled optimization, query function update and ordered key index selection in our prompt selection ablation Tab 3. We also study the impact of proposed additive prompt tuning as compared to other possible prompt attachments $f_{p}$ in Tab 4.", + "bbox": [ + 212, + 146, + 787, + 375 + ], + "page_idx": 10 + }, + { + "type": "text", + "text": "5.1 Comparison with State-of-the-Art", + "text_level": 1, + "bbox": [ + 214, + 393, + 542, + 407 + ], + "page_idx": 10 + }, + { + "type": "text", + "text": "POET sets the SOTA on existing prompt tuning works (Tab 1,2). We adapt three standard CIL works that prompt tune ViTs for images - L2P [49], CODA-P [41] and APT [4] to our setting. L2P and CODA-P share prompt pool across tasks (similar to us), whereas APT learns task-specific prompts. L2P decouples the optimization of keys from the prompt pool and concatenates the selected prompts. Since concatenation is not defined for our GNN backbone, we adapt these SOTA to our setting by concatenating along the temporal dimension $(\\mathbf{L2P}^*,\\mathbf{CODA - P}^*)$ . CODA-P [41] couples keys with the prompt pool by using a cosine similarity weighing over all prompts in the pool, forming a 'soft prompt selection', different from our 'ordered hard prompt selection'. In APT, we train prompt-classifier pairs for each continual task separately ( $\\hat{\\mathbf{\\Gamma}}$ denotes task-specific), and use task identity at test time. See details in Appendix A. These methods by design rely on extensively pretrained (ImageNet21k) query functions which does not require updates; and require full supervision on new classes, perhaps explaining their poor 'New' accuracy in our few-shot setting.", + "bbox": [ + 212, + 417, + 787, + 643 + ], + "page_idx": 10 + }, + { + "type": "text", + "text": "Standard Continual learning Baselines. We compared with two well established knowledge-distillation approaches, learning without forgetting (LWF) and LUCIR. Both of them perform poorly on both old and new classes. EWC [17] learns better on new but does not retain old knowledge. We conclude that any CL method that fine-tunes the backbone feature representation in subsequent sessions $t > 0$ will not be able to retain base/old class knowledge (a finding consistent with existing FSCIL literature for images [10,43]). We also adapt and compare with one of the latest FSCIL baselines ALICE [31], originally developed for image classification benchmarks on our gesture recognition benchmark in Table 2. Note the high retention of base task performance (due to non-parametric classifier on top of frozen base model). However, it suffers from poor plasticity and adaptation to new classes. This is the issue of feature-classifier misalignment that we hoped to alleviate through prompt tuning.", + "bbox": [ + 212, + 643, + 787, + 840 + ], + "page_idx": 10 + }, + { + "type": "header", + "text": "POET: Prompt Offset Tuning for Continual Human Action Adaptation", + "bbox": [ + 254, + 113, + 730, + 128 + ], + "page_idx": 10 + }, + { + "type": "page_number", + "text": "11", + "bbox": [ + 767, + 114, + 782, + 126 + ], + "page_idx": 10 + }, + { + "type": "table", + "img_path": "images/5ecc0bbf3092fcabfdfcad1ed96c79e801d80f9b5530c10715bfd2d644eb700b.jpg", + "table_caption": [ + "Table 1: Activity Recognition Results $(\\% ,\\uparrow)$ , Comparison with SOTA: NTU RGB+D [39] dataset on CTR-GCN [8] backbone. After training on each incremental task, we report Average of all classes seen so far ('Avg'). We also report (i) $A_{HM}$ , (ii) old classes accuracy ('Old'), (iii) new classes accuracy ('New') in the last session. We report Mean and STD across 10 sets of 5-shots. POET achieves the best stability-plasticity trade-off across all baselines indicated by the $A_{HM} = 56.3\\%$ . POET also has the highest Avg across all user sessions outside of upper bound baselines in orange." + ], + "table_footnote": [], + "table_body": "
Method\\( \\mathbf{U}\\mathbf{B}^{(0)} \\)\\( \\mathbf{U}\\mathbf{S}^{(1)} \\)\\( \\mathbf{U}\\mathbf{S}^{(2)} \\)\\( \\mathbf{U}\\mathbf{S}^{(3)} \\)\\( \\mathbf{U}\\mathbf{S}^{(4)} \\)
Base (↑)Avg (↑)Avg (↑)Avg (↑)Old (↑)New (↑)Avg (↑)\\( A_{HM} \\) (↑)
Upper Bounds
Joint (Oracle)88.479.071.066.863.5
Joint POET (Oracle)67.2
FE, Task-Specific^88.470.1 ± 2.652.5 ± 5.844.8 ± 5.070.3 ± 2.146.7 ± 2.0NANA
FE+Replay88.482.4 ± 1.178.2 ± 1.274.5 ± 1.273.1 ± 1.043.3 ± 3.370.6 ± 1.254.3 ± 2.6
Continual Linear Probing
FE88.472.0 ± 1.160.4 ± 2.447.7 ± 2.140.0 ± 1.651.0 ± 2.340.9 ± 1.444.8 ± 1.1
FE, Frozen88.476.1 ± 1.052.4 ± 4.138.3 ± 2.728.4 ± 1.622.4 ± 4.527.9 ± 1.424.8 ± 3.0
FE+Replay†88.472.0 ± 1.559.5 ± 4.058.7 ± 2.856.7 ± 2.534.7 ± 5.654.9 ± 2.742.8 ± 4.4
FT88.46.2 ± 1.44.3 ± 1.52.8 ± 1.00.2 ± 0.536.0 ± 10.13.2 ± 0.80.3 ± 1.0
Standard Continual Learning
LWF [25]88.46.2 ± 1.52.8 ± 0.73.7 ± 1.30.0 ± 0.038.9 ± 8.83.2 ± 0.70.0 ± 0.0
EWC [17]88.46.6 ± 1.54.1 ± 1.43.1 ± 0.90.0 ± 0.042.1 ± 9.53.5 ± 0.80.0 ± 0.0
Experience Replay88.435.1 ± 8.350.6 ± 5.060.6 ± 5.454.6 ± 6.543.7 ± 14.653.7 ± 7.147.8 ± 11.2
Experience \\( \\text{Replay}^{\\dagger} \\)88.46.2 ± 1.59.0 ± 2.611.2 ± 3.010.9 ± 2.634.6 ± 7.912.9 ± 3.016.3 ± 3.5
LUCIR [15]87.94.3 ± 2.14.1 ± 1.32.7 ± 0.80.2 ± 0.426.0 ± 9.22.3 ± 0.90.4 ± 0.8
Continual Prompt Tuning
CODA-P [41]*87.476.1 ± 1.066.7 ± 1.358.6 ± 2.756.5 ± 2.90.5 ± 0.451.8 ± 2.71.1 ± 0.7
L2P [49]*88.678.9 ± 0.171.0 ± 1.064.2 ± 0.162.0 ± 0.70.0 ± 0.056.8 ± 0.60.0 ± 0.0
APT [4]^86.627.3 ± 1.630.8 ± 3.437.6 ± 2.3NA33.4 ± 2.0NANA
POET (Ours)87.982.3 ± 0.676.8 ± 0.968.4 ± 0.757.2 ± 1.055.8 ± 5.957.1 ± 1.156.3 ± 3.2
", + "bbox": [ + 220, + 212, + 772, + 465 + ], + "page_idx": 11 + }, + { + "type": "text", + "text": "Fine-tuning (FE) and Feature Extraction (FE) Baselines. We implement standard continual learning baselines to understand stability-plasticity trade-offs in our new benchmarks. In all these baselines, we expand the classifier output dimension by $N$ new classes. In 'FT (Fine-Tuning)', we tune all model parameters on cross entropy loss of new task. FSCIL is challenging for this modality as old task performance sharply reduces to zero starting from $\\mathcal{U}\\mathcal{S}^{(1)}$ as model overfits to user's few-shots. 'FE (Feature Extraction)' differs from FT as we freeze the feature extractor to preserve base knowledge. This serves as a competitive baseline in our findings. In 'FE, frozen', we zero out the gradients of previous class weights in classifier $f_{c}$ to prevent forgetting from the classifier. 'FE' and 'FE, Frozen' exhibit different New-Old trade-offs in Tables 1, 2 because the scale of pretraining is different (gesture more lightweight than activity).", + "bbox": [ + 212, + 487, + 787, + 669 + ], + "page_idx": 11 + }, + { + "type": "text", + "text": "Upper-bound baselines, top section Tables 1, 2. In 'Joint (oracle)' experiment, we train on all task data at the same time in a multi-task (non-sequential) manner. Training POET in a multi-task manner ('Joint POET') outperforms 'Joint Oracle' demonstrating the strength of our approach. In addition to these generalist upper bounds, we point out that 'FE, Task-specific' is a competitive specialist upper bound. In this, we perform feature extraction from base model to each task individually, storing separate task-specific models $(\\mathcal{U}\\mathcal{S}^{(0)}\\to \\mathcal{U}\\mathcal{S}^{(i)},i > 0)$ . POET outperforms 'New' accuracy compared with this baseline, achieving a forward transfer on each $t > 0$ . This indicates that prompt", + "bbox": [ + 212, + 671, + 787, + 808 + ], + "page_idx": 11 + }, + { + "type": "page_number", + "text": "12", + "bbox": [ + 217, + 114, + 235, + 126 + ], + "page_idx": 11 + }, + { + "type": "header", + "text": "P. Garg et al.", + "bbox": [ + 271, + 114, + 362, + 128 + ], + "page_idx": 11 + }, + { + "type": "page_footnote", + "text": "9 'FE' is the same as 'w/o prompts' in Table 3. We highlight key baselines in gray color.", + "bbox": [ + 217, + 824, + 732, + 839 + ], + "page_idx": 11 + }, + { + "type": "table", + "img_path": "images/dd25c58ccaa6e1c39338adc55e993dd398d2ef5700630ebe08616e6dd57a4bc1.jpg", + "table_caption": [ + "Table 2: Gesture Recognition Results (\\%, ↑), Comparison with SOTA: SHREC 2017 [40] dataset on DG-STA [7] graph transformer backbone. Reporting mean and standard deviation across 5 runs. POET achieves best $A_{HM} = 56.2\\%$ ." + ], + "table_footnote": [], + "table_body": "
Method\\( \\mathbf{\\mathcal{U}\\mathbf{B}^{(0)}} \\)\\( \\mathbf{\\mathcal{U}\\mathbf{S}^{(1)}} \\)\\( \\mathbf{\\mathcal{U}\\mathbf{S}^{(2)}} \\)\\( \\mathbf{\\mathcal{U}\\mathbf{S}^{(3)}} \\)
Base (↑)Avg (↑)Avg (↑)Old (↑)New (↑)Avg (↑)\\( A_{HM} \\) (↑)
Joint (Oracle)88.879.4 ± 0.777.3 ± 2.170.9 ± 1.262.4 ± 0.4
FT88.820.3 ± 0.812.4 ± 2.10.0 ± 0.085.8 ± 9.413.4 ± 1.50.0 ± 0.0
FE88.862.7 ± 2.441.9 ± 6.917.5 ± 5.177.3 ± 8.826.8 ± 3.428.5 ± 6.4
FE, Frozen88.871.3 ± 1.961.4 ± 2.744.7 ± 3.254.5 ± 6.746.2 ± 2.749.1 ± 4.3
LWF [25]88.820.2 ± 1.412.5 ± 1.00.0 ± 0.088.4 ± 13.713.8 ± 2.10.0 ± 0.0
L2P [49]**88.820.3 ± 5.910.5 ± 4.88.2 ± 4.06.9 ± 8.57.9 ± 3.97.5 ± 5.5
CODA-P [41]**87.715.6 ± 4.511.6 ± 1.97.9 ± 1.814.1 ± 21.48.8 ± 2.410.1 ± 3.2
ALICE [31]92.172.4 ± 5.763.3 ± 7.662.5 ± 6.811.9 ± 9.954.6 ± 6.920.0 ± 8.1
POET (Ours)91.973.2 ± 3.761.9 ± 1.845.9 ± 2.672.4 ± 7.150.0 ± 1.656.2 ± 1.6
", + "bbox": [ + 254, + 183, + 743, + 318 + ], + "page_idx": 12 + }, + { + "type": "text", + "text": "tuning benefits New performance due to the pre-existing knowledge in the shared knowledge pool. Avg in sessions $0 < t < 4$ indicates New for task-specific $^{\\wedge}$ .", + "bbox": [ + 212, + 335, + 782, + 364 + ], + "page_idx": 12 + }, + { + "type": "text", + "text": "Experience Replay Baselines, Tab 1. Even though our privacy-aware setting prohibits previous data replay, we compare with 'Experience Replay' (store and replay 5-samples of base and incremental sessions) and 'Experience Replay†' (replay only previous incremental sessions) for completeness.", + "bbox": [ + 212, + 366, + 784, + 425 + ], + "page_idx": 12 + }, + { + "type": "text", + "text": "'FE+Replay' serves as the best upper bound (even better than Experience Replay as we are freezing backbone in addition to replay). It is noteworthy that POET (which is FE+prompts) learns an implicit 'data-free' form of prompt pool memory, and yet has a better $A_{HM}$ trade-off as compared to explicitly stored and replayed samples from previous classes in FE+replay.", + "bbox": [ + 212, + 426, + 784, + 501 + ], + "page_idx": 12 + }, + { + "type": "text", + "text": "6 Ablation Studies and Analysis", + "text_level": 1, + "bbox": [ + 215, + 510, + 545, + 527 + ], + "page_idx": 12 + }, + { + "type": "text", + "text": "Importance of prompts in POET. First, we consider the contribution of prompt offsets in POET. Since we only attach prompts to address continual learning in POET, removing prompts gives the Feature Extraction (FE) baseline ('w/o prompts', Table 3) where the backbone is frozen after base training and only the classifier is expanded and updated on classification loss of new classes. POET improves both, 'Old' ( $\\uparrow 20.1\\%$ ) and 'New' ( $\\uparrow 10.6\\%$ ) marked in blue.", + "bbox": [ + 212, + 535, + 787, + 625 + ], + "page_idx": 12 + }, + { + "type": "text", + "text": "Prompt Selection Mechanism. In Table 3, we investigate our prompt selection mechanism and optimization choices. The 'w/o coupled optim.' experiment is a direct comparison of our additive prompt attachment with the de-coupled optimization in L2P [49]. Updating key parameters but keeping only query adaptor", + "bbox": [ + 212, + 626, + 460, + 775 + ], + "page_idx": 12 + }, + { + "type": "text", + "text": "Table 3: Prompt Selection Mechanism Analysis on NTU RGB+D dataset $(\\% ,\\uparrow)$ .. $\\mathrm{w / o}$ denotes removing that component from POET, numbers in brackets are wrt POET $(M = T)$ experiment. 'Avg' accuracy is biased towards Old' classes accuracy, $A_{HM}$ is good indicator of trade-off between New' and Old'.", + "bbox": [ + 467, + 626, + 787, + 686 + ], + "page_idx": 12 + }, + { + "type": "table", + "img_path": "images/0ab5cb54fa4aa5aebe7c535490da1a25f9563d48a4a29b259b09df1711d07886.jpg", + "table_caption": [], + "table_footnote": [], + "table_body": "
NTU RGB+D\\( \\left| \\mathcal{U}\\mathcal{B}^{(0)}\\right| \\)\\( \\left| \\mathcal{U}\\mathcal{S}^{(1)}\\right| \\)\\( \\left| \\mathcal{U}\\mathcal{S}^{(2)}\\right| \\)\\( \\left| \\mathcal{U}\\mathcal{S}^{(3)}\\right| \\)\\( \\left| \\mathcal{U}\\mathcal{S}^{(4)}\\right| \\)
MethodBaseAvgAvgAvgOldNewAvg\\( A_{HM} \\)
w/o prompts88.474.566.349.539.2 (-20.1)46.8 (-10.6)39.942.7
w/o coupled optim.88.082.875.365.856.5 (-2.8)51.3 (-6.1)56.153.8
w/o clustering loss85.581.674.364.562.0 (+2.7)18.2 (-39.2)57.028.1
w/o QA update87.982.877.469.159.4 (+0.1)52.8 (-4.6)58.755.9
w/o sorting88.282.275.268.859.9 (+0.6)46.6 (-10.8)58.852.4
POET (M>T)87.982.777.268.860.3 (+1.0)54.4 (-3.0)59.857.2
POET (M=T)87.982.876.868.659.357.459.258.3
", + "bbox": [ + 468, + 688, + 792, + 773 + ], + "page_idx": 12 + }, + { + "type": "text", + "text": "$QA$ frozen after $\\mathcal{U}\\mathcal{B}^{(0)}$ training ('w/o QA update') reduces 'New' only performance of $\\mathcal{US}^{(4)}$ by $4.6\\%$ as the query function stays fixed at base session learning and is not discriminative towards new classes. 'W/o clustering loss' from Eq. 6, performance drops starting from $\\mathcal{UB}^{(0)}$ itself. The only difference", + "bbox": [ + 212, + 776, + 787, + 840 + ], + "page_idx": 12 + }, + { + "type": "header", + "text": "POET: Prompt Offset Tuning for Continual Human Action Adaptation", + "bbox": [ + 254, + 114, + 730, + 128 + ], + "page_idx": 12 + }, + { + "type": "page_number", + "text": "13", + "bbox": [ + 767, + 114, + 785, + 126 + ], + "page_idx": 12 + }, + { + "type": "text", + "text": "between the experiment 'w/o sorting' and 'POET (M=T)' is that we do not sort the cosine similarity before selecting top $T$ indices (same as Fig 5). The $10.8\\% \\uparrow$ in 'New' performance validates that our prompt selection mechanism is learning to choose a distinct temporal ordering for prompt tuning of new input samples. With pool expansion ('POET, $M > T$ ), we get more flexibility in the stability-plasticity trade-offs depending on how many new prompts we attach. For $R = 6$ , 'Old' is improved. In Table 3, we keep POET's additive prompt attachment and only vary prompt selection.", + "bbox": [ + 212, + 146, + 787, + 267 + ], + "page_idx": 13 + }, + { + "type": "text", + "text": "Prompt Attachment Mechanism. In Table 4, we keep our end-to-end optimization and ordered prompt selection as a constant and ablate prompt shape and attachment operator $f_{p}(.)$ . Drawing a parallel with transformers which concatenate prompts along the token dimension, we conduct experiments concatenating prompts along the (i) temporal dimension of the skeleton input feature embedding $\\mathbf{X}_{\\mathrm{e}}$ ('CONCAT temporal') and (ii) feature dimension $C_e$ ('CONCAT feature'). We find that addition works better than concatenation and", + "bbox": [ + 212, + 268, + 787, + 375 + ], + "page_idx": 13 + }, + { + "type": "text", + "text": "cross attention. We also verify our hypothesis that selecting the same number of prompts as the input temporal dimension ( $T = 64$ for NTU RGB+D and $T = 8$ for SHREC-2017) yields better results as compared to adding the same prompt frame to each input embedding frame ('Addition $T' = 1'$ ).", + "bbox": [ + 212, + 375, + 488, + 496 + ], + "page_idx": 13 + }, + { + "type": "table", + "img_path": "images/309e07f5e325f2a1338a7e5d3bf32495039abb30d144bfceb68b309f77749c42.jpg", + "table_caption": [ + "Table 4: Prompt Attachment Analysis (%, ↑): The best prompt attachment choice $f_{p}(.)$ is Adding #prompts same as #input frames (T=64)." + ], + "table_footnote": [], + "table_body": "
NTU RGB+D\\( \\left\\| u\\mathcal{B}^{(0)}\\right\\| \\)\\( \\left\\| u\\mathcal{S}^{(1)}\\right\\| \\)\\( \\left\\| u\\mathcal{S}^{(2)}\\right\\| \\)\\( \\left\\| u\\mathcal{S}^{(3)}\\right\\| \\)\\( \\left\\| u\\mathcal{S}^{(4)}\\right\\| \\)
MethodBaseAvgAvgAvgOld New Avg \\( A_{HM} \\)
CONCAT temporal, \\( T' = 64 \\)88.670.362.449.833.6 50.5 35.1 40.3
CONCAT feature, \\( T' = 64 \\)87.782.475.566.957.1 41.5 56.0 48.1
Cross Attention, \\( T' = 64 \\)82.977.472.265.057.1 32.3 55.0 41.2
ADD, \\( T' = 1 \\)88.773.362.745.533.7 47.0 34.8 39.3
ADD, \\( T' = 64 \\) (Ours)87.982.876.868.659.3 57.4 59.2 58.3
", + "bbox": [ + 500, + 412, + 784, + 474 + ], + "page_idx": 13 + }, + { + "type": "text", + "text": "7 Conclusions and Future Work", + "text_level": 1, + "bbox": [ + 215, + 511, + 540, + 527 + ], + "page_idx": 13 + }, + { + "type": "text", + "text": "The problem of continually adapting human action models to new user categories over time has gained prominence with the rising availability of XR devices. However, this setting poses unique challenges: (i) the user may be able to provide only a few samples for training, and (ii) accessing data from earlier sessions may violate privacy considerations. We hence propose a method based on prompt offset tuning to address this problem in this work. Prompt tuning to address learning over newer tasks has been attempted in recent years. However, these works have: (1) typically been designed for image-based tasks, (2) relied on strongly pretrained transformer backbones, (3) required full supervision for new tasks, and (4) exclusively applied prompt tuning to transformer architectures. This work departs from these four characteristics. Our work demonstrates that prompt offset tuning is a promising option to evolve and adapt skeleton-based human action models to new user classes. The careful design of each component of the proposed methodology finds validation in the promising results across well-known skeleton-based action recognition benchmarks. Our ablation studies and analysis corroborate our design choices in our implementation. Looking ahead, it will be interesting to explore how our approach and its design choices adapt when a \"generalist backbone\" trained on a large corpus of action recognition data becomes accessible. Extending our method for differential privacy is another interesting direction of future work.", + "bbox": [ + 212, + 537, + 787, + 840 + ], + "page_idx": 13 + }, + { + "type": "page_number", + "text": "14", + "bbox": [ + 217, + 114, + 235, + 126 + ], + "page_idx": 13 + }, + { + "type": "header", + "text": "P. Garg et al.", + "bbox": [ + 271, + 114, + 362, + 128 + ], + "page_idx": 13 + }, + { + "type": "text", + "text": "References", + "text_level": 1, + "bbox": [ + 217, + 143, + 321, + 159 + ], + "page_idx": 14 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "1. Aich, S., Ruiz-Santaquiteria, J., Lu, Z., Garg, P., Joseph, K.J., Fernandez, A., Balasubramanian, V.N., Kin, K., Wan, C., Camgoz, N.C., Ma, S., De la Torre, F.: Data-free class-incremental hand gesture recognition. In: Proceedings of the IEEE/CVF International Conference on Computer Vision (2023)", + "2. Albrecht, J.P.: How the gdpr will change the world. Eur. Data Prot. L. Rev. 2, 287 (2016)", + "3. Bengio, Y., Léonard, N., Courville, A.: Estimating or propagating gradients through stochastic neurons for conditional computation. arXiv preprint arXiv:1308.3432 (2013)", + "4. Bowman, B., Achille, A., Zancato, L., Trager, M., Perera, P., Paolini, G., Soatto, S.: a-la-carte prompt tuning (apt): Combining distinct data via composable prompting. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 14984-14993 (2023)", + "5. Chaudhry, A., Dokania, P.K., Ajanthan, T., Torr, P.H.: Riemannian walk for incremental learning: Understanding forgetting and intransigence. In: Proceedings of the European conference on computer vision (ECCV). pp. 532-547 (2018)", + "6. Chaudhry, A., Ranzato, M., Rohrbach, M., Elhoseiny, M.: Efficient lifelong learning with a-gem. arXiv preprint arXiv:1812.00420 (2018)", + "7. Chen, Y., Zhao, L., Peng, X., Yuan, J., Metaxas, D.N.: Construct dynamic graphs for hand gesture recognition via spatial-temporal attention. In: Proceedings of the British Machine Vision Conference (BMVC) (2019)", + "8. Chen, Y., Zhang, Z., Yuan, C., Li, B., Deng, Y., Hu, W.: Channel-wise topology refinement graph convolution for skeleton-based action recognition. In: Proceedings of the IEEE/CVF International Conference on Computer Vision. pp. 13359-13368 (2021)", + "9. Dhariwal, P., Jun, H., Payne, C., Kim, J.W., Radford, A., Sutskever, I.: Jukebox: A generative model for music. arXiv preprint arXiv:2005.00341 (2020)", + "0. Dong, S., Hong, X., Tao, X., Chang, X., Wei, X., Gong, Y.: Few-shot class incremental learning via relation knowledge distillation. In: Proceedings of the AAAI Conference on Artificial Intelligence. vol. 35, pp. 1255-1263 (2021)", + "1. Dwivedi, V.P., Bresson, X.: A generalization of transformer networks to graphs. arXiv preprint arXiv:2012.09699 (2020)", + "2. Dwivedi, V.P., Luu, A.T., Laurent, T., Bengio, Y., Bresson, X.: Graph neural networks with learnable structural and positional representations. In: International Conference on Learning Representations (2022), https://openreview.net/forum?id=wTTjnvGphYj", + "3. Hersche, M., Karunaratne, G., Cherubini, G., Benini, L., Sebastian, A., Rahimi, A.: Constrained few-shot class-incremental learning. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 9057-9067 (2022)", + "4. Hinojosa, C., Marquez, M., Arguello, H., Adeli, E., Fei-Fei, L., Niebles, J.C.: Privhar: Recognizing human actions from privacy-preserving lens. In: European Conference on Computer Vision. pp. 314-332. Springer (2022)", + "5. Hou, S., Pan, X., Loy, C.C., Wang, Z., Lin, D.: Learning a unified classifier incrementally via rebalancing. In: Proceedings of the IEEE/CVF conference on computer vision and pattern recognition. pp. 831-839 (2019)", + "6. Jia, M., Tang, L., Chen, B.C., Cardie, C., Belongie, S., Hariharan, B., Lim, S.N.: Visual prompt tuning. In: European Conference on Computer Vision. pp. 709-727. Springer (2022)" + ], + "bbox": [ + 225, + 176, + 785, + 839 + ], + "page_idx": 14 + }, + { + "type": "header", + "text": "POET: Prompt Offset Tuning for Continual Human Action Adaptation", + "bbox": [ + 254, + 114, + 730, + 128 + ], + "page_idx": 14 + }, + { + "type": "page_number", + "text": "15", + "bbox": [ + 767, + 116, + 785, + 126 + ], + "page_idx": 14 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "17. Kirkpatrick, J., Pascanu, R., Rabinowitz, N., Veness, J., Desjardins, G., Rusu, A.A., Milan, K., Quan, J., Ramalho, T., Grabska-Barwinska, A., et al.: Overcoming catastrophic forgetting in neural networks. Proceedings of the national academy of sciences 114(13), 3521-3526 (2017)", + "18. Kumawat, S., Nagahara, H.: Privacy-preserving action recognition via motion difference quantization. In: European Conference on Computer Vision. pp. 518-534. Springer (2022)", + "19. Lacoste, A., Luccioni, A., Schmidt, V., Dandes, T.: Quantifying the carbon emissions of machine learning. arXiv preprint arXiv:1910.09700 (2019)", + "20. Lester, B., Al-Rfou, R., Constant, N.: The power of scale for parameter-efficient prompt tuning. arXiv preprint arXiv:2104.08691 (2021)", + "21. Li, M., Xu, X., Fan, H., Zhou, P., Liu, J., Liu, J.W., Li, J., Keppo, J., Shou, M.Z., Yan, S.: Stprivacy: Spatio-temporal privacy-preserving action recognition. In: Proceedings of the IEEE/CVF International Conference on Computer Vision. pp. 5106-5115 (2023)", + "22. Li, T., Ke, Q., Rahmani, H., Ho, R.E., Ding, H., Liu, J.: Else-net: Elastic semantic network for continual action recognition from skeleton data. In: Proceedings of the IEEE/CVF International Conference on Computer Vision. pp. 13434-13443 (2021)", + "23. Li, X.L., Liang, P.: Prefix-tuning: Optimizing continuous prompts for generation. arXiv preprint arXiv:2101.00190 (2021)", + "24. Li, Y., Si, S., Li, G., Hsieh, C.J., Bengio, S.: Learnable fourier features for multidimensional spatial positional encoding. Advances in Neural Information Processing Systems 34, 15816-15829 (2021)", + "25. Li, Z., Hoiem, D.: Learning without forgetting. IEEE transactions on pattern analysis and machine intelligence 40(12), 2935-2947 (2017)", + "26. Liu, X., Yu, H.F., Dhillon, I., Hsieh, C.J.: Learning to encode position for transformer with continuous dynamical model. In: International conference on machine learning. pp. 6327-6335. PMLR (2020)", + "27. Ma, N., Zhang, H., Li, X., Zhou, S., Zhang, Z., Wen, J., Li, H., Gu, J., Bu, J.: Learning spatial-preserved skeleton representations for few-shot action recognition. In: European Conference on Computer Vision. pp. 174-191. Springer (2022)", + "28. Mallya, A., Lazebnik, S.: Packnet: Adding multiple tasks to a single network by iterative pruning. In: Proceedings of the IEEE conference on Computer Vision and Pattern Recognition. pp. 7765-7773 (2018)", + "29. Mialon, G., Chen, D., Selosse, M., Mairal, J.: Graphit: Encoding graph structure in transformers. arXiv preprint arXiv:2106.05667 (2021)", + "30. Park, J.J., Florence, P., Straub, J., Newcombe, R., Lovegrove, S.: Deepsdf: Learning continuous signed distance functions for shape representation. In: Proceedings of the IEEE/CVF conference on computer vision and pattern recognition. pp. 165-174 (2019)", + "31. Peng, C., Zhao, K., Wang, T., Li, M., Lovell, B.C.: Few-shot class-incremental learning from an open-set perspective. In: European Conference on Computer Vision. pp. 382-397. Springer (2022)", + "32. Pernici, F., Bruni, M., Baecchi, C., Turchini, F., Del Bimbo, A.: Class-incremental learning with pre-allocated fixed classifiers. In: 2020 25th International Conference on Pattern Recognition (ICPR). pp. 6259–6266. IEEE (2021)", + "33. Radford, A., Kim, J.W., Hallacy, C., Ramesh, A., Goh, G., Agarwal, S., Sastry, G., Askell, A., Mishkin, P., Clark, J., et al.: Learning transferable visual models from natural language supervision. In: International conference on machine learning. pp. 8748-8763. PMLR (2021)" + ], + "bbox": [ + 215, + 147, + 785, + 839 + ], + "page_idx": 15 + }, + { + "type": "page_number", + "text": "16", + "bbox": [ + 217, + 114, + 235, + 126 + ], + "page_idx": 15 + }, + { + "type": "header", + "text": "P. Garg et al.", + "bbox": [ + 271, + 114, + 362, + 128 + ], + "page_idx": 15 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "34. Razdaibiedina, A., Mao, Y., Hou, R., Khabsa, M., Lewis, M., Almahairi, A.: Progressive prompts: Continual learning for language models. arXiv preprint arXiv:2301.12314 (2023)", + "35. Rebuffi, S.A., Bilen, H., Vedaldi, A.: Efficient parametrization of multi-domain deep neural networks. In: Proceedings of the IEEE conference on computer vision and pattern recognition. pp. 8119-8127 (2018)", + "36. Ren, B., Liu, M., Ding, R., Liu, H.: A survey on 3d skeleton-based action recognition using learning method. Cyborg and Bionic Systems (2020)", + "37. Ridnik, T., Ben-Baruch, E., Noy, A., Zelnik-Manor, L.: Imagenet-21k pretraining for the masses. arXiv preprint arXiv:2104.10972 (2021)", + "38. Rusu, A.A., Rabinowitz, N.C., Desjardins, G., Soyer, H., Kirkpatrick, J., Kavukcuoglu, K., Pascanu, R., Hadsell, R.: Progressive neural networks. arXiv preprint arXiv:1606.04671 (2016)", + "39. Shahroudy, A., Liu, J., Ng, T.T., Wang, G.: Ntu rgb+ d: A large scale dataset for 3d human activity analysis. In: Proceedings of the IEEE conference on computer vision and pattern recognition. pp. 1010-1019 (2016)", + "40. Smedt, Q.D., Wannous, H., Vandeborre, J.P., Guerry, J., Saux, B.L., Filliat, D.: 3D Hand Gesture Recognition Using a Depth and Skeletal Dataset. In: Pratikakis, I., Dupont, F., Ovsjanikov, M. (eds.) Eurographics Workshop on 3D Object Retrieval. The Eurographics Association (2017). https://doi.org/10.2312/3dor.20171049", + "41. Smith, J.S., Karlinsky, L., Gutta, V., Cascante-Bonilla, P., Kim, D., Arbelle, A., Panda, R., Feris, R., Kira, Z.: Coda-prompt: Continual decomposed attention-based prompting for rehearsal-free continual learning. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 11909-11919 (2023)", + "42. Tang, Y.M., Peng, Y.X., Zheng, W.S.: When prompt-based incremental learning does not meet strong pretraining. arXiv preprint arXiv:2308.10445 (2023)", + "43. Tao, X., Hong, X., Chang, X., Dong, S., Wei, X., Gong, Y.: Few-shot class-incremental learning. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 12183-12192 (2020)", + "44. Van Den Oord, A., Vinyals, O., et al.: Neural discrete representation learning. Advances in neural information processing systems 30 (2017)", + "45. Villa, A., Alcazar, J.L., Alfarra, M., Alhamoud, K., Hurtado, J., Heilbron, F.C., Soto, A., Ghanem, B.: Pivot: Prompting for video continual learning. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 24214-24223 (2023)", + "46. Wang, X., Zhang, S., Qing, Z., Gao, C., Zhang, Y., Zhao, D., Sang, N.: Molo: Motion-augmented long-short contrastive learning for few-shot action recognition. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 18011-18021 (2023)", + "47. Wang, Y., Huang, Z., Hong, X.: S-prompts learning with pre-trained transformers: An occam's razor for domain incremental learning. Advances in Neural Information Processing Systems 35, 5682-5695 (2022)", + "48. Wang, Z., Zhang, Z., Ebrahimi, S., Sun, R., Zhang, H., Lee, C.Y., Ren, X., Su, G., Perot, V., Dy, J., et al.: Dualprompt: Complementary prompting for rehearsal-free continual learning. In: European Conference on Computer Vision. pp. 631-648. Springer (2022)", + "49. Wang, Z., Zhang, Z., Lee, C.Y., Zhang, H., Sun, R., Ren, X., Su, G., Perot, V., Dy, J., Pfister, T.: Learning to prompt for continual learning. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 139-149 (2022)" + ], + "bbox": [ + 212, + 146, + 787, + 839 + ], + "page_idx": 16 + }, + { + "type": "header", + "text": "POET: Prompt Offset Tuning for Continual Human Action Adaptation", + "bbox": [ + 254, + 114, + 730, + 128 + ], + "page_idx": 16 + }, + { + "type": "page_number", + "text": "17", + "bbox": [ + 767, + 116, + 785, + 126 + ], + "page_idx": 16 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "50. Williams, W., Ringer, S., Ash, T., MacLeod, D., Dougherty, J., Hughes, J.: Hierarchical quantized autoencoders. Advances in Neural Information Processing Systems 33, 4524-4535 (2020)", + "51. Yang, Y., Yuan, H., Li, X., Lin, Z., Torr, P., Tao, D.: Neural collapse inspired feature-classifier alignment for few-shot class-incremental learning. In: The Eleventh International Conference on Learning Representations (2023), https://openreview.net/forum?id=y5W8tpojhtJ", + "52. Yue, R., Tian, Z., Du, S.: Action recognition based on rgb and skeleton data sets: A survey. Neurocomputing (2022)", + "53. Zhang, H.B., Zhang, Y.X., Zhong, B., Lei, Q., Yang, L., Du, J.X., Chen, D.S.: A comprehensive survey of vision-based human action recognition methods. Sensors 19(5), 1005 (2019)", + "54. Zheng, C., Vedaldi, A.: Online clustered codebook. In: Proceedings of the IEEE/CVF International Conference on Computer Vision. pp. 22798-22807 (2023)", + "55. Zhou, D.W., Wang, F.Y., Ye, H.J., Ma, L., Pu, S., Zhan, D.C.: Forward compatible few-shot class-incremental learning. In: Proceedings of the IEEE/CVF conference on computer vision and pattern recognition. pp. 9046-9056 (2022)", + "56. Zhu, A., Ke, Q., Gong, M., Bailey, J.: Adaptive local-component-aware graph convolutional network for one-shot skeleton-based action recognition. In: Proceedings of the IEEE/CVF Winter Conference on Applications of Computer Vision. pp. 6038-6047 (2023)", + "57. Zhu, B., Niu, Y., Han, Y., Wu, Y., Zhang, H.: Prompt-aligned gradient for prompt tuning. In: Proceedings of the IEEE/CVF International Conference on Computer Vision. pp. 15659-15669 (2023)" + ], + "bbox": [ + 215, + 146, + 787, + 481 + ], + "page_idx": 17 + }, + { + "type": "page_number", + "text": "18", + "bbox": [ + 217, + 114, + 235, + 126 + ], + "page_idx": 17 + }, + { + "type": "header", + "text": "P. Garg et al.", + "bbox": [ + 271, + 114, + 362, + 128 + ], + "page_idx": 17 + } +] \ No newline at end of file diff --git a/2024/POET_ Prompt Offset Tuning for Continual Human Action Adaptation/9942db11-78b9-4c6d-815f-c44748e2ea7c_model.json b/2024/POET_ Prompt Offset Tuning for Continual Human Action Adaptation/9942db11-78b9-4c6d-815f-c44748e2ea7c_model.json new file mode 100644 index 0000000000000000000000000000000000000000..2404a0fc84df97262436f7680fc44f264ec689a4 --- /dev/null +++ b/2024/POET_ Prompt Offset Tuning for Continual Human Action Adaptation/9942db11-78b9-4c6d-815f-c44748e2ea7c_model.json @@ -0,0 +1,2392 @@ +[ + [ + { + "type": "title", + "bbox": [ + 0.251, + 0.141, + 0.756, + 0.187 + ], + "angle": 0, + "content": "POET: Prompt Offset Tuning for Continual Human Action Adaptation" + }, + { + "type": "text", + "bbox": [ + 0.238, + 0.213, + 0.767, + 0.259 + ], + "angle": 0, + "content": "Prachi Garg\\(^{1}\\), K J Joseph\\(^{3}\\), Vineeth N Balasubramanian\\(^{3}\\), Necati Cihan Camgoz\\(^{2}\\), Chengde Wan\\(^{2}\\), Kenrick Kin\\(^{2}\\), Weiguang Si\\(^{2}\\), Shugao Ma\\(^{2}\\), and Fernando De La Torre\\(^{1}\\)" + }, + { + "type": "text", + "bbox": [ + 0.382, + 0.27, + 0.621, + 0.286 + ], + "angle": 0, + "content": "1 Carnegie Mellon University, USA" + }, + { + "type": "text", + "bbox": [ + 0.432, + 0.286, + 0.573, + 0.299 + ], + "angle": 0, + "content": "Meta Reality Labs" + }, + { + "type": "text", + "bbox": [ + 0.353, + 0.299, + 0.651, + 0.313 + ], + "angle": 0, + "content": "3 Indian Institute of Technology, Hyderabad" + }, + { + "type": "text", + "bbox": [ + 0.261, + 0.357, + 0.744, + 0.662 + ], + "angle": 0, + "content": "Abstract. As extended reality (XR) is redefining how users interact with computing devices, research in human action recognition is gaining prominence. Typically, models deployed on immersive computing devices are static and limited to their default set of classes. The goal of our research is to provide users and developers with the capability to personalize their experience by adding new action classes to their device models continually. Importantly, a user should be able to add new classes in a low-shot and efficient manner, while this process should not require storing or replaying any of user's sensitive training data. We formalize this problem as privacy-aware few-shot continual action recognition. Towards this end, we propose POET: Prompt-offset Tuning. While existing prompt tuning approaches have shown great promise for continual learning of image, text, and video modalities; they demand access to extensively pretrained transformers. Breaking away from this assumption, POET demonstrates the efficacy of prompt tuning a significantly lightweight backbone, pretrained exclusively on the base class data. We propose a novel spatio-temporal learnable prompt offset tuning approach, and are the first to apply such prompt tuning to Graph Neural Networks. We contribute two new benchmarks for our new problem setting in human action recognition: (i) NTU RGB+D dataset for activity recognition, and (ii) SHREC-2017 dataset for hand gesture recognition. We find that POET consistently outperforms comprehensive benchmarks." + }, + { + "type": "text", + "bbox": [ + 0.262, + 0.675, + 0.741, + 0.704 + ], + "angle": 0, + "content": "Keywords: 3D Skeleton Activity Recognition \\(\\cdot\\) Extended Reality (XR) \\(\\cdot\\) Continual Learning \\(\\cdot\\) Prompt Tuning." + }, + { + "type": "title", + "bbox": [ + 0.217, + 0.734, + 0.377, + 0.75 + ], + "angle": 0, + "content": "1 Introduction" + }, + { + "type": "text", + "bbox": [ + 0.216, + 0.769, + 0.788, + 0.8 + ], + "angle": 0, + "content": "A key input modality to virtual, augmented and mixed reality (often together termed as extended reality, XR) devices today is through recognizing human" + }, + { + "type": "page_footnote", + "bbox": [ + 0.218, + 0.811, + 0.787, + 0.84 + ], + "angle": 0, + "content": "Source Code at https://github.com/humansensinglab/POET-continual-action-recognition" + } + ], + [ + { + "type": "page_number", + "bbox": [ + 0.218, + 0.116, + 0.23, + 0.127 + ], + "angle": 0, + "content": "2" + }, + { + "type": "header", + "bbox": [ + 0.272, + 0.115, + 0.364, + 0.129 + ], + "angle": 0, + "content": "P. Garg et al." + }, + { + "type": "image", + "bbox": [ + 0.223, + 0.144, + 0.784, + 0.279 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.214, + 0.288, + 0.789, + 0.359 + ], + "angle": 0, + "content": "Fig. 1: Proposed POET method continually adapts skeleton-based human action recognition models pretrained on a pre-defined set of categories to new user categories with few training examples. Users can thus expand the capabilities of XR systems with novel action classes by providing a few examples of each new class. We discard the user-sensitive data as soon as the model is updated on the new categories." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.369, + 0.789, + 0.52 + ], + "angle": 0, + "content": "activity and hand gestures based on body and hand pose estimates. Recognizing human actions5 facilitates seamless user interactions in head-mounted XR devices such as the Meta Quest 3 and Apple Vision Pro. If the provided action recognition models are static, then developers and users are limited to a predefined set of action categories. With the growing use of such devices in new contexts and the increasing demand for personalized technology delivery, there is an impending need to enable the action recognition models in such systems to adapt and learn new user actions over time. Defining their own action categories allows users to customize their experience and expand the functionality of their XR devices. Addressing this need is the primary objective of this work." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.521, + 0.789, + 0.686 + ], + "angle": 0, + "content": "Adapting human action models to new user categories over time faces a few challenges. Firstly, the model must be capable of learning new actions with minimal amount of training data so users can add new classes by providing just a few training examples per class. Secondly, due to the increasing use of XR devices for personal assistance, there is a need for privacy preservation in user action recognition-based pipelines [2, 14]. Hence, the adaptation of such action recognition models to new user categories must also be 'data-free', i.e., it cannot store and replay previously seen user training data in subsequent continual sessions. Considering these requirements, we leverage the recent success of 'data-free' prompt-based learning [49] and propose a new spatio-temporal prompt offset tuning approach to efficiently adapt the default model without finetuning." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.687, + 0.789, + 0.808 + ], + "angle": 0, + "content": "Human action recognition systems are moving to skeleton-based approaches, especially in applications that require low-shot action recognition capabilities such as medical action recognition [27, 56]. Skeletons offer a robust and compact alternative to videos in such low-shot regimes, due to their relatively low dimensionality and lesser variance under background conditions. While there have been a wide variety of efforts in skeleton-based human action recognition over the years [36, 52, 53], there have been fewer efforts on adapting such models to newer user categories. Efforts like [1, 22] attempted to continually learn new user" + }, + { + "type": "page_footnote", + "bbox": [ + 0.218, + 0.816, + 0.788, + 0.84 + ], + "angle": 0, + "content": "5 We use human action as an umbrella term for both hand gesture and body activity in this work for ease of presentation." + } + ], + [ + { + "type": "header", + "bbox": [ + 0.255, + 0.115, + 0.732, + 0.129 + ], + "angle": 0, + "content": "POET: Prompt Offset Tuning for Continual Human Action Adaptation" + }, + { + "type": "page_number", + "bbox": [ + 0.775, + 0.117, + 0.787, + 0.127 + ], + "angle": 0, + "content": "3" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.147, + 0.788, + 0.283 + ], + "angle": 0, + "content": "categories over time in skeleton-based human action recognition, but relied on fully-supervised data for the new classes. On the other hand, few-shot learning works [27, 46, 56] adapt a pretrained skeleton-based action recognition model to new data, but without explicitly retaining past categories. In this work, we seek to learn new user categories in trained human action models with very few labeled samples for the new classes, while being data-free (not storing samples from previously trained categories). Fig 1 summarizes our overall objective. One could view our setting as privacy-aware few-shot continual learning for skeleton-based action recognition." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.283, + 0.788, + 0.45 + ], + "angle": 0, + "content": "To this end, we propose a prompt offset tuning methodology that can be integrated with existing backbone architectures for skeleton-based human action recognition. Our learnable (soft) prompts are selected from a shared knowledge pool of prompts based on an input instance dependent attention mechanism. In particular, we propose prompt selection using an ordered query-key matching that enables a temporal prompt frame order selection consistent with the input instance. We show that such an approach allows us to learn new user categories without having to store data from past classes, without overwriting the pre-existing categories. To the best of our knowledge, this is the first effort on leveraging prompt tuning for skeleton-based models, as well as on spatiotemporal prompt selection and tuning." + }, + { + "type": "text", + "bbox": [ + 0.215, + 0.453, + 0.547, + 0.468 + ], + "angle": 0, + "content": "Our key contributions are summarized below:" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.474, + 0.784, + 0.5 + ], + "angle": 0, + "content": "- We formalize a novel problem setting which continually adapts human action models to new user categories over time, in a privacy aware manner." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.504, + 0.784, + 0.56 + ], + "angle": 0, + "content": "- To address this problem, we propose a novel spatio-temporal Prompt OffsEt Tuning methodology (POET). In particular, it is designed to seamlessly plug-and-play with a pretrained model's input embedding, without any significant architectural changes." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.563, + 0.784, + 0.591 + ], + "angle": 0, + "content": "- Our comprehensive experimental evaluation on two benchmark datasets brings out the efficacy of our proposed approach." + }, + { + "type": "title", + "bbox": [ + 0.215, + 0.616, + 0.398, + 0.632 + ], + "angle": 0, + "content": "2 Related Works" + }, + { + "type": "title", + "bbox": [ + 0.215, + 0.65, + 0.393, + 0.665 + ], + "angle": 0, + "content": "2.1 Prompt Tuning" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.674, + 0.788, + 0.841 + ], + "angle": 0, + "content": "The idea of prompting, as it originated from Large Language Models (LLMs), is to include additional information, known as a text prompt, to condition the model's input for generating an output relevant to the prompt. Instead of applying a discrete, pre-defined 'hard' language prompt token, prompt and prefix tuning [20,23] formalized the concept of applying 'soft prompts' to the input. A set of learnable parameters are prepended (concatenated) to the input text and trained along with the classifier while keeping the backbone parameters frozen. Similar to prompt tuning of LLMs, recent works have popularized prompt tuning of ViTs [16] as an effective way of adapting large pretrained models to downstream tasks [49,57]. However, it remains unexplored and undefined (to the best of our knowledge) for non-transformer architectures such as GNNs." + } + ], + [ + { + "type": "page_number", + "bbox": [ + 0.218, + 0.116, + 0.23, + 0.127 + ], + "angle": 0, + "content": "4" + }, + { + "type": "header", + "bbox": [ + 0.272, + 0.115, + 0.364, + 0.129 + ], + "angle": 0, + "content": "P. Garg et al." + }, + { + "type": "title", + "bbox": [ + 0.216, + 0.147, + 0.584, + 0.163 + ], + "angle": 0, + "content": "2.2 Prompt Tuning for Continual Learning" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.177, + 0.788, + 0.388 + ], + "angle": 0, + "content": "Prompt tuning provides a simple and cost-effective way of learning task-specific signal condensed into 'soft prompts'. For continual learning, training a set of prompts for each sequential task provides a natural alternative to storing privacy violating exemplars and replaying them. Training task-specific prompts for each sequential task is straightforward when authors assume access to task identity at both train and inference time, like in Progressive Prompts [34]. However, if task identity is unavailable at inference, the model will not know which task's prompts or classifier to use for evaluating a test sample. In this respect, S-prompts [47] and A-la-carte prompt tuning (APT) [4] learn an independent set of prompts for each domain/task and employ a KNN-based search for domain/task identity at test time. Since these methods learn stand-alone prompts for every task, the prompt feature space is task-specific, and there is no forgetting of old knowledge when learning new tasks (by design). At the same time however, these 'no forgetting' prompts cannot share knowledge across tasks." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.389, + 0.788, + 0.585 + ], + "angle": 0, + "content": "This leads to another ideology for continual prompt tuning, i.e., treat each prompt unit as being a part of a larger shared (knowledge) pool of prompts. Then the desired number of prompt units can be selected from the pool, conditioned on the input instance itself [41, 48, 49]. Given the scarcity of new data in our setting, we hypothesize that sharing of knowledge will benefit new tasks and draw inspiration from this line of works. Most recently, Adaptive Prompt Generator (APG) [42] challenges the intensive ImageNet21K pre-training assumption as it prompts a ViT pretrained only on the continual benchmark's base class data (similar to us). However, they use replay and knowledge distillation-style 'antiforgetting learning', in addition to using prompts. Even though our backbone is trained only on the base classes, we propose a simple prompt tuning-only strategy to counter forgetting. This implies that a prompt strategy is all we need to continually add new action semantics in a few-shot manner." + }, + { + "type": "title", + "bbox": [ + 0.216, + 0.611, + 0.573, + 0.627 + ], + "angle": 0, + "content": "2.3 Few-Shot Class Incremental Learning" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.641, + 0.788, + 0.853 + ], + "angle": 0, + "content": "FSCIL is a challenging continual learning setting where a model overfits to new classes, with the simultaneous heightened (often complete) forgetting of old knowledge as soon as the base model is fine-tuned on few-shot data [10,43]. Since the backbone feature extractor is the only source of previously seen knowledge, if it is updated, knowledge is lost forever. Typically, existing works decouple the learning of (backbone) feature representations from the classifier by learning the model only on the base data and relying on non-parametric class-mean classifiers for classification in subsequent steps [13,31,55]. This leads to a feature-classifier misalignment issue [32, 51] because new class prototypes are extracted from a backbone representation trained only on the base classes. We hypothesize that optimizing input prompt vectors along with a dynamically expanding parametric classifier on top of a frozen backbone can alleviate this misalignment issue. Our work not only provides a fresh perspective into FSCIL, but to our best knowledge is also the only work not designed for and evaluated on image benchmarks." + } + ], + [ + { + "type": "header", + "bbox": [ + 0.254, + 0.114, + 0.733, + 0.131 + ], + "angle": 0, + "content": "POET: Prompt Offset Tuning for Continual Human Action Adaptation" + }, + { + "type": "page_number", + "bbox": [ + 0.775, + 0.116, + 0.787, + 0.128 + ], + "angle": 0, + "content": "5" + }, + { + "type": "title", + "bbox": [ + 0.216, + 0.145, + 0.383, + 0.161 + ], + "angle": 0, + "content": "3 Preliminaries" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.173, + 0.788, + 0.386 + ], + "angle": 0, + "content": "Skeleton Action Recognition Using Graph Representations. Our input \\(\\mathbf{X} \\in \\mathbb{R}^{T \\times J \\times 3}\\) is a video sequence of \\(T\\) frames, each frame containing \\(J\\) joints of the human body (25 joints) or hand skeleton (22 joints) in 3D Cartesian coordinate system. Such a skeleton action sequence is naturally represented as a graph topology \\(G = \\{\\mathcal{V}, \\mathcal{E}\\}\\) with \\(\\mathcal{V}\\) vertices and \\(\\mathcal{E}\\) edges. Graphs are modeled using Graph Neural Networks (GNNs) [12], which can either be sparse graph convolutional networks (GCN) or fully connected graph transformers (GT). Our main model (a GNN) is defined as \\(f(\\mathbf{X}) = f_c \\circ f_g \\circ f_e(\\mathbf{X})\\) (as also shown in Fig. 2). Input \\(\\mathbf{X}\\) is first passed through an input embedding layer \\(f_e\\) to get an embedding of human joints \\(\\mathbf{X_e} = f_e(\\mathbf{X}), \\mathbf{X_e} \\in \\mathbb{R}^{T \\times J \\times C_e}\\), with feature dimension \\(C_e\\). \\(\\mathbf{X_e}\\) is further passed to a graph feature extractor \\(f_g\\) composed of a stack of convolutional layers (in GCNs) or attention layers (in GTs), and finally a classifier \\(f_c\\) which predicts the action class label \\(\\mathbf{y}\\). In POET, we propose to attach learnable parameters \\(\\mathbf{P_T}\\) (called prompt offsets) to the embedding \\(\\mathbf{X_e}\\)." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.389, + 0.789, + 0.688 + ], + "angle": 0, + "content": "Problem Definition. Given a default (pre)trained model deployed on a user's device, we would like to extend this model to new action classes over \\(T\\) subsequent user sessions (also called tasks) \\(\\{\\mathcal{US}^{(1)},\\dots,\\mathcal{US}^{(T)}\\}^6\\). In each user session \\(\\mathcal{US}^{(t)}\\), the model learns a dataset \\(\\mathcal{D}^{(t)} = (\\mathbf{X}_i^t,\\mathbf{y}_i^t)_{i = 1}^{|D^{(t)}|}\\) of skeleton action sequence and label pairs provided by the user, \\(\\mathbf{X}_i^t\\in \\mathbb{R}^{T\\times J\\times 3}\\), \\(y_i^t\\in \\mathbb{R}^{y^{(t)}}\\). In each session, the user typically provides a few training instances \\(F\\) (e.g. \\(F\\leq 5\\)) for each of the \\(N\\) new classes being added, such that \\(|\\mathcal{D}^{(t)}| = NF\\). The base (default) model's session \\(\\mathcal{UB}^{(0)}\\) is assumed to have a large number of default action classes \\(\\mathcal{V}^{(0)}\\) trained on sufficient data \\(\\mathcal{D}^{(0)}\\), which is most often proprietary and cannot be accessed in later user sessions. In each session, the user adds new action classes such that, \\(\\mathcal{V}^{(t)}\\cap \\mathcal{V}^{(t')} = \\emptyset ,\\forall t\\neq t'^7\\). Due to the aforementioned privacy constraints, in any training session \\(\\mathcal{US}^{(t)}\\), the model has access to only \\(\\mathcal{D}^{(t)}\\); after training this data is made inaccessible for use in subsequent sessions (no exemplar or prototypes stored). After training on every new session \\(\\mathcal{US}^{(t)}\\), the model is evaluated on the test set of all classes seen so far \\(\\cup_{i = 0}^{t}\\mathcal{V}^{(i)}\\). The challenge is to alleviate forgetting of old classes while not overfitting to the user-provided new class samples. One could view our setting as privacy-aware few-shot continual action recognition, a problem of practical relevance in human action recognition - which has not received adequate attention." + }, + { + "type": "title", + "bbox": [ + 0.216, + 0.704, + 0.694, + 0.723 + ], + "angle": 0, + "content": "4 Methodology: Prompt Offset Tuning (POET)" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.731, + 0.788, + 0.793 + ], + "angle": 0, + "content": "Overview. We propose to prompt tune a base GNN model \\( f(\\cdot) \\) by prompts \\( \\mathbf{P_T} \\) to address our overall objective. As shown in Fig. 3, for each input instance \\( \\mathbf{X} \\), corresponding prompts \\( \\mathbf{P_T} \\) are selected from a pool of prompt parameters, using an input-dependent query and key attention mechanism. The selected prompts" + }, + { + "type": "page_footnote", + "bbox": [ + 0.218, + 0.804, + 0.556, + 0.816 + ], + "angle": 0, + "content": "6 User sessions may be spaced at arbitrary time intervals." + }, + { + "type": "page_footnote", + "bbox": [ + 0.222, + 0.817, + 0.786, + 0.84 + ], + "angle": 0, + "content": "We make this assumption considering this is a first of such efforts; allowing for overlapping action classes and users to 'update' older classes would be interesting extensions of our proposed work." + }, + { + "type": "list", + "bbox": [ + 0.218, + 0.804, + 0.786, + 0.84 + ], + "angle": 0, + "content": null + } + ], + [ + { + "type": "page_number", + "bbox": [ + 0.218, + 0.116, + 0.23, + 0.127 + ], + "angle": 0, + "content": "6" + }, + { + "type": "header", + "bbox": [ + 0.272, + 0.115, + 0.364, + 0.129 + ], + "angle": 0, + "content": "P. Garg et al." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.147, + 0.784, + 0.178 + ], + "angle": 0, + "content": "are added to the input feature embedding (and hence the term 'prompt offsets'), before forwarding to the feature extractor and classifier (shown in Fig. 2)." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.178, + 0.574, + 0.33 + ], + "angle": 0, + "content": "To this end, our method, POET uses the same number of prompts as the number of temporal frames in the input, to maintain temporal consistency between the prompt and the input. Focusing solely on prompt offsets allows us to adapt the model to subsequent user sessions without having to update the input embeddings or the feature extraction backbone. Our prompt selection mechanism is learnable and trained along with the classifier to make this method simple and efficient." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.334, + 0.574, + 0.485 + ], + "angle": 0, + "content": "What are Prompt Offsets? Learnable (or soft [20]) prompts are parameter vectors in a continuous space which are optimized to adapt the pretrained frozen backbone \\( f_{g} \\) to each continual task. We define our spatio-temporal prompt offsets \\( \\mathbf{P_T} \\) as a set of \\( T \\) prompts (same in number as skeletal frames in input), each prompt \\( P_{i} \\) having length equal to the number of joints in a frame \\( J \\) and feature dimension same as input feature embedding \\( \\mathbf{X_e} \\), i.e., \\( P_{i} \\in \\mathbb{R}^{J \\times C_{e}} \\)." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.486, + 0.574, + 0.546 + ], + "angle": 0, + "content": "Existing prompt tuning efforts, for example in image classification, focus on concatenating learnable prompts to the input token sequence in transformer architectures [16, 20]. Even though trans" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.546, + 0.788, + 0.698 + ], + "angle": 0, + "content": "formers can be generalized to graphs [7, 11, 29], it is non-trivial to attach prompts to a GNN. This is because transformers can be viewed as treating sentences or images as fully connected graphs where any word (or image patch) can attend to any other word in the sentence [12]. However, our input is a spatio-temporal graph skeleton of the human joint-bone structure with its own edge connectivity. Concatenating prompts along spatial or temporal dimensions would affect the graph semantics, and also affect standard training strategies such as a forward pass or backpropagation (especially in GCNs). Hence, we attach the selected prompts \\(\\mathbf{P_T}\\) to the corresponding input feature embedding \\(\\mathbf{X_e}\\) via a prompt attachment operator \\(f_{p}(.)\\). The class logit distribution \\(\\mathbf{y}\\) is thus obtained as:" + }, + { + "type": "equation", + "bbox": [ + 0.355, + 0.703, + 0.787, + 0.721 + ], + "angle": 0, + "content": "\\[\n\\mathbf {y} = f (\\mathbf {X}, \\mathbf {P} _ {\\mathbf {T}}) = f _ {c} \\circ f _ {g} \\circ f _ {p} \\left(f _ {e} (\\mathbf {X}), \\mathbf {P} _ {\\mathbf {T}}\\right) \\tag {1}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.727, + 0.788, + 0.806 + ], + "angle": 0, + "content": "In every user session \\( t > 0 \\), the classifier output dimension expands by \\( N \\) to accommodate the new action classes. Unlike most existing continual prompt tuning works, our feature extractor backbone \\( f_{g} \\) is trained only on the base class data \\( \\mathcal{D}^{(0)} \\) and is never fine-tuned on classes from new user sessions \\( \\mathcal{U}\\mathcal{S}^{(t)}, t > 0 \\). After the base session training, parameters of \\( f_{g}, f_{e} \\) are frozen." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.81, + 0.788, + 0.842 + ], + "angle": 0, + "content": "Prompt Pool Design. As stated in Sec. 2.2, to encourage knowledge sharing across user sessions, we choose to construct a single prompt pool \\(\\mathbf{P}\\) which encodes" + }, + { + "type": "image", + "bbox": [ + 0.589, + 0.196, + 0.782, + 0.412 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.582, + 0.413, + 0.788, + 0.538 + ], + "angle": 0, + "content": "Fig.2: POET: Prompt-offset Tuning proposes to offset the input feature embedding \\(\\mathbf{X}_{\\mathrm{e}}\\) of the main model by learnable prompt parameters \\(\\mathbf{P_T}\\) for privacy-aware few-shot continual action recognition. We explain prompt selection mechanism in Fig. 3." + } + ], + [ + { + "type": "header", + "bbox": [ + 0.255, + 0.115, + 0.733, + 0.13 + ], + "angle": 0, + "content": "POET: Prompt Offset Tuning for Continual Human Action Adaptation" + }, + { + "type": "page_number", + "bbox": [ + 0.776, + 0.117, + 0.786, + 0.127 + ], + "angle": 0, + "content": "7" + }, + { + "type": "image", + "bbox": [ + 0.231, + 0.144, + 0.768, + 0.281 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.216, + 0.282, + 0.788, + 0.348 + ], + "angle": 0, + "content": "Fig. 3: Selection of our prompts \\(\\mathbf{P_T}\\): Input-dependent query \\(q\\) is matched with keys \\(K\\) using sorted cosine similarity to get an ordered index sequence \\((s_i)_{i=1}^T\\) of the top \\(T\\) keys. This ordered index sequence is used to select the corresponding ordered prompt sequence \\(\\mathbf{P_T}\\) from prompt pool \\(\\mathbf{P}\\). We add \\(\\mathbf{P_T}\\) to \\(\\mathbf{X_e}\\), thereby adding an offset to it. Our experimental evaluation confirms that such an additive spatio-temporal prompt offset can balance the plasticity to learn new classes from a few action samples, while maintaining stability on previously learned classes." + }, + { + "type": "text", + "bbox": [ + 0.216, + 0.361, + 0.435, + 0.375 + ], + "angle": 0, + "content": "knowledge across the sessions:" + }, + { + "type": "equation", + "bbox": [ + 0.261, + 0.38, + 0.786, + 0.399 + ], + "angle": 0, + "content": "\\[\n\\mathbf {P} = \\left\\{P _ {1},.. P _ {i},..., P _ {M} \\right\\}, \\qquad P _ {i} \\in \\mathbb {R} ^ {J \\times C _ {e}}; M = \\# \\text {p r o m p t s a t t i m e t} \\tag {2}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.215, + 0.405, + 0.784, + 0.48 + ], + "angle": 0, + "content": "For selecting prompts from this pool (Fig. 3), we construct a bijective key-value codebook, treating prompts in the pool \\(\\mathbf{P}\\) as values and defining learnable key vectors \\(\\pmb{K} = \\{\\pmb{k}_1, \\dots, \\pmb{k}_i, \\dots, \\pmb{k}_M\\}, \\pmb{k}_i \\in \\mathbb{R}^{C_e}\\). A cosine similarity matching \\(\\gamma(.)\\) between the query \\(\\pmb{q}\\) and keys \\(\\pmb{K}\\) is used to find indices of the \\(T\\) closest keys \\(\\mathbb{Z}\\), which in turn are used to select prompts from the pool:" + }, + { + "type": "equation", + "bbox": [ + 0.409, + 0.485, + 0.786, + 0.511 + ], + "angle": 0, + "content": "\\[\n\\mathbb {Z} = \\underset {T} {\\operatorname {a r g m a x}} \\gamma \\left(f _ {q} (\\mathbf {X}), \\mathbf {K}\\right) \\tag {3}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.216, + 0.518, + 0.784, + 0.549 + ], + "angle": 0, + "content": "This quantization process is enabled by a query function \\( f_{q}(.) \\), which is a pretrained encoder that maps an input instance \\( \\mathbf{X} \\) to a query \\( \\pmb{q} \\) as:" + }, + { + "type": "equation", + "bbox": [ + 0.298, + 0.552, + 0.786, + 0.573 + ], + "angle": 0, + "content": "\\[\n\\boldsymbol {q} = f _ {q} (\\mathbf {X}) = f _ {Q A} \\circ f _ {g} ^ {\\prime} \\circ f _ {e} ^ {\\prime} (\\mathbf {X}), \\quad f _ {q}: \\mathbb {R} ^ {T \\times J \\times 3} \\longrightarrow \\mathbb {R} ^ {C _ {e}} \\tag {4}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.215, + 0.579, + 0.784, + 0.609 + ], + "angle": 0, + "content": "where the query adaptor \\( f_{QA} \\) is a fully connected layer mapping the \\( f_{g}^{\\prime} \\) output dimension to the desired prompt embedding dimension \\( C_e \\)." + }, + { + "type": "text", + "bbox": [ + 0.215, + 0.614, + 0.789, + 0.844 + ], + "angle": 0, + "content": "Coupled Optimization in User Sessions \\( t > 0 \\). Typically, the argmax operator in Eq. 3 decouples the optimization of keys from the prompt pool and main model as it prevents backpropagation of gradients to the keys (also seen in earlier works such as [48, 49]). However, this approach does not work for our setting, even more so since we assume no large-scale pre-training of our base model. Due to the lack of off-the-shelf availability of large-scale models for skeletal action data, our query function \\( f_{q} \\) is pretrained only on base class data \\( \\mathcal{D}^{(0)} \\). Hence, it becomes important that \\( f_{q} \\) is updated as the model learns new classes. As shown in red boxes in Figs. 2 and 3, we propose to couple this optimization process such that the overall cross-entropy loss for new tasks updates: (i) the classifier \\( f_{c} \\), (ii) selected prompts in \\( \\mathbf{P} \\), (iii) selected keys in \\( K \\), as well as (iv) query adaptor \\( f_{QA} \\). We achieve this by approximating the gradient for \\( K \\) and \\( f_{QA} \\) by the straight-through estimator reparameterization trick as in [3, 44]. We freeze the query feature extractor layers \\( f_{g}^{\\prime}, f_{e}^{\\prime} \\) in \\( t > 0 \\) to prevent catastrophic forgetting of base knowledge in \\( f_{q} \\). Our cross-entropy loss is hence given by:" + } + ], + [ + { + "type": "page_number", + "bbox": [ + 0.218, + 0.116, + 0.23, + 0.127 + ], + "angle": 0, + "content": "8" + }, + { + "type": "header", + "bbox": [ + 0.272, + 0.115, + 0.364, + 0.129 + ], + "angle": 0, + "content": "P. Garg et al." + }, + { + "type": "equation", + "bbox": [ + 0.396, + 0.147, + 0.785, + 0.172 + ], + "angle": 0, + "content": "\\[\n\\min _ {\\theta_ {f _ {Q A}}, \\theta_ {K}, \\theta_ {\\mathbf {P}}, \\theta_ {f _ {c}}} \\mathcal {L} (f (\\mathbf {X}, \\mathbf {P} _ {\\mathbf {T}}), \\mathbf {y}) \\tag {5}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.178, + 0.784, + 0.208 + ], + "angle": 0, + "content": "To move queries closer to their aligned \\(T\\) keys during training, we use a vector quantization clustering loss inspired from VQ-VAE [44] as:" + }, + { + "type": "equation", + "bbox": [ + 0.404, + 0.211, + 0.786, + 0.243 + ], + "angle": 0, + "content": "\\[\n\\max _ {\\theta_ {f _ {Q A}}, \\theta_ {K}} \\lambda \\sum_ {i \\in \\mathbb {Z}} \\gamma \\left(f _ {q} (\\mathbf {X}), K _ {i}\\right) \\tag {6}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.248, + 0.784, + 0.308 + ], + "angle": 0, + "content": "where \\(\\lambda\\) is the clustering loss coefficient. Our end-to-end optimization thus establishes a prompt optimization framework which is amenable to prompt tuning when extensive pre-training is not possible. This sets the foundation for our spatio-temporal prompt selection module, described next." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.313, + 0.784, + 0.359 + ], + "angle": 0, + "content": "Spatio-Temporal Prompt Selection. In order to ensure that our learned prompts respect temporal information in the input video sequence, we choose the number of selected prompts to be equal to the number of frames in the input" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.36, + 0.488, + 0.553 + ], + "angle": 0, + "content": "video \\(T\\). After coupling the prompt pool and keys, we observed in our initial experiments with pool size \\(M > T\\) that the same set of prompts get selected across training iterations and user sessions (Fig. 4A). More concretely, as the vector quantization loss (Eqn 6) brings the query close to the selected keys, the same set of active prompts get selected and optimized in each iteration, not using other prompts at all. This is similar to the well-known issue of 'codebook col" + }, + { + "type": "image", + "bbox": [ + 0.501, + 0.361, + 0.787, + 0.43 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.5, + 0.436, + 0.785, + 0.504 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.497, + 0.504, + 0.786, + 0.547 + ], + "angle": 0, + "content": "Fig.4: \\(M > T\\) Case: Prompt Pool Collapse. (Top) Certain prompt indices remain unused across user sessions. (Bottom) Our POET pool expansion strategy alleviates pool collapse." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.555, + 0.784, + 0.584 + ], + "angle": 0, + "content": "lapse' in VQ-VAE [9, 50, 54]. Based on this observation, we design two prompt pool update mechanisms in user sessions \\( t > 0 \\) as below:" + }, + { + "type": "text", + "bbox": [ + 0.215, + 0.587, + 0.785, + 0.647 + ], + "angle": 0, + "content": "1. Case 1, \\( M = T \\forall t \\): No pool expansion, Algorithm 1. All prompts are selected in all tasks. But the order of their selection \\( (s_i)_{i=1}^T \\) varies with each input instance as we replace Eq. 3 by sorting the cosine similarity before selecting the top \\( T \\) indices as follows:" + }, + { + "type": "equation", + "bbox": [ + 0.421, + 0.65, + 0.785, + 0.678 + ], + "angle": 0, + "content": "\\[\n\\mathbb {Z} = \\underset {\\left(s _ {i}\\right) _ {i = 1} ^ {T}} {\\operatorname {a r g s o r t}} \\gamma \\left(f _ {q} (\\mathbf {X}), \\mathbf {K}\\right) \\tag {7}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.235, + 0.681, + 0.787, + 0.741 + ], + "angle": 0, + "content": "In Fig. 5, we visualize the positions occupied by indices in this (sorted) ordered key index sequence \\((s_i)_{i=1}^T\\). Entropy increase across tasks \\(t = 1\\) to \\(t = 4\\) (bottom row of figure) shows that our selection mechanism learns to select a unique temporal code for all inputs." + }, + { + "type": "text", + "bbox": [ + 0.215, + 0.741, + 0.784, + 0.802 + ], + "angle": 0, + "content": "2. Case 2, \\( M = T + (R * t), t > 0 \\). Expand pool with \\( R \\) prompts. We also propose an order-aware prompt pool expansion strategy (Appendix B) that selects prompts from an expanded pool in a temporally coherent manner, for \\( t > 0 \\). This alleviates prompt pool collapse as shown in Fig. 4B." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.81, + 0.784, + 0.84 + ], + "angle": 0, + "content": "Prompt Offset Attachment. Since concatenation is not meaningful for graph data, we use addition as our choice for the prompt attachment operator as:" + } + ], + [ + { + "type": "header", + "bbox": [ + 0.254, + 0.114, + 0.733, + 0.131 + ], + "angle": 0, + "content": "POET: Prompt Offset Tuning for Continual Human Action Adaptation" + }, + { + "type": "page_number", + "bbox": [ + 0.775, + 0.116, + 0.787, + 0.127 + ], + "angle": 0, + "content": "9" + }, + { + "type": "equation", + "bbox": [ + 0.414, + 0.146, + 0.786, + 0.164 + ], + "angle": 0, + "content": "\\[\nf _ {p} \\left(\\mathbf {X} _ {\\mathbf {e}}, \\mathbf {P} _ {\\mathbf {T}}\\right) = \\mathbf {X} _ {\\mathbf {e}} + \\mathbf {P} _ {\\mathbf {T}} \\tag {8}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.168, + 0.784, + 0.199 + ], + "angle": 0, + "content": "Hence, we call our approach as prompt offset tuning. We also study this empirically through experiments that support this choice in Sec. 6." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.202, + 0.785, + 0.248 + ], + "angle": 0, + "content": "Interpreting Prompt Offset Tuning of GNNs. Our additive prompt offsets are open to interpretation, as shown in Fig. 5. (i) Adding our selected prompts \\(\\mathbf{P_T}\\) to input feature embedding \\(\\mathbf{X_e}\\) acts like an input-dependent transformation" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.248, + 0.401, + 0.263 + ], + "angle": 0, + "content": "for spatio-temporal joints." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.264, + 0.389, + 0.475 + ], + "angle": 0, + "content": "(ii) As our prompts have same size as \\(\\mathbf{X}_{\\mathbf{e}}\\) it can also be thought of as a learned prompt encoding, bearing similarity with learnable position encoding works [12, 24, 26]. Our purpose is different however as prompt offsets seek to dynamically condition the input for adapting the backbone continually," + }, + { + "type": "image", + "bbox": [ + 0.39, + 0.25, + 0.788, + 0.384 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.384, + 0.391, + 0.788, + 0.467 + ], + "angle": 0, + "content": "Fig. 5: Here we visualize the order \\((s_i)_{i=1}^T\\) in which the \\(M = 64\\) prompts in the pool are selected at train time, across 4 user sessions \\(\\mathcal{U}\\mathcal{S}^{(t)}\\). X-axis: prompt index, Y-axis: index position in selected sequence. Top: The no sorting case uses the default sequence (hence diagonal matrices), giving equal importance to all prompts. Bottom (Our Method): Even though the same 64 prompts are selected and updated, the ordering is temporally unique and consistent with input." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.475, + 0.788, + 0.567 + ], + "angle": 0, + "content": "instead of learning positions. (iii) POET also bears similarity with auto-decoders like DeepSDF [30] which learn latent codes for each style or shape and use relevant codes along with a frozen decoder at inference. (iv) Prompt tuning can also be thought of as a parameter isolation technique for continual learning [28,34,35,38]. POET's ordered prompt selection as seen in Fig. 5 learns to isolate the relevant sequence of prompts for each input action sequence." + }, + { + "type": "code_caption", + "bbox": [ + 0.216, + 0.577, + 0.755, + 0.593 + ], + "angle": 0, + "content": "Algorithm 1 POET at Train Time, \\( t > 0 \\) (Case 1 \\( M = T \\), No pool expansion)" + }, + { + "type": "algorithm", + "bbox": [ + 0.216, + 0.594, + 0.789, + 0.733 + ], + "angle": 0, + "content": "Input: Query function \\(f_{q}\\) keys \\(\\pmb {K} = \\{\\pmb {k}_j\\}_{j = 1}^T\\) prompt pool \\(\\mathbf{P} = \\{\\pmb {P}_j\\}_{j = 1}^T\\) main model \\(f_{e},f_{g},f_{c}\\) Initialize: \\(\\mathbf{P},\\mathbf{K}\\) from \\(t - 1\\) ; Expand \\(f_{c}\\) by \\(N\\) new classes. Initialize \\(f_{c}\\) as: (i) copy \\(f_{c}^{old}\\) weights, (ii) \\(f_{c}^{new}\\gets Mean(f_{c}^{old})\\) Freeze: query layers \\(f_{g}^{\\prime},f_{e}^{\\prime}\\) ; main model layers \\(f_{e},f_{g}\\) for epochs and batch \\((\\mathbf{X}_i^t,\\mathbf{y}_i^t)_{i = 1}^{NK}\\) do 1. Get query feature \\(\\pmb{q}\\) (Eq.4); Compute \\(\\gamma (\\cdot)\\) b/w query \\(\\pmb{q}\\) and keys \\(\\pmb{K}\\) 2.Sort \\(\\gamma (\\cdot)\\) ; Get ordered key index sequence \\((s_i)_i = 1\\) Eq.7 3. Read pool memory \\(\\mathbf{P}\\) in order \\((s_i)^T_{i = 1}\\rightarrow\\) Get prompt offsets \\(\\mathbf{P_T}\\) 4.Get \\(\\mathbf{X_e};\\) Add \\(\\mathbf{P_T}\\) to it (Eq.8); get prediction y from prompted input (Eq. 1) 5.Use cross entropy loss (Equation 5) to update \\(f_{QA},K,\\mathbf{P},f_{c}\\) 6.Use clustering loss (Equation 6) to update \\(f_{QA}\\) and \\(\\pmb{K}\\) end // See \\(t = 0\\) training protocol in Algorithm 2 in Appendix" + }, + { + "type": "title", + "bbox": [ + 0.216, + 0.752, + 0.498, + 0.769 + ], + "angle": 0, + "content": "5 Experiments and Results" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.773, + 0.788, + 0.806 + ], + "angle": 0, + "content": "Datasets. We evaluated our method on well-known action recognition datasets: (i) activity recognition on the NTU RGB+D dataset [39]; and (ii) hand gesture" + }, + { + "type": "page_footnote", + "bbox": [ + 0.218, + 0.811, + 0.787, + 0.84 + ], + "angle": 0, + "content": "8 The datasets used in this work were accessed and processed at and by CMU. They were not accessed, processed, stored, or maintained at Meta." + } + ], + [ + { + "type": "page_number", + "bbox": [ + 0.218, + 0.116, + 0.236, + 0.127 + ], + "angle": 0, + "content": "10" + }, + { + "type": "header", + "bbox": [ + 0.272, + 0.115, + 0.364, + 0.129 + ], + "angle": 0, + "content": "P. Garg et al." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.147, + 0.785, + 0.193 + ], + "angle": 0, + "content": "recognition on the SHREC-2017 dataset [40]. As we introduce a new problem setting in human action recognition, we contribute two new benchmarks to the community for this setting, on the NTU RGB+D and SHREC-2017 datasets." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.196, + 0.788, + 0.318 + ], + "angle": 0, + "content": "For the NTU RGB+D dataset, we divide the 60 daily action categories into 40 base classes, learning the remaining 20 classes in subsequent user sessions. In few-shot learning parlance, our protocol is 4-task 5-way 5-shot, i.e. 5 novel classes using 5 user training instances in 4 user sessions. Each input 3D skeleton sequence has 64 temporal frames, each consisting of 25 body keypoints, such that \\( x \\in \\mathcal{R}^{64 \\times 25 \\times 3} \\). We use the spatio-temporal GCN, CTR-GCN [8], as the architecture for NTU RGB+D, where we choose the joint input modality for better interpretability of prompt tuning." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.32, + 0.788, + 0.442 + ], + "angle": 0, + "content": "For SHREC-2017, we divide the 14 fine-grained hand gesture classes into 8 base classes and 6 classes learned in subsequent user sessions. This is done in a 3-task 2-way 5-shot protocol, i.e. 2 novel classes using 5 user training instances in 3 user sessions. For each input instance of SHREC-2017, we use 8 temporal frames each having 22 hand keypoints, such that input \\( x \\in \\mathcal{R}^{8 \\times 22 \\times 3} \\). We use a fully-connected graph transformer backbone, DG-STA [7] for SHREC-2017. We select DG-STA due to easily reproducible code and to validate if our method POET works equally well across graph convolutional networks and graph transformers." + }, + { + "type": "text", + "bbox": [ + 0.215, + 0.444, + 0.788, + 0.567 + ], + "angle": 0, + "content": "Evaluation Metrics. Following earlier work in similar settings [31], we report: (i) Average accuracy 'Avg' of all classes seen so far, and (ii) Harmonic Mean \\( A_{HM} \\) between 'accuracy only on Old classes' and 'accuracy only on New classes' after learning each new user session. Note that the average accuracy tends to be biased towards the base session \\( \\mathcal{T}^{(0)} \\) performance due to more number of base classes. A higher \\( A_{HM} \\) implies better stability-plasticity trade-off between new task performance and old tasks' retention. Unlike many earlier CIL efforts, we report accuracy for both Old and New classes in each user session for transparency." + }, + { + "type": "text", + "bbox": [ + 0.215, + 0.569, + 0.789, + 0.842 + ], + "angle": 0, + "content": "Implementation Details. We observe that a key source of forgetting in our setting is from the classifier as the logits tend to become heavily biased towards the few-shot samples of new classes. We use a cosine classifier for activity recognition experiments on CTR-GCN. For gesture recognition on the lightweight DG-STA, we use a standard fully-connected layer as classifier, but freeze old class parameters in the classifier by zeroing their gradients. We attach prompts after the 1st layer of DG-STA and 1st CTR-GC block of CTR-GCN. For both datasets, we have equal or higher learning rates in user sessions when compared to the base model's training in order to accommodate new knowledge in the model (for better plasticity). For exact implementation details (including learning rates, epochs, hyperparameter analysis, and backward forgetting metric), see Appendix A. In earlier efforts that more generally tune prompts for class-incremental learning [41, 45, 47-49], it is common to rely on an ImageNet21K pretrained ViT [37] or CLIP [33] as the backbone. However, such backbones do not exist for skeleton-based human action recognition. Our base feature extractor is hence trained on the base session dataset itself without any pretraining, making this one of the first efforts of prompt tuning without extensive pretraining (scale of data 3-5 times lower order of magnitude)." + } + ], + [ + { + "type": "header", + "bbox": [ + 0.255, + 0.114, + 0.732, + 0.129 + ], + "angle": 0, + "content": "POET: Prompt Offset Tuning for Continual Human Action Adaptation" + }, + { + "type": "page_number", + "bbox": [ + 0.769, + 0.116, + 0.784, + 0.127 + ], + "angle": 0, + "content": "11" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.147, + 0.788, + 0.375 + ], + "angle": 0, + "content": "Results. Since there are no existing baselines for our proposed setting in skeletal action recognition, we compare our method by adapting continual learning (CL) baselines to skeletal data in Sec 5.1, Tables 1, 2. We first compare POET with prompt tuning based class-incremental learning (CIL) approaches originally designed for images (L2P [49], CODA-P [41], APT [4]) and find that it has very low performance on new classes as they do not update their query function. We find any fine-tuning or knowledge distillation based approaches (LWF [25], EWC [17], LUCIR [15]) lead to rapid forgetting of base knowledge as the model overfits to user's few-shots. We also compare with multiple variants of Feature Extraction (FE) to check if prompts truly have merit (POET=FE+Prompts) and provide upper bound baselines. In Sec 6, we first show the importance of prompts in POET by removing the prompts. We discuss the value of our coupled optimization, query function update and ordered key index selection in our prompt selection ablation Tab 3. We also study the impact of proposed additive prompt tuning as compared to other possible prompt attachments \\( f_{p} \\) in Tab 4." + }, + { + "type": "title", + "bbox": [ + 0.215, + 0.394, + 0.543, + 0.409 + ], + "angle": 0, + "content": "5.1 Comparison with State-of-the-Art" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.418, + 0.788, + 0.644 + ], + "angle": 0, + "content": "POET sets the SOTA on existing prompt tuning works (Tab 1,2). We adapt three standard CIL works that prompt tune ViTs for images - L2P [49], CODA-P [41] and APT [4] to our setting. L2P and CODA-P share prompt pool across tasks (similar to us), whereas APT learns task-specific prompts. L2P decouples the optimization of keys from the prompt pool and concatenates the selected prompts. Since concatenation is not defined for our GNN backbone, we adapt these SOTA to our setting by concatenating along the temporal dimension \\((\\mathbf{L2P}^*,\\mathbf{CODA - P}^*)\\). CODA-P [41] couples keys with the prompt pool by using a cosine similarity weighing over all prompts in the pool, forming a 'soft prompt selection', different from our 'ordered hard prompt selection'. In APT, we train prompt-classifier pairs for each continual task separately (\\(\\hat{\\mathbf{\\Gamma}}\\) denotes task-specific), and use task identity at test time. See details in Appendix A. These methods by design rely on extensively pretrained (ImageNet21k) query functions which does not require updates; and require full supervision on new classes, perhaps explaining their poor 'New' accuracy in our few-shot setting." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.645, + 0.788, + 0.841 + ], + "angle": 0, + "content": "Standard Continual learning Baselines. We compared with two well established knowledge-distillation approaches, learning without forgetting (LWF) and LUCIR. Both of them perform poorly on both old and new classes. EWC [17] learns better on new but does not retain old knowledge. We conclude that any CL method that fine-tunes the backbone feature representation in subsequent sessions \\( t > 0 \\) will not be able to retain base/old class knowledge (a finding consistent with existing FSCIL literature for images [10,43]). We also adapt and compare with one of the latest FSCIL baselines ALICE [31], originally developed for image classification benchmarks on our gesture recognition benchmark in Table 2. Note the high retention of base task performance (due to non-parametric classifier on top of frozen base model). However, it suffers from poor plasticity and adaptation to new classes. This is the issue of feature-classifier misalignment that we hoped to alleviate through prompt tuning." + } + ], + [ + { + "type": "page_number", + "bbox": [ + 0.218, + 0.116, + 0.236, + 0.127 + ], + "angle": 0, + "content": "12" + }, + { + "type": "header", + "bbox": [ + 0.272, + 0.115, + 0.364, + 0.129 + ], + "angle": 0, + "content": "P. Garg et al." + }, + { + "type": "table_caption", + "bbox": [ + 0.216, + 0.146, + 0.788, + 0.21 + ], + "angle": 0, + "content": "Table 1: Activity Recognition Results \\((\\% ,\\uparrow)\\), Comparison with SOTA: NTU RGB+D [39] dataset on CTR-GCN [8] backbone. After training on each incremental task, we report Average of all classes seen so far ('Avg'). We also report (i) \\(A_{HM}\\), (ii) old classes accuracy ('Old'), (iii) new classes accuracy ('New') in the last session. We report Mean and STD across 10 sets of 5-shots. POET achieves the best stability-plasticity trade-off across all baselines indicated by the \\(A_{HM} = 56.3\\%\\). POET also has the highest Avg across all user sessions outside of upper bound baselines in orange." + }, + { + "type": "table", + "bbox": [ + 0.222, + 0.213, + 0.774, + 0.467 + ], + "angle": 0, + "content": "
Method\\( \\mathbf{U}\\mathbf{B}^{(0)} \\)\\( \\mathbf{U}\\mathbf{S}^{(1)} \\)\\( \\mathbf{U}\\mathbf{S}^{(2)} \\)\\( \\mathbf{U}\\mathbf{S}^{(3)} \\)\\( \\mathbf{U}\\mathbf{S}^{(4)} \\)
Base (↑)Avg (↑)Avg (↑)Avg (↑)Old (↑)New (↑)Avg (↑)\\( A_{HM} \\) (↑)
Upper Bounds
Joint (Oracle)88.479.071.066.863.5
Joint POET (Oracle)67.2
FE, Task-Specific^88.470.1 ± 2.652.5 ± 5.844.8 ± 5.070.3 ± 2.146.7 ± 2.0NANA
FE+Replay88.482.4 ± 1.178.2 ± 1.274.5 ± 1.273.1 ± 1.043.3 ± 3.370.6 ± 1.254.3 ± 2.6
Continual Linear Probing
FE88.472.0 ± 1.160.4 ± 2.447.7 ± 2.140.0 ± 1.651.0 ± 2.340.9 ± 1.444.8 ± 1.1
FE, Frozen88.476.1 ± 1.052.4 ± 4.138.3 ± 2.728.4 ± 1.622.4 ± 4.527.9 ± 1.424.8 ± 3.0
FE+Replay†88.472.0 ± 1.559.5 ± 4.058.7 ± 2.856.7 ± 2.534.7 ± 5.654.9 ± 2.742.8 ± 4.4
FT88.46.2 ± 1.44.3 ± 1.52.8 ± 1.00.2 ± 0.536.0 ± 10.13.2 ± 0.80.3 ± 1.0
Standard Continual Learning
LWF [25]88.46.2 ± 1.52.8 ± 0.73.7 ± 1.30.0 ± 0.038.9 ± 8.83.2 ± 0.70.0 ± 0.0
EWC [17]88.46.6 ± 1.54.1 ± 1.43.1 ± 0.90.0 ± 0.042.1 ± 9.53.5 ± 0.80.0 ± 0.0
Experience Replay88.435.1 ± 8.350.6 ± 5.060.6 ± 5.454.6 ± 6.543.7 ± 14.653.7 ± 7.147.8 ± 11.2
Experience \\( \\text{Replay}^{\\dagger} \\)88.46.2 ± 1.59.0 ± 2.611.2 ± 3.010.9 ± 2.634.6 ± 7.912.9 ± 3.016.3 ± 3.5
LUCIR [15]87.94.3 ± 2.14.1 ± 1.32.7 ± 0.80.2 ± 0.426.0 ± 9.22.3 ± 0.90.4 ± 0.8
Continual Prompt Tuning
CODA-P [41]*87.476.1 ± 1.066.7 ± 1.358.6 ± 2.756.5 ± 2.90.5 ± 0.451.8 ± 2.71.1 ± 0.7
L2P [49]*88.678.9 ± 0.171.0 ± 1.064.2 ± 0.162.0 ± 0.70.0 ± 0.056.8 ± 0.60.0 ± 0.0
APT [4]^86.627.3 ± 1.630.8 ± 3.437.6 ± 2.3NA33.4 ± 2.0NANA
POET (Ours)87.982.3 ± 0.676.8 ± 0.968.4 ± 0.757.2 ± 1.055.8 ± 5.957.1 ± 1.156.3 ± 3.2
" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.488, + 0.788, + 0.67 + ], + "angle": 0, + "content": "Fine-tuning (FE) and Feature Extraction (FE) Baselines. We implement standard continual learning baselines to understand stability-plasticity trade-offs in our new benchmarks. In all these baselines, we expand the classifier output dimension by \\(N\\) new classes. In 'FT (Fine-Tuning)', we tune all model parameters on cross entropy loss of new task. FSCIL is challenging for this modality as old task performance sharply reduces to zero starting from \\(\\mathcal{U}\\mathcal{S}^{(1)}\\) as model overfits to user's few-shots. 'FE (Feature Extraction)' differs from FT as we freeze the feature extractor to preserve base knowledge. This serves as a competitive baseline in our findings. In 'FE, frozen', we zero out the gradients of previous class weights in classifier \\(f_{c}\\) to prevent forgetting from the classifier. 'FE' and 'FE, Frozen' exhibit different New-Old trade-offs in Tables 1, 2 because the scale of pretraining is different (gesture more lightweight than activity)." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.672, + 0.788, + 0.809 + ], + "angle": 0, + "content": "Upper-bound baselines, top section Tables 1, 2. In 'Joint (oracle)' experiment, we train on all task data at the same time in a multi-task (non-sequential) manner. Training POET in a multi-task manner ('Joint POET') outperforms 'Joint Oracle' demonstrating the strength of our approach. In addition to these generalist upper bounds, we point out that 'FE, Task-specific' is a competitive specialist upper bound. In this, we perform feature extraction from base model to each task individually, storing separate task-specific models \\((\\mathcal{U}\\mathcal{S}^{(0)}\\to \\mathcal{U}\\mathcal{S}^{(i)},i > 0)\\). POET outperforms 'New' accuracy compared with this baseline, achieving a forward transfer on each \\(t > 0\\). This indicates that prompt" + }, + { + "type": "page_footnote", + "bbox": [ + 0.218, + 0.825, + 0.733, + 0.84 + ], + "angle": 0, + "content": "9 'FE' is the same as 'w/o prompts' in Table 3. We highlight key baselines in gray color." + } + ], + [ + { + "type": "header", + "bbox": [ + 0.255, + 0.115, + 0.732, + 0.13 + ], + "angle": 0, + "content": "POET: Prompt Offset Tuning for Continual Human Action Adaptation" + }, + { + "type": "page_number", + "bbox": [ + 0.769, + 0.116, + 0.786, + 0.127 + ], + "angle": 0, + "content": "13" + }, + { + "type": "table_caption", + "bbox": [ + 0.216, + 0.146, + 0.788, + 0.179 + ], + "angle": 0, + "content": "Table 2: Gesture Recognition Results (\\%, ↑), Comparison with SOTA: SHREC 2017 [40] dataset on DG-STA [7] graph transformer backbone. Reporting mean and standard deviation across 5 runs. POET achieves best \\( A_{HM} = 56.2\\% \\)." + }, + { + "type": "table", + "bbox": [ + 0.255, + 0.184, + 0.744, + 0.319 + ], + "angle": 0, + "content": "
Method\\( \\mathbf{\\mathcal{U}\\mathbf{B}^{(0)}} \\)\\( \\mathbf{\\mathcal{U}\\mathbf{S}^{(1)}} \\)\\( \\mathbf{\\mathcal{U}\\mathbf{S}^{(2)}} \\)\\( \\mathbf{\\mathcal{U}\\mathbf{S}^{(3)}} \\)
Base (↑)Avg (↑)Avg (↑)Old (↑)New (↑)Avg (↑)\\( A_{HM} \\) (↑)
Joint (Oracle)88.879.4 ± 0.777.3 ± 2.170.9 ± 1.262.4 ± 0.4
FT88.820.3 ± 0.812.4 ± 2.10.0 ± 0.085.8 ± 9.413.4 ± 1.50.0 ± 0.0
FE88.862.7 ± 2.441.9 ± 6.917.5 ± 5.177.3 ± 8.826.8 ± 3.428.5 ± 6.4
FE, Frozen88.871.3 ± 1.961.4 ± 2.744.7 ± 3.254.5 ± 6.746.2 ± 2.749.1 ± 4.3
LWF [25]88.820.2 ± 1.412.5 ± 1.00.0 ± 0.088.4 ± 13.713.8 ± 2.10.0 ± 0.0
L2P [49]**88.820.3 ± 5.910.5 ± 4.88.2 ± 4.06.9 ± 8.57.9 ± 3.97.5 ± 5.5
CODA-P [41]**87.715.6 ± 4.511.6 ± 1.97.9 ± 1.814.1 ± 21.48.8 ± 2.410.1 ± 3.2
ALICE [31]92.172.4 ± 5.763.3 ± 7.662.5 ± 6.811.9 ± 9.954.6 ± 6.920.0 ± 8.1
POET (Ours)91.973.2 ± 3.761.9 ± 1.845.9 ± 2.672.4 ± 7.150.0 ± 1.656.2 ± 1.6
" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.336, + 0.784, + 0.365 + ], + "angle": 0, + "content": "tuning benefits New performance due to the pre-existing knowledge in the shared knowledge pool. Avg in sessions \\(0 < t < 4\\) indicates New for task-specific\\(^{\\wedge}\\)." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.367, + 0.785, + 0.426 + ], + "angle": 0, + "content": "Experience Replay Baselines, Tab 1. Even though our privacy-aware setting prohibits previous data replay, we compare with 'Experience Replay' (store and replay 5-samples of base and incremental sessions) and 'Experience Replay†' (replay only previous incremental sessions) for completeness." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.427, + 0.785, + 0.502 + ], + "angle": 0, + "content": "'FE+Replay' serves as the best upper bound (even better than Experience Replay as we are freezing backbone in addition to replay). It is noteworthy that POET (which is FE+prompts) learns an implicit 'data-free' form of prompt pool memory, and yet has a better \\( A_{HM} \\) trade-off as compared to explicitly stored and replayed samples from previous classes in FE+replay." + }, + { + "type": "title", + "bbox": [ + 0.216, + 0.511, + 0.546, + 0.529 + ], + "angle": 0, + "content": "6 Ablation Studies and Analysis" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.536, + 0.788, + 0.625 + ], + "angle": 0, + "content": "Importance of prompts in POET. First, we consider the contribution of prompt offsets in POET. Since we only attach prompts to address continual learning in POET, removing prompts gives the Feature Extraction (FE) baseline ('w/o prompts', Table 3) where the backbone is frozen after base training and only the classifier is expanded and updated on classification loss of new classes. POET improves both, 'Old' (\\( \\uparrow 20.1\\% \\)) and 'New' (\\( \\uparrow 10.6\\% \\)) marked in blue." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.627, + 0.461, + 0.776 + ], + "angle": 0, + "content": "Prompt Selection Mechanism. In Table 3, we investigate our prompt selection mechanism and optimization choices. The 'w/o coupled optim.' experiment is a direct comparison of our additive prompt attachment with the de-coupled optimization in L2P [49]. Updating key parameters but keeping only query adaptor" + }, + { + "type": "text", + "bbox": [ + 0.468, + 0.627, + 0.788, + 0.688 + ], + "angle": 0, + "content": "Table 3: Prompt Selection Mechanism Analysis on NTU RGB+D dataset \\((\\% ,\\uparrow)\\) .. \\(\\mathrm{w / o}\\) denotes removing that component from POET, numbers in brackets are wrt POET \\((M = T)\\) experiment. 'Avg' accuracy is biased towards Old' classes accuracy, \\(A_{HM}\\) is good indicator of trade-off between New' and Old'." + }, + { + "type": "table", + "bbox": [ + 0.47, + 0.689, + 0.794, + 0.774 + ], + "angle": 0, + "content": "
NTU RGB+D\\( \\left| \\mathcal{U}\\mathcal{B}^{(0)}\\right| \\)\\( \\left| \\mathcal{U}\\mathcal{S}^{(1)}\\right| \\)\\( \\left| \\mathcal{U}\\mathcal{S}^{(2)}\\right| \\)\\( \\left| \\mathcal{U}\\mathcal{S}^{(3)}\\right| \\)\\( \\left| \\mathcal{U}\\mathcal{S}^{(4)}\\right| \\)
MethodBaseAvgAvgAvgOldNewAvg\\( A_{HM} \\)
w/o prompts88.474.566.349.539.2 (-20.1)46.8 (-10.6)39.942.7
w/o coupled optim.88.082.875.365.856.5 (-2.8)51.3 (-6.1)56.153.8
w/o clustering loss85.581.674.364.562.0 (+2.7)18.2 (-39.2)57.028.1
w/o QA update87.982.877.469.159.4 (+0.1)52.8 (-4.6)58.755.9
w/o sorting88.282.275.268.859.9 (+0.6)46.6 (-10.8)58.852.4
POET (M>T)87.982.777.268.860.3 (+1.0)54.4 (-3.0)59.857.2
POET (M=T)87.982.876.868.659.357.459.258.3
" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.777, + 0.788, + 0.841 + ], + "angle": 0, + "content": "\\(QA\\) frozen after \\(\\mathcal{U}\\mathcal{B}^{(0)}\\) training ('w/o QA update') reduces 'New' only performance of \\(\\mathcal{US}^{(4)}\\) by \\(4.6\\%\\) as the query function stays fixed at base session learning and is not discriminative towards new classes. 'W/o clustering loss' from Eq. 6, performance drops starting from \\(\\mathcal{UB}^{(0)}\\) itself. The only difference" + } + ], + [ + { + "type": "page_number", + "bbox": [ + 0.218, + 0.116, + 0.236, + 0.127 + ], + "angle": 0, + "content": "14" + }, + { + "type": "header", + "bbox": [ + 0.272, + 0.115, + 0.364, + 0.129 + ], + "angle": 0, + "content": "P. Garg et al." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.147, + 0.788, + 0.268 + ], + "angle": 0, + "content": "between the experiment 'w/o sorting' and 'POET (M=T)' is that we do not sort the cosine similarity before selecting top \\(T\\) indices (same as Fig 5). The \\(10.8\\% \\uparrow\\) in 'New' performance validates that our prompt selection mechanism is learning to choose a distinct temporal ordering for prompt tuning of new input samples. With pool expansion ('POET, \\(M > T\\)), we get more flexibility in the stability-plasticity trade-offs depending on how many new prompts we attach. For \\(R = 6\\), 'Old' is improved. In Table 3, we keep POET's additive prompt attachment and only vary prompt selection." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.269, + 0.788, + 0.375 + ], + "angle": 0, + "content": "Prompt Attachment Mechanism. In Table 4, we keep our end-to-end optimization and ordered prompt selection as a constant and ablate prompt shape and attachment operator \\( f_{p}(.) \\). Drawing a parallel with transformers which concatenate prompts along the token dimension, we conduct experiments concatenating prompts along the (i) temporal dimension of the skeleton input feature embedding \\( \\mathbf{X}_{\\mathrm{e}} \\) ('CONCAT temporal') and (ii) feature dimension \\( C_e \\) ('CONCAT feature'). We find that addition works better than concatenation and" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.376, + 0.489, + 0.497 + ], + "angle": 0, + "content": "cross attention. We also verify our hypothesis that selecting the same number of prompts as the input temporal dimension (\\(T = 64\\) for NTU RGB+D and \\(T = 8\\) for SHREC-2017) yields better results as compared to adding the same prompt frame to each input embedding frame ('Addition \\(T' = 1'\\))." + }, + { + "type": "table_caption", + "bbox": [ + 0.499, + 0.378, + 0.788, + 0.412 + ], + "angle": 0, + "content": "Table 4: Prompt Attachment Analysis (%, ↑): The best prompt attachment choice \\(f_{p}(.)\\) is Adding #prompts same as #input frames (T=64)." + }, + { + "type": "table", + "bbox": [ + 0.5, + 0.413, + 0.785, + 0.476 + ], + "angle": 0, + "content": "
NTU RGB+D\\( \\left\\| u\\mathcal{B}^{(0)}\\right\\| \\)\\( \\left\\| u\\mathcal{S}^{(1)}\\right\\| \\)\\( \\left\\| u\\mathcal{S}^{(2)}\\right\\| \\)\\( \\left\\| u\\mathcal{S}^{(3)}\\right\\| \\)\\( \\left\\| u\\mathcal{S}^{(4)}\\right\\| \\)
MethodBaseAvgAvgAvgOld New Avg \\( A_{HM} \\)
CONCAT temporal, \\( T' = 64 \\)88.670.362.449.833.6 50.5 35.1 40.3
CONCAT feature, \\( T' = 64 \\)87.782.475.566.957.1 41.5 56.0 48.1
Cross Attention, \\( T' = 64 \\)82.977.472.265.057.1 32.3 55.0 41.2
ADD, \\( T' = 1 \\)88.773.362.745.533.7 47.0 34.8 39.3
ADD, \\( T' = 64 \\) (Ours)87.982.876.868.659.3 57.4 59.2 58.3
" + }, + { + "type": "title", + "bbox": [ + 0.216, + 0.512, + 0.542, + 0.529 + ], + "angle": 0, + "content": "7 Conclusions and Future Work" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.538, + 0.789, + 0.841 + ], + "angle": 0, + "content": "The problem of continually adapting human action models to new user categories over time has gained prominence with the rising availability of XR devices. However, this setting poses unique challenges: (i) the user may be able to provide only a few samples for training, and (ii) accessing data from earlier sessions may violate privacy considerations. We hence propose a method based on prompt offset tuning to address this problem in this work. Prompt tuning to address learning over newer tasks has been attempted in recent years. However, these works have: (1) typically been designed for image-based tasks, (2) relied on strongly pretrained transformer backbones, (3) required full supervision for new tasks, and (4) exclusively applied prompt tuning to transformer architectures. This work departs from these four characteristics. Our work demonstrates that prompt offset tuning is a promising option to evolve and adapt skeleton-based human action models to new user classes. The careful design of each component of the proposed methodology finds validation in the promising results across well-known skeleton-based action recognition benchmarks. Our ablation studies and analysis corroborate our design choices in our implementation. Looking ahead, it will be interesting to explore how our approach and its design choices adapt when a \"generalist backbone\" trained on a large corpus of action recognition data becomes accessible. Extending our method for differential privacy is another interesting direction of future work." + } + ], + [ + { + "type": "header", + "bbox": [ + 0.255, + 0.115, + 0.732, + 0.129 + ], + "angle": 0, + "content": "POET: Prompt Offset Tuning for Continual Human Action Adaptation" + }, + { + "type": "page_number", + "bbox": [ + 0.769, + 0.117, + 0.786, + 0.127 + ], + "angle": 0, + "content": "15" + }, + { + "type": "title", + "bbox": [ + 0.218, + 0.145, + 0.323, + 0.16 + ], + "angle": 0, + "content": "References" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.177, + 0.786, + 0.232 + ], + "angle": 0, + "content": "1. Aich, S., Ruiz-Santaquiteria, J., Lu, Z., Garg, P., Joseph, K.J., Fernandez, A., Balasubramanian, V.N., Kin, K., Wan, C., Camgoz, N.C., Ma, S., De la Torre, F.: Data-free class-incremental hand gesture recognition. In: Proceedings of the IEEE/CVF International Conference on Computer Vision (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.233, + 0.786, + 0.259 + ], + "angle": 0, + "content": "2. Albrecht, J.P.: How the gdpr will change the world. Eur. Data Prot. L. Rev. 2, 287 (2016)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.261, + 0.786, + 0.3 + ], + "angle": 0, + "content": "3. Bengio, Y., Léonard, N., Courville, A.: Estimating or propagating gradients through stochastic neurons for conditional computation. arXiv preprint arXiv:1308.3432 (2013)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.302, + 0.786, + 0.357 + ], + "angle": 0, + "content": "4. Bowman, B., Achille, A., Zancato, L., Trager, M., Perera, P., Paolini, G., Soatto, S.: a-la-carte prompt tuning (apt): Combining distinct data via composable prompting. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 14984-14993 (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.358, + 0.786, + 0.398 + ], + "angle": 0, + "content": "5. Chaudhry, A., Dokania, P.K., Ajanthan, T., Torr, P.H.: Riemannian walk for incremental learning: Understanding forgetting and intransigence. In: Proceedings of the European conference on computer vision (ECCV). pp. 532-547 (2018)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.399, + 0.786, + 0.425 + ], + "angle": 0, + "content": "6. Chaudhry, A., Ranzato, M., Rohrbach, M., Elhoseiny, M.: Efficient lifelong learning with a-gem. arXiv preprint arXiv:1812.00420 (2018)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.427, + 0.786, + 0.467 + ], + "angle": 0, + "content": "7. Chen, Y., Zhao, L., Peng, X., Yuan, J., Metaxas, D.N.: Construct dynamic graphs for hand gesture recognition via spatial-temporal attention. In: Proceedings of the British Machine Vision Conference (BMVC) (2019)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.468, + 0.786, + 0.522 + ], + "angle": 0, + "content": "8. Chen, Y., Zhang, Z., Yuan, C., Li, B., Deng, Y., Hu, W.: Channel-wise topology refinement graph convolution for skeleton-based action recognition. In: Proceedings of the IEEE/CVF International Conference on Computer Vision. pp. 13359-13368 (2021)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.524, + 0.786, + 0.55 + ], + "angle": 0, + "content": "9. Dhariwal, P., Jun, H., Payne, C., Kim, J.W., Radford, A., Sutskever, I.: Jukebox: A generative model for music. arXiv preprint arXiv:2005.00341 (2020)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.551, + 0.786, + 0.592 + ], + "angle": 0, + "content": "0. Dong, S., Hong, X., Tao, X., Chang, X., Wei, X., Gong, Y.: Few-shot class incremental learning via relation knowledge distillation. In: Proceedings of the AAAI Conference on Artificial Intelligence. vol. 35, pp. 1255-1263 (2021)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.593, + 0.786, + 0.619 + ], + "angle": 0, + "content": "1. Dwivedi, V.P., Bresson, X.: A generalization of transformer networks to graphs. arXiv preprint arXiv:2012.09699 (2020)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.62, + 0.786, + 0.674 + ], + "angle": 0, + "content": "2. Dwivedi, V.P., Luu, A.T., Laurent, T., Bengio, Y., Bresson, X.: Graph neural networks with learnable structural and positional representations. In: International Conference on Learning Representations (2022), https://openreview.net/forum?id=wTTjnvGphYj" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.676, + 0.786, + 0.716 + ], + "angle": 0, + "content": "3. Hersche, M., Karunaratne, G., Cherubini, G., Benini, L., Sebastian, A., Rahimi, A.: Constrained few-shot class-incremental learning. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 9057-9067 (2022)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.717, + 0.786, + 0.757 + ], + "angle": 0, + "content": "4. Hinojosa, C., Marquez, M., Arguello, H., Adeli, E., Fei-Fei, L., Niebles, J.C.: Privhar: Recognizing human actions from privacy-preserving lens. In: European Conference on Computer Vision. pp. 314-332. Springer (2022)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.758, + 0.786, + 0.799 + ], + "angle": 0, + "content": "5. Hou, S., Pan, X., Loy, C.C., Wang, Z., Lin, D.: Learning a unified classifier incrementally via rebalancing. In: Proceedings of the IEEE/CVF conference on computer vision and pattern recognition. pp. 831-839 (2019)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.8, + 0.786, + 0.84 + ], + "angle": 0, + "content": "6. Jia, M., Tang, L., Chen, B.C., Cardie, C., Belongie, S., Hariharan, B., Lim, S.N.: Visual prompt tuning. In: European Conference on Computer Vision. pp. 709-727. Springer (2022)" + }, + { + "type": "list", + "bbox": [ + 0.226, + 0.177, + 0.786, + 0.84 + ], + "angle": 0, + "content": null + } + ], + [ + { + "type": "page_number", + "bbox": [ + 0.218, + 0.116, + 0.236, + 0.127 + ], + "angle": 0, + "content": "16" + }, + { + "type": "header", + "bbox": [ + 0.272, + 0.115, + 0.364, + 0.129 + ], + "angle": 0, + "content": "P. Garg et al." + }, + { + "type": "ref_text", + "bbox": [ + 0.218, + 0.148, + 0.787, + 0.204 + ], + "angle": 0, + "content": "17. Kirkpatrick, J., Pascanu, R., Rabinowitz, N., Veness, J., Desjardins, G., Rusu, A.A., Milan, K., Quan, J., Ramalho, T., Grabska-Barwinska, A., et al.: Overcoming catastrophic forgetting in neural networks. Proceedings of the national academy of sciences 114(13), 3521-3526 (2017)" + }, + { + "type": "ref_text", + "bbox": [ + 0.218, + 0.205, + 0.787, + 0.245 + ], + "angle": 0, + "content": "18. Kumawat, S., Nagahara, H.: Privacy-preserving action recognition via motion difference quantization. In: European Conference on Computer Vision. pp. 518-534. Springer (2022)" + }, + { + "type": "ref_text", + "bbox": [ + 0.218, + 0.246, + 0.787, + 0.273 + ], + "angle": 0, + "content": "19. Lacoste, A., Luccioni, A., Schmidt, V., Dandes, T.: Quantifying the carbon emissions of machine learning. arXiv preprint arXiv:1910.09700 (2019)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.273, + 0.787, + 0.3 + ], + "angle": 0, + "content": "20. Lester, B., Al-Rfou, R., Constant, N.: The power of scale for parameter-efficient prompt tuning. arXiv preprint arXiv:2104.08691 (2021)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.3, + 0.787, + 0.356 + ], + "angle": 0, + "content": "21. Li, M., Xu, X., Fan, H., Zhou, P., Liu, J., Liu, J.W., Li, J., Keppo, J., Shou, M.Z., Yan, S.: Stprivacy: Spatio-temporal privacy-preserving action recognition. In: Proceedings of the IEEE/CVF International Conference on Computer Vision. pp. 5106-5115 (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.356, + 0.787, + 0.398 + ], + "angle": 0, + "content": "22. Li, T., Ke, Q., Rahmani, H., Ho, R.E., Ding, H., Liu, J.: Else-net: Elastic semantic network for continual action recognition from skeleton data. In: Proceedings of the IEEE/CVF International Conference on Computer Vision. pp. 13434-13443 (2021)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.398, + 0.787, + 0.424 + ], + "angle": 0, + "content": "23. Li, X.L., Liang, P.: Prefix-tuning: Optimizing continuous prompts for generation. arXiv preprint arXiv:2101.00190 (2021)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.425, + 0.787, + 0.466 + ], + "angle": 0, + "content": "24. Li, Y., Si, S., Li, G., Hsieh, C.J., Bengio, S.: Learnable fourier features for multidimensional spatial positional encoding. Advances in Neural Information Processing Systems 34, 15816-15829 (2021)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.466, + 0.787, + 0.494 + ], + "angle": 0, + "content": "25. Li, Z., Hoiem, D.: Learning without forgetting. IEEE transactions on pattern analysis and machine intelligence 40(12), 2935-2947 (2017)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.495, + 0.787, + 0.536 + ], + "angle": 0, + "content": "26. Liu, X., Yu, H.F., Dhillon, I., Hsieh, C.J.: Learning to encode position for transformer with continuous dynamical model. In: International conference on machine learning. pp. 6327-6335. PMLR (2020)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.536, + 0.787, + 0.578 + ], + "angle": 0, + "content": "27. Ma, N., Zhang, H., Li, X., Zhou, S., Zhang, Z., Wen, J., Li, H., Gu, J., Bu, J.: Learning spatial-preserved skeleton representations for few-shot action recognition. In: European Conference on Computer Vision. pp. 174-191. Springer (2022)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.578, + 0.787, + 0.619 + ], + "angle": 0, + "content": "28. Mallya, A., Lazebnik, S.: Packnet: Adding multiple tasks to a single network by iterative pruning. In: Proceedings of the IEEE conference on Computer Vision and Pattern Recognition. pp. 7765-7773 (2018)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.619, + 0.787, + 0.646 + ], + "angle": 0, + "content": "29. Mialon, G., Chen, D., Selosse, M., Mairal, J.: Graphit: Encoding graph structure in transformers. arXiv preprint arXiv:2106.05667 (2021)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.646, + 0.787, + 0.702 + ], + "angle": 0, + "content": "30. Park, J.J., Florence, P., Straub, J., Newcombe, R., Lovegrove, S.: Deepsdf: Learning continuous signed distance functions for shape representation. In: Proceedings of the IEEE/CVF conference on computer vision and pattern recognition. pp. 165-174 (2019)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.702, + 0.787, + 0.743 + ], + "angle": 0, + "content": "31. Peng, C., Zhao, K., Wang, T., Li, M., Lovell, B.C.: Few-shot class-incremental learning from an open-set perspective. In: European Conference on Computer Vision. pp. 382-397. Springer (2022)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.743, + 0.787, + 0.785 + ], + "angle": 0, + "content": "32. Pernici, F., Bruni, M., Baecchi, C., Turchini, F., Del Bimbo, A.: Class-incremental learning with pre-allocated fixed classifiers. In: 2020 25th International Conference on Pattern Recognition (ICPR). pp. 6259–6266. IEEE (2021)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.785, + 0.787, + 0.84 + ], + "angle": 0, + "content": "33. Radford, A., Kim, J.W., Hallacy, C., Ramesh, A., Goh, G., Agarwal, S., Sastry, G., Askell, A., Mishkin, P., Clark, J., et al.: Learning transferable visual models from natural language supervision. In: International conference on machine learning. pp. 8748-8763. PMLR (2021)" + }, + { + "type": "list", + "bbox": [ + 0.217, + 0.148, + 0.787, + 0.84 + ], + "angle": 0, + "content": null + } + ], + [ + { + "type": "header", + "bbox": [ + 0.255, + 0.115, + 0.732, + 0.129 + ], + "angle": 0, + "content": "POET: Prompt Offset Tuning for Continual Human Action Adaptation" + }, + { + "type": "page_number", + "bbox": [ + 0.769, + 0.117, + 0.786, + 0.127 + ], + "angle": 0, + "content": "17" + }, + { + "type": "ref_text", + "bbox": [ + 0.214, + 0.147, + 0.788, + 0.189 + ], + "angle": 0, + "content": "34. Razdaibiedina, A., Mao, Y., Hou, R., Khabsa, M., Lewis, M., Almahairi, A.: Progressive prompts: Continual learning for language models. arXiv preprint arXiv:2301.12314 (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.214, + 0.189, + 0.788, + 0.23 + ], + "angle": 0, + "content": "35. Rebuffi, S.A., Bilen, H., Vedaldi, A.: Efficient parametrization of multi-domain deep neural networks. In: Proceedings of the IEEE conference on computer vision and pattern recognition. pp. 8119-8127 (2018)" + }, + { + "type": "ref_text", + "bbox": [ + 0.214, + 0.23, + 0.788, + 0.257 + ], + "angle": 0, + "content": "36. Ren, B., Liu, M., Ding, R., Liu, H.: A survey on 3d skeleton-based action recognition using learning method. Cyborg and Bionic Systems (2020)" + }, + { + "type": "ref_text", + "bbox": [ + 0.214, + 0.257, + 0.788, + 0.283 + ], + "angle": 0, + "content": "37. Ridnik, T., Ben-Baruch, E., Noy, A., Zelnik-Manor, L.: Imagenet-21k pretraining for the masses. arXiv preprint arXiv:2104.10972 (2021)" + }, + { + "type": "ref_text", + "bbox": [ + 0.214, + 0.283, + 0.788, + 0.324 + ], + "angle": 0, + "content": "38. Rusu, A.A., Rabinowitz, N.C., Desjardins, G., Soyer, H., Kirkpatrick, J., Kavukcuoglu, K., Pascanu, R., Hadsell, R.: Progressive neural networks. arXiv preprint arXiv:1606.04671 (2016)" + }, + { + "type": "ref_text", + "bbox": [ + 0.214, + 0.324, + 0.788, + 0.365 + ], + "angle": 0, + "content": "39. Shahroudy, A., Liu, J., Ng, T.T., Wang, G.: Ntu rgb+ d: A large scale dataset for 3d human activity analysis. In: Proceedings of the IEEE conference on computer vision and pattern recognition. pp. 1010-1019 (2016)" + }, + { + "type": "ref_text", + "bbox": [ + 0.214, + 0.365, + 0.788, + 0.419 + ], + "angle": 0, + "content": "40. Smedt, Q.D., Wannous, H., Vandeborre, J.P., Guerry, J., Saux, B.L., Filliat, D.: 3D Hand Gesture Recognition Using a Depth and Skeletal Dataset. In: Pratikakis, I., Dupont, F., Ovsjanikov, M. (eds.) Eurographics Workshop on 3D Object Retrieval. The Eurographics Association (2017). https://doi.org/10.2312/3dor.20171049" + }, + { + "type": "ref_text", + "bbox": [ + 0.214, + 0.419, + 0.788, + 0.487 + ], + "angle": 0, + "content": "41. Smith, J.S., Karlinsky, L., Gutta, V., Cascante-Bonilla, P., Kim, D., Arbelle, A., Panda, R., Feris, R., Kira, Z.: Coda-prompt: Continual decomposed attention-based prompting for rehearsal-free continual learning. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 11909-11919 (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.214, + 0.487, + 0.788, + 0.514 + ], + "angle": 0, + "content": "42. Tang, Y.M., Peng, Y.X., Zheng, W.S.: When prompt-based incremental learning does not meet strong pretraining. arXiv preprint arXiv:2308.10445 (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.214, + 0.514, + 0.788, + 0.554 + ], + "angle": 0, + "content": "43. Tao, X., Hong, X., Chang, X., Dong, S., Wei, X., Gong, Y.: Few-shot class-incremental learning. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 12183-12192 (2020)" + }, + { + "type": "ref_text", + "bbox": [ + 0.214, + 0.554, + 0.788, + 0.581 + ], + "angle": 0, + "content": "44. Van Den Oord, A., Vinyals, O., et al.: Neural discrete representation learning. Advances in neural information processing systems 30 (2017)" + }, + { + "type": "ref_text", + "bbox": [ + 0.214, + 0.581, + 0.788, + 0.635 + ], + "angle": 0, + "content": "45. Villa, A., Alcazar, J.L., Alfarra, M., Alhamoud, K., Hurtado, J., Heilbron, F.C., Soto, A., Ghanem, B.: Pivot: Prompting for video continual learning. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 24214-24223 (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.214, + 0.635, + 0.788, + 0.69 + ], + "angle": 0, + "content": "46. Wang, X., Zhang, S., Qing, Z., Gao, C., Zhang, Y., Zhao, D., Sang, N.: Molo: Motion-augmented long-short contrastive learning for few-shot action recognition. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 18011-18021 (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.214, + 0.69, + 0.788, + 0.731 + ], + "angle": 0, + "content": "47. Wang, Y., Huang, Z., Hong, X.: S-prompts learning with pre-trained transformers: An occam's razor for domain incremental learning. Advances in Neural Information Processing Systems 35, 5682-5695 (2022)" + }, + { + "type": "ref_text", + "bbox": [ + 0.214, + 0.731, + 0.788, + 0.785 + ], + "angle": 0, + "content": "48. Wang, Z., Zhang, Z., Ebrahimi, S., Sun, R., Zhang, H., Lee, C.Y., Ren, X., Su, G., Perot, V., Dy, J., et al.: Dualprompt: Complementary prompting for rehearsal-free continual learning. In: European Conference on Computer Vision. pp. 631-648. Springer (2022)" + }, + { + "type": "ref_text", + "bbox": [ + 0.214, + 0.785, + 0.788, + 0.84 + ], + "angle": 0, + "content": "49. Wang, Z., Zhang, Z., Lee, C.Y., Zhang, H., Sun, R., Ren, X., Su, G., Perot, V., Dy, J., Pfister, T.: Learning to prompt for continual learning. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 139-149 (2022)" + }, + { + "type": "list", + "bbox": [ + 0.214, + 0.147, + 0.788, + 0.84 + ], + "angle": 0, + "content": null + } + ], + [ + { + "type": "page_number", + "bbox": [ + 0.218, + 0.116, + 0.236, + 0.127 + ], + "angle": 0, + "content": "18" + }, + { + "type": "header", + "bbox": [ + 0.272, + 0.115, + 0.364, + 0.129 + ], + "angle": 0, + "content": "P. Garg et al." + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.147, + 0.788, + 0.189 + ], + "angle": 0, + "content": "50. Williams, W., Ringer, S., Ash, T., MacLeod, D., Dougherty, J., Hughes, J.: Hierarchical quantized autoencoders. Advances in Neural Information Processing Systems 33, 4524-4535 (2020)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.19, + 0.788, + 0.245 + ], + "angle": 0, + "content": "51. Yang, Y., Yuan, H., Li, X., Lin, Z., Torr, P., Tao, D.: Neural collapse inspired feature-classifier alignment for few-shot class-incremental learning. In: The Eleventh International Conference on Learning Representations (2023), https://openreview.net/forum?id=y5W8tpojhtJ" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.246, + 0.787, + 0.273 + ], + "angle": 0, + "content": "52. Yue, R., Tian, Z., Du, S.: Action recognition based on rgb and skeleton data sets: A survey. Neurocomputing (2022)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.273, + 0.787, + 0.314 + ], + "angle": 0, + "content": "53. Zhang, H.B., Zhang, Y.X., Zhong, B., Lei, Q., Yang, L., Du, J.X., Chen, D.S.: A comprehensive survey of vision-based human action recognition methods. Sensors 19(5), 1005 (2019)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.315, + 0.787, + 0.342 + ], + "angle": 0, + "content": "54. Zheng, C., Vedaldi, A.: Online clustered codebook. In: Proceedings of the IEEE/CVF International Conference on Computer Vision. pp. 22798-22807 (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.343, + 0.787, + 0.384 + ], + "angle": 0, + "content": "55. Zhou, D.W., Wang, F.Y., Ye, H.J., Ma, L., Pu, S., Zhan, D.C.: Forward compatible few-shot class-incremental learning. In: Proceedings of the IEEE/CVF conference on computer vision and pattern recognition. pp. 9046-9056 (2022)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.384, + 0.787, + 0.438 + ], + "angle": 0, + "content": "56. Zhu, A., Ke, Q., Gong, M., Bailey, J.: Adaptive local-component-aware graph convolutional network for one-shot skeleton-based action recognition. In: Proceedings of the IEEE/CVF Winter Conference on Applications of Computer Vision. pp. 6038-6047 (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.439, + 0.787, + 0.482 + ], + "angle": 0, + "content": "57. Zhu, B., Niu, Y., Han, Y., Wu, Y., Zhang, H.: Prompt-aligned gradient for prompt tuning. In: Proceedings of the IEEE/CVF International Conference on Computer Vision. pp. 15659-15669 (2023)" + }, + { + "type": "list", + "bbox": [ + 0.217, + 0.147, + 0.788, + 0.482 + ], + "angle": 0, + "content": null + } + ] +] \ No newline at end of file diff --git a/2024/POET_ Prompt Offset Tuning for Continual Human Action Adaptation/9942db11-78b9-4c6d-815f-c44748e2ea7c_origin.pdf b/2024/POET_ Prompt Offset Tuning for Continual Human Action Adaptation/9942db11-78b9-4c6d-815f-c44748e2ea7c_origin.pdf new file mode 100644 index 0000000000000000000000000000000000000000..6c1a4e985f4e00f08bf9211f99cac9bad7e2f243 --- /dev/null +++ b/2024/POET_ Prompt Offset Tuning for Continual Human Action Adaptation/9942db11-78b9-4c6d-815f-c44748e2ea7c_origin.pdf @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:e3ec9a2086f41c89b2304859c33adcc9c63a7aed800bd4e7808dffb2e61f9e12 +size 2411438 diff --git a/2024/POET_ Prompt Offset Tuning for Continual Human Action Adaptation/full.md b/2024/POET_ Prompt Offset Tuning for Continual Human Action Adaptation/full.md new file mode 100644 index 0000000000000000000000000000000000000000..dfd3abe68113f0ec3a91fce5463f861c961f23cc --- /dev/null +++ b/2024/POET_ Prompt Offset Tuning for Continual Human Action Adaptation/full.md @@ -0,0 +1,294 @@ +# POET: Prompt Offset Tuning for Continual Human Action Adaptation + +Prachi Garg $^{1}$ , K J Joseph $^{3}$ , Vineeth N Balasubramanian $^{3}$ , Necati Cihan Camgoz $^{2}$ , Chengde Wan $^{2}$ , Kenrick Kin $^{2}$ , Weiguang Si $^{2}$ , Shugao Ma $^{2}$ , and Fernando De La Torre $^{1}$ + +1 Carnegie Mellon University, USA + +Meta Reality Labs + +3 Indian Institute of Technology, Hyderabad + +Abstract. As extended reality (XR) is redefining how users interact with computing devices, research in human action recognition is gaining prominence. Typically, models deployed on immersive computing devices are static and limited to their default set of classes. The goal of our research is to provide users and developers with the capability to personalize their experience by adding new action classes to their device models continually. Importantly, a user should be able to add new classes in a low-shot and efficient manner, while this process should not require storing or replaying any of user's sensitive training data. We formalize this problem as privacy-aware few-shot continual action recognition. Towards this end, we propose POET: Prompt-offset Tuning. While existing prompt tuning approaches have shown great promise for continual learning of image, text, and video modalities; they demand access to extensively pretrained transformers. Breaking away from this assumption, POET demonstrates the efficacy of prompt tuning a significantly lightweight backbone, pretrained exclusively on the base class data. We propose a novel spatio-temporal learnable prompt offset tuning approach, and are the first to apply such prompt tuning to Graph Neural Networks. We contribute two new benchmarks for our new problem setting in human action recognition: (i) NTU RGB+D dataset for activity recognition, and (ii) SHREC-2017 dataset for hand gesture recognition. We find that POET consistently outperforms comprehensive benchmarks. + +Keywords: 3D Skeleton Activity Recognition $\cdot$ Extended Reality (XR) $\cdot$ Continual Learning $\cdot$ Prompt Tuning. + +# 1 Introduction + +A key input modality to virtual, augmented and mixed reality (often together termed as extended reality, XR) devices today is through recognizing human + +![](images/30acfb9b2daf18ec6aef71a76a4b9c3a515194d19662e3e8544a9a71a9e5b0cf.jpg) +Fig. 1: Proposed POET method continually adapts skeleton-based human action recognition models pretrained on a pre-defined set of categories to new user categories with few training examples. Users can thus expand the capabilities of XR systems with novel action classes by providing a few examples of each new class. We discard the user-sensitive data as soon as the model is updated on the new categories. + +activity and hand gestures based on body and hand pose estimates. Recognizing human actions5 facilitates seamless user interactions in head-mounted XR devices such as the Meta Quest 3 and Apple Vision Pro. If the provided action recognition models are static, then developers and users are limited to a predefined set of action categories. With the growing use of such devices in new contexts and the increasing demand for personalized technology delivery, there is an impending need to enable the action recognition models in such systems to adapt and learn new user actions over time. Defining their own action categories allows users to customize their experience and expand the functionality of their XR devices. Addressing this need is the primary objective of this work. + +Adapting human action models to new user categories over time faces a few challenges. Firstly, the model must be capable of learning new actions with minimal amount of training data so users can add new classes by providing just a few training examples per class. Secondly, due to the increasing use of XR devices for personal assistance, there is a need for privacy preservation in user action recognition-based pipelines [2, 14]. Hence, the adaptation of such action recognition models to new user categories must also be 'data-free', i.e., it cannot store and replay previously seen user training data in subsequent continual sessions. Considering these requirements, we leverage the recent success of 'data-free' prompt-based learning [49] and propose a new spatio-temporal prompt offset tuning approach to efficiently adapt the default model without finetuning. + +Human action recognition systems are moving to skeleton-based approaches, especially in applications that require low-shot action recognition capabilities such as medical action recognition [27, 56]. Skeletons offer a robust and compact alternative to videos in such low-shot regimes, due to their relatively low dimensionality and lesser variance under background conditions. While there have been a wide variety of efforts in skeleton-based human action recognition over the years [36, 52, 53], there have been fewer efforts on adapting such models to newer user categories. Efforts like [1, 22] attempted to continually learn new user + +categories over time in skeleton-based human action recognition, but relied on fully-supervised data for the new classes. On the other hand, few-shot learning works [27, 46, 56] adapt a pretrained skeleton-based action recognition model to new data, but without explicitly retaining past categories. In this work, we seek to learn new user categories in trained human action models with very few labeled samples for the new classes, while being data-free (not storing samples from previously trained categories). Fig 1 summarizes our overall objective. One could view our setting as privacy-aware few-shot continual learning for skeleton-based action recognition. + +To this end, we propose a prompt offset tuning methodology that can be integrated with existing backbone architectures for skeleton-based human action recognition. Our learnable (soft) prompts are selected from a shared knowledge pool of prompts based on an input instance dependent attention mechanism. In particular, we propose prompt selection using an ordered query-key matching that enables a temporal prompt frame order selection consistent with the input instance. We show that such an approach allows us to learn new user categories without having to store data from past classes, without overwriting the pre-existing categories. To the best of our knowledge, this is the first effort on leveraging prompt tuning for skeleton-based models, as well as on spatiotemporal prompt selection and tuning. + +Our key contributions are summarized below: + +- We formalize a novel problem setting which continually adapts human action models to new user categories over time, in a privacy aware manner. + +- To address this problem, we propose a novel spatio-temporal Prompt OffsEt Tuning methodology (POET). In particular, it is designed to seamlessly plug-and-play with a pretrained model's input embedding, without any significant architectural changes. + +- Our comprehensive experimental evaluation on two benchmark datasets brings out the efficacy of our proposed approach. + +# 2 Related Works + +# 2.1 Prompt Tuning + +The idea of prompting, as it originated from Large Language Models (LLMs), is to include additional information, known as a text prompt, to condition the model's input for generating an output relevant to the prompt. Instead of applying a discrete, pre-defined 'hard' language prompt token, prompt and prefix tuning [20,23] formalized the concept of applying 'soft prompts' to the input. A set of learnable parameters are prepended (concatenated) to the input text and trained along with the classifier while keeping the backbone parameters frozen. Similar to prompt tuning of LLMs, recent works have popularized prompt tuning of ViTs [16] as an effective way of adapting large pretrained models to downstream tasks [49,57]. However, it remains unexplored and undefined (to the best of our knowledge) for non-transformer architectures such as GNNs. + +# 2.2 Prompt Tuning for Continual Learning + +Prompt tuning provides a simple and cost-effective way of learning task-specific signal condensed into 'soft prompts'. For continual learning, training a set of prompts for each sequential task provides a natural alternative to storing privacy violating exemplars and replaying them. Training task-specific prompts for each sequential task is straightforward when authors assume access to task identity at both train and inference time, like in Progressive Prompts [34]. However, if task identity is unavailable at inference, the model will not know which task's prompts or classifier to use for evaluating a test sample. In this respect, S-prompts [47] and A-la-carte prompt tuning (APT) [4] learn an independent set of prompts for each domain/task and employ a KNN-based search for domain/task identity at test time. Since these methods learn stand-alone prompts for every task, the prompt feature space is task-specific, and there is no forgetting of old knowledge when learning new tasks (by design). At the same time however, these 'no forgetting' prompts cannot share knowledge across tasks. + +This leads to another ideology for continual prompt tuning, i.e., treat each prompt unit as being a part of a larger shared (knowledge) pool of prompts. Then the desired number of prompt units can be selected from the pool, conditioned on the input instance itself [41, 48, 49]. Given the scarcity of new data in our setting, we hypothesize that sharing of knowledge will benefit new tasks and draw inspiration from this line of works. Most recently, Adaptive Prompt Generator (APG) [42] challenges the intensive ImageNet21K pre-training assumption as it prompts a ViT pretrained only on the continual benchmark's base class data (similar to us). However, they use replay and knowledge distillation-style 'antiforgetting learning', in addition to using prompts. Even though our backbone is trained only on the base classes, we propose a simple prompt tuning-only strategy to counter forgetting. This implies that a prompt strategy is all we need to continually add new action semantics in a few-shot manner. + +# 2.3 Few-Shot Class Incremental Learning + +FSCIL is a challenging continual learning setting where a model overfits to new classes, with the simultaneous heightened (often complete) forgetting of old knowledge as soon as the base model is fine-tuned on few-shot data [10,43]. Since the backbone feature extractor is the only source of previously seen knowledge, if it is updated, knowledge is lost forever. Typically, existing works decouple the learning of (backbone) feature representations from the classifier by learning the model only on the base data and relying on non-parametric class-mean classifiers for classification in subsequent steps [13,31,55]. This leads to a feature-classifier misalignment issue [32, 51] because new class prototypes are extracted from a backbone representation trained only on the base classes. We hypothesize that optimizing input prompt vectors along with a dynamically expanding parametric classifier on top of a frozen backbone can alleviate this misalignment issue. Our work not only provides a fresh perspective into FSCIL, but to our best knowledge is also the only work not designed for and evaluated on image benchmarks. + +# 3 Preliminaries + +Skeleton Action Recognition Using Graph Representations. Our input $\mathbf{X} \in \mathbb{R}^{T \times J \times 3}$ is a video sequence of $T$ frames, each frame containing $J$ joints of the human body (25 joints) or hand skeleton (22 joints) in 3D Cartesian coordinate system. Such a skeleton action sequence is naturally represented as a graph topology $G = \{\mathcal{V}, \mathcal{E}\}$ with $\mathcal{V}$ vertices and $\mathcal{E}$ edges. Graphs are modeled using Graph Neural Networks (GNNs) [12], which can either be sparse graph convolutional networks (GCN) or fully connected graph transformers (GT). Our main model (a GNN) is defined as $f(\mathbf{X}) = f_c \circ f_g \circ f_e(\mathbf{X})$ (as also shown in Fig. 2). Input $\mathbf{X}$ is first passed through an input embedding layer $f_e$ to get an embedding of human joints $\mathbf{X_e} = f_e(\mathbf{X}), \mathbf{X_e} \in \mathbb{R}^{T \times J \times C_e}$ , with feature dimension $C_e$ . $\mathbf{X_e}$ is further passed to a graph feature extractor $f_g$ composed of a stack of convolutional layers (in GCNs) or attention layers (in GTs), and finally a classifier $f_c$ which predicts the action class label $\mathbf{y}$ . In POET, we propose to attach learnable parameters $\mathbf{P_T}$ (called prompt offsets) to the embedding $\mathbf{X_e}$ . + +Problem Definition. Given a default (pre)trained model deployed on a user's device, we would like to extend this model to new action classes over $T$ subsequent user sessions (also called tasks) $\{\mathcal{US}^{(1)},\dots,\mathcal{US}^{(T)}\}^6$ . In each user session $\mathcal{US}^{(t)}$ , the model learns a dataset $\mathcal{D}^{(t)} = (\mathbf{X}_i^t,\mathbf{y}_i^t)_{i = 1}^{|D^{(t)}|}$ of skeleton action sequence and label pairs provided by the user, $\mathbf{X}_i^t\in \mathbb{R}^{T\times J\times 3}$ , $y_i^t\in \mathbb{R}^{y^{(t)}}$ . In each session, the user typically provides a few training instances $F$ (e.g. $F\leq 5$ ) for each of the $N$ new classes being added, such that $|\mathcal{D}^{(t)}| = NF$ . The base (default) model's session $\mathcal{UB}^{(0)}$ is assumed to have a large number of default action classes $\mathcal{V}^{(0)}$ trained on sufficient data $\mathcal{D}^{(0)}$ , which is most often proprietary and cannot be accessed in later user sessions. In each session, the user adds new action classes such that, $\mathcal{V}^{(t)}\cap \mathcal{V}^{(t')} = \emptyset ,\forall t\neq t'^7$ . Due to the aforementioned privacy constraints, in any training session $\mathcal{US}^{(t)}$ , the model has access to only $\mathcal{D}^{(t)}$ ; after training this data is made inaccessible for use in subsequent sessions (no exemplar or prototypes stored). After training on every new session $\mathcal{US}^{(t)}$ , the model is evaluated on the test set of all classes seen so far $\cup_{i = 0}^{t}\mathcal{V}^{(i)}$ . The challenge is to alleviate forgetting of old classes while not overfitting to the user-provided new class samples. One could view our setting as privacy-aware few-shot continual action recognition, a problem of practical relevance in human action recognition - which has not received adequate attention. + +# 4 Methodology: Prompt Offset Tuning (POET) + +Overview. We propose to prompt tune a base GNN model $f(\cdot)$ by prompts $\mathbf{P_T}$ to address our overall objective. As shown in Fig. 3, for each input instance $\mathbf{X}$ , corresponding prompts $\mathbf{P_T}$ are selected from a pool of prompt parameters, using an input-dependent query and key attention mechanism. The selected prompts + +are added to the input feature embedding (and hence the term 'prompt offsets'), before forwarding to the feature extractor and classifier (shown in Fig. 2). + +To this end, our method, POET uses the same number of prompts as the number of temporal frames in the input, to maintain temporal consistency between the prompt and the input. Focusing solely on prompt offsets allows us to adapt the model to subsequent user sessions without having to update the input embeddings or the feature extraction backbone. Our prompt selection mechanism is learnable and trained along with the classifier to make this method simple and efficient. + +What are Prompt Offsets? Learnable (or soft [20]) prompts are parameter vectors in a continuous space which are optimized to adapt the pretrained frozen backbone $f_{g}$ to each continual task. We define our spatio-temporal prompt offsets $\mathbf{P_T}$ as a set of $T$ prompts (same in number as skeletal frames in input), each prompt $P_{i}$ having length equal to the number of joints in a frame $J$ and feature dimension same as input feature embedding $\mathbf{X_e}$ , i.e., $P_{i} \in \mathbb{R}^{J \times C_{e}}$ . + +Existing prompt tuning efforts, for example in image classification, focus on concatenating learnable prompts to the input token sequence in transformer architectures [16, 20]. Even though trans + +formers can be generalized to graphs [7, 11, 29], it is non-trivial to attach prompts to a GNN. This is because transformers can be viewed as treating sentences or images as fully connected graphs where any word (or image patch) can attend to any other word in the sentence [12]. However, our input is a spatio-temporal graph skeleton of the human joint-bone structure with its own edge connectivity. Concatenating prompts along spatial or temporal dimensions would affect the graph semantics, and also affect standard training strategies such as a forward pass or backpropagation (especially in GCNs). Hence, we attach the selected prompts $\mathbf{P_T}$ to the corresponding input feature embedding $\mathbf{X_e}$ via a prompt attachment operator $f_{p}(.)$ . The class logit distribution $\mathbf{y}$ is thus obtained as: + +$$ +\mathbf {y} = f (\mathbf {X}, \mathbf {P} _ {\mathbf {T}}) = f _ {c} \circ f _ {g} \circ f _ {p} \left(f _ {e} (\mathbf {X}), \mathbf {P} _ {\mathbf {T}}\right) \tag {1} +$$ + +In every user session $t > 0$ , the classifier output dimension expands by $N$ to accommodate the new action classes. Unlike most existing continual prompt tuning works, our feature extractor backbone $f_{g}$ is trained only on the base class data $\mathcal{D}^{(0)}$ and is never fine-tuned on classes from new user sessions $\mathcal{U}\mathcal{S}^{(t)}, t > 0$ . After the base session training, parameters of $f_{g}, f_{e}$ are frozen. + +Prompt Pool Design. As stated in Sec. 2.2, to encourage knowledge sharing across user sessions, we choose to construct a single prompt pool $\mathbf{P}$ which encodes + +![](images/a2ff12afd4f1ec19e137cbe1431a41753ef0ef8f4958409d3b5efa49301b4fb8.jpg) +Fig.2: POET: Prompt-offset Tuning proposes to offset the input feature embedding $\mathbf{X}_{\mathrm{e}}$ of the main model by learnable prompt parameters $\mathbf{P_T}$ for privacy-aware few-shot continual action recognition. We explain prompt selection mechanism in Fig. 3. + +![](images/1d64d8218068d4851448335f14f345fe791b511cc7f3c49d714e715015bbf95a.jpg) +Fig. 3: Selection of our prompts $\mathbf{P_T}$ : Input-dependent query $q$ is matched with keys $K$ using sorted cosine similarity to get an ordered index sequence $(s_i)_{i=1}^T$ of the top $T$ keys. This ordered index sequence is used to select the corresponding ordered prompt sequence $\mathbf{P_T}$ from prompt pool $\mathbf{P}$ . We add $\mathbf{P_T}$ to $\mathbf{X_e}$ , thereby adding an offset to it. Our experimental evaluation confirms that such an additive spatio-temporal prompt offset can balance the plasticity to learn new classes from a few action samples, while maintaining stability on previously learned classes. + +knowledge across the sessions: + +$$ +\mathbf {P} = \left\{P _ {1},.. P _ {i},..., P _ {M} \right\}, \qquad P _ {i} \in \mathbb {R} ^ {J \times C _ {e}}; M = \# \text {p r o m p t s a t t i m e t} \tag {2} +$$ + +For selecting prompts from this pool (Fig. 3), we construct a bijective key-value codebook, treating prompts in the pool $\mathbf{P}$ as values and defining learnable key vectors $\pmb{K} = \{\pmb{k}_1, \dots, \pmb{k}_i, \dots, \pmb{k}_M\}, \pmb{k}_i \in \mathbb{R}^{C_e}$ . A cosine similarity matching $\gamma(.)$ between the query $\pmb{q}$ and keys $\pmb{K}$ is used to find indices of the $T$ closest keys $\mathbb{Z}$ , which in turn are used to select prompts from the pool: + +$$ +\mathbb {Z} = \underset {T} {\operatorname {a r g m a x}} \gamma \left(f _ {q} (\mathbf {X}), \mathbf {K}\right) \tag {3} +$$ + +This quantization process is enabled by a query function $f_{q}(.)$ , which is a pretrained encoder that maps an input instance $\mathbf{X}$ to a query $\pmb{q}$ as: + +$$ +\boldsymbol {q} = f _ {q} (\mathbf {X}) = f _ {Q A} \circ f _ {g} ^ {\prime} \circ f _ {e} ^ {\prime} (\mathbf {X}), \quad f _ {q}: \mathbb {R} ^ {T \times J \times 3} \longrightarrow \mathbb {R} ^ {C _ {e}} \tag {4} +$$ + +where the query adaptor $f_{QA}$ is a fully connected layer mapping the $f_{g}^{\prime}$ output dimension to the desired prompt embedding dimension $C_e$ . + +Coupled Optimization in User Sessions $t > 0$ . Typically, the argmax operator in Eq. 3 decouples the optimization of keys from the prompt pool and main model as it prevents backpropagation of gradients to the keys (also seen in earlier works such as [48, 49]). However, this approach does not work for our setting, even more so since we assume no large-scale pre-training of our base model. Due to the lack of off-the-shelf availability of large-scale models for skeletal action data, our query function $f_{q}$ is pretrained only on base class data $\mathcal{D}^{(0)}$ . Hence, it becomes important that $f_{q}$ is updated as the model learns new classes. As shown in red boxes in Figs. 2 and 3, we propose to couple this optimization process such that the overall cross-entropy loss for new tasks updates: (i) the classifier $f_{c}$ , (ii) selected prompts in $\mathbf{P}$ , (iii) selected keys in $K$ , as well as (iv) query adaptor $f_{QA}$ . We achieve this by approximating the gradient for $K$ and $f_{QA}$ by the straight-through estimator reparameterization trick as in [3, 44]. We freeze the query feature extractor layers $f_{g}^{\prime}, f_{e}^{\prime}$ in $t > 0$ to prevent catastrophic forgetting of base knowledge in $f_{q}$ . Our cross-entropy loss is hence given by: + +$$ +\min _ {\theta_ {f _ {Q A}}, \theta_ {K}, \theta_ {\mathbf {P}}, \theta_ {f _ {c}}} \mathcal {L} (f (\mathbf {X}, \mathbf {P} _ {\mathbf {T}}), \mathbf {y}) \tag {5} +$$ + +To move queries closer to their aligned $T$ keys during training, we use a vector quantization clustering loss inspired from VQ-VAE [44] as: + +$$ +\max _ {\theta_ {f _ {Q A}}, \theta_ {K}} \lambda \sum_ {i \in \mathbb {Z}} \gamma \left(f _ {q} (\mathbf {X}), K _ {i}\right) \tag {6} +$$ + +where $\lambda$ is the clustering loss coefficient. Our end-to-end optimization thus establishes a prompt optimization framework which is amenable to prompt tuning when extensive pre-training is not possible. This sets the foundation for our spatio-temporal prompt selection module, described next. + +Spatio-Temporal Prompt Selection. In order to ensure that our learned prompts respect temporal information in the input video sequence, we choose the number of selected prompts to be equal to the number of frames in the input + +video $T$ . After coupling the prompt pool and keys, we observed in our initial experiments with pool size $M > T$ that the same set of prompts get selected across training iterations and user sessions (Fig. 4A). More concretely, as the vector quantization loss (Eqn 6) brings the query close to the selected keys, the same set of active prompts get selected and optimized in each iteration, not using other prompts at all. This is similar to the well-known issue of 'codebook col + +![](images/655464329a10d8067e2e10803916ebc135458dab373e9d280dcd48d26a17afed.jpg) + +![](images/1183da4b6d6d075e6477a447af1a7ac4f6075e36005e75afa5cc5623a79d6a0d.jpg) +Fig.4: $M > T$ Case: Prompt Pool Collapse. (Top) Certain prompt indices remain unused across user sessions. (Bottom) Our POET pool expansion strategy alleviates pool collapse. + +lapse' in VQ-VAE [9, 50, 54]. Based on this observation, we design two prompt pool update mechanisms in user sessions $t > 0$ as below: + +1. Case 1, $M = T \forall t$ : No pool expansion, Algorithm 1. All prompts are selected in all tasks. But the order of their selection $(s_i)_{i=1}^T$ varies with each input instance as we replace Eq. 3 by sorting the cosine similarity before selecting the top $T$ indices as follows: + +$$ +\mathbb {Z} = \underset {\left(s _ {i}\right) _ {i = 1} ^ {T}} {\operatorname {a r g s o r t}} \gamma \left(f _ {q} (\mathbf {X}), \mathbf {K}\right) \tag {7} +$$ + +In Fig. 5, we visualize the positions occupied by indices in this (sorted) ordered key index sequence $(s_i)_{i=1}^T$ . Entropy increase across tasks $t = 1$ to $t = 4$ (bottom row of figure) shows that our selection mechanism learns to select a unique temporal code for all inputs. + +2. Case 2, $M = T + (R * t), t > 0$ . Expand pool with $R$ prompts. We also propose an order-aware prompt pool expansion strategy (Appendix B) that selects prompts from an expanded pool in a temporally coherent manner, for $t > 0$ . This alleviates prompt pool collapse as shown in Fig. 4B. + +Prompt Offset Attachment. Since concatenation is not meaningful for graph data, we use addition as our choice for the prompt attachment operator as: + +$$ +f _ {p} \left(\mathbf {X} _ {\mathbf {e}}, \mathbf {P} _ {\mathbf {T}}\right) = \mathbf {X} _ {\mathbf {e}} + \mathbf {P} _ {\mathbf {T}} \tag {8} +$$ + +Hence, we call our approach as prompt offset tuning. We also study this empirically through experiments that support this choice in Sec. 6. + +Interpreting Prompt Offset Tuning of GNNs. Our additive prompt offsets are open to interpretation, as shown in Fig. 5. (i) Adding our selected prompts $\mathbf{P_T}$ to input feature embedding $\mathbf{X_e}$ acts like an input-dependent transformation + +for spatio-temporal joints. + +(ii) As our prompts have same size as $\mathbf{X}_{\mathbf{e}}$ it can also be thought of as a learned prompt encoding, bearing similarity with learnable position encoding works [12, 24, 26]. Our purpose is different however as prompt offsets seek to dynamically condition the input for adapting the backbone continually, + +![](images/8b6936c5caba33ff9e421a5b926c64d9a3beb54d70507904b14af0dd6f349945.jpg) +Fig. 5: Here we visualize the order $(s_i)_{i=1}^T$ in which the $M = 64$ prompts in the pool are selected at train time, across 4 user sessions $\mathcal{U}\mathcal{S}^{(t)}$ . X-axis: prompt index, Y-axis: index position in selected sequence. Top: The no sorting case uses the default sequence (hence diagonal matrices), giving equal importance to all prompts. Bottom (Our Method): Even though the same 64 prompts are selected and updated, the ordering is temporally unique and consistent with input. + +instead of learning positions. (iii) POET also bears similarity with auto-decoders like DeepSDF [30] which learn latent codes for each style or shape and use relevant codes along with a frozen decoder at inference. (iv) Prompt tuning can also be thought of as a parameter isolation technique for continual learning [28,34,35,38]. POET's ordered prompt selection as seen in Fig. 5 learns to isolate the relevant sequence of prompts for each input action sequence. + +Algorithm 1 POET at Train Time, $t > 0$ (Case 1 $M = T$ , No pool expansion) +Input: Query function $f_{q}$ keys $\pmb {K} = \{\pmb {k}_j\}_{j = 1}^T$ prompt pool $\mathbf{P} = \{\pmb {P}_j\}_{j = 1}^T$ main model $f_{e},f_{g},f_{c}$ Initialize: $\mathbf{P},\mathbf{K}$ from $t - 1$ ; Expand $f_{c}$ by $N$ new classes. Initialize $f_{c}$ as: (i) copy $f_{c}^{old}$ weights, (ii) $f_{c}^{new}\gets Mean(f_{c}^{old})$ Freeze: query layers $f_{g}^{\prime},f_{e}^{\prime}$ ; main model layers $f_{e},f_{g}$ for epochs and batch $(\mathbf{X}_i^t,\mathbf{y}_i^t)_{i = 1}^{NK}$ do 1. Get query feature $\pmb{q}$ (Eq.4); Compute $\gamma (\cdot)$ b/w query $\pmb{q}$ and keys $\pmb{K}$ 2.Sort $\gamma (\cdot)$ ; Get ordered key index sequence $(s_i)_i = 1$ Eq.7 3. Read pool memory $\mathbf{P}$ in order $(s_i)^T_{i = 1}\rightarrow$ Get prompt offsets $\mathbf{P_T}$ 4.Get $\mathbf{X_e};$ Add $\mathbf{P_T}$ to it (Eq.8); get prediction y from prompted input (Eq. 1) 5.Use cross entropy loss (Equation 5) to update $f_{QA},K,\mathbf{P},f_{c}$ 6.Use clustering loss (Equation 6) to update $f_{QA}$ and $\pmb{K}$ end // See $t = 0$ training protocol in Algorithm 2 in Appendix + +# 5 Experiments and Results + +Datasets. We evaluated our method on well-known action recognition datasets: (i) activity recognition on the NTU RGB+D dataset [39]; and (ii) hand gesture + +recognition on the SHREC-2017 dataset [40]. As we introduce a new problem setting in human action recognition, we contribute two new benchmarks to the community for this setting, on the NTU RGB+D and SHREC-2017 datasets. + +For the NTU RGB+D dataset, we divide the 60 daily action categories into 40 base classes, learning the remaining 20 classes in subsequent user sessions. In few-shot learning parlance, our protocol is 4-task 5-way 5-shot, i.e. 5 novel classes using 5 user training instances in 4 user sessions. Each input 3D skeleton sequence has 64 temporal frames, each consisting of 25 body keypoints, such that $x \in \mathcal{R}^{64 \times 25 \times 3}$ . We use the spatio-temporal GCN, CTR-GCN [8], as the architecture for NTU RGB+D, where we choose the joint input modality for better interpretability of prompt tuning. + +For SHREC-2017, we divide the 14 fine-grained hand gesture classes into 8 base classes and 6 classes learned in subsequent user sessions. This is done in a 3-task 2-way 5-shot protocol, i.e. 2 novel classes using 5 user training instances in 3 user sessions. For each input instance of SHREC-2017, we use 8 temporal frames each having 22 hand keypoints, such that input $x \in \mathcal{R}^{8 \times 22 \times 3}$ . We use a fully-connected graph transformer backbone, DG-STA [7] for SHREC-2017. We select DG-STA due to easily reproducible code and to validate if our method POET works equally well across graph convolutional networks and graph transformers. + +Evaluation Metrics. Following earlier work in similar settings [31], we report: (i) Average accuracy 'Avg' of all classes seen so far, and (ii) Harmonic Mean $A_{HM}$ between 'accuracy only on Old classes' and 'accuracy only on New classes' after learning each new user session. Note that the average accuracy tends to be biased towards the base session $\mathcal{T}^{(0)}$ performance due to more number of base classes. A higher $A_{HM}$ implies better stability-plasticity trade-off between new task performance and old tasks' retention. Unlike many earlier CIL efforts, we report accuracy for both Old and New classes in each user session for transparency. + +Implementation Details. We observe that a key source of forgetting in our setting is from the classifier as the logits tend to become heavily biased towards the few-shot samples of new classes. We use a cosine classifier for activity recognition experiments on CTR-GCN. For gesture recognition on the lightweight DG-STA, we use a standard fully-connected layer as classifier, but freeze old class parameters in the classifier by zeroing their gradients. We attach prompts after the 1st layer of DG-STA and 1st CTR-GC block of CTR-GCN. For both datasets, we have equal or higher learning rates in user sessions when compared to the base model's training in order to accommodate new knowledge in the model (for better plasticity). For exact implementation details (including learning rates, epochs, hyperparameter analysis, and backward forgetting metric), see Appendix A. In earlier efforts that more generally tune prompts for class-incremental learning [41, 45, 47-49], it is common to rely on an ImageNet21K pretrained ViT [37] or CLIP [33] as the backbone. However, such backbones do not exist for skeleton-based human action recognition. Our base feature extractor is hence trained on the base session dataset itself without any pretraining, making this one of the first efforts of prompt tuning without extensive pretraining (scale of data 3-5 times lower order of magnitude). + +Results. Since there are no existing baselines for our proposed setting in skeletal action recognition, we compare our method by adapting continual learning (CL) baselines to skeletal data in Sec 5.1, Tables 1, 2. We first compare POET with prompt tuning based class-incremental learning (CIL) approaches originally designed for images (L2P [49], CODA-P [41], APT [4]) and find that it has very low performance on new classes as they do not update their query function. We find any fine-tuning or knowledge distillation based approaches (LWF [25], EWC [17], LUCIR [15]) lead to rapid forgetting of base knowledge as the model overfits to user's few-shots. We also compare with multiple variants of Feature Extraction (FE) to check if prompts truly have merit (POET=FE+Prompts) and provide upper bound baselines. In Sec 6, we first show the importance of prompts in POET by removing the prompts. We discuss the value of our coupled optimization, query function update and ordered key index selection in our prompt selection ablation Tab 3. We also study the impact of proposed additive prompt tuning as compared to other possible prompt attachments $f_{p}$ in Tab 4. + +# 5.1 Comparison with State-of-the-Art + +POET sets the SOTA on existing prompt tuning works (Tab 1,2). We adapt three standard CIL works that prompt tune ViTs for images - L2P [49], CODA-P [41] and APT [4] to our setting. L2P and CODA-P share prompt pool across tasks (similar to us), whereas APT learns task-specific prompts. L2P decouples the optimization of keys from the prompt pool and concatenates the selected prompts. Since concatenation is not defined for our GNN backbone, we adapt these SOTA to our setting by concatenating along the temporal dimension $(\mathbf{L2P}^*,\mathbf{CODA - P}^*)$ . CODA-P [41] couples keys with the prompt pool by using a cosine similarity weighing over all prompts in the pool, forming a 'soft prompt selection', different from our 'ordered hard prompt selection'. In APT, we train prompt-classifier pairs for each continual task separately ( $\hat{\mathbf{\Gamma}}$ denotes task-specific), and use task identity at test time. See details in Appendix A. These methods by design rely on extensively pretrained (ImageNet21k) query functions which does not require updates; and require full supervision on new classes, perhaps explaining their poor 'New' accuracy in our few-shot setting. + +Standard Continual learning Baselines. We compared with two well established knowledge-distillation approaches, learning without forgetting (LWF) and LUCIR. Both of them perform poorly on both old and new classes. EWC [17] learns better on new but does not retain old knowledge. We conclude that any CL method that fine-tunes the backbone feature representation in subsequent sessions $t > 0$ will not be able to retain base/old class knowledge (a finding consistent with existing FSCIL literature for images [10,43]). We also adapt and compare with one of the latest FSCIL baselines ALICE [31], originally developed for image classification benchmarks on our gesture recognition benchmark in Table 2. Note the high retention of base task performance (due to non-parametric classifier on top of frozen base model). However, it suffers from poor plasticity and adaptation to new classes. This is the issue of feature-classifier misalignment that we hoped to alleviate through prompt tuning. + +Table 1: Activity Recognition Results $(\% ,\uparrow)$ , Comparison with SOTA: NTU RGB+D [39] dataset on CTR-GCN [8] backbone. After training on each incremental task, we report Average of all classes seen so far ('Avg'). We also report (i) $A_{HM}$ , (ii) old classes accuracy ('Old'), (iii) new classes accuracy ('New') in the last session. We report Mean and STD across 10 sets of 5-shots. POET achieves the best stability-plasticity trade-off across all baselines indicated by the $A_{HM} = 56.3\%$ . POET also has the highest Avg across all user sessions outside of upper bound baselines in orange. + +
Method\( \mathbf{U}\mathbf{B}^{(0)} \)\( \mathbf{U}\mathbf{S}^{(1)} \)\( \mathbf{U}\mathbf{S}^{(2)} \)\( \mathbf{U}\mathbf{S}^{(3)} \)\( \mathbf{U}\mathbf{S}^{(4)} \)
Base (↑)Avg (↑)Avg (↑)Avg (↑)Old (↑)New (↑)Avg (↑)\( A_{HM} \) (↑)
Upper Bounds
Joint (Oracle)88.479.071.066.863.5
Joint POET (Oracle)67.2
FE, Task-Specific^88.470.1 ± 2.652.5 ± 5.844.8 ± 5.070.3 ± 2.146.7 ± 2.0NANA
FE+Replay88.482.4 ± 1.178.2 ± 1.274.5 ± 1.273.1 ± 1.043.3 ± 3.370.6 ± 1.254.3 ± 2.6
Continual Linear Probing
FE88.472.0 ± 1.160.4 ± 2.447.7 ± 2.140.0 ± 1.651.0 ± 2.340.9 ± 1.444.8 ± 1.1
FE, Frozen88.476.1 ± 1.052.4 ± 4.138.3 ± 2.728.4 ± 1.622.4 ± 4.527.9 ± 1.424.8 ± 3.0
FE+Replay†88.472.0 ± 1.559.5 ± 4.058.7 ± 2.856.7 ± 2.534.7 ± 5.654.9 ± 2.742.8 ± 4.4
FT88.46.2 ± 1.44.3 ± 1.52.8 ± 1.00.2 ± 0.536.0 ± 10.13.2 ± 0.80.3 ± 1.0
Standard Continual Learning
LWF [25]88.46.2 ± 1.52.8 ± 0.73.7 ± 1.30.0 ± 0.038.9 ± 8.83.2 ± 0.70.0 ± 0.0
EWC [17]88.46.6 ± 1.54.1 ± 1.43.1 ± 0.90.0 ± 0.042.1 ± 9.53.5 ± 0.80.0 ± 0.0
Experience Replay88.435.1 ± 8.350.6 ± 5.060.6 ± 5.454.6 ± 6.543.7 ± 14.653.7 ± 7.147.8 ± 11.2
Experience \( \text{Replay}^{\dagger} \)88.46.2 ± 1.59.0 ± 2.611.2 ± 3.010.9 ± 2.634.6 ± 7.912.9 ± 3.016.3 ± 3.5
LUCIR [15]87.94.3 ± 2.14.1 ± 1.32.7 ± 0.80.2 ± 0.426.0 ± 9.22.3 ± 0.90.4 ± 0.8
Continual Prompt Tuning
CODA-P [41]*87.476.1 ± 1.066.7 ± 1.358.6 ± 2.756.5 ± 2.90.5 ± 0.451.8 ± 2.71.1 ± 0.7
L2P [49]*88.678.9 ± 0.171.0 ± 1.064.2 ± 0.162.0 ± 0.70.0 ± 0.056.8 ± 0.60.0 ± 0.0
APT [4]^86.627.3 ± 1.630.8 ± 3.437.6 ± 2.3NA33.4 ± 2.0NANA
POET (Ours)87.982.3 ± 0.676.8 ± 0.968.4 ± 0.757.2 ± 1.055.8 ± 5.957.1 ± 1.156.3 ± 3.2
+ +Fine-tuning (FE) and Feature Extraction (FE) Baselines. We implement standard continual learning baselines to understand stability-plasticity trade-offs in our new benchmarks. In all these baselines, we expand the classifier output dimension by $N$ new classes. In 'FT (Fine-Tuning)', we tune all model parameters on cross entropy loss of new task. FSCIL is challenging for this modality as old task performance sharply reduces to zero starting from $\mathcal{U}\mathcal{S}^{(1)}$ as model overfits to user's few-shots. 'FE (Feature Extraction)' differs from FT as we freeze the feature extractor to preserve base knowledge. This serves as a competitive baseline in our findings. In 'FE, frozen', we zero out the gradients of previous class weights in classifier $f_{c}$ to prevent forgetting from the classifier. 'FE' and 'FE, Frozen' exhibit different New-Old trade-offs in Tables 1, 2 because the scale of pretraining is different (gesture more lightweight than activity). + +Upper-bound baselines, top section Tables 1, 2. In 'Joint (oracle)' experiment, we train on all task data at the same time in a multi-task (non-sequential) manner. Training POET in a multi-task manner ('Joint POET') outperforms 'Joint Oracle' demonstrating the strength of our approach. In addition to these generalist upper bounds, we point out that 'FE, Task-specific' is a competitive specialist upper bound. In this, we perform feature extraction from base model to each task individually, storing separate task-specific models $(\mathcal{U}\mathcal{S}^{(0)}\to \mathcal{U}\mathcal{S}^{(i)},i > 0)$ . POET outperforms 'New' accuracy compared with this baseline, achieving a forward transfer on each $t > 0$ . This indicates that prompt + +Table 2: Gesture Recognition Results (\%, ↑), Comparison with SOTA: SHREC 2017 [40] dataset on DG-STA [7] graph transformer backbone. Reporting mean and standard deviation across 5 runs. POET achieves best $A_{HM} = 56.2\%$ . + +
Method\( \mathbf{\mathcal{U}\mathbf{B}^{(0)}} \)\( \mathbf{\mathcal{U}\mathbf{S}^{(1)}} \)\( \mathbf{\mathcal{U}\mathbf{S}^{(2)}} \)\( \mathbf{\mathcal{U}\mathbf{S}^{(3)}} \)
Base (↑)Avg (↑)Avg (↑)Old (↑)New (↑)Avg (↑)\( A_{HM} \) (↑)
Joint (Oracle)88.879.4 ± 0.777.3 ± 2.170.9 ± 1.262.4 ± 0.4
FT88.820.3 ± 0.812.4 ± 2.10.0 ± 0.085.8 ± 9.413.4 ± 1.50.0 ± 0.0
FE88.862.7 ± 2.441.9 ± 6.917.5 ± 5.177.3 ± 8.826.8 ± 3.428.5 ± 6.4
FE, Frozen88.871.3 ± 1.961.4 ± 2.744.7 ± 3.254.5 ± 6.746.2 ± 2.749.1 ± 4.3
LWF [25]88.820.2 ± 1.412.5 ± 1.00.0 ± 0.088.4 ± 13.713.8 ± 2.10.0 ± 0.0
L2P [49]**88.820.3 ± 5.910.5 ± 4.88.2 ± 4.06.9 ± 8.57.9 ± 3.97.5 ± 5.5
CODA-P [41]**87.715.6 ± 4.511.6 ± 1.97.9 ± 1.814.1 ± 21.48.8 ± 2.410.1 ± 3.2
ALICE [31]92.172.4 ± 5.763.3 ± 7.662.5 ± 6.811.9 ± 9.954.6 ± 6.920.0 ± 8.1
POET (Ours)91.973.2 ± 3.761.9 ± 1.845.9 ± 2.672.4 ± 7.150.0 ± 1.656.2 ± 1.6
+ +tuning benefits New performance due to the pre-existing knowledge in the shared knowledge pool. Avg in sessions $0 < t < 4$ indicates New for task-specific $^{\wedge}$ . + +Experience Replay Baselines, Tab 1. Even though our privacy-aware setting prohibits previous data replay, we compare with 'Experience Replay' (store and replay 5-samples of base and incremental sessions) and 'Experience Replay†' (replay only previous incremental sessions) for completeness. + +'FE+Replay' serves as the best upper bound (even better than Experience Replay as we are freezing backbone in addition to replay). It is noteworthy that POET (which is FE+prompts) learns an implicit 'data-free' form of prompt pool memory, and yet has a better $A_{HM}$ trade-off as compared to explicitly stored and replayed samples from previous classes in FE+replay. + +# 6 Ablation Studies and Analysis + +Importance of prompts in POET. First, we consider the contribution of prompt offsets in POET. Since we only attach prompts to address continual learning in POET, removing prompts gives the Feature Extraction (FE) baseline ('w/o prompts', Table 3) where the backbone is frozen after base training and only the classifier is expanded and updated on classification loss of new classes. POET improves both, 'Old' ( $\uparrow 20.1\%$ ) and 'New' ( $\uparrow 10.6\%$ ) marked in blue. + +Prompt Selection Mechanism. In Table 3, we investigate our prompt selection mechanism and optimization choices. The 'w/o coupled optim.' experiment is a direct comparison of our additive prompt attachment with the de-coupled optimization in L2P [49]. Updating key parameters but keeping only query adaptor + +Table 3: Prompt Selection Mechanism Analysis on NTU RGB+D dataset $(\% ,\uparrow)$ .. $\mathrm{w / o}$ denotes removing that component from POET, numbers in brackets are wrt POET $(M = T)$ experiment. 'Avg' accuracy is biased towards Old' classes accuracy, $A_{HM}$ is good indicator of trade-off between New' and Old'. + +
NTU RGB+D\( \left| \mathcal{U}\mathcal{B}^{(0)}\right| \)\( \left| \mathcal{U}\mathcal{S}^{(1)}\right| \)\( \left| \mathcal{U}\mathcal{S}^{(2)}\right| \)\( \left| \mathcal{U}\mathcal{S}^{(3)}\right| \)\( \left| \mathcal{U}\mathcal{S}^{(4)}\right| \)
MethodBaseAvgAvgAvgOldNewAvg\( A_{HM} \)
w/o prompts88.474.566.349.539.2 (-20.1)46.8 (-10.6)39.942.7
w/o coupled optim.88.082.875.365.856.5 (-2.8)51.3 (-6.1)56.153.8
w/o clustering loss85.581.674.364.562.0 (+2.7)18.2 (-39.2)57.028.1
w/o QA update87.982.877.469.159.4 (+0.1)52.8 (-4.6)58.755.9
w/o sorting88.282.275.268.859.9 (+0.6)46.6 (-10.8)58.852.4
POET (M>T)87.982.777.268.860.3 (+1.0)54.4 (-3.0)59.857.2
POET (M=T)87.982.876.868.659.357.459.258.3
+ +$QA$ frozen after $\mathcal{U}\mathcal{B}^{(0)}$ training ('w/o QA update') reduces 'New' only performance of $\mathcal{US}^{(4)}$ by $4.6\%$ as the query function stays fixed at base session learning and is not discriminative towards new classes. 'W/o clustering loss' from Eq. 6, performance drops starting from $\mathcal{UB}^{(0)}$ itself. The only difference + +between the experiment 'w/o sorting' and 'POET (M=T)' is that we do not sort the cosine similarity before selecting top $T$ indices (same as Fig 5). The $10.8\% \uparrow$ in 'New' performance validates that our prompt selection mechanism is learning to choose a distinct temporal ordering for prompt tuning of new input samples. With pool expansion ('POET, $M > T$ ), we get more flexibility in the stability-plasticity trade-offs depending on how many new prompts we attach. For $R = 6$ , 'Old' is improved. In Table 3, we keep POET's additive prompt attachment and only vary prompt selection. + +Prompt Attachment Mechanism. In Table 4, we keep our end-to-end optimization and ordered prompt selection as a constant and ablate prompt shape and attachment operator $f_{p}(.)$ . Drawing a parallel with transformers which concatenate prompts along the token dimension, we conduct experiments concatenating prompts along the (i) temporal dimension of the skeleton input feature embedding $\mathbf{X}_{\mathrm{e}}$ ('CONCAT temporal') and (ii) feature dimension $C_e$ ('CONCAT feature'). We find that addition works better than concatenation and + +cross attention. We also verify our hypothesis that selecting the same number of prompts as the input temporal dimension ( $T = 64$ for NTU RGB+D and $T = 8$ for SHREC-2017) yields better results as compared to adding the same prompt frame to each input embedding frame ('Addition $T' = 1'$ ). + +Table 4: Prompt Attachment Analysis (%, ↑): The best prompt attachment choice $f_{p}(.)$ is Adding #prompts same as #input frames (T=64). + +
NTU RGB+D\( \left\| u\mathcal{B}^{(0)}\right\| \)\( \left\| u\mathcal{S}^{(1)}\right\| \)\( \left\| u\mathcal{S}^{(2)}\right\| \)\( \left\| u\mathcal{S}^{(3)}\right\| \)\( \left\| u\mathcal{S}^{(4)}\right\| \)
MethodBaseAvgAvgAvgOld New Avg \( A_{HM} \)
CONCAT temporal, \( T' = 64 \)88.670.362.449.833.6 50.5 35.1 40.3
CONCAT feature, \( T' = 64 \)87.782.475.566.957.1 41.5 56.0 48.1
Cross Attention, \( T' = 64 \)82.977.472.265.057.1 32.3 55.0 41.2
ADD, \( T' = 1 \)88.773.362.745.533.7 47.0 34.8 39.3
ADD, \( T' = 64 \) (Ours)87.982.876.868.659.3 57.4 59.2 58.3
+ +# 7 Conclusions and Future Work + +The problem of continually adapting human action models to new user categories over time has gained prominence with the rising availability of XR devices. However, this setting poses unique challenges: (i) the user may be able to provide only a few samples for training, and (ii) accessing data from earlier sessions may violate privacy considerations. We hence propose a method based on prompt offset tuning to address this problem in this work. Prompt tuning to address learning over newer tasks has been attempted in recent years. However, these works have: (1) typically been designed for image-based tasks, (2) relied on strongly pretrained transformer backbones, (3) required full supervision for new tasks, and (4) exclusively applied prompt tuning to transformer architectures. This work departs from these four characteristics. Our work demonstrates that prompt offset tuning is a promising option to evolve and adapt skeleton-based human action models to new user classes. The careful design of each component of the proposed methodology finds validation in the promising results across well-known skeleton-based action recognition benchmarks. Our ablation studies and analysis corroborate our design choices in our implementation. Looking ahead, it will be interesting to explore how our approach and its design choices adapt when a "generalist backbone" trained on a large corpus of action recognition data becomes accessible. Extending our method for differential privacy is another interesting direction of future work. + +# References + +1. Aich, S., Ruiz-Santaquiteria, J., Lu, Z., Garg, P., Joseph, K.J., Fernandez, A., Balasubramanian, V.N., Kin, K., Wan, C., Camgoz, N.C., Ma, S., De la Torre, F.: Data-free class-incremental hand gesture recognition. In: Proceedings of the IEEE/CVF International Conference on Computer Vision (2023) +2. Albrecht, J.P.: How the gdpr will change the world. Eur. Data Prot. L. Rev. 2, 287 (2016) +3. Bengio, Y., Léonard, N., Courville, A.: Estimating or propagating gradients through stochastic neurons for conditional computation. arXiv preprint arXiv:1308.3432 (2013) +4. Bowman, B., Achille, A., Zancato, L., Trager, M., Perera, P., Paolini, G., Soatto, S.: a-la-carte prompt tuning (apt): Combining distinct data via composable prompting. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 14984-14993 (2023) +5. Chaudhry, A., Dokania, P.K., Ajanthan, T., Torr, P.H.: Riemannian walk for incremental learning: Understanding forgetting and intransigence. In: Proceedings of the European conference on computer vision (ECCV). pp. 532-547 (2018) +6. Chaudhry, A., Ranzato, M., Rohrbach, M., Elhoseiny, M.: Efficient lifelong learning with a-gem. arXiv preprint arXiv:1812.00420 (2018) +7. Chen, Y., Zhao, L., Peng, X., Yuan, J., Metaxas, D.N.: Construct dynamic graphs for hand gesture recognition via spatial-temporal attention. In: Proceedings of the British Machine Vision Conference (BMVC) (2019) +8. Chen, Y., Zhang, Z., Yuan, C., Li, B., Deng, Y., Hu, W.: Channel-wise topology refinement graph convolution for skeleton-based action recognition. In: Proceedings of the IEEE/CVF International Conference on Computer Vision. pp. 13359-13368 (2021) +9. Dhariwal, P., Jun, H., Payne, C., Kim, J.W., Radford, A., Sutskever, I.: Jukebox: A generative model for music. arXiv preprint arXiv:2005.00341 (2020) +0. Dong, S., Hong, X., Tao, X., Chang, X., Wei, X., Gong, Y.: Few-shot class incremental learning via relation knowledge distillation. In: Proceedings of the AAAI Conference on Artificial Intelligence. vol. 35, pp. 1255-1263 (2021) +1. Dwivedi, V.P., Bresson, X.: A generalization of transformer networks to graphs. arXiv preprint arXiv:2012.09699 (2020) +2. Dwivedi, V.P., Luu, A.T., Laurent, T., Bengio, Y., Bresson, X.: Graph neural networks with learnable structural and positional representations. In: International Conference on Learning Representations (2022), https://openreview.net/forum?id=wTTjnvGphYj +3. Hersche, M., Karunaratne, G., Cherubini, G., Benini, L., Sebastian, A., Rahimi, A.: Constrained few-shot class-incremental learning. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 9057-9067 (2022) +4. Hinojosa, C., Marquez, M., Arguello, H., Adeli, E., Fei-Fei, L., Niebles, J.C.: Privhar: Recognizing human actions from privacy-preserving lens. In: European Conference on Computer Vision. pp. 314-332. Springer (2022) +5. Hou, S., Pan, X., Loy, C.C., Wang, Z., Lin, D.: Learning a unified classifier incrementally via rebalancing. In: Proceedings of the IEEE/CVF conference on computer vision and pattern recognition. pp. 831-839 (2019) +6. Jia, M., Tang, L., Chen, B.C., Cardie, C., Belongie, S., Hariharan, B., Lim, S.N.: Visual prompt tuning. In: European Conference on Computer Vision. pp. 709-727. Springer (2022) + +17. Kirkpatrick, J., Pascanu, R., Rabinowitz, N., Veness, J., Desjardins, G., Rusu, A.A., Milan, K., Quan, J., Ramalho, T., Grabska-Barwinska, A., et al.: Overcoming catastrophic forgetting in neural networks. Proceedings of the national academy of sciences 114(13), 3521-3526 (2017) +18. Kumawat, S., Nagahara, H.: Privacy-preserving action recognition via motion difference quantization. In: European Conference on Computer Vision. pp. 518-534. Springer (2022) +19. Lacoste, A., Luccioni, A., Schmidt, V., Dandes, T.: Quantifying the carbon emissions of machine learning. arXiv preprint arXiv:1910.09700 (2019) +20. Lester, B., Al-Rfou, R., Constant, N.: The power of scale for parameter-efficient prompt tuning. arXiv preprint arXiv:2104.08691 (2021) +21. Li, M., Xu, X., Fan, H., Zhou, P., Liu, J., Liu, J.W., Li, J., Keppo, J., Shou, M.Z., Yan, S.: Stprivacy: Spatio-temporal privacy-preserving action recognition. In: Proceedings of the IEEE/CVF International Conference on Computer Vision. pp. 5106-5115 (2023) +22. Li, T., Ke, Q., Rahmani, H., Ho, R.E., Ding, H., Liu, J.: Else-net: Elastic semantic network for continual action recognition from skeleton data. In: Proceedings of the IEEE/CVF International Conference on Computer Vision. pp. 13434-13443 (2021) +23. Li, X.L., Liang, P.: Prefix-tuning: Optimizing continuous prompts for generation. arXiv preprint arXiv:2101.00190 (2021) +24. Li, Y., Si, S., Li, G., Hsieh, C.J., Bengio, S.: Learnable fourier features for multidimensional spatial positional encoding. Advances in Neural Information Processing Systems 34, 15816-15829 (2021) +25. Li, Z., Hoiem, D.: Learning without forgetting. IEEE transactions on pattern analysis and machine intelligence 40(12), 2935-2947 (2017) +26. Liu, X., Yu, H.F., Dhillon, I., Hsieh, C.J.: Learning to encode position for transformer with continuous dynamical model. In: International conference on machine learning. pp. 6327-6335. PMLR (2020) +27. Ma, N., Zhang, H., Li, X., Zhou, S., Zhang, Z., Wen, J., Li, H., Gu, J., Bu, J.: Learning spatial-preserved skeleton representations for few-shot action recognition. In: European Conference on Computer Vision. pp. 174-191. Springer (2022) +28. Mallya, A., Lazebnik, S.: Packnet: Adding multiple tasks to a single network by iterative pruning. In: Proceedings of the IEEE conference on Computer Vision and Pattern Recognition. pp. 7765-7773 (2018) +29. Mialon, G., Chen, D., Selosse, M., Mairal, J.: Graphit: Encoding graph structure in transformers. arXiv preprint arXiv:2106.05667 (2021) +30. Park, J.J., Florence, P., Straub, J., Newcombe, R., Lovegrove, S.: Deepsdf: Learning continuous signed distance functions for shape representation. In: Proceedings of the IEEE/CVF conference on computer vision and pattern recognition. pp. 165-174 (2019) +31. Peng, C., Zhao, K., Wang, T., Li, M., Lovell, B.C.: Few-shot class-incremental learning from an open-set perspective. In: European Conference on Computer Vision. pp. 382-397. Springer (2022) +32. Pernici, F., Bruni, M., Baecchi, C., Turchini, F., Del Bimbo, A.: Class-incremental learning with pre-allocated fixed classifiers. In: 2020 25th International Conference on Pattern Recognition (ICPR). pp. 6259–6266. IEEE (2021) +33. Radford, A., Kim, J.W., Hallacy, C., Ramesh, A., Goh, G., Agarwal, S., Sastry, G., Askell, A., Mishkin, P., Clark, J., et al.: Learning transferable visual models from natural language supervision. In: International conference on machine learning. pp. 8748-8763. PMLR (2021) + +34. Razdaibiedina, A., Mao, Y., Hou, R., Khabsa, M., Lewis, M., Almahairi, A.: Progressive prompts: Continual learning for language models. arXiv preprint arXiv:2301.12314 (2023) +35. Rebuffi, S.A., Bilen, H., Vedaldi, A.: Efficient parametrization of multi-domain deep neural networks. In: Proceedings of the IEEE conference on computer vision and pattern recognition. pp. 8119-8127 (2018) +36. Ren, B., Liu, M., Ding, R., Liu, H.: A survey on 3d skeleton-based action recognition using learning method. Cyborg and Bionic Systems (2020) +37. Ridnik, T., Ben-Baruch, E., Noy, A., Zelnik-Manor, L.: Imagenet-21k pretraining for the masses. arXiv preprint arXiv:2104.10972 (2021) +38. Rusu, A.A., Rabinowitz, N.C., Desjardins, G., Soyer, H., Kirkpatrick, J., Kavukcuoglu, K., Pascanu, R., Hadsell, R.: Progressive neural networks. arXiv preprint arXiv:1606.04671 (2016) +39. Shahroudy, A., Liu, J., Ng, T.T., Wang, G.: Ntu rgb+ d: A large scale dataset for 3d human activity analysis. In: Proceedings of the IEEE conference on computer vision and pattern recognition. pp. 1010-1019 (2016) +40. Smedt, Q.D., Wannous, H., Vandeborre, J.P., Guerry, J., Saux, B.L., Filliat, D.: 3D Hand Gesture Recognition Using a Depth and Skeletal Dataset. In: Pratikakis, I., Dupont, F., Ovsjanikov, M. (eds.) Eurographics Workshop on 3D Object Retrieval. The Eurographics Association (2017). https://doi.org/10.2312/3dor.20171049 +41. Smith, J.S., Karlinsky, L., Gutta, V., Cascante-Bonilla, P., Kim, D., Arbelle, A., Panda, R., Feris, R., Kira, Z.: Coda-prompt: Continual decomposed attention-based prompting for rehearsal-free continual learning. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 11909-11919 (2023) +42. Tang, Y.M., Peng, Y.X., Zheng, W.S.: When prompt-based incremental learning does not meet strong pretraining. arXiv preprint arXiv:2308.10445 (2023) +43. Tao, X., Hong, X., Chang, X., Dong, S., Wei, X., Gong, Y.: Few-shot class-incremental learning. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 12183-12192 (2020) +44. Van Den Oord, A., Vinyals, O., et al.: Neural discrete representation learning. Advances in neural information processing systems 30 (2017) +45. Villa, A., Alcazar, J.L., Alfarra, M., Alhamoud, K., Hurtado, J., Heilbron, F.C., Soto, A., Ghanem, B.: Pivot: Prompting for video continual learning. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 24214-24223 (2023) +46. Wang, X., Zhang, S., Qing, Z., Gao, C., Zhang, Y., Zhao, D., Sang, N.: Molo: Motion-augmented long-short contrastive learning for few-shot action recognition. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 18011-18021 (2023) +47. Wang, Y., Huang, Z., Hong, X.: S-prompts learning with pre-trained transformers: An occam's razor for domain incremental learning. Advances in Neural Information Processing Systems 35, 5682-5695 (2022) +48. Wang, Z., Zhang, Z., Ebrahimi, S., Sun, R., Zhang, H., Lee, C.Y., Ren, X., Su, G., Perot, V., Dy, J., et al.: Dualprompt: Complementary prompting for rehearsal-free continual learning. In: European Conference on Computer Vision. pp. 631-648. Springer (2022) +49. Wang, Z., Zhang, Z., Lee, C.Y., Zhang, H., Sun, R., Ren, X., Su, G., Perot, V., Dy, J., Pfister, T.: Learning to prompt for continual learning. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 139-149 (2022) + +50. Williams, W., Ringer, S., Ash, T., MacLeod, D., Dougherty, J., Hughes, J.: Hierarchical quantized autoencoders. Advances in Neural Information Processing Systems 33, 4524-4535 (2020) +51. Yang, Y., Yuan, H., Li, X., Lin, Z., Torr, P., Tao, D.: Neural collapse inspired feature-classifier alignment for few-shot class-incremental learning. In: The Eleventh International Conference on Learning Representations (2023), https://openreview.net/forum?id=y5W8tpojhtJ +52. Yue, R., Tian, Z., Du, S.: Action recognition based on rgb and skeleton data sets: A survey. Neurocomputing (2022) +53. Zhang, H.B., Zhang, Y.X., Zhong, B., Lei, Q., Yang, L., Du, J.X., Chen, D.S.: A comprehensive survey of vision-based human action recognition methods. Sensors 19(5), 1005 (2019) +54. Zheng, C., Vedaldi, A.: Online clustered codebook. In: Proceedings of the IEEE/CVF International Conference on Computer Vision. pp. 22798-22807 (2023) +55. Zhou, D.W., Wang, F.Y., Ye, H.J., Ma, L., Pu, S., Zhan, D.C.: Forward compatible few-shot class-incremental learning. In: Proceedings of the IEEE/CVF conference on computer vision and pattern recognition. pp. 9046-9056 (2022) +56. Zhu, A., Ke, Q., Gong, M., Bailey, J.: Adaptive local-component-aware graph convolutional network for one-shot skeleton-based action recognition. In: Proceedings of the IEEE/CVF Winter Conference on Applications of Computer Vision. pp. 6038-6047 (2023) +57. Zhu, B., Niu, Y., Han, Y., Wu, Y., Zhang, H.: Prompt-aligned gradient for prompt tuning. In: Proceedings of the IEEE/CVF International Conference on Computer Vision. pp. 15659-15669 (2023) \ No newline at end of file diff --git a/2024/POET_ Prompt Offset Tuning for Continual Human Action Adaptation/images.zip b/2024/POET_ Prompt Offset Tuning for Continual Human Action Adaptation/images.zip new file mode 100644 index 0000000000000000000000000000000000000000..076dc2f2cdb6700e46ae5bbcbfd93f6b7c6c27ec --- /dev/null +++ b/2024/POET_ Prompt Offset Tuning for Continual Human Action Adaptation/images.zip @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:450b8bc8d5ffc2fbfea9a0bd18513b42b3ec7c838f519f2aef48096a5a087bb0 +size 416366 diff --git a/2024/POET_ Prompt Offset Tuning for Continual Human Action Adaptation/layout.json b/2024/POET_ Prompt Offset Tuning for Continual Human Action Adaptation/layout.json new file mode 100644 index 0000000000000000000000000000000000000000..ae9756d61c61297067e9f37cd59f81ef11e6d9bf --- /dev/null +++ b/2024/POET_ Prompt Offset Tuning for Continual Human Action Adaptation/layout.json @@ -0,0 +1,11302 @@ +{ + "pdf_info": [ + { + "para_blocks": [ + { + "bbox": [ + 153, + 111, + 462, + 148 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 153, + 111, + 462, + 148 + ], + "spans": [ + { + "bbox": [ + 153, + 111, + 462, + 148 + ], + "type": "text", + "content": "POET: Prompt Offset Tuning for Continual Human Action Adaptation" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 145, + 168, + 469, + 205 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 145, + 168, + 469, + 205 + ], + "spans": [ + { + "bbox": [ + 145, + 168, + 469, + 205 + ], + "type": "text", + "content": "Prachi Garg" + }, + { + "bbox": [ + 145, + 168, + 469, + 205 + ], + "type": "inline_equation", + "content": "^{1}" + }, + { + "bbox": [ + 145, + 168, + 469, + 205 + ], + "type": "text", + "content": ", K J Joseph" + }, + { + "bbox": [ + 145, + 168, + 469, + 205 + ], + "type": "inline_equation", + "content": "^{3}" + }, + { + "bbox": [ + 145, + 168, + 469, + 205 + ], + "type": "text", + "content": ", Vineeth N Balasubramanian" + }, + { + "bbox": [ + 145, + 168, + 469, + 205 + ], + "type": "inline_equation", + "content": "^{3}" + }, + { + "bbox": [ + 145, + 168, + 469, + 205 + ], + "type": "text", + "content": ", Necati Cihan Camgoz" + }, + { + "bbox": [ + 145, + 168, + 469, + 205 + ], + "type": "inline_equation", + "content": "^{2}" + }, + { + "bbox": [ + 145, + 168, + 469, + 205 + ], + "type": "text", + "content": ", Chengde Wan" + }, + { + "bbox": [ + 145, + 168, + 469, + 205 + ], + "type": "inline_equation", + "content": "^{2}" + }, + { + "bbox": [ + 145, + 168, + 469, + 205 + ], + "type": "text", + "content": ", Kenrick Kin" + }, + { + "bbox": [ + 145, + 168, + 469, + 205 + ], + "type": "inline_equation", + "content": "^{2}" + }, + { + "bbox": [ + 145, + 168, + 469, + 205 + ], + "type": "text", + "content": ", Weiguang Si" + }, + { + "bbox": [ + 145, + 168, + 469, + 205 + ], + "type": "inline_equation", + "content": "^{2}" + }, + { + "bbox": [ + 145, + 168, + 469, + 205 + ], + "type": "text", + "content": ", Shugao Ma" + }, + { + "bbox": [ + 145, + 168, + 469, + 205 + ], + "type": "inline_equation", + "content": "^{2}" + }, + { + "bbox": [ + 145, + 168, + 469, + 205 + ], + "type": "text", + "content": ", and Fernando De La Torre" + }, + { + "bbox": [ + 145, + 168, + 469, + 205 + ], + "type": "inline_equation", + "content": "^{1}" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 233, + 213, + 380, + 226 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 233, + 213, + 380, + 226 + ], + "spans": [ + { + "bbox": [ + 233, + 213, + 380, + 226 + ], + "type": "text", + "content": "1 Carnegie Mellon University, USA" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 264, + 226, + 350, + 236 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 264, + 226, + 350, + 236 + ], + "spans": [ + { + "bbox": [ + 264, + 226, + 350, + 236 + ], + "type": "text", + "content": "Meta Reality Labs" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 216, + 236, + 398, + 247 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 216, + 236, + 398, + 247 + ], + "spans": [ + { + "bbox": [ + 216, + 236, + 398, + 247 + ], + "type": "text", + "content": "3 Indian Institute of Technology, Hyderabad" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 159, + 282, + 455, + 524 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 159, + 282, + 455, + 524 + ], + "spans": [ + { + "bbox": [ + 159, + 282, + 455, + 524 + ], + "type": "text", + "content": "Abstract. As extended reality (XR) is redefining how users interact with computing devices, research in human action recognition is gaining prominence. Typically, models deployed on immersive computing devices are static and limited to their default set of classes. The goal of our research is to provide users and developers with the capability to personalize their experience by adding new action classes to their device models continually. Importantly, a user should be able to add new classes in a low-shot and efficient manner, while this process should not require storing or replaying any of user's sensitive training data. We formalize this problem as privacy-aware few-shot continual action recognition. Towards this end, we propose POET: Prompt-offset Tuning. While existing prompt tuning approaches have shown great promise for continual learning of image, text, and video modalities; they demand access to extensively pretrained transformers. Breaking away from this assumption, POET demonstrates the efficacy of prompt tuning a significantly lightweight backbone, pretrained exclusively on the base class data. We propose a novel spatio-temporal learnable prompt offset tuning approach, and are the first to apply such prompt tuning to Graph Neural Networks. We contribute two new benchmarks for our new problem setting in human action recognition: (i) NTU RGB+D dataset for activity recognition, and (ii) SHREC-2017 dataset for hand gesture recognition. We find that POET consistently outperforms comprehensive benchmarks." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 160, + 534, + 453, + 557 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 160, + 534, + 453, + 557 + ], + "spans": [ + { + "bbox": [ + 160, + 534, + 453, + 557 + ], + "type": "text", + "content": "Keywords: 3D Skeleton Activity Recognition " + }, + { + "bbox": [ + 160, + 534, + 453, + 557 + ], + "type": "inline_equation", + "content": "\\cdot" + }, + { + "bbox": [ + 160, + 534, + 453, + 557 + ], + "type": "text", + "content": " Extended Reality (XR) " + }, + { + "bbox": [ + 160, + 534, + 453, + 557 + ], + "type": "inline_equation", + "content": "\\cdot" + }, + { + "bbox": [ + 160, + 534, + 453, + 557 + ], + "type": "text", + "content": " Continual Learning " + }, + { + "bbox": [ + 160, + 534, + 453, + 557 + ], + "type": "inline_equation", + "content": "\\cdot" + }, + { + "bbox": [ + 160, + 534, + 453, + 557 + ], + "type": "text", + "content": " Prompt Tuning." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 132, + 581, + 230, + 594 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 581, + 230, + 594 + ], + "spans": [ + { + "bbox": [ + 132, + 581, + 230, + 594 + ], + "type": "text", + "content": "1 Introduction" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 132, + 609, + 482, + 633 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 609, + 482, + 633 + ], + "spans": [ + { + "bbox": [ + 132, + 609, + 482, + 633 + ], + "type": "text", + "content": "A key input modality to virtual, augmented and mixed reality (often together termed as extended reality, XR) devices today is through recognizing human" + } + ] + } + ], + "index": 8 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 133, + 642, + 481, + 665 + ], + "type": "page_footnote", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 642, + 481, + 665 + ], + "spans": [ + { + "bbox": [ + 133, + 642, + 481, + 665 + ], + "type": "text", + "content": "Source Code at https://github.com/humansensinglab/POET-continual-action-recognition" + } + ] + } + ], + "index": 9 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 0 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 136, + 114, + 479, + 220 + ], + "blocks": [ + { + "bbox": [ + 136, + 114, + 479, + 220 + ], + "lines": [ + { + "bbox": [ + 136, + 114, + 479, + 220 + ], + "spans": [ + { + "bbox": [ + 136, + 114, + 479, + 220 + ], + "type": "image", + "image_path": "30acfb9b2daf18ec6aef71a76a4b9c3a515194d19662e3e8544a9a71a9e5b0cf.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 130, + 228, + 482, + 284 + ], + "lines": [ + { + "bbox": [ + 130, + 228, + 482, + 284 + ], + "spans": [ + { + "bbox": [ + 130, + 228, + 482, + 284 + ], + "type": "text", + "content": "Fig. 1: Proposed POET method continually adapts skeleton-based human action recognition models pretrained on a pre-defined set of categories to new user categories with few training examples. Users can thus expand the capabilities of XR systems with novel action classes by providing a few examples of each new class. We discard the user-sensitive data as soon as the model is updated on the new categories." + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_caption" + } + ], + "index": 2 + }, + { + "bbox": [ + 130, + 292, + 482, + 411 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 292, + 482, + 411 + ], + "spans": [ + { + "bbox": [ + 130, + 292, + 482, + 411 + ], + "type": "text", + "content": "activity and hand gestures based on body and hand pose estimates. Recognizing human actions5 facilitates seamless user interactions in head-mounted XR devices such as the Meta Quest 3 and Apple Vision Pro. If the provided action recognition models are static, then developers and users are limited to a predefined set of action categories. With the growing use of such devices in new contexts and the increasing demand for personalized technology delivery, there is an impending need to enable the action recognition models in such systems to adapt and learn new user actions over time. Defining their own action categories allows users to customize their experience and expand the functionality of their XR devices. Addressing this need is the primary objective of this work." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 130, + 412, + 482, + 543 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 412, + 482, + 543 + ], + "spans": [ + { + "bbox": [ + 130, + 412, + 482, + 543 + ], + "type": "text", + "content": "Adapting human action models to new user categories over time faces a few challenges. Firstly, the model must be capable of learning new actions with minimal amount of training data so users can add new classes by providing just a few training examples per class. Secondly, due to the increasing use of XR devices for personal assistance, there is a need for privacy preservation in user action recognition-based pipelines [2, 14]. Hence, the adaptation of such action recognition models to new user categories must also be 'data-free', i.e., it cannot store and replay previously seen user training data in subsequent continual sessions. Considering these requirements, we leverage the recent success of 'data-free' prompt-based learning [49] and propose a new spatio-temporal prompt offset tuning approach to efficiently adapt the default model without finetuning." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 130, + 544, + 482, + 639 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 544, + 482, + 639 + ], + "spans": [ + { + "bbox": [ + 130, + 544, + 482, + 639 + ], + "type": "text", + "content": "Human action recognition systems are moving to skeleton-based approaches, especially in applications that require low-shot action recognition capabilities such as medical action recognition [27, 56]. Skeletons offer a robust and compact alternative to videos in such low-shot regimes, due to their relatively low dimensionality and lesser variance under background conditions. While there have been a wide variety of efforts in skeleton-based human action recognition over the years [36, 52, 53], there have been fewer efforts on adapting such models to newer user categories. Efforts like [1, 22] attempted to continually learn new user" + } + ] + } + ], + "index": 6 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "spans": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "type": "text", + "content": "2" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 166, + 91, + 222, + 102 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 166, + 91, + 222, + 102 + ], + "spans": [ + { + "bbox": [ + 166, + 91, + 222, + 102 + ], + "type": "text", + "content": "P. Garg et al." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 133, + 646, + 482, + 665 + ], + "type": "page_footnote", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 646, + 482, + 665 + ], + "spans": [ + { + "bbox": [ + 133, + 646, + 482, + 665 + ], + "type": "text", + "content": "5 We use human action as an umbrella term for both hand gesture and body activity in this work for ease of presentation." + } + ] + } + ], + "index": 7 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 1 + }, + { + "para_blocks": [ + { + "bbox": [ + 130, + 116, + 482, + 224 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 116, + 482, + 224 + ], + "spans": [ + { + "bbox": [ + 130, + 116, + 482, + 224 + ], + "type": "text", + "content": "categories over time in skeleton-based human action recognition, but relied on fully-supervised data for the new classes. On the other hand, few-shot learning works [27, 46, 56] adapt a pretrained skeleton-based action recognition model to new data, but without explicitly retaining past categories. In this work, we seek to learn new user categories in trained human action models with very few labeled samples for the new classes, while being data-free (not storing samples from previously trained categories). Fig 1 summarizes our overall objective. One could view our setting as privacy-aware few-shot continual learning for skeleton-based action recognition." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 130, + 224, + 482, + 356 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 224, + 482, + 356 + ], + "spans": [ + { + "bbox": [ + 130, + 224, + 482, + 356 + ], + "type": "text", + "content": "To this end, we propose a prompt offset tuning methodology that can be integrated with existing backbone architectures for skeleton-based human action recognition. Our learnable (soft) prompts are selected from a shared knowledge pool of prompts based on an input instance dependent attention mechanism. In particular, we propose prompt selection using an ordered query-key matching that enables a temporal prompt frame order selection consistent with the input instance. We show that such an approach allows us to learn new user categories without having to store data from past classes, without overwriting the pre-existing categories. To the best of our knowledge, this is the first effort on leveraging prompt tuning for skeleton-based models, as well as on spatiotemporal prompt selection and tuning." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 131, + 358, + 334, + 370 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 131, + 358, + 334, + 370 + ], + "spans": [ + { + "bbox": [ + 131, + 358, + 334, + 370 + ], + "type": "text", + "content": "Our key contributions are summarized below:" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 130, + 375, + 479, + 396 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 375, + 479, + 396 + ], + "spans": [ + { + "bbox": [ + 130, + 375, + 479, + 396 + ], + "type": "text", + "content": "- We formalize a novel problem setting which continually adapts human action models to new user categories over time, in a privacy aware manner." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 130, + 399, + 479, + 443 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 399, + 479, + 443 + ], + "spans": [ + { + "bbox": [ + 130, + 399, + 479, + 443 + ], + "type": "text", + "content": "- To address this problem, we propose a novel spatio-temporal Prompt OffsEt Tuning methodology (POET). In particular, it is designed to seamlessly plug-and-play with a pretrained model's input embedding, without any significant architectural changes." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 130, + 445, + 479, + 468 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 445, + 479, + 468 + ], + "spans": [ + { + "bbox": [ + 130, + 445, + 479, + 468 + ], + "type": "text", + "content": "- Our comprehensive experimental evaluation on two benchmark datasets brings out the efficacy of our proposed approach." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 131, + 487, + 243, + 500 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 131, + 487, + 243, + 500 + ], + "spans": [ + { + "bbox": [ + 131, + 487, + 243, + 500 + ], + "type": "text", + "content": "2 Related Works" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 131, + 514, + 240, + 526 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 131, + 514, + 240, + 526 + ], + "spans": [ + { + "bbox": [ + 131, + 514, + 240, + 526 + ], + "type": "text", + "content": "2.1 Prompt Tuning" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 130, + 533, + 482, + 666 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 533, + 482, + 666 + ], + "spans": [ + { + "bbox": [ + 130, + 533, + 482, + 666 + ], + "type": "text", + "content": "The idea of prompting, as it originated from Large Language Models (LLMs), is to include additional information, known as a text prompt, to condition the model's input for generating an output relevant to the prompt. Instead of applying a discrete, pre-defined 'hard' language prompt token, prompt and prefix tuning [20,23] formalized the concept of applying 'soft prompts' to the input. A set of learnable parameters are prepended (concatenated) to the input text and trained along with the classifier while keeping the backbone parameters frozen. Similar to prompt tuning of LLMs, recent works have popularized prompt tuning of ViTs [16] as an effective way of adapting large pretrained models to downstream tasks [49,57]. However, it remains unexplored and undefined (to the best of our knowledge) for non-transformer architectures such as GNNs." + } + ] + } + ], + "index": 10 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 156, + 91, + 447, + 102 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 156, + 91, + 447, + 102 + ], + "spans": [ + { + "bbox": [ + 156, + 91, + 447, + 102 + ], + "type": "text", + "content": "POET: Prompt Offset Tuning for Continual Human Action Adaptation" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 474, + 92, + 481, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 474, + 92, + 481, + 100 + ], + "spans": [ + { + "bbox": [ + 474, + 92, + 481, + 100 + ], + "type": "text", + "content": "3" + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 2 + }, + { + "para_blocks": [ + { + "bbox": [ + 132, + 116, + 357, + 129 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 116, + 357, + 129 + ], + "spans": [ + { + "bbox": [ + 132, + 116, + 357, + 129 + ], + "type": "text", + "content": "2.2 Prompt Tuning for Continual Learning" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 130, + 140, + 482, + 307 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 140, + 482, + 307 + ], + "spans": [ + { + "bbox": [ + 130, + 140, + 482, + 307 + ], + "type": "text", + "content": "Prompt tuning provides a simple and cost-effective way of learning task-specific signal condensed into 'soft prompts'. For continual learning, training a set of prompts for each sequential task provides a natural alternative to storing privacy violating exemplars and replaying them. Training task-specific prompts for each sequential task is straightforward when authors assume access to task identity at both train and inference time, like in Progressive Prompts [34]. However, if task identity is unavailable at inference, the model will not know which task's prompts or classifier to use for evaluating a test sample. In this respect, S-prompts [47] and A-la-carte prompt tuning (APT) [4] learn an independent set of prompts for each domain/task and employ a KNN-based search for domain/task identity at test time. Since these methods learn stand-alone prompts for every task, the prompt feature space is task-specific, and there is no forgetting of old knowledge when learning new tasks (by design). At the same time however, these 'no forgetting' prompts cannot share knowledge across tasks." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 130, + 308, + 482, + 463 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 308, + 482, + 463 + ], + "spans": [ + { + "bbox": [ + 130, + 308, + 482, + 463 + ], + "type": "text", + "content": "This leads to another ideology for continual prompt tuning, i.e., treat each prompt unit as being a part of a larger shared (knowledge) pool of prompts. Then the desired number of prompt units can be selected from the pool, conditioned on the input instance itself [41, 48, 49]. Given the scarcity of new data in our setting, we hypothesize that sharing of knowledge will benefit new tasks and draw inspiration from this line of works. Most recently, Adaptive Prompt Generator (APG) [42] challenges the intensive ImageNet21K pre-training assumption as it prompts a ViT pretrained only on the continual benchmark's base class data (similar to us). However, they use replay and knowledge distillation-style 'antiforgetting learning', in addition to using prompts. Even though our backbone is trained only on the base classes, we propose a simple prompt tuning-only strategy to counter forgetting. This implies that a prompt strategy is all we need to continually add new action semantics in a few-shot manner." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 132, + 483, + 350, + 496 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 483, + 350, + 496 + ], + "spans": [ + { + "bbox": [ + 132, + 483, + 350, + 496 + ], + "type": "text", + "content": "2.3 Few-Shot Class Incremental Learning" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 130, + 507, + 482, + 675 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 507, + 482, + 675 + ], + "spans": [ + { + "bbox": [ + 130, + 507, + 482, + 675 + ], + "type": "text", + "content": "FSCIL is a challenging continual learning setting where a model overfits to new classes, with the simultaneous heightened (often complete) forgetting of old knowledge as soon as the base model is fine-tuned on few-shot data [10,43]. Since the backbone feature extractor is the only source of previously seen knowledge, if it is updated, knowledge is lost forever. Typically, existing works decouple the learning of (backbone) feature representations from the classifier by learning the model only on the base data and relying on non-parametric class-mean classifiers for classification in subsequent steps [13,31,55]. This leads to a feature-classifier misalignment issue [32, 51] because new class prototypes are extracted from a backbone representation trained only on the base classes. We hypothesize that optimizing input prompt vectors along with a dynamically expanding parametric classifier on top of a frozen backbone can alleviate this misalignment issue. Our work not only provides a fresh perspective into FSCIL, but to our best knowledge is also the only work not designed for and evaluated on image benchmarks." + } + ] + } + ], + "index": 6 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "spans": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "type": "text", + "content": "4" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 166, + 91, + 222, + 102 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 166, + 91, + 222, + 102 + ], + "spans": [ + { + "bbox": [ + 166, + 91, + 222, + 102 + ], + "type": "text", + "content": "P. Garg et al." + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 3 + }, + { + "para_blocks": [ + { + "bbox": [ + 132, + 114, + 234, + 127 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 114, + 234, + 127 + ], + "spans": [ + { + "bbox": [ + 132, + 114, + 234, + 127 + ], + "type": "text", + "content": "3 Preliminaries" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 130, + 137, + 482, + 305 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 137, + 482, + 305 + ], + "spans": [ + { + "bbox": [ + 130, + 137, + 482, + 305 + ], + "type": "text", + "content": "Skeleton Action Recognition Using Graph Representations. Our input " + }, + { + "bbox": [ + 130, + 137, + 482, + 305 + ], + "type": "inline_equation", + "content": "\\mathbf{X} \\in \\mathbb{R}^{T \\times J \\times 3}" + }, + { + "bbox": [ + 130, + 137, + 482, + 305 + ], + "type": "text", + "content": " is a video sequence of " + }, + { + "bbox": [ + 130, + 137, + 482, + 305 + ], + "type": "inline_equation", + "content": "T" + }, + { + "bbox": [ + 130, + 137, + 482, + 305 + ], + "type": "text", + "content": " frames, each frame containing " + }, + { + "bbox": [ + 130, + 137, + 482, + 305 + ], + "type": "inline_equation", + "content": "J" + }, + { + "bbox": [ + 130, + 137, + 482, + 305 + ], + "type": "text", + "content": " joints of the human body (25 joints) or hand skeleton (22 joints) in 3D Cartesian coordinate system. Such a skeleton action sequence is naturally represented as a graph topology " + }, + { + "bbox": [ + 130, + 137, + 482, + 305 + ], + "type": "inline_equation", + "content": "G = \\{\\mathcal{V}, \\mathcal{E}\\}" + }, + { + "bbox": [ + 130, + 137, + 482, + 305 + ], + "type": "text", + "content": " with " + }, + { + "bbox": [ + 130, + 137, + 482, + 305 + ], + "type": "inline_equation", + "content": "\\mathcal{V}" + }, + { + "bbox": [ + 130, + 137, + 482, + 305 + ], + "type": "text", + "content": " vertices and " + }, + { + "bbox": [ + 130, + 137, + 482, + 305 + ], + "type": "inline_equation", + "content": "\\mathcal{E}" + }, + { + "bbox": [ + 130, + 137, + 482, + 305 + ], + "type": "text", + "content": " edges. Graphs are modeled using Graph Neural Networks (GNNs) [12], which can either be sparse graph convolutional networks (GCN) or fully connected graph transformers (GT). Our main model (a GNN) is defined as " + }, + { + "bbox": [ + 130, + 137, + 482, + 305 + ], + "type": "inline_equation", + "content": "f(\\mathbf{X}) = f_c \\circ f_g \\circ f_e(\\mathbf{X})" + }, + { + "bbox": [ + 130, + 137, + 482, + 305 + ], + "type": "text", + "content": " (as also shown in Fig. 2). Input " + }, + { + "bbox": [ + 130, + 137, + 482, + 305 + ], + "type": "inline_equation", + "content": "\\mathbf{X}" + }, + { + "bbox": [ + 130, + 137, + 482, + 305 + ], + "type": "text", + "content": " is first passed through an input embedding layer " + }, + { + "bbox": [ + 130, + 137, + 482, + 305 + ], + "type": "inline_equation", + "content": "f_e" + }, + { + "bbox": [ + 130, + 137, + 482, + 305 + ], + "type": "text", + "content": " to get an embedding of human joints " + }, + { + "bbox": [ + 130, + 137, + 482, + 305 + ], + "type": "inline_equation", + "content": "\\mathbf{X_e} = f_e(\\mathbf{X}), \\mathbf{X_e} \\in \\mathbb{R}^{T \\times J \\times C_e}" + }, + { + "bbox": [ + 130, + 137, + 482, + 305 + ], + "type": "text", + "content": ", with feature dimension " + }, + { + "bbox": [ + 130, + 137, + 482, + 305 + ], + "type": "inline_equation", + "content": "C_e" + }, + { + "bbox": [ + 130, + 137, + 482, + 305 + ], + "type": "text", + "content": ". " + }, + { + "bbox": [ + 130, + 137, + 482, + 305 + ], + "type": "inline_equation", + "content": "\\mathbf{X_e}" + }, + { + "bbox": [ + 130, + 137, + 482, + 305 + ], + "type": "text", + "content": " is further passed to a graph feature extractor " + }, + { + "bbox": [ + 130, + 137, + 482, + 305 + ], + "type": "inline_equation", + "content": "f_g" + }, + { + "bbox": [ + 130, + 137, + 482, + 305 + ], + "type": "text", + "content": " composed of a stack of convolutional layers (in GCNs) or attention layers (in GTs), and finally a classifier " + }, + { + "bbox": [ + 130, + 137, + 482, + 305 + ], + "type": "inline_equation", + "content": "f_c" + }, + { + "bbox": [ + 130, + 137, + 482, + 305 + ], + "type": "text", + "content": " which predicts the action class label " + }, + { + "bbox": [ + 130, + 137, + 482, + 305 + ], + "type": "inline_equation", + "content": "\\mathbf{y}" + }, + { + "bbox": [ + 130, + 137, + 482, + 305 + ], + "type": "text", + "content": ". In POET, we propose to attach learnable parameters " + }, + { + "bbox": [ + 130, + 137, + 482, + 305 + ], + "type": "inline_equation", + "content": "\\mathbf{P_T}" + }, + { + "bbox": [ + 130, + 137, + 482, + 305 + ], + "type": "text", + "content": " (called prompt offsets) to the embedding " + }, + { + "bbox": [ + 130, + 137, + 482, + 305 + ], + "type": "inline_equation", + "content": "\\mathbf{X_e}" + }, + { + "bbox": [ + 130, + 137, + 482, + 305 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 130, + 308, + 482, + 544 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 308, + 482, + 544 + ], + "spans": [ + { + "bbox": [ + 130, + 308, + 482, + 544 + ], + "type": "text", + "content": "Problem Definition. Given a default (pre)trained model deployed on a user's device, we would like to extend this model to new action classes over " + }, + { + "bbox": [ + 130, + 308, + 482, + 544 + ], + "type": "inline_equation", + "content": "T" + }, + { + "bbox": [ + 130, + 308, + 482, + 544 + ], + "type": "text", + "content": " subsequent user sessions (also called tasks) " + }, + { + "bbox": [ + 130, + 308, + 482, + 544 + ], + "type": "inline_equation", + "content": "\\{\\mathcal{US}^{(1)},\\dots,\\mathcal{US}^{(T)}\\}^6" + }, + { + "bbox": [ + 130, + 308, + 482, + 544 + ], + "type": "text", + "content": ". In each user session " + }, + { + "bbox": [ + 130, + 308, + 482, + 544 + ], + "type": "inline_equation", + "content": "\\mathcal{US}^{(t)}" + }, + { + "bbox": [ + 130, + 308, + 482, + 544 + ], + "type": "text", + "content": ", the model learns a dataset " + }, + { + "bbox": [ + 130, + 308, + 482, + 544 + ], + "type": "inline_equation", + "content": "\\mathcal{D}^{(t)} = (\\mathbf{X}_i^t,\\mathbf{y}_i^t)_{i = 1}^{|D^{(t)}|}" + }, + { + "bbox": [ + 130, + 308, + 482, + 544 + ], + "type": "text", + "content": " of skeleton action sequence and label pairs provided by the user, " + }, + { + "bbox": [ + 130, + 308, + 482, + 544 + ], + "type": "inline_equation", + "content": "\\mathbf{X}_i^t\\in \\mathbb{R}^{T\\times J\\times 3}" + }, + { + "bbox": [ + 130, + 308, + 482, + 544 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 130, + 308, + 482, + 544 + ], + "type": "inline_equation", + "content": "y_i^t\\in \\mathbb{R}^{y^{(t)}}" + }, + { + "bbox": [ + 130, + 308, + 482, + 544 + ], + "type": "text", + "content": ". In each session, the user typically provides a few training instances " + }, + { + "bbox": [ + 130, + 308, + 482, + 544 + ], + "type": "inline_equation", + "content": "F" + }, + { + "bbox": [ + 130, + 308, + 482, + 544 + ], + "type": "text", + "content": " (e.g. " + }, + { + "bbox": [ + 130, + 308, + 482, + 544 + ], + "type": "inline_equation", + "content": "F\\leq 5" + }, + { + "bbox": [ + 130, + 308, + 482, + 544 + ], + "type": "text", + "content": ") for each of the " + }, + { + "bbox": [ + 130, + 308, + 482, + 544 + ], + "type": "inline_equation", + "content": "N" + }, + { + "bbox": [ + 130, + 308, + 482, + 544 + ], + "type": "text", + "content": " new classes being added, such that " + }, + { + "bbox": [ + 130, + 308, + 482, + 544 + ], + "type": "inline_equation", + "content": "|\\mathcal{D}^{(t)}| = NF" + }, + { + "bbox": [ + 130, + 308, + 482, + 544 + ], + "type": "text", + "content": ". The base (default) model's session " + }, + { + "bbox": [ + 130, + 308, + 482, + 544 + ], + "type": "inline_equation", + "content": "\\mathcal{UB}^{(0)}" + }, + { + "bbox": [ + 130, + 308, + 482, + 544 + ], + "type": "text", + "content": " is assumed to have a large number of default action classes " + }, + { + "bbox": [ + 130, + 308, + 482, + 544 + ], + "type": "inline_equation", + "content": "\\mathcal{V}^{(0)}" + }, + { + "bbox": [ + 130, + 308, + 482, + 544 + ], + "type": "text", + "content": " trained on sufficient data " + }, + { + "bbox": [ + 130, + 308, + 482, + 544 + ], + "type": "inline_equation", + "content": "\\mathcal{D}^{(0)}" + }, + { + "bbox": [ + 130, + 308, + 482, + 544 + ], + "type": "text", + "content": ", which is most often proprietary and cannot be accessed in later user sessions. In each session, the user adds new action classes such that, " + }, + { + "bbox": [ + 130, + 308, + 482, + 544 + ], + "type": "inline_equation", + "content": "\\mathcal{V}^{(t)}\\cap \\mathcal{V}^{(t')} = \\emptyset ,\\forall t\\neq t'^7" + }, + { + "bbox": [ + 130, + 308, + 482, + 544 + ], + "type": "text", + "content": ". Due to the aforementioned privacy constraints, in any training session " + }, + { + "bbox": [ + 130, + 308, + 482, + 544 + ], + "type": "inline_equation", + "content": "\\mathcal{US}^{(t)}" + }, + { + "bbox": [ + 130, + 308, + 482, + 544 + ], + "type": "text", + "content": ", the model has access to only " + }, + { + "bbox": [ + 130, + 308, + 482, + 544 + ], + "type": "inline_equation", + "content": "\\mathcal{D}^{(t)}" + }, + { + "bbox": [ + 130, + 308, + 482, + 544 + ], + "type": "text", + "content": "; after training this data is made inaccessible for use in subsequent sessions (no exemplar or prototypes stored). After training on every new session " + }, + { + "bbox": [ + 130, + 308, + 482, + 544 + ], + "type": "inline_equation", + "content": "\\mathcal{US}^{(t)}" + }, + { + "bbox": [ + 130, + 308, + 482, + 544 + ], + "type": "text", + "content": ", the model is evaluated on the test set of all classes seen so far " + }, + { + "bbox": [ + 130, + 308, + 482, + 544 + ], + "type": "inline_equation", + "content": "\\cup_{i = 0}^{t}\\mathcal{V}^{(i)}" + }, + { + "bbox": [ + 130, + 308, + 482, + 544 + ], + "type": "text", + "content": ". The challenge is to alleviate forgetting of old classes while not overfitting to the user-provided new class samples. One could view our setting as privacy-aware few-shot continual action recognition, a problem of practical relevance in human action recognition - which has not received adequate attention." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 132, + 557, + 424, + 572 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 557, + 424, + 572 + ], + "spans": [ + { + "bbox": [ + 132, + 557, + 424, + 572 + ], + "type": "text", + "content": "4 Methodology: Prompt Offset Tuning (POET)" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 130, + 578, + 482, + 628 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 578, + 482, + 628 + ], + "spans": [ + { + "bbox": [ + 130, + 578, + 482, + 628 + ], + "type": "text", + "content": "Overview. We propose to prompt tune a base GNN model " + }, + { + "bbox": [ + 130, + 578, + 482, + 628 + ], + "type": "inline_equation", + "content": "f(\\cdot)" + }, + { + "bbox": [ + 130, + 578, + 482, + 628 + ], + "type": "text", + "content": " by prompts " + }, + { + "bbox": [ + 130, + 578, + 482, + 628 + ], + "type": "inline_equation", + "content": "\\mathbf{P_T}" + }, + { + "bbox": [ + 130, + 578, + 482, + 628 + ], + "type": "text", + "content": " to address our overall objective. As shown in Fig. 3, for each input instance " + }, + { + "bbox": [ + 130, + 578, + 482, + 628 + ], + "type": "inline_equation", + "content": "\\mathbf{X}" + }, + { + "bbox": [ + 130, + 578, + 482, + 628 + ], + "type": "text", + "content": ", corresponding prompts " + }, + { + "bbox": [ + 130, + 578, + 482, + 628 + ], + "type": "inline_equation", + "content": "\\mathbf{P_T}" + }, + { + "bbox": [ + 130, + 578, + 482, + 628 + ], + "type": "text", + "content": " are selected from a pool of prompt parameters, using an input-dependent query and key attention mechanism. The selected prompts" + } + ] + } + ], + "index": 6 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 155, + 90, + 448, + 103 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 155, + 90, + 448, + 103 + ], + "spans": [ + { + "bbox": [ + 155, + 90, + 448, + 103 + ], + "type": "text", + "content": "POET: Prompt Offset Tuning for Continual Human Action Adaptation" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 474, + 91, + 481, + 101 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 474, + 91, + 481, + 101 + ], + "spans": [ + { + "bbox": [ + 474, + 91, + 481, + 101 + ], + "type": "text", + "content": "5" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 133, + 636, + 340, + 646 + ], + "type": "page_footnote", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 636, + 340, + 646 + ], + "spans": [ + { + "bbox": [ + 133, + 636, + 340, + 646 + ], + "type": "text", + "content": "6 User sessions may be spaced at arbitrary time intervals." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 135, + 647, + 481, + 665 + ], + "type": "page_footnote", + "angle": 0, + "lines": [ + { + "bbox": [ + 135, + 647, + 481, + 665 + ], + "spans": [ + { + "bbox": [ + 135, + 647, + 481, + 665 + ], + "type": "text", + "content": "We make this assumption considering this is a first of such efforts; allowing for overlapping action classes and users to 'update' older classes would be interesting extensions of our proposed work." + } + ] + } + ], + "index": 8 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 4 + }, + { + "para_blocks": [ + { + "bbox": [ + 130, + 116, + 479, + 140 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 116, + 479, + 140 + ], + "spans": [ + { + "bbox": [ + 130, + 116, + 479, + 140 + ], + "type": "text", + "content": "are added to the input feature embedding (and hence the term 'prompt offsets'), before forwarding to the feature extractor and classifier (shown in Fig. 2)." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 130, + 140, + 351, + 261 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 140, + 351, + 261 + ], + "spans": [ + { + "bbox": [ + 130, + 140, + 351, + 261 + ], + "type": "text", + "content": "To this end, our method, POET uses the same number of prompts as the number of temporal frames in the input, to maintain temporal consistency between the prompt and the input. Focusing solely on prompt offsets allows us to adapt the model to subsequent user sessions without having to update the input embeddings or the feature extraction backbone. Our prompt selection mechanism is learnable and trained along with the classifier to make this method simple and efficient." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 130, + 264, + 351, + 384 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 264, + 351, + 384 + ], + "spans": [ + { + "bbox": [ + 130, + 264, + 351, + 384 + ], + "type": "text", + "content": "What are Prompt Offsets? Learnable (or soft [20]) prompts are parameter vectors in a continuous space which are optimized to adapt the pretrained frozen backbone " + }, + { + "bbox": [ + 130, + 264, + 351, + 384 + ], + "type": "inline_equation", + "content": "f_{g}" + }, + { + "bbox": [ + 130, + 264, + 351, + 384 + ], + "type": "text", + "content": " to each continual task. We define our spatio-temporal prompt offsets " + }, + { + "bbox": [ + 130, + 264, + 351, + 384 + ], + "type": "inline_equation", + "content": "\\mathbf{P_T}" + }, + { + "bbox": [ + 130, + 264, + 351, + 384 + ], + "type": "text", + "content": " as a set of " + }, + { + "bbox": [ + 130, + 264, + 351, + 384 + ], + "type": "inline_equation", + "content": "T" + }, + { + "bbox": [ + 130, + 264, + 351, + 384 + ], + "type": "text", + "content": " prompts (same in number as skeletal frames in input), each prompt " + }, + { + "bbox": [ + 130, + 264, + 351, + 384 + ], + "type": "inline_equation", + "content": "P_{i}" + }, + { + "bbox": [ + 130, + 264, + 351, + 384 + ], + "type": "text", + "content": " having length equal to the number of joints in a frame " + }, + { + "bbox": [ + 130, + 264, + 351, + 384 + ], + "type": "inline_equation", + "content": "J" + }, + { + "bbox": [ + 130, + 264, + 351, + 384 + ], + "type": "text", + "content": " and feature dimension same as input feature embedding " + }, + { + "bbox": [ + 130, + 264, + 351, + 384 + ], + "type": "inline_equation", + "content": "\\mathbf{X_e}" + }, + { + "bbox": [ + 130, + 264, + 351, + 384 + ], + "type": "text", + "content": ", i.e., " + }, + { + "bbox": [ + 130, + 264, + 351, + 384 + ], + "type": "inline_equation", + "content": "P_{i} \\in \\mathbb{R}^{J \\times C_{e}}" + }, + { + "bbox": [ + 130, + 264, + 351, + 384 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 130, + 384, + 351, + 432 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 384, + 351, + 432 + ], + "spans": [ + { + "bbox": [ + 130, + 384, + 351, + 432 + ], + "type": "text", + "content": "Existing prompt tuning efforts, for example in image classification, focus on concatenating learnable prompts to the input token sequence in transformer architectures [16, 20]. Even though trans" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 130, + 432, + 482, + 552 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 432, + 482, + 552 + ], + "spans": [ + { + "bbox": [ + 130, + 432, + 482, + 552 + ], + "type": "text", + "content": "formers can be generalized to graphs [7, 11, 29], it is non-trivial to attach prompts to a GNN. This is because transformers can be viewed as treating sentences or images as fully connected graphs where any word (or image patch) can attend to any other word in the sentence [12]. However, our input is a spatio-temporal graph skeleton of the human joint-bone structure with its own edge connectivity. Concatenating prompts along spatial or temporal dimensions would affect the graph semantics, and also affect standard training strategies such as a forward pass or backpropagation (especially in GCNs). Hence, we attach the selected prompts " + }, + { + "bbox": [ + 130, + 432, + 482, + 552 + ], + "type": "inline_equation", + "content": "\\mathbf{P_T}" + }, + { + "bbox": [ + 130, + 432, + 482, + 552 + ], + "type": "text", + "content": " to the corresponding input feature embedding " + }, + { + "bbox": [ + 130, + 432, + 482, + 552 + ], + "type": "inline_equation", + "content": "\\mathbf{X_e}" + }, + { + "bbox": [ + 130, + 432, + 482, + 552 + ], + "type": "text", + "content": " via a prompt attachment operator " + }, + { + "bbox": [ + 130, + 432, + 482, + 552 + ], + "type": "inline_equation", + "content": "f_{p}(.)" + }, + { + "bbox": [ + 130, + 432, + 482, + 552 + ], + "type": "text", + "content": ". The class logit distribution " + }, + { + "bbox": [ + 130, + 432, + 482, + 552 + ], + "type": "inline_equation", + "content": "\\mathbf{y}" + }, + { + "bbox": [ + 130, + 432, + 482, + 552 + ], + "type": "text", + "content": " is thus obtained as:" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 217, + 556, + 481, + 571 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 217, + 556, + 481, + 571 + ], + "spans": [ + { + "bbox": [ + 217, + 556, + 481, + 571 + ], + "type": "interline_equation", + "content": "\\mathbf {y} = f (\\mathbf {X}, \\mathbf {P} _ {\\mathbf {T}}) = f _ {c} \\circ f _ {g} \\circ f _ {p} \\left(f _ {e} (\\mathbf {X}), \\mathbf {P} _ {\\mathbf {T}}\\right) \\tag {1}", + "image_path": "045f1fc724cd352ab8650e4ba723013a3eb0502f4ab400609349c73bae946c41.jpg" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 130, + 575, + 482, + 638 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 575, + 482, + 638 + ], + "spans": [ + { + "bbox": [ + 130, + 575, + 482, + 638 + ], + "type": "text", + "content": "In every user session " + }, + { + "bbox": [ + 130, + 575, + 482, + 638 + ], + "type": "inline_equation", + "content": "t > 0" + }, + { + "bbox": [ + 130, + 575, + 482, + 638 + ], + "type": "text", + "content": ", the classifier output dimension expands by " + }, + { + "bbox": [ + 130, + 575, + 482, + 638 + ], + "type": "inline_equation", + "content": "N" + }, + { + "bbox": [ + 130, + 575, + 482, + 638 + ], + "type": "text", + "content": " to accommodate the new action classes. Unlike most existing continual prompt tuning works, our feature extractor backbone " + }, + { + "bbox": [ + 130, + 575, + 482, + 638 + ], + "type": "inline_equation", + "content": "f_{g}" + }, + { + "bbox": [ + 130, + 575, + 482, + 638 + ], + "type": "text", + "content": " is trained only on the base class data " + }, + { + "bbox": [ + 130, + 575, + 482, + 638 + ], + "type": "inline_equation", + "content": "\\mathcal{D}^{(0)}" + }, + { + "bbox": [ + 130, + 575, + 482, + 638 + ], + "type": "text", + "content": " and is never fine-tuned on classes from new user sessions " + }, + { + "bbox": [ + 130, + 575, + 482, + 638 + ], + "type": "inline_equation", + "content": "\\mathcal{U}\\mathcal{S}^{(t)}, t > 0" + }, + { + "bbox": [ + 130, + 575, + 482, + 638 + ], + "type": "text", + "content": ". After the base session training, parameters of " + }, + { + "bbox": [ + 130, + 575, + 482, + 638 + ], + "type": "inline_equation", + "content": "f_{g}, f_{e}" + }, + { + "bbox": [ + 130, + 575, + 482, + 638 + ], + "type": "text", + "content": " are frozen." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 130, + 641, + 482, + 666 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 641, + 482, + 666 + ], + "spans": [ + { + "bbox": [ + 130, + 641, + 482, + 666 + ], + "type": "text", + "content": "Prompt Pool Design. As stated in Sec. 2.2, to encourage knowledge sharing across user sessions, we choose to construct a single prompt pool " + }, + { + "bbox": [ + 130, + 641, + 482, + 666 + ], + "type": "inline_equation", + "content": "\\mathbf{P}" + }, + { + "bbox": [ + 130, + 641, + 482, + 666 + ], + "type": "text", + "content": " which encodes" + } + ] + } + ], + "index": 9 + }, + { + "type": "image", + "bbox": [ + 360, + 155, + 478, + 326 + ], + "blocks": [ + { + "bbox": [ + 360, + 155, + 478, + 326 + ], + "lines": [ + { + "bbox": [ + 360, + 155, + 478, + 326 + ], + "spans": [ + { + "bbox": [ + 360, + 155, + 478, + 326 + ], + "type": "image", + "image_path": "a2ff12afd4f1ec19e137cbe1431a41753ef0ef8f4958409d3b5efa49301b4fb8.jpg" + } + ] + } + ], + "index": 10, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 356, + 327, + 482, + 426 + ], + "lines": [ + { + "bbox": [ + 356, + 327, + 482, + 426 + ], + "spans": [ + { + "bbox": [ + 356, + 327, + 482, + 426 + ], + "type": "text", + "content": "Fig.2: POET: Prompt-offset Tuning proposes to offset the input feature embedding " + }, + { + "bbox": [ + 356, + 327, + 482, + 426 + ], + "type": "inline_equation", + "content": "\\mathbf{X}_{\\mathrm{e}}" + }, + { + "bbox": [ + 356, + 327, + 482, + 426 + ], + "type": "text", + "content": " of the main model by learnable prompt parameters " + }, + { + "bbox": [ + 356, + 327, + 482, + 426 + ], + "type": "inline_equation", + "content": "\\mathbf{P_T}" + }, + { + "bbox": [ + 356, + 327, + 482, + 426 + ], + "type": "text", + "content": " for privacy-aware few-shot continual action recognition. We explain prompt selection mechanism in Fig. 3." + } + ] + } + ], + "index": 11, + "angle": 0, + "type": "image_caption" + } + ], + "index": 10 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "spans": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "type": "text", + "content": "6" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 166, + 91, + 222, + 102 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 166, + 91, + 222, + 102 + ], + "spans": [ + { + "bbox": [ + 166, + 91, + 222, + 102 + ], + "type": "text", + "content": "P. Garg et al." + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 5 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 141, + 114, + 470, + 222 + ], + "blocks": [ + { + "bbox": [ + 141, + 114, + 470, + 222 + ], + "lines": [ + { + "bbox": [ + 141, + 114, + 470, + 222 + ], + "spans": [ + { + "bbox": [ + 141, + 114, + 470, + 222 + ], + "type": "image", + "image_path": "1d64d8218068d4851448335f14f345fe791b511cc7f3c49d714e715015bbf95a.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 132, + 223, + 482, + 275 + ], + "lines": [ + { + "bbox": [ + 132, + 223, + 482, + 275 + ], + "spans": [ + { + "bbox": [ + 132, + 223, + 482, + 275 + ], + "type": "text", + "content": "Fig. 3: Selection of our prompts " + }, + { + "bbox": [ + 132, + 223, + 482, + 275 + ], + "type": "inline_equation", + "content": "\\mathbf{P_T}" + }, + { + "bbox": [ + 132, + 223, + 482, + 275 + ], + "type": "text", + "content": ": Input-dependent query " + }, + { + "bbox": [ + 132, + 223, + 482, + 275 + ], + "type": "inline_equation", + "content": "q" + }, + { + "bbox": [ + 132, + 223, + 482, + 275 + ], + "type": "text", + "content": " is matched with keys " + }, + { + "bbox": [ + 132, + 223, + 482, + 275 + ], + "type": "inline_equation", + "content": "K" + }, + { + "bbox": [ + 132, + 223, + 482, + 275 + ], + "type": "text", + "content": " using sorted cosine similarity to get an ordered index sequence " + }, + { + "bbox": [ + 132, + 223, + 482, + 275 + ], + "type": "inline_equation", + "content": "(s_i)_{i=1}^T" + }, + { + "bbox": [ + 132, + 223, + 482, + 275 + ], + "type": "text", + "content": " of the top " + }, + { + "bbox": [ + 132, + 223, + 482, + 275 + ], + "type": "inline_equation", + "content": "T" + }, + { + "bbox": [ + 132, + 223, + 482, + 275 + ], + "type": "text", + "content": " keys. This ordered index sequence is used to select the corresponding ordered prompt sequence " + }, + { + "bbox": [ + 132, + 223, + 482, + 275 + ], + "type": "inline_equation", + "content": "\\mathbf{P_T}" + }, + { + "bbox": [ + 132, + 223, + 482, + 275 + ], + "type": "text", + "content": " from prompt pool " + }, + { + "bbox": [ + 132, + 223, + 482, + 275 + ], + "type": "inline_equation", + "content": "\\mathbf{P}" + }, + { + "bbox": [ + 132, + 223, + 482, + 275 + ], + "type": "text", + "content": ". We add " + }, + { + "bbox": [ + 132, + 223, + 482, + 275 + ], + "type": "inline_equation", + "content": "\\mathbf{P_T}" + }, + { + "bbox": [ + 132, + 223, + 482, + 275 + ], + "type": "text", + "content": " to " + }, + { + "bbox": [ + 132, + 223, + 482, + 275 + ], + "type": "inline_equation", + "content": "\\mathbf{X_e}" + }, + { + "bbox": [ + 132, + 223, + 482, + 275 + ], + "type": "text", + "content": ", thereby adding an offset to it. Our experimental evaluation confirms that such an additive spatio-temporal prompt offset can balance the plasticity to learn new classes from a few action samples, while maintaining stability on previously learned classes." + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_caption" + } + ], + "index": 2 + }, + { + "bbox": [ + 132, + 285, + 266, + 297 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 285, + 266, + 297 + ], + "spans": [ + { + "bbox": [ + 132, + 285, + 266, + 297 + ], + "type": "text", + "content": "knowledge across the sessions:" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 159, + 300, + 481, + 316 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 159, + 300, + 481, + 316 + ], + "spans": [ + { + "bbox": [ + 159, + 300, + 481, + 316 + ], + "type": "interline_equation", + "content": "\\mathbf {P} = \\left\\{P _ {1},.. P _ {i},..., P _ {M} \\right\\}, \\qquad P _ {i} \\in \\mathbb {R} ^ {J \\times C _ {e}}; M = \\# \\text {p r o m p t s a t t i m e t} \\tag {2}", + "image_path": "ee8e5026386ddcf92ba494279de5db863ad7c17766ef78b3b7e75122ed3dbeb6.jpg" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 131, + 320, + 479, + 380 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 131, + 320, + 479, + 380 + ], + "spans": [ + { + "bbox": [ + 131, + 320, + 479, + 380 + ], + "type": "text", + "content": "For selecting prompts from this pool (Fig. 3), we construct a bijective key-value codebook, treating prompts in the pool " + }, + { + "bbox": [ + 131, + 320, + 479, + 380 + ], + "type": "inline_equation", + "content": "\\mathbf{P}" + }, + { + "bbox": [ + 131, + 320, + 479, + 380 + ], + "type": "text", + "content": " as values and defining learnable key vectors " + }, + { + "bbox": [ + 131, + 320, + 479, + 380 + ], + "type": "inline_equation", + "content": "\\pmb{K} = \\{\\pmb{k}_1, \\dots, \\pmb{k}_i, \\dots, \\pmb{k}_M\\}, \\pmb{k}_i \\in \\mathbb{R}^{C_e}" + }, + { + "bbox": [ + 131, + 320, + 479, + 380 + ], + "type": "text", + "content": ". A cosine similarity matching " + }, + { + "bbox": [ + 131, + 320, + 479, + 380 + ], + "type": "inline_equation", + "content": "\\gamma(.)" + }, + { + "bbox": [ + 131, + 320, + 479, + 380 + ], + "type": "text", + "content": " between the query " + }, + { + "bbox": [ + 131, + 320, + 479, + 380 + ], + "type": "inline_equation", + "content": "\\pmb{q}" + }, + { + "bbox": [ + 131, + 320, + 479, + 380 + ], + "type": "text", + "content": " and keys " + }, + { + "bbox": [ + 131, + 320, + 479, + 380 + ], + "type": "inline_equation", + "content": "\\pmb{K}" + }, + { + "bbox": [ + 131, + 320, + 479, + 380 + ], + "type": "text", + "content": " is used to find indices of the " + }, + { + "bbox": [ + 131, + 320, + 479, + 380 + ], + "type": "inline_equation", + "content": "T" + }, + { + "bbox": [ + 131, + 320, + 479, + 380 + ], + "type": "text", + "content": " closest keys " + }, + { + "bbox": [ + 131, + 320, + 479, + 380 + ], + "type": "inline_equation", + "content": "\\mathbb{Z}" + }, + { + "bbox": [ + 131, + 320, + 479, + 380 + ], + "type": "text", + "content": ", which in turn are used to select prompts from the pool:" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 250, + 384, + 481, + 404 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 250, + 384, + 481, + 404 + ], + "spans": [ + { + "bbox": [ + 250, + 384, + 481, + 404 + ], + "type": "interline_equation", + "content": "\\mathbb {Z} = \\underset {T} {\\operatorname {a r g m a x}} \\gamma \\left(f _ {q} (\\mathbf {X}), \\mathbf {K}\\right) \\tag {3}", + "image_path": "5558c6e2c0d71a83351767a042ba7a6881f8b4aa96b09c4e42ea64f2e19e1cb6.jpg" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 132, + 410, + 479, + 434 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 410, + 479, + 434 + ], + "spans": [ + { + "bbox": [ + 132, + 410, + 479, + 434 + ], + "type": "text", + "content": "This quantization process is enabled by a query function " + }, + { + "bbox": [ + 132, + 410, + 479, + 434 + ], + "type": "inline_equation", + "content": "f_{q}(.)" + }, + { + "bbox": [ + 132, + 410, + 479, + 434 + ], + "type": "text", + "content": ", which is a pretrained encoder that maps an input instance " + }, + { + "bbox": [ + 132, + 410, + 479, + 434 + ], + "type": "inline_equation", + "content": "\\mathbf{X}" + }, + { + "bbox": [ + 132, + 410, + 479, + 434 + ], + "type": "text", + "content": " to a query " + }, + { + "bbox": [ + 132, + 410, + 479, + 434 + ], + "type": "inline_equation", + "content": "\\pmb{q}" + }, + { + "bbox": [ + 132, + 410, + 479, + 434 + ], + "type": "text", + "content": " as:" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 182, + 437, + 481, + 453 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 182, + 437, + 481, + 453 + ], + "spans": [ + { + "bbox": [ + 182, + 437, + 481, + 453 + ], + "type": "interline_equation", + "content": "\\boldsymbol {q} = f _ {q} (\\mathbf {X}) = f _ {Q A} \\circ f _ {g} ^ {\\prime} \\circ f _ {e} ^ {\\prime} (\\mathbf {X}), \\quad f _ {q}: \\mathbb {R} ^ {T \\times J \\times 3} \\longrightarrow \\mathbb {R} ^ {C _ {e}} \\tag {4}", + "image_path": "9d0d14114e92b3332d7e3ca317dec3cb06d29fe8e7cc6646ee5fe4ba4bca90e6.jpg" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 131, + 458, + 479, + 482 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 131, + 458, + 479, + 482 + ], + "spans": [ + { + "bbox": [ + 131, + 458, + 479, + 482 + ], + "type": "text", + "content": "where the query adaptor " + }, + { + "bbox": [ + 131, + 458, + 479, + 482 + ], + "type": "inline_equation", + "content": "f_{QA}" + }, + { + "bbox": [ + 131, + 458, + 479, + 482 + ], + "type": "text", + "content": " is a fully connected layer mapping the " + }, + { + "bbox": [ + 131, + 458, + 479, + 482 + ], + "type": "inline_equation", + "content": "f_{g}^{\\prime}" + }, + { + "bbox": [ + 131, + 458, + 479, + 482 + ], + "type": "text", + "content": " output dimension to the desired prompt embedding dimension " + }, + { + "bbox": [ + 131, + 458, + 479, + 482 + ], + "type": "inline_equation", + "content": "C_e" + }, + { + "bbox": [ + 131, + 458, + 479, + 482 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 131, + 486, + 482, + 668 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 131, + 486, + 482, + 668 + ], + "spans": [ + { + "bbox": [ + 131, + 486, + 482, + 668 + ], + "type": "text", + "content": "Coupled Optimization in User Sessions " + }, + { + "bbox": [ + 131, + 486, + 482, + 668 + ], + "type": "inline_equation", + "content": "t > 0" + }, + { + "bbox": [ + 131, + 486, + 482, + 668 + ], + "type": "text", + "content": ". Typically, the argmax operator in Eq. 3 decouples the optimization of keys from the prompt pool and main model as it prevents backpropagation of gradients to the keys (also seen in earlier works such as [48, 49]). However, this approach does not work for our setting, even more so since we assume no large-scale pre-training of our base model. Due to the lack of off-the-shelf availability of large-scale models for skeletal action data, our query function " + }, + { + "bbox": [ + 131, + 486, + 482, + 668 + ], + "type": "inline_equation", + "content": "f_{q}" + }, + { + "bbox": [ + 131, + 486, + 482, + 668 + ], + "type": "text", + "content": " is pretrained only on base class data " + }, + { + "bbox": [ + 131, + 486, + 482, + 668 + ], + "type": "inline_equation", + "content": "\\mathcal{D}^{(0)}" + }, + { + "bbox": [ + 131, + 486, + 482, + 668 + ], + "type": "text", + "content": ". Hence, it becomes important that " + }, + { + "bbox": [ + 131, + 486, + 482, + 668 + ], + "type": "inline_equation", + "content": "f_{q}" + }, + { + "bbox": [ + 131, + 486, + 482, + 668 + ], + "type": "text", + "content": " is updated as the model learns new classes. As shown in red boxes in Figs. 2 and 3, we propose to couple this optimization process such that the overall cross-entropy loss for new tasks updates: (i) the classifier " + }, + { + "bbox": [ + 131, + 486, + 482, + 668 + ], + "type": "inline_equation", + "content": "f_{c}" + }, + { + "bbox": [ + 131, + 486, + 482, + 668 + ], + "type": "text", + "content": ", (ii) selected prompts in " + }, + { + "bbox": [ + 131, + 486, + 482, + 668 + ], + "type": "inline_equation", + "content": "\\mathbf{P}" + }, + { + "bbox": [ + 131, + 486, + 482, + 668 + ], + "type": "text", + "content": ", (iii) selected keys in " + }, + { + "bbox": [ + 131, + 486, + 482, + 668 + ], + "type": "inline_equation", + "content": "K" + }, + { + "bbox": [ + 131, + 486, + 482, + 668 + ], + "type": "text", + "content": ", as well as (iv) query adaptor " + }, + { + "bbox": [ + 131, + 486, + 482, + 668 + ], + "type": "inline_equation", + "content": "f_{QA}" + }, + { + "bbox": [ + 131, + 486, + 482, + 668 + ], + "type": "text", + "content": ". We achieve this by approximating the gradient for " + }, + { + "bbox": [ + 131, + 486, + 482, + 668 + ], + "type": "inline_equation", + "content": "K" + }, + { + "bbox": [ + 131, + 486, + 482, + 668 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 131, + 486, + 482, + 668 + ], + "type": "inline_equation", + "content": "f_{QA}" + }, + { + "bbox": [ + 131, + 486, + 482, + 668 + ], + "type": "text", + "content": " by the straight-through estimator reparameterization trick as in [3, 44]. We freeze the query feature extractor layers " + }, + { + "bbox": [ + 131, + 486, + 482, + 668 + ], + "type": "inline_equation", + "content": "f_{g}^{\\prime}, f_{e}^{\\prime}" + }, + { + "bbox": [ + 131, + 486, + 482, + 668 + ], + "type": "text", + "content": " in " + }, + { + "bbox": [ + 131, + 486, + 482, + 668 + ], + "type": "inline_equation", + "content": "t > 0" + }, + { + "bbox": [ + 131, + 486, + 482, + 668 + ], + "type": "text", + "content": " to prevent catastrophic forgetting of base knowledge in " + }, + { + "bbox": [ + 131, + 486, + 482, + 668 + ], + "type": "inline_equation", + "content": "f_{q}" + }, + { + "bbox": [ + 131, + 486, + 482, + 668 + ], + "type": "text", + "content": ". Our cross-entropy loss is hence given by:" + } + ] + } + ], + "index": 11 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 156, + 91, + 448, + 102 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 156, + 91, + 448, + 102 + ], + "spans": [ + { + "bbox": [ + 156, + 91, + 448, + 102 + ], + "type": "text", + "content": "POET: Prompt Offset Tuning for Continual Human Action Adaptation" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 474, + 92, + 481, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 474, + 92, + 481, + 100 + ], + "spans": [ + { + "bbox": [ + 474, + 92, + 481, + 100 + ], + "type": "text", + "content": "7" + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 6 + }, + { + "para_blocks": [ + { + "bbox": [ + 242, + 116, + 480, + 136 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 242, + 116, + 480, + 136 + ], + "spans": [ + { + "bbox": [ + 242, + 116, + 480, + 136 + ], + "type": "interline_equation", + "content": "\\min _ {\\theta_ {f _ {Q A}}, \\theta_ {K}, \\theta_ {\\mathbf {P}}, \\theta_ {f _ {c}}} \\mathcal {L} (f (\\mathbf {X}, \\mathbf {P} _ {\\mathbf {T}}), \\mathbf {y}) \\tag {5}", + "image_path": "df497f5795e408f45a78c084eb92d408f8afe85a12ebd71e2e2b891b748a4b48.jpg" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 130, + 140, + 479, + 164 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 140, + 479, + 164 + ], + "spans": [ + { + "bbox": [ + 130, + 140, + 479, + 164 + ], + "type": "text", + "content": "To move queries closer to their aligned " + }, + { + "bbox": [ + 130, + 140, + 479, + 164 + ], + "type": "inline_equation", + "content": "T" + }, + { + "bbox": [ + 130, + 140, + 479, + 164 + ], + "type": "text", + "content": " keys during training, we use a vector quantization clustering loss inspired from VQ-VAE [44] as:" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 247, + 167, + 481, + 192 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 247, + 167, + 481, + 192 + ], + "spans": [ + { + "bbox": [ + 247, + 167, + 481, + 192 + ], + "type": "interline_equation", + "content": "\\max _ {\\theta_ {f _ {Q A}}, \\theta_ {K}} \\lambda \\sum_ {i \\in \\mathbb {Z}} \\gamma \\left(f _ {q} (\\mathbf {X}), K _ {i}\\right) \\tag {6}", + "image_path": "59026ae7dfc73fa0b78d4fb2a5cdad3306eb5df8250b4f19cbaeda4ccb9c0f44.jpg" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 130, + 196, + 479, + 243 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 196, + 479, + 243 + ], + "spans": [ + { + "bbox": [ + 130, + 196, + 479, + 243 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 130, + 196, + 479, + 243 + ], + "type": "inline_equation", + "content": "\\lambda" + }, + { + "bbox": [ + 130, + 196, + 479, + 243 + ], + "type": "text", + "content": " is the clustering loss coefficient. Our end-to-end optimization thus establishes a prompt optimization framework which is amenable to prompt tuning when extensive pre-training is not possible. This sets the foundation for our spatio-temporal prompt selection module, described next." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 130, + 247, + 479, + 284 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 247, + 479, + 284 + ], + "spans": [ + { + "bbox": [ + 130, + 247, + 479, + 284 + ], + "type": "text", + "content": "Spatio-Temporal Prompt Selection. In order to ensure that our learned prompts respect temporal information in the input video sequence, we choose the number of selected prompts to be equal to the number of frames in the input" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 130, + 285, + 298, + 437 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 285, + 298, + 437 + ], + "spans": [ + { + "bbox": [ + 130, + 285, + 298, + 437 + ], + "type": "text", + "content": "video " + }, + { + "bbox": [ + 130, + 285, + 298, + 437 + ], + "type": "inline_equation", + "content": "T" + }, + { + "bbox": [ + 130, + 285, + 298, + 437 + ], + "type": "text", + "content": ". After coupling the prompt pool and keys, we observed in our initial experiments with pool size " + }, + { + "bbox": [ + 130, + 285, + 298, + 437 + ], + "type": "inline_equation", + "content": "M > T" + }, + { + "bbox": [ + 130, + 285, + 298, + 437 + ], + "type": "text", + "content": " that the same set of prompts get selected across training iterations and user sessions (Fig. 4A). More concretely, as the vector quantization loss (Eqn 6) brings the query close to the selected keys, the same set of active prompts get selected and optimized in each iteration, not using other prompts at all. This is similar to the well-known issue of 'codebook col" + } + ] + } + ], + "index": 7 + }, + { + "type": "image", + "bbox": [ + 306, + 285, + 481, + 340 + ], + "blocks": [ + { + "bbox": [ + 306, + 285, + 481, + 340 + ], + "lines": [ + { + "bbox": [ + 306, + 285, + 481, + 340 + ], + "spans": [ + { + "bbox": [ + 306, + 285, + 481, + 340 + ], + "type": "image", + "image_path": "655464329a10d8067e2e10803916ebc135458dab373e9d280dcd48d26a17afed.jpg" + } + ] + } + ], + "index": 8, + "angle": 0, + "type": "image_body" + } + ], + "index": 8 + }, + { + "type": "image", + "bbox": [ + 306, + 345, + 480, + 399 + ], + "blocks": [ + { + "bbox": [ + 306, + 345, + 480, + 399 + ], + "lines": [ + { + "bbox": [ + 306, + 345, + 480, + 399 + ], + "spans": [ + { + "bbox": [ + 306, + 345, + 480, + 399 + ], + "type": "image", + "image_path": "1183da4b6d6d075e6477a447af1a7ac4f6075e36005e75afa5cc5623a79d6a0d.jpg" + } + ] + } + ], + "index": 9, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 304, + 399, + 481, + 433 + ], + "lines": [ + { + "bbox": [ + 304, + 399, + 481, + 433 + ], + "spans": [ + { + "bbox": [ + 304, + 399, + 481, + 433 + ], + "type": "text", + "content": "Fig.4: " + }, + { + "bbox": [ + 304, + 399, + 481, + 433 + ], + "type": "inline_equation", + "content": "M > T" + }, + { + "bbox": [ + 304, + 399, + 481, + 433 + ], + "type": "text", + "content": " Case: Prompt Pool Collapse. (Top) Certain prompt indices remain unused across user sessions. (Bottom) Our POET pool expansion strategy alleviates pool collapse." + } + ] + } + ], + "index": 10, + "angle": 0, + "type": "image_caption" + } + ], + "index": 9 + }, + { + "bbox": [ + 130, + 439, + 479, + 462 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 439, + 479, + 462 + ], + "spans": [ + { + "bbox": [ + 130, + 439, + 479, + 462 + ], + "type": "text", + "content": "lapse' in VQ-VAE [9, 50, 54]. Based on this observation, we design two prompt pool update mechanisms in user sessions " + }, + { + "bbox": [ + 130, + 439, + 479, + 462 + ], + "type": "inline_equation", + "content": "t > 0" + }, + { + "bbox": [ + 130, + 439, + 479, + 462 + ], + "type": "text", + "content": " as below:" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 131, + 464, + 480, + 512 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 131, + 464, + 480, + 512 + ], + "spans": [ + { + "bbox": [ + 131, + 464, + 480, + 512 + ], + "type": "text", + "content": "1. Case 1, " + }, + { + "bbox": [ + 131, + 464, + 480, + 512 + ], + "type": "inline_equation", + "content": "M = T \\forall t" + }, + { + "bbox": [ + 131, + 464, + 480, + 512 + ], + "type": "text", + "content": ": No pool expansion, Algorithm 1. All prompts are selected in all tasks. But the order of their selection " + }, + { + "bbox": [ + 131, + 464, + 480, + 512 + ], + "type": "inline_equation", + "content": "(s_i)_{i=1}^T" + }, + { + "bbox": [ + 131, + 464, + 480, + 512 + ], + "type": "text", + "content": " varies with each input instance as we replace Eq. 3 by sorting the cosine similarity before selecting the top " + }, + { + "bbox": [ + 131, + 464, + 480, + 512 + ], + "type": "inline_equation", + "content": "T" + }, + { + "bbox": [ + 131, + 464, + 480, + 512 + ], + "type": "text", + "content": " indices as follows:" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 257, + 514, + 480, + 536 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 257, + 514, + 480, + 536 + ], + "spans": [ + { + "bbox": [ + 257, + 514, + 480, + 536 + ], + "type": "interline_equation", + "content": "\\mathbb {Z} = \\underset {\\left(s _ {i}\\right) _ {i = 1} ^ {T}} {\\operatorname {a r g s o r t}} \\gamma \\left(f _ {q} (\\mathbf {X}), \\mathbf {K}\\right) \\tag {7}", + "image_path": "983dea89389e0a64812733258c29cb2602edf1de1dcaa03d7d744896e8dfc9f5.jpg" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 143, + 539, + 481, + 586 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 143, + 539, + 481, + 586 + ], + "spans": [ + { + "bbox": [ + 143, + 539, + 481, + 586 + ], + "type": "text", + "content": "In Fig. 5, we visualize the positions occupied by indices in this (sorted) ordered key index sequence " + }, + { + "bbox": [ + 143, + 539, + 481, + 586 + ], + "type": "inline_equation", + "content": "(s_i)_{i=1}^T" + }, + { + "bbox": [ + 143, + 539, + 481, + 586 + ], + "type": "text", + "content": ". Entropy increase across tasks " + }, + { + "bbox": [ + 143, + 539, + 481, + 586 + ], + "type": "inline_equation", + "content": "t = 1" + }, + { + "bbox": [ + 143, + 539, + 481, + 586 + ], + "type": "text", + "content": " to " + }, + { + "bbox": [ + 143, + 539, + 481, + 586 + ], + "type": "inline_equation", + "content": "t = 4" + }, + { + "bbox": [ + 143, + 539, + 481, + 586 + ], + "type": "text", + "content": " (bottom row of figure) shows that our selection mechanism learns to select a unique temporal code for all inputs." + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 131, + 586, + 479, + 635 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 131, + 586, + 479, + 635 + ], + "spans": [ + { + "bbox": [ + 131, + 586, + 479, + 635 + ], + "type": "text", + "content": "2. Case 2, " + }, + { + "bbox": [ + 131, + 586, + 479, + 635 + ], + "type": "inline_equation", + "content": "M = T + (R * t), t > 0" + }, + { + "bbox": [ + 131, + 586, + 479, + 635 + ], + "type": "text", + "content": ". Expand pool with " + }, + { + "bbox": [ + 131, + 586, + 479, + 635 + ], + "type": "inline_equation", + "content": "R" + }, + { + "bbox": [ + 131, + 586, + 479, + 635 + ], + "type": "text", + "content": " prompts. We also propose an order-aware prompt pool expansion strategy (Appendix B) that selects prompts from an expanded pool in a temporally coherent manner, for " + }, + { + "bbox": [ + 131, + 586, + 479, + 635 + ], + "type": "inline_equation", + "content": "t > 0" + }, + { + "bbox": [ + 131, + 586, + 479, + 635 + ], + "type": "text", + "content": ". This alleviates prompt pool collapse as shown in Fig. 4B." + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 130, + 641, + 479, + 665 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 641, + 479, + 665 + ], + "spans": [ + { + "bbox": [ + 130, + 641, + 479, + 665 + ], + "type": "text", + "content": "Prompt Offset Attachment. Since concatenation is not meaningful for graph data, we use addition as our choice for the prompt attachment operator as:" + } + ] + } + ], + "index": 16 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "spans": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "type": "text", + "content": "8" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 166, + 91, + 222, + 102 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 166, + 91, + 222, + 102 + ], + "spans": [ + { + "bbox": [ + 166, + 91, + 222, + 102 + ], + "type": "text", + "content": "P. Garg et al." + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 7 + }, + { + "para_blocks": [ + { + "bbox": [ + 253, + 115, + 481, + 129 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 253, + 115, + 481, + 129 + ], + "spans": [ + { + "bbox": [ + 253, + 115, + 481, + 129 + ], + "type": "interline_equation", + "content": "f _ {p} \\left(\\mathbf {X} _ {\\mathbf {e}}, \\mathbf {P} _ {\\mathbf {T}}\\right) = \\mathbf {X} _ {\\mathbf {e}} + \\mathbf {P} _ {\\mathbf {T}} \\tag {8}", + "image_path": "9e08d93bcf3bb1dfaa8be928dadffde624ab6e097a9a5f346d65788ad48b1656.jpg" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 130, + 133, + 479, + 157 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 133, + 479, + 157 + ], + "spans": [ + { + "bbox": [ + 130, + 133, + 479, + 157 + ], + "type": "text", + "content": "Hence, we call our approach as prompt offset tuning. We also study this empirically through experiments that support this choice in Sec. 6." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 130, + 159, + 480, + 196 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 159, + 480, + 196 + ], + "spans": [ + { + "bbox": [ + 130, + 159, + 480, + 196 + ], + "type": "text", + "content": "Interpreting Prompt Offset Tuning of GNNs. Our additive prompt offsets are open to interpretation, as shown in Fig. 5. (i) Adding our selected prompts " + }, + { + "bbox": [ + 130, + 159, + 480, + 196 + ], + "type": "inline_equation", + "content": "\\mathbf{P_T}" + }, + { + "bbox": [ + 130, + 159, + 480, + 196 + ], + "type": "text", + "content": " to input feature embedding " + }, + { + "bbox": [ + 130, + 159, + 480, + 196 + ], + "type": "inline_equation", + "content": "\\mathbf{X_e}" + }, + { + "bbox": [ + 130, + 159, + 480, + 196 + ], + "type": "text", + "content": " acts like an input-dependent transformation" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 130, + 196, + 245, + 208 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 196, + 245, + 208 + ], + "spans": [ + { + "bbox": [ + 130, + 196, + 245, + 208 + ], + "type": "text", + "content": "for spatio-temporal joints." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 130, + 209, + 238, + 376 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 209, + 238, + 376 + ], + "spans": [ + { + "bbox": [ + 130, + 209, + 238, + 376 + ], + "type": "text", + "content": "(ii) As our prompts have same size as " + }, + { + "bbox": [ + 130, + 209, + 238, + 376 + ], + "type": "inline_equation", + "content": "\\mathbf{X}_{\\mathbf{e}}" + }, + { + "bbox": [ + 130, + 209, + 238, + 376 + ], + "type": "text", + "content": " it can also be thought of as a learned prompt encoding, bearing similarity with learnable position encoding works [12, 24, 26]. Our purpose is different however as prompt offsets seek to dynamically condition the input for adapting the backbone continually," + } + ] + } + ], + "index": 6 + }, + { + "type": "image", + "bbox": [ + 238, + 198, + 482, + 304 + ], + "blocks": [ + { + "bbox": [ + 238, + 198, + 482, + 304 + ], + "lines": [ + { + "bbox": [ + 238, + 198, + 482, + 304 + ], + "spans": [ + { + "bbox": [ + 238, + 198, + 482, + 304 + ], + "type": "image", + "image_path": "8b6936c5caba33ff9e421a5b926c64d9a3beb54d70507904b14af0dd6f349945.jpg" + } + ] + } + ], + "index": 7, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 235, + 309, + 482, + 369 + ], + "lines": [ + { + "bbox": [ + 235, + 309, + 482, + 369 + ], + "spans": [ + { + "bbox": [ + 235, + 309, + 482, + 369 + ], + "type": "text", + "content": "Fig. 5: Here we visualize the order " + }, + { + "bbox": [ + 235, + 309, + 482, + 369 + ], + "type": "inline_equation", + "content": "(s_i)_{i=1}^T" + }, + { + "bbox": [ + 235, + 309, + 482, + 369 + ], + "type": "text", + "content": " in which the " + }, + { + "bbox": [ + 235, + 309, + 482, + 369 + ], + "type": "inline_equation", + "content": "M = 64" + }, + { + "bbox": [ + 235, + 309, + 482, + 369 + ], + "type": "text", + "content": " prompts in the pool are selected at train time, across 4 user sessions " + }, + { + "bbox": [ + 235, + 309, + 482, + 369 + ], + "type": "inline_equation", + "content": "\\mathcal{U}\\mathcal{S}^{(t)}" + }, + { + "bbox": [ + 235, + 309, + 482, + 369 + ], + "type": "text", + "content": ". X-axis: prompt index, Y-axis: index position in selected sequence. Top: The no sorting case uses the default sequence (hence diagonal matrices), giving equal importance to all prompts. Bottom (Our Method): Even though the same 64 prompts are selected and updated, the ordering is temporally unique and consistent with input." + } + ] + } + ], + "index": 8, + "angle": 0, + "type": "image_caption" + } + ], + "index": 7 + }, + { + "bbox": [ + 130, + 376, + 482, + 449 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 376, + 482, + 449 + ], + "spans": [ + { + "bbox": [ + 130, + 376, + 482, + 449 + ], + "type": "text", + "content": "instead of learning positions. (iii) POET also bears similarity with auto-decoders like DeepSDF [30] which learn latent codes for each style or shape and use relevant codes along with a frozen decoder at inference. (iv) Prompt tuning can also be thought of as a parameter isolation technique for continual learning [28,34,35,38]. POET's ordered prompt selection as seen in Fig. 5 learns to isolate the relevant sequence of prompts for each input action sequence." + } + ] + } + ], + "index": 9 + }, + { + "type": "code", + "bbox": [ + 132, + 470, + 482, + 580 + ], + "blocks": [ + { + "bbox": [ + 132, + 456, + 462, + 469 + ], + "lines": [ + { + "bbox": [ + 132, + 456, + 462, + 469 + ], + "spans": [ + { + "bbox": [ + 132, + 456, + 462, + 469 + ], + "type": "text", + "content": "Algorithm 1 POET at Train Time, " + }, + { + "bbox": [ + 132, + 456, + 462, + 469 + ], + "type": "inline_equation", + "content": "t > 0" + }, + { + "bbox": [ + 132, + 456, + 462, + 469 + ], + "type": "text", + "content": " (Case 1 " + }, + { + "bbox": [ + 132, + 456, + 462, + 469 + ], + "type": "inline_equation", + "content": "M = T" + }, + { + "bbox": [ + 132, + 456, + 462, + 469 + ], + "type": "text", + "content": ", No pool expansion)" + } + ] + } + ], + "index": 10, + "angle": 0, + "type": "code_caption" + }, + { + "bbox": [ + 132, + 470, + 482, + 580 + ], + "lines": [ + { + "bbox": [ + 132, + 470, + 482, + 580 + ], + "spans": [ + { + "bbox": [ + 132, + 470, + 482, + 580 + ], + "type": "text", + "content": "Input: Query function " + }, + { + "bbox": [ + 132, + 470, + 482, + 580 + ], + "type": "inline_equation", + "content": "f_{q}" + }, + { + "bbox": [ + 132, + 470, + 482, + 580 + ], + "type": "text", + "content": " keys " + }, + { + "bbox": [ + 132, + 470, + 482, + 580 + ], + "type": "inline_equation", + "content": "\\pmb {K} = \\{\\pmb {k}_j\\}_{j = 1}^T" + }, + { + "bbox": [ + 132, + 470, + 482, + 580 + ], + "type": "text", + "content": " prompt pool " + }, + { + "bbox": [ + 132, + 470, + 482, + 580 + ], + "type": "inline_equation", + "content": "\\mathbf{P} = \\{\\pmb {P}_j\\}_{j = 1}^T" + }, + { + "bbox": [ + 132, + 470, + 482, + 580 + ], + "type": "text", + "content": " main model " + }, + { + "bbox": [ + 132, + 470, + 482, + 580 + ], + "type": "inline_equation", + "content": "f_{e},f_{g},f_{c}" + }, + { + "bbox": [ + 132, + 470, + 482, + 580 + ], + "type": "text", + "content": " Initialize: " + }, + { + "bbox": [ + 132, + 470, + 482, + 580 + ], + "type": "inline_equation", + "content": "\\mathbf{P},\\mathbf{K}" + }, + { + "bbox": [ + 132, + 470, + 482, + 580 + ], + "type": "text", + "content": " from " + }, + { + "bbox": [ + 132, + 470, + 482, + 580 + ], + "type": "inline_equation", + "content": "t - 1" + }, + { + "bbox": [ + 132, + 470, + 482, + 580 + ], + "type": "text", + "content": " ; Expand " + }, + { + "bbox": [ + 132, + 470, + 482, + 580 + ], + "type": "inline_equation", + "content": "f_{c}" + }, + { + "bbox": [ + 132, + 470, + 482, + 580 + ], + "type": "text", + "content": " by " + }, + { + "bbox": [ + 132, + 470, + 482, + 580 + ], + "type": "inline_equation", + "content": "N" + }, + { + "bbox": [ + 132, + 470, + 482, + 580 + ], + "type": "text", + "content": " new classes. Initialize " + }, + { + "bbox": [ + 132, + 470, + 482, + 580 + ], + "type": "inline_equation", + "content": "f_{c}" + }, + { + "bbox": [ + 132, + 470, + 482, + 580 + ], + "type": "text", + "content": " as: (i) copy " + }, + { + "bbox": [ + 132, + 470, + 482, + 580 + ], + "type": "inline_equation", + "content": "f_{c}^{old}" + }, + { + "bbox": [ + 132, + 470, + 482, + 580 + ], + "type": "text", + "content": " weights, (ii) " + }, + { + "bbox": [ + 132, + 470, + 482, + 580 + ], + "type": "inline_equation", + "content": "f_{c}^{new}\\gets Mean(f_{c}^{old})" + }, + { + "bbox": [ + 132, + 470, + 482, + 580 + ], + "type": "text", + "content": " Freeze: query layers " + }, + { + "bbox": [ + 132, + 470, + 482, + 580 + ], + "type": "inline_equation", + "content": "f_{g}^{\\prime},f_{e}^{\\prime}" + }, + { + "bbox": [ + 132, + 470, + 482, + 580 + ], + "type": "text", + "content": " ; main model layers " + }, + { + "bbox": [ + 132, + 470, + 482, + 580 + ], + "type": "inline_equation", + "content": "f_{e},f_{g}" + }, + { + "bbox": [ + 132, + 470, + 482, + 580 + ], + "type": "text", + "content": " for epochs and batch " + }, + { + "bbox": [ + 132, + 470, + 482, + 580 + ], + "type": "inline_equation", + "content": "(\\mathbf{X}_i^t,\\mathbf{y}_i^t)_{i = 1}^{NK}" + }, + { + "bbox": [ + 132, + 470, + 482, + 580 + ], + "type": "text", + "content": " do 1. Get query feature " + }, + { + "bbox": [ + 132, + 470, + 482, + 580 + ], + "type": "inline_equation", + "content": "\\pmb{q}" + }, + { + "bbox": [ + 132, + 470, + 482, + 580 + ], + "type": "text", + "content": " (Eq.4); Compute " + }, + { + "bbox": [ + 132, + 470, + 482, + 580 + ], + "type": "inline_equation", + "content": "\\gamma (\\cdot)" + }, + { + "bbox": [ + 132, + 470, + 482, + 580 + ], + "type": "text", + "content": " b/w query " + }, + { + "bbox": [ + 132, + 470, + 482, + 580 + ], + "type": "inline_equation", + "content": "\\pmb{q}" + }, + { + "bbox": [ + 132, + 470, + 482, + 580 + ], + "type": "text", + "content": " and keys " + }, + { + "bbox": [ + 132, + 470, + 482, + 580 + ], + "type": "inline_equation", + "content": "\\pmb{K}" + }, + { + "bbox": [ + 132, + 470, + 482, + 580 + ], + "type": "text", + "content": " 2.Sort " + }, + { + "bbox": [ + 132, + 470, + 482, + 580 + ], + "type": "inline_equation", + "content": "\\gamma (\\cdot)" + }, + { + "bbox": [ + 132, + 470, + 482, + 580 + ], + "type": "text", + "content": " ; Get ordered key index sequence " + }, + { + "bbox": [ + 132, + 470, + 482, + 580 + ], + "type": "inline_equation", + "content": "(s_i)_i = 1" + }, + { + "bbox": [ + 132, + 470, + 482, + 580 + ], + "type": "text", + "content": " Eq.7 3. Read pool memory " + }, + { + "bbox": [ + 132, + 470, + 482, + 580 + ], + "type": "inline_equation", + "content": "\\mathbf{P}" + }, + { + "bbox": [ + 132, + 470, + 482, + 580 + ], + "type": "text", + "content": " in order " + }, + { + "bbox": [ + 132, + 470, + 482, + 580 + ], + "type": "inline_equation", + "content": "(s_i)^T_{i = 1}\\rightarrow" + }, + { + "bbox": [ + 132, + 470, + 482, + 580 + ], + "type": "text", + "content": " Get prompt offsets " + }, + { + "bbox": [ + 132, + 470, + 482, + 580 + ], + "type": "inline_equation", + "content": "\\mathbf{P_T}" + }, + { + "bbox": [ + 132, + 470, + 482, + 580 + ], + "type": "text", + "content": " 4.Get " + }, + { + "bbox": [ + 132, + 470, + 482, + 580 + ], + "type": "inline_equation", + "content": "\\mathbf{X_e};" + }, + { + "bbox": [ + 132, + 470, + 482, + 580 + ], + "type": "text", + "content": " Add " + }, + { + "bbox": [ + 132, + 470, + 482, + 580 + ], + "type": "inline_equation", + "content": "\\mathbf{P_T}" + }, + { + "bbox": [ + 132, + 470, + 482, + 580 + ], + "type": "text", + "content": " to it (Eq.8); get prediction y from prompted input (Eq. 1) 5.Use cross entropy loss (Equation 5) to update " + }, + { + "bbox": [ + 132, + 470, + 482, + 580 + ], + "type": "inline_equation", + "content": "f_{QA},K,\\mathbf{P},f_{c}" + }, + { + "bbox": [ + 132, + 470, + 482, + 580 + ], + "type": "text", + "content": " 6.Use clustering loss (Equation 6) to update " + }, + { + "bbox": [ + 132, + 470, + 482, + 580 + ], + "type": "inline_equation", + "content": "f_{QA}" + }, + { + "bbox": [ + 132, + 470, + 482, + 580 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 132, + 470, + 482, + 580 + ], + "type": "inline_equation", + "content": "\\pmb{K}" + }, + { + "bbox": [ + 132, + 470, + 482, + 580 + ], + "type": "text", + "content": " end // See " + }, + { + "bbox": [ + 132, + 470, + 482, + 580 + ], + "type": "inline_equation", + "content": "t = 0" + }, + { + "bbox": [ + 132, + 470, + 482, + 580 + ], + "type": "text", + "content": " training protocol in Algorithm 2 in Appendix" + } + ] + } + ], + "index": 11, + "angle": 0, + "type": "code_body" + } + ], + "index": 11, + "sub_type": "algorithm" + }, + { + "bbox": [ + 132, + 595, + 304, + 609 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 595, + 304, + 609 + ], + "spans": [ + { + "bbox": [ + 132, + 595, + 304, + 609 + ], + "type": "text", + "content": "5 Experiments and Results" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 130, + 612, + 482, + 638 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 612, + 482, + 638 + ], + "spans": [ + { + "bbox": [ + 130, + 612, + 482, + 638 + ], + "type": "text", + "content": "Datasets. We evaluated our method on well-known action recognition datasets: (i) activity recognition on the NTU RGB+D dataset [39]; and (ii) hand gesture" + } + ] + } + ], + "index": 13 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 155, + 90, + 448, + 103 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 155, + 90, + 448, + 103 + ], + "spans": [ + { + "bbox": [ + 155, + 90, + 448, + 103 + ], + "type": "text", + "content": "POET: Prompt Offset Tuning for Continual Human Action Adaptation" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 474, + 91, + 481, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 474, + 91, + 481, + 100 + ], + "spans": [ + { + "bbox": [ + 474, + 91, + 481, + 100 + ], + "type": "text", + "content": "9" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 133, + 642, + 481, + 665 + ], + "type": "page_footnote", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 642, + 481, + 665 + ], + "spans": [ + { + "bbox": [ + 133, + 642, + 481, + 665 + ], + "type": "text", + "content": "8 The datasets used in this work were accessed and processed at and by CMU. They were not accessed, processed, stored, or maintained at Meta." + } + ] + } + ], + "index": 14 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 8 + }, + { + "para_blocks": [ + { + "bbox": [ + 130, + 116, + 480, + 152 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 116, + 480, + 152 + ], + "spans": [ + { + "bbox": [ + 130, + 116, + 480, + 152 + ], + "type": "text", + "content": "recognition on the SHREC-2017 dataset [40]. As we introduce a new problem setting in human action recognition, we contribute two new benchmarks to the community for this setting, on the NTU RGB+D and SHREC-2017 datasets." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 130, + 155, + 482, + 251 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 155, + 482, + 251 + ], + "spans": [ + { + "bbox": [ + 130, + 155, + 482, + 251 + ], + "type": "text", + "content": "For the NTU RGB+D dataset, we divide the 60 daily action categories into 40 base classes, learning the remaining 20 classes in subsequent user sessions. In few-shot learning parlance, our protocol is 4-task 5-way 5-shot, i.e. 5 novel classes using 5 user training instances in 4 user sessions. Each input 3D skeleton sequence has 64 temporal frames, each consisting of 25 body keypoints, such that " + }, + { + "bbox": [ + 130, + 155, + 482, + 251 + ], + "type": "inline_equation", + "content": "x \\in \\mathcal{R}^{64 \\times 25 \\times 3}" + }, + { + "bbox": [ + 130, + 155, + 482, + 251 + ], + "type": "text", + "content": ". We use the spatio-temporal GCN, CTR-GCN [8], as the architecture for NTU RGB+D, where we choose the joint input modality for better interpretability of prompt tuning." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 130, + 253, + 482, + 350 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 253, + 482, + 350 + ], + "spans": [ + { + "bbox": [ + 130, + 253, + 482, + 350 + ], + "type": "text", + "content": "For SHREC-2017, we divide the 14 fine-grained hand gesture classes into 8 base classes and 6 classes learned in subsequent user sessions. This is done in a 3-task 2-way 5-shot protocol, i.e. 2 novel classes using 5 user training instances in 3 user sessions. For each input instance of SHREC-2017, we use 8 temporal frames each having 22 hand keypoints, such that input " + }, + { + "bbox": [ + 130, + 253, + 482, + 350 + ], + "type": "inline_equation", + "content": "x \\in \\mathcal{R}^{8 \\times 22 \\times 3}" + }, + { + "bbox": [ + 130, + 253, + 482, + 350 + ], + "type": "text", + "content": ". We use a fully-connected graph transformer backbone, DG-STA [7] for SHREC-2017. We select DG-STA due to easily reproducible code and to validate if our method POET works equally well across graph convolutional networks and graph transformers." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 131, + 351, + 482, + 449 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 131, + 351, + 482, + 449 + ], + "spans": [ + { + "bbox": [ + 131, + 351, + 482, + 449 + ], + "type": "text", + "content": "Evaluation Metrics. Following earlier work in similar settings [31], we report: (i) Average accuracy 'Avg' of all classes seen so far, and (ii) Harmonic Mean " + }, + { + "bbox": [ + 131, + 351, + 482, + 449 + ], + "type": "inline_equation", + "content": "A_{HM}" + }, + { + "bbox": [ + 131, + 351, + 482, + 449 + ], + "type": "text", + "content": " between 'accuracy only on Old classes' and 'accuracy only on New classes' after learning each new user session. Note that the average accuracy tends to be biased towards the base session " + }, + { + "bbox": [ + 131, + 351, + 482, + 449 + ], + "type": "inline_equation", + "content": "\\mathcal{T}^{(0)}" + }, + { + "bbox": [ + 131, + 351, + 482, + 449 + ], + "type": "text", + "content": " performance due to more number of base classes. A higher " + }, + { + "bbox": [ + 131, + 351, + 482, + 449 + ], + "type": "inline_equation", + "content": "A_{HM}" + }, + { + "bbox": [ + 131, + 351, + 482, + 449 + ], + "type": "text", + "content": " implies better stability-plasticity trade-off between new task performance and old tasks' retention. Unlike many earlier CIL efforts, we report accuracy for both Old and New classes in each user session for transparency." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 131, + 450, + 482, + 666 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 131, + 450, + 482, + 666 + ], + "spans": [ + { + "bbox": [ + 131, + 450, + 482, + 666 + ], + "type": "text", + "content": "Implementation Details. We observe that a key source of forgetting in our setting is from the classifier as the logits tend to become heavily biased towards the few-shot samples of new classes. We use a cosine classifier for activity recognition experiments on CTR-GCN. For gesture recognition on the lightweight DG-STA, we use a standard fully-connected layer as classifier, but freeze old class parameters in the classifier by zeroing their gradients. We attach prompts after the 1st layer of DG-STA and 1st CTR-GC block of CTR-GCN. For both datasets, we have equal or higher learning rates in user sessions when compared to the base model's training in order to accommodate new knowledge in the model (for better plasticity). For exact implementation details (including learning rates, epochs, hyperparameter analysis, and backward forgetting metric), see Appendix A. In earlier efforts that more generally tune prompts for class-incremental learning [41, 45, 47-49], it is common to rely on an ImageNet21K pretrained ViT [37] or CLIP [33] as the backbone. However, such backbones do not exist for skeleton-based human action recognition. Our base feature extractor is hence trained on the base session dataset itself without any pretraining, making this one of the first efforts of prompt tuning without extensive pretraining (scale of data 3-5 times lower order of magnitude)." + } + ] + } + ], + "index": 6 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "spans": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "type": "text", + "content": "10" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 166, + 91, + 222, + 102 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 166, + 91, + 222, + 102 + ], + "spans": [ + { + "bbox": [ + 166, + 91, + 222, + 102 + ], + "type": "text", + "content": "P. Garg et al." + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 9 + }, + { + "para_blocks": [ + { + "bbox": [ + 130, + 116, + 482, + 297 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 116, + 482, + 297 + ], + "spans": [ + { + "bbox": [ + 130, + 116, + 482, + 297 + ], + "type": "text", + "content": "Results. Since there are no existing baselines for our proposed setting in skeletal action recognition, we compare our method by adapting continual learning (CL) baselines to skeletal data in Sec 5.1, Tables 1, 2. We first compare POET with prompt tuning based class-incremental learning (CIL) approaches originally designed for images (L2P [49], CODA-P [41], APT [4]) and find that it has very low performance on new classes as they do not update their query function. We find any fine-tuning or knowledge distillation based approaches (LWF [25], EWC [17], LUCIR [15]) lead to rapid forgetting of base knowledge as the model overfits to user's few-shots. We also compare with multiple variants of Feature Extraction (FE) to check if prompts truly have merit (POET=FE+Prompts) and provide upper bound baselines. In Sec 6, we first show the importance of prompts in POET by removing the prompts. We discuss the value of our coupled optimization, query function update and ordered key index selection in our prompt selection ablation Tab 3. We also study the impact of proposed additive prompt tuning as compared to other possible prompt attachments " + }, + { + "bbox": [ + 130, + 116, + 482, + 297 + ], + "type": "inline_equation", + "content": "f_{p}" + }, + { + "bbox": [ + 130, + 116, + 482, + 297 + ], + "type": "text", + "content": " in Tab 4." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 131, + 312, + 332, + 323 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 131, + 312, + 332, + 323 + ], + "spans": [ + { + "bbox": [ + 131, + 312, + 332, + 323 + ], + "type": "text", + "content": "5.1 Comparison with State-of-the-Art" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 130, + 331, + 482, + 510 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 331, + 482, + 510 + ], + "spans": [ + { + "bbox": [ + 130, + 331, + 482, + 510 + ], + "type": "text", + "content": "POET sets the SOTA on existing prompt tuning works (Tab 1,2). We adapt three standard CIL works that prompt tune ViTs for images - L2P [49], CODA-P [41] and APT [4] to our setting. L2P and CODA-P share prompt pool across tasks (similar to us), whereas APT learns task-specific prompts. L2P decouples the optimization of keys from the prompt pool and concatenates the selected prompts. Since concatenation is not defined for our GNN backbone, we adapt these SOTA to our setting by concatenating along the temporal dimension " + }, + { + "bbox": [ + 130, + 331, + 482, + 510 + ], + "type": "inline_equation", + "content": "(\\mathbf{L2P}^*,\\mathbf{CODA - P}^*)" + }, + { + "bbox": [ + 130, + 331, + 482, + 510 + ], + "type": "text", + "content": ". CODA-P [41] couples keys with the prompt pool by using a cosine similarity weighing over all prompts in the pool, forming a 'soft prompt selection', different from our 'ordered hard prompt selection'. In APT, we train prompt-classifier pairs for each continual task separately (" + }, + { + "bbox": [ + 130, + 331, + 482, + 510 + ], + "type": "inline_equation", + "content": "\\hat{\\mathbf{\\Gamma}}" + }, + { + "bbox": [ + 130, + 331, + 482, + 510 + ], + "type": "text", + "content": " denotes task-specific), and use task identity at test time. See details in Appendix A. These methods by design rely on extensively pretrained (ImageNet21k) query functions which does not require updates; and require full supervision on new classes, perhaps explaining their poor 'New' accuracy in our few-shot setting." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 130, + 510, + 482, + 666 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 510, + 482, + 666 + ], + "spans": [ + { + "bbox": [ + 130, + 510, + 482, + 666 + ], + "type": "text", + "content": "Standard Continual learning Baselines. We compared with two well established knowledge-distillation approaches, learning without forgetting (LWF) and LUCIR. Both of them perform poorly on both old and new classes. EWC [17] learns better on new but does not retain old knowledge. We conclude that any CL method that fine-tunes the backbone feature representation in subsequent sessions " + }, + { + "bbox": [ + 130, + 510, + 482, + 666 + ], + "type": "inline_equation", + "content": "t > 0" + }, + { + "bbox": [ + 130, + 510, + 482, + 666 + ], + "type": "text", + "content": " will not be able to retain base/old class knowledge (a finding consistent with existing FSCIL literature for images [10,43]). We also adapt and compare with one of the latest FSCIL baselines ALICE [31], originally developed for image classification benchmarks on our gesture recognition benchmark in Table 2. Note the high retention of base task performance (due to non-parametric classifier on top of frozen base model). However, it suffers from poor plasticity and adaptation to new classes. This is the issue of feature-classifier misalignment that we hoped to alleviate through prompt tuning." + } + ] + } + ], + "index": 5 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 156, + 90, + 447, + 102 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 156, + 90, + 447, + 102 + ], + "spans": [ + { + "bbox": [ + 156, + 90, + 447, + 102 + ], + "type": "text", + "content": "POET: Prompt Offset Tuning for Continual Human Action Adaptation" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 470, + 91, + 479, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 470, + 91, + 479, + 100 + ], + "spans": [ + { + "bbox": [ + 470, + 91, + 479, + 100 + ], + "type": "text", + "content": "11" + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 10 + }, + { + "para_blocks": [ + { + "type": "table", + "bbox": [ + 135, + 168, + 473, + 369 + ], + "blocks": [ + { + "bbox": [ + 132, + 115, + 482, + 166 + ], + "lines": [ + { + "bbox": [ + 132, + 115, + 482, + 166 + ], + "spans": [ + { + "bbox": [ + 132, + 115, + 482, + 166 + ], + "type": "text", + "content": "Table 1: Activity Recognition Results " + }, + { + "bbox": [ + 132, + 115, + 482, + 166 + ], + "type": "inline_equation", + "content": "(\\% ,\\uparrow)" + }, + { + "bbox": [ + 132, + 115, + 482, + 166 + ], + "type": "text", + "content": ", Comparison with SOTA: NTU RGB+D [39] dataset on CTR-GCN [8] backbone. After training on each incremental task, we report Average of all classes seen so far ('Avg'). We also report (i) " + }, + { + "bbox": [ + 132, + 115, + 482, + 166 + ], + "type": "inline_equation", + "content": "A_{HM}" + }, + { + "bbox": [ + 132, + 115, + 482, + 166 + ], + "type": "text", + "content": ", (ii) old classes accuracy ('Old'), (iii) new classes accuracy ('New') in the last session. We report Mean and STD across 10 sets of 5-shots. POET achieves the best stability-plasticity trade-off across all baselines indicated by the " + }, + { + "bbox": [ + 132, + 115, + 482, + 166 + ], + "type": "inline_equation", + "content": "A_{HM} = 56.3\\%" + }, + { + "bbox": [ + 132, + 115, + 482, + 166 + ], + "type": "text", + "content": ". POET also has the highest Avg across all user sessions outside of upper bound baselines in orange." + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 135, + 168, + 473, + 369 + ], + "lines": [ + { + "bbox": [ + 135, + 168, + 473, + 369 + ], + "spans": [ + { + "bbox": [ + 135, + 168, + 473, + 369 + ], + "type": "table", + "html": "
Method\\( \\mathbf{U}\\mathbf{B}^{(0)} \\)\\( \\mathbf{U}\\mathbf{S}^{(1)} \\)\\( \\mathbf{U}\\mathbf{S}^{(2)} \\)\\( \\mathbf{U}\\mathbf{S}^{(3)} \\)\\( \\mathbf{U}\\mathbf{S}^{(4)} \\)
Base (↑)Avg (↑)Avg (↑)Avg (↑)Old (↑)New (↑)Avg (↑)\\( A_{HM} \\) (↑)
Upper Bounds
Joint (Oracle)88.479.071.066.863.5
Joint POET (Oracle)67.2
FE, Task-Specific^88.470.1 ± 2.652.5 ± 5.844.8 ± 5.070.3 ± 2.146.7 ± 2.0NANA
FE+Replay88.482.4 ± 1.178.2 ± 1.274.5 ± 1.273.1 ± 1.043.3 ± 3.370.6 ± 1.254.3 ± 2.6
Continual Linear Probing
FE88.472.0 ± 1.160.4 ± 2.447.7 ± 2.140.0 ± 1.651.0 ± 2.340.9 ± 1.444.8 ± 1.1
FE, Frozen88.476.1 ± 1.052.4 ± 4.138.3 ± 2.728.4 ± 1.622.4 ± 4.527.9 ± 1.424.8 ± 3.0
FE+Replay†88.472.0 ± 1.559.5 ± 4.058.7 ± 2.856.7 ± 2.534.7 ± 5.654.9 ± 2.742.8 ± 4.4
FT88.46.2 ± 1.44.3 ± 1.52.8 ± 1.00.2 ± 0.536.0 ± 10.13.2 ± 0.80.3 ± 1.0
Standard Continual Learning
LWF [25]88.46.2 ± 1.52.8 ± 0.73.7 ± 1.30.0 ± 0.038.9 ± 8.83.2 ± 0.70.0 ± 0.0
EWC [17]88.46.6 ± 1.54.1 ± 1.43.1 ± 0.90.0 ± 0.042.1 ± 9.53.5 ± 0.80.0 ± 0.0
Experience Replay88.435.1 ± 8.350.6 ± 5.060.6 ± 5.454.6 ± 6.543.7 ± 14.653.7 ± 7.147.8 ± 11.2
Experience \\( \\text{Replay}^{\\dagger} \\)88.46.2 ± 1.59.0 ± 2.611.2 ± 3.010.9 ± 2.634.6 ± 7.912.9 ± 3.016.3 ± 3.5
LUCIR [15]87.94.3 ± 2.14.1 ± 1.32.7 ± 0.80.2 ± 0.426.0 ± 9.22.3 ± 0.90.4 ± 0.8
Continual Prompt Tuning
CODA-P [41]*87.476.1 ± 1.066.7 ± 1.358.6 ± 2.756.5 ± 2.90.5 ± 0.451.8 ± 2.71.1 ± 0.7
L2P [49]*88.678.9 ± 0.171.0 ± 1.064.2 ± 0.162.0 ± 0.70.0 ± 0.056.8 ± 0.60.0 ± 0.0
APT [4]^86.627.3 ± 1.630.8 ± 3.437.6 ± 2.3NA33.4 ± 2.0NANA
POET (Ours)87.982.3 ± 0.676.8 ± 0.968.4 ± 0.757.2 ± 1.055.8 ± 5.957.1 ± 1.156.3 ± 3.2
", + "image_path": "5ecc0bbf3092fcabfdfcad1ed96c79e801d80f9b5530c10715bfd2d644eb700b.jpg" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "table_body" + } + ], + "index": 3 + }, + { + "bbox": [ + 130, + 386, + 482, + 530 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 386, + 482, + 530 + ], + "spans": [ + { + "bbox": [ + 130, + 386, + 482, + 530 + ], + "type": "text", + "content": "Fine-tuning (FE) and Feature Extraction (FE) Baselines. We implement standard continual learning baselines to understand stability-plasticity trade-offs in our new benchmarks. In all these baselines, we expand the classifier output dimension by " + }, + { + "bbox": [ + 130, + 386, + 482, + 530 + ], + "type": "inline_equation", + "content": "N" + }, + { + "bbox": [ + 130, + 386, + 482, + 530 + ], + "type": "text", + "content": " new classes. In 'FT (Fine-Tuning)', we tune all model parameters on cross entropy loss of new task. FSCIL is challenging for this modality as old task performance sharply reduces to zero starting from " + }, + { + "bbox": [ + 130, + 386, + 482, + 530 + ], + "type": "inline_equation", + "content": "\\mathcal{U}\\mathcal{S}^{(1)}" + }, + { + "bbox": [ + 130, + 386, + 482, + 530 + ], + "type": "text", + "content": " as model overfits to user's few-shots. 'FE (Feature Extraction)' differs from FT as we freeze the feature extractor to preserve base knowledge. This serves as a competitive baseline in our findings. In 'FE, frozen', we zero out the gradients of previous class weights in classifier " + }, + { + "bbox": [ + 130, + 386, + 482, + 530 + ], + "type": "inline_equation", + "content": "f_{c}" + }, + { + "bbox": [ + 130, + 386, + 482, + 530 + ], + "type": "text", + "content": " to prevent forgetting from the classifier. 'FE' and 'FE, Frozen' exhibit different New-Old trade-offs in Tables 1, 2 because the scale of pretraining is different (gesture more lightweight than activity)." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 130, + 532, + 482, + 640 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 532, + 482, + 640 + ], + "spans": [ + { + "bbox": [ + 130, + 532, + 482, + 640 + ], + "type": "text", + "content": "Upper-bound baselines, top section Tables 1, 2. In 'Joint (oracle)' experiment, we train on all task data at the same time in a multi-task (non-sequential) manner. Training POET in a multi-task manner ('Joint POET') outperforms 'Joint Oracle' demonstrating the strength of our approach. In addition to these generalist upper bounds, we point out that 'FE, Task-specific' is a competitive specialist upper bound. In this, we perform feature extraction from base model to each task individually, storing separate task-specific models " + }, + { + "bbox": [ + 130, + 532, + 482, + 640 + ], + "type": "inline_equation", + "content": "(\\mathcal{U}\\mathcal{S}^{(0)}\\to \\mathcal{U}\\mathcal{S}^{(i)},i > 0)" + }, + { + "bbox": [ + 130, + 532, + 482, + 640 + ], + "type": "text", + "content": ". POET outperforms 'New' accuracy compared with this baseline, achieving a forward transfer on each " + }, + { + "bbox": [ + 130, + 532, + 482, + 640 + ], + "type": "inline_equation", + "content": "t > 0" + }, + { + "bbox": [ + 130, + 532, + 482, + 640 + ], + "type": "text", + "content": ". This indicates that prompt" + } + ] + } + ], + "index": 5 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "spans": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "type": "text", + "content": "12" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 166, + 91, + 222, + 102 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 166, + 91, + 222, + 102 + ], + "spans": [ + { + "bbox": [ + 166, + 91, + 222, + 102 + ], + "type": "text", + "content": "P. Garg et al." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 133, + 653, + 448, + 665 + ], + "type": "page_footnote", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 653, + 448, + 665 + ], + "spans": [ + { + "bbox": [ + 133, + 653, + 448, + 665 + ], + "type": "text", + "content": "9 'FE' is the same as 'w/o prompts' in Table 3. We highlight key baselines in gray color." + } + ] + } + ], + "index": 6 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 11 + }, + { + "para_blocks": [ + { + "type": "table", + "bbox": [ + 156, + 145, + 455, + 252 + ], + "blocks": [ + { + "bbox": [ + 132, + 115, + 482, + 141 + ], + "lines": [ + { + "bbox": [ + 132, + 115, + 482, + 141 + ], + "spans": [ + { + "bbox": [ + 132, + 115, + 482, + 141 + ], + "type": "text", + "content": "Table 2: Gesture Recognition Results (\\%, ↑), Comparison with SOTA: SHREC 2017 [40] dataset on DG-STA [7] graph transformer backbone. Reporting mean and standard deviation across 5 runs. POET achieves best " + }, + { + "bbox": [ + 132, + 115, + 482, + 141 + ], + "type": "inline_equation", + "content": "A_{HM} = 56.2\\%" + }, + { + "bbox": [ + 132, + 115, + 482, + 141 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 156, + 145, + 455, + 252 + ], + "lines": [ + { + "bbox": [ + 156, + 145, + 455, + 252 + ], + "spans": [ + { + "bbox": [ + 156, + 145, + 455, + 252 + ], + "type": "table", + "html": "
Method\\( \\mathbf{\\mathcal{U}\\mathbf{B}^{(0)}} \\)\\( \\mathbf{\\mathcal{U}\\mathbf{S}^{(1)}} \\)\\( \\mathbf{\\mathcal{U}\\mathbf{S}^{(2)}} \\)\\( \\mathbf{\\mathcal{U}\\mathbf{S}^{(3)}} \\)
Base (↑)Avg (↑)Avg (↑)Old (↑)New (↑)Avg (↑)\\( A_{HM} \\) (↑)
Joint (Oracle)88.879.4 ± 0.777.3 ± 2.170.9 ± 1.262.4 ± 0.4
FT88.820.3 ± 0.812.4 ± 2.10.0 ± 0.085.8 ± 9.413.4 ± 1.50.0 ± 0.0
FE88.862.7 ± 2.441.9 ± 6.917.5 ± 5.177.3 ± 8.826.8 ± 3.428.5 ± 6.4
FE, Frozen88.871.3 ± 1.961.4 ± 2.744.7 ± 3.254.5 ± 6.746.2 ± 2.749.1 ± 4.3
LWF [25]88.820.2 ± 1.412.5 ± 1.00.0 ± 0.088.4 ± 13.713.8 ± 2.10.0 ± 0.0
L2P [49]**88.820.3 ± 5.910.5 ± 4.88.2 ± 4.06.9 ± 8.57.9 ± 3.97.5 ± 5.5
CODA-P [41]**87.715.6 ± 4.511.6 ± 1.97.9 ± 1.814.1 ± 21.48.8 ± 2.410.1 ± 3.2
ALICE [31]92.172.4 ± 5.763.3 ± 7.662.5 ± 6.811.9 ± 9.954.6 ± 6.920.0 ± 8.1
POET (Ours)91.973.2 ± 3.761.9 ± 1.845.9 ± 2.672.4 ± 7.150.0 ± 1.656.2 ± 1.6
", + "image_path": "dd25c58ccaa6e1c39338adc55e993dd398d2ef5700630ebe08616e6dd57a4bc1.jpg" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "table_body" + } + ], + "index": 3 + }, + { + "bbox": [ + 130, + 266, + 479, + 289 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 266, + 479, + 289 + ], + "spans": [ + { + "bbox": [ + 130, + 266, + 479, + 289 + ], + "type": "text", + "content": "tuning benefits New performance due to the pre-existing knowledge in the shared knowledge pool. Avg in sessions " + }, + { + "bbox": [ + 130, + 266, + 479, + 289 + ], + "type": "inline_equation", + "content": "0 < t < 4" + }, + { + "bbox": [ + 130, + 266, + 479, + 289 + ], + "type": "text", + "content": " indicates New for task-specific" + }, + { + "bbox": [ + 130, + 266, + 479, + 289 + ], + "type": "inline_equation", + "content": "^{\\wedge}" + }, + { + "bbox": [ + 130, + 266, + 479, + 289 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 130, + 290, + 480, + 337 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 290, + 480, + 337 + ], + "spans": [ + { + "bbox": [ + 130, + 290, + 480, + 337 + ], + "type": "text", + "content": "Experience Replay Baselines, Tab 1. Even though our privacy-aware setting prohibits previous data replay, we compare with 'Experience Replay' (store and replay 5-samples of base and incremental sessions) and 'Experience Replay†' (replay only previous incremental sessions) for completeness." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 130, + 338, + 480, + 397 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 338, + 480, + 397 + ], + "spans": [ + { + "bbox": [ + 130, + 338, + 480, + 397 + ], + "type": "text", + "content": "'FE+Replay' serves as the best upper bound (even better than Experience Replay as we are freezing backbone in addition to replay). It is noteworthy that POET (which is FE+prompts) learns an implicit 'data-free' form of prompt pool memory, and yet has a better " + }, + { + "bbox": [ + 130, + 338, + 480, + 397 + ], + "type": "inline_equation", + "content": "A_{HM}" + }, + { + "bbox": [ + 130, + 338, + 480, + 397 + ], + "type": "text", + "content": " trade-off as compared to explicitly stored and replayed samples from previous classes in FE+replay." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 132, + 404, + 334, + 418 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 404, + 334, + 418 + ], + "spans": [ + { + "bbox": [ + 132, + 404, + 334, + 418 + ], + "type": "text", + "content": "6 Ablation Studies and Analysis" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 130, + 424, + 482, + 495 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 424, + 482, + 495 + ], + "spans": [ + { + "bbox": [ + 130, + 424, + 482, + 495 + ], + "type": "text", + "content": "Importance of prompts in POET. First, we consider the contribution of prompt offsets in POET. Since we only attach prompts to address continual learning in POET, removing prompts gives the Feature Extraction (FE) baseline ('w/o prompts', Table 3) where the backbone is frozen after base training and only the classifier is expanded and updated on classification loss of new classes. POET improves both, 'Old' (" + }, + { + "bbox": [ + 130, + 424, + 482, + 495 + ], + "type": "inline_equation", + "content": "\\uparrow 20.1\\%" + }, + { + "bbox": [ + 130, + 424, + 482, + 495 + ], + "type": "text", + "content": ") and 'New' (" + }, + { + "bbox": [ + 130, + 424, + 482, + 495 + ], + "type": "inline_equation", + "content": "\\uparrow 10.6\\%" + }, + { + "bbox": [ + 130, + 424, + 482, + 495 + ], + "type": "text", + "content": ") marked in blue." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 130, + 496, + 282, + 614 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 496, + 282, + 614 + ], + "spans": [ + { + "bbox": [ + 130, + 496, + 282, + 614 + ], + "type": "text", + "content": "Prompt Selection Mechanism. In Table 3, we investigate our prompt selection mechanism and optimization choices. The 'w/o coupled optim.' experiment is a direct comparison of our additive prompt attachment with the de-coupled optimization in L2P [49]. Updating key parameters but keeping only query adaptor" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 286, + 496, + 482, + 544 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 286, + 496, + 482, + 544 + ], + "spans": [ + { + "bbox": [ + 286, + 496, + 482, + 544 + ], + "type": "text", + "content": "Table 3: Prompt Selection Mechanism Analysis on NTU RGB+D dataset " + }, + { + "bbox": [ + 286, + 496, + 482, + 544 + ], + "type": "inline_equation", + "content": "(\\% ,\\uparrow)" + }, + { + "bbox": [ + 286, + 496, + 482, + 544 + ], + "type": "text", + "content": " .. " + }, + { + "bbox": [ + 286, + 496, + 482, + 544 + ], + "type": "inline_equation", + "content": "\\mathrm{w / o}" + }, + { + "bbox": [ + 286, + 496, + 482, + 544 + ], + "type": "text", + "content": " denotes removing that component from POET, numbers in brackets are wrt POET " + }, + { + "bbox": [ + 286, + 496, + 482, + 544 + ], + "type": "inline_equation", + "content": "(M = T)" + }, + { + "bbox": [ + 286, + 496, + 482, + 544 + ], + "type": "text", + "content": " experiment. 'Avg' accuracy is biased towards Old' classes accuracy, " + }, + { + "bbox": [ + 286, + 496, + 482, + 544 + ], + "type": "inline_equation", + "content": "A_{HM}" + }, + { + "bbox": [ + 286, + 496, + 482, + 544 + ], + "type": "text", + "content": " is good indicator of trade-off between New' and Old'." + } + ] + } + ], + "index": 10 + }, + { + "type": "table", + "bbox": [ + 287, + 545, + 485, + 613 + ], + "blocks": [ + { + "bbox": [ + 287, + 545, + 485, + 613 + ], + "lines": [ + { + "bbox": [ + 287, + 545, + 485, + 613 + ], + "spans": [ + { + "bbox": [ + 287, + 545, + 485, + 613 + ], + "type": "table", + "html": "
NTU RGB+D\\( \\left| \\mathcal{U}\\mathcal{B}^{(0)}\\right| \\)\\( \\left| \\mathcal{U}\\mathcal{S}^{(1)}\\right| \\)\\( \\left| \\mathcal{U}\\mathcal{S}^{(2)}\\right| \\)\\( \\left| \\mathcal{U}\\mathcal{S}^{(3)}\\right| \\)\\( \\left| \\mathcal{U}\\mathcal{S}^{(4)}\\right| \\)
MethodBaseAvgAvgAvgOldNewAvg\\( A_{HM} \\)
w/o prompts88.474.566.349.539.2 (-20.1)46.8 (-10.6)39.942.7
w/o coupled optim.88.082.875.365.856.5 (-2.8)51.3 (-6.1)56.153.8
w/o clustering loss85.581.674.364.562.0 (+2.7)18.2 (-39.2)57.028.1
w/o QA update87.982.877.469.159.4 (+0.1)52.8 (-4.6)58.755.9
w/o sorting88.282.275.268.859.9 (+0.6)46.6 (-10.8)58.852.4
POET (M>T)87.982.777.268.860.3 (+1.0)54.4 (-3.0)59.857.2
POET (M=T)87.982.876.868.659.357.459.258.3
", + "image_path": "0ab5cb54fa4aa5aebe7c535490da1a25f9563d48a4a29b259b09df1711d07886.jpg" + } + ] + } + ], + "index": 11, + "angle": 0, + "type": "table_body" + } + ], + "index": 11 + }, + { + "bbox": [ + 130, + 615, + 482, + 666 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 615, + 482, + 666 + ], + "spans": [ + { + "bbox": [ + 130, + 615, + 482, + 666 + ], + "type": "inline_equation", + "content": "QA" + }, + { + "bbox": [ + 130, + 615, + 482, + 666 + ], + "type": "text", + "content": " frozen after " + }, + { + "bbox": [ + 130, + 615, + 482, + 666 + ], + "type": "inline_equation", + "content": "\\mathcal{U}\\mathcal{B}^{(0)}" + }, + { + "bbox": [ + 130, + 615, + 482, + 666 + ], + "type": "text", + "content": " training ('w/o QA update') reduces 'New' only performance of " + }, + { + "bbox": [ + 130, + 615, + 482, + 666 + ], + "type": "inline_equation", + "content": "\\mathcal{US}^{(4)}" + }, + { + "bbox": [ + 130, + 615, + 482, + 666 + ], + "type": "text", + "content": " by " + }, + { + "bbox": [ + 130, + 615, + 482, + 666 + ], + "type": "inline_equation", + "content": "4.6\\%" + }, + { + "bbox": [ + 130, + 615, + 482, + 666 + ], + "type": "text", + "content": " as the query function stays fixed at base session learning and is not discriminative towards new classes. 'W/o clustering loss' from Eq. 6, performance drops starting from " + }, + { + "bbox": [ + 130, + 615, + 482, + 666 + ], + "type": "inline_equation", + "content": "\\mathcal{UB}^{(0)}" + }, + { + "bbox": [ + 130, + 615, + 482, + 666 + ], + "type": "text", + "content": " itself. The only difference" + } + ] + } + ], + "index": 12 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 156, + 91, + 447, + 102 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 156, + 91, + 447, + 102 + ], + "spans": [ + { + "bbox": [ + 156, + 91, + 447, + 102 + ], + "type": "text", + "content": "POET: Prompt Offset Tuning for Continual Human Action Adaptation" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 470, + 91, + 481, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 470, + 91, + 481, + 100 + ], + "spans": [ + { + "bbox": [ + 470, + 91, + 481, + 100 + ], + "type": "text", + "content": "13" + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 12 + }, + { + "para_blocks": [ + { + "bbox": [ + 130, + 116, + 482, + 212 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 116, + 482, + 212 + ], + "spans": [ + { + "bbox": [ + 130, + 116, + 482, + 212 + ], + "type": "text", + "content": "between the experiment 'w/o sorting' and 'POET (M=T)' is that we do not sort the cosine similarity before selecting top " + }, + { + "bbox": [ + 130, + 116, + 482, + 212 + ], + "type": "inline_equation", + "content": "T" + }, + { + "bbox": [ + 130, + 116, + 482, + 212 + ], + "type": "text", + "content": " indices (same as Fig 5). The " + }, + { + "bbox": [ + 130, + 116, + 482, + 212 + ], + "type": "inline_equation", + "content": "10.8\\% \\uparrow" + }, + { + "bbox": [ + 130, + 116, + 482, + 212 + ], + "type": "text", + "content": " in 'New' performance validates that our prompt selection mechanism is learning to choose a distinct temporal ordering for prompt tuning of new input samples. With pool expansion ('POET, " + }, + { + "bbox": [ + 130, + 116, + 482, + 212 + ], + "type": "inline_equation", + "content": "M > T" + }, + { + "bbox": [ + 130, + 116, + 482, + 212 + ], + "type": "text", + "content": "), we get more flexibility in the stability-plasticity trade-offs depending on how many new prompts we attach. For " + }, + { + "bbox": [ + 130, + 116, + 482, + 212 + ], + "type": "inline_equation", + "content": "R = 6" + }, + { + "bbox": [ + 130, + 116, + 482, + 212 + ], + "type": "text", + "content": ", 'Old' is improved. In Table 3, we keep POET's additive prompt attachment and only vary prompt selection." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 130, + 213, + 482, + 297 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 213, + 482, + 297 + ], + "spans": [ + { + "bbox": [ + 130, + 213, + 482, + 297 + ], + "type": "text", + "content": "Prompt Attachment Mechanism. In Table 4, we keep our end-to-end optimization and ordered prompt selection as a constant and ablate prompt shape and attachment operator " + }, + { + "bbox": [ + 130, + 213, + 482, + 297 + ], + "type": "inline_equation", + "content": "f_{p}(.)" + }, + { + "bbox": [ + 130, + 213, + 482, + 297 + ], + "type": "text", + "content": ". Drawing a parallel with transformers which concatenate prompts along the token dimension, we conduct experiments concatenating prompts along the (i) temporal dimension of the skeleton input feature embedding " + }, + { + "bbox": [ + 130, + 213, + 482, + 297 + ], + "type": "inline_equation", + "content": "\\mathbf{X}_{\\mathrm{e}}" + }, + { + "bbox": [ + 130, + 213, + 482, + 297 + ], + "type": "text", + "content": " ('CONCAT temporal') and (ii) feature dimension " + }, + { + "bbox": [ + 130, + 213, + 482, + 297 + ], + "type": "inline_equation", + "content": "C_e" + }, + { + "bbox": [ + 130, + 213, + 482, + 297 + ], + "type": "text", + "content": " ('CONCAT feature'). We find that addition works better than concatenation and" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 130, + 297, + 299, + 393 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 297, + 299, + 393 + ], + "spans": [ + { + "bbox": [ + 130, + 297, + 299, + 393 + ], + "type": "text", + "content": "cross attention. We also verify our hypothesis that selecting the same number of prompts as the input temporal dimension (" + }, + { + "bbox": [ + 130, + 297, + 299, + 393 + ], + "type": "inline_equation", + "content": "T = 64" + }, + { + "bbox": [ + 130, + 297, + 299, + 393 + ], + "type": "text", + "content": " for NTU RGB+D and " + }, + { + "bbox": [ + 130, + 297, + 299, + 393 + ], + "type": "inline_equation", + "content": "T = 8" + }, + { + "bbox": [ + 130, + 297, + 299, + 393 + ], + "type": "text", + "content": " for SHREC-2017) yields better results as compared to adding the same prompt frame to each input embedding frame ('Addition " + }, + { + "bbox": [ + 130, + 297, + 299, + 393 + ], + "type": "inline_equation", + "content": "T' = 1'" + }, + { + "bbox": [ + 130, + 297, + 299, + 393 + ], + "type": "text", + "content": ")." + } + ] + } + ], + "index": 4 + }, + { + "type": "table", + "bbox": [ + 306, + 327, + 480, + 376 + ], + "blocks": [ + { + "bbox": [ + 305, + 299, + 482, + 326 + ], + "lines": [ + { + "bbox": [ + 305, + 299, + 482, + 326 + ], + "spans": [ + { + "bbox": [ + 305, + 299, + 482, + 326 + ], + "type": "text", + "content": "Table 4: Prompt Attachment Analysis (%, ↑): The best prompt attachment choice " + }, + { + "bbox": [ + 305, + 299, + 482, + 326 + ], + "type": "inline_equation", + "content": "f_{p}(.)" + }, + { + "bbox": [ + 305, + 299, + 482, + 326 + ], + "type": "text", + "content": " is Adding #prompts same as #input frames (T=64)." + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 306, + 327, + 480, + 376 + ], + "lines": [ + { + "bbox": [ + 306, + 327, + 480, + 376 + ], + "spans": [ + { + "bbox": [ + 306, + 327, + 480, + 376 + ], + "type": "table", + "html": "
NTU RGB+D\\( \\left\\| u\\mathcal{B}^{(0)}\\right\\| \\)\\( \\left\\| u\\mathcal{S}^{(1)}\\right\\| \\)\\( \\left\\| u\\mathcal{S}^{(2)}\\right\\| \\)\\( \\left\\| u\\mathcal{S}^{(3)}\\right\\| \\)\\( \\left\\| u\\mathcal{S}^{(4)}\\right\\| \\)
MethodBaseAvgAvgAvgOld New Avg \\( A_{HM} \\)
CONCAT temporal, \\( T' = 64 \\)88.670.362.449.833.6 50.5 35.1 40.3
CONCAT feature, \\( T' = 64 \\)87.782.475.566.957.1 41.5 56.0 48.1
Cross Attention, \\( T' = 64 \\)82.977.472.265.057.1 32.3 55.0 41.2
ADD, \\( T' = 1 \\)88.773.362.745.533.7 47.0 34.8 39.3
ADD, \\( T' = 64 \\) (Ours)87.982.876.868.659.3 57.4 59.2 58.3
", + "image_path": "309e07f5e325f2a1338a7e5d3bf32495039abb30d144bfceb68b309f77749c42.jpg" + } + ] + } + ], + "index": 6, + "angle": 0, + "type": "table_body" + } + ], + "index": 6 + }, + { + "bbox": [ + 132, + 405, + 331, + 418 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 405, + 331, + 418 + ], + "spans": [ + { + "bbox": [ + 132, + 405, + 331, + 418 + ], + "type": "text", + "content": "7 Conclusions and Future Work" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 130, + 426, + 482, + 666 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 426, + 482, + 666 + ], + "spans": [ + { + "bbox": [ + 130, + 426, + 482, + 666 + ], + "type": "text", + "content": "The problem of continually adapting human action models to new user categories over time has gained prominence with the rising availability of XR devices. However, this setting poses unique challenges: (i) the user may be able to provide only a few samples for training, and (ii) accessing data from earlier sessions may violate privacy considerations. We hence propose a method based on prompt offset tuning to address this problem in this work. Prompt tuning to address learning over newer tasks has been attempted in recent years. However, these works have: (1) typically been designed for image-based tasks, (2) relied on strongly pretrained transformer backbones, (3) required full supervision for new tasks, and (4) exclusively applied prompt tuning to transformer architectures. This work departs from these four characteristics. Our work demonstrates that prompt offset tuning is a promising option to evolve and adapt skeleton-based human action models to new user classes. The careful design of each component of the proposed methodology finds validation in the promising results across well-known skeleton-based action recognition benchmarks. Our ablation studies and analysis corroborate our design choices in our implementation. Looking ahead, it will be interesting to explore how our approach and its design choices adapt when a \"generalist backbone\" trained on a large corpus of action recognition data becomes accessible. Extending our method for differential privacy is another interesting direction of future work." + } + ] + } + ], + "index": 8 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "spans": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "type": "text", + "content": "14" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 166, + 91, + 222, + 102 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 166, + 91, + 222, + 102 + ], + "spans": [ + { + "bbox": [ + 166, + 91, + 222, + 102 + ], + "type": "text", + "content": "P. Garg et al." + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 13 + }, + { + "para_blocks": [ + { + "bbox": [ + 133, + 114, + 197, + 126 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 114, + 197, + 126 + ], + "spans": [ + { + "bbox": [ + 133, + 114, + 197, + 126 + ], + "type": "text", + "content": "References" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 138, + 140, + 481, + 665 + ], + "type": "list", + "angle": 0, + "index": 19, + "blocks": [ + { + "bbox": [ + 138, + 140, + 481, + 183 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 140, + 481, + 183 + ], + "spans": [ + { + "bbox": [ + 138, + 140, + 481, + 183 + ], + "type": "text", + "content": "1. Aich, S., Ruiz-Santaquiteria, J., Lu, Z., Garg, P., Joseph, K.J., Fernandez, A., Balasubramanian, V.N., Kin, K., Wan, C., Camgoz, N.C., Ma, S., De la Torre, F.: Data-free class-incremental hand gesture recognition. In: Proceedings of the IEEE/CVF International Conference on Computer Vision (2023)" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 138, + 184, + 481, + 205 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 184, + 481, + 205 + ], + "spans": [ + { + "bbox": [ + 138, + 184, + 481, + 205 + ], + "type": "text", + "content": "2. Albrecht, J.P.: How the gdpr will change the world. Eur. Data Prot. L. Rev. 2, 287 (2016)" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 138, + 206, + 481, + 237 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 206, + 481, + 237 + ], + "spans": [ + { + "bbox": [ + 138, + 206, + 481, + 237 + ], + "type": "text", + "content": "3. Bengio, Y., Léonard, N., Courville, A.: Estimating or propagating gradients through stochastic neurons for conditional computation. arXiv preprint arXiv:1308.3432 (2013)" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 138, + 239, + 481, + 282 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 239, + 481, + 282 + ], + "spans": [ + { + "bbox": [ + 138, + 239, + 481, + 282 + ], + "type": "text", + "content": "4. Bowman, B., Achille, A., Zancato, L., Trager, M., Perera, P., Paolini, G., Soatto, S.: a-la-carte prompt tuning (apt): Combining distinct data via composable prompting. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 14984-14993 (2023)" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 138, + 283, + 481, + 315 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 283, + 481, + 315 + ], + "spans": [ + { + "bbox": [ + 138, + 283, + 481, + 315 + ], + "type": "text", + "content": "5. Chaudhry, A., Dokania, P.K., Ajanthan, T., Torr, P.H.: Riemannian walk for incremental learning: Understanding forgetting and intransigence. In: Proceedings of the European conference on computer vision (ECCV). pp. 532-547 (2018)" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 138, + 316, + 481, + 336 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 316, + 481, + 336 + ], + "spans": [ + { + "bbox": [ + 138, + 316, + 481, + 336 + ], + "type": "text", + "content": "6. Chaudhry, A., Ranzato, M., Rohrbach, M., Elhoseiny, M.: Efficient lifelong learning with a-gem. arXiv preprint arXiv:1812.00420 (2018)" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 138, + 338, + 481, + 369 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 338, + 481, + 369 + ], + "spans": [ + { + "bbox": [ + 138, + 338, + 481, + 369 + ], + "type": "text", + "content": "7. Chen, Y., Zhao, L., Peng, X., Yuan, J., Metaxas, D.N.: Construct dynamic graphs for hand gesture recognition via spatial-temporal attention. In: Proceedings of the British Machine Vision Conference (BMVC) (2019)" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 138, + 370, + 481, + 413 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 370, + 481, + 413 + ], + "spans": [ + { + "bbox": [ + 138, + 370, + 481, + 413 + ], + "type": "text", + "content": "8. Chen, Y., Zhang, Z., Yuan, C., Li, B., Deng, Y., Hu, W.: Channel-wise topology refinement graph convolution for skeleton-based action recognition. In: Proceedings of the IEEE/CVF International Conference on Computer Vision. pp. 13359-13368 (2021)" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 138, + 415, + 481, + 435 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 415, + 481, + 435 + ], + "spans": [ + { + "bbox": [ + 138, + 415, + 481, + 435 + ], + "type": "text", + "content": "9. Dhariwal, P., Jun, H., Payne, C., Kim, J.W., Radford, A., Sutskever, I.: Jukebox: A generative model for music. arXiv preprint arXiv:2005.00341 (2020)" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 138, + 436, + 481, + 468 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 436, + 481, + 468 + ], + "spans": [ + { + "bbox": [ + 138, + 436, + 481, + 468 + ], + "type": "text", + "content": "0. Dong, S., Hong, X., Tao, X., Chang, X., Wei, X., Gong, Y.: Few-shot class incremental learning via relation knowledge distillation. In: Proceedings of the AAAI Conference on Artificial Intelligence. vol. 35, pp. 1255-1263 (2021)" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 138, + 469, + 481, + 490 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 469, + 481, + 490 + ], + "spans": [ + { + "bbox": [ + 138, + 469, + 481, + 490 + ], + "type": "text", + "content": "1. Dwivedi, V.P., Bresson, X.: A generalization of transformer networks to graphs. arXiv preprint arXiv:2012.09699 (2020)" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 138, + 491, + 481, + 533 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 491, + 481, + 533 + ], + "spans": [ + { + "bbox": [ + 138, + 491, + 481, + 533 + ], + "type": "text", + "content": "2. Dwivedi, V.P., Luu, A.T., Laurent, T., Bengio, Y., Bresson, X.: Graph neural networks with learnable structural and positional representations. In: International Conference on Learning Representations (2022), https://openreview.net/forum?id=wTTjnvGphYj" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 138, + 535, + 481, + 567 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 535, + 481, + 567 + ], + "spans": [ + { + "bbox": [ + 138, + 535, + 481, + 567 + ], + "type": "text", + "content": "3. Hersche, M., Karunaratne, G., Cherubini, G., Benini, L., Sebastian, A., Rahimi, A.: Constrained few-shot class-incremental learning. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 9057-9067 (2022)" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 138, + 567, + 481, + 599 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 567, + 481, + 599 + ], + "spans": [ + { + "bbox": [ + 138, + 567, + 481, + 599 + ], + "type": "text", + "content": "4. Hinojosa, C., Marquez, M., Arguello, H., Adeli, E., Fei-Fei, L., Niebles, J.C.: Privhar: Recognizing human actions from privacy-preserving lens. In: European Conference on Computer Vision. pp. 314-332. Springer (2022)" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 138, + 600, + 481, + 632 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 600, + 481, + 632 + ], + "spans": [ + { + "bbox": [ + 138, + 600, + 481, + 632 + ], + "type": "text", + "content": "5. Hou, S., Pan, X., Loy, C.C., Wang, Z., Lin, D.: Learning a unified classifier incrementally via rebalancing. In: Proceedings of the IEEE/CVF conference on computer vision and pattern recognition. pp. 831-839 (2019)" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 138, + 633, + 481, + 665 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 633, + 481, + 665 + ], + "spans": [ + { + "bbox": [ + 138, + 633, + 481, + 665 + ], + "type": "text", + "content": "6. Jia, M., Tang, L., Chen, B.C., Cardie, C., Belongie, S., Hariharan, B., Lim, S.N.: Visual prompt tuning. In: European Conference on Computer Vision. pp. 709-727. Springer (2022)" + } + ] + } + ], + "index": 18 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 156, + 91, + 447, + 102 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 156, + 91, + 447, + 102 + ], + "spans": [ + { + "bbox": [ + 156, + 91, + 447, + 102 + ], + "type": "text", + "content": "POET: Prompt Offset Tuning for Continual Human Action Adaptation" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 470, + 92, + 481, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 470, + 92, + 481, + 100 + ], + "spans": [ + { + "bbox": [ + 470, + 92, + 481, + 100 + ], + "type": "text", + "content": "15" + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 14 + }, + { + "para_blocks": [ + { + "bbox": [ + 132, + 117, + 481, + 665 + ], + "type": "list", + "angle": 0, + "index": 19, + "blocks": [ + { + "bbox": [ + 133, + 117, + 481, + 161 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 117, + 481, + 161 + ], + "spans": [ + { + "bbox": [ + 133, + 117, + 481, + 161 + ], + "type": "text", + "content": "17. Kirkpatrick, J., Pascanu, R., Rabinowitz, N., Veness, J., Desjardins, G., Rusu, A.A., Milan, K., Quan, J., Ramalho, T., Grabska-Barwinska, A., et al.: Overcoming catastrophic forgetting in neural networks. Proceedings of the national academy of sciences 114(13), 3521-3526 (2017)" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 133, + 162, + 481, + 194 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 162, + 481, + 194 + ], + "spans": [ + { + "bbox": [ + 133, + 162, + 481, + 194 + ], + "type": "text", + "content": "18. Kumawat, S., Nagahara, H.: Privacy-preserving action recognition via motion difference quantization. In: European Conference on Computer Vision. pp. 518-534. Springer (2022)" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 133, + 194, + 481, + 216 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 194, + 481, + 216 + ], + "spans": [ + { + "bbox": [ + 133, + 194, + 481, + 216 + ], + "type": "text", + "content": "19. Lacoste, A., Luccioni, A., Schmidt, V., Dandes, T.: Quantifying the carbon emissions of machine learning. arXiv preprint arXiv:1910.09700 (2019)" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 132, + 216, + 481, + 237 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 216, + 481, + 237 + ], + "spans": [ + { + "bbox": [ + 132, + 216, + 481, + 237 + ], + "type": "text", + "content": "20. Lester, B., Al-Rfou, R., Constant, N.: The power of scale for parameter-efficient prompt tuning. arXiv preprint arXiv:2104.08691 (2021)" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 132, + 237, + 481, + 281 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 237, + 481, + 281 + ], + "spans": [ + { + "bbox": [ + 132, + 237, + 481, + 281 + ], + "type": "text", + "content": "21. Li, M., Xu, X., Fan, H., Zhou, P., Liu, J., Liu, J.W., Li, J., Keppo, J., Shou, M.Z., Yan, S.: Stprivacy: Spatio-temporal privacy-preserving action recognition. In: Proceedings of the IEEE/CVF International Conference on Computer Vision. pp. 5106-5115 (2023)" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 132, + 281, + 481, + 315 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 281, + 481, + 315 + ], + "spans": [ + { + "bbox": [ + 132, + 281, + 481, + 315 + ], + "type": "text", + "content": "22. Li, T., Ke, Q., Rahmani, H., Ho, R.E., Ding, H., Liu, J.: Else-net: Elastic semantic network for continual action recognition from skeleton data. In: Proceedings of the IEEE/CVF International Conference on Computer Vision. pp. 13434-13443 (2021)" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 132, + 315, + 481, + 335 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 315, + 481, + 335 + ], + "spans": [ + { + "bbox": [ + 132, + 315, + 481, + 335 + ], + "type": "text", + "content": "23. Li, X.L., Liang, P.: Prefix-tuning: Optimizing continuous prompts for generation. arXiv preprint arXiv:2101.00190 (2021)" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 132, + 336, + 481, + 369 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 336, + 481, + 369 + ], + "spans": [ + { + "bbox": [ + 132, + 336, + 481, + 369 + ], + "type": "text", + "content": "24. Li, Y., Si, S., Li, G., Hsieh, C.J., Bengio, S.: Learnable fourier features for multidimensional spatial positional encoding. Advances in Neural Information Processing Systems 34, 15816-15829 (2021)" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 132, + 369, + 481, + 391 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 369, + 481, + 391 + ], + "spans": [ + { + "bbox": [ + 132, + 369, + 481, + 391 + ], + "type": "text", + "content": "25. Li, Z., Hoiem, D.: Learning without forgetting. IEEE transactions on pattern analysis and machine intelligence 40(12), 2935-2947 (2017)" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 132, + 392, + 481, + 424 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 392, + 481, + 424 + ], + "spans": [ + { + "bbox": [ + 132, + 392, + 481, + 424 + ], + "type": "text", + "content": "26. Liu, X., Yu, H.F., Dhillon, I., Hsieh, C.J.: Learning to encode position for transformer with continuous dynamical model. In: International conference on machine learning. pp. 6327-6335. PMLR (2020)" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 132, + 424, + 481, + 457 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 424, + 481, + 457 + ], + "spans": [ + { + "bbox": [ + 132, + 424, + 481, + 457 + ], + "type": "text", + "content": "27. Ma, N., Zhang, H., Li, X., Zhou, S., Zhang, Z., Wen, J., Li, H., Gu, J., Bu, J.: Learning spatial-preserved skeleton representations for few-shot action recognition. In: European Conference on Computer Vision. pp. 174-191. Springer (2022)" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 132, + 457, + 481, + 490 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 457, + 481, + 490 + ], + "spans": [ + { + "bbox": [ + 132, + 457, + 481, + 490 + ], + "type": "text", + "content": "28. Mallya, A., Lazebnik, S.: Packnet: Adding multiple tasks to a single network by iterative pruning. In: Proceedings of the IEEE conference on Computer Vision and Pattern Recognition. pp. 7765-7773 (2018)" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 132, + 490, + 481, + 511 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 490, + 481, + 511 + ], + "spans": [ + { + "bbox": [ + 132, + 490, + 481, + 511 + ], + "type": "text", + "content": "29. Mialon, G., Chen, D., Selosse, M., Mairal, J.: Graphit: Encoding graph structure in transformers. arXiv preprint arXiv:2106.05667 (2021)" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 132, + 511, + 481, + 555 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 511, + 481, + 555 + ], + "spans": [ + { + "bbox": [ + 132, + 511, + 481, + 555 + ], + "type": "text", + "content": "30. Park, J.J., Florence, P., Straub, J., Newcombe, R., Lovegrove, S.: Deepsdf: Learning continuous signed distance functions for shape representation. In: Proceedings of the IEEE/CVF conference on computer vision and pattern recognition. pp. 165-174 (2019)" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 132, + 555, + 481, + 588 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 555, + 481, + 588 + ], + "spans": [ + { + "bbox": [ + 132, + 555, + 481, + 588 + ], + "type": "text", + "content": "31. Peng, C., Zhao, K., Wang, T., Li, M., Lovell, B.C.: Few-shot class-incremental learning from an open-set perspective. In: European Conference on Computer Vision. pp. 382-397. Springer (2022)" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 132, + 588, + 481, + 621 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 588, + 481, + 621 + ], + "spans": [ + { + "bbox": [ + 132, + 588, + 481, + 621 + ], + "type": "text", + "content": "32. Pernici, F., Bruni, M., Baecchi, C., Turchini, F., Del Bimbo, A.: Class-incremental learning with pre-allocated fixed classifiers. In: 2020 25th International Conference on Pattern Recognition (ICPR). pp. 6259–6266. IEEE (2021)" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 132, + 621, + 481, + 665 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 621, + 481, + 665 + ], + "spans": [ + { + "bbox": [ + 132, + 621, + 481, + 665 + ], + "type": "text", + "content": "33. Radford, A., Kim, J.W., Hallacy, C., Ramesh, A., Goh, G., Agarwal, S., Sastry, G., Askell, A., Mishkin, P., Clark, J., et al.: Learning transferable visual models from natural language supervision. In: International conference on machine learning. pp. 8748-8763. PMLR (2021)" + } + ] + } + ], + "index": 18 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "spans": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "type": "text", + "content": "16" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 166, + 91, + 222, + 102 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 166, + 91, + 222, + 102 + ], + "spans": [ + { + "bbox": [ + 166, + 91, + 222, + 102 + ], + "type": "text", + "content": "P. Garg et al." + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 15 + }, + { + "para_blocks": [ + { + "bbox": [ + 130, + 116, + 482, + 665 + ], + "type": "list", + "angle": 0, + "index": 18, + "blocks": [ + { + "bbox": [ + 130, + 116, + 482, + 149 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 116, + 482, + 149 + ], + "spans": [ + { + "bbox": [ + 130, + 116, + 482, + 149 + ], + "type": "text", + "content": "34. Razdaibiedina, A., Mao, Y., Hou, R., Khabsa, M., Lewis, M., Almahairi, A.: Progressive prompts: Continual learning for language models. arXiv preprint arXiv:2301.12314 (2023)" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 130, + 149, + 482, + 182 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 149, + 482, + 182 + ], + "spans": [ + { + "bbox": [ + 130, + 149, + 482, + 182 + ], + "type": "text", + "content": "35. Rebuffi, S.A., Bilen, H., Vedaldi, A.: Efficient parametrization of multi-domain deep neural networks. In: Proceedings of the IEEE conference on computer vision and pattern recognition. pp. 8119-8127 (2018)" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 130, + 182, + 482, + 203 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 182, + 482, + 203 + ], + "spans": [ + { + "bbox": [ + 130, + 182, + 482, + 203 + ], + "type": "text", + "content": "36. Ren, B., Liu, M., Ding, R., Liu, H.: A survey on 3d skeleton-based action recognition using learning method. Cyborg and Bionic Systems (2020)" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 130, + 203, + 482, + 224 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 203, + 482, + 224 + ], + "spans": [ + { + "bbox": [ + 130, + 203, + 482, + 224 + ], + "type": "text", + "content": "37. Ridnik, T., Ben-Baruch, E., Noy, A., Zelnik-Manor, L.: Imagenet-21k pretraining for the masses. arXiv preprint arXiv:2104.10972 (2021)" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 130, + 224, + 482, + 256 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 224, + 482, + 256 + ], + "spans": [ + { + "bbox": [ + 130, + 224, + 482, + 256 + ], + "type": "text", + "content": "38. Rusu, A.A., Rabinowitz, N.C., Desjardins, G., Soyer, H., Kirkpatrick, J., Kavukcuoglu, K., Pascanu, R., Hadsell, R.: Progressive neural networks. arXiv preprint arXiv:1606.04671 (2016)" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 130, + 256, + 482, + 289 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 256, + 482, + 289 + ], + "spans": [ + { + "bbox": [ + 130, + 256, + 482, + 289 + ], + "type": "text", + "content": "39. Shahroudy, A., Liu, J., Ng, T.T., Wang, G.: Ntu rgb+ d: A large scale dataset for 3d human activity analysis. In: Proceedings of the IEEE conference on computer vision and pattern recognition. pp. 1010-1019 (2016)" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 130, + 289, + 482, + 331 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 289, + 482, + 331 + ], + "spans": [ + { + "bbox": [ + 130, + 289, + 482, + 331 + ], + "type": "text", + "content": "40. Smedt, Q.D., Wannous, H., Vandeborre, J.P., Guerry, J., Saux, B.L., Filliat, D.: 3D Hand Gesture Recognition Using a Depth and Skeletal Dataset. In: Pratikakis, I., Dupont, F., Ovsjanikov, M. (eds.) Eurographics Workshop on 3D Object Retrieval. The Eurographics Association (2017). https://doi.org/10.2312/3dor.20171049" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 130, + 331, + 482, + 385 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 331, + 482, + 385 + ], + "spans": [ + { + "bbox": [ + 130, + 331, + 482, + 385 + ], + "type": "text", + "content": "41. Smith, J.S., Karlinsky, L., Gutta, V., Cascante-Bonilla, P., Kim, D., Arbelle, A., Panda, R., Feris, R., Kira, Z.: Coda-prompt: Continual decomposed attention-based prompting for rehearsal-free continual learning. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 11909-11919 (2023)" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 130, + 385, + 482, + 407 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 385, + 482, + 407 + ], + "spans": [ + { + "bbox": [ + 130, + 385, + 482, + 407 + ], + "type": "text", + "content": "42. Tang, Y.M., Peng, Y.X., Zheng, W.S.: When prompt-based incremental learning does not meet strong pretraining. arXiv preprint arXiv:2308.10445 (2023)" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 130, + 407, + 482, + 438 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 407, + 482, + 438 + ], + "spans": [ + { + "bbox": [ + 130, + 407, + 482, + 438 + ], + "type": "text", + "content": "43. Tao, X., Hong, X., Chang, X., Dong, S., Wei, X., Gong, Y.: Few-shot class-incremental learning. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 12183-12192 (2020)" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 130, + 438, + 482, + 460 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 438, + 482, + 460 + ], + "spans": [ + { + "bbox": [ + 130, + 438, + 482, + 460 + ], + "type": "text", + "content": "44. Van Den Oord, A., Vinyals, O., et al.: Neural discrete representation learning. Advances in neural information processing systems 30 (2017)" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 130, + 460, + 482, + 502 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 460, + 482, + 502 + ], + "spans": [ + { + "bbox": [ + 130, + 460, + 482, + 502 + ], + "type": "text", + "content": "45. Villa, A., Alcazar, J.L., Alfarra, M., Alhamoud, K., Hurtado, J., Heilbron, F.C., Soto, A., Ghanem, B.: Pivot: Prompting for video continual learning. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 24214-24223 (2023)" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 130, + 502, + 482, + 546 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 502, + 482, + 546 + ], + "spans": [ + { + "bbox": [ + 130, + 502, + 482, + 546 + ], + "type": "text", + "content": "46. Wang, X., Zhang, S., Qing, Z., Gao, C., Zhang, Y., Zhao, D., Sang, N.: Molo: Motion-augmented long-short contrastive learning for few-shot action recognition. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 18011-18021 (2023)" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 130, + 546, + 482, + 578 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 546, + 482, + 578 + ], + "spans": [ + { + "bbox": [ + 130, + 546, + 482, + 578 + ], + "type": "text", + "content": "47. Wang, Y., Huang, Z., Hong, X.: S-prompts learning with pre-trained transformers: An occam's razor for domain incremental learning. Advances in Neural Information Processing Systems 35, 5682-5695 (2022)" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 130, + 578, + 482, + 621 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 578, + 482, + 621 + ], + "spans": [ + { + "bbox": [ + 130, + 578, + 482, + 621 + ], + "type": "text", + "content": "48. Wang, Z., Zhang, Z., Ebrahimi, S., Sun, R., Zhang, H., Lee, C.Y., Ren, X., Su, G., Perot, V., Dy, J., et al.: Dualprompt: Complementary prompting for rehearsal-free continual learning. In: European Conference on Computer Vision. pp. 631-648. Springer (2022)" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 130, + 621, + 482, + 665 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 621, + 482, + 665 + ], + "spans": [ + { + "bbox": [ + 130, + 621, + 482, + 665 + ], + "type": "text", + "content": "49. Wang, Z., Zhang, Z., Lee, C.Y., Zhang, H., Sun, R., Ren, X., Su, G., Perot, V., Dy, J., Pfister, T.: Learning to prompt for continual learning. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 139-149 (2022)" + } + ] + } + ], + "index": 17 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 156, + 91, + 447, + 102 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 156, + 91, + 447, + 102 + ], + "spans": [ + { + "bbox": [ + 156, + 91, + 447, + 102 + ], + "type": "text", + "content": "POET: Prompt Offset Tuning for Continual Human Action Adaptation" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 470, + 92, + 481, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 470, + 92, + 481, + 100 + ], + "spans": [ + { + "bbox": [ + 470, + 92, + 481, + 100 + ], + "type": "text", + "content": "17" + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 16 + }, + { + "para_blocks": [ + { + "bbox": [ + 132, + 116, + 482, + 381 + ], + "type": "list", + "angle": 0, + "index": 10, + "blocks": [ + { + "bbox": [ + 132, + 116, + 482, + 149 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 116, + 482, + 149 + ], + "spans": [ + { + "bbox": [ + 132, + 116, + 482, + 149 + ], + "type": "text", + "content": "50. Williams, W., Ringer, S., Ash, T., MacLeod, D., Dougherty, J., Hughes, J.: Hierarchical quantized autoencoders. Advances in Neural Information Processing Systems 33, 4524-4535 (2020)" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 132, + 150, + 482, + 194 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 150, + 482, + 194 + ], + "spans": [ + { + "bbox": [ + 132, + 150, + 482, + 194 + ], + "type": "text", + "content": "51. Yang, Y., Yuan, H., Li, X., Lin, Z., Torr, P., Tao, D.: Neural collapse inspired feature-classifier alignment for few-shot class-incremental learning. In: The Eleventh International Conference on Learning Representations (2023), https://openreview.net/forum?id=y5W8tpojhtJ" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 132, + 194, + 481, + 216 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 194, + 481, + 216 + ], + "spans": [ + { + "bbox": [ + 132, + 194, + 481, + 216 + ], + "type": "text", + "content": "52. Yue, R., Tian, Z., Du, S.: Action recognition based on rgb and skeleton data sets: A survey. Neurocomputing (2022)" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 132, + 216, + 481, + 248 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 216, + 481, + 248 + ], + "spans": [ + { + "bbox": [ + 132, + 216, + 481, + 248 + ], + "type": "text", + "content": "53. Zhang, H.B., Zhang, Y.X., Zhong, B., Lei, Q., Yang, L., Du, J.X., Chen, D.S.: A comprehensive survey of vision-based human action recognition methods. Sensors 19(5), 1005 (2019)" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 132, + 249, + 481, + 270 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 249, + 481, + 270 + ], + "spans": [ + { + "bbox": [ + 132, + 249, + 481, + 270 + ], + "type": "text", + "content": "54. Zheng, C., Vedaldi, A.: Online clustered codebook. In: Proceedings of the IEEE/CVF International Conference on Computer Vision. pp. 22798-22807 (2023)" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 132, + 271, + 481, + 304 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 271, + 481, + 304 + ], + "spans": [ + { + "bbox": [ + 132, + 271, + 481, + 304 + ], + "type": "text", + "content": "55. Zhou, D.W., Wang, F.Y., Ye, H.J., Ma, L., Pu, S., Zhan, D.C.: Forward compatible few-shot class-incremental learning. In: Proceedings of the IEEE/CVF conference on computer vision and pattern recognition. pp. 9046-9056 (2022)" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 132, + 304, + 481, + 346 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 304, + 481, + 346 + ], + "spans": [ + { + "bbox": [ + 132, + 304, + 481, + 346 + ], + "type": "text", + "content": "56. Zhu, A., Ke, Q., Gong, M., Bailey, J.: Adaptive local-component-aware graph convolutional network for one-shot skeleton-based action recognition. In: Proceedings of the IEEE/CVF Winter Conference on Applications of Computer Vision. pp. 6038-6047 (2023)" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 132, + 347, + 481, + 381 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 347, + 481, + 381 + ], + "spans": [ + { + "bbox": [ + 132, + 347, + 481, + 381 + ], + "type": "text", + "content": "57. Zhu, B., Niu, Y., Han, Y., Wu, Y., Zhang, H.: Prompt-aligned gradient for prompt tuning. In: Proceedings of the IEEE/CVF International Conference on Computer Vision. pp. 15659-15669 (2023)" + } + ] + } + ], + "index": 9 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "spans": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "type": "text", + "content": "18" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 166, + 91, + 222, + 102 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 166, + 91, + 222, + 102 + ], + "spans": [ + { + "bbox": [ + 166, + 91, + 222, + 102 + ], + "type": "text", + "content": "P. Garg et al." + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 17 + } + ], + "_backend": "vlm", + "_version_name": "2.6.4" +} \ No newline at end of file diff --git a/2024/PPAD_ Iterative Interactions of Prediction and Planning for End-to-end Autonomous Driving/2548ffc2-3e35-4690-aa4a-52b1b19d7a46_content_list.json b/2024/PPAD_ Iterative Interactions of Prediction and Planning for End-to-end Autonomous Driving/2548ffc2-3e35-4690-aa4a-52b1b19d7a46_content_list.json new file mode 100644 index 0000000000000000000000000000000000000000..3ec031bf4e33cc9824926de515aedd959a352db9 --- /dev/null +++ b/2024/PPAD_ Iterative Interactions of Prediction and Planning for End-to-end Autonomous Driving/2548ffc2-3e35-4690-aa4a-52b1b19d7a46_content_list.json @@ -0,0 +1,1728 @@ +[ + { + "type": "text", + "text": "PPAD: Iterative Interactions of Prediction and Planning for End-to-end Autonomous Driving", + "text_level": 1, + "bbox": [ + 233, + 140, + 772, + 186 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Zhili Chen $^{1\\dagger}$ , Maosheng Ye $^{1}$ , Shuangjie Xu $^{1}$ , Tongyi Cao $^{2}$ , and Qifeng Chen $^{1\\boxtimes}$", + "bbox": [ + 321, + 210, + 679, + 244 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "$^{1}$ HKUST $^{2}$ DeepRoute.AI {zchenei, myeag, shuangjie.xu}@connect.ust.hk, tongyicao@deeproute.ai, cqf@cse.ust.hk", + "bbox": [ + 323, + 253, + 679, + 297 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Abstract. We present a new interaction mechanism of prediction and planning for end-to-end autonomous driving, called PPAD (Iterative Interaction of Prediction and Planning Autonomous Driving), which considers the timestep-wise interaction to better integrate prediction and planning. An ego vehicle performs motion planning at each timestep based on the trajectory prediction of surrounding agents (e.g., vehicles and pedestrians) and its local road conditions. Unlike existing end-to-end autonomous driving frameworks, PPAD models the interactions among ego, agents, and the dynamic environment in an autoregressive manner by interleaving the Prediction and Planning processes at every timestep, instead of a single sequential process of prediction followed by planning. Specifically, we design ego-to-agent, ego-to-map, and ego-to-BEV interaction mechanisms with hierarchical dynamic key objects attention to better model the interactions. The experiments on the nuScenes benchmark show that our approach outperforms state-of-the-art methods. Project page at https://github.com/zlichen/PPAD.", + "bbox": [ + 259, + 337, + 743, + 561 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Keywords: End-to-end Autonomous Driving", + "bbox": [ + 261, + 573, + 571, + 587 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "1 Introduction", + "text_level": 1, + "bbox": [ + 215, + 614, + 375, + 630 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "The blossom of deep learning techniques has empowered autonomous driving, where many exciting milestones in autonomous driving have burst into our eyes owing to the convenient and interpretable discrete module designs. Recently, the planning-oriented [19] philosophy resonated with the community for pursuing a more effective end-to-end driving system, which is the focus of this work.", + "bbox": [ + 212, + 648, + 787, + 723 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Traditional methods in an autonomous driving system often break down the system into modular components, including localization, perception, tracking, prediction, planning, and control for interpretability and visibility. However, there are several drawbacks: 1) the accumulation of errors between modules becomes more significant as the system complexity increases. 2) the performance of", + "bbox": [ + 212, + 724, + 787, + 801 + ], + "page_idx": 0 + }, + { + "type": "page_footnote", + "text": "$^{\\dagger}$ Work done during an internship at DeepRoute.AI.", + "bbox": [ + 230, + 810, + 576, + 825 + ], + "page_idx": 0 + }, + { + "type": "page_footnote", + "text": "Corresponding author.", + "bbox": [ + 230, + 825, + 393, + 839 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/8e37dc15643bdf9d25811804db201a98863293da2849adb8288c0d85806510a8.jpg", + "image_caption": [ + "(a) Traditional one-shot approach." + ], + "image_footnote": [], + "bbox": [ + 217, + 143, + 408, + 196 + ], + "page_idx": 1 + }, + { + "type": "image", + "img_path": "images/e52a0eacb7fb6b059e97b0354127e1b1ed5780d7d8e706a927ef443b79522dee.jpg", + "image_caption": [ + "(b) Our proposed PPAD framework.", + "Fig. 1: A high-level illustration of our proposed PPAD framework. The agent (in blue) intends to drive straight, while the ego (in red) plans to change lanes. Fig. 1(a) presents the typical one-shot method that might result in invalid motion plans and lead to an accident because of a lack of in-depth interactions. Fig. 1(b) demonstrates the gaming process between the ego and the agent under the PPAD architecture. During the prediction process, the agent executes an assertive plan by accelerating to stop the ego from blocking its route. The planning process of the ego plans trajectory based on the previous prediction process of the agent. The ego decelerates to avoid a potential accident and then changes lanes to achieve its driving goal." + ], + "image_footnote": [], + "bbox": [ + 411, + 143, + 784, + 196 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "the downstream task is highly related to the upper stream module, which makes it very difficult to construct a unified data-driven infrastructure.", + "bbox": [ + 212, + 381, + 784, + 411 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Recently, end-to-end autonomous driving has gained popularity due to its simplicity. Two main lines are proposed based on the learning architecture. The first kind of method [10] takes the raw sensor data as input and directly outputs the planning trajectories or control command without any view transformation as intermediate representations for scene understanding. The other kinds of approaches [19,23] are built upon BEV representation and fully utilize the queries to generate the intermediate outputs as guidance for producing the planning results. One of the most significant advantages lies in the interpretability. In this work, we follow the design of the second kind of work.", + "bbox": [ + 212, + 412, + 787, + 549 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "VAD [23] and UniAD [19] are typical one-shot motion planning methods, which only consider a single-step interaction between agents, ego-agent and surrounding environment (e.g., map elements). Think Twice [22] makes it a two-stage framework to enhance the gaming or interaction procedure. QCNet [56] and GameFormer [21] also recurrently model the trajectory prediction task. Given motion planning is a computational problem that finds a sequence of valid trajectories, often based on surrounding agents' forecasting, environmental understanding, and historical and future contexts. It can also be viewed as a game in which agents continuously plan their next move according to other agents' intentions and the encountering environment, further achieving their ultimate goals through incremental actions. To model these dynamic interactions of prediction and planning in end-to-end autonomous driving, it is crucial to consider the possible variance of predicted trajectories through multi-step modeling for planning feasible trajectories.", + "bbox": [ + 212, + 551, + 787, + 763 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Inspired by VAD [23], we aim to introduce the step-by-step Prediction-Planning into a learning-based framework. Intuitively, the prediction and planning modules can be modeled as a motion forecasting task, which predicts future waypoints by the given historical information. The results of prediction and planning modules at each time step are highly dependent on each other. Therefore,", + "bbox": [ + 212, + 763, + 787, + 840 + ], + "page_idx": 1 + }, + { + "type": "page_number", + "text": "2", + "bbox": [ + 217, + 114, + 228, + 126 + ], + "page_idx": 1 + }, + { + "type": "header", + "text": "Z. Chen et al.", + "bbox": [ + 271, + 114, + 364, + 126 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "we need to consider the agent-agent and agent-environment interactions iteratively and bidirectionally to maximize the expectation of agents' prediction under the given observation of the other agents. We propose our PPAD to plan the ego agent's future trajectories step-by-step to model the timestep-wise bidirectional interaction or gaming in a vectorized learning framework as shown in Fig. 1. PPAD consists of the prediction and planning process. For each motion forecasting step, 1) Prediction process generates current step motion states by cross-attention and self-attention among agents and environment based on previous motion states to model the fine-grained bidirectional interactions. We take ego-agent-environment-BEV interaction into account to propagate features among all the traffic participants. 2) Planning process predicts the current step motion trajectories based on the expectation process. Our contributions are summarized as follows:", + "bbox": [ + 212, + 146, + 787, + 342 + ], + "page_idx": 2 + }, + { + "type": "list", + "sub_type": "text", + "list_items": [ + "- We propose PPAD that optimizes ego-agent-environment interactions in an iterative prediction-planning manner. Iterative optimization could model the interactions and gaming better and more naturally in a planning task. The prediction process deals with more fine-grained and complex future uncertainties for multi-agent context learning, while the planning process plans a one-step future trajectory for the ego vehicle.", + "- We model fine-grained interactions among the ego vehicle, agents, environment, and BEV features map, step-by-step with hierarchical dynamic key objects attention emphasizing on the spatial locality.", + "- The experiments conducted on the nuScenes [3] and Argoverse [6,44] datasets have demonstrated the effectiveness of our approach over state-of-the-art approaches." + ], + "bbox": [ + 225, + 354, + 782, + 532 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "2 Related Work", + "text_level": 1, + "bbox": [ + 215, + 556, + 387, + 573 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "2.1 Multi-stage Autonomous Driving", + "text_level": 1, + "bbox": [ + 215, + 589, + 535, + 604 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Most autonomous driving systems are built upon the multi-stage design philosophy, which commonly consists of localization, perception, and planning. The perception module has been well studied recently due to the emergence of deep learning. Camera-based [20, 28, 29], Lidar-based [26, 48-50, 55] or fused-based [36, 46] approaches are proposed to fully exploit the potential of raw sensor data in order to produce accurate 3D objects prediction, semantic segmentation or tracking velocity. Prediction takes the outputs of the perception module to generate the future waypoints for the ego and the agents. Current approaches [1, 5, 11-13, 16, 30, 32, 40, 42, 48, 52, 56, 57] explore different representations to encode surrounding environment (map information) and agent interactions to predict final trajectories by regression or postprocessing sampling strategies. Some other works [4, 37] propose a joint perception and prediction framework, which aggregates historical information to generate tracklets with future trajectories. This unified learning framework could help address the nondifferential process and alleviate the unstable perception problem, compared", + "bbox": [ + 212, + 613, + 787, + 840 + ], + "page_idx": 2 + }, + { + "type": "header", + "text": "PPAD", + "bbox": [ + 686, + 114, + 730, + 126 + ], + "page_idx": 2 + }, + { + "type": "page_number", + "text": "3", + "bbox": [ + 774, + 114, + 785, + 126 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "with previous works. Based on the perception and prediction results, planning module [1,7,12,39] plan its future behavior by cost-map optimization or learning-based approaches.", + "bbox": [ + 215, + 146, + 784, + 191 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "2.2 End-To-End Autonomous Driving", + "text_level": 1, + "bbox": [ + 215, + 212, + 540, + 228 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Recently, more and more works have focused on end-to-end autonomous driving due to its merits in reducing internal accumulative errors and direct yet simple learning objectives. Typical methods [2,45] take the raw visual inputs to regress the final control command or trajectory points without any view transformations. To embrace reinforcement learning, a series of following works [8, 45, 54] utilize policy-based or valued-based approaches to improve driving behavior. With the popularity of BEV representation, more advanced architectures [9, 22] are introduced to attach more interpretability and help deal with the complex interactions in the driving scenarios. Another merit of BEV representation lies in its simplicity in fusing multi-modality sensors. Moreover, the modularized approaches [14, 18, 19, 23] decouple the end-to-end learning-based methods into several submodules or subtasks, while in a multi-task learning manner. The unified design could propagate and share learning context between modules through queries or feature maps.", + "bbox": [ + 215, + 236, + 785, + 448 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "2.3 BEV Representation", + "text_level": 1, + "bbox": [ + 215, + 468, + 431, + 484 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "BEV representation has gained significant prominence in the field of autonomous driving systems due to its inherent distortion-free characteristics and its simplicity in facilitating multi-sensor fusion. There are two main lines for BEV representation, including bottom-up and top-down ways. LSS [38] stands out as a bottom-up pioneering work that explores depth distribution for the 3D space frustum sampling to form BEV representation. Works [20, 28] optimize the pipeline through better depth estimation or lightweight sampling design. BEVFormer [29] and its following works [20, 33-35, 43, 47] adopt the top-down architecture, which uses a deformable transformer for the view transformations without depth supervision.", + "bbox": [ + 215, + 492, + 785, + 643 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "3 Method", + "text_level": 1, + "bbox": [ + 215, + 665, + 330, + 681 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "3.1 Framework Overview", + "text_level": 1, + "bbox": [ + 215, + 695, + 437, + 709 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "We present the overall framework, PPAD, in Fig. 2, which comprises the principal modules of the Perception Transformer and our proposed Iterative Prediction-Planning Module. The Perception Transformer encodes the scene contexts into the BEV features map and further decodes as vectorized agents and map representations. The Iterative Prediction-Planning module consists of Prediction and Planning process in general. It dissects the dynamic interactions between the ego vehicle and the agents along the temporal dimension. Eventually, it predicts the motions of the agents and plans the future trajectory for the ego vehicle.", + "bbox": [ + 215, + 719, + 785, + 839 + ], + "page_idx": 3 + }, + { + "type": "page_number", + "text": "4", + "bbox": [ + 217, + 114, + 228, + 126 + ], + "page_idx": 3 + }, + { + "type": "header", + "text": "Z. Chen et al.", + "bbox": [ + 271, + 114, + 364, + 126 + ], + "page_idx": 3 + }, + { + "type": "image", + "img_path": "images/8e97feb90329d1f49e53f7e11f45c68f3e7f7caa7872baacdc6c73c24e31b0d1.jpg", + "image_caption": [ + "Fig. 2: Overall architecture of our proposed self-driving framework, PPAD. It consists of the Perception Transformer and the Iterative Prediction-Planning Module. The Perception Transformer encodes scene contexts into agent queries, map queries, and BEV queries. Then, the Prediction-Planning Module interleaves the processes of the agent motion prediction and the ego planning for $N$ times. Throughout the iterative Prediction and Planning processes, in-depth interactions are conducted among the ego, agents, map elements, and BEV features. In the Prediction process, the agent initially intends to go straight and is unaware of the potential motion of the ego. After interacting with the ego, map elements, and BEV features, the agent plans to be assertive and proceeds to accelerate. In the following Planning process, the ego knows the agent will accelerate through interacting with the updated agent query. It eventually plans to decelerate first and then conduct the lane change for safety reasons." + ], + "image_footnote": [], + "bbox": [ + 217, + 142, + 787, + 388 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Image Features Module uses a shared image backbone network (e.g., ResNet [15]) to extract image features for separate camera views.", + "bbox": [ + 212, + 595, + 784, + 626 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "BEV Features Module transform the semantic features from the multi-view cameras into a united bird's-eye-view. Specifically, we inherit the encoder from BEVFormer [29, 47] to construct the BEV features. The grid-shape learnable BEV queries $\\mathbf{B} \\in \\mathbb{R}^{H \\times W \\times C}$ are randomly initialized and learned to interact with the multi-view image features through deformable attention [58] to conduct spatial modeling. Temporal modeling is conducted in a recurrent manner, which applies the deformable attention between the current frame's BEV queries and the one from the previous time step.", + "bbox": [ + 212, + 627, + 787, + 747 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Vectorized Features Module Inspired by the VAD [23] paradigm, we also encode the scene contexts into vectorized representations through a detection decoder head [29, 58] and a map element decoding head [31], resulting in $N_{A}$ of learned agent queries $\\mathbf{A} \\in \\mathbb{R}^{N_A \\times C}$ and $N_{M}$ of learned map queries $\\mathbf{M} \\in \\mathbb{R}^{N_M \\times C}$ . Separate MLP-based decoders will be attached to produce side output, which takes the learned queries as inputs and predicts with the agent attributes (lo", + "bbox": [ + 212, + 750, + 787, + 840 + ], + "page_idx": 4 + }, + { + "type": "header", + "text": "PPAD", + "bbox": [ + 686, + 114, + 730, + 126 + ], + "page_idx": 4 + }, + { + "type": "page_number", + "text": "5", + "bbox": [ + 774, + 116, + 784, + 126 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "cations, dimensions, classes, etc.) or map attributes (classes and map vectors described by points). Additionally, the agent queries will be combined with the learnable motion embeddings for modeling the diverse motions of the agents. The agents with motions are represented as $\\mathbf{A} \\in \\mathbb{R}^{N_A \\times N_A^{mot} \\times C}$ . Similarly, the ego vehicle is modeled with three modes, representing the high-level driving commands of going straight, turn left, and turn right, in the form of $\\mathbf{E} \\in \\mathbb{R}^{N_E \\times N_E^{mot} \\times C}$ .", + "bbox": [ + 212, + 146, + 787, + 236 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Iterative Prediction-Planning Module predicts the future trajectories for the ego vehicle and the agents in an interleaved fashion. Different from the traditional practice that predicts all the trajectories in one go, our PPAD framework articulates each step of motion planning by iterating the agent motion prediction and the ego planning processes. Thanks to the PPAD framework, we can conduct in-depth design to enforce key objects interactions (in Sec. 3.3) on scene contexts in a coarse-to-fine manner. We further improve the driving performance for the ego vehicle by taking the noisy trajectory as each step prediction and training the PPAD framework to reconstruct its original position at the following time step (in Sec. 3.4).", + "bbox": [ + 212, + 237, + 787, + 388 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "3.2 Iterative Interactions of Prediction and Planning", + "text_level": 1, + "bbox": [ + 214, + 410, + 665, + 425 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "In the real world, the driving traffic changes constantly. Drivers plan and execute their decisions by ceaselessly reasoning the relationships among traffic participants in the scene. The planning task requires the self-driving system to have a good understanding of the scene and be capable of resolving the spatial-temporal causal factors. Therefore, we innovate the PPAD to dissect the planning task into multi-steps of the agent prediction and the ego planning processes and eventually promote consensuses among the ego's and the agents' future trajectories. The PPAD framework embodies the traffic interactions as gaming along space-time, producing a more accurate planning trajectory for the ego vehicle.", + "bbox": [ + 212, + 433, + 787, + 569 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Specifically, the ego and agents inherit the same philosophy of alternatively optimizing their motion behavior based on each other's motion forecasting at each future time step. In the following section, we will demonstrate the agent prediction process and elaborate on the details of the ego planning process.", + "bbox": [ + 212, + 570, + 787, + 631 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Prediction Process As illustrated in Fig 2, the agent predicts its subsequent step motion during the Prediction process, conditioned on the output of the ego vehicle's outcome from the previous Planning process. Specifically, the initial state of the agent query comprises its driving intention. It will then interact with the ego query updated from the previous planning process, which indicates the latest driving plan of the ego vehicle. After that, it will interact with map elements to choose the driving paths. At last, it gathers detailed geometric information by interacting with the BEV features and comes up with its precise next-step movement.", + "bbox": [ + 212, + 651, + 787, + 787 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Planning Process We consider the period of history with $T_{obs}$ steps and the future with $T_{fut}$ steps. The future trajectory of ego is denoted as $\\{p_E^t\\}^{t \\in T_{fut}}$ .", + "bbox": [ + 214, + 809, + 787, + 842 + ], + "page_idx": 5 + }, + { + "type": "page_number", + "text": "6", + "bbox": [ + 217, + 114, + 228, + 126 + ], + "page_idx": 5 + }, + { + "type": "header", + "text": "Z. Chen et al.", + "bbox": [ + 271, + 114, + 364, + 126 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "For each agent $a \\in \\mathbf{A}$ , the trajectory is represented as $\\{p_a^t\\}^{t \\in T_{fut}}$ . The positions of the detected map elements are denoted as $\\{p_m\\}_{m \\in \\mathbf{M}}$ . We define the operator of $\\mathcal{M}(p_E^t, p_{\\mathbf{A}} / p_{\\mathbf{M}}, s)$ to mask out the agents or map elements that are beyond a distance of $s$ towards the ego located at $p_E^t$ to conduct key objects attention, and we will discuss the algorithm details in Sec. 3.3.", + "bbox": [ + 212, + 145, + 782, + 220 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Agent Interaction The resulting agents' motions from the prediction process are located at $\\{p_a^{t + 1}\\}_{a\\in \\mathbf{A}}$ , comprising the motion states up to the time step of $t + 1$ . Moving the ego from $p_E^t$ to the future one-step of $p_E^{t + 1}$ , the ego vehicle should consider the agents' traffic globally and locally. From a more global perspective, the agents at a larger range provide more extensive information on traffic flow, which is essential in long-term trajectory planning. Regarding the spatial locality, the nearby agents are recognized as the key agents, which are supposed to be vitally related to the ego's driving decision.", + "bbox": [ + 212, + 222, + 784, + 342 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Therefore, we propose to conduct a hierarchical interaction with the agents through the attention mechanism to learn coarse-to-fine traffic context features for the ego. Centered at the ego's space-time position $p_E^t$ , we initially define a distance set of $\\mathbf{S}$ of $\\{+\\infty m, 15m, 7.5m\\}$ which covers the coarse-to-fine perception ranges. We formed the multi-scale agent sets by applying $\\mathcal{M}(p_E^t, p_{\\mathbf{A}}, s)$ with different ranges. Then, the ego query interacts with the agents hierarchically through multi-head cross-attention, MHCA. We take the sum of the learned hierarchical attention results as the final values:", + "bbox": [ + 212, + 343, + 784, + 462 + ], + "page_idx": 6 + }, + { + "type": "equation", + "text": "\n$$\n\\mathbf {E} ^ {k} = \\sum_ {s \\in \\mathbf {S}} \\operatorname {M H C A} (\\mathbf {E}, \\mathbf {A} ^ {k}, \\mathcal {M} \\left(p _ {\\mathbf {E}} ^ {t}, p _ {\\mathbf {A} ^ {k}} ^ {t + 1}, s\\right)), k \\in [ 1, N _ {A} ^ {\\text {m o t}} ], \\tag {1}\n$$\n", + "text_format": "latex", + "bbox": [ + 316, + 470, + 784, + 501 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "where the ego independently queries information from different modes $k \\in N_A^{mot}$ of agents, and then we stack the results output from different modes. Further, we apply the set operations to condense the features:", + "bbox": [ + 212, + 510, + 782, + 555 + ], + "page_idx": 6 + }, + { + "type": "equation", + "text": "\n$$\n\\mathbf {E} ^ {\\prime} = \\operatorname {M A X} \\left(\\left[ \\mathbf {E} ^ {1},.., \\mathbf {E} ^ {N _ {A} ^ {\\text {m o t}}} \\right]\\right) + \\operatorname {M E A N} \\left(\\left[ \\mathbf {E} ^ {1},.., \\mathbf {E} ^ {N _ {A} ^ {\\text {m o t}}} \\right]\\right), \\tag {2}\n$$\n", + "text_format": "latex", + "bbox": [ + 328, + 564, + 784, + 582 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "where the MAX and MEAN are applied to aggregate features along the agents' mode dimension and output with the updated ego query $\\mathbf{E}'$ .", + "bbox": [ + 212, + 590, + 782, + 619 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Map Interaction Existing works [19, 23] tried to summarize all the required map information for planning by simply applying the global-level interaction once. They overlook the complexity of the evolving motion dynamic and overrate that the ego can plan precisely in the longer term by a single interaction with the map information.", + "bbox": [ + 212, + 621, + 784, + 696 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "With our proposed PPAD framework, we can enrich the ego-map interaction by considering the ego's local road conditions based on its latest position. This results in better identifying the useful map information for each step of planning. The ego query interacts with the map queries in a similar practice as the interaction with the agents. The difference is that the map instances are not movable in the future time steps. The local and global map information can be abstracted into the ego query by MHCA:", + "bbox": [ + 212, + 696, + 784, + 801 + ], + "page_idx": 6 + }, + { + "type": "equation", + "text": "\n$$\n\\mathbf {E} ^ {\\prime \\prime} = \\sum_ {s \\in \\mathbf {S}} \\operatorname {M H C A} \\left(\\mathbf {E} ^ {\\prime}, \\mathbf {M}, \\mathcal {M} \\left(p _ {\\mathbf {E}} ^ {t}, p _ {m}, s\\right)\\right), \\tag {3}\n$$\n", + "text_format": "latex", + "bbox": [ + 370, + 811, + 784, + 842 + ], + "page_idx": 6 + }, + { + "type": "header", + "text": "PPAD", + "bbox": [ + 686, + 114, + 730, + 126 + ], + "page_idx": 6 + }, + { + "type": "page_number", + "text": "7", + "bbox": [ + 774, + 116, + 785, + 126 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "where $\\mathbf{E}''$ is the ego query updated with the critical map information for the next step of planning.", + "bbox": [ + 212, + 146, + 782, + 176 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "BEV Interaction BEV is the fundamental representation of the whole system, which is the abstraction of the multi-camera features. Beyond the vectorized representation, there is other non-structural environmental information, including roads and fences. UniAD [19] models these non-structural pieces of stuff with an occupancy grid map. There are several drawbacks in UniAD: 1) occupancy grids consume large memory considering the whole scene's range. 2) UniAD failed to build the explicit interactions with the grid map. Therefore, we propose the BEV interactions that dynamically query the surrounding environment for each possible future step. This query process could help agents understand and learn the effects of their actions. Specifically, after applying the interactions above, the ego vehicle understands the dynamic agents' traffic better and knows its fronting road conditions. Nevertheless, planning a more precise motion requires the ego to comprehend the local detail geometric information. Hence, the ego query further interacts with BEV features to extract low-level geometric information. Specifically, we achieve by the deformable attention [58]:", + "bbox": [ + 212, + 178, + 787, + 406 + ], + "page_idx": 7 + }, + { + "type": "equation", + "text": "\n$$\n\\mathbf {E} ^ {\\prime \\prime \\prime} = \\operatorname {D e f o r m A t t n} \\left(\\mathbf {E} ^ {\\prime \\prime}, p _ {\\mathbf {E}} ^ {t}, \\mathbf {B}\\right), \\tag {4}\n$$\n", + "text_format": "latex", + "bbox": [ + 393, + 421, + 784, + 438 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "where $p_{\\mathbf{E}}^t$ is the location of ego at time $t$ , and it serves as the reference point on the BEV features. The deformable attention DeformAttn applies sparse attention around the reference points $p_{\\mathbf{E}}^t$ and learns to pick up the low-level geometric information from the BEV for planning.", + "bbox": [ + 212, + 452, + 782, + 513 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Motion Planning PPAD follows the same practice as [18, 19, 23], which uses the information of the high-level driving commands: go straight, turn left, and turn right. The concatenated features of $\\mathbf{h}_{\\mathbf{E}} = [\\mathbf{E}', \\mathbf{E}'', \\mathbf{E}''']$ contain the information of the dynamic agent traffic, map semantics, and the precise environmental geometry. An MLP takes $\\mathbf{h}_{\\mathbf{E}}$ as input and predicts the future one-step waypoint offset $w_{\\mathbf{E}}^{t+1} = (x, y)$ . We then update the ego state by applying another MLP on $\\mathbf{h}_{\\mathbf{E}}$ for the next step of processing.", + "bbox": [ + 212, + 515, + 787, + 621 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "3.3 Hierarchical Feature Learning", + "text_level": 1, + "bbox": [ + 214, + 652, + 509, + 669 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Hierarchical structure has a better capability to capture and recognize fine-grained patterns. For the driving scenarios, the driving behavior is based on scene understanding both globally and locally. Driving tends to focus on only a few key objects, which demonstrates the spatial locality or local attention. Therefore, we design hierarchical key objects attention to exploit the coarse-to-fine scene contexts. Specifically, given a set of distance ranges, we first find the key objects (agents or map elements) within the given range. Consequently, we apply dynamic local attention, which only considers the interactions among agents or map elements in the local area. The pseudo code shown in Alg. 1 delineates the implementation of dynamic key objects attention.", + "bbox": [ + 212, + 688, + 787, + 840 + ], + "page_idx": 7 + }, + { + "type": "page_number", + "text": "8", + "bbox": [ + 217, + 114, + 228, + 126 + ], + "page_idx": 7 + }, + { + "type": "header", + "text": "Z. Chen et al.", + "bbox": [ + 271, + 114, + 364, + 126 + ], + "page_idx": 7 + }, + { + "type": "code", + "sub_type": "code", + "code_caption": [ + "Algorithm 1 Pseudo code of Key objects attention in a PyTorch-like style." + ], + "code_body": ">>> layer = nn.MultiheadAttention(embed_dim, num_head)\n>>> def forward(layer, query, key, q_pos, k_pos, max_dis):\n # q_pos: position of the query with shape [B,Lq,2]\n # k_pos: position of the key with shape [B,Lk,2]\n # layer: attention layer as initialization\n # max_dis: distance threshold\n diff = (q_pos unsqueeze(2) - k_pos unsqueeze(1))\n dist = (diff ** 2).sum(-1).sqrt()\n attn_mask = (dist > max_dis).repeat(num_head,1,1)\n return attention(query, key, attn_mask=attn_mask)", + "guess_lang": "python", + "bbox": [ + 300, + 210, + 697, + 359 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "3.4 Noisy Trajectory as Prediction", + "text_level": 1, + "bbox": [ + 215, + 402, + 514, + 417 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "PPAD interleaves the prediction and the planning processes to plan the ego and agents' trajectories step-by-step. Expert driving knowledge is then enforced into the model through imitation learning. Thanks to our multi-step framework and the inspiration from [27], we introduce the noisy trajectory as the prediction to the PPAD while training. Specifically, we perturb each step of the ground truth ego trajectory by adding noise. The ego is then trained to predict the original next step waypoint offset of the ego regardless of disturbance on its starting noisy positions. The system is learned to predict the accurate waypoint offset by interacting with vectorized instances and the environment even though it starts at an inaccurate position. This strategy brings improvement to the planning performance.", + "bbox": [ + 212, + 429, + 787, + 595 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "3.5 End-to-End Learning", + "text_level": 1, + "bbox": [ + 215, + 619, + 437, + 636 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "Scene Context Loss Similar to VAD [23], we formulate the loss for the agents' motion and map as follows:", + "bbox": [ + 212, + 646, + 784, + 676 + ], + "page_idx": 8 + }, + { + "type": "equation", + "text": "\n$$\n\\mathcal {L} _ {S} = \\lambda_ {1} \\mathcal {L} _ {\\text {a g e n t}} + \\lambda_ {2} \\mathcal {L} _ {\\text {m a p}}, \\tag {5}\n$$\n", + "text_format": "latex", + "bbox": [ + 406, + 690, + 785, + 708 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "where $\\lambda_{1}$ and $\\lambda_{2}$ are set as 1.0.", + "bbox": [ + 212, + 719, + 439, + 733 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "Constraint Loss Inspired by [23], we propose the confidence-aware collision loss $\\mathcal{L}_{CA - Col}$ , which considers the potential collision of all the agents' motion modalities instead of only computing the loss on the agents' most confident mode. We multiply the resulting collision loss from each mode with the predicted confidence score. For the trajectories having the potential to collide, it will penalize more when the predicted confidence score is higher. Combined with ego-boundary overstepping $\\mathcal{L}_{bd}$ and ego-lane directional $\\mathcal{L}_{dir}$ constraints proposed by [23], the", + "bbox": [ + 212, + 734, + 787, + 840 + ], + "page_idx": 8 + }, + { + "type": "header", + "text": "PPAD", + "bbox": [ + 686, + 114, + 730, + 126 + ], + "page_idx": 8 + }, + { + "type": "page_number", + "text": "9", + "bbox": [ + 774, + 116, + 785, + 126 + ], + "page_idx": 8 + }, + { + "type": "table", + "img_path": "images/1005f0d49a5dfc56a9d22c59eb8284f5632b74f29a5bc5b84a4b561d1d99705e.jpg", + "table_caption": [], + "table_footnote": [], + "table_body": "
MethodL2 (m) ↓Collision (%) ↓Latency (ms)FPS
1s2s3sAvg.1s2s3sAvg.
ST-P3 MetricsST-P3 [18]1.332.112.902.110.230.621.270.71628.31.6
VAD-Tiny [23]0.460.761.120.780.210.350.580.3859.516.8
VAD-Base [23]0.410.701.050.720.070.170.410.22224.34.5
OccNet [41]1.292.132.992.130.210.591.370.72--
FusionAD [51]---1.030.250.130.250.21--
Ours (Progress.)0.310.560.870.580.080.120.380.193852.6
Ours0.250.450.730.480.070.150.360.193852.6
UniAD MetricsNMP† [53]--2.31---1.92---
SA-NMP† [53]--2.05---1.59---
FF† [17]0.551.202.541.430.060.171.070.43--
EO† [24]0.671.362.781.600.040.090.880.33--
UniAD [19]0.480.961.651.030.050.170.710.31555.61.8
VAD-Base [23]0.501.021.691.070.000.300.950.42224.34.5
Ours (Progress.)0.380.831.450.890.020.200.930.383852.6
Ours0.300.691.260.750.030.220.730.333852.6
", + "bbox": [ + 228, + 143, + 782, + 340 + ], + "page_idx": 9 + }, + { + "type": "text", + "text": "Table 1: Open-loop planning results on the nuScenes dataset [3]. The results of other methods are obtained from the original paper. We faithfully re-evaluate VAD [23] based on the UniAD [19] metrics. As for our PPAD, we provide two versions of results that utilize different training strategies. Progress means that we follow the progressive training pipeline as proposed in VAD [23], which trained all tasks except the planning task in the first 48 epochs and then finetuned with another 12 epochs for the planning task. The second row for our method trains the whole network for 60 epochs and then finetuned for another 12 epochs incorporating noisy trajectories. The latency of STP3 [18], VAD [23], and ours are measured on one NVIDIA Geforce RTX 3090 GPU, while UniAD is measured on one NVIDIA Tesla A100 GPU.", + "bbox": [ + 212, + 351, + 787, + 491 + ], + "page_idx": 9 + }, + { + "type": "text", + "text": "overall constraint loss is", + "bbox": [ + 215, + 537, + 392, + 550 + ], + "page_idx": 9 + }, + { + "type": "equation", + "text": "\n$$\n\\mathcal {L} _ {C} = \\lambda_ {3} \\mathcal {L} _ {C A - C o l} + \\lambda_ {4} \\mathcal {L} _ {b d} + \\lambda_ {5} \\mathcal {L} _ {d i r}, \\tag {6}\n$$\n", + "text_format": "latex", + "bbox": [ + 369, + 566, + 785, + 583 + ], + "page_idx": 9 + }, + { + "type": "text", + "text": "where $\\lambda_3$ and $\\lambda_4$ are set as 1.0, and $\\lambda_5$ is set as 0.5.", + "bbox": [ + 212, + 598, + 588, + 613 + ], + "page_idx": 9 + }, + { + "type": "text", + "text": "Planning Loss We conduct $L_{1}$ loss between each step of the ego's prediction $w_{\\mathbf{E}}^{t}$ and the ground truth's waypoint offset $\\tilde{w}_{\\mathbf{E}}^{t}$ along the future time horizon:", + "bbox": [ + 212, + 614, + 785, + 646 + ], + "page_idx": 9 + }, + { + "type": "equation", + "text": "\n$$\n\\mathcal {L} _ {\\text {P l a n}} = \\frac {1}{T _ {f u t}} \\sum_ {t = 1} ^ {T _ {f u t}} \\| w _ {\\mathbf {E}} ^ {t} - \\tilde {w} _ {\\mathbf {E}} ^ {t} \\| _ {1}. \\tag {7}\n$$\n", + "text_format": "latex", + "bbox": [ + 385, + 660, + 785, + 702 + ], + "page_idx": 9 + }, + { + "type": "text", + "text": "The overall end-to-end trainable loss function is formed by the sum of the perception loss, constraint loss, and planning loss. The same constraint losses $\\mathcal{L}_C^{noisy}$ and planning losses $\\mathcal{L}_{plan}^{noisy}$ will be applied to the predictions taking the noisy trajectories as input:", + "bbox": [ + 212, + 717, + 787, + 779 + ], + "page_idx": 9 + }, + { + "type": "equation", + "text": "\n$$\n\\mathcal {L} = \\mathcal {L} _ {S} + \\zeta_ {1} \\left(\\mathcal {L} _ {C} + \\mathcal {L} _ {\\text {P l a n}}\\right) + \\zeta_ {2} \\left(\\mathcal {L} _ {C} ^ {\\text {n o i s y}} + \\mathcal {L} _ {\\text {P l a n}} ^ {\\text {n o i s y}}\\right), \\tag {8}\n$$\n", + "text_format": "latex", + "bbox": [ + 325, + 792, + 785, + 811 + ], + "page_idx": 9 + }, + { + "type": "text", + "text": "where $\\zeta_{1}$ is set as 0.6 and $\\zeta_{2}$ is set as 0.4.", + "bbox": [ + 212, + 824, + 516, + 840 + ], + "page_idx": 9 + }, + { + "type": "page_number", + "text": "10", + "bbox": [ + 217, + 114, + 235, + 126 + ], + "page_idx": 9 + }, + { + "type": "header", + "text": "Z. Chen et al.", + "bbox": [ + 271, + 114, + 364, + 126 + ], + "page_idx": 9 + }, + { + "type": "table", + "img_path": "images/33f2e788d68a74f385b10807f27f3262b7b1bf13e9e71d5a0ff40765f4be463c.jpg", + "table_caption": [], + "table_footnote": [], + "table_body": "
MethodL2 (m) ↓Collision (%) ↓
1s2s3sAvg.1s2s3sAvg.
VAD [23]1.041.792.601.810.611.262.251.37
Ours0.731.301.981.340.510.851.310.89
", + "bbox": [ + 328, + 143, + 683, + 200 + ], + "page_idx": 10 + }, + { + "type": "table", + "img_path": "images/2f112880a869062099ecff6ebd1ac88880980c68ae585008f0c2750200a77cc8.jpg", + "table_caption": [ + "Table 2: Open-loop planning results on the Argoverse dataset [3]. The results are evaluated under the ST-P3 [18] and VAD [23] metrics." + ], + "table_footnote": [], + "table_body": "
MethodDetectionMapMotion Forecasting
NDS ↑mAP↑mAP↑minADE (m)↓minFDE (m) ↓MR↓
VAD [23]0.4590.3290.4760.6780.8820.08
UniAD [19]0.4990.382-0.7081.020.13
Ours0.4650.3320.5190.6760.8890.07
", + "bbox": [ + 287, + 266, + 725, + 337 + ], + "page_idx": 10 + }, + { + "type": "text", + "text": "Table 3: Results comparison on the tasks beyond the planning task.", + "bbox": [ + 269, + 347, + 730, + 362 + ], + "page_idx": 10 + }, + { + "type": "text", + "text": "4 Experiments", + "text_level": 1, + "bbox": [ + 215, + 402, + 375, + 420 + ], + "page_idx": 10 + }, + { + "type": "text", + "text": "4.1 Experimental Setup", + "text_level": 1, + "bbox": [ + 215, + 436, + 428, + 452 + ], + "page_idx": 10 + }, + { + "type": "text", + "text": "Dataset We evaluate our method on two challenging large-scale real-world datasets, nuScenes [3] and Argoverse2 [44]. We conduct ablation studies to evaluate the effectiveness of our proposed components on the nuScenes [3] dataset. The nuScenes dataset [3] provides about 1K 20-second diverse driving scenes collected in Boston, Pittsburgh, Las Vegas, and Singapore for open-loop settings. Key samples which contain 6 camera images are annotated at $2\\mathrm{Hz}$ . The Argoverse2 dataset comprises 1K 15-second scenes. It captures 7 RGB images on each frame. We align the data sampling frequency to nuScenes [3] by downsampling the annotated frames by an interval of 5 frames.", + "bbox": [ + 212, + 460, + 787, + 597 + ], + "page_idx": 10 + }, + { + "type": "text", + "text": "Metrics We adopt the metrics of L2 Displacement Error (L2) and Collision Rate (CR) for evaluation [18]. L2 is measured between the prediction and ground-truth trajectories over the timesteps 1-s, 2-s, and 3-s, evaluating the trajectory quality. Collision Rate (CR) measures how often the collision occurs between the ego vehicle and the other agents along the planning horizon, reflecting the trajectory safety. We notice that the UniAD [19] adheres to different calculations from ST-P3 [18] and VAD [23]: the former [19] reports the evaluations at each second. In contrast, the latters [18, 23] reports the results of the cumulative average by each second. We faithfully make comparisons of our method to others with these two kinds of computations.", + "bbox": [ + 212, + 598, + 787, + 748 + ], + "page_idx": 10 + }, + { + "type": "text", + "text": "Implementation Details We strictly follow the standard settings as proposed by UniAD [19] and VAD [23], which did not use the information of the historical ego trajectory. Same as VAD [23], the perception range is $60m \\times 30m$ . PPAD also recurrently encodes 2-s historical information into BEV and predicts the 3-s trajectory in the future. It also conducts the tasks of motion prediction and map construction. Our PPAD trains with a batch size of 1 using the Adam [25]", + "bbox": [ + 212, + 750, + 787, + 840 + ], + "page_idx": 10 + }, + { + "type": "header", + "text": "PPAD", + "bbox": [ + 686, + 114, + 730, + 126 + ], + "page_idx": 10 + }, + { + "type": "page_number", + "text": "11", + "bbox": [ + 767, + 114, + 784, + 126 + ], + "page_idx": 10 + }, + { + "type": "text", + "text": "optimizer set with an initial learning rate of 2e-4. It takes us about six days to conduct the end-to-end training on eight NVIDIA A30 24GB GPUs.", + "bbox": [ + 212, + 146, + 782, + 176 + ], + "page_idx": 11 + }, + { + "type": "text", + "text": "4.2 Main Results", + "text_level": 1, + "bbox": [ + 215, + 196, + 372, + 210 + ], + "page_idx": 11 + }, + { + "type": "text", + "text": "Planning Results As shown in Tab. 1, our PPAD outperforms the current state-of-the-art performance by a large margin. Especially for the L2 distance metrics, there are about $20\\%$ of consistent improvements can be observed along the temporal horizon. Thanks to the iterative interaction of prediction and planning, PPAD can help avoid collisions, leading to better results on collision rate compared to the one-shot representative VAD [23]. At the same time, we maintain a competitive efficiency compared to UniAD [19].", + "bbox": [ + 212, + 220, + 784, + 325 + ], + "page_idx": 11 + }, + { + "type": "text", + "text": "We further make a fair comparison between our method and the baseline, VAD [23], on the other dataset of Argoverse 2 [6,44]. In Tab. 2, our method can consistently outperform the baseline with a clear margin in both L2 distance and collision rate metrics.", + "bbox": [ + 212, + 327, + 784, + 385 + ], + "page_idx": 11 + }, + { + "type": "text", + "text": "Subtasks Results To demonstrate the overall performance of our PPAD, we also provide the evaluation results besides planning metrics on the traditional perception and motion forecasting task in Tab. 3. Our PPAD also achieves promising performance in upstreaming perception and prediction tasks, which demonstrates that the whole system is jointly optimized.", + "bbox": [ + 212, + 387, + 784, + 462 + ], + "page_idx": 11 + }, + { + "type": "text", + "text": "Qualitative Results We provide qualitative results shown in Fig. 3. PPAD can perceive the scene precisely and predict with reasonable and diverse motions for the surrounding agents. It also plans a smooth and accurate trajectory for the ego vehicle.", + "bbox": [ + 212, + 463, + 784, + 522 + ], + "page_idx": 11 + }, + { + "type": "text", + "text": "4.3 Ablation Study", + "text_level": 1, + "bbox": [ + 215, + 544, + 387, + 559 + ], + "page_idx": 11 + }, + { + "type": "text", + "text": "The following experiments adhere to the progressive training pipeline as proposed in VAD [23].", + "bbox": [ + 212, + 568, + 784, + 597 + ], + "page_idx": 11 + }, + { + "type": "text", + "text": "Effectiveness of Designs We provide ablation studies to verify the effectiveness of our proposed components. As shown in Tab. 4, the proposed PPAD framework (row 2) brings a remarkable improvement compared with one-shoot methods [23] (row 1). The multi-step interactions help the ego agent better understand the intention and potential effects brought by its actions along the temporal horizon, leading to an over $10\\%$ L2 distance error reduction. We can observe a further improvement regarding the L2 distance with our proposed key objects attention (row 3). The slight degradation in collision rate might be due to the key objects attention being conducted on the ego with each mode of the agents, and the diverse modes of the agents mislead the behavior of the ego vehicle. From rows 5-7, the confidence-aware collision loss and noisy trajectory as the prediction can circumvent this phenomenon and further escalate the capability in planning accuracy and avoiding a collision. When we keep all of the remaining components while not applying key objects attention (row 4), we observed the degradations from row 7 in L2 $(0.58\\mathrm{m}$ vs. $0.59\\mathrm{m})$ and collision rate $(0.19\\%)$ vs. $0.21\\%$ , proving the effectiveness of the key objects attention.", + "bbox": [ + 212, + 598, + 785, + 839 + ], + "page_idx": 11 + }, + { + "type": "page_number", + "text": "12", + "bbox": [ + 217, + 114, + 235, + 126 + ], + "page_idx": 11 + }, + { + "type": "header", + "text": "Z. Chen et al.", + "bbox": [ + 271, + 114, + 364, + 126 + ], + "page_idx": 11 + }, + { + "type": "image", + "img_path": "images/6f6db780942a50b689bf835fc86986a19030e231705f53fa316ba8778258c77c.jpg", + "image_caption": [ + "Fig. 3: Qualitative results of PPAD. The green box in the figure demonstrates the ego agent, while the red ones are agents." + ], + "image_footnote": [], + "bbox": [ + 245, + 143, + 754, + 464 + ], + "page_idx": 12 + }, + { + "type": "table", + "img_path": "images/aad3653c19b9272d049bc9de00f321b578aabcb517e1926106ad4025c490f3e9.jpg", + "table_caption": [], + "table_footnote": [], + "table_body": "
PPADKey Objects Attn.\\( {\\mathcal{L}}_{CA - {Col}} \\)Noisy Traj.L2 (m) \\( \\downarrow \\)Collision (%) \\( \\downarrow \\)Latency (ms)
1s2s3sAvg.1s2s3sAvg.
1----0.410.701.050.720.070.170.410.22224
2✓---0.350.590.880.600.110.180.410.23318
3✓✓--0.320.550.830.570.190.280.580.35385
4✓-✓✓0.330.570.880.590.100.140.380.21318
5✓✓✓-0.350.590.890.610.080.140.320.18385
6✓✓-✓0.340.590.900.610.100.140.340.19385
7✓✓✓✓0.310.560.870.580.080.120.380.19385
", + "bbox": [ + 227, + 521, + 784, + 621 + ], + "page_idx": 12 + }, + { + "type": "text", + "text": "Table 4: Component study for PPAD. Models follow the progressive training pipeline. PPAD means the auto-regressive framework with the designed stepwise interactions. Key Objects Attention represents hierarchical feature learning for the key objects. $\\mathcal{L}_{CA - Col}$ represents the loss design for the confidence-aware collision loss. Noisy Traj. means that we incorporate noisy trajectories while training.", + "bbox": [ + 212, + 631, + 787, + 702 + ], + "page_idx": 12 + }, + { + "type": "text", + "text": "Effectiveness of Interactions Our PPAD framework enables richer interactions among scene contexts, introducing local and global understandings of the world to the model. We further conduct ablation studies to demonstrate the performance gains brought by the interactions. We conduct the ablation study on interactions under the setting (Tab. 4 in row 5) without using the noisy trajectory as the prediction for better comparison. As illustrated in Tab. 5, we can achieve", + "bbox": [ + 212, + 750, + 787, + 840 + ], + "page_idx": 12 + }, + { + "type": "header", + "text": "PPAD", + "bbox": [ + 686, + 114, + 730, + 126 + ], + "page_idx": 12 + }, + { + "type": "page_number", + "text": "13", + "bbox": [ + 767, + 114, + 785, + 126 + ], + "page_idx": 12 + }, + { + "type": "table", + "img_path": "images/d9bee3110fbb9459bde671ef2c1377b8d566a26f0ec76d58a6ed327de343badc.jpg", + "table_caption": [], + "table_footnote": [], + "table_body": "
EAMapBEVL2 (m) ↓Collision (%) ↓
1s2s3sAvg.1s2s3sAvg.
✓--0.350.580.890.610.230.240.490.32
✓✓-0.380.640.950.660.170.180.480.28
✓✓✓0.350.590.890.610.080.140.320.18
", + "bbox": [ + 338, + 143, + 669, + 209 + ], + "page_idx": 13 + }, + { + "type": "table", + "img_path": "images/1a70c634fc3e97aa242a3877cd6bcc5b0527fe7179c0cc187acfb47a4bbb3ef3.jpg", + "table_caption": [ + "Table 5: Interaction study for PPAD. EA, Map, BEV mean the interactions of the ego with agents, the ego and agents with the map, the ego and agents with the BEV, respectively." + ], + "table_footnote": [], + "table_body": "
Interaction IterationsL2 (m) ↓Collision (%) ↓Latency (ms)
1s2s3sAvg.1s2s3sAvg.
2(Every 1.5 sec)0.360.630.960.650.140.200.430.25306
3(Every 1.0 sec)0.370.630.950.650.100.150.390.21326
6(Every 0.5 sec)0.310.560.870.580.080.120.380.19385
", + "bbox": [ + 230, + 286, + 781, + 363 + ], + "page_idx": 13 + }, + { + "type": "text", + "text": "Table 6: The ablation study of conducting different interaction iterations in the future time horizons.", + "bbox": [ + 215, + 373, + 785, + 402 + ], + "page_idx": 13 + }, + { + "type": "text", + "text": "the best performance in terms of L2 distance and collision by incorporating all of the interactions.", + "bbox": [ + 212, + 441, + 784, + 470 + ], + "page_idx": 13 + }, + { + "type": "text", + "text": "Effect of Different Iterations on Prediction and Planning Interaction Our innovative PPAD interaction mechanism not only better models motion planning as gaming among the ego and the agents but also enriches the ego's / agents' interactions with their local environments. We further conducted the ablation study on the different iterations in applying the interactions of prediction and planning, as shown in Tab. 6. Specifically, PPAD plans the trajectories with 3, 2, and 1 steps of waypoints after each of the prediction-planning processes for the interaction iterations of 2, 3, and 6 in Tab. 6. It is demonstrated that the performance reaches the best as we conduct the interactions of prediction and planning processes at every future step.", + "bbox": [ + 212, + 472, + 785, + 625 + ], + "page_idx": 13 + }, + { + "type": "text", + "text": "5 Conclusion", + "text_level": 1, + "bbox": [ + 215, + 643, + 359, + 660 + ], + "page_idx": 13 + }, + { + "type": "text", + "text": "In this paper, we have presented a novel autonomous driving framework, PPAD. Different from the previous methods that lack in-depth modeling of interactions, we pose the planning problem as a multi-step Prediction and Planning gaming process among the ego vehicle and agents. With PPAD architecture, our proposed hierarchical dynamic key objects attention is incorporated to learn local and global scene contexts at each step and eventually plan with a more precise trajectory. The confidence-aware collision constraint and noisy trajectories are utilized while training to improve driving safety further. In general, our proposed novel PPAD achieves compelling performance upon the existing state-of-the-art methods, and we hope the PPAD framework can inspire the community to further exploration.", + "bbox": [ + 212, + 672, + 785, + 840 + ], + "page_idx": 13 + }, + { + "type": "page_number", + "text": "14", + "bbox": [ + 217, + 114, + 235, + 126 + ], + "page_idx": 13 + }, + { + "type": "header", + "text": "Z. Chen et al.", + "bbox": [ + 271, + 114, + 364, + 126 + ], + "page_idx": 13 + }, + { + "type": "text", + "text": "Acknowledgements", + "text_level": 1, + "bbox": [ + 217, + 143, + 401, + 162 + ], + "page_idx": 14 + }, + { + "type": "text", + "text": "The authors are thankful for the financial support from the Hetao Shenzhen-HongKong Science and Technology Innovation Cooperation Zone (HZQB-KCZYZ-2021055), this work was also supported by Shenzhen Deeproute.ai Co., Ltd (HZQB-KCZYZ-2021055).", + "bbox": [ + 212, + 178, + 795, + 239 + ], + "page_idx": 14 + }, + { + "type": "text", + "text": "References", + "text_level": 1, + "bbox": [ + 215, + 263, + 321, + 277 + ], + "page_idx": 14 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "1. Bansal, M., Krizhevsky, A., Ogale, A.: Chauffeurnet. In: Robotics: Science and Systems XV (2019)", + "2. Bojarski, M., Del Testa, D., Dworakowski, D., Firner, B., Flepp, B., Goyal, P., Jackel, L.D., Monfort, M., Muller, U., Zhang, J., et al.: End to end learning for self-driving cars. arXiv preprint arXiv:1604.07316 (2016)", + "3. Caesar, H., Bankiti, V., Lang, A.H., Vora, S., Liong, V.E., Xu, Q., Krishnan, A., Pan, Y., Baldan, G., Beijbom, O.: nuscenes: A multimodal dataset for autonomous driving. In: Proceedings of the IEEE/CVF conference on computer vision and pattern recognition. pp. 11621-11631 (2020)", + "4. Casas, S., Luo, W., Urtasun, R.: Intentnet: Learning to predict intention from raw sensor data. In: Conference on Robot Learning. pp. 947-956 (2018)", + "5. Chai, Y., Sapp, B., Bansal, M., Anguelov, D.: Multipath: Multiple probabilistic anchor trajectory hypotheses for behavior prediction. arXiv preprint arXiv:1910.05449 (2019)", + "6. Chang, M.F., Lambert, J., Sangkloy, P., Singh, J., Bak, S., Hartnett, A., Wang, D., Carr, P., Lucey, S., Ramanan, D., et al.: Argoverse: 3d tracking and forecasting with rich maps. In: Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition. pp. 8748-8757 (2019)", + "7. Chekroun, R., Toromanoff, M., Hornauer, S., Moutarde, F.: Gri: General reinforced imitation and its application to vision-based autonomous driving. arXiv preprint arXiv:2111.08575 (2021)", + "8. Chen, D., Zhou, B., Koltun, V., Krahenbuhl, P.: Learning by cheating. In: Conference on Robot Learning. pp. 66-75. PMLR (2020)", + "9. Chitta, K., Prakash, A., Jaeger, B., Yu, Z., Renz, K., Geiger, A.: Transfuser: Imitation with transformer-based sensor fusion for autonomous driving. IEEE Transactions on Pattern Analysis and Machine Intelligence (2022)", + "10. Codevilla, F., Santana, E., López, A.M., Gaidon, A.: Exploring the limitations of behavior cloning for autonomous driving. In: Proceedings of the IEEE/CVF International Conference on Computer Vision. pp. 9329-9338 (2019)", + "11. Duvenaud, D.K., Maclaurin, D., Iparraguirre, J., Bombarell, R., Hirzel, T., Aspuru-Guzik, A., Adams, R.P.: Convolutional networks on graphs for learning molecular fingerprints. In: Advances in neural information processing systems. pp. 2224-2232 (2015)", + "12. Fan, H., Zhu, F., Liu, C., Zhang, L., Zhuang, L., Li, D., Zhu, W., Hu, J., Li, H., Kong, Q.: BaiduApollo em motion planner. arXiv preprint arXiv:1807.08048 (2018)", + "13. Gao, J., Sun, C., Zhao, H., Shen, Y., Anguelov, D., Li, C., Schmid, C.: Vectornet: Encoding hd maps and agent dynamics from vectorized representation. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 11525-11533 (2020)" + ], + "bbox": [ + 218, + 295, + 784, + 839 + ], + "page_idx": 14 + }, + { + "type": "header", + "text": "PPAD", + "bbox": [ + 686, + 114, + 730, + 126 + ], + "page_idx": 14 + }, + { + "type": "page_number", + "text": "15", + "bbox": [ + 767, + 116, + 785, + 126 + ], + "page_idx": 14 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "14. Gu, J., Hu, C., Zhang, T., Chen, X., Wang, Y., Wang, Y., Zhao, H.: Vip3d: End-to-end visual trajectory prediction via 3d agent queries. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 5496-5506 (2023)", + "15. He, K., Zhang, X., Ren, S., Sun, J.: Deep residual learning for image recognition. In: Proceedings of the IEEE conference on computer vision and pattern recognition. pp. 770-778 (2016)", + "16. Henaff, M., Bruna, J., LeCun, Y.: Deep convolutional networks on graph-structured data. arXiv preprint arXiv:1506.05163 (2015)", + "17. Hu, P., Huang, A., Dolan, J., Held, D., Ramanan, D.: Safe local motion planning with self-supervised freespace forecasting. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 12732-12741 (2021)", + "18. Hu, S., Chen, L., Wu, P., Li, H., Yan, J., Tao, D.: St-p3: End-to-end vision-based autonomous driving via spatial-temporal feature learning. In: European Conference on Computer Vision. pp. 533-549. Springer (2022)", + "19. Hu, Y., Yang, J., Chen, L., Li, K., Sima, C., Zhu, X., Chai, S., Du, S., Lin, T., Wang, W., et al.: Planning-oriented autonomous driving. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 17853-17862 (2023)", + "20. Huang, J., Huang, G., Zhu, Z., Ye, Y., Du, D.: Bevdet: High-performance multicamera 3d object detection in bird-eye-view. arXiv preprint arXiv:2112.11790 (2021)", + "21. Huang, Z., Liu, H., Lv, C.: Gameformer: Game-theoretic modeling and learning of transformer-based interactive prediction and planning for autonomous driving. arXiv preprint arXiv:2303.05760 (2023)", + "22. Jia, X., Wu, P., Chen, L., Xie, J., He, C., Yan, J., Li, H.: Think twice before driving: Towards scalable decoders for end-to-end autonomous driving. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 21983-21994 (2023)", + "23. Jiang, B., Chen, S., Xu, Q., Liao, B., Chen, J., Zhou, H., Zhang, Q., Liu, W., Huang, C., Wang, X.: Vad: Vectorized scene representation for efficient autonomous driving. arXiv preprint arXiv:2303.12077 (2023)", + "24. Khurana, T., Hu, P., Dave, A., Ziglar, J., Held, D., Ramanan, D.: Differentiable raycasting for self-supervised occupancy forecasting. In: European Conference on Computer Vision. pp. 353-369. Springer (2022)", + "25. Kingma, D.P., Ba, J.: Adam: A method for stochastic optimization. arXiv preprint arXiv:1412.6980 (2014)", + "26. Lang, A.H., Vora, S., Caesar, H., Zhou, L., Yang, J., Beijbom, O.: PointPillars: Fast Encoders for Object Detection From Point Clouds. In: Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition. pp. 12697-12705 (2019)", + "27. Li, F., Zhang, H., Liu, S., Guo, J., Ni, L.M., Zhang, L.: Dn-detr: Accelerate detr training by introducing query denoising. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 13619-13627 (2022)", + "28. Li, Y., Ge, Z., Yu, G., Yang, J., Wang, Z., Shi, Y., Sun, J., Li, Z.: Bevdepth: Acquisition of reliable depth for multi-view 3d object detection. In: Proceedings of the AAAI Conference on Artificial Intelligence. pp. 1477-1485 (2023)", + "29. Li, Z., Wang, W., Li, H., Xie, E., Sima, C., Lu, T., Qiao, Y., Dai, J.: Bevformer: Learning bird's-eye-view representation from multi-camera images via spatiotemporal transformers. In: European conference on computer vision. pp. 1-18. Springer (2022)" + ], + "bbox": [ + 215, + 146, + 785, + 839 + ], + "page_idx": 15 + }, + { + "type": "page_number", + "text": "16", + "bbox": [ + 217, + 114, + 235, + 126 + ], + "page_idx": 15 + }, + { + "type": "header", + "text": "Z. Chen et al.", + "bbox": [ + 271, + 114, + 364, + 126 + ], + "page_idx": 15 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "30. Liang, M., Yang, B., Hu, R., Chen, Y., Liao, R., Feng, S., Urtasun, R.: Learning lane graph representations for motion forecasting. In: Proceedings of the European Conference on Computer Vision (ECCV). pp. 541-556 (2020)", + "31. Liao, B., Chen, S., Wang, X., Cheng, T., Zhang, Q., Liu, W., Huang, C.: Maptr: Structured modeling and learning for online vectorized hd map construction. arXiv preprint arXiv:2208.14437 (2022)", + "32. Liu, Y., Zhang, J., Fang, L., Jiang, Q., Zhou, B.: Multimodal motion prediction with stacked transformers. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 7577-7586 (2021)", + "33. Liu, Y., Wang, T., Zhang, X., Sun, J.: Petr: Position embedding transformation for multi-view 3d object detection. In: European Conference on Computer Vision. pp. 531-548. Springer (2022)", + "34. Liu, Y., Yan, J., Jia, F., Li, S., Gao, A., Wang, T., Zhang, X., Sun, J.: Petrv2: A unified framework for 3d perception from multi-camera images. arXiv preprint arXiv:2206.01256 (2022)", + "35. Liu, Z., Chen, S., Guo, X., Wang, X., Cheng, T., Zhu, H., Zhang, Q., Liu, W., Zhang, Y.: Vision-based uneven bev representation learning with polar rasterization and surface estimation. In: Conference on Robot Learning. pp. 437-446. PMLR (2023)", + "36. Liu, Z., Tang, H., Amini, A., Yang, X., Mao, H., Rus, D.L., Han, S.: Bevfusion: Multi-task multi-sensor fusion with unified bird's-eye view representation. In: 2023 IEEE International Conference on Robotics and Automation (ICRA). pp. 2774-2781. IEEE (2023)", + "37. Luo, W., Yang, B., Urtasun, R.: Fast and furious: Real time end-to-end 3d detection, tracking and motion forecasting with a single convolutional net. In: Proceedings of the IEEE conference on Computer Vision and Pattern Recognition. pp. 3569-3577 (2018)", + "38. Philion, J., Fidler, S.: Lift, splat, shoot: Encoding images from arbitrary camera rigs by implicitly unprojecting to 3d. In: Computer Vision-ECCV 2020: 16th European Conference, Glasgow, UK, August 23-28, 2020, Proceedings, Part XIV 16. pp. 194-210. Springer (2020)", + "39. Renz, K., Chitta, K., Mercea, O.B., Koepke, A., Akata, Z., Geiger, A.: Plant: Explainable planning transformers via object-level representations. arXiv preprint arXiv:2210.14222 (2022)", + "40. Shuman, D.I., Narang, S.K., Frossard, P., Ortega, A., Vandergheynst, P.: The emerging field of signal processing on graphs: Extending high-dimensional data analysis to networks and other irregular domains. IEEE signal processing magazine 30(3), 83-98 (2013)", + "41. Sima, C., Tong, W., Wang, T., Chen, L., Wu, S., Deng, H., Gu, Y., Lu, L., Luo, P., Lin, D., et al.: Scene as occupancy. arXiv e-prints pp. arXiv-2306 (2023)", + "42. Song, H., Luan, D., Ding, W., Wang, M.Y., Chen, Q.: Learning to predict vehicle trajectories with model-based planning. In: Conference on Robot Learning. pp. 1035-1045. PMLR (2021)", + "43. Wang, S., Liu, Y., Wang, T., Li, Y., Zhang, X.: Exploring object-centric temporal modeling for efficient multi-view 3d object detection. arXiv preprint arXiv:2303.11926 (2023)", + "44. Wilson, B., Qi, W., Agarwal, T., Lambert, J., Singh, J., Khandelwal, S., Pan, B., Kumar, R., Hartnett, A., Pontes, J.K., et al.: Argoverse 2: Next generation datasets for self-driving perception and forecasting. arXiv preprint arXiv:2301.00493 (2023)" + ], + "bbox": [ + 212, + 146, + 787, + 840 + ], + "page_idx": 16 + }, + { + "type": "header", + "text": "PPAD", + "bbox": [ + 686, + 114, + 730, + 126 + ], + "page_idx": 16 + }, + { + "type": "page_number", + "text": "17", + "bbox": [ + 767, + 114, + 785, + 126 + ], + "page_idx": 16 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "45. Wu, P., Jia, X., Chen, L., Yan, J., Li, H., Qiao, Y.: Trajectory-guided control prediction for end-to-end autonomous driving: A simple yet strong baseline. Advances in Neural Information Processing Systems 35, 6119-6132 (2022)", + "46. Xie, Y., Xu, C., Rakotosaona, M.J., Rim, P., Tombari, F., Keutzer, K., Tomizuka, M., Zhan, W.: Sparsefusion: Fusing multi-modal sparse representations for multisensor 3d object detection. arXiv preprint arXiv:2304.14340 (2023)", + "47. Yang, C., Chen, Y., Tian, H., Tao, C., Zhu, X., Zhang, Z., Huang, G., Li, H., Qiao, Y., Lu, L., et al.: Bevformer v2: Adapting modern image backbones to bird's-eye-view recognition via perspective supervision. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 17830-17839 (2023)", + "48. Ye, M., Cao, T., Chen, Q.: Tpcn: Temporal point cloud networks for motion forecasting. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 11318-11327 (2021)", + "49. Ye, M., Xu, S., Cao, T.: Hvnet: Hybrid voxel network for lidar based 3d object detection. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 1631-1640 (2020)", + "50. Ye, M., Xu, S., Cao, T., Chen, Q.: Drinet: A dual-representation iterative learning network for point cloud segmentation. In: Proceedings of the IEEE/CVF international conference on computer vision. pp. 7447-7456 (2021)", + "51. Ye, T., Jing, W., Hu, C., Huang, S., Gao, L., Li, F., Wang, J., Guo, K., Xiao, W., Mao, W., et al.: Fusionad: Multi-modality fusion for prediction and planning tasks of autonomous driving. arXiv preprint arXiv:2308.01006 (2023)", + "52. Zeng, W., Liang, M., Liao, R., Urtasun, R.: Lanercnn: Distributed representations for graph-centric motion forecasting. arXiv preprint arXiv:2101.06653 (2021)", + "53. Zeng, W., Luo, W., Suo, S., Sadat, A., Yang, B., Casas, S., Urtasun, R.: End-to-end interpretable neural motion planner. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 8660-8669 (2019)", + "54. Zhang, Z., Liniger, A., Dai, D., Yu, F., Van Gool, L.: End-to-end urban driving by imitating a reinforcement learning coach. In: Proceedings of the IEEE/CVF international conference on computer vision. pp. 15222-15232 (2021)", + "55. Zhou, Y., Tuzel, O.: VoxelNet: End-to-End Learning for Point Cloud Based 3D Object Detection. In: Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition. pp. 4490-4499 (2018)", + "56. Zhou, Z., Wang, J., Li, Y.H., Huang, Y.K.: Query-centric trajectory prediction. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 17863-17873 (2023)", + "57. Zhou, Z., Ye, L., Wang, J., Wu, K., Lu, K.: Hivt: Hierarchical vector transformer for multi-agent motion prediction. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 8823-8833 (2022)", + "58. Zhu, X., Su, W., Lu, L., Li, B., Wang, X., Dai, J.: Deformable detr: Deformable transformers for end-to-end object detection. arXiv preprint arXiv:2010.04159 (2020)" + ], + "bbox": [ + 215, + 146, + 784, + 728 + ], + "page_idx": 17 + }, + { + "type": "page_number", + "text": "18", + "bbox": [ + 217, + 114, + 235, + 126 + ], + "page_idx": 17 + }, + { + "type": "header", + "text": "Z. Chen et al.", + "bbox": [ + 271, + 114, + 364, + 126 + ], + "page_idx": 17 + } +] \ No newline at end of file diff --git a/2024/PPAD_ Iterative Interactions of Prediction and Planning for End-to-end Autonomous Driving/2548ffc2-3e35-4690-aa4a-52b1b19d7a46_model.json b/2024/PPAD_ Iterative Interactions of Prediction and Planning for End-to-end Autonomous Driving/2548ffc2-3e35-4690-aa4a-52b1b19d7a46_model.json new file mode 100644 index 0000000000000000000000000000000000000000..c526cfc11f18e6f199f4344d7ce5d8886889f3f5 --- /dev/null +++ b/2024/PPAD_ Iterative Interactions of Prediction and Planning for End-to-end Autonomous Driving/2548ffc2-3e35-4690-aa4a-52b1b19d7a46_model.json @@ -0,0 +1,2392 @@ +[ + [ + { + "type": "title", + "bbox": [ + 0.235, + 0.141, + 0.773, + 0.187 + ], + "angle": 0, + "content": "PPAD: Iterative Interactions of Prediction and Planning for End-to-end Autonomous Driving" + }, + { + "type": "text", + "bbox": [ + 0.322, + 0.212, + 0.681, + 0.245 + ], + "angle": 0, + "content": "Zhili Chen\\(^{1\\dagger}\\), Maosheng Ye\\(^{1}\\), Shuangjie Xu\\(^{1}\\), Tongyi Cao\\(^{2}\\), and Qifeng Chen\\(^{1\\boxtimes}\\)" + }, + { + "type": "text", + "bbox": [ + 0.324, + 0.255, + 0.681, + 0.298 + ], + "angle": 0, + "content": "\\(^{1}\\)HKUST \\(^{2}\\)DeepRoute.AI {zchenei, myeag, shuangjie.xu}@connect.ust.hk, tongyicao@deeproute.ai, cqf@cse.ust.hk" + }, + { + "type": "text", + "bbox": [ + 0.261, + 0.338, + 0.744, + 0.562 + ], + "angle": 0, + "content": "Abstract. We present a new interaction mechanism of prediction and planning for end-to-end autonomous driving, called PPAD (Iterative Interaction of Prediction and Planning Autonomous Driving), which considers the timestep-wise interaction to better integrate prediction and planning. An ego vehicle performs motion planning at each timestep based on the trajectory prediction of surrounding agents (e.g., vehicles and pedestrians) and its local road conditions. Unlike existing end-to-end autonomous driving frameworks, PPAD models the interactions among ego, agents, and the dynamic environment in an autoregressive manner by interleaving the Prediction and Planning processes at every timestep, instead of a single sequential process of prediction followed by planning. Specifically, we design ego-to-agent, ego-to-map, and ego-to-BEV interaction mechanisms with hierarchical dynamic key objects attention to better model the interactions. The experiments on the nuScenes benchmark show that our approach outperforms state-of-the-art methods. Project page at https://github.com/zlichen/PPAD." + }, + { + "type": "text", + "bbox": [ + 0.263, + 0.574, + 0.573, + 0.588 + ], + "angle": 0, + "content": "Keywords: End-to-end Autonomous Driving" + }, + { + "type": "title", + "bbox": [ + 0.216, + 0.616, + 0.377, + 0.631 + ], + "angle": 0, + "content": "1 Introduction" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.649, + 0.788, + 0.724 + ], + "angle": 0, + "content": "The blossom of deep learning techniques has empowered autonomous driving, where many exciting milestones in autonomous driving have burst into our eyes owing to the convenient and interpretable discrete module designs. Recently, the planning-oriented [19] philosophy resonated with the community for pursuing a more effective end-to-end driving system, which is the focus of this work." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.725, + 0.788, + 0.802 + ], + "angle": 0, + "content": "Traditional methods in an autonomous driving system often break down the system into modular components, including localization, perception, tracking, prediction, planning, and control for interpretability and visibility. However, there are several drawbacks: 1) the accumulation of errors between modules becomes more significant as the system complexity increases. 2) the performance of" + }, + { + "type": "page_footnote", + "bbox": [ + 0.232, + 0.811, + 0.577, + 0.826 + ], + "angle": 0, + "content": "\\(^{\\dagger}\\)Work done during an internship at DeepRoute.AI." + }, + { + "type": "page_footnote", + "bbox": [ + 0.232, + 0.826, + 0.395, + 0.84 + ], + "angle": 0, + "content": "Corresponding author." + }, + { + "type": "list", + "bbox": [ + 0.232, + 0.811, + 0.577, + 0.84 + ], + "angle": 0, + "content": null + } + ], + [ + { + "type": "page_number", + "bbox": [ + 0.218, + 0.116, + 0.23, + 0.127 + ], + "angle": 0, + "content": "2" + }, + { + "type": "header", + "bbox": [ + 0.272, + 0.115, + 0.366, + 0.127 + ], + "angle": 0, + "content": "Z. Chen et al." + }, + { + "type": "image", + "bbox": [ + 0.218, + 0.145, + 0.41, + 0.197 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.218, + 0.199, + 0.407, + 0.212 + ], + "angle": 0, + "content": "(a) Traditional one-shot approach." + }, + { + "type": "image", + "bbox": [ + 0.412, + 0.145, + 0.785, + 0.198 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.513, + 0.199, + 0.713, + 0.212 + ], + "angle": 0, + "content": "(b) Our proposed PPAD framework." + }, + { + "type": "image_caption", + "bbox": [ + 0.214, + 0.224, + 0.788, + 0.349 + ], + "angle": 0, + "content": "Fig. 1: A high-level illustration of our proposed PPAD framework. The agent (in blue) intends to drive straight, while the ego (in red) plans to change lanes. Fig. 1(a) presents the typical one-shot method that might result in invalid motion plans and lead to an accident because of a lack of in-depth interactions. Fig. 1(b) demonstrates the gaming process between the ego and the agent under the PPAD architecture. During the prediction process, the agent executes an assertive plan by accelerating to stop the ego from blocking its route. The planning process of the ego plans trajectory based on the previous prediction process of the agent. The ego decelerates to avoid a potential accident and then changes lanes to achieve its driving goal." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.382, + 0.785, + 0.412 + ], + "angle": 0, + "content": "the downstream task is highly related to the upper stream module, which makes it very difficult to construct a unified data-driven infrastructure." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.414, + 0.788, + 0.55 + ], + "angle": 0, + "content": "Recently, end-to-end autonomous driving has gained popularity due to its simplicity. Two main lines are proposed based on the learning architecture. The first kind of method [10] takes the raw sensor data as input and directly outputs the planning trajectories or control command without any view transformation as intermediate representations for scene understanding. The other kinds of approaches [19,23] are built upon BEV representation and fully utilize the queries to generate the intermediate outputs as guidance for producing the planning results. One of the most significant advantages lies in the interpretability. In this work, we follow the design of the second kind of work." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.552, + 0.788, + 0.764 + ], + "angle": 0, + "content": "VAD [23] and UniAD [19] are typical one-shot motion planning methods, which only consider a single-step interaction between agents, ego-agent and surrounding environment (e.g., map elements). Think Twice [22] makes it a two-stage framework to enhance the gaming or interaction procedure. QCNet [56] and GameFormer [21] also recurrently model the trajectory prediction task. Given motion planning is a computational problem that finds a sequence of valid trajectories, often based on surrounding agents' forecasting, environmental understanding, and historical and future contexts. It can also be viewed as a game in which agents continuously plan their next move according to other agents' intentions and the encountering environment, further achieving their ultimate goals through incremental actions. To model these dynamic interactions of prediction and planning in end-to-end autonomous driving, it is crucial to consider the possible variance of predicted trajectories through multi-step modeling for planning feasible trajectories." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.765, + 0.788, + 0.841 + ], + "angle": 0, + "content": "Inspired by VAD [23], we aim to introduce the step-by-step Prediction-Planning into a learning-based framework. Intuitively, the prediction and planning modules can be modeled as a motion forecasting task, which predicts future waypoints by the given historical information. The results of prediction and planning modules at each time step are highly dependent on each other. Therefore," + } + ], + [ + { + "type": "header", + "bbox": [ + 0.687, + 0.115, + 0.732, + 0.127 + ], + "angle": 0, + "content": "PPAD" + }, + { + "type": "page_number", + "bbox": [ + 0.775, + 0.116, + 0.787, + 0.127 + ], + "angle": 0, + "content": "3" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.147, + 0.788, + 0.343 + ], + "angle": 0, + "content": "we need to consider the agent-agent and agent-environment interactions iteratively and bidirectionally to maximize the expectation of agents' prediction under the given observation of the other agents. We propose our PPAD to plan the ego agent's future trajectories step-by-step to model the timestep-wise bidirectional interaction or gaming in a vectorized learning framework as shown in Fig. 1. PPAD consists of the prediction and planning process. For each motion forecasting step, 1) Prediction process generates current step motion states by cross-attention and self-attention among agents and environment based on previous motion states to model the fine-grained bidirectional interactions. We take ego-agent-environment-BEV interaction into account to propagate features among all the traffic participants. 2) Planning process predicts the current step motion trajectories based on the expectation process. Our contributions are summarized as follows:" + }, + { + "type": "text", + "bbox": [ + 0.227, + 0.355, + 0.784, + 0.442 + ], + "angle": 0, + "content": "- We propose PPAD that optimizes ego-agent-environment interactions in an iterative prediction-planning manner. Iterative optimization could model the interactions and gaming better and more naturally in a planning task. The prediction process deals with more fine-grained and complex future uncertainties for multi-agent context learning, while the planning process plans a one-step future trajectory for the ego vehicle." + }, + { + "type": "text", + "bbox": [ + 0.227, + 0.446, + 0.784, + 0.488 + ], + "angle": 0, + "content": "- We model fine-grained interactions among the ego vehicle, agents, environment, and BEV features map, step-by-step with hierarchical dynamic key objects attention emphasizing on the spatial locality." + }, + { + "type": "text", + "bbox": [ + 0.228, + 0.491, + 0.784, + 0.533 + ], + "angle": 0, + "content": "- The experiments conducted on the nuScenes [3] and Argoverse [6,44] datasets have demonstrated the effectiveness of our approach over state-of-the-art approaches." + }, + { + "type": "list", + "bbox": [ + 0.227, + 0.355, + 0.784, + 0.533 + ], + "angle": 0, + "content": null + }, + { + "type": "title", + "bbox": [ + 0.216, + 0.558, + 0.388, + 0.574 + ], + "angle": 0, + "content": "2 Related Work" + }, + { + "type": "title", + "bbox": [ + 0.216, + 0.59, + 0.536, + 0.605 + ], + "angle": 0, + "content": "2.1 Multi-stage Autonomous Driving" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.614, + 0.788, + 0.842 + ], + "angle": 0, + "content": "Most autonomous driving systems are built upon the multi-stage design philosophy, which commonly consists of localization, perception, and planning. The perception module has been well studied recently due to the emergence of deep learning. Camera-based [20, 28, 29], Lidar-based [26, 48-50, 55] or fused-based [36, 46] approaches are proposed to fully exploit the potential of raw sensor data in order to produce accurate 3D objects prediction, semantic segmentation or tracking velocity. Prediction takes the outputs of the perception module to generate the future waypoints for the ego and the agents. Current approaches [1, 5, 11-13, 16, 30, 32, 40, 42, 48, 52, 56, 57] explore different representations to encode surrounding environment (map information) and agent interactions to predict final trajectories by regression or postprocessing sampling strategies. Some other works [4, 37] propose a joint perception and prediction framework, which aggregates historical information to generate tracklets with future trajectories. This unified learning framework could help address the nondifferential process and alleviate the unstable perception problem, compared" + } + ], + [ + { + "type": "page_number", + "bbox": [ + 0.218, + 0.116, + 0.23, + 0.127 + ], + "angle": 0, + "content": "4" + }, + { + "type": "header", + "bbox": [ + 0.272, + 0.115, + 0.365, + 0.127 + ], + "angle": 0, + "content": "Z. Chen et al." + }, + { + "type": "text", + "bbox": [ + 0.217, + 0.147, + 0.785, + 0.193 + ], + "angle": 0, + "content": "with previous works. Based on the perception and prediction results, planning module [1,7,12,39] plan its future behavior by cost-map optimization or learning-based approaches." + }, + { + "type": "title", + "bbox": [ + 0.217, + 0.213, + 0.542, + 0.229 + ], + "angle": 0, + "content": "2.2 End-To-End Autonomous Driving" + }, + { + "type": "text", + "bbox": [ + 0.216, + 0.237, + 0.787, + 0.449 + ], + "angle": 0, + "content": "Recently, more and more works have focused on end-to-end autonomous driving due to its merits in reducing internal accumulative errors and direct yet simple learning objectives. Typical methods [2,45] take the raw visual inputs to regress the final control command or trajectory points without any view transformations. To embrace reinforcement learning, a series of following works [8, 45, 54] utilize policy-based or valued-based approaches to improve driving behavior. With the popularity of BEV representation, more advanced architectures [9, 22] are introduced to attach more interpretability and help deal with the complex interactions in the driving scenarios. Another merit of BEV representation lies in its simplicity in fusing multi-modality sensors. Moreover, the modularized approaches [14, 18, 19, 23] decouple the end-to-end learning-based methods into several submodules or subtasks, while in a multi-task learning manner. The unified design could propagate and share learning context between modules through queries or feature maps." + }, + { + "type": "title", + "bbox": [ + 0.217, + 0.469, + 0.433, + 0.485 + ], + "angle": 0, + "content": "2.3 BEV Representation" + }, + { + "type": "text", + "bbox": [ + 0.216, + 0.493, + 0.787, + 0.645 + ], + "angle": 0, + "content": "BEV representation has gained significant prominence in the field of autonomous driving systems due to its inherent distortion-free characteristics and its simplicity in facilitating multi-sensor fusion. There are two main lines for BEV representation, including bottom-up and top-down ways. LSS [38] stands out as a bottom-up pioneering work that explores depth distribution for the 3D space frustum sampling to form BEV representation. Works [20, 28] optimize the pipeline through better depth estimation or lightweight sampling design. BEVFormer [29] and its following works [20, 33-35, 43, 47] adopt the top-down architecture, which uses a deformable transformer for the view transformations without depth supervision." + }, + { + "type": "title", + "bbox": [ + 0.217, + 0.666, + 0.331, + 0.682 + ], + "angle": 0, + "content": "3 Method" + }, + { + "type": "title", + "bbox": [ + 0.217, + 0.696, + 0.438, + 0.71 + ], + "angle": 0, + "content": "3.1 Framework Overview" + }, + { + "type": "text", + "bbox": [ + 0.216, + 0.72, + 0.787, + 0.84 + ], + "angle": 0, + "content": "We present the overall framework, PPAD, in Fig. 2, which comprises the principal modules of the Perception Transformer and our proposed Iterative Prediction-Planning Module. The Perception Transformer encodes the scene contexts into the BEV features map and further decodes as vectorized agents and map representations. The Iterative Prediction-Planning module consists of Prediction and Planning process in general. It dissects the dynamic interactions between the ego vehicle and the agents along the temporal dimension. Eventually, it predicts the motions of the agents and plans the future trajectory for the ego vehicle." + } + ], + [ + { + "type": "header", + "bbox": [ + 0.687, + 0.115, + 0.732, + 0.127 + ], + "angle": 0, + "content": "PPAD" + }, + { + "type": "page_number", + "bbox": [ + 0.775, + 0.117, + 0.785, + 0.127 + ], + "angle": 0, + "content": "5" + }, + { + "type": "image", + "bbox": [ + 0.218, + 0.143, + 0.789, + 0.389 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.214, + 0.398, + 0.789, + 0.568 + ], + "angle": 0, + "content": "Fig. 2: Overall architecture of our proposed self-driving framework, PPAD. It consists of the Perception Transformer and the Iterative Prediction-Planning Module. The Perception Transformer encodes scene contexts into agent queries, map queries, and BEV queries. Then, the Prediction-Planning Module interleaves the processes of the agent motion prediction and the ego planning for \\( N \\) times. Throughout the iterative Prediction and Planning processes, in-depth interactions are conducted among the ego, agents, map elements, and BEV features. In the Prediction process, the agent initially intends to go straight and is unaware of the potential motion of the ego. After interacting with the ego, map elements, and BEV features, the agent plans to be assertive and proceeds to accelerate. In the following Planning process, the ego knows the agent will accelerate through interacting with the updated agent query. It eventually plans to decelerate first and then conduct the lane change for safety reasons." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.597, + 0.785, + 0.627 + ], + "angle": 0, + "content": "Image Features Module uses a shared image backbone network (e.g., ResNet [15]) to extract image features for separate camera views." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.628, + 0.788, + 0.748 + ], + "angle": 0, + "content": "BEV Features Module transform the semantic features from the multi-view cameras into a united bird's-eye-view. Specifically, we inherit the encoder from BEVFormer [29, 47] to construct the BEV features. The grid-shape learnable BEV queries \\(\\mathbf{B} \\in \\mathbb{R}^{H \\times W \\times C}\\) are randomly initialized and learned to interact with the multi-view image features through deformable attention [58] to conduct spatial modeling. Temporal modeling is conducted in a recurrent manner, which applies the deformable attention between the current frame's BEV queries and the one from the previous time step." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.75, + 0.788, + 0.841 + ], + "angle": 0, + "content": "Vectorized Features Module Inspired by the VAD [23] paradigm, we also encode the scene contexts into vectorized representations through a detection decoder head [29, 58] and a map element decoding head [31], resulting in \\( N_{A} \\) of learned agent queries \\( \\mathbf{A} \\in \\mathbb{R}^{N_A \\times C} \\) and \\( N_{M} \\) of learned map queries \\( \\mathbf{M} \\in \\mathbb{R}^{N_M \\times C} \\). Separate MLP-based decoders will be attached to produce side output, which takes the learned queries as inputs and predicts with the agent attributes (lo" + } + ], + [ + { + "type": "page_number", + "bbox": [ + 0.218, + 0.116, + 0.23, + 0.127 + ], + "angle": 0, + "content": "6" + }, + { + "type": "header", + "bbox": [ + 0.272, + 0.115, + 0.365, + 0.127 + ], + "angle": 0, + "content": "Z. Chen et al." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.147, + 0.788, + 0.237 + ], + "angle": 0, + "content": "cations, dimensions, classes, etc.) or map attributes (classes and map vectors described by points). Additionally, the agent queries will be combined with the learnable motion embeddings for modeling the diverse motions of the agents. The agents with motions are represented as \\(\\mathbf{A} \\in \\mathbb{R}^{N_A \\times N_A^{mot} \\times C}\\). Similarly, the ego vehicle is modeled with three modes, representing the high-level driving commands of going straight, turn left, and turn right, in the form of \\(\\mathbf{E} \\in \\mathbb{R}^{N_E \\times N_E^{mot} \\times C}\\)." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.238, + 0.788, + 0.389 + ], + "angle": 0, + "content": "Iterative Prediction-Planning Module predicts the future trajectories for the ego vehicle and the agents in an interleaved fashion. Different from the traditional practice that predicts all the trajectories in one go, our PPAD framework articulates each step of motion planning by iterating the agent motion prediction and the ego planning processes. Thanks to the PPAD framework, we can conduct in-depth design to enforce key objects interactions (in Sec. 3.3) on scene contexts in a coarse-to-fine manner. We further improve the driving performance for the ego vehicle by taking the noisy trajectory as each step prediction and training the PPAD framework to reconstruct its original position at the following time step (in Sec. 3.4)." + }, + { + "type": "title", + "bbox": [ + 0.215, + 0.411, + 0.666, + 0.426 + ], + "angle": 0, + "content": "3.2 Iterative Interactions of Prediction and Planning" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.434, + 0.788, + 0.57 + ], + "angle": 0, + "content": "In the real world, the driving traffic changes constantly. Drivers plan and execute their decisions by ceaselessly reasoning the relationships among traffic participants in the scene. The planning task requires the self-driving system to have a good understanding of the scene and be capable of resolving the spatial-temporal causal factors. Therefore, we innovate the PPAD to dissect the planning task into multi-steps of the agent prediction and the ego planning processes and eventually promote consensuses among the ego's and the agents' future trajectories. The PPAD framework embodies the traffic interactions as gaming along space-time, producing a more accurate planning trajectory for the ego vehicle." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.571, + 0.788, + 0.632 + ], + "angle": 0, + "content": "Specifically, the ego and agents inherit the same philosophy of alternatively optimizing their motion behavior based on each other's motion forecasting at each future time step. In the following section, we will demonstrate the agent prediction process and elaborate on the details of the ego planning process." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.652, + 0.788, + 0.789 + ], + "angle": 0, + "content": "Prediction Process As illustrated in Fig 2, the agent predicts its subsequent step motion during the Prediction process, conditioned on the output of the ego vehicle's outcome from the previous Planning process. Specifically, the initial state of the agent query comprises its driving intention. It will then interact with the ego query updated from the previous planning process, which indicates the latest driving plan of the ego vehicle. After that, it will interact with map elements to choose the driving paths. At last, it gathers detailed geometric information by interacting with the BEV features and comes up with its precise next-step movement." + }, + { + "type": "text", + "bbox": [ + 0.215, + 0.81, + 0.788, + 0.843 + ], + "angle": 0, + "content": "Planning Process We consider the period of history with \\( T_{obs} \\) steps and the future with \\( T_{fut} \\) steps. The future trajectory of ego is denoted as \\( \\{p_E^t\\}^{t \\in T_{fut}} \\)." + } + ], + [ + { + "type": "header", + "bbox": [ + 0.687, + 0.115, + 0.732, + 0.127 + ], + "angle": 0, + "content": "PPAD" + }, + { + "type": "page_number", + "bbox": [ + 0.776, + 0.117, + 0.786, + 0.127 + ], + "angle": 0, + "content": "7" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.146, + 0.784, + 0.221 + ], + "angle": 0, + "content": "For each agent \\(a \\in \\mathbf{A}\\), the trajectory is represented as \\(\\{p_a^t\\}^{t \\in T_{fut}}\\). The positions of the detected map elements are denoted as \\(\\{p_m\\}_{m \\in \\mathbf{M}}\\). We define the operator of \\(\\mathcal{M}(p_E^t, p_{\\mathbf{A}} / p_{\\mathbf{M}}, s)\\) to mask out the agents or map elements that are beyond a distance of \\(s\\) towards the ego located at \\(p_E^t\\) to conduct key objects attention, and we will discuss the algorithm details in Sec. 3.3." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.223, + 0.785, + 0.343 + ], + "angle": 0, + "content": "Agent Interaction The resulting agents' motions from the prediction process are located at \\(\\{p_a^{t + 1}\\}_{a\\in \\mathbf{A}}\\), comprising the motion states up to the time step of \\(t + 1\\). Moving the ego from \\(p_E^t\\) to the future one-step of \\(p_E^{t + 1}\\), the ego vehicle should consider the agents' traffic globally and locally. From a more global perspective, the agents at a larger range provide more extensive information on traffic flow, which is essential in long-term trajectory planning. Regarding the spatial locality, the nearby agents are recognized as the key agents, which are supposed to be vitally related to the ego's driving decision." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.344, + 0.785, + 0.463 + ], + "angle": 0, + "content": "Therefore, we propose to conduct a hierarchical interaction with the agents through the attention mechanism to learn coarse-to-fine traffic context features for the ego. Centered at the ego's space-time position \\( p_E^t \\), we initially define a distance set of \\( \\mathbf{S} \\) of \\( \\{+\\infty m, 15m, 7.5m\\} \\) which covers the coarse-to-fine perception ranges. We formed the multi-scale agent sets by applying \\( \\mathcal{M}(p_E^t, p_{\\mathbf{A}}, s) \\) with different ranges. Then, the ego query interacts with the agents hierarchically through multi-head cross-attention, MHCA. We take the sum of the learned hierarchical attention results as the final values:" + }, + { + "type": "equation", + "bbox": [ + 0.317, + 0.472, + 0.785, + 0.502 + ], + "angle": 0, + "content": "\\[\n\\mathbf {E} ^ {k} = \\sum_ {s \\in \\mathbf {S}} \\operatorname {M H C A} (\\mathbf {E}, \\mathbf {A} ^ {k}, \\mathcal {M} \\left(p _ {\\mathbf {E}} ^ {t}, p _ {\\mathbf {A} ^ {k}} ^ {t + 1}, s\\right)), k \\in [ 1, N _ {A} ^ {\\text {m o t}} ], \\tag {1}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.511, + 0.784, + 0.556 + ], + "angle": 0, + "content": "where the ego independently queries information from different modes \\( k \\in N_A^{mot} \\) of agents, and then we stack the results output from different modes. Further, we apply the set operations to condense the features:" + }, + { + "type": "equation", + "bbox": [ + 0.33, + 0.565, + 0.785, + 0.583 + ], + "angle": 0, + "content": "\\[\n\\mathbf {E} ^ {\\prime} = \\operatorname {M A X} \\left(\\left[ \\mathbf {E} ^ {1},.., \\mathbf {E} ^ {N _ {A} ^ {\\text {m o t}}} \\right]\\right) + \\operatorname {M E A N} \\left(\\left[ \\mathbf {E} ^ {1},.., \\mathbf {E} ^ {N _ {A} ^ {\\text {m o t}}} \\right]\\right), \\tag {2}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.592, + 0.784, + 0.621 + ], + "angle": 0, + "content": "where the MAX and MEAN are applied to aggregate features along the agents' mode dimension and output with the updated ego query \\(\\mathbf{E}'\\)." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.622, + 0.785, + 0.697 + ], + "angle": 0, + "content": "Map Interaction Existing works [19, 23] tried to summarize all the required map information for planning by simply applying the global-level interaction once. They overlook the complexity of the evolving motion dynamic and overrate that the ego can plan precisely in the longer term by a single interaction with the map information." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.698, + 0.785, + 0.803 + ], + "angle": 0, + "content": "With our proposed PPAD framework, we can enrich the ego-map interaction by considering the ego's local road conditions based on its latest position. This results in better identifying the useful map information for each step of planning. The ego query interacts with the map queries in a similar practice as the interaction with the agents. The difference is that the map instances are not movable in the future time steps. The local and global map information can be abstracted into the ego query by MHCA:" + }, + { + "type": "equation", + "bbox": [ + 0.371, + 0.813, + 0.785, + 0.843 + ], + "angle": 0, + "content": "\\[\n\\mathbf {E} ^ {\\prime \\prime} = \\sum_ {s \\in \\mathbf {S}} \\operatorname {M H C A} \\left(\\mathbf {E} ^ {\\prime}, \\mathbf {M}, \\mathcal {M} \\left(p _ {\\mathbf {E}} ^ {t}, p _ {m}, s\\right)\\right), \\tag {3}\n\\]" + } + ], + [ + { + "type": "page_number", + "bbox": [ + 0.218, + 0.116, + 0.23, + 0.127 + ], + "angle": 0, + "content": "8" + }, + { + "type": "header", + "bbox": [ + 0.272, + 0.115, + 0.365, + 0.127 + ], + "angle": 0, + "content": "Z. Chen et al." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.147, + 0.784, + 0.177 + ], + "angle": 0, + "content": "where \\(\\mathbf{E}''\\) is the ego query updated with the critical map information for the next step of planning." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.179, + 0.788, + 0.407 + ], + "angle": 0, + "content": "BEV Interaction BEV is the fundamental representation of the whole system, which is the abstraction of the multi-camera features. Beyond the vectorized representation, there is other non-structural environmental information, including roads and fences. UniAD [19] models these non-structural pieces of stuff with an occupancy grid map. There are several drawbacks in UniAD: 1) occupancy grids consume large memory considering the whole scene's range. 2) UniAD failed to build the explicit interactions with the grid map. Therefore, we propose the BEV interactions that dynamically query the surrounding environment for each possible future step. This query process could help agents understand and learn the effects of their actions. Specifically, after applying the interactions above, the ego vehicle understands the dynamic agents' traffic better and knows its fronting road conditions. Nevertheless, planning a more precise motion requires the ego to comprehend the local detail geometric information. Hence, the ego query further interacts with BEV features to extract low-level geometric information. Specifically, we achieve by the deformable attention [58]:" + }, + { + "type": "equation", + "bbox": [ + 0.395, + 0.422, + 0.785, + 0.439 + ], + "angle": 0, + "content": "\\[\n\\mathbf {E} ^ {\\prime \\prime \\prime} = \\operatorname {D e f o r m A t t n} \\left(\\mathbf {E} ^ {\\prime \\prime}, p _ {\\mathbf {E}} ^ {t}, \\mathbf {B}\\right), \\tag {4}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.453, + 0.784, + 0.514 + ], + "angle": 0, + "content": "where \\( p_{\\mathbf{E}}^t \\) is the location of ego at time \\( t \\), and it serves as the reference point on the BEV features. The deformable attention DeformAttn applies sparse attention around the reference points \\( p_{\\mathbf{E}}^t \\) and learns to pick up the low-level geometric information from the BEV for planning." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.516, + 0.788, + 0.622 + ], + "angle": 0, + "content": "Motion Planning PPAD follows the same practice as [18, 19, 23], which uses the information of the high-level driving commands: go straight, turn left, and turn right. The concatenated features of \\(\\mathbf{h}_{\\mathbf{E}} = [\\mathbf{E}', \\mathbf{E}'', \\mathbf{E}''']\\) contain the information of the dynamic agent traffic, map semantics, and the precise environmental geometry. An MLP takes \\(\\mathbf{h}_{\\mathbf{E}}\\) as input and predicts the future one-step waypoint offset \\(w_{\\mathbf{E}}^{t+1} = (x, y)\\). We then update the ego state by applying another MLP on \\(\\mathbf{h}_{\\mathbf{E}}\\) for the next step of processing." + }, + { + "type": "title", + "bbox": [ + 0.215, + 0.654, + 0.51, + 0.67 + ], + "angle": 0, + "content": "3.3 Hierarchical Feature Learning" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.689, + 0.788, + 0.841 + ], + "angle": 0, + "content": "Hierarchical structure has a better capability to capture and recognize fine-grained patterns. For the driving scenarios, the driving behavior is based on scene understanding both globally and locally. Driving tends to focus on only a few key objects, which demonstrates the spatial locality or local attention. Therefore, we design hierarchical key objects attention to exploit the coarse-to-fine scene contexts. Specifically, given a set of distance ranges, we first find the key objects (agents or map elements) within the given range. Consequently, we apply dynamic local attention, which only considers the interactions among agents or map elements in the local area. The pseudo code shown in Alg. 1 delineates the implementation of dynamic key objects attention." + } + ], + [ + { + "type": "header", + "bbox": [ + 0.687, + 0.115, + 0.732, + 0.127 + ], + "angle": 0, + "content": "PPAD" + }, + { + "type": "page_number", + "bbox": [ + 0.776, + 0.117, + 0.786, + 0.127 + ], + "angle": 0, + "content": "9" + }, + { + "type": "code_caption", + "bbox": [ + 0.302, + 0.175, + 0.701, + 0.204 + ], + "angle": 0, + "content": "Algorithm 1 Pseudo code of Key objects attention in a PyTorch-like style." + }, + { + "type": "code", + "bbox": [ + 0.302, + 0.212, + 0.699, + 0.361 + ], + "angle": 0, + "content": "```python\n>>> layer = nn.MultiheadAttention(embed_dim, num_head)\n>>> def forward(layer, query, key, q_pos, k_pos, max_dis):\n # q_pos: position of the query with shape [B,Lq,2]\n # k_pos: position of the key with shape [B,Lk,2]\n # layer: attention layer as initialization\n # max_dis: distance threshold\n diff = (q_pos unsqueeze(2) - k_pos unsqueeze(1))\n dist = (diff ** 2).sum(-1).sqrt()\n attn_mask = (dist > max_dis).repeat(num_head,1,1)\n return attention(query, key, attn_mask=attn_mask)" + }, + { + "type": "title", + "bbox": [ + 0.216, + 0.403, + 0.516, + 0.419 + ], + "angle": 0, + "content": "3.4 Noisy Trajectory as Prediction" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.43, + 0.788, + 0.597 + ], + "angle": 0, + "content": "PPAD interleaves the prediction and the planning processes to plan the ego and agents' trajectories step-by-step. Expert driving knowledge is then enforced into the model through imitation learning. Thanks to our multi-step framework and the inspiration from [27], we introduce the noisy trajectory as the prediction to the PPAD while training. Specifically, we perturb each step of the ground truth ego trajectory by adding noise. The ego is then trained to predict the original next step waypoint offset of the ego regardless of disturbance on its starting noisy positions. The system is learned to predict the accurate waypoint offset by interacting with vectorized instances and the environment even though it starts at an inaccurate position. This strategy brings improvement to the planning performance." + }, + { + "type": "title", + "bbox": [ + 0.216, + 0.621, + 0.438, + 0.637 + ], + "angle": 0, + "content": "3.5 End-to-End Learning" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.647, + 0.785, + 0.678 + ], + "angle": 0, + "content": "Scene Context Loss Similar to VAD [23], we formulate the loss for the agents' motion and map as follows:" + }, + { + "type": "equation", + "bbox": [ + 0.407, + 0.691, + 0.786, + 0.709 + ], + "angle": 0, + "content": "\\[\n\\mathcal {L} _ {S} = \\lambda_ {1} \\mathcal {L} _ {\\text {a g e n t}} + \\lambda_ {2} \\mathcal {L} _ {\\text {m a p}}, \\tag {5}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.72, + 0.441, + 0.734 + ], + "angle": 0, + "content": "where \\(\\lambda_{1}\\) and \\(\\lambda_{2}\\) are set as 1.0." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.735, + 0.788, + 0.842 + ], + "angle": 0, + "content": "Constraint Loss Inspired by [23], we propose the confidence-aware collision loss \\(\\mathcal{L}_{CA - Col}\\), which considers the potential collision of all the agents' motion modalities instead of only computing the loss on the agents' most confident mode. We multiply the resulting collision loss from each mode with the predicted confidence score. For the trajectories having the potential to collide, it will penalize more when the predicted confidence score is higher. Combined with ego-boundary overstepping \\(\\mathcal{L}_{bd}\\) and ego-lane directional \\(\\mathcal{L}_{dir}\\) constraints proposed by [23], the" + } + ], + [ + { + "type": "page_number", + "bbox": [ + 0.218, + 0.116, + 0.236, + 0.127 + ], + "angle": 0, + "content": "10" + }, + { + "type": "header", + "bbox": [ + 0.272, + 0.115, + 0.365, + 0.127 + ], + "angle": 0, + "content": "Z. Chen et al." + }, + { + "type": "table", + "bbox": [ + 0.229, + 0.145, + 0.784, + 0.342 + ], + "angle": 0, + "content": "
MethodL2 (m) ↓Collision (%) ↓Latency (ms)FPS
1s2s3sAvg.1s2s3sAvg.
ST-P3 MetricsST-P3 [18]1.332.112.902.110.230.621.270.71628.31.6
VAD-Tiny [23]0.460.761.120.780.210.350.580.3859.516.8
VAD-Base [23]0.410.701.050.720.070.170.410.22224.34.5
OccNet [41]1.292.132.992.130.210.591.370.72--
FusionAD [51]---1.030.250.130.250.21--
Ours (Progress.)0.310.560.870.580.080.120.380.193852.6
Ours0.250.450.730.480.070.150.360.193852.6
UniAD MetricsNMP† [53]--2.31---1.92---
SA-NMP† [53]--2.05---1.59---
FF† [17]0.551.202.541.430.060.171.070.43--
EO† [24]0.671.362.781.600.040.090.880.33--
UniAD [19]0.480.961.651.030.050.170.710.31555.61.8
VAD-Base [23]0.501.021.691.070.000.300.950.42224.34.5
Ours (Progress.)0.380.831.450.890.020.200.930.383852.6
Ours0.300.691.260.750.030.220.730.333852.6
" + }, + { + "type": "table_caption", + "bbox": [ + 0.214, + 0.352, + 0.788, + 0.492 + ], + "angle": 0, + "content": "Table 1: Open-loop planning results on the nuScenes dataset [3]. The results of other methods are obtained from the original paper. We faithfully re-evaluate VAD [23] based on the UniAD [19] metrics. As for our PPAD, we provide two versions of results that utilize different training strategies. Progress means that we follow the progressive training pipeline as proposed in VAD [23], which trained all tasks except the planning task in the first 48 epochs and then finetuned with another 12 epochs for the planning task. The second row for our method trains the whole network for 60 epochs and then finetuned for another 12 epochs incorporating noisy trajectories. The latency of STP3 [18], VAD [23], and ours are measured on one NVIDIA Geforce RTX 3090 GPU, while UniAD is measured on one NVIDIA Tesla A100 GPU." + }, + { + "type": "text", + "bbox": [ + 0.216, + 0.538, + 0.393, + 0.551 + ], + "angle": 0, + "content": "overall constraint loss is" + }, + { + "type": "equation", + "bbox": [ + 0.37, + 0.568, + 0.786, + 0.584 + ], + "angle": 0, + "content": "\\[\n\\mathcal {L} _ {C} = \\lambda_ {3} \\mathcal {L} _ {C A - C o l} + \\lambda_ {4} \\mathcal {L} _ {b d} + \\lambda_ {5} \\mathcal {L} _ {d i r}, \\tag {6}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.599, + 0.589, + 0.614 + ], + "angle": 0, + "content": "where \\(\\lambda_3\\) and \\(\\lambda_4\\) are set as 1.0, and \\(\\lambda_5\\) is set as 0.5." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.615, + 0.787, + 0.647 + ], + "angle": 0, + "content": "Planning Loss We conduct \\( L_{1} \\) loss between each step of the ego's prediction \\( w_{\\mathbf{E}}^{t} \\) and the ground truth's waypoint offset \\( \\tilde{w}_{\\mathbf{E}}^{t} \\) along the future time horizon:" + }, + { + "type": "equation", + "bbox": [ + 0.387, + 0.661, + 0.786, + 0.703 + ], + "angle": 0, + "content": "\\[\n\\mathcal {L} _ {\\text {P l a n}} = \\frac {1}{T _ {f u t}} \\sum_ {t = 1} ^ {T _ {f u t}} \\| w _ {\\mathbf {E}} ^ {t} - \\tilde {w} _ {\\mathbf {E}} ^ {t} \\| _ {1}. \\tag {7}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.718, + 0.788, + 0.78 + ], + "angle": 0, + "content": "The overall end-to-end trainable loss function is formed by the sum of the perception loss, constraint loss, and planning loss. The same constraint losses \\(\\mathcal{L}_C^{noisy}\\) and planning losses \\(\\mathcal{L}_{plan}^{noisy}\\) will be applied to the predictions taking the noisy trajectories as input:" + }, + { + "type": "equation", + "bbox": [ + 0.326, + 0.793, + 0.786, + 0.812 + ], + "angle": 0, + "content": "\\[\n\\mathcal {L} = \\mathcal {L} _ {S} + \\zeta_ {1} \\left(\\mathcal {L} _ {C} + \\mathcal {L} _ {\\text {P l a n}}\\right) + \\zeta_ {2} \\left(\\mathcal {L} _ {C} ^ {\\text {n o i s y}} + \\mathcal {L} _ {\\text {P l a n}} ^ {\\text {n o i s y}}\\right), \\tag {8}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.825, + 0.517, + 0.841 + ], + "angle": 0, + "content": "where \\(\\zeta_{1}\\) is set as 0.6 and \\(\\zeta_{2}\\) is set as 0.4." + } + ], + [ + { + "type": "header", + "bbox": [ + 0.687, + 0.115, + 0.732, + 0.127 + ], + "angle": 0, + "content": "PPAD" + }, + { + "type": "page_number", + "bbox": [ + 0.769, + 0.116, + 0.785, + 0.127 + ], + "angle": 0, + "content": "11" + }, + { + "type": "table", + "bbox": [ + 0.33, + 0.144, + 0.684, + 0.202 + ], + "angle": 0, + "content": "
MethodL2 (m) ↓Collision (%) ↓
1s2s3sAvg.1s2s3sAvg.
VAD [23]1.041.792.601.810.611.262.251.37
Ours0.731.301.981.340.510.851.310.89
" + }, + { + "type": "table_caption", + "bbox": [ + 0.216, + 0.212, + 0.787, + 0.24 + ], + "angle": 0, + "content": "Table 2: Open-loop planning results on the Argoverse dataset [3]. The results are evaluated under the ST-P3 [18] and VAD [23] metrics." + }, + { + "type": "table", + "bbox": [ + 0.288, + 0.267, + 0.726, + 0.338 + ], + "angle": 0, + "content": "
MethodDetectionMapMotion Forecasting
NDS ↑mAP↑mAP↑minADE (m)↓minFDE (m) ↓MR↓
VAD [23]0.4590.3290.4760.6780.8820.08
UniAD [19]0.4990.382-0.7081.020.13
Ours0.4650.3320.5190.6760.8890.07
" + }, + { + "type": "table_caption", + "bbox": [ + 0.271, + 0.348, + 0.731, + 0.363 + ], + "angle": 0, + "content": "Table 3: Results comparison on the tasks beyond the planning task." + }, + { + "type": "title", + "bbox": [ + 0.216, + 0.404, + 0.377, + 0.421 + ], + "angle": 0, + "content": "4 Experiments" + }, + { + "type": "title", + "bbox": [ + 0.216, + 0.437, + 0.429, + 0.453 + ], + "angle": 0, + "content": "4.1 Experimental Setup" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.462, + 0.788, + 0.598 + ], + "angle": 0, + "content": "Dataset We evaluate our method on two challenging large-scale real-world datasets, nuScenes [3] and Argoverse2 [44]. We conduct ablation studies to evaluate the effectiveness of our proposed components on the nuScenes [3] dataset. The nuScenes dataset [3] provides about 1K 20-second diverse driving scenes collected in Boston, Pittsburgh, Las Vegas, and Singapore for open-loop settings. Key samples which contain 6 camera images are annotated at \\(2\\mathrm{Hz}\\). The Argoverse2 dataset comprises 1K 15-second scenes. It captures 7 RGB images on each frame. We align the data sampling frequency to nuScenes [3] by downsampling the annotated frames by an interval of 5 frames." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.599, + 0.788, + 0.749 + ], + "angle": 0, + "content": "Metrics We adopt the metrics of L2 Displacement Error (L2) and Collision Rate (CR) for evaluation [18]. L2 is measured between the prediction and ground-truth trajectories over the timesteps 1-s, 2-s, and 3-s, evaluating the trajectory quality. Collision Rate (CR) measures how often the collision occurs between the ego vehicle and the other agents along the planning horizon, reflecting the trajectory safety. We notice that the UniAD [19] adheres to different calculations from ST-P3 [18] and VAD [23]: the former [19] reports the evaluations at each second. In contrast, the latters [18, 23] reports the results of the cumulative average by each second. We faithfully make comparisons of our method to others with these two kinds of computations." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.75, + 0.788, + 0.841 + ], + "angle": 0, + "content": "Implementation Details We strictly follow the standard settings as proposed by UniAD [19] and VAD [23], which did not use the information of the historical ego trajectory. Same as VAD [23], the perception range is \\(60m \\times 30m\\). PPAD also recurrently encodes 2-s historical information into BEV and predicts the 3-s trajectory in the future. It also conducts the tasks of motion prediction and map construction. Our PPAD trains with a batch size of 1 using the Adam [25]" + } + ], + [ + { + "type": "page_number", + "bbox": [ + 0.218, + 0.116, + 0.236, + 0.127 + ], + "angle": 0, + "content": "12" + }, + { + "type": "header", + "bbox": [ + 0.272, + 0.115, + 0.365, + 0.127 + ], + "angle": 0, + "content": "Z. Chen et al." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.147, + 0.784, + 0.177 + ], + "angle": 0, + "content": "optimizer set with an initial learning rate of 2e-4. It takes us about six days to conduct the end-to-end training on eight NVIDIA A30 24GB GPUs." + }, + { + "type": "title", + "bbox": [ + 0.216, + 0.198, + 0.373, + 0.212 + ], + "angle": 0, + "content": "4.2 Main Results" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.222, + 0.785, + 0.327 + ], + "angle": 0, + "content": "Planning Results As shown in Tab. 1, our PPAD outperforms the current state-of-the-art performance by a large margin. Especially for the L2 distance metrics, there are about \\(20\\%\\) of consistent improvements can be observed along the temporal horizon. Thanks to the iterative interaction of prediction and planning, PPAD can help avoid collisions, leading to better results on collision rate compared to the one-shot representative VAD [23]. At the same time, we maintain a competitive efficiency compared to UniAD [19]." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.328, + 0.785, + 0.386 + ], + "angle": 0, + "content": "We further make a fair comparison between our method and the baseline, VAD [23], on the other dataset of Argoverse 2 [6,44]. In Tab. 2, our method can consistently outperform the baseline with a clear margin in both L2 distance and collision rate metrics." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.388, + 0.785, + 0.463 + ], + "angle": 0, + "content": "Subtasks Results To demonstrate the overall performance of our PPAD, we also provide the evaluation results besides planning metrics on the traditional perception and motion forecasting task in Tab. 3. Our PPAD also achieves promising performance in upstreaming perception and prediction tasks, which demonstrates that the whole system is jointly optimized." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.464, + 0.785, + 0.523 + ], + "angle": 0, + "content": "Qualitative Results We provide qualitative results shown in Fig. 3. PPAD can perceive the scene precisely and predict with reasonable and diverse motions for the surrounding agents. It also plans a smooth and accurate trajectory for the ego vehicle." + }, + { + "type": "title", + "bbox": [ + 0.216, + 0.545, + 0.388, + 0.56 + ], + "angle": 0, + "content": "4.3 Ablation Study" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.569, + 0.785, + 0.598 + ], + "angle": 0, + "content": "The following experiments adhere to the progressive training pipeline as proposed in VAD [23]." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.599, + 0.787, + 0.84 + ], + "angle": 0, + "content": "Effectiveness of Designs We provide ablation studies to verify the effectiveness of our proposed components. As shown in Tab. 4, the proposed PPAD framework (row 2) brings a remarkable improvement compared with one-shoot methods [23] (row 1). The multi-step interactions help the ego agent better understand the intention and potential effects brought by its actions along the temporal horizon, leading to an over \\(10\\%\\) L2 distance error reduction. We can observe a further improvement regarding the L2 distance with our proposed key objects attention (row 3). The slight degradation in collision rate might be due to the key objects attention being conducted on the ego with each mode of the agents, and the diverse modes of the agents mislead the behavior of the ego vehicle. From rows 5-7, the confidence-aware collision loss and noisy trajectory as the prediction can circumvent this phenomenon and further escalate the capability in planning accuracy and avoiding a collision. When we keep all of the remaining components while not applying key objects attention (row 4), we observed the degradations from row 7 in L2 \\((0.58\\mathrm{m}\\) vs. \\(0.59\\mathrm{m})\\) and collision rate \\((0.19\\%)\\) vs. \\(0.21\\%\\) , proving the effectiveness of the key objects attention." + } + ], + [ + { + "type": "header", + "bbox": [ + 0.687, + 0.115, + 0.732, + 0.127 + ], + "angle": 0, + "content": "PPAD" + }, + { + "type": "page_number", + "bbox": [ + 0.769, + 0.116, + 0.786, + 0.127 + ], + "angle": 0, + "content": "13" + }, + { + "type": "image", + "bbox": [ + 0.246, + 0.144, + 0.756, + 0.465 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.214, + 0.474, + 0.788, + 0.503 + ], + "angle": 0, + "content": "Fig. 3: Qualitative results of PPAD. The green box in the figure demonstrates the ego agent, while the red ones are agents." + }, + { + "type": "table", + "bbox": [ + 0.228, + 0.522, + 0.785, + 0.622 + ], + "angle": 0, + "content": "
PPADKey Objects Attn.\\( {\\mathcal{L}}_{CA - {Col}} \\)Noisy Traj.L2 (m) \\( \\downarrow \\)Collision (%) \\( \\downarrow \\)Latency (ms)
1s2s3sAvg.1s2s3sAvg.
1----0.410.701.050.720.070.170.410.22224
2✓---0.350.590.880.600.110.180.410.23318
3✓✓--0.320.550.830.570.190.280.580.35385
4✓-✓✓0.330.570.880.590.100.140.380.21318
5✓✓✓-0.350.590.890.610.080.140.320.18385
6✓✓-✓0.340.590.900.610.100.140.340.19385
7✓✓✓✓0.310.560.870.580.080.120.380.19385
" + }, + { + "type": "table_caption", + "bbox": [ + 0.214, + 0.632, + 0.788, + 0.703 + ], + "angle": 0, + "content": "Table 4: Component study for PPAD. Models follow the progressive training pipeline. PPAD means the auto-regressive framework with the designed stepwise interactions. Key Objects Attention represents hierarchical feature learning for the key objects. \\(\\mathcal{L}_{CA - Col}\\) represents the loss design for the confidence-aware collision loss. Noisy Traj. means that we incorporate noisy trajectories while training." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.75, + 0.788, + 0.841 + ], + "angle": 0, + "content": "Effectiveness of Interactions Our PPAD framework enables richer interactions among scene contexts, introducing local and global understandings of the world to the model. We further conduct ablation studies to demonstrate the performance gains brought by the interactions. We conduct the ablation study on interactions under the setting (Tab. 4 in row 5) without using the noisy trajectory as the prediction for better comparison. As illustrated in Tab. 5, we can achieve" + } + ], + [ + { + "type": "page_number", + "bbox": [ + 0.218, + 0.116, + 0.236, + 0.127 + ], + "angle": 0, + "content": "14" + }, + { + "type": "header", + "bbox": [ + 0.272, + 0.115, + 0.365, + 0.127 + ], + "angle": 0, + "content": "Z. Chen et al." + }, + { + "type": "table", + "bbox": [ + 0.339, + 0.144, + 0.67, + 0.21 + ], + "angle": 0, + "content": "
EAMapBEVL2 (m) ↓Collision (%) ↓
1s2s3sAvg.1s2s3sAvg.
✓--0.350.580.890.610.230.240.490.32
✓✓-0.380.640.950.660.170.180.480.28
✓✓✓0.350.590.890.610.080.140.320.18
" + }, + { + "type": "table_caption", + "bbox": [ + 0.216, + 0.22, + 0.787, + 0.262 + ], + "angle": 0, + "content": "Table 5: Interaction study for PPAD. EA, Map, BEV mean the interactions of the ego with agents, the ego and agents with the map, the ego and agents with the BEV, respectively." + }, + { + "type": "table", + "bbox": [ + 0.232, + 0.287, + 0.782, + 0.364 + ], + "angle": 0, + "content": "
Interaction IterationsL2 (m) ↓Collision (%) ↓Latency (ms)
1s2s3sAvg.1s2s3sAvg.
2(Every 1.5 sec)0.360.630.960.650.140.200.430.25306
3(Every 1.0 sec)0.370.630.950.650.100.150.390.21326
6(Every 0.5 sec)0.310.560.870.580.080.120.380.19385
" + }, + { + "type": "table_caption", + "bbox": [ + 0.216, + 0.374, + 0.787, + 0.403 + ], + "angle": 0, + "content": "Table 6: The ablation study of conducting different interaction iterations in the future time horizons." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.443, + 0.785, + 0.472 + ], + "angle": 0, + "content": "the best performance in terms of L2 distance and collision by incorporating all of the interactions." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.473, + 0.787, + 0.625 + ], + "angle": 0, + "content": "Effect of Different Iterations on Prediction and Planning Interaction Our innovative PPAD interaction mechanism not only better models motion planning as gaming among the ego and the agents but also enriches the ego's / agents' interactions with their local environments. We further conducted the ablation study on the different iterations in applying the interactions of prediction and planning, as shown in Tab. 6. Specifically, PPAD plans the trajectories with 3, 2, and 1 steps of waypoints after each of the prediction-planning processes for the interaction iterations of 2, 3, and 6 in Tab. 6. It is demonstrated that the performance reaches the best as we conduct the interactions of prediction and planning processes at every future step." + }, + { + "type": "title", + "bbox": [ + 0.216, + 0.645, + 0.36, + 0.661 + ], + "angle": 0, + "content": "5 Conclusion" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.674, + 0.787, + 0.841 + ], + "angle": 0, + "content": "In this paper, we have presented a novel autonomous driving framework, PPAD. Different from the previous methods that lack in-depth modeling of interactions, we pose the planning problem as a multi-step Prediction and Planning gaming process among the ego vehicle and agents. With PPAD architecture, our proposed hierarchical dynamic key objects attention is incorporated to learn local and global scene contexts at each step and eventually plan with a more precise trajectory. The confidence-aware collision constraint and noisy trajectories are utilized while training to improve driving safety further. In general, our proposed novel PPAD achieves compelling performance upon the existing state-of-the-art methods, and we hope the PPAD framework can inspire the community to further exploration." + } + ], + [ + { + "type": "header", + "bbox": [ + 0.687, + 0.115, + 0.732, + 0.127 + ], + "angle": 0, + "content": "PPAD" + }, + { + "type": "page_number", + "bbox": [ + 0.769, + 0.117, + 0.786, + 0.127 + ], + "angle": 0, + "content": "15" + }, + { + "type": "title", + "bbox": [ + 0.218, + 0.145, + 0.403, + 0.163 + ], + "angle": 0, + "content": "Acknowledgements" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.179, + 0.796, + 0.24 + ], + "angle": 0, + "content": "The authors are thankful for the financial support from the Hetao Shenzhen-HongKong Science and Technology Innovation Cooperation Zone (HZQB-KCZYZ-2021055), this work was also supported by Shenzhen Deeproute.ai Co., Ltd (HZQB-KCZYZ-2021055)." + }, + { + "type": "title", + "bbox": [ + 0.217, + 0.264, + 0.323, + 0.279 + ], + "angle": 0, + "content": "References" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.296, + 0.785, + 0.325 + ], + "angle": 0, + "content": "1. Bansal, M., Krizhevsky, A., Ogale, A.: Chauffeurnet. In: Robotics: Science and Systems XV (2019)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.326, + 0.785, + 0.367 + ], + "angle": 0, + "content": "2. Bojarski, M., Del Testa, D., Dworakowski, D., Firner, B., Flepp, B., Goyal, P., Jackel, L.D., Monfort, M., Muller, U., Zhang, J., et al.: End to end learning for self-driving cars. arXiv preprint arXiv:1604.07316 (2016)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.368, + 0.785, + 0.422 + ], + "angle": 0, + "content": "3. Caesar, H., Bankiti, V., Lang, A.H., Vora, S., Liong, V.E., Xu, Q., Krishnan, A., Pan, Y., Baldan, G., Beijbom, O.: nuscenes: A multimodal dataset for autonomous driving. In: Proceedings of the IEEE/CVF conference on computer vision and pattern recognition. pp. 11621-11631 (2020)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.423, + 0.785, + 0.45 + ], + "angle": 0, + "content": "4. Casas, S., Luo, W., Urtasun, R.: Intentnet: Learning to predict intention from raw sensor data. In: Conference on Robot Learning. pp. 947-956 (2018)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.451, + 0.785, + 0.491 + ], + "angle": 0, + "content": "5. Chai, Y., Sapp, B., Bansal, M., Anguelov, D.: Multipath: Multiple probabilistic anchor trajectory hypotheses for behavior prediction. arXiv preprint arXiv:1910.05449 (2019)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.493, + 0.785, + 0.547 + ], + "angle": 0, + "content": "6. Chang, M.F., Lambert, J., Sangkloy, P., Singh, J., Bak, S., Hartnett, A., Wang, D., Carr, P., Lucey, S., Ramanan, D., et al.: Argoverse: 3d tracking and forecasting with rich maps. In: Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition. pp. 8748-8757 (2019)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.548, + 0.785, + 0.589 + ], + "angle": 0, + "content": "7. Chekroun, R., Toromanoff, M., Hornauer, S., Moutarde, F.: Gri: General reinforced imitation and its application to vision-based autonomous driving. arXiv preprint arXiv:2111.08575 (2021)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.59, + 0.785, + 0.617 + ], + "angle": 0, + "content": "8. Chen, D., Zhou, B., Koltun, V., Krahenbuhl, P.: Learning by cheating. In: Conference on Robot Learning. pp. 66-75. PMLR (2020)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.618, + 0.785, + 0.659 + ], + "angle": 0, + "content": "9. Chitta, K., Prakash, A., Jaeger, B., Yu, Z., Renz, K., Geiger, A.: Transfuser: Imitation with transformer-based sensor fusion for autonomous driving. IEEE Transactions on Pattern Analysis and Machine Intelligence (2022)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.66, + 0.785, + 0.701 + ], + "angle": 0, + "content": "10. Codevilla, F., Santana, E., López, A.M., Gaidon, A.: Exploring the limitations of behavior cloning for autonomous driving. In: Proceedings of the IEEE/CVF International Conference on Computer Vision. pp. 9329-9338 (2019)" + }, + { + "type": "ref_text", + "bbox": [ + 0.22, + 0.702, + 0.785, + 0.757 + ], + "angle": 0, + "content": "11. Duvenaud, D.K., Maclaurin, D., Iparraguirre, J., Bombarell, R., Hirzel, T., Aspuru-Guzik, A., Adams, R.P.: Convolutional networks on graphs for learning molecular fingerprints. In: Advances in neural information processing systems. pp. 2224-2232 (2015)" + }, + { + "type": "ref_text", + "bbox": [ + 0.22, + 0.758, + 0.785, + 0.785 + ], + "angle": 0, + "content": "12. Fan, H., Zhu, F., Liu, C., Zhang, L., Zhuang, L., Li, D., Zhu, W., Hu, J., Li, H., Kong, Q.: BaiduApollo em motion planner. arXiv preprint arXiv:1807.08048 (2018)" + }, + { + "type": "ref_text", + "bbox": [ + 0.22, + 0.786, + 0.785, + 0.84 + ], + "angle": 0, + "content": "13. Gao, J., Sun, C., Zhao, H., Shen, Y., Anguelov, D., Li, C., Schmid, C.: Vectornet: Encoding hd maps and agent dynamics from vectorized representation. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 11525-11533 (2020)" + }, + { + "type": "list", + "bbox": [ + 0.22, + 0.296, + 0.785, + 0.84 + ], + "angle": 0, + "content": null + } + ], + [ + { + "type": "page_number", + "bbox": [ + 0.218, + 0.116, + 0.236, + 0.127 + ], + "angle": 0, + "content": "16" + }, + { + "type": "header", + "bbox": [ + 0.272, + 0.115, + 0.365, + 0.127 + ], + "angle": 0, + "content": "Z. Chen et al." + }, + { + "type": "ref_text", + "bbox": [ + 0.218, + 0.147, + 0.787, + 0.203 + ], + "angle": 0, + "content": "14. Gu, J., Hu, C., Zhang, T., Chen, X., Wang, Y., Wang, Y., Zhao, H.: Vip3d: End-to-end visual trajectory prediction via 3d agent queries. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 5496-5506 (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.218, + 0.203, + 0.787, + 0.244 + ], + "angle": 0, + "content": "15. He, K., Zhang, X., Ren, S., Sun, J.: Deep residual learning for image recognition. In: Proceedings of the IEEE conference on computer vision and pattern recognition. pp. 770-778 (2016)" + }, + { + "type": "ref_text", + "bbox": [ + 0.218, + 0.244, + 0.787, + 0.271 + ], + "angle": 0, + "content": "16. Henaff, M., Bruna, J., LeCun, Y.: Deep convolutional networks on graph-structured data. arXiv preprint arXiv:1506.05163 (2015)" + }, + { + "type": "ref_text", + "bbox": [ + 0.218, + 0.271, + 0.787, + 0.311 + ], + "angle": 0, + "content": "17. Hu, P., Huang, A., Dolan, J., Held, D., Ramanan, D.: Safe local motion planning with self-supervised freespace forecasting. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 12732-12741 (2021)" + }, + { + "type": "ref_text", + "bbox": [ + 0.218, + 0.311, + 0.787, + 0.352 + ], + "angle": 0, + "content": "18. Hu, S., Chen, L., Wu, P., Li, H., Yan, J., Tao, D.: St-p3: End-to-end vision-based autonomous driving via spatial-temporal feature learning. In: European Conference on Computer Vision. pp. 533-549. Springer (2022)" + }, + { + "type": "ref_text", + "bbox": [ + 0.218, + 0.352, + 0.787, + 0.405 + ], + "angle": 0, + "content": "19. Hu, Y., Yang, J., Chen, L., Li, K., Sima, C., Zhu, X., Chai, S., Du, S., Lin, T., Wang, W., et al.: Planning-oriented autonomous driving. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 17853-17862 (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.406, + 0.787, + 0.447 + ], + "angle": 0, + "content": "20. Huang, J., Huang, G., Zhu, Z., Ye, Y., Du, D.: Bevdet: High-performance multicamera 3d object detection in bird-eye-view. arXiv preprint arXiv:2112.11790 (2021)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.447, + 0.787, + 0.488 + ], + "angle": 0, + "content": "21. Huang, Z., Liu, H., Lv, C.: Gameformer: Game-theoretic modeling and learning of transformer-based interactive prediction and planning for autonomous driving. arXiv preprint arXiv:2303.05760 (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.488, + 0.787, + 0.542 + ], + "angle": 0, + "content": "22. Jia, X., Wu, P., Chen, L., Xie, J., He, C., Yan, J., Li, H.: Think twice before driving: Towards scalable decoders for end-to-end autonomous driving. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 21983-21994 (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.542, + 0.787, + 0.583 + ], + "angle": 0, + "content": "23. Jiang, B., Chen, S., Xu, Q., Liao, B., Chen, J., Zhou, H., Zhang, Q., Liu, W., Huang, C., Wang, X.: Vad: Vectorized scene representation for efficient autonomous driving. arXiv preprint arXiv:2303.12077 (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.583, + 0.787, + 0.624 + ], + "angle": 0, + "content": "24. Khurana, T., Hu, P., Dave, A., Ziglar, J., Held, D., Ramanan, D.: Differentiable raycasting for self-supervised occupancy forecasting. In: European Conference on Computer Vision. pp. 353-369. Springer (2022)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.624, + 0.787, + 0.65 + ], + "angle": 0, + "content": "25. Kingma, D.P., Ba, J.: Adam: A method for stochastic optimization. arXiv preprint arXiv:1412.6980 (2014)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.65, + 0.787, + 0.704 + ], + "angle": 0, + "content": "26. Lang, A.H., Vora, S., Caesar, H., Zhou, L., Yang, J., Beijbom, O.: PointPillars: Fast Encoders for Object Detection From Point Clouds. In: Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition. pp. 12697-12705 (2019)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.704, + 0.787, + 0.745 + ], + "angle": 0, + "content": "27. Li, F., Zhang, H., Liu, S., Guo, J., Ni, L.M., Zhang, L.: Dn-detr: Accelerate detr training by introducing query denoising. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 13619-13627 (2022)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.745, + 0.787, + 0.786 + ], + "angle": 0, + "content": "28. Li, Y., Ge, Z., Yu, G., Yang, J., Wang, Z., Shi, Y., Sun, J., Li, Z.: Bevdepth: Acquisition of reliable depth for multi-view 3d object detection. In: Proceedings of the AAAI Conference on Artificial Intelligence. pp. 1477-1485 (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.786, + 0.787, + 0.84 + ], + "angle": 0, + "content": "29. Li, Z., Wang, W., Li, H., Xie, E., Sima, C., Lu, T., Qiao, Y., Dai, J.: Bevformer: Learning bird's-eye-view representation from multi-camera images via spatiotemporal transformers. In: European conference on computer vision. pp. 1-18. Springer (2022)" + }, + { + "type": "list", + "bbox": [ + 0.217, + 0.147, + 0.787, + 0.84 + ], + "angle": 0, + "content": null + } + ], + [ + { + "type": "header", + "bbox": [ + 0.687, + 0.115, + 0.732, + 0.127 + ], + "angle": 0, + "content": "PPAD" + }, + { + "type": "page_number", + "bbox": [ + 0.769, + 0.116, + 0.786, + 0.127 + ], + "angle": 0, + "content": "17" + }, + { + "type": "ref_text", + "bbox": [ + 0.214, + 0.147, + 0.788, + 0.19 + ], + "angle": 0, + "content": "30. Liang, M., Yang, B., Hu, R., Chen, Y., Liao, R., Feng, S., Urtasun, R.: Learning lane graph representations for motion forecasting. In: Proceedings of the European Conference on Computer Vision (ECCV). pp. 541-556 (2020)" + }, + { + "type": "ref_text", + "bbox": [ + 0.214, + 0.19, + 0.787, + 0.233 + ], + "angle": 0, + "content": "31. Liao, B., Chen, S., Wang, X., Cheng, T., Zhang, Q., Liu, W., Huang, C.: Maptr: Structured modeling and learning for online vectorized hd map construction. arXiv preprint arXiv:2208.14437 (2022)" + }, + { + "type": "ref_text", + "bbox": [ + 0.214, + 0.233, + 0.787, + 0.275 + ], + "angle": 0, + "content": "32. Liu, Y., Zhang, J., Fang, L., Jiang, Q., Zhou, B.: Multimodal motion prediction with stacked transformers. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 7577-7586 (2021)" + }, + { + "type": "ref_text", + "bbox": [ + 0.214, + 0.276, + 0.787, + 0.318 + ], + "angle": 0, + "content": "33. Liu, Y., Wang, T., Zhang, X., Sun, J.: Petr: Position embedding transformation for multi-view 3d object detection. In: European Conference on Computer Vision. pp. 531-548. Springer (2022)" + }, + { + "type": "ref_text", + "bbox": [ + 0.214, + 0.318, + 0.787, + 0.36 + ], + "angle": 0, + "content": "34. Liu, Y., Yan, J., Jia, F., Li, S., Gao, A., Wang, T., Zhang, X., Sun, J.: Petrv2: A unified framework for 3d perception from multi-camera images. arXiv preprint arXiv:2206.01256 (2022)" + }, + { + "type": "ref_text", + "bbox": [ + 0.214, + 0.36, + 0.787, + 0.416 + ], + "angle": 0, + "content": "35. Liu, Z., Chen, S., Guo, X., Wang, X., Cheng, T., Zhu, H., Zhang, Q., Liu, W., Zhang, Y.: Vision-based uneven bev representation learning with polar rasterization and surface estimation. In: Conference on Robot Learning. pp. 437-446. PMLR (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.214, + 0.416, + 0.787, + 0.473 + ], + "angle": 0, + "content": "36. Liu, Z., Tang, H., Amini, A., Yang, X., Mao, H., Rus, D.L., Han, S.: Bevfusion: Multi-task multi-sensor fusion with unified bird's-eye view representation. In: 2023 IEEE International Conference on Robotics and Automation (ICRA). pp. 2774-2781. IEEE (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.214, + 0.473, + 0.787, + 0.529 + ], + "angle": 0, + "content": "37. Luo, W., Yang, B., Urtasun, R.: Fast and furious: Real time end-to-end 3d detection, tracking and motion forecasting with a single convolutional net. In: Proceedings of the IEEE conference on Computer Vision and Pattern Recognition. pp. 3569-3577 (2018)" + }, + { + "type": "ref_text", + "bbox": [ + 0.214, + 0.529, + 0.787, + 0.585 + ], + "angle": 0, + "content": "38. Philion, J., Fidler, S.: Lift, splat, shoot: Encoding images from arbitrary camera rigs by implicitly unprojecting to 3d. In: Computer Vision-ECCV 2020: 16th European Conference, Glasgow, UK, August 23-28, 2020, Proceedings, Part XIV 16. pp. 194-210. Springer (2020)" + }, + { + "type": "ref_text", + "bbox": [ + 0.214, + 0.585, + 0.787, + 0.628 + ], + "angle": 0, + "content": "39. Renz, K., Chitta, K., Mercea, O.B., Koepke, A., Akata, Z., Geiger, A.: Plant: Explainable planning transformers via object-level representations. arXiv preprint arXiv:2210.14222 (2022)" + }, + { + "type": "ref_text", + "bbox": [ + 0.214, + 0.628, + 0.787, + 0.684 + ], + "angle": 0, + "content": "40. Shuman, D.I., Narang, S.K., Frossard, P., Ortega, A., Vandergheynst, P.: The emerging field of signal processing on graphs: Extending high-dimensional data analysis to networks and other irregular domains. IEEE signal processing magazine 30(3), 83-98 (2013)" + }, + { + "type": "ref_text", + "bbox": [ + 0.214, + 0.684, + 0.787, + 0.713 + ], + "angle": 0, + "content": "41. Sima, C., Tong, W., Wang, T., Chen, L., Wu, S., Deng, H., Gu, Y., Lu, L., Luo, P., Lin, D., et al.: Scene as occupancy. arXiv e-prints pp. arXiv-2306 (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.214, + 0.713, + 0.787, + 0.756 + ], + "angle": 0, + "content": "42. Song, H., Luan, D., Ding, W., Wang, M.Y., Chen, Q.: Learning to predict vehicle trajectories with model-based planning. In: Conference on Robot Learning. pp. 1035-1045. PMLR (2021)" + }, + { + "type": "ref_text", + "bbox": [ + 0.214, + 0.756, + 0.787, + 0.798 + ], + "angle": 0, + "content": "43. Wang, S., Liu, Y., Wang, T., Li, Y., Zhang, X.: Exploring object-centric temporal modeling for efficient multi-view 3d object detection. arXiv preprint arXiv:2303.11926 (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.214, + 0.798, + 0.787, + 0.842 + ], + "angle": 0, + "content": "44. Wilson, B., Qi, W., Agarwal, T., Lambert, J., Singh, J., Khandelwal, S., Pan, B., Kumar, R., Hartnett, A., Pontes, J.K., et al.: Argoverse 2: Next generation datasets for self-driving perception and forecasting. arXiv preprint arXiv:2301.00493 (2023)" + }, + { + "type": "list", + "bbox": [ + 0.214, + 0.147, + 0.788, + 0.842 + ], + "angle": 0, + "content": null + } + ], + [ + { + "type": "page_number", + "bbox": [ + 0.218, + 0.116, + 0.236, + 0.127 + ], + "angle": 0, + "content": "18" + }, + { + "type": "header", + "bbox": [ + 0.272, + 0.115, + 0.365, + 0.127 + ], + "angle": 0, + "content": "Z. Chen et al." + }, + { + "type": "ref_text", + "bbox": [ + 0.218, + 0.147, + 0.785, + 0.189 + ], + "angle": 0, + "content": "45. Wu, P., Jia, X., Chen, L., Yan, J., Li, H., Qiao, Y.: Trajectory-guided control prediction for end-to-end autonomous driving: A simple yet strong baseline. Advances in Neural Information Processing Systems 35, 6119-6132 (2022)" + }, + { + "type": "ref_text", + "bbox": [ + 0.218, + 0.19, + 0.785, + 0.231 + ], + "angle": 0, + "content": "46. Xie, Y., Xu, C., Rakotosaona, M.J., Rim, P., Tombari, F., Keutzer, K., Tomizuka, M., Zhan, W.: Sparsefusion: Fusing multi-modal sparse representations for multisensor 3d object detection. arXiv preprint arXiv:2304.14340 (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.232, + 0.785, + 0.286 + ], + "angle": 0, + "content": "47. Yang, C., Chen, Y., Tian, H., Tao, C., Zhu, X., Zhang, Z., Huang, G., Li, H., Qiao, Y., Lu, L., et al.: Bevformer v2: Adapting modern image backbones to bird's-eye-view recognition via perspective supervision. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 17830-17839 (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.288, + 0.785, + 0.327 + ], + "angle": 0, + "content": "48. Ye, M., Cao, T., Chen, Q.: Tpcn: Temporal point cloud networks for motion forecasting. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 11318-11327 (2021)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.328, + 0.785, + 0.37 + ], + "angle": 0, + "content": "49. Ye, M., Xu, S., Cao, T.: Hvnet: Hybrid voxel network for lidar based 3d object detection. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 1631-1640 (2020)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.371, + 0.785, + 0.411 + ], + "angle": 0, + "content": "50. Ye, M., Xu, S., Cao, T., Chen, Q.: Drinet: A dual-representation iterative learning network for point cloud segmentation. In: Proceedings of the IEEE/CVF international conference on computer vision. pp. 7447-7456 (2021)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.412, + 0.785, + 0.452 + ], + "angle": 0, + "content": "51. Ye, T., Jing, W., Hu, C., Huang, S., Gao, L., Li, F., Wang, J., Guo, K., Xiao, W., Mao, W., et al.: Fusionad: Multi-modality fusion for prediction and planning tasks of autonomous driving. arXiv preprint arXiv:2308.01006 (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.453, + 0.785, + 0.479 + ], + "angle": 0, + "content": "52. Zeng, W., Liang, M., Liao, R., Urtasun, R.: Lanercnn: Distributed representations for graph-centric motion forecasting. arXiv preprint arXiv:2101.06653 (2021)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.481, + 0.785, + 0.522 + ], + "angle": 0, + "content": "53. Zeng, W., Luo, W., Suo, S., Sadat, A., Yang, B., Casas, S., Urtasun, R.: End-to-end interpretable neural motion planner. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 8660-8669 (2019)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.523, + 0.785, + 0.563 + ], + "angle": 0, + "content": "54. Zhang, Z., Liniger, A., Dai, D., Yu, F., Van Gool, L.: End-to-end urban driving by imitating a reinforcement learning coach. In: Proceedings of the IEEE/CVF international conference on computer vision. pp. 15222-15232 (2021)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.565, + 0.785, + 0.605 + ], + "angle": 0, + "content": "55. Zhou, Y., Tuzel, O.: VoxelNet: End-to-End Learning for Point Cloud Based 3D Object Detection. In: Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition. pp. 4490-4499 (2018)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.606, + 0.785, + 0.646 + ], + "angle": 0, + "content": "56. Zhou, Z., Wang, J., Li, Y.H., Huang, Y.K.: Query-centric trajectory prediction. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 17863-17873 (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.647, + 0.785, + 0.688 + ], + "angle": 0, + "content": "57. Zhou, Z., Ye, L., Wang, J., Wu, K., Lu, K.: Hivt: Hierarchical vector transformer for multi-agent motion prediction. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 8823-8833 (2022)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.689, + 0.785, + 0.729 + ], + "angle": 0, + "content": "58. Zhu, X., Su, W., Lu, L., Li, B., Wang, X., Dai, J.: Deformable detr: Deformable transformers for end-to-end object detection. arXiv preprint arXiv:2010.04159 (2020)" + }, + { + "type": "list", + "bbox": [ + 0.217, + 0.147, + 0.785, + 0.729 + ], + "angle": 0, + "content": null + } + ] +] \ No newline at end of file diff --git a/2024/PPAD_ Iterative Interactions of Prediction and Planning for End-to-end Autonomous Driving/2548ffc2-3e35-4690-aa4a-52b1b19d7a46_origin.pdf b/2024/PPAD_ Iterative Interactions of Prediction and Planning for End-to-end Autonomous Driving/2548ffc2-3e35-4690-aa4a-52b1b19d7a46_origin.pdf new file mode 100644 index 0000000000000000000000000000000000000000..45960ef6d2dd94b1e5f4e8ff908660d7200486f6 --- /dev/null +++ b/2024/PPAD_ Iterative Interactions of Prediction and Planning for End-to-end Autonomous Driving/2548ffc2-3e35-4690-aa4a-52b1b19d7a46_origin.pdf @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:8a43b6832a3bab9673691e7837e92abedd6d0253c8ea67bbc55847a04b6a2eaf +size 3666417 diff --git a/2024/PPAD_ Iterative Interactions of Prediction and Planning for End-to-end Autonomous Driving/full.md b/2024/PPAD_ Iterative Interactions of Prediction and Planning for End-to-end Autonomous Driving/full.md new file mode 100644 index 0000000000000000000000000000000000000000..18a951ec6c55338187928d48dd7a68e405702b6c --- /dev/null +++ b/2024/PPAD_ Iterative Interactions of Prediction and Planning for End-to-end Autonomous Driving/full.md @@ -0,0 +1,308 @@ +# PPAD: Iterative Interactions of Prediction and Planning for End-to-end Autonomous Driving + +Zhili Chen $^{1\dagger}$ , Maosheng Ye $^{1}$ , Shuangjie Xu $^{1}$ , Tongyi Cao $^{2}$ , and Qifeng Chen $^{1\boxtimes}$ + +$^{1}$ HKUST $^{2}$ DeepRoute.AI {zchenei, myeag, shuangjie.xu}@connect.ust.hk, tongyicao@deeproute.ai, cqf@cse.ust.hk + +Abstract. We present a new interaction mechanism of prediction and planning for end-to-end autonomous driving, called PPAD (Iterative Interaction of Prediction and Planning Autonomous Driving), which considers the timestep-wise interaction to better integrate prediction and planning. An ego vehicle performs motion planning at each timestep based on the trajectory prediction of surrounding agents (e.g., vehicles and pedestrians) and its local road conditions. Unlike existing end-to-end autonomous driving frameworks, PPAD models the interactions among ego, agents, and the dynamic environment in an autoregressive manner by interleaving the Prediction and Planning processes at every timestep, instead of a single sequential process of prediction followed by planning. Specifically, we design ego-to-agent, ego-to-map, and ego-to-BEV interaction mechanisms with hierarchical dynamic key objects attention to better model the interactions. The experiments on the nuScenes benchmark show that our approach outperforms state-of-the-art methods. Project page at https://github.com/zlichen/PPAD. + +Keywords: End-to-end Autonomous Driving + +# 1 Introduction + +The blossom of deep learning techniques has empowered autonomous driving, where many exciting milestones in autonomous driving have burst into our eyes owing to the convenient and interpretable discrete module designs. Recently, the planning-oriented [19] philosophy resonated with the community for pursuing a more effective end-to-end driving system, which is the focus of this work. + +Traditional methods in an autonomous driving system often break down the system into modular components, including localization, perception, tracking, prediction, planning, and control for interpretability and visibility. However, there are several drawbacks: 1) the accumulation of errors between modules becomes more significant as the system complexity increases. 2) the performance of + +![](images/8e37dc15643bdf9d25811804db201a98863293da2849adb8288c0d85806510a8.jpg) +(a) Traditional one-shot approach. + +![](images/e52a0eacb7fb6b059e97b0354127e1b1ed5780d7d8e706a927ef443b79522dee.jpg) +(b) Our proposed PPAD framework. +Fig. 1: A high-level illustration of our proposed PPAD framework. The agent (in blue) intends to drive straight, while the ego (in red) plans to change lanes. Fig. 1(a) presents the typical one-shot method that might result in invalid motion plans and lead to an accident because of a lack of in-depth interactions. Fig. 1(b) demonstrates the gaming process between the ego and the agent under the PPAD architecture. During the prediction process, the agent executes an assertive plan by accelerating to stop the ego from blocking its route. The planning process of the ego plans trajectory based on the previous prediction process of the agent. The ego decelerates to avoid a potential accident and then changes lanes to achieve its driving goal. + +the downstream task is highly related to the upper stream module, which makes it very difficult to construct a unified data-driven infrastructure. + +Recently, end-to-end autonomous driving has gained popularity due to its simplicity. Two main lines are proposed based on the learning architecture. The first kind of method [10] takes the raw sensor data as input and directly outputs the planning trajectories or control command without any view transformation as intermediate representations for scene understanding. The other kinds of approaches [19,23] are built upon BEV representation and fully utilize the queries to generate the intermediate outputs as guidance for producing the planning results. One of the most significant advantages lies in the interpretability. In this work, we follow the design of the second kind of work. + +VAD [23] and UniAD [19] are typical one-shot motion planning methods, which only consider a single-step interaction between agents, ego-agent and surrounding environment (e.g., map elements). Think Twice [22] makes it a two-stage framework to enhance the gaming or interaction procedure. QCNet [56] and GameFormer [21] also recurrently model the trajectory prediction task. Given motion planning is a computational problem that finds a sequence of valid trajectories, often based on surrounding agents' forecasting, environmental understanding, and historical and future contexts. It can also be viewed as a game in which agents continuously plan their next move according to other agents' intentions and the encountering environment, further achieving their ultimate goals through incremental actions. To model these dynamic interactions of prediction and planning in end-to-end autonomous driving, it is crucial to consider the possible variance of predicted trajectories through multi-step modeling for planning feasible trajectories. + +Inspired by VAD [23], we aim to introduce the step-by-step Prediction-Planning into a learning-based framework. Intuitively, the prediction and planning modules can be modeled as a motion forecasting task, which predicts future waypoints by the given historical information. The results of prediction and planning modules at each time step are highly dependent on each other. Therefore, + +we need to consider the agent-agent and agent-environment interactions iteratively and bidirectionally to maximize the expectation of agents' prediction under the given observation of the other agents. We propose our PPAD to plan the ego agent's future trajectories step-by-step to model the timestep-wise bidirectional interaction or gaming in a vectorized learning framework as shown in Fig. 1. PPAD consists of the prediction and planning process. For each motion forecasting step, 1) Prediction process generates current step motion states by cross-attention and self-attention among agents and environment based on previous motion states to model the fine-grained bidirectional interactions. We take ego-agent-environment-BEV interaction into account to propagate features among all the traffic participants. 2) Planning process predicts the current step motion trajectories based on the expectation process. Our contributions are summarized as follows: + +- We propose PPAD that optimizes ego-agent-environment interactions in an iterative prediction-planning manner. Iterative optimization could model the interactions and gaming better and more naturally in a planning task. The prediction process deals with more fine-grained and complex future uncertainties for multi-agent context learning, while the planning process plans a one-step future trajectory for the ego vehicle. +- We model fine-grained interactions among the ego vehicle, agents, environment, and BEV features map, step-by-step with hierarchical dynamic key objects attention emphasizing on the spatial locality. +- The experiments conducted on the nuScenes [3] and Argoverse [6,44] datasets have demonstrated the effectiveness of our approach over state-of-the-art approaches. + +# 2 Related Work + +# 2.1 Multi-stage Autonomous Driving + +Most autonomous driving systems are built upon the multi-stage design philosophy, which commonly consists of localization, perception, and planning. The perception module has been well studied recently due to the emergence of deep learning. Camera-based [20, 28, 29], Lidar-based [26, 48-50, 55] or fused-based [36, 46] approaches are proposed to fully exploit the potential of raw sensor data in order to produce accurate 3D objects prediction, semantic segmentation or tracking velocity. Prediction takes the outputs of the perception module to generate the future waypoints for the ego and the agents. Current approaches [1, 5, 11-13, 16, 30, 32, 40, 42, 48, 52, 56, 57] explore different representations to encode surrounding environment (map information) and agent interactions to predict final trajectories by regression or postprocessing sampling strategies. Some other works [4, 37] propose a joint perception and prediction framework, which aggregates historical information to generate tracklets with future trajectories. This unified learning framework could help address the nondifferential process and alleviate the unstable perception problem, compared + +with previous works. Based on the perception and prediction results, planning module [1,7,12,39] plan its future behavior by cost-map optimization or learning-based approaches. + +# 2.2 End-To-End Autonomous Driving + +Recently, more and more works have focused on end-to-end autonomous driving due to its merits in reducing internal accumulative errors and direct yet simple learning objectives. Typical methods [2,45] take the raw visual inputs to regress the final control command or trajectory points without any view transformations. To embrace reinforcement learning, a series of following works [8, 45, 54] utilize policy-based or valued-based approaches to improve driving behavior. With the popularity of BEV representation, more advanced architectures [9, 22] are introduced to attach more interpretability and help deal with the complex interactions in the driving scenarios. Another merit of BEV representation lies in its simplicity in fusing multi-modality sensors. Moreover, the modularized approaches [14, 18, 19, 23] decouple the end-to-end learning-based methods into several submodules or subtasks, while in a multi-task learning manner. The unified design could propagate and share learning context between modules through queries or feature maps. + +# 2.3 BEV Representation + +BEV representation has gained significant prominence in the field of autonomous driving systems due to its inherent distortion-free characteristics and its simplicity in facilitating multi-sensor fusion. There are two main lines for BEV representation, including bottom-up and top-down ways. LSS [38] stands out as a bottom-up pioneering work that explores depth distribution for the 3D space frustum sampling to form BEV representation. Works [20, 28] optimize the pipeline through better depth estimation or lightweight sampling design. BEVFormer [29] and its following works [20, 33-35, 43, 47] adopt the top-down architecture, which uses a deformable transformer for the view transformations without depth supervision. + +# 3 Method + +# 3.1 Framework Overview + +We present the overall framework, PPAD, in Fig. 2, which comprises the principal modules of the Perception Transformer and our proposed Iterative Prediction-Planning Module. The Perception Transformer encodes the scene contexts into the BEV features map and further decodes as vectorized agents and map representations. The Iterative Prediction-Planning module consists of Prediction and Planning process in general. It dissects the dynamic interactions between the ego vehicle and the agents along the temporal dimension. Eventually, it predicts the motions of the agents and plans the future trajectory for the ego vehicle. + +![](images/8e97feb90329d1f49e53f7e11f45c68f3e7f7caa7872baacdc6c73c24e31b0d1.jpg) +Fig. 2: Overall architecture of our proposed self-driving framework, PPAD. It consists of the Perception Transformer and the Iterative Prediction-Planning Module. The Perception Transformer encodes scene contexts into agent queries, map queries, and BEV queries. Then, the Prediction-Planning Module interleaves the processes of the agent motion prediction and the ego planning for $N$ times. Throughout the iterative Prediction and Planning processes, in-depth interactions are conducted among the ego, agents, map elements, and BEV features. In the Prediction process, the agent initially intends to go straight and is unaware of the potential motion of the ego. After interacting with the ego, map elements, and BEV features, the agent plans to be assertive and proceeds to accelerate. In the following Planning process, the ego knows the agent will accelerate through interacting with the updated agent query. It eventually plans to decelerate first and then conduct the lane change for safety reasons. + +Image Features Module uses a shared image backbone network (e.g., ResNet [15]) to extract image features for separate camera views. + +BEV Features Module transform the semantic features from the multi-view cameras into a united bird's-eye-view. Specifically, we inherit the encoder from BEVFormer [29, 47] to construct the BEV features. The grid-shape learnable BEV queries $\mathbf{B} \in \mathbb{R}^{H \times W \times C}$ are randomly initialized and learned to interact with the multi-view image features through deformable attention [58] to conduct spatial modeling. Temporal modeling is conducted in a recurrent manner, which applies the deformable attention between the current frame's BEV queries and the one from the previous time step. + +Vectorized Features Module Inspired by the VAD [23] paradigm, we also encode the scene contexts into vectorized representations through a detection decoder head [29, 58] and a map element decoding head [31], resulting in $N_{A}$ of learned agent queries $\mathbf{A} \in \mathbb{R}^{N_A \times C}$ and $N_{M}$ of learned map queries $\mathbf{M} \in \mathbb{R}^{N_M \times C}$ . Separate MLP-based decoders will be attached to produce side output, which takes the learned queries as inputs and predicts with the agent attributes (lo + +cations, dimensions, classes, etc.) or map attributes (classes and map vectors described by points). Additionally, the agent queries will be combined with the learnable motion embeddings for modeling the diverse motions of the agents. The agents with motions are represented as $\mathbf{A} \in \mathbb{R}^{N_A \times N_A^{mot} \times C}$ . Similarly, the ego vehicle is modeled with three modes, representing the high-level driving commands of going straight, turn left, and turn right, in the form of $\mathbf{E} \in \mathbb{R}^{N_E \times N_E^{mot} \times C}$ . + +Iterative Prediction-Planning Module predicts the future trajectories for the ego vehicle and the agents in an interleaved fashion. Different from the traditional practice that predicts all the trajectories in one go, our PPAD framework articulates each step of motion planning by iterating the agent motion prediction and the ego planning processes. Thanks to the PPAD framework, we can conduct in-depth design to enforce key objects interactions (in Sec. 3.3) on scene contexts in a coarse-to-fine manner. We further improve the driving performance for the ego vehicle by taking the noisy trajectory as each step prediction and training the PPAD framework to reconstruct its original position at the following time step (in Sec. 3.4). + +# 3.2 Iterative Interactions of Prediction and Planning + +In the real world, the driving traffic changes constantly. Drivers plan and execute their decisions by ceaselessly reasoning the relationships among traffic participants in the scene. The planning task requires the self-driving system to have a good understanding of the scene and be capable of resolving the spatial-temporal causal factors. Therefore, we innovate the PPAD to dissect the planning task into multi-steps of the agent prediction and the ego planning processes and eventually promote consensuses among the ego's and the agents' future trajectories. The PPAD framework embodies the traffic interactions as gaming along space-time, producing a more accurate planning trajectory for the ego vehicle. + +Specifically, the ego and agents inherit the same philosophy of alternatively optimizing their motion behavior based on each other's motion forecasting at each future time step. In the following section, we will demonstrate the agent prediction process and elaborate on the details of the ego planning process. + +Prediction Process As illustrated in Fig 2, the agent predicts its subsequent step motion during the Prediction process, conditioned on the output of the ego vehicle's outcome from the previous Planning process. Specifically, the initial state of the agent query comprises its driving intention. It will then interact with the ego query updated from the previous planning process, which indicates the latest driving plan of the ego vehicle. After that, it will interact with map elements to choose the driving paths. At last, it gathers detailed geometric information by interacting with the BEV features and comes up with its precise next-step movement. + +Planning Process We consider the period of history with $T_{obs}$ steps and the future with $T_{fut}$ steps. The future trajectory of ego is denoted as $\{p_E^t\}^{t \in T_{fut}}$ . + +For each agent $a \in \mathbf{A}$ , the trajectory is represented as $\{p_a^t\}^{t \in T_{fut}}$ . The positions of the detected map elements are denoted as $\{p_m\}_{m \in \mathbf{M}}$ . We define the operator of $\mathcal{M}(p_E^t, p_{\mathbf{A}} / p_{\mathbf{M}}, s)$ to mask out the agents or map elements that are beyond a distance of $s$ towards the ego located at $p_E^t$ to conduct key objects attention, and we will discuss the algorithm details in Sec. 3.3. + +Agent Interaction The resulting agents' motions from the prediction process are located at $\{p_a^{t + 1}\}_{a\in \mathbf{A}}$ , comprising the motion states up to the time step of $t + 1$ . Moving the ego from $p_E^t$ to the future one-step of $p_E^{t + 1}$ , the ego vehicle should consider the agents' traffic globally and locally. From a more global perspective, the agents at a larger range provide more extensive information on traffic flow, which is essential in long-term trajectory planning. Regarding the spatial locality, the nearby agents are recognized as the key agents, which are supposed to be vitally related to the ego's driving decision. + +Therefore, we propose to conduct a hierarchical interaction with the agents through the attention mechanism to learn coarse-to-fine traffic context features for the ego. Centered at the ego's space-time position $p_E^t$ , we initially define a distance set of $\mathbf{S}$ of $\{+\infty m, 15m, 7.5m\}$ which covers the coarse-to-fine perception ranges. We formed the multi-scale agent sets by applying $\mathcal{M}(p_E^t, p_{\mathbf{A}}, s)$ with different ranges. Then, the ego query interacts with the agents hierarchically through multi-head cross-attention, MHCA. We take the sum of the learned hierarchical attention results as the final values: + +$$ +\mathbf {E} ^ {k} = \sum_ {s \in \mathbf {S}} \operatorname {M H C A} (\mathbf {E}, \mathbf {A} ^ {k}, \mathcal {M} \left(p _ {\mathbf {E}} ^ {t}, p _ {\mathbf {A} ^ {k}} ^ {t + 1}, s\right)), k \in [ 1, N _ {A} ^ {\text {m o t}} ], \tag {1} +$$ + +where the ego independently queries information from different modes $k \in N_A^{mot}$ of agents, and then we stack the results output from different modes. Further, we apply the set operations to condense the features: + +$$ +\mathbf {E} ^ {\prime} = \operatorname {M A X} \left(\left[ \mathbf {E} ^ {1},.., \mathbf {E} ^ {N _ {A} ^ {\text {m o t}}} \right]\right) + \operatorname {M E A N} \left(\left[ \mathbf {E} ^ {1},.., \mathbf {E} ^ {N _ {A} ^ {\text {m o t}}} \right]\right), \tag {2} +$$ + +where the MAX and MEAN are applied to aggregate features along the agents' mode dimension and output with the updated ego query $\mathbf{E}'$ . + +Map Interaction Existing works [19, 23] tried to summarize all the required map information for planning by simply applying the global-level interaction once. They overlook the complexity of the evolving motion dynamic and overrate that the ego can plan precisely in the longer term by a single interaction with the map information. + +With our proposed PPAD framework, we can enrich the ego-map interaction by considering the ego's local road conditions based on its latest position. This results in better identifying the useful map information for each step of planning. The ego query interacts with the map queries in a similar practice as the interaction with the agents. The difference is that the map instances are not movable in the future time steps. The local and global map information can be abstracted into the ego query by MHCA: + +$$ +\mathbf {E} ^ {\prime \prime} = \sum_ {s \in \mathbf {S}} \operatorname {M H C A} \left(\mathbf {E} ^ {\prime}, \mathbf {M}, \mathcal {M} \left(p _ {\mathbf {E}} ^ {t}, p _ {m}, s\right)\right), \tag {3} +$$ + +where $\mathbf{E}''$ is the ego query updated with the critical map information for the next step of planning. + +BEV Interaction BEV is the fundamental representation of the whole system, which is the abstraction of the multi-camera features. Beyond the vectorized representation, there is other non-structural environmental information, including roads and fences. UniAD [19] models these non-structural pieces of stuff with an occupancy grid map. There are several drawbacks in UniAD: 1) occupancy grids consume large memory considering the whole scene's range. 2) UniAD failed to build the explicit interactions with the grid map. Therefore, we propose the BEV interactions that dynamically query the surrounding environment for each possible future step. This query process could help agents understand and learn the effects of their actions. Specifically, after applying the interactions above, the ego vehicle understands the dynamic agents' traffic better and knows its fronting road conditions. Nevertheless, planning a more precise motion requires the ego to comprehend the local detail geometric information. Hence, the ego query further interacts with BEV features to extract low-level geometric information. Specifically, we achieve by the deformable attention [58]: + +$$ +\mathbf {E} ^ {\prime \prime \prime} = \operatorname {D e f o r m A t t n} \left(\mathbf {E} ^ {\prime \prime}, p _ {\mathbf {E}} ^ {t}, \mathbf {B}\right), \tag {4} +$$ + +where $p_{\mathbf{E}}^t$ is the location of ego at time $t$ , and it serves as the reference point on the BEV features. The deformable attention DeformAttn applies sparse attention around the reference points $p_{\mathbf{E}}^t$ and learns to pick up the low-level geometric information from the BEV for planning. + +Motion Planning PPAD follows the same practice as [18, 19, 23], which uses the information of the high-level driving commands: go straight, turn left, and turn right. The concatenated features of $\mathbf{h}_{\mathbf{E}} = [\mathbf{E}', \mathbf{E}'', \mathbf{E}''']$ contain the information of the dynamic agent traffic, map semantics, and the precise environmental geometry. An MLP takes $\mathbf{h}_{\mathbf{E}}$ as input and predicts the future one-step waypoint offset $w_{\mathbf{E}}^{t+1} = (x, y)$ . We then update the ego state by applying another MLP on $\mathbf{h}_{\mathbf{E}}$ for the next step of processing. + +# 3.3 Hierarchical Feature Learning + +Hierarchical structure has a better capability to capture and recognize fine-grained patterns. For the driving scenarios, the driving behavior is based on scene understanding both globally and locally. Driving tends to focus on only a few key objects, which demonstrates the spatial locality or local attention. Therefore, we design hierarchical key objects attention to exploit the coarse-to-fine scene contexts. Specifically, given a set of distance ranges, we first find the key objects (agents or map elements) within the given range. Consequently, we apply dynamic local attention, which only considers the interactions among agents or map elements in the local area. The pseudo code shown in Alg. 1 delineates the implementation of dynamic key objects attention. + +Algorithm 1 Pseudo code of Key objects attention in a PyTorch-like style. +```python +>>> layer = nn.MultiheadAttention(embed_dim, num_head) +>>> def forward(layer, query, key, q_pos, k_pos, max_dis): + # q_pos: position of the query with shape [B,Lq,2] + # k_pos: position of the key with shape [B,Lk,2] + # layer: attention layer as initialization + # max_dis: distance threshold + diff = (q_pos unsqueeze(2) - k_pos unsqueeze(1)) + dist = (diff ** 2).sum(-1).sqrt() + attn_mask = (dist > max_dis).repeat(num_head,1,1) + return attention(query, key, attn_mask=attn_mask) +``` + +# 3.4 Noisy Trajectory as Prediction + +PPAD interleaves the prediction and the planning processes to plan the ego and agents' trajectories step-by-step. Expert driving knowledge is then enforced into the model through imitation learning. Thanks to our multi-step framework and the inspiration from [27], we introduce the noisy trajectory as the prediction to the PPAD while training. Specifically, we perturb each step of the ground truth ego trajectory by adding noise. The ego is then trained to predict the original next step waypoint offset of the ego regardless of disturbance on its starting noisy positions. The system is learned to predict the accurate waypoint offset by interacting with vectorized instances and the environment even though it starts at an inaccurate position. This strategy brings improvement to the planning performance. + +# 3.5 End-to-End Learning + +Scene Context Loss Similar to VAD [23], we formulate the loss for the agents' motion and map as follows: + +$$ +\mathcal {L} _ {S} = \lambda_ {1} \mathcal {L} _ {\text {a g e n t}} + \lambda_ {2} \mathcal {L} _ {\text {m a p}}, \tag {5} +$$ + +where $\lambda_{1}$ and $\lambda_{2}$ are set as 1.0. + +Constraint Loss Inspired by [23], we propose the confidence-aware collision loss $\mathcal{L}_{CA - Col}$ , which considers the potential collision of all the agents' motion modalities instead of only computing the loss on the agents' most confident mode. We multiply the resulting collision loss from each mode with the predicted confidence score. For the trajectories having the potential to collide, it will penalize more when the predicted confidence score is higher. Combined with ego-boundary overstepping $\mathcal{L}_{bd}$ and ego-lane directional $\mathcal{L}_{dir}$ constraints proposed by [23], the + +
MethodL2 (m) ↓Collision (%) ↓Latency (ms)FPS
1s2s3sAvg.1s2s3sAvg.
ST-P3 MetricsST-P3 [18]1.332.112.902.110.230.621.270.71628.31.6
VAD-Tiny [23]0.460.761.120.780.210.350.580.3859.516.8
VAD-Base [23]0.410.701.050.720.070.170.410.22224.34.5
OccNet [41]1.292.132.992.130.210.591.370.72--
FusionAD [51]---1.030.250.130.250.21--
Ours (Progress.)0.310.560.870.580.080.120.380.193852.6
Ours0.250.450.730.480.070.150.360.193852.6
UniAD MetricsNMP† [53]--2.31---1.92---
SA-NMP† [53]--2.05---1.59---
FF† [17]0.551.202.541.430.060.171.070.43--
EO† [24]0.671.362.781.600.040.090.880.33--
UniAD [19]0.480.961.651.030.050.170.710.31555.61.8
VAD-Base [23]0.501.021.691.070.000.300.950.42224.34.5
Ours (Progress.)0.380.831.450.890.020.200.930.383852.6
Ours0.300.691.260.750.030.220.730.333852.6
+ +Table 1: Open-loop planning results on the nuScenes dataset [3]. The results of other methods are obtained from the original paper. We faithfully re-evaluate VAD [23] based on the UniAD [19] metrics. As for our PPAD, we provide two versions of results that utilize different training strategies. Progress means that we follow the progressive training pipeline as proposed in VAD [23], which trained all tasks except the planning task in the first 48 epochs and then finetuned with another 12 epochs for the planning task. The second row for our method trains the whole network for 60 epochs and then finetuned for another 12 epochs incorporating noisy trajectories. The latency of STP3 [18], VAD [23], and ours are measured on one NVIDIA Geforce RTX 3090 GPU, while UniAD is measured on one NVIDIA Tesla A100 GPU. + +overall constraint loss is + +$$ +\mathcal {L} _ {C} = \lambda_ {3} \mathcal {L} _ {C A - C o l} + \lambda_ {4} \mathcal {L} _ {b d} + \lambda_ {5} \mathcal {L} _ {d i r}, \tag {6} +$$ + +where $\lambda_3$ and $\lambda_4$ are set as 1.0, and $\lambda_5$ is set as 0.5. + +Planning Loss We conduct $L_{1}$ loss between each step of the ego's prediction $w_{\mathbf{E}}^{t}$ and the ground truth's waypoint offset $\tilde{w}_{\mathbf{E}}^{t}$ along the future time horizon: + +$$ +\mathcal {L} _ {\text {P l a n}} = \frac {1}{T _ {f u t}} \sum_ {t = 1} ^ {T _ {f u t}} \| w _ {\mathbf {E}} ^ {t} - \tilde {w} _ {\mathbf {E}} ^ {t} \| _ {1}. \tag {7} +$$ + +The overall end-to-end trainable loss function is formed by the sum of the perception loss, constraint loss, and planning loss. The same constraint losses $\mathcal{L}_C^{noisy}$ and planning losses $\mathcal{L}_{plan}^{noisy}$ will be applied to the predictions taking the noisy trajectories as input: + +$$ +\mathcal {L} = \mathcal {L} _ {S} + \zeta_ {1} \left(\mathcal {L} _ {C} + \mathcal {L} _ {\text {P l a n}}\right) + \zeta_ {2} \left(\mathcal {L} _ {C} ^ {\text {n o i s y}} + \mathcal {L} _ {\text {P l a n}} ^ {\text {n o i s y}}\right), \tag {8} +$$ + +where $\zeta_{1}$ is set as 0.6 and $\zeta_{2}$ is set as 0.4. + +
MethodL2 (m) ↓Collision (%) ↓
1s2s3sAvg.1s2s3sAvg.
VAD [23]1.041.792.601.810.611.262.251.37
Ours0.731.301.981.340.510.851.310.89
+ +Table 2: Open-loop planning results on the Argoverse dataset [3]. The results are evaluated under the ST-P3 [18] and VAD [23] metrics. + +
MethodDetectionMapMotion Forecasting
NDS ↑mAP↑mAP↑minADE (m)↓minFDE (m) ↓MR↓
VAD [23]0.4590.3290.4760.6780.8820.08
UniAD [19]0.4990.382-0.7081.020.13
Ours0.4650.3320.5190.6760.8890.07
+ +Table 3: Results comparison on the tasks beyond the planning task. + +# 4 Experiments + +# 4.1 Experimental Setup + +Dataset We evaluate our method on two challenging large-scale real-world datasets, nuScenes [3] and Argoverse2 [44]. We conduct ablation studies to evaluate the effectiveness of our proposed components on the nuScenes [3] dataset. The nuScenes dataset [3] provides about 1K 20-second diverse driving scenes collected in Boston, Pittsburgh, Las Vegas, and Singapore for open-loop settings. Key samples which contain 6 camera images are annotated at $2\mathrm{Hz}$ . The Argoverse2 dataset comprises 1K 15-second scenes. It captures 7 RGB images on each frame. We align the data sampling frequency to nuScenes [3] by downsampling the annotated frames by an interval of 5 frames. + +Metrics We adopt the metrics of L2 Displacement Error (L2) and Collision Rate (CR) for evaluation [18]. L2 is measured between the prediction and ground-truth trajectories over the timesteps 1-s, 2-s, and 3-s, evaluating the trajectory quality. Collision Rate (CR) measures how often the collision occurs between the ego vehicle and the other agents along the planning horizon, reflecting the trajectory safety. We notice that the UniAD [19] adheres to different calculations from ST-P3 [18] and VAD [23]: the former [19] reports the evaluations at each second. In contrast, the latters [18, 23] reports the results of the cumulative average by each second. We faithfully make comparisons of our method to others with these two kinds of computations. + +Implementation Details We strictly follow the standard settings as proposed by UniAD [19] and VAD [23], which did not use the information of the historical ego trajectory. Same as VAD [23], the perception range is $60m \times 30m$ . PPAD also recurrently encodes 2-s historical information into BEV and predicts the 3-s trajectory in the future. It also conducts the tasks of motion prediction and map construction. Our PPAD trains with a batch size of 1 using the Adam [25] + +optimizer set with an initial learning rate of 2e-4. It takes us about six days to conduct the end-to-end training on eight NVIDIA A30 24GB GPUs. + +# 4.2 Main Results + +Planning Results As shown in Tab. 1, our PPAD outperforms the current state-of-the-art performance by a large margin. Especially for the L2 distance metrics, there are about $20\%$ of consistent improvements can be observed along the temporal horizon. Thanks to the iterative interaction of prediction and planning, PPAD can help avoid collisions, leading to better results on collision rate compared to the one-shot representative VAD [23]. At the same time, we maintain a competitive efficiency compared to UniAD [19]. + +We further make a fair comparison between our method and the baseline, VAD [23], on the other dataset of Argoverse 2 [6,44]. In Tab. 2, our method can consistently outperform the baseline with a clear margin in both L2 distance and collision rate metrics. + +Subtasks Results To demonstrate the overall performance of our PPAD, we also provide the evaluation results besides planning metrics on the traditional perception and motion forecasting task in Tab. 3. Our PPAD also achieves promising performance in upstreaming perception and prediction tasks, which demonstrates that the whole system is jointly optimized. + +Qualitative Results We provide qualitative results shown in Fig. 3. PPAD can perceive the scene precisely and predict with reasonable and diverse motions for the surrounding agents. It also plans a smooth and accurate trajectory for the ego vehicle. + +# 4.3 Ablation Study + +The following experiments adhere to the progressive training pipeline as proposed in VAD [23]. + +Effectiveness of Designs We provide ablation studies to verify the effectiveness of our proposed components. As shown in Tab. 4, the proposed PPAD framework (row 2) brings a remarkable improvement compared with one-shoot methods [23] (row 1). The multi-step interactions help the ego agent better understand the intention and potential effects brought by its actions along the temporal horizon, leading to an over $10\%$ L2 distance error reduction. We can observe a further improvement regarding the L2 distance with our proposed key objects attention (row 3). The slight degradation in collision rate might be due to the key objects attention being conducted on the ego with each mode of the agents, and the diverse modes of the agents mislead the behavior of the ego vehicle. From rows 5-7, the confidence-aware collision loss and noisy trajectory as the prediction can circumvent this phenomenon and further escalate the capability in planning accuracy and avoiding a collision. When we keep all of the remaining components while not applying key objects attention (row 4), we observed the degradations from row 7 in L2 $(0.58\mathrm{m}$ vs. $0.59\mathrm{m})$ and collision rate $(0.19\%)$ vs. $0.21\%$ , proving the effectiveness of the key objects attention. + +![](images/6f6db780942a50b689bf835fc86986a19030e231705f53fa316ba8778258c77c.jpg) +Fig. 3: Qualitative results of PPAD. The green box in the figure demonstrates the ego agent, while the red ones are agents. + +
PPADKey Objects Attn.\( {\mathcal{L}}_{CA - {Col}} \)Noisy Traj.L2 (m) \( \downarrow \)Collision (%) \( \downarrow \)Latency (ms)
1s2s3sAvg.1s2s3sAvg.
1----0.410.701.050.720.070.170.410.22224
2✓---0.350.590.880.600.110.180.410.23318
3✓✓--0.320.550.830.570.190.280.580.35385
4✓-✓✓0.330.570.880.590.100.140.380.21318
5✓✓✓-0.350.590.890.610.080.140.320.18385
6✓✓-✓0.340.590.900.610.100.140.340.19385
7✓✓✓✓0.310.560.870.580.080.120.380.19385
+ +Table 4: Component study for PPAD. Models follow the progressive training pipeline. PPAD means the auto-regressive framework with the designed stepwise interactions. Key Objects Attention represents hierarchical feature learning for the key objects. $\mathcal{L}_{CA - Col}$ represents the loss design for the confidence-aware collision loss. Noisy Traj. means that we incorporate noisy trajectories while training. + +Effectiveness of Interactions Our PPAD framework enables richer interactions among scene contexts, introducing local and global understandings of the world to the model. We further conduct ablation studies to demonstrate the performance gains brought by the interactions. We conduct the ablation study on interactions under the setting (Tab. 4 in row 5) without using the noisy trajectory as the prediction for better comparison. As illustrated in Tab. 5, we can achieve + +
EAMapBEVL2 (m) ↓Collision (%) ↓
1s2s3sAvg.1s2s3sAvg.
✓--0.350.580.890.610.230.240.490.32
✓✓-0.380.640.950.660.170.180.480.28
✓✓✓0.350.590.890.610.080.140.320.18
+ +Table 5: Interaction study for PPAD. EA, Map, BEV mean the interactions of the ego with agents, the ego and agents with the map, the ego and agents with the BEV, respectively. + +
Interaction IterationsL2 (m) ↓Collision (%) ↓Latency (ms)
1s2s3sAvg.1s2s3sAvg.
2(Every 1.5 sec)0.360.630.960.650.140.200.430.25306
3(Every 1.0 sec)0.370.630.950.650.100.150.390.21326
6(Every 0.5 sec)0.310.560.870.580.080.120.380.19385
+ +Table 6: The ablation study of conducting different interaction iterations in the future time horizons. + +the best performance in terms of L2 distance and collision by incorporating all of the interactions. + +Effect of Different Iterations on Prediction and Planning Interaction Our innovative PPAD interaction mechanism not only better models motion planning as gaming among the ego and the agents but also enriches the ego's / agents' interactions with their local environments. We further conducted the ablation study on the different iterations in applying the interactions of prediction and planning, as shown in Tab. 6. Specifically, PPAD plans the trajectories with 3, 2, and 1 steps of waypoints after each of the prediction-planning processes for the interaction iterations of 2, 3, and 6 in Tab. 6. It is demonstrated that the performance reaches the best as we conduct the interactions of prediction and planning processes at every future step. + +# 5 Conclusion + +In this paper, we have presented a novel autonomous driving framework, PPAD. Different from the previous methods that lack in-depth modeling of interactions, we pose the planning problem as a multi-step Prediction and Planning gaming process among the ego vehicle and agents. With PPAD architecture, our proposed hierarchical dynamic key objects attention is incorporated to learn local and global scene contexts at each step and eventually plan with a more precise trajectory. The confidence-aware collision constraint and noisy trajectories are utilized while training to improve driving safety further. In general, our proposed novel PPAD achieves compelling performance upon the existing state-of-the-art methods, and we hope the PPAD framework can inspire the community to further exploration. + +# Acknowledgements + +The authors are thankful for the financial support from the Hetao Shenzhen-HongKong Science and Technology Innovation Cooperation Zone (HZQB-KCZYZ-2021055), this work was also supported by Shenzhen Deeproute.ai Co., Ltd (HZQB-KCZYZ-2021055). + +# References + +1. Bansal, M., Krizhevsky, A., Ogale, A.: Chauffeurnet. In: Robotics: Science and Systems XV (2019) +2. Bojarski, M., Del Testa, D., Dworakowski, D., Firner, B., Flepp, B., Goyal, P., Jackel, L.D., Monfort, M., Muller, U., Zhang, J., et al.: End to end learning for self-driving cars. arXiv preprint arXiv:1604.07316 (2016) +3. Caesar, H., Bankiti, V., Lang, A.H., Vora, S., Liong, V.E., Xu, Q., Krishnan, A., Pan, Y., Baldan, G., Beijbom, O.: nuscenes: A multimodal dataset for autonomous driving. In: Proceedings of the IEEE/CVF conference on computer vision and pattern recognition. pp. 11621-11631 (2020) +4. Casas, S., Luo, W., Urtasun, R.: Intentnet: Learning to predict intention from raw sensor data. In: Conference on Robot Learning. pp. 947-956 (2018) +5. Chai, Y., Sapp, B., Bansal, M., Anguelov, D.: Multipath: Multiple probabilistic anchor trajectory hypotheses for behavior prediction. arXiv preprint arXiv:1910.05449 (2019) +6. Chang, M.F., Lambert, J., Sangkloy, P., Singh, J., Bak, S., Hartnett, A., Wang, D., Carr, P., Lucey, S., Ramanan, D., et al.: Argoverse: 3d tracking and forecasting with rich maps. In: Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition. pp. 8748-8757 (2019) +7. Chekroun, R., Toromanoff, M., Hornauer, S., Moutarde, F.: Gri: General reinforced imitation and its application to vision-based autonomous driving. arXiv preprint arXiv:2111.08575 (2021) +8. Chen, D., Zhou, B., Koltun, V., Krahenbuhl, P.: Learning by cheating. In: Conference on Robot Learning. pp. 66-75. PMLR (2020) +9. Chitta, K., Prakash, A., Jaeger, B., Yu, Z., Renz, K., Geiger, A.: Transfuser: Imitation with transformer-based sensor fusion for autonomous driving. IEEE Transactions on Pattern Analysis and Machine Intelligence (2022) +10. Codevilla, F., Santana, E., López, A.M., Gaidon, A.: Exploring the limitations of behavior cloning for autonomous driving. In: Proceedings of the IEEE/CVF International Conference on Computer Vision. pp. 9329-9338 (2019) +11. Duvenaud, D.K., Maclaurin, D., Iparraguirre, J., Bombarell, R., Hirzel, T., Aspuru-Guzik, A., Adams, R.P.: Convolutional networks on graphs for learning molecular fingerprints. In: Advances in neural information processing systems. pp. 2224-2232 (2015) +12. Fan, H., Zhu, F., Liu, C., Zhang, L., Zhuang, L., Li, D., Zhu, W., Hu, J., Li, H., Kong, Q.: BaiduApollo em motion planner. arXiv preprint arXiv:1807.08048 (2018) +13. Gao, J., Sun, C., Zhao, H., Shen, Y., Anguelov, D., Li, C., Schmid, C.: Vectornet: Encoding hd maps and agent dynamics from vectorized representation. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 11525-11533 (2020) + +14. Gu, J., Hu, C., Zhang, T., Chen, X., Wang, Y., Wang, Y., Zhao, H.: Vip3d: End-to-end visual trajectory prediction via 3d agent queries. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 5496-5506 (2023) +15. He, K., Zhang, X., Ren, S., Sun, J.: Deep residual learning for image recognition. In: Proceedings of the IEEE conference on computer vision and pattern recognition. pp. 770-778 (2016) +16. Henaff, M., Bruna, J., LeCun, Y.: Deep convolutional networks on graph-structured data. arXiv preprint arXiv:1506.05163 (2015) +17. Hu, P., Huang, A., Dolan, J., Held, D., Ramanan, D.: Safe local motion planning with self-supervised freespace forecasting. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 12732-12741 (2021) +18. Hu, S., Chen, L., Wu, P., Li, H., Yan, J., Tao, D.: St-p3: End-to-end vision-based autonomous driving via spatial-temporal feature learning. In: European Conference on Computer Vision. pp. 533-549. Springer (2022) +19. Hu, Y., Yang, J., Chen, L., Li, K., Sima, C., Zhu, X., Chai, S., Du, S., Lin, T., Wang, W., et al.: Planning-oriented autonomous driving. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 17853-17862 (2023) +20. Huang, J., Huang, G., Zhu, Z., Ye, Y., Du, D.: Bevdet: High-performance multicamera 3d object detection in bird-eye-view. arXiv preprint arXiv:2112.11790 (2021) +21. Huang, Z., Liu, H., Lv, C.: Gameformer: Game-theoretic modeling and learning of transformer-based interactive prediction and planning for autonomous driving. arXiv preprint arXiv:2303.05760 (2023) +22. Jia, X., Wu, P., Chen, L., Xie, J., He, C., Yan, J., Li, H.: Think twice before driving: Towards scalable decoders for end-to-end autonomous driving. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 21983-21994 (2023) +23. Jiang, B., Chen, S., Xu, Q., Liao, B., Chen, J., Zhou, H., Zhang, Q., Liu, W., Huang, C., Wang, X.: Vad: Vectorized scene representation for efficient autonomous driving. arXiv preprint arXiv:2303.12077 (2023) +24. Khurana, T., Hu, P., Dave, A., Ziglar, J., Held, D., Ramanan, D.: Differentiable raycasting for self-supervised occupancy forecasting. In: European Conference on Computer Vision. pp. 353-369. Springer (2022) +25. Kingma, D.P., Ba, J.: Adam: A method for stochastic optimization. arXiv preprint arXiv:1412.6980 (2014) +26. Lang, A.H., Vora, S., Caesar, H., Zhou, L., Yang, J., Beijbom, O.: PointPillars: Fast Encoders for Object Detection From Point Clouds. In: Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition. pp. 12697-12705 (2019) +27. Li, F., Zhang, H., Liu, S., Guo, J., Ni, L.M., Zhang, L.: Dn-detr: Accelerate detr training by introducing query denoising. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 13619-13627 (2022) +28. Li, Y., Ge, Z., Yu, G., Yang, J., Wang, Z., Shi, Y., Sun, J., Li, Z.: Bevdepth: Acquisition of reliable depth for multi-view 3d object detection. In: Proceedings of the AAAI Conference on Artificial Intelligence. pp. 1477-1485 (2023) +29. Li, Z., Wang, W., Li, H., Xie, E., Sima, C., Lu, T., Qiao, Y., Dai, J.: Bevformer: Learning bird's-eye-view representation from multi-camera images via spatiotemporal transformers. In: European conference on computer vision. pp. 1-18. Springer (2022) + +30. Liang, M., Yang, B., Hu, R., Chen, Y., Liao, R., Feng, S., Urtasun, R.: Learning lane graph representations for motion forecasting. In: Proceedings of the European Conference on Computer Vision (ECCV). pp. 541-556 (2020) +31. Liao, B., Chen, S., Wang, X., Cheng, T., Zhang, Q., Liu, W., Huang, C.: Maptr: Structured modeling and learning for online vectorized hd map construction. arXiv preprint arXiv:2208.14437 (2022) +32. Liu, Y., Zhang, J., Fang, L., Jiang, Q., Zhou, B.: Multimodal motion prediction with stacked transformers. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 7577-7586 (2021) +33. Liu, Y., Wang, T., Zhang, X., Sun, J.: Petr: Position embedding transformation for multi-view 3d object detection. In: European Conference on Computer Vision. pp. 531-548. Springer (2022) +34. Liu, Y., Yan, J., Jia, F., Li, S., Gao, A., Wang, T., Zhang, X., Sun, J.: Petrv2: A unified framework for 3d perception from multi-camera images. arXiv preprint arXiv:2206.01256 (2022) +35. Liu, Z., Chen, S., Guo, X., Wang, X., Cheng, T., Zhu, H., Zhang, Q., Liu, W., Zhang, Y.: Vision-based uneven bev representation learning with polar rasterization and surface estimation. In: Conference on Robot Learning. pp. 437-446. PMLR (2023) +36. Liu, Z., Tang, H., Amini, A., Yang, X., Mao, H., Rus, D.L., Han, S.: Bevfusion: Multi-task multi-sensor fusion with unified bird's-eye view representation. In: 2023 IEEE International Conference on Robotics and Automation (ICRA). pp. 2774-2781. IEEE (2023) +37. Luo, W., Yang, B., Urtasun, R.: Fast and furious: Real time end-to-end 3d detection, tracking and motion forecasting with a single convolutional net. In: Proceedings of the IEEE conference on Computer Vision and Pattern Recognition. pp. 3569-3577 (2018) +38. Philion, J., Fidler, S.: Lift, splat, shoot: Encoding images from arbitrary camera rigs by implicitly unprojecting to 3d. In: Computer Vision-ECCV 2020: 16th European Conference, Glasgow, UK, August 23-28, 2020, Proceedings, Part XIV 16. pp. 194-210. Springer (2020) +39. Renz, K., Chitta, K., Mercea, O.B., Koepke, A., Akata, Z., Geiger, A.: Plant: Explainable planning transformers via object-level representations. arXiv preprint arXiv:2210.14222 (2022) +40. Shuman, D.I., Narang, S.K., Frossard, P., Ortega, A., Vandergheynst, P.: The emerging field of signal processing on graphs: Extending high-dimensional data analysis to networks and other irregular domains. IEEE signal processing magazine 30(3), 83-98 (2013) +41. Sima, C., Tong, W., Wang, T., Chen, L., Wu, S., Deng, H., Gu, Y., Lu, L., Luo, P., Lin, D., et al.: Scene as occupancy. arXiv e-prints pp. arXiv-2306 (2023) +42. Song, H., Luan, D., Ding, W., Wang, M.Y., Chen, Q.: Learning to predict vehicle trajectories with model-based planning. In: Conference on Robot Learning. pp. 1035-1045. PMLR (2021) +43. Wang, S., Liu, Y., Wang, T., Li, Y., Zhang, X.: Exploring object-centric temporal modeling for efficient multi-view 3d object detection. arXiv preprint arXiv:2303.11926 (2023) +44. Wilson, B., Qi, W., Agarwal, T., Lambert, J., Singh, J., Khandelwal, S., Pan, B., Kumar, R., Hartnett, A., Pontes, J.K., et al.: Argoverse 2: Next generation datasets for self-driving perception and forecasting. arXiv preprint arXiv:2301.00493 (2023) + +45. Wu, P., Jia, X., Chen, L., Yan, J., Li, H., Qiao, Y.: Trajectory-guided control prediction for end-to-end autonomous driving: A simple yet strong baseline. Advances in Neural Information Processing Systems 35, 6119-6132 (2022) +46. Xie, Y., Xu, C., Rakotosaona, M.J., Rim, P., Tombari, F., Keutzer, K., Tomizuka, M., Zhan, W.: Sparsefusion: Fusing multi-modal sparse representations for multisensor 3d object detection. arXiv preprint arXiv:2304.14340 (2023) +47. Yang, C., Chen, Y., Tian, H., Tao, C., Zhu, X., Zhang, Z., Huang, G., Li, H., Qiao, Y., Lu, L., et al.: Bevformer v2: Adapting modern image backbones to bird's-eye-view recognition via perspective supervision. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 17830-17839 (2023) +48. Ye, M., Cao, T., Chen, Q.: Tpcn: Temporal point cloud networks for motion forecasting. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 11318-11327 (2021) +49. Ye, M., Xu, S., Cao, T.: Hvnet: Hybrid voxel network for lidar based 3d object detection. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 1631-1640 (2020) +50. Ye, M., Xu, S., Cao, T., Chen, Q.: Drinet: A dual-representation iterative learning network for point cloud segmentation. In: Proceedings of the IEEE/CVF international conference on computer vision. pp. 7447-7456 (2021) +51. Ye, T., Jing, W., Hu, C., Huang, S., Gao, L., Li, F., Wang, J., Guo, K., Xiao, W., Mao, W., et al.: Fusionad: Multi-modality fusion for prediction and planning tasks of autonomous driving. arXiv preprint arXiv:2308.01006 (2023) +52. Zeng, W., Liang, M., Liao, R., Urtasun, R.: Lanercnn: Distributed representations for graph-centric motion forecasting. arXiv preprint arXiv:2101.06653 (2021) +53. Zeng, W., Luo, W., Suo, S., Sadat, A., Yang, B., Casas, S., Urtasun, R.: End-to-end interpretable neural motion planner. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 8660-8669 (2019) +54. Zhang, Z., Liniger, A., Dai, D., Yu, F., Van Gool, L.: End-to-end urban driving by imitating a reinforcement learning coach. In: Proceedings of the IEEE/CVF international conference on computer vision. pp. 15222-15232 (2021) +55. Zhou, Y., Tuzel, O.: VoxelNet: End-to-End Learning for Point Cloud Based 3D Object Detection. In: Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition. pp. 4490-4499 (2018) +56. Zhou, Z., Wang, J., Li, Y.H., Huang, Y.K.: Query-centric trajectory prediction. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 17863-17873 (2023) +57. Zhou, Z., Ye, L., Wang, J., Wu, K., Lu, K.: Hivt: Hierarchical vector transformer for multi-agent motion prediction. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 8823-8833 (2022) +58. Zhu, X., Su, W., Lu, L., Li, B., Wang, X., Dai, J.: Deformable detr: Deformable transformers for end-to-end object detection. arXiv preprint arXiv:2010.04159 (2020) \ No newline at end of file diff --git a/2024/PPAD_ Iterative Interactions of Prediction and Planning for End-to-end Autonomous Driving/images.zip b/2024/PPAD_ Iterative Interactions of Prediction and Planning for End-to-end Autonomous Driving/images.zip new file mode 100644 index 0000000000000000000000000000000000000000..9ea08648b70aec94da8303141a3394554f0f83b0 --- /dev/null +++ b/2024/PPAD_ Iterative Interactions of Prediction and Planning for End-to-end Autonomous Driving/images.zip @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:e70385db26b8ad2ecc2aed3842537620c225d6ca44e7ee5910a50887a94a7551 +size 465141 diff --git a/2024/PPAD_ Iterative Interactions of Prediction and Planning for End-to-end Autonomous Driving/layout.json b/2024/PPAD_ Iterative Interactions of Prediction and Planning for End-to-end Autonomous Driving/layout.json new file mode 100644 index 0000000000000000000000000000000000000000..bd7eaf73a26ad2333b075ae1fa53f81ab7653b44 --- /dev/null +++ b/2024/PPAD_ Iterative Interactions of Prediction and Planning for End-to-end Autonomous Driving/layout.json @@ -0,0 +1,8596 @@ +{ + "pdf_info": [ + { + "para_blocks": [ + { + "bbox": [ + 143, + 111, + 473, + 148 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 143, + 111, + 473, + 148 + ], + "spans": [ + { + "bbox": [ + 143, + 111, + 473, + 148 + ], + "type": "text", + "content": "PPAD: Iterative Interactions of Prediction and Planning for End-to-end Autonomous Driving" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 197, + 167, + 416, + 194 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 197, + 167, + 416, + 194 + ], + "spans": [ + { + "bbox": [ + 197, + 167, + 416, + 194 + ], + "type": "text", + "content": "Zhili Chen" + }, + { + "bbox": [ + 197, + 167, + 416, + 194 + ], + "type": "inline_equation", + "content": "^{1\\dagger}" + }, + { + "bbox": [ + 197, + 167, + 416, + 194 + ], + "type": "text", + "content": ", Maosheng Ye" + }, + { + "bbox": [ + 197, + 167, + 416, + 194 + ], + "type": "inline_equation", + "content": "^{1}" + }, + { + "bbox": [ + 197, + 167, + 416, + 194 + ], + "type": "text", + "content": ", Shuangjie Xu" + }, + { + "bbox": [ + 197, + 167, + 416, + 194 + ], + "type": "inline_equation", + "content": "^{1}" + }, + { + "bbox": [ + 197, + 167, + 416, + 194 + ], + "type": "text", + "content": ", Tongyi Cao" + }, + { + "bbox": [ + 197, + 167, + 416, + 194 + ], + "type": "inline_equation", + "content": "^{2}" + }, + { + "bbox": [ + 197, + 167, + 416, + 194 + ], + "type": "text", + "content": ", and Qifeng Chen" + }, + { + "bbox": [ + 197, + 167, + 416, + 194 + ], + "type": "inline_equation", + "content": "^{1\\boxtimes}" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 198, + 201, + 416, + 236 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 198, + 201, + 416, + 236 + ], + "spans": [ + { + "bbox": [ + 198, + 201, + 416, + 236 + ], + "type": "inline_equation", + "content": "^{1}" + }, + { + "bbox": [ + 198, + 201, + 416, + 236 + ], + "type": "text", + "content": "HKUST " + }, + { + "bbox": [ + 198, + 201, + 416, + 236 + ], + "type": "inline_equation", + "content": "^{2}" + }, + { + "bbox": [ + 198, + 201, + 416, + 236 + ], + "type": "text", + "content": "DeepRoute.AI {zchenei, myeag, shuangjie.xu}@connect.ust.hk, tongyicao@deeproute.ai, cqf@cse.ust.hk" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 159, + 267, + 455, + 445 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 159, + 267, + 455, + 445 + ], + "spans": [ + { + "bbox": [ + 159, + 267, + 455, + 445 + ], + "type": "text", + "content": "Abstract. We present a new interaction mechanism of prediction and planning for end-to-end autonomous driving, called PPAD (Iterative Interaction of Prediction and Planning Autonomous Driving), which considers the timestep-wise interaction to better integrate prediction and planning. An ego vehicle performs motion planning at each timestep based on the trajectory prediction of surrounding agents (e.g., vehicles and pedestrians) and its local road conditions. Unlike existing end-to-end autonomous driving frameworks, PPAD models the interactions among ego, agents, and the dynamic environment in an autoregressive manner by interleaving the Prediction and Planning processes at every timestep, instead of a single sequential process of prediction followed by planning. Specifically, we design ego-to-agent, ego-to-map, and ego-to-BEV interaction mechanisms with hierarchical dynamic key objects attention to better model the interactions. The experiments on the nuScenes benchmark show that our approach outperforms state-of-the-art methods. Project page at https://github.com/zlichen/PPAD." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 160, + 454, + 350, + 465 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 160, + 454, + 350, + 465 + ], + "spans": [ + { + "bbox": [ + 160, + 454, + 350, + 465 + ], + "type": "text", + "content": "Keywords: End-to-end Autonomous Driving" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 132, + 487, + 230, + 499 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 487, + 230, + 499 + ], + "spans": [ + { + "bbox": [ + 132, + 487, + 230, + 499 + ], + "type": "text", + "content": "1 Introduction" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 130, + 514, + 482, + 573 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 514, + 482, + 573 + ], + "spans": [ + { + "bbox": [ + 130, + 514, + 482, + 573 + ], + "type": "text", + "content": "The blossom of deep learning techniques has empowered autonomous driving, where many exciting milestones in autonomous driving have burst into our eyes owing to the convenient and interpretable discrete module designs. Recently, the planning-oriented [19] philosophy resonated with the community for pursuing a more effective end-to-end driving system, which is the focus of this work." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 130, + 574, + 482, + 635 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 574, + 482, + 635 + ], + "spans": [ + { + "bbox": [ + 130, + 574, + 482, + 635 + ], + "type": "text", + "content": "Traditional methods in an autonomous driving system often break down the system into modular components, including localization, perception, tracking, prediction, planning, and control for interpretability and visibility. However, there are several drawbacks: 1) the accumulation of errors between modules becomes more significant as the system complexity increases. 2) the performance of" + } + ] + } + ], + "index": 7 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 141, + 642, + 353, + 654 + ], + "type": "page_footnote", + "angle": 0, + "lines": [ + { + "bbox": [ + 141, + 642, + 353, + 654 + ], + "spans": [ + { + "bbox": [ + 141, + 642, + 353, + 654 + ], + "type": "inline_equation", + "content": "^{\\dagger}" + }, + { + "bbox": [ + 141, + 642, + 353, + 654 + ], + "type": "text", + "content": "Work done during an internship at DeepRoute.AI." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 141, + 654, + 241, + 665 + ], + "type": "page_footnote", + "angle": 0, + "lines": [ + { + "bbox": [ + 141, + 654, + 241, + 665 + ], + "spans": [ + { + "bbox": [ + 141, + 654, + 241, + 665 + ], + "type": "text", + "content": "Corresponding author." + } + ] + } + ], + "index": 9 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 0 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 133, + 114, + 250, + 156 + ], + "blocks": [ + { + "bbox": [ + 133, + 114, + 250, + 156 + ], + "lines": [ + { + "bbox": [ + 133, + 114, + 250, + 156 + ], + "spans": [ + { + "bbox": [ + 133, + 114, + 250, + 156 + ], + "type": "image", + "image_path": "8e37dc15643bdf9d25811804db201a98863293da2849adb8288c0d85806510a8.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 133, + 157, + 249, + 167 + ], + "lines": [ + { + "bbox": [ + 133, + 157, + 249, + 167 + ], + "spans": [ + { + "bbox": [ + 133, + 157, + 249, + 167 + ], + "type": "text", + "content": "(a) Traditional one-shot approach." + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_caption" + } + ], + "index": 2 + }, + { + "type": "image", + "bbox": [ + 252, + 114, + 480, + 156 + ], + "blocks": [ + { + "bbox": [ + 252, + 114, + 480, + 156 + ], + "lines": [ + { + "bbox": [ + 252, + 114, + 480, + 156 + ], + "spans": [ + { + "bbox": [ + 252, + 114, + 480, + 156 + ], + "type": "image", + "image_path": "e52a0eacb7fb6b059e97b0354127e1b1ed5780d7d8e706a927ef443b79522dee.jpg" + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 313, + 157, + 436, + 167 + ], + "lines": [ + { + "bbox": [ + 313, + 157, + 436, + 167 + ], + "spans": [ + { + "bbox": [ + 313, + 157, + 436, + 167 + ], + "type": "text", + "content": "(b) Our proposed PPAD framework." + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 130, + 177, + 482, + 276 + ], + "lines": [ + { + "bbox": [ + 130, + 177, + 482, + 276 + ], + "spans": [ + { + "bbox": [ + 130, + 177, + 482, + 276 + ], + "type": "text", + "content": "Fig. 1: A high-level illustration of our proposed PPAD framework. The agent (in blue) intends to drive straight, while the ego (in red) plans to change lanes. Fig. 1(a) presents the typical one-shot method that might result in invalid motion plans and lead to an accident because of a lack of in-depth interactions. Fig. 1(b) demonstrates the gaming process between the ego and the agent under the PPAD architecture. During the prediction process, the agent executes an assertive plan by accelerating to stop the ego from blocking its route. The planning process of the ego plans trajectory based on the previous prediction process of the agent. The ego decelerates to avoid a potential accident and then changes lanes to achieve its driving goal." + } + ] + } + ], + "index": 6, + "angle": 0, + "type": "image_caption" + } + ], + "index": 4 + }, + { + "bbox": [ + 130, + 302, + 480, + 326 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 302, + 480, + 326 + ], + "spans": [ + { + "bbox": [ + 130, + 302, + 480, + 326 + ], + "type": "text", + "content": "the downstream task is highly related to the upper stream module, which makes it very difficult to construct a unified data-driven infrastructure." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 130, + 327, + 482, + 435 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 327, + 482, + 435 + ], + "spans": [ + { + "bbox": [ + 130, + 327, + 482, + 435 + ], + "type": "text", + "content": "Recently, end-to-end autonomous driving has gained popularity due to its simplicity. Two main lines are proposed based on the learning architecture. The first kind of method [10] takes the raw sensor data as input and directly outputs the planning trajectories or control command without any view transformation as intermediate representations for scene understanding. The other kinds of approaches [19,23] are built upon BEV representation and fully utilize the queries to generate the intermediate outputs as guidance for producing the planning results. One of the most significant advantages lies in the interpretability. In this work, we follow the design of the second kind of work." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 130, + 437, + 482, + 605 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 437, + 482, + 605 + ], + "spans": [ + { + "bbox": [ + 130, + 437, + 482, + 605 + ], + "type": "text", + "content": "VAD [23] and UniAD [19] are typical one-shot motion planning methods, which only consider a single-step interaction between agents, ego-agent and surrounding environment (e.g., map elements). Think Twice [22] makes it a two-stage framework to enhance the gaming or interaction procedure. QCNet [56] and GameFormer [21] also recurrently model the trajectory prediction task. Given motion planning is a computational problem that finds a sequence of valid trajectories, often based on surrounding agents' forecasting, environmental understanding, and historical and future contexts. It can also be viewed as a game in which agents continuously plan their next move according to other agents' intentions and the encountering environment, further achieving their ultimate goals through incremental actions. To model these dynamic interactions of prediction and planning in end-to-end autonomous driving, it is crucial to consider the possible variance of predicted trajectories through multi-step modeling for planning feasible trajectories." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 130, + 605, + 482, + 666 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 605, + 482, + 666 + ], + "spans": [ + { + "bbox": [ + 130, + 605, + 482, + 666 + ], + "type": "text", + "content": "Inspired by VAD [23], we aim to introduce the step-by-step Prediction-Planning into a learning-based framework. Intuitively, the prediction and planning modules can be modeled as a motion forecasting task, which predicts future waypoints by the given historical information. The results of prediction and planning modules at each time step are highly dependent on each other. Therefore," + } + ] + } + ], + "index": 10 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "spans": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "type": "text", + "content": "2" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 166, + 91, + 223, + 100 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 166, + 91, + 223, + 100 + ], + "spans": [ + { + "bbox": [ + 166, + 91, + 223, + 100 + ], + "type": "text", + "content": "Z. Chen et al." + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 1 + }, + { + "para_blocks": [ + { + "bbox": [ + 130, + 116, + 482, + 271 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 116, + 482, + 271 + ], + "spans": [ + { + "bbox": [ + 130, + 116, + 482, + 271 + ], + "type": "text", + "content": "we need to consider the agent-agent and agent-environment interactions iteratively and bidirectionally to maximize the expectation of agents' prediction under the given observation of the other agents. We propose our PPAD to plan the ego agent's future trajectories step-by-step to model the timestep-wise bidirectional interaction or gaming in a vectorized learning framework as shown in Fig. 1. PPAD consists of the prediction and planning process. For each motion forecasting step, 1) Prediction process generates current step motion states by cross-attention and self-attention among agents and environment based on previous motion states to model the fine-grained bidirectional interactions. We take ego-agent-environment-BEV interaction into account to propagate features among all the traffic participants. 2) Planning process predicts the current step motion trajectories based on the expectation process. Our contributions are summarized as follows:" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 138, + 281, + 479, + 422 + ], + "type": "list", + "angle": 0, + "index": 6, + "blocks": [ + { + "bbox": [ + 138, + 281, + 479, + 350 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 281, + 479, + 350 + ], + "spans": [ + { + "bbox": [ + 138, + 281, + 479, + 350 + ], + "type": "text", + "content": "- We propose PPAD that optimizes ego-agent-environment interactions in an iterative prediction-planning manner. Iterative optimization could model the interactions and gaming better and more naturally in a planning task. The prediction process deals with more fine-grained and complex future uncertainties for multi-agent context learning, while the planning process plans a one-step future trajectory for the ego vehicle." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 138, + 353, + 479, + 386 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 353, + 479, + 386 + ], + "spans": [ + { + "bbox": [ + 138, + 353, + 479, + 386 + ], + "type": "text", + "content": "- We model fine-grained interactions among the ego vehicle, agents, environment, and BEV features map, step-by-step with hierarchical dynamic key objects attention emphasizing on the spatial locality." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 139, + 388, + 479, + 422 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 139, + 388, + 479, + 422 + ], + "spans": [ + { + "bbox": [ + 139, + 388, + 479, + 422 + ], + "type": "text", + "content": "- The experiments conducted on the nuScenes [3] and Argoverse [6,44] datasets have demonstrated the effectiveness of our approach over state-of-the-art approaches." + } + ] + } + ], + "index": 5 + } + ], + "sub_type": "text" + }, + { + "bbox": [ + 132, + 441, + 237, + 454 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 441, + 237, + 454 + ], + "spans": [ + { + "bbox": [ + 132, + 441, + 237, + 454 + ], + "type": "text", + "content": "2 Related Work" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 132, + 467, + 328, + 479 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 467, + 328, + 479 + ], + "spans": [ + { + "bbox": [ + 132, + 467, + 328, + 479 + ], + "type": "text", + "content": "2.1 Multi-stage Autonomous Driving" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 130, + 486, + 482, + 666 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 486, + 482, + 666 + ], + "spans": [ + { + "bbox": [ + 130, + 486, + 482, + 666 + ], + "type": "text", + "content": "Most autonomous driving systems are built upon the multi-stage design philosophy, which commonly consists of localization, perception, and planning. The perception module has been well studied recently due to the emergence of deep learning. Camera-based [20, 28, 29], Lidar-based [26, 48-50, 55] or fused-based [36, 46] approaches are proposed to fully exploit the potential of raw sensor data in order to produce accurate 3D objects prediction, semantic segmentation or tracking velocity. Prediction takes the outputs of the perception module to generate the future waypoints for the ego and the agents. Current approaches [1, 5, 11-13, 16, 30, 32, 40, 42, 48, 52, 56, 57] explore different representations to encode surrounding environment (map information) and agent interactions to predict final trajectories by regression or postprocessing sampling strategies. Some other works [4, 37] propose a joint perception and prediction framework, which aggregates historical information to generate tracklets with future trajectories. This unified learning framework could help address the nondifferential process and alleviate the unstable perception problem, compared" + } + ] + } + ], + "index": 9 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 420, + 91, + 447, + 100 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 420, + 91, + 447, + 100 + ], + "spans": [ + { + "bbox": [ + 420, + 91, + 447, + 100 + ], + "type": "text", + "content": "PPAD" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 474, + 91, + 481, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 474, + 91, + 481, + 100 + ], + "spans": [ + { + "bbox": [ + 474, + 91, + 481, + 100 + ], + "type": "text", + "content": "3" + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 2 + }, + { + "para_blocks": [ + { + "bbox": [ + 132, + 116, + 480, + 152 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 116, + 480, + 152 + ], + "spans": [ + { + "bbox": [ + 132, + 116, + 480, + 152 + ], + "type": "text", + "content": "with previous works. Based on the perception and prediction results, planning module [1,7,12,39] plan its future behavior by cost-map optimization or learning-based approaches." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 132, + 168, + 331, + 181 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 168, + 331, + 181 + ], + "spans": [ + { + "bbox": [ + 132, + 168, + 331, + 181 + ], + "type": "text", + "content": "2.2 End-To-End Autonomous Driving" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 132, + 187, + 481, + 355 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 187, + 481, + 355 + ], + "spans": [ + { + "bbox": [ + 132, + 187, + 481, + 355 + ], + "type": "text", + "content": "Recently, more and more works have focused on end-to-end autonomous driving due to its merits in reducing internal accumulative errors and direct yet simple learning objectives. Typical methods [2,45] take the raw visual inputs to regress the final control command or trajectory points without any view transformations. To embrace reinforcement learning, a series of following works [8, 45, 54] utilize policy-based or valued-based approaches to improve driving behavior. With the popularity of BEV representation, more advanced architectures [9, 22] are introduced to attach more interpretability and help deal with the complex interactions in the driving scenarios. Another merit of BEV representation lies in its simplicity in fusing multi-modality sensors. Moreover, the modularized approaches [14, 18, 19, 23] decouple the end-to-end learning-based methods into several submodules or subtasks, while in a multi-task learning manner. The unified design could propagate and share learning context between modules through queries or feature maps." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 132, + 371, + 264, + 384 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 371, + 264, + 384 + ], + "spans": [ + { + "bbox": [ + 132, + 371, + 264, + 384 + ], + "type": "text", + "content": "2.3 BEV Representation" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 132, + 390, + 481, + 510 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 390, + 481, + 510 + ], + "spans": [ + { + "bbox": [ + 132, + 390, + 481, + 510 + ], + "type": "text", + "content": "BEV representation has gained significant prominence in the field of autonomous driving systems due to its inherent distortion-free characteristics and its simplicity in facilitating multi-sensor fusion. There are two main lines for BEV representation, including bottom-up and top-down ways. LSS [38] stands out as a bottom-up pioneering work that explores depth distribution for the 3D space frustum sampling to form BEV representation. Works [20, 28] optimize the pipeline through better depth estimation or lightweight sampling design. BEVFormer [29] and its following works [20, 33-35, 43, 47] adopt the top-down architecture, which uses a deformable transformer for the view transformations without depth supervision." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 132, + 527, + 202, + 540 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 527, + 202, + 540 + ], + "spans": [ + { + "bbox": [ + 132, + 527, + 202, + 540 + ], + "type": "text", + "content": "3 Method" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 132, + 551, + 268, + 562 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 551, + 268, + 562 + ], + "spans": [ + { + "bbox": [ + 132, + 551, + 268, + 562 + ], + "type": "text", + "content": "3.1 Framework Overview" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 132, + 570, + 481, + 665 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 570, + 481, + 665 + ], + "spans": [ + { + "bbox": [ + 132, + 570, + 481, + 665 + ], + "type": "text", + "content": "We present the overall framework, PPAD, in Fig. 2, which comprises the principal modules of the Perception Transformer and our proposed Iterative Prediction-Planning Module. The Perception Transformer encodes the scene contexts into the BEV features map and further decodes as vectorized agents and map representations. The Iterative Prediction-Planning module consists of Prediction and Planning process in general. It dissects the dynamic interactions between the ego vehicle and the agents along the temporal dimension. Eventually, it predicts the motions of the agents and plans the future trajectory for the ego vehicle." + } + ] + } + ], + "index": 9 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "spans": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "type": "text", + "content": "4" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 166, + 91, + 223, + 100 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 166, + 91, + 223, + 100 + ], + "spans": [ + { + "bbox": [ + 166, + 91, + 223, + 100 + ], + "type": "text", + "content": "Z. Chen et al." + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 3 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 133, + 113, + 482, + 308 + ], + "blocks": [ + { + "bbox": [ + 133, + 113, + 482, + 308 + ], + "lines": [ + { + "bbox": [ + 133, + 113, + 482, + 308 + ], + "spans": [ + { + "bbox": [ + 133, + 113, + 482, + 308 + ], + "type": "image", + "image_path": "8e97feb90329d1f49e53f7e11f45c68f3e7f7caa7872baacdc6c73c24e31b0d1.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 130, + 315, + 482, + 449 + ], + "lines": [ + { + "bbox": [ + 130, + 315, + 482, + 449 + ], + "spans": [ + { + "bbox": [ + 130, + 315, + 482, + 449 + ], + "type": "text", + "content": "Fig. 2: Overall architecture of our proposed self-driving framework, PPAD. It consists of the Perception Transformer and the Iterative Prediction-Planning Module. The Perception Transformer encodes scene contexts into agent queries, map queries, and BEV queries. Then, the Prediction-Planning Module interleaves the processes of the agent motion prediction and the ego planning for " + }, + { + "bbox": [ + 130, + 315, + 482, + 449 + ], + "type": "inline_equation", + "content": "N" + }, + { + "bbox": [ + 130, + 315, + 482, + 449 + ], + "type": "text", + "content": " times. Throughout the iterative Prediction and Planning processes, in-depth interactions are conducted among the ego, agents, map elements, and BEV features. In the Prediction process, the agent initially intends to go straight and is unaware of the potential motion of the ego. After interacting with the ego, map elements, and BEV features, the agent plans to be assertive and proceeds to accelerate. In the following Planning process, the ego knows the agent will accelerate through interacting with the updated agent query. It eventually plans to decelerate first and then conduct the lane change for safety reasons." + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_caption" + } + ], + "index": 2 + }, + { + "bbox": [ + 130, + 472, + 480, + 496 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 472, + 480, + 496 + ], + "spans": [ + { + "bbox": [ + 130, + 472, + 480, + 496 + ], + "type": "text", + "content": "Image Features Module uses a shared image backbone network (e.g., ResNet [15]) to extract image features for separate camera views." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 130, + 497, + 482, + 592 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 497, + 482, + 592 + ], + "spans": [ + { + "bbox": [ + 130, + 497, + 482, + 592 + ], + "type": "text", + "content": "BEV Features Module transform the semantic features from the multi-view cameras into a united bird's-eye-view. Specifically, we inherit the encoder from BEVFormer [29, 47] to construct the BEV features. The grid-shape learnable BEV queries " + }, + { + "bbox": [ + 130, + 497, + 482, + 592 + ], + "type": "inline_equation", + "content": "\\mathbf{B} \\in \\mathbb{R}^{H \\times W \\times C}" + }, + { + "bbox": [ + 130, + 497, + 482, + 592 + ], + "type": "text", + "content": " are randomly initialized and learned to interact with the multi-view image features through deformable attention [58] to conduct spatial modeling. Temporal modeling is conducted in a recurrent manner, which applies the deformable attention between the current frame's BEV queries and the one from the previous time step." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 130, + 594, + 482, + 666 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 594, + 482, + 666 + ], + "spans": [ + { + "bbox": [ + 130, + 594, + 482, + 666 + ], + "type": "text", + "content": "Vectorized Features Module Inspired by the VAD [23] paradigm, we also encode the scene contexts into vectorized representations through a detection decoder head [29, 58] and a map element decoding head [31], resulting in " + }, + { + "bbox": [ + 130, + 594, + 482, + 666 + ], + "type": "inline_equation", + "content": "N_{A}" + }, + { + "bbox": [ + 130, + 594, + 482, + 666 + ], + "type": "text", + "content": " of learned agent queries " + }, + { + "bbox": [ + 130, + 594, + 482, + 666 + ], + "type": "inline_equation", + "content": "\\mathbf{A} \\in \\mathbb{R}^{N_A \\times C}" + }, + { + "bbox": [ + 130, + 594, + 482, + 666 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 130, + 594, + 482, + 666 + ], + "type": "inline_equation", + "content": "N_{M}" + }, + { + "bbox": [ + 130, + 594, + 482, + 666 + ], + "type": "text", + "content": " of learned map queries " + }, + { + "bbox": [ + 130, + 594, + 482, + 666 + ], + "type": "inline_equation", + "content": "\\mathbf{M} \\in \\mathbb{R}^{N_M \\times C}" + }, + { + "bbox": [ + 130, + 594, + 482, + 666 + ], + "type": "text", + "content": ". Separate MLP-based decoders will be attached to produce side output, which takes the learned queries as inputs and predicts with the agent attributes (lo" + } + ] + } + ], + "index": 6 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 420, + 91, + 447, + 100 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 420, + 91, + 447, + 100 + ], + "spans": [ + { + "bbox": [ + 420, + 91, + 447, + 100 + ], + "type": "text", + "content": "PPAD" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 474, + 92, + 480, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 474, + 92, + 480, + 100 + ], + "spans": [ + { + "bbox": [ + 474, + 92, + 480, + 100 + ], + "type": "text", + "content": "5" + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 4 + }, + { + "para_blocks": [ + { + "bbox": [ + 130, + 116, + 482, + 187 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 116, + 482, + 187 + ], + "spans": [ + { + "bbox": [ + 130, + 116, + 482, + 187 + ], + "type": "text", + "content": "cations, dimensions, classes, etc.) or map attributes (classes and map vectors described by points). Additionally, the agent queries will be combined with the learnable motion embeddings for modeling the diverse motions of the agents. The agents with motions are represented as " + }, + { + "bbox": [ + 130, + 116, + 482, + 187 + ], + "type": "inline_equation", + "content": "\\mathbf{A} \\in \\mathbb{R}^{N_A \\times N_A^{mot} \\times C}" + }, + { + "bbox": [ + 130, + 116, + 482, + 187 + ], + "type": "text", + "content": ". Similarly, the ego vehicle is modeled with three modes, representing the high-level driving commands of going straight, turn left, and turn right, in the form of " + }, + { + "bbox": [ + 130, + 116, + 482, + 187 + ], + "type": "inline_equation", + "content": "\\mathbf{E} \\in \\mathbb{R}^{N_E \\times N_E^{mot} \\times C}" + }, + { + "bbox": [ + 130, + 116, + 482, + 187 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 130, + 188, + 482, + 308 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 188, + 482, + 308 + ], + "spans": [ + { + "bbox": [ + 130, + 188, + 482, + 308 + ], + "type": "text", + "content": "Iterative Prediction-Planning Module predicts the future trajectories for the ego vehicle and the agents in an interleaved fashion. Different from the traditional practice that predicts all the trajectories in one go, our PPAD framework articulates each step of motion planning by iterating the agent motion prediction and the ego planning processes. Thanks to the PPAD framework, we can conduct in-depth design to enforce key objects interactions (in Sec. 3.3) on scene contexts in a coarse-to-fine manner. We further improve the driving performance for the ego vehicle by taking the noisy trajectory as each step prediction and training the PPAD framework to reconstruct its original position at the following time step (in Sec. 3.4)." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 131, + 325, + 407, + 337 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 131, + 325, + 407, + 337 + ], + "spans": [ + { + "bbox": [ + 131, + 325, + 407, + 337 + ], + "type": "text", + "content": "3.2 Iterative Interactions of Prediction and Planning" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 130, + 343, + 482, + 451 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 343, + 482, + 451 + ], + "spans": [ + { + "bbox": [ + 130, + 343, + 482, + 451 + ], + "type": "text", + "content": "In the real world, the driving traffic changes constantly. Drivers plan and execute their decisions by ceaselessly reasoning the relationships among traffic participants in the scene. The planning task requires the self-driving system to have a good understanding of the scene and be capable of resolving the spatial-temporal causal factors. Therefore, we innovate the PPAD to dissect the planning task into multi-steps of the agent prediction and the ego planning processes and eventually promote consensuses among the ego's and the agents' future trajectories. The PPAD framework embodies the traffic interactions as gaming along space-time, producing a more accurate planning trajectory for the ego vehicle." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 130, + 452, + 482, + 500 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 452, + 482, + 500 + ], + "spans": [ + { + "bbox": [ + 130, + 452, + 482, + 500 + ], + "type": "text", + "content": "Specifically, the ego and agents inherit the same philosophy of alternatively optimizing their motion behavior based on each other's motion forecasting at each future time step. In the following section, we will demonstrate the agent prediction process and elaborate on the details of the ego planning process." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 130, + 516, + 482, + 624 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 516, + 482, + 624 + ], + "spans": [ + { + "bbox": [ + 130, + 516, + 482, + 624 + ], + "type": "text", + "content": "Prediction Process As illustrated in Fig 2, the agent predicts its subsequent step motion during the Prediction process, conditioned on the output of the ego vehicle's outcome from the previous Planning process. Specifically, the initial state of the agent query comprises its driving intention. It will then interact with the ego query updated from the previous planning process, which indicates the latest driving plan of the ego vehicle. After that, it will interact with map elements to choose the driving paths. At last, it gathers detailed geometric information by interacting with the BEV features and comes up with its precise next-step movement." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 131, + 641, + 482, + 667 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 131, + 641, + 482, + 667 + ], + "spans": [ + { + "bbox": [ + 131, + 641, + 482, + 667 + ], + "type": "text", + "content": "Planning Process We consider the period of history with " + }, + { + "bbox": [ + 131, + 641, + 482, + 667 + ], + "type": "inline_equation", + "content": "T_{obs}" + }, + { + "bbox": [ + 131, + 641, + 482, + 667 + ], + "type": "text", + "content": " steps and the future with " + }, + { + "bbox": [ + 131, + 641, + 482, + 667 + ], + "type": "inline_equation", + "content": "T_{fut}" + }, + { + "bbox": [ + 131, + 641, + 482, + 667 + ], + "type": "text", + "content": " steps. The future trajectory of ego is denoted as " + }, + { + "bbox": [ + 131, + 641, + 482, + 667 + ], + "type": "inline_equation", + "content": "\\{p_E^t\\}^{t \\in T_{fut}}" + }, + { + "bbox": [ + 131, + 641, + 482, + 667 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 8 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "spans": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "type": "text", + "content": "6" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 166, + 91, + 223, + 100 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 166, + 91, + 223, + 100 + ], + "spans": [ + { + "bbox": [ + 166, + 91, + 223, + 100 + ], + "type": "text", + "content": "Z. Chen et al." + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 5 + }, + { + "para_blocks": [ + { + "bbox": [ + 130, + 115, + 479, + 175 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 115, + 479, + 175 + ], + "spans": [ + { + "bbox": [ + 130, + 115, + 479, + 175 + ], + "type": "text", + "content": "For each agent " + }, + { + "bbox": [ + 130, + 115, + 479, + 175 + ], + "type": "inline_equation", + "content": "a \\in \\mathbf{A}" + }, + { + "bbox": [ + 130, + 115, + 479, + 175 + ], + "type": "text", + "content": ", the trajectory is represented as " + }, + { + "bbox": [ + 130, + 115, + 479, + 175 + ], + "type": "inline_equation", + "content": "\\{p_a^t\\}^{t \\in T_{fut}}" + }, + { + "bbox": [ + 130, + 115, + 479, + 175 + ], + "type": "text", + "content": ". The positions of the detected map elements are denoted as " + }, + { + "bbox": [ + 130, + 115, + 479, + 175 + ], + "type": "inline_equation", + "content": "\\{p_m\\}_{m \\in \\mathbf{M}}" + }, + { + "bbox": [ + 130, + 115, + 479, + 175 + ], + "type": "text", + "content": ". We define the operator of " + }, + { + "bbox": [ + 130, + 115, + 479, + 175 + ], + "type": "inline_equation", + "content": "\\mathcal{M}(p_E^t, p_{\\mathbf{A}} / p_{\\mathbf{M}}, s)" + }, + { + "bbox": [ + 130, + 115, + 479, + 175 + ], + "type": "text", + "content": " to mask out the agents or map elements that are beyond a distance of " + }, + { + "bbox": [ + 130, + 115, + 479, + 175 + ], + "type": "inline_equation", + "content": "s" + }, + { + "bbox": [ + 130, + 115, + 479, + 175 + ], + "type": "text", + "content": " towards the ego located at " + }, + { + "bbox": [ + 130, + 115, + 479, + 175 + ], + "type": "inline_equation", + "content": "p_E^t" + }, + { + "bbox": [ + 130, + 115, + 479, + 175 + ], + "type": "text", + "content": " to conduct key objects attention, and we will discuss the algorithm details in Sec. 3.3." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 130, + 176, + 480, + 271 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 176, + 480, + 271 + ], + "spans": [ + { + "bbox": [ + 130, + 176, + 480, + 271 + ], + "type": "text", + "content": "Agent Interaction The resulting agents' motions from the prediction process are located at " + }, + { + "bbox": [ + 130, + 176, + 480, + 271 + ], + "type": "inline_equation", + "content": "\\{p_a^{t + 1}\\}_{a\\in \\mathbf{A}}" + }, + { + "bbox": [ + 130, + 176, + 480, + 271 + ], + "type": "text", + "content": ", comprising the motion states up to the time step of " + }, + { + "bbox": [ + 130, + 176, + 480, + 271 + ], + "type": "inline_equation", + "content": "t + 1" + }, + { + "bbox": [ + 130, + 176, + 480, + 271 + ], + "type": "text", + "content": ". Moving the ego from " + }, + { + "bbox": [ + 130, + 176, + 480, + 271 + ], + "type": "inline_equation", + "content": "p_E^t" + }, + { + "bbox": [ + 130, + 176, + 480, + 271 + ], + "type": "text", + "content": " to the future one-step of " + }, + { + "bbox": [ + 130, + 176, + 480, + 271 + ], + "type": "inline_equation", + "content": "p_E^{t + 1}" + }, + { + "bbox": [ + 130, + 176, + 480, + 271 + ], + "type": "text", + "content": ", the ego vehicle should consider the agents' traffic globally and locally. From a more global perspective, the agents at a larger range provide more extensive information on traffic flow, which is essential in long-term trajectory planning. Regarding the spatial locality, the nearby agents are recognized as the key agents, which are supposed to be vitally related to the ego's driving decision." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 130, + 272, + 480, + 366 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 272, + 480, + 366 + ], + "spans": [ + { + "bbox": [ + 130, + 272, + 480, + 366 + ], + "type": "text", + "content": "Therefore, we propose to conduct a hierarchical interaction with the agents through the attention mechanism to learn coarse-to-fine traffic context features for the ego. Centered at the ego's space-time position " + }, + { + "bbox": [ + 130, + 272, + 480, + 366 + ], + "type": "inline_equation", + "content": "p_E^t" + }, + { + "bbox": [ + 130, + 272, + 480, + 366 + ], + "type": "text", + "content": ", we initially define a distance set of " + }, + { + "bbox": [ + 130, + 272, + 480, + 366 + ], + "type": "inline_equation", + "content": "\\mathbf{S}" + }, + { + "bbox": [ + 130, + 272, + 480, + 366 + ], + "type": "text", + "content": " of " + }, + { + "bbox": [ + 130, + 272, + 480, + 366 + ], + "type": "inline_equation", + "content": "\\{+\\infty m, 15m, 7.5m\\}" + }, + { + "bbox": [ + 130, + 272, + 480, + 366 + ], + "type": "text", + "content": " which covers the coarse-to-fine perception ranges. We formed the multi-scale agent sets by applying " + }, + { + "bbox": [ + 130, + 272, + 480, + 366 + ], + "type": "inline_equation", + "content": "\\mathcal{M}(p_E^t, p_{\\mathbf{A}}, s)" + }, + { + "bbox": [ + 130, + 272, + 480, + 366 + ], + "type": "text", + "content": " with different ranges. Then, the ego query interacts with the agents hierarchically through multi-head cross-attention, MHCA. We take the sum of the learned hierarchical attention results as the final values:" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 194, + 373, + 480, + 397 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 194, + 373, + 480, + 397 + ], + "spans": [ + { + "bbox": [ + 194, + 373, + 480, + 397 + ], + "type": "interline_equation", + "content": "\\mathbf {E} ^ {k} = \\sum_ {s \\in \\mathbf {S}} \\operatorname {M H C A} (\\mathbf {E}, \\mathbf {A} ^ {k}, \\mathcal {M} \\left(p _ {\\mathbf {E}} ^ {t}, p _ {\\mathbf {A} ^ {k}} ^ {t + 1}, s\\right)), k \\in [ 1, N _ {A} ^ {\\text {m o t}} ], \\tag {1}", + "image_path": "ffd3d05661b71f9f01edac1340a2673c4cdb44e79584be77bdb540129cc24e5a.jpg" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 130, + 404, + 479, + 440 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 404, + 479, + 440 + ], + "spans": [ + { + "bbox": [ + 130, + 404, + 479, + 440 + ], + "type": "text", + "content": "where the ego independently queries information from different modes " + }, + { + "bbox": [ + 130, + 404, + 479, + 440 + ], + "type": "inline_equation", + "content": "k \\in N_A^{mot}" + }, + { + "bbox": [ + 130, + 404, + 479, + 440 + ], + "type": "text", + "content": " of agents, and then we stack the results output from different modes. Further, we apply the set operations to condense the features:" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 201, + 447, + 480, + 461 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 201, + 447, + 480, + 461 + ], + "spans": [ + { + "bbox": [ + 201, + 447, + 480, + 461 + ], + "type": "interline_equation", + "content": "\\mathbf {E} ^ {\\prime} = \\operatorname {M A X} \\left(\\left[ \\mathbf {E} ^ {1},.., \\mathbf {E} ^ {N _ {A} ^ {\\text {m o t}}} \\right]\\right) + \\operatorname {M E A N} \\left(\\left[ \\mathbf {E} ^ {1},.., \\mathbf {E} ^ {N _ {A} ^ {\\text {m o t}}} \\right]\\right), \\tag {2}", + "image_path": "ffa1eb7fee54a6a52abdc8a5b568e1efdff33f9ad326bfd984118addbe9fa347.jpg" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 130, + 468, + 479, + 491 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 468, + 479, + 491 + ], + "spans": [ + { + "bbox": [ + 130, + 468, + 479, + 491 + ], + "type": "text", + "content": "where the MAX and MEAN are applied to aggregate features along the agents' mode dimension and output with the updated ego query " + }, + { + "bbox": [ + 130, + 468, + 479, + 491 + ], + "type": "inline_equation", + "content": "\\mathbf{E}'" + }, + { + "bbox": [ + 130, + 468, + 479, + 491 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 130, + 492, + 480, + 552 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 492, + 480, + 552 + ], + "spans": [ + { + "bbox": [ + 130, + 492, + 480, + 552 + ], + "type": "text", + "content": "Map Interaction Existing works [19, 23] tried to summarize all the required map information for planning by simply applying the global-level interaction once. They overlook the complexity of the evolving motion dynamic and overrate that the ego can plan precisely in the longer term by a single interaction with the map information." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 130, + 552, + 480, + 635 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 552, + 480, + 635 + ], + "spans": [ + { + "bbox": [ + 130, + 552, + 480, + 635 + ], + "type": "text", + "content": "With our proposed PPAD framework, we can enrich the ego-map interaction by considering the ego's local road conditions based on its latest position. This results in better identifying the useful map information for each step of planning. The ego query interacts with the map queries in a similar practice as the interaction with the agents. The difference is that the map instances are not movable in the future time steps. The local and global map information can be abstracted into the ego query by MHCA:" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 227, + 643, + 480, + 667 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 227, + 643, + 480, + 667 + ], + "spans": [ + { + "bbox": [ + 227, + 643, + 480, + 667 + ], + "type": "interline_equation", + "content": "\\mathbf {E} ^ {\\prime \\prime} = \\sum_ {s \\in \\mathbf {S}} \\operatorname {M H C A} \\left(\\mathbf {E} ^ {\\prime}, \\mathbf {M}, \\mathcal {M} \\left(p _ {\\mathbf {E}} ^ {t}, p _ {m}, s\\right)\\right), \\tag {3}", + "image_path": "f0814564721f2ef771273c49fb052af0434841cf2ebc6e13ea0cbd6dea2548c7.jpg" + } + ] + } + ], + "index": 11 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 420, + 91, + 447, + 100 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 420, + 91, + 447, + 100 + ], + "spans": [ + { + "bbox": [ + 420, + 91, + 447, + 100 + ], + "type": "text", + "content": "PPAD" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 474, + 92, + 481, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 474, + 92, + 481, + 100 + ], + "spans": [ + { + "bbox": [ + 474, + 92, + 481, + 100 + ], + "type": "text", + "content": "7" + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 6 + }, + { + "para_blocks": [ + { + "bbox": [ + 130, + 116, + 479, + 140 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 116, + 479, + 140 + ], + "spans": [ + { + "bbox": [ + 130, + 116, + 479, + 140 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 130, + 116, + 479, + 140 + ], + "type": "inline_equation", + "content": "\\mathbf{E}''" + }, + { + "bbox": [ + 130, + 116, + 479, + 140 + ], + "type": "text", + "content": " is the ego query updated with the critical map information for the next step of planning." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 130, + 141, + 482, + 322 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 141, + 482, + 322 + ], + "spans": [ + { + "bbox": [ + 130, + 141, + 482, + 322 + ], + "type": "text", + "content": "BEV Interaction BEV is the fundamental representation of the whole system, which is the abstraction of the multi-camera features. Beyond the vectorized representation, there is other non-structural environmental information, including roads and fences. UniAD [19] models these non-structural pieces of stuff with an occupancy grid map. There are several drawbacks in UniAD: 1) occupancy grids consume large memory considering the whole scene's range. 2) UniAD failed to build the explicit interactions with the grid map. Therefore, we propose the BEV interactions that dynamically query the surrounding environment for each possible future step. This query process could help agents understand and learn the effects of their actions. Specifically, after applying the interactions above, the ego vehicle understands the dynamic agents' traffic better and knows its fronting road conditions. Nevertheless, planning a more precise motion requires the ego to comprehend the local detail geometric information. Hence, the ego query further interacts with BEV features to extract low-level geometric information. Specifically, we achieve by the deformable attention [58]:" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 241, + 334, + 480, + 347 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 241, + 334, + 480, + 347 + ], + "spans": [ + { + "bbox": [ + 241, + 334, + 480, + 347 + ], + "type": "interline_equation", + "content": "\\mathbf {E} ^ {\\prime \\prime \\prime} = \\operatorname {D e f o r m A t t n} \\left(\\mathbf {E} ^ {\\prime \\prime}, p _ {\\mathbf {E}} ^ {t}, \\mathbf {B}\\right), \\tag {4}", + "image_path": "ac73f99b9d04270b70c0a87118a68974b4a584ab671cd5608ad6f8cc1998dc19.jpg" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 130, + 358, + 479, + 407 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 358, + 479, + 407 + ], + "spans": [ + { + "bbox": [ + 130, + 358, + 479, + 407 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 130, + 358, + 479, + 407 + ], + "type": "inline_equation", + "content": "p_{\\mathbf{E}}^t" + }, + { + "bbox": [ + 130, + 358, + 479, + 407 + ], + "type": "text", + "content": " is the location of ego at time " + }, + { + "bbox": [ + 130, + 358, + 479, + 407 + ], + "type": "inline_equation", + "content": "t" + }, + { + "bbox": [ + 130, + 358, + 479, + 407 + ], + "type": "text", + "content": ", and it serves as the reference point on the BEV features. The deformable attention DeformAttn applies sparse attention around the reference points " + }, + { + "bbox": [ + 130, + 358, + 479, + 407 + ], + "type": "inline_equation", + "content": "p_{\\mathbf{E}}^t" + }, + { + "bbox": [ + 130, + 358, + 479, + 407 + ], + "type": "text", + "content": " and learns to pick up the low-level geometric information from the BEV for planning." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 130, + 408, + 482, + 492 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 408, + 482, + 492 + ], + "spans": [ + { + "bbox": [ + 130, + 408, + 482, + 492 + ], + "type": "text", + "content": "Motion Planning PPAD follows the same practice as [18, 19, 23], which uses the information of the high-level driving commands: go straight, turn left, and turn right. The concatenated features of " + }, + { + "bbox": [ + 130, + 408, + 482, + 492 + ], + "type": "inline_equation", + "content": "\\mathbf{h}_{\\mathbf{E}} = [\\mathbf{E}', \\mathbf{E}'', \\mathbf{E}''']" + }, + { + "bbox": [ + 130, + 408, + 482, + 492 + ], + "type": "text", + "content": " contain the information of the dynamic agent traffic, map semantics, and the precise environmental geometry. An MLP takes " + }, + { + "bbox": [ + 130, + 408, + 482, + 492 + ], + "type": "inline_equation", + "content": "\\mathbf{h}_{\\mathbf{E}}" + }, + { + "bbox": [ + 130, + 408, + 482, + 492 + ], + "type": "text", + "content": " as input and predicts the future one-step waypoint offset " + }, + { + "bbox": [ + 130, + 408, + 482, + 492 + ], + "type": "inline_equation", + "content": "w_{\\mathbf{E}}^{t+1} = (x, y)" + }, + { + "bbox": [ + 130, + 408, + 482, + 492 + ], + "type": "text", + "content": ". We then update the ego state by applying another MLP on " + }, + { + "bbox": [ + 130, + 408, + 482, + 492 + ], + "type": "inline_equation", + "content": "\\mathbf{h}_{\\mathbf{E}}" + }, + { + "bbox": [ + 130, + 408, + 482, + 492 + ], + "type": "text", + "content": " for the next step of processing." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 131, + 517, + 312, + 530 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 131, + 517, + 312, + 530 + ], + "spans": [ + { + "bbox": [ + 131, + 517, + 312, + 530 + ], + "type": "text", + "content": "3.3 Hierarchical Feature Learning" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 130, + 545, + 482, + 666 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 545, + 482, + 666 + ], + "spans": [ + { + "bbox": [ + 130, + 545, + 482, + 666 + ], + "type": "text", + "content": "Hierarchical structure has a better capability to capture and recognize fine-grained patterns. For the driving scenarios, the driving behavior is based on scene understanding both globally and locally. Driving tends to focus on only a few key objects, which demonstrates the spatial locality or local attention. Therefore, we design hierarchical key objects attention to exploit the coarse-to-fine scene contexts. Specifically, given a set of distance ranges, we first find the key objects (agents or map elements) within the given range. Consequently, we apply dynamic local attention, which only considers the interactions among agents or map elements in the local area. The pseudo code shown in Alg. 1 delineates the implementation of dynamic key objects attention." + } + ] + } + ], + "index": 8 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "spans": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "type": "text", + "content": "8" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 166, + 91, + 223, + 100 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 166, + 91, + 223, + 100 + ], + "spans": [ + { + "bbox": [ + 166, + 91, + 223, + 100 + ], + "type": "text", + "content": "Z. Chen et al." + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 7 + }, + { + "para_blocks": [ + { + "type": "code", + "bbox": [ + 184, + 167, + 427, + 285 + ], + "blocks": [ + { + "bbox": [ + 184, + 138, + 429, + 161 + ], + "lines": [ + { + "bbox": [ + 184, + 138, + 429, + 161 + ], + "spans": [ + { + "bbox": [ + 184, + 138, + 429, + 161 + ], + "type": "text", + "content": "Algorithm 1 Pseudo code of Key objects attention in a PyTorch-like style." + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "code_caption" + }, + { + "bbox": [ + 184, + 167, + 427, + 285 + ], + "lines": [ + { + "bbox": [ + 184, + 167, + 427, + 285 + ], + "spans": [ + { + "bbox": [ + 184, + 167, + 427, + 285 + ], + "type": "text", + "content": ">>> layer = nn.MultiheadAttention(embed_dim, num_head)\n>>> def forward(layer, query, key, q_pos, k_pos, max_dis):\n # q_pos: position of the query with shape [B,Lq,2]\n # k_pos: position of the key with shape [B,Lk,2]\n # layer: attention layer as initialization\n # max_dis: distance threshold\n diff = (q_pos unsqueeze(2) - k_pos unsqueeze(1))\n dist = (diff ** 2).sum(-1).sqrt()\n attn_mask = (dist > max_dis).repeat(num_head,1,1)\n return attention(query, key, attn_mask=attn_mask)" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "code_body" + } + ], + "index": 3, + "sub_type": "code", + "guess_lang": "python" + }, + { + "bbox": [ + 132, + 319, + 315, + 331 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 319, + 315, + 331 + ], + "spans": [ + { + "bbox": [ + 132, + 319, + 315, + 331 + ], + "type": "text", + "content": "3.4 Noisy Trajectory as Prediction" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 130, + 340, + 482, + 472 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 340, + 482, + 472 + ], + "spans": [ + { + "bbox": [ + 130, + 340, + 482, + 472 + ], + "type": "text", + "content": "PPAD interleaves the prediction and the planning processes to plan the ego and agents' trajectories step-by-step. Expert driving knowledge is then enforced into the model through imitation learning. Thanks to our multi-step framework and the inspiration from [27], we introduce the noisy trajectory as the prediction to the PPAD while training. Specifically, we perturb each step of the ground truth ego trajectory by adding noise. The ego is then trained to predict the original next step waypoint offset of the ego regardless of disturbance on its starting noisy positions. The system is learned to predict the accurate waypoint offset by interacting with vectorized instances and the environment even though it starts at an inaccurate position. This strategy brings improvement to the planning performance." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 132, + 491, + 268, + 504 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 491, + 268, + 504 + ], + "spans": [ + { + "bbox": [ + 132, + 491, + 268, + 504 + ], + "type": "text", + "content": "3.5 End-to-End Learning" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 130, + 512, + 480, + 536 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 512, + 480, + 536 + ], + "spans": [ + { + "bbox": [ + 130, + 512, + 480, + 536 + ], + "type": "text", + "content": "Scene Context Loss Similar to VAD [23], we formulate the loss for the agents' motion and map as follows:" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 249, + 547, + 481, + 561 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 249, + 547, + 481, + 561 + ], + "spans": [ + { + "bbox": [ + 249, + 547, + 481, + 561 + ], + "type": "interline_equation", + "content": "\\mathcal {L} _ {S} = \\lambda_ {1} \\mathcal {L} _ {\\text {a g e n t}} + \\lambda_ {2} \\mathcal {L} _ {\\text {m a p}}, \\tag {5}", + "image_path": "28759a74ea8ec452a9aecb3cffb15f1544df4b413befffeccfb561ccd11e5ce6.jpg" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 130, + 570, + 269, + 581 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 570, + 269, + 581 + ], + "spans": [ + { + "bbox": [ + 130, + 570, + 269, + 581 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 130, + 570, + 269, + 581 + ], + "type": "inline_equation", + "content": "\\lambda_{1}" + }, + { + "bbox": [ + 130, + 570, + 269, + 581 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 130, + 570, + 269, + 581 + ], + "type": "inline_equation", + "content": "\\lambda_{2}" + }, + { + "bbox": [ + 130, + 570, + 269, + 581 + ], + "type": "text", + "content": " are set as 1.0." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 130, + 582, + 482, + 666 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 582, + 482, + 666 + ], + "spans": [ + { + "bbox": [ + 130, + 582, + 482, + 666 + ], + "type": "text", + "content": "Constraint Loss Inspired by [23], we propose the confidence-aware collision loss " + }, + { + "bbox": [ + 130, + 582, + 482, + 666 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_{CA - Col}" + }, + { + "bbox": [ + 130, + 582, + 482, + 666 + ], + "type": "text", + "content": ", which considers the potential collision of all the agents' motion modalities instead of only computing the loss on the agents' most confident mode. We multiply the resulting collision loss from each mode with the predicted confidence score. For the trajectories having the potential to collide, it will penalize more when the predicted confidence score is higher. Combined with ego-boundary overstepping " + }, + { + "bbox": [ + 130, + 582, + 482, + 666 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_{bd}" + }, + { + "bbox": [ + 130, + 582, + 482, + 666 + ], + "type": "text", + "content": " and ego-lane directional " + }, + { + "bbox": [ + 130, + 582, + 482, + 666 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_{dir}" + }, + { + "bbox": [ + 130, + 582, + 482, + 666 + ], + "type": "text", + "content": " constraints proposed by [23], the" + } + ] + } + ], + "index": 10 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 420, + 91, + 447, + 100 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 420, + 91, + 447, + 100 + ], + "spans": [ + { + "bbox": [ + 420, + 91, + 447, + 100 + ], + "type": "text", + "content": "PPAD" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 474, + 92, + 481, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 474, + 92, + 481, + 100 + ], + "spans": [ + { + "bbox": [ + 474, + 92, + 481, + 100 + ], + "type": "text", + "content": "9" + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 8 + }, + { + "para_blocks": [ + { + "type": "table", + "bbox": [ + 140, + 114, + 479, + 270 + ], + "blocks": [ + { + "bbox": [ + 140, + 114, + 479, + 270 + ], + "lines": [ + { + "bbox": [ + 140, + 114, + 479, + 270 + ], + "spans": [ + { + "bbox": [ + 140, + 114, + 479, + 270 + ], + "type": "table", + "html": "
MethodL2 (m) ↓Collision (%) ↓Latency (ms)FPS
1s2s3sAvg.1s2s3sAvg.
ST-P3 MetricsST-P3 [18]1.332.112.902.110.230.621.270.71628.31.6
VAD-Tiny [23]0.460.761.120.780.210.350.580.3859.516.8
VAD-Base [23]0.410.701.050.720.070.170.410.22224.34.5
OccNet [41]1.292.132.992.130.210.591.370.72--
FusionAD [51]---1.030.250.130.250.21--
Ours (Progress.)0.310.560.870.580.080.120.380.193852.6
Ours0.250.450.730.480.070.150.360.193852.6
UniAD MetricsNMP† [53]--2.31---1.92---
SA-NMP† [53]--2.05---1.59---
FF† [17]0.551.202.541.430.060.171.070.43--
EO† [24]0.671.362.781.600.040.090.880.33--
UniAD [19]0.480.961.651.030.050.170.710.31555.61.8
VAD-Base [23]0.501.021.691.070.000.300.950.42224.34.5
Ours (Progress.)0.380.831.450.890.020.200.930.383852.6
Ours0.300.691.260.750.030.220.730.333852.6
", + "image_path": "1005f0d49a5dfc56a9d22c59eb8284f5632b74f29a5bc5b84a4b561d1d99705e.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "table_body" + } + ], + "index": 2 + }, + { + "bbox": [ + 130, + 278, + 482, + 389 + ], + "lines": [ + { + "bbox": [ + 130, + 278, + 482, + 389 + ], + "spans": [ + { + "bbox": [ + 130, + 278, + 482, + 389 + ], + "type": "text", + "content": "Table 1: Open-loop planning results on the nuScenes dataset [3]. The results of other methods are obtained from the original paper. We faithfully re-evaluate VAD [23] based on the UniAD [19] metrics. As for our PPAD, we provide two versions of results that utilize different training strategies. Progress means that we follow the progressive training pipeline as proposed in VAD [23], which trained all tasks except the planning task in the first 48 epochs and then finetuned with another 12 epochs for the planning task. The second row for our method trains the whole network for 60 epochs and then finetuned for another 12 epochs incorporating noisy trajectories. The latency of STP3 [18], VAD [23], and ours are measured on one NVIDIA Geforce RTX 3090 GPU, while UniAD is measured on one NVIDIA Tesla A100 GPU." + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "text" + }, + { + "bbox": [ + 132, + 426, + 240, + 436 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 426, + 240, + 436 + ], + "spans": [ + { + "bbox": [ + 132, + 426, + 240, + 436 + ], + "type": "text", + "content": "overall constraint loss is" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 226, + 449, + 481, + 462 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 226, + 449, + 481, + 462 + ], + "spans": [ + { + "bbox": [ + 226, + 449, + 481, + 462 + ], + "type": "interline_equation", + "content": "\\mathcal {L} _ {C} = \\lambda_ {3} \\mathcal {L} _ {C A - C o l} + \\lambda_ {4} \\mathcal {L} _ {b d} + \\lambda_ {5} \\mathcal {L} _ {d i r}, \\tag {6}", + "image_path": "38d4df5f5feb5f701590b739126eedf1a703391fe8f1de67d9f59d125d74b5cc.jpg" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 130, + 474, + 360, + 486 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 474, + 360, + 486 + ], + "spans": [ + { + "bbox": [ + 130, + 474, + 360, + 486 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 130, + 474, + 360, + 486 + ], + "type": "inline_equation", + "content": "\\lambda_3" + }, + { + "bbox": [ + 130, + 474, + 360, + 486 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 130, + 474, + 360, + 486 + ], + "type": "inline_equation", + "content": "\\lambda_4" + }, + { + "bbox": [ + 130, + 474, + 360, + 486 + ], + "type": "text", + "content": " are set as 1.0, and " + }, + { + "bbox": [ + 130, + 474, + 360, + 486 + ], + "type": "inline_equation", + "content": "\\lambda_5" + }, + { + "bbox": [ + 130, + 474, + 360, + 486 + ], + "type": "text", + "content": " is set as 0.5." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 130, + 487, + 481, + 512 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 487, + 481, + 512 + ], + "spans": [ + { + "bbox": [ + 130, + 487, + 481, + 512 + ], + "type": "text", + "content": "Planning Loss We conduct " + }, + { + "bbox": [ + 130, + 487, + 481, + 512 + ], + "type": "inline_equation", + "content": "L_{1}" + }, + { + "bbox": [ + 130, + 487, + 481, + 512 + ], + "type": "text", + "content": " loss between each step of the ego's prediction " + }, + { + "bbox": [ + 130, + 487, + 481, + 512 + ], + "type": "inline_equation", + "content": "w_{\\mathbf{E}}^{t}" + }, + { + "bbox": [ + 130, + 487, + 481, + 512 + ], + "type": "text", + "content": " and the ground truth's waypoint offset " + }, + { + "bbox": [ + 130, + 487, + 481, + 512 + ], + "type": "inline_equation", + "content": "\\tilde{w}_{\\mathbf{E}}^{t}" + }, + { + "bbox": [ + 130, + 487, + 481, + 512 + ], + "type": "text", + "content": " along the future time horizon:" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 236, + 523, + 481, + 556 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 236, + 523, + 481, + 556 + ], + "spans": [ + { + "bbox": [ + 236, + 523, + 481, + 556 + ], + "type": "interline_equation", + "content": "\\mathcal {L} _ {\\text {P l a n}} = \\frac {1}{T _ {f u t}} \\sum_ {t = 1} ^ {T _ {f u t}} \\| w _ {\\mathbf {E}} ^ {t} - \\tilde {w} _ {\\mathbf {E}} ^ {t} \\| _ {1}. \\tag {7}", + "image_path": "3227a2a8d5bd9d30c46a4513b999495d4ee332ce9136e83ac9ac62f9cce4bf0d.jpg" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 130, + 568, + 482, + 617 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 568, + 482, + 617 + ], + "spans": [ + { + "bbox": [ + 130, + 568, + 482, + 617 + ], + "type": "text", + "content": "The overall end-to-end trainable loss function is formed by the sum of the perception loss, constraint loss, and planning loss. The same constraint losses " + }, + { + "bbox": [ + 130, + 568, + 482, + 617 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_C^{noisy}" + }, + { + "bbox": [ + 130, + 568, + 482, + 617 + ], + "type": "text", + "content": " and planning losses " + }, + { + "bbox": [ + 130, + 568, + 482, + 617 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_{plan}^{noisy}" + }, + { + "bbox": [ + 130, + 568, + 482, + 617 + ], + "type": "text", + "content": " will be applied to the predictions taking the noisy trajectories as input:" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 199, + 628, + 481, + 643 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 199, + 628, + 481, + 643 + ], + "spans": [ + { + "bbox": [ + 199, + 628, + 481, + 643 + ], + "type": "interline_equation", + "content": "\\mathcal {L} = \\mathcal {L} _ {S} + \\zeta_ {1} \\left(\\mathcal {L} _ {C} + \\mathcal {L} _ {\\text {P l a n}}\\right) + \\zeta_ {2} \\left(\\mathcal {L} _ {C} ^ {\\text {n o i s y}} + \\mathcal {L} _ {\\text {P l a n}} ^ {\\text {n o i s y}}\\right), \\tag {8}", + "image_path": "6be6d26131f743241d71debe9c49d2c00c2e5932c30e6d7aca0a7779baf065e0.jpg" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 130, + 653, + 316, + 666 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 653, + 316, + 666 + ], + "spans": [ + { + "bbox": [ + 130, + 653, + 316, + 666 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 130, + 653, + 316, + 666 + ], + "type": "inline_equation", + "content": "\\zeta_{1}" + }, + { + "bbox": [ + 130, + 653, + 316, + 666 + ], + "type": "text", + "content": " is set as 0.6 and " + }, + { + "bbox": [ + 130, + 653, + 316, + 666 + ], + "type": "inline_equation", + "content": "\\zeta_{2}" + }, + { + "bbox": [ + 130, + 653, + 316, + 666 + ], + "type": "text", + "content": " is set as 0.4." + } + ] + } + ], + "index": 11 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "spans": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "type": "text", + "content": "10" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 166, + 91, + 223, + 100 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 166, + 91, + 223, + 100 + ], + "spans": [ + { + "bbox": [ + 166, + 91, + 223, + 100 + ], + "type": "text", + "content": "Z. Chen et al." + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 9 + }, + { + "para_blocks": [ + { + "type": "table", + "bbox": [ + 201, + 114, + 418, + 159 + ], + "blocks": [ + { + "bbox": [ + 201, + 114, + 418, + 159 + ], + "lines": [ + { + "bbox": [ + 201, + 114, + 418, + 159 + ], + "spans": [ + { + "bbox": [ + 201, + 114, + 418, + 159 + ], + "type": "table", + "html": "
MethodL2 (m) ↓Collision (%) ↓
1s2s3sAvg.1s2s3sAvg.
VAD [23]1.041.792.601.810.611.262.251.37
Ours0.731.301.981.340.510.851.310.89
", + "image_path": "33f2e788d68a74f385b10807f27f3262b7b1bf13e9e71d5a0ff40765f4be463c.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "table_body" + } + ], + "index": 2 + }, + { + "type": "table", + "bbox": [ + 176, + 211, + 444, + 267 + ], + "blocks": [ + { + "bbox": [ + 132, + 167, + 481, + 190 + ], + "lines": [ + { + "bbox": [ + 132, + 167, + 481, + 190 + ], + "spans": [ + { + "bbox": [ + 132, + 167, + 481, + 190 + ], + "type": "text", + "content": "Table 2: Open-loop planning results on the Argoverse dataset [3]. The results are evaluated under the ST-P3 [18] and VAD [23] metrics." + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 176, + 211, + 444, + 267 + ], + "lines": [ + { + "bbox": [ + 176, + 211, + 444, + 267 + ], + "spans": [ + { + "bbox": [ + 176, + 211, + 444, + 267 + ], + "type": "table", + "html": "
MethodDetectionMapMotion Forecasting
NDS ↑mAP↑mAP↑minADE (m)↓minFDE (m) ↓MR↓
VAD [23]0.4590.3290.4760.6780.8820.08
UniAD [19]0.4990.382-0.7081.020.13
Ours0.4650.3320.5190.6760.8890.07
", + "image_path": "2f112880a869062099ecff6ebd1ac88880980c68ae585008f0c2750200a77cc8.jpg" + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "table_body" + } + ], + "index": 4 + }, + { + "bbox": [ + 165, + 275, + 447, + 287 + ], + "lines": [ + { + "bbox": [ + 165, + 275, + 447, + 287 + ], + "spans": [ + { + "bbox": [ + 165, + 275, + 447, + 287 + ], + "type": "text", + "content": "Table 3: Results comparison on the tasks beyond the planning task." + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "text" + }, + { + "bbox": [ + 132, + 319, + 230, + 333 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 319, + 230, + 333 + ], + "spans": [ + { + "bbox": [ + 132, + 319, + 230, + 333 + ], + "type": "text", + "content": "4 Experiments" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 132, + 346, + 262, + 358 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 346, + 262, + 358 + ], + "spans": [ + { + "bbox": [ + 132, + 346, + 262, + 358 + ], + "type": "text", + "content": "4.1 Experimental Setup" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 130, + 365, + 482, + 473 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 365, + 482, + 473 + ], + "spans": [ + { + "bbox": [ + 130, + 365, + 482, + 473 + ], + "type": "text", + "content": "Dataset We evaluate our method on two challenging large-scale real-world datasets, nuScenes [3] and Argoverse2 [44]. We conduct ablation studies to evaluate the effectiveness of our proposed components on the nuScenes [3] dataset. The nuScenes dataset [3] provides about 1K 20-second diverse driving scenes collected in Boston, Pittsburgh, Las Vegas, and Singapore for open-loop settings. Key samples which contain 6 camera images are annotated at " + }, + { + "bbox": [ + 130, + 365, + 482, + 473 + ], + "type": "inline_equation", + "content": "2\\mathrm{Hz}" + }, + { + "bbox": [ + 130, + 365, + 482, + 473 + ], + "type": "text", + "content": ". The Argoverse2 dataset comprises 1K 15-second scenes. It captures 7 RGB images on each frame. We align the data sampling frequency to nuScenes [3] by downsampling the annotated frames by an interval of 5 frames." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 130, + 474, + 482, + 593 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 474, + 482, + 593 + ], + "spans": [ + { + "bbox": [ + 130, + 474, + 482, + 593 + ], + "type": "text", + "content": "Metrics We adopt the metrics of L2 Displacement Error (L2) and Collision Rate (CR) for evaluation [18]. L2 is measured between the prediction and ground-truth trajectories over the timesteps 1-s, 2-s, and 3-s, evaluating the trajectory quality. Collision Rate (CR) measures how often the collision occurs between the ego vehicle and the other agents along the planning horizon, reflecting the trajectory safety. We notice that the UniAD [19] adheres to different calculations from ST-P3 [18] and VAD [23]: the former [19] reports the evaluations at each second. In contrast, the latters [18, 23] reports the results of the cumulative average by each second. We faithfully make comparisons of our method to others with these two kinds of computations." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 130, + 594, + 482, + 666 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 594, + 482, + 666 + ], + "spans": [ + { + "bbox": [ + 130, + 594, + 482, + 666 + ], + "type": "text", + "content": "Implementation Details We strictly follow the standard settings as proposed by UniAD [19] and VAD [23], which did not use the information of the historical ego trajectory. Same as VAD [23], the perception range is " + }, + { + "bbox": [ + 130, + 594, + 482, + 666 + ], + "type": "inline_equation", + "content": "60m \\times 30m" + }, + { + "bbox": [ + 130, + 594, + 482, + 666 + ], + "type": "text", + "content": ". PPAD also recurrently encodes 2-s historical information into BEV and predicts the 3-s trajectory in the future. It also conducts the tasks of motion prediction and map construction. Our PPAD trains with a batch size of 1 using the Adam [25]" + } + ] + } + ], + "index": 10 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 420, + 91, + 447, + 100 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 420, + 91, + 447, + 100 + ], + "spans": [ + { + "bbox": [ + 420, + 91, + 447, + 100 + ], + "type": "text", + "content": "PPAD" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 470, + 91, + 480, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 470, + 91, + 480, + 100 + ], + "spans": [ + { + "bbox": [ + 470, + 91, + 480, + 100 + ], + "type": "text", + "content": "11" + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 10 + }, + { + "para_blocks": [ + { + "bbox": [ + 130, + 116, + 479, + 140 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 116, + 479, + 140 + ], + "spans": [ + { + "bbox": [ + 130, + 116, + 479, + 140 + ], + "type": "text", + "content": "optimizer set with an initial learning rate of 2e-4. It takes us about six days to conduct the end-to-end training on eight NVIDIA A30 24GB GPUs." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 132, + 156, + 228, + 167 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 156, + 228, + 167 + ], + "spans": [ + { + "bbox": [ + 132, + 156, + 228, + 167 + ], + "type": "text", + "content": "4.2 Main Results" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 130, + 175, + 480, + 258 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 175, + 480, + 258 + ], + "spans": [ + { + "bbox": [ + 130, + 175, + 480, + 258 + ], + "type": "text", + "content": "Planning Results As shown in Tab. 1, our PPAD outperforms the current state-of-the-art performance by a large margin. Especially for the L2 distance metrics, there are about " + }, + { + "bbox": [ + 130, + 175, + 480, + 258 + ], + "type": "inline_equation", + "content": "20\\%" + }, + { + "bbox": [ + 130, + 175, + 480, + 258 + ], + "type": "text", + "content": " of consistent improvements can be observed along the temporal horizon. Thanks to the iterative interaction of prediction and planning, PPAD can help avoid collisions, leading to better results on collision rate compared to the one-shot representative VAD [23]. At the same time, we maintain a competitive efficiency compared to UniAD [19]." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 130, + 259, + 480, + 305 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 259, + 480, + 305 + ], + "spans": [ + { + "bbox": [ + 130, + 259, + 480, + 305 + ], + "type": "text", + "content": "We further make a fair comparison between our method and the baseline, VAD [23], on the other dataset of Argoverse 2 [6,44]. In Tab. 2, our method can consistently outperform the baseline with a clear margin in both L2 distance and collision rate metrics." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 130, + 307, + 480, + 366 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 307, + 480, + 366 + ], + "spans": [ + { + "bbox": [ + 130, + 307, + 480, + 366 + ], + "type": "text", + "content": "Subtasks Results To demonstrate the overall performance of our PPAD, we also provide the evaluation results besides planning metrics on the traditional perception and motion forecasting task in Tab. 3. Our PPAD also achieves promising performance in upstreaming perception and prediction tasks, which demonstrates that the whole system is jointly optimized." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 130, + 367, + 480, + 414 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 367, + 480, + 414 + ], + "spans": [ + { + "bbox": [ + 130, + 367, + 480, + 414 + ], + "type": "text", + "content": "Qualitative Results We provide qualitative results shown in Fig. 3. PPAD can perceive the scene precisely and predict with reasonable and diverse motions for the surrounding agents. It also plans a smooth and accurate trajectory for the ego vehicle." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 132, + 431, + 237, + 443 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 431, + 237, + 443 + ], + "spans": [ + { + "bbox": [ + 132, + 431, + 237, + 443 + ], + "type": "text", + "content": "4.3 Ablation Study" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 130, + 450, + 480, + 473 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 450, + 480, + 473 + ], + "spans": [ + { + "bbox": [ + 130, + 450, + 480, + 473 + ], + "type": "text", + "content": "The following experiments adhere to the progressive training pipeline as proposed in VAD [23]." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 130, + 474, + 481, + 665 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 474, + 481, + 665 + ], + "spans": [ + { + "bbox": [ + 130, + 474, + 481, + 665 + ], + "type": "text", + "content": "Effectiveness of Designs We provide ablation studies to verify the effectiveness of our proposed components. As shown in Tab. 4, the proposed PPAD framework (row 2) brings a remarkable improvement compared with one-shoot methods [23] (row 1). The multi-step interactions help the ego agent better understand the intention and potential effects brought by its actions along the temporal horizon, leading to an over " + }, + { + "bbox": [ + 130, + 474, + 481, + 665 + ], + "type": "inline_equation", + "content": "10\\%" + }, + { + "bbox": [ + 130, + 474, + 481, + 665 + ], + "type": "text", + "content": " L2 distance error reduction. We can observe a further improvement regarding the L2 distance with our proposed key objects attention (row 3). The slight degradation in collision rate might be due to the key objects attention being conducted on the ego with each mode of the agents, and the diverse modes of the agents mislead the behavior of the ego vehicle. From rows 5-7, the confidence-aware collision loss and noisy trajectory as the prediction can circumvent this phenomenon and further escalate the capability in planning accuracy and avoiding a collision. When we keep all of the remaining components while not applying key objects attention (row 4), we observed the degradations from row 7 in L2 " + }, + { + "bbox": [ + 130, + 474, + 481, + 665 + ], + "type": "inline_equation", + "content": "(0.58\\mathrm{m}" + }, + { + "bbox": [ + 130, + 474, + 481, + 665 + ], + "type": "text", + "content": " vs. " + }, + { + "bbox": [ + 130, + 474, + 481, + 665 + ], + "type": "inline_equation", + "content": "0.59\\mathrm{m})" + }, + { + "bbox": [ + 130, + 474, + 481, + 665 + ], + "type": "text", + "content": " and collision rate " + }, + { + "bbox": [ + 130, + 474, + 481, + 665 + ], + "type": "inline_equation", + "content": "(0.19\\%)" + }, + { + "bbox": [ + 130, + 474, + 481, + 665 + ], + "type": "text", + "content": " vs. " + }, + { + "bbox": [ + 130, + 474, + 481, + 665 + ], + "type": "inline_equation", + "content": "0.21\\%" + }, + { + "bbox": [ + 130, + 474, + 481, + 665 + ], + "type": "text", + "content": " , proving the effectiveness of the key objects attention." + } + ] + } + ], + "index": 10 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "spans": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "type": "text", + "content": "12" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 166, + 91, + 223, + 100 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 166, + 91, + 223, + 100 + ], + "spans": [ + { + "bbox": [ + 166, + 91, + 223, + 100 + ], + "type": "text", + "content": "Z. Chen et al." + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 11 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 150, + 114, + 462, + 368 + ], + "blocks": [ + { + "bbox": [ + 150, + 114, + 462, + 368 + ], + "lines": [ + { + "bbox": [ + 150, + 114, + 462, + 368 + ], + "spans": [ + { + "bbox": [ + 150, + 114, + 462, + 368 + ], + "type": "image", + "image_path": "6f6db780942a50b689bf835fc86986a19030e231705f53fa316ba8778258c77c.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 130, + 375, + 482, + 398 + ], + "lines": [ + { + "bbox": [ + 130, + 375, + 482, + 398 + ], + "spans": [ + { + "bbox": [ + 130, + 375, + 482, + 398 + ], + "type": "text", + "content": "Fig. 3: Qualitative results of PPAD. The green box in the figure demonstrates the ego agent, while the red ones are agents." + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_caption" + } + ], + "index": 2 + }, + { + "type": "table", + "bbox": [ + 139, + 413, + 480, + 492 + ], + "blocks": [ + { + "bbox": [ + 139, + 413, + 480, + 492 + ], + "lines": [ + { + "bbox": [ + 139, + 413, + 480, + 492 + ], + "spans": [ + { + "bbox": [ + 139, + 413, + 480, + 492 + ], + "type": "table", + "html": "
PPADKey Objects Attn.\\( {\\mathcal{L}}_{CA - {Col}} \\)Noisy Traj.L2 (m) \\( \\downarrow \\)Collision (%) \\( \\downarrow \\)Latency (ms)
1s2s3sAvg.1s2s3sAvg.
1----0.410.701.050.720.070.170.410.22224
2✓---0.350.590.880.600.110.180.410.23318
3✓✓--0.320.550.830.570.190.280.580.35385
4✓-✓✓0.330.570.880.590.100.140.380.21318
5✓✓✓-0.350.590.890.610.080.140.320.18385
6✓✓-✓0.340.590.900.610.100.140.340.19385
7✓✓✓✓0.310.560.870.580.080.120.380.19385
", + "image_path": "aad3653c19b9272d049bc9de00f321b578aabcb517e1926106ad4025c490f3e9.jpg" + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "table_body" + } + ], + "index": 4 + }, + { + "bbox": [ + 130, + 500, + 482, + 556 + ], + "lines": [ + { + "bbox": [ + 130, + 500, + 482, + 556 + ], + "spans": [ + { + "bbox": [ + 130, + 500, + 482, + 556 + ], + "type": "text", + "content": "Table 4: Component study for PPAD. Models follow the progressive training pipeline. PPAD means the auto-regressive framework with the designed stepwise interactions. Key Objects Attention represents hierarchical feature learning for the key objects. " + }, + { + "bbox": [ + 130, + 500, + 482, + 556 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_{CA - Col}" + }, + { + "bbox": [ + 130, + 500, + 482, + 556 + ], + "type": "text", + "content": " represents the loss design for the confidence-aware collision loss. Noisy Traj. means that we incorporate noisy trajectories while training." + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "text" + }, + { + "bbox": [ + 130, + 594, + 482, + 666 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 594, + 482, + 666 + ], + "spans": [ + { + "bbox": [ + 130, + 594, + 482, + 666 + ], + "type": "text", + "content": "Effectiveness of Interactions Our PPAD framework enables richer interactions among scene contexts, introducing local and global understandings of the world to the model. We further conduct ablation studies to demonstrate the performance gains brought by the interactions. We conduct the ablation study on interactions under the setting (Tab. 4 in row 5) without using the noisy trajectory as the prediction for better comparison. As illustrated in Tab. 5, we can achieve" + } + ] + } + ], + "index": 6 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 420, + 91, + 447, + 100 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 420, + 91, + 447, + 100 + ], + "spans": [ + { + "bbox": [ + 420, + 91, + 447, + 100 + ], + "type": "text", + "content": "PPAD" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 470, + 91, + 481, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 470, + 91, + 481, + 100 + ], + "spans": [ + { + "bbox": [ + 470, + 91, + 481, + 100 + ], + "type": "text", + "content": "13" + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 12 + }, + { + "para_blocks": [ + { + "type": "table", + "bbox": [ + 207, + 114, + 410, + 166 + ], + "blocks": [ + { + "bbox": [ + 207, + 114, + 410, + 166 + ], + "lines": [ + { + "bbox": [ + 207, + 114, + 410, + 166 + ], + "spans": [ + { + "bbox": [ + 207, + 114, + 410, + 166 + ], + "type": "table", + "html": "
EAMapBEVL2 (m) ↓Collision (%) ↓
1s2s3sAvg.1s2s3sAvg.
✓--0.350.580.890.610.230.240.490.32
✓✓-0.380.640.950.660.170.180.480.28
✓✓✓0.350.590.890.610.080.140.320.18
", + "image_path": "d9bee3110fbb9459bde671ef2c1377b8d566a26f0ec76d58a6ed327de343badc.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "table_body" + } + ], + "index": 2 + }, + { + "type": "table", + "bbox": [ + 141, + 227, + 478, + 288 + ], + "blocks": [ + { + "bbox": [ + 132, + 174, + 481, + 207 + ], + "lines": [ + { + "bbox": [ + 132, + 174, + 481, + 207 + ], + "spans": [ + { + "bbox": [ + 132, + 174, + 481, + 207 + ], + "type": "text", + "content": "Table 5: Interaction study for PPAD. EA, Map, BEV mean the interactions of the ego with agents, the ego and agents with the map, the ego and agents with the BEV, respectively." + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 141, + 227, + 478, + 288 + ], + "lines": [ + { + "bbox": [ + 141, + 227, + 478, + 288 + ], + "spans": [ + { + "bbox": [ + 141, + 227, + 478, + 288 + ], + "type": "table", + "html": "
Interaction IterationsL2 (m) ↓Collision (%) ↓Latency (ms)
1s2s3sAvg.1s2s3sAvg.
2(Every 1.5 sec)0.360.630.960.650.140.200.430.25306
3(Every 1.0 sec)0.370.630.950.650.100.150.390.21326
6(Every 0.5 sec)0.310.560.870.580.080.120.380.19385
", + "image_path": "1a70c634fc3e97aa242a3877cd6bcc5b0527fe7179c0cc187acfb47a4bbb3ef3.jpg" + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "table_body" + } + ], + "index": 4 + }, + { + "bbox": [ + 132, + 296, + 481, + 319 + ], + "lines": [ + { + "bbox": [ + 132, + 296, + 481, + 319 + ], + "spans": [ + { + "bbox": [ + 132, + 296, + 481, + 319 + ], + "type": "text", + "content": "Table 6: The ablation study of conducting different interaction iterations in the future time horizons." + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "text" + }, + { + "bbox": [ + 130, + 350, + 480, + 373 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 350, + 480, + 373 + ], + "spans": [ + { + "bbox": [ + 130, + 350, + 480, + 373 + ], + "type": "text", + "content": "the best performance in terms of L2 distance and collision by incorporating all of the interactions." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 130, + 374, + 481, + 495 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 374, + 481, + 495 + ], + "spans": [ + { + "bbox": [ + 130, + 374, + 481, + 495 + ], + "type": "text", + "content": "Effect of Different Iterations on Prediction and Planning Interaction Our innovative PPAD interaction mechanism not only better models motion planning as gaming among the ego and the agents but also enriches the ego's / agents' interactions with their local environments. We further conducted the ablation study on the different iterations in applying the interactions of prediction and planning, as shown in Tab. 6. Specifically, PPAD plans the trajectories with 3, 2, and 1 steps of waypoints after each of the prediction-planning processes for the interaction iterations of 2, 3, and 6 in Tab. 6. It is demonstrated that the performance reaches the best as we conduct the interactions of prediction and planning processes at every future step." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 132, + 510, + 220, + 523 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 510, + 220, + 523 + ], + "spans": [ + { + "bbox": [ + 132, + 510, + 220, + 523 + ], + "type": "text", + "content": "5 Conclusion" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 130, + 533, + 481, + 666 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 533, + 481, + 666 + ], + "spans": [ + { + "bbox": [ + 130, + 533, + 481, + 666 + ], + "type": "text", + "content": "In this paper, we have presented a novel autonomous driving framework, PPAD. Different from the previous methods that lack in-depth modeling of interactions, we pose the planning problem as a multi-step Prediction and Planning gaming process among the ego vehicle and agents. With PPAD architecture, our proposed hierarchical dynamic key objects attention is incorporated to learn local and global scene contexts at each step and eventually plan with a more precise trajectory. The confidence-aware collision constraint and noisy trajectories are utilized while training to improve driving safety further. In general, our proposed novel PPAD achieves compelling performance upon the existing state-of-the-art methods, and we hope the PPAD framework can inspire the community to further exploration." + } + ] + } + ], + "index": 9 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "spans": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "type": "text", + "content": "14" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 166, + 91, + 223, + 100 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 166, + 91, + 223, + 100 + ], + "spans": [ + { + "bbox": [ + 166, + 91, + 223, + 100 + ], + "type": "text", + "content": "Z. Chen et al." + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 13 + }, + { + "para_blocks": [ + { + "bbox": [ + 133, + 114, + 246, + 129 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 114, + 246, + 129 + ], + "spans": [ + { + "bbox": [ + 133, + 114, + 246, + 129 + ], + "type": "text", + "content": "Acknowledgements" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 130, + 141, + 487, + 190 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 141, + 487, + 190 + ], + "spans": [ + { + "bbox": [ + 130, + 141, + 487, + 190 + ], + "type": "text", + "content": "The authors are thankful for the financial support from the Hetao Shenzhen-HongKong Science and Technology Innovation Cooperation Zone (HZQB-KCZYZ-2021055), this work was also supported by Shenzhen Deeproute.ai Co., Ltd (HZQB-KCZYZ-2021055)." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 132, + 209, + 197, + 220 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 209, + 197, + 220 + ], + "spans": [ + { + "bbox": [ + 132, + 209, + 197, + 220 + ], + "type": "text", + "content": "References" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 134, + 234, + 480, + 665 + ], + "type": "list", + "angle": 0, + "index": 18, + "blocks": [ + { + "bbox": [ + 138, + 234, + 480, + 257 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 234, + 480, + 257 + ], + "spans": [ + { + "bbox": [ + 138, + 234, + 480, + 257 + ], + "type": "text", + "content": "1. Bansal, M., Krizhevsky, A., Ogale, A.: Chauffeurnet. In: Robotics: Science and Systems XV (2019)" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 138, + 258, + 480, + 290 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 258, + 480, + 290 + ], + "spans": [ + { + "bbox": [ + 138, + 258, + 480, + 290 + ], + "type": "text", + "content": "2. Bojarski, M., Del Testa, D., Dworakowski, D., Firner, B., Flepp, B., Goyal, P., Jackel, L.D., Monfort, M., Muller, U., Zhang, J., et al.: End to end learning for self-driving cars. arXiv preprint arXiv:1604.07316 (2016)" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 138, + 291, + 480, + 334 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 291, + 480, + 334 + ], + "spans": [ + { + "bbox": [ + 138, + 291, + 480, + 334 + ], + "type": "text", + "content": "3. Caesar, H., Bankiti, V., Lang, A.H., Vora, S., Liong, V.E., Xu, Q., Krishnan, A., Pan, Y., Baldan, G., Beijbom, O.: nuscenes: A multimodal dataset for autonomous driving. In: Proceedings of the IEEE/CVF conference on computer vision and pattern recognition. pp. 11621-11631 (2020)" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 138, + 335, + 480, + 356 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 335, + 480, + 356 + ], + "spans": [ + { + "bbox": [ + 138, + 335, + 480, + 356 + ], + "type": "text", + "content": "4. Casas, S., Luo, W., Urtasun, R.: Intentnet: Learning to predict intention from raw sensor data. In: Conference on Robot Learning. pp. 947-956 (2018)" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 138, + 357, + 480, + 388 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 357, + 480, + 388 + ], + "spans": [ + { + "bbox": [ + 138, + 357, + 480, + 388 + ], + "type": "text", + "content": "5. Chai, Y., Sapp, B., Bansal, M., Anguelov, D.: Multipath: Multiple probabilistic anchor trajectory hypotheses for behavior prediction. arXiv preprint arXiv:1910.05449 (2019)" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 138, + 390, + 480, + 433 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 390, + 480, + 433 + ], + "spans": [ + { + "bbox": [ + 138, + 390, + 480, + 433 + ], + "type": "text", + "content": "6. Chang, M.F., Lambert, J., Sangkloy, P., Singh, J., Bak, S., Hartnett, A., Wang, D., Carr, P., Lucey, S., Ramanan, D., et al.: Argoverse: 3d tracking and forecasting with rich maps. In: Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition. pp. 8748-8757 (2019)" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 138, + 434, + 480, + 466 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 434, + 480, + 466 + ], + "spans": [ + { + "bbox": [ + 138, + 434, + 480, + 466 + ], + "type": "text", + "content": "7. Chekroun, R., Toromanoff, M., Hornauer, S., Moutarde, F.: Gri: General reinforced imitation and its application to vision-based autonomous driving. arXiv preprint arXiv:2111.08575 (2021)" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 138, + 467, + 480, + 488 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 467, + 480, + 488 + ], + "spans": [ + { + "bbox": [ + 138, + 467, + 480, + 488 + ], + "type": "text", + "content": "8. Chen, D., Zhou, B., Koltun, V., Krahenbuhl, P.: Learning by cheating. In: Conference on Robot Learning. pp. 66-75. PMLR (2020)" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 138, + 489, + 480, + 521 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 489, + 480, + 521 + ], + "spans": [ + { + "bbox": [ + 138, + 489, + 480, + 521 + ], + "type": "text", + "content": "9. Chitta, K., Prakash, A., Jaeger, B., Yu, Z., Renz, K., Geiger, A.: Transfuser: Imitation with transformer-based sensor fusion for autonomous driving. IEEE Transactions on Pattern Analysis and Machine Intelligence (2022)" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 138, + 522, + 480, + 555 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 522, + 480, + 555 + ], + "spans": [ + { + "bbox": [ + 138, + 522, + 480, + 555 + ], + "type": "text", + "content": "10. Codevilla, F., Santana, E., López, A.M., Gaidon, A.: Exploring the limitations of behavior cloning for autonomous driving. In: Proceedings of the IEEE/CVF International Conference on Computer Vision. pp. 9329-9338 (2019)" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 134, + 555, + 480, + 599 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 134, + 555, + 480, + 599 + ], + "spans": [ + { + "bbox": [ + 134, + 555, + 480, + 599 + ], + "type": "text", + "content": "11. Duvenaud, D.K., Maclaurin, D., Iparraguirre, J., Bombarell, R., Hirzel, T., Aspuru-Guzik, A., Adams, R.P.: Convolutional networks on graphs for learning molecular fingerprints. In: Advances in neural information processing systems. pp. 2224-2232 (2015)" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 134, + 600, + 480, + 621 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 134, + 600, + 480, + 621 + ], + "spans": [ + { + "bbox": [ + 134, + 600, + 480, + 621 + ], + "type": "text", + "content": "12. Fan, H., Zhu, F., Liu, C., Zhang, L., Zhuang, L., Li, D., Zhu, W., Hu, J., Li, H., Kong, Q.: BaiduApollo em motion planner. arXiv preprint arXiv:1807.08048 (2018)" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 134, + 622, + 480, + 665 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 134, + 622, + 480, + 665 + ], + "spans": [ + { + "bbox": [ + 134, + 622, + 480, + 665 + ], + "type": "text", + "content": "13. Gao, J., Sun, C., Zhao, H., Shen, Y., Anguelov, D., Li, C., Schmid, C.: Vectornet: Encoding hd maps and agent dynamics from vectorized representation. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 11525-11533 (2020)" + } + ] + } + ], + "index": 17 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 420, + 91, + 447, + 100 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 420, + 91, + 447, + 100 + ], + "spans": [ + { + "bbox": [ + 420, + 91, + 447, + 100 + ], + "type": "text", + "content": "PPAD" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 470, + 92, + 481, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 470, + 92, + 481, + 100 + ], + "spans": [ + { + "bbox": [ + 470, + 92, + 481, + 100 + ], + "type": "text", + "content": "15" + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 14 + }, + { + "para_blocks": [ + { + "bbox": [ + 132, + 116, + 481, + 665 + ], + "type": "list", + "angle": 0, + "index": 18, + "blocks": [ + { + "bbox": [ + 133, + 116, + 481, + 160 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 116, + 481, + 160 + ], + "spans": [ + { + "bbox": [ + 133, + 116, + 481, + 160 + ], + "type": "text", + "content": "14. Gu, J., Hu, C., Zhang, T., Chen, X., Wang, Y., Wang, Y., Zhao, H.: Vip3d: End-to-end visual trajectory prediction via 3d agent queries. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 5496-5506 (2023)" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 133, + 160, + 481, + 193 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 160, + 481, + 193 + ], + "spans": [ + { + "bbox": [ + 133, + 160, + 481, + 193 + ], + "type": "text", + "content": "15. He, K., Zhang, X., Ren, S., Sun, J.: Deep residual learning for image recognition. In: Proceedings of the IEEE conference on computer vision and pattern recognition. pp. 770-778 (2016)" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 133, + 193, + 481, + 214 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 193, + 481, + 214 + ], + "spans": [ + { + "bbox": [ + 133, + 193, + 481, + 214 + ], + "type": "text", + "content": "16. Henaff, M., Bruna, J., LeCun, Y.: Deep convolutional networks on graph-structured data. arXiv preprint arXiv:1506.05163 (2015)" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 133, + 214, + 481, + 246 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 214, + 481, + 246 + ], + "spans": [ + { + "bbox": [ + 133, + 214, + 481, + 246 + ], + "type": "text", + "content": "17. Hu, P., Huang, A., Dolan, J., Held, D., Ramanan, D.: Safe local motion planning with self-supervised freespace forecasting. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 12732-12741 (2021)" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 133, + 246, + 481, + 278 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 246, + 481, + 278 + ], + "spans": [ + { + "bbox": [ + 133, + 246, + 481, + 278 + ], + "type": "text", + "content": "18. Hu, S., Chen, L., Wu, P., Li, H., Yan, J., Tao, D.: St-p3: End-to-end vision-based autonomous driving via spatial-temporal feature learning. In: European Conference on Computer Vision. pp. 533-549. Springer (2022)" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 133, + 278, + 481, + 320 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 278, + 481, + 320 + ], + "spans": [ + { + "bbox": [ + 133, + 278, + 481, + 320 + ], + "type": "text", + "content": "19. Hu, Y., Yang, J., Chen, L., Li, K., Sima, C., Zhu, X., Chai, S., Du, S., Lin, T., Wang, W., et al.: Planning-oriented autonomous driving. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 17853-17862 (2023)" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 132, + 321, + 481, + 354 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 321, + 481, + 354 + ], + "spans": [ + { + "bbox": [ + 132, + 321, + 481, + 354 + ], + "type": "text", + "content": "20. Huang, J., Huang, G., Zhu, Z., Ye, Y., Du, D.: Bevdet: High-performance multicamera 3d object detection in bird-eye-view. arXiv preprint arXiv:2112.11790 (2021)" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 132, + 354, + 481, + 386 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 354, + 481, + 386 + ], + "spans": [ + { + "bbox": [ + 132, + 354, + 481, + 386 + ], + "type": "text", + "content": "21. Huang, Z., Liu, H., Lv, C.: Gameformer: Game-theoretic modeling and learning of transformer-based interactive prediction and planning for autonomous driving. arXiv preprint arXiv:2303.05760 (2023)" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 132, + 386, + 481, + 429 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 386, + 481, + 429 + ], + "spans": [ + { + "bbox": [ + 132, + 386, + 481, + 429 + ], + "type": "text", + "content": "22. Jia, X., Wu, P., Chen, L., Xie, J., He, C., Yan, J., Li, H.: Think twice before driving: Towards scalable decoders for end-to-end autonomous driving. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 21983-21994 (2023)" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 132, + 429, + 481, + 461 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 429, + 481, + 461 + ], + "spans": [ + { + "bbox": [ + 132, + 429, + 481, + 461 + ], + "type": "text", + "content": "23. Jiang, B., Chen, S., Xu, Q., Liao, B., Chen, J., Zhou, H., Zhang, Q., Liu, W., Huang, C., Wang, X.: Vad: Vectorized scene representation for efficient autonomous driving. arXiv preprint arXiv:2303.12077 (2023)" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 132, + 461, + 481, + 494 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 461, + 481, + 494 + ], + "spans": [ + { + "bbox": [ + 132, + 461, + 481, + 494 + ], + "type": "text", + "content": "24. Khurana, T., Hu, P., Dave, A., Ziglar, J., Held, D., Ramanan, D.: Differentiable raycasting for self-supervised occupancy forecasting. In: European Conference on Computer Vision. pp. 353-369. Springer (2022)" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 132, + 494, + 481, + 514 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 494, + 481, + 514 + ], + "spans": [ + { + "bbox": [ + 132, + 494, + 481, + 514 + ], + "type": "text", + "content": "25. Kingma, D.P., Ba, J.: Adam: A method for stochastic optimization. arXiv preprint arXiv:1412.6980 (2014)" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 132, + 514, + 481, + 557 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 514, + 481, + 557 + ], + "spans": [ + { + "bbox": [ + 132, + 514, + 481, + 557 + ], + "type": "text", + "content": "26. Lang, A.H., Vora, S., Caesar, H., Zhou, L., Yang, J., Beijbom, O.: PointPillars: Fast Encoders for Object Detection From Point Clouds. In: Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition. pp. 12697-12705 (2019)" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 132, + 557, + 481, + 590 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 557, + 481, + 590 + ], + "spans": [ + { + "bbox": [ + 132, + 557, + 481, + 590 + ], + "type": "text", + "content": "27. Li, F., Zhang, H., Liu, S., Guo, J., Ni, L.M., Zhang, L.: Dn-detr: Accelerate detr training by introducing query denoising. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 13619-13627 (2022)" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 132, + 590, + 481, + 622 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 590, + 481, + 622 + ], + "spans": [ + { + "bbox": [ + 132, + 590, + 481, + 622 + ], + "type": "text", + "content": "28. Li, Y., Ge, Z., Yu, G., Yang, J., Wang, Z., Shi, Y., Sun, J., Li, Z.: Bevdepth: Acquisition of reliable depth for multi-view 3d object detection. In: Proceedings of the AAAI Conference on Artificial Intelligence. pp. 1477-1485 (2023)" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 132, + 622, + 481, + 665 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 622, + 481, + 665 + ], + "spans": [ + { + "bbox": [ + 132, + 622, + 481, + 665 + ], + "type": "text", + "content": "29. Li, Z., Wang, W., Li, H., Xie, E., Sima, C., Lu, T., Qiao, Y., Dai, J.: Bevformer: Learning bird's-eye-view representation from multi-camera images via spatiotemporal transformers. In: European conference on computer vision. pp. 1-18. Springer (2022)" + } + ] + } + ], + "index": 17 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "spans": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "type": "text", + "content": "16" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 166, + 91, + 223, + 100 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 166, + 91, + 223, + 100 + ], + "spans": [ + { + "bbox": [ + 166, + 91, + 223, + 100 + ], + "type": "text", + "content": "Z. Chen et al." + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 15 + }, + { + "para_blocks": [ + { + "bbox": [ + 130, + 116, + 482, + 666 + ], + "type": "list", + "angle": 0, + "index": 17, + "blocks": [ + { + "bbox": [ + 130, + 116, + 482, + 150 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 116, + 482, + 150 + ], + "spans": [ + { + "bbox": [ + 130, + 116, + 482, + 150 + ], + "type": "text", + "content": "30. Liang, M., Yang, B., Hu, R., Chen, Y., Liao, R., Feng, S., Urtasun, R.: Learning lane graph representations for motion forecasting. In: Proceedings of the European Conference on Computer Vision (ECCV). pp. 541-556 (2020)" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 130, + 150, + 481, + 184 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 150, + 481, + 184 + ], + "spans": [ + { + "bbox": [ + 130, + 150, + 481, + 184 + ], + "type": "text", + "content": "31. Liao, B., Chen, S., Wang, X., Cheng, T., Zhang, Q., Liu, W., Huang, C.: Maptr: Structured modeling and learning for online vectorized hd map construction. arXiv preprint arXiv:2208.14437 (2022)" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 130, + 184, + 481, + 217 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 184, + 481, + 217 + ], + "spans": [ + { + "bbox": [ + 130, + 184, + 481, + 217 + ], + "type": "text", + "content": "32. Liu, Y., Zhang, J., Fang, L., Jiang, Q., Zhou, B.: Multimodal motion prediction with stacked transformers. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 7577-7586 (2021)" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 130, + 218, + 481, + 251 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 218, + 481, + 251 + ], + "spans": [ + { + "bbox": [ + 130, + 218, + 481, + 251 + ], + "type": "text", + "content": "33. Liu, Y., Wang, T., Zhang, X., Sun, J.: Petr: Position embedding transformation for multi-view 3d object detection. In: European Conference on Computer Vision. pp. 531-548. Springer (2022)" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 130, + 251, + 481, + 285 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 251, + 481, + 285 + ], + "spans": [ + { + "bbox": [ + 130, + 251, + 481, + 285 + ], + "type": "text", + "content": "34. Liu, Y., Yan, J., Jia, F., Li, S., Gao, A., Wang, T., Zhang, X., Sun, J.: Petrv2: A unified framework for 3d perception from multi-camera images. arXiv preprint arXiv:2206.01256 (2022)" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 130, + 285, + 481, + 329 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 285, + 481, + 329 + ], + "spans": [ + { + "bbox": [ + 130, + 285, + 481, + 329 + ], + "type": "text", + "content": "35. Liu, Z., Chen, S., Guo, X., Wang, X., Cheng, T., Zhu, H., Zhang, Q., Liu, W., Zhang, Y.: Vision-based uneven bev representation learning with polar rasterization and surface estimation. In: Conference on Robot Learning. pp. 437-446. PMLR (2023)" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 130, + 329, + 481, + 374 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 329, + 481, + 374 + ], + "spans": [ + { + "bbox": [ + 130, + 329, + 481, + 374 + ], + "type": "text", + "content": "36. Liu, Z., Tang, H., Amini, A., Yang, X., Mao, H., Rus, D.L., Han, S.: Bevfusion: Multi-task multi-sensor fusion with unified bird's-eye view representation. In: 2023 IEEE International Conference on Robotics and Automation (ICRA). pp. 2774-2781. IEEE (2023)" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 130, + 374, + 481, + 418 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 374, + 481, + 418 + ], + "spans": [ + { + "bbox": [ + 130, + 374, + 481, + 418 + ], + "type": "text", + "content": "37. Luo, W., Yang, B., Urtasun, R.: Fast and furious: Real time end-to-end 3d detection, tracking and motion forecasting with a single convolutional net. In: Proceedings of the IEEE conference on Computer Vision and Pattern Recognition. pp. 3569-3577 (2018)" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 130, + 418, + 481, + 463 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 418, + 481, + 463 + ], + "spans": [ + { + "bbox": [ + 130, + 418, + 481, + 463 + ], + "type": "text", + "content": "38. Philion, J., Fidler, S.: Lift, splat, shoot: Encoding images from arbitrary camera rigs by implicitly unprojecting to 3d. In: Computer Vision-ECCV 2020: 16th European Conference, Glasgow, UK, August 23-28, 2020, Proceedings, Part XIV 16. pp. 194-210. Springer (2020)" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 130, + 463, + 481, + 497 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 463, + 481, + 497 + ], + "spans": [ + { + "bbox": [ + 130, + 463, + 481, + 497 + ], + "type": "text", + "content": "39. Renz, K., Chitta, K., Mercea, O.B., Koepke, A., Akata, Z., Geiger, A.: Plant: Explainable planning transformers via object-level representations. arXiv preprint arXiv:2210.14222 (2022)" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 130, + 497, + 481, + 541 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 497, + 481, + 541 + ], + "spans": [ + { + "bbox": [ + 130, + 497, + 481, + 541 + ], + "type": "text", + "content": "40. Shuman, D.I., Narang, S.K., Frossard, P., Ortega, A., Vandergheynst, P.: The emerging field of signal processing on graphs: Extending high-dimensional data analysis to networks and other irregular domains. IEEE signal processing magazine 30(3), 83-98 (2013)" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 130, + 541, + 481, + 564 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 541, + 481, + 564 + ], + "spans": [ + { + "bbox": [ + 130, + 541, + 481, + 564 + ], + "type": "text", + "content": "41. Sima, C., Tong, W., Wang, T., Chen, L., Wu, S., Deng, H., Gu, Y., Lu, L., Luo, P., Lin, D., et al.: Scene as occupancy. arXiv e-prints pp. arXiv-2306 (2023)" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 130, + 564, + 481, + 598 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 564, + 481, + 598 + ], + "spans": [ + { + "bbox": [ + 130, + 564, + 481, + 598 + ], + "type": "text", + "content": "42. Song, H., Luan, D., Ding, W., Wang, M.Y., Chen, Q.: Learning to predict vehicle trajectories with model-based planning. In: Conference on Robot Learning. pp. 1035-1045. PMLR (2021)" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 130, + 598, + 481, + 632 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 598, + 481, + 632 + ], + "spans": [ + { + "bbox": [ + 130, + 598, + 481, + 632 + ], + "type": "text", + "content": "43. Wang, S., Liu, Y., Wang, T., Li, Y., Zhang, X.: Exploring object-centric temporal modeling for efficient multi-view 3d object detection. arXiv preprint arXiv:2303.11926 (2023)" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 130, + 632, + 481, + 666 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 632, + 481, + 666 + ], + "spans": [ + { + "bbox": [ + 130, + 632, + 481, + 666 + ], + "type": "text", + "content": "44. Wilson, B., Qi, W., Agarwal, T., Lambert, J., Singh, J., Khandelwal, S., Pan, B., Kumar, R., Hartnett, A., Pontes, J.K., et al.: Argoverse 2: Next generation datasets for self-driving perception and forecasting. arXiv preprint arXiv:2301.00493 (2023)" + } + ] + } + ], + "index": 16 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 420, + 91, + 447, + 100 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 420, + 91, + 447, + 100 + ], + "spans": [ + { + "bbox": [ + 420, + 91, + 447, + 100 + ], + "type": "text", + "content": "PPAD" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 470, + 91, + 481, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 470, + 91, + 481, + 100 + ], + "spans": [ + { + "bbox": [ + 470, + 91, + 481, + 100 + ], + "type": "text", + "content": "17" + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 16 + }, + { + "para_blocks": [ + { + "bbox": [ + 132, + 116, + 480, + 577 + ], + "type": "list", + "angle": 0, + "index": 16, + "blocks": [ + { + "bbox": [ + 133, + 116, + 480, + 149 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 116, + 480, + 149 + ], + "spans": [ + { + "bbox": [ + 133, + 116, + 480, + 149 + ], + "type": "text", + "content": "45. Wu, P., Jia, X., Chen, L., Yan, J., Li, H., Qiao, Y.: Trajectory-guided control prediction for end-to-end autonomous driving: A simple yet strong baseline. Advances in Neural Information Processing Systems 35, 6119-6132 (2022)" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 133, + 150, + 480, + 182 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 150, + 480, + 182 + ], + "spans": [ + { + "bbox": [ + 133, + 150, + 480, + 182 + ], + "type": "text", + "content": "46. Xie, Y., Xu, C., Rakotosaona, M.J., Rim, P., Tombari, F., Keutzer, K., Tomizuka, M., Zhan, W.: Sparsefusion: Fusing multi-modal sparse representations for multisensor 3d object detection. arXiv preprint arXiv:2304.14340 (2023)" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 132, + 183, + 480, + 226 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 183, + 480, + 226 + ], + "spans": [ + { + "bbox": [ + 132, + 183, + 480, + 226 + ], + "type": "text", + "content": "47. Yang, C., Chen, Y., Tian, H., Tao, C., Zhu, X., Zhang, Z., Huang, G., Li, H., Qiao, Y., Lu, L., et al.: Bevformer v2: Adapting modern image backbones to bird's-eye-view recognition via perspective supervision. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 17830-17839 (2023)" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 132, + 228, + 480, + 258 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 228, + 480, + 258 + ], + "spans": [ + { + "bbox": [ + 132, + 228, + 480, + 258 + ], + "type": "text", + "content": "48. Ye, M., Cao, T., Chen, Q.: Tpcn: Temporal point cloud networks for motion forecasting. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 11318-11327 (2021)" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 132, + 259, + 480, + 293 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 259, + 480, + 293 + ], + "spans": [ + { + "bbox": [ + 132, + 259, + 480, + 293 + ], + "type": "text", + "content": "49. Ye, M., Xu, S., Cao, T.: Hvnet: Hybrid voxel network for lidar based 3d object detection. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 1631-1640 (2020)" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 132, + 293, + 480, + 325 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 293, + 480, + 325 + ], + "spans": [ + { + "bbox": [ + 132, + 293, + 480, + 325 + ], + "type": "text", + "content": "50. Ye, M., Xu, S., Cao, T., Chen, Q.: Drinet: A dual-representation iterative learning network for point cloud segmentation. In: Proceedings of the IEEE/CVF international conference on computer vision. pp. 7447-7456 (2021)" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 132, + 326, + 480, + 357 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 326, + 480, + 357 + ], + "spans": [ + { + "bbox": [ + 132, + 326, + 480, + 357 + ], + "type": "text", + "content": "51. Ye, T., Jing, W., Hu, C., Huang, S., Gao, L., Li, F., Wang, J., Guo, K., Xiao, W., Mao, W., et al.: Fusionad: Multi-modality fusion for prediction and planning tasks of autonomous driving. arXiv preprint arXiv:2308.01006 (2023)" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 132, + 358, + 480, + 379 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 358, + 480, + 379 + ], + "spans": [ + { + "bbox": [ + 132, + 358, + 480, + 379 + ], + "type": "text", + "content": "52. Zeng, W., Liang, M., Liao, R., Urtasun, R.: Lanercnn: Distributed representations for graph-centric motion forecasting. arXiv preprint arXiv:2101.06653 (2021)" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 132, + 380, + 480, + 413 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 380, + 480, + 413 + ], + "spans": [ + { + "bbox": [ + 132, + 380, + 480, + 413 + ], + "type": "text", + "content": "53. Zeng, W., Luo, W., Suo, S., Sadat, A., Yang, B., Casas, S., Urtasun, R.: End-to-end interpretable neural motion planner. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 8660-8669 (2019)" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 132, + 414, + 480, + 445 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 414, + 480, + 445 + ], + "spans": [ + { + "bbox": [ + 132, + 414, + 480, + 445 + ], + "type": "text", + "content": "54. Zhang, Z., Liniger, A., Dai, D., Yu, F., Van Gool, L.: End-to-end urban driving by imitating a reinforcement learning coach. In: Proceedings of the IEEE/CVF international conference on computer vision. pp. 15222-15232 (2021)" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 132, + 447, + 480, + 479 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 447, + 480, + 479 + ], + "spans": [ + { + "bbox": [ + 132, + 447, + 480, + 479 + ], + "type": "text", + "content": "55. Zhou, Y., Tuzel, O.: VoxelNet: End-to-End Learning for Point Cloud Based 3D Object Detection. In: Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition. pp. 4490-4499 (2018)" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 132, + 479, + 480, + 511 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 479, + 480, + 511 + ], + "spans": [ + { + "bbox": [ + 132, + 479, + 480, + 511 + ], + "type": "text", + "content": "56. Zhou, Z., Wang, J., Li, Y.H., Huang, Y.K.: Query-centric trajectory prediction. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 17863-17873 (2023)" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 132, + 512, + 480, + 544 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 512, + 480, + 544 + ], + "spans": [ + { + "bbox": [ + 132, + 512, + 480, + 544 + ], + "type": "text", + "content": "57. Zhou, Z., Ye, L., Wang, J., Wu, K., Lu, K.: Hivt: Hierarchical vector transformer for multi-agent motion prediction. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 8823-8833 (2022)" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 132, + 545, + 480, + 577 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 545, + 480, + 577 + ], + "spans": [ + { + "bbox": [ + 132, + 545, + 480, + 577 + ], + "type": "text", + "content": "58. Zhu, X., Su, W., Lu, L., Li, B., Wang, X., Dai, J.: Deformable detr: Deformable transformers for end-to-end object detection. arXiv preprint arXiv:2010.04159 (2020)" + } + ] + } + ], + "index": 15 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "spans": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "type": "text", + "content": "18" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 166, + 91, + 223, + 100 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 166, + 91, + 223, + 100 + ], + "spans": [ + { + "bbox": [ + 166, + 91, + 223, + 100 + ], + "type": "text", + "content": "Z. Chen et al." + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 17 + } + ], + "_backend": "vlm", + "_version_name": "2.6.4" +} \ No newline at end of file diff --git a/2024/PQ-SAM_ Post-training Quantization for Segment Anything Model/5e412ebd-1f25-44d2-98a8-cb85fe86acb8_content_list.json b/2024/PQ-SAM_ Post-training Quantization for Segment Anything Model/5e412ebd-1f25-44d2-98a8-cb85fe86acb8_content_list.json new file mode 100644 index 0000000000000000000000000000000000000000..2e3b5b4f878668fbdedf0399226e16a90dd0306d --- /dev/null +++ b/2024/PQ-SAM_ Post-training Quantization for Segment Anything Model/5e412ebd-1f25-44d2-98a8-cb85fe86acb8_content_list.json @@ -0,0 +1,1807 @@ +[ + { + "type": "text", + "text": "PQ-SAM: Post-training Quantization for Segment Anything Model", + "text_level": 1, + "bbox": [ + 217, + 140, + 782, + 185 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Xiaoyu Liu $^{1}$ , Xin Ding $^{1}$ , Lei Yu $^{2}$ , Yuanyuan Xi $^{2}$ , Wei Li $^{2}$ , Zhijun Tu $^{2}$ , Jie Hu $^{2}$ , Hanting Chen $^{2}$ , Baoqun Yin $^{1}$ , and Zhiwei Xiong $^{1(\\boxtimes)}$", + "bbox": [ + 215, + 210, + 784, + 243 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "1 University of Science and Technology of China {liuxyu,xinding64}@mail.ustc.edu.cn, zwxiong@ustc.edu.cn \n2 Huawei Noah's Ark Lab {wei.lee,zhijun.tu}@huawei.com", + "bbox": [ + 282, + 253, + 715, + 310 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Abstract. Segment anything model (SAM) is a promising prompt-guided vision foundation model to segment objects of interest. However, the extensive computational requirements of SAM have limited its applicability in resource-constraint edge devices. Post-training quantization (PTQ) is an effective potential for fast-deploying SAM. Nevertheless, SAM's billion-scale pretraining creates a highly asymmetric activation distribution with detrimental outliers in excessive channels, resulting in significant performance degradation of the low-bit PTQ. In this paper, we propose PQ-SAM, the first PTQ method customized for SAM. To achieve a quantization-friendly tensor-wise distribution, PQ-SAM incorporates a novel grouped activation distribution transformation (GADT) based on a two-stage outlier hierarchical clustering (OHC) scheme to scale and shift each channel. Firstly, OHC identifies and truncates extreme outliers to reduce the scale variance of different channels. Secondly, OHC iteratively allocates learnable shifting and scaling sizes to each group of channels with similar distributions, reducing the number of learnable parameters and easing the optimization difficulty. These shifting and scaling sizes are used to adjust activation channels, and jointly optimized with quantization step sizes for optimal results. Extensive experiments demonstrate that PQ-SAM outperforms existing PTQ methods on nine zero-shot datasets, and pushes the 4-bit PTQ of SAM to a usable level.", + "bbox": [ + 261, + 348, + 743, + 638 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Keywords: Segment Anything Model-Post-training Quantization", + "bbox": [ + 261, + 652, + 707, + 667 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "1 Introduction", + "text_level": 1, + "bbox": [ + 215, + 694, + 374, + 709 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "The advent of large language models [2,7,52] has ushered in a new paradigm in natural language processing (NLP). These models exhibit exceptional zero-shot and few-shot generalization, learning broad competencies from sizable pretraining [14,17,33]. Inspired by their success, researchers have begun developing similarly large-scale foundation models for computer vision. A pioneering example", + "bbox": [ + 212, + 726, + 784, + 801 + ], + "page_idx": 0 + }, + { + "type": "page_footnote", + "text": "X. Liu and X. Ding—Equal contribution. This research was conducted while Xiaoyu Liu and Xin Ding were interns at Huawei Noah's Ark Lab.", + "bbox": [ + 228, + 811, + 784, + 839 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/c17837d20f33e4d9a58e77e2045f76af8c61cecfa1985761feb3972b1a6a67fc.jpg", + "image_caption": [ + "(a) SAM: block1" + ], + "image_footnote": [], + "bbox": [ + 302, + 143, + 498, + 212 + ], + "page_idx": 1 + }, + { + "type": "image", + "img_path": "images/204ef9c94144ea33b4232596c17feb658e4e9c8f68a10d6be457d30923e823cd.jpg", + "image_caption": [ + "(b) SAM: block3" + ], + "image_footnote": [], + "bbox": [ + 501, + 143, + 700, + 212 + ], + "page_idx": 1 + }, + { + "type": "image", + "img_path": "images/96d05616c1ddd8d2ce65628c8294b6f7496715ceb2a3a51b88ede47b5b1d6e1a.jpg", + "image_caption": [ + "(c) Conventional ViT: block1" + ], + "image_footnote": [], + "bbox": [ + 302, + 222, + 500, + 290 + ], + "page_idx": 1 + }, + { + "type": "image", + "img_path": "images/37cae7ad5b8b0b1513bf72c81b2ac5d94da74722e575c3a118adbc5a4b865407.jpg", + "image_caption": [ + "(d) Conventional ViT: block3", + "Fig. 1: Comparison of activation distributions between image encoder of SAM (a, b) and the conventional ViT (c, d) in two different transformer blocks. Both of these two models are based on the ViT-base version. Each distribution is visualized with the horizontal axis representing different channels and the vertical axis representing the intensity values within that channel." + ], + "image_footnote": [], + "bbox": [ + 501, + 222, + 700, + 290 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "is the recently proposed segment anything model (SAM) [19], representing the prompt-guided visual foundation model with impressive zero-shot capabilities. SAM focuses on the crucial task of segmenting objects of interest from an image, which underpins many vision applications [4, 11, 42, 45].", + "bbox": [ + 212, + 412, + 784, + 474 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "While SAM has demonstrated impressive advancements in segmentation abilities and flexibility, its practical deployment remains challenging. Specifically, SAM's reliance on image encoder of vision transformer (ViT) architecture induces extensive computational overhead that restricts applicability on edge devices with tight latency, power, and memory budgets. This problem necessitates the employment of model compression techniques such as network pruning [13, 49], knowledge distillation [15,26,27,47,48,53], model quantization [5,16,38] and compact architecture design [40,43] etc. Among these techniques, post-training quantization (PTQ) only requires a few unlabeled calibration images without updating model parameters, which enables fast deployment on various devices within a limited time. Given that training and fine-tuning SAM requires significant GPUs using the large-scale training dataset SA1B [19] with 11 million images, there is a growing demand for the PTQ technique.", + "bbox": [ + 212, + 476, + 787, + 672 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Nevertheless, exiting PTQ methods [22, 23, 25, 50] for ViT are focusing on conventional ViT models for classification, which are different from the ViT in SAM. SAM requires accurate prediction for each pixel on multiple zero-shot datasets after billion-scale pretraining, which is much more sensitive to low-bit compression for image embeddings, and its floating-point activations have a highly asymmetric distribution with detrimental outliers. As illustrated in Fig. 1, different from the conventional ViT, our analysis reveals two properties of SAM's activation distribution that pose challenges for quantization: The presence of detrimental outliers implies that the activation values in SAM are not symmetrically distributed around a central tendency. Moreover, SAM's activations often exhibit a concentration of extreme outliers in specific channels, resulting in", + "bbox": [ + 212, + 672, + 787, + 840 + ], + "page_idx": 1 + }, + { + "type": "page_number", + "text": "2", + "bbox": [ + 217, + 114, + 228, + 126 + ], + "page_idx": 1 + }, + { + "type": "header", + "text": "X. Liu et al.", + "bbox": [ + 271, + 114, + 354, + 126 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "a considerable number of channels that require adjustment. While the existing PTQ methods [36, 41], such as RepQ-Vit [23], propose to scale activations at the channel level, the practical challenge in SAM arises in two aspects. Firstly, extreme outliers introduce detrimental scale variance in specific activation channels. Secondly, adjusting an excessive number of channels simultaneously proves to be challenging. These distribution properties of SAM's activations pose new challenges for existing PTQ methods, especially for the ultra-low-bit (e.g., 4-bit, 2-bit) quantization.", + "bbox": [ + 212, + 146, + 787, + 267 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "In this paper, we propose PQ-SAM, the first PTQ method specially designed for SAM, incorporating a grouped activation distribution transformation (GADT) to reshape SAM's activations for quantization-friendly tensor-wise distribution guided by grouped learnable shifting and scaling sizes. Our proposed GADT has a specially designed outlier hierarchical clustering (OHC) scheme to suppress extreme outliers and reduce the total number of learnable parameters in GADT, easing optimization difficulty and preventing overfitting with limited training cost. Specifically, OHC first identifies and truncates tensor-wise extreme outliers across the whole activation tensor by examining interquartile range statistics at the tensor level. This effectively reduces the detrimental scale variance of different channels due to extreme outliers. Secondly, OHC iteratively allocates each pair of learnable shifting and scaling sizes for each group of channels with similar distributions by calculating their quantile range in each iteration. Critically, these shifting and scaling sizes are jointly optimized in an end-to-end manner together with the quantization step sizes for holistic optimal results, under the supervision of the full-precision (FP) SAM model. The optimized shifting and scaling sizes can be reparameterized into corresponding weights so that the proposed PQ-SAM is friendly for practical deployment.", + "bbox": [ + 212, + 271, + 787, + 544 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "We conduct extensive experiments to demonstrate the superiority of our PQ-SAM over existing advanced PTQ methods, on nine representative zero-shot datasets across three widely used segmentation modes of SAM, i.e., point prompt mode, automatic mode, and bounding box prompt mode. Our PQ-SAM preserves SAM's strong zero-shot performance and achieves usable 4-bit PTQ for SAM, which represents an important step towards unlocking the practical edge deployment of SAM.", + "bbox": [ + 212, + 547, + 789, + 654 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "The contributions of this paper are as follows:", + "bbox": [ + 238, + 657, + 573, + 674 + ], + "page_idx": 2 + }, + { + "type": "list", + "sub_type": "text", + "list_items": [ + "- We propose PQ-SAM, the first post-training quantization method customized for the segment anything model with limited computational cost.", + "- We propose a novel grouped activation distribution transformation (GADT) to reshape SAM's activations with learnable shifting and scaling sizes jointly end-to-end optimized together with quantization step sizes.", + "- We design a two-stage outlier hierarchical clustering scheme (OHC) to reduce the learnable parameters and ease the optimization difficulty in GADT.", + "- Extensive experiments demonstrate that PQ-SAM significantly outperforms previous PTQ methods on nine zero-shot datasets." + ], + "bbox": [ + 223, + 690, + 785, + 838 + ], + "page_idx": 2 + }, + { + "type": "header", + "text": "PQ-SAM", + "bbox": [ + 666, + 114, + 730, + 128 + ], + "page_idx": 2 + }, + { + "type": "page_number", + "text": "3", + "bbox": [ + 774, + 116, + 785, + 126 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "2 Related Work", + "text_level": 1, + "bbox": [ + 215, + 143, + 387, + 160 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "2.1 Model Compression for SAM", + "text_level": 1, + "bbox": [ + 215, + 172, + 503, + 186 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "To address the extensive computational requirements of SAM and make SAM suitable for edge devices, recent works MobileSAM [51] and FastSAM [54], EfficientSAM [44] propose different knowledge distillation based technological routes. MobileSAM reduces SAM's heavy image encoder through knowledge distillation. They replace SAM's encoder with a lightweight CNN and distil the heavyweight encoder's knowledge into the CNN. FastSAM avoids SAM's Transformer architecture entirely by adopting a regular convolution neural network (CNN) detector with a segmentation branch and retraining the CNN using 1/50th of SAM's SA1B dataset. EfficientSAM adopts a two-stage training strategy including preraining and finetuning.", + "bbox": [ + 212, + 193, + 785, + 344 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "While MobileSAM, FastSAM, and EfficientSAM can compress SAM models, they still require extensive SA1B data and training resources. In contrast, our proposed PQ-SAM technique only requires one millionth of SA1B data for calibration. With just 100 unlabeled images, PQ-SAM achieves efficient quantization of SAM with minimal GPU costs. PQ-SAM provides a highly practical solution to fast deploying SAM using commodity edge resources. Moreover, our PTQ method can be integrated into these methods for better deployment.", + "bbox": [ + 212, + 345, + 785, + 450 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "2.2 Model Quantization", + "text_level": 1, + "bbox": [ + 215, + 468, + 426, + 484 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Network quantization is an effective compression technique that transforms weight and activation values from floating points to low-bitwidth integers, substantially reducing memory, data movement, and energy costs. Existing methods fall into two categories: Quantization-Aware Training (QAT) [6,16,21] and Post-Training Quantization [10, 16]. However, QAT relies heavily on retraining with the full original dataset, which is infeasible for large vision models like SAM trained on massive data. While PTQ methods using limited calibration data are more practical, current techniques target classification and are not designed for SAM's unique segmentation task. Specifically, SAM exhibits a highly asymmetric distribution with detrimental outliers, which arises from its billion-scale pretraining. Prior PTQ methods for vision transformers [22, 25, 28, 50] do not address these characteristics of the large vision foundation model. To fill this gap, we develop the first customized PTQ approach, PQ-SAM, specifically tailored for quantizing the immense SAM model by accounting for its unusual activation properties.", + "bbox": [ + 212, + 489, + 789, + 702 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "3 Preliminary", + "text_level": 1, + "bbox": [ + 215, + 720, + 367, + 738 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "To provide context before detailing our proposed method, we briefly introduce uniform quantization. We formulate k-bit quantization and dequantization of tensor $x$ as", + "bbox": [ + 212, + 748, + 785, + 792 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\n\\text {Q u a n t}: x _ {q} = \\operatorname {c l i p} \\left(\\operatorname {r o u n d} \\left(\\frac {x}{\\Delta}\\right) + z, 0, 2 ^ {k} - 1\\right), \\tag {1}\n$$\n", + "text_format": "latex", + "bbox": [ + 344, + 797, + 785, + 827 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\nD e q u a n t: x ^ {\\prime} = \\Delta (x _ {q} - z) \\approx x,\n$$\n", + "text_format": "latex", + "bbox": [ + 328, + 827, + 552, + 842 + ], + "page_idx": 3 + }, + { + "type": "page_number", + "text": "4", + "bbox": [ + 217, + 114, + 228, + 126 + ], + "page_idx": 3 + }, + { + "type": "header", + "text": "X. Liu et al.", + "bbox": [ + 271, + 114, + 354, + 126 + ], + "page_idx": 3 + }, + { + "type": "image", + "img_path": "images/c34ea5eab4001cb0647d745609d082d058868384ee01079d66eebac38338aa0f.jpg", + "image_caption": [ + "Fig. 2: The workflow of our proposed PQ-SAM. PQ-SAM introduces grouped activation distribution transformation (GADT) operations into each transformer block of SAM to enable model quantization." + ], + "image_footnote": [], + "bbox": [ + 217, + 142, + 483, + 303 + ], + "page_idx": 4 + }, + { + "type": "image", + "img_path": "images/8e0e94ecd9a29678bcb388a37b7457eb34b576286bb8b6eaff157edde86f24fc.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 490, + 142, + 787, + 304 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "where $\\operatorname{Round}(\\cdot)$ projects a value to the k-bit nearest integer around the input, and $\\operatorname{clip}(x,l,u) = \\min (\\max (x,l),u)$ . $x_{q}$ represents the quantized value, and $x^{\\prime}$ are de-quantized value which approximately equals to $x$ . Note that $\\Delta \\in \\mathbb{R}^{+}$ represents the quantization step size, and $z\\in \\mathbb{Z}$ denotes the zero-point which is a per-tensor offset. The values of $\\Delta$ and $z$ are determined based on the lower and upper bounds of $x$ as follows:", + "bbox": [ + 212, + 371, + 787, + 462 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\n\\Delta = \\frac {\\operatorname* {m a x} (x) - \\operatorname* {m i n} (x)}{2 ^ {k} - 1}, \\quad z = \\operatorname {r o u n d} \\left(- \\frac {\\operatorname* {m i n} (x)}{\\Delta}\\right). \\tag {2}\n$$\n", + "text_format": "latex", + "bbox": [ + 318, + 470, + 785, + 505 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Post-training quantization typically focuses on determining the quantization step sizes $\\varDelta_W$ and $\\varDelta_X$ for weights and activations per layer.", + "bbox": [ + 212, + 513, + 785, + 545 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "4 Grouped Activation Distribution Transformation", + "text_level": 1, + "bbox": [ + 212, + 566, + 723, + 585 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Overview. As shown in Fig. 2, PQ-SAM inserts learnable grouped activation distribution transformation (GADT) operations into each transformer block of SAM to reshape the distribution of activations. The shifting and scaling sizes in GADT are jointly optimized with quantization step sizes in an end-to-end manner under full-precision SAM supervision. This achieves holistically optimized quantization parameters tailored for SAM. Importantly, the learnable ADT is readily absorbed into the preceding linear layer, thus introducing no additional computations or parameters after quantization.", + "bbox": [ + 212, + 597, + 787, + 717 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "4.1 Scaling and Shifting Sizes", + "text_level": 1, + "bbox": [ + 212, + 739, + 475, + 755 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "As shown in Fig. 1, the highly asymmetric activation distribution of SAM is problematic for the per-tensor quantization setting, making it challenging to properly determine the quantization parameters $\\varDelta_W$ and $\\varDelta_X$ . Moreover, the outliers concentrated in specific channels negatively bias $\\varDelta_W$ and $\\varDelta_X$ due to their disproportionate magnitude compared to the majority of activations. To", + "bbox": [ + 212, + 763, + 787, + 840 + ], + "page_idx": 4 + }, + { + "type": "header", + "text": "PQ-SAM", + "bbox": [ + 666, + 114, + 730, + 128 + ], + "page_idx": 4 + }, + { + "type": "page_number", + "text": "5", + "bbox": [ + 774, + 116, + 784, + 126 + ], + "page_idx": 4 + }, + { + "type": "code", + "sub_type": "algorithm", + "code_caption": [ + "Algorithm 1: outlier hierarchical clustering" + ], + "code_body": "Input: FP SAM with activation $X\\in \\mathbb{R}^{T\\times C_1}$ Output: $\\{v_{n},s_{n}\\}_{1\\leq n\\leq N}$ \n1 { Tensor-wise Outlier Truancting.} \n2 Count $X_{Q1}$ $X_{Q3}$ and IQR of $X$ \n3 $X^t\\gets X$ in Eq. 5. \n4 { Channel-wise Outlier Grouping.} \n5 Initialize the iteration number $n = 0$ \n6 Allocated channel Group $C_a = \\emptyset$ \n7 Remaining channel Group $C_r = \\{1,2,\\dots,C_1\\}$ \n8 while $C_r$ is not $\\emptyset$ do \n9 foreach c in $C_r$ do \n10 if $X_{1 - \\alpha}^{t,n - 1}\\leq X^t (c)\\leq X_\\alpha^{t,n - 1}$ then \n11 Allocate $\\{v_n,s_n\\} \\to X(c)$ in Eq. 6; \n12 $C_a.\\mathrm{push}(c);$ \n13 $C_r.\\mathrm{pop}(c);$ \n14 end \n15 end \n16 $n = n + 1$ \n17 end", + "bbox": [ + 217, + 166, + 553, + 438 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "address this, we propose to transform SAM's activations towards a quantization-friendly distribution at the tensor level.", + "bbox": [ + 212, + 470, + 784, + 500 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Straightforwardly, we introduce learnable channel-wise scaling size $S$ and shifting size $V$ to scale and shift the activation tensor on a channel-by-channel basis. For a linear layer, suppose that the input of an activation tensor $X \\in \\mathbb{R}^{T \\times C_1}$ has $T$ tokens and $C_1$ channels, and the weight matrix $W \\in \\mathbb{R}^{C_1 \\times C_2}$ has a bias $B \\in \\mathbb{R}^{1 \\times C_2}$ . We utilize scaling size $S \\in \\mathbb{R}^{1 \\times C_1}$ and shifting size $V \\in \\mathbb{R}^{1 \\times C_1}$ to reshape its distribution, which is formulated as", + "bbox": [ + 212, + 501, + 785, + 590 + ], + "page_idx": 5 + }, + { + "type": "equation", + "text": "\n$$\n\\tilde {X} = (X - V) \\oslash S, \\tag {3}\n$$\n", + "text_format": "latex", + "bbox": [ + 431, + 598, + 785, + 616 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "where $\\widetilde{X}$ represents the reshaped activation tensor and $\\varnothing$ represents the division operation. This process is mathematically equivalent to the original linear layer by updating the weight matrix:", + "bbox": [ + 212, + 625, + 784, + 672 + ], + "page_idx": 5 + }, + { + "type": "equation", + "text": "\n$$\nX W + B = \\widetilde {X} \\left(W ^ {\\top} \\odot S ^ {\\top}\\right) + \\left(V W ^ {\\top} + B\\right), \\tag {4}\n$$\n", + "text_format": "latex", + "bbox": [ + 344, + 678, + 785, + 698 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "where $\\odot$ represents the multiplication operation, and the updated linear layer has a weight matrix $\\widetilde{W} = W^{\\top} \\odot S^{\\top}$ and a bias $\\widetilde{B} = (VW^{\\top} + B)$ .", + "bbox": [ + 212, + 704, + 784, + 733 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Given the practical challenge of optimizing SAM's large number of activation channels, it is difficult to simultaneously optimize the shifting and scaling sizes at the channel level. To address this, we propose Grouped Activation Distribution Transformation (GADT). This technique utilizes our outlier hierarchical clustering scheme to group channels with similar distributions together. By sharing a set of learnable parameters within each group, we effectively reduce the optimization difficulty.", + "bbox": [ + 212, + 734, + 785, + 839 + ], + "page_idx": 5 + }, + { + "type": "page_number", + "text": "6", + "bbox": [ + 217, + 114, + 228, + 126 + ], + "page_idx": 5 + }, + { + "type": "header", + "text": "X. Liu et al.", + "bbox": [ + 271, + 114, + 354, + 127 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "4.2 Outlier Hierarchical Clustering", + "text_level": 1, + "bbox": [ + 215, + 146, + 519, + 161 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "The above-mentioned shifting and scaling sizes are one-on-one corresponding to each activation channel, which leads to excessive learned parameters being optimized and increases optimization difficulty within limited training cost. To reduce learning difficulty and prevent overfitting of the learnable shifting and scaling sizes, we propose an outlier hierarchical clustering (OHC) scheme to iteratively allocate a pair of shifting and scaling sizes to a group of channels. OHC consists of a tensor-wise outlier truncating stage and a channel-wise outlier grouping stage.", + "bbox": [ + 212, + 169, + 787, + 289 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Tensor-wise Outlier Truncating. Due to the influence of extreme outliers in the activation values, the grouping of channels with similar distributions can be affected. Additionally, these extreme anomalies can impact the effective usage of quantization intervals. Therefore, as a preliminary step, we perform outlier truncation at the tensor level by examining the outlier situation of numerical values across the entire tensor layer.", + "bbox": [ + 212, + 289, + 787, + 378 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Motivated by the statistic theory [34,39,55], we design an effective preprocessing scheme. For the activation tensor $X$ , we count the upper and lower quartile fractions $Q_{3}$ and $Q_{1}$ , and the interquartile range $IQR$ of its numerical distribution. We then introduce adjustable statistical thresholds, and the truncated activation $X^{t}$ is formulated as", + "bbox": [ + 212, + 378, + 787, + 455 + ], + "page_idx": 6 + }, + { + "type": "equation", + "text": "\n$$\nX ^ {t} = \\left\\{X \\mid X _ {Q _ {1}} - \\lambda I Q R < x < X _ {Q _ {3}} + \\lambda I Q R, x \\in X \\right\\}, \\tag {5}\n$$\n", + "text_format": "latex", + "bbox": [ + 316, + 460, + 785, + 486 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "where $\\lambda$ is a hyper-parameter, and $X_{Q_3}$ and $X_{Q_1}$ are the upper and lower quartile values at $Q_3$ and $Q_1$ , respectively.", + "bbox": [ + 212, + 502, + 785, + 532 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Channel-wise Outlier Grouping. For a truncated activation tensor $X_{t}$ consisting of $C$ channels, we allocate $N$ pairs of shifting and scaling sizes $\\{v_{n}, s_{n}\\}_{1 \\leq n \\leq N}$ to the channels, and $N \\ll C$ . As depicted in Fig. 2 (b), we perform this allocation iteratively, assigning a pair of shifting and scaling sizes $v_{n}, s_{n}$ to a group of channels that meet the pre-defined range between the upper quantile $\\alpha$ and lower quantile $1 - \\alpha$ of the non-allocating channels. In the $n^{th}$ iteration, we formulate the procedure of $\\{v_{n}, s_{n}\\}$ for activations of the $c^{th}$ channel $X(c)$ as", + "bbox": [ + 212, + 532, + 800, + 638 + ], + "page_idx": 6 + }, + { + "type": "equation", + "text": "\n$$\n\\widetilde {X} (c) = \\left(X (c) - v _ {n}\\right) \\oslash s _ {n}, \\text {i f} X _ {1 - \\alpha} ^ {t, n - 1} \\leq X ^ {t} (c) \\leq X _ {\\alpha} ^ {t, n - 1}, \\tag {6}\n$$\n", + "text_format": "latex", + "bbox": [ + 305, + 645, + 785, + 664 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "where $\\widetilde{X}(c)$ represents the $c^{th}$ channel of reshaped activation $\\widetilde{X}$ . Additionally, $X_{1 - \\alpha}^{t,n-1}$ and $X_{\\alpha}^{t,n-1}$ respectively denote the lower and upper quantile values, of the activations from these channels that are not allocated in previous iterations.", + "bbox": [ + 212, + 672, + 787, + 718 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Overall, the procedure of the proposed outlier hierarchical clustering is summarized in Algorithm 1. This iterative allocation process allows for the dynamic calculation of percentile values for the unallocated channels. By using a fixed value of $\\alpha$ , we ensure that more similar channels are grouped together at the beginning of the allocation process. As the allocation process advances, the groups gradually become more refined, resulting in finer granularity. This adaptive approach ultimately yields highly effective grouping outcomes, which prove advantageous for subsequent learning and optimization. By conducting the joint", + "bbox": [ + 212, + 719, + 787, + 840 + ], + "page_idx": 6 + }, + { + "type": "header", + "text": "PQ-SAM", + "bbox": [ + 666, + 114, + 730, + 128 + ], + "page_idx": 6 + }, + { + "type": "page_number", + "text": "7", + "bbox": [ + 774, + 116, + 785, + 126 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "To optimize of GADT and quantization step sizes, PQ-SAM quantizes SAM with limited overheads on unlabeled samples.", + "bbox": [ + 212, + 146, + 782, + 176 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "4.3 Joint Optimization with Quantization Step Sizes", + "text_level": 1, + "bbox": [ + 214, + 198, + 661, + 214 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "The shifting and scaling sizes in GADT are learnable and optimized jointly with $\\Delta_W$ and $\\Delta_X$ in our PQ-SAM. This integration of learnable shifting and scaling sizes allows for a more effective adjustment of the activation distribution. We use FP model supervision to learn the quantization step sizes $\\Delta_W$ and $\\Delta_X$ in an end-to-end manner, and adjust offsets via the above-mentioned GADT, which not only achieves channel-wise offset adjustment but also preserves the hardware-friendly per-tensor quantization setting.", + "bbox": [ + 212, + 222, + 784, + 328 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Given the floating-point weight tensor $W$ and activation tensor $X$ in one layer of SAM, we follow the Eq. 1 and quantize them to $k$ -bit integer values $W_{q}$ and $X_{q}$ with the corresponding weight and activation quantization step sizes $\\Delta_W$ and $\\Delta_X$ , respectively. Our optimization framework goes beyond previous approaches [23,36,41] that focus on determining $\\Delta_W$ and $\\Delta_X$ for each layer individually. Instead, our framework enables a holistic, FP-supervised quantization process that is customized for the entire model.", + "bbox": [ + 214, + 329, + 785, + 433 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "We jointly optimize the shifting and scaling sizes $\\{v_{n}, s_{n}\\}_{1 \\leq n \\leq N}$ , and quantization step sizes $\\Delta_{W}$ and $\\Delta_{X}$ for holistic optimal results. By leveraging the output $O(W, K)$ of the FP model as a guide, we can tailor the quantization steps specifically for the pre-trained weights and activations by minimizing the Mean Squared Error (MSE) loss between $O(W, X)$ and the output $O(\\widetilde{W}_q, \\widetilde{X}_q)$ of quantized model:", + "bbox": [ + 212, + 434, + 784, + 525 + ], + "page_idx": 7 + }, + { + "type": "equation", + "text": "\n$$\n\\underset {\\Delta_ {W}, \\Delta_ {X}, \\left\\{v _ {n}, s _ {n} \\right\\} _ {1 \\leq n \\leq N}} {\\arg \\min } \\| O (\\widetilde {W} _ {q}, \\widetilde {X} _ {q}) - O (W, X) \\| _ {2}, \\tag {7}\n$$\n", + "text_format": "latex", + "bbox": [ + 334, + 535, + 785, + 566 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "where $\\widetilde{W}_q$ represent the corresponding quantized weight of the updated weight by reparameterization as demonstrated in Eq. 4, and $\\widetilde{X}_q$ represents the quantized activation which has been reshaped.", + "bbox": [ + 212, + 577, + 782, + 626 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "5 Experiment", + "text_level": 1, + "bbox": [ + 215, + 648, + 366, + 666 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "5.1 Datasets and Metrics", + "text_level": 1, + "bbox": [ + 214, + 679, + 439, + 694 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "We conduct evaluations on nine representative zero-shot segmentation datasets, focusing on three widely used segmentation modes of SAM, i.e., the point prompt mode, the automatic mode, and the bounding box prompt mode. These datasets have been officially verified by SAM as mentioned in the work [19]. However, since there is no publicly available test code provided by SAM, we independently evaluate these datasets based on our own resources.", + "bbox": [ + 212, + 704, + 784, + 792 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Point Prompt Mode. The point prompt specifies what to segment in the image, with the goal of determining the mask containing that point. To fully validate the zero-shot performance of quantized SAM, we subsample six datasets", + "bbox": [ + 212, + 794, + 785, + 840 + ], + "page_idx": 7 + }, + { + "type": "page_number", + "text": "8", + "bbox": [ + 217, + 114, + 228, + 126 + ], + "page_idx": 7 + }, + { + "type": "header", + "text": "X. Liu et al.", + "bbox": [ + 271, + 114, + 354, + 127 + ], + "page_idx": 7 + }, + { + "type": "table", + "img_path": "images/bd9238fb990241f68ef05c9a3700be03d09f73f0353cd3cbd0a1569f3f9df32c.jpg", + "table_caption": [ + "Table 1: Quantitative comparison of different PTQ methods for the point prompt mode of SAM." + ], + "table_footnote": [], + "table_body": "
MethodBitBBBC038V1NDD20CityscapeDOORSiShapeNDISPark
IoUDiceIoUDiceIoUDiceIoUDiceIoUDiceIoUDice
FP320.77950.85000.84680.91000.59450.70610.84960.91260.37220.51080.81920.8938
ViT-H
MinMax [16]60.34850.45990.53940.66050.33280.43740.36780.50820.25900.38630.60950.7091
MSE [5]60.65980.73930.55110.63750.44230.55180.61150.70760.34090.48180.71360.8046
Percentile [20]60.72190.72900.76680.84940.50530.61550.78710.86170.36790.50760.78500.8672
FQ-ViT [25]60.44440.60080.25130.38050.28360.41860.28670.42980.10880.19260.41370.5715
RepQ-ViT [23]60.71740.79380.82230.89390.55980.67850.83830.90480.34440.48890.80100.8838
Ours60.76460.83760.82250.89330.56970.68610.84510.90920.33800.47760.81570.8919
MinMax [16]40.33090.46380.27180.40500.19200.29660.46340.61440.11330.20070.25150.3735
MSE [5]40.37610.51410.36460.51870.22650.33940.42650.58350.14450.24730.39330.5408
Percentile [20]40.37350.52000.34470.48110.26190.38910.38880.52730.12530.21690.41530.5657
FQ-ViT [25]40.22130.34840.09720.16360.14530.23870.11680.19830.03590.06780.21030.3341
RepQ-ViT [23]40.70130.78090.76850.85500.45900.57400.75900.84930.27360.40650.73990.8325
Ours40.70810.79730.78180.86700.50560.62840.79750.87660.29210.42940.76470.8541
FP320.77610.84760.84680.90750.59390.70680.85140.91380.34790.48560.81270.8893
ViT-L
MinMax [16]60.51290.62850.52620.63870.36810.48380.41960.56120.28210.41440.64900.7532
MSE [5]60.70100.77640.65210.74080.47570.58880.68570.78080.35440.49640.75830.8489
Percentile [20]60.73680.80660.75110.83200.52300.63640.78050.85740.37750.51950.79400.8759
FQ-ViT [25]60.42130.56600.35960.51430.28400.42060.38400.54530.12050.21130.40980.5675
RepQ-ViT [23]60.69520.77150.80040.87670.55800.67270.82700.89400.31050.44860.78680.8712
Ours60.74250.81540.81800.89020.55660.67350.83700.90360.31990.45930.79900.8786
MinMax [16]40.12560.20530.27640.41610.12850.20270.41510.57750.09480.17070.18180.2868
MSE [5]40.28360.40420.33040.47470.18410.27660.33410.49010.15540.26430.36410.4994
Percentile [20]40.41360.56370.30740.45310.27200.40530.37010.52590.13550.23390.40840.5631
FQ-ViT [25]40.27910.42380.16690.27470.17940.29030.18130.29790.07460.13760.28960.4406
RepQ-ViT [23]40.41290.49830.72350.82500.44220.56220.75520.84830.27910.41030.71000.8118
Ours40.68550.77270.73930.83390.48780.61140.78900.87120.25200.38330.73790.8332
FP320.76660.84170.83220.90000.56920.68840.82940.89870.35360.49870.79330.8760
ViT-B
MinMax [16]60.38290.47770.57270.69760.33560.44450.34950.47970.26850.40450.59040.7021
MSE [5]60.63250.72550.57670.69210.39760.51630.50300.63460.29430.43460.62940.7398
Percentile [20]60.72690.79970.73640.82600.51990.64060.79520.87470.36050.50680.76760.8564
FQ-ViT [25]60.61540.71520.47830.60100.38260.52170.35380.48590.20830.33100.58460.7196
RepQ-ViT [23]60.69000.76750.80700.87300.54490.66440.80530.88010.34390.48560.78360.8696
Ours60.75830.83310.80080.87820.53910.65930.80590.88860.34410.48860.78450.8705
MinMax [16]40.05560.09860.15330.24130.08970.13740.13970.22520.04470.08420.05890.0964
MSE [5]40.33590.46120.37560.53150.22780.33430.46350.62630.13300.23060.36370.5018
Percentile [20]40.39270.53810.40540.55830.27280.40340.42190.57610.13560.23450.41490.5564
FQ-ViT [25]40.34760.50100.16690.27470.21010.32920.22840.35910.11340.19950.31190.4647
RepQ-ViT [23]40.47380.56590.64790.76080.43380.56530.58780.71180.30640.44660.62530.7448
Ours40.70800.79690.67680.78370.47140.59860.73750.83610.31060.45280.68000.7983
", + "bbox": [ + 222, + 176, + 779, + 520 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "including BBBC038V1 [3], NDD20 [37], Cityscape [9], DOORS [32], iShape [46], and NDISPark [8]. These datasets contain 15,213 images and 29,978 masks and cover a broad range of domains, which is used for the very comprehensive evaluation experiment. Following the original SAM's experiments, we determine the single-point prompt by selecting the \"center\" of the ground truth. This center is identified as the point with the highest value in the mask's interior distance transform. We calculate the mean IoU and Dice [35] for all masks of each dataset as quantitative metrics, which are common metrics to quantify the overlap between predicted and ground truth masks.", + "bbox": [ + 212, + 535, + 784, + 671 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "Automatic Mode. SAM's automatic mask generation pipeline can automatically extract a series of masks from an input image with a regular grid of foreground points. We evaluate this automatic mode on the zero-shot edge detection task using the BSDS500 [30] dataset. These automatically generated masks are processed into object edges by the post-processing $^{1}$ . We conduct evaluations on the test subset, consisting of 200 images, and assess the performance of all methods using four standard metrics $^{2}$ for edge detection [1]: optimal dataset", + "bbox": [ + 212, + 672, + 785, + 780 + ], + "page_idx": 8 + }, + { + "type": "header", + "text": "PQ-SAM", + "bbox": [ + 666, + 114, + 730, + 128 + ], + "page_idx": 8 + }, + { + "type": "page_number", + "text": "9", + "bbox": [ + 774, + 116, + 784, + 126 + ], + "page_idx": 8 + }, + { + "type": "page_footnote", + "text": "1 Our experiments are based on the descriptions in SAM's paper since the official test code is not provided to the public.", + "bbox": [ + 217, + 797, + 782, + 824 + ], + "page_idx": 8 + }, + { + "type": "page_footnote", + "text": "$^{2}$ https://www2.eecs.berkeley.edu/Research/Projects/CS/vision/bsds/", + "bbox": [ + 218, + 824, + 687, + 839 + ], + "page_idx": 8 + }, + { + "type": "image", + "img_path": "images/ff09578659fe2a3db7d0f84b8d59ff6ef34ca858445b4cfbc222ec61a1d100e7.jpg", + "image_caption": [ + "Fig. 3: Visual comparison of different PTQ methods for SAM's point prompt mode. The red block and green star represent the object to be segmented and the position of the prompt point, respectively." + ], + "image_footnote": [], + "bbox": [ + 215, + 143, + 785, + 306 + ], + "page_idx": 9 + }, + { + "type": "text", + "text": "scale (ODS), optimal image scale (OIS), average precision (AP), and recall at $50\\%$ precision (R50).", + "bbox": [ + 212, + 359, + 782, + 388 + ], + "page_idx": 9 + }, + { + "type": "text", + "text": "Bounding Box Prompt Mode. Similar to the point prompt mode, the bounding box prompt mode of SAM focuses on segmenting the objects circled by the bounding box. We follow the original setting of SAM [19] and conduct experiments on the COCO [24] and LVIS [12] datasets. The results are evaluated by average precision (AP).", + "bbox": [ + 212, + 390, + 784, + 465 + ], + "page_idx": 9 + }, + { + "type": "text", + "text": "5.2 Implementation Details", + "text_level": 1, + "bbox": [ + 215, + 484, + 455, + 500 + ], + "page_idx": 9 + }, + { + "type": "text", + "text": "Training Details. We utilize the PyTorch [31] and conduct all of our experiments on NVIDIA Tesla V100 GPUs. To calibrate our model, we randomly select 100 samples from the SA1B dataset as calibration data and perform the calibration process over 100 epochs, using a batch size of 1. The learnable quantization parameters $\\Delta_W$ , $\\Delta_X$ , and $\\{v_n, s_n\\}_{1 \\ll n \\ll N}$ , are optimized using the Adam optimizer [18] with $\\beta_1 = 0.9$ and $\\beta_2 = 0.999$ , with a learning rate of $1e - 4$ . The learning rate scheduler is CosineAnnealingLR [29]. The hyperparameters $\\lambda$ and $\\alpha$ are set as 10 and 0.25, respectively.", + "bbox": [ + 212, + 507, + 785, + 628 + ], + "page_idx": 9 + }, + { + "type": "text", + "text": "Experimental Settings. In our evaluation experiments, we focus on challenging low-bit-width quantization settings, specifically W4A4 (4-bit quantization for weight and activation) and W6A6 (6-bit quantization), for the point prompt mode of SAM. It is important to note that the automatic mode of SAM, which is used for the edge detection task, is less sensitive to network prediction errors due to the incorporation of various post-processing algorithms. Therefore, we specifically perform 4-bit quantization experiments on the BSDS500 dataset for the automatic mode of SAM. Additionally, to validate the superiority of our PQ-SAM approach as discussed in Section 5.3, we conduct experiments in an ultra-low 2-bit setting specifically for the ViT-B version of SAM. This ultra-low-bit setting allowed us to assess the performance and effectiveness of PQ-SAM in scenarios with extremely limited precision.", + "bbox": [ + 212, + 628, + 785, + 809 + ], + "page_idx": 9 + }, + { + "type": "text", + "text": "Baseline Methods. In our comparative analysis, we include two categories of PTQ methods as baselines. The first category consists of classical post-training", + "bbox": [ + 212, + 809, + 785, + 839 + ], + "page_idx": 9 + }, + { + "type": "page_number", + "text": "10", + "bbox": [ + 217, + 114, + 235, + 126 + ], + "page_idx": 9 + }, + { + "type": "header", + "text": "X. Liu et al.", + "bbox": [ + 271, + 114, + 354, + 126 + ], + "page_idx": 9 + }, + { + "type": "table", + "img_path": "images/89fbc54915f6c2963b3af8b8a17775f1210b6e9baf3b6d6db1650d007c0b371f.jpg", + "table_caption": [ + "Table 2: Quantitative comparison for the automatic mode of SAM." + ], + "table_footnote": [], + "table_body": "
H-LVMethodBitODSOISAPR50
FP320.7540.7700.7290.865
MinMax [16]40.4300.4400.304-
MSE [5]40.4540.4680.3880.049
Percentile [20]40.4430.4570.339-
FQ-ViT [25]40.4120.4160.230-
RepQ-ViT [23]40.7060.7200.6470.801
Ours40.7360.7450.7070.859
T-LVFP320.7540.7730.7260.866
MinMax [16]40.4230.4240.247-
MSE [5]40.4420.4560.332-
Percentile [20]40.4440.4550.320-
RepQ-ViT [23]40.7090.7220.6450.769
FQ-ViT [25]40.4240.4290.255-
Ours40.7260.7350.6690.843
B-LVFP320.7450.7550.6900.871
MinMax [16]40.3560.3570.145-
MSE [5]40.4250.4800.4750.374
Percentile [20]40.4490.4670.4000.024
FQ-ViT [25]4----
RepQ-ViT [23]40.7010.7120.6320.806
Ours40.7060.7100.6420.861
", + "bbox": [ + 272, + 162, + 728, + 378 + ], + "page_idx": 10 + }, + { + "type": "text", + "text": "methods, namely MSE [5] and MinMax [16]. These methods are versatile and can be easily applied to different models. The second category comprises quantization methods specifically designed for Vision Transformer (ViT) models, including two state-of-the-art methods RepQ-ViT [23] and FQ-ViT [25]. These methods have gained popularity in post-training quantization and are directly relevant to our evaluation. Specifically, RepQ-ViT is designed for the post-LayerNorm activations with severe inter-channel variation but based on a hand-craft heuristic algorithm, which is not as flexible and effective as our learning-based method.", + "bbox": [ + 212, + 404, + 784, + 525 + ], + "page_idx": 10 + }, + { + "type": "text", + "text": "5.3 Experimental Results", + "text_level": 1, + "bbox": [ + 215, + 551, + 439, + 566 + ], + "page_idx": 10 + }, + { + "type": "text", + "text": "Point Prompt Mode. In Fig. 3, we present the quantitative results of different post-quantization methods for the SAM's point prompt mode. Overall, our PQ-SAM consistently outperforms existing methods across most datasets, which proves that our method can better maintain the strong generalization capacity of SAM. For the 6-bit quantization, PQ-SAM exhibits a performance gap to the full-precision model of less than 0.04 in terms of IoU and Dice metrics across most of the datasets. While PQ-SAM may exhibit slightly lower performance compared to strong baseline methods on a few datasets, it still demonstrates competitive results and preserves the generalization ability overall. For the 4-bit quantization, our method significantly surpasses the baseline methods across all datasets. For instance, PQ-SAM surpasses the second-best method by over $40\\%$ in terms of IoU and Dice metrics on the BBBC038V1 dataset for the SAM's ViT-L and ViT-B versions.", + "bbox": [ + 212, + 582, + 784, + 777 + ], + "page_idx": 10 + }, + { + "type": "text", + "text": "Automatic Mode. In Table 2, we present the quantitative results of different post-quantization methods for the SAM's automatic mode on the BSDS500 dataset. Our PQ-SAM surpasses the baseline methods across three SAM versions and archives a small performance gap to the FP model.", + "bbox": [ + 212, + 779, + 785, + 839 + ], + "page_idx": 10 + }, + { + "type": "header", + "text": "PQ-SAM", + "bbox": [ + 666, + 114, + 730, + 128 + ], + "page_idx": 10 + }, + { + "type": "page_number", + "text": "11", + "bbox": [ + 767, + 116, + 782, + 126 + ], + "page_idx": 10 + }, + { + "type": "image", + "img_path": "images/7d1ef5aa672446b279bf9fa865e446ad29cc0530d00cf917f25edef7176c9caf.jpg", + "image_caption": [ + "Fig. 4: Visual comparison of different PTQ methods for SAM's automatic mode. The blue circle indicates segmentation errors." + ], + "image_footnote": [], + "bbox": [ + 217, + 143, + 308, + 226 + ], + "page_idx": 11 + }, + { + "type": "image", + "img_path": "images/b2527f7c3411574f300f143e5a807592730fe1e159e79ff8ff84667eee6233cb.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 310, + 143, + 401, + 227 + ], + "page_idx": 11 + }, + { + "type": "image", + "img_path": "images/dd9ecf85e9720b0c641c22bc081c4960acb4fb2cad1b356bbf88b41d8af64ed9.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 406, + 143, + 496, + 227 + ], + "page_idx": 11 + }, + { + "type": "image", + "img_path": "images/1797122d2348301fef5f9bcbe3b94c1258832105d7abe2bec78715b07c119f2c.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 501, + 143, + 596, + 227 + ], + "page_idx": 11 + }, + { + "type": "image", + "img_path": "images/ed37b285a43dee42e0370b4bca16fcf77d7903b766db60c06809fa4de2a6f2ff.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 598, + 143, + 691, + 227 + ], + "page_idx": 11 + }, + { + "type": "image", + "img_path": "images/8695153b4341028ce561c947e97ef6787b7d512f2ca8dfcb8242bd2d8b359109.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 696, + 145, + 784, + 227 + ], + "page_idx": 11 + }, + { + "type": "table", + "img_path": "images/735acb3b91c63f6efdc01993a8f6e72da3db6434e673c0bcf8a69125bb44ae96.jpg", + "table_caption": [ + "Table 3: Quantitative Results for the bounding box prompt mode of SAM." + ], + "table_footnote": [], + "table_body": "
MethodBitCOCOLVIS
APAPSAPMAPLAPAPSAPMAPL
HVT-HFP3246.430.751.061.744.432.157.265.2
RepQ-Vit [23]643.527.648.159.242.430.054.963.8
Ours645.129.649.560.243.431.555.764.0
RepQ-Vit [23]438.224.741.752.536.025.546.554.1
Ours441.126.245.355.938.127.149.156.9
", + "bbox": [ + 243, + 286, + 759, + 375 + ], + "page_idx": 11 + }, + { + "type": "table", + "img_path": "images/f61db4dfa1ce6325cf40768cad24fa28659017288e5b411fbc2a1cba385a3e4e.jpg", + "table_caption": [ + "Table 4: Quantitative comparison for the ultra-low 2bit quantization setting." + ], + "table_footnote": [], + "table_body": "
MethodNDD20DOORSNDISPark
IoUDiceIoUDiceIoUDice
RepQ-ViT [23]0.26810.41460.36770.53010.27800.4224
Ours0.29410.44800.40990.56970.42130.5812
", + "bbox": [ + 277, + 401, + 725, + 454 + ], + "page_idx": 11 + }, + { + "type": "text", + "text": "Bounding Box Prompt Mode. In Table 3, we provide the quantitative results of different post-quantization methods for the SAM's bounding box prompt on the COCO and LVIS datasets. Our PQ-SAM outperforms the baseline methods and has a small performance gap compared to the FP model. In the 6-bit quantization setting, PQ-SAM exhibits a loss of accuracy within $3\\%$ compared to the FP model. Moreover, in the 4-bit quantization configuration, PQ-SAM shows a significant improvement of $6\\%$ compared to the RepQ-Vit method.", + "bbox": [ + 212, + 464, + 785, + 571 + ], + "page_idx": 11 + }, + { + "type": "text", + "text": "Ultra-low Bit Setting. As shown in Table 4, we validate the superiority of our PQ-SAM on the ultra-low 2-bit quantization setting, which significantly surpasses the state-of-the-art post-training quantization (PTQ) method RepQ-ViT. For instance, our method improves Dice by over $35\\%$ compared to RepQ-ViT, demonstrating the effectiveness of our approach for the ultra-low-bit quantization of SAM models.", + "bbox": [ + 212, + 571, + 787, + 661 + ], + "page_idx": 11 + }, + { + "type": "text", + "text": "Visualization. We provide the visual results of quantized SAM using different quantization methods for the W4A4 setting, including the point prompt mode in Fig. 3 and the auto mode in Fig. 4, respectively. Compared with existing PTQ methods, our proposed method could achieve better segmentation performance with more accurate segmentation masks and fewer segmentation errors. We further give more qualitative results in the supplementary material.", + "bbox": [ + 212, + 662, + 787, + 753 + ], + "page_idx": 11 + }, + { + "type": "text", + "text": "5.4 Ablation Studies", + "text_level": 1, + "bbox": [ + 215, + 780, + 401, + 794 + ], + "page_idx": 11 + }, + { + "type": "text", + "text": "We perform ablation studies to evaluate the impacts of each component in the proposed PQ-SAM approach using SAM's ViT-B on the BBBC039V1 dataset.", + "bbox": [ + 212, + 809, + 785, + 839 + ], + "page_idx": 11 + }, + { + "type": "page_number", + "text": "12", + "bbox": [ + 217, + 114, + 235, + 126 + ], + "page_idx": 11 + }, + { + "type": "header", + "text": "X. Liu et al.", + "bbox": [ + 271, + 114, + 354, + 127 + ], + "page_idx": 11 + }, + { + "type": "table", + "img_path": "images/10ebe85cdc3e41894d6acc2e6f900cdf334adef16e84a7c80160ee11ef60302c.jpg", + "table_caption": [ + "Table 5: Ablation study on learnable quantization parameters. $\\delta_W$ and $\\delta_X$ represent the weight and activation quantization step sizes, respectively. $\\{v_n,s_n\\}_{1\\leq n\\leq N}$ represents the allocated N pairs of shifting and scaling sizes." + ], + "table_footnote": [], + "table_body": "
ΔW,ΔX{vn}1≤n≤N{sn}1≤n≤NIoUDice
xxx0.05560.0986
x✓✓0.15150.2339
✓xx0.63420.7334
✓✓x0.68470.7721
✓✓✓0.70800.7969
", + "bbox": [ + 302, + 193, + 702, + 253 + ], + "page_idx": 12 + }, + { + "type": "table", + "img_path": "images/01af2aea8995f7c18cbb01d13d6a34d6d9fe77bff41b44d7299290a80712703f.jpg", + "table_caption": [ + "Table 6: Ablation study on activation distribution transformation (ADT) with the outlier hierarchical clustering (OHC). 'Truncating' represents the tensor-level outlier truncating operation. 'Even Groupsize' represents that we evenly allocate the same number of learnable shifting and scaling sizes as ours to each group of adjacent channels. 'Per-Channel' represents that we allocate each pair of learnable shifting and scaling sizes to each channel." + ], + "table_footnote": [], + "table_body": "
ADT | Truncating | Even Groupwise | Per-Channel | OHC (ours) | IoU | Dice
xxxxx0.63420.7334
x✓xxx0.65330.7602
✓✓x✓x0.68420.7720
✓✓✓xx0.68170.7797
✓✓xx✓0.70800.7969
", + "bbox": [ + 259, + 359, + 743, + 420 + ], + "page_idx": 12 + }, + { + "type": "text", + "text": "Learnable Quantization Parameters. As presented in Table 5, we perform an ablation study on the learnable quantization parameters in our PQ-SAM framework. This involves considering the weight and activation quantization step sizes, denoted as $\\delta_W$ and $\\delta_X$ respectively, as well as allocating a set of $N$ pairs of shifting and scaling sizes $\\{v_n,s_n\\}_{1\\leq n\\leq N}$ .", + "bbox": [ + 212, + 441, + 787, + 518 + ], + "page_idx": 12 + }, + { + "type": "text", + "text": "The results demonstrate that introducing learnable quantization step sizes, shifting sizes, or scaling sizes individually resulted in significantly improved performance compared to fixed quantization schemes. Moreover, it is observed that the learnable shifting and scaling sizes yielded even better quantization performance than the learnable quantization step sizes. By optimizing all of these quantization parameters during training, the model is able to effectively adapt to the underlying data distribution and achieve more accurate quantization.", + "bbox": [ + 212, + 520, + 787, + 626 + ], + "page_idx": 12 + }, + { + "type": "text", + "text": "Activation Distribution Transformation. In Table 6, we present the results of an ablation study to validate the effectiveness of the activation distribution transformation (ADT) in combination with the outlier hierarchical clustering (OHC) scheme. The ADT technique significantly improves the quantization performance, and the OHC scheme further enhances the optimization process and leads to additional improvements in quantization performance. Meanwhile, we validate that the tensor-level outlier truncating operation in ADT can significantly improve the quantization performance.", + "bbox": [ + 212, + 627, + 787, + 747 + ], + "page_idx": 12 + }, + { + "type": "text", + "text": "To validate the effectiveness of the outlier hierarchical clustering (OHC) scheme, we compare it with two alternative schemes: the \"Even Groupsize\" and \"Per-Channel\" schemes. In the \"Even Groupsize\" scheme, we allocate the same number of learnable shifting and scaling sizes to each group of adjacent channels. In the \"Per-Channel\" scheme, we allocate a unique pair of learnable shifting and scaling sizes to each individual channel, resulting in the same number of learn-", + "bbox": [ + 212, + 750, + 787, + 839 + ], + "page_idx": 12 + }, + { + "type": "header", + "text": "PQ-SAM", + "bbox": [ + 666, + 114, + 730, + 128 + ], + "page_idx": 12 + }, + { + "type": "page_number", + "text": "13", + "bbox": [ + 767, + 114, + 785, + 126 + ], + "page_idx": 12 + }, + { + "type": "table", + "img_path": "images/8d9f72073fef43cc03e0e4d93a7cd6191a65f2429164aa379776ef4d78c1859e.jpg", + "table_caption": [ + "Table 7: Ablation study on hyper-parameters and amounts of calibration data." + ], + "table_footnote": [], + "table_body": "
(a) Hyper-parameters of PQ-SAM.
λ61014
IoU0.69010.70800.6858
Dice0.78940.79690.7871
α0.200.250.30
IoU0.68840.70800.6841
Dice0.78120.79690.7693
", + "bbox": [ + 246, + 165, + 470, + 252 + ], + "page_idx": 13 + }, + { + "type": "table", + "img_path": "images/5a40c494d4ec1dd76c62ee62a8453c3bef167e4b677804bc6a66e539b2d14717.jpg", + "table_caption": [ + "(b) Amount of calibration data." + ], + "table_footnote": [], + "table_body": "
Number of data100200500
IoU0.70800.70910.7120
Dice0.79690.79800.8014
", + "bbox": [ + 521, + 188, + 759, + 227 + ], + "page_idx": 13 + }, + { + "type": "text", + "text": "able shifting and scaling size pairs as there are activation channels. The results of our ablation study demonstrate that the activation distribution transformation (ADT) with OHC in the PQ-SAM framework outperforms those with the \"Even Groupsize\" or \"Per-Channel\" schemes.", + "bbox": [ + 212, + 263, + 782, + 323 + ], + "page_idx": 13 + }, + { + "type": "text", + "text": "Hyperparameter $\\lambda$ and $\\alpha$ . We conduct ablation studies with extreme parameter settings to showcase the impact of hyperparameters $\\lambda$ and $\\alpha$ on performance, as shown in Table 7a. The first set of experiments focuses on evaluating the impact of the hyperparameter $\\lambda$ on tensor-level outlier truncation. Different values of $\\lambda$ are tested for various truncated ranges. The results demonstrate the effects of varying $\\lambda$ on the outlier truncation process and its subsequent influence on the overall performance of the PQ-SAM framework. The second set of experiments involves an ablation study on different upper quantiles $\\alpha$ for the proposed channel-wise outlier grouping in the outlier hierarchical clustering step. By testing different values of $\\alpha$ , different numbers of learnable shifting and scaling sizes are allocated to each group of channels.", + "bbox": [ + 212, + 324, + 784, + 489 + ], + "page_idx": 13 + }, + { + "type": "text", + "text": "Amount of Calibration Data. We conduct an ablation study on the amount of calibration data. The results in Table 7b show that 100 samples are sufficient for our method, and additional data does not yield significant improvements.", + "bbox": [ + 212, + 489, + 784, + 536 + ], + "page_idx": 13 + }, + { + "type": "text", + "text": "6 Conclusion", + "text_level": 1, + "bbox": [ + 214, + 559, + 359, + 574 + ], + "page_idx": 13 + }, + { + "type": "text", + "text": "Our proposed PQ-SAM addresses the challenge of resource constraints in deploying the segment anything model on edge devices. By incorporating the grouped activation distribution transformation with outlier hierarchical clustering techniques, PQ-SAM offers a customized post-training quantization (PTQ) solution for SAM. The GADT reshapes the asymmetric activation distribution of SAM, while the OHC scheme effectively handles extreme outliers and optimizes the allocation of shifting and scaling sizes for each group of channels. Our extensive experiments demonstrate that PQ-SAM surpasses existing PTQ methods on nine zero-shot datasets across three modes of SAM, especially for the low-bit quantization settings.", + "bbox": [ + 212, + 590, + 784, + 742 + ], + "page_idx": 13 + }, + { + "type": "text", + "text": "Acknowledgments. This work was supported in part by the National Natural Science Foundation of China under Grants 62131003 and 62021001. We gratefully acknowledge the support of MindSpore, CANN (Compute Architecture for Neural Networks) and Ascend AI Processor used for this research.", + "bbox": [ + 212, + 757, + 784, + 816 + ], + "page_idx": 13 + }, + { + "type": "page_number", + "text": "14", + "bbox": [ + 217, + 114, + 235, + 126 + ], + "page_idx": 13 + }, + { + "type": "header", + "text": "X. Liu et al.", + "bbox": [ + 271, + 114, + 354, + 126 + ], + "page_idx": 13 + }, + { + "type": "text", + "text": "References", + "text_level": 1, + "bbox": [ + 217, + 143, + 321, + 159 + ], + "page_idx": 14 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "1. Arbelaez, P., Maire, M., Fowlkes, C., Malik, J.: Contour detection and hierarchical image segmentation. IEEE transactions on pattern analysis and machine intelligence 33(5), 898-916 (2010)", + "2. Brown, T., Mann, B., Ryder, N., Subbiah, M., Kaplan, J.D., Dhariwal, P., Neelakantan, A., Shyam, P., Sastry, G., Askell, A., et al.: Language models are few-shot learners. NeurIPS (2020)", + "3. Caicedo, J.C., Goodman, A., Karhohs, K.W., Cimini, B.A., Ackerman, J., Haghighi, M., Heng, C., Becker, T., Doan, M., McQuin, C., et al.: Nucleus segmentation across imaging experiments: the 2018 data science bowl. Nature methods 16(12), 1247-1253 (2019)", + "4. Cheng, H.K., Oh, S.W., Price, B., Schwing, A., Lee, J.Y.: Tracking anything with decoupled video segmentation. In: ICCV (2023)", + "5. Choi, J., Chuang, P.I.J., Wang, Z., Venkataramani, S., Srinivasan, V., Gopalakrishnan, K.: Bridging the accuracy gap for 2-bit quantized neural networks (qnn). arXiv preprint arXiv:1807.06964 (2018)", + "6. Choi, J., Wang, Z., Venkataramani, S., Chuang, P.I.J., Srinivasan, V., Gopalakrishnan, K.: Pact: Parameterized clipping activation for quantized neural networks. arXiv preprint arXiv:1805.06085 (2018)", + "7. Chowdhery, A., Narang, S., Devlin, J., Bosma, M., Mishra, G., Roberts, A., Barham, P., Chung, H.W., Sutton, C., Gehrmann, S., et al.: Palm: Scaling language modeling with pathways. arXiv preprint arXiv:2204.02311 (2022)", + "8. Ciampi, L., Santiago, C., Costeira, J.P., Gennaro, C., Amato, G.: Domain adaptation for traffic density estimation. In: VISAPP (2021)", + "9. Cordts, M., Omran, M., Ramos, S., Rehfeld, T., Enzweiler, M., Benenson, R., Franke, U., Roth, S., Schiele, B.: The cityscapes dataset for semantic urban scene understanding. In: CVPR (2016)", + "10. Ding, X., Liu, X., Zhang, Y., Tu, Z., Li, W., Hu, J., Chen, H., Tang, Y., Xiong, Z., Yin, B., et al.: Cbq: Cross-block quantization for large language models. arXiv preprint arXiv:2312.07950 (2023)", + "1. Gong, S., Zhong, Y., Ma, W., Li, J., Wang, Z., Zhang, J., Heng, P.A., Dou, Q.: 3dsam-adapter: Holistic adaptation of sam from 2d to 3d for promptable medical image segmentation. arXiv preprint arXiv:2306.13465 (2023)", + "2. Gupta, A., Dollar, P., Girshick, R.: Lvis: A dataset for large vocabulary instance segmentation. In: CVPR (2019)", + "3. Han, S., Mao, H., Dally, W.J.: Deep compression: Compressing deep neural networks with pruning, trained quantization and huffman coding. arXiv preprint arXiv:1510.00149 (2015)", + "4. He, K., Chen, X., Xie, S., Li, Y., Dólár, P., Girshick, R.: Masked autoencoders are scalable vision learners. In: CVPR (2022)", + "5. Hinton, G., Vinyals, O., Dean, J.: Distilling the knowledge in a neural network. arXiv preprint arXiv:1503.02531 (2015)", + "6. Jacob, B., Kligys, S., Chen, B., Zhu, M., Tang, M., Howard, A., Adam, H., Kalenichenko, D.: Quantization and training of neural networks for efficient integer-arithmetic-only inference. In: CVPR (2018)", + "7. Jia, C., Yang, Y., Xia, Y., Chen, Y.T., Parekh, Z., Pham, H., Le, Q., Sung, Y.H., Li, Z., Duerig, T.: Scaling up visual and vision-language representation learning with noisy text supervision. In: ICML (2021)" + ], + "bbox": [ + 225, + 178, + 784, + 839 + ], + "page_idx": 14 + }, + { + "type": "header", + "text": "PQ-SAM", + "bbox": [ + 666, + 114, + 730, + 128 + ], + "page_idx": 14 + }, + { + "type": "page_number", + "text": "15", + "bbox": [ + 767, + 116, + 785, + 126 + ], + "page_idx": 14 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "18. Kingma, D.P., Ba, J.L.: Adam: A method for stochastic optimization. In: ICLR (2015)", + "19. Kirillov, A., Mintun, E., Ravi, N., Mao, H., Rolland, C., Gustafson, L., Xiao, T., Whitehead, S., Berg, A.C., Lo, W.Y., et al.: Segment anything. In: ICCV (2023)", + "20. Li, R., Wang, Y., Liang, F., Qin, H., Yan, J., Fan, R.: Fully quantized network for object detection. In: CVPR (2019)", + "21. Li, Z., Gu, Q.: I-vit: Integer-only quantization for efficient vision transformer inference. In: ICCV (2023)", + "22. Li, Z., Ma, L., Chen, M., Xiao, J., Gu, Q.: Patch similarity aware data-free quantization for vision transformers. In: ECCV (2022)", + "23. Li, Z., Xiao, J., Yang, L., Gu, Q.: Repq-vit: Scale reparameterization for posttraining quantization of vision transformers. In: ICCV (2023)", + "24. Lin, T.Y., Maire, M., Belongie, S., Hays, J., Perona, P., Ramanan, D., Dólar, P., Zitnick, C.L.: Microsoft coco: Common objects in context. In: ECCV (2014)", + "25. Lin, Y., Zhang, T., Sun, P., Li, Z., Zhou, S.: Fq-vit: Post-training quantization for fully quantized vision transformer. arXiv preprint arXiv:2111.13824 (2021)", + "26. Liu, X., Cai, M., Chen, Y., Zhang, Y., Shi, T., Zhang, R., Chen, X., Xiong, Z.: Cross-dimension affinity distillation for 3d em neuron segmentation. In: CVPR (2024)", + "27. Liu, X., Hu, B., Huang, W., Zhang, Y., Xiong, Z.: Efficient biomedical instance segmentation via knowledge distillation. In: MICCAI (2022)", + "28. Liu, Z., Wang, Y., Han, K., Zhang, W., Ma, S., Gao, W.: Post-training quantization for vision transformer. In: NeurIPS (2021)", + "29. Loshchilov, I., Hutter, F.: Sgdr: Stochastic gradient descent with warm restarts. arXiv preprint arXiv:1608.03983 (2016)", + "30. Martin, D., Fowlkes, C., Tal, D., Malik, J.: A database of human segmented natural images and its application to evaluating segmentation algorithms and measuring ecological statistics. In: ICCV (2001)", + "31. Paszke, A., Gross, S., Massa, F., Lerer, A., Bradbury, J., Chanan, G., Killeen, T., Lin, Z., Gimelshein, N., Antiga, L., et al.: Pytorch: An imperative style, high-performance deep learning library. In: NeurIPS (2019)", + "32. Pugliatti, M., Topputo, F.: Doors: Dataset for boulders segmentation. Zenodo 9, 20 (2022)", + "33. Radford, A., Kim, J.W., Hallacy, C., Ramesh, A., Goh, G., Agarwal, S., Sastry, G., Askell, A., Mishkin, P., Clark, J., et al.: Learning transferable visual models from natural language supervision. In: ICML (2021)", + "34. Ross, S.M.: Introductory statistics. Academic Press (2017)", + "35. Shamir, R.R., Duchin, Y., Kim, J., Sapiro, G., Harel, N.: Continuous dice coefficient: a method for evaluating probabilistic segmentations. arXiv preprint arXiv:1906.11031 (2019)", + "36. Shao, W., Chen, M., Zhang, Z., Xu, P., Zhao, L., Li, Z., Zhang, K., Gao, P., Qiao, Y., Luo, P.: Omniquant: Omnidirectionally calibrated quantization for large language models. In: ICLR (2023)", + "37. Trotter, C., Atkinson, G., Sharpe, M., Richardson, K., McGough, A.S., Wright, N., Burville, B., Berggren, P.: Ndd20: A large-scale few-shot dolphin dataset for coarse and fine-grained categorisation. arXiv preprint arXiv:2005.13359 (2020)", + "38. Tu, Z., Hu, J., Chen, H., Wang, Y.: Toward accurate post-training quantization for image super resolution. In: CVPR (2023)", + "39. Upton, G., Cook, I.: Understanding statistics. Oxford University Press (1996)", + "40. Vasu, P.K.A., Gabriel, J., Zhu, J., Tuzel, O., Ranjan, A.: Fastvit: A fast hybrid vision transformer using structural reparameterization. In: ICCV (2023)" + ], + "bbox": [ + 215, + 146, + 784, + 840 + ], + "page_idx": 15 + }, + { + "type": "page_number", + "text": "16", + "bbox": [ + 217, + 114, + 235, + 126 + ], + "page_idx": 15 + }, + { + "type": "header", + "text": "X. Liu et al.", + "bbox": [ + 271, + 114, + 354, + 126 + ], + "page_idx": 15 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "41. Wei, X., Zhang, Y., Li, Y., Zhang, X., Gong, R., Guo, J., Liu, X.: Outlier suppression+: Accurate quantization of large language models by equivalent and effective shifting and scaling. In: EMNLP (2023)", + "42. Xie, D., Wang, R., Ma, J., Chen, C., Lu, H., Yang, D., Shi, F., Lin, X.: Edit everything: A text-guided generative system for images editing. arXiv preprint arXiv:2304.14006 (2023)", + "43. Xie, Y., Liao, Y.: Efficient-vit: A light-weight classification model based on cnn and vit. In: ICIGP (2023)", + "44. Xiong, Y., Varadarajan, B., Wu, L., Xiang, X., Xiao, F., Zhu, C., Dai, X., Wang, D., Sun, F., Iandola, F., et al.: Efficientsam: Leveraged masked image pretraining for efficient segment anything. In: CVPR (2024)", + "45. Yang, J., Gao, M., Li, Z., Gao, S., Wang, F., Zheng, F.: Track anything: Segment anything meets videos. arXiv preprint arXiv:2304.11968 (2023)", + "46. Yang, L., Wei, Y.Z., He, Y., Sun, W., Huang, Z., Huang, H., Fan, H.: ishape: A first step towards irregular shape instance segmentation. arXiv preprint arXiv:2109.15068 (2021)", + "47. Yang, X., Ye, J., Wang, X.: Factorizing knowledge in neural networks. In: European Conference on Computer Vision. Springer (2022)", + "48. Yang, X., Zhou, D., Liu, S., Ye, J., Wang, X.: Deep model reassembly. NeurIPS 35, 25739-25753 (2022)", + "49. Yu, F., Huang, K., Wang, M., Cheng, Y., Chu, W., Cui, L.: Width & depth pruning for vision transformers. In: AAAI (2022)", + "50. Yuan, Z., Xue, C., Chen, Y., Wu, Q., Sun, G.: Ptq4vit: Post-training quantization for vision transformers with twin uniform quantization. In: ECCV (2022)", + "51. Zhang, C., Han, D., Qiao, Y., Kim, J.U., Bae, S.H., Lee, S., Hong, C.S.: Faster segment anything: Towards lightweight sam for mobile applications. arXiv preprint arXiv:2306.14289 (2023)", + "52. Zhang, C., Zhang, C., Li, C., Qiao, Y., Zheng, S., Dam, S.K., Zhang, M., Kim, J.U., Kim, S.T., Choi, J., et al.: One small step for generative ai, one giant leap for agi: A complete survey on chatgpt in aigc era. arXiv preprint arXiv:2304.06488 (2023)", + "53. Zhang, Q., Liu, X., Li, W., Chen, H., Liu, J., Hu, J., Xiong, Z., Yuan, C., Wang, Y.: Distilling semantic priors from sam to efficient image restoration models. In: CVPR (2024)", + "54. Zhao, X., Ding, W., An, Y., Du, Y., Yu, T., Li, M., Tang, M., Wang, J.: Fast segment anything. arXiv preprint arXiv:2306.12156 (2023)", + "55. Zwillinger, D., Kokoska, S.: CRC standard probability and statistics tables and formulae. *Crc Press* (1999)" + ], + "bbox": [ + 212, + 146, + 787, + 672 + ], + "page_idx": 16 + }, + { + "type": "header", + "text": "PQ-SAM", + "bbox": [ + 666, + 114, + 730, + 128 + ], + "page_idx": 16 + }, + { + "type": "page_number", + "text": "17", + "bbox": [ + 767, + 114, + 785, + 126 + ], + "page_idx": 16 + } +] \ No newline at end of file diff --git a/2024/PQ-SAM_ Post-training Quantization for Segment Anything Model/5e412ebd-1f25-44d2-98a8-cb85fe86acb8_model.json b/2024/PQ-SAM_ Post-training Quantization for Segment Anything Model/5e412ebd-1f25-44d2-98a8-cb85fe86acb8_model.json new file mode 100644 index 0000000000000000000000000000000000000000..300abe215ed41c7b0ddba491017105a7565ccc0d --- /dev/null +++ b/2024/PQ-SAM_ Post-training Quantization for Segment Anything Model/5e412ebd-1f25-44d2-98a8-cb85fe86acb8_model.json @@ -0,0 +1,2511 @@ +[ + [ + { + "type": "title", + "bbox": [ + 0.218, + 0.141, + 0.784, + 0.186 + ], + "angle": 0, + "content": "PQ-SAM: Post-training Quantization for Segment Anything Model" + }, + { + "type": "text", + "bbox": [ + 0.217, + 0.212, + 0.785, + 0.244 + ], + "angle": 0, + "content": "Xiaoyu Liu\\(^{1}\\), Xin Ding\\(^{1}\\), Lei Yu\\(^{2}\\), Yuanyuan Xi\\(^{2}\\), Wei Li\\(^{2}\\), Zhijun Tu\\(^{2}\\), Jie Hu\\(^{2}\\), Hanting Chen\\(^{2}\\), Baoqun Yin\\(^{1}\\), and Zhiwei Xiong\\(^{1(\\boxtimes)}\\)" + }, + { + "type": "text", + "bbox": [ + 0.284, + 0.255, + 0.717, + 0.311 + ], + "angle": 0, + "content": "1 University of Science and Technology of China {liuxyu,xinding64}@mail.ustc.edu.cn, zwxiong@ustc.edu.cn \n2 Huawei Noah's Ark Lab {wei.lee,zhijun.tu}@huawei.com" + }, + { + "type": "text", + "bbox": [ + 0.263, + 0.349, + 0.744, + 0.64 + ], + "angle": 0, + "content": "Abstract. Segment anything model (SAM) is a promising prompt-guided vision foundation model to segment objects of interest. However, the extensive computational requirements of SAM have limited its applicability in resource-constraint edge devices. Post-training quantization (PTQ) is an effective potential for fast-deploying SAM. Nevertheless, SAM's billion-scale pretraining creates a highly asymmetric activation distribution with detrimental outliers in excessive channels, resulting in significant performance degradation of the low-bit PTQ. In this paper, we propose PQ-SAM, the first PTQ method customized for SAM. To achieve a quantization-friendly tensor-wise distribution, PQ-SAM incorporates a novel grouped activation distribution transformation (GADT) based on a two-stage outlier hierarchical clustering (OHC) scheme to scale and shift each channel. Firstly, OHC identifies and truncates extreme outliers to reduce the scale variance of different channels. Secondly, OHC iteratively allocates learnable shifting and scaling sizes to each group of channels with similar distributions, reducing the number of learnable parameters and easing the optimization difficulty. These shifting and scaling sizes are used to adjust activation channels, and jointly optimized with quantization step sizes for optimal results. Extensive experiments demonstrate that PQ-SAM outperforms existing PTQ methods on nine zero-shot datasets, and pushes the 4-bit PTQ of SAM to a usable level." + }, + { + "type": "text", + "bbox": [ + 0.263, + 0.654, + 0.708, + 0.668 + ], + "angle": 0, + "content": "Keywords: Segment Anything Model-Post-training Quantization" + }, + { + "type": "title", + "bbox": [ + 0.217, + 0.695, + 0.375, + 0.71 + ], + "angle": 0, + "content": "1 Introduction" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.727, + 0.785, + 0.803 + ], + "angle": 0, + "content": "The advent of large language models [2,7,52] has ushered in a new paradigm in natural language processing (NLP). These models exhibit exceptional zero-shot and few-shot generalization, learning broad competencies from sizable pretraining [14,17,33]. Inspired by their success, researchers have begun developing similarly large-scale foundation models for computer vision. A pioneering example" + }, + { + "type": "page_footnote", + "bbox": [ + 0.23, + 0.812, + 0.785, + 0.84 + ], + "angle": 0, + "content": "X. Liu and X. Ding—Equal contribution. This research was conducted while Xiaoyu Liu and Xin Ding were interns at Huawei Noah's Ark Lab." + } + ], + [ + { + "type": "page_number", + "bbox": [ + 0.218, + 0.116, + 0.23, + 0.127 + ], + "angle": 0, + "content": "2" + }, + { + "type": "header", + "bbox": [ + 0.272, + 0.115, + 0.356, + 0.127 + ], + "angle": 0, + "content": "X. Liu et al." + }, + { + "type": "image", + "bbox": [ + 0.303, + 0.145, + 0.499, + 0.213 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.361, + 0.213, + 0.435, + 0.222 + ], + "angle": 0, + "content": "(a) SAM: block1" + }, + { + "type": "image", + "bbox": [ + 0.502, + 0.145, + 0.701, + 0.213 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.554, + 0.213, + 0.631, + 0.222 + ], + "angle": 0, + "content": "(b) SAM: block3" + }, + { + "type": "image", + "bbox": [ + 0.303, + 0.223, + 0.501, + 0.291 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.334, + 0.291, + 0.463, + 0.3 + ], + "angle": 0, + "content": "(c) Conventional ViT: block1" + }, + { + "type": "image", + "bbox": [ + 0.503, + 0.223, + 0.702, + 0.291 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.526, + 0.291, + 0.658, + 0.3 + ], + "angle": 0, + "content": "(d) Conventional ViT: block3" + }, + { + "type": "image_caption", + "bbox": [ + 0.214, + 0.311, + 0.788, + 0.381 + ], + "angle": 0, + "content": "Fig. 1: Comparison of activation distributions between image encoder of SAM (a, b) and the conventional ViT (c, d) in two different transformer blocks. Both of these two models are based on the ViT-base version. Each distribution is visualized with the horizontal axis representing different channels and the vertical axis representing the intensity values within that channel." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.414, + 0.785, + 0.476 + ], + "angle": 0, + "content": "is the recently proposed segment anything model (SAM) [19], representing the prompt-guided visual foundation model with impressive zero-shot capabilities. SAM focuses on the crucial task of segmenting objects of interest from an image, which underpins many vision applications [4, 11, 42, 45]." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.477, + 0.788, + 0.674 + ], + "angle": 0, + "content": "While SAM has demonstrated impressive advancements in segmentation abilities and flexibility, its practical deployment remains challenging. Specifically, SAM's reliance on image encoder of vision transformer (ViT) architecture induces extensive computational overhead that restricts applicability on edge devices with tight latency, power, and memory budgets. This problem necessitates the employment of model compression techniques such as network pruning [13, 49], knowledge distillation [15,26,27,47,48,53], model quantization [5,16,38] and compact architecture design [40,43] etc. Among these techniques, post-training quantization (PTQ) only requires a few unlabeled calibration images without updating model parameters, which enables fast deployment on various devices within a limited time. Given that training and fine-tuning SAM requires significant GPUs using the large-scale training dataset SA1B [19] with 11 million images, there is a growing demand for the PTQ technique." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.674, + 0.789, + 0.841 + ], + "angle": 0, + "content": "Nevertheless, exiting PTQ methods [22, 23, 25, 50] for ViT are focusing on conventional ViT models for classification, which are different from the ViT in SAM. SAM requires accurate prediction for each pixel on multiple zero-shot datasets after billion-scale pretraining, which is much more sensitive to low-bit compression for image embeddings, and its floating-point activations have a highly asymmetric distribution with detrimental outliers. As illustrated in Fig. 1, different from the conventional ViT, our analysis reveals two properties of SAM's activation distribution that pose challenges for quantization: The presence of detrimental outliers implies that the activation values in SAM are not symmetrically distributed around a central tendency. Moreover, SAM's activations often exhibit a concentration of extreme outliers in specific channels, resulting in" + } + ], + [ + { + "type": "header", + "bbox": [ + 0.668, + 0.115, + 0.732, + 0.129 + ], + "angle": 0, + "content": "PQ-SAM" + }, + { + "type": "page_number", + "bbox": [ + 0.775, + 0.117, + 0.787, + 0.127 + ], + "angle": 0, + "content": "3" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.147, + 0.788, + 0.268 + ], + "angle": 0, + "content": "a considerable number of channels that require adjustment. While the existing PTQ methods [36, 41], such as RepQ-Vit [23], propose to scale activations at the channel level, the practical challenge in SAM arises in two aspects. Firstly, extreme outliers introduce detrimental scale variance in specific activation channels. Secondly, adjusting an excessive number of channels simultaneously proves to be challenging. These distribution properties of SAM's activations pose new challenges for existing PTQ methods, especially for the ultra-low-bit (e.g., 4-bit, 2-bit) quantization." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.272, + 0.789, + 0.545 + ], + "angle": 0, + "content": "In this paper, we propose PQ-SAM, the first PTQ method specially designed for SAM, incorporating a grouped activation distribution transformation (GADT) to reshape SAM's activations for quantization-friendly tensor-wise distribution guided by grouped learnable shifting and scaling sizes. Our proposed GADT has a specially designed outlier hierarchical clustering (OHC) scheme to suppress extreme outliers and reduce the total number of learnable parameters in GADT, easing optimization difficulty and preventing overfitting with limited training cost. Specifically, OHC first identifies and truncates tensor-wise extreme outliers across the whole activation tensor by examining interquartile range statistics at the tensor level. This effectively reduces the detrimental scale variance of different channels due to extreme outliers. Secondly, OHC iteratively allocates each pair of learnable shifting and scaling sizes for each group of channels with similar distributions by calculating their quantile range in each iteration. Critically, these shifting and scaling sizes are jointly optimized in an end-to-end manner together with the quantization step sizes for holistic optimal results, under the supervision of the full-precision (FP) SAM model. The optimized shifting and scaling sizes can be reparameterized into corresponding weights so that the proposed PQ-SAM is friendly for practical deployment." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.549, + 0.79, + 0.655 + ], + "angle": 0, + "content": "We conduct extensive experiments to demonstrate the superiority of our PQ-SAM over existing advanced PTQ methods, on nine representative zero-shot datasets across three widely used segmentation modes of SAM, i.e., point prompt mode, automatic mode, and bounding box prompt mode. Our PQ-SAM preserves SAM's strong zero-shot performance and achieves usable 4-bit PTQ for SAM, which represents an important step towards unlocking the practical edge deployment of SAM." + }, + { + "type": "text", + "bbox": [ + 0.239, + 0.659, + 0.574, + 0.675 + ], + "angle": 0, + "content": "The contributions of this paper are as follows:" + }, + { + "type": "text", + "bbox": [ + 0.226, + 0.691, + 0.787, + 0.72 + ], + "angle": 0, + "content": "- We propose PQ-SAM, the first post-training quantization method customized for the segment anything model with limited computational cost." + }, + { + "type": "text", + "bbox": [ + 0.226, + 0.726, + 0.785, + 0.771 + ], + "angle": 0, + "content": "- We propose a novel grouped activation distribution transformation (GADT) to reshape SAM's activations with learnable shifting and scaling sizes jointly end-to-end optimized together with quantization step sizes." + }, + { + "type": "text", + "bbox": [ + 0.225, + 0.776, + 0.785, + 0.804 + ], + "angle": 0, + "content": "- We design a two-stage outlier hierarchical clustering scheme (OHC) to reduce the learnable parameters and ease the optimization difficulty in GADT." + }, + { + "type": "text", + "bbox": [ + 0.225, + 0.81, + 0.785, + 0.839 + ], + "angle": 0, + "content": "- Extensive experiments demonstrate that PQ-SAM significantly outperforms previous PTQ methods on nine zero-shot datasets." + }, + { + "type": "list", + "bbox": [ + 0.225, + 0.691, + 0.787, + 0.839 + ], + "angle": 0, + "content": null + } + ], + [ + { + "type": "page_number", + "bbox": [ + 0.218, + 0.116, + 0.23, + 0.127 + ], + "angle": 0, + "content": "4" + }, + { + "type": "header", + "bbox": [ + 0.272, + 0.115, + 0.355, + 0.127 + ], + "angle": 0, + "content": "X. Liu et al." + }, + { + "type": "title", + "bbox": [ + 0.216, + 0.145, + 0.388, + 0.161 + ], + "angle": 0, + "content": "2 Related Work" + }, + { + "type": "title", + "bbox": [ + 0.216, + 0.173, + 0.504, + 0.188 + ], + "angle": 0, + "content": "2.1 Model Compression for SAM" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.194, + 0.787, + 0.345 + ], + "angle": 0, + "content": "To address the extensive computational requirements of SAM and make SAM suitable for edge devices, recent works MobileSAM [51] and FastSAM [54], EfficientSAM [44] propose different knowledge distillation based technological routes. MobileSAM reduces SAM's heavy image encoder through knowledge distillation. They replace SAM's encoder with a lightweight CNN and distil the heavyweight encoder's knowledge into the CNN. FastSAM avoids SAM's Transformer architecture entirely by adopting a regular convolution neural network (CNN) detector with a segmentation branch and retraining the CNN using 1/50th of SAM's SA1B dataset. EfficientSAM adopts a two-stage training strategy including preraining and finetuning." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.346, + 0.787, + 0.452 + ], + "angle": 0, + "content": "While MobileSAM, FastSAM, and EfficientSAM can compress SAM models, they still require extensive SA1B data and training resources. In contrast, our proposed PQ-SAM technique only requires one millionth of SA1B data for calibration. With just 100 unlabeled images, PQ-SAM achieves efficient quantization of SAM with minimal GPU costs. PQ-SAM provides a highly practical solution to fast deploying SAM using commodity edge resources. Moreover, our PTQ method can be integrated into these methods for better deployment." + }, + { + "type": "title", + "bbox": [ + 0.216, + 0.469, + 0.427, + 0.485 + ], + "angle": 0, + "content": "2.2 Model Quantization" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.49, + 0.79, + 0.703 + ], + "angle": 0, + "content": "Network quantization is an effective compression technique that transforms weight and activation values from floating points to low-bitwidth integers, substantially reducing memory, data movement, and energy costs. Existing methods fall into two categories: Quantization-Aware Training (QAT) [6,16,21] and Post-Training Quantization [10, 16]. However, QAT relies heavily on retraining with the full original dataset, which is infeasible for large vision models like SAM trained on massive data. While PTQ methods using limited calibration data are more practical, current techniques target classification and are not designed for SAM's unique segmentation task. Specifically, SAM exhibits a highly asymmetric distribution with detrimental outliers, which arises from its billion-scale pretraining. Prior PTQ methods for vision transformers [22, 25, 28, 50] do not address these characteristics of the large vision foundation model. To fill this gap, we develop the first customized PTQ approach, PQ-SAM, specifically tailored for quantizing the immense SAM model by accounting for its unusual activation properties." + }, + { + "type": "title", + "bbox": [ + 0.216, + 0.721, + 0.369, + 0.739 + ], + "angle": 0, + "content": "3 Preliminary" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.749, + 0.787, + 0.793 + ], + "angle": 0, + "content": "To provide context before detailing our proposed method, we briefly introduce uniform quantization. We formulate k-bit quantization and dequantization of tensor \\( x \\) as" + }, + { + "type": "equation", + "bbox": [ + 0.345, + 0.798, + 0.786, + 0.828 + ], + "angle": 0, + "content": "\\[\n\\text {Q u a n t}: x _ {q} = \\operatorname {c l i p} \\left(\\operatorname {r o u n d} \\left(\\frac {x}{\\Delta}\\right) + z, 0, 2 ^ {k} - 1\\right), \\tag {1}\n\\]" + }, + { + "type": "equation", + "bbox": [ + 0.33, + 0.828, + 0.553, + 0.843 + ], + "angle": 0, + "content": "\\[\nD e q u a n t: x ^ {\\prime} = \\Delta (x _ {q} - z) \\approx x,\n\\]" + } + ], + [ + { + "type": "header", + "bbox": [ + 0.668, + 0.115, + 0.732, + 0.129 + ], + "angle": 0, + "content": "PQ-SAM" + }, + { + "type": "page_number", + "bbox": [ + 0.775, + 0.117, + 0.785, + 0.127 + ], + "angle": 0, + "content": "5" + }, + { + "type": "image", + "bbox": [ + 0.218, + 0.143, + 0.484, + 0.304 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.491, + 0.143, + 0.788, + 0.305 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.214, + 0.315, + 0.788, + 0.357 + ], + "angle": 0, + "content": "Fig. 2: The workflow of our proposed PQ-SAM. PQ-SAM introduces grouped activation distribution transformation (GADT) operations into each transformer block of SAM to enable model quantization." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.372, + 0.788, + 0.463 + ], + "angle": 0, + "content": "where \\(\\operatorname{Round}(\\cdot)\\) projects a value to the k-bit nearest integer around the input, and \\(\\operatorname{clip}(x,l,u) = \\min (\\max (x,l),u)\\). \\(x_{q}\\) represents the quantized value, and \\(x^{\\prime}\\) are de-quantized value which approximately equals to \\(x\\). Note that \\(\\Delta \\in \\mathbb{R}^{+}\\) represents the quantization step size, and \\(z\\in \\mathbb{Z}\\) denotes the zero-point which is a per-tensor offset. The values of \\(\\Delta\\) and \\(z\\) are determined based on the lower and upper bounds of \\(x\\) as follows:" + }, + { + "type": "equation", + "bbox": [ + 0.32, + 0.472, + 0.787, + 0.506 + ], + "angle": 0, + "content": "\\[\n\\Delta = \\frac {\\operatorname* {m a x} (x) - \\operatorname* {m i n} (x)}{2 ^ {k} - 1}, \\quad z = \\operatorname {r o u n d} \\left(- \\frac {\\operatorname* {m i n} (x)}{\\Delta}\\right). \\tag {2}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.514, + 0.787, + 0.546 + ], + "angle": 0, + "content": "Post-training quantization typically focuses on determining the quantization step sizes \\(\\varDelta_W\\) and \\(\\varDelta_X\\) for weights and activations per layer." + }, + { + "type": "title", + "bbox": [ + 0.214, + 0.567, + 0.725, + 0.586 + ], + "angle": 0, + "content": "4 Grouped Activation Distribution Transformation" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.598, + 0.788, + 0.718 + ], + "angle": 0, + "content": "Overview. As shown in Fig. 2, PQ-SAM inserts learnable grouped activation distribution transformation (GADT) operations into each transformer block of SAM to reshape the distribution of activations. The shifting and scaling sizes in GADT are jointly optimized with quantization step sizes in an end-to-end manner under full-precision SAM supervision. This achieves holistically optimized quantization parameters tailored for SAM. Importantly, the learnable ADT is readily absorbed into the preceding linear layer, thus introducing no additional computations or parameters after quantization." + }, + { + "type": "title", + "bbox": [ + 0.214, + 0.74, + 0.476, + 0.756 + ], + "angle": 0, + "content": "4.1 Scaling and Shifting Sizes" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.765, + 0.788, + 0.841 + ], + "angle": 0, + "content": "As shown in Fig. 1, the highly asymmetric activation distribution of SAM is problematic for the per-tensor quantization setting, making it challenging to properly determine the quantization parameters \\(\\varDelta_W\\) and \\(\\varDelta_X\\). Moreover, the outliers concentrated in specific channels negatively bias \\(\\varDelta_W\\) and \\(\\varDelta_X\\) due to their disproportionate magnitude compared to the majority of activations. To" + } + ], + [ + { + "type": "page_number", + "bbox": [ + 0.218, + 0.116, + 0.23, + 0.127 + ], + "angle": 0, + "content": "6" + }, + { + "type": "header", + "bbox": [ + 0.272, + 0.115, + 0.356, + 0.128 + ], + "angle": 0, + "content": "X. Liu et al." + }, + { + "type": "code_caption", + "bbox": [ + 0.229, + 0.15, + 0.553, + 0.166 + ], + "angle": 0, + "content": "Algorithm 1: outlier hierarchical clustering" + }, + { + "type": "algorithm", + "bbox": [ + 0.218, + 0.167, + 0.554, + 0.439 + ], + "angle": 0, + "content": "Input: FP SAM with activation \\(X\\in \\mathbb{R}^{T\\times C_1}\\) Output: \\(\\{v_{n},s_{n}\\}_{1\\leq n\\leq N}\\) \n1 { Tensor-wise Outlier Truancting.} \n2 Count \\(X_{Q1}\\) \\(X_{Q3}\\) and IQR of \\(X\\) \n3 \\(X^t\\gets X\\) in Eq. 5. \n4 { Channel-wise Outlier Grouping.} \n5 Initialize the iteration number \\(n = 0\\) \n6 Allocated channel Group \\(C_a = \\emptyset\\) \n7 Remaining channel Group \\(C_r = \\{1,2,\\dots,C_1\\}\\) \n8 while \\(C_r\\) is not \\(\\emptyset\\) do \n9 foreach c in \\(C_r\\) do \n10 if \\(X_{1 - \\alpha}^{t,n - 1}\\leq X^t (c)\\leq X_\\alpha^{t,n - 1}\\) then \n11 Allocate \\(\\{v_n,s_n\\} \\to X(c)\\) in Eq. 6; \n12 \\(C_a.\\mathrm{push}(c);\\) \n13 \\(C_r.\\mathrm{pop}(c);\\) \n14 end \n15 end \n16 \\(n = n + 1\\) \n17 end" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.471, + 0.785, + 0.5 + ], + "angle": 0, + "content": "address this, we propose to transform SAM's activations towards a quantization-friendly distribution at the tensor level." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.502, + 0.787, + 0.592 + ], + "angle": 0, + "content": "Straightforwardly, we introduce learnable channel-wise scaling size \\( S \\) and shifting size \\( V \\) to scale and shift the activation tensor on a channel-by-channel basis. For a linear layer, suppose that the input of an activation tensor \\( X \\in \\mathbb{R}^{T \\times C_1} \\) has \\( T \\) tokens and \\( C_1 \\) channels, and the weight matrix \\( W \\in \\mathbb{R}^{C_1 \\times C_2} \\) has a bias \\( B \\in \\mathbb{R}^{1 \\times C_2} \\). We utilize scaling size \\( S \\in \\mathbb{R}^{1 \\times C_1} \\) and shifting size \\( V \\in \\mathbb{R}^{1 \\times C_1} \\) to reshape its distribution, which is formulated as" + }, + { + "type": "equation", + "bbox": [ + 0.433, + 0.599, + 0.786, + 0.617 + ], + "angle": 0, + "content": "\\[\n\\tilde {X} = (X - V) \\oslash S, \\tag {3}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.626, + 0.785, + 0.673 + ], + "angle": 0, + "content": "where \\(\\widetilde{X}\\) represents the reshaped activation tensor and \\(\\varnothing\\) represents the division operation. This process is mathematically equivalent to the original linear layer by updating the weight matrix:" + }, + { + "type": "equation", + "bbox": [ + 0.345, + 0.679, + 0.786, + 0.699 + ], + "angle": 0, + "content": "\\[\nX W + B = \\widetilde {X} \\left(W ^ {\\top} \\odot S ^ {\\top}\\right) + \\left(V W ^ {\\top} + B\\right), \\tag {4}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.705, + 0.785, + 0.734 + ], + "angle": 0, + "content": "where \\(\\odot\\) represents the multiplication operation, and the updated linear layer has a weight matrix \\(\\widetilde{W} = W^{\\top} \\odot S^{\\top}\\) and a bias \\(\\widetilde{B} = (VW^{\\top} + B)\\)." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.735, + 0.787, + 0.84 + ], + "angle": 0, + "content": "Given the practical challenge of optimizing SAM's large number of activation channels, it is difficult to simultaneously optimize the shifting and scaling sizes at the channel level. To address this, we propose Grouped Activation Distribution Transformation (GADT). This technique utilizes our outlier hierarchical clustering scheme to group channels with similar distributions together. By sharing a set of learnable parameters within each group, we effectively reduce the optimization difficulty." + } + ], + [ + { + "type": "header", + "bbox": [ + 0.668, + 0.115, + 0.732, + 0.129 + ], + "angle": 0, + "content": "PQ-SAM" + }, + { + "type": "page_number", + "bbox": [ + 0.776, + 0.117, + 0.786, + 0.127 + ], + "angle": 0, + "content": "7" + }, + { + "type": "title", + "bbox": [ + 0.216, + 0.147, + 0.52, + 0.162 + ], + "angle": 0, + "content": "4.2 Outlier Hierarchical Clustering" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.17, + 0.788, + 0.29 + ], + "angle": 0, + "content": "The above-mentioned shifting and scaling sizes are one-on-one corresponding to each activation channel, which leads to excessive learned parameters being optimized and increases optimization difficulty within limited training cost. To reduce learning difficulty and prevent overfitting of the learnable shifting and scaling sizes, we propose an outlier hierarchical clustering (OHC) scheme to iteratively allocate a pair of shifting and scaling sizes to a group of channels. OHC consists of a tensor-wise outlier truncating stage and a channel-wise outlier grouping stage." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.29, + 0.788, + 0.38 + ], + "angle": 0, + "content": "Tensor-wise Outlier Truncating. Due to the influence of extreme outliers in the activation values, the grouping of channels with similar distributions can be affected. Additionally, these extreme anomalies can impact the effective usage of quantization intervals. Therefore, as a preliminary step, we perform outlier truncation at the tensor level by examining the outlier situation of numerical values across the entire tensor layer." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.38, + 0.788, + 0.456 + ], + "angle": 0, + "content": "Motivated by the statistic theory [34,39,55], we design an effective preprocessing scheme. For the activation tensor \\( X \\), we count the upper and lower quartile fractions \\( Q_{3} \\) and \\( Q_{1} \\), and the interquartile range \\( IQR \\) of its numerical distribution. We then introduce adjustable statistical thresholds, and the truncated activation \\( X^{t} \\) is formulated as" + }, + { + "type": "equation", + "bbox": [ + 0.318, + 0.461, + 0.786, + 0.487 + ], + "angle": 0, + "content": "\\[\nX ^ {t} = \\left\\{X \\mid X _ {Q _ {1}} - \\lambda I Q R < x < X _ {Q _ {3}} + \\lambda I Q R, x \\in X \\right\\}, \\tag {5}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.503, + 0.787, + 0.533 + ], + "angle": 0, + "content": "where \\(\\lambda\\) is a hyper-parameter, and \\(X_{Q_3}\\) and \\(X_{Q_1}\\) are the upper and lower quartile values at \\(Q_3\\) and \\(Q_1\\), respectively." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.533, + 0.802, + 0.64 + ], + "angle": 0, + "content": "Channel-wise Outlier Grouping. For a truncated activation tensor \\( X_{t} \\) consisting of \\( C \\) channels, we allocate \\( N \\) pairs of shifting and scaling sizes \\( \\{v_{n}, s_{n}\\}_{1 \\leq n \\leq N} \\) to the channels, and \\( N \\ll C \\). As depicted in Fig. 2 (b), we perform this allocation iteratively, assigning a pair of shifting and scaling sizes \\( v_{n}, s_{n} \\) to a group of channels that meet the pre-defined range between the upper quantile \\( \\alpha \\) and lower quantile \\( 1 - \\alpha \\) of the non-allocating channels. In the \\( n^{th} \\) iteration, we formulate the procedure of \\( \\{v_{n}, s_{n}\\} \\) for activations of the \\( c^{th} \\) channel \\( X(c) \\) as" + }, + { + "type": "equation", + "bbox": [ + 0.307, + 0.646, + 0.786, + 0.665 + ], + "angle": 0, + "content": "\\[\n\\widetilde {X} (c) = \\left(X (c) - v _ {n}\\right) \\oslash s _ {n}, \\text {i f} X _ {1 - \\alpha} ^ {t, n - 1} \\leq X ^ {t} (c) \\leq X _ {\\alpha} ^ {t, n - 1}, \\tag {6}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.673, + 0.788, + 0.719 + ], + "angle": 0, + "content": "where \\(\\widetilde{X}(c)\\) represents the \\(c^{th}\\) channel of reshaped activation \\(\\widetilde{X}\\). Additionally, \\(X_{1 - \\alpha}^{t,n-1}\\) and \\(X_{\\alpha}^{t,n-1}\\) respectively denote the lower and upper quantile values, of the activations from these channels that are not allocated in previous iterations." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.72, + 0.788, + 0.841 + ], + "angle": 0, + "content": "Overall, the procedure of the proposed outlier hierarchical clustering is summarized in Algorithm 1. This iterative allocation process allows for the dynamic calculation of percentile values for the unallocated channels. By using a fixed value of \\(\\alpha\\), we ensure that more similar channels are grouped together at the beginning of the allocation process. As the allocation process advances, the groups gradually become more refined, resulting in finer granularity. This adaptive approach ultimately yields highly effective grouping outcomes, which prove advantageous for subsequent learning and optimization. By conducting the joint" + } + ], + [ + { + "type": "page_number", + "bbox": [ + 0.218, + 0.116, + 0.23, + 0.127 + ], + "angle": 0, + "content": "8" + }, + { + "type": "header", + "bbox": [ + 0.272, + 0.115, + 0.355, + 0.128 + ], + "angle": 0, + "content": "X. Liu et al." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.147, + 0.784, + 0.178 + ], + "angle": 0, + "content": "To optimize of GADT and quantization step sizes, PQ-SAM quantizes SAM with limited overheads on unlabeled samples." + }, + { + "type": "title", + "bbox": [ + 0.215, + 0.199, + 0.663, + 0.215 + ], + "angle": 0, + "content": "4.3 Joint Optimization with Quantization Step Sizes" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.223, + 0.785, + 0.329 + ], + "angle": 0, + "content": "The shifting and scaling sizes in GADT are learnable and optimized jointly with \\(\\Delta_W\\) and \\(\\Delta_X\\) in our PQ-SAM. This integration of learnable shifting and scaling sizes allows for a more effective adjustment of the activation distribution. We use FP model supervision to learn the quantization step sizes \\(\\Delta_W\\) and \\(\\Delta_X\\) in an end-to-end manner, and adjust offsets via the above-mentioned GADT, which not only achieves channel-wise offset adjustment but also preserves the hardware-friendly per-tensor quantization setting." + }, + { + "type": "text", + "bbox": [ + 0.215, + 0.33, + 0.787, + 0.434 + ], + "angle": 0, + "content": "Given the floating-point weight tensor \\( W \\) and activation tensor \\( X \\) in one layer of SAM, we follow the Eq. 1 and quantize them to \\( k \\)-bit integer values \\( W_{q} \\) and \\( X_{q} \\) with the corresponding weight and activation quantization step sizes \\( \\Delta_W \\) and \\( \\Delta_X \\), respectively. Our optimization framework goes beyond previous approaches [23,36,41] that focus on determining \\( \\Delta_W \\) and \\( \\Delta_X \\) for each layer individually. Instead, our framework enables a holistic, FP-supervised quantization process that is customized for the entire model." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.435, + 0.785, + 0.526 + ], + "angle": 0, + "content": "We jointly optimize the shifting and scaling sizes \\(\\{v_{n}, s_{n}\\}_{1 \\leq n \\leq N}\\), and quantization step sizes \\(\\Delta_{W}\\) and \\(\\Delta_{X}\\) for holistic optimal results. By leveraging the output \\(O(W, K)\\) of the FP model as a guide, we can tailor the quantization steps specifically for the pre-trained weights and activations by minimizing the Mean Squared Error (MSE) loss between \\(O(W, X)\\) and the output \\(O(\\widetilde{W}_q, \\widetilde{X}_q)\\) of quantized model:" + }, + { + "type": "equation", + "bbox": [ + 0.335, + 0.536, + 0.786, + 0.567 + ], + "angle": 0, + "content": "\\[\n\\underset {\\Delta_ {W}, \\Delta_ {X}, \\left\\{v _ {n}, s _ {n} \\right\\} _ {1 \\leq n \\leq N}} {\\arg \\min } \\| O (\\widetilde {W} _ {q}, \\widetilde {X} _ {q}) - O (W, X) \\| _ {2}, \\tag {7}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.578, + 0.784, + 0.627 + ], + "angle": 0, + "content": "where \\(\\widetilde{W}_q\\) represent the corresponding quantized weight of the updated weight by reparameterization as demonstrated in Eq. 4, and \\(\\widetilde{X}_q\\) represents the quantized activation which has been reshaped." + }, + { + "type": "title", + "bbox": [ + 0.216, + 0.649, + 0.367, + 0.667 + ], + "angle": 0, + "content": "5 Experiment" + }, + { + "type": "title", + "bbox": [ + 0.215, + 0.68, + 0.44, + 0.695 + ], + "angle": 0, + "content": "5.1 Datasets and Metrics" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.705, + 0.785, + 0.794 + ], + "angle": 0, + "content": "We conduct evaluations on nine representative zero-shot segmentation datasets, focusing on three widely used segmentation modes of SAM, i.e., the point prompt mode, the automatic mode, and the bounding box prompt mode. These datasets have been officially verified by SAM as mentioned in the work [19]. However, since there is no publicly available test code provided by SAM, we independently evaluate these datasets based on our own resources." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.795, + 0.786, + 0.841 + ], + "angle": 0, + "content": "Point Prompt Mode. The point prompt specifies what to segment in the image, with the goal of determining the mask containing that point. To fully validate the zero-shot performance of quantized SAM, we subsample six datasets" + } + ], + [ + { + "type": "header", + "bbox": [ + 0.668, + 0.115, + 0.732, + 0.129 + ], + "angle": 0, + "content": "PQ-SAM" + }, + { + "type": "page_number", + "bbox": [ + 0.775, + 0.117, + 0.785, + 0.127 + ], + "angle": 0, + "content": "9" + }, + { + "type": "table_caption", + "bbox": [ + 0.217, + 0.145, + 0.784, + 0.172 + ], + "angle": 0, + "content": "Table 1: Quantitative comparison of different PTQ methods for the point prompt mode of SAM." + }, + { + "type": "table", + "bbox": [ + 0.223, + 0.178, + 0.78, + 0.521 + ], + "angle": 0, + "content": "
MethodBitBBBC038V1NDD20CityscapeDOORSiShapeNDISPark
IoUDiceIoUDiceIoUDiceIoUDiceIoUDiceIoUDice
FP320.77950.85000.84680.91000.59450.70610.84960.91260.37220.51080.81920.8938
ViT-H
MinMax [16]60.34850.45990.53940.66050.33280.43740.36780.50820.25900.38630.60950.7091
MSE [5]60.65980.73930.55110.63750.44230.55180.61150.70760.34090.48180.71360.8046
Percentile [20]60.72190.72900.76680.84940.50530.61550.78710.86170.36790.50760.78500.8672
FQ-ViT [25]60.44440.60080.25130.38050.28360.41860.28670.42980.10880.19260.41370.5715
RepQ-ViT [23]60.71740.79380.82230.89390.55980.67850.83830.90480.34440.48890.80100.8838
Ours60.76460.83760.82250.89330.56970.68610.84510.90920.33800.47760.81570.8919
MinMax [16]40.33090.46380.27180.40500.19200.29660.46340.61440.11330.20070.25150.3735
MSE [5]40.37610.51410.36460.51870.22650.33940.42650.58350.14450.24730.39330.5408
Percentile [20]40.37350.52000.34470.48110.26190.38910.38880.52730.12530.21690.41530.5657
FQ-ViT [25]40.22130.34840.09720.16360.14530.23870.11680.19830.03590.06780.21030.3341
RepQ-ViT [23]40.70130.78090.76850.85500.45900.57400.75900.84930.27360.40650.73990.8325
Ours40.70810.79730.78180.86700.50560.62840.79750.87660.29210.42940.76470.8541
FP320.77610.84760.84680.90750.59390.70680.85140.91380.34790.48560.81270.8893
ViT-L
MinMax [16]60.51290.62850.52620.63870.36810.48380.41960.56120.28210.41440.64900.7532
MSE [5]60.70100.77640.65210.74080.47570.58880.68570.78080.35440.49640.75830.8489
Percentile [20]60.73680.80660.75110.83200.52300.63640.78050.85740.37750.51950.79400.8759
FQ-ViT [25]60.42130.56600.35960.51430.28400.42060.38400.54530.12050.21130.40980.5675
RepQ-ViT [23]60.69520.77150.80040.87670.55800.67270.82700.89400.31050.44860.78680.8712
Ours60.74250.81540.81800.89020.55660.67350.83700.90360.31990.45930.79900.8786
MinMax [16]40.12560.20530.27640.41610.12850.20270.41510.57750.09480.17070.18180.2868
MSE [5]40.28360.40420.33040.47470.18410.27660.33410.49010.15540.26430.36410.4994
Percentile [20]40.41360.56370.30740.45310.27200.40530.37010.52590.13550.23390.40840.5631
FQ-ViT [25]40.27910.42380.16690.27470.17940.29030.18130.29790.07460.13760.28960.4406
RepQ-ViT [23]40.41290.49830.72350.82500.44220.56220.75520.84830.27910.41030.71000.8118
Ours40.68550.77270.73930.83390.48780.61140.78900.87120.25200.38330.73790.8332
FP320.76660.84170.83220.90000.56920.68840.82940.89870.35360.49870.79330.8760
ViT-B
MinMax [16]60.38290.47770.57270.69760.33560.44450.34950.47970.26850.40450.59040.7021
MSE [5]60.63250.72550.57670.69210.39760.51630.50300.63460.29430.43460.62940.7398
Percentile [20]60.72690.79970.73640.82600.51990.64060.79520.87470.36050.50680.76760.8564
FQ-ViT [25]60.61540.71520.47830.60100.38260.52170.35380.48590.20830.33100.58460.7196
RepQ-ViT [23]60.69000.76750.80700.87300.54490.66440.80530.88010.34390.48560.78360.8696
Ours60.75830.83310.80080.87820.53910.65930.80590.88860.34410.48860.78450.8705
MinMax [16]40.05560.09860.15330.24130.08970.13740.13970.22520.04470.08420.05890.0964
MSE [5]40.33590.46120.37560.53150.22780.33430.46350.62630.13300.23060.36370.5018
Percentile [20]40.39270.53810.40540.55830.27280.40340.42190.57610.13560.23450.41490.5564
FQ-ViT [25]40.34760.50100.16690.27470.21010.32920.22840.35910.11340.19950.31190.4647
RepQ-ViT [23]40.47380.56590.64790.76080.43380.56530.58780.71180.30640.44660.62530.7448
Ours40.70800.79690.67680.78370.47140.59860.73750.83610.31060.45280.68000.7983
" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.536, + 0.785, + 0.672 + ], + "angle": 0, + "content": "including BBBC038V1 [3], NDD20 [37], Cityscape [9], DOORS [32], iShape [46], and NDISPark [8]. These datasets contain 15,213 images and 29,978 masks and cover a broad range of domains, which is used for the very comprehensive evaluation experiment. Following the original SAM's experiments, we determine the single-point prompt by selecting the \"center\" of the ground truth. This center is identified as the point with the highest value in the mask's interior distance transform. We calculate the mean IoU and Dice [35] for all masks of each dataset as quantitative metrics, which are common metrics to quantify the overlap between predicted and ground truth masks." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.674, + 0.787, + 0.781 + ], + "angle": 0, + "content": "Automatic Mode. SAM's automatic mask generation pipeline can automatically extract a series of masks from an input image with a regular grid of foreground points. We evaluate this automatic mode on the zero-shot edge detection task using the BSDS500 [30] dataset. These automatically generated masks are processed into object edges by the post-processing \\(^{1}\\). We conduct evaluations on the test subset, consisting of 200 images, and assess the performance of all methods using four standard metrics \\(^{2}\\) for edge detection [1]: optimal dataset" + }, + { + "type": "page_footnote", + "bbox": [ + 0.218, + 0.798, + 0.783, + 0.825 + ], + "angle": 0, + "content": "1 Our experiments are based on the descriptions in SAM's paper since the official test code is not provided to the public." + }, + { + "type": "page_footnote", + "bbox": [ + 0.22, + 0.825, + 0.689, + 0.84 + ], + "angle": 0, + "content": "\\(^{2}\\) https://www2.eecs.berkeley.edu/Research/Projects/CS/vision/bsds/" + }, + { + "type": "list", + "bbox": [ + 0.218, + 0.798, + 0.783, + 0.84 + ], + "angle": 0, + "content": null + } + ], + [ + { + "type": "page_number", + "bbox": [ + 0.218, + 0.116, + 0.236, + 0.127 + ], + "angle": 0, + "content": "10" + }, + { + "type": "header", + "bbox": [ + 0.272, + 0.115, + 0.356, + 0.127 + ], + "angle": 0, + "content": "X. Liu et al." + }, + { + "type": "image", + "bbox": [ + 0.216, + 0.145, + 0.787, + 0.307 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.214, + 0.309, + 0.788, + 0.352 + ], + "angle": 0, + "content": "Fig. 3: Visual comparison of different PTQ methods for SAM's point prompt mode. The red block and green star represent the object to be segmented and the position of the prompt point, respectively." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.36, + 0.784, + 0.39 + ], + "angle": 0, + "content": "scale (ODS), optimal image scale (OIS), average precision (AP), and recall at \\(50\\%\\) precision (R50)." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.391, + 0.785, + 0.466 + ], + "angle": 0, + "content": "Bounding Box Prompt Mode. Similar to the point prompt mode, the bounding box prompt mode of SAM focuses on segmenting the objects circled by the bounding box. We follow the original setting of SAM [19] and conduct experiments on the COCO [24] and LVIS [12] datasets. The results are evaluated by average precision (AP)." + }, + { + "type": "title", + "bbox": [ + 0.216, + 0.486, + 0.456, + 0.501 + ], + "angle": 0, + "content": "5.2 Implementation Details" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.508, + 0.787, + 0.629 + ], + "angle": 0, + "content": "Training Details. We utilize the PyTorch [31] and conduct all of our experiments on NVIDIA Tesla V100 GPUs. To calibrate our model, we randomly select 100 samples from the SA1B dataset as calibration data and perform the calibration process over 100 epochs, using a batch size of 1. The learnable quantization parameters \\(\\Delta_W\\), \\(\\Delta_X\\), and \\(\\{v_n, s_n\\}_{1 \\ll n \\ll N}\\), are optimized using the Adam optimizer [18] with \\(\\beta_1 = 0.9\\) and \\(\\beta_2 = 0.999\\), with a learning rate of \\(1e - 4\\). The learning rate scheduler is CosineAnnealingLR [29]. The hyperparameters \\(\\lambda\\) and \\(\\alpha\\) are set as 10 and 0.25, respectively." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.63, + 0.787, + 0.81 + ], + "angle": 0, + "content": "Experimental Settings. In our evaluation experiments, we focus on challenging low-bit-width quantization settings, specifically W4A4 (4-bit quantization for weight and activation) and W6A6 (6-bit quantization), for the point prompt mode of SAM. It is important to note that the automatic mode of SAM, which is used for the edge detection task, is less sensitive to network prediction errors due to the incorporation of various post-processing algorithms. Therefore, we specifically perform 4-bit quantization experiments on the BSDS500 dataset for the automatic mode of SAM. Additionally, to validate the superiority of our PQ-SAM approach as discussed in Section 5.3, we conduct experiments in an ultra-low 2-bit setting specifically for the ViT-B version of SAM. This ultra-low-bit setting allowed us to assess the performance and effectiveness of PQ-SAM in scenarios with extremely limited precision." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.81, + 0.787, + 0.84 + ], + "angle": 0, + "content": "Baseline Methods. In our comparative analysis, we include two categories of PTQ methods as baselines. The first category consists of classical post-training" + } + ], + [ + { + "type": "header", + "bbox": [ + 0.668, + 0.115, + 0.732, + 0.129 + ], + "angle": 0, + "content": "PQ-SAM" + }, + { + "type": "page_number", + "bbox": [ + 0.769, + 0.117, + 0.784, + 0.127 + ], + "angle": 0, + "content": "11" + }, + { + "type": "table_caption", + "bbox": [ + 0.275, + 0.145, + 0.726, + 0.159 + ], + "angle": 0, + "content": "Table 2: Quantitative comparison for the automatic mode of SAM." + }, + { + "type": "table", + "bbox": [ + 0.274, + 0.163, + 0.73, + 0.38 + ], + "angle": 0, + "content": "
H-LVMethodBitODSOISAPR50
FP320.7540.7700.7290.865
MinMax [16]40.4300.4400.304-
MSE [5]40.4540.4680.3880.049
Percentile [20]40.4430.4570.339-
FQ-ViT [25]40.4120.4160.230-
RepQ-ViT [23]40.7060.7200.6470.801
Ours40.7360.7450.7070.859
T-LVFP320.7540.7730.7260.866
MinMax [16]40.4230.4240.247-
MSE [5]40.4420.4560.332-
Percentile [20]40.4440.4550.320-
RepQ-ViT [23]40.7090.7220.6450.769
FQ-ViT [25]40.4240.4290.255-
Ours40.7260.7350.6690.843
B-LVFP320.7450.7550.6900.871
MinMax [16]40.3560.3570.145-
MSE [5]40.4250.4800.4750.374
Percentile [20]40.4490.4670.4000.024
FQ-ViT [25]4----
RepQ-ViT [23]40.7010.7120.6320.806
Ours40.7060.7100.6420.861
" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.405, + 0.785, + 0.526 + ], + "angle": 0, + "content": "methods, namely MSE [5] and MinMax [16]. These methods are versatile and can be easily applied to different models. The second category comprises quantization methods specifically designed for Vision Transformer (ViT) models, including two state-of-the-art methods RepQ-ViT [23] and FQ-ViT [25]. These methods have gained popularity in post-training quantization and are directly relevant to our evaluation. Specifically, RepQ-ViT is designed for the post-LayerNorm activations with severe inter-channel variation but based on a hand-craft heuristic algorithm, which is not as flexible and effective as our learning-based method." + }, + { + "type": "title", + "bbox": [ + 0.216, + 0.553, + 0.441, + 0.568 + ], + "angle": 0, + "content": "5.3 Experimental Results" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.583, + 0.785, + 0.778 + ], + "angle": 0, + "content": "Point Prompt Mode. In Fig. 3, we present the quantitative results of different post-quantization methods for the SAM's point prompt mode. Overall, our PQ-SAM consistently outperforms existing methods across most datasets, which proves that our method can better maintain the strong generalization capacity of SAM. For the 6-bit quantization, PQ-SAM exhibits a performance gap to the full-precision model of less than 0.04 in terms of IoU and Dice metrics across most of the datasets. While PQ-SAM may exhibit slightly lower performance compared to strong baseline methods on a few datasets, it still demonstrates competitive results and preserves the generalization ability overall. For the 4-bit quantization, our method significantly surpasses the baseline methods across all datasets. For instance, PQ-SAM surpasses the second-best method by over \\(40\\%\\) in terms of IoU and Dice metrics on the BBBC038V1 dataset for the SAM's ViT-L and ViT-B versions." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.78, + 0.787, + 0.84 + ], + "angle": 0, + "content": "Automatic Mode. In Table 2, we present the quantitative results of different post-quantization methods for the SAM's automatic mode on the BSDS500 dataset. Our PQ-SAM surpasses the baseline methods across three SAM versions and archives a small performance gap to the FP model." + } + ], + [ + { + "type": "page_number", + "bbox": [ + 0.218, + 0.116, + 0.236, + 0.127 + ], + "angle": 0, + "content": "12" + }, + { + "type": "header", + "bbox": [ + 0.272, + 0.115, + 0.356, + 0.128 + ], + "angle": 0, + "content": "X. Liu et al." + }, + { + "type": "image", + "bbox": [ + 0.218, + 0.145, + 0.31, + 0.227 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.311, + 0.145, + 0.403, + 0.228 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.408, + 0.145, + 0.498, + 0.228 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.502, + 0.145, + 0.597, + 0.228 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.599, + 0.145, + 0.692, + 0.228 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.697, + 0.146, + 0.785, + 0.228 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.216, + 0.231, + 0.784, + 0.259 + ], + "angle": 0, + "content": "Fig. 4: Visual comparison of different PTQ methods for SAM's automatic mode. The blue circle indicates segmentation errors." + }, + { + "type": "table_caption", + "bbox": [ + 0.248, + 0.268, + 0.754, + 0.283 + ], + "angle": 0, + "content": "Table 3: Quantitative Results for the bounding box prompt mode of SAM." + }, + { + "type": "table", + "bbox": [ + 0.244, + 0.287, + 0.761, + 0.376 + ], + "angle": 0, + "content": "
MethodBitCOCOLVIS
APAPSAPMAPLAPAPSAPMAPL
HVT-HFP3246.430.751.061.744.432.157.265.2
RepQ-Vit [23]643.527.648.159.242.430.054.963.8
Ours645.129.649.560.243.431.555.764.0
RepQ-Vit [23]438.224.741.752.536.025.546.554.1
Ours441.126.245.355.938.127.149.156.9
" + }, + { + "type": "table_caption", + "bbox": [ + 0.241, + 0.385, + 0.759, + 0.399 + ], + "angle": 0, + "content": "Table 4: Quantitative comparison for the ultra-low 2bit quantization setting." + }, + { + "type": "table", + "bbox": [ + 0.279, + 0.402, + 0.726, + 0.455 + ], + "angle": 0, + "content": "
MethodNDD20DOORSNDISPark
IoUDiceIoUDiceIoUDice
RepQ-ViT [23]0.26810.41460.36770.53010.27800.4224
Ours0.29410.44800.40990.56970.42130.5812
" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.465, + 0.787, + 0.572 + ], + "angle": 0, + "content": "Bounding Box Prompt Mode. In Table 3, we provide the quantitative results of different post-quantization methods for the SAM's bounding box prompt on the COCO and LVIS datasets. Our PQ-SAM outperforms the baseline methods and has a small performance gap compared to the FP model. In the 6-bit quantization setting, PQ-SAM exhibits a loss of accuracy within \\(3\\%\\) compared to the FP model. Moreover, in the 4-bit quantization configuration, PQ-SAM shows a significant improvement of \\(6\\%\\) compared to the RepQ-Vit method." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.572, + 0.788, + 0.662 + ], + "angle": 0, + "content": "Ultra-low Bit Setting. As shown in Table 4, we validate the superiority of our PQ-SAM on the ultra-low 2-bit quantization setting, which significantly surpasses the state-of-the-art post-training quantization (PTQ) method RepQ-ViT. For instance, our method improves Dice by over \\(35\\%\\) compared to RepQ-ViT, demonstrating the effectiveness of our approach for the ultra-low-bit quantization of SAM models." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.663, + 0.788, + 0.755 + ], + "angle": 0, + "content": "Visualization. We provide the visual results of quantized SAM using different quantization methods for the W4A4 setting, including the point prompt mode in Fig. 3 and the auto mode in Fig. 4, respectively. Compared with existing PTQ methods, our proposed method could achieve better segmentation performance with more accurate segmentation masks and fewer segmentation errors. We further give more qualitative results in the supplementary material." + }, + { + "type": "title", + "bbox": [ + 0.216, + 0.781, + 0.402, + 0.795 + ], + "angle": 0, + "content": "5.4 Ablation Studies" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.81, + 0.787, + 0.84 + ], + "angle": 0, + "content": "We perform ablation studies to evaluate the impacts of each component in the proposed PQ-SAM approach using SAM's ViT-B on the BBBC039V1 dataset." + } + ], + [ + { + "type": "header", + "bbox": [ + 0.668, + 0.115, + 0.732, + 0.129 + ], + "angle": 0, + "content": "PQ-SAM" + }, + { + "type": "page_number", + "bbox": [ + 0.769, + 0.116, + 0.786, + 0.127 + ], + "angle": 0, + "content": "13" + }, + { + "type": "table_caption", + "bbox": [ + 0.214, + 0.145, + 0.788, + 0.187 + ], + "angle": 0, + "content": "Table 5: Ablation study on learnable quantization parameters. \\(\\delta_W\\) and \\(\\delta_X\\) represent the weight and activation quantization step sizes, respectively. \\(\\{v_n,s_n\\}_{1\\leq n\\leq N}\\) represents the allocated N pairs of shifting and scaling sizes." + }, + { + "type": "table", + "bbox": [ + 0.303, + 0.194, + 0.703, + 0.255 + ], + "angle": 0, + "content": "
ΔW,ΔX{vn}1≤n≤N{sn}1≤n≤NIoUDice
xxx0.05560.0986
x✓✓0.15150.2339
✓xx0.63420.7334
✓✓x0.68470.7721
✓✓✓0.70800.7969
" + }, + { + "type": "table_caption", + "bbox": [ + 0.214, + 0.27, + 0.788, + 0.354 + ], + "angle": 0, + "content": "Table 6: Ablation study on activation distribution transformation (ADT) with the outlier hierarchical clustering (OHC). 'Truncating' represents the tensor-level outlier truncating operation. 'Even Groupsize' represents that we evenly allocate the same number of learnable shifting and scaling sizes as ours to each group of adjacent channels. 'Per-Channel' represents that we allocate each pair of learnable shifting and scaling sizes to each channel." + }, + { + "type": "table", + "bbox": [ + 0.26, + 0.361, + 0.744, + 0.421 + ], + "angle": 0, + "content": "
ADT | Truncating | Even Groupwise | Per-Channel | OHC (ours) | IoU | Dice
xxxxx0.63420.7334
x✓xxx0.65330.7602
✓✓x✓x0.68420.7720
✓✓✓xx0.68170.7797
✓✓xx✓0.70800.7969
" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.443, + 0.788, + 0.52 + ], + "angle": 0, + "content": "Learnable Quantization Parameters. As presented in Table 5, we perform an ablation study on the learnable quantization parameters in our PQ-SAM framework. This involves considering the weight and activation quantization step sizes, denoted as \\(\\delta_W\\) and \\(\\delta_X\\) respectively, as well as allocating a set of \\(N\\) pairs of shifting and scaling sizes \\(\\{v_n,s_n\\}_{1\\leq n\\leq N}\\)." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.521, + 0.788, + 0.627 + ], + "angle": 0, + "content": "The results demonstrate that introducing learnable quantization step sizes, shifting sizes, or scaling sizes individually resulted in significantly improved performance compared to fixed quantization schemes. Moreover, it is observed that the learnable shifting and scaling sizes yielded even better quantization performance than the learnable quantization step sizes. By optimizing all of these quantization parameters during training, the model is able to effectively adapt to the underlying data distribution and achieve more accurate quantization." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.628, + 0.788, + 0.748 + ], + "angle": 0, + "content": "Activation Distribution Transformation. In Table 6, we present the results of an ablation study to validate the effectiveness of the activation distribution transformation (ADT) in combination with the outlier hierarchical clustering (OHC) scheme. The ADT technique significantly improves the quantization performance, and the OHC scheme further enhances the optimization process and leads to additional improvements in quantization performance. Meanwhile, we validate that the tensor-level outlier truncating operation in ADT can significantly improve the quantization performance." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.75, + 0.788, + 0.84 + ], + "angle": 0, + "content": "To validate the effectiveness of the outlier hierarchical clustering (OHC) scheme, we compare it with two alternative schemes: the \"Even Groupsize\" and \"Per-Channel\" schemes. In the \"Even Groupsize\" scheme, we allocate the same number of learnable shifting and scaling sizes to each group of adjacent channels. In the \"Per-Channel\" scheme, we allocate a unique pair of learnable shifting and scaling sizes to each individual channel, resulting in the same number of learn-" + } + ], + [ + { + "type": "page_number", + "bbox": [ + 0.218, + 0.116, + 0.236, + 0.127 + ], + "angle": 0, + "content": "14" + }, + { + "type": "header", + "bbox": [ + 0.272, + 0.115, + 0.355, + 0.127 + ], + "angle": 0, + "content": "X. Liu et al." + }, + { + "type": "table_caption", + "bbox": [ + 0.235, + 0.145, + 0.766, + 0.159 + ], + "angle": 0, + "content": "Table 7: Ablation study on hyper-parameters and amounts of calibration data." + }, + { + "type": "table", + "bbox": [ + 0.248, + 0.166, + 0.472, + 0.253 + ], + "angle": 0, + "content": "
(a) Hyper-parameters of PQ-SAM.
λ61014
IoU0.69010.70800.6858
Dice0.78940.79690.7871
α0.200.250.30
IoU0.68840.70800.6841
Dice0.78120.79690.7693
" + }, + { + "type": "table_caption", + "bbox": [ + 0.522, + 0.166, + 0.734, + 0.179 + ], + "angle": 0, + "content": "(b) Amount of calibration data." + }, + { + "type": "table", + "bbox": [ + 0.522, + 0.189, + 0.761, + 0.228 + ], + "angle": 0, + "content": "
Number of data100200500
IoU0.70800.70910.7120
Dice0.79690.79800.8014
" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.265, + 0.784, + 0.324 + ], + "angle": 0, + "content": "able shifting and scaling size pairs as there are activation channels. The results of our ablation study demonstrate that the activation distribution transformation (ADT) with OHC in the PQ-SAM framework outperforms those with the \"Even Groupsize\" or \"Per-Channel\" schemes." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.325, + 0.785, + 0.49 + ], + "angle": 0, + "content": "Hyperparameter \\(\\lambda\\) and \\(\\alpha\\). We conduct ablation studies with extreme parameter settings to showcase the impact of hyperparameters \\(\\lambda\\) and \\(\\alpha\\) on performance, as shown in Table 7a. The first set of experiments focuses on evaluating the impact of the hyperparameter \\(\\lambda\\) on tensor-level outlier truncation. Different values of \\(\\lambda\\) are tested for various truncated ranges. The results demonstrate the effects of varying \\(\\lambda\\) on the outlier truncation process and its subsequent influence on the overall performance of the PQ-SAM framework. The second set of experiments involves an ablation study on different upper quantiles \\(\\alpha\\) for the proposed channel-wise outlier grouping in the outlier hierarchical clustering step. By testing different values of \\(\\alpha\\), different numbers of learnable shifting and scaling sizes are allocated to each group of channels." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.491, + 0.785, + 0.537 + ], + "angle": 0, + "content": "Amount of Calibration Data. We conduct an ablation study on the amount of calibration data. The results in Table 7b show that 100 samples are sufficient for our method, and additional data does not yield significant improvements." + }, + { + "type": "title", + "bbox": [ + 0.215, + 0.56, + 0.36, + 0.575 + ], + "angle": 0, + "content": "6 Conclusion" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.592, + 0.785, + 0.743 + ], + "angle": 0, + "content": "Our proposed PQ-SAM addresses the challenge of resource constraints in deploying the segment anything model on edge devices. By incorporating the grouped activation distribution transformation with outlier hierarchical clustering techniques, PQ-SAM offers a customized post-training quantization (PTQ) solution for SAM. The GADT reshapes the asymmetric activation distribution of SAM, while the OHC scheme effectively handles extreme outliers and optimizes the allocation of shifting and scaling sizes for each group of channels. Our extensive experiments demonstrate that PQ-SAM surpasses existing PTQ methods on nine zero-shot datasets across three modes of SAM, especially for the low-bit quantization settings." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.758, + 0.785, + 0.818 + ], + "angle": 0, + "content": "Acknowledgments. This work was supported in part by the National Natural Science Foundation of China under Grants 62131003 and 62021001. We gratefully acknowledge the support of MindSpore, CANN (Compute Architecture for Neural Networks) and Ascend AI Processor used for this research." + } + ], + [ + { + "type": "header", + "bbox": [ + 0.668, + 0.115, + 0.732, + 0.129 + ], + "angle": 0, + "content": "PQ-SAM" + }, + { + "type": "page_number", + "bbox": [ + 0.769, + 0.117, + 0.786, + 0.127 + ], + "angle": 0, + "content": "15" + }, + { + "type": "title", + "bbox": [ + 0.218, + 0.145, + 0.323, + 0.16 + ], + "angle": 0, + "content": "References" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.179, + 0.785, + 0.221 + ], + "angle": 0, + "content": "1. Arbelaez, P., Maire, M., Fowlkes, C., Malik, J.: Contour detection and hierarchical image segmentation. IEEE transactions on pattern analysis and machine intelligence 33(5), 898-916 (2010)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.223, + 0.785, + 0.263 + ], + "angle": 0, + "content": "2. Brown, T., Mann, B., Ryder, N., Subbiah, M., Kaplan, J.D., Dhariwal, P., Neelakantan, A., Shyam, P., Sastry, G., Askell, A., et al.: Language models are few-shot learners. NeurIPS (2020)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.265, + 0.785, + 0.319 + ], + "angle": 0, + "content": "3. Caicedo, J.C., Goodman, A., Karhohs, K.W., Cimini, B.A., Ackerman, J., Haghighi, M., Heng, C., Becker, T., Doan, M., McQuin, C., et al.: Nucleus segmentation across imaging experiments: the 2018 data science bowl. Nature methods 16(12), 1247-1253 (2019)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.321, + 0.785, + 0.347 + ], + "angle": 0, + "content": "4. Cheng, H.K., Oh, S.W., Price, B., Schwing, A., Lee, J.Y.: Tracking anything with decoupled video segmentation. In: ICCV (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.349, + 0.785, + 0.389 + ], + "angle": 0, + "content": "5. Choi, J., Chuang, P.I.J., Wang, Z., Venkataramani, S., Srinivasan, V., Gopalakrishnan, K.: Bridging the accuracy gap for 2-bit quantized neural networks (qnn). arXiv preprint arXiv:1807.06964 (2018)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.391, + 0.785, + 0.432 + ], + "angle": 0, + "content": "6. Choi, J., Wang, Z., Venkataramani, S., Chuang, P.I.J., Srinivasan, V., Gopalakrishnan, K.: Pact: Parameterized clipping activation for quantized neural networks. arXiv preprint arXiv:1805.06085 (2018)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.434, + 0.785, + 0.473 + ], + "angle": 0, + "content": "7. Chowdhery, A., Narang, S., Devlin, J., Bosma, M., Mishra, G., Roberts, A., Barham, P., Chung, H.W., Sutton, C., Gehrmann, S., et al.: Palm: Scaling language modeling with pathways. arXiv preprint arXiv:2204.02311 (2022)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.476, + 0.785, + 0.502 + ], + "angle": 0, + "content": "8. Ciampi, L., Santiago, C., Costeira, J.P., Gennaro, C., Amato, G.: Domain adaptation for traffic density estimation. In: VISAPP (2021)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.504, + 0.785, + 0.544 + ], + "angle": 0, + "content": "9. Cordts, M., Omran, M., Ramos, S., Rehfeld, T., Enzweiler, M., Benenson, R., Franke, U., Roth, S., Schiele, B.: The cityscapes dataset for semantic urban scene understanding. In: CVPR (2016)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.546, + 0.785, + 0.586 + ], + "angle": 0, + "content": "10. Ding, X., Liu, X., Zhang, Y., Tu, Z., Li, W., Hu, J., Chen, H., Tang, Y., Xiong, Z., Yin, B., et al.: Cbq: Cross-block quantization for large language models. arXiv preprint arXiv:2312.07950 (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.588, + 0.785, + 0.628 + ], + "angle": 0, + "content": "1. Gong, S., Zhong, Y., Ma, W., Li, J., Wang, Z., Zhang, J., Heng, P.A., Dou, Q.: 3dsam-adapter: Holistic adaptation of sam from 2d to 3d for promptable medical image segmentation. arXiv preprint arXiv:2306.13465 (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.63, + 0.785, + 0.656 + ], + "angle": 0, + "content": "2. Gupta, A., Dollar, P., Girshick, R.: Lvis: A dataset for large vocabulary instance segmentation. In: CVPR (2019)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.659, + 0.785, + 0.698 + ], + "angle": 0, + "content": "3. Han, S., Mao, H., Dally, W.J.: Deep compression: Compressing deep neural networks with pruning, trained quantization and huffman coding. arXiv preprint arXiv:1510.00149 (2015)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.7, + 0.785, + 0.727 + ], + "angle": 0, + "content": "4. He, K., Chen, X., Xie, S., Li, Y., Dólár, P., Girshick, R.: Masked autoencoders are scalable vision learners. In: CVPR (2022)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.729, + 0.785, + 0.755 + ], + "angle": 0, + "content": "5. Hinton, G., Vinyals, O., Dean, J.: Distilling the knowledge in a neural network. arXiv preprint arXiv:1503.02531 (2015)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.757, + 0.785, + 0.797 + ], + "angle": 0, + "content": "6. Jacob, B., Kligys, S., Chen, B., Zhu, M., Tang, M., Howard, A., Adam, H., Kalenichenko, D.: Quantization and training of neural networks for efficient integer-arithmetic-only inference. In: CVPR (2018)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.799, + 0.785, + 0.84 + ], + "angle": 0, + "content": "7. Jia, C., Yang, Y., Xia, Y., Chen, Y.T., Parekh, Z., Pham, H., Le, Q., Sung, Y.H., Li, Z., Duerig, T.: Scaling up visual and vision-language representation learning with noisy text supervision. In: ICML (2021)" + }, + { + "type": "list", + "bbox": [ + 0.226, + 0.179, + 0.785, + 0.84 + ], + "angle": 0, + "content": null + } + ], + [ + { + "type": "page_number", + "bbox": [ + 0.218, + 0.116, + 0.236, + 0.127 + ], + "angle": 0, + "content": "16" + }, + { + "type": "header", + "bbox": [ + 0.272, + 0.115, + 0.356, + 0.127 + ], + "angle": 0, + "content": "X. Liu et al." + }, + { + "type": "ref_text", + "bbox": [ + 0.218, + 0.147, + 0.785, + 0.175 + ], + "angle": 0, + "content": "18. Kingma, D.P., Ba, J.L.: Adam: A method for stochastic optimization. In: ICLR (2015)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.175, + 0.785, + 0.202 + ], + "angle": 0, + "content": "19. Kirillov, A., Mintun, E., Ravi, N., Mao, H., Rolland, C., Gustafson, L., Xiao, T., Whitehead, S., Berg, A.C., Lo, W.Y., et al.: Segment anything. In: ICCV (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.203, + 0.785, + 0.23 + ], + "angle": 0, + "content": "20. Li, R., Wang, Y., Liang, F., Qin, H., Yan, J., Fan, R.: Fully quantized network for object detection. In: CVPR (2019)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.23, + 0.785, + 0.256 + ], + "angle": 0, + "content": "21. Li, Z., Gu, Q.: I-vit: Integer-only quantization for efficient vision transformer inference. In: ICCV (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.257, + 0.785, + 0.284 + ], + "angle": 0, + "content": "22. Li, Z., Ma, L., Chen, M., Xiao, J., Gu, Q.: Patch similarity aware data-free quantization for vision transformers. In: ECCV (2022)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.284, + 0.785, + 0.311 + ], + "angle": 0, + "content": "23. Li, Z., Xiao, J., Yang, L., Gu, Q.: Repq-vit: Scale reparameterization for posttraining quantization of vision transformers. In: ICCV (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.311, + 0.785, + 0.338 + ], + "angle": 0, + "content": "24. Lin, T.Y., Maire, M., Belongie, S., Hays, J., Perona, P., Ramanan, D., Dólar, P., Zitnick, C.L.: Microsoft coco: Common objects in context. In: ECCV (2014)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.338, + 0.785, + 0.365 + ], + "angle": 0, + "content": "25. Lin, Y., Zhang, T., Sun, P., Li, Z., Zhou, S.: Fq-vit: Post-training quantization for fully quantized vision transformer. arXiv preprint arXiv:2111.13824 (2021)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.365, + 0.785, + 0.405 + ], + "angle": 0, + "content": "26. Liu, X., Cai, M., Chen, Y., Zhang, Y., Shi, T., Zhang, R., Chen, X., Xiong, Z.: Cross-dimension affinity distillation for 3d em neuron segmentation. In: CVPR (2024)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.406, + 0.785, + 0.433 + ], + "angle": 0, + "content": "27. Liu, X., Hu, B., Huang, W., Zhang, Y., Xiong, Z.: Efficient biomedical instance segmentation via knowledge distillation. In: MICCAI (2022)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.433, + 0.785, + 0.46 + ], + "angle": 0, + "content": "28. Liu, Z., Wang, Y., Han, K., Zhang, W., Ma, S., Gao, W.: Post-training quantization for vision transformer. In: NeurIPS (2021)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.46, + 0.785, + 0.487 + ], + "angle": 0, + "content": "29. Loshchilov, I., Hutter, F.: Sgdr: Stochastic gradient descent with warm restarts. arXiv preprint arXiv:1608.03983 (2016)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.487, + 0.785, + 0.528 + ], + "angle": 0, + "content": "30. Martin, D., Fowlkes, C., Tal, D., Malik, J.: A database of human segmented natural images and its application to evaluating segmentation algorithms and measuring ecological statistics. In: ICCV (2001)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.528, + 0.785, + 0.569 + ], + "angle": 0, + "content": "31. Paszke, A., Gross, S., Massa, F., Lerer, A., Bradbury, J., Chanan, G., Killeen, T., Lin, Z., Gimelshein, N., Antiga, L., et al.: Pytorch: An imperative style, high-performance deep learning library. In: NeurIPS (2019)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.569, + 0.785, + 0.595 + ], + "angle": 0, + "content": "32. Pugliatti, M., Topputo, F.: Doors: Dataset for boulders segmentation. Zenodo 9, 20 (2022)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.596, + 0.785, + 0.637 + ], + "angle": 0, + "content": "33. Radford, A., Kim, J.W., Hallacy, C., Ramesh, A., Goh, G., Agarwal, S., Sastry, G., Askell, A., Mishkin, P., Clark, J., et al.: Learning transferable visual models from natural language supervision. In: ICML (2021)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.637, + 0.633, + 0.65 + ], + "angle": 0, + "content": "34. Ross, S.M.: Introductory statistics. Academic Press (2017)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.65, + 0.785, + 0.69 + ], + "angle": 0, + "content": "35. Shamir, R.R., Duchin, Y., Kim, J., Sapiro, G., Harel, N.: Continuous dice coefficient: a method for evaluating probabilistic segmentations. arXiv preprint arXiv:1906.11031 (2019)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.691, + 0.785, + 0.732 + ], + "angle": 0, + "content": "36. Shao, W., Chen, M., Zhang, Z., Xu, P., Zhao, L., Li, Z., Zhang, K., Gao, P., Qiao, Y., Luo, P.: Omniquant: Omnidirectionally calibrated quantization for large language models. In: ICLR (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.732, + 0.785, + 0.772 + ], + "angle": 0, + "content": "37. Trotter, C., Atkinson, G., Sharpe, M., Richardson, K., McGough, A.S., Wright, N., Burville, B., Berggren, P.: Ndd20: A large-scale few-shot dolphin dataset for coarse and fine-grained categorisation. arXiv preprint arXiv:2005.13359 (2020)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.772, + 0.785, + 0.8 + ], + "angle": 0, + "content": "38. Tu, Z., Hu, J., Chen, H., Wang, Y.: Toward accurate post-training quantization for image super resolution. In: CVPR (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.8, + 0.76, + 0.813 + ], + "angle": 0, + "content": "39. Upton, G., Cook, I.: Understanding statistics. Oxford University Press (1996)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.813, + 0.785, + 0.841 + ], + "angle": 0, + "content": "40. Vasu, P.K.A., Gabriel, J., Zhu, J., Tuzel, O., Ranjan, A.: Fastvit: A fast hybrid vision transformer using structural reparameterization. In: ICCV (2023)" + }, + { + "type": "list", + "bbox": [ + 0.217, + 0.147, + 0.785, + 0.841 + ], + "angle": 0, + "content": null + } + ], + [ + { + "type": "header", + "bbox": [ + 0.668, + 0.115, + 0.732, + 0.129 + ], + "angle": 0, + "content": "PQ-SAM" + }, + { + "type": "page_number", + "bbox": [ + 0.769, + 0.116, + 0.786, + 0.127 + ], + "angle": 0, + "content": "17" + }, + { + "type": "ref_text", + "bbox": [ + 0.214, + 0.147, + 0.788, + 0.189 + ], + "angle": 0, + "content": "41. Wei, X., Zhang, Y., Li, Y., Zhang, X., Gong, R., Guo, J., Liu, X.: Outlier suppression+: Accurate quantization of large language models by equivalent and effective shifting and scaling. In: EMNLP (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.214, + 0.19, + 0.788, + 0.232 + ], + "angle": 0, + "content": "42. Xie, D., Wang, R., Ma, J., Chen, C., Lu, H., Yang, D., Shi, F., Lin, X.: Edit everything: A text-guided generative system for images editing. arXiv preprint arXiv:2304.14006 (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.214, + 0.232, + 0.788, + 0.259 + ], + "angle": 0, + "content": "43. Xie, Y., Liao, Y.: Efficient-vit: A light-weight classification model based on cnn and vit. In: ICIGP (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.214, + 0.259, + 0.788, + 0.3 + ], + "angle": 0, + "content": "44. Xiong, Y., Varadarajan, B., Wu, L., Xiang, X., Xiao, F., Zhu, C., Dai, X., Wang, D., Sun, F., Iandola, F., et al.: Efficientsam: Leveraged masked image pretraining for efficient segment anything. In: CVPR (2024)" + }, + { + "type": "ref_text", + "bbox": [ + 0.214, + 0.3, + 0.788, + 0.328 + ], + "angle": 0, + "content": "45. Yang, J., Gao, M., Li, Z., Gao, S., Wang, F., Zheng, F.: Track anything: Segment anything meets videos. arXiv preprint arXiv:2304.11968 (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.214, + 0.328, + 0.788, + 0.37 + ], + "angle": 0, + "content": "46. Yang, L., Wei, Y.Z., He, Y., Sun, W., Huang, Z., Huang, H., Fan, H.: ishape: A first step towards irregular shape instance segmentation. arXiv preprint arXiv:2109.15068 (2021)" + }, + { + "type": "ref_text", + "bbox": [ + 0.214, + 0.371, + 0.788, + 0.398 + ], + "angle": 0, + "content": "47. Yang, X., Ye, J., Wang, X.: Factorizing knowledge in neural networks. In: European Conference on Computer Vision. Springer (2022)" + }, + { + "type": "ref_text", + "bbox": [ + 0.214, + 0.398, + 0.788, + 0.425 + ], + "angle": 0, + "content": "48. Yang, X., Zhou, D., Liu, S., Ye, J., Wang, X.: Deep model reassembly. NeurIPS 35, 25739-25753 (2022)" + }, + { + "type": "ref_text", + "bbox": [ + 0.214, + 0.425, + 0.788, + 0.452 + ], + "angle": 0, + "content": "49. Yu, F., Huang, K., Wang, M., Cheng, Y., Chu, W., Cui, L.: Width & depth pruning for vision transformers. In: AAAI (2022)" + }, + { + "type": "ref_text", + "bbox": [ + 0.214, + 0.453, + 0.788, + 0.481 + ], + "angle": 0, + "content": "50. Yuan, Z., Xue, C., Chen, Y., Wu, Q., Sun, G.: Ptq4vit: Post-training quantization for vision transformers with twin uniform quantization. In: ECCV (2022)" + }, + { + "type": "ref_text", + "bbox": [ + 0.214, + 0.481, + 0.788, + 0.522 + ], + "angle": 0, + "content": "51. Zhang, C., Han, D., Qiao, Y., Kim, J.U., Bae, S.H., Lee, S., Hong, C.S.: Faster segment anything: Towards lightweight sam for mobile applications. arXiv preprint arXiv:2306.14289 (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.214, + 0.522, + 0.788, + 0.577 + ], + "angle": 0, + "content": "52. Zhang, C., Zhang, C., Li, C., Qiao, Y., Zheng, S., Dam, S.K., Zhang, M., Kim, J.U., Kim, S.T., Choi, J., et al.: One small step for generative ai, one giant leap for agi: A complete survey on chatgpt in aigc era. arXiv preprint arXiv:2304.06488 (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.214, + 0.577, + 0.788, + 0.618 + ], + "angle": 0, + "content": "53. Zhang, Q., Liu, X., Li, W., Chen, H., Liu, J., Hu, J., Xiong, Z., Yuan, C., Wang, Y.: Distilling semantic priors from sam to efficient image restoration models. In: CVPR (2024)" + }, + { + "type": "ref_text", + "bbox": [ + 0.214, + 0.619, + 0.788, + 0.646 + ], + "angle": 0, + "content": "54. Zhao, X., Ding, W., An, Y., Du, Y., Yu, T., Li, M., Tang, M., Wang, J.: Fast segment anything. arXiv preprint arXiv:2306.12156 (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.214, + 0.646, + 0.788, + 0.674 + ], + "angle": 0, + "content": "55. Zwillinger, D., Kokoska, S.: CRC standard probability and statistics tables and formulae. *Crc Press* (1999)" + }, + { + "type": "list", + "bbox": [ + 0.214, + 0.147, + 0.788, + 0.674 + ], + "angle": 0, + "content": null + } + ] +] \ No newline at end of file diff --git a/2024/PQ-SAM_ Post-training Quantization for Segment Anything Model/5e412ebd-1f25-44d2-98a8-cb85fe86acb8_origin.pdf b/2024/PQ-SAM_ Post-training Quantization for Segment Anything Model/5e412ebd-1f25-44d2-98a8-cb85fe86acb8_origin.pdf new file mode 100644 index 0000000000000000000000000000000000000000..c172298f84ff3443b4ff27146fa3d3be10e5a9ac --- /dev/null +++ b/2024/PQ-SAM_ Post-training Quantization for Segment Anything Model/5e412ebd-1f25-44d2-98a8-cb85fe86acb8_origin.pdf @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:7175549509055236197f71a90ad956363c140d29254d0faf3e94401c003eb894 +size 1167789 diff --git a/2024/PQ-SAM_ Post-training Quantization for Segment Anything Model/full.md b/2024/PQ-SAM_ Post-training Quantization for Segment Anything Model/full.md new file mode 100644 index 0000000000000000000000000000000000000000..b506503d3f83ca4708bd6980f81408298b60ae27 --- /dev/null +++ b/2024/PQ-SAM_ Post-training Quantization for Segment Anything Model/full.md @@ -0,0 +1,339 @@ +# PQ-SAM: Post-training Quantization for Segment Anything Model + +Xiaoyu Liu $^{1}$ , Xin Ding $^{1}$ , Lei Yu $^{2}$ , Yuanyuan Xi $^{2}$ , Wei Li $^{2}$ , Zhijun Tu $^{2}$ , Jie Hu $^{2}$ , Hanting Chen $^{2}$ , Baoqun Yin $^{1}$ , and Zhiwei Xiong $^{1(\boxtimes)}$ + +1 University of Science and Technology of China {liuxyu,xinding64}@mail.ustc.edu.cn, zwxiong@ustc.edu.cn +2 Huawei Noah's Ark Lab {wei.lee,zhijun.tu}@huawei.com + +Abstract. Segment anything model (SAM) is a promising prompt-guided vision foundation model to segment objects of interest. However, the extensive computational requirements of SAM have limited its applicability in resource-constraint edge devices. Post-training quantization (PTQ) is an effective potential for fast-deploying SAM. Nevertheless, SAM's billion-scale pretraining creates a highly asymmetric activation distribution with detrimental outliers in excessive channels, resulting in significant performance degradation of the low-bit PTQ. In this paper, we propose PQ-SAM, the first PTQ method customized for SAM. To achieve a quantization-friendly tensor-wise distribution, PQ-SAM incorporates a novel grouped activation distribution transformation (GADT) based on a two-stage outlier hierarchical clustering (OHC) scheme to scale and shift each channel. Firstly, OHC identifies and truncates extreme outliers to reduce the scale variance of different channels. Secondly, OHC iteratively allocates learnable shifting and scaling sizes to each group of channels with similar distributions, reducing the number of learnable parameters and easing the optimization difficulty. These shifting and scaling sizes are used to adjust activation channels, and jointly optimized with quantization step sizes for optimal results. Extensive experiments demonstrate that PQ-SAM outperforms existing PTQ methods on nine zero-shot datasets, and pushes the 4-bit PTQ of SAM to a usable level. + +Keywords: Segment Anything Model-Post-training Quantization + +# 1 Introduction + +The advent of large language models [2,7,52] has ushered in a new paradigm in natural language processing (NLP). These models exhibit exceptional zero-shot and few-shot generalization, learning broad competencies from sizable pretraining [14,17,33]. Inspired by their success, researchers have begun developing similarly large-scale foundation models for computer vision. A pioneering example + +![](images/c17837d20f33e4d9a58e77e2045f76af8c61cecfa1985761feb3972b1a6a67fc.jpg) +(a) SAM: block1 + +![](images/204ef9c94144ea33b4232596c17feb658e4e9c8f68a10d6be457d30923e823cd.jpg) +(b) SAM: block3 + +![](images/96d05616c1ddd8d2ce65628c8294b6f7496715ceb2a3a51b88ede47b5b1d6e1a.jpg) +(c) Conventional ViT: block1 + +![](images/37cae7ad5b8b0b1513bf72c81b2ac5d94da74722e575c3a118adbc5a4b865407.jpg) +(d) Conventional ViT: block3 +Fig. 1: Comparison of activation distributions between image encoder of SAM (a, b) and the conventional ViT (c, d) in two different transformer blocks. Both of these two models are based on the ViT-base version. Each distribution is visualized with the horizontal axis representing different channels and the vertical axis representing the intensity values within that channel. + +is the recently proposed segment anything model (SAM) [19], representing the prompt-guided visual foundation model with impressive zero-shot capabilities. SAM focuses on the crucial task of segmenting objects of interest from an image, which underpins many vision applications [4, 11, 42, 45]. + +While SAM has demonstrated impressive advancements in segmentation abilities and flexibility, its practical deployment remains challenging. Specifically, SAM's reliance on image encoder of vision transformer (ViT) architecture induces extensive computational overhead that restricts applicability on edge devices with tight latency, power, and memory budgets. This problem necessitates the employment of model compression techniques such as network pruning [13, 49], knowledge distillation [15,26,27,47,48,53], model quantization [5,16,38] and compact architecture design [40,43] etc. Among these techniques, post-training quantization (PTQ) only requires a few unlabeled calibration images without updating model parameters, which enables fast deployment on various devices within a limited time. Given that training and fine-tuning SAM requires significant GPUs using the large-scale training dataset SA1B [19] with 11 million images, there is a growing demand for the PTQ technique. + +Nevertheless, exiting PTQ methods [22, 23, 25, 50] for ViT are focusing on conventional ViT models for classification, which are different from the ViT in SAM. SAM requires accurate prediction for each pixel on multiple zero-shot datasets after billion-scale pretraining, which is much more sensitive to low-bit compression for image embeddings, and its floating-point activations have a highly asymmetric distribution with detrimental outliers. As illustrated in Fig. 1, different from the conventional ViT, our analysis reveals two properties of SAM's activation distribution that pose challenges for quantization: The presence of detrimental outliers implies that the activation values in SAM are not symmetrically distributed around a central tendency. Moreover, SAM's activations often exhibit a concentration of extreme outliers in specific channels, resulting in + +a considerable number of channels that require adjustment. While the existing PTQ methods [36, 41], such as RepQ-Vit [23], propose to scale activations at the channel level, the practical challenge in SAM arises in two aspects. Firstly, extreme outliers introduce detrimental scale variance in specific activation channels. Secondly, adjusting an excessive number of channels simultaneously proves to be challenging. These distribution properties of SAM's activations pose new challenges for existing PTQ methods, especially for the ultra-low-bit (e.g., 4-bit, 2-bit) quantization. + +In this paper, we propose PQ-SAM, the first PTQ method specially designed for SAM, incorporating a grouped activation distribution transformation (GADT) to reshape SAM's activations for quantization-friendly tensor-wise distribution guided by grouped learnable shifting and scaling sizes. Our proposed GADT has a specially designed outlier hierarchical clustering (OHC) scheme to suppress extreme outliers and reduce the total number of learnable parameters in GADT, easing optimization difficulty and preventing overfitting with limited training cost. Specifically, OHC first identifies and truncates tensor-wise extreme outliers across the whole activation tensor by examining interquartile range statistics at the tensor level. This effectively reduces the detrimental scale variance of different channels due to extreme outliers. Secondly, OHC iteratively allocates each pair of learnable shifting and scaling sizes for each group of channels with similar distributions by calculating their quantile range in each iteration. Critically, these shifting and scaling sizes are jointly optimized in an end-to-end manner together with the quantization step sizes for holistic optimal results, under the supervision of the full-precision (FP) SAM model. The optimized shifting and scaling sizes can be reparameterized into corresponding weights so that the proposed PQ-SAM is friendly for practical deployment. + +We conduct extensive experiments to demonstrate the superiority of our PQ-SAM over existing advanced PTQ methods, on nine representative zero-shot datasets across three widely used segmentation modes of SAM, i.e., point prompt mode, automatic mode, and bounding box prompt mode. Our PQ-SAM preserves SAM's strong zero-shot performance and achieves usable 4-bit PTQ for SAM, which represents an important step towards unlocking the practical edge deployment of SAM. + +The contributions of this paper are as follows: + +- We propose PQ-SAM, the first post-training quantization method customized for the segment anything model with limited computational cost. +- We propose a novel grouped activation distribution transformation (GADT) to reshape SAM's activations with learnable shifting and scaling sizes jointly end-to-end optimized together with quantization step sizes. +- We design a two-stage outlier hierarchical clustering scheme (OHC) to reduce the learnable parameters and ease the optimization difficulty in GADT. +- Extensive experiments demonstrate that PQ-SAM significantly outperforms previous PTQ methods on nine zero-shot datasets. + +# 2 Related Work + +# 2.1 Model Compression for SAM + +To address the extensive computational requirements of SAM and make SAM suitable for edge devices, recent works MobileSAM [51] and FastSAM [54], EfficientSAM [44] propose different knowledge distillation based technological routes. MobileSAM reduces SAM's heavy image encoder through knowledge distillation. They replace SAM's encoder with a lightweight CNN and distil the heavyweight encoder's knowledge into the CNN. FastSAM avoids SAM's Transformer architecture entirely by adopting a regular convolution neural network (CNN) detector with a segmentation branch and retraining the CNN using 1/50th of SAM's SA1B dataset. EfficientSAM adopts a two-stage training strategy including preraining and finetuning. + +While MobileSAM, FastSAM, and EfficientSAM can compress SAM models, they still require extensive SA1B data and training resources. In contrast, our proposed PQ-SAM technique only requires one millionth of SA1B data for calibration. With just 100 unlabeled images, PQ-SAM achieves efficient quantization of SAM with minimal GPU costs. PQ-SAM provides a highly practical solution to fast deploying SAM using commodity edge resources. Moreover, our PTQ method can be integrated into these methods for better deployment. + +# 2.2 Model Quantization + +Network quantization is an effective compression technique that transforms weight and activation values from floating points to low-bitwidth integers, substantially reducing memory, data movement, and energy costs. Existing methods fall into two categories: Quantization-Aware Training (QAT) [6,16,21] and Post-Training Quantization [10, 16]. However, QAT relies heavily on retraining with the full original dataset, which is infeasible for large vision models like SAM trained on massive data. While PTQ methods using limited calibration data are more practical, current techniques target classification and are not designed for SAM's unique segmentation task. Specifically, SAM exhibits a highly asymmetric distribution with detrimental outliers, which arises from its billion-scale pretraining. Prior PTQ methods for vision transformers [22, 25, 28, 50] do not address these characteristics of the large vision foundation model. To fill this gap, we develop the first customized PTQ approach, PQ-SAM, specifically tailored for quantizing the immense SAM model by accounting for its unusual activation properties. + +# 3 Preliminary + +To provide context before detailing our proposed method, we briefly introduce uniform quantization. We formulate k-bit quantization and dequantization of tensor $x$ as + +$$ +\text {Q u a n t}: x _ {q} = \operatorname {c l i p} \left(\operatorname {r o u n d} \left(\frac {x}{\Delta}\right) + z, 0, 2 ^ {k} - 1\right), \tag {1} +$$ + +$$ +D e q u a n t: x ^ {\prime} = \Delta (x _ {q} - z) \approx x, +$$ + +![](images/c34ea5eab4001cb0647d745609d082d058868384ee01079d66eebac38338aa0f.jpg) +Fig. 2: The workflow of our proposed PQ-SAM. PQ-SAM introduces grouped activation distribution transformation (GADT) operations into each transformer block of SAM to enable model quantization. + +![](images/8e0e94ecd9a29678bcb388a37b7457eb34b576286bb8b6eaff157edde86f24fc.jpg) + +where $\operatorname{Round}(\cdot)$ projects a value to the k-bit nearest integer around the input, and $\operatorname{clip}(x,l,u) = \min (\max (x,l),u)$ . $x_{q}$ represents the quantized value, and $x^{\prime}$ are de-quantized value which approximately equals to $x$ . Note that $\Delta \in \mathbb{R}^{+}$ represents the quantization step size, and $z\in \mathbb{Z}$ denotes the zero-point which is a per-tensor offset. The values of $\Delta$ and $z$ are determined based on the lower and upper bounds of $x$ as follows: + +$$ +\Delta = \frac {\operatorname* {m a x} (x) - \operatorname* {m i n} (x)}{2 ^ {k} - 1}, \quad z = \operatorname {r o u n d} \left(- \frac {\operatorname* {m i n} (x)}{\Delta}\right). \tag {2} +$$ + +Post-training quantization typically focuses on determining the quantization step sizes $\varDelta_W$ and $\varDelta_X$ for weights and activations per layer. + +# 4 Grouped Activation Distribution Transformation + +Overview. As shown in Fig. 2, PQ-SAM inserts learnable grouped activation distribution transformation (GADT) operations into each transformer block of SAM to reshape the distribution of activations. The shifting and scaling sizes in GADT are jointly optimized with quantization step sizes in an end-to-end manner under full-precision SAM supervision. This achieves holistically optimized quantization parameters tailored for SAM. Importantly, the learnable ADT is readily absorbed into the preceding linear layer, thus introducing no additional computations or parameters after quantization. + +# 4.1 Scaling and Shifting Sizes + +As shown in Fig. 1, the highly asymmetric activation distribution of SAM is problematic for the per-tensor quantization setting, making it challenging to properly determine the quantization parameters $\varDelta_W$ and $\varDelta_X$ . Moreover, the outliers concentrated in specific channels negatively bias $\varDelta_W$ and $\varDelta_X$ due to their disproportionate magnitude compared to the majority of activations. To + +Algorithm 1: outlier hierarchical clustering +Input: FP SAM with activation $X\in \mathbb{R}^{T\times C_1}$ Output: $\{v_{n},s_{n}\}_{1\leq n\leq N}$ +1 { Tensor-wise Outlier Truancting.} +2 Count $X_{Q1}$ $X_{Q3}$ and IQR of $X$ +3 $X^t\gets X$ in Eq. 5. +4 { Channel-wise Outlier Grouping.} +5 Initialize the iteration number $n = 0$ +6 Allocated channel Group $C_a = \emptyset$ +7 Remaining channel Group $C_r = \{1,2,\dots,C_1\}$ +8 while $C_r$ is not $\emptyset$ do +9 foreach c in $C_r$ do +10 if $X_{1 - \alpha}^{t,n - 1}\leq X^t (c)\leq X_\alpha^{t,n - 1}$ then +11 Allocate $\{v_n,s_n\} \to X(c)$ in Eq. 6; +12 $C_a.\mathrm{push}(c);$ +13 $C_r.\mathrm{pop}(c);$ +14 end +15 end +16 $n = n + 1$ +17 end + +address this, we propose to transform SAM's activations towards a quantization-friendly distribution at the tensor level. + +Straightforwardly, we introduce learnable channel-wise scaling size $S$ and shifting size $V$ to scale and shift the activation tensor on a channel-by-channel basis. For a linear layer, suppose that the input of an activation tensor $X \in \mathbb{R}^{T \times C_1}$ has $T$ tokens and $C_1$ channels, and the weight matrix $W \in \mathbb{R}^{C_1 \times C_2}$ has a bias $B \in \mathbb{R}^{1 \times C_2}$ . We utilize scaling size $S \in \mathbb{R}^{1 \times C_1}$ and shifting size $V \in \mathbb{R}^{1 \times C_1}$ to reshape its distribution, which is formulated as + +$$ +\tilde {X} = (X - V) \oslash S, \tag {3} +$$ + +where $\widetilde{X}$ represents the reshaped activation tensor and $\varnothing$ represents the division operation. This process is mathematically equivalent to the original linear layer by updating the weight matrix: + +$$ +X W + B = \widetilde {X} \left(W ^ {\top} \odot S ^ {\top}\right) + \left(V W ^ {\top} + B\right), \tag {4} +$$ + +where $\odot$ represents the multiplication operation, and the updated linear layer has a weight matrix $\widetilde{W} = W^{\top} \odot S^{\top}$ and a bias $\widetilde{B} = (VW^{\top} + B)$ . + +Given the practical challenge of optimizing SAM's large number of activation channels, it is difficult to simultaneously optimize the shifting and scaling sizes at the channel level. To address this, we propose Grouped Activation Distribution Transformation (GADT). This technique utilizes our outlier hierarchical clustering scheme to group channels with similar distributions together. By sharing a set of learnable parameters within each group, we effectively reduce the optimization difficulty. + +# 4.2 Outlier Hierarchical Clustering + +The above-mentioned shifting and scaling sizes are one-on-one corresponding to each activation channel, which leads to excessive learned parameters being optimized and increases optimization difficulty within limited training cost. To reduce learning difficulty and prevent overfitting of the learnable shifting and scaling sizes, we propose an outlier hierarchical clustering (OHC) scheme to iteratively allocate a pair of shifting and scaling sizes to a group of channels. OHC consists of a tensor-wise outlier truncating stage and a channel-wise outlier grouping stage. + +Tensor-wise Outlier Truncating. Due to the influence of extreme outliers in the activation values, the grouping of channels with similar distributions can be affected. Additionally, these extreme anomalies can impact the effective usage of quantization intervals. Therefore, as a preliminary step, we perform outlier truncation at the tensor level by examining the outlier situation of numerical values across the entire tensor layer. + +Motivated by the statistic theory [34,39,55], we design an effective preprocessing scheme. For the activation tensor $X$ , we count the upper and lower quartile fractions $Q_{3}$ and $Q_{1}$ , and the interquartile range $IQR$ of its numerical distribution. We then introduce adjustable statistical thresholds, and the truncated activation $X^{t}$ is formulated as + +$$ +X ^ {t} = \left\{X \mid X _ {Q _ {1}} - \lambda I Q R < x < X _ {Q _ {3}} + \lambda I Q R, x \in X \right\}, \tag {5} +$$ + +where $\lambda$ is a hyper-parameter, and $X_{Q_3}$ and $X_{Q_1}$ are the upper and lower quartile values at $Q_3$ and $Q_1$ , respectively. + +Channel-wise Outlier Grouping. For a truncated activation tensor $X_{t}$ consisting of $C$ channels, we allocate $N$ pairs of shifting and scaling sizes $\{v_{n}, s_{n}\}_{1 \leq n \leq N}$ to the channels, and $N \ll C$ . As depicted in Fig. 2 (b), we perform this allocation iteratively, assigning a pair of shifting and scaling sizes $v_{n}, s_{n}$ to a group of channels that meet the pre-defined range between the upper quantile $\alpha$ and lower quantile $1 - \alpha$ of the non-allocating channels. In the $n^{th}$ iteration, we formulate the procedure of $\{v_{n}, s_{n}\}$ for activations of the $c^{th}$ channel $X(c)$ as + +$$ +\widetilde {X} (c) = \left(X (c) - v _ {n}\right) \oslash s _ {n}, \text {i f} X _ {1 - \alpha} ^ {t, n - 1} \leq X ^ {t} (c) \leq X _ {\alpha} ^ {t, n - 1}, \tag {6} +$$ + +where $\widetilde{X}(c)$ represents the $c^{th}$ channel of reshaped activation $\widetilde{X}$ . Additionally, $X_{1 - \alpha}^{t,n-1}$ and $X_{\alpha}^{t,n-1}$ respectively denote the lower and upper quantile values, of the activations from these channels that are not allocated in previous iterations. + +Overall, the procedure of the proposed outlier hierarchical clustering is summarized in Algorithm 1. This iterative allocation process allows for the dynamic calculation of percentile values for the unallocated channels. By using a fixed value of $\alpha$ , we ensure that more similar channels are grouped together at the beginning of the allocation process. As the allocation process advances, the groups gradually become more refined, resulting in finer granularity. This adaptive approach ultimately yields highly effective grouping outcomes, which prove advantageous for subsequent learning and optimization. By conducting the joint + +To optimize of GADT and quantization step sizes, PQ-SAM quantizes SAM with limited overheads on unlabeled samples. + +# 4.3 Joint Optimization with Quantization Step Sizes + +The shifting and scaling sizes in GADT are learnable and optimized jointly with $\Delta_W$ and $\Delta_X$ in our PQ-SAM. This integration of learnable shifting and scaling sizes allows for a more effective adjustment of the activation distribution. We use FP model supervision to learn the quantization step sizes $\Delta_W$ and $\Delta_X$ in an end-to-end manner, and adjust offsets via the above-mentioned GADT, which not only achieves channel-wise offset adjustment but also preserves the hardware-friendly per-tensor quantization setting. + +Given the floating-point weight tensor $W$ and activation tensor $X$ in one layer of SAM, we follow the Eq. 1 and quantize them to $k$ -bit integer values $W_{q}$ and $X_{q}$ with the corresponding weight and activation quantization step sizes $\Delta_W$ and $\Delta_X$ , respectively. Our optimization framework goes beyond previous approaches [23,36,41] that focus on determining $\Delta_W$ and $\Delta_X$ for each layer individually. Instead, our framework enables a holistic, FP-supervised quantization process that is customized for the entire model. + +We jointly optimize the shifting and scaling sizes $\{v_{n}, s_{n}\}_{1 \leq n \leq N}$ , and quantization step sizes $\Delta_{W}$ and $\Delta_{X}$ for holistic optimal results. By leveraging the output $O(W, K)$ of the FP model as a guide, we can tailor the quantization steps specifically for the pre-trained weights and activations by minimizing the Mean Squared Error (MSE) loss between $O(W, X)$ and the output $O(\widetilde{W}_q, \widetilde{X}_q)$ of quantized model: + +$$ +\underset {\Delta_ {W}, \Delta_ {X}, \left\{v _ {n}, s _ {n} \right\} _ {1 \leq n \leq N}} {\arg \min } \| O (\widetilde {W} _ {q}, \widetilde {X} _ {q}) - O (W, X) \| _ {2}, \tag {7} +$$ + +where $\widetilde{W}_q$ represent the corresponding quantized weight of the updated weight by reparameterization as demonstrated in Eq. 4, and $\widetilde{X}_q$ represents the quantized activation which has been reshaped. + +# 5 Experiment + +# 5.1 Datasets and Metrics + +We conduct evaluations on nine representative zero-shot segmentation datasets, focusing on three widely used segmentation modes of SAM, i.e., the point prompt mode, the automatic mode, and the bounding box prompt mode. These datasets have been officially verified by SAM as mentioned in the work [19]. However, since there is no publicly available test code provided by SAM, we independently evaluate these datasets based on our own resources. + +Point Prompt Mode. The point prompt specifies what to segment in the image, with the goal of determining the mask containing that point. To fully validate the zero-shot performance of quantized SAM, we subsample six datasets + +Table 1: Quantitative comparison of different PTQ methods for the point prompt mode of SAM. + +
MethodBitBBBC038V1NDD20CityscapeDOORSiShapeNDISPark
IoUDiceIoUDiceIoUDiceIoUDiceIoUDiceIoUDice
FP320.77950.85000.84680.91000.59450.70610.84960.91260.37220.51080.81920.8938
ViT-H
MinMax [16]60.34850.45990.53940.66050.33280.43740.36780.50820.25900.38630.60950.7091
MSE [5]60.65980.73930.55110.63750.44230.55180.61150.70760.34090.48180.71360.8046
Percentile [20]60.72190.72900.76680.84940.50530.61550.78710.86170.36790.50760.78500.8672
FQ-ViT [25]60.44440.60080.25130.38050.28360.41860.28670.42980.10880.19260.41370.5715
RepQ-ViT [23]60.71740.79380.82230.89390.55980.67850.83830.90480.34440.48890.80100.8838
Ours60.76460.83760.82250.89330.56970.68610.84510.90920.33800.47760.81570.8919
MinMax [16]40.33090.46380.27180.40500.19200.29660.46340.61440.11330.20070.25150.3735
MSE [5]40.37610.51410.36460.51870.22650.33940.42650.58350.14450.24730.39330.5408
Percentile [20]40.37350.52000.34470.48110.26190.38910.38880.52730.12530.21690.41530.5657
FQ-ViT [25]40.22130.34840.09720.16360.14530.23870.11680.19830.03590.06780.21030.3341
RepQ-ViT [23]40.70130.78090.76850.85500.45900.57400.75900.84930.27360.40650.73990.8325
Ours40.70810.79730.78180.86700.50560.62840.79750.87660.29210.42940.76470.8541
FP320.77610.84760.84680.90750.59390.70680.85140.91380.34790.48560.81270.8893
ViT-L
MinMax [16]60.51290.62850.52620.63870.36810.48380.41960.56120.28210.41440.64900.7532
MSE [5]60.70100.77640.65210.74080.47570.58880.68570.78080.35440.49640.75830.8489
Percentile [20]60.73680.80660.75110.83200.52300.63640.78050.85740.37750.51950.79400.8759
FQ-ViT [25]60.42130.56600.35960.51430.28400.42060.38400.54530.12050.21130.40980.5675
RepQ-ViT [23]60.69520.77150.80040.87670.55800.67270.82700.89400.31050.44860.78680.8712
Ours60.74250.81540.81800.89020.55660.67350.83700.90360.31990.45930.79900.8786
MinMax [16]40.12560.20530.27640.41610.12850.20270.41510.57750.09480.17070.18180.2868
MSE [5]40.28360.40420.33040.47470.18410.27660.33410.49010.15540.26430.36410.4994
Percentile [20]40.41360.56370.30740.45310.27200.40530.37010.52590.13550.23390.40840.5631
FQ-ViT [25]40.27910.42380.16690.27470.17940.29030.18130.29790.07460.13760.28960.4406
RepQ-ViT [23]40.41290.49830.72350.82500.44220.56220.75520.84830.27910.41030.71000.8118
Ours40.68550.77270.73930.83390.48780.61140.78900.87120.25200.38330.73790.8332
FP320.76660.84170.83220.90000.56920.68840.82940.89870.35360.49870.79330.8760
ViT-B
MinMax [16]60.38290.47770.57270.69760.33560.44450.34950.47970.26850.40450.59040.7021
MSE [5]60.63250.72550.57670.69210.39760.51630.50300.63460.29430.43460.62940.7398
Percentile [20]60.72690.79970.73640.82600.51990.64060.79520.87470.36050.50680.76760.8564
FQ-ViT [25]60.61540.71520.47830.60100.38260.52170.35380.48590.20830.33100.58460.7196
RepQ-ViT [23]60.69000.76750.80700.87300.54490.66440.80530.88010.34390.48560.78360.8696
Ours60.75830.83310.80080.87820.53910.65930.80590.88860.34410.48860.78450.8705
MinMax [16]40.05560.09860.15330.24130.08970.13740.13970.22520.04470.08420.05890.0964
MSE [5]40.33590.46120.37560.53150.22780.33430.46350.62630.13300.23060.36370.5018
Percentile [20]40.39270.53810.40540.55830.27280.40340.42190.57610.13560.23450.41490.5564
FQ-ViT [25]40.34760.50100.16690.27470.21010.32920.22840.35910.11340.19950.31190.4647
RepQ-ViT [23]40.47380.56590.64790.76080.43380.56530.58780.71180.30640.44660.62530.7448
Ours40.70800.79690.67680.78370.47140.59860.73750.83610.31060.45280.68000.7983
+ +including BBBC038V1 [3], NDD20 [37], Cityscape [9], DOORS [32], iShape [46], and NDISPark [8]. These datasets contain 15,213 images and 29,978 masks and cover a broad range of domains, which is used for the very comprehensive evaluation experiment. Following the original SAM's experiments, we determine the single-point prompt by selecting the "center" of the ground truth. This center is identified as the point with the highest value in the mask's interior distance transform. We calculate the mean IoU and Dice [35] for all masks of each dataset as quantitative metrics, which are common metrics to quantify the overlap between predicted and ground truth masks. + +Automatic Mode. SAM's automatic mask generation pipeline can automatically extract a series of masks from an input image with a regular grid of foreground points. We evaluate this automatic mode on the zero-shot edge detection task using the BSDS500 [30] dataset. These automatically generated masks are processed into object edges by the post-processing $^{1}$ . We conduct evaluations on the test subset, consisting of 200 images, and assess the performance of all methods using four standard metrics $^{2}$ for edge detection [1]: optimal dataset + +![](images/ff09578659fe2a3db7d0f84b8d59ff6ef34ca858445b4cfbc222ec61a1d100e7.jpg) +Fig. 3: Visual comparison of different PTQ methods for SAM's point prompt mode. The red block and green star represent the object to be segmented and the position of the prompt point, respectively. + +scale (ODS), optimal image scale (OIS), average precision (AP), and recall at $50\%$ precision (R50). + +Bounding Box Prompt Mode. Similar to the point prompt mode, the bounding box prompt mode of SAM focuses on segmenting the objects circled by the bounding box. We follow the original setting of SAM [19] and conduct experiments on the COCO [24] and LVIS [12] datasets. The results are evaluated by average precision (AP). + +# 5.2 Implementation Details + +Training Details. We utilize the PyTorch [31] and conduct all of our experiments on NVIDIA Tesla V100 GPUs. To calibrate our model, we randomly select 100 samples from the SA1B dataset as calibration data and perform the calibration process over 100 epochs, using a batch size of 1. The learnable quantization parameters $\Delta_W$ , $\Delta_X$ , and $\{v_n, s_n\}_{1 \ll n \ll N}$ , are optimized using the Adam optimizer [18] with $\beta_1 = 0.9$ and $\beta_2 = 0.999$ , with a learning rate of $1e - 4$ . The learning rate scheduler is CosineAnnealingLR [29]. The hyperparameters $\lambda$ and $\alpha$ are set as 10 and 0.25, respectively. + +Experimental Settings. In our evaluation experiments, we focus on challenging low-bit-width quantization settings, specifically W4A4 (4-bit quantization for weight and activation) and W6A6 (6-bit quantization), for the point prompt mode of SAM. It is important to note that the automatic mode of SAM, which is used for the edge detection task, is less sensitive to network prediction errors due to the incorporation of various post-processing algorithms. Therefore, we specifically perform 4-bit quantization experiments on the BSDS500 dataset for the automatic mode of SAM. Additionally, to validate the superiority of our PQ-SAM approach as discussed in Section 5.3, we conduct experiments in an ultra-low 2-bit setting specifically for the ViT-B version of SAM. This ultra-low-bit setting allowed us to assess the performance and effectiveness of PQ-SAM in scenarios with extremely limited precision. + +Baseline Methods. In our comparative analysis, we include two categories of PTQ methods as baselines. The first category consists of classical post-training + +Table 2: Quantitative comparison for the automatic mode of SAM. + +
H-LVMethodBitODSOISAPR50
FP320.7540.7700.7290.865
MinMax [16]40.4300.4400.304-
MSE [5]40.4540.4680.3880.049
Percentile [20]40.4430.4570.339-
FQ-ViT [25]40.4120.4160.230-
RepQ-ViT [23]40.7060.7200.6470.801
Ours40.7360.7450.7070.859
T-LVFP320.7540.7730.7260.866
MinMax [16]40.4230.4240.247-
MSE [5]40.4420.4560.332-
Percentile [20]40.4440.4550.320-
RepQ-ViT [23]40.7090.7220.6450.769
FQ-ViT [25]40.4240.4290.255-
Ours40.7260.7350.6690.843
B-LVFP320.7450.7550.6900.871
MinMax [16]40.3560.3570.145-
MSE [5]40.4250.4800.4750.374
Percentile [20]40.4490.4670.4000.024
FQ-ViT [25]4----
RepQ-ViT [23]40.7010.7120.6320.806
Ours40.7060.7100.6420.861
+ +methods, namely MSE [5] and MinMax [16]. These methods are versatile and can be easily applied to different models. The second category comprises quantization methods specifically designed for Vision Transformer (ViT) models, including two state-of-the-art methods RepQ-ViT [23] and FQ-ViT [25]. These methods have gained popularity in post-training quantization and are directly relevant to our evaluation. Specifically, RepQ-ViT is designed for the post-LayerNorm activations with severe inter-channel variation but based on a hand-craft heuristic algorithm, which is not as flexible and effective as our learning-based method. + +# 5.3 Experimental Results + +Point Prompt Mode. In Fig. 3, we present the quantitative results of different post-quantization methods for the SAM's point prompt mode. Overall, our PQ-SAM consistently outperforms existing methods across most datasets, which proves that our method can better maintain the strong generalization capacity of SAM. For the 6-bit quantization, PQ-SAM exhibits a performance gap to the full-precision model of less than 0.04 in terms of IoU and Dice metrics across most of the datasets. While PQ-SAM may exhibit slightly lower performance compared to strong baseline methods on a few datasets, it still demonstrates competitive results and preserves the generalization ability overall. For the 4-bit quantization, our method significantly surpasses the baseline methods across all datasets. For instance, PQ-SAM surpasses the second-best method by over $40\%$ in terms of IoU and Dice metrics on the BBBC038V1 dataset for the SAM's ViT-L and ViT-B versions. + +Automatic Mode. In Table 2, we present the quantitative results of different post-quantization methods for the SAM's automatic mode on the BSDS500 dataset. Our PQ-SAM surpasses the baseline methods across three SAM versions and archives a small performance gap to the FP model. + +![](images/7d1ef5aa672446b279bf9fa865e446ad29cc0530d00cf917f25edef7176c9caf.jpg) +Fig. 4: Visual comparison of different PTQ methods for SAM's automatic mode. The blue circle indicates segmentation errors. + +![](images/b2527f7c3411574f300f143e5a807592730fe1e159e79ff8ff84667eee6233cb.jpg) + +![](images/dd9ecf85e9720b0c641c22bc081c4960acb4fb2cad1b356bbf88b41d8af64ed9.jpg) + +![](images/1797122d2348301fef5f9bcbe3b94c1258832105d7abe2bec78715b07c119f2c.jpg) + +![](images/ed37b285a43dee42e0370b4bca16fcf77d7903b766db60c06809fa4de2a6f2ff.jpg) + +![](images/8695153b4341028ce561c947e97ef6787b7d512f2ca8dfcb8242bd2d8b359109.jpg) + +Table 3: Quantitative Results for the bounding box prompt mode of SAM. + +
MethodBitCOCOLVIS
APAPSAPMAPLAPAPSAPMAPL
HVT-HFP3246.430.751.061.744.432.157.265.2
RepQ-Vit [23]643.527.648.159.242.430.054.963.8
Ours645.129.649.560.243.431.555.764.0
RepQ-Vit [23]438.224.741.752.536.025.546.554.1
Ours441.126.245.355.938.127.149.156.9
+ +Table 4: Quantitative comparison for the ultra-low 2bit quantization setting. + +
MethodNDD20DOORSNDISPark
IoUDiceIoUDiceIoUDice
RepQ-ViT [23]0.26810.41460.36770.53010.27800.4224
Ours0.29410.44800.40990.56970.42130.5812
+ +Bounding Box Prompt Mode. In Table 3, we provide the quantitative results of different post-quantization methods for the SAM's bounding box prompt on the COCO and LVIS datasets. Our PQ-SAM outperforms the baseline methods and has a small performance gap compared to the FP model. In the 6-bit quantization setting, PQ-SAM exhibits a loss of accuracy within $3\%$ compared to the FP model. Moreover, in the 4-bit quantization configuration, PQ-SAM shows a significant improvement of $6\%$ compared to the RepQ-Vit method. + +Ultra-low Bit Setting. As shown in Table 4, we validate the superiority of our PQ-SAM on the ultra-low 2-bit quantization setting, which significantly surpasses the state-of-the-art post-training quantization (PTQ) method RepQ-ViT. For instance, our method improves Dice by over $35\%$ compared to RepQ-ViT, demonstrating the effectiveness of our approach for the ultra-low-bit quantization of SAM models. + +Visualization. We provide the visual results of quantized SAM using different quantization methods for the W4A4 setting, including the point prompt mode in Fig. 3 and the auto mode in Fig. 4, respectively. Compared with existing PTQ methods, our proposed method could achieve better segmentation performance with more accurate segmentation masks and fewer segmentation errors. We further give more qualitative results in the supplementary material. + +# 5.4 Ablation Studies + +We perform ablation studies to evaluate the impacts of each component in the proposed PQ-SAM approach using SAM's ViT-B on the BBBC039V1 dataset. + +Table 5: Ablation study on learnable quantization parameters. $\delta_W$ and $\delta_X$ represent the weight and activation quantization step sizes, respectively. $\{v_n,s_n\}_{1\leq n\leq N}$ represents the allocated N pairs of shifting and scaling sizes. + +
ΔW,ΔX{vn}1≤n≤N{sn}1≤n≤NIoUDice
xxx0.05560.0986
x✓✓0.15150.2339
✓xx0.63420.7334
✓✓x0.68470.7721
✓✓✓0.70800.7969
+ +Table 6: Ablation study on activation distribution transformation (ADT) with the outlier hierarchical clustering (OHC). 'Truncating' represents the tensor-level outlier truncating operation. 'Even Groupsize' represents that we evenly allocate the same number of learnable shifting and scaling sizes as ours to each group of adjacent channels. 'Per-Channel' represents that we allocate each pair of learnable shifting and scaling sizes to each channel. + +
ADT | Truncating | Even Groupwise | Per-Channel | OHC (ours) | IoU | Dice
xxxxx0.63420.7334
x✓xxx0.65330.7602
✓✓x✓x0.68420.7720
✓✓✓xx0.68170.7797
✓✓xx✓0.70800.7969
+ +Learnable Quantization Parameters. As presented in Table 5, we perform an ablation study on the learnable quantization parameters in our PQ-SAM framework. This involves considering the weight and activation quantization step sizes, denoted as $\delta_W$ and $\delta_X$ respectively, as well as allocating a set of $N$ pairs of shifting and scaling sizes $\{v_n,s_n\}_{1\leq n\leq N}$ . + +The results demonstrate that introducing learnable quantization step sizes, shifting sizes, or scaling sizes individually resulted in significantly improved performance compared to fixed quantization schemes. Moreover, it is observed that the learnable shifting and scaling sizes yielded even better quantization performance than the learnable quantization step sizes. By optimizing all of these quantization parameters during training, the model is able to effectively adapt to the underlying data distribution and achieve more accurate quantization. + +Activation Distribution Transformation. In Table 6, we present the results of an ablation study to validate the effectiveness of the activation distribution transformation (ADT) in combination with the outlier hierarchical clustering (OHC) scheme. The ADT technique significantly improves the quantization performance, and the OHC scheme further enhances the optimization process and leads to additional improvements in quantization performance. Meanwhile, we validate that the tensor-level outlier truncating operation in ADT can significantly improve the quantization performance. + +To validate the effectiveness of the outlier hierarchical clustering (OHC) scheme, we compare it with two alternative schemes: the "Even Groupsize" and "Per-Channel" schemes. In the "Even Groupsize" scheme, we allocate the same number of learnable shifting and scaling sizes to each group of adjacent channels. In the "Per-Channel" scheme, we allocate a unique pair of learnable shifting and scaling sizes to each individual channel, resulting in the same number of learn- + +Table 7: Ablation study on hyper-parameters and amounts of calibration data. + +
(a) Hyper-parameters of PQ-SAM.
λ61014
IoU0.69010.70800.6858
Dice0.78940.79690.7871
α0.200.250.30
IoU0.68840.70800.6841
Dice0.78120.79690.7693
+ +(b) Amount of calibration data. + +
Number of data100200500
IoU0.70800.70910.7120
Dice0.79690.79800.8014
+ +able shifting and scaling size pairs as there are activation channels. The results of our ablation study demonstrate that the activation distribution transformation (ADT) with OHC in the PQ-SAM framework outperforms those with the "Even Groupsize" or "Per-Channel" schemes. + +Hyperparameter $\lambda$ and $\alpha$ . We conduct ablation studies with extreme parameter settings to showcase the impact of hyperparameters $\lambda$ and $\alpha$ on performance, as shown in Table 7a. The first set of experiments focuses on evaluating the impact of the hyperparameter $\lambda$ on tensor-level outlier truncation. Different values of $\lambda$ are tested for various truncated ranges. The results demonstrate the effects of varying $\lambda$ on the outlier truncation process and its subsequent influence on the overall performance of the PQ-SAM framework. The second set of experiments involves an ablation study on different upper quantiles $\alpha$ for the proposed channel-wise outlier grouping in the outlier hierarchical clustering step. By testing different values of $\alpha$ , different numbers of learnable shifting and scaling sizes are allocated to each group of channels. + +Amount of Calibration Data. We conduct an ablation study on the amount of calibration data. The results in Table 7b show that 100 samples are sufficient for our method, and additional data does not yield significant improvements. + +# 6 Conclusion + +Our proposed PQ-SAM addresses the challenge of resource constraints in deploying the segment anything model on edge devices. By incorporating the grouped activation distribution transformation with outlier hierarchical clustering techniques, PQ-SAM offers a customized post-training quantization (PTQ) solution for SAM. The GADT reshapes the asymmetric activation distribution of SAM, while the OHC scheme effectively handles extreme outliers and optimizes the allocation of shifting and scaling sizes for each group of channels. Our extensive experiments demonstrate that PQ-SAM surpasses existing PTQ methods on nine zero-shot datasets across three modes of SAM, especially for the low-bit quantization settings. + +Acknowledgments. This work was supported in part by the National Natural Science Foundation of China under Grants 62131003 and 62021001. We gratefully acknowledge the support of MindSpore, CANN (Compute Architecture for Neural Networks) and Ascend AI Processor used for this research. + +# References + +1. Arbelaez, P., Maire, M., Fowlkes, C., Malik, J.: Contour detection and hierarchical image segmentation. IEEE transactions on pattern analysis and machine intelligence 33(5), 898-916 (2010) +2. Brown, T., Mann, B., Ryder, N., Subbiah, M., Kaplan, J.D., Dhariwal, P., Neelakantan, A., Shyam, P., Sastry, G., Askell, A., et al.: Language models are few-shot learners. NeurIPS (2020) +3. Caicedo, J.C., Goodman, A., Karhohs, K.W., Cimini, B.A., Ackerman, J., Haghighi, M., Heng, C., Becker, T., Doan, M., McQuin, C., et al.: Nucleus segmentation across imaging experiments: the 2018 data science bowl. Nature methods 16(12), 1247-1253 (2019) +4. Cheng, H.K., Oh, S.W., Price, B., Schwing, A., Lee, J.Y.: Tracking anything with decoupled video segmentation. In: ICCV (2023) +5. Choi, J., Chuang, P.I.J., Wang, Z., Venkataramani, S., Srinivasan, V., Gopalakrishnan, K.: Bridging the accuracy gap for 2-bit quantized neural networks (qnn). arXiv preprint arXiv:1807.06964 (2018) +6. Choi, J., Wang, Z., Venkataramani, S., Chuang, P.I.J., Srinivasan, V., Gopalakrishnan, K.: Pact: Parameterized clipping activation for quantized neural networks. arXiv preprint arXiv:1805.06085 (2018) +7. Chowdhery, A., Narang, S., Devlin, J., Bosma, M., Mishra, G., Roberts, A., Barham, P., Chung, H.W., Sutton, C., Gehrmann, S., et al.: Palm: Scaling language modeling with pathways. arXiv preprint arXiv:2204.02311 (2022) +8. Ciampi, L., Santiago, C., Costeira, J.P., Gennaro, C., Amato, G.: Domain adaptation for traffic density estimation. In: VISAPP (2021) +9. Cordts, M., Omran, M., Ramos, S., Rehfeld, T., Enzweiler, M., Benenson, R., Franke, U., Roth, S., Schiele, B.: The cityscapes dataset for semantic urban scene understanding. In: CVPR (2016) +10. Ding, X., Liu, X., Zhang, Y., Tu, Z., Li, W., Hu, J., Chen, H., Tang, Y., Xiong, Z., Yin, B., et al.: Cbq: Cross-block quantization for large language models. arXiv preprint arXiv:2312.07950 (2023) +1. Gong, S., Zhong, Y., Ma, W., Li, J., Wang, Z., Zhang, J., Heng, P.A., Dou, Q.: 3dsam-adapter: Holistic adaptation of sam from 2d to 3d for promptable medical image segmentation. arXiv preprint arXiv:2306.13465 (2023) +2. Gupta, A., Dollar, P., Girshick, R.: Lvis: A dataset for large vocabulary instance segmentation. In: CVPR (2019) +3. Han, S., Mao, H., Dally, W.J.: Deep compression: Compressing deep neural networks with pruning, trained quantization and huffman coding. arXiv preprint arXiv:1510.00149 (2015) +4. He, K., Chen, X., Xie, S., Li, Y., Dólár, P., Girshick, R.: Masked autoencoders are scalable vision learners. In: CVPR (2022) +5. Hinton, G., Vinyals, O., Dean, J.: Distilling the knowledge in a neural network. arXiv preprint arXiv:1503.02531 (2015) +6. Jacob, B., Kligys, S., Chen, B., Zhu, M., Tang, M., Howard, A., Adam, H., Kalenichenko, D.: Quantization and training of neural networks for efficient integer-arithmetic-only inference. In: CVPR (2018) +7. Jia, C., Yang, Y., Xia, Y., Chen, Y.T., Parekh, Z., Pham, H., Le, Q., Sung, Y.H., Li, Z., Duerig, T.: Scaling up visual and vision-language representation learning with noisy text supervision. In: ICML (2021) + +18. Kingma, D.P., Ba, J.L.: Adam: A method for stochastic optimization. In: ICLR (2015) +19. Kirillov, A., Mintun, E., Ravi, N., Mao, H., Rolland, C., Gustafson, L., Xiao, T., Whitehead, S., Berg, A.C., Lo, W.Y., et al.: Segment anything. In: ICCV (2023) +20. Li, R., Wang, Y., Liang, F., Qin, H., Yan, J., Fan, R.: Fully quantized network for object detection. In: CVPR (2019) +21. Li, Z., Gu, Q.: I-vit: Integer-only quantization for efficient vision transformer inference. In: ICCV (2023) +22. Li, Z., Ma, L., Chen, M., Xiao, J., Gu, Q.: Patch similarity aware data-free quantization for vision transformers. In: ECCV (2022) +23. Li, Z., Xiao, J., Yang, L., Gu, Q.: Repq-vit: Scale reparameterization for posttraining quantization of vision transformers. In: ICCV (2023) +24. Lin, T.Y., Maire, M., Belongie, S., Hays, J., Perona, P., Ramanan, D., Dólar, P., Zitnick, C.L.: Microsoft coco: Common objects in context. In: ECCV (2014) +25. Lin, Y., Zhang, T., Sun, P., Li, Z., Zhou, S.: Fq-vit: Post-training quantization for fully quantized vision transformer. arXiv preprint arXiv:2111.13824 (2021) +26. Liu, X., Cai, M., Chen, Y., Zhang, Y., Shi, T., Zhang, R., Chen, X., Xiong, Z.: Cross-dimension affinity distillation for 3d em neuron segmentation. In: CVPR (2024) +27. Liu, X., Hu, B., Huang, W., Zhang, Y., Xiong, Z.: Efficient biomedical instance segmentation via knowledge distillation. In: MICCAI (2022) +28. Liu, Z., Wang, Y., Han, K., Zhang, W., Ma, S., Gao, W.: Post-training quantization for vision transformer. In: NeurIPS (2021) +29. Loshchilov, I., Hutter, F.: Sgdr: Stochastic gradient descent with warm restarts. arXiv preprint arXiv:1608.03983 (2016) +30. Martin, D., Fowlkes, C., Tal, D., Malik, J.: A database of human segmented natural images and its application to evaluating segmentation algorithms and measuring ecological statistics. In: ICCV (2001) +31. Paszke, A., Gross, S., Massa, F., Lerer, A., Bradbury, J., Chanan, G., Killeen, T., Lin, Z., Gimelshein, N., Antiga, L., et al.: Pytorch: An imperative style, high-performance deep learning library. In: NeurIPS (2019) +32. Pugliatti, M., Topputo, F.: Doors: Dataset for boulders segmentation. Zenodo 9, 20 (2022) +33. Radford, A., Kim, J.W., Hallacy, C., Ramesh, A., Goh, G., Agarwal, S., Sastry, G., Askell, A., Mishkin, P., Clark, J., et al.: Learning transferable visual models from natural language supervision. In: ICML (2021) +34. Ross, S.M.: Introductory statistics. Academic Press (2017) +35. Shamir, R.R., Duchin, Y., Kim, J., Sapiro, G., Harel, N.: Continuous dice coefficient: a method for evaluating probabilistic segmentations. arXiv preprint arXiv:1906.11031 (2019) +36. Shao, W., Chen, M., Zhang, Z., Xu, P., Zhao, L., Li, Z., Zhang, K., Gao, P., Qiao, Y., Luo, P.: Omniquant: Omnidirectionally calibrated quantization for large language models. In: ICLR (2023) +37. Trotter, C., Atkinson, G., Sharpe, M., Richardson, K., McGough, A.S., Wright, N., Burville, B., Berggren, P.: Ndd20: A large-scale few-shot dolphin dataset for coarse and fine-grained categorisation. arXiv preprint arXiv:2005.13359 (2020) +38. Tu, Z., Hu, J., Chen, H., Wang, Y.: Toward accurate post-training quantization for image super resolution. In: CVPR (2023) +39. Upton, G., Cook, I.: Understanding statistics. Oxford University Press (1996) +40. Vasu, P.K.A., Gabriel, J., Zhu, J., Tuzel, O., Ranjan, A.: Fastvit: A fast hybrid vision transformer using structural reparameterization. In: ICCV (2023) + +41. Wei, X., Zhang, Y., Li, Y., Zhang, X., Gong, R., Guo, J., Liu, X.: Outlier suppression+: Accurate quantization of large language models by equivalent and effective shifting and scaling. In: EMNLP (2023) +42. Xie, D., Wang, R., Ma, J., Chen, C., Lu, H., Yang, D., Shi, F., Lin, X.: Edit everything: A text-guided generative system for images editing. arXiv preprint arXiv:2304.14006 (2023) +43. Xie, Y., Liao, Y.: Efficient-vit: A light-weight classification model based on cnn and vit. In: ICIGP (2023) +44. Xiong, Y., Varadarajan, B., Wu, L., Xiang, X., Xiao, F., Zhu, C., Dai, X., Wang, D., Sun, F., Iandola, F., et al.: Efficientsam: Leveraged masked image pretraining for efficient segment anything. In: CVPR (2024) +45. Yang, J., Gao, M., Li, Z., Gao, S., Wang, F., Zheng, F.: Track anything: Segment anything meets videos. arXiv preprint arXiv:2304.11968 (2023) +46. Yang, L., Wei, Y.Z., He, Y., Sun, W., Huang, Z., Huang, H., Fan, H.: ishape: A first step towards irregular shape instance segmentation. arXiv preprint arXiv:2109.15068 (2021) +47. Yang, X., Ye, J., Wang, X.: Factorizing knowledge in neural networks. In: European Conference on Computer Vision. Springer (2022) +48. Yang, X., Zhou, D., Liu, S., Ye, J., Wang, X.: Deep model reassembly. NeurIPS 35, 25739-25753 (2022) +49. Yu, F., Huang, K., Wang, M., Cheng, Y., Chu, W., Cui, L.: Width & depth pruning for vision transformers. In: AAAI (2022) +50. Yuan, Z., Xue, C., Chen, Y., Wu, Q., Sun, G.: Ptq4vit: Post-training quantization for vision transformers with twin uniform quantization. In: ECCV (2022) +51. Zhang, C., Han, D., Qiao, Y., Kim, J.U., Bae, S.H., Lee, S., Hong, C.S.: Faster segment anything: Towards lightweight sam for mobile applications. arXiv preprint arXiv:2306.14289 (2023) +52. Zhang, C., Zhang, C., Li, C., Qiao, Y., Zheng, S., Dam, S.K., Zhang, M., Kim, J.U., Kim, S.T., Choi, J., et al.: One small step for generative ai, one giant leap for agi: A complete survey on chatgpt in aigc era. arXiv preprint arXiv:2304.06488 (2023) +53. Zhang, Q., Liu, X., Li, W., Chen, H., Liu, J., Hu, J., Xiong, Z., Yuan, C., Wang, Y.: Distilling semantic priors from sam to efficient image restoration models. In: CVPR (2024) +54. Zhao, X., Ding, W., An, Y., Du, Y., Yu, T., Li, M., Tang, M., Wang, J.: Fast segment anything. arXiv preprint arXiv:2306.12156 (2023) +55. Zwillinger, D., Kokoska, S.: CRC standard probability and statistics tables and formulae. *Crc Press* (1999) \ No newline at end of file diff --git a/2024/PQ-SAM_ Post-training Quantization for Segment Anything Model/images.zip b/2024/PQ-SAM_ Post-training Quantization for Segment Anything Model/images.zip new file mode 100644 index 0000000000000000000000000000000000000000..64efac01101543b2eb23979bb73ecc539a22b33f --- /dev/null +++ b/2024/PQ-SAM_ Post-training Quantization for Segment Anything Model/images.zip @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:91d8c46f0478caf5a02f65c11acee2cfa82dbbc5a109dc3196957d0b0f104f61 +size 642724 diff --git a/2024/PQ-SAM_ Post-training Quantization for Segment Anything Model/layout.json b/2024/PQ-SAM_ Post-training Quantization for Segment Anything Model/layout.json new file mode 100644 index 0000000000000000000000000000000000000000..96747385f3a5dd59d37bc4d29b17a956f6abaae7 --- /dev/null +++ b/2024/PQ-SAM_ Post-training Quantization for Segment Anything Model/layout.json @@ -0,0 +1,10540 @@ +{ + "pdf_info": [ + { + "para_blocks": [ + { + "bbox": [ + 133, + 111, + 479, + 147 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 111, + 479, + 147 + ], + "spans": [ + { + "bbox": [ + 133, + 111, + 479, + 147 + ], + "type": "text", + "content": "PQ-SAM: Post-training Quantization for Segment Anything Model" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 132, + 167, + 480, + 193 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 167, + 480, + 193 + ], + "spans": [ + { + "bbox": [ + 132, + 167, + 480, + 193 + ], + "type": "text", + "content": "Xiaoyu Liu" + }, + { + "bbox": [ + 132, + 167, + 480, + 193 + ], + "type": "inline_equation", + "content": "^{1}" + }, + { + "bbox": [ + 132, + 167, + 480, + 193 + ], + "type": "text", + "content": ", Xin Ding" + }, + { + "bbox": [ + 132, + 167, + 480, + 193 + ], + "type": "inline_equation", + "content": "^{1}" + }, + { + "bbox": [ + 132, + 167, + 480, + 193 + ], + "type": "text", + "content": ", Lei Yu" + }, + { + "bbox": [ + 132, + 167, + 480, + 193 + ], + "type": "inline_equation", + "content": "^{2}" + }, + { + "bbox": [ + 132, + 167, + 480, + 193 + ], + "type": "text", + "content": ", Yuanyuan Xi" + }, + { + "bbox": [ + 132, + 167, + 480, + 193 + ], + "type": "inline_equation", + "content": "^{2}" + }, + { + "bbox": [ + 132, + 167, + 480, + 193 + ], + "type": "text", + "content": ", Wei Li" + }, + { + "bbox": [ + 132, + 167, + 480, + 193 + ], + "type": "inline_equation", + "content": "^{2}" + }, + { + "bbox": [ + 132, + 167, + 480, + 193 + ], + "type": "text", + "content": ", Zhijun Tu" + }, + { + "bbox": [ + 132, + 167, + 480, + 193 + ], + "type": "inline_equation", + "content": "^{2}" + }, + { + "bbox": [ + 132, + 167, + 480, + 193 + ], + "type": "text", + "content": ", Jie Hu" + }, + { + "bbox": [ + 132, + 167, + 480, + 193 + ], + "type": "inline_equation", + "content": "^{2}" + }, + { + "bbox": [ + 132, + 167, + 480, + 193 + ], + "type": "text", + "content": ", Hanting Chen" + }, + { + "bbox": [ + 132, + 167, + 480, + 193 + ], + "type": "inline_equation", + "content": "^{2}" + }, + { + "bbox": [ + 132, + 167, + 480, + 193 + ], + "type": "text", + "content": ", Baoqun Yin" + }, + { + "bbox": [ + 132, + 167, + 480, + 193 + ], + "type": "inline_equation", + "content": "^{1}" + }, + { + "bbox": [ + 132, + 167, + 480, + 193 + ], + "type": "text", + "content": ", and Zhiwei Xiong" + }, + { + "bbox": [ + 132, + 167, + 480, + 193 + ], + "type": "inline_equation", + "content": "^{1(\\boxtimes)}" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 173, + 201, + 438, + 246 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 173, + 201, + 438, + 246 + ], + "spans": [ + { + "bbox": [ + 173, + 201, + 438, + 246 + ], + "type": "text", + "content": "1 University of Science and Technology of China {liuxyu,xinding64}@mail.ustc.edu.cn, zwxiong@ustc.edu.cn \n2 Huawei Noah's Ark Lab {wei.lee,zhijun.tu}@huawei.com" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 160, + 276, + 455, + 506 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 160, + 276, + 455, + 506 + ], + "spans": [ + { + "bbox": [ + 160, + 276, + 455, + 506 + ], + "type": "text", + "content": "Abstract. Segment anything model (SAM) is a promising prompt-guided vision foundation model to segment objects of interest. However, the extensive computational requirements of SAM have limited its applicability in resource-constraint edge devices. Post-training quantization (PTQ) is an effective potential for fast-deploying SAM. Nevertheless, SAM's billion-scale pretraining creates a highly asymmetric activation distribution with detrimental outliers in excessive channels, resulting in significant performance degradation of the low-bit PTQ. In this paper, we propose PQ-SAM, the first PTQ method customized for SAM. To achieve a quantization-friendly tensor-wise distribution, PQ-SAM incorporates a novel grouped activation distribution transformation (GADT) based on a two-stage outlier hierarchical clustering (OHC) scheme to scale and shift each channel. Firstly, OHC identifies and truncates extreme outliers to reduce the scale variance of different channels. Secondly, OHC iteratively allocates learnable shifting and scaling sizes to each group of channels with similar distributions, reducing the number of learnable parameters and easing the optimization difficulty. These shifting and scaling sizes are used to adjust activation channels, and jointly optimized with quantization step sizes for optimal results. Extensive experiments demonstrate that PQ-SAM outperforms existing PTQ methods on nine zero-shot datasets, and pushes the 4-bit PTQ of SAM to a usable level." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 160, + 517, + 433, + 529 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 160, + 517, + 433, + 529 + ], + "spans": [ + { + "bbox": [ + 160, + 517, + 433, + 529 + ], + "type": "text", + "content": "Keywords: Segment Anything Model-Post-training Quantization" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 132, + 550, + 229, + 562 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 550, + 229, + 562 + ], + "spans": [ + { + "bbox": [ + 132, + 550, + 229, + 562 + ], + "type": "text", + "content": "1 Introduction" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 130, + 575, + 480, + 635 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 575, + 480, + 635 + ], + "spans": [ + { + "bbox": [ + 130, + 575, + 480, + 635 + ], + "type": "text", + "content": "The advent of large language models [2,7,52] has ushered in a new paradigm in natural language processing (NLP). These models exhibit exceptional zero-shot and few-shot generalization, learning broad competencies from sizable pretraining [14,17,33]. Inspired by their success, researchers have begun developing similarly large-scale foundation models for computer vision. A pioneering example" + } + ] + } + ], + "index": 6 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 140, + 643, + 480, + 665 + ], + "type": "page_footnote", + "angle": 0, + "lines": [ + { + "bbox": [ + 140, + 643, + 480, + 665 + ], + "spans": [ + { + "bbox": [ + 140, + 643, + 480, + 665 + ], + "type": "text", + "content": "X. Liu and X. Ding—Equal contribution. This research was conducted while Xiaoyu Liu and Xin Ding were interns at Huawei Noah's Ark Lab." + } + ] + } + ], + "index": 7 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 0 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 185, + 114, + 305, + 168 + ], + "blocks": [ + { + "bbox": [ + 185, + 114, + 305, + 168 + ], + "lines": [ + { + "bbox": [ + 185, + 114, + 305, + 168 + ], + "spans": [ + { + "bbox": [ + 185, + 114, + 305, + 168 + ], + "type": "image", + "image_path": "c17837d20f33e4d9a58e77e2045f76af8c61cecfa1985761feb3972b1a6a67fc.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 220, + 168, + 266, + 175 + ], + "lines": [ + { + "bbox": [ + 220, + 168, + 266, + 175 + ], + "spans": [ + { + "bbox": [ + 220, + 168, + 266, + 175 + ], + "type": "text", + "content": "(a) SAM: block1" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_caption" + } + ], + "index": 2 + }, + { + "type": "image", + "bbox": [ + 307, + 114, + 429, + 168 + ], + "blocks": [ + { + "bbox": [ + 307, + 114, + 429, + 168 + ], + "lines": [ + { + "bbox": [ + 307, + 114, + 429, + 168 + ], + "spans": [ + { + "bbox": [ + 307, + 114, + 429, + 168 + ], + "type": "image", + "image_path": "204ef9c94144ea33b4232596c17feb658e4e9c8f68a10d6be457d30923e823cd.jpg" + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 339, + 168, + 386, + 175 + ], + "lines": [ + { + "bbox": [ + 339, + 168, + 386, + 175 + ], + "spans": [ + { + "bbox": [ + 339, + 168, + 386, + 175 + ], + "type": "text", + "content": "(b) SAM: block3" + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "image_caption" + } + ], + "index": 4 + }, + { + "type": "image", + "bbox": [ + 185, + 176, + 306, + 230 + ], + "blocks": [ + { + "bbox": [ + 185, + 176, + 306, + 230 + ], + "lines": [ + { + "bbox": [ + 185, + 176, + 306, + 230 + ], + "spans": [ + { + "bbox": [ + 185, + 176, + 306, + 230 + ], + "type": "image", + "image_path": "96d05616c1ddd8d2ce65628c8294b6f7496715ceb2a3a51b88ede47b5b1d6e1a.jpg" + } + ] + } + ], + "index": 6, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 204, + 230, + 283, + 237 + ], + "lines": [ + { + "bbox": [ + 204, + 230, + 283, + 237 + ], + "spans": [ + { + "bbox": [ + 204, + 230, + 283, + 237 + ], + "type": "text", + "content": "(c) Conventional ViT: block1" + } + ] + } + ], + "index": 7, + "angle": 0, + "type": "image_caption" + } + ], + "index": 6 + }, + { + "type": "image", + "bbox": [ + 307, + 176, + 429, + 230 + ], + "blocks": [ + { + "bbox": [ + 307, + 176, + 429, + 230 + ], + "lines": [ + { + "bbox": [ + 307, + 176, + 429, + 230 + ], + "spans": [ + { + "bbox": [ + 307, + 176, + 429, + 230 + ], + "type": "image", + "image_path": "37cae7ad5b8b0b1513bf72c81b2ac5d94da74722e575c3a118adbc5a4b865407.jpg" + } + ] + } + ], + "index": 8, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 321, + 230, + 402, + 237 + ], + "lines": [ + { + "bbox": [ + 321, + 230, + 402, + 237 + ], + "spans": [ + { + "bbox": [ + 321, + 230, + 402, + 237 + ], + "type": "text", + "content": "(d) Conventional ViT: block3" + } + ] + } + ], + "index": 9, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 130, + 246, + 482, + 301 + ], + "lines": [ + { + "bbox": [ + 130, + 246, + 482, + 301 + ], + "spans": [ + { + "bbox": [ + 130, + 246, + 482, + 301 + ], + "type": "text", + "content": "Fig. 1: Comparison of activation distributions between image encoder of SAM (a, b) and the conventional ViT (c, d) in two different transformer blocks. Both of these two models are based on the ViT-base version. Each distribution is visualized with the horizontal axis representing different channels and the vertical axis representing the intensity values within that channel." + } + ] + } + ], + "index": 10, + "angle": 0, + "type": "image_caption" + } + ], + "index": 8 + }, + { + "bbox": [ + 130, + 327, + 480, + 376 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 327, + 480, + 376 + ], + "spans": [ + { + "bbox": [ + 130, + 327, + 480, + 376 + ], + "type": "text", + "content": "is the recently proposed segment anything model (SAM) [19], representing the prompt-guided visual foundation model with impressive zero-shot capabilities. SAM focuses on the crucial task of segmenting objects of interest from an image, which underpins many vision applications [4, 11, 42, 45]." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 130, + 377, + 482, + 533 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 377, + 482, + 533 + ], + "spans": [ + { + "bbox": [ + 130, + 377, + 482, + 533 + ], + "type": "text", + "content": "While SAM has demonstrated impressive advancements in segmentation abilities and flexibility, its practical deployment remains challenging. Specifically, SAM's reliance on image encoder of vision transformer (ViT) architecture induces extensive computational overhead that restricts applicability on edge devices with tight latency, power, and memory budgets. This problem necessitates the employment of model compression techniques such as network pruning [13, 49], knowledge distillation [15,26,27,47,48,53], model quantization [5,16,38] and compact architecture design [40,43] etc. Among these techniques, post-training quantization (PTQ) only requires a few unlabeled calibration images without updating model parameters, which enables fast deployment on various devices within a limited time. Given that training and fine-tuning SAM requires significant GPUs using the large-scale training dataset SA1B [19] with 11 million images, there is a growing demand for the PTQ technique." + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 130, + 533, + 482, + 666 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 533, + 482, + 666 + ], + "spans": [ + { + "bbox": [ + 130, + 533, + 482, + 666 + ], + "type": "text", + "content": "Nevertheless, exiting PTQ methods [22, 23, 25, 50] for ViT are focusing on conventional ViT models for classification, which are different from the ViT in SAM. SAM requires accurate prediction for each pixel on multiple zero-shot datasets after billion-scale pretraining, which is much more sensitive to low-bit compression for image embeddings, and its floating-point activations have a highly asymmetric distribution with detrimental outliers. As illustrated in Fig. 1, different from the conventional ViT, our analysis reveals two properties of SAM's activation distribution that pose challenges for quantization: The presence of detrimental outliers implies that the activation values in SAM are not symmetrically distributed around a central tendency. Moreover, SAM's activations often exhibit a concentration of extreme outliers in specific channels, resulting in" + } + ] + } + ], + "index": 13 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "spans": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "type": "text", + "content": "2" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 166, + 91, + 217, + 100 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 166, + 91, + 217, + 100 + ], + "spans": [ + { + "bbox": [ + 166, + 91, + 217, + 100 + ], + "type": "text", + "content": "X. Liu et al." + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 1 + }, + { + "para_blocks": [ + { + "bbox": [ + 130, + 116, + 482, + 212 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 116, + 482, + 212 + ], + "spans": [ + { + "bbox": [ + 130, + 116, + 482, + 212 + ], + "type": "text", + "content": "a considerable number of channels that require adjustment. While the existing PTQ methods [36, 41], such as RepQ-Vit [23], propose to scale activations at the channel level, the practical challenge in SAM arises in two aspects. Firstly, extreme outliers introduce detrimental scale variance in specific activation channels. Secondly, adjusting an excessive number of channels simultaneously proves to be challenging. These distribution properties of SAM's activations pose new challenges for existing PTQ methods, especially for the ultra-low-bit (e.g., 4-bit, 2-bit) quantization." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 130, + 215, + 482, + 431 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 215, + 482, + 431 + ], + "spans": [ + { + "bbox": [ + 130, + 215, + 482, + 431 + ], + "type": "text", + "content": "In this paper, we propose PQ-SAM, the first PTQ method specially designed for SAM, incorporating a grouped activation distribution transformation (GADT) to reshape SAM's activations for quantization-friendly tensor-wise distribution guided by grouped learnable shifting and scaling sizes. Our proposed GADT has a specially designed outlier hierarchical clustering (OHC) scheme to suppress extreme outliers and reduce the total number of learnable parameters in GADT, easing optimization difficulty and preventing overfitting with limited training cost. Specifically, OHC first identifies and truncates tensor-wise extreme outliers across the whole activation tensor by examining interquartile range statistics at the tensor level. This effectively reduces the detrimental scale variance of different channels due to extreme outliers. Secondly, OHC iteratively allocates each pair of learnable shifting and scaling sizes for each group of channels with similar distributions by calculating their quantile range in each iteration. Critically, these shifting and scaling sizes are jointly optimized in an end-to-end manner together with the quantization step sizes for holistic optimal results, under the supervision of the full-precision (FP) SAM model. The optimized shifting and scaling sizes can be reparameterized into corresponding weights so that the proposed PQ-SAM is friendly for practical deployment." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 130, + 434, + 483, + 518 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 434, + 483, + 518 + ], + "spans": [ + { + "bbox": [ + 130, + 434, + 483, + 518 + ], + "type": "text", + "content": "We conduct extensive experiments to demonstrate the superiority of our PQ-SAM over existing advanced PTQ methods, on nine representative zero-shot datasets across three widely used segmentation modes of SAM, i.e., point prompt mode, automatic mode, and bounding box prompt mode. Our PQ-SAM preserves SAM's strong zero-shot performance and achieves usable 4-bit PTQ for SAM, which represents an important step towards unlocking the practical edge deployment of SAM." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 146, + 521, + 351, + 534 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 146, + 521, + 351, + 534 + ], + "spans": [ + { + "bbox": [ + 146, + 521, + 351, + 534 + ], + "type": "text", + "content": "The contributions of this paper are as follows:" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 137, + 547, + 481, + 664 + ], + "type": "list", + "angle": 0, + "index": 10, + "blocks": [ + { + "bbox": [ + 138, + 547, + 481, + 570 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 547, + 481, + 570 + ], + "spans": [ + { + "bbox": [ + 138, + 547, + 481, + 570 + ], + "type": "text", + "content": "- We propose PQ-SAM, the first post-training quantization method customized for the segment anything model with limited computational cost." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 138, + 574, + 480, + 610 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 574, + 480, + 610 + ], + "spans": [ + { + "bbox": [ + 138, + 574, + 480, + 610 + ], + "type": "text", + "content": "- We propose a novel grouped activation distribution transformation (GADT) to reshape SAM's activations with learnable shifting and scaling sizes jointly end-to-end optimized together with quantization step sizes." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 137, + 614, + 480, + 636 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 137, + 614, + 480, + 636 + ], + "spans": [ + { + "bbox": [ + 137, + 614, + 480, + 636 + ], + "type": "text", + "content": "- We design a two-stage outlier hierarchical clustering scheme (OHC) to reduce the learnable parameters and ease the optimization difficulty in GADT." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 137, + 641, + 480, + 664 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 137, + 641, + 480, + 664 + ], + "spans": [ + { + "bbox": [ + 137, + 641, + 480, + 664 + ], + "type": "text", + "content": "- Extensive experiments demonstrate that PQ-SAM significantly outperforms previous PTQ methods on nine zero-shot datasets." + } + ] + } + ], + "index": 9 + } + ], + "sub_type": "text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 408, + 91, + 447, + 102 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 408, + 91, + 447, + 102 + ], + "spans": [ + { + "bbox": [ + 408, + 91, + 447, + 102 + ], + "type": "text", + "content": "PQ-SAM" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 474, + 92, + 481, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 474, + 92, + 481, + 100 + ], + "spans": [ + { + "bbox": [ + 474, + 92, + 481, + 100 + ], + "type": "text", + "content": "3" + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 2 + }, + { + "para_blocks": [ + { + "bbox": [ + 132, + 114, + 237, + 127 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 114, + 237, + 127 + ], + "spans": [ + { + "bbox": [ + 132, + 114, + 237, + 127 + ], + "type": "text", + "content": "2 Related Work" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 132, + 137, + 308, + 148 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 137, + 308, + 148 + ], + "spans": [ + { + "bbox": [ + 132, + 137, + 308, + 148 + ], + "type": "text", + "content": "2.1 Model Compression for SAM" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 130, + 153, + 481, + 273 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 153, + 481, + 273 + ], + "spans": [ + { + "bbox": [ + 130, + 153, + 481, + 273 + ], + "type": "text", + "content": "To address the extensive computational requirements of SAM and make SAM suitable for edge devices, recent works MobileSAM [51] and FastSAM [54], EfficientSAM [44] propose different knowledge distillation based technological routes. MobileSAM reduces SAM's heavy image encoder through knowledge distillation. They replace SAM's encoder with a lightweight CNN and distil the heavyweight encoder's knowledge into the CNN. FastSAM avoids SAM's Transformer architecture entirely by adopting a regular convolution neural network (CNN) detector with a segmentation branch and retraining the CNN using 1/50th of SAM's SA1B dataset. EfficientSAM adopts a two-stage training strategy including preraining and finetuning." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 130, + 274, + 481, + 357 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 274, + 481, + 357 + ], + "spans": [ + { + "bbox": [ + 130, + 274, + 481, + 357 + ], + "type": "text", + "content": "While MobileSAM, FastSAM, and EfficientSAM can compress SAM models, they still require extensive SA1B data and training resources. In contrast, our proposed PQ-SAM technique only requires one millionth of SA1B data for calibration. With just 100 unlabeled images, PQ-SAM achieves efficient quantization of SAM with minimal GPU costs. PQ-SAM provides a highly practical solution to fast deploying SAM using commodity edge resources. Moreover, our PTQ method can be integrated into these methods for better deployment." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 132, + 371, + 261, + 384 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 371, + 261, + 384 + ], + "spans": [ + { + "bbox": [ + 132, + 371, + 261, + 384 + ], + "type": "text", + "content": "2.2 Model Quantization" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 130, + 388, + 483, + 556 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 388, + 483, + 556 + ], + "spans": [ + { + "bbox": [ + 130, + 388, + 483, + 556 + ], + "type": "text", + "content": "Network quantization is an effective compression technique that transforms weight and activation values from floating points to low-bitwidth integers, substantially reducing memory, data movement, and energy costs. Existing methods fall into two categories: Quantization-Aware Training (QAT) [6,16,21] and Post-Training Quantization [10, 16]. However, QAT relies heavily on retraining with the full original dataset, which is infeasible for large vision models like SAM trained on massive data. While PTQ methods using limited calibration data are more practical, current techniques target classification and are not designed for SAM's unique segmentation task. Specifically, SAM exhibits a highly asymmetric distribution with detrimental outliers, which arises from its billion-scale pretraining. Prior PTQ methods for vision transformers [22, 25, 28, 50] do not address these characteristics of the large vision foundation model. To fill this gap, we develop the first customized PTQ approach, PQ-SAM, specifically tailored for quantizing the immense SAM model by accounting for its unusual activation properties." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 132, + 571, + 225, + 585 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 571, + 225, + 585 + ], + "spans": [ + { + "bbox": [ + 132, + 571, + 225, + 585 + ], + "type": "text", + "content": "3 Preliminary" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 130, + 593, + 481, + 628 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 593, + 481, + 628 + ], + "spans": [ + { + "bbox": [ + 130, + 593, + 481, + 628 + ], + "type": "text", + "content": "To provide context before detailing our proposed method, we briefly introduce uniform quantization. We formulate k-bit quantization and dequantization of tensor " + }, + { + "bbox": [ + 130, + 593, + 481, + 628 + ], + "type": "inline_equation", + "content": "x" + }, + { + "bbox": [ + 130, + 593, + 481, + 628 + ], + "type": "text", + "content": " as" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 211, + 632, + 481, + 655 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 211, + 632, + 481, + 655 + ], + "spans": [ + { + "bbox": [ + 211, + 632, + 481, + 655 + ], + "type": "interline_equation", + "content": "\\text {Q u a n t}: x _ {q} = \\operatorname {c l i p} \\left(\\operatorname {r o u n d} \\left(\\frac {x}{\\Delta}\\right) + z, 0, 2 ^ {k} - 1\\right), \\tag {1}", + "image_path": "b536e95c9a820eca64fd7141eed536218377848cbfe6364af22ddc1eddb03c73.jpg" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 201, + 655, + 338, + 667 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 201, + 655, + 338, + 667 + ], + "spans": [ + { + "bbox": [ + 201, + 655, + 338, + 667 + ], + "type": "interline_equation", + "content": "D e q u a n t: x ^ {\\prime} = \\Delta (x _ {q} - z) \\approx x,", + "image_path": "6633daf4258da3b71a971a56ab0a0f648936ed979ea6ebd5e6c009b180491397.jpg" + } + ] + } + ], + "index": 11 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "spans": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "type": "text", + "content": "4" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 166, + 91, + 217, + 100 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 166, + 91, + 217, + 100 + ], + "spans": [ + { + "bbox": [ + 166, + 91, + 217, + 100 + ], + "type": "text", + "content": "X. Liu et al." + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 3 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 133, + 113, + 296, + 240 + ], + "blocks": [ + { + "bbox": [ + 133, + 113, + 296, + 240 + ], + "lines": [ + { + "bbox": [ + 133, + 113, + 296, + 240 + ], + "spans": [ + { + "bbox": [ + 133, + 113, + 296, + 240 + ], + "type": "image", + "image_path": "c34ea5eab4001cb0647d745609d082d058868384ee01079d66eebac38338aa0f.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 130, + 249, + 482, + 282 + ], + "lines": [ + { + "bbox": [ + 130, + 249, + 482, + 282 + ], + "spans": [ + { + "bbox": [ + 130, + 249, + 482, + 282 + ], + "type": "text", + "content": "Fig. 2: The workflow of our proposed PQ-SAM. PQ-SAM introduces grouped activation distribution transformation (GADT) operations into each transformer block of SAM to enable model quantization." + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "image_caption" + } + ], + "index": 2 + }, + { + "type": "image", + "bbox": [ + 300, + 113, + 482, + 241 + ], + "blocks": [ + { + "bbox": [ + 300, + 113, + 482, + 241 + ], + "lines": [ + { + "bbox": [ + 300, + 113, + 482, + 241 + ], + "spans": [ + { + "bbox": [ + 300, + 113, + 482, + 241 + ], + "type": "image", + "image_path": "8e0e94ecd9a29678bcb388a37b7457eb34b576286bb8b6eaff157edde86f24fc.jpg" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_body" + } + ], + "index": 3 + }, + { + "bbox": [ + 130, + 294, + 482, + 366 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 294, + 482, + 366 + ], + "spans": [ + { + "bbox": [ + 130, + 294, + 482, + 366 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 130, + 294, + 482, + 366 + ], + "type": "inline_equation", + "content": "\\operatorname{Round}(\\cdot)" + }, + { + "bbox": [ + 130, + 294, + 482, + 366 + ], + "type": "text", + "content": " projects a value to the k-bit nearest integer around the input, and " + }, + { + "bbox": [ + 130, + 294, + 482, + 366 + ], + "type": "inline_equation", + "content": "\\operatorname{clip}(x,l,u) = \\min (\\max (x,l),u)" + }, + { + "bbox": [ + 130, + 294, + 482, + 366 + ], + "type": "text", + "content": ". " + }, + { + "bbox": [ + 130, + 294, + 482, + 366 + ], + "type": "inline_equation", + "content": "x_{q}" + }, + { + "bbox": [ + 130, + 294, + 482, + 366 + ], + "type": "text", + "content": " represents the quantized value, and " + }, + { + "bbox": [ + 130, + 294, + 482, + 366 + ], + "type": "inline_equation", + "content": "x^{\\prime}" + }, + { + "bbox": [ + 130, + 294, + 482, + 366 + ], + "type": "text", + "content": " are de-quantized value which approximately equals to " + }, + { + "bbox": [ + 130, + 294, + 482, + 366 + ], + "type": "inline_equation", + "content": "x" + }, + { + "bbox": [ + 130, + 294, + 482, + 366 + ], + "type": "text", + "content": ". Note that " + }, + { + "bbox": [ + 130, + 294, + 482, + 366 + ], + "type": "inline_equation", + "content": "\\Delta \\in \\mathbb{R}^{+}" + }, + { + "bbox": [ + 130, + 294, + 482, + 366 + ], + "type": "text", + "content": " represents the quantization step size, and " + }, + { + "bbox": [ + 130, + 294, + 482, + 366 + ], + "type": "inline_equation", + "content": "z\\in \\mathbb{Z}" + }, + { + "bbox": [ + 130, + 294, + 482, + 366 + ], + "type": "text", + "content": " denotes the zero-point which is a per-tensor offset. The values of " + }, + { + "bbox": [ + 130, + 294, + 482, + 366 + ], + "type": "inline_equation", + "content": "\\Delta" + }, + { + "bbox": [ + 130, + 294, + 482, + 366 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 130, + 294, + 482, + 366 + ], + "type": "inline_equation", + "content": "z" + }, + { + "bbox": [ + 130, + 294, + 482, + 366 + ], + "type": "text", + "content": " are determined based on the lower and upper bounds of " + }, + { + "bbox": [ + 130, + 294, + 482, + 366 + ], + "type": "inline_equation", + "content": "x" + }, + { + "bbox": [ + 130, + 294, + 482, + 366 + ], + "type": "text", + "content": " as follows:" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 195, + 373, + 481, + 400 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 195, + 373, + 481, + 400 + ], + "spans": [ + { + "bbox": [ + 195, + 373, + 481, + 400 + ], + "type": "interline_equation", + "content": "\\Delta = \\frac {\\operatorname* {m a x} (x) - \\operatorname* {m i n} (x)}{2 ^ {k} - 1}, \\quad z = \\operatorname {r o u n d} \\left(- \\frac {\\operatorname* {m i n} (x)}{\\Delta}\\right). \\tag {2}", + "image_path": "a9c87c2b71f254597dce2fe61e284eca7b006b1fb7b78059dd7b59f68856ae16.jpg" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 130, + 407, + 481, + 432 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 407, + 481, + 432 + ], + "spans": [ + { + "bbox": [ + 130, + 407, + 481, + 432 + ], + "type": "text", + "content": "Post-training quantization typically focuses on determining the quantization step sizes " + }, + { + "bbox": [ + 130, + 407, + 481, + 432 + ], + "type": "inline_equation", + "content": "\\varDelta_W" + }, + { + "bbox": [ + 130, + 407, + 481, + 432 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 130, + 407, + 481, + 432 + ], + "type": "inline_equation", + "content": "\\varDelta_X" + }, + { + "bbox": [ + 130, + 407, + 481, + 432 + ], + "type": "text", + "content": " for weights and activations per layer." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 130, + 449, + 443, + 464 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 449, + 443, + 464 + ], + "spans": [ + { + "bbox": [ + 130, + 449, + 443, + 464 + ], + "type": "text", + "content": "4 Grouped Activation Distribution Transformation" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 130, + 473, + 482, + 568 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 473, + 482, + 568 + ], + "spans": [ + { + "bbox": [ + 130, + 473, + 482, + 568 + ], + "type": "text", + "content": "Overview. As shown in Fig. 2, PQ-SAM inserts learnable grouped activation distribution transformation (GADT) operations into each transformer block of SAM to reshape the distribution of activations. The shifting and scaling sizes in GADT are jointly optimized with quantization step sizes in an end-to-end manner under full-precision SAM supervision. This achieves holistically optimized quantization parameters tailored for SAM. Importantly, the learnable ADT is readily absorbed into the preceding linear layer, thus introducing no additional computations or parameters after quantization." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 130, + 586, + 291, + 598 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 586, + 291, + 598 + ], + "spans": [ + { + "bbox": [ + 130, + 586, + 291, + 598 + ], + "type": "text", + "content": "4.1 Scaling and Shifting Sizes" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 130, + 605, + 482, + 666 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 605, + 482, + 666 + ], + "spans": [ + { + "bbox": [ + 130, + 605, + 482, + 666 + ], + "type": "text", + "content": "As shown in Fig. 1, the highly asymmetric activation distribution of SAM is problematic for the per-tensor quantization setting, making it challenging to properly determine the quantization parameters " + }, + { + "bbox": [ + 130, + 605, + 482, + 666 + ], + "type": "inline_equation", + "content": "\\varDelta_W" + }, + { + "bbox": [ + 130, + 605, + 482, + 666 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 130, + 605, + 482, + 666 + ], + "type": "inline_equation", + "content": "\\varDelta_X" + }, + { + "bbox": [ + 130, + 605, + 482, + 666 + ], + "type": "text", + "content": ". Moreover, the outliers concentrated in specific channels negatively bias " + }, + { + "bbox": [ + 130, + 605, + 482, + 666 + ], + "type": "inline_equation", + "content": "\\varDelta_W" + }, + { + "bbox": [ + 130, + 605, + 482, + 666 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 130, + 605, + 482, + 666 + ], + "type": "inline_equation", + "content": "\\varDelta_X" + }, + { + "bbox": [ + 130, + 605, + 482, + 666 + ], + "type": "text", + "content": " due to their disproportionate magnitude compared to the majority of activations. To" + } + ] + } + ], + "index": 11 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 408, + 91, + 447, + 102 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 408, + 91, + 447, + 102 + ], + "spans": [ + { + "bbox": [ + 408, + 91, + 447, + 102 + ], + "type": "text", + "content": "PQ-SAM" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 474, + 92, + 480, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 474, + 92, + 480, + 100 + ], + "spans": [ + { + "bbox": [ + 474, + 92, + 480, + 100 + ], + "type": "text", + "content": "5" + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 4 + }, + { + "para_blocks": [ + { + "type": "code", + "bbox": [ + 133, + 132, + 339, + 347 + ], + "blocks": [ + { + "bbox": [ + 140, + 118, + 338, + 131 + ], + "lines": [ + { + "bbox": [ + 140, + 118, + 338, + 131 + ], + "spans": [ + { + "bbox": [ + 140, + 118, + 338, + 131 + ], + "type": "text", + "content": "Algorithm 1: outlier hierarchical clustering" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "code_caption" + }, + { + "bbox": [ + 133, + 132, + 339, + 347 + ], + "lines": [ + { + "bbox": [ + 133, + 132, + 339, + 347 + ], + "spans": [ + { + "bbox": [ + 133, + 132, + 339, + 347 + ], + "type": "text", + "content": "Input: FP SAM with activation " + }, + { + "bbox": [ + 133, + 132, + 339, + 347 + ], + "type": "inline_equation", + "content": "X\\in \\mathbb{R}^{T\\times C_1}" + }, + { + "bbox": [ + 133, + 132, + 339, + 347 + ], + "type": "text", + "content": " Output: " + }, + { + "bbox": [ + 133, + 132, + 339, + 347 + ], + "type": "inline_equation", + "content": "\\{v_{n},s_{n}\\}_{1\\leq n\\leq N}" + }, + { + "bbox": [ + 133, + 132, + 339, + 347 + ], + "type": "text", + "content": " \n1 { Tensor-wise Outlier Truancting.} \n2 Count " + }, + { + "bbox": [ + 133, + 132, + 339, + 347 + ], + "type": "inline_equation", + "content": "X_{Q1}" + }, + { + "bbox": [ + 133, + 132, + 339, + 347 + ], + "type": "inline_equation", + "content": "X_{Q3}" + }, + { + "bbox": [ + 133, + 132, + 339, + 347 + ], + "type": "text", + "content": " and IQR of " + }, + { + "bbox": [ + 133, + 132, + 339, + 347 + ], + "type": "inline_equation", + "content": "X" + }, + { + "bbox": [ + 133, + 132, + 339, + 347 + ], + "type": "text", + "content": " \n3 " + }, + { + "bbox": [ + 133, + 132, + 339, + 347 + ], + "type": "inline_equation", + "content": "X^t\\gets X" + }, + { + "bbox": [ + 133, + 132, + 339, + 347 + ], + "type": "text", + "content": " in Eq. 5. \n4 { Channel-wise Outlier Grouping.} \n5 Initialize the iteration number " + }, + { + "bbox": [ + 133, + 132, + 339, + 347 + ], + "type": "inline_equation", + "content": "n = 0" + }, + { + "bbox": [ + 133, + 132, + 339, + 347 + ], + "type": "text", + "content": " \n6 Allocated channel Group " + }, + { + "bbox": [ + 133, + 132, + 339, + 347 + ], + "type": "inline_equation", + "content": "C_a = \\emptyset" + }, + { + "bbox": [ + 133, + 132, + 339, + 347 + ], + "type": "text", + "content": " \n7 Remaining channel Group " + }, + { + "bbox": [ + 133, + 132, + 339, + 347 + ], + "type": "inline_equation", + "content": "C_r = \\{1,2,\\dots,C_1\\}" + }, + { + "bbox": [ + 133, + 132, + 339, + 347 + ], + "type": "text", + "content": " \n8 while " + }, + { + "bbox": [ + 133, + 132, + 339, + 347 + ], + "type": "inline_equation", + "content": "C_r" + }, + { + "bbox": [ + 133, + 132, + 339, + 347 + ], + "type": "text", + "content": " is not " + }, + { + "bbox": [ + 133, + 132, + 339, + 347 + ], + "type": "inline_equation", + "content": "\\emptyset" + }, + { + "bbox": [ + 133, + 132, + 339, + 347 + ], + "type": "text", + "content": " do \n9 foreach c in " + }, + { + "bbox": [ + 133, + 132, + 339, + 347 + ], + "type": "inline_equation", + "content": "C_r" + }, + { + "bbox": [ + 133, + 132, + 339, + 347 + ], + "type": "text", + "content": " do \n10 if " + }, + { + "bbox": [ + 133, + 132, + 339, + 347 + ], + "type": "inline_equation", + "content": "X_{1 - \\alpha}^{t,n - 1}\\leq X^t (c)\\leq X_\\alpha^{t,n - 1}" + }, + { + "bbox": [ + 133, + 132, + 339, + 347 + ], + "type": "text", + "content": " then \n11 Allocate " + }, + { + "bbox": [ + 133, + 132, + 339, + 347 + ], + "type": "inline_equation", + "content": "\\{v_n,s_n\\} \\to X(c)" + }, + { + "bbox": [ + 133, + 132, + 339, + 347 + ], + "type": "text", + "content": " in Eq. 6; \n12 " + }, + { + "bbox": [ + 133, + 132, + 339, + 347 + ], + "type": "inline_equation", + "content": "C_a.\\mathrm{push}(c);" + }, + { + "bbox": [ + 133, + 132, + 339, + 347 + ], + "type": "text", + "content": " \n13 " + }, + { + "bbox": [ + 133, + 132, + 339, + 347 + ], + "type": "inline_equation", + "content": "C_r.\\mathrm{pop}(c);" + }, + { + "bbox": [ + 133, + 132, + 339, + 347 + ], + "type": "text", + "content": " \n14 end \n15 end \n16 " + }, + { + "bbox": [ + 133, + 132, + 339, + 347 + ], + "type": "inline_equation", + "content": "n = n + 1" + }, + { + "bbox": [ + 133, + 132, + 339, + 347 + ], + "type": "text", + "content": " \n17 end" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "code_body" + } + ], + "index": 3, + "sub_type": "algorithm" + }, + { + "bbox": [ + 130, + 373, + 480, + 396 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 373, + 480, + 396 + ], + "spans": [ + { + "bbox": [ + 130, + 373, + 480, + 396 + ], + "type": "text", + "content": "address this, we propose to transform SAM's activations towards a quantization-friendly distribution at the tensor level." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 130, + 397, + 481, + 468 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 397, + 481, + 468 + ], + "spans": [ + { + "bbox": [ + 130, + 397, + 481, + 468 + ], + "type": "text", + "content": "Straightforwardly, we introduce learnable channel-wise scaling size " + }, + { + "bbox": [ + 130, + 397, + 481, + 468 + ], + "type": "inline_equation", + "content": "S" + }, + { + "bbox": [ + 130, + 397, + 481, + 468 + ], + "type": "text", + "content": " and shifting size " + }, + { + "bbox": [ + 130, + 397, + 481, + 468 + ], + "type": "inline_equation", + "content": "V" + }, + { + "bbox": [ + 130, + 397, + 481, + 468 + ], + "type": "text", + "content": " to scale and shift the activation tensor on a channel-by-channel basis. For a linear layer, suppose that the input of an activation tensor " + }, + { + "bbox": [ + 130, + 397, + 481, + 468 + ], + "type": "inline_equation", + "content": "X \\in \\mathbb{R}^{T \\times C_1}" + }, + { + "bbox": [ + 130, + 397, + 481, + 468 + ], + "type": "text", + "content": " has " + }, + { + "bbox": [ + 130, + 397, + 481, + 468 + ], + "type": "inline_equation", + "content": "T" + }, + { + "bbox": [ + 130, + 397, + 481, + 468 + ], + "type": "text", + "content": " tokens and " + }, + { + "bbox": [ + 130, + 397, + 481, + 468 + ], + "type": "inline_equation", + "content": "C_1" + }, + { + "bbox": [ + 130, + 397, + 481, + 468 + ], + "type": "text", + "content": " channels, and the weight matrix " + }, + { + "bbox": [ + 130, + 397, + 481, + 468 + ], + "type": "inline_equation", + "content": "W \\in \\mathbb{R}^{C_1 \\times C_2}" + }, + { + "bbox": [ + 130, + 397, + 481, + 468 + ], + "type": "text", + "content": " has a bias " + }, + { + "bbox": [ + 130, + 397, + 481, + 468 + ], + "type": "inline_equation", + "content": "B \\in \\mathbb{R}^{1 \\times C_2}" + }, + { + "bbox": [ + 130, + 397, + 481, + 468 + ], + "type": "text", + "content": ". We utilize scaling size " + }, + { + "bbox": [ + 130, + 397, + 481, + 468 + ], + "type": "inline_equation", + "content": "S \\in \\mathbb{R}^{1 \\times C_1}" + }, + { + "bbox": [ + 130, + 397, + 481, + 468 + ], + "type": "text", + "content": " and shifting size " + }, + { + "bbox": [ + 130, + 397, + 481, + 468 + ], + "type": "inline_equation", + "content": "V \\in \\mathbb{R}^{1 \\times C_1}" + }, + { + "bbox": [ + 130, + 397, + 481, + 468 + ], + "type": "text", + "content": " to reshape its distribution, which is formulated as" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 264, + 474, + 481, + 488 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 264, + 474, + 481, + 488 + ], + "spans": [ + { + "bbox": [ + 264, + 474, + 481, + 488 + ], + "type": "interline_equation", + "content": "\\tilde {X} = (X - V) \\oslash S, \\tag {3}", + "image_path": "ad2acbd77eeb5d74c585c97a6a051b7e452b87132594cda62caa59bea80ad69b.jpg" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 130, + 495, + 480, + 533 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 495, + 480, + 533 + ], + "spans": [ + { + "bbox": [ + 130, + 495, + 480, + 533 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 130, + 495, + 480, + 533 + ], + "type": "inline_equation", + "content": "\\widetilde{X}" + }, + { + "bbox": [ + 130, + 495, + 480, + 533 + ], + "type": "text", + "content": " represents the reshaped activation tensor and " + }, + { + "bbox": [ + 130, + 495, + 480, + 533 + ], + "type": "inline_equation", + "content": "\\varnothing" + }, + { + "bbox": [ + 130, + 495, + 480, + 533 + ], + "type": "text", + "content": " represents the division operation. This process is mathematically equivalent to the original linear layer by updating the weight matrix:" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 211, + 537, + 481, + 553 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 211, + 537, + 481, + 553 + ], + "spans": [ + { + "bbox": [ + 211, + 537, + 481, + 553 + ], + "type": "interline_equation", + "content": "X W + B = \\widetilde {X} \\left(W ^ {\\top} \\odot S ^ {\\top}\\right) + \\left(V W ^ {\\top} + B\\right), \\tag {4}", + "image_path": "f773647104809f3ba1da1f2dd14b6ff6eba41387aa7ea8deef2919397a2201eb.jpg" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 130, + 558, + 480, + 581 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 558, + 480, + 581 + ], + "spans": [ + { + "bbox": [ + 130, + 558, + 480, + 581 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 130, + 558, + 480, + 581 + ], + "type": "inline_equation", + "content": "\\odot" + }, + { + "bbox": [ + 130, + 558, + 480, + 581 + ], + "type": "text", + "content": " represents the multiplication operation, and the updated linear layer has a weight matrix " + }, + { + "bbox": [ + 130, + 558, + 480, + 581 + ], + "type": "inline_equation", + "content": "\\widetilde{W} = W^{\\top} \\odot S^{\\top}" + }, + { + "bbox": [ + 130, + 558, + 480, + 581 + ], + "type": "text", + "content": " and a bias " + }, + { + "bbox": [ + 130, + 558, + 480, + 581 + ], + "type": "inline_equation", + "content": "\\widetilde{B} = (VW^{\\top} + B)" + }, + { + "bbox": [ + 130, + 558, + 480, + 581 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 130, + 582, + 481, + 665 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 582, + 481, + 665 + ], + "spans": [ + { + "bbox": [ + 130, + 582, + 481, + 665 + ], + "type": "text", + "content": "Given the practical challenge of optimizing SAM's large number of activation channels, it is difficult to simultaneously optimize the shifting and scaling sizes at the channel level. To address this, we propose Grouped Activation Distribution Transformation (GADT). This technique utilizes our outlier hierarchical clustering scheme to group channels with similar distributions together. By sharing a set of learnable parameters within each group, we effectively reduce the optimization difficulty." + } + ] + } + ], + "index": 10 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "spans": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "type": "text", + "content": "6" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 166, + 91, + 217, + 101 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 166, + 91, + 217, + 101 + ], + "spans": [ + { + "bbox": [ + 166, + 91, + 217, + 101 + ], + "type": "text", + "content": "X. Liu et al." + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 5 + }, + { + "para_blocks": [ + { + "bbox": [ + 132, + 116, + 318, + 128 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 116, + 318, + 128 + ], + "spans": [ + { + "bbox": [ + 132, + 116, + 318, + 128 + ], + "type": "text", + "content": "4.2 Outlier Hierarchical Clustering" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 130, + 134, + 482, + 229 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 134, + 482, + 229 + ], + "spans": [ + { + "bbox": [ + 130, + 134, + 482, + 229 + ], + "type": "text", + "content": "The above-mentioned shifting and scaling sizes are one-on-one corresponding to each activation channel, which leads to excessive learned parameters being optimized and increases optimization difficulty within limited training cost. To reduce learning difficulty and prevent overfitting of the learnable shifting and scaling sizes, we propose an outlier hierarchical clustering (OHC) scheme to iteratively allocate a pair of shifting and scaling sizes to a group of channels. OHC consists of a tensor-wise outlier truncating stage and a channel-wise outlier grouping stage." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 130, + 229, + 482, + 300 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 229, + 482, + 300 + ], + "spans": [ + { + "bbox": [ + 130, + 229, + 482, + 300 + ], + "type": "text", + "content": "Tensor-wise Outlier Truncating. Due to the influence of extreme outliers in the activation values, the grouping of channels with similar distributions can be affected. Additionally, these extreme anomalies can impact the effective usage of quantization intervals. Therefore, as a preliminary step, we perform outlier truncation at the tensor level by examining the outlier situation of numerical values across the entire tensor layer." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 130, + 300, + 482, + 361 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 300, + 482, + 361 + ], + "spans": [ + { + "bbox": [ + 130, + 300, + 482, + 361 + ], + "type": "text", + "content": "Motivated by the statistic theory [34,39,55], we design an effective preprocessing scheme. For the activation tensor " + }, + { + "bbox": [ + 130, + 300, + 482, + 361 + ], + "type": "inline_equation", + "content": "X" + }, + { + "bbox": [ + 130, + 300, + 482, + 361 + ], + "type": "text", + "content": ", we count the upper and lower quartile fractions " + }, + { + "bbox": [ + 130, + 300, + 482, + 361 + ], + "type": "inline_equation", + "content": "Q_{3}" + }, + { + "bbox": [ + 130, + 300, + 482, + 361 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 130, + 300, + 482, + 361 + ], + "type": "inline_equation", + "content": "Q_{1}" + }, + { + "bbox": [ + 130, + 300, + 482, + 361 + ], + "type": "text", + "content": ", and the interquartile range " + }, + { + "bbox": [ + 130, + 300, + 482, + 361 + ], + "type": "inline_equation", + "content": "IQR" + }, + { + "bbox": [ + 130, + 300, + 482, + 361 + ], + "type": "text", + "content": " of its numerical distribution. We then introduce adjustable statistical thresholds, and the truncated activation " + }, + { + "bbox": [ + 130, + 300, + 482, + 361 + ], + "type": "inline_equation", + "content": "X^{t}" + }, + { + "bbox": [ + 130, + 300, + 482, + 361 + ], + "type": "text", + "content": " is formulated as" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 194, + 365, + 481, + 385 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 194, + 365, + 481, + 385 + ], + "spans": [ + { + "bbox": [ + 194, + 365, + 481, + 385 + ], + "type": "interline_equation", + "content": "X ^ {t} = \\left\\{X \\mid X _ {Q _ {1}} - \\lambda I Q R < x < X _ {Q _ {3}} + \\lambda I Q R, x \\in X \\right\\}, \\tag {5}", + "image_path": "307b52bfcb198da9aa7bcbfb23e31cced1bcb1a2fc6e357d30f822ca746f2537.jpg" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 130, + 398, + 481, + 422 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 398, + 481, + 422 + ], + "spans": [ + { + "bbox": [ + 130, + 398, + 481, + 422 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 130, + 398, + 481, + 422 + ], + "type": "inline_equation", + "content": "\\lambda" + }, + { + "bbox": [ + 130, + 398, + 481, + 422 + ], + "type": "text", + "content": " is a hyper-parameter, and " + }, + { + "bbox": [ + 130, + 398, + 481, + 422 + ], + "type": "inline_equation", + "content": "X_{Q_3}" + }, + { + "bbox": [ + 130, + 398, + 481, + 422 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 130, + 398, + 481, + 422 + ], + "type": "inline_equation", + "content": "X_{Q_1}" + }, + { + "bbox": [ + 130, + 398, + 481, + 422 + ], + "type": "text", + "content": " are the upper and lower quartile values at " + }, + { + "bbox": [ + 130, + 398, + 481, + 422 + ], + "type": "inline_equation", + "content": "Q_3" + }, + { + "bbox": [ + 130, + 398, + 481, + 422 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 130, + 398, + 481, + 422 + ], + "type": "inline_equation", + "content": "Q_1" + }, + { + "bbox": [ + 130, + 398, + 481, + 422 + ], + "type": "text", + "content": ", respectively." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 130, + 422, + 490, + 506 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 422, + 490, + 506 + ], + "spans": [ + { + "bbox": [ + 130, + 422, + 490, + 506 + ], + "type": "text", + "content": "Channel-wise Outlier Grouping. For a truncated activation tensor " + }, + { + "bbox": [ + 130, + 422, + 490, + 506 + ], + "type": "inline_equation", + "content": "X_{t}" + }, + { + "bbox": [ + 130, + 422, + 490, + 506 + ], + "type": "text", + "content": " consisting of " + }, + { + "bbox": [ + 130, + 422, + 490, + 506 + ], + "type": "inline_equation", + "content": "C" + }, + { + "bbox": [ + 130, + 422, + 490, + 506 + ], + "type": "text", + "content": " channels, we allocate " + }, + { + "bbox": [ + 130, + 422, + 490, + 506 + ], + "type": "inline_equation", + "content": "N" + }, + { + "bbox": [ + 130, + 422, + 490, + 506 + ], + "type": "text", + "content": " pairs of shifting and scaling sizes " + }, + { + "bbox": [ + 130, + 422, + 490, + 506 + ], + "type": "inline_equation", + "content": "\\{v_{n}, s_{n}\\}_{1 \\leq n \\leq N}" + }, + { + "bbox": [ + 130, + 422, + 490, + 506 + ], + "type": "text", + "content": " to the channels, and " + }, + { + "bbox": [ + 130, + 422, + 490, + 506 + ], + "type": "inline_equation", + "content": "N \\ll C" + }, + { + "bbox": [ + 130, + 422, + 490, + 506 + ], + "type": "text", + "content": ". As depicted in Fig. 2 (b), we perform this allocation iteratively, assigning a pair of shifting and scaling sizes " + }, + { + "bbox": [ + 130, + 422, + 490, + 506 + ], + "type": "inline_equation", + "content": "v_{n}, s_{n}" + }, + { + "bbox": [ + 130, + 422, + 490, + 506 + ], + "type": "text", + "content": " to a group of channels that meet the pre-defined range between the upper quantile " + }, + { + "bbox": [ + 130, + 422, + 490, + 506 + ], + "type": "inline_equation", + "content": "\\alpha" + }, + { + "bbox": [ + 130, + 422, + 490, + 506 + ], + "type": "text", + "content": " and lower quantile " + }, + { + "bbox": [ + 130, + 422, + 490, + 506 + ], + "type": "inline_equation", + "content": "1 - \\alpha" + }, + { + "bbox": [ + 130, + 422, + 490, + 506 + ], + "type": "text", + "content": " of the non-allocating channels. In the " + }, + { + "bbox": [ + 130, + 422, + 490, + 506 + ], + "type": "inline_equation", + "content": "n^{th}" + }, + { + "bbox": [ + 130, + 422, + 490, + 506 + ], + "type": "text", + "content": " iteration, we formulate the procedure of " + }, + { + "bbox": [ + 130, + 422, + 490, + 506 + ], + "type": "inline_equation", + "content": "\\{v_{n}, s_{n}\\}" + }, + { + "bbox": [ + 130, + 422, + 490, + 506 + ], + "type": "text", + "content": " for activations of the " + }, + { + "bbox": [ + 130, + 422, + 490, + 506 + ], + "type": "inline_equation", + "content": "c^{th}" + }, + { + "bbox": [ + 130, + 422, + 490, + 506 + ], + "type": "text", + "content": " channel " + }, + { + "bbox": [ + 130, + 422, + 490, + 506 + ], + "type": "inline_equation", + "content": "X(c)" + }, + { + "bbox": [ + 130, + 422, + 490, + 506 + ], + "type": "text", + "content": " as" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 187, + 511, + 481, + 526 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 187, + 511, + 481, + 526 + ], + "spans": [ + { + "bbox": [ + 187, + 511, + 481, + 526 + ], + "type": "interline_equation", + "content": "\\widetilde {X} (c) = \\left(X (c) - v _ {n}\\right) \\oslash s _ {n}, \\text {i f} X _ {1 - \\alpha} ^ {t, n - 1} \\leq X ^ {t} (c) \\leq X _ {\\alpha} ^ {t, n - 1}, \\tag {6}", + "image_path": "16282e0de5122abb54cd5080e915a8be83d397a445bbded57fa1682abaecade5.jpg" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 130, + 533, + 482, + 569 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 533, + 482, + 569 + ], + "spans": [ + { + "bbox": [ + 130, + 533, + 482, + 569 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 130, + 533, + 482, + 569 + ], + "type": "inline_equation", + "content": "\\widetilde{X}(c)" + }, + { + "bbox": [ + 130, + 533, + 482, + 569 + ], + "type": "text", + "content": " represents the " + }, + { + "bbox": [ + 130, + 533, + 482, + 569 + ], + "type": "inline_equation", + "content": "c^{th}" + }, + { + "bbox": [ + 130, + 533, + 482, + 569 + ], + "type": "text", + "content": " channel of reshaped activation " + }, + { + "bbox": [ + 130, + 533, + 482, + 569 + ], + "type": "inline_equation", + "content": "\\widetilde{X}" + }, + { + "bbox": [ + 130, + 533, + 482, + 569 + ], + "type": "text", + "content": ". Additionally, " + }, + { + "bbox": [ + 130, + 533, + 482, + 569 + ], + "type": "inline_equation", + "content": "X_{1 - \\alpha}^{t,n-1}" + }, + { + "bbox": [ + 130, + 533, + 482, + 569 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 130, + 533, + 482, + 569 + ], + "type": "inline_equation", + "content": "X_{\\alpha}^{t,n-1}" + }, + { + "bbox": [ + 130, + 533, + 482, + 569 + ], + "type": "text", + "content": " respectively denote the lower and upper quantile values, of the activations from these channels that are not allocated in previous iterations." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 130, + 570, + 482, + 666 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 570, + 482, + 666 + ], + "spans": [ + { + "bbox": [ + 130, + 570, + 482, + 666 + ], + "type": "text", + "content": "Overall, the procedure of the proposed outlier hierarchical clustering is summarized in Algorithm 1. This iterative allocation process allows for the dynamic calculation of percentile values for the unallocated channels. By using a fixed value of " + }, + { + "bbox": [ + 130, + 570, + 482, + 666 + ], + "type": "inline_equation", + "content": "\\alpha" + }, + { + "bbox": [ + 130, + 570, + 482, + 666 + ], + "type": "text", + "content": ", we ensure that more similar channels are grouped together at the beginning of the allocation process. As the allocation process advances, the groups gradually become more refined, resulting in finer granularity. This adaptive approach ultimately yields highly effective grouping outcomes, which prove advantageous for subsequent learning and optimization. By conducting the joint" + } + ] + } + ], + "index": 11 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 408, + 91, + 447, + 102 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 408, + 91, + 447, + 102 + ], + "spans": [ + { + "bbox": [ + 408, + 91, + 447, + 102 + ], + "type": "text", + "content": "PQ-SAM" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 474, + 92, + 481, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 474, + 92, + 481, + 100 + ], + "spans": [ + { + "bbox": [ + 474, + 92, + 481, + 100 + ], + "type": "text", + "content": "7" + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 6 + }, + { + "para_blocks": [ + { + "bbox": [ + 130, + 116, + 479, + 140 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 116, + 479, + 140 + ], + "spans": [ + { + "bbox": [ + 130, + 116, + 479, + 140 + ], + "type": "text", + "content": "To optimize of GADT and quantization step sizes, PQ-SAM quantizes SAM with limited overheads on unlabeled samples." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 131, + 157, + 405, + 170 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 131, + 157, + 405, + 170 + ], + "spans": [ + { + "bbox": [ + 131, + 157, + 405, + 170 + ], + "type": "text", + "content": "4.3 Joint Optimization with Quantization Step Sizes" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 130, + 176, + 480, + 260 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 176, + 480, + 260 + ], + "spans": [ + { + "bbox": [ + 130, + 176, + 480, + 260 + ], + "type": "text", + "content": "The shifting and scaling sizes in GADT are learnable and optimized jointly with " + }, + { + "bbox": [ + 130, + 176, + 480, + 260 + ], + "type": "inline_equation", + "content": "\\Delta_W" + }, + { + "bbox": [ + 130, + 176, + 480, + 260 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 130, + 176, + 480, + 260 + ], + "type": "inline_equation", + "content": "\\Delta_X" + }, + { + "bbox": [ + 130, + 176, + 480, + 260 + ], + "type": "text", + "content": " in our PQ-SAM. This integration of learnable shifting and scaling sizes allows for a more effective adjustment of the activation distribution. We use FP model supervision to learn the quantization step sizes " + }, + { + "bbox": [ + 130, + 176, + 480, + 260 + ], + "type": "inline_equation", + "content": "\\Delta_W" + }, + { + "bbox": [ + 130, + 176, + 480, + 260 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 130, + 176, + 480, + 260 + ], + "type": "inline_equation", + "content": "\\Delta_X" + }, + { + "bbox": [ + 130, + 176, + 480, + 260 + ], + "type": "text", + "content": " in an end-to-end manner, and adjust offsets via the above-mentioned GADT, which not only achieves channel-wise offset adjustment but also preserves the hardware-friendly per-tensor quantization setting." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 131, + 261, + 481, + 343 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 131, + 261, + 481, + 343 + ], + "spans": [ + { + "bbox": [ + 131, + 261, + 481, + 343 + ], + "type": "text", + "content": "Given the floating-point weight tensor " + }, + { + "bbox": [ + 131, + 261, + 481, + 343 + ], + "type": "inline_equation", + "content": "W" + }, + { + "bbox": [ + 131, + 261, + 481, + 343 + ], + "type": "text", + "content": " and activation tensor " + }, + { + "bbox": [ + 131, + 261, + 481, + 343 + ], + "type": "inline_equation", + "content": "X" + }, + { + "bbox": [ + 131, + 261, + 481, + 343 + ], + "type": "text", + "content": " in one layer of SAM, we follow the Eq. 1 and quantize them to " + }, + { + "bbox": [ + 131, + 261, + 481, + 343 + ], + "type": "inline_equation", + "content": "k" + }, + { + "bbox": [ + 131, + 261, + 481, + 343 + ], + "type": "text", + "content": "-bit integer values " + }, + { + "bbox": [ + 131, + 261, + 481, + 343 + ], + "type": "inline_equation", + "content": "W_{q}" + }, + { + "bbox": [ + 131, + 261, + 481, + 343 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 131, + 261, + 481, + 343 + ], + "type": "inline_equation", + "content": "X_{q}" + }, + { + "bbox": [ + 131, + 261, + 481, + 343 + ], + "type": "text", + "content": " with the corresponding weight and activation quantization step sizes " + }, + { + "bbox": [ + 131, + 261, + 481, + 343 + ], + "type": "inline_equation", + "content": "\\Delta_W" + }, + { + "bbox": [ + 131, + 261, + 481, + 343 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 131, + 261, + 481, + 343 + ], + "type": "inline_equation", + "content": "\\Delta_X" + }, + { + "bbox": [ + 131, + 261, + 481, + 343 + ], + "type": "text", + "content": ", respectively. Our optimization framework goes beyond previous approaches [23,36,41] that focus on determining " + }, + { + "bbox": [ + 131, + 261, + 481, + 343 + ], + "type": "inline_equation", + "content": "\\Delta_W" + }, + { + "bbox": [ + 131, + 261, + 481, + 343 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 131, + 261, + 481, + 343 + ], + "type": "inline_equation", + "content": "\\Delta_X" + }, + { + "bbox": [ + 131, + 261, + 481, + 343 + ], + "type": "text", + "content": " for each layer individually. Instead, our framework enables a holistic, FP-supervised quantization process that is customized for the entire model." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 130, + 344, + 480, + 416 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 344, + 480, + 416 + ], + "spans": [ + { + "bbox": [ + 130, + 344, + 480, + 416 + ], + "type": "text", + "content": "We jointly optimize the shifting and scaling sizes " + }, + { + "bbox": [ + 130, + 344, + 480, + 416 + ], + "type": "inline_equation", + "content": "\\{v_{n}, s_{n}\\}_{1 \\leq n \\leq N}" + }, + { + "bbox": [ + 130, + 344, + 480, + 416 + ], + "type": "text", + "content": ", and quantization step sizes " + }, + { + "bbox": [ + 130, + 344, + 480, + 416 + ], + "type": "inline_equation", + "content": "\\Delta_{W}" + }, + { + "bbox": [ + 130, + 344, + 480, + 416 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 130, + 344, + 480, + 416 + ], + "type": "inline_equation", + "content": "\\Delta_{X}" + }, + { + "bbox": [ + 130, + 344, + 480, + 416 + ], + "type": "text", + "content": " for holistic optimal results. By leveraging the output " + }, + { + "bbox": [ + 130, + 344, + 480, + 416 + ], + "type": "inline_equation", + "content": "O(W, K)" + }, + { + "bbox": [ + 130, + 344, + 480, + 416 + ], + "type": "text", + "content": " of the FP model as a guide, we can tailor the quantization steps specifically for the pre-trained weights and activations by minimizing the Mean Squared Error (MSE) loss between " + }, + { + "bbox": [ + 130, + 344, + 480, + 416 + ], + "type": "inline_equation", + "content": "O(W, X)" + }, + { + "bbox": [ + 130, + 344, + 480, + 416 + ], + "type": "text", + "content": " and the output " + }, + { + "bbox": [ + 130, + 344, + 480, + 416 + ], + "type": "inline_equation", + "content": "O(\\widetilde{W}_q, \\widetilde{X}_q)" + }, + { + "bbox": [ + 130, + 344, + 480, + 416 + ], + "type": "text", + "content": " of quantized model:" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 205, + 424, + 481, + 449 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 205, + 424, + 481, + 449 + ], + "spans": [ + { + "bbox": [ + 205, + 424, + 481, + 449 + ], + "type": "interline_equation", + "content": "\\underset {\\Delta_ {W}, \\Delta_ {X}, \\left\\{v _ {n}, s _ {n} \\right\\} _ {1 \\leq n \\leq N}} {\\arg \\min } \\| O (\\widetilde {W} _ {q}, \\widetilde {X} _ {q}) - O (W, X) \\| _ {2}, \\tag {7}", + "image_path": "9a5cc82cb691fbac6b3dddbbe5dc9976a34623f6cbcb169d6e157063eda2a326.jpg" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 130, + 457, + 479, + 496 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 457, + 479, + 496 + ], + "spans": [ + { + "bbox": [ + 130, + 457, + 479, + 496 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 130, + 457, + 479, + 496 + ], + "type": "inline_equation", + "content": "\\widetilde{W}_q" + }, + { + "bbox": [ + 130, + 457, + 479, + 496 + ], + "type": "text", + "content": " represent the corresponding quantized weight of the updated weight by reparameterization as demonstrated in Eq. 4, and " + }, + { + "bbox": [ + 130, + 457, + 479, + 496 + ], + "type": "inline_equation", + "content": "\\widetilde{X}_q" + }, + { + "bbox": [ + 130, + 457, + 479, + 496 + ], + "type": "text", + "content": " represents the quantized activation which has been reshaped." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 132, + 514, + 224, + 528 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 514, + 224, + 528 + ], + "spans": [ + { + "bbox": [ + 132, + 514, + 224, + 528 + ], + "type": "text", + "content": "5 Experiment" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 131, + 538, + 269, + 550 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 131, + 538, + 269, + 550 + ], + "spans": [ + { + "bbox": [ + 131, + 538, + 269, + 550 + ], + "type": "text", + "content": "5.1 Datasets and Metrics" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 130, + 558, + 480, + 628 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 558, + 480, + 628 + ], + "spans": [ + { + "bbox": [ + 130, + 558, + 480, + 628 + ], + "type": "text", + "content": "We conduct evaluations on nine representative zero-shot segmentation datasets, focusing on three widely used segmentation modes of SAM, i.e., the point prompt mode, the automatic mode, and the bounding box prompt mode. These datasets have been officially verified by SAM as mentioned in the work [19]. However, since there is no publicly available test code provided by SAM, we independently evaluate these datasets based on our own resources." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 130, + 629, + 481, + 666 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 629, + 481, + 666 + ], + "spans": [ + { + "bbox": [ + 130, + 629, + 481, + 666 + ], + "type": "text", + "content": "Point Prompt Mode. The point prompt specifies what to segment in the image, with the goal of determining the mask containing that point. To fully validate the zero-shot performance of quantized SAM, we subsample six datasets" + } + ] + } + ], + "index": 12 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "spans": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "type": "text", + "content": "8" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 166, + 91, + 217, + 101 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 166, + 91, + 217, + 101 + ], + "spans": [ + { + "bbox": [ + 166, + 91, + 217, + 101 + ], + "type": "text", + "content": "X. Liu et al." + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 7 + }, + { + "para_blocks": [ + { + "type": "table", + "bbox": [ + 136, + 140, + 477, + 412 + ], + "blocks": [ + { + "bbox": [ + 132, + 114, + 479, + 136 + ], + "lines": [ + { + "bbox": [ + 132, + 114, + 479, + 136 + ], + "spans": [ + { + "bbox": [ + 132, + 114, + 479, + 136 + ], + "type": "text", + "content": "Table 1: Quantitative comparison of different PTQ methods for the point prompt mode of SAM." + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 136, + 140, + 477, + 412 + ], + "lines": [ + { + "bbox": [ + 136, + 140, + 477, + 412 + ], + "spans": [ + { + "bbox": [ + 136, + 140, + 477, + 412 + ], + "type": "table", + "html": "
MethodBitBBBC038V1NDD20CityscapeDOORSiShapeNDISPark
IoUDiceIoUDiceIoUDiceIoUDiceIoUDiceIoUDice
FP320.77950.85000.84680.91000.59450.70610.84960.91260.37220.51080.81920.8938
ViT-H
MinMax [16]60.34850.45990.53940.66050.33280.43740.36780.50820.25900.38630.60950.7091
MSE [5]60.65980.73930.55110.63750.44230.55180.61150.70760.34090.48180.71360.8046
Percentile [20]60.72190.72900.76680.84940.50530.61550.78710.86170.36790.50760.78500.8672
FQ-ViT [25]60.44440.60080.25130.38050.28360.41860.28670.42980.10880.19260.41370.5715
RepQ-ViT [23]60.71740.79380.82230.89390.55980.67850.83830.90480.34440.48890.80100.8838
Ours60.76460.83760.82250.89330.56970.68610.84510.90920.33800.47760.81570.8919
MinMax [16]40.33090.46380.27180.40500.19200.29660.46340.61440.11330.20070.25150.3735
MSE [5]40.37610.51410.36460.51870.22650.33940.42650.58350.14450.24730.39330.5408
Percentile [20]40.37350.52000.34470.48110.26190.38910.38880.52730.12530.21690.41530.5657
FQ-ViT [25]40.22130.34840.09720.16360.14530.23870.11680.19830.03590.06780.21030.3341
RepQ-ViT [23]40.70130.78090.76850.85500.45900.57400.75900.84930.27360.40650.73990.8325
Ours40.70810.79730.78180.86700.50560.62840.79750.87660.29210.42940.76470.8541
FP320.77610.84760.84680.90750.59390.70680.85140.91380.34790.48560.81270.8893
ViT-L
MinMax [16]60.51290.62850.52620.63870.36810.48380.41960.56120.28210.41440.64900.7532
MSE [5]60.70100.77640.65210.74080.47570.58880.68570.78080.35440.49640.75830.8489
Percentile [20]60.73680.80660.75110.83200.52300.63640.78050.85740.37750.51950.79400.8759
FQ-ViT [25]60.42130.56600.35960.51430.28400.42060.38400.54530.12050.21130.40980.5675
RepQ-ViT [23]60.69520.77150.80040.87670.55800.67270.82700.89400.31050.44860.78680.8712
Ours60.74250.81540.81800.89020.55660.67350.83700.90360.31990.45930.79900.8786
MinMax [16]40.12560.20530.27640.41610.12850.20270.41510.57750.09480.17070.18180.2868
MSE [5]40.28360.40420.33040.47470.18410.27660.33410.49010.15540.26430.36410.4994
Percentile [20]40.41360.56370.30740.45310.27200.40530.37010.52590.13550.23390.40840.5631
FQ-ViT [25]40.27910.42380.16690.27470.17940.29030.18130.29790.07460.13760.28960.4406
RepQ-ViT [23]40.41290.49830.72350.82500.44220.56220.75520.84830.27910.41030.71000.8118
Ours40.68550.77270.73930.83390.48780.61140.78900.87120.25200.38330.73790.8332
FP320.76660.84170.83220.90000.56920.68840.82940.89870.35360.49870.79330.8760
ViT-B
MinMax [16]60.38290.47770.57270.69760.33560.44450.34950.47970.26850.40450.59040.7021
MSE [5]60.63250.72550.57670.69210.39760.51630.50300.63460.29430.43460.62940.7398
Percentile [20]60.72690.79970.73640.82600.51990.64060.79520.87470.36050.50680.76760.8564
FQ-ViT [25]60.61540.71520.47830.60100.38260.52170.35380.48590.20830.33100.58460.7196
RepQ-ViT [23]60.69000.76750.80700.87300.54490.66440.80530.88010.34390.48560.78360.8696
Ours60.75830.83310.80080.87820.53910.65930.80590.88860.34410.48860.78450.8705
MinMax [16]40.05560.09860.15330.24130.08970.13740.13970.22520.04470.08420.05890.0964
MSE [5]40.33590.46120.37560.53150.22780.33430.46350.62630.13300.23060.36370.5018
Percentile [20]40.39270.53810.40540.55830.27280.40340.42190.57610.13560.23450.41490.5564
FQ-ViT [25]40.34760.50100.16690.27470.21010.32920.22840.35910.11340.19950.31190.4647
RepQ-ViT [23]40.47380.56590.64790.76080.43380.56530.58780.71180.30640.44660.62530.7448
Ours40.70800.79690.67680.78370.47140.59860.73750.83610.31060.45280.68000.7983
", + "image_path": "bd9238fb990241f68ef05c9a3700be03d09f73f0353cd3cbd0a1569f3f9df32c.jpg" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "table_body" + } + ], + "index": 3 + }, + { + "bbox": [ + 130, + 424, + 480, + 532 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 424, + 480, + 532 + ], + "spans": [ + { + "bbox": [ + 130, + 424, + 480, + 532 + ], + "type": "text", + "content": "including BBBC038V1 [3], NDD20 [37], Cityscape [9], DOORS [32], iShape [46], and NDISPark [8]. These datasets contain 15,213 images and 29,978 masks and cover a broad range of domains, which is used for the very comprehensive evaluation experiment. Following the original SAM's experiments, we determine the single-point prompt by selecting the \"center\" of the ground truth. This center is identified as the point with the highest value in the mask's interior distance transform. We calculate the mean IoU and Dice [35] for all masks of each dataset as quantitative metrics, which are common metrics to quantify the overlap between predicted and ground truth masks." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 130, + 533, + 481, + 618 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 533, + 481, + 618 + ], + "spans": [ + { + "bbox": [ + 130, + 533, + 481, + 618 + ], + "type": "text", + "content": "Automatic Mode. SAM's automatic mask generation pipeline can automatically extract a series of masks from an input image with a regular grid of foreground points. We evaluate this automatic mode on the zero-shot edge detection task using the BSDS500 [30] dataset. These automatically generated masks are processed into object edges by the post-processing " + }, + { + "bbox": [ + 130, + 533, + 481, + 618 + ], + "type": "inline_equation", + "content": "^{1}" + }, + { + "bbox": [ + 130, + 533, + 481, + 618 + ], + "type": "text", + "content": ". We conduct evaluations on the test subset, consisting of 200 images, and assess the performance of all methods using four standard metrics " + }, + { + "bbox": [ + 130, + 533, + 481, + 618 + ], + "type": "inline_equation", + "content": "^{2}" + }, + { + "bbox": [ + 130, + 533, + 481, + 618 + ], + "type": "text", + "content": " for edge detection [1]: optimal dataset" + } + ] + } + ], + "index": 5 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 408, + 91, + 447, + 102 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 408, + 91, + 447, + 102 + ], + "spans": [ + { + "bbox": [ + 408, + 91, + 447, + 102 + ], + "type": "text", + "content": "PQ-SAM" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 474, + 92, + 480, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 474, + 92, + 480, + 100 + ], + "spans": [ + { + "bbox": [ + 474, + 92, + 480, + 100 + ], + "type": "text", + "content": "9" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 133, + 632, + 479, + 653 + ], + "type": "page_footnote", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 632, + 479, + 653 + ], + "spans": [ + { + "bbox": [ + 133, + 632, + 479, + 653 + ], + "type": "text", + "content": "1 Our experiments are based on the descriptions in SAM's paper since the official test code is not provided to the public." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 134, + 653, + 421, + 665 + ], + "type": "page_footnote", + "angle": 0, + "lines": [ + { + "bbox": [ + 134, + 653, + 421, + 665 + ], + "spans": [ + { + "bbox": [ + 134, + 653, + 421, + 665 + ], + "type": "inline_equation", + "content": "^{2}" + }, + { + "bbox": [ + 134, + 653, + 421, + 665 + ], + "type": "text", + "content": " https://www2.eecs.berkeley.edu/Research/Projects/CS/vision/bsds/" + } + ] + } + ], + "index": 7 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 8 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 132, + 114, + 481, + 243 + ], + "blocks": [ + { + "bbox": [ + 132, + 114, + 481, + 243 + ], + "lines": [ + { + "bbox": [ + 132, + 114, + 481, + 243 + ], + "spans": [ + { + "bbox": [ + 132, + 114, + 481, + 243 + ], + "type": "image", + "image_path": "ff09578659fe2a3db7d0f84b8d59ff6ef34ca858445b4cfbc222ec61a1d100e7.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 130, + 244, + 482, + 278 + ], + "lines": [ + { + "bbox": [ + 130, + 244, + 482, + 278 + ], + "spans": [ + { + "bbox": [ + 130, + 244, + 482, + 278 + ], + "type": "text", + "content": "Fig. 3: Visual comparison of different PTQ methods for SAM's point prompt mode. The red block and green star represent the object to be segmented and the position of the prompt point, respectively." + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_caption" + } + ], + "index": 2 + }, + { + "bbox": [ + 130, + 285, + 479, + 308 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 285, + 479, + 308 + ], + "spans": [ + { + "bbox": [ + 130, + 285, + 479, + 308 + ], + "type": "text", + "content": "scale (ODS), optimal image scale (OIS), average precision (AP), and recall at " + }, + { + "bbox": [ + 130, + 285, + 479, + 308 + ], + "type": "inline_equation", + "content": "50\\%" + }, + { + "bbox": [ + 130, + 285, + 479, + 308 + ], + "type": "text", + "content": " precision (R50)." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 130, + 309, + 480, + 369 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 309, + 480, + 369 + ], + "spans": [ + { + "bbox": [ + 130, + 309, + 480, + 369 + ], + "type": "text", + "content": "Bounding Box Prompt Mode. Similar to the point prompt mode, the bounding box prompt mode of SAM focuses on segmenting the objects circled by the bounding box. We follow the original setting of SAM [19] and conduct experiments on the COCO [24] and LVIS [12] datasets. The results are evaluated by average precision (AP)." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 132, + 384, + 279, + 396 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 384, + 279, + 396 + ], + "spans": [ + { + "bbox": [ + 132, + 384, + 279, + 396 + ], + "type": "text", + "content": "5.2 Implementation Details" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 130, + 402, + 481, + 498 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 402, + 481, + 498 + ], + "spans": [ + { + "bbox": [ + 130, + 402, + 481, + 498 + ], + "type": "text", + "content": "Training Details. We utilize the PyTorch [31] and conduct all of our experiments on NVIDIA Tesla V100 GPUs. To calibrate our model, we randomly select 100 samples from the SA1B dataset as calibration data and perform the calibration process over 100 epochs, using a batch size of 1. The learnable quantization parameters " + }, + { + "bbox": [ + 130, + 402, + 481, + 498 + ], + "type": "inline_equation", + "content": "\\Delta_W" + }, + { + "bbox": [ + 130, + 402, + 481, + 498 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 130, + 402, + 481, + 498 + ], + "type": "inline_equation", + "content": "\\Delta_X" + }, + { + "bbox": [ + 130, + 402, + 481, + 498 + ], + "type": "text", + "content": ", and " + }, + { + "bbox": [ + 130, + 402, + 481, + 498 + ], + "type": "inline_equation", + "content": "\\{v_n, s_n\\}_{1 \\ll n \\ll N}" + }, + { + "bbox": [ + 130, + 402, + 481, + 498 + ], + "type": "text", + "content": ", are optimized using the Adam optimizer [18] with " + }, + { + "bbox": [ + 130, + 402, + 481, + 498 + ], + "type": "inline_equation", + "content": "\\beta_1 = 0.9" + }, + { + "bbox": [ + 130, + 402, + 481, + 498 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 130, + 402, + 481, + 498 + ], + "type": "inline_equation", + "content": "\\beta_2 = 0.999" + }, + { + "bbox": [ + 130, + 402, + 481, + 498 + ], + "type": "text", + "content": ", with a learning rate of " + }, + { + "bbox": [ + 130, + 402, + 481, + 498 + ], + "type": "inline_equation", + "content": "1e - 4" + }, + { + "bbox": [ + 130, + 402, + 481, + 498 + ], + "type": "text", + "content": ". The learning rate scheduler is CosineAnnealingLR [29]. The hyperparameters " + }, + { + "bbox": [ + 130, + 402, + 481, + 498 + ], + "type": "inline_equation", + "content": "\\lambda" + }, + { + "bbox": [ + 130, + 402, + 481, + 498 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 130, + 402, + 481, + 498 + ], + "type": "inline_equation", + "content": "\\alpha" + }, + { + "bbox": [ + 130, + 402, + 481, + 498 + ], + "type": "text", + "content": " are set as 10 and 0.25, respectively." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 130, + 498, + 481, + 641 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 498, + 481, + 641 + ], + "spans": [ + { + "bbox": [ + 130, + 498, + 481, + 641 + ], + "type": "text", + "content": "Experimental Settings. In our evaluation experiments, we focus on challenging low-bit-width quantization settings, specifically W4A4 (4-bit quantization for weight and activation) and W6A6 (6-bit quantization), for the point prompt mode of SAM. It is important to note that the automatic mode of SAM, which is used for the edge detection task, is less sensitive to network prediction errors due to the incorporation of various post-processing algorithms. Therefore, we specifically perform 4-bit quantization experiments on the BSDS500 dataset for the automatic mode of SAM. Additionally, to validate the superiority of our PQ-SAM approach as discussed in Section 5.3, we conduct experiments in an ultra-low 2-bit setting specifically for the ViT-B version of SAM. This ultra-low-bit setting allowed us to assess the performance and effectiveness of PQ-SAM in scenarios with extremely limited precision." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 130, + 641, + 481, + 665 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 641, + 481, + 665 + ], + "spans": [ + { + "bbox": [ + 130, + 641, + 481, + 665 + ], + "type": "text", + "content": "Baseline Methods. In our comparative analysis, we include two categories of PTQ methods as baselines. The first category consists of classical post-training" + } + ] + } + ], + "index": 9 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "spans": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "type": "text", + "content": "10" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 166, + 91, + 217, + 100 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 166, + 91, + 217, + 100 + ], + "spans": [ + { + "bbox": [ + 166, + 91, + 217, + 100 + ], + "type": "text", + "content": "X. Liu et al." + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 9 + }, + { + "para_blocks": [ + { + "type": "table", + "bbox": [ + 167, + 129, + 446, + 300 + ], + "blocks": [ + { + "bbox": [ + 168, + 114, + 444, + 125 + ], + "lines": [ + { + "bbox": [ + 168, + 114, + 444, + 125 + ], + "spans": [ + { + "bbox": [ + 168, + 114, + 444, + 125 + ], + "type": "text", + "content": "Table 2: Quantitative comparison for the automatic mode of SAM." + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 167, + 129, + 446, + 300 + ], + "lines": [ + { + "bbox": [ + 167, + 129, + 446, + 300 + ], + "spans": [ + { + "bbox": [ + 167, + 129, + 446, + 300 + ], + "type": "table", + "html": "
H-LVMethodBitODSOISAPR50
FP320.7540.7700.7290.865
MinMax [16]40.4300.4400.304-
MSE [5]40.4540.4680.3880.049
Percentile [20]40.4430.4570.339-
FQ-ViT [25]40.4120.4160.230-
RepQ-ViT [23]40.7060.7200.6470.801
Ours40.7360.7450.7070.859
T-LVFP320.7540.7730.7260.866
MinMax [16]40.4230.4240.247-
MSE [5]40.4420.4560.332-
Percentile [20]40.4440.4550.320-
RepQ-ViT [23]40.7090.7220.6450.769
FQ-ViT [25]40.4240.4290.255-
Ours40.7260.7350.6690.843
B-LVFP320.7450.7550.6900.871
MinMax [16]40.3560.3570.145-
MSE [5]40.4250.4800.4750.374
Percentile [20]40.4490.4670.4000.024
FQ-ViT [25]4----
RepQ-ViT [23]40.7010.7120.6320.806
Ours40.7060.7100.6420.861
", + "image_path": "89fbc54915f6c2963b3af8b8a17775f1210b6e9baf3b6d6db1650d007c0b371f.jpg" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "table_body" + } + ], + "index": 3 + }, + { + "bbox": [ + 130, + 320, + 480, + 416 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 320, + 480, + 416 + ], + "spans": [ + { + "bbox": [ + 130, + 320, + 480, + 416 + ], + "type": "text", + "content": "methods, namely MSE [5] and MinMax [16]. These methods are versatile and can be easily applied to different models. The second category comprises quantization methods specifically designed for Vision Transformer (ViT) models, including two state-of-the-art methods RepQ-ViT [23] and FQ-ViT [25]. These methods have gained popularity in post-training quantization and are directly relevant to our evaluation. Specifically, RepQ-ViT is designed for the post-LayerNorm activations with severe inter-channel variation but based on a hand-craft heuristic algorithm, which is not as flexible and effective as our learning-based method." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 132, + 437, + 269, + 449 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 437, + 269, + 449 + ], + "spans": [ + { + "bbox": [ + 132, + 437, + 269, + 449 + ], + "type": "text", + "content": "5.3 Experimental Results" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 130, + 461, + 480, + 616 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 461, + 480, + 616 + ], + "spans": [ + { + "bbox": [ + 130, + 461, + 480, + 616 + ], + "type": "text", + "content": "Point Prompt Mode. In Fig. 3, we present the quantitative results of different post-quantization methods for the SAM's point prompt mode. Overall, our PQ-SAM consistently outperforms existing methods across most datasets, which proves that our method can better maintain the strong generalization capacity of SAM. For the 6-bit quantization, PQ-SAM exhibits a performance gap to the full-precision model of less than 0.04 in terms of IoU and Dice metrics across most of the datasets. While PQ-SAM may exhibit slightly lower performance compared to strong baseline methods on a few datasets, it still demonstrates competitive results and preserves the generalization ability overall. For the 4-bit quantization, our method significantly surpasses the baseline methods across all datasets. For instance, PQ-SAM surpasses the second-best method by over " + }, + { + "bbox": [ + 130, + 461, + 480, + 616 + ], + "type": "inline_equation", + "content": "40\\%" + }, + { + "bbox": [ + 130, + 461, + 480, + 616 + ], + "type": "text", + "content": " in terms of IoU and Dice metrics on the BBBC038V1 dataset for the SAM's ViT-L and ViT-B versions." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 130, + 617, + 481, + 665 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 617, + 481, + 665 + ], + "spans": [ + { + "bbox": [ + 130, + 617, + 481, + 665 + ], + "type": "text", + "content": "Automatic Mode. In Table 2, we present the quantitative results of different post-quantization methods for the SAM's automatic mode on the BSDS500 dataset. Our PQ-SAM surpasses the baseline methods across three SAM versions and archives a small performance gap to the FP model." + } + ] + } + ], + "index": 7 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 408, + 91, + 447, + 102 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 408, + 91, + 447, + 102 + ], + "spans": [ + { + "bbox": [ + 408, + 91, + 447, + 102 + ], + "type": "text", + "content": "PQ-SAM" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 470, + 92, + 479, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 470, + 92, + 479, + 100 + ], + "spans": [ + { + "bbox": [ + 470, + 92, + 479, + 100 + ], + "type": "text", + "content": "11" + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 10 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 133, + 114, + 189, + 179 + ], + "blocks": [ + { + "bbox": [ + 133, + 114, + 189, + 179 + ], + "lines": [ + { + "bbox": [ + 133, + 114, + 189, + 179 + ], + "spans": [ + { + "bbox": [ + 133, + 114, + 189, + 179 + ], + "type": "image", + "image_path": "7d1ef5aa672446b279bf9fa865e446ad29cc0530d00cf917f25edef7176c9caf.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 132, + 182, + 479, + 205 + ], + "lines": [ + { + "bbox": [ + 132, + 182, + 479, + 205 + ], + "spans": [ + { + "bbox": [ + 132, + 182, + 479, + 205 + ], + "type": "text", + "content": "Fig. 4: Visual comparison of different PTQ methods for SAM's automatic mode. The blue circle indicates segmentation errors." + } + ] + } + ], + "index": 8, + "angle": 0, + "type": "image_caption" + } + ], + "index": 2 + }, + { + "type": "image", + "bbox": [ + 190, + 114, + 246, + 180 + ], + "blocks": [ + { + "bbox": [ + 190, + 114, + 246, + 180 + ], + "lines": [ + { + "bbox": [ + 190, + 114, + 246, + 180 + ], + "spans": [ + { + "bbox": [ + 190, + 114, + 246, + 180 + ], + "type": "image", + "image_path": "b2527f7c3411574f300f143e5a807592730fe1e159e79ff8ff84667eee6233cb.jpg" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_body" + } + ], + "index": 3 + }, + { + "type": "image", + "bbox": [ + 249, + 114, + 304, + 180 + ], + "blocks": [ + { + "bbox": [ + 249, + 114, + 304, + 180 + ], + "lines": [ + { + "bbox": [ + 249, + 114, + 304, + 180 + ], + "spans": [ + { + "bbox": [ + 249, + 114, + 304, + 180 + ], + "type": "image", + "image_path": "dd9ecf85e9720b0c641c22bc081c4960acb4fb2cad1b356bbf88b41d8af64ed9.jpg" + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "image_body" + } + ], + "index": 4 + }, + { + "type": "image", + "bbox": [ + 307, + 114, + 365, + 180 + ], + "blocks": [ + { + "bbox": [ + 307, + 114, + 365, + 180 + ], + "lines": [ + { + "bbox": [ + 307, + 114, + 365, + 180 + ], + "spans": [ + { + "bbox": [ + 307, + 114, + 365, + 180 + ], + "type": "image", + "image_path": "1797122d2348301fef5f9bcbe3b94c1258832105d7abe2bec78715b07c119f2c.jpg" + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "image_body" + } + ], + "index": 5 + }, + { + "type": "image", + "bbox": [ + 366, + 114, + 423, + 180 + ], + "blocks": [ + { + "bbox": [ + 366, + 114, + 423, + 180 + ], + "lines": [ + { + "bbox": [ + 366, + 114, + 423, + 180 + ], + "spans": [ + { + "bbox": [ + 366, + 114, + 423, + 180 + ], + "type": "image", + "image_path": "ed37b285a43dee42e0370b4bca16fcf77d7903b766db60c06809fa4de2a6f2ff.jpg" + } + ] + } + ], + "index": 6, + "angle": 0, + "type": "image_body" + } + ], + "index": 6 + }, + { + "type": "image", + "bbox": [ + 426, + 115, + 480, + 180 + ], + "blocks": [ + { + "bbox": [ + 426, + 115, + 480, + 180 + ], + "lines": [ + { + "bbox": [ + 426, + 115, + 480, + 180 + ], + "spans": [ + { + "bbox": [ + 426, + 115, + 480, + 180 + ], + "type": "image", + "image_path": "8695153b4341028ce561c947e97ef6787b7d512f2ca8dfcb8242bd2d8b359109.jpg" + } + ] + } + ], + "index": 7, + "angle": 0, + "type": "image_body" + } + ], + "index": 7 + }, + { + "type": "table", + "bbox": [ + 149, + 227, + 465, + 297 + ], + "blocks": [ + { + "bbox": [ + 151, + 212, + 461, + 224 + ], + "lines": [ + { + "bbox": [ + 151, + 212, + 461, + 224 + ], + "spans": [ + { + "bbox": [ + 151, + 212, + 461, + 224 + ], + "type": "text", + "content": "Table 3: Quantitative Results for the bounding box prompt mode of SAM." + } + ] + } + ], + "index": 9, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 149, + 227, + 465, + 297 + ], + "lines": [ + { + "bbox": [ + 149, + 227, + 465, + 297 + ], + "spans": [ + { + "bbox": [ + 149, + 227, + 465, + 297 + ], + "type": "table", + "html": "
MethodBitCOCOLVIS
APAPSAPMAPLAPAPSAPMAPL
HVT-HFP3246.430.751.061.744.432.157.265.2
RepQ-Vit [23]643.527.648.159.242.430.054.963.8
Ours645.129.649.560.243.431.555.764.0
RepQ-Vit [23]438.224.741.752.536.025.546.554.1
Ours441.126.245.355.938.127.149.156.9
", + "image_path": "735acb3b91c63f6efdc01993a8f6e72da3db6434e673c0bcf8a69125bb44ae96.jpg" + } + ] + } + ], + "index": 10, + "angle": 0, + "type": "table_body" + } + ], + "index": 10 + }, + { + "type": "table", + "bbox": [ + 170, + 318, + 444, + 360 + ], + "blocks": [ + { + "bbox": [ + 147, + 304, + 464, + 316 + ], + "lines": [ + { + "bbox": [ + 147, + 304, + 464, + 316 + ], + "spans": [ + { + "bbox": [ + 147, + 304, + 464, + 316 + ], + "type": "text", + "content": "Table 4: Quantitative comparison for the ultra-low 2bit quantization setting." + } + ] + } + ], + "index": 11, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 170, + 318, + 444, + 360 + ], + "lines": [ + { + "bbox": [ + 170, + 318, + 444, + 360 + ], + "spans": [ + { + "bbox": [ + 170, + 318, + 444, + 360 + ], + "type": "table", + "html": "
MethodNDD20DOORSNDISPark
IoUDiceIoUDiceIoUDice
RepQ-ViT [23]0.26810.41460.36770.53010.27800.4224
Ours0.29410.44800.40990.56970.42130.5812
", + "image_path": "f61db4dfa1ce6325cf40768cad24fa28659017288e5b411fbc2a1cba385a3e4e.jpg" + } + ] + } + ], + "index": 12, + "angle": 0, + "type": "table_body" + } + ], + "index": 12 + }, + { + "bbox": [ + 130, + 368, + 481, + 453 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 368, + 481, + 453 + ], + "spans": [ + { + "bbox": [ + 130, + 368, + 481, + 453 + ], + "type": "text", + "content": "Bounding Box Prompt Mode. In Table 3, we provide the quantitative results of different post-quantization methods for the SAM's bounding box prompt on the COCO and LVIS datasets. Our PQ-SAM outperforms the baseline methods and has a small performance gap compared to the FP model. In the 6-bit quantization setting, PQ-SAM exhibits a loss of accuracy within " + }, + { + "bbox": [ + 130, + 368, + 481, + 453 + ], + "type": "inline_equation", + "content": "3\\%" + }, + { + "bbox": [ + 130, + 368, + 481, + 453 + ], + "type": "text", + "content": " compared to the FP model. Moreover, in the 4-bit quantization configuration, PQ-SAM shows a significant improvement of " + }, + { + "bbox": [ + 130, + 368, + 481, + 453 + ], + "type": "inline_equation", + "content": "6\\%" + }, + { + "bbox": [ + 130, + 368, + 481, + 453 + ], + "type": "text", + "content": " compared to the RepQ-Vit method." + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 130, + 453, + 482, + 524 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 453, + 482, + 524 + ], + "spans": [ + { + "bbox": [ + 130, + 453, + 482, + 524 + ], + "type": "text", + "content": "Ultra-low Bit Setting. As shown in Table 4, we validate the superiority of our PQ-SAM on the ultra-low 2-bit quantization setting, which significantly surpasses the state-of-the-art post-training quantization (PTQ) method RepQ-ViT. For instance, our method improves Dice by over " + }, + { + "bbox": [ + 130, + 453, + 482, + 524 + ], + "type": "inline_equation", + "content": "35\\%" + }, + { + "bbox": [ + 130, + 453, + 482, + 524 + ], + "type": "text", + "content": " compared to RepQ-ViT, demonstrating the effectiveness of our approach for the ultra-low-bit quantization of SAM models." + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 130, + 525, + 482, + 597 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 525, + 482, + 597 + ], + "spans": [ + { + "bbox": [ + 130, + 525, + 482, + 597 + ], + "type": "text", + "content": "Visualization. We provide the visual results of quantized SAM using different quantization methods for the W4A4 setting, including the point prompt mode in Fig. 3 and the auto mode in Fig. 4, respectively. Compared with existing PTQ methods, our proposed method could achieve better segmentation performance with more accurate segmentation masks and fewer segmentation errors. We further give more qualitative results in the supplementary material." + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 132, + 618, + 246, + 629 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 618, + 246, + 629 + ], + "spans": [ + { + "bbox": [ + 132, + 618, + 246, + 629 + ], + "type": "text", + "content": "5.4 Ablation Studies" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 130, + 641, + 481, + 665 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 641, + 481, + 665 + ], + "spans": [ + { + "bbox": [ + 130, + 641, + 481, + 665 + ], + "type": "text", + "content": "We perform ablation studies to evaluate the impacts of each component in the proposed PQ-SAM approach using SAM's ViT-B on the BBBC039V1 dataset." + } + ] + } + ], + "index": 17 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "spans": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "type": "text", + "content": "12" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 166, + 91, + 217, + 101 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 166, + 91, + 217, + 101 + ], + "spans": [ + { + "bbox": [ + 166, + 91, + 217, + 101 + ], + "type": "text", + "content": "X. Liu et al." + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 11 + }, + { + "para_blocks": [ + { + "type": "table", + "bbox": [ + 185, + 153, + 430, + 201 + ], + "blocks": [ + { + "bbox": [ + 130, + 114, + 482, + 148 + ], + "lines": [ + { + "bbox": [ + 130, + 114, + 482, + 148 + ], + "spans": [ + { + "bbox": [ + 130, + 114, + 482, + 148 + ], + "type": "text", + "content": "Table 5: Ablation study on learnable quantization parameters. " + }, + { + "bbox": [ + 130, + 114, + 482, + 148 + ], + "type": "inline_equation", + "content": "\\delta_W" + }, + { + "bbox": [ + 130, + 114, + 482, + 148 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 130, + 114, + 482, + 148 + ], + "type": "inline_equation", + "content": "\\delta_X" + }, + { + "bbox": [ + 130, + 114, + 482, + 148 + ], + "type": "text", + "content": " represent the weight and activation quantization step sizes, respectively. " + }, + { + "bbox": [ + 130, + 114, + 482, + 148 + ], + "type": "inline_equation", + "content": "\\{v_n,s_n\\}_{1\\leq n\\leq N}" + }, + { + "bbox": [ + 130, + 114, + 482, + 148 + ], + "type": "text", + "content": " represents the allocated N pairs of shifting and scaling sizes." + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 185, + 153, + 430, + 201 + ], + "lines": [ + { + "bbox": [ + 185, + 153, + 430, + 201 + ], + "spans": [ + { + "bbox": [ + 185, + 153, + 430, + 201 + ], + "type": "table", + "html": "
ΔW,ΔX{vn}1≤n≤N{sn}1≤n≤NIoUDice
xxx0.05560.0986
x✓✓0.15150.2339
✓xx0.63420.7334
✓✓x0.68470.7721
✓✓✓0.70800.7969
", + "image_path": "10ebe85cdc3e41894d6acc2e6f900cdf334adef16e84a7c80160ee11ef60302c.jpg" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "table_body" + } + ], + "index": 3 + }, + { + "type": "table", + "bbox": [ + 159, + 285, + 455, + 333 + ], + "blocks": [ + { + "bbox": [ + 130, + 213, + 482, + 280 + ], + "lines": [ + { + "bbox": [ + 130, + 213, + 482, + 280 + ], + "spans": [ + { + "bbox": [ + 130, + 213, + 482, + 280 + ], + "type": "text", + "content": "Table 6: Ablation study on activation distribution transformation (ADT) with the outlier hierarchical clustering (OHC). 'Truncating' represents the tensor-level outlier truncating operation. 'Even Groupsize' represents that we evenly allocate the same number of learnable shifting and scaling sizes as ours to each group of adjacent channels. 'Per-Channel' represents that we allocate each pair of learnable shifting and scaling sizes to each channel." + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 159, + 285, + 455, + 333 + ], + "lines": [ + { + "bbox": [ + 159, + 285, + 455, + 333 + ], + "spans": [ + { + "bbox": [ + 159, + 285, + 455, + 333 + ], + "type": "table", + "html": "
ADT | Truncating | Even Groupwise | Per-Channel | OHC (ours) | IoU | Dice
xxxxx0.63420.7334
x✓xxx0.65330.7602
✓✓x✓x0.68420.7720
✓✓✓xx0.68170.7797
✓✓xx✓0.70800.7969
", + "image_path": "01af2aea8995f7c18cbb01d13d6a34d6d9fe77bff41b44d7299290a80712703f.jpg" + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "table_body" + } + ], + "index": 5 + }, + { + "bbox": [ + 130, + 350, + 482, + 411 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 350, + 482, + 411 + ], + "spans": [ + { + "bbox": [ + 130, + 350, + 482, + 411 + ], + "type": "text", + "content": "Learnable Quantization Parameters. As presented in Table 5, we perform an ablation study on the learnable quantization parameters in our PQ-SAM framework. This involves considering the weight and activation quantization step sizes, denoted as " + }, + { + "bbox": [ + 130, + 350, + 482, + 411 + ], + "type": "inline_equation", + "content": "\\delta_W" + }, + { + "bbox": [ + 130, + 350, + 482, + 411 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 130, + 350, + 482, + 411 + ], + "type": "inline_equation", + "content": "\\delta_X" + }, + { + "bbox": [ + 130, + 350, + 482, + 411 + ], + "type": "text", + "content": " respectively, as well as allocating a set of " + }, + { + "bbox": [ + 130, + 350, + 482, + 411 + ], + "type": "inline_equation", + "content": "N" + }, + { + "bbox": [ + 130, + 350, + 482, + 411 + ], + "type": "text", + "content": " pairs of shifting and scaling sizes " + }, + { + "bbox": [ + 130, + 350, + 482, + 411 + ], + "type": "inline_equation", + "content": "\\{v_n,s_n\\}_{1\\leq n\\leq N}" + }, + { + "bbox": [ + 130, + 350, + 482, + 411 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 130, + 412, + 482, + 496 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 412, + 482, + 496 + ], + "spans": [ + { + "bbox": [ + 130, + 412, + 482, + 496 + ], + "type": "text", + "content": "The results demonstrate that introducing learnable quantization step sizes, shifting sizes, or scaling sizes individually resulted in significantly improved performance compared to fixed quantization schemes. Moreover, it is observed that the learnable shifting and scaling sizes yielded even better quantization performance than the learnable quantization step sizes. By optimizing all of these quantization parameters during training, the model is able to effectively adapt to the underlying data distribution and achieve more accurate quantization." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 130, + 497, + 482, + 592 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 497, + 482, + 592 + ], + "spans": [ + { + "bbox": [ + 130, + 497, + 482, + 592 + ], + "type": "text", + "content": "Activation Distribution Transformation. In Table 6, we present the results of an ablation study to validate the effectiveness of the activation distribution transformation (ADT) in combination with the outlier hierarchical clustering (OHC) scheme. The ADT technique significantly improves the quantization performance, and the OHC scheme further enhances the optimization process and leads to additional improvements in quantization performance. Meanwhile, we validate that the tensor-level outlier truncating operation in ADT can significantly improve the quantization performance." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 130, + 594, + 482, + 665 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 594, + 482, + 665 + ], + "spans": [ + { + "bbox": [ + 130, + 594, + 482, + 665 + ], + "type": "text", + "content": "To validate the effectiveness of the outlier hierarchical clustering (OHC) scheme, we compare it with two alternative schemes: the \"Even Groupsize\" and \"Per-Channel\" schemes. In the \"Even Groupsize\" scheme, we allocate the same number of learnable shifting and scaling sizes to each group of adjacent channels. In the \"Per-Channel\" scheme, we allocate a unique pair of learnable shifting and scaling sizes to each individual channel, resulting in the same number of learn-" + } + ] + } + ], + "index": 9 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 408, + 91, + 447, + 102 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 408, + 91, + 447, + 102 + ], + "spans": [ + { + "bbox": [ + 408, + 91, + 447, + 102 + ], + "type": "text", + "content": "PQ-SAM" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 470, + 91, + 481, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 470, + 91, + 481, + 100 + ], + "spans": [ + { + "bbox": [ + 470, + 91, + 481, + 100 + ], + "type": "text", + "content": "13" + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 12 + }, + { + "para_blocks": [ + { + "type": "table", + "bbox": [ + 151, + 131, + 288, + 200 + ], + "blocks": [ + { + "bbox": [ + 143, + 114, + 468, + 125 + ], + "lines": [ + { + "bbox": [ + 143, + 114, + 468, + 125 + ], + "spans": [ + { + "bbox": [ + 143, + 114, + 468, + 125 + ], + "type": "text", + "content": "Table 7: Ablation study on hyper-parameters and amounts of calibration data." + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 151, + 131, + 288, + 200 + ], + "lines": [ + { + "bbox": [ + 151, + 131, + 288, + 200 + ], + "spans": [ + { + "bbox": [ + 151, + 131, + 288, + 200 + ], + "type": "table", + "html": "
(a) Hyper-parameters of PQ-SAM.
λ61014
IoU0.69010.70800.6858
Dice0.78940.79690.7871
α0.200.250.30
IoU0.68840.70800.6841
Dice0.78120.79690.7693
", + "image_path": "8d9f72073fef43cc03e0e4d93a7cd6191a65f2429164aa379776ef4d78c1859e.jpg" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "table_body" + } + ], + "index": 3 + }, + { + "type": "table", + "bbox": [ + 319, + 149, + 465, + 180 + ], + "blocks": [ + { + "bbox": [ + 319, + 131, + 449, + 141 + ], + "lines": [ + { + "bbox": [ + 319, + 131, + 449, + 141 + ], + "spans": [ + { + "bbox": [ + 319, + 131, + 449, + 141 + ], + "type": "text", + "content": "(b) Amount of calibration data." + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 319, + 149, + 465, + 180 + ], + "lines": [ + { + "bbox": [ + 319, + 149, + 465, + 180 + ], + "spans": [ + { + "bbox": [ + 319, + 149, + 465, + 180 + ], + "type": "table", + "html": "
Number of data100200500
IoU0.70800.70910.7120
Dice0.79690.79800.8014
", + "image_path": "5a40c494d4ec1dd76c62ee62a8453c3bef167e4b677804bc6a66e539b2d14717.jpg" + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "table_body" + } + ], + "index": 5 + }, + { + "bbox": [ + 130, + 209, + 479, + 256 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 209, + 479, + 256 + ], + "spans": [ + { + "bbox": [ + 130, + 209, + 479, + 256 + ], + "type": "text", + "content": "able shifting and scaling size pairs as there are activation channels. The results of our ablation study demonstrate that the activation distribution transformation (ADT) with OHC in the PQ-SAM framework outperforms those with the \"Even Groupsize\" or \"Per-Channel\" schemes." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 130, + 257, + 480, + 388 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 257, + 480, + 388 + ], + "spans": [ + { + "bbox": [ + 130, + 257, + 480, + 388 + ], + "type": "text", + "content": "Hyperparameter " + }, + { + "bbox": [ + 130, + 257, + 480, + 388 + ], + "type": "inline_equation", + "content": "\\lambda" + }, + { + "bbox": [ + 130, + 257, + 480, + 388 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 130, + 257, + 480, + 388 + ], + "type": "inline_equation", + "content": "\\alpha" + }, + { + "bbox": [ + 130, + 257, + 480, + 388 + ], + "type": "text", + "content": ". We conduct ablation studies with extreme parameter settings to showcase the impact of hyperparameters " + }, + { + "bbox": [ + 130, + 257, + 480, + 388 + ], + "type": "inline_equation", + "content": "\\lambda" + }, + { + "bbox": [ + 130, + 257, + 480, + 388 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 130, + 257, + 480, + 388 + ], + "type": "inline_equation", + "content": "\\alpha" + }, + { + "bbox": [ + 130, + 257, + 480, + 388 + ], + "type": "text", + "content": " on performance, as shown in Table 7a. The first set of experiments focuses on evaluating the impact of the hyperparameter " + }, + { + "bbox": [ + 130, + 257, + 480, + 388 + ], + "type": "inline_equation", + "content": "\\lambda" + }, + { + "bbox": [ + 130, + 257, + 480, + 388 + ], + "type": "text", + "content": " on tensor-level outlier truncation. Different values of " + }, + { + "bbox": [ + 130, + 257, + 480, + 388 + ], + "type": "inline_equation", + "content": "\\lambda" + }, + { + "bbox": [ + 130, + 257, + 480, + 388 + ], + "type": "text", + "content": " are tested for various truncated ranges. The results demonstrate the effects of varying " + }, + { + "bbox": [ + 130, + 257, + 480, + 388 + ], + "type": "inline_equation", + "content": "\\lambda" + }, + { + "bbox": [ + 130, + 257, + 480, + 388 + ], + "type": "text", + "content": " on the outlier truncation process and its subsequent influence on the overall performance of the PQ-SAM framework. The second set of experiments involves an ablation study on different upper quantiles " + }, + { + "bbox": [ + 130, + 257, + 480, + 388 + ], + "type": "inline_equation", + "content": "\\alpha" + }, + { + "bbox": [ + 130, + 257, + 480, + 388 + ], + "type": "text", + "content": " for the proposed channel-wise outlier grouping in the outlier hierarchical clustering step. By testing different values of " + }, + { + "bbox": [ + 130, + 257, + 480, + 388 + ], + "type": "inline_equation", + "content": "\\alpha" + }, + { + "bbox": [ + 130, + 257, + 480, + 388 + ], + "type": "text", + "content": ", different numbers of learnable shifting and scaling sizes are allocated to each group of channels." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 130, + 388, + 480, + 425 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 388, + 480, + 425 + ], + "spans": [ + { + "bbox": [ + 130, + 388, + 480, + 425 + ], + "type": "text", + "content": "Amount of Calibration Data. We conduct an ablation study on the amount of calibration data. The results in Table 7b show that 100 samples are sufficient for our method, and additional data does not yield significant improvements." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 131, + 443, + 220, + 455 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 131, + 443, + 220, + 455 + ], + "spans": [ + { + "bbox": [ + 131, + 443, + 220, + 455 + ], + "type": "text", + "content": "6 Conclusion" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 130, + 468, + 480, + 588 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 468, + 480, + 588 + ], + "spans": [ + { + "bbox": [ + 130, + 468, + 480, + 588 + ], + "type": "text", + "content": "Our proposed PQ-SAM addresses the challenge of resource constraints in deploying the segment anything model on edge devices. By incorporating the grouped activation distribution transformation with outlier hierarchical clustering techniques, PQ-SAM offers a customized post-training quantization (PTQ) solution for SAM. The GADT reshapes the asymmetric activation distribution of SAM, while the OHC scheme effectively handles extreme outliers and optimizes the allocation of shifting and scaling sizes for each group of channels. Our extensive experiments demonstrate that PQ-SAM surpasses existing PTQ methods on nine zero-shot datasets across three modes of SAM, especially for the low-bit quantization settings." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 130, + 600, + 480, + 647 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 600, + 480, + 647 + ], + "spans": [ + { + "bbox": [ + 130, + 600, + 480, + 647 + ], + "type": "text", + "content": "Acknowledgments. This work was supported in part by the National Natural Science Foundation of China under Grants 62131003 and 62021001. We gratefully acknowledge the support of MindSpore, CANN (Compute Architecture for Neural Networks) and Ascend AI Processor used for this research." + } + ] + } + ], + "index": 11 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "spans": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "type": "text", + "content": "14" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 166, + 91, + 217, + 100 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 166, + 91, + 217, + 100 + ], + "spans": [ + { + "bbox": [ + 166, + 91, + 217, + 100 + ], + "type": "text", + "content": "X. Liu et al." + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 13 + }, + { + "para_blocks": [ + { + "bbox": [ + 133, + 114, + 197, + 126 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 114, + 197, + 126 + ], + "spans": [ + { + "bbox": [ + 133, + 114, + 197, + 126 + ], + "type": "text", + "content": "References" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 138, + 141, + 480, + 665 + ], + "type": "list", + "angle": 0, + "index": 20, + "blocks": [ + { + "bbox": [ + 138, + 141, + 480, + 175 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 141, + 480, + 175 + ], + "spans": [ + { + "bbox": [ + 138, + 141, + 480, + 175 + ], + "type": "text", + "content": "1. Arbelaez, P., Maire, M., Fowlkes, C., Malik, J.: Contour detection and hierarchical image segmentation. IEEE transactions on pattern analysis and machine intelligence 33(5), 898-916 (2010)" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 138, + 176, + 480, + 208 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 176, + 480, + 208 + ], + "spans": [ + { + "bbox": [ + 138, + 176, + 480, + 208 + ], + "type": "text", + "content": "2. Brown, T., Mann, B., Ryder, N., Subbiah, M., Kaplan, J.D., Dhariwal, P., Neelakantan, A., Shyam, P., Sastry, G., Askell, A., et al.: Language models are few-shot learners. NeurIPS (2020)" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 138, + 209, + 480, + 252 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 209, + 480, + 252 + ], + "spans": [ + { + "bbox": [ + 138, + 209, + 480, + 252 + ], + "type": "text", + "content": "3. Caicedo, J.C., Goodman, A., Karhohs, K.W., Cimini, B.A., Ackerman, J., Haghighi, M., Heng, C., Becker, T., Doan, M., McQuin, C., et al.: Nucleus segmentation across imaging experiments: the 2018 data science bowl. Nature methods 16(12), 1247-1253 (2019)" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 138, + 254, + 480, + 274 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 254, + 480, + 274 + ], + "spans": [ + { + "bbox": [ + 138, + 254, + 480, + 274 + ], + "type": "text", + "content": "4. Cheng, H.K., Oh, S.W., Price, B., Schwing, A., Lee, J.Y.: Tracking anything with decoupled video segmentation. In: ICCV (2023)" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 138, + 276, + 480, + 308 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 276, + 480, + 308 + ], + "spans": [ + { + "bbox": [ + 138, + 276, + 480, + 308 + ], + "type": "text", + "content": "5. Choi, J., Chuang, P.I.J., Wang, Z., Venkataramani, S., Srinivasan, V., Gopalakrishnan, K.: Bridging the accuracy gap for 2-bit quantized neural networks (qnn). arXiv preprint arXiv:1807.06964 (2018)" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 138, + 309, + 480, + 342 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 309, + 480, + 342 + ], + "spans": [ + { + "bbox": [ + 138, + 309, + 480, + 342 + ], + "type": "text", + "content": "6. Choi, J., Wang, Z., Venkataramani, S., Chuang, P.I.J., Srinivasan, V., Gopalakrishnan, K.: Pact: Parameterized clipping activation for quantized neural networks. arXiv preprint arXiv:1805.06085 (2018)" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 138, + 343, + 480, + 374 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 343, + 480, + 374 + ], + "spans": [ + { + "bbox": [ + 138, + 343, + 480, + 374 + ], + "type": "text", + "content": "7. Chowdhery, A., Narang, S., Devlin, J., Bosma, M., Mishra, G., Roberts, A., Barham, P., Chung, H.W., Sutton, C., Gehrmann, S., et al.: Palm: Scaling language modeling with pathways. arXiv preprint arXiv:2204.02311 (2022)" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 138, + 376, + 480, + 397 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 376, + 480, + 397 + ], + "spans": [ + { + "bbox": [ + 138, + 376, + 480, + 397 + ], + "type": "text", + "content": "8. Ciampi, L., Santiago, C., Costeira, J.P., Gennaro, C., Amato, G.: Domain adaptation for traffic density estimation. In: VISAPP (2021)" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 138, + 399, + 480, + 430 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 399, + 480, + 430 + ], + "spans": [ + { + "bbox": [ + 138, + 399, + 480, + 430 + ], + "type": "text", + "content": "9. Cordts, M., Omran, M., Ramos, S., Rehfeld, T., Enzweiler, M., Benenson, R., Franke, U., Roth, S., Schiele, B.: The cityscapes dataset for semantic urban scene understanding. In: CVPR (2016)" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 138, + 432, + 480, + 464 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 432, + 480, + 464 + ], + "spans": [ + { + "bbox": [ + 138, + 432, + 480, + 464 + ], + "type": "text", + "content": "10. Ding, X., Liu, X., Zhang, Y., Tu, Z., Li, W., Hu, J., Chen, H., Tang, Y., Xiong, Z., Yin, B., et al.: Cbq: Cross-block quantization for large language models. arXiv preprint arXiv:2312.07950 (2023)" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 138, + 465, + 480, + 497 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 465, + 480, + 497 + ], + "spans": [ + { + "bbox": [ + 138, + 465, + 480, + 497 + ], + "type": "text", + "content": "1. Gong, S., Zhong, Y., Ma, W., Li, J., Wang, Z., Zhang, J., Heng, P.A., Dou, Q.: 3dsam-adapter: Holistic adaptation of sam from 2d to 3d for promptable medical image segmentation. arXiv preprint arXiv:2306.13465 (2023)" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 138, + 498, + 480, + 519 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 498, + 480, + 519 + ], + "spans": [ + { + "bbox": [ + 138, + 498, + 480, + 519 + ], + "type": "text", + "content": "2. Gupta, A., Dollar, P., Girshick, R.: Lvis: A dataset for large vocabulary instance segmentation. In: CVPR (2019)" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 138, + 521, + 480, + 552 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 521, + 480, + 552 + ], + "spans": [ + { + "bbox": [ + 138, + 521, + 480, + 552 + ], + "type": "text", + "content": "3. Han, S., Mao, H., Dally, W.J.: Deep compression: Compressing deep neural networks with pruning, trained quantization and huffman coding. arXiv preprint arXiv:1510.00149 (2015)" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 138, + 554, + 480, + 575 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 554, + 480, + 575 + ], + "spans": [ + { + "bbox": [ + 138, + 554, + 480, + 575 + ], + "type": "text", + "content": "4. He, K., Chen, X., Xie, S., Li, Y., Dólár, P., Girshick, R.: Masked autoencoders are scalable vision learners. In: CVPR (2022)" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 138, + 577, + 480, + 597 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 577, + 480, + 597 + ], + "spans": [ + { + "bbox": [ + 138, + 577, + 480, + 597 + ], + "type": "text", + "content": "5. Hinton, G., Vinyals, O., Dean, J.: Distilling the knowledge in a neural network. arXiv preprint arXiv:1503.02531 (2015)" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 138, + 599, + 480, + 631 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 599, + 480, + 631 + ], + "spans": [ + { + "bbox": [ + 138, + 599, + 480, + 631 + ], + "type": "text", + "content": "6. Jacob, B., Kligys, S., Chen, B., Zhu, M., Tang, M., Howard, A., Adam, H., Kalenichenko, D.: Quantization and training of neural networks for efficient integer-arithmetic-only inference. In: CVPR (2018)" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 138, + 632, + 480, + 665 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 632, + 480, + 665 + ], + "spans": [ + { + "bbox": [ + 138, + 632, + 480, + 665 + ], + "type": "text", + "content": "7. Jia, C., Yang, Y., Xia, Y., Chen, Y.T., Parekh, Z., Pham, H., Le, Q., Sung, Y.H., Li, Z., Duerig, T.: Scaling up visual and vision-language representation learning with noisy text supervision. In: ICML (2021)" + } + ] + } + ], + "index": 19 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 408, + 91, + 447, + 102 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 408, + 91, + 447, + 102 + ], + "spans": [ + { + "bbox": [ + 408, + 91, + 447, + 102 + ], + "type": "text", + "content": "PQ-SAM" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 470, + 92, + 481, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 470, + 92, + 481, + 100 + ], + "spans": [ + { + "bbox": [ + 470, + 92, + 481, + 100 + ], + "type": "text", + "content": "15" + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 14 + }, + { + "para_blocks": [ + { + "bbox": [ + 132, + 116, + 480, + 666 + ], + "type": "list", + "angle": 0, + "index": 25, + "blocks": [ + { + "bbox": [ + 133, + 116, + 480, + 138 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 116, + 480, + 138 + ], + "spans": [ + { + "bbox": [ + 133, + 116, + 480, + 138 + ], + "type": "text", + "content": "18. Kingma, D.P., Ba, J.L.: Adam: A method for stochastic optimization. In: ICLR (2015)" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 132, + 138, + 480, + 159 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 138, + 480, + 159 + ], + "spans": [ + { + "bbox": [ + 132, + 138, + 480, + 159 + ], + "type": "text", + "content": "19. Kirillov, A., Mintun, E., Ravi, N., Mao, H., Rolland, C., Gustafson, L., Xiao, T., Whitehead, S., Berg, A.C., Lo, W.Y., et al.: Segment anything. In: ICCV (2023)" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 132, + 160, + 480, + 182 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 160, + 480, + 182 + ], + "spans": [ + { + "bbox": [ + 132, + 160, + 480, + 182 + ], + "type": "text", + "content": "20. Li, R., Wang, Y., Liang, F., Qin, H., Yan, J., Fan, R.: Fully quantized network for object detection. In: CVPR (2019)" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 132, + 182, + 480, + 202 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 182, + 480, + 202 + ], + "spans": [ + { + "bbox": [ + 132, + 182, + 480, + 202 + ], + "type": "text", + "content": "21. Li, Z., Gu, Q.: I-vit: Integer-only quantization for efficient vision transformer inference. In: ICCV (2023)" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 132, + 203, + 480, + 224 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 203, + 480, + 224 + ], + "spans": [ + { + "bbox": [ + 132, + 203, + 480, + 224 + ], + "type": "text", + "content": "22. Li, Z., Ma, L., Chen, M., Xiao, J., Gu, Q.: Patch similarity aware data-free quantization for vision transformers. In: ECCV (2022)" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 132, + 224, + 480, + 246 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 224, + 480, + 246 + ], + "spans": [ + { + "bbox": [ + 132, + 224, + 480, + 246 + ], + "type": "text", + "content": "23. Li, Z., Xiao, J., Yang, L., Gu, Q.: Repq-vit: Scale reparameterization for posttraining quantization of vision transformers. In: ICCV (2023)" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 132, + 246, + 480, + 267 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 246, + 480, + 267 + ], + "spans": [ + { + "bbox": [ + 132, + 246, + 480, + 267 + ], + "type": "text", + "content": "24. Lin, T.Y., Maire, M., Belongie, S., Hays, J., Perona, P., Ramanan, D., Dólar, P., Zitnick, C.L.: Microsoft coco: Common objects in context. In: ECCV (2014)" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 132, + 267, + 480, + 289 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 267, + 480, + 289 + ], + "spans": [ + { + "bbox": [ + 132, + 267, + 480, + 289 + ], + "type": "text", + "content": "25. Lin, Y., Zhang, T., Sun, P., Li, Z., Zhou, S.: Fq-vit: Post-training quantization for fully quantized vision transformer. arXiv preprint arXiv:2111.13824 (2021)" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 132, + 289, + 480, + 320 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 289, + 480, + 320 + ], + "spans": [ + { + "bbox": [ + 132, + 289, + 480, + 320 + ], + "type": "text", + "content": "26. Liu, X., Cai, M., Chen, Y., Zhang, Y., Shi, T., Zhang, R., Chen, X., Xiong, Z.: Cross-dimension affinity distillation for 3d em neuron segmentation. In: CVPR (2024)" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 132, + 321, + 480, + 342 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 321, + 480, + 342 + ], + "spans": [ + { + "bbox": [ + 132, + 321, + 480, + 342 + ], + "type": "text", + "content": "27. Liu, X., Hu, B., Huang, W., Zhang, Y., Xiong, Z.: Efficient biomedical instance segmentation via knowledge distillation. In: MICCAI (2022)" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 132, + 342, + 480, + 364 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 342, + 480, + 364 + ], + "spans": [ + { + "bbox": [ + 132, + 342, + 480, + 364 + ], + "type": "text", + "content": "28. Liu, Z., Wang, Y., Han, K., Zhang, W., Ma, S., Gao, W.: Post-training quantization for vision transformer. In: NeurIPS (2021)" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 132, + 364, + 480, + 385 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 364, + 480, + 385 + ], + "spans": [ + { + "bbox": [ + 132, + 364, + 480, + 385 + ], + "type": "text", + "content": "29. Loshchilov, I., Hutter, F.: Sgdr: Stochastic gradient descent with warm restarts. arXiv preprint arXiv:1608.03983 (2016)" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 132, + 385, + 480, + 418 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 385, + 480, + 418 + ], + "spans": [ + { + "bbox": [ + 132, + 385, + 480, + 418 + ], + "type": "text", + "content": "30. Martin, D., Fowlkes, C., Tal, D., Malik, J.: A database of human segmented natural images and its application to evaluating segmentation algorithms and measuring ecological statistics. In: ICCV (2001)" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 132, + 418, + 480, + 450 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 418, + 480, + 450 + ], + "spans": [ + { + "bbox": [ + 132, + 418, + 480, + 450 + ], + "type": "text", + "content": "31. Paszke, A., Gross, S., Massa, F., Lerer, A., Bradbury, J., Chanan, G., Killeen, T., Lin, Z., Gimelshein, N., Antiga, L., et al.: Pytorch: An imperative style, high-performance deep learning library. In: NeurIPS (2019)" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 132, + 450, + 480, + 471 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 450, + 480, + 471 + ], + "spans": [ + { + "bbox": [ + 132, + 450, + 480, + 471 + ], + "type": "text", + "content": "32. Pugliatti, M., Topputo, F.: Doors: Dataset for boulders segmentation. Zenodo 9, 20 (2022)" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 132, + 472, + 480, + 504 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 472, + 480, + 504 + ], + "spans": [ + { + "bbox": [ + 132, + 472, + 480, + 504 + ], + "type": "text", + "content": "33. Radford, A., Kim, J.W., Hallacy, C., Ramesh, A., Goh, G., Agarwal, S., Sastry, G., Askell, A., Mishkin, P., Clark, J., et al.: Learning transferable visual models from natural language supervision. In: ICML (2021)" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 132, + 504, + 387, + 514 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 504, + 387, + 514 + ], + "spans": [ + { + "bbox": [ + 132, + 504, + 387, + 514 + ], + "type": "text", + "content": "34. Ross, S.M.: Introductory statistics. Academic Press (2017)" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 132, + 514, + 480, + 546 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 514, + 480, + 546 + ], + "spans": [ + { + "bbox": [ + 132, + 514, + 480, + 546 + ], + "type": "text", + "content": "35. Shamir, R.R., Duchin, Y., Kim, J., Sapiro, G., Harel, N.: Continuous dice coefficient: a method for evaluating probabilistic segmentations. arXiv preprint arXiv:1906.11031 (2019)" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 132, + 547, + 480, + 579 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 547, + 480, + 579 + ], + "spans": [ + { + "bbox": [ + 132, + 547, + 480, + 579 + ], + "type": "text", + "content": "36. Shao, W., Chen, M., Zhang, Z., Xu, P., Zhao, L., Li, Z., Zhang, K., Gao, P., Qiao, Y., Luo, P.: Omniquant: Omnidirectionally calibrated quantization for large language models. In: ICLR (2023)" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 132, + 579, + 480, + 611 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 579, + 480, + 611 + ], + "spans": [ + { + "bbox": [ + 132, + 579, + 480, + 611 + ], + "type": "text", + "content": "37. Trotter, C., Atkinson, G., Sharpe, M., Richardson, K., McGough, A.S., Wright, N., Burville, B., Berggren, P.: Ndd20: A large-scale few-shot dolphin dataset for coarse and fine-grained categorisation. arXiv preprint arXiv:2005.13359 (2020)" + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 132, + 611, + 480, + 633 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 611, + 480, + 633 + ], + "spans": [ + { + "bbox": [ + 132, + 611, + 480, + 633 + ], + "type": "text", + "content": "38. Tu, Z., Hu, J., Chen, H., Wang, Y.: Toward accurate post-training quantization for image super resolution. In: CVPR (2023)" + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 132, + 633, + 465, + 643 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 633, + 465, + 643 + ], + "spans": [ + { + "bbox": [ + 132, + 633, + 465, + 643 + ], + "type": "text", + "content": "39. Upton, G., Cook, I.: Understanding statistics. Oxford University Press (1996)" + } + ] + } + ], + "index": 23 + }, + { + "bbox": [ + 132, + 643, + 480, + 666 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 643, + 480, + 666 + ], + "spans": [ + { + "bbox": [ + 132, + 643, + 480, + 666 + ], + "type": "text", + "content": "40. Vasu, P.K.A., Gabriel, J., Zhu, J., Tuzel, O., Ranjan, A.: Fastvit: A fast hybrid vision transformer using structural reparameterization. In: ICCV (2023)" + } + ] + } + ], + "index": 24 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "spans": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "type": "text", + "content": "16" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 166, + 91, + 217, + 100 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 166, + 91, + 217, + 100 + ], + "spans": [ + { + "bbox": [ + 166, + 91, + 217, + 100 + ], + "type": "text", + "content": "X. Liu et al." + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 15 + }, + { + "para_blocks": [ + { + "bbox": [ + 130, + 116, + 482, + 533 + ], + "type": "list", + "angle": 0, + "index": 17, + "blocks": [ + { + "bbox": [ + 130, + 116, + 482, + 149 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 116, + 482, + 149 + ], + "spans": [ + { + "bbox": [ + 130, + 116, + 482, + 149 + ], + "type": "text", + "content": "41. Wei, X., Zhang, Y., Li, Y., Zhang, X., Gong, R., Guo, J., Liu, X.: Outlier suppression+: Accurate quantization of large language models by equivalent and effective shifting and scaling. In: EMNLP (2023)" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 130, + 150, + 482, + 183 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 150, + 482, + 183 + ], + "spans": [ + { + "bbox": [ + 130, + 150, + 482, + 183 + ], + "type": "text", + "content": "42. Xie, D., Wang, R., Ma, J., Chen, C., Lu, H., Yang, D., Shi, F., Lin, X.: Edit everything: A text-guided generative system for images editing. arXiv preprint arXiv:2304.14006 (2023)" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 130, + 183, + 482, + 205 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 183, + 482, + 205 + ], + "spans": [ + { + "bbox": [ + 130, + 183, + 482, + 205 + ], + "type": "text", + "content": "43. Xie, Y., Liao, Y.: Efficient-vit: A light-weight classification model based on cnn and vit. In: ICIGP (2023)" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 130, + 205, + 482, + 237 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 205, + 482, + 237 + ], + "spans": [ + { + "bbox": [ + 130, + 205, + 482, + 237 + ], + "type": "text", + "content": "44. Xiong, Y., Varadarajan, B., Wu, L., Xiang, X., Xiao, F., Zhu, C., Dai, X., Wang, D., Sun, F., Iandola, F., et al.: Efficientsam: Leveraged masked image pretraining for efficient segment anything. In: CVPR (2024)" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 130, + 237, + 482, + 259 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 237, + 482, + 259 + ], + "spans": [ + { + "bbox": [ + 130, + 237, + 482, + 259 + ], + "type": "text", + "content": "45. Yang, J., Gao, M., Li, Z., Gao, S., Wang, F., Zheng, F.: Track anything: Segment anything meets videos. arXiv preprint arXiv:2304.11968 (2023)" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 130, + 259, + 482, + 293 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 259, + 482, + 293 + ], + "spans": [ + { + "bbox": [ + 130, + 259, + 482, + 293 + ], + "type": "text", + "content": "46. Yang, L., Wei, Y.Z., He, Y., Sun, W., Huang, Z., Huang, H., Fan, H.: ishape: A first step towards irregular shape instance segmentation. arXiv preprint arXiv:2109.15068 (2021)" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 130, + 293, + 482, + 315 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 293, + 482, + 315 + ], + "spans": [ + { + "bbox": [ + 130, + 293, + 482, + 315 + ], + "type": "text", + "content": "47. Yang, X., Ye, J., Wang, X.: Factorizing knowledge in neural networks. In: European Conference on Computer Vision. Springer (2022)" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 130, + 315, + 482, + 336 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 315, + 482, + 336 + ], + "spans": [ + { + "bbox": [ + 130, + 315, + 482, + 336 + ], + "type": "text", + "content": "48. Yang, X., Zhou, D., Liu, S., Ye, J., Wang, X.: Deep model reassembly. NeurIPS 35, 25739-25753 (2022)" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 130, + 336, + 482, + 357 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 336, + 482, + 357 + ], + "spans": [ + { + "bbox": [ + 130, + 336, + 482, + 357 + ], + "type": "text", + "content": "49. Yu, F., Huang, K., Wang, M., Cheng, Y., Chu, W., Cui, L.: Width & depth pruning for vision transformers. In: AAAI (2022)" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 130, + 358, + 482, + 380 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 358, + 482, + 380 + ], + "spans": [ + { + "bbox": [ + 130, + 358, + 482, + 380 + ], + "type": "text", + "content": "50. Yuan, Z., Xue, C., Chen, Y., Wu, Q., Sun, G.: Ptq4vit: Post-training quantization for vision transformers with twin uniform quantization. In: ECCV (2022)" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 130, + 380, + 482, + 413 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 380, + 482, + 413 + ], + "spans": [ + { + "bbox": [ + 130, + 380, + 482, + 413 + ], + "type": "text", + "content": "51. Zhang, C., Han, D., Qiao, Y., Kim, J.U., Bae, S.H., Lee, S., Hong, C.S.: Faster segment anything: Towards lightweight sam for mobile applications. arXiv preprint arXiv:2306.14289 (2023)" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 130, + 413, + 482, + 456 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 413, + 482, + 456 + ], + "spans": [ + { + "bbox": [ + 130, + 413, + 482, + 456 + ], + "type": "text", + "content": "52. Zhang, C., Zhang, C., Li, C., Qiao, Y., Zheng, S., Dam, S.K., Zhang, M., Kim, J.U., Kim, S.T., Choi, J., et al.: One small step for generative ai, one giant leap for agi: A complete survey on chatgpt in aigc era. arXiv preprint arXiv:2304.06488 (2023)" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 130, + 456, + 482, + 489 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 456, + 482, + 489 + ], + "spans": [ + { + "bbox": [ + 130, + 456, + 482, + 489 + ], + "type": "text", + "content": "53. Zhang, Q., Liu, X., Li, W., Chen, H., Liu, J., Hu, J., Xiong, Z., Yuan, C., Wang, Y.: Distilling semantic priors from sam to efficient image restoration models. In: CVPR (2024)" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 130, + 490, + 482, + 511 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 490, + 482, + 511 + ], + "spans": [ + { + "bbox": [ + 130, + 490, + 482, + 511 + ], + "type": "text", + "content": "54. Zhao, X., Ding, W., An, Y., Du, Y., Yu, T., Li, M., Tang, M., Wang, J.: Fast segment anything. arXiv preprint arXiv:2306.12156 (2023)" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 130, + 511, + 482, + 533 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 511, + 482, + 533 + ], + "spans": [ + { + "bbox": [ + 130, + 511, + 482, + 533 + ], + "type": "text", + "content": "55. Zwillinger, D., Kokoska, S.: CRC standard probability and statistics tables and formulae. *Crc Press* (1999)" + } + ] + } + ], + "index": 16 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 408, + 91, + 447, + 102 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 408, + 91, + 447, + 102 + ], + "spans": [ + { + "bbox": [ + 408, + 91, + 447, + 102 + ], + "type": "text", + "content": "PQ-SAM" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 470, + 91, + 481, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 470, + 91, + 481, + 100 + ], + "spans": [ + { + "bbox": [ + 470, + 91, + 481, + 100 + ], + "type": "text", + "content": "17" + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 16 + } + ], + "_backend": "vlm", + "_version_name": "2.6.4" +} \ No newline at end of file diff --git a/2024/PRET_ Planning with Directed Fidelity Trajectory for Vision and Language Navigation/4e64d425-40e4-4880-bd30-e7c4e06b2cbc_content_list.json b/2024/PRET_ Planning with Directed Fidelity Trajectory for Vision and Language Navigation/4e64d425-40e4-4880-bd30-e7c4e06b2cbc_content_list.json new file mode 100644 index 0000000000000000000000000000000000000000..e79c589bb08c3e597148038ec2a1ccd61143007d --- /dev/null +++ b/2024/PRET_ Planning with Directed Fidelity Trajectory for Vision and Language Navigation/4e64d425-40e4-4880-bd30-e7c4e06b2cbc_content_list.json @@ -0,0 +1,1731 @@ +[ + { + "type": "text", + "text": "PRET: Planning with Directed Fidelity Trajectory for Vision and Language Navigation", + "text_level": 1, + "bbox": [ + 230, + 140, + 772, + 186 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Renjie Lu1, Jingke Meng1(3), and Wei-Shi Zheng1,2,3", + "bbox": [ + 290, + 212, + 712, + 229 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "$^{1}$ School of Computer Science and Engineering, Sun Yat-sen University, Guangzhou, China", + "bbox": [ + 261, + 239, + 738, + 268 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "lurj3@mail2.sysu.edu.cn,mengjke@gmail.com, wszheng@ieee.org", + "bbox": [ + 276, + 268, + 725, + 282 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "2 Peng Cheng Laboratory, Shenzhen, China", + "bbox": [ + 354, + 282, + 648, + 295 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "3 Key Laboratory of Machine Intelligence and Advanced Computing, Ministry of Education, Guangzhou, China", + "bbox": [ + 269, + 295, + 730, + 324 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Abstract. Vision and language navigation is a task that requires an agent to navigate according to a natural language instruction. Recent methods predict sub-goals on constructed topology map at each step to enable long-term action planning. However, they suffer from high computational cost when attempting to support such high-level predictions with GCN-like models. In this work, we propose an alternative method that facilitates navigation planning by considering the alignment between instructions and directed fidelity trajectories, which refers to a path from the initial node to the candidate locations on a directed graph without detours. This planning strategy leads to an efficient model while achieving strong performance. Specifically, we introduce a directed graph to illustrate the explored area of the environment, emphasizing directionality. Then, we firstly define the trajectory representation as a sequence of directed edge features, which are extracted from the panorama based on the corresponding orientation. Ultimately, we assess and compare the alignment between instruction and different trajectories during navigation to determine the next navigation target. Our method outperforms previous SOTA method BEVBert on RxR dataset and is comparable on R2R dataset while largely reducing the computational cost. Code is available: https://github.com/iSEE-Laboratory/VLN-PRET.", + "bbox": [ + 261, + 364, + 743, + 643 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Keywords: Vision-and-Language Navigation $\\cdot$ Planning", + "bbox": [ + 261, + 657, + 643, + 672 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "1 Introduction", + "text_level": 1, + "bbox": [ + 215, + 699, + 375, + 715 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Enabling a robot to perform tasks on behalf of humans has been a longstanding objective in AI research. One such task is vision-and-language navigation (VLN) [4, 23, 36, 41], where an agent is required to navigate to a desired location by following natural language instructions provided by humans. For example, given the instruction \"Go up stairs and stop at the top in front of a mirror.\", the agent needs to follow the instruction and stop at an appropriate location. VLN has attracted numerous research interests [2, 7, 8, 14, 16-18, 20, 26, 31, 40, 44, 47].", + "bbox": [ + 212, + 733, + 787, + 842 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/7a7ffb983c5a1fa866447a08f99871bf31d168a8f065d88a44faab7fbe398a88.jpg", + "image_caption": [ + "Fig. 1: Comparison of SPL [3] and GFLOPs on R2R test unseen split dataset. Our method is comparable with previous SOTA methods while being more computational efficient. The computational cost of text encoder and visual encoder is omitted for fair comparison." + ], + "image_footnote": [], + "bbox": [ + 271, + 143, + 733, + 335 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Recent methods [2, 8, 16, 20, 44] demonstrate the effectiveness of introducing maps to enlarge the decision space to improve planning strategies. Instead of predicting low-level actions limited to short movements, these methods construct a graph to keep track of all visited and navigable locations observed so far, which enables high-level planning by expanding the action space to encompass the entire explored area. With the constructed topological maps, these methods predict actions in the global space via GCN-like models, where node features aggregate neighboring information. However, repeatedly calculating the entire map to predict actions at every step, even if the topo-map has only minor changes, is inefficient. In addition, formulating vision information of an environment in graph structure is too coarse-grained for accurate decision-making. Previous methods [2, 8] address this problem by introducing an additional branch to incorporate fine-grained information, such as local image features and bird's-eye-view features. But this strategy increases the model complexity and further escalates computational cost.", + "bbox": [ + 212, + 415, + 787, + 641 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "In this work, we present an alternative way that supports the global decision space while achieving comparable performance. Our method, named Planning with DiReacted FidElity Trajectory(PRET), does not require calculating the entire graph at each step, nor does it rely on incorporating additional fine-grained information. The main idea is to determine the next location to navigate by evaluating the alignment between the instruction and the visual observations along different trajectories between the start point to all the unvisited nodes, as shown in Figure 2(b). Specifically, we maintain a directed fidelity trajectory-colored in red) for each unvisited node. The fidelity trajectory refers to a path from initial node to an unvisited node without detours. We assess the alignment between each trajectory and the instruction, and select the unvisited node with the highest alignment score to navigate. This planning strategy is efficient as we only need to compute for newly observed nodes. We estimate the instruction-trajectory", + "bbox": [ + 212, + 643, + 789, + 840 + ], + "page_idx": 1 + }, + { + "type": "page_number", + "text": "2", + "bbox": [ + 217, + 114, + 228, + 126 + ], + "page_idx": 1 + }, + { + "type": "header", + "text": "Lu et al.", + "bbox": [ + 271, + 114, + 330, + 126 + ], + "page_idx": 1 + }, + { + "type": "image", + "img_path": "images/5291bd7aeed0f2eafe974f85b8f83f4c9bd21f9a287954cbfa27807e2c1b4014.jpg", + "image_caption": [ + "Fig. 2: Illustration of our approaches. (a) shows our directed graph representation. Each edge is assigned with an orientation-aware panorama feature. (b) depict our planning method. We select an unvisited node-colored yellow to navigate towards next by choosing the fidelity path-colored red that best aligned with instruction." + ], + "image_footnote": [], + "bbox": [ + 305, + 148, + 689, + 305 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "alignment with transformer and further reduce the computational cost by utilizing the KV-cache [35] technique. It also enables us to make decisions by taking into full consideration of the instruction-trajectory alignment.", + "bbox": [ + 212, + 393, + 784, + 439 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "In order to enhance the alignment between instructions and navigation trajectories, we propose to incorporate directionality in path representation. Due to the inherent directional nature of the navigation process, it is crucial that the representation of trajectories in opposite directions is asymmetric, as visual observations are linked to observed orientation. In this work, we introduce the directed graph to depict the explored area of the environment and firstly define the trajectory representation as a sequence of directed edge features. These directed edge features are derived from the panorama based on the corresponding orientation (as depicted in Fig. 2(a)). By incorporating directionality in the graph edge features, we eliminate irrelevant panorama information and obtain a more accurate representation of the directed trajectory. Our method overcome existing methods [2,8,16] that commonly adopt the undirected graph to store panorama features on graph nodes, resulting in direction-invariant path representations, which fails to capture the important distinction between paths that traverse the same node from different directions. For instance, consider a scenario where agents pass through the same node from different directions. Despite the fundamentally different spatial contexts of these paths, the node representation remains unchanged.", + "bbox": [ + 212, + 443, + 787, + 715 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "We conduct experiments on R2R, RxR datasets to evaluate the efficacy of our proposed methods. The results demonstrate that PRET achieves strong performance while being more efficient than previous methods. Specifically, PRET achieves comparable performance compared to previous state-of-the-art method BEVBert [2] on R2R dataset with only $3\\%$ computational cost as shown in Fig. 1. On the RxR dataset, PRET outperforms previous methods and achieves the new state-of-the-art performance. Qualitative visualization also shows that our simplified model is able to learn complex backtracking strategies.", + "bbox": [ + 212, + 719, + 787, + 840 + ], + "page_idx": 2 + }, + { + "type": "header", + "text": "PRET", + "bbox": [ + 686, + 114, + 730, + 126 + ], + "page_idx": 2 + }, + { + "type": "page_number", + "text": "3", + "bbox": [ + 774, + 116, + 784, + 126 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "2 Related Work", + "text_level": 1, + "bbox": [ + 215, + 143, + 388, + 160 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Vision-and-Language Navigation(VLN). Early VLN methods use sequence-to-sequence LSTMs with attention mechanism to encode language features and predict local actions. SpeakerFollower [14] first introduces back translation [38] technique to ease the data scarcity problem in VLN. EnvDrop [40] proposes to augment environment by using dropout on feature space to avoid overfitting on visual input. Following this work, many environment augmentation approaches [24, 29] is proposed. Some auxiliary tasks is proposed by [31, 48] for better guidance. More recently, transformer is explored how to be adopted in VLN. PRESS [27] uses BERT [11] as the text encoder. PREVALENT [17] first propose to pretrain the transformer in VLN dataset, but still use it as encoder. More recent works [7, 18, 28] explore how to memorize navigation history and use transformer to learn strong planning strategy.", + "bbox": [ + 212, + 174, + 787, + 354 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Navigation Strategy in VLN. Navigation strategy plays a key role in Vision- and-Language Navigation, as VLN requires agents to navigate in unseen environments and thus needs to explore and familiarize the environment. Works like [32] investigate designing regretful agents to enable explicit exploration. Reinforcement learning methods [40, 46] have also been explored to enhance navigation strategies. SSM [44] constructs directed graphs to represent explored areas, with vision features on nodes and orientation on edges. DUET [8] employs a dual-scale transformer to make local and global predictions. AZHP [16] constructs hierarchy graphs to facilitate exploration. MetaExplore [20] explicitly predicts whether to backtrack using a separate module.", + "bbox": [ + 212, + 356, + 787, + 507 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Maps For Navigation. Navigation research has a long history of using SLAM [15] to construct maps [5, 6, 42] for planning. In the VLN literature, several methods [8, 44] adopt topological graphs for global planning. However, the visual representation on topological maps is too coarse-grained for decision-making. To address this limitation, [2, 47] introduce metric maps in VLN. While grid-based metric maps can precisely represent spatial layouts, they result in high computational costs. To balance representation ability and computational cost, BEVBert [2] utilizes learnable hybrid topological and metric maps. However, the computational cost remains high. In our work, we propose a novel approach that involves constructing a directed topo-map and planning with trajectory. Our method is more efficient as we perform planning with trajectory instead of directly relying on GCN-like models to encode the entire map for planning. This also allows us to incrementally calculate embeddings for new trajectories and further reduces computational cost.", + "bbox": [ + 212, + 507, + 787, + 719 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "3 Method", + "text_level": 1, + "bbox": [ + 215, + 739, + 330, + 755 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "3.1 Problem Formulation", + "text_level": 1, + "bbox": [ + 215, + 770, + 439, + 785 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "In VLN with discrete environments [4,23,36,41], an environment is an undirected graph $\\mathcal{G} = \\{\\mathcal{V},\\mathcal{E}\\}$ , where $\\mathcal{V} = \\{V_i\\}_{i=1}^N$ represents $N$ navigable nodes and $\\mathcal{E}$ denotes navigable edges. At the beginning of navigation, an agent is initialized", + "bbox": [ + 212, + 794, + 785, + 840 + ], + "page_idx": 3 + }, + { + "type": "page_number", + "text": "4", + "bbox": [ + 217, + 114, + 230, + 126 + ], + "page_idx": 3 + }, + { + "type": "header", + "text": "Lu et al.", + "bbox": [ + 271, + 114, + 330, + 127 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "at a starting node and given a natural language instruction $W = \\{w_{i}\\}_{i=1}^{L}$ with $L$ words. The agent is required to interpret this instruction to navigate to the target location.", + "bbox": [ + 212, + 145, + 782, + 190 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "At step $t$ on node $V_{t}$ , the agent observes (1) a panorama $\\mathcal{R}_t = \\{r_{t,i}\\}_{i=1}^K$ represented by $K$ images from different views, $r_{t,i}$ is the extracted image feature of the $i$ th view; (2) neighboring navigable nodes $\\mathcal{N}(V_t)$ , where $\\mathcal{N}(V_t) \\subset \\mathcal{V}$ ; (3) orientations and coordinates of these neighboring nodes. The agent should choose a neighboring node to step to or stop at current location. Navigation is considered successful if the agent stops within 3 meters of the target.", + "bbox": [ + 212, + 191, + 784, + 281 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "We treat navigation as a process of searching temporary target to navigate towards next among unvisited nodes on $\\mathcal{G}_t$ . A temporary target can be either a local neighboring node or remote unvisited node. By adding a virtual stop node connected to all nodes on the graph, we can also model the stop action. We refer fidelity trajectory to a path from initial node to an unvisited node without detours.", + "bbox": [ + 212, + 282, + 785, + 371 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "3.2 Model Overview", + "text_level": 1, + "bbox": [ + 215, + 393, + 398, + 407 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "The overall framework is shown in Fig. 3(a). Our planning method consists of three components: (1) We construct a directed graph for explored area. Considering the directional nature of navigation, we propose an Orientation-aware Panorama Encoder(OPE) to extract orientation-aware vision features for edges. Edge sequence is used to represent directed trajectory and align with the instruction. (2) We maintain a path (i.e. directed fidelity trajectory) to each unvisited node and assess the alignment between each path and the instruction at each step. Only few neighboring nodes are added at each step, we only compute for these relevant new paths. Since each node corresponds to a path, we calculate a path embedding with a matching assessment module(MAM) that encodes the instruction-path alignment and stores the embedding on the node. (3) We compare the path embeddings of all unvisited nodes(candidates) with candidate comparison module(CCM)(shown in Fig. 3(c)) to determine which aligns best with the instruction. We select the best aligned candidate node as the temporary target. Then the agent can navigate to the temporary target along the shortest path on constructed graph. Since the path embedding is an estimate of the alignment between instruction and a path to the node, CCM ensures that the agent navigates by fully considering the instruction-trajectory alignment.", + "bbox": [ + 212, + 417, + 787, + 690 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "3.3 Orientation-Aware Directed Graph Construction", + "text_level": 1, + "bbox": [ + 215, + 710, + 663, + 724 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "We construct a directed graph to support path representation, incrementally adding nodes and edges as exploration proceeds. Previous methods [2,8,44] represent visited nodes with panorama features. However, panoramas of unvisited nodes are inaccessible as the agent has not visited them. So they represent these nodes with views towards them, leading to inconsistent representation between visited and unvisited nodes. In contrast, we extract visual features for the directed edges on the graph. These edge features represent the visual observations", + "bbox": [ + 212, + 734, + 787, + 840 + ], + "page_idx": 4 + }, + { + "type": "header", + "text": "PRET", + "bbox": [ + 684, + 113, + 730, + 126 + ], + "page_idx": 4 + }, + { + "type": "page_number", + "text": "5", + "bbox": [ + 774, + 114, + 785, + 126 + ], + "page_idx": 4 + }, + { + "type": "image", + "img_path": "images/2909c5b6168339dd7a93e5197057553ab89b003042ea5ba060d8584919acd8b3.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 222, + 146, + 782, + 263 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/0e14ba6e264285d2340f19a06a9ea1d5e964bc88b1099d91ff4d1487bc904edc.jpg", + "image_caption": [ + "Fig. 3: Illustration of our model. (a) is the overall framework of our method. At each step, we update the graph, extract path embeddings, and predict actions. (b) depicts the matching assessment module(MAM). Each token is an edge feature. We compute path embeddings for each newly observed nodes with cross-modal transformer and impose a causal mask to reduce computational cost. (c) shows the candidate comparison module(CCM). We gather path embeddings of unvisited nodes and forward them into a single layer transformer followed by a MLP to predict temporary target." + ], + "image_footnote": [ + "visited node current node unvisited node start token shared edge token unvisited neighboring edge token stop token" + ], + "bbox": [ + 222, + 263, + 614, + 383 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/a9f78e2971a855100fe31657721cd55dc149d31d8a54902342afb2ab2d6ed1d3.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 625, + 266, + 751, + 377 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "when facing specific directions from each node. As edges rely only on observable views rather than node panoramas, they enable consistent representation for both visited and unvisited nodes.", + "bbox": [ + 212, + 529, + 784, + 573 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Let $\\mathcal{G}_t = \\{\\mathcal{V}_t, \\mathcal{E}_t\\}$ be the directed graph at step $t$ . Neighboring nodes $\\mathcal{N}(V_t)$ are observed, and some nodes in $\\mathcal{N}(V_t)$ have already existed in $\\mathcal{G}_t$ while others are newly observed and added to $\\mathcal{G}_t$ . Besides, directed edges from $V_t$ to nodes in $\\mathcal{N}(V_t)$ together with features extracted by orientation-aware panorama encoder(OPE) will also be added in $G_t$ , as shown in Fig. 3(a).", + "bbox": [ + 212, + 575, + 787, + 650 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Visual Representation For Edges. We extract orientation-aware panorama feature by using relative orientation as query to attend panorama features in cross-attention manner, as shown in the OPE module in Fig. 3(a). Assume that $(\\phi, \\theta)$ represents the relative heading and elevation of a specific edge when compared to the agent's current orientation. To extract feature for the edge, we first encode $(\\phi, \\theta)$ as follows:", + "bbox": [ + 212, + 675, + 787, + 765 + ], + "page_idx": 5 + }, + { + "type": "equation", + "text": "\n$$\nx ^ {a} = \\left[ \\sin (\\phi), \\cos (\\phi), \\sin (\\theta), \\cos (\\theta) \\right] W ^ {a}, \\tag {1}\n$$\n", + "text_format": "latex", + "bbox": [ + 357, + 777, + 785, + 797 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "where $W^{a}\\in \\mathbb{R}^{4\\times d}$ are learnable weights and $d$ is the dimension of feature. Then we denote $X_{t}^{a} = \\{x_{t,i}^{a}\\}_{i = 1}^{\\lfloor \\mathcal{N}(V_{t})\\rfloor}$ as orientation features of adjacent edges of $V_{t}$ .", + "bbox": [ + 212, + 806, + 787, + 842 + ], + "page_idx": 5 + }, + { + "type": "page_number", + "text": "6", + "bbox": [ + 217, + 114, + 228, + 126 + ], + "page_idx": 5 + }, + { + "type": "header", + "text": "Lu et al.", + "bbox": [ + 271, + 114, + 330, + 126 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "We then encode the panorama feature. As the panorama $\\mathcal{R}_t = \\{r_{t,i}\\}_{i=1}^K$ is represented by $K$ images from different views and each view corresponds to a fixed orientation, we can also get the relative heading and elevation of each view. To encode the panorama, we first concatenate each view feature $r_{t,i}$ with corresponding relative orientation $(\\phi_{t,i},\\theta_{t,i})$ and then map it to dimension $d$ via a linear projection:", + "bbox": [ + 212, + 145, + 787, + 238 + ], + "page_idx": 6 + }, + { + "type": "equation", + "text": "\n$$\nx _ {t, i} ^ {p} = \\left[ r _ {t, i}; \\sin \\left(\\phi_ {t, i}\\right), \\cos \\left(\\phi_ {t, i}\\right), \\sin \\left(\\theta_ {t, i}\\right), \\cos \\left(\\theta_ {t, i}\\right) \\right] W ^ {p}. \\tag {2}\n$$\n", + "text_format": "latex", + "bbox": [ + 321, + 247, + 785, + 263 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Then we denote $X_{t}^{p} = \\{x_{t,i}^{p}\\}_{i=1}^{K}$ as the panoramic view representation. By considering views as patches of the panorama and orientations as position embedding, $X_{t}^{p}$ is like the encoded ViT [12] input.", + "bbox": [ + 214, + 272, + 782, + 318 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "We adopt transformer decoder [43] to extract orientation-aware panorama feature:", + "bbox": [ + 214, + 319, + 782, + 347 + ], + "page_idx": 6 + }, + { + "type": "equation", + "text": "\n$$\nE _ {t} = \\operatorname {T r a n s f o r m e r D e c o d e r} \\left(X _ {t} ^ {a}, X _ {t} ^ {p}\\right), \\tag {3}\n$$\n", + "text_format": "latex", + "bbox": [ + 375, + 349, + 784, + 364 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "where $E_{t} = \\{e_{t,i}\\}_{i = 1}^{\\lfloor \\mathcal{N}(V_{t})\\rfloor}$ are extracted edge features. By taking $X_{t}^{a}$ as the query input and $X_{t}^{p}$ as the key-value input, panorama is attended by orientation in the first layer. In the subsequent layers, both orientation and vision information are used to query relevant views in the panorama. Therefore, each directed edge only focuses on a specific region of the panorama, thus orientation-aware panorama feature is extracted.", + "bbox": [ + 212, + 371, + 787, + 463 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "3.4 Planning with Fidelity Trajectory", + "text_level": 1, + "bbox": [ + 214, + 484, + 539, + 500 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "We maintain a directed fidelity trajectory with a stack during navigation for each unvisited node on $\\mathcal{G}_t$ . When a new node is added, it is pushed into the stack. When the agent needs to backtrack, the top node is popped off the stack until a different node than the top is reached. In this way, detours are removed by popping nodes off the stack. With the maintained trajectory of each node, we propose MAM to assess its alignment with instruction. We then determine an unvisited node on $\\mathcal{G}_t$ to navigate next by comparing this alignment with CCM.", + "bbox": [ + 212, + 507, + 787, + 613 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Matching Assessment Module(MAM). MAM is a multi-layer transformer decoder that extracts a path embedding for each path, which estimates how much the trajectory match with the instruction. These path embeddings are then stored on the corresponding nodes in graph $\\mathcal{G}_t$ .", + "bbox": [ + 212, + 633, + 785, + 694 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "For a single path with length $l$ , we represent the path with the edge features along it, denoted as $X^{h} = [e_{1}, e_{2}, \\dots, e_{l}]$ . $\\{e_{i}\\}_{i=1}^{l}$ are the edge features on the path. Then we add it with position embedding and forward it together with text embedding $X^{w}$ into MAM like follows:", + "bbox": [ + 212, + 694, + 787, + 753 + ], + "page_idx": 6 + }, + { + "type": "equation", + "text": "\n$$\nX ^ {h \\prime} = X ^ {h} + P _ {l},\n$$\n", + "text_format": "latex", + "bbox": [ + 366, + 762, + 480, + 779 + ], + "page_idx": 6 + }, + { + "type": "equation", + "text": "\n$$\nX ^ {o} = \\operatorname {T r a n s f o r m e r D e c o d e r} \\left(X ^ {h t}, X ^ {w}\\right), \\tag {4}\n$$\n", + "text_format": "latex", + "bbox": [ + 370, + 775, + 784, + 800 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "where $P_{l}$ is a matrix including $l$ position embeddings [43] and $X^{w}$ is the text embedding extracted by a text encoder, which is a multi-layer transformer [43].", + "bbox": [ + 212, + 809, + 785, + 840 + ], + "page_idx": 6 + }, + { + "type": "header", + "text": "PRET", + "bbox": [ + 684, + 114, + 730, + 126 + ], + "page_idx": 6 + }, + { + "type": "page_number", + "text": "7", + "bbox": [ + 774, + 114, + 785, + 125 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "By utilizing two sequences from two modalities as inputs, the transformer is capable of effectively estimating their alignment, as extensively demonstrated in Vision and Language Pretraining studies [9, 22, 25]. The last token of the encoded sequence $X^o$ provides the path embedding, which is an estimation of the text-path alignment.", + "bbox": [ + 212, + 146, + 787, + 222 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "To support STOP action, we add a stop node that connected to all nodes in $\\mathcal{G}_t$ . Since we represent a path using edge features, we assign a learnable stop token, initialized with a zero vector, to all edges that point to the stop node. The path embedding corresponds to the stop node is stored on current node. So the agent decides to stop at current location when it finds that the path up to here appending with stop token matches instruction the best. Here stop token serves as the End-of-Sentence token in text generation. For each path, we also add a learnable start token that serves as the Start-of-Sentence token.", + "bbox": [ + 212, + 220, + 787, + 343 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "At step $t$ , considering the stop node, assume there are $N$ neighboring unvisited nodes to be updated. Computing path embeddings separately for these nodes requires multiple forward passes, which is inefficient. Actually, we can compute these embeddings in a single forward pass, as shown in Fig. 3(b). Note that these paths share the same prefix from the initial node to the current node. We can avoid duplicated computation by imposing a causal mask in the self-attention block. The causal mask is a lower triangular matrix that only allows a token to attend to its previous tokens. We merge these causal masks together and forward all relevant tokens into the transformer block. The merged mask still satisfies the constraint that each token can only attend to its previous tokens. We further reduce the computational cost by adopting the KV-cache technique [35], which stores the shared prefix tokens to avoid unnecessary re-calculation. In this way, we can equivalently compute the path embeddings for multiple newly observed nodes in a single forward pass. The last several output tokens colored yellow and green are path embeddings and used to update the graph, i.e., stored on corresponding graph nodes.", + "bbox": [ + 212, + 343, + 787, + 585 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Candidate Comparison Module(CCM). CCM compares the path embeddings from candidate nodes to predict a temporary target. As each path embedding encodes the alignment between a trajectory and the instruction, the CCM actually selects a path that is most aligned with the instruction. In this way, we make navigation planning fully according to the instruction-trajectory alignment. As shown in Fig. 3(c), we collect path embeddings $X_{t}^{e}$ and input them into CCM. CCM consists of a single-layer transformer encoder [43] and a MLP. The transformer encoder is responsible for the comparison of the candidates and the MLP maps the encoded path embedding to a 1D score:", + "bbox": [ + 212, + 603, + 787, + 741 + ], + "page_idx": 7 + }, + { + "type": "equation", + "text": "\n$$\nX _ {t} ^ {e \\prime} = \\operatorname {T r a n s f o r m e r L a y e r} \\left(X _ {t} ^ {e}\\right),\n$$\n", + "text_format": "latex", + "bbox": [ + 395, + 748, + 609, + 763 + ], + "page_idx": 7 + }, + { + "type": "equation", + "text": "\n$$\ns _ {t} = \\operatorname {M L P} \\left(X _ {t} ^ {e ^ {\\prime}}\\right), \\tag {5}\n$$\n", + "text_format": "latex", + "bbox": [ + 406, + 767, + 784, + 782 + ], + "page_idx": 7 + }, + { + "type": "equation", + "text": "\n$$\np _ {t} = \\operatorname {s o f t m a x} \\left(s _ {t}\\right).\n$$\n", + "text_format": "latex", + "bbox": [ + 406, + 787, + 529, + 801 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Score $s_t$ reflects the relative text-path alignment after comparison. An alternative way is removing the transformer and directly compute the alignment score $s_t$ .", + "bbox": [ + 212, + 809, + 787, + 840 + ], + "page_idx": 7 + }, + { + "type": "page_number", + "text": "8", + "bbox": [ + 217, + 114, + 228, + 126 + ], + "page_idx": 7 + }, + { + "type": "header", + "text": "Lu et al.", + "bbox": [ + 271, + 114, + 330, + 127 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "However, this method assesses each path independently without any comparison, which can make the decision-making process more challenging and potentially result in a decrease in performance(demonstrated in Sec. 4). It is then normalized with softmax and we select the highest score node as the temporary target. The agent will navigate to it along the shortest path on $\\mathcal{G}_t$ . If the stop node is selected, the agent believes the instruction is completed and stops the navigation.", + "bbox": [ + 214, + 146, + 782, + 237 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "3.5 Pretraining and Fine-tuning", + "text_level": 1, + "bbox": [ + 215, + 257, + 495, + 272 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "Pretraining. Pretraining is proved to be helpful in previous works [2,7,17,33]. PREVALENT [17] synthesized a substantial amount of data for pretraining, and we utilized the same data as their work. We use Masked Language Modeling(MLM) [11] for pretraining. Concretely, we randomly replace $15\\%$ of the input tokens with a special [MASK] token and forward these masked tokens into the text encoder. We also extract a sequence of edge features to represent the path. Then the encoded text tokens and path representation are input into an additional transformer decoder. The corresponding outputs are used to predict the masked tokens. MLM helps the model learn aligned text and path representations. In VLN, a key difference from other vision-and-language tasks is that we align a text sequence with a sequence of vision representations rather than a single image.", + "bbox": [ + 214, + 279, + 784, + 460 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "Fine-tuning. We train the agent with a mixture of teacher-forcing and student-forcing as previous methods [2,8]. In the teacher-forcing stage, the agent navigate according the ground truth actions $a_{t}^{*}$ and the loss is calculated as follows:", + "bbox": [ + 214, + 479, + 782, + 526 + ], + "page_idx": 8 + }, + { + "type": "equation", + "text": "\n$$\n\\mathcal {L} _ {T F} = \\frac {1}{T _ {T F}} \\sum_ {t = 1} ^ {T _ {T F}} \\mathrm {C E} \\left(p _ {t}, a _ {t} ^ {*}\\right), \\tag {6}\n$$\n", + "text_format": "latex", + "bbox": [ + 403, + 536, + 784, + 574 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "CE is the cross entropy loss and $T_{TF}$ is the number of navigate steps. In the student-forcing stage, we sample an action from the distribution predicted by the agent so that the agent can explore the environment and reduce the exposure bias. We train the agent with heuristic pseudo label $a_t^{pseudo}$ :", + "bbox": [ + 214, + 583, + 782, + 643 + ], + "page_idx": 8 + }, + { + "type": "equation", + "text": "\n$$\n\\mathcal {L} _ {S F} = \\frac {1}{T _ {S F}} \\sum_ {t = 1} ^ {T _ {S F}} \\mathrm {C E} \\left(p _ {t}, a _ {t} ^ {\\text {p s e u d o}}\\right). \\tag {7}\n$$\n", + "text_format": "latex", + "bbox": [ + 388, + 654, + 782, + 691 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "When the agent deviate from ground truth path, we take the nearest node on ground truth path as the pseudo label to encourage agent to learn backtracking strategy. If there is no unvisited node on ground truth path, we take the nearest node on the shortest path from current node to target node as the pseudo label. The agent navigate two times to compute $\\mathcal{L}_{TF}$ and $\\mathcal{L}_{SF}$ . The total loss is the weighted sum of them:", + "bbox": [ + 214, + 700, + 782, + 790 + ], + "page_idx": 8 + }, + { + "type": "equation", + "text": "\n$$\n\\mathcal {L} = \\lambda \\mathcal {L} _ {T F} + (1 - \\lambda) \\mathcal {L} _ {S F}, \\tag {8}\n$$\n", + "text_format": "latex", + "bbox": [ + 411, + 801, + 782, + 815 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "where $\\lambda \\in (0,1)$ is the weight.", + "bbox": [ + 215, + 824, + 433, + 839 + ], + "page_idx": 8 + }, + { + "type": "header", + "text": "PRET", + "bbox": [ + 684, + 114, + 730, + 126 + ], + "page_idx": 8 + }, + { + "type": "page_number", + "text": "9", + "bbox": [ + 774, + 114, + 785, + 126 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "4 Experiments", + "text_level": 1, + "bbox": [ + 215, + 143, + 375, + 162 + ], + "page_idx": 9 + }, + { + "type": "text", + "text": "4.1 Datasets and Evaluation Metrics", + "text_level": 1, + "bbox": [ + 215, + 175, + 532, + 191 + ], + "page_idx": 9 + }, + { + "type": "text", + "text": "R2R. Room-to-Room (R2R) dataset [4] contains 7,189 trajectories, with 3 instructions per trajectory, across 90 scenes. These scenes are divided into train, val unseen and test unseen splits with 61, 11 and 18 scenes respectively. A val seen split with the same scenes as the train split is also provided. All paths in R2R are the shortest paths between the start and target nodes.", + "bbox": [ + 215, + 199, + 785, + 275 + ], + "page_idx": 9 + }, + { + "type": "text", + "text": "RxR. Room-across-Room(RxR) [23] is a large multilingual dataset. The dataset has 126,000 instructions total, with 42,000 instructions in each of the three languages: English, Hindi, and Telugu. The paths in RxR are longer than those in R2R and are not the shortest possible routes. Additionally, the instructions in RxR are longer and contain more detailed descriptions compared to R2R.", + "bbox": [ + 215, + 296, + 785, + 372 + ], + "page_idx": 9 + }, + { + "type": "text", + "text": "Evaluation Metrics. We adopt the following evaluation metrics on R2R: (1) Trajectory length (TL): the average length of the agent's path in meters. (2) Navigation error (NE): the average distance between the agent's final position and the target location. (3) Success rate (SR): the ratio of successful navigations, where $NE < 3m$ is considered successful. (4) SR penalized by path length (SPL) [3]: TL longer than the shortest path is penalized. As paths in RxR dataset are shortest paths, TL and SPL are unsuitable. Instead, on RxR we use nDTW and sDTW [21] to measure trajectory similarity between the agent and ground truth. Please note that the fidelity trajectory that is removed detours is solely used for planning. It is not the agent's actual trajectory used for evaluation.", + "bbox": [ + 215, + 393, + 785, + 545 + ], + "page_idx": 9 + }, + { + "type": "text", + "text": "4.2 Implementation Details", + "text_level": 1, + "bbox": [ + 215, + 566, + 455, + 582 + ], + "page_idx": 9 + }, + { + "type": "text", + "text": "Module architecture. Text backbone is a 6 layer transformer encoder initialized with pretrained ALBEF [25]. For the multilingual RxR dataset, we use 12 layer mRoberta [10] instead. The layer numbers of the OPE, MAM, and CCM are 2, 4, and 1 respectively. All the hidden layer size is 768. Image features is extracted by DINOv2 [34]. We also report results with CLIP-ViT-base [37] feature for fair comparison.", + "bbox": [ + 215, + 590, + 785, + 681 + ], + "page_idx": 9 + }, + { + "type": "text", + "text": "Training details. On R2R dataset, we first pretrain PRET with learning rate 2e-5 and batch size 16 on a single 24G 4090 GPU for 100,000 iterations( $\\sim$ 5 hours). Then we fine-tune the model with learning rate 1e-5 and batch size 8 on 4090 for 100,000 iterations( $\\sim$ 25 hours). Optimizer is AdamW [30]. $\\lambda$ is set to 0.2. Augmented dataset [17] is used in both pretraining and fine-tuning. The best result is selected by SPL on val unseen split. On RxR dataset, we also follow the pretrain and fine-tune paradigm. Due to the longer instruction and trajectory, we use a smaller batch size 4. Marky [45] is used as augmented data. The best result is selected by sDTW on val unseen split.", + "bbox": [ + 215, + 703, + 785, + 839 + ], + "page_idx": 9 + }, + { + "type": "page_number", + "text": "10", + "bbox": [ + 217, + 114, + 235, + 126 + ], + "page_idx": 9 + }, + { + "type": "header", + "text": "Lu et al.", + "bbox": [ + 271, + 114, + 330, + 126 + ], + "page_idx": 9 + }, + { + "type": "table", + "img_path": "images/b674c067151bf323d0e173be14674b2adb6246d5775a975807c865b412b3ea80.jpg", + "table_caption": [ + "Table 1: Comparison with other methods on R2R dataset. SPL is considered as the primary evaluation metric." + ], + "table_footnote": [], + "table_body": "
MethodsVal SeenVal UnseenTest Unseen
TLNE↓SR↑SPL↑TLNE↓SR↑SPL↑TLNE↓SR↑SPL↑
Seq2Seq-SF [4]11.336.0139-8.397.8122-8.137.852818
Speaker-Follower [14]-3.3666--6.6235-14.826.623528
RCM [46]10.653.5367-11.466.0943-11.976.124338
Regretful [32]-3.236963-5.325041-5.695640
EnvDrop [40]11.003.99625910.705.22524811.665.235147
PREVALENT [17]10.323.67696510.194.71585310.515.305451
NvEM [1]11.093.44696511.834.27605512.984.375854
SSM [44]14.703.10716220.704.32624520.404.576146
RecBert [18]11.132.90726812.013.93635712.354.096357
HAMT [7]11.152.51767211.462.29666112.273.936560
MTVM [28]-2.677469-3.736659-3.856559
DUET [8]12.322.28797313.943.31726014.733.656959
AZHP [16]----14.053.15726114.953.527160
Meta-Explore [20]11.952.11817513.093.22726214.253.577161
GridMM [47]----13.272.83756414.433.357362
BEVBert [2]13.562.17817414.552.81756415.873.137362
Ours(CLIP)11.482.60746912.213.12716313.873.127262
Ours(DINOv2)11.252.41787211.872.90746512.213.097264
", + "bbox": [ + 233, + 183, + 774, + 444 + ], + "page_idx": 10 + }, + { + "type": "text", + "text": "4.3 Results", + "text_level": 1, + "bbox": [ + 215, + 474, + 321, + 488 + ], + "page_idx": 10 + }, + { + "type": "text", + "text": "Comparison on R2R. Tab. 1 compares our approach with previous methods. On the val unseen and test unseen split, our method achieves comparable performance with previous SOTA method BEVBert on the primary metric SPL, while our method significantly reduces the computational cost. Besides, BEVBert [2] adopts additional depth information to construct bird's-eye view input. On the val seen split, our method performs sub-optimal as PRET does not take the environment layout as input, which means it less tends to overfit on the seen environment. Also, in the VLN literature, all visual features are pre-extracted and remains unchanged, the fixed feature may not contains appropriate feature for navigation. Therefore, we also investigate different vision features for VLN, as shown in Tab. 1. We find that DINOv2 [34] feature is better than CLIP feature. DINOv2 is trained on large-scale curated data in self-supervised manner, which leads to more general purpose feature. Moreover, DINOv2 feature contains geometric information as it can be used to predict depth map according to [34]. Geometric information is shown helpful in [19].", + "bbox": [ + 215, + 506, + 785, + 733 + ], + "page_idx": 10 + }, + { + "type": "text", + "text": "Tab. 2 shows the parameter count, computational cost, and latency of different graph-based methods. We present GFLOPs of different navigation steps. Our method contains fewer parameters as we use a single stream model unlike [2,8] use dual stream transformer. Besides, our decoder computational cost is only $19\\%$ of DUET and $3\\%$ of BEVBert at step 10. These suggests that our method is significantly more efficient than previous SOTA graph-based methods. As shown in Tab. 2, this results in lower training and inference latency.", + "bbox": [ + 215, + 734, + 785, + 839 + ], + "page_idx": 10 + }, + { + "type": "header", + "text": "PRET", + "bbox": [ + 686, + 114, + 730, + 126 + ], + "page_idx": 10 + }, + { + "type": "page_number", + "text": "11", + "bbox": [ + 767, + 114, + 784, + 126 + ], + "page_idx": 10 + }, + { + "type": "table", + "img_path": "images/f3cfc9ca84b97482117a084186c6eea3e3804cc9d12991c504b2dcc8e512016e.jpg", + "table_caption": [ + "Table 2: Comparison of latency and computational cost on R2R val unseen. * indicates we reproduce DUET with CLIP feature for fair comparison. When computing GFLOPs, we simplify the calculation by assuming that each vertex of the graph is connected to 4 adjacent vertices (average degree in R2R), and at each step, 3 new vertices are encountered. The computational cost of text encoder and visual encoder is omitted for fair comparison." + ], + "table_footnote": [], + "table_body": "
MethodsR2R val unseenLatency(ms)GFLOPsParams
TLNE↓SR↑SPL↑traininference1 step10 steps20 steps
DUET* [8]14.03.271.661.111.88.62.44.87.590M
BEVBert [2]14.62.874.963.616.210.633.735.537.590M
Ours12.23.171.062.77.74.10.60.91.264M
", + "bbox": [ + 259, + 239, + 738, + 306 + ], + "page_idx": 11 + }, + { + "type": "table", + "img_path": "images/45d92cff227dc4f30b11192666867175deebdcffe08fbbeb53b2ff788878af27.jpg", + "table_caption": [ + "Table 3: Comparison with previous methods on RxR dataset. †indicates approaches that utilize additional augmented data. nDTW represents the normalized DTW distance between the agent's trajectory and ground truth path. sDTW is the nDTW weighted by success rate." + ], + "table_footnote": [], + "table_body": "
MethodsVal SeenVal Unseen
NE↓SR↑nDTW↑sDTW↑NE↓SR↑nDTW↑sDTW↑
LSTM [23]10.725.242.220.710.922.838.918.2
EnvDrop+ [39]-----43.655.7-
HAMT [7]-59.465.350.9-56.563.148.3
EnvEdit [24]-67.271.158.5-62.868.554.6
MPM [13]-67.771.058.9-63.567.754.5
MARVEL [45]†3.075.979.168.84.564.870.857.5
BEVBert [2]†3.275.076.366.74.068.569.658.6
Ours(CLIP)†2.677.078.067.53.371.271.860.4
Ours(DINOv2)†2.479.380.470.73.272.873.462.4
", + "bbox": [ + 274, + 386, + 723, + 527 + ], + "page_idx": 11 + }, + { + "type": "text", + "text": "Comparison on RxR. Tab. 3 reports the results on the RxR dataset. RxR is challenging as instructions and paths in RxR are longer than R2R. While the alignment between path and detailed path description is what our method skills at, our method outperforms previous methods among all metrics at a lower computational cost. PRET achieves a $1.8\\%$ improvement on the main metrics sDTW on the val unseen split, demonstrating that our model better understands and follows instructions. This result highlights the advantage of our planning with trajectory strategy, which considers the alignment between instructions and paths and enables efficient decision-making in the global space. Additionally, with DINOv2 features, the result is further improved.", + "bbox": [ + 212, + 556, + 787, + 708 + ], + "page_idx": 11 + }, + { + "type": "text", + "text": "4.4 Ablation Study", + "text_level": 1, + "bbox": [ + 214, + 729, + 388, + 744 + ], + "page_idx": 11 + }, + { + "type": "text", + "text": "We conducted ablation experiments on various components of PRET, including the directed path representation and module ablation. Results are reported on the R2R val unseen split.", + "bbox": [ + 212, + 753, + 785, + 801 + ], + "page_idx": 11 + }, + { + "type": "page_number", + "text": "12", + "bbox": [ + 217, + 114, + 235, + 126 + ], + "page_idx": 11 + }, + { + "type": "header", + "text": "Lu et al.", + "bbox": [ + 271, + 114, + 330, + 126 + ], + "page_idx": 11 + }, + { + "type": "table", + "img_path": "images/84ce37c5c82808bb0b3cbbdc37558415d5474c3d9ef99e30d14d1019db09a5cf.jpg", + "table_caption": [ + "Table 4: Comparison of undirected and directed path representation." + ], + "table_footnote": [], + "table_body": "
MethodsTLNE↓SR↑SPL↑
undirected15.623.5968.2856.77
directed11.872.9073.7865.16
", + "bbox": [ + 230, + 189, + 439, + 232 + ], + "page_idx": 12 + }, + { + "type": "table", + "img_path": "images/ca045507a1821b5ad005432c4d45bfd40e59b1f14ee5ac3031f5e03763c4da09.jpg", + "table_caption": [ + "Table 5: Ablation study on modules." + ], + "table_footnote": [], + "table_body": "
MethodsTLNE↓SR↑SPL↑
1 MAM12.043.9962.3254.48
2 MAM+CCM12.153.5465.9457.32
3 MAM+OPE12.183.1571.6063.07
4 MAM+OPE+CCM11.872.9073.7865.16
", + "bbox": [ + 475, + 170, + 751, + 236 + ], + "page_idx": 12 + }, + { + "type": "image", + "img_path": "images/73164e5bbdd0ccea64413afcdab955178397f303e85f580b1c421d823f92f2ff.jpg", + "image_caption": [ + "Fig. 4: Comparison of orientation panoramic view and single candidate view." + ], + "image_footnote": [], + "bbox": [ + 357, + 250, + 645, + 301 + ], + "page_idx": 12 + }, + { + "type": "text", + "text": "Directionality in Path Representation. We compare whether to incorporate directionality for path representations in Tab. 4. Undirected path is represented by a sequence of node features. The node features is extracted by forwarding panoramic views into a 2-layer transformer encoder and averages the output tokens. Directed path is represented by a sequence of orientation-aware panorama features. According to the results, directed path representation outperforms undirected node features among all metrics. Specifically, it gains $8.39\\%$ improvement on SPL and $5.5\\%$ on SR. This indicates that edge feature is more suitable to represent the directional navigation process and align with the instruction. Node features does not distinguish different path directions and provide redundant information in a path, which hinders the alignment.", + "bbox": [ + 212, + 347, + 787, + 513 + ], + "page_idx": 12 + }, + { + "type": "text", + "text": "Module ablation. Tab. 5 shows ablation results on modules. Row 4 is the full model introduced in Sec. 3. OPE extracts orientation-aware features for edges, MAM extracts path embeddings while CCM compares them to select the temporary target. When removing CCM (Row 3), we only predict a single alignment score rather than a path embedding for each path without comparison. The score is then normalized and used to predict the temporary goal. Performance drops $2.09\\%$ in this case, as selecting a path based solely on a single score, without comparing different paths, is more difficult.", + "bbox": [ + 212, + 537, + 787, + 657 + ], + "page_idx": 12 + }, + { + "type": "text", + "text": "In row 2, we remove OPE, panorama features are not used and only a single 60-degree view towards each neighboring node is adopted as edge features. Performance drops $7.84\\%$ as the limited field-of-view cannot fully represent the path. The trajectory represented by these views is discrete and does not form a continuous change of views. In some cases, a single view towards candidates provides no information for navigation, as shown in Fig. 4. When go up the stairs, the upper view only sees the wall and does not help the navigation at all. While using the panorama, the agent can dynamically see a larger region. These results show our graph representation with panorama encoder provides rich representation for trajectories. Row 1 demonstrates the performance when using only MAM. The performance is poor without other model components, demonstrating the effectiveness of these modules.", + "bbox": [ + 212, + 659, + 787, + 839 + ], + "page_idx": 12 + }, + { + "type": "header", + "text": "PRET", + "bbox": [ + 684, + 114, + 730, + 126 + ], + "page_idx": 12 + }, + { + "type": "page_number", + "text": "13", + "bbox": [ + 767, + 114, + 785, + 126 + ], + "page_idx": 12 + }, + { + "type": "text", + "text": "4.5 Visualization", + "text_level": 1, + "bbox": [ + 215, + 146, + 370, + 159 + ], + "page_idx": 13 + }, + { + "type": "image", + "img_path": "images/1661bb8d1b75a64d5109f9fae4bdfd50c34a1793afbf6832f1f8d3d623e5c5d0.jpg", + "image_caption": [ + "Instruction: Face directly away from stairs and walk through doorway straight ahead, right immediately, walk past bed and turn right into closet, walk straight and stop in front of toilet.", + "(a)" + ], + "image_footnote": [], + "bbox": [ + 238, + 170, + 460, + 296 + ], + "page_idx": 13 + }, + { + "type": "image", + "img_path": "images/5d8baa7f63037a6f0be0e12bc335b6b7ba26e0b1574f0545206d7fd0feae89bc.jpg", + "image_caption": [ + "undirected trajectory" + ], + "image_footnote": [], + "bbox": [ + 498, + 181, + 629, + 213 + ], + "page_idx": 13 + }, + { + "type": "image", + "img_path": "images/3c9519642b448d44cbce4747794e4d79c64fbb0c3b15bd002592649ddb5f5a6a.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 498, + 213, + 629, + 257 + ], + "page_idx": 13 + }, + { + "type": "image", + "img_path": "images/83c08a5184cabee19094350062c18cae036b3b49ddf035843b4986b7a15d641e.jpg", + "image_caption": [ + "Instruction: Turn right and go up the stairs." + ], + "image_footnote": [], + "bbox": [ + 498, + 257, + 629, + 297 + ], + "page_idx": 13 + }, + { + "type": "image", + "img_path": "images/659a53cccecbd5e255c4cd2007a9b649f08041e684f66e66f9d4d6c16c76aba8.jpg", + "image_caption": [ + "directed trajectory" + ], + "image_footnote": [], + "bbox": [ + 642, + 181, + 771, + 213 + ], + "page_idx": 13 + }, + { + "type": "image", + "img_path": "images/9b7b767cb0e77eb4644e3e0a2791e9f95abb6ce5e8112d187235c43759854aaf.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 642, + 213, + 771, + 257 + ], + "page_idx": 13 + }, + { + "type": "image", + "img_path": "images/4574cc6c8740af637fc91b0614b2db35f5bf94ae75c5561d9135c162cf93a917.jpg", + "image_caption": [ + "(b)", + "Fig. 5: (a) Visualization of the agent's navigation process, showcasing its ability to learn a backtracking strategy. (b) Visualizing attention weights in OPE to illustrate the distinction between undirected and directed trajectory representations." + ], + "image_footnote": [], + "bbox": [ + 642, + 257, + 771, + 299 + ], + "page_idx": 13 + }, + { + "type": "text", + "text": "Fig. 5(a) shows a qualitative example of our model's behavior. Initially placed at the green starting node, the agent wrongly decides to explore the upper left area(blue path). It takes many unnecessary detours there, but through comparing the trajectory alignment with the instructions, realizes these paths do not match the text well. Therefore, it backtracks to the start and correctly navigates to the destination(red path). This demonstrates the capability of PRET to learn complex backtracking behaviors.", + "bbox": [ + 212, + 401, + 784, + 507 + ], + "page_idx": 13 + }, + { + "type": "text", + "text": "Fig. 5(b) shows the difference between undirected and directed trajectory representations. The former represents a path using panorama features stored on nodes, which eliminates directional discrepancies since the same node provides the same feature regardless of orientation, and contains redundant information compared to \"stairs\" in the instruction. In contrast, our proposed approach represents a path using orientation-aware panorama features on edges. This eliminates redundant information and focuses on the \"stairs\" better captures the directional navigation process.", + "bbox": [ + 212, + 507, + 787, + 628 + ], + "page_idx": 13 + }, + { + "type": "text", + "text": "5 Conclusion", + "text_level": 1, + "bbox": [ + 215, + 646, + 359, + 662 + ], + "page_idx": 13 + }, + { + "type": "text", + "text": "In this work, we present an efficient method that planning with directed fidelity trajectory, explicitly leveraging the alignment between instructions and trajectories for navigation. Additionally, we consider the directional nature of navigation and introduce a directed graph construction during navigation, storing vision information on directed edges. This approach provides rich directed path representation and enhances instruction-trajectory alignment. Experiments demonstrate that our method is achieves strong performance while being significantly more efficient than previous methods. Our method has certain limitations that we need to address. For example, navigation requires environment layout information in some cases. We recognize the need to investigate potential solutions to overcome these challenges.", + "bbox": [ + 212, + 672, + 787, + 840 + ], + "page_idx": 13 + }, + { + "type": "page_number", + "text": "14", + "bbox": [ + 217, + 114, + 235, + 126 + ], + "page_idx": 13 + }, + { + "type": "header", + "text": "Lu et al.", + "bbox": [ + 271, + 114, + 330, + 126 + ], + "page_idx": 13 + }, + { + "type": "text", + "text": "Acknowledgments", + "text_level": 1, + "bbox": [ + 217, + 143, + 392, + 162 + ], + "page_idx": 14 + }, + { + "type": "text", + "text": "This work was supported partially by the National Key Research and Development Program of China (2023YFA1008503), NSFC(U21A20471, 62206315), Guangdong NSF Project (No. 2023B1515040025, 2020B1515120085, 2024A1515-010101), Guangzhou Basic and Applied Basic Research Scheme(2024A04J4067).", + "bbox": [ + 212, + 176, + 787, + 238 + ], + "page_idx": 14 + }, + { + "type": "text", + "text": "References", + "text_level": 1, + "bbox": [ + 215, + 258, + 321, + 273 + ], + "page_idx": 14 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "1. An, D., Qi, Y., Huang, Y., Wu, Q., Wang, L., Tan, T.: Neighbor-view enhanced model for vision and language navigation. In: ACM MM (2021)", + "2. An, D., Qi, Y., Li, Y., Huang, Y., Wang, L., Tan, T., Shao, J.: Bevbert: Multimodal map pre-training for language-guided navigation. In: ICCV (2023)", + "3. Anderson, P., Chang, A.X., Chaplot, D.S., Dosovitskiy, A., Gupta, S., Koltun, V., Kosecka, J., Malik, J., Mottaghi, R., Savva, M., Zamir, A.R.: On evaluation of embodied navigation agents. CoRR (2018)", + "4. Anderson, P., Wu, Q., Teney, D., Bruce, J., Johnson, M., Sünderhauf, N., Reid, I., Gould, S., Van Den Hengel, A.: Vision-and-language navigation: Interpreting visually-grounded navigation instructions in real environments. In: CVPR (2018)", + "5. Chaplot, D.S., Gandhi, D., Gupta, A., Salakhutdinov, R.: Object goal navigation using goal-oriented semantic exploration. In: NeurIPS (2020)", + "6. Chaplot, D.S., Gandhi, D., Gupta, S., Gupta, A., Salakhutdinov, R.: Learning to explore using active neural SLAM. In: ICLR (2020)", + "7. Chen, S., Guhur, P.L., Schmid, C., Laptev, I.: History aware multimodal transformer for vision-and-language navigation. NeurIPS (2021)", + "8. Chen, S., Guhur, P., Tapaswi, M., Schmid, C., Laptev, I.: Think global, act local: Dual-scale graph transformer for vision-and-language navigation. In: CVPR (2022)", + "9. Chen, Y.C., Li, L., Yu, L., El Kholy, A., Ahmed, F., Gan, Z., Cheng, Y., Liu, J.: Uniter: Universal image-text representation learning. In: ECCV (2020)", + "0. Conneau, A., Khandelwal, K., Goyal, N., Chaudhary, V., Wenzek, G., Guzmán, F., Grave, E., Ott, M., Zettlemoyer, L., Stoyanov, V.: Unsupervised cross-lingual representation learning at scale. In: ACL (2020)", + "1. Devlin, J., Chang, M.W., Lee, K., Toutanova, K.: BERT: Pre-training of deep bidirectional transformers for language understanding. In: NAACL (2019)", + "12. Dosovitskiy, A., Beyer, L., Kolesnikov, A., Weissenborn, D., Zhai, X., Unterthiner, T., Dehghani, M., Minderer, M., Heigold, G., Gelly, S., Uszkoreit, J., Houlsby, N.: An image is worth 16x16 words: Transformers for image recognition at scale. In: ICLR (2021)", + "13. Dou, Z., Gao, F., Peng, N.: Masked path modeling for vision-and-language navigation. CoRR (2023)", + "14. Fried, D., Hu, R., Cirik, V., Rohrbach, A., Andreas, J., Morency, L.P., Berg-Kirkpatrick, T., Saenko, K., Klein, D., Darrell, T.: Speaker-follower models for vision-and-language navigation. In: NeurIPS (2018)", + "15. Fuentes-Pacheco, J., Ascencio, J.R., Rendon-Mancha, J.M.: Visual simultaneous localization and mapping: a survey. Artif. Intell. Rev. (2015)", + "16. Gao, C., Peng, X., Yan, M., Wang, H., Yang, L., Ren, H., Li, H., Liu, S.: Adaptive zone-aware hierarchical planner for vision-language navigation. In: CVPR (2023)", + "17. Hao, W., Li, C., Li, X., Carin, L., Gao, J.: Towards learning a generic agent for vision-and-language navigation via pre-training. In: CVPR (2020)" + ], + "bbox": [ + 218, + 287, + 785, + 839 + ], + "page_idx": 14 + }, + { + "type": "header", + "text": "PRET", + "bbox": [ + 684, + 114, + 730, + 126 + ], + "page_idx": 14 + }, + { + "type": "page_number", + "text": "15", + "bbox": [ + 767, + 114, + 785, + 126 + ], + "page_idx": 14 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "18. Hong, Y., Wu, Q., Qi, Y., Rodriguez-Opazo, C., Gould, S.: Vln bert: A recurrent vision-and-language bert for navigation. In: CVPR. pp. 1643-1653 (2021)", + "19. Huo, J., Sun, Q., Jiang, B., Lin, H., Fu, Y.: Geovln: Learning geometry-enhanced visual representation with slot attention for vision-and-language navigation. In: CVPR (2023)", + "20. Hwang, M., Jeong, J., Kim, M., Oh, Y., Oh, S.: Meta-exlore: Exploratory hierarchical vision-and-language navigation using scene object spectrum grounding. In: CVPR (2023)", + "21. Ilharco, G., Jain, V., Ku, A., Ie, E., Baldridge, J.: General evaluation for instruction conditioned navigation using dynamic time warping. In: NeurIPS Workshop (2019)", + "22. Kim, W., Son, B., Kim, I.: Vilt: Vision-and-language transformer without convolution or region supervision. In: ICML (2021)", + "23. Ku, A., Anderson, P., Patel, R., Ie, E., Baldridge, J.: Room-across-room: Multilingual vision-and-language navigation with dense spatiotemporal grounding. In: EMNLP (2020)", + "24. Li, J., Tan, H., Bansal, M.: Envedit: Environment editing for vision-and-language navigation. In: CVPR (2022)", + "25. Li, J., Selvaraju, R., Gotmare, A., Joty, S., Xiong, C., Hoi, S.C.H.: Align before fuse: Vision and language representation learning with momentum distillation. NeurIPS (2021)", + "26. Li, X., Li, C., Xia, Q., Bisk, Y., Celikylmaz, A., Gao, J., Smith, N.A., Choi, Y.: Robust navigation with language pretraining and stochastic sampling. In: EMNLP-IJCNLP (2019)", + "27. Li, X., Li, C., Xia, Q., Bisk, Y., Celikyilmaz, A., Gao, J., Smith, N.A., Choi, Y.: Robust navigation with language pretraining and stochastic sampling. In: EMNLP-IJCNLP (2019)", + "28. Lin, C., Jiang, Y., Cai, J., Qu, L., Haffari, G., Yuan, Z.: Multimodal transformer with variable-length memory for vision-and-language navigation. In: ECCV (2022)", + "29. Liu, C., Zhu, F., Chang, X., Liang, X., Ge, Z., Shen, Y.: Vision-language navigation with random environmental mixup. In: ICCV (2021)", + "30. Loshchilov, I., Hutter, F.: Decoupled weight decay regularization. In: ICLR (2019)", + "31. Ma, C.Y., Lu, J., Wu, Z., AlRegib, G., Kira, Z., Socher, R., Xiong, C.: Self-monitoring navigation agent via auxiliary progress estimation. In: ICLR (2019)", + "32. Ma, C., Wu, Z., AlRegib, G., Xiong, C., Kira, Z.: The regretful agent: Heuristic-aided navigation through progress estimation. In: CVPR (2019)", + "33. Majumdar, A., Shrivastava, A., Lee, S., Anderson, P., Parikh, D., Batra, D.: Improving vision-and-language navigation with image-text pairs from the web. In: ECCV (2020)", + "34. Oquab, M., Darcet, T., Moutakanni, T., Vo, H., Szafraniec, M., Khalidov, V., Fernandez, P., Haziza, D., Massa, F., El-Nouby, A., Assran, M., Ballas, N., Galuba, W., Howes, R., Huang, P., Li, S., Misra, I., Rabbat, M.G., Sharma, V., Synnaeve, G., Xu, H., Jegou, H., Mairal, J., Labatut, P., Joulin, A., Bojanowski, P.: Dinov2: Learning robust visual features without supervision. CoRR (2023)", + "35. Pope, R., Douglas, S., Chowdhery, A., Devlin, J., Bradbury, J., Levskaya, A., Heek, J., Xiao, K., Agrawal, S., Dean, J.: Efficiently scaling transformer inference. CoRR (2022)", + "36. Qi, Y., Wu, Q., Anderson, P., Wang, X., Wang, W.Y., Shen, C., Hengel, A.v.d.: Reverie: Remote embodied visual referring expression in real indoor environments. In: CVPR (2020)" + ], + "bbox": [ + 215, + 147, + 784, + 839 + ], + "page_idx": 15 + }, + { + "type": "page_number", + "text": "16", + "bbox": [ + 217, + 114, + 235, + 126 + ], + "page_idx": 15 + }, + { + "type": "header", + "text": "Lu et al.", + "bbox": [ + 271, + 114, + 330, + 126 + ], + "page_idx": 15 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "37. Radford, A., Kim, J.W., Hallacy, C., Ramesh, A., Goh, G., Agarwal, S., Sastry, G., Askell, A., Mishkin, P., Clark, J., et al.: Learning transferable visual models from natural language supervision. In: ICML (2021)", + "38. Sennrich, R., Haddow, B., Birch, A.: Improving neural machine translation models with monolingual data. In: ACL (2016)", + "39. Shen, S., Li, L.H., Tan, H., Bansal, M., Rohrbach, A., Chang, K., Yao, Z., Keutzer, K.: How much can CLIP benefit vision-and-language tasks? In: ICLR (2022)", + "40. Tan, H., Yu, L., Bansal, M.: Learning to navigate unseen environments: Back translation with environmental dropout. In: NAACL (2019)", + "41. Thomason, J., Murray, M., Cakmak, M., Zettlemoyer, L.: Vision-and-dialog navigation. In: CoRL (2020)", + "42. Thrun, S.: Learning metric-topological maps for indoor mobile robot navigation. Artif. Intell. (1998)", + "43. Vaswani, A., Shazeer, N., Parmar, N., Uszkoreit, J., Jones, L., Gomez, A.N., Kaiser, L.u., Polosukhin, I.: Attention is all you need. In: NeurIPS (2017)", + "44. Wang, H., Wang, W., Liang, W., Xiong, C., Shen, J.: Structured scene memory for vision-language navigation. In: CVPR (2021)", + "45. Wang, S., Montgomery, C., Orbay, J., Birodkar, V., Faust, A., Gur, I., Jaques, N., Waters, A., Baldridge, J., Anderson, P.: Less is more: Generating grounded navigation instructions from landmarks. In: CVPR (2022)", + "46. Wang, X., Huang, Q., Celikyilmaz, A., Gao, J., Shen, D., Wang, Y.F., Wang, W.Y., Zhang, L.: Reinforced cross-modal matching and self-supervised imitation learning for vision-language navigation. In: CVPR (2019)", + "47. Wang, Z., Li, X., Yang, J., Liu, Y., Jiang, S.: Gridmm: Grid memory map for vision-and-language navigation. In: ICCV (2023)", + "48. Zhu, F., Zhu, Y., Chang, X., Liang, X.: Vision-language navigation with self-supervised auxiliary reasoning tasks. In: CVPR (2020)" + ], + "bbox": [ + 212, + 146, + 787, + 522 + ], + "page_idx": 16 + }, + { + "type": "header", + "text": "PRET", + "bbox": [ + 686, + 114, + 730, + 126 + ], + "page_idx": 16 + }, + { + "type": "page_number", + "text": "17", + "bbox": [ + 767, + 114, + 785, + 126 + ], + "page_idx": 16 + } +] \ No newline at end of file diff --git a/2024/PRET_ Planning with Directed Fidelity Trajectory for Vision and Language Navigation/4e64d425-40e4-4880-bd30-e7c4e06b2cbc_model.json b/2024/PRET_ Planning with Directed Fidelity Trajectory for Vision and Language Navigation/4e64d425-40e4-4880-bd30-e7c4e06b2cbc_model.json new file mode 100644 index 0000000000000000000000000000000000000000..be3273b01710fc44c185de8954a463c5819c14da --- /dev/null +++ b/2024/PRET_ Planning with Directed Fidelity Trajectory for Vision and Language Navigation/4e64d425-40e4-4880-bd30-e7c4e06b2cbc_model.json @@ -0,0 +1,2324 @@ +[ + [ + { + "type": "title", + "bbox": [ + 0.232, + 0.141, + 0.773, + 0.187 + ], + "angle": 0, + "content": "PRET: Planning with Directed Fidelity Trajectory for Vision and Language Navigation" + }, + { + "type": "text", + "bbox": [ + 0.292, + 0.213, + 0.713, + 0.23 + ], + "angle": 0, + "content": "Renjie Lu1, Jingke Meng1(3), and Wei-Shi Zheng1,2,3" + }, + { + "type": "text", + "bbox": [ + 0.263, + 0.24, + 0.74, + 0.269 + ], + "angle": 0, + "content": "\\(^{1}\\) School of Computer Science and Engineering, Sun Yat-sen University, Guangzhou, China" + }, + { + "type": "text", + "bbox": [ + 0.277, + 0.27, + 0.726, + 0.283 + ], + "angle": 0, + "content": "lurj3@mail2.sysu.edu.cn,mengjke@gmail.com, wszheng@ieee.org" + }, + { + "type": "text", + "bbox": [ + 0.355, + 0.283, + 0.649, + 0.296 + ], + "angle": 0, + "content": "2 Peng Cheng Laboratory, Shenzhen, China" + }, + { + "type": "text", + "bbox": [ + 0.27, + 0.296, + 0.732, + 0.325 + ], + "angle": 0, + "content": "3 Key Laboratory of Machine Intelligence and Advanced Computing, Ministry of Education, Guangzhou, China" + }, + { + "type": "text", + "bbox": [ + 0.262, + 0.366, + 0.744, + 0.645 + ], + "angle": 0, + "content": "Abstract. Vision and language navigation is a task that requires an agent to navigate according to a natural language instruction. Recent methods predict sub-goals on constructed topology map at each step to enable long-term action planning. However, they suffer from high computational cost when attempting to support such high-level predictions with GCN-like models. In this work, we propose an alternative method that facilitates navigation planning by considering the alignment between instructions and directed fidelity trajectories, which refers to a path from the initial node to the candidate locations on a directed graph without detours. This planning strategy leads to an efficient model while achieving strong performance. Specifically, we introduce a directed graph to illustrate the explored area of the environment, emphasizing directionality. Then, we firstly define the trajectory representation as a sequence of directed edge features, which are extracted from the panorama based on the corresponding orientation. Ultimately, we assess and compare the alignment between instruction and different trajectories during navigation to determine the next navigation target. Our method outperforms previous SOTA method BEVBert on RxR dataset and is comparable on R2R dataset while largely reducing the computational cost. Code is available: https://github.com/iSEE-Laboratory/VLN-PRET." + }, + { + "type": "text", + "bbox": [ + 0.262, + 0.658, + 0.644, + 0.673 + ], + "angle": 0, + "content": "Keywords: Vision-and-Language Navigation \\(\\cdot\\) Planning" + }, + { + "type": "title", + "bbox": [ + 0.216, + 0.7, + 0.376, + 0.716 + ], + "angle": 0, + "content": "1 Introduction" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.734, + 0.788, + 0.843 + ], + "angle": 0, + "content": "Enabling a robot to perform tasks on behalf of humans has been a longstanding objective in AI research. One such task is vision-and-language navigation (VLN) [4, 23, 36, 41], where an agent is required to navigate to a desired location by following natural language instructions provided by humans. For example, given the instruction \"Go up stairs and stop at the top in front of a mirror.\", the agent needs to follow the instruction and stop at an appropriate location. VLN has attracted numerous research interests [2, 7, 8, 14, 16-18, 20, 26, 31, 40, 44, 47]." + } + ], + [ + { + "type": "page_number", + "bbox": [ + 0.218, + 0.116, + 0.23, + 0.127 + ], + "angle": 0, + "content": "2" + }, + { + "type": "header", + "bbox": [ + 0.272, + 0.116, + 0.331, + 0.127 + ], + "angle": 0, + "content": "Lu et al." + }, + { + "type": "image", + "bbox": [ + 0.272, + 0.144, + 0.734, + 0.336 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.214, + 0.338, + 0.788, + 0.395 + ], + "angle": 0, + "content": "Fig. 1: Comparison of SPL [3] and GFLOPs on R2R test unseen split dataset. Our method is comparable with previous SOTA methods while being more computational efficient. The computational cost of text encoder and visual encoder is omitted for fair comparison." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.416, + 0.789, + 0.642 + ], + "angle": 0, + "content": "Recent methods [2, 8, 16, 20, 44] demonstrate the effectiveness of introducing maps to enlarge the decision space to improve planning strategies. Instead of predicting low-level actions limited to short movements, these methods construct a graph to keep track of all visited and navigable locations observed so far, which enables high-level planning by expanding the action space to encompass the entire explored area. With the constructed topological maps, these methods predict actions in the global space via GCN-like models, where node features aggregate neighboring information. However, repeatedly calculating the entire map to predict actions at every step, even if the topo-map has only minor changes, is inefficient. In addition, formulating vision information of an environment in graph structure is too coarse-grained for accurate decision-making. Previous methods [2, 8] address this problem by introducing an additional branch to incorporate fine-grained information, such as local image features and bird's-eye-view features. But this strategy increases the model complexity and further escalates computational cost." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.644, + 0.79, + 0.842 + ], + "angle": 0, + "content": "In this work, we present an alternative way that supports the global decision space while achieving comparable performance. Our method, named Planning with DiReacted FidElity Trajectory(PRET), does not require calculating the entire graph at each step, nor does it rely on incorporating additional fine-grained information. The main idea is to determine the next location to navigate by evaluating the alignment between the instruction and the visual observations along different trajectories between the start point to all the unvisited nodes, as shown in Figure 2(b). Specifically, we maintain a directed fidelity trajectory-colored in red) for each unvisited node. The fidelity trajectory refers to a path from initial node to an unvisited node without detours. We assess the alignment between each trajectory and the instruction, and select the unvisited node with the highest alignment score to navigate. This planning strategy is efficient as we only need to compute for newly observed nodes. We estimate the instruction-trajectory" + } + ], + [ + { + "type": "header", + "bbox": [ + 0.687, + 0.115, + 0.732, + 0.127 + ], + "angle": 0, + "content": "PRET" + }, + { + "type": "page_number", + "bbox": [ + 0.775, + 0.117, + 0.785, + 0.127 + ], + "angle": 0, + "content": "3" + }, + { + "type": "image", + "bbox": [ + 0.307, + 0.149, + 0.691, + 0.306 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.215, + 0.312, + 0.788, + 0.368 + ], + "angle": 0, + "content": "Fig. 2: Illustration of our approaches. (a) shows our directed graph representation. Each edge is assigned with an orientation-aware panorama feature. (b) depict our planning method. We select an unvisited node-colored yellow to navigate towards next by choosing the fidelity path-colored red that best aligned with instruction." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.394, + 0.785, + 0.44 + ], + "angle": 0, + "content": "alignment with transformer and further reduce the computational cost by utilizing the KV-cache [35] technique. It also enables us to make decisions by taking into full consideration of the instruction-trajectory alignment." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.444, + 0.788, + 0.717 + ], + "angle": 0, + "content": "In order to enhance the alignment between instructions and navigation trajectories, we propose to incorporate directionality in path representation. Due to the inherent directional nature of the navigation process, it is crucial that the representation of trajectories in opposite directions is asymmetric, as visual observations are linked to observed orientation. In this work, we introduce the directed graph to depict the explored area of the environment and firstly define the trajectory representation as a sequence of directed edge features. These directed edge features are derived from the panorama based on the corresponding orientation (as depicted in Fig. 2(a)). By incorporating directionality in the graph edge features, we eliminate irrelevant panorama information and obtain a more accurate representation of the directed trajectory. Our method overcome existing methods [2,8,16] that commonly adopt the undirected graph to store panorama features on graph nodes, resulting in direction-invariant path representations, which fails to capture the important distinction between paths that traverse the same node from different directions. For instance, consider a scenario where agents pass through the same node from different directions. Despite the fundamentally different spatial contexts of these paths, the node representation remains unchanged." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.72, + 0.788, + 0.841 + ], + "angle": 0, + "content": "We conduct experiments on R2R, RxR datasets to evaluate the efficacy of our proposed methods. The results demonstrate that PRET achieves strong performance while being more efficient than previous methods. Specifically, PRET achieves comparable performance compared to previous state-of-the-art method BEVBert [2] on R2R dataset with only \\(3\\%\\) computational cost as shown in Fig. 1. On the RxR dataset, PRET outperforms previous methods and achieves the new state-of-the-art performance. Qualitative visualization also shows that our simplified model is able to learn complex backtracking strategies." + } + ], + [ + { + "type": "page_number", + "bbox": [ + 0.218, + 0.116, + 0.232, + 0.127 + ], + "angle": 0, + "content": "4" + }, + { + "type": "header", + "bbox": [ + 0.272, + 0.115, + 0.331, + 0.128 + ], + "angle": 0, + "content": "Lu et al." + }, + { + "type": "title", + "bbox": [ + 0.217, + 0.145, + 0.39, + 0.161 + ], + "angle": 0, + "content": "2 Related Work" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.175, + 0.788, + 0.356 + ], + "angle": 0, + "content": "Vision-and-Language Navigation(VLN). Early VLN methods use sequence-to-sequence LSTMs with attention mechanism to encode language features and predict local actions. SpeakerFollower [14] first introduces back translation [38] technique to ease the data scarcity problem in VLN. EnvDrop [40] proposes to augment environment by using dropout on feature space to avoid overfitting on visual input. Following this work, many environment augmentation approaches [24, 29] is proposed. Some auxiliary tasks is proposed by [31, 48] for better guidance. More recently, transformer is explored how to be adopted in VLN. PRESS [27] uses BERT [11] as the text encoder. PREVALENT [17] first propose to pretrain the transformer in VLN dataset, but still use it as encoder. More recent works [7, 18, 28] explore how to memorize navigation history and use transformer to learn strong planning strategy." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.357, + 0.788, + 0.508 + ], + "angle": 0, + "content": "Navigation Strategy in VLN. Navigation strategy plays a key role in Vision- and-Language Navigation, as VLN requires agents to navigate in unseen environments and thus needs to explore and familiarize the environment. Works like [32] investigate designing regretful agents to enable explicit exploration. Reinforcement learning methods [40, 46] have also been explored to enhance navigation strategies. SSM [44] constructs directed graphs to represent explored areas, with vision features on nodes and orientation on edges. DUET [8] employs a dual-scale transformer to make local and global predictions. AZHP [16] constructs hierarchy graphs to facilitate exploration. MetaExplore [20] explicitly predicts whether to backtrack using a separate module." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.508, + 0.788, + 0.72 + ], + "angle": 0, + "content": "Maps For Navigation. Navigation research has a long history of using SLAM [15] to construct maps [5, 6, 42] for planning. In the VLN literature, several methods [8, 44] adopt topological graphs for global planning. However, the visual representation on topological maps is too coarse-grained for decision-making. To address this limitation, [2, 47] introduce metric maps in VLN. While grid-based metric maps can precisely represent spatial layouts, they result in high computational costs. To balance representation ability and computational cost, BEVBert [2] utilizes learnable hybrid topological and metric maps. However, the computational cost remains high. In our work, we propose a novel approach that involves constructing a directed topo-map and planning with trajectory. Our method is more efficient as we perform planning with trajectory instead of directly relying on GCN-like models to encode the entire map for planning. This also allows us to incrementally calculate embeddings for new trajectories and further reduces computational cost." + }, + { + "type": "title", + "bbox": [ + 0.216, + 0.741, + 0.331, + 0.756 + ], + "angle": 0, + "content": "3 Method" + }, + { + "type": "title", + "bbox": [ + 0.216, + 0.771, + 0.44, + 0.786 + ], + "angle": 0, + "content": "3.1 Problem Formulation" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.795, + 0.787, + 0.842 + ], + "angle": 0, + "content": "In VLN with discrete environments [4,23,36,41], an environment is an undirected graph \\(\\mathcal{G} = \\{\\mathcal{V},\\mathcal{E}\\}\\), where \\(\\mathcal{V} = \\{V_i\\}_{i=1}^N\\) represents \\(N\\) navigable nodes and \\(\\mathcal{E}\\) denotes navigable edges. At the beginning of navigation, an agent is initialized" + } + ], + [ + { + "type": "header", + "bbox": [ + 0.686, + 0.114, + 0.732, + 0.127 + ], + "angle": 0, + "content": "PRET" + }, + { + "type": "page_number", + "bbox": [ + 0.776, + 0.116, + 0.786, + 0.127 + ], + "angle": 0, + "content": "5" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.146, + 0.784, + 0.191 + ], + "angle": 0, + "content": "at a starting node and given a natural language instruction \\( W = \\{w_{i}\\}_{i=1}^{L} \\) with \\( L \\) words. The agent is required to interpret this instruction to navigate to the target location." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.192, + 0.785, + 0.282 + ], + "angle": 0, + "content": "At step \\( t \\) on node \\( V_{t} \\), the agent observes (1) a panorama \\( \\mathcal{R}_t = \\{r_{t,i}\\}_{i=1}^K \\) represented by \\( K \\) images from different views, \\( r_{t,i} \\) is the extracted image feature of the \\( i \\)th view; (2) neighboring navigable nodes \\( \\mathcal{N}(V_t) \\), where \\( \\mathcal{N}(V_t) \\subset \\mathcal{V} \\); (3) orientations and coordinates of these neighboring nodes. The agent should choose a neighboring node to step to or stop at current location. Navigation is considered successful if the agent stops within 3 meters of the target." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.283, + 0.787, + 0.372 + ], + "angle": 0, + "content": "We treat navigation as a process of searching temporary target to navigate towards next among unvisited nodes on \\(\\mathcal{G}_t\\). A temporary target can be either a local neighboring node or remote unvisited node. By adding a virtual stop node connected to all nodes on the graph, we can also model the stop action. We refer fidelity trajectory to a path from initial node to an unvisited node without detours." + }, + { + "type": "title", + "bbox": [ + 0.216, + 0.394, + 0.4, + 0.408 + ], + "angle": 0, + "content": "3.2 Model Overview" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.418, + 0.788, + 0.691 + ], + "angle": 0, + "content": "The overall framework is shown in Fig. 3(a). Our planning method consists of three components: (1) We construct a directed graph for explored area. Considering the directional nature of navigation, we propose an Orientation-aware Panorama Encoder(OPE) to extract orientation-aware vision features for edges. Edge sequence is used to represent directed trajectory and align with the instruction. (2) We maintain a path (i.e. directed fidelity trajectory) to each unvisited node and assess the alignment between each path and the instruction at each step. Only few neighboring nodes are added at each step, we only compute for these relevant new paths. Since each node corresponds to a path, we calculate a path embedding with a matching assessment module(MAM) that encodes the instruction-path alignment and stores the embedding on the node. (3) We compare the path embeddings of all unvisited nodes(candidates) with candidate comparison module(CCM)(shown in Fig. 3(c)) to determine which aligns best with the instruction. We select the best aligned candidate node as the temporary target. Then the agent can navigate to the temporary target along the shortest path on constructed graph. Since the path embedding is an estimate of the alignment between instruction and a path to the node, CCM ensures that the agent navigates by fully considering the instruction-trajectory alignment." + }, + { + "type": "title", + "bbox": [ + 0.216, + 0.711, + 0.664, + 0.726 + ], + "angle": 0, + "content": "3.3 Orientation-Aware Directed Graph Construction" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.735, + 0.788, + 0.841 + ], + "angle": 0, + "content": "We construct a directed graph to support path representation, incrementally adding nodes and edges as exploration proceeds. Previous methods [2,8,44] represent visited nodes with panorama features. However, panoramas of unvisited nodes are inaccessible as the agent has not visited them. So they represent these nodes with views towards them, leading to inconsistent representation between visited and unvisited nodes. In contrast, we extract visual features for the directed edges on the graph. These edge features represent the visual observations" + } + ], + [ + { + "type": "page_number", + "bbox": [ + 0.218, + 0.116, + 0.23, + 0.127 + ], + "angle": 0, + "content": "6" + }, + { + "type": "header", + "bbox": [ + 0.272, + 0.115, + 0.331, + 0.127 + ], + "angle": 0, + "content": "Lu et al." + }, + { + "type": "image", + "bbox": [ + 0.223, + 0.147, + 0.784, + 0.265 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.223, + 0.265, + 0.616, + 0.385 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.626, + 0.267, + 0.752, + 0.378 + ], + "angle": 0, + "content": null + }, + { + "type": "image_footnote", + "bbox": [ + 0.221, + 0.387, + 0.782, + 0.398 + ], + "angle": 0, + "content": "visited node current node unvisited node start token shared edge token unvisited neighboring edge token stop token" + }, + { + "type": "image_caption", + "bbox": [ + 0.214, + 0.402, + 0.788, + 0.5 + ], + "angle": 0, + "content": "Fig. 3: Illustration of our model. (a) is the overall framework of our method. At each step, we update the graph, extract path embeddings, and predict actions. (b) depicts the matching assessment module(MAM). Each token is an edge feature. We compute path embeddings for each newly observed nodes with cross-modal transformer and impose a causal mask to reduce computational cost. (c) shows the candidate comparison module(CCM). We gather path embeddings of unvisited nodes and forward them into a single layer transformer followed by a MLP to predict temporary target." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.53, + 0.785, + 0.574 + ], + "angle": 0, + "content": "when facing specific directions from each node. As edges rely only on observable views rather than node panoramas, they enable consistent representation for both visited and unvisited nodes." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.576, + 0.788, + 0.651 + ], + "angle": 0, + "content": "Let \\(\\mathcal{G}_t = \\{\\mathcal{V}_t, \\mathcal{E}_t\\}\\) be the directed graph at step \\(t\\). Neighboring nodes \\(\\mathcal{N}(V_t)\\) are observed, and some nodes in \\(\\mathcal{N}(V_t)\\) have already existed in \\(\\mathcal{G}_t\\) while others are newly observed and added to \\(\\mathcal{G}_t\\). Besides, directed edges from \\(V_t\\) to nodes in \\(\\mathcal{N}(V_t)\\) together with features extracted by orientation-aware panorama encoder(OPE) will also be added in \\(G_t\\), as shown in Fig. 3(a)." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.676, + 0.788, + 0.766 + ], + "angle": 0, + "content": "Visual Representation For Edges. We extract orientation-aware panorama feature by using relative orientation as query to attend panorama features in cross-attention manner, as shown in the OPE module in Fig. 3(a). Assume that \\((\\phi, \\theta)\\) represents the relative heading and elevation of a specific edge when compared to the agent's current orientation. To extract feature for the edge, we first encode \\((\\phi, \\theta)\\) as follows:" + }, + { + "type": "equation", + "bbox": [ + 0.359, + 0.778, + 0.787, + 0.798 + ], + "angle": 0, + "content": "\\[\nx ^ {a} = \\left[ \\sin (\\phi), \\cos (\\phi), \\sin (\\theta), \\cos (\\theta) \\right] W ^ {a}, \\tag {1}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.807, + 0.788, + 0.843 + ], + "angle": 0, + "content": "where \\(W^{a}\\in \\mathbb{R}^{4\\times d}\\) are learnable weights and \\(d\\) is the dimension of feature. Then we denote \\(X_{t}^{a} = \\{x_{t,i}^{a}\\}_{i = 1}^{\\lfloor \\mathcal{N}(V_{t})\\rfloor}\\) as orientation features of adjacent edges of \\(V_{t}\\)." + } + ], + [ + { + "type": "header", + "bbox": [ + 0.686, + 0.115, + 0.732, + 0.127 + ], + "angle": 0, + "content": "PRET" + }, + { + "type": "page_number", + "bbox": [ + 0.776, + 0.116, + 0.786, + 0.126 + ], + "angle": 0, + "content": "7" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.146, + 0.788, + 0.239 + ], + "angle": 0, + "content": "We then encode the panorama feature. As the panorama \\(\\mathcal{R}_t = \\{r_{t,i}\\}_{i=1}^K\\) is represented by \\(K\\) images from different views and each view corresponds to a fixed orientation, we can also get the relative heading and elevation of each view. To encode the panorama, we first concatenate each view feature \\(r_{t,i}\\) with corresponding relative orientation \\((\\phi_{t,i},\\theta_{t,i})\\) and then map it to dimension \\(d\\) via a linear projection:" + }, + { + "type": "equation", + "bbox": [ + 0.322, + 0.248, + 0.786, + 0.265 + ], + "angle": 0, + "content": "\\[\nx _ {t, i} ^ {p} = \\left[ r _ {t, i}; \\sin \\left(\\phi_ {t, i}\\right), \\cos \\left(\\phi_ {t, i}\\right), \\sin \\left(\\theta_ {t, i}\\right), \\cos \\left(\\theta_ {t, i}\\right) \\right] W ^ {p}. \\tag {2}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.215, + 0.273, + 0.784, + 0.319 + ], + "angle": 0, + "content": "Then we denote \\( X_{t}^{p} = \\{x_{t,i}^{p}\\}_{i=1}^{K} \\) as the panoramic view representation. By considering views as patches of the panorama and orientations as position embedding, \\( X_{t}^{p} \\) is like the encoded ViT [12] input." + }, + { + "type": "text", + "bbox": [ + 0.215, + 0.32, + 0.784, + 0.348 + ], + "angle": 0, + "content": "We adopt transformer decoder [43] to extract orientation-aware panorama feature:" + }, + { + "type": "equation", + "bbox": [ + 0.376, + 0.35, + 0.785, + 0.366 + ], + "angle": 0, + "content": "\\[\nE _ {t} = \\operatorname {T r a n s f o r m e r D e c o d e r} \\left(X _ {t} ^ {a}, X _ {t} ^ {p}\\right), \\tag {3}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.372, + 0.788, + 0.464 + ], + "angle": 0, + "content": "where \\( E_{t} = \\{e_{t,i}\\}_{i = 1}^{\\lfloor \\mathcal{N}(V_{t})\\rfloor} \\) are extracted edge features. By taking \\( X_{t}^{a} \\) as the query input and \\( X_{t}^{p} \\) as the key-value input, panorama is attended by orientation in the first layer. In the subsequent layers, both orientation and vision information are used to query relevant views in the panorama. Therefore, each directed edge only focuses on a specific region of the panorama, thus orientation-aware panorama feature is extracted." + }, + { + "type": "title", + "bbox": [ + 0.215, + 0.485, + 0.54, + 0.501 + ], + "angle": 0, + "content": "3.4 Planning with Fidelity Trajectory" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.508, + 0.788, + 0.614 + ], + "angle": 0, + "content": "We maintain a directed fidelity trajectory with a stack during navigation for each unvisited node on \\(\\mathcal{G}_t\\). When a new node is added, it is pushed into the stack. When the agent needs to backtrack, the top node is popped off the stack until a different node than the top is reached. In this way, detours are removed by popping nodes off the stack. With the maintained trajectory of each node, we propose MAM to assess its alignment with instruction. We then determine an unvisited node on \\(\\mathcal{G}_t\\) to navigate next by comparing this alignment with CCM." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.634, + 0.787, + 0.695 + ], + "angle": 0, + "content": "Matching Assessment Module(MAM). MAM is a multi-layer transformer decoder that extracts a path embedding for each path, which estimates how much the trajectory match with the instruction. These path embeddings are then stored on the corresponding nodes in graph \\(\\mathcal{G}_t\\)." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.695, + 0.788, + 0.755 + ], + "angle": 0, + "content": "For a single path with length \\( l \\), we represent the path with the edge features along it, denoted as \\( X^{h} = [e_{1}, e_{2}, \\dots, e_{l}] \\). \\( \\{e_{i}\\}_{i=1}^{l} \\) are the edge features on the path. Then we add it with position embedding and forward it together with text embedding \\( X^{w} \\) into MAM like follows:" + }, + { + "type": "equation", + "bbox": [ + 0.367, + 0.763, + 0.482, + 0.78 + ], + "angle": 0, + "content": "\\[\nX ^ {h \\prime} = X ^ {h} + P _ {l},\n\\]" + }, + { + "type": "equation", + "bbox": [ + 0.371, + 0.776, + 0.785, + 0.801 + ], + "angle": 0, + "content": "\\[\nX ^ {o} = \\operatorname {T r a n s f o r m e r D e c o d e r} \\left(X ^ {h t}, X ^ {w}\\right), \\tag {4}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.81, + 0.787, + 0.842 + ], + "angle": 0, + "content": "where \\( P_{l} \\) is a matrix including \\( l \\) position embeddings [43] and \\( X^{w} \\) is the text embedding extracted by a text encoder, which is a multi-layer transformer [43]." + } + ], + [ + { + "type": "page_number", + "bbox": [ + 0.218, + 0.116, + 0.23, + 0.127 + ], + "angle": 0, + "content": "8" + }, + { + "type": "header", + "bbox": [ + 0.272, + 0.115, + 0.331, + 0.128 + ], + "angle": 0, + "content": "Lu et al." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.147, + 0.788, + 0.223 + ], + "angle": 0, + "content": "By utilizing two sequences from two modalities as inputs, the transformer is capable of effectively estimating their alignment, as extensively demonstrated in Vision and Language Pretraining studies [9, 22, 25]. The last token of the encoded sequence \\( X^o \\) provides the path embedding, which is an estimation of the text-path alignment." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.222, + 0.788, + 0.344 + ], + "angle": 0, + "content": "To support STOP action, we add a stop node that connected to all nodes in \\(\\mathcal{G}_t\\). Since we represent a path using edge features, we assign a learnable stop token, initialized with a zero vector, to all edges that point to the stop node. The path embedding corresponds to the stop node is stored on current node. So the agent decides to stop at current location when it finds that the path up to here appending with stop token matches instruction the best. Here stop token serves as the End-of-Sentence token in text generation. For each path, we also add a learnable start token that serves as the Start-of-Sentence token." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.344, + 0.789, + 0.587 + ], + "angle": 0, + "content": "At step \\( t \\), considering the stop node, assume there are \\( N \\) neighboring unvisited nodes to be updated. Computing path embeddings separately for these nodes requires multiple forward passes, which is inefficient. Actually, we can compute these embeddings in a single forward pass, as shown in Fig. 3(b). Note that these paths share the same prefix from the initial node to the current node. We can avoid duplicated computation by imposing a causal mask in the self-attention block. The causal mask is a lower triangular matrix that only allows a token to attend to its previous tokens. We merge these causal masks together and forward all relevant tokens into the transformer block. The merged mask still satisfies the constraint that each token can only attend to its previous tokens. We further reduce the computational cost by adopting the KV-cache technique [35], which stores the shared prefix tokens to avoid unnecessary re-calculation. In this way, we can equivalently compute the path embeddings for multiple newly observed nodes in a single forward pass. The last several output tokens colored yellow and green are path embeddings and used to update the graph, i.e., stored on corresponding graph nodes." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.604, + 0.789, + 0.742 + ], + "angle": 0, + "content": "Candidate Comparison Module(CCM). CCM compares the path embeddings from candidate nodes to predict a temporary target. As each path embedding encodes the alignment between a trajectory and the instruction, the CCM actually selects a path that is most aligned with the instruction. In this way, we make navigation planning fully according to the instruction-trajectory alignment. As shown in Fig. 3(c), we collect path embeddings \\( X_{t}^{e} \\) and input them into CCM. CCM consists of a single-layer transformer encoder [43] and a MLP. The transformer encoder is responsible for the comparison of the candidates and the MLP maps the encoded path embedding to a 1D score:" + }, + { + "type": "equation", + "bbox": [ + 0.397, + 0.749, + 0.61, + 0.765 + ], + "angle": 0, + "content": "\\[\nX _ {t} ^ {e \\prime} = \\operatorname {T r a n s f o r m e r L a y e r} \\left(X _ {t} ^ {e}\\right),\n\\]" + }, + { + "type": "equation", + "bbox": [ + 0.408, + 0.768, + 0.785, + 0.784 + ], + "angle": 0, + "content": "\\[\ns _ {t} = \\operatorname {M L P} \\left(X _ {t} ^ {e ^ {\\prime}}\\right), \\tag {5}\n\\]" + }, + { + "type": "equation", + "bbox": [ + 0.407, + 0.788, + 0.53, + 0.802 + ], + "angle": 0, + "content": "\\[\np _ {t} = \\operatorname {s o f t m a x} \\left(s _ {t}\\right).\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.81, + 0.788, + 0.842 + ], + "angle": 0, + "content": "Score \\( s_t \\) reflects the relative text-path alignment after comparison. An alternative way is removing the transformer and directly compute the alignment score \\( s_t \\)." + } + ], + [ + { + "type": "header", + "bbox": [ + 0.686, + 0.115, + 0.732, + 0.127 + ], + "angle": 0, + "content": "PRET" + }, + { + "type": "page_number", + "bbox": [ + 0.776, + 0.116, + 0.786, + 0.127 + ], + "angle": 0, + "content": "9" + }, + { + "type": "text", + "bbox": [ + 0.215, + 0.147, + 0.784, + 0.238 + ], + "angle": 0, + "content": "However, this method assesses each path independently without any comparison, which can make the decision-making process more challenging and potentially result in a decrease in performance(demonstrated in Sec. 4). It is then normalized with softmax and we select the highest score node as the temporary target. The agent will navigate to it along the shortest path on \\(\\mathcal{G}_t\\). If the stop node is selected, the agent believes the instruction is completed and stops the navigation." + }, + { + "type": "title", + "bbox": [ + 0.216, + 0.258, + 0.496, + 0.273 + ], + "angle": 0, + "content": "3.5 Pretraining and Fine-tuning" + }, + { + "type": "text", + "bbox": [ + 0.215, + 0.28, + 0.785, + 0.461 + ], + "angle": 0, + "content": "Pretraining. Pretraining is proved to be helpful in previous works [2,7,17,33]. PREVALENT [17] synthesized a substantial amount of data for pretraining, and we utilized the same data as their work. We use Masked Language Modeling(MLM) [11] for pretraining. Concretely, we randomly replace \\(15\\%\\) of the input tokens with a special [MASK] token and forward these masked tokens into the text encoder. We also extract a sequence of edge features to represent the path. Then the encoded text tokens and path representation are input into an additional transformer decoder. The corresponding outputs are used to predict the masked tokens. MLM helps the model learn aligned text and path representations. In VLN, a key difference from other vision-and-language tasks is that we align a text sequence with a sequence of vision representations rather than a single image." + }, + { + "type": "text", + "bbox": [ + 0.215, + 0.481, + 0.784, + 0.527 + ], + "angle": 0, + "content": "Fine-tuning. We train the agent with a mixture of teacher-forcing and student-forcing as previous methods [2,8]. In the teacher-forcing stage, the agent navigate according the ground truth actions \\(a_{t}^{*}\\) and the loss is calculated as follows:" + }, + { + "type": "equation", + "bbox": [ + 0.405, + 0.537, + 0.785, + 0.575 + ], + "angle": 0, + "content": "\\[\n\\mathcal {L} _ {T F} = \\frac {1}{T _ {T F}} \\sum_ {t = 1} ^ {T _ {T F}} \\mathrm {C E} \\left(p _ {t}, a _ {t} ^ {*}\\right), \\tag {6}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.215, + 0.584, + 0.784, + 0.644 + ], + "angle": 0, + "content": "CE is the cross entropy loss and \\( T_{TF} \\) is the number of navigate steps. In the student-forcing stage, we sample an action from the distribution predicted by the agent so that the agent can explore the environment and reduce the exposure bias. We train the agent with heuristic pseudo label \\( a_t^{pseudo} \\):" + }, + { + "type": "equation", + "bbox": [ + 0.389, + 0.655, + 0.784, + 0.692 + ], + "angle": 0, + "content": "\\[\n\\mathcal {L} _ {S F} = \\frac {1}{T _ {S F}} \\sum_ {t = 1} ^ {T _ {S F}} \\mathrm {C E} \\left(p _ {t}, a _ {t} ^ {\\text {p s e u d o}}\\right). \\tag {7}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.215, + 0.701, + 0.784, + 0.791 + ], + "angle": 0, + "content": "When the agent deviate from ground truth path, we take the nearest node on ground truth path as the pseudo label to encourage agent to learn backtracking strategy. If there is no unvisited node on ground truth path, we take the nearest node on the shortest path from current node to target node as the pseudo label. The agent navigate two times to compute \\(\\mathcal{L}_{TF}\\) and \\(\\mathcal{L}_{SF}\\). The total loss is the weighted sum of them:" + }, + { + "type": "equation", + "bbox": [ + 0.413, + 0.802, + 0.784, + 0.816 + ], + "angle": 0, + "content": "\\[\n\\mathcal {L} = \\lambda \\mathcal {L} _ {T F} + (1 - \\lambda) \\mathcal {L} _ {S F}, \\tag {8}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.216, + 0.825, + 0.434, + 0.84 + ], + "angle": 0, + "content": "where \\(\\lambda \\in (0,1)\\) is the weight." + } + ], + [ + { + "type": "page_number", + "bbox": [ + 0.218, + 0.116, + 0.236, + 0.127 + ], + "angle": 0, + "content": "10" + }, + { + "type": "header", + "bbox": [ + 0.272, + 0.115, + 0.331, + 0.127 + ], + "angle": 0, + "content": "Lu et al." + }, + { + "type": "title", + "bbox": [ + 0.217, + 0.145, + 0.377, + 0.163 + ], + "angle": 0, + "content": "4 Experiments" + }, + { + "type": "title", + "bbox": [ + 0.217, + 0.176, + 0.533, + 0.192 + ], + "angle": 0, + "content": "4.1 Datasets and Evaluation Metrics" + }, + { + "type": "text", + "bbox": [ + 0.217, + 0.2, + 0.787, + 0.276 + ], + "angle": 0, + "content": "R2R. Room-to-Room (R2R) dataset [4] contains 7,189 trajectories, with 3 instructions per trajectory, across 90 scenes. These scenes are divided into train, val unseen and test unseen splits with 61, 11 and 18 scenes respectively. A val seen split with the same scenes as the train split is also provided. All paths in R2R are the shortest paths between the start and target nodes." + }, + { + "type": "text", + "bbox": [ + 0.217, + 0.297, + 0.787, + 0.373 + ], + "angle": 0, + "content": "RxR. Room-across-Room(RxR) [23] is a large multilingual dataset. The dataset has 126,000 instructions total, with 42,000 instructions in each of the three languages: English, Hindi, and Telugu. The paths in RxR are longer than those in R2R and are not the shortest possible routes. Additionally, the instructions in RxR are longer and contain more detailed descriptions compared to R2R." + }, + { + "type": "text", + "bbox": [ + 0.217, + 0.394, + 0.787, + 0.546 + ], + "angle": 0, + "content": "Evaluation Metrics. We adopt the following evaluation metrics on R2R: (1) Trajectory length (TL): the average length of the agent's path in meters. (2) Navigation error (NE): the average distance between the agent's final position and the target location. (3) Success rate (SR): the ratio of successful navigations, where \\( NE < 3m \\) is considered successful. (4) SR penalized by path length (SPL) [3]: TL longer than the shortest path is penalized. As paths in RxR dataset are shortest paths, TL and SPL are unsuitable. Instead, on RxR we use nDTW and sDTW [21] to measure trajectory similarity between the agent and ground truth. Please note that the fidelity trajectory that is removed detours is solely used for planning. It is not the agent's actual trajectory used for evaluation." + }, + { + "type": "title", + "bbox": [ + 0.217, + 0.567, + 0.456, + 0.583 + ], + "angle": 0, + "content": "4.2 Implementation Details" + }, + { + "type": "text", + "bbox": [ + 0.217, + 0.592, + 0.787, + 0.682 + ], + "angle": 0, + "content": "Module architecture. Text backbone is a 6 layer transformer encoder initialized with pretrained ALBEF [25]. For the multilingual RxR dataset, we use 12 layer mRoberta [10] instead. The layer numbers of the OPE, MAM, and CCM are 2, 4, and 1 respectively. All the hidden layer size is 768. Image features is extracted by DINOv2 [34]. We also report results with CLIP-ViT-base [37] feature for fair comparison." + }, + { + "type": "text", + "bbox": [ + 0.217, + 0.704, + 0.787, + 0.84 + ], + "angle": 0, + "content": "Training details. On R2R dataset, we first pretrain PRET with learning rate 2e-5 and batch size 16 on a single 24G 4090 GPU for 100,000 iterations(\\(\\sim\\) 5 hours). Then we fine-tune the model with learning rate 1e-5 and batch size 8 on 4090 for 100,000 iterations(\\(\\sim\\) 25 hours). Optimizer is AdamW [30]. \\(\\lambda\\) is set to 0.2. Augmented dataset [17] is used in both pretraining and fine-tuning. The best result is selected by SPL on val unseen split. On RxR dataset, we also follow the pretrain and fine-tune paradigm. Due to the longer instruction and trajectory, we use a smaller batch size 4. Marky [45] is used as augmented data. The best result is selected by sDTW on val unseen split." + } + ], + [ + { + "type": "header", + "bbox": [ + 0.687, + 0.115, + 0.732, + 0.127 + ], + "angle": 0, + "content": "PRET" + }, + { + "type": "page_number", + "bbox": [ + 0.769, + 0.116, + 0.785, + 0.127 + ], + "angle": 0, + "content": "11" + }, + { + "type": "table_caption", + "bbox": [ + 0.217, + 0.145, + 0.785, + 0.173 + ], + "angle": 0, + "content": "Table 1: Comparison with other methods on R2R dataset. SPL is considered as the primary evaluation metric." + }, + { + "type": "table", + "bbox": [ + 0.235, + 0.184, + 0.775, + 0.445 + ], + "angle": 0, + "content": "
MethodsVal SeenVal UnseenTest Unseen
TLNE↓SR↑SPL↑TLNE↓SR↑SPL↑TLNE↓SR↑SPL↑
Seq2Seq-SF [4]11.336.0139-8.397.8122-8.137.852818
Speaker-Follower [14]-3.3666--6.6235-14.826.623528
RCM [46]10.653.5367-11.466.0943-11.976.124338
Regretful [32]-3.236963-5.325041-5.695640
EnvDrop [40]11.003.99625910.705.22524811.665.235147
PREVALENT [17]10.323.67696510.194.71585310.515.305451
NvEM [1]11.093.44696511.834.27605512.984.375854
SSM [44]14.703.10716220.704.32624520.404.576146
RecBert [18]11.132.90726812.013.93635712.354.096357
HAMT [7]11.152.51767211.462.29666112.273.936560
MTVM [28]-2.677469-3.736659-3.856559
DUET [8]12.322.28797313.943.31726014.733.656959
AZHP [16]----14.053.15726114.953.527160
Meta-Explore [20]11.952.11817513.093.22726214.253.577161
GridMM [47]----13.272.83756414.433.357362
BEVBert [2]13.562.17817414.552.81756415.873.137362
Ours(CLIP)11.482.60746912.213.12716313.873.127262
Ours(DINOv2)11.252.41787211.872.90746512.213.097264
" + }, + { + "type": "title", + "bbox": [ + 0.217, + 0.475, + 0.323, + 0.489 + ], + "angle": 0, + "content": "4.3 Results" + }, + { + "type": "text", + "bbox": [ + 0.216, + 0.507, + 0.787, + 0.734 + ], + "angle": 0, + "content": "Comparison on R2R. Tab. 1 compares our approach with previous methods. On the val unseen and test unseen split, our method achieves comparable performance with previous SOTA method BEVBert on the primary metric SPL, while our method significantly reduces the computational cost. Besides, BEVBert [2] adopts additional depth information to construct bird's-eye view input. On the val seen split, our method performs sub-optimal as PRET does not take the environment layout as input, which means it less tends to overfit on the seen environment. Also, in the VLN literature, all visual features are pre-extracted and remains unchanged, the fixed feature may not contains appropriate feature for navigation. Therefore, we also investigate different vision features for VLN, as shown in Tab. 1. We find that DINOv2 [34] feature is better than CLIP feature. DINOv2 is trained on large-scale curated data in self-supervised manner, which leads to more general purpose feature. Moreover, DINOv2 feature contains geometric information as it can be used to predict depth map according to [34]. Geometric information is shown helpful in [19]." + }, + { + "type": "text", + "bbox": [ + 0.216, + 0.735, + 0.787, + 0.84 + ], + "angle": 0, + "content": "Tab. 2 shows the parameter count, computational cost, and latency of different graph-based methods. We present GFLOPs of different navigation steps. Our method contains fewer parameters as we use a single stream model unlike [2,8] use dual stream transformer. Besides, our decoder computational cost is only \\(19\\%\\) of DUET and \\(3\\%\\) of BEVBert at step 10. These suggests that our method is significantly more efficient than previous SOTA graph-based methods. As shown in Tab. 2, this results in lower training and inference latency." + } + ], + [ + { + "type": "page_number", + "bbox": [ + 0.218, + 0.116, + 0.236, + 0.127 + ], + "angle": 0, + "content": "12" + }, + { + "type": "header", + "bbox": [ + 0.272, + 0.115, + 0.331, + 0.127 + ], + "angle": 0, + "content": "Lu et al." + }, + { + "type": "table_caption", + "bbox": [ + 0.214, + 0.145, + 0.788, + 0.228 + ], + "angle": 0, + "content": "Table 2: Comparison of latency and computational cost on R2R val unseen. * indicates we reproduce DUET with CLIP feature for fair comparison. When computing GFLOPs, we simplify the calculation by assuming that each vertex of the graph is connected to 4 adjacent vertices (average degree in R2R), and at each step, 3 new vertices are encountered. The computational cost of text encoder and visual encoder is omitted for fair comparison." + }, + { + "type": "table", + "bbox": [ + 0.261, + 0.24, + 0.739, + 0.307 + ], + "angle": 0, + "content": "
MethodsR2R val unseenLatency(ms)GFLOPsParams
TLNE↓SR↑SPL↑traininference1 step10 steps20 steps
DUET* [8]14.03.271.661.111.88.62.44.87.590M
BEVBert [2]14.62.874.963.616.210.633.735.537.590M
Ours12.23.171.062.77.74.10.60.91.264M
" + }, + { + "type": "table_caption", + "bbox": [ + 0.214, + 0.319, + 0.788, + 0.375 + ], + "angle": 0, + "content": "Table 3: Comparison with previous methods on RxR dataset. †indicates approaches that utilize additional augmented data. nDTW represents the normalized DTW distance between the agent's trajectory and ground truth path. sDTW is the nDTW weighted by success rate." + }, + { + "type": "table", + "bbox": [ + 0.276, + 0.387, + 0.724, + 0.528 + ], + "angle": 0, + "content": "
MethodsVal SeenVal Unseen
NE↓SR↑nDTW↑sDTW↑NE↓SR↑nDTW↑sDTW↑
LSTM [23]10.725.242.220.710.922.838.918.2
EnvDrop+ [39]-----43.655.7-
HAMT [7]-59.465.350.9-56.563.148.3
EnvEdit [24]-67.271.158.5-62.868.554.6
MPM [13]-67.771.058.9-63.567.754.5
MARVEL [45]†3.075.979.168.84.564.870.857.5
BEVBert [2]†3.275.076.366.74.068.569.658.6
Ours(CLIP)†2.677.078.067.53.371.271.860.4
Ours(DINOv2)†2.479.380.470.73.272.873.462.4
" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.557, + 0.788, + 0.709 + ], + "angle": 0, + "content": "Comparison on RxR. Tab. 3 reports the results on the RxR dataset. RxR is challenging as instructions and paths in RxR are longer than R2R. While the alignment between path and detailed path description is what our method skills at, our method outperforms previous methods among all metrics at a lower computational cost. PRET achieves a \\(1.8\\%\\) improvement on the main metrics sDTW on the val unseen split, demonstrating that our model better understands and follows instructions. This result highlights the advantage of our planning with trajectory strategy, which considers the alignment between instructions and paths and enables efficient decision-making in the global space. Additionally, with DINOv2 features, the result is further improved." + }, + { + "type": "title", + "bbox": [ + 0.215, + 0.731, + 0.39, + 0.745 + ], + "angle": 0, + "content": "4.4 Ablation Study" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.755, + 0.787, + 0.802 + ], + "angle": 0, + "content": "We conducted ablation experiments on various components of PRET, including the directed path representation and module ablation. Results are reported on the R2R val unseen split." + } + ], + [ + { + "type": "header", + "bbox": [ + 0.686, + 0.115, + 0.732, + 0.127 + ], + "angle": 0, + "content": "PRET" + }, + { + "type": "page_number", + "bbox": [ + 0.769, + 0.116, + 0.786, + 0.127 + ], + "angle": 0, + "content": "13" + }, + { + "type": "table_caption", + "bbox": [ + 0.216, + 0.15, + 0.457, + 0.178 + ], + "angle": 0, + "content": "Table 4: Comparison of undirected and directed path representation." + }, + { + "type": "table", + "bbox": [ + 0.231, + 0.19, + 0.44, + 0.233 + ], + "angle": 0, + "content": "
MethodsTLNE↓SR↑SPL↑
undirected15.623.5968.2856.77
directed11.872.9073.7865.16
" + }, + { + "type": "table_caption", + "bbox": [ + 0.49, + 0.145, + 0.744, + 0.159 + ], + "angle": 0, + "content": "Table 5: Ablation study on modules." + }, + { + "type": "table", + "bbox": [ + 0.477, + 0.171, + 0.752, + 0.237 + ], + "angle": 0, + "content": "
MethodsTLNE↓SR↑SPL↑
1 MAM12.043.9962.3254.48
2 MAM+CCM12.153.5465.9457.32
3 MAM+OPE12.183.1571.6063.07
4 MAM+OPE+CCM11.872.9073.7865.16
" + }, + { + "type": "image", + "bbox": [ + 0.359, + 0.251, + 0.646, + 0.302 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.243, + 0.303, + 0.757, + 0.318 + ], + "angle": 0, + "content": "Fig. 4: Comparison of orientation panoramic view and single candidate view." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.348, + 0.788, + 0.515 + ], + "angle": 0, + "content": "Directionality in Path Representation. We compare whether to incorporate directionality for path representations in Tab. 4. Undirected path is represented by a sequence of node features. The node features is extracted by forwarding panoramic views into a 2-layer transformer encoder and averages the output tokens. Directed path is represented by a sequence of orientation-aware panorama features. According to the results, directed path representation outperforms undirected node features among all metrics. Specifically, it gains \\(8.39\\%\\) improvement on SPL and \\(5.5\\%\\) on SR. This indicates that edge feature is more suitable to represent the directional navigation process and align with the instruction. Node features does not distinguish different path directions and provide redundant information in a path, which hinders the alignment." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.538, + 0.788, + 0.658 + ], + "angle": 0, + "content": "Module ablation. Tab. 5 shows ablation results on modules. Row 4 is the full model introduced in Sec. 3. OPE extracts orientation-aware features for edges, MAM extracts path embeddings while CCM compares them to select the temporary target. When removing CCM (Row 3), we only predict a single alignment score rather than a path embedding for each path without comparison. The score is then normalized and used to predict the temporary goal. Performance drops \\(2.09\\%\\) in this case, as selecting a path based solely on a single score, without comparing different paths, is more difficult." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.66, + 0.788, + 0.84 + ], + "angle": 0, + "content": "In row 2, we remove OPE, panorama features are not used and only a single 60-degree view towards each neighboring node is adopted as edge features. Performance drops \\(7.84\\%\\) as the limited field-of-view cannot fully represent the path. The trajectory represented by these views is discrete and does not form a continuous change of views. In some cases, a single view towards candidates provides no information for navigation, as shown in Fig. 4. When go up the stairs, the upper view only sees the wall and does not help the navigation at all. While using the panorama, the agent can dynamically see a larger region. These results show our graph representation with panorama encoder provides rich representation for trajectories. Row 1 demonstrates the performance when using only MAM. The performance is poor without other model components, demonstrating the effectiveness of these modules." + } + ], + [ + { + "type": "page_number", + "bbox": [ + 0.218, + 0.116, + 0.236, + 0.127 + ], + "angle": 0, + "content": "14" + }, + { + "type": "header", + "bbox": [ + 0.272, + 0.115, + 0.331, + 0.127 + ], + "angle": 0, + "content": "Lu et al." + }, + { + "type": "title", + "bbox": [ + 0.217, + 0.147, + 0.371, + 0.16 + ], + "angle": 0, + "content": "4.5 Visualization" + }, + { + "type": "image", + "bbox": [ + 0.24, + 0.171, + 0.462, + 0.297 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.236, + 0.298, + 0.458, + 0.323 + ], + "angle": 0, + "content": "Instruction: Face directly away from stairs and walk through doorway straight ahead, right immediately, walk past bed and turn right into closet, walk straight and stop in front of toilet." + }, + { + "type": "image_caption", + "bbox": [ + 0.343, + 0.325, + 0.359, + 0.335 + ], + "angle": 0, + "content": "(a)" + }, + { + "type": "image_caption", + "bbox": [ + 0.517, + 0.169, + 0.617, + 0.18 + ], + "angle": 0, + "content": "undirected trajectory" + }, + { + "type": "image", + "bbox": [ + 0.499, + 0.182, + 0.63, + 0.214 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.499, + 0.214, + 0.63, + 0.258 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.499, + 0.258, + 0.63, + 0.299 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.565, + 0.306, + 0.721, + 0.315 + ], + "angle": 0, + "content": "Instruction: Turn right and go up the stairs." + }, + { + "type": "image_caption", + "bbox": [ + 0.665, + 0.169, + 0.754, + 0.18 + ], + "angle": 0, + "content": "directed trajectory" + }, + { + "type": "image", + "bbox": [ + 0.643, + 0.182, + 0.772, + 0.214 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.643, + 0.214, + 0.772, + 0.258 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.643, + 0.258, + 0.772, + 0.3 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.629, + 0.324, + 0.646, + 0.335 + ], + "angle": 0, + "content": "(b)" + }, + { + "type": "image_caption", + "bbox": [ + 0.214, + 0.342, + 0.788, + 0.385 + ], + "angle": 0, + "content": "Fig. 5: (a) Visualization of the agent's navigation process, showcasing its ability to learn a backtracking strategy. (b) Visualizing attention weights in OPE to illustrate the distinction between undirected and directed trajectory representations." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.402, + 0.785, + 0.508 + ], + "angle": 0, + "content": "Fig. 5(a) shows a qualitative example of our model's behavior. Initially placed at the green starting node, the agent wrongly decides to explore the upper left area(blue path). It takes many unnecessary detours there, but through comparing the trajectory alignment with the instructions, realizes these paths do not match the text well. Therefore, it backtracks to the start and correctly navigates to the destination(red path). This demonstrates the capability of PRET to learn complex backtracking behaviors." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.508, + 0.788, + 0.629 + ], + "angle": 0, + "content": "Fig. 5(b) shows the difference between undirected and directed trajectory representations. The former represents a path using panorama features stored on nodes, which eliminates directional discrepancies since the same node provides the same feature regardless of orientation, and contains redundant information compared to \"stairs\" in the instruction. In contrast, our proposed approach represents a path using orientation-aware panorama features on edges. This eliminates redundant information and focuses on the \"stairs\" better captures the directional navigation process." + }, + { + "type": "title", + "bbox": [ + 0.216, + 0.647, + 0.36, + 0.663 + ], + "angle": 0, + "content": "5 Conclusion" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.674, + 0.789, + 0.841 + ], + "angle": 0, + "content": "In this work, we present an efficient method that planning with directed fidelity trajectory, explicitly leveraging the alignment between instructions and trajectories for navigation. Additionally, we consider the directional nature of navigation and introduce a directed graph construction during navigation, storing vision information on directed edges. This approach provides rich directed path representation and enhances instruction-trajectory alignment. Experiments demonstrate that our method is achieves strong performance while being significantly more efficient than previous methods. Our method has certain limitations that we need to address. For example, navigation requires environment layout information in some cases. We recognize the need to investigate potential solutions to overcome these challenges." + } + ], + [ + { + "type": "header", + "bbox": [ + 0.686, + 0.115, + 0.732, + 0.127 + ], + "angle": 0, + "content": "PRET" + }, + { + "type": "page_number", + "bbox": [ + 0.769, + 0.116, + 0.786, + 0.127 + ], + "angle": 0, + "content": "15" + }, + { + "type": "title", + "bbox": [ + 0.218, + 0.145, + 0.393, + 0.163 + ], + "angle": 0, + "content": "Acknowledgments" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.177, + 0.788, + 0.239 + ], + "angle": 0, + "content": "This work was supported partially by the National Key Research and Development Program of China (2023YFA1008503), NSFC(U21A20471, 62206315), Guangdong NSF Project (No. 2023B1515040025, 2020B1515120085, 2024A1515-010101), Guangzhou Basic and Applied Basic Research Scheme(2024A04J4067)." + }, + { + "type": "title", + "bbox": [ + 0.217, + 0.259, + 0.323, + 0.275 + ], + "angle": 0, + "content": "References" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.289, + 0.785, + 0.317 + ], + "angle": 0, + "content": "1. An, D., Qi, Y., Huang, Y., Wu, Q., Wang, L., Tan, T.: Neighbor-view enhanced model for vision and language navigation. In: ACM MM (2021)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.318, + 0.785, + 0.345 + ], + "angle": 0, + "content": "2. An, D., Qi, Y., Li, Y., Huang, Y., Wang, L., Tan, T., Shao, J.: Bevbert: Multimodal map pre-training for language-guided navigation. In: ICCV (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.346, + 0.787, + 0.386 + ], + "angle": 0, + "content": "3. Anderson, P., Chang, A.X., Chaplot, D.S., Dosovitskiy, A., Gupta, S., Koltun, V., Kosecka, J., Malik, J., Mottaghi, R., Savva, M., Zamir, A.R.: On evaluation of embodied navigation agents. CoRR (2018)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.387, + 0.785, + 0.428 + ], + "angle": 0, + "content": "4. Anderson, P., Wu, Q., Teney, D., Bruce, J., Johnson, M., Sünderhauf, N., Reid, I., Gould, S., Van Den Hengel, A.: Vision-and-language navigation: Interpreting visually-grounded navigation instructions in real environments. In: CVPR (2018)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.429, + 0.785, + 0.455 + ], + "angle": 0, + "content": "5. Chaplot, D.S., Gandhi, D., Gupta, A., Salakhutdinov, R.: Object goal navigation using goal-oriented semantic exploration. In: NeurIPS (2020)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.457, + 0.785, + 0.482 + ], + "angle": 0, + "content": "6. Chaplot, D.S., Gandhi, D., Gupta, S., Gupta, A., Salakhutdinov, R.: Learning to explore using active neural SLAM. In: ICLR (2020)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.484, + 0.785, + 0.51 + ], + "angle": 0, + "content": "7. Chen, S., Guhur, P.L., Schmid, C., Laptev, I.: History aware multimodal transformer for vision-and-language navigation. NeurIPS (2021)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.511, + 0.785, + 0.537 + ], + "angle": 0, + "content": "8. Chen, S., Guhur, P., Tapaswi, M., Schmid, C., Laptev, I.: Think global, act local: Dual-scale graph transformer for vision-and-language navigation. In: CVPR (2022)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.538, + 0.785, + 0.565 + ], + "angle": 0, + "content": "9. Chen, Y.C., Li, L., Yu, L., El Kholy, A., Ahmed, F., Gan, Z., Cheng, Y., Liu, J.: Uniter: Universal image-text representation learning. In: ECCV (2020)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.567, + 0.785, + 0.606 + ], + "angle": 0, + "content": "0. Conneau, A., Khandelwal, K., Goyal, N., Chaudhary, V., Wenzek, G., Guzmán, F., Grave, E., Ott, M., Zettlemoyer, L., Stoyanov, V.: Unsupervised cross-lingual representation learning at scale. In: ACL (2020)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.607, + 0.785, + 0.634 + ], + "angle": 0, + "content": "1. Devlin, J., Chang, M.W., Lee, K., Toutanova, K.: BERT: Pre-training of deep bidirectional transformers for language understanding. In: NAACL (2019)" + }, + { + "type": "ref_text", + "bbox": [ + 0.22, + 0.635, + 0.785, + 0.689 + ], + "angle": 0, + "content": "12. Dosovitskiy, A., Beyer, L., Kolesnikov, A., Weissenborn, D., Zhai, X., Unterthiner, T., Dehghani, M., Minderer, M., Heigold, G., Gelly, S., Uszkoreit, J., Houlsby, N.: An image is worth 16x16 words: Transformers for image recognition at scale. In: ICLR (2021)" + }, + { + "type": "ref_text", + "bbox": [ + 0.22, + 0.69, + 0.785, + 0.716 + ], + "angle": 0, + "content": "13. Dou, Z., Gao, F., Peng, N.: Masked path modeling for vision-and-language navigation. CoRR (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.22, + 0.717, + 0.785, + 0.758 + ], + "angle": 0, + "content": "14. Fried, D., Hu, R., Cirik, V., Rohrbach, A., Andreas, J., Morency, L.P., Berg-Kirkpatrick, T., Saenko, K., Klein, D., Darrell, T.: Speaker-follower models for vision-and-language navigation. In: NeurIPS (2018)" + }, + { + "type": "ref_text", + "bbox": [ + 0.22, + 0.759, + 0.785, + 0.785 + ], + "angle": 0, + "content": "15. Fuentes-Pacheco, J., Ascencio, J.R., Rendon-Mancha, J.M.: Visual simultaneous localization and mapping: a survey. Artif. Intell. Rev. (2015)" + }, + { + "type": "ref_text", + "bbox": [ + 0.22, + 0.786, + 0.785, + 0.813 + ], + "angle": 0, + "content": "16. Gao, C., Peng, X., Yan, M., Wang, H., Yang, L., Ren, H., Li, H., Liu, S.: Adaptive zone-aware hierarchical planner for vision-language navigation. In: CVPR (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.22, + 0.814, + 0.785, + 0.84 + ], + "angle": 0, + "content": "17. Hao, W., Li, C., Li, X., Carin, L., Gao, J.: Towards learning a generic agent for vision-and-language navigation via pre-training. In: CVPR (2020)" + }, + { + "type": "list", + "bbox": [ + 0.22, + 0.289, + 0.787, + 0.84 + ], + "angle": 0, + "content": null + } + ], + [ + { + "type": "page_number", + "bbox": [ + 0.218, + 0.116, + 0.236, + 0.127 + ], + "angle": 0, + "content": "16" + }, + { + "type": "header", + "bbox": [ + 0.272, + 0.115, + 0.331, + 0.127 + ], + "angle": 0, + "content": "Lu et al." + }, + { + "type": "ref_text", + "bbox": [ + 0.218, + 0.148, + 0.785, + 0.175 + ], + "angle": 0, + "content": "18. Hong, Y., Wu, Q., Qi, Y., Rodriguez-Opazo, C., Gould, S.: Vln bert: A recurrent vision-and-language bert for navigation. In: CVPR. pp. 1643-1653 (2021)" + }, + { + "type": "ref_text", + "bbox": [ + 0.218, + 0.177, + 0.785, + 0.217 + ], + "angle": 0, + "content": "19. Huo, J., Sun, Q., Jiang, B., Lin, H., Fu, Y.: Geovln: Learning geometry-enhanced visual representation with slot attention for vision-and-language navigation. In: CVPR (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.219, + 0.785, + 0.259 + ], + "angle": 0, + "content": "20. Hwang, M., Jeong, J., Kim, M., Oh, Y., Oh, S.: Meta-exlore: Exploratory hierarchical vision-and-language navigation using scene object spectrum grounding. In: CVPR (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.261, + 0.785, + 0.289 + ], + "angle": 0, + "content": "21. Ilharco, G., Jain, V., Ku, A., Ie, E., Baldridge, J.: General evaluation for instruction conditioned navigation using dynamic time warping. In: NeurIPS Workshop (2019)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.29, + 0.785, + 0.317 + ], + "angle": 0, + "content": "22. Kim, W., Son, B., Kim, I.: Vilt: Vision-and-language transformer without convolution or region supervision. In: ICML (2021)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.318, + 0.785, + 0.358 + ], + "angle": 0, + "content": "23. Ku, A., Anderson, P., Patel, R., Ie, E., Baldridge, J.: Room-across-room: Multilingual vision-and-language navigation with dense spatiotemporal grounding. In: EMNLP (2020)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.361, + 0.785, + 0.388 + ], + "angle": 0, + "content": "24. Li, J., Tan, H., Bansal, M.: Envedit: Environment editing for vision-and-language navigation. In: CVPR (2022)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.389, + 0.785, + 0.43 + ], + "angle": 0, + "content": "25. Li, J., Selvaraju, R., Gotmare, A., Joty, S., Xiong, C., Hoi, S.C.H.: Align before fuse: Vision and language representation learning with momentum distillation. NeurIPS (2021)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.431, + 0.785, + 0.472 + ], + "angle": 0, + "content": "26. Li, X., Li, C., Xia, Q., Bisk, Y., Celikylmaz, A., Gao, J., Smith, N.A., Choi, Y.: Robust navigation with language pretraining and stochastic sampling. In: EMNLP-IJCNLP (2019)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.474, + 0.785, + 0.514 + ], + "angle": 0, + "content": "27. Li, X., Li, C., Xia, Q., Bisk, Y., Celikyilmaz, A., Gao, J., Smith, N.A., Choi, Y.: Robust navigation with language pretraining and stochastic sampling. In: EMNLP-IJCNLP (2019)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.516, + 0.785, + 0.543 + ], + "angle": 0, + "content": "28. Lin, C., Jiang, Y., Cai, J., Qu, L., Haffari, G., Yuan, Z.: Multimodal transformer with variable-length memory for vision-and-language navigation. In: ECCV (2022)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.545, + 0.785, + 0.572 + ], + "angle": 0, + "content": "29. Liu, C., Zhu, F., Chang, X., Liang, X., Ge, Z., Shen, Y.: Vision-language navigation with random environmental mixup. In: ICCV (2021)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.573, + 0.785, + 0.586 + ], + "angle": 0, + "content": "30. Loshchilov, I., Hutter, F.: Decoupled weight decay regularization. In: ICLR (2019)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.588, + 0.785, + 0.615 + ], + "angle": 0, + "content": "31. Ma, C.Y., Lu, J., Wu, Z., AlRegib, G., Kira, Z., Socher, R., Xiong, C.: Self-monitoring navigation agent via auxiliary progress estimation. In: ICLR (2019)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.616, + 0.785, + 0.643 + ], + "angle": 0, + "content": "32. Ma, C., Wu, Z., AlRegib, G., Xiong, C., Kira, Z.: The regretful agent: Heuristic-aided navigation through progress estimation. In: CVPR (2019)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.645, + 0.785, + 0.685 + ], + "angle": 0, + "content": "33. Majumdar, A., Shrivastava, A., Lee, S., Anderson, P., Parikh, D., Batra, D.: Improving vision-and-language navigation with image-text pairs from the web. In: ECCV (2020)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.687, + 0.785, + 0.755 + ], + "angle": 0, + "content": "34. Oquab, M., Darcet, T., Moutakanni, T., Vo, H., Szafraniec, M., Khalidov, V., Fernandez, P., Haziza, D., Massa, F., El-Nouby, A., Assran, M., Ballas, N., Galuba, W., Howes, R., Huang, P., Li, S., Misra, I., Rabbat, M.G., Sharma, V., Synnaeve, G., Xu, H., Jegou, H., Mairal, J., Labatut, P., Joulin, A., Bojanowski, P.: Dinov2: Learning robust visual features without supervision. CoRR (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.757, + 0.785, + 0.797 + ], + "angle": 0, + "content": "35. Pope, R., Douglas, S., Chowdhery, A., Devlin, J., Bradbury, J., Levskaya, A., Heek, J., Xiao, K., Agrawal, S., Dean, J.: Efficiently scaling transformer inference. CoRR (2022)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.799, + 0.785, + 0.84 + ], + "angle": 0, + "content": "36. Qi, Y., Wu, Q., Anderson, P., Wang, X., Wang, W.Y., Shen, C., Hengel, A.v.d.: Reverie: Remote embodied visual referring expression in real indoor environments. In: CVPR (2020)" + }, + { + "type": "list", + "bbox": [ + 0.217, + 0.148, + 0.785, + 0.84 + ], + "angle": 0, + "content": null + } + ], + [ + { + "type": "header", + "bbox": [ + 0.687, + 0.115, + 0.732, + 0.127 + ], + "angle": 0, + "content": "PRET" + }, + { + "type": "page_number", + "bbox": [ + 0.769, + 0.116, + 0.786, + 0.127 + ], + "angle": 0, + "content": "17" + }, + { + "type": "ref_text", + "bbox": [ + 0.214, + 0.147, + 0.788, + 0.189 + ], + "angle": 0, + "content": "37. Radford, A., Kim, J.W., Hallacy, C., Ramesh, A., Goh, G., Agarwal, S., Sastry, G., Askell, A., Mishkin, P., Clark, J., et al.: Learning transferable visual models from natural language supervision. In: ICML (2021)" + }, + { + "type": "ref_text", + "bbox": [ + 0.214, + 0.19, + 0.787, + 0.218 + ], + "angle": 0, + "content": "38. Sennrich, R., Haddow, B., Birch, A.: Improving neural machine translation models with monolingual data. In: ACL (2016)" + }, + { + "type": "ref_text", + "bbox": [ + 0.216, + 0.218, + 0.786, + 0.245 + ], + "angle": 0, + "content": "39. Shen, S., Li, L.H., Tan, H., Bansal, M., Rohrbach, A., Chang, K., Yao, Z., Keutzer, K.: How much can CLIP benefit vision-and-language tasks? In: ICLR (2022)" + }, + { + "type": "ref_text", + "bbox": [ + 0.216, + 0.246, + 0.787, + 0.273 + ], + "angle": 0, + "content": "40. Tan, H., Yu, L., Bansal, M.: Learning to navigate unseen environments: Back translation with environmental dropout. In: NAACL (2019)" + }, + { + "type": "ref_text", + "bbox": [ + 0.216, + 0.273, + 0.787, + 0.3 + ], + "angle": 0, + "content": "41. Thomason, J., Murray, M., Cakmak, M., Zettlemoyer, L.: Vision-and-dialog navigation. In: CoRL (2020)" + }, + { + "type": "ref_text", + "bbox": [ + 0.216, + 0.301, + 0.787, + 0.328 + ], + "angle": 0, + "content": "42. Thrun, S.: Learning metric-topological maps for indoor mobile robot navigation. Artif. Intell. (1998)" + }, + { + "type": "ref_text", + "bbox": [ + 0.216, + 0.328, + 0.787, + 0.356 + ], + "angle": 0, + "content": "43. Vaswani, A., Shazeer, N., Parmar, N., Uszkoreit, J., Jones, L., Gomez, A.N., Kaiser, L.u., Polosukhin, I.: Attention is all you need. In: NeurIPS (2017)" + }, + { + "type": "ref_text", + "bbox": [ + 0.216, + 0.357, + 0.787, + 0.384 + ], + "angle": 0, + "content": "44. Wang, H., Wang, W., Liang, W., Xiong, C., Shen, J.: Structured scene memory for vision-language navigation. In: CVPR (2021)" + }, + { + "type": "ref_text", + "bbox": [ + 0.216, + 0.385, + 0.787, + 0.424 + ], + "angle": 0, + "content": "45. Wang, S., Montgomery, C., Orbay, J., Birodkar, V., Faust, A., Gur, I., Jaques, N., Waters, A., Baldridge, J., Anderson, P.: Less is more: Generating grounded navigation instructions from landmarks. In: CVPR (2022)" + }, + { + "type": "ref_text", + "bbox": [ + 0.216, + 0.425, + 0.787, + 0.467 + ], + "angle": 0, + "content": "46. Wang, X., Huang, Q., Celikyilmaz, A., Gao, J., Shen, D., Wang, Y.F., Wang, W.Y., Zhang, L.: Reinforced cross-modal matching and self-supervised imitation learning for vision-language navigation. In: CVPR (2019)" + }, + { + "type": "ref_text", + "bbox": [ + 0.216, + 0.468, + 0.787, + 0.495 + ], + "angle": 0, + "content": "47. Wang, Z., Li, X., Yang, J., Liu, Y., Jiang, S.: Gridmm: Grid memory map for vision-and-language navigation. In: ICCV (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.216, + 0.495, + 0.787, + 0.523 + ], + "angle": 0, + "content": "48. Zhu, F., Zhu, Y., Chang, X., Liang, X.: Vision-language navigation with self-supervised auxiliary reasoning tasks. In: CVPR (2020)" + }, + { + "type": "list", + "bbox": [ + 0.214, + 0.147, + 0.788, + 0.523 + ], + "angle": 0, + "content": null + } + ] +] \ No newline at end of file diff --git a/2024/PRET_ Planning with Directed Fidelity Trajectory for Vision and Language Navigation/4e64d425-40e4-4880-bd30-e7c4e06b2cbc_origin.pdf b/2024/PRET_ Planning with Directed Fidelity Trajectory for Vision and Language Navigation/4e64d425-40e4-4880-bd30-e7c4e06b2cbc_origin.pdf new file mode 100644 index 0000000000000000000000000000000000000000..1a27935ba9d02b7bd048764506168e1525e04a8b --- /dev/null +++ b/2024/PRET_ Planning with Directed Fidelity Trajectory for Vision and Language Navigation/4e64d425-40e4-4880-bd30-e7c4e06b2cbc_origin.pdf @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:53ac7b7080c1bb0f3c19ac24147623d5b7a09629f02b82ebf826fefbc4dbbf1c +size 1224470 diff --git a/2024/PRET_ Planning with Directed Fidelity Trajectory for Vision and Language Navigation/full.md b/2024/PRET_ Planning with Directed Fidelity Trajectory for Vision and Language Navigation/full.md new file mode 100644 index 0000000000000000000000000000000000000000..300c98697e91e9f25055248b147f0078e7054ff2 --- /dev/null +++ b/2024/PRET_ Planning with Directed Fidelity Trajectory for Vision and Language Navigation/full.md @@ -0,0 +1,310 @@ +# PRET: Planning with Directed Fidelity Trajectory for Vision and Language Navigation + +Renjie Lu1, Jingke Meng1(3), and Wei-Shi Zheng1,2,3 + +$^{1}$ School of Computer Science and Engineering, Sun Yat-sen University, Guangzhou, China + +lurj3@mail2.sysu.edu.cn,mengjke@gmail.com, wszheng@ieee.org + +2 Peng Cheng Laboratory, Shenzhen, China + +3 Key Laboratory of Machine Intelligence and Advanced Computing, Ministry of Education, Guangzhou, China + +Abstract. Vision and language navigation is a task that requires an agent to navigate according to a natural language instruction. Recent methods predict sub-goals on constructed topology map at each step to enable long-term action planning. However, they suffer from high computational cost when attempting to support such high-level predictions with GCN-like models. In this work, we propose an alternative method that facilitates navigation planning by considering the alignment between instructions and directed fidelity trajectories, which refers to a path from the initial node to the candidate locations on a directed graph without detours. This planning strategy leads to an efficient model while achieving strong performance. Specifically, we introduce a directed graph to illustrate the explored area of the environment, emphasizing directionality. Then, we firstly define the trajectory representation as a sequence of directed edge features, which are extracted from the panorama based on the corresponding orientation. Ultimately, we assess and compare the alignment between instruction and different trajectories during navigation to determine the next navigation target. Our method outperforms previous SOTA method BEVBert on RxR dataset and is comparable on R2R dataset while largely reducing the computational cost. Code is available: https://github.com/iSEE-Laboratory/VLN-PRET. + +Keywords: Vision-and-Language Navigation $\cdot$ Planning + +# 1 Introduction + +Enabling a robot to perform tasks on behalf of humans has been a longstanding objective in AI research. One such task is vision-and-language navigation (VLN) [4, 23, 36, 41], where an agent is required to navigate to a desired location by following natural language instructions provided by humans. For example, given the instruction "Go up stairs and stop at the top in front of a mirror.", the agent needs to follow the instruction and stop at an appropriate location. VLN has attracted numerous research interests [2, 7, 8, 14, 16-18, 20, 26, 31, 40, 44, 47]. + +![](images/7a7ffb983c5a1fa866447a08f99871bf31d168a8f065d88a44faab7fbe398a88.jpg) +Fig. 1: Comparison of SPL [3] and GFLOPs on R2R test unseen split dataset. Our method is comparable with previous SOTA methods while being more computational efficient. The computational cost of text encoder and visual encoder is omitted for fair comparison. + +Recent methods [2, 8, 16, 20, 44] demonstrate the effectiveness of introducing maps to enlarge the decision space to improve planning strategies. Instead of predicting low-level actions limited to short movements, these methods construct a graph to keep track of all visited and navigable locations observed so far, which enables high-level planning by expanding the action space to encompass the entire explored area. With the constructed topological maps, these methods predict actions in the global space via GCN-like models, where node features aggregate neighboring information. However, repeatedly calculating the entire map to predict actions at every step, even if the topo-map has only minor changes, is inefficient. In addition, formulating vision information of an environment in graph structure is too coarse-grained for accurate decision-making. Previous methods [2, 8] address this problem by introducing an additional branch to incorporate fine-grained information, such as local image features and bird's-eye-view features. But this strategy increases the model complexity and further escalates computational cost. + +In this work, we present an alternative way that supports the global decision space while achieving comparable performance. Our method, named Planning with DiReacted FidElity Trajectory(PRET), does not require calculating the entire graph at each step, nor does it rely on incorporating additional fine-grained information. The main idea is to determine the next location to navigate by evaluating the alignment between the instruction and the visual observations along different trajectories between the start point to all the unvisited nodes, as shown in Figure 2(b). Specifically, we maintain a directed fidelity trajectory-colored in red) for each unvisited node. The fidelity trajectory refers to a path from initial node to an unvisited node without detours. We assess the alignment between each trajectory and the instruction, and select the unvisited node with the highest alignment score to navigate. This planning strategy is efficient as we only need to compute for newly observed nodes. We estimate the instruction-trajectory + +![](images/5291bd7aeed0f2eafe974f85b8f83f4c9bd21f9a287954cbfa27807e2c1b4014.jpg) +Fig. 2: Illustration of our approaches. (a) shows our directed graph representation. Each edge is assigned with an orientation-aware panorama feature. (b) depict our planning method. We select an unvisited node-colored yellow to navigate towards next by choosing the fidelity path-colored red that best aligned with instruction. + +alignment with transformer and further reduce the computational cost by utilizing the KV-cache [35] technique. It also enables us to make decisions by taking into full consideration of the instruction-trajectory alignment. + +In order to enhance the alignment between instructions and navigation trajectories, we propose to incorporate directionality in path representation. Due to the inherent directional nature of the navigation process, it is crucial that the representation of trajectories in opposite directions is asymmetric, as visual observations are linked to observed orientation. In this work, we introduce the directed graph to depict the explored area of the environment and firstly define the trajectory representation as a sequence of directed edge features. These directed edge features are derived from the panorama based on the corresponding orientation (as depicted in Fig. 2(a)). By incorporating directionality in the graph edge features, we eliminate irrelevant panorama information and obtain a more accurate representation of the directed trajectory. Our method overcome existing methods [2,8,16] that commonly adopt the undirected graph to store panorama features on graph nodes, resulting in direction-invariant path representations, which fails to capture the important distinction between paths that traverse the same node from different directions. For instance, consider a scenario where agents pass through the same node from different directions. Despite the fundamentally different spatial contexts of these paths, the node representation remains unchanged. + +We conduct experiments on R2R, RxR datasets to evaluate the efficacy of our proposed methods. The results demonstrate that PRET achieves strong performance while being more efficient than previous methods. Specifically, PRET achieves comparable performance compared to previous state-of-the-art method BEVBert [2] on R2R dataset with only $3\%$ computational cost as shown in Fig. 1. On the RxR dataset, PRET outperforms previous methods and achieves the new state-of-the-art performance. Qualitative visualization also shows that our simplified model is able to learn complex backtracking strategies. + +# 2 Related Work + +Vision-and-Language Navigation(VLN). Early VLN methods use sequence-to-sequence LSTMs with attention mechanism to encode language features and predict local actions. SpeakerFollower [14] first introduces back translation [38] technique to ease the data scarcity problem in VLN. EnvDrop [40] proposes to augment environment by using dropout on feature space to avoid overfitting on visual input. Following this work, many environment augmentation approaches [24, 29] is proposed. Some auxiliary tasks is proposed by [31, 48] for better guidance. More recently, transformer is explored how to be adopted in VLN. PRESS [27] uses BERT [11] as the text encoder. PREVALENT [17] first propose to pretrain the transformer in VLN dataset, but still use it as encoder. More recent works [7, 18, 28] explore how to memorize navigation history and use transformer to learn strong planning strategy. + +Navigation Strategy in VLN. Navigation strategy plays a key role in Vision- and-Language Navigation, as VLN requires agents to navigate in unseen environments and thus needs to explore and familiarize the environment. Works like [32] investigate designing regretful agents to enable explicit exploration. Reinforcement learning methods [40, 46] have also been explored to enhance navigation strategies. SSM [44] constructs directed graphs to represent explored areas, with vision features on nodes and orientation on edges. DUET [8] employs a dual-scale transformer to make local and global predictions. AZHP [16] constructs hierarchy graphs to facilitate exploration. MetaExplore [20] explicitly predicts whether to backtrack using a separate module. + +Maps For Navigation. Navigation research has a long history of using SLAM [15] to construct maps [5, 6, 42] for planning. In the VLN literature, several methods [8, 44] adopt topological graphs for global planning. However, the visual representation on topological maps is too coarse-grained for decision-making. To address this limitation, [2, 47] introduce metric maps in VLN. While grid-based metric maps can precisely represent spatial layouts, they result in high computational costs. To balance representation ability and computational cost, BEVBert [2] utilizes learnable hybrid topological and metric maps. However, the computational cost remains high. In our work, we propose a novel approach that involves constructing a directed topo-map and planning with trajectory. Our method is more efficient as we perform planning with trajectory instead of directly relying on GCN-like models to encode the entire map for planning. This also allows us to incrementally calculate embeddings for new trajectories and further reduces computational cost. + +# 3 Method + +# 3.1 Problem Formulation + +In VLN with discrete environments [4,23,36,41], an environment is an undirected graph $\mathcal{G} = \{\mathcal{V},\mathcal{E}\}$ , where $\mathcal{V} = \{V_i\}_{i=1}^N$ represents $N$ navigable nodes and $\mathcal{E}$ denotes navigable edges. At the beginning of navigation, an agent is initialized + +at a starting node and given a natural language instruction $W = \{w_{i}\}_{i=1}^{L}$ with $L$ words. The agent is required to interpret this instruction to navigate to the target location. + +At step $t$ on node $V_{t}$ , the agent observes (1) a panorama $\mathcal{R}_t = \{r_{t,i}\}_{i=1}^K$ represented by $K$ images from different views, $r_{t,i}$ is the extracted image feature of the $i$ th view; (2) neighboring navigable nodes $\mathcal{N}(V_t)$ , where $\mathcal{N}(V_t) \subset \mathcal{V}$ ; (3) orientations and coordinates of these neighboring nodes. The agent should choose a neighboring node to step to or stop at current location. Navigation is considered successful if the agent stops within 3 meters of the target. + +We treat navigation as a process of searching temporary target to navigate towards next among unvisited nodes on $\mathcal{G}_t$ . A temporary target can be either a local neighboring node or remote unvisited node. By adding a virtual stop node connected to all nodes on the graph, we can also model the stop action. We refer fidelity trajectory to a path from initial node to an unvisited node without detours. + +# 3.2 Model Overview + +The overall framework is shown in Fig. 3(a). Our planning method consists of three components: (1) We construct a directed graph for explored area. Considering the directional nature of navigation, we propose an Orientation-aware Panorama Encoder(OPE) to extract orientation-aware vision features for edges. Edge sequence is used to represent directed trajectory and align with the instruction. (2) We maintain a path (i.e. directed fidelity trajectory) to each unvisited node and assess the alignment between each path and the instruction at each step. Only few neighboring nodes are added at each step, we only compute for these relevant new paths. Since each node corresponds to a path, we calculate a path embedding with a matching assessment module(MAM) that encodes the instruction-path alignment and stores the embedding on the node. (3) We compare the path embeddings of all unvisited nodes(candidates) with candidate comparison module(CCM)(shown in Fig. 3(c)) to determine which aligns best with the instruction. We select the best aligned candidate node as the temporary target. Then the agent can navigate to the temporary target along the shortest path on constructed graph. Since the path embedding is an estimate of the alignment between instruction and a path to the node, CCM ensures that the agent navigates by fully considering the instruction-trajectory alignment. + +# 3.3 Orientation-Aware Directed Graph Construction + +We construct a directed graph to support path representation, incrementally adding nodes and edges as exploration proceeds. Previous methods [2,8,44] represent visited nodes with panorama features. However, panoramas of unvisited nodes are inaccessible as the agent has not visited them. So they represent these nodes with views towards them, leading to inconsistent representation between visited and unvisited nodes. In contrast, we extract visual features for the directed edges on the graph. These edge features represent the visual observations + +![](images/2909c5b6168339dd7a93e5197057553ab89b003042ea5ba060d8584919acd8b3.jpg) + +Fig. 3: Illustration of our model. (a) is the overall framework of our method. At each step, we update the graph, extract path embeddings, and predict actions. (b) depicts the matching assessment module(MAM). Each token is an edge feature. We compute path embeddings for each newly observed nodes with cross-modal transformer and impose a causal mask to reduce computational cost. (c) shows the candidate comparison module(CCM). We gather path embeddings of unvisited nodes and forward them into a single layer transformer followed by a MLP to predict temporary target. +![](images/0e14ba6e264285d2340f19a06a9ea1d5e964bc88b1099d91ff4d1487bc904edc.jpg) +visited node current node unvisited node start token shared edge token unvisited neighboring edge token stop token + +![](images/a9f78e2971a855100fe31657721cd55dc149d31d8a54902342afb2ab2d6ed1d3.jpg) + +when facing specific directions from each node. As edges rely only on observable views rather than node panoramas, they enable consistent representation for both visited and unvisited nodes. + +Let $\mathcal{G}_t = \{\mathcal{V}_t, \mathcal{E}_t\}$ be the directed graph at step $t$ . Neighboring nodes $\mathcal{N}(V_t)$ are observed, and some nodes in $\mathcal{N}(V_t)$ have already existed in $\mathcal{G}_t$ while others are newly observed and added to $\mathcal{G}_t$ . Besides, directed edges from $V_t$ to nodes in $\mathcal{N}(V_t)$ together with features extracted by orientation-aware panorama encoder(OPE) will also be added in $G_t$ , as shown in Fig. 3(a). + +Visual Representation For Edges. We extract orientation-aware panorama feature by using relative orientation as query to attend panorama features in cross-attention manner, as shown in the OPE module in Fig. 3(a). Assume that $(\phi, \theta)$ represents the relative heading and elevation of a specific edge when compared to the agent's current orientation. To extract feature for the edge, we first encode $(\phi, \theta)$ as follows: + +$$ +x ^ {a} = \left[ \sin (\phi), \cos (\phi), \sin (\theta), \cos (\theta) \right] W ^ {a}, \tag {1} +$$ + +where $W^{a}\in \mathbb{R}^{4\times d}$ are learnable weights and $d$ is the dimension of feature. Then we denote $X_{t}^{a} = \{x_{t,i}^{a}\}_{i = 1}^{\lfloor \mathcal{N}(V_{t})\rfloor}$ as orientation features of adjacent edges of $V_{t}$ . + +We then encode the panorama feature. As the panorama $\mathcal{R}_t = \{r_{t,i}\}_{i=1}^K$ is represented by $K$ images from different views and each view corresponds to a fixed orientation, we can also get the relative heading and elevation of each view. To encode the panorama, we first concatenate each view feature $r_{t,i}$ with corresponding relative orientation $(\phi_{t,i},\theta_{t,i})$ and then map it to dimension $d$ via a linear projection: + +$$ +x _ {t, i} ^ {p} = \left[ r _ {t, i}; \sin \left(\phi_ {t, i}\right), \cos \left(\phi_ {t, i}\right), \sin \left(\theta_ {t, i}\right), \cos \left(\theta_ {t, i}\right) \right] W ^ {p}. \tag {2} +$$ + +Then we denote $X_{t}^{p} = \{x_{t,i}^{p}\}_{i=1}^{K}$ as the panoramic view representation. By considering views as patches of the panorama and orientations as position embedding, $X_{t}^{p}$ is like the encoded ViT [12] input. + +We adopt transformer decoder [43] to extract orientation-aware panorama feature: + +$$ +E _ {t} = \operatorname {T r a n s f o r m e r D e c o d e r} \left(X _ {t} ^ {a}, X _ {t} ^ {p}\right), \tag {3} +$$ + +where $E_{t} = \{e_{t,i}\}_{i = 1}^{\lfloor \mathcal{N}(V_{t})\rfloor}$ are extracted edge features. By taking $X_{t}^{a}$ as the query input and $X_{t}^{p}$ as the key-value input, panorama is attended by orientation in the first layer. In the subsequent layers, both orientation and vision information are used to query relevant views in the panorama. Therefore, each directed edge only focuses on a specific region of the panorama, thus orientation-aware panorama feature is extracted. + +# 3.4 Planning with Fidelity Trajectory + +We maintain a directed fidelity trajectory with a stack during navigation for each unvisited node on $\mathcal{G}_t$ . When a new node is added, it is pushed into the stack. When the agent needs to backtrack, the top node is popped off the stack until a different node than the top is reached. In this way, detours are removed by popping nodes off the stack. With the maintained trajectory of each node, we propose MAM to assess its alignment with instruction. We then determine an unvisited node on $\mathcal{G}_t$ to navigate next by comparing this alignment with CCM. + +Matching Assessment Module(MAM). MAM is a multi-layer transformer decoder that extracts a path embedding for each path, which estimates how much the trajectory match with the instruction. These path embeddings are then stored on the corresponding nodes in graph $\mathcal{G}_t$ . + +For a single path with length $l$ , we represent the path with the edge features along it, denoted as $X^{h} = [e_{1}, e_{2}, \dots, e_{l}]$ . $\{e_{i}\}_{i=1}^{l}$ are the edge features on the path. Then we add it with position embedding and forward it together with text embedding $X^{w}$ into MAM like follows: + +$$ +X ^ {h \prime} = X ^ {h} + P _ {l}, +$$ + +$$ +X ^ {o} = \operatorname {T r a n s f o r m e r D e c o d e r} \left(X ^ {h t}, X ^ {w}\right), \tag {4} +$$ + +where $P_{l}$ is a matrix including $l$ position embeddings [43] and $X^{w}$ is the text embedding extracted by a text encoder, which is a multi-layer transformer [43]. + +By utilizing two sequences from two modalities as inputs, the transformer is capable of effectively estimating their alignment, as extensively demonstrated in Vision and Language Pretraining studies [9, 22, 25]. The last token of the encoded sequence $X^o$ provides the path embedding, which is an estimation of the text-path alignment. + +To support STOP action, we add a stop node that connected to all nodes in $\mathcal{G}_t$ . Since we represent a path using edge features, we assign a learnable stop token, initialized with a zero vector, to all edges that point to the stop node. The path embedding corresponds to the stop node is stored on current node. So the agent decides to stop at current location when it finds that the path up to here appending with stop token matches instruction the best. Here stop token serves as the End-of-Sentence token in text generation. For each path, we also add a learnable start token that serves as the Start-of-Sentence token. + +At step $t$ , considering the stop node, assume there are $N$ neighboring unvisited nodes to be updated. Computing path embeddings separately for these nodes requires multiple forward passes, which is inefficient. Actually, we can compute these embeddings in a single forward pass, as shown in Fig. 3(b). Note that these paths share the same prefix from the initial node to the current node. We can avoid duplicated computation by imposing a causal mask in the self-attention block. The causal mask is a lower triangular matrix that only allows a token to attend to its previous tokens. We merge these causal masks together and forward all relevant tokens into the transformer block. The merged mask still satisfies the constraint that each token can only attend to its previous tokens. We further reduce the computational cost by adopting the KV-cache technique [35], which stores the shared prefix tokens to avoid unnecessary re-calculation. In this way, we can equivalently compute the path embeddings for multiple newly observed nodes in a single forward pass. The last several output tokens colored yellow and green are path embeddings and used to update the graph, i.e., stored on corresponding graph nodes. + +Candidate Comparison Module(CCM). CCM compares the path embeddings from candidate nodes to predict a temporary target. As each path embedding encodes the alignment between a trajectory and the instruction, the CCM actually selects a path that is most aligned with the instruction. In this way, we make navigation planning fully according to the instruction-trajectory alignment. As shown in Fig. 3(c), we collect path embeddings $X_{t}^{e}$ and input them into CCM. CCM consists of a single-layer transformer encoder [43] and a MLP. The transformer encoder is responsible for the comparison of the candidates and the MLP maps the encoded path embedding to a 1D score: + +$$ +X _ {t} ^ {e \prime} = \operatorname {T r a n s f o r m e r L a y e r} \left(X _ {t} ^ {e}\right), +$$ + +$$ +s _ {t} = \operatorname {M L P} \left(X _ {t} ^ {e ^ {\prime}}\right), \tag {5} +$$ + +$$ +p _ {t} = \operatorname {s o f t m a x} \left(s _ {t}\right). +$$ + +Score $s_t$ reflects the relative text-path alignment after comparison. An alternative way is removing the transformer and directly compute the alignment score $s_t$ . + +However, this method assesses each path independently without any comparison, which can make the decision-making process more challenging and potentially result in a decrease in performance(demonstrated in Sec. 4). It is then normalized with softmax and we select the highest score node as the temporary target. The agent will navigate to it along the shortest path on $\mathcal{G}_t$ . If the stop node is selected, the agent believes the instruction is completed and stops the navigation. + +# 3.5 Pretraining and Fine-tuning + +Pretraining. Pretraining is proved to be helpful in previous works [2,7,17,33]. PREVALENT [17] synthesized a substantial amount of data for pretraining, and we utilized the same data as their work. We use Masked Language Modeling(MLM) [11] for pretraining. Concretely, we randomly replace $15\%$ of the input tokens with a special [MASK] token and forward these masked tokens into the text encoder. We also extract a sequence of edge features to represent the path. Then the encoded text tokens and path representation are input into an additional transformer decoder. The corresponding outputs are used to predict the masked tokens. MLM helps the model learn aligned text and path representations. In VLN, a key difference from other vision-and-language tasks is that we align a text sequence with a sequence of vision representations rather than a single image. + +Fine-tuning. We train the agent with a mixture of teacher-forcing and student-forcing as previous methods [2,8]. In the teacher-forcing stage, the agent navigate according the ground truth actions $a_{t}^{*}$ and the loss is calculated as follows: + +$$ +\mathcal {L} _ {T F} = \frac {1}{T _ {T F}} \sum_ {t = 1} ^ {T _ {T F}} \mathrm {C E} \left(p _ {t}, a _ {t} ^ {*}\right), \tag {6} +$$ + +CE is the cross entropy loss and $T_{TF}$ is the number of navigate steps. In the student-forcing stage, we sample an action from the distribution predicted by the agent so that the agent can explore the environment and reduce the exposure bias. We train the agent with heuristic pseudo label $a_t^{pseudo}$ : + +$$ +\mathcal {L} _ {S F} = \frac {1}{T _ {S F}} \sum_ {t = 1} ^ {T _ {S F}} \mathrm {C E} \left(p _ {t}, a _ {t} ^ {\text {p s e u d o}}\right). \tag {7} +$$ + +When the agent deviate from ground truth path, we take the nearest node on ground truth path as the pseudo label to encourage agent to learn backtracking strategy. If there is no unvisited node on ground truth path, we take the nearest node on the shortest path from current node to target node as the pseudo label. The agent navigate two times to compute $\mathcal{L}_{TF}$ and $\mathcal{L}_{SF}$ . The total loss is the weighted sum of them: + +$$ +\mathcal {L} = \lambda \mathcal {L} _ {T F} + (1 - \lambda) \mathcal {L} _ {S F}, \tag {8} +$$ + +where $\lambda \in (0,1)$ is the weight. + +# 4 Experiments + +# 4.1 Datasets and Evaluation Metrics + +R2R. Room-to-Room (R2R) dataset [4] contains 7,189 trajectories, with 3 instructions per trajectory, across 90 scenes. These scenes are divided into train, val unseen and test unseen splits with 61, 11 and 18 scenes respectively. A val seen split with the same scenes as the train split is also provided. All paths in R2R are the shortest paths between the start and target nodes. + +RxR. Room-across-Room(RxR) [23] is a large multilingual dataset. The dataset has 126,000 instructions total, with 42,000 instructions in each of the three languages: English, Hindi, and Telugu. The paths in RxR are longer than those in R2R and are not the shortest possible routes. Additionally, the instructions in RxR are longer and contain more detailed descriptions compared to R2R. + +Evaluation Metrics. We adopt the following evaluation metrics on R2R: (1) Trajectory length (TL): the average length of the agent's path in meters. (2) Navigation error (NE): the average distance between the agent's final position and the target location. (3) Success rate (SR): the ratio of successful navigations, where $NE < 3m$ is considered successful. (4) SR penalized by path length (SPL) [3]: TL longer than the shortest path is penalized. As paths in RxR dataset are shortest paths, TL and SPL are unsuitable. Instead, on RxR we use nDTW and sDTW [21] to measure trajectory similarity between the agent and ground truth. Please note that the fidelity trajectory that is removed detours is solely used for planning. It is not the agent's actual trajectory used for evaluation. + +# 4.2 Implementation Details + +Module architecture. Text backbone is a 6 layer transformer encoder initialized with pretrained ALBEF [25]. For the multilingual RxR dataset, we use 12 layer mRoberta [10] instead. The layer numbers of the OPE, MAM, and CCM are 2, 4, and 1 respectively. All the hidden layer size is 768. Image features is extracted by DINOv2 [34]. We also report results with CLIP-ViT-base [37] feature for fair comparison. + +Training details. On R2R dataset, we first pretrain PRET with learning rate 2e-5 and batch size 16 on a single 24G 4090 GPU for 100,000 iterations( $\sim$ 5 hours). Then we fine-tune the model with learning rate 1e-5 and batch size 8 on 4090 for 100,000 iterations( $\sim$ 25 hours). Optimizer is AdamW [30]. $\lambda$ is set to 0.2. Augmented dataset [17] is used in both pretraining and fine-tuning. The best result is selected by SPL on val unseen split. On RxR dataset, we also follow the pretrain and fine-tune paradigm. Due to the longer instruction and trajectory, we use a smaller batch size 4. Marky [45] is used as augmented data. The best result is selected by sDTW on val unseen split. + +Table 1: Comparison with other methods on R2R dataset. SPL is considered as the primary evaluation metric. + +
MethodsVal SeenVal UnseenTest Unseen
TLNE↓SR↑SPL↑TLNE↓SR↑SPL↑TLNE↓SR↑SPL↑
Seq2Seq-SF [4]11.336.0139-8.397.8122-8.137.852818
Speaker-Follower [14]-3.3666--6.6235-14.826.623528
RCM [46]10.653.5367-11.466.0943-11.976.124338
Regretful [32]-3.236963-5.325041-5.695640
EnvDrop [40]11.003.99625910.705.22524811.665.235147
PREVALENT [17]10.323.67696510.194.71585310.515.305451
NvEM [1]11.093.44696511.834.27605512.984.375854
SSM [44]14.703.10716220.704.32624520.404.576146
RecBert [18]11.132.90726812.013.93635712.354.096357
HAMT [7]11.152.51767211.462.29666112.273.936560
MTVM [28]-2.677469-3.736659-3.856559
DUET [8]12.322.28797313.943.31726014.733.656959
AZHP [16]----14.053.15726114.953.527160
Meta-Explore [20]11.952.11817513.093.22726214.253.577161
GridMM [47]----13.272.83756414.433.357362
BEVBert [2]13.562.17817414.552.81756415.873.137362
Ours(CLIP)11.482.60746912.213.12716313.873.127262
Ours(DINOv2)11.252.41787211.872.90746512.213.097264
+ +# 4.3 Results + +Comparison on R2R. Tab. 1 compares our approach with previous methods. On the val unseen and test unseen split, our method achieves comparable performance with previous SOTA method BEVBert on the primary metric SPL, while our method significantly reduces the computational cost. Besides, BEVBert [2] adopts additional depth information to construct bird's-eye view input. On the val seen split, our method performs sub-optimal as PRET does not take the environment layout as input, which means it less tends to overfit on the seen environment. Also, in the VLN literature, all visual features are pre-extracted and remains unchanged, the fixed feature may not contains appropriate feature for navigation. Therefore, we also investigate different vision features for VLN, as shown in Tab. 1. We find that DINOv2 [34] feature is better than CLIP feature. DINOv2 is trained on large-scale curated data in self-supervised manner, which leads to more general purpose feature. Moreover, DINOv2 feature contains geometric information as it can be used to predict depth map according to [34]. Geometric information is shown helpful in [19]. + +Tab. 2 shows the parameter count, computational cost, and latency of different graph-based methods. We present GFLOPs of different navigation steps. Our method contains fewer parameters as we use a single stream model unlike [2,8] use dual stream transformer. Besides, our decoder computational cost is only $19\%$ of DUET and $3\%$ of BEVBert at step 10. These suggests that our method is significantly more efficient than previous SOTA graph-based methods. As shown in Tab. 2, this results in lower training and inference latency. + +Table 2: Comparison of latency and computational cost on R2R val unseen. * indicates we reproduce DUET with CLIP feature for fair comparison. When computing GFLOPs, we simplify the calculation by assuming that each vertex of the graph is connected to 4 adjacent vertices (average degree in R2R), and at each step, 3 new vertices are encountered. The computational cost of text encoder and visual encoder is omitted for fair comparison. + +
MethodsR2R val unseenLatency(ms)GFLOPsParams
TLNE↓SR↑SPL↑traininference1 step10 steps20 steps
DUET* [8]14.03.271.661.111.88.62.44.87.590M
BEVBert [2]14.62.874.963.616.210.633.735.537.590M
Ours12.23.171.062.77.74.10.60.91.264M
+ +Table 3: Comparison with previous methods on RxR dataset. †indicates approaches that utilize additional augmented data. nDTW represents the normalized DTW distance between the agent's trajectory and ground truth path. sDTW is the nDTW weighted by success rate. + +
MethodsVal SeenVal Unseen
NE↓SR↑nDTW↑sDTW↑NE↓SR↑nDTW↑sDTW↑
LSTM [23]10.725.242.220.710.922.838.918.2
EnvDrop+ [39]-----43.655.7-
HAMT [7]-59.465.350.9-56.563.148.3
EnvEdit [24]-67.271.158.5-62.868.554.6
MPM [13]-67.771.058.9-63.567.754.5
MARVEL [45]†3.075.979.168.84.564.870.857.5
BEVBert [2]†3.275.076.366.74.068.569.658.6
Ours(CLIP)†2.677.078.067.53.371.271.860.4
Ours(DINOv2)†2.479.380.470.73.272.873.462.4
+ +Comparison on RxR. Tab. 3 reports the results on the RxR dataset. RxR is challenging as instructions and paths in RxR are longer than R2R. While the alignment between path and detailed path description is what our method skills at, our method outperforms previous methods among all metrics at a lower computational cost. PRET achieves a $1.8\%$ improvement on the main metrics sDTW on the val unseen split, demonstrating that our model better understands and follows instructions. This result highlights the advantage of our planning with trajectory strategy, which considers the alignment between instructions and paths and enables efficient decision-making in the global space. Additionally, with DINOv2 features, the result is further improved. + +# 4.4 Ablation Study + +We conducted ablation experiments on various components of PRET, including the directed path representation and module ablation. Results are reported on the R2R val unseen split. + +Table 4: Comparison of undirected and directed path representation. + +
MethodsTLNE↓SR↑SPL↑
undirected15.623.5968.2856.77
directed11.872.9073.7865.16
+ +Table 5: Ablation study on modules. + +
MethodsTLNE↓SR↑SPL↑
1 MAM12.043.9962.3254.48
2 MAM+CCM12.153.5465.9457.32
3 MAM+OPE12.183.1571.6063.07
4 MAM+OPE+CCM11.872.9073.7865.16
+ +![](images/73164e5bbdd0ccea64413afcdab955178397f303e85f580b1c421d823f92f2ff.jpg) +Fig. 4: Comparison of orientation panoramic view and single candidate view. + +Directionality in Path Representation. We compare whether to incorporate directionality for path representations in Tab. 4. Undirected path is represented by a sequence of node features. The node features is extracted by forwarding panoramic views into a 2-layer transformer encoder and averages the output tokens. Directed path is represented by a sequence of orientation-aware panorama features. According to the results, directed path representation outperforms undirected node features among all metrics. Specifically, it gains $8.39\%$ improvement on SPL and $5.5\%$ on SR. This indicates that edge feature is more suitable to represent the directional navigation process and align with the instruction. Node features does not distinguish different path directions and provide redundant information in a path, which hinders the alignment. + +Module ablation. Tab. 5 shows ablation results on modules. Row 4 is the full model introduced in Sec. 3. OPE extracts orientation-aware features for edges, MAM extracts path embeddings while CCM compares them to select the temporary target. When removing CCM (Row 3), we only predict a single alignment score rather than a path embedding for each path without comparison. The score is then normalized and used to predict the temporary goal. Performance drops $2.09\%$ in this case, as selecting a path based solely on a single score, without comparing different paths, is more difficult. + +In row 2, we remove OPE, panorama features are not used and only a single 60-degree view towards each neighboring node is adopted as edge features. Performance drops $7.84\%$ as the limited field-of-view cannot fully represent the path. The trajectory represented by these views is discrete and does not form a continuous change of views. In some cases, a single view towards candidates provides no information for navigation, as shown in Fig. 4. When go up the stairs, the upper view only sees the wall and does not help the navigation at all. While using the panorama, the agent can dynamically see a larger region. These results show our graph representation with panorama encoder provides rich representation for trajectories. Row 1 demonstrates the performance when using only MAM. The performance is poor without other model components, demonstrating the effectiveness of these modules. + +# 4.5 Visualization + +![](images/1661bb8d1b75a64d5109f9fae4bdfd50c34a1793afbf6832f1f8d3d623e5c5d0.jpg) +Instruction: Face directly away from stairs and walk through doorway straight ahead, right immediately, walk past bed and turn right into closet, walk straight and stop in front of toilet. +(a) + +![](images/5d8baa7f63037a6f0be0e12bc335b6b7ba26e0b1574f0545206d7fd0feae89bc.jpg) +undirected trajectory + +![](images/3c9519642b448d44cbce4747794e4d79c64fbb0c3b15bd002592649ddb5f5a6a.jpg) + +![](images/83c08a5184cabee19094350062c18cae036b3b49ddf035843b4986b7a15d641e.jpg) +Instruction: Turn right and go up the stairs. + +![](images/659a53cccecbd5e255c4cd2007a9b649f08041e684f66e66f9d4d6c16c76aba8.jpg) +directed trajectory + +![](images/9b7b767cb0e77eb4644e3e0a2791e9f95abb6ce5e8112d187235c43759854aaf.jpg) + +![](images/4574cc6c8740af637fc91b0614b2db35f5bf94ae75c5561d9135c162cf93a917.jpg) +(b) +Fig. 5: (a) Visualization of the agent's navigation process, showcasing its ability to learn a backtracking strategy. (b) Visualizing attention weights in OPE to illustrate the distinction between undirected and directed trajectory representations. + +Fig. 5(a) shows a qualitative example of our model's behavior. Initially placed at the green starting node, the agent wrongly decides to explore the upper left area(blue path). It takes many unnecessary detours there, but through comparing the trajectory alignment with the instructions, realizes these paths do not match the text well. Therefore, it backtracks to the start and correctly navigates to the destination(red path). This demonstrates the capability of PRET to learn complex backtracking behaviors. + +Fig. 5(b) shows the difference between undirected and directed trajectory representations. The former represents a path using panorama features stored on nodes, which eliminates directional discrepancies since the same node provides the same feature regardless of orientation, and contains redundant information compared to "stairs" in the instruction. In contrast, our proposed approach represents a path using orientation-aware panorama features on edges. This eliminates redundant information and focuses on the "stairs" better captures the directional navigation process. + +# 5 Conclusion + +In this work, we present an efficient method that planning with directed fidelity trajectory, explicitly leveraging the alignment between instructions and trajectories for navigation. Additionally, we consider the directional nature of navigation and introduce a directed graph construction during navigation, storing vision information on directed edges. This approach provides rich directed path representation and enhances instruction-trajectory alignment. Experiments demonstrate that our method is achieves strong performance while being significantly more efficient than previous methods. Our method has certain limitations that we need to address. For example, navigation requires environment layout information in some cases. We recognize the need to investigate potential solutions to overcome these challenges. + +# Acknowledgments + +This work was supported partially by the National Key Research and Development Program of China (2023YFA1008503), NSFC(U21A20471, 62206315), Guangdong NSF Project (No. 2023B1515040025, 2020B1515120085, 2024A1515-010101), Guangzhou Basic and Applied Basic Research Scheme(2024A04J4067). + +# References + +1. An, D., Qi, Y., Huang, Y., Wu, Q., Wang, L., Tan, T.: Neighbor-view enhanced model for vision and language navigation. In: ACM MM (2021) +2. An, D., Qi, Y., Li, Y., Huang, Y., Wang, L., Tan, T., Shao, J.: Bevbert: Multimodal map pre-training for language-guided navigation. In: ICCV (2023) +3. Anderson, P., Chang, A.X., Chaplot, D.S., Dosovitskiy, A., Gupta, S., Koltun, V., Kosecka, J., Malik, J., Mottaghi, R., Savva, M., Zamir, A.R.: On evaluation of embodied navigation agents. CoRR (2018) +4. Anderson, P., Wu, Q., Teney, D., Bruce, J., Johnson, M., Sünderhauf, N., Reid, I., Gould, S., Van Den Hengel, A.: Vision-and-language navigation: Interpreting visually-grounded navigation instructions in real environments. In: CVPR (2018) +5. Chaplot, D.S., Gandhi, D., Gupta, A., Salakhutdinov, R.: Object goal navigation using goal-oriented semantic exploration. In: NeurIPS (2020) +6. Chaplot, D.S., Gandhi, D., Gupta, S., Gupta, A., Salakhutdinov, R.: Learning to explore using active neural SLAM. In: ICLR (2020) +7. Chen, S., Guhur, P.L., Schmid, C., Laptev, I.: History aware multimodal transformer for vision-and-language navigation. NeurIPS (2021) +8. Chen, S., Guhur, P., Tapaswi, M., Schmid, C., Laptev, I.: Think global, act local: Dual-scale graph transformer for vision-and-language navigation. In: CVPR (2022) +9. Chen, Y.C., Li, L., Yu, L., El Kholy, A., Ahmed, F., Gan, Z., Cheng, Y., Liu, J.: Uniter: Universal image-text representation learning. In: ECCV (2020) +0. Conneau, A., Khandelwal, K., Goyal, N., Chaudhary, V., Wenzek, G., Guzmán, F., Grave, E., Ott, M., Zettlemoyer, L., Stoyanov, V.: Unsupervised cross-lingual representation learning at scale. In: ACL (2020) +1. Devlin, J., Chang, M.W., Lee, K., Toutanova, K.: BERT: Pre-training of deep bidirectional transformers for language understanding. In: NAACL (2019) +12. Dosovitskiy, A., Beyer, L., Kolesnikov, A., Weissenborn, D., Zhai, X., Unterthiner, T., Dehghani, M., Minderer, M., Heigold, G., Gelly, S., Uszkoreit, J., Houlsby, N.: An image is worth 16x16 words: Transformers for image recognition at scale. In: ICLR (2021) +13. Dou, Z., Gao, F., Peng, N.: Masked path modeling for vision-and-language navigation. CoRR (2023) +14. Fried, D., Hu, R., Cirik, V., Rohrbach, A., Andreas, J., Morency, L.P., Berg-Kirkpatrick, T., Saenko, K., Klein, D., Darrell, T.: Speaker-follower models for vision-and-language navigation. In: NeurIPS (2018) +15. Fuentes-Pacheco, J., Ascencio, J.R., Rendon-Mancha, J.M.: Visual simultaneous localization and mapping: a survey. Artif. Intell. Rev. (2015) +16. Gao, C., Peng, X., Yan, M., Wang, H., Yang, L., Ren, H., Li, H., Liu, S.: Adaptive zone-aware hierarchical planner for vision-language navigation. In: CVPR (2023) +17. Hao, W., Li, C., Li, X., Carin, L., Gao, J.: Towards learning a generic agent for vision-and-language navigation via pre-training. In: CVPR (2020) + +18. Hong, Y., Wu, Q., Qi, Y., Rodriguez-Opazo, C., Gould, S.: Vln bert: A recurrent vision-and-language bert for navigation. In: CVPR. pp. 1643-1653 (2021) +19. Huo, J., Sun, Q., Jiang, B., Lin, H., Fu, Y.: Geovln: Learning geometry-enhanced visual representation with slot attention for vision-and-language navigation. In: CVPR (2023) +20. Hwang, M., Jeong, J., Kim, M., Oh, Y., Oh, S.: Meta-exlore: Exploratory hierarchical vision-and-language navigation using scene object spectrum grounding. In: CVPR (2023) +21. Ilharco, G., Jain, V., Ku, A., Ie, E., Baldridge, J.: General evaluation for instruction conditioned navigation using dynamic time warping. In: NeurIPS Workshop (2019) +22. Kim, W., Son, B., Kim, I.: Vilt: Vision-and-language transformer without convolution or region supervision. In: ICML (2021) +23. Ku, A., Anderson, P., Patel, R., Ie, E., Baldridge, J.: Room-across-room: Multilingual vision-and-language navigation with dense spatiotemporal grounding. In: EMNLP (2020) +24. Li, J., Tan, H., Bansal, M.: Envedit: Environment editing for vision-and-language navigation. In: CVPR (2022) +25. Li, J., Selvaraju, R., Gotmare, A., Joty, S., Xiong, C., Hoi, S.C.H.: Align before fuse: Vision and language representation learning with momentum distillation. NeurIPS (2021) +26. Li, X., Li, C., Xia, Q., Bisk, Y., Celikylmaz, A., Gao, J., Smith, N.A., Choi, Y.: Robust navigation with language pretraining and stochastic sampling. In: EMNLP-IJCNLP (2019) +27. Li, X., Li, C., Xia, Q., Bisk, Y., Celikyilmaz, A., Gao, J., Smith, N.A., Choi, Y.: Robust navigation with language pretraining and stochastic sampling. In: EMNLP-IJCNLP (2019) +28. Lin, C., Jiang, Y., Cai, J., Qu, L., Haffari, G., Yuan, Z.: Multimodal transformer with variable-length memory for vision-and-language navigation. In: ECCV (2022) +29. Liu, C., Zhu, F., Chang, X., Liang, X., Ge, Z., Shen, Y.: Vision-language navigation with random environmental mixup. In: ICCV (2021) +30. Loshchilov, I., Hutter, F.: Decoupled weight decay regularization. In: ICLR (2019) +31. Ma, C.Y., Lu, J., Wu, Z., AlRegib, G., Kira, Z., Socher, R., Xiong, C.: Self-monitoring navigation agent via auxiliary progress estimation. In: ICLR (2019) +32. Ma, C., Wu, Z., AlRegib, G., Xiong, C., Kira, Z.: The regretful agent: Heuristic-aided navigation through progress estimation. In: CVPR (2019) +33. Majumdar, A., Shrivastava, A., Lee, S., Anderson, P., Parikh, D., Batra, D.: Improving vision-and-language navigation with image-text pairs from the web. In: ECCV (2020) +34. Oquab, M., Darcet, T., Moutakanni, T., Vo, H., Szafraniec, M., Khalidov, V., Fernandez, P., Haziza, D., Massa, F., El-Nouby, A., Assran, M., Ballas, N., Galuba, W., Howes, R., Huang, P., Li, S., Misra, I., Rabbat, M.G., Sharma, V., Synnaeve, G., Xu, H., Jegou, H., Mairal, J., Labatut, P., Joulin, A., Bojanowski, P.: Dinov2: Learning robust visual features without supervision. CoRR (2023) +35. Pope, R., Douglas, S., Chowdhery, A., Devlin, J., Bradbury, J., Levskaya, A., Heek, J., Xiao, K., Agrawal, S., Dean, J.: Efficiently scaling transformer inference. CoRR (2022) +36. Qi, Y., Wu, Q., Anderson, P., Wang, X., Wang, W.Y., Shen, C., Hengel, A.v.d.: Reverie: Remote embodied visual referring expression in real indoor environments. In: CVPR (2020) + +37. Radford, A., Kim, J.W., Hallacy, C., Ramesh, A., Goh, G., Agarwal, S., Sastry, G., Askell, A., Mishkin, P., Clark, J., et al.: Learning transferable visual models from natural language supervision. In: ICML (2021) +38. Sennrich, R., Haddow, B., Birch, A.: Improving neural machine translation models with monolingual data. In: ACL (2016) +39. Shen, S., Li, L.H., Tan, H., Bansal, M., Rohrbach, A., Chang, K., Yao, Z., Keutzer, K.: How much can CLIP benefit vision-and-language tasks? In: ICLR (2022) +40. Tan, H., Yu, L., Bansal, M.: Learning to navigate unseen environments: Back translation with environmental dropout. In: NAACL (2019) +41. Thomason, J., Murray, M., Cakmak, M., Zettlemoyer, L.: Vision-and-dialog navigation. In: CoRL (2020) +42. Thrun, S.: Learning metric-topological maps for indoor mobile robot navigation. Artif. Intell. (1998) +43. Vaswani, A., Shazeer, N., Parmar, N., Uszkoreit, J., Jones, L., Gomez, A.N., Kaiser, L.u., Polosukhin, I.: Attention is all you need. In: NeurIPS (2017) +44. Wang, H., Wang, W., Liang, W., Xiong, C., Shen, J.: Structured scene memory for vision-language navigation. In: CVPR (2021) +45. Wang, S., Montgomery, C., Orbay, J., Birodkar, V., Faust, A., Gur, I., Jaques, N., Waters, A., Baldridge, J., Anderson, P.: Less is more: Generating grounded navigation instructions from landmarks. In: CVPR (2022) +46. Wang, X., Huang, Q., Celikyilmaz, A., Gao, J., Shen, D., Wang, Y.F., Wang, W.Y., Zhang, L.: Reinforced cross-modal matching and self-supervised imitation learning for vision-language navigation. In: CVPR (2019) +47. Wang, Z., Li, X., Yang, J., Liu, Y., Jiang, S.: Gridmm: Grid memory map for vision-and-language navigation. In: ICCV (2023) +48. Zhu, F., Zhu, Y., Chang, X., Liang, X.: Vision-language navigation with self-supervised auxiliary reasoning tasks. In: CVPR (2020) \ No newline at end of file diff --git a/2024/PRET_ Planning with Directed Fidelity Trajectory for Vision and Language Navigation/images.zip b/2024/PRET_ Planning with Directed Fidelity Trajectory for Vision and Language Navigation/images.zip new file mode 100644 index 0000000000000000000000000000000000000000..811cb734ba2332202e0e55431dd7990674463373 --- /dev/null +++ b/2024/PRET_ Planning with Directed Fidelity Trajectory for Vision and Language Navigation/images.zip @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:a3468a8851d13e75f7023ab79fd7e85e2b5d5e7b015cfb3beebfba17744a049e +size 439314 diff --git a/2024/PRET_ Planning with Directed Fidelity Trajectory for Vision and Language Navigation/layout.json b/2024/PRET_ Planning with Directed Fidelity Trajectory for Vision and Language Navigation/layout.json new file mode 100644 index 0000000000000000000000000000000000000000..f8ef5751b29f75ae96db8399a12de46d6d86f55f --- /dev/null +++ b/2024/PRET_ Planning with Directed Fidelity Trajectory for Vision and Language Navigation/layout.json @@ -0,0 +1,8801 @@ +{ + "pdf_info": [ + { + "para_blocks": [ + { + "bbox": [ + 141, + 111, + 473, + 148 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 141, + 111, + 473, + 148 + ], + "spans": [ + { + "bbox": [ + 141, + 111, + 473, + 148 + ], + "type": "text", + "content": "PRET: Planning with Directed Fidelity Trajectory for Vision and Language Navigation" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 178, + 168, + 436, + 182 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 178, + 168, + 436, + 182 + ], + "spans": [ + { + "bbox": [ + 178, + 168, + 436, + 182 + ], + "type": "text", + "content": "Renjie Lu1, Jingke Meng1(3), and Wei-Shi Zheng1,2,3" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 160, + 190, + 452, + 213 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 160, + 190, + 452, + 213 + ], + "spans": [ + { + "bbox": [ + 160, + 190, + 452, + 213 + ], + "type": "inline_equation", + "content": "^{1}" + }, + { + "bbox": [ + 160, + 190, + 452, + 213 + ], + "type": "text", + "content": " School of Computer Science and Engineering, Sun Yat-sen University, Guangzhou, China" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 169, + 213, + 444, + 224 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 169, + 213, + 444, + 224 + ], + "spans": [ + { + "bbox": [ + 169, + 213, + 444, + 224 + ], + "type": "text", + "content": "lurj3@mail2.sysu.edu.cn,mengjke@gmail.com, wszheng@ieee.org" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 217, + 224, + 397, + 234 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 217, + 224, + 397, + 234 + ], + "spans": [ + { + "bbox": [ + 217, + 224, + 397, + 234 + ], + "type": "text", + "content": "2 Peng Cheng Laboratory, Shenzhen, China" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 165, + 234, + 447, + 257 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 165, + 234, + 447, + 257 + ], + "spans": [ + { + "bbox": [ + 165, + 234, + 447, + 257 + ], + "type": "text", + "content": "3 Key Laboratory of Machine Intelligence and Advanced Computing, Ministry of Education, Guangzhou, China" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 160, + 289, + 455, + 510 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 160, + 289, + 455, + 510 + ], + "spans": [ + { + "bbox": [ + 160, + 289, + 455, + 510 + ], + "type": "text", + "content": "Abstract. Vision and language navigation is a task that requires an agent to navigate according to a natural language instruction. Recent methods predict sub-goals on constructed topology map at each step to enable long-term action planning. However, they suffer from high computational cost when attempting to support such high-level predictions with GCN-like models. In this work, we propose an alternative method that facilitates navigation planning by considering the alignment between instructions and directed fidelity trajectories, which refers to a path from the initial node to the candidate locations on a directed graph without detours. This planning strategy leads to an efficient model while achieving strong performance. Specifically, we introduce a directed graph to illustrate the explored area of the environment, emphasizing directionality. Then, we firstly define the trajectory representation as a sequence of directed edge features, which are extracted from the panorama based on the corresponding orientation. Ultimately, we assess and compare the alignment between instruction and different trajectories during navigation to determine the next navigation target. Our method outperforms previous SOTA method BEVBert on RxR dataset and is comparable on R2R dataset while largely reducing the computational cost. Code is available: https://github.com/iSEE-Laboratory/VLN-PRET." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 160, + 521, + 394, + 533 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 160, + 521, + 394, + 533 + ], + "spans": [ + { + "bbox": [ + 160, + 521, + 394, + 533 + ], + "type": "text", + "content": "Keywords: Vision-and-Language Navigation " + }, + { + "bbox": [ + 160, + 521, + 394, + 533 + ], + "type": "inline_equation", + "content": "\\cdot" + }, + { + "bbox": [ + 160, + 521, + 394, + 533 + ], + "type": "text", + "content": " Planning" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 132, + 554, + 230, + 567 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 554, + 230, + 567 + ], + "spans": [ + { + "bbox": [ + 132, + 554, + 230, + 567 + ], + "type": "text", + "content": "1 Introduction" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 130, + 581, + 482, + 667 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 581, + 482, + 667 + ], + "spans": [ + { + "bbox": [ + 130, + 581, + 482, + 667 + ], + "type": "text", + "content": "Enabling a robot to perform tasks on behalf of humans has been a longstanding objective in AI research. One such task is vision-and-language navigation (VLN) [4, 23, 36, 41], where an agent is required to navigate to a desired location by following natural language instructions provided by humans. For example, given the instruction \"Go up stairs and stop at the top in front of a mirror.\", the agent needs to follow the instruction and stop at an appropriate location. VLN has attracted numerous research interests [2, 7, 8, 14, 16-18, 20, 26, 31, 40, 44, 47]." + } + ] + } + ], + "index": 9 + } + ], + "discarded_blocks": [], + "page_size": [ + 612, + 792 + ], + "page_idx": 0 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 166, + 114, + 449, + 266 + ], + "blocks": [ + { + "bbox": [ + 166, + 114, + 449, + 266 + ], + "lines": [ + { + "bbox": [ + 166, + 114, + 449, + 266 + ], + "spans": [ + { + "bbox": [ + 166, + 114, + 449, + 266 + ], + "type": "image", + "image_path": "7a7ffb983c5a1fa866447a08f99871bf31d168a8f065d88a44faab7fbe398a88.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 130, + 267, + 482, + 312 + ], + "lines": [ + { + "bbox": [ + 130, + 267, + 482, + 312 + ], + "spans": [ + { + "bbox": [ + 130, + 267, + 482, + 312 + ], + "type": "text", + "content": "Fig. 1: Comparison of SPL [3] and GFLOPs on R2R test unseen split dataset. Our method is comparable with previous SOTA methods while being more computational efficient. The computational cost of text encoder and visual encoder is omitted for fair comparison." + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_caption" + } + ], + "index": 2 + }, + { + "bbox": [ + 130, + 329, + 482, + 508 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 329, + 482, + 508 + ], + "spans": [ + { + "bbox": [ + 130, + 329, + 482, + 508 + ], + "type": "text", + "content": "Recent methods [2, 8, 16, 20, 44] demonstrate the effectiveness of introducing maps to enlarge the decision space to improve planning strategies. Instead of predicting low-level actions limited to short movements, these methods construct a graph to keep track of all visited and navigable locations observed so far, which enables high-level planning by expanding the action space to encompass the entire explored area. With the constructed topological maps, these methods predict actions in the global space via GCN-like models, where node features aggregate neighboring information. However, repeatedly calculating the entire map to predict actions at every step, even if the topo-map has only minor changes, is inefficient. In addition, formulating vision information of an environment in graph structure is too coarse-grained for accurate decision-making. Previous methods [2, 8] address this problem by introducing an additional branch to incorporate fine-grained information, such as local image features and bird's-eye-view features. But this strategy increases the model complexity and further escalates computational cost." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 130, + 510, + 483, + 666 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 510, + 483, + 666 + ], + "spans": [ + { + "bbox": [ + 130, + 510, + 483, + 666 + ], + "type": "text", + "content": "In this work, we present an alternative way that supports the global decision space while achieving comparable performance. Our method, named Planning with DiReacted FidElity Trajectory(PRET), does not require calculating the entire graph at each step, nor does it rely on incorporating additional fine-grained information. The main idea is to determine the next location to navigate by evaluating the alignment between the instruction and the visual observations along different trajectories between the start point to all the unvisited nodes, as shown in Figure 2(b). Specifically, we maintain a directed fidelity trajectory-colored in red) for each unvisited node. The fidelity trajectory refers to a path from initial node to an unvisited node without detours. We assess the alignment between each trajectory and the instruction, and select the unvisited node with the highest alignment score to navigate. This planning strategy is efficient as we only need to compute for newly observed nodes. We estimate the instruction-trajectory" + } + ] + } + ], + "index": 5 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "spans": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "type": "text", + "content": "2" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 166, + 91, + 202, + 100 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 166, + 91, + 202, + 100 + ], + "spans": [ + { + "bbox": [ + 166, + 91, + 202, + 100 + ], + "type": "text", + "content": "Lu et al." + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 1 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 187, + 118, + 422, + 242 + ], + "blocks": [ + { + "bbox": [ + 187, + 118, + 422, + 242 + ], + "lines": [ + { + "bbox": [ + 187, + 118, + 422, + 242 + ], + "spans": [ + { + "bbox": [ + 187, + 118, + 422, + 242 + ], + "type": "image", + "image_path": "5291bd7aeed0f2eafe974f85b8f83f4c9bd21f9a287954cbfa27807e2c1b4014.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 131, + 247, + 482, + 291 + ], + "lines": [ + { + "bbox": [ + 131, + 247, + 482, + 291 + ], + "spans": [ + { + "bbox": [ + 131, + 247, + 482, + 291 + ], + "type": "text", + "content": "Fig. 2: Illustration of our approaches. (a) shows our directed graph representation. Each edge is assigned with an orientation-aware panorama feature. (b) depict our planning method. We select an unvisited node-colored yellow to navigate towards next by choosing the fidelity path-colored red that best aligned with instruction." + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_caption" + } + ], + "index": 2 + }, + { + "bbox": [ + 130, + 312, + 480, + 348 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 312, + 480, + 348 + ], + "spans": [ + { + "bbox": [ + 130, + 312, + 480, + 348 + ], + "type": "text", + "content": "alignment with transformer and further reduce the computational cost by utilizing the KV-cache [35] technique. It also enables us to make decisions by taking into full consideration of the instruction-trajectory alignment." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 130, + 351, + 482, + 567 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 351, + 482, + 567 + ], + "spans": [ + { + "bbox": [ + 130, + 351, + 482, + 567 + ], + "type": "text", + "content": "In order to enhance the alignment between instructions and navigation trajectories, we propose to incorporate directionality in path representation. Due to the inherent directional nature of the navigation process, it is crucial that the representation of trajectories in opposite directions is asymmetric, as visual observations are linked to observed orientation. In this work, we introduce the directed graph to depict the explored area of the environment and firstly define the trajectory representation as a sequence of directed edge features. These directed edge features are derived from the panorama based on the corresponding orientation (as depicted in Fig. 2(a)). By incorporating directionality in the graph edge features, we eliminate irrelevant panorama information and obtain a more accurate representation of the directed trajectory. Our method overcome existing methods [2,8,16] that commonly adopt the undirected graph to store panorama features on graph nodes, resulting in direction-invariant path representations, which fails to capture the important distinction between paths that traverse the same node from different directions. For instance, consider a scenario where agents pass through the same node from different directions. Despite the fundamentally different spatial contexts of these paths, the node representation remains unchanged." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 130, + 570, + 482, + 666 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 570, + 482, + 666 + ], + "spans": [ + { + "bbox": [ + 130, + 570, + 482, + 666 + ], + "type": "text", + "content": "We conduct experiments on R2R, RxR datasets to evaluate the efficacy of our proposed methods. The results demonstrate that PRET achieves strong performance while being more efficient than previous methods. Specifically, PRET achieves comparable performance compared to previous state-of-the-art method BEVBert [2] on R2R dataset with only " + }, + { + "bbox": [ + 130, + 570, + 482, + 666 + ], + "type": "inline_equation", + "content": "3\\%" + }, + { + "bbox": [ + 130, + 570, + 482, + 666 + ], + "type": "text", + "content": " computational cost as shown in Fig. 1. On the RxR dataset, PRET outperforms previous methods and achieves the new state-of-the-art performance. Qualitative visualization also shows that our simplified model is able to learn complex backtracking strategies." + } + ] + } + ], + "index": 6 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 420, + 91, + 447, + 100 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 420, + 91, + 447, + 100 + ], + "spans": [ + { + "bbox": [ + 420, + 91, + 447, + 100 + ], + "type": "text", + "content": "PRET" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 474, + 92, + 480, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 474, + 92, + 480, + 100 + ], + "spans": [ + { + "bbox": [ + 474, + 92, + 480, + 100 + ], + "type": "text", + "content": "3" + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 2 + }, + { + "para_blocks": [ + { + "bbox": [ + 132, + 114, + 238, + 127 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 114, + 238, + 127 + ], + "spans": [ + { + "bbox": [ + 132, + 114, + 238, + 127 + ], + "type": "text", + "content": "2 Related Work" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 130, + 138, + 482, + 281 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 138, + 482, + 281 + ], + "spans": [ + { + "bbox": [ + 130, + 138, + 482, + 281 + ], + "type": "text", + "content": "Vision-and-Language Navigation(VLN). Early VLN methods use sequence-to-sequence LSTMs with attention mechanism to encode language features and predict local actions. SpeakerFollower [14] first introduces back translation [38] technique to ease the data scarcity problem in VLN. EnvDrop [40] proposes to augment environment by using dropout on feature space to avoid overfitting on visual input. Following this work, many environment augmentation approaches [24, 29] is proposed. Some auxiliary tasks is proposed by [31, 48] for better guidance. More recently, transformer is explored how to be adopted in VLN. PRESS [27] uses BERT [11] as the text encoder. PREVALENT [17] first propose to pretrain the transformer in VLN dataset, but still use it as encoder. More recent works [7, 18, 28] explore how to memorize navigation history and use transformer to learn strong planning strategy." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 130, + 282, + 482, + 402 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 282, + 482, + 402 + ], + "spans": [ + { + "bbox": [ + 130, + 282, + 482, + 402 + ], + "type": "text", + "content": "Navigation Strategy in VLN. Navigation strategy plays a key role in Vision- and-Language Navigation, as VLN requires agents to navigate in unseen environments and thus needs to explore and familiarize the environment. Works like [32] investigate designing regretful agents to enable explicit exploration. Reinforcement learning methods [40, 46] have also been explored to enhance navigation strategies. SSM [44] constructs directed graphs to represent explored areas, with vision features on nodes and orientation on edges. DUET [8] employs a dual-scale transformer to make local and global predictions. AZHP [16] constructs hierarchy graphs to facilitate exploration. MetaExplore [20] explicitly predicts whether to backtrack using a separate module." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 130, + 402, + 482, + 570 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 402, + 482, + 570 + ], + "spans": [ + { + "bbox": [ + 130, + 402, + 482, + 570 + ], + "type": "text", + "content": "Maps For Navigation. Navigation research has a long history of using SLAM [15] to construct maps [5, 6, 42] for planning. In the VLN literature, several methods [8, 44] adopt topological graphs for global planning. However, the visual representation on topological maps is too coarse-grained for decision-making. To address this limitation, [2, 47] introduce metric maps in VLN. While grid-based metric maps can precisely represent spatial layouts, they result in high computational costs. To balance representation ability and computational cost, BEVBert [2] utilizes learnable hybrid topological and metric maps. However, the computational cost remains high. In our work, we propose a novel approach that involves constructing a directed topo-map and planning with trajectory. Our method is more efficient as we perform planning with trajectory instead of directly relying on GCN-like models to encode the entire map for planning. This also allows us to incrementally calculate embeddings for new trajectories and further reduces computational cost." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 132, + 586, + 202, + 598 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 586, + 202, + 598 + ], + "spans": [ + { + "bbox": [ + 132, + 586, + 202, + 598 + ], + "type": "text", + "content": "3 Method" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 132, + 610, + 269, + 622 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 610, + 269, + 622 + ], + "spans": [ + { + "bbox": [ + 132, + 610, + 269, + 622 + ], + "type": "text", + "content": "3.1 Problem Formulation" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 130, + 629, + 481, + 666 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 629, + 481, + 666 + ], + "spans": [ + { + "bbox": [ + 130, + 629, + 481, + 666 + ], + "type": "text", + "content": "In VLN with discrete environments [4,23,36,41], an environment is an undirected graph " + }, + { + "bbox": [ + 130, + 629, + 481, + 666 + ], + "type": "inline_equation", + "content": "\\mathcal{G} = \\{\\mathcal{V},\\mathcal{E}\\}" + }, + { + "bbox": [ + 130, + 629, + 481, + 666 + ], + "type": "text", + "content": ", where " + }, + { + "bbox": [ + 130, + 629, + 481, + 666 + ], + "type": "inline_equation", + "content": "\\mathcal{V} = \\{V_i\\}_{i=1}^N" + }, + { + "bbox": [ + 130, + 629, + 481, + 666 + ], + "type": "text", + "content": " represents " + }, + { + "bbox": [ + 130, + 629, + 481, + 666 + ], + "type": "inline_equation", + "content": "N" + }, + { + "bbox": [ + 130, + 629, + 481, + 666 + ], + "type": "text", + "content": " navigable nodes and " + }, + { + "bbox": [ + 130, + 629, + 481, + 666 + ], + "type": "inline_equation", + "content": "\\mathcal{E}" + }, + { + "bbox": [ + 130, + 629, + 481, + 666 + ], + "type": "text", + "content": " denotes navigable edges. At the beginning of navigation, an agent is initialized" + } + ] + } + ], + "index": 8 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 133, + 91, + 141, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 91, + 141, + 100 + ], + "spans": [ + { + "bbox": [ + 133, + 91, + 141, + 100 + ], + "type": "text", + "content": "4" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 166, + 91, + 202, + 101 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 166, + 91, + 202, + 101 + ], + "spans": [ + { + "bbox": [ + 166, + 91, + 202, + 101 + ], + "type": "text", + "content": "Lu et al." + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 3 + }, + { + "para_blocks": [ + { + "bbox": [ + 130, + 115, + 479, + 151 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 115, + 479, + 151 + ], + "spans": [ + { + "bbox": [ + 130, + 115, + 479, + 151 + ], + "type": "text", + "content": "at a starting node and given a natural language instruction " + }, + { + "bbox": [ + 130, + 115, + 479, + 151 + ], + "type": "inline_equation", + "content": "W = \\{w_{i}\\}_{i=1}^{L}" + }, + { + "bbox": [ + 130, + 115, + 479, + 151 + ], + "type": "text", + "content": " with " + }, + { + "bbox": [ + 130, + 115, + 479, + 151 + ], + "type": "inline_equation", + "content": "L" + }, + { + "bbox": [ + 130, + 115, + 479, + 151 + ], + "type": "text", + "content": " words. The agent is required to interpret this instruction to navigate to the target location." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 130, + 152, + 480, + 223 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 152, + 480, + 223 + ], + "spans": [ + { + "bbox": [ + 130, + 152, + 480, + 223 + ], + "type": "text", + "content": "At step " + }, + { + "bbox": [ + 130, + 152, + 480, + 223 + ], + "type": "inline_equation", + "content": "t" + }, + { + "bbox": [ + 130, + 152, + 480, + 223 + ], + "type": "text", + "content": " on node " + }, + { + "bbox": [ + 130, + 152, + 480, + 223 + ], + "type": "inline_equation", + "content": "V_{t}" + }, + { + "bbox": [ + 130, + 152, + 480, + 223 + ], + "type": "text", + "content": ", the agent observes (1) a panorama " + }, + { + "bbox": [ + 130, + 152, + 480, + 223 + ], + "type": "inline_equation", + "content": "\\mathcal{R}_t = \\{r_{t,i}\\}_{i=1}^K" + }, + { + "bbox": [ + 130, + 152, + 480, + 223 + ], + "type": "text", + "content": " represented by " + }, + { + "bbox": [ + 130, + 152, + 480, + 223 + ], + "type": "inline_equation", + "content": "K" + }, + { + "bbox": [ + 130, + 152, + 480, + 223 + ], + "type": "text", + "content": " images from different views, " + }, + { + "bbox": [ + 130, + 152, + 480, + 223 + ], + "type": "inline_equation", + "content": "r_{t,i}" + }, + { + "bbox": [ + 130, + 152, + 480, + 223 + ], + "type": "text", + "content": " is the extracted image feature of the " + }, + { + "bbox": [ + 130, + 152, + 480, + 223 + ], + "type": "inline_equation", + "content": "i" + }, + { + "bbox": [ + 130, + 152, + 480, + 223 + ], + "type": "text", + "content": "th view; (2) neighboring navigable nodes " + }, + { + "bbox": [ + 130, + 152, + 480, + 223 + ], + "type": "inline_equation", + "content": "\\mathcal{N}(V_t)" + }, + { + "bbox": [ + 130, + 152, + 480, + 223 + ], + "type": "text", + "content": ", where " + }, + { + "bbox": [ + 130, + 152, + 480, + 223 + ], + "type": "inline_equation", + "content": "\\mathcal{N}(V_t) \\subset \\mathcal{V}" + }, + { + "bbox": [ + 130, + 152, + 480, + 223 + ], + "type": "text", + "content": "; (3) orientations and coordinates of these neighboring nodes. The agent should choose a neighboring node to step to or stop at current location. Navigation is considered successful if the agent stops within 3 meters of the target." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 130, + 224, + 481, + 294 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 224, + 481, + 294 + ], + "spans": [ + { + "bbox": [ + 130, + 224, + 481, + 294 + ], + "type": "text", + "content": "We treat navigation as a process of searching temporary target to navigate towards next among unvisited nodes on " + }, + { + "bbox": [ + 130, + 224, + 481, + 294 + ], + "type": "inline_equation", + "content": "\\mathcal{G}_t" + }, + { + "bbox": [ + 130, + 224, + 481, + 294 + ], + "type": "text", + "content": ". A temporary target can be either a local neighboring node or remote unvisited node. By adding a virtual stop node connected to all nodes on the graph, we can also model the stop action. We refer fidelity trajectory to a path from initial node to an unvisited node without detours." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 132, + 312, + 244, + 323 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 312, + 244, + 323 + ], + "spans": [ + { + "bbox": [ + 132, + 312, + 244, + 323 + ], + "type": "text", + "content": "3.2 Model Overview" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 130, + 331, + 482, + 547 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 331, + 482, + 547 + ], + "spans": [ + { + "bbox": [ + 130, + 331, + 482, + 547 + ], + "type": "text", + "content": "The overall framework is shown in Fig. 3(a). Our planning method consists of three components: (1) We construct a directed graph for explored area. Considering the directional nature of navigation, we propose an Orientation-aware Panorama Encoder(OPE) to extract orientation-aware vision features for edges. Edge sequence is used to represent directed trajectory and align with the instruction. (2) We maintain a path (i.e. directed fidelity trajectory) to each unvisited node and assess the alignment between each path and the instruction at each step. Only few neighboring nodes are added at each step, we only compute for these relevant new paths. Since each node corresponds to a path, we calculate a path embedding with a matching assessment module(MAM) that encodes the instruction-path alignment and stores the embedding on the node. (3) We compare the path embeddings of all unvisited nodes(candidates) with candidate comparison module(CCM)(shown in Fig. 3(c)) to determine which aligns best with the instruction. We select the best aligned candidate node as the temporary target. Then the agent can navigate to the temporary target along the shortest path on constructed graph. Since the path embedding is an estimate of the alignment between instruction and a path to the node, CCM ensures that the agent navigates by fully considering the instruction-trajectory alignment." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 132, + 563, + 406, + 574 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 563, + 406, + 574 + ], + "spans": [ + { + "bbox": [ + 132, + 563, + 406, + 574 + ], + "type": "text", + "content": "3.3 Orientation-Aware Directed Graph Construction" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 130, + 582, + 482, + 666 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 582, + 482, + 666 + ], + "spans": [ + { + "bbox": [ + 130, + 582, + 482, + 666 + ], + "type": "text", + "content": "We construct a directed graph to support path representation, incrementally adding nodes and edges as exploration proceeds. Previous methods [2,8,44] represent visited nodes with panorama features. However, panoramas of unvisited nodes are inaccessible as the agent has not visited them. So they represent these nodes with views towards them, leading to inconsistent representation between visited and unvisited nodes. In contrast, we extract visual features for the directed edges on the graph. These edge features represent the visual observations" + } + ] + } + ], + "index": 8 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 419, + 90, + 447, + 100 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 419, + 90, + 447, + 100 + ], + "spans": [ + { + "bbox": [ + 419, + 90, + 447, + 100 + ], + "type": "text", + "content": "PRET" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 474, + 91, + 481, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 474, + 91, + 481, + 100 + ], + "spans": [ + { + "bbox": [ + 474, + 91, + 481, + 100 + ], + "type": "text", + "content": "5" + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 4 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 136, + 116, + 479, + 209 + ], + "blocks": [ + { + "bbox": [ + 136, + 116, + 479, + 209 + ], + "lines": [ + { + "bbox": [ + 136, + 116, + 479, + 209 + ], + "spans": [ + { + "bbox": [ + 136, + 116, + 479, + 209 + ], + "type": "image", + "image_path": "2909c5b6168339dd7a93e5197057553ab89b003042ea5ba060d8584919acd8b3.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_body" + } + ], + "index": 2 + }, + { + "type": "image", + "bbox": [ + 136, + 209, + 376, + 304 + ], + "blocks": [ + { + "bbox": [ + 136, + 209, + 376, + 304 + ], + "lines": [ + { + "bbox": [ + 136, + 209, + 376, + 304 + ], + "spans": [ + { + "bbox": [ + 136, + 209, + 376, + 304 + ], + "type": "image", + "image_path": "0e14ba6e264285d2340f19a06a9ea1d5e964bc88b1099d91ff4d1487bc904edc.jpg" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 135, + 306, + 478, + 315 + ], + "lines": [ + { + "bbox": [ + 135, + 306, + 478, + 315 + ], + "spans": [ + { + "bbox": [ + 135, + 306, + 478, + 315 + ], + "type": "text", + "content": "visited node current node unvisited node start token shared edge token unvisited neighboring edge token stop token" + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "image_footnote" + }, + { + "bbox": [ + 130, + 318, + 482, + 396 + ], + "lines": [ + { + "bbox": [ + 130, + 318, + 482, + 396 + ], + "spans": [ + { + "bbox": [ + 130, + 318, + 482, + 396 + ], + "type": "text", + "content": "Fig. 3: Illustration of our model. (a) is the overall framework of our method. At each step, we update the graph, extract path embeddings, and predict actions. (b) depicts the matching assessment module(MAM). Each token is an edge feature. We compute path embeddings for each newly observed nodes with cross-modal transformer and impose a causal mask to reduce computational cost. (c) shows the candidate comparison module(CCM). We gather path embeddings of unvisited nodes and forward them into a single layer transformer followed by a MLP to predict temporary target." + } + ] + } + ], + "index": 6, + "angle": 0, + "type": "image_caption" + } + ], + "index": 3 + }, + { + "type": "image", + "bbox": [ + 383, + 211, + 460, + 299 + ], + "blocks": [ + { + "bbox": [ + 383, + 211, + 460, + 299 + ], + "lines": [ + { + "bbox": [ + 383, + 211, + 460, + 299 + ], + "spans": [ + { + "bbox": [ + 383, + 211, + 460, + 299 + ], + "type": "image", + "image_path": "a9f78e2971a855100fe31657721cd55dc149d31d8a54902342afb2ab2d6ed1d3.jpg" + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "image_body" + } + ], + "index": 4 + }, + { + "bbox": [ + 130, + 419, + 480, + 454 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 419, + 480, + 454 + ], + "spans": [ + { + "bbox": [ + 130, + 419, + 480, + 454 + ], + "type": "text", + "content": "when facing specific directions from each node. As edges rely only on observable views rather than node panoramas, they enable consistent representation for both visited and unvisited nodes." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 130, + 456, + 482, + 515 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 456, + 482, + 515 + ], + "spans": [ + { + "bbox": [ + 130, + 456, + 482, + 515 + ], + "type": "text", + "content": "Let " + }, + { + "bbox": [ + 130, + 456, + 482, + 515 + ], + "type": "inline_equation", + "content": "\\mathcal{G}_t = \\{\\mathcal{V}_t, \\mathcal{E}_t\\}" + }, + { + "bbox": [ + 130, + 456, + 482, + 515 + ], + "type": "text", + "content": " be the directed graph at step " + }, + { + "bbox": [ + 130, + 456, + 482, + 515 + ], + "type": "inline_equation", + "content": "t" + }, + { + "bbox": [ + 130, + 456, + 482, + 515 + ], + "type": "text", + "content": ". Neighboring nodes " + }, + { + "bbox": [ + 130, + 456, + 482, + 515 + ], + "type": "inline_equation", + "content": "\\mathcal{N}(V_t)" + }, + { + "bbox": [ + 130, + 456, + 482, + 515 + ], + "type": "text", + "content": " are observed, and some nodes in " + }, + { + "bbox": [ + 130, + 456, + 482, + 515 + ], + "type": "inline_equation", + "content": "\\mathcal{N}(V_t)" + }, + { + "bbox": [ + 130, + 456, + 482, + 515 + ], + "type": "text", + "content": " have already existed in " + }, + { + "bbox": [ + 130, + 456, + 482, + 515 + ], + "type": "inline_equation", + "content": "\\mathcal{G}_t" + }, + { + "bbox": [ + 130, + 456, + 482, + 515 + ], + "type": "text", + "content": " while others are newly observed and added to " + }, + { + "bbox": [ + 130, + 456, + 482, + 515 + ], + "type": "inline_equation", + "content": "\\mathcal{G}_t" + }, + { + "bbox": [ + 130, + 456, + 482, + 515 + ], + "type": "text", + "content": ". Besides, directed edges from " + }, + { + "bbox": [ + 130, + 456, + 482, + 515 + ], + "type": "inline_equation", + "content": "V_t" + }, + { + "bbox": [ + 130, + 456, + 482, + 515 + ], + "type": "text", + "content": " to nodes in " + }, + { + "bbox": [ + 130, + 456, + 482, + 515 + ], + "type": "inline_equation", + "content": "\\mathcal{N}(V_t)" + }, + { + "bbox": [ + 130, + 456, + 482, + 515 + ], + "type": "text", + "content": " together with features extracted by orientation-aware panorama encoder(OPE) will also be added in " + }, + { + "bbox": [ + 130, + 456, + 482, + 515 + ], + "type": "inline_equation", + "content": "G_t" + }, + { + "bbox": [ + 130, + 456, + 482, + 515 + ], + "type": "text", + "content": ", as shown in Fig. 3(a)." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 130, + 535, + 482, + 606 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 535, + 482, + 606 + ], + "spans": [ + { + "bbox": [ + 130, + 535, + 482, + 606 + ], + "type": "text", + "content": "Visual Representation For Edges. We extract orientation-aware panorama feature by using relative orientation as query to attend panorama features in cross-attention manner, as shown in the OPE module in Fig. 3(a). Assume that " + }, + { + "bbox": [ + 130, + 535, + 482, + 606 + ], + "type": "inline_equation", + "content": "(\\phi, \\theta)" + }, + { + "bbox": [ + 130, + 535, + 482, + 606 + ], + "type": "text", + "content": " represents the relative heading and elevation of a specific edge when compared to the agent's current orientation. To extract feature for the edge, we first encode " + }, + { + "bbox": [ + 130, + 535, + 482, + 606 + ], + "type": "inline_equation", + "content": "(\\phi, \\theta)" + }, + { + "bbox": [ + 130, + 535, + 482, + 606 + ], + "type": "text", + "content": " as follows:" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 219, + 616, + 481, + 632 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 219, + 616, + 481, + 632 + ], + "spans": [ + { + "bbox": [ + 219, + 616, + 481, + 632 + ], + "type": "interline_equation", + "content": "x ^ {a} = \\left[ \\sin (\\phi), \\cos (\\phi), \\sin (\\theta), \\cos (\\theta) \\right] W ^ {a}, \\tag {1}", + "image_path": "adf892ca666a2be07f25938e4b870b4128ce3a289aff9689af5d53e761faf714.jpg" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 130, + 639, + 482, + 667 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 639, + 482, + 667 + ], + "spans": [ + { + "bbox": [ + 130, + 639, + 482, + 667 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 130, + 639, + 482, + 667 + ], + "type": "inline_equation", + "content": "W^{a}\\in \\mathbb{R}^{4\\times d}" + }, + { + "bbox": [ + 130, + 639, + 482, + 667 + ], + "type": "text", + "content": " are learnable weights and " + }, + { + "bbox": [ + 130, + 639, + 482, + 667 + ], + "type": "inline_equation", + "content": "d" + }, + { + "bbox": [ + 130, + 639, + 482, + 667 + ], + "type": "text", + "content": " is the dimension of feature. Then we denote " + }, + { + "bbox": [ + 130, + 639, + 482, + 667 + ], + "type": "inline_equation", + "content": "X_{t}^{a} = \\{x_{t,i}^{a}\\}_{i = 1}^{\\lfloor \\mathcal{N}(V_{t})\\rfloor}" + }, + { + "bbox": [ + 130, + 639, + 482, + 667 + ], + "type": "text", + "content": " as orientation features of adjacent edges of " + }, + { + "bbox": [ + 130, + 639, + 482, + 667 + ], + "type": "inline_equation", + "content": "V_{t}" + }, + { + "bbox": [ + 130, + 639, + 482, + 667 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 11 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "spans": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "type": "text", + "content": "6" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 166, + 91, + 202, + 100 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 166, + 91, + 202, + 100 + ], + "spans": [ + { + "bbox": [ + 166, + 91, + 202, + 100 + ], + "type": "text", + "content": "Lu et al." + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 5 + }, + { + "para_blocks": [ + { + "bbox": [ + 130, + 115, + 482, + 189 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 115, + 482, + 189 + ], + "spans": [ + { + "bbox": [ + 130, + 115, + 482, + 189 + ], + "type": "text", + "content": "We then encode the panorama feature. As the panorama " + }, + { + "bbox": [ + 130, + 115, + 482, + 189 + ], + "type": "inline_equation", + "content": "\\mathcal{R}_t = \\{r_{t,i}\\}_{i=1}^K" + }, + { + "bbox": [ + 130, + 115, + 482, + 189 + ], + "type": "text", + "content": " is represented by " + }, + { + "bbox": [ + 130, + 115, + 482, + 189 + ], + "type": "inline_equation", + "content": "K" + }, + { + "bbox": [ + 130, + 115, + 482, + 189 + ], + "type": "text", + "content": " images from different views and each view corresponds to a fixed orientation, we can also get the relative heading and elevation of each view. To encode the panorama, we first concatenate each view feature " + }, + { + "bbox": [ + 130, + 115, + 482, + 189 + ], + "type": "inline_equation", + "content": "r_{t,i}" + }, + { + "bbox": [ + 130, + 115, + 482, + 189 + ], + "type": "text", + "content": " with corresponding relative orientation " + }, + { + "bbox": [ + 130, + 115, + 482, + 189 + ], + "type": "inline_equation", + "content": "(\\phi_{t,i},\\theta_{t,i})" + }, + { + "bbox": [ + 130, + 115, + 482, + 189 + ], + "type": "text", + "content": " and then map it to dimension " + }, + { + "bbox": [ + 130, + 115, + 482, + 189 + ], + "type": "inline_equation", + "content": "d" + }, + { + "bbox": [ + 130, + 115, + 482, + 189 + ], + "type": "text", + "content": " via a linear projection:" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 197, + 196, + 481, + 209 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 197, + 196, + 481, + 209 + ], + "spans": [ + { + "bbox": [ + 197, + 196, + 481, + 209 + ], + "type": "interline_equation", + "content": "x _ {t, i} ^ {p} = \\left[ r _ {t, i}; \\sin \\left(\\phi_ {t, i}\\right), \\cos \\left(\\phi_ {t, i}\\right), \\sin \\left(\\theta_ {t, i}\\right), \\cos \\left(\\theta_ {t, i}\\right) \\right] W ^ {p}. \\tag {2}", + "image_path": "43ed0cf3cc75d31ac36796c72ffb87c9a9725af8cc5228b412c5fc361c0cd048.jpg" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 131, + 216, + 479, + 252 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 131, + 216, + 479, + 252 + ], + "spans": [ + { + "bbox": [ + 131, + 216, + 479, + 252 + ], + "type": "text", + "content": "Then we denote " + }, + { + "bbox": [ + 131, + 216, + 479, + 252 + ], + "type": "inline_equation", + "content": "X_{t}^{p} = \\{x_{t,i}^{p}\\}_{i=1}^{K}" + }, + { + "bbox": [ + 131, + 216, + 479, + 252 + ], + "type": "text", + "content": " as the panoramic view representation. By considering views as patches of the panorama and orientations as position embedding, " + }, + { + "bbox": [ + 131, + 216, + 479, + 252 + ], + "type": "inline_equation", + "content": "X_{t}^{p}" + }, + { + "bbox": [ + 131, + 216, + 479, + 252 + ], + "type": "text", + "content": " is like the encoded ViT [12] input." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 131, + 253, + 479, + 275 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 131, + 253, + 479, + 275 + ], + "spans": [ + { + "bbox": [ + 131, + 253, + 479, + 275 + ], + "type": "text", + "content": "We adopt transformer decoder [43] to extract orientation-aware panorama feature:" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 230, + 277, + 480, + 289 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 230, + 277, + 480, + 289 + ], + "spans": [ + { + "bbox": [ + 230, + 277, + 480, + 289 + ], + "type": "interline_equation", + "content": "E _ {t} = \\operatorname {T r a n s f o r m e r D e c o d e r} \\left(X _ {t} ^ {a}, X _ {t} ^ {p}\\right), \\tag {3}", + "image_path": "38108b3cffdc89cb16286b3b0b4429553e7e92da6e1e6b47f4ee8d08c07c5a3d.jpg" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 130, + 294, + 482, + 367 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 294, + 482, + 367 + ], + "spans": [ + { + "bbox": [ + 130, + 294, + 482, + 367 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 130, + 294, + 482, + 367 + ], + "type": "inline_equation", + "content": "E_{t} = \\{e_{t,i}\\}_{i = 1}^{\\lfloor \\mathcal{N}(V_{t})\\rfloor}" + }, + { + "bbox": [ + 130, + 294, + 482, + 367 + ], + "type": "text", + "content": " are extracted edge features. By taking " + }, + { + "bbox": [ + 130, + 294, + 482, + 367 + ], + "type": "inline_equation", + "content": "X_{t}^{a}" + }, + { + "bbox": [ + 130, + 294, + 482, + 367 + ], + "type": "text", + "content": " as the query input and " + }, + { + "bbox": [ + 130, + 294, + 482, + 367 + ], + "type": "inline_equation", + "content": "X_{t}^{p}" + }, + { + "bbox": [ + 130, + 294, + 482, + 367 + ], + "type": "text", + "content": " as the key-value input, panorama is attended by orientation in the first layer. In the subsequent layers, both orientation and vision information are used to query relevant views in the panorama. Therefore, each directed edge only focuses on a specific region of the panorama, thus orientation-aware panorama feature is extracted." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 131, + 384, + 330, + 396 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 131, + 384, + 330, + 396 + ], + "spans": [ + { + "bbox": [ + 131, + 384, + 330, + 396 + ], + "type": "text", + "content": "3.4 Planning with Fidelity Trajectory" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 130, + 402, + 482, + 486 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 402, + 482, + 486 + ], + "spans": [ + { + "bbox": [ + 130, + 402, + 482, + 486 + ], + "type": "text", + "content": "We maintain a directed fidelity trajectory with a stack during navigation for each unvisited node on " + }, + { + "bbox": [ + 130, + 402, + 482, + 486 + ], + "type": "inline_equation", + "content": "\\mathcal{G}_t" + }, + { + "bbox": [ + 130, + 402, + 482, + 486 + ], + "type": "text", + "content": ". When a new node is added, it is pushed into the stack. When the agent needs to backtrack, the top node is popped off the stack until a different node than the top is reached. In this way, detours are removed by popping nodes off the stack. With the maintained trajectory of each node, we propose MAM to assess its alignment with instruction. We then determine an unvisited node on " + }, + { + "bbox": [ + 130, + 402, + 482, + 486 + ], + "type": "inline_equation", + "content": "\\mathcal{G}_t" + }, + { + "bbox": [ + 130, + 402, + 482, + 486 + ], + "type": "text", + "content": " to navigate next by comparing this alignment with CCM." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 130, + 502, + 481, + 550 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 502, + 481, + 550 + ], + "spans": [ + { + "bbox": [ + 130, + 502, + 481, + 550 + ], + "type": "text", + "content": "Matching Assessment Module(MAM). MAM is a multi-layer transformer decoder that extracts a path embedding for each path, which estimates how much the trajectory match with the instruction. These path embeddings are then stored on the corresponding nodes in graph " + }, + { + "bbox": [ + 130, + 502, + 481, + 550 + ], + "type": "inline_equation", + "content": "\\mathcal{G}_t" + }, + { + "bbox": [ + 130, + 502, + 481, + 550 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 130, + 550, + 482, + 597 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 550, + 482, + 597 + ], + "spans": [ + { + "bbox": [ + 130, + 550, + 482, + 597 + ], + "type": "text", + "content": "For a single path with length " + }, + { + "bbox": [ + 130, + 550, + 482, + 597 + ], + "type": "inline_equation", + "content": "l" + }, + { + "bbox": [ + 130, + 550, + 482, + 597 + ], + "type": "text", + "content": ", we represent the path with the edge features along it, denoted as " + }, + { + "bbox": [ + 130, + 550, + 482, + 597 + ], + "type": "inline_equation", + "content": "X^{h} = [e_{1}, e_{2}, \\dots, e_{l}]" + }, + { + "bbox": [ + 130, + 550, + 482, + 597 + ], + "type": "text", + "content": ". " + }, + { + "bbox": [ + 130, + 550, + 482, + 597 + ], + "type": "inline_equation", + "content": "\\{e_{i}\\}_{i=1}^{l}" + }, + { + "bbox": [ + 130, + 550, + 482, + 597 + ], + "type": "text", + "content": " are the edge features on the path. Then we add it with position embedding and forward it together with text embedding " + }, + { + "bbox": [ + 130, + 550, + 482, + 597 + ], + "type": "inline_equation", + "content": "X^{w}" + }, + { + "bbox": [ + 130, + 550, + 482, + 597 + ], + "type": "text", + "content": " into MAM like follows:" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 224, + 604, + 294, + 617 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 224, + 604, + 294, + 617 + ], + "spans": [ + { + "bbox": [ + 224, + 604, + 294, + 617 + ], + "type": "interline_equation", + "content": "X ^ {h \\prime} = X ^ {h} + P _ {l},", + "image_path": "7ec9ab7a62acd8c6c53293b2d343baaade5a1601cd798a46ff57c8ad1f80ade0.jpg" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 227, + 614, + 480, + 634 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 227, + 614, + 480, + 634 + ], + "spans": [ + { + "bbox": [ + 227, + 614, + 480, + 634 + ], + "type": "interline_equation", + "content": "X ^ {o} = \\operatorname {T r a n s f o r m e r D e c o d e r} \\left(X ^ {h t}, X ^ {w}\\right), \\tag {4}", + "image_path": "92060d10717652e6bc2abf99684e8701f29483a91cc0419934e88d7b0ea33bbc.jpg" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 130, + 641, + 481, + 666 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 641, + 481, + 666 + ], + "spans": [ + { + "bbox": [ + 130, + 641, + 481, + 666 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 130, + 641, + 481, + 666 + ], + "type": "inline_equation", + "content": "P_{l}" + }, + { + "bbox": [ + 130, + 641, + 481, + 666 + ], + "type": "text", + "content": " is a matrix including " + }, + { + "bbox": [ + 130, + 641, + 481, + 666 + ], + "type": "inline_equation", + "content": "l" + }, + { + "bbox": [ + 130, + 641, + 481, + 666 + ], + "type": "text", + "content": " position embeddings [43] and " + }, + { + "bbox": [ + 130, + 641, + 481, + 666 + ], + "type": "inline_equation", + "content": "X^{w}" + }, + { + "bbox": [ + 130, + 641, + 481, + 666 + ], + "type": "text", + "content": " is the text embedding extracted by a text encoder, which is a multi-layer transformer [43]." + } + ] + } + ], + "index": 14 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 419, + 91, + 447, + 100 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 419, + 91, + 447, + 100 + ], + "spans": [ + { + "bbox": [ + 419, + 91, + 447, + 100 + ], + "type": "text", + "content": "PRET" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 474, + 91, + 481, + 99 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 474, + 91, + 481, + 99 + ], + "spans": [ + { + "bbox": [ + 474, + 91, + 481, + 99 + ], + "type": "text", + "content": "7" + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 6 + }, + { + "para_blocks": [ + { + "bbox": [ + 130, + 116, + 482, + 176 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 116, + 482, + 176 + ], + "spans": [ + { + "bbox": [ + 130, + 116, + 482, + 176 + ], + "type": "text", + "content": "By utilizing two sequences from two modalities as inputs, the transformer is capable of effectively estimating their alignment, as extensively demonstrated in Vision and Language Pretraining studies [9, 22, 25]. The last token of the encoded sequence " + }, + { + "bbox": [ + 130, + 116, + 482, + 176 + ], + "type": "inline_equation", + "content": "X^o" + }, + { + "bbox": [ + 130, + 116, + 482, + 176 + ], + "type": "text", + "content": " provides the path embedding, which is an estimation of the text-path alignment." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 130, + 175, + 482, + 272 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 175, + 482, + 272 + ], + "spans": [ + { + "bbox": [ + 130, + 175, + 482, + 272 + ], + "type": "text", + "content": "To support STOP action, we add a stop node that connected to all nodes in " + }, + { + "bbox": [ + 130, + 175, + 482, + 272 + ], + "type": "inline_equation", + "content": "\\mathcal{G}_t" + }, + { + "bbox": [ + 130, + 175, + 482, + 272 + ], + "type": "text", + "content": ". Since we represent a path using edge features, we assign a learnable stop token, initialized with a zero vector, to all edges that point to the stop node. The path embedding corresponds to the stop node is stored on current node. So the agent decides to stop at current location when it finds that the path up to here appending with stop token matches instruction the best. Here stop token serves as the End-of-Sentence token in text generation. For each path, we also add a learnable start token that serves as the Start-of-Sentence token." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 130, + 272, + 482, + 464 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 272, + 482, + 464 + ], + "spans": [ + { + "bbox": [ + 130, + 272, + 482, + 464 + ], + "type": "text", + "content": "At step " + }, + { + "bbox": [ + 130, + 272, + 482, + 464 + ], + "type": "inline_equation", + "content": "t" + }, + { + "bbox": [ + 130, + 272, + 482, + 464 + ], + "type": "text", + "content": ", considering the stop node, assume there are " + }, + { + "bbox": [ + 130, + 272, + 482, + 464 + ], + "type": "inline_equation", + "content": "N" + }, + { + "bbox": [ + 130, + 272, + 482, + 464 + ], + "type": "text", + "content": " neighboring unvisited nodes to be updated. Computing path embeddings separately for these nodes requires multiple forward passes, which is inefficient. Actually, we can compute these embeddings in a single forward pass, as shown in Fig. 3(b). Note that these paths share the same prefix from the initial node to the current node. We can avoid duplicated computation by imposing a causal mask in the self-attention block. The causal mask is a lower triangular matrix that only allows a token to attend to its previous tokens. We merge these causal masks together and forward all relevant tokens into the transformer block. The merged mask still satisfies the constraint that each token can only attend to its previous tokens. We further reduce the computational cost by adopting the KV-cache technique [35], which stores the shared prefix tokens to avoid unnecessary re-calculation. In this way, we can equivalently compute the path embeddings for multiple newly observed nodes in a single forward pass. The last several output tokens colored yellow and green are path embeddings and used to update the graph, i.e., stored on corresponding graph nodes." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 130, + 478, + 482, + 587 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 478, + 482, + 587 + ], + "spans": [ + { + "bbox": [ + 130, + 478, + 482, + 587 + ], + "type": "text", + "content": "Candidate Comparison Module(CCM). CCM compares the path embeddings from candidate nodes to predict a temporary target. As each path embedding encodes the alignment between a trajectory and the instruction, the CCM actually selects a path that is most aligned with the instruction. In this way, we make navigation planning fully according to the instruction-trajectory alignment. As shown in Fig. 3(c), we collect path embeddings " + }, + { + "bbox": [ + 130, + 478, + 482, + 587 + ], + "type": "inline_equation", + "content": "X_{t}^{e}" + }, + { + "bbox": [ + 130, + 478, + 482, + 587 + ], + "type": "text", + "content": " and input them into CCM. CCM consists of a single-layer transformer encoder [43] and a MLP. The transformer encoder is responsible for the comparison of the candidates and the MLP maps the encoded path embedding to a 1D score:" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 242, + 593, + 373, + 605 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 242, + 593, + 373, + 605 + ], + "spans": [ + { + "bbox": [ + 242, + 593, + 373, + 605 + ], + "type": "interline_equation", + "content": "X _ {t} ^ {e \\prime} = \\operatorname {T r a n s f o r m e r L a y e r} \\left(X _ {t} ^ {e}\\right),", + "image_path": "eae0a4728387a0e0a60a368e47208e0aa9ad635671b0fcd51a7c6472ca692859.jpg" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 249, + 608, + 480, + 620 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 249, + 608, + 480, + 620 + ], + "spans": [ + { + "bbox": [ + 249, + 608, + 480, + 620 + ], + "type": "interline_equation", + "content": "s _ {t} = \\operatorname {M L P} \\left(X _ {t} ^ {e ^ {\\prime}}\\right), \\tag {5}", + "image_path": "393a4d988002548f3038af7838e96ce0bfde1767097a63839c70c9350fc6070f.jpg" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 249, + 624, + 324, + 635 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 249, + 624, + 324, + 635 + ], + "spans": [ + { + "bbox": [ + 249, + 624, + 324, + 635 + ], + "type": "interline_equation", + "content": "p _ {t} = \\operatorname {s o f t m a x} \\left(s _ {t}\\right).", + "image_path": "685e059d97addc8191bd8079702ba9fe2dd89ac9a1bbc11f95bad24d76de0523.jpg" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 130, + 641, + 482, + 666 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 641, + 482, + 666 + ], + "spans": [ + { + "bbox": [ + 130, + 641, + 482, + 666 + ], + "type": "text", + "content": "Score " + }, + { + "bbox": [ + 130, + 641, + 482, + 666 + ], + "type": "inline_equation", + "content": "s_t" + }, + { + "bbox": [ + 130, + 641, + 482, + 666 + ], + "type": "text", + "content": " reflects the relative text-path alignment after comparison. An alternative way is removing the transformer and directly compute the alignment score " + }, + { + "bbox": [ + 130, + 641, + 482, + 666 + ], + "type": "inline_equation", + "content": "s_t" + }, + { + "bbox": [ + 130, + 641, + 482, + 666 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 9 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "spans": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "type": "text", + "content": "8" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 166, + 91, + 202, + 101 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 166, + 91, + 202, + 101 + ], + "spans": [ + { + "bbox": [ + 166, + 91, + 202, + 101 + ], + "type": "text", + "content": "Lu et al." + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 7 + }, + { + "para_blocks": [ + { + "bbox": [ + 131, + 116, + 479, + 188 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 131, + 116, + 479, + 188 + ], + "spans": [ + { + "bbox": [ + 131, + 116, + 479, + 188 + ], + "type": "text", + "content": "However, this method assesses each path independently without any comparison, which can make the decision-making process more challenging and potentially result in a decrease in performance(demonstrated in Sec. 4). It is then normalized with softmax and we select the highest score node as the temporary target. The agent will navigate to it along the shortest path on " + }, + { + "bbox": [ + 131, + 116, + 479, + 188 + ], + "type": "inline_equation", + "content": "\\mathcal{G}_t" + }, + { + "bbox": [ + 131, + 116, + 479, + 188 + ], + "type": "text", + "content": ". If the stop node is selected, the agent believes the instruction is completed and stops the navigation." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 132, + 204, + 303, + 216 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 204, + 303, + 216 + ], + "spans": [ + { + "bbox": [ + 132, + 204, + 303, + 216 + ], + "type": "text", + "content": "3.5 Pretraining and Fine-tuning" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 131, + 221, + 480, + 365 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 131, + 221, + 480, + 365 + ], + "spans": [ + { + "bbox": [ + 131, + 221, + 480, + 365 + ], + "type": "text", + "content": "Pretraining. Pretraining is proved to be helpful in previous works [2,7,17,33]. PREVALENT [17] synthesized a substantial amount of data for pretraining, and we utilized the same data as their work. We use Masked Language Modeling(MLM) [11] for pretraining. Concretely, we randomly replace " + }, + { + "bbox": [ + 131, + 221, + 480, + 365 + ], + "type": "inline_equation", + "content": "15\\%" + }, + { + "bbox": [ + 131, + 221, + 480, + 365 + ], + "type": "text", + "content": " of the input tokens with a special [MASK] token and forward these masked tokens into the text encoder. We also extract a sequence of edge features to represent the path. Then the encoded text tokens and path representation are input into an additional transformer decoder. The corresponding outputs are used to predict the masked tokens. MLM helps the model learn aligned text and path representations. In VLN, a key difference from other vision-and-language tasks is that we align a text sequence with a sequence of vision representations rather than a single image." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 131, + 380, + 479, + 417 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 131, + 380, + 479, + 417 + ], + "spans": [ + { + "bbox": [ + 131, + 380, + 479, + 417 + ], + "type": "text", + "content": "Fine-tuning. We train the agent with a mixture of teacher-forcing and student-forcing as previous methods [2,8]. In the teacher-forcing stage, the agent navigate according the ground truth actions " + }, + { + "bbox": [ + 131, + 380, + 479, + 417 + ], + "type": "inline_equation", + "content": "a_{t}^{*}" + }, + { + "bbox": [ + 131, + 380, + 479, + 417 + ], + "type": "text", + "content": " and the loss is calculated as follows:" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 247, + 425, + 480, + 455 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 247, + 425, + 480, + 455 + ], + "spans": [ + { + "bbox": [ + 247, + 425, + 480, + 455 + ], + "type": "interline_equation", + "content": "\\mathcal {L} _ {T F} = \\frac {1}{T _ {T F}} \\sum_ {t = 1} ^ {T _ {T F}} \\mathrm {C E} \\left(p _ {t}, a _ {t} ^ {*}\\right), \\tag {6}", + "image_path": "3172c0590e00714397bc4c43108e1e4160e011b398eca8a6bb5f8830d197f4b0.jpg" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 131, + 462, + 479, + 510 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 131, + 462, + 479, + 510 + ], + "spans": [ + { + "bbox": [ + 131, + 462, + 479, + 510 + ], + "type": "text", + "content": "CE is the cross entropy loss and " + }, + { + "bbox": [ + 131, + 462, + 479, + 510 + ], + "type": "inline_equation", + "content": "T_{TF}" + }, + { + "bbox": [ + 131, + 462, + 479, + 510 + ], + "type": "text", + "content": " is the number of navigate steps. In the student-forcing stage, we sample an action from the distribution predicted by the agent so that the agent can explore the environment and reduce the exposure bias. We train the agent with heuristic pseudo label " + }, + { + "bbox": [ + 131, + 462, + 479, + 510 + ], + "type": "inline_equation", + "content": "a_t^{pseudo}" + }, + { + "bbox": [ + 131, + 462, + 479, + 510 + ], + "type": "text", + "content": ":" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 238, + 518, + 479, + 548 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 238, + 518, + 479, + 548 + ], + "spans": [ + { + "bbox": [ + 238, + 518, + 479, + 548 + ], + "type": "interline_equation", + "content": "\\mathcal {L} _ {S F} = \\frac {1}{T _ {S F}} \\sum_ {t = 1} ^ {T _ {S F}} \\mathrm {C E} \\left(p _ {t}, a _ {t} ^ {\\text {p s e u d o}}\\right). \\tag {7}", + "image_path": "b6cb999ab4544dea3a000e6638d655c0dfc9adfc59ebd6a822ed33f4a5c85db5.jpg" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 131, + 555, + 479, + 626 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 131, + 555, + 479, + 626 + ], + "spans": [ + { + "bbox": [ + 131, + 555, + 479, + 626 + ], + "type": "text", + "content": "When the agent deviate from ground truth path, we take the nearest node on ground truth path as the pseudo label to encourage agent to learn backtracking strategy. If there is no unvisited node on ground truth path, we take the nearest node on the shortest path from current node to target node as the pseudo label. The agent navigate two times to compute " + }, + { + "bbox": [ + 131, + 555, + 479, + 626 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_{TF}" + }, + { + "bbox": [ + 131, + 555, + 479, + 626 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 131, + 555, + 479, + 626 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_{SF}" + }, + { + "bbox": [ + 131, + 555, + 479, + 626 + ], + "type": "text", + "content": ". The total loss is the weighted sum of them:" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 252, + 635, + 479, + 646 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 252, + 635, + 479, + 646 + ], + "spans": [ + { + "bbox": [ + 252, + 635, + 479, + 646 + ], + "type": "interline_equation", + "content": "\\mathcal {L} = \\lambda \\mathcal {L} _ {T F} + (1 - \\lambda) \\mathcal {L} _ {S F}, \\tag {8}", + "image_path": "fd99a0e55fd897a1b106fb9cbe7a919ce94ccd4338588fe565072a3ae9fa7166.jpg" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 132, + 653, + 265, + 665 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 653, + 265, + 665 + ], + "spans": [ + { + "bbox": [ + 132, + 653, + 265, + 665 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 132, + 653, + 265, + 665 + ], + "type": "inline_equation", + "content": "\\lambda \\in (0,1)" + }, + { + "bbox": [ + 132, + 653, + 265, + 665 + ], + "type": "text", + "content": " is the weight." + } + ] + } + ], + "index": 11 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 419, + 91, + 447, + 100 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 419, + 91, + 447, + 100 + ], + "spans": [ + { + "bbox": [ + 419, + 91, + 447, + 100 + ], + "type": "text", + "content": "PRET" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 474, + 91, + 481, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 474, + 91, + 481, + 100 + ], + "spans": [ + { + "bbox": [ + 474, + 91, + 481, + 100 + ], + "type": "text", + "content": "9" + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 8 + }, + { + "para_blocks": [ + { + "bbox": [ + 132, + 114, + 230, + 129 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 114, + 230, + 129 + ], + "spans": [ + { + "bbox": [ + 132, + 114, + 230, + 129 + ], + "type": "text", + "content": "4 Experiments" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 132, + 139, + 326, + 152 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 139, + 326, + 152 + ], + "spans": [ + { + "bbox": [ + 132, + 139, + 326, + 152 + ], + "type": "text", + "content": "4.1 Datasets and Evaluation Metrics" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 132, + 158, + 481, + 218 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 158, + 481, + 218 + ], + "spans": [ + { + "bbox": [ + 132, + 158, + 481, + 218 + ], + "type": "text", + "content": "R2R. Room-to-Room (R2R) dataset [4] contains 7,189 trajectories, with 3 instructions per trajectory, across 90 scenes. These scenes are divided into train, val unseen and test unseen splits with 61, 11 and 18 scenes respectively. A val seen split with the same scenes as the train split is also provided. All paths in R2R are the shortest paths between the start and target nodes." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 132, + 235, + 481, + 295 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 235, + 481, + 295 + ], + "spans": [ + { + "bbox": [ + 132, + 235, + 481, + 295 + ], + "type": "text", + "content": "RxR. Room-across-Room(RxR) [23] is a large multilingual dataset. The dataset has 126,000 instructions total, with 42,000 instructions in each of the three languages: English, Hindi, and Telugu. The paths in RxR are longer than those in R2R and are not the shortest possible routes. Additionally, the instructions in RxR are longer and contain more detailed descriptions compared to R2R." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 132, + 312, + 481, + 432 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 312, + 481, + 432 + ], + "spans": [ + { + "bbox": [ + 132, + 312, + 481, + 432 + ], + "type": "text", + "content": "Evaluation Metrics. We adopt the following evaluation metrics on R2R: (1) Trajectory length (TL): the average length of the agent's path in meters. (2) Navigation error (NE): the average distance between the agent's final position and the target location. (3) Success rate (SR): the ratio of successful navigations, where " + }, + { + "bbox": [ + 132, + 312, + 481, + 432 + ], + "type": "inline_equation", + "content": "NE < 3m" + }, + { + "bbox": [ + 132, + 312, + 481, + 432 + ], + "type": "text", + "content": " is considered successful. (4) SR penalized by path length (SPL) [3]: TL longer than the shortest path is penalized. As paths in RxR dataset are shortest paths, TL and SPL are unsuitable. Instead, on RxR we use nDTW and sDTW [21] to measure trajectory similarity between the agent and ground truth. Please note that the fidelity trajectory that is removed detours is solely used for planning. It is not the agent's actual trajectory used for evaluation." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 132, + 449, + 279, + 461 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 449, + 279, + 461 + ], + "spans": [ + { + "bbox": [ + 132, + 449, + 279, + 461 + ], + "type": "text", + "content": "4.2 Implementation Details" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 132, + 468, + 481, + 540 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 468, + 481, + 540 + ], + "spans": [ + { + "bbox": [ + 132, + 468, + 481, + 540 + ], + "type": "text", + "content": "Module architecture. Text backbone is a 6 layer transformer encoder initialized with pretrained ALBEF [25]. For the multilingual RxR dataset, we use 12 layer mRoberta [10] instead. The layer numbers of the OPE, MAM, and CCM are 2, 4, and 1 respectively. All the hidden layer size is 768. Image features is extracted by DINOv2 [34]. We also report results with CLIP-ViT-base [37] feature for fair comparison." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 132, + 557, + 481, + 665 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 557, + 481, + 665 + ], + "spans": [ + { + "bbox": [ + 132, + 557, + 481, + 665 + ], + "type": "text", + "content": "Training details. On R2R dataset, we first pretrain PRET with learning rate 2e-5 and batch size 16 on a single 24G 4090 GPU for 100,000 iterations(" + }, + { + "bbox": [ + 132, + 557, + 481, + 665 + ], + "type": "inline_equation", + "content": "\\sim" + }, + { + "bbox": [ + 132, + 557, + 481, + 665 + ], + "type": "text", + "content": " 5 hours). Then we fine-tune the model with learning rate 1e-5 and batch size 8 on 4090 for 100,000 iterations(" + }, + { + "bbox": [ + 132, + 557, + 481, + 665 + ], + "type": "inline_equation", + "content": "\\sim" + }, + { + "bbox": [ + 132, + 557, + 481, + 665 + ], + "type": "text", + "content": " 25 hours). Optimizer is AdamW [30]. " + }, + { + "bbox": [ + 132, + 557, + 481, + 665 + ], + "type": "inline_equation", + "content": "\\lambda" + }, + { + "bbox": [ + 132, + 557, + 481, + 665 + ], + "type": "text", + "content": " is set to 0.2. Augmented dataset [17] is used in both pretraining and fine-tuning. The best result is selected by SPL on val unseen split. On RxR dataset, we also follow the pretrain and fine-tune paradigm. Due to the longer instruction and trajectory, we use a smaller batch size 4. Marky [45] is used as augmented data. The best result is selected by sDTW on val unseen split." + } + ] + } + ], + "index": 9 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "spans": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "type": "text", + "content": "10" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 166, + 91, + 202, + 100 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 166, + 91, + 202, + 100 + ], + "spans": [ + { + "bbox": [ + 166, + 91, + 202, + 100 + ], + "type": "text", + "content": "Lu et al." + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 9 + }, + { + "para_blocks": [ + { + "type": "table", + "bbox": [ + 143, + 145, + 474, + 352 + ], + "blocks": [ + { + "bbox": [ + 132, + 114, + 480, + 137 + ], + "lines": [ + { + "bbox": [ + 132, + 114, + 480, + 137 + ], + "spans": [ + { + "bbox": [ + 132, + 114, + 480, + 137 + ], + "type": "text", + "content": "Table 1: Comparison with other methods on R2R dataset. SPL is considered as the primary evaluation metric." + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 143, + 145, + 474, + 352 + ], + "lines": [ + { + "bbox": [ + 143, + 145, + 474, + 352 + ], + "spans": [ + { + "bbox": [ + 143, + 145, + 474, + 352 + ], + "type": "table", + "html": "
MethodsVal SeenVal UnseenTest Unseen
TLNE↓SR↑SPL↑TLNE↓SR↑SPL↑TLNE↓SR↑SPL↑
Seq2Seq-SF [4]11.336.0139-8.397.8122-8.137.852818
Speaker-Follower [14]-3.3666--6.6235-14.826.623528
RCM [46]10.653.5367-11.466.0943-11.976.124338
Regretful [32]-3.236963-5.325041-5.695640
EnvDrop [40]11.003.99625910.705.22524811.665.235147
PREVALENT [17]10.323.67696510.194.71585310.515.305451
NvEM [1]11.093.44696511.834.27605512.984.375854
SSM [44]14.703.10716220.704.32624520.404.576146
RecBert [18]11.132.90726812.013.93635712.354.096357
HAMT [7]11.152.51767211.462.29666112.273.936560
MTVM [28]-2.677469-3.736659-3.856559
DUET [8]12.322.28797313.943.31726014.733.656959
AZHP [16]----14.053.15726114.953.527160
Meta-Explore [20]11.952.11817513.093.22726214.253.577161
GridMM [47]----13.272.83756414.433.357362
BEVBert [2]13.562.17817414.552.81756415.873.137362
Ours(CLIP)11.482.60746912.213.12716313.873.127262
Ours(DINOv2)11.252.41787211.872.90746512.213.097264
", + "image_path": "b674c067151bf323d0e173be14674b2adb6246d5775a975807c865b412b3ea80.jpg" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "table_body" + } + ], + "index": 3 + }, + { + "bbox": [ + 132, + 376, + 197, + 387 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 376, + 197, + 387 + ], + "spans": [ + { + "bbox": [ + 132, + 376, + 197, + 387 + ], + "type": "text", + "content": "4.3 Results" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 132, + 401, + 481, + 581 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 401, + 481, + 581 + ], + "spans": [ + { + "bbox": [ + 132, + 401, + 481, + 581 + ], + "type": "text", + "content": "Comparison on R2R. Tab. 1 compares our approach with previous methods. On the val unseen and test unseen split, our method achieves comparable performance with previous SOTA method BEVBert on the primary metric SPL, while our method significantly reduces the computational cost. Besides, BEVBert [2] adopts additional depth information to construct bird's-eye view input. On the val seen split, our method performs sub-optimal as PRET does not take the environment layout as input, which means it less tends to overfit on the seen environment. Also, in the VLN literature, all visual features are pre-extracted and remains unchanged, the fixed feature may not contains appropriate feature for navigation. Therefore, we also investigate different vision features for VLN, as shown in Tab. 1. We find that DINOv2 [34] feature is better than CLIP feature. DINOv2 is trained on large-scale curated data in self-supervised manner, which leads to more general purpose feature. Moreover, DINOv2 feature contains geometric information as it can be used to predict depth map according to [34]. Geometric information is shown helpful in [19]." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 132, + 582, + 481, + 665 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 582, + 481, + 665 + ], + "spans": [ + { + "bbox": [ + 132, + 582, + 481, + 665 + ], + "type": "text", + "content": "Tab. 2 shows the parameter count, computational cost, and latency of different graph-based methods. We present GFLOPs of different navigation steps. Our method contains fewer parameters as we use a single stream model unlike [2,8] use dual stream transformer. Besides, our decoder computational cost is only " + }, + { + "bbox": [ + 132, + 582, + 481, + 665 + ], + "type": "inline_equation", + "content": "19\\%" + }, + { + "bbox": [ + 132, + 582, + 481, + 665 + ], + "type": "text", + "content": " of DUET and " + }, + { + "bbox": [ + 132, + 582, + 481, + 665 + ], + "type": "inline_equation", + "content": "3\\%" + }, + { + "bbox": [ + 132, + 582, + 481, + 665 + ], + "type": "text", + "content": " of BEVBert at step 10. These suggests that our method is significantly more efficient than previous SOTA graph-based methods. As shown in Tab. 2, this results in lower training and inference latency." + } + ] + } + ], + "index": 6 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 420, + 91, + 447, + 100 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 420, + 91, + 447, + 100 + ], + "spans": [ + { + "bbox": [ + 420, + 91, + 447, + 100 + ], + "type": "text", + "content": "PRET" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 470, + 91, + 480, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 470, + 91, + 480, + 100 + ], + "spans": [ + { + "bbox": [ + 470, + 91, + 480, + 100 + ], + "type": "text", + "content": "11" + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 10 + }, + { + "para_blocks": [ + { + "type": "table", + "bbox": [ + 159, + 190, + 452, + 243 + ], + "blocks": [ + { + "bbox": [ + 130, + 114, + 482, + 180 + ], + "lines": [ + { + "bbox": [ + 130, + 114, + 482, + 180 + ], + "spans": [ + { + "bbox": [ + 130, + 114, + 482, + 180 + ], + "type": "text", + "content": "Table 2: Comparison of latency and computational cost on R2R val unseen. * indicates we reproduce DUET with CLIP feature for fair comparison. When computing GFLOPs, we simplify the calculation by assuming that each vertex of the graph is connected to 4 adjacent vertices (average degree in R2R), and at each step, 3 new vertices are encountered. The computational cost of text encoder and visual encoder is omitted for fair comparison." + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 159, + 190, + 452, + 243 + ], + "lines": [ + { + "bbox": [ + 159, + 190, + 452, + 243 + ], + "spans": [ + { + "bbox": [ + 159, + 190, + 452, + 243 + ], + "type": "table", + "html": "
MethodsR2R val unseenLatency(ms)GFLOPsParams
TLNE↓SR↑SPL↑traininference1 step10 steps20 steps
DUET* [8]14.03.271.661.111.88.62.44.87.590M
BEVBert [2]14.62.874.963.616.210.633.735.537.590M
Ours12.23.171.062.77.74.10.60.91.264M
", + "image_path": "f3cfc9ca84b97482117a084186c6eea3e3804cc9d12991c504b2dcc8e512016e.jpg" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "table_body" + } + ], + "index": 3 + }, + { + "type": "table", + "bbox": [ + 168, + 306, + 443, + 418 + ], + "blocks": [ + { + "bbox": [ + 130, + 252, + 482, + 297 + ], + "lines": [ + { + "bbox": [ + 130, + 252, + 482, + 297 + ], + "spans": [ + { + "bbox": [ + 130, + 252, + 482, + 297 + ], + "type": "text", + "content": "Table 3: Comparison with previous methods on RxR dataset. †indicates approaches that utilize additional augmented data. nDTW represents the normalized DTW distance between the agent's trajectory and ground truth path. sDTW is the nDTW weighted by success rate." + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 168, + 306, + 443, + 418 + ], + "lines": [ + { + "bbox": [ + 168, + 306, + 443, + 418 + ], + "spans": [ + { + "bbox": [ + 168, + 306, + 443, + 418 + ], + "type": "table", + "html": "
MethodsVal SeenVal Unseen
NE↓SR↑nDTW↑sDTW↑NE↓SR↑nDTW↑sDTW↑
LSTM [23]10.725.242.220.710.922.838.918.2
EnvDrop+ [39]-----43.655.7-
HAMT [7]-59.465.350.9-56.563.148.3
EnvEdit [24]-67.271.158.5-62.868.554.6
MPM [13]-67.771.058.9-63.567.754.5
MARVEL [45]†3.075.979.168.84.564.870.857.5
BEVBert [2]†3.275.076.366.74.068.569.658.6
Ours(CLIP)†2.677.078.067.53.371.271.860.4
Ours(DINOv2)†2.479.380.470.73.272.873.462.4
", + "image_path": "45d92cff227dc4f30b11192666867175deebdcffe08fbbeb53b2ff788878af27.jpg" + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "table_body" + } + ], + "index": 5 + }, + { + "bbox": [ + 130, + 441, + 482, + 561 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 441, + 482, + 561 + ], + "spans": [ + { + "bbox": [ + 130, + 441, + 482, + 561 + ], + "type": "text", + "content": "Comparison on RxR. Tab. 3 reports the results on the RxR dataset. RxR is challenging as instructions and paths in RxR are longer than R2R. While the alignment between path and detailed path description is what our method skills at, our method outperforms previous methods among all metrics at a lower computational cost. PRET achieves a " + }, + { + "bbox": [ + 130, + 441, + 482, + 561 + ], + "type": "inline_equation", + "content": "1.8\\%" + }, + { + "bbox": [ + 130, + 441, + 482, + 561 + ], + "type": "text", + "content": " improvement on the main metrics sDTW on the val unseen split, demonstrating that our model better understands and follows instructions. This result highlights the advantage of our planning with trajectory strategy, which considers the alignment between instructions and paths and enables efficient decision-making in the global space. Additionally, with DINOv2 features, the result is further improved." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 131, + 578, + 238, + 590 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 131, + 578, + 238, + 590 + ], + "spans": [ + { + "bbox": [ + 131, + 578, + 238, + 590 + ], + "type": "text", + "content": "4.4 Ablation Study" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 130, + 597, + 481, + 635 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 597, + 481, + 635 + ], + "spans": [ + { + "bbox": [ + 130, + 597, + 481, + 635 + ], + "type": "text", + "content": "We conducted ablation experiments on various components of PRET, including the directed path representation and module ablation. Results are reported on the R2R val unseen split." + } + ] + } + ], + "index": 8 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "spans": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "type": "text", + "content": "12" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 166, + 91, + 202, + 100 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 166, + 91, + 202, + 100 + ], + "spans": [ + { + "bbox": [ + 166, + 91, + 202, + 100 + ], + "type": "text", + "content": "Lu et al." + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 11 + }, + { + "para_blocks": [ + { + "type": "table", + "bbox": [ + 141, + 150, + 269, + 184 + ], + "blocks": [ + { + "bbox": [ + 132, + 118, + 279, + 140 + ], + "lines": [ + { + "bbox": [ + 132, + 118, + 279, + 140 + ], + "spans": [ + { + "bbox": [ + 132, + 118, + 279, + 140 + ], + "type": "text", + "content": "Table 4: Comparison of undirected and directed path representation." + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 141, + 150, + 269, + 184 + ], + "lines": [ + { + "bbox": [ + 141, + 150, + 269, + 184 + ], + "spans": [ + { + "bbox": [ + 141, + 150, + 269, + 184 + ], + "type": "table", + "html": "
MethodsTLNE↓SR↑SPL↑
undirected15.623.5968.2856.77
directed11.872.9073.7865.16
", + "image_path": "84ce37c5c82808bb0b3cbbdc37558415d5474c3d9ef99e30d14d1019db09a5cf.jpg" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "table_body" + } + ], + "index": 3 + }, + { + "type": "table", + "bbox": [ + 291, + 135, + 460, + 187 + ], + "blocks": [ + { + "bbox": [ + 299, + 114, + 455, + 125 + ], + "lines": [ + { + "bbox": [ + 299, + 114, + 455, + 125 + ], + "spans": [ + { + "bbox": [ + 299, + 114, + 455, + 125 + ], + "type": "text", + "content": "Table 5: Ablation study on modules." + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 291, + 135, + 460, + 187 + ], + "lines": [ + { + "bbox": [ + 291, + 135, + 460, + 187 + ], + "spans": [ + { + "bbox": [ + 291, + 135, + 460, + 187 + ], + "type": "table", + "html": "
MethodsTLNE↓SR↑SPL↑
1 MAM12.043.9962.3254.48
2 MAM+CCM12.153.5465.9457.32
3 MAM+OPE12.183.1571.6063.07
4 MAM+OPE+CCM11.872.9073.7865.16
", + "image_path": "ca045507a1821b5ad005432c4d45bfd40e59b1f14ee5ac3031f5e03763c4da09.jpg" + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "table_body" + } + ], + "index": 5 + }, + { + "type": "image", + "bbox": [ + 219, + 198, + 395, + 239 + ], + "blocks": [ + { + "bbox": [ + 219, + 198, + 395, + 239 + ], + "lines": [ + { + "bbox": [ + 219, + 198, + 395, + 239 + ], + "spans": [ + { + "bbox": [ + 219, + 198, + 395, + 239 + ], + "type": "image", + "image_path": "73164e5bbdd0ccea64413afcdab955178397f303e85f580b1c421d823f92f2ff.jpg" + } + ] + } + ], + "index": 6, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 148, + 239, + 463, + 251 + ], + "lines": [ + { + "bbox": [ + 148, + 239, + 463, + 251 + ], + "spans": [ + { + "bbox": [ + 148, + 239, + 463, + 251 + ], + "type": "text", + "content": "Fig. 4: Comparison of orientation panoramic view and single candidate view." + } + ] + } + ], + "index": 7, + "angle": 0, + "type": "image_caption" + } + ], + "index": 6 + }, + { + "bbox": [ + 130, + 275, + 482, + 407 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 275, + 482, + 407 + ], + "spans": [ + { + "bbox": [ + 130, + 275, + 482, + 407 + ], + "type": "text", + "content": "Directionality in Path Representation. We compare whether to incorporate directionality for path representations in Tab. 4. Undirected path is represented by a sequence of node features. The node features is extracted by forwarding panoramic views into a 2-layer transformer encoder and averages the output tokens. Directed path is represented by a sequence of orientation-aware panorama features. According to the results, directed path representation outperforms undirected node features among all metrics. Specifically, it gains " + }, + { + "bbox": [ + 130, + 275, + 482, + 407 + ], + "type": "inline_equation", + "content": "8.39\\%" + }, + { + "bbox": [ + 130, + 275, + 482, + 407 + ], + "type": "text", + "content": " improvement on SPL and " + }, + { + "bbox": [ + 130, + 275, + 482, + 407 + ], + "type": "inline_equation", + "content": "5.5\\%" + }, + { + "bbox": [ + 130, + 275, + 482, + 407 + ], + "type": "text", + "content": " on SR. This indicates that edge feature is more suitable to represent the directional navigation process and align with the instruction. Node features does not distinguish different path directions and provide redundant information in a path, which hinders the alignment." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 130, + 426, + 482, + 521 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 426, + 482, + 521 + ], + "spans": [ + { + "bbox": [ + 130, + 426, + 482, + 521 + ], + "type": "text", + "content": "Module ablation. Tab. 5 shows ablation results on modules. Row 4 is the full model introduced in Sec. 3. OPE extracts orientation-aware features for edges, MAM extracts path embeddings while CCM compares them to select the temporary target. When removing CCM (Row 3), we only predict a single alignment score rather than a path embedding for each path without comparison. The score is then normalized and used to predict the temporary goal. Performance drops " + }, + { + "bbox": [ + 130, + 426, + 482, + 521 + ], + "type": "inline_equation", + "content": "2.09\\%" + }, + { + "bbox": [ + 130, + 426, + 482, + 521 + ], + "type": "text", + "content": " in this case, as selecting a path based solely on a single score, without comparing different paths, is more difficult." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 130, + 522, + 482, + 665 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 522, + 482, + 665 + ], + "spans": [ + { + "bbox": [ + 130, + 522, + 482, + 665 + ], + "type": "text", + "content": "In row 2, we remove OPE, panorama features are not used and only a single 60-degree view towards each neighboring node is adopted as edge features. Performance drops " + }, + { + "bbox": [ + 130, + 522, + 482, + 665 + ], + "type": "inline_equation", + "content": "7.84\\%" + }, + { + "bbox": [ + 130, + 522, + 482, + 665 + ], + "type": "text", + "content": " as the limited field-of-view cannot fully represent the path. The trajectory represented by these views is discrete and does not form a continuous change of views. In some cases, a single view towards candidates provides no information for navigation, as shown in Fig. 4. When go up the stairs, the upper view only sees the wall and does not help the navigation at all. While using the panorama, the agent can dynamically see a larger region. These results show our graph representation with panorama encoder provides rich representation for trajectories. Row 1 demonstrates the performance when using only MAM. The performance is poor without other model components, demonstrating the effectiveness of these modules." + } + ] + } + ], + "index": 10 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 419, + 91, + 447, + 100 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 419, + 91, + 447, + 100 + ], + "spans": [ + { + "bbox": [ + 419, + 91, + 447, + 100 + ], + "type": "text", + "content": "PRET" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 470, + 91, + 481, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 470, + 91, + 481, + 100 + ], + "spans": [ + { + "bbox": [ + 470, + 91, + 481, + 100 + ], + "type": "text", + "content": "13" + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 12 + }, + { + "para_blocks": [ + { + "bbox": [ + 132, + 116, + 227, + 126 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 116, + 227, + 126 + ], + "spans": [ + { + "bbox": [ + 132, + 116, + 227, + 126 + ], + "type": "text", + "content": "4.5 Visualization" + } + ] + } + ], + "index": 2 + }, + { + "type": "image", + "bbox": [ + 146, + 135, + 282, + 235 + ], + "blocks": [ + { + "bbox": [ + 146, + 135, + 282, + 235 + ], + "lines": [ + { + "bbox": [ + 146, + 135, + 282, + 235 + ], + "spans": [ + { + "bbox": [ + 146, + 135, + 282, + 235 + ], + "type": "image", + "image_path": "1661bb8d1b75a64d5109f9fae4bdfd50c34a1793afbf6832f1f8d3d623e5c5d0.jpg" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 144, + 236, + 280, + 255 + ], + "lines": [ + { + "bbox": [ + 144, + 236, + 280, + 255 + ], + "spans": [ + { + "bbox": [ + 144, + 236, + 280, + 255 + ], + "type": "text", + "content": "Instruction: Face directly away from stairs and walk through doorway straight ahead, right immediately, walk past bed and turn right into closet, walk straight and stop in front of toilet." + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 209, + 257, + 219, + 265 + ], + "lines": [ + { + "bbox": [ + 209, + 257, + 219, + 265 + ], + "spans": [ + { + "bbox": [ + 209, + 257, + 219, + 265 + ], + "type": "text", + "content": "(a)" + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "image_caption" + } + ], + "index": 3 + }, + { + "type": "image", + "bbox": [ + 305, + 144, + 385, + 169 + ], + "blocks": [ + { + "bbox": [ + 316, + 133, + 377, + 142 + ], + "lines": [ + { + "bbox": [ + 316, + 133, + 377, + 142 + ], + "spans": [ + { + "bbox": [ + 316, + 133, + 377, + 142 + ], + "type": "text", + "content": "undirected trajectory" + } + ] + } + ], + "index": 6, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 305, + 144, + 385, + 169 + ], + "lines": [ + { + "bbox": [ + 305, + 144, + 385, + 169 + ], + "spans": [ + { + "bbox": [ + 305, + 144, + 385, + 169 + ], + "type": "image", + "image_path": "5d8baa7f63037a6f0be0e12bc335b6b7ba26e0b1574f0545206d7fd0feae89bc.jpg" + } + ] + } + ], + "index": 7, + "angle": 0, + "type": "image_body" + } + ], + "index": 7 + }, + { + "type": "image", + "bbox": [ + 305, + 169, + 385, + 204 + ], + "blocks": [ + { + "bbox": [ + 305, + 169, + 385, + 204 + ], + "lines": [ + { + "bbox": [ + 305, + 169, + 385, + 204 + ], + "spans": [ + { + "bbox": [ + 305, + 169, + 385, + 204 + ], + "type": "image", + "image_path": "3c9519642b448d44cbce4747794e4d79c64fbb0c3b15bd002592649ddb5f5a6a.jpg" + } + ] + } + ], + "index": 8, + "angle": 0, + "type": "image_body" + } + ], + "index": 8 + }, + { + "type": "image", + "bbox": [ + 305, + 204, + 385, + 236 + ], + "blocks": [ + { + "bbox": [ + 305, + 204, + 385, + 236 + ], + "lines": [ + { + "bbox": [ + 305, + 204, + 385, + 236 + ], + "spans": [ + { + "bbox": [ + 305, + 204, + 385, + 236 + ], + "type": "image", + "image_path": "83c08a5184cabee19094350062c18cae036b3b49ddf035843b4986b7a15d641e.jpg" + } + ] + } + ], + "index": 9, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 345, + 242, + 441, + 249 + ], + "lines": [ + { + "bbox": [ + 345, + 242, + 441, + 249 + ], + "spans": [ + { + "bbox": [ + 345, + 242, + 441, + 249 + ], + "type": "text", + "content": "Instruction: Turn right and go up the stairs." + } + ] + } + ], + "index": 10, + "angle": 0, + "type": "image_caption" + } + ], + "index": 9 + }, + { + "type": "image", + "bbox": [ + 393, + 144, + 472, + 169 + ], + "blocks": [ + { + "bbox": [ + 406, + 133, + 461, + 142 + ], + "lines": [ + { + "bbox": [ + 406, + 133, + 461, + 142 + ], + "spans": [ + { + "bbox": [ + 406, + 133, + 461, + 142 + ], + "type": "text", + "content": "directed trajectory" + } + ] + } + ], + "index": 11, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 393, + 144, + 472, + 169 + ], + "lines": [ + { + "bbox": [ + 393, + 144, + 472, + 169 + ], + "spans": [ + { + "bbox": [ + 393, + 144, + 472, + 169 + ], + "type": "image", + "image_path": "659a53cccecbd5e255c4cd2007a9b649f08041e684f66e66f9d4d6c16c76aba8.jpg" + } + ] + } + ], + "index": 12, + "angle": 0, + "type": "image_body" + } + ], + "index": 12 + }, + { + "type": "image", + "bbox": [ + 393, + 169, + 472, + 204 + ], + "blocks": [ + { + "bbox": [ + 393, + 169, + 472, + 204 + ], + "lines": [ + { + "bbox": [ + 393, + 169, + 472, + 204 + ], + "spans": [ + { + "bbox": [ + 393, + 169, + 472, + 204 + ], + "type": "image", + "image_path": "9b7b767cb0e77eb4644e3e0a2791e9f95abb6ce5e8112d187235c43759854aaf.jpg" + } + ] + } + ], + "index": 13, + "angle": 0, + "type": "image_body" + } + ], + "index": 13 + }, + { + "type": "image", + "bbox": [ + 393, + 204, + 472, + 237 + ], + "blocks": [ + { + "bbox": [ + 393, + 204, + 472, + 237 + ], + "lines": [ + { + "bbox": [ + 393, + 204, + 472, + 237 + ], + "spans": [ + { + "bbox": [ + 393, + 204, + 472, + 237 + ], + "type": "image", + "image_path": "4574cc6c8740af637fc91b0614b2db35f5bf94ae75c5561d9135c162cf93a917.jpg" + } + ] + } + ], + "index": 14, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 384, + 256, + 395, + 265 + ], + "lines": [ + { + "bbox": [ + 384, + 256, + 395, + 265 + ], + "spans": [ + { + "bbox": [ + 384, + 256, + 395, + 265 + ], + "type": "text", + "content": "(b)" + } + ] + } + ], + "index": 15, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 130, + 270, + 482, + 304 + ], + "lines": [ + { + "bbox": [ + 130, + 270, + 482, + 304 + ], + "spans": [ + { + "bbox": [ + 130, + 270, + 482, + 304 + ], + "type": "text", + "content": "Fig. 5: (a) Visualization of the agent's navigation process, showcasing its ability to learn a backtracking strategy. (b) Visualizing attention weights in OPE to illustrate the distinction between undirected and directed trajectory representations." + } + ] + } + ], + "index": 16, + "angle": 0, + "type": "image_caption" + } + ], + "index": 14 + }, + { + "bbox": [ + 130, + 318, + 480, + 402 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 318, + 480, + 402 + ], + "spans": [ + { + "bbox": [ + 130, + 318, + 480, + 402 + ], + "type": "text", + "content": "Fig. 5(a) shows a qualitative example of our model's behavior. Initially placed at the green starting node, the agent wrongly decides to explore the upper left area(blue path). It takes many unnecessary detours there, but through comparing the trajectory alignment with the instructions, realizes these paths do not match the text well. Therefore, it backtracks to the start and correctly navigates to the destination(red path). This demonstrates the capability of PRET to learn complex backtracking behaviors." + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 130, + 402, + 482, + 498 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 402, + 482, + 498 + ], + "spans": [ + { + "bbox": [ + 130, + 402, + 482, + 498 + ], + "type": "text", + "content": "Fig. 5(b) shows the difference between undirected and directed trajectory representations. The former represents a path using panorama features stored on nodes, which eliminates directional discrepancies since the same node provides the same feature regardless of orientation, and contains redundant information compared to \"stairs\" in the instruction. In contrast, our proposed approach represents a path using orientation-aware panorama features on edges. This eliminates redundant information and focuses on the \"stairs\" better captures the directional navigation process." + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 132, + 512, + 220, + 525 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 512, + 220, + 525 + ], + "spans": [ + { + "bbox": [ + 132, + 512, + 220, + 525 + ], + "type": "text", + "content": "5 Conclusion" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 130, + 533, + 482, + 666 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 533, + 482, + 666 + ], + "spans": [ + { + "bbox": [ + 130, + 533, + 482, + 666 + ], + "type": "text", + "content": "In this work, we present an efficient method that planning with directed fidelity trajectory, explicitly leveraging the alignment between instructions and trajectories for navigation. Additionally, we consider the directional nature of navigation and introduce a directed graph construction during navigation, storing vision information on directed edges. This approach provides rich directed path representation and enhances instruction-trajectory alignment. Experiments demonstrate that our method is achieves strong performance while being significantly more efficient than previous methods. Our method has certain limitations that we need to address. For example, navigation requires environment layout information in some cases. We recognize the need to investigate potential solutions to overcome these challenges." + } + ] + } + ], + "index": 20 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "spans": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "type": "text", + "content": "14" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 166, + 91, + 202, + 100 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 166, + 91, + 202, + 100 + ], + "spans": [ + { + "bbox": [ + 166, + 91, + 202, + 100 + ], + "type": "text", + "content": "Lu et al." + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 13 + }, + { + "para_blocks": [ + { + "bbox": [ + 133, + 114, + 240, + 129 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 114, + 240, + 129 + ], + "spans": [ + { + "bbox": [ + 133, + 114, + 240, + 129 + ], + "type": "text", + "content": "Acknowledgments" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 130, + 140, + 482, + 189 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 140, + 482, + 189 + ], + "spans": [ + { + "bbox": [ + 130, + 140, + 482, + 189 + ], + "type": "text", + "content": "This work was supported partially by the National Key Research and Development Program of China (2023YFA1008503), NSFC(U21A20471, 62206315), Guangdong NSF Project (No. 2023B1515040025, 2020B1515120085, 2024A1515-010101), Guangzhou Basic and Applied Basic Research Scheme(2024A04J4067)." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 132, + 205, + 197, + 217 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 205, + 197, + 217 + ], + "spans": [ + { + "bbox": [ + 132, + 205, + 197, + 217 + ], + "type": "text", + "content": "References" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 134, + 228, + 481, + 665 + ], + "type": "list", + "angle": 0, + "index": 22, + "blocks": [ + { + "bbox": [ + 138, + 228, + 480, + 251 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 228, + 480, + 251 + ], + "spans": [ + { + "bbox": [ + 138, + 228, + 480, + 251 + ], + "type": "text", + "content": "1. An, D., Qi, Y., Huang, Y., Wu, Q., Wang, L., Tan, T.: Neighbor-view enhanced model for vision and language navigation. In: ACM MM (2021)" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 138, + 251, + 480, + 273 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 251, + 480, + 273 + ], + "spans": [ + { + "bbox": [ + 138, + 251, + 480, + 273 + ], + "type": "text", + "content": "2. An, D., Qi, Y., Li, Y., Huang, Y., Wang, L., Tan, T., Shao, J.: Bevbert: Multimodal map pre-training for language-guided navigation. In: ICCV (2023)" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 138, + 274, + 481, + 305 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 274, + 481, + 305 + ], + "spans": [ + { + "bbox": [ + 138, + 274, + 481, + 305 + ], + "type": "text", + "content": "3. Anderson, P., Chang, A.X., Chaplot, D.S., Dosovitskiy, A., Gupta, S., Koltun, V., Kosecka, J., Malik, J., Mottaghi, R., Savva, M., Zamir, A.R.: On evaluation of embodied navigation agents. CoRR (2018)" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 138, + 306, + 480, + 338 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 306, + 480, + 338 + ], + "spans": [ + { + "bbox": [ + 138, + 306, + 480, + 338 + ], + "type": "text", + "content": "4. Anderson, P., Wu, Q., Teney, D., Bruce, J., Johnson, M., Sünderhauf, N., Reid, I., Gould, S., Van Den Hengel, A.: Vision-and-language navigation: Interpreting visually-grounded navigation instructions in real environments. In: CVPR (2018)" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 138, + 339, + 480, + 360 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 339, + 480, + 360 + ], + "spans": [ + { + "bbox": [ + 138, + 339, + 480, + 360 + ], + "type": "text", + "content": "5. Chaplot, D.S., Gandhi, D., Gupta, A., Salakhutdinov, R.: Object goal navigation using goal-oriented semantic exploration. In: NeurIPS (2020)" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 138, + 361, + 480, + 381 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 361, + 480, + 381 + ], + "spans": [ + { + "bbox": [ + 138, + 361, + 480, + 381 + ], + "type": "text", + "content": "6. Chaplot, D.S., Gandhi, D., Gupta, S., Gupta, A., Salakhutdinov, R.: Learning to explore using active neural SLAM. In: ICLR (2020)" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 138, + 383, + 480, + 403 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 383, + 480, + 403 + ], + "spans": [ + { + "bbox": [ + 138, + 383, + 480, + 403 + ], + "type": "text", + "content": "7. Chen, S., Guhur, P.L., Schmid, C., Laptev, I.: History aware multimodal transformer for vision-and-language navigation. NeurIPS (2021)" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 138, + 404, + 480, + 425 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 404, + 480, + 425 + ], + "spans": [ + { + "bbox": [ + 138, + 404, + 480, + 425 + ], + "type": "text", + "content": "8. Chen, S., Guhur, P., Tapaswi, M., Schmid, C., Laptev, I.: Think global, act local: Dual-scale graph transformer for vision-and-language navigation. In: CVPR (2022)" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 138, + 426, + 480, + 447 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 426, + 480, + 447 + ], + "spans": [ + { + "bbox": [ + 138, + 426, + 480, + 447 + ], + "type": "text", + "content": "9. Chen, Y.C., Li, L., Yu, L., El Kholy, A., Ahmed, F., Gan, Z., Cheng, Y., Liu, J.: Uniter: Universal image-text representation learning. In: ECCV (2020)" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 138, + 449, + 480, + 479 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 449, + 480, + 479 + ], + "spans": [ + { + "bbox": [ + 138, + 449, + 480, + 479 + ], + "type": "text", + "content": "0. Conneau, A., Khandelwal, K., Goyal, N., Chaudhary, V., Wenzek, G., Guzmán, F., Grave, E., Ott, M., Zettlemoyer, L., Stoyanov, V.: Unsupervised cross-lingual representation learning at scale. In: ACL (2020)" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 138, + 480, + 480, + 502 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 480, + 480, + 502 + ], + "spans": [ + { + "bbox": [ + 138, + 480, + 480, + 502 + ], + "type": "text", + "content": "1. Devlin, J., Chang, M.W., Lee, K., Toutanova, K.: BERT: Pre-training of deep bidirectional transformers for language understanding. In: NAACL (2019)" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 134, + 502, + 480, + 545 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 134, + 502, + 480, + 545 + ], + "spans": [ + { + "bbox": [ + 134, + 502, + 480, + 545 + ], + "type": "text", + "content": "12. Dosovitskiy, A., Beyer, L., Kolesnikov, A., Weissenborn, D., Zhai, X., Unterthiner, T., Dehghani, M., Minderer, M., Heigold, G., Gelly, S., Uszkoreit, J., Houlsby, N.: An image is worth 16x16 words: Transformers for image recognition at scale. In: ICLR (2021)" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 134, + 546, + 480, + 567 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 134, + 546, + 480, + 567 + ], + "spans": [ + { + "bbox": [ + 134, + 546, + 480, + 567 + ], + "type": "text", + "content": "13. Dou, Z., Gao, F., Peng, N.: Masked path modeling for vision-and-language navigation. CoRR (2023)" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 134, + 567, + 480, + 600 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 134, + 567, + 480, + 600 + ], + "spans": [ + { + "bbox": [ + 134, + 567, + 480, + 600 + ], + "type": "text", + "content": "14. Fried, D., Hu, R., Cirik, V., Rohrbach, A., Andreas, J., Morency, L.P., Berg-Kirkpatrick, T., Saenko, K., Klein, D., Darrell, T.: Speaker-follower models for vision-and-language navigation. In: NeurIPS (2018)" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 134, + 601, + 480, + 621 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 134, + 601, + 480, + 621 + ], + "spans": [ + { + "bbox": [ + 134, + 601, + 480, + 621 + ], + "type": "text", + "content": "15. Fuentes-Pacheco, J., Ascencio, J.R., Rendon-Mancha, J.M.: Visual simultaneous localization and mapping: a survey. Artif. Intell. Rev. (2015)" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 134, + 622, + 480, + 643 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 134, + 622, + 480, + 643 + ], + "spans": [ + { + "bbox": [ + 134, + 622, + 480, + 643 + ], + "type": "text", + "content": "16. Gao, C., Peng, X., Yan, M., Wang, H., Yang, L., Ren, H., Li, H., Liu, S.: Adaptive zone-aware hierarchical planner for vision-language navigation. In: CVPR (2023)" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 134, + 644, + 480, + 665 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 134, + 644, + 480, + 665 + ], + "spans": [ + { + "bbox": [ + 134, + 644, + 480, + 665 + ], + "type": "text", + "content": "17. Hao, W., Li, C., Li, X., Carin, L., Gao, J.: Towards learning a generic agent for vision-and-language navigation via pre-training. In: CVPR (2020)" + } + ] + } + ], + "index": 21 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 419, + 91, + 447, + 100 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 419, + 91, + 447, + 100 + ], + "spans": [ + { + "bbox": [ + 419, + 91, + 447, + 100 + ], + "type": "text", + "content": "PRET" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 470, + 91, + 481, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 470, + 91, + 481, + 100 + ], + "spans": [ + { + "bbox": [ + 470, + 91, + 481, + 100 + ], + "type": "text", + "content": "15" + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 14 + }, + { + "para_blocks": [ + { + "bbox": [ + 132, + 117, + 480, + 665 + ], + "type": "list", + "angle": 0, + "index": 21, + "blocks": [ + { + "bbox": [ + 133, + 117, + 480, + 138 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 117, + 480, + 138 + ], + "spans": [ + { + "bbox": [ + 133, + 117, + 480, + 138 + ], + "type": "text", + "content": "18. Hong, Y., Wu, Q., Qi, Y., Rodriguez-Opazo, C., Gould, S.: Vln bert: A recurrent vision-and-language bert for navigation. In: CVPR. pp. 1643-1653 (2021)" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 133, + 140, + 480, + 171 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 140, + 480, + 171 + ], + "spans": [ + { + "bbox": [ + 133, + 140, + 480, + 171 + ], + "type": "text", + "content": "19. Huo, J., Sun, Q., Jiang, B., Lin, H., Fu, Y.: Geovln: Learning geometry-enhanced visual representation with slot attention for vision-and-language navigation. In: CVPR (2023)" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 132, + 173, + 480, + 205 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 173, + 480, + 205 + ], + "spans": [ + { + "bbox": [ + 132, + 173, + 480, + 205 + ], + "type": "text", + "content": "20. Hwang, M., Jeong, J., Kim, M., Oh, Y., Oh, S.: Meta-exlore: Exploratory hierarchical vision-and-language navigation using scene object spectrum grounding. In: CVPR (2023)" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 132, + 206, + 480, + 228 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 206, + 480, + 228 + ], + "spans": [ + { + "bbox": [ + 132, + 206, + 480, + 228 + ], + "type": "text", + "content": "21. Ilharco, G., Jain, V., Ku, A., Ie, E., Baldridge, J.: General evaluation for instruction conditioned navigation using dynamic time warping. In: NeurIPS Workshop (2019)" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 132, + 229, + 480, + 251 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 229, + 480, + 251 + ], + "spans": [ + { + "bbox": [ + 132, + 229, + 480, + 251 + ], + "type": "text", + "content": "22. Kim, W., Son, B., Kim, I.: Vilt: Vision-and-language transformer without convolution or region supervision. In: ICML (2021)" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 132, + 251, + 480, + 283 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 251, + 480, + 283 + ], + "spans": [ + { + "bbox": [ + 132, + 251, + 480, + 283 + ], + "type": "text", + "content": "23. Ku, A., Anderson, P., Patel, R., Ie, E., Baldridge, J.: Room-across-room: Multilingual vision-and-language navigation with dense spatiotemporal grounding. In: EMNLP (2020)" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 132, + 285, + 480, + 307 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 285, + 480, + 307 + ], + "spans": [ + { + "bbox": [ + 132, + 285, + 480, + 307 + ], + "type": "text", + "content": "24. Li, J., Tan, H., Bansal, M.: Envedit: Environment editing for vision-and-language navigation. In: CVPR (2022)" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 132, + 308, + 480, + 340 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 308, + 480, + 340 + ], + "spans": [ + { + "bbox": [ + 132, + 308, + 480, + 340 + ], + "type": "text", + "content": "25. Li, J., Selvaraju, R., Gotmare, A., Joty, S., Xiong, C., Hoi, S.C.H.: Align before fuse: Vision and language representation learning with momentum distillation. NeurIPS (2021)" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 132, + 341, + 480, + 373 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 341, + 480, + 373 + ], + "spans": [ + { + "bbox": [ + 132, + 341, + 480, + 373 + ], + "type": "text", + "content": "26. Li, X., Li, C., Xia, Q., Bisk, Y., Celikylmaz, A., Gao, J., Smith, N.A., Choi, Y.: Robust navigation with language pretraining and stochastic sampling. In: EMNLP-IJCNLP (2019)" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 132, + 375, + 480, + 407 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 375, + 480, + 407 + ], + "spans": [ + { + "bbox": [ + 132, + 375, + 480, + 407 + ], + "type": "text", + "content": "27. Li, X., Li, C., Xia, Q., Bisk, Y., Celikyilmaz, A., Gao, J., Smith, N.A., Choi, Y.: Robust navigation with language pretraining and stochastic sampling. In: EMNLP-IJCNLP (2019)" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 132, + 408, + 480, + 430 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 408, + 480, + 430 + ], + "spans": [ + { + "bbox": [ + 132, + 408, + 480, + 430 + ], + "type": "text", + "content": "28. Lin, C., Jiang, Y., Cai, J., Qu, L., Haffari, G., Yuan, Z.: Multimodal transformer with variable-length memory for vision-and-language navigation. In: ECCV (2022)" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 132, + 431, + 480, + 453 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 431, + 480, + 453 + ], + "spans": [ + { + "bbox": [ + 132, + 431, + 480, + 453 + ], + "type": "text", + "content": "29. Liu, C., Zhu, F., Chang, X., Liang, X., Ge, Z., Shen, Y.: Vision-language navigation with random environmental mixup. In: ICCV (2021)" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 132, + 453, + 480, + 464 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 453, + 480, + 464 + ], + "spans": [ + { + "bbox": [ + 132, + 453, + 480, + 464 + ], + "type": "text", + "content": "30. Loshchilov, I., Hutter, F.: Decoupled weight decay regularization. In: ICLR (2019)" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 132, + 465, + 480, + 487 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 465, + 480, + 487 + ], + "spans": [ + { + "bbox": [ + 132, + 465, + 480, + 487 + ], + "type": "text", + "content": "31. Ma, C.Y., Lu, J., Wu, Z., AlRegib, G., Kira, Z., Socher, R., Xiong, C.: Self-monitoring navigation agent via auxiliary progress estimation. In: ICLR (2019)" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 132, + 487, + 480, + 509 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 487, + 480, + 509 + ], + "spans": [ + { + "bbox": [ + 132, + 487, + 480, + 509 + ], + "type": "text", + "content": "32. Ma, C., Wu, Z., AlRegib, G., Xiong, C., Kira, Z.: The regretful agent: Heuristic-aided navigation through progress estimation. In: CVPR (2019)" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 132, + 510, + 480, + 542 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 510, + 480, + 542 + ], + "spans": [ + { + "bbox": [ + 132, + 510, + 480, + 542 + ], + "type": "text", + "content": "33. Majumdar, A., Shrivastava, A., Lee, S., Anderson, P., Parikh, D., Batra, D.: Improving vision-and-language navigation with image-text pairs from the web. In: ECCV (2020)" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 132, + 544, + 480, + 597 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 544, + 480, + 597 + ], + "spans": [ + { + "bbox": [ + 132, + 544, + 480, + 597 + ], + "type": "text", + "content": "34. Oquab, M., Darcet, T., Moutakanni, T., Vo, H., Szafraniec, M., Khalidov, V., Fernandez, P., Haziza, D., Massa, F., El-Nouby, A., Assran, M., Ballas, N., Galuba, W., Howes, R., Huang, P., Li, S., Misra, I., Rabbat, M.G., Sharma, V., Synnaeve, G., Xu, H., Jegou, H., Mairal, J., Labatut, P., Joulin, A., Bojanowski, P.: Dinov2: Learning robust visual features without supervision. CoRR (2023)" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 132, + 599, + 480, + 631 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 599, + 480, + 631 + ], + "spans": [ + { + "bbox": [ + 132, + 599, + 480, + 631 + ], + "type": "text", + "content": "35. Pope, R., Douglas, S., Chowdhery, A., Devlin, J., Bradbury, J., Levskaya, A., Heek, J., Xiao, K., Agrawal, S., Dean, J.: Efficiently scaling transformer inference. CoRR (2022)" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 132, + 632, + 480, + 665 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 632, + 480, + 665 + ], + "spans": [ + { + "bbox": [ + 132, + 632, + 480, + 665 + ], + "type": "text", + "content": "36. Qi, Y., Wu, Q., Anderson, P., Wang, X., Wang, W.Y., Shen, C., Hengel, A.v.d.: Reverie: Remote embodied visual referring expression in real indoor environments. In: CVPR (2020)" + } + ] + } + ], + "index": 20 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "spans": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "type": "text", + "content": "16" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 166, + 91, + 202, + 100 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 166, + 91, + 202, + 100 + ], + "spans": [ + { + "bbox": [ + 166, + 91, + 202, + 100 + ], + "type": "text", + "content": "Lu et al." + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 15 + }, + { + "para_blocks": [ + { + "bbox": [ + 130, + 116, + 482, + 414 + ], + "type": "list", + "angle": 0, + "index": 14, + "blocks": [ + { + "bbox": [ + 130, + 116, + 482, + 149 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 116, + 482, + 149 + ], + "spans": [ + { + "bbox": [ + 130, + 116, + 482, + 149 + ], + "type": "text", + "content": "37. Radford, A., Kim, J.W., Hallacy, C., Ramesh, A., Goh, G., Agarwal, S., Sastry, G., Askell, A., Mishkin, P., Clark, J., et al.: Learning transferable visual models from natural language supervision. In: ICML (2021)" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 130, + 150, + 481, + 172 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 150, + 481, + 172 + ], + "spans": [ + { + "bbox": [ + 130, + 150, + 481, + 172 + ], + "type": "text", + "content": "38. Sennrich, R., Haddow, B., Birch, A.: Improving neural machine translation models with monolingual data. In: ACL (2016)" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 132, + 172, + 481, + 194 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 172, + 481, + 194 + ], + "spans": [ + { + "bbox": [ + 132, + 172, + 481, + 194 + ], + "type": "text", + "content": "39. Shen, S., Li, L.H., Tan, H., Bansal, M., Rohrbach, A., Chang, K., Yao, Z., Keutzer, K.: How much can CLIP benefit vision-and-language tasks? In: ICLR (2022)" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 132, + 194, + 481, + 216 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 194, + 481, + 216 + ], + "spans": [ + { + "bbox": [ + 132, + 194, + 481, + 216 + ], + "type": "text", + "content": "40. Tan, H., Yu, L., Bansal, M.: Learning to navigate unseen environments: Back translation with environmental dropout. In: NAACL (2019)" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 132, + 216, + 481, + 237 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 216, + 481, + 237 + ], + "spans": [ + { + "bbox": [ + 132, + 216, + 481, + 237 + ], + "type": "text", + "content": "41. Thomason, J., Murray, M., Cakmak, M., Zettlemoyer, L.: Vision-and-dialog navigation. In: CoRL (2020)" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 132, + 238, + 481, + 259 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 238, + 481, + 259 + ], + "spans": [ + { + "bbox": [ + 132, + 238, + 481, + 259 + ], + "type": "text", + "content": "42. Thrun, S.: Learning metric-topological maps for indoor mobile robot navigation. Artif. Intell. (1998)" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 132, + 259, + 481, + 281 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 259, + 481, + 281 + ], + "spans": [ + { + "bbox": [ + 132, + 259, + 481, + 281 + ], + "type": "text", + "content": "43. Vaswani, A., Shazeer, N., Parmar, N., Uszkoreit, J., Jones, L., Gomez, A.N., Kaiser, L.u., Polosukhin, I.: Attention is all you need. In: NeurIPS (2017)" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 132, + 282, + 481, + 304 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 282, + 481, + 304 + ], + "spans": [ + { + "bbox": [ + 132, + 282, + 481, + 304 + ], + "type": "text", + "content": "44. Wang, H., Wang, W., Liang, W., Xiong, C., Shen, J.: Structured scene memory for vision-language navigation. In: CVPR (2021)" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 132, + 304, + 481, + 335 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 304, + 481, + 335 + ], + "spans": [ + { + "bbox": [ + 132, + 304, + 481, + 335 + ], + "type": "text", + "content": "45. Wang, S., Montgomery, C., Orbay, J., Birodkar, V., Faust, A., Gur, I., Jaques, N., Waters, A., Baldridge, J., Anderson, P.: Less is more: Generating grounded navigation instructions from landmarks. In: CVPR (2022)" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 132, + 336, + 481, + 369 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 336, + 481, + 369 + ], + "spans": [ + { + "bbox": [ + 132, + 336, + 481, + 369 + ], + "type": "text", + "content": "46. Wang, X., Huang, Q., Celikyilmaz, A., Gao, J., Shen, D., Wang, Y.F., Wang, W.Y., Zhang, L.: Reinforced cross-modal matching and self-supervised imitation learning for vision-language navigation. In: CVPR (2019)" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 132, + 370, + 481, + 392 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 370, + 481, + 392 + ], + "spans": [ + { + "bbox": [ + 132, + 370, + 481, + 392 + ], + "type": "text", + "content": "47. Wang, Z., Li, X., Yang, J., Liu, Y., Jiang, S.: Gridmm: Grid memory map for vision-and-language navigation. In: ICCV (2023)" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 132, + 392, + 481, + 414 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 392, + 481, + 414 + ], + "spans": [ + { + "bbox": [ + 132, + 392, + 481, + 414 + ], + "type": "text", + "content": "48. Zhu, F., Zhu, Y., Chang, X., Liang, X.: Vision-language navigation with self-supervised auxiliary reasoning tasks. In: CVPR (2020)" + } + ] + } + ], + "index": 13 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 420, + 91, + 447, + 100 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 420, + 91, + 447, + 100 + ], + "spans": [ + { + "bbox": [ + 420, + 91, + 447, + 100 + ], + "type": "text", + "content": "PRET" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 470, + 91, + 481, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 470, + 91, + 481, + 100 + ], + "spans": [ + { + "bbox": [ + 470, + 91, + 481, + 100 + ], + "type": "text", + "content": "17" + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 16 + } + ], + "_backend": "vlm", + "_version_name": "2.6.4" +} \ No newline at end of file diff --git a/2024/PSALM_ Pixelwise Segmentation with Large Multi-modal Model/6e437084-134d-4cf9-a26b-095cfa791f83_content_list.json b/2024/PSALM_ Pixelwise Segmentation with Large Multi-modal Model/6e437084-134d-4cf9-a26b-095cfa791f83_content_list.json new file mode 100644 index 0000000000000000000000000000000000000000..4a851f262eb26482ef6fff88c0b61c85c812958c --- /dev/null +++ b/2024/PSALM_ Pixelwise Segmentation with Large Multi-modal Model/6e437084-134d-4cf9-a26b-095cfa791f83_content_list.json @@ -0,0 +1,1640 @@ +[ + { + "type": "text", + "text": "PSALM: Pixelwise SegmentAtion with Large Multi-Modal Model", + "text_level": 1, + "bbox": [ + 243, + 140, + 759, + 186 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Zheng Zhang $^{1,2*}$ , Yeyao Ma $^{1*}$ , Enming Zhang $^{1*}$ , and Xiang Bai $^{1}$", + "bbox": [ + 259, + 208, + 743, + 228 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "1 Huazhong University of Science and Technology \n2 Microsoft Research Asia", + "bbox": [ + 331, + 238, + 669, + 267 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Abstract. PSALM is a powerful extension of the Large Multi-modal Model (LMM) to address the segmentation task challenges. To overcome the limitation of the LMM being limited to textual output, PSALM incorporates a mask decoder and a well-designed input schema to handle a variety of segmentation tasks. This schema includes images, task instructions, conditional prompts, and mask tokens, which enable the model to generate and classify segmentation masks effectively. The flexible design of PSALM supports joint training across multiple datasets and tasks, leading to improved performance and task generalization. PSALM achieves superior results on several benchmarks, such as RefCOCO/RefCOCO+/RefCOCOg, COCO Panoptic Segmentation, and COCO-Interactive, and further exhibits zero-shot capabilities on unseen tasks, such as open-vocabulary segmentation, generalized referring expression segmentation and video object segmentation, making a significant step towards a GPT moment in computer vision. Through extensive experiments, PSALM demonstrates its potential to transform the domain of image segmentation, leveraging the robust visual understanding capabilities of LMMs as seen in natural language processing. Code and models are available at https://github.com/zamling/PSALM.", + "bbox": [ + 259, + 300, + 743, + 565 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Keywords: Segmentation $\\cdot$ Large Multimodal Model $\\cdot$ Visual Language", + "bbox": [ + 261, + 575, + 740, + 592 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "1 Introduction", + "text_level": 1, + "bbox": [ + 215, + 635, + 377, + 651 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Large multi-modal model (LMM) has ignited the dawn of the vision GPT [2] moment by making ground-breaking progress in various advanced visual understanding tasks by compressing image and language information into a single auto-regressive model. However, there are still many obstacles on the road to achieving vision GPT, and an important one being that current LMM can only perform text outputs, making it challenging to address the pixel-level image understanding problem directly, i.e., image segmentation, which is one of the most critical tasks in computer vision.", + "bbox": [ + 212, + 665, + 787, + 785 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Behind the obstacle is many challenges. First, the default output of the LMM is discrete tokens, and there is no apparent way to generate masks directly.", + "bbox": [ + 212, + 786, + 787, + 818 + ], + "page_idx": 0 + }, + { + "type": "page_footnote", + "text": "* Equal contribution. ${}^{ \\boxtimes }$ Corresponding author.", + "bbox": [ + 230, + 824, + 547, + 840 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/23d5814cdcbed14d1acaa71e7a63e13276e2b67be2a3ad5a5679354e03c77d9a.jpg", + "image_caption": [ + "Fig. 1: PSALM has capability to handle multiple segmentation tasks in only one single model. We visualize some tasks, including Panoptic segmentation in COCO [24]; Open-Vocabulary instance segmentation in ADE20K [59]; Interactive segmentation in COCO-Interactive; Referring segmentation in RefCOCO [57]; Generalized referring segmentation in gRefCOCO [25]; Ego-exo correspondence in Ego-Exo4d [12]; Video object segmentation in DAVIS2017 [36]." + ], + "image_footnote": [], + "bbox": [ + 248, + 145, + 416, + 351 + ], + "page_idx": 1 + }, + { + "type": "image", + "img_path": "images/83b9485f68a18c6a8252495eb4d5c0671a7c640098271a829cf37648d10c9078.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 421, + 145, + 586, + 351 + ], + "page_idx": 1 + }, + { + "type": "image", + "img_path": "images/a3f75599d0a30bf7e3f9c386a776707c0a0f38ecf1968a487738425cde68973c.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 589, + 146, + 754, + 351 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Second, the variety of image segmentation tasks requires different forms of inputs and outputs. For example, semantic segmentation needs support inputs with different categories. Instance segmentation demands generate object IDs and the class confidence score for each object. Referring segmentation is supposed to have a language sentence as input and interactive segmentation has more varied inputs than the other tasks, which can be points, scribbles, bounding boxes, or masks. Third, unifying different segmentation tasks with a shared weight model is also challenging because different tasks require varied capabilities.", + "bbox": [ + 212, + 477, + 787, + 598 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "In this work, we propose a method named PSALM (Pixelwise SegmentAtion with Large Multi-Modal Model) that aims to address the above challenges and extend the capabilities of LMM from text-output tasks to general segmentation tasks (Fig. 1 shows representative tasks). Specifically, PSALM externalizes a mask decoder on the top of LMM and designs a flexible input schema to unify different segmentation tasks into a single model.", + "bbox": [ + 212, + 598, + 787, + 688 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "The input schema consists of four different parts: images, task instruction prompt, condition prompt, and a set of mask tokens, where the instruction prompt is a text sentence describing the task itself, condition prompt contains the additional necessary information to solve the task, either in terms of category names, sentence or visual features, and mask tokens are a set of learnable embeddings. All these inputs are fed into the LMM, and the resulting output mask tokens are further used as input by the mask generator to present mask proposals. Apart from producing the mask proposals, it is also necessary to predict the class of each segmentation mask or estimate a confidence score, which can be achieved by using the output embedding of the condition prompt as the", + "bbox": [ + 212, + 689, + 787, + 840 + ], + "page_idx": 1 + }, + { + "type": "page_number", + "text": "2", + "bbox": [ + 217, + 114, + 228, + 126 + ], + "page_idx": 1 + }, + { + "type": "header", + "text": "Z. Zhang, Y. Ma, E. Zhang, X. Bai", + "bbox": [ + 269, + 114, + 506, + 128 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "classifier weights to classify each mask proposal. In practice, we categorize conditions into category condition, sentence condition, and visual-prior condition, and present the corresponding methods to build the classifier weights according to the properties of each type of condition.", + "bbox": [ + 212, + 146, + 782, + 205 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Some other methods, represented by LISA [17], also aim to use LMM for segmentation tasks. However, these methods are usually designed for referring segmentation and fail to justify their ability to solve generalized segmentation tasks (see Tab. 1). In contrast, thanks to the generality and flexibility of the proposed architecture, PSALM can not only solve a variety of segmentation tasks but also be able to joint train on different tasks, which makes the model task generalizable while allowing the model to take full advantage of the intrinsic connections of different datasets/tasks to achieve better performance. Specifically, with the joint training of COCO Panoptic Segmentation [24], RefCOCO [57]/RefCOCO+/RefCOCOg [31], and COCO Interactive, we observe a significant performance improvement compared to training at different tasks individually, and therefore result in even better performance than other task-specific methods. On referring segmentation tasks, we outperform other LLM-based pixel reasoning methods (e.g., LISA, PixelLM [40] and GSVA [48]) on RefCOCO, RefCOCO+, and RefCOCOg, and it worth noting that we only use Phi-1.5 1.3B model [22] while others adopt Vicuna-7B [7] or LLama2-13B model [45].", + "bbox": [ + 212, + 207, + 787, + 448 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "The flexible design of architecture and input schema, multi-task joint-training, and the strong visual understanding capability of LMM not only make PSALM perform well on trained in-domain tasks but also enable generalizability to out-of-domain tasks in a zero-shot manner, i.e., directly dealing with unseen tasks without additional training. We test on three tasks: open-vocabulary segmentation, generalized referring expression segmentation, and video object segmentation. PSALM achieves promising zero-shot performance on these tasks. We believe this task-level generalizability is crucial, which is one of the key properties of the large language model for its success in natural language processing.", + "bbox": [ + 212, + 449, + 787, + 584 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Through extensive experiments on a variety of segmentation tasks, we show that presented PSALM has strong potential to address general image segmentation tasks and exhibits a certain degree of task generalization capability as LLM does in NLP. We believe that this work can inspire extensive research on realizing the GPT moment in computer vision and facilitate its arrival.", + "bbox": [ + 212, + 584, + 787, + 659 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "2 Related works", + "text_level": 1, + "bbox": [ + 214, + 681, + 392, + 696 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "2.1 Large Multimodal Models", + "text_level": 1, + "bbox": [ + 214, + 710, + 478, + 724 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "With the release of GPT-V [32] and Gemini [44], more attention and efforts from open-source and research communities are shifting from large language models (LLM) to large multi-modal models (LMM). LLaVA [27], BLIP-2 [19], and Flamingo [1] are three representative works, where the core idea of both LLaVA and BLIP-2 is to map visual features into the input space of LLM to implement multi-modal capabilities, while Flamingo employs deeper feature fusion in the intermediate layers of LLM. Some works, such as Kosmos-2 [33],", + "bbox": [ + 212, + 733, + 787, + 840 + ], + "page_idx": 2 + }, + { + "type": "header", + "text": "PSALM: Pixelwise SegmentAtion with Large Multi-Modal Model", + "bbox": [ + 295, + 114, + 732, + 128 + ], + "page_idx": 2 + }, + { + "type": "page_number", + "text": "3", + "bbox": [ + 774, + 116, + 784, + 126 + ], + "page_idx": 2 + }, + { + "type": "table", + "img_path": "images/571443d84933cd8088259de6638780403b777685331013ce558ef44268098b40.jpg", + "table_caption": [ + "Table 1: Capability for different methods. Our proposed PSALM can handle more segmentation tasks than other LLM-centric methods. LLM-centric methods can also deal with text generation tasks, which is hard for most vision-centric methods." + ], + "table_footnote": [], + "table_body": "
MethodsGeneric Seg.Referring Seg.Interactive Seg.OV Seg.
Vision centricMask2Former [6]✓
ODISE [49]✓✓
UNINEXT [53]✓✓✓
SEEM [61]✓✓✓✓
OMG-Seg [21]✓✓✓
LLM centricLISA [17]✓
GLAMM [39]✓
PixelLM [40]✓
PSALM (Ours)✓✓✓✓
", + "bbox": [ + 267, + 198, + 733, + 342 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Shikra [4], and Ferret [55], further introduce object localization tasks into the LMM, while others, such as Emu [43], CogVLM [46], and DreamLLM [10], focus on how to integrate visual generation into the LMM. Monkey [23], OtterHD [18], and LLaVA-NeXT [26] explores mechanisms for processing large-size images and significantly improves performance on tasks such as OCR. However, these above methods are mainly designed for text output tasks or image generation and cannot directly deal with pixel-level understanding tasks such as image segmentation, which is different from ours and we can base on these models.", + "bbox": [ + 212, + 373, + 787, + 494 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "2.2 Pixel Reasoning with LMM", + "text_level": 1, + "bbox": [ + 215, + 527, + 491, + 544 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Similar to our goal, some existing works attempt to enable LMMs to generate segmentation masks. LISA [17] is a pioneering work that uses a special seg token to aggregate information of a given sentence and use it as a prompt embedding in a SAM decoder to predict the segmentation mask. u-LLaVA [50] further supports object grounding tasks on the basis of LISA, and NExT-Chat [58] introduces richer inputs, such as bounding boxes. Furthermore, since LISA can only deal with a single object, many subsequent works that attempt to address the multi-object case, such as GLaMM [39], PerceptionGPT [34], PixelLM [40], GSVA [48], and LISA++ [54], all of which share the basic idea of introducing a seg token for each sentence describing a different object, and except PixelLM, all other methods are based on SAM decoder [15].", + "bbox": [ + 212, + 565, + 787, + 733 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Although all these methods can generate masks, they are primarily designed for reference segmentation. In contrast, our method is designed for generalized segmentation tasks, which have diverse input and output requirements. In addition, the difference in goals also brings technical discrepancies: these methods use language models to directly generate the final segmentation masks, while our approach is closer to Mask2Former [6], which first generates mask proposals and then classifies the masks.", + "bbox": [ + 212, + 734, + 787, + 839 + ], + "page_idx": 3 + }, + { + "type": "page_number", + "text": "4", + "bbox": [ + 217, + 114, + 228, + 126 + ], + "page_idx": 3 + }, + { + "type": "header", + "text": "Z. Zhang, Y. Ma, E. Zhang, X. Bai", + "bbox": [ + 271, + 114, + 506, + 128 + ], + "page_idx": 3 + }, + { + "type": "image", + "img_path": "images/8fb1ade5589cb18ace88fbb6c48b90fd4d69c3bf5d21a24f69a3cfe6798dbd51.jpg", + "image_caption": [ + "Fig. 2: PSALM architecture overview." + ], + "image_footnote": [], + "bbox": [ + 243, + 143, + 759, + 287 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "2.3 Unified Segmentation Framework", + "text_level": 1, + "bbox": [ + 215, + 340, + 537, + 356 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Another class of relevant explorations [21, 47, 53, 61] studies how to unify different segmentation tasks into a single framework. For instance, Mask2Former [6] presents a unified architecture of generic segmentation $^3$ but with different models. OneFormer [14] further integrates these three tasks within a single model. UNINEXT [53] aims to unify instance-perception tasks and introduce text inputs and thus it can handle referring segmentation. X-Decoder [60] presents a flexible decoder architecture that can support generic segmentation, referring segmentation, retrieval, and image captioning. SEEM [61] designs a generic encoder-decoder to unify different segmentation tasks, where the encoder is used for projecting image, text, and human inputs into a joint visual-semantic space, and the decoder is used for mask prediction. However, all these works are not based on LMM, instead, they are mostly vision-centric models, i.e., usually designed for visual tasks only and thus cannot address language tasks very well.", + "bbox": [ + 212, + 366, + 787, + 563 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "3 Methods", + "text_level": 1, + "bbox": [ + 215, + 585, + 339, + 602 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Fig. 2 provides an overview schematic of PSALM, which consists of a large multimodal model (LMM), a mask generator, and a flexible input schema designed for general segmentation tasks. The input schema has four different types of inputs: image, task instruction prompt, condition prompt, and a set of mask tokens. LMM processes the input tokens and the output embedding of mask tokens is further fed into the mask generator to generate masks. In the following, we will introduce our approach in detail.", + "bbox": [ + 212, + 618, + 787, + 724 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "3.1 Large Multimodal Model and Input Schema", + "text_level": 1, + "bbox": [ + 215, + 746, + 627, + 762 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "PSALM is built on large multimodal models (LMM), and there are many different LMM architectures, such as LLaVA [27], BLIP [20], and Flamingo [1]. Here,", + "bbox": [ + 212, + 771, + 785, + 801 + ], + "page_idx": 4 + }, + { + "type": "header", + "text": "PSALM: Pixelwise SegmentAtion with Large Multi-Modal Model", + "bbox": [ + 295, + 114, + 732, + 128 + ], + "page_idx": 4 + }, + { + "type": "page_number", + "text": "5", + "bbox": [ + 774, + 116, + 784, + 126 + ], + "page_idx": 4 + }, + { + "type": "page_footnote", + "text": "3 Generic segmentation includes semantic segmentation, instance segmentation, and panoptic segmentation.", + "bbox": [ + 217, + 810, + 785, + 839 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "we adopt the design of LLaVA because of its proven performance and simplicity, but the other LMM architectures are also compatible with our approach without any theoretical difficulties.", + "bbox": [ + 212, + 146, + 782, + 191 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "The LMM used in our work has a visual encoder and pre-trained large language model (LLM). The two models are connected by a lightweight vision-language alignment model, which is a $3 \\times 3$ convolution layer followed by a linear layer. The official LLaVA model uses a frozen CLIP model [37] as a visual coder, whose features lack the fine-grained information that is required for segmentation tasks [52]. Therefore, we train a customized LLaVA model by using the Swin Transformer [28], and limited by resources, we additionally replace the LLM from the Vicuna 7B model [7] to a smaller Phi-1.5 1.3B model [22]. Here, we applied only the first visual-language alignment stage of LLaVA by following its default settings. In our ablations, we found the alignment stage is crucial for open-vocabulary segmentation and referring segmentation tasks (Tab. 6).", + "bbox": [ + 212, + 193, + 784, + 359 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Different segmentation tasks need different forms of inputs and outputs, which motivates us to present a flexible input schema to unify various requirements. In addition to the input image used in the visual encoder, our input schema has three other different types of inputs: task instruction prompt, condition prompt, and a set of mask tokens. We will introduce each of them and summarize the prompts used for all different tasks in the Appendix.", + "bbox": [ + 212, + 362, + 784, + 452 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Task Instruction Prompt. The task instruction prompt is usually a text sentence describing and specifying the model's task. For example, in panoptic segmentation, the task instruction can be \"You need to segment all objects. This is all the candidate categories.\" and in referring segmentation, the instruction can be \"Please segment according to the following instruction.\"", + "bbox": [ + 214, + 455, + 785, + 531 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Condition Prompt. Some tasks require additional information to perform, e.g., panoptic segmentation needs specifying the candidate set of categories to be segmented, and interactive segmentation needs interactive inputs. The condition prompt is designed for these tasks. In addition to providing information, condition prompt also plays an important role in predicting categories or estimating confidence scores for each segmentation mask. In Sec. 3.3, we will discuss the design of condition prompts for different tasks in detail.", + "bbox": [ + 214, + 532, + 785, + 638 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Mask Token. The LLM is designed for text output and cannot directly generate segmentation masks. To bypass this challenge, we append a set of mask tokens after other inputs, and then these mask tokens are decoded to segmentation masks by a mask generator (will be introduced in Sec. 3.2). This design is inspired by Mask2Former [6], with the difference that the mask tokens in Mask2Former are used directly in the mask generator, whereas the mask tokens in our approach are first updated by the LMM and then used in the mask generator, and we found our approach leads better performance in practice (see Tab. 2).", + "bbox": [ + 212, + 641, + 784, + 762 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Some works, such as LISA [17] and PixelLM [40], take similar seg tokens as input and employ a decoder to generate the masks. However, our objective is fundamentally different: in LISA and PixelLLM, seg tokens are used to generate the final prediction, while we generate mask proposals and further classify them based on condition prompts. Compared to the design of LISA and PixelLLM,", + "bbox": [ + 212, + 763, + 784, + 839 + ], + "page_idx": 5 + }, + { + "type": "page_number", + "text": "6", + "bbox": [ + 217, + 114, + 228, + 126 + ], + "page_idx": 5 + }, + { + "type": "header", + "text": "Z. Zhang, Y. Ma, E. Zhang, X. Bai", + "bbox": [ + 271, + 114, + 504, + 127 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/f125fa85c77e2aafe356b5cbf1636857c587eafbef960a94bbcba5e61c8dfef3.jpg", + "image_caption": [ + "Fig. 3: Detailed processing for different condition prompts. (a) shows the processing for category condition. (b) shows the processing for sentence condition. (c) shows the processing for visual-prior condition." + ], + "image_footnote": [], + "bbox": [ + 246, + 143, + 759, + 290 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "our approach is more flexible and adaptable to more tasks, yet decoupling mask generation and classification is more efficient (see Tab. 3 and Tab. 4).", + "bbox": [ + 212, + 369, + 784, + 402 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "3.2 Mask Generator", + "text_level": 1, + "bbox": [ + 215, + 421, + 397, + 435 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "The mask generator predicts the mask and their category probabilities from three inputs: a multi-level visual features $\\{v_{l}\\}_{l = 1}^{L}$ , a set of mask tokens $\\{q_i\\}_{i = 1}^N$ and a set of condition embeddings $\\{c_k\\}_{k = 1}^K$ . It can be formally defined as:", + "bbox": [ + 212, + 445, + 782, + 492 + ], + "page_idx": 6 + }, + { + "type": "equation", + "text": "\n$$\n\\{(m _ {i}, p _ {i}) \\} _ {i = 1} ^ {N} = \\operatorname {M a s k G e n e r a t o r} (\\{v _ {i} \\} _ {l = 1} ^ {L}, \\left\\{q _ {i} \\right\\} _ {i = 1} ^ {N}, \\left\\{c _ {k} \\right\\} _ {k = 1} ^ {K}), \\tag {1}\n$$\n", + "text_format": "latex", + "bbox": [ + 307, + 498, + 785, + 516 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "where $m_{i}\\in \\mathbb{R}^{H\\times W}$ is the i-th predicted segmentation mask and $p_i\\in \\mathbb{R}^K$ is the corresponding category probability. In practice, the multi-level visual features are the internal features of the Swin visual encoder used in LMM. The design of the mask generator follows Mask2Former, which employs multi-scale deformable attention as a pixel decoder and a transformer-based mask decoder to generate segmentation masks. The class of each mask is predicted by the condition embedding $\\{c_k\\}$ , which is basically obtained from the output of the condition prompt, the obtaining method is slightly different for different types of conditions.", + "bbox": [ + 212, + 522, + 787, + 645 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "3.3 Design of Condition Prompts", + "text_level": 1, + "bbox": [ + 215, + 665, + 503, + 681 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "In our approach, the conditional prompt serves two important purposes: First, it provides the necessary information required to solve the task; Second, we use the output embedding of the conditional prompt in the LLM as classifier weights to predict the class of each segmentation mask. The design of the conditional prompt is closely tied to the type of task, and based on the information required for different tasks, we summarize three condition types: category condition, sentence condition, and visual-prior condition.", + "bbox": [ + 212, + 689, + 787, + 794 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Category Condition This condition type is used for tasks that need specifying a set of categories to be segmented, such as semantic segmentation, instance segmentation, and panoptic segmentation, and often needs to predict the", + "bbox": [ + 212, + 794, + 787, + 840 + ], + "page_idx": 6 + }, + { + "type": "header", + "text": "PSALM: Pixelwise SegmentAtion with Large Multi-Modal Model", + "bbox": [ + 295, + 114, + 732, + 128 + ], + "page_idx": 6 + }, + { + "type": "page_number", + "text": "7", + "bbox": [ + 774, + 116, + 784, + 126 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "class probability of each segmentation mask. Specifically, given a set of category names, we join them into a sentence by a comma separator, e.g., given three categories: person, bicycle, and car, the joint sentence is \"person, bicycle, car\".", + "bbox": [ + 212, + 146, + 782, + 191 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "The joint sentences are then proceeded by LMM to get the output embeddings, which can be further used to classify the predicted segmentation masks. Specifically, for each category, we select its corresponding output embeddings and apply the avg_pooling over them to obtain a condition embedding $c \\in \\mathbb{R}^D$ , where $D$ is the embedding dimension, and thus condition embedding for all categories is a set $\\{c_k\\}_{k=1}^K$ , where $K$ is the number of categories (see Fig. 3 (a)). This embedding set can be used by the mask generator to predict the class.", + "bbox": [ + 212, + 191, + 784, + 297 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Sentence Condition This condition is usually used for referring segmentation. Unlike the category condition, where category names are usually short, sentences are much longer, and not every word in the sentence is useful, so the avg_pooling is not the optimal choice here. Instead, we introduce a special [REF] token, which is appended after the condition sentence as an anchor to aggregate useful information, and the output embedding of [REF] token, i.e., the output features of LMM on the location of [REF] token, is used as condition embedding $c$ and used by mask generator, as shown in Fig. 3 (b).", + "bbox": [ + 212, + 299, + 785, + 419 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Visual-Prior Condition We formulate most interaction (e.g., point, mask, box, or scribble) used in interactive segmentation tasks as the visual-prior condition. Taking the scribble as an example, we first generate a binary scribble curve with a width of 5 pixels on a binary map, whose size is the same as the input image of the Swin vision encoder, and then apply the binary map to the output features of the vision-language alignment model, performing average mask pooling by upsampling the output feature map to the size of the binary map. The average pooled feature is used as the visual-prior condition and fed into LMM. If there are multiple interactions, the process is repeated for each, yielding multiple pooled features as inputs, each feature separated using a strategy akin to the category condition. For other types of interaction, we adopt similar approaches. Specifically, for box or mask, we directly apply the pooling operator by treating them as binary masks, and for point, we bold it to a 10-pixel circle and then apply the mask pooling. After that, we use the output embedding of the visual-prior condition as the classifier weight in the mask generator to estimate the confidence of each mask proposal, as shown in Fig. 3 (c).", + "bbox": [ + 212, + 421, + 787, + 662 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "3.4 Training Objectives", + "text_level": 1, + "bbox": [ + 215, + 689, + 423, + 704 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "The training process of PSALM has two stages: In the first stage, we train the visual language alignment model following LLaVA setting by freezing the vision encoder and LMM; In the second stage, we only freeze the vision encoder and fine-tune all other modules, including the mask generator. Similarly to Mask2Former, we use matching loss in the second stage training, i.e., we use bipartite matching to find the optimal assignments between the mask proposals and the ground truth masks by minimizing the matching loss and use these assignments to perform training. The loss has two terms: $\\mathcal{L} = \\mathcal{L}_{mask} + \\mathcal{L}_{cls}$ ,", + "bbox": [ + 212, + 719, + 787, + 840 + ], + "page_idx": 7 + }, + { + "type": "page_number", + "text": "8", + "bbox": [ + 217, + 114, + 228, + 126 + ], + "page_idx": 7 + }, + { + "type": "header", + "text": "Z. Zhang, Y. Ma, E. Zhang, X. Bai", + "bbox": [ + 271, + 114, + 504, + 127 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "where $\\mathcal{L}_{mask}$ indicates the mask loss which comprises a pixel-level Binary Cross-Entropy (BCE) loss and Dice loss [42], while the $\\mathcal{L}_{cls}$ indicates the category classification loss, and we use Cross-Entropy loss for category condition and BCE loss for other cases.", + "bbox": [ + 212, + 146, + 787, + 205 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "4 Experiments", + "text_level": 1, + "bbox": [ + 214, + 229, + 375, + 247 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "We conduct extensive experiments on various datasets and tasks to analyze PSALM. Details of the datasets and implementations are given in the Appendix.", + "bbox": [ + 212, + 261, + 787, + 292 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "4.1 Ablations", + "text_level": 1, + "bbox": [ + 214, + 314, + 343, + 329 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "We first ablate key designs and present behind insights in this section. To better show how different designs affect the performance on a wide range of tasks, we mainly report the results on three in-domain benchmarks: COCO Panoptic Segmentation (COCO-Pan), RefCOCO-val (RefCOCO), and COCO Interactive Segmentation with point inputs (COCO-Point), and one out-of-domain benchmark: open-vocabulary instance segmentation on ADE20K-150 (A150-OV).", + "bbox": [ + 212, + 339, + 787, + 431 + ], + "page_idx": 8 + }, + { + "type": "table", + "img_path": "images/0f5995d076b5cbe46765139a7ccf7fd52bc4d1d975f628dd480c557806db89b1.jpg", + "table_caption": [ + "Table 2: Ablation on the design of mask tokens. w.LLM: use mask tokens as inputs of LLM. Prefix: place mask tokens at the front. Suffix: place mask tokens at the end." + ], + "table_footnote": [], + "table_body": "
Mask TokensCOCO-PanRefCOCOCOCO-PointA150-OV
w.LLMPos.PQcIoUmIoUmAP
✓Suffix55.176.153.39.3
✗-54.8(↓0.3)74.3(↓1.8)53.1(↓0.2)8.2(↓1.1)
✓Prefix55.0(↓0.1)75.1(↓1.0)53.0(↓0.3)7.8(↓1.5)
", + "bbox": [ + 220, + 523, + 552, + 602 + ], + "page_idx": 8 + }, + { + "type": "table", + "img_path": "images/fe907bd25d5875899faaaa398aee2788031d6f7ac8fa45ebec1f60b78eb8d2fa.jpg", + "table_caption": [ + "Table 3: Effect of decouple design in COCO Semantic Segmentation." + ], + "table_footnote": [], + "table_body": "
DecouplemIoUfwIoU
✓66.572.5
X42.735.0
", + "bbox": [ + 604, + 525, + 761, + 588 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "Design of Mask Tokens. In our approach, we use a set of mask tokens to predict the mask proposal. In practice, we have found that using mask tokens as inputs to the LLM leads to better performance than applying them directly to the mask generator, which is the default method of Mask2Former. Tab. 2 shows the results, where the direct use of mask tokens leads to a noticeable performance degradation in RefCOCO and A150-OV. We believe this is because using mask tokens as input leads to a better awareness of the information needed for the task and thus improves performance, which is essential for these two tasks. For better validation, we placed mask tokens before conditional prompts and task instruction prompts and found a similar performance drop to that of not using mask tokens in LMM, which further supports our hypothesis.", + "bbox": [ + 212, + 628, + 787, + 794 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "Compared to LISA and other methods that use seg token to generate final segmentation results directly, our mask proposal approach has three advantages: First, our design is more flexible and thus can be applied to a wider range", + "bbox": [ + 212, + 795, + 787, + 840 + ], + "page_idx": 8 + }, + { + "type": "header", + "text": "PSALM: Pixelwise SegmentAtion with Large Multi-Modal Model", + "bbox": [ + 295, + 113, + 732, + 128 + ], + "page_idx": 8 + }, + { + "type": "page_number", + "text": "9", + "bbox": [ + 774, + 114, + 785, + 127 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "of segmentation tasks, especially tasks that require predicting category or confidence scores; Second, our design decouples the mask prediction and classification, which alleviating the learning difficulties for some tasks. In Tab. 3, we study how the decouple design affects the semantic segmentation performance on COCO Semantic Segmentation $^{4}$ , and we found our decouple design is significantly better $^{5}$ . Third, the mask proposals allow multiple masks to be generated for a single instance, which makes the mask accuracy superior to solutions like LISA that only predict a single mask. Tab. 4 shows that using more mask proposals on RefCOCO gives a clear improvement over using a single mask.", + "bbox": [ + 212, + 146, + 787, + 282 + ], + "page_idx": 9 + }, + { + "type": "table", + "img_path": "images/165f472494cbfd4d544ad17166d0044030aee5e4226c3081348b87351fa6ede8.jpg", + "table_caption": [ + "Table 4: Effects of number of mask tokens in RefCOCO(cIoU). The results of LISA are listed as a reference." + ], + "table_footnote": [], + "table_body": "
#Mask TokenvaltestAtestB
LISA74.176.571.1
175.378.072.2
10076.578.573.4
", + "bbox": [ + 240, + 377, + 419, + 441 + ], + "page_idx": 9 + }, + { + "type": "table", + "img_path": "images/effd1c92afd062bb5888b8b65ab822a939625c828f9a2d00d52639c572b4f9c5.jpg", + "table_caption": [ + "Table 5: Ablation on different designs for condition prompts." + ], + "table_footnote": [], + "table_body": "
ConditionCOCO-PanRefCOCOA150-OV
CategorySentencePQcIoUmAP
avg_poolingavg_pooling55.175.39.2
[REF]54.976.18.4
avg_pooling[REF]55.176.19.3
", + "bbox": [ + 473, + 358, + 797, + 434 + ], + "page_idx": 9 + }, + { + "type": "text", + "text": "Design of Condition Prompts. Another key design in PSALM is the condition prompt, particularly the way we obtain the condition embeddings, which are used as classifier weights in the mask generator to predict the class of mask proposals. As described in Sec. 3.3, for category condition, we use the avg_pooling over output embeddings of each class name as the condition embedding, and for sentence condition, we adopt a [REF] token to aggregate useful information. Tab. 5 shows the ablation, where we first tried to use the same design for all conditions and found that avg_pooling performed slightly better on COCO-Pan, with a larger improvement on A150-OV, while [REF] worked better on RefCOCO. We further used different designs and found that the advantages of each design are preserved, and the best overall performance is achieved.", + "bbox": [ + 212, + 470, + 787, + 638 + ], + "page_idx": 9 + }, + { + "type": "table", + "img_path": "images/a25a87db8fe80a6deaace08e2a5caff4eea8604d48ef5aea3f40d8a10db1e1be.jpg", + "table_caption": [ + "Table 6: Ablation on effect of vision-language alingment." + ], + "table_footnote": [], + "table_body": "
VL AlignmentCOCO-Pan PQRefCOCO cIoUCOCO-Point mIoUA150-OV mAP
✓55.176.153.39.3
✗54.9(↓0.2)71.7(↓4.4)53.0 (↓0.3)8.2(↓1.1)
", + "bbox": [ + 218, + 704, + 488, + 756 + ], + "page_idx": 9 + }, + { + "type": "table", + "img_path": "images/3bbc623b9edd52ac54bbcf4b9398e5416470787643051b752fe5f8e05c966d8e.jpg", + "table_caption": [ + "Table 7: Ablation on joint training." + ], + "table_footnote": [], + "table_body": "
AblationCOCO-PanRefCOCOCOCO-PointA150-OV
PQcIoUmIoUmAP
Task Specific Train55.676.562.27.0
Joint Train55.983.664.19.0
Δ+0.3+7.1+1.9+2.0
", + "bbox": [ + 513, + 696, + 790, + 751 + ], + "page_idx": 9 + }, + { + "type": "page_number", + "text": "10", + "bbox": [ + 217, + 114, + 235, + 126 + ], + "page_idx": 9 + }, + { + "type": "header", + "text": "Z. Zhang, Y. Ma, E. Zhang, X. Bai", + "bbox": [ + 271, + 113, + 506, + 128 + ], + "page_idx": 9 + }, + { + "type": "page_footnote", + "text": "This benchmark is introduced by Mask2Former, which is composed by merging instances belonging to the same class together in COCO Panoptic Segmentation.", + "bbox": [ + 217, + 796, + 785, + 825 + ], + "page_idx": 9 + }, + { + "type": "page_footnote", + "text": "5 More experimental details are in Appendix.", + "bbox": [ + 220, + 825, + 524, + 839 + ], + "page_idx": 9 + }, + { + "type": "table", + "img_path": "images/091e917b94172ac571def0e770bf591fa0e6fedab9a80e7f503fec3f76d739ae.jpg", + "table_caption": [ + "Table 8: Comparison with the state-of-the-art methods on three referring image segmentation benchmarks with cIoU. (ft) denotes models further finetuned on RefCOCO+/g after mix training. We abbreviate the datasets: COCO(C) [24], LVIS(L) [13], RefCOCO(RC) [57], Object365(O365) [41], Video segmentation datasets(V), ADE20K(A) [59], COCO-Stuff(CS) [3], PACO-LVIS(PL) [38], PASCAL-Part(PP) [5], GranD(G) [39], VOC2010(VOC) [11], Visual Genome(VG) [16], Flicker30k(F30K) [35], MUSE(M) [40], gRefCOCO(gRC) [25], COCO-Interactive(Cl)." + ], + "table_footnote": [], + "table_body": "
MethodSegmentation DataLLM TypeRefCOCORefCOCO+RefCOCOg
valtestAtestBvaltestAtestBvaltest
SEEM-L [61]C, L, RC-------65.6-
UNINEXT-L [53]O365, C, RC, V-80.382.677.870.074.962.673.473.7
UNINEXT-H [53]O365, C, RC, V-82.283.481.372.576.466.274.776.4
LISA [17]A, CS, RC, PL, PPVicuna-7B74.176.571.162.467.456.566.468.5
LISA(ft) [17]A, CS, RC, PL, PPVicuna-7B74.979.172.365.170.858.167.970.6
GLaMM [39]G, RCVicuna-7B79.583.276.972.678.764.674.274.9
u-LLaVA [50]A, CS, RC, PL, VOCVicuna-7B80.482.777.872.276.666.874.875.6
PerceptionGPT [34]RC, VG, F30kVicuna-13B75.379.172.168.974.061.970.771.9
PixelLM [40]A, CS, RC, PL, MLlama2-13B73.076.568.266.371.758.369.370.5
GSVA [48]A, CS, RC, PL, PP, gRCLlama2-13B77.779.974.268.071.561.573.273.9
GSVA(ft) [48]A, CS, RC, PL, PP, gRCLlama2-13B79.281.777.170.373.863.675.777.0
PSALMC, RC, CIPhi-1.5 (1.3B)83.684.781.672.975.570.173.874.4
", + "bbox": [ + 241, + 253, + 754, + 402 + ], + "page_idx": 10 + }, + { + "type": "text", + "text": "Importance of VL-alignments. The visual-language alignment stage (i.e., first training stage) is to project the visual features to the text input space, and it is an important step towards making the LLM understand images. In Tab. 6, we examine the impact of this stage, and we found that without VL alignment, the performance of all four tasks becomes worse, with the performance of A150-OV and RefCOCO being significantly affected, for example, the A150-OV dropped by -1.1 mAP and RefCOCO even dropped by -4.4 cIoU, probably because these two tasks require a strong requirement on understanding the relationship between vision and language. This result also suggests that the VL alignment is essential, and the LMM-based segmentation models have strong potential.", + "bbox": [ + 212, + 430, + 787, + 580 + ], + "page_idx": 10 + }, + { + "type": "text", + "text": "Joint Training. Our architecture design and input schema help integrate various segmentation tasks so that they can be trained on one model. Tab. 7 shows the effect of this joint training on different tasks. For the task-specific models, we perform the training on the corresponding task data for $18\\mathrm{k}$ iterations. In contrast, the joint training setting (see implementation section for details) has a total of $56\\mathrm{k}$ training iterations, which corresponds to $14\\mathrm{k}$ iterations per task. The results show that joint training of different tasks greatly improves the performance. This suggests that learning between tasks is mutually beneficial, which is also the secret of success in LLM. For example, the generic segmentation task helps refine the mask prediction in the referring segmentation, and referring expressions also enhance the model's ability to recognize more unseen categories, which in turn improves the performance of open-vocabulary segmentation tasks.", + "bbox": [ + 212, + 580, + 787, + 763 + ], + "page_idx": 10 + }, + { + "type": "text", + "text": "4.2 System-Level Comparison on In-Domain Tasks", + "text_level": 1, + "bbox": [ + 214, + 785, + 648, + 800 + ], + "page_idx": 10 + }, + { + "type": "text", + "text": "In this section, we compared our model on three in-domain tasks with other state-of-the-arts to illustrate the effectiveness of our approach.", + "bbox": [ + 212, + 809, + 785, + 839 + ], + "page_idx": 10 + }, + { + "type": "header", + "text": "PSALM: Pixelwise SegmentAtion with Large Multi-Modal Model", + "bbox": [ + 295, + 114, + 730, + 128 + ], + "page_idx": 10 + }, + { + "type": "page_number", + "text": "11", + "bbox": [ + 767, + 116, + 782, + 126 + ], + "page_idx": 10 + }, + { + "type": "text", + "text": "Referring Segmentation. Most existing works aimed at getting LMMs to perform image segmentation are designed for reference segmentation tasks. We compare PSALM with other works on RefCOCO, RefCOCO+, and RefCOCOg, and Tab. 8 shows the results. Owing to the generalized and flexible design of PSALM and the advantages of joint training on multiple tasks and datasets, our system was able to achieve state-of-the-art (SOTA) performance on RefCOCO and RefCOCO+, and competitive performance on RefCOCOg with STOA, despite being driven by LLM with only 1.3B parameters. It is worth noting unlike methods such as LISA and GSVA, which may achieve improvements through task-specific fine-tuning (gray-labeled results), PSALM does not perform additional fine-tuning but still achieves better performance on RefCOCO and RefCOCO+ than their fine-tuned models.", + "bbox": [ + 212, + 146, + 787, + 327 + ], + "page_idx": 11 + }, + { + "type": "text", + "text": "Generic Segmentation. We evaluate PSALM with state-of-the-art methods on the COCO Panoptic Segmentation validation set (Tab. 9). Here, we follow the evaluation protocol used in Mask2Former to report PQ, which is the main metric for panoptic segmentation, mAP on thing classes for instance segmentation, and mIoU by merging instance masks from the same category for semantic segmentation. Compared to other methods, PSALM achieves comparable performance at similar visual backbone sizes, demonstrating that PSALM is a powerful architecture, even when compared to approaches designed for specific tasks.", + "bbox": [ + 212, + 330, + 787, + 454 + ], + "page_idx": 11 + }, + { + "type": "table", + "img_path": "images/9c47837cddcdba2939e4d52662a0ceffd8e535d3612978b497d5da89696ba2a8.jpg", + "table_caption": [ + "Table 9: Comparison with the state-of-the-art methods on Panoptic COCO-val. We abbreviate the datasets: COCO-SAM(CM) [21], VIPSeg (VIP) [29], while others following Tab. 8." + ], + "table_footnote": [], + "table_body": "
MethodBackboneSeg. DataPQmAPmIoU
Mask2Former [6]Swin-BC55.145.265.1
Mask2Former [6]Swin-LC57.848.667.4
X-Decoder [60]DaViT-BC, L, RC56.245.866.0
SEEM [61]DaViT-BC, L, RC56.146.466.3
OMG-Seg [21]ConvNeXt-XXLC, VIP, CM, V55.4--
PSALMSwin-BC, RC, CI55.945.766.6
", + "bbox": [ + 300, + 541, + 697, + 637 + ], + "page_idx": 11 + }, + { + "type": "text", + "text": "Interactive Segmentation. We also evaluate PSALM in the interactive segmentation tasks. Since the task does not have a well-developed dataset containing all four instructions, previous works have typically used the in-house dataset, thus we re-evaluate other methods on the COCO interactive validation set. The results are shown in Tab. 10, and PSALM achieves leading performance on point, scribble, and mask instructions than all other methods, while on box instruction, SAM performs better on mIoU but worse on cIoU, and we hypothesize might be caused by the different distribution of training data, and the fact that SAM is trained on SA-1B [15], which has a much larger data scale than what we used. In addition, we also report the official results of SEEM, which only evaluated 600 samples from the COCO validation set as a reference.", + "bbox": [ + 212, + 672, + 787, + 839 + ], + "page_idx": 11 + }, + { + "type": "page_number", + "text": "12", + "bbox": [ + 217, + 114, + 235, + 126 + ], + "page_idx": 11 + }, + { + "type": "header", + "text": "Z. Zhang, Y. Ma, E. Zhang, X. Bai", + "bbox": [ + 271, + 114, + 506, + 128 + ], + "page_idx": 11 + }, + { + "type": "table", + "img_path": "images/f3006fe4d43851b612ef9f3c3ce317f4e80a6113ccf35e4c7da12e95facd171d.jpg", + "table_caption": [ + "Table 10: Comparison with the state-of-the-art methods on COCO-Interactive. The results of SEEM-B* is the result reported in official paper, which is evaluated on 600 random samples of COCO-val, while all others are evaluated on all samples of COCO-val. Abbreviations for each dataset are the same as Tab. 9." + ], + "table_footnote": [], + "table_body": "
MethodSeg. DataPointScribbleBoxMask
mIoUcIoUmIoUcIoUmIoUcIoUmIoUcIoU
SAM-B [15]SA-1B48.733.6--73.768.7--
SAM-L [15]SA-1B51.837.7--76.671.6--
SEEM-B [61]C, L, RC47.857.843.044.044.942.148.465.0
SEEM-B* [61]C, L, RC81.7-83.5-75.7-76.0-
OMG-Seg [21]C, VIP, CM, V59.3-------
PSALMC, RC, CI64.374.066.980.067.380.967.682.4
", + "bbox": [ + 282, + 210, + 718, + 319 + ], + "page_idx": 12 + }, + { + "type": "text", + "text": "4.3 Generalizability on Out-of-Domain Tasks", + "text_level": 1, + "bbox": [ + 214, + 344, + 602, + 359 + ], + "page_idx": 12 + }, + { + "type": "text", + "text": "Thanks to the flexible design of architecture and input schema, multi-task joint-training, and the strong visual understanding capability of LMM, PSALM shows excellent performance on in-domain tasks, but more importantly, PSALM also demonstrates great potential to generalize to out-of-domain tasks in the zero-shot setting. In this section, we conduct experiments on three different out-of-domain tasks: open-vocabulary segmentation, generalized referring expression segmentation, and video object segmentation. We also tested the zero-shot result of the correspondence benchmark in Ego-Exo4D [12] and Video Object Segmentation in the Appendix.", + "bbox": [ + 212, + 368, + 787, + 505 + ], + "page_idx": 12 + }, + { + "type": "table", + "img_path": "images/8971ad6fb34aa6cf5de66b97eb72b0d84307fa19eadd464947f6582e0cad6a58.jpg", + "table_caption": [ + "Table 11: Comparison with the state-of-the-art methods on open-vocabulary instance segmentation and semantic segmentation benchmarks. We use mAP for instance segmentation and mIoU for semantic segmentation. We abbreviate the datasets: Pascal Context-459(PC459) [30], Pascal Context-59(PC59) [30], Pascal VOC-20(PAS20) [11]" + ], + "table_footnote": [], + "table_body": "
MethodOV Instance Seg.OV Semantic Seg.
A150CityscapesPC459A150PC59PAS20
MaskCLIP [9]6.0-10.023.745.9-
ODISE [49]14.4-14.529.957.3-
SAN [51]10.6-17.133.360.295.5
PSALM9.020.510.218.248.581.3
PSALM+LVIS13.919.314.024.457.295.0
", + "bbox": [ + 325, + 597, + 674, + 708 + ], + "page_idx": 12 + }, + { + "type": "text", + "text": "Open-Vocabulary Segmentation. We first evaluate PSALM in open-vocabulary segmentation tasks, which require the model to have the ability to deal with unseen categories in training. Here, we conduct experiments on both open-vocabulary instance segmentation and open-vocabulary semantic segmentation, and Tab. 11 shows the results. Without any special design, PSALM achieves reasonably good performance, although it is still worse than the best specific method in this task, such as SAN, but we believe that PSALM has a strong", + "bbox": [ + 212, + 734, + 802, + 840 + ], + "page_idx": 12 + }, + { + "type": "header", + "text": "PSALM: Pixelwise SegmentAtion with Large Multi-Modal Model", + "bbox": [ + 295, + 113, + 730, + 128 + ], + "page_idx": 12 + }, + { + "type": "page_number", + "text": "13", + "bbox": [ + 767, + 114, + 784, + 126 + ], + "page_idx": 12 + }, + { + "type": "text", + "text": "potential to be further improved by adding more diverse training data, which is advantages of our method. We have also made a preliminary attempt by further involving the LVIS dataset, and as we expected, the performance has significantly improvements. In addition, existing open-vocabulary segmentation methods are built upon the CLIP model or diffusion model, while our approach is based on LMM models, which is a new path and attempt to bring new inspiration to the community, which we believe is even more important than the performance.", + "bbox": [ + 212, + 146, + 787, + 252 + ], + "page_idx": 13 + }, + { + "type": "table", + "img_path": "images/60d51dc93f9282efd496d8c84eb9ec5935a24baf77684ee292b2825aa7d84a68.jpg", + "table_caption": [ + "Table 12: Our method's zero-shot performance on gRefCOCO. (ft) denotes models further fine-tuned on gRefCOCO after mix training." + ], + "table_footnote": [], + "table_body": "
MethodsLLM TypeZero-ShotvaltestAtestB
cIoUgIoUcIoUgIoUcIoUgIoU
MattNet [56]-X47.548.258.759.345.346.1
LTS [8]-X52.352.761.962.649.950.4
ReLA [25]-X62.463.669.370.059.961.0
LISA [17]Vicuna-7BX38.732.252.648.544.839.7
LISA(ft) [17]Vicuna-7BX61.763.369.270.160.361.3
GSVA [48]Vicuna-7BX61.763.369.270.160.361.3
GSVA(ft) [48]Vicuna-7BX63.366.569.971.160.562.2
PSALMPhi-1.5 (1.3B)✓42.043.352.454.550.652.5
", + "bbox": [ + 300, + 316, + 697, + 465 + ], + "page_idx": 13 + }, + { + "type": "text", + "text": "Generalized Referring Expression Segmentation. The referring segmentation datasets used in training contain only a single object, however, the design of the mask proposal allows PSALM to directly address multi-target without any further training or fine-tuning. We evaluate the gRefCOCO benchmark which contains multiple segment targets. In practice, given an expression, we compute the similarity with all mask proposals and retain all masks with similarity greater than 0.6 as foreground. Tab. 12 shows the results, PSALM also achieved very promising performance, even outperforming the LISA version that only pre-trained on gRefCOCO but without task-specific fine-tuning.", + "bbox": [ + 212, + 492, + 787, + 628 + ], + "page_idx": 13 + }, + { + "type": "text", + "text": "5 Conclusion", + "text_level": 1, + "bbox": [ + 214, + 650, + 359, + 666 + ], + "page_idx": 13 + }, + { + "type": "text", + "text": "The PSALM proposed in this study extends the capability of LMM from text output tasks to image segmentation, addresses the output limitations of the LMM, and unifies various segmentation tasks. PSALM exhibits excellent performance in multiple in-domain tasks, and its generalization ability in out-of-domain tasks further underscores its potential.", + "bbox": [ + 212, + 680, + 787, + 756 + ], + "page_idx": 13 + }, + { + "type": "text", + "text": "Acknowledgments", + "text_level": 1, + "bbox": [ + 214, + 777, + 393, + 796 + ], + "page_idx": 13 + }, + { + "type": "text", + "text": "This work was supported by the National Science Fund for Distinguished Young Scholars of China (Grant No.62225603).", + "bbox": [ + 212, + 809, + 785, + 839 + ], + "page_idx": 13 + }, + { + "type": "page_number", + "text": "14", + "bbox": [ + 217, + 114, + 235, + 126 + ], + "page_idx": 13 + }, + { + "type": "header", + "text": "Z. Zhang, Y. Ma, E. Zhang, X. Bai", + "bbox": [ + 271, + 113, + 506, + 128 + ], + "page_idx": 13 + }, + { + "type": "text", + "text": "References", + "text_level": 1, + "bbox": [ + 217, + 143, + 321, + 160 + ], + "page_idx": 14 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "1. Alayrac, J.B., Donahue, J., Luc, P., Miech, A., Barr, I., Hasson, Y., Lenc, K., Mensch, A., Millican, K., Reynolds, M., et al.: Flamingo: a visual language model for few-shot learning. Advances in Neural Information Processing Systems 35, 23716-23736 (2022)", + "2. Brown, T., Mann, B., Ryder, N., Subbiah, M., Kaplan, J.D., Dhariwal, P., Neelakantan, A., Shyam, P., Sastry, G., Askell, A., et al.: Language models are few-shot learners. Advances in neural information processing systems 33, 1877-1901 (2020)", + "3. Caesar, H., Uijlings, J., Ferrari, V.: Coco-stuff: Thing and stuff classes in context. In: Proceedings of the IEEE conference on computer vision and pattern recognition. pp. 1209-1218 (2018)", + "4. Chen, K., Zhang, Z., Zeng, W., Zhang, R., Zhu, F., Zhao, R.: Shikra: Unleashing multimodal llm's referential dialogue magic. arXiv preprint arXiv:2306.15195 (2023)", + "5. Chen, X., Mottaghi, R., Liu, X., Fidler, S., Urtasun, R., Yuille, A.: Detect what you can: Detecting and representing objects using holistic models and body parts. In: Proceedings of the IEEE conference on computer vision and pattern recognition. pp. 1971-1978 (2014)", + "6. Cheng, B., Misra, I., Schwing, A.G., Kirillov, A., Girdhar, R.: Masked-attention mask transformer for universal image segmentation. In: Proceedings of the IEEE/CVF conference on computer vision and pattern recognition. pp. 1290-1299 (2022)", + "7. Chiang, W.L., Li, Z., Lin, Z., Sheng, Y., Wu, Z., Zhang, H., Zheng, L., Zhuang, S., Zhuang, Y., Gonzalez, J.E., et al.: Vicuna: An open-source chatbot impressing gpt-4 with $90\\%$ chatgpt quality. See https://vicuna.lmsys.org (accessed 14 April 2023) (2023)", + "8. Ding, H., Liu, C., Wang, S., Jiang, X.: Vision-language transformer and query generation for referring segmentation. In: Proceedings of the IEEE/CVF International Conference on Computer Vision. pp. 16321-16330 (2021)", + "9. Ding, Z., Wang, J., Tu, Z.: Open-vocabulary universal image segmentation with maskclip (2023)", + "0. Dong, R., Han, C., Peng, Y., Qi, Z., Ge, Z., Yang, J., Zhao, L., Sun, J., Zhou, H., Wei, H., et al.: Dreamllm: Synergistic multimodal comprehension and creation. arXiv preprint arXiv:2309.11499 (2023)", + "1. Everingham, M., Van Gool, L., Williams, C.K.I., Winn, J., Zisserman, A.: The PASCAL Visual Object Classes Challenge 2010 (VOC2010) Results. http://www.pascal-network.org/challenges/VOC/voc2010/workshop/index.html", + "2. Grauman, K., Westbury, A., Torresani, L., Kitani, K., Malik, J., Afouras, T., Ashutosh, K., Baiyya, V., Bansal, S., Boote, B., et al.: Ego-exo4d: Understanding skilled human activity from first-and third-person perspectives. arXiv preprint arXiv:2311.18259 (2023)", + "3. Gupta, A., Dollar, P., Girshick, R.: Lvis: A dataset for large vocabulary instance segmentation. In: Proceedings of the IEEE/CVF conference on computer vision and pattern recognition. pp. 5356-5364 (2019)", + "4. Jain, J., Li, J., Chiu, M.T., Hassani, A., Orlov, N., Shi, H.: Oneformer: One transformer to rule universal image segmentation. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 2989-2998 (2023)", + "5. Kirillov, A., Mintun, E., Ravi, N., Mao, H., Rolland, C., Gustafson, L., Xiao, T., Whitehead, S., Berg, A.C., Lo, W.Y., et al.: Segment anything. arXiv preprint arXiv:2304.02643 (2023)" + ], + "bbox": [ + 225, + 172, + 785, + 839 + ], + "page_idx": 14 + }, + { + "type": "header", + "text": "PSALM: Pixelwise SegmentAtion with Large Multi-Modal Model", + "bbox": [ + 295, + 114, + 730, + 128 + ], + "page_idx": 14 + }, + { + "type": "page_number", + "text": "15", + "bbox": [ + 767, + 116, + 784, + 126 + ], + "page_idx": 14 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "16. Krishna, R., Zhu, Y., Groth, O., Johnson, J., Hata, K., Kravitz, J., Chen, S., Kalantidis, Y., Li, L.J., Shamma, D.A., et al.: Visual genome: Connecting language and vision using crowdsourced dense image annotations. International journal of computer vision 123, 32-73 (2017)", + "17. Lai, X., Tian, Z., Chen, Y., Li, Y., Yuan, Y., Liu, S., Jia, J.: Lisa: Reasoning segmentation via large language model. arXiv preprint arXiv:2308.00692 (2023)", + "18. Li, B., Zhang, P., Yang, J., Zhang, Y., Pu, F., Liu, Z.: Otterhd: A high-resolution multi-modality model. arXiv preprint arXiv:2311.04219 (2023)", + "19. Li, J., Li, D., Savarese, S., Hoi, S.: Blip-2: Bootstrapping language-image pretraining with frozen image encoders and large language models. arXiv preprint arXiv:2301.12597 (2023)", + "20. Li, J., Li, D., Xiong, C., Hoi, S.: Blip: Bootstrapping language-image pre-training for unified vision-language understanding and generation. In: International Conference on Machine Learning. pp. 12888-12900. PMLR (2022)", + "21. Li, X., Yuan, H., Li, W., Ding, H., Wu, S., Zhang, W., Li, Y., Chen, K., Loy, C.C.: Omg-seg: Is one model good enough for all segmentation? arXiv preprint arXiv:2401.10229 (2024)", + "22. Li, Y., Bubeck, S., Eldan, R., Del Giorno, A., Gunasekar, S., Lee, Y.T.: Textbooks are all you need ii: phi-1.5 technical report. arXiv preprint arXiv:2309.05463 (2023)", + "23. Li, Z., Yang, B., Liu, Q., Ma, Z., Zhang, S., Yang, J., Sun, Y., Liu, Y., Bai, X.: Monkey: Image resolution and text label are important things for large multi-modal models. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 26763-26773 (2024)", + "24. Lin, T.Y., Maire, M., Belongie, S., Hays, J., Perona, P., Ramanan, D., Dollar, P., Zitnick, C.L.: Microsoft coco: Common objects in context. In: Computer Vision-ECCV 2014: 13th European Conference, Zurich, Switzerland, September 6-12, 2014, Proceedings, Part V 13. pp. 740-755. Springer (2014)", + "25. Liu, C., Ding, H., Jiang, X.: Gres: Generalized referring expression segmentation. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 23592-23601 (2023)", + "26. Liu, H., Li, C., Li, Y., Li, B., Zhang, Y., Shen, S., Lee, Y.J.: Llava-last: Improved reasoning,OCR,and world knowledge (2024)", + "27. Liu, H., Li, C., Wu, Q., Lee, Y.J.: Visual instruction tuning. Advances in neural information processing systems 36 (2024)", + "28. Liu, Z., Lin, Y., Cao, Y., Hu, H., Wei, Y., Zhang, Z., Lin, S., Guo, B.: Swin transformer: Hierarchical vision transformer using shifted windows. In: Proceedings of the IEEE/CVF international conference on computer vision. pp. 10012-10022 (2021)", + "29. Miao, J., Wang, X., Wu, Y., Li, W., Zhang, X., Wei, Y., Yang, Y.: Large-scale video panoptic segmentation in the wild: A benchmark. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 21033-21043 (2022)", + "30. Mottaghi, R., Chen, X., Liu, X., Cho, N.G., Lee, S.W., Fidler, S., Urtasun, R., Yuille, A.: The role of context for object detection and semantic segmentation in the wild. In: IEEE Conference on Computer Vision and Pattern Recognition (CVPR) (2014)", + "31. Nagaraja, V.K., Morariu, V.I., Davis, L.S.: Modeling context between objects for referring expression understanding. In: Computer Vision-ECCV 2016: 14th European Conference, Amsterdam, The Netherlands, October 11-14, 2016, Proceedings, Part IV 14. pp. 792-807. Springer (2016)" + ], + "bbox": [ + 215, + 146, + 785, + 839 + ], + "page_idx": 15 + }, + { + "type": "page_number", + "text": "16", + "bbox": [ + 217, + 114, + 235, + 126 + ], + "page_idx": 15 + }, + { + "type": "header", + "text": "Z. Zhang, Y. Ma, E. Zhang, X. Bai", + "bbox": [ + 271, + 114, + 504, + 127 + ], + "page_idx": 15 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "32. OpenAI: Gpt-4v(ision) system card (2023)", + "33. Peng, Z., Wang, W., Dong, L., Hao, Y., Huang, S., Ma, S., Wei, F.: Kosmos-2: Grounding multimodal large language models to the world. arXiv preprint arXiv:2306.14824 (2023)", + "34. Pi, R., Yao, L., Gao, J., Zhang, J., Zhang, T.: Perceptiongpt: Effectively fusing visual perception into llm. arXiv preprint arXiv:2311.06612 (2023)", + "35. Plummer, B.A., Wang, L., Cervantes, C.M., Caicedo, J.C., Hockenmaier, J., Lazebnik, S.: Flickr30k entities: Collecting region-to-phrase correspondences for richer image-to-sentence models. In: Proceedings of the IEEE international conference on computer vision. pp. 2641-2649 (2015)", + "36. Pont-Tuset, J., Perazzi, F., Caelles, S., Arbeláez, P., Sorkine-Hornung, A., Van Gool, L.: The 2017 davis challenge on video object segmentation. arXiv preprint arXiv:1704.00675 (2017)", + "37. Radford, A., Kim, J.W., Hallacy, C., Ramesh, A., Goh, G., Agarwal, S., Sastry, G., Askell, A., Mishkin, P., Clark, J., et al.: Learning transferable visual models from natural language supervision. In: International conference on machine learning. pp. 8748-8763. PMLR (2021)", + "38. Ramanathan, V., Kalia, A., Petrovic, V., Wen, Y., Zheng, B., Guo, B., Wang, R., Marquez, A., Kovvuri, R., Kadian, A., et al.: Paco: Parts and attributes of common objects. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 7141-7151 (2023)", + "39. Rasheed, H., Maaz, M., Shaji, S., Shaker, A., Khan, S., Cholakkal, H., Anwer, R.M., Xing, E., Yang, M.H., Khan, F.S.: Glamm: Pixel grounding large multimodal model. arXiv preprint arXiv:2311.03356 (2023)", + "40. Ren, Z., Huang, Z., Wei, Y., Zhao, Y., Fu, D., Feng, J., Jin, X.: Pixelm: Pixel reasoning with large multimodal model. arXiv preprint arXiv:2312.02228 (2023)", + "41. Shao, S., Li, Z., Zhang, T., Peng, C., Yu, G., Zhang, X., Li, J., Sun, J.: Objects365: A large-scale, high-quality dataset for object detection. In: Proceedings of the IEEE/CVF international conference on computer vision. pp. 8430-8439 (2019)", + "42. Sudre, C.H., Li, W., Vercauteren, T., Ourselin, S., Jorge Cardoso, M.: Generalised dice overlap as a deep learning loss function for highly unbalanced segmentations. In: Deep Learning in Medical Image Analysis and Multimodal Learning for Clinical Decision Support: Third International Workshop, DLMIA 2017, and 7th International Workshop, ML-CDS 2017, Held in Conjunction with MICCAI 2017, Quebec City, QC, Canada, September 14, Proceedings 3. pp. 240-248. Springer (2017)", + "43. Sun, Q., Yu, Q., Cui, Y., Zhang, F., Zhang, X., Wang, Y., Gao, H., Liu, J., Huang, T., Wang, X.: Generative pretraining in multimodality. arXiv preprint arXiv:2307.05222 (2023)", + "44. Team, G., Anil, R., Borgeaud, S., Wu, Y., Alayrac, J.B., Yu, J., Soricut, R., Schalkwyk, J., Dai, A.M., Hauth, A., et al.: Gemini: a family of highly capable multimodal models. arXiv preprint arXiv:2312.11805 (2023)", + "45. Touvron, H., Martin, L., Stone, K., Albert, P., Almahairi, A., Babaei, Y., Bashlykov, N., Batra, S., Bhargava, P., Bhosale, S., et al.: Llama 2: Open foundation and fine-tuned chat models. arXiv preprint arXiv:2307.09288 (2023)", + "46. Wang, W., Lv, Q., Yu, W., Hong, W., Qi, J., Wang, Y., Ji, J., Yang, Z., Zhao, L., Song, X., et al.: Cogvlm: Visual expert for pretrained language models. arXiv preprint arXiv:2311.03079 (2023)", + "47. Wu, J., Jiang, Y., Liu, Q., Yuan, Z., Bai, X., Bai, S.: General object foundation model for images and videos at scale. arXiv preprint arXiv:2312.09158 (2023)" + ], + "bbox": [ + 215, + 146, + 785, + 840 + ], + "page_idx": 16 + }, + { + "type": "header", + "text": "PSALM: Pixelwise SegmentAtion with Large Multi-Modal Model", + "bbox": [ + 295, + 114, + 730, + 128 + ], + "page_idx": 16 + }, + { + "type": "page_number", + "text": "17", + "bbox": [ + 767, + 116, + 784, + 126 + ], + "page_idx": 16 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "48. Xia, Z., Han, D., Han, Y., Pan, X., Song, S., Huang, G.: Gsva: Generalized segmentation via multimodal large language models. arXiv preprint arXiv:2312.10103 (2023)", + "49. Xu, J., Liu, S., Vahdat, A., Byeon, W., Wang, X., De Mello, S.: Open-vocabulary panoptic segmentation with text-to-image diffusion models. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 2955-2966 (2023)", + "50. Xu, J., Xu, L., Yang, Y., Li, X., Xie, Y., Huang, Y.J., Li, Y.: u-llava: Unifying multi-modal tasks via large language model. arXiv preprint arXiv:2311.05348 (2023)", + "51. Xu, M., Zhang, Z., Wei, F., Hu, H., Bai, X.: Side adapter network for open-vocabulary semantic segmentation. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 2945-2954 (2023)", + "52. Xu, M., Zhang, Z., Wei, F., Lin, Y., Cao, Y., Hu, H., Bai, X.: A simple baseline for open-vocabulary semantic segmentation with pre-trained vision-language model. In: European Conference on Computer Vision. pp. 736-753. Springer (2022)", + "53. Yan, B., Jiang, Y., Wu, J., Wang, D., Luo, P., Yuan, Z., Lu, H.: Universal instance perception as object discovery and retrieval. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 15325-15336 (2023)", + "54. Yang, S., Qu, T., Lai, X., Tian, Z., Peng, B., Liu, S., Jia, J.: An improved baseline for reasoning segmentation with large language model. arXiv preprint arXiv:2312.17240 (2023)", + "55. You, H., Zhang, H., Gan, Z., Du, X., Zhang, B., Wang, Z., Cao, L., Chang, S.F., Yang, Y.: Ferret: Refer and ground anything anywhere at any granularity. arXiv preprint arXiv:2310.07704 (2023)", + "56. Yu, L., Lin, Z., Shen, X., Yang, J., Lu, X., Bansal, M., Berg, T.L.: Matttnet: Modular attention network for referring expression comprehension. In: Proceedings of the IEEE conference on computer vision and pattern recognition. pp. 1307-1315 (2018)", + "57. Yu, L., Poirson, P., Yang, S., Berg, A.C., Berg, T.L.: Modeling context in referring expressions. In: Computer Vision-ECCV 2016: 14th European Conference, Amsterdam, The Netherlands, October 11-14, 2016, Proceedings, Part II 14. pp. 69-85. Springer (2016)", + "58. Zhang, A., Zhao, L., Xie, C.W., Zheng, Y., Ji, W., Chua, T.S.: Next-chat: An lmm for chat, detection and segmentation. arXiv preprint arXiv:2311.04498 (2023)", + "59. Zhou, B., Zhao, H., Puig, X., Xiao, T., Fidler, S., Barriuso, A., Torralba, A.: Semantic understanding of scenes through the ade20k dataset. International Journal of Computer Vision 127, 302-321 (2019)", + "60. Zou, X., Dou, Z.Y., Yang, J., Gan, Z., Li, L., Li, C., Dai, X., Behl, H., Wang, J., Yuan, L., et al.: Generalized decoding for pixel, image, and language. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 15116-15127 (2023)", + "61. Zou, X., Yang, J., Zhang, H., Li, F., Li, L., Wang, J., Wang, L., Gao, J., Lee, Y.J.: Segment everything everywhere all at once. Advances in Neural Information Processing Systems 36 (2024)" + ], + "bbox": [ + 215, + 146, + 785, + 757 + ], + "page_idx": 17 + }, + { + "type": "page_number", + "text": "18", + "bbox": [ + 217, + 114, + 235, + 126 + ], + "page_idx": 17 + }, + { + "type": "header", + "text": "Z. Zhang, Y. Ma, E. Zhang, X. Bai", + "bbox": [ + 271, + 114, + 506, + 128 + ], + "page_idx": 17 + } +] \ No newline at end of file diff --git a/2024/PSALM_ Pixelwise Segmentation with Large Multi-modal Model/6e437084-134d-4cf9-a26b-095cfa791f83_model.json b/2024/PSALM_ Pixelwise Segmentation with Large Multi-modal Model/6e437084-134d-4cf9-a26b-095cfa791f83_model.json new file mode 100644 index 0000000000000000000000000000000000000000..c24d3dc3172ca2f2ff7a29f0790898ee434bbf42 --- /dev/null +++ b/2024/PSALM_ Pixelwise Segmentation with Large Multi-modal Model/6e437084-134d-4cf9-a26b-095cfa791f83_model.json @@ -0,0 +1,2359 @@ +[ + [ + { + "type": "title", + "bbox": [ + 0.245, + 0.141, + 0.761, + 0.187 + ], + "angle": 0, + "content": "PSALM: Pixelwise SegmentAtion with Large Multi-Modal Model" + }, + { + "type": "text", + "bbox": [ + 0.261, + 0.209, + 0.744, + 0.229 + ], + "angle": 0, + "content": "Zheng Zhang\\(^{1,2*}\\), Yeyao Ma\\(^{1*}\\), Enming Zhang\\(^{1*}\\), and Xiang Bai\\(^{1}\\)" + }, + { + "type": "text", + "bbox": [ + 0.333, + 0.239, + 0.67, + 0.268 + ], + "angle": 0, + "content": "1 Huazhong University of Science and Technology \n2 Microsoft Research Asia" + }, + { + "type": "text", + "bbox": [ + 0.261, + 0.301, + 0.744, + 0.566 + ], + "angle": 0, + "content": "Abstract. PSALM is a powerful extension of the Large Multi-modal Model (LMM) to address the segmentation task challenges. To overcome the limitation of the LMM being limited to textual output, PSALM incorporates a mask decoder and a well-designed input schema to handle a variety of segmentation tasks. This schema includes images, task instructions, conditional prompts, and mask tokens, which enable the model to generate and classify segmentation masks effectively. The flexible design of PSALM supports joint training across multiple datasets and tasks, leading to improved performance and task generalization. PSALM achieves superior results on several benchmarks, such as RefCOCO/RefCOCO+/RefCOCOg, COCO Panoptic Segmentation, and COCO-Interactive, and further exhibits zero-shot capabilities on unseen tasks, such as open-vocabulary segmentation, generalized referring expression segmentation and video object segmentation, making a significant step towards a GPT moment in computer vision. Through extensive experiments, PSALM demonstrates its potential to transform the domain of image segmentation, leveraging the robust visual understanding capabilities of LMMs as seen in natural language processing. Code and models are available at https://github.com/zamling/PSALM." + }, + { + "type": "text", + "bbox": [ + 0.262, + 0.577, + 0.741, + 0.593 + ], + "angle": 0, + "content": "Keywords: Segmentation \\(\\cdot\\) Large Multimodal Model \\(\\cdot\\) Visual Language" + }, + { + "type": "title", + "bbox": [ + 0.216, + 0.636, + 0.378, + 0.652 + ], + "angle": 0, + "content": "1 Introduction" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.666, + 0.788, + 0.786 + ], + "angle": 0, + "content": "Large multi-modal model (LMM) has ignited the dawn of the vision GPT [2] moment by making ground-breaking progress in various advanced visual understanding tasks by compressing image and language information into a single auto-regressive model. However, there are still many obstacles on the road to achieving vision GPT, and an important one being that current LMM can only perform text outputs, making it challenging to address the pixel-level image understanding problem directly, i.e., image segmentation, which is one of the most critical tasks in computer vision." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.787, + 0.788, + 0.819 + ], + "angle": 0, + "content": "Behind the obstacle is many challenges. First, the default output of the LMM is discrete tokens, and there is no apparent way to generate masks directly." + }, + { + "type": "page_footnote", + "bbox": [ + 0.232, + 0.825, + 0.548, + 0.841 + ], + "angle": 0, + "content": "* Equal contribution. \\( {}^{ \\boxtimes } \\) Corresponding author." + } + ], + [ + { + "type": "page_number", + "bbox": [ + 0.218, + 0.116, + 0.23, + 0.127 + ], + "angle": 0, + "content": "2" + }, + { + "type": "header", + "bbox": [ + 0.271, + 0.115, + 0.507, + 0.129 + ], + "angle": 0, + "content": "Z. Zhang, Y. Ma, E. Zhang, X. Bai" + }, + { + "type": "image", + "bbox": [ + 0.249, + 0.146, + 0.418, + 0.352 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.422, + 0.146, + 0.588, + 0.352 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.59, + 0.147, + 0.756, + 0.352 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.214, + 0.364, + 0.788, + 0.448 + ], + "angle": 0, + "content": "Fig. 1: PSALM has capability to handle multiple segmentation tasks in only one single model. We visualize some tasks, including Panoptic segmentation in COCO [24]; Open-Vocabulary instance segmentation in ADE20K [59]; Interactive segmentation in COCO-Interactive; Referring segmentation in RefCOCO [57]; Generalized referring segmentation in gRefCOCO [25]; Ego-exo correspondence in Ego-Exo4d [12]; Video object segmentation in DAVIS2017 [36]." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.478, + 0.788, + 0.599 + ], + "angle": 0, + "content": "Second, the variety of image segmentation tasks requires different forms of inputs and outputs. For example, semantic segmentation needs support inputs with different categories. Instance segmentation demands generate object IDs and the class confidence score for each object. Referring segmentation is supposed to have a language sentence as input and interactive segmentation has more varied inputs than the other tasks, which can be points, scribbles, bounding boxes, or masks. Third, unifying different segmentation tasks with a shared weight model is also challenging because different tasks require varied capabilities." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.599, + 0.788, + 0.689 + ], + "angle": 0, + "content": "In this work, we propose a method named PSALM (Pixelwise SegmentAtion with Large Multi-Modal Model) that aims to address the above challenges and extend the capabilities of LMM from text-output tasks to general segmentation tasks (Fig. 1 shows representative tasks). Specifically, PSALM externalizes a mask decoder on the top of LMM and designs a flexible input schema to unify different segmentation tasks into a single model." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.69, + 0.788, + 0.841 + ], + "angle": 0, + "content": "The input schema consists of four different parts: images, task instruction prompt, condition prompt, and a set of mask tokens, where the instruction prompt is a text sentence describing the task itself, condition prompt contains the additional necessary information to solve the task, either in terms of category names, sentence or visual features, and mask tokens are a set of learnable embeddings. All these inputs are fed into the LMM, and the resulting output mask tokens are further used as input by the mask generator to present mask proposals. Apart from producing the mask proposals, it is also necessary to predict the class of each segmentation mask or estimate a confidence score, which can be achieved by using the output embedding of the condition prompt as the" + } + ], + [ + { + "type": "header", + "bbox": [ + 0.297, + 0.115, + 0.733, + 0.129 + ], + "angle": 0, + "content": "PSALM: Pixelwise SegmentAtion with Large Multi-Modal Model" + }, + { + "type": "page_number", + "bbox": [ + 0.775, + 0.117, + 0.785, + 0.127 + ], + "angle": 0, + "content": "3" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.147, + 0.784, + 0.207 + ], + "angle": 0, + "content": "classifier weights to classify each mask proposal. In practice, we categorize conditions into category condition, sentence condition, and visual-prior condition, and present the corresponding methods to build the classifier weights according to the properties of each type of condition." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.208, + 0.789, + 0.449 + ], + "angle": 0, + "content": "Some other methods, represented by LISA [17], also aim to use LMM for segmentation tasks. However, these methods are usually designed for referring segmentation and fail to justify their ability to solve generalized segmentation tasks (see Tab. 1). In contrast, thanks to the generality and flexibility of the proposed architecture, PSALM can not only solve a variety of segmentation tasks but also be able to joint train on different tasks, which makes the model task generalizable while allowing the model to take full advantage of the intrinsic connections of different datasets/tasks to achieve better performance. Specifically, with the joint training of COCO Panoptic Segmentation [24], RefCOCO [57]/RefCOCO+/RefCOCOg [31], and COCO Interactive, we observe a significant performance improvement compared to training at different tasks individually, and therefore result in even better performance than other task-specific methods. On referring segmentation tasks, we outperform other LLM-based pixel reasoning methods (e.g., LISA, PixelLM [40] and GSVA [48]) on RefCOCO, RefCOCO+, and RefCOCOg, and it worth noting that we only use Phi-1.5 1.3B model [22] while others adopt Vicuna-7B [7] or LLama2-13B model [45]." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.45, + 0.789, + 0.585 + ], + "angle": 0, + "content": "The flexible design of architecture and input schema, multi-task joint-training, and the strong visual understanding capability of LMM not only make PSALM perform well on trained in-domain tasks but also enable generalizability to out-of-domain tasks in a zero-shot manner, i.e., directly dealing with unseen tasks without additional training. We test on three tasks: open-vocabulary segmentation, generalized referring expression segmentation, and video object segmentation. PSALM achieves promising zero-shot performance on these tasks. We believe this task-level generalizability is crucial, which is one of the key properties of the large language model for its success in natural language processing." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.585, + 0.789, + 0.66 + ], + "angle": 0, + "content": "Through extensive experiments on a variety of segmentation tasks, we show that presented PSALM has strong potential to address general image segmentation tasks and exhibits a certain degree of task generalization capability as LLM does in NLP. We believe that this work can inspire extensive research on realizing the GPT moment in computer vision and facilitate its arrival." + }, + { + "type": "title", + "bbox": [ + 0.215, + 0.682, + 0.393, + 0.697 + ], + "angle": 0, + "content": "2 Related works" + }, + { + "type": "title", + "bbox": [ + 0.215, + 0.711, + 0.479, + 0.726 + ], + "angle": 0, + "content": "2.1 Large Multimodal Models" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.734, + 0.788, + 0.841 + ], + "angle": 0, + "content": "With the release of GPT-V [32] and Gemini [44], more attention and efforts from open-source and research communities are shifting from large language models (LLM) to large multi-modal models (LMM). LLaVA [27], BLIP-2 [19], and Flamingo [1] are three representative works, where the core idea of both LLaVA and BLIP-2 is to map visual features into the input space of LLM to implement multi-modal capabilities, while Flamingo employs deeper feature fusion in the intermediate layers of LLM. Some works, such as Kosmos-2 [33]," + } + ], + [ + { + "type": "page_number", + "bbox": [ + 0.218, + 0.116, + 0.23, + 0.127 + ], + "angle": 0, + "content": "4" + }, + { + "type": "header", + "bbox": [ + 0.272, + 0.115, + 0.507, + 0.129 + ], + "angle": 0, + "content": "Z. Zhang, Y. Ma, E. Zhang, X. Bai" + }, + { + "type": "table_caption", + "bbox": [ + 0.216, + 0.145, + 0.788, + 0.187 + ], + "angle": 0, + "content": "Table 1: Capability for different methods. Our proposed PSALM can handle more segmentation tasks than other LLM-centric methods. LLM-centric methods can also deal with text generation tasks, which is hard for most vision-centric methods." + }, + { + "type": "table", + "bbox": [ + 0.268, + 0.199, + 0.734, + 0.343 + ], + "angle": 0, + "content": "
MethodsGeneric Seg.Referring Seg.Interactive Seg.OV Seg.
Vision centricMask2Former [6]✓
ODISE [49]✓✓
UNINEXT [53]✓✓✓
SEEM [61]✓✓✓✓
OMG-Seg [21]✓✓✓
LLM centricLISA [17]✓
GLAMM [39]✓
PixelLM [40]✓
PSALM (Ours)✓✓✓✓
" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.374, + 0.788, + 0.496 + ], + "angle": 0, + "content": "Shikra [4], and Ferret [55], further introduce object localization tasks into the LMM, while others, such as Emu [43], CogVLM [46], and DreamLLM [10], focus on how to integrate visual generation into the LMM. Monkey [23], OtterHD [18], and LLaVA-NeXT [26] explores mechanisms for processing large-size images and significantly improves performance on tasks such as OCR. However, these above methods are mainly designed for text output tasks or image generation and cannot directly deal with pixel-level understanding tasks such as image segmentation, which is different from ours and we can base on these models." + }, + { + "type": "title", + "bbox": [ + 0.216, + 0.529, + 0.492, + 0.545 + ], + "angle": 0, + "content": "2.2 Pixel Reasoning with LMM" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.566, + 0.788, + 0.734 + ], + "angle": 0, + "content": "Similar to our goal, some existing works attempt to enable LMMs to generate segmentation masks. LISA [17] is a pioneering work that uses a special seg token to aggregate information of a given sentence and use it as a prompt embedding in a SAM decoder to predict the segmentation mask. u-LLaVA [50] further supports object grounding tasks on the basis of LISA, and NExT-Chat [58] introduces richer inputs, such as bounding boxes. Furthermore, since LISA can only deal with a single object, many subsequent works that attempt to address the multi-object case, such as GLaMM [39], PerceptionGPT [34], PixelLM [40], GSVA [48], and LISA++ [54], all of which share the basic idea of introducing a seg token for each sentence describing a different object, and except PixelLM, all other methods are based on SAM decoder [15]." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.735, + 0.788, + 0.84 + ], + "angle": 0, + "content": "Although all these methods can generate masks, they are primarily designed for reference segmentation. In contrast, our method is designed for generalized segmentation tasks, which have diverse input and output requirements. In addition, the difference in goals also brings technical discrepancies: these methods use language models to directly generate the final segmentation masks, while our approach is closer to Mask2Former [6], which first generates mask proposals and then classifies the masks." + } + ], + [ + { + "type": "header", + "bbox": [ + 0.297, + 0.115, + 0.733, + 0.129 + ], + "angle": 0, + "content": "PSALM: Pixelwise SegmentAtion with Large Multi-Modal Model" + }, + { + "type": "page_number", + "bbox": [ + 0.775, + 0.117, + 0.785, + 0.127 + ], + "angle": 0, + "content": "5" + }, + { + "type": "image", + "bbox": [ + 0.245, + 0.145, + 0.761, + 0.289 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.371, + 0.298, + 0.63, + 0.311 + ], + "angle": 0, + "content": "Fig. 2: PSALM architecture overview." + }, + { + "type": "title", + "bbox": [ + 0.216, + 0.341, + 0.538, + 0.357 + ], + "angle": 0, + "content": "2.3 Unified Segmentation Framework" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.367, + 0.788, + 0.564 + ], + "angle": 0, + "content": "Another class of relevant explorations [21, 47, 53, 61] studies how to unify different segmentation tasks into a single framework. For instance, Mask2Former [6] presents a unified architecture of generic segmentation \\(^3\\) but with different models. OneFormer [14] further integrates these three tasks within a single model. UNINEXT [53] aims to unify instance-perception tasks and introduce text inputs and thus it can handle referring segmentation. X-Decoder [60] presents a flexible decoder architecture that can support generic segmentation, referring segmentation, retrieval, and image captioning. SEEM [61] designs a generic encoder-decoder to unify different segmentation tasks, where the encoder is used for projecting image, text, and human inputs into a joint visual-semantic space, and the decoder is used for mask prediction. However, all these works are not based on LMM, instead, they are mostly vision-centric models, i.e., usually designed for visual tasks only and thus cannot address language tasks very well." + }, + { + "type": "title", + "bbox": [ + 0.216, + 0.587, + 0.341, + 0.603 + ], + "angle": 0, + "content": "3 Methods" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.619, + 0.788, + 0.725 + ], + "angle": 0, + "content": "Fig. 2 provides an overview schematic of PSALM, which consists of a large multimodal model (LMM), a mask generator, and a flexible input schema designed for general segmentation tasks. The input schema has four different types of inputs: image, task instruction prompt, condition prompt, and a set of mask tokens. LMM processes the input tokens and the output embedding of mask tokens is further fed into the mask generator to generate masks. In the following, we will introduce our approach in detail." + }, + { + "type": "title", + "bbox": [ + 0.216, + 0.747, + 0.628, + 0.763 + ], + "angle": 0, + "content": "3.1 Large Multimodal Model and Input Schema" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.772, + 0.787, + 0.803 + ], + "angle": 0, + "content": "PSALM is built on large multimodal models (LMM), and there are many different LMM architectures, such as LLaVA [27], BLIP [20], and Flamingo [1]. Here," + }, + { + "type": "page_footnote", + "bbox": [ + 0.218, + 0.811, + 0.787, + 0.84 + ], + "angle": 0, + "content": "3 Generic segmentation includes semantic segmentation, instance segmentation, and panoptic segmentation." + } + ], + [ + { + "type": "page_number", + "bbox": [ + 0.218, + 0.116, + 0.23, + 0.127 + ], + "angle": 0, + "content": "6" + }, + { + "type": "header", + "bbox": [ + 0.272, + 0.115, + 0.506, + 0.128 + ], + "angle": 0, + "content": "Z. Zhang, Y. Ma, E. Zhang, X. Bai" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.147, + 0.784, + 0.192 + ], + "angle": 0, + "content": "we adopt the design of LLaVA because of its proven performance and simplicity, but the other LMM architectures are also compatible with our approach without any theoretical difficulties." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.194, + 0.785, + 0.36 + ], + "angle": 0, + "content": "The LMM used in our work has a visual encoder and pre-trained large language model (LLM). The two models are connected by a lightweight vision-language alignment model, which is a \\(3 \\times 3\\) convolution layer followed by a linear layer. The official LLaVA model uses a frozen CLIP model [37] as a visual coder, whose features lack the fine-grained information that is required for segmentation tasks [52]. Therefore, we train a customized LLaVA model by using the Swin Transformer [28], and limited by resources, we additionally replace the LLM from the Vicuna 7B model [7] to a smaller Phi-1.5 1.3B model [22]. Here, we applied only the first visual-language alignment stage of LLaVA by following its default settings. In our ablations, we found the alignment stage is crucial for open-vocabulary segmentation and referring segmentation tasks (Tab. 6)." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.363, + 0.785, + 0.453 + ], + "angle": 0, + "content": "Different segmentation tasks need different forms of inputs and outputs, which motivates us to present a flexible input schema to unify various requirements. In addition to the input image used in the visual encoder, our input schema has three other different types of inputs: task instruction prompt, condition prompt, and a set of mask tokens. We will introduce each of them and summarize the prompts used for all different tasks in the Appendix." + }, + { + "type": "text", + "bbox": [ + 0.215, + 0.456, + 0.787, + 0.532 + ], + "angle": 0, + "content": "Task Instruction Prompt. The task instruction prompt is usually a text sentence describing and specifying the model's task. For example, in panoptic segmentation, the task instruction can be \"You need to segment all objects. This is all the candidate categories.\" and in referring segmentation, the instruction can be \"Please segment according to the following instruction.\"" + }, + { + "type": "text", + "bbox": [ + 0.215, + 0.534, + 0.787, + 0.639 + ], + "angle": 0, + "content": "Condition Prompt. Some tasks require additional information to perform, e.g., panoptic segmentation needs specifying the candidate set of categories to be segmented, and interactive segmentation needs interactive inputs. The condition prompt is designed for these tasks. In addition to providing information, condition prompt also plays an important role in predicting categories or estimating confidence scores for each segmentation mask. In Sec. 3.3, we will discuss the design of condition prompts for different tasks in detail." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.642, + 0.785, + 0.763 + ], + "angle": 0, + "content": "Mask Token. The LLM is designed for text output and cannot directly generate segmentation masks. To bypass this challenge, we append a set of mask tokens after other inputs, and then these mask tokens are decoded to segmentation masks by a mask generator (will be introduced in Sec. 3.2). This design is inspired by Mask2Former [6], with the difference that the mask tokens in Mask2Former are used directly in the mask generator, whereas the mask tokens in our approach are first updated by the LMM and then used in the mask generator, and we found our approach leads better performance in practice (see Tab. 2)." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.765, + 0.785, + 0.84 + ], + "angle": 0, + "content": "Some works, such as LISA [17] and PixelLM [40], take similar seg tokens as input and employ a decoder to generate the masks. However, our objective is fundamentally different: in LISA and PixelLLM, seg tokens are used to generate the final prediction, while we generate mask proposals and further classify them based on condition prompts. Compared to the design of LISA and PixelLLM," + } + ], + [ + { + "type": "header", + "bbox": [ + 0.297, + 0.115, + 0.733, + 0.129 + ], + "angle": 0, + "content": "PSALM: Pixelwise SegmentAtion with Large Multi-Modal Model" + }, + { + "type": "page_number", + "bbox": [ + 0.775, + 0.117, + 0.785, + 0.127 + ], + "angle": 0, + "content": "7" + }, + { + "type": "image", + "bbox": [ + 0.248, + 0.144, + 0.761, + 0.291 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.215, + 0.301, + 0.788, + 0.343 + ], + "angle": 0, + "content": "Fig. 3: Detailed processing for different condition prompts. (a) shows the processing for category condition. (b) shows the processing for sentence condition. (c) shows the processing for visual-prior condition." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.371, + 0.785, + 0.403 + ], + "angle": 0, + "content": "our approach is more flexible and adaptable to more tasks, yet decoupling mask generation and classification is more efficient (see Tab. 3 and Tab. 4)." + }, + { + "type": "title", + "bbox": [ + 0.216, + 0.422, + 0.398, + 0.436 + ], + "angle": 0, + "content": "3.2 Mask Generator" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.446, + 0.784, + 0.493 + ], + "angle": 0, + "content": "The mask generator predicts the mask and their category probabilities from three inputs: a multi-level visual features \\(\\{v_{l}\\}_{l = 1}^{L}\\), a set of mask tokens \\(\\{q_i\\}_{i = 1}^N\\) and a set of condition embeddings \\(\\{c_k\\}_{k = 1}^K\\). It can be formally defined as:" + }, + { + "type": "equation", + "bbox": [ + 0.308, + 0.499, + 0.786, + 0.517 + ], + "angle": 0, + "content": "\\[\n\\{(m _ {i}, p _ {i}) \\} _ {i = 1} ^ {N} = \\operatorname {M a s k G e n e r a t o r} (\\{v _ {i} \\} _ {l = 1} ^ {L}, \\left\\{q _ {i} \\right\\} _ {i = 1} ^ {N}, \\left\\{c _ {k} \\right\\} _ {k = 1} ^ {K}), \\tag {1}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.523, + 0.788, + 0.646 + ], + "angle": 0, + "content": "where \\(m_{i}\\in \\mathbb{R}^{H\\times W}\\) is the i-th predicted segmentation mask and \\(p_i\\in \\mathbb{R}^K\\) is the corresponding category probability. In practice, the multi-level visual features are the internal features of the Swin visual encoder used in LMM. The design of the mask generator follows Mask2Former, which employs multi-scale deformable attention as a pixel decoder and a transformer-based mask decoder to generate segmentation masks. The class of each mask is predicted by the condition embedding \\(\\{c_k\\}\\), which is basically obtained from the output of the condition prompt, the obtaining method is slightly different for different types of conditions." + }, + { + "type": "title", + "bbox": [ + 0.216, + 0.666, + 0.504, + 0.683 + ], + "angle": 0, + "content": "3.3 Design of Condition Prompts" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.69, + 0.788, + 0.795 + ], + "angle": 0, + "content": "In our approach, the conditional prompt serves two important purposes: First, it provides the necessary information required to solve the task; Second, we use the output embedding of the conditional prompt in the LLM as classifier weights to predict the class of each segmentation mask. The design of the conditional prompt is closely tied to the type of task, and based on the information required for different tasks, we summarize three condition types: category condition, sentence condition, and visual-prior condition." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.795, + 0.788, + 0.842 + ], + "angle": 0, + "content": "Category Condition This condition type is used for tasks that need specifying a set of categories to be segmented, such as semantic segmentation, instance segmentation, and panoptic segmentation, and often needs to predict the" + } + ], + [ + { + "type": "page_number", + "bbox": [ + 0.218, + 0.116, + 0.23, + 0.127 + ], + "angle": 0, + "content": "8" + }, + { + "type": "header", + "bbox": [ + 0.272, + 0.115, + 0.506, + 0.128 + ], + "angle": 0, + "content": "Z. Zhang, Y. Ma, E. Zhang, X. Bai" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.147, + 0.784, + 0.192 + ], + "angle": 0, + "content": "class probability of each segmentation mask. Specifically, given a set of category names, we join them into a sentence by a comma separator, e.g., given three categories: person, bicycle, and car, the joint sentence is \"person, bicycle, car\"." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.193, + 0.785, + 0.299 + ], + "angle": 0, + "content": "The joint sentences are then proceeded by LMM to get the output embeddings, which can be further used to classify the predicted segmentation masks. Specifically, for each category, we select its corresponding output embeddings and apply the avg_pooling over them to obtain a condition embedding \\( c \\in \\mathbb{R}^D \\), where \\( D \\) is the embedding dimension, and thus condition embedding for all categories is a set \\( \\{c_k\\}_{k=1}^K \\), where \\( K \\) is the number of categories (see Fig. 3 (a)). This embedding set can be used by the mask generator to predict the class." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.3, + 0.787, + 0.42 + ], + "angle": 0, + "content": "Sentence Condition This condition is usually used for referring segmentation. Unlike the category condition, where category names are usually short, sentences are much longer, and not every word in the sentence is useful, so the avg_pooling is not the optimal choice here. Instead, we introduce a special [REF] token, which is appended after the condition sentence as an anchor to aggregate useful information, and the output embedding of [REF] token, i.e., the output features of LMM on the location of [REF] token, is used as condition embedding \\( c \\) and used by mask generator, as shown in Fig. 3 (b)." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.422, + 0.788, + 0.664 + ], + "angle": 0, + "content": "Visual-Prior Condition We formulate most interaction (e.g., point, mask, box, or scribble) used in interactive segmentation tasks as the visual-prior condition. Taking the scribble as an example, we first generate a binary scribble curve with a width of 5 pixels on a binary map, whose size is the same as the input image of the Swin vision encoder, and then apply the binary map to the output features of the vision-language alignment model, performing average mask pooling by upsampling the output feature map to the size of the binary map. The average pooled feature is used as the visual-prior condition and fed into LMM. If there are multiple interactions, the process is repeated for each, yielding multiple pooled features as inputs, each feature separated using a strategy akin to the category condition. For other types of interaction, we adopt similar approaches. Specifically, for box or mask, we directly apply the pooling operator by treating them as binary masks, and for point, we bold it to a 10-pixel circle and then apply the mask pooling. After that, we use the output embedding of the visual-prior condition as the classifier weight in the mask generator to estimate the confidence of each mask proposal, as shown in Fig. 3 (c)." + }, + { + "type": "title", + "bbox": [ + 0.216, + 0.69, + 0.424, + 0.705 + ], + "angle": 0, + "content": "3.4 Training Objectives" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.72, + 0.788, + 0.841 + ], + "angle": 0, + "content": "The training process of PSALM has two stages: In the first stage, we train the visual language alignment model following LLaVA setting by freezing the vision encoder and LMM; In the second stage, we only freeze the vision encoder and fine-tune all other modules, including the mask generator. Similarly to Mask2Former, we use matching loss in the second stage training, i.e., we use bipartite matching to find the optimal assignments between the mask proposals and the ground truth masks by minimizing the matching loss and use these assignments to perform training. The loss has two terms: \\(\\mathcal{L} = \\mathcal{L}_{mask} + \\mathcal{L}_{cls}\\)," + } + ], + [ + { + "type": "header", + "bbox": [ + 0.296, + 0.114, + 0.733, + 0.13 + ], + "angle": 0, + "content": "PSALM: Pixelwise SegmentAtion with Large Multi-Modal Model" + }, + { + "type": "page_number", + "bbox": [ + 0.775, + 0.116, + 0.787, + 0.128 + ], + "angle": 0, + "content": "9" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.147, + 0.788, + 0.207 + ], + "angle": 0, + "content": "where \\(\\mathcal{L}_{mask}\\) indicates the mask loss which comprises a pixel-level Binary Cross-Entropy (BCE) loss and Dice loss [42], while the \\(\\mathcal{L}_{cls}\\) indicates the category classification loss, and we use Cross-Entropy loss for category condition and BCE loss for other cases." + }, + { + "type": "title", + "bbox": [ + 0.215, + 0.231, + 0.376, + 0.248 + ], + "angle": 0, + "content": "4 Experiments" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.262, + 0.788, + 0.294 + ], + "angle": 0, + "content": "We conduct extensive experiments on various datasets and tasks to analyze PSALM. Details of the datasets and implementations are given in the Appendix." + }, + { + "type": "title", + "bbox": [ + 0.215, + 0.315, + 0.344, + 0.33 + ], + "angle": 0, + "content": "4.1 Ablations" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.34, + 0.788, + 0.432 + ], + "angle": 0, + "content": "We first ablate key designs and present behind insights in this section. To better show how different designs affect the performance on a wide range of tasks, we mainly report the results on three in-domain benchmarks: COCO Panoptic Segmentation (COCO-Pan), RefCOCO-val (RefCOCO), and COCO Interactive Segmentation with point inputs (COCO-Point), and one out-of-domain benchmark: open-vocabulary instance segmentation on ADE20K-150 (A150-OV)." + }, + { + "type": "table_caption", + "bbox": [ + 0.215, + 0.457, + 0.561, + 0.514 + ], + "angle": 0, + "content": "Table 2: Ablation on the design of mask tokens. w.LLM: use mask tokens as inputs of LLM. Prefix: place mask tokens at the front. Suffix: place mask tokens at the end." + }, + { + "type": "table", + "bbox": [ + 0.222, + 0.525, + 0.553, + 0.603 + ], + "angle": 0, + "content": "
Mask TokensCOCO-PanRefCOCOCOCO-PointA150-OV
w.LLMPos.PQcIoUmIoUmAP
✓Suffix55.176.153.39.3
✗-54.8(↓0.3)74.3(↓1.8)53.1(↓0.2)8.2(↓1.1)
✓Prefix55.0(↓0.1)75.1(↓1.0)53.0(↓0.3)7.8(↓1.5)
" + }, + { + "type": "table_caption", + "bbox": [ + 0.582, + 0.471, + 0.789, + 0.514 + ], + "angle": 0, + "content": "Table 3: Effect of decouple design in COCO Semantic Segmentation." + }, + { + "type": "table", + "bbox": [ + 0.606, + 0.526, + 0.763, + 0.589 + ], + "angle": 0, + "content": "
DecouplemIoUfwIoU
✓66.572.5
X42.735.0
" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.629, + 0.788, + 0.795 + ], + "angle": 0, + "content": "Design of Mask Tokens. In our approach, we use a set of mask tokens to predict the mask proposal. In practice, we have found that using mask tokens as inputs to the LLM leads to better performance than applying them directly to the mask generator, which is the default method of Mask2Former. Tab. 2 shows the results, where the direct use of mask tokens leads to a noticeable performance degradation in RefCOCO and A150-OV. We believe this is because using mask tokens as input leads to a better awareness of the information needed for the task and thus improves performance, which is essential for these two tasks. For better validation, we placed mask tokens before conditional prompts and task instruction prompts and found a similar performance drop to that of not using mask tokens in LMM, which further supports our hypothesis." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.796, + 0.788, + 0.842 + ], + "angle": 0, + "content": "Compared to LISA and other methods that use seg token to generate final segmentation results directly, our mask proposal approach has three advantages: First, our design is more flexible and thus can be applied to a wider range" + } + ], + [ + { + "type": "page_number", + "bbox": [ + 0.218, + 0.116, + 0.236, + 0.127 + ], + "angle": 0, + "content": "10" + }, + { + "type": "header", + "bbox": [ + 0.272, + 0.114, + 0.507, + 0.129 + ], + "angle": 0, + "content": "Z. Zhang, Y. Ma, E. Zhang, X. Bai" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.147, + 0.788, + 0.284 + ], + "angle": 0, + "content": "of segmentation tasks, especially tasks that require predicting category or confidence scores; Second, our design decouples the mask prediction and classification, which alleviating the learning difficulties for some tasks. In Tab. 3, we study how the decouple design affects the semantic segmentation performance on COCO Semantic Segmentation\\(^{4}\\), and we found our decouple design is significantly better\\(^{5}\\). Third, the mask proposals allow multiple masks to be generated for a single instance, which makes the mask accuracy superior to solutions like LISA that only predict a single mask. Tab. 4 shows that using more mask proposals on RefCOCO gives a clear improvement over using a single mask." + }, + { + "type": "table_caption", + "bbox": [ + 0.215, + 0.31, + 0.45, + 0.365 + ], + "angle": 0, + "content": "Table 4: Effects of number of mask tokens in RefCOCO(cIoU). The results of LISA are listed as a reference." + }, + { + "type": "table", + "bbox": [ + 0.241, + 0.378, + 0.421, + 0.442 + ], + "angle": 0, + "content": "
#Mask TokenvaltestAtestB
LISA74.176.571.1
175.378.072.2
10076.578.573.4
" + }, + { + "type": "table_caption", + "bbox": [ + 0.47, + 0.318, + 0.788, + 0.348 + ], + "angle": 0, + "content": "Table 5: Ablation on different designs for condition prompts." + }, + { + "type": "table", + "bbox": [ + 0.474, + 0.359, + 0.798, + 0.435 + ], + "angle": 0, + "content": "
ConditionCOCO-PanRefCOCOA150-OV
CategorySentencePQcIoUmAP
avg_poolingavg_pooling55.175.39.2
[REF]54.976.18.4
avg_pooling[REF]55.176.19.3
" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.472, + 0.789, + 0.64 + ], + "angle": 0, + "content": "Design of Condition Prompts. Another key design in PSALM is the condition prompt, particularly the way we obtain the condition embeddings, which are used as classifier weights in the mask generator to predict the class of mask proposals. As described in Sec. 3.3, for category condition, we use the avg_pooling over output embeddings of each class name as the condition embedding, and for sentence condition, we adopt a [REF] token to aggregate useful information. Tab. 5 shows the ablation, where we first tried to use the same design for all conditions and found that avg_pooling performed slightly better on COCO-Pan, with a larger improvement on A150-OV, while [REF] worked better on RefCOCO. We further used different designs and found that the advantages of each design are preserved, and the best overall performance is achieved." + }, + { + "type": "table_caption", + "bbox": [ + 0.215, + 0.666, + 0.492, + 0.694 + ], + "angle": 0, + "content": "Table 6: Ablation on effect of vision-language alingment." + }, + { + "type": "table", + "bbox": [ + 0.22, + 0.705, + 0.49, + 0.757 + ], + "angle": 0, + "content": "
VL AlignmentCOCO-Pan PQRefCOCO cIoUCOCO-Point mIoUA150-OV mAP
✓55.176.153.39.3
✗54.9(↓0.2)71.7(↓4.4)53.0 (↓0.3)8.2(↓1.1)
" + }, + { + "type": "table_caption", + "bbox": [ + 0.525, + 0.67, + 0.771, + 0.685 + ], + "angle": 0, + "content": "Table 7: Ablation on joint training." + }, + { + "type": "table", + "bbox": [ + 0.514, + 0.697, + 0.792, + 0.752 + ], + "angle": 0, + "content": "
AblationCOCO-PanRefCOCOCOCO-PointA150-OV
PQcIoUmIoUmAP
Task Specific Train55.676.562.27.0
Joint Train55.983.664.19.0
Δ+0.3+7.1+1.9+2.0
" + }, + { + "type": "page_footnote", + "bbox": [ + 0.218, + 0.797, + 0.786, + 0.826 + ], + "angle": 0, + "content": "This benchmark is introduced by Mask2Former, which is composed by merging instances belonging to the same class together in COCO Panoptic Segmentation." + }, + { + "type": "page_footnote", + "bbox": [ + 0.222, + 0.826, + 0.525, + 0.84 + ], + "angle": 0, + "content": "5 More experimental details are in Appendix." + }, + { + "type": "list", + "bbox": [ + 0.218, + 0.797, + 0.786, + 0.84 + ], + "angle": 0, + "content": null + } + ], + [ + { + "type": "header", + "bbox": [ + 0.297, + 0.115, + 0.732, + 0.129 + ], + "angle": 0, + "content": "PSALM: Pixelwise SegmentAtion with Large Multi-Modal Model" + }, + { + "type": "page_number", + "bbox": [ + 0.769, + 0.117, + 0.784, + 0.127 + ], + "angle": 0, + "content": "11" + }, + { + "type": "table_caption", + "bbox": [ + 0.214, + 0.145, + 0.788, + 0.244 + ], + "angle": 0, + "content": "Table 8: Comparison with the state-of-the-art methods on three referring image segmentation benchmarks with cIoU. (ft) denotes models further finetuned on RefCOCO+/g after mix training. We abbreviate the datasets: COCO(C) [24], LVIS(L) [13], RefCOCO(RC) [57], Object365(O365) [41], Video segmentation datasets(V), ADE20K(A) [59], COCO-Stuff(CS) [3], PACO-LVIS(PL) [38], PASCAL-Part(PP) [5], GranD(G) [39], VOC2010(VOC) [11], Visual Genome(VG) [16], Flicker30k(F30K) [35], MUSE(M) [40], gRefCOCO(gRC) [25], COCO-Interactive(Cl)." + }, + { + "type": "table", + "bbox": [ + 0.243, + 0.255, + 0.756, + 0.403 + ], + "angle": 0, + "content": "
MethodSegmentation DataLLM TypeRefCOCORefCOCO+RefCOCOg
valtestAtestBvaltestAtestBvaltest
SEEM-L [61]C, L, RC-------65.6-
UNINEXT-L [53]O365, C, RC, V-80.382.677.870.074.962.673.473.7
UNINEXT-H [53]O365, C, RC, V-82.283.481.372.576.466.274.776.4
LISA [17]A, CS, RC, PL, PPVicuna-7B74.176.571.162.467.456.566.468.5
LISA(ft) [17]A, CS, RC, PL, PPVicuna-7B74.979.172.365.170.858.167.970.6
GLaMM [39]G, RCVicuna-7B79.583.276.972.678.764.674.274.9
u-LLaVA [50]A, CS, RC, PL, VOCVicuna-7B80.482.777.872.276.666.874.875.6
PerceptionGPT [34]RC, VG, F30kVicuna-13B75.379.172.168.974.061.970.771.9
PixelLM [40]A, CS, RC, PL, MLlama2-13B73.076.568.266.371.758.369.370.5
GSVA [48]A, CS, RC, PL, PP, gRCLlama2-13B77.779.974.268.071.561.573.273.9
GSVA(ft) [48]A, CS, RC, PL, PP, gRCLlama2-13B79.281.777.170.373.863.675.777.0
PSALMC, RC, CIPhi-1.5 (1.3B)83.684.781.672.975.570.173.874.4
" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.431, + 0.788, + 0.581 + ], + "angle": 0, + "content": "Importance of VL-alignments. The visual-language alignment stage (i.e., first training stage) is to project the visual features to the text input space, and it is an important step towards making the LLM understand images. In Tab. 6, we examine the impact of this stage, and we found that without VL alignment, the performance of all four tasks becomes worse, with the performance of A150-OV and RefCOCO being significantly affected, for example, the A150-OV dropped by -1.1 mAP and RefCOCO even dropped by -4.4 cIoU, probably because these two tasks require a strong requirement on understanding the relationship between vision and language. This result also suggests that the VL alignment is essential, and the LMM-based segmentation models have strong potential." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.582, + 0.788, + 0.765 + ], + "angle": 0, + "content": "Joint Training. Our architecture design and input schema help integrate various segmentation tasks so that they can be trained on one model. Tab. 7 shows the effect of this joint training on different tasks. For the task-specific models, we perform the training on the corresponding task data for \\(18\\mathrm{k}\\) iterations. In contrast, the joint training setting (see implementation section for details) has a total of \\(56\\mathrm{k}\\) training iterations, which corresponds to \\(14\\mathrm{k}\\) iterations per task. The results show that joint training of different tasks greatly improves the performance. This suggests that learning between tasks is mutually beneficial, which is also the secret of success in LLM. For example, the generic segmentation task helps refine the mask prediction in the referring segmentation, and referring expressions also enhance the model's ability to recognize more unseen categories, which in turn improves the performance of open-vocabulary segmentation tasks." + }, + { + "type": "title", + "bbox": [ + 0.215, + 0.786, + 0.649, + 0.801 + ], + "angle": 0, + "content": "4.2 System-Level Comparison on In-Domain Tasks" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.81, + 0.787, + 0.84 + ], + "angle": 0, + "content": "In this section, we compared our model on three in-domain tasks with other state-of-the-arts to illustrate the effectiveness of our approach." + } + ], + [ + { + "type": "page_number", + "bbox": [ + 0.218, + 0.116, + 0.236, + 0.127 + ], + "angle": 0, + "content": "12" + }, + { + "type": "header", + "bbox": [ + 0.272, + 0.115, + 0.507, + 0.129 + ], + "angle": 0, + "content": "Z. Zhang, Y. Ma, E. Zhang, X. Bai" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.147, + 0.788, + 0.328 + ], + "angle": 0, + "content": "Referring Segmentation. Most existing works aimed at getting LMMs to perform image segmentation are designed for reference segmentation tasks. We compare PSALM with other works on RefCOCO, RefCOCO+, and RefCOCOg, and Tab. 8 shows the results. Owing to the generalized and flexible design of PSALM and the advantages of joint training on multiple tasks and datasets, our system was able to achieve state-of-the-art (SOTA) performance on RefCOCO and RefCOCO+, and competitive performance on RefCOCOg with STOA, despite being driven by LLM with only 1.3B parameters. It is worth noting unlike methods such as LISA and GSVA, which may achieve improvements through task-specific fine-tuning (gray-labeled results), PSALM does not perform additional fine-tuning but still achieves better performance on RefCOCO and RefCOCO+ than their fine-tuned models." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.332, + 0.788, + 0.455 + ], + "angle": 0, + "content": "Generic Segmentation. We evaluate PSALM with state-of-the-art methods on the COCO Panoptic Segmentation validation set (Tab. 9). Here, we follow the evaluation protocol used in Mask2Former to report PQ, which is the main metric for panoptic segmentation, mAP on thing classes for instance segmentation, and mIoU by merging instance masks from the same category for semantic segmentation. Compared to other methods, PSALM achieves comparable performance at similar visual backbone sizes, demonstrating that PSALM is a powerful architecture, even when compared to approaches designed for specific tasks." + }, + { + "type": "table_caption", + "bbox": [ + 0.214, + 0.488, + 0.788, + 0.53 + ], + "angle": 0, + "content": "Table 9: Comparison with the state-of-the-art methods on Panoptic COCO-val. We abbreviate the datasets: COCO-SAM(CM) [21], VIPSeg (VIP) [29], while others following Tab. 8." + }, + { + "type": "table", + "bbox": [ + 0.301, + 0.542, + 0.699, + 0.638 + ], + "angle": 0, + "content": "
MethodBackboneSeg. DataPQmAPmIoU
Mask2Former [6]Swin-BC55.145.265.1
Mask2Former [6]Swin-LC57.848.667.4
X-Decoder [60]DaViT-BC, L, RC56.245.866.0
SEEM [61]DaViT-BC, L, RC56.146.466.3
OMG-Seg [21]ConvNeXt-XXLC, VIP, CM, V55.4--
PSALMSwin-BC, RC, CI55.945.766.6
" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.674, + 0.788, + 0.84 + ], + "angle": 0, + "content": "Interactive Segmentation. We also evaluate PSALM in the interactive segmentation tasks. Since the task does not have a well-developed dataset containing all four instructions, previous works have typically used the in-house dataset, thus we re-evaluate other methods on the COCO interactive validation set. The results are shown in Tab. 10, and PSALM achieves leading performance on point, scribble, and mask instructions than all other methods, while on box instruction, SAM performs better on mIoU but worse on cIoU, and we hypothesize might be caused by the different distribution of training data, and the fact that SAM is trained on SA-1B [15], which has a much larger data scale than what we used. In addition, we also report the official results of SEEM, which only evaluated 600 samples from the COCO validation set as a reference." + } + ], + [ + { + "type": "header", + "bbox": [ + 0.296, + 0.114, + 0.732, + 0.129 + ], + "angle": 0, + "content": "PSALM: Pixelwise SegmentAtion with Large Multi-Modal Model" + }, + { + "type": "page_number", + "bbox": [ + 0.769, + 0.116, + 0.785, + 0.127 + ], + "angle": 0, + "content": "13" + }, + { + "type": "table_caption", + "bbox": [ + 0.214, + 0.145, + 0.788, + 0.2 + ], + "angle": 0, + "content": "Table 10: Comparison with the state-of-the-art methods on COCO-Interactive. The results of SEEM-B* is the result reported in official paper, which is evaluated on 600 random samples of COCO-val, while all others are evaluated on all samples of COCO-val. Abbreviations for each dataset are the same as Tab. 9." + }, + { + "type": "table", + "bbox": [ + 0.283, + 0.212, + 0.719, + 0.32 + ], + "angle": 0, + "content": "
MethodSeg. DataPointScribbleBoxMask
mIoUcIoUmIoUcIoUmIoUcIoUmIoUcIoU
SAM-B [15]SA-1B48.733.6--73.768.7--
SAM-L [15]SA-1B51.837.7--76.671.6--
SEEM-B [61]C, L, RC47.857.843.044.044.942.148.465.0
SEEM-B* [61]C, L, RC81.7-83.5-75.7-76.0-
OMG-Seg [21]C, VIP, CM, V59.3-------
PSALMC, RC, CI64.374.066.980.067.380.967.682.4
" + }, + { + "type": "title", + "bbox": [ + 0.215, + 0.345, + 0.603, + 0.361 + ], + "angle": 0, + "content": "4.3 Generalizability on Out-of-Domain Tasks" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.369, + 0.788, + 0.506 + ], + "angle": 0, + "content": "Thanks to the flexible design of architecture and input schema, multi-task joint-training, and the strong visual understanding capability of LMM, PSALM shows excellent performance on in-domain tasks, but more importantly, PSALM also demonstrates great potential to generalize to out-of-domain tasks in the zero-shot setting. In this section, we conduct experiments on three different out-of-domain tasks: open-vocabulary segmentation, generalized referring expression segmentation, and video object segmentation. We also tested the zero-shot result of the correspondence benchmark in Ego-Exo4D [12] and Video Object Segmentation in the Appendix." + }, + { + "type": "table_caption", + "bbox": [ + 0.214, + 0.53, + 0.788, + 0.588 + ], + "angle": 0, + "content": "Table 11: Comparison with the state-of-the-art methods on open-vocabulary instance segmentation and semantic segmentation benchmarks. We use mAP for instance segmentation and mIoU for semantic segmentation. We abbreviate the datasets: Pascal Context-459(PC459) [30], Pascal Context-59(PC59) [30], Pascal VOC-20(PAS20) [11]" + }, + { + "type": "table", + "bbox": [ + 0.326, + 0.598, + 0.675, + 0.709 + ], + "angle": 0, + "content": "
MethodOV Instance Seg.OV Semantic Seg.
A150CityscapesPC459A150PC59PAS20
MaskCLIP [9]6.0-10.023.745.9-
ODISE [49]14.4-14.529.957.3-
SAN [51]10.6-17.133.360.295.5
PSALM9.020.510.218.248.581.3
PSALM+LVIS13.919.314.024.457.295.0
" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.735, + 0.803, + 0.842 + ], + "angle": 0, + "content": "Open-Vocabulary Segmentation. We first evaluate PSALM in open-vocabulary segmentation tasks, which require the model to have the ability to deal with unseen categories in training. Here, we conduct experiments on both open-vocabulary instance segmentation and open-vocabulary semantic segmentation, and Tab. 11 shows the results. Without any special design, PSALM achieves reasonably good performance, although it is still worse than the best specific method in this task, such as SAN, but we believe that PSALM has a strong" + } + ], + [ + { + "type": "page_number", + "bbox": [ + 0.218, + 0.116, + 0.236, + 0.127 + ], + "angle": 0, + "content": "14" + }, + { + "type": "header", + "bbox": [ + 0.272, + 0.114, + 0.507, + 0.129 + ], + "angle": 0, + "content": "Z. Zhang, Y. Ma, E. Zhang, X. Bai" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.147, + 0.788, + 0.253 + ], + "angle": 0, + "content": "potential to be further improved by adding more diverse training data, which is advantages of our method. We have also made a preliminary attempt by further involving the LVIS dataset, and as we expected, the performance has significantly improvements. In addition, existing open-vocabulary segmentation methods are built upon the CLIP model or diffusion model, while our approach is based on LMM models, which is a new path and attempt to bring new inspiration to the community, which we believe is even more important than the performance." + }, + { + "type": "table_caption", + "bbox": [ + 0.214, + 0.278, + 0.788, + 0.308 + ], + "angle": 0, + "content": "Table 12: Our method's zero-shot performance on gRefCOCO. (ft) denotes models further fine-tuned on gRefCOCO after mix training." + }, + { + "type": "table", + "bbox": [ + 0.301, + 0.318, + 0.699, + 0.466 + ], + "angle": 0, + "content": "
MethodsLLM TypeZero-ShotvaltestAtestB
cIoUgIoUcIoUgIoUcIoUgIoU
MattNet [56]-X47.548.258.759.345.346.1
LTS [8]-X52.352.761.962.649.950.4
ReLA [25]-X62.463.669.370.059.961.0
LISA [17]Vicuna-7BX38.732.252.648.544.839.7
LISA(ft) [17]Vicuna-7BX61.763.369.270.160.361.3
GSVA [48]Vicuna-7BX61.763.369.270.160.361.3
GSVA(ft) [48]Vicuna-7BX63.366.569.971.160.562.2
PSALMPhi-1.5 (1.3B)✓42.043.352.454.550.652.5
" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.493, + 0.788, + 0.63 + ], + "angle": 0, + "content": "Generalized Referring Expression Segmentation. The referring segmentation datasets used in training contain only a single object, however, the design of the mask proposal allows PSALM to directly address multi-target without any further training or fine-tuning. We evaluate the gRefCOCO benchmark which contains multiple segment targets. In practice, given an expression, we compute the similarity with all mask proposals and retain all masks with similarity greater than 0.6 as foreground. Tab. 12 shows the results, PSALM also achieved very promising performance, even outperforming the LISA version that only pre-trained on gRefCOCO but without task-specific fine-tuning." + }, + { + "type": "title", + "bbox": [ + 0.215, + 0.651, + 0.36, + 0.667 + ], + "angle": 0, + "content": "5 Conclusion" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.681, + 0.788, + 0.757 + ], + "angle": 0, + "content": "The PSALM proposed in this study extends the capability of LMM from text output tasks to image segmentation, addresses the output limitations of the LMM, and unifies various segmentation tasks. PSALM exhibits excellent performance in multiple in-domain tasks, and its generalization ability in out-of-domain tasks further underscores its potential." + }, + { + "type": "title", + "bbox": [ + 0.215, + 0.779, + 0.395, + 0.797 + ], + "angle": 0, + "content": "Acknowledgments" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.81, + 0.787, + 0.84 + ], + "angle": 0, + "content": "This work was supported by the National Science Fund for Distinguished Young Scholars of China (Grant No.62225603)." + } + ], + [ + { + "type": "header", + "bbox": [ + 0.297, + 0.115, + 0.732, + 0.129 + ], + "angle": 0, + "content": "PSALM: Pixelwise SegmentAtion with Large Multi-Modal Model" + }, + { + "type": "page_number", + "bbox": [ + 0.769, + 0.117, + 0.785, + 0.127 + ], + "angle": 0, + "content": "15" + }, + { + "type": "title", + "bbox": [ + 0.218, + 0.145, + 0.323, + 0.161 + ], + "angle": 0, + "content": "References" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.173, + 0.786, + 0.228 + ], + "angle": 0, + "content": "1. Alayrac, J.B., Donahue, J., Luc, P., Miech, A., Barr, I., Hasson, Y., Lenc, K., Mensch, A., Millican, K., Reynolds, M., et al.: Flamingo: a visual language model for few-shot learning. Advances in Neural Information Processing Systems 35, 23716-23736 (2022)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.229, + 0.786, + 0.269 + ], + "angle": 0, + "content": "2. Brown, T., Mann, B., Ryder, N., Subbiah, M., Kaplan, J.D., Dhariwal, P., Neelakantan, A., Shyam, P., Sastry, G., Askell, A., et al.: Language models are few-shot learners. Advances in neural information processing systems 33, 1877-1901 (2020)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.27, + 0.786, + 0.31 + ], + "angle": 0, + "content": "3. Caesar, H., Uijlings, J., Ferrari, V.: Coco-stuff: Thing and stuff classes in context. In: Proceedings of the IEEE conference on computer vision and pattern recognition. pp. 1209-1218 (2018)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.31, + 0.786, + 0.35 + ], + "angle": 0, + "content": "4. Chen, K., Zhang, Z., Zeng, W., Zhang, R., Zhu, F., Zhao, R.: Shikra: Unleashing multimodal llm's referential dialogue magic. arXiv preprint arXiv:2306.15195 (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.35, + 0.786, + 0.405 + ], + "angle": 0, + "content": "5. Chen, X., Mottaghi, R., Liu, X., Fidler, S., Urtasun, R., Yuille, A.: Detect what you can: Detecting and representing objects using holistic models and body parts. In: Proceedings of the IEEE conference on computer vision and pattern recognition. pp. 1971-1978 (2014)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.405, + 0.786, + 0.459 + ], + "angle": 0, + "content": "6. Cheng, B., Misra, I., Schwing, A.G., Kirillov, A., Girdhar, R.: Masked-attention mask transformer for universal image segmentation. In: Proceedings of the IEEE/CVF conference on computer vision and pattern recognition. pp. 1290-1299 (2022)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.459, + 0.786, + 0.514 + ], + "angle": 0, + "content": "7. Chiang, W.L., Li, Z., Lin, Z., Sheng, Y., Wu, Z., Zhang, H., Zheng, L., Zhuang, S., Zhuang, Y., Gonzalez, J.E., et al.: Vicuna: An open-source chatbot impressing gpt-4 with \\(90\\%\\) chatgpt quality. See https://vicuna.lmsys.org (accessed 14 April 2023) (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.514, + 0.786, + 0.555 + ], + "angle": 0, + "content": "8. Ding, H., Liu, C., Wang, S., Jiang, X.: Vision-language transformer and query generation for referring segmentation. In: Proceedings of the IEEE/CVF International Conference on Computer Vision. pp. 16321-16330 (2021)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.555, + 0.786, + 0.582 + ], + "angle": 0, + "content": "9. Ding, Z., Wang, J., Tu, Z.: Open-vocabulary universal image segmentation with maskclip (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.582, + 0.786, + 0.623 + ], + "angle": 0, + "content": "0. Dong, R., Han, C., Peng, Y., Qi, Z., Ge, Z., Yang, J., Zhao, L., Sun, J., Zhou, H., Wei, H., et al.: Dreamllm: Synergistic multimodal comprehension and creation. arXiv preprint arXiv:2309.11499 (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.623, + 0.786, + 0.663 + ], + "angle": 0, + "content": "1. Everingham, M., Van Gool, L., Williams, C.K.I., Winn, J., Zisserman, A.: The PASCAL Visual Object Classes Challenge 2010 (VOC2010) Results. http://www.pascal-network.org/challenges/VOC/voc2010/workshop/index.html" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.663, + 0.786, + 0.718 + ], + "angle": 0, + "content": "2. Grauman, K., Westbury, A., Torresani, L., Kitani, K., Malik, J., Afouras, T., Ashutosh, K., Baiyya, V., Bansal, S., Boote, B., et al.: Ego-exo4d: Understanding skilled human activity from first-and third-person perspectives. arXiv preprint arXiv:2311.18259 (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.718, + 0.786, + 0.759 + ], + "angle": 0, + "content": "3. Gupta, A., Dollar, P., Girshick, R.: Lvis: A dataset for large vocabulary instance segmentation. In: Proceedings of the IEEE/CVF conference on computer vision and pattern recognition. pp. 5356-5364 (2019)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.759, + 0.786, + 0.8 + ], + "angle": 0, + "content": "4. Jain, J., Li, J., Chiu, M.T., Hassani, A., Orlov, N., Shi, H.: Oneformer: One transformer to rule universal image segmentation. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 2989-2998 (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.8, + 0.786, + 0.84 + ], + "angle": 0, + "content": "5. Kirillov, A., Mintun, E., Ravi, N., Mao, H., Rolland, C., Gustafson, L., Xiao, T., Whitehead, S., Berg, A.C., Lo, W.Y., et al.: Segment anything. arXiv preprint arXiv:2304.02643 (2023)" + }, + { + "type": "list", + "bbox": [ + 0.226, + 0.173, + 0.786, + 0.84 + ], + "angle": 0, + "content": null + } + ], + [ + { + "type": "page_number", + "bbox": [ + 0.218, + 0.116, + 0.236, + 0.127 + ], + "angle": 0, + "content": "16" + }, + { + "type": "header", + "bbox": [ + 0.272, + 0.115, + 0.506, + 0.128 + ], + "angle": 0, + "content": "Z. Zhang, Y. Ma, E. Zhang, X. Bai" + }, + { + "type": "ref_text", + "bbox": [ + 0.218, + 0.147, + 0.787, + 0.203 + ], + "angle": 0, + "content": "16. Krishna, R., Zhu, Y., Groth, O., Johnson, J., Hata, K., Kravitz, J., Chen, S., Kalantidis, Y., Li, L.J., Shamma, D.A., et al.: Visual genome: Connecting language and vision using crowdsourced dense image annotations. International journal of computer vision 123, 32-73 (2017)" + }, + { + "type": "ref_text", + "bbox": [ + 0.218, + 0.204, + 0.787, + 0.231 + ], + "angle": 0, + "content": "17. Lai, X., Tian, Z., Chen, Y., Li, Y., Yuan, Y., Liu, S., Jia, J.: Lisa: Reasoning segmentation via large language model. arXiv preprint arXiv:2308.00692 (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.218, + 0.232, + 0.786, + 0.259 + ], + "angle": 0, + "content": "18. Li, B., Zhang, P., Yang, J., Zhang, Y., Pu, F., Liu, Z.: Otterhd: A high-resolution multi-modality model. arXiv preprint arXiv:2311.04219 (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.219, + 0.26, + 0.787, + 0.3 + ], + "angle": 0, + "content": "19. Li, J., Li, D., Savarese, S., Hoi, S.: Blip-2: Bootstrapping language-image pretraining with frozen image encoders and large language models. arXiv preprint arXiv:2301.12597 (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.218, + 0.301, + 0.787, + 0.342 + ], + "angle": 0, + "content": "20. Li, J., Li, D., Xiong, C., Hoi, S.: Blip: Bootstrapping language-image pre-training for unified vision-language understanding and generation. In: International Conference on Machine Learning. pp. 12888-12900. PMLR (2022)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.343, + 0.786, + 0.383 + ], + "angle": 0, + "content": "21. Li, X., Yuan, H., Li, W., Ding, H., Wu, S., Zhang, W., Li, Y., Chen, K., Loy, C.C.: Omg-seg: Is one model good enough for all segmentation? arXiv preprint arXiv:2401.10229 (2024)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.384, + 0.786, + 0.411 + ], + "angle": 0, + "content": "22. Li, Y., Bubeck, S., Eldan, R., Del Giorno, A., Gunasekar, S., Lee, Y.T.: Textbooks are all you need ii: phi-1.5 technical report. arXiv preprint arXiv:2309.05463 (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.412, + 0.787, + 0.467 + ], + "angle": 0, + "content": "23. Li, Z., Yang, B., Liu, Q., Ma, Z., Zhang, S., Yang, J., Sun, Y., Liu, Y., Bai, X.: Monkey: Image resolution and text label are important things for large multi-modal models. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 26763-26773 (2024)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.468, + 0.787, + 0.522 + ], + "angle": 0, + "content": "24. Lin, T.Y., Maire, M., Belongie, S., Hays, J., Perona, P., Ramanan, D., Dollar, P., Zitnick, C.L.: Microsoft coco: Common objects in context. In: Computer Vision-ECCV 2014: 13th European Conference, Zurich, Switzerland, September 6-12, 2014, Proceedings, Part V 13. pp. 740-755. Springer (2014)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.523, + 0.787, + 0.563 + ], + "angle": 0, + "content": "25. Liu, C., Ding, H., Jiang, X.: Gres: Generalized referring expression segmentation. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 23592-23601 (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.565, + 0.787, + 0.591 + ], + "angle": 0, + "content": "26. Liu, H., Li, C., Li, Y., Li, B., Zhang, Y., Shen, S., Lee, Y.J.: Llava-last: Improved reasoning,OCR,and world knowledge (2024)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.592, + 0.787, + 0.619 + ], + "angle": 0, + "content": "27. Liu, H., Li, C., Wu, Q., Lee, Y.J.: Visual instruction tuning. Advances in neural information processing systems 36 (2024)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.62, + 0.787, + 0.674 + ], + "angle": 0, + "content": "28. Liu, Z., Lin, Y., Cao, Y., Hu, H., Wei, Y., Zhang, Z., Lin, S., Guo, B.: Swin transformer: Hierarchical vision transformer using shifted windows. In: Proceedings of the IEEE/CVF international conference on computer vision. pp. 10012-10022 (2021)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.675, + 0.787, + 0.729 + ], + "angle": 0, + "content": "29. Miao, J., Wang, X., Wu, Y., Li, W., Zhang, X., Wei, Y., Yang, Y.: Large-scale video panoptic segmentation in the wild: A benchmark. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 21033-21043 (2022)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.73, + 0.787, + 0.785 + ], + "angle": 0, + "content": "30. Mottaghi, R., Chen, X., Liu, X., Cho, N.G., Lee, S.W., Fidler, S., Urtasun, R., Yuille, A.: The role of context for object detection and semantic segmentation in the wild. In: IEEE Conference on Computer Vision and Pattern Recognition (CVPR) (2014)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.786, + 0.787, + 0.84 + ], + "angle": 0, + "content": "31. Nagaraja, V.K., Morariu, V.I., Davis, L.S.: Modeling context between objects for referring expression understanding. In: Computer Vision-ECCV 2016: 14th European Conference, Amsterdam, The Netherlands, October 11-14, 2016, Proceedings, Part IV 14. pp. 792-807. Springer (2016)" + }, + { + "type": "list", + "bbox": [ + 0.217, + 0.147, + 0.787, + 0.84 + ], + "angle": 0, + "content": null + } + ], + [ + { + "type": "header", + "bbox": [ + 0.297, + 0.115, + 0.732, + 0.129 + ], + "angle": 0, + "content": "PSALM: Pixelwise SegmentAtion with Large Multi-Modal Model" + }, + { + "type": "page_number", + "bbox": [ + 0.769, + 0.117, + 0.785, + 0.127 + ], + "angle": 0, + "content": "17" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.147, + 0.528, + 0.162 + ], + "angle": 0, + "content": "32. OpenAI: Gpt-4v(ision) system card (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.164, + 0.787, + 0.205 + ], + "angle": 0, + "content": "33. Peng, Z., Wang, W., Dong, L., Hao, Y., Huang, S., Ma, S., Wei, F.: Kosmos-2: Grounding multimodal large language models to the world. arXiv preprint arXiv:2306.14824 (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.206, + 0.787, + 0.234 + ], + "angle": 0, + "content": "34. Pi, R., Yao, L., Gao, J., Zhang, J., Zhang, T.: Perceptiongpt: Effectively fusing visual perception into llm. arXiv preprint arXiv:2311.06612 (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.235, + 0.787, + 0.29 + ], + "angle": 0, + "content": "35. Plummer, B.A., Wang, L., Cervantes, C.M., Caicedo, J.C., Hockenmaier, J., Lazebnik, S.: Flickr30k entities: Collecting region-to-phrase correspondences for richer image-to-sentence models. In: Proceedings of the IEEE international conference on computer vision. pp. 2641-2649 (2015)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.291, + 0.787, + 0.332 + ], + "angle": 0, + "content": "36. Pont-Tuset, J., Perazzi, F., Caelles, S., Arbeláez, P., Sorkine-Hornung, A., Van Gool, L.: The 2017 davis challenge on video object segmentation. arXiv preprint arXiv:1704.00675 (2017)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.334, + 0.787, + 0.389 + ], + "angle": 0, + "content": "37. Radford, A., Kim, J.W., Hallacy, C., Ramesh, A., Goh, G., Agarwal, S., Sastry, G., Askell, A., Mishkin, P., Clark, J., et al.: Learning transferable visual models from natural language supervision. In: International conference on machine learning. pp. 8748-8763. PMLR (2021)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.39, + 0.787, + 0.445 + ], + "angle": 0, + "content": "38. Ramanathan, V., Kalia, A., Petrovic, V., Wen, Y., Zheng, B., Guo, B., Wang, R., Marquez, A., Kovvuri, R., Kadian, A., et al.: Paco: Parts and attributes of common objects. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 7141-7151 (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.446, + 0.787, + 0.487 + ], + "angle": 0, + "content": "39. Rasheed, H., Maaz, M., Shaji, S., Shaker, A., Khan, S., Cholakkal, H., Anwer, R.M., Xing, E., Yang, M.H., Khan, F.S.: Glamm: Pixel grounding large multimodal model. arXiv preprint arXiv:2311.03356 (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.488, + 0.787, + 0.516 + ], + "angle": 0, + "content": "40. Ren, Z., Huang, Z., Wei, Y., Zhao, Y., Fu, D., Feng, J., Jin, X.: Pixelm: Pixel reasoning with large multimodal model. arXiv preprint arXiv:2312.02228 (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.517, + 0.787, + 0.559 + ], + "angle": 0, + "content": "41. Shao, S., Li, Z., Zhang, T., Peng, C., Yu, G., Zhang, X., Li, J., Sun, J.: Objects365: A large-scale, high-quality dataset for object detection. In: Proceedings of the IEEE/CVF international conference on computer vision. pp. 8430-8439 (2019)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.56, + 0.787, + 0.642 + ], + "angle": 0, + "content": "42. Sudre, C.H., Li, W., Vercauteren, T., Ourselin, S., Jorge Cardoso, M.: Generalised dice overlap as a deep learning loss function for highly unbalanced segmentations. In: Deep Learning in Medical Image Analysis and Multimodal Learning for Clinical Decision Support: Third International Workshop, DLMIA 2017, and 7th International Workshop, ML-CDS 2017, Held in Conjunction with MICCAI 2017, Quebec City, QC, Canada, September 14, Proceedings 3. pp. 240-248. Springer (2017)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.643, + 0.787, + 0.685 + ], + "angle": 0, + "content": "43. Sun, Q., Yu, Q., Cui, Y., Zhang, F., Zhang, X., Wang, Y., Gao, H., Liu, J., Huang, T., Wang, X.: Generative pretraining in multimodality. arXiv preprint arXiv:2307.05222 (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.686, + 0.787, + 0.727 + ], + "angle": 0, + "content": "44. Team, G., Anil, R., Borgeaud, S., Wu, Y., Alayrac, J.B., Yu, J., Soricut, R., Schalkwyk, J., Dai, A.M., Hauth, A., et al.: Gemini: a family of highly capable multimodal models. arXiv preprint arXiv:2312.11805 (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.728, + 0.787, + 0.77 + ], + "angle": 0, + "content": "45. Touvron, H., Martin, L., Stone, K., Albert, P., Almahairi, A., Babaei, Y., Bashlykov, N., Batra, S., Bhargava, P., Bhosale, S., et al.: Llama 2: Open foundation and fine-tuned chat models. arXiv preprint arXiv:2307.09288 (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.771, + 0.787, + 0.812 + ], + "angle": 0, + "content": "46. Wang, W., Lv, Q., Yu, W., Hong, W., Qi, J., Wang, Y., Ji, J., Yang, Z., Zhao, L., Song, X., et al.: Cogvlm: Visual expert for pretrained language models. arXiv preprint arXiv:2311.03079 (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.813, + 0.787, + 0.841 + ], + "angle": 0, + "content": "47. Wu, J., Jiang, Y., Liu, Q., Yuan, Z., Bai, X., Bai, S.: General object foundation model for images and videos at scale. arXiv preprint arXiv:2312.09158 (2023)" + }, + { + "type": "list", + "bbox": [ + 0.217, + 0.147, + 0.787, + 0.841 + ], + "angle": 0, + "content": null + } + ], + [ + { + "type": "page_number", + "bbox": [ + 0.218, + 0.116, + 0.236, + 0.127 + ], + "angle": 0, + "content": "18" + }, + { + "type": "header", + "bbox": [ + 0.272, + 0.115, + 0.507, + 0.129 + ], + "angle": 0, + "content": "Z. Zhang, Y. Ma, E. Zhang, X. Bai" + }, + { + "type": "ref_text", + "bbox": [ + 0.218, + 0.147, + 0.785, + 0.189 + ], + "angle": 0, + "content": "48. Xia, Z., Han, D., Han, Y., Pan, X., Song, S., Huang, G.: Gsva: Generalized segmentation via multimodal large language models. arXiv preprint arXiv:2312.10103 (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.218, + 0.19, + 0.787, + 0.245 + ], + "angle": 0, + "content": "49. Xu, J., Liu, S., Vahdat, A., Byeon, W., Wang, X., De Mello, S.: Open-vocabulary panoptic segmentation with text-to-image diffusion models. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 2955-2966 (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.246, + 0.785, + 0.273 + ], + "angle": 0, + "content": "50. Xu, J., Xu, L., Yang, Y., Li, X., Xie, Y., Huang, Y.J., Li, Y.: u-llava: Unifying multi-modal tasks via large language model. arXiv preprint arXiv:2311.05348 (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.273, + 0.785, + 0.314 + ], + "angle": 0, + "content": "51. Xu, M., Zhang, Z., Wei, F., Hu, H., Bai, X.: Side adapter network for open-vocabulary semantic segmentation. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 2945-2954 (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.315, + 0.785, + 0.357 + ], + "angle": 0, + "content": "52. Xu, M., Zhang, Z., Wei, F., Lin, Y., Cao, Y., Hu, H., Bai, X.: A simple baseline for open-vocabulary semantic segmentation with pre-trained vision-language model. In: European Conference on Computer Vision. pp. 736-753. Springer (2022)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.357, + 0.785, + 0.398 + ], + "angle": 0, + "content": "53. Yan, B., Jiang, Y., Wu, J., Wang, D., Luo, P., Yuan, Z., Lu, H.: Universal instance perception as object discovery and retrieval. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 15325-15336 (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.398, + 0.785, + 0.438 + ], + "angle": 0, + "content": "54. Yang, S., Qu, T., Lai, X., Tian, Z., Peng, B., Liu, S., Jia, J.: An improved baseline for reasoning segmentation with large language model. arXiv preprint arXiv:2312.17240 (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.439, + 0.785, + 0.481 + ], + "angle": 0, + "content": "55. You, H., Zhang, H., Gan, Z., Du, X., Zhang, B., Wang, Z., Cao, L., Chang, S.F., Yang, Y.: Ferret: Refer and ground anything anywhere at any granularity. arXiv preprint arXiv:2310.07704 (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.481, + 0.785, + 0.535 + ], + "angle": 0, + "content": "56. Yu, L., Lin, Z., Shen, X., Yang, J., Lu, X., Bansal, M., Berg, T.L.: Matttnet: Modular attention network for referring expression comprehension. In: Proceedings of the IEEE conference on computer vision and pattern recognition. pp. 1307-1315 (2018)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.536, + 0.785, + 0.59 + ], + "angle": 0, + "content": "57. Yu, L., Poirson, P., Yang, S., Berg, A.C., Berg, T.L.: Modeling context in referring expressions. In: Computer Vision-ECCV 2016: 14th European Conference, Amsterdam, The Netherlands, October 11-14, 2016, Proceedings, Part II 14. pp. 69-85. Springer (2016)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.591, + 0.785, + 0.619 + ], + "angle": 0, + "content": "58. Zhang, A., Zhao, L., Xie, C.W., Zheng, Y., Ji, W., Chua, T.S.: Next-chat: An lmm for chat, detection and segmentation. arXiv preprint arXiv:2311.04498 (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.62, + 0.785, + 0.66 + ], + "angle": 0, + "content": "59. Zhou, B., Zhao, H., Puig, X., Xiao, T., Fidler, S., Barriuso, A., Torralba, A.: Semantic understanding of scenes through the ade20k dataset. International Journal of Computer Vision 127, 302-321 (2019)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.66, + 0.785, + 0.716 + ], + "angle": 0, + "content": "60. Zou, X., Dou, Z.Y., Yang, J., Gan, Z., Li, L., Li, C., Dai, X., Behl, H., Wang, J., Yuan, L., et al.: Generalized decoding for pixel, image, and language. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 15116-15127 (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.716, + 0.785, + 0.758 + ], + "angle": 0, + "content": "61. Zou, X., Yang, J., Zhang, H., Li, F., Li, L., Wang, J., Wang, L., Gao, J., Lee, Y.J.: Segment everything everywhere all at once. Advances in Neural Information Processing Systems 36 (2024)" + }, + { + "type": "list", + "bbox": [ + 0.217, + 0.147, + 0.787, + 0.758 + ], + "angle": 0, + "content": null + } + ] +] \ No newline at end of file diff --git a/2024/PSALM_ Pixelwise Segmentation with Large Multi-modal Model/6e437084-134d-4cf9-a26b-095cfa791f83_origin.pdf b/2024/PSALM_ Pixelwise Segmentation with Large Multi-modal Model/6e437084-134d-4cf9-a26b-095cfa791f83_origin.pdf new file mode 100644 index 0000000000000000000000000000000000000000..c75da834e02611482a4ca3e5b98bb4cf47d4821b --- /dev/null +++ b/2024/PSALM_ Pixelwise Segmentation with Large Multi-modal Model/6e437084-134d-4cf9-a26b-095cfa791f83_origin.pdf @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:e197b757ba3143d58b18d3300bbd8e0a682dfcf9157700b88bc4f164dfc9a241 +size 2203420 diff --git a/2024/PSALM_ Pixelwise Segmentation with Large Multi-modal Model/full.md b/2024/PSALM_ Pixelwise Segmentation with Large Multi-modal Model/full.md new file mode 100644 index 0000000000000000000000000000000000000000..8cd77a2198d4c32abdc993b798ed1c84ecca4146 --- /dev/null +++ b/2024/PSALM_ Pixelwise Segmentation with Large Multi-modal Model/full.md @@ -0,0 +1,278 @@ +# PSALM: Pixelwise SegmentAtion with Large Multi-Modal Model + +Zheng Zhang $^{1,2*}$ , Yeyao Ma $^{1*}$ , Enming Zhang $^{1*}$ , and Xiang Bai $^{1}$ + +1 Huazhong University of Science and Technology +2 Microsoft Research Asia + +Abstract. PSALM is a powerful extension of the Large Multi-modal Model (LMM) to address the segmentation task challenges. To overcome the limitation of the LMM being limited to textual output, PSALM incorporates a mask decoder and a well-designed input schema to handle a variety of segmentation tasks. This schema includes images, task instructions, conditional prompts, and mask tokens, which enable the model to generate and classify segmentation masks effectively. The flexible design of PSALM supports joint training across multiple datasets and tasks, leading to improved performance and task generalization. PSALM achieves superior results on several benchmarks, such as RefCOCO/RefCOCO+/RefCOCOg, COCO Panoptic Segmentation, and COCO-Interactive, and further exhibits zero-shot capabilities on unseen tasks, such as open-vocabulary segmentation, generalized referring expression segmentation and video object segmentation, making a significant step towards a GPT moment in computer vision. Through extensive experiments, PSALM demonstrates its potential to transform the domain of image segmentation, leveraging the robust visual understanding capabilities of LMMs as seen in natural language processing. Code and models are available at https://github.com/zamling/PSALM. + +Keywords: Segmentation $\cdot$ Large Multimodal Model $\cdot$ Visual Language + +# 1 Introduction + +Large multi-modal model (LMM) has ignited the dawn of the vision GPT [2] moment by making ground-breaking progress in various advanced visual understanding tasks by compressing image and language information into a single auto-regressive model. However, there are still many obstacles on the road to achieving vision GPT, and an important one being that current LMM can only perform text outputs, making it challenging to address the pixel-level image understanding problem directly, i.e., image segmentation, which is one of the most critical tasks in computer vision. + +Behind the obstacle is many challenges. First, the default output of the LMM is discrete tokens, and there is no apparent way to generate masks directly. + +![](images/23d5814cdcbed14d1acaa71e7a63e13276e2b67be2a3ad5a5679354e03c77d9a.jpg) +Fig. 1: PSALM has capability to handle multiple segmentation tasks in only one single model. We visualize some tasks, including Panoptic segmentation in COCO [24]; Open-Vocabulary instance segmentation in ADE20K [59]; Interactive segmentation in COCO-Interactive; Referring segmentation in RefCOCO [57]; Generalized referring segmentation in gRefCOCO [25]; Ego-exo correspondence in Ego-Exo4d [12]; Video object segmentation in DAVIS2017 [36]. + +![](images/83b9485f68a18c6a8252495eb4d5c0671a7c640098271a829cf37648d10c9078.jpg) + +![](images/a3f75599d0a30bf7e3f9c386a776707c0a0f38ecf1968a487738425cde68973c.jpg) + +Second, the variety of image segmentation tasks requires different forms of inputs and outputs. For example, semantic segmentation needs support inputs with different categories. Instance segmentation demands generate object IDs and the class confidence score for each object. Referring segmentation is supposed to have a language sentence as input and interactive segmentation has more varied inputs than the other tasks, which can be points, scribbles, bounding boxes, or masks. Third, unifying different segmentation tasks with a shared weight model is also challenging because different tasks require varied capabilities. + +In this work, we propose a method named PSALM (Pixelwise SegmentAtion with Large Multi-Modal Model) that aims to address the above challenges and extend the capabilities of LMM from text-output tasks to general segmentation tasks (Fig. 1 shows representative tasks). Specifically, PSALM externalizes a mask decoder on the top of LMM and designs a flexible input schema to unify different segmentation tasks into a single model. + +The input schema consists of four different parts: images, task instruction prompt, condition prompt, and a set of mask tokens, where the instruction prompt is a text sentence describing the task itself, condition prompt contains the additional necessary information to solve the task, either in terms of category names, sentence or visual features, and mask tokens are a set of learnable embeddings. All these inputs are fed into the LMM, and the resulting output mask tokens are further used as input by the mask generator to present mask proposals. Apart from producing the mask proposals, it is also necessary to predict the class of each segmentation mask or estimate a confidence score, which can be achieved by using the output embedding of the condition prompt as the + +classifier weights to classify each mask proposal. In practice, we categorize conditions into category condition, sentence condition, and visual-prior condition, and present the corresponding methods to build the classifier weights according to the properties of each type of condition. + +Some other methods, represented by LISA [17], also aim to use LMM for segmentation tasks. However, these methods are usually designed for referring segmentation and fail to justify their ability to solve generalized segmentation tasks (see Tab. 1). In contrast, thanks to the generality and flexibility of the proposed architecture, PSALM can not only solve a variety of segmentation tasks but also be able to joint train on different tasks, which makes the model task generalizable while allowing the model to take full advantage of the intrinsic connections of different datasets/tasks to achieve better performance. Specifically, with the joint training of COCO Panoptic Segmentation [24], RefCOCO [57]/RefCOCO+/RefCOCOg [31], and COCO Interactive, we observe a significant performance improvement compared to training at different tasks individually, and therefore result in even better performance than other task-specific methods. On referring segmentation tasks, we outperform other LLM-based pixel reasoning methods (e.g., LISA, PixelLM [40] and GSVA [48]) on RefCOCO, RefCOCO+, and RefCOCOg, and it worth noting that we only use Phi-1.5 1.3B model [22] while others adopt Vicuna-7B [7] or LLama2-13B model [45]. + +The flexible design of architecture and input schema, multi-task joint-training, and the strong visual understanding capability of LMM not only make PSALM perform well on trained in-domain tasks but also enable generalizability to out-of-domain tasks in a zero-shot manner, i.e., directly dealing with unseen tasks without additional training. We test on three tasks: open-vocabulary segmentation, generalized referring expression segmentation, and video object segmentation. PSALM achieves promising zero-shot performance on these tasks. We believe this task-level generalizability is crucial, which is one of the key properties of the large language model for its success in natural language processing. + +Through extensive experiments on a variety of segmentation tasks, we show that presented PSALM has strong potential to address general image segmentation tasks and exhibits a certain degree of task generalization capability as LLM does in NLP. We believe that this work can inspire extensive research on realizing the GPT moment in computer vision and facilitate its arrival. + +# 2 Related works + +# 2.1 Large Multimodal Models + +With the release of GPT-V [32] and Gemini [44], more attention and efforts from open-source and research communities are shifting from large language models (LLM) to large multi-modal models (LMM). LLaVA [27], BLIP-2 [19], and Flamingo [1] are three representative works, where the core idea of both LLaVA and BLIP-2 is to map visual features into the input space of LLM to implement multi-modal capabilities, while Flamingo employs deeper feature fusion in the intermediate layers of LLM. Some works, such as Kosmos-2 [33], + +Table 1: Capability for different methods. Our proposed PSALM can handle more segmentation tasks than other LLM-centric methods. LLM-centric methods can also deal with text generation tasks, which is hard for most vision-centric methods. + +
MethodsGeneric Seg.Referring Seg.Interactive Seg.OV Seg.
Vision centricMask2Former [6]✓
ODISE [49]✓✓
UNINEXT [53]✓✓✓
SEEM [61]✓✓✓✓
OMG-Seg [21]✓✓✓
LLM centricLISA [17]✓
GLAMM [39]✓
PixelLM [40]✓
PSALM (Ours)✓✓✓✓
+ +Shikra [4], and Ferret [55], further introduce object localization tasks into the LMM, while others, such as Emu [43], CogVLM [46], and DreamLLM [10], focus on how to integrate visual generation into the LMM. Monkey [23], OtterHD [18], and LLaVA-NeXT [26] explores mechanisms for processing large-size images and significantly improves performance on tasks such as OCR. However, these above methods are mainly designed for text output tasks or image generation and cannot directly deal with pixel-level understanding tasks such as image segmentation, which is different from ours and we can base on these models. + +# 2.2 Pixel Reasoning with LMM + +Similar to our goal, some existing works attempt to enable LMMs to generate segmentation masks. LISA [17] is a pioneering work that uses a special seg token to aggregate information of a given sentence and use it as a prompt embedding in a SAM decoder to predict the segmentation mask. u-LLaVA [50] further supports object grounding tasks on the basis of LISA, and NExT-Chat [58] introduces richer inputs, such as bounding boxes. Furthermore, since LISA can only deal with a single object, many subsequent works that attempt to address the multi-object case, such as GLaMM [39], PerceptionGPT [34], PixelLM [40], GSVA [48], and LISA++ [54], all of which share the basic idea of introducing a seg token for each sentence describing a different object, and except PixelLM, all other methods are based on SAM decoder [15]. + +Although all these methods can generate masks, they are primarily designed for reference segmentation. In contrast, our method is designed for generalized segmentation tasks, which have diverse input and output requirements. In addition, the difference in goals also brings technical discrepancies: these methods use language models to directly generate the final segmentation masks, while our approach is closer to Mask2Former [6], which first generates mask proposals and then classifies the masks. + +![](images/8fb1ade5589cb18ace88fbb6c48b90fd4d69c3bf5d21a24f69a3cfe6798dbd51.jpg) +Fig. 2: PSALM architecture overview. + +# 2.3 Unified Segmentation Framework + +Another class of relevant explorations [21, 47, 53, 61] studies how to unify different segmentation tasks into a single framework. For instance, Mask2Former [6] presents a unified architecture of generic segmentation $^3$ but with different models. OneFormer [14] further integrates these three tasks within a single model. UNINEXT [53] aims to unify instance-perception tasks and introduce text inputs and thus it can handle referring segmentation. X-Decoder [60] presents a flexible decoder architecture that can support generic segmentation, referring segmentation, retrieval, and image captioning. SEEM [61] designs a generic encoder-decoder to unify different segmentation tasks, where the encoder is used for projecting image, text, and human inputs into a joint visual-semantic space, and the decoder is used for mask prediction. However, all these works are not based on LMM, instead, they are mostly vision-centric models, i.e., usually designed for visual tasks only and thus cannot address language tasks very well. + +# 3 Methods + +Fig. 2 provides an overview schematic of PSALM, which consists of a large multimodal model (LMM), a mask generator, and a flexible input schema designed for general segmentation tasks. The input schema has four different types of inputs: image, task instruction prompt, condition prompt, and a set of mask tokens. LMM processes the input tokens and the output embedding of mask tokens is further fed into the mask generator to generate masks. In the following, we will introduce our approach in detail. + +# 3.1 Large Multimodal Model and Input Schema + +PSALM is built on large multimodal models (LMM), and there are many different LMM architectures, such as LLaVA [27], BLIP [20], and Flamingo [1]. Here, + +we adopt the design of LLaVA because of its proven performance and simplicity, but the other LMM architectures are also compatible with our approach without any theoretical difficulties. + +The LMM used in our work has a visual encoder and pre-trained large language model (LLM). The two models are connected by a lightweight vision-language alignment model, which is a $3 \times 3$ convolution layer followed by a linear layer. The official LLaVA model uses a frozen CLIP model [37] as a visual coder, whose features lack the fine-grained information that is required for segmentation tasks [52]. Therefore, we train a customized LLaVA model by using the Swin Transformer [28], and limited by resources, we additionally replace the LLM from the Vicuna 7B model [7] to a smaller Phi-1.5 1.3B model [22]. Here, we applied only the first visual-language alignment stage of LLaVA by following its default settings. In our ablations, we found the alignment stage is crucial for open-vocabulary segmentation and referring segmentation tasks (Tab. 6). + +Different segmentation tasks need different forms of inputs and outputs, which motivates us to present a flexible input schema to unify various requirements. In addition to the input image used in the visual encoder, our input schema has three other different types of inputs: task instruction prompt, condition prompt, and a set of mask tokens. We will introduce each of them and summarize the prompts used for all different tasks in the Appendix. + +Task Instruction Prompt. The task instruction prompt is usually a text sentence describing and specifying the model's task. For example, in panoptic segmentation, the task instruction can be "You need to segment all objects. This is all the candidate categories." and in referring segmentation, the instruction can be "Please segment according to the following instruction." + +Condition Prompt. Some tasks require additional information to perform, e.g., panoptic segmentation needs specifying the candidate set of categories to be segmented, and interactive segmentation needs interactive inputs. The condition prompt is designed for these tasks. In addition to providing information, condition prompt also plays an important role in predicting categories or estimating confidence scores for each segmentation mask. In Sec. 3.3, we will discuss the design of condition prompts for different tasks in detail. + +Mask Token. The LLM is designed for text output and cannot directly generate segmentation masks. To bypass this challenge, we append a set of mask tokens after other inputs, and then these mask tokens are decoded to segmentation masks by a mask generator (will be introduced in Sec. 3.2). This design is inspired by Mask2Former [6], with the difference that the mask tokens in Mask2Former are used directly in the mask generator, whereas the mask tokens in our approach are first updated by the LMM and then used in the mask generator, and we found our approach leads better performance in practice (see Tab. 2). + +Some works, such as LISA [17] and PixelLM [40], take similar seg tokens as input and employ a decoder to generate the masks. However, our objective is fundamentally different: in LISA and PixelLLM, seg tokens are used to generate the final prediction, while we generate mask proposals and further classify them based on condition prompts. Compared to the design of LISA and PixelLLM, + +![](images/f125fa85c77e2aafe356b5cbf1636857c587eafbef960a94bbcba5e61c8dfef3.jpg) +Fig. 3: Detailed processing for different condition prompts. (a) shows the processing for category condition. (b) shows the processing for sentence condition. (c) shows the processing for visual-prior condition. + +our approach is more flexible and adaptable to more tasks, yet decoupling mask generation and classification is more efficient (see Tab. 3 and Tab. 4). + +# 3.2 Mask Generator + +The mask generator predicts the mask and their category probabilities from three inputs: a multi-level visual features $\{v_{l}\}_{l = 1}^{L}$ , a set of mask tokens $\{q_i\}_{i = 1}^N$ and a set of condition embeddings $\{c_k\}_{k = 1}^K$ . It can be formally defined as: + +$$ +\{(m _ {i}, p _ {i}) \} _ {i = 1} ^ {N} = \operatorname {M a s k G e n e r a t o r} (\{v _ {i} \} _ {l = 1} ^ {L}, \left\{q _ {i} \right\} _ {i = 1} ^ {N}, \left\{c _ {k} \right\} _ {k = 1} ^ {K}), \tag {1} +$$ + +where $m_{i}\in \mathbb{R}^{H\times W}$ is the i-th predicted segmentation mask and $p_i\in \mathbb{R}^K$ is the corresponding category probability. In practice, the multi-level visual features are the internal features of the Swin visual encoder used in LMM. The design of the mask generator follows Mask2Former, which employs multi-scale deformable attention as a pixel decoder and a transformer-based mask decoder to generate segmentation masks. The class of each mask is predicted by the condition embedding $\{c_k\}$ , which is basically obtained from the output of the condition prompt, the obtaining method is slightly different for different types of conditions. + +# 3.3 Design of Condition Prompts + +In our approach, the conditional prompt serves two important purposes: First, it provides the necessary information required to solve the task; Second, we use the output embedding of the conditional prompt in the LLM as classifier weights to predict the class of each segmentation mask. The design of the conditional prompt is closely tied to the type of task, and based on the information required for different tasks, we summarize three condition types: category condition, sentence condition, and visual-prior condition. + +Category Condition This condition type is used for tasks that need specifying a set of categories to be segmented, such as semantic segmentation, instance segmentation, and panoptic segmentation, and often needs to predict the + +class probability of each segmentation mask. Specifically, given a set of category names, we join them into a sentence by a comma separator, e.g., given three categories: person, bicycle, and car, the joint sentence is "person, bicycle, car". + +The joint sentences are then proceeded by LMM to get the output embeddings, which can be further used to classify the predicted segmentation masks. Specifically, for each category, we select its corresponding output embeddings and apply the avg_pooling over them to obtain a condition embedding $c \in \mathbb{R}^D$ , where $D$ is the embedding dimension, and thus condition embedding for all categories is a set $\{c_k\}_{k=1}^K$ , where $K$ is the number of categories (see Fig. 3 (a)). This embedding set can be used by the mask generator to predict the class. + +Sentence Condition This condition is usually used for referring segmentation. Unlike the category condition, where category names are usually short, sentences are much longer, and not every word in the sentence is useful, so the avg_pooling is not the optimal choice here. Instead, we introduce a special [REF] token, which is appended after the condition sentence as an anchor to aggregate useful information, and the output embedding of [REF] token, i.e., the output features of LMM on the location of [REF] token, is used as condition embedding $c$ and used by mask generator, as shown in Fig. 3 (b). + +Visual-Prior Condition We formulate most interaction (e.g., point, mask, box, or scribble) used in interactive segmentation tasks as the visual-prior condition. Taking the scribble as an example, we first generate a binary scribble curve with a width of 5 pixels on a binary map, whose size is the same as the input image of the Swin vision encoder, and then apply the binary map to the output features of the vision-language alignment model, performing average mask pooling by upsampling the output feature map to the size of the binary map. The average pooled feature is used as the visual-prior condition and fed into LMM. If there are multiple interactions, the process is repeated for each, yielding multiple pooled features as inputs, each feature separated using a strategy akin to the category condition. For other types of interaction, we adopt similar approaches. Specifically, for box or mask, we directly apply the pooling operator by treating them as binary masks, and for point, we bold it to a 10-pixel circle and then apply the mask pooling. After that, we use the output embedding of the visual-prior condition as the classifier weight in the mask generator to estimate the confidence of each mask proposal, as shown in Fig. 3 (c). + +# 3.4 Training Objectives + +The training process of PSALM has two stages: In the first stage, we train the visual language alignment model following LLaVA setting by freezing the vision encoder and LMM; In the second stage, we only freeze the vision encoder and fine-tune all other modules, including the mask generator. Similarly to Mask2Former, we use matching loss in the second stage training, i.e., we use bipartite matching to find the optimal assignments between the mask proposals and the ground truth masks by minimizing the matching loss and use these assignments to perform training. The loss has two terms: $\mathcal{L} = \mathcal{L}_{mask} + \mathcal{L}_{cls}$ , + +where $\mathcal{L}_{mask}$ indicates the mask loss which comprises a pixel-level Binary Cross-Entropy (BCE) loss and Dice loss [42], while the $\mathcal{L}_{cls}$ indicates the category classification loss, and we use Cross-Entropy loss for category condition and BCE loss for other cases. + +# 4 Experiments + +We conduct extensive experiments on various datasets and tasks to analyze PSALM. Details of the datasets and implementations are given in the Appendix. + +# 4.1 Ablations + +We first ablate key designs and present behind insights in this section. To better show how different designs affect the performance on a wide range of tasks, we mainly report the results on three in-domain benchmarks: COCO Panoptic Segmentation (COCO-Pan), RefCOCO-val (RefCOCO), and COCO Interactive Segmentation with point inputs (COCO-Point), and one out-of-domain benchmark: open-vocabulary instance segmentation on ADE20K-150 (A150-OV). + +Table 2: Ablation on the design of mask tokens. w.LLM: use mask tokens as inputs of LLM. Prefix: place mask tokens at the front. Suffix: place mask tokens at the end. + +
Mask TokensCOCO-PanRefCOCOCOCO-PointA150-OV
w.LLMPos.PQcIoUmIoUmAP
✓Suffix55.176.153.39.3
✗-54.8(↓0.3)74.3(↓1.8)53.1(↓0.2)8.2(↓1.1)
✓Prefix55.0(↓0.1)75.1(↓1.0)53.0(↓0.3)7.8(↓1.5)
+ +Table 3: Effect of decouple design in COCO Semantic Segmentation. + +
DecouplemIoUfwIoU
✓66.572.5
X42.735.0
+ +Design of Mask Tokens. In our approach, we use a set of mask tokens to predict the mask proposal. In practice, we have found that using mask tokens as inputs to the LLM leads to better performance than applying them directly to the mask generator, which is the default method of Mask2Former. Tab. 2 shows the results, where the direct use of mask tokens leads to a noticeable performance degradation in RefCOCO and A150-OV. We believe this is because using mask tokens as input leads to a better awareness of the information needed for the task and thus improves performance, which is essential for these two tasks. For better validation, we placed mask tokens before conditional prompts and task instruction prompts and found a similar performance drop to that of not using mask tokens in LMM, which further supports our hypothesis. + +Compared to LISA and other methods that use seg token to generate final segmentation results directly, our mask proposal approach has three advantages: First, our design is more flexible and thus can be applied to a wider range + +of segmentation tasks, especially tasks that require predicting category or confidence scores; Second, our design decouples the mask prediction and classification, which alleviating the learning difficulties for some tasks. In Tab. 3, we study how the decouple design affects the semantic segmentation performance on COCO Semantic Segmentation $^{4}$ , and we found our decouple design is significantly better $^{5}$ . Third, the mask proposals allow multiple masks to be generated for a single instance, which makes the mask accuracy superior to solutions like LISA that only predict a single mask. Tab. 4 shows that using more mask proposals on RefCOCO gives a clear improvement over using a single mask. + +Table 4: Effects of number of mask tokens in RefCOCO(cIoU). The results of LISA are listed as a reference. + +
#Mask TokenvaltestAtestB
LISA74.176.571.1
175.378.072.2
10076.578.573.4
+ +Table 5: Ablation on different designs for condition prompts. + +
ConditionCOCO-PanRefCOCOA150-OV
CategorySentencePQcIoUmAP
avg_poolingavg_pooling55.175.39.2
[REF]54.976.18.4
avg_pooling[REF]55.176.19.3
+ +Design of Condition Prompts. Another key design in PSALM is the condition prompt, particularly the way we obtain the condition embeddings, which are used as classifier weights in the mask generator to predict the class of mask proposals. As described in Sec. 3.3, for category condition, we use the avg_pooling over output embeddings of each class name as the condition embedding, and for sentence condition, we adopt a [REF] token to aggregate useful information. Tab. 5 shows the ablation, where we first tried to use the same design for all conditions and found that avg_pooling performed slightly better on COCO-Pan, with a larger improvement on A150-OV, while [REF] worked better on RefCOCO. We further used different designs and found that the advantages of each design are preserved, and the best overall performance is achieved. + +Table 6: Ablation on effect of vision-language alingment. + +
VL AlignmentCOCO-Pan PQRefCOCO cIoUCOCO-Point mIoUA150-OV mAP
✓55.176.153.39.3
✗54.9(↓0.2)71.7(↓4.4)53.0 (↓0.3)8.2(↓1.1)
+ +Table 7: Ablation on joint training. + +
AblationCOCO-PanRefCOCOCOCO-PointA150-OV
PQcIoUmIoUmAP
Task Specific Train55.676.562.27.0
Joint Train55.983.664.19.0
Δ+0.3+7.1+1.9+2.0
+ +Table 8: Comparison with the state-of-the-art methods on three referring image segmentation benchmarks with cIoU. (ft) denotes models further finetuned on RefCOCO+/g after mix training. We abbreviate the datasets: COCO(C) [24], LVIS(L) [13], RefCOCO(RC) [57], Object365(O365) [41], Video segmentation datasets(V), ADE20K(A) [59], COCO-Stuff(CS) [3], PACO-LVIS(PL) [38], PASCAL-Part(PP) [5], GranD(G) [39], VOC2010(VOC) [11], Visual Genome(VG) [16], Flicker30k(F30K) [35], MUSE(M) [40], gRefCOCO(gRC) [25], COCO-Interactive(Cl). + +
MethodSegmentation DataLLM TypeRefCOCORefCOCO+RefCOCOg
valtestAtestBvaltestAtestBvaltest
SEEM-L [61]C, L, RC-------65.6-
UNINEXT-L [53]O365, C, RC, V-80.382.677.870.074.962.673.473.7
UNINEXT-H [53]O365, C, RC, V-82.283.481.372.576.466.274.776.4
LISA [17]A, CS, RC, PL, PPVicuna-7B74.176.571.162.467.456.566.468.5
LISA(ft) [17]A, CS, RC, PL, PPVicuna-7B74.979.172.365.170.858.167.970.6
GLaMM [39]G, RCVicuna-7B79.583.276.972.678.764.674.274.9
u-LLaVA [50]A, CS, RC, PL, VOCVicuna-7B80.482.777.872.276.666.874.875.6
PerceptionGPT [34]RC, VG, F30kVicuna-13B75.379.172.168.974.061.970.771.9
PixelLM [40]A, CS, RC, PL, MLlama2-13B73.076.568.266.371.758.369.370.5
GSVA [48]A, CS, RC, PL, PP, gRCLlama2-13B77.779.974.268.071.561.573.273.9
GSVA(ft) [48]A, CS, RC, PL, PP, gRCLlama2-13B79.281.777.170.373.863.675.777.0
PSALMC, RC, CIPhi-1.5 (1.3B)83.684.781.672.975.570.173.874.4
+ +Importance of VL-alignments. The visual-language alignment stage (i.e., first training stage) is to project the visual features to the text input space, and it is an important step towards making the LLM understand images. In Tab. 6, we examine the impact of this stage, and we found that without VL alignment, the performance of all four tasks becomes worse, with the performance of A150-OV and RefCOCO being significantly affected, for example, the A150-OV dropped by -1.1 mAP and RefCOCO even dropped by -4.4 cIoU, probably because these two tasks require a strong requirement on understanding the relationship between vision and language. This result also suggests that the VL alignment is essential, and the LMM-based segmentation models have strong potential. + +Joint Training. Our architecture design and input schema help integrate various segmentation tasks so that they can be trained on one model. Tab. 7 shows the effect of this joint training on different tasks. For the task-specific models, we perform the training on the corresponding task data for $18\mathrm{k}$ iterations. In contrast, the joint training setting (see implementation section for details) has a total of $56\mathrm{k}$ training iterations, which corresponds to $14\mathrm{k}$ iterations per task. The results show that joint training of different tasks greatly improves the performance. This suggests that learning between tasks is mutually beneficial, which is also the secret of success in LLM. For example, the generic segmentation task helps refine the mask prediction in the referring segmentation, and referring expressions also enhance the model's ability to recognize more unseen categories, which in turn improves the performance of open-vocabulary segmentation tasks. + +# 4.2 System-Level Comparison on In-Domain Tasks + +In this section, we compared our model on three in-domain tasks with other state-of-the-arts to illustrate the effectiveness of our approach. + +Referring Segmentation. Most existing works aimed at getting LMMs to perform image segmentation are designed for reference segmentation tasks. We compare PSALM with other works on RefCOCO, RefCOCO+, and RefCOCOg, and Tab. 8 shows the results. Owing to the generalized and flexible design of PSALM and the advantages of joint training on multiple tasks and datasets, our system was able to achieve state-of-the-art (SOTA) performance on RefCOCO and RefCOCO+, and competitive performance on RefCOCOg with STOA, despite being driven by LLM with only 1.3B parameters. It is worth noting unlike methods such as LISA and GSVA, which may achieve improvements through task-specific fine-tuning (gray-labeled results), PSALM does not perform additional fine-tuning but still achieves better performance on RefCOCO and RefCOCO+ than their fine-tuned models. + +Generic Segmentation. We evaluate PSALM with state-of-the-art methods on the COCO Panoptic Segmentation validation set (Tab. 9). Here, we follow the evaluation protocol used in Mask2Former to report PQ, which is the main metric for panoptic segmentation, mAP on thing classes for instance segmentation, and mIoU by merging instance masks from the same category for semantic segmentation. Compared to other methods, PSALM achieves comparable performance at similar visual backbone sizes, demonstrating that PSALM is a powerful architecture, even when compared to approaches designed for specific tasks. + +Table 9: Comparison with the state-of-the-art methods on Panoptic COCO-val. We abbreviate the datasets: COCO-SAM(CM) [21], VIPSeg (VIP) [29], while others following Tab. 8. + +
MethodBackboneSeg. DataPQmAPmIoU
Mask2Former [6]Swin-BC55.145.265.1
Mask2Former [6]Swin-LC57.848.667.4
X-Decoder [60]DaViT-BC, L, RC56.245.866.0
SEEM [61]DaViT-BC, L, RC56.146.466.3
OMG-Seg [21]ConvNeXt-XXLC, VIP, CM, V55.4--
PSALMSwin-BC, RC, CI55.945.766.6
+ +Interactive Segmentation. We also evaluate PSALM in the interactive segmentation tasks. Since the task does not have a well-developed dataset containing all four instructions, previous works have typically used the in-house dataset, thus we re-evaluate other methods on the COCO interactive validation set. The results are shown in Tab. 10, and PSALM achieves leading performance on point, scribble, and mask instructions than all other methods, while on box instruction, SAM performs better on mIoU but worse on cIoU, and we hypothesize might be caused by the different distribution of training data, and the fact that SAM is trained on SA-1B [15], which has a much larger data scale than what we used. In addition, we also report the official results of SEEM, which only evaluated 600 samples from the COCO validation set as a reference. + +Table 10: Comparison with the state-of-the-art methods on COCO-Interactive. The results of SEEM-B* is the result reported in official paper, which is evaluated on 600 random samples of COCO-val, while all others are evaluated on all samples of COCO-val. Abbreviations for each dataset are the same as Tab. 9. + +
MethodSeg. DataPointScribbleBoxMask
mIoUcIoUmIoUcIoUmIoUcIoUmIoUcIoU
SAM-B [15]SA-1B48.733.6--73.768.7--
SAM-L [15]SA-1B51.837.7--76.671.6--
SEEM-B [61]C, L, RC47.857.843.044.044.942.148.465.0
SEEM-B* [61]C, L, RC81.7-83.5-75.7-76.0-
OMG-Seg [21]C, VIP, CM, V59.3-------
PSALMC, RC, CI64.374.066.980.067.380.967.682.4
+ +# 4.3 Generalizability on Out-of-Domain Tasks + +Thanks to the flexible design of architecture and input schema, multi-task joint-training, and the strong visual understanding capability of LMM, PSALM shows excellent performance on in-domain tasks, but more importantly, PSALM also demonstrates great potential to generalize to out-of-domain tasks in the zero-shot setting. In this section, we conduct experiments on three different out-of-domain tasks: open-vocabulary segmentation, generalized referring expression segmentation, and video object segmentation. We also tested the zero-shot result of the correspondence benchmark in Ego-Exo4D [12] and Video Object Segmentation in the Appendix. + +Table 11: Comparison with the state-of-the-art methods on open-vocabulary instance segmentation and semantic segmentation benchmarks. We use mAP for instance segmentation and mIoU for semantic segmentation. We abbreviate the datasets: Pascal Context-459(PC459) [30], Pascal Context-59(PC59) [30], Pascal VOC-20(PAS20) [11] + +
MethodOV Instance Seg.OV Semantic Seg.
A150CityscapesPC459A150PC59PAS20
MaskCLIP [9]6.0-10.023.745.9-
ODISE [49]14.4-14.529.957.3-
SAN [51]10.6-17.133.360.295.5
PSALM9.020.510.218.248.581.3
PSALM+LVIS13.919.314.024.457.295.0
+ +Open-Vocabulary Segmentation. We first evaluate PSALM in open-vocabulary segmentation tasks, which require the model to have the ability to deal with unseen categories in training. Here, we conduct experiments on both open-vocabulary instance segmentation and open-vocabulary semantic segmentation, and Tab. 11 shows the results. Without any special design, PSALM achieves reasonably good performance, although it is still worse than the best specific method in this task, such as SAN, but we believe that PSALM has a strong + +potential to be further improved by adding more diverse training data, which is advantages of our method. We have also made a preliminary attempt by further involving the LVIS dataset, and as we expected, the performance has significantly improvements. In addition, existing open-vocabulary segmentation methods are built upon the CLIP model or diffusion model, while our approach is based on LMM models, which is a new path and attempt to bring new inspiration to the community, which we believe is even more important than the performance. + +Table 12: Our method's zero-shot performance on gRefCOCO. (ft) denotes models further fine-tuned on gRefCOCO after mix training. + +
MethodsLLM TypeZero-ShotvaltestAtestB
cIoUgIoUcIoUgIoUcIoUgIoU
MattNet [56]-X47.548.258.759.345.346.1
LTS [8]-X52.352.761.962.649.950.4
ReLA [25]-X62.463.669.370.059.961.0
LISA [17]Vicuna-7BX38.732.252.648.544.839.7
LISA(ft) [17]Vicuna-7BX61.763.369.270.160.361.3
GSVA [48]Vicuna-7BX61.763.369.270.160.361.3
GSVA(ft) [48]Vicuna-7BX63.366.569.971.160.562.2
PSALMPhi-1.5 (1.3B)✓42.043.352.454.550.652.5
+ +Generalized Referring Expression Segmentation. The referring segmentation datasets used in training contain only a single object, however, the design of the mask proposal allows PSALM to directly address multi-target without any further training or fine-tuning. We evaluate the gRefCOCO benchmark which contains multiple segment targets. In practice, given an expression, we compute the similarity with all mask proposals and retain all masks with similarity greater than 0.6 as foreground. Tab. 12 shows the results, PSALM also achieved very promising performance, even outperforming the LISA version that only pre-trained on gRefCOCO but without task-specific fine-tuning. + +# 5 Conclusion + +The PSALM proposed in this study extends the capability of LMM from text output tasks to image segmentation, addresses the output limitations of the LMM, and unifies various segmentation tasks. PSALM exhibits excellent performance in multiple in-domain tasks, and its generalization ability in out-of-domain tasks further underscores its potential. + +# Acknowledgments + +This work was supported by the National Science Fund for Distinguished Young Scholars of China (Grant No.62225603). + +# References + +1. Alayrac, J.B., Donahue, J., Luc, P., Miech, A., Barr, I., Hasson, Y., Lenc, K., Mensch, A., Millican, K., Reynolds, M., et al.: Flamingo: a visual language model for few-shot learning. Advances in Neural Information Processing Systems 35, 23716-23736 (2022) +2. Brown, T., Mann, B., Ryder, N., Subbiah, M., Kaplan, J.D., Dhariwal, P., Neelakantan, A., Shyam, P., Sastry, G., Askell, A., et al.: Language models are few-shot learners. Advances in neural information processing systems 33, 1877-1901 (2020) +3. Caesar, H., Uijlings, J., Ferrari, V.: Coco-stuff: Thing and stuff classes in context. In: Proceedings of the IEEE conference on computer vision and pattern recognition. pp. 1209-1218 (2018) +4. Chen, K., Zhang, Z., Zeng, W., Zhang, R., Zhu, F., Zhao, R.: Shikra: Unleashing multimodal llm's referential dialogue magic. arXiv preprint arXiv:2306.15195 (2023) +5. Chen, X., Mottaghi, R., Liu, X., Fidler, S., Urtasun, R., Yuille, A.: Detect what you can: Detecting and representing objects using holistic models and body parts. In: Proceedings of the IEEE conference on computer vision and pattern recognition. pp. 1971-1978 (2014) +6. Cheng, B., Misra, I., Schwing, A.G., Kirillov, A., Girdhar, R.: Masked-attention mask transformer for universal image segmentation. In: Proceedings of the IEEE/CVF conference on computer vision and pattern recognition. pp. 1290-1299 (2022) +7. Chiang, W.L., Li, Z., Lin, Z., Sheng, Y., Wu, Z., Zhang, H., Zheng, L., Zhuang, S., Zhuang, Y., Gonzalez, J.E., et al.: Vicuna: An open-source chatbot impressing gpt-4 with $90\%$ chatgpt quality. See https://vicuna.lmsys.org (accessed 14 April 2023) (2023) +8. Ding, H., Liu, C., Wang, S., Jiang, X.: Vision-language transformer and query generation for referring segmentation. In: Proceedings of the IEEE/CVF International Conference on Computer Vision. pp. 16321-16330 (2021) +9. Ding, Z., Wang, J., Tu, Z.: Open-vocabulary universal image segmentation with maskclip (2023) +0. Dong, R., Han, C., Peng, Y., Qi, Z., Ge, Z., Yang, J., Zhao, L., Sun, J., Zhou, H., Wei, H., et al.: Dreamllm: Synergistic multimodal comprehension and creation. arXiv preprint arXiv:2309.11499 (2023) +1. Everingham, M., Van Gool, L., Williams, C.K.I., Winn, J., Zisserman, A.: The PASCAL Visual Object Classes Challenge 2010 (VOC2010) Results. http://www.pascal-network.org/challenges/VOC/voc2010/workshop/index.html +2. Grauman, K., Westbury, A., Torresani, L., Kitani, K., Malik, J., Afouras, T., Ashutosh, K., Baiyya, V., Bansal, S., Boote, B., et al.: Ego-exo4d: Understanding skilled human activity from first-and third-person perspectives. arXiv preprint arXiv:2311.18259 (2023) +3. Gupta, A., Dollar, P., Girshick, R.: Lvis: A dataset for large vocabulary instance segmentation. In: Proceedings of the IEEE/CVF conference on computer vision and pattern recognition. pp. 5356-5364 (2019) +4. Jain, J., Li, J., Chiu, M.T., Hassani, A., Orlov, N., Shi, H.: Oneformer: One transformer to rule universal image segmentation. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 2989-2998 (2023) +5. Kirillov, A., Mintun, E., Ravi, N., Mao, H., Rolland, C., Gustafson, L., Xiao, T., Whitehead, S., Berg, A.C., Lo, W.Y., et al.: Segment anything. arXiv preprint arXiv:2304.02643 (2023) + +16. Krishna, R., Zhu, Y., Groth, O., Johnson, J., Hata, K., Kravitz, J., Chen, S., Kalantidis, Y., Li, L.J., Shamma, D.A., et al.: Visual genome: Connecting language and vision using crowdsourced dense image annotations. International journal of computer vision 123, 32-73 (2017) +17. Lai, X., Tian, Z., Chen, Y., Li, Y., Yuan, Y., Liu, S., Jia, J.: Lisa: Reasoning segmentation via large language model. arXiv preprint arXiv:2308.00692 (2023) +18. Li, B., Zhang, P., Yang, J., Zhang, Y., Pu, F., Liu, Z.: Otterhd: A high-resolution multi-modality model. arXiv preprint arXiv:2311.04219 (2023) +19. Li, J., Li, D., Savarese, S., Hoi, S.: Blip-2: Bootstrapping language-image pretraining with frozen image encoders and large language models. arXiv preprint arXiv:2301.12597 (2023) +20. Li, J., Li, D., Xiong, C., Hoi, S.: Blip: Bootstrapping language-image pre-training for unified vision-language understanding and generation. In: International Conference on Machine Learning. pp. 12888-12900. PMLR (2022) +21. Li, X., Yuan, H., Li, W., Ding, H., Wu, S., Zhang, W., Li, Y., Chen, K., Loy, C.C.: Omg-seg: Is one model good enough for all segmentation? arXiv preprint arXiv:2401.10229 (2024) +22. Li, Y., Bubeck, S., Eldan, R., Del Giorno, A., Gunasekar, S., Lee, Y.T.: Textbooks are all you need ii: phi-1.5 technical report. arXiv preprint arXiv:2309.05463 (2023) +23. Li, Z., Yang, B., Liu, Q., Ma, Z., Zhang, S., Yang, J., Sun, Y., Liu, Y., Bai, X.: Monkey: Image resolution and text label are important things for large multi-modal models. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 26763-26773 (2024) +24. Lin, T.Y., Maire, M., Belongie, S., Hays, J., Perona, P., Ramanan, D., Dollar, P., Zitnick, C.L.: Microsoft coco: Common objects in context. In: Computer Vision-ECCV 2014: 13th European Conference, Zurich, Switzerland, September 6-12, 2014, Proceedings, Part V 13. pp. 740-755. Springer (2014) +25. Liu, C., Ding, H., Jiang, X.: Gres: Generalized referring expression segmentation. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 23592-23601 (2023) +26. Liu, H., Li, C., Li, Y., Li, B., Zhang, Y., Shen, S., Lee, Y.J.: Llava-last: Improved reasoning,OCR,and world knowledge (2024) +27. Liu, H., Li, C., Wu, Q., Lee, Y.J.: Visual instruction tuning. Advances in neural information processing systems 36 (2024) +28. Liu, Z., Lin, Y., Cao, Y., Hu, H., Wei, Y., Zhang, Z., Lin, S., Guo, B.: Swin transformer: Hierarchical vision transformer using shifted windows. In: Proceedings of the IEEE/CVF international conference on computer vision. pp. 10012-10022 (2021) +29. Miao, J., Wang, X., Wu, Y., Li, W., Zhang, X., Wei, Y., Yang, Y.: Large-scale video panoptic segmentation in the wild: A benchmark. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 21033-21043 (2022) +30. Mottaghi, R., Chen, X., Liu, X., Cho, N.G., Lee, S.W., Fidler, S., Urtasun, R., Yuille, A.: The role of context for object detection and semantic segmentation in the wild. In: IEEE Conference on Computer Vision and Pattern Recognition (CVPR) (2014) +31. Nagaraja, V.K., Morariu, V.I., Davis, L.S.: Modeling context between objects for referring expression understanding. In: Computer Vision-ECCV 2016: 14th European Conference, Amsterdam, The Netherlands, October 11-14, 2016, Proceedings, Part IV 14. pp. 792-807. Springer (2016) + +32. OpenAI: Gpt-4v(ision) system card (2023) +33. Peng, Z., Wang, W., Dong, L., Hao, Y., Huang, S., Ma, S., Wei, F.: Kosmos-2: Grounding multimodal large language models to the world. arXiv preprint arXiv:2306.14824 (2023) +34. Pi, R., Yao, L., Gao, J., Zhang, J., Zhang, T.: Perceptiongpt: Effectively fusing visual perception into llm. arXiv preprint arXiv:2311.06612 (2023) +35. Plummer, B.A., Wang, L., Cervantes, C.M., Caicedo, J.C., Hockenmaier, J., Lazebnik, S.: Flickr30k entities: Collecting region-to-phrase correspondences for richer image-to-sentence models. In: Proceedings of the IEEE international conference on computer vision. pp. 2641-2649 (2015) +36. Pont-Tuset, J., Perazzi, F., Caelles, S., Arbeláez, P., Sorkine-Hornung, A., Van Gool, L.: The 2017 davis challenge on video object segmentation. arXiv preprint arXiv:1704.00675 (2017) +37. Radford, A., Kim, J.W., Hallacy, C., Ramesh, A., Goh, G., Agarwal, S., Sastry, G., Askell, A., Mishkin, P., Clark, J., et al.: Learning transferable visual models from natural language supervision. In: International conference on machine learning. pp. 8748-8763. PMLR (2021) +38. Ramanathan, V., Kalia, A., Petrovic, V., Wen, Y., Zheng, B., Guo, B., Wang, R., Marquez, A., Kovvuri, R., Kadian, A., et al.: Paco: Parts and attributes of common objects. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 7141-7151 (2023) +39. Rasheed, H., Maaz, M., Shaji, S., Shaker, A., Khan, S., Cholakkal, H., Anwer, R.M., Xing, E., Yang, M.H., Khan, F.S.: Glamm: Pixel grounding large multimodal model. arXiv preprint arXiv:2311.03356 (2023) +40. Ren, Z., Huang, Z., Wei, Y., Zhao, Y., Fu, D., Feng, J., Jin, X.: Pixelm: Pixel reasoning with large multimodal model. arXiv preprint arXiv:2312.02228 (2023) +41. Shao, S., Li, Z., Zhang, T., Peng, C., Yu, G., Zhang, X., Li, J., Sun, J.: Objects365: A large-scale, high-quality dataset for object detection. In: Proceedings of the IEEE/CVF international conference on computer vision. pp. 8430-8439 (2019) +42. Sudre, C.H., Li, W., Vercauteren, T., Ourselin, S., Jorge Cardoso, M.: Generalised dice overlap as a deep learning loss function for highly unbalanced segmentations. In: Deep Learning in Medical Image Analysis and Multimodal Learning for Clinical Decision Support: Third International Workshop, DLMIA 2017, and 7th International Workshop, ML-CDS 2017, Held in Conjunction with MICCAI 2017, Quebec City, QC, Canada, September 14, Proceedings 3. pp. 240-248. Springer (2017) +43. Sun, Q., Yu, Q., Cui, Y., Zhang, F., Zhang, X., Wang, Y., Gao, H., Liu, J., Huang, T., Wang, X.: Generative pretraining in multimodality. arXiv preprint arXiv:2307.05222 (2023) +44. Team, G., Anil, R., Borgeaud, S., Wu, Y., Alayrac, J.B., Yu, J., Soricut, R., Schalkwyk, J., Dai, A.M., Hauth, A., et al.: Gemini: a family of highly capable multimodal models. arXiv preprint arXiv:2312.11805 (2023) +45. Touvron, H., Martin, L., Stone, K., Albert, P., Almahairi, A., Babaei, Y., Bashlykov, N., Batra, S., Bhargava, P., Bhosale, S., et al.: Llama 2: Open foundation and fine-tuned chat models. arXiv preprint arXiv:2307.09288 (2023) +46. Wang, W., Lv, Q., Yu, W., Hong, W., Qi, J., Wang, Y., Ji, J., Yang, Z., Zhao, L., Song, X., et al.: Cogvlm: Visual expert for pretrained language models. arXiv preprint arXiv:2311.03079 (2023) +47. Wu, J., Jiang, Y., Liu, Q., Yuan, Z., Bai, X., Bai, S.: General object foundation model for images and videos at scale. arXiv preprint arXiv:2312.09158 (2023) + +48. Xia, Z., Han, D., Han, Y., Pan, X., Song, S., Huang, G.: Gsva: Generalized segmentation via multimodal large language models. arXiv preprint arXiv:2312.10103 (2023) +49. Xu, J., Liu, S., Vahdat, A., Byeon, W., Wang, X., De Mello, S.: Open-vocabulary panoptic segmentation with text-to-image diffusion models. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 2955-2966 (2023) +50. Xu, J., Xu, L., Yang, Y., Li, X., Xie, Y., Huang, Y.J., Li, Y.: u-llava: Unifying multi-modal tasks via large language model. arXiv preprint arXiv:2311.05348 (2023) +51. Xu, M., Zhang, Z., Wei, F., Hu, H., Bai, X.: Side adapter network for open-vocabulary semantic segmentation. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 2945-2954 (2023) +52. Xu, M., Zhang, Z., Wei, F., Lin, Y., Cao, Y., Hu, H., Bai, X.: A simple baseline for open-vocabulary semantic segmentation with pre-trained vision-language model. In: European Conference on Computer Vision. pp. 736-753. Springer (2022) +53. Yan, B., Jiang, Y., Wu, J., Wang, D., Luo, P., Yuan, Z., Lu, H.: Universal instance perception as object discovery and retrieval. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 15325-15336 (2023) +54. Yang, S., Qu, T., Lai, X., Tian, Z., Peng, B., Liu, S., Jia, J.: An improved baseline for reasoning segmentation with large language model. arXiv preprint arXiv:2312.17240 (2023) +55. You, H., Zhang, H., Gan, Z., Du, X., Zhang, B., Wang, Z., Cao, L., Chang, S.F., Yang, Y.: Ferret: Refer and ground anything anywhere at any granularity. arXiv preprint arXiv:2310.07704 (2023) +56. Yu, L., Lin, Z., Shen, X., Yang, J., Lu, X., Bansal, M., Berg, T.L.: Matttnet: Modular attention network for referring expression comprehension. In: Proceedings of the IEEE conference on computer vision and pattern recognition. pp. 1307-1315 (2018) +57. Yu, L., Poirson, P., Yang, S., Berg, A.C., Berg, T.L.: Modeling context in referring expressions. In: Computer Vision-ECCV 2016: 14th European Conference, Amsterdam, The Netherlands, October 11-14, 2016, Proceedings, Part II 14. pp. 69-85. Springer (2016) +58. Zhang, A., Zhao, L., Xie, C.W., Zheng, Y., Ji, W., Chua, T.S.: Next-chat: An lmm for chat, detection and segmentation. arXiv preprint arXiv:2311.04498 (2023) +59. Zhou, B., Zhao, H., Puig, X., Xiao, T., Fidler, S., Barriuso, A., Torralba, A.: Semantic understanding of scenes through the ade20k dataset. International Journal of Computer Vision 127, 302-321 (2019) +60. Zou, X., Dou, Z.Y., Yang, J., Gan, Z., Li, L., Li, C., Dai, X., Behl, H., Wang, J., Yuan, L., et al.: Generalized decoding for pixel, image, and language. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 15116-15127 (2023) +61. Zou, X., Yang, J., Zhang, H., Li, F., Li, L., Wang, J., Wang, L., Gao, J., Lee, Y.J.: Segment everything everywhere all at once. Advances in Neural Information Processing Systems 36 (2024) \ No newline at end of file diff --git a/2024/PSALM_ Pixelwise Segmentation with Large Multi-modal Model/images.zip b/2024/PSALM_ Pixelwise Segmentation with Large Multi-modal Model/images.zip new file mode 100644 index 0000000000000000000000000000000000000000..2395d904251150f60e4fe7a1dd085c7cbf353a38 --- /dev/null +++ b/2024/PSALM_ Pixelwise Segmentation with Large Multi-modal Model/images.zip @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:20bd23a8c3fb6c6a42f37b539eb2324b1370f1858f15c7e6476d387128d9248c +size 458552 diff --git a/2024/PSALM_ Pixelwise Segmentation with Large Multi-modal Model/layout.json b/2024/PSALM_ Pixelwise Segmentation with Large Multi-modal Model/layout.json new file mode 100644 index 0000000000000000000000000000000000000000..09bfdb455bf5abfb09feb95d670a2b2ec2db12bb --- /dev/null +++ b/2024/PSALM_ Pixelwise Segmentation with Large Multi-modal Model/layout.json @@ -0,0 +1,7845 @@ +{ + "pdf_info": [ + { + "para_blocks": [ + { + "bbox": [ + 149, + 111, + 465, + 148 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 149, + 111, + 465, + 148 + ], + "spans": [ + { + "bbox": [ + 149, + 111, + 465, + 148 + ], + "type": "text", + "content": "PSALM: Pixelwise SegmentAtion with Large Multi-Modal Model" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 159, + 165, + 455, + 181 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 159, + 165, + 455, + 181 + ], + "spans": [ + { + "bbox": [ + 159, + 165, + 455, + 181 + ], + "type": "text", + "content": "Zheng Zhang" + }, + { + "bbox": [ + 159, + 165, + 455, + 181 + ], + "type": "inline_equation", + "content": "^{1,2*}" + }, + { + "bbox": [ + 159, + 165, + 455, + 181 + ], + "type": "text", + "content": ", Yeyao Ma" + }, + { + "bbox": [ + 159, + 165, + 455, + 181 + ], + "type": "inline_equation", + "content": "^{1*}" + }, + { + "bbox": [ + 159, + 165, + 455, + 181 + ], + "type": "text", + "content": ", Enming Zhang" + }, + { + "bbox": [ + 159, + 165, + 455, + 181 + ], + "type": "inline_equation", + "content": "^{1*}" + }, + { + "bbox": [ + 159, + 165, + 455, + 181 + ], + "type": "text", + "content": ", and Xiang Bai" + }, + { + "bbox": [ + 159, + 165, + 455, + 181 + ], + "type": "inline_equation", + "content": "^{1}" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 203, + 189, + 410, + 212 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 203, + 189, + 410, + 212 + ], + "spans": [ + { + "bbox": [ + 203, + 189, + 410, + 212 + ], + "type": "text", + "content": "1 Huazhong University of Science and Technology \n2 Microsoft Research Asia" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 159, + 238, + 455, + 448 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 159, + 238, + 455, + 448 + ], + "spans": [ + { + "bbox": [ + 159, + 238, + 455, + 448 + ], + "type": "text", + "content": "Abstract. PSALM is a powerful extension of the Large Multi-modal Model (LMM) to address the segmentation task challenges. To overcome the limitation of the LMM being limited to textual output, PSALM incorporates a mask decoder and a well-designed input schema to handle a variety of segmentation tasks. This schema includes images, task instructions, conditional prompts, and mask tokens, which enable the model to generate and classify segmentation masks effectively. The flexible design of PSALM supports joint training across multiple datasets and tasks, leading to improved performance and task generalization. PSALM achieves superior results on several benchmarks, such as RefCOCO/RefCOCO+/RefCOCOg, COCO Panoptic Segmentation, and COCO-Interactive, and further exhibits zero-shot capabilities on unseen tasks, such as open-vocabulary segmentation, generalized referring expression segmentation and video object segmentation, making a significant step towards a GPT moment in computer vision. Through extensive experiments, PSALM demonstrates its potential to transform the domain of image segmentation, leveraging the robust visual understanding capabilities of LMMs as seen in natural language processing. Code and models are available at https://github.com/zamling/PSALM." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 160, + 456, + 453, + 469 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 160, + 456, + 453, + 469 + ], + "spans": [ + { + "bbox": [ + 160, + 456, + 453, + 469 + ], + "type": "text", + "content": "Keywords: Segmentation " + }, + { + "bbox": [ + 160, + 456, + 453, + 469 + ], + "type": "inline_equation", + "content": "\\cdot" + }, + { + "bbox": [ + 160, + 456, + 453, + 469 + ], + "type": "text", + "content": " Large Multimodal Model " + }, + { + "bbox": [ + 160, + 456, + 453, + 469 + ], + "type": "inline_equation", + "content": "\\cdot" + }, + { + "bbox": [ + 160, + 456, + 453, + 469 + ], + "type": "text", + "content": " Visual Language" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 132, + 503, + 231, + 516 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 503, + 231, + 516 + ], + "spans": [ + { + "bbox": [ + 132, + 503, + 231, + 516 + ], + "type": "text", + "content": "1 Introduction" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 130, + 527, + 482, + 622 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 527, + 482, + 622 + ], + "spans": [ + { + "bbox": [ + 130, + 527, + 482, + 622 + ], + "type": "text", + "content": "Large multi-modal model (LMM) has ignited the dawn of the vision GPT [2] moment by making ground-breaking progress in various advanced visual understanding tasks by compressing image and language information into a single auto-regressive model. However, there are still many obstacles on the road to achieving vision GPT, and an important one being that current LMM can only perform text outputs, making it challenging to address the pixel-level image understanding problem directly, i.e., image segmentation, which is one of the most critical tasks in computer vision." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 130, + 623, + 482, + 648 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 623, + 482, + 648 + ], + "spans": [ + { + "bbox": [ + 130, + 623, + 482, + 648 + ], + "type": "text", + "content": "Behind the obstacle is many challenges. First, the default output of the LMM is discrete tokens, and there is no apparent way to generate masks directly." + } + ] + } + ], + "index": 7 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 141, + 653, + 335, + 666 + ], + "type": "page_footnote", + "angle": 0, + "lines": [ + { + "bbox": [ + 141, + 653, + 335, + 666 + ], + "spans": [ + { + "bbox": [ + 141, + 653, + 335, + 666 + ], + "type": "text", + "content": "* Equal contribution. " + }, + { + "bbox": [ + 141, + 653, + 335, + 666 + ], + "type": "inline_equation", + "content": "{}^{ \\boxtimes }" + }, + { + "bbox": [ + 141, + 653, + 335, + 666 + ], + "type": "text", + "content": " Corresponding author." + } + ] + } + ], + "index": 8 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 0 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 152, + 115, + 255, + 278 + ], + "blocks": [ + { + "bbox": [ + 152, + 115, + 255, + 278 + ], + "lines": [ + { + "bbox": [ + 152, + 115, + 255, + 278 + ], + "spans": [ + { + "bbox": [ + 152, + 115, + 255, + 278 + ], + "type": "image", + "image_path": "23d5814cdcbed14d1acaa71e7a63e13276e2b67be2a3ad5a5679354e03c77d9a.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 130, + 288, + 482, + 354 + ], + "lines": [ + { + "bbox": [ + 130, + 288, + 482, + 354 + ], + "spans": [ + { + "bbox": [ + 130, + 288, + 482, + 354 + ], + "type": "text", + "content": "Fig. 1: PSALM has capability to handle multiple segmentation tasks in only one single model. We visualize some tasks, including Panoptic segmentation in COCO [24]; Open-Vocabulary instance segmentation in ADE20K [59]; Interactive segmentation in COCO-Interactive; Referring segmentation in RefCOCO [57]; Generalized referring segmentation in gRefCOCO [25]; Ego-exo correspondence in Ego-Exo4d [12]; Video object segmentation in DAVIS2017 [36]." + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "image_caption" + } + ], + "index": 2 + }, + { + "type": "image", + "bbox": [ + 258, + 115, + 359, + 278 + ], + "blocks": [ + { + "bbox": [ + 258, + 115, + 359, + 278 + ], + "lines": [ + { + "bbox": [ + 258, + 115, + 359, + 278 + ], + "spans": [ + { + "bbox": [ + 258, + 115, + 359, + 278 + ], + "type": "image", + "image_path": "83b9485f68a18c6a8252495eb4d5c0671a7c640098271a829cf37648d10c9078.jpg" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_body" + } + ], + "index": 3 + }, + { + "type": "image", + "bbox": [ + 361, + 116, + 462, + 278 + ], + "blocks": [ + { + "bbox": [ + 361, + 116, + 462, + 278 + ], + "lines": [ + { + "bbox": [ + 361, + 116, + 462, + 278 + ], + "spans": [ + { + "bbox": [ + 361, + 116, + 462, + 278 + ], + "type": "image", + "image_path": "a3f75599d0a30bf7e3f9c386a776707c0a0f38ecf1968a487738425cde68973c.jpg" + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "image_body" + } + ], + "index": 4 + }, + { + "bbox": [ + 130, + 378, + 482, + 474 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 378, + 482, + 474 + ], + "spans": [ + { + "bbox": [ + 130, + 378, + 482, + 474 + ], + "type": "text", + "content": "Second, the variety of image segmentation tasks requires different forms of inputs and outputs. For example, semantic segmentation needs support inputs with different categories. Instance segmentation demands generate object IDs and the class confidence score for each object. Referring segmentation is supposed to have a language sentence as input and interactive segmentation has more varied inputs than the other tasks, which can be points, scribbles, bounding boxes, or masks. Third, unifying different segmentation tasks with a shared weight model is also challenging because different tasks require varied capabilities." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 130, + 474, + 482, + 545 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 474, + 482, + 545 + ], + "spans": [ + { + "bbox": [ + 130, + 474, + 482, + 545 + ], + "type": "text", + "content": "In this work, we propose a method named PSALM (Pixelwise SegmentAtion with Large Multi-Modal Model) that aims to address the above challenges and extend the capabilities of LMM from text-output tasks to general segmentation tasks (Fig. 1 shows representative tasks). Specifically, PSALM externalizes a mask decoder on the top of LMM and designs a flexible input schema to unify different segmentation tasks into a single model." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 130, + 546, + 482, + 666 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 546, + 482, + 666 + ], + "spans": [ + { + "bbox": [ + 130, + 546, + 482, + 666 + ], + "type": "text", + "content": "The input schema consists of four different parts: images, task instruction prompt, condition prompt, and a set of mask tokens, where the instruction prompt is a text sentence describing the task itself, condition prompt contains the additional necessary information to solve the task, either in terms of category names, sentence or visual features, and mask tokens are a set of learnable embeddings. All these inputs are fed into the LMM, and the resulting output mask tokens are further used as input by the mask generator to present mask proposals. Apart from producing the mask proposals, it is also necessary to predict the class of each segmentation mask or estimate a confidence score, which can be achieved by using the output embedding of the condition prompt as the" + } + ] + } + ], + "index": 8 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "spans": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "type": "text", + "content": "2" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 165, + 91, + 310, + 102 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 165, + 91, + 310, + 102 + ], + "spans": [ + { + "bbox": [ + 165, + 91, + 310, + 102 + ], + "type": "text", + "content": "Z. Zhang, Y. Ma, E. Zhang, X. Bai" + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 1 + }, + { + "para_blocks": [ + { + "bbox": [ + 130, + 116, + 479, + 163 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 116, + 479, + 163 + ], + "spans": [ + { + "bbox": [ + 130, + 116, + 479, + 163 + ], + "type": "text", + "content": "classifier weights to classify each mask proposal. In practice, we categorize conditions into category condition, sentence condition, and visual-prior condition, and present the corresponding methods to build the classifier weights according to the properties of each type of condition." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 130, + 164, + 482, + 355 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 164, + 482, + 355 + ], + "spans": [ + { + "bbox": [ + 130, + 164, + 482, + 355 + ], + "type": "text", + "content": "Some other methods, represented by LISA [17], also aim to use LMM for segmentation tasks. However, these methods are usually designed for referring segmentation and fail to justify their ability to solve generalized segmentation tasks (see Tab. 1). In contrast, thanks to the generality and flexibility of the proposed architecture, PSALM can not only solve a variety of segmentation tasks but also be able to joint train on different tasks, which makes the model task generalizable while allowing the model to take full advantage of the intrinsic connections of different datasets/tasks to achieve better performance. Specifically, with the joint training of COCO Panoptic Segmentation [24], RefCOCO [57]/RefCOCO+/RefCOCOg [31], and COCO Interactive, we observe a significant performance improvement compared to training at different tasks individually, and therefore result in even better performance than other task-specific methods. On referring segmentation tasks, we outperform other LLM-based pixel reasoning methods (e.g., LISA, PixelLM [40] and GSVA [48]) on RefCOCO, RefCOCO+, and RefCOCOg, and it worth noting that we only use Phi-1.5 1.3B model [22] while others adopt Vicuna-7B [7] or LLama2-13B model [45]." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 130, + 356, + 482, + 463 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 356, + 482, + 463 + ], + "spans": [ + { + "bbox": [ + 130, + 356, + 482, + 463 + ], + "type": "text", + "content": "The flexible design of architecture and input schema, multi-task joint-training, and the strong visual understanding capability of LMM not only make PSALM perform well on trained in-domain tasks but also enable generalizability to out-of-domain tasks in a zero-shot manner, i.e., directly dealing with unseen tasks without additional training. We test on three tasks: open-vocabulary segmentation, generalized referring expression segmentation, and video object segmentation. PSALM achieves promising zero-shot performance on these tasks. We believe this task-level generalizability is crucial, which is one of the key properties of the large language model for its success in natural language processing." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 130, + 463, + 482, + 522 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 463, + 482, + 522 + ], + "spans": [ + { + "bbox": [ + 130, + 463, + 482, + 522 + ], + "type": "text", + "content": "Through extensive experiments on a variety of segmentation tasks, we show that presented PSALM has strong potential to address general image segmentation tasks and exhibits a certain degree of task generalization capability as LLM does in NLP. We believe that this work can inspire extensive research on realizing the GPT moment in computer vision and facilitate its arrival." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 131, + 540, + 240, + 552 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 131, + 540, + 240, + 552 + ], + "spans": [ + { + "bbox": [ + 131, + 540, + 240, + 552 + ], + "type": "text", + "content": "2 Related works" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 131, + 563, + 293, + 574 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 131, + 563, + 293, + 574 + ], + "spans": [ + { + "bbox": [ + 131, + 563, + 293, + 574 + ], + "type": "text", + "content": "2.1 Large Multimodal Models" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 130, + 581, + 482, + 666 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 581, + 482, + 666 + ], + "spans": [ + { + "bbox": [ + 130, + 581, + 482, + 666 + ], + "type": "text", + "content": "With the release of GPT-V [32] and Gemini [44], more attention and efforts from open-source and research communities are shifting from large language models (LLM) to large multi-modal models (LMM). LLaVA [27], BLIP-2 [19], and Flamingo [1] are three representative works, where the core idea of both LLaVA and BLIP-2 is to map visual features into the input space of LLM to implement multi-modal capabilities, while Flamingo employs deeper feature fusion in the intermediate layers of LLM. Some works, such as Kosmos-2 [33]," + } + ] + } + ], + "index": 8 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 181, + 91, + 448, + 102 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 181, + 91, + 448, + 102 + ], + "spans": [ + { + "bbox": [ + 181, + 91, + 448, + 102 + ], + "type": "text", + "content": "PSALM: Pixelwise SegmentAtion with Large Multi-Modal Model" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 474, + 92, + 480, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 474, + 92, + 480, + 100 + ], + "spans": [ + { + "bbox": [ + 474, + 92, + 480, + 100 + ], + "type": "text", + "content": "3" + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 2 + }, + { + "para_blocks": [ + { + "type": "table", + "bbox": [ + 164, + 157, + 449, + 271 + ], + "blocks": [ + { + "bbox": [ + 132, + 114, + 482, + 148 + ], + "lines": [ + { + "bbox": [ + 132, + 114, + 482, + 148 + ], + "spans": [ + { + "bbox": [ + 132, + 114, + 482, + 148 + ], + "type": "text", + "content": "Table 1: Capability for different methods. Our proposed PSALM can handle more segmentation tasks than other LLM-centric methods. LLM-centric methods can also deal with text generation tasks, which is hard for most vision-centric methods." + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 164, + 157, + 449, + 271 + ], + "lines": [ + { + "bbox": [ + 164, + 157, + 449, + 271 + ], + "spans": [ + { + "bbox": [ + 164, + 157, + 449, + 271 + ], + "type": "table", + "html": "
MethodsGeneric Seg.Referring Seg.Interactive Seg.OV Seg.
Vision centricMask2Former [6]✓
ODISE [49]✓✓
UNINEXT [53]✓✓✓
SEEM [61]✓✓✓✓
OMG-Seg [21]✓✓✓
LLM centricLISA [17]✓
GLAMM [39]✓
PixelLM [40]✓
PSALM (Ours)✓✓✓✓
", + "image_path": "571443d84933cd8088259de6638780403b777685331013ce558ef44268098b40.jpg" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "table_body" + } + ], + "index": 3 + }, + { + "bbox": [ + 130, + 296, + 482, + 392 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 296, + 482, + 392 + ], + "spans": [ + { + "bbox": [ + 130, + 296, + 482, + 392 + ], + "type": "text", + "content": "Shikra [4], and Ferret [55], further introduce object localization tasks into the LMM, while others, such as Emu [43], CogVLM [46], and DreamLLM [10], focus on how to integrate visual generation into the LMM. Monkey [23], OtterHD [18], and LLaVA-NeXT [26] explores mechanisms for processing large-size images and significantly improves performance on tasks such as OCR. However, these above methods are mainly designed for text output tasks or image generation and cannot directly deal with pixel-level understanding tasks such as image segmentation, which is different from ours and we can base on these models." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 132, + 418, + 301, + 431 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 418, + 301, + 431 + ], + "spans": [ + { + "bbox": [ + 132, + 418, + 301, + 431 + ], + "type": "text", + "content": "2.2 Pixel Reasoning with LMM" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 130, + 448, + 482, + 581 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 448, + 482, + 581 + ], + "spans": [ + { + "bbox": [ + 130, + 448, + 482, + 581 + ], + "type": "text", + "content": "Similar to our goal, some existing works attempt to enable LMMs to generate segmentation masks. LISA [17] is a pioneering work that uses a special seg token to aggregate information of a given sentence and use it as a prompt embedding in a SAM decoder to predict the segmentation mask. u-LLaVA [50] further supports object grounding tasks on the basis of LISA, and NExT-Chat [58] introduces richer inputs, such as bounding boxes. Furthermore, since LISA can only deal with a single object, many subsequent works that attempt to address the multi-object case, such as GLaMM [39], PerceptionGPT [34], PixelLM [40], GSVA [48], and LISA++ [54], all of which share the basic idea of introducing a seg token for each sentence describing a different object, and except PixelLM, all other methods are based on SAM decoder [15]." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 130, + 582, + 482, + 665 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 582, + 482, + 665 + ], + "spans": [ + { + "bbox": [ + 130, + 582, + 482, + 665 + ], + "type": "text", + "content": "Although all these methods can generate masks, they are primarily designed for reference segmentation. In contrast, our method is designed for generalized segmentation tasks, which have diverse input and output requirements. In addition, the difference in goals also brings technical discrepancies: these methods use language models to directly generate the final segmentation masks, while our approach is closer to Mask2Former [6], which first generates mask proposals and then classifies the masks." + } + ] + } + ], + "index": 7 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "spans": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "type": "text", + "content": "4" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 166, + 91, + 310, + 102 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 166, + 91, + 310, + 102 + ], + "spans": [ + { + "bbox": [ + 166, + 91, + 310, + 102 + ], + "type": "text", + "content": "Z. Zhang, Y. Ma, E. Zhang, X. Bai" + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 3 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 149, + 114, + 465, + 228 + ], + "blocks": [ + { + "bbox": [ + 149, + 114, + 465, + 228 + ], + "lines": [ + { + "bbox": [ + 149, + 114, + 465, + 228 + ], + "spans": [ + { + "bbox": [ + 149, + 114, + 465, + 228 + ], + "type": "image", + "image_path": "8fb1ade5589cb18ace88fbb6c48b90fd4d69c3bf5d21a24f69a3cfe6798dbd51.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 227, + 236, + 385, + 246 + ], + "lines": [ + { + "bbox": [ + 227, + 236, + 385, + 246 + ], + "spans": [ + { + "bbox": [ + 227, + 236, + 385, + 246 + ], + "type": "text", + "content": "Fig. 2: PSALM architecture overview." + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_caption" + } + ], + "index": 2 + }, + { + "bbox": [ + 132, + 270, + 329, + 282 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 270, + 329, + 282 + ], + "spans": [ + { + "bbox": [ + 132, + 270, + 329, + 282 + ], + "type": "text", + "content": "2.3 Unified Segmentation Framework" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 130, + 290, + 482, + 446 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 290, + 482, + 446 + ], + "spans": [ + { + "bbox": [ + 130, + 290, + 482, + 446 + ], + "type": "text", + "content": "Another class of relevant explorations [21, 47, 53, 61] studies how to unify different segmentation tasks into a single framework. For instance, Mask2Former [6] presents a unified architecture of generic segmentation " + }, + { + "bbox": [ + 130, + 290, + 482, + 446 + ], + "type": "inline_equation", + "content": "^3" + }, + { + "bbox": [ + 130, + 290, + 482, + 446 + ], + "type": "text", + "content": " but with different models. OneFormer [14] further integrates these three tasks within a single model. UNINEXT [53] aims to unify instance-perception tasks and introduce text inputs and thus it can handle referring segmentation. X-Decoder [60] presents a flexible decoder architecture that can support generic segmentation, referring segmentation, retrieval, and image captioning. SEEM [61] designs a generic encoder-decoder to unify different segmentation tasks, where the encoder is used for projecting image, text, and human inputs into a joint visual-semantic space, and the decoder is used for mask prediction. However, all these works are not based on LMM, instead, they are mostly vision-centric models, i.e., usually designed for visual tasks only and thus cannot address language tasks very well." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 132, + 464, + 208, + 477 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 464, + 208, + 477 + ], + "spans": [ + { + "bbox": [ + 132, + 464, + 208, + 477 + ], + "type": "text", + "content": "3 Methods" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 130, + 490, + 482, + 574 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 490, + 482, + 574 + ], + "spans": [ + { + "bbox": [ + 130, + 490, + 482, + 574 + ], + "type": "text", + "content": "Fig. 2 provides an overview schematic of PSALM, which consists of a large multimodal model (LMM), a mask generator, and a flexible input schema designed for general segmentation tasks. The input schema has four different types of inputs: image, task instruction prompt, condition prompt, and a set of mask tokens. LMM processes the input tokens and the output embedding of mask tokens is further fed into the mask generator to generate masks. In the following, we will introduce our approach in detail." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 132, + 591, + 384, + 604 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 591, + 384, + 604 + ], + "spans": [ + { + "bbox": [ + 132, + 591, + 384, + 604 + ], + "type": "text", + "content": "3.1 Large Multimodal Model and Input Schema" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 130, + 611, + 481, + 635 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 611, + 481, + 635 + ], + "spans": [ + { + "bbox": [ + 130, + 611, + 481, + 635 + ], + "type": "text", + "content": "PSALM is built on large multimodal models (LMM), and there are many different LMM architectures, such as LLaVA [27], BLIP [20], and Flamingo [1]. Here," + } + ] + } + ], + "index": 9 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 181, + 91, + 448, + 102 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 181, + 91, + 448, + 102 + ], + "spans": [ + { + "bbox": [ + 181, + 91, + 448, + 102 + ], + "type": "text", + "content": "PSALM: Pixelwise SegmentAtion with Large Multi-Modal Model" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 474, + 92, + 480, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 474, + 92, + 480, + 100 + ], + "spans": [ + { + "bbox": [ + 474, + 92, + 480, + 100 + ], + "type": "text", + "content": "5" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 133, + 642, + 481, + 665 + ], + "type": "page_footnote", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 642, + 481, + 665 + ], + "spans": [ + { + "bbox": [ + 133, + 642, + 481, + 665 + ], + "type": "text", + "content": "3 Generic segmentation includes semantic segmentation, instance segmentation, and panoptic segmentation." + } + ] + } + ], + "index": 10 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 4 + }, + { + "para_blocks": [ + { + "bbox": [ + 130, + 116, + 479, + 152 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 116, + 479, + 152 + ], + "spans": [ + { + "bbox": [ + 130, + 116, + 479, + 152 + ], + "type": "text", + "content": "we adopt the design of LLaVA because of its proven performance and simplicity, but the other LMM architectures are also compatible with our approach without any theoretical difficulties." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 130, + 153, + 480, + 285 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 153, + 480, + 285 + ], + "spans": [ + { + "bbox": [ + 130, + 153, + 480, + 285 + ], + "type": "text", + "content": "The LMM used in our work has a visual encoder and pre-trained large language model (LLM). The two models are connected by a lightweight vision-language alignment model, which is a " + }, + { + "bbox": [ + 130, + 153, + 480, + 285 + ], + "type": "inline_equation", + "content": "3 \\times 3" + }, + { + "bbox": [ + 130, + 153, + 480, + 285 + ], + "type": "text", + "content": " convolution layer followed by a linear layer. The official LLaVA model uses a frozen CLIP model [37] as a visual coder, whose features lack the fine-grained information that is required for segmentation tasks [52]. Therefore, we train a customized LLaVA model by using the Swin Transformer [28], and limited by resources, we additionally replace the LLM from the Vicuna 7B model [7] to a smaller Phi-1.5 1.3B model [22]. Here, we applied only the first visual-language alignment stage of LLaVA by following its default settings. In our ablations, we found the alignment stage is crucial for open-vocabulary segmentation and referring segmentation tasks (Tab. 6)." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 130, + 287, + 480, + 358 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 287, + 480, + 358 + ], + "spans": [ + { + "bbox": [ + 130, + 287, + 480, + 358 + ], + "type": "text", + "content": "Different segmentation tasks need different forms of inputs and outputs, which motivates us to present a flexible input schema to unify various requirements. In addition to the input image used in the visual encoder, our input schema has three other different types of inputs: task instruction prompt, condition prompt, and a set of mask tokens. We will introduce each of them and summarize the prompts used for all different tasks in the Appendix." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 131, + 361, + 481, + 421 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 131, + 361, + 481, + 421 + ], + "spans": [ + { + "bbox": [ + 131, + 361, + 481, + 421 + ], + "type": "text", + "content": "Task Instruction Prompt. The task instruction prompt is usually a text sentence describing and specifying the model's task. For example, in panoptic segmentation, the task instruction can be \"You need to segment all objects. This is all the candidate categories.\" and in referring segmentation, the instruction can be \"Please segment according to the following instruction.\"" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 131, + 422, + 481, + 506 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 131, + 422, + 481, + 506 + ], + "spans": [ + { + "bbox": [ + 131, + 422, + 481, + 506 + ], + "type": "text", + "content": "Condition Prompt. Some tasks require additional information to perform, e.g., panoptic segmentation needs specifying the candidate set of categories to be segmented, and interactive segmentation needs interactive inputs. The condition prompt is designed for these tasks. In addition to providing information, condition prompt also plays an important role in predicting categories or estimating confidence scores for each segmentation mask. In Sec. 3.3, we will discuss the design of condition prompts for different tasks in detail." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 130, + 508, + 480, + 604 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 508, + 480, + 604 + ], + "spans": [ + { + "bbox": [ + 130, + 508, + 480, + 604 + ], + "type": "text", + "content": "Mask Token. The LLM is designed for text output and cannot directly generate segmentation masks. To bypass this challenge, we append a set of mask tokens after other inputs, and then these mask tokens are decoded to segmentation masks by a mask generator (will be introduced in Sec. 3.2). This design is inspired by Mask2Former [6], with the difference that the mask tokens in Mask2Former are used directly in the mask generator, whereas the mask tokens in our approach are first updated by the LMM and then used in the mask generator, and we found our approach leads better performance in practice (see Tab. 2)." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 130, + 605, + 480, + 665 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 605, + 480, + 665 + ], + "spans": [ + { + "bbox": [ + 130, + 605, + 480, + 665 + ], + "type": "text", + "content": "Some works, such as LISA [17] and PixelLM [40], take similar seg tokens as input and employ a decoder to generate the masks. However, our objective is fundamentally different: in LISA and PixelLLM, seg tokens are used to generate the final prediction, while we generate mask proposals and further classify them based on condition prompts. Compared to the design of LISA and PixelLLM," + } + ] + } + ], + "index": 8 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "spans": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "type": "text", + "content": "6" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 166, + 91, + 309, + 101 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 166, + 91, + 309, + 101 + ], + "spans": [ + { + "bbox": [ + 166, + 91, + 309, + 101 + ], + "type": "text", + "content": "Z. Zhang, Y. Ma, E. Zhang, X. Bai" + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 5 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 151, + 114, + 465, + 230 + ], + "blocks": [ + { + "bbox": [ + 151, + 114, + 465, + 230 + ], + "lines": [ + { + "bbox": [ + 151, + 114, + 465, + 230 + ], + "spans": [ + { + "bbox": [ + 151, + 114, + 465, + 230 + ], + "type": "image", + "image_path": "f125fa85c77e2aafe356b5cbf1636857c587eafbef960a94bbcba5e61c8dfef3.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 131, + 238, + 482, + 271 + ], + "lines": [ + { + "bbox": [ + 131, + 238, + 482, + 271 + ], + "spans": [ + { + "bbox": [ + 131, + 238, + 482, + 271 + ], + "type": "text", + "content": "Fig. 3: Detailed processing for different condition prompts. (a) shows the processing for category condition. (b) shows the processing for sentence condition. (c) shows the processing for visual-prior condition." + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_caption" + } + ], + "index": 2 + }, + { + "bbox": [ + 130, + 293, + 480, + 319 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 293, + 480, + 319 + ], + "spans": [ + { + "bbox": [ + 130, + 293, + 480, + 319 + ], + "type": "text", + "content": "our approach is more flexible and adaptable to more tasks, yet decoupling mask generation and classification is more efficient (see Tab. 3 and Tab. 4)." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 132, + 334, + 243, + 345 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 334, + 243, + 345 + ], + "spans": [ + { + "bbox": [ + 132, + 334, + 243, + 345 + ], + "type": "text", + "content": "3.2 Mask Generator" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 130, + 353, + 479, + 390 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 353, + 479, + 390 + ], + "spans": [ + { + "bbox": [ + 130, + 353, + 479, + 390 + ], + "type": "text", + "content": "The mask generator predicts the mask and their category probabilities from three inputs: a multi-level visual features " + }, + { + "bbox": [ + 130, + 353, + 479, + 390 + ], + "type": "inline_equation", + "content": "\\{v_{l}\\}_{l = 1}^{L}" + }, + { + "bbox": [ + 130, + 353, + 479, + 390 + ], + "type": "text", + "content": ", a set of mask tokens " + }, + { + "bbox": [ + 130, + 353, + 479, + 390 + ], + "type": "inline_equation", + "content": "\\{q_i\\}_{i = 1}^N" + }, + { + "bbox": [ + 130, + 353, + 479, + 390 + ], + "type": "text", + "content": " and a set of condition embeddings " + }, + { + "bbox": [ + 130, + 353, + 479, + 390 + ], + "type": "inline_equation", + "content": "\\{c_k\\}_{k = 1}^K" + }, + { + "bbox": [ + 130, + 353, + 479, + 390 + ], + "type": "text", + "content": ". It can be formally defined as:" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 188, + 395, + 481, + 409 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 188, + 395, + 481, + 409 + ], + "spans": [ + { + "bbox": [ + 188, + 395, + 481, + 409 + ], + "type": "interline_equation", + "content": "\\{(m _ {i}, p _ {i}) \\} _ {i = 1} ^ {N} = \\operatorname {M a s k G e n e r a t o r} (\\{v _ {i} \\} _ {l = 1} ^ {L}, \\left\\{q _ {i} \\right\\} _ {i = 1} ^ {N}, \\left\\{c _ {k} \\right\\} _ {k = 1} ^ {K}), \\tag {1}", + "image_path": "9e31025d3d924704b0024a843d2d402f56d4a99dee7fca62e4e6bc4c81e88d1d.jpg" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 130, + 414, + 482, + 511 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 414, + 482, + 511 + ], + "spans": [ + { + "bbox": [ + 130, + 414, + 482, + 511 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 130, + 414, + 482, + 511 + ], + "type": "inline_equation", + "content": "m_{i}\\in \\mathbb{R}^{H\\times W}" + }, + { + "bbox": [ + 130, + 414, + 482, + 511 + ], + "type": "text", + "content": " is the i-th predicted segmentation mask and " + }, + { + "bbox": [ + 130, + 414, + 482, + 511 + ], + "type": "inline_equation", + "content": "p_i\\in \\mathbb{R}^K" + }, + { + "bbox": [ + 130, + 414, + 482, + 511 + ], + "type": "text", + "content": " is the corresponding category probability. In practice, the multi-level visual features are the internal features of the Swin visual encoder used in LMM. The design of the mask generator follows Mask2Former, which employs multi-scale deformable attention as a pixel decoder and a transformer-based mask decoder to generate segmentation masks. The class of each mask is predicted by the condition embedding " + }, + { + "bbox": [ + 130, + 414, + 482, + 511 + ], + "type": "inline_equation", + "content": "\\{c_k\\}" + }, + { + "bbox": [ + 130, + 414, + 482, + 511 + ], + "type": "text", + "content": ", which is basically obtained from the output of the condition prompt, the obtaining method is slightly different for different types of conditions." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 132, + 527, + 308, + 540 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 527, + 308, + 540 + ], + "spans": [ + { + "bbox": [ + 132, + 527, + 308, + 540 + ], + "type": "text", + "content": "3.3 Design of Condition Prompts" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 130, + 546, + 482, + 629 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 546, + 482, + 629 + ], + "spans": [ + { + "bbox": [ + 130, + 546, + 482, + 629 + ], + "type": "text", + "content": "In our approach, the conditional prompt serves two important purposes: First, it provides the necessary information required to solve the task; Second, we use the output embedding of the conditional prompt in the LLM as classifier weights to predict the class of each segmentation mask. The design of the conditional prompt is closely tied to the type of task, and based on the information required for different tasks, we summarize three condition types: category condition, sentence condition, and visual-prior condition." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 130, + 629, + 482, + 666 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 629, + 482, + 666 + ], + "spans": [ + { + "bbox": [ + 130, + 629, + 482, + 666 + ], + "type": "text", + "content": "Category Condition This condition type is used for tasks that need specifying a set of categories to be segmented, such as semantic segmentation, instance segmentation, and panoptic segmentation, and often needs to predict the" + } + ] + } + ], + "index": 11 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 181, + 91, + 448, + 102 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 181, + 91, + 448, + 102 + ], + "spans": [ + { + "bbox": [ + 181, + 91, + 448, + 102 + ], + "type": "text", + "content": "PSALM: Pixelwise SegmentAtion with Large Multi-Modal Model" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 474, + 92, + 480, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 474, + 92, + 480, + 100 + ], + "spans": [ + { + "bbox": [ + 474, + 92, + 480, + 100 + ], + "type": "text", + "content": "7" + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 6 + }, + { + "para_blocks": [ + { + "bbox": [ + 130, + 116, + 479, + 152 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 116, + 479, + 152 + ], + "spans": [ + { + "bbox": [ + 130, + 116, + 479, + 152 + ], + "type": "text", + "content": "class probability of each segmentation mask. Specifically, given a set of category names, we join them into a sentence by a comma separator, e.g., given three categories: person, bicycle, and car, the joint sentence is \"person, bicycle, car\"." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 130, + 152, + 480, + 236 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 152, + 480, + 236 + ], + "spans": [ + { + "bbox": [ + 130, + 152, + 480, + 236 + ], + "type": "text", + "content": "The joint sentences are then proceeded by LMM to get the output embeddings, which can be further used to classify the predicted segmentation masks. Specifically, for each category, we select its corresponding output embeddings and apply the avg_pooling over them to obtain a condition embedding " + }, + { + "bbox": [ + 130, + 152, + 480, + 236 + ], + "type": "inline_equation", + "content": "c \\in \\mathbb{R}^D" + }, + { + "bbox": [ + 130, + 152, + 480, + 236 + ], + "type": "text", + "content": ", where " + }, + { + "bbox": [ + 130, + 152, + 480, + 236 + ], + "type": "inline_equation", + "content": "D" + }, + { + "bbox": [ + 130, + 152, + 480, + 236 + ], + "type": "text", + "content": " is the embedding dimension, and thus condition embedding for all categories is a set " + }, + { + "bbox": [ + 130, + 152, + 480, + 236 + ], + "type": "inline_equation", + "content": "\\{c_k\\}_{k=1}^K" + }, + { + "bbox": [ + 130, + 152, + 480, + 236 + ], + "type": "text", + "content": ", where " + }, + { + "bbox": [ + 130, + 152, + 480, + 236 + ], + "type": "inline_equation", + "content": "K" + }, + { + "bbox": [ + 130, + 152, + 480, + 236 + ], + "type": "text", + "content": " is the number of categories (see Fig. 3 (a)). This embedding set can be used by the mask generator to predict the class." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 130, + 237, + 481, + 332 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 237, + 481, + 332 + ], + "spans": [ + { + "bbox": [ + 130, + 237, + 481, + 332 + ], + "type": "text", + "content": "Sentence Condition This condition is usually used for referring segmentation. Unlike the category condition, where category names are usually short, sentences are much longer, and not every word in the sentence is useful, so the avg_pooling is not the optimal choice here. Instead, we introduce a special [REF] token, which is appended after the condition sentence as an anchor to aggregate useful information, and the output embedding of [REF] token, i.e., the output features of LMM on the location of [REF] token, is used as condition embedding " + }, + { + "bbox": [ + 130, + 237, + 481, + 332 + ], + "type": "inline_equation", + "content": "c" + }, + { + "bbox": [ + 130, + 237, + 481, + 332 + ], + "type": "text", + "content": " and used by mask generator, as shown in Fig. 3 (b)." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 130, + 334, + 482, + 525 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 334, + 482, + 525 + ], + "spans": [ + { + "bbox": [ + 130, + 334, + 482, + 525 + ], + "type": "text", + "content": "Visual-Prior Condition We formulate most interaction (e.g., point, mask, box, or scribble) used in interactive segmentation tasks as the visual-prior condition. Taking the scribble as an example, we first generate a binary scribble curve with a width of 5 pixels on a binary map, whose size is the same as the input image of the Swin vision encoder, and then apply the binary map to the output features of the vision-language alignment model, performing average mask pooling by upsampling the output feature map to the size of the binary map. The average pooled feature is used as the visual-prior condition and fed into LMM. If there are multiple interactions, the process is repeated for each, yielding multiple pooled features as inputs, each feature separated using a strategy akin to the category condition. For other types of interaction, we adopt similar approaches. Specifically, for box or mask, we directly apply the pooling operator by treating them as binary masks, and for point, we bold it to a 10-pixel circle and then apply the mask pooling. After that, we use the output embedding of the visual-prior condition as the classifier weight in the mask generator to estimate the confidence of each mask proposal, as shown in Fig. 3 (c)." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 132, + 546, + 259, + 558 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 546, + 259, + 558 + ], + "spans": [ + { + "bbox": [ + 132, + 546, + 259, + 558 + ], + "type": "text", + "content": "3.4 Training Objectives" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 130, + 570, + 482, + 666 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 570, + 482, + 666 + ], + "spans": [ + { + "bbox": [ + 130, + 570, + 482, + 666 + ], + "type": "text", + "content": "The training process of PSALM has two stages: In the first stage, we train the visual language alignment model following LLaVA setting by freezing the vision encoder and LMM; In the second stage, we only freeze the vision encoder and fine-tune all other modules, including the mask generator. Similarly to Mask2Former, we use matching loss in the second stage training, i.e., we use bipartite matching to find the optimal assignments between the mask proposals and the ground truth masks by minimizing the matching loss and use these assignments to perform training. The loss has two terms: " + }, + { + "bbox": [ + 130, + 570, + 482, + 666 + ], + "type": "inline_equation", + "content": "\\mathcal{L} = \\mathcal{L}_{mask} + \\mathcal{L}_{cls}" + }, + { + "bbox": [ + 130, + 570, + 482, + 666 + ], + "type": "text", + "content": "," + } + ] + } + ], + "index": 7 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "spans": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "type": "text", + "content": "8" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 166, + 91, + 309, + 101 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 166, + 91, + 309, + 101 + ], + "spans": [ + { + "bbox": [ + 166, + 91, + 309, + 101 + ], + "type": "text", + "content": "Z. Zhang, Y. Ma, E. Zhang, X. Bai" + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 7 + }, + { + "para_blocks": [ + { + "bbox": [ + 130, + 116, + 482, + 163 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 116, + 482, + 163 + ], + "spans": [ + { + "bbox": [ + 130, + 116, + 482, + 163 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 130, + 116, + 482, + 163 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_{mask}" + }, + { + "bbox": [ + 130, + 116, + 482, + 163 + ], + "type": "text", + "content": " indicates the mask loss which comprises a pixel-level Binary Cross-Entropy (BCE) loss and Dice loss [42], while the " + }, + { + "bbox": [ + 130, + 116, + 482, + 163 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_{cls}" + }, + { + "bbox": [ + 130, + 116, + 482, + 163 + ], + "type": "text", + "content": " indicates the category classification loss, and we use Cross-Entropy loss for category condition and BCE loss for other cases." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 131, + 182, + 230, + 196 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 131, + 182, + 230, + 196 + ], + "spans": [ + { + "bbox": [ + 131, + 182, + 230, + 196 + ], + "type": "text", + "content": "4 Experiments" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 130, + 207, + 482, + 232 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 207, + 482, + 232 + ], + "spans": [ + { + "bbox": [ + 130, + 207, + 482, + 232 + ], + "type": "text", + "content": "We conduct extensive experiments on various datasets and tasks to analyze PSALM. Details of the datasets and implementations are given in the Appendix." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 131, + 249, + 210, + 261 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 131, + 249, + 210, + 261 + ], + "spans": [ + { + "bbox": [ + 131, + 249, + 210, + 261 + ], + "type": "text", + "content": "4.1 Ablations" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 130, + 269, + 482, + 342 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 269, + 482, + 342 + ], + "spans": [ + { + "bbox": [ + 130, + 269, + 482, + 342 + ], + "type": "text", + "content": "We first ablate key designs and present behind insights in this section. To better show how different designs affect the performance on a wide range of tasks, we mainly report the results on three in-domain benchmarks: COCO Panoptic Segmentation (COCO-Pan), RefCOCO-val (RefCOCO), and COCO Interactive Segmentation with point inputs (COCO-Point), and one out-of-domain benchmark: open-vocabulary instance segmentation on ADE20K-150 (A150-OV)." + } + ] + } + ], + "index": 6 + }, + { + "type": "table", + "bbox": [ + 135, + 415, + 338, + 477 + ], + "blocks": [ + { + "bbox": [ + 131, + 361, + 343, + 407 + ], + "lines": [ + { + "bbox": [ + 131, + 361, + 343, + 407 + ], + "spans": [ + { + "bbox": [ + 131, + 361, + 343, + 407 + ], + "type": "text", + "content": "Table 2: Ablation on the design of mask tokens. w.LLM: use mask tokens as inputs of LLM. Prefix: place mask tokens at the front. Suffix: place mask tokens at the end." + } + ] + } + ], + "index": 7, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 135, + 415, + 338, + 477 + ], + "lines": [ + { + "bbox": [ + 135, + 415, + 338, + 477 + ], + "spans": [ + { + "bbox": [ + 135, + 415, + 338, + 477 + ], + "type": "table", + "html": "
Mask TokensCOCO-PanRefCOCOCOCO-PointA150-OV
w.LLMPos.PQcIoUmIoUmAP
✓Suffix55.176.153.39.3
✗-54.8(↓0.3)74.3(↓1.8)53.1(↓0.2)8.2(↓1.1)
✓Prefix55.0(↓0.1)75.1(↓1.0)53.0(↓0.3)7.8(↓1.5)
", + "image_path": "0f5995d076b5cbe46765139a7ccf7fd52bc4d1d975f628dd480c557806db89b1.jpg" + } + ] + } + ], + "index": 8, + "angle": 0, + "type": "table_body" + } + ], + "index": 8 + }, + { + "type": "table", + "bbox": [ + 370, + 416, + 466, + 466 + ], + "blocks": [ + { + "bbox": [ + 356, + 373, + 482, + 407 + ], + "lines": [ + { + "bbox": [ + 356, + 373, + 482, + 407 + ], + "spans": [ + { + "bbox": [ + 356, + 373, + 482, + 407 + ], + "type": "text", + "content": "Table 3: Effect of decouple design in COCO Semantic Segmentation." + } + ] + } + ], + "index": 9, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 370, + 416, + 466, + 466 + ], + "lines": [ + { + "bbox": [ + 370, + 416, + 466, + 466 + ], + "spans": [ + { + "bbox": [ + 370, + 416, + 466, + 466 + ], + "type": "table", + "html": "
DecouplemIoUfwIoU
✓66.572.5
X42.735.0
", + "image_path": "fe907bd25d5875899faaaa398aee2788031d6f7ac8fa45ebec1f60b78eb8d2fa.jpg" + } + ] + } + ], + "index": 10, + "angle": 0, + "type": "table_body" + } + ], + "index": 10 + }, + { + "bbox": [ + 130, + 498, + 482, + 629 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 498, + 482, + 629 + ], + "spans": [ + { + "bbox": [ + 130, + 498, + 482, + 629 + ], + "type": "text", + "content": "Design of Mask Tokens. In our approach, we use a set of mask tokens to predict the mask proposal. In practice, we have found that using mask tokens as inputs to the LLM leads to better performance than applying them directly to the mask generator, which is the default method of Mask2Former. Tab. 2 shows the results, where the direct use of mask tokens leads to a noticeable performance degradation in RefCOCO and A150-OV. We believe this is because using mask tokens as input leads to a better awareness of the information needed for the task and thus improves performance, which is essential for these two tasks. For better validation, we placed mask tokens before conditional prompts and task instruction prompts and found a similar performance drop to that of not using mask tokens in LMM, which further supports our hypothesis." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 130, + 630, + 482, + 666 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 630, + 482, + 666 + ], + "spans": [ + { + "bbox": [ + 130, + 630, + 482, + 666 + ], + "type": "text", + "content": "Compared to LISA and other methods that use seg token to generate final segmentation results directly, our mask proposal approach has three advantages: First, our design is more flexible and thus can be applied to a wider range" + } + ] + } + ], + "index": 12 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 181, + 90, + 448, + 102 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 181, + 90, + 448, + 102 + ], + "spans": [ + { + "bbox": [ + 181, + 90, + 448, + 102 + ], + "type": "text", + "content": "PSALM: Pixelwise SegmentAtion with Large Multi-Modal Model" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 474, + 91, + 481, + 101 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 474, + 91, + 481, + 101 + ], + "spans": [ + { + "bbox": [ + 474, + 91, + 481, + 101 + ], + "type": "text", + "content": "9" + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 8 + }, + { + "para_blocks": [ + { + "bbox": [ + 130, + 116, + 482, + 224 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 116, + 482, + 224 + ], + "spans": [ + { + "bbox": [ + 130, + 116, + 482, + 224 + ], + "type": "text", + "content": "of segmentation tasks, especially tasks that require predicting category or confidence scores; Second, our design decouples the mask prediction and classification, which alleviating the learning difficulties for some tasks. In Tab. 3, we study how the decouple design affects the semantic segmentation performance on COCO Semantic Segmentation" + }, + { + "bbox": [ + 130, + 116, + 482, + 224 + ], + "type": "inline_equation", + "content": "^{4}" + }, + { + "bbox": [ + 130, + 116, + 482, + 224 + ], + "type": "text", + "content": ", and we found our decouple design is significantly better" + }, + { + "bbox": [ + 130, + 116, + 482, + 224 + ], + "type": "inline_equation", + "content": "^{5}" + }, + { + "bbox": [ + 130, + 116, + 482, + 224 + ], + "type": "text", + "content": ". Third, the mask proposals allow multiple masks to be generated for a single instance, which makes the mask accuracy superior to solutions like LISA that only predict a single mask. Tab. 4 shows that using more mask proposals on RefCOCO gives a clear improvement over using a single mask." + } + ] + } + ], + "index": 2 + }, + { + "type": "table", + "bbox": [ + 147, + 299, + 257, + 350 + ], + "blocks": [ + { + "bbox": [ + 131, + 245, + 275, + 289 + ], + "lines": [ + { + "bbox": [ + 131, + 245, + 275, + 289 + ], + "spans": [ + { + "bbox": [ + 131, + 245, + 275, + 289 + ], + "type": "text", + "content": "Table 4: Effects of number of mask tokens in RefCOCO(cIoU). The results of LISA are listed as a reference." + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 147, + 299, + 257, + 350 + ], + "lines": [ + { + "bbox": [ + 147, + 299, + 257, + 350 + ], + "spans": [ + { + "bbox": [ + 147, + 299, + 257, + 350 + ], + "type": "table", + "html": "
#Mask TokenvaltestAtestB
LISA74.176.571.1
175.378.072.2
10076.578.573.4
", + "image_path": "165f472494cbfd4d544ad17166d0044030aee5e4226c3081348b87351fa6ede8.jpg" + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "table_body" + } + ], + "index": 4 + }, + { + "type": "table", + "bbox": [ + 290, + 284, + 488, + 344 + ], + "blocks": [ + { + "bbox": [ + 287, + 251, + 482, + 275 + ], + "lines": [ + { + "bbox": [ + 287, + 251, + 482, + 275 + ], + "spans": [ + { + "bbox": [ + 287, + 251, + 482, + 275 + ], + "type": "text", + "content": "Table 5: Ablation on different designs for condition prompts." + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 290, + 284, + 488, + 344 + ], + "lines": [ + { + "bbox": [ + 290, + 284, + 488, + 344 + ], + "spans": [ + { + "bbox": [ + 290, + 284, + 488, + 344 + ], + "type": "table", + "html": "
ConditionCOCO-PanRefCOCOA150-OV
CategorySentencePQcIoUmAP
avg_poolingavg_pooling55.175.39.2
[REF]54.976.18.4
avg_pooling[REF]55.176.19.3
", + "image_path": "effd1c92afd062bb5888b8b65ab822a939625c828f9a2d00d52639c572b4f9c5.jpg" + } + ] + } + ], + "index": 6, + "angle": 0, + "type": "table_body" + } + ], + "index": 6 + }, + { + "bbox": [ + 130, + 373, + 482, + 506 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 373, + 482, + 506 + ], + "spans": [ + { + "bbox": [ + 130, + 373, + 482, + 506 + ], + "type": "text", + "content": "Design of Condition Prompts. Another key design in PSALM is the condition prompt, particularly the way we obtain the condition embeddings, which are used as classifier weights in the mask generator to predict the class of mask proposals. As described in Sec. 3.3, for category condition, we use the avg_pooling over output embeddings of each class name as the condition embedding, and for sentence condition, we adopt a [REF] token to aggregate useful information. Tab. 5 shows the ablation, where we first tried to use the same design for all conditions and found that avg_pooling performed slightly better on COCO-Pan, with a larger improvement on A150-OV, while [REF] worked better on RefCOCO. We further used different designs and found that the advantages of each design are preserved, and the best overall performance is achieved." + } + ] + } + ], + "index": 7 + }, + { + "type": "table", + "bbox": [ + 134, + 558, + 299, + 599 + ], + "blocks": [ + { + "bbox": [ + 131, + 527, + 301, + 549 + ], + "lines": [ + { + "bbox": [ + 131, + 527, + 301, + 549 + ], + "spans": [ + { + "bbox": [ + 131, + 527, + 301, + 549 + ], + "type": "text", + "content": "Table 6: Ablation on effect of vision-language alingment." + } + ] + } + ], + "index": 8, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 134, + 558, + 299, + 599 + ], + "lines": [ + { + "bbox": [ + 134, + 558, + 299, + 599 + ], + "spans": [ + { + "bbox": [ + 134, + 558, + 299, + 599 + ], + "type": "table", + "html": "
VL AlignmentCOCO-Pan PQRefCOCO cIoUCOCO-Point mIoUA150-OV mAP
✓55.176.153.39.3
✗54.9(↓0.2)71.7(↓4.4)53.0 (↓0.3)8.2(↓1.1)
", + "image_path": "a25a87db8fe80a6deaace08e2a5caff4eea8604d48ef5aea3f40d8a10db1e1be.jpg" + } + ] + } + ], + "index": 9, + "angle": 0, + "type": "table_body" + } + ], + "index": 9 + }, + { + "type": "table", + "bbox": [ + 314, + 552, + 484, + 595 + ], + "blocks": [ + { + "bbox": [ + 321, + 530, + 471, + 542 + ], + "lines": [ + { + "bbox": [ + 321, + 530, + 471, + 542 + ], + "spans": [ + { + "bbox": [ + 321, + 530, + 471, + 542 + ], + "type": "text", + "content": "Table 7: Ablation on joint training." + } + ] + } + ], + "index": 10, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 314, + 552, + 484, + 595 + ], + "lines": [ + { + "bbox": [ + 314, + 552, + 484, + 595 + ], + "spans": [ + { + "bbox": [ + 314, + 552, + 484, + 595 + ], + "type": "table", + "html": "
AblationCOCO-PanRefCOCOCOCO-PointA150-OV
PQcIoUmIoUmAP
Task Specific Train55.676.562.27.0
Joint Train55.983.664.19.0
Δ+0.3+7.1+1.9+2.0
", + "image_path": "3bbc623b9edd52ac54bbcf4b9398e5416470787643051b752fe5f8e05c966d8e.jpg" + } + ] + } + ], + "index": 11, + "angle": 0, + "type": "table_body" + } + ], + "index": 11 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "spans": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "type": "text", + "content": "10" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 166, + 90, + 310, + 102 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 166, + 90, + 310, + 102 + ], + "spans": [ + { + "bbox": [ + 166, + 90, + 310, + 102 + ], + "type": "text", + "content": "Z. Zhang, Y. Ma, E. Zhang, X. Bai" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 133, + 631, + 481, + 654 + ], + "type": "page_footnote", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 631, + 481, + 654 + ], + "spans": [ + { + "bbox": [ + 133, + 631, + 481, + 654 + ], + "type": "text", + "content": "This benchmark is introduced by Mask2Former, which is composed by merging instances belonging to the same class together in COCO Panoptic Segmentation." + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 135, + 654, + 321, + 665 + ], + "type": "page_footnote", + "angle": 0, + "lines": [ + { + "bbox": [ + 135, + 654, + 321, + 665 + ], + "spans": [ + { + "bbox": [ + 135, + 654, + 321, + 665 + ], + "type": "text", + "content": "5 More experimental details are in Appendix." + } + ] + } + ], + "index": 13 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 9 + }, + { + "para_blocks": [ + { + "type": "table", + "bbox": [ + 148, + 201, + 462, + 319 + ], + "blocks": [ + { + "bbox": [ + 130, + 114, + 482, + 193 + ], + "lines": [ + { + "bbox": [ + 130, + 114, + 482, + 193 + ], + "spans": [ + { + "bbox": [ + 130, + 114, + 482, + 193 + ], + "type": "text", + "content": "Table 8: Comparison with the state-of-the-art methods on three referring image segmentation benchmarks with cIoU. (ft) denotes models further finetuned on RefCOCO+/g after mix training. We abbreviate the datasets: COCO(C) [24], LVIS(L) [13], RefCOCO(RC) [57], Object365(O365) [41], Video segmentation datasets(V), ADE20K(A) [59], COCO-Stuff(CS) [3], PACO-LVIS(PL) [38], PASCAL-Part(PP) [5], GranD(G) [39], VOC2010(VOC) [11], Visual Genome(VG) [16], Flicker30k(F30K) [35], MUSE(M) [40], gRefCOCO(gRC) [25], COCO-Interactive(Cl)." + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 148, + 201, + 462, + 319 + ], + "lines": [ + { + "bbox": [ + 148, + 201, + 462, + 319 + ], + "spans": [ + { + "bbox": [ + 148, + 201, + 462, + 319 + ], + "type": "table", + "html": "
MethodSegmentation DataLLM TypeRefCOCORefCOCO+RefCOCOg
valtestAtestBvaltestAtestBvaltest
SEEM-L [61]C, L, RC-------65.6-
UNINEXT-L [53]O365, C, RC, V-80.382.677.870.074.962.673.473.7
UNINEXT-H [53]O365, C, RC, V-82.283.481.372.576.466.274.776.4
LISA [17]A, CS, RC, PL, PPVicuna-7B74.176.571.162.467.456.566.468.5
LISA(ft) [17]A, CS, RC, PL, PPVicuna-7B74.979.172.365.170.858.167.970.6
GLaMM [39]G, RCVicuna-7B79.583.276.972.678.764.674.274.9
u-LLaVA [50]A, CS, RC, PL, VOCVicuna-7B80.482.777.872.276.666.874.875.6
PerceptionGPT [34]RC, VG, F30kVicuna-13B75.379.172.168.974.061.970.771.9
PixelLM [40]A, CS, RC, PL, MLlama2-13B73.076.568.266.371.758.369.370.5
GSVA [48]A, CS, RC, PL, PP, gRCLlama2-13B77.779.974.268.071.561.573.273.9
GSVA(ft) [48]A, CS, RC, PL, PP, gRCLlama2-13B79.281.777.170.373.863.675.777.0
PSALMC, RC, CIPhi-1.5 (1.3B)83.684.781.672.975.570.173.874.4
", + "image_path": "091e917b94172ac571def0e770bf591fa0e6fedab9a80e7f503fec3f76d739ae.jpg" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "table_body" + } + ], + "index": 3 + }, + { + "bbox": [ + 130, + 341, + 482, + 460 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 341, + 482, + 460 + ], + "spans": [ + { + "bbox": [ + 130, + 341, + 482, + 460 + ], + "type": "text", + "content": "Importance of VL-alignments. The visual-language alignment stage (i.e., first training stage) is to project the visual features to the text input space, and it is an important step towards making the LLM understand images. In Tab. 6, we examine the impact of this stage, and we found that without VL alignment, the performance of all four tasks becomes worse, with the performance of A150-OV and RefCOCO being significantly affected, for example, the A150-OV dropped by -1.1 mAP and RefCOCO even dropped by -4.4 cIoU, probably because these two tasks require a strong requirement on understanding the relationship between vision and language. This result also suggests that the VL alignment is essential, and the LMM-based segmentation models have strong potential." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 130, + 460, + 482, + 605 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 460, + 482, + 605 + ], + "spans": [ + { + "bbox": [ + 130, + 460, + 482, + 605 + ], + "type": "text", + "content": "Joint Training. Our architecture design and input schema help integrate various segmentation tasks so that they can be trained on one model. Tab. 7 shows the effect of this joint training on different tasks. For the task-specific models, we perform the training on the corresponding task data for " + }, + { + "bbox": [ + 130, + 460, + 482, + 605 + ], + "type": "inline_equation", + "content": "18\\mathrm{k}" + }, + { + "bbox": [ + 130, + 460, + 482, + 605 + ], + "type": "text", + "content": " iterations. In contrast, the joint training setting (see implementation section for details) has a total of " + }, + { + "bbox": [ + 130, + 460, + 482, + 605 + ], + "type": "inline_equation", + "content": "56\\mathrm{k}" + }, + { + "bbox": [ + 130, + 460, + 482, + 605 + ], + "type": "text", + "content": " training iterations, which corresponds to " + }, + { + "bbox": [ + 130, + 460, + 482, + 605 + ], + "type": "inline_equation", + "content": "14\\mathrm{k}" + }, + { + "bbox": [ + 130, + 460, + 482, + 605 + ], + "type": "text", + "content": " iterations per task. The results show that joint training of different tasks greatly improves the performance. This suggests that learning between tasks is mutually beneficial, which is also the secret of success in LLM. For example, the generic segmentation task helps refine the mask prediction in the referring segmentation, and referring expressions also enhance the model's ability to recognize more unseen categories, which in turn improves the performance of open-vocabulary segmentation tasks." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 131, + 622, + 397, + 634 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 131, + 622, + 397, + 634 + ], + "spans": [ + { + "bbox": [ + 131, + 622, + 397, + 634 + ], + "type": "text", + "content": "4.2 System-Level Comparison on In-Domain Tasks" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 130, + 641, + 481, + 665 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 641, + 481, + 665 + ], + "spans": [ + { + "bbox": [ + 130, + 641, + 481, + 665 + ], + "type": "text", + "content": "In this section, we compared our model on three in-domain tasks with other state-of-the-arts to illustrate the effectiveness of our approach." + } + ] + } + ], + "index": 7 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 181, + 91, + 447, + 102 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 181, + 91, + 447, + 102 + ], + "spans": [ + { + "bbox": [ + 181, + 91, + 447, + 102 + ], + "type": "text", + "content": "PSALM: Pixelwise SegmentAtion with Large Multi-Modal Model" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 470, + 92, + 479, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 470, + 92, + 479, + 100 + ], + "spans": [ + { + "bbox": [ + 470, + 92, + 479, + 100 + ], + "type": "text", + "content": "11" + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 10 + }, + { + "para_blocks": [ + { + "bbox": [ + 130, + 116, + 482, + 259 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 116, + 482, + 259 + ], + "spans": [ + { + "bbox": [ + 130, + 116, + 482, + 259 + ], + "type": "text", + "content": "Referring Segmentation. Most existing works aimed at getting LMMs to perform image segmentation are designed for reference segmentation tasks. We compare PSALM with other works on RefCOCO, RefCOCO+, and RefCOCOg, and Tab. 8 shows the results. Owing to the generalized and flexible design of PSALM and the advantages of joint training on multiple tasks and datasets, our system was able to achieve state-of-the-art (SOTA) performance on RefCOCO and RefCOCO+, and competitive performance on RefCOCOg with STOA, despite being driven by LLM with only 1.3B parameters. It is worth noting unlike methods such as LISA and GSVA, which may achieve improvements through task-specific fine-tuning (gray-labeled results), PSALM does not perform additional fine-tuning but still achieves better performance on RefCOCO and RefCOCO+ than their fine-tuned models." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 130, + 262, + 482, + 360 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 262, + 482, + 360 + ], + "spans": [ + { + "bbox": [ + 130, + 262, + 482, + 360 + ], + "type": "text", + "content": "Generic Segmentation. We evaluate PSALM with state-of-the-art methods on the COCO Panoptic Segmentation validation set (Tab. 9). Here, we follow the evaluation protocol used in Mask2Former to report PQ, which is the main metric for panoptic segmentation, mAP on thing classes for instance segmentation, and mIoU by merging instance masks from the same category for semantic segmentation. Compared to other methods, PSALM achieves comparable performance at similar visual backbone sizes, demonstrating that PSALM is a powerful architecture, even when compared to approaches designed for specific tasks." + } + ] + } + ], + "index": 3 + }, + { + "type": "table", + "bbox": [ + 184, + 429, + 427, + 505 + ], + "blocks": [ + { + "bbox": [ + 130, + 386, + 482, + 419 + ], + "lines": [ + { + "bbox": [ + 130, + 386, + 482, + 419 + ], + "spans": [ + { + "bbox": [ + 130, + 386, + 482, + 419 + ], + "type": "text", + "content": "Table 9: Comparison with the state-of-the-art methods on Panoptic COCO-val. We abbreviate the datasets: COCO-SAM(CM) [21], VIPSeg (VIP) [29], while others following Tab. 8." + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 184, + 429, + 427, + 505 + ], + "lines": [ + { + "bbox": [ + 184, + 429, + 427, + 505 + ], + "spans": [ + { + "bbox": [ + 184, + 429, + 427, + 505 + ], + "type": "table", + "html": "
MethodBackboneSeg. DataPQmAPmIoU
Mask2Former [6]Swin-BC55.145.265.1
Mask2Former [6]Swin-LC57.848.667.4
X-Decoder [60]DaViT-BC, L, RC56.245.866.0
SEEM [61]DaViT-BC, L, RC56.146.466.3
OMG-Seg [21]ConvNeXt-XXLC, VIP, CM, V55.4--
PSALMSwin-BC, RC, CI55.945.766.6
", + "image_path": "9c47837cddcdba2939e4d52662a0ceffd8e535d3612978b497d5da89696ba2a8.jpg" + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "table_body" + } + ], + "index": 5 + }, + { + "bbox": [ + 130, + 533, + 482, + 665 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 533, + 482, + 665 + ], + "spans": [ + { + "bbox": [ + 130, + 533, + 482, + 665 + ], + "type": "text", + "content": "Interactive Segmentation. We also evaluate PSALM in the interactive segmentation tasks. Since the task does not have a well-developed dataset containing all four instructions, previous works have typically used the in-house dataset, thus we re-evaluate other methods on the COCO interactive validation set. The results are shown in Tab. 10, and PSALM achieves leading performance on point, scribble, and mask instructions than all other methods, while on box instruction, SAM performs better on mIoU but worse on cIoU, and we hypothesize might be caused by the different distribution of training data, and the fact that SAM is trained on SA-1B [15], which has a much larger data scale than what we used. In addition, we also report the official results of SEEM, which only evaluated 600 samples from the COCO validation set as a reference." + } + ] + } + ], + "index": 6 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "spans": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "type": "text", + "content": "12" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 166, + 91, + 310, + 102 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 166, + 91, + 310, + 102 + ], + "spans": [ + { + "bbox": [ + 166, + 91, + 310, + 102 + ], + "type": "text", + "content": "Z. Zhang, Y. Ma, E. Zhang, X. Bai" + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 11 + }, + { + "para_blocks": [ + { + "type": "table", + "bbox": [ + 173, + 167, + 440, + 253 + ], + "blocks": [ + { + "bbox": [ + 130, + 114, + 482, + 158 + ], + "lines": [ + { + "bbox": [ + 130, + 114, + 482, + 158 + ], + "spans": [ + { + "bbox": [ + 130, + 114, + 482, + 158 + ], + "type": "text", + "content": "Table 10: Comparison with the state-of-the-art methods on COCO-Interactive. The results of SEEM-B* is the result reported in official paper, which is evaluated on 600 random samples of COCO-val, while all others are evaluated on all samples of COCO-val. Abbreviations for each dataset are the same as Tab. 9." + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 173, + 167, + 440, + 253 + ], + "lines": [ + { + "bbox": [ + 173, + 167, + 440, + 253 + ], + "spans": [ + { + "bbox": [ + 173, + 167, + 440, + 253 + ], + "type": "table", + "html": "
MethodSeg. DataPointScribbleBoxMask
mIoUcIoUmIoUcIoUmIoUcIoUmIoUcIoU
SAM-B [15]SA-1B48.733.6--73.768.7--
SAM-L [15]SA-1B51.837.7--76.671.6--
SEEM-B [61]C, L, RC47.857.843.044.044.942.148.465.0
SEEM-B* [61]C, L, RC81.7-83.5-75.7-76.0-
OMG-Seg [21]C, VIP, CM, V59.3-------
PSALMC, RC, CI64.374.066.980.067.380.967.682.4
", + "image_path": "f3006fe4d43851b612ef9f3c3ce317f4e80a6113ccf35e4c7da12e95facd171d.jpg" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "table_body" + } + ], + "index": 3 + }, + { + "bbox": [ + 131, + 273, + 369, + 285 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 131, + 273, + 369, + 285 + ], + "spans": [ + { + "bbox": [ + 131, + 273, + 369, + 285 + ], + "type": "text", + "content": "4.3 Generalizability on Out-of-Domain Tasks" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 130, + 292, + 482, + 400 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 292, + 482, + 400 + ], + "spans": [ + { + "bbox": [ + 130, + 292, + 482, + 400 + ], + "type": "text", + "content": "Thanks to the flexible design of architecture and input schema, multi-task joint-training, and the strong visual understanding capability of LMM, PSALM shows excellent performance on in-domain tasks, but more importantly, PSALM also demonstrates great potential to generalize to out-of-domain tasks in the zero-shot setting. In this section, we conduct experiments on three different out-of-domain tasks: open-vocabulary segmentation, generalized referring expression segmentation, and video object segmentation. We also tested the zero-shot result of the correspondence benchmark in Ego-Exo4D [12] and Video Object Segmentation in the Appendix." + } + ] + } + ], + "index": 5 + }, + { + "type": "table", + "bbox": [ + 199, + 473, + 413, + 561 + ], + "blocks": [ + { + "bbox": [ + 130, + 419, + 482, + 465 + ], + "lines": [ + { + "bbox": [ + 130, + 419, + 482, + 465 + ], + "spans": [ + { + "bbox": [ + 130, + 419, + 482, + 465 + ], + "type": "text", + "content": "Table 11: Comparison with the state-of-the-art methods on open-vocabulary instance segmentation and semantic segmentation benchmarks. We use mAP for instance segmentation and mIoU for semantic segmentation. We abbreviate the datasets: Pascal Context-459(PC459) [30], Pascal Context-59(PC59) [30], Pascal VOC-20(PAS20) [11]" + } + ] + } + ], + "index": 6, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 199, + 473, + 413, + 561 + ], + "lines": [ + { + "bbox": [ + 199, + 473, + 413, + 561 + ], + "spans": [ + { + "bbox": [ + 199, + 473, + 413, + 561 + ], + "type": "table", + "html": "
MethodOV Instance Seg.OV Semantic Seg.
A150CityscapesPC459A150PC59PAS20
MaskCLIP [9]6.0-10.023.745.9-
ODISE [49]14.4-14.529.957.3-
SAN [51]10.6-17.133.360.295.5
PSALM9.020.510.218.248.581.3
PSALM+LVIS13.919.314.024.457.295.0
", + "image_path": "8971ad6fb34aa6cf5de66b97eb72b0d84307fa19eadd464947f6582e0cad6a58.jpg" + } + ] + } + ], + "index": 7, + "angle": 0, + "type": "table_body" + } + ], + "index": 7 + }, + { + "bbox": [ + 130, + 582, + 491, + 666 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 582, + 491, + 666 + ], + "spans": [ + { + "bbox": [ + 130, + 582, + 491, + 666 + ], + "type": "text", + "content": "Open-Vocabulary Segmentation. We first evaluate PSALM in open-vocabulary segmentation tasks, which require the model to have the ability to deal with unseen categories in training. Here, we conduct experiments on both open-vocabulary instance segmentation and open-vocabulary semantic segmentation, and Tab. 11 shows the results. Without any special design, PSALM achieves reasonably good performance, although it is still worse than the best specific method in this task, such as SAN, but we believe that PSALM has a strong" + } + ] + } + ], + "index": 8 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 181, + 90, + 447, + 102 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 181, + 90, + 447, + 102 + ], + "spans": [ + { + "bbox": [ + 181, + 90, + 447, + 102 + ], + "type": "text", + "content": "PSALM: Pixelwise SegmentAtion with Large Multi-Modal Model" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 470, + 91, + 480, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 470, + 91, + 480, + 100 + ], + "spans": [ + { + "bbox": [ + 470, + 91, + 480, + 100 + ], + "type": "text", + "content": "13" + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 12 + }, + { + "para_blocks": [ + { + "bbox": [ + 130, + 116, + 482, + 200 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 116, + 482, + 200 + ], + "spans": [ + { + "bbox": [ + 130, + 116, + 482, + 200 + ], + "type": "text", + "content": "potential to be further improved by adding more diverse training data, which is advantages of our method. We have also made a preliminary attempt by further involving the LVIS dataset, and as we expected, the performance has significantly improvements. In addition, existing open-vocabulary segmentation methods are built upon the CLIP model or diffusion model, while our approach is based on LMM models, which is a new path and attempt to bring new inspiration to the community, which we believe is even more important than the performance." + } + ] + } + ], + "index": 2 + }, + { + "type": "table", + "bbox": [ + 184, + 251, + 427, + 369 + ], + "blocks": [ + { + "bbox": [ + 130, + 220, + 482, + 243 + ], + "lines": [ + { + "bbox": [ + 130, + 220, + 482, + 243 + ], + "spans": [ + { + "bbox": [ + 130, + 220, + 482, + 243 + ], + "type": "text", + "content": "Table 12: Our method's zero-shot performance on gRefCOCO. (ft) denotes models further fine-tuned on gRefCOCO after mix training." + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 184, + 251, + 427, + 369 + ], + "lines": [ + { + "bbox": [ + 184, + 251, + 427, + 369 + ], + "spans": [ + { + "bbox": [ + 184, + 251, + 427, + 369 + ], + "type": "table", + "html": "
MethodsLLM TypeZero-ShotvaltestAtestB
cIoUgIoUcIoUgIoUcIoUgIoU
MattNet [56]-X47.548.258.759.345.346.1
LTS [8]-X52.352.761.962.649.950.4
ReLA [25]-X62.463.669.370.059.961.0
LISA [17]Vicuna-7BX38.732.252.648.544.839.7
LISA(ft) [17]Vicuna-7BX61.763.369.270.160.361.3
GSVA [48]Vicuna-7BX61.763.369.270.160.361.3
GSVA(ft) [48]Vicuna-7BX63.366.569.971.160.562.2
PSALMPhi-1.5 (1.3B)✓42.043.352.454.550.652.5
", + "image_path": "60d51dc93f9282efd496d8c84eb9ec5935a24baf77684ee292b2825aa7d84a68.jpg" + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "table_body" + } + ], + "index": 4 + }, + { + "bbox": [ + 130, + 390, + 482, + 498 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 390, + 482, + 498 + ], + "spans": [ + { + "bbox": [ + 130, + 390, + 482, + 498 + ], + "type": "text", + "content": "Generalized Referring Expression Segmentation. The referring segmentation datasets used in training contain only a single object, however, the design of the mask proposal allows PSALM to directly address multi-target without any further training or fine-tuning. We evaluate the gRefCOCO benchmark which contains multiple segment targets. In practice, given an expression, we compute the similarity with all mask proposals and retain all masks with similarity greater than 0.6 as foreground. Tab. 12 shows the results, PSALM also achieved very promising performance, even outperforming the LISA version that only pre-trained on gRefCOCO but without task-specific fine-tuning." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 131, + 515, + 220, + 528 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 131, + 515, + 220, + 528 + ], + "spans": [ + { + "bbox": [ + 131, + 515, + 220, + 528 + ], + "type": "text", + "content": "5 Conclusion" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 130, + 539, + 482, + 599 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 539, + 482, + 599 + ], + "spans": [ + { + "bbox": [ + 130, + 539, + 482, + 599 + ], + "type": "text", + "content": "The PSALM proposed in this study extends the capability of LMM from text output tasks to image segmentation, addresses the output limitations of the LMM, and unifies various segmentation tasks. PSALM exhibits excellent performance in multiple in-domain tasks, and its generalization ability in out-of-domain tasks further underscores its potential." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 131, + 616, + 241, + 631 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 131, + 616, + 241, + 631 + ], + "spans": [ + { + "bbox": [ + 131, + 616, + 241, + 631 + ], + "type": "text", + "content": "Acknowledgments" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 130, + 641, + 481, + 665 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 641, + 481, + 665 + ], + "spans": [ + { + "bbox": [ + 130, + 641, + 481, + 665 + ], + "type": "text", + "content": "This work was supported by the National Science Fund for Distinguished Young Scholars of China (Grant No.62225603)." + } + ] + } + ], + "index": 9 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "spans": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "type": "text", + "content": "14" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 166, + 90, + 310, + 102 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 166, + 90, + 310, + 102 + ], + "spans": [ + { + "bbox": [ + 166, + 90, + 310, + 102 + ], + "type": "text", + "content": "Z. Zhang, Y. Ma, E. Zhang, X. Bai" + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 13 + }, + { + "para_blocks": [ + { + "bbox": [ + 133, + 114, + 197, + 127 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 114, + 197, + 127 + ], + "spans": [ + { + "bbox": [ + 133, + 114, + 197, + 127 + ], + "type": "text", + "content": "References" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 138, + 137, + 481, + 665 + ], + "type": "list", + "angle": 0, + "index": 18, + "blocks": [ + { + "bbox": [ + 138, + 137, + 481, + 180 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 137, + 481, + 180 + ], + "spans": [ + { + "bbox": [ + 138, + 137, + 481, + 180 + ], + "type": "text", + "content": "1. Alayrac, J.B., Donahue, J., Luc, P., Miech, A., Barr, I., Hasson, Y., Lenc, K., Mensch, A., Millican, K., Reynolds, M., et al.: Flamingo: a visual language model for few-shot learning. Advances in Neural Information Processing Systems 35, 23716-23736 (2022)" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 138, + 181, + 481, + 213 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 181, + 481, + 213 + ], + "spans": [ + { + "bbox": [ + 138, + 181, + 481, + 213 + ], + "type": "text", + "content": "2. Brown, T., Mann, B., Ryder, N., Subbiah, M., Kaplan, J.D., Dhariwal, P., Neelakantan, A., Shyam, P., Sastry, G., Askell, A., et al.: Language models are few-shot learners. Advances in neural information processing systems 33, 1877-1901 (2020)" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 138, + 213, + 481, + 245 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 213, + 481, + 245 + ], + "spans": [ + { + "bbox": [ + 138, + 213, + 481, + 245 + ], + "type": "text", + "content": "3. Caesar, H., Uijlings, J., Ferrari, V.: Coco-stuff: Thing and stuff classes in context. In: Proceedings of the IEEE conference on computer vision and pattern recognition. pp. 1209-1218 (2018)" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 138, + 245, + 481, + 277 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 245, + 481, + 277 + ], + "spans": [ + { + "bbox": [ + 138, + 245, + 481, + 277 + ], + "type": "text", + "content": "4. Chen, K., Zhang, Z., Zeng, W., Zhang, R., Zhu, F., Zhao, R.: Shikra: Unleashing multimodal llm's referential dialogue magic. arXiv preprint arXiv:2306.15195 (2023)" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 138, + 277, + 481, + 320 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 277, + 481, + 320 + ], + "spans": [ + { + "bbox": [ + 138, + 277, + 481, + 320 + ], + "type": "text", + "content": "5. Chen, X., Mottaghi, R., Liu, X., Fidler, S., Urtasun, R., Yuille, A.: Detect what you can: Detecting and representing objects using holistic models and body parts. In: Proceedings of the IEEE conference on computer vision and pattern recognition. pp. 1971-1978 (2014)" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 138, + 320, + 481, + 363 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 320, + 481, + 363 + ], + "spans": [ + { + "bbox": [ + 138, + 320, + 481, + 363 + ], + "type": "text", + "content": "6. Cheng, B., Misra, I., Schwing, A.G., Kirillov, A., Girdhar, R.: Masked-attention mask transformer for universal image segmentation. In: Proceedings of the IEEE/CVF conference on computer vision and pattern recognition. pp. 1290-1299 (2022)" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 138, + 363, + 481, + 407 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 363, + 481, + 407 + ], + "spans": [ + { + "bbox": [ + 138, + 363, + 481, + 407 + ], + "type": "text", + "content": "7. Chiang, W.L., Li, Z., Lin, Z., Sheng, Y., Wu, Z., Zhang, H., Zheng, L., Zhuang, S., Zhuang, Y., Gonzalez, J.E., et al.: Vicuna: An open-source chatbot impressing gpt-4 with " + }, + { + "bbox": [ + 138, + 363, + 481, + 407 + ], + "type": "inline_equation", + "content": "90\\%" + }, + { + "bbox": [ + 138, + 363, + 481, + 407 + ], + "type": "text", + "content": " chatgpt quality. See https://vicuna.lmsys.org (accessed 14 April 2023) (2023)" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 138, + 407, + 481, + 439 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 407, + 481, + 439 + ], + "spans": [ + { + "bbox": [ + 138, + 407, + 481, + 439 + ], + "type": "text", + "content": "8. Ding, H., Liu, C., Wang, S., Jiang, X.: Vision-language transformer and query generation for referring segmentation. In: Proceedings of the IEEE/CVF International Conference on Computer Vision. pp. 16321-16330 (2021)" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 138, + 439, + 481, + 460 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 439, + 481, + 460 + ], + "spans": [ + { + "bbox": [ + 138, + 439, + 481, + 460 + ], + "type": "text", + "content": "9. Ding, Z., Wang, J., Tu, Z.: Open-vocabulary universal image segmentation with maskclip (2023)" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 138, + 460, + 481, + 493 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 460, + 481, + 493 + ], + "spans": [ + { + "bbox": [ + 138, + 460, + 481, + 493 + ], + "type": "text", + "content": "0. Dong, R., Han, C., Peng, Y., Qi, Z., Ge, Z., Yang, J., Zhao, L., Sun, J., Zhou, H., Wei, H., et al.: Dreamllm: Synergistic multimodal comprehension and creation. arXiv preprint arXiv:2309.11499 (2023)" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 138, + 493, + 481, + 525 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 493, + 481, + 525 + ], + "spans": [ + { + "bbox": [ + 138, + 493, + 481, + 525 + ], + "type": "text", + "content": "1. Everingham, M., Van Gool, L., Williams, C.K.I., Winn, J., Zisserman, A.: The PASCAL Visual Object Classes Challenge 2010 (VOC2010) Results. http://www.pascal-network.org/challenges/VOC/voc2010/workshop/index.html" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 138, + 525, + 481, + 568 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 525, + 481, + 568 + ], + "spans": [ + { + "bbox": [ + 138, + 525, + 481, + 568 + ], + "type": "text", + "content": "2. Grauman, K., Westbury, A., Torresani, L., Kitani, K., Malik, J., Afouras, T., Ashutosh, K., Baiyya, V., Bansal, S., Boote, B., et al.: Ego-exo4d: Understanding skilled human activity from first-and third-person perspectives. arXiv preprint arXiv:2311.18259 (2023)" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 138, + 568, + 481, + 601 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 568, + 481, + 601 + ], + "spans": [ + { + "bbox": [ + 138, + 568, + 481, + 601 + ], + "type": "text", + "content": "3. Gupta, A., Dollar, P., Girshick, R.: Lvis: A dataset for large vocabulary instance segmentation. In: Proceedings of the IEEE/CVF conference on computer vision and pattern recognition. pp. 5356-5364 (2019)" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 138, + 601, + 481, + 633 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 601, + 481, + 633 + ], + "spans": [ + { + "bbox": [ + 138, + 601, + 481, + 633 + ], + "type": "text", + "content": "4. Jain, J., Li, J., Chiu, M.T., Hassani, A., Orlov, N., Shi, H.: Oneformer: One transformer to rule universal image segmentation. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 2989-2998 (2023)" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 138, + 633, + 481, + 665 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 633, + 481, + 665 + ], + "spans": [ + { + "bbox": [ + 138, + 633, + 481, + 665 + ], + "type": "text", + "content": "5. Kirillov, A., Mintun, E., Ravi, N., Mao, H., Rolland, C., Gustafson, L., Xiao, T., Whitehead, S., Berg, A.C., Lo, W.Y., et al.: Segment anything. arXiv preprint arXiv:2304.02643 (2023)" + } + ] + } + ], + "index": 17 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 181, + 91, + 447, + 102 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 181, + 91, + 447, + 102 + ], + "spans": [ + { + "bbox": [ + 181, + 91, + 447, + 102 + ], + "type": "text", + "content": "PSALM: Pixelwise SegmentAtion with Large Multi-Modal Model" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 470, + 92, + 480, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 470, + 92, + 480, + 100 + ], + "spans": [ + { + "bbox": [ + 470, + 92, + 480, + 100 + ], + "type": "text", + "content": "15" + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 14 + }, + { + "para_blocks": [ + { + "bbox": [ + 132, + 116, + 481, + 665 + ], + "type": "list", + "angle": 0, + "index": 18, + "blocks": [ + { + "bbox": [ + 133, + 116, + 481, + 160 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 116, + 481, + 160 + ], + "spans": [ + { + "bbox": [ + 133, + 116, + 481, + 160 + ], + "type": "text", + "content": "16. Krishna, R., Zhu, Y., Groth, O., Johnson, J., Hata, K., Kravitz, J., Chen, S., Kalantidis, Y., Li, L.J., Shamma, D.A., et al.: Visual genome: Connecting language and vision using crowdsourced dense image annotations. International journal of computer vision 123, 32-73 (2017)" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 133, + 161, + 481, + 182 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 161, + 481, + 182 + ], + "spans": [ + { + "bbox": [ + 133, + 161, + 481, + 182 + ], + "type": "text", + "content": "17. Lai, X., Tian, Z., Chen, Y., Li, Y., Yuan, Y., Liu, S., Jia, J.: Lisa: Reasoning segmentation via large language model. arXiv preprint arXiv:2308.00692 (2023)" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 133, + 183, + 481, + 205 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 183, + 481, + 205 + ], + "spans": [ + { + "bbox": [ + 133, + 183, + 481, + 205 + ], + "type": "text", + "content": "18. Li, B., Zhang, P., Yang, J., Zhang, Y., Pu, F., Liu, Z.: Otterhd: A high-resolution multi-modality model. arXiv preprint arXiv:2311.04219 (2023)" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 134, + 205, + 481, + 237 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 134, + 205, + 481, + 237 + ], + "spans": [ + { + "bbox": [ + 134, + 205, + 481, + 237 + ], + "type": "text", + "content": "19. Li, J., Li, D., Savarese, S., Hoi, S.: Blip-2: Bootstrapping language-image pretraining with frozen image encoders and large language models. arXiv preprint arXiv:2301.12597 (2023)" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 133, + 238, + 481, + 270 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 238, + 481, + 270 + ], + "spans": [ + { + "bbox": [ + 133, + 238, + 481, + 270 + ], + "type": "text", + "content": "20. Li, J., Li, D., Xiong, C., Hoi, S.: Blip: Bootstrapping language-image pre-training for unified vision-language understanding and generation. In: International Conference on Machine Learning. pp. 12888-12900. PMLR (2022)" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 132, + 271, + 481, + 303 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 271, + 481, + 303 + ], + "spans": [ + { + "bbox": [ + 132, + 271, + 481, + 303 + ], + "type": "text", + "content": "21. Li, X., Yuan, H., Li, W., Ding, H., Wu, S., Zhang, W., Li, Y., Chen, K., Loy, C.C.: Omg-seg: Is one model good enough for all segmentation? arXiv preprint arXiv:2401.10229 (2024)" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 132, + 304, + 481, + 325 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 304, + 481, + 325 + ], + "spans": [ + { + "bbox": [ + 132, + 304, + 481, + 325 + ], + "type": "text", + "content": "22. Li, Y., Bubeck, S., Eldan, R., Del Giorno, A., Gunasekar, S., Lee, Y.T.: Textbooks are all you need ii: phi-1.5 technical report. arXiv preprint arXiv:2309.05463 (2023)" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 132, + 326, + 481, + 369 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 326, + 481, + 369 + ], + "spans": [ + { + "bbox": [ + 132, + 326, + 481, + 369 + ], + "type": "text", + "content": "23. Li, Z., Yang, B., Liu, Q., Ma, Z., Zhang, S., Yang, J., Sun, Y., Liu, Y., Bai, X.: Monkey: Image resolution and text label are important things for large multi-modal models. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 26763-26773 (2024)" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 132, + 370, + 481, + 413 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 370, + 481, + 413 + ], + "spans": [ + { + "bbox": [ + 132, + 370, + 481, + 413 + ], + "type": "text", + "content": "24. Lin, T.Y., Maire, M., Belongie, S., Hays, J., Perona, P., Ramanan, D., Dollar, P., Zitnick, C.L.: Microsoft coco: Common objects in context. In: Computer Vision-ECCV 2014: 13th European Conference, Zurich, Switzerland, September 6-12, 2014, Proceedings, Part V 13. pp. 740-755. Springer (2014)" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 132, + 414, + 481, + 445 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 414, + 481, + 445 + ], + "spans": [ + { + "bbox": [ + 132, + 414, + 481, + 445 + ], + "type": "text", + "content": "25. Liu, C., Ding, H., Jiang, X.: Gres: Generalized referring expression segmentation. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 23592-23601 (2023)" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 132, + 447, + 481, + 468 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 447, + 481, + 468 + ], + "spans": [ + { + "bbox": [ + 132, + 447, + 481, + 468 + ], + "type": "text", + "content": "26. Liu, H., Li, C., Li, Y., Li, B., Zhang, Y., Shen, S., Lee, Y.J.: Llava-last: Improved reasoning,OCR,and world knowledge (2024)" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 132, + 468, + 481, + 490 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 468, + 481, + 490 + ], + "spans": [ + { + "bbox": [ + 132, + 468, + 481, + 490 + ], + "type": "text", + "content": "27. Liu, H., Li, C., Wu, Q., Lee, Y.J.: Visual instruction tuning. Advances in neural information processing systems 36 (2024)" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 132, + 491, + 481, + 533 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 491, + 481, + 533 + ], + "spans": [ + { + "bbox": [ + 132, + 491, + 481, + 533 + ], + "type": "text", + "content": "28. Liu, Z., Lin, Y., Cao, Y., Hu, H., Wei, Y., Zhang, Z., Lin, S., Guo, B.: Swin transformer: Hierarchical vision transformer using shifted windows. In: Proceedings of the IEEE/CVF international conference on computer vision. pp. 10012-10022 (2021)" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 132, + 534, + 481, + 577 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 534, + 481, + 577 + ], + "spans": [ + { + "bbox": [ + 132, + 534, + 481, + 577 + ], + "type": "text", + "content": "29. Miao, J., Wang, X., Wu, Y., Li, W., Zhang, X., Wei, Y., Yang, Y.: Large-scale video panoptic segmentation in the wild: A benchmark. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 21033-21043 (2022)" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 132, + 578, + 481, + 621 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 578, + 481, + 621 + ], + "spans": [ + { + "bbox": [ + 132, + 578, + 481, + 621 + ], + "type": "text", + "content": "30. Mottaghi, R., Chen, X., Liu, X., Cho, N.G., Lee, S.W., Fidler, S., Urtasun, R., Yuille, A.: The role of context for object detection and semantic segmentation in the wild. In: IEEE Conference on Computer Vision and Pattern Recognition (CVPR) (2014)" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 132, + 622, + 481, + 665 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 622, + 481, + 665 + ], + "spans": [ + { + "bbox": [ + 132, + 622, + 481, + 665 + ], + "type": "text", + "content": "31. Nagaraja, V.K., Morariu, V.I., Davis, L.S.: Modeling context between objects for referring expression understanding. In: Computer Vision-ECCV 2016: 14th European Conference, Amsterdam, The Netherlands, October 11-14, 2016, Proceedings, Part IV 14. pp. 792-807. Springer (2016)" + } + ] + } + ], + "index": 17 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "spans": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "type": "text", + "content": "16" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 166, + 91, + 309, + 101 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 166, + 91, + 309, + 101 + ], + "spans": [ + { + "bbox": [ + 166, + 91, + 309, + 101 + ], + "type": "text", + "content": "Z. Zhang, Y. Ma, E. Zhang, X. Bai" + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 15 + }, + { + "para_blocks": [ + { + "bbox": [ + 132, + 116, + 481, + 666 + ], + "type": "list", + "angle": 0, + "index": 18, + "blocks": [ + { + "bbox": [ + 132, + 116, + 323, + 128 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 116, + 323, + 128 + ], + "spans": [ + { + "bbox": [ + 132, + 116, + 323, + 128 + ], + "type": "text", + "content": "32. OpenAI: Gpt-4v(ision) system card (2023)" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 132, + 129, + 481, + 162 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 129, + 481, + 162 + ], + "spans": [ + { + "bbox": [ + 132, + 129, + 481, + 162 + ], + "type": "text", + "content": "33. Peng, Z., Wang, W., Dong, L., Hao, Y., Huang, S., Ma, S., Wei, F.: Kosmos-2: Grounding multimodal large language models to the world. arXiv preprint arXiv:2306.14824 (2023)" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 132, + 163, + 481, + 185 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 163, + 481, + 185 + ], + "spans": [ + { + "bbox": [ + 132, + 163, + 481, + 185 + ], + "type": "text", + "content": "34. Pi, R., Yao, L., Gao, J., Zhang, J., Zhang, T.: Perceptiongpt: Effectively fusing visual perception into llm. arXiv preprint arXiv:2311.06612 (2023)" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 132, + 186, + 481, + 229 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 186, + 481, + 229 + ], + "spans": [ + { + "bbox": [ + 132, + 186, + 481, + 229 + ], + "type": "text", + "content": "35. Plummer, B.A., Wang, L., Cervantes, C.M., Caicedo, J.C., Hockenmaier, J., Lazebnik, S.: Flickr30k entities: Collecting region-to-phrase correspondences for richer image-to-sentence models. In: Proceedings of the IEEE international conference on computer vision. pp. 2641-2649 (2015)" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 132, + 230, + 481, + 262 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 230, + 481, + 262 + ], + "spans": [ + { + "bbox": [ + 132, + 230, + 481, + 262 + ], + "type": "text", + "content": "36. Pont-Tuset, J., Perazzi, F., Caelles, S., Arbeláez, P., Sorkine-Hornung, A., Van Gool, L.: The 2017 davis challenge on video object segmentation. arXiv preprint arXiv:1704.00675 (2017)" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 132, + 264, + 481, + 308 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 264, + 481, + 308 + ], + "spans": [ + { + "bbox": [ + 132, + 264, + 481, + 308 + ], + "type": "text", + "content": "37. Radford, A., Kim, J.W., Hallacy, C., Ramesh, A., Goh, G., Agarwal, S., Sastry, G., Askell, A., Mishkin, P., Clark, J., et al.: Learning transferable visual models from natural language supervision. In: International conference on machine learning. pp. 8748-8763. PMLR (2021)" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 132, + 308, + 481, + 352 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 308, + 481, + 352 + ], + "spans": [ + { + "bbox": [ + 132, + 308, + 481, + 352 + ], + "type": "text", + "content": "38. Ramanathan, V., Kalia, A., Petrovic, V., Wen, Y., Zheng, B., Guo, B., Wang, R., Marquez, A., Kovvuri, R., Kadian, A., et al.: Paco: Parts and attributes of common objects. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 7141-7151 (2023)" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 132, + 353, + 481, + 385 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 353, + 481, + 385 + ], + "spans": [ + { + "bbox": [ + 132, + 353, + 481, + 385 + ], + "type": "text", + "content": "39. Rasheed, H., Maaz, M., Shaji, S., Shaker, A., Khan, S., Cholakkal, H., Anwer, R.M., Xing, E., Yang, M.H., Khan, F.S.: Glamm: Pixel grounding large multimodal model. arXiv preprint arXiv:2311.03356 (2023)" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 132, + 386, + 481, + 408 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 386, + 481, + 408 + ], + "spans": [ + { + "bbox": [ + 132, + 386, + 481, + 408 + ], + "type": "text", + "content": "40. Ren, Z., Huang, Z., Wei, Y., Zhao, Y., Fu, D., Feng, J., Jin, X.: Pixelm: Pixel reasoning with large multimodal model. arXiv preprint arXiv:2312.02228 (2023)" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 132, + 409, + 481, + 442 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 409, + 481, + 442 + ], + "spans": [ + { + "bbox": [ + 132, + 409, + 481, + 442 + ], + "type": "text", + "content": "41. Shao, S., Li, Z., Zhang, T., Peng, C., Yu, G., Zhang, X., Li, J., Sun, J.: Objects365: A large-scale, high-quality dataset for object detection. In: Proceedings of the IEEE/CVF international conference on computer vision. pp. 8430-8439 (2019)" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 132, + 443, + 481, + 508 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 443, + 481, + 508 + ], + "spans": [ + { + "bbox": [ + 132, + 443, + 481, + 508 + ], + "type": "text", + "content": "42. Sudre, C.H., Li, W., Vercauteren, T., Ourselin, S., Jorge Cardoso, M.: Generalised dice overlap as a deep learning loss function for highly unbalanced segmentations. In: Deep Learning in Medical Image Analysis and Multimodal Learning for Clinical Decision Support: Third International Workshop, DLMIA 2017, and 7th International Workshop, ML-CDS 2017, Held in Conjunction with MICCAI 2017, Quebec City, QC, Canada, September 14, Proceedings 3. pp. 240-248. Springer (2017)" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 132, + 509, + 481, + 542 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 509, + 481, + 542 + ], + "spans": [ + { + "bbox": [ + 132, + 509, + 481, + 542 + ], + "type": "text", + "content": "43. Sun, Q., Yu, Q., Cui, Y., Zhang, F., Zhang, X., Wang, Y., Gao, H., Liu, J., Huang, T., Wang, X.: Generative pretraining in multimodality. arXiv preprint arXiv:2307.05222 (2023)" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 132, + 543, + 481, + 575 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 543, + 481, + 575 + ], + "spans": [ + { + "bbox": [ + 132, + 543, + 481, + 575 + ], + "type": "text", + "content": "44. Team, G., Anil, R., Borgeaud, S., Wu, Y., Alayrac, J.B., Yu, J., Soricut, R., Schalkwyk, J., Dai, A.M., Hauth, A., et al.: Gemini: a family of highly capable multimodal models. arXiv preprint arXiv:2312.11805 (2023)" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 132, + 576, + 481, + 609 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 576, + 481, + 609 + ], + "spans": [ + { + "bbox": [ + 132, + 576, + 481, + 609 + ], + "type": "text", + "content": "45. Touvron, H., Martin, L., Stone, K., Albert, P., Almahairi, A., Babaei, Y., Bashlykov, N., Batra, S., Bhargava, P., Bhosale, S., et al.: Llama 2: Open foundation and fine-tuned chat models. arXiv preprint arXiv:2307.09288 (2023)" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 132, + 610, + 481, + 643 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 610, + 481, + 643 + ], + "spans": [ + { + "bbox": [ + 132, + 610, + 481, + 643 + ], + "type": "text", + "content": "46. Wang, W., Lv, Q., Yu, W., Hong, W., Qi, J., Wang, Y., Ji, J., Yang, Z., Zhao, L., Song, X., et al.: Cogvlm: Visual expert for pretrained language models. arXiv preprint arXiv:2311.03079 (2023)" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 132, + 643, + 481, + 666 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 643, + 481, + 666 + ], + "spans": [ + { + "bbox": [ + 132, + 643, + 481, + 666 + ], + "type": "text", + "content": "47. Wu, J., Jiang, Y., Liu, Q., Yuan, Z., Bai, X., Bai, S.: General object foundation model for images and videos at scale. arXiv preprint arXiv:2312.09158 (2023)" + } + ] + } + ], + "index": 17 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 181, + 91, + 447, + 102 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 181, + 91, + 447, + 102 + ], + "spans": [ + { + "bbox": [ + 181, + 91, + 447, + 102 + ], + "type": "text", + "content": "PSALM: Pixelwise SegmentAtion with Large Multi-Modal Model" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 470, + 92, + 480, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 470, + 92, + 480, + 100 + ], + "spans": [ + { + "bbox": [ + 470, + 92, + 480, + 100 + ], + "type": "text", + "content": "17" + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 16 + }, + { + "para_blocks": [ + { + "bbox": [ + 132, + 116, + 481, + 600 + ], + "type": "list", + "angle": 0, + "index": 16, + "blocks": [ + { + "bbox": [ + 133, + 116, + 480, + 149 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 116, + 480, + 149 + ], + "spans": [ + { + "bbox": [ + 133, + 116, + 480, + 149 + ], + "type": "text", + "content": "48. Xia, Z., Han, D., Han, Y., Pan, X., Song, S., Huang, G.: Gsva: Generalized segmentation via multimodal large language models. arXiv preprint arXiv:2312.10103 (2023)" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 133, + 150, + 481, + 194 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 150, + 481, + 194 + ], + "spans": [ + { + "bbox": [ + 133, + 150, + 481, + 194 + ], + "type": "text", + "content": "49. Xu, J., Liu, S., Vahdat, A., Byeon, W., Wang, X., De Mello, S.: Open-vocabulary panoptic segmentation with text-to-image diffusion models. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 2955-2966 (2023)" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 132, + 194, + 480, + 216 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 194, + 480, + 216 + ], + "spans": [ + { + "bbox": [ + 132, + 194, + 480, + 216 + ], + "type": "text", + "content": "50. Xu, J., Xu, L., Yang, Y., Li, X., Xie, Y., Huang, Y.J., Li, Y.: u-llava: Unifying multi-modal tasks via large language model. arXiv preprint arXiv:2311.05348 (2023)" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 132, + 216, + 480, + 248 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 216, + 480, + 248 + ], + "spans": [ + { + "bbox": [ + 132, + 216, + 480, + 248 + ], + "type": "text", + "content": "51. Xu, M., Zhang, Z., Wei, F., Hu, H., Bai, X.: Side adapter network for open-vocabulary semantic segmentation. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 2945-2954 (2023)" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 132, + 249, + 480, + 282 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 249, + 480, + 282 + ], + "spans": [ + { + "bbox": [ + 132, + 249, + 480, + 282 + ], + "type": "text", + "content": "52. Xu, M., Zhang, Z., Wei, F., Lin, Y., Cao, Y., Hu, H., Bai, X.: A simple baseline for open-vocabulary semantic segmentation with pre-trained vision-language model. In: European Conference on Computer Vision. pp. 736-753. Springer (2022)" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 132, + 282, + 480, + 315 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 282, + 480, + 315 + ], + "spans": [ + { + "bbox": [ + 132, + 282, + 480, + 315 + ], + "type": "text", + "content": "53. Yan, B., Jiang, Y., Wu, J., Wang, D., Luo, P., Yuan, Z., Lu, H.: Universal instance perception as object discovery and retrieval. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 15325-15336 (2023)" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 132, + 315, + 480, + 346 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 315, + 480, + 346 + ], + "spans": [ + { + "bbox": [ + 132, + 315, + 480, + 346 + ], + "type": "text", + "content": "54. Yang, S., Qu, T., Lai, X., Tian, Z., Peng, B., Liu, S., Jia, J.: An improved baseline for reasoning segmentation with large language model. arXiv preprint arXiv:2312.17240 (2023)" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 132, + 347, + 480, + 380 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 347, + 480, + 380 + ], + "spans": [ + { + "bbox": [ + 132, + 347, + 480, + 380 + ], + "type": "text", + "content": "55. You, H., Zhang, H., Gan, Z., Du, X., Zhang, B., Wang, Z., Cao, L., Chang, S.F., Yang, Y.: Ferret: Refer and ground anything anywhere at any granularity. arXiv preprint arXiv:2310.07704 (2023)" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 132, + 380, + 480, + 423 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 380, + 480, + 423 + ], + "spans": [ + { + "bbox": [ + 132, + 380, + 480, + 423 + ], + "type": "text", + "content": "56. Yu, L., Lin, Z., Shen, X., Yang, J., Lu, X., Bansal, M., Berg, T.L.: Matttnet: Modular attention network for referring expression comprehension. In: Proceedings of the IEEE conference on computer vision and pattern recognition. pp. 1307-1315 (2018)" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 132, + 424, + 480, + 467 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 424, + 480, + 467 + ], + "spans": [ + { + "bbox": [ + 132, + 424, + 480, + 467 + ], + "type": "text", + "content": "57. Yu, L., Poirson, P., Yang, S., Berg, A.C., Berg, T.L.: Modeling context in referring expressions. In: Computer Vision-ECCV 2016: 14th European Conference, Amsterdam, The Netherlands, October 11-14, 2016, Proceedings, Part II 14. pp. 69-85. Springer (2016)" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 132, + 468, + 480, + 490 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 468, + 480, + 490 + ], + "spans": [ + { + "bbox": [ + 132, + 468, + 480, + 490 + ], + "type": "text", + "content": "58. Zhang, A., Zhao, L., Xie, C.W., Zheng, Y., Ji, W., Chua, T.S.: Next-chat: An lmm for chat, detection and segmentation. arXiv preprint arXiv:2311.04498 (2023)" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 132, + 491, + 480, + 522 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 491, + 480, + 522 + ], + "spans": [ + { + "bbox": [ + 132, + 491, + 480, + 522 + ], + "type": "text", + "content": "59. Zhou, B., Zhao, H., Puig, X., Xiao, T., Fidler, S., Barriuso, A., Torralba, A.: Semantic understanding of scenes through the ade20k dataset. International Journal of Computer Vision 127, 302-321 (2019)" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 132, + 522, + 480, + 567 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 522, + 480, + 567 + ], + "spans": [ + { + "bbox": [ + 132, + 522, + 480, + 567 + ], + "type": "text", + "content": "60. Zou, X., Dou, Z.Y., Yang, J., Gan, Z., Li, L., Li, C., Dai, X., Behl, H., Wang, J., Yuan, L., et al.: Generalized decoding for pixel, image, and language. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 15116-15127 (2023)" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 132, + 567, + 480, + 600 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 567, + 480, + 600 + ], + "spans": [ + { + "bbox": [ + 132, + 567, + 480, + 600 + ], + "type": "text", + "content": "61. Zou, X., Yang, J., Zhang, H., Li, F., Li, L., Wang, J., Wang, L., Gao, J., Lee, Y.J.: Segment everything everywhere all at once. Advances in Neural Information Processing Systems 36 (2024)" + } + ] + } + ], + "index": 15 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "spans": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "type": "text", + "content": "18" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 166, + 91, + 310, + 102 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 166, + 91, + 310, + 102 + ], + "spans": [ + { + "bbox": [ + 166, + 91, + 310, + 102 + ], + "type": "text", + "content": "Z. Zhang, Y. Ma, E. Zhang, X. Bai" + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 17 + } + ], + "_backend": "vlm", + "_version_name": "2.6.4" +} \ No newline at end of file diff --git a/2024/PYRA_ Parallel Yielding Re-Activation for Training-Inference Efficient Task Adaptation/a96361d9-106b-4325-b9d1-39f6bd731ebb_content_list.json b/2024/PYRA_ Parallel Yielding Re-Activation for Training-Inference Efficient Task Adaptation/a96361d9-106b-4325-b9d1-39f6bd731ebb_content_list.json new file mode 100644 index 0000000000000000000000000000000000000000..cbdc77d2048ab1a5bbd6fab280edd2a4c30d040b --- /dev/null +++ b/2024/PYRA_ Parallel Yielding Re-Activation for Training-Inference Efficient Task Adaptation/a96361d9-106b-4325-b9d1-39f6bd731ebb_content_list.json @@ -0,0 +1,1828 @@ +[ + { + "type": "text", + "text": "PYRA: Parallel Yielding Re-Activation for Training-Inference Efficient Task Adaptation", + "text_level": 1, + "bbox": [ + 246, + 140, + 754, + 186 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Yizhe Xiong $^{1,2}$ , Hui Chen $^{2*}$ , Tianxiang Hao $^{1,2}$ , Zijia Lin $^{1}$ , Jungong Han $^{2,3}$ , Yuesong Zhang $^{4}$ , Guoxin Wang $^{4}$ , Yongjun Bao $^{4}$ , and Guiguang Ding $^{1,2}$", + "bbox": [ + 222, + 210, + 781, + 257 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "$^{1}$ School of Software, Tsinghua University, Beijing, China", + "bbox": [ + 307, + 268, + 692, + 284 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "$^{2}$ BNRist, Tsinghua University, Beijing, China", + "bbox": [ + 346, + 282, + 656, + 297 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "$^{3}$ Department of Automation, Tsinghua University, Beijing, China", + "bbox": [ + 279, + 296, + 720, + 311 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "4 JD.com, Beijing, China", + "bbox": [ + 416, + 311, + 586, + 325 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "{xiongyizhe2001,beyondtx,jungonghan77}@gmail.com", + "bbox": [ + 310, + 325, + 691, + 340 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "huichen@mail.tsinghua.edu.cn", + "bbox": [ + 390, + 340, + 609, + 353 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "{zhangyuesong1,wangguoxin14,baoyongjun}@jd.com", + "bbox": [ + 323, + 354, + 679, + 367 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "linzijia07@tsinghua.org.cn dinggg@tsinghua.edu.cn", + "bbox": [ + 307, + 368, + 694, + 381 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Abstract. Recently, the scale of transformers has grown rapidly, which introduces considerable challenges in terms of training overhead and inference efficiency in the scope of task adaptation. Existing works, namely Parameter-Efficient Fine-Tuning (PEFT) and model compression, have separately investigated the challenges. However, PEFT cannot guarantee the inference efficiency of the original backbone, especially for large-scale models. Model compression requires significant training costs for structure searching and re-training. Consequently, a simple combination of them cannot guarantee accomplishing both training efficiency and inference efficiency with minimal costs. In this paper, we propose a novel Parallel Yielding Re-Activation (PYRA) method for such a challenge of training-inference efficient task adaptation. PYRA first utilizes parallel yielding adaptive weights to comprehensively perceive the data distribution in downstream tasks. A re-activation strategy for token modulation is then applied for tokens to be merged, leading to calibrated token features. Extensive experiments demonstrate that PYRA outperforms all competing methods under both low compression rate and high compression rate, demonstrating its effectiveness and superiority in maintaining both training efficiency and inference efficiency for large-scale foundation models. Our code is available at https://github.com/THU-MIG/PyRA.", + "bbox": [ + 259, + 412, + 743, + 691 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Keywords: Vision Transformer $\\cdot$ Task Adaptation $\\cdot$ Model Compression", + "bbox": [ + 261, + 703, + 740, + 718 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "1 Introduction", + "text_level": 1, + "bbox": [ + 215, + 741, + 375, + 757 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Vision transformers [16] have made a profound impact across various domains of computer vision, such as image classification [6,15,16,50,55,66], object detection [4,40,47,48,74], image segmentation [11,49,53,70,72], etc. In recent years, the", + "bbox": [ + 212, + 771, + 787, + 816 + ], + "page_idx": 0 + }, + { + "type": "page_footnote", + "text": "* Corresponding Author", + "bbox": [ + 217, + 824, + 385, + 839 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/5664647c2b736e0e82a403b7c620536892af767f6bc456d67ad6a3f18ccce171.jpg", + "image_caption": [ + "(a) Full Fine-tuning" + ], + "image_footnote": [], + "bbox": [ + 222, + 146, + 380, + 208 + ], + "page_idx": 1 + }, + { + "type": "image", + "img_path": "images/2ab152f60759215e24818de68dbcf2edb03dee47b612c65aa58460f983ac19a7.jpg", + "image_caption": [ + "(c) PEFT" + ], + "image_footnote": [], + "bbox": [ + 222, + 223, + 380, + 282 + ], + "page_idx": 1 + }, + { + "type": "image", + "img_path": "images/36b7ea242bb2fc69de46fb442daa7344f5f931b70f4eefd2392f17f1de845c7d.jpg", + "image_caption": [ + "(b) Model compression" + ], + "image_footnote": [], + "bbox": [ + 400, + 148, + 560, + 209 + ], + "page_idx": 1 + }, + { + "type": "image", + "img_path": "images/e939f1d5fbe6cc22b2fc2ddb8383ad38f11f5c6b2de37f5afaa3082c0f1cf27b.jpg", + "image_caption": [ + "(d) Ours", + "Fig. 1: (a) Full fine-tuning trains all parameters on downstream tasks and utilizes the trained model for inference, thereby lacking efficiency in both training and inference stages. (b) Model compression employs pruning to enhance inference efficiency, but the pruned model necessitates extensive re-training on large-scale data. (c) PEFT freezes the model backbone and only fine-tunes a small amount of parameters, yet retains the inference complexity. (d) Our training-inference efficient task adaptation incorporates the advantages of all existing pipelines by training inference-efficient models with minimal tunable parameters." + ], + "image_footnote": [], + "bbox": [ + 400, + 223, + 558, + 282 + ], + "page_idx": 1 + }, + { + "type": "table", + "img_path": "images/d733460eb81370b56af7d9c23d8cbcde0c31a3a7d07b8fdc3776a55a00fff777.jpg", + "table_caption": [], + "table_footnote": [], + "table_body": "
Method(a)(b)(c)(d)
Non-vast Data Scale✔×✔✔
Training Efficient××✔✔
Inference Efficient×✔×✔
", + "bbox": [ + 573, + 146, + 782, + 285 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "scale of vision transformers has grown to be billion-parameters [29,61,66]. Consequently, adapting such models with vast scales into downstream tasks presents increasingly complex challenges, particularly in real-world deployment scenarios. Two critical concerns have been widely acknowledged as primary obstacles in implementing large-scale transformers for downstream applications [17,23,27,46]: (1) the training overhead when fine-tuning on downstream tasks, and (2) the inference efficiency after model deployment.", + "bbox": [ + 212, + 455, + 787, + 561 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Specifically, first, conventional fine-tuning methods, which necessitate adjusting all parameters of the model (i.e., Full Fine-tuning), suffer from unaffordable consumption of GPU resources and training time given the extensive scales of the foundation models [16, 29, 66]. Researchers have delved into Parameter-Efficient Fine-Tuning (PEFT) [17, 19, 23, 24, 26, 36] algorithms, which generally freeze the pre-trained models and only tune extra small parameters, leading to great reduction of training time and storage overhead. The second issue pertains to inference efficiency, requiring the deployed model to promptly process the input data. The computational complexity of the models significantly influences achieving satisfactory inference throughput. Representative solutions for this matter encompass model compression methods, including model pruning [7, 10, 56, 64], knowledge distillation [1, 21, 52, 57, 58], model quantization [3, 13, 37], etc.", + "bbox": [ + 212, + 565, + 787, + 747 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "In the literature, these two intriguing topics are investigated separately. PEFT methods either identify a subset of tunable parameters in the backbone for fine-tuning [65] or introduce learnable parameters to the frozen backbone during fine-tuning [17, 23, 24, 26, 36]. While effectively reducing training costs, most of these methods inevitably escalate computational complexity, resulting in inefficient inference. Model compression methods frequently require signifi", + "bbox": [ + 212, + 750, + 787, + 840 + ], + "page_idx": 1 + }, + { + "type": "page_number", + "text": "2", + "bbox": [ + 217, + 114, + 228, + 126 + ], + "page_idx": 1 + }, + { + "type": "header", + "text": "Y. Xiong et al.", + "bbox": [ + 271, + 114, + 372, + 128 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "cant computational resources to identify optimal structures for pruning. After pruning, a comprehensive re-training process using a substantial amount of data is crucial to prevent significant performance degradation. Therefore, model compression methods are typically inefficient in training efficiency.", + "bbox": [ + 212, + 146, + 787, + 207 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "These observations naturally lead to the question: can we achieve both training efficiency and inference efficiency simultaneously for downstream tasks? We refer to this challenge as Training-Inference Efficient Task Adaptation (seeing Fig. 1). Exploring this issue can enable us to conveniently deploy the advanced large-scale foundation models in real-world downstream applications with minimal costs, which is appealing and essential for the widespread implementation of foundation models. A straightforward solution is combining PEFT and model compression. However, for efficient task adaptation, a", + "bbox": [ + 212, + 208, + 433, + 537 + ], + "page_idx": 2 + }, + { + "type": "image", + "img_path": "images/d910bd2d0f40de2b7791e9e805efde4c4695c9c3c911f9b850789508a6ca9060.jpg", + "image_caption": [ + "Fig.2: Comparisons between simply combining ToMe [2] and LoRA [24] and our proposed PYRA. Red boxes represent the performance drop problem in low compression rates. Green boxes represent the adverse compression in high compression rates. See Sec. 4 for more results." + ], + "image_footnote": [], + "bbox": [ + 442, + 238, + 784, + 402 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "heavy re-training stage is unaffordable. Consequently, simply combining PEFT and model compression can easily suffer from substantial performance drops. For instance, we could integrate LoRA [24], a notable PEFT method that introduces a small low-rank adapter, with ToMe [2], a parameter-free model compression technique for vision transformers5. As shown in Fig. 2, under lower compression rates (around $1.7 \\times$ ), performance on both backbones (ViT-L/16 and ViT-B/16) present slight drops ( $< 1\\%$ ) compared to directly conducting PEFT on the backbone, indicating that the ToMe+LoRA combination serves as a basic solution within lower compression rate range, yet performance improvements are also demanded. Under high compression rates ( $>3.0 \\times$ ), the performance quickly drops and is even inferior to directly fine-tuning the small-scale backbone with corresponding throughput. We term this phenomenon as Adverse Compression. Both phenomena indicate that directly combining existing works cannot effectively address the new challenges.", + "bbox": [ + 212, + 539, + 787, + 750 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "In this paper, we propose a novel Parallel Yielding Re-Accivitation method (PYRA) designed for training-inference efficient task adaptation using vision transformers. Generally, the proposed PYRA follows the token merging paradigm [2, 8, 43] for inference efficiency. For effective task adaptation, we propose to mod-", + "bbox": [ + 212, + 752, + 807, + 813 + ], + "page_idx": 2 + }, + { + "type": "header", + "text": "PYRA: Parallel Yielding Re-Activation", + "bbox": [ + 468, + 114, + 730, + 128 + ], + "page_idx": 2 + }, + { + "type": "page_number", + "text": "3", + "bbox": [ + 774, + 116, + 784, + 126 + ], + "page_idx": 2 + }, + { + "type": "page_footnote", + "text": "In Sec. 4, we show that $\\mathrm{ToMe + LoRA}$ is a neat and strong solution.", + "bbox": [ + 217, + 824, + 679, + 839 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "ulate token features from both feature tokens and feature channels and parallelly yield weights for tokens to be merged by small modulation weight generators. These parallel yielding weights can comprehensively perceive the data distribution in downstream tasks. They are then applied to features through re-activation, resulting in adaptive token modulation. Thanks to such a token modulation strategy, PYRA can adaptively calibrate the learned feature distribution for downstream tasks with low computational complexity, ultimately leading to effective training and efficient inference.", + "bbox": [ + 212, + 146, + 787, + 267 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "We conduct extensive experiments to verify the effectiveness of our PYRA. We show that, under a low compression rate with $\\sim 1.7\\times$ speedup, PYRA introduces negligible performance drops. Under a high compression rate with $>3.0\\times$ speedup, PYRA eliminates the adverse compression gap. We emphasize that considering the scenario of high compression rates is practical, given the substantial size of current transformers. Concurrently, the specific model sizes required by downstream applications may not match any publicly available models, thus requiring the acquisition through high compression rates while keeping the performance comparable. In this regard, our approach represents an effective method for obtaining small models in the absence of pre-trained parameters for smaller-scale models.", + "bbox": [ + 212, + 267, + 787, + 431 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Overall, we summarize our contribution as follows.", + "bbox": [ + 238, + 433, + 604, + 448 + ], + "page_idx": 3 + }, + { + "type": "list", + "sub_type": "text", + "list_items": [ + "- We propose a novel challenge termed training-inference efficient task adaptation, in which the inference efficiency of large-scale transformers is escalated during parameter-efficient task adaptation.", + "- We propose PYRA for training-inference efficient task adaptation, which enhances the perception of feature distribution via token modulation. We generate parallel yielding decoupled weights to comprehensively perceive the feature distributions. We apply a re-activation strategy to modulate tokens to be merged for calibrated token features.", + "- Extensive experiments show that our PYRA outperforms all competing methods under both low compression rate and high compression rate. Further analysis shows that PYRA is effective across a series of different transformer backbones and model scales, well demonstrating the effectiveness and superiority of our method." + ], + "bbox": [ + 225, + 455, + 784, + 648 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "2 Related Works", + "text_level": 1, + "bbox": [ + 215, + 672, + 397, + 690 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Parameter-Efficient Fine-Tuning (PEFT) for Task Adaptation. Transferring large-scale transformers to downstream tasks has been a popular topic in computer vision [9, 17, 26, 28, 36, 60, 68, 69]. PEFT methods either locate a subset of parameters inside the model for fine-tuning [65], or inject human designed modules to the original model structure. Specifically, adapters [23, 28, 42, 45] are a type of MLP module with a bottleneck in the middle. Prompt-tuning [14, 18, 26, 32, 33, 39] insert learnable tokens to model input to generate task-specific outputs. SSF [36] tunes additional scaling and shifting parameters. LoRA [24], AdaptFormer [9], and Consolidator [17] add lightweight modules as", + "bbox": [ + 212, + 703, + 787, + 840 + ], + "page_idx": 3 + }, + { + "type": "page_number", + "text": "4", + "bbox": [ + 217, + 114, + 228, + 126 + ], + "page_idx": 3 + }, + { + "type": "header", + "text": "Y. Xiong et al.", + "bbox": [ + 271, + 114, + 370, + 128 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "bypasses. These modules can be merged with the original backbone for no extra inference cost. In the scope of training-inference efficient task adaptation, PEFT methods retain model inference cost, resulting in extreme difficulty for model deployment. To solve the problem, we conduct adaptive token merging for PEFT. While achieving promising adaptation performance under both low and high compression rates, our method inherits the advantages of PEFT methods. Model Compression. Model compression is often applied on large-scale models to acquire smaller-scale models with comparable performance. Mainstream approaches of model compression include model pruning [2,8,10,54,56,64,71], knowledge distillation [1,21,52] and model quantization [3,13,37,38]. For transformer models, model pruning methods can be roughly grouped into two categories: channel pruning and token pruning. Channel pruning methods [5,10, 64,73] reduce the number of parameters, channels, heads, or blocks. Recently, token pruning has emerged as another mainstream approach. Several works have attempted to prune tokens for vision transformers (ViTs). Among these methods, token pooling [43] uses a slow k-means approach that does not work for an off-the-shelf model. ToMe [2] constructs bipartite graphs and merges token pairs with the most weighted connections. DiffRate [8] combines token pruning and token merging with searched optimal token reduction rates for each layer. Although achieving promising results, most existing model compression methods fail when combining with PEFT for training-inference efficient task adaptation since they usually involve a heavy training stage. Model pruning methods demand full re-training after pruning to restore performance. Knowledge distillation necessitates training a small-scale model from scratch, demanding large amount of data. As for model quantization, although no heavy training is demanded, post-training quantization [41,63] achieves poor performance compared to quantization-aware training [25,34] that demands full re-training on the quantized model. As a comparison, our proposed PYRA achieves promising performance on compressed models under the restrictions of PEFT, surpassing the baselines of simply combining model compression and PEFT.", + "bbox": [ + 212, + 142, + 787, + 603 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "3 Methodology", + "text_level": 1, + "bbox": [ + 215, + 623, + 380, + 641 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "3.1 Preliminaries", + "text_level": 1, + "bbox": [ + 215, + 655, + 372, + 670 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "ViT Model. In this paper, we mainly focus on the training-inference efficient task adaptation of ViT models [16, 20, 51]. A ViT model consists of $L$ identical encoder blocks, each of which consists of a multi-head self-attention (MHSA) module and a feed-forward network (FFN). Formally, an input image $\\mathbf{x}$ is reshaped and linear projected to $N$ tokens $\\mathbf{x} = [t_1, t_2, \\dots, t_N]$ in $D$ dimensions. For simplicity, we omit the classification token ([CLS]) and the distillation token [51]. For encoder block $l$ , we denote the input as $\\mathbf{x}^l = [t_1^l, t_2^l, \\dots, t_{N^{l-1}}^l] \\in \\mathbb{R}^{N^{l-1} \\times D}$ and the output as $\\hat{\\mathbf{x}}^l = [\\hat{t}_1^l, \\hat{t}_2^l, \\dots, \\hat{t}_{N^l}^l] \\in \\mathbb{R}^{N^l \\times D}$ . For MHSA, the input tokens are first processed by three FC layers to generate $Q, K$ , and $V$ matrices, and the output is calculated by Softmax $(\\frac{QK^T}{\\sqrt{D}}) V$ before being projected by another FC", + "bbox": [ + 212, + 680, + 789, + 843 + ], + "page_idx": 4 + }, + { + "type": "header", + "text": "PYRA: Parallel Yielding Re-Activation", + "bbox": [ + 468, + 114, + 730, + 128 + ], + "page_idx": 4 + }, + { + "type": "page_number", + "text": "5", + "bbox": [ + 774, + 116, + 784, + 126 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "layer. For FFN, the tokens are projected by two FC layers. Our method mainly focuses on the input tokens $\\mathbf{x}^l$ before feeding them to the MHSA module.", + "bbox": [ + 212, + 146, + 782, + 176 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "LoRA. LoRA [24] is a widely employed PEFT method for task adaptation. Vision transformers consist of large dense parameter matrices. When adapting to a specific task, the updates to the matrices are in small subspaces and can be modeled with low-rank decompositions. LoRA trains only the decomposed matrices during fine-tuning. Specifically, for dense matrix $W_0 \\in \\mathbb{R}^{d \\times k}$ and input $x$ , the modified forward pass of updated $W_0$ is:", + "bbox": [ + 212, + 176, + 784, + 267 + ], + "page_idx": 5 + }, + { + "type": "equation", + "text": "\n$$\n\\hat {x} = W _ {0} x + B A x, \\tag {1}\n$$\n", + "text_format": "latex", + "bbox": [ + 439, + 279, + 784, + 294 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "where $B \\in \\mathbb{R}^{d \\times h}$ and $A \\in \\mathbb{R}^{h \\times k}$ . During inference, $BA$ can be merged with $W_0$ for no extra computation overhead.", + "bbox": [ + 212, + 301, + 782, + 333 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Token Merging. Token merging [2,8] is a parameter-free compression technique for vision transformers. It is orthogonal to the transformer structure and capable of flexibly changing the compression rate. Specifically, the input tokens $\\mathbf{x}^l$ of the $l$ -th ViT block are randomly separated before the MHSA:", + "bbox": [ + 212, + 334, + 782, + 393 + ], + "page_idx": 5 + }, + { + "type": "equation", + "text": "\n$$\nG _ {1} ^ {l} = \\left[ t _ {i _ {1}} ^ {l}, t _ {i _ {2}} ^ {l}, \\dots , t _ {i _ {s}} ^ {l} \\right], \\quad G _ {2} ^ {l} = \\left[ t _ {j _ {1}} ^ {l}, t _ {j _ {2}} ^ {l}, \\dots , t _ {j _ {s}} ^ {l} \\right], \\quad 2 s = N ^ {l - 1}. \\tag {2}\n$$\n", + "text_format": "latex", + "bbox": [ + 287, + 404, + 784, + 422 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Then, for each $t_{i_\\bullet}^l$ token, the most similar $t_{j_\\bullet}^l$ token is matched to it via cosine similarity. From $G_1^l$ , $r$ tokens with the most similar connections to tokens in $G_2^l$ are selected to form token pairs $(t_{m_k}^l, t_{n_k}^l)$ , where $t_{m_k}^l \\in G_1^l$ , $t_{n_k}^l \\in G_2^l$ , and $k = 1, 2, \\dots, r$ . Note that $\\{m_k\\}$ is a re-indexed subset of $\\{i_\\bullet\\}$ . Formally,", + "bbox": [ + 212, + 431, + 782, + 494 + ], + "page_idx": 5 + }, + { + "type": "equation", + "text": "\n$$\n\\left\\{t _ {m _ {k}} ^ {l} \\right\\} = \\arg \\operatorname {T o p r} _ {\\hat {t} _ {i} ^ {l}} \\left(\\max _ {j} \\frac {\\hat {t} _ {i} ^ {l} \\cdot t _ {j} ^ {l}}{\\| \\hat {t} _ {i} ^ {l} \\| \\cdot \\| t _ {j} ^ {l} \\|}\\right), \\quad t _ {n _ {k}} ^ {l} = \\arg \\max _ {t _ {j} ^ {l}} \\left(\\frac {t _ {m _ {k}} ^ {l} \\cdot t _ {j} ^ {l}}{\\| t _ {m _ {k}} ^ {l} \\| \\cdot \\| t _ {j} ^ {l} \\|}\\right) \\tag {3}\n$$\n", + "text_format": "latex", + "bbox": [ + 271, + 505, + 784, + 544 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Previous works [2, 8, 43], generally merge tokens via average pooling, cutting down the computational cost by decreasing the token number by $r$ in layer $l$ .", + "bbox": [ + 212, + 553, + 782, + 583 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "We employ the above techniques as our baseline method, in which we attach token merging while fine-tuning LoRA for task adaptation. These methods are selected due to their advantages being highly compatible with training-inference efficient task adaptation. First, token merging is parameter-free and training-free, and does not change model structures, which retains the storage-efficient advantage of PEFT. Besides, we choose LoRA for its popularity, simplicity, and mergeability (not introducing extra FLOPs during inference). Experimental results show that $\\mathrm{ToMe + LoRA}$ is a strong baseline (see Sec. 4).", + "bbox": [ + 212, + 584, + 784, + 705 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "3.2 PYRA: Parallel Yielding Re-Activation", + "text_level": 1, + "bbox": [ + 214, + 724, + 586, + 739 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Conventional fine-tuning methods excel at guiding the model to dynamically align with the target data distribution in downstream tasks by extensively adjusting parameters. However, in the context of training-inference efficient task adaptation, only a fraction of parameters can be fine-tuned, posing significant challenges in accurately capturing the nuances of data distribution. While reducing model complexity through token merging shows promise in enhancing", + "bbox": [ + 212, + 750, + 782, + 840 + ], + "page_idx": 5 + }, + { + "type": "page_number", + "text": "6", + "bbox": [ + 217, + 114, + 228, + 126 + ], + "page_idx": 5 + }, + { + "type": "header", + "text": "Y. Xiong et al.", + "bbox": [ + 271, + 114, + 370, + 128 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/9cfda996373e5beb11405400e7f727dd2c70b99c443e8b30987a29f36a116397.jpg", + "image_caption": [ + "Fig. 3: The pipeline of our PYRA. PYRA conducts token modulation before the MHSA module in each transformer block. Inside PYRA, a pair of learnable modulation weight generators are leveraged to generate adaptive modulation weights parallelly. After that, generated weights modulate tokens through re-activation. The generators in PYRA can be trained along with the LoRA module $\\theta_{p}$ in an end-to-end manner." + ], + "image_footnote": [], + "bbox": [ + 218, + 143, + 782, + 282 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "inference efficiency, it introduces the risk of information loss during layer-wise processing in vision transformers. This loss is difficult to rectify due to the constrained understanding of data distribution in the efficient task adaptation scenario. Therefore, a straightforward combination of token merging with PEFT algorithms for achieving training-inference efficient task adaptation may not yield optimal results, as depicted in Fig. 2.", + "bbox": [ + 212, + 397, + 782, + 487 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Here, we propose Parallel Yielding Re-Aviation (PYRA) to adaptively modulate token features to enhance the perception of data distribution during token merging. Specifically, inside PYRA, the weights for adaptive merging are first yielded in a parallel manner through a pair of lightweight learnable vectors in each ViT block. These generated weights are then applied to modulate tokens to be merged through re-activation. As a result, PYRA enables adaptive calibration of the learned feature distribution with low computational complexity.", + "bbox": [ + 212, + 488, + 784, + 594 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Parallel Yielding Adaptive Weights. We aim to optimize the merging process of each chosen token pair. Inspired by feature modulation [44, 59], we propose to modulate token features before merging. Formally, for encoder block $l$ with $r$ pairs of tokens in $D$ dimensions to be merged, we group the $t_{m_k}^l$ and $t_{n_k}^l$ tokens as token matrices $M_s^l = [t_{m_1}^l,\\dots ,t_{m_r}^l ]$ and $M_t^l = [t_{n_1}^l,\\dots ,t_{n_r}^l ]$ , where $M_s^l,M_t^l\\in \\mathbb{R}^{D\\times r}$ . We learn a modulation matrix $W^{l}\\in \\mathbb{R}^{D\\times r}$ that adaptively modulates tokens at the granularity of each channel. We emphasize that directly learning $W^{l}$ is redundant [24] and cannot adaptively satisfy the conditions with different images and token pairs. Therefore, we further specify the goal as learning decoupled weights $\\delta_D^l = f_D(M_s^l,M_t^l)\\in \\mathbb{R}^{D\\times 1}$ for feature channels and $\\delta_r^l = f_r(M_s^l,M_t^l)\\in \\mathbb{R}^{1\\times r}$ for feature tokens, where $W^{l} = \\delta_{D}^{l}\\delta_{r}^{l}$ , separately. We propose to generate $\\delta_D^l$ and $\\delta_r^l$ in a parallel yielding manner.", + "bbox": [ + 212, + 595, + 785, + 777 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Specifically, for layer $l$ with $r$ tokens to be merged, we create two learnable vectors as the modulation weight generator inside the transformer block: $W_r^l \\in \\mathbb{R}^{r \\times 1}$ and $W_D^l \\in \\mathbb{R}^{1 \\times D}$ . To guarantee that $W_r^l$ and $W_D^l$ digest the token features from both tokens in a token pair, we first normalize the sum of token pairs to", + "bbox": [ + 212, + 779, + 785, + 840 + ], + "page_idx": 6 + }, + { + "type": "header", + "text": "PYRA: Parallel Yielding Re-Activation", + "bbox": [ + 468, + 114, + 730, + 128 + ], + "page_idx": 6 + }, + { + "type": "page_number", + "text": "7", + "bbox": [ + 774, + 116, + 784, + 126 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "construct the token information matrix:", + "bbox": [ + 215, + 146, + 504, + 160 + ], + "page_idx": 7 + }, + { + "type": "equation", + "text": "\n$$\nM _ {\\text {i n f o}} ^ {l} = \\operatorname {L a y e r N o r m} \\left(M _ {s} ^ {l} + M _ {t} ^ {l}\\right) \\in \\mathbb {R} ^ {D \\times r}. \\tag {4}\n$$\n", + "text_format": "latex", + "bbox": [ + 359, + 172, + 785, + 191 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "We normalize the distribution of $M_{\\mathrm{info}}^{l}$ tokens by leveraging the LayerNorm(\\cdot) operation to enable smoother gradients when training $W_{r}^{l}$ and $W_{D}^{l}$ . With the token information matrix, we then yield the adaptive weights $\\delta_{D}^{l}$ and $\\delta_r^l$ parallelly:", + "bbox": [ + 214, + 200, + 785, + 247 + ], + "page_idx": 7 + }, + { + "type": "equation", + "text": "\n$$\n\\delta_ {D} ^ {l} = M _ {\\text {i n f o}} ^ {l} W _ {r} ^ {l} \\in \\mathbb {R} ^ {D \\times 1} \\tag {5}\n$$\n", + "text_format": "latex", + "bbox": [ + 418, + 257, + 785, + 284 + ], + "page_idx": 7 + }, + { + "type": "equation", + "text": "\n$$\n\\delta_ {r} ^ {l} = W _ {D} ^ {l} M _ {\\mathrm {i n f o}} ^ {l} \\in \\mathbb {R} ^ {1 \\times r}.\n$$\n", + "text_format": "latex", + "bbox": [ + 423, + 277, + 586, + 296 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Re-Activation for Token Modulation. Simply generating $\\delta_D^l$ and $\\delta_r^l$ via matrix multiplication still faces several possible issues. First, no measures have taken to ensure that $\\delta_D^l$ and $\\delta_r^l$ stay in a normal range. Second, decoupling weights to feature tokens and feature channels results in a low-rank modulation weight matrix $W^l$ , which exhibits limited expressive capacity and thus could be unable to optimally modulate tokens in complicated data distributions. To cope with these issues, we conduct token modulation in a re-activation strategy. Specifically, we first broadcast $\\delta_D^l$ to $\\hat{\\delta}_D^l \\in \\mathbb{R}^{D \\times r}$ and conduct sigmoid activation $\\sigma(\\cdot)$ on it, and then modulate $M_s^l$ for an intermediate modulation result:", + "bbox": [ + 214, + 301, + 787, + 441 + ], + "page_idx": 7 + }, + { + "type": "equation", + "text": "\n$$\n\\hat {M} _ {s} ^ {l} = 2 \\sigma \\left(\\hat {\\delta} _ {D} ^ {l}\\right) \\odot M _ {s} ^ {l}. \\tag {6}\n$$\n", + "text_format": "latex", + "bbox": [ + 431, + 450, + 785, + 469 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "where $\\odot$ is Hadamard product. $\\hat{M}_s^l$ is then modulated with sigmoid-activated broadcast weight $\\hat{\\delta}_r^l\\in \\mathbb{R}^{D\\times r}$ again to acquire the modulated tokens:", + "bbox": [ + 214, + 478, + 785, + 512 + ], + "page_idx": 7 + }, + { + "type": "equation", + "text": "\n$$\nM _ {s} ^ {l} \\leftarrow M _ {s} ^ {l} + \\left(2 \\sigma \\left(\\hat {\\delta} _ {r} ^ {l}\\right) - 1\\right) \\odot \\hat {M} _ {s} ^ {l}, \\tag {7}\n$$\n", + "text_format": "latex", + "bbox": [ + 390, + 523, + 785, + 541 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Note that we use the original tokens $M_{s}^{l}$ to create a residual connection that preserves the gradient flow during training. The modulated $M_{s}^{l}$ tokens are then merged with $M_{t}^{l}$ with average pooling. We use a random Gaussian initialization for generator $W_{r}^{l}$ and zero for generator $W_{D}^{l}$ , so re-activation is equivalent to identity transformation at the beginning of training. Token modulation is conducted only on $M_{s}^{l}$ to guarantee the parallelism in training and inference, as $t_{m_k}^l$ tokens are distinct while different $t_{n_k}^l$ might point to the same $t_j^l$ token.", + "bbox": [ + 214, + 551, + 787, + 657 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Discussion. Our PYRA effectively enhances the calibration of features for downstream tasks. Specifically, by utilizing parallel yielding adaptive weights, PYRA effectively decouples the modulation weight from the feature token and the feature channel, i.e., $\\delta_{D}$ and $\\delta_{r}$ , enabling comprehensive perception of the feature distribution in downstream tasks. Furthermore, considering the challenge of capturing an accurate feature distribution for parameter-efficient training, such smooth re-activations by Eq. (6) and Eq. (7) can constrain the negative impact on weight values brought by limited perception of feature distributions, thus leading to better token modulation for improving feature representations in downstream tasks. As a result, the proposed PYRA can maintain discriminative information to an utmost degree during token merging, leading to improved performance while achieving complexity reduction.", + "bbox": [ + 214, + 659, + 787, + 840 + ], + "page_idx": 7 + }, + { + "type": "page_number", + "text": "8", + "bbox": [ + 217, + 114, + 228, + 126 + ], + "page_idx": 7 + }, + { + "type": "header", + "text": "Y. Xiong et al.", + "bbox": [ + 271, + 114, + 372, + 128 + ], + "page_idx": 7 + }, + { + "type": "table", + "img_path": "images/8752dbd7461abdb46e100d4b77830a9112ec29d03940d6723d44b919b97722f9.jpg", + "table_caption": [ + "Table 1: The complexity comparisons between conducting PEFT with and without PYRA. The FLOPs metric is obtained during inference." + ], + "table_footnote": [], + "table_body": "
ModelMetricTotalPEFT w/o PYRA (%)PEFT w. PYRA (%)
ViT-Base# params86M0.29M0.34%0.30M0.35%
FLOPs16.37G16.37G100%8.15G49.79%
ViT-Large# params303M1.18M0.39%1.20M0.40%
FLOPs57.37G57.37G100%28.76G50.13%
", + "bbox": [ + 259, + 184, + 743, + 263 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "3.3 Complexity Analysis", + "text_level": 1, + "bbox": [ + 215, + 289, + 431, + 305 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "We present the parameter and computation complexity for introducing PYRA to PEFT for task adaptation. In each ViT block, PYRA introduces $D + r$ training parameters and $4rD$ extra FLOPs. For a ViT model with $L$ layers and $R$ total merged tokens, in total, our PYRA introduces $LD + R$ extra training parameters and conducts $4RD$ extra FLOPs beside PEFT.", + "bbox": [ + 215, + 311, + 784, + 386 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "To better demonstrate the training-inference efficiency of our PYRA, we compare the complexity between attaching PYRA to task adaptation via PEFT (here we employ LoRA [24]) and plainly conducting PEFT without PYRA. As shown in Tab. 1, PYRA keeps the training efficient feature of PEFT by introducing only a tiny amount of training parameters, which however, results in a substantial efficiency boost for inference, i.e., at around $50\\%$ for both ViT-B and ViT-L models. Results in Sec. 4 show that PYRA achieves comparable performance to the PEFT counterpart without PYRA. This well indicates that PYRA is an effective method for training-inference efficient task adaptation.", + "bbox": [ + 215, + 388, + 784, + 523 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "4 Experiments", + "text_level": 1, + "bbox": [ + 215, + 545, + 374, + 561 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "4.1 Experimental Setting", + "text_level": 1, + "bbox": [ + 215, + 575, + 437, + 590 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "Datasets. We conduct extensive experiments on task adaptation benchmarks to verify the effectiveness of PYRA within the challenge of training-inference efficient task adaptation. Specifically, we choose the VTAB-1k [67] benchmark for the evaluation. VTAB-1k is a challenging benchmark that consists of 19 different tasks from diverse domains: 1) natural images captured in the actual world; 2) specialized images from professional fields; and 3) structured synthesized images. Each task only contains 800 training samples and 200 validation samples.", + "bbox": [ + 215, + 598, + 784, + 703 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "Models. We choose two ViT backbones pre-trained on the ImageNet-21K [16], i.e., ViT-L/16 and ViT-B/16, for comparison. Additionally, we also generalize our PYRA to different backbones and pre-train methods, including DeiT-B [51] and ViT backbones pre-trained by MAE [20].", + "bbox": [ + 215, + 704, + 784, + 763 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "Implementation Details. We choose LoRA [24] as the PEFT module for all methods due to its simplicity and mergeability. For ease of explanation, we omit the \"+\"LoRA\" when comparing different methods. We append LoRA only on the $Q,K$ , and $V$ projection matrices, and apply the training schedule for LoRA following [17, 69]. The generators are trained along with LoRA", + "bbox": [ + 215, + 763, + 784, + 839 + ], + "page_idx": 8 + }, + { + "type": "header", + "text": "PYRA: Parallel Yielding Re-Activation", + "bbox": [ + 468, + 114, + 730, + 128 + ], + "page_idx": 8 + }, + { + "type": "page_number", + "text": "9", + "bbox": [ + 774, + 116, + 784, + 126 + ], + "page_idx": 8 + }, + { + "type": "table", + "img_path": "images/a79ca51c7c5e3144e040fba5e63a45739bfb8799139883ee1bf8117fd79b6aa5.jpg", + "table_caption": [ + "Table 2: Results on the VTAB-1k [67] benchmark under low compression rate. Bold and underline denote the best and second-best accuracy within compression methods." + ], + "table_footnote": [], + "table_body": "
Method# paramsThroughputNaturalSpecializedStructured
Cifar100Caltech101DTDFlowers102PetsSVHNSun397CamelyonEuroSATResisc45RetinopathyClevr-CountClevr-DistDMLABKITTI-DistdSpr-LocdSpr-OrisNORB-AzimsNORB-EleAverage
Model: ViT-B/16 (Throughput: 425)
PEFT0.34%42567.190.269.499.190.585.754.183.195.884.374.682.269.250.179.281.847.131.142.674.76
RaP3.43%65425.968.453.364.057.471.321.575.887.959.373.643.153.826.360.573.525.516.727.955.57
SPViT4.46%56741.675.461.183.266.256.128.379.394.273.373.670.661.542.467.875.450.528.931.364.16
DiffRate0.35%70937.184.663.796.786.232.648.278.985.867.073.732.929.834.155.712.616.013.121.555.82
ToMe0.34%75364.690.467.998.589.883.953.282.694.783.574.981.969.849.276.981.946.531.043.174.10
PYRA0.35%74567.590.369.398.990.084.653.183.395.783.375.282.668.950.880.081.845.832.242.874.69
Model: ViT-L/16 (Throughput: 130)
PEFT0.39%13077.191.473.499.591.389.657.685.996.187.376.183.163.050.782.181.753.532.236.676.52
RaP1.95%19643.287.962.652.881.786.734.778.492.473.373.668.059.646.982.475.543.624.525.765.64
SPViT2.47%18848.187.565.294.477.480.938.879.993.979.874.378.265.847.474.182.350.331.037.970.22
DiffRate0.39%22150.986.870.397.888.339.052.380.287.272.274.232.632.336.557.422.826.615.223.459.53
ToMe0.39%22776.191.172.399.291.789.256.486.495.186.675.182.461.950.981.481.653.533.436.876.11
PYRA0.40%22576.691.373.299.391.589.457.186.995.987.176.283.263.252.883.182.552.634.839.076.84
", + "bbox": [ + 199, + 185, + 802, + 388 + ], + "page_idx": 9 + }, + { + "type": "text", + "text": "modules. During inference, we merge the LoRA module to the backbone. All throughputs are measured during inference on a GeForce RTX 3090 GPU. More details can be found in the supplementary materials.", + "bbox": [ + 212, + 415, + 787, + 463 + ], + "page_idx": 9 + }, + { + "type": "text", + "text": "4.2 Performance comparison on Task Adaptation Benchmark", + "text_level": 1, + "bbox": [ + 212, + 483, + 733, + 500 + ], + "page_idx": 9 + }, + { + "type": "text", + "text": "We verify the effectiveness of PYRA for training-inference efficient task adaptation on (1) low compression rate: comparing the performance of different methods under the same sparsity ratio (here we set sparsity ratio $= 50\\%$ ); (2) high compression rate: comparing the performance between the competing methods and the smaller-scale model with similar throughput levels. We leverage ViT-L/16 and ViT-B/16 for these inspections. The competing methods include RaP [35], SPViT [30], DiffRate [8], and ToMe [2]. Comparisons to more baselines are in the supplementary materials. For RaP and SPViT, we first train the backbone with LoRA on downstream tasks, then we prune the models, and lastly we retrain the parameters attached by the pruning method and LoRA. DiffRate and ToMe can be employed with LoRA during fine-tuning, while DiffRate demands ImageNet-21K for searching the optimal compression schedule.", + "bbox": [ + 212, + 507, + 787, + 688 + ], + "page_idx": 9 + }, + { + "type": "text", + "text": "PYRA on low compression rate. Results on low compression rate are reported in Tab. 2. Overall, while achieving one of the best speedups on throughput, our PYRA is the best performed method compared to other competing methods using the lowest level of training parameters. Counting results on both backbones, our PYRA achieves the best or second-best performance on 37 of 38 dataset metrics. Compared to directly conducting PEFT on the backbone, while all competing methods cause worse results, our PYRA achieves comparable adaptation performance on ViT-B/16, and even outperforms ViT-L/16. The above results convincingly demonstrate that our PYRA successfully sets a new benchmark, i.e. reaching comparable performance as the uncompressed model,", + "bbox": [ + 212, + 689, + 787, + 840 + ], + "page_idx": 9 + }, + { + "type": "page_number", + "text": "10", + "bbox": [ + 217, + 114, + 235, + 126 + ], + "page_idx": 9 + }, + { + "type": "header", + "text": "Y. Xiong et al.", + "bbox": [ + 271, + 114, + 372, + 128 + ], + "page_idx": 9 + }, + { + "type": "table", + "img_path": "images/4b8c035c7014024e5a1ba0496be39bd046ede68e1f2568ba6113d60083c80ba5.jpg", + "table_caption": [ + "Table 3: Results on the VTAB-1k [67] benchmark under high compression rate. Bold and underline denote the best and second-best accuracy within compression methods. \\*: As a comparison of similar throughputs, we compare ViT-B/16 with PEFT on ViTS/16, and ViT-L/16 with PEFT on ViT-B/16." + ], + "table_footnote": [], + "table_body": "
Method# paramsThroughputNaturalSpecializedStructured
Cifar100Caltech101DTDFlowers102PetsSVHNSun397CamelyonEuroSATResisc45RetinopathyClevr-CountClevr-DistDMLABKITTI-DistdSpr-LocdSpr-OrisNORB-AzimsNORB-EleAverage
Model: ViT-B/16 (Throughput: 425)
PEFT*0.34%135057.788.270.198.788.785.744.981.494.784.673.681.664.148.180.072.938.422.937.771.85
RaP0.86%102924.340.134.541.840.521.711.475.886.535.173.849.649.728.139.413.815.412.426.942.60
SPViT4.46%94423.767.951.969.953.219.613.171.981.367.974.753.561.939.557.445.034.511.123.252.49
DiffRate0.35%130823.273.055.787.966.727.229.378.177.853.173.629.728.631.752.611.517.211.320.349.29
ToMe0.34%138154.287.865.596.181.779.745.279.493.676.373.878.365.748.071.380.045.830.941.270.43
PYRA0.35%136554.089.367.196.584.081.844.681.294.679.575.179.967.049.276.982.647.831.942.072.06
Model: ViT-L/16 (Throughput: 130)
PEFT*0.34%42567.190.269.499.190.585.754.183.195.884.374.682.269.250.179.281.847.131.142.674.76
RaP0.65%30117.737.127.046.233.323.213.376.574.254.473.650.431.425.749.853.125.513.426.044.11
SPViT2.47%28954.087.665.594.874.932.638.681.895.378.074.072.861.246.970.277.147.431.328.666.90
DiffRate0.39%41647.473.554.184.360.219.622.250.064.642.818.231.531.931.137.322.017.714.821.440.48
ToMe0.39%43171.090.970.498.388.587.252.482.994.583.175.080.761.148.976.980.853.032.135.274.10
PYRA0.40%42771.691.871.198.589.788.152.285.195.384.675.780.963.051.782.082.054.236.041.275.66
", + "bbox": [ + 196, + 212, + 802, + 412 + ], + "page_idx": 10 + }, + { + "type": "text", + "text": "on low compression rate for training-inference efficient task adaptation while accelerating the model to $1.75 \\times$ speedup with only $0.4\\%$ training parameters.", + "bbox": [ + 212, + 445, + 784, + 474 + ], + "page_idx": 10 + }, + { + "type": "text", + "text": "PYRA on high compression rate. Results are reported in Tab. 3. With comparable throughputs to the smaller-scale model and minimal training parameters, PYRA outperforms all competing methods and achieves the best performance on 35 of 38 dataset metrics. Compared to the smaller-scale model with comparable throughput, PYRA successfully outperforms it. On the compressed ViT-L/16, PYRA even surpasses ViT-B/16 by $0.9\\%$ . This shows that PYRA, as the state-of-the-art, effectively bridges the adverse compression gap between compressed large-scale models and small-scale models, as mentioned in Sec. 1. Therefore, PYRA is an applicable alternative for acquiring a smaller-scale model $(3.2\\times$ speedup) on downstream tasks through efficient training $(0.4\\%)$ training parameters) when no pre-trained small-scale model is available.", + "bbox": [ + 212, + 476, + 785, + 641 + ], + "page_idx": 10 + }, + { + "type": "text", + "text": "PYRA on self-supervised ViT backbone. We conduct the experiments on both compression rates on self-supervised ViT-L/16 (MAE) [20]. As shown in Tab. 4, PYRA significantly surpasses all competing methods. Under low compression rate, PYRA outperforms directly transferring the uncompressed backbone. Under high compression rate, PYRA also eliminates adverse compression. These results show that PYRA generalizes well for self-supervised visual models.", + "bbox": [ + 212, + 642, + 785, + 733 + ], + "page_idx": 10 + }, + { + "type": "text", + "text": "PYRA on different architectures. To further verify the generalizability of PYRA, we conduct the above experiments on the DeiT-B [51]. As shown in Tab. 5, PYRA achieves comparable performance to the uncompressed model under low compression rate and eliminates adverse compression under high compression rate, indicating its generalizability to other transformer architectures.", + "bbox": [ + 212, + 734, + 785, + 808 + ], + "page_idx": 10 + }, + { + "type": "text", + "text": "PYRA consistently yields better models of different throughputs. To further show the superiority of PYRA for different compression rates, we", + "bbox": [ + 212, + 809, + 785, + 839 + ], + "page_idx": 10 + }, + { + "type": "header", + "text": "PYRA: Parallel Yielding Re-Activation", + "bbox": [ + 468, + 114, + 730, + 128 + ], + "page_idx": 10 + }, + { + "type": "page_number", + "text": "11", + "bbox": [ + 767, + 116, + 782, + 126 + ], + "page_idx": 10 + }, + { + "type": "table", + "img_path": "images/19e5322a0012641ce6cf6e7ccc7a1ff9bfc42fb413e5f5dd156d94ffa36e510c.jpg", + "table_caption": [ + "Table 4: VTAB-1k [67] results on both compression rates for ViT-L (MAE) (Throughput: 130)." + ], + "table_footnote": [], + "table_body": "
ModelMethod# paramsThroughputAverage
ViT-L (MAE)PEFT0.39%13075.96
ViT-L (MAE)RaP [35]2.01%18267.55
DiffRate [8]0.39%22146.91
ToMe [2]0.39%22774.97
PYRA0.40%22576.13
ViT-B (MAE)PEFT0.34%42570.23
ViT-L (MAE)RaP [35]0.76%29852.91
DiffRate [8]0.39%41648.03
ToMe [2]0.39%43168.20
PYRA0.40%42770.33
", + "bbox": [ + 215, + 205, + 488, + 397 + ], + "page_idx": 11 + }, + { + "type": "table", + "img_path": "images/2942dbe10df70cc190cb45a65e1b800ae5efc72246ce9d4dbe6325c286d03945.jpg", + "table_caption": [ + "Table 5: VTAB-1k [67] results on both compression rates for DeiT-B (Throughput: 431)." + ], + "table_footnote": [], + "table_body": "
ModelMethod# paramsThroughputAverage
DeiT-BPEFT0.34%43173.76
DeiT-BRaP [35]3.57%69562.34
DiffRate [8]0.35%73452.10
ToMe [2]0.34%74772.83
PYRA0.35%74073.55
DeiT-SPEFT0.34%133270.01
DeiT-BRaP [35]1.09%118757.70
DiffRate [8]0.35%131444.51
ToMe [2]0.34%135168.68
PYRA0.35%134170.13
", + "bbox": [ + 514, + 205, + 787, + 397 + ], + "page_idx": 11 + }, + { + "type": "table", + "img_path": "images/457b41a3dc31fd83f96808fffd81cecc7fafb49350cb3e1665f91e94da05f7cc.jpg", + "table_caption": [ + "Table 6: Ablation study results for PYRA on ViT-B/16 under high compression rate. Here for # params we report only the parameters introduced by token modulation." + ], + "table_footnote": [], + "table_body": "
Method\\( W_r \\)\\( W_D \\)Activation# paramsNaturalSpecializedStructuredAverage
Baseline×××072.8780.7857.6470.43
Plain \\( W_r \\)√××0.19K72.9081.0757.6670.54
\\( σ(\\cdot) \\& W_r \\)√×√0.19K73.1881.6957.6570.84
Plain \\( W_D \\)×√×8.45K73.0981.1358.4370.88
\\( σ(\\cdot) \\& W_D \\)×√√8.45K73.3182.1758.4471.31
Plain \\( W_r \\& W_D \\)√√×8.64K73.7781.3758.8171.32
PYRA√√√8.64K73.9182.6059.6672.06
", + "bbox": [ + 235, + 450, + 769, + 554 + ], + "page_idx": 11 + }, + { + "type": "text", + "text": "compress the chosen backbones for a series of speedups and compare the results with the strongest baseline, ToMe [2], under similar values of throughputs. As shown in Figs. 2 and 4, PYRA consistently outperforms ToMe and flattens the accuracy-throughput curve. Therefore, our PYRA is applicable to acquire task-specific smaller-scale models of different throughputs consistently in the absence of pre-trained parameters for smaller-scale models.", + "bbox": [ + 212, + 582, + 787, + 674 + ], + "page_idx": 11 + }, + { + "type": "text", + "text": "4.3 Ablation Studies", + "text_level": 1, + "bbox": [ + 215, + 694, + 401, + 709 + ], + "page_idx": 11 + }, + { + "type": "text", + "text": "We do controlled experiments to identify the effect of individual components in PYRA, i.e., the generators $W_{r}$ and $W_{D}$ , and the sigmoid activation $\\sigma(\\cdot)$ applied on the generated modulation weights. Specifically, when using only a single generator, we omit Eq. (6) and replace $\\hat{\\delta}_r^l$ in Eq. (7) with the broadcast weight generated by the employed generator. When using the generators without sigmoid activation, we simply remove the $\\sigma$ in Eq. (6) and Eq. (7) while keeping other calculations intact. We choose the ImageNet-21K pre-trained ViT-B/16 to carry out the ablation studies under high compression rates as in Tab. 3.", + "bbox": [ + 212, + 718, + 787, + 840 + ], + "page_idx": 11 + }, + { + "type": "page_number", + "text": "12", + "bbox": [ + 217, + 114, + 235, + 126 + ], + "page_idx": 11 + }, + { + "type": "header", + "text": "Y. Xiong et al.", + "bbox": [ + 271, + 114, + 372, + 128 + ], + "page_idx": 11 + }, + { + "type": "image", + "img_path": "images/cd78e6c8c8c119120dbf7db427ac6705fe8c42780ab35658e2372770f9bfdbdb.jpg", + "image_caption": [ + "Fig. 4: Comparisons between PYRA and ToMe [2] under different compression rates for ViT-Large (MAE) and DeiT-Base. Red boxes: PYRA mitigates the performance drops under low compression rate. Green boxes: PYRA eliminates adverse compression under high compression rate." + ], + "image_footnote": [], + "bbox": [ + 223, + 150, + 480, + 313 + ], + "page_idx": 12 + }, + { + "type": "image", + "img_path": "images/93a8f85f613579036839ee19c25c863ffdd9b7d954343e0ea5711adc5ba0da07.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 514, + 150, + 769, + 313 + ], + "page_idx": 12 + }, + { + "type": "table", + "img_path": "images/d3a8e06b8fbba5a1aaec3229f243c1136f8a0d79489fb99355aec4f5376efabb.jpg", + "table_caption": [ + "Table 7: Adaptation performance comparison to $W_{r \\times D}$ unadaptable to tokens. We report token modulation parameters." + ], + "table_footnote": [], + "table_body": "
Method# paramsNaturalSpecializedStructuredAverage
Baseline072.980.857.670.43
WD×r70.66K73.881.559.271.49
PYRA8.64K73.982.659.772.06
", + "bbox": [ + 220, + 455, + 498, + 571 + ], + "page_idx": 12 + }, + { + "type": "table", + "img_path": "images/1d9b019acd9e93fc9735e447650d46ea0e4e33d8ded3a6a2373c9f019b4ba05b.jpg", + "table_caption": [ + "Table 8: Adaptation performance comparison to the common gated generator. We report token modulation parameters." + ], + "table_footnote": [], + "table_body": "
Method# paramsNaturalSpecializedStructuredAverage
Baseline072.980.857.670.43
Gated73.73K72.881.558.971.06
PYRA8.64K73.982.659.772.06
", + "bbox": [ + 509, + 455, + 787, + 570 + ], + "page_idx": 12 + }, + { + "type": "text", + "text": "Results are shown in Tab. 6. Here the baseline method refers to conducting token merging [2,8] while fine-tuning LoRA [24] for task adaptation. First, both $W_{r}$ and $W_{D}$ , no matter whether sigmoid activation is attached, lead to performance gains, and employing them simultaneously outperforms using them individually. This proves the superiority of our parallel yielding strategy. Second, compared to the corresponding counterpart without activations, employing sigmoid activation always leads to significant improvements, indicating the effectiveness of the reactivation. Overall, our PYRA yields the best adaptation performance, showing that our strategies are effective and complementary.", + "bbox": [ + 212, + 601, + 787, + 739 + ], + "page_idx": 12 + }, + { + "type": "text", + "text": "4.4 Further Analysis on Different Designs", + "text_level": 1, + "bbox": [ + 215, + 763, + 576, + 780 + ], + "page_idx": 12 + }, + { + "type": "text", + "text": "We conduct further analysis on ViT-B/16 model pre-trained on ImageNet-21K. The baseline here refers to training LoRA with token merging [2, 8] attached. More analysis experiments can be found in the supplementary materials.", + "bbox": [ + 212, + 794, + 785, + 840 + ], + "page_idx": 12 + }, + { + "type": "header", + "text": "PYRA: Parallel Yielding Re-Activation", + "bbox": [ + 468, + 114, + 730, + 128 + ], + "page_idx": 12 + }, + { + "type": "page_number", + "text": "13", + "bbox": [ + 767, + 114, + 784, + 126 + ], + "page_idx": 12 + }, + { + "type": "text", + "text": "Impact of Making Modulation Weights Adaptive. In our PYRA, we train modulation weight generators $W_{D}^{l}$ and $W_{r}^{l}$ to conduct adaptive modulation on different merging tokens. To prove the necessity of making modulation weights adaptive, we compare PYRA with the approach of directly training final modulation weights $W_{D\\times r}^{l}$ for each layer, and conduct token modulation as $M_{s}^{l}\\gets M_{s}^{l} + (2\\sigma (W_{D\\times r}^{l}) - 1)\\odot M_{s}^{l}$ . For fair comparisons, we inherit the sigmoid activation and the residual connection. As shown in Tab. 7, although with more training parameters, training $W_{D\\times r}$ is still inferior to our PYRA. This indicates that PYRA is a more effective strategy to conduct adaptive token modulation.", + "bbox": [ + 212, + 146, + 787, + 285 + ], + "page_idx": 13 + }, + { + "type": "text", + "text": "Compare with the Common Gated Generator. We compare PYRA with setting the commonly-applied gated-style trainable module [12, 22, 31, 62] as modulation weight generator. Formally, for each ViT block, we insert a learnable two-layer MLP module $(\\mathrm{MLP}^l(\\cdot))$ with input dimension $D$ , hidden dimension $d \\ll D$ to ensure parameter-efficient training and fast inference, and output dimension $D$ . To generate modulation weights, we feed the information matrix $M_{\\mathrm{info}}^l$ into the MLP, and modulate the merging tokens thereafter: $\\delta^l = \\mathrm{MLP}(M_{\\mathrm{info}}^l) \\in \\mathbb{R}^{D \\times r}$ , $M_s^l \\gets M_s^l + (2\\sigma(\\delta^l - 1) \\odot M_s^l)$ . We set $d = 4$ for all layers. As shown in Tab. 8, both the gated generator and our PYRA achieve performance gains, while our PYRA surpasses the gated generator by $1.0\\%$ with significantly fewer trainable parameters. This indicates that the decoupled weights generated by our $W_D$ and $W_r$ in PYRA are effective and parameter-efficient compared to the common gating strategy.", + "bbox": [ + 212, + 287, + 787, + 484 + ], + "page_idx": 13 + }, + { + "type": "text", + "text": "5 Conclusion", + "text_level": 1, + "bbox": [ + 215, + 527, + 359, + 542 + ], + "page_idx": 13 + }, + { + "type": "text", + "text": "In this work, we defined and investigated a new challenge named training-inference efficient task adaptation, in which the inference efficiency of large-scale transformers is enhanced during parameter-efficient task adaptation. We propose a novel Parallel Yielding Re-Activation method (PYRA) to effectively cope with the challenge by modulating token features during token merging. Specifically, PYRA generates decoupled parallel yielding modulation weights, and conducts token modulation through re-activation. Extensive experiments show that PYRA introduces negligible performance drops under low compression rate, and bridges the gap of adverse compression between compressed transformers and small-scale models under high compression rate. In real-world applications, our PYRA is highly suitable to the scenario of transferring large-scale vision transformers to downstream tasks where no small-scale model is presented.", + "bbox": [ + 212, + 578, + 787, + 760 + ], + "page_idx": 13 + }, + { + "type": "text", + "text": "Limitations. We have not yet validated the effectiveness of PYRA in object detection and image segmentation. We plan to extend these tasks in the future. Meanwhile, it is also applicable to attach pruning methods after training the vision transformers with PYRA to eliminate additional parameters attached by PYRA. We leave that topic to future studies.", + "bbox": [ + 212, + 763, + 787, + 840 + ], + "page_idx": 13 + }, + { + "type": "page_number", + "text": "14", + "bbox": [ + 217, + 114, + 235, + 126 + ], + "page_idx": 13 + }, + { + "type": "header", + "text": "Y. Xiong et al.", + "bbox": [ + 271, + 114, + 370, + 128 + ], + "page_idx": 13 + }, + { + "type": "text", + "text": "Acknowledgements", + "text_level": 1, + "bbox": [ + 217, + 143, + 401, + 162 + ], + "page_idx": 14 + }, + { + "type": "text", + "text": "This work was supported by National Science and Technology Major 2022ZD0119401, National Natural Science Foundation of China (Nos. 62271281, 61925107, 62021002). It is also sponsored by CAAI-CANN Open Fund, developed on OpenI Community.", + "bbox": [ + 212, + 176, + 815, + 239 + ], + "page_idx": 14 + }, + { + "type": "text", + "text": "References", + "text_level": 1, + "bbox": [ + 215, + 263, + 321, + 277 + ], + "page_idx": 14 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "1. Ahn, S., Hu, S.X., Damianou, A., Lawrence, N.D., Dai, Z.: Variational information distillation for knowledge transfer. In: Proceedings of the IEEE/CVF conference on computer vision and pattern recognition. pp. 9163-9171 (2019)", + "2. Bolya, D., Fu, C.Y., Dai, X., Zhang, P., Feichtenhofer, C., Hoffman, J.: Token merging: Your vit but faster. arXiv preprint arXiv:2210.09461 (2022)", + "3. Cai, Y., Yao, Z., Dong, Z., Gholami, A., Mahoney, M.W., Keutzer, K.: Zeroq: A novel zero shot quantization framework. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 13169-13178 (2020)", + "4. Carion, N., Massa, F., Synnaeve, G., Usunier, N., Kirillov, A., Zagoruyko, S.: End-to-end object detection with transformers. In: European conference on computer vision. pp. 213-229. Springer (2020)", + "5. Chavan, A., Shen, Z., Liu, Z., Liu, Z., Cheng, K.T., Xing, E.P.: Vision transformer slimming: Multi-dimension searching in continuous optimization space. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 4931-4941 (2022)", + "6. Chen, C.F.R., Fan, Q., Panda, R.: Crossvit: Cross-attention multi-scale vision transformer for image classification. In: Proceedings of the IEEE/CVF international conference on computer vision. pp. 357-366 (2021)", + "7. Chen, L., Zhao, H., Liu, T., Bai, S., Lin, J., Zhou, C., Chang, B.: An image is worth 1/2 tokens after layer 2: Plug-and-play inference acceleration for large vision-language models. arXiv preprint arXiv:2403.06764 (2024)", + "8. Chen, M., Shao, W., Xu, P., Lin, M., Zhang, K., Chao, F., Ji, R., Qiao, Y., Luo, P.: Diffrate: Differentiable compression rate for efficient vision transformers. arXiv preprint arXiv:2305.17997 (2023)", + "9. Chen, S., Ge, C., Tong, Z., Wang, J., Song, Y., Wang, J., Luo, P.: Adaptformer: Adapting vision transformers for scalable visual recognition. Advances in Neural Information Processing Systems 35, 16664-16678 (2022)", + "0. Chen, T., Cheng, Y., Gan, Z., Yuan, L., Zhang, L., Wang, Z.: Chasing sparsity in vision transformers: An end-to-end exploration. Advances in Neural Information Processing Systems 34, 19974-19988 (2021)", + "11. Cheng, B., Misra, I., Schwing, A.G., Kirillov, A., Girdhar, R.: Masked-attention mask transformer for universal image segmentation. In: Proceedings of the IEEE/CVF conference on computer vision and pattern recognition. pp. 1290-1299 (2022)", + "12. Chung, J., Gulcehre, C., Cho, K., Bengio, Y.: Empirical evaluation of gated recurrent neural networks on sequence modeling. arXiv preprint arXiv:1412.3555 (2014)", + "13. Courbariaux, M., Bengio, Y., David, J.P.: Binaryconnect: Training deep neural networks with binary weights during propagations. Advances in neural information processing systems 28 (2015)" + ], + "bbox": [ + 218, + 295, + 785, + 839 + ], + "page_idx": 14 + }, + { + "type": "header", + "text": "PYRA: Parallel Yielding Re-Activation", + "bbox": [ + 468, + 114, + 730, + 128 + ], + "page_idx": 14 + }, + { + "type": "page_number", + "text": "15", + "bbox": [ + 767, + 114, + 785, + 126 + ], + "page_idx": 14 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "14. Ding, N., Qin, Y., Yang, G., Wei, F., Yang, Z., Su, Y., Hu, S., Chen, Y., Chan, C.M., Chen, W., et al.: Delta tuning: A comprehensive study of parameter efficient methods for pre-trained language models. arXiv preprint arXiv:2203.06904 (2022)", + "15. Ding, Z., Wang, A., Chen, H., Zhang, Q., Liu, P., Bao, Y., Yan, W., Han, J.: Exploring structured semantic prior for multi label recognition with incomplete labels. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 3398-3407 (2023)", + "16. Dosovitskiy, A., Beyer, L., Kolesnikov, A., Weissenborn, D., Zhai, X., Unterthiner, T., Dehghani, M., Minderer, M., Heigold, G., Gelly, S., et al.: An image is worth 16x16 words: Transformers for image recognition at scale. arXiv preprint arXiv:2010.11929 (2020)", + "17. Hao, T., Chen, H., Guo, Y., Ding, G.: Consolidator: Mergeable adapter with grouped connections for visual adaptation. arXiv preprint arXiv:2305.00603 (2023)", + "18. Hao, T., Ding, X., Feng, J., Yang, Y., Chen, H., Ding, G.: Quantized prompt for efficient generalization of vision-language models. arXiv preprint arXiv:2407.10704 (2024)", + "19. Hao, T., Lyu, M., Chen, H., Zhao, S., Han, J., Ding, G.: Re-parameterized low-rank prompt: Generalize a vision-language model within $0.5\\mathrm{k}$ parameters. arXiv preprint arXiv:2312.10813 (2023)", + "20. He, K., Chen, X., Xie, S., Li, Y., Dóllar, P., Girshick, R.: Masked autoencoders are scalable vision learners. In: Proceedings of the IEEE/CVF conference on computer vision and pattern recognition. pp. 16000-16009 (2022)", + "21. Hinton, G., Vinyals, O., Dean, J.: Distilling the knowledge in a neural network. arXiv preprint arXiv:1503.02531 (2015)", + "22. Hochreiter, S., Schmidhuber, J.: Long short-term memory. Neural Computation 9(8), 1735-1780 (1997)", + "23. Houlsby, N., Giurgiu, A., Jastrzebski, S., Morrone, B., De Laroussilhe, Q., Gesmundo, A., Attariyan, M., Gelly, S.: Parameter-efficient transfer learning for nlp. In: International Conference on Machine Learning. pp. 2790-2799. PMLR (2019)", + "24. Hu, E.J., Shen, Y., Wallis, P., Allen-Zhu, Z., Li, Y., Wang, S., Wang, L., Chen, W.: Lora: Low-rank adaptation of large language models. arXiv preprint arXiv:2106.09685 (2021)", + "25. Huang, X., Shen, Z., Cheng, K.T.: Variation-aware vision transformer quantization. arXiv preprint arXiv:2307.00331 (2023)", + "26. Jia, M., Tang, L., Chen, B.C., Cardie, C., Belongie, S., Hariharan, B., Lim, S.N.: Visual prompt tuning. In: European Conference on Computer Vision. pp. 709-727. Springer (2022)", + "27. Jiang, H., Wu, Q., Lin, C.Y., Yang, Y., Qiu, L.: Llmlingua: Compressing prompts for accelerated inference of large language models. arXiv preprint arXiv:2310.05736 (2023)", + "28. Jie, S., Deng, Z.H.: Convolutional bypasses are better vision transformer adapters. arXiv preprint arXiv:2207.07039 (2022)", + "29. Kirillov, A., Mintun, E., Ravi, N., Mao, H., Rolland, C., Gustafson, L., Xiao, T., Whitehead, S., Berg, A.C., Lo, W.Y., et al.: Segment anything. arXiv preprint arXiv:2304.02643 (2023)", + "30. Kong, Z., Dong, P., Ma, X., Meng, X., Niu, W., Sun, M., Shen, X., Yuan, G., Ren, B., Tang, H., Qin, M., Wang, Y.: Spvit: Enabling faster vision transformers via latency-aware soft token pruning. In: Avidan, S., Brostow, G.J., Cissé, M., Farinella, G.M., Hassner, T. (eds.) Computer Vision - ECCV 2022 - 17th European Conference, Tel Aviv, Israel, October 23-27, 2022, Proceedings, Part XI. Lecture" + ], + "bbox": [ + 217, + 146, + 785, + 840 + ], + "page_idx": 15 + }, + { + "type": "page_number", + "text": "16", + "bbox": [ + 217, + 114, + 235, + 126 + ], + "page_idx": 15 + }, + { + "type": "header", + "text": "Y. Xiong et al.", + "bbox": [ + 271, + 114, + 370, + 127 + ], + "page_idx": 15 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "Notes in Computer Science, vol. 13671, pp. 620-640. Springer (2022). https://doi.org/10.1007/978-3-031-20083-0_37, https://doi.org/10.1007/978-3-031-20083-0_37", + "31. Kong, Z., Dong, P., Ma, X., Meng, X., Sun, M., Niu, W., Shen, X., Yuan, G., Ren, B., Qin, M., et al.: Spvit: Enabling faster vision transformers via soft token pruning. arXiv preprint arXiv:2112.13890 (2021)", + "32. Lester, B., Al-Rfou, R., Constant, N.: The power of scale for parameter-efficient prompt tuning. arXiv preprint arXiv:2104.08691 (2021)", + "33. Li, X.L., Liang, P.: Prefix-tuning: Optimizing continuous prompts for generation. arXiv preprint arXiv:2101.00190 (2021)", + "34. Li, Y., Xu, S., Zhang, B., Cao, X., Gao, P., Guo, G.: Q-vit: Accurate and fully quantized low-bit vision transformer. Advances in Neural Information Processing Systems 35, 34451-34463 (2022)", + "35. Li, Y., Adamczewski, K., Li, W., Gu, S., Timofte, R., Gool, L.V.: Revisiting random channel pruning for neural network compression. In: IEEE/CVF Conference on Computer Vision and Pattern Recognition, CVPR 2022, New Orleans, LA, USA, June 18-24, 2022. pp. 191-201. IEEE (2022). https://doi.org/10.1109/CVPR52688.2022.00029, https://doi.org/10.1109/CVPR52688.2022.00029", + "36. Lian, D., Zhou, D., Feng, J., Wang, X.: Scaling & shifting your features: A new baseline for efficient model tuning. Advances in Neural Information Processing Systems 35, 109-123 (2022)", + "37. Lin, X., Zhao, C., Pan, W.: Towards accurate binary convolutional neural network. Advances in neural information processing systems 30 (2017)", + "38. Lin, Z., Courbariaux, M., Memisevic, R., Bengio, Y.: Neural networks with few multiplications. arXiv preprint arXiv:1510.03009 (2015)", + "39. Liu, X., Ji, K., Fu, Y., Tam, W.L., Du, Z., Yang, Z., Tang, J.: P-tuning v2: Prompt tuning can be comparable to fine-tuning universally across scales and tasks. arXiv preprint arXiv:2110.07602 (2021)", + "40. Liu, Z., Lin, Y., Cao, Y., Hu, H., Wei, Y., Zhang, Z., Lin, S., Guo, B.: Swin transformer: Hierarchical vision transformer using shifted windows. In: Proceedings of the IEEE/CVF international conference on computer vision. pp. 10012-10022 (2021)", + "41. Liu, Z., Wang, Y., Han, K., Zhang, W., Ma, S., Gao, W.: Post-training quantization for vision transformer. Advances in Neural Information Processing Systems 34, 28092-28103 (2021)", + "42. Lyu, M., Yang, Y., Hong, H., Chen, H., Jin, X., He, Y., Xue, H., Han, J., Ding, G.: One-dimensional adapter to rule them all: Concepts, diffusion models and erasing applications. arXiv preprint arXiv:2312.16145 (2023)", + "43. Marin, D., Chang, J.H.R., Ranjan, A., Prabhu, A., Rastegari, M., Tuzel, O.: Token pooling in vision transformers. arXiv preprint arXiv:2110.03860 (2021)", + "44. Perez, E., Strub, F., De Vries, H., Dumoulin, V., Courville, A.: Film: Visual reasoning with a general conditioning layer. In: Proceedings of the AAAI conference on artificial intelligence (2018)", + "45. Pfeiffer, J., Kamath, A., Rücklé, A., Cho, K., Gurevych, I.: Adapterfusion: Nondestructive task composition for transfer learning. arXiv preprint arXiv:2005.00247 (2020)", + "46. Song, Y., Zhou, Q., Li, X., Fan, D.P., Lu, X., Ma, L.: Ba-sam: Scalable bias-mode attention mask for segment anything model. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 3162-3173 (2024)" + ], + "bbox": [ + 215, + 146, + 785, + 840 + ], + "page_idx": 16 + }, + { + "type": "header", + "text": "PYRA: Parallel Yielding Re-Activation", + "bbox": [ + 468, + 114, + 730, + 128 + ], + "page_idx": 16 + }, + { + "type": "page_number", + "text": "17", + "bbox": [ + 767, + 114, + 785, + 126 + ], + "page_idx": 16 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "47. Song, Z., Jia, C., Yang, L., Wei, H., Liu, L.: Graphalign++: An accurate feature alignment by graph matching for multi-modal 3d object detection. IEEE Transactions on Circuits and Systems for Video Technology (2023)", + "48. Song, Z., Wei, H., Bai, L., Yang, L., Jia, C.: Graphalign: Enhancing accurate feature alignment by graph matching for multi-modal 3d object detection. In: Proceedings of the IEEE/CVF International Conference on Computer Vision. pp. 3358-3369 (2023)", + "49. Strudel, R., Garcia, R., Laptev, I., Schmid, C.: Segmenter: Transformer for semantic segmentation. In: Proceedings of the IEEE/CVF international conference on computer vision. pp. 7262-7272 (2021)", + "50. Tian, L., Ye, M., Zhou, L., He, Q.: Clip-guided black-box domain adaptation of image classification. Signal, Image and Video Processing 18(5), 4637-4646 (2024)", + "51. Touvron, H., Cord, M., Douze, M., Massa, F., Sablayrolles, A., Jégou, H.: Training data-efficient image transformers & distillation through attention. In: International conference on machine learning. pp. 10347-10357. PMLR (2021)", + "52. Tung, F., Mori, G.: Similarity-preserving knowledge distillation. In: Proceedings of the IEEE/CVF international conference on computer vision. pp. 1365-1374 (2019)", + "53. Wang, A., Chen, H., Lin, Z., Pu, H., Ding, G.: Repvit: Revisiting mobile cnn from vit perspective. arXiv preprint arXiv:2307.09283 (2023)", + "54. Wang, A., Chen, H., Lin, Z., Zhao, S., Han, J., Ding, G.: Cait: Triple-win compression towards high accuracy, fast inference, and favorable transferability for vits. arXiv preprint arXiv:2309.15755 (2023)", + "55. Wang, W., Xie, E., Li, X., Fan, D.P., Song, K., Liang, D., Lu, T., Luo, P., Shao, L.: Pyramid vision transformer: A versatile backbone for dense prediction without convolutions. In: Proceedings of the IEEE/CVF international conference on computer vision. pp. 568-578 (2021)", + "56. Wang, Z., Luo, H., Wang, P., Ding, F., Wang, F., Li, H.: Vtc-lfc: Vision transformer compression with low-frequency components. Advances in Neural Information Processing Systems 35, 13974-13988 (2022)", + "57. Wei, K., Du, R., Jin, L., Liu, J., Yin, J., Zhang, L., Liu, J., Liu, N., Zhang, J., Guo, Z.: Video event extraction with multi-view interaction knowledge distillation. In: Proceedings of the AAAI Conference on Artificial Intelligence. vol. 38, pp. 19224-19233 (2024)", + "58. Wei, K., Jin, L., Zhang, Z., Guo, Z., Li, X., Liu, Q., Feng, W.: More than syntaxes: Investigating semantics to zero-shot cross-lingual relation extraction and event argument role labelling. ACM Transactions on Asian and Low-Resource Language Information Processing 23(5), 1-21 (2024)", + "59. Wu, Y., He, K.: Group normalization. In: Proceedings of the European conference on computer vision (ECCV). pp. 3-19 (2018)", + "60. Xiong, Y., Chen, H., Lin, Z., Zhao, S., Ding, G.: Confidence-based visual dispersal for few-shot unsupervised domain adaptation. In: Proceedings of the IEEE/CVF International Conference on Computer Vision. pp. 11621-11631 (2023)", + "61. Xiong, Y., Chen, X., Ye, X., Chen, H., Lin, Z., Lian, H., Niu, J., Ding, G.: Temporal scaling law for large language models. arXiv preprint arXiv:2404.17785 (2024)", + "62. Xu, L., Xie, H., Qin, S.Z.J., Tao, X., Wang, F.L.: Parameter-efficient fine-tuning methods for pretrained language models: A critical review and assessment. arXiv preprint arXiv:2312.12148 (2023)", + "63. Yao, Z., Yazdani Aminabadi, R., Zhang, M., Wu, X., Li, C., He, Y.: Zeroquant: Efficient and affordable post-training quantization for large-scale transformers. Advances in Neural Information Processing Systems 35, 27168-27183 (2022)" + ], + "bbox": [ + 215, + 147, + 785, + 839 + ], + "page_idx": 17 + }, + { + "type": "page_number", + "text": "18", + "bbox": [ + 217, + 114, + 235, + 126 + ], + "page_idx": 17 + }, + { + "type": "header", + "text": "Y. Xiong et al.", + "bbox": [ + 271, + 114, + 370, + 127 + ], + "page_idx": 17 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "64. Yu, S., Chen, T., Shen, J., Yuan, H., Tan, J., Yang, S., Liu, J., Wang, Z.: Unified visual transformer compression. arXiv preprint arXiv:2203.08243 (2022)", + "65. Zaken, E.B., Ravfogel, S., Goldberg, Y.: Bitfit: Simple parameter-efficient fine-tuning for transformer-based masked language-models. arXiv preprint arXiv:2106.10199 (2021)", + "66. Zhai, X., Kolesnikov, A., Houlsby, N., Beyer, L.: Scaling vision transformers. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 12104-12113 (2022)", + "67. Zhai, X., Puigcerver, J., Kolesnikov, A., Ruyssen, P., Riquelme, C., Lucic, M., Djolonga, J., Pinto, A.S., Neumann, M., Dosovitskiy, A., et al.: A large-scale study of representation learning with the visual task adaptation benchmark. arXiv preprint arXiv:1910.04867 (2019)", + "68. Zhang, Y., Qiu, L., Zhu, Y., Wen, L., Luo, X.: A new childhood pneumonia diagnosis method based on fine-grained convolutional neural network. Computer Modeling in Engineering & Sciences 133, 873-894 (2022)", + "69. Zhang, Y., Zhou, K., Liu, Z.: Neural prompt search. arXiv preprint arXiv:2206.04673 (2022)", + "70. Zheng, S., Lu, J., Zhao, H., Zhu, X., Luo, Z., Wang, Y., Fu, Y., Feng, J., Xiang, T., Torr, P.H., et al.: Rethinking semantic segmentation from a sequence-to-sequence perspective with transformers. In: Proceedings of the IEEE/CVF conference on computer vision and pattern recognition. pp. 6881-6890 (2021)", + "71. Zhou, F., Zhou, Q., Yin, B., Zheng, H., Lu, X., Ma, L., Ling, H.: Rethinking impersonation and dodging attacks on face recognition systems. In: Proceedings of the 30th ACM International Conference on Multimedia (ACM MM) (2024)", + "72. Zhou, Q., Feng, Z., Gu, Q., Pang, J., Cheng, G., Lu, X., Shi, J., Ma, L.: Context-aware mixup for domain adaptive semantic segmentation. IEEE Transactions on Circuits and Systems for Video Technology 33(2), 804-817 (2022)", + "73. Zhu, M., Tang, Y., Han, K.: Vision transformer pruning. arXiv preprint arXiv:2104.08500 (2021)", + "74. Zhu, X., Su, W., Lu, L., Li, B., Wang, X., Dai, J.: Deformable detr: Deformable transformers for end-to-end object detection. arXiv preprint arXiv:2010.04159 (2020)" + ], + "bbox": [ + 212, + 146, + 787, + 590 + ], + "page_idx": 18 + }, + { + "type": "header", + "text": "PYRA: Parallel Yielding Re-Activation", + "bbox": [ + 468, + 114, + 730, + 128 + ], + "page_idx": 18 + }, + { + "type": "page_number", + "text": "19", + "bbox": [ + 767, + 114, + 785, + 126 + ], + "page_idx": 18 + } +] \ No newline at end of file diff --git a/2024/PYRA_ Parallel Yielding Re-Activation for Training-Inference Efficient Task Adaptation/a96361d9-106b-4325-b9d1-39f6bd731ebb_model.json b/2024/PYRA_ Parallel Yielding Re-Activation for Training-Inference Efficient Task Adaptation/a96361d9-106b-4325-b9d1-39f6bd731ebb_model.json new file mode 100644 index 0000000000000000000000000000000000000000..4d2d338d415d582700f2389eda633bb99ed31bc8 --- /dev/null +++ b/2024/PYRA_ Parallel Yielding Re-Activation for Training-Inference Efficient Task Adaptation/a96361d9-106b-4325-b9d1-39f6bd731ebb_model.json @@ -0,0 +1,2713 @@ +[ + [ + { + "type": "title", + "bbox": [ + 0.248, + 0.141, + 0.756, + 0.187 + ], + "angle": 0, + "content": "PYRA: Parallel Yielding Re-Activation for Training-Inference Efficient Task Adaptation" + }, + { + "type": "text", + "bbox": [ + 0.223, + 0.212, + 0.782, + 0.258 + ], + "angle": 0, + "content": "Yizhe Xiong\\(^{1,2}\\), Hui Chen\\(^{2*}\\), Tianxiang Hao\\(^{1,2}\\), Zijia Lin\\(^{1}\\), Jungong Han\\(^{2,3}\\), Yuesong Zhang\\(^{4}\\), Guoxin Wang\\(^{4}\\), Yongjun Bao\\(^{4}\\), and Guiguang Ding\\(^{1,2}\\)" + }, + { + "type": "text", + "bbox": [ + 0.308, + 0.269, + 0.694, + 0.285 + ], + "angle": 0, + "content": "\\(^{1}\\) School of Software, Tsinghua University, Beijing, China" + }, + { + "type": "text", + "bbox": [ + 0.348, + 0.284, + 0.657, + 0.298 + ], + "angle": 0, + "content": "\\(^{2}\\) BNRist, Tsinghua University, Beijing, China" + }, + { + "type": "text", + "bbox": [ + 0.281, + 0.297, + 0.722, + 0.312 + ], + "angle": 0, + "content": "\\(^{3}\\) Department of Automation, Tsinghua University, Beijing, China" + }, + { + "type": "text", + "bbox": [ + 0.417, + 0.312, + 0.588, + 0.326 + ], + "angle": 0, + "content": "4 JD.com, Beijing, China" + }, + { + "type": "text", + "bbox": [ + 0.312, + 0.327, + 0.692, + 0.341 + ], + "angle": 0, + "content": "{xiongyizhe2001,beyondtx,jungonghan77}@gmail.com" + }, + { + "type": "text", + "bbox": [ + 0.392, + 0.341, + 0.61, + 0.354 + ], + "angle": 0, + "content": "huichen@mail.tsinghua.edu.cn" + }, + { + "type": "text", + "bbox": [ + 0.325, + 0.355, + 0.68, + 0.368 + ], + "angle": 0, + "content": "{zhangyuesong1,wangguoxin14,baoyongjun}@jd.com" + }, + { + "type": "text", + "bbox": [ + 0.308, + 0.369, + 0.696, + 0.382 + ], + "angle": 0, + "content": "linzijia07@tsinghua.org.cn dinggg@tsinghua.edu.cn" + }, + { + "type": "text", + "bbox": [ + 0.261, + 0.414, + 0.744, + 0.692 + ], + "angle": 0, + "content": "Abstract. Recently, the scale of transformers has grown rapidly, which introduces considerable challenges in terms of training overhead and inference efficiency in the scope of task adaptation. Existing works, namely Parameter-Efficient Fine-Tuning (PEFT) and model compression, have separately investigated the challenges. However, PEFT cannot guarantee the inference efficiency of the original backbone, especially for large-scale models. Model compression requires significant training costs for structure searching and re-training. Consequently, a simple combination of them cannot guarantee accomplishing both training efficiency and inference efficiency with minimal costs. In this paper, we propose a novel Parallel Yielding Re-Activation (PYRA) method for such a challenge of training-inference efficient task adaptation. PYRA first utilizes parallel yielding adaptive weights to comprehensively perceive the data distribution in downstream tasks. A re-activation strategy for token modulation is then applied for tokens to be merged, leading to calibrated token features. Extensive experiments demonstrate that PYRA outperforms all competing methods under both low compression rate and high compression rate, demonstrating its effectiveness and superiority in maintaining both training efficiency and inference efficiency for large-scale foundation models. Our code is available at https://github.com/THU-MIG/PyRA." + }, + { + "type": "text", + "bbox": [ + 0.262, + 0.704, + 0.741, + 0.719 + ], + "angle": 0, + "content": "Keywords: Vision Transformer \\(\\cdot\\) Task Adaptation \\(\\cdot\\) Model Compression" + }, + { + "type": "title", + "bbox": [ + 0.216, + 0.742, + 0.376, + 0.758 + ], + "angle": 0, + "content": "1 Introduction" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.772, + 0.788, + 0.818 + ], + "angle": 0, + "content": "Vision transformers [16] have made a profound impact across various domains of computer vision, such as image classification [6,15,16,50,55,66], object detection [4,40,47,48,74], image segmentation [11,49,53,70,72], etc. In recent years, the" + }, + { + "type": "page_footnote", + "bbox": [ + 0.218, + 0.825, + 0.386, + 0.84 + ], + "angle": 0, + "content": "* Corresponding Author" + } + ], + [ + { + "type": "page_number", + "bbox": [ + 0.218, + 0.116, + 0.23, + 0.127 + ], + "angle": 0, + "content": "2" + }, + { + "type": "header", + "bbox": [ + 0.272, + 0.115, + 0.373, + 0.129 + ], + "angle": 0, + "content": "Y. Xiong et al." + }, + { + "type": "image", + "bbox": [ + 0.223, + 0.147, + 0.382, + 0.209 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.248, + 0.211, + 0.36, + 0.223 + ], + "angle": 0, + "content": "(a) Full Fine-tuning" + }, + { + "type": "image", + "bbox": [ + 0.223, + 0.224, + 0.382, + 0.284 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.277, + 0.286, + 0.33, + 0.297 + ], + "angle": 0, + "content": "(c) PEFT" + }, + { + "type": "image", + "bbox": [ + 0.401, + 0.149, + 0.561, + 0.21 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.419, + 0.211, + 0.548, + 0.223 + ], + "angle": 0, + "content": "(b) Model compression" + }, + { + "type": "image", + "bbox": [ + 0.401, + 0.224, + 0.56, + 0.284 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.459, + 0.286, + 0.509, + 0.297 + ], + "angle": 0, + "content": "(d) Ours" + }, + { + "type": "table", + "bbox": [ + 0.574, + 0.147, + 0.783, + 0.286 + ], + "angle": 0, + "content": "
Method(a)(b)(c)(d)
Non-vast Data Scale✔×✔✔
Training Efficient××✔✔
Inference Efficient×✔×✔
" + }, + { + "type": "image_caption", + "bbox": [ + 0.214, + 0.309, + 0.788, + 0.42 + ], + "angle": 0, + "content": "Fig. 1: (a) Full fine-tuning trains all parameters on downstream tasks and utilizes the trained model for inference, thereby lacking efficiency in both training and inference stages. (b) Model compression employs pruning to enhance inference efficiency, but the pruned model necessitates extensive re-training on large-scale data. (c) PEFT freezes the model backbone and only fine-tunes a small amount of parameters, yet retains the inference complexity. (d) Our training-inference efficient task adaptation incorporates the advantages of all existing pipelines by training inference-efficient models with minimal tunable parameters." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.457, + 0.788, + 0.563 + ], + "angle": 0, + "content": "scale of vision transformers has grown to be billion-parameters [29,61,66]. Consequently, adapting such models with vast scales into downstream tasks presents increasingly complex challenges, particularly in real-world deployment scenarios. Two critical concerns have been widely acknowledged as primary obstacles in implementing large-scale transformers for downstream applications [17,23,27,46]: (1) the training overhead when fine-tuning on downstream tasks, and (2) the inference efficiency after model deployment." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.566, + 0.788, + 0.748 + ], + "angle": 0, + "content": "Specifically, first, conventional fine-tuning methods, which necessitate adjusting all parameters of the model (i.e., Full Fine-tuning), suffer from unaffordable consumption of GPU resources and training time given the extensive scales of the foundation models [16, 29, 66]. Researchers have delved into Parameter-Efficient Fine-Tuning (PEFT) [17, 19, 23, 24, 26, 36] algorithms, which generally freeze the pre-trained models and only tune extra small parameters, leading to great reduction of training time and storage overhead. The second issue pertains to inference efficiency, requiring the deployed model to promptly process the input data. The computational complexity of the models significantly influences achieving satisfactory inference throughput. Representative solutions for this matter encompass model compression methods, including model pruning [7, 10, 56, 64], knowledge distillation [1, 21, 52, 57, 58], model quantization [3, 13, 37], etc." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.75, + 0.788, + 0.841 + ], + "angle": 0, + "content": "In the literature, these two intriguing topics are investigated separately. PEFT methods either identify a subset of tunable parameters in the backbone for fine-tuning [65] or introduce learnable parameters to the frozen backbone during fine-tuning [17, 23, 24, 26, 36]. While effectively reducing training costs, most of these methods inevitably escalate computational complexity, resulting in inefficient inference. Model compression methods frequently require signifi" + } + ], + [ + { + "type": "header", + "bbox": [ + 0.469, + 0.115, + 0.732, + 0.129 + ], + "angle": 0, + "content": "PYRA: Parallel Yielding Re-Activation" + }, + { + "type": "page_number", + "bbox": [ + 0.775, + 0.117, + 0.785, + 0.127 + ], + "angle": 0, + "content": "3" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.147, + 0.788, + 0.208 + ], + "angle": 0, + "content": "cant computational resources to identify optimal structures for pruning. After pruning, a comprehensive re-training process using a substantial amount of data is crucial to prevent significant performance degradation. Therefore, model compression methods are typically inefficient in training efficiency." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.209, + 0.434, + 0.539 + ], + "angle": 0, + "content": "These observations naturally lead to the question: can we achieve both training efficiency and inference efficiency simultaneously for downstream tasks? We refer to this challenge as Training-Inference Efficient Task Adaptation (seeing Fig. 1). Exploring this issue can enable us to conveniently deploy the advanced large-scale foundation models in real-world downstream applications with minimal costs, which is appealing and essential for the widespread implementation of foundation models. A straightforward solution is combining PEFT and model compression. However, for efficient task adaptation, a" + }, + { + "type": "image", + "bbox": [ + 0.443, + 0.239, + 0.785, + 0.404 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.44, + 0.415, + 0.788, + 0.499 + ], + "angle": 0, + "content": "Fig.2: Comparisons between simply combining ToMe [2] and LoRA [24] and our proposed PYRA. Red boxes represent the performance drop problem in low compression rates. Green boxes represent the adverse compression in high compression rates. See Sec. 4 for more results." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.54, + 0.788, + 0.751 + ], + "angle": 0, + "content": "heavy re-training stage is unaffordable. Consequently, simply combining PEFT and model compression can easily suffer from substantial performance drops. For instance, we could integrate LoRA [24], a notable PEFT method that introduces a small low-rank adapter, with ToMe [2], a parameter-free model compression technique for vision transformers5. As shown in Fig. 2, under lower compression rates (around \\(1.7 \\times\\)), performance on both backbones (ViT-L/16 and ViT-B/16) present slight drops (\\(< 1\\%\\)) compared to directly conducting PEFT on the backbone, indicating that the ToMe+LoRA combination serves as a basic solution within lower compression rate range, yet performance improvements are also demanded. Under high compression rates (\\(>3.0 \\times\\)), the performance quickly drops and is even inferior to directly fine-tuning the small-scale backbone with corresponding throughput. We term this phenomenon as Adverse Compression. Both phenomena indicate that directly combining existing works cannot effectively address the new challenges." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.753, + 0.808, + 0.814 + ], + "angle": 0, + "content": "In this paper, we propose a novel Parallel Yielding Re-Accivitation method (PYRA) designed for training-inference efficient task adaptation using vision transformers. Generally, the proposed PYRA follows the token merging paradigm [2, 8, 43] for inference efficiency. For effective task adaptation, we propose to mod-" + }, + { + "type": "page_footnote", + "bbox": [ + 0.218, + 0.825, + 0.681, + 0.84 + ], + "angle": 0, + "content": "In Sec. 4, we show that \\(\\mathrm{ToMe + LoRA}\\) is a neat and strong solution." + } + ], + [ + { + "type": "page_number", + "bbox": [ + 0.218, + 0.116, + 0.23, + 0.127 + ], + "angle": 0, + "content": "4" + }, + { + "type": "header", + "bbox": [ + 0.272, + 0.115, + 0.372, + 0.129 + ], + "angle": 0, + "content": "Y. Xiong et al." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.147, + 0.788, + 0.268 + ], + "angle": 0, + "content": "ulate token features from both feature tokens and feature channels and parallelly yield weights for tokens to be merged by small modulation weight generators. These parallel yielding weights can comprehensively perceive the data distribution in downstream tasks. They are then applied to features through re-activation, resulting in adaptive token modulation. Thanks to such a token modulation strategy, PYRA can adaptively calibrate the learned feature distribution for downstream tasks with low computational complexity, ultimately leading to effective training and efficient inference." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.268, + 0.788, + 0.433 + ], + "angle": 0, + "content": "We conduct extensive experiments to verify the effectiveness of our PYRA. We show that, under a low compression rate with \\(\\sim 1.7\\times\\) speedup, PYRA introduces negligible performance drops. Under a high compression rate with \\(>3.0\\times\\) speedup, PYRA eliminates the adverse compression gap. We emphasize that considering the scenario of high compression rates is practical, given the substantial size of current transformers. Concurrently, the specific model sizes required by downstream applications may not match any publicly available models, thus requiring the acquisition through high compression rates while keeping the performance comparable. In this regard, our approach represents an effective method for obtaining small models in the absence of pre-trained parameters for smaller-scale models." + }, + { + "type": "text", + "bbox": [ + 0.24, + 0.434, + 0.605, + 0.449 + ], + "angle": 0, + "content": "Overall, we summarize our contribution as follows." + }, + { + "type": "text", + "bbox": [ + 0.226, + 0.457, + 0.785, + 0.5 + ], + "angle": 0, + "content": "- We propose a novel challenge termed training-inference efficient task adaptation, in which the inference efficiency of large-scale transformers is escalated during parameter-efficient task adaptation." + }, + { + "type": "text", + "bbox": [ + 0.226, + 0.502, + 0.785, + 0.574 + ], + "angle": 0, + "content": "- We propose PYRA for training-inference efficient task adaptation, which enhances the perception of feature distribution via token modulation. We generate parallel yielding decoupled weights to comprehensively perceive the feature distributions. We apply a re-activation strategy to modulate tokens to be merged for calibrated token features." + }, + { + "type": "text", + "bbox": [ + 0.226, + 0.577, + 0.785, + 0.65 + ], + "angle": 0, + "content": "- Extensive experiments show that our PYRA outperforms all competing methods under both low compression rate and high compression rate. Further analysis shows that PYRA is effective across a series of different transformer backbones and model scales, well demonstrating the effectiveness and superiority of our method." + }, + { + "type": "list", + "bbox": [ + 0.226, + 0.457, + 0.785, + 0.65 + ], + "angle": 0, + "content": null + }, + { + "type": "title", + "bbox": [ + 0.216, + 0.674, + 0.398, + 0.691 + ], + "angle": 0, + "content": "2 Related Works" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.704, + 0.788, + 0.841 + ], + "angle": 0, + "content": "Parameter-Efficient Fine-Tuning (PEFT) for Task Adaptation. Transferring large-scale transformers to downstream tasks has been a popular topic in computer vision [9, 17, 26, 28, 36, 60, 68, 69]. PEFT methods either locate a subset of parameters inside the model for fine-tuning [65], or inject human designed modules to the original model structure. Specifically, adapters [23, 28, 42, 45] are a type of MLP module with a bottleneck in the middle. Prompt-tuning [14, 18, 26, 32, 33, 39] insert learnable tokens to model input to generate task-specific outputs. SSF [36] tunes additional scaling and shifting parameters. LoRA [24], AdaptFormer [9], and Consolidator [17] add lightweight modules as" + } + ], + [ + { + "type": "header", + "bbox": [ + 0.469, + 0.115, + 0.732, + 0.129 + ], + "angle": 0, + "content": "PYRA: Parallel Yielding Re-Activation" + }, + { + "type": "page_number", + "bbox": [ + 0.775, + 0.117, + 0.785, + 0.127 + ], + "angle": 0, + "content": "5" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.143, + 0.788, + 0.604 + ], + "angle": 0, + "content": "bypasses. These modules can be merged with the original backbone for no extra inference cost. In the scope of training-inference efficient task adaptation, PEFT methods retain model inference cost, resulting in extreme difficulty for model deployment. To solve the problem, we conduct adaptive token merging for PEFT. While achieving promising adaptation performance under both low and high compression rates, our method inherits the advantages of PEFT methods. Model Compression. Model compression is often applied on large-scale models to acquire smaller-scale models with comparable performance. Mainstream approaches of model compression include model pruning [2,8,10,54,56,64,71], knowledge distillation [1,21,52] and model quantization [3,13,37,38]. For transformer models, model pruning methods can be roughly grouped into two categories: channel pruning and token pruning. Channel pruning methods [5,10, 64,73] reduce the number of parameters, channels, heads, or blocks. Recently, token pruning has emerged as another mainstream approach. Several works have attempted to prune tokens for vision transformers (ViTs). Among these methods, token pooling [43] uses a slow k-means approach that does not work for an off-the-shelf model. ToMe [2] constructs bipartite graphs and merges token pairs with the most weighted connections. DiffRate [8] combines token pruning and token merging with searched optimal token reduction rates for each layer. Although achieving promising results, most existing model compression methods fail when combining with PEFT for training-inference efficient task adaptation since they usually involve a heavy training stage. Model pruning methods demand full re-training after pruning to restore performance. Knowledge distillation necessitates training a small-scale model from scratch, demanding large amount of data. As for model quantization, although no heavy training is demanded, post-training quantization [41,63] achieves poor performance compared to quantization-aware training [25,34] that demands full re-training on the quantized model. As a comparison, our proposed PYRA achieves promising performance on compressed models under the restrictions of PEFT, surpassing the baselines of simply combining model compression and PEFT." + }, + { + "type": "title", + "bbox": [ + 0.216, + 0.624, + 0.381, + 0.642 + ], + "angle": 0, + "content": "3 Methodology" + }, + { + "type": "title", + "bbox": [ + 0.216, + 0.656, + 0.374, + 0.671 + ], + "angle": 0, + "content": "3.1 Preliminaries" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.681, + 0.79, + 0.844 + ], + "angle": 0, + "content": "ViT Model. In this paper, we mainly focus on the training-inference efficient task adaptation of ViT models [16, 20, 51]. A ViT model consists of \\(L\\) identical encoder blocks, each of which consists of a multi-head self-attention (MHSA) module and a feed-forward network (FFN). Formally, an input image \\(\\mathbf{x}\\) is reshaped and linear projected to \\(N\\) tokens \\(\\mathbf{x} = [t_1, t_2, \\dots, t_N]\\) in \\(D\\) dimensions. For simplicity, we omit the classification token ([CLS]) and the distillation token [51]. For encoder block \\(l\\), we denote the input as \\(\\mathbf{x}^l = [t_1^l, t_2^l, \\dots, t_{N^{l-1}}^l] \\in \\mathbb{R}^{N^{l-1} \\times D}\\) and the output as \\(\\hat{\\mathbf{x}}^l = [\\hat{t}_1^l, \\hat{t}_2^l, \\dots, \\hat{t}_{N^l}^l] \\in \\mathbb{R}^{N^l \\times D}\\). For MHSA, the input tokens are first processed by three FC layers to generate \\(Q, K\\), and \\(V\\) matrices, and the output is calculated by Softmax \\((\\frac{QK^T}{\\sqrt{D}}) V\\) before being projected by another FC" + } + ], + [ + { + "type": "page_number", + "bbox": [ + 0.218, + 0.116, + 0.23, + 0.127 + ], + "angle": 0, + "content": "6" + }, + { + "type": "header", + "bbox": [ + 0.272, + 0.115, + 0.372, + 0.129 + ], + "angle": 0, + "content": "Y. Xiong et al." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.147, + 0.784, + 0.177 + ], + "angle": 0, + "content": "layer. For FFN, the tokens are projected by two FC layers. Our method mainly focuses on the input tokens \\(\\mathbf{x}^l\\) before feeding them to the MHSA module." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.178, + 0.785, + 0.268 + ], + "angle": 0, + "content": "LoRA. LoRA [24] is a widely employed PEFT method for task adaptation. Vision transformers consist of large dense parameter matrices. When adapting to a specific task, the updates to the matrices are in small subspaces and can be modeled with low-rank decompositions. LoRA trains only the decomposed matrices during fine-tuning. Specifically, for dense matrix \\( W_0 \\in \\mathbb{R}^{d \\times k} \\) and input \\( x \\), the modified forward pass of updated \\( W_0 \\) is:" + }, + { + "type": "equation", + "bbox": [ + 0.441, + 0.28, + 0.785, + 0.295 + ], + "angle": 0, + "content": "\\[\n\\hat {x} = W _ {0} x + B A x, \\tag {1}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.303, + 0.784, + 0.334 + ], + "angle": 0, + "content": "where \\(B \\in \\mathbb{R}^{d \\times h}\\) and \\(A \\in \\mathbb{R}^{h \\times k}\\). During inference, \\(BA\\) can be merged with \\(W_0\\) for no extra computation overhead." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.335, + 0.784, + 0.395 + ], + "angle": 0, + "content": "Token Merging. Token merging [2,8] is a parameter-free compression technique for vision transformers. It is orthogonal to the transformer structure and capable of flexibly changing the compression rate. Specifically, the input tokens \\(\\mathbf{x}^l\\) of the \\(l\\)-th ViT block are randomly separated before the MHSA:" + }, + { + "type": "equation", + "bbox": [ + 0.289, + 0.405, + 0.785, + 0.423 + ], + "angle": 0, + "content": "\\[\nG _ {1} ^ {l} = \\left[ t _ {i _ {1}} ^ {l}, t _ {i _ {2}} ^ {l}, \\dots , t _ {i _ {s}} ^ {l} \\right], \\quad G _ {2} ^ {l} = \\left[ t _ {j _ {1}} ^ {l}, t _ {j _ {2}} ^ {l}, \\dots , t _ {j _ {s}} ^ {l} \\right], \\quad 2 s = N ^ {l - 1}. \\tag {2}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.433, + 0.784, + 0.496 + ], + "angle": 0, + "content": "Then, for each \\( t_{i_\\bullet}^l \\) token, the most similar \\( t_{j_\\bullet}^l \\) token is matched to it via cosine similarity. From \\( G_1^l \\), \\( r \\) tokens with the most similar connections to tokens in \\( G_2^l \\) are selected to form token pairs \\( (t_{m_k}^l, t_{n_k}^l) \\), where \\( t_{m_k}^l \\in G_1^l \\), \\( t_{n_k}^l \\in G_2^l \\), and \\( k = 1, 2, \\dots, r \\). Note that \\( \\{m_k\\} \\) is a re-indexed subset of \\( \\{i_\\bullet\\} \\). Formally," + }, + { + "type": "equation", + "bbox": [ + 0.272, + 0.506, + 0.785, + 0.545 + ], + "angle": 0, + "content": "\\[\n\\left\\{t _ {m _ {k}} ^ {l} \\right\\} = \\arg \\operatorname {T o p r} _ {\\hat {t} _ {i} ^ {l}} \\left(\\max _ {j} \\frac {\\hat {t} _ {i} ^ {l} \\cdot t _ {j} ^ {l}}{\\| \\hat {t} _ {i} ^ {l} \\| \\cdot \\| t _ {j} ^ {l} \\|}\\right), \\quad t _ {n _ {k}} ^ {l} = \\arg \\max _ {t _ {j} ^ {l}} \\left(\\frac {t _ {m _ {k}} ^ {l} \\cdot t _ {j} ^ {l}}{\\| t _ {m _ {k}} ^ {l} \\| \\cdot \\| t _ {j} ^ {l} \\|}\\right) \\tag {3}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.554, + 0.784, + 0.584 + ], + "angle": 0, + "content": "Previous works [2, 8, 43], generally merge tokens via average pooling, cutting down the computational cost by decreasing the token number by \\( r \\) in layer \\( l \\)." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.585, + 0.785, + 0.706 + ], + "angle": 0, + "content": "We employ the above techniques as our baseline method, in which we attach token merging while fine-tuning LoRA for task adaptation. These methods are selected due to their advantages being highly compatible with training-inference efficient task adaptation. First, token merging is parameter-free and training-free, and does not change model structures, which retains the storage-efficient advantage of PEFT. Besides, we choose LoRA for its popularity, simplicity, and mergeability (not introducing extra FLOPs during inference). Experimental results show that \\(\\mathrm{ToMe + LoRA}\\) is a strong baseline (see Sec. 4)." + }, + { + "type": "title", + "bbox": [ + 0.215, + 0.726, + 0.587, + 0.741 + ], + "angle": 0, + "content": "3.2 PYRA: Parallel Yielding Re-Activation" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.75, + 0.784, + 0.841 + ], + "angle": 0, + "content": "Conventional fine-tuning methods excel at guiding the model to dynamically align with the target data distribution in downstream tasks by extensively adjusting parameters. However, in the context of training-inference efficient task adaptation, only a fraction of parameters can be fine-tuned, posing significant challenges in accurately capturing the nuances of data distribution. While reducing model complexity through token merging shows promise in enhancing" + } + ], + [ + { + "type": "header", + "bbox": [ + 0.469, + 0.115, + 0.731, + 0.129 + ], + "angle": 0, + "content": "PYRA: Parallel Yielding Re-Activation" + }, + { + "type": "page_number", + "bbox": [ + 0.775, + 0.117, + 0.785, + 0.127 + ], + "angle": 0, + "content": "7" + }, + { + "type": "image", + "bbox": [ + 0.219, + 0.145, + 0.784, + 0.284 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.214, + 0.295, + 0.785, + 0.366 + ], + "angle": 0, + "content": "Fig. 3: The pipeline of our PYRA. PYRA conducts token modulation before the MHSA module in each transformer block. Inside PYRA, a pair of learnable modulation weight generators are leveraged to generate adaptive modulation weights parallelly. After that, generated weights modulate tokens through re-activation. The generators in PYRA can be trained along with the LoRA module \\(\\theta_{p}\\) in an end-to-end manner." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.398, + 0.784, + 0.488 + ], + "angle": 0, + "content": "inference efficiency, it introduces the risk of information loss during layer-wise processing in vision transformers. This loss is difficult to rectify due to the constrained understanding of data distribution in the efficient task adaptation scenario. Therefore, a straightforward combination of token merging with PEFT algorithms for achieving training-inference efficient task adaptation may not yield optimal results, as depicted in Fig. 2." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.489, + 0.785, + 0.595 + ], + "angle": 0, + "content": "Here, we propose Parallel Yielding Re-Aviation (PYRA) to adaptively modulate token features to enhance the perception of data distribution during token merging. Specifically, inside PYRA, the weights for adaptive merging are first yielded in a parallel manner through a pair of lightweight learnable vectors in each ViT block. These generated weights are then applied to modulate tokens to be merged through re-activation. As a result, PYRA enables adaptive calibration of the learned feature distribution with low computational complexity." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.597, + 0.786, + 0.779 + ], + "angle": 0, + "content": "Parallel Yielding Adaptive Weights. We aim to optimize the merging process of each chosen token pair. Inspired by feature modulation [44, 59], we propose to modulate token features before merging. Formally, for encoder block \\(l\\) with \\(r\\) pairs of tokens in \\(D\\) dimensions to be merged, we group the \\(t_{m_k}^l\\) and \\(t_{n_k}^l\\) tokens as token matrices \\(M_s^l = [t_{m_1}^l,\\dots ,t_{m_r}^l ]\\) and \\(M_t^l = [t_{n_1}^l,\\dots ,t_{n_r}^l ]\\), where \\(M_s^l,M_t^l\\in \\mathbb{R}^{D\\times r}\\). We learn a modulation matrix \\(W^{l}\\in \\mathbb{R}^{D\\times r}\\) that adaptively modulates tokens at the granularity of each channel. We emphasize that directly learning \\(W^{l}\\) is redundant [24] and cannot adaptively satisfy the conditions with different images and token pairs. Therefore, we further specify the goal as learning decoupled weights \\(\\delta_D^l = f_D(M_s^l,M_t^l)\\in \\mathbb{R}^{D\\times 1}\\) for feature channels and \\(\\delta_r^l = f_r(M_s^l,M_t^l)\\in \\mathbb{R}^{1\\times r}\\) for feature tokens, where \\(W^{l} = \\delta_{D}^{l}\\delta_{r}^{l}\\), separately. We propose to generate \\(\\delta_D^l\\) and \\(\\delta_r^l\\) in a parallel yielding manner." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.78, + 0.787, + 0.841 + ], + "angle": 0, + "content": "Specifically, for layer \\( l \\) with \\( r \\) tokens to be merged, we create two learnable vectors as the modulation weight generator inside the transformer block: \\( W_r^l \\in \\mathbb{R}^{r \\times 1} \\) and \\( W_D^l \\in \\mathbb{R}^{1 \\times D} \\). To guarantee that \\( W_r^l \\) and \\( W_D^l \\) digest the token features from both tokens in a token pair, we first normalize the sum of token pairs to" + } + ], + [ + { + "type": "page_number", + "bbox": [ + 0.218, + 0.116, + 0.23, + 0.127 + ], + "angle": 0, + "content": "8" + }, + { + "type": "header", + "bbox": [ + 0.272, + 0.115, + 0.373, + 0.129 + ], + "angle": 0, + "content": "Y. Xiong et al." + }, + { + "type": "text", + "bbox": [ + 0.216, + 0.147, + 0.506, + 0.161 + ], + "angle": 0, + "content": "construct the token information matrix:" + }, + { + "type": "equation", + "bbox": [ + 0.361, + 0.173, + 0.786, + 0.192 + ], + "angle": 0, + "content": "\\[\nM _ {\\text {i n f o}} ^ {l} = \\operatorname {L a y e r N o r m} \\left(M _ {s} ^ {l} + M _ {t} ^ {l}\\right) \\in \\mathbb {R} ^ {D \\times r}. \\tag {4}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.215, + 0.201, + 0.787, + 0.248 + ], + "angle": 0, + "content": "We normalize the distribution of \\( M_{\\mathrm{info}}^{l} \\) tokens by leveraging the LayerNorm(\\cdot) operation to enable smoother gradients when training \\( W_{r}^{l} \\) and \\( W_{D}^{l} \\). With the token information matrix, we then yield the adaptive weights \\( \\delta_{D}^{l} \\) and \\( \\delta_r^l \\) parallelly:" + }, + { + "type": "equation", + "bbox": [ + 0.419, + 0.258, + 0.786, + 0.285 + ], + "angle": 0, + "content": "\\[\n\\delta_ {D} ^ {l} = M _ {\\text {i n f o}} ^ {l} W _ {r} ^ {l} \\in \\mathbb {R} ^ {D \\times 1} \\tag {5}\n\\]" + }, + { + "type": "equation", + "bbox": [ + 0.424, + 0.278, + 0.588, + 0.297 + ], + "angle": 0, + "content": "\\[\n\\delta_ {r} ^ {l} = W _ {D} ^ {l} M _ {\\mathrm {i n f o}} ^ {l} \\in \\mathbb {R} ^ {1 \\times r}.\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.215, + 0.303, + 0.789, + 0.442 + ], + "angle": 0, + "content": "Re-Activation for Token Modulation. Simply generating \\(\\delta_D^l\\) and \\(\\delta_r^l\\) via matrix multiplication still faces several possible issues. First, no measures have taken to ensure that \\(\\delta_D^l\\) and \\(\\delta_r^l\\) stay in a normal range. Second, decoupling weights to feature tokens and feature channels results in a low-rank modulation weight matrix \\(W^l\\), which exhibits limited expressive capacity and thus could be unable to optimally modulate tokens in complicated data distributions. To cope with these issues, we conduct token modulation in a re-activation strategy. Specifically, we first broadcast \\(\\delta_D^l\\) to \\(\\hat{\\delta}_D^l \\in \\mathbb{R}^{D \\times r}\\) and conduct sigmoid activation \\(\\sigma(\\cdot)\\) on it, and then modulate \\(M_s^l\\) for an intermediate modulation result:" + }, + { + "type": "equation", + "bbox": [ + 0.432, + 0.452, + 0.786, + 0.47 + ], + "angle": 0, + "content": "\\[\n\\hat {M} _ {s} ^ {l} = 2 \\sigma \\left(\\hat {\\delta} _ {D} ^ {l}\\right) \\odot M _ {s} ^ {l}. \\tag {6}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.215, + 0.479, + 0.787, + 0.513 + ], + "angle": 0, + "content": "where \\(\\odot\\) is Hadamard product. \\(\\hat{M}_s^l\\) is then modulated with sigmoid-activated broadcast weight \\(\\hat{\\delta}_r^l\\in \\mathbb{R}^{D\\times r}\\) again to acquire the modulated tokens:" + }, + { + "type": "equation", + "bbox": [ + 0.391, + 0.524, + 0.786, + 0.542 + ], + "angle": 0, + "content": "\\[\nM _ {s} ^ {l} \\leftarrow M _ {s} ^ {l} + \\left(2 \\sigma \\left(\\hat {\\delta} _ {r} ^ {l}\\right) - 1\\right) \\odot \\hat {M} _ {s} ^ {l}, \\tag {7}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.215, + 0.552, + 0.788, + 0.659 + ], + "angle": 0, + "content": "Note that we use the original tokens \\( M_{s}^{l} \\) to create a residual connection that preserves the gradient flow during training. The modulated \\( M_{s}^{l} \\) tokens are then merged with \\( M_{t}^{l} \\) with average pooling. We use a random Gaussian initialization for generator \\( W_{r}^{l} \\) and zero for generator \\( W_{D}^{l} \\), so re-activation is equivalent to identity transformation at the beginning of training. Token modulation is conducted only on \\( M_{s}^{l} \\) to guarantee the parallelism in training and inference, as \\( t_{m_k}^l \\) tokens are distinct while different \\( t_{n_k}^l \\) might point to the same \\( t_j^l \\) token." + }, + { + "type": "text", + "bbox": [ + 0.215, + 0.66, + 0.789, + 0.841 + ], + "angle": 0, + "content": "Discussion. Our PYRA effectively enhances the calibration of features for downstream tasks. Specifically, by utilizing parallel yielding adaptive weights, PYRA effectively decouples the modulation weight from the feature token and the feature channel, i.e., \\(\\delta_{D}\\) and \\(\\delta_{r}\\), enabling comprehensive perception of the feature distribution in downstream tasks. Furthermore, considering the challenge of capturing an accurate feature distribution for parameter-efficient training, such smooth re-activations by Eq. (6) and Eq. (7) can constrain the negative impact on weight values brought by limited perception of feature distributions, thus leading to better token modulation for improving feature representations in downstream tasks. As a result, the proposed PYRA can maintain discriminative information to an utmost degree during token merging, leading to improved performance while achieving complexity reduction." + } + ], + [ + { + "type": "header", + "bbox": [ + 0.469, + 0.115, + 0.732, + 0.129 + ], + "angle": 0, + "content": "PYRA: Parallel Yielding Re-Activation" + }, + { + "type": "page_number", + "bbox": [ + 0.775, + 0.117, + 0.785, + 0.127 + ], + "angle": 0, + "content": "9" + }, + { + "type": "table_caption", + "bbox": [ + 0.217, + 0.145, + 0.784, + 0.173 + ], + "angle": 0, + "content": "Table 1: The complexity comparisons between conducting PEFT with and without PYRA. The FLOPs metric is obtained during inference." + }, + { + "type": "table", + "bbox": [ + 0.261, + 0.185, + 0.744, + 0.264 + ], + "angle": 0, + "content": "
ModelMetricTotalPEFT w/o PYRA (%)PEFT w. PYRA (%)
ViT-Base# params86M0.29M0.34%0.30M0.35%
FLOPs16.37G16.37G100%8.15G49.79%
ViT-Large# params303M1.18M0.39%1.20M0.40%
FLOPs57.37G57.37G100%28.76G50.13%
" + }, + { + "type": "title", + "bbox": [ + 0.217, + 0.29, + 0.432, + 0.306 + ], + "angle": 0, + "content": "3.3 Complexity Analysis" + }, + { + "type": "text", + "bbox": [ + 0.217, + 0.313, + 0.785, + 0.387 + ], + "angle": 0, + "content": "We present the parameter and computation complexity for introducing PYRA to PEFT for task adaptation. In each ViT block, PYRA introduces \\( D + r \\) training parameters and \\( 4rD \\) extra FLOPs. For a ViT model with \\( L \\) layers and \\( R \\) total merged tokens, in total, our PYRA introduces \\( LD + R \\) extra training parameters and conducts \\( 4RD \\) extra FLOPs beside PEFT." + }, + { + "type": "text", + "bbox": [ + 0.217, + 0.389, + 0.785, + 0.525 + ], + "angle": 0, + "content": "To better demonstrate the training-inference efficiency of our PYRA, we compare the complexity between attaching PYRA to task adaptation via PEFT (here we employ LoRA [24]) and plainly conducting PEFT without PYRA. As shown in Tab. 1, PYRA keeps the training efficient feature of PEFT by introducing only a tiny amount of training parameters, which however, results in a substantial efficiency boost for inference, i.e., at around \\(50\\%\\) for both ViT-B and ViT-L models. Results in Sec. 4 show that PYRA achieves comparable performance to the PEFT counterpart without PYRA. This well indicates that PYRA is an effective method for training-inference efficient task adaptation." + }, + { + "type": "title", + "bbox": [ + 0.217, + 0.546, + 0.375, + 0.563 + ], + "angle": 0, + "content": "4 Experiments" + }, + { + "type": "title", + "bbox": [ + 0.217, + 0.576, + 0.438, + 0.592 + ], + "angle": 0, + "content": "4.1 Experimental Setting" + }, + { + "type": "text", + "bbox": [ + 0.217, + 0.599, + 0.785, + 0.704 + ], + "angle": 0, + "content": "Datasets. We conduct extensive experiments on task adaptation benchmarks to verify the effectiveness of PYRA within the challenge of training-inference efficient task adaptation. Specifically, we choose the VTAB-1k [67] benchmark for the evaluation. VTAB-1k is a challenging benchmark that consists of 19 different tasks from diverse domains: 1) natural images captured in the actual world; 2) specialized images from professional fields; and 3) structured synthesized images. Each task only contains 800 training samples and 200 validation samples." + }, + { + "type": "text", + "bbox": [ + 0.217, + 0.705, + 0.785, + 0.764 + ], + "angle": 0, + "content": "Models. We choose two ViT backbones pre-trained on the ImageNet-21K [16], i.e., ViT-L/16 and ViT-B/16, for comparison. Additionally, we also generalize our PYRA to different backbones and pre-train methods, including DeiT-B [51] and ViT backbones pre-trained by MAE [20]." + }, + { + "type": "text", + "bbox": [ + 0.217, + 0.765, + 0.785, + 0.84 + ], + "angle": 0, + "content": "Implementation Details. We choose LoRA [24] as the PEFT module for all methods due to its simplicity and mergeability. For ease of explanation, we omit the \"+\"LoRA\" when comparing different methods. We append LoRA only on the \\(Q,K\\), and \\(V\\) projection matrices, and apply the training schedule for LoRA following [17, 69]. The generators are trained along with LoRA" + } + ], + [ + { + "type": "page_number", + "bbox": [ + 0.218, + 0.116, + 0.236, + 0.127 + ], + "angle": 0, + "content": "10" + }, + { + "type": "header", + "bbox": [ + 0.272, + 0.115, + 0.373, + 0.129 + ], + "angle": 0, + "content": "Y. Xiong et al." + }, + { + "type": "table_caption", + "bbox": [ + 0.216, + 0.145, + 0.787, + 0.174 + ], + "angle": 0, + "content": "Table 2: Results on the VTAB-1k [67] benchmark under low compression rate. Bold and underline denote the best and second-best accuracy within compression methods." + }, + { + "type": "table", + "bbox": [ + 0.2, + 0.186, + 0.803, + 0.39 + ], + "angle": 0, + "content": "
Method# paramsThroughputNaturalSpecializedStructured
Cifar100Caltech101DTDFlowers102PetsSVHNSun397CamelyonEuroSATResisc45RetinopathyClevr-CountClevr-DistDMLABKITTI-DistdSpr-LocdSpr-OrisNORB-AzimsNORB-EleAverage
Model: ViT-B/16 (Throughput: 425)
PEFT0.34%42567.190.269.499.190.585.754.183.195.884.374.682.269.250.179.281.847.131.142.674.76
RaP3.43%65425.968.453.364.057.471.321.575.887.959.373.643.153.826.360.573.525.516.727.955.57
SPViT4.46%56741.675.461.183.266.256.128.379.394.273.373.670.661.542.467.875.450.528.931.364.16
DiffRate0.35%70937.184.663.796.786.232.648.278.985.867.073.732.929.834.155.712.616.013.121.555.82
ToMe0.34%75364.690.467.998.589.883.953.282.694.783.574.981.969.849.276.981.946.531.043.174.10
PYRA0.35%74567.590.369.398.990.084.653.183.395.783.375.282.668.950.880.081.845.832.242.874.69
Model: ViT-L/16 (Throughput: 130)
PEFT0.39%13077.191.473.499.591.389.657.685.996.187.376.183.163.050.782.181.753.532.236.676.52
RaP1.95%19643.287.962.652.881.786.734.778.492.473.373.668.059.646.982.475.543.624.525.765.64
SPViT2.47%18848.187.565.294.477.480.938.879.993.979.874.378.265.847.474.182.350.331.037.970.22
DiffRate0.39%22150.986.870.397.888.339.052.380.287.272.274.232.632.336.557.422.826.615.223.459.53
ToMe0.39%22776.191.172.399.291.789.256.486.495.186.675.182.461.950.981.481.653.533.436.876.11
PYRA0.40%22576.691.373.299.391.589.457.186.995.987.176.283.263.252.883.182.552.634.839.076.84
" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.416, + 0.788, + 0.464 + ], + "angle": 0, + "content": "modules. During inference, we merge the LoRA module to the backbone. All throughputs are measured during inference on a GeForce RTX 3090 GPU. More details can be found in the supplementary materials." + }, + { + "type": "title", + "bbox": [ + 0.214, + 0.484, + 0.734, + 0.5 + ], + "angle": 0, + "content": "4.2 Performance comparison on Task Adaptation Benchmark" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.508, + 0.788, + 0.689 + ], + "angle": 0, + "content": "We verify the effectiveness of PYRA for training-inference efficient task adaptation on (1) low compression rate: comparing the performance of different methods under the same sparsity ratio (here we set sparsity ratio \\(= 50\\%\\)); (2) high compression rate: comparing the performance between the competing methods and the smaller-scale model with similar throughput levels. We leverage ViT-L/16 and ViT-B/16 for these inspections. The competing methods include RaP [35], SPViT [30], DiffRate [8], and ToMe [2]. Comparisons to more baselines are in the supplementary materials. For RaP and SPViT, we first train the backbone with LoRA on downstream tasks, then we prune the models, and lastly we retrain the parameters attached by the pruning method and LoRA. DiffRate and ToMe can be employed with LoRA during fine-tuning, while DiffRate demands ImageNet-21K for searching the optimal compression schedule." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.69, + 0.788, + 0.842 + ], + "angle": 0, + "content": "PYRA on low compression rate. Results on low compression rate are reported in Tab. 2. Overall, while achieving one of the best speedups on throughput, our PYRA is the best performed method compared to other competing methods using the lowest level of training parameters. Counting results on both backbones, our PYRA achieves the best or second-best performance on 37 of 38 dataset metrics. Compared to directly conducting PEFT on the backbone, while all competing methods cause worse results, our PYRA achieves comparable adaptation performance on ViT-B/16, and even outperforms ViT-L/16. The above results convincingly demonstrate that our PYRA successfully sets a new benchmark, i.e. reaching comparable performance as the uncompressed model," + } + ], + [ + { + "type": "header", + "bbox": [ + 0.469, + 0.115, + 0.732, + 0.129 + ], + "angle": 0, + "content": "PYRA: Parallel Yielding Re-Activation" + }, + { + "type": "page_number", + "bbox": [ + 0.768, + 0.117, + 0.784, + 0.127 + ], + "angle": 0, + "content": "11" + }, + { + "type": "table_caption", + "bbox": [ + 0.214, + 0.145, + 0.788, + 0.201 + ], + "angle": 0, + "content": "Table 3: Results on the VTAB-1k [67] benchmark under high compression rate. Bold and underline denote the best and second-best accuracy within compression methods. \\*: As a comparison of similar throughputs, we compare ViT-B/16 with PEFT on ViTS/16, and ViT-L/16 with PEFT on ViT-B/16." + }, + { + "type": "table", + "bbox": [ + 0.197, + 0.213, + 0.803, + 0.414 + ], + "angle": 0, + "content": "
Method# paramsThroughputNaturalSpecializedStructured
Cifar100Caltech101DTDFlowers102PetsSVHNSun397CamelyonEuroSATResisc45RetinopathyClevr-CountClevr-DistDMLABKITTI-DistdSpr-LocdSpr-OrisNORB-AzimsNORB-EleAverage
Model: ViT-B/16 (Throughput: 425)
PEFT*0.34%135057.788.270.198.788.785.744.981.494.784.673.681.664.148.180.072.938.422.937.771.85
RaP0.86%102924.340.134.541.840.521.711.475.886.535.173.849.649.728.139.413.815.412.426.942.60
SPViT4.46%94423.767.951.969.953.219.613.171.981.367.974.753.561.939.557.445.034.511.123.252.49
DiffRate0.35%130823.273.055.787.966.727.229.378.177.853.173.629.728.631.752.611.517.211.320.349.29
ToMe0.34%138154.287.865.596.181.779.745.279.493.676.373.878.365.748.071.380.045.830.941.270.43
PYRA0.35%136554.089.367.196.584.081.844.681.294.679.575.179.967.049.276.982.647.831.942.072.06
Model: ViT-L/16 (Throughput: 130)
PEFT*0.34%42567.190.269.499.190.585.754.183.195.884.374.682.269.250.179.281.847.131.142.674.76
RaP0.65%30117.737.127.046.233.323.213.376.574.254.473.650.431.425.749.853.125.513.426.044.11
SPViT2.47%28954.087.665.594.874.932.638.681.895.378.074.072.861.246.970.277.147.431.328.666.90
DiffRate0.39%41647.473.554.184.360.219.622.250.064.642.818.231.531.931.137.322.017.714.821.440.48
ToMe0.39%43171.090.970.498.388.587.252.482.994.583.175.080.761.148.976.980.853.032.135.274.10
PYRA0.40%42771.691.871.198.589.788.152.285.195.384.675.780.963.051.782.082.054.236.041.275.66
" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.446, + 0.785, + 0.476 + ], + "angle": 0, + "content": "on low compression rate for training-inference efficient task adaptation while accelerating the model to \\(1.75 \\times\\) speedup with only \\(0.4\\%\\) training parameters." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.477, + 0.787, + 0.642 + ], + "angle": 0, + "content": "PYRA on high compression rate. Results are reported in Tab. 3. With comparable throughputs to the smaller-scale model and minimal training parameters, PYRA outperforms all competing methods and achieves the best performance on 35 of 38 dataset metrics. Compared to the smaller-scale model with comparable throughput, PYRA successfully outperforms it. On the compressed ViT-L/16, PYRA even surpasses ViT-B/16 by \\(0.9\\%\\). This shows that PYRA, as the state-of-the-art, effectively bridges the adverse compression gap between compressed large-scale models and small-scale models, as mentioned in Sec. 1. Therefore, PYRA is an applicable alternative for acquiring a smaller-scale model \\((3.2\\times\\) speedup) on downstream tasks through efficient training \\((0.4\\%)\\) training parameters) when no pre-trained small-scale model is available." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.643, + 0.787, + 0.734 + ], + "angle": 0, + "content": "PYRA on self-supervised ViT backbone. We conduct the experiments on both compression rates on self-supervised ViT-L/16 (MAE) [20]. As shown in Tab. 4, PYRA significantly surpasses all competing methods. Under low compression rate, PYRA outperforms directly transferring the uncompressed backbone. Under high compression rate, PYRA also eliminates adverse compression. These results show that PYRA generalizes well for self-supervised visual models." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.735, + 0.787, + 0.809 + ], + "angle": 0, + "content": "PYRA on different architectures. To further verify the generalizability of PYRA, we conduct the above experiments on the DeiT-B [51]. As shown in Tab. 5, PYRA achieves comparable performance to the uncompressed model under low compression rate and eliminates adverse compression under high compression rate, indicating its generalizability to other transformer architectures." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.81, + 0.787, + 0.84 + ], + "angle": 0, + "content": "PYRA consistently yields better models of different throughputs. To further show the superiority of PYRA for different compression rates, we" + } + ], + [ + { + "type": "page_number", + "bbox": [ + 0.218, + 0.116, + 0.236, + 0.127 + ], + "angle": 0, + "content": "12" + }, + { + "type": "header", + "bbox": [ + 0.272, + 0.115, + 0.373, + 0.129 + ], + "angle": 0, + "content": "Y. Xiong et al." + }, + { + "type": "table_caption", + "bbox": [ + 0.216, + 0.144, + 0.488, + 0.188 + ], + "angle": 0, + "content": "Table 4: VTAB-1k [67] results on both compression rates for ViT-L (MAE) (Throughput: 130)." + }, + { + "type": "table", + "bbox": [ + 0.217, + 0.206, + 0.49, + 0.398 + ], + "angle": 0, + "content": "
ModelMethod# paramsThroughputAverage
ViT-L (MAE)PEFT0.39%13075.96
ViT-L (MAE)RaP [35]2.01%18267.55
DiffRate [8]0.39%22146.91
ToMe [2]0.39%22774.97
PYRA0.40%22576.13
ViT-B (MAE)PEFT0.34%42570.23
ViT-L (MAE)RaP [35]0.76%29852.91
DiffRate [8]0.39%41648.03
ToMe [2]0.39%43168.20
PYRA0.40%42770.33
" + }, + { + "type": "table_caption", + "bbox": [ + 0.515, + 0.144, + 0.788, + 0.188 + ], + "angle": 0, + "content": "Table 5: VTAB-1k [67] results on both compression rates for DeiT-B (Throughput: 431)." + }, + { + "type": "table", + "bbox": [ + 0.516, + 0.206, + 0.789, + 0.398 + ], + "angle": 0, + "content": "
ModelMethod# paramsThroughputAverage
DeiT-BPEFT0.34%43173.76
DeiT-BRaP [35]3.57%69562.34
DiffRate [8]0.35%73452.10
ToMe [2]0.34%74772.83
PYRA0.35%74073.55
DeiT-SPEFT0.34%133270.01
DeiT-BRaP [35]1.09%118757.70
DiffRate [8]0.35%131444.51
ToMe [2]0.34%135168.68
PYRA0.35%134170.13
" + }, + { + "type": "table_caption", + "bbox": [ + 0.216, + 0.41, + 0.784, + 0.439 + ], + "angle": 0, + "content": "Table 6: Ablation study results for PYRA on ViT-B/16 under high compression rate. Here for # params we report only the parameters introduced by token modulation." + }, + { + "type": "table", + "bbox": [ + 0.236, + 0.451, + 0.77, + 0.555 + ], + "angle": 0, + "content": "
Method\\( W_r \\)\\( W_D \\)Activation# paramsNaturalSpecializedStructuredAverage
Baseline×××072.8780.7857.6470.43
Plain \\( W_r \\)√××0.19K72.9081.0757.6670.54
\\( σ(\\cdot) \\& W_r \\)√×√0.19K73.1881.6957.6570.84
Plain \\( W_D \\)×√×8.45K73.0981.1358.4370.88
\\( σ(\\cdot) \\& W_D \\)×√√8.45K73.3182.1758.4471.31
Plain \\( W_r \\& W_D \\)√√×8.64K73.7781.3758.8171.32
PYRA√√√8.64K73.9182.6059.6672.06
" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.583, + 0.788, + 0.675 + ], + "angle": 0, + "content": "compress the chosen backbones for a series of speedups and compare the results with the strongest baseline, ToMe [2], under similar values of throughputs. As shown in Figs. 2 and 4, PYRA consistently outperforms ToMe and flattens the accuracy-throughput curve. Therefore, our PYRA is applicable to acquire task-specific smaller-scale models of different throughputs consistently in the absence of pre-trained parameters for smaller-scale models." + }, + { + "type": "title", + "bbox": [ + 0.216, + 0.695, + 0.403, + 0.71 + ], + "angle": 0, + "content": "4.3 Ablation Studies" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.719, + 0.789, + 0.841 + ], + "angle": 0, + "content": "We do controlled experiments to identify the effect of individual components in PYRA, i.e., the generators \\( W_{r} \\) and \\( W_{D} \\), and the sigmoid activation \\( \\sigma(\\cdot) \\) applied on the generated modulation weights. Specifically, when using only a single generator, we omit Eq. (6) and replace \\( \\hat{\\delta}_r^l \\) in Eq. (7) with the broadcast weight generated by the employed generator. When using the generators without sigmoid activation, we simply remove the \\( \\sigma \\) in Eq. (6) and Eq. (7) while keeping other calculations intact. We choose the ImageNet-21K pre-trained ViT-B/16 to carry out the ablation studies under high compression rates as in Tab. 3." + } + ], + [ + { + "type": "header", + "bbox": [ + 0.469, + 0.115, + 0.732, + 0.129 + ], + "angle": 0, + "content": "PYRA: Parallel Yielding Re-Activation" + }, + { + "type": "page_number", + "bbox": [ + 0.769, + 0.116, + 0.785, + 0.127 + ], + "angle": 0, + "content": "13" + }, + { + "type": "image", + "bbox": [ + 0.225, + 0.151, + 0.481, + 0.314 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.515, + 0.151, + 0.77, + 0.314 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.214, + 0.331, + 0.788, + 0.388 + ], + "angle": 0, + "content": "Fig. 4: Comparisons between PYRA and ToMe [2] under different compression rates for ViT-Large (MAE) and DeiT-Base. Red boxes: PYRA mitigates the performance drops under low compression rate. Green boxes: PYRA eliminates adverse compression under high compression rate." + }, + { + "type": "table_caption", + "bbox": [ + 0.216, + 0.403, + 0.499, + 0.446 + ], + "angle": 0, + "content": "Table 7: Adaptation performance comparison to \\( W_{r \\times D} \\) unadaptable to tokens. We report token modulation parameters." + }, + { + "type": "table_caption", + "bbox": [ + 0.504, + 0.403, + 0.788, + 0.447 + ], + "angle": 0, + "content": "Table 8: Adaptation performance comparison to the common gated generator. We report token modulation parameters." + }, + { + "type": "table", + "bbox": [ + 0.222, + 0.457, + 0.499, + 0.572 + ], + "angle": 0, + "content": "
Method# paramsNaturalSpecializedStructuredAverage
Baseline072.980.857.670.43
WD×r70.66K73.881.559.271.49
PYRA8.64K73.982.659.772.06
" + }, + { + "type": "table", + "bbox": [ + 0.51, + 0.457, + 0.789, + 0.571 + ], + "angle": 0, + "content": "
Method# paramsNaturalSpecializedStructuredAverage
Baseline072.980.857.670.43
Gated73.73K72.881.558.971.06
PYRA8.64K73.982.659.772.06
" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.602, + 0.789, + 0.74 + ], + "angle": 0, + "content": "Results are shown in Tab. 6. Here the baseline method refers to conducting token merging [2,8] while fine-tuning LoRA [24] for task adaptation. First, both \\( W_{r} \\) and \\( W_{D} \\), no matter whether sigmoid activation is attached, lead to performance gains, and employing them simultaneously outperforms using them individually. This proves the superiority of our parallel yielding strategy. Second, compared to the corresponding counterpart without activations, employing sigmoid activation always leads to significant improvements, indicating the effectiveness of the reactivation. Overall, our PYRA yields the best adaptation performance, showing that our strategies are effective and complementary." + }, + { + "type": "title", + "bbox": [ + 0.216, + 0.765, + 0.577, + 0.781 + ], + "angle": 0, + "content": "4.4 Further Analysis on Different Designs" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.795, + 0.786, + 0.842 + ], + "angle": 0, + "content": "We conduct further analysis on ViT-B/16 model pre-trained on ImageNet-21K. The baseline here refers to training LoRA with token merging [2, 8] attached. More analysis experiments can be found in the supplementary materials." + } + ], + [ + { + "type": "page_number", + "bbox": [ + 0.218, + 0.116, + 0.236, + 0.127 + ], + "angle": 0, + "content": "14" + }, + { + "type": "header", + "bbox": [ + 0.272, + 0.115, + 0.372, + 0.129 + ], + "angle": 0, + "content": "Y. Xiong et al." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.147, + 0.788, + 0.286 + ], + "angle": 0, + "content": "Impact of Making Modulation Weights Adaptive. In our PYRA, we train modulation weight generators \\( W_{D}^{l} \\) and \\( W_{r}^{l} \\) to conduct adaptive modulation on different merging tokens. To prove the necessity of making modulation weights adaptive, we compare PYRA with the approach of directly training final modulation weights \\( W_{D\\times r}^{l} \\) for each layer, and conduct token modulation as \\( M_{s}^{l}\\gets M_{s}^{l} + (2\\sigma (W_{D\\times r}^{l}) - 1)\\odot M_{s}^{l} \\). For fair comparisons, we inherit the sigmoid activation and the residual connection. As shown in Tab. 7, although with more training parameters, training \\( W_{D\\times r} \\) is still inferior to our PYRA. This indicates that PYRA is a more effective strategy to conduct adaptive token modulation." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.288, + 0.789, + 0.486 + ], + "angle": 0, + "content": "Compare with the Common Gated Generator. We compare PYRA with setting the commonly-applied gated-style trainable module [12, 22, 31, 62] as modulation weight generator. Formally, for each ViT block, we insert a learnable two-layer MLP module \\((\\mathrm{MLP}^l(\\cdot))\\) with input dimension \\(D\\), hidden dimension \\(d \\ll D\\) to ensure parameter-efficient training and fast inference, and output dimension \\(D\\). To generate modulation weights, we feed the information matrix \\(M_{\\mathrm{info}}^l\\) into the MLP, and modulate the merging tokens thereafter: \\(\\delta^l = \\mathrm{MLP}(M_{\\mathrm{info}}^l) \\in \\mathbb{R}^{D \\times r}\\), \\(M_s^l \\gets M_s^l + (2\\sigma(\\delta^l - 1) \\odot M_s^l)\\). We set \\(d = 4\\) for all layers. As shown in Tab. 8, both the gated generator and our PYRA achieve performance gains, while our PYRA surpasses the gated generator by \\(1.0\\%\\) with significantly fewer trainable parameters. This indicates that the decoupled weights generated by our \\(W_D\\) and \\(W_r\\) in PYRA are effective and parameter-efficient compared to the common gating strategy." + }, + { + "type": "title", + "bbox": [ + 0.216, + 0.528, + 0.36, + 0.544 + ], + "angle": 0, + "content": "5 Conclusion" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.579, + 0.788, + 0.761 + ], + "angle": 0, + "content": "In this work, we defined and investigated a new challenge named training-inference efficient task adaptation, in which the inference efficiency of large-scale transformers is enhanced during parameter-efficient task adaptation. We propose a novel Parallel Yielding Re-Activation method (PYRA) to effectively cope with the challenge by modulating token features during token merging. Specifically, PYRA generates decoupled parallel yielding modulation weights, and conducts token modulation through re-activation. Extensive experiments show that PYRA introduces negligible performance drops under low compression rate, and bridges the gap of adverse compression between compressed transformers and small-scale models under high compression rate. In real-world applications, our PYRA is highly suitable to the scenario of transferring large-scale vision transformers to downstream tasks where no small-scale model is presented." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.765, + 0.788, + 0.841 + ], + "angle": 0, + "content": "Limitations. We have not yet validated the effectiveness of PYRA in object detection and image segmentation. We plan to extend these tasks in the future. Meanwhile, it is also applicable to attach pruning methods after training the vision transformers with PYRA to eliminate additional parameters attached by PYRA. We leave that topic to future studies." + } + ], + [ + { + "type": "header", + "bbox": [ + 0.469, + 0.115, + 0.732, + 0.129 + ], + "angle": 0, + "content": "PYRA: Parallel Yielding Re-Activation" + }, + { + "type": "page_number", + "bbox": [ + 0.768, + 0.116, + 0.786, + 0.127 + ], + "angle": 0, + "content": "15" + }, + { + "type": "title", + "bbox": [ + 0.218, + 0.145, + 0.403, + 0.163 + ], + "angle": 0, + "content": "Acknowledgements" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.178, + 0.816, + 0.24 + ], + "angle": 0, + "content": "This work was supported by National Science and Technology Major 2022ZD0119401, National Natural Science Foundation of China (Nos. 62271281, 61925107, 62021002). It is also sponsored by CAAI-CANN Open Fund, developed on OpenI Community." + }, + { + "type": "title", + "bbox": [ + 0.217, + 0.264, + 0.323, + 0.279 + ], + "angle": 0, + "content": "References" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.296, + 0.787, + 0.338 + ], + "angle": 0, + "content": "1. Ahn, S., Hu, S.X., Damianou, A., Lawrence, N.D., Dai, Z.: Variational information distillation for knowledge transfer. In: Proceedings of the IEEE/CVF conference on computer vision and pattern recognition. pp. 9163-9171 (2019)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.34, + 0.787, + 0.367 + ], + "angle": 0, + "content": "2. Bolya, D., Fu, C.Y., Dai, X., Zhang, P., Feichtenhofer, C., Hoffman, J.: Token merging: Your vit but faster. arXiv preprint arXiv:2210.09461 (2022)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.368, + 0.787, + 0.408 + ], + "angle": 0, + "content": "3. Cai, Y., Yao, Z., Dong, Z., Gholami, A., Mahoney, M.W., Keutzer, K.: Zeroq: A novel zero shot quantization framework. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 13169-13178 (2020)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.409, + 0.787, + 0.45 + ], + "angle": 0, + "content": "4. Carion, N., Massa, F., Synnaeve, G., Usunier, N., Kirillov, A., Zagoruyko, S.: End-to-end object detection with transformers. In: European conference on computer vision. pp. 213-229. Springer (2020)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.451, + 0.787, + 0.506 + ], + "angle": 0, + "content": "5. Chavan, A., Shen, Z., Liu, Z., Liu, Z., Cheng, K.T., Xing, E.P.: Vision transformer slimming: Multi-dimension searching in continuous optimization space. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 4931-4941 (2022)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.507, + 0.787, + 0.548 + ], + "angle": 0, + "content": "6. Chen, C.F.R., Fan, Q., Panda, R.: Crossvit: Cross-attention multi-scale vision transformer for image classification. In: Proceedings of the IEEE/CVF international conference on computer vision. pp. 357-366 (2021)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.549, + 0.787, + 0.589 + ], + "angle": 0, + "content": "7. Chen, L., Zhao, H., Liu, T., Bai, S., Lin, J., Zhou, C., Chang, B.: An image is worth 1/2 tokens after layer 2: Plug-and-play inference acceleration for large vision-language models. arXiv preprint arXiv:2403.06764 (2024)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.59, + 0.787, + 0.631 + ], + "angle": 0, + "content": "8. Chen, M., Shao, W., Xu, P., Lin, M., Zhang, K., Chao, F., Ji, R., Qiao, Y., Luo, P.: Diffrate: Differentiable compression rate for efficient vision transformers. arXiv preprint arXiv:2305.17997 (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.632, + 0.787, + 0.673 + ], + "angle": 0, + "content": "9. Chen, S., Ge, C., Tong, Z., Wang, J., Song, Y., Wang, J., Luo, P.: Adaptformer: Adapting vision transformers for scalable visual recognition. Advances in Neural Information Processing Systems 35, 16664-16678 (2022)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.674, + 0.787, + 0.715 + ], + "angle": 0, + "content": "0. Chen, T., Cheng, Y., Gan, Z., Yuan, L., Zhang, L., Wang, Z.: Chasing sparsity in vision transformers: An end-to-end exploration. Advances in Neural Information Processing Systems 34, 19974-19988 (2021)" + }, + { + "type": "ref_text", + "bbox": [ + 0.22, + 0.716, + 0.787, + 0.77 + ], + "angle": 0, + "content": "11. Cheng, B., Misra, I., Schwing, A.G., Kirillov, A., Girdhar, R.: Masked-attention mask transformer for universal image segmentation. In: Proceedings of the IEEE/CVF conference on computer vision and pattern recognition. pp. 1290-1299 (2022)" + }, + { + "type": "ref_text", + "bbox": [ + 0.22, + 0.771, + 0.787, + 0.798 + ], + "angle": 0, + "content": "12. Chung, J., Gulcehre, C., Cho, K., Bengio, Y.: Empirical evaluation of gated recurrent neural networks on sequence modeling. arXiv preprint arXiv:1412.3555 (2014)" + }, + { + "type": "ref_text", + "bbox": [ + 0.22, + 0.799, + 0.787, + 0.84 + ], + "angle": 0, + "content": "13. Courbariaux, M., Bengio, Y., David, J.P.: Binaryconnect: Training deep neural networks with binary weights during propagations. Advances in neural information processing systems 28 (2015)" + }, + { + "type": "list", + "bbox": [ + 0.22, + 0.296, + 0.787, + 0.84 + ], + "angle": 0, + "content": null + } + ], + [ + { + "type": "page_number", + "bbox": [ + 0.218, + 0.116, + 0.236, + 0.127 + ], + "angle": 0, + "content": "16" + }, + { + "type": "header", + "bbox": [ + 0.272, + 0.115, + 0.372, + 0.128 + ], + "angle": 0, + "content": "Y. Xiong et al." + }, + { + "type": "ref_text", + "bbox": [ + 0.218, + 0.147, + 0.787, + 0.189 + ], + "angle": 0, + "content": "14. Ding, N., Qin, Y., Yang, G., Wei, F., Yang, Z., Su, Y., Hu, S., Chen, Y., Chan, C.M., Chen, W., et al.: Delta tuning: A comprehensive study of parameter efficient methods for pre-trained language models. arXiv preprint arXiv:2203.06904 (2022)" + }, + { + "type": "ref_text", + "bbox": [ + 0.218, + 0.189, + 0.787, + 0.245 + ], + "angle": 0, + "content": "15. Ding, Z., Wang, A., Chen, H., Zhang, Q., Liu, P., Bao, Y., Yan, W., Han, J.: Exploring structured semantic prior for multi label recognition with incomplete labels. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 3398-3407 (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.218, + 0.245, + 0.787, + 0.3 + ], + "angle": 0, + "content": "16. Dosovitskiy, A., Beyer, L., Kolesnikov, A., Weissenborn, D., Zhai, X., Unterthiner, T., Dehghani, M., Minderer, M., Heigold, G., Gelly, S., et al.: An image is worth 16x16 words: Transformers for image recognition at scale. arXiv preprint arXiv:2010.11929 (2020)" + }, + { + "type": "ref_text", + "bbox": [ + 0.218, + 0.3, + 0.787, + 0.328 + ], + "angle": 0, + "content": "17. Hao, T., Chen, H., Guo, Y., Ding, G.: Consolidator: Mergeable adapter with grouped connections for visual adaptation. arXiv preprint arXiv:2305.00603 (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.218, + 0.328, + 0.787, + 0.37 + ], + "angle": 0, + "content": "18. Hao, T., Ding, X., Feng, J., Yang, Y., Chen, H., Ding, G.: Quantized prompt for efficient generalization of vision-language models. arXiv preprint arXiv:2407.10704 (2024)" + }, + { + "type": "ref_text", + "bbox": [ + 0.218, + 0.37, + 0.787, + 0.411 + ], + "angle": 0, + "content": "19. Hao, T., Lyu, M., Chen, H., Zhao, S., Han, J., Ding, G.: Re-parameterized low-rank prompt: Generalize a vision-language model within \\(0.5\\mathrm{k}\\) parameters. arXiv preprint arXiv:2312.10813 (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.218, + 0.411, + 0.787, + 0.453 + ], + "angle": 0, + "content": "20. He, K., Chen, X., Xie, S., Li, Y., Dóllar, P., Girshick, R.: Masked autoencoders are scalable vision learners. In: Proceedings of the IEEE/CVF conference on computer vision and pattern recognition. pp. 16000-16009 (2022)" + }, + { + "type": "ref_text", + "bbox": [ + 0.218, + 0.453, + 0.787, + 0.481 + ], + "angle": 0, + "content": "21. Hinton, G., Vinyals, O., Dean, J.: Distilling the knowledge in a neural network. arXiv preprint arXiv:1503.02531 (2015)" + }, + { + "type": "ref_text", + "bbox": [ + 0.218, + 0.481, + 0.787, + 0.509 + ], + "angle": 0, + "content": "22. Hochreiter, S., Schmidhuber, J.: Long short-term memory. Neural Computation 9(8), 1735-1780 (1997)" + }, + { + "type": "ref_text", + "bbox": [ + 0.218, + 0.509, + 0.787, + 0.55 + ], + "angle": 0, + "content": "23. Houlsby, N., Giurgiu, A., Jastrzebski, S., Morrone, B., De Laroussilhe, Q., Gesmundo, A., Attariyan, M., Gelly, S.: Parameter-efficient transfer learning for nlp. In: International Conference on Machine Learning. pp. 2790-2799. PMLR (2019)" + }, + { + "type": "ref_text", + "bbox": [ + 0.218, + 0.55, + 0.787, + 0.592 + ], + "angle": 0, + "content": "24. Hu, E.J., Shen, Y., Wallis, P., Allen-Zhu, Z., Li, Y., Wang, S., Wang, L., Chen, W.: Lora: Low-rank adaptation of large language models. arXiv preprint arXiv:2106.09685 (2021)" + }, + { + "type": "ref_text", + "bbox": [ + 0.218, + 0.592, + 0.787, + 0.619 + ], + "angle": 0, + "content": "25. Huang, X., Shen, Z., Cheng, K.T.: Variation-aware vision transformer quantization. arXiv preprint arXiv:2307.00331 (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.218, + 0.619, + 0.787, + 0.661 + ], + "angle": 0, + "content": "26. Jia, M., Tang, L., Chen, B.C., Cardie, C., Belongie, S., Hariharan, B., Lim, S.N.: Visual prompt tuning. In: European Conference on Computer Vision. pp. 709-727. Springer (2022)" + }, + { + "type": "ref_text", + "bbox": [ + 0.218, + 0.661, + 0.787, + 0.702 + ], + "angle": 0, + "content": "27. Jiang, H., Wu, Q., Lin, C.Y., Yang, Y., Qiu, L.: Llmlingua: Compressing prompts for accelerated inference of large language models. arXiv preprint arXiv:2310.05736 (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.218, + 0.702, + 0.787, + 0.731 + ], + "angle": 0, + "content": "28. Jie, S., Deng, Z.H.: Convolutional bypasses are better vision transformer adapters. arXiv preprint arXiv:2207.07039 (2022)" + }, + { + "type": "ref_text", + "bbox": [ + 0.218, + 0.73, + 0.787, + 0.772 + ], + "angle": 0, + "content": "29. Kirillov, A., Mintun, E., Ravi, N., Mao, H., Rolland, C., Gustafson, L., Xiao, T., Whitehead, S., Berg, A.C., Lo, W.Y., et al.: Segment anything. arXiv preprint arXiv:2304.02643 (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.218, + 0.772, + 0.787, + 0.841 + ], + "angle": 0, + "content": "30. Kong, Z., Dong, P., Ma, X., Meng, X., Niu, W., Sun, M., Shen, X., Yuan, G., Ren, B., Tang, H., Qin, M., Wang, Y.: Spvit: Enabling faster vision transformers via latency-aware soft token pruning. In: Avidan, S., Brostow, G.J., Cissé, M., Farinella, G.M., Hassner, T. (eds.) Computer Vision - ECCV 2022 - 17th European Conference, Tel Aviv, Israel, October 23-27, 2022, Proceedings, Part XI. Lecture" + }, + { + "type": "list", + "bbox": [ + 0.218, + 0.147, + 0.787, + 0.841 + ], + "angle": 0, + "content": null + } + ], + [ + { + "type": "header", + "bbox": [ + 0.469, + 0.115, + 0.732, + 0.129 + ], + "angle": 0, + "content": "PYRA: Parallel Yielding Re-Activation" + }, + { + "type": "page_number", + "bbox": [ + 0.768, + 0.116, + 0.786, + 0.127 + ], + "angle": 0, + "content": "17" + }, + { + "type": "ref_text", + "bbox": [ + 0.241, + 0.147, + 0.786, + 0.188 + ], + "angle": 0, + "content": "Notes in Computer Science, vol. 13671, pp. 620-640. Springer (2022). https://doi.org/10.1007/978-3-031-20083-0_37, https://doi.org/10.1007/978-3-031-20083-0_37" + }, + { + "type": "ref_text", + "bbox": [ + 0.218, + 0.19, + 0.787, + 0.233 + ], + "angle": 0, + "content": "31. Kong, Z., Dong, P., Ma, X., Meng, X., Sun, M., Niu, W., Shen, X., Yuan, G., Ren, B., Qin, M., et al.: Spvit: Enabling faster vision transformers via soft token pruning. arXiv preprint arXiv:2112.13890 (2021)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.234, + 0.786, + 0.261 + ], + "angle": 0, + "content": "32. Lester, B., Al-Rfou, R., Constant, N.: The power of scale for parameter-efficient prompt tuning. arXiv preprint arXiv:2104.08691 (2021)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.262, + 0.786, + 0.289 + ], + "angle": 0, + "content": "33. Li, X.L., Liang, P.: Prefix-tuning: Optimizing continuous prompts for generation. arXiv preprint arXiv:2101.00190 (2021)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.29, + 0.787, + 0.331 + ], + "angle": 0, + "content": "34. Li, Y., Xu, S., Zhang, B., Cao, X., Gao, P., Guo, G.: Q-vit: Accurate and fully quantized low-bit vision transformer. Advances in Neural Information Processing Systems 35, 34451-34463 (2022)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.333, + 0.787, + 0.402 + ], + "angle": 0, + "content": "35. Li, Y., Adamczewski, K., Li, W., Gu, S., Timofte, R., Gool, L.V.: Revisiting random channel pruning for neural network compression. In: IEEE/CVF Conference on Computer Vision and Pattern Recognition, CVPR 2022, New Orleans, LA, USA, June 18-24, 2022. pp. 191-201. IEEE (2022). https://doi.org/10.1109/CVPR52688.2022.00029, https://doi.org/10.1109/CVPR52688.2022.00029" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.403, + 0.787, + 0.444 + ], + "angle": 0, + "content": "36. Lian, D., Zhou, D., Feng, J., Wang, X.: Scaling & shifting your features: A new baseline for efficient model tuning. Advances in Neural Information Processing Systems 35, 109-123 (2022)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.445, + 0.786, + 0.473 + ], + "angle": 0, + "content": "37. Lin, X., Zhao, C., Pan, W.: Towards accurate binary convolutional neural network. Advances in neural information processing systems 30 (2017)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.474, + 0.786, + 0.502 + ], + "angle": 0, + "content": "38. Lin, Z., Courbariaux, M., Memisevic, R., Bengio, Y.: Neural networks with few multiplications. arXiv preprint arXiv:1510.03009 (2015)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.503, + 0.787, + 0.544 + ], + "angle": 0, + "content": "39. Liu, X., Ji, K., Fu, Y., Tam, W.L., Du, Z., Yang, Z., Tang, J.: P-tuning v2: Prompt tuning can be comparable to fine-tuning universally across scales and tasks. arXiv preprint arXiv:2110.07602 (2021)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.545, + 0.787, + 0.599 + ], + "angle": 0, + "content": "40. Liu, Z., Lin, Y., Cao, Y., Hu, H., Wei, Y., Zhang, Z., Lin, S., Guo, B.: Swin transformer: Hierarchical vision transformer using shifted windows. In: Proceedings of the IEEE/CVF international conference on computer vision. pp. 10012-10022 (2021)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.601, + 0.787, + 0.642 + ], + "angle": 0, + "content": "41. Liu, Z., Wang, Y., Han, K., Zhang, W., Ma, S., Gao, W.: Post-training quantization for vision transformer. Advances in Neural Information Processing Systems 34, 28092-28103 (2021)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.643, + 0.787, + 0.685 + ], + "angle": 0, + "content": "42. Lyu, M., Yang, Y., Hong, H., Chen, H., Jin, X., He, Y., Xue, H., Han, J., Ding, G.: One-dimensional adapter to rule them all: Concepts, diffusion models and erasing applications. arXiv preprint arXiv:2312.16145 (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.686, + 0.787, + 0.713 + ], + "angle": 0, + "content": "43. Marin, D., Chang, J.H.R., Ranjan, A., Prabhu, A., Rastegari, M., Tuzel, O.: Token pooling in vision transformers. arXiv preprint arXiv:2110.03860 (2021)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.714, + 0.787, + 0.755 + ], + "angle": 0, + "content": "44. Perez, E., Strub, F., De Vries, H., Dumoulin, V., Courville, A.: Film: Visual reasoning with a general conditioning layer. In: Proceedings of the AAAI conference on artificial intelligence (2018)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.757, + 0.787, + 0.798 + ], + "angle": 0, + "content": "45. Pfeiffer, J., Kamath, A., Rücklé, A., Cho, K., Gurevych, I.: Adapterfusion: Nondestructive task composition for transfer learning. arXiv preprint arXiv:2005.00247 (2020)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.799, + 0.787, + 0.841 + ], + "angle": 0, + "content": "46. Song, Y., Zhou, Q., Li, X., Fan, D.P., Lu, X., Ma, L.: Ba-sam: Scalable bias-mode attention mask for segment anything model. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 3162-3173 (2024)" + }, + { + "type": "list", + "bbox": [ + 0.217, + 0.147, + 0.787, + 0.841 + ], + "angle": 0, + "content": null + } + ], + [ + { + "type": "page_number", + "bbox": [ + 0.218, + 0.116, + 0.236, + 0.127 + ], + "angle": 0, + "content": "18" + }, + { + "type": "header", + "bbox": [ + 0.272, + 0.115, + 0.372, + 0.128 + ], + "angle": 0, + "content": "Y. Xiong et al." + }, + { + "type": "ref_text", + "bbox": [ + 0.218, + 0.148, + 0.785, + 0.189 + ], + "angle": 0, + "content": "47. Song, Z., Jia, C., Yang, L., Wei, H., Liu, L.: Graphalign++: An accurate feature alignment by graph matching for multi-modal 3d object detection. IEEE Transactions on Circuits and Systems for Video Technology (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.218, + 0.19, + 0.787, + 0.245 + ], + "angle": 0, + "content": "48. Song, Z., Wei, H., Bai, L., Yang, L., Jia, C.: Graphalign: Enhancing accurate feature alignment by graph matching for multi-modal 3d object detection. In: Proceedings of the IEEE/CVF International Conference on Computer Vision. pp. 3358-3369 (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.246, + 0.785, + 0.286 + ], + "angle": 0, + "content": "49. Strudel, R., Garcia, R., Laptev, I., Schmid, C.: Segmenter: Transformer for semantic segmentation. In: Proceedings of the IEEE/CVF international conference on computer vision. pp. 7262-7272 (2021)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.287, + 0.785, + 0.314 + ], + "angle": 0, + "content": "50. Tian, L., Ye, M., Zhou, L., He, Q.: Clip-guided black-box domain adaptation of image classification. Signal, Image and Video Processing 18(5), 4637-4646 (2024)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.315, + 0.785, + 0.357 + ], + "angle": 0, + "content": "51. Touvron, H., Cord, M., Douze, M., Massa, F., Sablayrolles, A., Jégou, H.: Training data-efficient image transformers & distillation through attention. In: International conference on machine learning. pp. 10347-10357. PMLR (2021)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.357, + 0.785, + 0.384 + ], + "angle": 0, + "content": "52. Tung, F., Mori, G.: Similarity-preserving knowledge distillation. In: Proceedings of the IEEE/CVF international conference on computer vision. pp. 1365-1374 (2019)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.385, + 0.785, + 0.411 + ], + "angle": 0, + "content": "53. Wang, A., Chen, H., Lin, Z., Pu, H., Ding, G.: Repvit: Revisiting mobile cnn from vit perspective. arXiv preprint arXiv:2307.09283 (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.412, + 0.785, + 0.452 + ], + "angle": 0, + "content": "54. Wang, A., Chen, H., Lin, Z., Zhao, S., Han, J., Ding, G.: Cait: Triple-win compression towards high accuracy, fast inference, and favorable transferability for vits. arXiv preprint arXiv:2309.15755 (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.453, + 0.785, + 0.508 + ], + "angle": 0, + "content": "55. Wang, W., Xie, E., Li, X., Fan, D.P., Song, K., Liang, D., Lu, T., Luo, P., Shao, L.: Pyramid vision transformer: A versatile backbone for dense prediction without convolutions. In: Proceedings of the IEEE/CVF international conference on computer vision. pp. 568-578 (2021)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.509, + 0.785, + 0.549 + ], + "angle": 0, + "content": "56. Wang, Z., Luo, H., Wang, P., Ding, F., Wang, F., Li, H.: Vtc-lfc: Vision transformer compression with low-frequency components. Advances in Neural Information Processing Systems 35, 13974-13988 (2022)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.55, + 0.785, + 0.605 + ], + "angle": 0, + "content": "57. Wei, K., Du, R., Jin, L., Liu, J., Yin, J., Zhang, L., Liu, J., Liu, N., Zhang, J., Guo, Z.: Video event extraction with multi-view interaction knowledge distillation. In: Proceedings of the AAAI Conference on Artificial Intelligence. vol. 38, pp. 19224-19233 (2024)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.606, + 0.785, + 0.66 + ], + "angle": 0, + "content": "58. Wei, K., Jin, L., Zhang, Z., Guo, Z., Li, X., Liu, Q., Feng, W.: More than syntaxes: Investigating semantics to zero-shot cross-lingual relation extraction and event argument role labelling. ACM Transactions on Asian and Low-Resource Language Information Processing 23(5), 1-21 (2024)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.661, + 0.785, + 0.688 + ], + "angle": 0, + "content": "59. Wu, Y., He, K.: Group normalization. In: Proceedings of the European conference on computer vision (ECCV). pp. 3-19 (2018)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.689, + 0.785, + 0.73 + ], + "angle": 0, + "content": "60. Xiong, Y., Chen, H., Lin, Z., Zhao, S., Ding, G.: Confidence-based visual dispersal for few-shot unsupervised domain adaptation. In: Proceedings of the IEEE/CVF International Conference on Computer Vision. pp. 11621-11631 (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.731, + 0.785, + 0.757 + ], + "angle": 0, + "content": "61. Xiong, Y., Chen, X., Ye, X., Chen, H., Lin, Z., Lian, H., Niu, J., Ding, G.: Temporal scaling law for large language models. arXiv preprint arXiv:2404.17785 (2024)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.758, + 0.785, + 0.799 + ], + "angle": 0, + "content": "62. Xu, L., Xie, H., Qin, S.Z.J., Tao, X., Wang, F.L.: Parameter-efficient fine-tuning methods for pretrained language models: A critical review and assessment. arXiv preprint arXiv:2312.12148 (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.8, + 0.785, + 0.84 + ], + "angle": 0, + "content": "63. Yao, Z., Yazdani Aminabadi, R., Zhang, M., Wu, X., Li, C., He, Y.: Zeroquant: Efficient and affordable post-training quantization for large-scale transformers. Advances in Neural Information Processing Systems 35, 27168-27183 (2022)" + }, + { + "type": "list", + "bbox": [ + 0.217, + 0.148, + 0.787, + 0.84 + ], + "angle": 0, + "content": null + } + ], + [ + { + "type": "header", + "bbox": [ + 0.469, + 0.115, + 0.732, + 0.129 + ], + "angle": 0, + "content": "PYRA: Parallel Yielding Re-Activation" + }, + { + "type": "page_number", + "bbox": [ + 0.769, + 0.116, + 0.786, + 0.127 + ], + "angle": 0, + "content": "19" + }, + { + "type": "ref_text", + "bbox": [ + 0.214, + 0.147, + 0.788, + 0.175 + ], + "angle": 0, + "content": "64. Yu, S., Chen, T., Shen, J., Yuan, H., Tan, J., Yang, S., Liu, J., Wang, Z.: Unified visual transformer compression. arXiv preprint arXiv:2203.08243 (2022)" + }, + { + "type": "ref_text", + "bbox": [ + 0.214, + 0.175, + 0.788, + 0.217 + ], + "angle": 0, + "content": "65. Zaken, E.B., Ravfogel, S., Goldberg, Y.: Bitfit: Simple parameter-efficient fine-tuning for transformer-based masked language-models. arXiv preprint arXiv:2106.10199 (2021)" + }, + { + "type": "ref_text", + "bbox": [ + 0.216, + 0.218, + 0.787, + 0.259 + ], + "angle": 0, + "content": "66. Zhai, X., Kolesnikov, A., Houlsby, N., Beyer, L.: Scaling vision transformers. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 12104-12113 (2022)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.259, + 0.787, + 0.313 + ], + "angle": 0, + "content": "67. Zhai, X., Puigcerver, J., Kolesnikov, A., Ruyssen, P., Riquelme, C., Lucic, M., Djolonga, J., Pinto, A.S., Neumann, M., Dosovitskiy, A., et al.: A large-scale study of representation learning with the visual task adaptation benchmark. arXiv preprint arXiv:1910.04867 (2019)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.314, + 0.787, + 0.355 + ], + "angle": 0, + "content": "68. Zhang, Y., Qiu, L., Zhu, Y., Wen, L., Luo, X.: A new childhood pneumonia diagnosis method based on fine-grained convolutional neural network. Computer Modeling in Engineering & Sciences 133, 873-894 (2022)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.356, + 0.787, + 0.383 + ], + "angle": 0, + "content": "69. Zhang, Y., Zhou, K., Liu, Z.: Neural prompt search. arXiv preprint arXiv:2206.04673 (2022)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.383, + 0.787, + 0.438 + ], + "angle": 0, + "content": "70. Zheng, S., Lu, J., Zhao, H., Zhu, X., Luo, Z., Wang, Y., Fu, Y., Feng, J., Xiang, T., Torr, P.H., et al.: Rethinking semantic segmentation from a sequence-to-sequence perspective with transformers. In: Proceedings of the IEEE/CVF conference on computer vision and pattern recognition. pp. 6881-6890 (2021)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.439, + 0.787, + 0.481 + ], + "angle": 0, + "content": "71. Zhou, F., Zhou, Q., Yin, B., Zheng, H., Lu, X., Ma, L., Ling, H.: Rethinking impersonation and dodging attacks on face recognition systems. In: Proceedings of the 30th ACM International Conference on Multimedia (ACM MM) (2024)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.481, + 0.787, + 0.522 + ], + "angle": 0, + "content": "72. Zhou, Q., Feng, Z., Gu, Q., Pang, J., Cheng, G., Lu, X., Shi, J., Ma, L.: Context-aware mixup for domain adaptive semantic segmentation. IEEE Transactions on Circuits and Systems for Video Technology 33(2), 804-817 (2022)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.522, + 0.787, + 0.549 + ], + "angle": 0, + "content": "73. Zhu, M., Tang, Y., Han, K.: Vision transformer pruning. arXiv preprint arXiv:2104.08500 (2021)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.549, + 0.787, + 0.591 + ], + "angle": 0, + "content": "74. Zhu, X., Su, W., Lu, L., Li, B., Wang, X., Dai, J.: Deformable detr: Deformable transformers for end-to-end object detection. arXiv preprint arXiv:2010.04159 (2020)" + }, + { + "type": "list", + "bbox": [ + 0.214, + 0.147, + 0.788, + 0.591 + ], + "angle": 0, + "content": null + } + ] +] \ No newline at end of file diff --git a/2024/PYRA_ Parallel Yielding Re-Activation for Training-Inference Efficient Task Adaptation/a96361d9-106b-4325-b9d1-39f6bd731ebb_origin.pdf b/2024/PYRA_ Parallel Yielding Re-Activation for Training-Inference Efficient Task Adaptation/a96361d9-106b-4325-b9d1-39f6bd731ebb_origin.pdf new file mode 100644 index 0000000000000000000000000000000000000000..474fe96b3399a75df4d02ecf5b68b6df5e552525 --- /dev/null +++ b/2024/PYRA_ Parallel Yielding Re-Activation for Training-Inference Efficient Task Adaptation/a96361d9-106b-4325-b9d1-39f6bd731ebb_origin.pdf @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:af24ceaf00d0452fffaba13fe70d2a11ecf9be5c73225e93956c2776c5a8c7e5 +size 1096235 diff --git a/2024/PYRA_ Parallel Yielding Re-Activation for Training-Inference Efficient Task Adaptation/full.md b/2024/PYRA_ Parallel Yielding Re-Activation for Training-Inference Efficient Task Adaptation/full.md new file mode 100644 index 0000000000000000000000000000000000000000..63a7ff418636a482761b19d58795de40e968a54b --- /dev/null +++ b/2024/PYRA_ Parallel Yielding Re-Activation for Training-Inference Efficient Task Adaptation/full.md @@ -0,0 +1,333 @@ +# PYRA: Parallel Yielding Re-Activation for Training-Inference Efficient Task Adaptation + +Yizhe Xiong $^{1,2}$ , Hui Chen $^{2*}$ , Tianxiang Hao $^{1,2}$ , Zijia Lin $^{1}$ , Jungong Han $^{2,3}$ , Yuesong Zhang $^{4}$ , Guoxin Wang $^{4}$ , Yongjun Bao $^{4}$ , and Guiguang Ding $^{1,2}$ + +$^{1}$ School of Software, Tsinghua University, Beijing, China + +$^{2}$ BNRist, Tsinghua University, Beijing, China + +$^{3}$ Department of Automation, Tsinghua University, Beijing, China + +4 JD.com, Beijing, China + +{xiongyizhe2001,beyondtx,jungonghan77}@gmail.com + +huichen@mail.tsinghua.edu.cn + +{zhangyuesong1,wangguoxin14,baoyongjun}@jd.com + +linzijia07@tsinghua.org.cn dinggg@tsinghua.edu.cn + +Abstract. Recently, the scale of transformers has grown rapidly, which introduces considerable challenges in terms of training overhead and inference efficiency in the scope of task adaptation. Existing works, namely Parameter-Efficient Fine-Tuning (PEFT) and model compression, have separately investigated the challenges. However, PEFT cannot guarantee the inference efficiency of the original backbone, especially for large-scale models. Model compression requires significant training costs for structure searching and re-training. Consequently, a simple combination of them cannot guarantee accomplishing both training efficiency and inference efficiency with minimal costs. In this paper, we propose a novel Parallel Yielding Re-Activation (PYRA) method for such a challenge of training-inference efficient task adaptation. PYRA first utilizes parallel yielding adaptive weights to comprehensively perceive the data distribution in downstream tasks. A re-activation strategy for token modulation is then applied for tokens to be merged, leading to calibrated token features. Extensive experiments demonstrate that PYRA outperforms all competing methods under both low compression rate and high compression rate, demonstrating its effectiveness and superiority in maintaining both training efficiency and inference efficiency for large-scale foundation models. Our code is available at https://github.com/THU-MIG/PyRA. + +Keywords: Vision Transformer $\cdot$ Task Adaptation $\cdot$ Model Compression + +# 1 Introduction + +Vision transformers [16] have made a profound impact across various domains of computer vision, such as image classification [6,15,16,50,55,66], object detection [4,40,47,48,74], image segmentation [11,49,53,70,72], etc. In recent years, the + +![](images/5664647c2b736e0e82a403b7c620536892af767f6bc456d67ad6a3f18ccce171.jpg) +(a) Full Fine-tuning + +![](images/2ab152f60759215e24818de68dbcf2edb03dee47b612c65aa58460f983ac19a7.jpg) +(c) PEFT + +![](images/36b7ea242bb2fc69de46fb442daa7344f5f931b70f4eefd2392f17f1de845c7d.jpg) +(b) Model compression + +![](images/e939f1d5fbe6cc22b2fc2ddb8383ad38f11f5c6b2de37f5afaa3082c0f1cf27b.jpg) +(d) Ours +Fig. 1: (a) Full fine-tuning trains all parameters on downstream tasks and utilizes the trained model for inference, thereby lacking efficiency in both training and inference stages. (b) Model compression employs pruning to enhance inference efficiency, but the pruned model necessitates extensive re-training on large-scale data. (c) PEFT freezes the model backbone and only fine-tunes a small amount of parameters, yet retains the inference complexity. (d) Our training-inference efficient task adaptation incorporates the advantages of all existing pipelines by training inference-efficient models with minimal tunable parameters. + +
Method(a)(b)(c)(d)
Non-vast Data Scale✔×✔✔
Training Efficient××✔✔
Inference Efficient×✔×✔
+ +scale of vision transformers has grown to be billion-parameters [29,61,66]. Consequently, adapting such models with vast scales into downstream tasks presents increasingly complex challenges, particularly in real-world deployment scenarios. Two critical concerns have been widely acknowledged as primary obstacles in implementing large-scale transformers for downstream applications [17,23,27,46]: (1) the training overhead when fine-tuning on downstream tasks, and (2) the inference efficiency after model deployment. + +Specifically, first, conventional fine-tuning methods, which necessitate adjusting all parameters of the model (i.e., Full Fine-tuning), suffer from unaffordable consumption of GPU resources and training time given the extensive scales of the foundation models [16, 29, 66]. Researchers have delved into Parameter-Efficient Fine-Tuning (PEFT) [17, 19, 23, 24, 26, 36] algorithms, which generally freeze the pre-trained models and only tune extra small parameters, leading to great reduction of training time and storage overhead. The second issue pertains to inference efficiency, requiring the deployed model to promptly process the input data. The computational complexity of the models significantly influences achieving satisfactory inference throughput. Representative solutions for this matter encompass model compression methods, including model pruning [7, 10, 56, 64], knowledge distillation [1, 21, 52, 57, 58], model quantization [3, 13, 37], etc. + +In the literature, these two intriguing topics are investigated separately. PEFT methods either identify a subset of tunable parameters in the backbone for fine-tuning [65] or introduce learnable parameters to the frozen backbone during fine-tuning [17, 23, 24, 26, 36]. While effectively reducing training costs, most of these methods inevitably escalate computational complexity, resulting in inefficient inference. Model compression methods frequently require signifi + +cant computational resources to identify optimal structures for pruning. After pruning, a comprehensive re-training process using a substantial amount of data is crucial to prevent significant performance degradation. Therefore, model compression methods are typically inefficient in training efficiency. + +These observations naturally lead to the question: can we achieve both training efficiency and inference efficiency simultaneously for downstream tasks? We refer to this challenge as Training-Inference Efficient Task Adaptation (seeing Fig. 1). Exploring this issue can enable us to conveniently deploy the advanced large-scale foundation models in real-world downstream applications with minimal costs, which is appealing and essential for the widespread implementation of foundation models. A straightforward solution is combining PEFT and model compression. However, for efficient task adaptation, a + +![](images/d910bd2d0f40de2b7791e9e805efde4c4695c9c3c911f9b850789508a6ca9060.jpg) +Fig.2: Comparisons between simply combining ToMe [2] and LoRA [24] and our proposed PYRA. Red boxes represent the performance drop problem in low compression rates. Green boxes represent the adverse compression in high compression rates. See Sec. 4 for more results. + +heavy re-training stage is unaffordable. Consequently, simply combining PEFT and model compression can easily suffer from substantial performance drops. For instance, we could integrate LoRA [24], a notable PEFT method that introduces a small low-rank adapter, with ToMe [2], a parameter-free model compression technique for vision transformers5. As shown in Fig. 2, under lower compression rates (around $1.7 \times$ ), performance on both backbones (ViT-L/16 and ViT-B/16) present slight drops ( $< 1\%$ ) compared to directly conducting PEFT on the backbone, indicating that the ToMe+LoRA combination serves as a basic solution within lower compression rate range, yet performance improvements are also demanded. Under high compression rates ( $>3.0 \times$ ), the performance quickly drops and is even inferior to directly fine-tuning the small-scale backbone with corresponding throughput. We term this phenomenon as Adverse Compression. Both phenomena indicate that directly combining existing works cannot effectively address the new challenges. + +In this paper, we propose a novel Parallel Yielding Re-Accivitation method (PYRA) designed for training-inference efficient task adaptation using vision transformers. Generally, the proposed PYRA follows the token merging paradigm [2, 8, 43] for inference efficiency. For effective task adaptation, we propose to mod- + +ulate token features from both feature tokens and feature channels and parallelly yield weights for tokens to be merged by small modulation weight generators. These parallel yielding weights can comprehensively perceive the data distribution in downstream tasks. They are then applied to features through re-activation, resulting in adaptive token modulation. Thanks to such a token modulation strategy, PYRA can adaptively calibrate the learned feature distribution for downstream tasks with low computational complexity, ultimately leading to effective training and efficient inference. + +We conduct extensive experiments to verify the effectiveness of our PYRA. We show that, under a low compression rate with $\sim 1.7\times$ speedup, PYRA introduces negligible performance drops. Under a high compression rate with $>3.0\times$ speedup, PYRA eliminates the adverse compression gap. We emphasize that considering the scenario of high compression rates is practical, given the substantial size of current transformers. Concurrently, the specific model sizes required by downstream applications may not match any publicly available models, thus requiring the acquisition through high compression rates while keeping the performance comparable. In this regard, our approach represents an effective method for obtaining small models in the absence of pre-trained parameters for smaller-scale models. + +Overall, we summarize our contribution as follows. + +- We propose a novel challenge termed training-inference efficient task adaptation, in which the inference efficiency of large-scale transformers is escalated during parameter-efficient task adaptation. +- We propose PYRA for training-inference efficient task adaptation, which enhances the perception of feature distribution via token modulation. We generate parallel yielding decoupled weights to comprehensively perceive the feature distributions. We apply a re-activation strategy to modulate tokens to be merged for calibrated token features. +- Extensive experiments show that our PYRA outperforms all competing methods under both low compression rate and high compression rate. Further analysis shows that PYRA is effective across a series of different transformer backbones and model scales, well demonstrating the effectiveness and superiority of our method. + +# 2 Related Works + +Parameter-Efficient Fine-Tuning (PEFT) for Task Adaptation. Transferring large-scale transformers to downstream tasks has been a popular topic in computer vision [9, 17, 26, 28, 36, 60, 68, 69]. PEFT methods either locate a subset of parameters inside the model for fine-tuning [65], or inject human designed modules to the original model structure. Specifically, adapters [23, 28, 42, 45] are a type of MLP module with a bottleneck in the middle. Prompt-tuning [14, 18, 26, 32, 33, 39] insert learnable tokens to model input to generate task-specific outputs. SSF [36] tunes additional scaling and shifting parameters. LoRA [24], AdaptFormer [9], and Consolidator [17] add lightweight modules as + +bypasses. These modules can be merged with the original backbone for no extra inference cost. In the scope of training-inference efficient task adaptation, PEFT methods retain model inference cost, resulting in extreme difficulty for model deployment. To solve the problem, we conduct adaptive token merging for PEFT. While achieving promising adaptation performance under both low and high compression rates, our method inherits the advantages of PEFT methods. Model Compression. Model compression is often applied on large-scale models to acquire smaller-scale models with comparable performance. Mainstream approaches of model compression include model pruning [2,8,10,54,56,64,71], knowledge distillation [1,21,52] and model quantization [3,13,37,38]. For transformer models, model pruning methods can be roughly grouped into two categories: channel pruning and token pruning. Channel pruning methods [5,10, 64,73] reduce the number of parameters, channels, heads, or blocks. Recently, token pruning has emerged as another mainstream approach. Several works have attempted to prune tokens for vision transformers (ViTs). Among these methods, token pooling [43] uses a slow k-means approach that does not work for an off-the-shelf model. ToMe [2] constructs bipartite graphs and merges token pairs with the most weighted connections. DiffRate [8] combines token pruning and token merging with searched optimal token reduction rates for each layer. Although achieving promising results, most existing model compression methods fail when combining with PEFT for training-inference efficient task adaptation since they usually involve a heavy training stage. Model pruning methods demand full re-training after pruning to restore performance. Knowledge distillation necessitates training a small-scale model from scratch, demanding large amount of data. As for model quantization, although no heavy training is demanded, post-training quantization [41,63] achieves poor performance compared to quantization-aware training [25,34] that demands full re-training on the quantized model. As a comparison, our proposed PYRA achieves promising performance on compressed models under the restrictions of PEFT, surpassing the baselines of simply combining model compression and PEFT. + +# 3 Methodology + +# 3.1 Preliminaries + +ViT Model. In this paper, we mainly focus on the training-inference efficient task adaptation of ViT models [16, 20, 51]. A ViT model consists of $L$ identical encoder blocks, each of which consists of a multi-head self-attention (MHSA) module and a feed-forward network (FFN). Formally, an input image $\mathbf{x}$ is reshaped and linear projected to $N$ tokens $\mathbf{x} = [t_1, t_2, \dots, t_N]$ in $D$ dimensions. For simplicity, we omit the classification token ([CLS]) and the distillation token [51]. For encoder block $l$ , we denote the input as $\mathbf{x}^l = [t_1^l, t_2^l, \dots, t_{N^{l-1}}^l] \in \mathbb{R}^{N^{l-1} \times D}$ and the output as $\hat{\mathbf{x}}^l = [\hat{t}_1^l, \hat{t}_2^l, \dots, \hat{t}_{N^l}^l] \in \mathbb{R}^{N^l \times D}$ . For MHSA, the input tokens are first processed by three FC layers to generate $Q, K$ , and $V$ matrices, and the output is calculated by Softmax $(\frac{QK^T}{\sqrt{D}}) V$ before being projected by another FC + +layer. For FFN, the tokens are projected by two FC layers. Our method mainly focuses on the input tokens $\mathbf{x}^l$ before feeding them to the MHSA module. + +LoRA. LoRA [24] is a widely employed PEFT method for task adaptation. Vision transformers consist of large dense parameter matrices. When adapting to a specific task, the updates to the matrices are in small subspaces and can be modeled with low-rank decompositions. LoRA trains only the decomposed matrices during fine-tuning. Specifically, for dense matrix $W_0 \in \mathbb{R}^{d \times k}$ and input $x$ , the modified forward pass of updated $W_0$ is: + +$$ +\hat {x} = W _ {0} x + B A x, \tag {1} +$$ + +where $B \in \mathbb{R}^{d \times h}$ and $A \in \mathbb{R}^{h \times k}$ . During inference, $BA$ can be merged with $W_0$ for no extra computation overhead. + +Token Merging. Token merging [2,8] is a parameter-free compression technique for vision transformers. It is orthogonal to the transformer structure and capable of flexibly changing the compression rate. Specifically, the input tokens $\mathbf{x}^l$ of the $l$ -th ViT block are randomly separated before the MHSA: + +$$ +G _ {1} ^ {l} = \left[ t _ {i _ {1}} ^ {l}, t _ {i _ {2}} ^ {l}, \dots , t _ {i _ {s}} ^ {l} \right], \quad G _ {2} ^ {l} = \left[ t _ {j _ {1}} ^ {l}, t _ {j _ {2}} ^ {l}, \dots , t _ {j _ {s}} ^ {l} \right], \quad 2 s = N ^ {l - 1}. \tag {2} +$$ + +Then, for each $t_{i_\bullet}^l$ token, the most similar $t_{j_\bullet}^l$ token is matched to it via cosine similarity. From $G_1^l$ , $r$ tokens with the most similar connections to tokens in $G_2^l$ are selected to form token pairs $(t_{m_k}^l, t_{n_k}^l)$ , where $t_{m_k}^l \in G_1^l$ , $t_{n_k}^l \in G_2^l$ , and $k = 1, 2, \dots, r$ . Note that $\{m_k\}$ is a re-indexed subset of $\{i_\bullet\}$ . Formally, + +$$ +\left\{t _ {m _ {k}} ^ {l} \right\} = \arg \operatorname {T o p r} _ {\hat {t} _ {i} ^ {l}} \left(\max _ {j} \frac {\hat {t} _ {i} ^ {l} \cdot t _ {j} ^ {l}}{\| \hat {t} _ {i} ^ {l} \| \cdot \| t _ {j} ^ {l} \|}\right), \quad t _ {n _ {k}} ^ {l} = \arg \max _ {t _ {j} ^ {l}} \left(\frac {t _ {m _ {k}} ^ {l} \cdot t _ {j} ^ {l}}{\| t _ {m _ {k}} ^ {l} \| \cdot \| t _ {j} ^ {l} \|}\right) \tag {3} +$$ + +Previous works [2, 8, 43], generally merge tokens via average pooling, cutting down the computational cost by decreasing the token number by $r$ in layer $l$ . + +We employ the above techniques as our baseline method, in which we attach token merging while fine-tuning LoRA for task adaptation. These methods are selected due to their advantages being highly compatible with training-inference efficient task adaptation. First, token merging is parameter-free and training-free, and does not change model structures, which retains the storage-efficient advantage of PEFT. Besides, we choose LoRA for its popularity, simplicity, and mergeability (not introducing extra FLOPs during inference). Experimental results show that $\mathrm{ToMe + LoRA}$ is a strong baseline (see Sec. 4). + +# 3.2 PYRA: Parallel Yielding Re-Activation + +Conventional fine-tuning methods excel at guiding the model to dynamically align with the target data distribution in downstream tasks by extensively adjusting parameters. However, in the context of training-inference efficient task adaptation, only a fraction of parameters can be fine-tuned, posing significant challenges in accurately capturing the nuances of data distribution. While reducing model complexity through token merging shows promise in enhancing + +![](images/9cfda996373e5beb11405400e7f727dd2c70b99c443e8b30987a29f36a116397.jpg) +Fig. 3: The pipeline of our PYRA. PYRA conducts token modulation before the MHSA module in each transformer block. Inside PYRA, a pair of learnable modulation weight generators are leveraged to generate adaptive modulation weights parallelly. After that, generated weights modulate tokens through re-activation. The generators in PYRA can be trained along with the LoRA module $\theta_{p}$ in an end-to-end manner. + +inference efficiency, it introduces the risk of information loss during layer-wise processing in vision transformers. This loss is difficult to rectify due to the constrained understanding of data distribution in the efficient task adaptation scenario. Therefore, a straightforward combination of token merging with PEFT algorithms for achieving training-inference efficient task adaptation may not yield optimal results, as depicted in Fig. 2. + +Here, we propose Parallel Yielding Re-Aviation (PYRA) to adaptively modulate token features to enhance the perception of data distribution during token merging. Specifically, inside PYRA, the weights for adaptive merging are first yielded in a parallel manner through a pair of lightweight learnable vectors in each ViT block. These generated weights are then applied to modulate tokens to be merged through re-activation. As a result, PYRA enables adaptive calibration of the learned feature distribution with low computational complexity. + +Parallel Yielding Adaptive Weights. We aim to optimize the merging process of each chosen token pair. Inspired by feature modulation [44, 59], we propose to modulate token features before merging. Formally, for encoder block $l$ with $r$ pairs of tokens in $D$ dimensions to be merged, we group the $t_{m_k}^l$ and $t_{n_k}^l$ tokens as token matrices $M_s^l = [t_{m_1}^l,\dots ,t_{m_r}^l ]$ and $M_t^l = [t_{n_1}^l,\dots ,t_{n_r}^l ]$ , where $M_s^l,M_t^l\in \mathbb{R}^{D\times r}$ . We learn a modulation matrix $W^{l}\in \mathbb{R}^{D\times r}$ that adaptively modulates tokens at the granularity of each channel. We emphasize that directly learning $W^{l}$ is redundant [24] and cannot adaptively satisfy the conditions with different images and token pairs. Therefore, we further specify the goal as learning decoupled weights $\delta_D^l = f_D(M_s^l,M_t^l)\in \mathbb{R}^{D\times 1}$ for feature channels and $\delta_r^l = f_r(M_s^l,M_t^l)\in \mathbb{R}^{1\times r}$ for feature tokens, where $W^{l} = \delta_{D}^{l}\delta_{r}^{l}$ , separately. We propose to generate $\delta_D^l$ and $\delta_r^l$ in a parallel yielding manner. + +Specifically, for layer $l$ with $r$ tokens to be merged, we create two learnable vectors as the modulation weight generator inside the transformer block: $W_r^l \in \mathbb{R}^{r \times 1}$ and $W_D^l \in \mathbb{R}^{1 \times D}$ . To guarantee that $W_r^l$ and $W_D^l$ digest the token features from both tokens in a token pair, we first normalize the sum of token pairs to + +construct the token information matrix: + +$$ +M _ {\text {i n f o}} ^ {l} = \operatorname {L a y e r N o r m} \left(M _ {s} ^ {l} + M _ {t} ^ {l}\right) \in \mathbb {R} ^ {D \times r}. \tag {4} +$$ + +We normalize the distribution of $M_{\mathrm{info}}^{l}$ tokens by leveraging the LayerNorm(\cdot) operation to enable smoother gradients when training $W_{r}^{l}$ and $W_{D}^{l}$ . With the token information matrix, we then yield the adaptive weights $\delta_{D}^{l}$ and $\delta_r^l$ parallelly: + +$$ +\delta_ {D} ^ {l} = M _ {\text {i n f o}} ^ {l} W _ {r} ^ {l} \in \mathbb {R} ^ {D \times 1} \tag {5} +$$ + +$$ +\delta_ {r} ^ {l} = W _ {D} ^ {l} M _ {\mathrm {i n f o}} ^ {l} \in \mathbb {R} ^ {1 \times r}. +$$ + +Re-Activation for Token Modulation. Simply generating $\delta_D^l$ and $\delta_r^l$ via matrix multiplication still faces several possible issues. First, no measures have taken to ensure that $\delta_D^l$ and $\delta_r^l$ stay in a normal range. Second, decoupling weights to feature tokens and feature channels results in a low-rank modulation weight matrix $W^l$ , which exhibits limited expressive capacity and thus could be unable to optimally modulate tokens in complicated data distributions. To cope with these issues, we conduct token modulation in a re-activation strategy. Specifically, we first broadcast $\delta_D^l$ to $\hat{\delta}_D^l \in \mathbb{R}^{D \times r}$ and conduct sigmoid activation $\sigma(\cdot)$ on it, and then modulate $M_s^l$ for an intermediate modulation result: + +$$ +\hat {M} _ {s} ^ {l} = 2 \sigma \left(\hat {\delta} _ {D} ^ {l}\right) \odot M _ {s} ^ {l}. \tag {6} +$$ + +where $\odot$ is Hadamard product. $\hat{M}_s^l$ is then modulated with sigmoid-activated broadcast weight $\hat{\delta}_r^l\in \mathbb{R}^{D\times r}$ again to acquire the modulated tokens: + +$$ +M _ {s} ^ {l} \leftarrow M _ {s} ^ {l} + \left(2 \sigma \left(\hat {\delta} _ {r} ^ {l}\right) - 1\right) \odot \hat {M} _ {s} ^ {l}, \tag {7} +$$ + +Note that we use the original tokens $M_{s}^{l}$ to create a residual connection that preserves the gradient flow during training. The modulated $M_{s}^{l}$ tokens are then merged with $M_{t}^{l}$ with average pooling. We use a random Gaussian initialization for generator $W_{r}^{l}$ and zero for generator $W_{D}^{l}$ , so re-activation is equivalent to identity transformation at the beginning of training. Token modulation is conducted only on $M_{s}^{l}$ to guarantee the parallelism in training and inference, as $t_{m_k}^l$ tokens are distinct while different $t_{n_k}^l$ might point to the same $t_j^l$ token. + +Discussion. Our PYRA effectively enhances the calibration of features for downstream tasks. Specifically, by utilizing parallel yielding adaptive weights, PYRA effectively decouples the modulation weight from the feature token and the feature channel, i.e., $\delta_{D}$ and $\delta_{r}$ , enabling comprehensive perception of the feature distribution in downstream tasks. Furthermore, considering the challenge of capturing an accurate feature distribution for parameter-efficient training, such smooth re-activations by Eq. (6) and Eq. (7) can constrain the negative impact on weight values brought by limited perception of feature distributions, thus leading to better token modulation for improving feature representations in downstream tasks. As a result, the proposed PYRA can maintain discriminative information to an utmost degree during token merging, leading to improved performance while achieving complexity reduction. + +Table 1: The complexity comparisons between conducting PEFT with and without PYRA. The FLOPs metric is obtained during inference. + +
ModelMetricTotalPEFT w/o PYRA (%)PEFT w. PYRA (%)
ViT-Base# params86M0.29M0.34%0.30M0.35%
FLOPs16.37G16.37G100%8.15G49.79%
ViT-Large# params303M1.18M0.39%1.20M0.40%
FLOPs57.37G57.37G100%28.76G50.13%
+ +# 3.3 Complexity Analysis + +We present the parameter and computation complexity for introducing PYRA to PEFT for task adaptation. In each ViT block, PYRA introduces $D + r$ training parameters and $4rD$ extra FLOPs. For a ViT model with $L$ layers and $R$ total merged tokens, in total, our PYRA introduces $LD + R$ extra training parameters and conducts $4RD$ extra FLOPs beside PEFT. + +To better demonstrate the training-inference efficiency of our PYRA, we compare the complexity between attaching PYRA to task adaptation via PEFT (here we employ LoRA [24]) and plainly conducting PEFT without PYRA. As shown in Tab. 1, PYRA keeps the training efficient feature of PEFT by introducing only a tiny amount of training parameters, which however, results in a substantial efficiency boost for inference, i.e., at around $50\%$ for both ViT-B and ViT-L models. Results in Sec. 4 show that PYRA achieves comparable performance to the PEFT counterpart without PYRA. This well indicates that PYRA is an effective method for training-inference efficient task adaptation. + +# 4 Experiments + +# 4.1 Experimental Setting + +Datasets. We conduct extensive experiments on task adaptation benchmarks to verify the effectiveness of PYRA within the challenge of training-inference efficient task adaptation. Specifically, we choose the VTAB-1k [67] benchmark for the evaluation. VTAB-1k is a challenging benchmark that consists of 19 different tasks from diverse domains: 1) natural images captured in the actual world; 2) specialized images from professional fields; and 3) structured synthesized images. Each task only contains 800 training samples and 200 validation samples. + +Models. We choose two ViT backbones pre-trained on the ImageNet-21K [16], i.e., ViT-L/16 and ViT-B/16, for comparison. Additionally, we also generalize our PYRA to different backbones and pre-train methods, including DeiT-B [51] and ViT backbones pre-trained by MAE [20]. + +Implementation Details. We choose LoRA [24] as the PEFT module for all methods due to its simplicity and mergeability. For ease of explanation, we omit the "+"LoRA" when comparing different methods. We append LoRA only on the $Q,K$ , and $V$ projection matrices, and apply the training schedule for LoRA following [17, 69]. The generators are trained along with LoRA + +Table 2: Results on the VTAB-1k [67] benchmark under low compression rate. Bold and underline denote the best and second-best accuracy within compression methods. + +
Method# paramsThroughputNaturalSpecializedStructured
Cifar100Caltech101DTDFlowers102PetsSVHNSun397CamelyonEuroSATResisc45RetinopathyClevr-CountClevr-DistDMLABKITTI-DistdSpr-LocdSpr-OrisNORB-AzimsNORB-EleAverage
Model: ViT-B/16 (Throughput: 425)
PEFT0.34%42567.190.269.499.190.585.754.183.195.884.374.682.269.250.179.281.847.131.142.674.76
RaP3.43%65425.968.453.364.057.471.321.575.887.959.373.643.153.826.360.573.525.516.727.955.57
SPViT4.46%56741.675.461.183.266.256.128.379.394.273.373.670.661.542.467.875.450.528.931.364.16
DiffRate0.35%70937.184.663.796.786.232.648.278.985.867.073.732.929.834.155.712.616.013.121.555.82
ToMe0.34%75364.690.467.998.589.883.953.282.694.783.574.981.969.849.276.981.946.531.043.174.10
PYRA0.35%74567.590.369.398.990.084.653.183.395.783.375.282.668.950.880.081.845.832.242.874.69
Model: ViT-L/16 (Throughput: 130)
PEFT0.39%13077.191.473.499.591.389.657.685.996.187.376.183.163.050.782.181.753.532.236.676.52
RaP1.95%19643.287.962.652.881.786.734.778.492.473.373.668.059.646.982.475.543.624.525.765.64
SPViT2.47%18848.187.565.294.477.480.938.879.993.979.874.378.265.847.474.182.350.331.037.970.22
DiffRate0.39%22150.986.870.397.888.339.052.380.287.272.274.232.632.336.557.422.826.615.223.459.53
ToMe0.39%22776.191.172.399.291.789.256.486.495.186.675.182.461.950.981.481.653.533.436.876.11
PYRA0.40%22576.691.373.299.391.589.457.186.995.987.176.283.263.252.883.182.552.634.839.076.84
+ +modules. During inference, we merge the LoRA module to the backbone. All throughputs are measured during inference on a GeForce RTX 3090 GPU. More details can be found in the supplementary materials. + +# 4.2 Performance comparison on Task Adaptation Benchmark + +We verify the effectiveness of PYRA for training-inference efficient task adaptation on (1) low compression rate: comparing the performance of different methods under the same sparsity ratio (here we set sparsity ratio $= 50\%$ ); (2) high compression rate: comparing the performance between the competing methods and the smaller-scale model with similar throughput levels. We leverage ViT-L/16 and ViT-B/16 for these inspections. The competing methods include RaP [35], SPViT [30], DiffRate [8], and ToMe [2]. Comparisons to more baselines are in the supplementary materials. For RaP and SPViT, we first train the backbone with LoRA on downstream tasks, then we prune the models, and lastly we retrain the parameters attached by the pruning method and LoRA. DiffRate and ToMe can be employed with LoRA during fine-tuning, while DiffRate demands ImageNet-21K for searching the optimal compression schedule. + +PYRA on low compression rate. Results on low compression rate are reported in Tab. 2. Overall, while achieving one of the best speedups on throughput, our PYRA is the best performed method compared to other competing methods using the lowest level of training parameters. Counting results on both backbones, our PYRA achieves the best or second-best performance on 37 of 38 dataset metrics. Compared to directly conducting PEFT on the backbone, while all competing methods cause worse results, our PYRA achieves comparable adaptation performance on ViT-B/16, and even outperforms ViT-L/16. The above results convincingly demonstrate that our PYRA successfully sets a new benchmark, i.e. reaching comparable performance as the uncompressed model, + +Table 3: Results on the VTAB-1k [67] benchmark under high compression rate. Bold and underline denote the best and second-best accuracy within compression methods. \*: As a comparison of similar throughputs, we compare ViT-B/16 with PEFT on ViTS/16, and ViT-L/16 with PEFT on ViT-B/16. + +
Method# paramsThroughputNaturalSpecializedStructured
Cifar100Caltech101DTDFlowers102PetsSVHNSun397CamelyonEuroSATResisc45RetinopathyClevr-CountClevr-DistDMLABKITTI-DistdSpr-LocdSpr-OrisNORB-AzimsNORB-EleAverage
Model: ViT-B/16 (Throughput: 425)
PEFT*0.34%135057.788.270.198.788.785.744.981.494.784.673.681.664.148.180.072.938.422.937.771.85
RaP0.86%102924.340.134.541.840.521.711.475.886.535.173.849.649.728.139.413.815.412.426.942.60
SPViT4.46%94423.767.951.969.953.219.613.171.981.367.974.753.561.939.557.445.034.511.123.252.49
DiffRate0.35%130823.273.055.787.966.727.229.378.177.853.173.629.728.631.752.611.517.211.320.349.29
ToMe0.34%138154.287.865.596.181.779.745.279.493.676.373.878.365.748.071.380.045.830.941.270.43
PYRA0.35%136554.089.367.196.584.081.844.681.294.679.575.179.967.049.276.982.647.831.942.072.06
Model: ViT-L/16 (Throughput: 130)
PEFT*0.34%42567.190.269.499.190.585.754.183.195.884.374.682.269.250.179.281.847.131.142.674.76
RaP0.65%30117.737.127.046.233.323.213.376.574.254.473.650.431.425.749.853.125.513.426.044.11
SPViT2.47%28954.087.665.594.874.932.638.681.895.378.074.072.861.246.970.277.147.431.328.666.90
DiffRate0.39%41647.473.554.184.360.219.622.250.064.642.818.231.531.931.137.322.017.714.821.440.48
ToMe0.39%43171.090.970.498.388.587.252.482.994.583.175.080.761.148.976.980.853.032.135.274.10
PYRA0.40%42771.691.871.198.589.788.152.285.195.384.675.780.963.051.782.082.054.236.041.275.66
+ +on low compression rate for training-inference efficient task adaptation while accelerating the model to $1.75 \times$ speedup with only $0.4\%$ training parameters. + +PYRA on high compression rate. Results are reported in Tab. 3. With comparable throughputs to the smaller-scale model and minimal training parameters, PYRA outperforms all competing methods and achieves the best performance on 35 of 38 dataset metrics. Compared to the smaller-scale model with comparable throughput, PYRA successfully outperforms it. On the compressed ViT-L/16, PYRA even surpasses ViT-B/16 by $0.9\%$ . This shows that PYRA, as the state-of-the-art, effectively bridges the adverse compression gap between compressed large-scale models and small-scale models, as mentioned in Sec. 1. Therefore, PYRA is an applicable alternative for acquiring a smaller-scale model $(3.2\times$ speedup) on downstream tasks through efficient training $(0.4\%)$ training parameters) when no pre-trained small-scale model is available. + +PYRA on self-supervised ViT backbone. We conduct the experiments on both compression rates on self-supervised ViT-L/16 (MAE) [20]. As shown in Tab. 4, PYRA significantly surpasses all competing methods. Under low compression rate, PYRA outperforms directly transferring the uncompressed backbone. Under high compression rate, PYRA also eliminates adverse compression. These results show that PYRA generalizes well for self-supervised visual models. + +PYRA on different architectures. To further verify the generalizability of PYRA, we conduct the above experiments on the DeiT-B [51]. As shown in Tab. 5, PYRA achieves comparable performance to the uncompressed model under low compression rate and eliminates adverse compression under high compression rate, indicating its generalizability to other transformer architectures. + +PYRA consistently yields better models of different throughputs. To further show the superiority of PYRA for different compression rates, we + +Table 4: VTAB-1k [67] results on both compression rates for ViT-L (MAE) (Throughput: 130). + +
ModelMethod# paramsThroughputAverage
ViT-L (MAE)PEFT0.39%13075.96
ViT-L (MAE)RaP [35]2.01%18267.55
DiffRate [8]0.39%22146.91
ToMe [2]0.39%22774.97
PYRA0.40%22576.13
ViT-B (MAE)PEFT0.34%42570.23
ViT-L (MAE)RaP [35]0.76%29852.91
DiffRate [8]0.39%41648.03
ToMe [2]0.39%43168.20
PYRA0.40%42770.33
+ +Table 5: VTAB-1k [67] results on both compression rates for DeiT-B (Throughput: 431). + +
ModelMethod# paramsThroughputAverage
DeiT-BPEFT0.34%43173.76
DeiT-BRaP [35]3.57%69562.34
DiffRate [8]0.35%73452.10
ToMe [2]0.34%74772.83
PYRA0.35%74073.55
DeiT-SPEFT0.34%133270.01
DeiT-BRaP [35]1.09%118757.70
DiffRate [8]0.35%131444.51
ToMe [2]0.34%135168.68
PYRA0.35%134170.13
+ +Table 6: Ablation study results for PYRA on ViT-B/16 under high compression rate. Here for # params we report only the parameters introduced by token modulation. + +
Method\( W_r \)\( W_D \)Activation# paramsNaturalSpecializedStructuredAverage
Baseline×××072.8780.7857.6470.43
Plain \( W_r \)√××0.19K72.9081.0757.6670.54
\( σ(\cdot) \& W_r \)√×√0.19K73.1881.6957.6570.84
Plain \( W_D \)×√×8.45K73.0981.1358.4370.88
\( σ(\cdot) \& W_D \)×√√8.45K73.3182.1758.4471.31
Plain \( W_r \& W_D \)√√×8.64K73.7781.3758.8171.32
PYRA√√√8.64K73.9182.6059.6672.06
+ +compress the chosen backbones for a series of speedups and compare the results with the strongest baseline, ToMe [2], under similar values of throughputs. As shown in Figs. 2 and 4, PYRA consistently outperforms ToMe and flattens the accuracy-throughput curve. Therefore, our PYRA is applicable to acquire task-specific smaller-scale models of different throughputs consistently in the absence of pre-trained parameters for smaller-scale models. + +# 4.3 Ablation Studies + +We do controlled experiments to identify the effect of individual components in PYRA, i.e., the generators $W_{r}$ and $W_{D}$ , and the sigmoid activation $\sigma(\cdot)$ applied on the generated modulation weights. Specifically, when using only a single generator, we omit Eq. (6) and replace $\hat{\delta}_r^l$ in Eq. (7) with the broadcast weight generated by the employed generator. When using the generators without sigmoid activation, we simply remove the $\sigma$ in Eq. (6) and Eq. (7) while keeping other calculations intact. We choose the ImageNet-21K pre-trained ViT-B/16 to carry out the ablation studies under high compression rates as in Tab. 3. + +![](images/cd78e6c8c8c119120dbf7db427ac6705fe8c42780ab35658e2372770f9bfdbdb.jpg) +Fig. 4: Comparisons between PYRA and ToMe [2] under different compression rates for ViT-Large (MAE) and DeiT-Base. Red boxes: PYRA mitigates the performance drops under low compression rate. Green boxes: PYRA eliminates adverse compression under high compression rate. + +![](images/93a8f85f613579036839ee19c25c863ffdd9b7d954343e0ea5711adc5ba0da07.jpg) + +Table 7: Adaptation performance comparison to $W_{r \times D}$ unadaptable to tokens. We report token modulation parameters. + +
Method# paramsNaturalSpecializedStructuredAverage
Baseline072.980.857.670.43
WD×r70.66K73.881.559.271.49
PYRA8.64K73.982.659.772.06
+ +Table 8: Adaptation performance comparison to the common gated generator. We report token modulation parameters. + +
Method# paramsNaturalSpecializedStructuredAverage
Baseline072.980.857.670.43
Gated73.73K72.881.558.971.06
PYRA8.64K73.982.659.772.06
+ +Results are shown in Tab. 6. Here the baseline method refers to conducting token merging [2,8] while fine-tuning LoRA [24] for task adaptation. First, both $W_{r}$ and $W_{D}$ , no matter whether sigmoid activation is attached, lead to performance gains, and employing them simultaneously outperforms using them individually. This proves the superiority of our parallel yielding strategy. Second, compared to the corresponding counterpart without activations, employing sigmoid activation always leads to significant improvements, indicating the effectiveness of the reactivation. Overall, our PYRA yields the best adaptation performance, showing that our strategies are effective and complementary. + +# 4.4 Further Analysis on Different Designs + +We conduct further analysis on ViT-B/16 model pre-trained on ImageNet-21K. The baseline here refers to training LoRA with token merging [2, 8] attached. More analysis experiments can be found in the supplementary materials. + +Impact of Making Modulation Weights Adaptive. In our PYRA, we train modulation weight generators $W_{D}^{l}$ and $W_{r}^{l}$ to conduct adaptive modulation on different merging tokens. To prove the necessity of making modulation weights adaptive, we compare PYRA with the approach of directly training final modulation weights $W_{D\times r}^{l}$ for each layer, and conduct token modulation as $M_{s}^{l}\gets M_{s}^{l} + (2\sigma (W_{D\times r}^{l}) - 1)\odot M_{s}^{l}$ . For fair comparisons, we inherit the sigmoid activation and the residual connection. As shown in Tab. 7, although with more training parameters, training $W_{D\times r}$ is still inferior to our PYRA. This indicates that PYRA is a more effective strategy to conduct adaptive token modulation. + +Compare with the Common Gated Generator. We compare PYRA with setting the commonly-applied gated-style trainable module [12, 22, 31, 62] as modulation weight generator. Formally, for each ViT block, we insert a learnable two-layer MLP module $(\mathrm{MLP}^l(\cdot))$ with input dimension $D$ , hidden dimension $d \ll D$ to ensure parameter-efficient training and fast inference, and output dimension $D$ . To generate modulation weights, we feed the information matrix $M_{\mathrm{info}}^l$ into the MLP, and modulate the merging tokens thereafter: $\delta^l = \mathrm{MLP}(M_{\mathrm{info}}^l) \in \mathbb{R}^{D \times r}$ , $M_s^l \gets M_s^l + (2\sigma(\delta^l - 1) \odot M_s^l)$ . We set $d = 4$ for all layers. As shown in Tab. 8, both the gated generator and our PYRA achieve performance gains, while our PYRA surpasses the gated generator by $1.0\%$ with significantly fewer trainable parameters. This indicates that the decoupled weights generated by our $W_D$ and $W_r$ in PYRA are effective and parameter-efficient compared to the common gating strategy. + +# 5 Conclusion + +In this work, we defined and investigated a new challenge named training-inference efficient task adaptation, in which the inference efficiency of large-scale transformers is enhanced during parameter-efficient task adaptation. We propose a novel Parallel Yielding Re-Activation method (PYRA) to effectively cope with the challenge by modulating token features during token merging. Specifically, PYRA generates decoupled parallel yielding modulation weights, and conducts token modulation through re-activation. Extensive experiments show that PYRA introduces negligible performance drops under low compression rate, and bridges the gap of adverse compression between compressed transformers and small-scale models under high compression rate. In real-world applications, our PYRA is highly suitable to the scenario of transferring large-scale vision transformers to downstream tasks where no small-scale model is presented. + +Limitations. We have not yet validated the effectiveness of PYRA in object detection and image segmentation. We plan to extend these tasks in the future. Meanwhile, it is also applicable to attach pruning methods after training the vision transformers with PYRA to eliminate additional parameters attached by PYRA. We leave that topic to future studies. + +# Acknowledgements + +This work was supported by National Science and Technology Major 2022ZD0119401, National Natural Science Foundation of China (Nos. 62271281, 61925107, 62021002). It is also sponsored by CAAI-CANN Open Fund, developed on OpenI Community. + +# References + +1. Ahn, S., Hu, S.X., Damianou, A., Lawrence, N.D., Dai, Z.: Variational information distillation for knowledge transfer. In: Proceedings of the IEEE/CVF conference on computer vision and pattern recognition. pp. 9163-9171 (2019) +2. Bolya, D., Fu, C.Y., Dai, X., Zhang, P., Feichtenhofer, C., Hoffman, J.: Token merging: Your vit but faster. arXiv preprint arXiv:2210.09461 (2022) +3. Cai, Y., Yao, Z., Dong, Z., Gholami, A., Mahoney, M.W., Keutzer, K.: Zeroq: A novel zero shot quantization framework. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 13169-13178 (2020) +4. Carion, N., Massa, F., Synnaeve, G., Usunier, N., Kirillov, A., Zagoruyko, S.: End-to-end object detection with transformers. In: European conference on computer vision. pp. 213-229. Springer (2020) +5. Chavan, A., Shen, Z., Liu, Z., Liu, Z., Cheng, K.T., Xing, E.P.: Vision transformer slimming: Multi-dimension searching in continuous optimization space. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 4931-4941 (2022) +6. Chen, C.F.R., Fan, Q., Panda, R.: Crossvit: Cross-attention multi-scale vision transformer for image classification. In: Proceedings of the IEEE/CVF international conference on computer vision. pp. 357-366 (2021) +7. Chen, L., Zhao, H., Liu, T., Bai, S., Lin, J., Zhou, C., Chang, B.: An image is worth 1/2 tokens after layer 2: Plug-and-play inference acceleration for large vision-language models. arXiv preprint arXiv:2403.06764 (2024) +8. Chen, M., Shao, W., Xu, P., Lin, M., Zhang, K., Chao, F., Ji, R., Qiao, Y., Luo, P.: Diffrate: Differentiable compression rate for efficient vision transformers. arXiv preprint arXiv:2305.17997 (2023) +9. Chen, S., Ge, C., Tong, Z., Wang, J., Song, Y., Wang, J., Luo, P.: Adaptformer: Adapting vision transformers for scalable visual recognition. Advances in Neural Information Processing Systems 35, 16664-16678 (2022) +0. Chen, T., Cheng, Y., Gan, Z., Yuan, L., Zhang, L., Wang, Z.: Chasing sparsity in vision transformers: An end-to-end exploration. Advances in Neural Information Processing Systems 34, 19974-19988 (2021) +11. Cheng, B., Misra, I., Schwing, A.G., Kirillov, A., Girdhar, R.: Masked-attention mask transformer for universal image segmentation. In: Proceedings of the IEEE/CVF conference on computer vision and pattern recognition. pp. 1290-1299 (2022) +12. Chung, J., Gulcehre, C., Cho, K., Bengio, Y.: Empirical evaluation of gated recurrent neural networks on sequence modeling. arXiv preprint arXiv:1412.3555 (2014) +13. Courbariaux, M., Bengio, Y., David, J.P.: Binaryconnect: Training deep neural networks with binary weights during propagations. Advances in neural information processing systems 28 (2015) + +14. Ding, N., Qin, Y., Yang, G., Wei, F., Yang, Z., Su, Y., Hu, S., Chen, Y., Chan, C.M., Chen, W., et al.: Delta tuning: A comprehensive study of parameter efficient methods for pre-trained language models. arXiv preprint arXiv:2203.06904 (2022) +15. Ding, Z., Wang, A., Chen, H., Zhang, Q., Liu, P., Bao, Y., Yan, W., Han, J.: Exploring structured semantic prior for multi label recognition with incomplete labels. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 3398-3407 (2023) +16. Dosovitskiy, A., Beyer, L., Kolesnikov, A., Weissenborn, D., Zhai, X., Unterthiner, T., Dehghani, M., Minderer, M., Heigold, G., Gelly, S., et al.: An image is worth 16x16 words: Transformers for image recognition at scale. arXiv preprint arXiv:2010.11929 (2020) +17. Hao, T., Chen, H., Guo, Y., Ding, G.: Consolidator: Mergeable adapter with grouped connections for visual adaptation. arXiv preprint arXiv:2305.00603 (2023) +18. Hao, T., Ding, X., Feng, J., Yang, Y., Chen, H., Ding, G.: Quantized prompt for efficient generalization of vision-language models. arXiv preprint arXiv:2407.10704 (2024) +19. Hao, T., Lyu, M., Chen, H., Zhao, S., Han, J., Ding, G.: Re-parameterized low-rank prompt: Generalize a vision-language model within $0.5\mathrm{k}$ parameters. arXiv preprint arXiv:2312.10813 (2023) +20. He, K., Chen, X., Xie, S., Li, Y., Dóllar, P., Girshick, R.: Masked autoencoders are scalable vision learners. In: Proceedings of the IEEE/CVF conference on computer vision and pattern recognition. pp. 16000-16009 (2022) +21. Hinton, G., Vinyals, O., Dean, J.: Distilling the knowledge in a neural network. arXiv preprint arXiv:1503.02531 (2015) +22. Hochreiter, S., Schmidhuber, J.: Long short-term memory. Neural Computation 9(8), 1735-1780 (1997) +23. Houlsby, N., Giurgiu, A., Jastrzebski, S., Morrone, B., De Laroussilhe, Q., Gesmundo, A., Attariyan, M., Gelly, S.: Parameter-efficient transfer learning for nlp. In: International Conference on Machine Learning. pp. 2790-2799. PMLR (2019) +24. Hu, E.J., Shen, Y., Wallis, P., Allen-Zhu, Z., Li, Y., Wang, S., Wang, L., Chen, W.: Lora: Low-rank adaptation of large language models. arXiv preprint arXiv:2106.09685 (2021) +25. Huang, X., Shen, Z., Cheng, K.T.: Variation-aware vision transformer quantization. arXiv preprint arXiv:2307.00331 (2023) +26. Jia, M., Tang, L., Chen, B.C., Cardie, C., Belongie, S., Hariharan, B., Lim, S.N.: Visual prompt tuning. In: European Conference on Computer Vision. pp. 709-727. Springer (2022) +27. Jiang, H., Wu, Q., Lin, C.Y., Yang, Y., Qiu, L.: Llmlingua: Compressing prompts for accelerated inference of large language models. arXiv preprint arXiv:2310.05736 (2023) +28. Jie, S., Deng, Z.H.: Convolutional bypasses are better vision transformer adapters. arXiv preprint arXiv:2207.07039 (2022) +29. Kirillov, A., Mintun, E., Ravi, N., Mao, H., Rolland, C., Gustafson, L., Xiao, T., Whitehead, S., Berg, A.C., Lo, W.Y., et al.: Segment anything. arXiv preprint arXiv:2304.02643 (2023) +30. Kong, Z., Dong, P., Ma, X., Meng, X., Niu, W., Sun, M., Shen, X., Yuan, G., Ren, B., Tang, H., Qin, M., Wang, Y.: Spvit: Enabling faster vision transformers via latency-aware soft token pruning. In: Avidan, S., Brostow, G.J., Cissé, M., Farinella, G.M., Hassner, T. (eds.) Computer Vision - ECCV 2022 - 17th European Conference, Tel Aviv, Israel, October 23-27, 2022, Proceedings, Part XI. Lecture + +Notes in Computer Science, vol. 13671, pp. 620-640. Springer (2022). https://doi.org/10.1007/978-3-031-20083-0_37, https://doi.org/10.1007/978-3-031-20083-0_37 +31. Kong, Z., Dong, P., Ma, X., Meng, X., Sun, M., Niu, W., Shen, X., Yuan, G., Ren, B., Qin, M., et al.: Spvit: Enabling faster vision transformers via soft token pruning. arXiv preprint arXiv:2112.13890 (2021) +32. Lester, B., Al-Rfou, R., Constant, N.: The power of scale for parameter-efficient prompt tuning. arXiv preprint arXiv:2104.08691 (2021) +33. Li, X.L., Liang, P.: Prefix-tuning: Optimizing continuous prompts for generation. arXiv preprint arXiv:2101.00190 (2021) +34. Li, Y., Xu, S., Zhang, B., Cao, X., Gao, P., Guo, G.: Q-vit: Accurate and fully quantized low-bit vision transformer. Advances in Neural Information Processing Systems 35, 34451-34463 (2022) +35. Li, Y., Adamczewski, K., Li, W., Gu, S., Timofte, R., Gool, L.V.: Revisiting random channel pruning for neural network compression. In: IEEE/CVF Conference on Computer Vision and Pattern Recognition, CVPR 2022, New Orleans, LA, USA, June 18-24, 2022. pp. 191-201. IEEE (2022). https://doi.org/10.1109/CVPR52688.2022.00029, https://doi.org/10.1109/CVPR52688.2022.00029 +36. Lian, D., Zhou, D., Feng, J., Wang, X.: Scaling & shifting your features: A new baseline for efficient model tuning. Advances in Neural Information Processing Systems 35, 109-123 (2022) +37. Lin, X., Zhao, C., Pan, W.: Towards accurate binary convolutional neural network. Advances in neural information processing systems 30 (2017) +38. Lin, Z., Courbariaux, M., Memisevic, R., Bengio, Y.: Neural networks with few multiplications. arXiv preprint arXiv:1510.03009 (2015) +39. Liu, X., Ji, K., Fu, Y., Tam, W.L., Du, Z., Yang, Z., Tang, J.: P-tuning v2: Prompt tuning can be comparable to fine-tuning universally across scales and tasks. arXiv preprint arXiv:2110.07602 (2021) +40. Liu, Z., Lin, Y., Cao, Y., Hu, H., Wei, Y., Zhang, Z., Lin, S., Guo, B.: Swin transformer: Hierarchical vision transformer using shifted windows. In: Proceedings of the IEEE/CVF international conference on computer vision. pp. 10012-10022 (2021) +41. Liu, Z., Wang, Y., Han, K., Zhang, W., Ma, S., Gao, W.: Post-training quantization for vision transformer. Advances in Neural Information Processing Systems 34, 28092-28103 (2021) +42. Lyu, M., Yang, Y., Hong, H., Chen, H., Jin, X., He, Y., Xue, H., Han, J., Ding, G.: One-dimensional adapter to rule them all: Concepts, diffusion models and erasing applications. arXiv preprint arXiv:2312.16145 (2023) +43. Marin, D., Chang, J.H.R., Ranjan, A., Prabhu, A., Rastegari, M., Tuzel, O.: Token pooling in vision transformers. arXiv preprint arXiv:2110.03860 (2021) +44. Perez, E., Strub, F., De Vries, H., Dumoulin, V., Courville, A.: Film: Visual reasoning with a general conditioning layer. In: Proceedings of the AAAI conference on artificial intelligence (2018) +45. Pfeiffer, J., Kamath, A., Rücklé, A., Cho, K., Gurevych, I.: Adapterfusion: Nondestructive task composition for transfer learning. arXiv preprint arXiv:2005.00247 (2020) +46. Song, Y., Zhou, Q., Li, X., Fan, D.P., Lu, X., Ma, L.: Ba-sam: Scalable bias-mode attention mask for segment anything model. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 3162-3173 (2024) + +47. Song, Z., Jia, C., Yang, L., Wei, H., Liu, L.: Graphalign++: An accurate feature alignment by graph matching for multi-modal 3d object detection. IEEE Transactions on Circuits and Systems for Video Technology (2023) +48. Song, Z., Wei, H., Bai, L., Yang, L., Jia, C.: Graphalign: Enhancing accurate feature alignment by graph matching for multi-modal 3d object detection. In: Proceedings of the IEEE/CVF International Conference on Computer Vision. pp. 3358-3369 (2023) +49. Strudel, R., Garcia, R., Laptev, I., Schmid, C.: Segmenter: Transformer for semantic segmentation. In: Proceedings of the IEEE/CVF international conference on computer vision. pp. 7262-7272 (2021) +50. Tian, L., Ye, M., Zhou, L., He, Q.: Clip-guided black-box domain adaptation of image classification. Signal, Image and Video Processing 18(5), 4637-4646 (2024) +51. Touvron, H., Cord, M., Douze, M., Massa, F., Sablayrolles, A., Jégou, H.: Training data-efficient image transformers & distillation through attention. In: International conference on machine learning. pp. 10347-10357. PMLR (2021) +52. Tung, F., Mori, G.: Similarity-preserving knowledge distillation. In: Proceedings of the IEEE/CVF international conference on computer vision. pp. 1365-1374 (2019) +53. Wang, A., Chen, H., Lin, Z., Pu, H., Ding, G.: Repvit: Revisiting mobile cnn from vit perspective. arXiv preprint arXiv:2307.09283 (2023) +54. Wang, A., Chen, H., Lin, Z., Zhao, S., Han, J., Ding, G.: Cait: Triple-win compression towards high accuracy, fast inference, and favorable transferability for vits. arXiv preprint arXiv:2309.15755 (2023) +55. Wang, W., Xie, E., Li, X., Fan, D.P., Song, K., Liang, D., Lu, T., Luo, P., Shao, L.: Pyramid vision transformer: A versatile backbone for dense prediction without convolutions. In: Proceedings of the IEEE/CVF international conference on computer vision. pp. 568-578 (2021) +56. Wang, Z., Luo, H., Wang, P., Ding, F., Wang, F., Li, H.: Vtc-lfc: Vision transformer compression with low-frequency components. Advances in Neural Information Processing Systems 35, 13974-13988 (2022) +57. Wei, K., Du, R., Jin, L., Liu, J., Yin, J., Zhang, L., Liu, J., Liu, N., Zhang, J., Guo, Z.: Video event extraction with multi-view interaction knowledge distillation. In: Proceedings of the AAAI Conference on Artificial Intelligence. vol. 38, pp. 19224-19233 (2024) +58. Wei, K., Jin, L., Zhang, Z., Guo, Z., Li, X., Liu, Q., Feng, W.: More than syntaxes: Investigating semantics to zero-shot cross-lingual relation extraction and event argument role labelling. ACM Transactions on Asian and Low-Resource Language Information Processing 23(5), 1-21 (2024) +59. Wu, Y., He, K.: Group normalization. In: Proceedings of the European conference on computer vision (ECCV). pp. 3-19 (2018) +60. Xiong, Y., Chen, H., Lin, Z., Zhao, S., Ding, G.: Confidence-based visual dispersal for few-shot unsupervised domain adaptation. In: Proceedings of the IEEE/CVF International Conference on Computer Vision. pp. 11621-11631 (2023) +61. Xiong, Y., Chen, X., Ye, X., Chen, H., Lin, Z., Lian, H., Niu, J., Ding, G.: Temporal scaling law for large language models. arXiv preprint arXiv:2404.17785 (2024) +62. Xu, L., Xie, H., Qin, S.Z.J., Tao, X., Wang, F.L.: Parameter-efficient fine-tuning methods for pretrained language models: A critical review and assessment. arXiv preprint arXiv:2312.12148 (2023) +63. Yao, Z., Yazdani Aminabadi, R., Zhang, M., Wu, X., Li, C., He, Y.: Zeroquant: Efficient and affordable post-training quantization for large-scale transformers. Advances in Neural Information Processing Systems 35, 27168-27183 (2022) + +64. Yu, S., Chen, T., Shen, J., Yuan, H., Tan, J., Yang, S., Liu, J., Wang, Z.: Unified visual transformer compression. arXiv preprint arXiv:2203.08243 (2022) +65. Zaken, E.B., Ravfogel, S., Goldberg, Y.: Bitfit: Simple parameter-efficient fine-tuning for transformer-based masked language-models. arXiv preprint arXiv:2106.10199 (2021) +66. Zhai, X., Kolesnikov, A., Houlsby, N., Beyer, L.: Scaling vision transformers. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 12104-12113 (2022) +67. Zhai, X., Puigcerver, J., Kolesnikov, A., Ruyssen, P., Riquelme, C., Lucic, M., Djolonga, J., Pinto, A.S., Neumann, M., Dosovitskiy, A., et al.: A large-scale study of representation learning with the visual task adaptation benchmark. arXiv preprint arXiv:1910.04867 (2019) +68. Zhang, Y., Qiu, L., Zhu, Y., Wen, L., Luo, X.: A new childhood pneumonia diagnosis method based on fine-grained convolutional neural network. Computer Modeling in Engineering & Sciences 133, 873-894 (2022) +69. Zhang, Y., Zhou, K., Liu, Z.: Neural prompt search. arXiv preprint arXiv:2206.04673 (2022) +70. Zheng, S., Lu, J., Zhao, H., Zhu, X., Luo, Z., Wang, Y., Fu, Y., Feng, J., Xiang, T., Torr, P.H., et al.: Rethinking semantic segmentation from a sequence-to-sequence perspective with transformers. In: Proceedings of the IEEE/CVF conference on computer vision and pattern recognition. pp. 6881-6890 (2021) +71. Zhou, F., Zhou, Q., Yin, B., Zheng, H., Lu, X., Ma, L., Ling, H.: Rethinking impersonation and dodging attacks on face recognition systems. In: Proceedings of the 30th ACM International Conference on Multimedia (ACM MM) (2024) +72. Zhou, Q., Feng, Z., Gu, Q., Pang, J., Cheng, G., Lu, X., Shi, J., Ma, L.: Context-aware mixup for domain adaptive semantic segmentation. IEEE Transactions on Circuits and Systems for Video Technology 33(2), 804-817 (2022) +73. Zhu, M., Tang, Y., Han, K.: Vision transformer pruning. arXiv preprint arXiv:2104.08500 (2021) +74. Zhu, X., Su, W., Lu, L., Li, B., Wang, X., Dai, J.: Deformable detr: Deformable transformers for end-to-end object detection. arXiv preprint arXiv:2010.04159 (2020) \ No newline at end of file diff --git a/2024/PYRA_ Parallel Yielding Re-Activation for Training-Inference Efficient Task Adaptation/images.zip b/2024/PYRA_ Parallel Yielding Re-Activation for Training-Inference Efficient Task Adaptation/images.zip new file mode 100644 index 0000000000000000000000000000000000000000..21034717d77a0eec7eb517ca2083736fcd1cc39c --- /dev/null +++ b/2024/PYRA_ Parallel Yielding Re-Activation for Training-Inference Efficient Task Adaptation/images.zip @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:63d0b4a688746355e4925fdadf48e64df310b2b25a9a911abc881e491343ad92 +size 616060 diff --git a/2024/PYRA_ Parallel Yielding Re-Activation for Training-Inference Efficient Task Adaptation/layout.json b/2024/PYRA_ Parallel Yielding Re-Activation for Training-Inference Efficient Task Adaptation/layout.json new file mode 100644 index 0000000000000000000000000000000000000000..04ce940f13d66ffc26a0021196fa562be9fb4190 --- /dev/null +++ b/2024/PYRA_ Parallel Yielding Re-Activation for Training-Inference Efficient Task Adaptation/layout.json @@ -0,0 +1,11201 @@ +{ + "pdf_info": [ + { + "para_blocks": [ + { + "bbox": [ + 151, + 111, + 462, + 148 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 151, + 111, + 462, + 148 + ], + "spans": [ + { + "bbox": [ + 151, + 111, + 462, + 148 + ], + "type": "text", + "content": "PYRA: Parallel Yielding Re-Activation for Training-Inference Efficient Task Adaptation" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 136, + 167, + 478, + 204 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 136, + 167, + 478, + 204 + ], + "spans": [ + { + "bbox": [ + 136, + 167, + 478, + 204 + ], + "type": "text", + "content": "Yizhe Xiong" + }, + { + "bbox": [ + 136, + 167, + 478, + 204 + ], + "type": "inline_equation", + "content": "^{1,2}" + }, + { + "bbox": [ + 136, + 167, + 478, + 204 + ], + "type": "text", + "content": ", Hui Chen" + }, + { + "bbox": [ + 136, + 167, + 478, + 204 + ], + "type": "inline_equation", + "content": "^{2*}" + }, + { + "bbox": [ + 136, + 167, + 478, + 204 + ], + "type": "text", + "content": ", Tianxiang Hao" + }, + { + "bbox": [ + 136, + 167, + 478, + 204 + ], + "type": "inline_equation", + "content": "^{1,2}" + }, + { + "bbox": [ + 136, + 167, + 478, + 204 + ], + "type": "text", + "content": ", Zijia Lin" + }, + { + "bbox": [ + 136, + 167, + 478, + 204 + ], + "type": "inline_equation", + "content": "^{1}" + }, + { + "bbox": [ + 136, + 167, + 478, + 204 + ], + "type": "text", + "content": ", Jungong Han" + }, + { + "bbox": [ + 136, + 167, + 478, + 204 + ], + "type": "inline_equation", + "content": "^{2,3}" + }, + { + "bbox": [ + 136, + 167, + 478, + 204 + ], + "type": "text", + "content": ", Yuesong Zhang" + }, + { + "bbox": [ + 136, + 167, + 478, + 204 + ], + "type": "inline_equation", + "content": "^{4}" + }, + { + "bbox": [ + 136, + 167, + 478, + 204 + ], + "type": "text", + "content": ", Guoxin Wang" + }, + { + "bbox": [ + 136, + 167, + 478, + 204 + ], + "type": "inline_equation", + "content": "^{4}" + }, + { + "bbox": [ + 136, + 167, + 478, + 204 + ], + "type": "text", + "content": ", Yongjun Bao" + }, + { + "bbox": [ + 136, + 167, + 478, + 204 + ], + "type": "inline_equation", + "content": "^{4}" + }, + { + "bbox": [ + 136, + 167, + 478, + 204 + ], + "type": "text", + "content": ", and Guiguang Ding" + }, + { + "bbox": [ + 136, + 167, + 478, + 204 + ], + "type": "inline_equation", + "content": "^{1,2}" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 188, + 213, + 424, + 225 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 188, + 213, + 424, + 225 + ], + "spans": [ + { + "bbox": [ + 188, + 213, + 424, + 225 + ], + "type": "inline_equation", + "content": "^{1}" + }, + { + "bbox": [ + 188, + 213, + 424, + 225 + ], + "type": "text", + "content": " School of Software, Tsinghua University, Beijing, China" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 212, + 224, + 402, + 236 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 212, + 224, + 402, + 236 + ], + "spans": [ + { + "bbox": [ + 212, + 224, + 402, + 236 + ], + "type": "inline_equation", + "content": "^{2}" + }, + { + "bbox": [ + 212, + 224, + 402, + 236 + ], + "type": "text", + "content": " BNRist, Tsinghua University, Beijing, China" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 171, + 235, + 441, + 247 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 171, + 235, + 441, + 247 + ], + "spans": [ + { + "bbox": [ + 171, + 235, + 441, + 247 + ], + "type": "inline_equation", + "content": "^{3}" + }, + { + "bbox": [ + 171, + 235, + 441, + 247 + ], + "type": "text", + "content": " Department of Automation, Tsinghua University, Beijing, China" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 255, + 247, + 359, + 258 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 255, + 247, + 359, + 258 + ], + "spans": [ + { + "bbox": [ + 255, + 247, + 359, + 258 + ], + "type": "text", + "content": "4 JD.com, Beijing, China" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 190, + 258, + 423, + 270 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 190, + 258, + 423, + 270 + ], + "spans": [ + { + "bbox": [ + 190, + 258, + 423, + 270 + ], + "type": "text", + "content": "{xiongyizhe2001,beyondtx,jungonghan77}@gmail.com" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 239, + 270, + 373, + 280 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 239, + 270, + 373, + 280 + ], + "spans": [ + { + "bbox": [ + 239, + 270, + 373, + 280 + ], + "type": "text", + "content": "huichen@mail.tsinghua.edu.cn" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 198, + 281, + 416, + 291 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 198, + 281, + 416, + 291 + ], + "spans": [ + { + "bbox": [ + 198, + 281, + 416, + 291 + ], + "type": "text", + "content": "{zhangyuesong1,wangguoxin14,baoyongjun}@jd.com" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 188, + 292, + 425, + 302 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 188, + 292, + 425, + 302 + ], + "spans": [ + { + "bbox": [ + 188, + 292, + 425, + 302 + ], + "type": "text", + "content": "linzijia07@tsinghua.org.cn dinggg@tsinghua.edu.cn" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 159, + 327, + 455, + 548 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 159, + 327, + 455, + 548 + ], + "spans": [ + { + "bbox": [ + 159, + 327, + 455, + 548 + ], + "type": "text", + "content": "Abstract. Recently, the scale of transformers has grown rapidly, which introduces considerable challenges in terms of training overhead and inference efficiency in the scope of task adaptation. Existing works, namely Parameter-Efficient Fine-Tuning (PEFT) and model compression, have separately investigated the challenges. However, PEFT cannot guarantee the inference efficiency of the original backbone, especially for large-scale models. Model compression requires significant training costs for structure searching and re-training. Consequently, a simple combination of them cannot guarantee accomplishing both training efficiency and inference efficiency with minimal costs. In this paper, we propose a novel Parallel Yielding Re-Activation (PYRA) method for such a challenge of training-inference efficient task adaptation. PYRA first utilizes parallel yielding adaptive weights to comprehensively perceive the data distribution in downstream tasks. A re-activation strategy for token modulation is then applied for tokens to be merged, leading to calibrated token features. Extensive experiments demonstrate that PYRA outperforms all competing methods under both low compression rate and high compression rate, demonstrating its effectiveness and superiority in maintaining both training efficiency and inference efficiency for large-scale foundation models. Our code is available at https://github.com/THU-MIG/PyRA." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 160, + 557, + 453, + 569 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 160, + 557, + 453, + 569 + ], + "spans": [ + { + "bbox": [ + 160, + 557, + 453, + 569 + ], + "type": "text", + "content": "Keywords: Vision Transformer " + }, + { + "bbox": [ + 160, + 557, + 453, + 569 + ], + "type": "inline_equation", + "content": "\\cdot" + }, + { + "bbox": [ + 160, + 557, + 453, + 569 + ], + "type": "text", + "content": " Task Adaptation " + }, + { + "bbox": [ + 160, + 557, + 453, + 569 + ], + "type": "inline_equation", + "content": "\\cdot" + }, + { + "bbox": [ + 160, + 557, + 453, + 569 + ], + "type": "text", + "content": " Model Compression" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 132, + 587, + 230, + 600 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 587, + 230, + 600 + ], + "spans": [ + { + "bbox": [ + 132, + 587, + 230, + 600 + ], + "type": "text", + "content": "1 Introduction" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 130, + 611, + 482, + 647 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 611, + 482, + 647 + ], + "spans": [ + { + "bbox": [ + 130, + 611, + 482, + 647 + ], + "type": "text", + "content": "Vision transformers [16] have made a profound impact across various domains of computer vision, such as image classification [6,15,16,50,55,66], object detection [4,40,47,48,74], image segmentation [11,49,53,70,72], etc. In recent years, the" + } + ] + } + ], + "index": 13 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 133, + 653, + 236, + 665 + ], + "type": "page_footnote", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 653, + 236, + 665 + ], + "spans": [ + { + "bbox": [ + 133, + 653, + 236, + 665 + ], + "type": "text", + "content": "* Corresponding Author" + } + ] + } + ], + "index": 14 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 0 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 136, + 116, + 233, + 165 + ], + "blocks": [ + { + "bbox": [ + 136, + 116, + 233, + 165 + ], + "lines": [ + { + "bbox": [ + 136, + 116, + 233, + 165 + ], + "spans": [ + { + "bbox": [ + 136, + 116, + 233, + 165 + ], + "type": "image", + "image_path": "5664647c2b736e0e82a403b7c620536892af767f6bc456d67ad6a3f18ccce171.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 151, + 167, + 220, + 176 + ], + "lines": [ + { + "bbox": [ + 151, + 167, + 220, + 176 + ], + "spans": [ + { + "bbox": [ + 151, + 167, + 220, + 176 + ], + "type": "text", + "content": "(a) Full Fine-tuning" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_caption" + } + ], + "index": 2 + }, + { + "type": "image", + "bbox": [ + 136, + 177, + 233, + 224 + ], + "blocks": [ + { + "bbox": [ + 136, + 177, + 233, + 224 + ], + "lines": [ + { + "bbox": [ + 136, + 177, + 233, + 224 + ], + "spans": [ + { + "bbox": [ + 136, + 177, + 233, + 224 + ], + "type": "image", + "image_path": "2ab152f60759215e24818de68dbcf2edb03dee47b612c65aa58460f983ac19a7.jpg" + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 169, + 226, + 201, + 235 + ], + "lines": [ + { + "bbox": [ + 169, + 226, + 201, + 235 + ], + "spans": [ + { + "bbox": [ + 169, + 226, + 201, + 235 + ], + "type": "text", + "content": "(c) PEFT" + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "image_caption" + } + ], + "index": 4 + }, + { + "type": "image", + "bbox": [ + 245, + 118, + 343, + 166 + ], + "blocks": [ + { + "bbox": [ + 245, + 118, + 343, + 166 + ], + "lines": [ + { + "bbox": [ + 245, + 118, + 343, + 166 + ], + "spans": [ + { + "bbox": [ + 245, + 118, + 343, + 166 + ], + "type": "image", + "image_path": "36b7ea242bb2fc69de46fb442daa7344f5f931b70f4eefd2392f17f1de845c7d.jpg" + } + ] + } + ], + "index": 6, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 256, + 167, + 335, + 176 + ], + "lines": [ + { + "bbox": [ + 256, + 167, + 335, + 176 + ], + "spans": [ + { + "bbox": [ + 256, + 167, + 335, + 176 + ], + "type": "text", + "content": "(b) Model compression" + } + ] + } + ], + "index": 7, + "angle": 0, + "type": "image_caption" + } + ], + "index": 6 + }, + { + "type": "image", + "bbox": [ + 245, + 177, + 342, + 224 + ], + "blocks": [ + { + "bbox": [ + 245, + 177, + 342, + 224 + ], + "lines": [ + { + "bbox": [ + 245, + 177, + 342, + 224 + ], + "spans": [ + { + "bbox": [ + 245, + 177, + 342, + 224 + ], + "type": "image", + "image_path": "e939f1d5fbe6cc22b2fc2ddb8383ad38f11f5c6b2de37f5afaa3082c0f1cf27b.jpg" + } + ] + } + ], + "index": 8, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 280, + 226, + 311, + 235 + ], + "lines": [ + { + "bbox": [ + 280, + 226, + 311, + 235 + ], + "spans": [ + { + "bbox": [ + 280, + 226, + 311, + 235 + ], + "type": "text", + "content": "(d) Ours" + } + ] + } + ], + "index": 9, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 130, + 244, + 482, + 332 + ], + "lines": [ + { + "bbox": [ + 130, + 244, + 482, + 332 + ], + "spans": [ + { + "bbox": [ + 130, + 244, + 482, + 332 + ], + "type": "text", + "content": "Fig. 1: (a) Full fine-tuning trains all parameters on downstream tasks and utilizes the trained model for inference, thereby lacking efficiency in both training and inference stages. (b) Model compression employs pruning to enhance inference efficiency, but the pruned model necessitates extensive re-training on large-scale data. (c) PEFT freezes the model backbone and only fine-tunes a small amount of parameters, yet retains the inference complexity. (d) Our training-inference efficient task adaptation incorporates the advantages of all existing pipelines by training inference-efficient models with minimal tunable parameters." + } + ] + } + ], + "index": 11, + "angle": 0, + "type": "image_caption" + } + ], + "index": 8 + }, + { + "type": "table", + "bbox": [ + 351, + 116, + 479, + 226 + ], + "blocks": [ + { + "bbox": [ + 351, + 116, + 479, + 226 + ], + "lines": [ + { + "bbox": [ + 351, + 116, + 479, + 226 + ], + "spans": [ + { + "bbox": [ + 351, + 116, + 479, + 226 + ], + "type": "table", + "html": "
Method(a)(b)(c)(d)
Non-vast Data Scale✔×✔✔
Training Efficient××✔✔
Inference Efficient×✔×✔
", + "image_path": "d733460eb81370b56af7d9c23d8cbcde0c31a3a7d07b8fdc3776a55a00fff777.jpg" + } + ] + } + ], + "index": 10, + "angle": 0, + "type": "table_body" + } + ], + "index": 10 + }, + { + "bbox": [ + 130, + 361, + 482, + 445 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 361, + 482, + 445 + ], + "spans": [ + { + "bbox": [ + 130, + 361, + 482, + 445 + ], + "type": "text", + "content": "scale of vision transformers has grown to be billion-parameters [29,61,66]. Consequently, adapting such models with vast scales into downstream tasks presents increasingly complex challenges, particularly in real-world deployment scenarios. Two critical concerns have been widely acknowledged as primary obstacles in implementing large-scale transformers for downstream applications [17,23,27,46]: (1) the training overhead when fine-tuning on downstream tasks, and (2) the inference efficiency after model deployment." + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 130, + 448, + 482, + 592 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 448, + 482, + 592 + ], + "spans": [ + { + "bbox": [ + 130, + 448, + 482, + 592 + ], + "type": "text", + "content": "Specifically, first, conventional fine-tuning methods, which necessitate adjusting all parameters of the model (i.e., Full Fine-tuning), suffer from unaffordable consumption of GPU resources and training time given the extensive scales of the foundation models [16, 29, 66]. Researchers have delved into Parameter-Efficient Fine-Tuning (PEFT) [17, 19, 23, 24, 26, 36] algorithms, which generally freeze the pre-trained models and only tune extra small parameters, leading to great reduction of training time and storage overhead. The second issue pertains to inference efficiency, requiring the deployed model to promptly process the input data. The computational complexity of the models significantly influences achieving satisfactory inference throughput. Representative solutions for this matter encompass model compression methods, including model pruning [7, 10, 56, 64], knowledge distillation [1, 21, 52, 57, 58], model quantization [3, 13, 37], etc." + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 130, + 594, + 482, + 666 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 594, + 482, + 666 + ], + "spans": [ + { + "bbox": [ + 130, + 594, + 482, + 666 + ], + "type": "text", + "content": "In the literature, these two intriguing topics are investigated separately. PEFT methods either identify a subset of tunable parameters in the backbone for fine-tuning [65] or introduce learnable parameters to the frozen backbone during fine-tuning [17, 23, 24, 26, 36]. While effectively reducing training costs, most of these methods inevitably escalate computational complexity, resulting in inefficient inference. Model compression methods frequently require signifi" + } + ] + } + ], + "index": 14 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "spans": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "type": "text", + "content": "2" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 166, + 91, + 228, + 102 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 166, + 91, + 228, + 102 + ], + "spans": [ + { + "bbox": [ + 166, + 91, + 228, + 102 + ], + "type": "text", + "content": "Y. Xiong et al." + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 1 + }, + { + "para_blocks": [ + { + "bbox": [ + 130, + 116, + 482, + 164 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 116, + 482, + 164 + ], + "spans": [ + { + "bbox": [ + 130, + 116, + 482, + 164 + ], + "type": "text", + "content": "cant computational resources to identify optimal structures for pruning. After pruning, a comprehensive re-training process using a substantial amount of data is crucial to prevent significant performance degradation. Therefore, model compression methods are typically inefficient in training efficiency." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 130, + 165, + 265, + 426 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 165, + 265, + 426 + ], + "spans": [ + { + "bbox": [ + 130, + 165, + 265, + 426 + ], + "type": "text", + "content": "These observations naturally lead to the question: can we achieve both training efficiency and inference efficiency simultaneously for downstream tasks? We refer to this challenge as Training-Inference Efficient Task Adaptation (seeing Fig. 1). Exploring this issue can enable us to conveniently deploy the advanced large-scale foundation models in real-world downstream applications with minimal costs, which is appealing and essential for the widespread implementation of foundation models. A straightforward solution is combining PEFT and model compression. However, for efficient task adaptation, a" + } + ] + } + ], + "index": 3 + }, + { + "type": "image", + "bbox": [ + 271, + 189, + 480, + 319 + ], + "blocks": [ + { + "bbox": [ + 271, + 189, + 480, + 319 + ], + "lines": [ + { + "bbox": [ + 271, + 189, + 480, + 319 + ], + "spans": [ + { + "bbox": [ + 271, + 189, + 480, + 319 + ], + "type": "image", + "image_path": "d910bd2d0f40de2b7791e9e805efde4c4695c9c3c911f9b850789508a6ca9060.jpg" + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 269, + 328, + 482, + 395 + ], + "lines": [ + { + "bbox": [ + 269, + 328, + 482, + 395 + ], + "spans": [ + { + "bbox": [ + 269, + 328, + 482, + 395 + ], + "type": "text", + "content": "Fig.2: Comparisons between simply combining ToMe [2] and LoRA [24] and our proposed PYRA. Red boxes represent the performance drop problem in low compression rates. Green boxes represent the adverse compression in high compression rates. See Sec. 4 for more results." + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "image_caption" + } + ], + "index": 4 + }, + { + "bbox": [ + 130, + 427, + 482, + 594 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 427, + 482, + 594 + ], + "spans": [ + { + "bbox": [ + 130, + 427, + 482, + 594 + ], + "type": "text", + "content": "heavy re-training stage is unaffordable. Consequently, simply combining PEFT and model compression can easily suffer from substantial performance drops. For instance, we could integrate LoRA [24], a notable PEFT method that introduces a small low-rank adapter, with ToMe [2], a parameter-free model compression technique for vision transformers5. As shown in Fig. 2, under lower compression rates (around " + }, + { + "bbox": [ + 130, + 427, + 482, + 594 + ], + "type": "inline_equation", + "content": "1.7 \\times" + }, + { + "bbox": [ + 130, + 427, + 482, + 594 + ], + "type": "text", + "content": "), performance on both backbones (ViT-L/16 and ViT-B/16) present slight drops (" + }, + { + "bbox": [ + 130, + 427, + 482, + 594 + ], + "type": "inline_equation", + "content": "< 1\\%" + }, + { + "bbox": [ + 130, + 427, + 482, + 594 + ], + "type": "text", + "content": ") compared to directly conducting PEFT on the backbone, indicating that the ToMe+LoRA combination serves as a basic solution within lower compression rate range, yet performance improvements are also demanded. Under high compression rates (" + }, + { + "bbox": [ + 130, + 427, + 482, + 594 + ], + "type": "inline_equation", + "content": ">3.0 \\times" + }, + { + "bbox": [ + 130, + 427, + 482, + 594 + ], + "type": "text", + "content": "), the performance quickly drops and is even inferior to directly fine-tuning the small-scale backbone with corresponding throughput. We term this phenomenon as Adverse Compression. Both phenomena indicate that directly combining existing works cannot effectively address the new challenges." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 130, + 596, + 494, + 644 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 596, + 494, + 644 + ], + "spans": [ + { + "bbox": [ + 130, + 596, + 494, + 644 + ], + "type": "text", + "content": "In this paper, we propose a novel Parallel Yielding Re-Accivitation method (PYRA) designed for training-inference efficient task adaptation using vision transformers. Generally, the proposed PYRA follows the token merging paradigm [2, 8, 43] for inference efficiency. For effective task adaptation, we propose to mod-" + } + ] + } + ], + "index": 7 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 287, + 91, + 447, + 102 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 287, + 91, + 447, + 102 + ], + "spans": [ + { + "bbox": [ + 287, + 91, + 447, + 102 + ], + "type": "text", + "content": "PYRA: Parallel Yielding Re-Activation" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 474, + 92, + 480, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 474, + 92, + 480, + 100 + ], + "spans": [ + { + "bbox": [ + 474, + 92, + 480, + 100 + ], + "type": "text", + "content": "3" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 133, + 653, + 416, + 665 + ], + "type": "page_footnote", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 653, + 416, + 665 + ], + "spans": [ + { + "bbox": [ + 133, + 653, + 416, + 665 + ], + "type": "text", + "content": "In Sec. 4, we show that " + }, + { + "bbox": [ + 133, + 653, + 416, + 665 + ], + "type": "inline_equation", + "content": "\\mathrm{ToMe + LoRA}" + }, + { + "bbox": [ + 133, + 653, + 416, + 665 + ], + "type": "text", + "content": " is a neat and strong solution." + } + ] + } + ], + "index": 8 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 2 + }, + { + "para_blocks": [ + { + "bbox": [ + 130, + 116, + 482, + 212 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 116, + 482, + 212 + ], + "spans": [ + { + "bbox": [ + 130, + 116, + 482, + 212 + ], + "type": "text", + "content": "ulate token features from both feature tokens and feature channels and parallelly yield weights for tokens to be merged by small modulation weight generators. These parallel yielding weights can comprehensively perceive the data distribution in downstream tasks. They are then applied to features through re-activation, resulting in adaptive token modulation. Thanks to such a token modulation strategy, PYRA can adaptively calibrate the learned feature distribution for downstream tasks with low computational complexity, ultimately leading to effective training and efficient inference." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 130, + 212, + 482, + 342 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 212, + 482, + 342 + ], + "spans": [ + { + "bbox": [ + 130, + 212, + 482, + 342 + ], + "type": "text", + "content": "We conduct extensive experiments to verify the effectiveness of our PYRA. We show that, under a low compression rate with " + }, + { + "bbox": [ + 130, + 212, + 482, + 342 + ], + "type": "inline_equation", + "content": "\\sim 1.7\\times" + }, + { + "bbox": [ + 130, + 212, + 482, + 342 + ], + "type": "text", + "content": " speedup, PYRA introduces negligible performance drops. Under a high compression rate with " + }, + { + "bbox": [ + 130, + 212, + 482, + 342 + ], + "type": "inline_equation", + "content": ">3.0\\times" + }, + { + "bbox": [ + 130, + 212, + 482, + 342 + ], + "type": "text", + "content": " speedup, PYRA eliminates the adverse compression gap. We emphasize that considering the scenario of high compression rates is practical, given the substantial size of current transformers. Concurrently, the specific model sizes required by downstream applications may not match any publicly available models, thus requiring the acquisition through high compression rates while keeping the performance comparable. In this regard, our approach represents an effective method for obtaining small models in the absence of pre-trained parameters for smaller-scale models." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 146, + 343, + 370, + 355 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 146, + 343, + 370, + 355 + ], + "spans": [ + { + "bbox": [ + 146, + 343, + 370, + 355 + ], + "type": "text", + "content": "Overall, we summarize our contribution as follows." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 138, + 361, + 480, + 514 + ], + "type": "list", + "angle": 0, + "index": 8, + "blocks": [ + { + "bbox": [ + 138, + 361, + 480, + 396 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 361, + 480, + 396 + ], + "spans": [ + { + "bbox": [ + 138, + 361, + 480, + 396 + ], + "type": "text", + "content": "- We propose a novel challenge termed training-inference efficient task adaptation, in which the inference efficiency of large-scale transformers is escalated during parameter-efficient task adaptation." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 138, + 397, + 480, + 454 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 397, + 480, + 454 + ], + "spans": [ + { + "bbox": [ + 138, + 397, + 480, + 454 + ], + "type": "text", + "content": "- We propose PYRA for training-inference efficient task adaptation, which enhances the perception of feature distribution via token modulation. We generate parallel yielding decoupled weights to comprehensively perceive the feature distributions. We apply a re-activation strategy to modulate tokens to be merged for calibrated token features." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 138, + 456, + 480, + 514 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 456, + 480, + 514 + ], + "spans": [ + { + "bbox": [ + 138, + 456, + 480, + 514 + ], + "type": "text", + "content": "- Extensive experiments show that our PYRA outperforms all competing methods under both low compression rate and high compression rate. Further analysis shows that PYRA is effective across a series of different transformer backbones and model scales, well demonstrating the effectiveness and superiority of our method." + } + ] + } + ], + "index": 7 + } + ], + "sub_type": "text" + }, + { + "bbox": [ + 132, + 533, + 243, + 547 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 533, + 243, + 547 + ], + "spans": [ + { + "bbox": [ + 132, + 533, + 243, + 547 + ], + "type": "text", + "content": "2 Related Works" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 130, + 557, + 482, + 666 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 557, + 482, + 666 + ], + "spans": [ + { + "bbox": [ + 130, + 557, + 482, + 666 + ], + "type": "text", + "content": "Parameter-Efficient Fine-Tuning (PEFT) for Task Adaptation. Transferring large-scale transformers to downstream tasks has been a popular topic in computer vision [9, 17, 26, 28, 36, 60, 68, 69]. PEFT methods either locate a subset of parameters inside the model for fine-tuning [65], or inject human designed modules to the original model structure. Specifically, adapters [23, 28, 42, 45] are a type of MLP module with a bottleneck in the middle. Prompt-tuning [14, 18, 26, 32, 33, 39] insert learnable tokens to model input to generate task-specific outputs. SSF [36] tunes additional scaling and shifting parameters. LoRA [24], AdaptFormer [9], and Consolidator [17] add lightweight modules as" + } + ] + } + ], + "index": 10 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "spans": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "type": "text", + "content": "4" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 166, + 91, + 227, + 102 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 166, + 91, + 227, + 102 + ], + "spans": [ + { + "bbox": [ + 166, + 91, + 227, + 102 + ], + "type": "text", + "content": "Y. Xiong et al." + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 3 + }, + { + "para_blocks": [ + { + "bbox": [ + 130, + 113, + 482, + 478 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 113, + 482, + 478 + ], + "spans": [ + { + "bbox": [ + 130, + 113, + 482, + 478 + ], + "type": "text", + "content": "bypasses. These modules can be merged with the original backbone for no extra inference cost. In the scope of training-inference efficient task adaptation, PEFT methods retain model inference cost, resulting in extreme difficulty for model deployment. To solve the problem, we conduct adaptive token merging for PEFT. While achieving promising adaptation performance under both low and high compression rates, our method inherits the advantages of PEFT methods. Model Compression. Model compression is often applied on large-scale models to acquire smaller-scale models with comparable performance. Mainstream approaches of model compression include model pruning [2,8,10,54,56,64,71], knowledge distillation [1,21,52] and model quantization [3,13,37,38]. For transformer models, model pruning methods can be roughly grouped into two categories: channel pruning and token pruning. Channel pruning methods [5,10, 64,73] reduce the number of parameters, channels, heads, or blocks. Recently, token pruning has emerged as another mainstream approach. Several works have attempted to prune tokens for vision transformers (ViTs). Among these methods, token pooling [43] uses a slow k-means approach that does not work for an off-the-shelf model. ToMe [2] constructs bipartite graphs and merges token pairs with the most weighted connections. DiffRate [8] combines token pruning and token merging with searched optimal token reduction rates for each layer. Although achieving promising results, most existing model compression methods fail when combining with PEFT for training-inference efficient task adaptation since they usually involve a heavy training stage. Model pruning methods demand full re-training after pruning to restore performance. Knowledge distillation necessitates training a small-scale model from scratch, demanding large amount of data. As for model quantization, although no heavy training is demanded, post-training quantization [41,63] achieves poor performance compared to quantization-aware training [25,34] that demands full re-training on the quantized model. As a comparison, our proposed PYRA achieves promising performance on compressed models under the restrictions of PEFT, surpassing the baselines of simply combining model compression and PEFT." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 132, + 494, + 233, + 508 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 494, + 233, + 508 + ], + "spans": [ + { + "bbox": [ + 132, + 494, + 233, + 508 + ], + "type": "text", + "content": "3 Methodology" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 132, + 519, + 228, + 531 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 519, + 228, + 531 + ], + "spans": [ + { + "bbox": [ + 132, + 519, + 228, + 531 + ], + "type": "text", + "content": "3.1 Preliminaries" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 130, + 539, + 483, + 668 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 539, + 483, + 668 + ], + "spans": [ + { + "bbox": [ + 130, + 539, + 483, + 668 + ], + "type": "text", + "content": "ViT Model. In this paper, we mainly focus on the training-inference efficient task adaptation of ViT models [16, 20, 51]. A ViT model consists of " + }, + { + "bbox": [ + 130, + 539, + 483, + 668 + ], + "type": "inline_equation", + "content": "L" + }, + { + "bbox": [ + 130, + 539, + 483, + 668 + ], + "type": "text", + "content": " identical encoder blocks, each of which consists of a multi-head self-attention (MHSA) module and a feed-forward network (FFN). Formally, an input image " + }, + { + "bbox": [ + 130, + 539, + 483, + 668 + ], + "type": "inline_equation", + "content": "\\mathbf{x}" + }, + { + "bbox": [ + 130, + 539, + 483, + 668 + ], + "type": "text", + "content": " is reshaped and linear projected to " + }, + { + "bbox": [ + 130, + 539, + 483, + 668 + ], + "type": "inline_equation", + "content": "N" + }, + { + "bbox": [ + 130, + 539, + 483, + 668 + ], + "type": "text", + "content": " tokens " + }, + { + "bbox": [ + 130, + 539, + 483, + 668 + ], + "type": "inline_equation", + "content": "\\mathbf{x} = [t_1, t_2, \\dots, t_N]" + }, + { + "bbox": [ + 130, + 539, + 483, + 668 + ], + "type": "text", + "content": " in " + }, + { + "bbox": [ + 130, + 539, + 483, + 668 + ], + "type": "inline_equation", + "content": "D" + }, + { + "bbox": [ + 130, + 539, + 483, + 668 + ], + "type": "text", + "content": " dimensions. For simplicity, we omit the classification token ([CLS]) and the distillation token [51]. For encoder block " + }, + { + "bbox": [ + 130, + 539, + 483, + 668 + ], + "type": "inline_equation", + "content": "l" + }, + { + "bbox": [ + 130, + 539, + 483, + 668 + ], + "type": "text", + "content": ", we denote the input as " + }, + { + "bbox": [ + 130, + 539, + 483, + 668 + ], + "type": "inline_equation", + "content": "\\mathbf{x}^l = [t_1^l, t_2^l, \\dots, t_{N^{l-1}}^l] \\in \\mathbb{R}^{N^{l-1} \\times D}" + }, + { + "bbox": [ + 130, + 539, + 483, + 668 + ], + "type": "text", + "content": " and the output as " + }, + { + "bbox": [ + 130, + 539, + 483, + 668 + ], + "type": "inline_equation", + "content": "\\hat{\\mathbf{x}}^l = [\\hat{t}_1^l, \\hat{t}_2^l, \\dots, \\hat{t}_{N^l}^l] \\in \\mathbb{R}^{N^l \\times D}" + }, + { + "bbox": [ + 130, + 539, + 483, + 668 + ], + "type": "text", + "content": ". For MHSA, the input tokens are first processed by three FC layers to generate " + }, + { + "bbox": [ + 130, + 539, + 483, + 668 + ], + "type": "inline_equation", + "content": "Q, K" + }, + { + "bbox": [ + 130, + 539, + 483, + 668 + ], + "type": "text", + "content": ", and " + }, + { + "bbox": [ + 130, + 539, + 483, + 668 + ], + "type": "inline_equation", + "content": "V" + }, + { + "bbox": [ + 130, + 539, + 483, + 668 + ], + "type": "text", + "content": " matrices, and the output is calculated by Softmax " + }, + { + "bbox": [ + 130, + 539, + 483, + 668 + ], + "type": "inline_equation", + "content": "(\\frac{QK^T}{\\sqrt{D}}) V" + }, + { + "bbox": [ + 130, + 539, + 483, + 668 + ], + "type": "text", + "content": " before being projected by another FC" + } + ] + } + ], + "index": 5 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 287, + 91, + 447, + 102 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 287, + 91, + 447, + 102 + ], + "spans": [ + { + "bbox": [ + 287, + 91, + 447, + 102 + ], + "type": "text", + "content": "PYRA: Parallel Yielding Re-Activation" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 474, + 92, + 480, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 474, + 92, + 480, + 100 + ], + "spans": [ + { + "bbox": [ + 474, + 92, + 480, + 100 + ], + "type": "text", + "content": "5" + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 4 + }, + { + "para_blocks": [ + { + "bbox": [ + 130, + 116, + 479, + 140 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 116, + 479, + 140 + ], + "spans": [ + { + "bbox": [ + 130, + 116, + 479, + 140 + ], + "type": "text", + "content": "layer. For FFN, the tokens are projected by two FC layers. Our method mainly focuses on the input tokens " + }, + { + "bbox": [ + 130, + 116, + 479, + 140 + ], + "type": "inline_equation", + "content": "\\mathbf{x}^l" + }, + { + "bbox": [ + 130, + 116, + 479, + 140 + ], + "type": "text", + "content": " before feeding them to the MHSA module." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 130, + 140, + 480, + 212 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 140, + 480, + 212 + ], + "spans": [ + { + "bbox": [ + 130, + 140, + 480, + 212 + ], + "type": "text", + "content": "LoRA. LoRA [24] is a widely employed PEFT method for task adaptation. Vision transformers consist of large dense parameter matrices. When adapting to a specific task, the updates to the matrices are in small subspaces and can be modeled with low-rank decompositions. LoRA trains only the decomposed matrices during fine-tuning. Specifically, for dense matrix " + }, + { + "bbox": [ + 130, + 140, + 480, + 212 + ], + "type": "inline_equation", + "content": "W_0 \\in \\mathbb{R}^{d \\times k}" + }, + { + "bbox": [ + 130, + 140, + 480, + 212 + ], + "type": "text", + "content": " and input " + }, + { + "bbox": [ + 130, + 140, + 480, + 212 + ], + "type": "inline_equation", + "content": "x" + }, + { + "bbox": [ + 130, + 140, + 480, + 212 + ], + "type": "text", + "content": ", the modified forward pass of updated " + }, + { + "bbox": [ + 130, + 140, + 480, + 212 + ], + "type": "inline_equation", + "content": "W_0" + }, + { + "bbox": [ + 130, + 140, + 480, + 212 + ], + "type": "text", + "content": " is:" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 269, + 221, + 480, + 233 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 269, + 221, + 480, + 233 + ], + "spans": [ + { + "bbox": [ + 269, + 221, + 480, + 233 + ], + "type": "interline_equation", + "content": "\\hat {x} = W _ {0} x + B A x, \\tag {1}", + "image_path": "20ae2f2b771d242f4af0275d2bef3aef26069d0c493f29771c8442b3383dddc7.jpg" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 130, + 239, + 479, + 264 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 239, + 479, + 264 + ], + "spans": [ + { + "bbox": [ + 130, + 239, + 479, + 264 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 130, + 239, + 479, + 264 + ], + "type": "inline_equation", + "content": "B \\in \\mathbb{R}^{d \\times h}" + }, + { + "bbox": [ + 130, + 239, + 479, + 264 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 130, + 239, + 479, + 264 + ], + "type": "inline_equation", + "content": "A \\in \\mathbb{R}^{h \\times k}" + }, + { + "bbox": [ + 130, + 239, + 479, + 264 + ], + "type": "text", + "content": ". During inference, " + }, + { + "bbox": [ + 130, + 239, + 479, + 264 + ], + "type": "inline_equation", + "content": "BA" + }, + { + "bbox": [ + 130, + 239, + 479, + 264 + ], + "type": "text", + "content": " can be merged with " + }, + { + "bbox": [ + 130, + 239, + 479, + 264 + ], + "type": "inline_equation", + "content": "W_0" + }, + { + "bbox": [ + 130, + 239, + 479, + 264 + ], + "type": "text", + "content": " for no extra computation overhead." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 130, + 265, + 479, + 312 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 265, + 479, + 312 + ], + "spans": [ + { + "bbox": [ + 130, + 265, + 479, + 312 + ], + "type": "text", + "content": "Token Merging. Token merging [2,8] is a parameter-free compression technique for vision transformers. It is orthogonal to the transformer structure and capable of flexibly changing the compression rate. Specifically, the input tokens " + }, + { + "bbox": [ + 130, + 265, + 479, + 312 + ], + "type": "inline_equation", + "content": "\\mathbf{x}^l" + }, + { + "bbox": [ + 130, + 265, + 479, + 312 + ], + "type": "text", + "content": " of the " + }, + { + "bbox": [ + 130, + 265, + 479, + 312 + ], + "type": "inline_equation", + "content": "l" + }, + { + "bbox": [ + 130, + 265, + 479, + 312 + ], + "type": "text", + "content": "-th ViT block are randomly separated before the MHSA:" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 176, + 320, + 480, + 335 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 176, + 320, + 480, + 335 + ], + "spans": [ + { + "bbox": [ + 176, + 320, + 480, + 335 + ], + "type": "interline_equation", + "content": "G _ {1} ^ {l} = \\left[ t _ {i _ {1}} ^ {l}, t _ {i _ {2}} ^ {l}, \\dots , t _ {i _ {s}} ^ {l} \\right], \\quad G _ {2} ^ {l} = \\left[ t _ {j _ {1}} ^ {l}, t _ {j _ {2}} ^ {l}, \\dots , t _ {j _ {s}} ^ {l} \\right], \\quad 2 s = N ^ {l - 1}. \\tag {2}", + "image_path": "7d9a7f77e42b171d61d89dcba00f3811a3859d01834f5d57d12a40720c21c858.jpg" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 130, + 342, + 479, + 392 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 342, + 479, + 392 + ], + "spans": [ + { + "bbox": [ + 130, + 342, + 479, + 392 + ], + "type": "text", + "content": "Then, for each " + }, + { + "bbox": [ + 130, + 342, + 479, + 392 + ], + "type": "inline_equation", + "content": "t_{i_\\bullet}^l" + }, + { + "bbox": [ + 130, + 342, + 479, + 392 + ], + "type": "text", + "content": " token, the most similar " + }, + { + "bbox": [ + 130, + 342, + 479, + 392 + ], + "type": "inline_equation", + "content": "t_{j_\\bullet}^l" + }, + { + "bbox": [ + 130, + 342, + 479, + 392 + ], + "type": "text", + "content": " token is matched to it via cosine similarity. From " + }, + { + "bbox": [ + 130, + 342, + 479, + 392 + ], + "type": "inline_equation", + "content": "G_1^l" + }, + { + "bbox": [ + 130, + 342, + 479, + 392 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 130, + 342, + 479, + 392 + ], + "type": "inline_equation", + "content": "r" + }, + { + "bbox": [ + 130, + 342, + 479, + 392 + ], + "type": "text", + "content": " tokens with the most similar connections to tokens in " + }, + { + "bbox": [ + 130, + 342, + 479, + 392 + ], + "type": "inline_equation", + "content": "G_2^l" + }, + { + "bbox": [ + 130, + 342, + 479, + 392 + ], + "type": "text", + "content": " are selected to form token pairs " + }, + { + "bbox": [ + 130, + 342, + 479, + 392 + ], + "type": "inline_equation", + "content": "(t_{m_k}^l, t_{n_k}^l)" + }, + { + "bbox": [ + 130, + 342, + 479, + 392 + ], + "type": "text", + "content": ", where " + }, + { + "bbox": [ + 130, + 342, + 479, + 392 + ], + "type": "inline_equation", + "content": "t_{m_k}^l \\in G_1^l" + }, + { + "bbox": [ + 130, + 342, + 479, + 392 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 130, + 342, + 479, + 392 + ], + "type": "inline_equation", + "content": "t_{n_k}^l \\in G_2^l" + }, + { + "bbox": [ + 130, + 342, + 479, + 392 + ], + "type": "text", + "content": ", and " + }, + { + "bbox": [ + 130, + 342, + 479, + 392 + ], + "type": "inline_equation", + "content": "k = 1, 2, \\dots, r" + }, + { + "bbox": [ + 130, + 342, + 479, + 392 + ], + "type": "text", + "content": ". Note that " + }, + { + "bbox": [ + 130, + 342, + 479, + 392 + ], + "type": "inline_equation", + "content": "\\{m_k\\}" + }, + { + "bbox": [ + 130, + 342, + 479, + 392 + ], + "type": "text", + "content": " is a re-indexed subset of " + }, + { + "bbox": [ + 130, + 342, + 479, + 392 + ], + "type": "inline_equation", + "content": "\\{i_\\bullet\\}" + }, + { + "bbox": [ + 130, + 342, + 479, + 392 + ], + "type": "text", + "content": ". Formally," + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 166, + 400, + 480, + 431 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 166, + 400, + 480, + 431 + ], + "spans": [ + { + "bbox": [ + 166, + 400, + 480, + 431 + ], + "type": "interline_equation", + "content": "\\left\\{t _ {m _ {k}} ^ {l} \\right\\} = \\arg \\operatorname {T o p r} _ {\\hat {t} _ {i} ^ {l}} \\left(\\max _ {j} \\frac {\\hat {t} _ {i} ^ {l} \\cdot t _ {j} ^ {l}}{\\| \\hat {t} _ {i} ^ {l} \\| \\cdot \\| t _ {j} ^ {l} \\|}\\right), \\quad t _ {n _ {k}} ^ {l} = \\arg \\max _ {t _ {j} ^ {l}} \\left(\\frac {t _ {m _ {k}} ^ {l} \\cdot t _ {j} ^ {l}}{\\| t _ {m _ {k}} ^ {l} \\| \\cdot \\| t _ {j} ^ {l} \\|}\\right) \\tag {3}", + "image_path": "0c85c5bb527248cf9087f4a905de80d0f2095c74483f150493034c9b5299ccad.jpg" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 130, + 438, + 479, + 462 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 438, + 479, + 462 + ], + "spans": [ + { + "bbox": [ + 130, + 438, + 479, + 462 + ], + "type": "text", + "content": "Previous works [2, 8, 43], generally merge tokens via average pooling, cutting down the computational cost by decreasing the token number by " + }, + { + "bbox": [ + 130, + 438, + 479, + 462 + ], + "type": "inline_equation", + "content": "r" + }, + { + "bbox": [ + 130, + 438, + 479, + 462 + ], + "type": "text", + "content": " in layer " + }, + { + "bbox": [ + 130, + 438, + 479, + 462 + ], + "type": "inline_equation", + "content": "l" + }, + { + "bbox": [ + 130, + 438, + 479, + 462 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 130, + 463, + 480, + 559 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 463, + 480, + 559 + ], + "spans": [ + { + "bbox": [ + 130, + 463, + 480, + 559 + ], + "type": "text", + "content": "We employ the above techniques as our baseline method, in which we attach token merging while fine-tuning LoRA for task adaptation. These methods are selected due to their advantages being highly compatible with training-inference efficient task adaptation. First, token merging is parameter-free and training-free, and does not change model structures, which retains the storage-efficient advantage of PEFT. Besides, we choose LoRA for its popularity, simplicity, and mergeability (not introducing extra FLOPs during inference). Experimental results show that " + }, + { + "bbox": [ + 130, + 463, + 480, + 559 + ], + "type": "inline_equation", + "content": "\\mathrm{ToMe + LoRA}" + }, + { + "bbox": [ + 130, + 463, + 480, + 559 + ], + "type": "text", + "content": " is a strong baseline (see Sec. 4)." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 131, + 574, + 359, + 586 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 131, + 574, + 359, + 586 + ], + "spans": [ + { + "bbox": [ + 131, + 574, + 359, + 586 + ], + "type": "text", + "content": "3.2 PYRA: Parallel Yielding Re-Activation" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 130, + 594, + 479, + 666 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 594, + 479, + 666 + ], + "spans": [ + { + "bbox": [ + 130, + 594, + 479, + 666 + ], + "type": "text", + "content": "Conventional fine-tuning methods excel at guiding the model to dynamically align with the target data distribution in downstream tasks by extensively adjusting parameters. However, in the context of training-inference efficient task adaptation, only a fraction of parameters can be fine-tuned, posing significant challenges in accurately capturing the nuances of data distribution. While reducing model complexity through token merging shows promise in enhancing" + } + ] + } + ], + "index": 13 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "spans": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "type": "text", + "content": "6" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 166, + 91, + 227, + 102 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 166, + 91, + 227, + 102 + ], + "spans": [ + { + "bbox": [ + 166, + 91, + 227, + 102 + ], + "type": "text", + "content": "Y. Xiong et al." + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 5 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 134, + 114, + 479, + 224 + ], + "blocks": [ + { + "bbox": [ + 134, + 114, + 479, + 224 + ], + "lines": [ + { + "bbox": [ + 134, + 114, + 479, + 224 + ], + "spans": [ + { + "bbox": [ + 134, + 114, + 479, + 224 + ], + "type": "image", + "image_path": "9cfda996373e5beb11405400e7f727dd2c70b99c443e8b30987a29f36a116397.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 130, + 233, + 480, + 289 + ], + "lines": [ + { + "bbox": [ + 130, + 233, + 480, + 289 + ], + "spans": [ + { + "bbox": [ + 130, + 233, + 480, + 289 + ], + "type": "text", + "content": "Fig. 3: The pipeline of our PYRA. PYRA conducts token modulation before the MHSA module in each transformer block. Inside PYRA, a pair of learnable modulation weight generators are leveraged to generate adaptive modulation weights parallelly. After that, generated weights modulate tokens through re-activation. The generators in PYRA can be trained along with the LoRA module " + }, + { + "bbox": [ + 130, + 233, + 480, + 289 + ], + "type": "inline_equation", + "content": "\\theta_{p}" + }, + { + "bbox": [ + 130, + 233, + 480, + 289 + ], + "type": "text", + "content": " in an end-to-end manner." + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_caption" + } + ], + "index": 2 + }, + { + "bbox": [ + 130, + 315, + 479, + 386 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 315, + 479, + 386 + ], + "spans": [ + { + "bbox": [ + 130, + 315, + 479, + 386 + ], + "type": "text", + "content": "inference efficiency, it introduces the risk of information loss during layer-wise processing in vision transformers. This loss is difficult to rectify due to the constrained understanding of data distribution in the efficient task adaptation scenario. Therefore, a straightforward combination of token merging with PEFT algorithms for achieving training-inference efficient task adaptation may not yield optimal results, as depicted in Fig. 2." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 130, + 387, + 480, + 471 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 387, + 480, + 471 + ], + "spans": [ + { + "bbox": [ + 130, + 387, + 480, + 471 + ], + "type": "text", + "content": "Here, we propose Parallel Yielding Re-Aviation (PYRA) to adaptively modulate token features to enhance the perception of data distribution during token merging. Specifically, inside PYRA, the weights for adaptive merging are first yielded in a parallel manner through a pair of lightweight learnable vectors in each ViT block. These generated weights are then applied to modulate tokens to be merged through re-activation. As a result, PYRA enables adaptive calibration of the learned feature distribution with low computational complexity." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 130, + 472, + 481, + 616 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 472, + 481, + 616 + ], + "spans": [ + { + "bbox": [ + 130, + 472, + 481, + 616 + ], + "type": "text", + "content": "Parallel Yielding Adaptive Weights. We aim to optimize the merging process of each chosen token pair. Inspired by feature modulation [44, 59], we propose to modulate token features before merging. Formally, for encoder block " + }, + { + "bbox": [ + 130, + 472, + 481, + 616 + ], + "type": "inline_equation", + "content": "l" + }, + { + "bbox": [ + 130, + 472, + 481, + 616 + ], + "type": "text", + "content": " with " + }, + { + "bbox": [ + 130, + 472, + 481, + 616 + ], + "type": "inline_equation", + "content": "r" + }, + { + "bbox": [ + 130, + 472, + 481, + 616 + ], + "type": "text", + "content": " pairs of tokens in " + }, + { + "bbox": [ + 130, + 472, + 481, + 616 + ], + "type": "inline_equation", + "content": "D" + }, + { + "bbox": [ + 130, + 472, + 481, + 616 + ], + "type": "text", + "content": " dimensions to be merged, we group the " + }, + { + "bbox": [ + 130, + 472, + 481, + 616 + ], + "type": "inline_equation", + "content": "t_{m_k}^l" + }, + { + "bbox": [ + 130, + 472, + 481, + 616 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 130, + 472, + 481, + 616 + ], + "type": "inline_equation", + "content": "t_{n_k}^l" + }, + { + "bbox": [ + 130, + 472, + 481, + 616 + ], + "type": "text", + "content": " tokens as token matrices " + }, + { + "bbox": [ + 130, + 472, + 481, + 616 + ], + "type": "inline_equation", + "content": "M_s^l = [t_{m_1}^l,\\dots ,t_{m_r}^l ]" + }, + { + "bbox": [ + 130, + 472, + 481, + 616 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 130, + 472, + 481, + 616 + ], + "type": "inline_equation", + "content": "M_t^l = [t_{n_1}^l,\\dots ,t_{n_r}^l ]" + }, + { + "bbox": [ + 130, + 472, + 481, + 616 + ], + "type": "text", + "content": ", where " + }, + { + "bbox": [ + 130, + 472, + 481, + 616 + ], + "type": "inline_equation", + "content": "M_s^l,M_t^l\\in \\mathbb{R}^{D\\times r}" + }, + { + "bbox": [ + 130, + 472, + 481, + 616 + ], + "type": "text", + "content": ". We learn a modulation matrix " + }, + { + "bbox": [ + 130, + 472, + 481, + 616 + ], + "type": "inline_equation", + "content": "W^{l}\\in \\mathbb{R}^{D\\times r}" + }, + { + "bbox": [ + 130, + 472, + 481, + 616 + ], + "type": "text", + "content": " that adaptively modulates tokens at the granularity of each channel. We emphasize that directly learning " + }, + { + "bbox": [ + 130, + 472, + 481, + 616 + ], + "type": "inline_equation", + "content": "W^{l}" + }, + { + "bbox": [ + 130, + 472, + 481, + 616 + ], + "type": "text", + "content": " is redundant [24] and cannot adaptively satisfy the conditions with different images and token pairs. Therefore, we further specify the goal as learning decoupled weights " + }, + { + "bbox": [ + 130, + 472, + 481, + 616 + ], + "type": "inline_equation", + "content": "\\delta_D^l = f_D(M_s^l,M_t^l)\\in \\mathbb{R}^{D\\times 1}" + }, + { + "bbox": [ + 130, + 472, + 481, + 616 + ], + "type": "text", + "content": " for feature channels and " + }, + { + "bbox": [ + 130, + 472, + 481, + 616 + ], + "type": "inline_equation", + "content": "\\delta_r^l = f_r(M_s^l,M_t^l)\\in \\mathbb{R}^{1\\times r}" + }, + { + "bbox": [ + 130, + 472, + 481, + 616 + ], + "type": "text", + "content": " for feature tokens, where " + }, + { + "bbox": [ + 130, + 472, + 481, + 616 + ], + "type": "inline_equation", + "content": "W^{l} = \\delta_{D}^{l}\\delta_{r}^{l}" + }, + { + "bbox": [ + 130, + 472, + 481, + 616 + ], + "type": "text", + "content": ", separately. We propose to generate " + }, + { + "bbox": [ + 130, + 472, + 481, + 616 + ], + "type": "inline_equation", + "content": "\\delta_D^l" + }, + { + "bbox": [ + 130, + 472, + 481, + 616 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 130, + 472, + 481, + 616 + ], + "type": "inline_equation", + "content": "\\delta_r^l" + }, + { + "bbox": [ + 130, + 472, + 481, + 616 + ], + "type": "text", + "content": " in a parallel yielding manner." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 130, + 617, + 481, + 666 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 617, + 481, + 666 + ], + "spans": [ + { + "bbox": [ + 130, + 617, + 481, + 666 + ], + "type": "text", + "content": "Specifically, for layer " + }, + { + "bbox": [ + 130, + 617, + 481, + 666 + ], + "type": "inline_equation", + "content": "l" + }, + { + "bbox": [ + 130, + 617, + 481, + 666 + ], + "type": "text", + "content": " with " + }, + { + "bbox": [ + 130, + 617, + 481, + 666 + ], + "type": "inline_equation", + "content": "r" + }, + { + "bbox": [ + 130, + 617, + 481, + 666 + ], + "type": "text", + "content": " tokens to be merged, we create two learnable vectors as the modulation weight generator inside the transformer block: " + }, + { + "bbox": [ + 130, + 617, + 481, + 666 + ], + "type": "inline_equation", + "content": "W_r^l \\in \\mathbb{R}^{r \\times 1}" + }, + { + "bbox": [ + 130, + 617, + 481, + 666 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 130, + 617, + 481, + 666 + ], + "type": "inline_equation", + "content": "W_D^l \\in \\mathbb{R}^{1 \\times D}" + }, + { + "bbox": [ + 130, + 617, + 481, + 666 + ], + "type": "text", + "content": ". To guarantee that " + }, + { + "bbox": [ + 130, + 617, + 481, + 666 + ], + "type": "inline_equation", + "content": "W_r^l" + }, + { + "bbox": [ + 130, + 617, + 481, + 666 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 130, + 617, + 481, + 666 + ], + "type": "inline_equation", + "content": "W_D^l" + }, + { + "bbox": [ + 130, + 617, + 481, + 666 + ], + "type": "text", + "content": " digest the token features from both tokens in a token pair, we first normalize the sum of token pairs to" + } + ] + } + ], + "index": 7 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 287, + 91, + 447, + 102 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 287, + 91, + 447, + 102 + ], + "spans": [ + { + "bbox": [ + 287, + 91, + 447, + 102 + ], + "type": "text", + "content": "PYRA: Parallel Yielding Re-Activation" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 474, + 92, + 480, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 474, + 92, + 480, + 100 + ], + "spans": [ + { + "bbox": [ + 474, + 92, + 480, + 100 + ], + "type": "text", + "content": "7" + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 6 + }, + { + "para_blocks": [ + { + "bbox": [ + 132, + 116, + 309, + 127 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 116, + 309, + 127 + ], + "spans": [ + { + "bbox": [ + 132, + 116, + 309, + 127 + ], + "type": "text", + "content": "construct the token information matrix:" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 220, + 137, + 481, + 152 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 220, + 137, + 481, + 152 + ], + "spans": [ + { + "bbox": [ + 220, + 137, + 481, + 152 + ], + "type": "interline_equation", + "content": "M _ {\\text {i n f o}} ^ {l} = \\operatorname {L a y e r N o r m} \\left(M _ {s} ^ {l} + M _ {t} ^ {l}\\right) \\in \\mathbb {R} ^ {D \\times r}. \\tag {4}", + "image_path": "9c2772c36e3ed43adcacdba6c885846fd6ea53eb7a93f45b054929e525a9e799.jpg" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 131, + 159, + 481, + 196 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 131, + 159, + 481, + 196 + ], + "spans": [ + { + "bbox": [ + 131, + 159, + 481, + 196 + ], + "type": "text", + "content": "We normalize the distribution of " + }, + { + "bbox": [ + 131, + 159, + 481, + 196 + ], + "type": "inline_equation", + "content": "M_{\\mathrm{info}}^{l}" + }, + { + "bbox": [ + 131, + 159, + 481, + 196 + ], + "type": "text", + "content": " tokens by leveraging the LayerNorm(\\cdot) operation to enable smoother gradients when training " + }, + { + "bbox": [ + 131, + 159, + 481, + 196 + ], + "type": "inline_equation", + "content": "W_{r}^{l}" + }, + { + "bbox": [ + 131, + 159, + 481, + 196 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 131, + 159, + 481, + 196 + ], + "type": "inline_equation", + "content": "W_{D}^{l}" + }, + { + "bbox": [ + 131, + 159, + 481, + 196 + ], + "type": "text", + "content": ". With the token information matrix, we then yield the adaptive weights " + }, + { + "bbox": [ + 131, + 159, + 481, + 196 + ], + "type": "inline_equation", + "content": "\\delta_{D}^{l}" + }, + { + "bbox": [ + 131, + 159, + 481, + 196 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 131, + 159, + 481, + 196 + ], + "type": "inline_equation", + "content": "\\delta_r^l" + }, + { + "bbox": [ + 131, + 159, + 481, + 196 + ], + "type": "text", + "content": " parallelly:" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 256, + 204, + 481, + 225 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 256, + 204, + 481, + 225 + ], + "spans": [ + { + "bbox": [ + 256, + 204, + 481, + 225 + ], + "type": "interline_equation", + "content": "\\delta_ {D} ^ {l} = M _ {\\text {i n f o}} ^ {l} W _ {r} ^ {l} \\in \\mathbb {R} ^ {D \\times 1} \\tag {5}", + "image_path": "3398af4776049a43589322259ab67247e54424cbf06d835d52a7af9c89adc859.jpg" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 259, + 220, + 359, + 235 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 259, + 220, + 359, + 235 + ], + "spans": [ + { + "bbox": [ + 259, + 220, + 359, + 235 + ], + "type": "interline_equation", + "content": "\\delta_ {r} ^ {l} = W _ {D} ^ {l} M _ {\\mathrm {i n f o}} ^ {l} \\in \\mathbb {R} ^ {1 \\times r}.", + "image_path": "f3d51405fcf4c659efe6c33d732dc6e1c7aac936dc956524c78fff61a936b95c.jpg" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 131, + 239, + 482, + 350 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 131, + 239, + 482, + 350 + ], + "spans": [ + { + "bbox": [ + 131, + 239, + 482, + 350 + ], + "type": "text", + "content": "Re-Activation for Token Modulation. Simply generating " + }, + { + "bbox": [ + 131, + 239, + 482, + 350 + ], + "type": "inline_equation", + "content": "\\delta_D^l" + }, + { + "bbox": [ + 131, + 239, + 482, + 350 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 131, + 239, + 482, + 350 + ], + "type": "inline_equation", + "content": "\\delta_r^l" + }, + { + "bbox": [ + 131, + 239, + 482, + 350 + ], + "type": "text", + "content": " via matrix multiplication still faces several possible issues. First, no measures have taken to ensure that " + }, + { + "bbox": [ + 131, + 239, + 482, + 350 + ], + "type": "inline_equation", + "content": "\\delta_D^l" + }, + { + "bbox": [ + 131, + 239, + 482, + 350 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 131, + 239, + 482, + 350 + ], + "type": "inline_equation", + "content": "\\delta_r^l" + }, + { + "bbox": [ + 131, + 239, + 482, + 350 + ], + "type": "text", + "content": " stay in a normal range. Second, decoupling weights to feature tokens and feature channels results in a low-rank modulation weight matrix " + }, + { + "bbox": [ + 131, + 239, + 482, + 350 + ], + "type": "inline_equation", + "content": "W^l" + }, + { + "bbox": [ + 131, + 239, + 482, + 350 + ], + "type": "text", + "content": ", which exhibits limited expressive capacity and thus could be unable to optimally modulate tokens in complicated data distributions. To cope with these issues, we conduct token modulation in a re-activation strategy. Specifically, we first broadcast " + }, + { + "bbox": [ + 131, + 239, + 482, + 350 + ], + "type": "inline_equation", + "content": "\\delta_D^l" + }, + { + "bbox": [ + 131, + 239, + 482, + 350 + ], + "type": "text", + "content": " to " + }, + { + "bbox": [ + 131, + 239, + 482, + 350 + ], + "type": "inline_equation", + "content": "\\hat{\\delta}_D^l \\in \\mathbb{R}^{D \\times r}" + }, + { + "bbox": [ + 131, + 239, + 482, + 350 + ], + "type": "text", + "content": " and conduct sigmoid activation " + }, + { + "bbox": [ + 131, + 239, + 482, + 350 + ], + "type": "inline_equation", + "content": "\\sigma(\\cdot)" + }, + { + "bbox": [ + 131, + 239, + 482, + 350 + ], + "type": "text", + "content": " on it, and then modulate " + }, + { + "bbox": [ + 131, + 239, + 482, + 350 + ], + "type": "inline_equation", + "content": "M_s^l" + }, + { + "bbox": [ + 131, + 239, + 482, + 350 + ], + "type": "text", + "content": " for an intermediate modulation result:" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 264, + 357, + 481, + 372 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 264, + 357, + 481, + 372 + ], + "spans": [ + { + "bbox": [ + 264, + 357, + 481, + 372 + ], + "type": "interline_equation", + "content": "\\hat {M} _ {s} ^ {l} = 2 \\sigma \\left(\\hat {\\delta} _ {D} ^ {l}\\right) \\odot M _ {s} ^ {l}. \\tag {6}", + "image_path": "b4d5b3a444639cc02ffd209983cfabbe3a904d030a562160272533c337488099.jpg" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 131, + 379, + 481, + 406 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 131, + 379, + 481, + 406 + ], + "spans": [ + { + "bbox": [ + 131, + 379, + 481, + 406 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 131, + 379, + 481, + 406 + ], + "type": "inline_equation", + "content": "\\odot" + }, + { + "bbox": [ + 131, + 379, + 481, + 406 + ], + "type": "text", + "content": " is Hadamard product. " + }, + { + "bbox": [ + 131, + 379, + 481, + 406 + ], + "type": "inline_equation", + "content": "\\hat{M}_s^l" + }, + { + "bbox": [ + 131, + 379, + 481, + 406 + ], + "type": "text", + "content": " is then modulated with sigmoid-activated broadcast weight " + }, + { + "bbox": [ + 131, + 379, + 481, + 406 + ], + "type": "inline_equation", + "content": "\\hat{\\delta}_r^l\\in \\mathbb{R}^{D\\times r}" + }, + { + "bbox": [ + 131, + 379, + 481, + 406 + ], + "type": "text", + "content": " again to acquire the modulated tokens:" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 239, + 415, + 481, + 429 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 239, + 415, + 481, + 429 + ], + "spans": [ + { + "bbox": [ + 239, + 415, + 481, + 429 + ], + "type": "interline_equation", + "content": "M _ {s} ^ {l} \\leftarrow M _ {s} ^ {l} + \\left(2 \\sigma \\left(\\hat {\\delta} _ {r} ^ {l}\\right) - 1\\right) \\odot \\hat {M} _ {s} ^ {l}, \\tag {7}", + "image_path": "9f530329eaf20e0f8b2d981f0fd6b712ec6b4055509c02ba8bd4df8a3f0caccd.jpg" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 131, + 437, + 482, + 521 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 131, + 437, + 482, + 521 + ], + "spans": [ + { + "bbox": [ + 131, + 437, + 482, + 521 + ], + "type": "text", + "content": "Note that we use the original tokens " + }, + { + "bbox": [ + 131, + 437, + 482, + 521 + ], + "type": "inline_equation", + "content": "M_{s}^{l}" + }, + { + "bbox": [ + 131, + 437, + 482, + 521 + ], + "type": "text", + "content": " to create a residual connection that preserves the gradient flow during training. The modulated " + }, + { + "bbox": [ + 131, + 437, + 482, + 521 + ], + "type": "inline_equation", + "content": "M_{s}^{l}" + }, + { + "bbox": [ + 131, + 437, + 482, + 521 + ], + "type": "text", + "content": " tokens are then merged with " + }, + { + "bbox": [ + 131, + 437, + 482, + 521 + ], + "type": "inline_equation", + "content": "M_{t}^{l}" + }, + { + "bbox": [ + 131, + 437, + 482, + 521 + ], + "type": "text", + "content": " with average pooling. We use a random Gaussian initialization for generator " + }, + { + "bbox": [ + 131, + 437, + 482, + 521 + ], + "type": "inline_equation", + "content": "W_{r}^{l}" + }, + { + "bbox": [ + 131, + 437, + 482, + 521 + ], + "type": "text", + "content": " and zero for generator " + }, + { + "bbox": [ + 131, + 437, + 482, + 521 + ], + "type": "inline_equation", + "content": "W_{D}^{l}" + }, + { + "bbox": [ + 131, + 437, + 482, + 521 + ], + "type": "text", + "content": ", so re-activation is equivalent to identity transformation at the beginning of training. Token modulation is conducted only on " + }, + { + "bbox": [ + 131, + 437, + 482, + 521 + ], + "type": "inline_equation", + "content": "M_{s}^{l}" + }, + { + "bbox": [ + 131, + 437, + 482, + 521 + ], + "type": "text", + "content": " to guarantee the parallelism in training and inference, as " + }, + { + "bbox": [ + 131, + 437, + 482, + 521 + ], + "type": "inline_equation", + "content": "t_{m_k}^l" + }, + { + "bbox": [ + 131, + 437, + 482, + 521 + ], + "type": "text", + "content": " tokens are distinct while different " + }, + { + "bbox": [ + 131, + 437, + 482, + 521 + ], + "type": "inline_equation", + "content": "t_{n_k}^l" + }, + { + "bbox": [ + 131, + 437, + 482, + 521 + ], + "type": "text", + "content": " might point to the same " + }, + { + "bbox": [ + 131, + 437, + 482, + 521 + ], + "type": "inline_equation", + "content": "t_j^l" + }, + { + "bbox": [ + 131, + 437, + 482, + 521 + ], + "type": "text", + "content": " token." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 131, + 522, + 482, + 666 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 131, + 522, + 482, + 666 + ], + "spans": [ + { + "bbox": [ + 131, + 522, + 482, + 666 + ], + "type": "text", + "content": "Discussion. Our PYRA effectively enhances the calibration of features for downstream tasks. Specifically, by utilizing parallel yielding adaptive weights, PYRA effectively decouples the modulation weight from the feature token and the feature channel, i.e., " + }, + { + "bbox": [ + 131, + 522, + 482, + 666 + ], + "type": "inline_equation", + "content": "\\delta_{D}" + }, + { + "bbox": [ + 131, + 522, + 482, + 666 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 131, + 522, + 482, + 666 + ], + "type": "inline_equation", + "content": "\\delta_{r}" + }, + { + "bbox": [ + 131, + 522, + 482, + 666 + ], + "type": "text", + "content": ", enabling comprehensive perception of the feature distribution in downstream tasks. Furthermore, considering the challenge of capturing an accurate feature distribution for parameter-efficient training, such smooth re-activations by Eq. (6) and Eq. (7) can constrain the negative impact on weight values brought by limited perception of feature distributions, thus leading to better token modulation for improving feature representations in downstream tasks. As a result, the proposed PYRA can maintain discriminative information to an utmost degree during token merging, leading to improved performance while achieving complexity reduction." + } + ] + } + ], + "index": 12 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "spans": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "type": "text", + "content": "8" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 166, + 91, + 228, + 102 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 166, + 91, + 228, + 102 + ], + "spans": [ + { + "bbox": [ + 166, + 91, + 228, + 102 + ], + "type": "text", + "content": "Y. Xiong et al." + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 7 + }, + { + "para_blocks": [ + { + "type": "table", + "bbox": [ + 159, + 146, + 455, + 209 + ], + "blocks": [ + { + "bbox": [ + 132, + 114, + 479, + 137 + ], + "lines": [ + { + "bbox": [ + 132, + 114, + 479, + 137 + ], + "spans": [ + { + "bbox": [ + 132, + 114, + 479, + 137 + ], + "type": "text", + "content": "Table 1: The complexity comparisons between conducting PEFT with and without PYRA. The FLOPs metric is obtained during inference." + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 159, + 146, + 455, + 209 + ], + "lines": [ + { + "bbox": [ + 159, + 146, + 455, + 209 + ], + "spans": [ + { + "bbox": [ + 159, + 146, + 455, + 209 + ], + "type": "table", + "html": "
ModelMetricTotalPEFT w/o PYRA (%)PEFT w. PYRA (%)
ViT-Base# params86M0.29M0.34%0.30M0.35%
FLOPs16.37G16.37G100%8.15G49.79%
ViT-Large# params303M1.18M0.39%1.20M0.40%
FLOPs57.37G57.37G100%28.76G50.13%
", + "image_path": "8752dbd7461abdb46e100d4b77830a9112ec29d03940d6723d44b919b97722f9.jpg" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "table_body" + } + ], + "index": 3 + }, + { + "bbox": [ + 132, + 229, + 264, + 242 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 229, + 264, + 242 + ], + "spans": [ + { + "bbox": [ + 132, + 229, + 264, + 242 + ], + "type": "text", + "content": "3.3 Complexity Analysis" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 132, + 247, + 480, + 306 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 247, + 480, + 306 + ], + "spans": [ + { + "bbox": [ + 132, + 247, + 480, + 306 + ], + "type": "text", + "content": "We present the parameter and computation complexity for introducing PYRA to PEFT for task adaptation. In each ViT block, PYRA introduces " + }, + { + "bbox": [ + 132, + 247, + 480, + 306 + ], + "type": "inline_equation", + "content": "D + r" + }, + { + "bbox": [ + 132, + 247, + 480, + 306 + ], + "type": "text", + "content": " training parameters and " + }, + { + "bbox": [ + 132, + 247, + 480, + 306 + ], + "type": "inline_equation", + "content": "4rD" + }, + { + "bbox": [ + 132, + 247, + 480, + 306 + ], + "type": "text", + "content": " extra FLOPs. For a ViT model with " + }, + { + "bbox": [ + 132, + 247, + 480, + 306 + ], + "type": "inline_equation", + "content": "L" + }, + { + "bbox": [ + 132, + 247, + 480, + 306 + ], + "type": "text", + "content": " layers and " + }, + { + "bbox": [ + 132, + 247, + 480, + 306 + ], + "type": "inline_equation", + "content": "R" + }, + { + "bbox": [ + 132, + 247, + 480, + 306 + ], + "type": "text", + "content": " total merged tokens, in total, our PYRA introduces " + }, + { + "bbox": [ + 132, + 247, + 480, + 306 + ], + "type": "inline_equation", + "content": "LD + R" + }, + { + "bbox": [ + 132, + 247, + 480, + 306 + ], + "type": "text", + "content": " extra training parameters and conducts " + }, + { + "bbox": [ + 132, + 247, + 480, + 306 + ], + "type": "inline_equation", + "content": "4RD" + }, + { + "bbox": [ + 132, + 247, + 480, + 306 + ], + "type": "text", + "content": " extra FLOPs beside PEFT." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 132, + 308, + 480, + 415 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 308, + 480, + 415 + ], + "spans": [ + { + "bbox": [ + 132, + 308, + 480, + 415 + ], + "type": "text", + "content": "To better demonstrate the training-inference efficiency of our PYRA, we compare the complexity between attaching PYRA to task adaptation via PEFT (here we employ LoRA [24]) and plainly conducting PEFT without PYRA. As shown in Tab. 1, PYRA keeps the training efficient feature of PEFT by introducing only a tiny amount of training parameters, which however, results in a substantial efficiency boost for inference, i.e., at around " + }, + { + "bbox": [ + 132, + 308, + 480, + 415 + ], + "type": "inline_equation", + "content": "50\\%" + }, + { + "bbox": [ + 132, + 308, + 480, + 415 + ], + "type": "text", + "content": " for both ViT-B and ViT-L models. Results in Sec. 4 show that PYRA achieves comparable performance to the PEFT counterpart without PYRA. This well indicates that PYRA is an effective method for training-inference efficient task adaptation." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 132, + 432, + 229, + 445 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 432, + 229, + 445 + ], + "spans": [ + { + "bbox": [ + 132, + 432, + 229, + 445 + ], + "type": "text", + "content": "4 Experiments" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 132, + 456, + 268, + 468 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 456, + 268, + 468 + ], + "spans": [ + { + "bbox": [ + 132, + 456, + 268, + 468 + ], + "type": "text", + "content": "4.1 Experimental Setting" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 132, + 474, + 480, + 557 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 474, + 480, + 557 + ], + "spans": [ + { + "bbox": [ + 132, + 474, + 480, + 557 + ], + "type": "text", + "content": "Datasets. We conduct extensive experiments on task adaptation benchmarks to verify the effectiveness of PYRA within the challenge of training-inference efficient task adaptation. Specifically, we choose the VTAB-1k [67] benchmark for the evaluation. VTAB-1k is a challenging benchmark that consists of 19 different tasks from diverse domains: 1) natural images captured in the actual world; 2) specialized images from professional fields; and 3) structured synthesized images. Each task only contains 800 training samples and 200 validation samples." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 132, + 558, + 480, + 605 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 558, + 480, + 605 + ], + "spans": [ + { + "bbox": [ + 132, + 558, + 480, + 605 + ], + "type": "text", + "content": "Models. We choose two ViT backbones pre-trained on the ImageNet-21K [16], i.e., ViT-L/16 and ViT-B/16, for comparison. Additionally, we also generalize our PYRA to different backbones and pre-train methods, including DeiT-B [51] and ViT backbones pre-trained by MAE [20]." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 132, + 605, + 480, + 665 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 605, + 480, + 665 + ], + "spans": [ + { + "bbox": [ + 132, + 605, + 480, + 665 + ], + "type": "text", + "content": "Implementation Details. We choose LoRA [24] as the PEFT module for all methods due to its simplicity and mergeability. For ease of explanation, we omit the \"+\"LoRA\" when comparing different methods. We append LoRA only on the " + }, + { + "bbox": [ + 132, + 605, + 480, + 665 + ], + "type": "inline_equation", + "content": "Q,K" + }, + { + "bbox": [ + 132, + 605, + 480, + 665 + ], + "type": "text", + "content": ", and " + }, + { + "bbox": [ + 132, + 605, + 480, + 665 + ], + "type": "inline_equation", + "content": "V" + }, + { + "bbox": [ + 132, + 605, + 480, + 665 + ], + "type": "text", + "content": " projection matrices, and apply the training schedule for LoRA following [17, 69]. The generators are trained along with LoRA" + } + ] + } + ], + "index": 11 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 287, + 91, + 447, + 102 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 287, + 91, + 447, + 102 + ], + "spans": [ + { + "bbox": [ + 287, + 91, + 447, + 102 + ], + "type": "text", + "content": "PYRA: Parallel Yielding Re-Activation" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 474, + 92, + 480, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 474, + 92, + 480, + 100 + ], + "spans": [ + { + "bbox": [ + 474, + 92, + 480, + 100 + ], + "type": "text", + "content": "9" + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 8 + }, + { + "para_blocks": [ + { + "type": "table", + "bbox": [ + 122, + 147, + 491, + 308 + ], + "blocks": [ + { + "bbox": [ + 132, + 114, + 481, + 137 + ], + "lines": [ + { + "bbox": [ + 132, + 114, + 481, + 137 + ], + "spans": [ + { + "bbox": [ + 132, + 114, + 481, + 137 + ], + "type": "text", + "content": "Table 2: Results on the VTAB-1k [67] benchmark under low compression rate. Bold and underline denote the best and second-best accuracy within compression methods." + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 122, + 147, + 491, + 308 + ], + "lines": [ + { + "bbox": [ + 122, + 147, + 491, + 308 + ], + "spans": [ + { + "bbox": [ + 122, + 147, + 491, + 308 + ], + "type": "table", + "html": "
Method# paramsThroughputNaturalSpecializedStructured
Cifar100Caltech101DTDFlowers102PetsSVHNSun397CamelyonEuroSATResisc45RetinopathyClevr-CountClevr-DistDMLABKITTI-DistdSpr-LocdSpr-OrisNORB-AzimsNORB-EleAverage
Model: ViT-B/16 (Throughput: 425)
PEFT0.34%42567.190.269.499.190.585.754.183.195.884.374.682.269.250.179.281.847.131.142.674.76
RaP3.43%65425.968.453.364.057.471.321.575.887.959.373.643.153.826.360.573.525.516.727.955.57
SPViT4.46%56741.675.461.183.266.256.128.379.394.273.373.670.661.542.467.875.450.528.931.364.16
DiffRate0.35%70937.184.663.796.786.232.648.278.985.867.073.732.929.834.155.712.616.013.121.555.82
ToMe0.34%75364.690.467.998.589.883.953.282.694.783.574.981.969.849.276.981.946.531.043.174.10
PYRA0.35%74567.590.369.398.990.084.653.183.395.783.375.282.668.950.880.081.845.832.242.874.69
Model: ViT-L/16 (Throughput: 130)
PEFT0.39%13077.191.473.499.591.389.657.685.996.187.376.183.163.050.782.181.753.532.236.676.52
RaP1.95%19643.287.962.652.881.786.734.778.492.473.373.668.059.646.982.475.543.624.525.765.64
SPViT2.47%18848.187.565.294.477.480.938.879.993.979.874.378.265.847.474.182.350.331.037.970.22
DiffRate0.39%22150.986.870.397.888.339.052.380.287.272.274.232.632.336.557.422.826.615.223.459.53
ToMe0.39%22776.191.172.399.291.789.256.486.495.186.675.182.461.950.981.481.653.533.436.876.11
PYRA0.40%22576.691.373.299.391.589.457.186.995.987.176.283.263.252.883.182.552.634.839.076.84
", + "image_path": "a79ca51c7c5e3144e040fba5e63a45739bfb8799139883ee1bf8117fd79b6aa5.jpg" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "table_body" + } + ], + "index": 3 + }, + { + "bbox": [ + 130, + 329, + 482, + 367 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 329, + 482, + 367 + ], + "spans": [ + { + "bbox": [ + 130, + 329, + 482, + 367 + ], + "type": "text", + "content": "modules. During inference, we merge the LoRA module to the backbone. All throughputs are measured during inference on a GeForce RTX 3090 GPU. More details can be found in the supplementary materials." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 130, + 383, + 449, + 396 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 383, + 449, + 396 + ], + "spans": [ + { + "bbox": [ + 130, + 383, + 449, + 396 + ], + "type": "text", + "content": "4.2 Performance comparison on Task Adaptation Benchmark" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 130, + 402, + 482, + 545 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 402, + 482, + 545 + ], + "spans": [ + { + "bbox": [ + 130, + 402, + 482, + 545 + ], + "type": "text", + "content": "We verify the effectiveness of PYRA for training-inference efficient task adaptation on (1) low compression rate: comparing the performance of different methods under the same sparsity ratio (here we set sparsity ratio " + }, + { + "bbox": [ + 130, + 402, + 482, + 545 + ], + "type": "inline_equation", + "content": "= 50\\%" + }, + { + "bbox": [ + 130, + 402, + 482, + 545 + ], + "type": "text", + "content": "); (2) high compression rate: comparing the performance between the competing methods and the smaller-scale model with similar throughput levels. We leverage ViT-L/16 and ViT-B/16 for these inspections. The competing methods include RaP [35], SPViT [30], DiffRate [8], and ToMe [2]. Comparisons to more baselines are in the supplementary materials. For RaP and SPViT, we first train the backbone with LoRA on downstream tasks, then we prune the models, and lastly we retrain the parameters attached by the pruning method and LoRA. DiffRate and ToMe can be employed with LoRA during fine-tuning, while DiffRate demands ImageNet-21K for searching the optimal compression schedule." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 130, + 546, + 482, + 666 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 546, + 482, + 666 + ], + "spans": [ + { + "bbox": [ + 130, + 546, + 482, + 666 + ], + "type": "text", + "content": "PYRA on low compression rate. Results on low compression rate are reported in Tab. 2. Overall, while achieving one of the best speedups on throughput, our PYRA is the best performed method compared to other competing methods using the lowest level of training parameters. Counting results on both backbones, our PYRA achieves the best or second-best performance on 37 of 38 dataset metrics. Compared to directly conducting PEFT on the backbone, while all competing methods cause worse results, our PYRA achieves comparable adaptation performance on ViT-B/16, and even outperforms ViT-L/16. The above results convincingly demonstrate that our PYRA successfully sets a new benchmark, i.e. reaching comparable performance as the uncompressed model," + } + ] + } + ], + "index": 7 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "spans": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "type": "text", + "content": "10" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 166, + 91, + 228, + 102 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 166, + 91, + 228, + 102 + ], + "spans": [ + { + "bbox": [ + 166, + 91, + 228, + 102 + ], + "type": "text", + "content": "Y. Xiong et al." + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 9 + }, + { + "para_blocks": [ + { + "type": "table", + "bbox": [ + 120, + 168, + 491, + 327 + ], + "blocks": [ + { + "bbox": [ + 130, + 114, + 482, + 159 + ], + "lines": [ + { + "bbox": [ + 130, + 114, + 482, + 159 + ], + "spans": [ + { + "bbox": [ + 130, + 114, + 482, + 159 + ], + "type": "text", + "content": "Table 3: Results on the VTAB-1k [67] benchmark under high compression rate. Bold and underline denote the best and second-best accuracy within compression methods. \\*: As a comparison of similar throughputs, we compare ViT-B/16 with PEFT on ViTS/16, and ViT-L/16 with PEFT on ViT-B/16." + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 120, + 168, + 491, + 327 + ], + "lines": [ + { + "bbox": [ + 120, + 168, + 491, + 327 + ], + "spans": [ + { + "bbox": [ + 120, + 168, + 491, + 327 + ], + "type": "table", + "html": "
Method# paramsThroughputNaturalSpecializedStructured
Cifar100Caltech101DTDFlowers102PetsSVHNSun397CamelyonEuroSATResisc45RetinopathyClevr-CountClevr-DistDMLABKITTI-DistdSpr-LocdSpr-OrisNORB-AzimsNORB-EleAverage
Model: ViT-B/16 (Throughput: 425)
PEFT*0.34%135057.788.270.198.788.785.744.981.494.784.673.681.664.148.180.072.938.422.937.771.85
RaP0.86%102924.340.134.541.840.521.711.475.886.535.173.849.649.728.139.413.815.412.426.942.60
SPViT4.46%94423.767.951.969.953.219.613.171.981.367.974.753.561.939.557.445.034.511.123.252.49
DiffRate0.35%130823.273.055.787.966.727.229.378.177.853.173.629.728.631.752.611.517.211.320.349.29
ToMe0.34%138154.287.865.596.181.779.745.279.493.676.373.878.365.748.071.380.045.830.941.270.43
PYRA0.35%136554.089.367.196.584.081.844.681.294.679.575.179.967.049.276.982.647.831.942.072.06
Model: ViT-L/16 (Throughput: 130)
PEFT*0.34%42567.190.269.499.190.585.754.183.195.884.374.682.269.250.179.281.847.131.142.674.76
RaP0.65%30117.737.127.046.233.323.213.376.574.254.473.650.431.425.749.853.125.513.426.044.11
SPViT2.47%28954.087.665.594.874.932.638.681.895.378.074.072.861.246.970.277.147.431.328.666.90
DiffRate0.39%41647.473.554.184.360.219.622.250.064.642.818.231.531.931.137.322.017.714.821.440.48
ToMe0.39%43171.090.970.498.388.587.252.482.994.583.175.080.761.148.976.980.853.032.135.274.10
PYRA0.40%42771.691.871.198.589.788.152.285.195.384.675.780.963.051.782.082.054.236.041.275.66
", + "image_path": "4b8c035c7014024e5a1ba0496be39bd046ede68e1f2568ba6113d60083c80ba5.jpg" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "table_body" + } + ], + "index": 3 + }, + { + "bbox": [ + 130, + 353, + 480, + 376 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 353, + 480, + 376 + ], + "spans": [ + { + "bbox": [ + 130, + 353, + 480, + 376 + ], + "type": "text", + "content": "on low compression rate for training-inference efficient task adaptation while accelerating the model to " + }, + { + "bbox": [ + 130, + 353, + 480, + 376 + ], + "type": "inline_equation", + "content": "1.75 \\times" + }, + { + "bbox": [ + 130, + 353, + 480, + 376 + ], + "type": "text", + "content": " speedup with only " + }, + { + "bbox": [ + 130, + 353, + 480, + 376 + ], + "type": "inline_equation", + "content": "0.4\\%" + }, + { + "bbox": [ + 130, + 353, + 480, + 376 + ], + "type": "text", + "content": " training parameters." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 130, + 377, + 481, + 508 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 377, + 481, + 508 + ], + "spans": [ + { + "bbox": [ + 130, + 377, + 481, + 508 + ], + "type": "text", + "content": "PYRA on high compression rate. Results are reported in Tab. 3. With comparable throughputs to the smaller-scale model and minimal training parameters, PYRA outperforms all competing methods and achieves the best performance on 35 of 38 dataset metrics. Compared to the smaller-scale model with comparable throughput, PYRA successfully outperforms it. On the compressed ViT-L/16, PYRA even surpasses ViT-B/16 by " + }, + { + "bbox": [ + 130, + 377, + 481, + 508 + ], + "type": "inline_equation", + "content": "0.9\\%" + }, + { + "bbox": [ + 130, + 377, + 481, + 508 + ], + "type": "text", + "content": ". This shows that PYRA, as the state-of-the-art, effectively bridges the adverse compression gap between compressed large-scale models and small-scale models, as mentioned in Sec. 1. Therefore, PYRA is an applicable alternative for acquiring a smaller-scale model " + }, + { + "bbox": [ + 130, + 377, + 481, + 508 + ], + "type": "inline_equation", + "content": "(3.2\\times" + }, + { + "bbox": [ + 130, + 377, + 481, + 508 + ], + "type": "text", + "content": " speedup) on downstream tasks through efficient training " + }, + { + "bbox": [ + 130, + 377, + 481, + 508 + ], + "type": "inline_equation", + "content": "(0.4\\%)" + }, + { + "bbox": [ + 130, + 377, + 481, + 508 + ], + "type": "text", + "content": " training parameters) when no pre-trained small-scale model is available." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 130, + 509, + 481, + 581 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 509, + 481, + 581 + ], + "spans": [ + { + "bbox": [ + 130, + 509, + 481, + 581 + ], + "type": "text", + "content": "PYRA on self-supervised ViT backbone. We conduct the experiments on both compression rates on self-supervised ViT-L/16 (MAE) [20]. As shown in Tab. 4, PYRA significantly surpasses all competing methods. Under low compression rate, PYRA outperforms directly transferring the uncompressed backbone. Under high compression rate, PYRA also eliminates adverse compression. These results show that PYRA generalizes well for self-supervised visual models." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 130, + 582, + 481, + 640 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 582, + 481, + 640 + ], + "spans": [ + { + "bbox": [ + 130, + 582, + 481, + 640 + ], + "type": "text", + "content": "PYRA on different architectures. To further verify the generalizability of PYRA, we conduct the above experiments on the DeiT-B [51]. As shown in Tab. 5, PYRA achieves comparable performance to the uncompressed model under low compression rate and eliminates adverse compression under high compression rate, indicating its generalizability to other transformer architectures." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 130, + 641, + 481, + 665 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 641, + 481, + 665 + ], + "spans": [ + { + "bbox": [ + 130, + 641, + 481, + 665 + ], + "type": "text", + "content": "PYRA consistently yields better models of different throughputs. To further show the superiority of PYRA for different compression rates, we" + } + ] + } + ], + "index": 8 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 287, + 91, + 447, + 102 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 287, + 91, + 447, + 102 + ], + "spans": [ + { + "bbox": [ + 287, + 91, + 447, + 102 + ], + "type": "text", + "content": "PYRA: Parallel Yielding Re-Activation" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 470, + 92, + 479, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 470, + 92, + 479, + 100 + ], + "spans": [ + { + "bbox": [ + 470, + 92, + 479, + 100 + ], + "type": "text", + "content": "11" + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 10 + }, + { + "para_blocks": [ + { + "type": "table", + "bbox": [ + 132, + 163, + 299, + 315 + ], + "blocks": [ + { + "bbox": [ + 132, + 114, + 298, + 148 + ], + "lines": [ + { + "bbox": [ + 132, + 114, + 298, + 148 + ], + "spans": [ + { + "bbox": [ + 132, + 114, + 298, + 148 + ], + "type": "text", + "content": "Table 4: VTAB-1k [67] results on both compression rates for ViT-L (MAE) (Throughput: 130)." + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 132, + 163, + 299, + 315 + ], + "lines": [ + { + "bbox": [ + 132, + 163, + 299, + 315 + ], + "spans": [ + { + "bbox": [ + 132, + 163, + 299, + 315 + ], + "type": "table", + "html": "
ModelMethod# paramsThroughputAverage
ViT-L (MAE)PEFT0.39%13075.96
ViT-L (MAE)RaP [35]2.01%18267.55
DiffRate [8]0.39%22146.91
ToMe [2]0.39%22774.97
PYRA0.40%22576.13
ViT-B (MAE)PEFT0.34%42570.23
ViT-L (MAE)RaP [35]0.76%29852.91
DiffRate [8]0.39%41648.03
ToMe [2]0.39%43168.20
PYRA0.40%42770.33
", + "image_path": "19e5322a0012641ce6cf6e7ccc7a1ff9bfc42fb413e5f5dd156d94ffa36e510c.jpg" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "table_body" + } + ], + "index": 3 + }, + { + "type": "table", + "bbox": [ + 315, + 163, + 482, + 315 + ], + "blocks": [ + { + "bbox": [ + 315, + 114, + 482, + 148 + ], + "lines": [ + { + "bbox": [ + 315, + 114, + 482, + 148 + ], + "spans": [ + { + "bbox": [ + 315, + 114, + 482, + 148 + ], + "type": "text", + "content": "Table 5: VTAB-1k [67] results on both compression rates for DeiT-B (Throughput: 431)." + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 315, + 163, + 482, + 315 + ], + "lines": [ + { + "bbox": [ + 315, + 163, + 482, + 315 + ], + "spans": [ + { + "bbox": [ + 315, + 163, + 482, + 315 + ], + "type": "table", + "html": "
ModelMethod# paramsThroughputAverage
DeiT-BPEFT0.34%43173.76
DeiT-BRaP [35]3.57%69562.34
DiffRate [8]0.35%73452.10
ToMe [2]0.34%74772.83
PYRA0.35%74073.55
DeiT-SPEFT0.34%133270.01
DeiT-BRaP [35]1.09%118757.70
DiffRate [8]0.35%131444.51
ToMe [2]0.34%135168.68
PYRA0.35%134170.13
", + "image_path": "2942dbe10df70cc190cb45a65e1b800ae5efc72246ce9d4dbe6325c286d03945.jpg" + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "table_body" + } + ], + "index": 5 + }, + { + "type": "table", + "bbox": [ + 144, + 357, + 471, + 439 + ], + "blocks": [ + { + "bbox": [ + 132, + 324, + 479, + 347 + ], + "lines": [ + { + "bbox": [ + 132, + 324, + 479, + 347 + ], + "spans": [ + { + "bbox": [ + 132, + 324, + 479, + 347 + ], + "type": "text", + "content": "Table 6: Ablation study results for PYRA on ViT-B/16 under high compression rate. Here for # params we report only the parameters introduced by token modulation." + } + ] + } + ], + "index": 6, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 144, + 357, + 471, + 439 + ], + "lines": [ + { + "bbox": [ + 144, + 357, + 471, + 439 + ], + "spans": [ + { + "bbox": [ + 144, + 357, + 471, + 439 + ], + "type": "table", + "html": "
Method\\( W_r \\)\\( W_D \\)Activation# paramsNaturalSpecializedStructuredAverage
Baseline×××072.8780.7857.6470.43
Plain \\( W_r \\)√××0.19K72.9081.0757.6670.54
\\( σ(\\cdot) \\& W_r \\)√×√0.19K73.1881.6957.6570.84
Plain \\( W_D \\)×√×8.45K73.0981.1358.4370.88
\\( σ(\\cdot) \\& W_D \\)×√√8.45K73.3182.1758.4471.31
Plain \\( W_r \\& W_D \\)√√×8.64K73.7781.3758.8171.32
PYRA√√√8.64K73.9182.6059.6672.06
", + "image_path": "457b41a3dc31fd83f96808fffd81cecc7fafb49350cb3e1665f91e94da05f7cc.jpg" + } + ] + } + ], + "index": 7, + "angle": 0, + "type": "table_body" + } + ], + "index": 7 + }, + { + "bbox": [ + 130, + 461, + 482, + 534 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 461, + 482, + 534 + ], + "spans": [ + { + "bbox": [ + 130, + 461, + 482, + 534 + ], + "type": "text", + "content": "compress the chosen backbones for a series of speedups and compare the results with the strongest baseline, ToMe [2], under similar values of throughputs. As shown in Figs. 2 and 4, PYRA consistently outperforms ToMe and flattens the accuracy-throughput curve. Therefore, our PYRA is applicable to acquire task-specific smaller-scale models of different throughputs consistently in the absence of pre-trained parameters for smaller-scale models." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 132, + 550, + 246, + 562 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 550, + 246, + 562 + ], + "spans": [ + { + "bbox": [ + 132, + 550, + 246, + 562 + ], + "type": "text", + "content": "4.3 Ablation Studies" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 130, + 569, + 482, + 666 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 569, + 482, + 666 + ], + "spans": [ + { + "bbox": [ + 130, + 569, + 482, + 666 + ], + "type": "text", + "content": "We do controlled experiments to identify the effect of individual components in PYRA, i.e., the generators " + }, + { + "bbox": [ + 130, + 569, + 482, + 666 + ], + "type": "inline_equation", + "content": "W_{r}" + }, + { + "bbox": [ + 130, + 569, + 482, + 666 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 130, + 569, + 482, + 666 + ], + "type": "inline_equation", + "content": "W_{D}" + }, + { + "bbox": [ + 130, + 569, + 482, + 666 + ], + "type": "text", + "content": ", and the sigmoid activation " + }, + { + "bbox": [ + 130, + 569, + 482, + 666 + ], + "type": "inline_equation", + "content": "\\sigma(\\cdot)" + }, + { + "bbox": [ + 130, + 569, + 482, + 666 + ], + "type": "text", + "content": " applied on the generated modulation weights. Specifically, when using only a single generator, we omit Eq. (6) and replace " + }, + { + "bbox": [ + 130, + 569, + 482, + 666 + ], + "type": "inline_equation", + "content": "\\hat{\\delta}_r^l" + }, + { + "bbox": [ + 130, + 569, + 482, + 666 + ], + "type": "text", + "content": " in Eq. (7) with the broadcast weight generated by the employed generator. When using the generators without sigmoid activation, we simply remove the " + }, + { + "bbox": [ + 130, + 569, + 482, + 666 + ], + "type": "inline_equation", + "content": "\\sigma" + }, + { + "bbox": [ + 130, + 569, + 482, + 666 + ], + "type": "text", + "content": " in Eq. (6) and Eq. (7) while keeping other calculations intact. We choose the ImageNet-21K pre-trained ViT-B/16 to carry out the ablation studies under high compression rates as in Tab. 3." + } + ] + } + ], + "index": 10 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "spans": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "type": "text", + "content": "12" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 166, + 91, + 228, + 102 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 166, + 91, + 228, + 102 + ], + "spans": [ + { + "bbox": [ + 166, + 91, + 228, + 102 + ], + "type": "text", + "content": "Y. Xiong et al." + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 11 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 137, + 119, + 294, + 248 + ], + "blocks": [ + { + "bbox": [ + 137, + 119, + 294, + 248 + ], + "lines": [ + { + "bbox": [ + 137, + 119, + 294, + 248 + ], + "spans": [ + { + "bbox": [ + 137, + 119, + 294, + 248 + ], + "type": "image", + "image_path": "cd78e6c8c8c119120dbf7db427ac6705fe8c42780ab35658e2372770f9bfdbdb.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 130, + 262, + 482, + 307 + ], + "lines": [ + { + "bbox": [ + 130, + 262, + 482, + 307 + ], + "spans": [ + { + "bbox": [ + 130, + 262, + 482, + 307 + ], + "type": "text", + "content": "Fig. 4: Comparisons between PYRA and ToMe [2] under different compression rates for ViT-Large (MAE) and DeiT-Base. Red boxes: PYRA mitigates the performance drops under low compression rate. Green boxes: PYRA eliminates adverse compression under high compression rate." + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "image_caption" + } + ], + "index": 2 + }, + { + "type": "image", + "bbox": [ + 315, + 119, + 471, + 248 + ], + "blocks": [ + { + "bbox": [ + 315, + 119, + 471, + 248 + ], + "lines": [ + { + "bbox": [ + 315, + 119, + 471, + 248 + ], + "spans": [ + { + "bbox": [ + 315, + 119, + 471, + 248 + ], + "type": "image", + "image_path": "93a8f85f613579036839ee19c25c863ffdd9b7d954343e0ea5711adc5ba0da07.jpg" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_body" + } + ], + "index": 3 + }, + { + "type": "table", + "bbox": [ + 135, + 361, + 305, + 453 + ], + "blocks": [ + { + "bbox": [ + 132, + 319, + 305, + 353 + ], + "lines": [ + { + "bbox": [ + 132, + 319, + 305, + 353 + ], + "spans": [ + { + "bbox": [ + 132, + 319, + 305, + 353 + ], + "type": "text", + "content": "Table 7: Adaptation performance comparison to " + }, + { + "bbox": [ + 132, + 319, + 305, + 353 + ], + "type": "inline_equation", + "content": "W_{r \\times D}" + }, + { + "bbox": [ + 132, + 319, + 305, + 353 + ], + "type": "text", + "content": " unadaptable to tokens. We report token modulation parameters." + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 135, + 361, + 305, + 453 + ], + "lines": [ + { + "bbox": [ + 135, + 361, + 305, + 453 + ], + "spans": [ + { + "bbox": [ + 135, + 361, + 305, + 453 + ], + "type": "table", + "html": "
Method# paramsNaturalSpecializedStructuredAverage
Baseline072.980.857.670.43
WD×r70.66K73.881.559.271.49
PYRA8.64K73.982.659.772.06
", + "image_path": "d3a8e06b8fbba5a1aaec3229f243c1136f8a0d79489fb99355aec4f5376efabb.jpg" + } + ] + } + ], + "index": 7, + "angle": 0, + "type": "table_body" + } + ], + "index": 7 + }, + { + "type": "table", + "bbox": [ + 312, + 361, + 482, + 452 + ], + "blocks": [ + { + "bbox": [ + 308, + 319, + 482, + 354 + ], + "lines": [ + { + "bbox": [ + 308, + 319, + 482, + 354 + ], + "spans": [ + { + "bbox": [ + 308, + 319, + 482, + 354 + ], + "type": "text", + "content": "Table 8: Adaptation performance comparison to the common gated generator. We report token modulation parameters." + } + ] + } + ], + "index": 6, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 312, + 361, + 482, + 452 + ], + "lines": [ + { + "bbox": [ + 312, + 361, + 482, + 452 + ], + "spans": [ + { + "bbox": [ + 312, + 361, + 482, + 452 + ], + "type": "table", + "html": "
Method# paramsNaturalSpecializedStructuredAverage
Baseline072.980.857.670.43
Gated73.73K72.881.558.971.06
PYRA8.64K73.982.659.772.06
", + "image_path": "1d9b019acd9e93fc9735e447650d46ea0e4e33d8ded3a6a2373c9f019b4ba05b.jpg" + } + ] + } + ], + "index": 8, + "angle": 0, + "type": "table_body" + } + ], + "index": 8 + }, + { + "bbox": [ + 130, + 476, + 482, + 586 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 476, + 482, + 586 + ], + "spans": [ + { + "bbox": [ + 130, + 476, + 482, + 586 + ], + "type": "text", + "content": "Results are shown in Tab. 6. Here the baseline method refers to conducting token merging [2,8] while fine-tuning LoRA [24] for task adaptation. First, both " + }, + { + "bbox": [ + 130, + 476, + 482, + 586 + ], + "type": "inline_equation", + "content": "W_{r}" + }, + { + "bbox": [ + 130, + 476, + 482, + 586 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 130, + 476, + 482, + 586 + ], + "type": "inline_equation", + "content": "W_{D}" + }, + { + "bbox": [ + 130, + 476, + 482, + 586 + ], + "type": "text", + "content": ", no matter whether sigmoid activation is attached, lead to performance gains, and employing them simultaneously outperforms using them individually. This proves the superiority of our parallel yielding strategy. Second, compared to the corresponding counterpart without activations, employing sigmoid activation always leads to significant improvements, indicating the effectiveness of the reactivation. Overall, our PYRA yields the best adaptation performance, showing that our strategies are effective and complementary." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 132, + 605, + 353, + 618 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 605, + 353, + 618 + ], + "spans": [ + { + "bbox": [ + 132, + 605, + 353, + 618 + ], + "type": "text", + "content": "4.4 Further Analysis on Different Designs" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 130, + 629, + 481, + 666 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 629, + 481, + 666 + ], + "spans": [ + { + "bbox": [ + 130, + 629, + 481, + 666 + ], + "type": "text", + "content": "We conduct further analysis on ViT-B/16 model pre-trained on ImageNet-21K. The baseline here refers to training LoRA with token merging [2, 8] attached. More analysis experiments can be found in the supplementary materials." + } + ] + } + ], + "index": 11 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 287, + 91, + 447, + 102 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 287, + 91, + 447, + 102 + ], + "spans": [ + { + "bbox": [ + 287, + 91, + 447, + 102 + ], + "type": "text", + "content": "PYRA: Parallel Yielding Re-Activation" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 470, + 91, + 480, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 470, + 91, + 480, + 100 + ], + "spans": [ + { + "bbox": [ + 470, + 91, + 480, + 100 + ], + "type": "text", + "content": "13" + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 12 + }, + { + "para_blocks": [ + { + "bbox": [ + 130, + 116, + 482, + 226 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 116, + 482, + 226 + ], + "spans": [ + { + "bbox": [ + 130, + 116, + 482, + 226 + ], + "type": "text", + "content": "Impact of Making Modulation Weights Adaptive. In our PYRA, we train modulation weight generators " + }, + { + "bbox": [ + 130, + 116, + 482, + 226 + ], + "type": "inline_equation", + "content": "W_{D}^{l}" + }, + { + "bbox": [ + 130, + 116, + 482, + 226 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 130, + 116, + 482, + 226 + ], + "type": "inline_equation", + "content": "W_{r}^{l}" + }, + { + "bbox": [ + 130, + 116, + 482, + 226 + ], + "type": "text", + "content": " to conduct adaptive modulation on different merging tokens. To prove the necessity of making modulation weights adaptive, we compare PYRA with the approach of directly training final modulation weights " + }, + { + "bbox": [ + 130, + 116, + 482, + 226 + ], + "type": "inline_equation", + "content": "W_{D\\times r}^{l}" + }, + { + "bbox": [ + 130, + 116, + 482, + 226 + ], + "type": "text", + "content": " for each layer, and conduct token modulation as " + }, + { + "bbox": [ + 130, + 116, + 482, + 226 + ], + "type": "inline_equation", + "content": "M_{s}^{l}\\gets M_{s}^{l} + (2\\sigma (W_{D\\times r}^{l}) - 1)\\odot M_{s}^{l}" + }, + { + "bbox": [ + 130, + 116, + 482, + 226 + ], + "type": "text", + "content": ". For fair comparisons, we inherit the sigmoid activation and the residual connection. As shown in Tab. 7, although with more training parameters, training " + }, + { + "bbox": [ + 130, + 116, + 482, + 226 + ], + "type": "inline_equation", + "content": "W_{D\\times r}" + }, + { + "bbox": [ + 130, + 116, + 482, + 226 + ], + "type": "text", + "content": " is still inferior to our PYRA. This indicates that PYRA is a more effective strategy to conduct adaptive token modulation." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 130, + 228, + 482, + 384 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 228, + 482, + 384 + ], + "spans": [ + { + "bbox": [ + 130, + 228, + 482, + 384 + ], + "type": "text", + "content": "Compare with the Common Gated Generator. We compare PYRA with setting the commonly-applied gated-style trainable module [12, 22, 31, 62] as modulation weight generator. Formally, for each ViT block, we insert a learnable two-layer MLP module " + }, + { + "bbox": [ + 130, + 228, + 482, + 384 + ], + "type": "inline_equation", + "content": "(\\mathrm{MLP}^l(\\cdot))" + }, + { + "bbox": [ + 130, + 228, + 482, + 384 + ], + "type": "text", + "content": " with input dimension " + }, + { + "bbox": [ + 130, + 228, + 482, + 384 + ], + "type": "inline_equation", + "content": "D" + }, + { + "bbox": [ + 130, + 228, + 482, + 384 + ], + "type": "text", + "content": ", hidden dimension " + }, + { + "bbox": [ + 130, + 228, + 482, + 384 + ], + "type": "inline_equation", + "content": "d \\ll D" + }, + { + "bbox": [ + 130, + 228, + 482, + 384 + ], + "type": "text", + "content": " to ensure parameter-efficient training and fast inference, and output dimension " + }, + { + "bbox": [ + 130, + 228, + 482, + 384 + ], + "type": "inline_equation", + "content": "D" + }, + { + "bbox": [ + 130, + 228, + 482, + 384 + ], + "type": "text", + "content": ". To generate modulation weights, we feed the information matrix " + }, + { + "bbox": [ + 130, + 228, + 482, + 384 + ], + "type": "inline_equation", + "content": "M_{\\mathrm{info}}^l" + }, + { + "bbox": [ + 130, + 228, + 482, + 384 + ], + "type": "text", + "content": " into the MLP, and modulate the merging tokens thereafter: " + }, + { + "bbox": [ + 130, + 228, + 482, + 384 + ], + "type": "inline_equation", + "content": "\\delta^l = \\mathrm{MLP}(M_{\\mathrm{info}}^l) \\in \\mathbb{R}^{D \\times r}" + }, + { + "bbox": [ + 130, + 228, + 482, + 384 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 130, + 228, + 482, + 384 + ], + "type": "inline_equation", + "content": "M_s^l \\gets M_s^l + (2\\sigma(\\delta^l - 1) \\odot M_s^l)" + }, + { + "bbox": [ + 130, + 228, + 482, + 384 + ], + "type": "text", + "content": ". We set " + }, + { + "bbox": [ + 130, + 228, + 482, + 384 + ], + "type": "inline_equation", + "content": "d = 4" + }, + { + "bbox": [ + 130, + 228, + 482, + 384 + ], + "type": "text", + "content": " for all layers. As shown in Tab. 8, both the gated generator and our PYRA achieve performance gains, while our PYRA surpasses the gated generator by " + }, + { + "bbox": [ + 130, + 228, + 482, + 384 + ], + "type": "inline_equation", + "content": "1.0\\%" + }, + { + "bbox": [ + 130, + 228, + 482, + 384 + ], + "type": "text", + "content": " with significantly fewer trainable parameters. This indicates that the decoupled weights generated by our " + }, + { + "bbox": [ + 130, + 228, + 482, + 384 + ], + "type": "inline_equation", + "content": "W_D" + }, + { + "bbox": [ + 130, + 228, + 482, + 384 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 130, + 228, + 482, + 384 + ], + "type": "inline_equation", + "content": "W_r" + }, + { + "bbox": [ + 130, + 228, + 482, + 384 + ], + "type": "text", + "content": " in PYRA are effective and parameter-efficient compared to the common gating strategy." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 132, + 418, + 220, + 430 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 418, + 220, + 430 + ], + "spans": [ + { + "bbox": [ + 132, + 418, + 220, + 430 + ], + "type": "text", + "content": "5 Conclusion" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 130, + 458, + 482, + 602 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 458, + 482, + 602 + ], + "spans": [ + { + "bbox": [ + 130, + 458, + 482, + 602 + ], + "type": "text", + "content": "In this work, we defined and investigated a new challenge named training-inference efficient task adaptation, in which the inference efficiency of large-scale transformers is enhanced during parameter-efficient task adaptation. We propose a novel Parallel Yielding Re-Activation method (PYRA) to effectively cope with the challenge by modulating token features during token merging. Specifically, PYRA generates decoupled parallel yielding modulation weights, and conducts token modulation through re-activation. Extensive experiments show that PYRA introduces negligible performance drops under low compression rate, and bridges the gap of adverse compression between compressed transformers and small-scale models under high compression rate. In real-world applications, our PYRA is highly suitable to the scenario of transferring large-scale vision transformers to downstream tasks where no small-scale model is presented." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 130, + 605, + 482, + 666 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 605, + 482, + 666 + ], + "spans": [ + { + "bbox": [ + 130, + 605, + 482, + 666 + ], + "type": "text", + "content": "Limitations. We have not yet validated the effectiveness of PYRA in object detection and image segmentation. We plan to extend these tasks in the future. Meanwhile, it is also applicable to attach pruning methods after training the vision transformers with PYRA to eliminate additional parameters attached by PYRA. We leave that topic to future studies." + } + ] + } + ], + "index": 6 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "spans": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "type": "text", + "content": "14" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 166, + 91, + 227, + 102 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 166, + 91, + 227, + 102 + ], + "spans": [ + { + "bbox": [ + 166, + 91, + 227, + 102 + ], + "type": "text", + "content": "Y. Xiong et al." + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 13 + }, + { + "para_blocks": [ + { + "bbox": [ + 133, + 114, + 246, + 129 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 114, + 246, + 129 + ], + "spans": [ + { + "bbox": [ + 133, + 114, + 246, + 129 + ], + "type": "text", + "content": "Acknowledgements" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 130, + 140, + 499, + 190 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 140, + 499, + 190 + ], + "spans": [ + { + "bbox": [ + 130, + 140, + 499, + 190 + ], + "type": "text", + "content": "This work was supported by National Science and Technology Major 2022ZD0119401, National Natural Science Foundation of China (Nos. 62271281, 61925107, 62021002). It is also sponsored by CAAI-CANN Open Fund, developed on OpenI Community." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 132, + 209, + 197, + 220 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 209, + 197, + 220 + ], + "spans": [ + { + "bbox": [ + 132, + 209, + 197, + 220 + ], + "type": "text", + "content": "References" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 134, + 234, + 481, + 665 + ], + "type": "list", + "angle": 0, + "index": 18, + "blocks": [ + { + "bbox": [ + 138, + 234, + 481, + 267 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 234, + 481, + 267 + ], + "spans": [ + { + "bbox": [ + 138, + 234, + 481, + 267 + ], + "type": "text", + "content": "1. Ahn, S., Hu, S.X., Damianou, A., Lawrence, N.D., Dai, Z.: Variational information distillation for knowledge transfer. In: Proceedings of the IEEE/CVF conference on computer vision and pattern recognition. pp. 9163-9171 (2019)" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 138, + 269, + 481, + 290 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 269, + 481, + 290 + ], + "spans": [ + { + "bbox": [ + 138, + 269, + 481, + 290 + ], + "type": "text", + "content": "2. Bolya, D., Fu, C.Y., Dai, X., Zhang, P., Feichtenhofer, C., Hoffman, J.: Token merging: Your vit but faster. arXiv preprint arXiv:2210.09461 (2022)" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 138, + 291, + 481, + 323 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 291, + 481, + 323 + ], + "spans": [ + { + "bbox": [ + 138, + 291, + 481, + 323 + ], + "type": "text", + "content": "3. Cai, Y., Yao, Z., Dong, Z., Gholami, A., Mahoney, M.W., Keutzer, K.: Zeroq: A novel zero shot quantization framework. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 13169-13178 (2020)" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 138, + 323, + 481, + 356 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 323, + 481, + 356 + ], + "spans": [ + { + "bbox": [ + 138, + 323, + 481, + 356 + ], + "type": "text", + "content": "4. Carion, N., Massa, F., Synnaeve, G., Usunier, N., Kirillov, A., Zagoruyko, S.: End-to-end object detection with transformers. In: European conference on computer vision. pp. 213-229. Springer (2020)" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 138, + 357, + 481, + 400 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 357, + 481, + 400 + ], + "spans": [ + { + "bbox": [ + 138, + 357, + 481, + 400 + ], + "type": "text", + "content": "5. Chavan, A., Shen, Z., Liu, Z., Liu, Z., Cheng, K.T., Xing, E.P.: Vision transformer slimming: Multi-dimension searching in continuous optimization space. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 4931-4941 (2022)" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 138, + 401, + 481, + 434 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 401, + 481, + 434 + ], + "spans": [ + { + "bbox": [ + 138, + 401, + 481, + 434 + ], + "type": "text", + "content": "6. Chen, C.F.R., Fan, Q., Panda, R.: Crossvit: Cross-attention multi-scale vision transformer for image classification. In: Proceedings of the IEEE/CVF international conference on computer vision. pp. 357-366 (2021)" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 138, + 434, + 481, + 466 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 434, + 481, + 466 + ], + "spans": [ + { + "bbox": [ + 138, + 434, + 481, + 466 + ], + "type": "text", + "content": "7. Chen, L., Zhao, H., Liu, T., Bai, S., Lin, J., Zhou, C., Chang, B.: An image is worth 1/2 tokens after layer 2: Plug-and-play inference acceleration for large vision-language models. arXiv preprint arXiv:2403.06764 (2024)" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 138, + 467, + 481, + 499 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 467, + 481, + 499 + ], + "spans": [ + { + "bbox": [ + 138, + 467, + 481, + 499 + ], + "type": "text", + "content": "8. Chen, M., Shao, W., Xu, P., Lin, M., Zhang, K., Chao, F., Ji, R., Qiao, Y., Luo, P.: Diffrate: Differentiable compression rate for efficient vision transformers. arXiv preprint arXiv:2305.17997 (2023)" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 138, + 500, + 481, + 533 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 500, + 481, + 533 + ], + "spans": [ + { + "bbox": [ + 138, + 500, + 481, + 533 + ], + "type": "text", + "content": "9. Chen, S., Ge, C., Tong, Z., Wang, J., Song, Y., Wang, J., Luo, P.: Adaptformer: Adapting vision transformers for scalable visual recognition. Advances in Neural Information Processing Systems 35, 16664-16678 (2022)" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 138, + 533, + 481, + 566 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 533, + 481, + 566 + ], + "spans": [ + { + "bbox": [ + 138, + 533, + 481, + 566 + ], + "type": "text", + "content": "0. Chen, T., Cheng, Y., Gan, Z., Yuan, L., Zhang, L., Wang, Z.: Chasing sparsity in vision transformers: An end-to-end exploration. Advances in Neural Information Processing Systems 34, 19974-19988 (2021)" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 134, + 567, + 481, + 609 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 134, + 567, + 481, + 609 + ], + "spans": [ + { + "bbox": [ + 134, + 567, + 481, + 609 + ], + "type": "text", + "content": "11. Cheng, B., Misra, I., Schwing, A.G., Kirillov, A., Girdhar, R.: Masked-attention mask transformer for universal image segmentation. In: Proceedings of the IEEE/CVF conference on computer vision and pattern recognition. pp. 1290-1299 (2022)" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 134, + 610, + 481, + 632 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 134, + 610, + 481, + 632 + ], + "spans": [ + { + "bbox": [ + 134, + 610, + 481, + 632 + ], + "type": "text", + "content": "12. Chung, J., Gulcehre, C., Cho, K., Bengio, Y.: Empirical evaluation of gated recurrent neural networks on sequence modeling. arXiv preprint arXiv:1412.3555 (2014)" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 134, + 632, + 481, + 665 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 134, + 632, + 481, + 665 + ], + "spans": [ + { + "bbox": [ + 134, + 632, + 481, + 665 + ], + "type": "text", + "content": "13. Courbariaux, M., Bengio, Y., David, J.P.: Binaryconnect: Training deep neural networks with binary weights during propagations. Advances in neural information processing systems 28 (2015)" + } + ] + } + ], + "index": 17 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 287, + 91, + 447, + 102 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 287, + 91, + 447, + 102 + ], + "spans": [ + { + "bbox": [ + 287, + 91, + 447, + 102 + ], + "type": "text", + "content": "PYRA: Parallel Yielding Re-Activation" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 470, + 91, + 481, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 470, + 91, + 481, + 100 + ], + "spans": [ + { + "bbox": [ + 470, + 91, + 481, + 100 + ], + "type": "text", + "content": "15" + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 14 + }, + { + "para_blocks": [ + { + "bbox": [ + 133, + 116, + 481, + 666 + ], + "type": "list", + "angle": 0, + "index": 19, + "blocks": [ + { + "bbox": [ + 133, + 116, + 481, + 149 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 116, + 481, + 149 + ], + "spans": [ + { + "bbox": [ + 133, + 116, + 481, + 149 + ], + "type": "text", + "content": "14. Ding, N., Qin, Y., Yang, G., Wei, F., Yang, Z., Su, Y., Hu, S., Chen, Y., Chan, C.M., Chen, W., et al.: Delta tuning: A comprehensive study of parameter efficient methods for pre-trained language models. arXiv preprint arXiv:2203.06904 (2022)" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 133, + 149, + 481, + 194 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 149, + 481, + 194 + ], + "spans": [ + { + "bbox": [ + 133, + 149, + 481, + 194 + ], + "type": "text", + "content": "15. Ding, Z., Wang, A., Chen, H., Zhang, Q., Liu, P., Bao, Y., Yan, W., Han, J.: Exploring structured semantic prior for multi label recognition with incomplete labels. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 3398-3407 (2023)" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 133, + 194, + 481, + 237 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 194, + 481, + 237 + ], + "spans": [ + { + "bbox": [ + 133, + 194, + 481, + 237 + ], + "type": "text", + "content": "16. Dosovitskiy, A., Beyer, L., Kolesnikov, A., Weissenborn, D., Zhai, X., Unterthiner, T., Dehghani, M., Minderer, M., Heigold, G., Gelly, S., et al.: An image is worth 16x16 words: Transformers for image recognition at scale. arXiv preprint arXiv:2010.11929 (2020)" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 133, + 237, + 481, + 259 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 237, + 481, + 259 + ], + "spans": [ + { + "bbox": [ + 133, + 237, + 481, + 259 + ], + "type": "text", + "content": "17. Hao, T., Chen, H., Guo, Y., Ding, G.: Consolidator: Mergeable adapter with grouped connections for visual adaptation. arXiv preprint arXiv:2305.00603 (2023)" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 133, + 259, + 481, + 293 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 259, + 481, + 293 + ], + "spans": [ + { + "bbox": [ + 133, + 259, + 481, + 293 + ], + "type": "text", + "content": "18. Hao, T., Ding, X., Feng, J., Yang, Y., Chen, H., Ding, G.: Quantized prompt for efficient generalization of vision-language models. arXiv preprint arXiv:2407.10704 (2024)" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 133, + 293, + 481, + 325 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 293, + 481, + 325 + ], + "spans": [ + { + "bbox": [ + 133, + 293, + 481, + 325 + ], + "type": "text", + "content": "19. Hao, T., Lyu, M., Chen, H., Zhao, S., Han, J., Ding, G.: Re-parameterized low-rank prompt: Generalize a vision-language model within " + }, + { + "bbox": [ + 133, + 293, + 481, + 325 + ], + "type": "inline_equation", + "content": "0.5\\mathrm{k}" + }, + { + "bbox": [ + 133, + 293, + 481, + 325 + ], + "type": "text", + "content": " parameters. arXiv preprint arXiv:2312.10813 (2023)" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 133, + 325, + 481, + 358 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 325, + 481, + 358 + ], + "spans": [ + { + "bbox": [ + 133, + 325, + 481, + 358 + ], + "type": "text", + "content": "20. He, K., Chen, X., Xie, S., Li, Y., Dóllar, P., Girshick, R.: Masked autoencoders are scalable vision learners. In: Proceedings of the IEEE/CVF conference on computer vision and pattern recognition. pp. 16000-16009 (2022)" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 133, + 358, + 481, + 380 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 358, + 481, + 380 + ], + "spans": [ + { + "bbox": [ + 133, + 358, + 481, + 380 + ], + "type": "text", + "content": "21. Hinton, G., Vinyals, O., Dean, J.: Distilling the knowledge in a neural network. arXiv preprint arXiv:1503.02531 (2015)" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 133, + 380, + 481, + 403 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 380, + 481, + 403 + ], + "spans": [ + { + "bbox": [ + 133, + 380, + 481, + 403 + ], + "type": "text", + "content": "22. Hochreiter, S., Schmidhuber, J.: Long short-term memory. Neural Computation 9(8), 1735-1780 (1997)" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 133, + 403, + 481, + 435 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 403, + 481, + 435 + ], + "spans": [ + { + "bbox": [ + 133, + 403, + 481, + 435 + ], + "type": "text", + "content": "23. Houlsby, N., Giurgiu, A., Jastrzebski, S., Morrone, B., De Laroussilhe, Q., Gesmundo, A., Attariyan, M., Gelly, S.: Parameter-efficient transfer learning for nlp. In: International Conference on Machine Learning. pp. 2790-2799. PMLR (2019)" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 133, + 435, + 481, + 468 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 435, + 481, + 468 + ], + "spans": [ + { + "bbox": [ + 133, + 435, + 481, + 468 + ], + "type": "text", + "content": "24. Hu, E.J., Shen, Y., Wallis, P., Allen-Zhu, Z., Li, Y., Wang, S., Wang, L., Chen, W.: Lora: Low-rank adaptation of large language models. arXiv preprint arXiv:2106.09685 (2021)" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 133, + 468, + 481, + 490 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 468, + 481, + 490 + ], + "spans": [ + { + "bbox": [ + 133, + 468, + 481, + 490 + ], + "type": "text", + "content": "25. Huang, X., Shen, Z., Cheng, K.T.: Variation-aware vision transformer quantization. arXiv preprint arXiv:2307.00331 (2023)" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 133, + 490, + 481, + 523 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 490, + 481, + 523 + ], + "spans": [ + { + "bbox": [ + 133, + 490, + 481, + 523 + ], + "type": "text", + "content": "26. Jia, M., Tang, L., Chen, B.C., Cardie, C., Belongie, S., Hariharan, B., Lim, S.N.: Visual prompt tuning. In: European Conference on Computer Vision. pp. 709-727. Springer (2022)" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 133, + 523, + 481, + 555 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 523, + 481, + 555 + ], + "spans": [ + { + "bbox": [ + 133, + 523, + 481, + 555 + ], + "type": "text", + "content": "27. Jiang, H., Wu, Q., Lin, C.Y., Yang, Y., Qiu, L.: Llmlingua: Compressing prompts for accelerated inference of large language models. arXiv preprint arXiv:2310.05736 (2023)" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 133, + 555, + 481, + 578 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 555, + 481, + 578 + ], + "spans": [ + { + "bbox": [ + 133, + 555, + 481, + 578 + ], + "type": "text", + "content": "28. Jie, S., Deng, Z.H.: Convolutional bypasses are better vision transformer adapters. arXiv preprint arXiv:2207.07039 (2022)" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 133, + 578, + 481, + 611 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 578, + 481, + 611 + ], + "spans": [ + { + "bbox": [ + 133, + 578, + 481, + 611 + ], + "type": "text", + "content": "29. Kirillov, A., Mintun, E., Ravi, N., Mao, H., Rolland, C., Gustafson, L., Xiao, T., Whitehead, S., Berg, A.C., Lo, W.Y., et al.: Segment anything. arXiv preprint arXiv:2304.02643 (2023)" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 133, + 611, + 481, + 666 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 611, + 481, + 666 + ], + "spans": [ + { + "bbox": [ + 133, + 611, + 481, + 666 + ], + "type": "text", + "content": "30. Kong, Z., Dong, P., Ma, X., Meng, X., Niu, W., Sun, M., Shen, X., Yuan, G., Ren, B., Tang, H., Qin, M., Wang, Y.: Spvit: Enabling faster vision transformers via latency-aware soft token pruning. In: Avidan, S., Brostow, G.J., Cissé, M., Farinella, G.M., Hassner, T. (eds.) Computer Vision - ECCV 2022 - 17th European Conference, Tel Aviv, Israel, October 23-27, 2022, Proceedings, Part XI. Lecture" + } + ] + } + ], + "index": 18 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "spans": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "type": "text", + "content": "16" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 166, + 91, + 227, + 101 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 166, + 91, + 227, + 101 + ], + "spans": [ + { + "bbox": [ + 166, + 91, + 227, + 101 + ], + "type": "text", + "content": "Y. Xiong et al." + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 15 + }, + { + "para_blocks": [ + { + "bbox": [ + 132, + 116, + 481, + 666 + ], + "type": "list", + "angle": 0, + "index": 19, + "blocks": [ + { + "bbox": [ + 147, + 116, + 481, + 148 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 147, + 116, + 481, + 148 + ], + "spans": [ + { + "bbox": [ + 147, + 116, + 481, + 148 + ], + "type": "text", + "content": "Notes in Computer Science, vol. 13671, pp. 620-640. Springer (2022). https://doi.org/10.1007/978-3-031-20083-0_37, https://doi.org/10.1007/978-3-031-20083-0_37" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 133, + 150, + 481, + 184 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 150, + 481, + 184 + ], + "spans": [ + { + "bbox": [ + 133, + 150, + 481, + 184 + ], + "type": "text", + "content": "31. Kong, Z., Dong, P., Ma, X., Meng, X., Sun, M., Niu, W., Shen, X., Yuan, G., Ren, B., Qin, M., et al.: Spvit: Enabling faster vision transformers via soft token pruning. arXiv preprint arXiv:2112.13890 (2021)" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 132, + 185, + 481, + 206 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 185, + 481, + 206 + ], + "spans": [ + { + "bbox": [ + 132, + 185, + 481, + 206 + ], + "type": "text", + "content": "32. Lester, B., Al-Rfou, R., Constant, N.: The power of scale for parameter-efficient prompt tuning. arXiv preprint arXiv:2104.08691 (2021)" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 132, + 207, + 481, + 228 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 207, + 481, + 228 + ], + "spans": [ + { + "bbox": [ + 132, + 207, + 481, + 228 + ], + "type": "text", + "content": "33. Li, X.L., Liang, P.: Prefix-tuning: Optimizing continuous prompts for generation. arXiv preprint arXiv:2101.00190 (2021)" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 132, + 229, + 481, + 262 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 229, + 481, + 262 + ], + "spans": [ + { + "bbox": [ + 132, + 229, + 481, + 262 + ], + "type": "text", + "content": "34. Li, Y., Xu, S., Zhang, B., Cao, X., Gao, P., Guo, G.: Q-vit: Accurate and fully quantized low-bit vision transformer. Advances in Neural Information Processing Systems 35, 34451-34463 (2022)" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 132, + 263, + 481, + 318 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 263, + 481, + 318 + ], + "spans": [ + { + "bbox": [ + 132, + 263, + 481, + 318 + ], + "type": "text", + "content": "35. Li, Y., Adamczewski, K., Li, W., Gu, S., Timofte, R., Gool, L.V.: Revisiting random channel pruning for neural network compression. In: IEEE/CVF Conference on Computer Vision and Pattern Recognition, CVPR 2022, New Orleans, LA, USA, June 18-24, 2022. pp. 191-201. IEEE (2022). https://doi.org/10.1109/CVPR52688.2022.00029, https://doi.org/10.1109/CVPR52688.2022.00029" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 132, + 319, + 481, + 351 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 319, + 481, + 351 + ], + "spans": [ + { + "bbox": [ + 132, + 319, + 481, + 351 + ], + "type": "text", + "content": "36. Lian, D., Zhou, D., Feng, J., Wang, X.: Scaling & shifting your features: A new baseline for efficient model tuning. Advances in Neural Information Processing Systems 35, 109-123 (2022)" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 132, + 352, + 481, + 374 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 352, + 481, + 374 + ], + "spans": [ + { + "bbox": [ + 132, + 352, + 481, + 374 + ], + "type": "text", + "content": "37. Lin, X., Zhao, C., Pan, W.: Towards accurate binary convolutional neural network. Advances in neural information processing systems 30 (2017)" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 132, + 375, + 481, + 397 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 375, + 481, + 397 + ], + "spans": [ + { + "bbox": [ + 132, + 375, + 481, + 397 + ], + "type": "text", + "content": "38. Lin, Z., Courbariaux, M., Memisevic, R., Bengio, Y.: Neural networks with few multiplications. arXiv preprint arXiv:1510.03009 (2015)" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 132, + 398, + 481, + 430 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 398, + 481, + 430 + ], + "spans": [ + { + "bbox": [ + 132, + 398, + 481, + 430 + ], + "type": "text", + "content": "39. Liu, X., Ji, K., Fu, Y., Tam, W.L., Du, Z., Yang, Z., Tang, J.: P-tuning v2: Prompt tuning can be comparable to fine-tuning universally across scales and tasks. arXiv preprint arXiv:2110.07602 (2021)" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 132, + 431, + 481, + 474 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 431, + 481, + 474 + ], + "spans": [ + { + "bbox": [ + 132, + 431, + 481, + 474 + ], + "type": "text", + "content": "40. Liu, Z., Lin, Y., Cao, Y., Hu, H., Wei, Y., Zhang, Z., Lin, S., Guo, B.: Swin transformer: Hierarchical vision transformer using shifted windows. In: Proceedings of the IEEE/CVF international conference on computer vision. pp. 10012-10022 (2021)" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 132, + 475, + 481, + 508 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 475, + 481, + 508 + ], + "spans": [ + { + "bbox": [ + 132, + 475, + 481, + 508 + ], + "type": "text", + "content": "41. Liu, Z., Wang, Y., Han, K., Zhang, W., Ma, S., Gao, W.: Post-training quantization for vision transformer. Advances in Neural Information Processing Systems 34, 28092-28103 (2021)" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 132, + 509, + 481, + 542 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 509, + 481, + 542 + ], + "spans": [ + { + "bbox": [ + 132, + 509, + 481, + 542 + ], + "type": "text", + "content": "42. Lyu, M., Yang, Y., Hong, H., Chen, H., Jin, X., He, Y., Xue, H., Han, J., Ding, G.: One-dimensional adapter to rule them all: Concepts, diffusion models and erasing applications. arXiv preprint arXiv:2312.16145 (2023)" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 132, + 543, + 481, + 564 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 543, + 481, + 564 + ], + "spans": [ + { + "bbox": [ + 132, + 543, + 481, + 564 + ], + "type": "text", + "content": "43. Marin, D., Chang, J.H.R., Ranjan, A., Prabhu, A., Rastegari, M., Tuzel, O.: Token pooling in vision transformers. arXiv preprint arXiv:2110.03860 (2021)" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 132, + 565, + 481, + 597 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 565, + 481, + 597 + ], + "spans": [ + { + "bbox": [ + 132, + 565, + 481, + 597 + ], + "type": "text", + "content": "44. Perez, E., Strub, F., De Vries, H., Dumoulin, V., Courville, A.: Film: Visual reasoning with a general conditioning layer. In: Proceedings of the AAAI conference on artificial intelligence (2018)" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 132, + 599, + 481, + 632 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 599, + 481, + 632 + ], + "spans": [ + { + "bbox": [ + 132, + 599, + 481, + 632 + ], + "type": "text", + "content": "45. Pfeiffer, J., Kamath, A., Rücklé, A., Cho, K., Gurevych, I.: Adapterfusion: Nondestructive task composition for transfer learning. arXiv preprint arXiv:2005.00247 (2020)" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 132, + 632, + 481, + 666 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 632, + 481, + 666 + ], + "spans": [ + { + "bbox": [ + 132, + 632, + 481, + 666 + ], + "type": "text", + "content": "46. Song, Y., Zhou, Q., Li, X., Fan, D.P., Lu, X., Ma, L.: Ba-sam: Scalable bias-mode attention mask for segment anything model. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 3162-3173 (2024)" + } + ] + } + ], + "index": 18 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 287, + 91, + 447, + 102 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 287, + 91, + 447, + 102 + ], + "spans": [ + { + "bbox": [ + 287, + 91, + 447, + 102 + ], + "type": "text", + "content": "PYRA: Parallel Yielding Re-Activation" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 470, + 91, + 481, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 470, + 91, + 481, + 100 + ], + "spans": [ + { + "bbox": [ + 470, + 91, + 481, + 100 + ], + "type": "text", + "content": "17" + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 16 + }, + { + "para_blocks": [ + { + "bbox": [ + 132, + 117, + 481, + 665 + ], + "type": "list", + "angle": 0, + "index": 19, + "blocks": [ + { + "bbox": [ + 133, + 117, + 480, + 149 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 117, + 480, + 149 + ], + "spans": [ + { + "bbox": [ + 133, + 117, + 480, + 149 + ], + "type": "text", + "content": "47. Song, Z., Jia, C., Yang, L., Wei, H., Liu, L.: Graphalign++: An accurate feature alignment by graph matching for multi-modal 3d object detection. IEEE Transactions on Circuits and Systems for Video Technology (2023)" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 133, + 150, + 481, + 194 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 150, + 481, + 194 + ], + "spans": [ + { + "bbox": [ + 133, + 150, + 481, + 194 + ], + "type": "text", + "content": "48. Song, Z., Wei, H., Bai, L., Yang, L., Jia, C.: Graphalign: Enhancing accurate feature alignment by graph matching for multi-modal 3d object detection. In: Proceedings of the IEEE/CVF International Conference on Computer Vision. pp. 3358-3369 (2023)" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 132, + 194, + 480, + 226 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 194, + 480, + 226 + ], + "spans": [ + { + "bbox": [ + 132, + 194, + 480, + 226 + ], + "type": "text", + "content": "49. Strudel, R., Garcia, R., Laptev, I., Schmid, C.: Segmenter: Transformer for semantic segmentation. In: Proceedings of the IEEE/CVF international conference on computer vision. pp. 7262-7272 (2021)" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 132, + 227, + 480, + 248 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 227, + 480, + 248 + ], + "spans": [ + { + "bbox": [ + 132, + 227, + 480, + 248 + ], + "type": "text", + "content": "50. Tian, L., Ye, M., Zhou, L., He, Q.: Clip-guided black-box domain adaptation of image classification. Signal, Image and Video Processing 18(5), 4637-4646 (2024)" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 132, + 249, + 480, + 282 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 249, + 480, + 282 + ], + "spans": [ + { + "bbox": [ + 132, + 249, + 480, + 282 + ], + "type": "text", + "content": "51. Touvron, H., Cord, M., Douze, M., Massa, F., Sablayrolles, A., Jégou, H.: Training data-efficient image transformers & distillation through attention. In: International conference on machine learning. pp. 10347-10357. PMLR (2021)" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 132, + 282, + 480, + 304 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 282, + 480, + 304 + ], + "spans": [ + { + "bbox": [ + 132, + 282, + 480, + 304 + ], + "type": "text", + "content": "52. Tung, F., Mori, G.: Similarity-preserving knowledge distillation. In: Proceedings of the IEEE/CVF international conference on computer vision. pp. 1365-1374 (2019)" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 132, + 304, + 480, + 325 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 304, + 480, + 325 + ], + "spans": [ + { + "bbox": [ + 132, + 304, + 480, + 325 + ], + "type": "text", + "content": "53. Wang, A., Chen, H., Lin, Z., Pu, H., Ding, G.: Repvit: Revisiting mobile cnn from vit perspective. arXiv preprint arXiv:2307.09283 (2023)" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 132, + 326, + 480, + 357 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 326, + 480, + 357 + ], + "spans": [ + { + "bbox": [ + 132, + 326, + 480, + 357 + ], + "type": "text", + "content": "54. Wang, A., Chen, H., Lin, Z., Zhao, S., Han, J., Ding, G.: Cait: Triple-win compression towards high accuracy, fast inference, and favorable transferability for vits. arXiv preprint arXiv:2309.15755 (2023)" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 132, + 358, + 480, + 402 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 358, + 480, + 402 + ], + "spans": [ + { + "bbox": [ + 132, + 358, + 480, + 402 + ], + "type": "text", + "content": "55. Wang, W., Xie, E., Li, X., Fan, D.P., Song, K., Liang, D., Lu, T., Luo, P., Shao, L.: Pyramid vision transformer: A versatile backbone for dense prediction without convolutions. In: Proceedings of the IEEE/CVF international conference on computer vision. pp. 568-578 (2021)" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 132, + 403, + 480, + 434 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 403, + 480, + 434 + ], + "spans": [ + { + "bbox": [ + 132, + 403, + 480, + 434 + ], + "type": "text", + "content": "56. Wang, Z., Luo, H., Wang, P., Ding, F., Wang, F., Li, H.: Vtc-lfc: Vision transformer compression with low-frequency components. Advances in Neural Information Processing Systems 35, 13974-13988 (2022)" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 132, + 435, + 480, + 479 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 435, + 480, + 479 + ], + "spans": [ + { + "bbox": [ + 132, + 435, + 480, + 479 + ], + "type": "text", + "content": "57. Wei, K., Du, R., Jin, L., Liu, J., Yin, J., Zhang, L., Liu, J., Liu, N., Zhang, J., Guo, Z.: Video event extraction with multi-view interaction knowledge distillation. In: Proceedings of the AAAI Conference on Artificial Intelligence. vol. 38, pp. 19224-19233 (2024)" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 132, + 479, + 480, + 522 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 479, + 480, + 522 + ], + "spans": [ + { + "bbox": [ + 132, + 479, + 480, + 522 + ], + "type": "text", + "content": "58. Wei, K., Jin, L., Zhang, Z., Guo, Z., Li, X., Liu, Q., Feng, W.: More than syntaxes: Investigating semantics to zero-shot cross-lingual relation extraction and event argument role labelling. ACM Transactions on Asian and Low-Resource Language Information Processing 23(5), 1-21 (2024)" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 132, + 523, + 480, + 544 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 523, + 480, + 544 + ], + "spans": [ + { + "bbox": [ + 132, + 523, + 480, + 544 + ], + "type": "text", + "content": "59. Wu, Y., He, K.: Group normalization. In: Proceedings of the European conference on computer vision (ECCV). pp. 3-19 (2018)" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 132, + 545, + 480, + 578 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 545, + 480, + 578 + ], + "spans": [ + { + "bbox": [ + 132, + 545, + 480, + 578 + ], + "type": "text", + "content": "60. Xiong, Y., Chen, H., Lin, Z., Zhao, S., Ding, G.: Confidence-based visual dispersal for few-shot unsupervised domain adaptation. In: Proceedings of the IEEE/CVF International Conference on Computer Vision. pp. 11621-11631 (2023)" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 132, + 578, + 480, + 599 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 578, + 480, + 599 + ], + "spans": [ + { + "bbox": [ + 132, + 578, + 480, + 599 + ], + "type": "text", + "content": "61. Xiong, Y., Chen, X., Ye, X., Chen, H., Lin, Z., Lian, H., Niu, J., Ding, G.: Temporal scaling law for large language models. arXiv preprint arXiv:2404.17785 (2024)" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 132, + 600, + 480, + 632 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 600, + 480, + 632 + ], + "spans": [ + { + "bbox": [ + 132, + 600, + 480, + 632 + ], + "type": "text", + "content": "62. Xu, L., Xie, H., Qin, S.Z.J., Tao, X., Wang, F.L.: Parameter-efficient fine-tuning methods for pretrained language models: A critical review and assessment. arXiv preprint arXiv:2312.12148 (2023)" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 132, + 633, + 480, + 665 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 633, + 480, + 665 + ], + "spans": [ + { + "bbox": [ + 132, + 633, + 480, + 665 + ], + "type": "text", + "content": "63. Yao, Z., Yazdani Aminabadi, R., Zhang, M., Wu, X., Li, C., He, Y.: Zeroquant: Efficient and affordable post-training quantization for large-scale transformers. Advances in Neural Information Processing Systems 35, 27168-27183 (2022)" + } + ] + } + ], + "index": 18 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "spans": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "type": "text", + "content": "18" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 166, + 91, + 227, + 101 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 166, + 91, + 227, + 101 + ], + "spans": [ + { + "bbox": [ + 166, + 91, + 227, + 101 + ], + "type": "text", + "content": "Y. Xiong et al." + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 17 + }, + { + "para_blocks": [ + { + "bbox": [ + 130, + 116, + 482, + 468 + ], + "type": "list", + "angle": 0, + "index": 13, + "blocks": [ + { + "bbox": [ + 130, + 116, + 482, + 138 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 116, + 482, + 138 + ], + "spans": [ + { + "bbox": [ + 130, + 116, + 482, + 138 + ], + "type": "text", + "content": "64. Yu, S., Chen, T., Shen, J., Yuan, H., Tan, J., Yang, S., Liu, J., Wang, Z.: Unified visual transformer compression. arXiv preprint arXiv:2203.08243 (2022)" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 130, + 138, + 482, + 171 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 138, + 482, + 171 + ], + "spans": [ + { + "bbox": [ + 130, + 138, + 482, + 171 + ], + "type": "text", + "content": "65. Zaken, E.B., Ravfogel, S., Goldberg, Y.: Bitfit: Simple parameter-efficient fine-tuning for transformer-based masked language-models. arXiv preprint arXiv:2106.10199 (2021)" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 132, + 172, + 481, + 205 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 172, + 481, + 205 + ], + "spans": [ + { + "bbox": [ + 132, + 172, + 481, + 205 + ], + "type": "text", + "content": "66. Zhai, X., Kolesnikov, A., Houlsby, N., Beyer, L.: Scaling vision transformers. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 12104-12113 (2022)" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 132, + 205, + 481, + 247 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 205, + 481, + 247 + ], + "spans": [ + { + "bbox": [ + 132, + 205, + 481, + 247 + ], + "type": "text", + "content": "67. Zhai, X., Puigcerver, J., Kolesnikov, A., Ruyssen, P., Riquelme, C., Lucic, M., Djolonga, J., Pinto, A.S., Neumann, M., Dosovitskiy, A., et al.: A large-scale study of representation learning with the visual task adaptation benchmark. arXiv preprint arXiv:1910.04867 (2019)" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 132, + 248, + 481, + 281 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 248, + 481, + 281 + ], + "spans": [ + { + "bbox": [ + 132, + 248, + 481, + 281 + ], + "type": "text", + "content": "68. Zhang, Y., Qiu, L., Zhu, Y., Wen, L., Luo, X.: A new childhood pneumonia diagnosis method based on fine-grained convolutional neural network. Computer Modeling in Engineering & Sciences 133, 873-894 (2022)" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 132, + 281, + 481, + 303 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 281, + 481, + 303 + ], + "spans": [ + { + "bbox": [ + 132, + 281, + 481, + 303 + ], + "type": "text", + "content": "69. Zhang, Y., Zhou, K., Liu, Z.: Neural prompt search. arXiv preprint arXiv:2206.04673 (2022)" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 132, + 303, + 481, + 346 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 303, + 481, + 346 + ], + "spans": [ + { + "bbox": [ + 132, + 303, + 481, + 346 + ], + "type": "text", + "content": "70. Zheng, S., Lu, J., Zhao, H., Zhu, X., Luo, Z., Wang, Y., Fu, Y., Feng, J., Xiang, T., Torr, P.H., et al.: Rethinking semantic segmentation from a sequence-to-sequence perspective with transformers. In: Proceedings of the IEEE/CVF conference on computer vision and pattern recognition. pp. 6881-6890 (2021)" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 132, + 347, + 481, + 380 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 347, + 481, + 380 + ], + "spans": [ + { + "bbox": [ + 132, + 347, + 481, + 380 + ], + "type": "text", + "content": "71. Zhou, F., Zhou, Q., Yin, B., Zheng, H., Lu, X., Ma, L., Ling, H.: Rethinking impersonation and dodging attacks on face recognition systems. In: Proceedings of the 30th ACM International Conference on Multimedia (ACM MM) (2024)" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 132, + 380, + 481, + 413 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 380, + 481, + 413 + ], + "spans": [ + { + "bbox": [ + 132, + 380, + 481, + 413 + ], + "type": "text", + "content": "72. Zhou, Q., Feng, Z., Gu, Q., Pang, J., Cheng, G., Lu, X., Shi, J., Ma, L.: Context-aware mixup for domain adaptive semantic segmentation. IEEE Transactions on Circuits and Systems for Video Technology 33(2), 804-817 (2022)" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 132, + 413, + 481, + 434 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 413, + 481, + 434 + ], + "spans": [ + { + "bbox": [ + 132, + 413, + 481, + 434 + ], + "type": "text", + "content": "73. Zhu, M., Tang, Y., Han, K.: Vision transformer pruning. arXiv preprint arXiv:2104.08500 (2021)" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 132, + 434, + 481, + 468 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 434, + 481, + 468 + ], + "spans": [ + { + "bbox": [ + 132, + 434, + 481, + 468 + ], + "type": "text", + "content": "74. Zhu, X., Su, W., Lu, L., Li, B., Wang, X., Dai, J.: Deformable detr: Deformable transformers for end-to-end object detection. arXiv preprint arXiv:2010.04159 (2020)" + } + ] + } + ], + "index": 12 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 287, + 91, + 447, + 102 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 287, + 91, + 447, + 102 + ], + "spans": [ + { + "bbox": [ + 287, + 91, + 447, + 102 + ], + "type": "text", + "content": "PYRA: Parallel Yielding Re-Activation" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 470, + 91, + 481, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 470, + 91, + 481, + 100 + ], + "spans": [ + { + "bbox": [ + 470, + 91, + 481, + 100 + ], + "type": "text", + "content": "19" + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 18 + } + ], + "_backend": "vlm", + "_version_name": "2.6.4" +} \ No newline at end of file diff --git a/2024/PaPr_ Training-Free One-Step Patch Pruning with Lightweight ConvNets for Faster Inference/f59c8716-3e2f-464a-8e1c-4f54e686a19f_content_list.json b/2024/PaPr_ Training-Free One-Step Patch Pruning with Lightweight ConvNets for Faster Inference/f59c8716-3e2f-464a-8e1c-4f54e686a19f_content_list.json new file mode 100644 index 0000000000000000000000000000000000000000..6056259a55d26e49509268cf51f7f061dddf1682 --- /dev/null +++ b/2024/PaPr_ Training-Free One-Step Patch Pruning with Lightweight ConvNets for Faster Inference/f59c8716-3e2f-464a-8e1c-4f54e686a19f_content_list.json @@ -0,0 +1,1727 @@ +[ + { + "type": "text", + "text": "PaPr: Training-Free One-Step Patch Pruning with Lightweight ConvNets for Faster Inference", + "text_level": 1, + "bbox": [ + 218, + 140, + 784, + 185 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Tanvir Mahmud1 Burhaneddin Yaman2 Chun-Hao Liu3 Diana Marculescu1", + "bbox": [ + 222, + 210, + 777, + 226 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "1University of Texas at Austin 2Bosch Research 3Amazon Prime Video 1{tanvirmahmud, dianam}@utexas.edu 2yaman013@umn.edu 3liuch37@gmail.com", + "bbox": [ + 238, + 238, + 761, + 267 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Abstract. As deep neural networks evolve from convolutional neural networks (ConvNets) to advanced vision transformers (ViTs), there is an increased need to eliminate redundant data for faster processing without compromising accuracy. Previous methods are often architecture-specific or necessitate re-training, restricting their applicability with frequent model updates. To solve this, we first introduce a novel property of lightweight ConvNets: their ability to identify key discriminative patch regions in images, irrespective of model's final accuracy or size. We demonstrate that fully-connected layers are the primary bottleneck for ConvNets performance, and their suppression with simple weight recalibration markedly enhances discriminative patch localization performance. Using this insight, we introduce PaPr, a method for substantially pruning redundant patches with minimal accuracy loss using lightweight ConvNets across a variety of deep learning architectures, including ViTs, ConvNets, and hybrid transformers, without any re-training. Moreover, the simple early-stage one-step patch pruning with PaPr enhances existing patch reduction methods. Through extensive testing on diverse architectures, PaPr achieves significantly higher accuracy over state-of-the-art patch reduction methods with similar FLOP count reduction. More specifically, PaPr reduces about $70\\%$ of redundant patches in videos with less than $0.8\\%$ drop in accuracy, and up to $3.7\\times$ FLOPs reduction, which is a $15\\%$ more reduction with $2.5\\%$ higher accuracy. Code is available at https://github.com/tanvir-utexas/PaPr.", + "bbox": [ + 261, + 300, + 738, + 618 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "1 Introduction", + "text_level": 1, + "bbox": [ + 217, + 643, + 374, + 659 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Deep neural networks have grown from simple convolutional neural networks (ConvNets) to complex transformer models, aiming for better accuracy with more computation [10, 18]. Vision transformers (ViTs) excel by focusing on important parts of images with long-range attention and large-scale pre-training [17,39]. This success, however, comes at the cost of increased computational cost. Operating these large networks efficiently for downstream applications is crucial. Most visual tasks demand precisely pinpointing key image regions against complex backgrounds. Higher image resolution improves accuracy by capturing more details but also adds unnecessary background processing, burdening large models [44]. This highlights the need to cut down on redundant data in high-resolution images to maintain both speed and performance with advanced techniques [2,34].", + "bbox": [ + 217, + 672, + 787, + 840 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/2ebd885cf12a80b9567181d16380666f4a52e1d5d91ad631a18b3dd54be74dd6.jpg", + "image_caption": [ + "(b) Proposed PaPr", + "Fig. 1: (a) Existing patch pruning methods gradually reduce patches over the model. This requires additional training of mask generators in intermediate layers. (b) Proposed PaPr directly prunes redundant patches early in the network by leveraging pretrained lightweight ConvNets and directly speeds-up off-the-shelf models without re-training." + ], + "image_footnote": [], + "bbox": [ + 261, + 143, + 743, + 291 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Identifying key image regions demands a comprehensive understanding of the image and model operations. Incorrect key region estimation can impair pretraining performance by eliminating crucial areas. Traditional approaches [2,25,28, 34,56] for pruning redundant patch regions are hindered by three main limitations: (1) They often necessitate complex training of extra modules, which becomes increasingly difficult as baseline models evolve with more data, enhanced training methods, and deeper structures [25,41,56]. Re-training these modules for each model update is impractical. (2) Without a complete image understanding, these methods incrementally prune patches across the network, leading to unnecessary computations in early layers, which is particularly problematic for deeper models. (3) Many rely on specific architectural features for patch reduction, such as class tokens or attention maps, limiting their use to a narrow set of network designs [12,28]. Hence, there is a pressing need for a patch pruning solution that is adaptable to various modern architectures without additional training, can eliminate redundant patches in a single-step, thereby making it suitable for a broad spectrum of networks while streamlining each model update (See Fig. 1).", + "bbox": [ + 212, + 398, + 787, + 641 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Recent work on patch reduction [2, 23, 55] focuses mostly on transformers rather than ConvNets due to their impressive performance on various tasks, their aforementioned limitations notwithstanding. While ConvNets achieve lower ImageNet-1k top-1 accuracy than large ViTs (68.7% in MobileOne-S0 [49] vs. 88.7% in ViT-Huge [44]), they exhibit a remarkable ability to efficiently process the key image regions with hierarchical inductive bias. Our empirical investigation reveals that, as we broaden the evaluation metric (increasing $k$ in top- $k$ evaluations, see Fig. 4), the benefits of deeper models diminish, especially with a large number of image classes (e.g., 1000 in ImageNet). This suggests that shallower models excel at identifying discriminative areas as their bigger counterparts, rendering them ideal for patch pruning.", + "bbox": [ + 212, + 642, + 789, + 808 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Leveraging this insight, we propose PaPr, a novel Patch Pruning method that employs pretrained lightweight ConvNets for efficient, one-step patch pruning in", + "bbox": [ + 212, + 809, + 787, + 840 + ], + "page_idx": 1 + }, + { + "type": "page_number", + "text": "2", + "bbox": [ + 217, + 114, + 228, + 126 + ], + "page_idx": 1 + }, + { + "type": "header", + "text": "T. Mahmud et al.", + "bbox": [ + 271, + 114, + 392, + 127 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "a wide variety of deep learning models, maintaining accuracy while significantly cutting computational demands. Our findings show lightweight ConvNets have remarkable ability in identifying discriminative image regions but struggle in fine-grained prediction. To address this, PaPr relies on Patch Significance Maps (PSMs), which are generated using only the convolutional layers of ConvNets through uniform class weight recalibration in FC layers. Astonishingly, PSMs consistently highlight critical image regions across ConvNets of varying sizes and accuracies (Fig. 6), thereby amplifying the efficacy of ultra lightweight ConvNets.", + "bbox": [ + 212, + 146, + 787, + 267 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "ConvNets inherently preserve the positional property of patches for their inductive bias, while ViTs use cross-attention to blend patch features, leading to variability in patch relevance across models. Unlike previous methods that prune redundant patches gradually over multiple steps across intermediate layers of ViTs [2, 23, 41], our approach, PaPr, simplifies this process by eliminating non-essential patches at once, immediately after extraction (see Fig. 3), by leveraging lightweight ConvNets to assess patch significance. This direct, one-step pruning approach significantly cuts computational demands and is also compatible with other patch reduction techniques (see Fig. 5). PaPr's ability to separate crucial patch identification from fine-grained class prediction enhances a wide range of pre-trained models without further training, ensuring high accuracy with notable speed ups. By bypassing the complex training required for conventional patch selectors, PaPr capitalizes on the comprehensive capabilities of minimalist ConvNets for efficient patch pruning in larger models.", + "bbox": [ + 212, + 268, + 787, + 479 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Our experiments demonstrate PaPr's effectiveness across various architectures and pre-training methods, achieving significant reduction in redundant patches for ViTs, large-scale ConvNets (e.g., ConvNext [32]), and hybrid transformers (e.g., Swin [31]), outperforming state-of-the-art (SOTA) patch reduction methods by a large margin. Notably, PaPr can be easily integrated with most existing patch reduction methods to reduce patches early in operation. More specifically, PaPr boosts ToMe [2] accuracy by $4.5\\%$ for a computational budget similar to ViT-B. Moreover, PaPr can accelerate training akin to token merging techniques, unlike most patch pruning methods that fail to boost training speed [34, 41, 56]. PaPr shows robust patch localization performance with ultra lightweight ConvNets ( $<0.3\\%$ accuracy loss) for $42\\times$ reduction in proposal FLOPs, thereby enabling its use for larger off-the-shelf models. Remarkably, in video recognition, PaPr cuts down around $70\\%$ of redundant patches, resulting in up to $3.7\\times$ FLOPs reduction with minimal impact on accuracy ( $\\approx 0.8\\%$ ).", + "bbox": [ + 212, + 483, + 787, + 694 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "We summarize our main contributions as follows:", + "bbox": [ + 240, + 696, + 594, + 709 + ], + "page_idx": 2 + }, + { + "type": "list", + "sub_type": "text", + "list_items": [ + "- We propose PaPr, a novel background patch pruning method that can seamlessly operate with ViTs, ConvNets, and hybrid transformers, without further training while leveraging batch processing.", + "- We propose a simple weight recalibration method in ConvNets to precisely and efficiently locate discriminative patches, irrespective of model size.", + "- We facilitate the use of ultra-lightweight ConvNets to speed-up large models, such as ViTs, with a seamless framework and negligible accuracy loss." + ], + "bbox": [ + 223, + 729, + 784, + 838 + ], + "page_idx": 2 + }, + { + "type": "header", + "text": "PaPr for Faster Inference", + "bbox": [ + 560, + 114, + 730, + 127 + ], + "page_idx": 2 + }, + { + "type": "page_number", + "text": "3", + "bbox": [ + 774, + 116, + 785, + 126 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "- We present extensive qualitative and quantitative results across numerous model architectures in both image and video applications.", + "bbox": [ + 223, + 146, + 787, + 176 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "2 Related Work", + "text_level": 1, + "bbox": [ + 215, + 196, + 388, + 213 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "2.1 From ConvNets to Vision Transformers", + "text_level": 1, + "bbox": [ + 215, + 224, + 588, + 239 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "ConvNets have been pivotal in computer vision, offering computational efficiency through kernel reuse and localization [8, 18-20, 29, 42, 46, 49, 59]. Yet, they struggle in capturing long-range dependencies, a gap bridged by ViT [10]. Inspired by the success of transformers from natural language processing (NLP), ViTs excel in long-range feature modeling through cross-attention, outperforming ConvNets albeit with higher computational demands and extensive training data requirements [7, 48, 51, 57]. Efforts to alleviate these issues include self-supervised [17, 47, 52, 60] and weakly-supervised [39, 44] pre-training, although challenges in computational complexity and optimization remain. Hybrid architectures, merging ConvNets' inductive biases with ViTs' cross-attention, offer a balanced solution by reducing computational load while maintaining performance [6, 9, 11, 15, 27, 30, 31, 33, 54]. Recent ConvNets advancements [32, 53], with improved training and large-scale data, now rival ViTs, questioning whether architectural innovations or enhanced training primarily drive performance gains.", + "bbox": [ + 212, + 248, + 787, + 458 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Addressing the computational demands of these models, we focus on enhancing operational efficiency by pruning redundant patches without architectural modifications or re-training, maintaining performance while streamlining updates.", + "bbox": [ + 212, + 459, + 789, + 505 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "2.2 Class Activation Mapping for Explainable Deep Learning", + "text_level": 1, + "bbox": [ + 215, + 523, + 733, + 539 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Class Activation Mapping (CAM) techniques provide explainable visual reasoning for neural network predictions by highlighting activation regions1 crucial for decisions [1,21,38,43,61]. While ConvNets utilize convolutional operations to preserve spatial information, making them apt for such visualizations, ViTs link attention weights to class tokens for prediction [4,36]. However, CAM's reliance on accurate class predictions for effective feature localization is a significant drawback, as ConvNets' lower accuracy when compared to ViTs may lead to incorrect object localization. Despite ViTs' superior fine-grained classification capabilities, it is unclear if they prioritize the same discriminative regions as ConvNets. Additionally, CAM approaches often require gradient tracking [5,43] or complex feature map decomposition [35], complicating batch processing and necessitating extra optimization steps.", + "bbox": [ + 212, + 546, + 787, + 727 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Our work diverges from traditional CAM by aiming to consistently identify discriminative patches across different architectures, focusing computational resources on the most relevant areas without being constrained by class prediction accuracy. This approach enables more efficient processing by allowing larger models to concentrate on the most significant regions, identified by lighter networks, thus significantly reducing computational demands.", + "bbox": [ + 212, + 727, + 787, + 818 + ], + "page_idx": 3 + }, + { + "type": "page_number", + "text": "4", + "bbox": [ + 217, + 114, + 228, + 126 + ], + "page_idx": 3 + }, + { + "type": "header", + "text": "T. Mahmud et al.", + "bbox": [ + 271, + 114, + 392, + 127 + ], + "page_idx": 3 + }, + { + "type": "page_footnote", + "text": "1 \"Patch\" and \"region\" have been used interchangeably based on the context.", + "bbox": [ + 217, + 824, + 730, + 839 + ], + "page_idx": 3 + }, + { + "type": "image", + "img_path": "images/9ac1ccb35defa0ba0d4eae72669c9558242aed3ac197e5bc228178f70f6056c2.jpg", + "image_caption": [ + "(a) Baseline Convolutional Model Feature Flow" + ], + "image_footnote": [], + "bbox": [ + 261, + 143, + 743, + 222 + ], + "page_idx": 4 + }, + { + "type": "image", + "img_path": "images/3279da1ba742f58bf7165335ee79f9be4bb66d64ceadb6dce907d0cad97645b7.jpg", + "image_caption": [ + "(b) Proposed Pruning Mask Extraction", + "Fig.2: (a) Baseline ConvNet gradually reduces the feature map to produce $\\mathcal{F} = \\{f_k(x,y)\\}_{k=1}^K$ , followed by global average pooling and fully connected (FC) layers to predict $y_p$ . (b) In PaPr, we operate on $\\mathcal{F}$ by suppressing the FC layer. Initially, we extract pixel mean over $K$ channels to produce discriminative region proposal $\\mathcal{R}$ . Later, simple upsampling operation generates the patch significance map (PSM) $\\mathcal{P}$ of target dimension. Finally, patch mask $\\mathcal{M}$ for top $z\\%$ patches is obtained from $\\mathcal{P}$ ." + ], + "image_footnote": [], + "bbox": [ + 308, + 241, + 702, + 316 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "2.3 Patch Reduction for Faster Inference", + "text_level": 1, + "bbox": [ + 215, + 452, + 568, + 465 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Several approaches have sought to enhance computational efficiency by reducing redundant patches in neural networks, with early strategies involving additional adapters or controllers to identify and prune less significant patches [24,26,41,58]. These methods, however, necessitate separate adapter training for each network and are slow to adapt due to the need to learn from the dynamics of other layers. In the context of ViTs, efforts have leveraged architectural features such as class tokens and attention maps for patch relevance, yet these solutions often fail to generalize across various architectures like hybrid transformers [31] or larger ConvNets [32, 53], thus limiting their applicability [12, 25, 34, 56]. Furthermore, while some have explored patch merging in ViTs and a combination of merging and pruning [2, 3, 23], these approaches lack a holistic image understanding by not considering the entire image context in their optimizations, leading to incremental and sub-optimal patch reduction.", + "bbox": [ + 212, + 473, + 787, + 669 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Our work diverges from these traditional patch reduction methods by proposing a single-step, early-network patch removal strategy, that seamlessly integrates with any architecture without re-training.", + "bbox": [ + 212, + 670, + 789, + 715 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "3 Methodology", + "text_level": 1, + "bbox": [ + 215, + 734, + 380, + 753 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Our methodology introduces a novel approach to discriminative patch pruning across various deep learning architectures, leveraging the innate capabilities of lightweight ConvNets. By generating a PSM, we efficiently identify and prune non-essential patches in a single step, enhancing computational efficiency without compromising accuracy. This process is universally applicable, seamlessly", + "bbox": [ + 212, + 763, + 787, + 840 + ], + "page_idx": 4 + }, + { + "type": "header", + "text": "PaPr for Faster Inference", + "bbox": [ + 560, + 114, + 730, + 126 + ], + "page_idx": 4 + }, + { + "type": "page_number", + "text": "5", + "bbox": [ + 774, + 116, + 784, + 126 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "integrating with ViTs, ConvNets, and hybrid models, thereby addressing the limitations of previous methods with a scalable, architecture-agnostic solution.", + "bbox": [ + 212, + 146, + 782, + 176 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "3.1 Extracting Discriminative Regions with ConvNets", + "text_level": 1, + "bbox": [ + 214, + 198, + 674, + 214 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Despite achieving relatively lower top-1 accuracy on the ImageNet benchmark, lightweight ConvNets exhibit a competitive edge in top-10 accuracy when compared to their larger ViT counterparts (Fig. 4). This phenomenon underscores ConvNets' capability to effectively localize regions of interest through their convolutional layers, despite potential limitations in fine-grained classification attributed to the fully-connected (FC) layer. Our methodology leverages this insight by proposing discriminative patch regions while minimizing the influence of the FC layer (see Fig. 2), thereby enhancing the large off-the-shelf model's focus on most salient image regions proposed by lightweight ConvNets (see Fig. 3).", + "bbox": [ + 212, + 222, + 787, + 357 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Given an input image $X \\in \\mathbb{R}^{H \\times W \\times 3}$ , with height $H$ and width $W$ , we denote by $f_{k}(x,y)$ the feature map generated by the $k^{th}$ kernel in the last convolutional layer of typical ConvNets, $\\forall k \\in \\{1,2,\\dots,K\\}$ . Here, each pixel $(x,y)$ in $f_{k}(x,y)$ corresponds to a patch window of size $(H / d \\times W / d)$ in the input image $X$ , reflecting a spatial down-scaling factor of $d$ through the convolutional layer stack. Typically, global average pooling (GAP) is applied to each feature map $f_{k}(x,y)$ , yielding $F = \\{F_{k}\\}_{k=1}^{K} \\in \\mathbb{R}^{K}$ , where $F_{k} = \\sum_{x,y} f_{k}(x,y) \\in \\mathbb{R}$ . Finally, a fully connected (FC) layer $W \\in \\mathbb{R}^{C \\times K}$ with elements $w_{c}^{k}$ processes $F$ to generate class predictions $y_{p} \\in \\mathbb{R}^{C}$ across $C$ classes as follows:", + "bbox": [ + 212, + 358, + 787, + 506 + ], + "page_idx": 5 + }, + { + "type": "equation", + "text": "\n$$\ny _ {p} = W F = \\sum_ {c} \\sum_ {k} w _ {c} ^ {k} \\sum_ {x, y} f _ {k} (x, y) = \\sum_ {x, y} R (x, y), \\tag {1}\n$$\n", + "text_format": "latex", + "bbox": [ + 326, + 516, + 787, + 547 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "where $R \\in \\mathbb{R}^{h \\times w}$ encapsulates the weighted mean class activation mapping of the convolutional feature map.", + "bbox": [ + 212, + 559, + 787, + 589 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "We note that the reliance on class activation weights $w_{c}^{k}$ is influenced by the model's final accuracy, thereby posing a challenge for smaller models. Our objective transcends mere accuracy enhancement, aiming to precisely locate discriminative patches rich in information irrespective of the model size or its final classification performance. Recognizing that discriminative region localization is pivotal for detailed classification, and acknowledging the competitive top-10 accuracy of lighter ConvNets, we posit that an extremely lightweight ConvNet suffices for initial discriminative region proposal.", + "bbox": [ + 212, + 589, + 787, + 710 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "To counteract the influence of the weak linear FC layer $W$ in convolutional region proposal, we propose an adjustment where $w_{c}^{k} = 1 / KC$ , $\\forall k\\in \\{1,\\dots ,K\\}$ , facilitating the generation of a discriminative region proposal $\\mathcal{R}\\in \\mathbb{R}^{h\\times w}$ , where", + "bbox": [ + 212, + 710, + 787, + 756 + ], + "page_idx": 5 + }, + { + "type": "equation", + "text": "\n$$\n\\mathcal {R} (x, y) = \\sum_ {c} \\sum_ {k} \\frac {1}{K C} f _ {k} (x, y) = \\frac {1}{K} \\sum_ {k} f _ {k} (x, y). \\tag {2}\n$$\n", + "text_format": "latex", + "bbox": [ + 331, + 766, + 787, + 800 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "This strategy allows us to leverage ConvNets for what they excel at: pinpointing critical image areas. By reducing reliance on class activation weights, we", + "bbox": [ + 214, + 809, + 787, + 840 + ], + "page_idx": 5 + }, + { + "type": "page_number", + "text": "6", + "bbox": [ + 217, + 114, + 228, + 126 + ], + "page_idx": 5 + }, + { + "type": "header", + "text": "T. Mahmud et al.", + "bbox": [ + 271, + 114, + 392, + 127 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "efficiently generate discriminative region proposals directly from convolutional outputs. This approach not only enhances the interpretability and efficiency of the localization process but also enables the application of more complex models for subsequent detailed analysis, optimizing the use of computational resources.", + "bbox": [ + 212, + 146, + 787, + 208 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "3.2 Patch Significance Map", + "text_level": 1, + "bbox": [ + 214, + 229, + 457, + 244 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "With the discriminative region proposal $\\mathcal{R} \\in \\mathbb{R}^{h \\times w}$ , where each pixel $(x, y)$ quantifies the significance of corresponding patches in the original image, we proceed to establish a precise mapping to our intended feature map (Fig. 2). This mapping ensures the preservation of spatial relationships within the feature map.", + "bbox": [ + 212, + 253, + 787, + 314 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Consider a target feature map $\\mathcal{F} \\in \\mathbb{R}^{h' \\times w' \\times K}$ , with each pixel $(x', y')$ encapsulating a feature vector from a specific image patch. To align $\\mathcal{R}$ with $\\mathcal{F}$ , we employ an upsampling operation $U: \\mathbb{R}^{h \\times w} \\to \\mathbb{R}^{h' \\times w'}$ , transforming $\\mathcal{R}$ into the Patch Significance Map (PSM) $\\mathcal{P} \\in \\mathbb{R}^{h' \\times w'}$ . Consequently, each element of $\\mathcal{P}$ directly corresponds to the patch significance within $\\mathcal{F}$ for the given discriminative task. The next step involves utilizing $\\mathcal{P}$ to discern and prune non-essential patch features from $\\mathcal{F}$ . By sorting the values within $\\mathcal{P}$ , we acquire a pruning mask $\\mathcal{M}$ :", + "bbox": [ + 212, + 314, + 787, + 421 + ], + "page_idx": 6 + }, + { + "type": "equation", + "text": "\n$$\n\\mathcal {M} = \\operatorname {r e s h a p e} (\\operatorname {a r g s o r t} (\\operatorname {f l a t t e n} (\\mathcal {P}))), \\mathcal {M} \\in \\mathbb {R} ^ {h ^ {\\prime} \\times w ^ {\\prime}}, \\tag {3}\n$$\n", + "text_format": "latex", + "bbox": [ + 330, + 430, + 787, + 450 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "enabling batch-wise patch pruning. Specifically, we identify the indices corresponding to the top- $z\\%$ patches as per $\\mathcal{M}$ , facilitating the retention of only the most salient patches in $\\mathcal{F}$ , thereby enhancing computational efficiency in subsequent processing stages.", + "bbox": [ + 212, + 460, + 787, + 518 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "As part of our discussion on various architectures, we detail the application of this patch reduction technique. For enhanced visualization of the PSM, we apply min-max normalization to $\\mathcal{P}$ , adjusting for outliers and scaling the significance scores to visually depict the importance of different patches.", + "bbox": [ + 212, + 520, + 787, + 580 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "3.3 Integrating PSM with Vision Transformers", + "text_level": 1, + "bbox": [ + 214, + 603, + 617, + 619 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "The ViT processes an input image $X \\in \\mathbb{R}^{H \\times W \\times 3}$ by extracting $N$ -dimensional patch token features $X_{p} \\in \\mathbb{R}^{N \\times d}$ from a $(k \\times k)$ patch window using strided convolutions, where $N = HW / k^2$ (Fig. 3). Positional embeddings are then added to these patch tokens $X_{p}$ to retain spatial information, followed by the application of successive cross-attention mechanisms. For classification tasks, the model either introduces a class token $t_{cls} \\in \\mathbb{R}^{d}$ or utilizes the mean of the output patch tokens.", + "bbox": [ + 212, + 628, + 787, + 719 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "The encoder employs multi-headed cross-attention (MHA) on the patch token embeddings $X_{p}$ , with each MHA block consisting of multiple linear layers and a cross-attention layer. The computational complexity of these operations suggests that reducing the number of tokens $N$ can significantly decrease computational demands. A key strength of the ViT architecture is its adaptability to varying numbers of tokens $N$ , though at a higher computational cost. Prior work [2,41,56] has primarily aimed at reducing patch tokens within intermediate encoder layers, often requiring additional training due to the variability in token representations", + "bbox": [ + 212, + 719, + 787, + 840 + ], + "page_idx": 6 + }, + { + "type": "header", + "text": "PaPr for Faster Inference", + "bbox": [ + 560, + 114, + 730, + 127 + ], + "page_idx": 6 + }, + { + "type": "page_number", + "text": "7", + "bbox": [ + 774, + 114, + 785, + 126 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/3636b67499797712f50e8b886624e4e8bfefb4d7a91e409f8147d30015cc21a2.jpg", + "image_caption": [ + "Fig. 3: (a) In vanilla ViT, PaPr operates right after the patch extractor module. Hence, all transformer blocks can operate only with the most discriminative patches. (b) Hierarchical model blocks comprise of window based kernel operator (e.g., Convk $\\times$ $k$ /local attention), followed by pixel operator (e.g., linear layer, Conv1x1). Pixel operator consumes more than $60\\%$ of total computation. PaPr is used to split the foreground patches to be operated with pixel operator. Background patches are zero-ed out, and finally, re-assembled with foreground output patches." + ], + "image_footnote": [], + "bbox": [ + 259, + 143, + 743, + 296 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "across different architectures. These methods typically achieve only gradual token reduction, limited by the distributed nature of information across tokens.", + "bbox": [ + 212, + 428, + 785, + 458 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "In contrast, our approach seeks to eliminate redundant tokens early in the processing pipeline, immediately following initial patch extraction. This strategy offers multiple advantages: (1) Leveraging the convolutional nature of the initial patch extraction allows for a direct mapping between our PSM $\\mathcal{P}$ and the patch tokens $X_{p}$ , obviating the need for additional mechanisms to determine token redundancy. (2) It facilitates generalizability across transformer models without the need for retraining, potentially accelerating training by concentrating on essential patches. (3) It permits seamless integration of existing intermediate layer token reduction techniques following our initial patch pruning process.", + "bbox": [ + 212, + 459, + 787, + 595 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "3.4 Integrating PSM with Hierarchical Models", + "text_level": 1, + "bbox": [ + 214, + 617, + 614, + 633 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "ConvNets and hybrid transformers, such as Swin [31], primarily operate with window-based local operations in contrast to full-attention based operation in vanilla transformers. Such window-based local processing makes patch pruning particularly complicated, in contrast to vanilla transformers. However, these models maintain the location property of representative image patches all through the network, which leaves the door open to prune redundant patches. Nevertheless, the window-based operation is particularly difficult to prune.", + "bbox": [ + 212, + 643, + 787, + 750 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "In general, each block of hierarchical layers consists of window-based spatial kernel/attention operators, followed by pixel operators (usually, modeled with $1 \\times 1$ convolutional layers or linear layers). In contrast to windowed convolutions or cross-attention, these pixel operators are particularly suitable for patch pruning. Interestingly, in the SOTA ConvNets and hybrid transformers, more than $60\\%$ of total computations are performed with such pixel operators (63.3% in Swin [31],", + "bbox": [ + 212, + 750, + 787, + 840 + ], + "page_idx": 7 + }, + { + "type": "page_number", + "text": "8", + "bbox": [ + 217, + 114, + 228, + 126 + ], + "page_idx": 7 + }, + { + "type": "header", + "text": "T. Mahmud et al.", + "bbox": [ + 271, + 114, + 392, + 127 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/c427eb00d0e8b6575f9866740cad876e19f4ab819010fc3fc68aec68cebd876d.jpg", + "image_caption": [ + "Fig. 4: ImageNet-1k evaluation for varying top-k accuracy targets. The accuracy gain with bigger model largely shrinks, as $k$ increases. This suggests shallower ConvNets have understanding of object locations and visual property, despite their lower top-1 accuracy." + ], + "image_footnote": [], + "bbox": [ + 223, + 145, + 496, + 268 + ], + "page_idx": 8 + }, + { + "type": "image", + "img_path": "images/8780e33834425319c90bbbf763f15d7615eff67316943ce4be7a9aec94d968a4.jpg", + "image_caption": [ + "Fig.5: Integrating PaPr with ToMe [2]. We use the Augreg pretrained ViT-B-16 architecture as the baseline. We sweep token merging ratio (r) for different pruning ratio (z). Integration of PaPr achieves Pareto-optimal performance, thus, PaPr can enhance existing patch reduction methods." + ], + "image_footnote": [], + "bbox": [ + 506, + 143, + 779, + 268 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "$96.2\\%$ in ConvNext [32]). Based on the patch significance map from PaPr, we only use the pixel operator on most significant patches, and simply perform zero-padding on the remaining patch regions. The zero padding operations are performed to mostly recover the feature map shape to be used with subsequent spatial operators. Hence, the speed-up is mostly achieved by eliminating redundant computations in pixel operators (See Fig. 3).", + "bbox": [ + 212, + 402, + 787, + 492 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "Let's assume that the feature map before applying pixel operators is given by $\\mathcal{A}_w$ . Based on the PSM $\\mathcal{P}$ , we initially identify the foreground and background pixel features as $A_{f}$ , and $A_{b}$ , respectively. Later, the pixel operator layers are applied on foreground pixels $A_{f}$ , and a zero-padded representation is used for the background pixels $A_{b}$ . Finally, the output representation $\\mathcal{A}_c$ is reassembled with modulated foreground pixels $A_{f}$ and zero-padded background pixels $A_{z}$ , given by Eq. (4):", + "bbox": [ + 212, + 494, + 787, + 599 + ], + "page_idx": 8 + }, + { + "type": "equation", + "text": "\n$$\nA _ {f}, A _ {b} = \\operatorname {S p l i t} \\left(\\mathcal {A} _ {w}, \\mathcal {P}\\right),\n$$\n", + "text_format": "latex", + "bbox": [ + 460, + 599, + 624, + 616 + ], + "page_idx": 8 + }, + { + "type": "equation", + "text": "\n$$\nA _ {z} = \\operatorname {Z e r o s} \\left(A _ {b}\\right), A _ {f} = \\operatorname {L i n e a r} \\left(A _ {f}\\right), \\tag {4}\n$$\n", + "text_format": "latex", + "bbox": [ + 380, + 618, + 785, + 635 + ], + "page_idx": 8 + }, + { + "type": "equation", + "text": "\n$$\n\\mathcal {A} _ {c} = \\operatorname {R e a s s e m b l e} \\left(A _ {f}, A _ {z}\\right).\n$$\n", + "text_format": "latex", + "bbox": [ + 434, + 638, + 624, + 652 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "4 Image Experiments", + "text_level": 1, + "bbox": [ + 215, + 675, + 441, + 693 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "4.1 Experimental Setup", + "text_level": 1, + "bbox": [ + 215, + 708, + 426, + 724 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "We experiment on image classification task on the ImageNet-1k benchmark dataset, following prior work [2, 12, 23, 34, 55]. We use the MobileOne-s0 [49] model as the proposal model for all architectures, unless otherwise specified. We report training-free results of PaPr, unless otherwise specified. For the performance metric, we report top-1 accuracy, GFLOPs, and the throughput (img/s). Throughput is measured on a single RTX-A5000 GPU with 24GB VRAM. We reproduced all baseline models under the same setup for a fair comparison.", + "bbox": [ + 212, + 733, + 787, + 840 + ], + "page_idx": 8 + }, + { + "type": "header", + "text": "PaPr for Faster Inference", + "bbox": [ + 560, + 114, + 730, + 126 + ], + "page_idx": 8 + }, + { + "type": "page_number", + "text": "9", + "bbox": [ + 774, + 114, + 784, + 126 + ], + "page_idx": 8 + }, + { + "type": "table", + "img_path": "images/763a884e2d972a0963e4b8753d236ba9ef32318929a0fb1a7fc98fd5cea0101e.jpg", + "table_caption": [ + "Table 1: Performance comparison on AugReg models. PaPr achieves the best performance, while operating with token merging. $z$ denotes patch keeping ratio in PaPr." + ], + "table_footnote": [], + "table_body": "
ModelsMethodsAcc1GFLOPsImg/s
ViT-S-16Baseline81.394.61975
ToMe [2]76.962.291978
TokenFusion [23]77.122.291982
GTP-ViT [55]71.032.311970
PaPr (z=0.45)76.212.291988
PaPr (z=0.55) (+ToMe)77.762.192073
PaPr (z=0.5) (+ToMe)76.271.972315
ViT-B-16Baseline84.5917.59307
ToMe [2]80.388.78615
TokenFusion [23]80.78.78618
GTP-ViT [55]80.988.78610
PaPr (z=0.5)82.118.98605
PaPr (z=0.55) (+ToMe)82.348.21660
PaPr (z=0.5) (+ToMe)80.886.82785
ViT-L-16Baseline85.8261.6191
ToMe [2]83.530.99180
TokenFusion [23]83.9130.99182
GTP-ViT [55]81.5630.99181
PaPr (z=0.5)83.8730.83183
PaPr (z=0.55) (+ToMe)83.9926.33210
PaPr (z=0.55) (+ToMe)83.525.12224
", + "bbox": [ + 223, + 214, + 506, + 450 + ], + "page_idx": 9 + }, + { + "type": "table", + "img_path": "images/b47354f95bd87e820185c52ea83fa2e43a945bb224a8295bc17bac26ae6b192a.jpg", + "table_caption": [ + "Table 2: Performance comparison on MAE models. Since MAE uses masked pretraining, PaPr is particularly suitable for MAE inference. PaPr achieves significantly higher performance than others. $z$ denotes patch keeping ratio in PaPr." + ], + "table_footnote": [], + "table_body": "
ModelsMethodsAcc1GFLOPsImg/s
ViT-B-16Baseline83.7417.59307
ToMe [2]78.828.78615
TokenFusion [23]79.238.78618
GTP-ViT [55]79.148.78610
PaPr (z=0.5)82.48.98605
PaPr (z=0.4)81.47.72700
ViT-L-16Baseline85.9561.6191
ToMe [2]84.2430.99180
TokenFusion [23]84.3330.99182
GTP-ViT [55]84.1530.99181
PaPr (z=0.5)85.0630.83183
PaPr (z=0.4)84.7627.72201
ViT-H-16Baseline86.89167.436
ToMe [2]85.4882.5372
TokenFusion [23]85.7182.5373
GTP-ViT [55]85.5482.5371
PaPr (z=0.5)86.483.0471
PaPr (z=0.4)86.1374.5981
", + "bbox": [ + 519, + 239, + 777, + 452 + ], + "page_idx": 9 + }, + { + "type": "text", + "text": "4.2 Performance on Various Vision Transformers", + "text_level": 1, + "bbox": [ + 215, + 479, + 633, + 493 + ], + "page_idx": 9 + }, + { + "type": "text", + "text": "We study performance of PaPr in diverse ViT architectures, along with various pre-training methods. We also present training results, along with training-free results to compare with existing methods.", + "bbox": [ + 212, + 502, + 785, + 547 + ], + "page_idx": 9 + }, + { + "type": "text", + "text": "Training-free method comparison. We consider two different pre-training methods for comparative analysis on various ViT architectures. We use supervised pretrained Augreg models [45], and self-supervised pre-trained masked autoencoder (MAE) models [17], following prior work [2]. We compare with three recent training-free patch reduction methods, such as ToMe [2], TokenFusion [23], and GTP-ViT [55]. Additionally, we show the performance comparison on class-token free ViT models to highlight the architecture agnostic performance. In general, PaPr achieves better accuracy, with lower computational costs compared to state-of-the-art (SOTA) methods.", + "bbox": [ + 212, + 569, + 787, + 705 + ], + "page_idx": 9 + }, + { + "type": "text", + "text": "Augreg models: For AugReg pre-training with ViT-Base, PaPr achieves $2.14\\%$ higher accuracy over ToMe, with comparable FLOPs as shown in Tab. 1. In addition, by combining PaPr with ToMe, we achieve additional $22.3\\%$ FLOP reduction, while maintaining higher accuracy. We further study the compatibility of PaPr with existing patch reduction methods as ToMe. After the initial patch pruning with PaPr, we integrate ToMe only at the bottom-half layers due to its higher sensitivity in earlier layers. As shown in Fig. 5, integration of PaPr can directly boost ToMe performance thereby achieving Pareto optimality.", + "bbox": [ + 212, + 718, + 787, + 840 + ], + "page_idx": 9 + }, + { + "type": "page_number", + "text": "10", + "bbox": [ + 217, + 114, + 235, + 126 + ], + "page_idx": 9 + }, + { + "type": "header", + "text": "T. Mahmud et al.", + "bbox": [ + 271, + 114, + 392, + 127 + ], + "page_idx": 9 + }, + { + "type": "table", + "img_path": "images/b6fb9adb44c49c914e1c3c4a144082584c2794a10e78d201af053f12aa33c030.jpg", + "table_caption": [ + "Table 3: Performance analysis on class-token free ViT models. PaPr performance gain is not limited to specific architectures." + ], + "table_footnote": [], + "table_body": "
ModelMethodAcc1GFLOPsImg/s
ViT-MediumBaseline84.3310.58484
GAP-16GTP-ViT [55]80.55.73880
256ToMe [2]80.215.73890
PaPr (z=0.5)81.85.5932
ViT-MediumBaseline85.626.06182
GAP-16GTP-ViT [55]81.9513.59340
384ToMe [2]82.513.59344
PaPr83.9512.94360
", + "bbox": [ + 228, + 198, + 496, + 309 + ], + "page_idx": 10 + }, + { + "type": "table", + "img_path": "images/87ea40266a13d60a8025276ff69aff57573389808ea19476e6c205ccfafbcfae.jpg", + "table_caption": [ + "Table 4: Training performance analysis on DeIT-s. PaPr achieves competitive performance with higher training speed." + ], + "table_footnote": [], + "table_body": "
MethodsAcc1GFLOPsIm/sBatch Proc.Train Speed
Baseline79.84.6960✓1x
DynamicViT [41]79.32.91510X1x
A-ViT [56]78.62.9-X1x
ATS [12]79.52.91512-1x
SP-ViT [25]79.32.6-X1x
ToMe [2]79.42.71575✓1.5x
TokenFusion [23]79.52.71580✓1.5x
PaPr (z=0.55)79.22.71585✓1.6x
", + "bbox": [ + 511, + 199, + 777, + 309 + ], + "page_idx": 10 + }, + { + "type": "table", + "img_path": "images/f62733fddaa68b57f51440ec3d9800cb7b5ddeaa608c169cecdb485546816c6c.jpg", + "table_caption": [ + "Table 5: Performance comparison on ConvNext CNN models. PaPr can achieve competitive performance without training, and can seamlessly adapt to model updates." + ], + "table_footnote": [], + "table_body": "
ModelsMethodsTrain-FreeAcc1GFLOPsImg/s
ConvNeXtBaselineN/A83.8415.38265
Base-1kDynCNN [40]✘83.0810.21375
PaPr (z=0.65)✓82.7510.42395
ConvNeXtBaselineN/A85.8115.38265
Base-22kPaPr (z=0.65)✓84.2710.42395
ConvNeXtBaselineN/A84.3134.4135
Large-1kPaPr (z=0.65)✓83.2622.9203
ConvNeXtBaselineN/A86.6134.4135
Large-22kPaPr (z=0.65)✓85.6722.9203
", + "bbox": [ + 225, + 395, + 496, + 506 + ], + "page_idx": 10 + }, + { + "type": "table", + "img_path": "images/25edff39f045372724e390fa2e27cb83dcbdfe970019b82cdd1caf281da76d6a.jpg", + "table_caption": [ + "Table 6: Performance comparison on Swin hybrid transformer models. PaPr can adapt to much bigger models with higher operating resolutions without training." + ], + "table_footnote": [], + "table_body": "
ModelsMethodsTrain-FreeAcc1GFLOPsImg/s
Swin B-1kBaselineN/A83.4215.47258
DynSwin [40]✗83.1812.1327
PaPr (z=0.65)✓81.712.25325
Swin B-22kBaselineN/A85.1615.47258
PaPr (z=0.65)✓82.2712.25325
Swin L-22kBaselineN/A86.2534.53135
PaPr (z=0.65)✓84.5326.96175
Swin L-22k-384BaselineN/A87.25104.0842
PaPr (z=0.65)✓86.4781.4454
", + "bbox": [ + 511, + 395, + 777, + 506 + ], + "page_idx": 10 + }, + { + "type": "text", + "text": "MAE models: MAE used self-supervised pre-training by reconstructing masked image patches to train large ViT models [17]. Later, the model is fine-tuned with full resolution images without masking, which cannot exploit its latent ability to learn from fewer patches. Interestingly, PaPr introduces masked inference, which is particularly suitable for MAE models. Hence, we observe significantly higher accuracy with PaPr compared to other training-free methods, e.g., with ViT-B, PaPr achieves $4.5\\%$ higher accuracy over ToMe for similar FLOPs (see Tab. 2).", + "bbox": [ + 212, + 537, + 787, + 645 + ], + "page_idx": 10 + }, + { + "type": "text", + "text": "Class token-free models: Several existing patch reduction methods operate with the class token in ViTs to evaluate the patch relevance [12,28,34]. However, instead of class tokens, global pooling of patch tokens is used in many cases [16, 37]. Results presented in Tab. 3 demonstrate the superiority of PaPr for such cases.", + "bbox": [ + 212, + 670, + 787, + 732 + ], + "page_idx": 10 + }, + { + "type": "text", + "text": "Training-based methods comparison. We compare training performance of existing methods for the patch reduction on DeiT-s [48] model in Tab. 4. Prior pruning based methods cannot speed-up the training for learning the mask predictor [25,41,56]. However, PaPr achieves competitive performance as prior methods, while achieving large speed-up as other token merging methods [2,23,55].", + "bbox": [ + 212, + 763, + 787, + 842 + ], + "page_idx": 10 + }, + { + "type": "header", + "text": "PaPr for Faster Inference", + "bbox": [ + 560, + 114, + 730, + 127 + ], + "page_idx": 10 + }, + { + "type": "page_number", + "text": "11", + "bbox": [ + 767, + 114, + 782, + 126 + ], + "page_idx": 10 + }, + { + "type": "table", + "img_path": "images/d78f4788d79411df097b55d05c9d8fc460101ec8d5382733ac373bb2d8c490fe.jpg", + "table_caption": [ + "Table 7: Sweeping ConvNet proposal model in PaPr with $z = 0.5$ . PaPr can achieve similar performance irrespective of proposal model size. Thus, PaPr can use a much smaller and faster proposal model to speed-up larger models." + ], + "table_footnote": [], + "table_body": "
ProposalAccuracy1(%)
ModelGFLOPsViT-B-16ViT-L-16
ResNet-18 [18]1.8181.183.84
ResNet-50 [18]4.0982.3384.09
ResNet-152 [18]11.5182.5184.08
MobileOne-S0 [49]0.2782.2484.06
MobileOne-S2 [49]1.3582.2884.16
MobileOne-S4 [49]2.9882.3584.32
", + "bbox": [ + 235, + 237, + 488, + 339 + ], + "page_idx": 11 + }, + { + "type": "image", + "img_path": "images/772790e3fcf8e48f2e99138562f527bb7eb88b103832ad836105ed30f82e88b5.jpg", + "image_caption": [ + "Fig.6: Proposal models locate similar patches with PaPr for different pruning ratio $(z)$ irrespective of model size." + ], + "image_footnote": [], + "bbox": [ + 511, + 143, + 769, + 294 + ], + "page_idx": 11 + }, + { + "type": "table", + "img_path": "images/259bdc422f17add6668acccca35de691553f0c97b88bf7723b16e9da1b9abc20.jpg", + "table_caption": [ + "Table 8: Training-free performance comparison on Kinetics-400 video evaluation. Each video has $16 \\times 224^2$ input size. PaPr achieves significantly better performance for reducing spatio-temporal redundancy in videos." + ], + "table_footnote": [], + "table_body": "
ModelMethodAcc1GFLOPsViews
XViTATS [12]80.02591x3
TimeSformer-LATS [12]80.535101x3
ViT-B MAEBaseline81.211803x5
PaPr(z=0.45)81.18763x5
PaPr(z=0.35)80.15593x5
ViT-L MAEBaseline85.265983x5
ToMe [2]84.52811x10
PaPr(z=0.5)85.122753x5
ToMe [2]82.51841x10
PaPr(z=0.3)84.531603x5
", + "bbox": [ + 235, + 440, + 504, + 578 + ], + "page_idx": 11 + }, + { + "type": "table", + "img_path": "images/99195efc99a2eb99a856fc53d0173d0c576c5e6800af19c84e1261f494ce2528.jpg", + "table_caption": [ + "Table 9: Proposal ConvNet sweep in video recognition with ViT-B-MAE model as the baseline. PaPr achieves competitive performance with lighter proposal models for different $z$ values." + ], + "table_footnote": [], + "table_body": "
ProposalKeep Ratio(z)Acc1 (%)Views
ModelGFLOPs
X3d-s [13]1.250.581.113x5
X3d-m [13]2.450.581.193x5
ResNet-50 [14]41.90.579.963x5
ResNet-101 [14]85.670.579.973x5
X3d-s [13]1.250.480.863x5
X3d-m [13]2.450.480.933x5
ResNet-50 [14]41.90.478.853x5
ResNet-101 [14]85.670.479.323x5
", + "bbox": [ + 521, + 441, + 781, + 577 + ], + "page_idx": 11 + }, + { + "type": "text", + "text": "4.3 Performance on Various Hierarchical Models", + "text_level": 1, + "bbox": [ + 215, + 609, + 632, + 625 + ], + "page_idx": 11 + }, + { + "type": "text", + "text": "We study the performance on two variants of hierarchical models, such as pure convolutional models, and hybrid transformer models. We compare with training based DynamicCNN [40], and DynamicSwin [40] methods (trained for 120 epochs), whereas PaPr operates without training.", + "bbox": [ + 212, + 642, + 787, + 704 + ], + "page_idx": 11 + }, + { + "type": "text", + "text": "Analysis on convolutional models. We use the SOTA ConvNext [32] architecture for the analysis as shown in Tab. 5. The training-free operation in PaPr makes it seamlessly usable with new model updates. We analyze ImageNet-1k and ImageNet-22k performance of same models. For the ConvNext-Base model, PaPr achieves $99.6\\%$ accuracy of DynamicCNN with $7.1\\%$ higher throughput. By simply using ImageNet-22k weights, PaPr can achieve additional $2\\%$ accuracy improvement without training, while having the same computation cost.", + "bbox": [ + 212, + 733, + 789, + 840 + ], + "page_idx": 11 + }, + { + "type": "page_number", + "text": "12", + "bbox": [ + 217, + 114, + 235, + 126 + ], + "page_idx": 11 + }, + { + "type": "header", + "text": "T. Mahmud et al.", + "bbox": [ + 271, + 114, + 392, + 127 + ], + "page_idx": 11 + }, + { + "type": "image", + "img_path": "images/fe51584d2962fd0367d1dce2c621ada35d997467a3e559bfa97aa8b08f5a3857.jpg", + "image_caption": [ + "Fig. 7: Robustness of PaPr compared to CAM based methods. PaPr can perform even when the ConvNet proposal confidence (c) is very low. In contrast, existing CAM based methods fail in such cases, despite being significantly slower while not enabling batch processing and use of gradients in some cases." + ], + "image_footnote": [], + "bbox": [ + 256, + 143, + 743, + 371 + ], + "page_idx": 12 + }, + { + "type": "text", + "text": "Analysis on hybrid transformer models. We study the hierarchical Swin transformer models for patch reduction with PaPr, as given in Tab. 6. PaPr achieves $98\\%$ of DynamicSwin accuracy using similar FLOPs without re-training. Nevertheless, by leveraging bigger models, larger pre-training, and higher resolution, PaPr can seamlessly adapt to achieve higher accuracy.", + "bbox": [ + 212, + 460, + 789, + 537 + ], + "page_idx": 12 + }, + { + "type": "text", + "text": "4.4 Robustness of PaPr across various ConvNet proposals", + "text_level": 1, + "bbox": [ + 212, + 558, + 707, + 574 + ], + "page_idx": 12 + }, + { + "type": "text", + "text": "PaPr can operate with ultra-lightweight ConvNets for generating robust proposals. We study different ConvNet architectures for proposal generation, as shown in Tab. 7. Interestingly, PaPr shows small reductions of final accuracy $(0.3\\%)$ , when using MobileOne-S0 based proposals compared to ResNet-152 $(42\\times$ higher FLOPs) in ViT-B. When visualizing the PSM for different pruning ratio as in Fig. 6, we notice the similar PSMs irrespective of model size. Hence, PaPr can utilize ultra-lightweight ConvNets to mask patches without sacrificing accuracy.", + "bbox": [ + 212, + 582, + 787, + 689 + ], + "page_idx": 12 + }, + { + "type": "text", + "text": "4.5 Comparison with Class Activation Mappings (CAMs)", + "text_level": 1, + "bbox": [ + 212, + 709, + 702, + 726 + ], + "page_idx": 12 + }, + { + "type": "text", + "text": "Existing CAM methods mostly focus on explainability to highlight image regions responsible for final prediction. However, such objectives rely on final prediction performance, which can be much lower for light ConvNets. Moreover, for localization, many of these methods use gradients, and optimization methods that limit batch processing. Nevertheless, we study the impact of such CAM methods ([43, 50, 61]) in PaPr framework, in challenging scenarios where the baseline ViT-B has higher prediction confidence (c) on target class, and the proposal", + "bbox": [ + 212, + 734, + 787, + 842 + ], + "page_idx": 12 + }, + { + "type": "header", + "text": "PaPr for Faster Inference", + "bbox": [ + 560, + 114, + 730, + 126 + ], + "page_idx": 12 + }, + { + "type": "page_number", + "text": "13", + "bbox": [ + 767, + 114, + 785, + 126 + ], + "page_idx": 12 + }, + { + "type": "image", + "img_path": "images/58041a1a8ab5a9067821c623a7a89ccb36876571cfd2c6359efec62cb69c21d8.jpg", + "image_caption": [ + "Fig. 8: Visualization of PaPr localization in videos. Video has inherent high sparsity. PaPr effectively localizes the discriminative regions for using holistic spatio-temporal understanding with small ConvNets. Thus, it significantly reduces the computational burden for larger models in downstream video recognition tasks." + ], + "image_footnote": [], + "bbox": [ + 259, + 143, + 743, + 239 + ], + "page_idx": 13 + }, + { + "type": "text", + "text": "MobileOne-S0 has lower confidence (see Fig. 7). In most cases, the baseline CAM method significantly lowers the final accuracy after patch pruning. In contrast, PaPr maintains robust confidence for its precise localization. Moreover, PaPr can enhance confidence of baseline models in several scenarios by suppressing the redundant patches, e.g., in Sample 5, the $c$ increased by $22\\%$ with PaPr.", + "bbox": [ + 212, + 325, + 787, + 402 + ], + "page_idx": 13 + }, + { + "type": "text", + "text": "5 Video Experiments", + "text_level": 1, + "bbox": [ + 215, + 424, + 439, + 443 + ], + "page_idx": 13 + }, + { + "type": "text", + "text": "We study the training-free performance on Kinetics-400 [22] validation sets, as shown in Tab. 8. We use SOTA ATS [12], and ToMe [2] methods for comparison. We use PaPr on SOTA ViT-MAE [47] models with lightweight X3d-s [13] model for proposal generation. We follow the baseline [47] $3 \\times 5$ view (3 spatial view and 5 temporal view) approach, and separate the views from model FLOP counts similar to other works as it can be arbitrarily chosen. Since ViT-MAE removes class tokens in fine-tuning, ATS [12] cannot be adapted to these models. For ViT-L, PaPr achieves $3.7 \\times$ FLOPs reduction of the baseline model for a negligible $0.8\\%$ accuracy drop. We also study the impact of various proposal models on the final performance, as shown in Tab. 9. We use SlowOnly-ResNet [14] and X3d [13] models for comparison. Increasing model size for different pruning shows minimal impact on final performance. Finally, we visualize the patch masking of PaPr in videos, as shown in Fig. 8. By removing the redundant details, PaPr significantly boosts the performance of bigger models for fine-grained predictions.", + "bbox": [ + 212, + 454, + 789, + 667 + ], + "page_idx": 13 + }, + { + "type": "text", + "text": "6 Conclusion and Future Work", + "text_level": 1, + "bbox": [ + 215, + 688, + 532, + 704 + ], + "page_idx": 13 + }, + { + "type": "text", + "text": "In this paper, we introduce a novel patch pruning method, namely PaPr, that can effectively speed-up off-the-shelf pre-trained models inference without re-training. We propose a simple modification of ConvNets, that allows extracting a precise discriminative PSM from ultra lightweight ConvNets with unparalleled speed. In PaPr, such PSMs are leveraged to directly reduce the redundant data regions in a single-step to guide larger model computations on most discriminative regions for achieving computational efficiency. Moreover, PaPr can be easily integrated with most existing patch reduction methods due to its simple structure.", + "bbox": [ + 212, + 719, + 787, + 840 + ], + "page_idx": 13 + }, + { + "type": "page_number", + "text": "14", + "bbox": [ + 217, + 114, + 235, + 126 + ], + "page_idx": 13 + }, + { + "type": "header", + "text": "T. Mahmud et al.", + "bbox": [ + 271, + 114, + 392, + 126 + ], + "page_idx": 13 + }, + { + "type": "text", + "text": "Acknowledgement", + "text_level": 1, + "bbox": [ + 217, + 143, + 393, + 162 + ], + "page_idx": 14 + }, + { + "type": "text", + "text": "This work was supported in part by ONR Minerva program, iMAGiNE - the Intelligent Machine Engineering Consortium at UT Austin, and a UT Cockrell School of Engineering Doctoral Fellowship.", + "bbox": [ + 212, + 172, + 787, + 219 + ], + "page_idx": 14 + }, + { + "type": "text", + "text": "References", + "text_level": 1, + "bbox": [ + 217, + 239, + 323, + 255 + ], + "page_idx": 14 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "1. Belharbi, S., Sarraf, A., Pedersoli, M., Ben Ayed, I., McCaffrey, L., Granger, E.: F-cam: Full resolution class activation maps via guided parametric upscaling. In: Proceedings of the IEEE/CVF Winter Conference on Applications of Computer Vision. pp. 3490-3499 (2022)", + "2. Bolya, D., Fu, C.Y., Dai, X., Zhang, P., Feichtenhofer, C., Hoffman, J.: Token merging: Your vit but faster. In: The Twelfth International Conference on Learning Representations (2023)", + "3. Cao, Q., Paranjape, B., Hajishirzi, H.: Pumer: Pruning and merging tokens for efficient vision language models. arXiv preprint arXiv:2305.17530 (2023)", + "4. Caron, M., Touvron, H., Misra, I., Jégou, H., Mairal, J., Bojanowski, P., Joulin, A.: Emerging properties in self-supervised vision transformers. In: Proceedings of the IEEE/CVF international conference on computer vision. pp. 9650-9660 (2021)", + "5. Chattopadhyay, A., Sarkar, A., Howlader, P., Balasubramanian, V.N.: Grad-cam++: Generalized gradient-based visual explanations for deep convolutional networks. In: 2018 IEEE winter conference on applications of computer vision (WACV). pp. 839-847. IEEE (2018)", + "6. Chen, Y., Liu, H., Yin, H., Fan, B.: Building vision transformers with hierarchy aware feature aggregation. In: Proceedings of the IEEE/CVF International Conference on Computer Vision (ICCV). pp. 5908-5918 (October 2023)", + "7. Chu, X., Tian, Z., Wang, Y., Zhang, B., Ren, H., Wei, X., Xia, H., Shen, C.: Twins: Revisiting the design of spatial attention in vision transformers. Advances in Neural Information Processing Systems 34, 9355-9366 (2021)", + "8. Ding, X., Zhang, X., Ma, N., Han, J., Ding, G., Sun, J.: Repvgg: Making vgg-style convnets great again. In: Proceedings of the IEEE/CVF conference on computer vision and pattern recognition. pp. 13733-13742 (2021)", + "9. Dong, X., Bao, J., Chen, D., Zhang, W., Yu, N., Yuan, L., Chen, D., Guo, B.: Cswin transformer: A general vision transformer backbone with cross-shaped windows. In: CVPR (2022)", + "0. Dosovitskiy, A., Beyer, L., Kolesnikov, A., Weissenborn, D., Zhai, X., Unterthiner, T., Dehghani, M., Minderer, M., Heigold, G., Gelly, S., et al.: An image is worth 16x16 words: Transformers for image recognition at scale. In: ICLR (2020)", + "1. Fan, H., Xiong, B., Mangalam, K., Li, Y., Yan, Z., Malik, J., Feichtenhofer, C.: Multiscale vision transformers. In: ICCV (2021)", + "2. Fayyaz, M., Kouhpayegani, S.A., Jafari, F.R., Sommerlade, E., Joze, H.R.V., Pirsivash, H., Gall, J.: Ats: Adaptive token sampling for efficient vision transformers. In: ECCV (2022)", + "3. Feichtenhofer, C.: X3d: Expanding architectures for efficient video recognition. In: Proceedings of the IEEE/CVF conference on computer vision and pattern recognition. pp. 203-213 (2020)", + "4. Feichtenhofer, C., Fan, H., Malik, J., He, K.: Slowfast networks for video recognition. In: Proceedings of the IEEE/CVF international conference on computer vision. pp. 6202-6211 (2019)" + ], + "bbox": [ + 225, + 266, + 785, + 839 + ], + "page_idx": 14 + }, + { + "type": "header", + "text": "PaPr for Faster Inference", + "bbox": [ + 560, + 114, + 730, + 127 + ], + "page_idx": 14 + }, + { + "type": "page_number", + "text": "15", + "bbox": [ + 767, + 116, + 785, + 126 + ], + "page_idx": 14 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "15. Graham, B., El-Nouby, A., Touvron, H., Stock, P., Joulin, A., Jégou, H., Douze, M.: Levit: a vision transformer in convnet's clothing for faster inference. In: ICCV (2021)", + "16. Hatamizadeh, A., Yin, H., Heinrich, G., Kautz, J., Molchanov, P.: Global context vision transformers. In: International Conference on Machine Learning. pp. 12633-12646. PMLR (2023)", + "17. He, K., Chen, X., Xie, S., Li, Y., Dollar, P., Girshick, R.: Masked autoencoders are scalable vision learners. In: CVPR (2022)", + "18. He, K., Zhang, X., Ren, S., Sun, J.: Deep residual learning for image recognition. In: Proceedings of the IEEE conference on computer vision and pattern recognition. pp. 770-778 (2016)", + "19. Howard, A., Sandler, M., Chu, G., Chen, L.C., Chen, B., Tan, M., Wang, W., Zhu, Y., Pang, R., Vasudevan, V., et al.: Searching for mobilenetv3. In: Proceedings of the IEEE/CVF international conference on computer vision. pp. 1314-1324 (2019)", + "20. Howard, A.G., Zhu, M., Chen, B., Kalenichenko, D., Wang, W., Weyand, T., Andreetto, M., Adam, H.: Mobilenets: Efficient convolutional neural networks for mobile vision applications. arXiv preprint arXiv:1704.04861 (2017)", + "21. Jung, H., Oh, Y.: Towards better explanations of class activation mapping. In: Proceedings of the IEEE/CVF International Conference on Computer Vision. pp. 1336-1344 (2021)", + "22. Kay, W., Carreira, J., Simonyan, K., Zhang, B., Hillier, C., Vijayanarasimhan, S., Viola, F., Green, T., Back, T., Natsev, P., et al.: The kinetics human action video dataset. arXiv:1705.06950 [cs.CV] (2017)", + "23. Kim, M., Gao, S., Hsu, Y.C., Shen, Y., Jin, H.: Token fusion: Bridging the gap between token pruning and token merging. In: Proceedings of the IEEE/CVF Winter Conference on Applications of Computer Vision. pp. 1383-1392 (2024)", + "24. Kim, S., Shen, S., Thorsley, D., Gholami, A., Kwon, W., Hassoun, J., Keutzer, K.: Learned token pruning for transformers. arXiv:2107.00910 [cs.CL] (2021)", + "25. Kong, Z., Dong, P., Ma, X., Meng, X., Niu, W., Sun, M., Ren, B., Qin, M., Tang, H., Wang, Y.: Spvit: Enabling faster vision transformers via soft token pruning. In: ECCV (2022)", + "26. Lassance, C., Maachou, M., Park, J., Clinchant, S.: A study on token pruning for colbert. arXiv:2112.06540 [cs.CL] (2021)", + "27. Li, Y., Wu, C.Y., Fan, H., Mangalam, K., Xiong, B., Malik, J., Feichtenhofer, C.: Mvittv2: Improved multiscale vision transformers for classification and detection. In: CVPR (2022)", + "28. Liang, Y., Ge, C., Tong, Z., Song, Y., Wang, J., Xie, P.: Not all patches are what you need: Expediting vision transformers via token reorganizations. ICLR (2022)", + "29. Lin, J., Chen, W.M., Cai, H., Gan, C., Han, S.: Memory-efficient patch-based inference for tiny deep learning. Advances in Neural Information Processing Systems 34, 2346-2358 (2021)", + "30. Liu, Z., Hu, H., Lin, Y., Yao, Z., Xie, Z., Wei, Y., Ning, J., Cao, Y., Zhang, Z., Dong, L., et al.: Swin transformer v2: Scaling up capacity and resolution. In: CVPR (2022)", + "31. Liu, Z., Lin, Y., Cao, Y., Hu, H., Wei, Y., Zhang, Z., Lin, S., Guo, B.: Swin transformer: Hierarchical vision transformer using shifted windows. In: ICCV (2021)", + "32. Liu, Z., Mao, H., Wu, C.Y., Feichtenhofer, C., Darrell, T., Xie, S.: A convnet for the 2020s. In: Proceedings of the IEEE/CVF conference on computer vision and pattern recognition. pp. 11976-11986 (2022)" + ], + "bbox": [ + 215, + 146, + 785, + 839 + ], + "page_idx": 15 + }, + { + "type": "page_number", + "text": "16", + "bbox": [ + 217, + 114, + 235, + 126 + ], + "page_idx": 15 + }, + { + "type": "header", + "text": "T. Mahmud et al.", + "bbox": [ + 271, + 114, + 390, + 127 + ], + "page_idx": 15 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "33. Mehta, S., Rastegari, M.: Mobilevit: light-weight, general-purpose, and mobile-friendly vision transformer. In: ICLR (2021)", + "34. Meng, L., Li, H., Chen, B.C., Lan, S., Wu, Z., Jiang, Y.G., Lim, S.N.: Adavit: Adaptive vision transformers for efficient image recognition. In: CVPR (2022)", + "35. Muhammad, M.B., Yeasin, M.: Eigen-cam: Class activation map using principal components. In: 2020 international joint conference on neural networks (IJCNN). pp. 1-7. IEEE (2020)", + "36. Oquab, M., Darcet, T., Moutakanni, T., Vo, H., Szafraniec, M., Khalidov, V., Fernandez, P., Haziza, D., Massa, F., El-Nouby, A., et al.: Dinov2: Learning robust visual features without supervision. arXiv preprint arXiv:2304.07193 (2023)", + "37. Pan, Z., Zhuang, B., Liu, J., He, H., Cai, J.: Scalable vision transformers with hierarchical pooling. In: Proceedings of the IEEE/cvf international conference on computer vision. pp. 377-386 (2021)", + "38. Patro, B.N., Lunayach, M., Patel, S., Namboodiri, V.P.: U-cam: Visual explanation using uncertainty based class activation maps. In: Proceedings of the IEEE/CVF International Conference on Computer Vision. pp. 7444-7453 (2019)", + "39. Radford, A., Kim, J.W., Hallacy, C., Ramesh, A., Goh, G., Agarwal, S., Sastry, G., Askell, A., Mishkin, P., Clark, J., et al.: Learning transferable visual models from natural language supervision. In: International conference on machine learning. pp. 8748-8763. PMLR (2021)", + "40. Rao, Y., Liu, Z., Zhao, W., Zhou, J., Lu, J.: Dynamic spatial sparsification for efficient vision transformers and convolutional neural networks. IEEE Transactions on Pattern Analysis and Machine Intelligence (2023)", + "41. Rao, Y., Zhao, W., Liu, B., Lu, J., Zhou, J., Hsieh, C.J.: Dynamicvit: Efficient vision transformers with dynamic token sparsification. Advances in neural information processing systems 34, 13937-13949 (2021)", + "42. Sandler, M., Howard, A., Zhu, M., Zhmoginov, A., Chen, L.C.: Mobilenetv2: Inverted residuals and linear bottlenecks. In: Proceedings of the IEEE conference on computer vision and pattern recognition. pp. 4510-4520 (2018)", + "43. Selvaraju, R.R., Cogswell, M., Das, A., Vedantam, R., Parikh, D., Batra, D.: Gradcam: Visual explanations from deep networks via gradient-based localization. In: ICCV (2017)", + "44. Singh, M., Gustafson, L., Adcock, A., de Freitas Reis, V., Gedik, B., Kosaraju, R.P., Mahajan, D., Girshick, R., Dollar, P., van der Maaten, L.: Revisiting weakly supervised pre-training of visual perception models. In: CVPR (2022)", + "45. Steiner, A., Kolesnikov, A., Zhai, X., Wightman, R., Uszkoreit, J., Beyer, L.: How to train your vit? data, augmentation, and regularization in vision transformers. TMLR (2022)", + "46. Tan, M., Le, Q.: Efficientnet: Rethinking model scaling for convolutional neural networks. In: ICML (2019)", + "47. Tong, Z., Song, Y., Wang, J., Wang, L.: Videomae: Masked autoencoders are data-efficient learners for self-supervised video pre-training. arXiv:2203.12602 [cs.CV] (2022)", + "48. Touvron, H., Cord, M., Douze, M., Massa, F., Sablayrolles, A., Jégou, H.: Training data-efficient image transformers & distillation through attention. In: ICML (2021)", + "49. Vasu, P.K.A., Gabriel, J., Zhu, J., Tuzel, O., Ranjan, A.: Mobileone: An improved one millisecond mobile backbone. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 7907-7917 (2023)", + "50. Wang, H., Wang, Z., Du, M., Yang, F., Zhang, Z., Ding, S., Mardziel, P., Hu, X.: Score-cam: Score-weighted visual explanations for convolutional neural networks." + ], + "bbox": [ + 215, + 146, + 787, + 839 + ], + "page_idx": 16 + }, + { + "type": "header", + "text": "PaPr for Faster Inference", + "bbox": [ + 560, + 114, + 730, + 126 + ], + "page_idx": 16 + }, + { + "type": "page_number", + "text": "17", + "bbox": [ + 767, + 114, + 785, + 126 + ], + "page_idx": 16 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "In: Proceedings of the IEEE/CVF conference on computer vision and pattern recognition workshops. pp. 24-25 (2020)", + "51. Wang, S., Li, B.Z., Khabsa, M., Fang, H., Ma, H.: Linformer: Self-attention with linear complexity. arXiv preprint arXiv:2006.04768 (2020)", + "52. Wei, C., Fan, H., Xie, S., Wu, C.Y., Yuille, A., Feichtenhofer, C.: Masked feature prediction for self-supervised visual pre-training. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 14668-14678 (2022)", + "53. Woo, S., Debnath, S., Hu, R., Chen, X., Liu, Z., Kweon, I.S., Xie, S.: Convnext v2: Co-designing and scaling convnets with masked autoencoders. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 16133-16142 (2023)", + "54. Wu, H., Xiao, B., Codella, N., Liu, M., Dai, X., Yuan, L., Zhang, L.: Cvt: Introducing convolutions to vision transformers. In: Proceedings of the IEEE/CVF international conference on computer vision. pp. 22-31 (2021)", + "55. Xu, X., Wang, S., Chen, Y., Zheng, Y., Wei, Z., Liu, J.: Gtp-vit: Efficient vision transformers via graph-based token propagation. In: Proceedings of the IEEE/CVF Winter Conference on Applications of Computer Vision. pp. 86-95 (2024)", + "56. Yin, H., Vahdat, A., Alvarez, J.M., Mallya, A., Kautz, J., Molchanov, P.: A-vit: Adaptive tokens for efficient vision transformer. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 10809-10818 (2022)", + "57. You, H., Xiong, Y., Dai, X., Wu, B., Zhang, P., Fan, H., Vajda, P., Lin, Y.C.: Castling-vit: Compressing self-attention via switching towards linear-angular attention at vision transformer inference. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 14431-14442 (2023)", + "58. Yu, H., Wu, J.: A unified pruning framework for vision transformers. arXiv:2111.15127 [cs.CV] (2021)", + "59. Zhang, X., Zhou, X., Lin, M., Sun, J.: Shufflenet: An extremely efficient convolutional neural network for mobile devices. In: Proceedings of the IEEE conference on computer vision and pattern recognition. pp. 6848-6856 (2018)", + "60. Zhou, A., Li, Y., Qin, Z., Liu, J., Pan, J., Zhang, R., Zhao, R., Gao, P., Li, H.: Sparsemae: Sparse training meets masked autoencoders. In: Proceedings of the IEEE/CVF International Conference on Computer Vision. pp. 16176-16186 (2023)", + "61. Zhou, B., Khosla, A., Lapedriza, A., Oliva, A., Torralba, A.: Learning deep features for discriminative localization. In: Proceedings of the IEEE conference on computer vision and pattern recognition. pp. 2921-2929 (2016)" + ], + "bbox": [ + 215, + 146, + 785, + 632 + ], + "page_idx": 17 + }, + { + "type": "page_number", + "text": "18", + "bbox": [ + 217, + 114, + 235, + 126 + ], + "page_idx": 17 + }, + { + "type": "header", + "text": "T. Mahmud et al.", + "bbox": [ + 271, + 114, + 390, + 127 + ], + "page_idx": 17 + } +] \ No newline at end of file diff --git a/2024/PaPr_ Training-Free One-Step Patch Pruning with Lightweight ConvNets for Faster Inference/f59c8716-3e2f-464a-8e1c-4f54e686a19f_model.json b/2024/PaPr_ Training-Free One-Step Patch Pruning with Lightweight ConvNets for Faster Inference/f59c8716-3e2f-464a-8e1c-4f54e686a19f_model.json new file mode 100644 index 0000000000000000000000000000000000000000..bff8e04d3ffd30642a25fe32753b6b6e8e335311 --- /dev/null +++ b/2024/PaPr_ Training-Free One-Step Patch Pruning with Lightweight ConvNets for Faster Inference/f59c8716-3e2f-464a-8e1c-4f54e686a19f_model.json @@ -0,0 +1,2513 @@ +[ + [ + { + "type": "title", + "bbox": [ + 0.219, + 0.141, + 0.785, + 0.186 + ], + "angle": 0, + "content": "PaPr: Training-Free One-Step Patch Pruning with Lightweight ConvNets for Faster Inference" + }, + { + "type": "text", + "bbox": [ + 0.223, + 0.212, + 0.778, + 0.227 + ], + "angle": 0, + "content": "Tanvir Mahmud1 Burhaneddin Yaman2 Chun-Hao Liu3 Diana Marculescu1" + }, + { + "type": "text", + "bbox": [ + 0.24, + 0.239, + 0.763, + 0.268 + ], + "angle": 0, + "content": "1University of Texas at Austin 2Bosch Research 3Amazon Prime Video 1{tanvirmahmud, dianam}@utexas.edu 2yaman013@umn.edu 3liuch37@gmail.com" + }, + { + "type": "text", + "bbox": [ + 0.263, + 0.301, + 0.74, + 0.619 + ], + "angle": 0, + "content": "Abstract. As deep neural networks evolve from convolutional neural networks (ConvNets) to advanced vision transformers (ViTs), there is an increased need to eliminate redundant data for faster processing without compromising accuracy. Previous methods are often architecture-specific or necessitate re-training, restricting their applicability with frequent model updates. To solve this, we first introduce a novel property of lightweight ConvNets: their ability to identify key discriminative patch regions in images, irrespective of model's final accuracy or size. We demonstrate that fully-connected layers are the primary bottleneck for ConvNets performance, and their suppression with simple weight recalibration markedly enhances discriminative patch localization performance. Using this insight, we introduce PaPr, a method for substantially pruning redundant patches with minimal accuracy loss using lightweight ConvNets across a variety of deep learning architectures, including ViTs, ConvNets, and hybrid transformers, without any re-training. Moreover, the simple early-stage one-step patch pruning with PaPr enhances existing patch reduction methods. Through extensive testing on diverse architectures, PaPr achieves significantly higher accuracy over state-of-the-art patch reduction methods with similar FLOP count reduction. More specifically, PaPr reduces about \\(70\\%\\) of redundant patches in videos with less than \\(0.8\\%\\) drop in accuracy, and up to \\(3.7\\times\\) FLOPs reduction, which is a \\(15\\%\\) more reduction with \\(2.5\\%\\) higher accuracy. Code is available at https://github.com/tanvir-utexas/PaPr." + }, + { + "type": "title", + "bbox": [ + 0.218, + 0.644, + 0.375, + 0.66 + ], + "angle": 0, + "content": "1 Introduction" + }, + { + "type": "text", + "bbox": [ + 0.218, + 0.674, + 0.788, + 0.841 + ], + "angle": 0, + "content": "Deep neural networks have grown from simple convolutional neural networks (ConvNets) to complex transformer models, aiming for better accuracy with more computation [10, 18]. Vision transformers (ViTs) excel by focusing on important parts of images with long-range attention and large-scale pre-training [17,39]. This success, however, comes at the cost of increased computational cost. Operating these large networks efficiently for downstream applications is crucial. Most visual tasks demand precisely pinpointing key image regions against complex backgrounds. Higher image resolution improves accuracy by capturing more details but also adds unnecessary background processing, burdening large models [44]. This highlights the need to cut down on redundant data in high-resolution images to maintain both speed and performance with advanced techniques [2,34]." + } + ], + [ + { + "type": "page_number", + "bbox": [ + 0.218, + 0.116, + 0.23, + 0.127 + ], + "angle": 0, + "content": "2" + }, + { + "type": "header", + "bbox": [ + 0.272, + 0.115, + 0.393, + 0.128 + ], + "angle": 0, + "content": "T. Mahmud et al." + }, + { + "type": "image", + "bbox": [ + 0.262, + 0.145, + 0.744, + 0.292 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.439, + 0.295, + 0.544, + 0.308 + ], + "angle": 0, + "content": "(b) Proposed PaPr" + }, + { + "type": "image_caption", + "bbox": [ + 0.214, + 0.312, + 0.789, + 0.368 + ], + "angle": 0, + "content": "Fig. 1: (a) Existing patch pruning methods gradually reduce patches over the model. This requires additional training of mask generators in intermediate layers. (b) Proposed PaPr directly prunes redundant patches early in the network by leveraging pretrained lightweight ConvNets and directly speeds-up off-the-shelf models without re-training." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.4, + 0.789, + 0.642 + ], + "angle": 0, + "content": "Identifying key image regions demands a comprehensive understanding of the image and model operations. Incorrect key region estimation can impair pretraining performance by eliminating crucial areas. Traditional approaches [2,25,28, 34,56] for pruning redundant patch regions are hindered by three main limitations: (1) They often necessitate complex training of extra modules, which becomes increasingly difficult as baseline models evolve with more data, enhanced training methods, and deeper structures [25,41,56]. Re-training these modules for each model update is impractical. (2) Without a complete image understanding, these methods incrementally prune patches across the network, leading to unnecessary computations in early layers, which is particularly problematic for deeper models. (3) Many rely on specific architectural features for patch reduction, such as class tokens or attention maps, limiting their use to a narrow set of network designs [12,28]. Hence, there is a pressing need for a patch pruning solution that is adaptable to various modern architectures without additional training, can eliminate redundant patches in a single-step, thereby making it suitable for a broad spectrum of networks while streamlining each model update (See Fig. 1)." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.643, + 0.79, + 0.809 + ], + "angle": 0, + "content": "Recent work on patch reduction [2, 23, 55] focuses mostly on transformers rather than ConvNets due to their impressive performance on various tasks, their aforementioned limitations notwithstanding. While ConvNets achieve lower ImageNet-1k top-1 accuracy than large ViTs (68.7% in MobileOne-S0 [49] vs. 88.7% in ViT-Huge [44]), they exhibit a remarkable ability to efficiently process the key image regions with hierarchical inductive bias. Our empirical investigation reveals that, as we broaden the evaluation metric (increasing \\( k \\) in top- \\( k \\) evaluations, see Fig. 4), the benefits of deeper models diminish, especially with a large number of image classes (e.g., 1000 in ImageNet). This suggests that shallower models excel at identifying discriminative areas as their bigger counterparts, rendering them ideal for patch pruning." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.81, + 0.789, + 0.842 + ], + "angle": 0, + "content": "Leveraging this insight, we propose PaPr, a novel Patch Pruning method that employs pretrained lightweight ConvNets for efficient, one-step patch pruning in" + } + ], + [ + { + "type": "header", + "bbox": [ + 0.561, + 0.115, + 0.732, + 0.128 + ], + "angle": 0, + "content": "PaPr for Faster Inference" + }, + { + "type": "page_number", + "bbox": [ + 0.775, + 0.117, + 0.787, + 0.127 + ], + "angle": 0, + "content": "3" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.147, + 0.788, + 0.268 + ], + "angle": 0, + "content": "a wide variety of deep learning models, maintaining accuracy while significantly cutting computational demands. Our findings show lightweight ConvNets have remarkable ability in identifying discriminative image regions but struggle in fine-grained prediction. To address this, PaPr relies on Patch Significance Maps (PSMs), which are generated using only the convolutional layers of ConvNets through uniform class weight recalibration in FC layers. Astonishingly, PSMs consistently highlight critical image regions across ConvNets of varying sizes and accuracies (Fig. 6), thereby amplifying the efficacy of ultra lightweight ConvNets." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.27, + 0.789, + 0.481 + ], + "angle": 0, + "content": "ConvNets inherently preserve the positional property of patches for their inductive bias, while ViTs use cross-attention to blend patch features, leading to variability in patch relevance across models. Unlike previous methods that prune redundant patches gradually over multiple steps across intermediate layers of ViTs [2, 23, 41], our approach, PaPr, simplifies this process by eliminating non-essential patches at once, immediately after extraction (see Fig. 3), by leveraging lightweight ConvNets to assess patch significance. This direct, one-step pruning approach significantly cuts computational demands and is also compatible with other patch reduction techniques (see Fig. 5). PaPr's ability to separate crucial patch identification from fine-grained class prediction enhances a wide range of pre-trained models without further training, ensuring high accuracy with notable speed ups. By bypassing the complex training required for conventional patch selectors, PaPr capitalizes on the comprehensive capabilities of minimalist ConvNets for efficient patch pruning in larger models." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.484, + 0.789, + 0.695 + ], + "angle": 0, + "content": "Our experiments demonstrate PaPr's effectiveness across various architectures and pre-training methods, achieving significant reduction in redundant patches for ViTs, large-scale ConvNets (e.g., ConvNext [32]), and hybrid transformers (e.g., Swin [31]), outperforming state-of-the-art (SOTA) patch reduction methods by a large margin. Notably, PaPr can be easily integrated with most existing patch reduction methods to reduce patches early in operation. More specifically, PaPr boosts ToMe [2] accuracy by \\(4.5\\%\\) for a computational budget similar to ViT-B. Moreover, PaPr can accelerate training akin to token merging techniques, unlike most patch pruning methods that fail to boost training speed [34, 41, 56]. PaPr shows robust patch localization performance with ultra lightweight ConvNets (\\(<0.3\\%\\) accuracy loss) for \\(42\\times\\) reduction in proposal FLOPs, thereby enabling its use for larger off-the-shelf models. Remarkably, in video recognition, PaPr cuts down around \\(70\\%\\) of redundant patches, resulting in up to \\(3.7\\times\\) FLOPs reduction with minimal impact on accuracy (\\(\\approx 0.8\\%\\))." + }, + { + "type": "text", + "bbox": [ + 0.241, + 0.697, + 0.595, + 0.71 + ], + "angle": 0, + "content": "We summarize our main contributions as follows:" + }, + { + "type": "text", + "bbox": [ + 0.226, + 0.731, + 0.784, + 0.775 + ], + "angle": 0, + "content": "- We propose PaPr, a novel background patch pruning method that can seamlessly operate with ViTs, ConvNets, and hybrid transformers, without further training while leveraging batch processing." + }, + { + "type": "text", + "bbox": [ + 0.225, + 0.778, + 0.784, + 0.806 + ], + "angle": 0, + "content": "- We propose a simple weight recalibration method in ConvNets to precisely and efficiently locate discriminative patches, irrespective of model size." + }, + { + "type": "text", + "bbox": [ + 0.225, + 0.81, + 0.785, + 0.839 + ], + "angle": 0, + "content": "- We facilitate the use of ultra-lightweight ConvNets to speed-up large models, such as ViTs, with a seamless framework and negligible accuracy loss." + }, + { + "type": "list", + "bbox": [ + 0.225, + 0.731, + 0.785, + 0.839 + ], + "angle": 0, + "content": null + } + ], + [ + { + "type": "page_number", + "bbox": [ + 0.218, + 0.116, + 0.23, + 0.127 + ], + "angle": 0, + "content": "4" + }, + { + "type": "header", + "bbox": [ + 0.272, + 0.115, + 0.393, + 0.128 + ], + "angle": 0, + "content": "T. Mahmud et al." + }, + { + "type": "text", + "bbox": [ + 0.225, + 0.147, + 0.788, + 0.178 + ], + "angle": 0, + "content": "- We present extensive qualitative and quantitative results across numerous model architectures in both image and video applications." + }, + { + "type": "title", + "bbox": [ + 0.216, + 0.197, + 0.389, + 0.214 + ], + "angle": 0, + "content": "2 Related Work" + }, + { + "type": "title", + "bbox": [ + 0.216, + 0.226, + 0.589, + 0.241 + ], + "angle": 0, + "content": "2.1 From ConvNets to Vision Transformers" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.249, + 0.789, + 0.459 + ], + "angle": 0, + "content": "ConvNets have been pivotal in computer vision, offering computational efficiency through kernel reuse and localization [8, 18-20, 29, 42, 46, 49, 59]. Yet, they struggle in capturing long-range dependencies, a gap bridged by ViT [10]. Inspired by the success of transformers from natural language processing (NLP), ViTs excel in long-range feature modeling through cross-attention, outperforming ConvNets albeit with higher computational demands and extensive training data requirements [7, 48, 51, 57]. Efforts to alleviate these issues include self-supervised [17, 47, 52, 60] and weakly-supervised [39, 44] pre-training, although challenges in computational complexity and optimization remain. Hybrid architectures, merging ConvNets' inductive biases with ViTs' cross-attention, offer a balanced solution by reducing computational load while maintaining performance [6, 9, 11, 15, 27, 30, 31, 33, 54]. Recent ConvNets advancements [32, 53], with improved training and large-scale data, now rival ViTs, questioning whether architectural innovations or enhanced training primarily drive performance gains." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.46, + 0.79, + 0.506 + ], + "angle": 0, + "content": "Addressing the computational demands of these models, we focus on enhancing operational efficiency by pruning redundant patches without architectural modifications or re-training, maintaining performance while streamlining updates." + }, + { + "type": "title", + "bbox": [ + 0.216, + 0.524, + 0.734, + 0.54 + ], + "angle": 0, + "content": "2.2 Class Activation Mapping for Explainable Deep Learning" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.547, + 0.789, + 0.728 + ], + "angle": 0, + "content": "Class Activation Mapping (CAM) techniques provide explainable visual reasoning for neural network predictions by highlighting activation regions1 crucial for decisions [1,21,38,43,61]. While ConvNets utilize convolutional operations to preserve spatial information, making them apt for such visualizations, ViTs link attention weights to class tokens for prediction [4,36]. However, CAM's reliance on accurate class predictions for effective feature localization is a significant drawback, as ConvNets' lower accuracy when compared to ViTs may lead to incorrect object localization. Despite ViTs' superior fine-grained classification capabilities, it is unclear if they prioritize the same discriminative regions as ConvNets. Additionally, CAM approaches often require gradient tracking [5,43] or complex feature map decomposition [35], complicating batch processing and necessitating extra optimization steps." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.728, + 0.789, + 0.819 + ], + "angle": 0, + "content": "Our work diverges from traditional CAM by aiming to consistently identify discriminative patches across different architectures, focusing computational resources on the most relevant areas without being constrained by class prediction accuracy. This approach enables more efficient processing by allowing larger models to concentrate on the most significant regions, identified by lighter networks, thus significantly reducing computational demands." + }, + { + "type": "page_footnote", + "bbox": [ + 0.218, + 0.825, + 0.731, + 0.84 + ], + "angle": 0, + "content": "1 \"Patch\" and \"region\" have been used interchangeably based on the context." + } + ], + [ + { + "type": "header", + "bbox": [ + 0.561, + 0.115, + 0.732, + 0.127 + ], + "angle": 0, + "content": "PaPr for Faster Inference" + }, + { + "type": "page_number", + "bbox": [ + 0.775, + 0.117, + 0.785, + 0.127 + ], + "angle": 0, + "content": "5" + }, + { + "type": "image", + "bbox": [ + 0.262, + 0.144, + 0.745, + 0.223 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.371, + 0.224, + 0.65, + 0.237 + ], + "angle": 0, + "content": "(a) Baseline Convolutional Model Feature Flow" + }, + { + "type": "image", + "bbox": [ + 0.31, + 0.242, + 0.703, + 0.317 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.391, + 0.325, + 0.621, + 0.337 + ], + "angle": 0, + "content": "(b) Proposed Pruning Mask Extraction" + }, + { + "type": "image_caption", + "bbox": [ + 0.215, + 0.341, + 0.789, + 0.426 + ], + "angle": 0, + "content": "Fig.2: (a) Baseline ConvNet gradually reduces the feature map to produce \\(\\mathcal{F} = \\{f_k(x,y)\\}_{k=1}^K\\), followed by global average pooling and fully connected (FC) layers to predict \\(y_p\\). (b) In PaPr, we operate on \\(\\mathcal{F}\\) by suppressing the FC layer. Initially, we extract pixel mean over \\(K\\) channels to produce discriminative region proposal \\(\\mathcal{R}\\). Later, simple upsampling operation generates the patch significance map (PSM) \\(\\mathcal{P}\\) of target dimension. Finally, patch mask \\(\\mathcal{M}\\) for top \\(z\\%\\) patches is obtained from \\(\\mathcal{P}\\)." + }, + { + "type": "title", + "bbox": [ + 0.216, + 0.453, + 0.569, + 0.466 + ], + "angle": 0, + "content": "2.3 Patch Reduction for Faster Inference" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.474, + 0.789, + 0.67 + ], + "angle": 0, + "content": "Several approaches have sought to enhance computational efficiency by reducing redundant patches in neural networks, with early strategies involving additional adapters or controllers to identify and prune less significant patches [24,26,41,58]. These methods, however, necessitate separate adapter training for each network and are slow to adapt due to the need to learn from the dynamics of other layers. In the context of ViTs, efforts have leveraged architectural features such as class tokens and attention maps for patch relevance, yet these solutions often fail to generalize across various architectures like hybrid transformers [31] or larger ConvNets [32, 53], thus limiting their applicability [12, 25, 34, 56]. Furthermore, while some have explored patch merging in ViTs and a combination of merging and pruning [2, 3, 23], these approaches lack a holistic image understanding by not considering the entire image context in their optimizations, leading to incremental and sub-optimal patch reduction." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.671, + 0.79, + 0.716 + ], + "angle": 0, + "content": "Our work diverges from these traditional patch reduction methods by proposing a single-step, early-network patch removal strategy, that seamlessly integrates with any architecture without re-training." + }, + { + "type": "title", + "bbox": [ + 0.216, + 0.736, + 0.381, + 0.755 + ], + "angle": 0, + "content": "3 Methodology" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.765, + 0.789, + 0.842 + ], + "angle": 0, + "content": "Our methodology introduces a novel approach to discriminative patch pruning across various deep learning architectures, leveraging the innate capabilities of lightweight ConvNets. By generating a PSM, we efficiently identify and prune non-essential patches in a single step, enhancing computational efficiency without compromising accuracy. This process is universally applicable, seamlessly" + } + ], + [ + { + "type": "page_number", + "bbox": [ + 0.218, + 0.116, + 0.23, + 0.127 + ], + "angle": 0, + "content": "6" + }, + { + "type": "header", + "bbox": [ + 0.272, + 0.115, + 0.393, + 0.128 + ], + "angle": 0, + "content": "T. Mahmud et al." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.147, + 0.784, + 0.178 + ], + "angle": 0, + "content": "integrating with ViTs, ConvNets, and hybrid models, thereby addressing the limitations of previous methods with a scalable, architecture-agnostic solution." + }, + { + "type": "title", + "bbox": [ + 0.215, + 0.199, + 0.676, + 0.215 + ], + "angle": 0, + "content": "3.1 Extracting Discriminative Regions with ConvNets" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.223, + 0.788, + 0.358 + ], + "angle": 0, + "content": "Despite achieving relatively lower top-1 accuracy on the ImageNet benchmark, lightweight ConvNets exhibit a competitive edge in top-10 accuracy when compared to their larger ViT counterparts (Fig. 4). This phenomenon underscores ConvNets' capability to effectively localize regions of interest through their convolutional layers, despite potential limitations in fine-grained classification attributed to the fully-connected (FC) layer. Our methodology leverages this insight by proposing discriminative patch regions while minimizing the influence of the FC layer (see Fig. 2), thereby enhancing the large off-the-shelf model's focus on most salient image regions proposed by lightweight ConvNets (see Fig. 3)." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.359, + 0.788, + 0.507 + ], + "angle": 0, + "content": "Given an input image \\( X \\in \\mathbb{R}^{H \\times W \\times 3} \\), with height \\( H \\) and width \\( W \\), we denote by \\( f_{k}(x,y) \\) the feature map generated by the \\( k^{th} \\) kernel in the last convolutional layer of typical ConvNets, \\( \\forall k \\in \\{1,2,\\dots,K\\} \\). Here, each pixel \\( (x,y) \\) in \\( f_{k}(x,y) \\) corresponds to a patch window of size \\( (H / d \\times W / d) \\) in the input image \\( X \\), reflecting a spatial down-scaling factor of \\( d \\) through the convolutional layer stack. Typically, global average pooling (GAP) is applied to each feature map \\( f_{k}(x,y) \\), yielding \\( F = \\{F_{k}\\}_{k=1}^{K} \\in \\mathbb{R}^{K} \\), where \\( F_{k} = \\sum_{x,y} f_{k}(x,y) \\in \\mathbb{R} \\). Finally, a fully connected (FC) layer \\( W \\in \\mathbb{R}^{C \\times K} \\) with elements \\( w_{c}^{k} \\) processes \\( F \\) to generate class predictions \\( y_{p} \\in \\mathbb{R}^{C} \\) across \\( C \\) classes as follows:" + }, + { + "type": "equation", + "bbox": [ + 0.327, + 0.517, + 0.788, + 0.548 + ], + "angle": 0, + "content": "\\[\ny _ {p} = W F = \\sum_ {c} \\sum_ {k} w _ {c} ^ {k} \\sum_ {x, y} f _ {k} (x, y) = \\sum_ {x, y} R (x, y), \\tag {1}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.56, + 0.788, + 0.59 + ], + "angle": 0, + "content": "where \\( R \\in \\mathbb{R}^{h \\times w} \\) encapsulates the weighted mean class activation mapping of the convolutional feature map." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.59, + 0.788, + 0.711 + ], + "angle": 0, + "content": "We note that the reliance on class activation weights \\( w_{c}^{k} \\) is influenced by the model's final accuracy, thereby posing a challenge for smaller models. Our objective transcends mere accuracy enhancement, aiming to precisely locate discriminative patches rich in information irrespective of the model size or its final classification performance. Recognizing that discriminative region localization is pivotal for detailed classification, and acknowledging the competitive top-10 accuracy of lighter ConvNets, we posit that an extremely lightweight ConvNet suffices for initial discriminative region proposal." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.711, + 0.788, + 0.757 + ], + "angle": 0, + "content": "To counteract the influence of the weak linear FC layer \\(W\\) in convolutional region proposal, we propose an adjustment where \\(w_{c}^{k} = 1 / KC\\), \\(\\forall k\\in \\{1,\\dots ,K\\}\\), facilitating the generation of a discriminative region proposal \\(\\mathcal{R}\\in \\mathbb{R}^{h\\times w}\\), where" + }, + { + "type": "equation", + "bbox": [ + 0.332, + 0.767, + 0.788, + 0.801 + ], + "angle": 0, + "content": "\\[\n\\mathcal {R} (x, y) = \\sum_ {c} \\sum_ {k} \\frac {1}{K C} f _ {k} (x, y) = \\frac {1}{K} \\sum_ {k} f _ {k} (x, y). \\tag {2}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.215, + 0.81, + 0.788, + 0.841 + ], + "angle": 0, + "content": "This strategy allows us to leverage ConvNets for what they excel at: pinpointing critical image areas. By reducing reliance on class activation weights, we" + } + ], + [ + { + "type": "header", + "bbox": [ + 0.561, + 0.115, + 0.732, + 0.128 + ], + "angle": 0, + "content": "PaPr for Faster Inference" + }, + { + "type": "page_number", + "bbox": [ + 0.775, + 0.116, + 0.787, + 0.127 + ], + "angle": 0, + "content": "7" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.147, + 0.788, + 0.209 + ], + "angle": 0, + "content": "efficiently generate discriminative region proposals directly from convolutional outputs. This approach not only enhances the interpretability and efficiency of the localization process but also enables the application of more complex models for subsequent detailed analysis, optimizing the use of computational resources." + }, + { + "type": "title", + "bbox": [ + 0.215, + 0.23, + 0.458, + 0.246 + ], + "angle": 0, + "content": "3.2 Patch Significance Map" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.255, + 0.788, + 0.315 + ], + "angle": 0, + "content": "With the discriminative region proposal \\(\\mathcal{R} \\in \\mathbb{R}^{h \\times w}\\), where each pixel \\((x, y)\\) quantifies the significance of corresponding patches in the original image, we proceed to establish a precise mapping to our intended feature map (Fig. 2). This mapping ensures the preservation of spatial relationships within the feature map." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.315, + 0.789, + 0.422 + ], + "angle": 0, + "content": "Consider a target feature map \\(\\mathcal{F} \\in \\mathbb{R}^{h' \\times w' \\times K}\\), with each pixel \\((x', y')\\) encapsulating a feature vector from a specific image patch. To align \\(\\mathcal{R}\\) with \\(\\mathcal{F}\\), we employ an upsampling operation \\(U: \\mathbb{R}^{h \\times w} \\to \\mathbb{R}^{h' \\times w'}\\), transforming \\(\\mathcal{R}\\) into the Patch Significance Map (PSM) \\(\\mathcal{P} \\in \\mathbb{R}^{h' \\times w'}\\). Consequently, each element of \\(\\mathcal{P}\\) directly corresponds to the patch significance within \\(\\mathcal{F}\\) for the given discriminative task. The next step involves utilizing \\(\\mathcal{P}\\) to discern and prune non-essential patch features from \\(\\mathcal{F}\\). By sorting the values within \\(\\mathcal{P}\\), we acquire a pruning mask \\(\\mathcal{M}\\):" + }, + { + "type": "equation", + "bbox": [ + 0.331, + 0.431, + 0.788, + 0.451 + ], + "angle": 0, + "content": "\\[\n\\mathcal {M} = \\operatorname {r e s h a p e} (\\operatorname {a r g s o r t} (\\operatorname {f l a t t e n} (\\mathcal {P}))), \\mathcal {M} \\in \\mathbb {R} ^ {h ^ {\\prime} \\times w ^ {\\prime}}, \\tag {3}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.461, + 0.788, + 0.52 + ], + "angle": 0, + "content": "enabling batch-wise patch pruning. Specifically, we identify the indices corresponding to the top-\\(z\\%\\) patches as per \\(\\mathcal{M}\\), facilitating the retention of only the most salient patches in \\(\\mathcal{F}\\), thereby enhancing computational efficiency in subsequent processing stages." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.521, + 0.788, + 0.582 + ], + "angle": 0, + "content": "As part of our discussion on various architectures, we detail the application of this patch reduction technique. For enhanced visualization of the PSM, we apply min-max normalization to \\(\\mathcal{P}\\), adjusting for outliers and scaling the significance scores to visually depict the importance of different patches." + }, + { + "type": "title", + "bbox": [ + 0.215, + 0.604, + 0.618, + 0.62 + ], + "angle": 0, + "content": "3.3 Integrating PSM with Vision Transformers" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.629, + 0.788, + 0.72 + ], + "angle": 0, + "content": "The ViT processes an input image \\(X \\in \\mathbb{R}^{H \\times W \\times 3}\\) by extracting \\(N\\)-dimensional patch token features \\(X_{p} \\in \\mathbb{R}^{N \\times d}\\) from a \\((k \\times k)\\) patch window using strided convolutions, where \\(N = HW / k^2\\) (Fig. 3). Positional embeddings are then added to these patch tokens \\(X_{p}\\) to retain spatial information, followed by the application of successive cross-attention mechanisms. For classification tasks, the model either introduces a class token \\(t_{cls} \\in \\mathbb{R}^{d}\\) or utilizes the mean of the output patch tokens." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.72, + 0.789, + 0.841 + ], + "angle": 0, + "content": "The encoder employs multi-headed cross-attention (MHA) on the patch token embeddings \\( X_{p} \\), with each MHA block consisting of multiple linear layers and a cross-attention layer. The computational complexity of these operations suggests that reducing the number of tokens \\( N \\) can significantly decrease computational demands. A key strength of the ViT architecture is its adaptability to varying numbers of tokens \\( N \\), though at a higher computational cost. Prior work [2,41,56] has primarily aimed at reducing patch tokens within intermediate encoder layers, often requiring additional training due to the variability in token representations" + } + ], + [ + { + "type": "page_number", + "bbox": [ + 0.218, + 0.116, + 0.23, + 0.127 + ], + "angle": 0, + "content": "8" + }, + { + "type": "header", + "bbox": [ + 0.272, + 0.115, + 0.393, + 0.128 + ], + "angle": 0, + "content": "T. Mahmud et al." + }, + { + "type": "image", + "bbox": [ + 0.26, + 0.145, + 0.744, + 0.297 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.214, + 0.301, + 0.789, + 0.4 + ], + "angle": 0, + "content": "Fig. 3: (a) In vanilla ViT, PaPr operates right after the patch extractor module. Hence, all transformer blocks can operate only with the most discriminative patches. (b) Hierarchical model blocks comprise of window based kernel operator (e.g., Convk \\(\\times\\) \\(k\\)/local attention), followed by pixel operator (e.g., linear layer, Conv1x1). Pixel operator consumes more than \\(60\\%\\) of total computation. PaPr is used to split the foreground patches to be operated with pixel operator. Background patches are zero-ed out, and finally, re-assembled with foreground output patches." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.429, + 0.787, + 0.459 + ], + "angle": 0, + "content": "across different architectures. These methods typically achieve only gradual token reduction, limited by the distributed nature of information across tokens." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.46, + 0.789, + 0.596 + ], + "angle": 0, + "content": "In contrast, our approach seeks to eliminate redundant tokens early in the processing pipeline, immediately following initial patch extraction. This strategy offers multiple advantages: (1) Leveraging the convolutional nature of the initial patch extraction allows for a direct mapping between our PSM \\(\\mathcal{P}\\) and the patch tokens \\(X_{p}\\), obviating the need for additional mechanisms to determine token redundancy. (2) It facilitates generalizability across transformer models without the need for retraining, potentially accelerating training by concentrating on essential patches. (3) It permits seamless integration of existing intermediate layer token reduction techniques following our initial patch pruning process." + }, + { + "type": "title", + "bbox": [ + 0.215, + 0.618, + 0.616, + 0.634 + ], + "angle": 0, + "content": "3.4 Integrating PSM with Hierarchical Models" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.644, + 0.788, + 0.75 + ], + "angle": 0, + "content": "ConvNets and hybrid transformers, such as Swin [31], primarily operate with window-based local operations in contrast to full-attention based operation in vanilla transformers. Such window-based local processing makes patch pruning particularly complicated, in contrast to vanilla transformers. However, these models maintain the location property of representative image patches all through the network, which leaves the door open to prune redundant patches. Nevertheless, the window-based operation is particularly difficult to prune." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.75, + 0.789, + 0.842 + ], + "angle": 0, + "content": "In general, each block of hierarchical layers consists of window-based spatial kernel/attention operators, followed by pixel operators (usually, modeled with \\(1 \\times 1\\) convolutional layers or linear layers). In contrast to windowed convolutions or cross-attention, these pixel operators are particularly suitable for patch pruning. Interestingly, in the SOTA ConvNets and hybrid transformers, more than \\(60\\%\\) of total computations are performed with such pixel operators (63.3% in Swin [31]," + } + ], + [ + { + "type": "header", + "bbox": [ + 0.561, + 0.115, + 0.732, + 0.127 + ], + "angle": 0, + "content": "PaPr for Faster Inference" + }, + { + "type": "page_number", + "bbox": [ + 0.775, + 0.116, + 0.785, + 0.127 + ], + "angle": 0, + "content": "9" + }, + { + "type": "image", + "bbox": [ + 0.225, + 0.146, + 0.497, + 0.27 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.22, + 0.276, + 0.499, + 0.373 + ], + "angle": 0, + "content": "Fig. 4: ImageNet-1k evaluation for varying top-k accuracy targets. The accuracy gain with bigger model largely shrinks, as \\( k \\) increases. This suggests shallower ConvNets have understanding of object locations and visual property, despite their lower top-1 accuracy." + }, + { + "type": "image", + "bbox": [ + 0.507, + 0.145, + 0.78, + 0.27 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.503, + 0.276, + 0.784, + 0.373 + ], + "angle": 0, + "content": "Fig.5: Integrating PaPr with ToMe [2]. We use the Augreg pretrained ViT-B-16 architecture as the baseline. We sweep token merging ratio (r) for different pruning ratio (z). Integration of PaPr achieves Pareto-optimal performance, thus, PaPr can enhance existing patch reduction methods." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.403, + 0.788, + 0.493 + ], + "angle": 0, + "content": "\\(96.2\\%\\) in ConvNext [32]). Based on the patch significance map from PaPr, we only use the pixel operator on most significant patches, and simply perform zero-padding on the remaining patch regions. The zero padding operations are performed to mostly recover the feature map shape to be used with subsequent spatial operators. Hence, the speed-up is mostly achieved by eliminating redundant computations in pixel operators (See Fig. 3)." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.495, + 0.788, + 0.601 + ], + "angle": 0, + "content": "Let's assume that the feature map before applying pixel operators is given by \\(\\mathcal{A}_w\\). Based on the PSM \\(\\mathcal{P}\\), we initially identify the foreground and background pixel features as \\(A_{f}\\), and \\(A_{b}\\), respectively. Later, the pixel operator layers are applied on foreground pixels \\(A_{f}\\), and a zero-padded representation is used for the background pixels \\(A_{b}\\). Finally, the output representation \\(\\mathcal{A}_c\\) is reassembled with modulated foreground pixels \\(A_{f}\\) and zero-padded background pixels \\(A_{z}\\), given by Eq. (4):" + }, + { + "type": "equation", + "bbox": [ + 0.461, + 0.601, + 0.625, + 0.617 + ], + "angle": 0, + "content": "\\[\nA _ {f}, A _ {b} = \\operatorname {S p l i t} \\left(\\mathcal {A} _ {w}, \\mathcal {P}\\right),\n\\]" + }, + { + "type": "equation", + "bbox": [ + 0.381, + 0.619, + 0.787, + 0.636 + ], + "angle": 0, + "content": "\\[\nA _ {z} = \\operatorname {Z e r o s} \\left(A _ {b}\\right), A _ {f} = \\operatorname {L i n e a r} \\left(A _ {f}\\right), \\tag {4}\n\\]" + }, + { + "type": "equation", + "bbox": [ + 0.436, + 0.639, + 0.625, + 0.654 + ], + "angle": 0, + "content": "\\[\n\\mathcal {A} _ {c} = \\operatorname {R e a s s e m b l e} \\left(A _ {f}, A _ {z}\\right).\n\\]" + }, + { + "type": "title", + "bbox": [ + 0.216, + 0.676, + 0.442, + 0.694 + ], + "angle": 0, + "content": "4 Image Experiments" + }, + { + "type": "title", + "bbox": [ + 0.216, + 0.709, + 0.428, + 0.725 + ], + "angle": 0, + "content": "4.1 Experimental Setup" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.734, + 0.788, + 0.841 + ], + "angle": 0, + "content": "We experiment on image classification task on the ImageNet-1k benchmark dataset, following prior work [2, 12, 23, 34, 55]. We use the MobileOne-s0 [49] model as the proposal model for all architectures, unless otherwise specified. We report training-free results of PaPr, unless otherwise specified. For the performance metric, we report top-1 accuracy, GFLOPs, and the throughput (img/s). Throughput is measured on a single RTX-A5000 GPU with 24GB VRAM. We reproduced all baseline models under the same setup for a fair comparison." + } + ], + [ + { + "type": "page_number", + "bbox": [ + 0.218, + 0.116, + 0.236, + 0.127 + ], + "angle": 0, + "content": "10" + }, + { + "type": "header", + "bbox": [ + 0.272, + 0.115, + 0.393, + 0.128 + ], + "angle": 0, + "content": "T. Mahmud et al." + }, + { + "type": "table_caption", + "bbox": [ + 0.218, + 0.147, + 0.508, + 0.203 + ], + "angle": 0, + "content": "Table 1: Performance comparison on AugReg models. PaPr achieves the best performance, while operating with token merging. \\( z \\) denotes patch keeping ratio in PaPr." + }, + { + "type": "table", + "bbox": [ + 0.224, + 0.215, + 0.507, + 0.451 + ], + "angle": 0, + "content": "
ModelsMethodsAcc1GFLOPsImg/s
ViT-S-16Baseline81.394.61975
ToMe [2]76.962.291978
TokenFusion [23]77.122.291982
GTP-ViT [55]71.032.311970
PaPr (z=0.45)76.212.291988
PaPr (z=0.55) (+ToMe)77.762.192073
PaPr (z=0.5) (+ToMe)76.271.972315
ViT-B-16Baseline84.5917.59307
ToMe [2]80.388.78615
TokenFusion [23]80.78.78618
GTP-ViT [55]80.988.78610
PaPr (z=0.5)82.118.98605
PaPr (z=0.55) (+ToMe)82.348.21660
PaPr (z=0.5) (+ToMe)80.886.82785
ViT-L-16Baseline85.8261.6191
ToMe [2]83.530.99180
TokenFusion [23]83.9130.99182
GTP-ViT [55]81.5630.99181
PaPr (z=0.5)83.8730.83183
PaPr (z=0.55) (+ToMe)83.9926.33210
PaPr (z=0.55) (+ToMe)83.525.12224
" + }, + { + "type": "table_caption", + "bbox": [ + 0.513, + 0.145, + 0.782, + 0.228 + ], + "angle": 0, + "content": "Table 2: Performance comparison on MAE models. Since MAE uses masked pretraining, PaPr is particularly suitable for MAE inference. PaPr achieves significantly higher performance than others. \\( z \\) denotes patch keeping ratio in PaPr." + }, + { + "type": "table", + "bbox": [ + 0.521, + 0.241, + 0.778, + 0.453 + ], + "angle": 0, + "content": "
ModelsMethodsAcc1GFLOPsImg/s
ViT-B-16Baseline83.7417.59307
ToMe [2]78.828.78615
TokenFusion [23]79.238.78618
GTP-ViT [55]79.148.78610
PaPr (z=0.5)82.48.98605
PaPr (z=0.4)81.47.72700
ViT-L-16Baseline85.9561.6191
ToMe [2]84.2430.99180
TokenFusion [23]84.3330.99182
GTP-ViT [55]84.1530.99181
PaPr (z=0.5)85.0630.83183
PaPr (z=0.4)84.7627.72201
ViT-H-16Baseline86.89167.436
ToMe [2]85.4882.5372
TokenFusion [23]85.7182.5373
GTP-ViT [55]85.5482.5371
PaPr (z=0.5)86.483.0471
PaPr (z=0.4)86.1374.5981
" + }, + { + "type": "title", + "bbox": [ + 0.216, + 0.48, + 0.634, + 0.494 + ], + "angle": 0, + "content": "4.2 Performance on Various Vision Transformers" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.503, + 0.787, + 0.549 + ], + "angle": 0, + "content": "We study performance of PaPr in diverse ViT architectures, along with various pre-training methods. We also present training results, along with training-free results to compare with existing methods." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.57, + 0.788, + 0.707 + ], + "angle": 0, + "content": "Training-free method comparison. We consider two different pre-training methods for comparative analysis on various ViT architectures. We use supervised pretrained Augreg models [45], and self-supervised pre-trained masked autoencoder (MAE) models [17], following prior work [2]. We compare with three recent training-free patch reduction methods, such as ToMe [2], TokenFusion [23], and GTP-ViT [55]. Additionally, we show the performance comparison on class-token free ViT models to highlight the architecture agnostic performance. In general, PaPr achieves better accuracy, with lower computational costs compared to state-of-the-art (SOTA) methods." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.719, + 0.788, + 0.841 + ], + "angle": 0, + "content": "Augreg models: For AugReg pre-training with ViT-Base, PaPr achieves \\(2.14\\%\\) higher accuracy over ToMe, with comparable FLOPs as shown in Tab. 1. In addition, by combining PaPr with ToMe, we achieve additional \\(22.3\\%\\) FLOP reduction, while maintaining higher accuracy. We further study the compatibility of PaPr with existing patch reduction methods as ToMe. After the initial patch pruning with PaPr, we integrate ToMe only at the bottom-half layers due to its higher sensitivity in earlier layers. As shown in Fig. 5, integration of PaPr can directly boost ToMe performance thereby achieving Pareto optimality." + } + ], + [ + { + "type": "header", + "bbox": [ + 0.561, + 0.115, + 0.732, + 0.128 + ], + "angle": 0, + "content": "PaPr for Faster Inference" + }, + { + "type": "page_number", + "bbox": [ + 0.769, + 0.116, + 0.784, + 0.127 + ], + "angle": 0, + "content": "11" + }, + { + "type": "table_caption", + "bbox": [ + 0.221, + 0.145, + 0.499, + 0.187 + ], + "angle": 0, + "content": "Table 3: Performance analysis on class-token free ViT models. PaPr performance gain is not limited to specific architectures." + }, + { + "type": "table", + "bbox": [ + 0.229, + 0.199, + 0.497, + 0.31 + ], + "angle": 0, + "content": "
ModelMethodAcc1GFLOPsImg/s
ViT-MediumBaseline84.3310.58484
GAP-16GTP-ViT [55]80.55.73880
256ToMe [2]80.215.73890
PaPr (z=0.5)81.85.5932
ViT-MediumBaseline85.626.06182
GAP-16GTP-ViT [55]81.9513.59340
384ToMe [2]82.513.59344
PaPr83.9512.94360
" + }, + { + "type": "table_caption", + "bbox": [ + 0.504, + 0.145, + 0.782, + 0.188 + ], + "angle": 0, + "content": "Table 4: Training performance analysis on DeIT-s. PaPr achieves competitive performance with higher training speed." + }, + { + "type": "table", + "bbox": [ + 0.513, + 0.2, + 0.778, + 0.31 + ], + "angle": 0, + "content": "
MethodsAcc1GFLOPsIm/sBatch Proc.Train Speed
Baseline79.84.6960✓1x
DynamicViT [41]79.32.91510X1x
A-ViT [56]78.62.9-X1x
ATS [12]79.52.91512-1x
SP-ViT [25]79.32.6-X1x
ToMe [2]79.42.71575✓1.5x
TokenFusion [23]79.52.71580✓1.5x
PaPr (z=0.55)79.22.71585✓1.6x
" + }, + { + "type": "table_caption", + "bbox": [ + 0.22, + 0.327, + 0.501, + 0.384 + ], + "angle": 0, + "content": "Table 5: Performance comparison on ConvNext CNN models. PaPr can achieve competitive performance without training, and can seamlessly adapt to model updates." + }, + { + "type": "table", + "bbox": [ + 0.226, + 0.396, + 0.498, + 0.507 + ], + "angle": 0, + "content": "
ModelsMethodsTrain-FreeAcc1GFLOPsImg/s
ConvNeXtBaselineN/A83.8415.38265
Base-1kDynCNN [40]✘83.0810.21375
PaPr (z=0.65)✓82.7510.42395
ConvNeXtBaselineN/A85.8115.38265
Base-22kPaPr (z=0.65)✓84.2710.42395
ConvNeXtBaselineN/A84.3134.4135
Large-1kPaPr (z=0.65)✓83.2622.9203
ConvNeXtBaselineN/A86.6134.4135
Large-22kPaPr (z=0.65)✓85.6722.9203
" + }, + { + "type": "table_caption", + "bbox": [ + 0.504, + 0.327, + 0.782, + 0.384 + ], + "angle": 0, + "content": "Table 6: Performance comparison on Swin hybrid transformer models. PaPr can adapt to much bigger models with higher operating resolutions without training." + }, + { + "type": "table", + "bbox": [ + 0.512, + 0.396, + 0.779, + 0.507 + ], + "angle": 0, + "content": "
ModelsMethodsTrain-FreeAcc1GFLOPsImg/s
Swin B-1kBaselineN/A83.4215.47258
DynSwin [40]✗83.1812.1327
PaPr (z=0.65)✓81.712.25325
Swin B-22kBaselineN/A85.1615.47258
PaPr (z=0.65)✓82.2712.25325
Swin L-22kBaselineN/A86.2534.53135
PaPr (z=0.65)✓84.5326.96175
Swin L-22k-384BaselineN/A87.25104.0842
PaPr (z=0.65)✓86.4781.4454
" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.539, + 0.788, + 0.646 + ], + "angle": 0, + "content": "MAE models: MAE used self-supervised pre-training by reconstructing masked image patches to train large ViT models [17]. Later, the model is fine-tuned with full resolution images without masking, which cannot exploit its latent ability to learn from fewer patches. Interestingly, PaPr introduces masked inference, which is particularly suitable for MAE models. Hence, we observe significantly higher accuracy with PaPr compared to other training-free methods, e.g., with ViT-B, PaPr achieves \\(4.5\\%\\) higher accuracy over ToMe for similar FLOPs (see Tab. 2)." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.671, + 0.788, + 0.733 + ], + "angle": 0, + "content": "Class token-free models: Several existing patch reduction methods operate with the class token in ViTs to evaluate the patch relevance [12,28,34]. However, instead of class tokens, global pooling of patch tokens is used in many cases [16, 37]. Results presented in Tab. 3 demonstrate the superiority of PaPr for such cases." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.765, + 0.788, + 0.843 + ], + "angle": 0, + "content": "Training-based methods comparison. We compare training performance of existing methods for the patch reduction on DeiT-s [48] model in Tab. 4. Prior pruning based methods cannot speed-up the training for learning the mask predictor [25,41,56]. However, PaPr achieves competitive performance as prior methods, while achieving large speed-up as other token merging methods [2,23,55]." + } + ], + [ + { + "type": "page_number", + "bbox": [ + 0.218, + 0.116, + 0.236, + 0.127 + ], + "angle": 0, + "content": "12" + }, + { + "type": "header", + "bbox": [ + 0.272, + 0.115, + 0.393, + 0.128 + ], + "angle": 0, + "content": "T. Mahmud et al." + }, + { + "type": "table_caption", + "bbox": [ + 0.232, + 0.154, + 0.493, + 0.237 + ], + "angle": 0, + "content": "Table 7: Sweeping ConvNet proposal model in PaPr with \\( z = 0.5 \\). PaPr can achieve similar performance irrespective of proposal model size. Thus, PaPr can use a much smaller and faster proposal model to speed-up larger models." + }, + { + "type": "table", + "bbox": [ + 0.236, + 0.238, + 0.49, + 0.34 + ], + "angle": 0, + "content": "
ProposalAccuracy1(%)
ModelGFLOPsViT-B-16ViT-L-16
ResNet-18 [18]1.8181.183.84
ResNet-50 [18]4.0982.3384.09
ResNet-152 [18]11.5182.5184.08
MobileOne-S0 [49]0.2782.2484.06
MobileOne-S2 [49]1.3582.2884.16
MobileOne-S4 [49]2.9882.3584.32
" + }, + { + "type": "image", + "bbox": [ + 0.513, + 0.144, + 0.77, + 0.295 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.51, + 0.299, + 0.772, + 0.342 + ], + "angle": 0, + "content": "Fig.6: Proposal models locate similar patches with PaPr for different pruning ratio \\((z)\\) irrespective of model size." + }, + { + "type": "table_caption", + "bbox": [ + 0.223, + 0.359, + 0.514, + 0.43 + ], + "angle": 0, + "content": "Table 8: Training-free performance comparison on Kinetics-400 video evaluation. Each video has \\(16 \\times 224^2\\) input size. PaPr achieves significantly better performance for reducing spatio-temporal redundancy in videos." + }, + { + "type": "table", + "bbox": [ + 0.236, + 0.441, + 0.505, + 0.579 + ], + "angle": 0, + "content": "
ModelMethodAcc1GFLOPsViews
XViTATS [12]80.02591x3
TimeSformer-LATS [12]80.535101x3
ViT-B MAEBaseline81.211803x5
PaPr(z=0.45)81.18763x5
PaPr(z=0.35)80.15593x5
ViT-L MAEBaseline85.265983x5
ToMe [2]84.52811x10
PaPr(z=0.5)85.122753x5
ToMe [2]82.51841x10
PaPr(z=0.3)84.531603x5
" + }, + { + "type": "table_caption", + "bbox": [ + 0.517, + 0.36, + 0.78, + 0.431 + ], + "angle": 0, + "content": "Table 9: Proposal ConvNet sweep in video recognition with ViT-B-MAE model as the baseline. PaPr achieves competitive performance with lighter proposal models for different \\( z \\) values." + }, + { + "type": "table", + "bbox": [ + 0.522, + 0.442, + 0.782, + 0.578 + ], + "angle": 0, + "content": "
ProposalKeep Ratio(z)Acc1 (%)Views
ModelGFLOPs
X3d-s [13]1.250.581.113x5
X3d-m [13]2.450.581.193x5
ResNet-50 [14]41.90.579.963x5
ResNet-101 [14]85.670.579.973x5
X3d-s [13]1.250.480.863x5
X3d-m [13]2.450.480.933x5
ResNet-50 [14]41.90.478.853x5
ResNet-101 [14]85.670.479.323x5
" + }, + { + "type": "title", + "bbox": [ + 0.216, + 0.61, + 0.633, + 0.625 + ], + "angle": 0, + "content": "4.3 Performance on Various Hierarchical Models" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.643, + 0.789, + 0.705 + ], + "angle": 0, + "content": "We study the performance on two variants of hierarchical models, such as pure convolutional models, and hybrid transformer models. We compare with training based DynamicCNN [40], and DynamicSwin [40] methods (trained for 120 epochs), whereas PaPr operates without training." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.734, + 0.79, + 0.842 + ], + "angle": 0, + "content": "Analysis on convolutional models. We use the SOTA ConvNext [32] architecture for the analysis as shown in Tab. 5. The training-free operation in PaPr makes it seamlessly usable with new model updates. We analyze ImageNet-1k and ImageNet-22k performance of same models. For the ConvNext-Base model, PaPr achieves \\(99.6\\%\\) accuracy of DynamicCNN with \\(7.1\\%\\) higher throughput. By simply using ImageNet-22k weights, PaPr can achieve additional \\(2\\%\\) accuracy improvement without training, while having the same computation cost." + } + ], + [ + { + "type": "header", + "bbox": [ + 0.562, + 0.115, + 0.732, + 0.127 + ], + "angle": 0, + "content": "PaPr for Faster Inference" + }, + { + "type": "page_number", + "bbox": [ + 0.769, + 0.116, + 0.786, + 0.127 + ], + "angle": 0, + "content": "13" + }, + { + "type": "image", + "bbox": [ + 0.258, + 0.145, + 0.744, + 0.372 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.214, + 0.377, + 0.788, + 0.434 + ], + "angle": 0, + "content": "Fig. 7: Robustness of PaPr compared to CAM based methods. PaPr can perform even when the ConvNet proposal confidence (c) is very low. In contrast, existing CAM based methods fail in such cases, despite being significantly slower while not enabling batch processing and use of gradients in some cases." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.461, + 0.79, + 0.538 + ], + "angle": 0, + "content": "Analysis on hybrid transformer models. We study the hierarchical Swin transformer models for patch reduction with PaPr, as given in Tab. 6. PaPr achieves \\(98\\%\\) of DynamicSwin accuracy using similar FLOPs without re-training. Nevertheless, by leveraging bigger models, larger pre-training, and higher resolution, PaPr can seamlessly adapt to achieve higher accuracy." + }, + { + "type": "title", + "bbox": [ + 0.214, + 0.559, + 0.709, + 0.575 + ], + "angle": 0, + "content": "4.4 Robustness of PaPr across various ConvNet proposals" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.583, + 0.789, + 0.69 + ], + "angle": 0, + "content": "PaPr can operate with ultra-lightweight ConvNets for generating robust proposals. We study different ConvNet architectures for proposal generation, as shown in Tab. 7. Interestingly, PaPr shows small reductions of final accuracy \\((0.3\\%)\\), when using MobileOne-S0 based proposals compared to ResNet-152 \\((42\\times\\) higher FLOPs) in ViT-B. When visualizing the PSM for different pruning ratio as in Fig. 6, we notice the similar PSMs irrespective of model size. Hence, PaPr can utilize ultra-lightweight ConvNets to mask patches without sacrificing accuracy." + }, + { + "type": "title", + "bbox": [ + 0.214, + 0.71, + 0.704, + 0.727 + ], + "angle": 0, + "content": "4.5 Comparison with Class Activation Mappings (CAMs)" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.735, + 0.789, + 0.843 + ], + "angle": 0, + "content": "Existing CAM methods mostly focus on explainability to highlight image regions responsible for final prediction. However, such objectives rely on final prediction performance, which can be much lower for light ConvNets. Moreover, for localization, many of these methods use gradients, and optimization methods that limit batch processing. Nevertheless, we study the impact of such CAM methods ([43, 50, 61]) in PaPr framework, in challenging scenarios where the baseline ViT-B has higher prediction confidence (c) on target class, and the proposal" + } + ], + [ + { + "type": "page_number", + "bbox": [ + 0.218, + 0.116, + 0.236, + 0.127 + ], + "angle": 0, + "content": "14" + }, + { + "type": "header", + "bbox": [ + 0.272, + 0.115, + 0.393, + 0.127 + ], + "angle": 0, + "content": "T. Mahmud et al." + }, + { + "type": "image", + "bbox": [ + 0.26, + 0.144, + 0.744, + 0.24 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.214, + 0.242, + 0.788, + 0.299 + ], + "angle": 0, + "content": "Fig. 8: Visualization of PaPr localization in videos. Video has inherent high sparsity. PaPr effectively localizes the discriminative regions for using holistic spatio-temporal understanding with small ConvNets. Thus, it significantly reduces the computational burden for larger models in downstream video recognition tasks." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.327, + 0.788, + 0.404 + ], + "angle": 0, + "content": "MobileOne-S0 has lower confidence (see Fig. 7). In most cases, the baseline CAM method significantly lowers the final accuracy after patch pruning. In contrast, PaPr maintains robust confidence for its precise localization. Moreover, PaPr can enhance confidence of baseline models in several scenarios by suppressing the redundant patches, e.g., in Sample 5, the \\( c \\) increased by \\( 22\\% \\) with PaPr." + }, + { + "type": "title", + "bbox": [ + 0.216, + 0.425, + 0.44, + 0.444 + ], + "angle": 0, + "content": "5 Video Experiments" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.455, + 0.79, + 0.669 + ], + "angle": 0, + "content": "We study the training-free performance on Kinetics-400 [22] validation sets, as shown in Tab. 8. We use SOTA ATS [12], and ToMe [2] methods for comparison. We use PaPr on SOTA ViT-MAE [47] models with lightweight X3d-s [13] model for proposal generation. We follow the baseline [47] \\(3 \\times 5\\) view (3 spatial view and 5 temporal view) approach, and separate the views from model FLOP counts similar to other works as it can be arbitrarily chosen. Since ViT-MAE removes class tokens in fine-tuning, ATS [12] cannot be adapted to these models. For ViT-L, PaPr achieves \\(3.7 \\times\\) FLOPs reduction of the baseline model for a negligible \\(0.8\\%\\) accuracy drop. We also study the impact of various proposal models on the final performance, as shown in Tab. 9. We use SlowOnly-ResNet [14] and X3d [13] models for comparison. Increasing model size for different pruning shows minimal impact on final performance. Finally, we visualize the patch masking of PaPr in videos, as shown in Fig. 8. By removing the redundant details, PaPr significantly boosts the performance of bigger models for fine-grained predictions." + }, + { + "type": "title", + "bbox": [ + 0.216, + 0.689, + 0.533, + 0.705 + ], + "angle": 0, + "content": "6 Conclusion and Future Work" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.72, + 0.788, + 0.841 + ], + "angle": 0, + "content": "In this paper, we introduce a novel patch pruning method, namely PaPr, that can effectively speed-up off-the-shelf pre-trained models inference without re-training. We propose a simple modification of ConvNets, that allows extracting a precise discriminative PSM from ultra lightweight ConvNets with unparalleled speed. In PaPr, such PSMs are leveraged to directly reduce the redundant data regions in a single-step to guide larger model computations on most discriminative regions for achieving computational efficiency. Moreover, PaPr can be easily integrated with most existing patch reduction methods due to its simple structure." + } + ], + [ + { + "type": "header", + "bbox": [ + 0.561, + 0.115, + 0.732, + 0.128 + ], + "angle": 0, + "content": "PaPr for Faster Inference" + }, + { + "type": "page_number", + "bbox": [ + 0.769, + 0.117, + 0.786, + 0.127 + ], + "angle": 0, + "content": "15" + }, + { + "type": "title", + "bbox": [ + 0.218, + 0.145, + 0.395, + 0.163 + ], + "angle": 0, + "content": "Acknowledgement" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.174, + 0.788, + 0.22 + ], + "angle": 0, + "content": "This work was supported in part by ONR Minerva program, iMAGiNE - the Intelligent Machine Engineering Consortium at UT Austin, and a UT Cockrell School of Engineering Doctoral Fellowship." + }, + { + "type": "title", + "bbox": [ + 0.218, + 0.24, + 0.324, + 0.256 + ], + "angle": 0, + "content": "References" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.267, + 0.787, + 0.323 + ], + "angle": 0, + "content": "1. Belharbi, S., Sarraf, A., Pedersoli, M., Ben Ayed, I., McCaffrey, L., Granger, E.: F-cam: Full resolution class activation maps via guided parametric upscaling. In: Proceedings of the IEEE/CVF Winter Conference on Applications of Computer Vision. pp. 3490-3499 (2022)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.324, + 0.787, + 0.364 + ], + "angle": 0, + "content": "2. Bolya, D., Fu, C.Y., Dai, X., Zhang, P., Feichtenhofer, C., Hoffman, J.: Token merging: Your vit but faster. In: The Twelfth International Conference on Learning Representations (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.365, + 0.787, + 0.391 + ], + "angle": 0, + "content": "3. Cao, Q., Paranjape, B., Hajishirzi, H.: Pumer: Pruning and merging tokens for efficient vision language models. arXiv preprint arXiv:2305.17530 (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.392, + 0.787, + 0.432 + ], + "angle": 0, + "content": "4. Caron, M., Touvron, H., Misra, I., Jégou, H., Mairal, J., Bojanowski, P., Joulin, A.: Emerging properties in self-supervised vision transformers. In: Proceedings of the IEEE/CVF international conference on computer vision. pp. 9650-9660 (2021)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.433, + 0.787, + 0.486 + ], + "angle": 0, + "content": "5. Chattopadhyay, A., Sarkar, A., Howlader, P., Balasubramanian, V.N.: Grad-cam++: Generalized gradient-based visual explanations for deep convolutional networks. In: 2018 IEEE winter conference on applications of computer vision (WACV). pp. 839-847. IEEE (2018)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.487, + 0.787, + 0.528 + ], + "angle": 0, + "content": "6. Chen, Y., Liu, H., Yin, H., Fan, B.: Building vision transformers with hierarchy aware feature aggregation. In: Proceedings of the IEEE/CVF International Conference on Computer Vision (ICCV). pp. 5908-5918 (October 2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.529, + 0.787, + 0.568 + ], + "angle": 0, + "content": "7. Chu, X., Tian, Z., Wang, Y., Zhang, B., Ren, H., Wei, X., Xia, H., Shen, C.: Twins: Revisiting the design of spatial attention in vision transformers. Advances in Neural Information Processing Systems 34, 9355-9366 (2021)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.569, + 0.787, + 0.609 + ], + "angle": 0, + "content": "8. Ding, X., Zhang, X., Ma, N., Han, J., Ding, G., Sun, J.: Repvgg: Making vgg-style convnets great again. In: Proceedings of the IEEE/CVF conference on computer vision and pattern recognition. pp. 13733-13742 (2021)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.61, + 0.787, + 0.649 + ], + "angle": 0, + "content": "9. Dong, X., Bao, J., Chen, D., Zhang, W., Yu, N., Yuan, L., Chen, D., Guo, B.: Cswin transformer: A general vision transformer backbone with cross-shaped windows. In: CVPR (2022)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.65, + 0.787, + 0.691 + ], + "angle": 0, + "content": "0. Dosovitskiy, A., Beyer, L., Kolesnikov, A., Weissenborn, D., Zhai, X., Unterthiner, T., Dehghani, M., Minderer, M., Heigold, G., Gelly, S., et al.: An image is worth 16x16 words: Transformers for image recognition at scale. In: ICLR (2020)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.691, + 0.787, + 0.718 + ], + "angle": 0, + "content": "1. Fan, H., Xiong, B., Mangalam, K., Li, Y., Yan, Z., Malik, J., Feichtenhofer, C.: Multiscale vision transformers. In: ICCV (2021)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.719, + 0.787, + 0.759 + ], + "angle": 0, + "content": "2. Fayyaz, M., Kouhpayegani, S.A., Jafari, F.R., Sommerlade, E., Joze, H.R.V., Pirsivash, H., Gall, J.: Ats: Adaptive token sampling for efficient vision transformers. In: ECCV (2022)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.759, + 0.787, + 0.8 + ], + "angle": 0, + "content": "3. Feichtenhofer, C.: X3d: Expanding architectures for efficient video recognition. In: Proceedings of the IEEE/CVF conference on computer vision and pattern recognition. pp. 203-213 (2020)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.801, + 0.787, + 0.84 + ], + "angle": 0, + "content": "4. Feichtenhofer, C., Fan, H., Malik, J., He, K.: Slowfast networks for video recognition. In: Proceedings of the IEEE/CVF international conference on computer vision. pp. 6202-6211 (2019)" + }, + { + "type": "list", + "bbox": [ + 0.226, + 0.267, + 0.787, + 0.84 + ], + "angle": 0, + "content": null + } + ], + [ + { + "type": "page_number", + "bbox": [ + 0.218, + 0.116, + 0.236, + 0.127 + ], + "angle": 0, + "content": "16" + }, + { + "type": "header", + "bbox": [ + 0.272, + 0.115, + 0.392, + 0.128 + ], + "angle": 0, + "content": "T. Mahmud et al." + }, + { + "type": "ref_text", + "bbox": [ + 0.218, + 0.147, + 0.787, + 0.189 + ], + "angle": 0, + "content": "15. Graham, B., El-Nouby, A., Touvron, H., Stock, P., Joulin, A., Jégou, H., Douze, M.: Levit: a vision transformer in convnet's clothing for faster inference. In: ICCV (2021)" + }, + { + "type": "ref_text", + "bbox": [ + 0.218, + 0.19, + 0.787, + 0.232 + ], + "angle": 0, + "content": "16. Hatamizadeh, A., Yin, H., Heinrich, G., Kautz, J., Molchanov, P.: Global context vision transformers. In: International Conference on Machine Learning. pp. 12633-12646. PMLR (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.232, + 0.787, + 0.259 + ], + "angle": 0, + "content": "17. He, K., Chen, X., Xie, S., Li, Y., Dollar, P., Girshick, R.: Masked autoencoders are scalable vision learners. In: CVPR (2022)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.26, + 0.787, + 0.3 + ], + "angle": 0, + "content": "18. He, K., Zhang, X., Ren, S., Sun, J.: Deep residual learning for image recognition. In: Proceedings of the IEEE conference on computer vision and pattern recognition. pp. 770-778 (2016)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.301, + 0.787, + 0.342 + ], + "angle": 0, + "content": "19. Howard, A., Sandler, M., Chu, G., Chen, L.C., Chen, B., Tan, M., Wang, W., Zhu, Y., Pang, R., Vasudevan, V., et al.: Searching for mobilenetv3. In: Proceedings of the IEEE/CVF international conference on computer vision. pp. 1314-1324 (2019)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.343, + 0.787, + 0.384 + ], + "angle": 0, + "content": "20. Howard, A.G., Zhu, M., Chen, B., Kalenichenko, D., Wang, W., Weyand, T., Andreetto, M., Adam, H.: Mobilenets: Efficient convolutional neural networks for mobile vision applications. arXiv preprint arXiv:1704.04861 (2017)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.385, + 0.787, + 0.424 + ], + "angle": 0, + "content": "21. Jung, H., Oh, Y.: Towards better explanations of class activation mapping. In: Proceedings of the IEEE/CVF International Conference on Computer Vision. pp. 1336-1344 (2021)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.425, + 0.787, + 0.467 + ], + "angle": 0, + "content": "22. Kay, W., Carreira, J., Simonyan, K., Zhang, B., Hillier, C., Vijayanarasimhan, S., Viola, F., Green, T., Back, T., Natsev, P., et al.: The kinetics human action video dataset. arXiv:1705.06950 [cs.CV] (2017)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.468, + 0.787, + 0.509 + ], + "angle": 0, + "content": "23. Kim, M., Gao, S., Hsu, Y.C., Shen, Y., Jin, H.: Token fusion: Bridging the gap between token pruning and token merging. In: Proceedings of the IEEE/CVF Winter Conference on Applications of Computer Vision. pp. 1383-1392 (2024)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.509, + 0.787, + 0.536 + ], + "angle": 0, + "content": "24. Kim, S., Shen, S., Thorsley, D., Gholami, A., Kwon, W., Hassoun, J., Keutzer, K.: Learned token pruning for transformers. arXiv:2107.00910 [cs.CL] (2021)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.537, + 0.787, + 0.577 + ], + "angle": 0, + "content": "25. Kong, Z., Dong, P., Ma, X., Meng, X., Niu, W., Sun, M., Ren, B., Qin, M., Tang, H., Wang, Y.: Spvit: Enabling faster vision transformers via soft token pruning. In: ECCV (2022)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.578, + 0.787, + 0.605 + ], + "angle": 0, + "content": "26. Lassance, C., Maachou, M., Park, J., Clinchant, S.: A study on token pruning for colbert. arXiv:2112.06540 [cs.CL] (2021)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.606, + 0.787, + 0.646 + ], + "angle": 0, + "content": "27. Li, Y., Wu, C.Y., Fan, H., Mangalam, K., Xiong, B., Malik, J., Feichtenhofer, C.: Mvittv2: Improved multiscale vision transformers for classification and detection. In: CVPR (2022)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.647, + 0.787, + 0.674 + ], + "angle": 0, + "content": "28. Liang, Y., Ge, C., Tong, Z., Song, Y., Wang, J., Xie, P.: Not all patches are what you need: Expediting vision transformers via token reorganizations. ICLR (2022)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.675, + 0.787, + 0.715 + ], + "angle": 0, + "content": "29. Lin, J., Chen, W.M., Cai, H., Gan, C., Han, S.: Memory-efficient patch-based inference for tiny deep learning. Advances in Neural Information Processing Systems 34, 2346-2358 (2021)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.716, + 0.787, + 0.757 + ], + "angle": 0, + "content": "30. Liu, Z., Hu, H., Lin, Y., Yao, Z., Xie, Z., Wei, Y., Ning, J., Cao, Y., Zhang, Z., Dong, L., et al.: Swin transformer v2: Scaling up capacity and resolution. In: CVPR (2022)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.758, + 0.787, + 0.799 + ], + "angle": 0, + "content": "31. Liu, Z., Lin, Y., Cao, Y., Hu, H., Wei, Y., Zhang, Z., Lin, S., Guo, B.: Swin transformer: Hierarchical vision transformer using shifted windows. In: ICCV (2021)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.799, + 0.787, + 0.84 + ], + "angle": 0, + "content": "32. Liu, Z., Mao, H., Wu, C.Y., Feichtenhofer, C., Darrell, T., Xie, S.: A convnet for the 2020s. In: Proceedings of the IEEE/CVF conference on computer vision and pattern recognition. pp. 11976-11986 (2022)" + }, + { + "type": "list", + "bbox": [ + 0.217, + 0.147, + 0.787, + 0.84 + ], + "angle": 0, + "content": null + } + ], + [ + { + "type": "header", + "bbox": [ + 0.562, + 0.115, + 0.732, + 0.127 + ], + "angle": 0, + "content": "PaPr for Faster Inference" + }, + { + "type": "page_number", + "bbox": [ + 0.769, + 0.116, + 0.786, + 0.127 + ], + "angle": 0, + "content": "17" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.147, + 0.788, + 0.175 + ], + "angle": 0, + "content": "33. Mehta, S., Rastegari, M.: Mobilevit: light-weight, general-purpose, and mobile-friendly vision transformer. In: ICLR (2021)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.175, + 0.788, + 0.203 + ], + "angle": 0, + "content": "34. Meng, L., Li, H., Chen, B.C., Lan, S., Wu, Z., Jiang, Y.G., Lim, S.N.: Adavit: Adaptive vision transformers for efficient image recognition. In: CVPR (2022)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.203, + 0.788, + 0.245 + ], + "angle": 0, + "content": "35. Muhammad, M.B., Yeasin, M.: Eigen-cam: Class activation map using principal components. In: 2020 international joint conference on neural networks (IJCNN). pp. 1-7. IEEE (2020)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.245, + 0.788, + 0.286 + ], + "angle": 0, + "content": "36. Oquab, M., Darcet, T., Moutakanni, T., Vo, H., Szafraniec, M., Khalidov, V., Fernandez, P., Haziza, D., Massa, F., El-Nouby, A., et al.: Dinov2: Learning robust visual features without supervision. arXiv preprint arXiv:2304.07193 (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.286, + 0.788, + 0.327 + ], + "angle": 0, + "content": "37. Pan, Z., Zhuang, B., Liu, J., He, H., Cai, J.: Scalable vision transformers with hierarchical pooling. In: Proceedings of the IEEE/cvf international conference on computer vision. pp. 377-386 (2021)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.328, + 0.788, + 0.37 + ], + "angle": 0, + "content": "38. Patro, B.N., Lunayach, M., Patel, S., Namboodiri, V.P.: U-cam: Visual explanation using uncertainty based class activation maps. In: Proceedings of the IEEE/CVF International Conference on Computer Vision. pp. 7444-7453 (2019)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.37, + 0.788, + 0.424 + ], + "angle": 0, + "content": "39. Radford, A., Kim, J.W., Hallacy, C., Ramesh, A., Goh, G., Agarwal, S., Sastry, G., Askell, A., Mishkin, P., Clark, J., et al.: Learning transferable visual models from natural language supervision. In: International conference on machine learning. pp. 8748-8763. PMLR (2021)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.424, + 0.788, + 0.466 + ], + "angle": 0, + "content": "40. Rao, Y., Liu, Z., Zhao, W., Zhou, J., Lu, J.: Dynamic spatial sparsification for efficient vision transformers and convolutional neural networks. IEEE Transactions on Pattern Analysis and Machine Intelligence (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.466, + 0.788, + 0.509 + ], + "angle": 0, + "content": "41. Rao, Y., Zhao, W., Liu, B., Lu, J., Zhou, J., Hsieh, C.J.: Dynamicvit: Efficient vision transformers with dynamic token sparsification. Advances in neural information processing systems 34, 13937-13949 (2021)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.509, + 0.788, + 0.55 + ], + "angle": 0, + "content": "42. Sandler, M., Howard, A., Zhu, M., Zhmoginov, A., Chen, L.C.: Mobilenetv2: Inverted residuals and linear bottlenecks. In: Proceedings of the IEEE conference on computer vision and pattern recognition. pp. 4510-4520 (2018)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.55, + 0.788, + 0.591 + ], + "angle": 0, + "content": "43. Selvaraju, R.R., Cogswell, M., Das, A., Vedantam, R., Parikh, D., Batra, D.: Gradcam: Visual explanations from deep networks via gradient-based localization. In: ICCV (2017)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.591, + 0.788, + 0.633 + ], + "angle": 0, + "content": "44. Singh, M., Gustafson, L., Adcock, A., de Freitas Reis, V., Gedik, B., Kosaraju, R.P., Mahajan, D., Girshick, R., Dollar, P., van der Maaten, L.: Revisiting weakly supervised pre-training of visual perception models. In: CVPR (2022)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.633, + 0.788, + 0.674 + ], + "angle": 0, + "content": "45. Steiner, A., Kolesnikov, A., Zhai, X., Wightman, R., Uszkoreit, J., Beyer, L.: How to train your vit? data, augmentation, and regularization in vision transformers. TMLR (2022)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.674, + 0.788, + 0.702 + ], + "angle": 0, + "content": "46. Tan, M., Le, Q.: Efficientnet: Rethinking model scaling for convolutional neural networks. In: ICML (2019)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.702, + 0.788, + 0.743 + ], + "angle": 0, + "content": "47. Tong, Z., Song, Y., Wang, J., Wang, L.: Videomae: Masked autoencoders are data-efficient learners for self-supervised video pre-training. arXiv:2203.12602 [cs.CV] (2022)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.743, + 0.788, + 0.772 + ], + "angle": 0, + "content": "48. Touvron, H., Cord, M., Douze, M., Massa, F., Sablayrolles, A., Jégou, H.: Training data-efficient image transformers & distillation through attention. In: ICML (2021)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.772, + 0.788, + 0.813 + ], + "angle": 0, + "content": "49. Vasu, P.K.A., Gabriel, J., Zhu, J., Tuzel, O., Ranjan, A.: Mobileone: An improved one millisecond mobile backbone. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 7907-7917 (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.813, + 0.788, + 0.84 + ], + "angle": 0, + "content": "50. Wang, H., Wang, Z., Du, M., Yang, F., Zhang, Z., Ding, S., Mardziel, P., Hu, X.: Score-cam: Score-weighted visual explanations for convolutional neural networks." + }, + { + "type": "list", + "bbox": [ + 0.217, + 0.147, + 0.788, + 0.84 + ], + "angle": 0, + "content": null + } + ], + [ + { + "type": "page_number", + "bbox": [ + 0.218, + 0.116, + 0.236, + 0.127 + ], + "angle": 0, + "content": "18" + }, + { + "type": "header", + "bbox": [ + 0.272, + 0.115, + 0.392, + 0.128 + ], + "angle": 0, + "content": "T. Mahmud et al." + }, + { + "type": "ref_text", + "bbox": [ + 0.241, + 0.147, + 0.786, + 0.175 + ], + "angle": 0, + "content": "In: Proceedings of the IEEE/CVF conference on computer vision and pattern recognition workshops. pp. 24-25 (2020)" + }, + { + "type": "ref_text", + "bbox": [ + 0.218, + 0.176, + 0.786, + 0.203 + ], + "angle": 0, + "content": "51. Wang, S., Li, B.Z., Khabsa, M., Fang, H., Ma, H.: Linformer: Self-attention with linear complexity. arXiv preprint arXiv:2006.04768 (2020)" + }, + { + "type": "ref_text", + "bbox": [ + 0.218, + 0.204, + 0.787, + 0.245 + ], + "angle": 0, + "content": "52. Wei, C., Fan, H., Xie, S., Wu, C.Y., Yuille, A., Feichtenhofer, C.: Masked feature prediction for self-supervised visual pre-training. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 14668-14678 (2022)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.246, + 0.787, + 0.299 + ], + "angle": 0, + "content": "53. Woo, S., Debnath, S., Hu, R., Chen, X., Liu, Z., Kweon, I.S., Xie, S.: Convnext v2: Co-designing and scaling convnets with masked autoencoders. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 16133-16142 (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.3, + 0.787, + 0.342 + ], + "angle": 0, + "content": "54. Wu, H., Xiao, B., Codella, N., Liu, M., Dai, X., Yuan, L., Zhang, L.: Cvt: Introducing convolutions to vision transformers. In: Proceedings of the IEEE/CVF international conference on computer vision. pp. 22-31 (2021)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.343, + 0.787, + 0.384 + ], + "angle": 0, + "content": "55. Xu, X., Wang, S., Chen, Y., Zheng, Y., Wei, Z., Liu, J.: Gtp-vit: Efficient vision transformers via graph-based token propagation. In: Proceedings of the IEEE/CVF Winter Conference on Applications of Computer Vision. pp. 86-95 (2024)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.384, + 0.787, + 0.424 + ], + "angle": 0, + "content": "56. Yin, H., Vahdat, A., Alvarez, J.M., Mallya, A., Kautz, J., Molchanov, P.: A-vit: Adaptive tokens for efficient vision transformer. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 10809-10818 (2022)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.425, + 0.787, + 0.48 + ], + "angle": 0, + "content": "57. You, H., Xiong, Y., Dai, X., Wu, B., Zhang, P., Fan, H., Vajda, P., Lin, Y.C.: Castling-vit: Compressing self-attention via switching towards linear-angular attention at vision transformer inference. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 14431-14442 (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.481, + 0.787, + 0.508 + ], + "angle": 0, + "content": "58. Yu, H., Wu, J.: A unified pruning framework for vision transformers. arXiv:2111.15127 [cs.CV] (2021)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.509, + 0.787, + 0.55 + ], + "angle": 0, + "content": "59. Zhang, X., Zhou, X., Lin, M., Sun, J.: Shufflenet: An extremely efficient convolutional neural network for mobile devices. In: Proceedings of the IEEE conference on computer vision and pattern recognition. pp. 6848-6856 (2018)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.55, + 0.787, + 0.591 + ], + "angle": 0, + "content": "60. Zhou, A., Li, Y., Qin, Z., Liu, J., Pan, J., Zhang, R., Zhao, R., Gao, P., Li, H.: Sparsemae: Sparse training meets masked autoencoders. In: Proceedings of the IEEE/CVF International Conference on Computer Vision. pp. 16176-16186 (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.592, + 0.787, + 0.633 + ], + "angle": 0, + "content": "61. Zhou, B., Khosla, A., Lapedriza, A., Oliva, A., Torralba, A.: Learning deep features for discriminative localization. In: Proceedings of the IEEE conference on computer vision and pattern recognition. pp. 2921-2929 (2016)" + }, + { + "type": "list", + "bbox": [ + 0.217, + 0.147, + 0.787, + 0.633 + ], + "angle": 0, + "content": null + } + ] +] \ No newline at end of file diff --git a/2024/PaPr_ Training-Free One-Step Patch Pruning with Lightweight ConvNets for Faster Inference/f59c8716-3e2f-464a-8e1c-4f54e686a19f_origin.pdf b/2024/PaPr_ Training-Free One-Step Patch Pruning with Lightweight ConvNets for Faster Inference/f59c8716-3e2f-464a-8e1c-4f54e686a19f_origin.pdf new file mode 100644 index 0000000000000000000000000000000000000000..741156e00a3a1e82aabcab2813ae58e96d7d48ba --- /dev/null +++ b/2024/PaPr_ Training-Free One-Step Patch Pruning with Lightweight ConvNets for Faster Inference/f59c8716-3e2f-464a-8e1c-4f54e686a19f_origin.pdf @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:942c582f2bc65dcca963d1beb4f7a920715c8ca494e508155e5cbbdeabbd6993 +size 15562315 diff --git a/2024/PaPr_ Training-Free One-Step Patch Pruning with Lightweight ConvNets for Faster Inference/full.md b/2024/PaPr_ Training-Free One-Step Patch Pruning with Lightweight ConvNets for Faster Inference/full.md new file mode 100644 index 0000000000000000000000000000000000000000..d814bcd5eb5cf710a04c2d162d33e347d7ee1ad4 --- /dev/null +++ b/2024/PaPr_ Training-Free One-Step Patch Pruning with Lightweight ConvNets for Faster Inference/full.md @@ -0,0 +1,310 @@ +# PaPr: Training-Free One-Step Patch Pruning with Lightweight ConvNets for Faster Inference + +Tanvir Mahmud1 Burhaneddin Yaman2 Chun-Hao Liu3 Diana Marculescu1 + +1University of Texas at Austin 2Bosch Research 3Amazon Prime Video 1{tanvirmahmud, dianam}@utexas.edu 2yaman013@umn.edu 3liuch37@gmail.com + +Abstract. As deep neural networks evolve from convolutional neural networks (ConvNets) to advanced vision transformers (ViTs), there is an increased need to eliminate redundant data for faster processing without compromising accuracy. Previous methods are often architecture-specific or necessitate re-training, restricting their applicability with frequent model updates. To solve this, we first introduce a novel property of lightweight ConvNets: their ability to identify key discriminative patch regions in images, irrespective of model's final accuracy or size. We demonstrate that fully-connected layers are the primary bottleneck for ConvNets performance, and their suppression with simple weight recalibration markedly enhances discriminative patch localization performance. Using this insight, we introduce PaPr, a method for substantially pruning redundant patches with minimal accuracy loss using lightweight ConvNets across a variety of deep learning architectures, including ViTs, ConvNets, and hybrid transformers, without any re-training. Moreover, the simple early-stage one-step patch pruning with PaPr enhances existing patch reduction methods. Through extensive testing on diverse architectures, PaPr achieves significantly higher accuracy over state-of-the-art patch reduction methods with similar FLOP count reduction. More specifically, PaPr reduces about $70\%$ of redundant patches in videos with less than $0.8\%$ drop in accuracy, and up to $3.7\times$ FLOPs reduction, which is a $15\%$ more reduction with $2.5\%$ higher accuracy. Code is available at https://github.com/tanvir-utexas/PaPr. + +# 1 Introduction + +Deep neural networks have grown from simple convolutional neural networks (ConvNets) to complex transformer models, aiming for better accuracy with more computation [10, 18]. Vision transformers (ViTs) excel by focusing on important parts of images with long-range attention and large-scale pre-training [17,39]. This success, however, comes at the cost of increased computational cost. Operating these large networks efficiently for downstream applications is crucial. Most visual tasks demand precisely pinpointing key image regions against complex backgrounds. Higher image resolution improves accuracy by capturing more details but also adds unnecessary background processing, burdening large models [44]. This highlights the need to cut down on redundant data in high-resolution images to maintain both speed and performance with advanced techniques [2,34]. + +![](images/2ebd885cf12a80b9567181d16380666f4a52e1d5d91ad631a18b3dd54be74dd6.jpg) +(b) Proposed PaPr +Fig. 1: (a) Existing patch pruning methods gradually reduce patches over the model. This requires additional training of mask generators in intermediate layers. (b) Proposed PaPr directly prunes redundant patches early in the network by leveraging pretrained lightweight ConvNets and directly speeds-up off-the-shelf models without re-training. + +Identifying key image regions demands a comprehensive understanding of the image and model operations. Incorrect key region estimation can impair pretraining performance by eliminating crucial areas. Traditional approaches [2,25,28, 34,56] for pruning redundant patch regions are hindered by three main limitations: (1) They often necessitate complex training of extra modules, which becomes increasingly difficult as baseline models evolve with more data, enhanced training methods, and deeper structures [25,41,56]. Re-training these modules for each model update is impractical. (2) Without a complete image understanding, these methods incrementally prune patches across the network, leading to unnecessary computations in early layers, which is particularly problematic for deeper models. (3) Many rely on specific architectural features for patch reduction, such as class tokens or attention maps, limiting their use to a narrow set of network designs [12,28]. Hence, there is a pressing need for a patch pruning solution that is adaptable to various modern architectures without additional training, can eliminate redundant patches in a single-step, thereby making it suitable for a broad spectrum of networks while streamlining each model update (See Fig. 1). + +Recent work on patch reduction [2, 23, 55] focuses mostly on transformers rather than ConvNets due to their impressive performance on various tasks, their aforementioned limitations notwithstanding. While ConvNets achieve lower ImageNet-1k top-1 accuracy than large ViTs (68.7% in MobileOne-S0 [49] vs. 88.7% in ViT-Huge [44]), they exhibit a remarkable ability to efficiently process the key image regions with hierarchical inductive bias. Our empirical investigation reveals that, as we broaden the evaluation metric (increasing $k$ in top- $k$ evaluations, see Fig. 4), the benefits of deeper models diminish, especially with a large number of image classes (e.g., 1000 in ImageNet). This suggests that shallower models excel at identifying discriminative areas as their bigger counterparts, rendering them ideal for patch pruning. + +Leveraging this insight, we propose PaPr, a novel Patch Pruning method that employs pretrained lightweight ConvNets for efficient, one-step patch pruning in + +a wide variety of deep learning models, maintaining accuracy while significantly cutting computational demands. Our findings show lightweight ConvNets have remarkable ability in identifying discriminative image regions but struggle in fine-grained prediction. To address this, PaPr relies on Patch Significance Maps (PSMs), which are generated using only the convolutional layers of ConvNets through uniform class weight recalibration in FC layers. Astonishingly, PSMs consistently highlight critical image regions across ConvNets of varying sizes and accuracies (Fig. 6), thereby amplifying the efficacy of ultra lightweight ConvNets. + +ConvNets inherently preserve the positional property of patches for their inductive bias, while ViTs use cross-attention to blend patch features, leading to variability in patch relevance across models. Unlike previous methods that prune redundant patches gradually over multiple steps across intermediate layers of ViTs [2, 23, 41], our approach, PaPr, simplifies this process by eliminating non-essential patches at once, immediately after extraction (see Fig. 3), by leveraging lightweight ConvNets to assess patch significance. This direct, one-step pruning approach significantly cuts computational demands and is also compatible with other patch reduction techniques (see Fig. 5). PaPr's ability to separate crucial patch identification from fine-grained class prediction enhances a wide range of pre-trained models without further training, ensuring high accuracy with notable speed ups. By bypassing the complex training required for conventional patch selectors, PaPr capitalizes on the comprehensive capabilities of minimalist ConvNets for efficient patch pruning in larger models. + +Our experiments demonstrate PaPr's effectiveness across various architectures and pre-training methods, achieving significant reduction in redundant patches for ViTs, large-scale ConvNets (e.g., ConvNext [32]), and hybrid transformers (e.g., Swin [31]), outperforming state-of-the-art (SOTA) patch reduction methods by a large margin. Notably, PaPr can be easily integrated with most existing patch reduction methods to reduce patches early in operation. More specifically, PaPr boosts ToMe [2] accuracy by $4.5\%$ for a computational budget similar to ViT-B. Moreover, PaPr can accelerate training akin to token merging techniques, unlike most patch pruning methods that fail to boost training speed [34, 41, 56]. PaPr shows robust patch localization performance with ultra lightweight ConvNets ( $<0.3\%$ accuracy loss) for $42\times$ reduction in proposal FLOPs, thereby enabling its use for larger off-the-shelf models. Remarkably, in video recognition, PaPr cuts down around $70\%$ of redundant patches, resulting in up to $3.7\times$ FLOPs reduction with minimal impact on accuracy ( $\approx 0.8\%$ ). + +We summarize our main contributions as follows: + +- We propose PaPr, a novel background patch pruning method that can seamlessly operate with ViTs, ConvNets, and hybrid transformers, without further training while leveraging batch processing. +- We propose a simple weight recalibration method in ConvNets to precisely and efficiently locate discriminative patches, irrespective of model size. +- We facilitate the use of ultra-lightweight ConvNets to speed-up large models, such as ViTs, with a seamless framework and negligible accuracy loss. + +- We present extensive qualitative and quantitative results across numerous model architectures in both image and video applications. + +# 2 Related Work + +# 2.1 From ConvNets to Vision Transformers + +ConvNets have been pivotal in computer vision, offering computational efficiency through kernel reuse and localization [8, 18-20, 29, 42, 46, 49, 59]. Yet, they struggle in capturing long-range dependencies, a gap bridged by ViT [10]. Inspired by the success of transformers from natural language processing (NLP), ViTs excel in long-range feature modeling through cross-attention, outperforming ConvNets albeit with higher computational demands and extensive training data requirements [7, 48, 51, 57]. Efforts to alleviate these issues include self-supervised [17, 47, 52, 60] and weakly-supervised [39, 44] pre-training, although challenges in computational complexity and optimization remain. Hybrid architectures, merging ConvNets' inductive biases with ViTs' cross-attention, offer a balanced solution by reducing computational load while maintaining performance [6, 9, 11, 15, 27, 30, 31, 33, 54]. Recent ConvNets advancements [32, 53], with improved training and large-scale data, now rival ViTs, questioning whether architectural innovations or enhanced training primarily drive performance gains. + +Addressing the computational demands of these models, we focus on enhancing operational efficiency by pruning redundant patches without architectural modifications or re-training, maintaining performance while streamlining updates. + +# 2.2 Class Activation Mapping for Explainable Deep Learning + +Class Activation Mapping (CAM) techniques provide explainable visual reasoning for neural network predictions by highlighting activation regions1 crucial for decisions [1,21,38,43,61]. While ConvNets utilize convolutional operations to preserve spatial information, making them apt for such visualizations, ViTs link attention weights to class tokens for prediction [4,36]. However, CAM's reliance on accurate class predictions for effective feature localization is a significant drawback, as ConvNets' lower accuracy when compared to ViTs may lead to incorrect object localization. Despite ViTs' superior fine-grained classification capabilities, it is unclear if they prioritize the same discriminative regions as ConvNets. Additionally, CAM approaches often require gradient tracking [5,43] or complex feature map decomposition [35], complicating batch processing and necessitating extra optimization steps. + +Our work diverges from traditional CAM by aiming to consistently identify discriminative patches across different architectures, focusing computational resources on the most relevant areas without being constrained by class prediction accuracy. This approach enables more efficient processing by allowing larger models to concentrate on the most significant regions, identified by lighter networks, thus significantly reducing computational demands. + +![](images/9ac1ccb35defa0ba0d4eae72669c9558242aed3ac197e5bc228178f70f6056c2.jpg) +(a) Baseline Convolutional Model Feature Flow + +![](images/3279da1ba742f58bf7165335ee79f9be4bb66d64ceadb6dce907d0cad97645b7.jpg) +(b) Proposed Pruning Mask Extraction +Fig.2: (a) Baseline ConvNet gradually reduces the feature map to produce $\mathcal{F} = \{f_k(x,y)\}_{k=1}^K$ , followed by global average pooling and fully connected (FC) layers to predict $y_p$ . (b) In PaPr, we operate on $\mathcal{F}$ by suppressing the FC layer. Initially, we extract pixel mean over $K$ channels to produce discriminative region proposal $\mathcal{R}$ . Later, simple upsampling operation generates the patch significance map (PSM) $\mathcal{P}$ of target dimension. Finally, patch mask $\mathcal{M}$ for top $z\%$ patches is obtained from $\mathcal{P}$ . + +# 2.3 Patch Reduction for Faster Inference + +Several approaches have sought to enhance computational efficiency by reducing redundant patches in neural networks, with early strategies involving additional adapters or controllers to identify and prune less significant patches [24,26,41,58]. These methods, however, necessitate separate adapter training for each network and are slow to adapt due to the need to learn from the dynamics of other layers. In the context of ViTs, efforts have leveraged architectural features such as class tokens and attention maps for patch relevance, yet these solutions often fail to generalize across various architectures like hybrid transformers [31] or larger ConvNets [32, 53], thus limiting their applicability [12, 25, 34, 56]. Furthermore, while some have explored patch merging in ViTs and a combination of merging and pruning [2, 3, 23], these approaches lack a holistic image understanding by not considering the entire image context in their optimizations, leading to incremental and sub-optimal patch reduction. + +Our work diverges from these traditional patch reduction methods by proposing a single-step, early-network patch removal strategy, that seamlessly integrates with any architecture without re-training. + +# 3 Methodology + +Our methodology introduces a novel approach to discriminative patch pruning across various deep learning architectures, leveraging the innate capabilities of lightweight ConvNets. By generating a PSM, we efficiently identify and prune non-essential patches in a single step, enhancing computational efficiency without compromising accuracy. This process is universally applicable, seamlessly + +integrating with ViTs, ConvNets, and hybrid models, thereby addressing the limitations of previous methods with a scalable, architecture-agnostic solution. + +# 3.1 Extracting Discriminative Regions with ConvNets + +Despite achieving relatively lower top-1 accuracy on the ImageNet benchmark, lightweight ConvNets exhibit a competitive edge in top-10 accuracy when compared to their larger ViT counterparts (Fig. 4). This phenomenon underscores ConvNets' capability to effectively localize regions of interest through their convolutional layers, despite potential limitations in fine-grained classification attributed to the fully-connected (FC) layer. Our methodology leverages this insight by proposing discriminative patch regions while minimizing the influence of the FC layer (see Fig. 2), thereby enhancing the large off-the-shelf model's focus on most salient image regions proposed by lightweight ConvNets (see Fig. 3). + +Given an input image $X \in \mathbb{R}^{H \times W \times 3}$ , with height $H$ and width $W$ , we denote by $f_{k}(x,y)$ the feature map generated by the $k^{th}$ kernel in the last convolutional layer of typical ConvNets, $\forall k \in \{1,2,\dots,K\}$ . Here, each pixel $(x,y)$ in $f_{k}(x,y)$ corresponds to a patch window of size $(H / d \times W / d)$ in the input image $X$ , reflecting a spatial down-scaling factor of $d$ through the convolutional layer stack. Typically, global average pooling (GAP) is applied to each feature map $f_{k}(x,y)$ , yielding $F = \{F_{k}\}_{k=1}^{K} \in \mathbb{R}^{K}$ , where $F_{k} = \sum_{x,y} f_{k}(x,y) \in \mathbb{R}$ . Finally, a fully connected (FC) layer $W \in \mathbb{R}^{C \times K}$ with elements $w_{c}^{k}$ processes $F$ to generate class predictions $y_{p} \in \mathbb{R}^{C}$ across $C$ classes as follows: + +$$ +y _ {p} = W F = \sum_ {c} \sum_ {k} w _ {c} ^ {k} \sum_ {x, y} f _ {k} (x, y) = \sum_ {x, y} R (x, y), \tag {1} +$$ + +where $R \in \mathbb{R}^{h \times w}$ encapsulates the weighted mean class activation mapping of the convolutional feature map. + +We note that the reliance on class activation weights $w_{c}^{k}$ is influenced by the model's final accuracy, thereby posing a challenge for smaller models. Our objective transcends mere accuracy enhancement, aiming to precisely locate discriminative patches rich in information irrespective of the model size or its final classification performance. Recognizing that discriminative region localization is pivotal for detailed classification, and acknowledging the competitive top-10 accuracy of lighter ConvNets, we posit that an extremely lightweight ConvNet suffices for initial discriminative region proposal. + +To counteract the influence of the weak linear FC layer $W$ in convolutional region proposal, we propose an adjustment where $w_{c}^{k} = 1 / KC$ , $\forall k\in \{1,\dots ,K\}$ , facilitating the generation of a discriminative region proposal $\mathcal{R}\in \mathbb{R}^{h\times w}$ , where + +$$ +\mathcal {R} (x, y) = \sum_ {c} \sum_ {k} \frac {1}{K C} f _ {k} (x, y) = \frac {1}{K} \sum_ {k} f _ {k} (x, y). \tag {2} +$$ + +This strategy allows us to leverage ConvNets for what they excel at: pinpointing critical image areas. By reducing reliance on class activation weights, we + +efficiently generate discriminative region proposals directly from convolutional outputs. This approach not only enhances the interpretability and efficiency of the localization process but also enables the application of more complex models for subsequent detailed analysis, optimizing the use of computational resources. + +# 3.2 Patch Significance Map + +With the discriminative region proposal $\mathcal{R} \in \mathbb{R}^{h \times w}$ , where each pixel $(x, y)$ quantifies the significance of corresponding patches in the original image, we proceed to establish a precise mapping to our intended feature map (Fig. 2). This mapping ensures the preservation of spatial relationships within the feature map. + +Consider a target feature map $\mathcal{F} \in \mathbb{R}^{h' \times w' \times K}$ , with each pixel $(x', y')$ encapsulating a feature vector from a specific image patch. To align $\mathcal{R}$ with $\mathcal{F}$ , we employ an upsampling operation $U: \mathbb{R}^{h \times w} \to \mathbb{R}^{h' \times w'}$ , transforming $\mathcal{R}$ into the Patch Significance Map (PSM) $\mathcal{P} \in \mathbb{R}^{h' \times w'}$ . Consequently, each element of $\mathcal{P}$ directly corresponds to the patch significance within $\mathcal{F}$ for the given discriminative task. The next step involves utilizing $\mathcal{P}$ to discern and prune non-essential patch features from $\mathcal{F}$ . By sorting the values within $\mathcal{P}$ , we acquire a pruning mask $\mathcal{M}$ : + +$$ +\mathcal {M} = \operatorname {r e s h a p e} (\operatorname {a r g s o r t} (\operatorname {f l a t t e n} (\mathcal {P}))), \mathcal {M} \in \mathbb {R} ^ {h ^ {\prime} \times w ^ {\prime}}, \tag {3} +$$ + +enabling batch-wise patch pruning. Specifically, we identify the indices corresponding to the top- $z\%$ patches as per $\mathcal{M}$ , facilitating the retention of only the most salient patches in $\mathcal{F}$ , thereby enhancing computational efficiency in subsequent processing stages. + +As part of our discussion on various architectures, we detail the application of this patch reduction technique. For enhanced visualization of the PSM, we apply min-max normalization to $\mathcal{P}$ , adjusting for outliers and scaling the significance scores to visually depict the importance of different patches. + +# 3.3 Integrating PSM with Vision Transformers + +The ViT processes an input image $X \in \mathbb{R}^{H \times W \times 3}$ by extracting $N$ -dimensional patch token features $X_{p} \in \mathbb{R}^{N \times d}$ from a $(k \times k)$ patch window using strided convolutions, where $N = HW / k^2$ (Fig. 3). Positional embeddings are then added to these patch tokens $X_{p}$ to retain spatial information, followed by the application of successive cross-attention mechanisms. For classification tasks, the model either introduces a class token $t_{cls} \in \mathbb{R}^{d}$ or utilizes the mean of the output patch tokens. + +The encoder employs multi-headed cross-attention (MHA) on the patch token embeddings $X_{p}$ , with each MHA block consisting of multiple linear layers and a cross-attention layer. The computational complexity of these operations suggests that reducing the number of tokens $N$ can significantly decrease computational demands. A key strength of the ViT architecture is its adaptability to varying numbers of tokens $N$ , though at a higher computational cost. Prior work [2,41,56] has primarily aimed at reducing patch tokens within intermediate encoder layers, often requiring additional training due to the variability in token representations + +![](images/3636b67499797712f50e8b886624e4e8bfefb4d7a91e409f8147d30015cc21a2.jpg) +Fig. 3: (a) In vanilla ViT, PaPr operates right after the patch extractor module. Hence, all transformer blocks can operate only with the most discriminative patches. (b) Hierarchical model blocks comprise of window based kernel operator (e.g., Convk $\times$ $k$ /local attention), followed by pixel operator (e.g., linear layer, Conv1x1). Pixel operator consumes more than $60\%$ of total computation. PaPr is used to split the foreground patches to be operated with pixel operator. Background patches are zero-ed out, and finally, re-assembled with foreground output patches. + +across different architectures. These methods typically achieve only gradual token reduction, limited by the distributed nature of information across tokens. + +In contrast, our approach seeks to eliminate redundant tokens early in the processing pipeline, immediately following initial patch extraction. This strategy offers multiple advantages: (1) Leveraging the convolutional nature of the initial patch extraction allows for a direct mapping between our PSM $\mathcal{P}$ and the patch tokens $X_{p}$ , obviating the need for additional mechanisms to determine token redundancy. (2) It facilitates generalizability across transformer models without the need for retraining, potentially accelerating training by concentrating on essential patches. (3) It permits seamless integration of existing intermediate layer token reduction techniques following our initial patch pruning process. + +# 3.4 Integrating PSM with Hierarchical Models + +ConvNets and hybrid transformers, such as Swin [31], primarily operate with window-based local operations in contrast to full-attention based operation in vanilla transformers. Such window-based local processing makes patch pruning particularly complicated, in contrast to vanilla transformers. However, these models maintain the location property of representative image patches all through the network, which leaves the door open to prune redundant patches. Nevertheless, the window-based operation is particularly difficult to prune. + +In general, each block of hierarchical layers consists of window-based spatial kernel/attention operators, followed by pixel operators (usually, modeled with $1 \times 1$ convolutional layers or linear layers). In contrast to windowed convolutions or cross-attention, these pixel operators are particularly suitable for patch pruning. Interestingly, in the SOTA ConvNets and hybrid transformers, more than $60\%$ of total computations are performed with such pixel operators (63.3% in Swin [31], + +![](images/c427eb00d0e8b6575f9866740cad876e19f4ab819010fc3fc68aec68cebd876d.jpg) +Fig. 4: ImageNet-1k evaluation for varying top-k accuracy targets. The accuracy gain with bigger model largely shrinks, as $k$ increases. This suggests shallower ConvNets have understanding of object locations and visual property, despite their lower top-1 accuracy. + +![](images/8780e33834425319c90bbbf763f15d7615eff67316943ce4be7a9aec94d968a4.jpg) +Fig.5: Integrating PaPr with ToMe [2]. We use the Augreg pretrained ViT-B-16 architecture as the baseline. We sweep token merging ratio (r) for different pruning ratio (z). Integration of PaPr achieves Pareto-optimal performance, thus, PaPr can enhance existing patch reduction methods. + +$96.2\%$ in ConvNext [32]). Based on the patch significance map from PaPr, we only use the pixel operator on most significant patches, and simply perform zero-padding on the remaining patch regions. The zero padding operations are performed to mostly recover the feature map shape to be used with subsequent spatial operators. Hence, the speed-up is mostly achieved by eliminating redundant computations in pixel operators (See Fig. 3). + +Let's assume that the feature map before applying pixel operators is given by $\mathcal{A}_w$ . Based on the PSM $\mathcal{P}$ , we initially identify the foreground and background pixel features as $A_{f}$ , and $A_{b}$ , respectively. Later, the pixel operator layers are applied on foreground pixels $A_{f}$ , and a zero-padded representation is used for the background pixels $A_{b}$ . Finally, the output representation $\mathcal{A}_c$ is reassembled with modulated foreground pixels $A_{f}$ and zero-padded background pixels $A_{z}$ , given by Eq. (4): + +$$ +A _ {f}, A _ {b} = \operatorname {S p l i t} \left(\mathcal {A} _ {w}, \mathcal {P}\right), +$$ + +$$ +A _ {z} = \operatorname {Z e r o s} \left(A _ {b}\right), A _ {f} = \operatorname {L i n e a r} \left(A _ {f}\right), \tag {4} +$$ + +$$ +\mathcal {A} _ {c} = \operatorname {R e a s s e m b l e} \left(A _ {f}, A _ {z}\right). +$$ + +# 4 Image Experiments + +# 4.1 Experimental Setup + +We experiment on image classification task on the ImageNet-1k benchmark dataset, following prior work [2, 12, 23, 34, 55]. We use the MobileOne-s0 [49] model as the proposal model for all architectures, unless otherwise specified. We report training-free results of PaPr, unless otherwise specified. For the performance metric, we report top-1 accuracy, GFLOPs, and the throughput (img/s). Throughput is measured on a single RTX-A5000 GPU with 24GB VRAM. We reproduced all baseline models under the same setup for a fair comparison. + +Table 1: Performance comparison on AugReg models. PaPr achieves the best performance, while operating with token merging. $z$ denotes patch keeping ratio in PaPr. + +
ModelsMethodsAcc1GFLOPsImg/s
ViT-S-16Baseline81.394.61975
ToMe [2]76.962.291978
TokenFusion [23]77.122.291982
GTP-ViT [55]71.032.311970
PaPr (z=0.45)76.212.291988
PaPr (z=0.55) (+ToMe)77.762.192073
PaPr (z=0.5) (+ToMe)76.271.972315
ViT-B-16Baseline84.5917.59307
ToMe [2]80.388.78615
TokenFusion [23]80.78.78618
GTP-ViT [55]80.988.78610
PaPr (z=0.5)82.118.98605
PaPr (z=0.55) (+ToMe)82.348.21660
PaPr (z=0.5) (+ToMe)80.886.82785
ViT-L-16Baseline85.8261.6191
ToMe [2]83.530.99180
TokenFusion [23]83.9130.99182
GTP-ViT [55]81.5630.99181
PaPr (z=0.5)83.8730.83183
PaPr (z=0.55) (+ToMe)83.9926.33210
PaPr (z=0.55) (+ToMe)83.525.12224
+ +Table 2: Performance comparison on MAE models. Since MAE uses masked pretraining, PaPr is particularly suitable for MAE inference. PaPr achieves significantly higher performance than others. $z$ denotes patch keeping ratio in PaPr. + +
ModelsMethodsAcc1GFLOPsImg/s
ViT-B-16Baseline83.7417.59307
ToMe [2]78.828.78615
TokenFusion [23]79.238.78618
GTP-ViT [55]79.148.78610
PaPr (z=0.5)82.48.98605
PaPr (z=0.4)81.47.72700
ViT-L-16Baseline85.9561.6191
ToMe [2]84.2430.99180
TokenFusion [23]84.3330.99182
GTP-ViT [55]84.1530.99181
PaPr (z=0.5)85.0630.83183
PaPr (z=0.4)84.7627.72201
ViT-H-16Baseline86.89167.436
ToMe [2]85.4882.5372
TokenFusion [23]85.7182.5373
GTP-ViT [55]85.5482.5371
PaPr (z=0.5)86.483.0471
PaPr (z=0.4)86.1374.5981
+ +# 4.2 Performance on Various Vision Transformers + +We study performance of PaPr in diverse ViT architectures, along with various pre-training methods. We also present training results, along with training-free results to compare with existing methods. + +Training-free method comparison. We consider two different pre-training methods for comparative analysis on various ViT architectures. We use supervised pretrained Augreg models [45], and self-supervised pre-trained masked autoencoder (MAE) models [17], following prior work [2]. We compare with three recent training-free patch reduction methods, such as ToMe [2], TokenFusion [23], and GTP-ViT [55]. Additionally, we show the performance comparison on class-token free ViT models to highlight the architecture agnostic performance. In general, PaPr achieves better accuracy, with lower computational costs compared to state-of-the-art (SOTA) methods. + +Augreg models: For AugReg pre-training with ViT-Base, PaPr achieves $2.14\%$ higher accuracy over ToMe, with comparable FLOPs as shown in Tab. 1. In addition, by combining PaPr with ToMe, we achieve additional $22.3\%$ FLOP reduction, while maintaining higher accuracy. We further study the compatibility of PaPr with existing patch reduction methods as ToMe. After the initial patch pruning with PaPr, we integrate ToMe only at the bottom-half layers due to its higher sensitivity in earlier layers. As shown in Fig. 5, integration of PaPr can directly boost ToMe performance thereby achieving Pareto optimality. + +Table 3: Performance analysis on class-token free ViT models. PaPr performance gain is not limited to specific architectures. + +
ModelMethodAcc1GFLOPsImg/s
ViT-MediumBaseline84.3310.58484
GAP-16GTP-ViT [55]80.55.73880
256ToMe [2]80.215.73890
PaPr (z=0.5)81.85.5932
ViT-MediumBaseline85.626.06182
GAP-16GTP-ViT [55]81.9513.59340
384ToMe [2]82.513.59344
PaPr83.9512.94360
+ +Table 4: Training performance analysis on DeIT-s. PaPr achieves competitive performance with higher training speed. + +
MethodsAcc1GFLOPsIm/sBatch Proc.Train Speed
Baseline79.84.6960✓1x
DynamicViT [41]79.32.91510X1x
A-ViT [56]78.62.9-X1x
ATS [12]79.52.91512-1x
SP-ViT [25]79.32.6-X1x
ToMe [2]79.42.71575✓1.5x
TokenFusion [23]79.52.71580✓1.5x
PaPr (z=0.55)79.22.71585✓1.6x
+ +Table 5: Performance comparison on ConvNext CNN models. PaPr can achieve competitive performance without training, and can seamlessly adapt to model updates. + +
ModelsMethodsTrain-FreeAcc1GFLOPsImg/s
ConvNeXtBaselineN/A83.8415.38265
Base-1kDynCNN [40]✘83.0810.21375
PaPr (z=0.65)✓82.7510.42395
ConvNeXtBaselineN/A85.8115.38265
Base-22kPaPr (z=0.65)✓84.2710.42395
ConvNeXtBaselineN/A84.3134.4135
Large-1kPaPr (z=0.65)✓83.2622.9203
ConvNeXtBaselineN/A86.6134.4135
Large-22kPaPr (z=0.65)✓85.6722.9203
+ +Table 6: Performance comparison on Swin hybrid transformer models. PaPr can adapt to much bigger models with higher operating resolutions without training. + +
ModelsMethodsTrain-FreeAcc1GFLOPsImg/s
Swin B-1kBaselineN/A83.4215.47258
DynSwin [40]✗83.1812.1327
PaPr (z=0.65)✓81.712.25325
Swin B-22kBaselineN/A85.1615.47258
PaPr (z=0.65)✓82.2712.25325
Swin L-22kBaselineN/A86.2534.53135
PaPr (z=0.65)✓84.5326.96175
Swin L-22k-384BaselineN/A87.25104.0842
PaPr (z=0.65)✓86.4781.4454
+ +MAE models: MAE used self-supervised pre-training by reconstructing masked image patches to train large ViT models [17]. Later, the model is fine-tuned with full resolution images without masking, which cannot exploit its latent ability to learn from fewer patches. Interestingly, PaPr introduces masked inference, which is particularly suitable for MAE models. Hence, we observe significantly higher accuracy with PaPr compared to other training-free methods, e.g., with ViT-B, PaPr achieves $4.5\%$ higher accuracy over ToMe for similar FLOPs (see Tab. 2). + +Class token-free models: Several existing patch reduction methods operate with the class token in ViTs to evaluate the patch relevance [12,28,34]. However, instead of class tokens, global pooling of patch tokens is used in many cases [16, 37]. Results presented in Tab. 3 demonstrate the superiority of PaPr for such cases. + +Training-based methods comparison. We compare training performance of existing methods for the patch reduction on DeiT-s [48] model in Tab. 4. Prior pruning based methods cannot speed-up the training for learning the mask predictor [25,41,56]. However, PaPr achieves competitive performance as prior methods, while achieving large speed-up as other token merging methods [2,23,55]. + +Table 7: Sweeping ConvNet proposal model in PaPr with $z = 0.5$ . PaPr can achieve similar performance irrespective of proposal model size. Thus, PaPr can use a much smaller and faster proposal model to speed-up larger models. + +
ProposalAccuracy1(%)
ModelGFLOPsViT-B-16ViT-L-16
ResNet-18 [18]1.8181.183.84
ResNet-50 [18]4.0982.3384.09
ResNet-152 [18]11.5182.5184.08
MobileOne-S0 [49]0.2782.2484.06
MobileOne-S2 [49]1.3582.2884.16
MobileOne-S4 [49]2.9882.3584.32
+ +![](images/772790e3fcf8e48f2e99138562f527bb7eb88b103832ad836105ed30f82e88b5.jpg) +Fig.6: Proposal models locate similar patches with PaPr for different pruning ratio $(z)$ irrespective of model size. + +Table 8: Training-free performance comparison on Kinetics-400 video evaluation. Each video has $16 \times 224^2$ input size. PaPr achieves significantly better performance for reducing spatio-temporal redundancy in videos. + +
ModelMethodAcc1GFLOPsViews
XViTATS [12]80.02591x3
TimeSformer-LATS [12]80.535101x3
ViT-B MAEBaseline81.211803x5
PaPr(z=0.45)81.18763x5
PaPr(z=0.35)80.15593x5
ViT-L MAEBaseline85.265983x5
ToMe [2]84.52811x10
PaPr(z=0.5)85.122753x5
ToMe [2]82.51841x10
PaPr(z=0.3)84.531603x5
+ +Table 9: Proposal ConvNet sweep in video recognition with ViT-B-MAE model as the baseline. PaPr achieves competitive performance with lighter proposal models for different $z$ values. + +
ProposalKeep Ratio(z)Acc1 (%)Views
ModelGFLOPs
X3d-s [13]1.250.581.113x5
X3d-m [13]2.450.581.193x5
ResNet-50 [14]41.90.579.963x5
ResNet-101 [14]85.670.579.973x5
X3d-s [13]1.250.480.863x5
X3d-m [13]2.450.480.933x5
ResNet-50 [14]41.90.478.853x5
ResNet-101 [14]85.670.479.323x5
+ +# 4.3 Performance on Various Hierarchical Models + +We study the performance on two variants of hierarchical models, such as pure convolutional models, and hybrid transformer models. We compare with training based DynamicCNN [40], and DynamicSwin [40] methods (trained for 120 epochs), whereas PaPr operates without training. + +Analysis on convolutional models. We use the SOTA ConvNext [32] architecture for the analysis as shown in Tab. 5. The training-free operation in PaPr makes it seamlessly usable with new model updates. We analyze ImageNet-1k and ImageNet-22k performance of same models. For the ConvNext-Base model, PaPr achieves $99.6\%$ accuracy of DynamicCNN with $7.1\%$ higher throughput. By simply using ImageNet-22k weights, PaPr can achieve additional $2\%$ accuracy improvement without training, while having the same computation cost. + +![](images/fe51584d2962fd0367d1dce2c621ada35d997467a3e559bfa97aa8b08f5a3857.jpg) +Fig. 7: Robustness of PaPr compared to CAM based methods. PaPr can perform even when the ConvNet proposal confidence (c) is very low. In contrast, existing CAM based methods fail in such cases, despite being significantly slower while not enabling batch processing and use of gradients in some cases. + +Analysis on hybrid transformer models. We study the hierarchical Swin transformer models for patch reduction with PaPr, as given in Tab. 6. PaPr achieves $98\%$ of DynamicSwin accuracy using similar FLOPs without re-training. Nevertheless, by leveraging bigger models, larger pre-training, and higher resolution, PaPr can seamlessly adapt to achieve higher accuracy. + +# 4.4 Robustness of PaPr across various ConvNet proposals + +PaPr can operate with ultra-lightweight ConvNets for generating robust proposals. We study different ConvNet architectures for proposal generation, as shown in Tab. 7. Interestingly, PaPr shows small reductions of final accuracy $(0.3\%)$ , when using MobileOne-S0 based proposals compared to ResNet-152 $(42\times$ higher FLOPs) in ViT-B. When visualizing the PSM for different pruning ratio as in Fig. 6, we notice the similar PSMs irrespective of model size. Hence, PaPr can utilize ultra-lightweight ConvNets to mask patches without sacrificing accuracy. + +# 4.5 Comparison with Class Activation Mappings (CAMs) + +Existing CAM methods mostly focus on explainability to highlight image regions responsible for final prediction. However, such objectives rely on final prediction performance, which can be much lower for light ConvNets. Moreover, for localization, many of these methods use gradients, and optimization methods that limit batch processing. Nevertheless, we study the impact of such CAM methods ([43, 50, 61]) in PaPr framework, in challenging scenarios where the baseline ViT-B has higher prediction confidence (c) on target class, and the proposal + +![](images/58041a1a8ab5a9067821c623a7a89ccb36876571cfd2c6359efec62cb69c21d8.jpg) +Fig. 8: Visualization of PaPr localization in videos. Video has inherent high sparsity. PaPr effectively localizes the discriminative regions for using holistic spatio-temporal understanding with small ConvNets. Thus, it significantly reduces the computational burden for larger models in downstream video recognition tasks. + +MobileOne-S0 has lower confidence (see Fig. 7). In most cases, the baseline CAM method significantly lowers the final accuracy after patch pruning. In contrast, PaPr maintains robust confidence for its precise localization. Moreover, PaPr can enhance confidence of baseline models in several scenarios by suppressing the redundant patches, e.g., in Sample 5, the $c$ increased by $22\%$ with PaPr. + +# 5 Video Experiments + +We study the training-free performance on Kinetics-400 [22] validation sets, as shown in Tab. 8. We use SOTA ATS [12], and ToMe [2] methods for comparison. We use PaPr on SOTA ViT-MAE [47] models with lightweight X3d-s [13] model for proposal generation. We follow the baseline [47] $3 \times 5$ view (3 spatial view and 5 temporal view) approach, and separate the views from model FLOP counts similar to other works as it can be arbitrarily chosen. Since ViT-MAE removes class tokens in fine-tuning, ATS [12] cannot be adapted to these models. For ViT-L, PaPr achieves $3.7 \times$ FLOPs reduction of the baseline model for a negligible $0.8\%$ accuracy drop. We also study the impact of various proposal models on the final performance, as shown in Tab. 9. We use SlowOnly-ResNet [14] and X3d [13] models for comparison. Increasing model size for different pruning shows minimal impact on final performance. Finally, we visualize the patch masking of PaPr in videos, as shown in Fig. 8. By removing the redundant details, PaPr significantly boosts the performance of bigger models for fine-grained predictions. + +# 6 Conclusion and Future Work + +In this paper, we introduce a novel patch pruning method, namely PaPr, that can effectively speed-up off-the-shelf pre-trained models inference without re-training. We propose a simple modification of ConvNets, that allows extracting a precise discriminative PSM from ultra lightweight ConvNets with unparalleled speed. In PaPr, such PSMs are leveraged to directly reduce the redundant data regions in a single-step to guide larger model computations on most discriminative regions for achieving computational efficiency. Moreover, PaPr can be easily integrated with most existing patch reduction methods due to its simple structure. + +# Acknowledgement + +This work was supported in part by ONR Minerva program, iMAGiNE - the Intelligent Machine Engineering Consortium at UT Austin, and a UT Cockrell School of Engineering Doctoral Fellowship. + +# References + +1. Belharbi, S., Sarraf, A., Pedersoli, M., Ben Ayed, I., McCaffrey, L., Granger, E.: F-cam: Full resolution class activation maps via guided parametric upscaling. In: Proceedings of the IEEE/CVF Winter Conference on Applications of Computer Vision. pp. 3490-3499 (2022) +2. Bolya, D., Fu, C.Y., Dai, X., Zhang, P., Feichtenhofer, C., Hoffman, J.: Token merging: Your vit but faster. In: The Twelfth International Conference on Learning Representations (2023) +3. Cao, Q., Paranjape, B., Hajishirzi, H.: Pumer: Pruning and merging tokens for efficient vision language models. arXiv preprint arXiv:2305.17530 (2023) +4. Caron, M., Touvron, H., Misra, I., Jégou, H., Mairal, J., Bojanowski, P., Joulin, A.: Emerging properties in self-supervised vision transformers. In: Proceedings of the IEEE/CVF international conference on computer vision. pp. 9650-9660 (2021) +5. Chattopadhyay, A., Sarkar, A., Howlader, P., Balasubramanian, V.N.: Grad-cam++: Generalized gradient-based visual explanations for deep convolutional networks. In: 2018 IEEE winter conference on applications of computer vision (WACV). pp. 839-847. IEEE (2018) +6. Chen, Y., Liu, H., Yin, H., Fan, B.: Building vision transformers with hierarchy aware feature aggregation. In: Proceedings of the IEEE/CVF International Conference on Computer Vision (ICCV). pp. 5908-5918 (October 2023) +7. Chu, X., Tian, Z., Wang, Y., Zhang, B., Ren, H., Wei, X., Xia, H., Shen, C.: Twins: Revisiting the design of spatial attention in vision transformers. Advances in Neural Information Processing Systems 34, 9355-9366 (2021) +8. Ding, X., Zhang, X., Ma, N., Han, J., Ding, G., Sun, J.: Repvgg: Making vgg-style convnets great again. In: Proceedings of the IEEE/CVF conference on computer vision and pattern recognition. pp. 13733-13742 (2021) +9. Dong, X., Bao, J., Chen, D., Zhang, W., Yu, N., Yuan, L., Chen, D., Guo, B.: Cswin transformer: A general vision transformer backbone with cross-shaped windows. In: CVPR (2022) +0. Dosovitskiy, A., Beyer, L., Kolesnikov, A., Weissenborn, D., Zhai, X., Unterthiner, T., Dehghani, M., Minderer, M., Heigold, G., Gelly, S., et al.: An image is worth 16x16 words: Transformers for image recognition at scale. In: ICLR (2020) +1. Fan, H., Xiong, B., Mangalam, K., Li, Y., Yan, Z., Malik, J., Feichtenhofer, C.: Multiscale vision transformers. In: ICCV (2021) +2. Fayyaz, M., Kouhpayegani, S.A., Jafari, F.R., Sommerlade, E., Joze, H.R.V., Pirsivash, H., Gall, J.: Ats: Adaptive token sampling for efficient vision transformers. In: ECCV (2022) +3. Feichtenhofer, C.: X3d: Expanding architectures for efficient video recognition. In: Proceedings of the IEEE/CVF conference on computer vision and pattern recognition. pp. 203-213 (2020) +4. Feichtenhofer, C., Fan, H., Malik, J., He, K.: Slowfast networks for video recognition. In: Proceedings of the IEEE/CVF international conference on computer vision. pp. 6202-6211 (2019) + +15. Graham, B., El-Nouby, A., Touvron, H., Stock, P., Joulin, A., Jégou, H., Douze, M.: Levit: a vision transformer in convnet's clothing for faster inference. In: ICCV (2021) +16. Hatamizadeh, A., Yin, H., Heinrich, G., Kautz, J., Molchanov, P.: Global context vision transformers. In: International Conference on Machine Learning. pp. 12633-12646. PMLR (2023) +17. He, K., Chen, X., Xie, S., Li, Y., Dollar, P., Girshick, R.: Masked autoencoders are scalable vision learners. In: CVPR (2022) +18. He, K., Zhang, X., Ren, S., Sun, J.: Deep residual learning for image recognition. In: Proceedings of the IEEE conference on computer vision and pattern recognition. pp. 770-778 (2016) +19. Howard, A., Sandler, M., Chu, G., Chen, L.C., Chen, B., Tan, M., Wang, W., Zhu, Y., Pang, R., Vasudevan, V., et al.: Searching for mobilenetv3. In: Proceedings of the IEEE/CVF international conference on computer vision. pp. 1314-1324 (2019) +20. Howard, A.G., Zhu, M., Chen, B., Kalenichenko, D., Wang, W., Weyand, T., Andreetto, M., Adam, H.: Mobilenets: Efficient convolutional neural networks for mobile vision applications. arXiv preprint arXiv:1704.04861 (2017) +21. Jung, H., Oh, Y.: Towards better explanations of class activation mapping. In: Proceedings of the IEEE/CVF International Conference on Computer Vision. pp. 1336-1344 (2021) +22. Kay, W., Carreira, J., Simonyan, K., Zhang, B., Hillier, C., Vijayanarasimhan, S., Viola, F., Green, T., Back, T., Natsev, P., et al.: The kinetics human action video dataset. arXiv:1705.06950 [cs.CV] (2017) +23. Kim, M., Gao, S., Hsu, Y.C., Shen, Y., Jin, H.: Token fusion: Bridging the gap between token pruning and token merging. In: Proceedings of the IEEE/CVF Winter Conference on Applications of Computer Vision. pp. 1383-1392 (2024) +24. Kim, S., Shen, S., Thorsley, D., Gholami, A., Kwon, W., Hassoun, J., Keutzer, K.: Learned token pruning for transformers. arXiv:2107.00910 [cs.CL] (2021) +25. Kong, Z., Dong, P., Ma, X., Meng, X., Niu, W., Sun, M., Ren, B., Qin, M., Tang, H., Wang, Y.: Spvit: Enabling faster vision transformers via soft token pruning. In: ECCV (2022) +26. Lassance, C., Maachou, M., Park, J., Clinchant, S.: A study on token pruning for colbert. arXiv:2112.06540 [cs.CL] (2021) +27. Li, Y., Wu, C.Y., Fan, H., Mangalam, K., Xiong, B., Malik, J., Feichtenhofer, C.: Mvittv2: Improved multiscale vision transformers for classification and detection. In: CVPR (2022) +28. Liang, Y., Ge, C., Tong, Z., Song, Y., Wang, J., Xie, P.: Not all patches are what you need: Expediting vision transformers via token reorganizations. ICLR (2022) +29. Lin, J., Chen, W.M., Cai, H., Gan, C., Han, S.: Memory-efficient patch-based inference for tiny deep learning. Advances in Neural Information Processing Systems 34, 2346-2358 (2021) +30. Liu, Z., Hu, H., Lin, Y., Yao, Z., Xie, Z., Wei, Y., Ning, J., Cao, Y., Zhang, Z., Dong, L., et al.: Swin transformer v2: Scaling up capacity and resolution. In: CVPR (2022) +31. Liu, Z., Lin, Y., Cao, Y., Hu, H., Wei, Y., Zhang, Z., Lin, S., Guo, B.: Swin transformer: Hierarchical vision transformer using shifted windows. In: ICCV (2021) +32. Liu, Z., Mao, H., Wu, C.Y., Feichtenhofer, C., Darrell, T., Xie, S.: A convnet for the 2020s. In: Proceedings of the IEEE/CVF conference on computer vision and pattern recognition. pp. 11976-11986 (2022) + +33. Mehta, S., Rastegari, M.: Mobilevit: light-weight, general-purpose, and mobile-friendly vision transformer. In: ICLR (2021) +34. Meng, L., Li, H., Chen, B.C., Lan, S., Wu, Z., Jiang, Y.G., Lim, S.N.: Adavit: Adaptive vision transformers for efficient image recognition. In: CVPR (2022) +35. Muhammad, M.B., Yeasin, M.: Eigen-cam: Class activation map using principal components. In: 2020 international joint conference on neural networks (IJCNN). pp. 1-7. IEEE (2020) +36. Oquab, M., Darcet, T., Moutakanni, T., Vo, H., Szafraniec, M., Khalidov, V., Fernandez, P., Haziza, D., Massa, F., El-Nouby, A., et al.: Dinov2: Learning robust visual features without supervision. arXiv preprint arXiv:2304.07193 (2023) +37. Pan, Z., Zhuang, B., Liu, J., He, H., Cai, J.: Scalable vision transformers with hierarchical pooling. In: Proceedings of the IEEE/cvf international conference on computer vision. pp. 377-386 (2021) +38. Patro, B.N., Lunayach, M., Patel, S., Namboodiri, V.P.: U-cam: Visual explanation using uncertainty based class activation maps. In: Proceedings of the IEEE/CVF International Conference on Computer Vision. pp. 7444-7453 (2019) +39. Radford, A., Kim, J.W., Hallacy, C., Ramesh, A., Goh, G., Agarwal, S., Sastry, G., Askell, A., Mishkin, P., Clark, J., et al.: Learning transferable visual models from natural language supervision. In: International conference on machine learning. pp. 8748-8763. PMLR (2021) +40. Rao, Y., Liu, Z., Zhao, W., Zhou, J., Lu, J.: Dynamic spatial sparsification for efficient vision transformers and convolutional neural networks. IEEE Transactions on Pattern Analysis and Machine Intelligence (2023) +41. Rao, Y., Zhao, W., Liu, B., Lu, J., Zhou, J., Hsieh, C.J.: Dynamicvit: Efficient vision transformers with dynamic token sparsification. Advances in neural information processing systems 34, 13937-13949 (2021) +42. Sandler, M., Howard, A., Zhu, M., Zhmoginov, A., Chen, L.C.: Mobilenetv2: Inverted residuals and linear bottlenecks. In: Proceedings of the IEEE conference on computer vision and pattern recognition. pp. 4510-4520 (2018) +43. Selvaraju, R.R., Cogswell, M., Das, A., Vedantam, R., Parikh, D., Batra, D.: Gradcam: Visual explanations from deep networks via gradient-based localization. In: ICCV (2017) +44. Singh, M., Gustafson, L., Adcock, A., de Freitas Reis, V., Gedik, B., Kosaraju, R.P., Mahajan, D., Girshick, R., Dollar, P., van der Maaten, L.: Revisiting weakly supervised pre-training of visual perception models. In: CVPR (2022) +45. Steiner, A., Kolesnikov, A., Zhai, X., Wightman, R., Uszkoreit, J., Beyer, L.: How to train your vit? data, augmentation, and regularization in vision transformers. TMLR (2022) +46. Tan, M., Le, Q.: Efficientnet: Rethinking model scaling for convolutional neural networks. In: ICML (2019) +47. Tong, Z., Song, Y., Wang, J., Wang, L.: Videomae: Masked autoencoders are data-efficient learners for self-supervised video pre-training. arXiv:2203.12602 [cs.CV] (2022) +48. Touvron, H., Cord, M., Douze, M., Massa, F., Sablayrolles, A., Jégou, H.: Training data-efficient image transformers & distillation through attention. In: ICML (2021) +49. Vasu, P.K.A., Gabriel, J., Zhu, J., Tuzel, O., Ranjan, A.: Mobileone: An improved one millisecond mobile backbone. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 7907-7917 (2023) +50. Wang, H., Wang, Z., Du, M., Yang, F., Zhang, Z., Ding, S., Mardziel, P., Hu, X.: Score-cam: Score-weighted visual explanations for convolutional neural networks. + +In: Proceedings of the IEEE/CVF conference on computer vision and pattern recognition workshops. pp. 24-25 (2020) +51. Wang, S., Li, B.Z., Khabsa, M., Fang, H., Ma, H.: Linformer: Self-attention with linear complexity. arXiv preprint arXiv:2006.04768 (2020) +52. Wei, C., Fan, H., Xie, S., Wu, C.Y., Yuille, A., Feichtenhofer, C.: Masked feature prediction for self-supervised visual pre-training. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 14668-14678 (2022) +53. Woo, S., Debnath, S., Hu, R., Chen, X., Liu, Z., Kweon, I.S., Xie, S.: Convnext v2: Co-designing and scaling convnets with masked autoencoders. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 16133-16142 (2023) +54. Wu, H., Xiao, B., Codella, N., Liu, M., Dai, X., Yuan, L., Zhang, L.: Cvt: Introducing convolutions to vision transformers. In: Proceedings of the IEEE/CVF international conference on computer vision. pp. 22-31 (2021) +55. Xu, X., Wang, S., Chen, Y., Zheng, Y., Wei, Z., Liu, J.: Gtp-vit: Efficient vision transformers via graph-based token propagation. In: Proceedings of the IEEE/CVF Winter Conference on Applications of Computer Vision. pp. 86-95 (2024) +56. Yin, H., Vahdat, A., Alvarez, J.M., Mallya, A., Kautz, J., Molchanov, P.: A-vit: Adaptive tokens for efficient vision transformer. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 10809-10818 (2022) +57. You, H., Xiong, Y., Dai, X., Wu, B., Zhang, P., Fan, H., Vajda, P., Lin, Y.C.: Castling-vit: Compressing self-attention via switching towards linear-angular attention at vision transformer inference. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 14431-14442 (2023) +58. Yu, H., Wu, J.: A unified pruning framework for vision transformers. arXiv:2111.15127 [cs.CV] (2021) +59. Zhang, X., Zhou, X., Lin, M., Sun, J.: Shufflenet: An extremely efficient convolutional neural network for mobile devices. In: Proceedings of the IEEE conference on computer vision and pattern recognition. pp. 6848-6856 (2018) +60. Zhou, A., Li, Y., Qin, Z., Liu, J., Pan, J., Zhang, R., Zhao, R., Gao, P., Li, H.: Sparsemae: Sparse training meets masked autoencoders. In: Proceedings of the IEEE/CVF International Conference on Computer Vision. pp. 16176-16186 (2023) +61. Zhou, B., Khosla, A., Lapedriza, A., Oliva, A., Torralba, A.: Learning deep features for discriminative localization. In: Proceedings of the IEEE conference on computer vision and pattern recognition. pp. 2921-2929 (2016) \ No newline at end of file diff --git a/2024/PaPr_ Training-Free One-Step Patch Pruning with Lightweight ConvNets for Faster Inference/images.zip b/2024/PaPr_ Training-Free One-Step Patch Pruning with Lightweight ConvNets for Faster Inference/images.zip new file mode 100644 index 0000000000000000000000000000000000000000..7a47ef7efa7ea048e624098be29efac1c35deca7 --- /dev/null +++ b/2024/PaPr_ Training-Free One-Step Patch Pruning with Lightweight ConvNets for Faster Inference/images.zip @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:cbfaa6092fc02cedc1e8348872b97b34beac8cc58a6f5f29f354003b963ba839 +size 596199 diff --git a/2024/PaPr_ Training-Free One-Step Patch Pruning with Lightweight ConvNets for Faster Inference/layout.json b/2024/PaPr_ Training-Free One-Step Patch Pruning with Lightweight ConvNets for Faster Inference/layout.json new file mode 100644 index 0000000000000000000000000000000000000000..d6cbb4516272b1d69a04e34e2ccf3efe3317b6b3 --- /dev/null +++ b/2024/PaPr_ Training-Free One-Step Patch Pruning with Lightweight ConvNets for Faster Inference/layout.json @@ -0,0 +1,10024 @@ +{ + "pdf_info": [ + { + "para_blocks": [ + { + "bbox": [ + 134, + 111, + 480, + 147 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 134, + 111, + 480, + 147 + ], + "spans": [ + { + "bbox": [ + 134, + 111, + 480, + 147 + ], + "type": "text", + "content": "PaPr: Training-Free One-Step Patch Pruning with Lightweight ConvNets for Faster Inference" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 136, + 167, + 476, + 179 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 136, + 167, + 476, + 179 + ], + "spans": [ + { + "bbox": [ + 136, + 167, + 476, + 179 + ], + "type": "text", + "content": "Tanvir Mahmud1 Burhaneddin Yaman2 Chun-Hao Liu3 Diana Marculescu1" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 146, + 189, + 466, + 212 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 146, + 189, + 466, + 212 + ], + "spans": [ + { + "bbox": [ + 146, + 189, + 466, + 212 + ], + "type": "text", + "content": "1University of Texas at Austin 2Bosch Research 3Amazon Prime Video 1{tanvirmahmud, dianam}@utexas.edu 2yaman013@umn.edu 3liuch37@gmail.com" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 160, + 238, + 452, + 490 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 160, + 238, + 452, + 490 + ], + "spans": [ + { + "bbox": [ + 160, + 238, + 452, + 490 + ], + "type": "text", + "content": "Abstract. As deep neural networks evolve from convolutional neural networks (ConvNets) to advanced vision transformers (ViTs), there is an increased need to eliminate redundant data for faster processing without compromising accuracy. Previous methods are often architecture-specific or necessitate re-training, restricting their applicability with frequent model updates. To solve this, we first introduce a novel property of lightweight ConvNets: their ability to identify key discriminative patch regions in images, irrespective of model's final accuracy or size. We demonstrate that fully-connected layers are the primary bottleneck for ConvNets performance, and their suppression with simple weight recalibration markedly enhances discriminative patch localization performance. Using this insight, we introduce PaPr, a method for substantially pruning redundant patches with minimal accuracy loss using lightweight ConvNets across a variety of deep learning architectures, including ViTs, ConvNets, and hybrid transformers, without any re-training. Moreover, the simple early-stage one-step patch pruning with PaPr enhances existing patch reduction methods. Through extensive testing on diverse architectures, PaPr achieves significantly higher accuracy over state-of-the-art patch reduction methods with similar FLOP count reduction. More specifically, PaPr reduces about " + }, + { + "bbox": [ + 160, + 238, + 452, + 490 + ], + "type": "inline_equation", + "content": "70\\%" + }, + { + "bbox": [ + 160, + 238, + 452, + 490 + ], + "type": "text", + "content": " of redundant patches in videos with less than " + }, + { + "bbox": [ + 160, + 238, + 452, + 490 + ], + "type": "inline_equation", + "content": "0.8\\%" + }, + { + "bbox": [ + 160, + 238, + 452, + 490 + ], + "type": "text", + "content": " drop in accuracy, and up to " + }, + { + "bbox": [ + 160, + 238, + 452, + 490 + ], + "type": "inline_equation", + "content": "3.7\\times" + }, + { + "bbox": [ + 160, + 238, + 452, + 490 + ], + "type": "text", + "content": " FLOPs reduction, which is a " + }, + { + "bbox": [ + 160, + 238, + 452, + 490 + ], + "type": "inline_equation", + "content": "15\\%" + }, + { + "bbox": [ + 160, + 238, + 452, + 490 + ], + "type": "text", + "content": " more reduction with " + }, + { + "bbox": [ + 160, + 238, + 452, + 490 + ], + "type": "inline_equation", + "content": "2.5\\%" + }, + { + "bbox": [ + 160, + 238, + 452, + 490 + ], + "type": "text", + "content": " higher accuracy. Code is available at https://github.com/tanvir-utexas/PaPr." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 133, + 510, + 229, + 522 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 510, + 229, + 522 + ], + "spans": [ + { + "bbox": [ + 133, + 510, + 229, + 522 + ], + "type": "text", + "content": "1 Introduction" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 133, + 533, + 482, + 666 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 533, + 482, + 666 + ], + "spans": [ + { + "bbox": [ + 133, + 533, + 482, + 666 + ], + "type": "text", + "content": "Deep neural networks have grown from simple convolutional neural networks (ConvNets) to complex transformer models, aiming for better accuracy with more computation [10, 18]. Vision transformers (ViTs) excel by focusing on important parts of images with long-range attention and large-scale pre-training [17,39]. This success, however, comes at the cost of increased computational cost. Operating these large networks efficiently for downstream applications is crucial. Most visual tasks demand precisely pinpointing key image regions against complex backgrounds. Higher image resolution improves accuracy by capturing more details but also adds unnecessary background processing, burdening large models [44]. This highlights the need to cut down on redundant data in high-resolution images to maintain both speed and performance with advanced techniques [2,34]." + } + ] + } + ], + "index": 5 + } + ], + "discarded_blocks": [], + "page_size": [ + 612, + 792 + ], + "page_idx": 0 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 160, + 114, + 455, + 231 + ], + "blocks": [ + { + "bbox": [ + 160, + 114, + 455, + 231 + ], + "lines": [ + { + "bbox": [ + 160, + 114, + 455, + 231 + ], + "spans": [ + { + "bbox": [ + 160, + 114, + 455, + 231 + ], + "type": "image", + "image_path": "2ebd885cf12a80b9567181d16380666f4a52e1d5d91ad631a18b3dd54be74dd6.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 268, + 233, + 332, + 243 + ], + "lines": [ + { + "bbox": [ + 268, + 233, + 332, + 243 + ], + "spans": [ + { + "bbox": [ + 268, + 233, + 332, + 243 + ], + "type": "text", + "content": "(b) Proposed PaPr" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 130, + 247, + 482, + 291 + ], + "lines": [ + { + "bbox": [ + 130, + 247, + 482, + 291 + ], + "spans": [ + { + "bbox": [ + 130, + 247, + 482, + 291 + ], + "type": "text", + "content": "Fig. 1: (a) Existing patch pruning methods gradually reduce patches over the model. This requires additional training of mask generators in intermediate layers. (b) Proposed PaPr directly prunes redundant patches early in the network by leveraging pretrained lightweight ConvNets and directly speeds-up off-the-shelf models without re-training." + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "image_caption" + } + ], + "index": 2 + }, + { + "bbox": [ + 130, + 316, + 482, + 508 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 316, + 482, + 508 + ], + "spans": [ + { + "bbox": [ + 130, + 316, + 482, + 508 + ], + "type": "text", + "content": "Identifying key image regions demands a comprehensive understanding of the image and model operations. Incorrect key region estimation can impair pretraining performance by eliminating crucial areas. Traditional approaches [2,25,28, 34,56] for pruning redundant patch regions are hindered by three main limitations: (1) They often necessitate complex training of extra modules, which becomes increasingly difficult as baseline models evolve with more data, enhanced training methods, and deeper structures [25,41,56]. Re-training these modules for each model update is impractical. (2) Without a complete image understanding, these methods incrementally prune patches across the network, leading to unnecessary computations in early layers, which is particularly problematic for deeper models. (3) Many rely on specific architectural features for patch reduction, such as class tokens or attention maps, limiting their use to a narrow set of network designs [12,28]. Hence, there is a pressing need for a patch pruning solution that is adaptable to various modern architectures without additional training, can eliminate redundant patches in a single-step, thereby making it suitable for a broad spectrum of networks while streamlining each model update (See Fig. 1)." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 130, + 509, + 483, + 640 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 509, + 483, + 640 + ], + "spans": [ + { + "bbox": [ + 130, + 509, + 483, + 640 + ], + "type": "text", + "content": "Recent work on patch reduction [2, 23, 55] focuses mostly on transformers rather than ConvNets due to their impressive performance on various tasks, their aforementioned limitations notwithstanding. While ConvNets achieve lower ImageNet-1k top-1 accuracy than large ViTs (68.7% in MobileOne-S0 [49] vs. 88.7% in ViT-Huge [44]), they exhibit a remarkable ability to efficiently process the key image regions with hierarchical inductive bias. Our empirical investigation reveals that, as we broaden the evaluation metric (increasing " + }, + { + "bbox": [ + 130, + 509, + 483, + 640 + ], + "type": "inline_equation", + "content": "k" + }, + { + "bbox": [ + 130, + 509, + 483, + 640 + ], + "type": "text", + "content": " in top- " + }, + { + "bbox": [ + 130, + 509, + 483, + 640 + ], + "type": "inline_equation", + "content": "k" + }, + { + "bbox": [ + 130, + 509, + 483, + 640 + ], + "type": "text", + "content": " evaluations, see Fig. 4), the benefits of deeper models diminish, especially with a large number of image classes (e.g., 1000 in ImageNet). This suggests that shallower models excel at identifying discriminative areas as their bigger counterparts, rendering them ideal for patch pruning." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 130, + 641, + 482, + 666 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 641, + 482, + 666 + ], + "spans": [ + { + "bbox": [ + 130, + 641, + 482, + 666 + ], + "type": "text", + "content": "Leveraging this insight, we propose PaPr, a novel Patch Pruning method that employs pretrained lightweight ConvNets for efficient, one-step patch pruning in" + } + ] + } + ], + "index": 7 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "spans": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "type": "text", + "content": "2" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 166, + 91, + 240, + 101 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 166, + 91, + 240, + 101 + ], + "spans": [ + { + "bbox": [ + 166, + 91, + 240, + 101 + ], + "type": "text", + "content": "T. Mahmud et al." + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 1 + }, + { + "para_blocks": [ + { + "bbox": [ + 130, + 116, + 482, + 212 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 116, + 482, + 212 + ], + "spans": [ + { + "bbox": [ + 130, + 116, + 482, + 212 + ], + "type": "text", + "content": "a wide variety of deep learning models, maintaining accuracy while significantly cutting computational demands. Our findings show lightweight ConvNets have remarkable ability in identifying discriminative image regions but struggle in fine-grained prediction. To address this, PaPr relies on Patch Significance Maps (PSMs), which are generated using only the convolutional layers of ConvNets through uniform class weight recalibration in FC layers. Astonishingly, PSMs consistently highlight critical image regions across ConvNets of varying sizes and accuracies (Fig. 6), thereby amplifying the efficacy of ultra lightweight ConvNets." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 130, + 213, + 482, + 380 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 213, + 482, + 380 + ], + "spans": [ + { + "bbox": [ + 130, + 213, + 482, + 380 + ], + "type": "text", + "content": "ConvNets inherently preserve the positional property of patches for their inductive bias, while ViTs use cross-attention to blend patch features, leading to variability in patch relevance across models. Unlike previous methods that prune redundant patches gradually over multiple steps across intermediate layers of ViTs [2, 23, 41], our approach, PaPr, simplifies this process by eliminating non-essential patches at once, immediately after extraction (see Fig. 3), by leveraging lightweight ConvNets to assess patch significance. This direct, one-step pruning approach significantly cuts computational demands and is also compatible with other patch reduction techniques (see Fig. 5). PaPr's ability to separate crucial patch identification from fine-grained class prediction enhances a wide range of pre-trained models without further training, ensuring high accuracy with notable speed ups. By bypassing the complex training required for conventional patch selectors, PaPr capitalizes on the comprehensive capabilities of minimalist ConvNets for efficient patch pruning in larger models." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 130, + 383, + 482, + 550 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 383, + 482, + 550 + ], + "spans": [ + { + "bbox": [ + 130, + 383, + 482, + 550 + ], + "type": "text", + "content": "Our experiments demonstrate PaPr's effectiveness across various architectures and pre-training methods, achieving significant reduction in redundant patches for ViTs, large-scale ConvNets (e.g., ConvNext [32]), and hybrid transformers (e.g., Swin [31]), outperforming state-of-the-art (SOTA) patch reduction methods by a large margin. Notably, PaPr can be easily integrated with most existing patch reduction methods to reduce patches early in operation. More specifically, PaPr boosts ToMe [2] accuracy by " + }, + { + "bbox": [ + 130, + 383, + 482, + 550 + ], + "type": "inline_equation", + "content": "4.5\\%" + }, + { + "bbox": [ + 130, + 383, + 482, + 550 + ], + "type": "text", + "content": " for a computational budget similar to ViT-B. Moreover, PaPr can accelerate training akin to token merging techniques, unlike most patch pruning methods that fail to boost training speed [34, 41, 56]. PaPr shows robust patch localization performance with ultra lightweight ConvNets (" + }, + { + "bbox": [ + 130, + 383, + 482, + 550 + ], + "type": "inline_equation", + "content": "<0.3\\%" + }, + { + "bbox": [ + 130, + 383, + 482, + 550 + ], + "type": "text", + "content": " accuracy loss) for " + }, + { + "bbox": [ + 130, + 383, + 482, + 550 + ], + "type": "inline_equation", + "content": "42\\times" + }, + { + "bbox": [ + 130, + 383, + 482, + 550 + ], + "type": "text", + "content": " reduction in proposal FLOPs, thereby enabling its use for larger off-the-shelf models. Remarkably, in video recognition, PaPr cuts down around " + }, + { + "bbox": [ + 130, + 383, + 482, + 550 + ], + "type": "inline_equation", + "content": "70\\%" + }, + { + "bbox": [ + 130, + 383, + 482, + 550 + ], + "type": "text", + "content": " of redundant patches, resulting in up to " + }, + { + "bbox": [ + 130, + 383, + 482, + 550 + ], + "type": "inline_equation", + "content": "3.7\\times" + }, + { + "bbox": [ + 130, + 383, + 482, + 550 + ], + "type": "text", + "content": " FLOPs reduction with minimal impact on accuracy (" + }, + { + "bbox": [ + 130, + 383, + 482, + 550 + ], + "type": "inline_equation", + "content": "\\approx 0.8\\%" + }, + { + "bbox": [ + 130, + 383, + 482, + 550 + ], + "type": "text", + "content": ")." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 147, + 552, + 364, + 562 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 147, + 552, + 364, + 562 + ], + "spans": [ + { + "bbox": [ + 147, + 552, + 364, + 562 + ], + "type": "text", + "content": "We summarize our main contributions as follows:" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 137, + 578, + 480, + 664 + ], + "type": "list", + "angle": 0, + "index": 9, + "blocks": [ + { + "bbox": [ + 138, + 578, + 479, + 613 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 578, + 479, + 613 + ], + "spans": [ + { + "bbox": [ + 138, + 578, + 479, + 613 + ], + "type": "text", + "content": "- We propose PaPr, a novel background patch pruning method that can seamlessly operate with ViTs, ConvNets, and hybrid transformers, without further training while leveraging batch processing." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 137, + 616, + 479, + 638 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 137, + 616, + 479, + 638 + ], + "spans": [ + { + "bbox": [ + 137, + 616, + 479, + 638 + ], + "type": "text", + "content": "- We propose a simple weight recalibration method in ConvNets to precisely and efficiently locate discriminative patches, irrespective of model size." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 137, + 641, + 480, + 664 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 137, + 641, + 480, + 664 + ], + "spans": [ + { + "bbox": [ + 137, + 641, + 480, + 664 + ], + "type": "text", + "content": "- We facilitate the use of ultra-lightweight ConvNets to speed-up large models, such as ViTs, with a seamless framework and negligible accuracy loss." + } + ] + } + ], + "index": 8 + } + ], + "sub_type": "text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 343, + 91, + 447, + 101 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 343, + 91, + 447, + 101 + ], + "spans": [ + { + "bbox": [ + 343, + 91, + 447, + 101 + ], + "type": "text", + "content": "PaPr for Faster Inference" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 474, + 92, + 481, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 474, + 92, + 481, + 100 + ], + "spans": [ + { + "bbox": [ + 474, + 92, + 481, + 100 + ], + "type": "text", + "content": "3" + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 2 + }, + { + "para_blocks": [ + { + "bbox": [ + 137, + 116, + 482, + 140 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 137, + 116, + 482, + 140 + ], + "spans": [ + { + "bbox": [ + 137, + 116, + 482, + 140 + ], + "type": "text", + "content": "- We present extensive qualitative and quantitative results across numerous model architectures in both image and video applications." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 132, + 156, + 238, + 169 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 156, + 238, + 169 + ], + "spans": [ + { + "bbox": [ + 132, + 156, + 238, + 169 + ], + "type": "text", + "content": "2 Related Work" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 132, + 178, + 360, + 190 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 178, + 360, + 190 + ], + "spans": [ + { + "bbox": [ + 132, + 178, + 360, + 190 + ], + "type": "text", + "content": "2.1 From ConvNets to Vision Transformers" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 130, + 197, + 482, + 363 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 197, + 482, + 363 + ], + "spans": [ + { + "bbox": [ + 130, + 197, + 482, + 363 + ], + "type": "text", + "content": "ConvNets have been pivotal in computer vision, offering computational efficiency through kernel reuse and localization [8, 18-20, 29, 42, 46, 49, 59]. Yet, they struggle in capturing long-range dependencies, a gap bridged by ViT [10]. Inspired by the success of transformers from natural language processing (NLP), ViTs excel in long-range feature modeling through cross-attention, outperforming ConvNets albeit with higher computational demands and extensive training data requirements [7, 48, 51, 57]. Efforts to alleviate these issues include self-supervised [17, 47, 52, 60] and weakly-supervised [39, 44] pre-training, although challenges in computational complexity and optimization remain. Hybrid architectures, merging ConvNets' inductive biases with ViTs' cross-attention, offer a balanced solution by reducing computational load while maintaining performance [6, 9, 11, 15, 27, 30, 31, 33, 54]. Recent ConvNets advancements [32, 53], with improved training and large-scale data, now rival ViTs, questioning whether architectural innovations or enhanced training primarily drive performance gains." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 130, + 364, + 483, + 400 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 364, + 483, + 400 + ], + "spans": [ + { + "bbox": [ + 130, + 364, + 483, + 400 + ], + "type": "text", + "content": "Addressing the computational demands of these models, we focus on enhancing operational efficiency by pruning redundant patches without architectural modifications or re-training, maintaining performance while streamlining updates." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 132, + 415, + 449, + 427 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 415, + 449, + 427 + ], + "spans": [ + { + "bbox": [ + 132, + 415, + 449, + 427 + ], + "type": "text", + "content": "2.2 Class Activation Mapping for Explainable Deep Learning" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 130, + 433, + 482, + 576 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 433, + 482, + 576 + ], + "spans": [ + { + "bbox": [ + 130, + 433, + 482, + 576 + ], + "type": "text", + "content": "Class Activation Mapping (CAM) techniques provide explainable visual reasoning for neural network predictions by highlighting activation regions1 crucial for decisions [1,21,38,43,61]. While ConvNets utilize convolutional operations to preserve spatial information, making them apt for such visualizations, ViTs link attention weights to class tokens for prediction [4,36]. However, CAM's reliance on accurate class predictions for effective feature localization is a significant drawback, as ConvNets' lower accuracy when compared to ViTs may lead to incorrect object localization. Despite ViTs' superior fine-grained classification capabilities, it is unclear if they prioritize the same discriminative regions as ConvNets. Additionally, CAM approaches often require gradient tracking [5,43] or complex feature map decomposition [35], complicating batch processing and necessitating extra optimization steps." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 130, + 576, + 482, + 648 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 576, + 482, + 648 + ], + "spans": [ + { + "bbox": [ + 130, + 576, + 482, + 648 + ], + "type": "text", + "content": "Our work diverges from traditional CAM by aiming to consistently identify discriminative patches across different architectures, focusing computational resources on the most relevant areas without being constrained by class prediction accuracy. This approach enables more efficient processing by allowing larger models to concentrate on the most significant regions, identified by lighter networks, thus significantly reducing computational demands." + } + ] + } + ], + "index": 9 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "spans": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "type": "text", + "content": "4" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 166, + 91, + 240, + 101 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 166, + 91, + 240, + 101 + ], + "spans": [ + { + "bbox": [ + 166, + 91, + 240, + 101 + ], + "type": "text", + "content": "T. Mahmud et al." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 133, + 653, + 447, + 665 + ], + "type": "page_footnote", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 653, + 447, + 665 + ], + "spans": [ + { + "bbox": [ + 133, + 653, + 447, + 665 + ], + "type": "text", + "content": "1 \"Patch\" and \"region\" have been used interchangeably based on the context." + } + ] + } + ], + "index": 10 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 3 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 160, + 114, + 455, + 176 + ], + "blocks": [ + { + "bbox": [ + 160, + 114, + 455, + 176 + ], + "lines": [ + { + "bbox": [ + 160, + 114, + 455, + 176 + ], + "spans": [ + { + "bbox": [ + 160, + 114, + 455, + 176 + ], + "type": "image", + "image_path": "9ac1ccb35defa0ba0d4eae72669c9558242aed3ac197e5bc228178f70f6056c2.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 227, + 177, + 397, + 187 + ], + "lines": [ + { + "bbox": [ + 227, + 177, + 397, + 187 + ], + "spans": [ + { + "bbox": [ + 227, + 177, + 397, + 187 + ], + "type": "text", + "content": "(a) Baseline Convolutional Model Feature Flow" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_caption" + } + ], + "index": 2 + }, + { + "type": "image", + "bbox": [ + 189, + 191, + 430, + 251 + ], + "blocks": [ + { + "bbox": [ + 189, + 191, + 430, + 251 + ], + "lines": [ + { + "bbox": [ + 189, + 191, + 430, + 251 + ], + "spans": [ + { + "bbox": [ + 189, + 191, + 430, + 251 + ], + "type": "image", + "image_path": "3279da1ba742f58bf7165335ee79f9be4bb66d64ceadb6dce907d0cad97645b7.jpg" + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 239, + 257, + 380, + 266 + ], + "lines": [ + { + "bbox": [ + 239, + 257, + 380, + 266 + ], + "spans": [ + { + "bbox": [ + 239, + 257, + 380, + 266 + ], + "type": "text", + "content": "(b) Proposed Pruning Mask Extraction" + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 131, + 270, + 482, + 337 + ], + "lines": [ + { + "bbox": [ + 131, + 270, + 482, + 337 + ], + "spans": [ + { + "bbox": [ + 131, + 270, + 482, + 337 + ], + "type": "text", + "content": "Fig.2: (a) Baseline ConvNet gradually reduces the feature map to produce " + }, + { + "bbox": [ + 131, + 270, + 482, + 337 + ], + "type": "inline_equation", + "content": "\\mathcal{F} = \\{f_k(x,y)\\}_{k=1}^K" + }, + { + "bbox": [ + 131, + 270, + 482, + 337 + ], + "type": "text", + "content": ", followed by global average pooling and fully connected (FC) layers to predict " + }, + { + "bbox": [ + 131, + 270, + 482, + 337 + ], + "type": "inline_equation", + "content": "y_p" + }, + { + "bbox": [ + 131, + 270, + 482, + 337 + ], + "type": "text", + "content": ". (b) In PaPr, we operate on " + }, + { + "bbox": [ + 131, + 270, + 482, + 337 + ], + "type": "inline_equation", + "content": "\\mathcal{F}" + }, + { + "bbox": [ + 131, + 270, + 482, + 337 + ], + "type": "text", + "content": " by suppressing the FC layer. Initially, we extract pixel mean over " + }, + { + "bbox": [ + 131, + 270, + 482, + 337 + ], + "type": "inline_equation", + "content": "K" + }, + { + "bbox": [ + 131, + 270, + 482, + 337 + ], + "type": "text", + "content": " channels to produce discriminative region proposal " + }, + { + "bbox": [ + 131, + 270, + 482, + 337 + ], + "type": "inline_equation", + "content": "\\mathcal{R}" + }, + { + "bbox": [ + 131, + 270, + 482, + 337 + ], + "type": "text", + "content": ". Later, simple upsampling operation generates the patch significance map (PSM) " + }, + { + "bbox": [ + 131, + 270, + 482, + 337 + ], + "type": "inline_equation", + "content": "\\mathcal{P}" + }, + { + "bbox": [ + 131, + 270, + 482, + 337 + ], + "type": "text", + "content": " of target dimension. Finally, patch mask " + }, + { + "bbox": [ + 131, + 270, + 482, + 337 + ], + "type": "inline_equation", + "content": "\\mathcal{M}" + }, + { + "bbox": [ + 131, + 270, + 482, + 337 + ], + "type": "text", + "content": " for top " + }, + { + "bbox": [ + 131, + 270, + 482, + 337 + ], + "type": "inline_equation", + "content": "z\\%" + }, + { + "bbox": [ + 131, + 270, + 482, + 337 + ], + "type": "text", + "content": " patches is obtained from " + }, + { + "bbox": [ + 131, + 270, + 482, + 337 + ], + "type": "inline_equation", + "content": "\\mathcal{P}" + }, + { + "bbox": [ + 131, + 270, + 482, + 337 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 6, + "angle": 0, + "type": "image_caption" + } + ], + "index": 4 + }, + { + "bbox": [ + 132, + 358, + 348, + 369 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 358, + 348, + 369 + ], + "spans": [ + { + "bbox": [ + 132, + 358, + 348, + 369 + ], + "type": "text", + "content": "2.3 Patch Reduction for Faster Inference" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 130, + 375, + 482, + 530 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 375, + 482, + 530 + ], + "spans": [ + { + "bbox": [ + 130, + 375, + 482, + 530 + ], + "type": "text", + "content": "Several approaches have sought to enhance computational efficiency by reducing redundant patches in neural networks, with early strategies involving additional adapters or controllers to identify and prune less significant patches [24,26,41,58]. These methods, however, necessitate separate adapter training for each network and are slow to adapt due to the need to learn from the dynamics of other layers. In the context of ViTs, efforts have leveraged architectural features such as class tokens and attention maps for patch relevance, yet these solutions often fail to generalize across various architectures like hybrid transformers [31] or larger ConvNets [32, 53], thus limiting their applicability [12, 25, 34, 56]. Furthermore, while some have explored patch merging in ViTs and a combination of merging and pruning [2, 3, 23], these approaches lack a holistic image understanding by not considering the entire image context in their optimizations, leading to incremental and sub-optimal patch reduction." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 130, + 531, + 483, + 567 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 531, + 483, + 567 + ], + "spans": [ + { + "bbox": [ + 130, + 531, + 483, + 567 + ], + "type": "text", + "content": "Our work diverges from these traditional patch reduction methods by proposing a single-step, early-network patch removal strategy, that seamlessly integrates with any architecture without re-training." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 132, + 582, + 233, + 597 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 582, + 233, + 597 + ], + "spans": [ + { + "bbox": [ + 132, + 582, + 233, + 597 + ], + "type": "text", + "content": "3 Methodology" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 130, + 605, + 482, + 666 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 605, + 482, + 666 + ], + "spans": [ + { + "bbox": [ + 130, + 605, + 482, + 666 + ], + "type": "text", + "content": "Our methodology introduces a novel approach to discriminative patch pruning across various deep learning architectures, leveraging the innate capabilities of lightweight ConvNets. By generating a PSM, we efficiently identify and prune non-essential patches in a single step, enhancing computational efficiency without compromising accuracy. This process is universally applicable, seamlessly" + } + ] + } + ], + "index": 11 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 343, + 91, + 447, + 100 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 343, + 91, + 447, + 100 + ], + "spans": [ + { + "bbox": [ + 343, + 91, + 447, + 100 + ], + "type": "text", + "content": "PaPr for Faster Inference" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 474, + 92, + 480, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 474, + 92, + 480, + 100 + ], + "spans": [ + { + "bbox": [ + 474, + 92, + 480, + 100 + ], + "type": "text", + "content": "5" + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 4 + }, + { + "para_blocks": [ + { + "bbox": [ + 130, + 116, + 479, + 140 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 116, + 479, + 140 + ], + "spans": [ + { + "bbox": [ + 130, + 116, + 479, + 140 + ], + "type": "text", + "content": "integrating with ViTs, ConvNets, and hybrid models, thereby addressing the limitations of previous methods with a scalable, architecture-agnostic solution." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 131, + 157, + 413, + 170 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 131, + 157, + 413, + 170 + ], + "spans": [ + { + "bbox": [ + 131, + 157, + 413, + 170 + ], + "type": "text", + "content": "3.1 Extracting Discriminative Regions with ConvNets" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 130, + 176, + 482, + 283 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 176, + 482, + 283 + ], + "spans": [ + { + "bbox": [ + 130, + 176, + 482, + 283 + ], + "type": "text", + "content": "Despite achieving relatively lower top-1 accuracy on the ImageNet benchmark, lightweight ConvNets exhibit a competitive edge in top-10 accuracy when compared to their larger ViT counterparts (Fig. 4). This phenomenon underscores ConvNets' capability to effectively localize regions of interest through their convolutional layers, despite potential limitations in fine-grained classification attributed to the fully-connected (FC) layer. Our methodology leverages this insight by proposing discriminative patch regions while minimizing the influence of the FC layer (see Fig. 2), thereby enhancing the large off-the-shelf model's focus on most salient image regions proposed by lightweight ConvNets (see Fig. 3)." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 130, + 284, + 482, + 401 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 284, + 482, + 401 + ], + "spans": [ + { + "bbox": [ + 130, + 284, + 482, + 401 + ], + "type": "text", + "content": "Given an input image " + }, + { + "bbox": [ + 130, + 284, + 482, + 401 + ], + "type": "inline_equation", + "content": "X \\in \\mathbb{R}^{H \\times W \\times 3}" + }, + { + "bbox": [ + 130, + 284, + 482, + 401 + ], + "type": "text", + "content": ", with height " + }, + { + "bbox": [ + 130, + 284, + 482, + 401 + ], + "type": "inline_equation", + "content": "H" + }, + { + "bbox": [ + 130, + 284, + 482, + 401 + ], + "type": "text", + "content": " and width " + }, + { + "bbox": [ + 130, + 284, + 482, + 401 + ], + "type": "inline_equation", + "content": "W" + }, + { + "bbox": [ + 130, + 284, + 482, + 401 + ], + "type": "text", + "content": ", we denote by " + }, + { + "bbox": [ + 130, + 284, + 482, + 401 + ], + "type": "inline_equation", + "content": "f_{k}(x,y)" + }, + { + "bbox": [ + 130, + 284, + 482, + 401 + ], + "type": "text", + "content": " the feature map generated by the " + }, + { + "bbox": [ + 130, + 284, + 482, + 401 + ], + "type": "inline_equation", + "content": "k^{th}" + }, + { + "bbox": [ + 130, + 284, + 482, + 401 + ], + "type": "text", + "content": " kernel in the last convolutional layer of typical ConvNets, " + }, + { + "bbox": [ + 130, + 284, + 482, + 401 + ], + "type": "inline_equation", + "content": "\\forall k \\in \\{1,2,\\dots,K\\}" + }, + { + "bbox": [ + 130, + 284, + 482, + 401 + ], + "type": "text", + "content": ". Here, each pixel " + }, + { + "bbox": [ + 130, + 284, + 482, + 401 + ], + "type": "inline_equation", + "content": "(x,y)" + }, + { + "bbox": [ + 130, + 284, + 482, + 401 + ], + "type": "text", + "content": " in " + }, + { + "bbox": [ + 130, + 284, + 482, + 401 + ], + "type": "inline_equation", + "content": "f_{k}(x,y)" + }, + { + "bbox": [ + 130, + 284, + 482, + 401 + ], + "type": "text", + "content": " corresponds to a patch window of size " + }, + { + "bbox": [ + 130, + 284, + 482, + 401 + ], + "type": "inline_equation", + "content": "(H / d \\times W / d)" + }, + { + "bbox": [ + 130, + 284, + 482, + 401 + ], + "type": "text", + "content": " in the input image " + }, + { + "bbox": [ + 130, + 284, + 482, + 401 + ], + "type": "inline_equation", + "content": "X" + }, + { + "bbox": [ + 130, + 284, + 482, + 401 + ], + "type": "text", + "content": ", reflecting a spatial down-scaling factor of " + }, + { + "bbox": [ + 130, + 284, + 482, + 401 + ], + "type": "inline_equation", + "content": "d" + }, + { + "bbox": [ + 130, + 284, + 482, + 401 + ], + "type": "text", + "content": " through the convolutional layer stack. Typically, global average pooling (GAP) is applied to each feature map " + }, + { + "bbox": [ + 130, + 284, + 482, + 401 + ], + "type": "inline_equation", + "content": "f_{k}(x,y)" + }, + { + "bbox": [ + 130, + 284, + 482, + 401 + ], + "type": "text", + "content": ", yielding " + }, + { + "bbox": [ + 130, + 284, + 482, + 401 + ], + "type": "inline_equation", + "content": "F = \\{F_{k}\\}_{k=1}^{K} \\in \\mathbb{R}^{K}" + }, + { + "bbox": [ + 130, + 284, + 482, + 401 + ], + "type": "text", + "content": ", where " + }, + { + "bbox": [ + 130, + 284, + 482, + 401 + ], + "type": "inline_equation", + "content": "F_{k} = \\sum_{x,y} f_{k}(x,y) \\in \\mathbb{R}" + }, + { + "bbox": [ + 130, + 284, + 482, + 401 + ], + "type": "text", + "content": ". Finally, a fully connected (FC) layer " + }, + { + "bbox": [ + 130, + 284, + 482, + 401 + ], + "type": "inline_equation", + "content": "W \\in \\mathbb{R}^{C \\times K}" + }, + { + "bbox": [ + 130, + 284, + 482, + 401 + ], + "type": "text", + "content": " with elements " + }, + { + "bbox": [ + 130, + 284, + 482, + 401 + ], + "type": "inline_equation", + "content": "w_{c}^{k}" + }, + { + "bbox": [ + 130, + 284, + 482, + 401 + ], + "type": "text", + "content": " processes " + }, + { + "bbox": [ + 130, + 284, + 482, + 401 + ], + "type": "inline_equation", + "content": "F" + }, + { + "bbox": [ + 130, + 284, + 482, + 401 + ], + "type": "text", + "content": " to generate class predictions " + }, + { + "bbox": [ + 130, + 284, + 482, + 401 + ], + "type": "inline_equation", + "content": "y_{p} \\in \\mathbb{R}^{C}" + }, + { + "bbox": [ + 130, + 284, + 482, + 401 + ], + "type": "text", + "content": " across " + }, + { + "bbox": [ + 130, + 284, + 482, + 401 + ], + "type": "inline_equation", + "content": "C" + }, + { + "bbox": [ + 130, + 284, + 482, + 401 + ], + "type": "text", + "content": " classes as follows:" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 200, + 409, + 482, + 434 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 200, + 409, + 482, + 434 + ], + "spans": [ + { + "bbox": [ + 200, + 409, + 482, + 434 + ], + "type": "interline_equation", + "content": "y _ {p} = W F = \\sum_ {c} \\sum_ {k} w _ {c} ^ {k} \\sum_ {x, y} f _ {k} (x, y) = \\sum_ {x, y} R (x, y), \\tag {1}", + "image_path": "e75a9caefdad720a1543831a2a6b727cb369a668df87ef854d6e7836d4da0c20.jpg" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 130, + 443, + 482, + 467 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 443, + 482, + 467 + ], + "spans": [ + { + "bbox": [ + 130, + 443, + 482, + 467 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 130, + 443, + 482, + 467 + ], + "type": "inline_equation", + "content": "R \\in \\mathbb{R}^{h \\times w}" + }, + { + "bbox": [ + 130, + 443, + 482, + 467 + ], + "type": "text", + "content": " encapsulates the weighted mean class activation mapping of the convolutional feature map." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 130, + 467, + 482, + 563 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 467, + 482, + 563 + ], + "spans": [ + { + "bbox": [ + 130, + 467, + 482, + 563 + ], + "type": "text", + "content": "We note that the reliance on class activation weights " + }, + { + "bbox": [ + 130, + 467, + 482, + 563 + ], + "type": "inline_equation", + "content": "w_{c}^{k}" + }, + { + "bbox": [ + 130, + 467, + 482, + 563 + ], + "type": "text", + "content": " is influenced by the model's final accuracy, thereby posing a challenge for smaller models. Our objective transcends mere accuracy enhancement, aiming to precisely locate discriminative patches rich in information irrespective of the model size or its final classification performance. Recognizing that discriminative region localization is pivotal for detailed classification, and acknowledging the competitive top-10 accuracy of lighter ConvNets, we posit that an extremely lightweight ConvNet suffices for initial discriminative region proposal." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 130, + 563, + 482, + 599 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 563, + 482, + 599 + ], + "spans": [ + { + "bbox": [ + 130, + 563, + 482, + 599 + ], + "type": "text", + "content": "To counteract the influence of the weak linear FC layer " + }, + { + "bbox": [ + 130, + 563, + 482, + 599 + ], + "type": "inline_equation", + "content": "W" + }, + { + "bbox": [ + 130, + 563, + 482, + 599 + ], + "type": "text", + "content": " in convolutional region proposal, we propose an adjustment where " + }, + { + "bbox": [ + 130, + 563, + 482, + 599 + ], + "type": "inline_equation", + "content": "w_{c}^{k} = 1 / KC" + }, + { + "bbox": [ + 130, + 563, + 482, + 599 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 130, + 563, + 482, + 599 + ], + "type": "inline_equation", + "content": "\\forall k\\in \\{1,\\dots ,K\\}" + }, + { + "bbox": [ + 130, + 563, + 482, + 599 + ], + "type": "text", + "content": ", facilitating the generation of a discriminative region proposal " + }, + { + "bbox": [ + 130, + 563, + 482, + 599 + ], + "type": "inline_equation", + "content": "\\mathcal{R}\\in \\mathbb{R}^{h\\times w}" + }, + { + "bbox": [ + 130, + 563, + 482, + 599 + ], + "type": "text", + "content": ", where" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 203, + 607, + 482, + 634 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 203, + 607, + 482, + 634 + ], + "spans": [ + { + "bbox": [ + 203, + 607, + 482, + 634 + ], + "type": "interline_equation", + "content": "\\mathcal {R} (x, y) = \\sum_ {c} \\sum_ {k} \\frac {1}{K C} f _ {k} (x, y) = \\frac {1}{K} \\sum_ {k} f _ {k} (x, y). \\tag {2}", + "image_path": "771d955707dc5e4e619d7d762286ba095af0eb682ec112bdc48fe7190aa26d93.jpg" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 131, + 641, + 482, + 666 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 131, + 641, + 482, + 666 + ], + "spans": [ + { + "bbox": [ + 131, + 641, + 482, + 666 + ], + "type": "text", + "content": "This strategy allows us to leverage ConvNets for what they excel at: pinpointing critical image areas. By reducing reliance on class activation weights, we" + } + ] + } + ], + "index": 11 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "spans": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "type": "text", + "content": "6" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 166, + 91, + 240, + 101 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 166, + 91, + 240, + 101 + ], + "spans": [ + { + "bbox": [ + 166, + 91, + 240, + 101 + ], + "type": "text", + "content": "T. Mahmud et al." + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 5 + }, + { + "para_blocks": [ + { + "bbox": [ + 130, + 116, + 482, + 165 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 116, + 482, + 165 + ], + "spans": [ + { + "bbox": [ + 130, + 116, + 482, + 165 + ], + "type": "text", + "content": "efficiently generate discriminative region proposals directly from convolutional outputs. This approach not only enhances the interpretability and efficiency of the localization process but also enables the application of more complex models for subsequent detailed analysis, optimizing the use of computational resources." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 131, + 182, + 280, + 194 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 131, + 182, + 280, + 194 + ], + "spans": [ + { + "bbox": [ + 131, + 182, + 280, + 194 + ], + "type": "text", + "content": "3.2 Patch Significance Map" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 130, + 201, + 482, + 249 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 201, + 482, + 249 + ], + "spans": [ + { + "bbox": [ + 130, + 201, + 482, + 249 + ], + "type": "text", + "content": "With the discriminative region proposal " + }, + { + "bbox": [ + 130, + 201, + 482, + 249 + ], + "type": "inline_equation", + "content": "\\mathcal{R} \\in \\mathbb{R}^{h \\times w}" + }, + { + "bbox": [ + 130, + 201, + 482, + 249 + ], + "type": "text", + "content": ", where each pixel " + }, + { + "bbox": [ + 130, + 201, + 482, + 249 + ], + "type": "inline_equation", + "content": "(x, y)" + }, + { + "bbox": [ + 130, + 201, + 482, + 249 + ], + "type": "text", + "content": " quantifies the significance of corresponding patches in the original image, we proceed to establish a precise mapping to our intended feature map (Fig. 2). This mapping ensures the preservation of spatial relationships within the feature map." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 130, + 249, + 482, + 334 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 249, + 482, + 334 + ], + "spans": [ + { + "bbox": [ + 130, + 249, + 482, + 334 + ], + "type": "text", + "content": "Consider a target feature map " + }, + { + "bbox": [ + 130, + 249, + 482, + 334 + ], + "type": "inline_equation", + "content": "\\mathcal{F} \\in \\mathbb{R}^{h' \\times w' \\times K}" + }, + { + "bbox": [ + 130, + 249, + 482, + 334 + ], + "type": "text", + "content": ", with each pixel " + }, + { + "bbox": [ + 130, + 249, + 482, + 334 + ], + "type": "inline_equation", + "content": "(x', y')" + }, + { + "bbox": [ + 130, + 249, + 482, + 334 + ], + "type": "text", + "content": " encapsulating a feature vector from a specific image patch. To align " + }, + { + "bbox": [ + 130, + 249, + 482, + 334 + ], + "type": "inline_equation", + "content": "\\mathcal{R}" + }, + { + "bbox": [ + 130, + 249, + 482, + 334 + ], + "type": "text", + "content": " with " + }, + { + "bbox": [ + 130, + 249, + 482, + 334 + ], + "type": "inline_equation", + "content": "\\mathcal{F}" + }, + { + "bbox": [ + 130, + 249, + 482, + 334 + ], + "type": "text", + "content": ", we employ an upsampling operation " + }, + { + "bbox": [ + 130, + 249, + 482, + 334 + ], + "type": "inline_equation", + "content": "U: \\mathbb{R}^{h \\times w} \\to \\mathbb{R}^{h' \\times w'}" + }, + { + "bbox": [ + 130, + 249, + 482, + 334 + ], + "type": "text", + "content": ", transforming " + }, + { + "bbox": [ + 130, + 249, + 482, + 334 + ], + "type": "inline_equation", + "content": "\\mathcal{R}" + }, + { + "bbox": [ + 130, + 249, + 482, + 334 + ], + "type": "text", + "content": " into the Patch Significance Map (PSM) " + }, + { + "bbox": [ + 130, + 249, + 482, + 334 + ], + "type": "inline_equation", + "content": "\\mathcal{P} \\in \\mathbb{R}^{h' \\times w'}" + }, + { + "bbox": [ + 130, + 249, + 482, + 334 + ], + "type": "text", + "content": ". Consequently, each element of " + }, + { + "bbox": [ + 130, + 249, + 482, + 334 + ], + "type": "inline_equation", + "content": "\\mathcal{P}" + }, + { + "bbox": [ + 130, + 249, + 482, + 334 + ], + "type": "text", + "content": " directly corresponds to the patch significance within " + }, + { + "bbox": [ + 130, + 249, + 482, + 334 + ], + "type": "inline_equation", + "content": "\\mathcal{F}" + }, + { + "bbox": [ + 130, + 249, + 482, + 334 + ], + "type": "text", + "content": " for the given discriminative task. The next step involves utilizing " + }, + { + "bbox": [ + 130, + 249, + 482, + 334 + ], + "type": "inline_equation", + "content": "\\mathcal{P}" + }, + { + "bbox": [ + 130, + 249, + 482, + 334 + ], + "type": "text", + "content": " to discern and prune non-essential patch features from " + }, + { + "bbox": [ + 130, + 249, + 482, + 334 + ], + "type": "inline_equation", + "content": "\\mathcal{F}" + }, + { + "bbox": [ + 130, + 249, + 482, + 334 + ], + "type": "text", + "content": ". By sorting the values within " + }, + { + "bbox": [ + 130, + 249, + 482, + 334 + ], + "type": "inline_equation", + "content": "\\mathcal{P}" + }, + { + "bbox": [ + 130, + 249, + 482, + 334 + ], + "type": "text", + "content": ", we acquire a pruning mask " + }, + { + "bbox": [ + 130, + 249, + 482, + 334 + ], + "type": "inline_equation", + "content": "\\mathcal{M}" + }, + { + "bbox": [ + 130, + 249, + 482, + 334 + ], + "type": "text", + "content": ":" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 202, + 341, + 482, + 357 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 202, + 341, + 482, + 357 + ], + "spans": [ + { + "bbox": [ + 202, + 341, + 482, + 357 + ], + "type": "interline_equation", + "content": "\\mathcal {M} = \\operatorname {r e s h a p e} (\\operatorname {a r g s o r t} (\\operatorname {f l a t t e n} (\\mathcal {P}))), \\mathcal {M} \\in \\mathbb {R} ^ {h ^ {\\prime} \\times w ^ {\\prime}}, \\tag {3}", + "image_path": "5b07dae3612743c3699154d0deee07b5ef3fc7349018c37fc5dc6b91be5a976b.jpg" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 130, + 365, + 482, + 411 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 365, + 482, + 411 + ], + "spans": [ + { + "bbox": [ + 130, + 365, + 482, + 411 + ], + "type": "text", + "content": "enabling batch-wise patch pruning. Specifically, we identify the indices corresponding to the top-" + }, + { + "bbox": [ + 130, + 365, + 482, + 411 + ], + "type": "inline_equation", + "content": "z\\%" + }, + { + "bbox": [ + 130, + 365, + 482, + 411 + ], + "type": "text", + "content": " patches as per " + }, + { + "bbox": [ + 130, + 365, + 482, + 411 + ], + "type": "inline_equation", + "content": "\\mathcal{M}" + }, + { + "bbox": [ + 130, + 365, + 482, + 411 + ], + "type": "text", + "content": ", facilitating the retention of only the most salient patches in " + }, + { + "bbox": [ + 130, + 365, + 482, + 411 + ], + "type": "inline_equation", + "content": "\\mathcal{F}" + }, + { + "bbox": [ + 130, + 365, + 482, + 411 + ], + "type": "text", + "content": ", thereby enhancing computational efficiency in subsequent processing stages." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 130, + 412, + 482, + 460 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 412, + 482, + 460 + ], + "spans": [ + { + "bbox": [ + 130, + 412, + 482, + 460 + ], + "type": "text", + "content": "As part of our discussion on various architectures, we detail the application of this patch reduction technique. For enhanced visualization of the PSM, we apply min-max normalization to " + }, + { + "bbox": [ + 130, + 412, + 482, + 460 + ], + "type": "inline_equation", + "content": "\\mathcal{P}" + }, + { + "bbox": [ + 130, + 412, + 482, + 460 + ], + "type": "text", + "content": ", adjusting for outliers and scaling the significance scores to visually depict the importance of different patches." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 131, + 478, + 378, + 491 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 131, + 478, + 378, + 491 + ], + "spans": [ + { + "bbox": [ + 131, + 478, + 378, + 491 + ], + "type": "text", + "content": "3.3 Integrating PSM with Vision Transformers" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 130, + 498, + 482, + 570 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 498, + 482, + 570 + ], + "spans": [ + { + "bbox": [ + 130, + 498, + 482, + 570 + ], + "type": "text", + "content": "The ViT processes an input image " + }, + { + "bbox": [ + 130, + 498, + 482, + 570 + ], + "type": "inline_equation", + "content": "X \\in \\mathbb{R}^{H \\times W \\times 3}" + }, + { + "bbox": [ + 130, + 498, + 482, + 570 + ], + "type": "text", + "content": " by extracting " + }, + { + "bbox": [ + 130, + 498, + 482, + 570 + ], + "type": "inline_equation", + "content": "N" + }, + { + "bbox": [ + 130, + 498, + 482, + 570 + ], + "type": "text", + "content": "-dimensional patch token features " + }, + { + "bbox": [ + 130, + 498, + 482, + 570 + ], + "type": "inline_equation", + "content": "X_{p} \\in \\mathbb{R}^{N \\times d}" + }, + { + "bbox": [ + 130, + 498, + 482, + 570 + ], + "type": "text", + "content": " from a " + }, + { + "bbox": [ + 130, + 498, + 482, + 570 + ], + "type": "inline_equation", + "content": "(k \\times k)" + }, + { + "bbox": [ + 130, + 498, + 482, + 570 + ], + "type": "text", + "content": " patch window using strided convolutions, where " + }, + { + "bbox": [ + 130, + 498, + 482, + 570 + ], + "type": "inline_equation", + "content": "N = HW / k^2" + }, + { + "bbox": [ + 130, + 498, + 482, + 570 + ], + "type": "text", + "content": " (Fig. 3). Positional embeddings are then added to these patch tokens " + }, + { + "bbox": [ + 130, + 498, + 482, + 570 + ], + "type": "inline_equation", + "content": "X_{p}" + }, + { + "bbox": [ + 130, + 498, + 482, + 570 + ], + "type": "text", + "content": " to retain spatial information, followed by the application of successive cross-attention mechanisms. For classification tasks, the model either introduces a class token " + }, + { + "bbox": [ + 130, + 498, + 482, + 570 + ], + "type": "inline_equation", + "content": "t_{cls} \\in \\mathbb{R}^{d}" + }, + { + "bbox": [ + 130, + 498, + 482, + 570 + ], + "type": "text", + "content": " or utilizes the mean of the output patch tokens." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 130, + 570, + 482, + 666 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 570, + 482, + 666 + ], + "spans": [ + { + "bbox": [ + 130, + 570, + 482, + 666 + ], + "type": "text", + "content": "The encoder employs multi-headed cross-attention (MHA) on the patch token embeddings " + }, + { + "bbox": [ + 130, + 570, + 482, + 666 + ], + "type": "inline_equation", + "content": "X_{p}" + }, + { + "bbox": [ + 130, + 570, + 482, + 666 + ], + "type": "text", + "content": ", with each MHA block consisting of multiple linear layers and a cross-attention layer. The computational complexity of these operations suggests that reducing the number of tokens " + }, + { + "bbox": [ + 130, + 570, + 482, + 666 + ], + "type": "inline_equation", + "content": "N" + }, + { + "bbox": [ + 130, + 570, + 482, + 666 + ], + "type": "text", + "content": " can significantly decrease computational demands. A key strength of the ViT architecture is its adaptability to varying numbers of tokens " + }, + { + "bbox": [ + 130, + 570, + 482, + 666 + ], + "type": "inline_equation", + "content": "N" + }, + { + "bbox": [ + 130, + 570, + 482, + 666 + ], + "type": "text", + "content": ", though at a higher computational cost. Prior work [2,41,56] has primarily aimed at reducing patch tokens within intermediate encoder layers, often requiring additional training due to the variability in token representations" + } + ] + } + ], + "index": 11 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 343, + 91, + 447, + 101 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 343, + 91, + 447, + 101 + ], + "spans": [ + { + "bbox": [ + 343, + 91, + 447, + 101 + ], + "type": "text", + "content": "PaPr for Faster Inference" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 474, + 91, + 481, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 474, + 91, + 481, + 100 + ], + "spans": [ + { + "bbox": [ + 474, + 91, + 481, + 100 + ], + "type": "text", + "content": "7" + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 6 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 159, + 114, + 455, + 235 + ], + "blocks": [ + { + "bbox": [ + 159, + 114, + 455, + 235 + ], + "lines": [ + { + "bbox": [ + 159, + 114, + 455, + 235 + ], + "spans": [ + { + "bbox": [ + 159, + 114, + 455, + 235 + ], + "type": "image", + "image_path": "3636b67499797712f50e8b886624e4e8bfefb4d7a91e409f8147d30015cc21a2.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 130, + 238, + 482, + 316 + ], + "lines": [ + { + "bbox": [ + 130, + 238, + 482, + 316 + ], + "spans": [ + { + "bbox": [ + 130, + 238, + 482, + 316 + ], + "type": "text", + "content": "Fig. 3: (a) In vanilla ViT, PaPr operates right after the patch extractor module. Hence, all transformer blocks can operate only with the most discriminative patches. (b) Hierarchical model blocks comprise of window based kernel operator (e.g., Convk " + }, + { + "bbox": [ + 130, + 238, + 482, + 316 + ], + "type": "inline_equation", + "content": "\\times" + }, + { + "bbox": [ + 130, + 238, + 482, + 316 + ], + "type": "inline_equation", + "content": "k" + }, + { + "bbox": [ + 130, + 238, + 482, + 316 + ], + "type": "text", + "content": "/local attention), followed by pixel operator (e.g., linear layer, Conv1x1). Pixel operator consumes more than " + }, + { + "bbox": [ + 130, + 238, + 482, + 316 + ], + "type": "inline_equation", + "content": "60\\%" + }, + { + "bbox": [ + 130, + 238, + 482, + 316 + ], + "type": "text", + "content": " of total computation. PaPr is used to split the foreground patches to be operated with pixel operator. Background patches are zero-ed out, and finally, re-assembled with foreground output patches." + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_caption" + } + ], + "index": 2 + }, + { + "bbox": [ + 130, + 339, + 481, + 363 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 339, + 481, + 363 + ], + "spans": [ + { + "bbox": [ + 130, + 339, + 481, + 363 + ], + "type": "text", + "content": "across different architectures. These methods typically achieve only gradual token reduction, limited by the distributed nature of information across tokens." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 130, + 364, + 482, + 472 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 364, + 482, + 472 + ], + "spans": [ + { + "bbox": [ + 130, + 364, + 482, + 472 + ], + "type": "text", + "content": "In contrast, our approach seeks to eliminate redundant tokens early in the processing pipeline, immediately following initial patch extraction. This strategy offers multiple advantages: (1) Leveraging the convolutional nature of the initial patch extraction allows for a direct mapping between our PSM " + }, + { + "bbox": [ + 130, + 364, + 482, + 472 + ], + "type": "inline_equation", + "content": "\\mathcal{P}" + }, + { + "bbox": [ + 130, + 364, + 482, + 472 + ], + "type": "text", + "content": " and the patch tokens " + }, + { + "bbox": [ + 130, + 364, + 482, + 472 + ], + "type": "inline_equation", + "content": "X_{p}" + }, + { + "bbox": [ + 130, + 364, + 482, + 472 + ], + "type": "text", + "content": ", obviating the need for additional mechanisms to determine token redundancy. (2) It facilitates generalizability across transformer models without the need for retraining, potentially accelerating training by concentrating on essential patches. (3) It permits seamless integration of existing intermediate layer token reduction techniques following our initial patch pruning process." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 131, + 489, + 376, + 502 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 131, + 489, + 376, + 502 + ], + "spans": [ + { + "bbox": [ + 131, + 489, + 376, + 502 + ], + "type": "text", + "content": "3.4 Integrating PSM with Hierarchical Models" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 130, + 510, + 482, + 594 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 510, + 482, + 594 + ], + "spans": [ + { + "bbox": [ + 130, + 510, + 482, + 594 + ], + "type": "text", + "content": "ConvNets and hybrid transformers, such as Swin [31], primarily operate with window-based local operations in contrast to full-attention based operation in vanilla transformers. Such window-based local processing makes patch pruning particularly complicated, in contrast to vanilla transformers. However, these models maintain the location property of representative image patches all through the network, which leaves the door open to prune redundant patches. Nevertheless, the window-based operation is particularly difficult to prune." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 130, + 594, + 482, + 666 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 594, + 482, + 666 + ], + "spans": [ + { + "bbox": [ + 130, + 594, + 482, + 666 + ], + "type": "text", + "content": "In general, each block of hierarchical layers consists of window-based spatial kernel/attention operators, followed by pixel operators (usually, modeled with " + }, + { + "bbox": [ + 130, + 594, + 482, + 666 + ], + "type": "inline_equation", + "content": "1 \\times 1" + }, + { + "bbox": [ + 130, + 594, + 482, + 666 + ], + "type": "text", + "content": " convolutional layers or linear layers). In contrast to windowed convolutions or cross-attention, these pixel operators are particularly suitable for patch pruning. Interestingly, in the SOTA ConvNets and hybrid transformers, more than " + }, + { + "bbox": [ + 130, + 594, + 482, + 666 + ], + "type": "inline_equation", + "content": "60\\%" + }, + { + "bbox": [ + 130, + 594, + 482, + 666 + ], + "type": "text", + "content": " of total computations are performed with such pixel operators (63.3% in Swin [31]," + } + ] + } + ], + "index": 8 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "spans": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "type": "text", + "content": "8" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 166, + 91, + 240, + 101 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 166, + 91, + 240, + 101 + ], + "spans": [ + { + "bbox": [ + 166, + 91, + 240, + 101 + ], + "type": "text", + "content": "T. Mahmud et al." + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 7 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 137, + 115, + 304, + 213 + ], + "blocks": [ + { + "bbox": [ + 137, + 115, + 304, + 213 + ], + "lines": [ + { + "bbox": [ + 137, + 115, + 304, + 213 + ], + "spans": [ + { + "bbox": [ + 137, + 115, + 304, + 213 + ], + "type": "image", + "image_path": "c427eb00d0e8b6575f9866740cad876e19f4ab819010fc3fc68aec68cebd876d.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 134, + 218, + 305, + 295 + ], + "lines": [ + { + "bbox": [ + 134, + 218, + 305, + 295 + ], + "spans": [ + { + "bbox": [ + 134, + 218, + 305, + 295 + ], + "type": "text", + "content": "Fig. 4: ImageNet-1k evaluation for varying top-k accuracy targets. The accuracy gain with bigger model largely shrinks, as " + }, + { + "bbox": [ + 134, + 218, + 305, + 295 + ], + "type": "inline_equation", + "content": "k" + }, + { + "bbox": [ + 134, + 218, + 305, + 295 + ], + "type": "text", + "content": " increases. This suggests shallower ConvNets have understanding of object locations and visual property, despite their lower top-1 accuracy." + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_caption" + } + ], + "index": 2 + }, + { + "type": "image", + "bbox": [ + 310, + 114, + 477, + 213 + ], + "blocks": [ + { + "bbox": [ + 310, + 114, + 477, + 213 + ], + "lines": [ + { + "bbox": [ + 310, + 114, + 477, + 213 + ], + "spans": [ + { + "bbox": [ + 310, + 114, + 477, + 213 + ], + "type": "image", + "image_path": "8780e33834425319c90bbbf763f15d7615eff67316943ce4be7a9aec94d968a4.jpg" + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 307, + 218, + 479, + 295 + ], + "lines": [ + { + "bbox": [ + 307, + 218, + 479, + 295 + ], + "spans": [ + { + "bbox": [ + 307, + 218, + 479, + 295 + ], + "type": "text", + "content": "Fig.5: Integrating PaPr with ToMe [2]. We use the Augreg pretrained ViT-B-16 architecture as the baseline. We sweep token merging ratio (r) for different pruning ratio (z). Integration of PaPr achieves Pareto-optimal performance, thus, PaPr can enhance existing patch reduction methods." + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "image_caption" + } + ], + "index": 4 + }, + { + "bbox": [ + 130, + 319, + 482, + 390 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 319, + 482, + 390 + ], + "spans": [ + { + "bbox": [ + 130, + 319, + 482, + 390 + ], + "type": "inline_equation", + "content": "96.2\\%" + }, + { + "bbox": [ + 130, + 319, + 482, + 390 + ], + "type": "text", + "content": " in ConvNext [32]). Based on the patch significance map from PaPr, we only use the pixel operator on most significant patches, and simply perform zero-padding on the remaining patch regions. The zero padding operations are performed to mostly recover the feature map shape to be used with subsequent spatial operators. Hence, the speed-up is mostly achieved by eliminating redundant computations in pixel operators (See Fig. 3)." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 130, + 392, + 482, + 475 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 392, + 482, + 475 + ], + "spans": [ + { + "bbox": [ + 130, + 392, + 482, + 475 + ], + "type": "text", + "content": "Let's assume that the feature map before applying pixel operators is given by " + }, + { + "bbox": [ + 130, + 392, + 482, + 475 + ], + "type": "inline_equation", + "content": "\\mathcal{A}_w" + }, + { + "bbox": [ + 130, + 392, + 482, + 475 + ], + "type": "text", + "content": ". Based on the PSM " + }, + { + "bbox": [ + 130, + 392, + 482, + 475 + ], + "type": "inline_equation", + "content": "\\mathcal{P}" + }, + { + "bbox": [ + 130, + 392, + 482, + 475 + ], + "type": "text", + "content": ", we initially identify the foreground and background pixel features as " + }, + { + "bbox": [ + 130, + 392, + 482, + 475 + ], + "type": "inline_equation", + "content": "A_{f}" + }, + { + "bbox": [ + 130, + 392, + 482, + 475 + ], + "type": "text", + "content": ", and " + }, + { + "bbox": [ + 130, + 392, + 482, + 475 + ], + "type": "inline_equation", + "content": "A_{b}" + }, + { + "bbox": [ + 130, + 392, + 482, + 475 + ], + "type": "text", + "content": ", respectively. Later, the pixel operator layers are applied on foreground pixels " + }, + { + "bbox": [ + 130, + 392, + 482, + 475 + ], + "type": "inline_equation", + "content": "A_{f}" + }, + { + "bbox": [ + 130, + 392, + 482, + 475 + ], + "type": "text", + "content": ", and a zero-padded representation is used for the background pixels " + }, + { + "bbox": [ + 130, + 392, + 482, + 475 + ], + "type": "inline_equation", + "content": "A_{b}" + }, + { + "bbox": [ + 130, + 392, + 482, + 475 + ], + "type": "text", + "content": ". Finally, the output representation " + }, + { + "bbox": [ + 130, + 392, + 482, + 475 + ], + "type": "inline_equation", + "content": "\\mathcal{A}_c" + }, + { + "bbox": [ + 130, + 392, + 482, + 475 + ], + "type": "text", + "content": " is reassembled with modulated foreground pixels " + }, + { + "bbox": [ + 130, + 392, + 482, + 475 + ], + "type": "inline_equation", + "content": "A_{f}" + }, + { + "bbox": [ + 130, + 392, + 482, + 475 + ], + "type": "text", + "content": " and zero-padded background pixels " + }, + { + "bbox": [ + 130, + 392, + 482, + 475 + ], + "type": "inline_equation", + "content": "A_{z}" + }, + { + "bbox": [ + 130, + 392, + 482, + 475 + ], + "type": "text", + "content": ", given by Eq. (4):" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 282, + 475, + 382, + 488 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 282, + 475, + 382, + 488 + ], + "spans": [ + { + "bbox": [ + 282, + 475, + 382, + 488 + ], + "type": "interline_equation", + "content": "A _ {f}, A _ {b} = \\operatorname {S p l i t} \\left(\\mathcal {A} _ {w}, \\mathcal {P}\\right),", + "image_path": "dac35674564b0b728b1ae5a9612993da465f59a97545809b40d4136a4fcae557.jpg" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 233, + 490, + 481, + 503 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 233, + 490, + 481, + 503 + ], + "spans": [ + { + "bbox": [ + 233, + 490, + 481, + 503 + ], + "type": "interline_equation", + "content": "A _ {z} = \\operatorname {Z e r o s} \\left(A _ {b}\\right), A _ {f} = \\operatorname {L i n e a r} \\left(A _ {f}\\right), \\tag {4}", + "image_path": "6ad82878b058f2b5d15e1607fd5268a8fbdf728ebc97cd3604084ba217991a7f.jpg" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 266, + 506, + 382, + 517 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 266, + 506, + 382, + 517 + ], + "spans": [ + { + "bbox": [ + 266, + 506, + 382, + 517 + ], + "type": "interline_equation", + "content": "\\mathcal {A} _ {c} = \\operatorname {R e a s s e m b l e} \\left(A _ {f}, A _ {z}\\right).", + "image_path": "f3a662f74a383434dae184639c1f2bf121e79b0b36f34a50d0696d8375c48c3d.jpg" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 132, + 535, + 270, + 549 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 535, + 270, + 549 + ], + "spans": [ + { + "bbox": [ + 132, + 535, + 270, + 549 + ], + "type": "text", + "content": "4 Image Experiments" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 132, + 561, + 261, + 574 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 561, + 261, + 574 + ], + "spans": [ + { + "bbox": [ + 132, + 561, + 261, + 574 + ], + "type": "text", + "content": "4.1 Experimental Setup" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 130, + 581, + 482, + 666 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 581, + 482, + 666 + ], + "spans": [ + { + "bbox": [ + 130, + 581, + 482, + 666 + ], + "type": "text", + "content": "We experiment on image classification task on the ImageNet-1k benchmark dataset, following prior work [2, 12, 23, 34, 55]. We use the MobileOne-s0 [49] model as the proposal model for all architectures, unless otherwise specified. We report training-free results of PaPr, unless otherwise specified. For the performance metric, we report top-1 accuracy, GFLOPs, and the throughput (img/s). Throughput is measured on a single RTX-A5000 GPU with 24GB VRAM. We reproduced all baseline models under the same setup for a fair comparison." + } + ] + } + ], + "index": 13 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 343, + 91, + 447, + 100 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 343, + 91, + 447, + 100 + ], + "spans": [ + { + "bbox": [ + 343, + 91, + 447, + 100 + ], + "type": "text", + "content": "PaPr for Faster Inference" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 474, + 91, + 480, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 474, + 91, + 480, + 100 + ], + "spans": [ + { + "bbox": [ + 474, + 91, + 480, + 100 + ], + "type": "text", + "content": "9" + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 8 + }, + { + "para_blocks": [ + { + "type": "table", + "bbox": [ + 137, + 170, + 310, + 357 + ], + "blocks": [ + { + "bbox": [ + 133, + 116, + 310, + 160 + ], + "lines": [ + { + "bbox": [ + 133, + 116, + 310, + 160 + ], + "spans": [ + { + "bbox": [ + 133, + 116, + 310, + 160 + ], + "type": "text", + "content": "Table 1: Performance comparison on AugReg models. PaPr achieves the best performance, while operating with token merging. " + }, + { + "bbox": [ + 133, + 116, + 310, + 160 + ], + "type": "inline_equation", + "content": "z" + }, + { + "bbox": [ + 133, + 116, + 310, + 160 + ], + "type": "text", + "content": " denotes patch keeping ratio in PaPr." + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 137, + 170, + 310, + 357 + ], + "lines": [ + { + "bbox": [ + 137, + 170, + 310, + 357 + ], + "spans": [ + { + "bbox": [ + 137, + 170, + 310, + 357 + ], + "type": "table", + "html": "
ModelsMethodsAcc1GFLOPsImg/s
ViT-S-16Baseline81.394.61975
ToMe [2]76.962.291978
TokenFusion [23]77.122.291982
GTP-ViT [55]71.032.311970
PaPr (z=0.45)76.212.291988
PaPr (z=0.55) (+ToMe)77.762.192073
PaPr (z=0.5) (+ToMe)76.271.972315
ViT-B-16Baseline84.5917.59307
ToMe [2]80.388.78615
TokenFusion [23]80.78.78618
GTP-ViT [55]80.988.78610
PaPr (z=0.5)82.118.98605
PaPr (z=0.55) (+ToMe)82.348.21660
PaPr (z=0.5) (+ToMe)80.886.82785
ViT-L-16Baseline85.8261.6191
ToMe [2]83.530.99180
TokenFusion [23]83.9130.99182
GTP-ViT [55]81.5630.99181
PaPr (z=0.5)83.8730.83183
PaPr (z=0.55) (+ToMe)83.9926.33210
PaPr (z=0.55) (+ToMe)83.525.12224
", + "image_path": "763a884e2d972a0963e4b8753d236ba9ef32318929a0fb1a7fc98fd5cea0101e.jpg" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "table_body" + } + ], + "index": 3 + }, + { + "type": "table", + "bbox": [ + 318, + 190, + 476, + 358 + ], + "blocks": [ + { + "bbox": [ + 313, + 114, + 478, + 180 + ], + "lines": [ + { + "bbox": [ + 313, + 114, + 478, + 180 + ], + "spans": [ + { + "bbox": [ + 313, + 114, + 478, + 180 + ], + "type": "text", + "content": "Table 2: Performance comparison on MAE models. Since MAE uses masked pretraining, PaPr is particularly suitable for MAE inference. PaPr achieves significantly higher performance than others. " + }, + { + "bbox": [ + 313, + 114, + 478, + 180 + ], + "type": "inline_equation", + "content": "z" + }, + { + "bbox": [ + 313, + 114, + 478, + 180 + ], + "type": "text", + "content": " denotes patch keeping ratio in PaPr." + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 318, + 190, + 476, + 358 + ], + "lines": [ + { + "bbox": [ + 318, + 190, + 476, + 358 + ], + "spans": [ + { + "bbox": [ + 318, + 190, + 476, + 358 + ], + "type": "table", + "html": "
ModelsMethodsAcc1GFLOPsImg/s
ViT-B-16Baseline83.7417.59307
ToMe [2]78.828.78615
TokenFusion [23]79.238.78618
GTP-ViT [55]79.148.78610
PaPr (z=0.5)82.48.98605
PaPr (z=0.4)81.47.72700
ViT-L-16Baseline85.9561.6191
ToMe [2]84.2430.99180
TokenFusion [23]84.3330.99182
GTP-ViT [55]84.1530.99181
PaPr (z=0.5)85.0630.83183
PaPr (z=0.4)84.7627.72201
ViT-H-16Baseline86.89167.436
ToMe [2]85.4882.5372
TokenFusion [23]85.7182.5373
GTP-ViT [55]85.5482.5371
PaPr (z=0.5)86.483.0471
PaPr (z=0.4)86.1374.5981
", + "image_path": "b47354f95bd87e820185c52ea83fa2e43a945bb224a8295bc17bac26ae6b192a.jpg" + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "table_body" + } + ], + "index": 5 + }, + { + "bbox": [ + 132, + 380, + 388, + 391 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 380, + 388, + 391 + ], + "spans": [ + { + "bbox": [ + 132, + 380, + 388, + 391 + ], + "type": "text", + "content": "4.2 Performance on Various Vision Transformers" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 130, + 398, + 481, + 434 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 398, + 481, + 434 + ], + "spans": [ + { + "bbox": [ + 130, + 398, + 481, + 434 + ], + "type": "text", + "content": "We study performance of PaPr in diverse ViT architectures, along with various pre-training methods. We also present training results, along with training-free results to compare with existing methods." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 130, + 451, + 482, + 559 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 451, + 482, + 559 + ], + "spans": [ + { + "bbox": [ + 130, + 451, + 482, + 559 + ], + "type": "text", + "content": "Training-free method comparison. We consider two different pre-training methods for comparative analysis on various ViT architectures. We use supervised pretrained Augreg models [45], and self-supervised pre-trained masked autoencoder (MAE) models [17], following prior work [2]. We compare with three recent training-free patch reduction methods, such as ToMe [2], TokenFusion [23], and GTP-ViT [55]. Additionally, we show the performance comparison on class-token free ViT models to highlight the architecture agnostic performance. In general, PaPr achieves better accuracy, with lower computational costs compared to state-of-the-art (SOTA) methods." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 130, + 569, + 482, + 666 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 569, + 482, + 666 + ], + "spans": [ + { + "bbox": [ + 130, + 569, + 482, + 666 + ], + "type": "text", + "content": "Augreg models: For AugReg pre-training with ViT-Base, PaPr achieves " + }, + { + "bbox": [ + 130, + 569, + 482, + 666 + ], + "type": "inline_equation", + "content": "2.14\\%" + }, + { + "bbox": [ + 130, + 569, + 482, + 666 + ], + "type": "text", + "content": " higher accuracy over ToMe, with comparable FLOPs as shown in Tab. 1. In addition, by combining PaPr with ToMe, we achieve additional " + }, + { + "bbox": [ + 130, + 569, + 482, + 666 + ], + "type": "inline_equation", + "content": "22.3\\%" + }, + { + "bbox": [ + 130, + 569, + 482, + 666 + ], + "type": "text", + "content": " FLOP reduction, while maintaining higher accuracy. We further study the compatibility of PaPr with existing patch reduction methods as ToMe. After the initial patch pruning with PaPr, we integrate ToMe only at the bottom-half layers due to its higher sensitivity in earlier layers. As shown in Fig. 5, integration of PaPr can directly boost ToMe performance thereby achieving Pareto optimality." + } + ] + } + ], + "index": 9 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "spans": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "type": "text", + "content": "10" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 166, + 91, + 240, + 101 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 166, + 91, + 240, + 101 + ], + "spans": [ + { + "bbox": [ + 166, + 91, + 240, + 101 + ], + "type": "text", + "content": "T. Mahmud et al." + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 9 + }, + { + "para_blocks": [ + { + "type": "table", + "bbox": [ + 140, + 157, + 304, + 245 + ], + "blocks": [ + { + "bbox": [ + 135, + 114, + 305, + 148 + ], + "lines": [ + { + "bbox": [ + 135, + 114, + 305, + 148 + ], + "spans": [ + { + "bbox": [ + 135, + 114, + 305, + 148 + ], + "type": "text", + "content": "Table 3: Performance analysis on class-token free ViT models. PaPr performance gain is not limited to specific architectures." + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 140, + 157, + 304, + 245 + ], + "lines": [ + { + "bbox": [ + 140, + 157, + 304, + 245 + ], + "spans": [ + { + "bbox": [ + 140, + 157, + 304, + 245 + ], + "type": "table", + "html": "
ModelMethodAcc1GFLOPsImg/s
ViT-MediumBaseline84.3310.58484
GAP-16GTP-ViT [55]80.55.73880
256ToMe [2]80.215.73890
PaPr (z=0.5)81.85.5932
ViT-MediumBaseline85.626.06182
GAP-16GTP-ViT [55]81.9513.59340
384ToMe [2]82.513.59344
PaPr83.9512.94360
", + "image_path": "b6fb9adb44c49c914e1c3c4a144082584c2794a10e78d201af053f12aa33c030.jpg" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "table_body" + } + ], + "index": 3 + }, + { + "type": "table", + "bbox": [ + 313, + 158, + 476, + 245 + ], + "blocks": [ + { + "bbox": [ + 308, + 114, + 478, + 148 + ], + "lines": [ + { + "bbox": [ + 308, + 114, + 478, + 148 + ], + "spans": [ + { + "bbox": [ + 308, + 114, + 478, + 148 + ], + "type": "text", + "content": "Table 4: Training performance analysis on DeIT-s. PaPr achieves competitive performance with higher training speed." + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 313, + 158, + 476, + 245 + ], + "lines": [ + { + "bbox": [ + 313, + 158, + 476, + 245 + ], + "spans": [ + { + "bbox": [ + 313, + 158, + 476, + 245 + ], + "type": "table", + "html": "
MethodsAcc1GFLOPsIm/sBatch Proc.Train Speed
Baseline79.84.6960✓1x
DynamicViT [41]79.32.91510X1x
A-ViT [56]78.62.9-X1x
ATS [12]79.52.91512-1x
SP-ViT [25]79.32.6-X1x
ToMe [2]79.42.71575✓1.5x
TokenFusion [23]79.52.71580✓1.5x
PaPr (z=0.55)79.22.71585✓1.6x
", + "image_path": "87ea40266a13d60a8025276ff69aff57573389808ea19476e6c205ccfafbcfae.jpg" + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "table_body" + } + ], + "index": 5 + }, + { + "type": "table", + "bbox": [ + 138, + 313, + 304, + 401 + ], + "blocks": [ + { + "bbox": [ + 134, + 258, + 306, + 304 + ], + "lines": [ + { + "bbox": [ + 134, + 258, + 306, + 304 + ], + "spans": [ + { + "bbox": [ + 134, + 258, + 306, + 304 + ], + "type": "text", + "content": "Table 5: Performance comparison on ConvNext CNN models. PaPr can achieve competitive performance without training, and can seamlessly adapt to model updates." + } + ] + } + ], + "index": 6, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 138, + 313, + 304, + 401 + ], + "lines": [ + { + "bbox": [ + 138, + 313, + 304, + 401 + ], + "spans": [ + { + "bbox": [ + 138, + 313, + 304, + 401 + ], + "type": "table", + "html": "
ModelsMethodsTrain-FreeAcc1GFLOPsImg/s
ConvNeXtBaselineN/A83.8415.38265
Base-1kDynCNN [40]✘83.0810.21375
PaPr (z=0.65)✓82.7510.42395
ConvNeXtBaselineN/A85.8115.38265
Base-22kPaPr (z=0.65)✓84.2710.42395
ConvNeXtBaselineN/A84.3134.4135
Large-1kPaPr (z=0.65)✓83.2622.9203
ConvNeXtBaselineN/A86.6134.4135
Large-22kPaPr (z=0.65)✓85.6722.9203
", + "image_path": "f62733fddaa68b57f51440ec3d9800cb7b5ddeaa608c169cecdb485546816c6c.jpg" + } + ] + } + ], + "index": 7, + "angle": 0, + "type": "table_body" + } + ], + "index": 7 + }, + { + "type": "table", + "bbox": [ + 313, + 313, + 476, + 401 + ], + "blocks": [ + { + "bbox": [ + 308, + 258, + 478, + 304 + ], + "lines": [ + { + "bbox": [ + 308, + 258, + 478, + 304 + ], + "spans": [ + { + "bbox": [ + 308, + 258, + 478, + 304 + ], + "type": "text", + "content": "Table 6: Performance comparison on Swin hybrid transformer models. PaPr can adapt to much bigger models with higher operating resolutions without training." + } + ] + } + ], + "index": 8, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 313, + 313, + 476, + 401 + ], + "lines": [ + { + "bbox": [ + 313, + 313, + 476, + 401 + ], + "spans": [ + { + "bbox": [ + 313, + 313, + 476, + 401 + ], + "type": "table", + "html": "
ModelsMethodsTrain-FreeAcc1GFLOPsImg/s
Swin B-1kBaselineN/A83.4215.47258
DynSwin [40]✗83.1812.1327
PaPr (z=0.65)✓81.712.25325
Swin B-22kBaselineN/A85.1615.47258
PaPr (z=0.65)✓82.2712.25325
Swin L-22kBaselineN/A86.2534.53135
PaPr (z=0.65)✓84.5326.96175
Swin L-22k-384BaselineN/A87.25104.0842
PaPr (z=0.65)✓86.4781.4454
", + "image_path": "25edff39f045372724e390fa2e27cb83dcbdfe970019b82cdd1caf281da76d6a.jpg" + } + ] + } + ], + "index": 9, + "angle": 0, + "type": "table_body" + } + ], + "index": 9 + }, + { + "bbox": [ + 130, + 426, + 482, + 511 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 426, + 482, + 511 + ], + "spans": [ + { + "bbox": [ + 130, + 426, + 482, + 511 + ], + "type": "text", + "content": "MAE models: MAE used self-supervised pre-training by reconstructing masked image patches to train large ViT models [17]. Later, the model is fine-tuned with full resolution images without masking, which cannot exploit its latent ability to learn from fewer patches. Interestingly, PaPr introduces masked inference, which is particularly suitable for MAE models. Hence, we observe significantly higher accuracy with PaPr compared to other training-free methods, e.g., with ViT-B, PaPr achieves " + }, + { + "bbox": [ + 130, + 426, + 482, + 511 + ], + "type": "inline_equation", + "content": "4.5\\%" + }, + { + "bbox": [ + 130, + 426, + 482, + 511 + ], + "type": "text", + "content": " higher accuracy over ToMe for similar FLOPs (see Tab. 2)." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 130, + 531, + 482, + 580 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 531, + 482, + 580 + ], + "spans": [ + { + "bbox": [ + 130, + 531, + 482, + 580 + ], + "type": "text", + "content": "Class token-free models: Several existing patch reduction methods operate with the class token in ViTs to evaluate the patch relevance [12,28,34]. However, instead of class tokens, global pooling of patch tokens is used in many cases [16, 37]. Results presented in Tab. 3 demonstrate the superiority of PaPr for such cases." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 130, + 605, + 482, + 667 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 605, + 482, + 667 + ], + "spans": [ + { + "bbox": [ + 130, + 605, + 482, + 667 + ], + "type": "text", + "content": "Training-based methods comparison. We compare training performance of existing methods for the patch reduction on DeiT-s [48] model in Tab. 4. Prior pruning based methods cannot speed-up the training for learning the mask predictor [25,41,56]. However, PaPr achieves competitive performance as prior methods, while achieving large speed-up as other token merging methods [2,23,55]." + } + ] + } + ], + "index": 12 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 343, + 91, + 447, + 101 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 343, + 91, + 447, + 101 + ], + "spans": [ + { + "bbox": [ + 343, + 91, + 447, + 101 + ], + "type": "text", + "content": "PaPr for Faster Inference" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 470, + 91, + 479, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 470, + 91, + 479, + 100 + ], + "spans": [ + { + "bbox": [ + 470, + 91, + 479, + 100 + ], + "type": "text", + "content": "11" + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 10 + }, + { + "para_blocks": [ + { + "type": "table", + "bbox": [ + 144, + 188, + 299, + 269 + ], + "blocks": [ + { + "bbox": [ + 141, + 121, + 301, + 187 + ], + "lines": [ + { + "bbox": [ + 141, + 121, + 301, + 187 + ], + "spans": [ + { + "bbox": [ + 141, + 121, + 301, + 187 + ], + "type": "text", + "content": "Table 7: Sweeping ConvNet proposal model in PaPr with " + }, + { + "bbox": [ + 141, + 121, + 301, + 187 + ], + "type": "inline_equation", + "content": "z = 0.5" + }, + { + "bbox": [ + 141, + 121, + 301, + 187 + ], + "type": "text", + "content": ". PaPr can achieve similar performance irrespective of proposal model size. Thus, PaPr can use a much smaller and faster proposal model to speed-up larger models." + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 144, + 188, + 299, + 269 + ], + "lines": [ + { + "bbox": [ + 144, + 188, + 299, + 269 + ], + "spans": [ + { + "bbox": [ + 144, + 188, + 299, + 269 + ], + "type": "table", + "html": "
ProposalAccuracy1(%)
ModelGFLOPsViT-B-16ViT-L-16
ResNet-18 [18]1.8181.183.84
ResNet-50 [18]4.0982.3384.09
ResNet-152 [18]11.5182.5184.08
MobileOne-S0 [49]0.2782.2484.06
MobileOne-S2 [49]1.3582.2884.16
MobileOne-S4 [49]2.9882.3584.32
", + "image_path": "d78f4788d79411df097b55d05c9d8fc460101ec8d5382733ac373bb2d8c490fe.jpg" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "table_body" + } + ], + "index": 3 + }, + { + "type": "image", + "bbox": [ + 313, + 114, + 471, + 233 + ], + "blocks": [ + { + "bbox": [ + 313, + 114, + 471, + 233 + ], + "lines": [ + { + "bbox": [ + 313, + 114, + 471, + 233 + ], + "spans": [ + { + "bbox": [ + 313, + 114, + 471, + 233 + ], + "type": "image", + "image_path": "772790e3fcf8e48f2e99138562f527bb7eb88b103832ad836105ed30f82e88b5.jpg" + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 312, + 236, + 472, + 270 + ], + "lines": [ + { + "bbox": [ + 312, + 236, + 472, + 270 + ], + "spans": [ + { + "bbox": [ + 312, + 236, + 472, + 270 + ], + "type": "text", + "content": "Fig.6: Proposal models locate similar patches with PaPr for different pruning ratio " + }, + { + "bbox": [ + 312, + 236, + 472, + 270 + ], + "type": "inline_equation", + "content": "(z)" + }, + { + "bbox": [ + 312, + 236, + 472, + 270 + ], + "type": "text", + "content": " irrespective of model size." + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "image_caption" + } + ], + "index": 4 + }, + { + "type": "table", + "bbox": [ + 144, + 349, + 309, + 458 + ], + "blocks": [ + { + "bbox": [ + 136, + 284, + 314, + 340 + ], + "lines": [ + { + "bbox": [ + 136, + 284, + 314, + 340 + ], + "spans": [ + { + "bbox": [ + 136, + 284, + 314, + 340 + ], + "type": "text", + "content": "Table 8: Training-free performance comparison on Kinetics-400 video evaluation. Each video has " + }, + { + "bbox": [ + 136, + 284, + 314, + 340 + ], + "type": "inline_equation", + "content": "16 \\times 224^2" + }, + { + "bbox": [ + 136, + 284, + 314, + 340 + ], + "type": "text", + "content": " input size. PaPr achieves significantly better performance for reducing spatio-temporal redundancy in videos." + } + ] + } + ], + "index": 6, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 144, + 349, + 309, + 458 + ], + "lines": [ + { + "bbox": [ + 144, + 349, + 309, + 458 + ], + "spans": [ + { + "bbox": [ + 144, + 349, + 309, + 458 + ], + "type": "table", + "html": "
ModelMethodAcc1GFLOPsViews
XViTATS [12]80.02591x3
TimeSformer-LATS [12]80.535101x3
ViT-B MAEBaseline81.211803x5
PaPr(z=0.45)81.18763x5
PaPr(z=0.35)80.15593x5
ViT-L MAEBaseline85.265983x5
ToMe [2]84.52811x10
PaPr(z=0.5)85.122753x5
ToMe [2]82.51841x10
PaPr(z=0.3)84.531603x5
", + "image_path": "259bdc422f17add6668acccca35de691553f0c97b88bf7723b16e9da1b9abc20.jpg" + } + ] + } + ], + "index": 7, + "angle": 0, + "type": "table_body" + } + ], + "index": 7 + }, + { + "type": "table", + "bbox": [ + 319, + 350, + 478, + 457 + ], + "blocks": [ + { + "bbox": [ + 316, + 285, + 477, + 341 + ], + "lines": [ + { + "bbox": [ + 316, + 285, + 477, + 341 + ], + "spans": [ + { + "bbox": [ + 316, + 285, + 477, + 341 + ], + "type": "text", + "content": "Table 9: Proposal ConvNet sweep in video recognition with ViT-B-MAE model as the baseline. PaPr achieves competitive performance with lighter proposal models for different " + }, + { + "bbox": [ + 316, + 285, + 477, + 341 + ], + "type": "inline_equation", + "content": "z" + }, + { + "bbox": [ + 316, + 285, + 477, + 341 + ], + "type": "text", + "content": " values." + } + ] + } + ], + "index": 8, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 319, + 350, + 478, + 457 + ], + "lines": [ + { + "bbox": [ + 319, + 350, + 478, + 457 + ], + "spans": [ + { + "bbox": [ + 319, + 350, + 478, + 457 + ], + "type": "table", + "html": "
ProposalKeep Ratio(z)Acc1 (%)Views
ModelGFLOPs
X3d-s [13]1.250.581.113x5
X3d-m [13]2.450.581.193x5
ResNet-50 [14]41.90.579.963x5
ResNet-101 [14]85.670.579.973x5
X3d-s [13]1.250.480.863x5
X3d-m [13]2.450.480.933x5
ResNet-50 [14]41.90.478.853x5
ResNet-101 [14]85.670.479.323x5
", + "image_path": "99195efc99a2eb99a856fc53d0173d0c576c5e6800af19c84e1261f494ce2528.jpg" + } + ] + } + ], + "index": 9, + "angle": 0, + "type": "table_body" + } + ], + "index": 9 + }, + { + "bbox": [ + 132, + 483, + 387, + 495 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 483, + 387, + 495 + ], + "spans": [ + { + "bbox": [ + 132, + 483, + 387, + 495 + ], + "type": "text", + "content": "4.3 Performance on Various Hierarchical Models" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 130, + 509, + 482, + 558 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 509, + 482, + 558 + ], + "spans": [ + { + "bbox": [ + 130, + 509, + 482, + 558 + ], + "type": "text", + "content": "We study the performance on two variants of hierarchical models, such as pure convolutional models, and hybrid transformer models. We compare with training based DynamicCNN [40], and DynamicSwin [40] methods (trained for 120 epochs), whereas PaPr operates without training." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 130, + 581, + 483, + 666 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 581, + 483, + 666 + ], + "spans": [ + { + "bbox": [ + 130, + 581, + 483, + 666 + ], + "type": "text", + "content": "Analysis on convolutional models. We use the SOTA ConvNext [32] architecture for the analysis as shown in Tab. 5. The training-free operation in PaPr makes it seamlessly usable with new model updates. We analyze ImageNet-1k and ImageNet-22k performance of same models. For the ConvNext-Base model, PaPr achieves " + }, + { + "bbox": [ + 130, + 581, + 483, + 666 + ], + "type": "inline_equation", + "content": "99.6\\%" + }, + { + "bbox": [ + 130, + 581, + 483, + 666 + ], + "type": "text", + "content": " accuracy of DynamicCNN with " + }, + { + "bbox": [ + 130, + 581, + 483, + 666 + ], + "type": "inline_equation", + "content": "7.1\\%" + }, + { + "bbox": [ + 130, + 581, + 483, + 666 + ], + "type": "text", + "content": " higher throughput. By simply using ImageNet-22k weights, PaPr can achieve additional " + }, + { + "bbox": [ + 130, + 581, + 483, + 666 + ], + "type": "inline_equation", + "content": "2\\%" + }, + { + "bbox": [ + 130, + 581, + 483, + 666 + ], + "type": "text", + "content": " accuracy improvement without training, while having the same computation cost." + } + ] + } + ], + "index": 12 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "spans": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "type": "text", + "content": "12" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 166, + 91, + 240, + 101 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 166, + 91, + 240, + 101 + ], + "spans": [ + { + "bbox": [ + 166, + 91, + 240, + 101 + ], + "type": "text", + "content": "T. Mahmud et al." + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 11 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 157, + 114, + 455, + 294 + ], + "blocks": [ + { + "bbox": [ + 157, + 114, + 455, + 294 + ], + "lines": [ + { + "bbox": [ + 157, + 114, + 455, + 294 + ], + "spans": [ + { + "bbox": [ + 157, + 114, + 455, + 294 + ], + "type": "image", + "image_path": "fe51584d2962fd0367d1dce2c621ada35d997467a3e559bfa97aa8b08f5a3857.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 130, + 298, + 482, + 343 + ], + "lines": [ + { + "bbox": [ + 130, + 298, + 482, + 343 + ], + "spans": [ + { + "bbox": [ + 130, + 298, + 482, + 343 + ], + "type": "text", + "content": "Fig. 7: Robustness of PaPr compared to CAM based methods. PaPr can perform even when the ConvNet proposal confidence (c) is very low. In contrast, existing CAM based methods fail in such cases, despite being significantly slower while not enabling batch processing and use of gradients in some cases." + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_caption" + } + ], + "index": 2 + }, + { + "bbox": [ + 130, + 365, + 483, + 426 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 365, + 483, + 426 + ], + "spans": [ + { + "bbox": [ + 130, + 365, + 483, + 426 + ], + "type": "text", + "content": "Analysis on hybrid transformer models. We study the hierarchical Swin transformer models for patch reduction with PaPr, as given in Tab. 6. PaPr achieves " + }, + { + "bbox": [ + 130, + 365, + 483, + 426 + ], + "type": "inline_equation", + "content": "98\\%" + }, + { + "bbox": [ + 130, + 365, + 483, + 426 + ], + "type": "text", + "content": " of DynamicSwin accuracy using similar FLOPs without re-training. Nevertheless, by leveraging bigger models, larger pre-training, and higher resolution, PaPr can seamlessly adapt to achieve higher accuracy." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 130, + 442, + 433, + 455 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 442, + 433, + 455 + ], + "spans": [ + { + "bbox": [ + 130, + 442, + 433, + 455 + ], + "type": "text", + "content": "4.4 Robustness of PaPr across various ConvNet proposals" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 130, + 461, + 482, + 546 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 461, + 482, + 546 + ], + "spans": [ + { + "bbox": [ + 130, + 461, + 482, + 546 + ], + "type": "text", + "content": "PaPr can operate with ultra-lightweight ConvNets for generating robust proposals. We study different ConvNet architectures for proposal generation, as shown in Tab. 7. Interestingly, PaPr shows small reductions of final accuracy " + }, + { + "bbox": [ + 130, + 461, + 482, + 546 + ], + "type": "inline_equation", + "content": "(0.3\\%)" + }, + { + "bbox": [ + 130, + 461, + 482, + 546 + ], + "type": "text", + "content": ", when using MobileOne-S0 based proposals compared to ResNet-152 " + }, + { + "bbox": [ + 130, + 461, + 482, + 546 + ], + "type": "inline_equation", + "content": "(42\\times" + }, + { + "bbox": [ + 130, + 461, + 482, + 546 + ], + "type": "text", + "content": " higher FLOPs) in ViT-B. When visualizing the PSM for different pruning ratio as in Fig. 6, we notice the similar PSMs irrespective of model size. Hence, PaPr can utilize ultra-lightweight ConvNets to mask patches without sacrificing accuracy." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 130, + 562, + 430, + 575 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 562, + 430, + 575 + ], + "spans": [ + { + "bbox": [ + 130, + 562, + 430, + 575 + ], + "type": "text", + "content": "4.5 Comparison with Class Activation Mappings (CAMs)" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 130, + 582, + 482, + 667 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 582, + 482, + 667 + ], + "spans": [ + { + "bbox": [ + 130, + 582, + 482, + 667 + ], + "type": "text", + "content": "Existing CAM methods mostly focus on explainability to highlight image regions responsible for final prediction. However, such objectives rely on final prediction performance, which can be much lower for light ConvNets. Moreover, for localization, many of these methods use gradients, and optimization methods that limit batch processing. Nevertheless, we study the impact of such CAM methods ([43, 50, 61]) in PaPr framework, in challenging scenarios where the baseline ViT-B has higher prediction confidence (c) on target class, and the proposal" + } + ] + } + ], + "index": 8 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 343, + 91, + 447, + 100 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 343, + 91, + 447, + 100 + ], + "spans": [ + { + "bbox": [ + 343, + 91, + 447, + 100 + ], + "type": "text", + "content": "PaPr for Faster Inference" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 470, + 91, + 481, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 470, + 91, + 481, + 100 + ], + "spans": [ + { + "bbox": [ + 470, + 91, + 481, + 100 + ], + "type": "text", + "content": "13" + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 12 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 159, + 114, + 455, + 190 + ], + "blocks": [ + { + "bbox": [ + 159, + 114, + 455, + 190 + ], + "lines": [ + { + "bbox": [ + 159, + 114, + 455, + 190 + ], + "spans": [ + { + "bbox": [ + 159, + 114, + 455, + 190 + ], + "type": "image", + "image_path": "58041a1a8ab5a9067821c623a7a89ccb36876571cfd2c6359efec62cb69c21d8.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 130, + 191, + 482, + 236 + ], + "lines": [ + { + "bbox": [ + 130, + 191, + 482, + 236 + ], + "spans": [ + { + "bbox": [ + 130, + 191, + 482, + 236 + ], + "type": "text", + "content": "Fig. 8: Visualization of PaPr localization in videos. Video has inherent high sparsity. PaPr effectively localizes the discriminative regions for using holistic spatio-temporal understanding with small ConvNets. Thus, it significantly reduces the computational burden for larger models in downstream video recognition tasks." + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_caption" + } + ], + "index": 2 + }, + { + "bbox": [ + 130, + 258, + 482, + 319 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 258, + 482, + 319 + ], + "spans": [ + { + "bbox": [ + 130, + 258, + 482, + 319 + ], + "type": "text", + "content": "MobileOne-S0 has lower confidence (see Fig. 7). In most cases, the baseline CAM method significantly lowers the final accuracy after patch pruning. In contrast, PaPr maintains robust confidence for its precise localization. Moreover, PaPr can enhance confidence of baseline models in several scenarios by suppressing the redundant patches, e.g., in Sample 5, the " + }, + { + "bbox": [ + 130, + 258, + 482, + 319 + ], + "type": "inline_equation", + "content": "c" + }, + { + "bbox": [ + 130, + 258, + 482, + 319 + ], + "type": "text", + "content": " increased by " + }, + { + "bbox": [ + 130, + 258, + 482, + 319 + ], + "type": "inline_equation", + "content": "22\\%" + }, + { + "bbox": [ + 130, + 258, + 482, + 319 + ], + "type": "text", + "content": " with PaPr." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 132, + 336, + 269, + 351 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 336, + 269, + 351 + ], + "spans": [ + { + "bbox": [ + 132, + 336, + 269, + 351 + ], + "type": "text", + "content": "5 Video Experiments" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 130, + 360, + 483, + 529 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 360, + 483, + 529 + ], + "spans": [ + { + "bbox": [ + 130, + 360, + 483, + 529 + ], + "type": "text", + "content": "We study the training-free performance on Kinetics-400 [22] validation sets, as shown in Tab. 8. We use SOTA ATS [12], and ToMe [2] methods for comparison. We use PaPr on SOTA ViT-MAE [47] models with lightweight X3d-s [13] model for proposal generation. We follow the baseline [47] " + }, + { + "bbox": [ + 130, + 360, + 483, + 529 + ], + "type": "inline_equation", + "content": "3 \\times 5" + }, + { + "bbox": [ + 130, + 360, + 483, + 529 + ], + "type": "text", + "content": " view (3 spatial view and 5 temporal view) approach, and separate the views from model FLOP counts similar to other works as it can be arbitrarily chosen. Since ViT-MAE removes class tokens in fine-tuning, ATS [12] cannot be adapted to these models. For ViT-L, PaPr achieves " + }, + { + "bbox": [ + 130, + 360, + 483, + 529 + ], + "type": "inline_equation", + "content": "3.7 \\times" + }, + { + "bbox": [ + 130, + 360, + 483, + 529 + ], + "type": "text", + "content": " FLOPs reduction of the baseline model for a negligible " + }, + { + "bbox": [ + 130, + 360, + 483, + 529 + ], + "type": "inline_equation", + "content": "0.8\\%" + }, + { + "bbox": [ + 130, + 360, + 483, + 529 + ], + "type": "text", + "content": " accuracy drop. We also study the impact of various proposal models on the final performance, as shown in Tab. 9. We use SlowOnly-ResNet [14] and X3d [13] models for comparison. Increasing model size for different pruning shows minimal impact on final performance. Finally, we visualize the patch masking of PaPr in videos, as shown in Fig. 8. By removing the redundant details, PaPr significantly boosts the performance of bigger models for fine-grained predictions." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 132, + 545, + 326, + 558 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 545, + 326, + 558 + ], + "spans": [ + { + "bbox": [ + 132, + 545, + 326, + 558 + ], + "type": "text", + "content": "6 Conclusion and Future Work" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 130, + 570, + 482, + 666 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 570, + 482, + 666 + ], + "spans": [ + { + "bbox": [ + 130, + 570, + 482, + 666 + ], + "type": "text", + "content": "In this paper, we introduce a novel patch pruning method, namely PaPr, that can effectively speed-up off-the-shelf pre-trained models inference without re-training. We propose a simple modification of ConvNets, that allows extracting a precise discriminative PSM from ultra lightweight ConvNets with unparalleled speed. In PaPr, such PSMs are leveraged to directly reduce the redundant data regions in a single-step to guide larger model computations on most discriminative regions for achieving computational efficiency. Moreover, PaPr can be easily integrated with most existing patch reduction methods due to its simple structure." + } + ] + } + ], + "index": 8 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "spans": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "type": "text", + "content": "14" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 166, + 91, + 240, + 100 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 166, + 91, + 240, + 100 + ], + "spans": [ + { + "bbox": [ + 166, + 91, + 240, + 100 + ], + "type": "text", + "content": "T. Mahmud et al." + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 13 + }, + { + "para_blocks": [ + { + "bbox": [ + 133, + 114, + 241, + 129 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 114, + 241, + 129 + ], + "spans": [ + { + "bbox": [ + 133, + 114, + 241, + 129 + ], + "type": "text", + "content": "Acknowledgement" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 130, + 137, + 482, + 174 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 137, + 482, + 174 + ], + "spans": [ + { + "bbox": [ + 130, + 137, + 482, + 174 + ], + "type": "text", + "content": "This work was supported in part by ONR Minerva program, iMAGiNE - the Intelligent Machine Engineering Consortium at UT Austin, and a UT Cockrell School of Engineering Doctoral Fellowship." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 133, + 190, + 198, + 202 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 190, + 198, + 202 + ], + "spans": [ + { + "bbox": [ + 133, + 190, + 198, + 202 + ], + "type": "text", + "content": "References" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 138, + 211, + 481, + 665 + ], + "type": "list", + "angle": 0, + "index": 19, + "blocks": [ + { + "bbox": [ + 138, + 211, + 481, + 255 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 211, + 481, + 255 + ], + "spans": [ + { + "bbox": [ + 138, + 211, + 481, + 255 + ], + "type": "text", + "content": "1. Belharbi, S., Sarraf, A., Pedersoli, M., Ben Ayed, I., McCaffrey, L., Granger, E.: F-cam: Full resolution class activation maps via guided parametric upscaling. In: Proceedings of the IEEE/CVF Winter Conference on Applications of Computer Vision. pp. 3490-3499 (2022)" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 138, + 256, + 481, + 288 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 256, + 481, + 288 + ], + "spans": [ + { + "bbox": [ + 138, + 256, + 481, + 288 + ], + "type": "text", + "content": "2. Bolya, D., Fu, C.Y., Dai, X., Zhang, P., Feichtenhofer, C., Hoffman, J.: Token merging: Your vit but faster. In: The Twelfth International Conference on Learning Representations (2023)" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 138, + 289, + 481, + 309 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 289, + 481, + 309 + ], + "spans": [ + { + "bbox": [ + 138, + 289, + 481, + 309 + ], + "type": "text", + "content": "3. Cao, Q., Paranjape, B., Hajishirzi, H.: Pumer: Pruning and merging tokens for efficient vision language models. arXiv preprint arXiv:2305.17530 (2023)" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 138, + 310, + 481, + 342 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 310, + 481, + 342 + ], + "spans": [ + { + "bbox": [ + 138, + 310, + 481, + 342 + ], + "type": "text", + "content": "4. Caron, M., Touvron, H., Misra, I., Jégou, H., Mairal, J., Bojanowski, P., Joulin, A.: Emerging properties in self-supervised vision transformers. In: Proceedings of the IEEE/CVF international conference on computer vision. pp. 9650-9660 (2021)" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 138, + 342, + 481, + 384 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 342, + 481, + 384 + ], + "spans": [ + { + "bbox": [ + 138, + 342, + 481, + 384 + ], + "type": "text", + "content": "5. Chattopadhyay, A., Sarkar, A., Howlader, P., Balasubramanian, V.N.: Grad-cam++: Generalized gradient-based visual explanations for deep convolutional networks. In: 2018 IEEE winter conference on applications of computer vision (WACV). pp. 839-847. IEEE (2018)" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 138, + 385, + 481, + 418 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 385, + 481, + 418 + ], + "spans": [ + { + "bbox": [ + 138, + 385, + 481, + 418 + ], + "type": "text", + "content": "6. Chen, Y., Liu, H., Yin, H., Fan, B.: Building vision transformers with hierarchy aware feature aggregation. In: Proceedings of the IEEE/CVF International Conference on Computer Vision (ICCV). pp. 5908-5918 (October 2023)" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 138, + 418, + 481, + 449 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 418, + 481, + 449 + ], + "spans": [ + { + "bbox": [ + 138, + 418, + 481, + 449 + ], + "type": "text", + "content": "7. Chu, X., Tian, Z., Wang, Y., Zhang, B., Ren, H., Wei, X., Xia, H., Shen, C.: Twins: Revisiting the design of spatial attention in vision transformers. Advances in Neural Information Processing Systems 34, 9355-9366 (2021)" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 138, + 450, + 481, + 482 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 450, + 481, + 482 + ], + "spans": [ + { + "bbox": [ + 138, + 450, + 481, + 482 + ], + "type": "text", + "content": "8. Ding, X., Zhang, X., Ma, N., Han, J., Ding, G., Sun, J.: Repvgg: Making vgg-style convnets great again. In: Proceedings of the IEEE/CVF conference on computer vision and pattern recognition. pp. 13733-13742 (2021)" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 138, + 483, + 481, + 514 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 483, + 481, + 514 + ], + "spans": [ + { + "bbox": [ + 138, + 483, + 481, + 514 + ], + "type": "text", + "content": "9. Dong, X., Bao, J., Chen, D., Zhang, W., Yu, N., Yuan, L., Chen, D., Guo, B.: Cswin transformer: A general vision transformer backbone with cross-shaped windows. In: CVPR (2022)" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 138, + 514, + 481, + 547 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 514, + 481, + 547 + ], + "spans": [ + { + "bbox": [ + 138, + 514, + 481, + 547 + ], + "type": "text", + "content": "0. Dosovitskiy, A., Beyer, L., Kolesnikov, A., Weissenborn, D., Zhai, X., Unterthiner, T., Dehghani, M., Minderer, M., Heigold, G., Gelly, S., et al.: An image is worth 16x16 words: Transformers for image recognition at scale. In: ICLR (2020)" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 138, + 547, + 481, + 568 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 547, + 481, + 568 + ], + "spans": [ + { + "bbox": [ + 138, + 547, + 481, + 568 + ], + "type": "text", + "content": "1. Fan, H., Xiong, B., Mangalam, K., Li, Y., Yan, Z., Malik, J., Feichtenhofer, C.: Multiscale vision transformers. In: ICCV (2021)" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 138, + 569, + 481, + 601 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 569, + 481, + 601 + ], + "spans": [ + { + "bbox": [ + 138, + 569, + 481, + 601 + ], + "type": "text", + "content": "2. Fayyaz, M., Kouhpayegani, S.A., Jafari, F.R., Sommerlade, E., Joze, H.R.V., Pirsivash, H., Gall, J.: Ats: Adaptive token sampling for efficient vision transformers. In: ECCV (2022)" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 138, + 601, + 481, + 633 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 601, + 481, + 633 + ], + "spans": [ + { + "bbox": [ + 138, + 601, + 481, + 633 + ], + "type": "text", + "content": "3. Feichtenhofer, C.: X3d: Expanding architectures for efficient video recognition. In: Proceedings of the IEEE/CVF conference on computer vision and pattern recognition. pp. 203-213 (2020)" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 138, + 634, + 481, + 665 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 634, + 481, + 665 + ], + "spans": [ + { + "bbox": [ + 138, + 634, + 481, + 665 + ], + "type": "text", + "content": "4. Feichtenhofer, C., Fan, H., Malik, J., He, K.: Slowfast networks for video recognition. In: Proceedings of the IEEE/CVF international conference on computer vision. pp. 6202-6211 (2019)" + } + ] + } + ], + "index": 18 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 343, + 91, + 447, + 101 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 343, + 91, + 447, + 101 + ], + "spans": [ + { + "bbox": [ + 343, + 91, + 447, + 101 + ], + "type": "text", + "content": "PaPr for Faster Inference" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 470, + 92, + 481, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 470, + 92, + 481, + 100 + ], + "spans": [ + { + "bbox": [ + 470, + 92, + 481, + 100 + ], + "type": "text", + "content": "15" + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 14 + }, + { + "para_blocks": [ + { + "bbox": [ + 132, + 116, + 481, + 665 + ], + "type": "list", + "angle": 0, + "index": 20, + "blocks": [ + { + "bbox": [ + 133, + 116, + 481, + 149 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 116, + 481, + 149 + ], + "spans": [ + { + "bbox": [ + 133, + 116, + 481, + 149 + ], + "type": "text", + "content": "15. Graham, B., El-Nouby, A., Touvron, H., Stock, P., Joulin, A., Jégou, H., Douze, M.: Levit: a vision transformer in convnet's clothing for faster inference. In: ICCV (2021)" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 133, + 150, + 481, + 183 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 150, + 481, + 183 + ], + "spans": [ + { + "bbox": [ + 133, + 150, + 481, + 183 + ], + "type": "text", + "content": "16. Hatamizadeh, A., Yin, H., Heinrich, G., Kautz, J., Molchanov, P.: Global context vision transformers. In: International Conference on Machine Learning. pp. 12633-12646. PMLR (2023)" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 132, + 183, + 481, + 205 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 183, + 481, + 205 + ], + "spans": [ + { + "bbox": [ + 132, + 183, + 481, + 205 + ], + "type": "text", + "content": "17. He, K., Chen, X., Xie, S., Li, Y., Dollar, P., Girshick, R.: Masked autoencoders are scalable vision learners. In: CVPR (2022)" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 132, + 205, + 481, + 237 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 205, + 481, + 237 + ], + "spans": [ + { + "bbox": [ + 132, + 205, + 481, + 237 + ], + "type": "text", + "content": "18. He, K., Zhang, X., Ren, S., Sun, J.: Deep residual learning for image recognition. In: Proceedings of the IEEE conference on computer vision and pattern recognition. pp. 770-778 (2016)" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 132, + 238, + 481, + 270 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 238, + 481, + 270 + ], + "spans": [ + { + "bbox": [ + 132, + 238, + 481, + 270 + ], + "type": "text", + "content": "19. Howard, A., Sandler, M., Chu, G., Chen, L.C., Chen, B., Tan, M., Wang, W., Zhu, Y., Pang, R., Vasudevan, V., et al.: Searching for mobilenetv3. In: Proceedings of the IEEE/CVF international conference on computer vision. pp. 1314-1324 (2019)" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 132, + 271, + 481, + 304 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 271, + 481, + 304 + ], + "spans": [ + { + "bbox": [ + 132, + 271, + 481, + 304 + ], + "type": "text", + "content": "20. Howard, A.G., Zhu, M., Chen, B., Kalenichenko, D., Wang, W., Weyand, T., Andreetto, M., Adam, H.: Mobilenets: Efficient convolutional neural networks for mobile vision applications. arXiv preprint arXiv:1704.04861 (2017)" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 132, + 304, + 481, + 335 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 304, + 481, + 335 + ], + "spans": [ + { + "bbox": [ + 132, + 304, + 481, + 335 + ], + "type": "text", + "content": "21. Jung, H., Oh, Y.: Towards better explanations of class activation mapping. In: Proceedings of the IEEE/CVF International Conference on Computer Vision. pp. 1336-1344 (2021)" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 132, + 336, + 481, + 369 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 336, + 481, + 369 + ], + "spans": [ + { + "bbox": [ + 132, + 336, + 481, + 369 + ], + "type": "text", + "content": "22. Kay, W., Carreira, J., Simonyan, K., Zhang, B., Hillier, C., Vijayanarasimhan, S., Viola, F., Green, T., Back, T., Natsev, P., et al.: The kinetics human action video dataset. arXiv:1705.06950 [cs.CV] (2017)" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 132, + 370, + 481, + 403 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 370, + 481, + 403 + ], + "spans": [ + { + "bbox": [ + 132, + 370, + 481, + 403 + ], + "type": "text", + "content": "23. Kim, M., Gao, S., Hsu, Y.C., Shen, Y., Jin, H.: Token fusion: Bridging the gap between token pruning and token merging. In: Proceedings of the IEEE/CVF Winter Conference on Applications of Computer Vision. pp. 1383-1392 (2024)" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 132, + 403, + 481, + 424 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 403, + 481, + 424 + ], + "spans": [ + { + "bbox": [ + 132, + 403, + 481, + 424 + ], + "type": "text", + "content": "24. Kim, S., Shen, S., Thorsley, D., Gholami, A., Kwon, W., Hassoun, J., Keutzer, K.: Learned token pruning for transformers. arXiv:2107.00910 [cs.CL] (2021)" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 132, + 425, + 481, + 456 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 425, + 481, + 456 + ], + "spans": [ + { + "bbox": [ + 132, + 425, + 481, + 456 + ], + "type": "text", + "content": "25. Kong, Z., Dong, P., Ma, X., Meng, X., Niu, W., Sun, M., Ren, B., Qin, M., Tang, H., Wang, Y.: Spvit: Enabling faster vision transformers via soft token pruning. In: ECCV (2022)" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 132, + 457, + 481, + 479 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 457, + 481, + 479 + ], + "spans": [ + { + "bbox": [ + 132, + 457, + 481, + 479 + ], + "type": "text", + "content": "26. Lassance, C., Maachou, M., Park, J., Clinchant, S.: A study on token pruning for colbert. arXiv:2112.06540 [cs.CL] (2021)" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 132, + 479, + 481, + 511 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 479, + 481, + 511 + ], + "spans": [ + { + "bbox": [ + 132, + 479, + 481, + 511 + ], + "type": "text", + "content": "27. Li, Y., Wu, C.Y., Fan, H., Mangalam, K., Xiong, B., Malik, J., Feichtenhofer, C.: Mvittv2: Improved multiscale vision transformers for classification and detection. In: CVPR (2022)" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 132, + 512, + 481, + 533 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 512, + 481, + 533 + ], + "spans": [ + { + "bbox": [ + 132, + 512, + 481, + 533 + ], + "type": "text", + "content": "28. Liang, Y., Ge, C., Tong, Z., Song, Y., Wang, J., Xie, P.: Not all patches are what you need: Expediting vision transformers via token reorganizations. ICLR (2022)" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 132, + 534, + 481, + 566 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 534, + 481, + 566 + ], + "spans": [ + { + "bbox": [ + 132, + 534, + 481, + 566 + ], + "type": "text", + "content": "29. Lin, J., Chen, W.M., Cai, H., Gan, C., Han, S.: Memory-efficient patch-based inference for tiny deep learning. Advances in Neural Information Processing Systems 34, 2346-2358 (2021)" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 132, + 567, + 481, + 599 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 567, + 481, + 599 + ], + "spans": [ + { + "bbox": [ + 132, + 567, + 481, + 599 + ], + "type": "text", + "content": "30. Liu, Z., Hu, H., Lin, Y., Yao, Z., Xie, Z., Wei, Y., Ning, J., Cao, Y., Zhang, Z., Dong, L., et al.: Swin transformer v2: Scaling up capacity and resolution. In: CVPR (2022)" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 132, + 600, + 481, + 632 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 600, + 481, + 632 + ], + "spans": [ + { + "bbox": [ + 132, + 600, + 481, + 632 + ], + "type": "text", + "content": "31. Liu, Z., Lin, Y., Cao, Y., Hu, H., Wei, Y., Zhang, Z., Lin, S., Guo, B.: Swin transformer: Hierarchical vision transformer using shifted windows. In: ICCV (2021)" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 132, + 632, + 481, + 665 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 632, + 481, + 665 + ], + "spans": [ + { + "bbox": [ + 132, + 632, + 481, + 665 + ], + "type": "text", + "content": "32. Liu, Z., Mao, H., Wu, C.Y., Feichtenhofer, C., Darrell, T., Xie, S.: A convnet for the 2020s. In: Proceedings of the IEEE/CVF conference on computer vision and pattern recognition. pp. 11976-11986 (2022)" + } + ] + } + ], + "index": 19 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "spans": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "type": "text", + "content": "16" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 166, + 91, + 239, + 101 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 166, + 91, + 239, + 101 + ], + "spans": [ + { + "bbox": [ + 166, + 91, + 239, + 101 + ], + "type": "text", + "content": "T. Mahmud et al." + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 15 + }, + { + "para_blocks": [ + { + "bbox": [ + 132, + 116, + 482, + 665 + ], + "type": "list", + "angle": 0, + "index": 20, + "blocks": [ + { + "bbox": [ + 132, + 116, + 482, + 138 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 116, + 482, + 138 + ], + "spans": [ + { + "bbox": [ + 132, + 116, + 482, + 138 + ], + "type": "text", + "content": "33. Mehta, S., Rastegari, M.: Mobilevit: light-weight, general-purpose, and mobile-friendly vision transformer. In: ICLR (2021)" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 132, + 138, + 482, + 160 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 138, + 482, + 160 + ], + "spans": [ + { + "bbox": [ + 132, + 138, + 482, + 160 + ], + "type": "text", + "content": "34. Meng, L., Li, H., Chen, B.C., Lan, S., Wu, Z., Jiang, Y.G., Lim, S.N.: Adavit: Adaptive vision transformers for efficient image recognition. In: CVPR (2022)" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 132, + 160, + 482, + 194 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 160, + 482, + 194 + ], + "spans": [ + { + "bbox": [ + 132, + 160, + 482, + 194 + ], + "type": "text", + "content": "35. Muhammad, M.B., Yeasin, M.: Eigen-cam: Class activation map using principal components. In: 2020 international joint conference on neural networks (IJCNN). pp. 1-7. IEEE (2020)" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 132, + 194, + 482, + 226 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 194, + 482, + 226 + ], + "spans": [ + { + "bbox": [ + 132, + 194, + 482, + 226 + ], + "type": "text", + "content": "36. Oquab, M., Darcet, T., Moutakanni, T., Vo, H., Szafraniec, M., Khalidov, V., Fernandez, P., Haziza, D., Massa, F., El-Nouby, A., et al.: Dinov2: Learning robust visual features without supervision. arXiv preprint arXiv:2304.07193 (2023)" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 132, + 226, + 482, + 258 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 226, + 482, + 258 + ], + "spans": [ + { + "bbox": [ + 132, + 226, + 482, + 258 + ], + "type": "text", + "content": "37. Pan, Z., Zhuang, B., Liu, J., He, H., Cai, J.: Scalable vision transformers with hierarchical pooling. In: Proceedings of the IEEE/cvf international conference on computer vision. pp. 377-386 (2021)" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 132, + 259, + 482, + 293 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 259, + 482, + 293 + ], + "spans": [ + { + "bbox": [ + 132, + 259, + 482, + 293 + ], + "type": "text", + "content": "38. Patro, B.N., Lunayach, M., Patel, S., Namboodiri, V.P.: U-cam: Visual explanation using uncertainty based class activation maps. In: Proceedings of the IEEE/CVF International Conference on Computer Vision. pp. 7444-7453 (2019)" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 132, + 293, + 482, + 335 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 293, + 482, + 335 + ], + "spans": [ + { + "bbox": [ + 132, + 293, + 482, + 335 + ], + "type": "text", + "content": "39. Radford, A., Kim, J.W., Hallacy, C., Ramesh, A., Goh, G., Agarwal, S., Sastry, G., Askell, A., Mishkin, P., Clark, J., et al.: Learning transferable visual models from natural language supervision. In: International conference on machine learning. pp. 8748-8763. PMLR (2021)" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 132, + 335, + 482, + 369 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 335, + 482, + 369 + ], + "spans": [ + { + "bbox": [ + 132, + 335, + 482, + 369 + ], + "type": "text", + "content": "40. Rao, Y., Liu, Z., Zhao, W., Zhou, J., Lu, J.: Dynamic spatial sparsification for efficient vision transformers and convolutional neural networks. IEEE Transactions on Pattern Analysis and Machine Intelligence (2023)" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 132, + 369, + 482, + 403 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 369, + 482, + 403 + ], + "spans": [ + { + "bbox": [ + 132, + 369, + 482, + 403 + ], + "type": "text", + "content": "41. Rao, Y., Zhao, W., Liu, B., Lu, J., Zhou, J., Hsieh, C.J.: Dynamicvit: Efficient vision transformers with dynamic token sparsification. Advances in neural information processing systems 34, 13937-13949 (2021)" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 132, + 403, + 482, + 435 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 403, + 482, + 435 + ], + "spans": [ + { + "bbox": [ + 132, + 403, + 482, + 435 + ], + "type": "text", + "content": "42. Sandler, M., Howard, A., Zhu, M., Zhmoginov, A., Chen, L.C.: Mobilenetv2: Inverted residuals and linear bottlenecks. In: Proceedings of the IEEE conference on computer vision and pattern recognition. pp. 4510-4520 (2018)" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 132, + 435, + 482, + 468 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 435, + 482, + 468 + ], + "spans": [ + { + "bbox": [ + 132, + 435, + 482, + 468 + ], + "type": "text", + "content": "43. Selvaraju, R.R., Cogswell, M., Das, A., Vedantam, R., Parikh, D., Batra, D.: Gradcam: Visual explanations from deep networks via gradient-based localization. In: ICCV (2017)" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 132, + 468, + 482, + 501 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 468, + 482, + 501 + ], + "spans": [ + { + "bbox": [ + 132, + 468, + 482, + 501 + ], + "type": "text", + "content": "44. Singh, M., Gustafson, L., Adcock, A., de Freitas Reis, V., Gedik, B., Kosaraju, R.P., Mahajan, D., Girshick, R., Dollar, P., van der Maaten, L.: Revisiting weakly supervised pre-training of visual perception models. In: CVPR (2022)" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 132, + 501, + 482, + 533 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 501, + 482, + 533 + ], + "spans": [ + { + "bbox": [ + 132, + 501, + 482, + 533 + ], + "type": "text", + "content": "45. Steiner, A., Kolesnikov, A., Zhai, X., Wightman, R., Uszkoreit, J., Beyer, L.: How to train your vit? data, augmentation, and regularization in vision transformers. TMLR (2022)" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 132, + 533, + 482, + 555 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 533, + 482, + 555 + ], + "spans": [ + { + "bbox": [ + 132, + 533, + 482, + 555 + ], + "type": "text", + "content": "46. Tan, M., Le, Q.: Efficientnet: Rethinking model scaling for convolutional neural networks. In: ICML (2019)" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 132, + 555, + 482, + 588 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 555, + 482, + 588 + ], + "spans": [ + { + "bbox": [ + 132, + 555, + 482, + 588 + ], + "type": "text", + "content": "47. Tong, Z., Song, Y., Wang, J., Wang, L.: Videomae: Masked autoencoders are data-efficient learners for self-supervised video pre-training. arXiv:2203.12602 [cs.CV] (2022)" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 132, + 588, + 482, + 611 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 588, + 482, + 611 + ], + "spans": [ + { + "bbox": [ + 132, + 588, + 482, + 611 + ], + "type": "text", + "content": "48. Touvron, H., Cord, M., Douze, M., Massa, F., Sablayrolles, A., Jégou, H.: Training data-efficient image transformers & distillation through attention. In: ICML (2021)" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 132, + 611, + 482, + 643 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 611, + 482, + 643 + ], + "spans": [ + { + "bbox": [ + 132, + 611, + 482, + 643 + ], + "type": "text", + "content": "49. Vasu, P.K.A., Gabriel, J., Zhu, J., Tuzel, O., Ranjan, A.: Mobileone: An improved one millisecond mobile backbone. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 7907-7917 (2023)" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 132, + 643, + 482, + 665 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 643, + 482, + 665 + ], + "spans": [ + { + "bbox": [ + 132, + 643, + 482, + 665 + ], + "type": "text", + "content": "50. Wang, H., Wang, Z., Du, M., Yang, F., Zhang, Z., Ding, S., Mardziel, P., Hu, X.: Score-cam: Score-weighted visual explanations for convolutional neural networks." + } + ] + } + ], + "index": 19 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 343, + 91, + 447, + 100 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 343, + 91, + 447, + 100 + ], + "spans": [ + { + "bbox": [ + 343, + 91, + 447, + 100 + ], + "type": "text", + "content": "PaPr for Faster Inference" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 470, + 91, + 481, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 470, + 91, + 481, + 100 + ], + "spans": [ + { + "bbox": [ + 470, + 91, + 481, + 100 + ], + "type": "text", + "content": "17" + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 16 + }, + { + "para_blocks": [ + { + "bbox": [ + 132, + 116, + 481, + 501 + ], + "type": "list", + "angle": 0, + "index": 14, + "blocks": [ + { + "bbox": [ + 147, + 116, + 481, + 138 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 147, + 116, + 481, + 138 + ], + "spans": [ + { + "bbox": [ + 147, + 116, + 481, + 138 + ], + "type": "text", + "content": "In: Proceedings of the IEEE/CVF conference on computer vision and pattern recognition workshops. pp. 24-25 (2020)" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 133, + 139, + 481, + 160 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 139, + 481, + 160 + ], + "spans": [ + { + "bbox": [ + 133, + 139, + 481, + 160 + ], + "type": "text", + "content": "51. Wang, S., Li, B.Z., Khabsa, M., Fang, H., Ma, H.: Linformer: Self-attention with linear complexity. arXiv preprint arXiv:2006.04768 (2020)" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 133, + 161, + 481, + 194 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 161, + 481, + 194 + ], + "spans": [ + { + "bbox": [ + 133, + 161, + 481, + 194 + ], + "type": "text", + "content": "52. Wei, C., Fan, H., Xie, S., Wu, C.Y., Yuille, A., Feichtenhofer, C.: Masked feature prediction for self-supervised visual pre-training. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 14668-14678 (2022)" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 132, + 194, + 481, + 236 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 194, + 481, + 236 + ], + "spans": [ + { + "bbox": [ + 132, + 194, + 481, + 236 + ], + "type": "text", + "content": "53. Woo, S., Debnath, S., Hu, R., Chen, X., Liu, Z., Kweon, I.S., Xie, S.: Convnext v2: Co-designing and scaling convnets with masked autoencoders. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 16133-16142 (2023)" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 132, + 237, + 481, + 270 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 237, + 481, + 270 + ], + "spans": [ + { + "bbox": [ + 132, + 237, + 481, + 270 + ], + "type": "text", + "content": "54. Wu, H., Xiao, B., Codella, N., Liu, M., Dai, X., Yuan, L., Zhang, L.: Cvt: Introducing convolutions to vision transformers. In: Proceedings of the IEEE/CVF international conference on computer vision. pp. 22-31 (2021)" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 132, + 271, + 481, + 304 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 271, + 481, + 304 + ], + "spans": [ + { + "bbox": [ + 132, + 271, + 481, + 304 + ], + "type": "text", + "content": "55. Xu, X., Wang, S., Chen, Y., Zheng, Y., Wei, Z., Liu, J.: Gtp-vit: Efficient vision transformers via graph-based token propagation. In: Proceedings of the IEEE/CVF Winter Conference on Applications of Computer Vision. pp. 86-95 (2024)" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 132, + 304, + 481, + 335 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 304, + 481, + 335 + ], + "spans": [ + { + "bbox": [ + 132, + 304, + 481, + 335 + ], + "type": "text", + "content": "56. Yin, H., Vahdat, A., Alvarez, J.M., Mallya, A., Kautz, J., Molchanov, P.: A-vit: Adaptive tokens for efficient vision transformer. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 10809-10818 (2022)" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 132, + 336, + 481, + 380 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 336, + 481, + 380 + ], + "spans": [ + { + "bbox": [ + 132, + 336, + 481, + 380 + ], + "type": "text", + "content": "57. You, H., Xiong, Y., Dai, X., Wu, B., Zhang, P., Fan, H., Vajda, P., Lin, Y.C.: Castling-vit: Compressing self-attention via switching towards linear-angular attention at vision transformer inference. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 14431-14442 (2023)" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 132, + 380, + 481, + 402 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 380, + 481, + 402 + ], + "spans": [ + { + "bbox": [ + 132, + 380, + 481, + 402 + ], + "type": "text", + "content": "58. Yu, H., Wu, J.: A unified pruning framework for vision transformers. arXiv:2111.15127 [cs.CV] (2021)" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 132, + 403, + 481, + 435 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 403, + 481, + 435 + ], + "spans": [ + { + "bbox": [ + 132, + 403, + 481, + 435 + ], + "type": "text", + "content": "59. Zhang, X., Zhou, X., Lin, M., Sun, J.: Shufflenet: An extremely efficient convolutional neural network for mobile devices. In: Proceedings of the IEEE conference on computer vision and pattern recognition. pp. 6848-6856 (2018)" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 132, + 435, + 481, + 468 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 435, + 481, + 468 + ], + "spans": [ + { + "bbox": [ + 132, + 435, + 481, + 468 + ], + "type": "text", + "content": "60. Zhou, A., Li, Y., Qin, Z., Liu, J., Pan, J., Zhang, R., Zhao, R., Gao, P., Li, H.: Sparsemae: Sparse training meets masked autoencoders. In: Proceedings of the IEEE/CVF International Conference on Computer Vision. pp. 16176-16186 (2023)" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 132, + 468, + 481, + 501 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 468, + 481, + 501 + ], + "spans": [ + { + "bbox": [ + 132, + 468, + 481, + 501 + ], + "type": "text", + "content": "61. Zhou, B., Khosla, A., Lapedriza, A., Oliva, A., Torralba, A.: Learning deep features for discriminative localization. In: Proceedings of the IEEE conference on computer vision and pattern recognition. pp. 2921-2929 (2016)" + } + ] + } + ], + "index": 13 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "spans": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "type": "text", + "content": "18" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 166, + 91, + 239, + 101 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 166, + 91, + 239, + 101 + ], + "spans": [ + { + "bbox": [ + 166, + 91, + 239, + 101 + ], + "type": "text", + "content": "T. Mahmud et al." + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 17 + } + ], + "_backend": "vlm", + "_version_name": "2.6.4" +} \ No newline at end of file diff --git a/2024/PairingNet_ A Learning-based Pair-searching and -matching Network for Image Fragments/3b568531-7960-492d-9c00-7cccb8dd34e1_content_list.json b/2024/PairingNet_ A Learning-based Pair-searching and -matching Network for Image Fragments/3b568531-7960-492d-9c00-7cccb8dd34e1_content_list.json new file mode 100644 index 0000000000000000000000000000000000000000..8fcb5c423f2c4868a04dce11d9bcfcb22092c756 --- /dev/null +++ b/2024/PairingNet_ A Learning-based Pair-searching and -matching Network for Image Fragments/3b568531-7960-492d-9c00-7cccb8dd34e1_content_list.json @@ -0,0 +1,1784 @@ +[ + { + "type": "text", + "text": "PairingNet: A Learning-based Pair-searching and -matching Network for Image Fragments", + "text_level": 1, + "bbox": [ + 220, + 140, + 782, + 186 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Rixin Zhou $^{1}$ , Ding Xia $^{2}$ , Yi Zhang $^{3}$ , Honglin Pang $^{1}$ , Xi Yang $^{1,*,\\dagger, \\ddagger}$ , and Chuntao Li $^{4,*, \\ddagger}$", + "bbox": [ + 228, + 210, + 772, + 244 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "$^{1}$ School of Artificial Intelligence, Jilin University", + "bbox": [ + 333, + 253, + 666, + 268 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "2 The University of Tokyo", + "bbox": [ + 411, + 268, + 589, + 282 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "$^{3}$ National University of Defense Technology, College of Electronic Engineering", + "bbox": [ + 235, + 282, + 766, + 297 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "$^{4}$ School of Archaeology, Jilin University", + "bbox": [ + 364, + 296, + 638, + 310 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "{zhourx22}@mails.jlu.edu.cn {1ct33}@jlu.edu.cn", + "bbox": [ + 325, + 311, + 676, + 324 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "{dingxia1995, panghlwork, earthyangxi}@gmail.com {2356711993}@qq.com", + "bbox": [ + 246, + 325, + 754, + 339 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Abstract. In this paper, we propose a learning-based image fragment pair-searching and -matching approach to solve the challenging restoration problem. Existing works use rule-based methods to match similar contour shapes or textures, which are always difficult to tune hyperparameters for extensive data and computationally time-consuming. Therefore, we propose a neural network that can effectively utilize neighbor textures with contour shape information to fundamentally improve performance. First, we employ a graph-based network to extract the local contour and texture features of fragments. Then, for the pair-searching task, we adopt a linear transformer-based module to integrate these local features and use contrastive loss to encode the global features of each fragment. For the pair-matching task, we design a weighted fusion module to dynamically fuse extracted local contour and texture features, and formulate a similarity matrix for each pair of fragments to calculate the matching score and infer the adjacent segment of contours. To faithfully evaluate our proposed network, we collect a real dataset and generate a simulated image fragment dataset through an algorithm we designed that tears complete images into irregular fragments. The experimental results show that our proposed network achieves excellent pair-searching accuracy, reduces matching errors, and significantly reduces computational time. Source codes and data are available at here.", + "bbox": [ + 259, + 376, + 743, + 667 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Keywords: image fragment dataset $\\cdot$ image fragment pair-searching and -matching", + "bbox": [ + 259, + 680, + 743, + 709 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "1 Introduction", + "text_level": 1, + "bbox": [ + 215, + 739, + 375, + 755 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Pair-searching and -matching of massive unorganized fragments is a critical problem in the research fields of computer vision and computer graphics, especially in the field of cultural relic restoration [1,15,17,23]. Traditional methods recover geometry transformation between fragments by calculating the similarity of contour segments based on geometric invariants such as curvature and manually", + "bbox": [ + 212, + 763, + 787, + 840 + ], + "page_idx": 0 + }, + { + "type": "page_footnote", + "text": "*Corresponding authors", + "bbox": [ + 215, + 866, + 338, + 875 + ], + "page_idx": 0 + }, + { + "type": "page_footnote", + "text": "$\\dagger$ Engineering Research Center of Knowledge-Driven Human-Machine Intelligence, MoE, China", + "bbox": [ + 215, + 875, + 683, + 883 + ], + "page_idx": 0 + }, + { + "type": "page_footnote", + "text": "$\\ddagger$ Key Laboratory of Ancient Chinese Script, Culture Relics and Artificial Intelligence, Jilin Province", + "bbox": [ + 215, + 883, + 715, + 891 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/df094cbf5cd95bc8f5564e41d54071b551ce95439722ac23f906d28fce579c9f.jpg", + "image_caption": [ + "Fig. 1: Our pair-searching and -matching task. Given a large number of mixed image fragments, our proposed network can search and match pairs of fragments, which is important for restoration problems. We also designed an algorithm to generate a dataset of image fragments by tearing a set of complete images." + ], + "image_footnote": [], + "bbox": [ + 220, + 143, + 781, + 223 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "designed descriptors [18, 33]. It has a common core with the registration problem [3], which is to establish the correspondences of local features between related image pairs of the same scene. However, fragment assembly is more challenging because they do have overlapping regions. Thus, existing rule-based methods are difficult to generalize to various complex situations. Furthermore, they are time-consuming since the computational complexity of finding matching pairs from a set of fragments is typically $O(N^2)$ in global search, and in practice, only a few pairs are hidden in a large number of mixed fragments.", + "bbox": [ + 212, + 324, + 782, + 445 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Solving jigsaw puzzles [14, 37, 42] is also a similar task and learning-based approaches have been applied [2, 36]. However, two differences make these algorithms not generalizable. First, puzzle pieces usually have regular shapes, so they are best matched based on content information rather than contour features. Second, this problem usually studies matching all fragments into a complete block. Their input fragments do not come from mixed groups, and the matching results are dense and not lost. Therefore, they often use neighbor relationships to build graphs to constrain the search for the correct alignment of fragments. Once a fragment is lost, their closed-loop strategy fails.", + "bbox": [ + 212, + 446, + 787, + 583 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "In this paper, we propose PairingNet, an end-to-end deep learning network, for searching and matching pairs from a large number of image fragments. To address this problem, we first explore the use of Graph Convolutional Networks (GCNs) to extract more robust contour and texture features for our tasks. Second, we design a feature fusion way to learn local matching features by calculating the similarity matrix of each fragment pair. Third, we build a module based on the linear transformer to learn the global searching features for fragments using contrastive learning. Finally, we collect a real dataset and generate a image fragment dataset to provide data-driven and validate the effectiveness of our proposed network. The main contributions of our work are as follows:", + "bbox": [ + 212, + 585, + 787, + 736 + ], + "page_idx": 1 + }, + { + "type": "list", + "sub_type": "text", + "list_items": [ + "- We propose a novel network to leverage the local contour and texture features for solving the pair-searching and -matching of image fragments. It provides an attempt to use advanced deep-learning techniques to solve this long-standing problem.", + "- We provide a real dataset and a generated dataset with a carefully designed algorithm to simulate real-world fragments. The algorithm generates diverse" + ], + "bbox": [ + 223, + 750, + 782, + 839 + ], + "page_idx": 1 + }, + { + "type": "page_number", + "text": "2", + "bbox": [ + 217, + 114, + 228, + 126 + ], + "page_idx": 1 + }, + { + "type": "header", + "text": "Zhou et al.", + "bbox": [ + 271, + 114, + 346, + 126 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "image fragments to facilitate the application of subsequent learning-based methods.", + "bbox": [ + 240, + 146, + 784, + 175 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "- We conduct extensive experiments, ablation studies, and visual analysis to demonstrate the effectiveness of our proposed network.", + "bbox": [ + 225, + 176, + 784, + 205 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "2 Related Work", + "text_level": 1, + "bbox": [ + 215, + 229, + 387, + 247 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "2.1 Fragments Assembly (Irregular Shape)", + "text_level": 1, + "bbox": [ + 215, + 263, + 581, + 282 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Here, we focus on the pairwise matching algorithms of related representative works, although few of them also consider fragment relationships. Methods based on deep neural networks have been studied for object fragments; however, we could not find representative works for image fragments or even suitable datasets.", + "bbox": [ + 212, + 297, + 784, + 359 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Image fragments. Kong and Kimia [25] solved fragment matching using geometric features (partial curve matching). Leitao and Stolfi [17] described a specific multiscale matching method for the reassembly of 2D fragmented objects. Tsamoura and Pitas [43] presented a color-based method to automatically reassemble image fragments. Huang et al. [22] relied on salient curves detected inside the different image pieces to align gapped fragments. Subsequently, both shape and appearance information along the boundaries were utilized and extracted for each piece for matching [32,49]. Furthermore, Derech et al. [12] proposed to extrapolate each fragment to address fragment abrasion. For learning-based methods, traditional machine learning (SVM/Random Forest) was used as a classifier to process extracted local features [1,38]. Then Le and Li [27] built a deep neural network to detect the compatibility of pairwise stitching.", + "bbox": [ + 212, + 372, + 787, + 554 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Object fragments. Huang et al. [23] analyzed the geometry of the fracture surfaces by computing multi-scale surface characteristics. Yang et al. [46] matched stone tools based on contour points and their mean normals. Funkhouser et al. [15] learned a decision tree classifier to filter the candidates. Hong et al. [21] solved the problem of indistinguishable false matches by employing beam search to explore multiple registration possibilities. Ye et al. [47] presented an immersive system that can handle complex and ambiguous fragment shapes interactively with experts. In recent years, deep learning networks have been rapidly applied in 3D assembly. Huang et al. [48] proposed an assembly-oriented dynamic graph learning framework to estimate the pose of 3D parts. Chen et al. [8] proposed Neural Shape Mating (NSM) to tackle pairwise 3D geometric shape mating. Wu et al. [45] leveraged SE(3) equivariance for such shape pose disentanglement.", + "bbox": [ + 212, + 570, + 787, + 752 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Datasets. Existing image assembly works typically collect data in two ways: manual digitalization by photographing or scanning broken fragments, or using simple algorithms to tear complete images [7, 35, 38]. The lack of a large-scale dataset prevents the emergence of deep learning networks for this task. However, multiple datasets have been introduced in 3D assembly problems [24, 26, 40].", + "bbox": [ + 212, + 768, + 787, + 844 + ], + "page_idx": 2 + }, + { + "type": "header", + "text": "PairingNet", + "bbox": [ + 656, + 114, + 730, + 128 + ], + "page_idx": 2 + }, + { + "type": "page_number", + "text": "3", + "bbox": [ + 774, + 116, + 785, + 126 + ], + "page_idx": 2 + }, + { + "type": "image", + "img_path": "images/cadd093b8b7545134418df87cf2d950341591b2930aa4852b5775a42b4bf0fd9.jpg", + "image_caption": [ + "Fig. 2: Pipeline of our proposed PairingNet. (a) Based on patches, we first employ a binary encoding to describe contours and use ResGCN as the backbone to extract the local contour and texture features of image fragments. (b) We design self-gated fusion to fuse the extracted features in a weighted manner to calculate the similarity matrix for each pair of fragments, and (c) a linear transformer-based encoder for learning the global features of each fragment. (d) We use a two-step strategy to train our proposed network. (e) During inference, we calculate cosine similarity to find the adjacent fragment pairs and further process the matching similarity matrix to establish better correspondences between a fragment pair." + ], + "image_footnote": [], + "bbox": [ + 217, + 143, + 787, + 348 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "2.2 Jigsaw Puzzles (Regular Shape)", + "text_level": 1, + "bbox": [ + 215, + 513, + 524, + 530 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Solving jigsaw puzzles was first proposed in [14], and then regular pieces became a key research object in computer vision. Many existing methods [10, 16, 19, 37, 41, 42] focused on building relationships of pieces based on content information rather than contour-based pair-searching/-matching. And deep learning techniques [2, 36] was also used to solve this problem.", + "bbox": [ + 212, + 540, + 787, + 617 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "3 Method", + "text_level": 1, + "bbox": [ + 215, + 630, + 330, + 645 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "3.1 Problem Statement", + "text_level": 1, + "bbox": [ + 215, + 659, + 423, + 674 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Let us consider a fragment dataset $\\mathbf{F} = \\{f^{(i)}\\}_{i=1}^{N}$ , which consists of fragments $f$ torn from a certain image dataset. We define the contour points $c_{j}^{(i)} \\in \\mathbf{C}^{(i)}$ for each $f^{(i)}$ , where $j = 1,2,\\dots,M^{(i)}$ , $M^{(i)}$ is the contour length and $\\mathbf{C}^{(i)}$ is an ordered set that contains all contour points of $f^{(i)}$ .", + "bbox": [ + 212, + 681, + 787, + 748 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Regarding any two neighbor fragments sampled from a pairset $\\mathbf{P}$ , we denote the pair as $f^{(m)}$ and $f^{(n)}$ . The pair-searching task aims to retrieve the pairset $\\mathbf{P}$ from a given fragment dataset $\\mathbf{F}$ . Meanwhile, for the contours of the paired fragments, $\\mathbf{C}^{(m)}$ and $\\mathbf{C}^{(n)}$ , we can always find the overlapped part $\\mathbf{C}^{(m\\& n)}$ . The pair-matching task is designed to find the adjacent part from any given fragment pair.", + "bbox": [ + 212, + 750, + 787, + 840 + ], + "page_idx": 3 + }, + { + "type": "page_number", + "text": "4", + "bbox": [ + 217, + 114, + 228, + 126 + ], + "page_idx": 3 + }, + { + "type": "header", + "text": "Zhou et al.", + "bbox": [ + 271, + 114, + 346, + 126 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "3.2 Network Architecture", + "text_level": 1, + "bbox": [ + 215, + 146, + 442, + 159 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "In the pair-matching task, a model has the capability to discern differences between two fragments, regardless of whether these discrepancies are on a large or minimal scale. Simultaneously, the features extracted by the model can be used to quantify the likelihood of adjacency between any given two fragments, assisting in the identification of matched pairs. However, the process of comparing every two fragments is not only laborious, but also inefficient. Therefore, utilizing the rich features the matching model extracts, we conceived a pair-searching module. This module can adeptly convert any fragment into a unique feature vector, thereby significantly accelerating the pair-searching process.", + "bbox": [ + 212, + 172, + 787, + 309 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Feature Extraction", + "text_level": 1, + "bbox": [ + 215, + 323, + 375, + 335 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Contour encoding. In this module, we aim to direct the model's attention to the unique features of contours. We design distinct contour encodings that also may be applied in other sketch-based fields. We tested various patch sizes and all these distinct encodings, finalizing a binary encoding and a patch size of $7 \\times 7$ for our network. Subsequently, these contour patches are processed through a convolutional layer, followed by an average pooling layer and a fully connected network. We then adopt ResGCN [28] to amalgamate adjacent features, further extracting comprehensive global features. As such, for any provided fragment input denoted as $f^{(i)}$ with its corresponding contour $\\mathbf{C}^{(i)}$ , this module yields a return of $f_{c}^{(i)} \\in \\mathbb{R}^{M^{(i)} \\times 64}$ .", + "bbox": [ + 212, + 357, + 787, + 511 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Textural encoding. Our objective is to accentuate the distinguishing characteristics of textures in this module. We begin by cropping the image using a patch centered on each contour point. This is followed by filtering these patches through two consecutive convolutional layers, coupled with a single average pooling operation. Then, a ResGCN [28] is integrated to compile adjacent features, thereby enhancing the receptive domain of local textural characteristics. For any given fragment input represented as $f^{(i)}$ along with its associated contour $\\mathbf{C}^{(i)}$ , this module produces an output of $f_{t}^{(i)} \\in \\mathbb{R}^{M^{(i)} \\times 64}$ .", + "bbox": [ + 212, + 520, + 787, + 643 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Pair-matching Module", + "text_level": 1, + "bbox": [ + 215, + 652, + 405, + 667 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Feature Fusion. Regarding the decision-making process for pair-matching and -searching tasks, the dependence on contour and textural features varies. For instance, in scenarios where both fragments are solid colors, texture serves as the primary identifying factor suggesting adjacency. However, texture alone cannot identify the matched contours. In contrast, simple contours, like straight lines, require texture features for accurate matching position determination, as contour information alone falls short.", + "bbox": [ + 212, + 686, + 787, + 792 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Given this, we propose the incorporation of a Feature Fusion module [11] which would be capable of adaptively fine-tuning the weights assigned to these two features. The specific structure of this integral component is depicted in", + "bbox": [ + 212, + 794, + 787, + 839 + ], + "page_idx": 4 + }, + { + "type": "header", + "text": "PairingNet", + "bbox": [ + 656, + 114, + 730, + 128 + ], + "page_idx": 4 + }, + { + "type": "page_number", + "text": "5", + "bbox": [ + 774, + 116, + 784, + 126 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Figure 2 (b). First, we concatenate $f_{t}^{(i)}$ and $f_{c}^{(i)}$ , then the weight $\\mathbf{w}^{(i)}$ for each contour point $c_{j}^{(i)} \\in \\mathbf{C}^{(i)}$ is computed by:", + "bbox": [ + 212, + 143, + 782, + 181 + ], + "page_idx": 5 + }, + { + "type": "equation", + "text": "\n$$\n\\mathbf {w} ^ {(i)} = \\sigma \\left(\\mathbf {W} _ {g} \\left(f _ {t} ^ {(i)} \\odot f _ {c} ^ {(i)}\\right) + \\mathbf {b} _ {g}\\right) ^ {(i)} \\tag {1}\n$$\n", + "text_format": "latex", + "bbox": [ + 392, + 191, + 784, + 210 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "where $\\sigma$ is sigmoid the activation function, $(\\widehat{\\mathbf{c}})$ means channel-wise concatenation, and $\\mathbf{w} \\in \\mathbb{R}^{M^{(i)} \\times 64}$ . Second, we fuse features of two modalities:", + "bbox": [ + 214, + 220, + 782, + 252 + ], + "page_idx": 5 + }, + { + "type": "equation", + "text": "\n$$\n\\mathbf {F} _ {f} ^ {(i)} = \\mathbf {w} ^ {(i)} \\odot \\mathbf {F} _ {t} ^ {(i)} + (1 - \\mathbf {w} ^ {(i)}) \\odot \\mathbf {F} _ {c} ^ {(i)} \\tag {2}\n$$\n", + "text_format": "latex", + "bbox": [ + 377, + 262, + 784, + 281 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "where $\\mathbf{F}_f^{(i)}\\in \\mathbb{R}^{M^{(i)}\\times 64}$ , and $\\odot$ is element-wise multiplication. With this module, we let the model adaptively decide the weights between textual and contour features, which is proven effective in later experiments.", + "bbox": [ + 214, + 292, + 782, + 339 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Similarity Matrix. As shown in Figure 2 (b-2), for every pair of fragments $f^{(m)}$ and $f^{(n)}$ from the pairwise matrix $\\mathbf{S}^{(m\\& n)} \\in \\mathbb{R}^{M^{(m)} \\times M^{(n)}}$ [9,39].", + "bbox": [ + 214, + 348, + 782, + 395 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "We first calculate the similarity matrix $\\tilde{\\mathbf{S}}^{(m\\& n)}$ by matrix multiplication:", + "bbox": [ + 238, + 396, + 761, + 411 + ], + "page_idx": 5 + }, + { + "type": "equation", + "text": "\n$$\n\\tilde {\\mathbf {S}} ^ {(m \\& n)} = \\frac {1}{\\sqrt {D}} \\mathbf {F} _ {f} ^ {(m)} \\left(\\mathbf {F} _ {f} ^ {(n)}\\right) ^ {T} \\tag {3}\n$$\n", + "text_format": "latex", + "bbox": [ + 411, + 421, + 784, + 448 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "where $D$ is the dimension of the feature, in our model, $D$ is 64. Then, each entry in the matrix $\\mathbf{S}^{(m\\& n)}$ is normalized:", + "bbox": [ + 214, + 457, + 782, + 486 + ], + "page_idx": 5 + }, + { + "type": "equation", + "text": "\n$$\n\\mathbf {S} ^ {(m \\& n)} (i, j) = \\frac {\\exp \\left(\\tilde {\\mathbf {S}} _ {i , j} ^ {(m \\& n)}\\right)}{\\sum_ {i} \\exp \\left(\\tilde {\\mathbf {S}} _ {i , j} ^ {(m \\& n)}\\right)} \\frac {\\exp \\left(\\tilde {\\mathbf {S}} _ {i , j} ^ {(m \\& n)}\\right)}{\\sum_ {j} \\exp \\left(\\tilde {\\mathbf {S}} _ {i , j} ^ {(m \\& n)}\\right)} \\tag {4}\n$$\n", + "text_format": "latex", + "bbox": [ + 341, + 494, + 784, + 534 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Pair-searching Module We introduce a modified feature fusion module to process features extracted by the backbone network and encode any fragment $f^{(i)}$ into a feature vector $v_{s}^{(i)} \\in \\mathbb{R}^{128}$ . It is based on the discovery that using fused features in the pair-matching task will lower the performance.", + "bbox": [ + 214, + 558, + 782, + 619 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "To be specific, as shown in Figure 2 (c), every feature branch utilizes a Linear Transformer Encoder [44] to enhance the encoding of both contour and textural features. Subsequent to the modified fusion module, we add a fully-connected layer to encode the fused feature regarding per contour point. Then an average pooling layer is applied to map the feature matrix to a singular vector. Finally, another fully-connected layer is used to encode the feature.", + "bbox": [ + 214, + 621, + 782, + 710 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "For every pair of fragment $f^{(i)}$ and $f^{(j)}$ from the fragment set $\\mathbf{F}$ , this module allows them to be encoded as $v_{s}^{(i)}$ and $v_{s}^{(j)}$ . By calculating their inner product, we can efficiently quantify and contrast their similarity, thereby expediently identifying neighbor pairs.", + "bbox": [ + 214, + 710, + 782, + 773 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "3.3 Loss Function", + "text_level": 1, + "bbox": [ + 214, + 790, + 375, + 804 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "We use two loss functions to train our network, and the training process is shown in Figure 2 (d).", + "bbox": [ + 214, + 809, + 782, + 839 + ], + "page_idx": 5 + }, + { + "type": "page_number", + "text": "6", + "bbox": [ + 217, + 114, + 228, + 126 + ], + "page_idx": 5 + }, + { + "type": "header", + "text": "Zhou et al.", + "bbox": [ + 271, + 114, + 346, + 127 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Pair-matching Loss Given the extreme imbalance in the ratio of positive to negative samples in the pair-matching task, we adopt the concept of focal loss [30] to concentrate on learning challenging correspondences.", + "bbox": [ + 212, + 146, + 782, + 191 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "For any paired fragments $f^{(m)}$ and $f^{(n)}$ , entries in the ground truth similarity matrix $\\mathbf{S}_{gt}^{(m\\& n)}$ are 1 for matched contour points and 0 for unmatched ones. We calculate the matching loss $\\mathcal{L}_{match}^{(m\\& n)}$ as follows:", + "bbox": [ + 212, + 191, + 784, + 244 + ], + "page_idx": 6 + }, + { + "type": "equation", + "text": "\n$$\n\\begin{array}{l} \\mathcal {L} _ {\\text {m a t c h}} ^ {(m \\& n)} = - \\sum_ {(0, 0)} (\\beta_ {1} (\\mathbf {1} - \\mathbf {S} ^ {(m \\& n)}) ^ {\\gamma} \\log \\mathbf {S} ^ {(m \\& n)} \\mathbf {S} _ {g t} ^ {(m \\& n)}) \\tag {5} \\\\ + \\beta_ {2} (\\mathbf {S} ^ {(m \\& n)}) ^ {\\gamma} \\log \\left(\\mathbf {1} - \\mathbf {S} ^ {(m \\& n)}\\right) (\\mathbf {1} - \\mathbf {S} _ {g t} ^ {(m \\& n)}) \\\\ \\end{array}\n$$\n", + "text_format": "latex", + "bbox": [ + 326, + 248, + 784, + 290 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "where $\\beta_{1}$ and $\\beta_{2}$ are the balancing weights and $\\beta_{1} + \\beta_{2} = 1$ , $\\gamma$ is the decay factor. $\\sum$ adds all the entries in the input matrix and all operations except $\\sum$ are element-wise. By calculating and adding the pair-matching loss for all paired fragments $f^{(m)}, f^{(n)}$ from the pairwise $\\mathbf{P}$ , we get $\\mathcal{L}_{match}$ .", + "bbox": [ + 212, + 300, + 782, + 361 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Pair-searching Loss In the pair-searching task, we employ InfoNCE loss [34] as the contrastive loss function $\\mathcal{L}_{search}$ . We utilize paired samples for constructing positive sample pairs, and unpaired samples for negative sample pairs.", + "bbox": [ + 212, + 373, + 784, + 420 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "3.4 Inference", + "text_level": 1, + "bbox": [ + 214, + 436, + 339, + 450 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Pair-matching In this section, we target to retrieve matched contour from the similarity matrix $\\mathbf{S}^{(m\\& n)}$ .", + "bbox": [ + 212, + 455, + 782, + 484 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "We first apply a threshold $\\epsilon = 0.006$ to filter out noisy entries. Entries with value lower than the threshold is set to 0. Because contour pointset $\\mathbf{C}^{(m)}$ and $\\mathbf{C}^{(n)}$ are ordered sets, for any ground truth similarity matrix $\\mathbf{S}_{gt}^{(m\\& n)}$ , if the start for contour points is the left-top corner, we can observe that the part with a value of 1 in the matrix must be arranged in the direction from upper left to lower right and connected into a line segment. Therefore, this inspires us to enhance the filtered similarity matrix. We design an eroding kernel $\\mathbf{K}_e$ to filter out not dominant segments, which is adopted once. Then a dilating kernel $\\mathbf{K}_d$ is applied once to connect close but split segments. The definitions of two kernels are as follows:", + "bbox": [ + 212, + 486, + 784, + 637 + ], + "page_idx": 6 + }, + { + "type": "equation", + "text": "\n$$\n\\mathbf {K} _ {e} = \\left[ \\begin{array}{l l l} 0 & 0 & 1 \\\\ 0 & 0 & 0 \\\\ 1 & 0 & 0 \\end{array} \\right], \\quad \\mathbf {K} _ {d} = \\left[ \\begin{array}{l l l} 0 & 0 & 1 \\\\ 0 & 1 & 0 \\\\ 1 & 0 & 0 \\end{array} \\right]. \\tag {6}\n$$\n", + "text_format": "latex", + "bbox": [ + 393, + 651, + 784, + 694 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Finally, we use RANSAC [13] to estimate the correspondences of two fragments and obtain the transformation matrix.", + "bbox": [ + 212, + 698, + 782, + 729 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Pair-searching For N fragments $f^{(i)} \\in \\mathbf{F}, i = 1,2,\\dots,N$ , using the searching module, we can get a feature matrix $\\mathbf{V}_s = [v_1,v_2,\\ldots ,v_N]^T \\in \\mathbb{R}^{N\\times 128}$ . Subsequently, we compute the cosine similarity between all global features to form the cosine similarity matrix. This allows us to obtain the similarity score for any pair of fragments.", + "bbox": [ + 212, + 741, + 782, + 818 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "4 Dataset", + "text_level": 1, + "bbox": [ + 214, + 828, + 328, + 844 + ], + "page_idx": 6 + }, + { + "type": "header", + "text": "PairingNet", + "bbox": [ + 656, + 114, + 730, + 128 + ], + "page_idx": 6 + }, + { + "type": "page_number", + "text": "7", + "bbox": [ + 774, + 116, + 785, + 126 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/3f9ae64a0768296ac6a18f424d31230e30347881c843d7b2754bfeef048f9d2a.jpg", + "image_caption": [ + "Fig. 3: Dataset creation. (a) The process of generating our dataset. (b) Illustration of Algorithm 1 for tearing a complete image." + ], + "image_footnote": [], + "bbox": [ + 220, + 148, + 781, + 251 + ], + "page_idx": 7 + }, + { + "type": "table", + "img_path": "images/5787ad808ac71d71d8895da02e2410f9f63198efe1baa45b104e71346fe5195c.jpg", + "table_caption": [ + "Table 1: Dataset division. We divide the entire test set into three difficulty levels according to the contour overlapping proportion of fragment pairs." + ], + "table_footnote": [], + "table_body": "
TotalTrainValidationTest
FullHighMediumLow
Fragments819640988193279107116711090
Pairs14951365413723706631103604
", + "bbox": [ + 333, + 347, + 671, + 402 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Generated Dataset. We design Algorithm 1 to tear a series of complete images and create a new image fragment dataset with data close to the real world. The process is also shown in Figure 3. To control the shape and area of the fragments, we first find the circumscribed circle of a fragment and take two points at an appropriate distance on the circle to determine the starting and ending points of the segmentation. Then, to simulate a more realistic fragment contour, we randomly divide the cutting line into several segments, each segment randomly choosing a straight line or an irregular curve. The irregular curve is synthesized by a series of Fourier orthogonal bases to make the shape of the fragments more diverse and similar to the edge of a real torn paper. We set a minimum threshold for the fragments, and the fragment is re-segmented if a very small fragment appears.", + "bbox": [ + 212, + 434, + 787, + 617 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "To create our dataset, we collect 390 images of ten themes from Pexels [4] and empirically set the generation parameters shown in Algorithm 1 to obtain fragments. We calculate the distributions of our generated data, as shown in Figure 4. The distribution of fragment areas exhibits a long-tail pattern, which is consistent with the distribution of fragment areas observed when tearing large amounts of paper in real-life scenarios. The contour overlapping proportion of fragment pairs is closer to a uniform distribution over most ranges, indicating that our algorithm generates various types of fragment pairs. Then, we divide the dataset into training set, validation set, and test set in 5:1:4. We further divide the full test set into three difficulty levels based on the overlapping proportion of contour segments between fragment pairs. Low difficulty means that the pair of fragments has larger overlapping contour segments, and vice versa. The details are shown in Table 1.", + "bbox": [ + 212, + 618, + 789, + 815 + ], + "page_idx": 7 + }, + { + "type": "page_number", + "text": "8", + "bbox": [ + 217, + 114, + 228, + 126 + ], + "page_idx": 7 + }, + { + "type": "header", + "text": "Zhou et al.", + "bbox": [ + 271, + 114, + 346, + 126 + ], + "page_idx": 7 + }, + { + "type": "code", + "sub_type": "algorithm", + "code_caption": [ + "Algorithm 1 Generate fragments of an image" + ], + "code_body": "Input: List of fragments $L_{frag} = \\{\\}$ , a complete image $I$ Max iterations $t_{max} = 40$ min arc length ratio $\\tau = 0.9$ max number of cutting points $N_{max} =$ 3, min distance from cutting point to edge of fragment $D_{min} = 100$ , the number of Fourier orthogonal basis $n = 20$ scaling ratios $s_1 = 0.25$ $s_2 = 0.0067$ $s_3 = 1.5$ $s_4 = 0.3$ , probability of curve type $\\rho = 0.5$ , min fragment size $h_{min} = 150$ $w_{min} = 150$ \nOutput: Collection of fragments $L_{frag}$ \n1: put I in $L_{frag}$ \n2: while t < tmax do \n3: f<-randomly_pick(Lfrag), R<-bounding_rectangle(f) \n4: W,H<-width_and_height(R), C<-circumcircle(R) \n5: ldia<-diameter(C) \n6: \n7: // generate the cutting points \n8: repeat \n9: pa, pb<-randomly_pick(C) \n10: larc<-smaller_ar_c_length(pa,pb) \n11: until larc>τ·half_perimeter(C) \n12: pstart, pend<-intersection(f,pa,pb) \n13: m<-randomly_pick{0,...,Nmax} \n14: repeat \n15: [p1, ..., pm]<-random_select_insidepoints(f,m) \n16: d<-min_distance_to_edge(f,p1, ..., pm) \n17: until d > Dmin \n18: put pstart, pend in [p1, ..., pm] \n19: \n20: // generate the cutting segments \n21: for each pi∈[pstart, p1, ..., pm, pend] do \n22: irregular line $l_{irr} = \\{\\}$ , phase $\\phi \\sim \\mathcal{U}(-\\pi ,\\pi)$ amplitude A $\\sim \\mathcal{N}(s_1H,s_2H)$ period T $\\sim \\mathcal{N}(s_3W,s_4W)$ \n23: for each x-coordinate between (pi,pi+1) do \n24: y<∑i=0nA sin(2πix+φ)+H/2 \n25: put {x,y} in lirr \n26: end for \n27: lstr← straight_line(pi,pi+1) \n28: lcut← select_and_connect(lirr,lstr,ρ,lcut) \n29: end for \n30: f1, f2<-cut_structure(f,lcut) \n31: if size(f1)>hmin,wmin and size(f2)>hmin,wmin then \n32: put f1, f2 in Lfrag \n33: end if \n34: end while", + "bbox": [ + 215, + 159, + 785, + 577 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "Real Dataset. We also collect a series of real image fragments to construct a real dataset. We printed 34 complete images (not included in our generated dataset), manually cut them into fragments, and then scanned them. Then, we apply some post-processing steps to the scanned image fragments to obtain the contour of the fragments and the annotation of the matching points. More details about real dataset can be found in the supplementary materials.", + "bbox": [ + 215, + 604, + 785, + 696 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "5 Experiments", + "text_level": 1, + "bbox": [ + 215, + 709, + 375, + 726 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "5.1 Comparison Methods", + "text_level": 1, + "bbox": [ + 215, + 729, + 439, + 744 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "Finding comparison methods is not easy. Because rule-based methods have been around for quite some time, lack source code, and tend to exhibit lower performance compared to advanced deep learning methods. However, most existing deep learning methods require extensive modifications if applied to our task. Therefore, we employ Jigsawnet [27] (without loop consistency) and a rule-based method [49] (Step 1) as our comparison methods.", + "bbox": [ + 212, + 748, + 785, + 840 + ], + "page_idx": 8 + }, + { + "type": "header", + "text": "PairingNet", + "bbox": [ + 656, + 114, + 730, + 128 + ], + "page_idx": 8 + }, + { + "type": "page_number", + "text": "9", + "bbox": [ + 774, + 116, + 785, + 126 + ], + "page_idx": 8 + }, + { + "type": "image", + "img_path": "images/32bc207a37df5bb428d522db3af308c10cc8263dc6e5e54facc03970e7d2f24f.jpg", + "image_caption": [ + "Fig. 4: Dataset statistics. (a) Image area distribution of the generated fragments. (b) Proportional distribution of overlapping contour segments of the generated fragments. (c) Image area distribution of the scanned real fragments. (b) Proportional distribution of overlapping contour segments of the scanned real fragments." + ], + "image_footnote": [], + "bbox": [ + 215, + 143, + 504, + 215 + ], + "page_idx": 9 + }, + { + "type": "image", + "img_path": "images/34f244442e704b1028672b5542af0b539a890a23fef6ee6f71c96849fbee82c2.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 517, + 143, + 784, + 215 + ], + "page_idx": 9 + }, + { + "type": "image", + "img_path": "images/538965f920f9182e8f4cb9f7d734e2ab0305744a7a550aab2da822fe1cc18cf9.jpg", + "image_caption": [ + "Fig. 5: Violin plots of the matching results of our method and the comparison method: (a) RE, NTE, and HD on the test set. (b) HD on the test sets of different difficulties." + ], + "image_footnote": [], + "bbox": [ + 222, + 295, + 488, + 368 + ], + "page_idx": 9 + }, + { + "type": "image", + "img_path": "images/a821ca8844adee84036eda4afb9a636a3a266e58cc67cc9591536e033fff6f64.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 496, + 297, + 781, + 368 + ], + "page_idx": 9 + }, + { + "type": "text", + "text": "5.2 Implementation Details", + "text_level": 1, + "bbox": [ + 214, + 436, + 455, + 452 + ], + "page_idx": 9 + }, + { + "type": "text", + "text": "Our PairingNet was implemented by PyTorch, and conducted on a server with 4 RTX A40 GPUs and Intel® Xeon® Gold 5220 CPUs (72 cores). We used the Adam optimizer with an initial learning rate of 0.001, which was adjusted using a cosine annealing strategy. For training pair-matching module, the batch size was set to 20, and for the pair-searching module, the batch size was set to 175.", + "bbox": [ + 212, + 463, + 787, + 539 + ], + "page_idx": 9 + }, + { + "type": "text", + "text": "For comparison methods, given that the performance bottleneck of them resides in the CPU, utilizing GPUs for computational acceleration is deemed inappropriate. Thus we use two more powerful CPUs computing platforms to run the comparison method. We used the official sourcecode (implemented by TensorFlow) of Jigsawnet and conducted experiments on a server with 4 Quadro RTX 6000 GPUs and Intel® Xeon® Gold 5220 CPUs (144 cores). We reproduced the rule-based method by Python and conducted experiments on a node of a Supercomputing platform with Intel® Xeon® Gold 6258R CPUs (128 cores).", + "bbox": [ + 212, + 539, + 787, + 660 + ], + "page_idx": 9 + }, + { + "type": "text", + "text": "For evaluation metrics, we employ Recall@k and NDCG@k [5] for the pair-searching task, and Registration recall (RR), Hausdorff distance (HD), Radians Error (RE) and Normalized Translation Error (NTE) for the pair-matching task. We normalize Translation Error (TE) by the sum of the areas of fragment pairs to obtain the NTE, considering the large difference in area between fragments.", + "bbox": [ + 212, + 660, + 787, + 736 + ], + "page_idx": 9 + }, + { + "type": "text", + "text": "5.3 Results", + "text_level": 1, + "bbox": [ + 214, + 758, + 323, + 773 + ], + "page_idx": 9 + }, + { + "type": "text", + "text": "Performance on Generated Dataset. As shown in Table 2, our network outperforms the comparison methods on both the full test set and on every set of difficulty levels. Jigsawnet shows a certain superiority in pair-searching on the high-difficulty set, probably because it treats the problem as a simple binary", + "bbox": [ + 212, + 779, + 787, + 840 + ], + "page_idx": 9 + }, + { + "type": "page_number", + "text": "10", + "bbox": [ + 217, + 114, + 235, + 126 + ], + "page_idx": 9 + }, + { + "type": "header", + "text": "Zhou et al.", + "bbox": [ + 271, + 114, + 346, + 126 + ], + "page_idx": 9 + }, + { + "type": "table", + "img_path": "images/664d8e90f78f6d0a7151000f7f9ae616abb74fba9b0a0187423257a2c85ab855.jpg", + "table_caption": [ + "Table 2: Comparison Results of various metrics. The best results in our full dataset are highlighted in bold font, and the best results for each difficulty are marked in different colors." + ], + "table_footnote": [], + "table_body": "
MethodDifficultyRecall@5Recall@10Recall@20NDCG@5NDCG@10NDCG@20RR ↑HD ↓RE ↓NTE ↓
Rule-based [49]High0.2610.3540.4430.1850.2180.2410.107212.0031.28937.100×10-4
Medium0.3290.4270.5290.2420.2750.3020.153205.4631.24933.742×10-4
Low0.3580.4640.5550.2610.2960.320.188210.2741.23132.692×10-4
All0.3170.4160.5110.2480.2840.3110.149208.5191.25634.414×10-4
Jigsawnet [27]High0.3050.3940.4780.2310.2610.2840.117226.0261.30637.011×10-4
Medium0.3830.4780.5650.3090.340.3640.213206.0401.15531.467×10-4
Low0.4590.5480.6310.3680.3970.4190.241202.9421.18630.500×10-4
All0.380.4720.5570.3260.3590.3830.194210.8421.20532.772×10-4
OursHigh0.2760.4130.5350.1890.2340.2680.60692.2290.80718.335×10-4
Medium0.5010.6190.7320.3820.4230.4540.87935.6170.2365.685×10-4
Low0.7170.8050.8790.5810.610.6290.96112.3360.091.936×10-4
All0.4930.6080.7140.4170.4580.4870.83543.1160.3527.735×10-4
", + "bbox": [ + 218, + 198, + 784, + 325 + ], + "page_idx": 10 + }, + { + "type": "image", + "img_path": "images/7993877d11836b7c58a1b541d38db1fb231752dde2eb26188128b3850ca75c4a.jpg", + "image_caption": [ + "Fig. 6: Examples of pair-searching results. Our network is able to accurately identify the corresponding target fragment (highlighted with a green box)." + ], + "image_footnote": [], + "bbox": [ + 222, + 342, + 782, + 512 + ], + "page_idx": 10 + }, + { + "type": "text", + "text": "classification task; however, their pair-matching results are very unsatisfactory. It should be noted that we did not evaluate Recall@1 and NDCG@1 here due to the many-to-many pairings in our dataset, making it unreasonable to search only for the fragment with the highest similarity of searching. We can also observe that the performance of our proposed network shows reasonable variation with the difficulty of the data. This shows that our proposed network learns useful information and has more potential in complex situations. However, Jigsawnet and the rule-based method show relatively identical performance on test sets of different difficulty levels. Furthermore, we also plot the distributions of HD, RE, and NTE for each method for a more intuitive comparison, as shown in Figure 5.", + "bbox": [ + 212, + 585, + 787, + 738 + ], + "page_idx": 10 + }, + { + "type": "text", + "text": "Performance on Real Dataset. We also tested our method on the real dataset. In this experiment, both our method and the comparison method were trained on the generated dataset without retraining or fine-tuning. As shown in Table 3 and Figure 8, PairingNet obtained satisfactory results for both pair-searching and -matching tasks. These results confirmed that Algorithm 1 simulates real data well and our method has good generalization to real data.", + "bbox": [ + 212, + 750, + 787, + 840 + ], + "page_idx": 10 + }, + { + "type": "header", + "text": "PairingNet", + "bbox": [ + 656, + 114, + 730, + 128 + ], + "page_idx": 10 + }, + { + "type": "page_number", + "text": "11", + "bbox": [ + 767, + 114, + 784, + 126 + ], + "page_idx": 10 + }, + { + "type": "image", + "img_path": "images/8d02581339b5ca491fca5e0f094c297ff8eceedbbce1bdab5c1f9989cd7af6e8.jpg", + "image_caption": [ + "Fig. 7: Examples of pair-matching results for three difficulty levels. Our proposed network achieves satisfactory matching results. However, for low-difficulty cases, Jigsawnet and the Rule-based method find approximate matching positions, while the matchings are not accurate enough. For medium- and high-difficulty cases, both comparison methods are difficult to obtain correct matching results." + ], + "image_footnote": [], + "bbox": [ + 217, + 143, + 787, + 200 + ], + "page_idx": 11 + }, + { + "type": "table", + "img_path": "images/34d55e7f1f2fd12735dda1f12ffc76364b47d7a395dee0887e9e9f61d78040fc.jpg", + "table_caption": [ + "Table 3: Comparison Results of various metrics on our real dataset. The best results are highlighted in bold font." + ], + "table_footnote": [], + "table_body": "
MethodRecall@5Recall@10Recall@20NDCG@5NDCG@10NDCG@20RR↑HD↓RE↓NTE↓
Rule-based [49]0.1380.2470.3660.0790.1140.1440.1287.6981.61430.746×10-4
Jigsawnet [27]0.2370.3260.4250.1850.2130.2380.1537.5701.54529.061×10-4
Ours0.4150.5640.6780.3210.3690.3990.7524.1820.4809.075×10-4
", + "bbox": [ + 246, + 335, + 754, + 380 + ], + "page_idx": 11 + }, + { + "type": "text", + "text": "Visualization. We show the top five search results of different methods for the examples of fragment pairs in Figure 6. Given a source fragment, our model can accurately identify the corresponding target fragment (marked with a green box). Furthermore, in situations where a source fragment can be paired with multiple target fragments, our method identifies all matching fragments within the top five search results. We also show examples of matching results for three difficulty levels in Figure 7. Our method produces matching results that are consistent with the Ground Truth (GT), while the comparison methods yield numerous incorrect results.", + "bbox": [ + 212, + 405, + 787, + 540 + ], + "page_idx": 11 + }, + { + "type": "text", + "text": "We also visualize the normalized self-gated weights to analyze their changes under different situations. As shown in Figure 9 (a), we map the normalized weights to the edges of the fragments to show the influence of the texture/contour features. We classify the local parts of fragments into four types based on the information contained in texture and contour: those with rich texture and contour information (first quadrant), those with rich texture information (second quadrant), those with little texture and contour information (third quadrant), and those with rich contour information but with little texture information (fourth quadrant). The red color indicates that the weights of the texture features are relatively large, while the green color indicates that the contour feature weights are relatively large. The results show that our proposed network learned varying weights of the features for fragments with different characteristics.", + "bbox": [ + 212, + 541, + 787, + 722 + ], + "page_idx": 11 + }, + { + "type": "text", + "text": "Inference Time. We test the inference time of different methods on the pairwise searching task on our generated test set. Our network retrieves all fragment pairs in 73 seconds, while the rule-based method takes 223 hours, and JigsawNet takes 10 hours.", + "bbox": [ + 212, + 727, + 787, + 787 + ], + "page_idx": 11 + }, + { + "type": "text", + "text": "5.4 Ablation Study", + "text_level": 1, + "bbox": [ + 214, + 794, + 390, + 808 + ], + "page_idx": 11 + }, + { + "type": "text", + "text": "Network Structures. We explore the performance of our network under different structures, as shown in Table 4. First, we test three feature fusion ways.", + "bbox": [ + 212, + 809, + 787, + 840 + ], + "page_idx": 11 + }, + { + "type": "page_number", + "text": "12", + "bbox": [ + 217, + 114, + 235, + 126 + ], + "page_idx": 11 + }, + { + "type": "header", + "text": "Zhou et al.", + "bbox": [ + 271, + 114, + 346, + 126 + ], + "page_idx": 11 + }, + { + "type": "image", + "img_path": "images/0886c5dd15063f233ed70d809b3cd7bfa9cdb31d25e8cdc060b4ee3ba9acc216.jpg", + "image_caption": [ + "Fig. 8: Examples of pair-matching results of PairingNet on real dataset. Without retraining, our method can also effectively match the scanned real fragments." + ], + "image_footnote": [], + "bbox": [ + 222, + 143, + 782, + 305 + ], + "page_idx": 12 + }, + { + "type": "image", + "img_path": "images/e8bcd3996eef8dc74cd042054d0a5545b5cdfb3d6425de7b5b14d4af8a59fea9.jpg", + "image_caption": [ + "(a)" + ], + "image_footnote": [], + "bbox": [ + 349, + 358, + 522, + 460 + ], + "page_idx": 12 + }, + { + "type": "image", + "img_path": "images/795386562a2688255b455a2b03f0895e0874d88a3c5d56daabfab22a0ed06e38.jpg", + "image_caption": [ + "(b)", + "Fig.9: Visual analysis of Texture/Contour. (a) We visualize the weights of texture and contour features in our network, with colors closer to red representing more use of texture features and closer to green representing more use of contour features. We can see that on edges with rich texture or close to straight lines, the weights of the extracted texture features are greater, while on edges with single texture, the weights of the contour features are greater. (b) Examples of Textur/Contour ablation study. When the local contour shapes are similar, our network has difficulty matching them correctly using only contour features, and vice versa. However, these cases can be solved when the two features are fused." + ], + "image_footnote": [], + "bbox": [ + 522, + 358, + 653, + 464 + ], + "page_idx": 12 + }, + { + "type": "text", + "text": "In early fusion, instead of using two separate GCNs for feature extraction, we concatenate the two modalities and use only one GCN to learn the features. We also test using a simple concatenation to replace the self-gated fusion module. The results show that the performance of both pair-searching and -matching decreases. Additionally, we explore the impact of different numbers of layers on our network.", + "bbox": [ + 212, + 638, + 787, + 731 + ], + "page_idx": 12 + }, + { + "type": "text", + "text": "Texture/Contour. To study the impact of contour and texture features on the performance of our network, we conducted experiments using only one modality (either contour or texture) and compared the final results. As shown in Table 4, the comparison results indicate that both modalities perform slightly worse when used independently. When both modalities are used in conjunction, our network achieves the best performance. We also show examples of failed pair-matching when the two modalities are used individually in Figure 9 (b).", + "bbox": [ + 212, + 734, + 787, + 840 + ], + "page_idx": 12 + }, + { + "type": "header", + "text": "PairingNet", + "bbox": [ + 656, + 114, + 730, + 128 + ], + "page_idx": 12 + }, + { + "type": "page_number", + "text": "13", + "bbox": [ + 767, + 114, + 785, + 126 + ], + "page_idx": 12 + }, + { + "type": "table", + "img_path": "images/17bd2d05a4f6b932881cf253f5b29b135a1aa47ccf7b46b64e85ac0dbf3065c1.jpg", + "table_caption": [ + "Table 4: Results of ablation studies. Comparison of the impact of different feature fusion ways, feature modalities, and network layers employed in our proposed network on the performance of pair-searching and pair-matching tasks." + ], + "table_footnote": [], + "table_body": "
Recall@5Recall@10NDCG@5NDCG@10RR
Fusion WayEarly fusion0.4480.5590.3760.4130.658
Concat0.4890.6050.4180.4580.824
Weighted0.4930.6080.4170.4580.835
ResGCN10----0.680
14----0.835
18----0.759
Transformer Encoder30.4880.6000.4100.450-
40.4880.6060.4110.452-
50.4930.6080.4170.458-
Feature ModalityContour-only0.0160.0250.0130.0160.584
Texture-only0.4790.5970.3940.4360.657
Contour + Texture0.4930.6080.4170.4580.835
", + "bbox": [ + 354, + 198, + 647, + 305 + ], + "page_idx": 13 + }, + { + "type": "text", + "text": "Besides, as shown in Figure 9 (a), it can be seen numerically that our network tends to assign more weight to the texture features. This is reasonable because textures usually contain richer information than the contours of image fragments, enabling the network to learn more discriminative features.", + "bbox": [ + 215, + 333, + 784, + 393 + ], + "page_idx": 13 + }, + { + "type": "text", + "text": "Others. We report the results of ablation studies on patch size, contour encoding, and hyperparameters in our supplementary material.", + "bbox": [ + 215, + 398, + 784, + 428 + ], + "page_idx": 13 + }, + { + "type": "text", + "text": "6 Conclusions", + "text_level": 1, + "bbox": [ + 215, + 445, + 367, + 459 + ], + "page_idx": 13 + }, + { + "type": "text", + "text": "In this paper, to tackle the image fragment pair-searching and -matching tasks, we propose a novel network, PairingNet, to deduce the adjacent segments of fragments and encode the global features of each fragment. In PairingNet, ResGCN is employed as the backbone to extract the local contour and texture features of fragments. For pair-searching, a linear transformer-based module processes these local features leveraging the contrast of paired fragments and unpaired fragments. For pair-matching, we design a weighted fusion module to dynamically fuse the extracted local contour and texture features. Furthermore, to provide a foundation for learning-based methods, we design a cutting algorithm to generate a dataset by simulating real image fragments. Comprehensive experiments are conducted on our generated dataset and a real dataset, the results of which demonstrate the effectiveness of our PairingNet compared to existing methods. In future work, we will explore how to extend our work to other general technique, e.g. puzzles inspired MAE [20], JiGen [6] for self-supervised learning and domain generalization. The fragment generation algorithm that we designed can also create more diverse samples for image inpainting tasks [29,31].", + "bbox": [ + 215, + 462, + 785, + 703 + ], + "page_idx": 13 + }, + { + "type": "text", + "text": "Limitations While our research has made significant strides, fragment restoration remains a challenging task. Our dataset provides superior edge matching features. The task of fragment restoration becomes particularly arduous when the edges of the fragments are smooth and the texture is monotonous (for instance, in the case of shredded white paper), or when there is a substantial gap in the damaged edges.", + "bbox": [ + 215, + 704, + 785, + 794 + ], + "page_idx": 13 + }, + { + "type": "text", + "text": "Acknowledgements This work was supported by the Young Scientists Fund of the National Natural Science Foundation of China (Grant No.62206106). We also feel grateful to Yifan Jin for his exploration in the preliminary work.", + "bbox": [ + 215, + 795, + 785, + 839 + ], + "page_idx": 13 + }, + { + "type": "page_number", + "text": "14", + "bbox": [ + 217, + 114, + 235, + 126 + ], + "page_idx": 13 + }, + { + "type": "header", + "text": "Zhou et al.", + "bbox": [ + 271, + 114, + 344, + 126 + ], + "page_idx": 13 + }, + { + "type": "text", + "text": "References", + "text_level": 1, + "bbox": [ + 217, + 143, + 321, + 159 + ], + "page_idx": 14 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "1. Abitbol, R., Shimshoni, I., Ben-Dov, J.: Machine learning based assembly of fragments of ancient papyrus. Journal on Computing and Cultural Heritage (JOCCH) 14(3), 1-21 (2021)", + "2. Bridger, D., Danon, D., Tal, A.: Solving jigsaw puzzles with eroded boundaries. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 3526-3535 (2020)", + "3. Brown, L.G.: A survey of image registration techniques. ACM computing surveys (CSUR) 24(4), 325-376 (1992)", + "4. Bruno Joseph, I.J., Frese, D.: Pexels. https://www.pexels.com/ (2023)", + "5. Busa-Fekete, R., Szarvas, G., Elteto, T., Kégl, B.: An apple-to-apple comparison of learning-to-rank algorithms in terms of normalized discounted cumulative gain. In: ECAI 2012-20th European Conference on Artificial Intelligence: Preference Learning: Problems and Applications in AI Workshop. vol. 242. IOS Press (2012)", + "6. Carlucci, F.M., D'Innocente, A., Bucci, S., Caputo, B., Tommasi, T.: Domain generalization by solving jigsaw puzzles. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 2229-2238 (2019)", + "7. Chen, J., Tian, M., Qi, X., Wang, W., Liu, Y.: A solution to reconstruct cross-cut shredded text documents based on constrained seed k-means algorithm and ant colony algorithm. Expert Systems with Applications 127(AUG.), 35-46 (2019)", + "8. Chen, Y.C., Li, H., Turpin, D., Jacobson, A., Garg, A.: Neural shape mating: Self-supervised object assembly with adversarial shape priors. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 12724-12733 (2022)", + "9. Cheng, X., Lin, H., Wu, X., Yang, F., Shen, D.: Improving video-text retrieval by multi-stream corpus alignment and dual softmax loss. arXiv preprint arXiv:2109.04290 (2021)", + "0. Cho, T.S., Avidan, S., Freeman, W.T.: A probabilistic image jigsaw puzzle solver. In: 2010 IEEE Computer society conference on computer vision and pattern recognition. pp. 183-190. IEEE (2010)", + "1. Dai, Y., Gieseke, F., Oehmcke, S., Wu, Y., Barnard, K.: Attentional feature fusion. In: Proceedings of the IEEE/CVF winter conference on applications of computer vision. pp. 3560-3569 (2021)", + "2. Derech, N., Tal, A., Shimshoni, I.: Solving archaeological puzzles. Pattern Recognition 119, 108065 (2021)", + "3. Fischler, M.A., Bolles, R.C.: Random sample consensus: a paradigm for model fitting with applications to image analysis and automated cartography. Communications of the ACM 24(6), 381-395 (1981)", + "4. Freeman, H., Gardner, L.: Apictorial jigsaw puzzles: The computer solution of a problem in pattern recognition. IEEE Transactions on Electronic Computers (2), 118-127 (1964)", + "5. Funkhouser, T., Shin, H., Toler-Franklin, C., Castañeda, A.G., Brown, B., Dobkin, D., Rusinkiewicz, S., Weyrich, T.: Learning how to match fresco fragments. Journal on Computing and Cultural Heritage (JOCCH) 4(2), 1-13 (2011)", + "6. Gallagher, A.C.: Jigsaw puzzles with pieces of unknown orientation. In: 2012 IEEE Conference on computer vision and pattern recognition. pp. 382-389. IEEE (2012)", + "7. da Gama Leitao, H.C., Stolfi, J.: A multiscale method for the reassembly of two-dimensional fragmented objects. IEEE Transactions on Pattern Analysis and Machine Intelligence 24(9), 1239-1251 (2002)" + ], + "bbox": [ + 225, + 176, + 784, + 840 + ], + "page_idx": 14 + }, + { + "type": "header", + "text": "PairingNet", + "bbox": [ + 656, + 114, + 730, + 128 + ], + "page_idx": 14 + }, + { + "type": "page_number", + "text": "15", + "bbox": [ + 767, + 114, + 785, + 126 + ], + "page_idx": 14 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "18. Guo, Y., Bennamoun, M., Sohel, F., Lu, M., Wan, J., Kwok, N.M.: A comprehensive performance evaluation of 3d local feature descriptors. International Journal of Computer Vision 116, 66-89 (2016)", + "19. Gur, S., Ben-Shahar, O.: From square pieces to brick walls: The next challenge in solving jigsaw puzzles. In: Proceedings of the IEEE International Conference on Computer Vision. pp. 4029-4037 (2017)", + "20. He, K., Chen, X., Xie, S., Li, Y., Dollar, P., Girshick, R.: Masked autoencoders are scalable vision learners. In: Proceedings of the IEEE/CVF conference on computer vision and pattern recognition. pp. 16000-16009 (2022)", + "21. Hong, J.H., Yoo, S.J., Zeeshan, M.A., Kim, Y.M., Kim, J.: Structure-from-sherds: Incremental 3d reassembly of axially symmetric pots from unordered and mixed fragment collections. In: Proceedings of the IEEE/CVF International Conference on Computer Vision. pp. 5443-5451 (2021)", + "22. Huang, H., Yin, K., Gong, M., Lischinski, D., Cohen-Or, D., Ascher, U.M., Chen, B.: \"mind the gap\": tele-registration for structure-driven image completion. ACM Trans. Graph. 32(6), 174-1 (2013)", + "23. Huang, Q.X., Flóry, S., Gelfand, N., Hofer, M., Pottmann, H.: Reassembling fractured objects by geometric matching. In: ACM SIGGRAPH 2006 Papers, pp. 569-578 (2006)", + "24. Jones, B., Hildreth, D., Chen, D., Baran, I., Kim, V.G., Schulz, A.: Automate: A dataset and learning approach for automatic mating of cad assemblies. ACM Transactions on Graphics (TOG) 40(6), 1-18 (2021)", + "25. Kong, W., Kimia, B.B.: On solving 2d and 3d puzzles using curve matching. In: Proceedings of the 2001 IEEE Computer Society Conference on Computer Vision and Pattern Recognition. CVPR 2001. vol. 2, pp. II-II. IEEE (2001)", + "26. Lamb, N., Palmer, C., Molloy, B., Banerjee, S., Banerjee, N.K.: Fantastic breaks: A dataset of paired 3d scans of real-world broken objects and their complete counterparts. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 4681-4691 (2023)", + "27. Le, C., Li, X.: Jigsawnet: Shredded image reassembly using convolutional neural network and loop-based composition. IEEE Transactions on Image Processing 28(8), 4000-4015 (2019)", + "28. Li, G., Muller, M., Thabet, A., Ghanem, B.: Deep GCs: Can GCs go as deep as cnns? In: Proceedings of the IEEE/CVF international conference on computer vision. pp. 9267-9276 (2019)", + "29. Li, W., Lin, Z., Zhou, K., Qi, L., Wang, Y., Jia, J.: Mat: Mask-aware transformer for large hole image inpainting. In: Proceedings of the IEEE/CVF conference on computer vision and pattern recognition. pp. 10758-10768 (2022)", + "30. Li, Y., Harada, T.: Lepard: Learning partial point cloud matching in rigid and deformable scenes. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 5554-5564 (2022)", + "31. Liu, G., Reda, F.A., Shih, K.J., Wang, T.C., Tao, A., Catanzaro, B.: Image inpainting for irregular holes using partial convolutions. In: Proceedings of the European conference on computer vision (ECCV). pp. 85-100 (2018)", + "32. Liu, H., Cao, S., Yan, S.: Automated assembly of shredded pieces from multiple photos. IEEE transactions on multimedia 13(5), 1154-1162 (2011)", + "33. Mikolajczyk, K., Schmid, C.: A performance evaluation of local descriptors. IEEE transactions on pattern analysis and machine intelligence 27(10), 1615-1630 (2005)", + "34. Oord, A.v.d., Li, Y., Vinyals, O.: Representation learning with contrastive predictive coding. arXiv preprint arXiv:1807.03748 (2018)" + ], + "bbox": [ + 215, + 147, + 785, + 840 + ], + "page_idx": 15 + }, + { + "type": "page_number", + "text": "16", + "bbox": [ + 217, + 114, + 235, + 126 + ], + "page_idx": 15 + }, + { + "type": "header", + "text": "Zhou et al.", + "bbox": [ + 271, + 114, + 346, + 127 + ], + "page_idx": 15 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "35. Paixao, T.M., Boeres, M., Freitas, C., Oliveira-Santos, T.: Exploring character shapes for unsupervised reconstruction of strip-shredded text documents. IEEE Transactions on Information Forensics and Security pp. 1-1 (2018)", + "36. Paumard, M.M., Picard, D., Tabia, H.: Image reassembly combining deep learning and shortest path problem. In: Proceedings of the European conference on computer vision (ECCV). pp. 153-167 (2018)", + "37. Pomeranz, D., Shemesh, M., Ben-Shahar, O.: A fully automated greedy square jigsaw puzzle solver. In: CVPR 2011. pp. 9-16. IEEE (2011)", + "38. Richter, F., Ries, C.X., Cebron, N., Lienhart, R.: Learning to reassemble shredded documents. IEEE Transactions on Multimedia 15(3), 582-593 (2013)", + "39. Rocco, I., Cimpoi, M., Arandjelovic, R., Torii, A., Pajdla, T., Sivic, J.: Neighbourhood consensus networks. Advances in neural information processing systems 31 (2018)", + "40. Sellan, S., Chen, Y.C., Wu, Z., Garg, A., Jacobson, A.: Breaking bad: A dataset for geometric fracture and reassembly. Advances in Neural Information Processing Systems 35, 38885-38898 (2022)", + "41. Son, K., Hays, J., Cooper, D.B.: Solving square jigsaw puzzles with loop constraints. In: European conference on computer vision. pp. 32-46. Springer (2014)", + "42. Son, K., Hays, J., Cooper, D.B.: Solving square jigsaw puzzle by hierarchical loop constraints. IEEE transactions on pattern analysis and machine intelligence 41(9), 2222-2235 (2018)", + "43. Tsamoura, E., Pitas, I.: Automatic color based reassembly of fragmented images and paintings. IEEE Transactions on Image Processing 19(3), 680-690 (2009)", + "44. Wang, S., Li, B.Z., Khabsa, M., Fang, H., Ma, H.: Linformer: Self-attention with linear complexity. arXiv preprint arXiv:2006.04768 (2020)", + "45. Wu, R., Tie, C., Du, Y., Zhao, Y., Dong, H.: Leveraging se (3) equivariance for learning 3d geometric shape assembly. In: Proceedings of the IEEE/CVF International Conference on Computer Vision. pp. 14311-14320 (2023)", + "46. Yang, X., Matsuyama, K., Konno, K.: Pairwise matching of stone tools based on flake-surface contour points and normals. In: Proceedings of the Eurographics Workshop on Graphics and Cultural Heritage. pp. 125-129 (2017)", + "47. Ye, S., Chen, Z., Chu, X., Li, K., Luo, J., Li, Y., Geng, G., Wu, Y.: Puzzlefixer: A visual reassembly system for immersive fragments restoration. IEEE transactions on visualization and computer graphics 29(1), 429-439 (2022)", + "48. Zhan, G., Fan, Q., Mo, K., Shao, L., Chen, B., Guibas, L.J., Dong, H., et al.: Generative 3d part assembly via dynamic graph learning. Advances in Neural Information Processing Systems 33, 6315-6326 (2020)", + "49. Zhang, K., Li, X.: A graph-based optimization algorithm for fragmented image reassembly. Graphical Models 76(5), 484-495 (2014)" + ], + "bbox": [ + 212, + 146, + 787, + 686 + ], + "page_idx": 16 + }, + { + "type": "header", + "text": "PairingNet", + "bbox": [ + 656, + 114, + 730, + 128 + ], + "page_idx": 16 + }, + { + "type": "page_number", + "text": "17", + "bbox": [ + 767, + 114, + 785, + 126 + ], + "page_idx": 16 + } +] \ No newline at end of file diff --git a/2024/PairingNet_ A Learning-based Pair-searching and -matching Network for Image Fragments/3b568531-7960-492d-9c00-7cccb8dd34e1_model.json b/2024/PairingNet_ A Learning-based Pair-searching and -matching Network for Image Fragments/3b568531-7960-492d-9c00-7cccb8dd34e1_model.json new file mode 100644 index 0000000000000000000000000000000000000000..ce1d6d5cb2234a7cb7721c814ea46414aeb38e88 --- /dev/null +++ b/2024/PairingNet_ A Learning-based Pair-searching and -matching Network for Image Fragments/3b568531-7960-492d-9c00-7cccb8dd34e1_model.json @@ -0,0 +1,2401 @@ +[ + [ + { + "type": "title", + "bbox": [ + 0.221, + 0.141, + 0.783, + 0.187 + ], + "angle": 0, + "content": "PairingNet: A Learning-based Pair-searching and -matching Network for Image Fragments" + }, + { + "type": "text", + "bbox": [ + 0.23, + 0.211, + 0.773, + 0.245 + ], + "angle": 0, + "content": "Rixin Zhou\\(^{1}\\), Ding Xia\\(^{2}\\), Yi Zhang\\(^{3}\\), Honglin Pang\\(^{1}\\), Xi Yang\\(^{1,*,\\dagger, \\ddagger}\\), and Chuntao Li\\(^{4,*, \\ddagger}\\)" + }, + { + "type": "text", + "bbox": [ + 0.334, + 0.254, + 0.668, + 0.269 + ], + "angle": 0, + "content": "\\(^{1}\\) School of Artificial Intelligence, Jilin University" + }, + { + "type": "text", + "bbox": [ + 0.412, + 0.269, + 0.591, + 0.284 + ], + "angle": 0, + "content": "2 The University of Tokyo" + }, + { + "type": "text", + "bbox": [ + 0.236, + 0.283, + 0.767, + 0.298 + ], + "angle": 0, + "content": "\\(^{3}\\) National University of Defense Technology, College of Electronic Engineering" + }, + { + "type": "text", + "bbox": [ + 0.365, + 0.297, + 0.639, + 0.311 + ], + "angle": 0, + "content": "\\(^{4}\\) School of Archaeology, Jilin University" + }, + { + "type": "text", + "bbox": [ + 0.326, + 0.312, + 0.678, + 0.325 + ], + "angle": 0, + "content": "{zhourx22}@mails.jlu.edu.cn {1ct33}@jlu.edu.cn" + }, + { + "type": "text", + "bbox": [ + 0.248, + 0.326, + 0.756, + 0.34 + ], + "angle": 0, + "content": "{dingxia1995, panghlwork, earthyangxi}@gmail.com {2356711993}@qq.com" + }, + { + "type": "text", + "bbox": [ + 0.261, + 0.377, + 0.744, + 0.668 + ], + "angle": 0, + "content": "Abstract. In this paper, we propose a learning-based image fragment pair-searching and -matching approach to solve the challenging restoration problem. Existing works use rule-based methods to match similar contour shapes or textures, which are always difficult to tune hyperparameters for extensive data and computationally time-consuming. Therefore, we propose a neural network that can effectively utilize neighbor textures with contour shape information to fundamentally improve performance. First, we employ a graph-based network to extract the local contour and texture features of fragments. Then, for the pair-searching task, we adopt a linear transformer-based module to integrate these local features and use contrastive loss to encode the global features of each fragment. For the pair-matching task, we design a weighted fusion module to dynamically fuse extracted local contour and texture features, and formulate a similarity matrix for each pair of fragments to calculate the matching score and infer the adjacent segment of contours. To faithfully evaluate our proposed network, we collect a real dataset and generate a simulated image fragment dataset through an algorithm we designed that tears complete images into irregular fragments. The experimental results show that our proposed network achieves excellent pair-searching accuracy, reduces matching errors, and significantly reduces computational time. Source codes and data are available at here." + }, + { + "type": "text", + "bbox": [ + 0.261, + 0.681, + 0.744, + 0.71 + ], + "angle": 0, + "content": "Keywords: image fragment dataset \\(\\cdot\\) image fragment pair-searching and -matching" + }, + { + "type": "title", + "bbox": [ + 0.216, + 0.74, + 0.377, + 0.756 + ], + "angle": 0, + "content": "1 Introduction" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.765, + 0.788, + 0.841 + ], + "angle": 0, + "content": "Pair-searching and -matching of massive unorganized fragments is a critical problem in the research fields of computer vision and computer graphics, especially in the field of cultural relic restoration [1,15,17,23]. Traditional methods recover geometry transformation between fragments by calculating the similarity of contour segments based on geometric invariants such as curvature and manually" + }, + { + "type": "page_footnote", + "bbox": [ + 0.216, + 0.867, + 0.339, + 0.876 + ], + "angle": 0, + "content": "*Corresponding authors" + }, + { + "type": "page_footnote", + "bbox": [ + 0.216, + 0.876, + 0.684, + 0.884 + ], + "angle": 0, + "content": "\\(\\dagger\\) Engineering Research Center of Knowledge-Driven Human-Machine Intelligence, MoE, China" + }, + { + "type": "page_footnote", + "bbox": [ + 0.216, + 0.884, + 0.716, + 0.892 + ], + "angle": 0, + "content": "\\(\\ddagger\\) Key Laboratory of Ancient Chinese Script, Culture Relics and Artificial Intelligence, Jilin Province" + } + ], + [ + { + "type": "page_number", + "bbox": [ + 0.218, + 0.116, + 0.23, + 0.127 + ], + "angle": 0, + "content": "2" + }, + { + "type": "header", + "bbox": [ + 0.272, + 0.115, + 0.347, + 0.127 + ], + "angle": 0, + "content": "Zhou et al." + }, + { + "type": "image", + "bbox": [ + 0.222, + 0.145, + 0.782, + 0.224 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.214, + 0.235, + 0.788, + 0.292 + ], + "angle": 0, + "content": "Fig. 1: Our pair-searching and -matching task. Given a large number of mixed image fragments, our proposed network can search and match pairs of fragments, which is important for restoration problems. We also designed an algorithm to generate a dataset of image fragments by tearing a set of complete images." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.325, + 0.784, + 0.446 + ], + "angle": 0, + "content": "designed descriptors [18, 33]. It has a common core with the registration problem [3], which is to establish the correspondences of local features between related image pairs of the same scene. However, fragment assembly is more challenging because they do have overlapping regions. Thus, existing rule-based methods are difficult to generalize to various complex situations. Furthermore, they are time-consuming since the computational complexity of finding matching pairs from a set of fragments is typically \\( O(N^2) \\) in global search, and in practice, only a few pairs are hidden in a large number of mixed fragments." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.448, + 0.788, + 0.584 + ], + "angle": 0, + "content": "Solving jigsaw puzzles [14, 37, 42] is also a similar task and learning-based approaches have been applied [2, 36]. However, two differences make these algorithms not generalizable. First, puzzle pieces usually have regular shapes, so they are best matched based on content information rather than contour features. Second, this problem usually studies matching all fragments into a complete block. Their input fragments do not come from mixed groups, and the matching results are dense and not lost. Therefore, they often use neighbor relationships to build graphs to constrain the search for the correct alignment of fragments. Once a fragment is lost, their closed-loop strategy fails." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.587, + 0.788, + 0.737 + ], + "angle": 0, + "content": "In this paper, we propose PairingNet, an end-to-end deep learning network, for searching and matching pairs from a large number of image fragments. To address this problem, we first explore the use of Graph Convolutional Networks (GCNs) to extract more robust contour and texture features for our tasks. Second, we design a feature fusion way to learn local matching features by calculating the similarity matrix of each fragment pair. Third, we build a module based on the linear transformer to learn the global searching features for fragments using contrastive learning. Finally, we collect a real dataset and generate a image fragment dataset to provide data-driven and validate the effectiveness of our proposed network. The main contributions of our work are as follows:" + }, + { + "type": "text", + "bbox": [ + 0.225, + 0.75, + 0.784, + 0.809 + ], + "angle": 0, + "content": "- We propose a novel network to leverage the local contour and texture features for solving the pair-searching and -matching of image fragments. It provides an attempt to use advanced deep-learning techniques to solve this long-standing problem." + }, + { + "type": "text", + "bbox": [ + 0.225, + 0.81, + 0.784, + 0.84 + ], + "angle": 0, + "content": "- We provide a real dataset and a generated dataset with a carefully designed algorithm to simulate real-world fragments. The algorithm generates diverse" + }, + { + "type": "list", + "bbox": [ + 0.225, + 0.75, + 0.784, + 0.84 + ], + "angle": 0, + "content": null + } + ], + [ + { + "type": "header", + "bbox": [ + 0.657, + 0.115, + 0.732, + 0.129 + ], + "angle": 0, + "content": "PairingNet" + }, + { + "type": "page_number", + "bbox": [ + 0.775, + 0.117, + 0.787, + 0.127 + ], + "angle": 0, + "content": "3" + }, + { + "type": "text", + "bbox": [ + 0.241, + 0.147, + 0.785, + 0.176 + ], + "angle": 0, + "content": "image fragments to facilitate the application of subsequent learning-based methods." + }, + { + "type": "text", + "bbox": [ + 0.226, + 0.178, + 0.785, + 0.207 + ], + "angle": 0, + "content": "- We conduct extensive experiments, ablation studies, and visual analysis to demonstrate the effectiveness of our proposed network." + }, + { + "type": "title", + "bbox": [ + 0.216, + 0.231, + 0.388, + 0.248 + ], + "angle": 0, + "content": "2 Related Work" + }, + { + "type": "title", + "bbox": [ + 0.216, + 0.265, + 0.582, + 0.283 + ], + "angle": 0, + "content": "2.1 Fragments Assembly (Irregular Shape)" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.299, + 0.785, + 0.36 + ], + "angle": 0, + "content": "Here, we focus on the pairwise matching algorithms of related representative works, although few of them also consider fragment relationships. Methods based on deep neural networks have been studied for object fragments; however, we could not find representative works for image fragments or even suitable datasets." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.373, + 0.788, + 0.555 + ], + "angle": 0, + "content": "Image fragments. Kong and Kimia [25] solved fragment matching using geometric features (partial curve matching). Leitao and Stolfi [17] described a specific multiscale matching method for the reassembly of 2D fragmented objects. Tsamoura and Pitas [43] presented a color-based method to automatically reassemble image fragments. Huang et al. [22] relied on salient curves detected inside the different image pieces to align gapped fragments. Subsequently, both shape and appearance information along the boundaries were utilized and extracted for each piece for matching [32,49]. Furthermore, Derech et al. [12] proposed to extrapolate each fragment to address fragment abrasion. For learning-based methods, traditional machine learning (SVM/Random Forest) was used as a classifier to process extracted local features [1,38]. Then Le and Li [27] built a deep neural network to detect the compatibility of pairwise stitching." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.571, + 0.788, + 0.753 + ], + "angle": 0, + "content": "Object fragments. Huang et al. [23] analyzed the geometry of the fracture surfaces by computing multi-scale surface characteristics. Yang et al. [46] matched stone tools based on contour points and their mean normals. Funkhouser et al. [15] learned a decision tree classifier to filter the candidates. Hong et al. [21] solved the problem of indistinguishable false matches by employing beam search to explore multiple registration possibilities. Ye et al. [47] presented an immersive system that can handle complex and ambiguous fragment shapes interactively with experts. In recent years, deep learning networks have been rapidly applied in 3D assembly. Huang et al. [48] proposed an assembly-oriented dynamic graph learning framework to estimate the pose of 3D parts. Chen et al. [8] proposed Neural Shape Mating (NSM) to tackle pairwise 3D geometric shape mating. Wu et al. [45] leveraged SE(3) equivariance for such shape pose disentanglement." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.769, + 0.788, + 0.845 + ], + "angle": 0, + "content": "Datasets. Existing image assembly works typically collect data in two ways: manual digitalization by photographing or scanning broken fragments, or using simple algorithms to tear complete images [7, 35, 38]. The lack of a large-scale dataset prevents the emergence of deep learning networks for this task. However, multiple datasets have been introduced in 3D assembly problems [24, 26, 40]." + } + ], + [ + { + "type": "page_number", + "bbox": [ + 0.218, + 0.116, + 0.23, + 0.127 + ], + "angle": 0, + "content": "4" + }, + { + "type": "header", + "bbox": [ + 0.272, + 0.115, + 0.347, + 0.127 + ], + "angle": 0, + "content": "Zhou et al." + }, + { + "type": "image", + "bbox": [ + 0.218, + 0.144, + 0.789, + 0.349 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.214, + 0.357, + 0.788, + 0.483 + ], + "angle": 0, + "content": "Fig. 2: Pipeline of our proposed PairingNet. (a) Based on patches, we first employ a binary encoding to describe contours and use ResGCN as the backbone to extract the local contour and texture features of image fragments. (b) We design self-gated fusion to fuse the extracted features in a weighted manner to calculate the similarity matrix for each pair of fragments, and (c) a linear transformer-based encoder for learning the global features of each fragment. (d) We use a two-step strategy to train our proposed network. (e) During inference, we calculate cosine similarity to find the adjacent fragment pairs and further process the matching similarity matrix to establish better correspondences between a fragment pair." + }, + { + "type": "title", + "bbox": [ + 0.216, + 0.514, + 0.525, + 0.531 + ], + "angle": 0, + "content": "2.2 Jigsaw Puzzles (Regular Shape)" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.541, + 0.788, + 0.618 + ], + "angle": 0, + "content": "Solving jigsaw puzzles was first proposed in [14], and then regular pieces became a key research object in computer vision. Many existing methods [10, 16, 19, 37, 41, 42] focused on building relationships of pieces based on content information rather than contour-based pair-searching/-matching. And deep learning techniques [2, 36] was also used to solve this problem." + }, + { + "type": "title", + "bbox": [ + 0.216, + 0.631, + 0.331, + 0.646 + ], + "angle": 0, + "content": "3 Method" + }, + { + "type": "title", + "bbox": [ + 0.216, + 0.66, + 0.424, + 0.675 + ], + "angle": 0, + "content": "3.1 Problem Statement" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.683, + 0.788, + 0.749 + ], + "angle": 0, + "content": "Let us consider a fragment dataset \\(\\mathbf{F} = \\{f^{(i)}\\}_{i=1}^{N}\\), which consists of fragments \\(f\\) torn from a certain image dataset. We define the contour points \\(c_{j}^{(i)} \\in \\mathbf{C}^{(i)}\\) for each \\(f^{(i)}\\), where \\(j = 1,2,\\dots,M^{(i)}\\), \\(M^{(i)}\\) is the contour length and \\(\\mathbf{C}^{(i)}\\) is an ordered set that contains all contour points of \\(f^{(i)}\\)." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.75, + 0.788, + 0.841 + ], + "angle": 0, + "content": "Regarding any two neighbor fragments sampled from a pairset \\(\\mathbf{P}\\), we denote the pair as \\(f^{(m)}\\) and \\(f^{(n)}\\). The pair-searching task aims to retrieve the pairset \\(\\mathbf{P}\\) from a given fragment dataset \\(\\mathbf{F}\\). Meanwhile, for the contours of the paired fragments, \\(\\mathbf{C}^{(m)}\\) and \\(\\mathbf{C}^{(n)}\\), we can always find the overlapped part \\(\\mathbf{C}^{(m\\& n)}\\). The pair-matching task is designed to find the adjacent part from any given fragment pair." + } + ], + [ + { + "type": "header", + "bbox": [ + 0.657, + 0.115, + 0.732, + 0.129 + ], + "angle": 0, + "content": "PairingNet" + }, + { + "type": "page_number", + "bbox": [ + 0.775, + 0.117, + 0.785, + 0.127 + ], + "angle": 0, + "content": "5" + }, + { + "type": "title", + "bbox": [ + 0.216, + 0.147, + 0.444, + 0.16 + ], + "angle": 0, + "content": "3.2 Network Architecture" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.174, + 0.788, + 0.31 + ], + "angle": 0, + "content": "In the pair-matching task, a model has the capability to discern differences between two fragments, regardless of whether these discrepancies are on a large or minimal scale. Simultaneously, the features extracted by the model can be used to quantify the likelihood of adjacency between any given two fragments, assisting in the identification of matched pairs. However, the process of comparing every two fragments is not only laborious, but also inefficient. Therefore, utilizing the rich features the matching model extracts, we conceived a pair-searching module. This module can adeptly convert any fragment into a unique feature vector, thereby significantly accelerating the pair-searching process." + }, + { + "type": "title", + "bbox": [ + 0.216, + 0.324, + 0.376, + 0.337 + ], + "angle": 0, + "content": "Feature Extraction" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.358, + 0.788, + 0.512 + ], + "angle": 0, + "content": "Contour encoding. In this module, we aim to direct the model's attention to the unique features of contours. We design distinct contour encodings that also may be applied in other sketch-based fields. We tested various patch sizes and all these distinct encodings, finalizing a binary encoding and a patch size of \\(7 \\times 7\\) for our network. Subsequently, these contour patches are processed through a convolutional layer, followed by an average pooling layer and a fully connected network. We then adopt ResGCN [28] to amalgamate adjacent features, further extracting comprehensive global features. As such, for any provided fragment input denoted as \\(f^{(i)}\\) with its corresponding contour \\(\\mathbf{C}^{(i)}\\), this module yields a return of \\(f_{c}^{(i)} \\in \\mathbb{R}^{M^{(i)} \\times 64}\\)." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.521, + 0.788, + 0.645 + ], + "angle": 0, + "content": "Textural encoding. Our objective is to accentuate the distinguishing characteristics of textures in this module. We begin by cropping the image using a patch centered on each contour point. This is followed by filtering these patches through two consecutive convolutional layers, coupled with a single average pooling operation. Then, a ResGCN [28] is integrated to compile adjacent features, thereby enhancing the receptive domain of local textural characteristics. For any given fragment input represented as \\( f^{(i)} \\) along with its associated contour \\( \\mathbf{C}^{(i)} \\), this module produces an output of \\( f_{t}^{(i)} \\in \\mathbb{R}^{M^{(i)} \\times 64} \\)." + }, + { + "type": "title", + "bbox": [ + 0.216, + 0.653, + 0.406, + 0.669 + ], + "angle": 0, + "content": "Pair-matching Module" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.688, + 0.788, + 0.793 + ], + "angle": 0, + "content": "Feature Fusion. Regarding the decision-making process for pair-matching and -searching tasks, the dependence on contour and textural features varies. For instance, in scenarios where both fragments are solid colors, texture serves as the primary identifying factor suggesting adjacency. However, texture alone cannot identify the matched contours. In contrast, simple contours, like straight lines, require texture features for accurate matching position determination, as contour information alone falls short." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.795, + 0.788, + 0.84 + ], + "angle": 0, + "content": "Given this, we propose the incorporation of a Feature Fusion module [11] which would be capable of adaptively fine-tuning the weights assigned to these two features. The specific structure of this integral component is depicted in" + } + ], + [ + { + "type": "page_number", + "bbox": [ + 0.218, + 0.116, + 0.23, + 0.127 + ], + "angle": 0, + "content": "6" + }, + { + "type": "header", + "bbox": [ + 0.272, + 0.115, + 0.347, + 0.128 + ], + "angle": 0, + "content": "Zhou et al." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.145, + 0.784, + 0.182 + ], + "angle": 0, + "content": "Figure 2 (b). First, we concatenate \\( f_{t}^{(i)} \\) and \\( f_{c}^{(i)} \\), then the weight \\( \\mathbf{w}^{(i)} \\) for each contour point \\( c_{j}^{(i)} \\in \\mathbf{C}^{(i)} \\) is computed by:" + }, + { + "type": "equation", + "bbox": [ + 0.393, + 0.193, + 0.785, + 0.211 + ], + "angle": 0, + "content": "\\[\n\\mathbf {w} ^ {(i)} = \\sigma \\left(\\mathbf {W} _ {g} \\left(f _ {t} ^ {(i)} \\odot f _ {c} ^ {(i)}\\right) + \\mathbf {b} _ {g}\\right) ^ {(i)} \\tag {1}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.215, + 0.221, + 0.784, + 0.253 + ], + "angle": 0, + "content": "where \\(\\sigma\\) is sigmoid the activation function, \\((\\widehat{\\mathbf{c}})\\) means channel-wise concatenation, and \\(\\mathbf{w} \\in \\mathbb{R}^{M^{(i)} \\times 64}\\). Second, we fuse features of two modalities:" + }, + { + "type": "equation", + "bbox": [ + 0.379, + 0.263, + 0.785, + 0.282 + ], + "angle": 0, + "content": "\\[\n\\mathbf {F} _ {f} ^ {(i)} = \\mathbf {w} ^ {(i)} \\odot \\mathbf {F} _ {t} ^ {(i)} + (1 - \\mathbf {w} ^ {(i)}) \\odot \\mathbf {F} _ {c} ^ {(i)} \\tag {2}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.215, + 0.293, + 0.784, + 0.34 + ], + "angle": 0, + "content": "where \\(\\mathbf{F}_f^{(i)}\\in \\mathbb{R}^{M^{(i)}\\times 64}\\), and \\(\\odot\\) is element-wise multiplication. With this module, we let the model adaptively decide the weights between textual and contour features, which is proven effective in later experiments." + }, + { + "type": "text", + "bbox": [ + 0.215, + 0.349, + 0.784, + 0.396 + ], + "angle": 0, + "content": "Similarity Matrix. As shown in Figure 2 (b-2), for every pair of fragments \\( f^{(m)} \\) and \\( f^{(n)} \\) from the pairwise matrix \\( \\mathbf{S}^{(m\\& n)} \\in \\mathbb{R}^{M^{(m)} \\times M^{(n)}} \\) [9,39]." + }, + { + "type": "text", + "bbox": [ + 0.24, + 0.397, + 0.763, + 0.412 + ], + "angle": 0, + "content": "We first calculate the similarity matrix \\(\\tilde{\\mathbf{S}}^{(m\\& n)}\\) by matrix multiplication:" + }, + { + "type": "equation", + "bbox": [ + 0.413, + 0.422, + 0.785, + 0.449 + ], + "angle": 0, + "content": "\\[\n\\tilde {\\mathbf {S}} ^ {(m \\& n)} = \\frac {1}{\\sqrt {D}} \\mathbf {F} _ {f} ^ {(m)} \\left(\\mathbf {F} _ {f} ^ {(n)}\\right) ^ {T} \\tag {3}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.215, + 0.458, + 0.784, + 0.487 + ], + "angle": 0, + "content": "where \\(D\\) is the dimension of the feature, in our model, \\(D\\) is 64. Then, each entry in the matrix \\(\\mathbf{S}^{(m\\& n)}\\) is normalized:" + }, + { + "type": "equation", + "bbox": [ + 0.342, + 0.496, + 0.785, + 0.535 + ], + "angle": 0, + "content": "\\[\n\\mathbf {S} ^ {(m \\& n)} (i, j) = \\frac {\\exp \\left(\\tilde {\\mathbf {S}} _ {i , j} ^ {(m \\& n)}\\right)}{\\sum_ {i} \\exp \\left(\\tilde {\\mathbf {S}} _ {i , j} ^ {(m \\& n)}\\right)} \\frac {\\exp \\left(\\tilde {\\mathbf {S}} _ {i , j} ^ {(m \\& n)}\\right)}{\\sum_ {j} \\exp \\left(\\tilde {\\mathbf {S}} _ {i , j} ^ {(m \\& n)}\\right)} \\tag {4}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.215, + 0.559, + 0.784, + 0.621 + ], + "angle": 0, + "content": "Pair-searching Module We introduce a modified feature fusion module to process features extracted by the backbone network and encode any fragment \\( f^{(i)} \\) into a feature vector \\( v_{s}^{(i)} \\in \\mathbb{R}^{128} \\). It is based on the discovery that using fused features in the pair-matching task will lower the performance." + }, + { + "type": "text", + "bbox": [ + 0.215, + 0.622, + 0.784, + 0.711 + ], + "angle": 0, + "content": "To be specific, as shown in Figure 2 (c), every feature branch utilizes a Linear Transformer Encoder [44] to enhance the encoding of both contour and textural features. Subsequent to the modified fusion module, we add a fully-connected layer to encode the fused feature regarding per contour point. Then an average pooling layer is applied to map the feature matrix to a singular vector. Finally, another fully-connected layer is used to encode the feature." + }, + { + "type": "text", + "bbox": [ + 0.215, + 0.712, + 0.784, + 0.775 + ], + "angle": 0, + "content": "For every pair of fragment \\( f^{(i)} \\) and \\( f^{(j)} \\) from the fragment set \\( \\mathbf{F} \\), this module allows them to be encoded as \\( v_{s}^{(i)} \\) and \\( v_{s}^{(j)} \\). By calculating their inner product, we can efficiently quantify and contrast their similarity, thereby expediently identifying neighbor pairs." + }, + { + "type": "title", + "bbox": [ + 0.215, + 0.791, + 0.377, + 0.805 + ], + "angle": 0, + "content": "3.3 Loss Function" + }, + { + "type": "text", + "bbox": [ + 0.215, + 0.81, + 0.784, + 0.84 + ], + "angle": 0, + "content": "We use two loss functions to train our network, and the training process is shown in Figure 2 (d)." + } + ], + [ + { + "type": "header", + "bbox": [ + 0.657, + 0.115, + 0.732, + 0.129 + ], + "angle": 0, + "content": "PairingNet" + }, + { + "type": "page_number", + "bbox": [ + 0.776, + 0.117, + 0.786, + 0.127 + ], + "angle": 0, + "content": "7" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.147, + 0.784, + 0.192 + ], + "angle": 0, + "content": "Pair-matching Loss Given the extreme imbalance in the ratio of positive to negative samples in the pair-matching task, we adopt the concept of focal loss [30] to concentrate on learning challenging correspondences." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.192, + 0.785, + 0.245 + ], + "angle": 0, + "content": "For any paired fragments \\( f^{(m)} \\) and \\( f^{(n)} \\), entries in the ground truth similarity matrix \\( \\mathbf{S}_{gt}^{(m\\& n)} \\) are 1 for matched contour points and 0 for unmatched ones. We calculate the matching loss \\( \\mathcal{L}_{match}^{(m\\& n)} \\) as follows:" + }, + { + "type": "equation", + "bbox": [ + 0.328, + 0.249, + 0.785, + 0.291 + ], + "angle": 0, + "content": "\\[\n\\begin{array}{l} \\mathcal {L} _ {\\text {m a t c h}} ^ {(m \\& n)} = - \\sum_ {(0, 0)} (\\beta_ {1} (\\mathbf {1} - \\mathbf {S} ^ {(m \\& n)}) ^ {\\gamma} \\log \\mathbf {S} ^ {(m \\& n)} \\mathbf {S} _ {g t} ^ {(m \\& n)}) \\tag {5} \\\\ + \\beta_ {2} (\\mathbf {S} ^ {(m \\& n)}) ^ {\\gamma} \\log \\left(\\mathbf {1} - \\mathbf {S} ^ {(m \\& n)}\\right) (\\mathbf {1} - \\mathbf {S} _ {g t} ^ {(m \\& n)}) \\\\ \\end{array}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.301, + 0.784, + 0.362 + ], + "angle": 0, + "content": "where \\(\\beta_{1}\\) and \\(\\beta_{2}\\) are the balancing weights and \\(\\beta_{1} + \\beta_{2} = 1\\), \\(\\gamma\\) is the decay factor. \\(\\sum\\) adds all the entries in the input matrix and all operations except \\(\\sum\\) are element-wise. By calculating and adding the pair-matching loss for all paired fragments \\(f^{(m)}, f^{(n)}\\) from the pairwise \\(\\mathbf{P}\\), we get \\(\\mathcal{L}_{match}\\)." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.374, + 0.785, + 0.421 + ], + "angle": 0, + "content": "Pair-searching Loss In the pair-searching task, we employ InfoNCE loss [34] as the contrastive loss function \\(\\mathcal{L}_{search}\\). We utilize paired samples for constructing positive sample pairs, and unpaired samples for negative sample pairs." + }, + { + "type": "title", + "bbox": [ + 0.215, + 0.437, + 0.34, + 0.451 + ], + "angle": 0, + "content": "3.4 Inference" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.456, + 0.784, + 0.485 + ], + "angle": 0, + "content": "Pair-matching In this section, we target to retrieve matched contour from the similarity matrix \\(\\mathbf{S}^{(m\\& n)}\\)." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.487, + 0.785, + 0.638 + ], + "angle": 0, + "content": "We first apply a threshold \\(\\epsilon = 0.006\\) to filter out noisy entries. Entries with value lower than the threshold is set to 0. Because contour pointset \\(\\mathbf{C}^{(m)}\\) and \\(\\mathbf{C}^{(n)}\\) are ordered sets, for any ground truth similarity matrix \\(\\mathbf{S}_{gt}^{(m\\& n)}\\), if the start for contour points is the left-top corner, we can observe that the part with a value of 1 in the matrix must be arranged in the direction from upper left to lower right and connected into a line segment. Therefore, this inspires us to enhance the filtered similarity matrix. We design an eroding kernel \\(\\mathbf{K}_e\\) to filter out not dominant segments, which is adopted once. Then a dilating kernel \\(\\mathbf{K}_d\\) is applied once to connect close but split segments. The definitions of two kernels are as follows:" + }, + { + "type": "equation", + "bbox": [ + 0.394, + 0.652, + 0.785, + 0.695 + ], + "angle": 0, + "content": "\\[\n\\mathbf {K} _ {e} = \\left[ \\begin{array}{l l l} 0 & 0 & 1 \\\\ 0 & 0 & 0 \\\\ 1 & 0 & 0 \\end{array} \\right], \\quad \\mathbf {K} _ {d} = \\left[ \\begin{array}{l l l} 0 & 0 & 1 \\\\ 0 & 1 & 0 \\\\ 1 & 0 & 0 \\end{array} \\right]. \\tag {6}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.699, + 0.784, + 0.73 + ], + "angle": 0, + "content": "Finally, we use RANSAC [13] to estimate the correspondences of two fragments and obtain the transformation matrix." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.742, + 0.784, + 0.819 + ], + "angle": 0, + "content": "Pair-searching For N fragments \\( f^{(i)} \\in \\mathbf{F}, i = 1,2,\\dots,N \\), using the searching module, we can get a feature matrix \\( \\mathbf{V}_s = [v_1,v_2,\\ldots ,v_N]^T \\in \\mathbb{R}^{N\\times 128} \\). Subsequently, we compute the cosine similarity between all global features to form the cosine similarity matrix. This allows us to obtain the similarity score for any pair of fragments." + }, + { + "type": "title", + "bbox": [ + 0.215, + 0.829, + 0.329, + 0.845 + ], + "angle": 0, + "content": "4 Dataset" + } + ], + [ + { + "type": "page_number", + "bbox": [ + 0.218, + 0.116, + 0.23, + 0.127 + ], + "angle": 0, + "content": "8" + }, + { + "type": "header", + "bbox": [ + 0.272, + 0.115, + 0.347, + 0.127 + ], + "angle": 0, + "content": "Zhou et al." + }, + { + "type": "image", + "bbox": [ + 0.222, + 0.149, + 0.782, + 0.252 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.216, + 0.26, + 0.788, + 0.289 + ], + "angle": 0, + "content": "Fig. 3: Dataset creation. (a) The process of generating our dataset. (b) Illustration of Algorithm 1 for tearing a complete image." + }, + { + "type": "table_caption", + "bbox": [ + 0.216, + 0.308, + 0.789, + 0.339 + ], + "angle": 0, + "content": "Table 1: Dataset division. We divide the entire test set into three difficulty levels according to the contour overlapping proportion of fragment pairs." + }, + { + "type": "table", + "bbox": [ + 0.334, + 0.348, + 0.673, + 0.404 + ], + "angle": 0, + "content": "
TotalTrainValidationTest
FullHighMediumLow
Fragments819640988193279107116711090
Pairs14951365413723706631103604
" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.435, + 0.788, + 0.618 + ], + "angle": 0, + "content": "Generated Dataset. We design Algorithm 1 to tear a series of complete images and create a new image fragment dataset with data close to the real world. The process is also shown in Figure 3. To control the shape and area of the fragments, we first find the circumscribed circle of a fragment and take two points at an appropriate distance on the circle to determine the starting and ending points of the segmentation. Then, to simulate a more realistic fragment contour, we randomly divide the cutting line into several segments, each segment randomly choosing a straight line or an irregular curve. The irregular curve is synthesized by a series of Fourier orthogonal bases to make the shape of the fragments more diverse and similar to the edge of a real torn paper. We set a minimum threshold for the fragments, and the fragment is re-segmented if a very small fragment appears." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.619, + 0.79, + 0.816 + ], + "angle": 0, + "content": "To create our dataset, we collect 390 images of ten themes from Pexels [4] and empirically set the generation parameters shown in Algorithm 1 to obtain fragments. We calculate the distributions of our generated data, as shown in Figure 4. The distribution of fragment areas exhibits a long-tail pattern, which is consistent with the distribution of fragment areas observed when tearing large amounts of paper in real-life scenarios. The contour overlapping proportion of fragment pairs is closer to a uniform distribution over most ranges, indicating that our algorithm generates various types of fragment pairs. Then, we divide the dataset into training set, validation set, and test set in 5:1:4. We further divide the full test set into three difficulty levels based on the overlapping proportion of contour segments between fragment pairs. Low difficulty means that the pair of fragments has larger overlapping contour segments, and vice versa. The details are shown in Table 1." + } + ], + [ + { + "type": "header", + "bbox": [ + 0.657, + 0.115, + 0.732, + 0.129 + ], + "angle": 0, + "content": "PairingNet" + }, + { + "type": "page_number", + "bbox": [ + 0.776, + 0.117, + 0.786, + 0.127 + ], + "angle": 0, + "content": "9" + }, + { + "type": "code_caption", + "bbox": [ + 0.218, + 0.146, + 0.486, + 0.159 + ], + "angle": 0, + "content": "Algorithm 1 Generate fragments of an image" + }, + { + "type": "algorithm", + "bbox": [ + 0.217, + 0.16, + 0.787, + 0.578 + ], + "angle": 0, + "content": "Input: List of fragments \\(L_{frag} = \\{\\}\\) , a complete image \\(I\\) Max iterations \\(t_{max} = 40\\) min arc length ratio \\(\\tau = 0.9\\) max number of cutting points \\(N_{max} =\\) 3, min distance from cutting point to edge of fragment \\(D_{min} = 100\\) , the number of Fourier orthogonal basis \\(n = 20\\) scaling ratios \\(s_1 = 0.25\\) \\(s_2 = 0.0067\\) \\(s_3 = 1.5\\) \\(s_4 = 0.3\\) , probability of curve type \\(\\rho = 0.5\\) , min fragment size \\(h_{min} = 150\\) \\(w_{min} = 150\\) \nOutput: Collection of fragments \\(L_{frag}\\) \n1: put I in \\(L_{frag}\\) \n2: while t < tmax do \n3: f<-randomly_pick(Lfrag), R<-bounding_rectangle(f) \n4: W,H<-width_and_height(R), C<-circumcircle(R) \n5: ldia<-diameter(C) \n6: \n7: // generate the cutting points \n8: repeat \n9: pa, pb<-randomly_pick(C) \n10: larc<-smaller_ar_c_length(pa,pb) \n11: until larc>τ·half_perimeter(C) \n12: pstart, pend<-intersection(f,pa,pb) \n13: m<-randomly_pick{0,...,Nmax} \n14: repeat \n15: [p1, ..., pm]<-random_select_insidepoints(f,m) \n16: d<-min_distance_to_edge(f,p1, ..., pm) \n17: until d > Dmin \n18: put pstart, pend in [p1, ..., pm] \n19: \n20: // generate the cutting segments \n21: for each pi∈[pstart, p1, ..., pm, pend] do \n22: irregular line \\(l_{irr} = \\{\\}\\) , phase \\(\\phi \\sim \\mathcal{U}(-\\pi ,\\pi)\\) amplitude A \\(\\sim \\mathcal{N}(s_1H,s_2H)\\) period T \\(\\sim \\mathcal{N}(s_3W,s_4W)\\) \n23: for each x-coordinate between (pi,pi+1) do \n24: y<∑i=0nA sin(2πix+φ)+H/2 \n25: put {x,y} in lirr \n26: end for \n27: lstr← straight_line(pi,pi+1) \n28: lcut← select_and_connect(lirr,lstr,ρ,lcut) \n29: end for \n30: f1, f2<-cut_structure(f,lcut) \n31: if size(f1)>hmin,wmin and size(f2)>hmin,wmin then \n32: put f1, f2 in Lfrag \n33: end if \n34: end while" + }, + { + "type": "text", + "bbox": [ + 0.216, + 0.606, + 0.787, + 0.697 + ], + "angle": 0, + "content": "Real Dataset. We also collect a series of real image fragments to construct a real dataset. We printed 34 complete images (not included in our generated dataset), manually cut them into fragments, and then scanned them. Then, we apply some post-processing steps to the scanned image fragments to obtain the contour of the fragments and the annotation of the matching points. More details about real dataset can be found in the supplementary materials." + }, + { + "type": "title", + "bbox": [ + 0.216, + 0.71, + 0.376, + 0.727 + ], + "angle": 0, + "content": "5 Experiments" + }, + { + "type": "title", + "bbox": [ + 0.216, + 0.731, + 0.44, + 0.746 + ], + "angle": 0, + "content": "5.1 Comparison Methods" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.749, + 0.787, + 0.841 + ], + "angle": 0, + "content": "Finding comparison methods is not easy. Because rule-based methods have been around for quite some time, lack source code, and tend to exhibit lower performance compared to advanced deep learning methods. However, most existing deep learning methods require extensive modifications if applied to our task. Therefore, we employ Jigsawnet [27] (without loop consistency) and a rule-based method [49] (Step 1) as our comparison methods." + } + ], + [ + { + "type": "page_number", + "bbox": [ + 0.218, + 0.116, + 0.236, + 0.127 + ], + "angle": 0, + "content": "10" + }, + { + "type": "header", + "bbox": [ + 0.272, + 0.115, + 0.347, + 0.127 + ], + "angle": 0, + "content": "Zhou et al." + }, + { + "type": "image", + "bbox": [ + 0.217, + 0.145, + 0.505, + 0.217 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.519, + 0.144, + 0.785, + 0.217 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.214, + 0.226, + 0.788, + 0.283 + ], + "angle": 0, + "content": "Fig. 4: Dataset statistics. (a) Image area distribution of the generated fragments. (b) Proportional distribution of overlapping contour segments of the generated fragments. (c) Image area distribution of the scanned real fragments. (b) Proportional distribution of overlapping contour segments of the scanned real fragments." + }, + { + "type": "image", + "bbox": [ + 0.223, + 0.296, + 0.49, + 0.369 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.498, + 0.298, + 0.782, + 0.369 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.215, + 0.379, + 0.784, + 0.407 + ], + "angle": 0, + "content": "Fig. 5: Violin plots of the matching results of our method and the comparison method: (a) RE, NTE, and HD on the test set. (b) HD on the test sets of different difficulties." + }, + { + "type": "title", + "bbox": [ + 0.215, + 0.438, + 0.456, + 0.453 + ], + "angle": 0, + "content": "5.2 Implementation Details" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.464, + 0.788, + 0.54 + ], + "angle": 0, + "content": "Our PairingNet was implemented by PyTorch, and conducted on a server with 4 RTX A40 GPUs and Intel® Xeon® Gold 5220 CPUs (72 cores). We used the Adam optimizer with an initial learning rate of 0.001, which was adjusted using a cosine annealing strategy. For training pair-matching module, the batch size was set to 20, and for the pair-searching module, the batch size was set to 175." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.54, + 0.788, + 0.661 + ], + "angle": 0, + "content": "For comparison methods, given that the performance bottleneck of them resides in the CPU, utilizing GPUs for computational acceleration is deemed inappropriate. Thus we use two more powerful CPUs computing platforms to run the comparison method. We used the official sourcecode (implemented by TensorFlow) of Jigsawnet and conducted experiments on a server with 4 Quadro RTX 6000 GPUs and Intel® Xeon® Gold 5220 CPUs (144 cores). We reproduced the rule-based method by Python and conducted experiments on a node of a Supercomputing platform with Intel® Xeon® Gold 6258R CPUs (128 cores)." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.661, + 0.788, + 0.737 + ], + "angle": 0, + "content": "For evaluation metrics, we employ Recall@k and NDCG@k [5] for the pair-searching task, and Registration recall (RR), Hausdorff distance (HD), Radians Error (RE) and Normalized Translation Error (NTE) for the pair-matching task. We normalize Translation Error (TE) by the sum of the areas of fragment pairs to obtain the NTE, considering the large difference in area between fragments." + }, + { + "type": "title", + "bbox": [ + 0.215, + 0.76, + 0.324, + 0.774 + ], + "angle": 0, + "content": "5.3 Results" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.78, + 0.789, + 0.842 + ], + "angle": 0, + "content": "Performance on Generated Dataset. As shown in Table 2, our network outperforms the comparison methods on both the full test set and on every set of difficulty levels. Jigsawnet shows a certain superiority in pair-searching on the high-difficulty set, probably because it treats the problem as a simple binary" + } + ], + [ + { + "type": "header", + "bbox": [ + 0.657, + 0.115, + 0.732, + 0.129 + ], + "angle": 0, + "content": "PairingNet" + }, + { + "type": "page_number", + "bbox": [ + 0.769, + 0.116, + 0.785, + 0.127 + ], + "angle": 0, + "content": "11" + }, + { + "type": "table_caption", + "bbox": [ + 0.216, + 0.145, + 0.788, + 0.187 + ], + "angle": 0, + "content": "Table 2: Comparison Results of various metrics. The best results in our full dataset are highlighted in bold font, and the best results for each difficulty are marked in different colors." + }, + { + "type": "table", + "bbox": [ + 0.22, + 0.199, + 0.785, + 0.327 + ], + "angle": 0, + "content": "
MethodDifficultyRecall@5Recall@10Recall@20NDCG@5NDCG@10NDCG@20RR ↑HD ↓RE ↓NTE ↓
Rule-based [49]High0.2610.3540.4430.1850.2180.2410.107212.0031.28937.100×10-4
Medium0.3290.4270.5290.2420.2750.3020.153205.4631.24933.742×10-4
Low0.3580.4640.5550.2610.2960.320.188210.2741.23132.692×10-4
All0.3170.4160.5110.2480.2840.3110.149208.5191.25634.414×10-4
Jigsawnet [27]High0.3050.3940.4780.2310.2610.2840.117226.0261.30637.011×10-4
Medium0.3830.4780.5650.3090.340.3640.213206.0401.15531.467×10-4
Low0.4590.5480.6310.3680.3970.4190.241202.9421.18630.500×10-4
All0.380.4720.5570.3260.3590.3830.194210.8421.20532.772×10-4
OursHigh0.2760.4130.5350.1890.2340.2680.60692.2290.80718.335×10-4
Medium0.5010.6190.7320.3820.4230.4540.87935.6170.2365.685×10-4
Low0.7170.8050.8790.5810.610.6290.96112.3360.091.936×10-4
All0.4930.6080.7140.4170.4580.4870.83543.1160.3527.735×10-4
" + }, + { + "type": "image", + "bbox": [ + 0.223, + 0.343, + 0.784, + 0.513 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.216, + 0.525, + 0.788, + 0.555 + ], + "angle": 0, + "content": "Fig. 6: Examples of pair-searching results. Our network is able to accurately identify the corresponding target fragment (highlighted with a green box)." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.587, + 0.789, + 0.739 + ], + "angle": 0, + "content": "classification task; however, their pair-matching results are very unsatisfactory. It should be noted that we did not evaluate Recall@1 and NDCG@1 here due to the many-to-many pairings in our dataset, making it unreasonable to search only for the fragment with the highest similarity of searching. We can also observe that the performance of our proposed network shows reasonable variation with the difficulty of the data. This shows that our proposed network learns useful information and has more potential in complex situations. However, Jigsawnet and the rule-based method show relatively identical performance on test sets of different difficulty levels. Furthermore, we also plot the distributions of HD, RE, and NTE for each method for a more intuitive comparison, as shown in Figure 5." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.75, + 0.789, + 0.841 + ], + "angle": 0, + "content": "Performance on Real Dataset. We also tested our method on the real dataset. In this experiment, both our method and the comparison method were trained on the generated dataset without retraining or fine-tuning. As shown in Table 3 and Figure 8, PairingNet obtained satisfactory results for both pair-searching and -matching tasks. These results confirmed that Algorithm 1 simulates real data well and our method has good generalization to real data." + } + ], + [ + { + "type": "page_number", + "bbox": [ + 0.218, + 0.116, + 0.236, + 0.127 + ], + "angle": 0, + "content": "12" + }, + { + "type": "header", + "bbox": [ + 0.272, + 0.115, + 0.347, + 0.127 + ], + "angle": 0, + "content": "Zhou et al." + }, + { + "type": "image", + "bbox": [ + 0.218, + 0.145, + 0.788, + 0.202 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.214, + 0.213, + 0.788, + 0.285 + ], + "angle": 0, + "content": "Fig. 7: Examples of pair-matching results for three difficulty levels. Our proposed network achieves satisfactory matching results. However, for low-difficulty cases, Jigsawnet and the Rule-based method find approximate matching positions, while the matchings are not accurate enough. For medium- and high-difficulty cases, both comparison methods are difficult to obtain correct matching results." + }, + { + "type": "table_caption", + "bbox": [ + 0.215, + 0.296, + 0.789, + 0.327 + ], + "angle": 0, + "content": "Table 3: Comparison Results of various metrics on our real dataset. The best results are highlighted in bold font." + }, + { + "type": "table", + "bbox": [ + 0.248, + 0.336, + 0.756, + 0.381 + ], + "angle": 0, + "content": "
MethodRecall@5Recall@10Recall@20NDCG@5NDCG@10NDCG@20RR↑HD↓RE↓NTE↓
Rule-based [49]0.1380.2470.3660.0790.1140.1440.1287.6981.61430.746×10-4
Jigsawnet [27]0.2370.3260.4250.1850.2130.2380.1537.5701.54529.061×10-4
Ours0.4150.5640.6780.3210.3690.3990.7524.1820.4809.075×10-4
" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.406, + 0.789, + 0.541 + ], + "angle": 0, + "content": "Visualization. We show the top five search results of different methods for the examples of fragment pairs in Figure 6. Given a source fragment, our model can accurately identify the corresponding target fragment (marked with a green box). Furthermore, in situations where a source fragment can be paired with multiple target fragments, our method identifies all matching fragments within the top five search results. We also show examples of matching results for three difficulty levels in Figure 7. Our method produces matching results that are consistent with the Ground Truth (GT), while the comparison methods yield numerous incorrect results." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.542, + 0.789, + 0.723 + ], + "angle": 0, + "content": "We also visualize the normalized self-gated weights to analyze their changes under different situations. As shown in Figure 9 (a), we map the normalized weights to the edges of the fragments to show the influence of the texture/contour features. We classify the local parts of fragments into four types based on the information contained in texture and contour: those with rich texture and contour information (first quadrant), those with rich texture information (second quadrant), those with little texture and contour information (third quadrant), and those with rich contour information but with little texture information (fourth quadrant). The red color indicates that the weights of the texture features are relatively large, while the green color indicates that the contour feature weights are relatively large. The results show that our proposed network learned varying weights of the features for fragments with different characteristics." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.728, + 0.789, + 0.789 + ], + "angle": 0, + "content": "Inference Time. We test the inference time of different methods on the pairwise searching task on our generated test set. Our network retrieves all fragment pairs in 73 seconds, while the rule-based method takes 223 hours, and JigsawNet takes 10 hours." + }, + { + "type": "title", + "bbox": [ + 0.215, + 0.795, + 0.391, + 0.809 + ], + "angle": 0, + "content": "5.4 Ablation Study" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.81, + 0.789, + 0.842 + ], + "angle": 0, + "content": "Network Structures. We explore the performance of our network under different structures, as shown in Table 4. First, we test three feature fusion ways." + } + ], + [ + { + "type": "header", + "bbox": [ + 0.657, + 0.115, + 0.732, + 0.129 + ], + "angle": 0, + "content": "PairingNet" + }, + { + "type": "page_number", + "bbox": [ + 0.769, + 0.116, + 0.786, + 0.127 + ], + "angle": 0, + "content": "13" + }, + { + "type": "image", + "bbox": [ + 0.223, + 0.145, + 0.784, + 0.306 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.216, + 0.317, + 0.785, + 0.345 + ], + "angle": 0, + "content": "Fig. 8: Examples of pair-matching results of PairingNet on real dataset. Without retraining, our method can also effectively match the scanned real fragments." + }, + { + "type": "image", + "bbox": [ + 0.351, + 0.359, + 0.523, + 0.462 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.423, + 0.467, + 0.438, + 0.476 + ], + "angle": 0, + "content": "(a)" + }, + { + "type": "image", + "bbox": [ + 0.524, + 0.359, + 0.655, + 0.465 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.58, + 0.467, + 0.594, + 0.476 + ], + "angle": 0, + "content": "(b)" + }, + { + "type": "image_caption", + "bbox": [ + 0.214, + 0.486, + 0.788, + 0.61 + ], + "angle": 0, + "content": "Fig.9: Visual analysis of Texture/Contour. (a) We visualize the weights of texture and contour features in our network, with colors closer to red representing more use of texture features and closer to green representing more use of contour features. We can see that on edges with rich texture or close to straight lines, the weights of the extracted texture features are greater, while on edges with single texture, the weights of the contour features are greater. (b) Examples of Textur/Contour ablation study. When the local contour shapes are similar, our network has difficulty matching them correctly using only contour features, and vice versa. However, these cases can be solved when the two features are fused." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.64, + 0.788, + 0.732 + ], + "angle": 0, + "content": "In early fusion, instead of using two separate GCNs for feature extraction, we concatenate the two modalities and use only one GCN to learn the features. We also test using a simple concatenation to replace the self-gated fusion module. The results show that the performance of both pair-searching and -matching decreases. Additionally, we explore the impact of different numbers of layers on our network." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.735, + 0.788, + 0.842 + ], + "angle": 0, + "content": "Texture/Contour. To study the impact of contour and texture features on the performance of our network, we conducted experiments using only one modality (either contour or texture) and compared the final results. As shown in Table 4, the comparison results indicate that both modalities perform slightly worse when used independently. When both modalities are used in conjunction, our network achieves the best performance. We also show examples of failed pair-matching when the two modalities are used individually in Figure 9 (b)." + } + ], + [ + { + "type": "page_number", + "bbox": [ + 0.218, + 0.116, + 0.236, + 0.127 + ], + "angle": 0, + "content": "14" + }, + { + "type": "header", + "bbox": [ + 0.272, + 0.115, + 0.346, + 0.127 + ], + "angle": 0, + "content": "Zhou et al." + }, + { + "type": "table_caption", + "bbox": [ + 0.216, + 0.145, + 0.788, + 0.187 + ], + "angle": 0, + "content": "Table 4: Results of ablation studies. Comparison of the impact of different feature fusion ways, feature modalities, and network layers employed in our proposed network on the performance of pair-searching and pair-matching tasks." + }, + { + "type": "table", + "bbox": [ + 0.356, + 0.199, + 0.648, + 0.306 + ], + "angle": 0, + "content": "
Recall@5Recall@10NDCG@5NDCG@10RR
Fusion WayEarly fusion0.4480.5590.3760.4130.658
Concat0.4890.6050.4180.4580.824
Weighted0.4930.6080.4170.4580.835
ResGCN10----0.680
14----0.835
18----0.759
Transformer Encoder30.4880.6000.4100.450-
40.4880.6060.4110.452-
50.4930.6080.4170.458-
Feature ModalityContour-only0.0160.0250.0130.0160.584
Texture-only0.4790.5970.3940.4360.657
Contour + Texture0.4930.6080.4170.4580.835
" + }, + { + "type": "text", + "bbox": [ + 0.216, + 0.334, + 0.785, + 0.394 + ], + "angle": 0, + "content": "Besides, as shown in Figure 9 (a), it can be seen numerically that our network tends to assign more weight to the texture features. This is reasonable because textures usually contain richer information than the contours of image fragments, enabling the network to learn more discriminative features." + }, + { + "type": "text", + "bbox": [ + 0.216, + 0.399, + 0.785, + 0.429 + ], + "angle": 0, + "content": "Others. We report the results of ablation studies on patch size, contour encoding, and hyperparameters in our supplementary material." + }, + { + "type": "title", + "bbox": [ + 0.216, + 0.446, + 0.368, + 0.46 + ], + "angle": 0, + "content": "6 Conclusions" + }, + { + "type": "text", + "bbox": [ + 0.216, + 0.463, + 0.787, + 0.704 + ], + "angle": 0, + "content": "In this paper, to tackle the image fragment pair-searching and -matching tasks, we propose a novel network, PairingNet, to deduce the adjacent segments of fragments and encode the global features of each fragment. In PairingNet, ResGCN is employed as the backbone to extract the local contour and texture features of fragments. For pair-searching, a linear transformer-based module processes these local features leveraging the contrast of paired fragments and unpaired fragments. For pair-matching, we design a weighted fusion module to dynamically fuse the extracted local contour and texture features. Furthermore, to provide a foundation for learning-based methods, we design a cutting algorithm to generate a dataset by simulating real image fragments. Comprehensive experiments are conducted on our generated dataset and a real dataset, the results of which demonstrate the effectiveness of our PairingNet compared to existing methods. In future work, we will explore how to extend our work to other general technique, e.g. puzzles inspired MAE [20], JiGen [6] for self-supervised learning and domain generalization. The fragment generation algorithm that we designed can also create more diverse samples for image inpainting tasks [29,31]." + }, + { + "type": "text", + "bbox": [ + 0.216, + 0.705, + 0.787, + 0.795 + ], + "angle": 0, + "content": "Limitations While our research has made significant strides, fragment restoration remains a challenging task. Our dataset provides superior edge matching features. The task of fragment restoration becomes particularly arduous when the edges of the fragments are smooth and the texture is monotonous (for instance, in the case of shredded white paper), or when there is a substantial gap in the damaged edges." + }, + { + "type": "text", + "bbox": [ + 0.216, + 0.796, + 0.787, + 0.84 + ], + "angle": 0, + "content": "Acknowledgements This work was supported by the Young Scientists Fund of the National Natural Science Foundation of China (Grant No.62206106). We also feel grateful to Yifan Jin for his exploration in the preliminary work." + } + ], + [ + { + "type": "header", + "bbox": [ + 0.657, + 0.115, + 0.732, + 0.129 + ], + "angle": 0, + "content": "PairingNet" + }, + { + "type": "page_number", + "bbox": [ + 0.769, + 0.116, + 0.786, + 0.127 + ], + "angle": 0, + "content": "15" + }, + { + "type": "title", + "bbox": [ + 0.218, + 0.145, + 0.323, + 0.16 + ], + "angle": 0, + "content": "References" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.177, + 0.785, + 0.218 + ], + "angle": 0, + "content": "1. Abitbol, R., Shimshoni, I., Ben-Dov, J.: Machine learning based assembly of fragments of ancient papyrus. Journal on Computing and Cultural Heritage (JOCCH) 14(3), 1-21 (2021)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.219, + 0.785, + 0.26 + ], + "angle": 0, + "content": "2. Bridger, D., Danon, D., Tal, A.: Solving jigsaw puzzles with eroded boundaries. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 3526-3535 (2020)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.261, + 0.785, + 0.287 + ], + "angle": 0, + "content": "3. Brown, L.G.: A survey of image registration techniques. ACM computing surveys (CSUR) 24(4), 325-376 (1992)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.288, + 0.721, + 0.302 + ], + "angle": 0, + "content": "4. Bruno Joseph, I.J., Frese, D.: Pexels. https://www.pexels.com/ (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.302, + 0.785, + 0.357 + ], + "angle": 0, + "content": "5. Busa-Fekete, R., Szarvas, G., Elteto, T., Kégl, B.: An apple-to-apple comparison of learning-to-rank algorithms in terms of normalized discounted cumulative gain. In: ECAI 2012-20th European Conference on Artificial Intelligence: Preference Learning: Problems and Applications in AI Workshop. vol. 242. IOS Press (2012)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.358, + 0.785, + 0.398 + ], + "angle": 0, + "content": "6. Carlucci, F.M., D'Innocente, A., Bucci, S., Caputo, B., Tommasi, T.: Domain generalization by solving jigsaw puzzles. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 2229-2238 (2019)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.399, + 0.785, + 0.44 + ], + "angle": 0, + "content": "7. Chen, J., Tian, M., Qi, X., Wang, W., Liu, Y.: A solution to reconstruct cross-cut shredded text documents based on constrained seed k-means algorithm and ant colony algorithm. Expert Systems with Applications 127(AUG.), 35-46 (2019)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.44, + 0.785, + 0.495 + ], + "angle": 0, + "content": "8. Chen, Y.C., Li, H., Turpin, D., Jacobson, A., Garg, A.: Neural shape mating: Self-supervised object assembly with adversarial shape priors. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 12724-12733 (2022)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.495, + 0.785, + 0.536 + ], + "angle": 0, + "content": "9. Cheng, X., Lin, H., Wu, X., Yang, F., Shen, D.: Improving video-text retrieval by multi-stream corpus alignment and dual softmax loss. arXiv preprint arXiv:2109.04290 (2021)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.537, + 0.785, + 0.578 + ], + "angle": 0, + "content": "0. Cho, T.S., Avidan, S., Freeman, W.T.: A probabilistic image jigsaw puzzle solver. In: 2010 IEEE Computer society conference on computer vision and pattern recognition. pp. 183-190. IEEE (2010)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.579, + 0.785, + 0.62 + ], + "angle": 0, + "content": "1. Dai, Y., Gieseke, F., Oehmcke, S., Wu, Y., Barnard, K.: Attentional feature fusion. In: Proceedings of the IEEE/CVF winter conference on applications of computer vision. pp. 3560-3569 (2021)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.62, + 0.785, + 0.647 + ], + "angle": 0, + "content": "2. Derech, N., Tal, A., Shimshoni, I.: Solving archaeological puzzles. Pattern Recognition 119, 108065 (2021)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.647, + 0.785, + 0.688 + ], + "angle": 0, + "content": "3. Fischler, M.A., Bolles, R.C.: Random sample consensus: a paradigm for model fitting with applications to image analysis and automated cartography. Communications of the ACM 24(6), 381-395 (1981)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.689, + 0.785, + 0.73 + ], + "angle": 0, + "content": "4. Freeman, H., Gardner, L.: Apictorial jigsaw puzzles: The computer solution of a problem in pattern recognition. IEEE Transactions on Electronic Computers (2), 118-127 (1964)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.73, + 0.785, + 0.772 + ], + "angle": 0, + "content": "5. Funkhouser, T., Shin, H., Toler-Franklin, C., Castañeda, A.G., Brown, B., Dobkin, D., Rusinkiewicz, S., Weyrich, T.: Learning how to match fresco fragments. Journal on Computing and Cultural Heritage (JOCCH) 4(2), 1-13 (2011)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.772, + 0.785, + 0.799 + ], + "angle": 0, + "content": "6. Gallagher, A.C.: Jigsaw puzzles with pieces of unknown orientation. In: 2012 IEEE Conference on computer vision and pattern recognition. pp. 382-389. IEEE (2012)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.799, + 0.785, + 0.841 + ], + "angle": 0, + "content": "7. da Gama Leitao, H.C., Stolfi, J.: A multiscale method for the reassembly of two-dimensional fragmented objects. IEEE Transactions on Pattern Analysis and Machine Intelligence 24(9), 1239-1251 (2002)" + }, + { + "type": "list", + "bbox": [ + 0.226, + 0.177, + 0.785, + 0.841 + ], + "angle": 0, + "content": null + } + ], + [ + { + "type": "page_number", + "bbox": [ + 0.218, + 0.116, + 0.236, + 0.127 + ], + "angle": 0, + "content": "16" + }, + { + "type": "header", + "bbox": [ + 0.272, + 0.115, + 0.347, + 0.128 + ], + "angle": 0, + "content": "Zhou et al." + }, + { + "type": "ref_text", + "bbox": [ + 0.218, + 0.148, + 0.787, + 0.189 + ], + "angle": 0, + "content": "18. Guo, Y., Bennamoun, M., Sohel, F., Lu, M., Wan, J., Kwok, N.M.: A comprehensive performance evaluation of 3d local feature descriptors. International Journal of Computer Vision 116, 66-89 (2016)" + }, + { + "type": "ref_text", + "bbox": [ + 0.218, + 0.19, + 0.787, + 0.232 + ], + "angle": 0, + "content": "19. Gur, S., Ben-Shahar, O.: From square pieces to brick walls: The next challenge in solving jigsaw puzzles. In: Proceedings of the IEEE International Conference on Computer Vision. pp. 4029-4037 (2017)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.232, + 0.787, + 0.273 + ], + "angle": 0, + "content": "20. He, K., Chen, X., Xie, S., Li, Y., Dollar, P., Girshick, R.: Masked autoencoders are scalable vision learners. In: Proceedings of the IEEE/CVF conference on computer vision and pattern recognition. pp. 16000-16009 (2022)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.273, + 0.787, + 0.328 + ], + "angle": 0, + "content": "21. Hong, J.H., Yoo, S.J., Zeeshan, M.A., Kim, Y.M., Kim, J.: Structure-from-sherds: Incremental 3d reassembly of axially symmetric pots from unordered and mixed fragment collections. In: Proceedings of the IEEE/CVF International Conference on Computer Vision. pp. 5443-5451 (2021)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.328, + 0.787, + 0.37 + ], + "angle": 0, + "content": "22. Huang, H., Yin, K., Gong, M., Lischinski, D., Cohen-Or, D., Ascher, U.M., Chen, B.: \"mind the gap\": tele-registration for structure-driven image completion. ACM Trans. Graph. 32(6), 174-1 (2013)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.371, + 0.787, + 0.412 + ], + "angle": 0, + "content": "23. Huang, Q.X., Flóry, S., Gelfand, N., Hofer, M., Pottmann, H.: Reassembling fractured objects by geometric matching. In: ACM SIGGRAPH 2006 Papers, pp. 569-578 (2006)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.412, + 0.787, + 0.453 + ], + "angle": 0, + "content": "24. Jones, B., Hildreth, D., Chen, D., Baran, I., Kim, V.G., Schulz, A.: Automate: A dataset and learning approach for automatic mating of cad assemblies. ACM Transactions on Graphics (TOG) 40(6), 1-18 (2021)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.453, + 0.787, + 0.495 + ], + "angle": 0, + "content": "25. Kong, W., Kimia, B.B.: On solving 2d and 3d puzzles using curve matching. In: Proceedings of the 2001 IEEE Computer Society Conference on Computer Vision and Pattern Recognition. CVPR 2001. vol. 2, pp. II-II. IEEE (2001)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.495, + 0.787, + 0.55 + ], + "angle": 0, + "content": "26. Lamb, N., Palmer, C., Molloy, B., Banerjee, S., Banerjee, N.K.: Fantastic breaks: A dataset of paired 3d scans of real-world broken objects and their complete counterparts. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 4681-4691 (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.55, + 0.787, + 0.592 + ], + "angle": 0, + "content": "27. Le, C., Li, X.: Jigsawnet: Shredded image reassembly using convolutional neural network and loop-based composition. IEEE Transactions on Image Processing 28(8), 4000-4015 (2019)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.592, + 0.787, + 0.633 + ], + "angle": 0, + "content": "28. Li, G., Muller, M., Thabet, A., Ghanem, B.: Deep GCs: Can GCs go as deep as cnns? In: Proceedings of the IEEE/CVF international conference on computer vision. pp. 9267-9276 (2019)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.633, + 0.787, + 0.675 + ], + "angle": 0, + "content": "29. Li, W., Lin, Z., Zhou, K., Qi, L., Wang, Y., Jia, J.: Mat: Mask-aware transformer for large hole image inpainting. In: Proceedings of the IEEE/CVF conference on computer vision and pattern recognition. pp. 10758-10768 (2022)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.675, + 0.787, + 0.716 + ], + "angle": 0, + "content": "30. Li, Y., Harada, T.: Lepard: Learning partial point cloud matching in rigid and deformable scenes. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 5554-5564 (2022)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.716, + 0.787, + 0.758 + ], + "angle": 0, + "content": "31. Liu, G., Reda, F.A., Shih, K.J., Wang, T.C., Tao, A., Catanzaro, B.: Image inpainting for irregular holes using partial convolutions. In: Proceedings of the European conference on computer vision (ECCV). pp. 85-100 (2018)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.758, + 0.787, + 0.786 + ], + "angle": 0, + "content": "32. Liu, H., Cao, S., Yan, S.: Automated assembly of shredded pieces from multiple photos. IEEE transactions on multimedia 13(5), 1154-1162 (2011)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.786, + 0.787, + 0.814 + ], + "angle": 0, + "content": "33. Mikolajczyk, K., Schmid, C.: A performance evaluation of local descriptors. IEEE transactions on pattern analysis and machine intelligence 27(10), 1615-1630 (2005)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.814, + 0.787, + 0.842 + ], + "angle": 0, + "content": "34. Oord, A.v.d., Li, Y., Vinyals, O.: Representation learning with contrastive predictive coding. arXiv preprint arXiv:1807.03748 (2018)" + }, + { + "type": "list", + "bbox": [ + 0.217, + 0.148, + 0.787, + 0.842 + ], + "angle": 0, + "content": null + } + ], + [ + { + "type": "header", + "bbox": [ + 0.657, + 0.115, + 0.732, + 0.129 + ], + "angle": 0, + "content": "PairingNet" + }, + { + "type": "page_number", + "bbox": [ + 0.769, + 0.116, + 0.786, + 0.127 + ], + "angle": 0, + "content": "17" + }, + { + "type": "ref_text", + "bbox": [ + 0.214, + 0.147, + 0.788, + 0.189 + ], + "angle": 0, + "content": "35. Paixao, T.M., Boeres, M., Freitas, C., Oliveira-Santos, T.: Exploring character shapes for unsupervised reconstruction of strip-shredded text documents. IEEE Transactions on Information Forensics and Security pp. 1-1 (2018)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.19, + 0.787, + 0.232 + ], + "angle": 0, + "content": "36. Paumard, M.M., Picard, D., Tabia, H.: Image reassembly combining deep learning and shortest path problem. In: Proceedings of the European conference on computer vision (ECCV). pp. 153-167 (2018)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.232, + 0.786, + 0.259 + ], + "angle": 0, + "content": "37. Pomeranz, D., Shemesh, M., Ben-Shahar, O.: A fully automated greedy square jigsaw puzzle solver. In: CVPR 2011. pp. 9-16. IEEE (2011)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.26, + 0.786, + 0.285 + ], + "angle": 0, + "content": "38. Richter, F., Ries, C.X., Cebron, N., Lienhart, R.: Learning to reassemble shredded documents. IEEE Transactions on Multimedia 15(3), 582-593 (2013)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.287, + 0.787, + 0.327 + ], + "angle": 0, + "content": "39. Rocco, I., Cimpoi, M., Arandjelovic, R., Torii, A., Pajdla, T., Sivic, J.: Neighbourhood consensus networks. Advances in neural information processing systems 31 (2018)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.328, + 0.787, + 0.37 + ], + "angle": 0, + "content": "40. Sellan, S., Chen, Y.C., Wu, Z., Garg, A., Jacobson, A.: Breaking bad: A dataset for geometric fracture and reassembly. Advances in Neural Information Processing Systems 35, 38885-38898 (2022)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.371, + 0.787, + 0.397 + ], + "angle": 0, + "content": "41. Son, K., Hays, J., Cooper, D.B.: Solving square jigsaw puzzles with loop constraints. In: European conference on computer vision. pp. 32-46. Springer (2014)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.398, + 0.787, + 0.438 + ], + "angle": 0, + "content": "42. Son, K., Hays, J., Cooper, D.B.: Solving square jigsaw puzzle by hierarchical loop constraints. IEEE transactions on pattern analysis and machine intelligence 41(9), 2222-2235 (2018)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.439, + 0.787, + 0.466 + ], + "angle": 0, + "content": "43. Tsamoura, E., Pitas, I.: Automatic color based reassembly of fragmented images and paintings. IEEE Transactions on Image Processing 19(3), 680-690 (2009)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.467, + 0.787, + 0.494 + ], + "angle": 0, + "content": "44. Wang, S., Li, B.Z., Khabsa, M., Fang, H., Ma, H.: Linformer: Self-attention with linear complexity. arXiv preprint arXiv:2006.04768 (2020)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.495, + 0.787, + 0.535 + ], + "angle": 0, + "content": "45. Wu, R., Tie, C., Du, Y., Zhao, Y., Dong, H.: Leveraging se (3) equivariance for learning 3d geometric shape assembly. In: Proceedings of the IEEE/CVF International Conference on Computer Vision. pp. 14311-14320 (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.536, + 0.787, + 0.577 + ], + "angle": 0, + "content": "46. Yang, X., Matsuyama, K., Konno, K.: Pairwise matching of stone tools based on flake-surface contour points and normals. In: Proceedings of the Eurographics Workshop on Graphics and Cultural Heritage. pp. 125-129 (2017)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.578, + 0.787, + 0.618 + ], + "angle": 0, + "content": "47. Ye, S., Chen, Z., Chu, X., Li, K., Luo, J., Li, Y., Geng, G., Wu, Y.: Puzzlefixer: A visual reassembly system for immersive fragments restoration. IEEE transactions on visualization and computer graphics 29(1), 429-439 (2022)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.619, + 0.787, + 0.66 + ], + "angle": 0, + "content": "48. Zhan, G., Fan, Q., Mo, K., Shao, L., Chen, B., Guibas, L.J., Dong, H., et al.: Generative 3d part assembly via dynamic graph learning. Advances in Neural Information Processing Systems 33, 6315-6326 (2020)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.661, + 0.787, + 0.688 + ], + "angle": 0, + "content": "49. Zhang, K., Li, X.: A graph-based optimization algorithm for fragmented image reassembly. Graphical Models 76(5), 484-495 (2014)" + }, + { + "type": "list", + "bbox": [ + 0.214, + 0.147, + 0.788, + 0.688 + ], + "angle": 0, + "content": null + } + ] +] \ No newline at end of file diff --git a/2024/PairingNet_ A Learning-based Pair-searching and -matching Network for Image Fragments/3b568531-7960-492d-9c00-7cccb8dd34e1_origin.pdf b/2024/PairingNet_ A Learning-based Pair-searching and -matching Network for Image Fragments/3b568531-7960-492d-9c00-7cccb8dd34e1_origin.pdf new file mode 100644 index 0000000000000000000000000000000000000000..5984f3709fae96343588ab876e4727d0f636ccc9 --- /dev/null +++ b/2024/PairingNet_ A Learning-based Pair-searching and -matching Network for Image Fragments/3b568531-7960-492d-9c00-7cccb8dd34e1_origin.pdf @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:424ec87b410ee151f361b446156c3877c75fa5cb5b4e45f57a2ea94d385f823d +size 11979552 diff --git a/2024/PairingNet_ A Learning-based Pair-searching and -matching Network for Image Fragments/full.md b/2024/PairingNet_ A Learning-based Pair-searching and -matching Network for Image Fragments/full.md new file mode 100644 index 0000000000000000000000000000000000000000..64867fea2c8e50e4e39662d4423f0b247bff9ab0 --- /dev/null +++ b/2024/PairingNet_ A Learning-based Pair-searching and -matching Network for Image Fragments/full.md @@ -0,0 +1,339 @@ +# PairingNet: A Learning-based Pair-searching and -matching Network for Image Fragments + +Rixin Zhou $^{1}$ , Ding Xia $^{2}$ , Yi Zhang $^{3}$ , Honglin Pang $^{1}$ , Xi Yang $^{1,*,\dagger, \ddagger}$ , and Chuntao Li $^{4,*, \ddagger}$ + +$^{1}$ School of Artificial Intelligence, Jilin University + +2 The University of Tokyo + +$^{3}$ National University of Defense Technology, College of Electronic Engineering + +$^{4}$ School of Archaeology, Jilin University + +{zhourx22}@mails.jlu.edu.cn {1ct33}@jlu.edu.cn + +{dingxia1995, panghlwork, earthyangxi}@gmail.com {2356711993}@qq.com + +Abstract. In this paper, we propose a learning-based image fragment pair-searching and -matching approach to solve the challenging restoration problem. Existing works use rule-based methods to match similar contour shapes or textures, which are always difficult to tune hyperparameters for extensive data and computationally time-consuming. Therefore, we propose a neural network that can effectively utilize neighbor textures with contour shape information to fundamentally improve performance. First, we employ a graph-based network to extract the local contour and texture features of fragments. Then, for the pair-searching task, we adopt a linear transformer-based module to integrate these local features and use contrastive loss to encode the global features of each fragment. For the pair-matching task, we design a weighted fusion module to dynamically fuse extracted local contour and texture features, and formulate a similarity matrix for each pair of fragments to calculate the matching score and infer the adjacent segment of contours. To faithfully evaluate our proposed network, we collect a real dataset and generate a simulated image fragment dataset through an algorithm we designed that tears complete images into irregular fragments. The experimental results show that our proposed network achieves excellent pair-searching accuracy, reduces matching errors, and significantly reduces computational time. Source codes and data are available at here. + +Keywords: image fragment dataset $\cdot$ image fragment pair-searching and -matching + +# 1 Introduction + +Pair-searching and -matching of massive unorganized fragments is a critical problem in the research fields of computer vision and computer graphics, especially in the field of cultural relic restoration [1,15,17,23]. Traditional methods recover geometry transformation between fragments by calculating the similarity of contour segments based on geometric invariants such as curvature and manually + +![](images/df094cbf5cd95bc8f5564e41d54071b551ce95439722ac23f906d28fce579c9f.jpg) +Fig. 1: Our pair-searching and -matching task. Given a large number of mixed image fragments, our proposed network can search and match pairs of fragments, which is important for restoration problems. We also designed an algorithm to generate a dataset of image fragments by tearing a set of complete images. + +designed descriptors [18, 33]. It has a common core with the registration problem [3], which is to establish the correspondences of local features between related image pairs of the same scene. However, fragment assembly is more challenging because they do have overlapping regions. Thus, existing rule-based methods are difficult to generalize to various complex situations. Furthermore, they are time-consuming since the computational complexity of finding matching pairs from a set of fragments is typically $O(N^2)$ in global search, and in practice, only a few pairs are hidden in a large number of mixed fragments. + +Solving jigsaw puzzles [14, 37, 42] is also a similar task and learning-based approaches have been applied [2, 36]. However, two differences make these algorithms not generalizable. First, puzzle pieces usually have regular shapes, so they are best matched based on content information rather than contour features. Second, this problem usually studies matching all fragments into a complete block. Their input fragments do not come from mixed groups, and the matching results are dense and not lost. Therefore, they often use neighbor relationships to build graphs to constrain the search for the correct alignment of fragments. Once a fragment is lost, their closed-loop strategy fails. + +In this paper, we propose PairingNet, an end-to-end deep learning network, for searching and matching pairs from a large number of image fragments. To address this problem, we first explore the use of Graph Convolutional Networks (GCNs) to extract more robust contour and texture features for our tasks. Second, we design a feature fusion way to learn local matching features by calculating the similarity matrix of each fragment pair. Third, we build a module based on the linear transformer to learn the global searching features for fragments using contrastive learning. Finally, we collect a real dataset and generate a image fragment dataset to provide data-driven and validate the effectiveness of our proposed network. The main contributions of our work are as follows: + +- We propose a novel network to leverage the local contour and texture features for solving the pair-searching and -matching of image fragments. It provides an attempt to use advanced deep-learning techniques to solve this long-standing problem. +- We provide a real dataset and a generated dataset with a carefully designed algorithm to simulate real-world fragments. The algorithm generates diverse + +image fragments to facilitate the application of subsequent learning-based methods. + +- We conduct extensive experiments, ablation studies, and visual analysis to demonstrate the effectiveness of our proposed network. + +# 2 Related Work + +# 2.1 Fragments Assembly (Irregular Shape) + +Here, we focus on the pairwise matching algorithms of related representative works, although few of them also consider fragment relationships. Methods based on deep neural networks have been studied for object fragments; however, we could not find representative works for image fragments or even suitable datasets. + +Image fragments. Kong and Kimia [25] solved fragment matching using geometric features (partial curve matching). Leitao and Stolfi [17] described a specific multiscale matching method for the reassembly of 2D fragmented objects. Tsamoura and Pitas [43] presented a color-based method to automatically reassemble image fragments. Huang et al. [22] relied on salient curves detected inside the different image pieces to align gapped fragments. Subsequently, both shape and appearance information along the boundaries were utilized and extracted for each piece for matching [32,49]. Furthermore, Derech et al. [12] proposed to extrapolate each fragment to address fragment abrasion. For learning-based methods, traditional machine learning (SVM/Random Forest) was used as a classifier to process extracted local features [1,38]. Then Le and Li [27] built a deep neural network to detect the compatibility of pairwise stitching. + +Object fragments. Huang et al. [23] analyzed the geometry of the fracture surfaces by computing multi-scale surface characteristics. Yang et al. [46] matched stone tools based on contour points and their mean normals. Funkhouser et al. [15] learned a decision tree classifier to filter the candidates. Hong et al. [21] solved the problem of indistinguishable false matches by employing beam search to explore multiple registration possibilities. Ye et al. [47] presented an immersive system that can handle complex and ambiguous fragment shapes interactively with experts. In recent years, deep learning networks have been rapidly applied in 3D assembly. Huang et al. [48] proposed an assembly-oriented dynamic graph learning framework to estimate the pose of 3D parts. Chen et al. [8] proposed Neural Shape Mating (NSM) to tackle pairwise 3D geometric shape mating. Wu et al. [45] leveraged SE(3) equivariance for such shape pose disentanglement. + +Datasets. Existing image assembly works typically collect data in two ways: manual digitalization by photographing or scanning broken fragments, or using simple algorithms to tear complete images [7, 35, 38]. The lack of a large-scale dataset prevents the emergence of deep learning networks for this task. However, multiple datasets have been introduced in 3D assembly problems [24, 26, 40]. + +![](images/cadd093b8b7545134418df87cf2d950341591b2930aa4852b5775a42b4bf0fd9.jpg) +Fig. 2: Pipeline of our proposed PairingNet. (a) Based on patches, we first employ a binary encoding to describe contours and use ResGCN as the backbone to extract the local contour and texture features of image fragments. (b) We design self-gated fusion to fuse the extracted features in a weighted manner to calculate the similarity matrix for each pair of fragments, and (c) a linear transformer-based encoder for learning the global features of each fragment. (d) We use a two-step strategy to train our proposed network. (e) During inference, we calculate cosine similarity to find the adjacent fragment pairs and further process the matching similarity matrix to establish better correspondences between a fragment pair. + +# 2.2 Jigsaw Puzzles (Regular Shape) + +Solving jigsaw puzzles was first proposed in [14], and then regular pieces became a key research object in computer vision. Many existing methods [10, 16, 19, 37, 41, 42] focused on building relationships of pieces based on content information rather than contour-based pair-searching/-matching. And deep learning techniques [2, 36] was also used to solve this problem. + +# 3 Method + +# 3.1 Problem Statement + +Let us consider a fragment dataset $\mathbf{F} = \{f^{(i)}\}_{i=1}^{N}$ , which consists of fragments $f$ torn from a certain image dataset. We define the contour points $c_{j}^{(i)} \in \mathbf{C}^{(i)}$ for each $f^{(i)}$ , where $j = 1,2,\dots,M^{(i)}$ , $M^{(i)}$ is the contour length and $\mathbf{C}^{(i)}$ is an ordered set that contains all contour points of $f^{(i)}$ . + +Regarding any two neighbor fragments sampled from a pairset $\mathbf{P}$ , we denote the pair as $f^{(m)}$ and $f^{(n)}$ . The pair-searching task aims to retrieve the pairset $\mathbf{P}$ from a given fragment dataset $\mathbf{F}$ . Meanwhile, for the contours of the paired fragments, $\mathbf{C}^{(m)}$ and $\mathbf{C}^{(n)}$ , we can always find the overlapped part $\mathbf{C}^{(m\& n)}$ . The pair-matching task is designed to find the adjacent part from any given fragment pair. + +# 3.2 Network Architecture + +In the pair-matching task, a model has the capability to discern differences between two fragments, regardless of whether these discrepancies are on a large or minimal scale. Simultaneously, the features extracted by the model can be used to quantify the likelihood of adjacency between any given two fragments, assisting in the identification of matched pairs. However, the process of comparing every two fragments is not only laborious, but also inefficient. Therefore, utilizing the rich features the matching model extracts, we conceived a pair-searching module. This module can adeptly convert any fragment into a unique feature vector, thereby significantly accelerating the pair-searching process. + +# Feature Extraction + +Contour encoding. In this module, we aim to direct the model's attention to the unique features of contours. We design distinct contour encodings that also may be applied in other sketch-based fields. We tested various patch sizes and all these distinct encodings, finalizing a binary encoding and a patch size of $7 \times 7$ for our network. Subsequently, these contour patches are processed through a convolutional layer, followed by an average pooling layer and a fully connected network. We then adopt ResGCN [28] to amalgamate adjacent features, further extracting comprehensive global features. As such, for any provided fragment input denoted as $f^{(i)}$ with its corresponding contour $\mathbf{C}^{(i)}$ , this module yields a return of $f_{c}^{(i)} \in \mathbb{R}^{M^{(i)} \times 64}$ . + +Textural encoding. Our objective is to accentuate the distinguishing characteristics of textures in this module. We begin by cropping the image using a patch centered on each contour point. This is followed by filtering these patches through two consecutive convolutional layers, coupled with a single average pooling operation. Then, a ResGCN [28] is integrated to compile adjacent features, thereby enhancing the receptive domain of local textural characteristics. For any given fragment input represented as $f^{(i)}$ along with its associated contour $\mathbf{C}^{(i)}$ , this module produces an output of $f_{t}^{(i)} \in \mathbb{R}^{M^{(i)} \times 64}$ . + +# Pair-matching Module + +Feature Fusion. Regarding the decision-making process for pair-matching and -searching tasks, the dependence on contour and textural features varies. For instance, in scenarios where both fragments are solid colors, texture serves as the primary identifying factor suggesting adjacency. However, texture alone cannot identify the matched contours. In contrast, simple contours, like straight lines, require texture features for accurate matching position determination, as contour information alone falls short. + +Given this, we propose the incorporation of a Feature Fusion module [11] which would be capable of adaptively fine-tuning the weights assigned to these two features. The specific structure of this integral component is depicted in + +Figure 2 (b). First, we concatenate $f_{t}^{(i)}$ and $f_{c}^{(i)}$ , then the weight $\mathbf{w}^{(i)}$ for each contour point $c_{j}^{(i)} \in \mathbf{C}^{(i)}$ is computed by: + +$$ +\mathbf {w} ^ {(i)} = \sigma \left(\mathbf {W} _ {g} \left(f _ {t} ^ {(i)} \odot f _ {c} ^ {(i)}\right) + \mathbf {b} _ {g}\right) ^ {(i)} \tag {1} +$$ + +where $\sigma$ is sigmoid the activation function, $(\widehat{\mathbf{c}})$ means channel-wise concatenation, and $\mathbf{w} \in \mathbb{R}^{M^{(i)} \times 64}$ . Second, we fuse features of two modalities: + +$$ +\mathbf {F} _ {f} ^ {(i)} = \mathbf {w} ^ {(i)} \odot \mathbf {F} _ {t} ^ {(i)} + (1 - \mathbf {w} ^ {(i)}) \odot \mathbf {F} _ {c} ^ {(i)} \tag {2} +$$ + +where $\mathbf{F}_f^{(i)}\in \mathbb{R}^{M^{(i)}\times 64}$ , and $\odot$ is element-wise multiplication. With this module, we let the model adaptively decide the weights between textual and contour features, which is proven effective in later experiments. + +Similarity Matrix. As shown in Figure 2 (b-2), for every pair of fragments $f^{(m)}$ and $f^{(n)}$ from the pairwise matrix $\mathbf{S}^{(m\& n)} \in \mathbb{R}^{M^{(m)} \times M^{(n)}}$ [9,39]. + +We first calculate the similarity matrix $\tilde{\mathbf{S}}^{(m\& n)}$ by matrix multiplication: + +$$ +\tilde {\mathbf {S}} ^ {(m \& n)} = \frac {1}{\sqrt {D}} \mathbf {F} _ {f} ^ {(m)} \left(\mathbf {F} _ {f} ^ {(n)}\right) ^ {T} \tag {3} +$$ + +where $D$ is the dimension of the feature, in our model, $D$ is 64. Then, each entry in the matrix $\mathbf{S}^{(m\& n)}$ is normalized: + +$$ +\mathbf {S} ^ {(m \& n)} (i, j) = \frac {\exp \left(\tilde {\mathbf {S}} _ {i , j} ^ {(m \& n)}\right)}{\sum_ {i} \exp \left(\tilde {\mathbf {S}} _ {i , j} ^ {(m \& n)}\right)} \frac {\exp \left(\tilde {\mathbf {S}} _ {i , j} ^ {(m \& n)}\right)}{\sum_ {j} \exp \left(\tilde {\mathbf {S}} _ {i , j} ^ {(m \& n)}\right)} \tag {4} +$$ + +Pair-searching Module We introduce a modified feature fusion module to process features extracted by the backbone network and encode any fragment $f^{(i)}$ into a feature vector $v_{s}^{(i)} \in \mathbb{R}^{128}$ . It is based on the discovery that using fused features in the pair-matching task will lower the performance. + +To be specific, as shown in Figure 2 (c), every feature branch utilizes a Linear Transformer Encoder [44] to enhance the encoding of both contour and textural features. Subsequent to the modified fusion module, we add a fully-connected layer to encode the fused feature regarding per contour point. Then an average pooling layer is applied to map the feature matrix to a singular vector. Finally, another fully-connected layer is used to encode the feature. + +For every pair of fragment $f^{(i)}$ and $f^{(j)}$ from the fragment set $\mathbf{F}$ , this module allows them to be encoded as $v_{s}^{(i)}$ and $v_{s}^{(j)}$ . By calculating their inner product, we can efficiently quantify and contrast their similarity, thereby expediently identifying neighbor pairs. + +# 3.3 Loss Function + +We use two loss functions to train our network, and the training process is shown in Figure 2 (d). + +Pair-matching Loss Given the extreme imbalance in the ratio of positive to negative samples in the pair-matching task, we adopt the concept of focal loss [30] to concentrate on learning challenging correspondences. + +For any paired fragments $f^{(m)}$ and $f^{(n)}$ , entries in the ground truth similarity matrix $\mathbf{S}_{gt}^{(m\& n)}$ are 1 for matched contour points and 0 for unmatched ones. We calculate the matching loss $\mathcal{L}_{match}^{(m\& n)}$ as follows: + +$$ +\begin{array}{l} \mathcal {L} _ {\text {m a t c h}} ^ {(m \& n)} = - \sum_ {(0, 0)} (\beta_ {1} (\mathbf {1} - \mathbf {S} ^ {(m \& n)}) ^ {\gamma} \log \mathbf {S} ^ {(m \& n)} \mathbf {S} _ {g t} ^ {(m \& n)}) \tag {5} \\ + \beta_ {2} (\mathbf {S} ^ {(m \& n)}) ^ {\gamma} \log \left(\mathbf {1} - \mathbf {S} ^ {(m \& n)}\right) (\mathbf {1} - \mathbf {S} _ {g t} ^ {(m \& n)}) \\ \end{array} +$$ + +where $\beta_{1}$ and $\beta_{2}$ are the balancing weights and $\beta_{1} + \beta_{2} = 1$ , $\gamma$ is the decay factor. $\sum$ adds all the entries in the input matrix and all operations except $\sum$ are element-wise. By calculating and adding the pair-matching loss for all paired fragments $f^{(m)}, f^{(n)}$ from the pairwise $\mathbf{P}$ , we get $\mathcal{L}_{match}$ . + +Pair-searching Loss In the pair-searching task, we employ InfoNCE loss [34] as the contrastive loss function $\mathcal{L}_{search}$ . We utilize paired samples for constructing positive sample pairs, and unpaired samples for negative sample pairs. + +# 3.4 Inference + +Pair-matching In this section, we target to retrieve matched contour from the similarity matrix $\mathbf{S}^{(m\& n)}$ . + +We first apply a threshold $\epsilon = 0.006$ to filter out noisy entries. Entries with value lower than the threshold is set to 0. Because contour pointset $\mathbf{C}^{(m)}$ and $\mathbf{C}^{(n)}$ are ordered sets, for any ground truth similarity matrix $\mathbf{S}_{gt}^{(m\& n)}$ , if the start for contour points is the left-top corner, we can observe that the part with a value of 1 in the matrix must be arranged in the direction from upper left to lower right and connected into a line segment. Therefore, this inspires us to enhance the filtered similarity matrix. We design an eroding kernel $\mathbf{K}_e$ to filter out not dominant segments, which is adopted once. Then a dilating kernel $\mathbf{K}_d$ is applied once to connect close but split segments. The definitions of two kernels are as follows: + +$$ +\mathbf {K} _ {e} = \left[ \begin{array}{l l l} 0 & 0 & 1 \\ 0 & 0 & 0 \\ 1 & 0 & 0 \end{array} \right], \quad \mathbf {K} _ {d} = \left[ \begin{array}{l l l} 0 & 0 & 1 \\ 0 & 1 & 0 \\ 1 & 0 & 0 \end{array} \right]. \tag {6} +$$ + +Finally, we use RANSAC [13] to estimate the correspondences of two fragments and obtain the transformation matrix. + +Pair-searching For N fragments $f^{(i)} \in \mathbf{F}, i = 1,2,\dots,N$ , using the searching module, we can get a feature matrix $\mathbf{V}_s = [v_1,v_2,\ldots ,v_N]^T \in \mathbb{R}^{N\times 128}$ . Subsequently, we compute the cosine similarity between all global features to form the cosine similarity matrix. This allows us to obtain the similarity score for any pair of fragments. + +# 4 Dataset + +![](images/3f9ae64a0768296ac6a18f424d31230e30347881c843d7b2754bfeef048f9d2a.jpg) +Fig. 3: Dataset creation. (a) The process of generating our dataset. (b) Illustration of Algorithm 1 for tearing a complete image. + +Table 1: Dataset division. We divide the entire test set into three difficulty levels according to the contour overlapping proportion of fragment pairs. + +
TotalTrainValidationTest
FullHighMediumLow
Fragments819640988193279107116711090
Pairs14951365413723706631103604
+ +Generated Dataset. We design Algorithm 1 to tear a series of complete images and create a new image fragment dataset with data close to the real world. The process is also shown in Figure 3. To control the shape and area of the fragments, we first find the circumscribed circle of a fragment and take two points at an appropriate distance on the circle to determine the starting and ending points of the segmentation. Then, to simulate a more realistic fragment contour, we randomly divide the cutting line into several segments, each segment randomly choosing a straight line or an irregular curve. The irregular curve is synthesized by a series of Fourier orthogonal bases to make the shape of the fragments more diverse and similar to the edge of a real torn paper. We set a minimum threshold for the fragments, and the fragment is re-segmented if a very small fragment appears. + +To create our dataset, we collect 390 images of ten themes from Pexels [4] and empirically set the generation parameters shown in Algorithm 1 to obtain fragments. We calculate the distributions of our generated data, as shown in Figure 4. The distribution of fragment areas exhibits a long-tail pattern, which is consistent with the distribution of fragment areas observed when tearing large amounts of paper in real-life scenarios. The contour overlapping proportion of fragment pairs is closer to a uniform distribution over most ranges, indicating that our algorithm generates various types of fragment pairs. Then, we divide the dataset into training set, validation set, and test set in 5:1:4. We further divide the full test set into three difficulty levels based on the overlapping proportion of contour segments between fragment pairs. Low difficulty means that the pair of fragments has larger overlapping contour segments, and vice versa. The details are shown in Table 1. + +Algorithm 1 Generate fragments of an image +Input: List of fragments $L_{frag} = \{\}$ , a complete image $I$ Max iterations $t_{max} = 40$ min arc length ratio $\tau = 0.9$ max number of cutting points $N_{max} =$ 3, min distance from cutting point to edge of fragment $D_{min} = 100$ , the number of Fourier orthogonal basis $n = 20$ scaling ratios $s_1 = 0.25$ $s_2 = 0.0067$ $s_3 = 1.5$ $s_4 = 0.3$ , probability of curve type $\rho = 0.5$ , min fragment size $h_{min} = 150$ $w_{min} = 150$ +Output: Collection of fragments $L_{frag}$ +1: put I in $L_{frag}$ +2: while t < tmax do +3: f<-randomly_pick(Lfrag), R<-bounding_rectangle(f) +4: W,H<-width_and_height(R), C<-circumcircle(R) +5: ldia<-diameter(C) +6: +7: // generate the cutting points +8: repeat +9: pa, pb<-randomly_pick(C) +10: larc<-smaller_ar_c_length(pa,pb) +11: until larc>τ·half_perimeter(C) +12: pstart, pend<-intersection(f,pa,pb) +13: m<-randomly_pick{0,...,Nmax} +14: repeat +15: [p1, ..., pm]<-random_select_insidepoints(f,m) +16: d<-min_distance_to_edge(f,p1, ..., pm) +17: until d > Dmin +18: put pstart, pend in [p1, ..., pm] +19: +20: // generate the cutting segments +21: for each pi∈[pstart, p1, ..., pm, pend] do +22: irregular line $l_{irr} = \{\}$ , phase $\phi \sim \mathcal{U}(-\pi ,\pi)$ amplitude A $\sim \mathcal{N}(s_1H,s_2H)$ period T $\sim \mathcal{N}(s_3W,s_4W)$ +23: for each x-coordinate between (pi,pi+1) do +24: y<∑i=0nA sin(2πix+φ)+H/2 +25: put {x,y} in lirr +26: end for +27: lstr← straight_line(pi,pi+1) +28: lcut← select_and_connect(lirr,lstr,ρ,lcut) +29: end for +30: f1, f2<-cut_structure(f,lcut) +31: if size(f1)>hmin,wmin and size(f2)>hmin,wmin then +32: put f1, f2 in Lfrag +33: end if +34: end while + +Real Dataset. We also collect a series of real image fragments to construct a real dataset. We printed 34 complete images (not included in our generated dataset), manually cut them into fragments, and then scanned them. Then, we apply some post-processing steps to the scanned image fragments to obtain the contour of the fragments and the annotation of the matching points. More details about real dataset can be found in the supplementary materials. + +# 5 Experiments + +# 5.1 Comparison Methods + +Finding comparison methods is not easy. Because rule-based methods have been around for quite some time, lack source code, and tend to exhibit lower performance compared to advanced deep learning methods. However, most existing deep learning methods require extensive modifications if applied to our task. Therefore, we employ Jigsawnet [27] (without loop consistency) and a rule-based method [49] (Step 1) as our comparison methods. + +![](images/32bc207a37df5bb428d522db3af308c10cc8263dc6e5e54facc03970e7d2f24f.jpg) +Fig. 4: Dataset statistics. (a) Image area distribution of the generated fragments. (b) Proportional distribution of overlapping contour segments of the generated fragments. (c) Image area distribution of the scanned real fragments. (b) Proportional distribution of overlapping contour segments of the scanned real fragments. + +![](images/34f244442e704b1028672b5542af0b539a890a23fef6ee6f71c96849fbee82c2.jpg) + +![](images/538965f920f9182e8f4cb9f7d734e2ab0305744a7a550aab2da822fe1cc18cf9.jpg) +Fig. 5: Violin plots of the matching results of our method and the comparison method: (a) RE, NTE, and HD on the test set. (b) HD on the test sets of different difficulties. + +![](images/a821ca8844adee84036eda4afb9a636a3a266e58cc67cc9591536e033fff6f64.jpg) + +# 5.2 Implementation Details + +Our PairingNet was implemented by PyTorch, and conducted on a server with 4 RTX A40 GPUs and Intel® Xeon® Gold 5220 CPUs (72 cores). We used the Adam optimizer with an initial learning rate of 0.001, which was adjusted using a cosine annealing strategy. For training pair-matching module, the batch size was set to 20, and for the pair-searching module, the batch size was set to 175. + +For comparison methods, given that the performance bottleneck of them resides in the CPU, utilizing GPUs for computational acceleration is deemed inappropriate. Thus we use two more powerful CPUs computing platforms to run the comparison method. We used the official sourcecode (implemented by TensorFlow) of Jigsawnet and conducted experiments on a server with 4 Quadro RTX 6000 GPUs and Intel® Xeon® Gold 5220 CPUs (144 cores). We reproduced the rule-based method by Python and conducted experiments on a node of a Supercomputing platform with Intel® Xeon® Gold 6258R CPUs (128 cores). + +For evaluation metrics, we employ Recall@k and NDCG@k [5] for the pair-searching task, and Registration recall (RR), Hausdorff distance (HD), Radians Error (RE) and Normalized Translation Error (NTE) for the pair-matching task. We normalize Translation Error (TE) by the sum of the areas of fragment pairs to obtain the NTE, considering the large difference in area between fragments. + +# 5.3 Results + +Performance on Generated Dataset. As shown in Table 2, our network outperforms the comparison methods on both the full test set and on every set of difficulty levels. Jigsawnet shows a certain superiority in pair-searching on the high-difficulty set, probably because it treats the problem as a simple binary + +Table 2: Comparison Results of various metrics. The best results in our full dataset are highlighted in bold font, and the best results for each difficulty are marked in different colors. + +
MethodDifficultyRecall@5Recall@10Recall@20NDCG@5NDCG@10NDCG@20RR ↑HD ↓RE ↓NTE ↓
Rule-based [49]High0.2610.3540.4430.1850.2180.2410.107212.0031.28937.100×10-4
Medium0.3290.4270.5290.2420.2750.3020.153205.4631.24933.742×10-4
Low0.3580.4640.5550.2610.2960.320.188210.2741.23132.692×10-4
All0.3170.4160.5110.2480.2840.3110.149208.5191.25634.414×10-4
Jigsawnet [27]High0.3050.3940.4780.2310.2610.2840.117226.0261.30637.011×10-4
Medium0.3830.4780.5650.3090.340.3640.213206.0401.15531.467×10-4
Low0.4590.5480.6310.3680.3970.4190.241202.9421.18630.500×10-4
All0.380.4720.5570.3260.3590.3830.194210.8421.20532.772×10-4
OursHigh0.2760.4130.5350.1890.2340.2680.60692.2290.80718.335×10-4
Medium0.5010.6190.7320.3820.4230.4540.87935.6170.2365.685×10-4
Low0.7170.8050.8790.5810.610.6290.96112.3360.091.936×10-4
All0.4930.6080.7140.4170.4580.4870.83543.1160.3527.735×10-4
+ +![](images/7993877d11836b7c58a1b541d38db1fb231752dde2eb26188128b3850ca75c4a.jpg) +Fig. 6: Examples of pair-searching results. Our network is able to accurately identify the corresponding target fragment (highlighted with a green box). + +classification task; however, their pair-matching results are very unsatisfactory. It should be noted that we did not evaluate Recall@1 and NDCG@1 here due to the many-to-many pairings in our dataset, making it unreasonable to search only for the fragment with the highest similarity of searching. We can also observe that the performance of our proposed network shows reasonable variation with the difficulty of the data. This shows that our proposed network learns useful information and has more potential in complex situations. However, Jigsawnet and the rule-based method show relatively identical performance on test sets of different difficulty levels. Furthermore, we also plot the distributions of HD, RE, and NTE for each method for a more intuitive comparison, as shown in Figure 5. + +Performance on Real Dataset. We also tested our method on the real dataset. In this experiment, both our method and the comparison method were trained on the generated dataset without retraining or fine-tuning. As shown in Table 3 and Figure 8, PairingNet obtained satisfactory results for both pair-searching and -matching tasks. These results confirmed that Algorithm 1 simulates real data well and our method has good generalization to real data. + +![](images/8d02581339b5ca491fca5e0f094c297ff8eceedbbce1bdab5c1f9989cd7af6e8.jpg) +Fig. 7: Examples of pair-matching results for three difficulty levels. Our proposed network achieves satisfactory matching results. However, for low-difficulty cases, Jigsawnet and the Rule-based method find approximate matching positions, while the matchings are not accurate enough. For medium- and high-difficulty cases, both comparison methods are difficult to obtain correct matching results. + +Table 3: Comparison Results of various metrics on our real dataset. The best results are highlighted in bold font. + +
MethodRecall@5Recall@10Recall@20NDCG@5NDCG@10NDCG@20RR↑HD↓RE↓NTE↓
Rule-based [49]0.1380.2470.3660.0790.1140.1440.1287.6981.61430.746×10-4
Jigsawnet [27]0.2370.3260.4250.1850.2130.2380.1537.5701.54529.061×10-4
Ours0.4150.5640.6780.3210.3690.3990.7524.1820.4809.075×10-4
+ +Visualization. We show the top five search results of different methods for the examples of fragment pairs in Figure 6. Given a source fragment, our model can accurately identify the corresponding target fragment (marked with a green box). Furthermore, in situations where a source fragment can be paired with multiple target fragments, our method identifies all matching fragments within the top five search results. We also show examples of matching results for three difficulty levels in Figure 7. Our method produces matching results that are consistent with the Ground Truth (GT), while the comparison methods yield numerous incorrect results. + +We also visualize the normalized self-gated weights to analyze their changes under different situations. As shown in Figure 9 (a), we map the normalized weights to the edges of the fragments to show the influence of the texture/contour features. We classify the local parts of fragments into four types based on the information contained in texture and contour: those with rich texture and contour information (first quadrant), those with rich texture information (second quadrant), those with little texture and contour information (third quadrant), and those with rich contour information but with little texture information (fourth quadrant). The red color indicates that the weights of the texture features are relatively large, while the green color indicates that the contour feature weights are relatively large. The results show that our proposed network learned varying weights of the features for fragments with different characteristics. + +Inference Time. We test the inference time of different methods on the pairwise searching task on our generated test set. Our network retrieves all fragment pairs in 73 seconds, while the rule-based method takes 223 hours, and JigsawNet takes 10 hours. + +# 5.4 Ablation Study + +Network Structures. We explore the performance of our network under different structures, as shown in Table 4. First, we test three feature fusion ways. + +![](images/0886c5dd15063f233ed70d809b3cd7bfa9cdb31d25e8cdc060b4ee3ba9acc216.jpg) +Fig. 8: Examples of pair-matching results of PairingNet on real dataset. Without retraining, our method can also effectively match the scanned real fragments. + +![](images/e8bcd3996eef8dc74cd042054d0a5545b5cdfb3d6425de7b5b14d4af8a59fea9.jpg) +(a) + +![](images/795386562a2688255b455a2b03f0895e0874d88a3c5d56daabfab22a0ed06e38.jpg) +(b) +Fig.9: Visual analysis of Texture/Contour. (a) We visualize the weights of texture and contour features in our network, with colors closer to red representing more use of texture features and closer to green representing more use of contour features. We can see that on edges with rich texture or close to straight lines, the weights of the extracted texture features are greater, while on edges with single texture, the weights of the contour features are greater. (b) Examples of Textur/Contour ablation study. When the local contour shapes are similar, our network has difficulty matching them correctly using only contour features, and vice versa. However, these cases can be solved when the two features are fused. + +In early fusion, instead of using two separate GCNs for feature extraction, we concatenate the two modalities and use only one GCN to learn the features. We also test using a simple concatenation to replace the self-gated fusion module. The results show that the performance of both pair-searching and -matching decreases. Additionally, we explore the impact of different numbers of layers on our network. + +Texture/Contour. To study the impact of contour and texture features on the performance of our network, we conducted experiments using only one modality (either contour or texture) and compared the final results. As shown in Table 4, the comparison results indicate that both modalities perform slightly worse when used independently. When both modalities are used in conjunction, our network achieves the best performance. We also show examples of failed pair-matching when the two modalities are used individually in Figure 9 (b). + +Table 4: Results of ablation studies. Comparison of the impact of different feature fusion ways, feature modalities, and network layers employed in our proposed network on the performance of pair-searching and pair-matching tasks. + +
Recall@5Recall@10NDCG@5NDCG@10RR
Fusion WayEarly fusion0.4480.5590.3760.4130.658
Concat0.4890.6050.4180.4580.824
Weighted0.4930.6080.4170.4580.835
ResGCN10----0.680
14----0.835
18----0.759
Transformer Encoder30.4880.6000.4100.450-
40.4880.6060.4110.452-
50.4930.6080.4170.458-
Feature ModalityContour-only0.0160.0250.0130.0160.584
Texture-only0.4790.5970.3940.4360.657
Contour + Texture0.4930.6080.4170.4580.835
+ +Besides, as shown in Figure 9 (a), it can be seen numerically that our network tends to assign more weight to the texture features. This is reasonable because textures usually contain richer information than the contours of image fragments, enabling the network to learn more discriminative features. + +Others. We report the results of ablation studies on patch size, contour encoding, and hyperparameters in our supplementary material. + +# 6 Conclusions + +In this paper, to tackle the image fragment pair-searching and -matching tasks, we propose a novel network, PairingNet, to deduce the adjacent segments of fragments and encode the global features of each fragment. In PairingNet, ResGCN is employed as the backbone to extract the local contour and texture features of fragments. For pair-searching, a linear transformer-based module processes these local features leveraging the contrast of paired fragments and unpaired fragments. For pair-matching, we design a weighted fusion module to dynamically fuse the extracted local contour and texture features. Furthermore, to provide a foundation for learning-based methods, we design a cutting algorithm to generate a dataset by simulating real image fragments. Comprehensive experiments are conducted on our generated dataset and a real dataset, the results of which demonstrate the effectiveness of our PairingNet compared to existing methods. In future work, we will explore how to extend our work to other general technique, e.g. puzzles inspired MAE [20], JiGen [6] for self-supervised learning and domain generalization. The fragment generation algorithm that we designed can also create more diverse samples for image inpainting tasks [29,31]. + +Limitations While our research has made significant strides, fragment restoration remains a challenging task. Our dataset provides superior edge matching features. The task of fragment restoration becomes particularly arduous when the edges of the fragments are smooth and the texture is monotonous (for instance, in the case of shredded white paper), or when there is a substantial gap in the damaged edges. + +Acknowledgements This work was supported by the Young Scientists Fund of the National Natural Science Foundation of China (Grant No.62206106). We also feel grateful to Yifan Jin for his exploration in the preliminary work. + +# References + +1. Abitbol, R., Shimshoni, I., Ben-Dov, J.: Machine learning based assembly of fragments of ancient papyrus. Journal on Computing and Cultural Heritage (JOCCH) 14(3), 1-21 (2021) +2. Bridger, D., Danon, D., Tal, A.: Solving jigsaw puzzles with eroded boundaries. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 3526-3535 (2020) +3. Brown, L.G.: A survey of image registration techniques. ACM computing surveys (CSUR) 24(4), 325-376 (1992) +4. Bruno Joseph, I.J., Frese, D.: Pexels. https://www.pexels.com/ (2023) +5. Busa-Fekete, R., Szarvas, G., Elteto, T., Kégl, B.: An apple-to-apple comparison of learning-to-rank algorithms in terms of normalized discounted cumulative gain. In: ECAI 2012-20th European Conference on Artificial Intelligence: Preference Learning: Problems and Applications in AI Workshop. vol. 242. IOS Press (2012) +6. Carlucci, F.M., D'Innocente, A., Bucci, S., Caputo, B., Tommasi, T.: Domain generalization by solving jigsaw puzzles. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 2229-2238 (2019) +7. Chen, J., Tian, M., Qi, X., Wang, W., Liu, Y.: A solution to reconstruct cross-cut shredded text documents based on constrained seed k-means algorithm and ant colony algorithm. Expert Systems with Applications 127(AUG.), 35-46 (2019) +8. Chen, Y.C., Li, H., Turpin, D., Jacobson, A., Garg, A.: Neural shape mating: Self-supervised object assembly with adversarial shape priors. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 12724-12733 (2022) +9. Cheng, X., Lin, H., Wu, X., Yang, F., Shen, D.: Improving video-text retrieval by multi-stream corpus alignment and dual softmax loss. arXiv preprint arXiv:2109.04290 (2021) +0. Cho, T.S., Avidan, S., Freeman, W.T.: A probabilistic image jigsaw puzzle solver. In: 2010 IEEE Computer society conference on computer vision and pattern recognition. pp. 183-190. IEEE (2010) +1. Dai, Y., Gieseke, F., Oehmcke, S., Wu, Y., Barnard, K.: Attentional feature fusion. In: Proceedings of the IEEE/CVF winter conference on applications of computer vision. pp. 3560-3569 (2021) +2. Derech, N., Tal, A., Shimshoni, I.: Solving archaeological puzzles. Pattern Recognition 119, 108065 (2021) +3. Fischler, M.A., Bolles, R.C.: Random sample consensus: a paradigm for model fitting with applications to image analysis and automated cartography. Communications of the ACM 24(6), 381-395 (1981) +4. Freeman, H., Gardner, L.: Apictorial jigsaw puzzles: The computer solution of a problem in pattern recognition. IEEE Transactions on Electronic Computers (2), 118-127 (1964) +5. Funkhouser, T., Shin, H., Toler-Franklin, C., Castañeda, A.G., Brown, B., Dobkin, D., Rusinkiewicz, S., Weyrich, T.: Learning how to match fresco fragments. Journal on Computing and Cultural Heritage (JOCCH) 4(2), 1-13 (2011) +6. Gallagher, A.C.: Jigsaw puzzles with pieces of unknown orientation. In: 2012 IEEE Conference on computer vision and pattern recognition. pp. 382-389. IEEE (2012) +7. da Gama Leitao, H.C., Stolfi, J.: A multiscale method for the reassembly of two-dimensional fragmented objects. IEEE Transactions on Pattern Analysis and Machine Intelligence 24(9), 1239-1251 (2002) + +18. Guo, Y., Bennamoun, M., Sohel, F., Lu, M., Wan, J., Kwok, N.M.: A comprehensive performance evaluation of 3d local feature descriptors. International Journal of Computer Vision 116, 66-89 (2016) +19. Gur, S., Ben-Shahar, O.: From square pieces to brick walls: The next challenge in solving jigsaw puzzles. In: Proceedings of the IEEE International Conference on Computer Vision. pp. 4029-4037 (2017) +20. He, K., Chen, X., Xie, S., Li, Y., Dollar, P., Girshick, R.: Masked autoencoders are scalable vision learners. In: Proceedings of the IEEE/CVF conference on computer vision and pattern recognition. pp. 16000-16009 (2022) +21. Hong, J.H., Yoo, S.J., Zeeshan, M.A., Kim, Y.M., Kim, J.: Structure-from-sherds: Incremental 3d reassembly of axially symmetric pots from unordered and mixed fragment collections. In: Proceedings of the IEEE/CVF International Conference on Computer Vision. pp. 5443-5451 (2021) +22. Huang, H., Yin, K., Gong, M., Lischinski, D., Cohen-Or, D., Ascher, U.M., Chen, B.: "mind the gap": tele-registration for structure-driven image completion. ACM Trans. Graph. 32(6), 174-1 (2013) +23. Huang, Q.X., Flóry, S., Gelfand, N., Hofer, M., Pottmann, H.: Reassembling fractured objects by geometric matching. In: ACM SIGGRAPH 2006 Papers, pp. 569-578 (2006) +24. Jones, B., Hildreth, D., Chen, D., Baran, I., Kim, V.G., Schulz, A.: Automate: A dataset and learning approach for automatic mating of cad assemblies. ACM Transactions on Graphics (TOG) 40(6), 1-18 (2021) +25. Kong, W., Kimia, B.B.: On solving 2d and 3d puzzles using curve matching. In: Proceedings of the 2001 IEEE Computer Society Conference on Computer Vision and Pattern Recognition. CVPR 2001. vol. 2, pp. II-II. IEEE (2001) +26. Lamb, N., Palmer, C., Molloy, B., Banerjee, S., Banerjee, N.K.: Fantastic breaks: A dataset of paired 3d scans of real-world broken objects and their complete counterparts. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 4681-4691 (2023) +27. Le, C., Li, X.: Jigsawnet: Shredded image reassembly using convolutional neural network and loop-based composition. IEEE Transactions on Image Processing 28(8), 4000-4015 (2019) +28. Li, G., Muller, M., Thabet, A., Ghanem, B.: Deep GCs: Can GCs go as deep as cnns? In: Proceedings of the IEEE/CVF international conference on computer vision. pp. 9267-9276 (2019) +29. Li, W., Lin, Z., Zhou, K., Qi, L., Wang, Y., Jia, J.: Mat: Mask-aware transformer for large hole image inpainting. In: Proceedings of the IEEE/CVF conference on computer vision and pattern recognition. pp. 10758-10768 (2022) +30. Li, Y., Harada, T.: Lepard: Learning partial point cloud matching in rigid and deformable scenes. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 5554-5564 (2022) +31. Liu, G., Reda, F.A., Shih, K.J., Wang, T.C., Tao, A., Catanzaro, B.: Image inpainting for irregular holes using partial convolutions. In: Proceedings of the European conference on computer vision (ECCV). pp. 85-100 (2018) +32. Liu, H., Cao, S., Yan, S.: Automated assembly of shredded pieces from multiple photos. IEEE transactions on multimedia 13(5), 1154-1162 (2011) +33. Mikolajczyk, K., Schmid, C.: A performance evaluation of local descriptors. IEEE transactions on pattern analysis and machine intelligence 27(10), 1615-1630 (2005) +34. Oord, A.v.d., Li, Y., Vinyals, O.: Representation learning with contrastive predictive coding. arXiv preprint arXiv:1807.03748 (2018) + +35. Paixao, T.M., Boeres, M., Freitas, C., Oliveira-Santos, T.: Exploring character shapes for unsupervised reconstruction of strip-shredded text documents. IEEE Transactions on Information Forensics and Security pp. 1-1 (2018) +36. Paumard, M.M., Picard, D., Tabia, H.: Image reassembly combining deep learning and shortest path problem. In: Proceedings of the European conference on computer vision (ECCV). pp. 153-167 (2018) +37. Pomeranz, D., Shemesh, M., Ben-Shahar, O.: A fully automated greedy square jigsaw puzzle solver. In: CVPR 2011. pp. 9-16. IEEE (2011) +38. Richter, F., Ries, C.X., Cebron, N., Lienhart, R.: Learning to reassemble shredded documents. IEEE Transactions on Multimedia 15(3), 582-593 (2013) +39. Rocco, I., Cimpoi, M., Arandjelovic, R., Torii, A., Pajdla, T., Sivic, J.: Neighbourhood consensus networks. Advances in neural information processing systems 31 (2018) +40. Sellan, S., Chen, Y.C., Wu, Z., Garg, A., Jacobson, A.: Breaking bad: A dataset for geometric fracture and reassembly. Advances in Neural Information Processing Systems 35, 38885-38898 (2022) +41. Son, K., Hays, J., Cooper, D.B.: Solving square jigsaw puzzles with loop constraints. In: European conference on computer vision. pp. 32-46. Springer (2014) +42. Son, K., Hays, J., Cooper, D.B.: Solving square jigsaw puzzle by hierarchical loop constraints. IEEE transactions on pattern analysis and machine intelligence 41(9), 2222-2235 (2018) +43. Tsamoura, E., Pitas, I.: Automatic color based reassembly of fragmented images and paintings. IEEE Transactions on Image Processing 19(3), 680-690 (2009) +44. Wang, S., Li, B.Z., Khabsa, M., Fang, H., Ma, H.: Linformer: Self-attention with linear complexity. arXiv preprint arXiv:2006.04768 (2020) +45. Wu, R., Tie, C., Du, Y., Zhao, Y., Dong, H.: Leveraging se (3) equivariance for learning 3d geometric shape assembly. In: Proceedings of the IEEE/CVF International Conference on Computer Vision. pp. 14311-14320 (2023) +46. Yang, X., Matsuyama, K., Konno, K.: Pairwise matching of stone tools based on flake-surface contour points and normals. In: Proceedings of the Eurographics Workshop on Graphics and Cultural Heritage. pp. 125-129 (2017) +47. Ye, S., Chen, Z., Chu, X., Li, K., Luo, J., Li, Y., Geng, G., Wu, Y.: Puzzlefixer: A visual reassembly system for immersive fragments restoration. IEEE transactions on visualization and computer graphics 29(1), 429-439 (2022) +48. Zhan, G., Fan, Q., Mo, K., Shao, L., Chen, B., Guibas, L.J., Dong, H., et al.: Generative 3d part assembly via dynamic graph learning. Advances in Neural Information Processing Systems 33, 6315-6326 (2020) +49. Zhang, K., Li, X.: A graph-based optimization algorithm for fragmented image reassembly. Graphical Models 76(5), 484-495 (2014) \ No newline at end of file diff --git a/2024/PairingNet_ A Learning-based Pair-searching and -matching Network for Image Fragments/images.zip b/2024/PairingNet_ A Learning-based Pair-searching and -matching Network for Image Fragments/images.zip new file mode 100644 index 0000000000000000000000000000000000000000..74d11d43f4081d29f5032091171bbb91b93cdd54 --- /dev/null +++ b/2024/PairingNet_ A Learning-based Pair-searching and -matching Network for Image Fragments/images.zip @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:4a7182d0c7e3dabe5788e979aca4567cc56746db5176da6fc263ad01926a7028 +size 548335 diff --git a/2024/PairingNet_ A Learning-based Pair-searching and -matching Network for Image Fragments/layout.json b/2024/PairingNet_ A Learning-based Pair-searching and -matching Network for Image Fragments/layout.json new file mode 100644 index 0000000000000000000000000000000000000000..07c28b33c84d1e7a6c4cf02698f0affad04dae08 --- /dev/null +++ b/2024/PairingNet_ A Learning-based Pair-searching and -matching Network for Image Fragments/layout.json @@ -0,0 +1,9367 @@ +{ + "pdf_info": [ + { + "para_blocks": [ + { + "bbox": [ + 135, + 111, + 479, + 148 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 135, + 111, + 479, + 148 + ], + "spans": [ + { + "bbox": [ + 135, + 111, + 479, + 148 + ], + "type": "text", + "content": "PairingNet: A Learning-based Pair-searching and -matching Network for Image Fragments" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 140, + 167, + 473, + 194 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 140, + 167, + 473, + 194 + ], + "spans": [ + { + "bbox": [ + 140, + 167, + 473, + 194 + ], + "type": "text", + "content": "Rixin Zhou" + }, + { + "bbox": [ + 140, + 167, + 473, + 194 + ], + "type": "inline_equation", + "content": "^{1}" + }, + { + "bbox": [ + 140, + 167, + 473, + 194 + ], + "type": "text", + "content": ", Ding Xia" + }, + { + "bbox": [ + 140, + 167, + 473, + 194 + ], + "type": "inline_equation", + "content": "^{2}" + }, + { + "bbox": [ + 140, + 167, + 473, + 194 + ], + "type": "text", + "content": ", Yi Zhang" + }, + { + "bbox": [ + 140, + 167, + 473, + 194 + ], + "type": "inline_equation", + "content": "^{3}" + }, + { + "bbox": [ + 140, + 167, + 473, + 194 + ], + "type": "text", + "content": ", Honglin Pang" + }, + { + "bbox": [ + 140, + 167, + 473, + 194 + ], + "type": "inline_equation", + "content": "^{1}" + }, + { + "bbox": [ + 140, + 167, + 473, + 194 + ], + "type": "text", + "content": ", Xi Yang" + }, + { + "bbox": [ + 140, + 167, + 473, + 194 + ], + "type": "inline_equation", + "content": "^{1,*,\\dagger, \\ddagger}" + }, + { + "bbox": [ + 140, + 167, + 473, + 194 + ], + "type": "text", + "content": ", and Chuntao Li" + }, + { + "bbox": [ + 140, + 167, + 473, + 194 + ], + "type": "inline_equation", + "content": "^{4,*, \\ddagger}" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 204, + 201, + 408, + 213 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 204, + 201, + 408, + 213 + ], + "spans": [ + { + "bbox": [ + 204, + 201, + 408, + 213 + ], + "type": "inline_equation", + "content": "^{1}" + }, + { + "bbox": [ + 204, + 201, + 408, + 213 + ], + "type": "text", + "content": " School of Artificial Intelligence, Jilin University" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 252, + 213, + 361, + 224 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 252, + 213, + 361, + 224 + ], + "spans": [ + { + "bbox": [ + 252, + 213, + 361, + 224 + ], + "type": "text", + "content": "2 The University of Tokyo" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 144, + 224, + 469, + 236 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 144, + 224, + 469, + 236 + ], + "spans": [ + { + "bbox": [ + 144, + 224, + 469, + 236 + ], + "type": "inline_equation", + "content": "^{3}" + }, + { + "bbox": [ + 144, + 224, + 469, + 236 + ], + "type": "text", + "content": " National University of Defense Technology, College of Electronic Engineering" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 223, + 235, + 391, + 246 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 223, + 235, + 391, + 246 + ], + "spans": [ + { + "bbox": [ + 223, + 235, + 391, + 246 + ], + "type": "inline_equation", + "content": "^{4}" + }, + { + "bbox": [ + 223, + 235, + 391, + 246 + ], + "type": "text", + "content": " School of Archaeology, Jilin University" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 199, + 247, + 414, + 257 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 199, + 247, + 414, + 257 + ], + "spans": [ + { + "bbox": [ + 199, + 247, + 414, + 257 + ], + "type": "text", + "content": "{zhourx22}@mails.jlu.edu.cn {1ct33}@jlu.edu.cn" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 151, + 258, + 462, + 269 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 151, + 258, + 462, + 269 + ], + "spans": [ + { + "bbox": [ + 151, + 258, + 462, + 269 + ], + "type": "text", + "content": "{dingxia1995, panghlwork, earthyangxi}@gmail.com {2356711993}@qq.com" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 159, + 298, + 455, + 529 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 159, + 298, + 455, + 529 + ], + "spans": [ + { + "bbox": [ + 159, + 298, + 455, + 529 + ], + "type": "text", + "content": "Abstract. In this paper, we propose a learning-based image fragment pair-searching and -matching approach to solve the challenging restoration problem. Existing works use rule-based methods to match similar contour shapes or textures, which are always difficult to tune hyperparameters for extensive data and computationally time-consuming. Therefore, we propose a neural network that can effectively utilize neighbor textures with contour shape information to fundamentally improve performance. First, we employ a graph-based network to extract the local contour and texture features of fragments. Then, for the pair-searching task, we adopt a linear transformer-based module to integrate these local features and use contrastive loss to encode the global features of each fragment. For the pair-matching task, we design a weighted fusion module to dynamically fuse extracted local contour and texture features, and formulate a similarity matrix for each pair of fragments to calculate the matching score and infer the adjacent segment of contours. To faithfully evaluate our proposed network, we collect a real dataset and generate a simulated image fragment dataset through an algorithm we designed that tears complete images into irregular fragments. The experimental results show that our proposed network achieves excellent pair-searching accuracy, reduces matching errors, and significantly reduces computational time. Source codes and data are available at here." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 159, + 539, + 455, + 562 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 159, + 539, + 455, + 562 + ], + "spans": [ + { + "bbox": [ + 159, + 539, + 455, + 562 + ], + "type": "text", + "content": "Keywords: image fragment dataset " + }, + { + "bbox": [ + 159, + 539, + 455, + 562 + ], + "type": "inline_equation", + "content": "\\cdot" + }, + { + "bbox": [ + 159, + 539, + 455, + 562 + ], + "type": "text", + "content": " image fragment pair-searching and -matching" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 132, + 586, + 230, + 598 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 586, + 230, + 598 + ], + "spans": [ + { + "bbox": [ + 132, + 586, + 230, + 598 + ], + "type": "text", + "content": "1 Introduction" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 130, + 605, + 482, + 666 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 605, + 482, + 666 + ], + "spans": [ + { + "bbox": [ + 130, + 605, + 482, + 666 + ], + "type": "text", + "content": "Pair-searching and -matching of massive unorganized fragments is a critical problem in the research fields of computer vision and computer graphics, especially in the field of cultural relic restoration [1,15,17,23]. Traditional methods recover geometry transformation between fragments by calculating the similarity of contour segments based on geometric invariants such as curvature and manually" + } + ] + } + ], + "index": 11 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 132, + 686, + 207, + 693 + ], + "type": "page_footnote", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 686, + 207, + 693 + ], + "spans": [ + { + "bbox": [ + 132, + 686, + 207, + 693 + ], + "type": "text", + "content": "*Corresponding authors" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 132, + 693, + 418, + 700 + ], + "type": "page_footnote", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 693, + 418, + 700 + ], + "spans": [ + { + "bbox": [ + 132, + 693, + 418, + 700 + ], + "type": "inline_equation", + "content": "\\dagger" + }, + { + "bbox": [ + 132, + 693, + 418, + 700 + ], + "type": "text", + "content": " Engineering Research Center of Knowledge-Driven Human-Machine Intelligence, MoE, China" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 132, + 700, + 438, + 706 + ], + "type": "page_footnote", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 700, + 438, + 706 + ], + "spans": [ + { + "bbox": [ + 132, + 700, + 438, + 706 + ], + "type": "inline_equation", + "content": "\\ddagger" + }, + { + "bbox": [ + 132, + 700, + 438, + 706 + ], + "type": "text", + "content": " Key Laboratory of Ancient Chinese Script, Culture Relics and Artificial Intelligence, Jilin Province" + } + ] + } + ], + "index": 14 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 0 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 135, + 114, + 478, + 177 + ], + "blocks": [ + { + "bbox": [ + 135, + 114, + 478, + 177 + ], + "lines": [ + { + "bbox": [ + 135, + 114, + 478, + 177 + ], + "spans": [ + { + "bbox": [ + 135, + 114, + 478, + 177 + ], + "type": "image", + "image_path": "df094cbf5cd95bc8f5564e41d54071b551ce95439722ac23f906d28fce579c9f.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 130, + 186, + 482, + 231 + ], + "lines": [ + { + "bbox": [ + 130, + 186, + 482, + 231 + ], + "spans": [ + { + "bbox": [ + 130, + 186, + 482, + 231 + ], + "type": "text", + "content": "Fig. 1: Our pair-searching and -matching task. Given a large number of mixed image fragments, our proposed network can search and match pairs of fragments, which is important for restoration problems. We also designed an algorithm to generate a dataset of image fragments by tearing a set of complete images." + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_caption" + } + ], + "index": 2 + }, + { + "bbox": [ + 130, + 257, + 479, + 353 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 257, + 479, + 353 + ], + "spans": [ + { + "bbox": [ + 130, + 257, + 479, + 353 + ], + "type": "text", + "content": "designed descriptors [18, 33]. It has a common core with the registration problem [3], which is to establish the correspondences of local features between related image pairs of the same scene. However, fragment assembly is more challenging because they do have overlapping regions. Thus, existing rule-based methods are difficult to generalize to various complex situations. Furthermore, they are time-consuming since the computational complexity of finding matching pairs from a set of fragments is typically " + }, + { + "bbox": [ + 130, + 257, + 479, + 353 + ], + "type": "inline_equation", + "content": "O(N^2)" + }, + { + "bbox": [ + 130, + 257, + 479, + 353 + ], + "type": "text", + "content": " in global search, and in practice, only a few pairs are hidden in a large number of mixed fragments." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 130, + 354, + 482, + 462 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 354, + 482, + 462 + ], + "spans": [ + { + "bbox": [ + 130, + 354, + 482, + 462 + ], + "type": "text", + "content": "Solving jigsaw puzzles [14, 37, 42] is also a similar task and learning-based approaches have been applied [2, 36]. However, two differences make these algorithms not generalizable. First, puzzle pieces usually have regular shapes, so they are best matched based on content information rather than contour features. Second, this problem usually studies matching all fragments into a complete block. Their input fragments do not come from mixed groups, and the matching results are dense and not lost. Therefore, they often use neighbor relationships to build graphs to constrain the search for the correct alignment of fragments. Once a fragment is lost, their closed-loop strategy fails." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 130, + 464, + 482, + 583 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 464, + 482, + 583 + ], + "spans": [ + { + "bbox": [ + 130, + 464, + 482, + 583 + ], + "type": "text", + "content": "In this paper, we propose PairingNet, an end-to-end deep learning network, for searching and matching pairs from a large number of image fragments. To address this problem, we first explore the use of Graph Convolutional Networks (GCNs) to extract more robust contour and texture features for our tasks. Second, we design a feature fusion way to learn local matching features by calculating the similarity matrix of each fragment pair. Third, we build a module based on the linear transformer to learn the global searching features for fragments using contrastive learning. Finally, we collect a real dataset and generate a image fragment dataset to provide data-driven and validate the effectiveness of our proposed network. The main contributions of our work are as follows:" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 137, + 594, + 479, + 665 + ], + "type": "list", + "angle": 0, + "index": 9, + "blocks": [ + { + "bbox": [ + 137, + 594, + 479, + 640 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 137, + 594, + 479, + 640 + ], + "spans": [ + { + "bbox": [ + 137, + 594, + 479, + 640 + ], + "type": "text", + "content": "- We propose a novel network to leverage the local contour and texture features for solving the pair-searching and -matching of image fragments. It provides an attempt to use advanced deep-learning techniques to solve this long-standing problem." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 137, + 641, + 479, + 665 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 137, + 641, + 479, + 665 + ], + "spans": [ + { + "bbox": [ + 137, + 641, + 479, + 665 + ], + "type": "text", + "content": "- We provide a real dataset and a generated dataset with a carefully designed algorithm to simulate real-world fragments. The algorithm generates diverse" + } + ] + } + ], + "index": 8 + } + ], + "sub_type": "text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "spans": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "type": "text", + "content": "2" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 166, + 91, + 212, + 100 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 166, + 91, + 212, + 100 + ], + "spans": [ + { + "bbox": [ + 166, + 91, + 212, + 100 + ], + "type": "text", + "content": "Zhou et al." + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 1 + }, + { + "para_blocks": [ + { + "bbox": [ + 147, + 116, + 480, + 139 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 147, + 116, + 480, + 139 + ], + "spans": [ + { + "bbox": [ + 147, + 116, + 480, + 139 + ], + "type": "text", + "content": "image fragments to facilitate the application of subsequent learning-based methods." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 138, + 140, + 480, + 163 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 140, + 480, + 163 + ], + "spans": [ + { + "bbox": [ + 138, + 140, + 480, + 163 + ], + "type": "text", + "content": "- We conduct extensive experiments, ablation studies, and visual analysis to demonstrate the effectiveness of our proposed network." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 132, + 182, + 237, + 196 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 182, + 237, + 196 + ], + "spans": [ + { + "bbox": [ + 132, + 182, + 237, + 196 + ], + "type": "text", + "content": "2 Related Work" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 132, + 209, + 356, + 224 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 209, + 356, + 224 + ], + "spans": [ + { + "bbox": [ + 132, + 209, + 356, + 224 + ], + "type": "text", + "content": "2.1 Fragments Assembly (Irregular Shape)" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 130, + 236, + 480, + 285 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 236, + 480, + 285 + ], + "spans": [ + { + "bbox": [ + 130, + 236, + 480, + 285 + ], + "type": "text", + "content": "Here, we focus on the pairwise matching algorithms of related representative works, although few of them also consider fragment relationships. Methods based on deep neural networks have been studied for object fragments; however, we could not find representative works for image fragments or even suitable datasets." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 130, + 295, + 482, + 439 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 295, + 482, + 439 + ], + "spans": [ + { + "bbox": [ + 130, + 295, + 482, + 439 + ], + "type": "text", + "content": "Image fragments. Kong and Kimia [25] solved fragment matching using geometric features (partial curve matching). Leitao and Stolfi [17] described a specific multiscale matching method for the reassembly of 2D fragmented objects. Tsamoura and Pitas [43] presented a color-based method to automatically reassemble image fragments. Huang et al. [22] relied on salient curves detected inside the different image pieces to align gapped fragments. Subsequently, both shape and appearance information along the boundaries were utilized and extracted for each piece for matching [32,49]. Furthermore, Derech et al. [12] proposed to extrapolate each fragment to address fragment abrasion. For learning-based methods, traditional machine learning (SVM/Random Forest) was used as a classifier to process extracted local features [1,38]. Then Le and Li [27] built a deep neural network to detect the compatibility of pairwise stitching." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 130, + 452, + 482, + 596 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 452, + 482, + 596 + ], + "spans": [ + { + "bbox": [ + 130, + 452, + 482, + 596 + ], + "type": "text", + "content": "Object fragments. Huang et al. [23] analyzed the geometry of the fracture surfaces by computing multi-scale surface characteristics. Yang et al. [46] matched stone tools based on contour points and their mean normals. Funkhouser et al. [15] learned a decision tree classifier to filter the candidates. Hong et al. [21] solved the problem of indistinguishable false matches by employing beam search to explore multiple registration possibilities. Ye et al. [47] presented an immersive system that can handle complex and ambiguous fragment shapes interactively with experts. In recent years, deep learning networks have been rapidly applied in 3D assembly. Huang et al. [48] proposed an assembly-oriented dynamic graph learning framework to estimate the pose of 3D parts. Chen et al. [8] proposed Neural Shape Mating (NSM) to tackle pairwise 3D geometric shape mating. Wu et al. [45] leveraged SE(3) equivariance for such shape pose disentanglement." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 130, + 609, + 482, + 669 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 609, + 482, + 669 + ], + "spans": [ + { + "bbox": [ + 130, + 609, + 482, + 669 + ], + "type": "text", + "content": "Datasets. Existing image assembly works typically collect data in two ways: manual digitalization by photographing or scanning broken fragments, or using simple algorithms to tear complete images [7, 35, 38]. The lack of a large-scale dataset prevents the emergence of deep learning networks for this task. However, multiple datasets have been introduced in 3D assembly problems [24, 26, 40]." + } + ] + } + ], + "index": 9 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 402, + 91, + 447, + 102 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 402, + 91, + 447, + 102 + ], + "spans": [ + { + "bbox": [ + 402, + 91, + 447, + 102 + ], + "type": "text", + "content": "PairingNet" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 474, + 92, + 481, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 474, + 92, + 481, + 100 + ], + "spans": [ + { + "bbox": [ + 474, + 92, + 481, + 100 + ], + "type": "text", + "content": "3" + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 2 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 133, + 114, + 482, + 276 + ], + "blocks": [ + { + "bbox": [ + 133, + 114, + 482, + 276 + ], + "lines": [ + { + "bbox": [ + 133, + 114, + 482, + 276 + ], + "spans": [ + { + "bbox": [ + 133, + 114, + 482, + 276 + ], + "type": "image", + "image_path": "cadd093b8b7545134418df87cf2d950341591b2930aa4852b5775a42b4bf0fd9.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 130, + 282, + 482, + 382 + ], + "lines": [ + { + "bbox": [ + 130, + 282, + 482, + 382 + ], + "spans": [ + { + "bbox": [ + 130, + 282, + 482, + 382 + ], + "type": "text", + "content": "Fig. 2: Pipeline of our proposed PairingNet. (a) Based on patches, we first employ a binary encoding to describe contours and use ResGCN as the backbone to extract the local contour and texture features of image fragments. (b) We design self-gated fusion to fuse the extracted features in a weighted manner to calculate the similarity matrix for each pair of fragments, and (c) a linear transformer-based encoder for learning the global features of each fragment. (d) We use a two-step strategy to train our proposed network. (e) During inference, we calculate cosine similarity to find the adjacent fragment pairs and further process the matching similarity matrix to establish better correspondences between a fragment pair." + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_caption" + } + ], + "index": 2 + }, + { + "bbox": [ + 132, + 407, + 321, + 420 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 407, + 321, + 420 + ], + "spans": [ + { + "bbox": [ + 132, + 407, + 321, + 420 + ], + "type": "text", + "content": "2.2 Jigsaw Puzzles (Regular Shape)" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 130, + 428, + 482, + 489 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 428, + 482, + 489 + ], + "spans": [ + { + "bbox": [ + 130, + 428, + 482, + 489 + ], + "type": "text", + "content": "Solving jigsaw puzzles was first proposed in [14], and then regular pieces became a key research object in computer vision. Many existing methods [10, 16, 19, 37, 41, 42] focused on building relationships of pieces based on content information rather than contour-based pair-searching/-matching. And deep learning techniques [2, 36] was also used to solve this problem." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 132, + 499, + 202, + 511 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 499, + 202, + 511 + ], + "spans": [ + { + "bbox": [ + 132, + 499, + 202, + 511 + ], + "type": "text", + "content": "3 Method" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 132, + 522, + 259, + 534 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 522, + 259, + 534 + ], + "spans": [ + { + "bbox": [ + 132, + 522, + 259, + 534 + ], + "type": "text", + "content": "3.1 Problem Statement" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 130, + 540, + 482, + 593 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 540, + 482, + 593 + ], + "spans": [ + { + "bbox": [ + 130, + 540, + 482, + 593 + ], + "type": "text", + "content": "Let us consider a fragment dataset " + }, + { + "bbox": [ + 130, + 540, + 482, + 593 + ], + "type": "inline_equation", + "content": "\\mathbf{F} = \\{f^{(i)}\\}_{i=1}^{N}" + }, + { + "bbox": [ + 130, + 540, + 482, + 593 + ], + "type": "text", + "content": ", which consists of fragments " + }, + { + "bbox": [ + 130, + 540, + 482, + 593 + ], + "type": "inline_equation", + "content": "f" + }, + { + "bbox": [ + 130, + 540, + 482, + 593 + ], + "type": "text", + "content": " torn from a certain image dataset. We define the contour points " + }, + { + "bbox": [ + 130, + 540, + 482, + 593 + ], + "type": "inline_equation", + "content": "c_{j}^{(i)} \\in \\mathbf{C}^{(i)}" + }, + { + "bbox": [ + 130, + 540, + 482, + 593 + ], + "type": "text", + "content": " for each " + }, + { + "bbox": [ + 130, + 540, + 482, + 593 + ], + "type": "inline_equation", + "content": "f^{(i)}" + }, + { + "bbox": [ + 130, + 540, + 482, + 593 + ], + "type": "text", + "content": ", where " + }, + { + "bbox": [ + 130, + 540, + 482, + 593 + ], + "type": "inline_equation", + "content": "j = 1,2,\\dots,M^{(i)}" + }, + { + "bbox": [ + 130, + 540, + 482, + 593 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 130, + 540, + 482, + 593 + ], + "type": "inline_equation", + "content": "M^{(i)}" + }, + { + "bbox": [ + 130, + 540, + 482, + 593 + ], + "type": "text", + "content": " is the contour length and " + }, + { + "bbox": [ + 130, + 540, + 482, + 593 + ], + "type": "inline_equation", + "content": "\\mathbf{C}^{(i)}" + }, + { + "bbox": [ + 130, + 540, + 482, + 593 + ], + "type": "text", + "content": " is an ordered set that contains all contour points of " + }, + { + "bbox": [ + 130, + 540, + 482, + 593 + ], + "type": "inline_equation", + "content": "f^{(i)}" + }, + { + "bbox": [ + 130, + 540, + 482, + 593 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 130, + 594, + 482, + 666 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 594, + 482, + 666 + ], + "spans": [ + { + "bbox": [ + 130, + 594, + 482, + 666 + ], + "type": "text", + "content": "Regarding any two neighbor fragments sampled from a pairset " + }, + { + "bbox": [ + 130, + 594, + 482, + 666 + ], + "type": "inline_equation", + "content": "\\mathbf{P}" + }, + { + "bbox": [ + 130, + 594, + 482, + 666 + ], + "type": "text", + "content": ", we denote the pair as " + }, + { + "bbox": [ + 130, + 594, + 482, + 666 + ], + "type": "inline_equation", + "content": "f^{(m)}" + }, + { + "bbox": [ + 130, + 594, + 482, + 666 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 130, + 594, + 482, + 666 + ], + "type": "inline_equation", + "content": "f^{(n)}" + }, + { + "bbox": [ + 130, + 594, + 482, + 666 + ], + "type": "text", + "content": ". The pair-searching task aims to retrieve the pairset " + }, + { + "bbox": [ + 130, + 594, + 482, + 666 + ], + "type": "inline_equation", + "content": "\\mathbf{P}" + }, + { + "bbox": [ + 130, + 594, + 482, + 666 + ], + "type": "text", + "content": " from a given fragment dataset " + }, + { + "bbox": [ + 130, + 594, + 482, + 666 + ], + "type": "inline_equation", + "content": "\\mathbf{F}" + }, + { + "bbox": [ + 130, + 594, + 482, + 666 + ], + "type": "text", + "content": ". Meanwhile, for the contours of the paired fragments, " + }, + { + "bbox": [ + 130, + 594, + 482, + 666 + ], + "type": "inline_equation", + "content": "\\mathbf{C}^{(m)}" + }, + { + "bbox": [ + 130, + 594, + 482, + 666 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 130, + 594, + 482, + 666 + ], + "type": "inline_equation", + "content": "\\mathbf{C}^{(n)}" + }, + { + "bbox": [ + 130, + 594, + 482, + 666 + ], + "type": "text", + "content": ", we can always find the overlapped part " + }, + { + "bbox": [ + 130, + 594, + 482, + 666 + ], + "type": "inline_equation", + "content": "\\mathbf{C}^{(m\\& n)}" + }, + { + "bbox": [ + 130, + 594, + 482, + 666 + ], + "type": "text", + "content": ". The pair-matching task is designed to find the adjacent part from any given fragment pair." + } + ] + } + ], + "index": 9 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "spans": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "type": "text", + "content": "4" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 166, + 91, + 212, + 100 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 166, + 91, + 212, + 100 + ], + "spans": [ + { + "bbox": [ + 166, + 91, + 212, + 100 + ], + "type": "text", + "content": "Zhou et al." + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 3 + }, + { + "para_blocks": [ + { + "bbox": [ + 132, + 116, + 271, + 126 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 116, + 271, + 126 + ], + "spans": [ + { + "bbox": [ + 132, + 116, + 271, + 126 + ], + "type": "text", + "content": "3.2 Network Architecture" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 130, + 137, + 482, + 245 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 137, + 482, + 245 + ], + "spans": [ + { + "bbox": [ + 130, + 137, + 482, + 245 + ], + "type": "text", + "content": "In the pair-matching task, a model has the capability to discern differences between two fragments, regardless of whether these discrepancies are on a large or minimal scale. Simultaneously, the features extracted by the model can be used to quantify the likelihood of adjacency between any given two fragments, assisting in the identification of matched pairs. However, the process of comparing every two fragments is not only laborious, but also inefficient. Therefore, utilizing the rich features the matching model extracts, we conceived a pair-searching module. This module can adeptly convert any fragment into a unique feature vector, thereby significantly accelerating the pair-searching process." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 132, + 256, + 230, + 266 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 256, + 230, + 266 + ], + "spans": [ + { + "bbox": [ + 132, + 256, + 230, + 266 + ], + "type": "text", + "content": "Feature Extraction" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 130, + 283, + 482, + 405 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 283, + 482, + 405 + ], + "spans": [ + { + "bbox": [ + 130, + 283, + 482, + 405 + ], + "type": "text", + "content": "Contour encoding. In this module, we aim to direct the model's attention to the unique features of contours. We design distinct contour encodings that also may be applied in other sketch-based fields. We tested various patch sizes and all these distinct encodings, finalizing a binary encoding and a patch size of " + }, + { + "bbox": [ + 130, + 283, + 482, + 405 + ], + "type": "inline_equation", + "content": "7 \\times 7" + }, + { + "bbox": [ + 130, + 283, + 482, + 405 + ], + "type": "text", + "content": " for our network. Subsequently, these contour patches are processed through a convolutional layer, followed by an average pooling layer and a fully connected network. We then adopt ResGCN [28] to amalgamate adjacent features, further extracting comprehensive global features. As such, for any provided fragment input denoted as " + }, + { + "bbox": [ + 130, + 283, + 482, + 405 + ], + "type": "inline_equation", + "content": "f^{(i)}" + }, + { + "bbox": [ + 130, + 283, + 482, + 405 + ], + "type": "text", + "content": " with its corresponding contour " + }, + { + "bbox": [ + 130, + 283, + 482, + 405 + ], + "type": "inline_equation", + "content": "\\mathbf{C}^{(i)}" + }, + { + "bbox": [ + 130, + 283, + 482, + 405 + ], + "type": "text", + "content": ", this module yields a return of " + }, + { + "bbox": [ + 130, + 283, + 482, + 405 + ], + "type": "inline_equation", + "content": "f_{c}^{(i)} \\in \\mathbb{R}^{M^{(i)} \\times 64}" + }, + { + "bbox": [ + 130, + 283, + 482, + 405 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 130, + 412, + 482, + 510 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 412, + 482, + 510 + ], + "spans": [ + { + "bbox": [ + 130, + 412, + 482, + 510 + ], + "type": "text", + "content": "Textural encoding. Our objective is to accentuate the distinguishing characteristics of textures in this module. We begin by cropping the image using a patch centered on each contour point. This is followed by filtering these patches through two consecutive convolutional layers, coupled with a single average pooling operation. Then, a ResGCN [28] is integrated to compile adjacent features, thereby enhancing the receptive domain of local textural characteristics. For any given fragment input represented as " + }, + { + "bbox": [ + 130, + 412, + 482, + 510 + ], + "type": "inline_equation", + "content": "f^{(i)}" + }, + { + "bbox": [ + 130, + 412, + 482, + 510 + ], + "type": "text", + "content": " along with its associated contour " + }, + { + "bbox": [ + 130, + 412, + 482, + 510 + ], + "type": "inline_equation", + "content": "\\mathbf{C}^{(i)}" + }, + { + "bbox": [ + 130, + 412, + 482, + 510 + ], + "type": "text", + "content": ", this module produces an output of " + }, + { + "bbox": [ + 130, + 412, + 482, + 510 + ], + "type": "inline_equation", + "content": "f_{t}^{(i)} \\in \\mathbb{R}^{M^{(i)} \\times 64}" + }, + { + "bbox": [ + 130, + 412, + 482, + 510 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 132, + 517, + 248, + 529 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 517, + 248, + 529 + ], + "spans": [ + { + "bbox": [ + 132, + 517, + 248, + 529 + ], + "type": "text", + "content": "Pair-matching Module" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 130, + 544, + 482, + 628 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 544, + 482, + 628 + ], + "spans": [ + { + "bbox": [ + 130, + 544, + 482, + 628 + ], + "type": "text", + "content": "Feature Fusion. Regarding the decision-making process for pair-matching and -searching tasks, the dependence on contour and textural features varies. For instance, in scenarios where both fragments are solid colors, texture serves as the primary identifying factor suggesting adjacency. However, texture alone cannot identify the matched contours. In contrast, simple contours, like straight lines, require texture features for accurate matching position determination, as contour information alone falls short." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 130, + 629, + 482, + 665 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 629, + 482, + 665 + ], + "spans": [ + { + "bbox": [ + 130, + 629, + 482, + 665 + ], + "type": "text", + "content": "Given this, we propose the incorporation of a Feature Fusion module [11] which would be capable of adaptively fine-tuning the weights assigned to these two features. The specific structure of this integral component is depicted in" + } + ] + } + ], + "index": 9 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 402, + 91, + 447, + 102 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 402, + 91, + 447, + 102 + ], + "spans": [ + { + "bbox": [ + 402, + 91, + 447, + 102 + ], + "type": "text", + "content": "PairingNet" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 474, + 92, + 480, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 474, + 92, + 480, + 100 + ], + "spans": [ + { + "bbox": [ + 474, + 92, + 480, + 100 + ], + "type": "text", + "content": "5" + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 4 + }, + { + "para_blocks": [ + { + "bbox": [ + 130, + 114, + 479, + 144 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 114, + 479, + 144 + ], + "spans": [ + { + "bbox": [ + 130, + 114, + 479, + 144 + ], + "type": "text", + "content": "Figure 2 (b). First, we concatenate " + }, + { + "bbox": [ + 130, + 114, + 479, + 144 + ], + "type": "inline_equation", + "content": "f_{t}^{(i)}" + }, + { + "bbox": [ + 130, + 114, + 479, + 144 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 130, + 114, + 479, + 144 + ], + "type": "inline_equation", + "content": "f_{c}^{(i)}" + }, + { + "bbox": [ + 130, + 114, + 479, + 144 + ], + "type": "text", + "content": ", then the weight " + }, + { + "bbox": [ + 130, + 114, + 479, + 144 + ], + "type": "inline_equation", + "content": "\\mathbf{w}^{(i)}" + }, + { + "bbox": [ + 130, + 114, + 479, + 144 + ], + "type": "text", + "content": " for each contour point " + }, + { + "bbox": [ + 130, + 114, + 479, + 144 + ], + "type": "inline_equation", + "content": "c_{j}^{(i)} \\in \\mathbf{C}^{(i)}" + }, + { + "bbox": [ + 130, + 114, + 479, + 144 + ], + "type": "text", + "content": " is computed by:" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 240, + 152, + 480, + 167 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 240, + 152, + 480, + 167 + ], + "spans": [ + { + "bbox": [ + 240, + 152, + 480, + 167 + ], + "type": "interline_equation", + "content": "\\mathbf {w} ^ {(i)} = \\sigma \\left(\\mathbf {W} _ {g} \\left(f _ {t} ^ {(i)} \\odot f _ {c} ^ {(i)}\\right) + \\mathbf {b} _ {g}\\right) ^ {(i)} \\tag {1}", + "image_path": "d24243760ddd11f9ac3650ef3b75192fb47947caed25f52b048735e798486602.jpg" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 131, + 175, + 479, + 200 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 131, + 175, + 479, + 200 + ], + "spans": [ + { + "bbox": [ + 131, + 175, + 479, + 200 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 131, + 175, + 479, + 200 + ], + "type": "inline_equation", + "content": "\\sigma" + }, + { + "bbox": [ + 131, + 175, + 479, + 200 + ], + "type": "text", + "content": " is sigmoid the activation function, " + }, + { + "bbox": [ + 131, + 175, + 479, + 200 + ], + "type": "inline_equation", + "content": "(\\widehat{\\mathbf{c}})" + }, + { + "bbox": [ + 131, + 175, + 479, + 200 + ], + "type": "text", + "content": " means channel-wise concatenation, and " + }, + { + "bbox": [ + 131, + 175, + 479, + 200 + ], + "type": "inline_equation", + "content": "\\mathbf{w} \\in \\mathbb{R}^{M^{(i)} \\times 64}" + }, + { + "bbox": [ + 131, + 175, + 479, + 200 + ], + "type": "text", + "content": ". Second, we fuse features of two modalities:" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 231, + 208, + 480, + 223 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 231, + 208, + 480, + 223 + ], + "spans": [ + { + "bbox": [ + 231, + 208, + 480, + 223 + ], + "type": "interline_equation", + "content": "\\mathbf {F} _ {f} ^ {(i)} = \\mathbf {w} ^ {(i)} \\odot \\mathbf {F} _ {t} ^ {(i)} + (1 - \\mathbf {w} ^ {(i)}) \\odot \\mathbf {F} _ {c} ^ {(i)} \\tag {2}", + "image_path": "e9039378c91d84dc783d83cd7bf17ec84b6837d34c5c72b43d4d9d745ae40038.jpg" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 131, + 232, + 479, + 269 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 131, + 232, + 479, + 269 + ], + "spans": [ + { + "bbox": [ + 131, + 232, + 479, + 269 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 131, + 232, + 479, + 269 + ], + "type": "inline_equation", + "content": "\\mathbf{F}_f^{(i)}\\in \\mathbb{R}^{M^{(i)}\\times 64}" + }, + { + "bbox": [ + 131, + 232, + 479, + 269 + ], + "type": "text", + "content": ", and " + }, + { + "bbox": [ + 131, + 232, + 479, + 269 + ], + "type": "inline_equation", + "content": "\\odot" + }, + { + "bbox": [ + 131, + 232, + 479, + 269 + ], + "type": "text", + "content": " is element-wise multiplication. With this module, we let the model adaptively decide the weights between textual and contour features, which is proven effective in later experiments." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 131, + 276, + 479, + 313 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 131, + 276, + 479, + 313 + ], + "spans": [ + { + "bbox": [ + 131, + 276, + 479, + 313 + ], + "type": "text", + "content": "Similarity Matrix. As shown in Figure 2 (b-2), for every pair of fragments " + }, + { + "bbox": [ + 131, + 276, + 479, + 313 + ], + "type": "inline_equation", + "content": "f^{(m)}" + }, + { + "bbox": [ + 131, + 276, + 479, + 313 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 131, + 276, + 479, + 313 + ], + "type": "inline_equation", + "content": "f^{(n)}" + }, + { + "bbox": [ + 131, + 276, + 479, + 313 + ], + "type": "text", + "content": " from the pairwise matrix " + }, + { + "bbox": [ + 131, + 276, + 479, + 313 + ], + "type": "inline_equation", + "content": "\\mathbf{S}^{(m\\& n)} \\in \\mathbb{R}^{M^{(m)} \\times M^{(n)}}" + }, + { + "bbox": [ + 131, + 276, + 479, + 313 + ], + "type": "text", + "content": " [9,39]." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 146, + 314, + 466, + 326 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 146, + 314, + 466, + 326 + ], + "spans": [ + { + "bbox": [ + 146, + 314, + 466, + 326 + ], + "type": "text", + "content": "We first calculate the similarity matrix " + }, + { + "bbox": [ + 146, + 314, + 466, + 326 + ], + "type": "inline_equation", + "content": "\\tilde{\\mathbf{S}}^{(m\\& n)}" + }, + { + "bbox": [ + 146, + 314, + 466, + 326 + ], + "type": "text", + "content": " by matrix multiplication:" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 252, + 334, + 480, + 355 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 252, + 334, + 480, + 355 + ], + "spans": [ + { + "bbox": [ + 252, + 334, + 480, + 355 + ], + "type": "interline_equation", + "content": "\\tilde {\\mathbf {S}} ^ {(m \\& n)} = \\frac {1}{\\sqrt {D}} \\mathbf {F} _ {f} ^ {(m)} \\left(\\mathbf {F} _ {f} ^ {(n)}\\right) ^ {T} \\tag {3}", + "image_path": "91614b00dcc96d51161b184f064c195c80966a07b506b99163bbb7112e5dae59.jpg" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 131, + 362, + 479, + 385 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 131, + 362, + 479, + 385 + ], + "spans": [ + { + "bbox": [ + 131, + 362, + 479, + 385 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 131, + 362, + 479, + 385 + ], + "type": "inline_equation", + "content": "D" + }, + { + "bbox": [ + 131, + 362, + 479, + 385 + ], + "type": "text", + "content": " is the dimension of the feature, in our model, " + }, + { + "bbox": [ + 131, + 362, + 479, + 385 + ], + "type": "inline_equation", + "content": "D" + }, + { + "bbox": [ + 131, + 362, + 479, + 385 + ], + "type": "text", + "content": " is 64. Then, each entry in the matrix " + }, + { + "bbox": [ + 131, + 362, + 479, + 385 + ], + "type": "inline_equation", + "content": "\\mathbf{S}^{(m\\& n)}" + }, + { + "bbox": [ + 131, + 362, + 479, + 385 + ], + "type": "text", + "content": " is normalized:" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 209, + 392, + 480, + 423 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 209, + 392, + 480, + 423 + ], + "spans": [ + { + "bbox": [ + 209, + 392, + 480, + 423 + ], + "type": "interline_equation", + "content": "\\mathbf {S} ^ {(m \\& n)} (i, j) = \\frac {\\exp \\left(\\tilde {\\mathbf {S}} _ {i , j} ^ {(m \\& n)}\\right)}{\\sum_ {i} \\exp \\left(\\tilde {\\mathbf {S}} _ {i , j} ^ {(m \\& n)}\\right)} \\frac {\\exp \\left(\\tilde {\\mathbf {S}} _ {i , j} ^ {(m \\& n)}\\right)}{\\sum_ {j} \\exp \\left(\\tilde {\\mathbf {S}} _ {i , j} ^ {(m \\& n)}\\right)} \\tag {4}", + "image_path": "aea7df3b7c9ad381ce69aa78b21632549428761d9da11c4339541e7e232cfa29.jpg" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 131, + 442, + 479, + 491 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 131, + 442, + 479, + 491 + ], + "spans": [ + { + "bbox": [ + 131, + 442, + 479, + 491 + ], + "type": "text", + "content": "Pair-searching Module We introduce a modified feature fusion module to process features extracted by the backbone network and encode any fragment " + }, + { + "bbox": [ + 131, + 442, + 479, + 491 + ], + "type": "inline_equation", + "content": "f^{(i)}" + }, + { + "bbox": [ + 131, + 442, + 479, + 491 + ], + "type": "text", + "content": " into a feature vector " + }, + { + "bbox": [ + 131, + 442, + 479, + 491 + ], + "type": "inline_equation", + "content": "v_{s}^{(i)} \\in \\mathbb{R}^{128}" + }, + { + "bbox": [ + 131, + 442, + 479, + 491 + ], + "type": "text", + "content": ". It is based on the discovery that using fused features in the pair-matching task will lower the performance." + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 131, + 492, + 479, + 563 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 131, + 492, + 479, + 563 + ], + "spans": [ + { + "bbox": [ + 131, + 492, + 479, + 563 + ], + "type": "text", + "content": "To be specific, as shown in Figure 2 (c), every feature branch utilizes a Linear Transformer Encoder [44] to enhance the encoding of both contour and textural features. Subsequent to the modified fusion module, we add a fully-connected layer to encode the fused feature regarding per contour point. Then an average pooling layer is applied to map the feature matrix to a singular vector. Finally, another fully-connected layer is used to encode the feature." + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 131, + 563, + 479, + 613 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 131, + 563, + 479, + 613 + ], + "spans": [ + { + "bbox": [ + 131, + 563, + 479, + 613 + ], + "type": "text", + "content": "For every pair of fragment " + }, + { + "bbox": [ + 131, + 563, + 479, + 613 + ], + "type": "inline_equation", + "content": "f^{(i)}" + }, + { + "bbox": [ + 131, + 563, + 479, + 613 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 131, + 563, + 479, + 613 + ], + "type": "inline_equation", + "content": "f^{(j)}" + }, + { + "bbox": [ + 131, + 563, + 479, + 613 + ], + "type": "text", + "content": " from the fragment set " + }, + { + "bbox": [ + 131, + 563, + 479, + 613 + ], + "type": "inline_equation", + "content": "\\mathbf{F}" + }, + { + "bbox": [ + 131, + 563, + 479, + 613 + ], + "type": "text", + "content": ", this module allows them to be encoded as " + }, + { + "bbox": [ + 131, + 563, + 479, + 613 + ], + "type": "inline_equation", + "content": "v_{s}^{(i)}" + }, + { + "bbox": [ + 131, + 563, + 479, + 613 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 131, + 563, + 479, + 613 + ], + "type": "inline_equation", + "content": "v_{s}^{(j)}" + }, + { + "bbox": [ + 131, + 563, + 479, + 613 + ], + "type": "text", + "content": ". By calculating their inner product, we can efficiently quantify and contrast their similarity, thereby expediently identifying neighbor pairs." + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 131, + 626, + 230, + 637 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 131, + 626, + 230, + 637 + ], + "spans": [ + { + "bbox": [ + 131, + 626, + 230, + 637 + ], + "type": "text", + "content": "3.3 Loss Function" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 131, + 641, + 479, + 665 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 131, + 641, + 479, + 665 + ], + "spans": [ + { + "bbox": [ + 131, + 641, + 479, + 665 + ], + "type": "text", + "content": "We use two loss functions to train our network, and the training process is shown in Figure 2 (d)." + } + ] + } + ], + "index": 16 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "spans": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "type": "text", + "content": "6" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 166, + 91, + 212, + 101 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 166, + 91, + 212, + 101 + ], + "spans": [ + { + "bbox": [ + 166, + 91, + 212, + 101 + ], + "type": "text", + "content": "Zhou et al." + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 5 + }, + { + "para_blocks": [ + { + "bbox": [ + 130, + 116, + 479, + 152 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 116, + 479, + 152 + ], + "spans": [ + { + "bbox": [ + 130, + 116, + 479, + 152 + ], + "type": "text", + "content": "Pair-matching Loss Given the extreme imbalance in the ratio of positive to negative samples in the pair-matching task, we adopt the concept of focal loss [30] to concentrate on learning challenging correspondences." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 130, + 152, + 480, + 194 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 152, + 480, + 194 + ], + "spans": [ + { + "bbox": [ + 130, + 152, + 480, + 194 + ], + "type": "text", + "content": "For any paired fragments " + }, + { + "bbox": [ + 130, + 152, + 480, + 194 + ], + "type": "inline_equation", + "content": "f^{(m)}" + }, + { + "bbox": [ + 130, + 152, + 480, + 194 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 130, + 152, + 480, + 194 + ], + "type": "inline_equation", + "content": "f^{(n)}" + }, + { + "bbox": [ + 130, + 152, + 480, + 194 + ], + "type": "text", + "content": ", entries in the ground truth similarity matrix " + }, + { + "bbox": [ + 130, + 152, + 480, + 194 + ], + "type": "inline_equation", + "content": "\\mathbf{S}_{gt}^{(m\\& n)}" + }, + { + "bbox": [ + 130, + 152, + 480, + 194 + ], + "type": "text", + "content": " are 1 for matched contour points and 0 for unmatched ones. We calculate the matching loss " + }, + { + "bbox": [ + 130, + 152, + 480, + 194 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_{match}^{(m\\& n)}" + }, + { + "bbox": [ + 130, + 152, + 480, + 194 + ], + "type": "text", + "content": " as follows:" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 200, + 197, + 480, + 230 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 200, + 197, + 480, + 230 + ], + "spans": [ + { + "bbox": [ + 200, + 197, + 480, + 230 + ], + "type": "interline_equation", + "content": "\\begin{array}{l} \\mathcal {L} _ {\\text {m a t c h}} ^ {(m \\& n)} = - \\sum_ {(0, 0)} (\\beta_ {1} (\\mathbf {1} - \\mathbf {S} ^ {(m \\& n)}) ^ {\\gamma} \\log \\mathbf {S} ^ {(m \\& n)} \\mathbf {S} _ {g t} ^ {(m \\& n)}) \\tag {5} \\\\ + \\beta_ {2} (\\mathbf {S} ^ {(m \\& n)}) ^ {\\gamma} \\log \\left(\\mathbf {1} - \\mathbf {S} ^ {(m \\& n)}\\right) (\\mathbf {1} - \\mathbf {S} _ {g t} ^ {(m \\& n)}) \\\\ \\end{array}", + "image_path": "a6ec7f16d7cd113c2e2d0749f1acdf4eee068cf10918868a5a683272c659c835.jpg" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 130, + 238, + 479, + 286 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 238, + 479, + 286 + ], + "spans": [ + { + "bbox": [ + 130, + 238, + 479, + 286 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 130, + 238, + 479, + 286 + ], + "type": "inline_equation", + "content": "\\beta_{1}" + }, + { + "bbox": [ + 130, + 238, + 479, + 286 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 130, + 238, + 479, + 286 + ], + "type": "inline_equation", + "content": "\\beta_{2}" + }, + { + "bbox": [ + 130, + 238, + 479, + 286 + ], + "type": "text", + "content": " are the balancing weights and " + }, + { + "bbox": [ + 130, + 238, + 479, + 286 + ], + "type": "inline_equation", + "content": "\\beta_{1} + \\beta_{2} = 1" + }, + { + "bbox": [ + 130, + 238, + 479, + 286 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 130, + 238, + 479, + 286 + ], + "type": "inline_equation", + "content": "\\gamma" + }, + { + "bbox": [ + 130, + 238, + 479, + 286 + ], + "type": "text", + "content": " is the decay factor. " + }, + { + "bbox": [ + 130, + 238, + 479, + 286 + ], + "type": "inline_equation", + "content": "\\sum" + }, + { + "bbox": [ + 130, + 238, + 479, + 286 + ], + "type": "text", + "content": " adds all the entries in the input matrix and all operations except " + }, + { + "bbox": [ + 130, + 238, + 479, + 286 + ], + "type": "inline_equation", + "content": "\\sum" + }, + { + "bbox": [ + 130, + 238, + 479, + 286 + ], + "type": "text", + "content": " are element-wise. By calculating and adding the pair-matching loss for all paired fragments " + }, + { + "bbox": [ + 130, + 238, + 479, + 286 + ], + "type": "inline_equation", + "content": "f^{(m)}, f^{(n)}" + }, + { + "bbox": [ + 130, + 238, + 479, + 286 + ], + "type": "text", + "content": " from the pairwise " + }, + { + "bbox": [ + 130, + 238, + 479, + 286 + ], + "type": "inline_equation", + "content": "\\mathbf{P}" + }, + { + "bbox": [ + 130, + 238, + 479, + 286 + ], + "type": "text", + "content": ", we get " + }, + { + "bbox": [ + 130, + 238, + 479, + 286 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_{match}" + }, + { + "bbox": [ + 130, + 238, + 479, + 286 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 130, + 296, + 480, + 333 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 296, + 480, + 333 + ], + "spans": [ + { + "bbox": [ + 130, + 296, + 480, + 333 + ], + "type": "text", + "content": "Pair-searching Loss In the pair-searching task, we employ InfoNCE loss [34] as the contrastive loss function " + }, + { + "bbox": [ + 130, + 296, + 480, + 333 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_{search}" + }, + { + "bbox": [ + 130, + 296, + 480, + 333 + ], + "type": "text", + "content": ". We utilize paired samples for constructing positive sample pairs, and unpaired samples for negative sample pairs." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 131, + 346, + 208, + 357 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 131, + 346, + 208, + 357 + ], + "spans": [ + { + "bbox": [ + 131, + 346, + 208, + 357 + ], + "type": "text", + "content": "3.4 Inference" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 130, + 361, + 479, + 384 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 361, + 479, + 384 + ], + "spans": [ + { + "bbox": [ + 130, + 361, + 479, + 384 + ], + "type": "text", + "content": "Pair-matching In this section, we target to retrieve matched contour from the similarity matrix " + }, + { + "bbox": [ + 130, + 361, + 479, + 384 + ], + "type": "inline_equation", + "content": "\\mathbf{S}^{(m\\& n)}" + }, + { + "bbox": [ + 130, + 361, + 479, + 384 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 130, + 385, + 480, + 505 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 385, + 480, + 505 + ], + "spans": [ + { + "bbox": [ + 130, + 385, + 480, + 505 + ], + "type": "text", + "content": "We first apply a threshold " + }, + { + "bbox": [ + 130, + 385, + 480, + 505 + ], + "type": "inline_equation", + "content": "\\epsilon = 0.006" + }, + { + "bbox": [ + 130, + 385, + 480, + 505 + ], + "type": "text", + "content": " to filter out noisy entries. Entries with value lower than the threshold is set to 0. Because contour pointset " + }, + { + "bbox": [ + 130, + 385, + 480, + 505 + ], + "type": "inline_equation", + "content": "\\mathbf{C}^{(m)}" + }, + { + "bbox": [ + 130, + 385, + 480, + 505 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 130, + 385, + 480, + 505 + ], + "type": "inline_equation", + "content": "\\mathbf{C}^{(n)}" + }, + { + "bbox": [ + 130, + 385, + 480, + 505 + ], + "type": "text", + "content": " are ordered sets, for any ground truth similarity matrix " + }, + { + "bbox": [ + 130, + 385, + 480, + 505 + ], + "type": "inline_equation", + "content": "\\mathbf{S}_{gt}^{(m\\& n)}" + }, + { + "bbox": [ + 130, + 385, + 480, + 505 + ], + "type": "text", + "content": ", if the start for contour points is the left-top corner, we can observe that the part with a value of 1 in the matrix must be arranged in the direction from upper left to lower right and connected into a line segment. Therefore, this inspires us to enhance the filtered similarity matrix. We design an eroding kernel " + }, + { + "bbox": [ + 130, + 385, + 480, + 505 + ], + "type": "inline_equation", + "content": "\\mathbf{K}_e" + }, + { + "bbox": [ + 130, + 385, + 480, + 505 + ], + "type": "text", + "content": " to filter out not dominant segments, which is adopted once. Then a dilating kernel " + }, + { + "bbox": [ + 130, + 385, + 480, + 505 + ], + "type": "inline_equation", + "content": "\\mathbf{K}_d" + }, + { + "bbox": [ + 130, + 385, + 480, + 505 + ], + "type": "text", + "content": " is applied once to connect close but split segments. The definitions of two kernels are as follows:" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 241, + 516, + 480, + 550 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 241, + 516, + 480, + 550 + ], + "spans": [ + { + "bbox": [ + 241, + 516, + 480, + 550 + ], + "type": "interline_equation", + "content": "\\mathbf {K} _ {e} = \\left[ \\begin{array}{l l l} 0 & 0 & 1 \\\\ 0 & 0 & 0 \\\\ 1 & 0 & 0 \\end{array} \\right], \\quad \\mathbf {K} _ {d} = \\left[ \\begin{array}{l l l} 0 & 0 & 1 \\\\ 0 & 1 & 0 \\\\ 1 & 0 & 0 \\end{array} \\right]. \\tag {6}", + "image_path": "27aa8e00e5246361e46972ca44e4be069e2de665ccea557235f5ebd1465bc715.jpg" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 130, + 553, + 479, + 578 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 553, + 479, + 578 + ], + "spans": [ + { + "bbox": [ + 130, + 553, + 479, + 578 + ], + "type": "text", + "content": "Finally, we use RANSAC [13] to estimate the correspondences of two fragments and obtain the transformation matrix." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 130, + 587, + 479, + 648 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 587, + 479, + 648 + ], + "spans": [ + { + "bbox": [ + 130, + 587, + 479, + 648 + ], + "type": "text", + "content": "Pair-searching For N fragments " + }, + { + "bbox": [ + 130, + 587, + 479, + 648 + ], + "type": "inline_equation", + "content": "f^{(i)} \\in \\mathbf{F}, i = 1,2,\\dots,N" + }, + { + "bbox": [ + 130, + 587, + 479, + 648 + ], + "type": "text", + "content": ", using the searching module, we can get a feature matrix " + }, + { + "bbox": [ + 130, + 587, + 479, + 648 + ], + "type": "inline_equation", + "content": "\\mathbf{V}_s = [v_1,v_2,\\ldots ,v_N]^T \\in \\mathbb{R}^{N\\times 128}" + }, + { + "bbox": [ + 130, + 587, + 479, + 648 + ], + "type": "text", + "content": ". Subsequently, we compute the cosine similarity between all global features to form the cosine similarity matrix. This allows us to obtain the similarity score for any pair of fragments." + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 131, + 656, + 201, + 669 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 131, + 656, + 201, + 669 + ], + "spans": [ + { + "bbox": [ + 131, + 656, + 201, + 669 + ], + "type": "text", + "content": "4 Dataset" + } + ] + } + ], + "index": 13 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 402, + 91, + 447, + 102 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 402, + 91, + 447, + 102 + ], + "spans": [ + { + "bbox": [ + 402, + 91, + 447, + 102 + ], + "type": "text", + "content": "PairingNet" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 474, + 92, + 481, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 474, + 92, + 481, + 100 + ], + "spans": [ + { + "bbox": [ + 474, + 92, + 481, + 100 + ], + "type": "text", + "content": "7" + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 6 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 135, + 118, + 478, + 199 + ], + "blocks": [ + { + "bbox": [ + 135, + 118, + 478, + 199 + ], + "lines": [ + { + "bbox": [ + 135, + 118, + 478, + 199 + ], + "spans": [ + { + "bbox": [ + 135, + 118, + 478, + 199 + ], + "type": "image", + "image_path": "3f9ae64a0768296ac6a18f424d31230e30347881c843d7b2754bfeef048f9d2a.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 132, + 205, + 482, + 228 + ], + "lines": [ + { + "bbox": [ + 132, + 205, + 482, + 228 + ], + "spans": [ + { + "bbox": [ + 132, + 205, + 482, + 228 + ], + "type": "text", + "content": "Fig. 3: Dataset creation. (a) The process of generating our dataset. (b) Illustration of Algorithm 1 for tearing a complete image." + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_caption" + } + ], + "index": 2 + }, + { + "type": "table", + "bbox": [ + 204, + 275, + 411, + 319 + ], + "blocks": [ + { + "bbox": [ + 132, + 243, + 482, + 268 + ], + "lines": [ + { + "bbox": [ + 132, + 243, + 482, + 268 + ], + "spans": [ + { + "bbox": [ + 132, + 243, + 482, + 268 + ], + "type": "text", + "content": "Table 1: Dataset division. We divide the entire test set into three difficulty levels according to the contour overlapping proportion of fragment pairs." + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 204, + 275, + 411, + 319 + ], + "lines": [ + { + "bbox": [ + 204, + 275, + 411, + 319 + ], + "spans": [ + { + "bbox": [ + 204, + 275, + 411, + 319 + ], + "type": "table", + "html": "
TotalTrainValidationTest
FullHighMediumLow
Fragments819640988193279107116711090
Pairs14951365413723706631103604
", + "image_path": "5787ad808ac71d71d8895da02e2410f9f63198efe1baa45b104e71346fe5195c.jpg" + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "table_body" + } + ], + "index": 5 + }, + { + "bbox": [ + 130, + 344, + 482, + 489 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 344, + 482, + 489 + ], + "spans": [ + { + "bbox": [ + 130, + 344, + 482, + 489 + ], + "type": "text", + "content": "Generated Dataset. We design Algorithm 1 to tear a series of complete images and create a new image fragment dataset with data close to the real world. The process is also shown in Figure 3. To control the shape and area of the fragments, we first find the circumscribed circle of a fragment and take two points at an appropriate distance on the circle to determine the starting and ending points of the segmentation. Then, to simulate a more realistic fragment contour, we randomly divide the cutting line into several segments, each segment randomly choosing a straight line or an irregular curve. The irregular curve is synthesized by a series of Fourier orthogonal bases to make the shape of the fragments more diverse and similar to the edge of a real torn paper. We set a minimum threshold for the fragments, and the fragment is re-segmented if a very small fragment appears." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 130, + 490, + 483, + 646 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 490, + 483, + 646 + ], + "spans": [ + { + "bbox": [ + 130, + 490, + 483, + 646 + ], + "type": "text", + "content": "To create our dataset, we collect 390 images of ten themes from Pexels [4] and empirically set the generation parameters shown in Algorithm 1 to obtain fragments. We calculate the distributions of our generated data, as shown in Figure 4. The distribution of fragment areas exhibits a long-tail pattern, which is consistent with the distribution of fragment areas observed when tearing large amounts of paper in real-life scenarios. The contour overlapping proportion of fragment pairs is closer to a uniform distribution over most ranges, indicating that our algorithm generates various types of fragment pairs. Then, we divide the dataset into training set, validation set, and test set in 5:1:4. We further divide the full test set into three difficulty levels based on the overlapping proportion of contour segments between fragment pairs. Low difficulty means that the pair of fragments has larger overlapping contour segments, and vice versa. The details are shown in Table 1." + } + ] + } + ], + "index": 7 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "spans": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "type": "text", + "content": "8" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 166, + 91, + 212, + 100 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 166, + 91, + 212, + 100 + ], + "spans": [ + { + "bbox": [ + 166, + 91, + 212, + 100 + ], + "type": "text", + "content": "Zhou et al." + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 7 + }, + { + "para_blocks": [ + { + "type": "code", + "bbox": [ + 132, + 126, + 481, + 457 + ], + "blocks": [ + { + "bbox": [ + 133, + 115, + 297, + 125 + ], + "lines": [ + { + "bbox": [ + 133, + 115, + 297, + 125 + ], + "spans": [ + { + "bbox": [ + 133, + 115, + 297, + 125 + ], + "type": "text", + "content": "Algorithm 1 Generate fragments of an image" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "code_caption" + }, + { + "bbox": [ + 132, + 126, + 481, + 457 + ], + "lines": [ + { + "bbox": [ + 132, + 126, + 481, + 457 + ], + "spans": [ + { + "bbox": [ + 132, + 126, + 481, + 457 + ], + "type": "text", + "content": "Input: List of fragments " + }, + { + "bbox": [ + 132, + 126, + 481, + 457 + ], + "type": "inline_equation", + "content": "L_{frag} = \\{\\}" + }, + { + "bbox": [ + 132, + 126, + 481, + 457 + ], + "type": "text", + "content": " , a complete image " + }, + { + "bbox": [ + 132, + 126, + 481, + 457 + ], + "type": "inline_equation", + "content": "I" + }, + { + "bbox": [ + 132, + 126, + 481, + 457 + ], + "type": "text", + "content": " Max iterations " + }, + { + "bbox": [ + 132, + 126, + 481, + 457 + ], + "type": "inline_equation", + "content": "t_{max} = 40" + }, + { + "bbox": [ + 132, + 126, + 481, + 457 + ], + "type": "text", + "content": " min arc length ratio " + }, + { + "bbox": [ + 132, + 126, + 481, + 457 + ], + "type": "inline_equation", + "content": "\\tau = 0.9" + }, + { + "bbox": [ + 132, + 126, + 481, + 457 + ], + "type": "text", + "content": " max number of cutting points " + }, + { + "bbox": [ + 132, + 126, + 481, + 457 + ], + "type": "inline_equation", + "content": "N_{max} =" + }, + { + "bbox": [ + 132, + 126, + 481, + 457 + ], + "type": "text", + "content": " 3, min distance from cutting point to edge of fragment " + }, + { + "bbox": [ + 132, + 126, + 481, + 457 + ], + "type": "inline_equation", + "content": "D_{min} = 100" + }, + { + "bbox": [ + 132, + 126, + 481, + 457 + ], + "type": "text", + "content": " , the number of Fourier orthogonal basis " + }, + { + "bbox": [ + 132, + 126, + 481, + 457 + ], + "type": "inline_equation", + "content": "n = 20" + }, + { + "bbox": [ + 132, + 126, + 481, + 457 + ], + "type": "text", + "content": " scaling ratios " + }, + { + "bbox": [ + 132, + 126, + 481, + 457 + ], + "type": "inline_equation", + "content": "s_1 = 0.25" + }, + { + "bbox": [ + 132, + 126, + 481, + 457 + ], + "type": "inline_equation", + "content": "s_2 = 0.0067" + }, + { + "bbox": [ + 132, + 126, + 481, + 457 + ], + "type": "inline_equation", + "content": "s_3 = 1.5" + }, + { + "bbox": [ + 132, + 126, + 481, + 457 + ], + "type": "inline_equation", + "content": "s_4 = 0.3" + }, + { + "bbox": [ + 132, + 126, + 481, + 457 + ], + "type": "text", + "content": " , probability of curve type " + }, + { + "bbox": [ + 132, + 126, + 481, + 457 + ], + "type": "inline_equation", + "content": "\\rho = 0.5" + }, + { + "bbox": [ + 132, + 126, + 481, + 457 + ], + "type": "text", + "content": " , min fragment size " + }, + { + "bbox": [ + 132, + 126, + 481, + 457 + ], + "type": "inline_equation", + "content": "h_{min} = 150" + }, + { + "bbox": [ + 132, + 126, + 481, + 457 + ], + "type": "inline_equation", + "content": "w_{min} = 150" + }, + { + "bbox": [ + 132, + 126, + 481, + 457 + ], + "type": "text", + "content": " \nOutput: Collection of fragments " + }, + { + "bbox": [ + 132, + 126, + 481, + 457 + ], + "type": "inline_equation", + "content": "L_{frag}" + }, + { + "bbox": [ + 132, + 126, + 481, + 457 + ], + "type": "text", + "content": " \n1: put I in " + }, + { + "bbox": [ + 132, + 126, + 481, + 457 + ], + "type": "inline_equation", + "content": "L_{frag}" + }, + { + "bbox": [ + 132, + 126, + 481, + 457 + ], + "type": "text", + "content": " \n2: while t < tmax do \n3: f<-randomly_pick(Lfrag), R<-bounding_rectangle(f) \n4: W,H<-width_and_height(R), C<-circumcircle(R) \n5: ldia<-diameter(C) \n6: \n7: // generate the cutting points \n8: repeat \n9: pa, pb<-randomly_pick(C) \n10: larc<-smaller_ar_c_length(pa,pb) \n11: until larc>τ·half_perimeter(C) \n12: pstart, pend<-intersection(f,pa,pb) \n13: m<-randomly_pick{0,...,Nmax} \n14: repeat \n15: [p1, ..., pm]<-random_select_insidepoints(f,m) \n16: d<-min_distance_to_edge(f,p1, ..., pm) \n17: until d > Dmin \n18: put pstart, pend in [p1, ..., pm] \n19: \n20: // generate the cutting segments \n21: for each pi∈[pstart, p1, ..., pm, pend] do \n22: irregular line " + }, + { + "bbox": [ + 132, + 126, + 481, + 457 + ], + "type": "inline_equation", + "content": "l_{irr} = \\{\\}" + }, + { + "bbox": [ + 132, + 126, + 481, + 457 + ], + "type": "text", + "content": " , phase " + }, + { + "bbox": [ + 132, + 126, + 481, + 457 + ], + "type": "inline_equation", + "content": "\\phi \\sim \\mathcal{U}(-\\pi ,\\pi)" + }, + { + "bbox": [ + 132, + 126, + 481, + 457 + ], + "type": "text", + "content": " amplitude A " + }, + { + "bbox": [ + 132, + 126, + 481, + 457 + ], + "type": "inline_equation", + "content": "\\sim \\mathcal{N}(s_1H,s_2H)" + }, + { + "bbox": [ + 132, + 126, + 481, + 457 + ], + "type": "text", + "content": " period T " + }, + { + "bbox": [ + 132, + 126, + 481, + 457 + ], + "type": "inline_equation", + "content": "\\sim \\mathcal{N}(s_3W,s_4W)" + }, + { + "bbox": [ + 132, + 126, + 481, + 457 + ], + "type": "text", + "content": " \n23: for each x-coordinate between (pi,pi+1) do \n24: y<∑i=0nA sin(2πix+φ)+H/2 \n25: put {x,y} in lirr \n26: end for \n27: lstr← straight_line(pi,pi+1) \n28: lcut← select_and_connect(lirr,lstr,ρ,lcut) \n29: end for \n30: f1, f2<-cut_structure(f,lcut) \n31: if size(f1)>hmin,wmin and size(f2)>hmin,wmin then \n32: put f1, f2 in Lfrag \n33: end if \n34: end while" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "code_body" + } + ], + "index": 3, + "sub_type": "algorithm" + }, + { + "bbox": [ + 132, + 479, + 481, + 552 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 479, + 481, + 552 + ], + "spans": [ + { + "bbox": [ + 132, + 479, + 481, + 552 + ], + "type": "text", + "content": "Real Dataset. We also collect a series of real image fragments to construct a real dataset. We printed 34 complete images (not included in our generated dataset), manually cut them into fragments, and then scanned them. Then, we apply some post-processing steps to the scanned image fragments to obtain the contour of the fragments and the annotation of the matching points. More details about real dataset can be found in the supplementary materials." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 132, + 562, + 230, + 575 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 562, + 230, + 575 + ], + "spans": [ + { + "bbox": [ + 132, + 562, + 230, + 575 + ], + "type": "text", + "content": "5 Experiments" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 132, + 578, + 269, + 590 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 578, + 269, + 590 + ], + "spans": [ + { + "bbox": [ + 132, + 578, + 269, + 590 + ], + "type": "text", + "content": "5.1 Comparison Methods" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 130, + 593, + 481, + 666 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 593, + 481, + 666 + ], + "spans": [ + { + "bbox": [ + 130, + 593, + 481, + 666 + ], + "type": "text", + "content": "Finding comparison methods is not easy. Because rule-based methods have been around for quite some time, lack source code, and tend to exhibit lower performance compared to advanced deep learning methods. However, most existing deep learning methods require extensive modifications if applied to our task. Therefore, we employ Jigsawnet [27] (without loop consistency) and a rule-based method [49] (Step 1) as our comparison methods." + } + ] + } + ], + "index": 7 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 402, + 91, + 447, + 102 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 402, + 91, + 447, + 102 + ], + "spans": [ + { + "bbox": [ + 402, + 91, + 447, + 102 + ], + "type": "text", + "content": "PairingNet" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 474, + 92, + 481, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 474, + 92, + 481, + 100 + ], + "spans": [ + { + "bbox": [ + 474, + 92, + 481, + 100 + ], + "type": "text", + "content": "9" + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 8 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 132, + 114, + 309, + 171 + ], + "blocks": [ + { + "bbox": [ + 132, + 114, + 309, + 171 + ], + "lines": [ + { + "bbox": [ + 132, + 114, + 309, + 171 + ], + "spans": [ + { + "bbox": [ + 132, + 114, + 309, + 171 + ], + "type": "image", + "image_path": "32bc207a37df5bb428d522db3af308c10cc8263dc6e5e54facc03970e7d2f24f.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 130, + 178, + 482, + 224 + ], + "lines": [ + { + "bbox": [ + 130, + 178, + 482, + 224 + ], + "spans": [ + { + "bbox": [ + 130, + 178, + 482, + 224 + ], + "type": "text", + "content": "Fig. 4: Dataset statistics. (a) Image area distribution of the generated fragments. (b) Proportional distribution of overlapping contour segments of the generated fragments. (c) Image area distribution of the scanned real fragments. (b) Proportional distribution of overlapping contour segments of the scanned real fragments." + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "image_caption" + } + ], + "index": 2 + }, + { + "type": "image", + "bbox": [ + 317, + 114, + 480, + 171 + ], + "blocks": [ + { + "bbox": [ + 317, + 114, + 480, + 171 + ], + "lines": [ + { + "bbox": [ + 317, + 114, + 480, + 171 + ], + "spans": [ + { + "bbox": [ + 317, + 114, + 480, + 171 + ], + "type": "image", + "image_path": "34f244442e704b1028672b5542af0b539a890a23fef6ee6f71c96849fbee82c2.jpg" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_body" + } + ], + "index": 3 + }, + { + "type": "image", + "bbox": [ + 136, + 234, + 299, + 292 + ], + "blocks": [ + { + "bbox": [ + 136, + 234, + 299, + 292 + ], + "lines": [ + { + "bbox": [ + 136, + 234, + 299, + 292 + ], + "spans": [ + { + "bbox": [ + 136, + 234, + 299, + 292 + ], + "type": "image", + "image_path": "538965f920f9182e8f4cb9f7d734e2ab0305744a7a550aab2da822fe1cc18cf9.jpg" + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 131, + 300, + 479, + 322 + ], + "lines": [ + { + "bbox": [ + 131, + 300, + 479, + 322 + ], + "spans": [ + { + "bbox": [ + 131, + 300, + 479, + 322 + ], + "type": "text", + "content": "Fig. 5: Violin plots of the matching results of our method and the comparison method: (a) RE, NTE, and HD on the test set. (b) HD on the test sets of different difficulties." + } + ] + } + ], + "index": 7, + "angle": 0, + "type": "image_caption" + } + ], + "index": 5 + }, + { + "type": "image", + "bbox": [ + 304, + 236, + 478, + 292 + ], + "blocks": [ + { + "bbox": [ + 304, + 236, + 478, + 292 + ], + "lines": [ + { + "bbox": [ + 304, + 236, + 478, + 292 + ], + "spans": [ + { + "bbox": [ + 304, + 236, + 478, + 292 + ], + "type": "image", + "image_path": "a821ca8844adee84036eda4afb9a636a3a266e58cc67cc9591536e033fff6f64.jpg" + } + ] + } + ], + "index": 6, + "angle": 0, + "type": "image_body" + } + ], + "index": 6 + }, + { + "bbox": [ + 131, + 346, + 279, + 358 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 131, + 346, + 279, + 358 + ], + "spans": [ + { + "bbox": [ + 131, + 346, + 279, + 358 + ], + "type": "text", + "content": "5.2 Implementation Details" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 130, + 367, + 482, + 427 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 367, + 482, + 427 + ], + "spans": [ + { + "bbox": [ + 130, + 367, + 482, + 427 + ], + "type": "text", + "content": "Our PairingNet was implemented by PyTorch, and conducted on a server with 4 RTX A40 GPUs and Intel® Xeon® Gold 5220 CPUs (72 cores). We used the Adam optimizer with an initial learning rate of 0.001, which was adjusted using a cosine annealing strategy. For training pair-matching module, the batch size was set to 20, and for the pair-searching module, the batch size was set to 175." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 130, + 427, + 482, + 523 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 427, + 482, + 523 + ], + "spans": [ + { + "bbox": [ + 130, + 427, + 482, + 523 + ], + "type": "text", + "content": "For comparison methods, given that the performance bottleneck of them resides in the CPU, utilizing GPUs for computational acceleration is deemed inappropriate. Thus we use two more powerful CPUs computing platforms to run the comparison method. We used the official sourcecode (implemented by TensorFlow) of Jigsawnet and conducted experiments on a server with 4 Quadro RTX 6000 GPUs and Intel® Xeon® Gold 5220 CPUs (144 cores). We reproduced the rule-based method by Python and conducted experiments on a node of a Supercomputing platform with Intel® Xeon® Gold 6258R CPUs (128 cores)." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 130, + 523, + 482, + 583 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 523, + 482, + 583 + ], + "spans": [ + { + "bbox": [ + 130, + 523, + 482, + 583 + ], + "type": "text", + "content": "For evaluation metrics, we employ Recall@k and NDCG@k [5] for the pair-searching task, and Registration recall (RR), Hausdorff distance (HD), Radians Error (RE) and Normalized Translation Error (NTE) for the pair-matching task. We normalize Translation Error (TE) by the sum of the areas of fragment pairs to obtain the NTE, considering the large difference in area between fragments." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 131, + 601, + 198, + 613 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 131, + 601, + 198, + 613 + ], + "spans": [ + { + "bbox": [ + 131, + 601, + 198, + 613 + ], + "type": "text", + "content": "5.3 Results" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 130, + 617, + 482, + 666 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 617, + 482, + 666 + ], + "spans": [ + { + "bbox": [ + 130, + 617, + 482, + 666 + ], + "type": "text", + "content": "Performance on Generated Dataset. As shown in Table 2, our network outperforms the comparison methods on both the full test set and on every set of difficulty levels. Jigsawnet shows a certain superiority in pair-searching on the high-difficulty set, probably because it treats the problem as a simple binary" + } + ] + } + ], + "index": 13 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "spans": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "type": "text", + "content": "10" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 166, + 91, + 212, + 100 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 166, + 91, + 212, + 100 + ], + "spans": [ + { + "bbox": [ + 166, + 91, + 212, + 100 + ], + "type": "text", + "content": "Zhou et al." + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 9 + }, + { + "para_blocks": [ + { + "type": "table", + "bbox": [ + 134, + 157, + 480, + 258 + ], + "blocks": [ + { + "bbox": [ + 132, + 114, + 482, + 148 + ], + "lines": [ + { + "bbox": [ + 132, + 114, + 482, + 148 + ], + "spans": [ + { + "bbox": [ + 132, + 114, + 482, + 148 + ], + "type": "text", + "content": "Table 2: Comparison Results of various metrics. The best results in our full dataset are highlighted in bold font, and the best results for each difficulty are marked in different colors." + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 134, + 157, + 480, + 258 + ], + "lines": [ + { + "bbox": [ + 134, + 157, + 480, + 258 + ], + "spans": [ + { + "bbox": [ + 134, + 157, + 480, + 258 + ], + "type": "table", + "html": "
MethodDifficultyRecall@5Recall@10Recall@20NDCG@5NDCG@10NDCG@20RR ↑HD ↓RE ↓NTE ↓
Rule-based [49]High0.2610.3540.4430.1850.2180.2410.107212.0031.28937.100×10-4
Medium0.3290.4270.5290.2420.2750.3020.153205.4631.24933.742×10-4
Low0.3580.4640.5550.2610.2960.320.188210.2741.23132.692×10-4
All0.3170.4160.5110.2480.2840.3110.149208.5191.25634.414×10-4
Jigsawnet [27]High0.3050.3940.4780.2310.2610.2840.117226.0261.30637.011×10-4
Medium0.3830.4780.5650.3090.340.3640.213206.0401.15531.467×10-4
Low0.4590.5480.6310.3680.3970.4190.241202.9421.18630.500×10-4
All0.380.4720.5570.3260.3590.3830.194210.8421.20532.772×10-4
OursHigh0.2760.4130.5350.1890.2340.2680.60692.2290.80718.335×10-4
Medium0.5010.6190.7320.3820.4230.4540.87935.6170.2365.685×10-4
Low0.7170.8050.8790.5810.610.6290.96112.3360.091.936×10-4
All0.4930.6080.7140.4170.4580.4870.83543.1160.3527.735×10-4
", + "image_path": "664d8e90f78f6d0a7151000f7f9ae616abb74fba9b0a0187423257a2c85ab855.jpg" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "table_body" + } + ], + "index": 3 + }, + { + "type": "image", + "bbox": [ + 136, + 271, + 479, + 406 + ], + "blocks": [ + { + "bbox": [ + 136, + 271, + 479, + 406 + ], + "lines": [ + { + "bbox": [ + 136, + 271, + 479, + 406 + ], + "spans": [ + { + "bbox": [ + 136, + 271, + 479, + 406 + ], + "type": "image", + "image_path": "7993877d11836b7c58a1b541d38db1fb231752dde2eb26188128b3850ca75c4a.jpg" + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 132, + 415, + 482, + 439 + ], + "lines": [ + { + "bbox": [ + 132, + 415, + 482, + 439 + ], + "spans": [ + { + "bbox": [ + 132, + 415, + 482, + 439 + ], + "type": "text", + "content": "Fig. 6: Examples of pair-searching results. Our network is able to accurately identify the corresponding target fragment (highlighted with a green box)." + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "image_caption" + } + ], + "index": 4 + }, + { + "bbox": [ + 130, + 464, + 482, + 585 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 464, + 482, + 585 + ], + "spans": [ + { + "bbox": [ + 130, + 464, + 482, + 585 + ], + "type": "text", + "content": "classification task; however, their pair-matching results are very unsatisfactory. It should be noted that we did not evaluate Recall@1 and NDCG@1 here due to the many-to-many pairings in our dataset, making it unreasonable to search only for the fragment with the highest similarity of searching. We can also observe that the performance of our proposed network shows reasonable variation with the difficulty of the data. This shows that our proposed network learns useful information and has more potential in complex situations. However, Jigsawnet and the rule-based method show relatively identical performance on test sets of different difficulty levels. Furthermore, we also plot the distributions of HD, RE, and NTE for each method for a more intuitive comparison, as shown in Figure 5." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 130, + 594, + 482, + 666 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 594, + 482, + 666 + ], + "spans": [ + { + "bbox": [ + 130, + 594, + 482, + 666 + ], + "type": "text", + "content": "Performance on Real Dataset. We also tested our method on the real dataset. In this experiment, both our method and the comparison method were trained on the generated dataset without retraining or fine-tuning. As shown in Table 3 and Figure 8, PairingNet obtained satisfactory results for both pair-searching and -matching tasks. These results confirmed that Algorithm 1 simulates real data well and our method has good generalization to real data." + } + ] + } + ], + "index": 7 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 402, + 91, + 447, + 102 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 402, + 91, + 447, + 102 + ], + "spans": [ + { + "bbox": [ + 402, + 91, + 447, + 102 + ], + "type": "text", + "content": "PairingNet" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 470, + 91, + 480, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 470, + 91, + 480, + 100 + ], + "spans": [ + { + "bbox": [ + 470, + 91, + 480, + 100 + ], + "type": "text", + "content": "11" + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 10 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 133, + 114, + 482, + 159 + ], + "blocks": [ + { + "bbox": [ + 133, + 114, + 482, + 159 + ], + "lines": [ + { + "bbox": [ + 133, + 114, + 482, + 159 + ], + "spans": [ + { + "bbox": [ + 133, + 114, + 482, + 159 + ], + "type": "image", + "image_path": "8d02581339b5ca491fca5e0f094c297ff8eceedbbce1bdab5c1f9989cd7af6e8.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 130, + 168, + 482, + 225 + ], + "lines": [ + { + "bbox": [ + 130, + 168, + 482, + 225 + ], + "spans": [ + { + "bbox": [ + 130, + 168, + 482, + 225 + ], + "type": "text", + "content": "Fig. 7: Examples of pair-matching results for three difficulty levels. Our proposed network achieves satisfactory matching results. However, for low-difficulty cases, Jigsawnet and the Rule-based method find approximate matching positions, while the matchings are not accurate enough. For medium- and high-difficulty cases, both comparison methods are difficult to obtain correct matching results." + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_caption" + } + ], + "index": 2 + }, + { + "type": "table", + "bbox": [ + 151, + 266, + 462, + 301 + ], + "blocks": [ + { + "bbox": [ + 131, + 234, + 482, + 258 + ], + "lines": [ + { + "bbox": [ + 131, + 234, + 482, + 258 + ], + "spans": [ + { + "bbox": [ + 131, + 234, + 482, + 258 + ], + "type": "text", + "content": "Table 3: Comparison Results of various metrics on our real dataset. The best results are highlighted in bold font." + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 151, + 266, + 462, + 301 + ], + "lines": [ + { + "bbox": [ + 151, + 266, + 462, + 301 + ], + "spans": [ + { + "bbox": [ + 151, + 266, + 462, + 301 + ], + "type": "table", + "html": "
MethodRecall@5Recall@10Recall@20NDCG@5NDCG@10NDCG@20RR↑HD↓RE↓NTE↓
Rule-based [49]0.1380.2470.3660.0790.1140.1440.1287.6981.61430.746×10-4
Jigsawnet [27]0.2370.3260.4250.1850.2130.2380.1537.5701.54529.061×10-4
Ours0.4150.5640.6780.3210.3690.3990.7524.1820.4809.075×10-4
", + "image_path": "34d55e7f1f2fd12735dda1f12ffc76364b47d7a395dee0887e9e9f61d78040fc.jpg" + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "table_body" + } + ], + "index": 5 + }, + { + "bbox": [ + 130, + 321, + 482, + 428 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 321, + 482, + 428 + ], + "spans": [ + { + "bbox": [ + 130, + 321, + 482, + 428 + ], + "type": "text", + "content": "Visualization. We show the top five search results of different methods for the examples of fragment pairs in Figure 6. Given a source fragment, our model can accurately identify the corresponding target fragment (marked with a green box). Furthermore, in situations where a source fragment can be paired with multiple target fragments, our method identifies all matching fragments within the top five search results. We also show examples of matching results for three difficulty levels in Figure 7. Our method produces matching results that are consistent with the Ground Truth (GT), while the comparison methods yield numerous incorrect results." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 130, + 429, + 482, + 572 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 429, + 482, + 572 + ], + "spans": [ + { + "bbox": [ + 130, + 429, + 482, + 572 + ], + "type": "text", + "content": "We also visualize the normalized self-gated weights to analyze their changes under different situations. As shown in Figure 9 (a), we map the normalized weights to the edges of the fragments to show the influence of the texture/contour features. We classify the local parts of fragments into four types based on the information contained in texture and contour: those with rich texture and contour information (first quadrant), those with rich texture information (second quadrant), those with little texture and contour information (third quadrant), and those with rich contour information but with little texture information (fourth quadrant). The red color indicates that the weights of the texture features are relatively large, while the green color indicates that the contour feature weights are relatively large. The results show that our proposed network learned varying weights of the features for fragments with different characteristics." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 130, + 576, + 482, + 624 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 576, + 482, + 624 + ], + "spans": [ + { + "bbox": [ + 130, + 576, + 482, + 624 + ], + "type": "text", + "content": "Inference Time. We test the inference time of different methods on the pairwise searching task on our generated test set. Our network retrieves all fragment pairs in 73 seconds, while the rule-based method takes 223 hours, and JigsawNet takes 10 hours." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 131, + 629, + 239, + 640 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 131, + 629, + 239, + 640 + ], + "spans": [ + { + "bbox": [ + 131, + 629, + 239, + 640 + ], + "type": "text", + "content": "5.4 Ablation Study" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 130, + 641, + 482, + 666 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 641, + 482, + 666 + ], + "spans": [ + { + "bbox": [ + 130, + 641, + 482, + 666 + ], + "type": "text", + "content": "Network Structures. We explore the performance of our network under different structures, as shown in Table 4. First, we test three feature fusion ways." + } + ] + } + ], + "index": 10 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "spans": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "type": "text", + "content": "12" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 166, + 91, + 212, + 100 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 166, + 91, + 212, + 100 + ], + "spans": [ + { + "bbox": [ + 166, + 91, + 212, + 100 + ], + "type": "text", + "content": "Zhou et al." + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 11 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 136, + 114, + 479, + 242 + ], + "blocks": [ + { + "bbox": [ + 136, + 114, + 479, + 242 + ], + "lines": [ + { + "bbox": [ + 136, + 114, + 479, + 242 + ], + "spans": [ + { + "bbox": [ + 136, + 114, + 479, + 242 + ], + "type": "image", + "image_path": "0886c5dd15063f233ed70d809b3cd7bfa9cdb31d25e8cdc060b4ee3ba9acc216.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 132, + 251, + 480, + 273 + ], + "lines": [ + { + "bbox": [ + 132, + 251, + 480, + 273 + ], + "spans": [ + { + "bbox": [ + 132, + 251, + 480, + 273 + ], + "type": "text", + "content": "Fig. 8: Examples of pair-matching results of PairingNet on real dataset. Without retraining, our method can also effectively match the scanned real fragments." + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_caption" + } + ], + "index": 2 + }, + { + "type": "image", + "bbox": [ + 214, + 284, + 320, + 365 + ], + "blocks": [ + { + "bbox": [ + 214, + 284, + 320, + 365 + ], + "lines": [ + { + "bbox": [ + 214, + 284, + 320, + 365 + ], + "spans": [ + { + "bbox": [ + 214, + 284, + 320, + 365 + ], + "type": "image", + "image_path": "e8bcd3996eef8dc74cd042054d0a5545b5cdfb3d6425de7b5b14d4af8a59fea9.jpg" + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 258, + 369, + 268, + 376 + ], + "lines": [ + { + "bbox": [ + 258, + 369, + 268, + 376 + ], + "spans": [ + { + "bbox": [ + 258, + 369, + 268, + 376 + ], + "type": "text", + "content": "(a)" + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "image_caption" + } + ], + "index": 4 + }, + { + "type": "image", + "bbox": [ + 320, + 284, + 400, + 368 + ], + "blocks": [ + { + "bbox": [ + 320, + 284, + 400, + 368 + ], + "lines": [ + { + "bbox": [ + 320, + 284, + 400, + 368 + ], + "spans": [ + { + "bbox": [ + 320, + 284, + 400, + 368 + ], + "type": "image", + "image_path": "795386562a2688255b455a2b03f0895e0874d88a3c5d56daabfab22a0ed06e38.jpg" + } + ] + } + ], + "index": 6, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 354, + 369, + 363, + 376 + ], + "lines": [ + { + "bbox": [ + 354, + 369, + 363, + 376 + ], + "spans": [ + { + "bbox": [ + 354, + 369, + 363, + 376 + ], + "type": "text", + "content": "(b)" + } + ] + } + ], + "index": 7, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 130, + 384, + 482, + 483 + ], + "lines": [ + { + "bbox": [ + 130, + 384, + 482, + 483 + ], + "spans": [ + { + "bbox": [ + 130, + 384, + 482, + 483 + ], + "type": "text", + "content": "Fig.9: Visual analysis of Texture/Contour. (a) We visualize the weights of texture and contour features in our network, with colors closer to red representing more use of texture features and closer to green representing more use of contour features. We can see that on edges with rich texture or close to straight lines, the weights of the extracted texture features are greater, while on edges with single texture, the weights of the contour features are greater. (b) Examples of Textur/Contour ablation study. When the local contour shapes are similar, our network has difficulty matching them correctly using only contour features, and vice versa. However, these cases can be solved when the two features are fused." + } + ] + } + ], + "index": 8, + "angle": 0, + "type": "image_caption" + } + ], + "index": 6 + }, + { + "bbox": [ + 130, + 506, + 482, + 579 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 506, + 482, + 579 + ], + "spans": [ + { + "bbox": [ + 130, + 506, + 482, + 579 + ], + "type": "text", + "content": "In early fusion, instead of using two separate GCNs for feature extraction, we concatenate the two modalities and use only one GCN to learn the features. We also test using a simple concatenation to replace the self-gated fusion module. The results show that the performance of both pair-searching and -matching decreases. Additionally, we explore the impact of different numbers of layers on our network." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 130, + 582, + 482, + 666 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 582, + 482, + 666 + ], + "spans": [ + { + "bbox": [ + 130, + 582, + 482, + 666 + ], + "type": "text", + "content": "Texture/Contour. To study the impact of contour and texture features on the performance of our network, we conducted experiments using only one modality (either contour or texture) and compared the final results. As shown in Table 4, the comparison results indicate that both modalities perform slightly worse when used independently. When both modalities are used in conjunction, our network achieves the best performance. We also show examples of failed pair-matching when the two modalities are used individually in Figure 9 (b)." + } + ] + } + ], + "index": 10 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 402, + 91, + 447, + 102 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 402, + 91, + 447, + 102 + ], + "spans": [ + { + "bbox": [ + 402, + 91, + 447, + 102 + ], + "type": "text", + "content": "PairingNet" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 470, + 91, + 481, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 470, + 91, + 481, + 100 + ], + "spans": [ + { + "bbox": [ + 470, + 91, + 481, + 100 + ], + "type": "text", + "content": "13" + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 12 + }, + { + "para_blocks": [ + { + "type": "table", + "bbox": [ + 217, + 157, + 396, + 242 + ], + "blocks": [ + { + "bbox": [ + 132, + 114, + 482, + 148 + ], + "lines": [ + { + "bbox": [ + 132, + 114, + 482, + 148 + ], + "spans": [ + { + "bbox": [ + 132, + 114, + 482, + 148 + ], + "type": "text", + "content": "Table 4: Results of ablation studies. Comparison of the impact of different feature fusion ways, feature modalities, and network layers employed in our proposed network on the performance of pair-searching and pair-matching tasks." + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 217, + 157, + 396, + 242 + ], + "lines": [ + { + "bbox": [ + 217, + 157, + 396, + 242 + ], + "spans": [ + { + "bbox": [ + 217, + 157, + 396, + 242 + ], + "type": "table", + "html": "
Recall@5Recall@10NDCG@5NDCG@10RR
Fusion WayEarly fusion0.4480.5590.3760.4130.658
Concat0.4890.6050.4180.4580.824
Weighted0.4930.6080.4170.4580.835
ResGCN10----0.680
14----0.835
18----0.759
Transformer Encoder30.4880.6000.4100.450-
40.4880.6060.4110.452-
50.4930.6080.4170.458-
Feature ModalityContour-only0.0160.0250.0130.0160.584
Texture-only0.4790.5970.3940.4360.657
Contour + Texture0.4930.6080.4170.4580.835
", + "image_path": "17bd2d05a4f6b932881cf253f5b29b135a1aa47ccf7b46b64e85ac0dbf3065c1.jpg" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "table_body" + } + ], + "index": 3 + }, + { + "bbox": [ + 132, + 264, + 480, + 312 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 264, + 480, + 312 + ], + "spans": [ + { + "bbox": [ + 132, + 264, + 480, + 312 + ], + "type": "text", + "content": "Besides, as shown in Figure 9 (a), it can be seen numerically that our network tends to assign more weight to the texture features. This is reasonable because textures usually contain richer information than the contours of image fragments, enabling the network to learn more discriminative features." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 132, + 316, + 480, + 339 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 316, + 480, + 339 + ], + "spans": [ + { + "bbox": [ + 132, + 316, + 480, + 339 + ], + "type": "text", + "content": "Others. We report the results of ablation studies on patch size, contour encoding, and hyperparameters in our supplementary material." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 132, + 353, + 225, + 364 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 353, + 225, + 364 + ], + "spans": [ + { + "bbox": [ + 132, + 353, + 225, + 364 + ], + "type": "text", + "content": "6 Conclusions" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 132, + 366, + 481, + 557 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 366, + 481, + 557 + ], + "spans": [ + { + "bbox": [ + 132, + 366, + 481, + 557 + ], + "type": "text", + "content": "In this paper, to tackle the image fragment pair-searching and -matching tasks, we propose a novel network, PairingNet, to deduce the adjacent segments of fragments and encode the global features of each fragment. In PairingNet, ResGCN is employed as the backbone to extract the local contour and texture features of fragments. For pair-searching, a linear transformer-based module processes these local features leveraging the contrast of paired fragments and unpaired fragments. For pair-matching, we design a weighted fusion module to dynamically fuse the extracted local contour and texture features. Furthermore, to provide a foundation for learning-based methods, we design a cutting algorithm to generate a dataset by simulating real image fragments. Comprehensive experiments are conducted on our generated dataset and a real dataset, the results of which demonstrate the effectiveness of our PairingNet compared to existing methods. In future work, we will explore how to extend our work to other general technique, e.g. puzzles inspired MAE [20], JiGen [6] for self-supervised learning and domain generalization. The fragment generation algorithm that we designed can also create more diverse samples for image inpainting tasks [29,31]." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 132, + 558, + 481, + 629 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 558, + 481, + 629 + ], + "spans": [ + { + "bbox": [ + 132, + 558, + 481, + 629 + ], + "type": "text", + "content": "Limitations While our research has made significant strides, fragment restoration remains a challenging task. Our dataset provides superior edge matching features. The task of fragment restoration becomes particularly arduous when the edges of the fragments are smooth and the texture is monotonous (for instance, in the case of shredded white paper), or when there is a substantial gap in the damaged edges." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 132, + 630, + 481, + 665 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 630, + 481, + 665 + ], + "spans": [ + { + "bbox": [ + 132, + 630, + 481, + 665 + ], + "type": "text", + "content": "Acknowledgements This work was supported by the Young Scientists Fund of the National Natural Science Foundation of China (Grant No.62206106). We also feel grateful to Yifan Jin for his exploration in the preliminary work." + } + ] + } + ], + "index": 9 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "spans": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "type": "text", + "content": "14" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 166, + 91, + 211, + 100 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 166, + 91, + 211, + 100 + ], + "spans": [ + { + "bbox": [ + 166, + 91, + 211, + 100 + ], + "type": "text", + "content": "Zhou et al." + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 13 + }, + { + "para_blocks": [ + { + "bbox": [ + 133, + 114, + 197, + 126 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 114, + 197, + 126 + ], + "spans": [ + { + "bbox": [ + 133, + 114, + 197, + 126 + ], + "type": "text", + "content": "References" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 138, + 140, + 480, + 666 + ], + "type": "list", + "angle": 0, + "index": 20, + "blocks": [ + { + "bbox": [ + 138, + 140, + 480, + 172 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 140, + 480, + 172 + ], + "spans": [ + { + "bbox": [ + 138, + 140, + 480, + 172 + ], + "type": "text", + "content": "1. Abitbol, R., Shimshoni, I., Ben-Dov, J.: Machine learning based assembly of fragments of ancient papyrus. Journal on Computing and Cultural Heritage (JOCCH) 14(3), 1-21 (2021)" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 138, + 173, + 480, + 205 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 173, + 480, + 205 + ], + "spans": [ + { + "bbox": [ + 138, + 173, + 480, + 205 + ], + "type": "text", + "content": "2. Bridger, D., Danon, D., Tal, A.: Solving jigsaw puzzles with eroded boundaries. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 3526-3535 (2020)" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 138, + 206, + 480, + 227 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 206, + 480, + 227 + ], + "spans": [ + { + "bbox": [ + 138, + 206, + 480, + 227 + ], + "type": "text", + "content": "3. Brown, L.G.: A survey of image registration techniques. ACM computing surveys (CSUR) 24(4), 325-376 (1992)" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 138, + 228, + 441, + 239 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 228, + 441, + 239 + ], + "spans": [ + { + "bbox": [ + 138, + 228, + 441, + 239 + ], + "type": "text", + "content": "4. Bruno Joseph, I.J., Frese, D.: Pexels. https://www.pexels.com/ (2023)" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 138, + 239, + 480, + 282 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 239, + 480, + 282 + ], + "spans": [ + { + "bbox": [ + 138, + 239, + 480, + 282 + ], + "type": "text", + "content": "5. Busa-Fekete, R., Szarvas, G., Elteto, T., Kégl, B.: An apple-to-apple comparison of learning-to-rank algorithms in terms of normalized discounted cumulative gain. In: ECAI 2012-20th European Conference on Artificial Intelligence: Preference Learning: Problems and Applications in AI Workshop. vol. 242. IOS Press (2012)" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 138, + 283, + 480, + 315 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 283, + 480, + 315 + ], + "spans": [ + { + "bbox": [ + 138, + 283, + 480, + 315 + ], + "type": "text", + "content": "6. Carlucci, F.M., D'Innocente, A., Bucci, S., Caputo, B., Tommasi, T.: Domain generalization by solving jigsaw puzzles. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 2229-2238 (2019)" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 138, + 316, + 480, + 348 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 316, + 480, + 348 + ], + "spans": [ + { + "bbox": [ + 138, + 316, + 480, + 348 + ], + "type": "text", + "content": "7. Chen, J., Tian, M., Qi, X., Wang, W., Liu, Y.: A solution to reconstruct cross-cut shredded text documents based on constrained seed k-means algorithm and ant colony algorithm. Expert Systems with Applications 127(AUG.), 35-46 (2019)" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 138, + 348, + 480, + 392 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 348, + 480, + 392 + ], + "spans": [ + { + "bbox": [ + 138, + 348, + 480, + 392 + ], + "type": "text", + "content": "8. Chen, Y.C., Li, H., Turpin, D., Jacobson, A., Garg, A.: Neural shape mating: Self-supervised object assembly with adversarial shape priors. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 12724-12733 (2022)" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 138, + 392, + 480, + 424 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 392, + 480, + 424 + ], + "spans": [ + { + "bbox": [ + 138, + 392, + 480, + 424 + ], + "type": "text", + "content": "9. Cheng, X., Lin, H., Wu, X., Yang, F., Shen, D.: Improving video-text retrieval by multi-stream corpus alignment and dual softmax loss. arXiv preprint arXiv:2109.04290 (2021)" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 138, + 425, + 480, + 457 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 425, + 480, + 457 + ], + "spans": [ + { + "bbox": [ + 138, + 425, + 480, + 457 + ], + "type": "text", + "content": "0. Cho, T.S., Avidan, S., Freeman, W.T.: A probabilistic image jigsaw puzzle solver. In: 2010 IEEE Computer society conference on computer vision and pattern recognition. pp. 183-190. IEEE (2010)" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 138, + 458, + 480, + 491 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 458, + 480, + 491 + ], + "spans": [ + { + "bbox": [ + 138, + 458, + 480, + 491 + ], + "type": "text", + "content": "1. Dai, Y., Gieseke, F., Oehmcke, S., Wu, Y., Barnard, K.: Attentional feature fusion. In: Proceedings of the IEEE/CVF winter conference on applications of computer vision. pp. 3560-3569 (2021)" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 138, + 491, + 480, + 512 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 491, + 480, + 512 + ], + "spans": [ + { + "bbox": [ + 138, + 491, + 480, + 512 + ], + "type": "text", + "content": "2. Derech, N., Tal, A., Shimshoni, I.: Solving archaeological puzzles. Pattern Recognition 119, 108065 (2021)" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 138, + 512, + 480, + 544 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 512, + 480, + 544 + ], + "spans": [ + { + "bbox": [ + 138, + 512, + 480, + 544 + ], + "type": "text", + "content": "3. Fischler, M.A., Bolles, R.C.: Random sample consensus: a paradigm for model fitting with applications to image analysis and automated cartography. Communications of the ACM 24(6), 381-395 (1981)" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 138, + 545, + 480, + 578 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 545, + 480, + 578 + ], + "spans": [ + { + "bbox": [ + 138, + 545, + 480, + 578 + ], + "type": "text", + "content": "4. Freeman, H., Gardner, L.: Apictorial jigsaw puzzles: The computer solution of a problem in pattern recognition. IEEE Transactions on Electronic Computers (2), 118-127 (1964)" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 138, + 578, + 480, + 611 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 578, + 480, + 611 + ], + "spans": [ + { + "bbox": [ + 138, + 578, + 480, + 611 + ], + "type": "text", + "content": "5. Funkhouser, T., Shin, H., Toler-Franklin, C., Castañeda, A.G., Brown, B., Dobkin, D., Rusinkiewicz, S., Weyrich, T.: Learning how to match fresco fragments. Journal on Computing and Cultural Heritage (JOCCH) 4(2), 1-13 (2011)" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 138, + 611, + 480, + 632 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 611, + 480, + 632 + ], + "spans": [ + { + "bbox": [ + 138, + 611, + 480, + 632 + ], + "type": "text", + "content": "6. Gallagher, A.C.: Jigsaw puzzles with pieces of unknown orientation. In: 2012 IEEE Conference on computer vision and pattern recognition. pp. 382-389. IEEE (2012)" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 138, + 632, + 480, + 666 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 632, + 480, + 666 + ], + "spans": [ + { + "bbox": [ + 138, + 632, + 480, + 666 + ], + "type": "text", + "content": "7. da Gama Leitao, H.C., Stolfi, J.: A multiscale method for the reassembly of two-dimensional fragmented objects. IEEE Transactions on Pattern Analysis and Machine Intelligence 24(9), 1239-1251 (2002)" + } + ] + } + ], + "index": 19 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 402, + 91, + 447, + 102 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 402, + 91, + 447, + 102 + ], + "spans": [ + { + "bbox": [ + 402, + 91, + 447, + 102 + ], + "type": "text", + "content": "PairingNet" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 470, + 91, + 481, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 470, + 91, + 481, + 100 + ], + "spans": [ + { + "bbox": [ + 470, + 91, + 481, + 100 + ], + "type": "text", + "content": "15" + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 14 + }, + { + "para_blocks": [ + { + "bbox": [ + 132, + 117, + 481, + 666 + ], + "type": "list", + "angle": 0, + "index": 19, + "blocks": [ + { + "bbox": [ + 133, + 117, + 481, + 149 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 117, + 481, + 149 + ], + "spans": [ + { + "bbox": [ + 133, + 117, + 481, + 149 + ], + "type": "text", + "content": "18. Guo, Y., Bennamoun, M., Sohel, F., Lu, M., Wan, J., Kwok, N.M.: A comprehensive performance evaluation of 3d local feature descriptors. International Journal of Computer Vision 116, 66-89 (2016)" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 133, + 150, + 481, + 183 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 150, + 481, + 183 + ], + "spans": [ + { + "bbox": [ + 133, + 150, + 481, + 183 + ], + "type": "text", + "content": "19. Gur, S., Ben-Shahar, O.: From square pieces to brick walls: The next challenge in solving jigsaw puzzles. In: Proceedings of the IEEE International Conference on Computer Vision. pp. 4029-4037 (2017)" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 132, + 183, + 481, + 216 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 183, + 481, + 216 + ], + "spans": [ + { + "bbox": [ + 132, + 183, + 481, + 216 + ], + "type": "text", + "content": "20. He, K., Chen, X., Xie, S., Li, Y., Dollar, P., Girshick, R.: Masked autoencoders are scalable vision learners. In: Proceedings of the IEEE/CVF conference on computer vision and pattern recognition. pp. 16000-16009 (2022)" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 132, + 216, + 481, + 259 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 216, + 481, + 259 + ], + "spans": [ + { + "bbox": [ + 132, + 216, + 481, + 259 + ], + "type": "text", + "content": "21. Hong, J.H., Yoo, S.J., Zeeshan, M.A., Kim, Y.M., Kim, J.: Structure-from-sherds: Incremental 3d reassembly of axially symmetric pots from unordered and mixed fragment collections. In: Proceedings of the IEEE/CVF International Conference on Computer Vision. pp. 5443-5451 (2021)" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 132, + 259, + 481, + 293 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 259, + 481, + 293 + ], + "spans": [ + { + "bbox": [ + 132, + 259, + 481, + 293 + ], + "type": "text", + "content": "22. Huang, H., Yin, K., Gong, M., Lischinski, D., Cohen-Or, D., Ascher, U.M., Chen, B.: \"mind the gap\": tele-registration for structure-driven image completion. ACM Trans. Graph. 32(6), 174-1 (2013)" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 132, + 293, + 481, + 326 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 293, + 481, + 326 + ], + "spans": [ + { + "bbox": [ + 132, + 293, + 481, + 326 + ], + "type": "text", + "content": "23. Huang, Q.X., Flóry, S., Gelfand, N., Hofer, M., Pottmann, H.: Reassembling fractured objects by geometric matching. In: ACM SIGGRAPH 2006 Papers, pp. 569-578 (2006)" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 132, + 326, + 481, + 358 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 326, + 481, + 358 + ], + "spans": [ + { + "bbox": [ + 132, + 326, + 481, + 358 + ], + "type": "text", + "content": "24. Jones, B., Hildreth, D., Chen, D., Baran, I., Kim, V.G., Schulz, A.: Automate: A dataset and learning approach for automatic mating of cad assemblies. ACM Transactions on Graphics (TOG) 40(6), 1-18 (2021)" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 132, + 358, + 481, + 392 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 358, + 481, + 392 + ], + "spans": [ + { + "bbox": [ + 132, + 358, + 481, + 392 + ], + "type": "text", + "content": "25. Kong, W., Kimia, B.B.: On solving 2d and 3d puzzles using curve matching. In: Proceedings of the 2001 IEEE Computer Society Conference on Computer Vision and Pattern Recognition. CVPR 2001. vol. 2, pp. II-II. IEEE (2001)" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 132, + 392, + 481, + 435 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 392, + 481, + 435 + ], + "spans": [ + { + "bbox": [ + 132, + 392, + 481, + 435 + ], + "type": "text", + "content": "26. Lamb, N., Palmer, C., Molloy, B., Banerjee, S., Banerjee, N.K.: Fantastic breaks: A dataset of paired 3d scans of real-world broken objects and their complete counterparts. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 4681-4691 (2023)" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 132, + 435, + 481, + 468 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 435, + 481, + 468 + ], + "spans": [ + { + "bbox": [ + 132, + 435, + 481, + 468 + ], + "type": "text", + "content": "27. Le, C., Li, X.: Jigsawnet: Shredded image reassembly using convolutional neural network and loop-based composition. IEEE Transactions on Image Processing 28(8), 4000-4015 (2019)" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 132, + 468, + 481, + 501 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 468, + 481, + 501 + ], + "spans": [ + { + "bbox": [ + 132, + 468, + 481, + 501 + ], + "type": "text", + "content": "28. Li, G., Muller, M., Thabet, A., Ghanem, B.: Deep GCs: Can GCs go as deep as cnns? In: Proceedings of the IEEE/CVF international conference on computer vision. pp. 9267-9276 (2019)" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 132, + 501, + 481, + 534 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 501, + 481, + 534 + ], + "spans": [ + { + "bbox": [ + 132, + 501, + 481, + 534 + ], + "type": "text", + "content": "29. Li, W., Lin, Z., Zhou, K., Qi, L., Wang, Y., Jia, J.: Mat: Mask-aware transformer for large hole image inpainting. In: Proceedings of the IEEE/CVF conference on computer vision and pattern recognition. pp. 10758-10768 (2022)" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 132, + 534, + 481, + 567 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 534, + 481, + 567 + ], + "spans": [ + { + "bbox": [ + 132, + 534, + 481, + 567 + ], + "type": "text", + "content": "30. Li, Y., Harada, T.: Lepard: Learning partial point cloud matching in rigid and deformable scenes. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 5554-5564 (2022)" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 132, + 567, + 481, + 600 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 567, + 481, + 600 + ], + "spans": [ + { + "bbox": [ + 132, + 567, + 481, + 600 + ], + "type": "text", + "content": "31. Liu, G., Reda, F.A., Shih, K.J., Wang, T.C., Tao, A., Catanzaro, B.: Image inpainting for irregular holes using partial convolutions. In: Proceedings of the European conference on computer vision (ECCV). pp. 85-100 (2018)" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 132, + 600, + 481, + 622 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 600, + 481, + 622 + ], + "spans": [ + { + "bbox": [ + 132, + 600, + 481, + 622 + ], + "type": "text", + "content": "32. Liu, H., Cao, S., Yan, S.: Automated assembly of shredded pieces from multiple photos. IEEE transactions on multimedia 13(5), 1154-1162 (2011)" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 132, + 622, + 481, + 644 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 622, + 481, + 644 + ], + "spans": [ + { + "bbox": [ + 132, + 622, + 481, + 644 + ], + "type": "text", + "content": "33. Mikolajczyk, K., Schmid, C.: A performance evaluation of local descriptors. IEEE transactions on pattern analysis and machine intelligence 27(10), 1615-1630 (2005)" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 132, + 644, + 481, + 666 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 644, + 481, + 666 + ], + "spans": [ + { + "bbox": [ + 132, + 644, + 481, + 666 + ], + "type": "text", + "content": "34. Oord, A.v.d., Li, Y., Vinyals, O.: Representation learning with contrastive predictive coding. arXiv preprint arXiv:1807.03748 (2018)" + } + ] + } + ], + "index": 18 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "spans": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "type": "text", + "content": "16" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 166, + 91, + 212, + 101 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 166, + 91, + 212, + 101 + ], + "spans": [ + { + "bbox": [ + 166, + 91, + 212, + 101 + ], + "type": "text", + "content": "Zhou et al." + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 15 + }, + { + "para_blocks": [ + { + "bbox": [ + 130, + 116, + 482, + 544 + ], + "type": "list", + "angle": 0, + "index": 17, + "blocks": [ + { + "bbox": [ + 130, + 116, + 482, + 149 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 116, + 482, + 149 + ], + "spans": [ + { + "bbox": [ + 130, + 116, + 482, + 149 + ], + "type": "text", + "content": "35. Paixao, T.M., Boeres, M., Freitas, C., Oliveira-Santos, T.: Exploring character shapes for unsupervised reconstruction of strip-shredded text documents. IEEE Transactions on Information Forensics and Security pp. 1-1 (2018)" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 132, + 150, + 481, + 183 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 150, + 481, + 183 + ], + "spans": [ + { + "bbox": [ + 132, + 150, + 481, + 183 + ], + "type": "text", + "content": "36. Paumard, M.M., Picard, D., Tabia, H.: Image reassembly combining deep learning and shortest path problem. In: Proceedings of the European conference on computer vision (ECCV). pp. 153-167 (2018)" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 132, + 183, + 481, + 205 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 183, + 481, + 205 + ], + "spans": [ + { + "bbox": [ + 132, + 183, + 481, + 205 + ], + "type": "text", + "content": "37. Pomeranz, D., Shemesh, M., Ben-Shahar, O.: A fully automated greedy square jigsaw puzzle solver. In: CVPR 2011. pp. 9-16. IEEE (2011)" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 132, + 205, + 481, + 225 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 205, + 481, + 225 + ], + "spans": [ + { + "bbox": [ + 132, + 205, + 481, + 225 + ], + "type": "text", + "content": "38. Richter, F., Ries, C.X., Cebron, N., Lienhart, R.: Learning to reassemble shredded documents. IEEE Transactions on Multimedia 15(3), 582-593 (2013)" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 132, + 227, + 481, + 258 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 227, + 481, + 258 + ], + "spans": [ + { + "bbox": [ + 132, + 227, + 481, + 258 + ], + "type": "text", + "content": "39. Rocco, I., Cimpoi, M., Arandjelovic, R., Torii, A., Pajdla, T., Sivic, J.: Neighbourhood consensus networks. Advances in neural information processing systems 31 (2018)" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 132, + 259, + 481, + 293 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 259, + 481, + 293 + ], + "spans": [ + { + "bbox": [ + 132, + 259, + 481, + 293 + ], + "type": "text", + "content": "40. Sellan, S., Chen, Y.C., Wu, Z., Garg, A., Jacobson, A.: Breaking bad: A dataset for geometric fracture and reassembly. Advances in Neural Information Processing Systems 35, 38885-38898 (2022)" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 132, + 293, + 481, + 314 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 293, + 481, + 314 + ], + "spans": [ + { + "bbox": [ + 132, + 293, + 481, + 314 + ], + "type": "text", + "content": "41. Son, K., Hays, J., Cooper, D.B.: Solving square jigsaw puzzles with loop constraints. In: European conference on computer vision. pp. 32-46. Springer (2014)" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 132, + 315, + 481, + 346 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 315, + 481, + 346 + ], + "spans": [ + { + "bbox": [ + 132, + 315, + 481, + 346 + ], + "type": "text", + "content": "42. Son, K., Hays, J., Cooper, D.B.: Solving square jigsaw puzzle by hierarchical loop constraints. IEEE transactions on pattern analysis and machine intelligence 41(9), 2222-2235 (2018)" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 132, + 347, + 481, + 369 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 347, + 481, + 369 + ], + "spans": [ + { + "bbox": [ + 132, + 347, + 481, + 369 + ], + "type": "text", + "content": "43. Tsamoura, E., Pitas, I.: Automatic color based reassembly of fragmented images and paintings. IEEE Transactions on Image Processing 19(3), 680-690 (2009)" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 132, + 369, + 481, + 391 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 369, + 481, + 391 + ], + "spans": [ + { + "bbox": [ + 132, + 369, + 481, + 391 + ], + "type": "text", + "content": "44. Wang, S., Li, B.Z., Khabsa, M., Fang, H., Ma, H.: Linformer: Self-attention with linear complexity. arXiv preprint arXiv:2006.04768 (2020)" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 132, + 392, + 481, + 423 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 392, + 481, + 423 + ], + "spans": [ + { + "bbox": [ + 132, + 392, + 481, + 423 + ], + "type": "text", + "content": "45. Wu, R., Tie, C., Du, Y., Zhao, Y., Dong, H.: Leveraging se (3) equivariance for learning 3d geometric shape assembly. In: Proceedings of the IEEE/CVF International Conference on Computer Vision. pp. 14311-14320 (2023)" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 132, + 424, + 481, + 456 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 424, + 481, + 456 + ], + "spans": [ + { + "bbox": [ + 132, + 424, + 481, + 456 + ], + "type": "text", + "content": "46. Yang, X., Matsuyama, K., Konno, K.: Pairwise matching of stone tools based on flake-surface contour points and normals. In: Proceedings of the Eurographics Workshop on Graphics and Cultural Heritage. pp. 125-129 (2017)" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 132, + 457, + 481, + 489 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 457, + 481, + 489 + ], + "spans": [ + { + "bbox": [ + 132, + 457, + 481, + 489 + ], + "type": "text", + "content": "47. Ye, S., Chen, Z., Chu, X., Li, K., Luo, J., Li, Y., Geng, G., Wu, Y.: Puzzlefixer: A visual reassembly system for immersive fragments restoration. IEEE transactions on visualization and computer graphics 29(1), 429-439 (2022)" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 132, + 490, + 481, + 522 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 490, + 481, + 522 + ], + "spans": [ + { + "bbox": [ + 132, + 490, + 481, + 522 + ], + "type": "text", + "content": "48. Zhan, G., Fan, Q., Mo, K., Shao, L., Chen, B., Guibas, L.J., Dong, H., et al.: Generative 3d part assembly via dynamic graph learning. Advances in Neural Information Processing Systems 33, 6315-6326 (2020)" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 132, + 523, + 481, + 544 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 523, + 481, + 544 + ], + "spans": [ + { + "bbox": [ + 132, + 523, + 481, + 544 + ], + "type": "text", + "content": "49. Zhang, K., Li, X.: A graph-based optimization algorithm for fragmented image reassembly. Graphical Models 76(5), 484-495 (2014)" + } + ] + } + ], + "index": 16 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 402, + 91, + 447, + 102 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 402, + 91, + 447, + 102 + ], + "spans": [ + { + "bbox": [ + 402, + 91, + 447, + 102 + ], + "type": "text", + "content": "PairingNet" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 470, + 91, + 481, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 470, + 91, + 481, + 100 + ], + "spans": [ + { + "bbox": [ + 470, + 91, + 481, + 100 + ], + "type": "text", + "content": "17" + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 16 + } + ], + "_backend": "vlm", + "_version_name": "2.6.4" +} \ No newline at end of file diff --git a/2024/Pairwise Distance Distillation for Unsupervised Real-World Image Super-Resolution/821e13b2-5f5a-4742-b51d-28c69e766f13_content_list.json b/2024/Pairwise Distance Distillation for Unsupervised Real-World Image Super-Resolution/821e13b2-5f5a-4742-b51d-28c69e766f13_content_list.json new file mode 100644 index 0000000000000000000000000000000000000000..15b4100fc5cf6a36b718a1afec0c646ed7b00985 --- /dev/null +++ b/2024/Pairwise Distance Distillation for Unsupervised Real-World Image Super-Resolution/821e13b2-5f5a-4742-b51d-28c69e766f13_content_list.json @@ -0,0 +1,1991 @@ +[ + { + "type": "text", + "text": "Pairwise Distance Distillation for Unsupervised Real-World Image Super-Resolution", + "text_level": 1, + "bbox": [ + 230, + 140, + 772, + 186 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Yuehan Zhang $^{1}$ , Seungjun Lee $^{2}$ , and Angela Yao $^{1}$", + "bbox": [ + 305, + 212, + 697, + 228 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "$^{1}$ National University of Singapore", + "bbox": [ + 383, + 239, + 617, + 253 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "{zyuehan,ayao}@comp.nus.edu.sg", + "bbox": [ + 383, + 253, + 617, + 268 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "2 Korea University", + "bbox": [ + 436, + 268, + 565, + 281 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "9penguin9@korea.ac.kr", + "bbox": [ + 419, + 282, + 583, + 295 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Abstract. Standard single-image super-resolution creates paired training data from high-resolution images through fixed downsampling kernels. However, real-world super-resolution (RWSR) faces unknown degradations in the low-resolution inputs, all the while lacking paired training data. Existing methods approach this problem by learning blind general models through complex synthetic augmentations on training inputs; they sacrifice the performance on specific degradation for broader generalization to many possible ones. We address the unsupervised RWSR for a targeted real-world degradation. We study from a distillation perspective and introduce a novel pairwise distance distillation framework. Through our framework, a model specialized in synthetic degradation adapts to target real-world degradations by distilling intra- and intermodel distances across the specialized model and an auxiliary generalized model. Experiments on diverse datasets demonstrate that our method significantly enhances fidelity and perceptual quality, surpassing state-of-the-art approaches in RWSR. The source code is available at https://github.com/Yuehan717/PDD.", + "bbox": [ + 261, + 334, + 738, + 569 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "1 Introduction", + "text_level": 1, + "bbox": [ + 217, + 595, + 374, + 611 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Single-image super-resolution (SISR) predicts high-resolution (HR) images from low-resolution (LR) counterparts. The standard SISR model addresses predefined downsampling kernels, e.g. bicubic interpolation. However, real-world scenarios of SISR (RWSR) encompass unknown degradations in LR images, including blur, noise, and JPEG compression artifacts [35,43,45] with diverse combinations. The various and unknown degradations pose additional challenges when learning an RWSR model.", + "bbox": [ + 212, + 628, + 784, + 733 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Existing RWSR methods focus on blind generalization. Such methods synthesize paired training data with extensive and harsh degradations, involving multiple rounds of random blurring, added noise, resizing, and compression [35, 43]. The complex degradation pipeline allows the model to generalize to diverse unknown conditions. We refer to such a model addressing various degradations as a generalist. However, as much of the model's capacity is devoted to handling multiple conditions, these \"jack-of-all-trades\" models come with performance trade-", + "bbox": [ + 212, + 734, + 785, + 839 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/1a2f83a7888a9039e41899decc9bde8830ec36ac79d7e346fe43ee3b272ab5a5.jpg", + "image_caption": [ + "Fig. 1: $\\hat{Y}_G^U$ and $\\hat{Y}_G^L$ are reconstructions for real-world and bicubic-interpolated (BI) inputs using a blind generalized model (generalist); $\\hat{Y}_S^U$ and $\\hat{Y}_S^L$ are counterparts using a standard SR model (bicubic specialist). The specialist enhances the BI inputs with clearer details, while the generalist does better for the real-world one. We distill the intra- and inter-model distances for an improved real-world reconstruction $\\hat{Y}_{S^{*}}^{U}$ ." + ], + "image_footnote": [], + "bbox": [ + 261, + 142, + 736, + 258 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "offs [45], i.e., they have inferior performance compared to the model only optimized for the tested degradation $(\\hat{Y}_G^L$ and $\\hat{Y}_S^L$ in Fig. 1).", + "bbox": [ + 212, + 375, + 782, + 407 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "SR models can be optimized for specific and known input degradations through supervised learning [9, 11, 18, 36]. The LR-HR pairs for training are created by degrading HR images with fixed kernels, e.g., bicubic interpolation, and Gaussian blur. We refer to models adept at the specific input degradation as a specialist. Degradations in RWSR datasets [2, 39] always belong to a specific domain [45]. For example, images captured by a specific camera model tend to feature consistent sensor noise [2]. However, the formulation of a real-world degradation is always unknown, posing difficulties in creating paired training data. As such, learning specialist models for such real-world domains is challenging, and strategies that do not require direct supervision are crucial.", + "bbox": [ + 212, + 409, + 787, + 559 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "This paper tackles unsupervised RWSR from a knowledge distillation perspective. Distillation has been successful for domain adaptation in high-level tasks [6, 22, 38], but it is under-explored in low-level vision. Our distillation framework adapts a specialist model of known synthetic degradation to target real-world degradations. We use an auxiliary generalist model to provide generalization knowledge of real-world degradations that the specialist lacks. However, naive distillation only transfers knowledge from the generalist model; we aim to improve beyond the generalist's moderate performance. To that end, we also consider the low-level characteristics of the specialist's predictions on synthetic samples as a reference for the distillation.", + "bbox": [ + 212, + 564, + 787, + 715 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Our method uses the strength of both the specialist and the generalist by exploring the relationships between their predictions. Given a generalist and a specialist for synthetic degradation, we consider predictions of synthetic samples and real-world samples from both models (see Fig. 1). The generalist model has smaller gaps in qualities of real-world and synthetic predictions, featuring close low-level characteristics between domains [20,21]. In contrast, the specialist outputs distanced low-level characteristics but high-quality predictions for synthetic samples $(\\hat{Y}_S^L$ in Fig. 1), providing valuable low-level characteristics for reference.", + "bbox": [ + 212, + 719, + 787, + 842 + ], + "page_idx": 1 + }, + { + "type": "page_number", + "text": "2", + "bbox": [ + 217, + 114, + 228, + 126 + ], + "page_idx": 1 + }, + { + "type": "header", + "text": "Y. Zhang et al.", + "bbox": [ + 271, + 114, + 374, + 128 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "We propose to let the specialist imitate prediction relationships from the generalist - that is, push the specialist's real-world predictions $(\\hat{Y}_S^U$ in Fig. 1) to have low-level characteristics similar to its high-quality synthetic predictions. If such similarity is satisfied by the specialist, it would exhibit two consistencies. First, for two input samples, the difference between their predictions from the same model is intra-model distance; such distances would be consistent for the specialist and the generalist. Secondly, for a given input sample, the difference between generalist and specialist predictions is inter-model distance; such distances would be consistent whether the sample is synthetic or real-world. These intra- and inter-model distance consistencies form the basis of our pairwise distance distillation framework. Together with designs on configurations of specialist and generalist models, our adapted specialist model experimentally shows significant improvements over the generalist model.", + "bbox": [ + 212, + 146, + 787, + 340 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Contributions. To summarize, the contributions of our work are three-fold:", + "bbox": [ + 214, + 343, + 771, + 357 + ], + "page_idx": 2 + }, + { + "type": "list", + "sub_type": "text", + "list_items": [ + "- To the best of our knowledge, we are the first to propose a distillation perspective for unsupervised RWSR to combine generalist and specialist models.", + "- We propose a novel pairwise distance distillation framework, emphasizing transferring intra-model and inter-model distances to enhance the specialist's performance in real-world scenarios.", + "- Experiments on three benchmarks demonstrate that our approach improves the off-the-shelf models regarding fidelity and perceptual performances." + ], + "bbox": [ + 225, + 369, + 784, + 474 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "2 Related Works", + "text_level": 1, + "bbox": [ + 215, + 502, + 395, + 517 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Single-Image Super-Resolution (SISR) methods use deep neural networks of various architectures, such as residual networks [19,36] and transformers [18, 33]. Despite improving the standard SISR performance with dedicated architectures, these methods struggle to generalize to real-world scenarios [34,40].", + "bbox": [ + 212, + 536, + 784, + 595 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Real-World Image Super-Resolution (RWSR) aims to address unknown input degradations. The infeasibility of creating paired training data without known kernels presents a significant challenge in learning specialized models for RWSR. There are two primary strategies in existing RWSR. The first assumes that networks trained on diverse and challenging synthetic degradations will effectively generalize to real-world data [35, 43]. Even though real-world degradations are complex, within a specific setting or deployment, they tend to be limited in the domain and less diverse than the training pipelines. Consequently, [45] proposes customizing the synthetic pipeline to better match specific real-world data. Nonetheless, an unresolved gap exists between synthesized and real-world degradations. Instead, our method operates directly on real-world data and avoids the intrinsic gap from synthetic degradation.", + "bbox": [ + 212, + 598, + 785, + 777 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "The second strategy connects real-world and synthetic degradation with transfer learning. Most works use image-to-image translation [5,30,40] to adapt the synthetic degradation to real-world ones. Such methods resort to intricate designs like CycleGAN [50], which struggle to replicate real-world degradations [34]", + "bbox": [ + 212, + 779, + 785, + 840 + ], + "page_idx": 2 + }, + { + "type": "header", + "text": "Pairwise Distance Distillation", + "bbox": [ + 524, + 114, + 730, + 127 + ], + "page_idx": 2 + }, + { + "type": "page_number", + "text": "3", + "bbox": [ + 774, + 116, + 784, + 126 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "reliably. Our proposed method explores feature distances between model outputs and bypasses the need to mimic real-world degradations.", + "bbox": [ + 212, + 146, + 782, + 176 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Knowledge Distillation is initially introduced for model compression [10]. In low-level vision, several works [7, 28, 47] distill large models into more efficient ones by enforcing similarity between their internal features or predictions. Recently, knowledge distillation has served diverse purposes in high-level tasks like segmentation, including domain adaptation [6, 15, 27] and transfer learning [38, 41]. However, such applications are less discussed in low-level vision.", + "bbox": [ + 212, + 176, + 784, + 268 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "3 Method", + "text_level": 1, + "bbox": [ + 215, + 290, + 330, + 306 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Our method adapts the specialist in synthetic degradation to an unlabeled real-world domain. We use the low-level characteristics of synthetic predictions from the specialists and the knowledge of the generalist model. Fig. 2 is the overview of our approach and the novel Pairwise Distance Distillation (PDD) method, which enforces the consistency of the intra- and inter-model distances to improve the specialist $(M_S)$ 's real-world predictions. Sec. 3.1 gives notions used in this section and Sec. 3.2 illustrates the unsupervised formulation. Sec. 3.3 explains the details of PDD and Sec. 3.4 provides the full optimization aims and discusses the static and Exponential Moving Average (EMA) versions of our method.", + "bbox": [ + 212, + 321, + 782, + 458 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "3.1 Definitions & Setup", + "text_level": 1, + "bbox": [ + 215, + 479, + 426, + 494 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "When provided with an LR input $X \\in \\mathbb{R}^{H \\times W \\times C}$ - where $H \\times W$ is the spatial resolution and $C$ is the number of color channels - a SISR model aims to reconstruct an HR image $\\hat{Y} \\in \\mathbb{R}^{rH \\times rW \\times C}$ , scaling by a factor $r$ . Due to the difficulty in obtaining paired training data, the conventional approach in SISR involves generating an LR image $X$ by downsampling the HR ground-truth $Y$ , i.e., $X = D_{\\downarrow}(Y)$ . This downsampling operation $D_{\\downarrow}$ typically refers to bicubic interpolation, although other alternatives exist [1].", + "bbox": [ + 212, + 503, + 784, + 609 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "In a more extensive setting, it is typical for the LR image to be affected by various factors, such as blurring, noise, compression artifacts, etc. To exhibit these degradations, the LR image $X$ can be formed by applying $D_{\\downarrow}$ along with one or multiple degradations $D_{i}$ , in any order, i.e.,", + "bbox": [ + 212, + 609, + 784, + 671 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\nX = D _ {1} \\circ \\dots \\circ D _ {n} (Y). \\tag {1}\n$$\n", + "text_format": "latex", + "bbox": [ + 421, + 683, + 784, + 699 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Here, we include $D_{\\downarrow}$ as one of the degradations $\\{D_i\\}_{i \\in [1,n]}$ for simplicity. Now, let's consider two categories of degradation sets consisting of known degradations: $\\mathcal{D}_S$ , which comprises simple degradations with a few factors, and $\\mathcal{D}_G$ characterized by complex degradations with a wide range of factors:", + "bbox": [ + 212, + 709, + 782, + 771 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\n\\mathcal {D} _ {S} = \\left\\{D _ {i} \\right\\} _ {i \\in \\mathcal {S}}, \\mathcal {D} _ {G} = \\left\\{D _ {i} \\right\\} _ {i \\in \\mathcal {G}}, \\quad \\text {w h e r e} | \\mathcal {G} | \\gg | \\mathcal {S} |. \\tag {2}\n$$\n", + "text_format": "latex", + "bbox": [ + 323, + 782, + 784, + 799 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "In the simplest case for $\\mathcal{D}_S$ , the set $\\mathcal{D}_S = \\{D_\\downarrow\\}$ , which represents the standard SISR setup. For the broader case, $\\mathcal{D}_G$ should encompass diverse and challenging", + "bbox": [ + 212, + 809, + 784, + 840 + ], + "page_idx": 3 + }, + { + "type": "page_number", + "text": "4", + "bbox": [ + 217, + 114, + 228, + 126 + ], + "page_idx": 3 + }, + { + "type": "header", + "text": "Y. Zhang et al.", + "bbox": [ + 271, + 114, + 372, + 128 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "degradations beyond $\\{D_{\\downarrow}\\}$ to ensure that models trained on this data generalize well to blind settings [35, 43].", + "bbox": [ + 212, + 146, + 782, + 176 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Consider two models, $M_S$ and $M_G$ , trained on datasets constructed with $\\mathcal{D}_S$ and $\\mathcal{D}_G$ respectively. Both models are optimized with common supervised SR learning losses [13, 36] including MSE-based loss, VGG loss, and adversarial training (see Eq. (9)). Since $\\mathcal{D}_S$ comprises a narrower range of degradations compared to $\\mathcal{D}_G$ , $M_S$ is a \"specialist\" model for the limited domain defined by $\\mathcal{D}_S$ . On the other hand, $M_G$ is a \"generalist\" model as it achieves moderate performance across the broader domain derived by $\\mathcal{D}_G$ . Comparing $M_S$ and $M_G$ , we assume the specialist's superiority to the generalist in $\\mathcal{D}_S$ and the generalist's better handling of unknown degradations. These assumptions are empirically validated in Tab. 1.", + "bbox": [ + 212, + 176, + 787, + 327 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "3.2 Unsupervised Learning Through Distillation", + "text_level": 1, + "bbox": [ + 214, + 349, + 625, + 366 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Our approach aligns with the unsupervised setup [5, 30, 40], wherein we work with a collection of real-world LR images alongside unpaired HR clean images. We denote the set of LR images as $\\{X^U\\}$ with the same unknown degradations $\\{D_i\\}_{i\\in \\mathcal{U}}$ . Additionally, we generate LR counterparts $\\{X^L\\}$ from HR images using specific degradations from $\\mathcal{D}_S$ , e.g. bicubic interpolation. The specialist $M_S$ is adept in restoring $X^L$ , while $M_G$ performs moderately in restoring $X^L$ and $X^U$ . We aim to adapt $M_S$ to $\\{X^U\\}$ with the aid of $M_G$ . As in [10], a naive distillation approach is having $M_S$ imitate the predictions from $M_G$ for unlabeled inputs:", + "bbox": [ + 214, + 376, + 787, + 500 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\n\\begin{array}{l} \\hat {Y} _ {S} ^ {L} = M _ {S} (X ^ {L}), \\hat {Y} _ {S} ^ {U} = M _ {S} (X ^ {U}), \\hat {Y} _ {G} ^ {U} = M _ {G} (X ^ {U}), \\\\ \\mathcal {L} _ {N D} = \\mathcal {L} _ {L} \\left(\\hat {Y} _ {S} ^ {U}, \\hat {Y} _ {G} ^ {U}\\right) + \\lambda \\mathcal {L} _ {L} \\left(\\hat {Y} _ {S} ^ {L}, Y ^ {L}\\right), \\tag {3} \\\\ \\end{array}\n$$\n", + "text_format": "latex", + "bbox": [ + 318, + 508, + 784, + 549 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "where $Y^{L}$ represents the ground-truth image in the labeled domain, $\\mathcal{L}_L$ denotes the supervised loss formulated in Eq. (9), and $\\lambda$ is a scale factor to balance between the distillation and primary objective of $M_S$ . However, this simple imitation approach relies solely on information from the generalist $M_G$ and fails to harness the strengths of both models.", + "bbox": [ + 212, + 561, + 787, + 638 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "3.3 Pairwise Distance Distillation (PDD)", + "text_level": 1, + "bbox": [ + 214, + 660, + 568, + 676 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Consider any pair of LR images - a real-world $X^U$ and a synthesized $X^L$ . As shown in Fig. 2, we obtain four predictions by applying $M_S$ and $M_G$ :", + "bbox": [ + 212, + 685, + 787, + 718 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\n\\begin{array}{l} \\hat {Y} _ {S} ^ {U} = M _ {S} (X ^ {U}), \\hat {Y} _ {G} ^ {U} = M _ {G} (X ^ {U}), \\\\ \\text {a n d} \\hat {Y} _ {S} ^ {L} = M _ {S} \\left(X ^ {L}\\right), \\hat {Y} _ {G} ^ {L} = M _ {G} \\left(X ^ {L}\\right). \\tag {4} \\\\ \\end{array}\n$$\n", + "text_format": "latex", + "bbox": [ + 366, + 728, + 784, + 767 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Our distillation explores the relationship between predictions in Eq. (4) to seek for knowledge combination. We use VGG features [31] of the predictions as our basis for exploration. VGG, as a classification model, inherently captures semantic information. Yet it also captures low-level characteristics related to", + "bbox": [ + 212, + 779, + 787, + 840 + ], + "page_idx": 4 + }, + { + "type": "header", + "text": "Pairwise Distance Distillation", + "bbox": [ + 524, + 114, + 730, + 127 + ], + "page_idx": 4 + }, + { + "type": "page_number", + "text": "5", + "bbox": [ + 774, + 116, + 784, + 126 + ], + "page_idx": 4 + }, + { + "type": "image", + "img_path": "images/12350873bd7515b36d1bb9eac7095c5779bdad28b5317afcef3ab01b1e90f8ef.jpg", + "image_caption": [ + "Fig. 2: Schematic of Pairwise Distance Distillation (PDD). $X^{U}$ and $X^{L}$ are unlabeled (real-world) and labeled (synthetic) inputs. Model $M_{G}$ is a generalist trained with extensive synthetic pipeline, while $M_{S}$ specializes in $X^{L}$ . The prediction $\\hat{Y}_S^L$ is supervised by its ground truth throughout training. PDD enforces the consistency between $\\{\\varDelta_U, \\varDelta_L\\}$ and between $\\{d_S, d_G\\}$ to improve $M_{S}$ 's real-world performance." + ], + "image_footnote": [], + "bbox": [ + 246, + 145, + 756, + 247 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "image quality [37], e.g. blur and sharpness, and is widely used in the perceptual loss for super-resolution tasks [13, 35, 36].", + "bbox": [ + 212, + 359, + 784, + 390 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "As a generalist, $M_G$ outputs synthetic and real-world predictions with small quality gaps, which is characterized by the overlapping distributions of low-level characteristics of $\\{\\hat{Y}_G^L\\}$ and $\\{\\hat{Y}_G^U\\}$ [20, 21]. Following [20], Fig. 5a projects the low-level features of $\\{\\hat{Y}_G^L\\}$ (DIV2K) and $\\{\\hat{Y}_G^U\\}$ (NTIRE20). The generalist's coverage strongly overlaps while the specialist's predictions are well-separated (see Fig. 5b). The within-model relationship of the generalist exhibits the generalizability to real-world degradation.", + "bbox": [ + 212, + 391, + 784, + 497 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "On the other side, the high image quality of $\\{\\hat{Y}_S^L\\}$ , due to $M_S$ 's specialization, makes its low-level characteristics a valuable reference. To integrate the knowledge for real-world performance, we propose to make the specialist imitate the generalist's within-model relationship, that is, pushing the low-level characteristics of real-world predictions from $M_S$ , $\\{\\hat{Y}_S^U\\}$ , to be similar to that of its synthetic predictions $\\{\\hat{Y}_S^L\\}$ . To achieve this aim, we first assume such similarity is satisfied for $M_S$ and conclude the following consistencies about VGG feature distances between predictions:", + "bbox": [ + 212, + 498, + 785, + 619 + ], + "page_idx": 5 + }, + { + "type": "list", + "sub_type": "text", + "list_items": [ + "1. Intra-model distances for predictions of the same input pair should be consistent across $M_S$ and $M_G$ (green shade in Fig. 2).", + "2. Inter-model distances for predictions of a single input should be consistent across synthetic and real-world domains (orange shade in Fig. 2)." + ], + "bbox": [ + 220, + 632, + 782, + 691 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Our rationales for these consistencies are as follows. First, we recall the low-level characteristics of predictions from the same model are similar. We discussed it for $M_G$ with Fig. 5a and made such an assumption for $M_S$ . Thus, the distance between the same model's predictions $\\{\\hat{Y}_G^L,\\hat{Y}_G^U\\}$ (or $\\{\\hat{Y}_S^L,\\hat{Y}_S^U\\}$ ), i.e. intra-model distance, reflects mainly semantic differences due to the close low-level characteristics. As it naturally follows that the applied SR models should not change semantics, intra-model distances should be consistent for $M_S$ and $M_G$ .", + "bbox": [ + 212, + 703, + 784, + 808 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Second, we note the low-level characteristics of predictions from the two models differ, e.g. $\\hat{Y}_S^L$ has higher quality than the counterpart $\\hat{Y}_G^L$ . The distance", + "bbox": [ + 212, + 809, + 785, + 840 + ], + "page_idx": 5 + }, + { + "type": "page_number", + "text": "6", + "bbox": [ + 217, + 114, + 228, + 126 + ], + "page_idx": 5 + }, + { + "type": "header", + "text": "Y. Zhang et al.", + "bbox": [ + 271, + 114, + 374, + 128 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "between the two predictions $\\{\\hat{Y}_G^L,\\hat{Y}_S^L\\}$ (or $\\{\\hat{Y}_G^U,\\hat{Y}_S^U\\}$ ) of a given sample, i.e. inter-model distance, captures such differences in low-level characteristics, as semantics stay constant for the predictions of the same input. Given the similarity of low-level characteristics within a single model, the inter-model distances for synthetic and real-world samples should be consistent.", + "bbox": [ + 212, + 143, + 782, + 220 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "We structure the two consistencies as distillations on the intra- and inter-model distances. Our method encourages $M_S$ 's real-world predictions $\\{\\hat{Y}_S^U\\}$ to have similar low-level characteristics as its synthetic predictions $\\{\\hat{Y}_S^L\\}$ , taking it as a reference from high-quality images.", + "bbox": [ + 212, + 222, + 782, + 282 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Intra-model Distance Distillation enforces consistency on the distances between predictions from the same model (green shading in Fig. 2). Consider the $\\ell_1$ distance $d_G^{ij}$ between predictions from the generalist $M_G$ , i.e. $\\hat{Y}_G^L$ and $\\hat{Y}_G^U$ , in VGG feature space, and the distance $d_S^{ij}$ for the specialist $M_S$ is similarly defined:", + "bbox": [ + 214, + 284, + 784, + 357 + ], + "page_idx": 6 + }, + { + "type": "equation", + "text": "\n$$\nd _ {G} ^ {i j} = \\left\\| \\Phi_ {i j} \\left(\\hat {Y} _ {G} ^ {L}\\right) - \\Phi_ {i j} \\left(\\hat {Y} _ {G} ^ {U}\\right) \\right\\| _ {1}, \\tag {5}\n$$\n", + "text_format": "latex", + "bbox": [ + 398, + 358, + 784, + 385 + ], + "page_idx": 6 + }, + { + "type": "equation", + "text": "\n$$\nd _ {S} ^ {i j} = \\left\\| \\Phi_ {i j} \\left(\\hat {Y} _ {S} ^ {L}\\right) - \\Phi_ {i j} \\left(\\hat {Y} _ {S} ^ {U}\\right) \\right\\| _ {1},\n$$\n", + "text_format": "latex", + "bbox": [ + 398, + 378, + 604, + 396 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "where $\\varPhi_{ij}$ refers to the $j$ -th layer in the $i$ -th residual block of VGG19. For $d_G^{ij}, d_S^{ij} \\in \\mathbb{R}^{c \\times h \\times w}, h \\times w$ is the spatial resolution and $c$ is the number of channels. We enforce the consistency between $d_G^{ij}$ and $d_S^{ij}$ by minimizing their difference measured with the Cross-Entropy (CE) $R_{intra}$ :", + "bbox": [ + 214, + 398, + 782, + 462 + ], + "page_idx": 6 + }, + { + "type": "equation", + "text": "\n$$\nR _ {i n t r a} = - \\frac {1}{h w} \\sum_ {i, j} \\sum_ {m, n} \\mathbf {S} \\left(d _ {G} ^ {i j} [ m, n ]\\right) \\log \\mathbf {S} \\left(d _ {S} ^ {i j} [ m, n ]\\right), \\tag {6}\n$$\n", + "text_format": "latex", + "bbox": [ + 323, + 469, + 782, + 503 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "where $[m,n]$ is a spatial index of the feature map, and $\\mathbf{S}(\\cdot)$ denotes the SoftMax function with the input size being $\\mathbb{R}^c$ . The element-wise CE reflects the negative log-likelihood of distances locally. While other measures are feasible, CE is empirically a good choice in favoring the overall sharpness (see Sec. 4.4).", + "bbox": [ + 214, + 510, + 782, + 570 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Inter-model Distance Distillation enforces the consistency of changes in low-level characteristics between predictions from different models (orange shading in Fig. 2). Given the two predictions for a single input, i.e. $\\{\\hat{Y}_G^L,\\hat{Y}_S^L\\}$ (or $\\{\\hat{Y}_G^U,\\hat{Y}_S^U\\}$ ), we first calculate their feature distance $\\varDelta_L^{ij}$ (or $\\varDelta_U^{ij}$ ) to represent the differences in low-level characteristics. However, enforcing consistency on $\\varDelta_L^{ij}$ and $\\varDelta_U^{ij}$ is less straightforward, as the spatial layout of corresponding inputs can differ. Thus, we compute the inter-model distances with the following:", + "bbox": [ + 214, + 570, + 784, + 676 + ], + "page_idx": 6 + }, + { + "type": "equation", + "text": "\n$$\n\\Delta_ {L} ^ {i j} = \\operatorname {G r a m} \\left(\\Phi_ {i j} \\left(\\hat {Y} _ {S} ^ {L}\\right) - \\Phi_ {i j} \\left(\\hat {Y} _ {G} ^ {L}\\right)\\right), \\tag {7}\n$$\n", + "text_format": "latex", + "bbox": [ + 380, + 683, + 784, + 709 + ], + "page_idx": 6 + }, + { + "type": "equation", + "text": "\n$$\n\\Delta_ {U} ^ {i j} = \\operatorname {G r a m} \\left(\\Phi_ {i j} \\left(\\hat {Y} _ {S} ^ {U}\\right) - \\Phi_ {i j} \\left(\\hat {Y} _ {G} ^ {U}\\right)\\right).\n$$\n", + "text_format": "latex", + "bbox": [ + 380, + 705, + 622, + 722 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Gram(\\cdot) refers to the Gram matrix, which calculates the correlations among vectorized feature maps along the channel dimension. The Gram matrix captures the statistics while collapsing the spatial layout [8].", + "bbox": [ + 214, + 728, + 782, + 772 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "For distillation, we ensure the consistency between inter-model distances by minimizing $R_{inter}$ , the Frobenius norm between $\\Delta_U^{ij}$ and $\\Delta_L^{ij}$ :", + "bbox": [ + 214, + 773, + 782, + 805 + ], + "page_idx": 6 + }, + { + "type": "equation", + "text": "\n$$\nR _ {i n t e r} = \\sum_ {i, j} \\| \\Delta_ {U} ^ {i j} - \\Delta_ {L} ^ {i j} \\| _ {\\mathcal {F}}, \\tag {8}\n$$\n", + "text_format": "latex", + "bbox": [ + 405, + 811, + 782, + 842 + ], + "page_idx": 6 + }, + { + "type": "header", + "text": "Pairwise Distance Distillation", + "bbox": [ + 524, + 114, + 730, + 127 + ], + "page_idx": 6 + }, + { + "type": "page_number", + "text": "7", + "bbox": [ + 774, + 114, + 784, + 126 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "where $\\| \\cdot \\|_{\\mathcal{F}}$ computes the Frobenius norm. The Frobenius norm of Gram Matrix difference is interpreted in [8, 16] as measuring distribution discrepancy.", + "bbox": [ + 212, + 146, + 787, + 179 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "3.4 Full Method", + "text_level": 1, + "bbox": [ + 215, + 198, + 366, + 213 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "The network $M_S$ is fully optimized with supervised losses for labeled synthetic data $X^L$ and unsupervised losses for real-world data $X^U$ .", + "bbox": [ + 212, + 223, + 782, + 252 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Supervised Losses maintain the specialization in synthetic degradation and reduce overfitting to the distillation. With the ground-truth image for prediction $\\hat{Y}_S^L$ , denoted as $Y^{L}$ , we implement a combination of supervised loss functions. It includes a wavelet-based loss $L_{\\mathrm{wv}}$ , a perceptual loss $L_{\\mathrm{vgg}}$ , and generative loss $L_{\\mathrm{gan}}$ for adversarial training:", + "bbox": [ + 214, + 253, + 787, + 330 + ], + "page_idx": 7 + }, + { + "type": "equation", + "text": "\n$$\n\\mathcal {L} _ {L} \\left(\\hat {Y} _ {S} ^ {L}, Y ^ {L}\\right) = \\alpha_ {1} L _ {\\mathrm {w v}} \\left(\\hat {Y} _ {S} ^ {L}, Y ^ {L}\\right) + \\alpha_ {2} L _ {\\mathrm {v g g}} \\left(\\hat {Y} _ {S} ^ {L}, Y ^ {L}\\right) + \\alpha_ {3} L _ {\\mathrm {g a n}} \\left(\\hat {Y} _ {S} ^ {L}\\right), \\tag {9}\n$$\n", + "text_format": "latex", + "bbox": [ + 269, + 342, + 785, + 359 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "where $L_{\\mathrm{vgg}}$ and $L_{\\mathrm{gan}}$ are adopted from RealESRGAN [35], with balancing weights $\\alpha_{1}$ , $\\alpha_{2}$ , and $\\alpha_{3}$ . We use but omit writing the discriminator loss for brevity. $L_{\\mathrm{wv}}$ refers to an $\\ell_{1}$ -based wavelet loss [4, 49]:", + "bbox": [ + 214, + 369, + 787, + 417 + ], + "page_idx": 7 + }, + { + "type": "equation", + "text": "\n$$\nL _ {\\mathrm {w v}} \\left(\\hat {Y} _ {S} ^ {L}, Y ^ {L}\\right) = \\sum_ {i} \\omega_ {i} \\| W \\left(\\hat {Y} _ {S} ^ {L}\\right) _ {i} - W \\left(Y ^ {L}\\right) _ {i} \\| _ {1}, \\tag {10}\n$$\n", + "text_format": "latex", + "bbox": [ + 339, + 428, + 785, + 457 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "where $W(\\cdot)_i$ extracts the $i$ -th wavelet channel, and $\\omega_i$ weights its importance. For the wavelet transform, we apply the Haar wavelet [14] for all output channels.", + "bbox": [ + 212, + 468, + 782, + 497 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Unsupervised Losses. In the absence of ground truth for $\\dot{Y}_S^U$ , we opt to optimize for the consistency outlined in Eq. (6) and Eq. (8) as regularizers. To fully utilize the discriminator's knowledge of realness, we also incorporate the generative loss in Eq. (9):", + "bbox": [ + 212, + 500, + 784, + 559 + ], + "page_idx": 7 + }, + { + "type": "equation", + "text": "\n$$\n\\mathcal {L} _ {U} = \\lambda_ {1} R _ {\\text {i n t r a}} + \\lambda_ {2} R _ {\\text {i n t e r}} + \\lambda_ {3} L _ {\\text {g a n}} \\left(\\hat {Y} _ {S} ^ {U}\\right), \\tag {11}\n$$\n", + "text_format": "latex", + "bbox": [ + 352, + 570, + 785, + 588 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "where $\\lambda_1, \\lambda_2$ , and $\\lambda_3$ balance the optimization aims. $L_{gan}$ here uses the same discriminator in Eq. (9). The supervised loss in Eq. (9) and the unsupervised distillation loss in Eq. (11) are combined as our final optimization objective, expressed as $\\mathcal{L} = \\mathcal{L}_L + \\mathcal{L}_U$ .", + "bbox": [ + 212, + 598, + 782, + 657 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Color Correction. In practice, our approach often leads to color shifts due to the regularization of distances within the feature space. To address this, we rectify the output by normalizing the mean and variance of each color channel with those of the corresponding input channels [35]. Further details are provided in Supplementary Sec. E.", + "bbox": [ + 212, + 659, + 784, + 734 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Learning Configurations. Our proposed PDD is a learning framework that can utilize pre-trained networks. It is recommended that $M_G$ is pre-trained with extensive degradations $\\mathcal{D}_G$ , as in [35,43]. The static configuration (refer to Fig. 3 (a)) initializes $M_S$ with the model pre-trained on specific synthetic degradation $(\\mathcal{D}_S)$ , and $M_G$ is frozen throughout the training process.", + "bbox": [ + 214, + 734, + 785, + 809 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Alternative configuration initializes $M_G$ and $M_S$ with the same model pretrained with $\\mathcal{D}_G$ and maintain $M_G$ as an Exponential Moving Average (EMA)", + "bbox": [ + 214, + 809, + 785, + 840 + ], + "page_idx": 7 + }, + { + "type": "page_number", + "text": "8", + "bbox": [ + 217, + 114, + 228, + 126 + ], + "page_idx": 7 + }, + { + "type": "header", + "text": "Y. Zhang et al.", + "bbox": [ + 271, + 114, + 374, + 128 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/9a06f91c09709db24ce244043fae065c542b7217eea9c97661dd4c811c64823c.jpg", + "image_caption": [ + "(a) Static Version" + ], + "image_footnote": [], + "bbox": [ + 346, + 145, + 500, + 239 + ], + "page_idx": 8 + }, + { + "type": "image", + "img_path": "images/42699b1275ee8fb378a964dd43c03b25f4764c12587b85096f877b1d62b73e1c.jpg", + "image_caption": [ + "(b) EMA Version", + "Fig. 3: Two initialization options for the generalist and specialist models. (a) The static configuration initializes $M_G$ with a model pre-trained by the complex synthetic pipeline and $M_S$ with a model pre-trained by simple degradation in $X^L$ . Weights of $M_G$ are frozen during distillation. (b) Both $M_S$ and $M_G$ are initialized with a pre-trained generalized model. Weights of $M_G$ are the EMA version of $M_S$ ." + ], + "image_footnote": [], + "bbox": [ + 517, + 143, + 660, + 239 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "version of $M_S$ (see Fig. 3 (b)). Despite $M_S$ not being initialized as a specialist, it can readily specialize to the synthetic degradation $\\mathcal{D}_S$ through supervised learning in Eq. (9). Empirically, the EMA version tends to be more effective, likely because the updated knowledge of real-world data in $M_S$ also benefits $M_G$ through EMA, enabling $M_G$ to yield better references for optimization.", + "bbox": [ + 214, + 367, + 787, + 444 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "4 Experiments", + "text_level": 1, + "bbox": [ + 215, + 465, + 375, + 483 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "4.1 Settings", + "text_level": 1, + "bbox": [ + 215, + 497, + 330, + 513 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "Datasets. We experiment on real-world datasets, RealSR [2] and DRealSR [39], and a synthetic dataset without released degradation settings, NTIRE20 [23]. For experiments on RealSR and DRealSR, we synthesize labeled training data $\\{X^L\\}$ from high-resolution images in the DF2K dataset [1,32]. For the NTIRE20 dataset, we use the provided target clean images to synthesize low-resolution counterparts. For synthesizing $\\{X^L\\}$ , we use only bicubic interpolation as the default and investigate other options for $\\mathcal{D}_S$ in the Supplementary Sec. D.b. RealSR or DRealSR has multiple data sources, and we optimize our methods for each source separately. For DRealSR, we report results for Panasonic and Olympus sources, as the others do not have sufficient training samples.", + "bbox": [ + 214, + 522, + 787, + 672 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "Training Details. We crop patches of size $48 \\times 48$ as inputs and train the model with the ADAM optimizer [12] using the settings $\\beta_{1} = 0.9$ , $\\beta_{2} = 0.999$ , and $\\epsilon = 10^{-8}$ . The batch size is 16, including equal numbers of labeled and unlabeled inputs. The initial learning rate is $1e - 4$ and halved after $25\\mathrm{K}$ iterations. The static version needs $50\\mathrm{K}$ iterations, and the EMA version is trained for $100\\mathrm{K}$ iterations due to the initial specialist underfitting to the synthetic domain. All models are for $\\times 4$ super-resolution. Other details are in Supplementary Sec. A.", + "bbox": [ + 214, + 674, + 787, + 777 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "Evaluation. We evaluate models on testing sets of RealSR, DRealSR, and the validation set of NTIRE20. We crop non-overlapped 512x512 patches from LR images in DRealSR to avoid memory exhaustion. For other datasets, we use the full images. To fully use the provided high-resolution counterparts, we report", + "bbox": [ + 214, + 779, + 787, + 839 + ], + "page_idx": 8 + }, + { + "type": "header", + "text": "Pairwise Distance Distillation", + "bbox": [ + 524, + 114, + 730, + 127 + ], + "page_idx": 8 + }, + { + "type": "page_number", + "text": "9", + "bbox": [ + 774, + 114, + 785, + 126 + ], + "page_idx": 8 + }, + { + "type": "table", + "img_path": "images/c10b71c38239897b3a46ceae9f83c3798ffbf9f5589b6338c7d3063808597a12.jpg", + "table_caption": [ + "Table 1: $M_G$ and $M_S$ (static) columns show initial performances on the bicubic-interpolated data $\\{X^L\\}$ from Set14 [42] and $\\{X^U\\}$ with unknown degradation from NTIRE20 [23]. $M_S$ (EMA) column shows the performance of $M_S$ in the EMA version on $\\{X^L\\}$ after training process." + ], + "table_footnote": [], + "table_body": "
DomainMetricMGMS(static)MS(EMA)DomainMetricMGMS(static)
{X^L}PSNR↑23.9824.8126.26{X^U}PSNR↑25.0819.82
LPIPS↓0.23490.13370.1471LPIPS↓0.25040.7552
", + "bbox": [ + 261, + 212, + 741, + 252 + ], + "page_idx": 9 + }, + { + "type": "image", + "img_path": "images/adac3a713f37c17d3700c1664a4f5998e292bc09811b0b9568eb473d7e07adaa.jpg", + "image_caption": [ + "Fig. 4: Improvements over the Generalist on five unlabeled data domains, where a lower LPIPS score is better. ND improves fidelity scores (PSNR) but dramatically drops perceptual scores (LPIPS and NRQM). Both versions of our method achieve better improvements than ND for all reported metrics. For each domain, the Static version improves at least one of the perceptual metrics; the EMA version improves all." + ], + "image_footnote": [], + "bbox": [ + 227, + 266, + 405, + 371 + ], + "page_idx": 9 + }, + { + "type": "image", + "img_path": "images/e5caf7f5458acc358aa53c142ed322b3cd65192f3994ba1b8ff29c2aa6b504dc.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 406, + 268, + 594, + 372 + ], + "page_idx": 9 + }, + { + "type": "image", + "img_path": "images/ce96daf2ed4534ed3889474a07f1593f55cf0449e01b92eddd8b124e53975af4.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 599, + 267, + 784, + 371 + ], + "page_idx": 9 + }, + { + "type": "text", + "text": "scores of full-reference metrics, PSNR, SSIM, and LPIPS [44]. The PSNR and SSIM scores are calculated on the Y-channel of YCbCr format images. Additionally, we follow BSRGAN [43] and supplement with the no-reference metric NRQM [24] to evaluate the perceptual clearness, as the ground truth in real-world captured datasets [2,39] also feature some blur. We use the implementation in IQA-PyTorch3 for all metrics. It is important to consider both fidelity and perceptual scores; fidelity metrics are worse at detecting blurriness while perceptual metrics are not sensitive to high-frequency artifacts [25, 26, 29, 48].", + "bbox": [ + 212, + 483, + 787, + 606 + ], + "page_idx": 9 + }, + { + "type": "text", + "text": "4.2 Effectiveness of PDD", + "text_level": 1, + "bbox": [ + 214, + 631, + 437, + 645 + ], + "page_idx": 9 + }, + { + "type": "text", + "text": "Comparisons. We compare the two versions of our method in Fig. 3 with the naive distillation defined in Eq. (3) regarding their improvements over the generalist network $M_G$ . We choose RealESRGAN [35] as the pre-trained generalist model and ESRGAN [36], a standard SISR model, as the pre-trained specialist for the naive distillation and the static version in Fig. 3 (a). Tab. 1 shows the initial performances of the generalist and specialist for $\\{X^L\\}$ and $\\{X^U\\}$ . As we assumed in Sec. 3.1, the specialist performs better than the generalist for $\\{X^L\\}$ while $M_G$ generalizes better for $\\{X^U\\}$ . For the EMA version in Fig. 3 (b), $M_S$ as well as $M_G$ is initialized with RealESRGAN. Through the supervised learning in Eq. (9), $M_S(\\mathrm{EMA})$ shows specialized performance for labeled data $\\{X^L\\}$ .", + "bbox": [ + 212, + 660, + 787, + 813 + ], + "page_idx": 9 + }, + { + "type": "page_number", + "text": "10", + "bbox": [ + 217, + 114, + 235, + 126 + ], + "page_idx": 9 + }, + { + "type": "header", + "text": "Y. Zhang et al.", + "bbox": [ + 271, + 114, + 374, + 128 + ], + "page_idx": 9 + }, + { + "type": "page_footnote", + "text": "3 https://github.com/chaofengc/IQA-PyTorch", + "bbox": [ + 217, + 824, + 532, + 840 + ], + "page_idx": 9 + }, + { + "type": "image", + "img_path": "images/b9b0e83fd39785c5a23976a9f9c8b8420f148cee9965642b117a07c7173992fe.jpg", + "image_caption": [ + "(a) Generalist" + ], + "image_footnote": [], + "bbox": [ + 263, + 145, + 415, + 231 + ], + "page_idx": 10 + }, + { + "type": "image", + "img_path": "images/13111ea952a2ff2f9334a41a885eb6adab64666cff078714724ab6b817f11af1.jpg", + "image_caption": [ + "(b) Bicubic Specialist", + "Fig. 5: Visualization of low-level features for predictions of DIV2K [1] (bicubic) and NTIRE20 (unknown) following [20] (a) Generalist's predictions for the two domains has overlapped distribution. (b) The predictions from the specialist that only adept in bicubic interpolation have separated distributions. (c) After applying our method (static version), predictions for two domains are pushed close." + ], + "image_footnote": [], + "bbox": [ + 426, + 145, + 578, + 231 + ], + "page_idx": 10 + }, + { + "type": "image", + "img_path": "images/0dbef2d3aa8ff9f165ed726739e40ab8bc9098996d292124aeefc0f65688c46b.jpg", + "image_caption": [ + "(c) Adapted Specialist" + ], + "image_footnote": [], + "bbox": [ + 589, + 145, + 741, + 231 + ], + "page_idx": 10 + }, + { + "type": "text", + "text": "Fig. 4 plots the improvements over the pretrained generalist for Naive Distillation (ND), Static, and EMA versions of our method. Compared to the generalist, ND dramatically drops in perceptual scores (LPIPS, NRQM) despite maintaining or improving the PSNR. Both versions of our method perform better than ND; compared to the generalist, our static version improves PSNR significantly and at least one perceptual score for each data domain. The EMA version improves all metrics with the highest magnitudes.", + "bbox": [ + 212, + 357, + 787, + 462 + ], + "page_idx": 10 + }, + { + "type": "text", + "text": "Change in Low-Level Characteristics. We extract low-level features for visualization following [20]. Fig. 5b visualize predictions generated by the bicubic specialist (ESRGAN), which has separate distributions. Fig. 5c verifies the predictions from the specialist after applying our method (static version), where the separated distributions are pushed closer. Visualizing low-level differences between images with arbitrary content is challenging [20]. As such, we also quantitatively compare the KL divergence between the low-level features. As shown in Fig. 6, applying our method decreases the KL divergence between labeled (DIV2K-bicubic) and unlabeled predictions. Details of the visualization and KL-diversity calculations are in Supplementary Sec. B.", + "bbox": [ + 212, + 463, + 787, + 613 + ], + "page_idx": 10 + }, + { + "type": "text", + "text": "4.3 Comparisons with Other Methods", + "text_level": 1, + "bbox": [ + 214, + 633, + 542, + 648 + ], + "page_idx": 10 + }, + { + "type": "text", + "text": "Quantitative Comparison. We consider two variants of our method by applying the EMA version on RealESRGAN [35] and BSRGAN [43]. In Tab. 2, we compare their performances with other state-of-the-art RWSR methods, including FeMaSR [3], EDAN [17], DASR [40], DAN [11], RealESRGAN [35], BSRGAN [43], and TG [46], which released code and model weights that enable reproduction. We regenerate results to evaluate these methods, while Supplementary Sec. C discusses related but not directly comparable methods.", + "bbox": [ + 212, + 657, + 787, + 763 + ], + "page_idx": 10 + }, + { + "type": "text", + "text": "As shown in Tab. 2, EDAN, DAN, and DASR perform strongly on the fidelity metrics (PSNR, SSIM) but are poor perceptually, especially on the NRQM score. The qualitative comparison highlights the blurry effects. Our method enhances RealESRGAN and BSRGAN across all metrics and demonstrates superior performance compared to TG when both are applied to RealESRGAN. For FeMaSR,", + "bbox": [ + 212, + 763, + 787, + 839 + ], + "page_idx": 10 + }, + { + "type": "header", + "text": "Pairwise Distance Distillation", + "bbox": [ + 524, + 114, + 730, + 127 + ], + "page_idx": 10 + }, + { + "type": "page_number", + "text": "11", + "bbox": [ + 767, + 114, + 782, + 126 + ], + "page_idx": 10 + }, + { + "type": "image", + "img_path": "images/fc424c549e010f2a02e0bd6b0132326a4670d21da322796fc5bcbb28db78de02.jpg", + "image_caption": [ + "Fig. 6: KL divergence between specialist's unlabeled and labeled predictions. The KL divergence between labeled and unlabeled predictions is lower after applying our method (red) than that in the specialist before adaption (blue)." + ], + "image_footnote": [], + "bbox": [ + 263, + 152, + 416, + 250 + ], + "page_idx": 11 + }, + { + "type": "image", + "img_path": "images/999982b588e56a6e40aa14c057731539c91f904e67ea2cf4826f3b6c3b7788c0.jpg", + "image_caption": [ + "Fig.7: User study on outputs of 30 real-world images. We recruit 30 evaluators for each image. Each vote is for a better quality within a pair of options. We compared our method to RealESRGAN and BSRGAN by using them as pre-trained models; FeMaSR is compared to RealESRGAN+ours. Our method gains at least $65\\%$ votes." + ], + "image_footnote": [], + "bbox": [ + 506, + 146, + 787, + 227 + ], + "page_idx": 11 + }, + { + "type": "table", + "img_path": "images/e0b1ae9458556336027dc9be96d181e5769ffec273c81d57a06512b79ace2fd4.jpg", + "table_caption": [ + "Table 2: Quantitative comparison with state-of-the-art methods. DAN, EDAN, and DASR are separately listed as they favor fidelity metrics but yield poor perceptual scores. For other methods, **bold** and **underline** mark the best and second-best scores. Subscripts in the last two columns show their difference from the corresponding pretrained generalist, where improvement is colored red and drop is blue." + ], + "table_footnote": [], + "table_body": "
DatasetMetricDAN [11]EDAN [17]DASR [40]FeMaSR [3]RealESRGAN [43]BSRGAN [35]RealESRGAN +TG [46]RealESRGAN +OursBSRGAN +Ours
NTIRE20 -ValidPSNR↑26.8326.59-23.5425.0825.4425.5626.42(+1.340)26.37(+0.930)
SSIM↑0.71710.7400-0.66650.70610.69840.71930.7297(+0.024)0.7238(+0.025)
LPIPS↓0.57470.2678-0.23600.25040.26450.24200.2475(-0.003)0.2595(-0.005)
NRQM↑4.87845.8481-6.62456.12136.27795.79266.2263(+0.105)6.2781(+0.0002)
RealSR -CanonPSNR↑26.6726.3626.7124.2924.7425.5725.0325.94(+1.200)26.13(+0.560)
SSIM↑0.77400.77740.77820.74600.76340.76830.77200.7733(+0.010)0.7626(-0.006)
LPIPS↓0.40950.30260.25070.28090.26070.25730.26150.2516(-0.009)0.2517(-0.006)
NRQM↑3.04073.63603.93836.01516.06496.02935.42506.0647(-0.0002)6.1323(+0.103)
RealSR -NikonPSNR↑26.2925.3725.6323.9424.3124.7924.5625.28(+0.970)25.07(+0.280)
SSIM↑0.75330.74600.74730.70970.74060.73340.74060.7470(+0.014)0.7294(-0.001)
LPIPS↓0.41330.32000.27850.30460.28510.27970.29490.2802(-0.005)0.2713(-0.008)
NRQM↑3.20564.48724.49525.94885.66855.93875.19425.9044(+0.236)6.1276(+0.189)
DRealSR -PanasonicPSNR↑28.9928.12-25.0227.0326.9327.5127.87(+0.840)27.69(+0.760)
SSIM↑0.81230.8006-0.71180.77820.76210.78430.7939(+0.016)0.7752(+0.013)
LPIPS↓0.42350.2747-0.30880.27030.28200.28360.2682(-0.002)0.2687(-0.013)
NRQM↑2.81314.5915-6.03425.37515.58654.64835.3540(-0.021)5.5608(-0.026)
DRealSR -OlympusPSNR↑28.7427.85-24.3826.7326.4927.4327.40(+0.670)27.25(+0.764)
SSIM↑0.80630.7979-0.66910.76670.75680.78430.7846(+0.018)0.7635(+0.007)
LPIPS↓0.45950.3555-0.40070.31590.32830.34160.3139(-0.002)0.3268(-0.001)
NRQM↑2.79613.7989-6.25905.25305.53764.14595.3722(+0.120)5.5957(+0.058)
", + "bbox": [ + 220, + 460, + 787, + 686 + ], + "page_idx": 11 + }, + { + "type": "text", + "text": "we achieve competitive perceptual scores and significantly surpass their fidelity scores.", + "bbox": [ + 212, + 728, + 784, + 758 + ], + "page_idx": 11 + }, + { + "type": "text", + "text": "User Study is conducted by presenting two reconstructions of the same LR image and asking for the better one. We compare the EMA version of our method with methods RealESRGAN, BSRGAN, and FeMaSR. Each comparison is conducted on 30 different real-world results among 30 evaluators. As shown in Fig. 7, at least $65\\%$ evaluators vote for our method in each comparison. Although quan-", + "bbox": [ + 212, + 763, + 787, + 840 + ], + "page_idx": 11 + }, + { + "type": "page_number", + "text": "12", + "bbox": [ + 217, + 114, + 235, + 126 + ], + "page_idx": 11 + }, + { + "type": "header", + "text": "Y. Zhang et al.", + "bbox": [ + 271, + 114, + 374, + 128 + ], + "page_idx": 11 + }, + { + "type": "image", + "img_path": "images/08c40686f7cdf46f11179acf5ee3b4d46bf665dcd2a82aa0347d71b3d2327def.jpg", + "image_caption": [ + "(a) LR" + ], + "image_footnote": [], + "bbox": [ + 220, + 143, + 307, + 378 + ], + "page_idx": 12 + }, + { + "type": "image", + "img_path": "images/fe4126d5806061733da59ca32caea78ffea8581cc5a5d75aadc5a60df29e959a.jpg", + "image_caption": [ + "(b) FeMaSR" + ], + "image_footnote": [], + "bbox": [ + 315, + 143, + 401, + 378 + ], + "page_idx": 12 + }, + { + "type": "image", + "img_path": "images/a701cbffed2b69cac700b3bfe630a36a0ea7ecf6909494a40d05b33a34f62112.jpg", + "image_caption": [ + "(c) DASR" + ], + "image_footnote": [], + "bbox": [ + 410, + 143, + 496, + 378 + ], + "page_idx": 12 + }, + { + "type": "image", + "img_path": "images/94a365716f9add06d7e9deef1b9d16872979103513cd73fdecf30118425c7765.jpg", + "image_caption": [ + "(d) RE" + ], + "image_footnote": [], + "bbox": [ + 504, + 143, + 591, + 378 + ], + "page_idx": 12 + }, + { + "type": "image", + "img_path": "images/cc55b2e784cee3de2f35c78937d0beaea3ab304e47ca312581fb3aed6c3fa279.jpg", + "image_caption": [ + "(e) RE+ours", + "Fig. 8: Qualitative comparisons. LRs are from the RealSR-Canon dataset. Columns (e) and (f) are the results of applying our methods on RealESRGAN (RE) and BSRGAN (BG). FeMaSR and DASR can not predict reasonable details." + ], + "image_footnote": [], + "bbox": [ + 599, + 143, + 687, + 378 + ], + "page_idx": 12 + }, + { + "type": "image", + "img_path": "images/76210f2e54e46d87b870239539695486ecac9b1d4fac627b1b821560a83a56a0.jpg", + "image_caption": [ + "(f) BG+ours" + ], + "image_footnote": [], + "bbox": [ + 694, + 143, + 782, + 378 + ], + "page_idx": 12 + }, + { + "type": "text", + "text": "titatively competing, FeMaSR yields unpleasant artifacts to human eyes, further shown in qualitative comparison.", + "bbox": [ + 212, + 476, + 782, + 506 + ], + "page_idx": 12 + }, + { + "type": "text", + "text": "Qualitative Comparison is in Fig. 8. FeMaSR [3] produces high-frequency artifacts, while outputs from DASR [40] suffer blur or artifacts. Compared to RealESRGAN, our method predicts sharper patterns without introducing distortion. More comparisons are in Supplementary Sec. F.", + "bbox": [ + 212, + 507, + 784, + 566 + ], + "page_idx": 12 + }, + { + "type": "text", + "text": "4.4 Ablations", + "text_level": 1, + "bbox": [ + 214, + 588, + 341, + 602 + ], + "page_idx": 12 + }, + { + "type": "text", + "text": "Ablation on loss terms reveals the effectiveness of each term in Eq. (11). All experiments take RealESRGAN as the pre-trained model and follow the EMA implementation of our method in Fig. 3 (b). Fig. 9 shows that $R_{intra}$ and $R_{inter}$ relate differently to fidelity and perception. Their combination achieves balanced improvements in both directions, and the $L_{\\mathrm{gan}}$ increases pattern sharpness with a slight trade-off in PSNR. Controlling the ratio between weights of $R_{intra}$ and $R_{inter}$ results in fidelity and perceptual quality trade-offs. We choose their balance in Sec. 4.3 and refer to Supplementary Sec. D.a for more discussion.", + "bbox": [ + 212, + 613, + 784, + 733 + ], + "page_idx": 12 + }, + { + "type": "text", + "text": "Effectiveness of EMA is further validated by comparing to a SingleFixed version, which also uses RealESRGAN as the pre-trained model for both $M_S$ and $M_G$ but fixes the $M_G$ during training. As shown in Tab. 3, updating the weights of the generalist network through EMA enables higher performance in both fidelity and perception.", + "bbox": [ + 212, + 734, + 784, + 808 + ], + "page_idx": 12 + }, + { + "type": "text", + "text": "Choice of $R_{intra}$ measurement. Eq. (6) use Cross-Entropy (CE) for measuring the difference between $d_S^{ij}$ and $d_G^{ij}$ . Here, we alternate the CE to $\\ell_1$ distance", + "bbox": [ + 214, + 809, + 784, + 840 + ], + "page_idx": 12 + }, + { + "type": "header", + "text": "Pairwise Distance Distillation", + "bbox": [ + 526, + 114, + 730, + 126 + ], + "page_idx": 12 + }, + { + "type": "page_number", + "text": "13", + "bbox": [ + 767, + 114, + 784, + 126 + ], + "page_idx": 12 + }, + { + "type": "table", + "img_path": "images/8cd1517ede0fac8fcb7376f3a78dd5775236007edd18f747c7382e0a418cd902.jpg", + "table_caption": [ + "Table 3: Collection of ablations on EMA and choice of measurement in Eq. (6). Bold and underline mark the best and second best scores. SingleFixed uses the same prepretrained model to initialize $M_S$ and $M_G$ as the EMA version but fixes the weights of $M_G$ during training, resulting in inferior performance. Substituting CE in Eq. (6) to $\\ell_1$ biases the fidelity scores." + ], + "table_footnote": [], + "table_body": "
MethodCanonNTIRE20Olympus
PSNRNRQMPSNRNRQMPSNRNRQM
Generalist24.746.064925.086.121326.735.2530
SingleFixed25.715.980325.756.093827.475.2071
\\( \\ell_1 \\)26.175.806126.446.199727.724.9602
Ours25.946.064726.426.226327.405.3722
", + "bbox": [ + 318, + 227, + 686, + 315 + ], + "page_idx": 13 + }, + { + "type": "image", + "img_path": "images/88a535fed39f63c499dc3cda070746fde06c7777cc173d5c05a8a2c3d542523a.jpg", + "image_caption": [ + "Fig.9: Ablations of loss terms for distillation in Eq. (11). Using one of $R_{intra}$ (blue) and $R_{inter}$ (orange) bias fidelity or perceptual quality (a.-b.), while the combination of them (c.-d.) achieves a balanced improvement (see green and red dots)." + ], + "image_footnote": [], + "bbox": [ + 302, + 330, + 697, + 450 + ], + "page_idx": 13 + }, + { + "type": "text", + "text": "without changing the ratio between weights of $R_{intra}$ and $R_{inter}$ . As shown in Tab. 3, CE favors the overall sharpness while $\\ell_1$ focuses more on fidelity. Although $\\ell_1$ is a feasible choice, we use CE in our method due to the importance of sharpness for human eyes [13].", + "bbox": [ + 212, + 537, + 787, + 599 + ], + "page_idx": 13 + }, + { + "type": "text", + "text": "5 Conclusion", + "text_level": 1, + "bbox": [ + 215, + 631, + 359, + 647 + ], + "page_idx": 13 + }, + { + "type": "text", + "text": "This paper studies unsupervised real-world image super-resolution from the distillation perspective. The approach considers both the real-world knowledge of a generalized SR model and the low-level characteristics of synthetic predictions of the specialist model, adapting a synthetic domain specialist to target real-world degradation. The core of our method is the Pairwise Distance Distillation. By enforcing the consistencies for intra- and inter-model distances, our method aims at pushing the low-level characteristics of the specialist's real-world predictions towards its synthetic predictions. As a learning scheme, our method can improve off-the-shelf models regarding fidelity and perception on multiple real-world datasets. Additionally, we emphasize our approach to offering a new perspective toward addressing real-world super-resolution challenges.", + "bbox": [ + 212, + 672, + 787, + 840 + ], + "page_idx": 13 + }, + { + "type": "page_number", + "text": "14", + "bbox": [ + 217, + 114, + 235, + 126 + ], + "page_idx": 13 + }, + { + "type": "header", + "text": "Y. Zhang et al.", + "bbox": [ + 271, + 114, + 374, + 128 + ], + "page_idx": 13 + }, + { + "type": "text", + "text": "Acknowledgement This research/project is supported by the Ministry of Education, Singapore, under the Academic Research Fund Tier 1 (FY2022).", + "bbox": [ + 215, + 146, + 784, + 176 + ], + "page_idx": 14 + }, + { + "type": "text", + "text": "References", + "text_level": 1, + "bbox": [ + 217, + 196, + 321, + 213 + ], + "page_idx": 14 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "1. Agustsson, E., Timofte, R.: Ntire 2017 challenge on single image super-resolution: Dataset and study. In: The IEEE Conference on Computer Vision and Pattern Recognition (CVPR) Workshops (July 2017)", + "2. Cai, J., Zeng, H., Yong, H., Cao, Z., Zhang, L.: Toward real-world single image super-resolution: A new benchmark and a new model. In: Proceedings of the IEEE/CVF International Conference on Computer Vision. pp. 3086-3095 (2019)", + "3. Chen, C., Shi, X., Qin, Y., Li, X., Han, X., Yang, T., Guo, S.: Real-world blind super-resolution via feature matching with implicit high-resolution priors. In: Proceedings of the 30th ACM International Conference on Multimedia. pp. 1329-1338 (2022)", + "4. Deng, X., Yang, R., Xu, M., Dragotti, P.L.: Wavelet domain style transfer for an effective perception-distortion tradeoff in single image super-resolution. In: Proceedings of the IEEE/CVF international conference on computer vision. pp. 3076-3085 (2019)", + "5. Fritsche, M., Gu, S., Timofte, R.: Frequency separation for real-world superresolution. In: 2019 IEEE/CVF International Conference on Computer Vision Workshop (ICCVW). pp. 3599-3608. IEEE (2019)", + "6. Gao, H., Guo, J., Wang, G., Zhang, Q.: Cross-domain correlation distillation for unsupervised domain adaptation in nighttime semantic segmentation. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 9913-9923 (2022)", + "7. Gao, Q., Zhao, Y., Li, G., Tong, T.: Image super-resolution using knowledge distillation. In: Asian Conference on Computer Vision. pp. 527-541. Springer (2018)", + "8. Gatys, L.A., Ecker, A.S., Bethge, M.: Image style transfer using convolutional neural networks. In: Proceedings of the IEEE conference on computer vision and pattern recognition. pp. 2414-2423 (2016)", + "9. Gu, J., Lu, H., Zuo, W., Dong, C.: Blind super-resolution with iterative kernel correction. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 1604-1613 (2019)", + "0. Hinton, G., Vinyals, O., Dean, J.: Distilling the knowledge in a neural network. arXiv preprint arXiv:1503.02531 (2015)", + "1. Huang, Y., Li, S., Wang, L., Tan, T., et al.: Unfolding the alternating optimization for blind super resolution. Advances in Neural Information Processing Systems 33, 5632-5643 (2020)", + "2. Kingma, D.P., Ba, J.: Adam: A method for stochastic optimization. arXiv preprint arXiv:1412.6980 (2014)", + "3. Ledig, C., Theis, L., Huszár, F., Caballero, J., Cunningham, A., Acosta, A., Aitken, A., Tejani, A., Totz, J., Wang, Z., et al.: Photo-realistic single image superresolution using a generative adversarial network. In: Proceedings of the IEEE conference on computer vision and pattern recognition. pp. 4681-4690 (2017)", + "4. Lepik, U., Hein, H.: Haar wavelets. In: Haar Wavelets: With Applications, pp. 7-20. Springer (2014)", + "5. Li, W., Li, L., Yang, H.: Progressive cross-domain knowledge distillation for efficient unsupervised domain adaptive object detection. Engineering Applications of Artificial Intelligence 119, 105774 (2023)" + ], + "bbox": [ + 225, + 224, + 784, + 839 + ], + "page_idx": 14 + }, + { + "type": "header", + "text": "Pairwise Distance Distillation", + "bbox": [ + 524, + 114, + 730, + 127 + ], + "page_idx": 14 + }, + { + "type": "page_number", + "text": "15", + "bbox": [ + 767, + 116, + 785, + 126 + ], + "page_idx": 14 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "16. Li, Y., Wang, N., Liu, J., Hou, X.: Demystifying neural style transfer. arXiv preprint arXiv:1701.01036 (2017)", + "17. Liang, J., Zeng, H., Zhang, L.: Efficient and degradation-adaptive network for real-world image super-resolution. In: European Conference on Computer Vision. pp. 574-591. Springer (2022)", + "18. Liang, J., Cao, J., Sun, G., Zhang, K., Van Gool, L., Timofte, R.: Swinir: Image restoration using swim transformer. In: Proceedings of the IEEE/CVF international conference on computer vision. pp. 1833-1844 (2021)", + "19. Lim, B., Son, S., Kim, H., Nah, S., Mu Lee, K.: Enhanced deep residual networks for single image super-resolution. In: Proceedings of the IEEE conference on computer vision and pattern recognition workshops. pp. 136-144 (2017)", + "20. Liu, Y., Liu, A., Gu, J., Zhang, Z., Wu, W., Qiao, Y., Dong, C.: Discovering distinctive\" semantics\" in super-resolution networks. arXiv preprint arXiv:2108.00406 (2021)", + "21. Liu, Y., Zhao, H., Gu, J., Qiao, Y., Dong, C.: Evaluating the generalization ability of super-resolution networks. IEEE Transactions on pattern analysis and machine intelligence (2023)", + "22. Liu, Y., Zhang, W., Wang, J.: Adaptive multi-teacher multi-level knowledge distillation. Neurocomputing 415, 106-113 (2020)", + "23. Lugmayr, A., Danelljan, M., Timofte, R.: Ntire 2020 challenge on real-world image super-resolution: Methods and results. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition Workshops. pp. 494-495 (2020)", + "24. Ma, C., Yang, C.Y., Yang, X., Yang, M.H.: Learning a no-reference quality metric for single-image super-resolution. Computer Vision and Image Understanding 158, 1-16 (2017)", + "25. Ma, C., Rao, Y., Cheng, Y., Chen, C., Lu, J., Zhou, J.: Structure-preserving super resolution with gradient guidance. In: Proceedings of the IEEE/CVF conference on computer vision and pattern recognition. pp. 7769-7778 (2020)", + "26. Maggioni, M., Tanay, T., Babiloni, F., McDonagh, S., Leonardis, A.: Tunable convolutions with parametric multi-loss optimization. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 20226-20236 (2023)", + "27. Nguyen-Meidine, L.T., Belal, A., Kiran, M., Dolz, J., Blais-Morin, L.A., Granger, E.: Unsupervised multi-target domain adaptation through knowledge distillation. In: Proceedings of the IEEE/CVF Winter Conference on Applications of Computer Vision. pp. 1339-1347 (2021)", + "28. Park, S., Kwak, N.: Local-selective feature distillation for single image superresolution. arXiv preprint arXiv:2111.10988 (2021)", + "29. Park, S.H., Moon, Y.S., Cho, N.I.: Perception-oriented single image superresolution using optimal objective estimation. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 1725-1735 (2023)", + "30. Romero, A., Van Gool, L., Timofte, R.: Unpaired real-world super-resolution with pseudo controllable restoration. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 798-807 (2022)", + "31. Simonyan, K., Zisserman, A.: Very deep convolutional networks for large-scale image recognition. arXiv preprint arXiv:1409.1556 (2014)", + "32. Timofte, R., Agustsson, E., Van Gool, L., Yang, M.H., Zhang, L.: Ntire 2017 challenge on single image super-resolution: Methods and results. In: Proceedings of the IEEE conference on computer vision and pattern recognition workshops. pp. 114-125 (2017)" + ], + "bbox": [ + 215, + 146, + 784, + 839 + ], + "page_idx": 15 + }, + { + "type": "page_number", + "text": "16", + "bbox": [ + 217, + 114, + 235, + 126 + ], + "page_idx": 15 + }, + { + "type": "header", + "text": "Y. Zhang et al.", + "bbox": [ + 271, + 114, + 374, + 128 + ], + "page_idx": 15 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "33. Wang, H., Chen, X., Ni, B., Liu, Y., Liu, J.: Omni aggregation networks for lightweight image super-resolution. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 22378-22387 (2023)", + "34. Wang, W., Zhang, H., Yuan, Z., Wang, C.: Unsupervised real-world superresolution: A domain adaptation perspective. In: Proceedings of the IEEE/CVF International Conference on Computer Vision. pp. 4318-4327 (2021)", + "35. Wang, X., Xie, L., Dong, C., Shan, Y.: Real-esrgan: Training real-world blind super-resolution with pure synthetic data. In: Proceedings of the IEEE/CVF international conference on computer vision. pp. 1905-1914 (2021)", + "36. Wang, X., Yu, K., Wu, S., Gu, J., Liu, Y., Dong, C., Qiao, Y., Change Loy, C.: Esrgan: Enhanced super-resolution generative adversarial networks. In: Proceedings of the European conference on computer vision (ECCV) workshops. pp. 0-0 (2018)", + "37. Wang, Y., Cao, Y., Zha, Z.J., Zhang, J., Xiong, Z.: Deep degradation prior for low-quality image classification. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 11049-11058 (2020)", + "38. Wang, Y., Wang, L., Shi, S., Li, V.O., Tu, Z.: Go from the general to the particular: Multi-domain translation with domain transformation networks. In: Proceedings of the AAAI Conference on Artificial Intelligence. vol. 34, pp. 9233-9241 (2020)", + "39. Wei, P., Xie, Z., Lu, H., Zhan, Z., Ye, Q., Zuo, W., Lin, L.: Component divide-and-conquer for real-world image super-resolution. In: Computer Vision-ECCV 2020: 16th European Conference, Glasgow, UK, August 23-28, 2020, Proceedings, Part VIII 16. pp. 101-117. Springer (2020)", + "40. Wei, Y., Gu, S., Li, Y., Timofte, R., Jin, L., Song, H.: Unsupervised real-world image super resolution via domain-distance aware training. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 13385-13394 (2021)", + "41. Yang, C., Pan, J., Gao, X., Jiang, T., Liu, D., Chen, G.: Cross-task knowledge distillation in multi-task recommendation. In: Proceedings of the AAAI Conference on Artificial Intelligence. vol. 36, pp. 4318-4326 (2022)", + "42. Zeyde, R., Elad, M., Protter, M.: On single image scale-up using sparse-representations. In: Curves and Surfaces: 7th International Conference, Avignon, France, June 24-30, 2010, Revised Selected Papers 7. pp. 711-730. Springer (2012)", + "43. Zhang, K., Liang, J., Van Gool, L., Timofte, R.: Designing a practical degradation model for deep blind image super-resolution. In: Proceedings of the IEEE/CVF International Conference on Computer Vision. pp. 4791-4800 (2021)", + "44. Zhang, R., Isola, P., Efros, A.A., Shechtman, E., Wang, O.: The unreasonable effectiveness of deep features as a perceptual metric. In: Proceedings of the IEEE conference on computer vision and pattern recognition. pp. 586-595 (2018)", + "45. Zhang, R., Gu, J., Chen, H., Dong, C., Zhang, Y., Yang, W.: Crafting training degradation distribution for the accuracy-generalization trade-off. arXiv preprint arXiv:2305.18107 (2023)", + "46. Zhang, W., Li, X., Shi, G., Chen, X., Qiao, Y., Zhang, X., Wu, X.M., Dong, C.: Real-world image super-resolution as multi-task learning. Advances in Neural Information Processing Systems 36 (2024)", + "47. Zhang, Y., Chen, H., Chen, X., Deng, Y., Xu, C., Wang, Y.: Data-free knowledge distillation for image super-resolution. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 7852-7861 (2021)", + "48. Zhang, Y., Ji, B., Hao, J., Yao, A.: Perception-distortion balanced admm optimization for single-image super-resolution. In: European Conference on Computer Vision. pp. 108-125. Springer (2022)" + ], + "bbox": [ + 212, + 146, + 787, + 840 + ], + "page_idx": 16 + }, + { + "type": "header", + "text": "Pairwise Distance Distillation", + "bbox": [ + 524, + 114, + 730, + 127 + ], + "page_idx": 16 + }, + { + "type": "page_number", + "text": "17", + "bbox": [ + 767, + 114, + 785, + 126 + ], + "page_idx": 16 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "49. Zhong, Z., Shen, T., Yang, Y., Lin, Z., Zhang, C.: Joint sub-bands learning with clique structures for wavelet domain super-resolution. Advances in neural information processing systems 31 (2018)", + "50. Zhu, J.Y., Park, T., Isola, P., Efros, A.A.: Unpaired image-to-image translation using cycle-consistent adversarial networks. In: Proceedings of the IEEE international conference on computer vision. pp. 2223-2232 (2017)" + ], + "bbox": [ + 215, + 146, + 787, + 231 + ], + "page_idx": 17 + }, + { + "type": "page_number", + "text": "18", + "bbox": [ + 217, + 114, + 235, + 126 + ], + "page_idx": 17 + }, + { + "type": "header", + "text": "Y. Zhang et al.", + "bbox": [ + 271, + 114, + 374, + 128 + ], + "page_idx": 17 + } +] \ No newline at end of file diff --git a/2024/Pairwise Distance Distillation for Unsupervised Real-World Image Super-Resolution/821e13b2-5f5a-4742-b51d-28c69e766f13_model.json b/2024/Pairwise Distance Distillation for Unsupervised Real-World Image Super-Resolution/821e13b2-5f5a-4742-b51d-28c69e766f13_model.json new file mode 100644 index 0000000000000000000000000000000000000000..22de7181bc6c26eeb401b5575f86d826c5bbed98 --- /dev/null +++ b/2024/Pairwise Distance Distillation for Unsupervised Real-World Image Super-Resolution/821e13b2-5f5a-4742-b51d-28c69e766f13_model.json @@ -0,0 +1,2700 @@ +[ + [ + { + "type": "title", + "bbox": [ + 0.232, + 0.141, + 0.773, + 0.187 + ], + "angle": 0, + "content": "Pairwise Distance Distillation for Unsupervised Real-World Image Super-Resolution" + }, + { + "type": "text", + "bbox": [ + 0.306, + 0.213, + 0.699, + 0.229 + ], + "angle": 0, + "content": "Yuehan Zhang\\(^{1}\\), Seungjun Lee\\(^{2}\\), and Angela Yao\\(^{1}\\)" + }, + { + "type": "text", + "bbox": [ + 0.384, + 0.24, + 0.618, + 0.255 + ], + "angle": 0, + "content": "\\(^{1}\\) National University of Singapore" + }, + { + "type": "text", + "bbox": [ + 0.385, + 0.255, + 0.618, + 0.269 + ], + "angle": 0, + "content": "{zyuehan,ayao}@comp.nus.edu.sg" + }, + { + "type": "text", + "bbox": [ + 0.437, + 0.269, + 0.566, + 0.282 + ], + "angle": 0, + "content": "2 Korea University" + }, + { + "type": "text", + "bbox": [ + 0.42, + 0.284, + 0.584, + 0.296 + ], + "angle": 0, + "content": "9penguin9@korea.ac.kr" + }, + { + "type": "text", + "bbox": [ + 0.263, + 0.335, + 0.74, + 0.57 + ], + "angle": 0, + "content": "Abstract. Standard single-image super-resolution creates paired training data from high-resolution images through fixed downsampling kernels. However, real-world super-resolution (RWSR) faces unknown degradations in the low-resolution inputs, all the while lacking paired training data. Existing methods approach this problem by learning blind general models through complex synthetic augmentations on training inputs; they sacrifice the performance on specific degradation for broader generalization to many possible ones. We address the unsupervised RWSR for a targeted real-world degradation. We study from a distillation perspective and introduce a novel pairwise distance distillation framework. Through our framework, a model specialized in synthetic degradation adapts to target real-world degradations by distilling intra- and intermodel distances across the specialized model and an auxiliary generalized model. Experiments on diverse datasets demonstrate that our method significantly enhances fidelity and perceptual quality, surpassing state-of-the-art approaches in RWSR. The source code is available at https://github.com/Yuehan717/PDD." + }, + { + "type": "title", + "bbox": [ + 0.218, + 0.597, + 0.375, + 0.612 + ], + "angle": 0, + "content": "1 Introduction" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.629, + 0.785, + 0.734 + ], + "angle": 0, + "content": "Single-image super-resolution (SISR) predicts high-resolution (HR) images from low-resolution (LR) counterparts. The standard SISR model addresses predefined downsampling kernels, e.g. bicubic interpolation. However, real-world scenarios of SISR (RWSR) encompass unknown degradations in LR images, including blur, noise, and JPEG compression artifacts [35,43,45] with diverse combinations. The various and unknown degradations pose additional challenges when learning an RWSR model." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.735, + 0.787, + 0.84 + ], + "angle": 0, + "content": "Existing RWSR methods focus on blind generalization. Such methods synthesize paired training data with extensive and harsh degradations, involving multiple rounds of random blurring, added noise, resizing, and compression [35, 43]. The complex degradation pipeline allows the model to generalize to diverse unknown conditions. We refer to such a model addressing various degradations as a generalist. However, as much of the model's capacity is devoted to handling multiple conditions, these \"jack-of-all-trades\" models come with performance trade-" + } + ], + [ + { + "type": "page_number", + "bbox": [ + 0.218, + 0.116, + 0.23, + 0.127 + ], + "angle": 0, + "content": "2" + }, + { + "type": "header", + "bbox": [ + 0.272, + 0.115, + 0.375, + 0.129 + ], + "angle": 0, + "content": "Y. Zhang et al." + }, + { + "type": "image", + "bbox": [ + 0.263, + 0.143, + 0.737, + 0.259 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.214, + 0.266, + 0.789, + 0.34 + ], + "angle": 0, + "content": "Fig. 1: \\(\\hat{Y}_G^U\\) and \\(\\hat{Y}_G^L\\) are reconstructions for real-world and bicubic-interpolated (BI) inputs using a blind generalized model (generalist); \\(\\hat{Y}_S^U\\) and \\(\\hat{Y}_S^L\\) are counterparts using a standard SR model (bicubic specialist). The specialist enhances the BI inputs with clearer details, while the generalist does better for the real-world one. We distill the intra- and inter-model distances for an improved real-world reconstruction \\(\\hat{Y}_{S^{*}}^{U}\\)." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.375, + 0.784, + 0.408 + ], + "angle": 0, + "content": "offs [45], i.e., they have inferior performance compared to the model only optimized for the tested degradation \\((\\hat{Y}_G^L\\) and \\(\\hat{Y}_S^L\\) in Fig. 1)." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.41, + 0.789, + 0.56 + ], + "angle": 0, + "content": "SR models can be optimized for specific and known input degradations through supervised learning [9, 11, 18, 36]. The LR-HR pairs for training are created by degrading HR images with fixed kernels, e.g., bicubic interpolation, and Gaussian blur. We refer to models adept at the specific input degradation as a specialist. Degradations in RWSR datasets [2, 39] always belong to a specific domain [45]. For example, images captured by a specific camera model tend to feature consistent sensor noise [2]. However, the formulation of a real-world degradation is always unknown, posing difficulties in creating paired training data. As such, learning specialist models for such real-world domains is challenging, and strategies that do not require direct supervision are crucial." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.565, + 0.789, + 0.716 + ], + "angle": 0, + "content": "This paper tackles unsupervised RWSR from a knowledge distillation perspective. Distillation has been successful for domain adaptation in high-level tasks [6, 22, 38], but it is under-explored in low-level vision. Our distillation framework adapts a specialist model of known synthetic degradation to target real-world degradations. We use an auxiliary generalist model to provide generalization knowledge of real-world degradations that the specialist lacks. However, naive distillation only transfers knowledge from the generalist model; we aim to improve beyond the generalist's moderate performance. To that end, we also consider the low-level characteristics of the specialist's predictions on synthetic samples as a reference for the distillation." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.72, + 0.789, + 0.843 + ], + "angle": 0, + "content": "Our method uses the strength of both the specialist and the generalist by exploring the relationships between their predictions. Given a generalist and a specialist for synthetic degradation, we consider predictions of synthetic samples and real-world samples from both models (see Fig. 1). The generalist model has smaller gaps in qualities of real-world and synthetic predictions, featuring close low-level characteristics between domains [20,21]. In contrast, the specialist outputs distanced low-level characteristics but high-quality predictions for synthetic samples \\((\\hat{Y}_S^L\\) in Fig. 1), providing valuable low-level characteristics for reference." + } + ], + [ + { + "type": "header", + "bbox": [ + 0.526, + 0.115, + 0.732, + 0.128 + ], + "angle": 0, + "content": "Pairwise Distance Distillation" + }, + { + "type": "page_number", + "bbox": [ + 0.775, + 0.117, + 0.785, + 0.127 + ], + "angle": 0, + "content": "3" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.147, + 0.788, + 0.342 + ], + "angle": 0, + "content": "We propose to let the specialist imitate prediction relationships from the generalist - that is, push the specialist's real-world predictions \\((\\hat{Y}_S^U\\) in Fig. 1) to have low-level characteristics similar to its high-quality synthetic predictions. If such similarity is satisfied by the specialist, it would exhibit two consistencies. First, for two input samples, the difference between their predictions from the same model is intra-model distance; such distances would be consistent for the specialist and the generalist. Secondly, for a given input sample, the difference between generalist and specialist predictions is inter-model distance; such distances would be consistent whether the sample is synthetic or real-world. These intra- and inter-model distance consistencies form the basis of our pairwise distance distillation framework. Together with designs on configurations of specialist and generalist models, our adapted specialist model experimentally shows significant improvements over the generalist model." + }, + { + "type": "text", + "bbox": [ + 0.215, + 0.344, + 0.772, + 0.358 + ], + "angle": 0, + "content": "Contributions. To summarize, the contributions of our work are three-fold:" + }, + { + "type": "text", + "bbox": [ + 0.226, + 0.371, + 0.784, + 0.4 + ], + "angle": 0, + "content": "- To the best of our knowledge, we are the first to propose a distillation perspective for unsupervised RWSR to combine generalist and specialist models." + }, + { + "type": "text", + "bbox": [ + 0.226, + 0.402, + 0.785, + 0.445 + ], + "angle": 0, + "content": "- We propose a novel pairwise distance distillation framework, emphasizing transferring intra-model and inter-model distances to enhance the specialist's performance in real-world scenarios." + }, + { + "type": "text", + "bbox": [ + 0.226, + 0.448, + 0.784, + 0.476 + ], + "angle": 0, + "content": "- Experiments on three benchmarks demonstrate that our approach improves the off-the-shelf models regarding fidelity and perceptual performances." + }, + { + "type": "list", + "bbox": [ + 0.226, + 0.371, + 0.785, + 0.476 + ], + "angle": 0, + "content": null + }, + { + "type": "title", + "bbox": [ + 0.216, + 0.503, + 0.396, + 0.518 + ], + "angle": 0, + "content": "2 Related Works" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.537, + 0.785, + 0.597 + ], + "angle": 0, + "content": "Single-Image Super-Resolution (SISR) methods use deep neural networks of various architectures, such as residual networks [19,36] and transformers [18, 33]. Despite improving the standard SISR performance with dedicated architectures, these methods struggle to generalize to real-world scenarios [34,40]." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.599, + 0.787, + 0.779 + ], + "angle": 0, + "content": "Real-World Image Super-Resolution (RWSR) aims to address unknown input degradations. The infeasibility of creating paired training data without known kernels presents a significant challenge in learning specialized models for RWSR. There are two primary strategies in existing RWSR. The first assumes that networks trained on diverse and challenging synthetic degradations will effectively generalize to real-world data [35, 43]. Even though real-world degradations are complex, within a specific setting or deployment, they tend to be limited in the domain and less diverse than the training pipelines. Consequently, [45] proposes customizing the synthetic pipeline to better match specific real-world data. Nonetheless, an unresolved gap exists between synthesized and real-world degradations. Instead, our method operates directly on real-world data and avoids the intrinsic gap from synthetic degradation." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.78, + 0.787, + 0.841 + ], + "angle": 0, + "content": "The second strategy connects real-world and synthetic degradation with transfer learning. Most works use image-to-image translation [5,30,40] to adapt the synthetic degradation to real-world ones. Such methods resort to intricate designs like CycleGAN [50], which struggle to replicate real-world degradations [34]" + } + ], + [ + { + "type": "page_number", + "bbox": [ + 0.218, + 0.116, + 0.23, + 0.127 + ], + "angle": 0, + "content": "4" + }, + { + "type": "header", + "bbox": [ + 0.272, + 0.115, + 0.374, + 0.129 + ], + "angle": 0, + "content": "Y. Zhang et al." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.147, + 0.784, + 0.177 + ], + "angle": 0, + "content": "reliably. Our proposed method explores feature distances between model outputs and bypasses the need to mimic real-world degradations." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.177, + 0.785, + 0.269 + ], + "angle": 0, + "content": "Knowledge Distillation is initially introduced for model compression [10]. In low-level vision, several works [7, 28, 47] distill large models into more efficient ones by enforcing similarity between their internal features or predictions. Recently, knowledge distillation has served diverse purposes in high-level tasks like segmentation, including domain adaptation [6, 15, 27] and transfer learning [38, 41]. However, such applications are less discussed in low-level vision." + }, + { + "type": "title", + "bbox": [ + 0.216, + 0.291, + 0.331, + 0.307 + ], + "angle": 0, + "content": "3 Method" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.322, + 0.784, + 0.459 + ], + "angle": 0, + "content": "Our method adapts the specialist in synthetic degradation to an unlabeled real-world domain. We use the low-level characteristics of synthetic predictions from the specialists and the knowledge of the generalist model. Fig. 2 is the overview of our approach and the novel Pairwise Distance Distillation (PDD) method, which enforces the consistency of the intra- and inter-model distances to improve the specialist \\((M_S)\\)'s real-world predictions. Sec. 3.1 gives notions used in this section and Sec. 3.2 illustrates the unsupervised formulation. Sec. 3.3 explains the details of PDD and Sec. 3.4 provides the full optimization aims and discusses the static and Exponential Moving Average (EMA) versions of our method." + }, + { + "type": "title", + "bbox": [ + 0.216, + 0.481, + 0.427, + 0.496 + ], + "angle": 0, + "content": "3.1 Definitions & Setup" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.504, + 0.785, + 0.611 + ], + "angle": 0, + "content": "When provided with an LR input \\( X \\in \\mathbb{R}^{H \\times W \\times C} \\) - where \\( H \\times W \\) is the spatial resolution and \\( C \\) is the number of color channels - a SISR model aims to reconstruct an HR image \\( \\hat{Y} \\in \\mathbb{R}^{rH \\times rW \\times C} \\), scaling by a factor \\( r \\). Due to the difficulty in obtaining paired training data, the conventional approach in SISR involves generating an LR image \\( X \\) by downsampling the HR ground-truth \\( Y \\), i.e., \\( X = D_{\\downarrow}(Y) \\). This downsampling operation \\( D_{\\downarrow} \\) typically refers to bicubic interpolation, although other alternatives exist [1]." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.611, + 0.785, + 0.672 + ], + "angle": 0, + "content": "In a more extensive setting, it is typical for the LR image to be affected by various factors, such as blurring, noise, compression artifacts, etc. To exhibit these degradations, the LR image \\( X \\) can be formed by applying \\( D_{\\downarrow} \\) along with one or multiple degradations \\( D_{i} \\), in any order, i.e.," + }, + { + "type": "equation", + "bbox": [ + 0.423, + 0.684, + 0.785, + 0.7 + ], + "angle": 0, + "content": "\\[\nX = D _ {1} \\circ \\dots \\circ D _ {n} (Y). \\tag {1}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.71, + 0.784, + 0.772 + ], + "angle": 0, + "content": "Here, we include \\( D_{\\downarrow} \\) as one of the degradations \\( \\{D_i\\}_{i \\in [1,n]} \\) for simplicity. Now, let's consider two categories of degradation sets consisting of known degradations: \\( \\mathcal{D}_S \\), which comprises simple degradations with a few factors, and \\( \\mathcal{D}_G \\) characterized by complex degradations with a wide range of factors:" + }, + { + "type": "equation", + "bbox": [ + 0.324, + 0.784, + 0.785, + 0.8 + ], + "angle": 0, + "content": "\\[\n\\mathcal {D} _ {S} = \\left\\{D _ {i} \\right\\} _ {i \\in \\mathcal {S}}, \\mathcal {D} _ {G} = \\left\\{D _ {i} \\right\\} _ {i \\in \\mathcal {G}}, \\quad \\text {w h e r e} | \\mathcal {G} | \\gg | \\mathcal {S} |. \\tag {2}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.81, + 0.785, + 0.842 + ], + "angle": 0, + "content": "In the simplest case for \\(\\mathcal{D}_S\\), the set \\(\\mathcal{D}_S = \\{D_\\downarrow\\}\\), which represents the standard SISR setup. For the broader case, \\(\\mathcal{D}_G\\) should encompass diverse and challenging" + } + ], + [ + { + "type": "header", + "bbox": [ + 0.526, + 0.115, + 0.732, + 0.128 + ], + "angle": 0, + "content": "Pairwise Distance Distillation" + }, + { + "type": "page_number", + "bbox": [ + 0.775, + 0.117, + 0.785, + 0.127 + ], + "angle": 0, + "content": "5" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.147, + 0.784, + 0.177 + ], + "angle": 0, + "content": "degradations beyond \\(\\{D_{\\downarrow}\\}\\) to ensure that models trained on this data generalize well to blind settings [35, 43]." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.178, + 0.788, + 0.328 + ], + "angle": 0, + "content": "Consider two models, \\( M_S \\) and \\( M_G \\), trained on datasets constructed with \\( \\mathcal{D}_S \\) and \\( \\mathcal{D}_G \\) respectively. Both models are optimized with common supervised SR learning losses [13, 36] including MSE-based loss, VGG loss, and adversarial training (see Eq. (9)). Since \\( \\mathcal{D}_S \\) comprises a narrower range of degradations compared to \\( \\mathcal{D}_G \\), \\( M_S \\) is a \"specialist\" model for the limited domain defined by \\( \\mathcal{D}_S \\). On the other hand, \\( M_G \\) is a \"generalist\" model as it achieves moderate performance across the broader domain derived by \\( \\mathcal{D}_G \\). Comparing \\( M_S \\) and \\( M_G \\), we assume the specialist's superiority to the generalist in \\( \\mathcal{D}_S \\) and the generalist's better handling of unknown degradations. These assumptions are empirically validated in Tab. 1." + }, + { + "type": "title", + "bbox": [ + 0.215, + 0.351, + 0.627, + 0.367 + ], + "angle": 0, + "content": "3.2 Unsupervised Learning Through Distillation" + }, + { + "type": "text", + "bbox": [ + 0.215, + 0.377, + 0.789, + 0.501 + ], + "angle": 0, + "content": "Our approach aligns with the unsupervised setup [5, 30, 40], wherein we work with a collection of real-world LR images alongside unpaired HR clean images. We denote the set of LR images as \\(\\{X^U\\}\\) with the same unknown degradations \\(\\{D_i\\}_{i\\in \\mathcal{U}}\\). Additionally, we generate LR counterparts \\(\\{X^L\\}\\) from HR images using specific degradations from \\(\\mathcal{D}_S\\), e.g. bicubic interpolation. The specialist \\(M_S\\) is adept in restoring \\(X^L\\), while \\(M_G\\) performs moderately in restoring \\(X^L\\) and \\(X^U\\). We aim to adapt \\(M_S\\) to \\(\\{X^U\\}\\) with the aid of \\(M_G\\). As in [10], a naive distillation approach is having \\(M_S\\) imitate the predictions from \\(M_G\\) for unlabeled inputs:" + }, + { + "type": "equation", + "bbox": [ + 0.319, + 0.51, + 0.785, + 0.55 + ], + "angle": 0, + "content": "\\[\n\\begin{array}{l} \\hat {Y} _ {S} ^ {L} = M _ {S} (X ^ {L}), \\hat {Y} _ {S} ^ {U} = M _ {S} (X ^ {U}), \\hat {Y} _ {G} ^ {U} = M _ {G} (X ^ {U}), \\\\ \\mathcal {L} _ {N D} = \\mathcal {L} _ {L} \\left(\\hat {Y} _ {S} ^ {U}, \\hat {Y} _ {G} ^ {U}\\right) + \\lambda \\mathcal {L} _ {L} \\left(\\hat {Y} _ {S} ^ {L}, Y ^ {L}\\right), \\tag {3} \\\\ \\end{array}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.562, + 0.788, + 0.639 + ], + "angle": 0, + "content": "where \\( Y^{L} \\) represents the ground-truth image in the labeled domain, \\( \\mathcal{L}_L \\) denotes the supervised loss formulated in Eq. (9), and \\( \\lambda \\) is a scale factor to balance between the distillation and primary objective of \\( M_S \\). However, this simple imitation approach relies solely on information from the generalist \\( M_G \\) and fails to harness the strengths of both models." + }, + { + "type": "title", + "bbox": [ + 0.215, + 0.661, + 0.57, + 0.677 + ], + "angle": 0, + "content": "3.3 Pairwise Distance Distillation (PDD)" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.686, + 0.789, + 0.719 + ], + "angle": 0, + "content": "Consider any pair of LR images - a real-world \\( X^U \\) and a synthesized \\( X^L \\). As shown in Fig. 2, we obtain four predictions by applying \\( M_S \\) and \\( M_G \\):" + }, + { + "type": "equation", + "bbox": [ + 0.367, + 0.729, + 0.785, + 0.768 + ], + "angle": 0, + "content": "\\[\n\\begin{array}{l} \\hat {Y} _ {S} ^ {U} = M _ {S} (X ^ {U}), \\hat {Y} _ {G} ^ {U} = M _ {G} (X ^ {U}), \\\\ \\text {a n d} \\hat {Y} _ {S} ^ {L} = M _ {S} \\left(X ^ {L}\\right), \\hat {Y} _ {G} ^ {L} = M _ {G} \\left(X ^ {L}\\right). \\tag {4} \\\\ \\end{array}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.78, + 0.789, + 0.841 + ], + "angle": 0, + "content": "Our distillation explores the relationship between predictions in Eq. (4) to seek for knowledge combination. We use VGG features [31] of the predictions as our basis for exploration. VGG, as a classification model, inherently captures semantic information. Yet it also captures low-level characteristics related to" + } + ], + [ + { + "type": "page_number", + "bbox": [ + 0.218, + 0.116, + 0.23, + 0.127 + ], + "angle": 0, + "content": "6" + }, + { + "type": "header", + "bbox": [ + 0.272, + 0.115, + 0.375, + 0.129 + ], + "angle": 0, + "content": "Y. Zhang et al." + }, + { + "type": "image", + "bbox": [ + 0.248, + 0.146, + 0.758, + 0.248 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.214, + 0.26, + 0.788, + 0.332 + ], + "angle": 0, + "content": "Fig. 2: Schematic of Pairwise Distance Distillation (PDD). \\( X^{U} \\) and \\( X^{L} \\) are unlabeled (real-world) and labeled (synthetic) inputs. Model \\( M_{G} \\) is a generalist trained with extensive synthetic pipeline, while \\( M_{S} \\) specializes in \\( X^{L} \\). The prediction \\( \\hat{Y}_S^L \\) is supervised by its ground truth throughout training. PDD enforces the consistency between \\( \\{\\varDelta_U, \\varDelta_L\\} \\) and between \\( \\{d_S, d_G\\} \\) to improve \\( M_{S} \\)'s real-world performance." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.361, + 0.785, + 0.391 + ], + "angle": 0, + "content": "image quality [37], e.g. blur and sharpness, and is widely used in the perceptual loss for super-resolution tasks [13, 35, 36]." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.392, + 0.785, + 0.498 + ], + "angle": 0, + "content": "As a generalist, \\(M_G\\) outputs synthetic and real-world predictions with small quality gaps, which is characterized by the overlapping distributions of low-level characteristics of \\(\\{\\hat{Y}_G^L\\}\\) and \\(\\{\\hat{Y}_G^U\\}\\) [20, 21]. Following [20], Fig. 5a projects the low-level features of \\(\\{\\hat{Y}_G^L\\}\\) (DIV2K) and \\(\\{\\hat{Y}_G^U\\}\\) (NTIRE20). The generalist's coverage strongly overlaps while the specialist's predictions are well-separated (see Fig. 5b). The within-model relationship of the generalist exhibits the generalizability to real-world degradation." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.499, + 0.787, + 0.621 + ], + "angle": 0, + "content": "On the other side, the high image quality of \\(\\{\\hat{Y}_S^L\\}\\), due to \\(M_S\\)'s specialization, makes its low-level characteristics a valuable reference. To integrate the knowledge for real-world performance, we propose to make the specialist imitate the generalist's within-model relationship, that is, pushing the low-level characteristics of real-world predictions from \\(M_S\\), \\(\\{\\hat{Y}_S^U\\}\\), to be similar to that of its synthetic predictions \\(\\{\\hat{Y}_S^L\\}\\). To achieve this aim, we first assume such similarity is satisfied for \\(M_S\\) and conclude the following consistencies about VGG feature distances between predictions:" + }, + { + "type": "text", + "bbox": [ + 0.222, + 0.633, + 0.784, + 0.661 + ], + "angle": 0, + "content": "1. Intra-model distances for predictions of the same input pair should be consistent across \\( M_S \\) and \\( M_G \\) (green shade in Fig. 2)." + }, + { + "type": "text", + "bbox": [ + 0.222, + 0.663, + 0.784, + 0.692 + ], + "angle": 0, + "content": "2. Inter-model distances for predictions of a single input should be consistent across synthetic and real-world domains (orange shade in Fig. 2)." + }, + { + "type": "list", + "bbox": [ + 0.222, + 0.633, + 0.784, + 0.692 + ], + "angle": 0, + "content": null + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.704, + 0.785, + 0.809 + ], + "angle": 0, + "content": "Our rationales for these consistencies are as follows. First, we recall the low-level characteristics of predictions from the same model are similar. We discussed it for \\(M_G\\) with Fig. 5a and made such an assumption for \\(M_S\\). Thus, the distance between the same model's predictions \\(\\{\\hat{Y}_G^L,\\hat{Y}_G^U\\}\\) (or \\(\\{\\hat{Y}_S^L,\\hat{Y}_S^U\\}\\)), i.e. intra-model distance, reflects mainly semantic differences due to the close low-level characteristics. As it naturally follows that the applied SR models should not change semantics, intra-model distances should be consistent for \\(M_S\\) and \\(M_G\\)." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.81, + 0.787, + 0.842 + ], + "angle": 0, + "content": "Second, we note the low-level characteristics of predictions from the two models differ, e.g. \\(\\hat{Y}_S^L\\) has higher quality than the counterpart \\(\\hat{Y}_G^L\\). The distance" + } + ], + [ + { + "type": "header", + "bbox": [ + 0.526, + 0.115, + 0.732, + 0.128 + ], + "angle": 0, + "content": "Pairwise Distance Distillation" + }, + { + "type": "page_number", + "bbox": [ + 0.775, + 0.116, + 0.785, + 0.127 + ], + "angle": 0, + "content": "7" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.145, + 0.784, + 0.221 + ], + "angle": 0, + "content": "between the two predictions \\(\\{\\hat{Y}_G^L,\\hat{Y}_S^L\\}\\) (or \\(\\{\\hat{Y}_G^U,\\hat{Y}_S^U\\}\\)) of a given sample, i.e. inter-model distance, captures such differences in low-level characteristics, as semantics stay constant for the predictions of the same input. Given the similarity of low-level characteristics within a single model, the inter-model distances for synthetic and real-world samples should be consistent." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.223, + 0.784, + 0.284 + ], + "angle": 0, + "content": "We structure the two consistencies as distillations on the intra- and inter-model distances. Our method encourages \\( M_S \\)'s real-world predictions \\(\\{\\hat{Y}_S^U\\}\\) to have similar low-level characteristics as its synthetic predictions \\(\\{\\hat{Y}_S^L\\}\\), taking it as a reference from high-quality images." + }, + { + "type": "text", + "bbox": [ + 0.215, + 0.285, + 0.785, + 0.358 + ], + "angle": 0, + "content": "Intra-model Distance Distillation enforces consistency on the distances between predictions from the same model (green shading in Fig. 2). Consider the \\(\\ell_1\\) distance \\(d_G^{ij}\\) between predictions from the generalist \\(M_G\\), i.e. \\(\\hat{Y}_G^L\\) and \\(\\hat{Y}_G^U\\), in VGG feature space, and the distance \\(d_S^{ij}\\) for the specialist \\(M_S\\) is similarly defined:" + }, + { + "type": "equation", + "bbox": [ + 0.399, + 0.359, + 0.785, + 0.386 + ], + "angle": 0, + "content": "\\[\nd _ {G} ^ {i j} = \\left\\| \\Phi_ {i j} \\left(\\hat {Y} _ {G} ^ {L}\\right) - \\Phi_ {i j} \\left(\\hat {Y} _ {G} ^ {U}\\right) \\right\\| _ {1}, \\tag {5}\n\\]" + }, + { + "type": "equation", + "bbox": [ + 0.399, + 0.38, + 0.606, + 0.397 + ], + "angle": 0, + "content": "\\[\nd _ {S} ^ {i j} = \\left\\| \\Phi_ {i j} \\left(\\hat {Y} _ {S} ^ {L}\\right) - \\Phi_ {i j} \\left(\\hat {Y} _ {S} ^ {U}\\right) \\right\\| _ {1},\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.215, + 0.399, + 0.784, + 0.463 + ], + "angle": 0, + "content": "where \\(\\varPhi_{ij}\\) refers to the \\(j\\)-th layer in the \\(i\\)-th residual block of VGG19. For \\(d_G^{ij}, d_S^{ij} \\in \\mathbb{R}^{c \\times h \\times w}, h \\times w\\) is the spatial resolution and \\(c\\) is the number of channels. We enforce the consistency between \\(d_G^{ij}\\) and \\(d_S^{ij}\\) by minimizing their difference measured with the Cross-Entropy (CE) \\(R_{intra}\\):" + }, + { + "type": "equation", + "bbox": [ + 0.324, + 0.47, + 0.784, + 0.505 + ], + "angle": 0, + "content": "\\[\nR _ {i n t r a} = - \\frac {1}{h w} \\sum_ {i, j} \\sum_ {m, n} \\mathbf {S} \\left(d _ {G} ^ {i j} [ m, n ]\\right) \\log \\mathbf {S} \\left(d _ {S} ^ {i j} [ m, n ]\\right), \\tag {6}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.215, + 0.511, + 0.784, + 0.571 + ], + "angle": 0, + "content": "where \\([m,n]\\) is a spatial index of the feature map, and \\(\\mathbf{S}(\\cdot)\\) denotes the SoftMax function with the input size being \\(\\mathbb{R}^c\\). The element-wise CE reflects the negative log-likelihood of distances locally. While other measures are feasible, CE is empirically a good choice in favoring the overall sharpness (see Sec. 4.4)." + }, + { + "type": "text", + "bbox": [ + 0.215, + 0.571, + 0.785, + 0.678 + ], + "angle": 0, + "content": "Inter-model Distance Distillation enforces the consistency of changes in low-level characteristics between predictions from different models (orange shading in Fig. 2). Given the two predictions for a single input, i.e. \\(\\{\\hat{Y}_G^L,\\hat{Y}_S^L\\}\\) (or \\(\\{\\hat{Y}_G^U,\\hat{Y}_S^U\\}\\)), we first calculate their feature distance \\(\\varDelta_L^{ij}\\) (or \\(\\varDelta_U^{ij}\\)) to represent the differences in low-level characteristics. However, enforcing consistency on \\(\\varDelta_L^{ij}\\) and \\(\\varDelta_U^{ij}\\) is less straightforward, as the spatial layout of corresponding inputs can differ. Thus, we compute the inter-model distances with the following:" + }, + { + "type": "equation", + "bbox": [ + 0.381, + 0.684, + 0.785, + 0.71 + ], + "angle": 0, + "content": "\\[\n\\Delta_ {L} ^ {i j} = \\operatorname {G r a m} \\left(\\Phi_ {i j} \\left(\\hat {Y} _ {S} ^ {L}\\right) - \\Phi_ {i j} \\left(\\hat {Y} _ {G} ^ {L}\\right)\\right), \\tag {7}\n\\]" + }, + { + "type": "equation", + "bbox": [ + 0.381, + 0.706, + 0.623, + 0.723 + ], + "angle": 0, + "content": "\\[\n\\Delta_ {U} ^ {i j} = \\operatorname {G r a m} \\left(\\Phi_ {i j} \\left(\\hat {Y} _ {S} ^ {U}\\right) - \\Phi_ {i j} \\left(\\hat {Y} _ {G} ^ {U}\\right)\\right).\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.215, + 0.729, + 0.784, + 0.773 + ], + "angle": 0, + "content": "Gram(\\cdot) refers to the Gram matrix, which calculates the correlations among vectorized feature maps along the channel dimension. The Gram matrix captures the statistics while collapsing the spatial layout [8]." + }, + { + "type": "text", + "bbox": [ + 0.215, + 0.774, + 0.784, + 0.806 + ], + "angle": 0, + "content": "For distillation, we ensure the consistency between inter-model distances by minimizing \\( R_{inter} \\), the Frobenius norm between \\( \\Delta_U^{ij} \\) and \\( \\Delta_L^{ij} \\):" + }, + { + "type": "equation", + "bbox": [ + 0.406, + 0.812, + 0.784, + 0.843 + ], + "angle": 0, + "content": "\\[\nR _ {i n t e r} = \\sum_ {i, j} \\| \\Delta_ {U} ^ {i j} - \\Delta_ {L} ^ {i j} \\| _ {\\mathcal {F}}, \\tag {8}\n\\]" + } + ], + [ + { + "type": "page_number", + "bbox": [ + 0.218, + 0.116, + 0.23, + 0.127 + ], + "angle": 0, + "content": "8" + }, + { + "type": "header", + "bbox": [ + 0.272, + 0.115, + 0.375, + 0.129 + ], + "angle": 0, + "content": "Y. Zhang et al." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.147, + 0.788, + 0.18 + ], + "angle": 0, + "content": "where \\(\\| \\cdot \\|_{\\mathcal{F}}\\) computes the Frobenius norm. The Frobenius norm of Gram Matrix difference is interpreted in [8, 16] as measuring distribution discrepancy." + }, + { + "type": "title", + "bbox": [ + 0.216, + 0.199, + 0.367, + 0.214 + ], + "angle": 0, + "content": "3.4 Full Method" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.224, + 0.784, + 0.253 + ], + "angle": 0, + "content": "The network \\( M_S \\) is fully optimized with supervised losses for labeled synthetic data \\( X^L \\) and unsupervised losses for real-world data \\( X^U \\)." + }, + { + "type": "text", + "bbox": [ + 0.215, + 0.254, + 0.788, + 0.332 + ], + "angle": 0, + "content": "Supervised Losses maintain the specialization in synthetic degradation and reduce overfitting to the distillation. With the ground-truth image for prediction \\(\\hat{Y}_S^L\\), denoted as \\(Y^{L}\\), we implement a combination of supervised loss functions. It includes a wavelet-based loss \\(L_{\\mathrm{wv}}\\), a perceptual loss \\(L_{\\mathrm{vgg}}\\), and generative loss \\(L_{\\mathrm{gan}}\\) for adversarial training:" + }, + { + "type": "equation", + "bbox": [ + 0.271, + 0.343, + 0.786, + 0.361 + ], + "angle": 0, + "content": "\\[\n\\mathcal {L} _ {L} \\left(\\hat {Y} _ {S} ^ {L}, Y ^ {L}\\right) = \\alpha_ {1} L _ {\\mathrm {w v}} \\left(\\hat {Y} _ {S} ^ {L}, Y ^ {L}\\right) + \\alpha_ {2} L _ {\\mathrm {v g g}} \\left(\\hat {Y} _ {S} ^ {L}, Y ^ {L}\\right) + \\alpha_ {3} L _ {\\mathrm {g a n}} \\left(\\hat {Y} _ {S} ^ {L}\\right), \\tag {9}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.215, + 0.37, + 0.788, + 0.418 + ], + "angle": 0, + "content": "where \\( L_{\\mathrm{vgg}} \\) and \\( L_{\\mathrm{gan}} \\) are adopted from RealESRGAN [35], with balancing weights \\( \\alpha_{1} \\), \\( \\alpha_{2} \\), and \\( \\alpha_{3} \\). We use but omit writing the discriminator loss for brevity. \\( L_{\\mathrm{wv}} \\) refers to an \\( \\ell_{1} \\)-based wavelet loss [4, 49]:" + }, + { + "type": "equation", + "bbox": [ + 0.341, + 0.429, + 0.786, + 0.458 + ], + "angle": 0, + "content": "\\[\nL _ {\\mathrm {w v}} \\left(\\hat {Y} _ {S} ^ {L}, Y ^ {L}\\right) = \\sum_ {i} \\omega_ {i} \\| W \\left(\\hat {Y} _ {S} ^ {L}\\right) _ {i} - W \\left(Y ^ {L}\\right) _ {i} \\| _ {1}, \\tag {10}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.469, + 0.784, + 0.498 + ], + "angle": 0, + "content": "where \\( W(\\cdot)_i \\) extracts the \\( i \\)-th wavelet channel, and \\( \\omega_i \\) weights its importance. For the wavelet transform, we apply the Haar wavelet [14] for all output channels." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.5, + 0.785, + 0.56 + ], + "angle": 0, + "content": "Unsupervised Losses. In the absence of ground truth for \\(\\dot{Y}_S^U\\), we opt to optimize for the consistency outlined in Eq. (6) and Eq. (8) as regularizers. To fully utilize the discriminator's knowledge of realness, we also incorporate the generative loss in Eq. (9):" + }, + { + "type": "equation", + "bbox": [ + 0.354, + 0.571, + 0.786, + 0.589 + ], + "angle": 0, + "content": "\\[\n\\mathcal {L} _ {U} = \\lambda_ {1} R _ {\\text {i n t r a}} + \\lambda_ {2} R _ {\\text {i n t e r}} + \\lambda_ {3} L _ {\\text {g a n}} \\left(\\hat {Y} _ {S} ^ {U}\\right), \\tag {11}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.599, + 0.784, + 0.659 + ], + "angle": 0, + "content": "where \\(\\lambda_1, \\lambda_2\\), and \\(\\lambda_3\\) balance the optimization aims. \\(L_{gan}\\) here uses the same discriminator in Eq. (9). The supervised loss in Eq. (9) and the unsupervised distillation loss in Eq. (11) are combined as our final optimization objective, expressed as \\(\\mathcal{L} = \\mathcal{L}_L + \\mathcal{L}_U\\)." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.66, + 0.785, + 0.735 + ], + "angle": 0, + "content": "Color Correction. In practice, our approach often leads to color shifts due to the regularization of distances within the feature space. To address this, we rectify the output by normalizing the mean and variance of each color channel with those of the corresponding input channels [35]. Further details are provided in Supplementary Sec. E." + }, + { + "type": "text", + "bbox": [ + 0.215, + 0.735, + 0.786, + 0.81 + ], + "angle": 0, + "content": "Learning Configurations. Our proposed PDD is a learning framework that can utilize pre-trained networks. It is recommended that \\(M_G\\) is pre-trained with extensive degradations \\(\\mathcal{D}_G\\), as in [35,43]. The static configuration (refer to Fig. 3 (a)) initializes \\(M_S\\) with the model pre-trained on specific synthetic degradation \\((\\mathcal{D}_S)\\), and \\(M_G\\) is frozen throughout the training process." + }, + { + "type": "text", + "bbox": [ + 0.215, + 0.81, + 0.787, + 0.842 + ], + "angle": 0, + "content": "Alternative configuration initializes \\( M_G \\) and \\( M_S \\) with the same model pretrained with \\( \\mathcal{D}_G \\) and maintain \\( M_G \\) as an Exponential Moving Average (EMA)" + } + ], + [ + { + "type": "header", + "bbox": [ + 0.526, + 0.115, + 0.732, + 0.128 + ], + "angle": 0, + "content": "Pairwise Distance Distillation" + }, + { + "type": "page_number", + "bbox": [ + 0.775, + 0.116, + 0.787, + 0.127 + ], + "angle": 0, + "content": "9" + }, + { + "type": "image", + "bbox": [ + 0.348, + 0.146, + 0.501, + 0.241 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.39, + 0.243, + 0.46, + 0.253 + ], + "angle": 0, + "content": "(a) Static Version" + }, + { + "type": "image", + "bbox": [ + 0.519, + 0.144, + 0.661, + 0.241 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.556, + 0.244, + 0.625, + 0.253 + ], + "angle": 0, + "content": "(b) EMA Version" + }, + { + "type": "image_caption", + "bbox": [ + 0.215, + 0.269, + 0.788, + 0.34 + ], + "angle": 0, + "content": "Fig. 3: Two initialization options for the generalist and specialist models. (a) The static configuration initializes \\( M_G \\) with a model pre-trained by the complex synthetic pipeline and \\( M_S \\) with a model pre-trained by simple degradation in \\( X^L \\). Weights of \\( M_G \\) are frozen during distillation. (b) Both \\( M_S \\) and \\( M_G \\) are initialized with a pre-trained generalized model. Weights of \\( M_G \\) are the EMA version of \\( M_S \\)." + }, + { + "type": "text", + "bbox": [ + 0.215, + 0.368, + 0.788, + 0.445 + ], + "angle": 0, + "content": "version of \\( M_S \\) (see Fig. 3 (b)). Despite \\( M_S \\) not being initialized as a specialist, it can readily specialize to the synthetic degradation \\( \\mathcal{D}_S \\) through supervised learning in Eq. (9). Empirically, the EMA version tends to be more effective, likely because the updated knowledge of real-world data in \\( M_S \\) also benefits \\( M_G \\) through EMA, enabling \\( M_G \\) to yield better references for optimization." + }, + { + "type": "title", + "bbox": [ + 0.216, + 0.467, + 0.376, + 0.484 + ], + "angle": 0, + "content": "4 Experiments" + }, + { + "type": "title", + "bbox": [ + 0.216, + 0.498, + 0.331, + 0.514 + ], + "angle": 0, + "content": "4.1 Settings" + }, + { + "type": "text", + "bbox": [ + 0.215, + 0.523, + 0.788, + 0.674 + ], + "angle": 0, + "content": "Datasets. We experiment on real-world datasets, RealSR [2] and DRealSR [39], and a synthetic dataset without released degradation settings, NTIRE20 [23]. For experiments on RealSR and DRealSR, we synthesize labeled training data \\(\\{X^L\\}\\) from high-resolution images in the DF2K dataset [1,32]. For the NTIRE20 dataset, we use the provided target clean images to synthesize low-resolution counterparts. For synthesizing \\(\\{X^L\\}\\), we use only bicubic interpolation as the default and investigate other options for \\(\\mathcal{D}_S\\) in the Supplementary Sec. D.b. RealSR or DRealSR has multiple data sources, and we optimize our methods for each source separately. For DRealSR, we report results for Panasonic and Olympus sources, as the others do not have sufficient training samples." + }, + { + "type": "text", + "bbox": [ + 0.215, + 0.675, + 0.788, + 0.779 + ], + "angle": 0, + "content": "Training Details. We crop patches of size \\(48 \\times 48\\) as inputs and train the model with the ADAM optimizer [12] using the settings \\(\\beta_{1} = 0.9\\), \\(\\beta_{2} = 0.999\\), and \\(\\epsilon = 10^{-8}\\). The batch size is 16, including equal numbers of labeled and unlabeled inputs. The initial learning rate is \\(1e - 4\\) and halved after \\(25\\mathrm{K}\\) iterations. The static version needs \\(50\\mathrm{K}\\) iterations, and the EMA version is trained for \\(100\\mathrm{K}\\) iterations due to the initial specialist underfitting to the synthetic domain. All models are for \\(\\times 4\\) super-resolution. Other details are in Supplementary Sec. A." + }, + { + "type": "text", + "bbox": [ + 0.215, + 0.78, + 0.788, + 0.84 + ], + "angle": 0, + "content": "Evaluation. We evaluate models on testing sets of RealSR, DRealSR, and the validation set of NTIRE20. We crop non-overlapped 512x512 patches from LR images in DRealSR to avoid memory exhaustion. For other datasets, we use the full images. To fully use the provided high-resolution counterparts, we report" + } + ], + [ + { + "type": "page_number", + "bbox": [ + 0.218, + 0.116, + 0.236, + 0.127 + ], + "angle": 0, + "content": "10" + }, + { + "type": "header", + "bbox": [ + 0.272, + 0.115, + 0.375, + 0.129 + ], + "angle": 0, + "content": "Y. Zhang et al." + }, + { + "type": "table_caption", + "bbox": [ + 0.214, + 0.145, + 0.788, + 0.203 + ], + "angle": 0, + "content": "Table 1: \\( M_G \\) and \\( M_S \\) (static) columns show initial performances on the bicubic-interpolated data \\( \\{X^L\\} \\) from Set14 [42] and \\( \\{X^U\\} \\) with unknown degradation from NTIRE20 [23]. \\( M_S \\) (EMA) column shows the performance of \\( M_S \\) in the EMA version on \\( \\{X^L\\} \\) after training process." + }, + { + "type": "table", + "bbox": [ + 0.262, + 0.213, + 0.742, + 0.253 + ], + "angle": 0, + "content": "
DomainMetricMGMS(static)MS(EMA)DomainMetricMGMS(static)
{X^L}PSNR↑23.9824.8126.26{X^U}PSNR↑25.0819.82
LPIPS↓0.23490.13370.1471LPIPS↓0.25040.7552
" + }, + { + "type": "image", + "bbox": [ + 0.228, + 0.267, + 0.406, + 0.372 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.408, + 0.269, + 0.596, + 0.373 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.601, + 0.268, + 0.785, + 0.372 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.214, + 0.383, + 0.788, + 0.454 + ], + "angle": 0, + "content": "Fig. 4: Improvements over the Generalist on five unlabeled data domains, where a lower LPIPS score is better. ND improves fidelity scores (PSNR) but dramatically drops perceptual scores (LPIPS and NRQM). Both versions of our method achieve better improvements than ND for all reported metrics. For each domain, the Static version improves at least one of the perceptual metrics; the EMA version improves all." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.484, + 0.788, + 0.607 + ], + "angle": 0, + "content": "scores of full-reference metrics, PSNR, SSIM, and LPIPS [44]. The PSNR and SSIM scores are calculated on the Y-channel of YCbCr format images. Additionally, we follow BSRGAN [43] and supplement with the no-reference metric NRQM [24] to evaluate the perceptual clearness, as the ground truth in real-world captured datasets [2,39] also feature some blur. We use the implementation in IQA-PyTorch3 for all metrics. It is important to consider both fidelity and perceptual scores; fidelity metrics are worse at detecting blurriness while perceptual metrics are not sensitive to high-frequency artifacts [25, 26, 29, 48]." + }, + { + "type": "title", + "bbox": [ + 0.215, + 0.632, + 0.439, + 0.646 + ], + "angle": 0, + "content": "4.2 Effectiveness of PDD" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.661, + 0.788, + 0.814 + ], + "angle": 0, + "content": "Comparisons. We compare the two versions of our method in Fig. 3 with the naive distillation defined in Eq. (3) regarding their improvements over the generalist network \\( M_G \\). We choose RealESRGAN [35] as the pre-trained generalist model and ESRGAN [36], a standard SISR model, as the pre-trained specialist for the naive distillation and the static version in Fig. 3 (a). Tab. 1 shows the initial performances of the generalist and specialist for \\( \\{X^L\\} \\) and \\( \\{X^U\\} \\). As we assumed in Sec. 3.1, the specialist performs better than the generalist for \\( \\{X^L\\} \\) while \\( M_G \\) generalizes better for \\( \\{X^U\\} \\). For the EMA version in Fig. 3 (b), \\( M_S \\) as well as \\( M_G \\) is initialized with RealESRGAN. Through the supervised learning in Eq. (9), \\( M_S(\\mathrm{EMA}) \\) shows specialized performance for labeled data \\( \\{X^L\\} \\)." + }, + { + "type": "page_footnote", + "bbox": [ + 0.218, + 0.825, + 0.533, + 0.841 + ], + "angle": 0, + "content": "3 https://github.com/chaofengc/IQA-PyTorch" + } + ], + [ + { + "type": "header", + "bbox": [ + 0.526, + 0.115, + 0.732, + 0.128 + ], + "angle": 0, + "content": "Pairwise Distance Distillation" + }, + { + "type": "page_number", + "bbox": [ + 0.769, + 0.116, + 0.784, + 0.127 + ], + "angle": 0, + "content": "11" + }, + { + "type": "image", + "bbox": [ + 0.264, + 0.146, + 0.416, + 0.232 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.297, + 0.236, + 0.381, + 0.248 + ], + "angle": 0, + "content": "(a) Generalist" + }, + { + "type": "image", + "bbox": [ + 0.427, + 0.146, + 0.579, + 0.232 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.438, + 0.236, + 0.566, + 0.248 + ], + "angle": 0, + "content": "(b) Bicubic Specialist" + }, + { + "type": "image", + "bbox": [ + 0.59, + 0.146, + 0.742, + 0.232 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.599, + 0.236, + 0.731, + 0.248 + ], + "angle": 0, + "content": "(c) Adapted Specialist" + }, + { + "type": "image_caption", + "bbox": [ + 0.214, + 0.26, + 0.788, + 0.33 + ], + "angle": 0, + "content": "Fig. 5: Visualization of low-level features for predictions of DIV2K [1] (bicubic) and NTIRE20 (unknown) following [20] (a) Generalist's predictions for the two domains has overlapped distribution. (b) The predictions from the specialist that only adept in bicubic interpolation have separated distributions. (c) After applying our method (static version), predictions for two domains are pushed close." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.358, + 0.788, + 0.463 + ], + "angle": 0, + "content": "Fig. 4 plots the improvements over the pretrained generalist for Naive Distillation (ND), Static, and EMA versions of our method. Compared to the generalist, ND dramatically drops in perceptual scores (LPIPS, NRQM) despite maintaining or improving the PSNR. Both versions of our method perform better than ND; compared to the generalist, our static version improves PSNR significantly and at least one perceptual score for each data domain. The EMA version improves all metrics with the highest magnitudes." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.464, + 0.788, + 0.614 + ], + "angle": 0, + "content": "Change in Low-Level Characteristics. We extract low-level features for visualization following [20]. Fig. 5b visualize predictions generated by the bicubic specialist (ESRGAN), which has separate distributions. Fig. 5c verifies the predictions from the specialist after applying our method (static version), where the separated distributions are pushed closer. Visualizing low-level differences between images with arbitrary content is challenging [20]. As such, we also quantitatively compare the KL divergence between the low-level features. As shown in Fig. 6, applying our method decreases the KL divergence between labeled (DIV2K-bicubic) and unlabeled predictions. Details of the visualization and KL-diversity calculations are in Supplementary Sec. B." + }, + { + "type": "title", + "bbox": [ + 0.215, + 0.635, + 0.544, + 0.65 + ], + "angle": 0, + "content": "4.3 Comparisons with Other Methods" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.659, + 0.788, + 0.764 + ], + "angle": 0, + "content": "Quantitative Comparison. We consider two variants of our method by applying the EMA version on RealESRGAN [35] and BSRGAN [43]. In Tab. 2, we compare their performances with other state-of-the-art RWSR methods, including FeMaSR [3], EDAN [17], DASR [40], DAN [11], RealESRGAN [35], BSRGAN [43], and TG [46], which released code and model weights that enable reproduction. We regenerate results to evaluate these methods, while Supplementary Sec. C discusses related but not directly comparable methods." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.765, + 0.788, + 0.84 + ], + "angle": 0, + "content": "As shown in Tab. 2, EDAN, DAN, and DASR perform strongly on the fidelity metrics (PSNR, SSIM) but are poor perceptually, especially on the NRQM score. The qualitative comparison highlights the blurry effects. Our method enhances RealESRGAN and BSRGAN across all metrics and demonstrates superior performance compared to TG when both are applied to RealESRGAN. For FeMaSR," + } + ], + [ + { + "type": "page_number", + "bbox": [ + 0.218, + 0.116, + 0.236, + 0.127 + ], + "angle": 0, + "content": "12" + }, + { + "type": "header", + "bbox": [ + 0.272, + 0.115, + 0.375, + 0.129 + ], + "angle": 0, + "content": "Y. Zhang et al." + }, + { + "type": "image", + "bbox": [ + 0.264, + 0.153, + 0.418, + 0.25 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.215, + 0.264, + 0.477, + 0.349 + ], + "angle": 0, + "content": "Fig. 6: KL divergence between specialist's unlabeled and labeled predictions. The KL divergence between labeled and unlabeled predictions is lower after applying our method (red) than that in the specialist before adaption (blue)." + }, + { + "type": "image", + "bbox": [ + 0.507, + 0.147, + 0.788, + 0.228 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.497, + 0.24, + 0.788, + 0.352 + ], + "angle": 0, + "content": "Fig.7: User study on outputs of 30 real-world images. We recruit 30 evaluators for each image. Each vote is for a better quality within a pair of options. We compared our method to RealESRGAN and BSRGAN by using them as pre-trained models; FeMaSR is compared to RealESRGAN+ours. Our method gains at least \\(65\\%\\) votes." + }, + { + "type": "table_caption", + "bbox": [ + 0.215, + 0.378, + 0.788, + 0.448 + ], + "angle": 0, + "content": "Table 2: Quantitative comparison with state-of-the-art methods. DAN, EDAN, and DASR are separately listed as they favor fidelity metrics but yield poor perceptual scores. For other methods, **bold** and **underline** mark the best and second-best scores. Subscripts in the last two columns show their difference from the corresponding pretrained generalist, where improvement is colored red and drop is blue." + }, + { + "type": "table", + "bbox": [ + 0.221, + 0.461, + 0.789, + 0.688 + ], + "angle": 0, + "content": "
DatasetMetricDAN [11]EDAN [17]DASR [40]FeMaSR [3]RealESRGAN [43]BSRGAN [35]RealESRGAN +TG [46]RealESRGAN +OursBSRGAN +Ours
NTIRE20 -ValidPSNR↑26.8326.59-23.5425.0825.4425.5626.42(+1.340)26.37(+0.930)
SSIM↑0.71710.7400-0.66650.70610.69840.71930.7297(+0.024)0.7238(+0.025)
LPIPS↓0.57470.2678-0.23600.25040.26450.24200.2475(-0.003)0.2595(-0.005)
NRQM↑4.87845.8481-6.62456.12136.27795.79266.2263(+0.105)6.2781(+0.0002)
RealSR -CanonPSNR↑26.6726.3626.7124.2924.7425.5725.0325.94(+1.200)26.13(+0.560)
SSIM↑0.77400.77740.77820.74600.76340.76830.77200.7733(+0.010)0.7626(-0.006)
LPIPS↓0.40950.30260.25070.28090.26070.25730.26150.2516(-0.009)0.2517(-0.006)
NRQM↑3.04073.63603.93836.01516.06496.02935.42506.0647(-0.0002)6.1323(+0.103)
RealSR -NikonPSNR↑26.2925.3725.6323.9424.3124.7924.5625.28(+0.970)25.07(+0.280)
SSIM↑0.75330.74600.74730.70970.74060.73340.74060.7470(+0.014)0.7294(-0.001)
LPIPS↓0.41330.32000.27850.30460.28510.27970.29490.2802(-0.005)0.2713(-0.008)
NRQM↑3.20564.48724.49525.94885.66855.93875.19425.9044(+0.236)6.1276(+0.189)
DRealSR -PanasonicPSNR↑28.9928.12-25.0227.0326.9327.5127.87(+0.840)27.69(+0.760)
SSIM↑0.81230.8006-0.71180.77820.76210.78430.7939(+0.016)0.7752(+0.013)
LPIPS↓0.42350.2747-0.30880.27030.28200.28360.2682(-0.002)0.2687(-0.013)
NRQM↑2.81314.5915-6.03425.37515.58654.64835.3540(-0.021)5.5608(-0.026)
DRealSR -OlympusPSNR↑28.7427.85-24.3826.7326.4927.4327.40(+0.670)27.25(+0.764)
SSIM↑0.80630.7979-0.66910.76670.75680.78430.7846(+0.018)0.7635(+0.007)
LPIPS↓0.45950.3555-0.40070.31590.32830.34160.3139(-0.002)0.3268(-0.001)
NRQM↑2.79613.7989-6.25905.25305.53764.14595.3722(+0.120)5.5957(+0.058)
" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.729, + 0.785, + 0.759 + ], + "angle": 0, + "content": "we achieve competitive perceptual scores and significantly surpass their fidelity scores." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.765, + 0.788, + 0.841 + ], + "angle": 0, + "content": "User Study is conducted by presenting two reconstructions of the same LR image and asking for the better one. We compare the EMA version of our method with methods RealESRGAN, BSRGAN, and FeMaSR. Each comparison is conducted on 30 different real-world results among 30 evaluators. As shown in Fig. 7, at least \\(65\\%\\) evaluators vote for our method in each comparison. Although quan-" + } + ], + [ + { + "type": "header", + "bbox": [ + 0.527, + 0.115, + 0.732, + 0.127 + ], + "angle": 0, + "content": "Pairwise Distance Distillation" + }, + { + "type": "page_number", + "bbox": [ + 0.769, + 0.116, + 0.785, + 0.127 + ], + "angle": 0, + "content": "13" + }, + { + "type": "image", + "bbox": [ + 0.221, + 0.144, + 0.308, + 0.38 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.244, + 0.382, + 0.286, + 0.393 + ], + "angle": 0, + "content": "(a) LR" + }, + { + "type": "image", + "bbox": [ + 0.316, + 0.144, + 0.403, + 0.38 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.323, + 0.382, + 0.396, + 0.394 + ], + "angle": 0, + "content": "(b) FeMaSR" + }, + { + "type": "image", + "bbox": [ + 0.411, + 0.144, + 0.498, + 0.38 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.425, + 0.382, + 0.484, + 0.393 + ], + "angle": 0, + "content": "(c) DASR" + }, + { + "type": "image", + "bbox": [ + 0.506, + 0.144, + 0.593, + 0.38 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.528, + 0.382, + 0.571, + 0.394 + ], + "angle": 0, + "content": "(d) RE" + }, + { + "type": "image", + "bbox": [ + 0.601, + 0.144, + 0.688, + 0.38 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.605, + 0.382, + 0.686, + 0.393 + ], + "angle": 0, + "content": "(e) RE+ours" + }, + { + "type": "image", + "bbox": [ + 0.696, + 0.144, + 0.783, + 0.38 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.697, + 0.382, + 0.779, + 0.393 + ], + "angle": 0, + "content": "(f) BG+ours" + }, + { + "type": "image_caption", + "bbox": [ + 0.215, + 0.405, + 0.785, + 0.447 + ], + "angle": 0, + "content": "Fig. 8: Qualitative comparisons. LRs are from the RealSR-Canon dataset. Columns (e) and (f) are the results of applying our methods on RealESRGAN (RE) and BSRGAN (BG). FeMaSR and DASR can not predict reasonable details." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.477, + 0.784, + 0.507 + ], + "angle": 0, + "content": "titatively competing, FeMaSR yields unpleasant artifacts to human eyes, further shown in qualitative comparison." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.508, + 0.785, + 0.568 + ], + "angle": 0, + "content": "Qualitative Comparison is in Fig. 8. FeMaSR [3] produces high-frequency artifacts, while outputs from DASR [40] suffer blur or artifacts. Compared to RealESRGAN, our method predicts sharper patterns without introducing distortion. More comparisons are in Supplementary Sec. F." + }, + { + "type": "title", + "bbox": [ + 0.215, + 0.589, + 0.342, + 0.603 + ], + "angle": 0, + "content": "4.4 Ablations" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.614, + 0.785, + 0.734 + ], + "angle": 0, + "content": "Ablation on loss terms reveals the effectiveness of each term in Eq. (11). All experiments take RealESRGAN as the pre-trained model and follow the EMA implementation of our method in Fig. 3 (b). Fig. 9 shows that \\( R_{intra} \\) and \\( R_{inter} \\) relate differently to fidelity and perception. Their combination achieves balanced improvements in both directions, and the \\( L_{\\mathrm{gan}} \\) increases pattern sharpness with a slight trade-off in PSNR. Controlling the ratio between weights of \\( R_{intra} \\) and \\( R_{inter} \\) results in fidelity and perceptual quality trade-offs. We choose their balance in Sec. 4.3 and refer to Supplementary Sec. D.a for more discussion." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.735, + 0.785, + 0.809 + ], + "angle": 0, + "content": "Effectiveness of EMA is further validated by comparing to a SingleFixed version, which also uses RealESRGAN as the pre-trained model for both \\( M_S \\) and \\( M_G \\) but fixes the \\( M_G \\) during training. As shown in Tab. 3, updating the weights of the generalist network through EMA enables higher performance in both fidelity and perception." + }, + { + "type": "text", + "bbox": [ + 0.215, + 0.81, + 0.785, + 0.842 + ], + "angle": 0, + "content": "Choice of \\( R_{intra} \\) measurement. Eq. (6) use Cross-Entropy (CE) for measuring the difference between \\( d_S^{ij} \\) and \\( d_G^{ij} \\). Here, we alternate the CE to \\( \\ell_1 \\) distance" + } + ], + [ + { + "type": "page_number", + "bbox": [ + 0.218, + 0.116, + 0.236, + 0.127 + ], + "angle": 0, + "content": "14" + }, + { + "type": "header", + "bbox": [ + 0.272, + 0.115, + 0.375, + 0.129 + ], + "angle": 0, + "content": "Y. Zhang et al." + }, + { + "type": "table_caption", + "bbox": [ + 0.214, + 0.145, + 0.788, + 0.215 + ], + "angle": 0, + "content": "Table 3: Collection of ablations on EMA and choice of measurement in Eq. (6). Bold and underline mark the best and second best scores. SingleFixed uses the same prepretrained model to initialize \\( M_S \\) and \\( M_G \\) as the EMA version but fixes the weights of \\( M_G \\) during training, resulting in inferior performance. Substituting CE in Eq. (6) to \\( \\ell_1 \\) biases the fidelity scores." + }, + { + "type": "table", + "bbox": [ + 0.319, + 0.228, + 0.687, + 0.316 + ], + "angle": 0, + "content": "
MethodCanonNTIRE20Olympus
PSNRNRQMPSNRNRQMPSNRNRQM
Generalist24.746.064925.086.121326.735.2530
SingleFixed25.715.980325.756.093827.475.2071
\\( \\ell_1 \\)26.175.806126.446.199727.724.9602
Ours25.946.064726.426.226327.405.3722
" + }, + { + "type": "image", + "bbox": [ + 0.303, + 0.332, + 0.699, + 0.451 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.214, + 0.461, + 0.788, + 0.504 + ], + "angle": 0, + "content": "Fig.9: Ablations of loss terms for distillation in Eq. (11). Using one of \\( R_{intra} \\) (blue) and \\( R_{inter} \\) (orange) bias fidelity or perceptual quality (a.-b.), while the combination of them (c.-d.) achieves a balanced improvement (see green and red dots)." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.538, + 0.788, + 0.6 + ], + "angle": 0, + "content": "without changing the ratio between weights of \\( R_{intra} \\) and \\( R_{inter} \\). As shown in Tab. 3, CE favors the overall sharpness while \\( \\ell_1 \\) focuses more on fidelity. Although \\( \\ell_1 \\) is a feasible choice, we use CE in our method due to the importance of sharpness for human eyes [13]." + }, + { + "type": "title", + "bbox": [ + 0.216, + 0.632, + 0.36, + 0.648 + ], + "angle": 0, + "content": "5 Conclusion" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.674, + 0.789, + 0.842 + ], + "angle": 0, + "content": "This paper studies unsupervised real-world image super-resolution from the distillation perspective. The approach considers both the real-world knowledge of a generalized SR model and the low-level characteristics of synthetic predictions of the specialist model, adapting a synthetic domain specialist to target real-world degradation. The core of our method is the Pairwise Distance Distillation. By enforcing the consistencies for intra- and inter-model distances, our method aims at pushing the low-level characteristics of the specialist's real-world predictions towards its synthetic predictions. As a learning scheme, our method can improve off-the-shelf models regarding fidelity and perception on multiple real-world datasets. Additionally, we emphasize our approach to offering a new perspective toward addressing real-world super-resolution challenges." + } + ], + [ + { + "type": "header", + "bbox": [ + 0.526, + 0.115, + 0.732, + 0.128 + ], + "angle": 0, + "content": "Pairwise Distance Distillation" + }, + { + "type": "page_number", + "bbox": [ + 0.769, + 0.117, + 0.786, + 0.127 + ], + "angle": 0, + "content": "15" + }, + { + "type": "text", + "bbox": [ + 0.217, + 0.147, + 0.785, + 0.178 + ], + "angle": 0, + "content": "Acknowledgement This research/project is supported by the Ministry of Education, Singapore, under the Academic Research Fund Tier 1 (FY2022)." + }, + { + "type": "title", + "bbox": [ + 0.218, + 0.198, + 0.323, + 0.214 + ], + "angle": 0, + "content": "References" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.226, + 0.785, + 0.268 + ], + "angle": 0, + "content": "1. Agustsson, E., Timofte, R.: Ntire 2017 challenge on single image super-resolution: Dataset and study. In: The IEEE Conference on Computer Vision and Pattern Recognition (CVPR) Workshops (July 2017)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.269, + 0.785, + 0.309 + ], + "angle": 0, + "content": "2. Cai, J., Zeng, H., Yong, H., Cao, Z., Zhang, L.: Toward real-world single image super-resolution: A new benchmark and a new model. In: Proceedings of the IEEE/CVF International Conference on Computer Vision. pp. 3086-3095 (2019)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.31, + 0.785, + 0.363 + ], + "angle": 0, + "content": "3. Chen, C., Shi, X., Qin, Y., Li, X., Han, X., Yang, T., Guo, S.: Real-world blind super-resolution via feature matching with implicit high-resolution priors. In: Proceedings of the 30th ACM International Conference on Multimedia. pp. 1329-1338 (2022)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.364, + 0.785, + 0.417 + ], + "angle": 0, + "content": "4. Deng, X., Yang, R., Xu, M., Dragotti, P.L.: Wavelet domain style transfer for an effective perception-distortion tradeoff in single image super-resolution. In: Proceedings of the IEEE/CVF international conference on computer vision. pp. 3076-3085 (2019)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.418, + 0.785, + 0.459 + ], + "angle": 0, + "content": "5. Fritsche, M., Gu, S., Timofte, R.: Frequency separation for real-world superresolution. In: 2019 IEEE/CVF International Conference on Computer Vision Workshop (ICCVW). pp. 3599-3608. IEEE (2019)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.46, + 0.785, + 0.514 + ], + "angle": 0, + "content": "6. Gao, H., Guo, J., Wang, G., Zhang, Q.: Cross-domain correlation distillation for unsupervised domain adaptation in nighttime semantic segmentation. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 9913-9923 (2022)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.515, + 0.785, + 0.542 + ], + "angle": 0, + "content": "7. Gao, Q., Zhao, Y., Li, G., Tong, T.: Image super-resolution using knowledge distillation. In: Asian Conference on Computer Vision. pp. 527-541. Springer (2018)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.542, + 0.785, + 0.582 + ], + "angle": 0, + "content": "8. Gatys, L.A., Ecker, A.S., Bethge, M.: Image style transfer using convolutional neural networks. In: Proceedings of the IEEE conference on computer vision and pattern recognition. pp. 2414-2423 (2016)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.583, + 0.785, + 0.623 + ], + "angle": 0, + "content": "9. Gu, J., Lu, H., Zuo, W., Dong, C.: Blind super-resolution with iterative kernel correction. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 1604-1613 (2019)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.624, + 0.785, + 0.65 + ], + "angle": 0, + "content": "0. Hinton, G., Vinyals, O., Dean, J.: Distilling the knowledge in a neural network. arXiv preprint arXiv:1503.02531 (2015)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.65, + 0.785, + 0.69 + ], + "angle": 0, + "content": "1. Huang, Y., Li, S., Wang, L., Tan, T., et al.: Unfolding the alternating optimization for blind super resolution. Advances in Neural Information Processing Systems 33, 5632-5643 (2020)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.691, + 0.785, + 0.718 + ], + "angle": 0, + "content": "2. Kingma, D.P., Ba, J.: Adam: A method for stochastic optimization. arXiv preprint arXiv:1412.6980 (2014)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.719, + 0.785, + 0.772 + ], + "angle": 0, + "content": "3. Ledig, C., Theis, L., Huszár, F., Caballero, J., Cunningham, A., Acosta, A., Aitken, A., Tejani, A., Totz, J., Wang, Z., et al.: Photo-realistic single image superresolution using a generative adversarial network. In: Proceedings of the IEEE conference on computer vision and pattern recognition. pp. 4681-4690 (2017)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.773, + 0.785, + 0.8 + ], + "angle": 0, + "content": "4. Lepik, U., Hein, H.: Haar wavelets. In: Haar Wavelets: With Applications, pp. 7-20. Springer (2014)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.801, + 0.785, + 0.84 + ], + "angle": 0, + "content": "5. Li, W., Li, L., Yang, H.: Progressive cross-domain knowledge distillation for efficient unsupervised domain adaptive object detection. Engineering Applications of Artificial Intelligence 119, 105774 (2023)" + }, + { + "type": "list", + "bbox": [ + 0.226, + 0.226, + 0.785, + 0.84 + ], + "angle": 0, + "content": null + } + ], + [ + { + "type": "page_number", + "bbox": [ + 0.218, + 0.116, + 0.236, + 0.127 + ], + "angle": 0, + "content": "16" + }, + { + "type": "header", + "bbox": [ + 0.272, + 0.115, + 0.375, + 0.129 + ], + "angle": 0, + "content": "Y. Zhang et al." + }, + { + "type": "ref_text", + "bbox": [ + 0.218, + 0.147, + 0.785, + 0.175 + ], + "angle": 0, + "content": "16. Li, Y., Wang, N., Liu, J., Hou, X.: Demystifying neural style transfer. arXiv preprint arXiv:1701.01036 (2017)" + }, + { + "type": "ref_text", + "bbox": [ + 0.218, + 0.177, + 0.785, + 0.217 + ], + "angle": 0, + "content": "17. Liang, J., Zeng, H., Zhang, L.: Efficient and degradation-adaptive network for real-world image super-resolution. In: European Conference on Computer Vision. pp. 574-591. Springer (2022)" + }, + { + "type": "ref_text", + "bbox": [ + 0.218, + 0.218, + 0.785, + 0.259 + ], + "angle": 0, + "content": "18. Liang, J., Cao, J., Sun, G., Zhang, K., Van Gool, L., Timofte, R.: Swinir: Image restoration using swim transformer. In: Proceedings of the IEEE/CVF international conference on computer vision. pp. 1833-1844 (2021)" + }, + { + "type": "ref_text", + "bbox": [ + 0.218, + 0.26, + 0.785, + 0.3 + ], + "angle": 0, + "content": "19. Lim, B., Son, S., Kim, H., Nah, S., Mu Lee, K.: Enhanced deep residual networks for single image super-resolution. In: Proceedings of the IEEE conference on computer vision and pattern recognition workshops. pp. 136-144 (2017)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.301, + 0.785, + 0.341 + ], + "angle": 0, + "content": "20. Liu, Y., Liu, A., Gu, J., Zhang, Z., Wu, W., Qiao, Y., Dong, C.: Discovering distinctive\" semantics\" in super-resolution networks. arXiv preprint arXiv:2108.00406 (2021)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.342, + 0.785, + 0.383 + ], + "angle": 0, + "content": "21. Liu, Y., Zhao, H., Gu, J., Qiao, Y., Dong, C.: Evaluating the generalization ability of super-resolution networks. IEEE Transactions on pattern analysis and machine intelligence (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.384, + 0.785, + 0.411 + ], + "angle": 0, + "content": "22. Liu, Y., Zhang, W., Wang, J.: Adaptive multi-teacher multi-level knowledge distillation. Neurocomputing 415, 106-113 (2020)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.412, + 0.785, + 0.452 + ], + "angle": 0, + "content": "23. Lugmayr, A., Danelljan, M., Timofte, R.: Ntire 2020 challenge on real-world image super-resolution: Methods and results. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition Workshops. pp. 494-495 (2020)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.453, + 0.785, + 0.493 + ], + "angle": 0, + "content": "24. Ma, C., Yang, C.Y., Yang, X., Yang, M.H.: Learning a no-reference quality metric for single-image super-resolution. Computer Vision and Image Understanding 158, 1-16 (2017)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.494, + 0.785, + 0.536 + ], + "angle": 0, + "content": "25. Ma, C., Rao, Y., Cheng, Y., Chen, C., Lu, J., Zhou, J.: Structure-preserving super resolution with gradient guidance. In: Proceedings of the IEEE/CVF conference on computer vision and pattern recognition. pp. 7769-7778 (2020)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.537, + 0.785, + 0.59 + ], + "angle": 0, + "content": "26. Maggioni, M., Tanay, T., Babiloni, F., McDonagh, S., Leonardis, A.: Tunable convolutions with parametric multi-loss optimization. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 20226-20236 (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.591, + 0.785, + 0.646 + ], + "angle": 0, + "content": "27. Nguyen-Meidine, L.T., Belal, A., Kiran, M., Dolz, J., Blais-Morin, L.A., Granger, E.: Unsupervised multi-target domain adaptation through knowledge distillation. In: Proceedings of the IEEE/CVF Winter Conference on Applications of Computer Vision. pp. 1339-1347 (2021)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.647, + 0.785, + 0.673 + ], + "angle": 0, + "content": "28. Park, S., Kwak, N.: Local-selective feature distillation for single image superresolution. arXiv preprint arXiv:2111.10988 (2021)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.675, + 0.785, + 0.715 + ], + "angle": 0, + "content": "29. Park, S.H., Moon, Y.S., Cho, N.I.: Perception-oriented single image superresolution using optimal objective estimation. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 1725-1735 (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.716, + 0.785, + 0.757 + ], + "angle": 0, + "content": "30. Romero, A., Van Gool, L., Timofte, R.: Unpaired real-world super-resolution with pseudo controllable restoration. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 798-807 (2022)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.758, + 0.785, + 0.785 + ], + "angle": 0, + "content": "31. Simonyan, K., Zisserman, A.: Very deep convolutional networks for large-scale image recognition. arXiv preprint arXiv:1409.1556 (2014)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.786, + 0.785, + 0.84 + ], + "angle": 0, + "content": "32. Timofte, R., Agustsson, E., Van Gool, L., Yang, M.H., Zhang, L.: Ntire 2017 challenge on single image super-resolution: Methods and results. In: Proceedings of the IEEE conference on computer vision and pattern recognition workshops. pp. 114-125 (2017)" + }, + { + "type": "list", + "bbox": [ + 0.217, + 0.147, + 0.785, + 0.84 + ], + "angle": 0, + "content": null + } + ], + [ + { + "type": "header", + "bbox": [ + 0.526, + 0.115, + 0.732, + 0.128 + ], + "angle": 0, + "content": "Pairwise Distance Distillation" + }, + { + "type": "page_number", + "bbox": [ + 0.769, + 0.116, + 0.786, + 0.127 + ], + "angle": 0, + "content": "17" + }, + { + "type": "ref_text", + "bbox": [ + 0.214, + 0.147, + 0.788, + 0.189 + ], + "angle": 0, + "content": "33. Wang, H., Chen, X., Ni, B., Liu, Y., Liu, J.: Omni aggregation networks for lightweight image super-resolution. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 22378-22387 (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.214, + 0.189, + 0.788, + 0.23 + ], + "angle": 0, + "content": "34. Wang, W., Zhang, H., Yuan, Z., Wang, C.: Unsupervised real-world superresolution: A domain adaptation perspective. In: Proceedings of the IEEE/CVF International Conference on Computer Vision. pp. 4318-4327 (2021)" + }, + { + "type": "ref_text", + "bbox": [ + 0.214, + 0.23, + 0.788, + 0.27 + ], + "angle": 0, + "content": "35. Wang, X., Xie, L., Dong, C., Shan, Y.: Real-esrgan: Training real-world blind super-resolution with pure synthetic data. In: Proceedings of the IEEE/CVF international conference on computer vision. pp. 1905-1914 (2021)" + }, + { + "type": "ref_text", + "bbox": [ + 0.214, + 0.27, + 0.788, + 0.325 + ], + "angle": 0, + "content": "36. Wang, X., Yu, K., Wu, S., Gu, J., Liu, Y., Dong, C., Qiao, Y., Change Loy, C.: Esrgan: Enhanced super-resolution generative adversarial networks. In: Proceedings of the European conference on computer vision (ECCV) workshops. pp. 0-0 (2018)" + }, + { + "type": "ref_text", + "bbox": [ + 0.214, + 0.325, + 0.788, + 0.365 + ], + "angle": 0, + "content": "37. Wang, Y., Cao, Y., Zha, Z.J., Zhang, J., Xiong, Z.: Deep degradation prior for low-quality image classification. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 11049-11058 (2020)" + }, + { + "type": "ref_text", + "bbox": [ + 0.214, + 0.365, + 0.788, + 0.407 + ], + "angle": 0, + "content": "38. Wang, Y., Wang, L., Shi, S., Li, V.O., Tu, Z.: Go from the general to the particular: Multi-domain translation with domain transformation networks. In: Proceedings of the AAAI Conference on Artificial Intelligence. vol. 34, pp. 9233-9241 (2020)" + }, + { + "type": "ref_text", + "bbox": [ + 0.214, + 0.407, + 0.788, + 0.461 + ], + "angle": 0, + "content": "39. Wei, P., Xie, Z., Lu, H., Zhan, Z., Ye, Q., Zuo, W., Lin, L.: Component divide-and-conquer for real-world image super-resolution. In: Computer Vision-ECCV 2020: 16th European Conference, Glasgow, UK, August 23-28, 2020, Proceedings, Part VIII 16. pp. 101-117. Springer (2020)" + }, + { + "type": "ref_text", + "bbox": [ + 0.214, + 0.461, + 0.788, + 0.515 + ], + "angle": 0, + "content": "40. Wei, Y., Gu, S., Li, Y., Timofte, R., Jin, L., Song, H.: Unsupervised real-world image super resolution via domain-distance aware training. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 13385-13394 (2021)" + }, + { + "type": "ref_text", + "bbox": [ + 0.214, + 0.515, + 0.788, + 0.556 + ], + "angle": 0, + "content": "41. Yang, C., Pan, J., Gao, X., Jiang, T., Liu, D., Chen, G.: Cross-task knowledge distillation in multi-task recommendation. In: Proceedings of the AAAI Conference on Artificial Intelligence. vol. 36, pp. 4318-4326 (2022)" + }, + { + "type": "ref_text", + "bbox": [ + 0.214, + 0.556, + 0.788, + 0.597 + ], + "angle": 0, + "content": "42. Zeyde, R., Elad, M., Protter, M.: On single image scale-up using sparse-representations. In: Curves and Surfaces: 7th International Conference, Avignon, France, June 24-30, 2010, Revised Selected Papers 7. pp. 711-730. Springer (2012)" + }, + { + "type": "ref_text", + "bbox": [ + 0.214, + 0.597, + 0.788, + 0.638 + ], + "angle": 0, + "content": "43. Zhang, K., Liang, J., Van Gool, L., Timofte, R.: Designing a practical degradation model for deep blind image super-resolution. In: Proceedings of the IEEE/CVF International Conference on Computer Vision. pp. 4791-4800 (2021)" + }, + { + "type": "ref_text", + "bbox": [ + 0.214, + 0.638, + 0.788, + 0.678 + ], + "angle": 0, + "content": "44. Zhang, R., Isola, P., Efros, A.A., Shechtman, E., Wang, O.: The unreasonable effectiveness of deep features as a perceptual metric. In: Proceedings of the IEEE conference on computer vision and pattern recognition. pp. 586-595 (2018)" + }, + { + "type": "ref_text", + "bbox": [ + 0.214, + 0.678, + 0.788, + 0.718 + ], + "angle": 0, + "content": "45. Zhang, R., Gu, J., Chen, H., Dong, C., Zhang, Y., Yang, W.: Crafting training degradation distribution for the accuracy-generalization trade-off. arXiv preprint arXiv:2305.18107 (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.214, + 0.718, + 0.788, + 0.759 + ], + "angle": 0, + "content": "46. Zhang, W., Li, X., Shi, G., Chen, X., Qiao, Y., Zhang, X., Wu, X.M., Dong, C.: Real-world image super-resolution as multi-task learning. Advances in Neural Information Processing Systems 36 (2024)" + }, + { + "type": "ref_text", + "bbox": [ + 0.214, + 0.759, + 0.788, + 0.8 + ], + "angle": 0, + "content": "47. Zhang, Y., Chen, H., Chen, X., Deng, Y., Xu, C., Wang, Y.: Data-free knowledge distillation for image super-resolution. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 7852-7861 (2021)" + }, + { + "type": "ref_text", + "bbox": [ + 0.214, + 0.8, + 0.788, + 0.841 + ], + "angle": 0, + "content": "48. Zhang, Y., Ji, B., Hao, J., Yao, A.: Perception-distortion balanced admm optimization for single-image super-resolution. In: European Conference on Computer Vision. pp. 108-125. Springer (2022)" + }, + { + "type": "list", + "bbox": [ + 0.214, + 0.147, + 0.788, + 0.841 + ], + "angle": 0, + "content": null + } + ], + [ + { + "type": "page_number", + "bbox": [ + 0.218, + 0.116, + 0.236, + 0.127 + ], + "angle": 0, + "content": "18" + }, + { + "type": "header", + "bbox": [ + 0.272, + 0.115, + 0.375, + 0.129 + ], + "angle": 0, + "content": "Y. Zhang et al." + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.147, + 0.788, + 0.189 + ], + "angle": 0, + "content": "49. Zhong, Z., Shen, T., Yang, Y., Lin, Z., Zhang, C.: Joint sub-bands learning with clique structures for wavelet domain super-resolution. Advances in neural information processing systems 31 (2018)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.19, + 0.788, + 0.232 + ], + "angle": 0, + "content": "50. Zhu, J.Y., Park, T., Isola, P., Efros, A.A.: Unpaired image-to-image translation using cycle-consistent adversarial networks. In: Proceedings of the IEEE international conference on computer vision. pp. 2223-2232 (2017)" + }, + { + "type": "list", + "bbox": [ + 0.217, + 0.147, + 0.788, + 0.232 + ], + "angle": 0, + "content": null + } + ] +] \ No newline at end of file diff --git a/2024/Pairwise Distance Distillation for Unsupervised Real-World Image Super-Resolution/821e13b2-5f5a-4742-b51d-28c69e766f13_origin.pdf b/2024/Pairwise Distance Distillation for Unsupervised Real-World Image Super-Resolution/821e13b2-5f5a-4742-b51d-28c69e766f13_origin.pdf new file mode 100644 index 0000000000000000000000000000000000000000..fb6081d9f7a8ab70df4ce6909d0bfabe9fe9c86b --- /dev/null +++ b/2024/Pairwise Distance Distillation for Unsupervised Real-World Image Super-Resolution/821e13b2-5f5a-4742-b51d-28c69e766f13_origin.pdf @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:a34c1e98ccc11594dd0dd7039fb407e7394e1a755ee847f747cba518166126dd +size 7944409 diff --git a/2024/Pairwise Distance Distillation for Unsupervised Real-World Image Super-Resolution/full.md b/2024/Pairwise Distance Distillation for Unsupervised Real-World Image Super-Resolution/full.md new file mode 100644 index 0000000000000000000000000000000000000000..63e457d4c0a9c3fadfe71f107d418462b1317805 --- /dev/null +++ b/2024/Pairwise Distance Distillation for Unsupervised Real-World Image Super-Resolution/full.md @@ -0,0 +1,358 @@ +# Pairwise Distance Distillation for Unsupervised Real-World Image Super-Resolution + +Yuehan Zhang $^{1}$ , Seungjun Lee $^{2}$ , and Angela Yao $^{1}$ + +$^{1}$ National University of Singapore + +{zyuehan,ayao}@comp.nus.edu.sg + +2 Korea University + +9penguin9@korea.ac.kr + +Abstract. Standard single-image super-resolution creates paired training data from high-resolution images through fixed downsampling kernels. However, real-world super-resolution (RWSR) faces unknown degradations in the low-resolution inputs, all the while lacking paired training data. Existing methods approach this problem by learning blind general models through complex synthetic augmentations on training inputs; they sacrifice the performance on specific degradation for broader generalization to many possible ones. We address the unsupervised RWSR for a targeted real-world degradation. We study from a distillation perspective and introduce a novel pairwise distance distillation framework. Through our framework, a model specialized in synthetic degradation adapts to target real-world degradations by distilling intra- and intermodel distances across the specialized model and an auxiliary generalized model. Experiments on diverse datasets demonstrate that our method significantly enhances fidelity and perceptual quality, surpassing state-of-the-art approaches in RWSR. The source code is available at https://github.com/Yuehan717/PDD. + +# 1 Introduction + +Single-image super-resolution (SISR) predicts high-resolution (HR) images from low-resolution (LR) counterparts. The standard SISR model addresses predefined downsampling kernels, e.g. bicubic interpolation. However, real-world scenarios of SISR (RWSR) encompass unknown degradations in LR images, including blur, noise, and JPEG compression artifacts [35,43,45] with diverse combinations. The various and unknown degradations pose additional challenges when learning an RWSR model. + +Existing RWSR methods focus on blind generalization. Such methods synthesize paired training data with extensive and harsh degradations, involving multiple rounds of random blurring, added noise, resizing, and compression [35, 43]. The complex degradation pipeline allows the model to generalize to diverse unknown conditions. We refer to such a model addressing various degradations as a generalist. However, as much of the model's capacity is devoted to handling multiple conditions, these "jack-of-all-trades" models come with performance trade- + +![](images/1a2f83a7888a9039e41899decc9bde8830ec36ac79d7e346fe43ee3b272ab5a5.jpg) +Fig. 1: $\hat{Y}_G^U$ and $\hat{Y}_G^L$ are reconstructions for real-world and bicubic-interpolated (BI) inputs using a blind generalized model (generalist); $\hat{Y}_S^U$ and $\hat{Y}_S^L$ are counterparts using a standard SR model (bicubic specialist). The specialist enhances the BI inputs with clearer details, while the generalist does better for the real-world one. We distill the intra- and inter-model distances for an improved real-world reconstruction $\hat{Y}_{S^{*}}^{U}$ . + +offs [45], i.e., they have inferior performance compared to the model only optimized for the tested degradation $(\hat{Y}_G^L$ and $\hat{Y}_S^L$ in Fig. 1). + +SR models can be optimized for specific and known input degradations through supervised learning [9, 11, 18, 36]. The LR-HR pairs for training are created by degrading HR images with fixed kernels, e.g., bicubic interpolation, and Gaussian blur. We refer to models adept at the specific input degradation as a specialist. Degradations in RWSR datasets [2, 39] always belong to a specific domain [45]. For example, images captured by a specific camera model tend to feature consistent sensor noise [2]. However, the formulation of a real-world degradation is always unknown, posing difficulties in creating paired training data. As such, learning specialist models for such real-world domains is challenging, and strategies that do not require direct supervision are crucial. + +This paper tackles unsupervised RWSR from a knowledge distillation perspective. Distillation has been successful for domain adaptation in high-level tasks [6, 22, 38], but it is under-explored in low-level vision. Our distillation framework adapts a specialist model of known synthetic degradation to target real-world degradations. We use an auxiliary generalist model to provide generalization knowledge of real-world degradations that the specialist lacks. However, naive distillation only transfers knowledge from the generalist model; we aim to improve beyond the generalist's moderate performance. To that end, we also consider the low-level characteristics of the specialist's predictions on synthetic samples as a reference for the distillation. + +Our method uses the strength of both the specialist and the generalist by exploring the relationships between their predictions. Given a generalist and a specialist for synthetic degradation, we consider predictions of synthetic samples and real-world samples from both models (see Fig. 1). The generalist model has smaller gaps in qualities of real-world and synthetic predictions, featuring close low-level characteristics between domains [20,21]. In contrast, the specialist outputs distanced low-level characteristics but high-quality predictions for synthetic samples $(\hat{Y}_S^L$ in Fig. 1), providing valuable low-level characteristics for reference. + +We propose to let the specialist imitate prediction relationships from the generalist - that is, push the specialist's real-world predictions $(\hat{Y}_S^U$ in Fig. 1) to have low-level characteristics similar to its high-quality synthetic predictions. If such similarity is satisfied by the specialist, it would exhibit two consistencies. First, for two input samples, the difference between their predictions from the same model is intra-model distance; such distances would be consistent for the specialist and the generalist. Secondly, for a given input sample, the difference between generalist and specialist predictions is inter-model distance; such distances would be consistent whether the sample is synthetic or real-world. These intra- and inter-model distance consistencies form the basis of our pairwise distance distillation framework. Together with designs on configurations of specialist and generalist models, our adapted specialist model experimentally shows significant improvements over the generalist model. + +Contributions. To summarize, the contributions of our work are three-fold: + +- To the best of our knowledge, we are the first to propose a distillation perspective for unsupervised RWSR to combine generalist and specialist models. +- We propose a novel pairwise distance distillation framework, emphasizing transferring intra-model and inter-model distances to enhance the specialist's performance in real-world scenarios. +- Experiments on three benchmarks demonstrate that our approach improves the off-the-shelf models regarding fidelity and perceptual performances. + +# 2 Related Works + +Single-Image Super-Resolution (SISR) methods use deep neural networks of various architectures, such as residual networks [19,36] and transformers [18, 33]. Despite improving the standard SISR performance with dedicated architectures, these methods struggle to generalize to real-world scenarios [34,40]. + +Real-World Image Super-Resolution (RWSR) aims to address unknown input degradations. The infeasibility of creating paired training data without known kernels presents a significant challenge in learning specialized models for RWSR. There are two primary strategies in existing RWSR. The first assumes that networks trained on diverse and challenging synthetic degradations will effectively generalize to real-world data [35, 43]. Even though real-world degradations are complex, within a specific setting or deployment, they tend to be limited in the domain and less diverse than the training pipelines. Consequently, [45] proposes customizing the synthetic pipeline to better match specific real-world data. Nonetheless, an unresolved gap exists between synthesized and real-world degradations. Instead, our method operates directly on real-world data and avoids the intrinsic gap from synthetic degradation. + +The second strategy connects real-world and synthetic degradation with transfer learning. Most works use image-to-image translation [5,30,40] to adapt the synthetic degradation to real-world ones. Such methods resort to intricate designs like CycleGAN [50], which struggle to replicate real-world degradations [34] + +reliably. Our proposed method explores feature distances between model outputs and bypasses the need to mimic real-world degradations. + +Knowledge Distillation is initially introduced for model compression [10]. In low-level vision, several works [7, 28, 47] distill large models into more efficient ones by enforcing similarity between their internal features or predictions. Recently, knowledge distillation has served diverse purposes in high-level tasks like segmentation, including domain adaptation [6, 15, 27] and transfer learning [38, 41]. However, such applications are less discussed in low-level vision. + +# 3 Method + +Our method adapts the specialist in synthetic degradation to an unlabeled real-world domain. We use the low-level characteristics of synthetic predictions from the specialists and the knowledge of the generalist model. Fig. 2 is the overview of our approach and the novel Pairwise Distance Distillation (PDD) method, which enforces the consistency of the intra- and inter-model distances to improve the specialist $(M_S)$ 's real-world predictions. Sec. 3.1 gives notions used in this section and Sec. 3.2 illustrates the unsupervised formulation. Sec. 3.3 explains the details of PDD and Sec. 3.4 provides the full optimization aims and discusses the static and Exponential Moving Average (EMA) versions of our method. + +# 3.1 Definitions & Setup + +When provided with an LR input $X \in \mathbb{R}^{H \times W \times C}$ - where $H \times W$ is the spatial resolution and $C$ is the number of color channels - a SISR model aims to reconstruct an HR image $\hat{Y} \in \mathbb{R}^{rH \times rW \times C}$ , scaling by a factor $r$ . Due to the difficulty in obtaining paired training data, the conventional approach in SISR involves generating an LR image $X$ by downsampling the HR ground-truth $Y$ , i.e., $X = D_{\downarrow}(Y)$ . This downsampling operation $D_{\downarrow}$ typically refers to bicubic interpolation, although other alternatives exist [1]. + +In a more extensive setting, it is typical for the LR image to be affected by various factors, such as blurring, noise, compression artifacts, etc. To exhibit these degradations, the LR image $X$ can be formed by applying $D_{\downarrow}$ along with one or multiple degradations $D_{i}$ , in any order, i.e., + +$$ +X = D _ {1} \circ \dots \circ D _ {n} (Y). \tag {1} +$$ + +Here, we include $D_{\downarrow}$ as one of the degradations $\{D_i\}_{i \in [1,n]}$ for simplicity. Now, let's consider two categories of degradation sets consisting of known degradations: $\mathcal{D}_S$ , which comprises simple degradations with a few factors, and $\mathcal{D}_G$ characterized by complex degradations with a wide range of factors: + +$$ +\mathcal {D} _ {S} = \left\{D _ {i} \right\} _ {i \in \mathcal {S}}, \mathcal {D} _ {G} = \left\{D _ {i} \right\} _ {i \in \mathcal {G}}, \quad \text {w h e r e} | \mathcal {G} | \gg | \mathcal {S} |. \tag {2} +$$ + +In the simplest case for $\mathcal{D}_S$ , the set $\mathcal{D}_S = \{D_\downarrow\}$ , which represents the standard SISR setup. For the broader case, $\mathcal{D}_G$ should encompass diverse and challenging + +degradations beyond $\{D_{\downarrow}\}$ to ensure that models trained on this data generalize well to blind settings [35, 43]. + +Consider two models, $M_S$ and $M_G$ , trained on datasets constructed with $\mathcal{D}_S$ and $\mathcal{D}_G$ respectively. Both models are optimized with common supervised SR learning losses [13, 36] including MSE-based loss, VGG loss, and adversarial training (see Eq. (9)). Since $\mathcal{D}_S$ comprises a narrower range of degradations compared to $\mathcal{D}_G$ , $M_S$ is a "specialist" model for the limited domain defined by $\mathcal{D}_S$ . On the other hand, $M_G$ is a "generalist" model as it achieves moderate performance across the broader domain derived by $\mathcal{D}_G$ . Comparing $M_S$ and $M_G$ , we assume the specialist's superiority to the generalist in $\mathcal{D}_S$ and the generalist's better handling of unknown degradations. These assumptions are empirically validated in Tab. 1. + +# 3.2 Unsupervised Learning Through Distillation + +Our approach aligns with the unsupervised setup [5, 30, 40], wherein we work with a collection of real-world LR images alongside unpaired HR clean images. We denote the set of LR images as $\{X^U\}$ with the same unknown degradations $\{D_i\}_{i\in \mathcal{U}}$ . Additionally, we generate LR counterparts $\{X^L\}$ from HR images using specific degradations from $\mathcal{D}_S$ , e.g. bicubic interpolation. The specialist $M_S$ is adept in restoring $X^L$ , while $M_G$ performs moderately in restoring $X^L$ and $X^U$ . We aim to adapt $M_S$ to $\{X^U\}$ with the aid of $M_G$ . As in [10], a naive distillation approach is having $M_S$ imitate the predictions from $M_G$ for unlabeled inputs: + +$$ +\begin{array}{l} \hat {Y} _ {S} ^ {L} = M _ {S} (X ^ {L}), \hat {Y} _ {S} ^ {U} = M _ {S} (X ^ {U}), \hat {Y} _ {G} ^ {U} = M _ {G} (X ^ {U}), \\ \mathcal {L} _ {N D} = \mathcal {L} _ {L} \left(\hat {Y} _ {S} ^ {U}, \hat {Y} _ {G} ^ {U}\right) + \lambda \mathcal {L} _ {L} \left(\hat {Y} _ {S} ^ {L}, Y ^ {L}\right), \tag {3} \\ \end{array} +$$ + +where $Y^{L}$ represents the ground-truth image in the labeled domain, $\mathcal{L}_L$ denotes the supervised loss formulated in Eq. (9), and $\lambda$ is a scale factor to balance between the distillation and primary objective of $M_S$ . However, this simple imitation approach relies solely on information from the generalist $M_G$ and fails to harness the strengths of both models. + +# 3.3 Pairwise Distance Distillation (PDD) + +Consider any pair of LR images - a real-world $X^U$ and a synthesized $X^L$ . As shown in Fig. 2, we obtain four predictions by applying $M_S$ and $M_G$ : + +$$ +\begin{array}{l} \hat {Y} _ {S} ^ {U} = M _ {S} (X ^ {U}), \hat {Y} _ {G} ^ {U} = M _ {G} (X ^ {U}), \\ \text {a n d} \hat {Y} _ {S} ^ {L} = M _ {S} \left(X ^ {L}\right), \hat {Y} _ {G} ^ {L} = M _ {G} \left(X ^ {L}\right). \tag {4} \\ \end{array} +$$ + +Our distillation explores the relationship between predictions in Eq. (4) to seek for knowledge combination. We use VGG features [31] of the predictions as our basis for exploration. VGG, as a classification model, inherently captures semantic information. Yet it also captures low-level characteristics related to + +![](images/12350873bd7515b36d1bb9eac7095c5779bdad28b5317afcef3ab01b1e90f8ef.jpg) +Fig. 2: Schematic of Pairwise Distance Distillation (PDD). $X^{U}$ and $X^{L}$ are unlabeled (real-world) and labeled (synthetic) inputs. Model $M_{G}$ is a generalist trained with extensive synthetic pipeline, while $M_{S}$ specializes in $X^{L}$ . The prediction $\hat{Y}_S^L$ is supervised by its ground truth throughout training. PDD enforces the consistency between $\{\varDelta_U, \varDelta_L\}$ and between $\{d_S, d_G\}$ to improve $M_{S}$ 's real-world performance. + +image quality [37], e.g. blur and sharpness, and is widely used in the perceptual loss for super-resolution tasks [13, 35, 36]. + +As a generalist, $M_G$ outputs synthetic and real-world predictions with small quality gaps, which is characterized by the overlapping distributions of low-level characteristics of $\{\hat{Y}_G^L\}$ and $\{\hat{Y}_G^U\}$ [20, 21]. Following [20], Fig. 5a projects the low-level features of $\{\hat{Y}_G^L\}$ (DIV2K) and $\{\hat{Y}_G^U\}$ (NTIRE20). The generalist's coverage strongly overlaps while the specialist's predictions are well-separated (see Fig. 5b). The within-model relationship of the generalist exhibits the generalizability to real-world degradation. + +On the other side, the high image quality of $\{\hat{Y}_S^L\}$ , due to $M_S$ 's specialization, makes its low-level characteristics a valuable reference. To integrate the knowledge for real-world performance, we propose to make the specialist imitate the generalist's within-model relationship, that is, pushing the low-level characteristics of real-world predictions from $M_S$ , $\{\hat{Y}_S^U\}$ , to be similar to that of its synthetic predictions $\{\hat{Y}_S^L\}$ . To achieve this aim, we first assume such similarity is satisfied for $M_S$ and conclude the following consistencies about VGG feature distances between predictions: + +1. Intra-model distances for predictions of the same input pair should be consistent across $M_S$ and $M_G$ (green shade in Fig. 2). +2. Inter-model distances for predictions of a single input should be consistent across synthetic and real-world domains (orange shade in Fig. 2). + +Our rationales for these consistencies are as follows. First, we recall the low-level characteristics of predictions from the same model are similar. We discussed it for $M_G$ with Fig. 5a and made such an assumption for $M_S$ . Thus, the distance between the same model's predictions $\{\hat{Y}_G^L,\hat{Y}_G^U\}$ (or $\{\hat{Y}_S^L,\hat{Y}_S^U\}$ ), i.e. intra-model distance, reflects mainly semantic differences due to the close low-level characteristics. As it naturally follows that the applied SR models should not change semantics, intra-model distances should be consistent for $M_S$ and $M_G$ . + +Second, we note the low-level characteristics of predictions from the two models differ, e.g. $\hat{Y}_S^L$ has higher quality than the counterpart $\hat{Y}_G^L$ . The distance + +between the two predictions $\{\hat{Y}_G^L,\hat{Y}_S^L\}$ (or $\{\hat{Y}_G^U,\hat{Y}_S^U\}$ ) of a given sample, i.e. inter-model distance, captures such differences in low-level characteristics, as semantics stay constant for the predictions of the same input. Given the similarity of low-level characteristics within a single model, the inter-model distances for synthetic and real-world samples should be consistent. + +We structure the two consistencies as distillations on the intra- and inter-model distances. Our method encourages $M_S$ 's real-world predictions $\{\hat{Y}_S^U\}$ to have similar low-level characteristics as its synthetic predictions $\{\hat{Y}_S^L\}$ , taking it as a reference from high-quality images. + +Intra-model Distance Distillation enforces consistency on the distances between predictions from the same model (green shading in Fig. 2). Consider the $\ell_1$ distance $d_G^{ij}$ between predictions from the generalist $M_G$ , i.e. $\hat{Y}_G^L$ and $\hat{Y}_G^U$ , in VGG feature space, and the distance $d_S^{ij}$ for the specialist $M_S$ is similarly defined: + +$$ +d _ {G} ^ {i j} = \left\| \Phi_ {i j} \left(\hat {Y} _ {G} ^ {L}\right) - \Phi_ {i j} \left(\hat {Y} _ {G} ^ {U}\right) \right\| _ {1}, \tag {5} +$$ + +$$ +d _ {S} ^ {i j} = \left\| \Phi_ {i j} \left(\hat {Y} _ {S} ^ {L}\right) - \Phi_ {i j} \left(\hat {Y} _ {S} ^ {U}\right) \right\| _ {1}, +$$ + +where $\varPhi_{ij}$ refers to the $j$ -th layer in the $i$ -th residual block of VGG19. For $d_G^{ij}, d_S^{ij} \in \mathbb{R}^{c \times h \times w}, h \times w$ is the spatial resolution and $c$ is the number of channels. We enforce the consistency between $d_G^{ij}$ and $d_S^{ij}$ by minimizing their difference measured with the Cross-Entropy (CE) $R_{intra}$ : + +$$ +R _ {i n t r a} = - \frac {1}{h w} \sum_ {i, j} \sum_ {m, n} \mathbf {S} \left(d _ {G} ^ {i j} [ m, n ]\right) \log \mathbf {S} \left(d _ {S} ^ {i j} [ m, n ]\right), \tag {6} +$$ + +where $[m,n]$ is a spatial index of the feature map, and $\mathbf{S}(\cdot)$ denotes the SoftMax function with the input size being $\mathbb{R}^c$ . The element-wise CE reflects the negative log-likelihood of distances locally. While other measures are feasible, CE is empirically a good choice in favoring the overall sharpness (see Sec. 4.4). + +Inter-model Distance Distillation enforces the consistency of changes in low-level characteristics between predictions from different models (orange shading in Fig. 2). Given the two predictions for a single input, i.e. $\{\hat{Y}_G^L,\hat{Y}_S^L\}$ (or $\{\hat{Y}_G^U,\hat{Y}_S^U\}$ ), we first calculate their feature distance $\varDelta_L^{ij}$ (or $\varDelta_U^{ij}$ ) to represent the differences in low-level characteristics. However, enforcing consistency on $\varDelta_L^{ij}$ and $\varDelta_U^{ij}$ is less straightforward, as the spatial layout of corresponding inputs can differ. Thus, we compute the inter-model distances with the following: + +$$ +\Delta_ {L} ^ {i j} = \operatorname {G r a m} \left(\Phi_ {i j} \left(\hat {Y} _ {S} ^ {L}\right) - \Phi_ {i j} \left(\hat {Y} _ {G} ^ {L}\right)\right), \tag {7} +$$ + +$$ +\Delta_ {U} ^ {i j} = \operatorname {G r a m} \left(\Phi_ {i j} \left(\hat {Y} _ {S} ^ {U}\right) - \Phi_ {i j} \left(\hat {Y} _ {G} ^ {U}\right)\right). +$$ + +Gram(\cdot) refers to the Gram matrix, which calculates the correlations among vectorized feature maps along the channel dimension. The Gram matrix captures the statistics while collapsing the spatial layout [8]. + +For distillation, we ensure the consistency between inter-model distances by minimizing $R_{inter}$ , the Frobenius norm between $\Delta_U^{ij}$ and $\Delta_L^{ij}$ : + +$$ +R _ {i n t e r} = \sum_ {i, j} \| \Delta_ {U} ^ {i j} - \Delta_ {L} ^ {i j} \| _ {\mathcal {F}}, \tag {8} +$$ + +where $\| \cdot \|_{\mathcal{F}}$ computes the Frobenius norm. The Frobenius norm of Gram Matrix difference is interpreted in [8, 16] as measuring distribution discrepancy. + +# 3.4 Full Method + +The network $M_S$ is fully optimized with supervised losses for labeled synthetic data $X^L$ and unsupervised losses for real-world data $X^U$ . + +Supervised Losses maintain the specialization in synthetic degradation and reduce overfitting to the distillation. With the ground-truth image for prediction $\hat{Y}_S^L$ , denoted as $Y^{L}$ , we implement a combination of supervised loss functions. It includes a wavelet-based loss $L_{\mathrm{wv}}$ , a perceptual loss $L_{\mathrm{vgg}}$ , and generative loss $L_{\mathrm{gan}}$ for adversarial training: + +$$ +\mathcal {L} _ {L} \left(\hat {Y} _ {S} ^ {L}, Y ^ {L}\right) = \alpha_ {1} L _ {\mathrm {w v}} \left(\hat {Y} _ {S} ^ {L}, Y ^ {L}\right) + \alpha_ {2} L _ {\mathrm {v g g}} \left(\hat {Y} _ {S} ^ {L}, Y ^ {L}\right) + \alpha_ {3} L _ {\mathrm {g a n}} \left(\hat {Y} _ {S} ^ {L}\right), \tag {9} +$$ + +where $L_{\mathrm{vgg}}$ and $L_{\mathrm{gan}}$ are adopted from RealESRGAN [35], with balancing weights $\alpha_{1}$ , $\alpha_{2}$ , and $\alpha_{3}$ . We use but omit writing the discriminator loss for brevity. $L_{\mathrm{wv}}$ refers to an $\ell_{1}$ -based wavelet loss [4, 49]: + +$$ +L _ {\mathrm {w v}} \left(\hat {Y} _ {S} ^ {L}, Y ^ {L}\right) = \sum_ {i} \omega_ {i} \| W \left(\hat {Y} _ {S} ^ {L}\right) _ {i} - W \left(Y ^ {L}\right) _ {i} \| _ {1}, \tag {10} +$$ + +where $W(\cdot)_i$ extracts the $i$ -th wavelet channel, and $\omega_i$ weights its importance. For the wavelet transform, we apply the Haar wavelet [14] for all output channels. + +Unsupervised Losses. In the absence of ground truth for $\dot{Y}_S^U$ , we opt to optimize for the consistency outlined in Eq. (6) and Eq. (8) as regularizers. To fully utilize the discriminator's knowledge of realness, we also incorporate the generative loss in Eq. (9): + +$$ +\mathcal {L} _ {U} = \lambda_ {1} R _ {\text {i n t r a}} + \lambda_ {2} R _ {\text {i n t e r}} + \lambda_ {3} L _ {\text {g a n}} \left(\hat {Y} _ {S} ^ {U}\right), \tag {11} +$$ + +where $\lambda_1, \lambda_2$ , and $\lambda_3$ balance the optimization aims. $L_{gan}$ here uses the same discriminator in Eq. (9). The supervised loss in Eq. (9) and the unsupervised distillation loss in Eq. (11) are combined as our final optimization objective, expressed as $\mathcal{L} = \mathcal{L}_L + \mathcal{L}_U$ . + +Color Correction. In practice, our approach often leads to color shifts due to the regularization of distances within the feature space. To address this, we rectify the output by normalizing the mean and variance of each color channel with those of the corresponding input channels [35]. Further details are provided in Supplementary Sec. E. + +Learning Configurations. Our proposed PDD is a learning framework that can utilize pre-trained networks. It is recommended that $M_G$ is pre-trained with extensive degradations $\mathcal{D}_G$ , as in [35,43]. The static configuration (refer to Fig. 3 (a)) initializes $M_S$ with the model pre-trained on specific synthetic degradation $(\mathcal{D}_S)$ , and $M_G$ is frozen throughout the training process. + +Alternative configuration initializes $M_G$ and $M_S$ with the same model pretrained with $\mathcal{D}_G$ and maintain $M_G$ as an Exponential Moving Average (EMA) + +![](images/9a06f91c09709db24ce244043fae065c542b7217eea9c97661dd4c811c64823c.jpg) +(a) Static Version + +![](images/42699b1275ee8fb378a964dd43c03b25f4764c12587b85096f877b1d62b73e1c.jpg) +(b) EMA Version +Fig. 3: Two initialization options for the generalist and specialist models. (a) The static configuration initializes $M_G$ with a model pre-trained by the complex synthetic pipeline and $M_S$ with a model pre-trained by simple degradation in $X^L$ . Weights of $M_G$ are frozen during distillation. (b) Both $M_S$ and $M_G$ are initialized with a pre-trained generalized model. Weights of $M_G$ are the EMA version of $M_S$ . + +version of $M_S$ (see Fig. 3 (b)). Despite $M_S$ not being initialized as a specialist, it can readily specialize to the synthetic degradation $\mathcal{D}_S$ through supervised learning in Eq. (9). Empirically, the EMA version tends to be more effective, likely because the updated knowledge of real-world data in $M_S$ also benefits $M_G$ through EMA, enabling $M_G$ to yield better references for optimization. + +# 4 Experiments + +# 4.1 Settings + +Datasets. We experiment on real-world datasets, RealSR [2] and DRealSR [39], and a synthetic dataset without released degradation settings, NTIRE20 [23]. For experiments on RealSR and DRealSR, we synthesize labeled training data $\{X^L\}$ from high-resolution images in the DF2K dataset [1,32]. For the NTIRE20 dataset, we use the provided target clean images to synthesize low-resolution counterparts. For synthesizing $\{X^L\}$ , we use only bicubic interpolation as the default and investigate other options for $\mathcal{D}_S$ in the Supplementary Sec. D.b. RealSR or DRealSR has multiple data sources, and we optimize our methods for each source separately. For DRealSR, we report results for Panasonic and Olympus sources, as the others do not have sufficient training samples. + +Training Details. We crop patches of size $48 \times 48$ as inputs and train the model with the ADAM optimizer [12] using the settings $\beta_{1} = 0.9$ , $\beta_{2} = 0.999$ , and $\epsilon = 10^{-8}$ . The batch size is 16, including equal numbers of labeled and unlabeled inputs. The initial learning rate is $1e - 4$ and halved after $25\mathrm{K}$ iterations. The static version needs $50\mathrm{K}$ iterations, and the EMA version is trained for $100\mathrm{K}$ iterations due to the initial specialist underfitting to the synthetic domain. All models are for $\times 4$ super-resolution. Other details are in Supplementary Sec. A. + +Evaluation. We evaluate models on testing sets of RealSR, DRealSR, and the validation set of NTIRE20. We crop non-overlapped 512x512 patches from LR images in DRealSR to avoid memory exhaustion. For other datasets, we use the full images. To fully use the provided high-resolution counterparts, we report + +Table 1: $M_G$ and $M_S$ (static) columns show initial performances on the bicubic-interpolated data $\{X^L\}$ from Set14 [42] and $\{X^U\}$ with unknown degradation from NTIRE20 [23]. $M_S$ (EMA) column shows the performance of $M_S$ in the EMA version on $\{X^L\}$ after training process. + +
DomainMetricMGMS(static)MS(EMA)DomainMetricMGMS(static)
{X^L}PSNR↑23.9824.8126.26{X^U}PSNR↑25.0819.82
LPIPS↓0.23490.13370.1471LPIPS↓0.25040.7552
+ +![](images/adac3a713f37c17d3700c1664a4f5998e292bc09811b0b9568eb473d7e07adaa.jpg) +Fig. 4: Improvements over the Generalist on five unlabeled data domains, where a lower LPIPS score is better. ND improves fidelity scores (PSNR) but dramatically drops perceptual scores (LPIPS and NRQM). Both versions of our method achieve better improvements than ND for all reported metrics. For each domain, the Static version improves at least one of the perceptual metrics; the EMA version improves all. + +![](images/e5caf7f5458acc358aa53c142ed322b3cd65192f3994ba1b8ff29c2aa6b504dc.jpg) + +![](images/ce96daf2ed4534ed3889474a07f1593f55cf0449e01b92eddd8b124e53975af4.jpg) + +scores of full-reference metrics, PSNR, SSIM, and LPIPS [44]. The PSNR and SSIM scores are calculated on the Y-channel of YCbCr format images. Additionally, we follow BSRGAN [43] and supplement with the no-reference metric NRQM [24] to evaluate the perceptual clearness, as the ground truth in real-world captured datasets [2,39] also feature some blur. We use the implementation in IQA-PyTorch3 for all metrics. It is important to consider both fidelity and perceptual scores; fidelity metrics are worse at detecting blurriness while perceptual metrics are not sensitive to high-frequency artifacts [25, 26, 29, 48]. + +# 4.2 Effectiveness of PDD + +Comparisons. We compare the two versions of our method in Fig. 3 with the naive distillation defined in Eq. (3) regarding their improvements over the generalist network $M_G$ . We choose RealESRGAN [35] as the pre-trained generalist model and ESRGAN [36], a standard SISR model, as the pre-trained specialist for the naive distillation and the static version in Fig. 3 (a). Tab. 1 shows the initial performances of the generalist and specialist for $\{X^L\}$ and $\{X^U\}$ . As we assumed in Sec. 3.1, the specialist performs better than the generalist for $\{X^L\}$ while $M_G$ generalizes better for $\{X^U\}$ . For the EMA version in Fig. 3 (b), $M_S$ as well as $M_G$ is initialized with RealESRGAN. Through the supervised learning in Eq. (9), $M_S(\mathrm{EMA})$ shows specialized performance for labeled data $\{X^L\}$ . + +![](images/b9b0e83fd39785c5a23976a9f9c8b8420f148cee9965642b117a07c7173992fe.jpg) +(a) Generalist + +![](images/13111ea952a2ff2f9334a41a885eb6adab64666cff078714724ab6b817f11af1.jpg) +(b) Bicubic Specialist +Fig. 5: Visualization of low-level features for predictions of DIV2K [1] (bicubic) and NTIRE20 (unknown) following [20] (a) Generalist's predictions for the two domains has overlapped distribution. (b) The predictions from the specialist that only adept in bicubic interpolation have separated distributions. (c) After applying our method (static version), predictions for two domains are pushed close. + +![](images/0dbef2d3aa8ff9f165ed726739e40ab8bc9098996d292124aeefc0f65688c46b.jpg) +(c) Adapted Specialist + +Fig. 4 plots the improvements over the pretrained generalist for Naive Distillation (ND), Static, and EMA versions of our method. Compared to the generalist, ND dramatically drops in perceptual scores (LPIPS, NRQM) despite maintaining or improving the PSNR. Both versions of our method perform better than ND; compared to the generalist, our static version improves PSNR significantly and at least one perceptual score for each data domain. The EMA version improves all metrics with the highest magnitudes. + +Change in Low-Level Characteristics. We extract low-level features for visualization following [20]. Fig. 5b visualize predictions generated by the bicubic specialist (ESRGAN), which has separate distributions. Fig. 5c verifies the predictions from the specialist after applying our method (static version), where the separated distributions are pushed closer. Visualizing low-level differences between images with arbitrary content is challenging [20]. As such, we also quantitatively compare the KL divergence between the low-level features. As shown in Fig. 6, applying our method decreases the KL divergence between labeled (DIV2K-bicubic) and unlabeled predictions. Details of the visualization and KL-diversity calculations are in Supplementary Sec. B. + +# 4.3 Comparisons with Other Methods + +Quantitative Comparison. We consider two variants of our method by applying the EMA version on RealESRGAN [35] and BSRGAN [43]. In Tab. 2, we compare their performances with other state-of-the-art RWSR methods, including FeMaSR [3], EDAN [17], DASR [40], DAN [11], RealESRGAN [35], BSRGAN [43], and TG [46], which released code and model weights that enable reproduction. We regenerate results to evaluate these methods, while Supplementary Sec. C discusses related but not directly comparable methods. + +As shown in Tab. 2, EDAN, DAN, and DASR perform strongly on the fidelity metrics (PSNR, SSIM) but are poor perceptually, especially on the NRQM score. The qualitative comparison highlights the blurry effects. Our method enhances RealESRGAN and BSRGAN across all metrics and demonstrates superior performance compared to TG when both are applied to RealESRGAN. For FeMaSR, + +![](images/fc424c549e010f2a02e0bd6b0132326a4670d21da322796fc5bcbb28db78de02.jpg) +Fig. 6: KL divergence between specialist's unlabeled and labeled predictions. The KL divergence between labeled and unlabeled predictions is lower after applying our method (red) than that in the specialist before adaption (blue). + +![](images/999982b588e56a6e40aa14c057731539c91f904e67ea2cf4826f3b6c3b7788c0.jpg) +Fig.7: User study on outputs of 30 real-world images. We recruit 30 evaluators for each image. Each vote is for a better quality within a pair of options. We compared our method to RealESRGAN and BSRGAN by using them as pre-trained models; FeMaSR is compared to RealESRGAN+ours. Our method gains at least $65\%$ votes. + +Table 2: Quantitative comparison with state-of-the-art methods. DAN, EDAN, and DASR are separately listed as they favor fidelity metrics but yield poor perceptual scores. For other methods, **bold** and **underline** mark the best and second-best scores. Subscripts in the last two columns show their difference from the corresponding pretrained generalist, where improvement is colored red and drop is blue. + +
DatasetMetricDAN [11]EDAN [17]DASR [40]FeMaSR [3]RealESRGAN [43]BSRGAN [35]RealESRGAN +TG [46]RealESRGAN +OursBSRGAN +Ours
NTIRE20 -ValidPSNR↑26.8326.59-23.5425.0825.4425.5626.42(+1.340)26.37(+0.930)
SSIM↑0.71710.7400-0.66650.70610.69840.71930.7297(+0.024)0.7238(+0.025)
LPIPS↓0.57470.2678-0.23600.25040.26450.24200.2475(-0.003)0.2595(-0.005)
NRQM↑4.87845.8481-6.62456.12136.27795.79266.2263(+0.105)6.2781(+0.0002)
RealSR -CanonPSNR↑26.6726.3626.7124.2924.7425.5725.0325.94(+1.200)26.13(+0.560)
SSIM↑0.77400.77740.77820.74600.76340.76830.77200.7733(+0.010)0.7626(-0.006)
LPIPS↓0.40950.30260.25070.28090.26070.25730.26150.2516(-0.009)0.2517(-0.006)
NRQM↑3.04073.63603.93836.01516.06496.02935.42506.0647(-0.0002)6.1323(+0.103)
RealSR -NikonPSNR↑26.2925.3725.6323.9424.3124.7924.5625.28(+0.970)25.07(+0.280)
SSIM↑0.75330.74600.74730.70970.74060.73340.74060.7470(+0.014)0.7294(-0.001)
LPIPS↓0.41330.32000.27850.30460.28510.27970.29490.2802(-0.005)0.2713(-0.008)
NRQM↑3.20564.48724.49525.94885.66855.93875.19425.9044(+0.236)6.1276(+0.189)
DRealSR -PanasonicPSNR↑28.9928.12-25.0227.0326.9327.5127.87(+0.840)27.69(+0.760)
SSIM↑0.81230.8006-0.71180.77820.76210.78430.7939(+0.016)0.7752(+0.013)
LPIPS↓0.42350.2747-0.30880.27030.28200.28360.2682(-0.002)0.2687(-0.013)
NRQM↑2.81314.5915-6.03425.37515.58654.64835.3540(-0.021)5.5608(-0.026)
DRealSR -OlympusPSNR↑28.7427.85-24.3826.7326.4927.4327.40(+0.670)27.25(+0.764)
SSIM↑0.80630.7979-0.66910.76670.75680.78430.7846(+0.018)0.7635(+0.007)
LPIPS↓0.45950.3555-0.40070.31590.32830.34160.3139(-0.002)0.3268(-0.001)
NRQM↑2.79613.7989-6.25905.25305.53764.14595.3722(+0.120)5.5957(+0.058)
+ +we achieve competitive perceptual scores and significantly surpass their fidelity scores. + +User Study is conducted by presenting two reconstructions of the same LR image and asking for the better one. We compare the EMA version of our method with methods RealESRGAN, BSRGAN, and FeMaSR. Each comparison is conducted on 30 different real-world results among 30 evaluators. As shown in Fig. 7, at least $65\%$ evaluators vote for our method in each comparison. Although quan- + +![](images/08c40686f7cdf46f11179acf5ee3b4d46bf665dcd2a82aa0347d71b3d2327def.jpg) +(a) LR + +![](images/fe4126d5806061733da59ca32caea78ffea8581cc5a5d75aadc5a60df29e959a.jpg) +(b) FeMaSR + +![](images/a701cbffed2b69cac700b3bfe630a36a0ea7ecf6909494a40d05b33a34f62112.jpg) +(c) DASR + +![](images/94a365716f9add06d7e9deef1b9d16872979103513cd73fdecf30118425c7765.jpg) +(d) RE + +![](images/cc55b2e784cee3de2f35c78937d0beaea3ab304e47ca312581fb3aed6c3fa279.jpg) +(e) RE+ours +Fig. 8: Qualitative comparisons. LRs are from the RealSR-Canon dataset. Columns (e) and (f) are the results of applying our methods on RealESRGAN (RE) and BSRGAN (BG). FeMaSR and DASR can not predict reasonable details. + +![](images/76210f2e54e46d87b870239539695486ecac9b1d4fac627b1b821560a83a56a0.jpg) +(f) BG+ours + +titatively competing, FeMaSR yields unpleasant artifacts to human eyes, further shown in qualitative comparison. + +Qualitative Comparison is in Fig. 8. FeMaSR [3] produces high-frequency artifacts, while outputs from DASR [40] suffer blur or artifacts. Compared to RealESRGAN, our method predicts sharper patterns without introducing distortion. More comparisons are in Supplementary Sec. F. + +# 4.4 Ablations + +Ablation on loss terms reveals the effectiveness of each term in Eq. (11). All experiments take RealESRGAN as the pre-trained model and follow the EMA implementation of our method in Fig. 3 (b). Fig. 9 shows that $R_{intra}$ and $R_{inter}$ relate differently to fidelity and perception. Their combination achieves balanced improvements in both directions, and the $L_{\mathrm{gan}}$ increases pattern sharpness with a slight trade-off in PSNR. Controlling the ratio between weights of $R_{intra}$ and $R_{inter}$ results in fidelity and perceptual quality trade-offs. We choose their balance in Sec. 4.3 and refer to Supplementary Sec. D.a for more discussion. + +Effectiveness of EMA is further validated by comparing to a SingleFixed version, which also uses RealESRGAN as the pre-trained model for both $M_S$ and $M_G$ but fixes the $M_G$ during training. As shown in Tab. 3, updating the weights of the generalist network through EMA enables higher performance in both fidelity and perception. + +Choice of $R_{intra}$ measurement. Eq. (6) use Cross-Entropy (CE) for measuring the difference between $d_S^{ij}$ and $d_G^{ij}$ . Here, we alternate the CE to $\ell_1$ distance + +Table 3: Collection of ablations on EMA and choice of measurement in Eq. (6). Bold and underline mark the best and second best scores. SingleFixed uses the same prepretrained model to initialize $M_S$ and $M_G$ as the EMA version but fixes the weights of $M_G$ during training, resulting in inferior performance. Substituting CE in Eq. (6) to $\ell_1$ biases the fidelity scores. + +
MethodCanonNTIRE20Olympus
PSNRNRQMPSNRNRQMPSNRNRQM
Generalist24.746.064925.086.121326.735.2530
SingleFixed25.715.980325.756.093827.475.2071
\( \ell_1 \)26.175.806126.446.199727.724.9602
Ours25.946.064726.426.226327.405.3722
+ +![](images/88a535fed39f63c499dc3cda070746fde06c7777cc173d5c05a8a2c3d542523a.jpg) +Fig.9: Ablations of loss terms for distillation in Eq. (11). Using one of $R_{intra}$ (blue) and $R_{inter}$ (orange) bias fidelity or perceptual quality (a.-b.), while the combination of them (c.-d.) achieves a balanced improvement (see green and red dots). + +without changing the ratio between weights of $R_{intra}$ and $R_{inter}$ . As shown in Tab. 3, CE favors the overall sharpness while $\ell_1$ focuses more on fidelity. Although $\ell_1$ is a feasible choice, we use CE in our method due to the importance of sharpness for human eyes [13]. + +# 5 Conclusion + +This paper studies unsupervised real-world image super-resolution from the distillation perspective. The approach considers both the real-world knowledge of a generalized SR model and the low-level characteristics of synthetic predictions of the specialist model, adapting a synthetic domain specialist to target real-world degradation. The core of our method is the Pairwise Distance Distillation. By enforcing the consistencies for intra- and inter-model distances, our method aims at pushing the low-level characteristics of the specialist's real-world predictions towards its synthetic predictions. As a learning scheme, our method can improve off-the-shelf models regarding fidelity and perception on multiple real-world datasets. Additionally, we emphasize our approach to offering a new perspective toward addressing real-world super-resolution challenges. + +Acknowledgement This research/project is supported by the Ministry of Education, Singapore, under the Academic Research Fund Tier 1 (FY2022). + +# References + +1. Agustsson, E., Timofte, R.: Ntire 2017 challenge on single image super-resolution: Dataset and study. In: The IEEE Conference on Computer Vision and Pattern Recognition (CVPR) Workshops (July 2017) +2. Cai, J., Zeng, H., Yong, H., Cao, Z., Zhang, L.: Toward real-world single image super-resolution: A new benchmark and a new model. In: Proceedings of the IEEE/CVF International Conference on Computer Vision. pp. 3086-3095 (2019) +3. Chen, C., Shi, X., Qin, Y., Li, X., Han, X., Yang, T., Guo, S.: Real-world blind super-resolution via feature matching with implicit high-resolution priors. In: Proceedings of the 30th ACM International Conference on Multimedia. pp. 1329-1338 (2022) +4. Deng, X., Yang, R., Xu, M., Dragotti, P.L.: Wavelet domain style transfer for an effective perception-distortion tradeoff in single image super-resolution. In: Proceedings of the IEEE/CVF international conference on computer vision. pp. 3076-3085 (2019) +5. Fritsche, M., Gu, S., Timofte, R.: Frequency separation for real-world superresolution. In: 2019 IEEE/CVF International Conference on Computer Vision Workshop (ICCVW). pp. 3599-3608. IEEE (2019) +6. Gao, H., Guo, J., Wang, G., Zhang, Q.: Cross-domain correlation distillation for unsupervised domain adaptation in nighttime semantic segmentation. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 9913-9923 (2022) +7. Gao, Q., Zhao, Y., Li, G., Tong, T.: Image super-resolution using knowledge distillation. In: Asian Conference on Computer Vision. pp. 527-541. Springer (2018) +8. Gatys, L.A., Ecker, A.S., Bethge, M.: Image style transfer using convolutional neural networks. In: Proceedings of the IEEE conference on computer vision and pattern recognition. pp. 2414-2423 (2016) +9. Gu, J., Lu, H., Zuo, W., Dong, C.: Blind super-resolution with iterative kernel correction. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 1604-1613 (2019) +0. Hinton, G., Vinyals, O., Dean, J.: Distilling the knowledge in a neural network. arXiv preprint arXiv:1503.02531 (2015) +1. Huang, Y., Li, S., Wang, L., Tan, T., et al.: Unfolding the alternating optimization for blind super resolution. Advances in Neural Information Processing Systems 33, 5632-5643 (2020) +2. Kingma, D.P., Ba, J.: Adam: A method for stochastic optimization. arXiv preprint arXiv:1412.6980 (2014) +3. Ledig, C., Theis, L., Huszár, F., Caballero, J., Cunningham, A., Acosta, A., Aitken, A., Tejani, A., Totz, J., Wang, Z., et al.: Photo-realistic single image superresolution using a generative adversarial network. In: Proceedings of the IEEE conference on computer vision and pattern recognition. pp. 4681-4690 (2017) +4. Lepik, U., Hein, H.: Haar wavelets. In: Haar Wavelets: With Applications, pp. 7-20. Springer (2014) +5. Li, W., Li, L., Yang, H.: Progressive cross-domain knowledge distillation for efficient unsupervised domain adaptive object detection. Engineering Applications of Artificial Intelligence 119, 105774 (2023) + +16. Li, Y., Wang, N., Liu, J., Hou, X.: Demystifying neural style transfer. arXiv preprint arXiv:1701.01036 (2017) +17. Liang, J., Zeng, H., Zhang, L.: Efficient and degradation-adaptive network for real-world image super-resolution. In: European Conference on Computer Vision. pp. 574-591. Springer (2022) +18. Liang, J., Cao, J., Sun, G., Zhang, K., Van Gool, L., Timofte, R.: Swinir: Image restoration using swim transformer. In: Proceedings of the IEEE/CVF international conference on computer vision. pp. 1833-1844 (2021) +19. Lim, B., Son, S., Kim, H., Nah, S., Mu Lee, K.: Enhanced deep residual networks for single image super-resolution. In: Proceedings of the IEEE conference on computer vision and pattern recognition workshops. pp. 136-144 (2017) +20. Liu, Y., Liu, A., Gu, J., Zhang, Z., Wu, W., Qiao, Y., Dong, C.: Discovering distinctive" semantics" in super-resolution networks. arXiv preprint arXiv:2108.00406 (2021) +21. Liu, Y., Zhao, H., Gu, J., Qiao, Y., Dong, C.: Evaluating the generalization ability of super-resolution networks. IEEE Transactions on pattern analysis and machine intelligence (2023) +22. Liu, Y., Zhang, W., Wang, J.: Adaptive multi-teacher multi-level knowledge distillation. Neurocomputing 415, 106-113 (2020) +23. Lugmayr, A., Danelljan, M., Timofte, R.: Ntire 2020 challenge on real-world image super-resolution: Methods and results. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition Workshops. pp. 494-495 (2020) +24. Ma, C., Yang, C.Y., Yang, X., Yang, M.H.: Learning a no-reference quality metric for single-image super-resolution. Computer Vision and Image Understanding 158, 1-16 (2017) +25. Ma, C., Rao, Y., Cheng, Y., Chen, C., Lu, J., Zhou, J.: Structure-preserving super resolution with gradient guidance. In: Proceedings of the IEEE/CVF conference on computer vision and pattern recognition. pp. 7769-7778 (2020) +26. Maggioni, M., Tanay, T., Babiloni, F., McDonagh, S., Leonardis, A.: Tunable convolutions with parametric multi-loss optimization. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 20226-20236 (2023) +27. Nguyen-Meidine, L.T., Belal, A., Kiran, M., Dolz, J., Blais-Morin, L.A., Granger, E.: Unsupervised multi-target domain adaptation through knowledge distillation. In: Proceedings of the IEEE/CVF Winter Conference on Applications of Computer Vision. pp. 1339-1347 (2021) +28. Park, S., Kwak, N.: Local-selective feature distillation for single image superresolution. arXiv preprint arXiv:2111.10988 (2021) +29. Park, S.H., Moon, Y.S., Cho, N.I.: Perception-oriented single image superresolution using optimal objective estimation. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 1725-1735 (2023) +30. Romero, A., Van Gool, L., Timofte, R.: Unpaired real-world super-resolution with pseudo controllable restoration. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 798-807 (2022) +31. Simonyan, K., Zisserman, A.: Very deep convolutional networks for large-scale image recognition. arXiv preprint arXiv:1409.1556 (2014) +32. Timofte, R., Agustsson, E., Van Gool, L., Yang, M.H., Zhang, L.: Ntire 2017 challenge on single image super-resolution: Methods and results. In: Proceedings of the IEEE conference on computer vision and pattern recognition workshops. pp. 114-125 (2017) + +33. Wang, H., Chen, X., Ni, B., Liu, Y., Liu, J.: Omni aggregation networks for lightweight image super-resolution. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 22378-22387 (2023) +34. Wang, W., Zhang, H., Yuan, Z., Wang, C.: Unsupervised real-world superresolution: A domain adaptation perspective. In: Proceedings of the IEEE/CVF International Conference on Computer Vision. pp. 4318-4327 (2021) +35. Wang, X., Xie, L., Dong, C., Shan, Y.: Real-esrgan: Training real-world blind super-resolution with pure synthetic data. In: Proceedings of the IEEE/CVF international conference on computer vision. pp. 1905-1914 (2021) +36. Wang, X., Yu, K., Wu, S., Gu, J., Liu, Y., Dong, C., Qiao, Y., Change Loy, C.: Esrgan: Enhanced super-resolution generative adversarial networks. In: Proceedings of the European conference on computer vision (ECCV) workshops. pp. 0-0 (2018) +37. Wang, Y., Cao, Y., Zha, Z.J., Zhang, J., Xiong, Z.: Deep degradation prior for low-quality image classification. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 11049-11058 (2020) +38. Wang, Y., Wang, L., Shi, S., Li, V.O., Tu, Z.: Go from the general to the particular: Multi-domain translation with domain transformation networks. In: Proceedings of the AAAI Conference on Artificial Intelligence. vol. 34, pp. 9233-9241 (2020) +39. Wei, P., Xie, Z., Lu, H., Zhan, Z., Ye, Q., Zuo, W., Lin, L.: Component divide-and-conquer for real-world image super-resolution. In: Computer Vision-ECCV 2020: 16th European Conference, Glasgow, UK, August 23-28, 2020, Proceedings, Part VIII 16. pp. 101-117. Springer (2020) +40. Wei, Y., Gu, S., Li, Y., Timofte, R., Jin, L., Song, H.: Unsupervised real-world image super resolution via domain-distance aware training. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 13385-13394 (2021) +41. Yang, C., Pan, J., Gao, X., Jiang, T., Liu, D., Chen, G.: Cross-task knowledge distillation in multi-task recommendation. In: Proceedings of the AAAI Conference on Artificial Intelligence. vol. 36, pp. 4318-4326 (2022) +42. Zeyde, R., Elad, M., Protter, M.: On single image scale-up using sparse-representations. In: Curves and Surfaces: 7th International Conference, Avignon, France, June 24-30, 2010, Revised Selected Papers 7. pp. 711-730. Springer (2012) +43. Zhang, K., Liang, J., Van Gool, L., Timofte, R.: Designing a practical degradation model for deep blind image super-resolution. In: Proceedings of the IEEE/CVF International Conference on Computer Vision. pp. 4791-4800 (2021) +44. Zhang, R., Isola, P., Efros, A.A., Shechtman, E., Wang, O.: The unreasonable effectiveness of deep features as a perceptual metric. In: Proceedings of the IEEE conference on computer vision and pattern recognition. pp. 586-595 (2018) +45. Zhang, R., Gu, J., Chen, H., Dong, C., Zhang, Y., Yang, W.: Crafting training degradation distribution for the accuracy-generalization trade-off. arXiv preprint arXiv:2305.18107 (2023) +46. Zhang, W., Li, X., Shi, G., Chen, X., Qiao, Y., Zhang, X., Wu, X.M., Dong, C.: Real-world image super-resolution as multi-task learning. Advances in Neural Information Processing Systems 36 (2024) +47. Zhang, Y., Chen, H., Chen, X., Deng, Y., Xu, C., Wang, Y.: Data-free knowledge distillation for image super-resolution. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 7852-7861 (2021) +48. Zhang, Y., Ji, B., Hao, J., Yao, A.: Perception-distortion balanced admm optimization for single-image super-resolution. In: European Conference on Computer Vision. pp. 108-125. Springer (2022) + +49. Zhong, Z., Shen, T., Yang, Y., Lin, Z., Zhang, C.: Joint sub-bands learning with clique structures for wavelet domain super-resolution. Advances in neural information processing systems 31 (2018) +50. Zhu, J.Y., Park, T., Isola, P., Efros, A.A.: Unpaired image-to-image translation using cycle-consistent adversarial networks. In: Proceedings of the IEEE international conference on computer vision. pp. 2223-2232 (2017) \ No newline at end of file diff --git a/2024/Pairwise Distance Distillation for Unsupervised Real-World Image Super-Resolution/images.zip b/2024/Pairwise Distance Distillation for Unsupervised Real-World Image Super-Resolution/images.zip new file mode 100644 index 0000000000000000000000000000000000000000..0a48e203e6e39e38ac3bf87d4dca54b956c6044d --- /dev/null +++ b/2024/Pairwise Distance Distillation for Unsupervised Real-World Image Super-Resolution/images.zip @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:cc3f23ab314e8a218d7e372be8c54bb45dec897c5162601b3f898705d1e891c4 +size 534476 diff --git a/2024/Pairwise Distance Distillation for Unsupervised Real-World Image Super-Resolution/layout.json b/2024/Pairwise Distance Distillation for Unsupervised Real-World Image Super-Resolution/layout.json new file mode 100644 index 0000000000000000000000000000000000000000..1abf124a1e2d1bba11a4b2df2ddb59feb311ee94 --- /dev/null +++ b/2024/Pairwise Distance Distillation for Unsupervised Real-World Image Super-Resolution/layout.json @@ -0,0 +1,12975 @@ +{ + "pdf_info": [ + { + "para_blocks": [ + { + "bbox": [ + 141, + 111, + 473, + 148 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 141, + 111, + 473, + 148 + ], + "spans": [ + { + "bbox": [ + 141, + 111, + 473, + 148 + ], + "type": "text", + "content": "Pairwise Distance Distillation for Unsupervised Real-World Image Super-Resolution" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 187, + 168, + 427, + 181 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 187, + 168, + 427, + 181 + ], + "spans": [ + { + "bbox": [ + 187, + 168, + 427, + 181 + ], + "type": "text", + "content": "Yuehan Zhang" + }, + { + "bbox": [ + 187, + 168, + 427, + 181 + ], + "type": "inline_equation", + "content": "^{1}" + }, + { + "bbox": [ + 187, + 168, + 427, + 181 + ], + "type": "text", + "content": ", Seungjun Lee" + }, + { + "bbox": [ + 187, + 168, + 427, + 181 + ], + "type": "inline_equation", + "content": "^{2}" + }, + { + "bbox": [ + 187, + 168, + 427, + 181 + ], + "type": "text", + "content": ", and Angela Yao" + }, + { + "bbox": [ + 187, + 168, + 427, + 181 + ], + "type": "inline_equation", + "content": "^{1}" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 235, + 190, + 378, + 201 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 235, + 190, + 378, + 201 + ], + "spans": [ + { + "bbox": [ + 235, + 190, + 378, + 201 + ], + "type": "inline_equation", + "content": "^{1}" + }, + { + "bbox": [ + 235, + 190, + 378, + 201 + ], + "type": "text", + "content": " National University of Singapore" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 235, + 201, + 378, + 213 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 235, + 201, + 378, + 213 + ], + "spans": [ + { + "bbox": [ + 235, + 201, + 378, + 213 + ], + "type": "text", + "content": "{zyuehan,ayao}@comp.nus.edu.sg" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 267, + 213, + 346, + 223 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 267, + 213, + 346, + 223 + ], + "spans": [ + { + "bbox": [ + 267, + 213, + 346, + 223 + ], + "type": "text", + "content": "2 Korea University" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 257, + 224, + 357, + 234 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 257, + 224, + 357, + 234 + ], + "spans": [ + { + "bbox": [ + 257, + 224, + 357, + 234 + ], + "type": "text", + "content": "9penguin9@korea.ac.kr" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 160, + 265, + 452, + 451 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 160, + 265, + 452, + 451 + ], + "spans": [ + { + "bbox": [ + 160, + 265, + 452, + 451 + ], + "type": "text", + "content": "Abstract. Standard single-image super-resolution creates paired training data from high-resolution images through fixed downsampling kernels. However, real-world super-resolution (RWSR) faces unknown degradations in the low-resolution inputs, all the while lacking paired training data. Existing methods approach this problem by learning blind general models through complex synthetic augmentations on training inputs; they sacrifice the performance on specific degradation for broader generalization to many possible ones. We address the unsupervised RWSR for a targeted real-world degradation. We study from a distillation perspective and introduce a novel pairwise distance distillation framework. Through our framework, a model specialized in synthetic degradation adapts to target real-world degradations by distilling intra- and intermodel distances across the specialized model and an auxiliary generalized model. Experiments on diverse datasets demonstrate that our method significantly enhances fidelity and perceptual quality, surpassing state-of-the-art approaches in RWSR. The source code is available at https://github.com/Yuehan717/PDD." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 133, + 472, + 229, + 484 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 472, + 229, + 484 + ], + "spans": [ + { + "bbox": [ + 133, + 472, + 229, + 484 + ], + "type": "text", + "content": "1 Introduction" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 130, + 498, + 480, + 581 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 498, + 480, + 581 + ], + "spans": [ + { + "bbox": [ + 130, + 498, + 480, + 581 + ], + "type": "text", + "content": "Single-image super-resolution (SISR) predicts high-resolution (HR) images from low-resolution (LR) counterparts. The standard SISR model addresses predefined downsampling kernels, e.g. bicubic interpolation. However, real-world scenarios of SISR (RWSR) encompass unknown degradations in LR images, including blur, noise, and JPEG compression artifacts [35,43,45] with diverse combinations. The various and unknown degradations pose additional challenges when learning an RWSR model." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 130, + 582, + 481, + 665 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 582, + 481, + 665 + ], + "spans": [ + { + "bbox": [ + 130, + 582, + 481, + 665 + ], + "type": "text", + "content": "Existing RWSR methods focus on blind generalization. Such methods synthesize paired training data with extensive and harsh degradations, involving multiple rounds of random blurring, added noise, resizing, and compression [35, 43]. The complex degradation pipeline allows the model to generalize to diverse unknown conditions. We refer to such a model addressing various degradations as a generalist. However, as much of the model's capacity is devoted to handling multiple conditions, these \"jack-of-all-trades\" models come with performance trade-" + } + ] + } + ], + "index": 9 + } + ], + "discarded_blocks": [], + "page_size": [ + 612, + 792 + ], + "page_idx": 0 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 160, + 113, + 451, + 205 + ], + "blocks": [ + { + "bbox": [ + 160, + 113, + 451, + 205 + ], + "lines": [ + { + "bbox": [ + 160, + 113, + 451, + 205 + ], + "spans": [ + { + "bbox": [ + 160, + 113, + 451, + 205 + ], + "type": "image", + "image_path": "1a2f83a7888a9039e41899decc9bde8830ec36ac79d7e346fe43ee3b272ab5a5.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 130, + 210, + 482, + 269 + ], + "lines": [ + { + "bbox": [ + 130, + 210, + 482, + 269 + ], + "spans": [ + { + "bbox": [ + 130, + 210, + 482, + 269 + ], + "type": "text", + "content": "Fig. 1: " + }, + { + "bbox": [ + 130, + 210, + 482, + 269 + ], + "type": "inline_equation", + "content": "\\hat{Y}_G^U" + }, + { + "bbox": [ + 130, + 210, + 482, + 269 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 130, + 210, + 482, + 269 + ], + "type": "inline_equation", + "content": "\\hat{Y}_G^L" + }, + { + "bbox": [ + 130, + 210, + 482, + 269 + ], + "type": "text", + "content": " are reconstructions for real-world and bicubic-interpolated (BI) inputs using a blind generalized model (generalist); " + }, + { + "bbox": [ + 130, + 210, + 482, + 269 + ], + "type": "inline_equation", + "content": "\\hat{Y}_S^U" + }, + { + "bbox": [ + 130, + 210, + 482, + 269 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 130, + 210, + 482, + 269 + ], + "type": "inline_equation", + "content": "\\hat{Y}_S^L" + }, + { + "bbox": [ + 130, + 210, + 482, + 269 + ], + "type": "text", + "content": " are counterparts using a standard SR model (bicubic specialist). The specialist enhances the BI inputs with clearer details, while the generalist does better for the real-world one. We distill the intra- and inter-model distances for an improved real-world reconstruction " + }, + { + "bbox": [ + 130, + 210, + 482, + 269 + ], + "type": "inline_equation", + "content": "\\hat{Y}_{S^{*}}^{U}" + }, + { + "bbox": [ + 130, + 210, + 482, + 269 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_caption" + } + ], + "index": 2 + }, + { + "bbox": [ + 130, + 297, + 479, + 323 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 297, + 479, + 323 + ], + "spans": [ + { + "bbox": [ + 130, + 297, + 479, + 323 + ], + "type": "text", + "content": "offs [45], i.e., they have inferior performance compared to the model only optimized for the tested degradation " + }, + { + "bbox": [ + 130, + 297, + 479, + 323 + ], + "type": "inline_equation", + "content": "(\\hat{Y}_G^L" + }, + { + "bbox": [ + 130, + 297, + 479, + 323 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 130, + 297, + 479, + 323 + ], + "type": "inline_equation", + "content": "\\hat{Y}_S^L" + }, + { + "bbox": [ + 130, + 297, + 479, + 323 + ], + "type": "text", + "content": " in Fig. 1)." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 130, + 324, + 482, + 443 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 324, + 482, + 443 + ], + "spans": [ + { + "bbox": [ + 130, + 324, + 482, + 443 + ], + "type": "text", + "content": "SR models can be optimized for specific and known input degradations through supervised learning [9, 11, 18, 36]. The LR-HR pairs for training are created by degrading HR images with fixed kernels, e.g., bicubic interpolation, and Gaussian blur. We refer to models adept at the specific input degradation as a specialist. Degradations in RWSR datasets [2, 39] always belong to a specific domain [45]. For example, images captured by a specific camera model tend to feature consistent sensor noise [2]. However, the formulation of a real-world degradation is always unknown, posing difficulties in creating paired training data. As such, learning specialist models for such real-world domains is challenging, and strategies that do not require direct supervision are crucial." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 130, + 447, + 482, + 567 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 447, + 482, + 567 + ], + "spans": [ + { + "bbox": [ + 130, + 447, + 482, + 567 + ], + "type": "text", + "content": "This paper tackles unsupervised RWSR from a knowledge distillation perspective. Distillation has been successful for domain adaptation in high-level tasks [6, 22, 38], but it is under-explored in low-level vision. Our distillation framework adapts a specialist model of known synthetic degradation to target real-world degradations. We use an auxiliary generalist model to provide generalization knowledge of real-world degradations that the specialist lacks. However, naive distillation only transfers knowledge from the generalist model; we aim to improve beyond the generalist's moderate performance. To that end, we also consider the low-level characteristics of the specialist's predictions on synthetic samples as a reference for the distillation." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 130, + 570, + 482, + 667 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 570, + 482, + 667 + ], + "spans": [ + { + "bbox": [ + 130, + 570, + 482, + 667 + ], + "type": "text", + "content": "Our method uses the strength of both the specialist and the generalist by exploring the relationships between their predictions. Given a generalist and a specialist for synthetic degradation, we consider predictions of synthetic samples and real-world samples from both models (see Fig. 1). The generalist model has smaller gaps in qualities of real-world and synthetic predictions, featuring close low-level characteristics between domains [20,21]. In contrast, the specialist outputs distanced low-level characteristics but high-quality predictions for synthetic samples " + }, + { + "bbox": [ + 130, + 570, + 482, + 667 + ], + "type": "inline_equation", + "content": "(\\hat{Y}_S^L" + }, + { + "bbox": [ + 130, + 570, + 482, + 667 + ], + "type": "text", + "content": " in Fig. 1), providing valuable low-level characteristics for reference." + } + ] + } + ], + "index": 7 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "spans": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "type": "text", + "content": "2" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 166, + 91, + 229, + 102 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 166, + 91, + 229, + 102 + ], + "spans": [ + { + "bbox": [ + 166, + 91, + 229, + 102 + ], + "type": "text", + "content": "Y. Zhang et al." + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 1 + }, + { + "para_blocks": [ + { + "bbox": [ + 130, + 116, + 482, + 270 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 116, + 482, + 270 + ], + "spans": [ + { + "bbox": [ + 130, + 116, + 482, + 270 + ], + "type": "text", + "content": "We propose to let the specialist imitate prediction relationships from the generalist - that is, push the specialist's real-world predictions " + }, + { + "bbox": [ + 130, + 116, + 482, + 270 + ], + "type": "inline_equation", + "content": "(\\hat{Y}_S^U" + }, + { + "bbox": [ + 130, + 116, + 482, + 270 + ], + "type": "text", + "content": " in Fig. 1) to have low-level characteristics similar to its high-quality synthetic predictions. If such similarity is satisfied by the specialist, it would exhibit two consistencies. First, for two input samples, the difference between their predictions from the same model is intra-model distance; such distances would be consistent for the specialist and the generalist. Secondly, for a given input sample, the difference between generalist and specialist predictions is inter-model distance; such distances would be consistent whether the sample is synthetic or real-world. These intra- and inter-model distance consistencies form the basis of our pairwise distance distillation framework. Together with designs on configurations of specialist and generalist models, our adapted specialist model experimentally shows significant improvements over the generalist model." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 131, + 272, + 472, + 283 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 131, + 272, + 472, + 283 + ], + "spans": [ + { + "bbox": [ + 131, + 272, + 472, + 283 + ], + "type": "text", + "content": "Contributions. To summarize, the contributions of our work are three-fold:" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 138, + 293, + 480, + 376 + ], + "type": "list", + "angle": 0, + "index": 7, + "blocks": [ + { + "bbox": [ + 138, + 293, + 479, + 316 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 293, + 479, + 316 + ], + "spans": [ + { + "bbox": [ + 138, + 293, + 479, + 316 + ], + "type": "text", + "content": "- To the best of our knowledge, we are the first to propose a distillation perspective for unsupervised RWSR to combine generalist and specialist models." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 138, + 318, + 480, + 352 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 318, + 480, + 352 + ], + "spans": [ + { + "bbox": [ + 138, + 318, + 480, + 352 + ], + "type": "text", + "content": "- We propose a novel pairwise distance distillation framework, emphasizing transferring intra-model and inter-model distances to enhance the specialist's performance in real-world scenarios." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 138, + 354, + 479, + 376 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 354, + 479, + 376 + ], + "spans": [ + { + "bbox": [ + 138, + 354, + 479, + 376 + ], + "type": "text", + "content": "- Experiments on three benchmarks demonstrate that our approach improves the off-the-shelf models regarding fidelity and perceptual performances." + } + ] + } + ], + "index": 6 + } + ], + "sub_type": "text" + }, + { + "bbox": [ + 132, + 398, + 242, + 410 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 398, + 242, + 410 + ], + "spans": [ + { + "bbox": [ + 132, + 398, + 242, + 410 + ], + "type": "text", + "content": "2 Related Works" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 130, + 425, + 480, + 472 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 425, + 480, + 472 + ], + "spans": [ + { + "bbox": [ + 130, + 425, + 480, + 472 + ], + "type": "text", + "content": "Single-Image Super-Resolution (SISR) methods use deep neural networks of various architectures, such as residual networks [19,36] and transformers [18, 33]. Despite improving the standard SISR performance with dedicated architectures, these methods struggle to generalize to real-world scenarios [34,40]." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 130, + 474, + 481, + 616 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 474, + 481, + 616 + ], + "spans": [ + { + "bbox": [ + 130, + 474, + 481, + 616 + ], + "type": "text", + "content": "Real-World Image Super-Resolution (RWSR) aims to address unknown input degradations. The infeasibility of creating paired training data without known kernels presents a significant challenge in learning specialized models for RWSR. There are two primary strategies in existing RWSR. The first assumes that networks trained on diverse and challenging synthetic degradations will effectively generalize to real-world data [35, 43]. Even though real-world degradations are complex, within a specific setting or deployment, they tend to be limited in the domain and less diverse than the training pipelines. Consequently, [45] proposes customizing the synthetic pipeline to better match specific real-world data. Nonetheless, an unresolved gap exists between synthesized and real-world degradations. Instead, our method operates directly on real-world data and avoids the intrinsic gap from synthetic degradation." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 130, + 617, + 481, + 666 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 617, + 481, + 666 + ], + "spans": [ + { + "bbox": [ + 130, + 617, + 481, + 666 + ], + "type": "text", + "content": "The second strategy connects real-world and synthetic degradation with transfer learning. Most works use image-to-image translation [5,30,40] to adapt the synthetic degradation to real-world ones. Such methods resort to intricate designs like CycleGAN [50], which struggle to replicate real-world degradations [34]" + } + ] + } + ], + "index": 11 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 321, + 91, + 447, + 101 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 321, + 91, + 447, + 101 + ], + "spans": [ + { + "bbox": [ + 321, + 91, + 447, + 101 + ], + "type": "text", + "content": "Pairwise Distance Distillation" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 474, + 92, + 480, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 474, + 92, + 480, + 100 + ], + "spans": [ + { + "bbox": [ + 474, + 92, + 480, + 100 + ], + "type": "text", + "content": "3" + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 2 + }, + { + "para_blocks": [ + { + "bbox": [ + 130, + 116, + 479, + 140 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 116, + 479, + 140 + ], + "spans": [ + { + "bbox": [ + 130, + 116, + 479, + 140 + ], + "type": "text", + "content": "reliably. Our proposed method explores feature distances between model outputs and bypasses the need to mimic real-world degradations." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 130, + 140, + 480, + 213 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 140, + 480, + 213 + ], + "spans": [ + { + "bbox": [ + 130, + 140, + 480, + 213 + ], + "type": "text", + "content": "Knowledge Distillation is initially introduced for model compression [10]. In low-level vision, several works [7, 28, 47] distill large models into more efficient ones by enforcing similarity between their internal features or predictions. Recently, knowledge distillation has served diverse purposes in high-level tasks like segmentation, including domain adaptation [6, 15, 27] and transfer learning [38, 41]. However, such applications are less discussed in low-level vision." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 132, + 230, + 202, + 243 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 230, + 202, + 243 + ], + "spans": [ + { + "bbox": [ + 132, + 230, + 202, + 243 + ], + "type": "text", + "content": "3 Method" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 130, + 255, + 479, + 363 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 255, + 479, + 363 + ], + "spans": [ + { + "bbox": [ + 130, + 255, + 479, + 363 + ], + "type": "text", + "content": "Our method adapts the specialist in synthetic degradation to an unlabeled real-world domain. We use the low-level characteristics of synthetic predictions from the specialists and the knowledge of the generalist model. Fig. 2 is the overview of our approach and the novel Pairwise Distance Distillation (PDD) method, which enforces the consistency of the intra- and inter-model distances to improve the specialist " + }, + { + "bbox": [ + 130, + 255, + 479, + 363 + ], + "type": "inline_equation", + "content": "(M_S)" + }, + { + "bbox": [ + 130, + 255, + 479, + 363 + ], + "type": "text", + "content": "'s real-world predictions. Sec. 3.1 gives notions used in this section and Sec. 3.2 illustrates the unsupervised formulation. Sec. 3.3 explains the details of PDD and Sec. 3.4 provides the full optimization aims and discusses the static and Exponential Moving Average (EMA) versions of our method." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 132, + 380, + 261, + 392 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 380, + 261, + 392 + ], + "spans": [ + { + "bbox": [ + 132, + 380, + 261, + 392 + ], + "type": "text", + "content": "3.1 Definitions & Setup" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 130, + 399, + 480, + 483 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 399, + 480, + 483 + ], + "spans": [ + { + "bbox": [ + 130, + 399, + 480, + 483 + ], + "type": "text", + "content": "When provided with an LR input " + }, + { + "bbox": [ + 130, + 399, + 480, + 483 + ], + "type": "inline_equation", + "content": "X \\in \\mathbb{R}^{H \\times W \\times C}" + }, + { + "bbox": [ + 130, + 399, + 480, + 483 + ], + "type": "text", + "content": " - where " + }, + { + "bbox": [ + 130, + 399, + 480, + 483 + ], + "type": "inline_equation", + "content": "H \\times W" + }, + { + "bbox": [ + 130, + 399, + 480, + 483 + ], + "type": "text", + "content": " is the spatial resolution and " + }, + { + "bbox": [ + 130, + 399, + 480, + 483 + ], + "type": "inline_equation", + "content": "C" + }, + { + "bbox": [ + 130, + 399, + 480, + 483 + ], + "type": "text", + "content": " is the number of color channels - a SISR model aims to reconstruct an HR image " + }, + { + "bbox": [ + 130, + 399, + 480, + 483 + ], + "type": "inline_equation", + "content": "\\hat{Y} \\in \\mathbb{R}^{rH \\times rW \\times C}" + }, + { + "bbox": [ + 130, + 399, + 480, + 483 + ], + "type": "text", + "content": ", scaling by a factor " + }, + { + "bbox": [ + 130, + 399, + 480, + 483 + ], + "type": "inline_equation", + "content": "r" + }, + { + "bbox": [ + 130, + 399, + 480, + 483 + ], + "type": "text", + "content": ". Due to the difficulty in obtaining paired training data, the conventional approach in SISR involves generating an LR image " + }, + { + "bbox": [ + 130, + 399, + 480, + 483 + ], + "type": "inline_equation", + "content": "X" + }, + { + "bbox": [ + 130, + 399, + 480, + 483 + ], + "type": "text", + "content": " by downsampling the HR ground-truth " + }, + { + "bbox": [ + 130, + 399, + 480, + 483 + ], + "type": "inline_equation", + "content": "Y" + }, + { + "bbox": [ + 130, + 399, + 480, + 483 + ], + "type": "text", + "content": ", i.e., " + }, + { + "bbox": [ + 130, + 399, + 480, + 483 + ], + "type": "inline_equation", + "content": "X = D_{\\downarrow}(Y)" + }, + { + "bbox": [ + 130, + 399, + 480, + 483 + ], + "type": "text", + "content": ". This downsampling operation " + }, + { + "bbox": [ + 130, + 399, + 480, + 483 + ], + "type": "inline_equation", + "content": "D_{\\downarrow}" + }, + { + "bbox": [ + 130, + 399, + 480, + 483 + ], + "type": "text", + "content": " typically refers to bicubic interpolation, although other alternatives exist [1]." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 130, + 483, + 480, + 532 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 483, + 480, + 532 + ], + "spans": [ + { + "bbox": [ + 130, + 483, + 480, + 532 + ], + "type": "text", + "content": "In a more extensive setting, it is typical for the LR image to be affected by various factors, such as blurring, noise, compression artifacts, etc. To exhibit these degradations, the LR image " + }, + { + "bbox": [ + 130, + 483, + 480, + 532 + ], + "type": "inline_equation", + "content": "X" + }, + { + "bbox": [ + 130, + 483, + 480, + 532 + ], + "type": "text", + "content": " can be formed by applying " + }, + { + "bbox": [ + 130, + 483, + 480, + 532 + ], + "type": "inline_equation", + "content": "D_{\\downarrow}" + }, + { + "bbox": [ + 130, + 483, + 480, + 532 + ], + "type": "text", + "content": " along with one or multiple degradations " + }, + { + "bbox": [ + 130, + 483, + 480, + 532 + ], + "type": "inline_equation", + "content": "D_{i}" + }, + { + "bbox": [ + 130, + 483, + 480, + 532 + ], + "type": "text", + "content": ", in any order, i.e.," + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 258, + 541, + 480, + 554 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 258, + 541, + 480, + 554 + ], + "spans": [ + { + "bbox": [ + 258, + 541, + 480, + 554 + ], + "type": "interline_equation", + "content": "X = D _ {1} \\circ \\dots \\circ D _ {n} (Y). \\tag {1}", + "image_path": "a8fcc9e420ecef6a9784f1b06f76962b32de7f8bf5f6687ba402ebe41040b20a.jpg" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 130, + 562, + 479, + 611 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 562, + 479, + 611 + ], + "spans": [ + { + "bbox": [ + 130, + 562, + 479, + 611 + ], + "type": "text", + "content": "Here, we include " + }, + { + "bbox": [ + 130, + 562, + 479, + 611 + ], + "type": "inline_equation", + "content": "D_{\\downarrow}" + }, + { + "bbox": [ + 130, + 562, + 479, + 611 + ], + "type": "text", + "content": " as one of the degradations " + }, + { + "bbox": [ + 130, + 562, + 479, + 611 + ], + "type": "inline_equation", + "content": "\\{D_i\\}_{i \\in [1,n]}" + }, + { + "bbox": [ + 130, + 562, + 479, + 611 + ], + "type": "text", + "content": " for simplicity. Now, let's consider two categories of degradation sets consisting of known degradations: " + }, + { + "bbox": [ + 130, + 562, + 479, + 611 + ], + "type": "inline_equation", + "content": "\\mathcal{D}_S" + }, + { + "bbox": [ + 130, + 562, + 479, + 611 + ], + "type": "text", + "content": ", which comprises simple degradations with a few factors, and " + }, + { + "bbox": [ + 130, + 562, + 479, + 611 + ], + "type": "inline_equation", + "content": "\\mathcal{D}_G" + }, + { + "bbox": [ + 130, + 562, + 479, + 611 + ], + "type": "text", + "content": " characterized by complex degradations with a wide range of factors:" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 198, + 620, + 480, + 633 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 198, + 620, + 480, + 633 + ], + "spans": [ + { + "bbox": [ + 198, + 620, + 480, + 633 + ], + "type": "interline_equation", + "content": "\\mathcal {D} _ {S} = \\left\\{D _ {i} \\right\\} _ {i \\in \\mathcal {S}}, \\mathcal {D} _ {G} = \\left\\{D _ {i} \\right\\} _ {i \\in \\mathcal {G}}, \\quad \\text {w h e r e} | \\mathcal {G} | \\gg | \\mathcal {S} |. \\tag {2}", + "image_path": "b082eca583a9b31d5d6446a429f091a8dbbb0a6b63f643e43177668bfe8d10c8.jpg" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 130, + 641, + 480, + 666 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 641, + 480, + 666 + ], + "spans": [ + { + "bbox": [ + 130, + 641, + 480, + 666 + ], + "type": "text", + "content": "In the simplest case for " + }, + { + "bbox": [ + 130, + 641, + 480, + 666 + ], + "type": "inline_equation", + "content": "\\mathcal{D}_S" + }, + { + "bbox": [ + 130, + 641, + 480, + 666 + ], + "type": "text", + "content": ", the set " + }, + { + "bbox": [ + 130, + 641, + 480, + 666 + ], + "type": "inline_equation", + "content": "\\mathcal{D}_S = \\{D_\\downarrow\\}" + }, + { + "bbox": [ + 130, + 641, + 480, + 666 + ], + "type": "text", + "content": ", which represents the standard SISR setup. For the broader case, " + }, + { + "bbox": [ + 130, + 641, + 480, + 666 + ], + "type": "inline_equation", + "content": "\\mathcal{D}_G" + }, + { + "bbox": [ + 130, + 641, + 480, + 666 + ], + "type": "text", + "content": " should encompass diverse and challenging" + } + ] + } + ], + "index": 12 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "spans": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "type": "text", + "content": "4" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 166, + 91, + 228, + 102 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 166, + 91, + 228, + 102 + ], + "spans": [ + { + "bbox": [ + 166, + 91, + 228, + 102 + ], + "type": "text", + "content": "Y. Zhang et al." + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 3 + }, + { + "para_blocks": [ + { + "bbox": [ + 130, + 116, + 479, + 140 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 116, + 479, + 140 + ], + "spans": [ + { + "bbox": [ + 130, + 116, + 479, + 140 + ], + "type": "text", + "content": "degradations beyond " + }, + { + "bbox": [ + 130, + 116, + 479, + 140 + ], + "type": "inline_equation", + "content": "\\{D_{\\downarrow}\\}" + }, + { + "bbox": [ + 130, + 116, + 479, + 140 + ], + "type": "text", + "content": " to ensure that models trained on this data generalize well to blind settings [35, 43]." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 130, + 140, + 482, + 259 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 140, + 482, + 259 + ], + "spans": [ + { + "bbox": [ + 130, + 140, + 482, + 259 + ], + "type": "text", + "content": "Consider two models, " + }, + { + "bbox": [ + 130, + 140, + 482, + 259 + ], + "type": "inline_equation", + "content": "M_S" + }, + { + "bbox": [ + 130, + 140, + 482, + 259 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 130, + 140, + 482, + 259 + ], + "type": "inline_equation", + "content": "M_G" + }, + { + "bbox": [ + 130, + 140, + 482, + 259 + ], + "type": "text", + "content": ", trained on datasets constructed with " + }, + { + "bbox": [ + 130, + 140, + 482, + 259 + ], + "type": "inline_equation", + "content": "\\mathcal{D}_S" + }, + { + "bbox": [ + 130, + 140, + 482, + 259 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 130, + 140, + 482, + 259 + ], + "type": "inline_equation", + "content": "\\mathcal{D}_G" + }, + { + "bbox": [ + 130, + 140, + 482, + 259 + ], + "type": "text", + "content": " respectively. Both models are optimized with common supervised SR learning losses [13, 36] including MSE-based loss, VGG loss, and adversarial training (see Eq. (9)). Since " + }, + { + "bbox": [ + 130, + 140, + 482, + 259 + ], + "type": "inline_equation", + "content": "\\mathcal{D}_S" + }, + { + "bbox": [ + 130, + 140, + 482, + 259 + ], + "type": "text", + "content": " comprises a narrower range of degradations compared to " + }, + { + "bbox": [ + 130, + 140, + 482, + 259 + ], + "type": "inline_equation", + "content": "\\mathcal{D}_G" + }, + { + "bbox": [ + 130, + 140, + 482, + 259 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 130, + 140, + 482, + 259 + ], + "type": "inline_equation", + "content": "M_S" + }, + { + "bbox": [ + 130, + 140, + 482, + 259 + ], + "type": "text", + "content": " is a \"specialist\" model for the limited domain defined by " + }, + { + "bbox": [ + 130, + 140, + 482, + 259 + ], + "type": "inline_equation", + "content": "\\mathcal{D}_S" + }, + { + "bbox": [ + 130, + 140, + 482, + 259 + ], + "type": "text", + "content": ". On the other hand, " + }, + { + "bbox": [ + 130, + 140, + 482, + 259 + ], + "type": "inline_equation", + "content": "M_G" + }, + { + "bbox": [ + 130, + 140, + 482, + 259 + ], + "type": "text", + "content": " is a \"generalist\" model as it achieves moderate performance across the broader domain derived by " + }, + { + "bbox": [ + 130, + 140, + 482, + 259 + ], + "type": "inline_equation", + "content": "\\mathcal{D}_G" + }, + { + "bbox": [ + 130, + 140, + 482, + 259 + ], + "type": "text", + "content": ". Comparing " + }, + { + "bbox": [ + 130, + 140, + 482, + 259 + ], + "type": "inline_equation", + "content": "M_S" + }, + { + "bbox": [ + 130, + 140, + 482, + 259 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 130, + 140, + 482, + 259 + ], + "type": "inline_equation", + "content": "M_G" + }, + { + "bbox": [ + 130, + 140, + 482, + 259 + ], + "type": "text", + "content": ", we assume the specialist's superiority to the generalist in " + }, + { + "bbox": [ + 130, + 140, + 482, + 259 + ], + "type": "inline_equation", + "content": "\\mathcal{D}_S" + }, + { + "bbox": [ + 130, + 140, + 482, + 259 + ], + "type": "text", + "content": " and the generalist's better handling of unknown degradations. These assumptions are empirically validated in Tab. 1." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 131, + 277, + 383, + 290 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 131, + 277, + 383, + 290 + ], + "spans": [ + { + "bbox": [ + 131, + 277, + 383, + 290 + ], + "type": "text", + "content": "3.2 Unsupervised Learning Through Distillation" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 131, + 298, + 482, + 396 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 131, + 298, + 482, + 396 + ], + "spans": [ + { + "bbox": [ + 131, + 298, + 482, + 396 + ], + "type": "text", + "content": "Our approach aligns with the unsupervised setup [5, 30, 40], wherein we work with a collection of real-world LR images alongside unpaired HR clean images. We denote the set of LR images as " + }, + { + "bbox": [ + 131, + 298, + 482, + 396 + ], + "type": "inline_equation", + "content": "\\{X^U\\}" + }, + { + "bbox": [ + 131, + 298, + 482, + 396 + ], + "type": "text", + "content": " with the same unknown degradations " + }, + { + "bbox": [ + 131, + 298, + 482, + 396 + ], + "type": "inline_equation", + "content": "\\{D_i\\}_{i\\in \\mathcal{U}}" + }, + { + "bbox": [ + 131, + 298, + 482, + 396 + ], + "type": "text", + "content": ". Additionally, we generate LR counterparts " + }, + { + "bbox": [ + 131, + 298, + 482, + 396 + ], + "type": "inline_equation", + "content": "\\{X^L\\}" + }, + { + "bbox": [ + 131, + 298, + 482, + 396 + ], + "type": "text", + "content": " from HR images using specific degradations from " + }, + { + "bbox": [ + 131, + 298, + 482, + 396 + ], + "type": "inline_equation", + "content": "\\mathcal{D}_S" + }, + { + "bbox": [ + 131, + 298, + 482, + 396 + ], + "type": "text", + "content": ", e.g. bicubic interpolation. The specialist " + }, + { + "bbox": [ + 131, + 298, + 482, + 396 + ], + "type": "inline_equation", + "content": "M_S" + }, + { + "bbox": [ + 131, + 298, + 482, + 396 + ], + "type": "text", + "content": " is adept in restoring " + }, + { + "bbox": [ + 131, + 298, + 482, + 396 + ], + "type": "inline_equation", + "content": "X^L" + }, + { + "bbox": [ + 131, + 298, + 482, + 396 + ], + "type": "text", + "content": ", while " + }, + { + "bbox": [ + 131, + 298, + 482, + 396 + ], + "type": "inline_equation", + "content": "M_G" + }, + { + "bbox": [ + 131, + 298, + 482, + 396 + ], + "type": "text", + "content": " performs moderately in restoring " + }, + { + "bbox": [ + 131, + 298, + 482, + 396 + ], + "type": "inline_equation", + "content": "X^L" + }, + { + "bbox": [ + 131, + 298, + 482, + 396 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 131, + 298, + 482, + 396 + ], + "type": "inline_equation", + "content": "X^U" + }, + { + "bbox": [ + 131, + 298, + 482, + 396 + ], + "type": "text", + "content": ". We aim to adapt " + }, + { + "bbox": [ + 131, + 298, + 482, + 396 + ], + "type": "inline_equation", + "content": "M_S" + }, + { + "bbox": [ + 131, + 298, + 482, + 396 + ], + "type": "text", + "content": " to " + }, + { + "bbox": [ + 131, + 298, + 482, + 396 + ], + "type": "inline_equation", + "content": "\\{X^U\\}" + }, + { + "bbox": [ + 131, + 298, + 482, + 396 + ], + "type": "text", + "content": " with the aid of " + }, + { + "bbox": [ + 131, + 298, + 482, + 396 + ], + "type": "inline_equation", + "content": "M_G" + }, + { + "bbox": [ + 131, + 298, + 482, + 396 + ], + "type": "text", + "content": ". As in [10], a naive distillation approach is having " + }, + { + "bbox": [ + 131, + 298, + 482, + 396 + ], + "type": "inline_equation", + "content": "M_S" + }, + { + "bbox": [ + 131, + 298, + 482, + 396 + ], + "type": "text", + "content": " imitate the predictions from " + }, + { + "bbox": [ + 131, + 298, + 482, + 396 + ], + "type": "inline_equation", + "content": "M_G" + }, + { + "bbox": [ + 131, + 298, + 482, + 396 + ], + "type": "text", + "content": " for unlabeled inputs:" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 195, + 403, + 480, + 435 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 195, + 403, + 480, + 435 + ], + "spans": [ + { + "bbox": [ + 195, + 403, + 480, + 435 + ], + "type": "interline_equation", + "content": "\\begin{array}{l} \\hat {Y} _ {S} ^ {L} = M _ {S} (X ^ {L}), \\hat {Y} _ {S} ^ {U} = M _ {S} (X ^ {U}), \\hat {Y} _ {G} ^ {U} = M _ {G} (X ^ {U}), \\\\ \\mathcal {L} _ {N D} = \\mathcal {L} _ {L} \\left(\\hat {Y} _ {S} ^ {U}, \\hat {Y} _ {G} ^ {U}\\right) + \\lambda \\mathcal {L} _ {L} \\left(\\hat {Y} _ {S} ^ {L}, Y ^ {L}\\right), \\tag {3} \\\\ \\end{array}", + "image_path": "e199eee94c2b36ba0ae21a970d4fa8ce08d9e3741d4e10bbdfffdd3431d44975.jpg" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 130, + 445, + 482, + 506 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 445, + 482, + 506 + ], + "spans": [ + { + "bbox": [ + 130, + 445, + 482, + 506 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 130, + 445, + 482, + 506 + ], + "type": "inline_equation", + "content": "Y^{L}" + }, + { + "bbox": [ + 130, + 445, + 482, + 506 + ], + "type": "text", + "content": " represents the ground-truth image in the labeled domain, " + }, + { + "bbox": [ + 130, + 445, + 482, + 506 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_L" + }, + { + "bbox": [ + 130, + 445, + 482, + 506 + ], + "type": "text", + "content": " denotes the supervised loss formulated in Eq. (9), and " + }, + { + "bbox": [ + 130, + 445, + 482, + 506 + ], + "type": "inline_equation", + "content": "\\lambda" + }, + { + "bbox": [ + 130, + 445, + 482, + 506 + ], + "type": "text", + "content": " is a scale factor to balance between the distillation and primary objective of " + }, + { + "bbox": [ + 130, + 445, + 482, + 506 + ], + "type": "inline_equation", + "content": "M_S" + }, + { + "bbox": [ + 130, + 445, + 482, + 506 + ], + "type": "text", + "content": ". However, this simple imitation approach relies solely on information from the generalist " + }, + { + "bbox": [ + 130, + 445, + 482, + 506 + ], + "type": "inline_equation", + "content": "M_G" + }, + { + "bbox": [ + 130, + 445, + 482, + 506 + ], + "type": "text", + "content": " and fails to harness the strengths of both models." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 131, + 523, + 348, + 536 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 131, + 523, + 348, + 536 + ], + "spans": [ + { + "bbox": [ + 131, + 523, + 348, + 536 + ], + "type": "text", + "content": "3.3 Pairwise Distance Distillation (PDD)" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 130, + 543, + 482, + 569 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 543, + 482, + 569 + ], + "spans": [ + { + "bbox": [ + 130, + 543, + 482, + 569 + ], + "type": "text", + "content": "Consider any pair of LR images - a real-world " + }, + { + "bbox": [ + 130, + 543, + 482, + 569 + ], + "type": "inline_equation", + "content": "X^U" + }, + { + "bbox": [ + 130, + 543, + 482, + 569 + ], + "type": "text", + "content": " and a synthesized " + }, + { + "bbox": [ + 130, + 543, + 482, + 569 + ], + "type": "inline_equation", + "content": "X^L" + }, + { + "bbox": [ + 130, + 543, + 482, + 569 + ], + "type": "text", + "content": ". As shown in Fig. 2, we obtain four predictions by applying " + }, + { + "bbox": [ + 130, + 543, + 482, + 569 + ], + "type": "inline_equation", + "content": "M_S" + }, + { + "bbox": [ + 130, + 543, + 482, + 569 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 130, + 543, + 482, + 569 + ], + "type": "inline_equation", + "content": "M_G" + }, + { + "bbox": [ + 130, + 543, + 482, + 569 + ], + "type": "text", + "content": ":" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 224, + 577, + 480, + 608 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 224, + 577, + 480, + 608 + ], + "spans": [ + { + "bbox": [ + 224, + 577, + 480, + 608 + ], + "type": "interline_equation", + "content": "\\begin{array}{l} \\hat {Y} _ {S} ^ {U} = M _ {S} (X ^ {U}), \\hat {Y} _ {G} ^ {U} = M _ {G} (X ^ {U}), \\\\ \\text {a n d} \\hat {Y} _ {S} ^ {L} = M _ {S} \\left(X ^ {L}\\right), \\hat {Y} _ {G} ^ {L} = M _ {G} \\left(X ^ {L}\\right). \\tag {4} \\\\ \\end{array}", + "image_path": "b62cb2888b73848c2f72fa9ee855bcf2f1debd221dc27927050c7f058fbb8f4f.jpg" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 130, + 617, + 482, + 666 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 617, + 482, + 666 + ], + "spans": [ + { + "bbox": [ + 130, + 617, + 482, + 666 + ], + "type": "text", + "content": "Our distillation explores the relationship between predictions in Eq. (4) to seek for knowledge combination. We use VGG features [31] of the predictions as our basis for exploration. VGG, as a classification model, inherently captures semantic information. Yet it also captures low-level characteristics related to" + } + ] + } + ], + "index": 11 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 321, + 91, + 447, + 101 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 321, + 91, + 447, + 101 + ], + "spans": [ + { + "bbox": [ + 321, + 91, + 447, + 101 + ], + "type": "text", + "content": "Pairwise Distance Distillation" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 474, + 92, + 480, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 474, + 92, + 480, + 100 + ], + "spans": [ + { + "bbox": [ + 474, + 92, + 480, + 100 + ], + "type": "text", + "content": "5" + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 4 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 151, + 115, + 463, + 196 + ], + "blocks": [ + { + "bbox": [ + 151, + 115, + 463, + 196 + ], + "lines": [ + { + "bbox": [ + 151, + 115, + 463, + 196 + ], + "spans": [ + { + "bbox": [ + 151, + 115, + 463, + 196 + ], + "type": "image", + "image_path": "12350873bd7515b36d1bb9eac7095c5779bdad28b5317afcef3ab01b1e90f8ef.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 130, + 205, + 482, + 262 + ], + "lines": [ + { + "bbox": [ + 130, + 205, + 482, + 262 + ], + "spans": [ + { + "bbox": [ + 130, + 205, + 482, + 262 + ], + "type": "text", + "content": "Fig. 2: Schematic of Pairwise Distance Distillation (PDD). " + }, + { + "bbox": [ + 130, + 205, + 482, + 262 + ], + "type": "inline_equation", + "content": "X^{U}" + }, + { + "bbox": [ + 130, + 205, + 482, + 262 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 130, + 205, + 482, + 262 + ], + "type": "inline_equation", + "content": "X^{L}" + }, + { + "bbox": [ + 130, + 205, + 482, + 262 + ], + "type": "text", + "content": " are unlabeled (real-world) and labeled (synthetic) inputs. Model " + }, + { + "bbox": [ + 130, + 205, + 482, + 262 + ], + "type": "inline_equation", + "content": "M_{G}" + }, + { + "bbox": [ + 130, + 205, + 482, + 262 + ], + "type": "text", + "content": " is a generalist trained with extensive synthetic pipeline, while " + }, + { + "bbox": [ + 130, + 205, + 482, + 262 + ], + "type": "inline_equation", + "content": "M_{S}" + }, + { + "bbox": [ + 130, + 205, + 482, + 262 + ], + "type": "text", + "content": " specializes in " + }, + { + "bbox": [ + 130, + 205, + 482, + 262 + ], + "type": "inline_equation", + "content": "X^{L}" + }, + { + "bbox": [ + 130, + 205, + 482, + 262 + ], + "type": "text", + "content": ". The prediction " + }, + { + "bbox": [ + 130, + 205, + 482, + 262 + ], + "type": "inline_equation", + "content": "\\hat{Y}_S^L" + }, + { + "bbox": [ + 130, + 205, + 482, + 262 + ], + "type": "text", + "content": " is supervised by its ground truth throughout training. PDD enforces the consistency between " + }, + { + "bbox": [ + 130, + 205, + 482, + 262 + ], + "type": "inline_equation", + "content": "\\{\\varDelta_U, \\varDelta_L\\}" + }, + { + "bbox": [ + 130, + 205, + 482, + 262 + ], + "type": "text", + "content": " and between " + }, + { + "bbox": [ + 130, + 205, + 482, + 262 + ], + "type": "inline_equation", + "content": "\\{d_S, d_G\\}" + }, + { + "bbox": [ + 130, + 205, + 482, + 262 + ], + "type": "text", + "content": " to improve " + }, + { + "bbox": [ + 130, + 205, + 482, + 262 + ], + "type": "inline_equation", + "content": "M_{S}" + }, + { + "bbox": [ + 130, + 205, + 482, + 262 + ], + "type": "text", + "content": "'s real-world performance." + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_caption" + } + ], + "index": 2 + }, + { + "bbox": [ + 130, + 285, + 480, + 309 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 285, + 480, + 309 + ], + "spans": [ + { + "bbox": [ + 130, + 285, + 480, + 309 + ], + "type": "text", + "content": "image quality [37], e.g. blur and sharpness, and is widely used in the perceptual loss for super-resolution tasks [13, 35, 36]." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 130, + 310, + 480, + 394 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 310, + 480, + 394 + ], + "spans": [ + { + "bbox": [ + 130, + 310, + 480, + 394 + ], + "type": "text", + "content": "As a generalist, " + }, + { + "bbox": [ + 130, + 310, + 480, + 394 + ], + "type": "inline_equation", + "content": "M_G" + }, + { + "bbox": [ + 130, + 310, + 480, + 394 + ], + "type": "text", + "content": " outputs synthetic and real-world predictions with small quality gaps, which is characterized by the overlapping distributions of low-level characteristics of " + }, + { + "bbox": [ + 130, + 310, + 480, + 394 + ], + "type": "inline_equation", + "content": "\\{\\hat{Y}_G^L\\}" + }, + { + "bbox": [ + 130, + 310, + 480, + 394 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 130, + 310, + 480, + 394 + ], + "type": "inline_equation", + "content": "\\{\\hat{Y}_G^U\\}" + }, + { + "bbox": [ + 130, + 310, + 480, + 394 + ], + "type": "text", + "content": " [20, 21]. Following [20], Fig. 5a projects the low-level features of " + }, + { + "bbox": [ + 130, + 310, + 480, + 394 + ], + "type": "inline_equation", + "content": "\\{\\hat{Y}_G^L\\}" + }, + { + "bbox": [ + 130, + 310, + 480, + 394 + ], + "type": "text", + "content": " (DIV2K) and " + }, + { + "bbox": [ + 130, + 310, + 480, + 394 + ], + "type": "inline_equation", + "content": "\\{\\hat{Y}_G^U\\}" + }, + { + "bbox": [ + 130, + 310, + 480, + 394 + ], + "type": "text", + "content": " (NTIRE20). The generalist's coverage strongly overlaps while the specialist's predictions are well-separated (see Fig. 5b). The within-model relationship of the generalist exhibits the generalizability to real-world degradation." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 130, + 395, + 481, + 491 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 395, + 481, + 491 + ], + "spans": [ + { + "bbox": [ + 130, + 395, + 481, + 491 + ], + "type": "text", + "content": "On the other side, the high image quality of " + }, + { + "bbox": [ + 130, + 395, + 481, + 491 + ], + "type": "inline_equation", + "content": "\\{\\hat{Y}_S^L\\}" + }, + { + "bbox": [ + 130, + 395, + 481, + 491 + ], + "type": "text", + "content": ", due to " + }, + { + "bbox": [ + 130, + 395, + 481, + 491 + ], + "type": "inline_equation", + "content": "M_S" + }, + { + "bbox": [ + 130, + 395, + 481, + 491 + ], + "type": "text", + "content": "'s specialization, makes its low-level characteristics a valuable reference. To integrate the knowledge for real-world performance, we propose to make the specialist imitate the generalist's within-model relationship, that is, pushing the low-level characteristics of real-world predictions from " + }, + { + "bbox": [ + 130, + 395, + 481, + 491 + ], + "type": "inline_equation", + "content": "M_S" + }, + { + "bbox": [ + 130, + 395, + 481, + 491 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 130, + 395, + 481, + 491 + ], + "type": "inline_equation", + "content": "\\{\\hat{Y}_S^U\\}" + }, + { + "bbox": [ + 130, + 395, + 481, + 491 + ], + "type": "text", + "content": ", to be similar to that of its synthetic predictions " + }, + { + "bbox": [ + 130, + 395, + 481, + 491 + ], + "type": "inline_equation", + "content": "\\{\\hat{Y}_S^L\\}" + }, + { + "bbox": [ + 130, + 395, + 481, + 491 + ], + "type": "text", + "content": ". To achieve this aim, we first assume such similarity is satisfied for " + }, + { + "bbox": [ + 130, + 395, + 481, + 491 + ], + "type": "inline_equation", + "content": "M_S" + }, + { + "bbox": [ + 130, + 395, + 481, + 491 + ], + "type": "text", + "content": " and conclude the following consistencies about VGG feature distances between predictions:" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 135, + 501, + 479, + 548 + ], + "type": "list", + "angle": 0, + "index": 9, + "blocks": [ + { + "bbox": [ + 135, + 501, + 479, + 523 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 135, + 501, + 479, + 523 + ], + "spans": [ + { + "bbox": [ + 135, + 501, + 479, + 523 + ], + "type": "text", + "content": "1. Intra-model distances for predictions of the same input pair should be consistent across " + }, + { + "bbox": [ + 135, + 501, + 479, + 523 + ], + "type": "inline_equation", + "content": "M_S" + }, + { + "bbox": [ + 135, + 501, + 479, + 523 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 135, + 501, + 479, + 523 + ], + "type": "inline_equation", + "content": "M_G" + }, + { + "bbox": [ + 135, + 501, + 479, + 523 + ], + "type": "text", + "content": " (green shade in Fig. 2)." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 135, + 525, + 479, + 548 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 135, + 525, + 479, + 548 + ], + "spans": [ + { + "bbox": [ + 135, + 525, + 479, + 548 + ], + "type": "text", + "content": "2. Inter-model distances for predictions of a single input should be consistent across synthetic and real-world domains (orange shade in Fig. 2)." + } + ] + } + ], + "index": 8 + } + ], + "sub_type": "text" + }, + { + "bbox": [ + 130, + 557, + 480, + 640 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 557, + 480, + 640 + ], + "spans": [ + { + "bbox": [ + 130, + 557, + 480, + 640 + ], + "type": "text", + "content": "Our rationales for these consistencies are as follows. First, we recall the low-level characteristics of predictions from the same model are similar. We discussed it for " + }, + { + "bbox": [ + 130, + 557, + 480, + 640 + ], + "type": "inline_equation", + "content": "M_G" + }, + { + "bbox": [ + 130, + 557, + 480, + 640 + ], + "type": "text", + "content": " with Fig. 5a and made such an assumption for " + }, + { + "bbox": [ + 130, + 557, + 480, + 640 + ], + "type": "inline_equation", + "content": "M_S" + }, + { + "bbox": [ + 130, + 557, + 480, + 640 + ], + "type": "text", + "content": ". Thus, the distance between the same model's predictions " + }, + { + "bbox": [ + 130, + 557, + 480, + 640 + ], + "type": "inline_equation", + "content": "\\{\\hat{Y}_G^L,\\hat{Y}_G^U\\}" + }, + { + "bbox": [ + 130, + 557, + 480, + 640 + ], + "type": "text", + "content": " (or " + }, + { + "bbox": [ + 130, + 557, + 480, + 640 + ], + "type": "inline_equation", + "content": "\\{\\hat{Y}_S^L,\\hat{Y}_S^U\\}" + }, + { + "bbox": [ + 130, + 557, + 480, + 640 + ], + "type": "text", + "content": "), i.e. intra-model distance, reflects mainly semantic differences due to the close low-level characteristics. As it naturally follows that the applied SR models should not change semantics, intra-model distances should be consistent for " + }, + { + "bbox": [ + 130, + 557, + 480, + 640 + ], + "type": "inline_equation", + "content": "M_S" + }, + { + "bbox": [ + 130, + 557, + 480, + 640 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 130, + 557, + 480, + 640 + ], + "type": "inline_equation", + "content": "M_G" + }, + { + "bbox": [ + 130, + 557, + 480, + 640 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 130, + 641, + 481, + 666 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 641, + 481, + 666 + ], + "spans": [ + { + "bbox": [ + 130, + 641, + 481, + 666 + ], + "type": "text", + "content": "Second, we note the low-level characteristics of predictions from the two models differ, e.g. " + }, + { + "bbox": [ + 130, + 641, + 481, + 666 + ], + "type": "inline_equation", + "content": "\\hat{Y}_S^L" + }, + { + "bbox": [ + 130, + 641, + 481, + 666 + ], + "type": "text", + "content": " has higher quality than the counterpart " + }, + { + "bbox": [ + 130, + 641, + 481, + 666 + ], + "type": "inline_equation", + "content": "\\hat{Y}_G^L" + }, + { + "bbox": [ + 130, + 641, + 481, + 666 + ], + "type": "text", + "content": ". The distance" + } + ] + } + ], + "index": 11 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "spans": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "type": "text", + "content": "6" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 166, + 91, + 229, + 102 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 166, + 91, + 229, + 102 + ], + "spans": [ + { + "bbox": [ + 166, + 91, + 229, + 102 + ], + "type": "text", + "content": "Y. Zhang et al." + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 5 + }, + { + "para_blocks": [ + { + "bbox": [ + 130, + 114, + 479, + 175 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 114, + 479, + 175 + ], + "spans": [ + { + "bbox": [ + 130, + 114, + 479, + 175 + ], + "type": "text", + "content": "between the two predictions " + }, + { + "bbox": [ + 130, + 114, + 479, + 175 + ], + "type": "inline_equation", + "content": "\\{\\hat{Y}_G^L,\\hat{Y}_S^L\\}" + }, + { + "bbox": [ + 130, + 114, + 479, + 175 + ], + "type": "text", + "content": " (or " + }, + { + "bbox": [ + 130, + 114, + 479, + 175 + ], + "type": "inline_equation", + "content": "\\{\\hat{Y}_G^U,\\hat{Y}_S^U\\}" + }, + { + "bbox": [ + 130, + 114, + 479, + 175 + ], + "type": "text", + "content": ") of a given sample, i.e. inter-model distance, captures such differences in low-level characteristics, as semantics stay constant for the predictions of the same input. Given the similarity of low-level characteristics within a single model, the inter-model distances for synthetic and real-world samples should be consistent." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 130, + 176, + 479, + 224 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 176, + 479, + 224 + ], + "spans": [ + { + "bbox": [ + 130, + 176, + 479, + 224 + ], + "type": "text", + "content": "We structure the two consistencies as distillations on the intra- and inter-model distances. Our method encourages " + }, + { + "bbox": [ + 130, + 176, + 479, + 224 + ], + "type": "inline_equation", + "content": "M_S" + }, + { + "bbox": [ + 130, + 176, + 479, + 224 + ], + "type": "text", + "content": "'s real-world predictions " + }, + { + "bbox": [ + 130, + 176, + 479, + 224 + ], + "type": "inline_equation", + "content": "\\{\\hat{Y}_S^U\\}" + }, + { + "bbox": [ + 130, + 176, + 479, + 224 + ], + "type": "text", + "content": " to have similar low-level characteristics as its synthetic predictions " + }, + { + "bbox": [ + 130, + 176, + 479, + 224 + ], + "type": "inline_equation", + "content": "\\{\\hat{Y}_S^L\\}" + }, + { + "bbox": [ + 130, + 176, + 479, + 224 + ], + "type": "text", + "content": ", taking it as a reference from high-quality images." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 131, + 225, + 480, + 283 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 131, + 225, + 480, + 283 + ], + "spans": [ + { + "bbox": [ + 131, + 225, + 480, + 283 + ], + "type": "text", + "content": "Intra-model Distance Distillation enforces consistency on the distances between predictions from the same model (green shading in Fig. 2). Consider the " + }, + { + "bbox": [ + 131, + 225, + 480, + 283 + ], + "type": "inline_equation", + "content": "\\ell_1" + }, + { + "bbox": [ + 131, + 225, + 480, + 283 + ], + "type": "text", + "content": " distance " + }, + { + "bbox": [ + 131, + 225, + 480, + 283 + ], + "type": "inline_equation", + "content": "d_G^{ij}" + }, + { + "bbox": [ + 131, + 225, + 480, + 283 + ], + "type": "text", + "content": " between predictions from the generalist " + }, + { + "bbox": [ + 131, + 225, + 480, + 283 + ], + "type": "inline_equation", + "content": "M_G" + }, + { + "bbox": [ + 131, + 225, + 480, + 283 + ], + "type": "text", + "content": ", i.e. " + }, + { + "bbox": [ + 131, + 225, + 480, + 283 + ], + "type": "inline_equation", + "content": "\\hat{Y}_G^L" + }, + { + "bbox": [ + 131, + 225, + 480, + 283 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 131, + 225, + 480, + 283 + ], + "type": "inline_equation", + "content": "\\hat{Y}_G^U" + }, + { + "bbox": [ + 131, + 225, + 480, + 283 + ], + "type": "text", + "content": ", in VGG feature space, and the distance " + }, + { + "bbox": [ + 131, + 225, + 480, + 283 + ], + "type": "inline_equation", + "content": "d_S^{ij}" + }, + { + "bbox": [ + 131, + 225, + 480, + 283 + ], + "type": "text", + "content": " for the specialist " + }, + { + "bbox": [ + 131, + 225, + 480, + 283 + ], + "type": "inline_equation", + "content": "M_S" + }, + { + "bbox": [ + 131, + 225, + 480, + 283 + ], + "type": "text", + "content": " is similarly defined:" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 244, + 284, + 480, + 305 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 244, + 284, + 480, + 305 + ], + "spans": [ + { + "bbox": [ + 244, + 284, + 480, + 305 + ], + "type": "interline_equation", + "content": "d _ {G} ^ {i j} = \\left\\| \\Phi_ {i j} \\left(\\hat {Y} _ {G} ^ {L}\\right) - \\Phi_ {i j} \\left(\\hat {Y} _ {G} ^ {U}\\right) \\right\\| _ {1}, \\tag {5}", + "image_path": "fe1bb05442c094e17ce2a88d53fc2e3791ad295798d096614ac872ad54f2a735.jpg" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 244, + 300, + 370, + 314 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 244, + 300, + 370, + 314 + ], + "spans": [ + { + "bbox": [ + 244, + 300, + 370, + 314 + ], + "type": "interline_equation", + "content": "d _ {S} ^ {i j} = \\left\\| \\Phi_ {i j} \\left(\\hat {Y} _ {S} ^ {L}\\right) - \\Phi_ {i j} \\left(\\hat {Y} _ {S} ^ {U}\\right) \\right\\| _ {1},", + "image_path": "df4eaabacfc6e40095fe662951cee9be1ecee24ececc4080be44d63ece1ce9b1.jpg" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 131, + 316, + 479, + 366 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 131, + 316, + 479, + 366 + ], + "spans": [ + { + "bbox": [ + 131, + 316, + 479, + 366 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 131, + 316, + 479, + 366 + ], + "type": "inline_equation", + "content": "\\varPhi_{ij}" + }, + { + "bbox": [ + 131, + 316, + 479, + 366 + ], + "type": "text", + "content": " refers to the " + }, + { + "bbox": [ + 131, + 316, + 479, + 366 + ], + "type": "inline_equation", + "content": "j" + }, + { + "bbox": [ + 131, + 316, + 479, + 366 + ], + "type": "text", + "content": "-th layer in the " + }, + { + "bbox": [ + 131, + 316, + 479, + 366 + ], + "type": "inline_equation", + "content": "i" + }, + { + "bbox": [ + 131, + 316, + 479, + 366 + ], + "type": "text", + "content": "-th residual block of VGG19. For " + }, + { + "bbox": [ + 131, + 316, + 479, + 366 + ], + "type": "inline_equation", + "content": "d_G^{ij}, d_S^{ij} \\in \\mathbb{R}^{c \\times h \\times w}, h \\times w" + }, + { + "bbox": [ + 131, + 316, + 479, + 366 + ], + "type": "text", + "content": " is the spatial resolution and " + }, + { + "bbox": [ + 131, + 316, + 479, + 366 + ], + "type": "inline_equation", + "content": "c" + }, + { + "bbox": [ + 131, + 316, + 479, + 366 + ], + "type": "text", + "content": " is the number of channels. We enforce the consistency between " + }, + { + "bbox": [ + 131, + 316, + 479, + 366 + ], + "type": "inline_equation", + "content": "d_G^{ij}" + }, + { + "bbox": [ + 131, + 316, + 479, + 366 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 131, + 316, + 479, + 366 + ], + "type": "inline_equation", + "content": "d_S^{ij}" + }, + { + "bbox": [ + 131, + 316, + 479, + 366 + ], + "type": "text", + "content": " by minimizing their difference measured with the Cross-Entropy (CE) " + }, + { + "bbox": [ + 131, + 316, + 479, + 366 + ], + "type": "inline_equation", + "content": "R_{intra}" + }, + { + "bbox": [ + 131, + 316, + 479, + 366 + ], + "type": "text", + "content": ":" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 198, + 372, + 479, + 399 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 198, + 372, + 479, + 399 + ], + "spans": [ + { + "bbox": [ + 198, + 372, + 479, + 399 + ], + "type": "interline_equation", + "content": "R _ {i n t r a} = - \\frac {1}{h w} \\sum_ {i, j} \\sum_ {m, n} \\mathbf {S} \\left(d _ {G} ^ {i j} [ m, n ]\\right) \\log \\mathbf {S} \\left(d _ {S} ^ {i j} [ m, n ]\\right), \\tag {6}", + "image_path": "09129e3ead3e81a94d5730e1969101511b09f16d0c7f71fd184c5e65b984175c.jpg" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 131, + 404, + 479, + 452 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 131, + 404, + 479, + 452 + ], + "spans": [ + { + "bbox": [ + 131, + 404, + 479, + 452 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 131, + 404, + 479, + 452 + ], + "type": "inline_equation", + "content": "[m,n]" + }, + { + "bbox": [ + 131, + 404, + 479, + 452 + ], + "type": "text", + "content": " is a spatial index of the feature map, and " + }, + { + "bbox": [ + 131, + 404, + 479, + 452 + ], + "type": "inline_equation", + "content": "\\mathbf{S}(\\cdot)" + }, + { + "bbox": [ + 131, + 404, + 479, + 452 + ], + "type": "text", + "content": " denotes the SoftMax function with the input size being " + }, + { + "bbox": [ + 131, + 404, + 479, + 452 + ], + "type": "inline_equation", + "content": "\\mathbb{R}^c" + }, + { + "bbox": [ + 131, + 404, + 479, + 452 + ], + "type": "text", + "content": ". The element-wise CE reflects the negative log-likelihood of distances locally. While other measures are feasible, CE is empirically a good choice in favoring the overall sharpness (see Sec. 4.4)." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 131, + 452, + 480, + 536 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 131, + 452, + 480, + 536 + ], + "spans": [ + { + "bbox": [ + 131, + 452, + 480, + 536 + ], + "type": "text", + "content": "Inter-model Distance Distillation enforces the consistency of changes in low-level characteristics between predictions from different models (orange shading in Fig. 2). Given the two predictions for a single input, i.e. " + }, + { + "bbox": [ + 131, + 452, + 480, + 536 + ], + "type": "inline_equation", + "content": "\\{\\hat{Y}_G^L,\\hat{Y}_S^L\\}" + }, + { + "bbox": [ + 131, + 452, + 480, + 536 + ], + "type": "text", + "content": " (or " + }, + { + "bbox": [ + 131, + 452, + 480, + 536 + ], + "type": "inline_equation", + "content": "\\{\\hat{Y}_G^U,\\hat{Y}_S^U\\}" + }, + { + "bbox": [ + 131, + 452, + 480, + 536 + ], + "type": "text", + "content": "), we first calculate their feature distance " + }, + { + "bbox": [ + 131, + 452, + 480, + 536 + ], + "type": "inline_equation", + "content": "\\varDelta_L^{ij}" + }, + { + "bbox": [ + 131, + 452, + 480, + 536 + ], + "type": "text", + "content": " (or " + }, + { + "bbox": [ + 131, + 452, + 480, + 536 + ], + "type": "inline_equation", + "content": "\\varDelta_U^{ij}" + }, + { + "bbox": [ + 131, + 452, + 480, + 536 + ], + "type": "text", + "content": ") to represent the differences in low-level characteristics. However, enforcing consistency on " + }, + { + "bbox": [ + 131, + 452, + 480, + 536 + ], + "type": "inline_equation", + "content": "\\varDelta_L^{ij}" + }, + { + "bbox": [ + 131, + 452, + 480, + 536 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 131, + 452, + 480, + 536 + ], + "type": "inline_equation", + "content": "\\varDelta_U^{ij}" + }, + { + "bbox": [ + 131, + 452, + 480, + 536 + ], + "type": "text", + "content": " is less straightforward, as the spatial layout of corresponding inputs can differ. Thus, we compute the inter-model distances with the following:" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 233, + 541, + 480, + 562 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 233, + 541, + 480, + 562 + ], + "spans": [ + { + "bbox": [ + 233, + 541, + 480, + 562 + ], + "type": "interline_equation", + "content": "\\Delta_ {L} ^ {i j} = \\operatorname {G r a m} \\left(\\Phi_ {i j} \\left(\\hat {Y} _ {S} ^ {L}\\right) - \\Phi_ {i j} \\left(\\hat {Y} _ {G} ^ {L}\\right)\\right), \\tag {7}", + "image_path": "3f7ada1caaf19c12436075af11808ef2a01dc61229bc764e8906fdb6c609ad5c.jpg" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 233, + 559, + 381, + 572 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 233, + 559, + 381, + 572 + ], + "spans": [ + { + "bbox": [ + 233, + 559, + 381, + 572 + ], + "type": "interline_equation", + "content": "\\Delta_ {U} ^ {i j} = \\operatorname {G r a m} \\left(\\Phi_ {i j} \\left(\\hat {Y} _ {S} ^ {U}\\right) - \\Phi_ {i j} \\left(\\hat {Y} _ {G} ^ {U}\\right)\\right).", + "image_path": "738e8fac144ad1dadacae035be0776919c70d93983933ef85b6bb19581171108.jpg" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 131, + 577, + 479, + 612 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 131, + 577, + 479, + 612 + ], + "spans": [ + { + "bbox": [ + 131, + 577, + 479, + 612 + ], + "type": "text", + "content": "Gram(\\cdot) refers to the Gram matrix, which calculates the correlations among vectorized feature maps along the channel dimension. The Gram matrix captures the statistics while collapsing the spatial layout [8]." + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 131, + 613, + 479, + 638 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 131, + 613, + 479, + 638 + ], + "spans": [ + { + "bbox": [ + 131, + 613, + 479, + 638 + ], + "type": "text", + "content": "For distillation, we ensure the consistency between inter-model distances by minimizing " + }, + { + "bbox": [ + 131, + 613, + 479, + 638 + ], + "type": "inline_equation", + "content": "R_{inter}" + }, + { + "bbox": [ + 131, + 613, + 479, + 638 + ], + "type": "text", + "content": ", the Frobenius norm between " + }, + { + "bbox": [ + 131, + 613, + 479, + 638 + ], + "type": "inline_equation", + "content": "\\Delta_U^{ij}" + }, + { + "bbox": [ + 131, + 613, + 479, + 638 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 131, + 613, + 479, + 638 + ], + "type": "inline_equation", + "content": "\\Delta_L^{ij}" + }, + { + "bbox": [ + 131, + 613, + 479, + 638 + ], + "type": "text", + "content": ":" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 248, + 643, + 479, + 667 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 248, + 643, + 479, + 667 + ], + "spans": [ + { + "bbox": [ + 248, + 643, + 479, + 667 + ], + "type": "interline_equation", + "content": "R _ {i n t e r} = \\sum_ {i, j} \\| \\Delta_ {U} ^ {i j} - \\Delta_ {L} ^ {i j} \\| _ {\\mathcal {F}}, \\tag {8}", + "image_path": "3ef3796152258b94eea8160f9a8e768ecbbb24fb218ef36ade56adc52e3cf8a8.jpg" + } + ] + } + ], + "index": 15 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 321, + 91, + 447, + 101 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 321, + 91, + 447, + 101 + ], + "spans": [ + { + "bbox": [ + 321, + 91, + 447, + 101 + ], + "type": "text", + "content": "Pairwise Distance Distillation" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 474, + 91, + 480, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 474, + 91, + 480, + 100 + ], + "spans": [ + { + "bbox": [ + 474, + 91, + 480, + 100 + ], + "type": "text", + "content": "7" + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 6 + }, + { + "para_blocks": [ + { + "bbox": [ + 130, + 116, + 482, + 142 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 116, + 482, + 142 + ], + "spans": [ + { + "bbox": [ + 130, + 116, + 482, + 142 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 130, + 116, + 482, + 142 + ], + "type": "inline_equation", + "content": "\\| \\cdot \\|_{\\mathcal{F}}" + }, + { + "bbox": [ + 130, + 116, + 482, + 142 + ], + "type": "text", + "content": " computes the Frobenius norm. The Frobenius norm of Gram Matrix difference is interpreted in [8, 16] as measuring distribution discrepancy." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 132, + 157, + 224, + 169 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 157, + 224, + 169 + ], + "spans": [ + { + "bbox": [ + 132, + 157, + 224, + 169 + ], + "type": "text", + "content": "3.4 Full Method" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 130, + 177, + 479, + 200 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 177, + 479, + 200 + ], + "spans": [ + { + "bbox": [ + 130, + 177, + 479, + 200 + ], + "type": "text", + "content": "The network " + }, + { + "bbox": [ + 130, + 177, + 479, + 200 + ], + "type": "inline_equation", + "content": "M_S" + }, + { + "bbox": [ + 130, + 177, + 479, + 200 + ], + "type": "text", + "content": " is fully optimized with supervised losses for labeled synthetic data " + }, + { + "bbox": [ + 130, + 177, + 479, + 200 + ], + "type": "inline_equation", + "content": "X^L" + }, + { + "bbox": [ + 130, + 177, + 479, + 200 + ], + "type": "text", + "content": " and unsupervised losses for real-world data " + }, + { + "bbox": [ + 130, + 177, + 479, + 200 + ], + "type": "inline_equation", + "content": "X^U" + }, + { + "bbox": [ + 130, + 177, + 479, + 200 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 131, + 201, + 482, + 262 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 131, + 201, + 482, + 262 + ], + "spans": [ + { + "bbox": [ + 131, + 201, + 482, + 262 + ], + "type": "text", + "content": "Supervised Losses maintain the specialization in synthetic degradation and reduce overfitting to the distillation. With the ground-truth image for prediction " + }, + { + "bbox": [ + 131, + 201, + 482, + 262 + ], + "type": "inline_equation", + "content": "\\hat{Y}_S^L" + }, + { + "bbox": [ + 131, + 201, + 482, + 262 + ], + "type": "text", + "content": ", denoted as " + }, + { + "bbox": [ + 131, + 201, + 482, + 262 + ], + "type": "inline_equation", + "content": "Y^{L}" + }, + { + "bbox": [ + 131, + 201, + 482, + 262 + ], + "type": "text", + "content": ", we implement a combination of supervised loss functions. It includes a wavelet-based loss " + }, + { + "bbox": [ + 131, + 201, + 482, + 262 + ], + "type": "inline_equation", + "content": "L_{\\mathrm{wv}}" + }, + { + "bbox": [ + 131, + 201, + 482, + 262 + ], + "type": "text", + "content": ", a perceptual loss " + }, + { + "bbox": [ + 131, + 201, + 482, + 262 + ], + "type": "inline_equation", + "content": "L_{\\mathrm{vgg}}" + }, + { + "bbox": [ + 131, + 201, + 482, + 262 + ], + "type": "text", + "content": ", and generative loss " + }, + { + "bbox": [ + 131, + 201, + 482, + 262 + ], + "type": "inline_equation", + "content": "L_{\\mathrm{gan}}" + }, + { + "bbox": [ + 131, + 201, + 482, + 262 + ], + "type": "text", + "content": " for adversarial training:" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 165, + 271, + 481, + 285 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 165, + 271, + 481, + 285 + ], + "spans": [ + { + "bbox": [ + 165, + 271, + 481, + 285 + ], + "type": "interline_equation", + "content": "\\mathcal {L} _ {L} \\left(\\hat {Y} _ {S} ^ {L}, Y ^ {L}\\right) = \\alpha_ {1} L _ {\\mathrm {w v}} \\left(\\hat {Y} _ {S} ^ {L}, Y ^ {L}\\right) + \\alpha_ {2} L _ {\\mathrm {v g g}} \\left(\\hat {Y} _ {S} ^ {L}, Y ^ {L}\\right) + \\alpha_ {3} L _ {\\mathrm {g a n}} \\left(\\hat {Y} _ {S} ^ {L}\\right), \\tag {9}", + "image_path": "8e955231fd41550f7bb5c2d57610d365a17b88990fa9473aa4066a556116cfbf.jpg" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 131, + 293, + 482, + 331 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 131, + 293, + 482, + 331 + ], + "spans": [ + { + "bbox": [ + 131, + 293, + 482, + 331 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 131, + 293, + 482, + 331 + ], + "type": "inline_equation", + "content": "L_{\\mathrm{vgg}}" + }, + { + "bbox": [ + 131, + 293, + 482, + 331 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 131, + 293, + 482, + 331 + ], + "type": "inline_equation", + "content": "L_{\\mathrm{gan}}" + }, + { + "bbox": [ + 131, + 293, + 482, + 331 + ], + "type": "text", + "content": " are adopted from RealESRGAN [35], with balancing weights " + }, + { + "bbox": [ + 131, + 293, + 482, + 331 + ], + "type": "inline_equation", + "content": "\\alpha_{1}" + }, + { + "bbox": [ + 131, + 293, + 482, + 331 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 131, + 293, + 482, + 331 + ], + "type": "inline_equation", + "content": "\\alpha_{2}" + }, + { + "bbox": [ + 131, + 293, + 482, + 331 + ], + "type": "text", + "content": ", and " + }, + { + "bbox": [ + 131, + 293, + 482, + 331 + ], + "type": "inline_equation", + "content": "\\alpha_{3}" + }, + { + "bbox": [ + 131, + 293, + 482, + 331 + ], + "type": "text", + "content": ". We use but omit writing the discriminator loss for brevity. " + }, + { + "bbox": [ + 131, + 293, + 482, + 331 + ], + "type": "inline_equation", + "content": "L_{\\mathrm{wv}}" + }, + { + "bbox": [ + 131, + 293, + 482, + 331 + ], + "type": "text", + "content": " refers to an " + }, + { + "bbox": [ + 131, + 293, + 482, + 331 + ], + "type": "inline_equation", + "content": "\\ell_{1}" + }, + { + "bbox": [ + 131, + 293, + 482, + 331 + ], + "type": "text", + "content": "-based wavelet loss [4, 49]:" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 208, + 339, + 481, + 362 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 208, + 339, + 481, + 362 + ], + "spans": [ + { + "bbox": [ + 208, + 339, + 481, + 362 + ], + "type": "interline_equation", + "content": "L _ {\\mathrm {w v}} \\left(\\hat {Y} _ {S} ^ {L}, Y ^ {L}\\right) = \\sum_ {i} \\omega_ {i} \\| W \\left(\\hat {Y} _ {S} ^ {L}\\right) _ {i} - W \\left(Y ^ {L}\\right) _ {i} \\| _ {1}, \\tag {10}", + "image_path": "4e9d932a82fe3f25446940f8a9f734247d9e86ddbec54cf4c1731bacfa9bc0a1.jpg" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 130, + 371, + 479, + 394 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 371, + 479, + 394 + ], + "spans": [ + { + "bbox": [ + 130, + 371, + 479, + 394 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 130, + 371, + 479, + 394 + ], + "type": "inline_equation", + "content": "W(\\cdot)_i" + }, + { + "bbox": [ + 130, + 371, + 479, + 394 + ], + "type": "text", + "content": " extracts the " + }, + { + "bbox": [ + 130, + 371, + 479, + 394 + ], + "type": "inline_equation", + "content": "i" + }, + { + "bbox": [ + 130, + 371, + 479, + 394 + ], + "type": "text", + "content": "-th wavelet channel, and " + }, + { + "bbox": [ + 130, + 371, + 479, + 394 + ], + "type": "inline_equation", + "content": "\\omega_i" + }, + { + "bbox": [ + 130, + 371, + 479, + 394 + ], + "type": "text", + "content": " weights its importance. For the wavelet transform, we apply the Haar wavelet [14] for all output channels." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 130, + 396, + 480, + 443 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 396, + 480, + 443 + ], + "spans": [ + { + "bbox": [ + 130, + 396, + 480, + 443 + ], + "type": "text", + "content": "Unsupervised Losses. In the absence of ground truth for " + }, + { + "bbox": [ + 130, + 396, + 480, + 443 + ], + "type": "inline_equation", + "content": "\\dot{Y}_S^U" + }, + { + "bbox": [ + 130, + 396, + 480, + 443 + ], + "type": "text", + "content": ", we opt to optimize for the consistency outlined in Eq. (6) and Eq. (8) as regularizers. To fully utilize the discriminator's knowledge of realness, we also incorporate the generative loss in Eq. (9):" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 216, + 452, + 481, + 466 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 216, + 452, + 481, + 466 + ], + "spans": [ + { + "bbox": [ + 216, + 452, + 481, + 466 + ], + "type": "interline_equation", + "content": "\\mathcal {L} _ {U} = \\lambda_ {1} R _ {\\text {i n t r a}} + \\lambda_ {2} R _ {\\text {i n t e r}} + \\lambda_ {3} L _ {\\text {g a n}} \\left(\\hat {Y} _ {S} ^ {U}\\right), \\tag {11}", + "image_path": "a4222e93c14de86d08b2b49ef5532dd5c37de9fa54463b9953ae7dc6f1983dc4.jpg" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 130, + 474, + 479, + 521 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 474, + 479, + 521 + ], + "spans": [ + { + "bbox": [ + 130, + 474, + 479, + 521 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 130, + 474, + 479, + 521 + ], + "type": "inline_equation", + "content": "\\lambda_1, \\lambda_2" + }, + { + "bbox": [ + 130, + 474, + 479, + 521 + ], + "type": "text", + "content": ", and " + }, + { + "bbox": [ + 130, + 474, + 479, + 521 + ], + "type": "inline_equation", + "content": "\\lambda_3" + }, + { + "bbox": [ + 130, + 474, + 479, + 521 + ], + "type": "text", + "content": " balance the optimization aims. " + }, + { + "bbox": [ + 130, + 474, + 479, + 521 + ], + "type": "inline_equation", + "content": "L_{gan}" + }, + { + "bbox": [ + 130, + 474, + 479, + 521 + ], + "type": "text", + "content": " here uses the same discriminator in Eq. (9). The supervised loss in Eq. (9) and the unsupervised distillation loss in Eq. (11) are combined as our final optimization objective, expressed as " + }, + { + "bbox": [ + 130, + 474, + 479, + 521 + ], + "type": "inline_equation", + "content": "\\mathcal{L} = \\mathcal{L}_L + \\mathcal{L}_U" + }, + { + "bbox": [ + 130, + 474, + 479, + 521 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 130, + 522, + 480, + 582 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 522, + 480, + 582 + ], + "spans": [ + { + "bbox": [ + 130, + 522, + 480, + 582 + ], + "type": "text", + "content": "Color Correction. In practice, our approach often leads to color shifts due to the regularization of distances within the feature space. To address this, we rectify the output by normalizing the mean and variance of each color channel with those of the corresponding input channels [35]. Further details are provided in Supplementary Sec. E." + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 131, + 582, + 481, + 641 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 131, + 582, + 481, + 641 + ], + "spans": [ + { + "bbox": [ + 131, + 582, + 481, + 641 + ], + "type": "text", + "content": "Learning Configurations. Our proposed PDD is a learning framework that can utilize pre-trained networks. It is recommended that " + }, + { + "bbox": [ + 131, + 582, + 481, + 641 + ], + "type": "inline_equation", + "content": "M_G" + }, + { + "bbox": [ + 131, + 582, + 481, + 641 + ], + "type": "text", + "content": " is pre-trained with extensive degradations " + }, + { + "bbox": [ + 131, + 582, + 481, + 641 + ], + "type": "inline_equation", + "content": "\\mathcal{D}_G" + }, + { + "bbox": [ + 131, + 582, + 481, + 641 + ], + "type": "text", + "content": ", as in [35,43]. The static configuration (refer to Fig. 3 (a)) initializes " + }, + { + "bbox": [ + 131, + 582, + 481, + 641 + ], + "type": "inline_equation", + "content": "M_S" + }, + { + "bbox": [ + 131, + 582, + 481, + 641 + ], + "type": "text", + "content": " with the model pre-trained on specific synthetic degradation " + }, + { + "bbox": [ + 131, + 582, + 481, + 641 + ], + "type": "inline_equation", + "content": "(\\mathcal{D}_S)" + }, + { + "bbox": [ + 131, + 582, + 481, + 641 + ], + "type": "text", + "content": ", and " + }, + { + "bbox": [ + 131, + 582, + 481, + 641 + ], + "type": "inline_equation", + "content": "M_G" + }, + { + "bbox": [ + 131, + 582, + 481, + 641 + ], + "type": "text", + "content": " is frozen throughout the training process." + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 131, + 641, + 481, + 666 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 131, + 641, + 481, + 666 + ], + "spans": [ + { + "bbox": [ + 131, + 641, + 481, + 666 + ], + "type": "text", + "content": "Alternative configuration initializes " + }, + { + "bbox": [ + 131, + 641, + 481, + 666 + ], + "type": "inline_equation", + "content": "M_G" + }, + { + "bbox": [ + 131, + 641, + 481, + 666 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 131, + 641, + 481, + 666 + ], + "type": "inline_equation", + "content": "M_S" + }, + { + "bbox": [ + 131, + 641, + 481, + 666 + ], + "type": "text", + "content": " with the same model pretrained with " + }, + { + "bbox": [ + 131, + 641, + 481, + 666 + ], + "type": "inline_equation", + "content": "\\mathcal{D}_G" + }, + { + "bbox": [ + 131, + 641, + 481, + 666 + ], + "type": "text", + "content": " and maintain " + }, + { + "bbox": [ + 131, + 641, + 481, + 666 + ], + "type": "inline_equation", + "content": "M_G" + }, + { + "bbox": [ + 131, + 641, + 481, + 666 + ], + "type": "text", + "content": " as an Exponential Moving Average (EMA)" + } + ] + } + ], + "index": 15 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "spans": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "type": "text", + "content": "8" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 166, + 91, + 229, + 102 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 166, + 91, + 229, + 102 + ], + "spans": [ + { + "bbox": [ + 166, + 91, + 229, + 102 + ], + "type": "text", + "content": "Y. Zhang et al." + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 7 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 212, + 115, + 306, + 190 + ], + "blocks": [ + { + "bbox": [ + 212, + 115, + 306, + 190 + ], + "lines": [ + { + "bbox": [ + 212, + 115, + 306, + 190 + ], + "spans": [ + { + "bbox": [ + 212, + 115, + 306, + 190 + ], + "type": "image", + "image_path": "9a06f91c09709db24ce244043fae065c542b7217eea9c97661dd4c811c64823c.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 238, + 192, + 281, + 200 + ], + "lines": [ + { + "bbox": [ + 238, + 192, + 281, + 200 + ], + "spans": [ + { + "bbox": [ + 238, + 192, + 281, + 200 + ], + "type": "text", + "content": "(a) Static Version" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_caption" + } + ], + "index": 2 + }, + { + "type": "image", + "bbox": [ + 317, + 114, + 404, + 190 + ], + "blocks": [ + { + "bbox": [ + 317, + 114, + 404, + 190 + ], + "lines": [ + { + "bbox": [ + 317, + 114, + 404, + 190 + ], + "spans": [ + { + "bbox": [ + 317, + 114, + 404, + 190 + ], + "type": "image", + "image_path": "42699b1275ee8fb378a964dd43c03b25f4764c12587b85096f877b1d62b73e1c.jpg" + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 340, + 193, + 382, + 200 + ], + "lines": [ + { + "bbox": [ + 340, + 193, + 382, + 200 + ], + "spans": [ + { + "bbox": [ + 340, + 193, + 382, + 200 + ], + "type": "text", + "content": "(b) EMA Version" + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 131, + 213, + 482, + 269 + ], + "lines": [ + { + "bbox": [ + 131, + 213, + 482, + 269 + ], + "spans": [ + { + "bbox": [ + 131, + 213, + 482, + 269 + ], + "type": "text", + "content": "Fig. 3: Two initialization options for the generalist and specialist models. (a) The static configuration initializes " + }, + { + "bbox": [ + 131, + 213, + 482, + 269 + ], + "type": "inline_equation", + "content": "M_G" + }, + { + "bbox": [ + 131, + 213, + 482, + 269 + ], + "type": "text", + "content": " with a model pre-trained by the complex synthetic pipeline and " + }, + { + "bbox": [ + 131, + 213, + 482, + 269 + ], + "type": "inline_equation", + "content": "M_S" + }, + { + "bbox": [ + 131, + 213, + 482, + 269 + ], + "type": "text", + "content": " with a model pre-trained by simple degradation in " + }, + { + "bbox": [ + 131, + 213, + 482, + 269 + ], + "type": "inline_equation", + "content": "X^L" + }, + { + "bbox": [ + 131, + 213, + 482, + 269 + ], + "type": "text", + "content": ". Weights of " + }, + { + "bbox": [ + 131, + 213, + 482, + 269 + ], + "type": "inline_equation", + "content": "M_G" + }, + { + "bbox": [ + 131, + 213, + 482, + 269 + ], + "type": "text", + "content": " are frozen during distillation. (b) Both " + }, + { + "bbox": [ + 131, + 213, + 482, + 269 + ], + "type": "inline_equation", + "content": "M_S" + }, + { + "bbox": [ + 131, + 213, + 482, + 269 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 131, + 213, + 482, + 269 + ], + "type": "inline_equation", + "content": "M_G" + }, + { + "bbox": [ + 131, + 213, + 482, + 269 + ], + "type": "text", + "content": " are initialized with a pre-trained generalized model. Weights of " + }, + { + "bbox": [ + 131, + 213, + 482, + 269 + ], + "type": "inline_equation", + "content": "M_G" + }, + { + "bbox": [ + 131, + 213, + 482, + 269 + ], + "type": "text", + "content": " are the EMA version of " + }, + { + "bbox": [ + 131, + 213, + 482, + 269 + ], + "type": "inline_equation", + "content": "M_S" + }, + { + "bbox": [ + 131, + 213, + 482, + 269 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 6, + "angle": 0, + "type": "image_caption" + } + ], + "index": 4 + }, + { + "bbox": [ + 131, + 291, + 482, + 352 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 131, + 291, + 482, + 352 + ], + "spans": [ + { + "bbox": [ + 131, + 291, + 482, + 352 + ], + "type": "text", + "content": "version of " + }, + { + "bbox": [ + 131, + 291, + 482, + 352 + ], + "type": "inline_equation", + "content": "M_S" + }, + { + "bbox": [ + 131, + 291, + 482, + 352 + ], + "type": "text", + "content": " (see Fig. 3 (b)). Despite " + }, + { + "bbox": [ + 131, + 291, + 482, + 352 + ], + "type": "inline_equation", + "content": "M_S" + }, + { + "bbox": [ + 131, + 291, + 482, + 352 + ], + "type": "text", + "content": " not being initialized as a specialist, it can readily specialize to the synthetic degradation " + }, + { + "bbox": [ + 131, + 291, + 482, + 352 + ], + "type": "inline_equation", + "content": "\\mathcal{D}_S" + }, + { + "bbox": [ + 131, + 291, + 482, + 352 + ], + "type": "text", + "content": " through supervised learning in Eq. (9). Empirically, the EMA version tends to be more effective, likely because the updated knowledge of real-world data in " + }, + { + "bbox": [ + 131, + 291, + 482, + 352 + ], + "type": "inline_equation", + "content": "M_S" + }, + { + "bbox": [ + 131, + 291, + 482, + 352 + ], + "type": "text", + "content": " also benefits " + }, + { + "bbox": [ + 131, + 291, + 482, + 352 + ], + "type": "inline_equation", + "content": "M_G" + }, + { + "bbox": [ + 131, + 291, + 482, + 352 + ], + "type": "text", + "content": " through EMA, enabling " + }, + { + "bbox": [ + 131, + 291, + 482, + 352 + ], + "type": "inline_equation", + "content": "M_G" + }, + { + "bbox": [ + 131, + 291, + 482, + 352 + ], + "type": "text", + "content": " to yield better references for optimization." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 132, + 369, + 230, + 383 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 369, + 230, + 383 + ], + "spans": [ + { + "bbox": [ + 132, + 369, + 230, + 383 + ], + "type": "text", + "content": "4 Experiments" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 132, + 394, + 202, + 407 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 394, + 202, + 407 + ], + "spans": [ + { + "bbox": [ + 132, + 394, + 202, + 407 + ], + "type": "text", + "content": "4.1 Settings" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 131, + 414, + 482, + 533 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 131, + 414, + 482, + 533 + ], + "spans": [ + { + "bbox": [ + 131, + 414, + 482, + 533 + ], + "type": "text", + "content": "Datasets. We experiment on real-world datasets, RealSR [2] and DRealSR [39], and a synthetic dataset without released degradation settings, NTIRE20 [23]. For experiments on RealSR and DRealSR, we synthesize labeled training data " + }, + { + "bbox": [ + 131, + 414, + 482, + 533 + ], + "type": "inline_equation", + "content": "\\{X^L\\}" + }, + { + "bbox": [ + 131, + 414, + 482, + 533 + ], + "type": "text", + "content": " from high-resolution images in the DF2K dataset [1,32]. For the NTIRE20 dataset, we use the provided target clean images to synthesize low-resolution counterparts. For synthesizing " + }, + { + "bbox": [ + 131, + 414, + 482, + 533 + ], + "type": "inline_equation", + "content": "\\{X^L\\}" + }, + { + "bbox": [ + 131, + 414, + 482, + 533 + ], + "type": "text", + "content": ", we use only bicubic interpolation as the default and investigate other options for " + }, + { + "bbox": [ + 131, + 414, + 482, + 533 + ], + "type": "inline_equation", + "content": "\\mathcal{D}_S" + }, + { + "bbox": [ + 131, + 414, + 482, + 533 + ], + "type": "text", + "content": " in the Supplementary Sec. D.b. RealSR or DRealSR has multiple data sources, and we optimize our methods for each source separately. For DRealSR, we report results for Panasonic and Olympus sources, as the others do not have sufficient training samples." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 131, + 534, + 482, + 616 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 131, + 534, + 482, + 616 + ], + "spans": [ + { + "bbox": [ + 131, + 534, + 482, + 616 + ], + "type": "text", + "content": "Training Details. We crop patches of size " + }, + { + "bbox": [ + 131, + 534, + 482, + 616 + ], + "type": "inline_equation", + "content": "48 \\times 48" + }, + { + "bbox": [ + 131, + 534, + 482, + 616 + ], + "type": "text", + "content": " as inputs and train the model with the ADAM optimizer [12] using the settings " + }, + { + "bbox": [ + 131, + 534, + 482, + 616 + ], + "type": "inline_equation", + "content": "\\beta_{1} = 0.9" + }, + { + "bbox": [ + 131, + 534, + 482, + 616 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 131, + 534, + 482, + 616 + ], + "type": "inline_equation", + "content": "\\beta_{2} = 0.999" + }, + { + "bbox": [ + 131, + 534, + 482, + 616 + ], + "type": "text", + "content": ", and " + }, + { + "bbox": [ + 131, + 534, + 482, + 616 + ], + "type": "inline_equation", + "content": "\\epsilon = 10^{-8}" + }, + { + "bbox": [ + 131, + 534, + 482, + 616 + ], + "type": "text", + "content": ". The batch size is 16, including equal numbers of labeled and unlabeled inputs. The initial learning rate is " + }, + { + "bbox": [ + 131, + 534, + 482, + 616 + ], + "type": "inline_equation", + "content": "1e - 4" + }, + { + "bbox": [ + 131, + 534, + 482, + 616 + ], + "type": "text", + "content": " and halved after " + }, + { + "bbox": [ + 131, + 534, + 482, + 616 + ], + "type": "inline_equation", + "content": "25\\mathrm{K}" + }, + { + "bbox": [ + 131, + 534, + 482, + 616 + ], + "type": "text", + "content": " iterations. The static version needs " + }, + { + "bbox": [ + 131, + 534, + 482, + 616 + ], + "type": "inline_equation", + "content": "50\\mathrm{K}" + }, + { + "bbox": [ + 131, + 534, + 482, + 616 + ], + "type": "text", + "content": " iterations, and the EMA version is trained for " + }, + { + "bbox": [ + 131, + 534, + 482, + 616 + ], + "type": "inline_equation", + "content": "100\\mathrm{K}" + }, + { + "bbox": [ + 131, + 534, + 482, + 616 + ], + "type": "text", + "content": " iterations due to the initial specialist underfitting to the synthetic domain. All models are for " + }, + { + "bbox": [ + 131, + 534, + 482, + 616 + ], + "type": "inline_equation", + "content": "\\times 4" + }, + { + "bbox": [ + 131, + 534, + 482, + 616 + ], + "type": "text", + "content": " super-resolution. Other details are in Supplementary Sec. A." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 131, + 617, + 482, + 665 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 131, + 617, + 482, + 665 + ], + "spans": [ + { + "bbox": [ + 131, + 617, + 482, + 665 + ], + "type": "text", + "content": "Evaluation. We evaluate models on testing sets of RealSR, DRealSR, and the validation set of NTIRE20. We crop non-overlapped 512x512 patches from LR images in DRealSR to avoid memory exhaustion. For other datasets, we use the full images. To fully use the provided high-resolution counterparts, we report" + } + ] + } + ], + "index": 12 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 321, + 91, + 447, + 101 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 321, + 91, + 447, + 101 + ], + "spans": [ + { + "bbox": [ + 321, + 91, + 447, + 101 + ], + "type": "text", + "content": "Pairwise Distance Distillation" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 474, + 91, + 481, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 474, + 91, + 481, + 100 + ], + "spans": [ + { + "bbox": [ + 474, + 91, + 481, + 100 + ], + "type": "text", + "content": "9" + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 8 + }, + { + "para_blocks": [ + { + "type": "table", + "bbox": [ + 160, + 168, + 454, + 200 + ], + "blocks": [ + { + "bbox": [ + 130, + 114, + 482, + 160 + ], + "lines": [ + { + "bbox": [ + 130, + 114, + 482, + 160 + ], + "spans": [ + { + "bbox": [ + 130, + 114, + 482, + 160 + ], + "type": "text", + "content": "Table 1: " + }, + { + "bbox": [ + 130, + 114, + 482, + 160 + ], + "type": "inline_equation", + "content": "M_G" + }, + { + "bbox": [ + 130, + 114, + 482, + 160 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 130, + 114, + 482, + 160 + ], + "type": "inline_equation", + "content": "M_S" + }, + { + "bbox": [ + 130, + 114, + 482, + 160 + ], + "type": "text", + "content": " (static) columns show initial performances on the bicubic-interpolated data " + }, + { + "bbox": [ + 130, + 114, + 482, + 160 + ], + "type": "inline_equation", + "content": "\\{X^L\\}" + }, + { + "bbox": [ + 130, + 114, + 482, + 160 + ], + "type": "text", + "content": " from Set14 [42] and " + }, + { + "bbox": [ + 130, + 114, + 482, + 160 + ], + "type": "inline_equation", + "content": "\\{X^U\\}" + }, + { + "bbox": [ + 130, + 114, + 482, + 160 + ], + "type": "text", + "content": " with unknown degradation from NTIRE20 [23]. " + }, + { + "bbox": [ + 130, + 114, + 482, + 160 + ], + "type": "inline_equation", + "content": "M_S" + }, + { + "bbox": [ + 130, + 114, + 482, + 160 + ], + "type": "text", + "content": " (EMA) column shows the performance of " + }, + { + "bbox": [ + 130, + 114, + 482, + 160 + ], + "type": "inline_equation", + "content": "M_S" + }, + { + "bbox": [ + 130, + 114, + 482, + 160 + ], + "type": "text", + "content": " in the EMA version on " + }, + { + "bbox": [ + 130, + 114, + 482, + 160 + ], + "type": "inline_equation", + "content": "\\{X^L\\}" + }, + { + "bbox": [ + 130, + 114, + 482, + 160 + ], + "type": "text", + "content": " after training process." + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 160, + 168, + 454, + 200 + ], + "lines": [ + { + "bbox": [ + 160, + 168, + 454, + 200 + ], + "spans": [ + { + "bbox": [ + 160, + 168, + 454, + 200 + ], + "type": "table", + "html": "
DomainMetricMGMS(static)MS(EMA)DomainMetricMGMS(static)
{X^L}PSNR↑23.9824.8126.26{X^U}PSNR↑25.0819.82
LPIPS↓0.23490.13370.1471LPIPS↓0.25040.7552
", + "image_path": "c10b71c38239897b3a46ceae9f83c3798ffbf9f5589b6338c7d3063808597a12.jpg" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "table_body" + } + ], + "index": 3 + }, + { + "type": "image", + "bbox": [ + 139, + 211, + 248, + 294 + ], + "blocks": [ + { + "bbox": [ + 139, + 211, + 248, + 294 + ], + "lines": [ + { + "bbox": [ + 139, + 211, + 248, + 294 + ], + "spans": [ + { + "bbox": [ + 139, + 211, + 248, + 294 + ], + "type": "image", + "image_path": "adac3a713f37c17d3700c1664a4f5998e292bc09811b0b9568eb473d7e07adaa.jpg" + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 130, + 303, + 482, + 359 + ], + "lines": [ + { + "bbox": [ + 130, + 303, + 482, + 359 + ], + "spans": [ + { + "bbox": [ + 130, + 303, + 482, + 359 + ], + "type": "text", + "content": "Fig. 4: Improvements over the Generalist on five unlabeled data domains, where a lower LPIPS score is better. ND improves fidelity scores (PSNR) but dramatically drops perceptual scores (LPIPS and NRQM). Both versions of our method achieve better improvements than ND for all reported metrics. For each domain, the Static version improves at least one of the perceptual metrics; the EMA version improves all." + } + ] + } + ], + "index": 7, + "angle": 0, + "type": "image_caption" + } + ], + "index": 4 + }, + { + "type": "image", + "bbox": [ + 249, + 213, + 364, + 295 + ], + "blocks": [ + { + "bbox": [ + 249, + 213, + 364, + 295 + ], + "lines": [ + { + "bbox": [ + 249, + 213, + 364, + 295 + ], + "spans": [ + { + "bbox": [ + 249, + 213, + 364, + 295 + ], + "type": "image", + "image_path": "e5caf7f5458acc358aa53c142ed322b3cd65192f3994ba1b8ff29c2aa6b504dc.jpg" + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "image_body" + } + ], + "index": 5 + }, + { + "type": "image", + "bbox": [ + 367, + 212, + 480, + 294 + ], + "blocks": [ + { + "bbox": [ + 367, + 212, + 480, + 294 + ], + "lines": [ + { + "bbox": [ + 367, + 212, + 480, + 294 + ], + "spans": [ + { + "bbox": [ + 367, + 212, + 480, + 294 + ], + "type": "image", + "image_path": "ce96daf2ed4534ed3889474a07f1593f55cf0449e01b92eddd8b124e53975af4.jpg" + } + ] + } + ], + "index": 6, + "angle": 0, + "type": "image_body" + } + ], + "index": 6 + }, + { + "bbox": [ + 130, + 383, + 482, + 480 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 383, + 482, + 480 + ], + "spans": [ + { + "bbox": [ + 130, + 383, + 482, + 480 + ], + "type": "text", + "content": "scores of full-reference metrics, PSNR, SSIM, and LPIPS [44]. The PSNR and SSIM scores are calculated on the Y-channel of YCbCr format images. Additionally, we follow BSRGAN [43] and supplement with the no-reference metric NRQM [24] to evaluate the perceptual clearness, as the ground truth in real-world captured datasets [2,39] also feature some blur. We use the implementation in IQA-PyTorch3 for all metrics. It is important to consider both fidelity and perceptual scores; fidelity metrics are worse at detecting blurriness while perceptual metrics are not sensitive to high-frequency artifacts [25, 26, 29, 48]." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 131, + 500, + 268, + 511 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 131, + 500, + 268, + 511 + ], + "spans": [ + { + "bbox": [ + 131, + 500, + 268, + 511 + ], + "type": "text", + "content": "4.2 Effectiveness of PDD" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 130, + 523, + 482, + 644 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 523, + 482, + 644 + ], + "spans": [ + { + "bbox": [ + 130, + 523, + 482, + 644 + ], + "type": "text", + "content": "Comparisons. We compare the two versions of our method in Fig. 3 with the naive distillation defined in Eq. (3) regarding their improvements over the generalist network " + }, + { + "bbox": [ + 130, + 523, + 482, + 644 + ], + "type": "inline_equation", + "content": "M_G" + }, + { + "bbox": [ + 130, + 523, + 482, + 644 + ], + "type": "text", + "content": ". We choose RealESRGAN [35] as the pre-trained generalist model and ESRGAN [36], a standard SISR model, as the pre-trained specialist for the naive distillation and the static version in Fig. 3 (a). Tab. 1 shows the initial performances of the generalist and specialist for " + }, + { + "bbox": [ + 130, + 523, + 482, + 644 + ], + "type": "inline_equation", + "content": "\\{X^L\\}" + }, + { + "bbox": [ + 130, + 523, + 482, + 644 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 130, + 523, + 482, + 644 + ], + "type": "inline_equation", + "content": "\\{X^U\\}" + }, + { + "bbox": [ + 130, + 523, + 482, + 644 + ], + "type": "text", + "content": ". As we assumed in Sec. 3.1, the specialist performs better than the generalist for " + }, + { + "bbox": [ + 130, + 523, + 482, + 644 + ], + "type": "inline_equation", + "content": "\\{X^L\\}" + }, + { + "bbox": [ + 130, + 523, + 482, + 644 + ], + "type": "text", + "content": " while " + }, + { + "bbox": [ + 130, + 523, + 482, + 644 + ], + "type": "inline_equation", + "content": "M_G" + }, + { + "bbox": [ + 130, + 523, + 482, + 644 + ], + "type": "text", + "content": " generalizes better for " + }, + { + "bbox": [ + 130, + 523, + 482, + 644 + ], + "type": "inline_equation", + "content": "\\{X^U\\}" + }, + { + "bbox": [ + 130, + 523, + 482, + 644 + ], + "type": "text", + "content": ". For the EMA version in Fig. 3 (b), " + }, + { + "bbox": [ + 130, + 523, + 482, + 644 + ], + "type": "inline_equation", + "content": "M_S" + }, + { + "bbox": [ + 130, + 523, + 482, + 644 + ], + "type": "text", + "content": " as well as " + }, + { + "bbox": [ + 130, + 523, + 482, + 644 + ], + "type": "inline_equation", + "content": "M_G" + }, + { + "bbox": [ + 130, + 523, + 482, + 644 + ], + "type": "text", + "content": " is initialized with RealESRGAN. Through the supervised learning in Eq. (9), " + }, + { + "bbox": [ + 130, + 523, + 482, + 644 + ], + "type": "inline_equation", + "content": "M_S(\\mathrm{EMA})" + }, + { + "bbox": [ + 130, + 523, + 482, + 644 + ], + "type": "text", + "content": " shows specialized performance for labeled data " + }, + { + "bbox": [ + 130, + 523, + 482, + 644 + ], + "type": "inline_equation", + "content": "\\{X^L\\}" + }, + { + "bbox": [ + 130, + 523, + 482, + 644 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 10 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "spans": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "type": "text", + "content": "10" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 166, + 91, + 229, + 102 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 166, + 91, + 229, + 102 + ], + "spans": [ + { + "bbox": [ + 166, + 91, + 229, + 102 + ], + "type": "text", + "content": "Y. Zhang et al." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 133, + 653, + 326, + 666 + ], + "type": "page_footnote", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 653, + 326, + 666 + ], + "spans": [ + { + "bbox": [ + 133, + 653, + 326, + 666 + ], + "type": "text", + "content": "3 https://github.com/chaofengc/IQA-PyTorch" + } + ] + } + ], + "index": 11 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 9 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 161, + 115, + 254, + 183 + ], + "blocks": [ + { + "bbox": [ + 161, + 115, + 254, + 183 + ], + "lines": [ + { + "bbox": [ + 161, + 115, + 254, + 183 + ], + "spans": [ + { + "bbox": [ + 161, + 115, + 254, + 183 + ], + "type": "image", + "image_path": "b9b0e83fd39785c5a23976a9f9c8b8420f148cee9965642b117a07c7173992fe.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 181, + 186, + 233, + 196 + ], + "lines": [ + { + "bbox": [ + 181, + 186, + 233, + 196 + ], + "spans": [ + { + "bbox": [ + 181, + 186, + 233, + 196 + ], + "type": "text", + "content": "(a) Generalist" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_caption" + } + ], + "index": 2 + }, + { + "type": "image", + "bbox": [ + 261, + 115, + 354, + 183 + ], + "blocks": [ + { + "bbox": [ + 261, + 115, + 354, + 183 + ], + "lines": [ + { + "bbox": [ + 261, + 115, + 354, + 183 + ], + "spans": [ + { + "bbox": [ + 261, + 115, + 354, + 183 + ], + "type": "image", + "image_path": "13111ea952a2ff2f9334a41a885eb6adab64666cff078714724ab6b817f11af1.jpg" + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 268, + 186, + 346, + 196 + ], + "lines": [ + { + "bbox": [ + 268, + 186, + 346, + 196 + ], + "spans": [ + { + "bbox": [ + 268, + 186, + 346, + 196 + ], + "type": "text", + "content": "(b) Bicubic Specialist" + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 130, + 205, + 482, + 261 + ], + "lines": [ + { + "bbox": [ + 130, + 205, + 482, + 261 + ], + "spans": [ + { + "bbox": [ + 130, + 205, + 482, + 261 + ], + "type": "text", + "content": "Fig. 5: Visualization of low-level features for predictions of DIV2K [1] (bicubic) and NTIRE20 (unknown) following [20] (a) Generalist's predictions for the two domains has overlapped distribution. (b) The predictions from the specialist that only adept in bicubic interpolation have separated distributions. (c) After applying our method (static version), predictions for two domains are pushed close." + } + ] + } + ], + "index": 8, + "angle": 0, + "type": "image_caption" + } + ], + "index": 4 + }, + { + "type": "image", + "bbox": [ + 361, + 115, + 454, + 183 + ], + "blocks": [ + { + "bbox": [ + 361, + 115, + 454, + 183 + ], + "lines": [ + { + "bbox": [ + 361, + 115, + 454, + 183 + ], + "spans": [ + { + "bbox": [ + 361, + 115, + 454, + 183 + ], + "type": "image", + "image_path": "0dbef2d3aa8ff9f165ed726739e40ab8bc9098996d292124aeefc0f65688c46b.jpg" + } + ] + } + ], + "index": 6, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 366, + 186, + 447, + 196 + ], + "lines": [ + { + "bbox": [ + 366, + 186, + 447, + 196 + ], + "spans": [ + { + "bbox": [ + 366, + 186, + 447, + 196 + ], + "type": "text", + "content": "(c) Adapted Specialist" + } + ] + } + ], + "index": 7, + "angle": 0, + "type": "image_caption" + } + ], + "index": 6 + }, + { + "bbox": [ + 130, + 283, + 482, + 366 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 283, + 482, + 366 + ], + "spans": [ + { + "bbox": [ + 130, + 283, + 482, + 366 + ], + "type": "text", + "content": "Fig. 4 plots the improvements over the pretrained generalist for Naive Distillation (ND), Static, and EMA versions of our method. Compared to the generalist, ND dramatically drops in perceptual scores (LPIPS, NRQM) despite maintaining or improving the PSNR. Both versions of our method perform better than ND; compared to the generalist, our static version improves PSNR significantly and at least one perceptual score for each data domain. The EMA version improves all metrics with the highest magnitudes." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 130, + 367, + 482, + 486 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 367, + 482, + 486 + ], + "spans": [ + { + "bbox": [ + 130, + 367, + 482, + 486 + ], + "type": "text", + "content": "Change in Low-Level Characteristics. We extract low-level features for visualization following [20]. Fig. 5b visualize predictions generated by the bicubic specialist (ESRGAN), which has separate distributions. Fig. 5c verifies the predictions from the specialist after applying our method (static version), where the separated distributions are pushed closer. Visualizing low-level differences between images with arbitrary content is challenging [20]. As such, we also quantitatively compare the KL divergence between the low-level features. As shown in Fig. 6, applying our method decreases the KL divergence between labeled (DIV2K-bicubic) and unlabeled predictions. Details of the visualization and KL-diversity calculations are in Supplementary Sec. B." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 131, + 502, + 332, + 514 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 131, + 502, + 332, + 514 + ], + "spans": [ + { + "bbox": [ + 131, + 502, + 332, + 514 + ], + "type": "text", + "content": "4.3 Comparisons with Other Methods" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 130, + 521, + 482, + 605 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 521, + 482, + 605 + ], + "spans": [ + { + "bbox": [ + 130, + 521, + 482, + 605 + ], + "type": "text", + "content": "Quantitative Comparison. We consider two variants of our method by applying the EMA version on RealESRGAN [35] and BSRGAN [43]. In Tab. 2, we compare their performances with other state-of-the-art RWSR methods, including FeMaSR [3], EDAN [17], DASR [40], DAN [11], RealESRGAN [35], BSRGAN [43], and TG [46], which released code and model weights that enable reproduction. We regenerate results to evaluate these methods, while Supplementary Sec. C discusses related but not directly comparable methods." + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 130, + 605, + 482, + 665 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 605, + 482, + 665 + ], + "spans": [ + { + "bbox": [ + 130, + 605, + 482, + 665 + ], + "type": "text", + "content": "As shown in Tab. 2, EDAN, DAN, and DASR perform strongly on the fidelity metrics (PSNR, SSIM) but are poor perceptually, especially on the NRQM score. The qualitative comparison highlights the blurry effects. Our method enhances RealESRGAN and BSRGAN across all metrics and demonstrates superior performance compared to TG when both are applied to RealESRGAN. For FeMaSR," + } + ] + } + ], + "index": 13 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 321, + 91, + 447, + 101 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 321, + 91, + 447, + 101 + ], + "spans": [ + { + "bbox": [ + 321, + 91, + 447, + 101 + ], + "type": "text", + "content": "Pairwise Distance Distillation" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 470, + 91, + 479, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 470, + 91, + 479, + 100 + ], + "spans": [ + { + "bbox": [ + 470, + 91, + 479, + 100 + ], + "type": "text", + "content": "11" + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 10 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 161, + 121, + 255, + 198 + ], + "blocks": [ + { + "bbox": [ + 161, + 121, + 255, + 198 + ], + "lines": [ + { + "bbox": [ + 161, + 121, + 255, + 198 + ], + "spans": [ + { + "bbox": [ + 161, + 121, + 255, + 198 + ], + "type": "image", + "image_path": "fc424c549e010f2a02e0bd6b0132326a4670d21da322796fc5bcbb28db78de02.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 131, + 209, + 291, + 276 + ], + "lines": [ + { + "bbox": [ + 131, + 209, + 291, + 276 + ], + "spans": [ + { + "bbox": [ + 131, + 209, + 291, + 276 + ], + "type": "text", + "content": "Fig. 6: KL divergence between specialist's unlabeled and labeled predictions. The KL divergence between labeled and unlabeled predictions is lower after applying our method (red) than that in the specialist before adaption (blue)." + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_caption" + } + ], + "index": 2 + }, + { + "type": "image", + "bbox": [ + 310, + 116, + 482, + 180 + ], + "blocks": [ + { + "bbox": [ + 310, + 116, + 482, + 180 + ], + "lines": [ + { + "bbox": [ + 310, + 116, + 482, + 180 + ], + "spans": [ + { + "bbox": [ + 310, + 116, + 482, + 180 + ], + "type": "image", + "image_path": "999982b588e56a6e40aa14c057731539c91f904e67ea2cf4826f3b6c3b7788c0.jpg" + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 304, + 190, + 482, + 278 + ], + "lines": [ + { + "bbox": [ + 304, + 190, + 482, + 278 + ], + "spans": [ + { + "bbox": [ + 304, + 190, + 482, + 278 + ], + "type": "text", + "content": "Fig.7: User study on outputs of 30 real-world images. We recruit 30 evaluators for each image. Each vote is for a better quality within a pair of options. We compared our method to RealESRGAN and BSRGAN by using them as pre-trained models; FeMaSR is compared to RealESRGAN+ours. Our method gains at least " + }, + { + "bbox": [ + 304, + 190, + 482, + 278 + ], + "type": "inline_equation", + "content": "65\\%" + }, + { + "bbox": [ + 304, + 190, + 482, + 278 + ], + "type": "text", + "content": " votes." + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "image_caption" + } + ], + "index": 4 + }, + { + "type": "table", + "bbox": [ + 135, + 365, + 482, + 544 + ], + "blocks": [ + { + "bbox": [ + 131, + 299, + 482, + 354 + ], + "lines": [ + { + "bbox": [ + 131, + 299, + 482, + 354 + ], + "spans": [ + { + "bbox": [ + 131, + 299, + 482, + 354 + ], + "type": "text", + "content": "Table 2: Quantitative comparison with state-of-the-art methods. DAN, EDAN, and DASR are separately listed as they favor fidelity metrics but yield poor perceptual scores. For other methods, **bold** and **underline** mark the best and second-best scores. Subscripts in the last two columns show their difference from the corresponding pretrained generalist, where improvement is colored red and drop is blue." + } + ] + } + ], + "index": 6, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 135, + 365, + 482, + 544 + ], + "lines": [ + { + "bbox": [ + 135, + 365, + 482, + 544 + ], + "spans": [ + { + "bbox": [ + 135, + 365, + 482, + 544 + ], + "type": "table", + "html": "
DatasetMetricDAN [11]EDAN [17]DASR [40]FeMaSR [3]RealESRGAN [43]BSRGAN [35]RealESRGAN +TG [46]RealESRGAN +OursBSRGAN +Ours
NTIRE20 -ValidPSNR↑26.8326.59-23.5425.0825.4425.5626.42(+1.340)26.37(+0.930)
SSIM↑0.71710.7400-0.66650.70610.69840.71930.7297(+0.024)0.7238(+0.025)
LPIPS↓0.57470.2678-0.23600.25040.26450.24200.2475(-0.003)0.2595(-0.005)
NRQM↑4.87845.8481-6.62456.12136.27795.79266.2263(+0.105)6.2781(+0.0002)
RealSR -CanonPSNR↑26.6726.3626.7124.2924.7425.5725.0325.94(+1.200)26.13(+0.560)
SSIM↑0.77400.77740.77820.74600.76340.76830.77200.7733(+0.010)0.7626(-0.006)
LPIPS↓0.40950.30260.25070.28090.26070.25730.26150.2516(-0.009)0.2517(-0.006)
NRQM↑3.04073.63603.93836.01516.06496.02935.42506.0647(-0.0002)6.1323(+0.103)
RealSR -NikonPSNR↑26.2925.3725.6323.9424.3124.7924.5625.28(+0.970)25.07(+0.280)
SSIM↑0.75330.74600.74730.70970.74060.73340.74060.7470(+0.014)0.7294(-0.001)
LPIPS↓0.41330.32000.27850.30460.28510.27970.29490.2802(-0.005)0.2713(-0.008)
NRQM↑3.20564.48724.49525.94885.66855.93875.19425.9044(+0.236)6.1276(+0.189)
DRealSR -PanasonicPSNR↑28.9928.12-25.0227.0326.9327.5127.87(+0.840)27.69(+0.760)
SSIM↑0.81230.8006-0.71180.77820.76210.78430.7939(+0.016)0.7752(+0.013)
LPIPS↓0.42350.2747-0.30880.27030.28200.28360.2682(-0.002)0.2687(-0.013)
NRQM↑2.81314.5915-6.03425.37515.58654.64835.3540(-0.021)5.5608(-0.026)
DRealSR -OlympusPSNR↑28.7427.85-24.3826.7326.4927.4327.40(+0.670)27.25(+0.764)
SSIM↑0.80630.7979-0.66910.76670.75680.78430.7846(+0.018)0.7635(+0.007)
LPIPS↓0.45950.3555-0.40070.31590.32830.34160.3139(-0.002)0.3268(-0.001)
NRQM↑2.79613.7989-6.25905.25305.53764.14595.3722(+0.120)5.5957(+0.058)
", + "image_path": "e0b1ae9458556336027dc9be96d181e5769ffec273c81d57a06512b79ace2fd4.jpg" + } + ] + } + ], + "index": 7, + "angle": 0, + "type": "table_body" + } + ], + "index": 7 + }, + { + "bbox": [ + 130, + 577, + 480, + 601 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 577, + 480, + 601 + ], + "spans": [ + { + "bbox": [ + 130, + 577, + 480, + 601 + ], + "type": "text", + "content": "we achieve competitive perceptual scores and significantly surpass their fidelity scores." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 130, + 605, + 482, + 666 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 605, + 482, + 666 + ], + "spans": [ + { + "bbox": [ + 130, + 605, + 482, + 666 + ], + "type": "text", + "content": "User Study is conducted by presenting two reconstructions of the same LR image and asking for the better one. We compare the EMA version of our method with methods RealESRGAN, BSRGAN, and FeMaSR. Each comparison is conducted on 30 different real-world results among 30 evaluators. As shown in Fig. 7, at least " + }, + { + "bbox": [ + 130, + 605, + 482, + 666 + ], + "type": "inline_equation", + "content": "65\\%" + }, + { + "bbox": [ + 130, + 605, + 482, + 666 + ], + "type": "text", + "content": " evaluators vote for our method in each comparison. Although quan-" + } + ] + } + ], + "index": 9 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "spans": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "type": "text", + "content": "12" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 166, + 91, + 229, + 102 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 166, + 91, + 229, + 102 + ], + "spans": [ + { + "bbox": [ + 166, + 91, + 229, + 102 + ], + "type": "text", + "content": "Y. Zhang et al." + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 11 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 135, + 114, + 188, + 300 + ], + "blocks": [ + { + "bbox": [ + 135, + 114, + 188, + 300 + ], + "lines": [ + { + "bbox": [ + 135, + 114, + 188, + 300 + ], + "spans": [ + { + "bbox": [ + 135, + 114, + 188, + 300 + ], + "type": "image", + "image_path": "08c40686f7cdf46f11179acf5ee3b4d46bf665dcd2a82aa0347d71b3d2327def.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 149, + 302, + 175, + 311 + ], + "lines": [ + { + "bbox": [ + 149, + 302, + 175, + 311 + ], + "spans": [ + { + "bbox": [ + 149, + 302, + 175, + 311 + ], + "type": "text", + "content": "(a) LR" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_caption" + } + ], + "index": 2 + }, + { + "type": "image", + "bbox": [ + 193, + 114, + 246, + 300 + ], + "blocks": [ + { + "bbox": [ + 193, + 114, + 246, + 300 + ], + "lines": [ + { + "bbox": [ + 193, + 114, + 246, + 300 + ], + "spans": [ + { + "bbox": [ + 193, + 114, + 246, + 300 + ], + "type": "image", + "image_path": "fe4126d5806061733da59ca32caea78ffea8581cc5a5d75aadc5a60df29e959a.jpg" + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 197, + 302, + 242, + 312 + ], + "lines": [ + { + "bbox": [ + 197, + 302, + 242, + 312 + ], + "spans": [ + { + "bbox": [ + 197, + 302, + 242, + 312 + ], + "type": "text", + "content": "(b) FeMaSR" + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "image_caption" + } + ], + "index": 4 + }, + { + "type": "image", + "bbox": [ + 251, + 114, + 304, + 300 + ], + "blocks": [ + { + "bbox": [ + 251, + 114, + 304, + 300 + ], + "lines": [ + { + "bbox": [ + 251, + 114, + 304, + 300 + ], + "spans": [ + { + "bbox": [ + 251, + 114, + 304, + 300 + ], + "type": "image", + "image_path": "a701cbffed2b69cac700b3bfe630a36a0ea7ecf6909494a40d05b33a34f62112.jpg" + } + ] + } + ], + "index": 6, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 260, + 302, + 296, + 311 + ], + "lines": [ + { + "bbox": [ + 260, + 302, + 296, + 311 + ], + "spans": [ + { + "bbox": [ + 260, + 302, + 296, + 311 + ], + "type": "text", + "content": "(c) DASR" + } + ] + } + ], + "index": 7, + "angle": 0, + "type": "image_caption" + } + ], + "index": 6 + }, + { + "type": "image", + "bbox": [ + 309, + 114, + 362, + 300 + ], + "blocks": [ + { + "bbox": [ + 309, + 114, + 362, + 300 + ], + "lines": [ + { + "bbox": [ + 309, + 114, + 362, + 300 + ], + "spans": [ + { + "bbox": [ + 309, + 114, + 362, + 300 + ], + "type": "image", + "image_path": "94a365716f9add06d7e9deef1b9d16872979103513cd73fdecf30118425c7765.jpg" + } + ] + } + ], + "index": 8, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 323, + 302, + 349, + 312 + ], + "lines": [ + { + "bbox": [ + 323, + 302, + 349, + 312 + ], + "spans": [ + { + "bbox": [ + 323, + 302, + 349, + 312 + ], + "type": "text", + "content": "(d) RE" + } + ] + } + ], + "index": 9, + "angle": 0, + "type": "image_caption" + } + ], + "index": 8 + }, + { + "type": "image", + "bbox": [ + 367, + 114, + 421, + 300 + ], + "blocks": [ + { + "bbox": [ + 367, + 114, + 421, + 300 + ], + "lines": [ + { + "bbox": [ + 367, + 114, + 421, + 300 + ], + "spans": [ + { + "bbox": [ + 367, + 114, + 421, + 300 + ], + "type": "image", + "image_path": "cc55b2e784cee3de2f35c78937d0beaea3ab304e47ca312581fb3aed6c3fa279.jpg" + } + ] + } + ], + "index": 10, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 370, + 302, + 419, + 311 + ], + "lines": [ + { + "bbox": [ + 370, + 302, + 419, + 311 + ], + "spans": [ + { + "bbox": [ + 370, + 302, + 419, + 311 + ], + "type": "text", + "content": "(e) RE+ours" + } + ] + } + ], + "index": 11, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 131, + 320, + 480, + 354 + ], + "lines": [ + { + "bbox": [ + 131, + 320, + 480, + 354 + ], + "spans": [ + { + "bbox": [ + 131, + 320, + 480, + 354 + ], + "type": "text", + "content": "Fig. 8: Qualitative comparisons. LRs are from the RealSR-Canon dataset. Columns (e) and (f) are the results of applying our methods on RealESRGAN (RE) and BSRGAN (BG). FeMaSR and DASR can not predict reasonable details." + } + ] + } + ], + "index": 14, + "angle": 0, + "type": "image_caption" + } + ], + "index": 10 + }, + { + "type": "image", + "bbox": [ + 425, + 114, + 479, + 300 + ], + "blocks": [ + { + "bbox": [ + 425, + 114, + 479, + 300 + ], + "lines": [ + { + "bbox": [ + 425, + 114, + 479, + 300 + ], + "spans": [ + { + "bbox": [ + 425, + 114, + 479, + 300 + ], + "type": "image", + "image_path": "76210f2e54e46d87b870239539695486ecac9b1d4fac627b1b821560a83a56a0.jpg" + } + ] + } + ], + "index": 12, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 426, + 302, + 476, + 311 + ], + "lines": [ + { + "bbox": [ + 426, + 302, + 476, + 311 + ], + "spans": [ + { + "bbox": [ + 426, + 302, + 476, + 311 + ], + "type": "text", + "content": "(f) BG+ours" + } + ] + } + ], + "index": 13, + "angle": 0, + "type": "image_caption" + } + ], + "index": 12 + }, + { + "bbox": [ + 130, + 377, + 479, + 401 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 377, + 479, + 401 + ], + "spans": [ + { + "bbox": [ + 130, + 377, + 479, + 401 + ], + "type": "text", + "content": "titatively competing, FeMaSR yields unpleasant artifacts to human eyes, further shown in qualitative comparison." + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 130, + 402, + 480, + 449 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 402, + 480, + 449 + ], + "spans": [ + { + "bbox": [ + 130, + 402, + 480, + 449 + ], + "type": "text", + "content": "Qualitative Comparison is in Fig. 8. FeMaSR [3] produces high-frequency artifacts, while outputs from DASR [40] suffer blur or artifacts. Compared to RealESRGAN, our method predicts sharper patterns without introducing distortion. More comparisons are in Supplementary Sec. F." + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 131, + 466, + 209, + 477 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 131, + 466, + 209, + 477 + ], + "spans": [ + { + "bbox": [ + 131, + 466, + 209, + 477 + ], + "type": "text", + "content": "4.4 Ablations" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 130, + 486, + 480, + 581 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 486, + 480, + 581 + ], + "spans": [ + { + "bbox": [ + 130, + 486, + 480, + 581 + ], + "type": "text", + "content": "Ablation on loss terms reveals the effectiveness of each term in Eq. (11). All experiments take RealESRGAN as the pre-trained model and follow the EMA implementation of our method in Fig. 3 (b). Fig. 9 shows that " + }, + { + "bbox": [ + 130, + 486, + 480, + 581 + ], + "type": "inline_equation", + "content": "R_{intra}" + }, + { + "bbox": [ + 130, + 486, + 480, + 581 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 130, + 486, + 480, + 581 + ], + "type": "inline_equation", + "content": "R_{inter}" + }, + { + "bbox": [ + 130, + 486, + 480, + 581 + ], + "type": "text", + "content": " relate differently to fidelity and perception. Their combination achieves balanced improvements in both directions, and the " + }, + { + "bbox": [ + 130, + 486, + 480, + 581 + ], + "type": "inline_equation", + "content": "L_{\\mathrm{gan}}" + }, + { + "bbox": [ + 130, + 486, + 480, + 581 + ], + "type": "text", + "content": " increases pattern sharpness with a slight trade-off in PSNR. Controlling the ratio between weights of " + }, + { + "bbox": [ + 130, + 486, + 480, + 581 + ], + "type": "inline_equation", + "content": "R_{intra}" + }, + { + "bbox": [ + 130, + 486, + 480, + 581 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 130, + 486, + 480, + 581 + ], + "type": "inline_equation", + "content": "R_{inter}" + }, + { + "bbox": [ + 130, + 486, + 480, + 581 + ], + "type": "text", + "content": " results in fidelity and perceptual quality trade-offs. We choose their balance in Sec. 4.3 and refer to Supplementary Sec. D.a for more discussion." + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 130, + 582, + 480, + 640 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 582, + 480, + 640 + ], + "spans": [ + { + "bbox": [ + 130, + 582, + 480, + 640 + ], + "type": "text", + "content": "Effectiveness of EMA is further validated by comparing to a SingleFixed version, which also uses RealESRGAN as the pre-trained model for both " + }, + { + "bbox": [ + 130, + 582, + 480, + 640 + ], + "type": "inline_equation", + "content": "M_S" + }, + { + "bbox": [ + 130, + 582, + 480, + 640 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 130, + 582, + 480, + 640 + ], + "type": "inline_equation", + "content": "M_G" + }, + { + "bbox": [ + 130, + 582, + 480, + 640 + ], + "type": "text", + "content": " but fixes the " + }, + { + "bbox": [ + 130, + 582, + 480, + 640 + ], + "type": "inline_equation", + "content": "M_G" + }, + { + "bbox": [ + 130, + 582, + 480, + 640 + ], + "type": "text", + "content": " during training. As shown in Tab. 3, updating the weights of the generalist network through EMA enables higher performance in both fidelity and perception." + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 131, + 641, + 480, + 666 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 131, + 641, + 480, + 666 + ], + "spans": [ + { + "bbox": [ + 131, + 641, + 480, + 666 + ], + "type": "text", + "content": "Choice of " + }, + { + "bbox": [ + 131, + 641, + 480, + 666 + ], + "type": "inline_equation", + "content": "R_{intra}" + }, + { + "bbox": [ + 131, + 641, + 480, + 666 + ], + "type": "text", + "content": " measurement. Eq. (6) use Cross-Entropy (CE) for measuring the difference between " + }, + { + "bbox": [ + 131, + 641, + 480, + 666 + ], + "type": "inline_equation", + "content": "d_S^{ij}" + }, + { + "bbox": [ + 131, + 641, + 480, + 666 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 131, + 641, + 480, + 666 + ], + "type": "inline_equation", + "content": "d_G^{ij}" + }, + { + "bbox": [ + 131, + 641, + 480, + 666 + ], + "type": "text", + "content": ". Here, we alternate the CE to " + }, + { + "bbox": [ + 131, + 641, + 480, + 666 + ], + "type": "inline_equation", + "content": "\\ell_1" + }, + { + "bbox": [ + 131, + 641, + 480, + 666 + ], + "type": "text", + "content": " distance" + } + ] + } + ], + "index": 20 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 322, + 91, + 447, + 100 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 322, + 91, + 447, + 100 + ], + "spans": [ + { + "bbox": [ + 322, + 91, + 447, + 100 + ], + "type": "text", + "content": "Pairwise Distance Distillation" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 470, + 91, + 480, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 470, + 91, + 480, + 100 + ], + "spans": [ + { + "bbox": [ + 470, + 91, + 480, + 100 + ], + "type": "text", + "content": "13" + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 12 + }, + { + "para_blocks": [ + { + "type": "table", + "bbox": [ + 195, + 180, + 420, + 250 + ], + "blocks": [ + { + "bbox": [ + 130, + 114, + 482, + 170 + ], + "lines": [ + { + "bbox": [ + 130, + 114, + 482, + 170 + ], + "spans": [ + { + "bbox": [ + 130, + 114, + 482, + 170 + ], + "type": "text", + "content": "Table 3: Collection of ablations on EMA and choice of measurement in Eq. (6). Bold and underline mark the best and second best scores. SingleFixed uses the same prepretrained model to initialize " + }, + { + "bbox": [ + 130, + 114, + 482, + 170 + ], + "type": "inline_equation", + "content": "M_S" + }, + { + "bbox": [ + 130, + 114, + 482, + 170 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 130, + 114, + 482, + 170 + ], + "type": "inline_equation", + "content": "M_G" + }, + { + "bbox": [ + 130, + 114, + 482, + 170 + ], + "type": "text", + "content": " as the EMA version but fixes the weights of " + }, + { + "bbox": [ + 130, + 114, + 482, + 170 + ], + "type": "inline_equation", + "content": "M_G" + }, + { + "bbox": [ + 130, + 114, + 482, + 170 + ], + "type": "text", + "content": " during training, resulting in inferior performance. Substituting CE in Eq. (6) to " + }, + { + "bbox": [ + 130, + 114, + 482, + 170 + ], + "type": "inline_equation", + "content": "\\ell_1" + }, + { + "bbox": [ + 130, + 114, + 482, + 170 + ], + "type": "text", + "content": " biases the fidelity scores." + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 195, + 180, + 420, + 250 + ], + "lines": [ + { + "bbox": [ + 195, + 180, + 420, + 250 + ], + "spans": [ + { + "bbox": [ + 195, + 180, + 420, + 250 + ], + "type": "table", + "html": "
MethodCanonNTIRE20Olympus
PSNRNRQMPSNRNRQMPSNRNRQM
Generalist24.746.064925.086.121326.735.2530
SingleFixed25.715.980325.756.093827.475.2071
\\( \\ell_1 \\)26.175.806126.446.199727.724.9602
Ours25.946.064726.426.226327.405.3722
", + "image_path": "8cd1517ede0fac8fcb7376f3a78dd5775236007edd18f747c7382e0a418cd902.jpg" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "table_body" + } + ], + "index": 3 + }, + { + "type": "image", + "bbox": [ + 185, + 262, + 427, + 357 + ], + "blocks": [ + { + "bbox": [ + 185, + 262, + 427, + 357 + ], + "lines": [ + { + "bbox": [ + 185, + 262, + 427, + 357 + ], + "spans": [ + { + "bbox": [ + 185, + 262, + 427, + 357 + ], + "type": "image", + "image_path": "88a535fed39f63c499dc3cda070746fde06c7777cc173d5c05a8a2c3d542523a.jpg" + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 130, + 365, + 482, + 399 + ], + "lines": [ + { + "bbox": [ + 130, + 365, + 482, + 399 + ], + "spans": [ + { + "bbox": [ + 130, + 365, + 482, + 399 + ], + "type": "text", + "content": "Fig.9: Ablations of loss terms for distillation in Eq. (11). Using one of " + }, + { + "bbox": [ + 130, + 365, + 482, + 399 + ], + "type": "inline_equation", + "content": "R_{intra}" + }, + { + "bbox": [ + 130, + 365, + 482, + 399 + ], + "type": "text", + "content": " (blue) and " + }, + { + "bbox": [ + 130, + 365, + 482, + 399 + ], + "type": "inline_equation", + "content": "R_{inter}" + }, + { + "bbox": [ + 130, + 365, + 482, + 399 + ], + "type": "text", + "content": " (orange) bias fidelity or perceptual quality (a.-b.), while the combination of them (c.-d.) achieves a balanced improvement (see green and red dots)." + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "image_caption" + } + ], + "index": 4 + }, + { + "bbox": [ + 130, + 426, + 482, + 475 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 426, + 482, + 475 + ], + "spans": [ + { + "bbox": [ + 130, + 426, + 482, + 475 + ], + "type": "text", + "content": "without changing the ratio between weights of " + }, + { + "bbox": [ + 130, + 426, + 482, + 475 + ], + "type": "inline_equation", + "content": "R_{intra}" + }, + { + "bbox": [ + 130, + 426, + 482, + 475 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 130, + 426, + 482, + 475 + ], + "type": "inline_equation", + "content": "R_{inter}" + }, + { + "bbox": [ + 130, + 426, + 482, + 475 + ], + "type": "text", + "content": ". As shown in Tab. 3, CE favors the overall sharpness while " + }, + { + "bbox": [ + 130, + 426, + 482, + 475 + ], + "type": "inline_equation", + "content": "\\ell_1" + }, + { + "bbox": [ + 130, + 426, + 482, + 475 + ], + "type": "text", + "content": " focuses more on fidelity. Although " + }, + { + "bbox": [ + 130, + 426, + 482, + 475 + ], + "type": "inline_equation", + "content": "\\ell_1" + }, + { + "bbox": [ + 130, + 426, + 482, + 475 + ], + "type": "text", + "content": " is a feasible choice, we use CE in our method due to the importance of sharpness for human eyes [13]." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 132, + 500, + 220, + 513 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 500, + 220, + 513 + ], + "spans": [ + { + "bbox": [ + 132, + 500, + 220, + 513 + ], + "type": "text", + "content": "5 Conclusion" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 130, + 533, + 482, + 666 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 533, + 482, + 666 + ], + "spans": [ + { + "bbox": [ + 130, + 533, + 482, + 666 + ], + "type": "text", + "content": "This paper studies unsupervised real-world image super-resolution from the distillation perspective. The approach considers both the real-world knowledge of a generalized SR model and the low-level characteristics of synthetic predictions of the specialist model, adapting a synthetic domain specialist to target real-world degradation. The core of our method is the Pairwise Distance Distillation. By enforcing the consistencies for intra- and inter-model distances, our method aims at pushing the low-level characteristics of the specialist's real-world predictions towards its synthetic predictions. As a learning scheme, our method can improve off-the-shelf models regarding fidelity and perception on multiple real-world datasets. Additionally, we emphasize our approach to offering a new perspective toward addressing real-world super-resolution challenges." + } + ] + } + ], + "index": 8 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "spans": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "type": "text", + "content": "14" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 166, + 91, + 229, + 102 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 166, + 91, + 229, + 102 + ], + "spans": [ + { + "bbox": [ + 166, + 91, + 229, + 102 + ], + "type": "text", + "content": "Y. Zhang et al." + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 13 + }, + { + "para_blocks": [ + { + "bbox": [ + 132, + 116, + 480, + 140 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 116, + 480, + 140 + ], + "spans": [ + { + "bbox": [ + 132, + 116, + 480, + 140 + ], + "type": "text", + "content": "Acknowledgement This research/project is supported by the Ministry of Education, Singapore, under the Academic Research Fund Tier 1 (FY2022)." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 133, + 156, + 197, + 169 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 156, + 197, + 169 + ], + "spans": [ + { + "bbox": [ + 133, + 156, + 197, + 169 + ], + "type": "text", + "content": "References" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 138, + 178, + 480, + 665 + ], + "type": "list", + "angle": 0, + "index": 19, + "blocks": [ + { + "bbox": [ + 138, + 178, + 480, + 212 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 178, + 480, + 212 + ], + "spans": [ + { + "bbox": [ + 138, + 178, + 480, + 212 + ], + "type": "text", + "content": "1. Agustsson, E., Timofte, R.: Ntire 2017 challenge on single image super-resolution: Dataset and study. In: The IEEE Conference on Computer Vision and Pattern Recognition (CVPR) Workshops (July 2017)" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 138, + 213, + 480, + 244 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 213, + 480, + 244 + ], + "spans": [ + { + "bbox": [ + 138, + 213, + 480, + 244 + ], + "type": "text", + "content": "2. Cai, J., Zeng, H., Yong, H., Cao, Z., Zhang, L.: Toward real-world single image super-resolution: A new benchmark and a new model. In: Proceedings of the IEEE/CVF International Conference on Computer Vision. pp. 3086-3095 (2019)" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 138, + 245, + 480, + 287 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 245, + 480, + 287 + ], + "spans": [ + { + "bbox": [ + 138, + 245, + 480, + 287 + ], + "type": "text", + "content": "3. Chen, C., Shi, X., Qin, Y., Li, X., Han, X., Yang, T., Guo, S.: Real-world blind super-resolution via feature matching with implicit high-resolution priors. In: Proceedings of the 30th ACM International Conference on Multimedia. pp. 1329-1338 (2022)" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 138, + 288, + 480, + 330 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 288, + 480, + 330 + ], + "spans": [ + { + "bbox": [ + 138, + 288, + 480, + 330 + ], + "type": "text", + "content": "4. Deng, X., Yang, R., Xu, M., Dragotti, P.L.: Wavelet domain style transfer for an effective perception-distortion tradeoff in single image super-resolution. In: Proceedings of the IEEE/CVF international conference on computer vision. pp. 3076-3085 (2019)" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 138, + 331, + 480, + 363 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 331, + 480, + 363 + ], + "spans": [ + { + "bbox": [ + 138, + 331, + 480, + 363 + ], + "type": "text", + "content": "5. Fritsche, M., Gu, S., Timofte, R.: Frequency separation for real-world superresolution. In: 2019 IEEE/CVF International Conference on Computer Vision Workshop (ICCVW). pp. 3599-3608. IEEE (2019)" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 138, + 364, + 480, + 407 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 364, + 480, + 407 + ], + "spans": [ + { + "bbox": [ + 138, + 364, + 480, + 407 + ], + "type": "text", + "content": "6. Gao, H., Guo, J., Wang, G., Zhang, Q.: Cross-domain correlation distillation for unsupervised domain adaptation in nighttime semantic segmentation. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 9913-9923 (2022)" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 138, + 407, + 480, + 429 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 407, + 480, + 429 + ], + "spans": [ + { + "bbox": [ + 138, + 407, + 480, + 429 + ], + "type": "text", + "content": "7. Gao, Q., Zhao, Y., Li, G., Tong, T.: Image super-resolution using knowledge distillation. In: Asian Conference on Computer Vision. pp. 527-541. Springer (2018)" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 138, + 429, + 480, + 460 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 429, + 480, + 460 + ], + "spans": [ + { + "bbox": [ + 138, + 429, + 480, + 460 + ], + "type": "text", + "content": "8. Gatys, L.A., Ecker, A.S., Bethge, M.: Image style transfer using convolutional neural networks. In: Proceedings of the IEEE conference on computer vision and pattern recognition. pp. 2414-2423 (2016)" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 138, + 461, + 480, + 493 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 461, + 480, + 493 + ], + "spans": [ + { + "bbox": [ + 138, + 461, + 480, + 493 + ], + "type": "text", + "content": "9. Gu, J., Lu, H., Zuo, W., Dong, C.: Blind super-resolution with iterative kernel correction. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 1604-1613 (2019)" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 138, + 494, + 480, + 514 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 494, + 480, + 514 + ], + "spans": [ + { + "bbox": [ + 138, + 494, + 480, + 514 + ], + "type": "text", + "content": "0. Hinton, G., Vinyals, O., Dean, J.: Distilling the knowledge in a neural network. arXiv preprint arXiv:1503.02531 (2015)" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 138, + 514, + 480, + 546 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 514, + 480, + 546 + ], + "spans": [ + { + "bbox": [ + 138, + 514, + 480, + 546 + ], + "type": "text", + "content": "1. Huang, Y., Li, S., Wang, L., Tan, T., et al.: Unfolding the alternating optimization for blind super resolution. Advances in Neural Information Processing Systems 33, 5632-5643 (2020)" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 138, + 547, + 480, + 568 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 547, + 480, + 568 + ], + "spans": [ + { + "bbox": [ + 138, + 547, + 480, + 568 + ], + "type": "text", + "content": "2. Kingma, D.P., Ba, J.: Adam: A method for stochastic optimization. arXiv preprint arXiv:1412.6980 (2014)" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 138, + 569, + 480, + 611 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 569, + 480, + 611 + ], + "spans": [ + { + "bbox": [ + 138, + 569, + 480, + 611 + ], + "type": "text", + "content": "3. Ledig, C., Theis, L., Huszár, F., Caballero, J., Cunningham, A., Acosta, A., Aitken, A., Tejani, A., Totz, J., Wang, Z., et al.: Photo-realistic single image superresolution using a generative adversarial network. In: Proceedings of the IEEE conference on computer vision and pattern recognition. pp. 4681-4690 (2017)" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 138, + 612, + 480, + 633 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 612, + 480, + 633 + ], + "spans": [ + { + "bbox": [ + 138, + 612, + 480, + 633 + ], + "type": "text", + "content": "4. Lepik, U., Hein, H.: Haar wavelets. In: Haar Wavelets: With Applications, pp. 7-20. Springer (2014)" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 138, + 634, + 480, + 665 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 634, + 480, + 665 + ], + "spans": [ + { + "bbox": [ + 138, + 634, + 480, + 665 + ], + "type": "text", + "content": "5. Li, W., Li, L., Yang, H.: Progressive cross-domain knowledge distillation for efficient unsupervised domain adaptive object detection. Engineering Applications of Artificial Intelligence 119, 105774 (2023)" + } + ] + } + ], + "index": 18 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 321, + 91, + 447, + 101 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 321, + 91, + 447, + 101 + ], + "spans": [ + { + "bbox": [ + 321, + 91, + 447, + 101 + ], + "type": "text", + "content": "Pairwise Distance Distillation" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 470, + 92, + 481, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 470, + 92, + 481, + 100 + ], + "spans": [ + { + "bbox": [ + 470, + 92, + 481, + 100 + ], + "type": "text", + "content": "15" + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 14 + }, + { + "para_blocks": [ + { + "bbox": [ + 132, + 116, + 480, + 665 + ], + "type": "list", + "angle": 0, + "index": 19, + "blocks": [ + { + "bbox": [ + 133, + 116, + 480, + 138 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 116, + 480, + 138 + ], + "spans": [ + { + "bbox": [ + 133, + 116, + 480, + 138 + ], + "type": "text", + "content": "16. Li, Y., Wang, N., Liu, J., Hou, X.: Demystifying neural style transfer. arXiv preprint arXiv:1701.01036 (2017)" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 133, + 140, + 480, + 171 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 140, + 480, + 171 + ], + "spans": [ + { + "bbox": [ + 133, + 140, + 480, + 171 + ], + "type": "text", + "content": "17. Liang, J., Zeng, H., Zhang, L.: Efficient and degradation-adaptive network for real-world image super-resolution. In: European Conference on Computer Vision. pp. 574-591. Springer (2022)" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 133, + 172, + 480, + 205 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 172, + 480, + 205 + ], + "spans": [ + { + "bbox": [ + 133, + 172, + 480, + 205 + ], + "type": "text", + "content": "18. Liang, J., Cao, J., Sun, G., Zhang, K., Van Gool, L., Timofte, R.: Swinir: Image restoration using swim transformer. In: Proceedings of the IEEE/CVF international conference on computer vision. pp. 1833-1844 (2021)" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 133, + 205, + 480, + 237 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 205, + 480, + 237 + ], + "spans": [ + { + "bbox": [ + 133, + 205, + 480, + 237 + ], + "type": "text", + "content": "19. Lim, B., Son, S., Kim, H., Nah, S., Mu Lee, K.: Enhanced deep residual networks for single image super-resolution. In: Proceedings of the IEEE conference on computer vision and pattern recognition workshops. pp. 136-144 (2017)" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 132, + 238, + 480, + 270 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 238, + 480, + 270 + ], + "spans": [ + { + "bbox": [ + 132, + 238, + 480, + 270 + ], + "type": "text", + "content": "20. Liu, Y., Liu, A., Gu, J., Zhang, Z., Wu, W., Qiao, Y., Dong, C.: Discovering distinctive\" semantics\" in super-resolution networks. arXiv preprint arXiv:2108.00406 (2021)" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 132, + 270, + 480, + 303 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 270, + 480, + 303 + ], + "spans": [ + { + "bbox": [ + 132, + 270, + 480, + 303 + ], + "type": "text", + "content": "21. Liu, Y., Zhao, H., Gu, J., Qiao, Y., Dong, C.: Evaluating the generalization ability of super-resolution networks. IEEE Transactions on pattern analysis and machine intelligence (2023)" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 132, + 304, + 480, + 325 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 304, + 480, + 325 + ], + "spans": [ + { + "bbox": [ + 132, + 304, + 480, + 325 + ], + "type": "text", + "content": "22. Liu, Y., Zhang, W., Wang, J.: Adaptive multi-teacher multi-level knowledge distillation. Neurocomputing 415, 106-113 (2020)" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 132, + 326, + 480, + 357 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 326, + 480, + 357 + ], + "spans": [ + { + "bbox": [ + 132, + 326, + 480, + 357 + ], + "type": "text", + "content": "23. Lugmayr, A., Danelljan, M., Timofte, R.: Ntire 2020 challenge on real-world image super-resolution: Methods and results. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition Workshops. pp. 494-495 (2020)" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 132, + 358, + 480, + 390 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 358, + 480, + 390 + ], + "spans": [ + { + "bbox": [ + 132, + 358, + 480, + 390 + ], + "type": "text", + "content": "24. Ma, C., Yang, C.Y., Yang, X., Yang, M.H.: Learning a no-reference quality metric for single-image super-resolution. Computer Vision and Image Understanding 158, 1-16 (2017)" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 132, + 391, + 480, + 424 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 391, + 480, + 424 + ], + "spans": [ + { + "bbox": [ + 132, + 391, + 480, + 424 + ], + "type": "text", + "content": "25. Ma, C., Rao, Y., Cheng, Y., Chen, C., Lu, J., Zhou, J.: Structure-preserving super resolution with gradient guidance. In: Proceedings of the IEEE/CVF conference on computer vision and pattern recognition. pp. 7769-7778 (2020)" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 132, + 425, + 480, + 467 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 425, + 480, + 467 + ], + "spans": [ + { + "bbox": [ + 132, + 425, + 480, + 467 + ], + "type": "text", + "content": "26. Maggioni, M., Tanay, T., Babiloni, F., McDonagh, S., Leonardis, A.: Tunable convolutions with parametric multi-loss optimization. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 20226-20236 (2023)" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 132, + 468, + 480, + 511 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 468, + 480, + 511 + ], + "spans": [ + { + "bbox": [ + 132, + 468, + 480, + 511 + ], + "type": "text", + "content": "27. Nguyen-Meidine, L.T., Belal, A., Kiran, M., Dolz, J., Blais-Morin, L.A., Granger, E.: Unsupervised multi-target domain adaptation through knowledge distillation. In: Proceedings of the IEEE/CVF Winter Conference on Applications of Computer Vision. pp. 1339-1347 (2021)" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 132, + 512, + 480, + 533 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 512, + 480, + 533 + ], + "spans": [ + { + "bbox": [ + 132, + 512, + 480, + 533 + ], + "type": "text", + "content": "28. Park, S., Kwak, N.: Local-selective feature distillation for single image superresolution. arXiv preprint arXiv:2111.10988 (2021)" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 132, + 534, + 480, + 566 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 534, + 480, + 566 + ], + "spans": [ + { + "bbox": [ + 132, + 534, + 480, + 566 + ], + "type": "text", + "content": "29. Park, S.H., Moon, Y.S., Cho, N.I.: Perception-oriented single image superresolution using optimal objective estimation. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 1725-1735 (2023)" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 132, + 567, + 480, + 599 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 567, + 480, + 599 + ], + "spans": [ + { + "bbox": [ + 132, + 567, + 480, + 599 + ], + "type": "text", + "content": "30. Romero, A., Van Gool, L., Timofte, R.: Unpaired real-world super-resolution with pseudo controllable restoration. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 798-807 (2022)" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 132, + 600, + 480, + 621 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 600, + 480, + 621 + ], + "spans": [ + { + "bbox": [ + 132, + 600, + 480, + 621 + ], + "type": "text", + "content": "31. Simonyan, K., Zisserman, A.: Very deep convolutional networks for large-scale image recognition. arXiv preprint arXiv:1409.1556 (2014)" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 132, + 622, + 480, + 665 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 622, + 480, + 665 + ], + "spans": [ + { + "bbox": [ + 132, + 622, + 480, + 665 + ], + "type": "text", + "content": "32. Timofte, R., Agustsson, E., Van Gool, L., Yang, M.H., Zhang, L.: Ntire 2017 challenge on single image super-resolution: Methods and results. In: Proceedings of the IEEE conference on computer vision and pattern recognition workshops. pp. 114-125 (2017)" + } + ] + } + ], + "index": 18 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "spans": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "type": "text", + "content": "16" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 166, + 91, + 229, + 102 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 166, + 91, + 229, + 102 + ], + "spans": [ + { + "bbox": [ + 166, + 91, + 229, + 102 + ], + "type": "text", + "content": "Y. Zhang et al." + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 15 + }, + { + "para_blocks": [ + { + "bbox": [ + 130, + 116, + 482, + 666 + ], + "type": "list", + "angle": 0, + "index": 18, + "blocks": [ + { + "bbox": [ + 130, + 116, + 482, + 149 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 116, + 482, + 149 + ], + "spans": [ + { + "bbox": [ + 130, + 116, + 482, + 149 + ], + "type": "text", + "content": "33. Wang, H., Chen, X., Ni, B., Liu, Y., Liu, J.: Omni aggregation networks for lightweight image super-resolution. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 22378-22387 (2023)" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 130, + 149, + 482, + 182 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 149, + 482, + 182 + ], + "spans": [ + { + "bbox": [ + 130, + 149, + 482, + 182 + ], + "type": "text", + "content": "34. Wang, W., Zhang, H., Yuan, Z., Wang, C.: Unsupervised real-world superresolution: A domain adaptation perspective. In: Proceedings of the IEEE/CVF International Conference on Computer Vision. pp. 4318-4327 (2021)" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 130, + 182, + 482, + 213 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 182, + 482, + 213 + ], + "spans": [ + { + "bbox": [ + 130, + 182, + 482, + 213 + ], + "type": "text", + "content": "35. Wang, X., Xie, L., Dong, C., Shan, Y.: Real-esrgan: Training real-world blind super-resolution with pure synthetic data. In: Proceedings of the IEEE/CVF international conference on computer vision. pp. 1905-1914 (2021)" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 130, + 213, + 482, + 257 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 213, + 482, + 257 + ], + "spans": [ + { + "bbox": [ + 130, + 213, + 482, + 257 + ], + "type": "text", + "content": "36. Wang, X., Yu, K., Wu, S., Gu, J., Liu, Y., Dong, C., Qiao, Y., Change Loy, C.: Esrgan: Enhanced super-resolution generative adversarial networks. In: Proceedings of the European conference on computer vision (ECCV) workshops. pp. 0-0 (2018)" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 130, + 257, + 482, + 289 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 257, + 482, + 289 + ], + "spans": [ + { + "bbox": [ + 130, + 257, + 482, + 289 + ], + "type": "text", + "content": "37. Wang, Y., Cao, Y., Zha, Z.J., Zhang, J., Xiong, Z.: Deep degradation prior for low-quality image classification. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 11049-11058 (2020)" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 130, + 289, + 482, + 322 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 289, + 482, + 322 + ], + "spans": [ + { + "bbox": [ + 130, + 289, + 482, + 322 + ], + "type": "text", + "content": "38. Wang, Y., Wang, L., Shi, S., Li, V.O., Tu, Z.: Go from the general to the particular: Multi-domain translation with domain transformation networks. In: Proceedings of the AAAI Conference on Artificial Intelligence. vol. 34, pp. 9233-9241 (2020)" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 130, + 322, + 482, + 365 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 322, + 482, + 365 + ], + "spans": [ + { + "bbox": [ + 130, + 322, + 482, + 365 + ], + "type": "text", + "content": "39. Wei, P., Xie, Z., Lu, H., Zhan, Z., Ye, Q., Zuo, W., Lin, L.: Component divide-and-conquer for real-world image super-resolution. In: Computer Vision-ECCV 2020: 16th European Conference, Glasgow, UK, August 23-28, 2020, Proceedings, Part VIII 16. pp. 101-117. Springer (2020)" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 130, + 365, + 482, + 407 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 365, + 482, + 407 + ], + "spans": [ + { + "bbox": [ + 130, + 365, + 482, + 407 + ], + "type": "text", + "content": "40. Wei, Y., Gu, S., Li, Y., Timofte, R., Jin, L., Song, H.: Unsupervised real-world image super resolution via domain-distance aware training. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 13385-13394 (2021)" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 130, + 407, + 482, + 440 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 407, + 482, + 440 + ], + "spans": [ + { + "bbox": [ + 130, + 407, + 482, + 440 + ], + "type": "text", + "content": "41. Yang, C., Pan, J., Gao, X., Jiang, T., Liu, D., Chen, G.: Cross-task knowledge distillation in multi-task recommendation. In: Proceedings of the AAAI Conference on Artificial Intelligence. vol. 36, pp. 4318-4326 (2022)" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 130, + 440, + 482, + 472 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 440, + 482, + 472 + ], + "spans": [ + { + "bbox": [ + 130, + 440, + 482, + 472 + ], + "type": "text", + "content": "42. Zeyde, R., Elad, M., Protter, M.: On single image scale-up using sparse-representations. In: Curves and Surfaces: 7th International Conference, Avignon, France, June 24-30, 2010, Revised Selected Papers 7. pp. 711-730. Springer (2012)" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 130, + 472, + 482, + 505 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 472, + 482, + 505 + ], + "spans": [ + { + "bbox": [ + 130, + 472, + 482, + 505 + ], + "type": "text", + "content": "43. Zhang, K., Liang, J., Van Gool, L., Timofte, R.: Designing a practical degradation model for deep blind image super-resolution. In: Proceedings of the IEEE/CVF International Conference on Computer Vision. pp. 4791-4800 (2021)" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 130, + 505, + 482, + 536 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 505, + 482, + 536 + ], + "spans": [ + { + "bbox": [ + 130, + 505, + 482, + 536 + ], + "type": "text", + "content": "44. Zhang, R., Isola, P., Efros, A.A., Shechtman, E., Wang, O.: The unreasonable effectiveness of deep features as a perceptual metric. In: Proceedings of the IEEE conference on computer vision and pattern recognition. pp. 586-595 (2018)" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 130, + 536, + 482, + 568 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 536, + 482, + 568 + ], + "spans": [ + { + "bbox": [ + 130, + 536, + 482, + 568 + ], + "type": "text", + "content": "45. Zhang, R., Gu, J., Chen, H., Dong, C., Zhang, Y., Yang, W.: Crafting training degradation distribution for the accuracy-generalization trade-off. arXiv preprint arXiv:2305.18107 (2023)" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 130, + 568, + 482, + 601 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 568, + 482, + 601 + ], + "spans": [ + { + "bbox": [ + 130, + 568, + 482, + 601 + ], + "type": "text", + "content": "46. Zhang, W., Li, X., Shi, G., Chen, X., Qiao, Y., Zhang, X., Wu, X.M., Dong, C.: Real-world image super-resolution as multi-task learning. Advances in Neural Information Processing Systems 36 (2024)" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 130, + 601, + 482, + 633 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 601, + 482, + 633 + ], + "spans": [ + { + "bbox": [ + 130, + 601, + 482, + 633 + ], + "type": "text", + "content": "47. Zhang, Y., Chen, H., Chen, X., Deng, Y., Xu, C., Wang, Y.: Data-free knowledge distillation for image super-resolution. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 7852-7861 (2021)" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 130, + 633, + 482, + 666 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 633, + 482, + 666 + ], + "spans": [ + { + "bbox": [ + 130, + 633, + 482, + 666 + ], + "type": "text", + "content": "48. Zhang, Y., Ji, B., Hao, J., Yao, A.: Perception-distortion balanced admm optimization for single-image super-resolution. In: European Conference on Computer Vision. pp. 108-125. Springer (2022)" + } + ] + } + ], + "index": 17 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 321, + 91, + 447, + 101 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 321, + 91, + 447, + 101 + ], + "spans": [ + { + "bbox": [ + 321, + 91, + 447, + 101 + ], + "type": "text", + "content": "Pairwise Distance Distillation" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 470, + 91, + 481, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 470, + 91, + 481, + 100 + ], + "spans": [ + { + "bbox": [ + 470, + 91, + 481, + 100 + ], + "type": "text", + "content": "17" + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 16 + }, + { + "para_blocks": [ + { + "bbox": [ + 132, + 116, + 482, + 183 + ], + "type": "list", + "angle": 0, + "index": 4, + "blocks": [ + { + "bbox": [ + 132, + 116, + 482, + 149 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 116, + 482, + 149 + ], + "spans": [ + { + "bbox": [ + 132, + 116, + 482, + 149 + ], + "type": "text", + "content": "49. Zhong, Z., Shen, T., Yang, Y., Lin, Z., Zhang, C.: Joint sub-bands learning with clique structures for wavelet domain super-resolution. Advances in neural information processing systems 31 (2018)" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 132, + 150, + 482, + 183 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 150, + 482, + 183 + ], + "spans": [ + { + "bbox": [ + 132, + 150, + 482, + 183 + ], + "type": "text", + "content": "50. Zhu, J.Y., Park, T., Isola, P., Efros, A.A.: Unpaired image-to-image translation using cycle-consistent adversarial networks. In: Proceedings of the IEEE international conference on computer vision. pp. 2223-2232 (2017)" + } + ] + } + ], + "index": 3 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "spans": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "type": "text", + "content": "18" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 166, + 91, + 229, + 102 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 166, + 91, + 229, + 102 + ], + "spans": [ + { + "bbox": [ + 166, + 91, + 229, + 102 + ], + "type": "text", + "content": "Y. Zhang et al." + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 17 + } + ], + "_backend": "vlm", + "_version_name": "2.6.4" +} \ No newline at end of file diff --git a/2024/PanGu-Draw_ Advancing Resource-Efficient Text-to-Image Synthesis with Time-Decoupled Training and Reusable Coop-Diffusion/c75d5546-18a8-4ad8-982c-ff094dd050a6_content_list.json b/2024/PanGu-Draw_ Advancing Resource-Efficient Text-to-Image Synthesis with Time-Decoupled Training and Reusable Coop-Diffusion/c75d5546-18a8-4ad8-982c-ff094dd050a6_content_list.json new file mode 100644 index 0000000000000000000000000000000000000000..5038543eec19ca3c12139b7cc064dd5c3f1291c7 --- /dev/null +++ b/2024/PanGu-Draw_ Advancing Resource-Efficient Text-to-Image Synthesis with Time-Decoupled Training and Reusable Coop-Diffusion/c75d5546-18a8-4ad8-982c-ff094dd050a6_content_list.json @@ -0,0 +1,1818 @@ +[ + { + "type": "text", + "text": "PanGu-Draw: Advancing Resource-Efficient Text-to-Image Synthesis with Time-Decoupled Training and Reusable Coop-Diffusion", + "text_level": 1, + "bbox": [ + 238, + 140, + 766, + 205 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Guansong Lu $^{1}$ , Yuanfan Guo $^{1}$ , Jianhua Han $^{1}$ , Minzhe Niu $^{1}$ , Yihan Zeng $^{1}$ , Songcen Xu $^{1}$ , Zeyi Huang $^{2}$ , Zhao Zhong $^{2}$ , Wei Zhang $^{1}$ , and Hang Xu $^{1\\star}$", + "bbox": [ + 235, + 231, + 767, + 263 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "1 Huawei Noah's Ark Lab 2 Huawei", + "bbox": [ + 380, + 272, + 622, + 287 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Abstract. Current large-scale diffusion models represent a giant leap forward in conditional image synthesis, capable of interpreting diverse cues like text, human poses, and edges. However, their reliance on substantial computational resources and extensive data collection remains a bottleneck. On the other hand, the integration of existing diffusion models, each specialized for different controls and operating in unique latent spaces, poses a challenge due to incompatible image resolutions and latent space embedding structures, hindering their joint use. Addressing these constraints, we present \"PanGu-Draw\", a novel latent diffusion model designed for resource-efficient text-to-image synthesis that adeptly accommodates multiple control signals. We first propose a resource-efficient Time-Decoupling Training Strategy, which splits the monolithic text-to-image model into structure and texture generators. Each generator is trained using a regimen that maximizes data utilization and computational efficiency, cutting data preparation by $48\\%$ and reducing training resources by $51\\%$ . Secondly, we introduce \"Coop-Diffusion\", an algorithm that enables the cooperative use of various pre-trained diffusion models with different latent spaces and predefined resolutions within a unified denoising process. This allows for multi-control image synthesis at arbitrary resolutions without the necessity for additional data or retraining. Empirical validations of Pangu-Draw show its exceptional prowess in text-to-image and multi-control image generation, suggesting a promising direction for future model training efficiencies and generation versatility. The largest 5B T2I PanGu-Draw model is released on the Ascend platform. Project page: https://pangu- draw.github.io", + "bbox": [ + 261, + 318, + 743, + 652 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Keywords: Text-to-Image Synthesis $\\cdot$ Resource-Efficient $\\cdot$ Diffusion Model", + "bbox": [ + 261, + 662, + 754, + 678 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "1 Introduction", + "text_level": 1, + "bbox": [ + 215, + 700, + 375, + 715 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "The Denoising Diffusion Probabilistic Models (DDPMs) [11] and their subsequent enhancements [5, 14, 21] have established diffusion models as a leading approach for image generation. These advancements excel in the application of diffusion models to text-to-image synthesis, yielding high-fidelity results with large-scale models and datasets, supported by substantial computational resources [13, 20, 24, 26, 29]. These foundational models, capable of understanding", + "bbox": [ + 212, + 728, + 787, + 818 + ], + "page_idx": 0 + }, + { + "type": "page_footnote", + "text": "* Corresponding author: xuHang@huawei.com", + "bbox": [ + 217, + 824, + 532, + 839 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/0cb2654b4516e86a69532b16ceedc69bbf26fbaba52d2b1bace1ef47193282f1.jpg", + "image_caption": [ + "Fig. 1: Illustration of three multi-stage training strategies and comparison between them in resource efficiency in data, training and inference aspects. Our time-decoupling training strategy significantly surpasses the representative methods in Cascaded Training [20,31] and Resolution Boost Training [26,37] in resource efficiency." + ], + "image_footnote": [], + "bbox": [ + 261, + 142, + 751, + 305 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "and rendering complex semantics, have paved the way for diverse image generation tasks, accommodating various control signals such as reference images, edges [39], and poses [39].", + "bbox": [ + 212, + 404, + 782, + 446 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "However, the extensive computational demand and significant data collection required by these models pose a substantial challenge. The ambitious goal of higher fidelity and increased resolution in image synthesis pushes the boundaries of model and dataset sizes, escalating computational costs, and environmental impact. Moreover, the aspiration for versatile control and multi-resolution in image generation introduces additional complexity. Existing diffusion models, each tailored for specific controls and operating within distinct latent spaces, face the challenge of integration due to incompatible image resolutions and latent space embeddings, obstructing their concurrent utilization. This incompatibility not only leads to more resource consumption of retraining but also impedes the joint synthesis of images controlled by multiple factors, thereby limiting the scalability and practical application of such existing generative models. In response to these challenges, our work introduces a novel paradigm named \"PanGu-Draw\" that judiciously conserves training resources while enhancing data efficiency, thereby proposing a resource-efficient pathway forward for diffusion model scalability.", + "bbox": [ + 212, + 448, + 784, + 665 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "As shown in Figure 1, the training strategies of predecessors like DeepFloyd [31] and GLIDE [20], which employ a cascaded approach, excel in leveraging data across resolutions but suffer from inefficient inference due to their reliance on multiple models. Alternatively, Stable Diffusion [26] and AltDiffusion [37] use a Resolution Boost Training strategy aiming for cost-effectiveness by refining a single model. However, this strategy falls short on data efficiency.", + "bbox": [ + 212, + 666, + 784, + 752 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "In light of these considerations, our PanGu-Draw framework advances the field by presenting a Time-Decoupling Training Strategy that segments the training of a comprehensive text-to-image model into two distinct generators: one dedicated to structural outlines and another to textural details. This division not only concentrates on training efforts but also enhances data efficacy. The structural generator is adept at crafting the initial outlines of images, offering", + "bbox": [ + 212, + 752, + 785, + 839 + ], + "page_idx": 1 + }, + { + "type": "page_number", + "text": "2", + "bbox": [ + 217, + 114, + 228, + 126 + ], + "page_idx": 1 + }, + { + "type": "header", + "text": "G. Lu et al.", + "bbox": [ + 271, + 114, + 351, + 126 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "flexibility in data quality and enabling training across a spectrum of data calibers; the textural generator, in contrast, is fine-tuned using low-resolution data to infuse these outlines with fine-grained details, ensuring optimal performance even during high-resolution synthesis. This focused approach not only accelerates the training process of our 5B model but also significantly reduces the reliance on extensive data collection and computational resources, as evidenced by a $48\\%$ reduction in data preparation and a $51\\%$ reduction in resource consumption.", + "bbox": [ + 212, + 146, + 787, + 247 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Furthermore, we introduce a pioneering algorithm named Coop-Diffusion, which facilitates the cooperative integration of diverse pre-trained diffusion models. Each model, conditioned on different controls and pre-defined resolutions, contributes to a seamless denoising process. The first algorithmic sub-module addresses inconsistencies in VAE decoders that arise during the denoising process across different latent spaces, ensuring cohesive image quality by effectively reconciling disparate latent space representations. The second sub-module confronts the challenges associated with multi-resolution denoising. Traditional bilinear upsampling for the intermediate noise map, introduced during the denoising process, can undesirably amplify the correlation between pixels. This amplification deviates from the initial Independent and Identically Distributed (IID) assumption, leading to severe artifacts in the final output image. However, our innovative approach circumvents this issue with a single-step sampling method that preserves the integrity of pixel independence, thus preventing the introduction of artifacts. Coop-Diffusion obviates the need for additional data or model retraining, addressing the challenges of multi-control and multi-resolution image generation with scalability and efficiency.", + "bbox": [ + 212, + 248, + 787, + 494 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "PanGu-Draw excels in text-to-image (T2I) generation, outperforming established models like DALL-E 2 and SDXL, as evidenced by its FID of 7.99 in English T2I. It also leads in Chinese T2I across metrics like FID, IS, and CNCLIP-score. User feedback highlights a strong preference for PanGu-Draw, aligning well with human visual perceptions.", + "bbox": [ + 212, + 494, + 787, + 566 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "In summary, our contributions are manifold:", + "bbox": [ + 238, + 566, + 560, + 580 + ], + "page_idx": 2 + }, + { + "type": "list", + "sub_type": "text", + "list_items": [ + "- PanGu-Draw: A resource-efficient diffusion model with a Time-Decoupling Training Strategy, reducing data and training resources for T2I synthesis.", + "- **Coop-Diffusion:** A novel approach for integrating multiple diffusion models, enabling efficient multi-control image synthesis at multi-resolutions within a unified denoising process.", + "- Comprehensive evaluations demonstrate PanGu-Draw's (5B model) can produce high-quality images aligned with text and various controls, advancing the scalability and flexibility of diffusion-based image generation." + ], + "bbox": [ + 214, + 580, + 784, + 696 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "2 Related Work", + "text_level": 1, + "bbox": [ + 215, + 720, + 387, + 736 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Text-to-Image Generation. The integration of diffusion models into the realm of text-to-image (T2I) generation marks a significant stride in computational creativity [5,7,12,13,20,21,24,26,29,31-33]. Models like GLIDE [20] and DALL-E 2 [24] have significantly advanced in generating diverse and semantically aligned images from textual descriptions. LDM [26] addresses computational challenges by creating images from text-conditioned low-dimensional latent representations.", + "bbox": [ + 212, + 752, + 787, + 840 + ], + "page_idx": 2 + }, + { + "type": "header", + "text": "PanGu-Draw", + "bbox": [ + 640, + 114, + 730, + 127 + ], + "page_idx": 2 + }, + { + "type": "page_number", + "text": "3", + "bbox": [ + 774, + 116, + 785, + 126 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "ControlNet [39] introduces spatial conditioning controls, offering flexibility in image generation under varied conditions like edges and depth. Despite the proliferation of such specialized models, a unified framework that consolidates these disparate capabilities remains absent, limiting the potential for multi-control and complex editing in image synthesis.", + "bbox": [ + 212, + 146, + 782, + 218 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Model Efficient Training and Scaling Up Strategies. Efficient training and scaling of models are pivotal for advancing large-scale neural networks. Previous models like DeepFloyd [31] and GLIDE [20] capitalize on cascaded approaches that proficiently utilize data across various resolutions, which results in less efficient inference processes. Contrastingly, models like Stable Diffusion [26] and AltDiffusion [37] adopt Resolution Boost Training strategies that refine a single model for cost-effectiveness, which however do not fully exploit data efficiency. In scaling up strategies, training efficiency is also important. Efficient adaptation and scaling are explored in [3] through distillation, and in [23] by marrying model expansion with domain-specific prompts. Serial scaling and knowledge distillation reduce training times significantly as demonstrated by [8], while [6] proposes progressive network expansion for faster training with minimal loss. We propose a novel Time-Decoupling training strategy to diffusion model scaling that enhances efficiency. eDiff-I [1] similarly proposes to split a diffusion model across time steps to boost generation quality without increasing inference cost. However, they do not consider about data and training efficiencies.", + "bbox": [ + 212, + 219, + 784, + 450 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "3 Preliminary", + "text_level": 1, + "bbox": [ + 215, + 474, + 367, + 492 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Given an image $x_0$ , diffusion models first produce a series of noisy images $x_1, \\ldots, x_T$ by adding Gaussian noise to $x_0$ according to some noise schedule given by $\\bar{\\alpha}_t$ as follows:", + "bbox": [ + 212, + 506, + 782, + 550 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\nx _ {t} = \\sqrt {\\bar {\\alpha} _ {t}} x _ {0} + \\sqrt {1 - \\bar {\\alpha} _ {t}} \\epsilon , \\epsilon \\sim \\mathcal {N} (0, I). \\tag {1}\n$$\n", + "text_format": "latex", + "bbox": [ + 367, + 561, + 784, + 578 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Diffusion models then learn a denoising model $\\epsilon_{\\theta}(x_t,t)$ to predict the added noise of a noisy image $x_{t}$ with the following training objective:", + "bbox": [ + 214, + 588, + 782, + 618 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\n\\mathcal {L} = \\mathbb {E} _ {x _ {0} \\sim q (x _ {0}), \\epsilon \\sim \\mathcal {N} (0, I), t \\sim [ 1, T ]} \\| \\epsilon - \\epsilon_ {\\theta} (x _ {t}, t) \\| ^ {2}, \\tag {2}\n$$\n", + "text_format": "latex", + "bbox": [ + 339, + 628, + 784, + 646 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "where $t$ is uniformly sampled from $\\{1, \\dots, T\\}$ . Once the denoising model $\\epsilon_{\\theta}(x_t, t)$ is learned, starting from a random noise $x_T \\sim \\mathcal{N}(0, I)$ , one can iteratively predict and reduce the noise in $x_t$ to get a real image $x_0$ . During the sampling process, we can predict the clean data $x_0$ from $\\epsilon_{\\theta}(x_t, t)$ with single-step sampling as:", + "bbox": [ + 212, + 656, + 782, + 715 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\n\\hat {x} _ {0, t} = \\frac {1}{\\sqrt {\\bar {\\alpha} _ {t}}} \\left(x _ {t} - \\sqrt {1 - \\bar {\\alpha} _ {t}} \\epsilon_ {\\theta} \\left(x _ {t}, t\\right)\\right). \\tag {3}\n$$\n", + "text_format": "latex", + "bbox": [ + 377, + 726, + 784, + 757 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Our text-to-image generation model is built on the model architecture proposed in Latent Diffusion Model [26]. In this model, a real image $x_0$ is first down-sampled 8 times as a lower-dimension latent code $z_0$ with an image encoder model $E$ , which can be decoded with a latent decoder model $D$ back to a real image $x_0$ . The denoising network $\\epsilon_{\\theta}(z_t,t,c)$ is parameterized as a U-Net [28]", + "bbox": [ + 212, + 767, + 784, + 840 + ], + "page_idx": 3 + }, + { + "type": "page_number", + "text": "4", + "bbox": [ + 217, + 114, + 228, + 126 + ], + "page_idx": 3 + }, + { + "type": "header", + "text": "G. Lu et al.", + "bbox": [ + 271, + 114, + 351, + 127 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "model, where embedding of time step $t$ is injected with adaptive normalization layers and embedding of input text $c$ is injected with cross-attention layers.", + "bbox": [ + 212, + 146, + 782, + 176 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "4 PanGu-Draw", + "text_level": 1, + "bbox": [ + 214, + 200, + 380, + 217 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "In this section, we first illustrate our resource-efficient 5B text-to-image generation model, trained with a time-decoupling training strategy and further enhanced with a prompt enhancement LLM. Then, we present our Coop-Diffusion algorithm for the cooperative integration of diverse pre-trained diffusion models, enabling multi-control and multi-resolution image generation.", + "bbox": [ + 212, + 234, + 787, + 309 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "4.1 Time-Decoupling Training Strategy", + "text_level": 1, + "bbox": [ + 214, + 332, + 555, + 349 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Enhancing data, training, and inference efficiency is vital for text-to-image models' practical use. Figure 1 shows two existing training strategies: (a) Cascaded Training, using three models to incrementally improve resolution, is data-efficient but triples training and inference time. (b) Resolution Boost Training starts at $512 \\times 512$ and then $1024 \\times 1024$ resolution, discarding lower resolution data and offering moderate efficiency with higher training costs and single-model inference across all timesteps. These approaches differ from our time-decoupling strategy, detailed below.", + "bbox": [ + 212, + 359, + 787, + 474 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Responding to the need for enhanced efficiencies, we draw inspiration from the denoising trajectory of diffusion processes, where initial denoising stages primarily shape the image's structural foundation, and later stages refine its textural complexity. With this insight, we introduce the Time-Decoupling Training Strategy. This approach divides a comprehensive text-to-image model, denoted as $\\epsilon_{\\theta}$ , into two specialized sub-models operating across different temporal intervals: a structure generator, $\\epsilon_{struct}$ , and a texture generator, $\\epsilon_{texture}$ . Each sub-model is half the size of the original, thus enhancing manageability and reducing computational load.", + "bbox": [ + 212, + 476, + 787, + 606 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "As illustrated in Figure 1(c), the structure generator, $\\epsilon_{\\text{struct}}$ , is responsible for early-stage denoising across larger time steps, specifically within the range $T, \\dots, T_{\\text{struct}}$ , where $0 < T_{\\text{struct}} < T$ . This stage focuses on establishing the foundational outlines of the image. Conversely, the texture generator, $\\epsilon_{\\text{texture}}$ , operates during the latter, smaller time steps, denoted by $T_{\\text{struct}}, \\dots, 0$ , to elaborate on the textural details. Each generator is trained in isolation, which not only alleviates the need for high-memory computation devices but also avoids the complexities associated with model sharding and its accompanying inter-machine communication overhead.", + "bbox": [ + 212, + 607, + 787, + 737 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "In the inference phase, $\\epsilon_{\\text{struct}}$ initially constructs a base structural image, $z_{T_{\\text{struct}}}$ , from an initial random noise vector, $z_T$ . Subsequently, $\\epsilon_{\\text{texture}}$ refines this base to enhance textural details, culminating in the final output, $z_0$ . This sequential processing facilitates a more resource-efficient workflow, significantly reducing the hardware footprint and expediting the generation process without compromising the model's performance or output quality, as demonstrated in our ablated experiment in Sec. 5.3.", + "bbox": [ + 212, + 738, + 787, + 839 + ], + "page_idx": 4 + }, + { + "type": "header", + "text": "PanGu-Draw", + "bbox": [ + 640, + 114, + 730, + 127 + ], + "page_idx": 4 + }, + { + "type": "page_number", + "text": "5", + "bbox": [ + 774, + 114, + 785, + 126 + ], + "page_idx": 4 + }, + { + "type": "image", + "img_path": "images/b26313a4b7eaefc2071c2ca00ab616296f64a56bceb23db74fe5e2378bf971e2.jpg", + "image_caption": [ + "Fig. 2: Visualization of our Coop-Diffusion algorithm for the cooperative integration of diverse pre-trained diffusion models. (a) Existing pre-trained diffusion models, each tailored for specific controls and operating within distinct latent spaces and image resolutions. (b) This sub-module bridges the gap arising from different latent spaces by transforming $\\epsilon_t'$ in latent space B to the target latent space A as $\\tilde{\\epsilon}_t$ . (c) This submodule bridges the gap arising from different resolutions by performing upsampling on the predicted clean data $\\hat{x}_{0,t}'$ ." + ], + "image_footnote": [], + "bbox": [ + 250, + 143, + 743, + 338 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Resource-Efficient Specialized Training Regime. We further adopt specialized training designs for the above two models. The structure generator $\\epsilon_{struct}$ , which derives image structures from text, requires training on an extensive dataset encompassing a wide range of concepts. Traditional methods, like Stable Diffusion, often eliminate low-resolution images, discarding about $48\\%$ of training data and thereby inflating dataset costs. Contrarily, we integrate high-resolution images with upscaled lower-resolution ones. This approach, as proven by our ablated experiments in Sec. 5.3, shows no performance drop, as the predicted $z_{T_{struct}}$ still contains substantial noise. In this way, we achieve higher data efficiency and avoid the problem of semantic degeneration.", + "bbox": [ + 212, + 474, + 787, + 619 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Since the image structure is determined in $z_{T_{\\text{struct}}}$ and the texture generator $\\epsilon_{\\text{texture}}$ focuses on refining texture, we propose training $\\epsilon_{\\text{texture}}$ at a lower resolution while still sampling at high resolution. This strategy, as demonstrated in our ablated experiments in Sec. 5.3, results in no performance drop and no structural problems (e.g., repetitive presentation [15]). Consequently, we achieved an overall $51\\%$ improvement in training efficiency. Figure 1 summarizes the data, training, and inference efficiency of different training strategies. Besides higher data and training efficiency, our strategy also achieves higher inference efficiency with fewer inference steps compared to the Cascaded Training strategy and a smaller per-step model compared to the Resolution Boost Training strategy.", + "bbox": [ + 212, + 621, + 787, + 767 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "4.2 Coop-Diffusion: Multi-Diffusion Fusion", + "text_level": 1, + "bbox": [ + 215, + 786, + 581, + 801 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "As shown in Figure 2(a), there are numerous pre-trained diffusion models, such as various SD, ControlNet, image variation, etc., each tailored for specific con", + "bbox": [ + 212, + 810, + 785, + 840 + ], + "page_idx": 5 + }, + { + "type": "page_number", + "text": "6", + "bbox": [ + 217, + 114, + 228, + 126 + ], + "page_idx": 5 + }, + { + "type": "header", + "text": "G. Lu et al.", + "bbox": [ + 271, + 114, + 351, + 126 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "trols and image resolutions. It is promising to fuse these pre-trained models for multi-control or multi-resolution image generation without needing to train a new model. However, the different latent spaces and resolutions of these models impede joint synthesis of images controlled by different models, thereby limiting their practical applications. In response to these challenges, we propose the Coop-Diffusion algorithm with two key sub-modules, as shown in Figures 2(b) and (c), to bridge the latent space gap and the resolution gap, and to unite the denoising process in the same space.", + "bbox": [ + 212, + 146, + 782, + 263 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Bridging the Latent Space Gap. To bridge the latent space gap between two different latent spaces A and B, we propose to unify the model prediction in latent space A by transforming the model prediction $\\epsilon_t^\\prime$ in latent space B to latent space A using the image space as an intermediate. This is done in the following way: first, we predict the clean data $\\hat{z}_{0,t}^{\\prime}$ using Equation (3) as:", + "bbox": [ + 212, + 263, + 784, + 333 + ], + "page_idx": 6 + }, + { + "type": "equation", + "text": "\n$$\n\\dot {z} _ {0, t} ^ {\\prime} = \\frac {1}{\\sqrt {\\bar {\\alpha} _ {t}}} (z _ {t} ^ {\\prime} - \\sqrt {1 - \\bar {\\alpha} _ {t}} \\epsilon_ {t} ^ {\\prime}), \\tag {4}\n$$\n", + "text_format": "latex", + "bbox": [ + 403, + 342, + 785, + 372 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "which is then decoded into a pixel-level image $\\hat{x}_{0,t}^{\\prime}$ using the latent decoder model $D^{\\prime}$ . This image is encoded into latent space A using the image encoder model $E$ , as $\\tilde{z}_{0,t} = E(\\hat{x}_{0,t}^{\\prime})$ , and finally transformed into a model prediction by inverting Equation (3) as:", + "bbox": [ + 214, + 378, + 782, + 439 + ], + "page_idx": 6 + }, + { + "type": "equation", + "text": "\n$$\n\\tilde {\\epsilon} _ {t} = \\frac {1}{\\sqrt {1 - \\bar {\\alpha} _ {t}}} \\left(z _ {t} - \\sqrt {\\bar {\\alpha} _ {t}} \\tilde {z} _ {0, t}\\right). \\tag {5}\n$$\n", + "text_format": "latex", + "bbox": [ + 403, + 446, + 785, + 477 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "With the united $\\tilde{\\epsilon}_t$ , we can now perform multi-control fusion between $\\tilde{\\epsilon}_t$ and $\\epsilon_t$ (the prediction from model $\\epsilon_{\\theta}$ with $z_{t}$ in latent space A, omitted in Figure 2 for brevity). In this paper, we adopt the fusion method proposed in Composable-Diffusion [19] as: $\\epsilon_{t,fuse} = d\\cdot \\tilde{\\epsilon}_t + (1 - d)\\cdot \\epsilon_t$ , where $d$ and $1 - d$ are the guidance strengths of each model with $d\\in [0,1]$ , to guide the denoising process jointly with these two models for multi-control image generation. Algorithm 1 further illustrates this unification and fusion process.", + "bbox": [ + 214, + 483, + 782, + 585 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Bridging Resolution Gap. To integrate the denoising processes of a low-resolution model with a high-resolution model, upsampling and/or downsam", + "bbox": [ + 214, + 587, + 782, + 617 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/41797f8253c6970c94b7ba047045df5204348de442f8bf20547e0ab89ec42daf.jpg", + "image_caption": [ + "(a) Upsampling from intermediate $z_{t}$ ." + ], + "image_footnote": [], + "bbox": [ + 330, + 648, + 493, + 771 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/f97620877ba2582d5e0440664a4d1cb1668cfa21e2e9ec676fb290d9c023b68d.jpg", + "image_caption": [ + "(b) Our upsampling algorithm.", + "Fig. 3: Results of fusing a low-resolution model and a high-resolution model with different upsampling methods. Upsampling from intermediate $z_{t}$ results in severe artifacts, while our upsampling algorithm results in high-fidelity image." + ], + "image_footnote": [], + "bbox": [ + 508, + 648, + 671, + 772 + ], + "page_idx": 6 + }, + { + "type": "header", + "text": "PanGu-Draw", + "bbox": [ + 640, + 114, + 730, + 126 + ], + "page_idx": 6 + }, + { + "type": "page_number", + "text": "7", + "bbox": [ + 774, + 114, + 784, + 125 + ], + "page_idx": 6 + }, + { + "type": "code", + "sub_type": "algorithm", + "code_caption": [ + "Algorithm 1 Coop-Diffusion: Multi-Diffusion Fusing" + ], + "code_body": "Sub-Module 1. Bridging Latent Space Gap \nInput: random noise $z_{T} \\sim \\mathcal{N}(0, I)$ , diffusion model $\\epsilon_{\\theta}$ , decoder $D$ , encoder $E$ in latent space A; random noise $z_{T}' = z_{T}$ , diffusion model $\\epsilon_{\\theta}'$ , decoder $D'$ , encoder $E'$ in latent space B; guidance strength $d$ , sampling method $S$ . \n1: for $t = T, \\ldots, 1$ do \n2: $\\epsilon_t' = \\epsilon_\\theta'(z_t')$ , $\\hat{z}_{0,t}' = \\frac{1}{\\sqrt{\\alpha_t}} (z_t' - \\sqrt{1 - \\bar{\\alpha}_t} \\epsilon_t')$ \n3: $\\hat{x}_{0,t}' = D'(\\hat{z}_{0,t}')$ , $\\tilde{z}_{0,t} = E(\\hat{x}_{0,t}')$ \n4: $\\tilde{\\epsilon}_t = \\frac{1}{\\sqrt{1 - \\bar{\\alpha}_t}} (z_t - \\sqrt{\\bar{\\alpha}_t} \\tilde{z}_{0,t})$ \n5: $\\epsilon_t = \\epsilon_\\theta(z_t)$ , $\\epsilon_{t,fuse} = d \\cdot \\tilde{\\epsilon}_t + (1 - d) \\cdot \\epsilon_t$ \n6: $z_{t-1} = S(z_t, t, \\epsilon_{t,fuse})$ \n7: $z_{t-1}' = S(z_t', t, \\epsilon_{t,fuse}') \\triangleright \\epsilon_{t,fuse}'$ from $\\epsilon_{t,fuse}'$ similar to the process from $\\epsilon_t'$ to $\\tilde{\\epsilon}_t$ , omitted for brevity \n8: end for \n9: return $D(z_0)$ \nSub-Module 2. Bridging Resolution Gap \nInput: diffusion model $\\epsilon_{\\theta}$ , decoder $D$ , encoder $E$ in high-resolution space; random noise $z_T' \\sim \\mathcal{N}(0, I)$ , diffusion model $\\epsilon_{\\theta}'$ , decoder $D'$ , encoder $E'$ in low-resolution space; low-resolution sampling end step $T_{low}$ , sampling method $S$ . \n1: for $t = T, \\ldots, T_{low} + 1$ do \n2: $\\epsilon_t' = \\epsilon_\\theta'(z_t')$ , $z_{t-1}' = S(z_t', t, \\epsilon_t')$ \n3: end for \n4: $\\hat{z}_{0,T_{low}}' = \\frac{1}{\\sqrt{\\bar{\\alpha}_{T_{low}}}} (z_{T_{low}}' - \\sqrt{1 - \\bar{\\alpha}_{T_{low}}} \\epsilon_{T_{low}}')$ \n5: $\\hat{x}_{0,T_{low}}' = D'(\\hat{z}_{0,T_{low}}')$ , $\\hat{x}_{0,T_{low}} = Upsample(\\hat{x}_{0,T_{low}}')$ \n6: $\\hat{z}_{0,T_{low}} = E(\\hat{x}_{0,T_{low}})$ \n7: $z_{T_{low}} = \\sqrt{\\bar{\\alpha}_{T_{low}}} \\hat{z}_{0,T_{low}} + \\sqrt{1 - \\bar{\\alpha}_{T_{low}}} \\epsilon, \\epsilon \\sim \\mathcal{N}(0, I)$ \n8: for $t = T_{low}, \\ldots, 1$ do \n9: $\\epsilon_t = \\epsilon_\\theta(z_t)$ , $z_{t-1} = S(z_t, t, \\epsilon_t)$ \n10: end for \n11: return $D(z_0)$", + "bbox": [ + 215, + 162, + 785, + 542 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "pling is necessary. Traditional bilinear upsampling, often applied to the intermediate result $z_{t}$ during the denoising process, can undesirably amplify pixel correlation. This amplification deviates from the initial Independent and Identically Distributed (IID) assumption, leading to severe artifacts in the final images, as shown in Figure 3(a). Conversely, downsampling does not present this issue. To address the IID issue in upsampling, we propose a new upsampling algorithm that preserves the IID assumption, thereby bridging the resolution gap between models with different pre-trained resolutions.", + "bbox": [ + 212, + 575, + 787, + 691 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Figure 2(c) visualizes our upsampling algorithm. Specifically, for a low-resolution $z_{t}^{\\prime}$ , we use the image space as an intermediate space to transform $z_{t}^{\\prime}$ in low-resolution space into high-resolution space as $\\tilde{z}_{t}$ . We first predict the noise $\\epsilon_{t}^{\\prime}$ with the denoising model $\\epsilon_{\\theta}^{\\prime}$ and then predict the clean data $\\hat{z}_{0,t}^{\\prime}$ as described in Eq. 4. This is decoded into an image $\\hat{x}_{0,t}^{\\prime}$ using decoder $D^{\\prime}$ . We then perform upsampling on $\\hat{x}_{0,t}^{\\prime}$ to obtain its high-resolution counterpart $\\hat{x}_{0,t}$ . Finally, $\\hat{x}_{0,t}$ is encoded into the latent space with encoder $E$ as $\\hat{z}_{0,t}$ , and $t$ -step noise is added to get the final result $\\tilde{z}_{t}$ using Eq. 1.", + "bbox": [ + 212, + 693, + 805, + 809 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "With the unified $\\tilde{z}_t$ , we can now perform multi-resolution fusion. First, we denoise with a low-resolution model to obtain the intermediate $z_t'$ and its high-", + "bbox": [ + 212, + 810, + 785, + 840 + ], + "page_idx": 7 + }, + { + "type": "page_number", + "text": "8", + "bbox": [ + 217, + 114, + 228, + 126 + ], + "page_idx": 7 + }, + { + "type": "header", + "text": "G. Lu et al.", + "bbox": [ + 271, + 114, + 351, + 127 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "resolution counterpart $\\tilde{z}_t$ . Then, we perform denoising with a high-resolution model starting from $\\tilde{z}_t$ , and vice versa. This approach allows us to conduct one-stage super-resolution without undergoing all the low-resolution denoising steps, thereby improving inference efficiency. Algorithm 1 further illustrates this unification and fusion process.", + "bbox": [ + 212, + 146, + 787, + 220 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "5 Experiments", + "text_level": 1, + "bbox": [ + 214, + 244, + 375, + 263 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "Implementation Details. We adopt the pretrained Variational Autoencoder (VAE) model from SDXL [22], and we build our structure and texture generator based on the architecture of its U-Net model with the following modifications. To achieve bilingual text-to-image generation (Chinese and English), we pretrain a Chinese text encoder [9, 36] on our Chinese training dataset. We then concatenate the text embeddings from this Chinese text encoder with those from a pretrained English text encoder, serving as the final text embeddings for the denoising models. For multi-resolution image generation, we select a range of image resolutions around $1024 \\times 1024$ and further condition the denoising model on the sinusoidal positional embeddings corresponding to the index of image resolutions. The $T_{struct}$ parameter is set to 500, as suggested by our ablation study.", + "bbox": [ + 212, + 279, + 787, + 439 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "Dataset Construction. To encompass the abundant concepts in the world, we collect images in various styles from multiple sources, including Noah-Wukong [9], LAION [27], and others, such as photography, cartoons, portraits, and gaming assets. The collected images are filtered based on CLIP score, aesthetic score, watermark presence, resolution, and aspect ratio. To improve the semantic alignment of PanGu-Draw, we discard parts of the noisy captions that are meaningless or mismatched to the image, sourced from the Internet. Instead, we recapture the collected images by first employing an open-vocabulary detector [35] to locate the primary subjects within the images. These subjects are then processed by LLaVA [18], a high-performance vision-language model, along with prompting templates, to yield detailed image descriptions. These English annotations are subsequently translated into Chinese.", + "bbox": [ + 212, + 439, + 787, + 613 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "Evaluation Metrics. We evaluate PanGu-Draw's text-to-image generation on COCO [17] with 30k images for English, and COCO-CN [16] with 10k images for Chinese. The Frechet Inception Distance (FID [10]) is utilized to evaluate image quality and diversity. For Chinese, additional metrics include the Inception Score (IS [30]) and CN-CLIP-score [34], assessing image quality and text-image alignment. Besides, a user study is conducted to evaluate image-text alignment, fidelity, and aesthetics using ImageEval-prompt1 across 339 prompts.", + "bbox": [ + 212, + 614, + 787, + 717 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "5.1 Text-to-Image Generation", + "text_level": 1, + "bbox": [ + 214, + 741, + 478, + 757 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "Evaluation on COCO. As shown in Table 1, PanGu-Draw achieves a FID of 7.99, which is superior to compared methods such as DALL-E 2 and SDXL. It also achieves competitive FID with SOTA methods, indicating the effectiveness", + "bbox": [ + 212, + 768, + 787, + 814 + ], + "page_idx": 8 + }, + { + "type": "header", + "text": "PanGu-Draw", + "bbox": [ + 640, + 114, + 730, + 127 + ], + "page_idx": 8 + }, + { + "type": "page_number", + "text": "9", + "bbox": [ + 774, + 114, + 785, + 126 + ], + "page_idx": 8 + }, + { + "type": "page_footnote", + "text": "1 https://github.com/FlagOpen/FlagEval/tree/master/imageEval", + "bbox": [ + 217, + 824, + 661, + 839 + ], + "page_idx": 8 + }, + { + "type": "table", + "img_path": "images/294bd8c65fd7edcdbb520777904f3bf5721646004ff33194c243f1ce475ae47d.jpg", + "table_caption": [ + "Table 1: Comparisons of PanGu-Draw with recent representative English text-to-image generation models on COCO dataset in terms of FID." + ], + "table_footnote": [], + "table_body": "
MethodFID↓Model SizeRelease
DALL-E [25]27.5012BN
LDM [26]12.631.5BY
GLIDE [20]12.245BN
SDXL [22]11.932.5BY
PixArt-α [4]10.650.6BY
DALL-E 2 [24]10.395.5BN
Imagen [29]7.273BN
RAPHAEL [33]6.613BN
PanGu-Draw7.995BY
", + "bbox": [ + 348, + 184, + 648, + 340 + ], + "page_idx": 9 + }, + { + "type": "table", + "img_path": "images/b0e7a9fde7a6900a3624779614277c5e2b276cc3d422882f206bbb7dc1b934f1.jpg", + "table_caption": [ + "Table 2: Comparisons of PanGu-Draw with Chinese text-to-image generation models on COCO-CN dataset in terms of FID, IS and CN-CLIP-score. The classifier-free guidance scales are set as 9 following AltDiffusion [37]." + ], + "table_footnote": [], + "table_body": "
ModelFID↓IS↑CN-CLIP-score↑
AltDiffusion [37]25.3129.1635.12
Taiyi-Bilingual [38]24.6134.2932.26
Taiyi-CN [38]23.9934.2934.22
PanGu-Draw21.8137.0036.62
", + "bbox": [ + 318, + 411, + 679, + 493 + ], + "page_idx": 9 + }, + { + "type": "table", + "img_path": "images/5ab0b0f1b7d8a1abdd0bb83276f1b8bb868a235d06def14a1a3d034a3697521b.jpg", + "table_caption": [ + "Table 3: Results of a User study on ImageVal-prompt in terms of image-text alignment, image fidelity, and aesthetics." + ], + "table_footnote": [], + "table_body": "
MethodAlign↑Fidelity↑Aesthetics↑Ave↑
DALL-E 3 [2]4.724.594.764.69
MJ 5.24.634.544.754.64
SDXL [22]4.414.374.594.46
SD [26]4.173.994.204.12
PanGu-Draw4.54.524.724.58
", + "bbox": [ + 325, + 547, + 671, + 645 + ], + "page_idx": 9 + }, + { + "type": "text", + "text": "of our time-decoupling training strategy and its outstanding data and training efficiencies. Our 5B PanGu model is the best-released model in terms of FID.", + "bbox": [ + 212, + 676, + 784, + 704 + ], + "page_idx": 9 + }, + { + "type": "text", + "text": "Evaluation on COCO-CN. As shown in Table 2, PanGu-Draw outperforms other released Chinese text-to-image models, including Taiyi-CN, Taiyi-Bilingual, and AltDiffusion, across all three metrics. This performance highlights PanGu-Draw's exceptional Chinese text-to-image generation capabilities and the effectiveness of our bilingual text encoder architecture.", + "bbox": [ + 212, + 707, + 785, + 779 + ], + "page_idx": 9 + }, + { + "type": "text", + "text": "User Study. We conducted a user study to compare PanGu-Draw with top-performing methods, including SDXL [22], Midjourney 5.2, and DALL-E 3 [2]. As shown in Table 3, PanGu-Draw achieves better results than SD and SDXL across all three metrics. It also attains approximately $99\\% / 98\\%$ of the performance of", + "bbox": [ + 212, + 781, + 785, + 839 + ], + "page_idx": 9 + }, + { + "type": "page_number", + "text": "10", + "bbox": [ + 217, + 114, + 235, + 126 + ], + "page_idx": 9 + }, + { + "type": "header", + "text": "G. Lu et al.", + "bbox": [ + 271, + 114, + 351, + 126 + ], + "page_idx": 9 + }, + { + "type": "image", + "img_path": "images/2201a44d5e082f7f387e156237a94b6d62479a7b00fd67e29d751a18ac5fa91f.jpg", + "image_caption": [ + "一位老年女性，身着优雅西服，佩戴菲黛丝帽，身穿夹紧对联鞋。伸出手臂，她的背景是夕阳的校园：“一个 young woman, wearing a elegant gown and traditional dress, carrying a heavy bag of luggage and arms, with the sunset on campus in the background.”" + ], + "image_footnote": [], + "bbox": [ + 243, + 143, + 370, + 239 + ], + "page_idx": 10 + }, + { + "type": "image", + "img_path": "images/88a7bac765b94cc8235d430ba29e851f16ba26e1044b25bd561a6550234ae2da.jpg", + "image_caption": [ + "一面像微笑的女子，身穿白色针织，红色长衣在阳光下翩翩生舞，面庞清新，风姿绰约，细尘一片。“A smiling woman wearing a white T-shirt and a red blouse with a black skirt, a red blouse, a red blouse, a red blouse, a red blouse, a red blouse, a red blouse, a red blouse, a red blouse, a red blouse, a red blouse, a red blouse, a red blouse, a red blouse, a red blouse, a red blouse, a red blouse, a red blouse, a red blouse, a red blouse, a red blouse, a red blouse, a red blouse, a red blouse, a red blouse, a red blouse, a red dress, a red dress, a red dress, a red dress, a red dress, a red dress, a red dress, a red dress, a red dress, a red dress, a red dress, a red dress, a red dress, a red dress" + ], + "image_footnote": [], + "bbox": [ + 374, + 143, + 500, + 241 + ], + "page_idx": 10 + }, + { + "type": "image", + "img_path": "images/7fdf2a7a3ac4de0e0c409df4c9fd0d2f94eb60f9d4229340e33de5f392f93325.jpg", + "image_caption": [ + "“一个巨大的水球体，内部蕴含着一个微型雨林，雨林中翅膀翱翔，阳光透过树洒洒洒。”“Auge crystal ball contains a miniature rainforest inside, with butterflies flying in the air, and sunlight shining through the leaves.”" + ], + "image_footnote": [], + "bbox": [ + 501, + 143, + 627, + 241 + ], + "page_idx": 10 + }, + { + "type": "image", + "img_path": "images/9d8b18b9bb763e9e709a4d4d81227bed8a74d9800437ae841b1863b2805558f8.jpg", + "image_caption": [ + "“一只穿着中世纪短发的兔子，手长颈尖站在一座古董城的城墙上，背后是满目的杂草。” \n\"A rabbit wearing medieval armor and holding a sword stands on the wall of the tower, with a blade with the setting sun behind him.\"" + ], + "image_footnote": [], + "bbox": [ + 630, + 143, + 756, + 241 + ], + "page_idx": 10 + }, + { + "type": "image", + "img_path": "images/2eeafd652288659ce2f5e1d7c7e8f27eaff1ea04694c96eabb732f80ab7d24f4.jpg", + "image_caption": [ + "“赛博办风面摄影的抓机人，依托在夜空飞行平安，以给予水面文明风貌，具有强烈的光影效果。在2019年北京冬奥会上，运动员们穿梭在的night sky, presented in particle ink painting style, with strong light and shadow effects.\"" + ], + "image_footnote": [], + "bbox": [ + 243, + 273, + 370, + 465 + ], + "page_idx": 10 + }, + { + "type": "image", + "img_path": "images/81b252e3489c342a012d7293b1eaedf5056fa2a236ccb9593b24496e0e236f0a.jpg", + "image_caption": [ + "“旅行者们正奔在光激染湖的湖面上，周围是雄伟的山脉，以中国水墨画风格描绘，画面也彩淡雅致，具有古朴诗志。”（1980年） \n“旅行者们正奔在湖面上，周围是美丽的山林，以自然、静谧、清新、宁静、和谐、庄严、庄严、庄严、庄严、庄严、庄严、庄严、庄严、庄严、庄严、庄严、庄严、庄严、庄严、庄严、庄严、庄严、庄严、庄严、庄严、庄严、庄严、庄严、庄严、庄严、庄严、庄严、庄严、庄严、庄严、庄严、庄严、庄严、庄严、庄严、庄严、庄严、庄严、庄严、庄严、庄严、庄严、庄严、庄严、庄严、庄严、庄严、庄严、庄严、庄严、庄重的白色。" + ], + "image_footnote": [], + "bbox": [ + 375, + 270, + 627, + 364 + ], + "page_idx": 10 + }, + { + "type": "image", + "img_path": "images/78246ef7da538a87b6398557cb88e04794f8bdc28b5f9a41f07772874ce19e1e.jpg", + "image_caption": [ + "“一台未来风格的摩托车，闪耀着霓虹灯，停在夜晚的东京街头。” \n“A futuristic motorcycle, shining with neon lights, is parked on the streets of Tokyo at night.”", + "Fig. 4: Images generated with PanGu-Draw, our 5B multi-lingual text-to-image generation model. PanGu-Draw is able to generate multi-resolution high-fidelity images semantically aligned with the input prompts." + ], + "image_footnote": [], + "bbox": [ + 374, + 383, + 500, + 482 + ], + "page_idx": 10 + }, + { + "type": "image", + "img_path": "images/dda03bb00997887131d6f8ce25a19e7cfc341b1dae790a6de919503eabdffd62.jpg", + "image_caption": [ + "一座由冰晶和雪花构成的精致城堡，坐落在在北极的冰原上。 \n“An exquisite castle made of ice crystals and snowflakes, located on the Arctic ice sheet.”" + ], + "image_footnote": [], + "bbox": [ + 503, + 383, + 627, + 482 + ], + "page_idx": 10 + }, + { + "type": "image", + "img_path": "images/4e5071405aae96c871b8e94442ea96f379d0f4d60c8ce883e167b9323215585a.jpg", + "image_caption": [ + "一袋老火的糖盔，完全由糖果和巧克力制成。一个 ancient pot was filled of candy and chocolate." + ], + "image_footnote": [], + "bbox": [ + 630, + 275, + 756, + 465 + ], + "page_idx": 10 + }, + { + "type": "text", + "text": "Midjourney 5.2 and DALL-E 3, respectively, indicating PanGu-Draw's excellent text-to-image capabilities. Figure 4 shows some high-fidelity multi-resolution images generated by PanGu-Draw. As we can see, the generated images of PanGu-Draw are of high aesthetics and semantically aligned with the input prompts.", + "bbox": [ + 214, + 583, + 787, + 643 + ], + "page_idx": 10 + }, + { + "type": "text", + "text": "5.2 Multi-Diffusion Fusing Results", + "text_level": 1, + "bbox": [ + 215, + 662, + 516, + 678 + ], + "page_idx": 10 + }, + { + "type": "text", + "text": "Multi-Control Image Generation. To demonstrate the effectiveness of the proposed reusable multi-diffusion fusing algorithm, Coop-Diffusion, we first present multiple results of multi-control image generation. Figure 5 displays results from fusing an image variation model $^2$ with PanGu-Draw. The fusing results maintain a style similar to that of the reference image, matching the texture described by the input prompt. Figure 6 shows results from fusing PanGu-Draw with a pose/edge-to-image ControlNet model, which operates in guess mode without input prompts. Here, the fusing results combine the structure of the pose/edge image with the texture described by the input prompt.", + "bbox": [ + 212, + 686, + 799, + 818 + ], + "page_idx": 10 + }, + { + "type": "header", + "text": "PanGu-Draw", + "bbox": [ + 640, + 114, + 730, + 126 + ], + "page_idx": 10 + }, + { + "type": "page_number", + "text": "11", + "bbox": [ + 767, + 114, + 784, + 126 + ], + "page_idx": 10 + }, + { + "type": "page_footnote", + "text": "$^{2}$ https://huggingface.co/lambdalabs/sd-image-variations-diffusers", + "bbox": [ + 217, + 823, + 661, + 840 + ], + "page_idx": 10 + }, + { + "type": "image", + "img_path": "images/7adbcae4a56790853f9fae971c5ffb910d46fb82f33d32a5639812f891912d04.jpg", + "image_caption": [ + "(a) Input image and generation results of image variation." + ], + "image_footnote": [], + "bbox": [ + 302, + 143, + 380, + 205 + ], + "page_idx": 11 + }, + { + "type": "image", + "img_path": "images/0f0e5f856510b0ed5aaff8f431bc6d24ad8fbe5d7825e09f232bfdbb36637d71.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 383, + 143, + 537, + 205 + ], + "page_idx": 11 + }, + { + "type": "image", + "img_path": "images/bc637749dfbb0cca22b06351b7c54cd300f3577a43dbb47c3f8f617e1b5a7733.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 315, + 214, + 367, + 255 + ], + "page_idx": 11 + }, + { + "type": "image", + "img_path": "images/c4d7506482d107c66fda628776989c67a3519d04d2f6efd960a1a096806b99f6.jpg", + "image_caption": [ + "(b) Input prompt and generation results of PanGu-Draw." + ], + "image_footnote": [], + "bbox": [ + 383, + 217, + 537, + 277 + ], + "page_idx": 11 + }, + { + "type": "image", + "img_path": "images/7f225c96c36d5735ec6785e7c15c7612048b2fb49e3abc9830d9ad078d997480.jpg", + "image_caption": [ + "(c) Generation results of fusing image variation and PanGu-Draw with Coop-Diffusion algorithm." + ], + "image_footnote": [], + "bbox": [ + 545, + 143, + 700, + 263 + ], + "page_idx": 11 + }, + { + "type": "image", + "img_path": "images/5b2abb2a30c0943316d2a6728ea7d6f55a1a4ccaa7b2c866e9cf5d94c53d915f.jpg", + "image_caption": [ + "Fig. 5: Generation results of the fusing of an image variation model and PanGu-Draw and with the proposed Coop-Diffusion algorithm." + ], + "image_footnote": [], + "bbox": [ + 308, + 358, + 385, + 419 + ], + "page_idx": 11 + }, + { + "type": "image", + "img_path": "images/8417a5359dcaf7329a496568c9e1950911c95755f385dfc9f541411a150d953a.jpg", + "image_caption": [ + "(b) Input prompt and generation results of PanGu-Draw." + ], + "image_footnote": [], + "bbox": [ + 300, + 420, + 517, + 489 + ], + "page_idx": 11 + }, + { + "type": "image", + "img_path": "images/edb622cc78a1a3d9f5df08900d8b98db1466a4cb1981153de1d85351f8f86a99.jpg", + "image_caption": [ + "(a) Input edge map and generation results of edge-to-image" + ], + "image_footnote": [], + "bbox": [ + 308, + 510, + 385, + 570 + ], + "page_idx": 11 + }, + { + "type": "image", + "img_path": "images/92dbdb4fb327017281fde2c184adfe6856c0d6a88d7633492ae3453cc52a8b39.jpg", + "image_caption": [ + "(b) Input prompt and generation results of PanGu-Draw.", + "Fig. 6: Generation results guided by fusing signals of text and pose/edge map by our Coop-Diffusion." + ], + "image_footnote": [], + "bbox": [ + 310, + 582, + 537, + 642 + ], + "page_idx": 11 + }, + { + "type": "image", + "img_path": "images/9e7f9916bc96a6607b1de3f6b2ce885429ecfef8dd4b009046a9cac1d3d183f8.jpg", + "image_caption": [ + "(c) Generation results of fusing pose-to-image and PanGu-Draw with Coop-Diffusion algorithm." + ], + "image_footnote": [], + "bbox": [ + 547, + 358, + 699, + 474 + ], + "page_idx": 11 + }, + { + "type": "image", + "img_path": "images/1eec92eebddea7cf18c9ea5ccfb8d7f32b866c6cc03741855dd723db25d1d1f1.jpg", + "image_caption": [ + "(c) Generation results of fusing edge-to-image and PanGu-Draw with CoopDiffusion algorithm." + ], + "image_footnote": [], + "bbox": [ + 547, + 510, + 697, + 627 + ], + "page_idx": 11 + }, + { + "type": "text", + "text": "Multi-Resolution Image Generation. We also present multi-resolution image generation results of fusing PanGu-Draw with low-resolution text-to-image and edge-to-image ControlNet model by first denoising with the low-resolution model to get the intermediate $z_{t}^{\\prime}$ and the high-resolution counterpart $\\tilde{z}_{t}$ , and then perform denoising in high resolution with PanGu-Draw. Figure 7 shows the results from the low-resolution model and our fusing algorithm Coop-Diffusion. As we can see, PanGu-Draw adds much details to the low-resolution predictions", + "bbox": [ + 212, + 737, + 787, + 840 + ], + "page_idx": 11 + }, + { + "type": "page_number", + "text": "12", + "bbox": [ + 217, + 114, + 235, + 126 + ], + "page_idx": 11 + }, + { + "type": "header", + "text": "G. Lu et al.", + "bbox": [ + 271, + 114, + 351, + 126 + ], + "page_idx": 11 + }, + { + "type": "image", + "img_path": "images/e0bcc639f1e4855ccf61fed739fa729c95380f9dcad585e1c9d243b10c6105c5.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 357, + 143, + 500, + 253 + ], + "page_idx": 12 + }, + { + "type": "image", + "img_path": "images/e0bc84a0d5e802d8500382c3c483ea9d1359513aab38ab16f640ff6ccef25bab.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 501, + 143, + 643, + 253 + ], + "page_idx": 12 + }, + { + "type": "image", + "img_path": "images/96ae32efe5cb26a9e555f8471c91d3dcc1eac5e8dfe36b104dec01b847a1951d.jpg", + "image_caption": [ + "LR model" + ], + "image_footnote": [], + "bbox": [ + 357, + 255, + 500, + 364 + ], + "page_idx": 12 + }, + { + "type": "image", + "img_path": "images/c50b523831a80171738df1ab3f8872dbaaf9291deeff5220f4380bcc71a9afb8.jpg", + "image_caption": [ + "LR model + HR model", + "Fig. 7: Images generated with a low-resolution (LR) model (1st row: T2I model; 2nd row: edge-to-image ControlNet) and the fusion of the LR model and HR PanGu-Draw with our Coop-Diffusion. The resolutions are $512 \\times 512$ and $1024 \\times 1024$ respectively. This allows for single-stage super-resolution for better details and higher inference efficiency." + ], + "image_footnote": [], + "bbox": [ + 501, + 253, + 643, + 364 + ], + "page_idx": 12 + }, + { + "type": "text", + "text": "leading to high-fidelity high-resolution results. Besides, compared with the common practice of super-resolution with diffusion model, which carries out all the low-resolution denoising steps, our method achieve higher inference efficiency.", + "bbox": [ + 214, + 484, + 787, + 529 + ], + "page_idx": 12 + }, + { + "type": "text", + "text": "5.3 Ablation Study", + "text_level": 1, + "bbox": [ + 215, + 551, + 388, + 566 + ], + "page_idx": 12 + }, + { + "type": "text", + "text": "In this section, we perform ablation studies to analyze our time-decoupling training strategy. The baseline model has $1B$ parameters while the structure and texture generators both have $0.5B$ parameters. During the training process, the latter two models only train half the steps of the baseline model with $T_{struct}$ set as 500. Both settings of the models are trained from scratch on a subset of the LAION dataset containing images with all sizes. After training, FID, IS and CLIP-score on COCO are reported for comparison.", + "bbox": [ + 212, + 577, + 787, + 679 + ], + "page_idx": 12 + }, + { + "type": "text", + "text": "Time-Decoupling Training Strategy. We compare the performance of models trained with the Resolution Boost strategy and our time-decoupling strategy in Table 4. We found that models trained with our strategy achieves better performance in all three criteria, indicating the effectiveness of our strategy.", + "bbox": [ + 212, + 680, + 787, + 737 + ], + "page_idx": 12 + }, + { + "type": "text", + "text": "Training Designs. The structure and texture generators ( $\\epsilon_{\\mathrm{struct}}$ and $\\epsilon_{\\mathrm{texture}}$ ) are designed to train on different resolutions to improve data and training efficiency. However, this approach may negatively influence the final performance. In Table 5, we compare such a design with a traditional training process, where $\\epsilon_{\\mathrm{struct}}$ discards low-resolution images, or $\\epsilon_{\\mathrm{texture}}$ trains with high resolution. Results on COCO show that $\\epsilon_{\\mathrm{struct}}$ benefits from these extra up-scaled data, and $\\epsilon_{\\mathrm{texture}}$ learns enough texture patterns at a smaller resolution.", + "bbox": [ + 212, + 738, + 787, + 839 + ], + "page_idx": 12 + }, + { + "type": "header", + "text": "PanGu-Draw", + "bbox": [ + 640, + 114, + 730, + 126 + ], + "page_idx": 12 + }, + { + "type": "page_number", + "text": "13", + "bbox": [ + 767, + 114, + 785, + 126 + ], + "page_idx": 12 + }, + { + "type": "table", + "img_path": "images/3d98fbda8ae48c71611ed7afff6a223c71d31d671095a9e250b6ef9f25afe147.jpg", + "table_caption": [ + "Table 4: Comparison of models across Resolution Boost (1B parameters) and Time-Decoupling training strategies (0.5B parameters for structure and texture generators)" + ], + "table_footnote": [], + "table_body": "
ModelFID↓IS↑CLIP-score↑
Resolution Boost106.1210.4622.9
Time-Decoupling87.6611.0723.4
", + "bbox": [ + 338, + 184, + 658, + 234 + ], + "page_idx": 13 + }, + { + "type": "table", + "img_path": "images/aff72fa8c6a45c239a4237d4b257d8bcd78f46b95c6b92c7dd055b92d5771693.jpg", + "table_caption": [ + "Table 5: Performance of structure and texture models training with images of different resolutions." + ], + "table_footnote": [], + "table_body": "
Structure DataTexture ResolutionFID↓IS↑CLIP-score↑
All data25687.6611.0723.4
Only high resolution25689.5210.9623.2
All data51290.9810.5923.3
", + "bbox": [ + 256, + 287, + 741, + 357 + ], + "page_idx": 13 + }, + { + "type": "table", + "img_path": "images/53258af43e3eb3bc47428d38168c69683774adc2576278989d320f3d7175d18e.jpg", + "table_caption": [ + "Table 6: Comparisons of PanGu-Draw inference performance with different time step splitting point $T_{struct}$ settings." + ], + "table_footnote": [], + "table_body": "
TstructFID↓IS↑CLIP-score↑
200105.0810.5922.98
30098.0810.7223.12
50087.6611.0723.40
70089.4811.0223.32
", + "bbox": [ + 375, + 410, + 620, + 491 + ], + "page_idx": 13 + }, + { + "type": "text", + "text": "Timestep Splitting Point. The timestep splitting point $T_{\\mathrm{struct}}$ between the structure and texture generators also influences the final performance. To this end, we set $T_{\\mathrm{struct}}$ to 200, 300, 500, and 700, while keeping the other settings of the structure and texture generators unchanged. As shown in Table 6, as $T_{\\mathrm{struct}}$ increases from 200 to 700, the performance initially increases and then decreases continuously. $T_{\\mathrm{struct}} = 500$ is the optimal value, and we adopt it as the default setting in all other experiments.", + "bbox": [ + 212, + 518, + 784, + 619 + ], + "page_idx": 13 + }, + { + "type": "text", + "text": "6 Conclusion", + "text_level": 1, + "bbox": [ + 215, + 643, + 357, + 659 + ], + "page_idx": 13 + }, + { + "type": "text", + "text": "In this paper, we present \"PanGu-Draw\", a new latent diffusion model for efficient text-to-image generation that effectively integrates multiple control signals. Our approach includes a Time-Decoupling Training Strategy to separate the text-to-image process into structure and texture generation, enhancing data use and computational efficiency. Additionally, \"Coop-Diffusion\" is introduced, an algorithm allowing cooperative use of different pre-trained diffusion models in a unified denoising process for multi-control image synthesis at various resolutions without extra data or retraining. PanGu-Draw outperforms models like DALL-E 2 and SDXL in English T2I, achieves superior FID, IS, and CN-CLIP-scores in Chinese T2I, and receives favorable user feedback. This positions PanGu-Draw as a versatile and efficient state-of-the-art method.", + "bbox": [ + 212, + 674, + 785, + 834 + ], + "page_idx": 13 + }, + { + "type": "page_number", + "text": "14", + "bbox": [ + 217, + 114, + 235, + 126 + ], + "page_idx": 13 + }, + { + "type": "header", + "text": "G. Lu et al.", + "bbox": [ + 271, + 114, + 351, + 126 + ], + "page_idx": 13 + }, + { + "type": "text", + "text": "References", + "text_level": 1, + "bbox": [ + 217, + 143, + 321, + 159 + ], + "page_idx": 14 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "1. Balaji, Y., Nah, S., Huang, X., Vahdat, A., Song, J., Kreis, K., Aittala, M., Aila, T., Laine, S., Catanzaro, B., et al.: edifi: Text-to-image diffusion models with an ensemble of expert denoisers. arXiv preprint arXiv:2211.01324 (2022)", + "2. Betker, J., Goh, G., Jing, L., Brooks, T., Wang, J., Li, L., Ouyang, L., Zhuang, J., Lee, J., Guo, Y., et al.: Improving image generation with better captions. Computer Science. https://cdn.opena.com/papers/dall-e-3.pdf (2023)", + "3. Chen, C., Yin, Y., Shang, L., Jiang, X., Qin, Y., Wang, F., Wang, Z., Chen, X., Liu, Z., Liu, Q.: bert2bert: Towards reusable pretrained language models. arXiv preprint arXiv:2110.07143 (2021)", + "4. Chen, J., Yu, J., Ge, C., Yao, L., Xie, E., Wu, Y., Wang, Z., Kwok, J., Luo, P., Lu, H., Li, Z.: Fast training of diffusion transformer for photorealistic text-to-image synthesis (2023)", + "5. Dhariwal, P., Nichol, A.: Diffusion models beat gans on image synthesis. Advances in Neural Information Processing Systems 34, 8780-8794 (2021)", + "6. Ding, N., Tang, Y., Han, K., Xu, C., Wang, Y.: Network expansion for practical training acceleration. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 20269-20279 (2023)", + "7. Feng, Z., Zhang, Z., Yu, X., Fang, Y., Li, L., Chen, X., Lu, Y., Liu, J., Yin, W., Feng, S., et al.: Ernie-vilg 2.0: Improving text-to-image diffusion model with knowledge-enhanced mixture-of-denoising-experts. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 10135-10145 (2023)", + "8. Fu, C., Huang, H., Jiang, Z., Ni, Y., Nai, L., Wu, G., Cheng, L., Zhou, Y., Li, S., Li, A., et al.: Triple: Revisiting pretrained model reuse and progressive learning for efficient vision transformer scaling and searching. In: ICCV (2023)", + "9. Gu, J., Meng, X., Lu, G., Hou, L., Minzhe, N., Liang, X., Yao, L., Huang, R., Zhang, W., Jiang, X., et al.: Wukong: A 100 million large-scale chinese cross-modal pre-training benchmark. Advances in Neural Information Processing Systems 35, 26418-26431 (2022)", + "0. Heusel, M., Ramsauer, H., Unterthiner, T., Nessler, B., Hochreiter, S.: Gans trained by a two time-scale update rule converge to a local nash equilibrium. Advances in neural information processing systems 30 (2017)", + "1. Ho, J., Jain, A., Abbeel, P.: Denoising diffusion probabilistic models. Advances in Neural Information Processing Systems 33, 6840-6851 (2020)", + "2. Ho, J., Jain, A., Abbeel, P.: Denoising diffusion probabilistic models. Advances in Neural Information Processing Systems 33, 6840-6851 (2020)", + "3. Ho, J., Sahara, C., Chan, W., Fleet, D.J., Norouzi, M., Salimans, T.: Cascaded diffusion models for high fidelity image generation. J. Mach. Learn. Res. 23, 47-1 (2022)", + "4. Ho, J., Salimans, T.: Classifier-free diffusion guidance. arXiv preprint arXiv:2207.12598 (2022)", + "5. Jin, Z., Shen, X., Li, B., Xue, X.: Training-free diffusion model adaptation for variable-sized text-to-image synthesis. arXiv preprint arXiv:2306.08645 (2023)", + "6. Li, X., Xu, C., Wang, X., Lan, W., Jia, Z., Yang, G., Xu, J.: Coco-cn for cross-lingual image tagging, captioning, and retrieval. IEEE Transactions on Multimedia 21(9), 2347-2360 (2019)", + "7. Lin, T.Y., Maire, M., Belongie, S., Hays, J., Perona, P., Ramanan, D., Dólar, P., Zitnick, C.L.: Microsoft coco: Common objects in context. In: Computer Vision-ECCV 2014: 13th European Conference, Zurich, Switzerland, September 6-12, 2014, Proceedings, Part V 13. pp. 740-755. Springer (2014)" + ], + "bbox": [ + 225, + 175, + 784, + 839 + ], + "page_idx": 14 + }, + { + "type": "header", + "text": "PanGu-Draw", + "bbox": [ + 640, + 114, + 730, + 126 + ], + "page_idx": 14 + }, + { + "type": "page_number", + "text": "15", + "bbox": [ + 767, + 116, + 784, + 126 + ], + "page_idx": 14 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "18. Liu, H., Li, C., Wu, Q., Lee, Y.J.: Visual instruction tuning. arXiv preprint arXiv:2304.08485 (2023)", + "19. Liu, N., Li, S., Du, Y., Torralba, A., Tenenbaum, J.B.: Compositional visual generation with composable diffusion models. In: European Conference on Computer Vision. pp. 423-439. Springer (2022)", + "20. Nichol, A., Dhariwal, P., Ramesh, A., Shyam, P., Mishkin, P., McGrew, B., Sutskever, I., Chen, M.: Glide: Towards photorealistic image generation and editing with text-guided diffusion models. arXiv preprint arXiv:2112.10741 (2021)", + "21. Nichol, A.Q., Dhariwal, P.: Improved denoising diffusion probabilistic models. In: ICML. pp. 8162-8171. PMLR (2021)", + "22. Podell, D., English, Z., Lacey, K., Blattmann, A., Dockhorn, T., Müller, J., Penna, J., Rombach, R.: Sdxl: improving latent diffusion models for high-resolution image synthesis. arXiv preprint arXiv:2307.01952 (2023)", + "23. Qin, Y., Zhang, J., Lin, Y., Liu, Z., Li, P., Sun, M., Zhou, J.: Elle: Efficient lifelong pre-training for emerging data. arXiv preprint arXiv:2203.06311 (2022)", + "24. Ramesh, A., Dhariwal, P., Nichol, A., Chu, C., Chen, M.: Hierarchical text-conditional image generation with clip latents. arXiv preprint arXiv:2204.06125 (2022)", + "25. Ramesh, A., Pavlov, M., Goh, G., Gray, S., Voss, C., Radford, A., Chen, M., Sutskever, I.: Zero-shot text-to-image generation. In: International Conference on Machine Learning. pp. 8821-8831. PMLR (2021)", + "26. Rombach, R., Blattmann, A., Lorenz, D., Esser, P., Ommer, B.: High-resolution image synthesis with latent diffusion models. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR). pp. 10684-10695 (June 2022)", + "27. Rombach, R., Blattmann, A., Lorenz, D., Esser, P., Ommer, B.: High-resolution image synthesis with latent diffusion models. In: Proceedings of the IEEE/CVF conference on computer vision and pattern recognition. pp. 10684-10695 (2022)", + "28. Ronneberger, O., Fischer, P., Brox, T.: U-net: Convolutional networks for biomedical image segmentation. In: Medical Image Computing and Computer-Assisted Intervention-MICCAI 2015: 18th International Conference, Munich, Germany, October 5-9, 2015, Proceedings, Part III 18. pp. 234-241. Springer (2015)", + "29. Sahara, C., Chan, W., Saxena, S., Li, L., Whang, J., Denton, E.L., Ghasemipour, K., Gontijo Lopes, R., Karagol Ayan, B., Salimans, T., et al.: Photorealistic text-to-image diffusion models with deep language understanding. Advances in Neural Information Processing Systems 35, 36479-36494 (2022)", + "30. Salimans, T., Goodfellow, I., Zaremba, W., Cheung, V., Radford, A., Chen, X.: Improved techniques for training gans. Advances in neural information processing systems 29, 2234-2242 (2016)", + "31. Shonenkov, A., Konstantinov, M., Bakshandaeva, D., Schuhmann, C., Ivanova, K., Klokova, N.: Deepfloyd if: A powerful text-to-image model that can smartly integrate text into images (2023), https://www deepfloyd.ai/deepfloyd-if, online; accessed 16-November-2023", + "32. Xu, X., Wang, Z., Zhang, E., Wang, K., Shi, H.: Versatile diffusion: Text, images and variations all in one diffusion model. arXiv preprint arXiv:2211.08332 (2022)", + "33. Xue, Z., Song, G., Guo, Q., Liu, B., Zong, Z., Liu, Y., Luo, P.: Raphael: Text-to-image generation via large mixture of diffusion paths. arXiv preprint arXiv:2305.18295 (2023)", + "34. Yang, A., Pan, J., Lin, J., Men, R., Zhang, Y., Zhou, J., Zhou, C.: Chinese clip: Contrastive vision-language pretraining in chinese. arXiv preprint arXiv:2211.01335 (2022)" + ], + "bbox": [ + 215, + 146, + 784, + 839 + ], + "page_idx": 15 + }, + { + "type": "page_number", + "text": "16", + "bbox": [ + 217, + 114, + 235, + 126 + ], + "page_idx": 15 + }, + { + "type": "header", + "text": "G. Lu et al.", + "bbox": [ + 271, + 114, + 351, + 126 + ], + "page_idx": 15 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "35. Yao, L., Han, J., Liang, X., Xu, D., Zhang, W., Li, Z., Xu, H.: Detclipv2: Scalable open-vocabulary object detection pre-training via word-region alignment. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 23497-23506 (2023)", + "36. Yao, L., Huang, R., Hou, L., Lu, G., Niu, M., Xu, H., Liang, X., Li, Z., Jiang, X., Xu, C.: Filip: Fine-grained interactive language-image pre-training. arXiv preprint arXiv:2111.07783 (2021)", + "37. Ye, F., Liu, G., Wu, X., Wu, L.: Altdiffusion: A multilingual text-to-image diffusion model (2023)", + "38. Zhang, J., Gan, R., Wang, J., Zhang, Y., Zhang, L., Yang, P., Gao, X., Wu, Z., Dong, X., He, J., Zhuo, J., Yang, Q., Huang, Y., Li, X., Wu, Y., Lu, J., Zhu, X., Chen, W., Han, T., Pan, K., Wang, R., Wang, H., Wu, X., Zeng, Z., Chen, C.: Fengshenbang 1.0: Being the foundation of chinese cognitive intelligence. CoRR abs/2209.02970 (2022)", + "39. Zhang, L., Rao, A., Agrawala, M.: Adding conditional control to text-to-image diffusion models. In: Proceedings of the IEEE/CVF International Conference on Computer Vision. pp. 3836-3847 (2023)" + ], + "bbox": [ + 212, + 146, + 787, + 375 + ], + "page_idx": 16 + }, + { + "type": "header", + "text": "PanGu-Draw", + "bbox": [ + 640, + 114, + 730, + 126 + ], + "page_idx": 16 + }, + { + "type": "page_number", + "text": "17", + "bbox": [ + 767, + 114, + 785, + 126 + ], + "page_idx": 16 + } +] \ No newline at end of file diff --git a/2024/PanGu-Draw_ Advancing Resource-Efficient Text-to-Image Synthesis with Time-Decoupled Training and Reusable Coop-Diffusion/c75d5546-18a8-4ad8-982c-ff094dd050a6_model.json b/2024/PanGu-Draw_ Advancing Resource-Efficient Text-to-Image Synthesis with Time-Decoupled Training and Reusable Coop-Diffusion/c75d5546-18a8-4ad8-982c-ff094dd050a6_model.json new file mode 100644 index 0000000000000000000000000000000000000000..bc8421cb7992ed9b0fcec456ca6e8c982205711f --- /dev/null +++ b/2024/PanGu-Draw_ Advancing Resource-Efficient Text-to-Image Synthesis with Time-Decoupled Training and Reusable Coop-Diffusion/c75d5546-18a8-4ad8-982c-ff094dd050a6_model.json @@ -0,0 +1,2489 @@ +[ + [ + { + "type": "title", + "bbox": [ + 0.239, + 0.141, + 0.767, + 0.207 + ], + "angle": 0, + "content": "PanGu-Draw: Advancing Resource-Efficient Text-to-Image Synthesis with Time-Decoupled Training and Reusable Coop-Diffusion" + }, + { + "type": "text", + "bbox": [ + 0.236, + 0.232, + 0.768, + 0.265 + ], + "angle": 0, + "content": "Guansong Lu\\(^{1}\\), Yuanfan Guo\\(^{1}\\), Jianhua Han\\(^{1}\\), Minzhe Niu\\(^{1}\\), Yihan Zeng\\(^{1}\\), Songcen Xu\\(^{1}\\), Zeyi Huang\\(^{2}\\), Zhao Zhong\\(^{2}\\), Wei Zhang\\(^{1}\\), and Hang Xu\\(^{1\\star}\\)" + }, + { + "type": "text", + "bbox": [ + 0.381, + 0.273, + 0.624, + 0.288 + ], + "angle": 0, + "content": "1 Huawei Noah's Ark Lab 2 Huawei" + }, + { + "type": "text", + "bbox": [ + 0.262, + 0.319, + 0.744, + 0.653 + ], + "angle": 0, + "content": "Abstract. Current large-scale diffusion models represent a giant leap forward in conditional image synthesis, capable of interpreting diverse cues like text, human poses, and edges. However, their reliance on substantial computational resources and extensive data collection remains a bottleneck. On the other hand, the integration of existing diffusion models, each specialized for different controls and operating in unique latent spaces, poses a challenge due to incompatible image resolutions and latent space embedding structures, hindering their joint use. Addressing these constraints, we present \"PanGu-Draw\", a novel latent diffusion model designed for resource-efficient text-to-image synthesis that adeptly accommodates multiple control signals. We first propose a resource-efficient Time-Decoupling Training Strategy, which splits the monolithic text-to-image model into structure and texture generators. Each generator is trained using a regimen that maximizes data utilization and computational efficiency, cutting data preparation by \\(48\\%\\) and reducing training resources by \\(51\\%\\). Secondly, we introduce \"Coop-Diffusion\", an algorithm that enables the cooperative use of various pre-trained diffusion models with different latent spaces and predefined resolutions within a unified denoising process. This allows for multi-control image synthesis at arbitrary resolutions without the necessity for additional data or retraining. Empirical validations of Pangu-Draw show its exceptional prowess in text-to-image and multi-control image generation, suggesting a promising direction for future model training efficiencies and generation versatility. The largest 5B T2I PanGu-Draw model is released on the Ascend platform. Project page: https://pangu- draw.github.io" + }, + { + "type": "text", + "bbox": [ + 0.262, + 0.664, + 0.756, + 0.679 + ], + "angle": 0, + "content": "Keywords: Text-to-Image Synthesis \\(\\cdot\\) Resource-Efficient \\(\\cdot\\) Diffusion Model" + }, + { + "type": "title", + "bbox": [ + 0.216, + 0.701, + 0.377, + 0.716 + ], + "angle": 0, + "content": "1 Introduction" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.729, + 0.788, + 0.819 + ], + "angle": 0, + "content": "The Denoising Diffusion Probabilistic Models (DDPMs) [11] and their subsequent enhancements [5, 14, 21] have established diffusion models as a leading approach for image generation. These advancements excel in the application of diffusion models to text-to-image synthesis, yielding high-fidelity results with large-scale models and datasets, supported by substantial computational resources [13, 20, 24, 26, 29]. These foundational models, capable of understanding" + }, + { + "type": "page_footnote", + "bbox": [ + 0.218, + 0.825, + 0.533, + 0.84 + ], + "angle": 0, + "content": "* Corresponding author: xuHang@huawei.com" + } + ], + [ + { + "type": "page_number", + "bbox": [ + 0.218, + 0.116, + 0.23, + 0.127 + ], + "angle": 0, + "content": "2" + }, + { + "type": "header", + "bbox": [ + 0.272, + 0.115, + 0.352, + 0.127 + ], + "angle": 0, + "content": "G. Lu et al." + }, + { + "type": "image", + "bbox": [ + 0.263, + 0.143, + 0.753, + 0.306 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.214, + 0.318, + 0.785, + 0.375 + ], + "angle": 0, + "content": "Fig. 1: Illustration of three multi-stage training strategies and comparison between them in resource efficiency in data, training and inference aspects. Our time-decoupling training strategy significantly surpasses the representative methods in Cascaded Training [20,31] and Resolution Boost Training [26,37] in resource efficiency." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.405, + 0.784, + 0.448 + ], + "angle": 0, + "content": "and rendering complex semantics, have paved the way for diverse image generation tasks, accommodating various control signals such as reference images, edges [39], and poses [39]." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.449, + 0.785, + 0.666 + ], + "angle": 0, + "content": "However, the extensive computational demand and significant data collection required by these models pose a substantial challenge. The ambitious goal of higher fidelity and increased resolution in image synthesis pushes the boundaries of model and dataset sizes, escalating computational costs, and environmental impact. Moreover, the aspiration for versatile control and multi-resolution in image generation introduces additional complexity. Existing diffusion models, each tailored for specific controls and operating within distinct latent spaces, face the challenge of integration due to incompatible image resolutions and latent space embeddings, obstructing their concurrent utilization. This incompatibility not only leads to more resource consumption of retraining but also impedes the joint synthesis of images controlled by multiple factors, thereby limiting the scalability and practical application of such existing generative models. In response to these challenges, our work introduces a novel paradigm named \"PanGu-Draw\" that judiciously conserves training resources while enhancing data efficiency, thereby proposing a resource-efficient pathway forward for diffusion model scalability." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.667, + 0.785, + 0.753 + ], + "angle": 0, + "content": "As shown in Figure 1, the training strategies of predecessors like DeepFloyd [31] and GLIDE [20], which employ a cascaded approach, excel in leveraging data across resolutions but suffer from inefficient inference due to their reliance on multiple models. Alternatively, Stable Diffusion [26] and AltDiffusion [37] use a Resolution Boost Training strategy aiming for cost-effectiveness by refining a single model. However, this strategy falls short on data efficiency." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.753, + 0.787, + 0.84 + ], + "angle": 0, + "content": "In light of these considerations, our PanGu-Draw framework advances the field by presenting a Time-Decoupling Training Strategy that segments the training of a comprehensive text-to-image model into two distinct generators: one dedicated to structural outlines and another to textural details. This division not only concentrates on training efforts but also enhances data efficacy. The structural generator is adept at crafting the initial outlines of images, offering" + } + ], + [ + { + "type": "header", + "bbox": [ + 0.642, + 0.115, + 0.732, + 0.128 + ], + "angle": 0, + "content": "PanGu-Draw" + }, + { + "type": "page_number", + "bbox": [ + 0.775, + 0.117, + 0.787, + 0.127 + ], + "angle": 0, + "content": "3" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.147, + 0.788, + 0.248 + ], + "angle": 0, + "content": "flexibility in data quality and enabling training across a spectrum of data calibers; the textural generator, in contrast, is fine-tuned using low-resolution data to infuse these outlines with fine-grained details, ensuring optimal performance even during high-resolution synthesis. This focused approach not only accelerates the training process of our 5B model but also significantly reduces the reliance on extensive data collection and computational resources, as evidenced by a \\(48\\%\\) reduction in data preparation and a \\(51\\%\\) reduction in resource consumption." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.249, + 0.788, + 0.495 + ], + "angle": 0, + "content": "Furthermore, we introduce a pioneering algorithm named Coop-Diffusion, which facilitates the cooperative integration of diverse pre-trained diffusion models. Each model, conditioned on different controls and pre-defined resolutions, contributes to a seamless denoising process. The first algorithmic sub-module addresses inconsistencies in VAE decoders that arise during the denoising process across different latent spaces, ensuring cohesive image quality by effectively reconciling disparate latent space representations. The second sub-module confronts the challenges associated with multi-resolution denoising. Traditional bilinear upsampling for the intermediate noise map, introduced during the denoising process, can undesirably amplify the correlation between pixels. This amplification deviates from the initial Independent and Identically Distributed (IID) assumption, leading to severe artifacts in the final output image. However, our innovative approach circumvents this issue with a single-step sampling method that preserves the integrity of pixel independence, thus preventing the introduction of artifacts. Coop-Diffusion obviates the need for additional data or model retraining, addressing the challenges of multi-control and multi-resolution image generation with scalability and efficiency." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.496, + 0.788, + 0.568 + ], + "angle": 0, + "content": "PanGu-Draw excels in text-to-image (T2I) generation, outperforming established models like DALL-E 2 and SDXL, as evidenced by its FID of 7.99 in English T2I. It also leads in Chinese T2I across metrics like FID, IS, and CNCLIP-score. User feedback highlights a strong preference for PanGu-Draw, aligning well with human visual perceptions." + }, + { + "type": "text", + "bbox": [ + 0.24, + 0.568, + 0.561, + 0.581 + ], + "angle": 0, + "content": "In summary, our contributions are manifold:" + }, + { + "type": "text", + "bbox": [ + 0.215, + 0.582, + 0.785, + 0.611 + ], + "angle": 0, + "content": "- PanGu-Draw: A resource-efficient diffusion model with a Time-Decoupling Training Strategy, reducing data and training resources for T2I synthesis." + }, + { + "type": "text", + "bbox": [ + 0.215, + 0.612, + 0.785, + 0.654 + ], + "angle": 0, + "content": "- **Coop-Diffusion:** A novel approach for integrating multiple diffusion models, enabling efficient multi-control image synthesis at multi-resolutions within a unified denoising process." + }, + { + "type": "text", + "bbox": [ + 0.215, + 0.655, + 0.785, + 0.698 + ], + "angle": 0, + "content": "- Comprehensive evaluations demonstrate PanGu-Draw's (5B model) can produce high-quality images aligned with text and various controls, advancing the scalability and flexibility of diffusion-based image generation." + }, + { + "type": "list", + "bbox": [ + 0.215, + 0.582, + 0.785, + 0.698 + ], + "angle": 0, + "content": null + }, + { + "type": "title", + "bbox": [ + 0.216, + 0.721, + 0.388, + 0.737 + ], + "angle": 0, + "content": "2 Related Work" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.753, + 0.788, + 0.841 + ], + "angle": 0, + "content": "Text-to-Image Generation. The integration of diffusion models into the realm of text-to-image (T2I) generation marks a significant stride in computational creativity [5,7,12,13,20,21,24,26,29,31-33]. Models like GLIDE [20] and DALL-E 2 [24] have significantly advanced in generating diverse and semantically aligned images from textual descriptions. LDM [26] addresses computational challenges by creating images from text-conditioned low-dimensional latent representations." + } + ], + [ + { + "type": "page_number", + "bbox": [ + 0.218, + 0.116, + 0.23, + 0.127 + ], + "angle": 0, + "content": "4" + }, + { + "type": "header", + "bbox": [ + 0.272, + 0.115, + 0.352, + 0.128 + ], + "angle": 0, + "content": "G. Lu et al." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.147, + 0.784, + 0.219 + ], + "angle": 0, + "content": "ControlNet [39] introduces spatial conditioning controls, offering flexibility in image generation under varied conditions like edges and depth. Despite the proliferation of such specialized models, a unified framework that consolidates these disparate capabilities remains absent, limiting the potential for multi-control and complex editing in image synthesis." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.22, + 0.785, + 0.452 + ], + "angle": 0, + "content": "Model Efficient Training and Scaling Up Strategies. Efficient training and scaling of models are pivotal for advancing large-scale neural networks. Previous models like DeepFloyd [31] and GLIDE [20] capitalize on cascaded approaches that proficiently utilize data across various resolutions, which results in less efficient inference processes. Contrastingly, models like Stable Diffusion [26] and AltDiffusion [37] adopt Resolution Boost Training strategies that refine a single model for cost-effectiveness, which however do not fully exploit data efficiency. In scaling up strategies, training efficiency is also important. Efficient adaptation and scaling are explored in [3] through distillation, and in [23] by marrying model expansion with domain-specific prompts. Serial scaling and knowledge distillation reduce training times significantly as demonstrated by [8], while [6] proposes progressive network expansion for faster training with minimal loss. We propose a novel Time-Decoupling training strategy to diffusion model scaling that enhances efficiency. eDiff-I [1] similarly proposes to split a diffusion model across time steps to boost generation quality without increasing inference cost. However, they do not consider about data and training efficiencies." + }, + { + "type": "title", + "bbox": [ + 0.216, + 0.475, + 0.369, + 0.493 + ], + "angle": 0, + "content": "3 Preliminary" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.507, + 0.784, + 0.551 + ], + "angle": 0, + "content": "Given an image \\( x_0 \\), diffusion models first produce a series of noisy images \\( x_1, \\ldots, x_T \\) by adding Gaussian noise to \\( x_0 \\) according to some noise schedule given by \\( \\bar{\\alpha}_t \\) as follows:" + }, + { + "type": "equation", + "bbox": [ + 0.368, + 0.562, + 0.785, + 0.579 + ], + "angle": 0, + "content": "\\[\nx _ {t} = \\sqrt {\\bar {\\alpha} _ {t}} x _ {0} + \\sqrt {1 - \\bar {\\alpha} _ {t}} \\epsilon , \\epsilon \\sim \\mathcal {N} (0, I). \\tag {1}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.215, + 0.589, + 0.784, + 0.619 + ], + "angle": 0, + "content": "Diffusion models then learn a denoising model \\(\\epsilon_{\\theta}(x_t,t)\\) to predict the added noise of a noisy image \\(x_{t}\\) with the following training objective:" + }, + { + "type": "equation", + "bbox": [ + 0.34, + 0.629, + 0.785, + 0.647 + ], + "angle": 0, + "content": "\\[\n\\mathcal {L} = \\mathbb {E} _ {x _ {0} \\sim q (x _ {0}), \\epsilon \\sim \\mathcal {N} (0, I), t \\sim [ 1, T ]} \\| \\epsilon - \\epsilon_ {\\theta} (x _ {t}, t) \\| ^ {2}, \\tag {2}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.657, + 0.784, + 0.716 + ], + "angle": 0, + "content": "where \\( t \\) is uniformly sampled from \\( \\{1, \\dots, T\\} \\). Once the denoising model \\( \\epsilon_{\\theta}(x_t, t) \\) is learned, starting from a random noise \\( x_T \\sim \\mathcal{N}(0, I) \\), one can iteratively predict and reduce the noise in \\( x_t \\) to get a real image \\( x_0 \\). During the sampling process, we can predict the clean data \\( x_0 \\) from \\( \\epsilon_{\\theta}(x_t, t) \\) with single-step sampling as:" + }, + { + "type": "equation", + "bbox": [ + 0.378, + 0.727, + 0.785, + 0.758 + ], + "angle": 0, + "content": "\\[\n\\hat {x} _ {0, t} = \\frac {1}{\\sqrt {\\bar {\\alpha} _ {t}}} \\left(x _ {t} - \\sqrt {1 - \\bar {\\alpha} _ {t}} \\epsilon_ {\\theta} \\left(x _ {t}, t\\right)\\right). \\tag {3}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.768, + 0.785, + 0.841 + ], + "angle": 0, + "content": "Our text-to-image generation model is built on the model architecture proposed in Latent Diffusion Model [26]. In this model, a real image \\( x_0 \\) is first down-sampled 8 times as a lower-dimension latent code \\( z_0 \\) with an image encoder model \\( E \\), which can be decoded with a latent decoder model \\( D \\) back to a real image \\( x_0 \\). The denoising network \\( \\epsilon_{\\theta}(z_t,t,c) \\) is parameterized as a U-Net [28]" + } + ], + [ + { + "type": "header", + "bbox": [ + 0.641, + 0.115, + 0.732, + 0.128 + ], + "angle": 0, + "content": "PanGu-Draw" + }, + { + "type": "page_number", + "bbox": [ + 0.775, + 0.116, + 0.787, + 0.127 + ], + "angle": 0, + "content": "5" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.147, + 0.784, + 0.177 + ], + "angle": 0, + "content": "model, where embedding of time step \\( t \\) is injected with adaptive normalization layers and embedding of input text \\( c \\) is injected with cross-attention layers." + }, + { + "type": "title", + "bbox": [ + 0.215, + 0.202, + 0.381, + 0.218 + ], + "angle": 0, + "content": "4 PanGu-Draw" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.235, + 0.788, + 0.31 + ], + "angle": 0, + "content": "In this section, we first illustrate our resource-efficient 5B text-to-image generation model, trained with a time-decoupling training strategy and further enhanced with a prompt enhancement LLM. Then, we present our Coop-Diffusion algorithm for the cooperative integration of diverse pre-trained diffusion models, enabling multi-control and multi-resolution image generation." + }, + { + "type": "title", + "bbox": [ + 0.215, + 0.333, + 0.557, + 0.35 + ], + "angle": 0, + "content": "4.1 Time-Decoupling Training Strategy" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.36, + 0.788, + 0.476 + ], + "angle": 0, + "content": "Enhancing data, training, and inference efficiency is vital for text-to-image models' practical use. Figure 1 shows two existing training strategies: (a) Cascaded Training, using three models to incrementally improve resolution, is data-efficient but triples training and inference time. (b) Resolution Boost Training starts at \\(512 \\times 512\\) and then \\(1024 \\times 1024\\) resolution, discarding lower resolution data and offering moderate efficiency with higher training costs and single-model inference across all timesteps. These approaches differ from our time-decoupling strategy, detailed below." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.477, + 0.788, + 0.607 + ], + "angle": 0, + "content": "Responding to the need for enhanced efficiencies, we draw inspiration from the denoising trajectory of diffusion processes, where initial denoising stages primarily shape the image's structural foundation, and later stages refine its textural complexity. With this insight, we introduce the Time-Decoupling Training Strategy. This approach divides a comprehensive text-to-image model, denoted as \\(\\epsilon_{\\theta}\\), into two specialized sub-models operating across different temporal intervals: a structure generator, \\(\\epsilon_{struct}\\), and a texture generator, \\(\\epsilon_{texture}\\). Each sub-model is half the size of the original, thus enhancing manageability and reducing computational load." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.608, + 0.788, + 0.738 + ], + "angle": 0, + "content": "As illustrated in Figure 1(c), the structure generator, \\(\\epsilon_{\\text{struct}}\\), is responsible for early-stage denoising across larger time steps, specifically within the range \\(T, \\dots, T_{\\text{struct}}\\), where \\(0 < T_{\\text{struct}} < T\\). This stage focuses on establishing the foundational outlines of the image. Conversely, the texture generator, \\(\\epsilon_{\\text{texture}}\\), operates during the latter, smaller time steps, denoted by \\(T_{\\text{struct}}, \\dots, 0\\), to elaborate on the textural details. Each generator is trained in isolation, which not only alleviates the need for high-memory computation devices but also avoids the complexities associated with model sharding and its accompanying inter-machine communication overhead." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.739, + 0.788, + 0.84 + ], + "angle": 0, + "content": "In the inference phase, \\(\\epsilon_{\\text{struct}}\\) initially constructs a base structural image, \\(z_{T_{\\text{struct}}}\\), from an initial random noise vector, \\(z_T\\). Subsequently, \\(\\epsilon_{\\text{texture}}\\) refines this base to enhance textural details, culminating in the final output, \\(z_0\\). This sequential processing facilitates a more resource-efficient workflow, significantly reducing the hardware footprint and expediting the generation process without compromising the model's performance or output quality, as demonstrated in our ablated experiment in Sec. 5.3." + } + ], + [ + { + "type": "page_number", + "bbox": [ + 0.218, + 0.116, + 0.23, + 0.127 + ], + "angle": 0, + "content": "6" + }, + { + "type": "header", + "bbox": [ + 0.272, + 0.115, + 0.352, + 0.127 + ], + "angle": 0, + "content": "G. Lu et al." + }, + { + "type": "image", + "bbox": [ + 0.251, + 0.144, + 0.744, + 0.339 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.214, + 0.35, + 0.788, + 0.448 + ], + "angle": 0, + "content": "Fig. 2: Visualization of our Coop-Diffusion algorithm for the cooperative integration of diverse pre-trained diffusion models. (a) Existing pre-trained diffusion models, each tailored for specific controls and operating within distinct latent spaces and image resolutions. (b) This sub-module bridges the gap arising from different latent spaces by transforming \\(\\epsilon_t'\\) in latent space B to the target latent space A as \\(\\tilde{\\epsilon}_t\\). (c) This submodule bridges the gap arising from different resolutions by performing upsampling on the predicted clean data \\(\\hat{x}_{0,t}'\\)." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.476, + 0.788, + 0.621 + ], + "angle": 0, + "content": "Resource-Efficient Specialized Training Regime. We further adopt specialized training designs for the above two models. The structure generator \\(\\epsilon_{struct}\\), which derives image structures from text, requires training on an extensive dataset encompassing a wide range of concepts. Traditional methods, like Stable Diffusion, often eliminate low-resolution images, discarding about \\(48\\%\\) of training data and thereby inflating dataset costs. Contrarily, we integrate high-resolution images with upscaled lower-resolution ones. This approach, as proven by our ablated experiments in Sec. 5.3, shows no performance drop, as the predicted \\(z_{T_{struct}}\\) still contains substantial noise. In this way, we achieve higher data efficiency and avoid the problem of semantic degeneration." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.622, + 0.788, + 0.768 + ], + "angle": 0, + "content": "Since the image structure is determined in \\( z_{T_{\\text{struct}}} \\) and the texture generator \\( \\epsilon_{\\text{texture}} \\) focuses on refining texture, we propose training \\( \\epsilon_{\\text{texture}} \\) at a lower resolution while still sampling at high resolution. This strategy, as demonstrated in our ablated experiments in Sec. 5.3, results in no performance drop and no structural problems (e.g., repetitive presentation [15]). Consequently, we achieved an overall \\( 51\\% \\) improvement in training efficiency. Figure 1 summarizes the data, training, and inference efficiency of different training strategies. Besides higher data and training efficiency, our strategy also achieves higher inference efficiency with fewer inference steps compared to the Cascaded Training strategy and a smaller per-step model compared to the Resolution Boost Training strategy." + }, + { + "type": "title", + "bbox": [ + 0.216, + 0.787, + 0.583, + 0.803 + ], + "angle": 0, + "content": "4.2 Coop-Diffusion: Multi-Diffusion Fusion" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.811, + 0.787, + 0.841 + ], + "angle": 0, + "content": "As shown in Figure 2(a), there are numerous pre-trained diffusion models, such as various SD, ControlNet, image variation, etc., each tailored for specific con" + } + ], + [ + { + "type": "header", + "bbox": [ + 0.642, + 0.115, + 0.732, + 0.127 + ], + "angle": 0, + "content": "PanGu-Draw" + }, + { + "type": "page_number", + "bbox": [ + 0.775, + 0.116, + 0.785, + 0.126 + ], + "angle": 0, + "content": "7" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.147, + 0.784, + 0.264 + ], + "angle": 0, + "content": "trols and image resolutions. It is promising to fuse these pre-trained models for multi-control or multi-resolution image generation without needing to train a new model. However, the different latent spaces and resolutions of these models impede joint synthesis of images controlled by different models, thereby limiting their practical applications. In response to these challenges, we propose the Coop-Diffusion algorithm with two key sub-modules, as shown in Figures 2(b) and (c), to bridge the latent space gap and the resolution gap, and to unite the denoising process in the same space." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.265, + 0.785, + 0.334 + ], + "angle": 0, + "content": "Bridging the Latent Space Gap. To bridge the latent space gap between two different latent spaces A and B, we propose to unify the model prediction in latent space A by transforming the model prediction \\(\\epsilon_t^\\prime\\) in latent space B to latent space A using the image space as an intermediate. This is done in the following way: first, we predict the clean data \\(\\hat{z}_{0,t}^{\\prime}\\) using Equation (3) as:" + }, + { + "type": "equation", + "bbox": [ + 0.405, + 0.343, + 0.786, + 0.373 + ], + "angle": 0, + "content": "\\[\n\\dot {z} _ {0, t} ^ {\\prime} = \\frac {1}{\\sqrt {\\bar {\\alpha} _ {t}}} (z _ {t} ^ {\\prime} - \\sqrt {1 - \\bar {\\alpha} _ {t}} \\epsilon_ {t} ^ {\\prime}), \\tag {4}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.215, + 0.38, + 0.784, + 0.44 + ], + "angle": 0, + "content": "which is then decoded into a pixel-level image \\(\\hat{x}_{0,t}^{\\prime}\\) using the latent decoder model \\(D^{\\prime}\\). This image is encoded into latent space A using the image encoder model \\(E\\), as \\(\\tilde{z}_{0,t} = E(\\hat{x}_{0,t}^{\\prime})\\), and finally transformed into a model prediction by inverting Equation (3) as:" + }, + { + "type": "equation", + "bbox": [ + 0.404, + 0.448, + 0.786, + 0.478 + ], + "angle": 0, + "content": "\\[\n\\tilde {\\epsilon} _ {t} = \\frac {1}{\\sqrt {1 - \\bar {\\alpha} _ {t}}} \\left(z _ {t} - \\sqrt {\\bar {\\alpha} _ {t}} \\tilde {z} _ {0, t}\\right). \\tag {5}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.215, + 0.484, + 0.784, + 0.587 + ], + "angle": 0, + "content": "With the united \\(\\tilde{\\epsilon}_t\\), we can now perform multi-control fusion between \\(\\tilde{\\epsilon}_t\\) and \\(\\epsilon_t\\) (the prediction from model \\(\\epsilon_{\\theta}\\) with \\(z_{t}\\) in latent space A, omitted in Figure 2 for brevity). In this paper, we adopt the fusion method proposed in Composable-Diffusion [19] as: \\(\\epsilon_{t,fuse} = d\\cdot \\tilde{\\epsilon}_t + (1 - d)\\cdot \\epsilon_t\\), where \\(d\\) and \\(1 - d\\) are the guidance strengths of each model with \\(d\\in [0,1]\\), to guide the denoising process jointly with these two models for multi-control image generation. Algorithm 1 further illustrates this unification and fusion process." + }, + { + "type": "text", + "bbox": [ + 0.215, + 0.588, + 0.784, + 0.618 + ], + "angle": 0, + "content": "Bridging Resolution Gap. To integrate the denoising processes of a low-resolution model with a high-resolution model, upsampling and/or downsam" + }, + { + "type": "image", + "bbox": [ + 0.331, + 0.65, + 0.495, + 0.772 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.336, + 0.773, + 0.489, + 0.783 + ], + "angle": 0, + "content": "(a) Upsampling from intermediate \\( z_{t} \\)." + }, + { + "type": "image", + "bbox": [ + 0.509, + 0.65, + 0.673, + 0.773 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.528, + 0.773, + 0.654, + 0.783 + ], + "angle": 0, + "content": "(b) Our upsampling algorithm." + }, + { + "type": "image_caption", + "bbox": [ + 0.215, + 0.794, + 0.784, + 0.837 + ], + "angle": 0, + "content": "Fig. 3: Results of fusing a low-resolution model and a high-resolution model with different upsampling methods. Upsampling from intermediate \\( z_{t} \\) results in severe artifacts, while our upsampling algorithm results in high-fidelity image." + } + ], + [ + { + "type": "page_number", + "bbox": [ + 0.218, + 0.116, + 0.23, + 0.127 + ], + "angle": 0, + "content": "8" + }, + { + "type": "header", + "bbox": [ + 0.272, + 0.115, + 0.352, + 0.128 + ], + "angle": 0, + "content": "G. Lu et al." + }, + { + "type": "code_caption", + "bbox": [ + 0.218, + 0.147, + 0.621, + 0.162 + ], + "angle": 0, + "content": "Algorithm 1 Coop-Diffusion: Multi-Diffusion Fusing" + }, + { + "type": "algorithm", + "bbox": [ + 0.217, + 0.164, + 0.787, + 0.544 + ], + "angle": 0, + "content": "Sub-Module 1. Bridging Latent Space Gap \nInput: random noise \\( z_{T} \\sim \\mathcal{N}(0, I) \\), diffusion model \\( \\epsilon_{\\theta} \\), decoder \\( D \\), encoder \\( E \\) in latent space A; random noise \\( z_{T}' = z_{T} \\), diffusion model \\( \\epsilon_{\\theta}' \\), decoder \\( D' \\), encoder \\( E' \\) in latent space B; guidance strength \\( d \\), sampling method \\( S \\). \n1: for \\( t = T, \\ldots, 1 \\) do \n2: \\( \\epsilon_t' = \\epsilon_\\theta'(z_t') \\), \\( \\hat{z}_{0,t}' = \\frac{1}{\\sqrt{\\alpha_t}} (z_t' - \\sqrt{1 - \\bar{\\alpha}_t} \\epsilon_t') \\) \n3: \\( \\hat{x}_{0,t}' = D'(\\hat{z}_{0,t}') \\), \\( \\tilde{z}_{0,t} = E(\\hat{x}_{0,t}') \\) \n4: \\( \\tilde{\\epsilon}_t = \\frac{1}{\\sqrt{1 - \\bar{\\alpha}_t}} (z_t - \\sqrt{\\bar{\\alpha}_t} \\tilde{z}_{0,t}) \\) \n5: \\( \\epsilon_t = \\epsilon_\\theta(z_t) \\), \\( \\epsilon_{t,fuse} = d \\cdot \\tilde{\\epsilon}_t + (1 - d) \\cdot \\epsilon_t \\) \n6: \\( z_{t-1} = S(z_t, t, \\epsilon_{t,fuse}) \\) \n7: \\( z_{t-1}' = S(z_t', t, \\epsilon_{t,fuse}') \\triangleright \\epsilon_{t,fuse}' \\) from \\( \\epsilon_{t,fuse}' \\) similar to the process from \\( \\epsilon_t' \\) to \\( \\tilde{\\epsilon}_t \\), omitted for brevity \n8: end for \n9: return \\( D(z_0) \\) \nSub-Module 2. Bridging Resolution Gap \nInput: diffusion model \\( \\epsilon_{\\theta} \\), decoder \\( D \\), encoder \\( E \\) in high-resolution space; random noise \\( z_T' \\sim \\mathcal{N}(0, I) \\), diffusion model \\( \\epsilon_{\\theta}' \\), decoder \\( D' \\), encoder \\( E' \\) in low-resolution space; low-resolution sampling end step \\( T_{low} \\), sampling method \\( S \\). \n1: for \\( t = T, \\ldots, T_{low} + 1 \\) do \n2: \\( \\epsilon_t' = \\epsilon_\\theta'(z_t') \\), \\( z_{t-1}' = S(z_t', t, \\epsilon_t') \\) \n3: end for \n4: \\( \\hat{z}_{0,T_{low}}' = \\frac{1}{\\sqrt{\\bar{\\alpha}_{T_{low}}}} (z_{T_{low}}' - \\sqrt{1 - \\bar{\\alpha}_{T_{low}}} \\epsilon_{T_{low}}') \\) \n5: \\( \\hat{x}_{0,T_{low}}' = D'(\\hat{z}_{0,T_{low}}') \\), \\( \\hat{x}_{0,T_{low}} = Upsample(\\hat{x}_{0,T_{low}}') \\) \n6: \\( \\hat{z}_{0,T_{low}} = E(\\hat{x}_{0,T_{low}}) \\) \n7: \\( z_{T_{low}} = \\sqrt{\\bar{\\alpha}_{T_{low}}} \\hat{z}_{0,T_{low}} + \\sqrt{1 - \\bar{\\alpha}_{T_{low}}} \\epsilon, \\epsilon \\sim \\mathcal{N}(0, I) \\) \n8: for \\( t = T_{low}, \\ldots, 1 \\) do \n9: \\( \\epsilon_t = \\epsilon_\\theta(z_t) \\), \\( z_{t-1} = S(z_t, t, \\epsilon_t) \\) \n10: end for \n11: return \\( D(z_0) \\)" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.576, + 0.788, + 0.692 + ], + "angle": 0, + "content": "pling is necessary. Traditional bilinear upsampling, often applied to the intermediate result \\( z_{t} \\) during the denoising process, can undesirably amplify pixel correlation. This amplification deviates from the initial Independent and Identically Distributed (IID) assumption, leading to severe artifacts in the final images, as shown in Figure 3(a). Conversely, downsampling does not present this issue. To address the IID issue in upsampling, we propose a new upsampling algorithm that preserves the IID assumption, thereby bridging the resolution gap between models with different pre-trained resolutions." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.694, + 0.806, + 0.81 + ], + "angle": 0, + "content": "Figure 2(c) visualizes our upsampling algorithm. Specifically, for a low-resolution \\( z_{t}^{\\prime} \\), we use the image space as an intermediate space to transform \\( z_{t}^{\\prime} \\) in low-resolution space into high-resolution space as \\( \\tilde{z}_{t} \\). We first predict the noise \\( \\epsilon_{t}^{\\prime} \\) with the denoising model \\( \\epsilon_{\\theta}^{\\prime} \\) and then predict the clean data \\( \\hat{z}_{0,t}^{\\prime} \\) as described in Eq. 4. This is decoded into an image \\( \\hat{x}_{0,t}^{\\prime} \\) using decoder \\( D^{\\prime} \\). We then perform upsampling on \\( \\hat{x}_{0,t}^{\\prime} \\) to obtain its high-resolution counterpart \\( \\hat{x}_{0,t} \\). Finally, \\( \\hat{x}_{0,t} \\) is encoded into the latent space with encoder \\( E \\) as \\( \\hat{z}_{0,t} \\), and \\( t \\)-step noise is added to get the final result \\( \\tilde{z}_{t} \\) using Eq. 1." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.811, + 0.787, + 0.842 + ], + "angle": 0, + "content": "With the unified \\(\\tilde{z}_t\\), we can now perform multi-resolution fusion. First, we denoise with a low-resolution model to obtain the intermediate \\(z_t'\\) and its high-" + } + ], + [ + { + "type": "header", + "bbox": [ + 0.641, + 0.115, + 0.732, + 0.128 + ], + "angle": 0, + "content": "PanGu-Draw" + }, + { + "type": "page_number", + "bbox": [ + 0.775, + 0.116, + 0.787, + 0.127 + ], + "angle": 0, + "content": "9" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.147, + 0.788, + 0.221 + ], + "angle": 0, + "content": "resolution counterpart \\(\\tilde{z}_t\\). Then, we perform denoising with a high-resolution model starting from \\(\\tilde{z}_t\\), and vice versa. This approach allows us to conduct one-stage super-resolution without undergoing all the low-resolution denoising steps, thereby improving inference efficiency. Algorithm 1 further illustrates this unification and fusion process." + }, + { + "type": "title", + "bbox": [ + 0.215, + 0.246, + 0.377, + 0.264 + ], + "angle": 0, + "content": "5 Experiments" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.28, + 0.788, + 0.44 + ], + "angle": 0, + "content": "Implementation Details. We adopt the pretrained Variational Autoencoder (VAE) model from SDXL [22], and we build our structure and texture generator based on the architecture of its U-Net model with the following modifications. To achieve bilingual text-to-image generation (Chinese and English), we pretrain a Chinese text encoder [9, 36] on our Chinese training dataset. We then concatenate the text embeddings from this Chinese text encoder with those from a pretrained English text encoder, serving as the final text embeddings for the denoising models. For multi-resolution image generation, we select a range of image resolutions around \\(1024 \\times 1024\\) and further condition the denoising model on the sinusoidal positional embeddings corresponding to the index of image resolutions. The \\(T_{struct}\\) parameter is set to 500, as suggested by our ablation study." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.44, + 0.788, + 0.614 + ], + "angle": 0, + "content": "Dataset Construction. To encompass the abundant concepts in the world, we collect images in various styles from multiple sources, including Noah-Wukong [9], LAION [27], and others, such as photography, cartoons, portraits, and gaming assets. The collected images are filtered based on CLIP score, aesthetic score, watermark presence, resolution, and aspect ratio. To improve the semantic alignment of PanGu-Draw, we discard parts of the noisy captions that are meaningless or mismatched to the image, sourced from the Internet. Instead, we recapture the collected images by first employing an open-vocabulary detector [35] to locate the primary subjects within the images. These subjects are then processed by LLaVA [18], a high-performance vision-language model, along with prompting templates, to yield detailed image descriptions. These English annotations are subsequently translated into Chinese." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.615, + 0.788, + 0.718 + ], + "angle": 0, + "content": "Evaluation Metrics. We evaluate PanGu-Draw's text-to-image generation on COCO [17] with 30k images for English, and COCO-CN [16] with 10k images for Chinese. The Frechet Inception Distance (FID [10]) is utilized to evaluate image quality and diversity. For Chinese, additional metrics include the Inception Score (IS [30]) and CN-CLIP-score [34], assessing image quality and text-image alignment. Besides, a user study is conducted to evaluate image-text alignment, fidelity, and aesthetics using ImageEval-prompt1 across 339 prompts." + }, + { + "type": "title", + "bbox": [ + 0.215, + 0.742, + 0.48, + 0.758 + ], + "angle": 0, + "content": "5.1 Text-to-Image Generation" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.77, + 0.788, + 0.815 + ], + "angle": 0, + "content": "Evaluation on COCO. As shown in Table 1, PanGu-Draw achieves a FID of 7.99, which is superior to compared methods such as DALL-E 2 and SDXL. It also achieves competitive FID with SOTA methods, indicating the effectiveness" + }, + { + "type": "page_footnote", + "bbox": [ + 0.218, + 0.825, + 0.663, + 0.84 + ], + "angle": 0, + "content": "1 https://github.com/FlagOpen/FlagEval/tree/master/imageEval" + } + ], + [ + { + "type": "page_number", + "bbox": [ + 0.218, + 0.116, + 0.236, + 0.127 + ], + "angle": 0, + "content": "10" + }, + { + "type": "header", + "bbox": [ + 0.272, + 0.115, + 0.352, + 0.127 + ], + "angle": 0, + "content": "G. Lu et al." + }, + { + "type": "table_caption", + "bbox": [ + 0.216, + 0.145, + 0.784, + 0.173 + ], + "angle": 0, + "content": "Table 1: Comparisons of PanGu-Draw with recent representative English text-to-image generation models on COCO dataset in terms of FID." + }, + { + "type": "table", + "bbox": [ + 0.349, + 0.185, + 0.65, + 0.341 + ], + "angle": 0, + "content": "
MethodFID↓Model SizeRelease
DALL-E [25]27.5012BN
LDM [26]12.631.5BY
GLIDE [20]12.245BN
SDXL [22]11.932.5BY
PixArt-α [4]10.650.6BY
DALL-E 2 [24]10.395.5BN
Imagen [29]7.273BN
RAPHAEL [33]6.613BN
PanGu-Draw7.995BY
" + }, + { + "type": "table_caption", + "bbox": [ + 0.214, + 0.358, + 0.784, + 0.4 + ], + "angle": 0, + "content": "Table 2: Comparisons of PanGu-Draw with Chinese text-to-image generation models on COCO-CN dataset in terms of FID, IS and CN-CLIP-score. The classifier-free guidance scales are set as 9 following AltDiffusion [37]." + }, + { + "type": "table", + "bbox": [ + 0.319, + 0.412, + 0.68, + 0.494 + ], + "angle": 0, + "content": "
ModelFID↓IS↑CN-CLIP-score↑
AltDiffusion [37]25.3129.1635.12
Taiyi-Bilingual [38]24.6134.2932.26
Taiyi-CN [38]23.9934.2934.22
PanGu-Draw21.8137.0036.62
" + }, + { + "type": "table_caption", + "bbox": [ + 0.214, + 0.509, + 0.784, + 0.536 + ], + "angle": 0, + "content": "Table 3: Results of a User study on ImageVal-prompt in terms of image-text alignment, image fidelity, and aesthetics." + }, + { + "type": "table", + "bbox": [ + 0.326, + 0.549, + 0.673, + 0.646 + ], + "angle": 0, + "content": "
MethodAlign↑Fidelity↑Aesthetics↑Ave↑
DALL-E 3 [2]4.724.594.764.69
MJ 5.24.634.544.754.64
SDXL [22]4.414.374.594.46
SD [26]4.173.994.204.12
PanGu-Draw4.54.524.724.58
" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.677, + 0.785, + 0.705 + ], + "angle": 0, + "content": "of our time-decoupling training strategy and its outstanding data and training efficiencies. Our 5B PanGu model is the best-released model in terms of FID." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.708, + 0.787, + 0.78 + ], + "angle": 0, + "content": "Evaluation on COCO-CN. As shown in Table 2, PanGu-Draw outperforms other released Chinese text-to-image models, including Taiyi-CN, Taiyi-Bilingual, and AltDiffusion, across all three metrics. This performance highlights PanGu-Draw's exceptional Chinese text-to-image generation capabilities and the effectiveness of our bilingual text encoder architecture." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.782, + 0.787, + 0.84 + ], + "angle": 0, + "content": "User Study. We conducted a user study to compare PanGu-Draw with top-performing methods, including SDXL [22], Midjourney 5.2, and DALL-E 3 [2]. As shown in Table 3, PanGu-Draw achieves better results than SD and SDXL across all three metrics. It also attains approximately \\(99\\% / 98\\%\\) of the performance of" + } + ], + [ + { + "type": "header", + "bbox": [ + 0.642, + 0.115, + 0.732, + 0.127 + ], + "angle": 0, + "content": "PanGu-Draw" + }, + { + "type": "page_number", + "bbox": [ + 0.769, + 0.116, + 0.785, + 0.127 + ], + "angle": 0, + "content": "11" + }, + { + "type": "image", + "bbox": [ + 0.245, + 0.144, + 0.371, + 0.241 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.25, + 0.242, + 0.37, + 0.267 + ], + "angle": 0, + "content": "一位老年女性，身着优雅西服，佩戴菲黛丝帽，身穿夹紧对联鞋。伸出手臂，她的背景是夕阳的校园：“一个 young woman, wearing a elegant gown and traditional dress, carrying a heavy bag of luggage and arms, with the sunset on campus in the background.”" + }, + { + "type": "image", + "bbox": [ + 0.375, + 0.144, + 0.501, + 0.242 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.377, + 0.242, + 0.498, + 0.267 + ], + "angle": 0, + "content": "一面像微笑的女子，身穿白色针织，红色长衣在阳光下翩翩生舞，面庞清新，风姿绰约，细尘一片。“A smiling woman wearing a white T-shirt and a red blouse with a black skirt, a red blouse, a red blouse, a red blouse, a red blouse, a red blouse, a red blouse, a red blouse, a red blouse, a red blouse, a red blouse, a red blouse, a red blouse, a red blouse, a red blouse, a red blouse, a red blouse, a red blouse, a red blouse, a red blouse, a red blouse, a red blouse, a red blouse, a red blouse, a red blouse, a red blouse, a red dress, a red dress, a red dress, a red dress, a red dress, a red dress, a red dress, a red dress, a red dress, a red dress, a red dress, a red dress, a red dress, a red dress" + }, + { + "type": "image", + "bbox": [ + 0.503, + 0.144, + 0.629, + 0.242 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.507, + 0.242, + 0.626, + 0.267 + ], + "angle": 0, + "content": "“一个巨大的水球体，内部蕴含着一个微型雨林，雨林中翅膀翱翔，阳光透过树洒洒洒。”“Auge crystal ball contains a miniature rainforest inside, with butterflies flying in the air, and sunlight shining through the leaves.”" + }, + { + "type": "image", + "bbox": [ + 0.632, + 0.144, + 0.758, + 0.242 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.636, + 0.242, + 0.755, + 0.267 + ], + "angle": 0, + "content": "“一只穿着中世纪短发的兔子，手长颈尖站在一座古董城的城墙上，背后是满目的杂草。” \n\"A rabbit wearing medieval armor and holding a sword stands on the wall of the tower, with a blade with the setting sun behind him.\"" + }, + { + "type": "image", + "bbox": [ + 0.245, + 0.275, + 0.372, + 0.467 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.251, + 0.468, + 0.368, + 0.493 + ], + "angle": 0, + "content": "“赛博办风面摄影的抓机人，依托在夜空飞行平安，以给予水面文明风貌，具有强烈的光影效果。在2019年北京冬奥会上，运动员们穿梭在的night sky, presented in particle ink painting style, with strong light and shadow effects.\"" + }, + { + "type": "image", + "bbox": [ + 0.376, + 0.271, + 0.628, + 0.366 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.381, + 0.367, + 0.624, + 0.383 + ], + "angle": 0, + "content": "“旅行者们正奔在光激染湖的湖面上，周围是雄伟的山脉，以中国水墨画风格描绘，画面也彩淡雅致，具有古朴诗志。”（1980年） \n“旅行者们正奔在湖面上，周围是美丽的山林，以自然、静谧、清新、宁静、和谐、庄严、庄严、庄严、庄严、庄严、庄严、庄严、庄严、庄严、庄严、庄严、庄严、庄严、庄严、庄严、庄严、庄严、庄严、庄严、庄严、庄严、庄严、庄严、庄严、庄严、庄严、庄严、庄严、庄严、庄严、庄严、庄严、庄严、庄严、庄严、庄严、庄严、庄严、庄严、庄严、庄严、庄严、庄严、庄严、庄严、庄严、庄严、庄严、庄严、庄严、庄重的白色。" + }, + { + "type": "image", + "bbox": [ + 0.375, + 0.385, + 0.501, + 0.483 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.381, + 0.483, + 0.498, + 0.504 + ], + "angle": 0, + "content": "“一台未来风格的摩托车，闪耀着霓虹灯，停在夜晚的东京街头。” \n“A futuristic motorcycle, shining with neon lights, is parked on the streets of Tokyo at night.”" + }, + { + "type": "image", + "bbox": [ + 0.504, + 0.385, + 0.629, + 0.483 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.517, + 0.483, + 0.627, + 0.503 + ], + "angle": 0, + "content": "一座由冰晶和雪花构成的精致城堡，坐落在在北极的冰原上。 \n“An exquisite castle made of ice crystals and snowflakes, located on the Arctic ice sheet.”" + }, + { + "type": "image", + "bbox": [ + 0.632, + 0.276, + 0.758, + 0.467 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.638, + 0.468, + 0.75, + 0.483 + ], + "angle": 0, + "content": "一袋老火的糖盔，完全由糖果和巧克力制成。一个 ancient pot was filled of candy and chocolate." + }, + { + "type": "image_caption", + "bbox": [ + 0.215, + 0.514, + 0.788, + 0.557 + ], + "angle": 0, + "content": "Fig. 4: Images generated with PanGu-Draw, our 5B multi-lingual text-to-image generation model. PanGu-Draw is able to generate multi-resolution high-fidelity images semantically aligned with the input prompts." + }, + { + "type": "text", + "bbox": [ + 0.215, + 0.584, + 0.788, + 0.645 + ], + "angle": 0, + "content": "Midjourney 5.2 and DALL-E 3, respectively, indicating PanGu-Draw's excellent text-to-image capabilities. Figure 4 shows some high-fidelity multi-resolution images generated by PanGu-Draw. As we can see, the generated images of PanGu-Draw are of high aesthetics and semantically aligned with the input prompts." + }, + { + "type": "title", + "bbox": [ + 0.216, + 0.664, + 0.517, + 0.679 + ], + "angle": 0, + "content": "5.2 Multi-Diffusion Fusing Results" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.687, + 0.8, + 0.819 + ], + "angle": 0, + "content": "Multi-Control Image Generation. To demonstrate the effectiveness of the proposed reusable multi-diffusion fusing algorithm, Coop-Diffusion, we first present multiple results of multi-control image generation. Figure 5 displays results from fusing an image variation model\\(^2\\) with PanGu-Draw. The fusing results maintain a style similar to that of the reference image, matching the texture described by the input prompt. Figure 6 shows results from fusing PanGu-Draw with a pose/edge-to-image ControlNet model, which operates in guess mode without input prompts. Here, the fusing results combine the structure of the pose/edge image with the texture described by the input prompt." + }, + { + "type": "page_footnote", + "bbox": [ + 0.218, + 0.824, + 0.663, + 0.841 + ], + "angle": 0, + "content": "\\(^{2}\\) https://huggingface.co/lambdalabs/sd-image-variations-diffusers" + } + ], + [ + { + "type": "page_number", + "bbox": [ + 0.218, + 0.116, + 0.236, + 0.127 + ], + "angle": 0, + "content": "12" + }, + { + "type": "header", + "bbox": [ + 0.272, + 0.115, + 0.352, + 0.127 + ], + "angle": 0, + "content": "G. Lu et al." + }, + { + "type": "image", + "bbox": [ + 0.303, + 0.145, + 0.382, + 0.206 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.384, + 0.145, + 0.538, + 0.206 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.319, + 0.207, + 0.521, + 0.216 + ], + "angle": 0, + "content": "(a) Input image and generation results of image variation." + }, + { + "type": "image", + "bbox": [ + 0.316, + 0.215, + 0.369, + 0.256 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.384, + 0.218, + 0.538, + 0.278 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.321, + 0.28, + 0.52, + 0.289 + ], + "angle": 0, + "content": "(b) Input prompt and generation results of PanGu-Draw." + }, + { + "type": "image", + "bbox": [ + 0.547, + 0.145, + 0.701, + 0.264 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.553, + 0.265, + 0.693, + 0.292 + ], + "angle": 0, + "content": "(c) Generation results of fusing image variation and PanGu-Draw with Coop-Diffusion algorithm." + }, + { + "type": "image_caption", + "bbox": [ + 0.215, + 0.31, + 0.785, + 0.339 + ], + "angle": 0, + "content": "Fig. 5: Generation results of the fusing of an image variation model and PanGu-Draw and with the proposed Coop-Diffusion algorithm." + }, + { + "type": "image", + "bbox": [ + 0.309, + 0.359, + 0.386, + 0.42 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.302, + 0.421, + 0.519, + 0.49 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.326, + 0.492, + 0.521, + 0.5 + ], + "angle": 0, + "content": "(b) Input prompt and generation results of PanGu-Draw." + }, + { + "type": "image", + "bbox": [ + 0.309, + 0.511, + 0.386, + 0.571 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.302, + 0.573, + 0.52, + 0.582 + ], + "angle": 0, + "content": "(a) Input edge map and generation results of edge-to-image" + }, + { + "type": "image", + "bbox": [ + 0.312, + 0.583, + 0.539, + 0.643 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.326, + 0.644, + 0.521, + 0.653 + ], + "angle": 0, + "content": "(b) Input prompt and generation results of PanGu-Draw." + }, + { + "type": "image", + "bbox": [ + 0.548, + 0.359, + 0.7, + 0.476 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.553, + 0.478, + 0.694, + 0.503 + ], + "angle": 0, + "content": "(c) Generation results of fusing pose-to-image and PanGu-Draw with Coop-Diffusion algorithm." + }, + { + "type": "image", + "bbox": [ + 0.548, + 0.511, + 0.699, + 0.628 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.553, + 0.629, + 0.693, + 0.654 + ], + "angle": 0, + "content": "(c) Generation results of fusing edge-to-image and PanGu-Draw with CoopDiffusion algorithm." + }, + { + "type": "image_caption", + "bbox": [ + 0.215, + 0.673, + 0.787, + 0.702 + ], + "angle": 0, + "content": "Fig. 6: Generation results guided by fusing signals of text and pose/edge map by our Coop-Diffusion." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.738, + 0.788, + 0.841 + ], + "angle": 0, + "content": "Multi-Resolution Image Generation. We also present multi-resolution image generation results of fusing PanGu-Draw with low-resolution text-to-image and edge-to-image ControlNet model by first denoising with the low-resolution model to get the intermediate \\( z_{t}^{\\prime} \\) and the high-resolution counterpart \\( \\tilde{z}_{t} \\), and then perform denoising in high resolution with PanGu-Draw. Figure 7 shows the results from the low-resolution model and our fusing algorithm Coop-Diffusion. As we can see, PanGu-Draw adds much details to the low-resolution predictions" + } + ], + [ + { + "type": "header", + "bbox": [ + 0.642, + 0.115, + 0.732, + 0.127 + ], + "angle": 0, + "content": "PanGu-Draw" + }, + { + "type": "page_number", + "bbox": [ + 0.769, + 0.116, + 0.786, + 0.127 + ], + "angle": 0, + "content": "13" + }, + { + "type": "image", + "bbox": [ + 0.359, + 0.144, + 0.501, + 0.254 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.503, + 0.144, + 0.645, + 0.254 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.359, + 0.256, + 0.501, + 0.365 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.413, + 0.366, + 0.449, + 0.374 + ], + "angle": 0, + "content": "LR model" + }, + { + "type": "image", + "bbox": [ + 0.503, + 0.255, + 0.645, + 0.365 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.534, + 0.366, + 0.613, + 0.374 + ], + "angle": 0, + "content": "LR model + HR model" + }, + { + "type": "image_caption", + "bbox": [ + 0.215, + 0.384, + 0.788, + 0.455 + ], + "angle": 0, + "content": "Fig. 7: Images generated with a low-resolution (LR) model (1st row: T2I model; 2nd row: edge-to-image ControlNet) and the fusion of the LR model and HR PanGu-Draw with our Coop-Diffusion. The resolutions are \\(512 \\times 512\\) and \\(1024 \\times 1024\\) respectively. This allows for single-stage super-resolution for better details and higher inference efficiency." + }, + { + "type": "text", + "bbox": [ + 0.215, + 0.485, + 0.788, + 0.53 + ], + "angle": 0, + "content": "leading to high-fidelity high-resolution results. Besides, compared with the common practice of super-resolution with diffusion model, which carries out all the low-resolution denoising steps, our method achieve higher inference efficiency." + }, + { + "type": "title", + "bbox": [ + 0.216, + 0.552, + 0.39, + 0.568 + ], + "angle": 0, + "content": "5.3 Ablation Study" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.578, + 0.788, + 0.68 + ], + "angle": 0, + "content": "In this section, we perform ablation studies to analyze our time-decoupling training strategy. The baseline model has \\(1B\\) parameters while the structure and texture generators both have \\(0.5B\\) parameters. During the training process, the latter two models only train half the steps of the baseline model with \\(T_{struct}\\) set as 500. Both settings of the models are trained from scratch on a subset of the LAION dataset containing images with all sizes. After training, FID, IS and CLIP-score on COCO are reported for comparison." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.681, + 0.788, + 0.738 + ], + "angle": 0, + "content": "Time-Decoupling Training Strategy. We compare the performance of models trained with the Resolution Boost strategy and our time-decoupling strategy in Table 4. We found that models trained with our strategy achieves better performance in all three criteria, indicating the effectiveness of our strategy." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.739, + 0.788, + 0.84 + ], + "angle": 0, + "content": "Training Designs. The structure and texture generators (\\(\\epsilon_{\\mathrm{struct}}\\) and \\(\\epsilon_{\\mathrm{texture}}\\)) are designed to train on different resolutions to improve data and training efficiency. However, this approach may negatively influence the final performance. In Table 5, we compare such a design with a traditional training process, where \\(\\epsilon_{\\mathrm{struct}}\\) discards low-resolution images, or \\(\\epsilon_{\\mathrm{texture}}\\) trains with high resolution. Results on COCO show that \\(\\epsilon_{\\mathrm{struct}}\\) benefits from these extra up-scaled data, and \\(\\epsilon_{\\mathrm{texture}}\\) learns enough texture patterns at a smaller resolution." + } + ], + [ + { + "type": "page_number", + "bbox": [ + 0.218, + 0.116, + 0.236, + 0.127 + ], + "angle": 0, + "content": "14" + }, + { + "type": "header", + "bbox": [ + 0.272, + 0.115, + 0.352, + 0.127 + ], + "angle": 0, + "content": "G. Lu et al." + }, + { + "type": "table_caption", + "bbox": [ + 0.216, + 0.145, + 0.784, + 0.174 + ], + "angle": 0, + "content": "Table 4: Comparison of models across Resolution Boost (1B parameters) and Time-Decoupling training strategies (0.5B parameters for structure and texture generators)" + }, + { + "type": "table", + "bbox": [ + 0.339, + 0.185, + 0.66, + 0.235 + ], + "angle": 0, + "content": "
ModelFID↓IS↑CLIP-score↑
Resolution Boost106.1210.4622.9
Time-Decoupling87.6611.0723.4
" + }, + { + "type": "table_caption", + "bbox": [ + 0.216, + 0.249, + 0.784, + 0.275 + ], + "angle": 0, + "content": "Table 5: Performance of structure and texture models training with images of different resolutions." + }, + { + "type": "table", + "bbox": [ + 0.258, + 0.288, + 0.742, + 0.358 + ], + "angle": 0, + "content": "
Structure DataTexture ResolutionFID↓IS↑CLIP-score↑
All data25687.6611.0723.4
Only high resolution25689.5210.9623.2
All data51290.9810.5923.3
" + }, + { + "type": "table_caption", + "bbox": [ + 0.216, + 0.37, + 0.784, + 0.398 + ], + "angle": 0, + "content": "Table 6: Comparisons of PanGu-Draw inference performance with different time step splitting point \\( T_{struct} \\) settings." + }, + { + "type": "table", + "bbox": [ + 0.377, + 0.411, + 0.622, + 0.492 + ], + "angle": 0, + "content": "
TstructFID↓IS↑CLIP-score↑
200105.0810.5922.98
30098.0810.7223.12
50087.6611.0723.40
70089.4811.0223.32
" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.519, + 0.785, + 0.621 + ], + "angle": 0, + "content": "Timestep Splitting Point. The timestep splitting point \\( T_{\\mathrm{struct}} \\) between the structure and texture generators also influences the final performance. To this end, we set \\( T_{\\mathrm{struct}} \\) to 200, 300, 500, and 700, while keeping the other settings of the structure and texture generators unchanged. As shown in Table 6, as \\( T_{\\mathrm{struct}} \\) increases from 200 to 700, the performance initially increases and then decreases continuously. \\( T_{\\mathrm{struct}} = 500 \\) is the optimal value, and we adopt it as the default setting in all other experiments." + }, + { + "type": "title", + "bbox": [ + 0.216, + 0.644, + 0.359, + 0.66 + ], + "angle": 0, + "content": "6 Conclusion" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.675, + 0.787, + 0.835 + ], + "angle": 0, + "content": "In this paper, we present \"PanGu-Draw\", a new latent diffusion model for efficient text-to-image generation that effectively integrates multiple control signals. Our approach includes a Time-Decoupling Training Strategy to separate the text-to-image process into structure and texture generation, enhancing data use and computational efficiency. Additionally, \"Coop-Diffusion\" is introduced, an algorithm allowing cooperative use of different pre-trained diffusion models in a unified denoising process for multi-control image synthesis at various resolutions without extra data or retraining. PanGu-Draw outperforms models like DALL-E 2 and SDXL in English T2I, achieves superior FID, IS, and CN-CLIP-scores in Chinese T2I, and receives favorable user feedback. This positions PanGu-Draw as a versatile and efficient state-of-the-art method." + } + ], + [ + { + "type": "header", + "bbox": [ + 0.642, + 0.115, + 0.732, + 0.127 + ], + "angle": 0, + "content": "PanGu-Draw" + }, + { + "type": "page_number", + "bbox": [ + 0.769, + 0.117, + 0.785, + 0.127 + ], + "angle": 0, + "content": "15" + }, + { + "type": "title", + "bbox": [ + 0.218, + 0.145, + 0.323, + 0.16 + ], + "angle": 0, + "content": "References" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.176, + 0.785, + 0.216 + ], + "angle": 0, + "content": "1. Balaji, Y., Nah, S., Huang, X., Vahdat, A., Song, J., Kreis, K., Aittala, M., Aila, T., Laine, S., Catanzaro, B., et al.: edifi: Text-to-image diffusion models with an ensemble of expert denoisers. arXiv preprint arXiv:2211.01324 (2022)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.217, + 0.785, + 0.257 + ], + "angle": 0, + "content": "2. Betker, J., Goh, G., Jing, L., Brooks, T., Wang, J., Li, L., Ouyang, L., Zhuang, J., Lee, J., Guo, Y., et al.: Improving image generation with better captions. Computer Science. https://cdn.opena.com/papers/dall-e-3.pdf (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.257, + 0.785, + 0.296 + ], + "angle": 0, + "content": "3. Chen, C., Yin, Y., Shang, L., Jiang, X., Qin, Y., Wang, F., Wang, Z., Chen, X., Liu, Z., Liu, Q.: bert2bert: Towards reusable pretrained language models. arXiv preprint arXiv:2110.07143 (2021)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.296, + 0.785, + 0.336 + ], + "angle": 0, + "content": "4. Chen, J., Yu, J., Ge, C., Yao, L., Xie, E., Wu, Y., Wang, Z., Kwok, J., Luo, P., Lu, H., Li, Z.: Fast training of diffusion transformer for photorealistic text-to-image synthesis (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.336, + 0.785, + 0.362 + ], + "angle": 0, + "content": "5. Dhariwal, P., Nichol, A.: Diffusion models beat gans on image synthesis. Advances in Neural Information Processing Systems 34, 8780-8794 (2021)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.362, + 0.785, + 0.402 + ], + "angle": 0, + "content": "6. Ding, N., Tang, Y., Han, K., Xu, C., Wang, Y.: Network expansion for practical training acceleration. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 20269-20279 (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.402, + 0.785, + 0.468 + ], + "angle": 0, + "content": "7. Feng, Z., Zhang, Z., Yu, X., Fang, Y., Li, L., Chen, X., Lu, Y., Liu, J., Yin, W., Feng, S., et al.: Ernie-vilg 2.0: Improving text-to-image diffusion model with knowledge-enhanced mixture-of-denoising-experts. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 10135-10145 (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.469, + 0.785, + 0.509 + ], + "angle": 0, + "content": "8. Fu, C., Huang, H., Jiang, Z., Ni, Y., Nai, L., Wu, G., Cheng, L., Zhou, Y., Li, S., Li, A., et al.: Triple: Revisiting pretrained model reuse and progressive learning for efficient vision transformer scaling and searching. In: ICCV (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.509, + 0.785, + 0.561 + ], + "angle": 0, + "content": "9. Gu, J., Meng, X., Lu, G., Hou, L., Minzhe, N., Liang, X., Yao, L., Huang, R., Zhang, W., Jiang, X., et al.: Wukong: A 100 million large-scale chinese cross-modal pre-training benchmark. Advances in Neural Information Processing Systems 35, 26418-26431 (2022)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.561, + 0.785, + 0.601 + ], + "angle": 0, + "content": "0. Heusel, M., Ramsauer, H., Unterthiner, T., Nessler, B., Hochreiter, S.: Gans trained by a two time-scale update rule converge to a local nash equilibrium. Advances in neural information processing systems 30 (2017)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.601, + 0.785, + 0.627 + ], + "angle": 0, + "content": "1. Ho, J., Jain, A., Abbeel, P.: Denoising diffusion probabilistic models. Advances in Neural Information Processing Systems 33, 6840-6851 (2020)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.628, + 0.785, + 0.654 + ], + "angle": 0, + "content": "2. Ho, J., Jain, A., Abbeel, P.: Denoising diffusion probabilistic models. Advances in Neural Information Processing Systems 33, 6840-6851 (2020)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.654, + 0.785, + 0.693 + ], + "angle": 0, + "content": "3. Ho, J., Sahara, C., Chan, W., Fleet, D.J., Norouzi, M., Salimans, T.: Cascaded diffusion models for high fidelity image generation. J. Mach. Learn. Res. 23, 47-1 (2022)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.693, + 0.785, + 0.72 + ], + "angle": 0, + "content": "4. Ho, J., Salimans, T.: Classifier-free diffusion guidance. arXiv preprint arXiv:2207.12598 (2022)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.721, + 0.785, + 0.746 + ], + "angle": 0, + "content": "5. Jin, Z., Shen, X., Li, B., Xue, X.: Training-free diffusion model adaptation for variable-sized text-to-image synthesis. arXiv preprint arXiv:2306.08645 (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.746, + 0.785, + 0.787 + ], + "angle": 0, + "content": "6. Li, X., Xu, C., Wang, X., Lan, W., Jia, Z., Yang, G., Xu, J.: Coco-cn for cross-lingual image tagging, captioning, and retrieval. IEEE Transactions on Multimedia 21(9), 2347-2360 (2019)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.787, + 0.785, + 0.84 + ], + "angle": 0, + "content": "7. Lin, T.Y., Maire, M., Belongie, S., Hays, J., Perona, P., Ramanan, D., Dólar, P., Zitnick, C.L.: Microsoft coco: Common objects in context. In: Computer Vision-ECCV 2014: 13th European Conference, Zurich, Switzerland, September 6-12, 2014, Proceedings, Part V 13. pp. 740-755. Springer (2014)" + }, + { + "type": "list", + "bbox": [ + 0.226, + 0.176, + 0.785, + 0.84 + ], + "angle": 0, + "content": null + } + ], + [ + { + "type": "page_number", + "bbox": [ + 0.218, + 0.116, + 0.236, + 0.127 + ], + "angle": 0, + "content": "16" + }, + { + "type": "header", + "bbox": [ + 0.272, + 0.115, + 0.352, + 0.127 + ], + "angle": 0, + "content": "G. Lu et al." + }, + { + "type": "ref_text", + "bbox": [ + 0.218, + 0.147, + 0.785, + 0.175 + ], + "angle": 0, + "content": "18. Liu, H., Li, C., Wu, Q., Lee, Y.J.: Visual instruction tuning. arXiv preprint arXiv:2304.08485 (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.176, + 0.785, + 0.215 + ], + "angle": 0, + "content": "19. Liu, N., Li, S., Du, Y., Torralba, A., Tenenbaum, J.B.: Compositional visual generation with composable diffusion models. In: European Conference on Computer Vision. pp. 423-439. Springer (2022)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.217, + 0.785, + 0.257 + ], + "angle": 0, + "content": "20. Nichol, A., Dhariwal, P., Ramesh, A., Shyam, P., Mishkin, P., McGrew, B., Sutskever, I., Chen, M.: Glide: Towards photorealistic image generation and editing with text-guided diffusion models. arXiv preprint arXiv:2112.10741 (2021)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.258, + 0.785, + 0.284 + ], + "angle": 0, + "content": "21. Nichol, A.Q., Dhariwal, P.: Improved denoising diffusion probabilistic models. In: ICML. pp. 8162-8171. PMLR (2021)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.285, + 0.785, + 0.325 + ], + "angle": 0, + "content": "22. Podell, D., English, Z., Lacey, K., Blattmann, A., Dockhorn, T., Müller, J., Penna, J., Rombach, R.: Sdxl: improving latent diffusion models for high-resolution image synthesis. arXiv preprint arXiv:2307.01952 (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.326, + 0.785, + 0.353 + ], + "angle": 0, + "content": "23. Qin, Y., Zhang, J., Lin, Y., Liu, Z., Li, P., Sun, M., Zhou, J.: Elle: Efficient lifelong pre-training for emerging data. arXiv preprint arXiv:2203.06311 (2022)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.354, + 0.785, + 0.393 + ], + "angle": 0, + "content": "24. Ramesh, A., Dhariwal, P., Nichol, A., Chu, C., Chen, M.: Hierarchical text-conditional image generation with clip latents. arXiv preprint arXiv:2204.06125 (2022)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.394, + 0.785, + 0.434 + ], + "angle": 0, + "content": "25. Ramesh, A., Pavlov, M., Goh, G., Gray, S., Voss, C., Radford, A., Chen, M., Sutskever, I.: Zero-shot text-to-image generation. In: International Conference on Machine Learning. pp. 8821-8831. PMLR (2021)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.435, + 0.785, + 0.488 + ], + "angle": 0, + "content": "26. Rombach, R., Blattmann, A., Lorenz, D., Esser, P., Ommer, B.: High-resolution image synthesis with latent diffusion models. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR). pp. 10684-10695 (June 2022)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.489, + 0.785, + 0.529 + ], + "angle": 0, + "content": "27. Rombach, R., Blattmann, A., Lorenz, D., Esser, P., Ommer, B.: High-resolution image synthesis with latent diffusion models. In: Proceedings of the IEEE/CVF conference on computer vision and pattern recognition. pp. 10684-10695 (2022)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.53, + 0.785, + 0.583 + ], + "angle": 0, + "content": "28. Ronneberger, O., Fischer, P., Brox, T.: U-net: Convolutional networks for biomedical image segmentation. In: Medical Image Computing and Computer-Assisted Intervention-MICCAI 2015: 18th International Conference, Munich, Germany, October 5-9, 2015, Proceedings, Part III 18. pp. 234-241. Springer (2015)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.584, + 0.785, + 0.636 + ], + "angle": 0, + "content": "29. Sahara, C., Chan, W., Saxena, S., Li, L., Whang, J., Denton, E.L., Ghasemipour, K., Gontijo Lopes, R., Karagol Ayan, B., Salimans, T., et al.: Photorealistic text-to-image diffusion models with deep language understanding. Advances in Neural Information Processing Systems 35, 36479-36494 (2022)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.637, + 0.785, + 0.677 + ], + "angle": 0, + "content": "30. Salimans, T., Goodfellow, I., Zaremba, W., Cheung, V., Radford, A., Chen, X.: Improved techniques for training gans. Advances in neural information processing systems 29, 2234-2242 (2016)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.678, + 0.785, + 0.731 + ], + "angle": 0, + "content": "31. Shonenkov, A., Konstantinov, M., Bakshandaeva, D., Schuhmann, C., Ivanova, K., Klokova, N.: Deepfloyd if: A powerful text-to-image model that can smartly integrate text into images (2023), https://www deepfloyd.ai/deepfloyd-if, online; accessed 16-November-2023" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.732, + 0.785, + 0.759 + ], + "angle": 0, + "content": "32. Xu, X., Wang, Z., Zhang, E., Wang, K., Shi, H.: Versatile diffusion: Text, images and variations all in one diffusion model. arXiv preprint arXiv:2211.08332 (2022)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.76, + 0.785, + 0.799 + ], + "angle": 0, + "content": "33. Xue, Z., Song, G., Guo, Q., Liu, B., Zong, Z., Liu, Y., Luo, P.: Raphael: Text-to-image generation via large mixture of diffusion paths. arXiv preprint arXiv:2305.18295 (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.8, + 0.785, + 0.84 + ], + "angle": 0, + "content": "34. Yang, A., Pan, J., Lin, J., Men, R., Zhang, Y., Zhou, J., Zhou, C.: Chinese clip: Contrastive vision-language pretraining in chinese. arXiv preprint arXiv:2211.01335 (2022)" + }, + { + "type": "list", + "bbox": [ + 0.217, + 0.147, + 0.785, + 0.84 + ], + "angle": 0, + "content": null + } + ], + [ + { + "type": "header", + "bbox": [ + 0.641, + 0.115, + 0.732, + 0.127 + ], + "angle": 0, + "content": "PanGu-Draw" + }, + { + "type": "page_number", + "bbox": [ + 0.769, + 0.116, + 0.786, + 0.127 + ], + "angle": 0, + "content": "17" + }, + { + "type": "ref_text", + "bbox": [ + 0.214, + 0.147, + 0.788, + 0.202 + ], + "angle": 0, + "content": "35. Yao, L., Han, J., Liang, X., Xu, D., Zhang, W., Li, Z., Xu, H.: Detclipv2: Scalable open-vocabulary object detection pre-training via word-region alignment. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 23497-23506 (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.214, + 0.203, + 0.787, + 0.242 + ], + "angle": 0, + "content": "36. Yao, L., Huang, R., Hou, L., Lu, G., Niu, M., Xu, H., Liang, X., Li, Z., Jiang, X., Xu, C.: Filip: Fine-grained interactive language-image pre-training. arXiv preprint arXiv:2111.07783 (2021)" + }, + { + "type": "ref_text", + "bbox": [ + 0.215, + 0.242, + 0.787, + 0.268 + ], + "angle": 0, + "content": "37. Ye, F., Liu, G., Wu, X., Wu, L.: Altdiffusion: A multilingual text-to-image diffusion model (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.215, + 0.268, + 0.787, + 0.335 + ], + "angle": 0, + "content": "38. Zhang, J., Gan, R., Wang, J., Zhang, Y., Zhang, L., Yang, P., Gao, X., Wu, Z., Dong, X., He, J., Zhuo, J., Yang, Q., Huang, Y., Li, X., Wu, Y., Lu, J., Zhu, X., Chen, W., Han, T., Pan, K., Wang, R., Wang, H., Wu, X., Zeng, Z., Chen, C.: Fengshenbang 1.0: Being the foundation of chinese cognitive intelligence. CoRR abs/2209.02970 (2022)" + }, + { + "type": "ref_text", + "bbox": [ + 0.215, + 0.335, + 0.787, + 0.375 + ], + "angle": 0, + "content": "39. Zhang, L., Rao, A., Agrawala, M.: Adding conditional control to text-to-image diffusion models. In: Proceedings of the IEEE/CVF International Conference on Computer Vision. pp. 3836-3847 (2023)" + }, + { + "type": "list", + "bbox": [ + 0.214, + 0.147, + 0.788, + 0.375 + ], + "angle": 0, + "content": null + } + ] +] \ No newline at end of file diff --git a/2024/PanGu-Draw_ Advancing Resource-Efficient Text-to-Image Synthesis with Time-Decoupled Training and Reusable Coop-Diffusion/c75d5546-18a8-4ad8-982c-ff094dd050a6_origin.pdf b/2024/PanGu-Draw_ Advancing Resource-Efficient Text-to-Image Synthesis with Time-Decoupled Training and Reusable Coop-Diffusion/c75d5546-18a8-4ad8-982c-ff094dd050a6_origin.pdf new file mode 100644 index 0000000000000000000000000000000000000000..28b03ddb53ad0ffb8301511680677d0ab5fab84e --- /dev/null +++ b/2024/PanGu-Draw_ Advancing Resource-Efficient Text-to-Image Synthesis with Time-Decoupled Training and Reusable Coop-Diffusion/c75d5546-18a8-4ad8-982c-ff094dd050a6_origin.pdf @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:da896b6b0ac542038b3a00e3dd895bbfd68073a110e28b944bff23f2b28a5077 +size 22479358 diff --git a/2024/PanGu-Draw_ Advancing Resource-Efficient Text-to-Image Synthesis with Time-Decoupled Training and Reusable Coop-Diffusion/full.md b/2024/PanGu-Draw_ Advancing Resource-Efficient Text-to-Image Synthesis with Time-Decoupled Training and Reusable Coop-Diffusion/full.md new file mode 100644 index 0000000000000000000000000000000000000000..0e907995c6d89e3d7a43ac9244d774d0611bfa43 --- /dev/null +++ b/2024/PanGu-Draw_ Advancing Resource-Efficient Text-to-Image Synthesis with Time-Decoupled Training and Reusable Coop-Diffusion/full.md @@ -0,0 +1,337 @@ +# PanGu-Draw: Advancing Resource-Efficient Text-to-Image Synthesis with Time-Decoupled Training and Reusable Coop-Diffusion + +Guansong Lu $^{1}$ , Yuanfan Guo $^{1}$ , Jianhua Han $^{1}$ , Minzhe Niu $^{1}$ , Yihan Zeng $^{1}$ , Songcen Xu $^{1}$ , Zeyi Huang $^{2}$ , Zhao Zhong $^{2}$ , Wei Zhang $^{1}$ , and Hang Xu $^{1\star}$ + +1 Huawei Noah's Ark Lab 2 Huawei + +Abstract. Current large-scale diffusion models represent a giant leap forward in conditional image synthesis, capable of interpreting diverse cues like text, human poses, and edges. However, their reliance on substantial computational resources and extensive data collection remains a bottleneck. On the other hand, the integration of existing diffusion models, each specialized for different controls and operating in unique latent spaces, poses a challenge due to incompatible image resolutions and latent space embedding structures, hindering their joint use. Addressing these constraints, we present "PanGu-Draw", a novel latent diffusion model designed for resource-efficient text-to-image synthesis that adeptly accommodates multiple control signals. We first propose a resource-efficient Time-Decoupling Training Strategy, which splits the monolithic text-to-image model into structure and texture generators. Each generator is trained using a regimen that maximizes data utilization and computational efficiency, cutting data preparation by $48\%$ and reducing training resources by $51\%$ . Secondly, we introduce "Coop-Diffusion", an algorithm that enables the cooperative use of various pre-trained diffusion models with different latent spaces and predefined resolutions within a unified denoising process. This allows for multi-control image synthesis at arbitrary resolutions without the necessity for additional data or retraining. Empirical validations of Pangu-Draw show its exceptional prowess in text-to-image and multi-control image generation, suggesting a promising direction for future model training efficiencies and generation versatility. The largest 5B T2I PanGu-Draw model is released on the Ascend platform. Project page: https://pangu- draw.github.io + +Keywords: Text-to-Image Synthesis $\cdot$ Resource-Efficient $\cdot$ Diffusion Model + +# 1 Introduction + +The Denoising Diffusion Probabilistic Models (DDPMs) [11] and their subsequent enhancements [5, 14, 21] have established diffusion models as a leading approach for image generation. These advancements excel in the application of diffusion models to text-to-image synthesis, yielding high-fidelity results with large-scale models and datasets, supported by substantial computational resources [13, 20, 24, 26, 29]. These foundational models, capable of understanding + +![](images/0cb2654b4516e86a69532b16ceedc69bbf26fbaba52d2b1bace1ef47193282f1.jpg) +Fig. 1: Illustration of three multi-stage training strategies and comparison between them in resource efficiency in data, training and inference aspects. Our time-decoupling training strategy significantly surpasses the representative methods in Cascaded Training [20,31] and Resolution Boost Training [26,37] in resource efficiency. + +and rendering complex semantics, have paved the way for diverse image generation tasks, accommodating various control signals such as reference images, edges [39], and poses [39]. + +However, the extensive computational demand and significant data collection required by these models pose a substantial challenge. The ambitious goal of higher fidelity and increased resolution in image synthesis pushes the boundaries of model and dataset sizes, escalating computational costs, and environmental impact. Moreover, the aspiration for versatile control and multi-resolution in image generation introduces additional complexity. Existing diffusion models, each tailored for specific controls and operating within distinct latent spaces, face the challenge of integration due to incompatible image resolutions and latent space embeddings, obstructing their concurrent utilization. This incompatibility not only leads to more resource consumption of retraining but also impedes the joint synthesis of images controlled by multiple factors, thereby limiting the scalability and practical application of such existing generative models. In response to these challenges, our work introduces a novel paradigm named "PanGu-Draw" that judiciously conserves training resources while enhancing data efficiency, thereby proposing a resource-efficient pathway forward for diffusion model scalability. + +As shown in Figure 1, the training strategies of predecessors like DeepFloyd [31] and GLIDE [20], which employ a cascaded approach, excel in leveraging data across resolutions but suffer from inefficient inference due to their reliance on multiple models. Alternatively, Stable Diffusion [26] and AltDiffusion [37] use a Resolution Boost Training strategy aiming for cost-effectiveness by refining a single model. However, this strategy falls short on data efficiency. + +In light of these considerations, our PanGu-Draw framework advances the field by presenting a Time-Decoupling Training Strategy that segments the training of a comprehensive text-to-image model into two distinct generators: one dedicated to structural outlines and another to textural details. This division not only concentrates on training efforts but also enhances data efficacy. The structural generator is adept at crafting the initial outlines of images, offering + +flexibility in data quality and enabling training across a spectrum of data calibers; the textural generator, in contrast, is fine-tuned using low-resolution data to infuse these outlines with fine-grained details, ensuring optimal performance even during high-resolution synthesis. This focused approach not only accelerates the training process of our 5B model but also significantly reduces the reliance on extensive data collection and computational resources, as evidenced by a $48\%$ reduction in data preparation and a $51\%$ reduction in resource consumption. + +Furthermore, we introduce a pioneering algorithm named Coop-Diffusion, which facilitates the cooperative integration of diverse pre-trained diffusion models. Each model, conditioned on different controls and pre-defined resolutions, contributes to a seamless denoising process. The first algorithmic sub-module addresses inconsistencies in VAE decoders that arise during the denoising process across different latent spaces, ensuring cohesive image quality by effectively reconciling disparate latent space representations. The second sub-module confronts the challenges associated with multi-resolution denoising. Traditional bilinear upsampling for the intermediate noise map, introduced during the denoising process, can undesirably amplify the correlation between pixels. This amplification deviates from the initial Independent and Identically Distributed (IID) assumption, leading to severe artifacts in the final output image. However, our innovative approach circumvents this issue with a single-step sampling method that preserves the integrity of pixel independence, thus preventing the introduction of artifacts. Coop-Diffusion obviates the need for additional data or model retraining, addressing the challenges of multi-control and multi-resolution image generation with scalability and efficiency. + +PanGu-Draw excels in text-to-image (T2I) generation, outperforming established models like DALL-E 2 and SDXL, as evidenced by its FID of 7.99 in English T2I. It also leads in Chinese T2I across metrics like FID, IS, and CNCLIP-score. User feedback highlights a strong preference for PanGu-Draw, aligning well with human visual perceptions. + +In summary, our contributions are manifold: + +- PanGu-Draw: A resource-efficient diffusion model with a Time-Decoupling Training Strategy, reducing data and training resources for T2I synthesis. +- **Coop-Diffusion:** A novel approach for integrating multiple diffusion models, enabling efficient multi-control image synthesis at multi-resolutions within a unified denoising process. +- Comprehensive evaluations demonstrate PanGu-Draw's (5B model) can produce high-quality images aligned with text and various controls, advancing the scalability and flexibility of diffusion-based image generation. + +# 2 Related Work + +Text-to-Image Generation. The integration of diffusion models into the realm of text-to-image (T2I) generation marks a significant stride in computational creativity [5,7,12,13,20,21,24,26,29,31-33]. Models like GLIDE [20] and DALL-E 2 [24] have significantly advanced in generating diverse and semantically aligned images from textual descriptions. LDM [26] addresses computational challenges by creating images from text-conditioned low-dimensional latent representations. + +ControlNet [39] introduces spatial conditioning controls, offering flexibility in image generation under varied conditions like edges and depth. Despite the proliferation of such specialized models, a unified framework that consolidates these disparate capabilities remains absent, limiting the potential for multi-control and complex editing in image synthesis. + +Model Efficient Training and Scaling Up Strategies. Efficient training and scaling of models are pivotal for advancing large-scale neural networks. Previous models like DeepFloyd [31] and GLIDE [20] capitalize on cascaded approaches that proficiently utilize data across various resolutions, which results in less efficient inference processes. Contrastingly, models like Stable Diffusion [26] and AltDiffusion [37] adopt Resolution Boost Training strategies that refine a single model for cost-effectiveness, which however do not fully exploit data efficiency. In scaling up strategies, training efficiency is also important. Efficient adaptation and scaling are explored in [3] through distillation, and in [23] by marrying model expansion with domain-specific prompts. Serial scaling and knowledge distillation reduce training times significantly as demonstrated by [8], while [6] proposes progressive network expansion for faster training with minimal loss. We propose a novel Time-Decoupling training strategy to diffusion model scaling that enhances efficiency. eDiff-I [1] similarly proposes to split a diffusion model across time steps to boost generation quality without increasing inference cost. However, they do not consider about data and training efficiencies. + +# 3 Preliminary + +Given an image $x_0$ , diffusion models first produce a series of noisy images $x_1, \ldots, x_T$ by adding Gaussian noise to $x_0$ according to some noise schedule given by $\bar{\alpha}_t$ as follows: + +$$ +x _ {t} = \sqrt {\bar {\alpha} _ {t}} x _ {0} + \sqrt {1 - \bar {\alpha} _ {t}} \epsilon , \epsilon \sim \mathcal {N} (0, I). \tag {1} +$$ + +Diffusion models then learn a denoising model $\epsilon_{\theta}(x_t,t)$ to predict the added noise of a noisy image $x_{t}$ with the following training objective: + +$$ +\mathcal {L} = \mathbb {E} _ {x _ {0} \sim q (x _ {0}), \epsilon \sim \mathcal {N} (0, I), t \sim [ 1, T ]} \| \epsilon - \epsilon_ {\theta} (x _ {t}, t) \| ^ {2}, \tag {2} +$$ + +where $t$ is uniformly sampled from $\{1, \dots, T\}$ . Once the denoising model $\epsilon_{\theta}(x_t, t)$ is learned, starting from a random noise $x_T \sim \mathcal{N}(0, I)$ , one can iteratively predict and reduce the noise in $x_t$ to get a real image $x_0$ . During the sampling process, we can predict the clean data $x_0$ from $\epsilon_{\theta}(x_t, t)$ with single-step sampling as: + +$$ +\hat {x} _ {0, t} = \frac {1}{\sqrt {\bar {\alpha} _ {t}}} \left(x _ {t} - \sqrt {1 - \bar {\alpha} _ {t}} \epsilon_ {\theta} \left(x _ {t}, t\right)\right). \tag {3} +$$ + +Our text-to-image generation model is built on the model architecture proposed in Latent Diffusion Model [26]. In this model, a real image $x_0$ is first down-sampled 8 times as a lower-dimension latent code $z_0$ with an image encoder model $E$ , which can be decoded with a latent decoder model $D$ back to a real image $x_0$ . The denoising network $\epsilon_{\theta}(z_t,t,c)$ is parameterized as a U-Net [28] + +model, where embedding of time step $t$ is injected with adaptive normalization layers and embedding of input text $c$ is injected with cross-attention layers. + +# 4 PanGu-Draw + +In this section, we first illustrate our resource-efficient 5B text-to-image generation model, trained with a time-decoupling training strategy and further enhanced with a prompt enhancement LLM. Then, we present our Coop-Diffusion algorithm for the cooperative integration of diverse pre-trained diffusion models, enabling multi-control and multi-resolution image generation. + +# 4.1 Time-Decoupling Training Strategy + +Enhancing data, training, and inference efficiency is vital for text-to-image models' practical use. Figure 1 shows two existing training strategies: (a) Cascaded Training, using three models to incrementally improve resolution, is data-efficient but triples training and inference time. (b) Resolution Boost Training starts at $512 \times 512$ and then $1024 \times 1024$ resolution, discarding lower resolution data and offering moderate efficiency with higher training costs and single-model inference across all timesteps. These approaches differ from our time-decoupling strategy, detailed below. + +Responding to the need for enhanced efficiencies, we draw inspiration from the denoising trajectory of diffusion processes, where initial denoising stages primarily shape the image's structural foundation, and later stages refine its textural complexity. With this insight, we introduce the Time-Decoupling Training Strategy. This approach divides a comprehensive text-to-image model, denoted as $\epsilon_{\theta}$ , into two specialized sub-models operating across different temporal intervals: a structure generator, $\epsilon_{struct}$ , and a texture generator, $\epsilon_{texture}$ . Each sub-model is half the size of the original, thus enhancing manageability and reducing computational load. + +As illustrated in Figure 1(c), the structure generator, $\epsilon_{\text{struct}}$ , is responsible for early-stage denoising across larger time steps, specifically within the range $T, \dots, T_{\text{struct}}$ , where $0 < T_{\text{struct}} < T$ . This stage focuses on establishing the foundational outlines of the image. Conversely, the texture generator, $\epsilon_{\text{texture}}$ , operates during the latter, smaller time steps, denoted by $T_{\text{struct}}, \dots, 0$ , to elaborate on the textural details. Each generator is trained in isolation, which not only alleviates the need for high-memory computation devices but also avoids the complexities associated with model sharding and its accompanying inter-machine communication overhead. + +In the inference phase, $\epsilon_{\text{struct}}$ initially constructs a base structural image, $z_{T_{\text{struct}}}$ , from an initial random noise vector, $z_T$ . Subsequently, $\epsilon_{\text{texture}}$ refines this base to enhance textural details, culminating in the final output, $z_0$ . This sequential processing facilitates a more resource-efficient workflow, significantly reducing the hardware footprint and expediting the generation process without compromising the model's performance or output quality, as demonstrated in our ablated experiment in Sec. 5.3. + +![](images/b26313a4b7eaefc2071c2ca00ab616296f64a56bceb23db74fe5e2378bf971e2.jpg) +Fig. 2: Visualization of our Coop-Diffusion algorithm for the cooperative integration of diverse pre-trained diffusion models. (a) Existing pre-trained diffusion models, each tailored for specific controls and operating within distinct latent spaces and image resolutions. (b) This sub-module bridges the gap arising from different latent spaces by transforming $\epsilon_t'$ in latent space B to the target latent space A as $\tilde{\epsilon}_t$ . (c) This submodule bridges the gap arising from different resolutions by performing upsampling on the predicted clean data $\hat{x}_{0,t}'$ . + +Resource-Efficient Specialized Training Regime. We further adopt specialized training designs for the above two models. The structure generator $\epsilon_{struct}$ , which derives image structures from text, requires training on an extensive dataset encompassing a wide range of concepts. Traditional methods, like Stable Diffusion, often eliminate low-resolution images, discarding about $48\%$ of training data and thereby inflating dataset costs. Contrarily, we integrate high-resolution images with upscaled lower-resolution ones. This approach, as proven by our ablated experiments in Sec. 5.3, shows no performance drop, as the predicted $z_{T_{struct}}$ still contains substantial noise. In this way, we achieve higher data efficiency and avoid the problem of semantic degeneration. + +Since the image structure is determined in $z_{T_{\text{struct}}}$ and the texture generator $\epsilon_{\text{texture}}$ focuses on refining texture, we propose training $\epsilon_{\text{texture}}$ at a lower resolution while still sampling at high resolution. This strategy, as demonstrated in our ablated experiments in Sec. 5.3, results in no performance drop and no structural problems (e.g., repetitive presentation [15]). Consequently, we achieved an overall $51\%$ improvement in training efficiency. Figure 1 summarizes the data, training, and inference efficiency of different training strategies. Besides higher data and training efficiency, our strategy also achieves higher inference efficiency with fewer inference steps compared to the Cascaded Training strategy and a smaller per-step model compared to the Resolution Boost Training strategy. + +# 4.2 Coop-Diffusion: Multi-Diffusion Fusion + +As shown in Figure 2(a), there are numerous pre-trained diffusion models, such as various SD, ControlNet, image variation, etc., each tailored for specific con + +trols and image resolutions. It is promising to fuse these pre-trained models for multi-control or multi-resolution image generation without needing to train a new model. However, the different latent spaces and resolutions of these models impede joint synthesis of images controlled by different models, thereby limiting their practical applications. In response to these challenges, we propose the Coop-Diffusion algorithm with two key sub-modules, as shown in Figures 2(b) and (c), to bridge the latent space gap and the resolution gap, and to unite the denoising process in the same space. + +Bridging the Latent Space Gap. To bridge the latent space gap between two different latent spaces A and B, we propose to unify the model prediction in latent space A by transforming the model prediction $\epsilon_t^\prime$ in latent space B to latent space A using the image space as an intermediate. This is done in the following way: first, we predict the clean data $\hat{z}_{0,t}^{\prime}$ using Equation (3) as: + +$$ +\dot {z} _ {0, t} ^ {\prime} = \frac {1}{\sqrt {\bar {\alpha} _ {t}}} (z _ {t} ^ {\prime} - \sqrt {1 - \bar {\alpha} _ {t}} \epsilon_ {t} ^ {\prime}), \tag {4} +$$ + +which is then decoded into a pixel-level image $\hat{x}_{0,t}^{\prime}$ using the latent decoder model $D^{\prime}$ . This image is encoded into latent space A using the image encoder model $E$ , as $\tilde{z}_{0,t} = E(\hat{x}_{0,t}^{\prime})$ , and finally transformed into a model prediction by inverting Equation (3) as: + +$$ +\tilde {\epsilon} _ {t} = \frac {1}{\sqrt {1 - \bar {\alpha} _ {t}}} \left(z _ {t} - \sqrt {\bar {\alpha} _ {t}} \tilde {z} _ {0, t}\right). \tag {5} +$$ + +With the united $\tilde{\epsilon}_t$ , we can now perform multi-control fusion between $\tilde{\epsilon}_t$ and $\epsilon_t$ (the prediction from model $\epsilon_{\theta}$ with $z_{t}$ in latent space A, omitted in Figure 2 for brevity). In this paper, we adopt the fusion method proposed in Composable-Diffusion [19] as: $\epsilon_{t,fuse} = d\cdot \tilde{\epsilon}_t + (1 - d)\cdot \epsilon_t$ , where $d$ and $1 - d$ are the guidance strengths of each model with $d\in [0,1]$ , to guide the denoising process jointly with these two models for multi-control image generation. Algorithm 1 further illustrates this unification and fusion process. + +Bridging Resolution Gap. To integrate the denoising processes of a low-resolution model with a high-resolution model, upsampling and/or downsam + +![](images/41797f8253c6970c94b7ba047045df5204348de442f8bf20547e0ab89ec42daf.jpg) +(a) Upsampling from intermediate $z_{t}$ . + +![](images/f97620877ba2582d5e0440664a4d1cb1668cfa21e2e9ec676fb290d9c023b68d.jpg) +(b) Our upsampling algorithm. +Fig. 3: Results of fusing a low-resolution model and a high-resolution model with different upsampling methods. Upsampling from intermediate $z_{t}$ results in severe artifacts, while our upsampling algorithm results in high-fidelity image. + +Algorithm 1 Coop-Diffusion: Multi-Diffusion Fusing +Sub-Module 1. Bridging Latent Space Gap +Input: random noise $z_{T} \sim \mathcal{N}(0, I)$ , diffusion model $\epsilon_{\theta}$ , decoder $D$ , encoder $E$ in latent space A; random noise $z_{T}' = z_{T}$ , diffusion model $\epsilon_{\theta}'$ , decoder $D'$ , encoder $E'$ in latent space B; guidance strength $d$ , sampling method $S$ . +1: for $t = T, \ldots, 1$ do +2: $\epsilon_t' = \epsilon_\theta'(z_t')$ , $\hat{z}_{0,t}' = \frac{1}{\sqrt{\alpha_t}} (z_t' - \sqrt{1 - \bar{\alpha}_t} \epsilon_t')$ +3: $\hat{x}_{0,t}' = D'(\hat{z}_{0,t}')$ , $\tilde{z}_{0,t} = E(\hat{x}_{0,t}')$ +4: $\tilde{\epsilon}_t = \frac{1}{\sqrt{1 - \bar{\alpha}_t}} (z_t - \sqrt{\bar{\alpha}_t} \tilde{z}_{0,t})$ +5: $\epsilon_t = \epsilon_\theta(z_t)$ , $\epsilon_{t,fuse} = d \cdot \tilde{\epsilon}_t + (1 - d) \cdot \epsilon_t$ +6: $z_{t-1} = S(z_t, t, \epsilon_{t,fuse})$ +7: $z_{t-1}' = S(z_t', t, \epsilon_{t,fuse}') \triangleright \epsilon_{t,fuse}'$ from $\epsilon_{t,fuse}'$ similar to the process from $\epsilon_t'$ to $\tilde{\epsilon}_t$ , omitted for brevity +8: end for +9: return $D(z_0)$ +Sub-Module 2. Bridging Resolution Gap +Input: diffusion model $\epsilon_{\theta}$ , decoder $D$ , encoder $E$ in high-resolution space; random noise $z_T' \sim \mathcal{N}(0, I)$ , diffusion model $\epsilon_{\theta}'$ , decoder $D'$ , encoder $E'$ in low-resolution space; low-resolution sampling end step $T_{low}$ , sampling method $S$ . +1: for $t = T, \ldots, T_{low} + 1$ do +2: $\epsilon_t' = \epsilon_\theta'(z_t')$ , $z_{t-1}' = S(z_t', t, \epsilon_t')$ +3: end for +4: $\hat{z}_{0,T_{low}}' = \frac{1}{\sqrt{\bar{\alpha}_{T_{low}}}} (z_{T_{low}}' - \sqrt{1 - \bar{\alpha}_{T_{low}}} \epsilon_{T_{low}}')$ +5: $\hat{x}_{0,T_{low}}' = D'(\hat{z}_{0,T_{low}}')$ , $\hat{x}_{0,T_{low}} = Upsample(\hat{x}_{0,T_{low}}')$ +6: $\hat{z}_{0,T_{low}} = E(\hat{x}_{0,T_{low}})$ +7: $z_{T_{low}} = \sqrt{\bar{\alpha}_{T_{low}}} \hat{z}_{0,T_{low}} + \sqrt{1 - \bar{\alpha}_{T_{low}}} \epsilon, \epsilon \sim \mathcal{N}(0, I)$ +8: for $t = T_{low}, \ldots, 1$ do +9: $\epsilon_t = \epsilon_\theta(z_t)$ , $z_{t-1} = S(z_t, t, \epsilon_t)$ +10: end for +11: return $D(z_0)$ + +pling is necessary. Traditional bilinear upsampling, often applied to the intermediate result $z_{t}$ during the denoising process, can undesirably amplify pixel correlation. This amplification deviates from the initial Independent and Identically Distributed (IID) assumption, leading to severe artifacts in the final images, as shown in Figure 3(a). Conversely, downsampling does not present this issue. To address the IID issue in upsampling, we propose a new upsampling algorithm that preserves the IID assumption, thereby bridging the resolution gap between models with different pre-trained resolutions. + +Figure 2(c) visualizes our upsampling algorithm. Specifically, for a low-resolution $z_{t}^{\prime}$ , we use the image space as an intermediate space to transform $z_{t}^{\prime}$ in low-resolution space into high-resolution space as $\tilde{z}_{t}$ . We first predict the noise $\epsilon_{t}^{\prime}$ with the denoising model $\epsilon_{\theta}^{\prime}$ and then predict the clean data $\hat{z}_{0,t}^{\prime}$ as described in Eq. 4. This is decoded into an image $\hat{x}_{0,t}^{\prime}$ using decoder $D^{\prime}$ . We then perform upsampling on $\hat{x}_{0,t}^{\prime}$ to obtain its high-resolution counterpart $\hat{x}_{0,t}$ . Finally, $\hat{x}_{0,t}$ is encoded into the latent space with encoder $E$ as $\hat{z}_{0,t}$ , and $t$ -step noise is added to get the final result $\tilde{z}_{t}$ using Eq. 1. + +With the unified $\tilde{z}_t$ , we can now perform multi-resolution fusion. First, we denoise with a low-resolution model to obtain the intermediate $z_t'$ and its high- + +resolution counterpart $\tilde{z}_t$ . Then, we perform denoising with a high-resolution model starting from $\tilde{z}_t$ , and vice versa. This approach allows us to conduct one-stage super-resolution without undergoing all the low-resolution denoising steps, thereby improving inference efficiency. Algorithm 1 further illustrates this unification and fusion process. + +# 5 Experiments + +Implementation Details. We adopt the pretrained Variational Autoencoder (VAE) model from SDXL [22], and we build our structure and texture generator based on the architecture of its U-Net model with the following modifications. To achieve bilingual text-to-image generation (Chinese and English), we pretrain a Chinese text encoder [9, 36] on our Chinese training dataset. We then concatenate the text embeddings from this Chinese text encoder with those from a pretrained English text encoder, serving as the final text embeddings for the denoising models. For multi-resolution image generation, we select a range of image resolutions around $1024 \times 1024$ and further condition the denoising model on the sinusoidal positional embeddings corresponding to the index of image resolutions. The $T_{struct}$ parameter is set to 500, as suggested by our ablation study. + +Dataset Construction. To encompass the abundant concepts in the world, we collect images in various styles from multiple sources, including Noah-Wukong [9], LAION [27], and others, such as photography, cartoons, portraits, and gaming assets. The collected images are filtered based on CLIP score, aesthetic score, watermark presence, resolution, and aspect ratio. To improve the semantic alignment of PanGu-Draw, we discard parts of the noisy captions that are meaningless or mismatched to the image, sourced from the Internet. Instead, we recapture the collected images by first employing an open-vocabulary detector [35] to locate the primary subjects within the images. These subjects are then processed by LLaVA [18], a high-performance vision-language model, along with prompting templates, to yield detailed image descriptions. These English annotations are subsequently translated into Chinese. + +Evaluation Metrics. We evaluate PanGu-Draw's text-to-image generation on COCO [17] with 30k images for English, and COCO-CN [16] with 10k images for Chinese. The Frechet Inception Distance (FID [10]) is utilized to evaluate image quality and diversity. For Chinese, additional metrics include the Inception Score (IS [30]) and CN-CLIP-score [34], assessing image quality and text-image alignment. Besides, a user study is conducted to evaluate image-text alignment, fidelity, and aesthetics using ImageEval-prompt1 across 339 prompts. + +# 5.1 Text-to-Image Generation + +Evaluation on COCO. As shown in Table 1, PanGu-Draw achieves a FID of 7.99, which is superior to compared methods such as DALL-E 2 and SDXL. It also achieves competitive FID with SOTA methods, indicating the effectiveness + +Table 1: Comparisons of PanGu-Draw with recent representative English text-to-image generation models on COCO dataset in terms of FID. + +
MethodFID↓Model SizeRelease
DALL-E [25]27.5012BN
LDM [26]12.631.5BY
GLIDE [20]12.245BN
SDXL [22]11.932.5BY
PixArt-α [4]10.650.6BY
DALL-E 2 [24]10.395.5BN
Imagen [29]7.273BN
RAPHAEL [33]6.613BN
PanGu-Draw7.995BY
+ +Table 2: Comparisons of PanGu-Draw with Chinese text-to-image generation models on COCO-CN dataset in terms of FID, IS and CN-CLIP-score. The classifier-free guidance scales are set as 9 following AltDiffusion [37]. + +
ModelFID↓IS↑CN-CLIP-score↑
AltDiffusion [37]25.3129.1635.12
Taiyi-Bilingual [38]24.6134.2932.26
Taiyi-CN [38]23.9934.2934.22
PanGu-Draw21.8137.0036.62
+ +Table 3: Results of a User study on ImageVal-prompt in terms of image-text alignment, image fidelity, and aesthetics. + +
MethodAlign↑Fidelity↑Aesthetics↑Ave↑
DALL-E 3 [2]4.724.594.764.69
MJ 5.24.634.544.754.64
SDXL [22]4.414.374.594.46
SD [26]4.173.994.204.12
PanGu-Draw4.54.524.724.58
+ +of our time-decoupling training strategy and its outstanding data and training efficiencies. Our 5B PanGu model is the best-released model in terms of FID. + +Evaluation on COCO-CN. As shown in Table 2, PanGu-Draw outperforms other released Chinese text-to-image models, including Taiyi-CN, Taiyi-Bilingual, and AltDiffusion, across all three metrics. This performance highlights PanGu-Draw's exceptional Chinese text-to-image generation capabilities and the effectiveness of our bilingual text encoder architecture. + +User Study. We conducted a user study to compare PanGu-Draw with top-performing methods, including SDXL [22], Midjourney 5.2, and DALL-E 3 [2]. As shown in Table 3, PanGu-Draw achieves better results than SD and SDXL across all three metrics. It also attains approximately $99\% / 98\%$ of the performance of + +![](images/2201a44d5e082f7f387e156237a94b6d62479a7b00fd67e29d751a18ac5fa91f.jpg) +一位老年女性，身着优雅西服，佩戴菲黛丝帽，身穿夹紧对联鞋。伸出手臂，她的背景是夕阳的校园：“一个 young woman, wearing a elegant gown and traditional dress, carrying a heavy bag of luggage and arms, with the sunset on campus in the background.” + +![](images/88a7bac765b94cc8235d430ba29e851f16ba26e1044b25bd561a6550234ae2da.jpg) +一面像微笑的女子，身穿白色针织，红色长衣在阳光下翩翩生舞，面庞清新，风姿绰约，细尘一片。“A smiling woman wearing a white T-shirt and a red blouse with a black skirt, a red blouse, a red blouse, a red blouse, a red blouse, a red blouse, a red blouse, a red blouse, a red blouse, a red blouse, a red blouse, a red blouse, a red blouse, a red blouse, a red blouse, a red blouse, a red blouse, a red blouse, a red blouse, a red blouse, a red blouse, a red blouse, a red blouse, a red blouse, a red blouse, a red blouse, a red dress, a red dress, a red dress, a red dress, a red dress, a red dress, a red dress, a red dress, a red dress, a red dress, a red dress, a red dress, a red dress, a red dress + +![](images/7fdf2a7a3ac4de0e0c409df4c9fd0d2f94eb60f9d4229340e33de5f392f93325.jpg) +“一个巨大的水球体，内部蕴含着一个微型雨林，雨林中翅膀翱翔，阳光透过树洒洒洒。”“Auge crystal ball contains a miniature rainforest inside, with butterflies flying in the air, and sunlight shining through the leaves.” + +![](images/9d8b18b9bb763e9e709a4d4d81227bed8a74d9800437ae841b1863b2805558f8.jpg) +“一只穿着中世纪短发的兔子，手长颈尖站在一座古董城的城墙上，背后是满目的杂草。” +"A rabbit wearing medieval armor and holding a sword stands on the wall of the tower, with a blade with the setting sun behind him." + +![](images/2eeafd652288659ce2f5e1d7c7e8f27eaff1ea04694c96eabb732f80ab7d24f4.jpg) +“赛博办风面摄影的抓机人，依托在夜空飞行平安，以给予水面文明风貌，具有强烈的光影效果。在2019年北京冬奥会上，运动员们穿梭在的night sky, presented in particle ink painting style, with strong light and shadow effects." + +![](images/81b252e3489c342a012d7293b1eaedf5056fa2a236ccb9593b24496e0e236f0a.jpg) +“旅行者们正奔在光激染湖的湖面上，周围是雄伟的山脉，以中国水墨画风格描绘，画面也彩淡雅致，具有古朴诗志。”（1980年） +“旅行者们正奔在湖面上，周围是美丽的山林，以自然、静谧、清新、宁静、和谐、庄严、庄严、庄严、庄严、庄严、庄严、庄严、庄严、庄严、庄严、庄严、庄严、庄严、庄严、庄严、庄严、庄严、庄严、庄严、庄严、庄严、庄严、庄严、庄严、庄严、庄严、庄严、庄严、庄严、庄严、庄严、庄严、庄严、庄严、庄严、庄严、庄严、庄严、庄严、庄严、庄严、庄严、庄严、庄严、庄严、庄严、庄严、庄严、庄严、庄严、庄重的白色。 + +![](images/78246ef7da538a87b6398557cb88e04794f8bdc28b5f9a41f07772874ce19e1e.jpg) +“一台未来风格的摩托车，闪耀着霓虹灯，停在夜晚的东京街头。” +“A futuristic motorcycle, shining with neon lights, is parked on the streets of Tokyo at night.” +Fig. 4: Images generated with PanGu-Draw, our 5B multi-lingual text-to-image generation model. PanGu-Draw is able to generate multi-resolution high-fidelity images semantically aligned with the input prompts. + +![](images/dda03bb00997887131d6f8ce25a19e7cfc341b1dae790a6de919503eabdffd62.jpg) +一座由冰晶和雪花构成的精致城堡，坐落在在北极的冰原上。 +“An exquisite castle made of ice crystals and snowflakes, located on the Arctic ice sheet.” + +![](images/4e5071405aae96c871b8e94442ea96f379d0f4d60c8ce883e167b9323215585a.jpg) +一袋老火的糖盔，完全由糖果和巧克力制成。一个 ancient pot was filled of candy and chocolate. + +Midjourney 5.2 and DALL-E 3, respectively, indicating PanGu-Draw's excellent text-to-image capabilities. Figure 4 shows some high-fidelity multi-resolution images generated by PanGu-Draw. As we can see, the generated images of PanGu-Draw are of high aesthetics and semantically aligned with the input prompts. + +# 5.2 Multi-Diffusion Fusing Results + +Multi-Control Image Generation. To demonstrate the effectiveness of the proposed reusable multi-diffusion fusing algorithm, Coop-Diffusion, we first present multiple results of multi-control image generation. Figure 5 displays results from fusing an image variation model $^2$ with PanGu-Draw. The fusing results maintain a style similar to that of the reference image, matching the texture described by the input prompt. Figure 6 shows results from fusing PanGu-Draw with a pose/edge-to-image ControlNet model, which operates in guess mode without input prompts. Here, the fusing results combine the structure of the pose/edge image with the texture described by the input prompt. + +![](images/7adbcae4a56790853f9fae971c5ffb910d46fb82f33d32a5639812f891912d04.jpg) +(a) Input image and generation results of image variation. + +![](images/0f0e5f856510b0ed5aaff8f431bc6d24ad8fbe5d7825e09f232bfdbb36637d71.jpg) + +![](images/bc637749dfbb0cca22b06351b7c54cd300f3577a43dbb47c3f8f617e1b5a7733.jpg) + +![](images/c4d7506482d107c66fda628776989c67a3519d04d2f6efd960a1a096806b99f6.jpg) +(b) Input prompt and generation results of PanGu-Draw. + +![](images/7f225c96c36d5735ec6785e7c15c7612048b2fb49e3abc9830d9ad078d997480.jpg) +(c) Generation results of fusing image variation and PanGu-Draw with Coop-Diffusion algorithm. + +![](images/5b2abb2a30c0943316d2a6728ea7d6f55a1a4ccaa7b2c866e9cf5d94c53d915f.jpg) +Fig. 5: Generation results of the fusing of an image variation model and PanGu-Draw and with the proposed Coop-Diffusion algorithm. + +![](images/8417a5359dcaf7329a496568c9e1950911c95755f385dfc9f541411a150d953a.jpg) +(b) Input prompt and generation results of PanGu-Draw. + +![](images/edb622cc78a1a3d9f5df08900d8b98db1466a4cb1981153de1d85351f8f86a99.jpg) +(a) Input edge map and generation results of edge-to-image + +![](images/92dbdb4fb327017281fde2c184adfe6856c0d6a88d7633492ae3453cc52a8b39.jpg) +(b) Input prompt and generation results of PanGu-Draw. +Fig. 6: Generation results guided by fusing signals of text and pose/edge map by our Coop-Diffusion. + +![](images/9e7f9916bc96a6607b1de3f6b2ce885429ecfef8dd4b009046a9cac1d3d183f8.jpg) +(c) Generation results of fusing pose-to-image and PanGu-Draw with Coop-Diffusion algorithm. + +![](images/1eec92eebddea7cf18c9ea5ccfb8d7f32b866c6cc03741855dd723db25d1d1f1.jpg) +(c) Generation results of fusing edge-to-image and PanGu-Draw with CoopDiffusion algorithm. + +Multi-Resolution Image Generation. We also present multi-resolution image generation results of fusing PanGu-Draw with low-resolution text-to-image and edge-to-image ControlNet model by first denoising with the low-resolution model to get the intermediate $z_{t}^{\prime}$ and the high-resolution counterpart $\tilde{z}_{t}$ , and then perform denoising in high resolution with PanGu-Draw. Figure 7 shows the results from the low-resolution model and our fusing algorithm Coop-Diffusion. As we can see, PanGu-Draw adds much details to the low-resolution predictions + +![](images/e0bcc639f1e4855ccf61fed739fa729c95380f9dcad585e1c9d243b10c6105c5.jpg) + +![](images/e0bc84a0d5e802d8500382c3c483ea9d1359513aab38ab16f640ff6ccef25bab.jpg) + +![](images/96ae32efe5cb26a9e555f8471c91d3dcc1eac5e8dfe36b104dec01b847a1951d.jpg) +LR model + +![](images/c50b523831a80171738df1ab3f8872dbaaf9291deeff5220f4380bcc71a9afb8.jpg) +LR model + HR model +Fig. 7: Images generated with a low-resolution (LR) model (1st row: T2I model; 2nd row: edge-to-image ControlNet) and the fusion of the LR model and HR PanGu-Draw with our Coop-Diffusion. The resolutions are $512 \times 512$ and $1024 \times 1024$ respectively. This allows for single-stage super-resolution for better details and higher inference efficiency. + +leading to high-fidelity high-resolution results. Besides, compared with the common practice of super-resolution with diffusion model, which carries out all the low-resolution denoising steps, our method achieve higher inference efficiency. + +# 5.3 Ablation Study + +In this section, we perform ablation studies to analyze our time-decoupling training strategy. The baseline model has $1B$ parameters while the structure and texture generators both have $0.5B$ parameters. During the training process, the latter two models only train half the steps of the baseline model with $T_{struct}$ set as 500. Both settings of the models are trained from scratch on a subset of the LAION dataset containing images with all sizes. After training, FID, IS and CLIP-score on COCO are reported for comparison. + +Time-Decoupling Training Strategy. We compare the performance of models trained with the Resolution Boost strategy and our time-decoupling strategy in Table 4. We found that models trained with our strategy achieves better performance in all three criteria, indicating the effectiveness of our strategy. + +Training Designs. The structure and texture generators ( $\epsilon_{\mathrm{struct}}$ and $\epsilon_{\mathrm{texture}}$ ) are designed to train on different resolutions to improve data and training efficiency. However, this approach may negatively influence the final performance. In Table 5, we compare such a design with a traditional training process, where $\epsilon_{\mathrm{struct}}$ discards low-resolution images, or $\epsilon_{\mathrm{texture}}$ trains with high resolution. Results on COCO show that $\epsilon_{\mathrm{struct}}$ benefits from these extra up-scaled data, and $\epsilon_{\mathrm{texture}}$ learns enough texture patterns at a smaller resolution. + +Table 4: Comparison of models across Resolution Boost (1B parameters) and Time-Decoupling training strategies (0.5B parameters for structure and texture generators) + +
ModelFID↓IS↑CLIP-score↑
Resolution Boost106.1210.4622.9
Time-Decoupling87.6611.0723.4
+ +Table 5: Performance of structure and texture models training with images of different resolutions. + +
Structure DataTexture ResolutionFID↓IS↑CLIP-score↑
All data25687.6611.0723.4
Only high resolution25689.5210.9623.2
All data51290.9810.5923.3
+ +Table 6: Comparisons of PanGu-Draw inference performance with different time step splitting point $T_{struct}$ settings. + +
TstructFID↓IS↑CLIP-score↑
200105.0810.5922.98
30098.0810.7223.12
50087.6611.0723.40
70089.4811.0223.32
+ +Timestep Splitting Point. The timestep splitting point $T_{\mathrm{struct}}$ between the structure and texture generators also influences the final performance. To this end, we set $T_{\mathrm{struct}}$ to 200, 300, 500, and 700, while keeping the other settings of the structure and texture generators unchanged. As shown in Table 6, as $T_{\mathrm{struct}}$ increases from 200 to 700, the performance initially increases and then decreases continuously. $T_{\mathrm{struct}} = 500$ is the optimal value, and we adopt it as the default setting in all other experiments. + +# 6 Conclusion + +In this paper, we present "PanGu-Draw", a new latent diffusion model for efficient text-to-image generation that effectively integrates multiple control signals. Our approach includes a Time-Decoupling Training Strategy to separate the text-to-image process into structure and texture generation, enhancing data use and computational efficiency. Additionally, "Coop-Diffusion" is introduced, an algorithm allowing cooperative use of different pre-trained diffusion models in a unified denoising process for multi-control image synthesis at various resolutions without extra data or retraining. PanGu-Draw outperforms models like DALL-E 2 and SDXL in English T2I, achieves superior FID, IS, and CN-CLIP-scores in Chinese T2I, and receives favorable user feedback. This positions PanGu-Draw as a versatile and efficient state-of-the-art method. + +# References + +1. Balaji, Y., Nah, S., Huang, X., Vahdat, A., Song, J., Kreis, K., Aittala, M., Aila, T., Laine, S., Catanzaro, B., et al.: edifi: Text-to-image diffusion models with an ensemble of expert denoisers. arXiv preprint arXiv:2211.01324 (2022) +2. Betker, J., Goh, G., Jing, L., Brooks, T., Wang, J., Li, L., Ouyang, L., Zhuang, J., Lee, J., Guo, Y., et al.: Improving image generation with better captions. Computer Science. https://cdn.opena.com/papers/dall-e-3.pdf (2023) +3. Chen, C., Yin, Y., Shang, L., Jiang, X., Qin, Y., Wang, F., Wang, Z., Chen, X., Liu, Z., Liu, Q.: bert2bert: Towards reusable pretrained language models. arXiv preprint arXiv:2110.07143 (2021) +4. Chen, J., Yu, J., Ge, C., Yao, L., Xie, E., Wu, Y., Wang, Z., Kwok, J., Luo, P., Lu, H., Li, Z.: Fast training of diffusion transformer for photorealistic text-to-image synthesis (2023) +5. Dhariwal, P., Nichol, A.: Diffusion models beat gans on image synthesis. Advances in Neural Information Processing Systems 34, 8780-8794 (2021) +6. Ding, N., Tang, Y., Han, K., Xu, C., Wang, Y.: Network expansion for practical training acceleration. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 20269-20279 (2023) +7. Feng, Z., Zhang, Z., Yu, X., Fang, Y., Li, L., Chen, X., Lu, Y., Liu, J., Yin, W., Feng, S., et al.: Ernie-vilg 2.0: Improving text-to-image diffusion model with knowledge-enhanced mixture-of-denoising-experts. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 10135-10145 (2023) +8. Fu, C., Huang, H., Jiang, Z., Ni, Y., Nai, L., Wu, G., Cheng, L., Zhou, Y., Li, S., Li, A., et al.: Triple: Revisiting pretrained model reuse and progressive learning for efficient vision transformer scaling and searching. In: ICCV (2023) +9. Gu, J., Meng, X., Lu, G., Hou, L., Minzhe, N., Liang, X., Yao, L., Huang, R., Zhang, W., Jiang, X., et al.: Wukong: A 100 million large-scale chinese cross-modal pre-training benchmark. Advances in Neural Information Processing Systems 35, 26418-26431 (2022) +0. Heusel, M., Ramsauer, H., Unterthiner, T., Nessler, B., Hochreiter, S.: Gans trained by a two time-scale update rule converge to a local nash equilibrium. Advances in neural information processing systems 30 (2017) +1. Ho, J., Jain, A., Abbeel, P.: Denoising diffusion probabilistic models. Advances in Neural Information Processing Systems 33, 6840-6851 (2020) +2. Ho, J., Jain, A., Abbeel, P.: Denoising diffusion probabilistic models. Advances in Neural Information Processing Systems 33, 6840-6851 (2020) +3. Ho, J., Sahara, C., Chan, W., Fleet, D.J., Norouzi, M., Salimans, T.: Cascaded diffusion models for high fidelity image generation. J. Mach. Learn. Res. 23, 47-1 (2022) +4. Ho, J., Salimans, T.: Classifier-free diffusion guidance. arXiv preprint arXiv:2207.12598 (2022) +5. Jin, Z., Shen, X., Li, B., Xue, X.: Training-free diffusion model adaptation for variable-sized text-to-image synthesis. arXiv preprint arXiv:2306.08645 (2023) +6. Li, X., Xu, C., Wang, X., Lan, W., Jia, Z., Yang, G., Xu, J.: Coco-cn for cross-lingual image tagging, captioning, and retrieval. IEEE Transactions on Multimedia 21(9), 2347-2360 (2019) +7. Lin, T.Y., Maire, M., Belongie, S., Hays, J., Perona, P., Ramanan, D., Dólar, P., Zitnick, C.L.: Microsoft coco: Common objects in context. In: Computer Vision-ECCV 2014: 13th European Conference, Zurich, Switzerland, September 6-12, 2014, Proceedings, Part V 13. pp. 740-755. Springer (2014) + +18. Liu, H., Li, C., Wu, Q., Lee, Y.J.: Visual instruction tuning. arXiv preprint arXiv:2304.08485 (2023) +19. Liu, N., Li, S., Du, Y., Torralba, A., Tenenbaum, J.B.: Compositional visual generation with composable diffusion models. In: European Conference on Computer Vision. pp. 423-439. Springer (2022) +20. Nichol, A., Dhariwal, P., Ramesh, A., Shyam, P., Mishkin, P., McGrew, B., Sutskever, I., Chen, M.: Glide: Towards photorealistic image generation and editing with text-guided diffusion models. arXiv preprint arXiv:2112.10741 (2021) +21. Nichol, A.Q., Dhariwal, P.: Improved denoising diffusion probabilistic models. In: ICML. pp. 8162-8171. PMLR (2021) +22. Podell, D., English, Z., Lacey, K., Blattmann, A., Dockhorn, T., Müller, J., Penna, J., Rombach, R.: Sdxl: improving latent diffusion models for high-resolution image synthesis. arXiv preprint arXiv:2307.01952 (2023) +23. Qin, Y., Zhang, J., Lin, Y., Liu, Z., Li, P., Sun, M., Zhou, J.: Elle: Efficient lifelong pre-training for emerging data. arXiv preprint arXiv:2203.06311 (2022) +24. Ramesh, A., Dhariwal, P., Nichol, A., Chu, C., Chen, M.: Hierarchical text-conditional image generation with clip latents. arXiv preprint arXiv:2204.06125 (2022) +25. Ramesh, A., Pavlov, M., Goh, G., Gray, S., Voss, C., Radford, A., Chen, M., Sutskever, I.: Zero-shot text-to-image generation. In: International Conference on Machine Learning. pp. 8821-8831. PMLR (2021) +26. Rombach, R., Blattmann, A., Lorenz, D., Esser, P., Ommer, B.: High-resolution image synthesis with latent diffusion models. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR). pp. 10684-10695 (June 2022) +27. Rombach, R., Blattmann, A., Lorenz, D., Esser, P., Ommer, B.: High-resolution image synthesis with latent diffusion models. In: Proceedings of the IEEE/CVF conference on computer vision and pattern recognition. pp. 10684-10695 (2022) +28. Ronneberger, O., Fischer, P., Brox, T.: U-net: Convolutional networks for biomedical image segmentation. In: Medical Image Computing and Computer-Assisted Intervention-MICCAI 2015: 18th International Conference, Munich, Germany, October 5-9, 2015, Proceedings, Part III 18. pp. 234-241. Springer (2015) +29. Sahara, C., Chan, W., Saxena, S., Li, L., Whang, J., Denton, E.L., Ghasemipour, K., Gontijo Lopes, R., Karagol Ayan, B., Salimans, T., et al.: Photorealistic text-to-image diffusion models with deep language understanding. Advances in Neural Information Processing Systems 35, 36479-36494 (2022) +30. Salimans, T., Goodfellow, I., Zaremba, W., Cheung, V., Radford, A., Chen, X.: Improved techniques for training gans. Advances in neural information processing systems 29, 2234-2242 (2016) +31. Shonenkov, A., Konstantinov, M., Bakshandaeva, D., Schuhmann, C., Ivanova, K., Klokova, N.: Deepfloyd if: A powerful text-to-image model that can smartly integrate text into images (2023), https://www deepfloyd.ai/deepfloyd-if, online; accessed 16-November-2023 +32. Xu, X., Wang, Z., Zhang, E., Wang, K., Shi, H.: Versatile diffusion: Text, images and variations all in one diffusion model. arXiv preprint arXiv:2211.08332 (2022) +33. Xue, Z., Song, G., Guo, Q., Liu, B., Zong, Z., Liu, Y., Luo, P.: Raphael: Text-to-image generation via large mixture of diffusion paths. arXiv preprint arXiv:2305.18295 (2023) +34. Yang, A., Pan, J., Lin, J., Men, R., Zhang, Y., Zhou, J., Zhou, C.: Chinese clip: Contrastive vision-language pretraining in chinese. arXiv preprint arXiv:2211.01335 (2022) + +35. Yao, L., Han, J., Liang, X., Xu, D., Zhang, W., Li, Z., Xu, H.: Detclipv2: Scalable open-vocabulary object detection pre-training via word-region alignment. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 23497-23506 (2023) +36. Yao, L., Huang, R., Hou, L., Lu, G., Niu, M., Xu, H., Liang, X., Li, Z., Jiang, X., Xu, C.: Filip: Fine-grained interactive language-image pre-training. arXiv preprint arXiv:2111.07783 (2021) +37. Ye, F., Liu, G., Wu, X., Wu, L.: Altdiffusion: A multilingual text-to-image diffusion model (2023) +38. Zhang, J., Gan, R., Wang, J., Zhang, Y., Zhang, L., Yang, P., Gao, X., Wu, Z., Dong, X., He, J., Zhuo, J., Yang, Q., Huang, Y., Li, X., Wu, Y., Lu, J., Zhu, X., Chen, W., Han, T., Pan, K., Wang, R., Wang, H., Wu, X., Zeng, Z., Chen, C.: Fengshenbang 1.0: Being the foundation of chinese cognitive intelligence. CoRR abs/2209.02970 (2022) +39. Zhang, L., Rao, A., Agrawala, M.: Adding conditional control to text-to-image diffusion models. In: Proceedings of the IEEE/CVF International Conference on Computer Vision. pp. 3836-3847 (2023) \ No newline at end of file diff --git a/2024/PanGu-Draw_ Advancing Resource-Efficient Text-to-Image Synthesis with Time-Decoupled Training and Reusable Coop-Diffusion/images.zip b/2024/PanGu-Draw_ Advancing Resource-Efficient Text-to-Image Synthesis with Time-Decoupled Training and Reusable Coop-Diffusion/images.zip new file mode 100644 index 0000000000000000000000000000000000000000..1c2f0ba1e9194b65f4b110ff8c5c2c3177a825ce --- /dev/null +++ b/2024/PanGu-Draw_ Advancing Resource-Efficient Text-to-Image Synthesis with Time-Decoupled Training and Reusable Coop-Diffusion/images.zip @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:46ecc171a1359649cf9aa376436078516cd53b8316a8933de762c50e9b9d6057 +size 591793 diff --git a/2024/PanGu-Draw_ Advancing Resource-Efficient Text-to-Image Synthesis with Time-Decoupled Training and Reusable Coop-Diffusion/layout.json b/2024/PanGu-Draw_ Advancing Resource-Efficient Text-to-Image Synthesis with Time-Decoupled Training and Reusable Coop-Diffusion/layout.json new file mode 100644 index 0000000000000000000000000000000000000000..2ab07e07849c2ec2a7ae1ccf8c04411d28f00c1b --- /dev/null +++ b/2024/PanGu-Draw_ Advancing Resource-Efficient Text-to-Image Synthesis with Time-Decoupled Training and Reusable Coop-Diffusion/layout.json @@ -0,0 +1,11238 @@ +{ + "pdf_info": [ + { + "para_blocks": [ + { + "bbox": [ + 146, + 111, + 469, + 163 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 146, + 111, + 469, + 163 + ], + "spans": [ + { + "bbox": [ + 146, + 111, + 469, + 163 + ], + "type": "text", + "content": "PanGu-Draw: Advancing Resource-Efficient Text-to-Image Synthesis with Time-Decoupled Training and Reusable Coop-Diffusion" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 144, + 183, + 470, + 209 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 144, + 183, + 470, + 209 + ], + "spans": [ + { + "bbox": [ + 144, + 183, + 470, + 209 + ], + "type": "text", + "content": "Guansong Lu" + }, + { + "bbox": [ + 144, + 183, + 470, + 209 + ], + "type": "inline_equation", + "content": "^{1}" + }, + { + "bbox": [ + 144, + 183, + 470, + 209 + ], + "type": "text", + "content": ", Yuanfan Guo" + }, + { + "bbox": [ + 144, + 183, + 470, + 209 + ], + "type": "inline_equation", + "content": "^{1}" + }, + { + "bbox": [ + 144, + 183, + 470, + 209 + ], + "type": "text", + "content": ", Jianhua Han" + }, + { + "bbox": [ + 144, + 183, + 470, + 209 + ], + "type": "inline_equation", + "content": "^{1}" + }, + { + "bbox": [ + 144, + 183, + 470, + 209 + ], + "type": "text", + "content": ", Minzhe Niu" + }, + { + "bbox": [ + 144, + 183, + 470, + 209 + ], + "type": "inline_equation", + "content": "^{1}" + }, + { + "bbox": [ + 144, + 183, + 470, + 209 + ], + "type": "text", + "content": ", Yihan Zeng" + }, + { + "bbox": [ + 144, + 183, + 470, + 209 + ], + "type": "inline_equation", + "content": "^{1}" + }, + { + "bbox": [ + 144, + 183, + 470, + 209 + ], + "type": "text", + "content": ", Songcen Xu" + }, + { + "bbox": [ + 144, + 183, + 470, + 209 + ], + "type": "inline_equation", + "content": "^{1}" + }, + { + "bbox": [ + 144, + 183, + 470, + 209 + ], + "type": "text", + "content": ", Zeyi Huang" + }, + { + "bbox": [ + 144, + 183, + 470, + 209 + ], + "type": "inline_equation", + "content": "^{2}" + }, + { + "bbox": [ + 144, + 183, + 470, + 209 + ], + "type": "text", + "content": ", Zhao Zhong" + }, + { + "bbox": [ + 144, + 183, + 470, + 209 + ], + "type": "inline_equation", + "content": "^{2}" + }, + { + "bbox": [ + 144, + 183, + 470, + 209 + ], + "type": "text", + "content": ", Wei Zhang" + }, + { + "bbox": [ + 144, + 183, + 470, + 209 + ], + "type": "inline_equation", + "content": "^{1}" + }, + { + "bbox": [ + 144, + 183, + 470, + 209 + ], + "type": "text", + "content": ", and Hang Xu" + }, + { + "bbox": [ + 144, + 183, + 470, + 209 + ], + "type": "inline_equation", + "content": "^{1\\star}" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 233, + 216, + 381, + 228 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 233, + 216, + 381, + 228 + ], + "spans": [ + { + "bbox": [ + 233, + 216, + 381, + 228 + ], + "type": "text", + "content": "1 Huawei Noah's Ark Lab 2 Huawei" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 160, + 252, + 455, + 517 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 160, + 252, + 455, + 517 + ], + "spans": [ + { + "bbox": [ + 160, + 252, + 455, + 517 + ], + "type": "text", + "content": "Abstract. Current large-scale diffusion models represent a giant leap forward in conditional image synthesis, capable of interpreting diverse cues like text, human poses, and edges. However, their reliance on substantial computational resources and extensive data collection remains a bottleneck. On the other hand, the integration of existing diffusion models, each specialized for different controls and operating in unique latent spaces, poses a challenge due to incompatible image resolutions and latent space embedding structures, hindering their joint use. Addressing these constraints, we present \"PanGu-Draw\", a novel latent diffusion model designed for resource-efficient text-to-image synthesis that adeptly accommodates multiple control signals. We first propose a resource-efficient Time-Decoupling Training Strategy, which splits the monolithic text-to-image model into structure and texture generators. Each generator is trained using a regimen that maximizes data utilization and computational efficiency, cutting data preparation by " + }, + { + "bbox": [ + 160, + 252, + 455, + 517 + ], + "type": "inline_equation", + "content": "48\\%" + }, + { + "bbox": [ + 160, + 252, + 455, + 517 + ], + "type": "text", + "content": " and reducing training resources by " + }, + { + "bbox": [ + 160, + 252, + 455, + 517 + ], + "type": "inline_equation", + "content": "51\\%" + }, + { + "bbox": [ + 160, + 252, + 455, + 517 + ], + "type": "text", + "content": ". Secondly, we introduce \"Coop-Diffusion\", an algorithm that enables the cooperative use of various pre-trained diffusion models with different latent spaces and predefined resolutions within a unified denoising process. This allows for multi-control image synthesis at arbitrary resolutions without the necessity for additional data or retraining. Empirical validations of Pangu-Draw show its exceptional prowess in text-to-image and multi-control image generation, suggesting a promising direction for future model training efficiencies and generation versatility. The largest 5B T2I PanGu-Draw model is released on the Ascend platform. Project page: https://pangu- draw.github.io" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 160, + 525, + 462, + 537 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 160, + 525, + 462, + 537 + ], + "spans": [ + { + "bbox": [ + 160, + 525, + 462, + 537 + ], + "type": "text", + "content": "Keywords: Text-to-Image Synthesis " + }, + { + "bbox": [ + 160, + 525, + 462, + 537 + ], + "type": "inline_equation", + "content": "\\cdot" + }, + { + "bbox": [ + 160, + 525, + 462, + 537 + ], + "type": "text", + "content": " Resource-Efficient " + }, + { + "bbox": [ + 160, + 525, + 462, + 537 + ], + "type": "inline_equation", + "content": "\\cdot" + }, + { + "bbox": [ + 160, + 525, + 462, + 537 + ], + "type": "text", + "content": " Diffusion Model" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 132, + 555, + 230, + 567 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 555, + 230, + 567 + ], + "spans": [ + { + "bbox": [ + 132, + 555, + 230, + 567 + ], + "type": "text", + "content": "1 Introduction" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 130, + 577, + 482, + 648 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 577, + 482, + 648 + ], + "spans": [ + { + "bbox": [ + 130, + 577, + 482, + 648 + ], + "type": "text", + "content": "The Denoising Diffusion Probabilistic Models (DDPMs) [11] and their subsequent enhancements [5, 14, 21] have established diffusion models as a leading approach for image generation. These advancements excel in the application of diffusion models to text-to-image synthesis, yielding high-fidelity results with large-scale models and datasets, supported by substantial computational resources [13, 20, 24, 26, 29]. These foundational models, capable of understanding" + } + ] + } + ], + "index": 6 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 133, + 653, + 326, + 665 + ], + "type": "page_footnote", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 653, + 326, + 665 + ], + "spans": [ + { + "bbox": [ + 133, + 653, + 326, + 665 + ], + "type": "text", + "content": "* Corresponding author: xuHang@huawei.com" + } + ] + } + ], + "index": 7 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 0 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 160, + 113, + 460, + 242 + ], + "blocks": [ + { + "bbox": [ + 160, + 113, + 460, + 242 + ], + "lines": [ + { + "bbox": [ + 160, + 113, + 460, + 242 + ], + "spans": [ + { + "bbox": [ + 160, + 113, + 460, + 242 + ], + "type": "image", + "image_path": "0cb2654b4516e86a69532b16ceedc69bbf26fbaba52d2b1bace1ef47193282f1.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 130, + 251, + 480, + 297 + ], + "lines": [ + { + "bbox": [ + 130, + 251, + 480, + 297 + ], + "spans": [ + { + "bbox": [ + 130, + 251, + 480, + 297 + ], + "type": "text", + "content": "Fig. 1: Illustration of three multi-stage training strategies and comparison between them in resource efficiency in data, training and inference aspects. Our time-decoupling training strategy significantly surpasses the representative methods in Cascaded Training [20,31] and Resolution Boost Training [26,37] in resource efficiency." + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_caption" + } + ], + "index": 2 + }, + { + "bbox": [ + 130, + 320, + 479, + 354 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 320, + 479, + 354 + ], + "spans": [ + { + "bbox": [ + 130, + 320, + 479, + 354 + ], + "type": "text", + "content": "and rendering complex semantics, have paved the way for diverse image generation tasks, accommodating various control signals such as reference images, edges [39], and poses [39]." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 130, + 355, + 480, + 527 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 355, + 480, + 527 + ], + "spans": [ + { + "bbox": [ + 130, + 355, + 480, + 527 + ], + "type": "text", + "content": "However, the extensive computational demand and significant data collection required by these models pose a substantial challenge. The ambitious goal of higher fidelity and increased resolution in image synthesis pushes the boundaries of model and dataset sizes, escalating computational costs, and environmental impact. Moreover, the aspiration for versatile control and multi-resolution in image generation introduces additional complexity. Existing diffusion models, each tailored for specific controls and operating within distinct latent spaces, face the challenge of integration due to incompatible image resolutions and latent space embeddings, obstructing their concurrent utilization. This incompatibility not only leads to more resource consumption of retraining but also impedes the joint synthesis of images controlled by multiple factors, thereby limiting the scalability and practical application of such existing generative models. In response to these challenges, our work introduces a novel paradigm named \"PanGu-Draw\" that judiciously conserves training resources while enhancing data efficiency, thereby proposing a resource-efficient pathway forward for diffusion model scalability." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 130, + 528, + 480, + 596 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 528, + 480, + 596 + ], + "spans": [ + { + "bbox": [ + 130, + 528, + 480, + 596 + ], + "type": "text", + "content": "As shown in Figure 1, the training strategies of predecessors like DeepFloyd [31] and GLIDE [20], which employ a cascaded approach, excel in leveraging data across resolutions but suffer from inefficient inference due to their reliance on multiple models. Alternatively, Stable Diffusion [26] and AltDiffusion [37] use a Resolution Boost Training strategy aiming for cost-effectiveness by refining a single model. However, this strategy falls short on data efficiency." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 130, + 596, + 481, + 665 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 596, + 481, + 665 + ], + "spans": [ + { + "bbox": [ + 130, + 596, + 481, + 665 + ], + "type": "text", + "content": "In light of these considerations, our PanGu-Draw framework advances the field by presenting a Time-Decoupling Training Strategy that segments the training of a comprehensive text-to-image model into two distinct generators: one dedicated to structural outlines and another to textural details. This division not only concentrates on training efforts but also enhances data efficacy. The structural generator is adept at crafting the initial outlines of images, offering" + } + ] + } + ], + "index": 7 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "spans": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "type": "text", + "content": "2" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 166, + 91, + 215, + 100 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 166, + 91, + 215, + 100 + ], + "spans": [ + { + "bbox": [ + 166, + 91, + 215, + 100 + ], + "type": "text", + "content": "G. Lu et al." + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 1 + }, + { + "para_blocks": [ + { + "bbox": [ + 130, + 116, + 482, + 196 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 116, + 482, + 196 + ], + "spans": [ + { + "bbox": [ + 130, + 116, + 482, + 196 + ], + "type": "text", + "content": "flexibility in data quality and enabling training across a spectrum of data calibers; the textural generator, in contrast, is fine-tuned using low-resolution data to infuse these outlines with fine-grained details, ensuring optimal performance even during high-resolution synthesis. This focused approach not only accelerates the training process of our 5B model but also significantly reduces the reliance on extensive data collection and computational resources, as evidenced by a " + }, + { + "bbox": [ + 130, + 116, + 482, + 196 + ], + "type": "inline_equation", + "content": "48\\%" + }, + { + "bbox": [ + 130, + 116, + 482, + 196 + ], + "type": "text", + "content": " reduction in data preparation and a " + }, + { + "bbox": [ + 130, + 116, + 482, + 196 + ], + "type": "inline_equation", + "content": "51\\%" + }, + { + "bbox": [ + 130, + 116, + 482, + 196 + ], + "type": "text", + "content": " reduction in resource consumption." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 130, + 197, + 482, + 392 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 197, + 482, + 392 + ], + "spans": [ + { + "bbox": [ + 130, + 197, + 482, + 392 + ], + "type": "text", + "content": "Furthermore, we introduce a pioneering algorithm named Coop-Diffusion, which facilitates the cooperative integration of diverse pre-trained diffusion models. Each model, conditioned on different controls and pre-defined resolutions, contributes to a seamless denoising process. The first algorithmic sub-module addresses inconsistencies in VAE decoders that arise during the denoising process across different latent spaces, ensuring cohesive image quality by effectively reconciling disparate latent space representations. The second sub-module confronts the challenges associated with multi-resolution denoising. Traditional bilinear upsampling for the intermediate noise map, introduced during the denoising process, can undesirably amplify the correlation between pixels. This amplification deviates from the initial Independent and Identically Distributed (IID) assumption, leading to severe artifacts in the final output image. However, our innovative approach circumvents this issue with a single-step sampling method that preserves the integrity of pixel independence, thus preventing the introduction of artifacts. Coop-Diffusion obviates the need for additional data or model retraining, addressing the challenges of multi-control and multi-resolution image generation with scalability and efficiency." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 130, + 392, + 482, + 449 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 392, + 482, + 449 + ], + "spans": [ + { + "bbox": [ + 130, + 392, + 482, + 449 + ], + "type": "text", + "content": "PanGu-Draw excels in text-to-image (T2I) generation, outperforming established models like DALL-E 2 and SDXL, as evidenced by its FID of 7.99 in English T2I. It also leads in Chinese T2I across metrics like FID, IS, and CNCLIP-score. User feedback highlights a strong preference for PanGu-Draw, aligning well with human visual perceptions." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 146, + 449, + 343, + 460 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 146, + 449, + 343, + 460 + ], + "spans": [ + { + "bbox": [ + 146, + 449, + 343, + 460 + ], + "type": "text", + "content": "In summary, our contributions are manifold:" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 131, + 460, + 480, + 552 + ], + "type": "list", + "angle": 0, + "index": 9, + "blocks": [ + { + "bbox": [ + 131, + 460, + 480, + 483 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 131, + 460, + 480, + 483 + ], + "spans": [ + { + "bbox": [ + 131, + 460, + 480, + 483 + ], + "type": "text", + "content": "- PanGu-Draw: A resource-efficient diffusion model with a Time-Decoupling Training Strategy, reducing data and training resources for T2I synthesis." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 131, + 484, + 480, + 517 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 131, + 484, + 480, + 517 + ], + "spans": [ + { + "bbox": [ + 131, + 484, + 480, + 517 + ], + "type": "text", + "content": "- **Coop-Diffusion:** A novel approach for integrating multiple diffusion models, enabling efficient multi-control image synthesis at multi-resolutions within a unified denoising process." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 131, + 518, + 480, + 552 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 131, + 518, + 480, + 552 + ], + "spans": [ + { + "bbox": [ + 131, + 518, + 480, + 552 + ], + "type": "text", + "content": "- Comprehensive evaluations demonstrate PanGu-Draw's (5B model) can produce high-quality images aligned with text and various controls, advancing the scalability and flexibility of diffusion-based image generation." + } + ] + } + ], + "index": 8 + } + ], + "sub_type": "text" + }, + { + "bbox": [ + 132, + 571, + 237, + 583 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 571, + 237, + 583 + ], + "spans": [ + { + "bbox": [ + 132, + 571, + 237, + 583 + ], + "type": "text", + "content": "2 Related Work" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 130, + 596, + 482, + 666 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 596, + 482, + 666 + ], + "spans": [ + { + "bbox": [ + 130, + 596, + 482, + 666 + ], + "type": "text", + "content": "Text-to-Image Generation. The integration of diffusion models into the realm of text-to-image (T2I) generation marks a significant stride in computational creativity [5,7,12,13,20,21,24,26,29,31-33]. Models like GLIDE [20] and DALL-E 2 [24] have significantly advanced in generating diverse and semantically aligned images from textual descriptions. LDM [26] addresses computational challenges by creating images from text-conditioned low-dimensional latent representations." + } + ] + } + ], + "index": 11 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 392, + 91, + 447, + 101 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 392, + 91, + 447, + 101 + ], + "spans": [ + { + "bbox": [ + 392, + 91, + 447, + 101 + ], + "type": "text", + "content": "PanGu-Draw" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 474, + 92, + 481, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 474, + 92, + 481, + 100 + ], + "spans": [ + { + "bbox": [ + 474, + 92, + 481, + 100 + ], + "type": "text", + "content": "3" + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 2 + }, + { + "para_blocks": [ + { + "bbox": [ + 130, + 116, + 479, + 173 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 116, + 479, + 173 + ], + "spans": [ + { + "bbox": [ + 130, + 116, + 479, + 173 + ], + "type": "text", + "content": "ControlNet [39] introduces spatial conditioning controls, offering flexibility in image generation under varied conditions like edges and depth. Despite the proliferation of such specialized models, a unified framework that consolidates these disparate capabilities remains absent, limiting the potential for multi-control and complex editing in image synthesis." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 130, + 174, + 480, + 357 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 174, + 480, + 357 + ], + "spans": [ + { + "bbox": [ + 130, + 174, + 480, + 357 + ], + "type": "text", + "content": "Model Efficient Training and Scaling Up Strategies. Efficient training and scaling of models are pivotal for advancing large-scale neural networks. Previous models like DeepFloyd [31] and GLIDE [20] capitalize on cascaded approaches that proficiently utilize data across various resolutions, which results in less efficient inference processes. Contrastingly, models like Stable Diffusion [26] and AltDiffusion [37] adopt Resolution Boost Training strategies that refine a single model for cost-effectiveness, which however do not fully exploit data efficiency. In scaling up strategies, training efficiency is also important. Efficient adaptation and scaling are explored in [3] through distillation, and in [23] by marrying model expansion with domain-specific prompts. Serial scaling and knowledge distillation reduce training times significantly as demonstrated by [8], while [6] proposes progressive network expansion for faster training with minimal loss. We propose a novel Time-Decoupling training strategy to diffusion model scaling that enhances efficiency. eDiff-I [1] similarly proposes to split a diffusion model across time steps to boost generation quality without increasing inference cost. However, they do not consider about data and training efficiencies." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 132, + 376, + 225, + 390 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 376, + 225, + 390 + ], + "spans": [ + { + "bbox": [ + 132, + 376, + 225, + 390 + ], + "type": "text", + "content": "3 Preliminary" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 130, + 401, + 479, + 436 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 401, + 479, + 436 + ], + "spans": [ + { + "bbox": [ + 130, + 401, + 479, + 436 + ], + "type": "text", + "content": "Given an image " + }, + { + "bbox": [ + 130, + 401, + 479, + 436 + ], + "type": "inline_equation", + "content": "x_0" + }, + { + "bbox": [ + 130, + 401, + 479, + 436 + ], + "type": "text", + "content": ", diffusion models first produce a series of noisy images " + }, + { + "bbox": [ + 130, + 401, + 479, + 436 + ], + "type": "inline_equation", + "content": "x_1, \\ldots, x_T" + }, + { + "bbox": [ + 130, + 401, + 479, + 436 + ], + "type": "text", + "content": " by adding Gaussian noise to " + }, + { + "bbox": [ + 130, + 401, + 479, + 436 + ], + "type": "inline_equation", + "content": "x_0" + }, + { + "bbox": [ + 130, + 401, + 479, + 436 + ], + "type": "text", + "content": " according to some noise schedule given by " + }, + { + "bbox": [ + 130, + 401, + 479, + 436 + ], + "type": "inline_equation", + "content": "\\bar{\\alpha}_t" + }, + { + "bbox": [ + 130, + 401, + 479, + 436 + ], + "type": "text", + "content": " as follows:" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 225, + 445, + 480, + 458 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 225, + 445, + 480, + 458 + ], + "spans": [ + { + "bbox": [ + 225, + 445, + 480, + 458 + ], + "type": "interline_equation", + "content": "x _ {t} = \\sqrt {\\bar {\\alpha} _ {t}} x _ {0} + \\sqrt {1 - \\bar {\\alpha} _ {t}} \\epsilon , \\epsilon \\sim \\mathcal {N} (0, I). \\tag {1}", + "image_path": "9b6d224a7c6f6349fdde29290acba39b6ad830e5f16c0c327e0acb137f880771.jpg" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 131, + 466, + 479, + 490 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 131, + 466, + 479, + 490 + ], + "spans": [ + { + "bbox": [ + 131, + 466, + 479, + 490 + ], + "type": "text", + "content": "Diffusion models then learn a denoising model " + }, + { + "bbox": [ + 131, + 466, + 479, + 490 + ], + "type": "inline_equation", + "content": "\\epsilon_{\\theta}(x_t,t)" + }, + { + "bbox": [ + 131, + 466, + 479, + 490 + ], + "type": "text", + "content": " to predict the added noise of a noisy image " + }, + { + "bbox": [ + 131, + 466, + 479, + 490 + ], + "type": "inline_equation", + "content": "x_{t}" + }, + { + "bbox": [ + 131, + 466, + 479, + 490 + ], + "type": "text", + "content": " with the following training objective:" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 208, + 498, + 480, + 512 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 208, + 498, + 480, + 512 + ], + "spans": [ + { + "bbox": [ + 208, + 498, + 480, + 512 + ], + "type": "interline_equation", + "content": "\\mathcal {L} = \\mathbb {E} _ {x _ {0} \\sim q (x _ {0}), \\epsilon \\sim \\mathcal {N} (0, I), t \\sim [ 1, T ]} \\| \\epsilon - \\epsilon_ {\\theta} (x _ {t}, t) \\| ^ {2}, \\tag {2}", + "image_path": "e550a76d0d0d8895916504ce563168b85b44e222069932ebf15d6c77bf445124.jpg" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 130, + 520, + 479, + 567 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 520, + 479, + 567 + ], + "spans": [ + { + "bbox": [ + 130, + 520, + 479, + 567 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 130, + 520, + 479, + 567 + ], + "type": "inline_equation", + "content": "t" + }, + { + "bbox": [ + 130, + 520, + 479, + 567 + ], + "type": "text", + "content": " is uniformly sampled from " + }, + { + "bbox": [ + 130, + 520, + 479, + 567 + ], + "type": "inline_equation", + "content": "\\{1, \\dots, T\\}" + }, + { + "bbox": [ + 130, + 520, + 479, + 567 + ], + "type": "text", + "content": ". Once the denoising model " + }, + { + "bbox": [ + 130, + 520, + 479, + 567 + ], + "type": "inline_equation", + "content": "\\epsilon_{\\theta}(x_t, t)" + }, + { + "bbox": [ + 130, + 520, + 479, + 567 + ], + "type": "text", + "content": " is learned, starting from a random noise " + }, + { + "bbox": [ + 130, + 520, + 479, + 567 + ], + "type": "inline_equation", + "content": "x_T \\sim \\mathcal{N}(0, I)" + }, + { + "bbox": [ + 130, + 520, + 479, + 567 + ], + "type": "text", + "content": ", one can iteratively predict and reduce the noise in " + }, + { + "bbox": [ + 130, + 520, + 479, + 567 + ], + "type": "inline_equation", + "content": "x_t" + }, + { + "bbox": [ + 130, + 520, + 479, + 567 + ], + "type": "text", + "content": " to get a real image " + }, + { + "bbox": [ + 130, + 520, + 479, + 567 + ], + "type": "inline_equation", + "content": "x_0" + }, + { + "bbox": [ + 130, + 520, + 479, + 567 + ], + "type": "text", + "content": ". During the sampling process, we can predict the clean data " + }, + { + "bbox": [ + 130, + 520, + 479, + 567 + ], + "type": "inline_equation", + "content": "x_0" + }, + { + "bbox": [ + 130, + 520, + 479, + 567 + ], + "type": "text", + "content": " from " + }, + { + "bbox": [ + 130, + 520, + 479, + 567 + ], + "type": "inline_equation", + "content": "\\epsilon_{\\theta}(x_t, t)" + }, + { + "bbox": [ + 130, + 520, + 479, + 567 + ], + "type": "text", + "content": " with single-step sampling as:" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 231, + 575, + 480, + 600 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 231, + 575, + 480, + 600 + ], + "spans": [ + { + "bbox": [ + 231, + 575, + 480, + 600 + ], + "type": "interline_equation", + "content": "\\hat {x} _ {0, t} = \\frac {1}{\\sqrt {\\bar {\\alpha} _ {t}}} \\left(x _ {t} - \\sqrt {1 - \\bar {\\alpha} _ {t}} \\epsilon_ {\\theta} \\left(x _ {t}, t\\right)\\right). \\tag {3}", + "image_path": "5c5764d299a76089fc9155f7e86bfc69d689adf8bb4c41dfab1849be5f4cf56c.jpg" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 130, + 608, + 480, + 666 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 608, + 480, + 666 + ], + "spans": [ + { + "bbox": [ + 130, + 608, + 480, + 666 + ], + "type": "text", + "content": "Our text-to-image generation model is built on the model architecture proposed in Latent Diffusion Model [26]. In this model, a real image " + }, + { + "bbox": [ + 130, + 608, + 480, + 666 + ], + "type": "inline_equation", + "content": "x_0" + }, + { + "bbox": [ + 130, + 608, + 480, + 666 + ], + "type": "text", + "content": " is first down-sampled 8 times as a lower-dimension latent code " + }, + { + "bbox": [ + 130, + 608, + 480, + 666 + ], + "type": "inline_equation", + "content": "z_0" + }, + { + "bbox": [ + 130, + 608, + 480, + 666 + ], + "type": "text", + "content": " with an image encoder model " + }, + { + "bbox": [ + 130, + 608, + 480, + 666 + ], + "type": "inline_equation", + "content": "E" + }, + { + "bbox": [ + 130, + 608, + 480, + 666 + ], + "type": "text", + "content": ", which can be decoded with a latent decoder model " + }, + { + "bbox": [ + 130, + 608, + 480, + 666 + ], + "type": "inline_equation", + "content": "D" + }, + { + "bbox": [ + 130, + 608, + 480, + 666 + ], + "type": "text", + "content": " back to a real image " + }, + { + "bbox": [ + 130, + 608, + 480, + 666 + ], + "type": "inline_equation", + "content": "x_0" + }, + { + "bbox": [ + 130, + 608, + 480, + 666 + ], + "type": "text", + "content": ". The denoising network " + }, + { + "bbox": [ + 130, + 608, + 480, + 666 + ], + "type": "inline_equation", + "content": "\\epsilon_{\\theta}(z_t,t,c)" + }, + { + "bbox": [ + 130, + 608, + 480, + 666 + ], + "type": "text", + "content": " is parameterized as a U-Net [28]" + } + ] + } + ], + "index": 11 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "spans": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "type": "text", + "content": "4" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 166, + 91, + 215, + 101 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 166, + 91, + 215, + 101 + ], + "spans": [ + { + "bbox": [ + 166, + 91, + 215, + 101 + ], + "type": "text", + "content": "G. Lu et al." + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 3 + }, + { + "para_blocks": [ + { + "bbox": [ + 130, + 116, + 479, + 140 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 116, + 479, + 140 + ], + "spans": [ + { + "bbox": [ + 130, + 116, + 479, + 140 + ], + "type": "text", + "content": "model, where embedding of time step " + }, + { + "bbox": [ + 130, + 116, + 479, + 140 + ], + "type": "inline_equation", + "content": "t" + }, + { + "bbox": [ + 130, + 116, + 479, + 140 + ], + "type": "text", + "content": " is injected with adaptive normalization layers and embedding of input text " + }, + { + "bbox": [ + 130, + 116, + 479, + 140 + ], + "type": "inline_equation", + "content": "c" + }, + { + "bbox": [ + 130, + 116, + 479, + 140 + ], + "type": "text", + "content": " is injected with cross-attention layers." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 131, + 159, + 233, + 172 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 131, + 159, + 233, + 172 + ], + "spans": [ + { + "bbox": [ + 131, + 159, + 233, + 172 + ], + "type": "text", + "content": "4 PanGu-Draw" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 130, + 186, + 482, + 245 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 186, + 482, + 245 + ], + "spans": [ + { + "bbox": [ + 130, + 186, + 482, + 245 + ], + "type": "text", + "content": "In this section, we first illustrate our resource-efficient 5B text-to-image generation model, trained with a time-decoupling training strategy and further enhanced with a prompt enhancement LLM. Then, we present our Coop-Diffusion algorithm for the cooperative integration of diverse pre-trained diffusion models, enabling multi-control and multi-resolution image generation." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 131, + 263, + 340, + 277 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 131, + 263, + 340, + 277 + ], + "spans": [ + { + "bbox": [ + 131, + 263, + 340, + 277 + ], + "type": "text", + "content": "4.1 Time-Decoupling Training Strategy" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 130, + 285, + 482, + 376 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 285, + 482, + 376 + ], + "spans": [ + { + "bbox": [ + 130, + 285, + 482, + 376 + ], + "type": "text", + "content": "Enhancing data, training, and inference efficiency is vital for text-to-image models' practical use. Figure 1 shows two existing training strategies: (a) Cascaded Training, using three models to incrementally improve resolution, is data-efficient but triples training and inference time. (b) Resolution Boost Training starts at " + }, + { + "bbox": [ + 130, + 285, + 482, + 376 + ], + "type": "inline_equation", + "content": "512 \\times 512" + }, + { + "bbox": [ + 130, + 285, + 482, + 376 + ], + "type": "text", + "content": " and then " + }, + { + "bbox": [ + 130, + 285, + 482, + 376 + ], + "type": "inline_equation", + "content": "1024 \\times 1024" + }, + { + "bbox": [ + 130, + 285, + 482, + 376 + ], + "type": "text", + "content": " resolution, discarding lower resolution data and offering moderate efficiency with higher training costs and single-model inference across all timesteps. These approaches differ from our time-decoupling strategy, detailed below." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 130, + 377, + 482, + 480 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 377, + 482, + 480 + ], + "spans": [ + { + "bbox": [ + 130, + 377, + 482, + 480 + ], + "type": "text", + "content": "Responding to the need for enhanced efficiencies, we draw inspiration from the denoising trajectory of diffusion processes, where initial denoising stages primarily shape the image's structural foundation, and later stages refine its textural complexity. With this insight, we introduce the Time-Decoupling Training Strategy. This approach divides a comprehensive text-to-image model, denoted as " + }, + { + "bbox": [ + 130, + 377, + 482, + 480 + ], + "type": "inline_equation", + "content": "\\epsilon_{\\theta}" + }, + { + "bbox": [ + 130, + 377, + 482, + 480 + ], + "type": "text", + "content": ", into two specialized sub-models operating across different temporal intervals: a structure generator, " + }, + { + "bbox": [ + 130, + 377, + 482, + 480 + ], + "type": "inline_equation", + "content": "\\epsilon_{struct}" + }, + { + "bbox": [ + 130, + 377, + 482, + 480 + ], + "type": "text", + "content": ", and a texture generator, " + }, + { + "bbox": [ + 130, + 377, + 482, + 480 + ], + "type": "inline_equation", + "content": "\\epsilon_{texture}" + }, + { + "bbox": [ + 130, + 377, + 482, + 480 + ], + "type": "text", + "content": ". Each sub-model is half the size of the original, thus enhancing manageability and reducing computational load." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 130, + 481, + 482, + 584 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 481, + 482, + 584 + ], + "spans": [ + { + "bbox": [ + 130, + 481, + 482, + 584 + ], + "type": "text", + "content": "As illustrated in Figure 1(c), the structure generator, " + }, + { + "bbox": [ + 130, + 481, + 482, + 584 + ], + "type": "inline_equation", + "content": "\\epsilon_{\\text{struct}}" + }, + { + "bbox": [ + 130, + 481, + 482, + 584 + ], + "type": "text", + "content": ", is responsible for early-stage denoising across larger time steps, specifically within the range " + }, + { + "bbox": [ + 130, + 481, + 482, + 584 + ], + "type": "inline_equation", + "content": "T, \\dots, T_{\\text{struct}}" + }, + { + "bbox": [ + 130, + 481, + 482, + 584 + ], + "type": "text", + "content": ", where " + }, + { + "bbox": [ + 130, + 481, + 482, + 584 + ], + "type": "inline_equation", + "content": "0 < T_{\\text{struct}} < T" + }, + { + "bbox": [ + 130, + 481, + 482, + 584 + ], + "type": "text", + "content": ". This stage focuses on establishing the foundational outlines of the image. Conversely, the texture generator, " + }, + { + "bbox": [ + 130, + 481, + 482, + 584 + ], + "type": "inline_equation", + "content": "\\epsilon_{\\text{texture}}" + }, + { + "bbox": [ + 130, + 481, + 482, + 584 + ], + "type": "text", + "content": ", operates during the latter, smaller time steps, denoted by " + }, + { + "bbox": [ + 130, + 481, + 482, + 584 + ], + "type": "inline_equation", + "content": "T_{\\text{struct}}, \\dots, 0" + }, + { + "bbox": [ + 130, + 481, + 482, + 584 + ], + "type": "text", + "content": ", to elaborate on the textural details. Each generator is trained in isolation, which not only alleviates the need for high-memory computation devices but also avoids the complexities associated with model sharding and its accompanying inter-machine communication overhead." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 130, + 585, + 482, + 665 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 585, + 482, + 665 + ], + "spans": [ + { + "bbox": [ + 130, + 585, + 482, + 665 + ], + "type": "text", + "content": "In the inference phase, " + }, + { + "bbox": [ + 130, + 585, + 482, + 665 + ], + "type": "inline_equation", + "content": "\\epsilon_{\\text{struct}}" + }, + { + "bbox": [ + 130, + 585, + 482, + 665 + ], + "type": "text", + "content": " initially constructs a base structural image, " + }, + { + "bbox": [ + 130, + 585, + 482, + 665 + ], + "type": "inline_equation", + "content": "z_{T_{\\text{struct}}}" + }, + { + "bbox": [ + 130, + 585, + 482, + 665 + ], + "type": "text", + "content": ", from an initial random noise vector, " + }, + { + "bbox": [ + 130, + 585, + 482, + 665 + ], + "type": "inline_equation", + "content": "z_T" + }, + { + "bbox": [ + 130, + 585, + 482, + 665 + ], + "type": "text", + "content": ". Subsequently, " + }, + { + "bbox": [ + 130, + 585, + 482, + 665 + ], + "type": "inline_equation", + "content": "\\epsilon_{\\text{texture}}" + }, + { + "bbox": [ + 130, + 585, + 482, + 665 + ], + "type": "text", + "content": " refines this base to enhance textural details, culminating in the final output, " + }, + { + "bbox": [ + 130, + 585, + 482, + 665 + ], + "type": "inline_equation", + "content": "z_0" + }, + { + "bbox": [ + 130, + 585, + 482, + 665 + ], + "type": "text", + "content": ". This sequential processing facilitates a more resource-efficient workflow, significantly reducing the hardware footprint and expediting the generation process without compromising the model's performance or output quality, as demonstrated in our ablated experiment in Sec. 5.3." + } + ] + } + ], + "index": 9 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 392, + 91, + 447, + 101 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 392, + 91, + 447, + 101 + ], + "spans": [ + { + "bbox": [ + 392, + 91, + 447, + 101 + ], + "type": "text", + "content": "PanGu-Draw" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 474, + 91, + 481, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 474, + 91, + 481, + 100 + ], + "spans": [ + { + "bbox": [ + 474, + 91, + 481, + 100 + ], + "type": "text", + "content": "5" + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 4 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 153, + 114, + 455, + 268 + ], + "blocks": [ + { + "bbox": [ + 153, + 114, + 455, + 268 + ], + "lines": [ + { + "bbox": [ + 153, + 114, + 455, + 268 + ], + "spans": [ + { + "bbox": [ + 153, + 114, + 455, + 268 + ], + "type": "image", + "image_path": "b26313a4b7eaefc2071c2ca00ab616296f64a56bceb23db74fe5e2378bf971e2.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 130, + 277, + 482, + 354 + ], + "lines": [ + { + "bbox": [ + 130, + 277, + 482, + 354 + ], + "spans": [ + { + "bbox": [ + 130, + 277, + 482, + 354 + ], + "type": "text", + "content": "Fig. 2: Visualization of our Coop-Diffusion algorithm for the cooperative integration of diverse pre-trained diffusion models. (a) Existing pre-trained diffusion models, each tailored for specific controls and operating within distinct latent spaces and image resolutions. (b) This sub-module bridges the gap arising from different latent spaces by transforming " + }, + { + "bbox": [ + 130, + 277, + 482, + 354 + ], + "type": "inline_equation", + "content": "\\epsilon_t'" + }, + { + "bbox": [ + 130, + 277, + 482, + 354 + ], + "type": "text", + "content": " in latent space B to the target latent space A as " + }, + { + "bbox": [ + 130, + 277, + 482, + 354 + ], + "type": "inline_equation", + "content": "\\tilde{\\epsilon}_t" + }, + { + "bbox": [ + 130, + 277, + 482, + 354 + ], + "type": "text", + "content": ". (c) This submodule bridges the gap arising from different resolutions by performing upsampling on the predicted clean data " + }, + { + "bbox": [ + 130, + 277, + 482, + 354 + ], + "type": "inline_equation", + "content": "\\hat{x}_{0,t}'" + }, + { + "bbox": [ + 130, + 277, + 482, + 354 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_caption" + } + ], + "index": 2 + }, + { + "bbox": [ + 130, + 376, + 482, + 491 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 376, + 482, + 491 + ], + "spans": [ + { + "bbox": [ + 130, + 376, + 482, + 491 + ], + "type": "text", + "content": "Resource-Efficient Specialized Training Regime. We further adopt specialized training designs for the above two models. The structure generator " + }, + { + "bbox": [ + 130, + 376, + 482, + 491 + ], + "type": "inline_equation", + "content": "\\epsilon_{struct}" + }, + { + "bbox": [ + 130, + 376, + 482, + 491 + ], + "type": "text", + "content": ", which derives image structures from text, requires training on an extensive dataset encompassing a wide range of concepts. Traditional methods, like Stable Diffusion, often eliminate low-resolution images, discarding about " + }, + { + "bbox": [ + 130, + 376, + 482, + 491 + ], + "type": "inline_equation", + "content": "48\\%" + }, + { + "bbox": [ + 130, + 376, + 482, + 491 + ], + "type": "text", + "content": " of training data and thereby inflating dataset costs. Contrarily, we integrate high-resolution images with upscaled lower-resolution ones. This approach, as proven by our ablated experiments in Sec. 5.3, shows no performance drop, as the predicted " + }, + { + "bbox": [ + 130, + 376, + 482, + 491 + ], + "type": "inline_equation", + "content": "z_{T_{struct}}" + }, + { + "bbox": [ + 130, + 376, + 482, + 491 + ], + "type": "text", + "content": " still contains substantial noise. In this way, we achieve higher data efficiency and avoid the problem of semantic degeneration." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 130, + 492, + 482, + 608 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 492, + 482, + 608 + ], + "spans": [ + { + "bbox": [ + 130, + 492, + 482, + 608 + ], + "type": "text", + "content": "Since the image structure is determined in " + }, + { + "bbox": [ + 130, + 492, + 482, + 608 + ], + "type": "inline_equation", + "content": "z_{T_{\\text{struct}}}" + }, + { + "bbox": [ + 130, + 492, + 482, + 608 + ], + "type": "text", + "content": " and the texture generator " + }, + { + "bbox": [ + 130, + 492, + 482, + 608 + ], + "type": "inline_equation", + "content": "\\epsilon_{\\text{texture}}" + }, + { + "bbox": [ + 130, + 492, + 482, + 608 + ], + "type": "text", + "content": " focuses on refining texture, we propose training " + }, + { + "bbox": [ + 130, + 492, + 482, + 608 + ], + "type": "inline_equation", + "content": "\\epsilon_{\\text{texture}}" + }, + { + "bbox": [ + 130, + 492, + 482, + 608 + ], + "type": "text", + "content": " at a lower resolution while still sampling at high resolution. This strategy, as demonstrated in our ablated experiments in Sec. 5.3, results in no performance drop and no structural problems (e.g., repetitive presentation [15]). Consequently, we achieved an overall " + }, + { + "bbox": [ + 130, + 492, + 482, + 608 + ], + "type": "inline_equation", + "content": "51\\%" + }, + { + "bbox": [ + 130, + 492, + 482, + 608 + ], + "type": "text", + "content": " improvement in training efficiency. Figure 1 summarizes the data, training, and inference efficiency of different training strategies. Besides higher data and training efficiency, our strategy also achieves higher inference efficiency with fewer inference steps compared to the Cascaded Training strategy and a smaller per-step model compared to the Resolution Boost Training strategy." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 132, + 623, + 356, + 635 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 623, + 356, + 635 + ], + "spans": [ + { + "bbox": [ + 132, + 623, + 356, + 635 + ], + "type": "text", + "content": "4.2 Coop-Diffusion: Multi-Diffusion Fusion" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 130, + 642, + 481, + 666 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 642, + 481, + 666 + ], + "spans": [ + { + "bbox": [ + 130, + 642, + 481, + 666 + ], + "type": "text", + "content": "As shown in Figure 2(a), there are numerous pre-trained diffusion models, such as various SD, ControlNet, image variation, etc., each tailored for specific con" + } + ] + } + ], + "index": 7 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "spans": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "type": "text", + "content": "6" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 166, + 91, + 215, + 100 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 166, + 91, + 215, + 100 + ], + "spans": [ + { + "bbox": [ + 166, + 91, + 215, + 100 + ], + "type": "text", + "content": "G. Lu et al." + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 5 + }, + { + "para_blocks": [ + { + "bbox": [ + 130, + 116, + 479, + 209 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 116, + 479, + 209 + ], + "spans": [ + { + "bbox": [ + 130, + 116, + 479, + 209 + ], + "type": "text", + "content": "trols and image resolutions. It is promising to fuse these pre-trained models for multi-control or multi-resolution image generation without needing to train a new model. However, the different latent spaces and resolutions of these models impede joint synthesis of images controlled by different models, thereby limiting their practical applications. In response to these challenges, we propose the Coop-Diffusion algorithm with two key sub-modules, as shown in Figures 2(b) and (c), to bridge the latent space gap and the resolution gap, and to unite the denoising process in the same space." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 130, + 209, + 480, + 264 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 209, + 480, + 264 + ], + "spans": [ + { + "bbox": [ + 130, + 209, + 480, + 264 + ], + "type": "text", + "content": "Bridging the Latent Space Gap. To bridge the latent space gap between two different latent spaces A and B, we propose to unify the model prediction in latent space A by transforming the model prediction " + }, + { + "bbox": [ + 130, + 209, + 480, + 264 + ], + "type": "inline_equation", + "content": "\\epsilon_t^\\prime" + }, + { + "bbox": [ + 130, + 209, + 480, + 264 + ], + "type": "text", + "content": " in latent space B to latent space A using the image space as an intermediate. This is done in the following way: first, we predict the clean data " + }, + { + "bbox": [ + 130, + 209, + 480, + 264 + ], + "type": "inline_equation", + "content": "\\hat{z}_{0,t}^{\\prime}" + }, + { + "bbox": [ + 130, + 209, + 480, + 264 + ], + "type": "text", + "content": " using Equation (3) as:" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 247, + 271, + 481, + 295 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 247, + 271, + 481, + 295 + ], + "spans": [ + { + "bbox": [ + 247, + 271, + 481, + 295 + ], + "type": "interline_equation", + "content": "\\dot {z} _ {0, t} ^ {\\prime} = \\frac {1}{\\sqrt {\\bar {\\alpha} _ {t}}} (z _ {t} ^ {\\prime} - \\sqrt {1 - \\bar {\\alpha} _ {t}} \\epsilon_ {t} ^ {\\prime}), \\tag {4}", + "image_path": "05f4d5eece0ef4b20ac75de718b81dd255dcb07f018264d0f1ac6ecf720e8348.jpg" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 131, + 300, + 479, + 348 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 131, + 300, + 479, + 348 + ], + "spans": [ + { + "bbox": [ + 131, + 300, + 479, + 348 + ], + "type": "text", + "content": "which is then decoded into a pixel-level image " + }, + { + "bbox": [ + 131, + 300, + 479, + 348 + ], + "type": "inline_equation", + "content": "\\hat{x}_{0,t}^{\\prime}" + }, + { + "bbox": [ + 131, + 300, + 479, + 348 + ], + "type": "text", + "content": " using the latent decoder model " + }, + { + "bbox": [ + 131, + 300, + 479, + 348 + ], + "type": "inline_equation", + "content": "D^{\\prime}" + }, + { + "bbox": [ + 131, + 300, + 479, + 348 + ], + "type": "text", + "content": ". This image is encoded into latent space A using the image encoder model " + }, + { + "bbox": [ + 131, + 300, + 479, + 348 + ], + "type": "inline_equation", + "content": "E" + }, + { + "bbox": [ + 131, + 300, + 479, + 348 + ], + "type": "text", + "content": ", as " + }, + { + "bbox": [ + 131, + 300, + 479, + 348 + ], + "type": "inline_equation", + "content": "\\tilde{z}_{0,t} = E(\\hat{x}_{0,t}^{\\prime})" + }, + { + "bbox": [ + 131, + 300, + 479, + 348 + ], + "type": "text", + "content": ", and finally transformed into a model prediction by inverting Equation (3) as:" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 247, + 354, + 481, + 378 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 247, + 354, + 481, + 378 + ], + "spans": [ + { + "bbox": [ + 247, + 354, + 481, + 378 + ], + "type": "interline_equation", + "content": "\\tilde {\\epsilon} _ {t} = \\frac {1}{\\sqrt {1 - \\bar {\\alpha} _ {t}}} \\left(z _ {t} - \\sqrt {\\bar {\\alpha} _ {t}} \\tilde {z} _ {0, t}\\right). \\tag {5}", + "image_path": "d8f445c17e333f86a436d02b7c193429cd533577b59d8b425546766dec0fece4.jpg" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 131, + 383, + 479, + 464 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 131, + 383, + 479, + 464 + ], + "spans": [ + { + "bbox": [ + 131, + 383, + 479, + 464 + ], + "type": "text", + "content": "With the united " + }, + { + "bbox": [ + 131, + 383, + 479, + 464 + ], + "type": "inline_equation", + "content": "\\tilde{\\epsilon}_t" + }, + { + "bbox": [ + 131, + 383, + 479, + 464 + ], + "type": "text", + "content": ", we can now perform multi-control fusion between " + }, + { + "bbox": [ + 131, + 383, + 479, + 464 + ], + "type": "inline_equation", + "content": "\\tilde{\\epsilon}_t" + }, + { + "bbox": [ + 131, + 383, + 479, + 464 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 131, + 383, + 479, + 464 + ], + "type": "inline_equation", + "content": "\\epsilon_t" + }, + { + "bbox": [ + 131, + 383, + 479, + 464 + ], + "type": "text", + "content": " (the prediction from model " + }, + { + "bbox": [ + 131, + 383, + 479, + 464 + ], + "type": "inline_equation", + "content": "\\epsilon_{\\theta}" + }, + { + "bbox": [ + 131, + 383, + 479, + 464 + ], + "type": "text", + "content": " with " + }, + { + "bbox": [ + 131, + 383, + 479, + 464 + ], + "type": "inline_equation", + "content": "z_{t}" + }, + { + "bbox": [ + 131, + 383, + 479, + 464 + ], + "type": "text", + "content": " in latent space A, omitted in Figure 2 for brevity). In this paper, we adopt the fusion method proposed in Composable-Diffusion [19] as: " + }, + { + "bbox": [ + 131, + 383, + 479, + 464 + ], + "type": "inline_equation", + "content": "\\epsilon_{t,fuse} = d\\cdot \\tilde{\\epsilon}_t + (1 - d)\\cdot \\epsilon_t" + }, + { + "bbox": [ + 131, + 383, + 479, + 464 + ], + "type": "text", + "content": ", where " + }, + { + "bbox": [ + 131, + 383, + 479, + 464 + ], + "type": "inline_equation", + "content": "d" + }, + { + "bbox": [ + 131, + 383, + 479, + 464 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 131, + 383, + 479, + 464 + ], + "type": "inline_equation", + "content": "1 - d" + }, + { + "bbox": [ + 131, + 383, + 479, + 464 + ], + "type": "text", + "content": " are the guidance strengths of each model with " + }, + { + "bbox": [ + 131, + 383, + 479, + 464 + ], + "type": "inline_equation", + "content": "d\\in [0,1]" + }, + { + "bbox": [ + 131, + 383, + 479, + 464 + ], + "type": "text", + "content": ", to guide the denoising process jointly with these two models for multi-control image generation. Algorithm 1 further illustrates this unification and fusion process." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 131, + 465, + 479, + 489 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 131, + 465, + 479, + 489 + ], + "spans": [ + { + "bbox": [ + 131, + 465, + 479, + 489 + ], + "type": "text", + "content": "Bridging Resolution Gap. To integrate the denoising processes of a low-resolution model with a high-resolution model, upsampling and/or downsam" + } + ] + } + ], + "index": 8 + }, + { + "type": "image", + "bbox": [ + 202, + 514, + 302, + 611 + ], + "blocks": [ + { + "bbox": [ + 202, + 514, + 302, + 611 + ], + "lines": [ + { + "bbox": [ + 202, + 514, + 302, + 611 + ], + "spans": [ + { + "bbox": [ + 202, + 514, + 302, + 611 + ], + "type": "image", + "image_path": "41797f8253c6970c94b7ba047045df5204348de442f8bf20547e0ab89ec42daf.jpg" + } + ] + } + ], + "index": 9, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 205, + 612, + 299, + 620 + ], + "lines": [ + { + "bbox": [ + 205, + 612, + 299, + 620 + ], + "spans": [ + { + "bbox": [ + 205, + 612, + 299, + 620 + ], + "type": "text", + "content": "(a) Upsampling from intermediate " + }, + { + "bbox": [ + 205, + 612, + 299, + 620 + ], + "type": "inline_equation", + "content": "z_{t}" + }, + { + "bbox": [ + 205, + 612, + 299, + 620 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 10, + "angle": 0, + "type": "image_caption" + } + ], + "index": 9 + }, + { + "type": "image", + "bbox": [ + 311, + 514, + 411, + 612 + ], + "blocks": [ + { + "bbox": [ + 311, + 514, + 411, + 612 + ], + "lines": [ + { + "bbox": [ + 311, + 514, + 411, + 612 + ], + "spans": [ + { + "bbox": [ + 311, + 514, + 411, + 612 + ], + "type": "image", + "image_path": "f97620877ba2582d5e0440664a4d1cb1668cfa21e2e9ec676fb290d9c023b68d.jpg" + } + ] + } + ], + "index": 11, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 323, + 612, + 400, + 620 + ], + "lines": [ + { + "bbox": [ + 323, + 612, + 400, + 620 + ], + "spans": [ + { + "bbox": [ + 323, + 612, + 400, + 620 + ], + "type": "text", + "content": "(b) Our upsampling algorithm." + } + ] + } + ], + "index": 12, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 131, + 628, + 479, + 662 + ], + "lines": [ + { + "bbox": [ + 131, + 628, + 479, + 662 + ], + "spans": [ + { + "bbox": [ + 131, + 628, + 479, + 662 + ], + "type": "text", + "content": "Fig. 3: Results of fusing a low-resolution model and a high-resolution model with different upsampling methods. Upsampling from intermediate " + }, + { + "bbox": [ + 131, + 628, + 479, + 662 + ], + "type": "inline_equation", + "content": "z_{t}" + }, + { + "bbox": [ + 131, + 628, + 479, + 662 + ], + "type": "text", + "content": " results in severe artifacts, while our upsampling algorithm results in high-fidelity image." + } + ] + } + ], + "index": 13, + "angle": 0, + "type": "image_caption" + } + ], + "index": 11 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 392, + 91, + 447, + 100 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 392, + 91, + 447, + 100 + ], + "spans": [ + { + "bbox": [ + 392, + 91, + 447, + 100 + ], + "type": "text", + "content": "PanGu-Draw" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 474, + 91, + 480, + 99 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 474, + 91, + 480, + 99 + ], + "spans": [ + { + "bbox": [ + 474, + 91, + 480, + 99 + ], + "type": "text", + "content": "7" + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 6 + }, + { + "para_blocks": [ + { + "type": "code", + "bbox": [ + 132, + 129, + 481, + 430 + ], + "blocks": [ + { + "bbox": [ + 133, + 116, + 380, + 128 + ], + "lines": [ + { + "bbox": [ + 133, + 116, + 380, + 128 + ], + "spans": [ + { + "bbox": [ + 133, + 116, + 380, + 128 + ], + "type": "text", + "content": "Algorithm 1 Coop-Diffusion: Multi-Diffusion Fusing" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "code_caption" + }, + { + "bbox": [ + 132, + 129, + 481, + 430 + ], + "lines": [ + { + "bbox": [ + 132, + 129, + 481, + 430 + ], + "spans": [ + { + "bbox": [ + 132, + 129, + 481, + 430 + ], + "type": "text", + "content": "Sub-Module 1. Bridging Latent Space Gap \nInput: random noise " + }, + { + "bbox": [ + 132, + 129, + 481, + 430 + ], + "type": "inline_equation", + "content": "z_{T} \\sim \\mathcal{N}(0, I)" + }, + { + "bbox": [ + 132, + 129, + 481, + 430 + ], + "type": "text", + "content": ", diffusion model " + }, + { + "bbox": [ + 132, + 129, + 481, + 430 + ], + "type": "inline_equation", + "content": "\\epsilon_{\\theta}" + }, + { + "bbox": [ + 132, + 129, + 481, + 430 + ], + "type": "text", + "content": ", decoder " + }, + { + "bbox": [ + 132, + 129, + 481, + 430 + ], + "type": "inline_equation", + "content": "D" + }, + { + "bbox": [ + 132, + 129, + 481, + 430 + ], + "type": "text", + "content": ", encoder " + }, + { + "bbox": [ + 132, + 129, + 481, + 430 + ], + "type": "inline_equation", + "content": "E" + }, + { + "bbox": [ + 132, + 129, + 481, + 430 + ], + "type": "text", + "content": " in latent space A; random noise " + }, + { + "bbox": [ + 132, + 129, + 481, + 430 + ], + "type": "inline_equation", + "content": "z_{T}' = z_{T}" + }, + { + "bbox": [ + 132, + 129, + 481, + 430 + ], + "type": "text", + "content": ", diffusion model " + }, + { + "bbox": [ + 132, + 129, + 481, + 430 + ], + "type": "inline_equation", + "content": "\\epsilon_{\\theta}'" + }, + { + "bbox": [ + 132, + 129, + 481, + 430 + ], + "type": "text", + "content": ", decoder " + }, + { + "bbox": [ + 132, + 129, + 481, + 430 + ], + "type": "inline_equation", + "content": "D'" + }, + { + "bbox": [ + 132, + 129, + 481, + 430 + ], + "type": "text", + "content": ", encoder " + }, + { + "bbox": [ + 132, + 129, + 481, + 430 + ], + "type": "inline_equation", + "content": "E'" + }, + { + "bbox": [ + 132, + 129, + 481, + 430 + ], + "type": "text", + "content": " in latent space B; guidance strength " + }, + { + "bbox": [ + 132, + 129, + 481, + 430 + ], + "type": "inline_equation", + "content": "d" + }, + { + "bbox": [ + 132, + 129, + 481, + 430 + ], + "type": "text", + "content": ", sampling method " + }, + { + "bbox": [ + 132, + 129, + 481, + 430 + ], + "type": "inline_equation", + "content": "S" + }, + { + "bbox": [ + 132, + 129, + 481, + 430 + ], + "type": "text", + "content": ". \n1: for " + }, + { + "bbox": [ + 132, + 129, + 481, + 430 + ], + "type": "inline_equation", + "content": "t = T, \\ldots, 1" + }, + { + "bbox": [ + 132, + 129, + 481, + 430 + ], + "type": "text", + "content": " do \n2: " + }, + { + "bbox": [ + 132, + 129, + 481, + 430 + ], + "type": "inline_equation", + "content": "\\epsilon_t' = \\epsilon_\\theta'(z_t')" + }, + { + "bbox": [ + 132, + 129, + 481, + 430 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 132, + 129, + 481, + 430 + ], + "type": "inline_equation", + "content": "\\hat{z}_{0,t}' = \\frac{1}{\\sqrt{\\alpha_t}} (z_t' - \\sqrt{1 - \\bar{\\alpha}_t} \\epsilon_t')" + }, + { + "bbox": [ + 132, + 129, + 481, + 430 + ], + "type": "text", + "content": " \n3: " + }, + { + "bbox": [ + 132, + 129, + 481, + 430 + ], + "type": "inline_equation", + "content": "\\hat{x}_{0,t}' = D'(\\hat{z}_{0,t}')" + }, + { + "bbox": [ + 132, + 129, + 481, + 430 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 132, + 129, + 481, + 430 + ], + "type": "inline_equation", + "content": "\\tilde{z}_{0,t} = E(\\hat{x}_{0,t}')" + }, + { + "bbox": [ + 132, + 129, + 481, + 430 + ], + "type": "text", + "content": " \n4: " + }, + { + "bbox": [ + 132, + 129, + 481, + 430 + ], + "type": "inline_equation", + "content": "\\tilde{\\epsilon}_t = \\frac{1}{\\sqrt{1 - \\bar{\\alpha}_t}} (z_t - \\sqrt{\\bar{\\alpha}_t} \\tilde{z}_{0,t})" + }, + { + "bbox": [ + 132, + 129, + 481, + 430 + ], + "type": "text", + "content": " \n5: " + }, + { + "bbox": [ + 132, + 129, + 481, + 430 + ], + "type": "inline_equation", + "content": "\\epsilon_t = \\epsilon_\\theta(z_t)" + }, + { + "bbox": [ + 132, + 129, + 481, + 430 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 132, + 129, + 481, + 430 + ], + "type": "inline_equation", + "content": "\\epsilon_{t,fuse} = d \\cdot \\tilde{\\epsilon}_t + (1 - d) \\cdot \\epsilon_t" + }, + { + "bbox": [ + 132, + 129, + 481, + 430 + ], + "type": "text", + "content": " \n6: " + }, + { + "bbox": [ + 132, + 129, + 481, + 430 + ], + "type": "inline_equation", + "content": "z_{t-1} = S(z_t, t, \\epsilon_{t,fuse})" + }, + { + "bbox": [ + 132, + 129, + 481, + 430 + ], + "type": "text", + "content": " \n7: " + }, + { + "bbox": [ + 132, + 129, + 481, + 430 + ], + "type": "inline_equation", + "content": "z_{t-1}' = S(z_t', t, \\epsilon_{t,fuse}') \\triangleright \\epsilon_{t,fuse}'" + }, + { + "bbox": [ + 132, + 129, + 481, + 430 + ], + "type": "text", + "content": " from " + }, + { + "bbox": [ + 132, + 129, + 481, + 430 + ], + "type": "inline_equation", + "content": "\\epsilon_{t,fuse}'" + }, + { + "bbox": [ + 132, + 129, + 481, + 430 + ], + "type": "text", + "content": " similar to the process from " + }, + { + "bbox": [ + 132, + 129, + 481, + 430 + ], + "type": "inline_equation", + "content": "\\epsilon_t'" + }, + { + "bbox": [ + 132, + 129, + 481, + 430 + ], + "type": "text", + "content": " to " + }, + { + "bbox": [ + 132, + 129, + 481, + 430 + ], + "type": "inline_equation", + "content": "\\tilde{\\epsilon}_t" + }, + { + "bbox": [ + 132, + 129, + 481, + 430 + ], + "type": "text", + "content": ", omitted for brevity \n8: end for \n9: return " + }, + { + "bbox": [ + 132, + 129, + 481, + 430 + ], + "type": "inline_equation", + "content": "D(z_0)" + }, + { + "bbox": [ + 132, + 129, + 481, + 430 + ], + "type": "text", + "content": " \nSub-Module 2. Bridging Resolution Gap \nInput: diffusion model " + }, + { + "bbox": [ + 132, + 129, + 481, + 430 + ], + "type": "inline_equation", + "content": "\\epsilon_{\\theta}" + }, + { + "bbox": [ + 132, + 129, + 481, + 430 + ], + "type": "text", + "content": ", decoder " + }, + { + "bbox": [ + 132, + 129, + 481, + 430 + ], + "type": "inline_equation", + "content": "D" + }, + { + "bbox": [ + 132, + 129, + 481, + 430 + ], + "type": "text", + "content": ", encoder " + }, + { + "bbox": [ + 132, + 129, + 481, + 430 + ], + "type": "inline_equation", + "content": "E" + }, + { + "bbox": [ + 132, + 129, + 481, + 430 + ], + "type": "text", + "content": " in high-resolution space; random noise " + }, + { + "bbox": [ + 132, + 129, + 481, + 430 + ], + "type": "inline_equation", + "content": "z_T' \\sim \\mathcal{N}(0, I)" + }, + { + "bbox": [ + 132, + 129, + 481, + 430 + ], + "type": "text", + "content": ", diffusion model " + }, + { + "bbox": [ + 132, + 129, + 481, + 430 + ], + "type": "inline_equation", + "content": "\\epsilon_{\\theta}'" + }, + { + "bbox": [ + 132, + 129, + 481, + 430 + ], + "type": "text", + "content": ", decoder " + }, + { + "bbox": [ + 132, + 129, + 481, + 430 + ], + "type": "inline_equation", + "content": "D'" + }, + { + "bbox": [ + 132, + 129, + 481, + 430 + ], + "type": "text", + "content": ", encoder " + }, + { + "bbox": [ + 132, + 129, + 481, + 430 + ], + "type": "inline_equation", + "content": "E'" + }, + { + "bbox": [ + 132, + 129, + 481, + 430 + ], + "type": "text", + "content": " in low-resolution space; low-resolution sampling end step " + }, + { + "bbox": [ + 132, + 129, + 481, + 430 + ], + "type": "inline_equation", + "content": "T_{low}" + }, + { + "bbox": [ + 132, + 129, + 481, + 430 + ], + "type": "text", + "content": ", sampling method " + }, + { + "bbox": [ + 132, + 129, + 481, + 430 + ], + "type": "inline_equation", + "content": "S" + }, + { + "bbox": [ + 132, + 129, + 481, + 430 + ], + "type": "text", + "content": ". \n1: for " + }, + { + "bbox": [ + 132, + 129, + 481, + 430 + ], + "type": "inline_equation", + "content": "t = T, \\ldots, T_{low} + 1" + }, + { + "bbox": [ + 132, + 129, + 481, + 430 + ], + "type": "text", + "content": " do \n2: " + }, + { + "bbox": [ + 132, + 129, + 481, + 430 + ], + "type": "inline_equation", + "content": "\\epsilon_t' = \\epsilon_\\theta'(z_t')" + }, + { + "bbox": [ + 132, + 129, + 481, + 430 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 132, + 129, + 481, + 430 + ], + "type": "inline_equation", + "content": "z_{t-1}' = S(z_t', t, \\epsilon_t')" + }, + { + "bbox": [ + 132, + 129, + 481, + 430 + ], + "type": "text", + "content": " \n3: end for \n4: " + }, + { + "bbox": [ + 132, + 129, + 481, + 430 + ], + "type": "inline_equation", + "content": "\\hat{z}_{0,T_{low}}' = \\frac{1}{\\sqrt{\\bar{\\alpha}_{T_{low}}}} (z_{T_{low}}' - \\sqrt{1 - \\bar{\\alpha}_{T_{low}}} \\epsilon_{T_{low}}')" + }, + { + "bbox": [ + 132, + 129, + 481, + 430 + ], + "type": "text", + "content": " \n5: " + }, + { + "bbox": [ + 132, + 129, + 481, + 430 + ], + "type": "inline_equation", + "content": "\\hat{x}_{0,T_{low}}' = D'(\\hat{z}_{0,T_{low}}')" + }, + { + "bbox": [ + 132, + 129, + 481, + 430 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 132, + 129, + 481, + 430 + ], + "type": "inline_equation", + "content": "\\hat{x}_{0,T_{low}} = Upsample(\\hat{x}_{0,T_{low}}')" + }, + { + "bbox": [ + 132, + 129, + 481, + 430 + ], + "type": "text", + "content": " \n6: " + }, + { + "bbox": [ + 132, + 129, + 481, + 430 + ], + "type": "inline_equation", + "content": "\\hat{z}_{0,T_{low}} = E(\\hat{x}_{0,T_{low}})" + }, + { + "bbox": [ + 132, + 129, + 481, + 430 + ], + "type": "text", + "content": " \n7: " + }, + { + "bbox": [ + 132, + 129, + 481, + 430 + ], + "type": "inline_equation", + "content": "z_{T_{low}} = \\sqrt{\\bar{\\alpha}_{T_{low}}} \\hat{z}_{0,T_{low}} + \\sqrt{1 - \\bar{\\alpha}_{T_{low}}} \\epsilon, \\epsilon \\sim \\mathcal{N}(0, I)" + }, + { + "bbox": [ + 132, + 129, + 481, + 430 + ], + "type": "text", + "content": " \n8: for " + }, + { + "bbox": [ + 132, + 129, + 481, + 430 + ], + "type": "inline_equation", + "content": "t = T_{low}, \\ldots, 1" + }, + { + "bbox": [ + 132, + 129, + 481, + 430 + ], + "type": "text", + "content": " do \n9: " + }, + { + "bbox": [ + 132, + 129, + 481, + 430 + ], + "type": "inline_equation", + "content": "\\epsilon_t = \\epsilon_\\theta(z_t)" + }, + { + "bbox": [ + 132, + 129, + 481, + 430 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 132, + 129, + 481, + 430 + ], + "type": "inline_equation", + "content": "z_{t-1} = S(z_t, t, \\epsilon_t)" + }, + { + "bbox": [ + 132, + 129, + 481, + 430 + ], + "type": "text", + "content": " \n10: end for \n11: return " + }, + { + "bbox": [ + 132, + 129, + 481, + 430 + ], + "type": "inline_equation", + "content": "D(z_0)" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "code_body" + } + ], + "index": 3, + "sub_type": "algorithm" + }, + { + "bbox": [ + 130, + 456, + 482, + 548 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 456, + 482, + 548 + ], + "spans": [ + { + "bbox": [ + 130, + 456, + 482, + 548 + ], + "type": "text", + "content": "pling is necessary. Traditional bilinear upsampling, often applied to the intermediate result " + }, + { + "bbox": [ + 130, + 456, + 482, + 548 + ], + "type": "inline_equation", + "content": "z_{t}" + }, + { + "bbox": [ + 130, + 456, + 482, + 548 + ], + "type": "text", + "content": " during the denoising process, can undesirably amplify pixel correlation. This amplification deviates from the initial Independent and Identically Distributed (IID) assumption, leading to severe artifacts in the final images, as shown in Figure 3(a). Conversely, downsampling does not present this issue. To address the IID issue in upsampling, we propose a new upsampling algorithm that preserves the IID assumption, thereby bridging the resolution gap between models with different pre-trained resolutions." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 130, + 549, + 493, + 641 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 549, + 493, + 641 + ], + "spans": [ + { + "bbox": [ + 130, + 549, + 493, + 641 + ], + "type": "text", + "content": "Figure 2(c) visualizes our upsampling algorithm. Specifically, for a low-resolution " + }, + { + "bbox": [ + 130, + 549, + 493, + 641 + ], + "type": "inline_equation", + "content": "z_{t}^{\\prime}" + }, + { + "bbox": [ + 130, + 549, + 493, + 641 + ], + "type": "text", + "content": ", we use the image space as an intermediate space to transform " + }, + { + "bbox": [ + 130, + 549, + 493, + 641 + ], + "type": "inline_equation", + "content": "z_{t}^{\\prime}" + }, + { + "bbox": [ + 130, + 549, + 493, + 641 + ], + "type": "text", + "content": " in low-resolution space into high-resolution space as " + }, + { + "bbox": [ + 130, + 549, + 493, + 641 + ], + "type": "inline_equation", + "content": "\\tilde{z}_{t}" + }, + { + "bbox": [ + 130, + 549, + 493, + 641 + ], + "type": "text", + "content": ". We first predict the noise " + }, + { + "bbox": [ + 130, + 549, + 493, + 641 + ], + "type": "inline_equation", + "content": "\\epsilon_{t}^{\\prime}" + }, + { + "bbox": [ + 130, + 549, + 493, + 641 + ], + "type": "text", + "content": " with the denoising model " + }, + { + "bbox": [ + 130, + 549, + 493, + 641 + ], + "type": "inline_equation", + "content": "\\epsilon_{\\theta}^{\\prime}" + }, + { + "bbox": [ + 130, + 549, + 493, + 641 + ], + "type": "text", + "content": " and then predict the clean data " + }, + { + "bbox": [ + 130, + 549, + 493, + 641 + ], + "type": "inline_equation", + "content": "\\hat{z}_{0,t}^{\\prime}" + }, + { + "bbox": [ + 130, + 549, + 493, + 641 + ], + "type": "text", + "content": " as described in Eq. 4. This is decoded into an image " + }, + { + "bbox": [ + 130, + 549, + 493, + 641 + ], + "type": "inline_equation", + "content": "\\hat{x}_{0,t}^{\\prime}" + }, + { + "bbox": [ + 130, + 549, + 493, + 641 + ], + "type": "text", + "content": " using decoder " + }, + { + "bbox": [ + 130, + 549, + 493, + 641 + ], + "type": "inline_equation", + "content": "D^{\\prime}" + }, + { + "bbox": [ + 130, + 549, + 493, + 641 + ], + "type": "text", + "content": ". We then perform upsampling on " + }, + { + "bbox": [ + 130, + 549, + 493, + 641 + ], + "type": "inline_equation", + "content": "\\hat{x}_{0,t}^{\\prime}" + }, + { + "bbox": [ + 130, + 549, + 493, + 641 + ], + "type": "text", + "content": " to obtain its high-resolution counterpart " + }, + { + "bbox": [ + 130, + 549, + 493, + 641 + ], + "type": "inline_equation", + "content": "\\hat{x}_{0,t}" + }, + { + "bbox": [ + 130, + 549, + 493, + 641 + ], + "type": "text", + "content": ". Finally, " + }, + { + "bbox": [ + 130, + 549, + 493, + 641 + ], + "type": "inline_equation", + "content": "\\hat{x}_{0,t}" + }, + { + "bbox": [ + 130, + 549, + 493, + 641 + ], + "type": "text", + "content": " is encoded into the latent space with encoder " + }, + { + "bbox": [ + 130, + 549, + 493, + 641 + ], + "type": "inline_equation", + "content": "E" + }, + { + "bbox": [ + 130, + 549, + 493, + 641 + ], + "type": "text", + "content": " as " + }, + { + "bbox": [ + 130, + 549, + 493, + 641 + ], + "type": "inline_equation", + "content": "\\hat{z}_{0,t}" + }, + { + "bbox": [ + 130, + 549, + 493, + 641 + ], + "type": "text", + "content": ", and " + }, + { + "bbox": [ + 130, + 549, + 493, + 641 + ], + "type": "inline_equation", + "content": "t" + }, + { + "bbox": [ + 130, + 549, + 493, + 641 + ], + "type": "text", + "content": "-step noise is added to get the final result " + }, + { + "bbox": [ + 130, + 549, + 493, + 641 + ], + "type": "inline_equation", + "content": "\\tilde{z}_{t}" + }, + { + "bbox": [ + 130, + 549, + 493, + 641 + ], + "type": "text", + "content": " using Eq. 1." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 130, + 642, + 481, + 666 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 642, + 481, + 666 + ], + "spans": [ + { + "bbox": [ + 130, + 642, + 481, + 666 + ], + "type": "text", + "content": "With the unified " + }, + { + "bbox": [ + 130, + 642, + 481, + 666 + ], + "type": "inline_equation", + "content": "\\tilde{z}_t" + }, + { + "bbox": [ + 130, + 642, + 481, + 666 + ], + "type": "text", + "content": ", we can now perform multi-resolution fusion. First, we denoise with a low-resolution model to obtain the intermediate " + }, + { + "bbox": [ + 130, + 642, + 481, + 666 + ], + "type": "inline_equation", + "content": "z_t'" + }, + { + "bbox": [ + 130, + 642, + 481, + 666 + ], + "type": "text", + "content": " and its high-" + } + ] + } + ], + "index": 6 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "spans": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "type": "text", + "content": "8" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 166, + 91, + 215, + 101 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 166, + 91, + 215, + 101 + ], + "spans": [ + { + "bbox": [ + 166, + 91, + 215, + 101 + ], + "type": "text", + "content": "G. Lu et al." + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 7 + }, + { + "para_blocks": [ + { + "bbox": [ + 130, + 116, + 482, + 175 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 116, + 482, + 175 + ], + "spans": [ + { + "bbox": [ + 130, + 116, + 482, + 175 + ], + "type": "text", + "content": "resolution counterpart " + }, + { + "bbox": [ + 130, + 116, + 482, + 175 + ], + "type": "inline_equation", + "content": "\\tilde{z}_t" + }, + { + "bbox": [ + 130, + 116, + 482, + 175 + ], + "type": "text", + "content": ". Then, we perform denoising with a high-resolution model starting from " + }, + { + "bbox": [ + 130, + 116, + 482, + 175 + ], + "type": "inline_equation", + "content": "\\tilde{z}_t" + }, + { + "bbox": [ + 130, + 116, + 482, + 175 + ], + "type": "text", + "content": ", and vice versa. This approach allows us to conduct one-stage super-resolution without undergoing all the low-resolution denoising steps, thereby improving inference efficiency. Algorithm 1 further illustrates this unification and fusion process." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 131, + 194, + 230, + 209 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 131, + 194, + 230, + 209 + ], + "spans": [ + { + "bbox": [ + 131, + 194, + 230, + 209 + ], + "type": "text", + "content": "5 Experiments" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 130, + 221, + 482, + 348 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 221, + 482, + 348 + ], + "spans": [ + { + "bbox": [ + 130, + 221, + 482, + 348 + ], + "type": "text", + "content": "Implementation Details. We adopt the pretrained Variational Autoencoder (VAE) model from SDXL [22], and we build our structure and texture generator based on the architecture of its U-Net model with the following modifications. To achieve bilingual text-to-image generation (Chinese and English), we pretrain a Chinese text encoder [9, 36] on our Chinese training dataset. We then concatenate the text embeddings from this Chinese text encoder with those from a pretrained English text encoder, serving as the final text embeddings for the denoising models. For multi-resolution image generation, we select a range of image resolutions around " + }, + { + "bbox": [ + 130, + 221, + 482, + 348 + ], + "type": "inline_equation", + "content": "1024 \\times 1024" + }, + { + "bbox": [ + 130, + 221, + 482, + 348 + ], + "type": "text", + "content": " and further condition the denoising model on the sinusoidal positional embeddings corresponding to the index of image resolutions. The " + }, + { + "bbox": [ + 130, + 221, + 482, + 348 + ], + "type": "inline_equation", + "content": "T_{struct}" + }, + { + "bbox": [ + 130, + 221, + 482, + 348 + ], + "type": "text", + "content": " parameter is set to 500, as suggested by our ablation study." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 130, + 348, + 482, + 486 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 348, + 482, + 486 + ], + "spans": [ + { + "bbox": [ + 130, + 348, + 482, + 486 + ], + "type": "text", + "content": "Dataset Construction. To encompass the abundant concepts in the world, we collect images in various styles from multiple sources, including Noah-Wukong [9], LAION [27], and others, such as photography, cartoons, portraits, and gaming assets. The collected images are filtered based on CLIP score, aesthetic score, watermark presence, resolution, and aspect ratio. To improve the semantic alignment of PanGu-Draw, we discard parts of the noisy captions that are meaningless or mismatched to the image, sourced from the Internet. Instead, we recapture the collected images by first employing an open-vocabulary detector [35] to locate the primary subjects within the images. These subjects are then processed by LLaVA [18], a high-performance vision-language model, along with prompting templates, to yield detailed image descriptions. These English annotations are subsequently translated into Chinese." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 130, + 487, + 482, + 568 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 487, + 482, + 568 + ], + "spans": [ + { + "bbox": [ + 130, + 487, + 482, + 568 + ], + "type": "text", + "content": "Evaluation Metrics. We evaluate PanGu-Draw's text-to-image generation on COCO [17] with 30k images for English, and COCO-CN [16] with 10k images for Chinese. The Frechet Inception Distance (FID [10]) is utilized to evaluate image quality and diversity. For Chinese, additional metrics include the Inception Score (IS [30]) and CN-CLIP-score [34], assessing image quality and text-image alignment. Besides, a user study is conducted to evaluate image-text alignment, fidelity, and aesthetics using ImageEval-prompt1 across 339 prompts." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 131, + 587, + 293, + 600 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 131, + 587, + 293, + 600 + ], + "spans": [ + { + "bbox": [ + 131, + 587, + 293, + 600 + ], + "type": "text", + "content": "5.1 Text-to-Image Generation" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 130, + 609, + 482, + 645 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 609, + 482, + 645 + ], + "spans": [ + { + "bbox": [ + 130, + 609, + 482, + 645 + ], + "type": "text", + "content": "Evaluation on COCO. As shown in Table 1, PanGu-Draw achieves a FID of 7.99, which is superior to compared methods such as DALL-E 2 and SDXL. It also achieves competitive FID with SOTA methods, indicating the effectiveness" + } + ] + } + ], + "index": 8 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 392, + 91, + 447, + 101 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 392, + 91, + 447, + 101 + ], + "spans": [ + { + "bbox": [ + 392, + 91, + 447, + 101 + ], + "type": "text", + "content": "PanGu-Draw" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 474, + 91, + 481, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 474, + 91, + 481, + 100 + ], + "spans": [ + { + "bbox": [ + 474, + 91, + 481, + 100 + ], + "type": "text", + "content": "9" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 133, + 653, + 405, + 665 + ], + "type": "page_footnote", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 653, + 405, + 665 + ], + "spans": [ + { + "bbox": [ + 133, + 653, + 405, + 665 + ], + "type": "text", + "content": "1 https://github.com/FlagOpen/FlagEval/tree/master/imageEval" + } + ] + } + ], + "index": 9 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 8 + }, + { + "para_blocks": [ + { + "type": "table", + "bbox": [ + 213, + 146, + 397, + 270 + ], + "blocks": [ + { + "bbox": [ + 132, + 114, + 479, + 137 + ], + "lines": [ + { + "bbox": [ + 132, + 114, + 479, + 137 + ], + "spans": [ + { + "bbox": [ + 132, + 114, + 479, + 137 + ], + "type": "text", + "content": "Table 1: Comparisons of PanGu-Draw with recent representative English text-to-image generation models on COCO dataset in terms of FID." + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 213, + 146, + 397, + 270 + ], + "lines": [ + { + "bbox": [ + 213, + 146, + 397, + 270 + ], + "spans": [ + { + "bbox": [ + 213, + 146, + 397, + 270 + ], + "type": "table", + "html": "
MethodFID↓Model SizeRelease
DALL-E [25]27.5012BN
LDM [26]12.631.5BY
GLIDE [20]12.245BN
SDXL [22]11.932.5BY
PixArt-α [4]10.650.6BY
DALL-E 2 [24]10.395.5BN
Imagen [29]7.273BN
RAPHAEL [33]6.613BN
PanGu-Draw7.995BY
", + "image_path": "294bd8c65fd7edcdbb520777904f3bf5721646004ff33194c243f1ce475ae47d.jpg" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "table_body" + } + ], + "index": 3 + }, + { + "type": "table", + "bbox": [ + 195, + 326, + 416, + 391 + ], + "blocks": [ + { + "bbox": [ + 130, + 283, + 479, + 316 + ], + "lines": [ + { + "bbox": [ + 130, + 283, + 479, + 316 + ], + "spans": [ + { + "bbox": [ + 130, + 283, + 479, + 316 + ], + "type": "text", + "content": "Table 2: Comparisons of PanGu-Draw with Chinese text-to-image generation models on COCO-CN dataset in terms of FID, IS and CN-CLIP-score. The classifier-free guidance scales are set as 9 following AltDiffusion [37]." + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 195, + 326, + 416, + 391 + ], + "lines": [ + { + "bbox": [ + 195, + 326, + 416, + 391 + ], + "spans": [ + { + "bbox": [ + 195, + 326, + 416, + 391 + ], + "type": "table", + "html": "
ModelFID↓IS↑CN-CLIP-score↑
AltDiffusion [37]25.3129.1635.12
Taiyi-Bilingual [38]24.6134.2932.26
Taiyi-CN [38]23.9934.2934.22
PanGu-Draw21.8137.0036.62
", + "image_path": "b0e7a9fde7a6900a3624779614277c5e2b276cc3d422882f206bbb7dc1b934f1.jpg" + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "table_body" + } + ], + "index": 5 + }, + { + "type": "table", + "bbox": [ + 199, + 434, + 411, + 511 + ], + "blocks": [ + { + "bbox": [ + 130, + 403, + 479, + 424 + ], + "lines": [ + { + "bbox": [ + 130, + 403, + 479, + 424 + ], + "spans": [ + { + "bbox": [ + 130, + 403, + 479, + 424 + ], + "type": "text", + "content": "Table 3: Results of a User study on ImageVal-prompt in terms of image-text alignment, image fidelity, and aesthetics." + } + ] + } + ], + "index": 6, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 199, + 434, + 411, + 511 + ], + "lines": [ + { + "bbox": [ + 199, + 434, + 411, + 511 + ], + "spans": [ + { + "bbox": [ + 199, + 434, + 411, + 511 + ], + "type": "table", + "html": "
MethodAlign↑Fidelity↑Aesthetics↑Ave↑
DALL-E 3 [2]4.724.594.764.69
MJ 5.24.634.544.754.64
SDXL [22]4.414.374.594.46
SD [26]4.173.994.204.12
PanGu-Draw4.54.524.724.58
", + "image_path": "5ab0b0f1b7d8a1abdd0bb83276f1b8bb868a235d06def14a1a3d034a3697521b.jpg" + } + ] + } + ], + "index": 7, + "angle": 0, + "type": "table_body" + } + ], + "index": 7 + }, + { + "bbox": [ + 130, + 536, + 480, + 558 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 536, + 480, + 558 + ], + "spans": [ + { + "bbox": [ + 130, + 536, + 480, + 558 + ], + "type": "text", + "content": "of our time-decoupling training strategy and its outstanding data and training efficiencies. Our 5B PanGu model is the best-released model in terms of FID." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 130, + 560, + 481, + 617 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 560, + 481, + 617 + ], + "spans": [ + { + "bbox": [ + 130, + 560, + 481, + 617 + ], + "type": "text", + "content": "Evaluation on COCO-CN. As shown in Table 2, PanGu-Draw outperforms other released Chinese text-to-image models, including Taiyi-CN, Taiyi-Bilingual, and AltDiffusion, across all three metrics. This performance highlights PanGu-Draw's exceptional Chinese text-to-image generation capabilities and the effectiveness of our bilingual text encoder architecture." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 130, + 619, + 481, + 665 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 619, + 481, + 665 + ], + "spans": [ + { + "bbox": [ + 130, + 619, + 481, + 665 + ], + "type": "text", + "content": "User Study. We conducted a user study to compare PanGu-Draw with top-performing methods, including SDXL [22], Midjourney 5.2, and DALL-E 3 [2]. As shown in Table 3, PanGu-Draw achieves better results than SD and SDXL across all three metrics. It also attains approximately " + }, + { + "bbox": [ + 130, + 619, + 481, + 665 + ], + "type": "inline_equation", + "content": "99\\% / 98\\%" + }, + { + "bbox": [ + 130, + 619, + 481, + 665 + ], + "type": "text", + "content": " of the performance of" + } + ] + } + ], + "index": 10 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "spans": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "type": "text", + "content": "10" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 166, + 91, + 215, + 100 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 166, + 91, + 215, + 100 + ], + "spans": [ + { + "bbox": [ + 166, + 91, + 215, + 100 + ], + "type": "text", + "content": "G. Lu et al." + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 9 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 149, + 114, + 227, + 190 + ], + "blocks": [ + { + "bbox": [ + 149, + 114, + 227, + 190 + ], + "lines": [ + { + "bbox": [ + 149, + 114, + 227, + 190 + ], + "spans": [ + { + "bbox": [ + 149, + 114, + 227, + 190 + ], + "type": "image", + "image_path": "2201a44d5e082f7f387e156237a94b6d62479a7b00fd67e29d751a18ac5fa91f.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 153, + 191, + 226, + 211 + ], + "lines": [ + { + "bbox": [ + 153, + 191, + 226, + 211 + ], + "spans": [ + { + "bbox": [ + 153, + 191, + 226, + 211 + ], + "type": "text", + "content": "一位老年女性，身着优雅西服，佩戴菲黛丝帽，身穿夹紧对联鞋。伸出手臂，她的背景是夕阳的校园：“一个 young woman, wearing a elegant gown and traditional dress, carrying a heavy bag of luggage and arms, with the sunset on campus in the background.”" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_caption" + } + ], + "index": 2 + }, + { + "type": "image", + "bbox": [ + 229, + 114, + 306, + 191 + ], + "blocks": [ + { + "bbox": [ + 229, + 114, + 306, + 191 + ], + "lines": [ + { + "bbox": [ + 229, + 114, + 306, + 191 + ], + "spans": [ + { + "bbox": [ + 229, + 114, + 306, + 191 + ], + "type": "image", + "image_path": "88a7bac765b94cc8235d430ba29e851f16ba26e1044b25bd561a6550234ae2da.jpg" + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 230, + 191, + 304, + 211 + ], + "lines": [ + { + "bbox": [ + 230, + 191, + 304, + 211 + ], + "spans": [ + { + "bbox": [ + 230, + 191, + 304, + 211 + ], + "type": "text", + "content": "一面像微笑的女子，身穿白色针织，红色长衣在阳光下翩翩生舞，面庞清新，风姿绰约，细尘一片。“A smiling woman wearing a white T-shirt and a red blouse with a black skirt, a red blouse, a red blouse, a red blouse, a red blouse, a red blouse, a red blouse, a red blouse, a red blouse, a red blouse, a red blouse, a red blouse, a red blouse, a red blouse, a red blouse, a red blouse, a red blouse, a red blouse, a red blouse, a red blouse, a red blouse, a red blouse, a red blouse, a red blouse, a red blouse, a red blouse, a red dress, a red dress, a red dress, a red dress, a red dress, a red dress, a red dress, a red dress, a red dress, a red dress, a red dress, a red dress, a red dress, a red dress" + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "image_caption" + } + ], + "index": 4 + }, + { + "type": "image", + "bbox": [ + 307, + 114, + 384, + 191 + ], + "blocks": [ + { + "bbox": [ + 307, + 114, + 384, + 191 + ], + "lines": [ + { + "bbox": [ + 307, + 114, + 384, + 191 + ], + "spans": [ + { + "bbox": [ + 307, + 114, + 384, + 191 + ], + "type": "image", + "image_path": "7fdf2a7a3ac4de0e0c409df4c9fd0d2f94eb60f9d4229340e33de5f392f93325.jpg" + } + ] + } + ], + "index": 6, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 310, + 191, + 383, + 211 + ], + "lines": [ + { + "bbox": [ + 310, + 191, + 383, + 211 + ], + "spans": [ + { + "bbox": [ + 310, + 191, + 383, + 211 + ], + "type": "text", + "content": "“一个巨大的水球体，内部蕴含着一个微型雨林，雨林中翅膀翱翔，阳光透过树洒洒洒。”“Auge crystal ball contains a miniature rainforest inside, with butterflies flying in the air, and sunlight shining through the leaves.”" + } + ] + } + ], + "index": 7, + "angle": 0, + "type": "image_caption" + } + ], + "index": 6 + }, + { + "type": "image", + "bbox": [ + 386, + 114, + 463, + 191 + ], + "blocks": [ + { + "bbox": [ + 386, + 114, + 463, + 191 + ], + "lines": [ + { + "bbox": [ + 386, + 114, + 463, + 191 + ], + "spans": [ + { + "bbox": [ + 386, + 114, + 463, + 191 + ], + "type": "image", + "image_path": "9d8b18b9bb763e9e709a4d4d81227bed8a74d9800437ae841b1863b2805558f8.jpg" + } + ] + } + ], + "index": 8, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 389, + 191, + 462, + 211 + ], + "lines": [ + { + "bbox": [ + 389, + 191, + 462, + 211 + ], + "spans": [ + { + "bbox": [ + 389, + 191, + 462, + 211 + ], + "type": "text", + "content": "“一只穿着中世纪短发的兔子，手长颈尖站在一座古董城的城墙上，背后是满目的杂草。” \n\"A rabbit wearing medieval armor and holding a sword stands on the wall of the tower, with a blade with the setting sun behind him.\"" + } + ] + } + ], + "index": 9, + "angle": 0, + "type": "image_caption" + } + ], + "index": 8 + }, + { + "type": "image", + "bbox": [ + 149, + 217, + 227, + 369 + ], + "blocks": [ + { + "bbox": [ + 149, + 217, + 227, + 369 + ], + "lines": [ + { + "bbox": [ + 149, + 217, + 227, + 369 + ], + "spans": [ + { + "bbox": [ + 149, + 217, + 227, + 369 + ], + "type": "image", + "image_path": "2eeafd652288659ce2f5e1d7c7e8f27eaff1ea04694c96eabb732f80ab7d24f4.jpg" + } + ] + } + ], + "index": 10, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 153, + 370, + 225, + 390 + ], + "lines": [ + { + "bbox": [ + 153, + 370, + 225, + 390 + ], + "spans": [ + { + "bbox": [ + 153, + 370, + 225, + 390 + ], + "type": "text", + "content": "“赛博办风面摄影的抓机人，依托在夜空飞行平安，以给予水面文明风貌，具有强烈的光影效果。在2019年北京冬奥会上，运动员们穿梭在的night sky, presented in particle ink painting style, with strong light and shadow effects.\"" + } + ] + } + ], + "index": 11, + "angle": 0, + "type": "image_caption" + } + ], + "index": 10 + }, + { + "type": "image", + "bbox": [ + 230, + 214, + 384, + 289 + ], + "blocks": [ + { + "bbox": [ + 230, + 214, + 384, + 289 + ], + "lines": [ + { + "bbox": [ + 230, + 214, + 384, + 289 + ], + "spans": [ + { + "bbox": [ + 230, + 214, + 384, + 289 + ], + "type": "image", + "image_path": "81b252e3489c342a012d7293b1eaedf5056fa2a236ccb9593b24496e0e236f0a.jpg" + } + ] + } + ], + "index": 12, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 233, + 290, + 381, + 303 + ], + "lines": [ + { + "bbox": [ + 233, + 290, + 381, + 303 + ], + "spans": [ + { + "bbox": [ + 233, + 290, + 381, + 303 + ], + "type": "text", + "content": "“旅行者们正奔在光激染湖的湖面上，周围是雄伟的山脉，以中国水墨画风格描绘，画面也彩淡雅致，具有古朴诗志。”（1980年） \n“旅行者们正奔在湖面上，周围是美丽的山林，以自然、静谧、清新、宁静、和谐、庄严、庄严、庄严、庄严、庄严、庄严、庄严、庄严、庄严、庄严、庄严、庄严、庄严、庄严、庄严、庄严、庄严、庄严、庄严、庄严、庄严、庄严、庄严、庄严、庄严、庄严、庄严、庄严、庄严、庄严、庄严、庄严、庄严、庄严、庄严、庄严、庄严、庄严、庄严、庄严、庄严、庄严、庄严、庄严、庄严、庄严、庄严、庄严、庄严、庄严、庄重的白色。" + } + ] + } + ], + "index": 13, + "angle": 0, + "type": "image_caption" + } + ], + "index": 12 + }, + { + "type": "image", + "bbox": [ + 229, + 304, + 306, + 382 + ], + "blocks": [ + { + "bbox": [ + 229, + 304, + 306, + 382 + ], + "lines": [ + { + "bbox": [ + 229, + 304, + 306, + 382 + ], + "spans": [ + { + "bbox": [ + 229, + 304, + 306, + 382 + ], + "type": "image", + "image_path": "78246ef7da538a87b6398557cb88e04794f8bdc28b5f9a41f07772874ce19e1e.jpg" + } + ] + } + ], + "index": 14, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 233, + 382, + 304, + 399 + ], + "lines": [ + { + "bbox": [ + 233, + 382, + 304, + 399 + ], + "spans": [ + { + "bbox": [ + 233, + 382, + 304, + 399 + ], + "type": "text", + "content": "“一台未来风格的摩托车，闪耀着霓虹灯，停在夜晚的东京街头。” \n“A futuristic motorcycle, shining with neon lights, is parked on the streets of Tokyo at night.”" + } + ] + } + ], + "index": 15, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 131, + 407, + 482, + 441 + ], + "lines": [ + { + "bbox": [ + 131, + 407, + 482, + 441 + ], + "spans": [ + { + "bbox": [ + 131, + 407, + 482, + 441 + ], + "type": "text", + "content": "Fig. 4: Images generated with PanGu-Draw, our 5B multi-lingual text-to-image generation model. PanGu-Draw is able to generate multi-resolution high-fidelity images semantically aligned with the input prompts." + } + ] + } + ], + "index": 20, + "angle": 0, + "type": "image_caption" + } + ], + "index": 14 + }, + { + "type": "image", + "bbox": [ + 308, + 304, + 384, + 382 + ], + "blocks": [ + { + "bbox": [ + 308, + 304, + 384, + 382 + ], + "lines": [ + { + "bbox": [ + 308, + 304, + 384, + 382 + ], + "spans": [ + { + "bbox": [ + 308, + 304, + 384, + 382 + ], + "type": "image", + "image_path": "dda03bb00997887131d6f8ce25a19e7cfc341b1dae790a6de919503eabdffd62.jpg" + } + ] + } + ], + "index": 16, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 316, + 382, + 383, + 398 + ], + "lines": [ + { + "bbox": [ + 316, + 382, + 383, + 398 + ], + "spans": [ + { + "bbox": [ + 316, + 382, + 383, + 398 + ], + "type": "text", + "content": "一座由冰晶和雪花构成的精致城堡，坐落在在北极的冰原上。 \n“An exquisite castle made of ice crystals and snowflakes, located on the Arctic ice sheet.”" + } + ] + } + ], + "index": 17, + "angle": 0, + "type": "image_caption" + } + ], + "index": 16 + }, + { + "type": "image", + "bbox": [ + 386, + 218, + 463, + 369 + ], + "blocks": [ + { + "bbox": [ + 386, + 218, + 463, + 369 + ], + "lines": [ + { + "bbox": [ + 386, + 218, + 463, + 369 + ], + "spans": [ + { + "bbox": [ + 386, + 218, + 463, + 369 + ], + "type": "image", + "image_path": "4e5071405aae96c871b8e94442ea96f379d0f4d60c8ce883e167b9323215585a.jpg" + } + ] + } + ], + "index": 18, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 390, + 370, + 459, + 382 + ], + "lines": [ + { + "bbox": [ + 390, + 370, + 459, + 382 + ], + "spans": [ + { + "bbox": [ + 390, + 370, + 459, + 382 + ], + "type": "text", + "content": "一袋老火的糖盔，完全由糖果和巧克力制成。一个 ancient pot was filled of candy and chocolate." + } + ] + } + ], + "index": 19, + "angle": 0, + "type": "image_caption" + } + ], + "index": 18 + }, + { + "bbox": [ + 131, + 462, + 482, + 510 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 131, + 462, + 482, + 510 + ], + "spans": [ + { + "bbox": [ + 131, + 462, + 482, + 510 + ], + "type": "text", + "content": "Midjourney 5.2 and DALL-E 3, respectively, indicating PanGu-Draw's excellent text-to-image capabilities. Figure 4 shows some high-fidelity multi-resolution images generated by PanGu-Draw. As we can see, the generated images of PanGu-Draw are of high aesthetics and semantically aligned with the input prompts." + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 132, + 525, + 316, + 537 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 525, + 316, + 537 + ], + "spans": [ + { + "bbox": [ + 132, + 525, + 316, + 537 + ], + "type": "text", + "content": "5.2 Multi-Diffusion Fusing Results" + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 130, + 544, + 489, + 648 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 544, + 489, + 648 + ], + "spans": [ + { + "bbox": [ + 130, + 544, + 489, + 648 + ], + "type": "text", + "content": "Multi-Control Image Generation. To demonstrate the effectiveness of the proposed reusable multi-diffusion fusing algorithm, Coop-Diffusion, we first present multiple results of multi-control image generation. Figure 5 displays results from fusing an image variation model" + }, + { + "bbox": [ + 130, + 544, + 489, + 648 + ], + "type": "inline_equation", + "content": "^2" + }, + { + "bbox": [ + 130, + 544, + 489, + 648 + ], + "type": "text", + "content": " with PanGu-Draw. The fusing results maintain a style similar to that of the reference image, matching the texture described by the input prompt. Figure 6 shows results from fusing PanGu-Draw with a pose/edge-to-image ControlNet model, which operates in guess mode without input prompts. Here, the fusing results combine the structure of the pose/edge image with the texture described by the input prompt." + } + ] + } + ], + "index": 23 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 392, + 91, + 447, + 100 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 392, + 91, + 447, + 100 + ], + "spans": [ + { + "bbox": [ + 392, + 91, + 447, + 100 + ], + "type": "text", + "content": "PanGu-Draw" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 470, + 91, + 480, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 470, + 91, + 480, + 100 + ], + "spans": [ + { + "bbox": [ + 470, + 91, + 480, + 100 + ], + "type": "text", + "content": "11" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 133, + 652, + 405, + 666 + ], + "type": "page_footnote", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 652, + 405, + 666 + ], + "spans": [ + { + "bbox": [ + 133, + 652, + 405, + 666 + ], + "type": "inline_equation", + "content": "^{2}" + }, + { + "bbox": [ + 133, + 652, + 405, + 666 + ], + "type": "text", + "content": " https://huggingface.co/lambdalabs/sd-image-variations-diffusers" + } + ] + } + ], + "index": 24 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 10 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 185, + 114, + 233, + 163 + ], + "blocks": [ + { + "bbox": [ + 185, + 114, + 233, + 163 + ], + "lines": [ + { + "bbox": [ + 185, + 114, + 233, + 163 + ], + "spans": [ + { + "bbox": [ + 185, + 114, + 233, + 163 + ], + "type": "image", + "image_path": "7adbcae4a56790853f9fae971c5ffb910d46fb82f33d32a5639812f891912d04.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 195, + 163, + 318, + 171 + ], + "lines": [ + { + "bbox": [ + 195, + 163, + 318, + 171 + ], + "spans": [ + { + "bbox": [ + 195, + 163, + 318, + 171 + ], + "type": "text", + "content": "(a) Input image and generation results of image variation." + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "image_caption" + } + ], + "index": 2 + }, + { + "type": "image", + "bbox": [ + 235, + 114, + 329, + 163 + ], + "blocks": [ + { + "bbox": [ + 235, + 114, + 329, + 163 + ], + "lines": [ + { + "bbox": [ + 235, + 114, + 329, + 163 + ], + "spans": [ + { + "bbox": [ + 235, + 114, + 329, + 163 + ], + "type": "image", + "image_path": "0f0e5f856510b0ed5aaff8f431bc6d24ad8fbe5d7825e09f232bfdbb36637d71.jpg" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_body" + } + ], + "index": 3 + }, + { + "type": "image", + "bbox": [ + 193, + 170, + 225, + 202 + ], + "blocks": [ + { + "bbox": [ + 193, + 170, + 225, + 202 + ], + "lines": [ + { + "bbox": [ + 193, + 170, + 225, + 202 + ], + "spans": [ + { + "bbox": [ + 193, + 170, + 225, + 202 + ], + "type": "image", + "image_path": "bc637749dfbb0cca22b06351b7c54cd300f3577a43dbb47c3f8f617e1b5a7733.jpg" + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "image_body" + } + ], + "index": 5 + }, + { + "type": "image", + "bbox": [ + 235, + 172, + 329, + 220 + ], + "blocks": [ + { + "bbox": [ + 235, + 172, + 329, + 220 + ], + "lines": [ + { + "bbox": [ + 235, + 172, + 329, + 220 + ], + "spans": [ + { + "bbox": [ + 235, + 172, + 329, + 220 + ], + "type": "image", + "image_path": "c4d7506482d107c66fda628776989c67a3519d04d2f6efd960a1a096806b99f6.jpg" + } + ] + } + ], + "index": 6, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 196, + 221, + 318, + 228 + ], + "lines": [ + { + "bbox": [ + 196, + 221, + 318, + 228 + ], + "spans": [ + { + "bbox": [ + 196, + 221, + 318, + 228 + ], + "type": "text", + "content": "(b) Input prompt and generation results of PanGu-Draw." + } + ] + } + ], + "index": 7, + "angle": 0, + "type": "image_caption" + } + ], + "index": 6 + }, + { + "type": "image", + "bbox": [ + 334, + 114, + 429, + 209 + ], + "blocks": [ + { + "bbox": [ + 334, + 114, + 429, + 209 + ], + "lines": [ + { + "bbox": [ + 334, + 114, + 429, + 209 + ], + "spans": [ + { + "bbox": [ + 334, + 114, + 429, + 209 + ], + "type": "image", + "image_path": "7f225c96c36d5735ec6785e7c15c7612048b2fb49e3abc9830d9ad078d997480.jpg" + } + ] + } + ], + "index": 8, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 338, + 209, + 424, + 231 + ], + "lines": [ + { + "bbox": [ + 338, + 209, + 424, + 231 + ], + "spans": [ + { + "bbox": [ + 338, + 209, + 424, + 231 + ], + "type": "text", + "content": "(c) Generation results of fusing image variation and PanGu-Draw with Coop-Diffusion algorithm." + } + ] + } + ], + "index": 9, + "angle": 0, + "type": "image_caption" + } + ], + "index": 8 + }, + { + "type": "image", + "bbox": [ + 189, + 284, + 236, + 332 + ], + "blocks": [ + { + "bbox": [ + 131, + 245, + 480, + 268 + ], + "lines": [ + { + "bbox": [ + 131, + 245, + 480, + 268 + ], + "spans": [ + { + "bbox": [ + 131, + 245, + 480, + 268 + ], + "type": "text", + "content": "Fig. 5: Generation results of the fusing of an image variation model and PanGu-Draw and with the proposed Coop-Diffusion algorithm." + } + ] + } + ], + "index": 10, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 189, + 284, + 236, + 332 + ], + "lines": [ + { + "bbox": [ + 189, + 284, + 236, + 332 + ], + "spans": [ + { + "bbox": [ + 189, + 284, + 236, + 332 + ], + "type": "image", + "image_path": "5b2abb2a30c0943316d2a6728ea7d6f55a1a4ccaa7b2c866e9cf5d94c53d915f.jpg" + } + ] + } + ], + "index": 11, + "angle": 0, + "type": "image_body" + } + ], + "index": 11 + }, + { + "type": "image", + "bbox": [ + 184, + 333, + 317, + 388 + ], + "blocks": [ + { + "bbox": [ + 184, + 333, + 317, + 388 + ], + "lines": [ + { + "bbox": [ + 184, + 333, + 317, + 388 + ], + "spans": [ + { + "bbox": [ + 184, + 333, + 317, + 388 + ], + "type": "image", + "image_path": "8417a5359dcaf7329a496568c9e1950911c95755f385dfc9f541411a150d953a.jpg" + } + ] + } + ], + "index": 12, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 199, + 389, + 318, + 396 + ], + "lines": [ + { + "bbox": [ + 199, + 389, + 318, + 396 + ], + "spans": [ + { + "bbox": [ + 199, + 389, + 318, + 396 + ], + "type": "text", + "content": "(b) Input prompt and generation results of PanGu-Draw." + } + ] + } + ], + "index": 13, + "angle": 0, + "type": "image_caption" + } + ], + "index": 12 + }, + { + "type": "image", + "bbox": [ + 189, + 404, + 236, + 452 + ], + "blocks": [ + { + "bbox": [ + 189, + 404, + 236, + 452 + ], + "lines": [ + { + "bbox": [ + 189, + 404, + 236, + 452 + ], + "spans": [ + { + "bbox": [ + 189, + 404, + 236, + 452 + ], + "type": "image", + "image_path": "edb622cc78a1a3d9f5df08900d8b98db1466a4cb1981153de1d85351f8f86a99.jpg" + } + ] + } + ], + "index": 14, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 184, + 453, + 318, + 460 + ], + "lines": [ + { + "bbox": [ + 184, + 453, + 318, + 460 + ], + "spans": [ + { + "bbox": [ + 184, + 453, + 318, + 460 + ], + "type": "text", + "content": "(a) Input edge map and generation results of edge-to-image" + } + ] + } + ], + "index": 15, + "angle": 0, + "type": "image_caption" + } + ], + "index": 14 + }, + { + "type": "image", + "bbox": [ + 190, + 461, + 329, + 509 + ], + "blocks": [ + { + "bbox": [ + 190, + 461, + 329, + 509 + ], + "lines": [ + { + "bbox": [ + 190, + 461, + 329, + 509 + ], + "spans": [ + { + "bbox": [ + 190, + 461, + 329, + 509 + ], + "type": "image", + "image_path": "92dbdb4fb327017281fde2c184adfe6856c0d6a88d7633492ae3453cc52a8b39.jpg" + } + ] + } + ], + "index": 16, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 199, + 510, + 318, + 517 + ], + "lines": [ + { + "bbox": [ + 199, + 510, + 318, + 517 + ], + "spans": [ + { + "bbox": [ + 199, + 510, + 318, + 517 + ], + "type": "text", + "content": "(b) Input prompt and generation results of PanGu-Draw." + } + ] + } + ], + "index": 17, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 131, + 533, + 481, + 555 + ], + "lines": [ + { + "bbox": [ + 131, + 533, + 481, + 555 + ], + "spans": [ + { + "bbox": [ + 131, + 533, + 481, + 555 + ], + "type": "text", + "content": "Fig. 6: Generation results guided by fusing signals of text and pose/edge map by our Coop-Diffusion." + } + ] + } + ], + "index": 22, + "angle": 0, + "type": "image_caption" + } + ], + "index": 16 + }, + { + "type": "image", + "bbox": [ + 335, + 284, + 428, + 376 + ], + "blocks": [ + { + "bbox": [ + 335, + 284, + 428, + 376 + ], + "lines": [ + { + "bbox": [ + 335, + 284, + 428, + 376 + ], + "spans": [ + { + "bbox": [ + 335, + 284, + 428, + 376 + ], + "type": "image", + "image_path": "9e7f9916bc96a6607b1de3f6b2ce885429ecfef8dd4b009046a9cac1d3d183f8.jpg" + } + ] + } + ], + "index": 18, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 338, + 378, + 424, + 398 + ], + "lines": [ + { + "bbox": [ + 338, + 378, + 424, + 398 + ], + "spans": [ + { + "bbox": [ + 338, + 378, + 424, + 398 + ], + "type": "text", + "content": "(c) Generation results of fusing pose-to-image and PanGu-Draw with Coop-Diffusion algorithm." + } + ] + } + ], + "index": 19, + "angle": 0, + "type": "image_caption" + } + ], + "index": 18 + }, + { + "type": "image", + "bbox": [ + 335, + 404, + 427, + 497 + ], + "blocks": [ + { + "bbox": [ + 335, + 404, + 427, + 497 + ], + "lines": [ + { + "bbox": [ + 335, + 404, + 427, + 497 + ], + "spans": [ + { + "bbox": [ + 335, + 404, + 427, + 497 + ], + "type": "image", + "image_path": "1eec92eebddea7cf18c9ea5ccfb8d7f32b866c6cc03741855dd723db25d1d1f1.jpg" + } + ] + } + ], + "index": 20, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 338, + 498, + 424, + 517 + ], + "lines": [ + { + "bbox": [ + 338, + 498, + 424, + 517 + ], + "spans": [ + { + "bbox": [ + 338, + 498, + 424, + 517 + ], + "type": "text", + "content": "(c) Generation results of fusing edge-to-image and PanGu-Draw with CoopDiffusion algorithm." + } + ] + } + ], + "index": 21, + "angle": 0, + "type": "image_caption" + } + ], + "index": 20 + }, + { + "bbox": [ + 130, + 584, + 482, + 666 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 584, + 482, + 666 + ], + "spans": [ + { + "bbox": [ + 130, + 584, + 482, + 666 + ], + "type": "text", + "content": "Multi-Resolution Image Generation. We also present multi-resolution image generation results of fusing PanGu-Draw with low-resolution text-to-image and edge-to-image ControlNet model by first denoising with the low-resolution model to get the intermediate " + }, + { + "bbox": [ + 130, + 584, + 482, + 666 + ], + "type": "inline_equation", + "content": "z_{t}^{\\prime}" + }, + { + "bbox": [ + 130, + 584, + 482, + 666 + ], + "type": "text", + "content": " and the high-resolution counterpart " + }, + { + "bbox": [ + 130, + 584, + 482, + 666 + ], + "type": "inline_equation", + "content": "\\tilde{z}_{t}" + }, + { + "bbox": [ + 130, + 584, + 482, + 666 + ], + "type": "text", + "content": ", and then perform denoising in high resolution with PanGu-Draw. Figure 7 shows the results from the low-resolution model and our fusing algorithm Coop-Diffusion. As we can see, PanGu-Draw adds much details to the low-resolution predictions" + } + ] + } + ], + "index": 23 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "spans": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "type": "text", + "content": "12" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 166, + 91, + 215, + 100 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 166, + 91, + 215, + 100 + ], + "spans": [ + { + "bbox": [ + 166, + 91, + 215, + 100 + ], + "type": "text", + "content": "G. Lu et al." + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 11 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 219, + 114, + 306, + 201 + ], + "blocks": [ + { + "bbox": [ + 219, + 114, + 306, + 201 + ], + "lines": [ + { + "bbox": [ + 219, + 114, + 306, + 201 + ], + "spans": [ + { + "bbox": [ + 219, + 114, + 306, + 201 + ], + "type": "image", + "image_path": "e0bcc639f1e4855ccf61fed739fa729c95380f9dcad585e1c9d243b10c6105c5.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_body" + } + ], + "index": 2 + }, + { + "type": "image", + "bbox": [ + 307, + 114, + 394, + 201 + ], + "blocks": [ + { + "bbox": [ + 307, + 114, + 394, + 201 + ], + "lines": [ + { + "bbox": [ + 307, + 114, + 394, + 201 + ], + "spans": [ + { + "bbox": [ + 307, + 114, + 394, + 201 + ], + "type": "image", + "image_path": "e0bc84a0d5e802d8500382c3c483ea9d1359513aab38ab16f640ff6ccef25bab.jpg" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_body" + } + ], + "index": 3 + }, + { + "type": "image", + "bbox": [ + 219, + 202, + 306, + 289 + ], + "blocks": [ + { + "bbox": [ + 219, + 202, + 306, + 289 + ], + "lines": [ + { + "bbox": [ + 219, + 202, + 306, + 289 + ], + "spans": [ + { + "bbox": [ + 219, + 202, + 306, + 289 + ], + "type": "image", + "image_path": "96ae32efe5cb26a9e555f8471c91d3dcc1eac5e8dfe36b104dec01b847a1951d.jpg" + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 252, + 289, + 274, + 296 + ], + "lines": [ + { + "bbox": [ + 252, + 289, + 274, + 296 + ], + "spans": [ + { + "bbox": [ + 252, + 289, + 274, + 296 + ], + "type": "text", + "content": "LR model" + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "image_caption" + } + ], + "index": 4 + }, + { + "type": "image", + "bbox": [ + 307, + 201, + 394, + 289 + ], + "blocks": [ + { + "bbox": [ + 307, + 201, + 394, + 289 + ], + "lines": [ + { + "bbox": [ + 307, + 201, + 394, + 289 + ], + "spans": [ + { + "bbox": [ + 307, + 201, + 394, + 289 + ], + "type": "image", + "image_path": "c50b523831a80171738df1ab3f8872dbaaf9291deeff5220f4380bcc71a9afb8.jpg" + } + ] + } + ], + "index": 6, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 326, + 289, + 375, + 296 + ], + "lines": [ + { + "bbox": [ + 326, + 289, + 375, + 296 + ], + "spans": [ + { + "bbox": [ + 326, + 289, + 375, + 296 + ], + "type": "text", + "content": "LR model + HR model" + } + ] + } + ], + "index": 7, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 131, + 304, + 482, + 360 + ], + "lines": [ + { + "bbox": [ + 131, + 304, + 482, + 360 + ], + "spans": [ + { + "bbox": [ + 131, + 304, + 482, + 360 + ], + "type": "text", + "content": "Fig. 7: Images generated with a low-resolution (LR) model (1st row: T2I model; 2nd row: edge-to-image ControlNet) and the fusion of the LR model and HR PanGu-Draw with our Coop-Diffusion. The resolutions are " + }, + { + "bbox": [ + 131, + 304, + 482, + 360 + ], + "type": "inline_equation", + "content": "512 \\times 512" + }, + { + "bbox": [ + 131, + 304, + 482, + 360 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 131, + 304, + 482, + 360 + ], + "type": "inline_equation", + "content": "1024 \\times 1024" + }, + { + "bbox": [ + 131, + 304, + 482, + 360 + ], + "type": "text", + "content": " respectively. This allows for single-stage super-resolution for better details and higher inference efficiency." + } + ] + } + ], + "index": 8, + "angle": 0, + "type": "image_caption" + } + ], + "index": 6 + }, + { + "bbox": [ + 131, + 384, + 482, + 419 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 131, + 384, + 482, + 419 + ], + "spans": [ + { + "bbox": [ + 131, + 384, + 482, + 419 + ], + "type": "text", + "content": "leading to high-fidelity high-resolution results. Besides, compared with the common practice of super-resolution with diffusion model, which carries out all the low-resolution denoising steps, our method achieve higher inference efficiency." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 132, + 437, + 238, + 449 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 437, + 238, + 449 + ], + "spans": [ + { + "bbox": [ + 132, + 437, + 238, + 449 + ], + "type": "text", + "content": "5.3 Ablation Study" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 130, + 457, + 482, + 538 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 457, + 482, + 538 + ], + "spans": [ + { + "bbox": [ + 130, + 457, + 482, + 538 + ], + "type": "text", + "content": "In this section, we perform ablation studies to analyze our time-decoupling training strategy. The baseline model has " + }, + { + "bbox": [ + 130, + 457, + 482, + 538 + ], + "type": "inline_equation", + "content": "1B" + }, + { + "bbox": [ + 130, + 457, + 482, + 538 + ], + "type": "text", + "content": " parameters while the structure and texture generators both have " + }, + { + "bbox": [ + 130, + 457, + 482, + 538 + ], + "type": "inline_equation", + "content": "0.5B" + }, + { + "bbox": [ + 130, + 457, + 482, + 538 + ], + "type": "text", + "content": " parameters. During the training process, the latter two models only train half the steps of the baseline model with " + }, + { + "bbox": [ + 130, + 457, + 482, + 538 + ], + "type": "inline_equation", + "content": "T_{struct}" + }, + { + "bbox": [ + 130, + 457, + 482, + 538 + ], + "type": "text", + "content": " set as 500. Both settings of the models are trained from scratch on a subset of the LAION dataset containing images with all sizes. After training, FID, IS and CLIP-score on COCO are reported for comparison." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 130, + 539, + 482, + 584 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 539, + 482, + 584 + ], + "spans": [ + { + "bbox": [ + 130, + 539, + 482, + 584 + ], + "type": "text", + "content": "Time-Decoupling Training Strategy. We compare the performance of models trained with the Resolution Boost strategy and our time-decoupling strategy in Table 4. We found that models trained with our strategy achieves better performance in all three criteria, indicating the effectiveness of our strategy." + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 130, + 585, + 482, + 665 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 585, + 482, + 665 + ], + "spans": [ + { + "bbox": [ + 130, + 585, + 482, + 665 + ], + "type": "text", + "content": "Training Designs. The structure and texture generators (" + }, + { + "bbox": [ + 130, + 585, + 482, + 665 + ], + "type": "inline_equation", + "content": "\\epsilon_{\\mathrm{struct}}" + }, + { + "bbox": [ + 130, + 585, + 482, + 665 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 130, + 585, + 482, + 665 + ], + "type": "inline_equation", + "content": "\\epsilon_{\\mathrm{texture}}" + }, + { + "bbox": [ + 130, + 585, + 482, + 665 + ], + "type": "text", + "content": ") are designed to train on different resolutions to improve data and training efficiency. However, this approach may negatively influence the final performance. In Table 5, we compare such a design with a traditional training process, where " + }, + { + "bbox": [ + 130, + 585, + 482, + 665 + ], + "type": "inline_equation", + "content": "\\epsilon_{\\mathrm{struct}}" + }, + { + "bbox": [ + 130, + 585, + 482, + 665 + ], + "type": "text", + "content": " discards low-resolution images, or " + }, + { + "bbox": [ + 130, + 585, + 482, + 665 + ], + "type": "inline_equation", + "content": "\\epsilon_{\\mathrm{texture}}" + }, + { + "bbox": [ + 130, + 585, + 482, + 665 + ], + "type": "text", + "content": " trains with high resolution. Results on COCO show that " + }, + { + "bbox": [ + 130, + 585, + 482, + 665 + ], + "type": "inline_equation", + "content": "\\epsilon_{\\mathrm{struct}}" + }, + { + "bbox": [ + 130, + 585, + 482, + 665 + ], + "type": "text", + "content": " benefits from these extra up-scaled data, and " + }, + { + "bbox": [ + 130, + 585, + 482, + 665 + ], + "type": "inline_equation", + "content": "\\epsilon_{\\mathrm{texture}}" + }, + { + "bbox": [ + 130, + 585, + 482, + 665 + ], + "type": "text", + "content": " learns enough texture patterns at a smaller resolution." + } + ] + } + ], + "index": 13 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 392, + 91, + 447, + 100 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 392, + 91, + 447, + 100 + ], + "spans": [ + { + "bbox": [ + 392, + 91, + 447, + 100 + ], + "type": "text", + "content": "PanGu-Draw" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 470, + 91, + 481, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 470, + 91, + 481, + 100 + ], + "spans": [ + { + "bbox": [ + 470, + 91, + 481, + 100 + ], + "type": "text", + "content": "13" + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 12 + }, + { + "para_blocks": [ + { + "type": "table", + "bbox": [ + 207, + 146, + 403, + 186 + ], + "blocks": [ + { + "bbox": [ + 132, + 114, + 479, + 137 + ], + "lines": [ + { + "bbox": [ + 132, + 114, + 479, + 137 + ], + "spans": [ + { + "bbox": [ + 132, + 114, + 479, + 137 + ], + "type": "text", + "content": "Table 4: Comparison of models across Resolution Boost (1B parameters) and Time-Decoupling training strategies (0.5B parameters for structure and texture generators)" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 207, + 146, + 403, + 186 + ], + "lines": [ + { + "bbox": [ + 207, + 146, + 403, + 186 + ], + "spans": [ + { + "bbox": [ + 207, + 146, + 403, + 186 + ], + "type": "table", + "html": "
ModelFID↓IS↑CLIP-score↑
Resolution Boost106.1210.4622.9
Time-Decoupling87.6611.0723.4
", + "image_path": "3d98fbda8ae48c71611ed7afff6a223c71d31d671095a9e250b6ef9f25afe147.jpg" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "table_body" + } + ], + "index": 3 + }, + { + "type": "table", + "bbox": [ + 157, + 228, + 454, + 283 + ], + "blocks": [ + { + "bbox": [ + 132, + 197, + 479, + 217 + ], + "lines": [ + { + "bbox": [ + 132, + 197, + 479, + 217 + ], + "spans": [ + { + "bbox": [ + 132, + 197, + 479, + 217 + ], + "type": "text", + "content": "Table 5: Performance of structure and texture models training with images of different resolutions." + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 157, + 228, + 454, + 283 + ], + "lines": [ + { + "bbox": [ + 157, + 228, + 454, + 283 + ], + "spans": [ + { + "bbox": [ + 157, + 228, + 454, + 283 + ], + "type": "table", + "html": "
Structure DataTexture ResolutionFID↓IS↑CLIP-score↑
All data25687.6611.0723.4
Only high resolution25689.5210.9623.2
All data51290.9810.5923.3
", + "image_path": "aff72fa8c6a45c239a4237d4b257d8bcd78f46b95c6b92c7dd055b92d5771693.jpg" + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "table_body" + } + ], + "index": 5 + }, + { + "type": "table", + "bbox": [ + 230, + 325, + 380, + 389 + ], + "blocks": [ + { + "bbox": [ + 132, + 293, + 479, + 315 + ], + "lines": [ + { + "bbox": [ + 132, + 293, + 479, + 315 + ], + "spans": [ + { + "bbox": [ + 132, + 293, + 479, + 315 + ], + "type": "text", + "content": "Table 6: Comparisons of PanGu-Draw inference performance with different time step splitting point " + }, + { + "bbox": [ + 132, + 293, + 479, + 315 + ], + "type": "inline_equation", + "content": "T_{struct}" + }, + { + "bbox": [ + 132, + 293, + 479, + 315 + ], + "type": "text", + "content": " settings." + } + ] + } + ], + "index": 6, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 230, + 325, + 380, + 389 + ], + "lines": [ + { + "bbox": [ + 230, + 325, + 380, + 389 + ], + "spans": [ + { + "bbox": [ + 230, + 325, + 380, + 389 + ], + "type": "table", + "html": "
TstructFID↓IS↑CLIP-score↑
200105.0810.5922.98
30098.0810.7223.12
50087.6611.0723.40
70089.4811.0223.32
", + "image_path": "53258af43e3eb3bc47428d38168c69683774adc2576278989d320f3d7175d18e.jpg" + } + ] + } + ], + "index": 7, + "angle": 0, + "type": "table_body" + } + ], + "index": 7 + }, + { + "bbox": [ + 130, + 411, + 480, + 491 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 411, + 480, + 491 + ], + "spans": [ + { + "bbox": [ + 130, + 411, + 480, + 491 + ], + "type": "text", + "content": "Timestep Splitting Point. The timestep splitting point " + }, + { + "bbox": [ + 130, + 411, + 480, + 491 + ], + "type": "inline_equation", + "content": "T_{\\mathrm{struct}}" + }, + { + "bbox": [ + 130, + 411, + 480, + 491 + ], + "type": "text", + "content": " between the structure and texture generators also influences the final performance. To this end, we set " + }, + { + "bbox": [ + 130, + 411, + 480, + 491 + ], + "type": "inline_equation", + "content": "T_{\\mathrm{struct}}" + }, + { + "bbox": [ + 130, + 411, + 480, + 491 + ], + "type": "text", + "content": " to 200, 300, 500, and 700, while keeping the other settings of the structure and texture generators unchanged. As shown in Table 6, as " + }, + { + "bbox": [ + 130, + 411, + 480, + 491 + ], + "type": "inline_equation", + "content": "T_{\\mathrm{struct}}" + }, + { + "bbox": [ + 130, + 411, + 480, + 491 + ], + "type": "text", + "content": " increases from 200 to 700, the performance initially increases and then decreases continuously. " + }, + { + "bbox": [ + 130, + 411, + 480, + 491 + ], + "type": "inline_equation", + "content": "T_{\\mathrm{struct}} = 500" + }, + { + "bbox": [ + 130, + 411, + 480, + 491 + ], + "type": "text", + "content": " is the optimal value, and we adopt it as the default setting in all other experiments." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 132, + 510, + 219, + 522 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 510, + 219, + 522 + ], + "spans": [ + { + "bbox": [ + 132, + 510, + 219, + 522 + ], + "type": "text", + "content": "6 Conclusion" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 130, + 534, + 481, + 661 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 534, + 481, + 661 + ], + "spans": [ + { + "bbox": [ + 130, + 534, + 481, + 661 + ], + "type": "text", + "content": "In this paper, we present \"PanGu-Draw\", a new latent diffusion model for efficient text-to-image generation that effectively integrates multiple control signals. Our approach includes a Time-Decoupling Training Strategy to separate the text-to-image process into structure and texture generation, enhancing data use and computational efficiency. Additionally, \"Coop-Diffusion\" is introduced, an algorithm allowing cooperative use of different pre-trained diffusion models in a unified denoising process for multi-control image synthesis at various resolutions without extra data or retraining. PanGu-Draw outperforms models like DALL-E 2 and SDXL in English T2I, achieves superior FID, IS, and CN-CLIP-scores in Chinese T2I, and receives favorable user feedback. This positions PanGu-Draw as a versatile and efficient state-of-the-art method." + } + ] + } + ], + "index": 10 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "spans": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "type": "text", + "content": "14" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 166, + 91, + 215, + 100 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 166, + 91, + 215, + 100 + ], + "spans": [ + { + "bbox": [ + 166, + 91, + 215, + 100 + ], + "type": "text", + "content": "G. Lu et al." + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 13 + }, + { + "para_blocks": [ + { + "bbox": [ + 133, + 114, + 197, + 126 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 114, + 197, + 126 + ], + "spans": [ + { + "bbox": [ + 133, + 114, + 197, + 126 + ], + "type": "text", + "content": "References" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 138, + 139, + 480, + 665 + ], + "type": "list", + "angle": 0, + "index": 20, + "blocks": [ + { + "bbox": [ + 138, + 139, + 480, + 171 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 139, + 480, + 171 + ], + "spans": [ + { + "bbox": [ + 138, + 139, + 480, + 171 + ], + "type": "text", + "content": "1. Balaji, Y., Nah, S., Huang, X., Vahdat, A., Song, J., Kreis, K., Aittala, M., Aila, T., Laine, S., Catanzaro, B., et al.: edifi: Text-to-image diffusion models with an ensemble of expert denoisers. arXiv preprint arXiv:2211.01324 (2022)" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 138, + 171, + 480, + 203 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 171, + 480, + 203 + ], + "spans": [ + { + "bbox": [ + 138, + 171, + 480, + 203 + ], + "type": "text", + "content": "2. Betker, J., Goh, G., Jing, L., Brooks, T., Wang, J., Li, L., Ouyang, L., Zhuang, J., Lee, J., Guo, Y., et al.: Improving image generation with better captions. Computer Science. https://cdn.opena.com/papers/dall-e-3.pdf (2023)" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 138, + 203, + 480, + 234 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 203, + 480, + 234 + ], + "spans": [ + { + "bbox": [ + 138, + 203, + 480, + 234 + ], + "type": "text", + "content": "3. Chen, C., Yin, Y., Shang, L., Jiang, X., Qin, Y., Wang, F., Wang, Z., Chen, X., Liu, Z., Liu, Q.: bert2bert: Towards reusable pretrained language models. arXiv preprint arXiv:2110.07143 (2021)" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 138, + 234, + 480, + 266 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 234, + 480, + 266 + ], + "spans": [ + { + "bbox": [ + 138, + 234, + 480, + 266 + ], + "type": "text", + "content": "4. Chen, J., Yu, J., Ge, C., Yao, L., Xie, E., Wu, Y., Wang, Z., Kwok, J., Luo, P., Lu, H., Li, Z.: Fast training of diffusion transformer for photorealistic text-to-image synthesis (2023)" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 138, + 266, + 480, + 286 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 266, + 480, + 286 + ], + "spans": [ + { + "bbox": [ + 138, + 266, + 480, + 286 + ], + "type": "text", + "content": "5. Dhariwal, P., Nichol, A.: Diffusion models beat gans on image synthesis. Advances in Neural Information Processing Systems 34, 8780-8794 (2021)" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 138, + 286, + 480, + 318 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 286, + 480, + 318 + ], + "spans": [ + { + "bbox": [ + 138, + 286, + 480, + 318 + ], + "type": "text", + "content": "6. Ding, N., Tang, Y., Han, K., Xu, C., Wang, Y.: Network expansion for practical training acceleration. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 20269-20279 (2023)" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 138, + 318, + 480, + 370 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 318, + 480, + 370 + ], + "spans": [ + { + "bbox": [ + 138, + 318, + 480, + 370 + ], + "type": "text", + "content": "7. Feng, Z., Zhang, Z., Yu, X., Fang, Y., Li, L., Chen, X., Lu, Y., Liu, J., Yin, W., Feng, S., et al.: Ernie-vilg 2.0: Improving text-to-image diffusion model with knowledge-enhanced mixture-of-denoising-experts. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 10135-10145 (2023)" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 138, + 371, + 480, + 403 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 371, + 480, + 403 + ], + "spans": [ + { + "bbox": [ + 138, + 371, + 480, + 403 + ], + "type": "text", + "content": "8. Fu, C., Huang, H., Jiang, Z., Ni, Y., Nai, L., Wu, G., Cheng, L., Zhou, Y., Li, S., Li, A., et al.: Triple: Revisiting pretrained model reuse and progressive learning for efficient vision transformer scaling and searching. In: ICCV (2023)" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 138, + 403, + 480, + 444 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 403, + 480, + 444 + ], + "spans": [ + { + "bbox": [ + 138, + 403, + 480, + 444 + ], + "type": "text", + "content": "9. Gu, J., Meng, X., Lu, G., Hou, L., Minzhe, N., Liang, X., Yao, L., Huang, R., Zhang, W., Jiang, X., et al.: Wukong: A 100 million large-scale chinese cross-modal pre-training benchmark. Advances in Neural Information Processing Systems 35, 26418-26431 (2022)" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 138, + 444, + 480, + 475 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 444, + 480, + 475 + ], + "spans": [ + { + "bbox": [ + 138, + 444, + 480, + 475 + ], + "type": "text", + "content": "0. Heusel, M., Ramsauer, H., Unterthiner, T., Nessler, B., Hochreiter, S.: Gans trained by a two time-scale update rule converge to a local nash equilibrium. Advances in neural information processing systems 30 (2017)" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 138, + 475, + 480, + 496 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 475, + 480, + 496 + ], + "spans": [ + { + "bbox": [ + 138, + 475, + 480, + 496 + ], + "type": "text", + "content": "1. Ho, J., Jain, A., Abbeel, P.: Denoising diffusion probabilistic models. Advances in Neural Information Processing Systems 33, 6840-6851 (2020)" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 138, + 497, + 480, + 517 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 497, + 480, + 517 + ], + "spans": [ + { + "bbox": [ + 138, + 497, + 480, + 517 + ], + "type": "text", + "content": "2. Ho, J., Jain, A., Abbeel, P.: Denoising diffusion probabilistic models. Advances in Neural Information Processing Systems 33, 6840-6851 (2020)" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 138, + 517, + 480, + 548 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 517, + 480, + 548 + ], + "spans": [ + { + "bbox": [ + 138, + 517, + 480, + 548 + ], + "type": "text", + "content": "3. Ho, J., Sahara, C., Chan, W., Fleet, D.J., Norouzi, M., Salimans, T.: Cascaded diffusion models for high fidelity image generation. J. Mach. Learn. Res. 23, 47-1 (2022)" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 138, + 548, + 480, + 570 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 548, + 480, + 570 + ], + "spans": [ + { + "bbox": [ + 138, + 548, + 480, + 570 + ], + "type": "text", + "content": "4. Ho, J., Salimans, T.: Classifier-free diffusion guidance. arXiv preprint arXiv:2207.12598 (2022)" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 138, + 571, + 480, + 590 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 571, + 480, + 590 + ], + "spans": [ + { + "bbox": [ + 138, + 571, + 480, + 590 + ], + "type": "text", + "content": "5. Jin, Z., Shen, X., Li, B., Xue, X.: Training-free diffusion model adaptation for variable-sized text-to-image synthesis. arXiv preprint arXiv:2306.08645 (2023)" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 138, + 590, + 480, + 623 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 590, + 480, + 623 + ], + "spans": [ + { + "bbox": [ + 138, + 590, + 480, + 623 + ], + "type": "text", + "content": "6. Li, X., Xu, C., Wang, X., Lan, W., Jia, Z., Yang, G., Xu, J.: Coco-cn for cross-lingual image tagging, captioning, and retrieval. IEEE Transactions on Multimedia 21(9), 2347-2360 (2019)" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 138, + 623, + 480, + 665 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 623, + 480, + 665 + ], + "spans": [ + { + "bbox": [ + 138, + 623, + 480, + 665 + ], + "type": "text", + "content": "7. Lin, T.Y., Maire, M., Belongie, S., Hays, J., Perona, P., Ramanan, D., Dólar, P., Zitnick, C.L.: Microsoft coco: Common objects in context. In: Computer Vision-ECCV 2014: 13th European Conference, Zurich, Switzerland, September 6-12, 2014, Proceedings, Part V 13. pp. 740-755. Springer (2014)" + } + ] + } + ], + "index": 19 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 392, + 91, + 447, + 100 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 392, + 91, + 447, + 100 + ], + "spans": [ + { + "bbox": [ + 392, + 91, + 447, + 100 + ], + "type": "text", + "content": "PanGu-Draw" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 470, + 92, + 480, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 470, + 92, + 480, + 100 + ], + "spans": [ + { + "bbox": [ + 470, + 92, + 480, + 100 + ], + "type": "text", + "content": "15" + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 14 + }, + { + "para_blocks": [ + { + "bbox": [ + 132, + 116, + 480, + 665 + ], + "type": "list", + "angle": 0, + "index": 19, + "blocks": [ + { + "bbox": [ + 133, + 116, + 480, + 138 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 116, + 480, + 138 + ], + "spans": [ + { + "bbox": [ + 133, + 116, + 480, + 138 + ], + "type": "text", + "content": "18. Liu, H., Li, C., Wu, Q., Lee, Y.J.: Visual instruction tuning. arXiv preprint arXiv:2304.08485 (2023)" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 132, + 139, + 480, + 170 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 139, + 480, + 170 + ], + "spans": [ + { + "bbox": [ + 132, + 139, + 480, + 170 + ], + "type": "text", + "content": "19. Liu, N., Li, S., Du, Y., Torralba, A., Tenenbaum, J.B.: Compositional visual generation with composable diffusion models. In: European Conference on Computer Vision. pp. 423-439. Springer (2022)" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 132, + 171, + 480, + 203 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 171, + 480, + 203 + ], + "spans": [ + { + "bbox": [ + 132, + 171, + 480, + 203 + ], + "type": "text", + "content": "20. Nichol, A., Dhariwal, P., Ramesh, A., Shyam, P., Mishkin, P., McGrew, B., Sutskever, I., Chen, M.: Glide: Towards photorealistic image generation and editing with text-guided diffusion models. arXiv preprint arXiv:2112.10741 (2021)" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 132, + 204, + 480, + 224 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 204, + 480, + 224 + ], + "spans": [ + { + "bbox": [ + 132, + 204, + 480, + 224 + ], + "type": "text", + "content": "21. Nichol, A.Q., Dhariwal, P.: Improved denoising diffusion probabilistic models. In: ICML. pp. 8162-8171. PMLR (2021)" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 132, + 225, + 480, + 257 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 225, + 480, + 257 + ], + "spans": [ + { + "bbox": [ + 132, + 225, + 480, + 257 + ], + "type": "text", + "content": "22. Podell, D., English, Z., Lacey, K., Blattmann, A., Dockhorn, T., Müller, J., Penna, J., Rombach, R.: Sdxl: improving latent diffusion models for high-resolution image synthesis. arXiv preprint arXiv:2307.01952 (2023)" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 132, + 258, + 480, + 279 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 258, + 480, + 279 + ], + "spans": [ + { + "bbox": [ + 132, + 258, + 480, + 279 + ], + "type": "text", + "content": "23. Qin, Y., Zhang, J., Lin, Y., Liu, Z., Li, P., Sun, M., Zhou, J.: Elle: Efficient lifelong pre-training for emerging data. arXiv preprint arXiv:2203.06311 (2022)" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 132, + 280, + 480, + 311 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 280, + 480, + 311 + ], + "spans": [ + { + "bbox": [ + 132, + 280, + 480, + 311 + ], + "type": "text", + "content": "24. Ramesh, A., Dhariwal, P., Nichol, A., Chu, C., Chen, M.: Hierarchical text-conditional image generation with clip latents. arXiv preprint arXiv:2204.06125 (2022)" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 132, + 312, + 480, + 343 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 312, + 480, + 343 + ], + "spans": [ + { + "bbox": [ + 132, + 312, + 480, + 343 + ], + "type": "text", + "content": "25. Ramesh, A., Pavlov, M., Goh, G., Gray, S., Voss, C., Radford, A., Chen, M., Sutskever, I.: Zero-shot text-to-image generation. In: International Conference on Machine Learning. pp. 8821-8831. PMLR (2021)" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 132, + 344, + 480, + 386 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 344, + 480, + 386 + ], + "spans": [ + { + "bbox": [ + 132, + 344, + 480, + 386 + ], + "type": "text", + "content": "26. Rombach, R., Blattmann, A., Lorenz, D., Esser, P., Ommer, B.: High-resolution image synthesis with latent diffusion models. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR). pp. 10684-10695 (June 2022)" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 132, + 387, + 480, + 418 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 387, + 480, + 418 + ], + "spans": [ + { + "bbox": [ + 132, + 387, + 480, + 418 + ], + "type": "text", + "content": "27. Rombach, R., Blattmann, A., Lorenz, D., Esser, P., Ommer, B.: High-resolution image synthesis with latent diffusion models. In: Proceedings of the IEEE/CVF conference on computer vision and pattern recognition. pp. 10684-10695 (2022)" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 132, + 419, + 480, + 461 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 419, + 480, + 461 + ], + "spans": [ + { + "bbox": [ + 132, + 419, + 480, + 461 + ], + "type": "text", + "content": "28. Ronneberger, O., Fischer, P., Brox, T.: U-net: Convolutional networks for biomedical image segmentation. In: Medical Image Computing and Computer-Assisted Intervention-MICCAI 2015: 18th International Conference, Munich, Germany, October 5-9, 2015, Proceedings, Part III 18. pp. 234-241. Springer (2015)" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 132, + 462, + 480, + 503 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 462, + 480, + 503 + ], + "spans": [ + { + "bbox": [ + 132, + 462, + 480, + 503 + ], + "type": "text", + "content": "29. Sahara, C., Chan, W., Saxena, S., Li, L., Whang, J., Denton, E.L., Ghasemipour, K., Gontijo Lopes, R., Karagol Ayan, B., Salimans, T., et al.: Photorealistic text-to-image diffusion models with deep language understanding. Advances in Neural Information Processing Systems 35, 36479-36494 (2022)" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 132, + 504, + 480, + 536 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 504, + 480, + 536 + ], + "spans": [ + { + "bbox": [ + 132, + 504, + 480, + 536 + ], + "type": "text", + "content": "30. Salimans, T., Goodfellow, I., Zaremba, W., Cheung, V., Radford, A., Chen, X.: Improved techniques for training gans. Advances in neural information processing systems 29, 2234-2242 (2016)" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 132, + 536, + 480, + 578 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 536, + 480, + 578 + ], + "spans": [ + { + "bbox": [ + 132, + 536, + 480, + 578 + ], + "type": "text", + "content": "31. Shonenkov, A., Konstantinov, M., Bakshandaeva, D., Schuhmann, C., Ivanova, K., Klokova, N.: Deepfloyd if: A powerful text-to-image model that can smartly integrate text into images (2023), https://www deepfloyd.ai/deepfloyd-if, online; accessed 16-November-2023" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 132, + 579, + 480, + 601 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 579, + 480, + 601 + ], + "spans": [ + { + "bbox": [ + 132, + 579, + 480, + 601 + ], + "type": "text", + "content": "32. Xu, X., Wang, Z., Zhang, E., Wang, K., Shi, H.: Versatile diffusion: Text, images and variations all in one diffusion model. arXiv preprint arXiv:2211.08332 (2022)" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 132, + 601, + 480, + 632 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 601, + 480, + 632 + ], + "spans": [ + { + "bbox": [ + 132, + 601, + 480, + 632 + ], + "type": "text", + "content": "33. Xue, Z., Song, G., Guo, Q., Liu, B., Zong, Z., Liu, Y., Luo, P.: Raphael: Text-to-image generation via large mixture of diffusion paths. arXiv preprint arXiv:2305.18295 (2023)" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 132, + 633, + 480, + 665 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 633, + 480, + 665 + ], + "spans": [ + { + "bbox": [ + 132, + 633, + 480, + 665 + ], + "type": "text", + "content": "34. Yang, A., Pan, J., Lin, J., Men, R., Zhang, Y., Zhou, J., Zhou, C.: Chinese clip: Contrastive vision-language pretraining in chinese. arXiv preprint arXiv:2211.01335 (2022)" + } + ] + } + ], + "index": 18 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "spans": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "type": "text", + "content": "16" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 166, + 91, + 215, + 100 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 166, + 91, + 215, + 100 + ], + "spans": [ + { + "bbox": [ + 166, + 91, + 215, + 100 + ], + "type": "text", + "content": "G. Lu et al." + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 15 + }, + { + "para_blocks": [ + { + "bbox": [ + 130, + 116, + 482, + 297 + ], + "type": "list", + "angle": 0, + "index": 7, + "blocks": [ + { + "bbox": [ + 130, + 116, + 482, + 159 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 116, + 482, + 159 + ], + "spans": [ + { + "bbox": [ + 130, + 116, + 482, + 159 + ], + "type": "text", + "content": "35. Yao, L., Han, J., Liang, X., Xu, D., Zhang, W., Li, Z., Xu, H.: Detclipv2: Scalable open-vocabulary object detection pre-training via word-region alignment. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 23497-23506 (2023)" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 130, + 160, + 481, + 191 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 160, + 481, + 191 + ], + "spans": [ + { + "bbox": [ + 130, + 160, + 481, + 191 + ], + "type": "text", + "content": "36. Yao, L., Huang, R., Hou, L., Lu, G., Niu, M., Xu, H., Liang, X., Li, Z., Jiang, X., Xu, C.: Filip: Fine-grained interactive language-image pre-training. arXiv preprint arXiv:2111.07783 (2021)" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 131, + 191, + 481, + 212 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 131, + 191, + 481, + 212 + ], + "spans": [ + { + "bbox": [ + 131, + 191, + 481, + 212 + ], + "type": "text", + "content": "37. Ye, F., Liu, G., Wu, X., Wu, L.: Altdiffusion: A multilingual text-to-image diffusion model (2023)" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 131, + 212, + 481, + 265 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 131, + 212, + 481, + 265 + ], + "spans": [ + { + "bbox": [ + 131, + 212, + 481, + 265 + ], + "type": "text", + "content": "38. Zhang, J., Gan, R., Wang, J., Zhang, Y., Zhang, L., Yang, P., Gao, X., Wu, Z., Dong, X., He, J., Zhuo, J., Yang, Q., Huang, Y., Li, X., Wu, Y., Lu, J., Zhu, X., Chen, W., Han, T., Pan, K., Wang, R., Wang, H., Wu, X., Zeng, Z., Chen, C.: Fengshenbang 1.0: Being the foundation of chinese cognitive intelligence. CoRR abs/2209.02970 (2022)" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 131, + 265, + 481, + 297 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 131, + 265, + 481, + 297 + ], + "spans": [ + { + "bbox": [ + 131, + 265, + 481, + 297 + ], + "type": "text", + "content": "39. Zhang, L., Rao, A., Agrawala, M.: Adding conditional control to text-to-image diffusion models. In: Proceedings of the IEEE/CVF International Conference on Computer Vision. pp. 3836-3847 (2023)" + } + ] + } + ], + "index": 6 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 392, + 91, + 447, + 100 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 392, + 91, + 447, + 100 + ], + "spans": [ + { + "bbox": [ + 392, + 91, + 447, + 100 + ], + "type": "text", + "content": "PanGu-Draw" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 470, + 91, + 481, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 470, + 91, + 481, + 100 + ], + "spans": [ + { + "bbox": [ + 470, + 91, + 481, + 100 + ], + "type": "text", + "content": "17" + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 16 + } + ], + "_backend": "vlm", + "_version_name": "2.6.4" +} \ No newline at end of file diff --git a/2024/Panel-Specific Degradation Representation for Raw Under-Display Camera Image Restoration/c089de88-5e00-46c3-9141-1e16159753ca_content_list.json b/2024/Panel-Specific Degradation Representation for Raw Under-Display Camera Image Restoration/c089de88-5e00-46c3-9141-1e16159753ca_content_list.json new file mode 100644 index 0000000000000000000000000000000000000000..1b983ff1e74511ec3ffea094e5ba1beb34d69041 --- /dev/null +++ b/2024/Panel-Specific Degradation Representation for Raw Under-Display Camera Image Restoration/c089de88-5e00-46c3-9141-1e16159753ca_content_list.json @@ -0,0 +1,2260 @@ +[ + { + "type": "text", + "text": "Panel-Specific Degradation Representation for Raw Under-Display Camera Image Restoration", + "text_level": 1, + "bbox": [ + 230, + 140, + 772, + 186 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Youngjin Oh $^{1}$ , Keuntek Lee $^{1}$ , Jooyoung Lee $^{3}$ , Dae-Hyun Lee $^{3}$ , and Nam Ik Cho $^{1,2}$", + "bbox": [ + 230, + 210, + 772, + 242 + ], + "page_idx": 0 + }, + { + "type": "list", + "sub_type": "text", + "list_items": [ + "$^{1}$ Department of ECE & INMC, Seoul National University, Seoul, Korea {yjymoh0211,leekt000,nicho}@snu.ac.kr", + "$^{2}$ IPAI, Seoul National University, Seoul, Korea", + "3 SK hynix Inc., Korea {jooyoung2.lee, daehyun1.lee}@sk.com" + ], + "bbox": [ + 256, + 253, + 741, + 324 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Abstract. Under-display camera (UDC) image restoration aims to restore images distorted by the OLED display panel covering the frontal camera on a smartphone. Previous deep learning-based UDC restoration methods focused on restoring the image within the RGB domain with the collection of real or synthetic RGB datasets. However, UDC images in these datasets exhibit domain differences from real commercial smartphone UDC images while inherently constraining the problem and solution within the RGB domain. To address this issue, we collect well-aligned sensor-level real UDC images using panels from two commercial smartphones equipped with UDC. We also propose a new UDC restoration method to exploit the disparities between degradations caused by different panels, considering that UDC degradations are specific to the type of OLED panel. For this purpose, we train an encoder with an unsupervised learning scheme using triplet loss that aims to extract the inherent degradations caused by different panels from degraded UDC images as implicit representations. The learned panel-specific degradation representations are then provided as priors to our restoration network based on an efficient Transformer network. Extensive experiments show that our proposed method achieves state-of-the-art performance on our real raw image dataset and generalizes well to previous datasets. Our dataset and code is available at https://github.com/OBAKSA/DREUDC.", + "bbox": [ + 261, + 359, + 738, + 651 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Keywords: Under-Display Camera $\\cdot$ Image Restoration $\\cdot$ Representation Learning", + "bbox": [ + 259, + 664, + 740, + 693 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "1 Introduction", + "text_level": 1, + "bbox": [ + 215, + 718, + 375, + 734 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "To hide the camera hole in the frontal side of smartphones, under-display camera (UDC) systems have been developed, enabled by recent advancements in imaging and display panel technologies. With the UDC technology, smartphone users can enjoy a full-screen display without any camera holes or notches on the front side of their device. However, since the camera is hidden under the panel, the light that reaches the camera is altered by the complex layers of the panel, causing", + "bbox": [ + 212, + 750, + 787, + 840 + ], + "page_idx": 0 + }, + { + "type": "table", + "img_path": "images/55f3bc253a196e3460cf0c79f466c984027730b3af48f2349a4d18eed6efe7d9.jpg", + "table_caption": [ + "Table 1: Comparison of previous datasets [9,10,30,38] and our commercial dataset." + ], + "table_footnote": [], + "table_body": "
DatasetZhou et al. [38]Feng et al. [10]Feng et al. [9]Song et al. [30]Ours
FormatRGBRGBRGBRGBRGB/RAW
UDC ModelingReal (prototype)SyntheticRealSyntheticReal
Aligned GT✓✓✗✓✓
Multiple Display✓✗✗✗✓
", + "bbox": [ + 228, + 170, + 767, + 236 + ], + "page_idx": 1 + }, + { + "type": "image", + "img_path": "images/d948b0a515a4d41a88c7fbc0d82e2f66ae1c6d20a86b66a90d3aa0bcf36e02d1.jpg", + "image_caption": [ + "(a) Feng et al. [9], GT" + ], + "image_footnote": [], + "bbox": [ + 318, + 253, + 392, + 311 + ], + "page_idx": 1 + }, + { + "type": "image", + "img_path": "images/6ffebcb9c500395872be2223822ed775976f904faa0de2179265e107609aa0cf.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 395, + 256, + 455, + 309 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "l. [9], GT", + "bbox": [ + 398, + 311, + 460, + 325 + ], + "page_idx": 1 + }, + { + "type": "image", + "img_path": "images/c44ea9092d19ab626987bfc963fe0afc1d6cea238d5dd4ba61c644430f212330.jpg", + "image_caption": [ + "(b) Feng et al. [9], UDC" + ], + "image_footnote": [], + "bbox": [ + 526, + 253, + 598, + 311 + ], + "page_idx": 1 + }, + { + "type": "image", + "img_path": "images/a07cd45ac42d6e56651124ea01d51ba65f9194188f40d955e0344d4b1a7628b4.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 602, + 256, + 663, + 309 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "al. [9], UDC", + "bbox": [ + 599, + 311, + 676, + 325 + ], + "page_idx": 1 + }, + { + "type": "image", + "img_path": "images/f6255a2f4b4f8f70c9e59f1bedbc4a12154fac0f34dba31677604939441d38e0.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 241, + 325, + 316, + 381 + ], + "page_idx": 1 + }, + { + "type": "image", + "img_path": "images/3f4084fbd01901d35c7e56711e655a381b04282122bdd002b908db7c1fcf519c.jpg", + "image_caption": [ + "(d) Feng et al. [10], GT" + ], + "image_footnote": [], + "bbox": [ + 318, + 327, + 379, + 382 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "l. [10], GT", + "bbox": [ + 318, + 383, + 387, + 396 + ], + "page_idx": 1 + }, + { + "type": "image", + "img_path": "images/852a895706697611d4b76c13c5b46fb97831956a03d1e9d819fccdffb10d5c7c.jpg", + "image_caption": [ + "(e) Feng et al. [10], UDC" + ], + "image_footnote": [], + "bbox": [ + 421, + 327, + 495, + 383 + ], + "page_idx": 1 + }, + { + "type": "image", + "img_path": "images/d0f96c5e388cf4d5cb82d4beea9c506c816e7c1ad43d65e6ee4aa8c7b31b07f6.jpg", + "image_caption": [ + "(f) Song et al. [30], UDC" + ], + "image_footnote": [], + "bbox": [ + 602, + 327, + 676, + 383 + ], + "page_idx": 1 + }, + { + "type": "image", + "img_path": "images/656f2a6209f476ac2a4613a836bbb4c45d6e89bf1c78caf87f96dd154cb4d8a9.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 241, + 398, + 316, + 453 + ], + "page_idx": 1 + }, + { + "type": "image", + "img_path": "images/aacadb8f0484e70271a5c0dc591bda948a147c3661a8a22a0376a7491093fac0.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 318, + 398, + 379, + 452 + ], + "page_idx": 1 + }, + { + "type": "image", + "img_path": "images/b78df607efece1427d9a1c6fb5dab173aa63b7e57b631e540e202f62b710ed46.jpg", + "image_caption": [ + "(h) Zho", + "Zhou et al. [38], T-OLED" + ], + "image_footnote": [], + "bbox": [ + 421, + 398, + 495, + 453 + ], + "page_idx": 1 + }, + { + "type": "image", + "img_path": "images/eca13371b9bbecceb84ed27e6ba1cee8f122782f1df832359620a0647d8f5708.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 602, + 398, + 676, + 453 + ], + "page_idx": 1 + }, + { + "type": "image", + "img_path": "images/9e0966a407f913e42c0d6c843a91c52368e6507478027bf8554a990c5da152de.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 679, + 398, + 741, + 452 + ], + "page_idx": 1 + }, + { + "type": "image", + "img_path": "images/175e3e054a157fa6399689707e170996e0f102d7afa2ae76bd82a96b99bc5865.jpg", + "image_caption": [ + "(g) Zhou et al. [38],", + "(j) Ours, GT", + "Fig. 1: Samples and their RGB channel histograms from previous UDC datasets [9, 10,30,38] and our commercial dataset." + ], + "image_footnote": [], + "bbox": [ + 241, + 470, + 316, + 525 + ], + "page_idx": 1 + }, + { + "type": "image", + "img_path": "images/33eb15ba359e820dc42a731fafa38907df0578b97b4b51c168ed0a9761c3298c.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 318, + 470, + 379, + 525 + ], + "page_idx": 1 + }, + { + "type": "image", + "img_path": "images/b79214c6f053dea6614bebde267b73bcbdf0893279ac48494551b4ccadfd0bbc.jpg", + "image_caption": [ + "(k) Ours, Axon 30" + ], + "image_footnote": [], + "bbox": [ + 421, + 470, + 493, + 525 + ], + "page_idx": 1 + }, + { + "type": "image", + "img_path": "images/5cb59ca95b7eb7347ba8642a6a8fcb253962c30ccffa8eb2c6203640970b2d43.jpg", + "image_caption": [ + ") Zhou et al", + "[38], P-OLED", + "(1) Ours, Z-Fold 3" + ], + "image_footnote": [], + "bbox": [ + 602, + 469, + 676, + 525 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "severe degradation in the resulting image. Degradations include blur, diffraction, noise, low-lightness, color shift, haziness, etc., and their spatially-variant nature makes the UDC image restoration very challenging [10, 17, 38].", + "bbox": [ + 212, + 613, + 784, + 657 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Most UDC image restoration methods used deep neural networks [9-11, 15, 17, 23, 24, 30, 31, 37, 38], providing promising results on public benchmark datasets [9, 10, 30, 38] gathered to facilitate research on UDC image restoration. However, these existing datasets either use prototype display panels to inadequately represent the degradations observed in real UDC images caused by commercial panels [38], lack realistic degradations due to incomplete UDC modeling (e.g., noise and color shift) in the synthesis pipeline [10,30], or mishandles aligning image pairs with occlusion and parallax due to the mechanism of the capturing imaging system [9]. In addition, the potential for enhanced image restoration performance through sensor-level restoration [1, 19, 25, 35] cannot be exploited, because previous datasets predominantly restrict experiments to RGB domain as raw data is unreleased. Tab. 1 shows comparison of the datasets.", + "bbox": [ + 212, + 659, + 787, + 840 + ], + "page_idx": 1 + }, + { + "type": "page_number", + "text": "2", + "bbox": [ + 217, + 114, + 228, + 126 + ], + "page_idx": 1 + }, + { + "type": "header", + "text": "Y. Oh et al.", + "bbox": [ + 271, + 114, + 354, + 127 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "In this work, we propose a new UDC dataset that is collected in the raw domain. The proposed dataset contains pairs of real UDC images and clean images that are well-aligned using a monitor-based imaging system with two different panels from commercial UDC smartphones, ZTE Axon 30 5G and Samsung Galaxy Z-Fold 3. In our experiments, we find that training with raw data enhances the performance of UDC image restoration. This raw-domain approach with real data also has the advantage of enabling the use of end-to-end neural ISPs that are being installed in recent smartphones [26], allowing a practical implementation on devices. Additionally, as shown in Fig. 1, real UDC data captured with commercial smartphone panels display noticeable distinct degradations visually and statistically when compared to previous datasets [10,30,38].", + "bbox": [ + 212, + 146, + 782, + 311 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Furthermore, we develop a two-stage restoration method for UDC images. First, we introduce a framework for learning the implicit representation of degradations in the UDC images. We notice that degradation in a patch of a UDC image is similar to that of other patches of the same UDC image while being distinct from degradations of the same scene taken with a different panel. With this observation, we develop a novel unsupervised learning scheme involving triplet loss [28] to train our encoder. The learning scheme aims to leverage the information of the data collected with different types of panels, ultimately condensing the degradation information into an implicit representation vector that reflects the panel-specific degradation moderately. We also propose an efficient Transformer-based restoration network that incorporates the learned panel-specific degradation representations as a prior by generating channel-wise calibration coefficients for attention. We discover that our strategy yields improved results for UDC image restoration. To the best of our knowledge, this is the first case in UDC image restoration where a dataset collected with one panel is utilized to enhance the performance on a dataset collected with another panel.", + "bbox": [ + 212, + 313, + 784, + 553 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "To summarize, our contributions are as follows:", + "bbox": [ + 238, + 554, + 581, + 568 + ], + "page_idx": 2 + }, + { + "type": "list", + "sub_type": "text", + "list_items": [ + "- We introduce a new real UDC dataset, which is collected with commercial smartphone panels. Our dataset is provided in raw format, which can improve the performance of the UDC restoration task.", + "- We propose an unsupervised learning approach that allows for implicit representations of degradations caused by various display panels in the embedding space. The panel-specific degradation representations are leveraged as priors in the restoration network, which is an efficient Transformer and brings notable improvement.", + "- Extensive experiments demonstrate that our restoration network embedded with learned representations achieves state-of-the-art performance on the proposed dataset and is consistently applicable to previous datasets." + ], + "bbox": [ + 225, + 574, + 784, + 737 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "2 Related Works", + "text_level": 1, + "bbox": [ + 215, + 758, + 395, + 773 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "2.1 UDC Image Restoration", + "text_level": 1, + "bbox": [ + 215, + 787, + 464, + 801 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Since UDC image restoration is a relatively new area, research is still ongoing to provide users with a satisfying experience when UDC is applied in real situa-", + "bbox": [ + 212, + 809, + 782, + 839 + ], + "page_idx": 2 + }, + { + "type": "header", + "text": "Panel-Specific UDC Image Restoration", + "bbox": [ + 470, + 114, + 730, + 128 + ], + "page_idx": 2 + }, + { + "type": "page_number", + "text": "3", + "bbox": [ + 774, + 116, + 784, + 126 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "tions. Research on UDC image restoration was sparked by the pioneering work of [38]. They defined a UDC imaging system and released the first UDC dataset for researchers to experiment with. Synthetic datasets were also made available by simulating the camera pipeline using a pre-calculated point spread function (PSF), which adequately represents the diffraction [10], with further improvements on scattering effects in [30] to account for the haziness of UDC images. More recently, a real UDC dataset generated by pseudo-aligning geometrically misaligned image pairs for a single UDC panel [9] was presented. Although the proposed datasets have some drawbacks, several works [9-11, 15, 17, 23, 24, 30, 31, 37, 38] have investigated UDC image restoration using deep neural networks and have shown promising results.", + "bbox": [ + 212, + 146, + 787, + 311 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Among these deep-learning-based methods, several of them use priors to improve performance. For instance, in [10], PCA is employed on a PSF of the UDC display panel to stretch it as a 'kernel code' and condition the network with it as a prior. In [30], the network is divided into a scattering branch and an image branch, and the former is trained to estimate the scattering parameters, which is prior information that can be used. However, these methods require precise priors; otherwise, the performance is degraded. Our approach differs from previous methods because we do not need strict parameters or ground truth settings. Instead, we aim to learn an acceptable representation of the degradation induced by panel-specific properties in the embedding space rather than a rigorous estimation of the degradation. We only need images taken with different panels because our method is unsupervised, and learn the implicit priors by leveraging the discrepancies of degradation caused by different panels.", + "bbox": [ + 212, + 313, + 787, + 508 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "2.2 Image Restoration with Raw Data", + "text_level": 1, + "bbox": [ + 214, + 529, + 545, + 544 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Most of the previous research on image restoration relies on RGB images as training data as they are easier to collect. However, some recent studies have shown that using raw sensor data can lead to improved performance. Specifically, [1, 25] are the works on image denoising where they collected DND and SIDD datasets that are available as raw and RGB images, where denoising with raw images usually yields higher PSNR. CycleISP [35] demonstrated that noise is more complex in RGB images and showed that noise modeling is more realistic in the sensor domain. This is because compared to the noise samples in the sensor pixels, which are nearly uncorrelated to each other, the noise samples become highly correlated in RGB images as they pass through the camera ISP pipeline, which involves processes like demosaicing, gamma correction, and white balancing [18, 22]. DeepISP [29] proposed an end-to-end camera ISP modeled with deep neural networks, suggesting that a deep network-based camera ISP is promising. Furthermore, [4] solved the issue of low-light enhancement problem by operating in the sensor domain, and [19] presented a dataset for raw image deblurring and proved that operating on raw sensor data achieves better performance than RGB-based methods. These researches suggest that it is easier to restore raw data than RGB images with deep neural networks, as sensor data contains the linear representation of incoming light intensity of the scene. This ensures that", + "bbox": [ + 212, + 553, + 787, + 840 + ], + "page_idx": 3 + }, + { + "type": "page_number", + "text": "4", + "bbox": [ + 217, + 114, + 228, + 126 + ], + "page_idx": 3 + }, + { + "type": "header", + "text": "Y. Oh et al.", + "bbox": [ + 271, + 114, + 354, + 127 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "the scene's authenticity is maintained more faithfully in raw data than in heavily processed RGB pixel values that are nonlinear with respect to light intensity.", + "bbox": [ + 212, + 146, + 782, + 176 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "3 Proposed Method", + "text_level": 1, + "bbox": [ + 214, + 199, + 426, + 217 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "3.1 Sensor-level Real UDC Dataset", + "text_level": 1, + "bbox": [ + 214, + 229, + 521, + 244 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "A raw dataset is necessary to achieve end-to-end image restoration using sensor images. However, the available UDC datasets are limited to the RGB domain. Moreover, public UDC datasets do not show the types of degradation that are present in images taken with commercially available smartphones equipped with UDC, limiting their assessment of practical scenarios. Therefore, we capture real raw UDC data by using OLED panels obtained from commercial UDC smartphones, a smartphone camera, and a monitor-based imaging system [38].", + "bbox": [ + 212, + 253, + 787, + 359 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "UDC data acquisition process Capturing a raw UDC image requires that we model the UDC imaging system as similarly as possible. Consequently, our data-capturing process involves rendering an image onto a high-resolution monitor, followed by capturing the displayed image using a smartphone camera sensor located closely behind a UDC panel. The panels used to capture the data are from two commercial UDC smartphones: ZTE Axon 30 5G and Samsung Galaxy Z-Fold 3. They introduce low-light/color shift, scattering, diffraction, and blur. As noise is usually generated from the sensor, we chose Samsung Galaxy S23 Ultra as our camera device to capture the authentic noise of smartphone camera sensors. Raw images are collected using the internal software of the smartphone that automatically saves the minimally processed raw data. The preference for monitor-displayed images [17,38] over real natural scenes is driven by the need to control additional factors beyond the panels' presence. Specifically, we can enforce the target scene to be static as dynamic scenes are prone to motion blur, and therefore lead to misaligned data pairs. We can also eliminate interference from external ambient sources of light other than the monitor's emitted light.", + "bbox": [ + 212, + 380, + 787, + 621 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "The monitor, UDC panel, and camera sensor are aligned in sequence, and the principal axis of the camera is aligned perpendicular to the plane of the monitor and the panel. The camera is mounted on a sturdy tripod, and the panel is situated close to the sensor using a firm holding device. The displayed images are then captured remotely with/without the smartphone panels to ensure the alignment of clean/UDC image pairs during capture while keeping the camera settings the same. We provide more details in the Supplementary Material.", + "bbox": [ + 212, + 622, + 787, + 729 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Our UDC dataset The dataset is captured using 400 images from the DIV2K dataset [2], which is comprised of natural scene images from various themes, encompassing diverse objects captured under different lighting conditions. This makes our dataset 1,200 images in total, as there are three different image sets (Axon 30, Z-Fold 3, ground truth) in raw format. The images are stored in 16 bits with a Bayer pattern of G-B-R-G. For the ground truth images, we capture", + "bbox": [ + 212, + 750, + 787, + 840 + ], + "page_idx": 4 + }, + { + "type": "header", + "text": "Panel-Specific UDC Image Restoration", + "bbox": [ + 470, + 114, + 730, + 128 + ], + "page_idx": 4 + }, + { + "type": "page_number", + "text": "5", + "bbox": [ + 774, + 116, + 784, + 126 + ], + "page_idx": 4 + }, + { + "type": "image", + "img_path": "images/8d7838f33a0fc9b19804d0939e843031ad78f3fb51bac4fbeaf68210bf52f226.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 282, + 148, + 413, + 251 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/86ac0c9bc19cf47552625d9c3439435a1bffb1f18c2061258c88989b772b2d54.jpg", + "image_caption": [ + "(a) Clean image", + "Fig. 2: Cropped examples from our captured dataset. All images are visualized with the same procedure that includes normalization, demosaicing, channel gain adjustment, and gamma correction. Due to the difference in smartphone display panels, noise and haze are more apparent in Axon 30 UDC images, while the Z-Fold 3 UDC images show blurrier results. The diffraction induces flare artifacts when the image intensity becomes too strong that the sensor saturates." + ], + "image_footnote": [], + "bbox": [ + 282, + 252, + 413, + 354 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/f4e1113a491cc030422183607632a0f7bac0a31aaa6b9e9f1806dd50a70b365f.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 434, + 148, + 568, + 251 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/17fdc077e18fb78a35e80a41fead853513427d56d89a7d63ed5091f8ef20e8fe.jpg", + "image_caption": [ + "(b) Axon 30 image" + ], + "image_footnote": [], + "bbox": [ + 434, + 252, + 566, + 354 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/ed02638ba7fa4724740b8648c53514befc953003693debced1e1d41f68a3fc53.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 588, + 148, + 720, + 251 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/dedcff742a8060f4fa0c23ca613882a2884f8fc74f73d5220d5e832fa096a21c.jpg", + "image_caption": [ + "(c) Z-Fold 3 image" + ], + "image_footnote": [], + "bbox": [ + 588, + 252, + 720, + 354 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "bursts of a scene without the display panel and average them. To check for misalignments that can be induced by optical image stabilization in the smartphone hardware [1], we utilize registration and feature matching algorithms [3, 12].", + "bbox": [ + 212, + 493, + 784, + 539 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Samples from our dataset are displayed in Fig. 2, with each UDC image taken with Axon 30 and Z-Fold 3 display panels exhibiting panel-specific degradations. As such, our dataset can be used to evaluate if a restoration method generally performs well in the restoration of various UDC systems. Also, our real commercial UDC data show different behaviors from previous datasets, as shown in Fig. 1. Specifically, low-light, color shift, blur, and noise are observed, and an object is seen repeated faintly near the main object due to diffraction. We also detect a phenomenon similar to haze due to the scattering effects of the panel [30]. Synthetic datasets that are generated using UDC modeling [10, 30] do not take noise into account. However, including noise in UDC images is vital as the level of noise captured with frontal smartphone camera sensors, which are relatively inferior to DSLR camera sensors [4, 26], is amplified when normalizing sensor data. This amplification is related to the attenuation caused by the panel which reduces light intensity. More samples are available in the Supplementary.", + "bbox": [ + 212, + 539, + 787, + 750 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "3.2 Proposed UDC Restoration Framework", + "text_level": 1, + "bbox": [ + 214, + 771, + 586, + 785 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "In this section, we propose our restoration framework called Degradation Representation Embedded Transformer for UDC image restoration (DREUDC), which consists of a degradation representation encoder and a restoration network. We", + "bbox": [ + 212, + 794, + 785, + 839 + ], + "page_idx": 5 + }, + { + "type": "page_number", + "text": "6", + "bbox": [ + 217, + 114, + 228, + 126 + ], + "page_idx": 5 + }, + { + "type": "header", + "text": "Y. Oh et al.", + "bbox": [ + 271, + 114, + 354, + 126 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "first train the encoder with our proposed unsupervised scheme using triplet loss to implicitly learn the degradations of a display panel. Then, we utilize the representations generated from the encoder by integrating them in the restoration network as a prior. It is worth noting that our approach's novelty lies in how we take advantage of UDC data from one display panel to restore a UDC image from another panel rather than in the restoration network's architecture itself.", + "bbox": [ + 212, + 146, + 787, + 238 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Motivation Representation learning has been demonstrated to be effective in blind super-resolution, as demonstrated in [32]. Inspired by this, we explore the use of representation learning to enhance the restoration of UDC images by extracting useful prior information about the panels. Since each UDC system has a different panel, it yields panel-specific degradations. For example, the UDC images presented in Fig. 2, which are captured using two different panels of UDC smartphones, show differences in the degree of degradations. This implies that the degradations present in an image captured using one panel are similar to each other and that they differ significantly from those of the image with the same content but are captured using another panel. Using this information, our approach intends to utilize these differences and train an encoder to encode them in an unsupervised manner into the embedding space.", + "bbox": [ + 212, + 258, + 789, + 441 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Unsupervised representation encoder training Our proposed unsupervised training scheme is visualized in Fig. 3. As mentioned above, the objective of training the panel-specific degradation representation encoder is to learn the unique properties of the panels and produce implicit representations of the degradations to be used as priors in UDC image restoration. To achieve this, the encoder should learn the degradations using attributes related to a specific panel, such as the level of noise, blur, and color shift. On the other hand, the encoder should be discouraged from learning degradations with features in an image that are panel-agnostic, such as textures and contents of a scene.", + "bbox": [ + 212, + 462, + 787, + 597 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "We believe that this condition is analogous to the situation of tasks where triplet loss [28] is exploited to learn favorable representations. Triplet loss is widely applied to various tasks of computer vision such as facial recognition [28], person re-identification [13,34], and domain adaptation [8] for its ability to learn useful embeddings by leveraging the distances between embeddings of similar instances and dissimilar instances. For instance, in person re-identification, networks are guided to identify among different people using features that are unique to a specific person, not with features that are shared among people, by learning suitable representations using triplet loss. In our case, the features unique to a specific person would correspond to properties unique to a specific panel, and the features that are shared among people would correspond to panel-agnostic features. Therefore, we opt to apply triplet loss in the training of our encoder to capture the panel-specific degradation representations.", + "bbox": [ + 212, + 598, + 787, + 792 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Nevertheless, triplet loss is only effective when the network is provided with valid triplets [28, 33]. To collect meaningful triplets, we impose constraints on the training pairs that will be used to train our encoder. We choose a pair of", + "bbox": [ + 212, + 795, + 787, + 840 + ], + "page_idx": 6 + }, + { + "type": "header", + "text": "Panel-Specific UDC Image Restoration", + "bbox": [ + 470, + 114, + 730, + 128 + ], + "page_idx": 6 + }, + { + "type": "page_number", + "text": "7", + "bbox": [ + 774, + 116, + 784, + 126 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/102c7da16a0a2a33a483935078135635c3dfde019631cf67c6a2d2f0a9d21ada.jpg", + "image_caption": [ + "$x_{i}$" + ], + "image_footnote": [], + "bbox": [ + 236, + 151, + 352, + 250 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/be880941c16cdc39543dd9a8e4879c36736fe48e2f0fbda52fa472477312ae56.jpg", + "image_caption": [ + "$x_{j}$ Raw UDC images (visualized for figure)" + ], + "image_footnote": [], + "bbox": [ + 236, + 263, + 351, + 359 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/04272cd6e3cf06c4a0814180c5e155924f18263db7ffbc4c7aa71ef051d4ee0f.jpg", + "image_caption": [ + "Training patches", + "Panel-specific degradation representation", + "Projected representations", + "Fig. 3: Our proposed unsupervised panel-specific degradation representation learning scheme. By using triplet loss, representations of the same panel-specific degradations with different image content are pulled closer together, while representations of different panel-specific degradations with the same image content are pushed away." + ], + "image_footnote": [], + "bbox": [ + 367, + 150, + 774, + 371 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "raw images $x_{i}$ and $x_{j}$ from our dataset that contains the same content but have been taken with different display panels. Then, we randomly crop two pairs of patches $(x_{i}^{1}, x_{i}^{2})$ and $(x_{j}^{1}, x_{j}^{2})$ from each image, where $x_{i}^{1}$ and $x_{j}^{1}$ , and $x_{i}^{2}$ and $x_{j}^{2}$ , respectively, are cropped from the same location of $x_{i}$ and $x_{j}$ . We choose patch $x_{i}^{1}$ as the anchor, and the positive and negative samples are decided to be $x_{i}^{2}$ and $x_{j}^{1}$ . The same process is repeated with $x_{j}^{1}$ as the anchor.", + "bbox": [ + 212, + 489, + 787, + 580 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "With this triplet selection, the encoder is guided to learn the intrinsic properties of the degradations caused by the display panel by pulling representation of images that show similar degradation while disregarding the features that are unrelated to degradations by pushing the representation of the same contents. The encoder $f$ is trained with triplet loss using the selected triplets to obtain the panel-specific degradation representation $y \\in \\mathbb{R}^{256}$ . The representations are further projected to a projection $z$ with two fully connected layers, following [5, 6]. The triplet loss for training the encoder is as follows:", + "bbox": [ + 212, + 580, + 787, + 700 + ], + "page_idx": 7 + }, + { + "type": "equation", + "text": "\n$$\n\\begin{array}{l} L = \\max \\left(\\left\\| z _ {i} ^ {1} - z _ {i} ^ {2} \\right\\| _ {2} ^ {2} - \\left\\| z _ {i} ^ {1} - z _ {j} ^ {1} \\right\\| _ {2} ^ {2} + \\alpha , 0\\right) \\\\ + m a x (| | z _ {j} ^ {1} - z _ {j} ^ {2} | | _ {2} ^ {2} - | | z _ {j} ^ {1} - z _ {i} ^ {1} | | _ {2} ^ {2} + \\alpha , 0), \\quad (1) \\\\ \\end{array}\n$$\n", + "text_format": "latex", + "bbox": [ + 230, + 710, + 785, + 753 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "where $\\alpha$ is the margin of the triplet loss. The margin is set as $\\alpha = 1$ . Following [32], the encoder is a simple network with six convolutional layers.", + "bbox": [ + 214, + 758, + 787, + 790 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Representation embedded restoration network For our restoration network, we design it to be as simple and efficient as possible to emphasize the ef-", + "bbox": [ + 214, + 809, + 787, + 839 + ], + "page_idx": 7 + }, + { + "type": "page_number", + "text": "8", + "bbox": [ + 217, + 114, + 228, + 126 + ], + "page_idx": 7 + }, + { + "type": "header", + "text": "Y. Oh et al.", + "bbox": [ + 271, + 114, + 354, + 126 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/baca668d5f8b5a85e243d7382e7c330b0b3f5b757a74ed434aec8ba3b75c838f.jpg", + "image_caption": [ + "(a) Our proposed framework, DREUDC", + "(b) DREblock", + "Fig. 4: Our proposed DREUDC. The framework consists of an encoder that produces an implicit representation of the panel-specific degradations of a UDC image and a restoration network that exploits the representations via the proposed DREblocks." + ], + "image_footnote": [], + "bbox": [ + 240, + 146, + 759, + 305 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "fectiveness of our contribution of how we leverage the difference in data collected with various panels to enhance the restoration performance rather than focusing on the impact of complex restoration network architectures. Therefore, we adopt a U-shaped [27] Transformer with Frequency domain-based Self-Attention Solver (FSAS) [16] as our restoration architecture. In previous works on UDC image restoration [9, 10, 17, 31], researchers have highlighted the importance of a wide receptive field. As FSAS utilizes the convolutional theorem of Fourier transform to model global contexts in the frequency domain efficiently [16], we choose it as a component of the building block for our restoration network.", + "bbox": [ + 212, + 407, + 787, + 542 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "Our restoration network is based on stacking repeated Transformer blocks, which we name DREblock (Degradation Representation Embedded Transformer block, Fig. 4(b)). The proposed DREblock efficiently integrates the degradation representation $y$ of the input UDC image as a condition of our restoration network by calibrating the features in the block in a channel-wise manner, i.e.,", + "bbox": [ + 212, + 544, + 787, + 619 + ], + "page_idx": 8 + }, + { + "type": "equation", + "text": "\n$$\nF _ {o} = F _ {a t t} \\otimes \\beta , \\tag {2}\n$$\n", + "text_format": "latex", + "bbox": [ + 446, + 633, + 785, + 651 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "where $F_{att}$ is the feature maps of the estimated attention of FSAS, $F_{o}$ is the feature after the conditioning, $\\otimes$ denotes channel-wise multiplication, and $\\beta$ is the channel conditioning coefficients that are produced by reshaping the panel-specific degradation representation $y$ using a single layer of $1 \\times 1$ convolution. The features $F_{o}$ are then processed by the point-wise feed-forward network.", + "bbox": [ + 212, + 657, + 787, + 734 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "The entire framework DREUDC is illustrated in Fig. 4. In training, the encoder is first trained with Eq. (1). Then, the encoder is frozen while the restoration network is trained with $\\mathcal{L}_1$ loss. During inference, the encoder takes the input raw UDC image and outputs a panel-specific degradation representation vector that contains prior information on UDC degradation. The representation is seamlessly embedded into the restoration network through DREblocks, finally producing a clean restored RGB image.", + "bbox": [ + 212, + 734, + 787, + 840 + ], + "page_idx": 8 + }, + { + "type": "header", + "text": "Panel-Specific UDC Image Restoration", + "bbox": [ + 470, + 114, + 730, + 128 + ], + "page_idx": 8 + }, + { + "type": "page_number", + "text": "9", + "bbox": [ + 774, + 114, + 785, + 126 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "4 Experiment and Analysis", + "text_level": 1, + "bbox": [ + 215, + 143, + 496, + 162 + ], + "page_idx": 9 + }, + { + "type": "text", + "text": "4.1 Datasets", + "text_level": 1, + "bbox": [ + 215, + 176, + 333, + 189 + ], + "page_idx": 9 + }, + { + "type": "text", + "text": "We evaluate the restoration performance using our captured data with display panels of ZTE Axon 30 5G and Samsung Galaxy Z-Fold 3. The raw images are normalized to have values within the range of [0,1]. To train and evaluate RGB-based methods, we use a simplified ISP algorithm written with MATLAB to visualize the normalized raw images to RGB images. The algorithm contains demosaicing, channel-wise gain control, and gamma correction in sequence. Note that the visualized images are not sRGB images, as actual camera ISPs in practice involve additional steps to render the image in sRGB domain. Please check the Supplementary for more discussion on the usage of simplified ISP. For both display panels, we split the dataset into 300 images for training and 100 images for testing. The training set is cropped into patches of $512 \\times 512$ . We also use previous UDC datasets [38], which consists of 300 images of both T-OLED and P-OLED scenes, to verify that our method is applicable to other real datasets.", + "bbox": [ + 215, + 200, + 787, + 398 + ], + "page_idx": 9 + }, + { + "type": "text", + "text": "4.2 Training and Implementation Details", + "text_level": 1, + "bbox": [ + 215, + 420, + 568, + 435 + ], + "page_idx": 9 + }, + { + "type": "text", + "text": "We use NVIDIA RTX 3090 GPUs in our experiments. To train the encoder, we use Adam optimizer $(\\beta_{1} = 0.9, \\beta_{2} = 0.999)$ for $1 \\times 10^{5}$ iterations with a batch size of 64, and the loss is given as Eq. (1). The restoration networks are trained with $\\mathcal{L}_1$ loss to minimize the distance between the restored image of the network and the ground truth. The initial number of channels $C$ of our restoration network is 32, and the number of blocks in each layer $L_1, L_2$ , and $L_3$ are [5,5,6]. We use Adam optimizer $(\\beta_{1} = 0.9, \\beta_{2} = 0.999)$ for $2 \\times 10^{5}$ iterations with a batch size of 2. For both networks, the initial learning rate is set as $2 \\times 10^{-4}$ and gradually decayed to $1 \\times 10^{-6}$ with cosine annealing [20]. This setting is applied to all experiments on our dataset, and every previous method reported in this paper is also reproduced by training on our dataset with the same setting for fairness.", + "bbox": [ + 215, + 444, + 787, + 612 + ], + "page_idx": 9 + }, + { + "type": "text", + "text": "4.3 Raw UDC Image Restoration", + "text_level": 1, + "bbox": [ + 215, + 633, + 506, + 648 + ], + "page_idx": 9 + }, + { + "type": "text", + "text": "First, we experiment to validate that using raw images can improve the UDC image restoration process. For this purpose, we employ a baseline neural network called DE-UNET [38] and use our captured data. We compare the performance of the network under three different settings: RAW-to-RAW, RAW-to-RGB, and RGB-to-RGB. We evaluate the performance of the network using the PSNR and SSIM metrics. To assess the performance of the RAW-to-RAW setting, we convert the output restored raw images to RGB using the same ISP process mentioned in Sec. 4.1 and measure them in the RGB domain. The results are presented in Tab. 2, showing that using raw sensor images as input (RAW-to-RAW, RAW-to-RGB) yields superior results compared to only using RGB images. Among the two settings that use raw images as input, the latter shows better performance overall, so we choose the method of converting and restoring", + "bbox": [ + 215, + 657, + 787, + 840 + ], + "page_idx": 9 + }, + { + "type": "page_number", + "text": "10", + "bbox": [ + 217, + 114, + 235, + 126 + ], + "page_idx": 9 + }, + { + "type": "header", + "text": "Y. Oh et al.", + "bbox": [ + 271, + 114, + 354, + 127 + ], + "page_idx": 9 + }, + { + "type": "table", + "img_path": "images/ae2963b391961012ee355c8771a89b6ae6f5f9c660dbdcf2aaef34b175a7db4e.jpg", + "table_caption": [ + "Table 2: Comparison of different input-to-output settings of DE-UNET [38] on our captured datasets. The best results are bold-faced." + ], + "table_footnote": [], + "table_body": "
MethodAxon 30 (raw)Z-Fold 3 (raw)Axon 30 (rgb)Z-Fold 3 (rgb)
PSNRSSIMPSNRSSIMPSNRSSIMPSNRSSIM
DE-UNET(RGB-to-RGB)----23.330.910926.900.9390
DE-UNET(RAW-to-RAW)27.360.933529.660.963724.210.917027.180.9388
DE-UNET(RAW-to-RGB)----24.200.919627.290.9419
", + "bbox": [ + 217, + 184, + 782, + 257 + ], + "page_idx": 10 + }, + { + "type": "table", + "img_path": "images/a3f698db37d5ced7ed48e35fed0ce9601d8c7cb0cb14a15fc37545d3383104f4.jpg", + "table_caption": [ + "Table 3: Results on our captured datasets for ZTE Axon 30 5G and Samsung Galaxy Z-Fold 3 smartphone display panels on various methods of UDC image restoration. The best results are bold-faced." + ], + "table_footnote": [], + "table_body": "
Input-to-outputMethodParamsMACsAxon 30 (rgb)Z-Fold 3 (rgb)
PSNRSSIMPSNRSSIM
RGB-to-RGBDE-UNET [38]8.93M273.84G23.330.910926.900.9390
DAGF [31]1.09M73.84G24.490.915028.420.9404
DISCNET [10]3.80M596.95G23.110.906326.670.9361
BNUDC [15]4.57M993.38G25.850.928328.540.9458
DWFormer [36]1.45M216.77G21.520.892326.360.9338
SRUDC [30]6.70M371.32G22.670.899827.340.9336
PPM-UNET [9]4.23M409.29G18.210.850223.850.9188
AWNet [7]47.00M1.53T20.270.883925.600.9316
RAW-to-RGBDE-UNET [38]8.94M69.29G24.200.919627.290.9419
PyNet [14]47.56M1.79T22.580.906927.350.9377
AWNet [7]49.07M392.25G20.980.888527.250.9369
DREUDC6.44M200.07G26.660.936829.040.9511
", + "bbox": [ + 217, + 325, + 782, + 537 + ], + "page_idx": 10 + }, + { + "type": "text", + "text": "raw UDC images to a clean RGB image (RAW-to-RGB). This is also the setting of most neural ISPs currently being used in smartphones [26] since a successful network is able to model the image processing pipeline as well.", + "bbox": [ + 215, + 566, + 784, + 612 + ], + "page_idx": 10 + }, + { + "type": "text", + "text": "4.4 Evaluation on Our Dataset", + "text_level": 1, + "bbox": [ + 215, + 633, + 483, + 647 + ], + "page_idx": 10 + }, + { + "type": "text", + "text": "To provide experimental results on our dataset and demonstrate the performance of DREUDC on UDC image restoration, we train numerous previous methods using our dataset for comparison. The quality of the restoration performance is evaluated with PSNR and SSIM as performance metrics. Additionally, we assess the efficiency of the methods by comparing the number of parameters and MACs. To calculate the MACs, we use a dummy image with a resolution of $1024 \\times 1024$ .", + "bbox": [ + 215, + 657, + 785, + 748 + ], + "page_idx": 10 + }, + { + "type": "text", + "text": "Tab. 3 shows experimental results for raw and RGB image restoration using our dataset. Because there are no neural networks on UDC image restoration that take raw sensor images as input except [38] as previous datasets were provided only in the RGB domain, we choose [7] and [14] for further comparison, which are RAW-to-RGB image conversion models. We observe that DREUDC achieves state-of-the-art performance on both Axon 30 and Z-fold 3 datasets,", + "bbox": [ + 215, + 750, + 785, + 839 + ], + "page_idx": 10 + }, + { + "type": "header", + "text": "Panel-Specific UDC Image Restoration", + "bbox": [ + 470, + 114, + 730, + 128 + ], + "page_idx": 10 + }, + { + "type": "page_number", + "text": "11", + "bbox": [ + 767, + 114, + 782, + 126 + ], + "page_idx": 10 + }, + { + "type": "image", + "img_path": "images/cf58ad7779c04e55c3e89ea859c1eaaca0bbc928c454a5744b52b5ec35cd483f.jpg", + "image_caption": [ + "Axon 30" + ], + "image_footnote": [], + "bbox": [ + 230, + 147, + 302, + 203 + ], + "page_idx": 11 + }, + { + "type": "image", + "img_path": "images/5a1a8cb18d978c4675bdfe9e419cdcd277ec46436c5f66f118596ce7557bcb4c.jpg", + "image_caption": [ + "DE-UNET" + ], + "image_footnote": [], + "bbox": [ + 307, + 147, + 382, + 203 + ], + "page_idx": 11 + }, + { + "type": "image", + "img_path": "images/443c069547bae31500f4df171a91cef2cb0d9ece0a885cf0b621b624b1fe3eaa.jpg", + "image_caption": [ + "DAGF" + ], + "image_footnote": [], + "bbox": [ + 387, + 147, + 460, + 203 + ], + "page_idx": 11 + }, + { + "type": "image", + "img_path": "images/1caa6d41d56031b6e8ea6abf11ee2cfdfc93a7366a8b91dd6a93cddc0c73735e.jpg", + "image_caption": [ + "DISCNET" + ], + "image_footnote": [], + "bbox": [ + 467, + 147, + 537, + 203 + ], + "page_idx": 11 + }, + { + "type": "image", + "img_path": "images/999eaef93d5b29bc7c11ce9721f9aa5977a0822732c42b96c9321c8dd2e228a2.jpg", + "image_caption": [ + "BNUDC" + ], + "image_footnote": [], + "bbox": [ + 544, + 147, + 617, + 203 + ], + "page_idx": 11 + }, + { + "type": "image", + "img_path": "images/de2b7ff4b043a6512ebb7b040e478633b48541ee459f2fc7c90449b255b5603e.jpg", + "image_caption": [ + "DWFormer" + ], + "image_footnote": [], + "bbox": [ + 624, + 147, + 696, + 203 + ], + "page_idx": 11 + }, + { + "type": "image", + "img_path": "images/6096c77c949deacc39372e624b262af9bf37fc56928e6b2e9e860f9bf079b002.jpg", + "image_caption": [ + "SRUDC" + ], + "image_footnote": [], + "bbox": [ + 702, + 147, + 774, + 203 + ], + "page_idx": 11 + }, + { + "type": "image", + "img_path": "images/51089fc1f79f4c8142433ecfc7d382fe91b547626d99c32fd262a77dc8e5b7f6.jpg", + "image_caption": [ + "PPM-UNet" + ], + "image_footnote": [], + "bbox": [ + 230, + 213, + 302, + 267 + ], + "page_idx": 11 + }, + { + "type": "image", + "img_path": "images/4ff5e71daf11dd9b66559d9a4f3362ff0ada86b13a0d4b0fafe821c8b084d69d.jpg", + "image_caption": [ + "AWNet" + ], + "image_footnote": [], + "bbox": [ + 308, + 213, + 382, + 267 + ], + "page_idx": 11 + }, + { + "type": "image", + "img_path": "images/36927a5073b0582a182a878734401017125dcf564e9b5c2a37e9203de448c778.jpg", + "image_caption": [ + "DE-UNET*" + ], + "image_footnote": [], + "bbox": [ + 387, + 213, + 460, + 267 + ], + "page_idx": 11 + }, + { + "type": "image", + "img_path": "images/3bfee9f0cd52fd79fe809ff4a57f8eaf3bc47d042de60068cc678c81f4e468b7.jpg", + "image_caption": [ + "PyNet*" + ], + "image_footnote": [], + "bbox": [ + 467, + 213, + 537, + 267 + ], + "page_idx": 11 + }, + { + "type": "image", + "img_path": "images/da2042856f64d2302393156e591606e5692b43e16cc78a33c0837b612ab2e325.jpg", + "image_caption": [ + "AWNet*" + ], + "image_footnote": [], + "bbox": [ + 545, + 213, + 617, + 267 + ], + "page_idx": 11 + }, + { + "type": "image", + "img_path": "images/cb53a85ef3ac108165db10d937ea1362965ad65f48b7ce111eef22a1d6ecbeae.jpg", + "image_caption": [ + "DREUDC*" + ], + "image_footnote": [], + "bbox": [ + 624, + 213, + 696, + 267 + ], + "page_idx": 11 + }, + { + "type": "image", + "img_path": "images/4f45c0a2f1c303941e6fb978baae763cedccde3d64323f982c5031a416288961.jpg", + "image_caption": [ + "GT" + ], + "image_footnote": [], + "bbox": [ + 702, + 213, + 774, + 267 + ], + "page_idx": 11 + }, + { + "type": "image", + "img_path": "images/28de6ae19e42b77bbbc157d038977b00bd492172412b94a6d946b5f1c3bac0f8.jpg", + "image_caption": [ + "Z-Fold 3" + ], + "image_footnote": [], + "bbox": [ + 230, + 277, + 302, + 332 + ], + "page_idx": 11 + }, + { + "type": "image", + "img_path": "images/64fecfff12cf6f2be39cd3bbba86cdc45930ac18b4a453adcd1d90557a229167.jpg", + "image_caption": [ + "DE-UNET" + ], + "image_footnote": [], + "bbox": [ + 308, + 277, + 382, + 332 + ], + "page_idx": 11 + }, + { + "type": "image", + "img_path": "images/11f2f6831a9fc818a427e74d9dc74aac14688a42aaee80aad7c7615d7ffe8788.jpg", + "image_caption": [ + "DAGE" + ], + "image_footnote": [], + "bbox": [ + 387, + 277, + 460, + 332 + ], + "page_idx": 11 + }, + { + "type": "image", + "img_path": "images/c913bb6114da146be22a155aac2b9b910b8f85fda682cf58f9da7f04660978f3.jpg", + "image_caption": [ + "DISCNET" + ], + "image_footnote": [], + "bbox": [ + 467, + 277, + 537, + 332 + ], + "page_idx": 11 + }, + { + "type": "image", + "img_path": "images/e91c89e4cf8ed108ccdc67db7c4c6484fa1f7f3eb632bb30a0af0b8ae862ef5d.jpg", + "image_caption": [ + "BNUDC" + ], + "image_footnote": [], + "bbox": [ + 545, + 277, + 617, + 332 + ], + "page_idx": 11 + }, + { + "type": "image", + "img_path": "images/3606fcd252b55853f99f3b3f00923e2e553763b8da48090f01d5f27500e00c16.jpg", + "image_caption": [ + "DWFormer" + ], + "image_footnote": [], + "bbox": [ + 624, + 277, + 696, + 332 + ], + "page_idx": 11 + }, + { + "type": "image", + "img_path": "images/fa9928952d2f6f5c2d057093b78ab1d74df365c69a455761edf511511e6b478c.jpg", + "image_caption": [ + "SRUDC" + ], + "image_footnote": [], + "bbox": [ + 702, + 277, + 774, + 332 + ], + "page_idx": 11 + }, + { + "type": "image", + "img_path": "images/ab01dbcaa3f52cceddbbb6f3978fec19d414d9599db3f1b4e1bfaf9e4b9508a8.jpg", + "image_caption": [ + "PPM-UNet", + "Fig. 5: Visual comparison of restoration examples for our dataset. Zoom in for a better comparison between the methods. The top two rows are results for Axon 30 images, and the bottom two rows are for Z-Fold 3 images. Methods that take raw input and produce RGB images are annotated with the asterisk *. Our method is capable of removing UDC degradations more clearly than previous UDC image restoration networks." + ], + "image_footnote": [], + "bbox": [ + 230, + 340, + 302, + 396 + ], + "page_idx": 11 + }, + { + "type": "image", + "img_path": "images/917c1e9d880f2e4772a68b9194a7e0a054e967976951b4c4f3a8799838b2f09e.jpg", + "image_caption": [ + "AWNet" + ], + "image_footnote": [], + "bbox": [ + 308, + 340, + 382, + 396 + ], + "page_idx": 11 + }, + { + "type": "image", + "img_path": "images/7581d9d8b76118ca94beef42656713ac9c3892ab949ceeb6eb406fa43a4e9d72.jpg", + "image_caption": [ + "DE-UNET*" + ], + "image_footnote": [], + "bbox": [ + 387, + 340, + 460, + 396 + ], + "page_idx": 11 + }, + { + "type": "image", + "img_path": "images/bb7cb2031bd8d94490db9686bc8a0a5393d58c215d8a4192d976d1750caae689.jpg", + "image_caption": [ + "PyNet*" + ], + "image_footnote": [], + "bbox": [ + 467, + 340, + 537, + 396 + ], + "page_idx": 11 + }, + { + "type": "image", + "img_path": "images/551417cccfd2aa5f9376bbda18fa2f3829c48d51fba0842194d24b60effe900c.jpg", + "image_caption": [ + "AWNet*" + ], + "image_footnote": [], + "bbox": [ + 545, + 340, + 617, + 396 + ], + "page_idx": 11 + }, + { + "type": "image", + "img_path": "images/0eebec71e83df9cc38e8ef93358879ed3008d6c100e0e3f08ccf2defd69afe42.jpg", + "image_caption": [ + "DREUDC*" + ], + "image_footnote": [], + "bbox": [ + 624, + 340, + 696, + 396 + ], + "page_idx": 11 + }, + { + "type": "image", + "img_path": "images/720186f975f12819cef15bb5695abc5a413fc91798a8ab844a3c80122a87a805.jpg", + "image_caption": [ + "GT" + ], + "image_footnote": [], + "bbox": [ + 702, + 340, + 774, + 396 + ], + "page_idx": 11 + }, + { + "type": "text", + "text": "surpassing BNUDC [15] by 0.81dB and 0.50dB, demonstrating that our method can successfully restore various UDC degradations. With the exception of a lightweight model [31] and the baseline model using raw data input [38], our method is the most efficient when it comes to computation while showing superior performance. Qualitative results are illustrated in Fig. 5, and more results including restoration of scenes that are not displayed on a monitor are available in the Supplementary Material.", + "bbox": [ + 212, + 521, + 787, + 628 + ], + "page_idx": 11 + }, + { + "type": "text", + "text": "4.5 Analysis on Panel-Specific Degradation Representation", + "text_level": 1, + "bbox": [ + 214, + 648, + 715, + 666 + ], + "page_idx": 11 + }, + { + "type": "text", + "text": "In this subsection, we analyze the effectiveness of DREUDC and panel-specific degradation representations learned with our proposed method. The strength of DREUDC is reported in Tab. 4 with an ablation study. We observe that our strategy of embedding an implicit representation of degraded UDC images is indeed effective by comparing (a) and (c) of Tab. 4, where Model 1 is our framework with only the restoration network without channel calibration and representation learning, while DREUDC is our entire framework trained with the protocol explained in Sec. 3.2.", + "bbox": [ + 212, + 672, + 787, + 794 + ], + "page_idx": 11 + }, + { + "type": "text", + "text": "However, with only this comparison, we cannot rule out the possibility that the rise in performance is solely because of increased parameters. Therefore, we further compare (a) and (c) with (b), where Model 2 is our framework trained in", + "bbox": [ + 212, + 795, + 787, + 840 + ], + "page_idx": 11 + }, + { + "type": "page_number", + "text": "12", + "bbox": [ + 217, + 114, + 235, + 126 + ], + "page_idx": 11 + }, + { + "type": "header", + "text": "Y. Oh et al.", + "bbox": [ + 271, + 114, + 354, + 127 + ], + "page_idx": 11 + }, + { + "type": "table", + "img_path": "images/42873a23ad1479e4f9c967e611ab7056ead45740aa386a48684e9312db625d39.jpg", + "table_caption": [ + "Table 4: Results on ablation of our method." + ], + "table_footnote": [], + "table_body": "
MethodChannel calibrationRepresentation learningParams MACsAxon 30Z-Fold 3
PNSRSSIMPNSRSSIM
(a) Model 1XX4.52M160.57G26.090.933828.66
(b) Model 2✓X6.44M200.07G26.380.934828.71
(c) DREUDC✓✓6.44M200.07G26.660.936829.04
", + "bbox": [ + 222, + 170, + 777, + 243 + ], + "page_idx": 12 + }, + { + "type": "image", + "img_path": "images/30a8160d912299f840c2c91ff8310327c227fc8485795b99133bf6aaf5c4acfa.jpg", + "image_caption": [ + "(a)", + "Fig. 6: T-SNE [21] visualization of representation generated from the encoder. (a) and (b) show the representations generated from encoders of Model 2 in Tab. 4 for Axon 30 and Z-fold 3 images, respectively, which are trained without representation learning. (d) and (e) show the representations generated from encoders of Model 2 in Tab. 5 for T-OLLED and P-OLED images, respectively, which are trained without representation learning. (c) and (f) illustrate the representations that are generated with our unsupervised triplet learning on our dataset and Zhou's dataset [38]." + ], + "image_footnote": [], + "bbox": [ + 251, + 282, + 336, + 340 + ], + "page_idx": 12 + }, + { + "type": "image", + "img_path": "images/ca577c75228f3381d81ff06c21f555e00c31b65ad3bae9bbb1b223ee79c211ab.jpg", + "image_caption": [ + "(b)" + ], + "image_footnote": [], + "bbox": [ + 341, + 282, + 418, + 340 + ], + "page_idx": 12 + }, + { + "type": "image", + "img_path": "images/c9c552d6708ed8c7b94a7aa5ae4cf7aec4b7a17cf944c8f105965aca7810b75a.jpg", + "image_caption": [ + "(c)" + ], + "image_footnote": [], + "bbox": [ + 423, + 284, + 500, + 340 + ], + "page_idx": 12 + }, + { + "type": "image", + "img_path": "images/985738607ba8d6416cbec90cb2aa1ee6e37818d715dca0d1e4e90c8bbebf302c.jpg", + "image_caption": [ + "(d)" + ], + "image_footnote": [], + "bbox": [ + 504, + 284, + 581, + 340 + ], + "page_idx": 12 + }, + { + "type": "image", + "img_path": "images/766f86b111c0330b006cea3df6e848b7d2b79161ad9c9acaa15edfa69e5e8252.jpg", + "image_caption": [ + "(e)" + ], + "image_footnote": [], + "bbox": [ + 589, + 284, + 665, + 339 + ], + "page_idx": 12 + }, + { + "type": "image", + "img_path": "images/e2b861872f2ccfa4529898572c04002f18e2d97e8ef9c3044a66afb75ae79f25.jpg", + "image_caption": [ + "(f)" + ], + "image_footnote": [], + "bbox": [ + 669, + 284, + 746, + 339 + ], + "page_idx": 12 + }, + { + "type": "text", + "text": "an end-to-end manner by training the encoder and the restoration network jointly with $\\mathcal{L}_1$ loss without our proposed representation learning. By comparing Model 1 and Model 2, we see that channel calibration with representations learned from a single dataset without triplet training increases the performance marginally. On the contrary, when given an appropriate representation learned from two datasets by leveraging the difference in degradations with our proposed method, we notice that the restoration model benefits substantially from it. Moreover, it is noteworthy that our proposed embedding scheme has a negligible effect on the computational increase (MACs) of the restoration network, as most of the increase is due to the encoder.", + "bbox": [ + 212, + 508, + 787, + 660 + ], + "page_idx": 12 + }, + { + "type": "text", + "text": "The validity of our panel-specific degradation representation learning scheme is visualized in Fig. 6. Without representation learning (Fig. 6(a) and (b)), the encoders are unable to distinguish different degradations. Conversely, with representation learning (Fig. 6(c)), the encoder identifies the degradation and discriminates them in clusters, offering advantageous priors for the restoration network.", + "bbox": [ + 212, + 667, + 787, + 743 + ], + "page_idx": 12 + }, + { + "type": "text", + "text": "This finding suggests an approach to train a restoration network by making good use of datasets other than the one gathered for a specific UDC system. Typically, a specific dataset is used to train a restoration network tailored for a particular UDC system. However, this dataset would be sub-optimal or ineffective when future UDC smartphones have different panels. Our method offers a way to utilize other datasets collected with different panels, mediating this issue.", + "bbox": [ + 212, + 750, + 787, + 840 + ], + "page_idx": 12 + }, + { + "type": "header", + "text": "Panel-Specific UDC Image Restoration", + "bbox": [ + 470, + 114, + 730, + 128 + ], + "page_idx": 12 + }, + { + "type": "page_number", + "text": "13", + "bbox": [ + 767, + 114, + 785, + 126 + ], + "page_idx": 12 + }, + { + "type": "table", + "img_path": "images/0ee50de85ce3425557997c1b7b2b04db5dc15488b67f4a483068f3e31c9e619b.jpg", + "table_caption": [ + "Table 5: Results of our method and ablation on T-OLED/P-OLED datasets [38]." + ], + "table_footnote": [], + "table_body": "
MethodChannel calibrationRepresentation learningParams MACsT-OLEDP-OLED
PSNR SSIMPSNR SSIM
DE-UNET [38]--8.94M 69.29G36.70 0.974528.58 0.9344
Model 1XX4.52M 160.57G37.37 0.977430.85 0.9510
Model 2✓X6.44M 200.07G37.46 0.977731.01 0.9515
DREUDC✓✓6.44M 200.07G37.55 0.977831.17 0.9521
", + "bbox": [ + 220, + 170, + 777, + 256 + ], + "page_idx": 13 + }, + { + "type": "text", + "text": "Application to previous datasets Our restoration method using panel-specific degradation representations is also applied to public UDC datasets to show that it generalizes well to other datasets. Our method requires a real dataset comprised of degraded images taken with more than a single panel from the same capturing environment, so we choose Zhou's P-OLED/T-OLED datasets [38]. DREUDC is designed to take raw images as input, so we make a pseudo-raw version of the dataset by mosaicing the RGB images. This pseudo-raw version of the RGB images that used to train the methods, including a new encoder, is created by reversing the in-camera pipeline mentioned in [38]. We only reversed the demosaicing step as it was the only mentioned process; hence we call it 'pseudo-raw' as the images have not been fully reversed to the original high-bit raw data. We choose DE-UNET [38] to compare with our results because it is a model that can be trained with RAW-to-RGB setting. Additional results on RGB image restoration are provided in the Supplementary Material.", + "bbox": [ + 212, + 285, + 785, + 496 + ], + "page_idx": 13 + }, + { + "type": "text", + "text": "Tab. 5 shows the results on T-OLED/P-OLED datasets. Our method outperforms DE-UNET on both panels with slightly fewer parameters. Most importantly, our strategy of representation embedding is also effective on [38] when we compare Model 1 to DREUDC, showing an increase of $0.18\\mathrm{dB}$ and $0.32\\mathrm{dB}$ in the restoration of T-OLED and P-OLED images. We also prove that the boost in performance is not only due to increased parameters but also because we enhance the restoration network with effective representations of the degradations learned using our training scheme by comparing Model 2 and DREUDC. Visualization of the representations are provided in Fig. 6(d), (e), and (f). Overall, the results imply that our method is also effective on other UDC datasets.", + "bbox": [ + 212, + 497, + 787, + 647 + ], + "page_idx": 13 + }, + { + "type": "text", + "text": "5 Conclusion", + "text_level": 1, + "bbox": [ + 215, + 671, + 359, + 686 + ], + "page_idx": 13 + }, + { + "type": "text", + "text": "We have collected a new UDC dataset using commercial UDC-equipped smartphones and demonstrated that UDC image restoration using raw sensor images achieves better performance than using RGB images. We have also proposed a UDC image restoration framework that generates panel-specific degradation representations in the embedding space with a novel unsupervised representation learning scheme. Our restoration network efficiently accommodates the learned representations as a prior with our proposed Transformer blocks, improving the performance. Experiments show that our method achieves state-of-the-art performance on our dataset and is generalizable to previous datasets.", + "bbox": [ + 212, + 703, + 787, + 840 + ], + "page_idx": 13 + }, + { + "type": "page_number", + "text": "14", + "bbox": [ + 217, + 114, + 235, + 126 + ], + "page_idx": 13 + }, + { + "type": "header", + "text": "Y. Oh et al.", + "bbox": [ + 271, + 114, + 354, + 127 + ], + "page_idx": 13 + }, + { + "type": "text", + "text": "Acknowledgment", + "text_level": 1, + "bbox": [ + 217, + 143, + 382, + 162 + ], + "page_idx": 14 + }, + { + "type": "text", + "text": "This paper was result of the research project supported by SK hynix Inc., including the support of camera equipments and lab environment for collecting the dataset. It was also partially supported by Institute of Information & communications Technology Planning & Evaluation (IITP) grant funded by the Korea government (MSIT) (No.2021-0-01062), and in part by the BK21 FOUR program of the Education and Research Program for Future ICT Pioneers, Seoul National University in 2024.", + "bbox": [ + 212, + 174, + 787, + 282 + ], + "page_idx": 14 + }, + { + "type": "text", + "text": "References", + "text_level": 1, + "bbox": [ + 217, + 301, + 321, + 316 + ], + "page_idx": 14 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "1. Abdelhamed, A., Lin, S., Brown, M.S.: A high-quality denoising dataset for smartphone cameras. In: Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition. pp. 1692-1700 (2018)", + "2. Agustsson, E., Timofte, R.: Ntire 2017 challenge on single image super-resolution: Dataset and study. In: Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition Workshops. pp. 126-135 (2017)", + "3. Bay, H., Tuytelaars, T., Van Gool, L.: Surf: Speeded up robust features. In: Computer Vision-ECCV 2006: 9th European Conference on Computer Vision, Graz, Austria, May 7-13, 2006. Proceedings, Part I 9. pp. 404-417. Springer (2006)", + "4. Chen, C., Chen, Q., Xu, J., Koltun, V.: Learning to see in the dark. In: Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition. pp. 3291-3300 (2018)", + "5. Chen, T., Kornblith, S., Norouzi, M., Hinton, G.: A simple framework for contrastive learning of visual representations. In: International Conference on Machine Learning. pp. 1597-1607. PMLR (2020)", + "6. Chen, X., Fan, H., Girshick, R., He, K.: Improved baselines with momentum contrastive learning. arXiv preprint arXiv:2003.04297 (2020)", + "7. Dai, L., Liu, X., Li, C., Chen, J.: Attnective wavelet network for image isp. In: Computer Vision-ECCV 2020 Workshops: Glasgow, UK, August 23-28, 2020, Proceedings, Part III 16. pp. 185-201. Springer (2020)", + "8. Deng, W., Zheng, L., Sun, Y., Jiao, J.: Rethinking triplet loss for domain adaptation. IEEE Transactions on Circuits and Systems for Video Technology 31(1), 29-37 (2020)", + "9. Feng, R., Li, C., Chen, H., Li, S., Gu, J., Loy, C.C.: Generating aligned pseudosupervision from non-aligned data for image restoration in under-display camera. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 5013-5022 (2023)", + "10. Feng, R., Li, C., Chen, H., Li, S., Loy, C.C., Gu, J.: Removing diffraction image artifacts in under-display camera via dynamic skip connection network. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 662-671 (2021)", + "11. Feng, R., Li, C., Zhou, S., Sun, W., Zhu, Q., Jiang, J., Yang, Q., Loy, C.C., Gu, J., Zhu, Y., et al.: Mipi 2022 challenge on under-display camera image restoration: Methods and results. In: European Conference on Computer Vision. pp. 60-77. Springer (2022)", + "12. Guizar-Sicairos, M., Thurman, S.T., Fienup, J.R.: Efficient subpixel image registration algorithms. Optics letters 33(2), 156-158 (2008)" + ], + "bbox": [ + 218, + 330, + 784, + 839 + ], + "page_idx": 14 + }, + { + "type": "header", + "text": "Panel-Specific UDC Image Restoration", + "bbox": [ + 470, + 114, + 730, + 128 + ], + "page_idx": 14 + }, + { + "type": "page_number", + "text": "15", + "bbox": [ + 767, + 116, + 784, + 126 + ], + "page_idx": 14 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "13. Hermans, A., Beyer, L., Leibe, B.: In defense of the triplet loss for person re-identification. arXiv preprint arXiv:1703.07737 (2017)", + "14. Ignatov, A., Van Gool, L., Timofte, R.: Replacing mobile camera isp with a single deep learning model. arXiv preprint arXiv:2002.05509 (2020)", + "15. Koh, J., Lee, J., Yoon, S.: Bnudc: a two-branched deep neural network for restoring images from under-display cameras. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 1950-1959 (2022)", + "16. Kong, L., Dong, J., Ge, J., Li, M., Pan, J.: Efficient frequency domain-based transformers for high-quality image deblurring. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 5886-5895 (2023)", + "17. Kwon, K., Kang, E., Lee, S., Lee, S.J., Lee, H.E., Yoo, B., Han, J.J.: Controllable image restoration for under-display camera in smartphones. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 2073-2082 (2021)", + "18. Lee, W., Son, S., Lee, K.M.: Ap-bsn: Self-supervised denoising for real-world images via asymmetric pd and blind-spot network. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 17725-17734 (2022)", + "19. Liang, C.H., Chen, Y.A., Liu, Y.C., Hsu, W.H.: Raw image deblurring. IEEE Transactions on Multimedia 24, 61-72 (2020)", + "20. Loshchilov, I., Hutter, F.: Sgdr: Stochastic gradient descent with warm restarts. arXiv preprint arXiv:1608.03983 (2016)", + "21. Van der Maaten, L., Hinton, G.: Visualizing data using t-sne. Journal of machine learning research 9(11) (2008)", + "22. Nam, S., Hwang, Y., Matsushita, Y., Kim, S.J.: A holistic approach to cross-channel image noise modeling and its application to image denoising. In: Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition. pp. 1683-1691 (2016)", + "23. Oh, Y., Park, G.Y., Cho, N.I.: Restoration of high-frequency components in under display camera images. In: 2022 Asia-Pacific Signal and Information Processing Association Annual Summit and Conference (APSIPA ASC). pp. 1040-1046. IEEE (2022)", + "24. Oh, Y., Park, G.Y., Chung, H., Cho, S., Cho, N.I.: Residual dilated u-net with spatially adaptive normalization for the restoration of under display camera images. In: 2021 Asia-Pacific Signal and Information Processing Association Annual Summit and Conference (APSIPA ASC). pp. 151-157. IEEE (2021)", + "25. Plotz, T., Roth, S.: Benchmarking denoising algorithms with real photographs. In: Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition. pp. 1586-1595 (2017)", + "26. Punnappurath, A., Abuolaim, A., Abdelhamed, A., Levinshtein, A., Brown, M.S.: Day-to-night image synthesis for training nighttime neural isps. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 10769-10778 (2022)", + "27. Ronneberger, O., Fischer, P., Brox, T.: U-net: Convolutional networks for biomedical image segmentation. In: Medical Image Computing and Computer-Assisted Intervention-MICCAI 2015: 18th International Conference, Munich, Germany, October 5-9, 2015, Proceedings, Part III 18. pp. 234-241. Springer (2015)", + "28. Schroff, F., Kalenichenko, D., Philbin, J.: Facenet: A unified embedding for face recognition and clustering. In: Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition. pp. 815-823 (2015)" + ], + "bbox": [ + 217, + 147, + 784, + 839 + ], + "page_idx": 15 + }, + { + "type": "page_number", + "text": "16", + "bbox": [ + 217, + 114, + 235, + 126 + ], + "page_idx": 15 + }, + { + "type": "header", + "text": "Y. Oh et al.", + "bbox": [ + 271, + 114, + 352, + 126 + ], + "page_idx": 15 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "29. Schwartz, E., Giryes, R., Bronstein, A.M.: Deepisp: Toward learning an end-to-end image processing pipeline. IEEE Transactions on Image Processing 28(2), 912-923 (2018)", + "30. Song, B., Chen, X., Xu, S., Zhou, J.: Under-display camera image restoration with scattering effect. In: Proceedings of the IEEE/CVF International Conference on Computer Vision. pp. 12580-12589 (2023)", + "31. Sundar, V., Hegde, S., Kothandaraman, D., Mitra, K.: Deep atrous guided filter for image restoration in under display cameras. In: Computer Vision-ECCV 2020 Workshops: Glasgow, UK, August 23-28, 2020, Proceedings, Part V 16. pp. 379-397. Springer (2020)", + "32. Wang, L., Wang, Y., Dong, X., Xu, Q., Yang, J., An, W., Guo, Y.: Unsupervised degradation representation learning for blind super-resolution. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 10581-10590 (2021)", + "33. Yu, B., Liu, T., Gong, M., Ding, C., Tao, D.: Correcting the triplet selection bias for triplet loss. In: Proceedings of the European Conference on Computer Vision (ECCV). pp. 71-87 (2018)", + "34. Yuan, Y., Chen, W., Yang, Y., Wang, Z.: In defense of the triplet loss again: Learning robust person re-identification with fast approximated triplet loss and label distillation. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition Workshops. pp. 354-355 (2020)", + "35. Zamir, S.W., Arora, A., Khan, S., Hayat, M., Khan, F.S., Yang, M.H., Shao, L.: Cycleisp: Real image restoration via improved data synthesis. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 2696-2705 (2020)", + "36. Zhou, Y., Song, Y., Du, X.: Modular degradation simulation and restoration for under-display camera. In: Proceedings of the Asian Conference on Computer Vision. pp. 265-282 (2022)", + "37. Zhou, Y., Kwan, M., Tolentino, K., Emerton, N., Lim, S., Large, T., Fu, L., Pan, Z., Li, B., Yang, Q., et al.: Udc 2020 challenge on image restoration of under-display camera: Methods and results. In: Computer Vision-ECCV 2020 Workshops: Glasgow, UK, August 23-28, 2020, Proceedings, Part V 16. pp. 337-351. Springer (2020)", + "38. Zhou, Y., Ren, D., Emerton, N., Lim, S., Large, T.: Image restoration for underdisplay camera. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 9179-9188 (2021)" + ], + "bbox": [ + 212, + 146, + 787, + 646 + ], + "page_idx": 16 + }, + { + "type": "header", + "text": "Panel-Specific UDC Image Restoration", + "bbox": [ + 470, + 114, + 730, + 128 + ], + "page_idx": 16 + }, + { + "type": "page_number", + "text": "17", + "bbox": [ + 767, + 114, + 785, + 126 + ], + "page_idx": 16 + } +] \ No newline at end of file diff --git a/2024/Panel-Specific Degradation Representation for Raw Under-Display Camera Image Restoration/c089de88-5e00-46c3-9141-1e16159753ca_model.json b/2024/Panel-Specific Degradation Representation for Raw Under-Display Camera Image Restoration/c089de88-5e00-46c3-9141-1e16159753ca_model.json new file mode 100644 index 0000000000000000000000000000000000000000..dbc33335d7ca563466280d837c44aa51ad3f0afc --- /dev/null +++ b/2024/Panel-Specific Degradation Representation for Raw Under-Display Camera Image Restoration/c089de88-5e00-46c3-9141-1e16159753ca_model.json @@ -0,0 +1,3193 @@ +[ + [ + { + "type": "title", + "bbox": [ + 0.232, + 0.141, + 0.773, + 0.187 + ], + "angle": 0, + "content": "Panel-Specific Degradation Representation for Raw Under-Display Camera Image Restoration" + }, + { + "type": "text", + "bbox": [ + 0.232, + 0.212, + 0.773, + 0.243 + ], + "angle": 0, + "content": "Youngjin Oh\\(^{1}\\), Keuntek Lee\\(^{1}\\), Jooyoung Lee\\(^{3}\\), Dae-Hyun Lee\\(^{3}\\), and Nam Ik Cho\\(^{1,2}\\)" + }, + { + "type": "text", + "bbox": [ + 0.258, + 0.254, + 0.743, + 0.282 + ], + "angle": 0, + "content": "\\(^{1}\\) Department of ECE & INMC, Seoul National University, Seoul, Korea {yjymoh0211,leekt000,nicho}@snu.ac.kr" + }, + { + "type": "text", + "bbox": [ + 0.342, + 0.283, + 0.661, + 0.297 + ], + "angle": 0, + "content": "\\(^{2}\\) IPAI, Seoul National University, Seoul, Korea" + }, + { + "type": "text", + "bbox": [ + 0.366, + 0.297, + 0.639, + 0.325 + ], + "angle": 0, + "content": "3 SK hynix Inc., Korea {jooyoung2.lee, daehyun1.lee}@sk.com" + }, + { + "type": "list", + "bbox": [ + 0.258, + 0.254, + 0.743, + 0.325 + ], + "angle": 0, + "content": null + }, + { + "type": "text", + "bbox": [ + 0.262, + 0.36, + 0.74, + 0.652 + ], + "angle": 0, + "content": "Abstract. Under-display camera (UDC) image restoration aims to restore images distorted by the OLED display panel covering the frontal camera on a smartphone. Previous deep learning-based UDC restoration methods focused on restoring the image within the RGB domain with the collection of real or synthetic RGB datasets. However, UDC images in these datasets exhibit domain differences from real commercial smartphone UDC images while inherently constraining the problem and solution within the RGB domain. To address this issue, we collect well-aligned sensor-level real UDC images using panels from two commercial smartphones equipped with UDC. We also propose a new UDC restoration method to exploit the disparities between degradations caused by different panels, considering that UDC degradations are specific to the type of OLED panel. For this purpose, we train an encoder with an unsupervised learning scheme using triplet loss that aims to extract the inherent degradations caused by different panels from degraded UDC images as implicit representations. The learned panel-specific degradation representations are then provided as priors to our restoration network based on an efficient Transformer network. Extensive experiments show that our proposed method achieves state-of-the-art performance on our real raw image dataset and generalizes well to previous datasets. Our dataset and code is available at https://github.com/OBAKSA/DREUDC." + }, + { + "type": "text", + "bbox": [ + 0.261, + 0.665, + 0.741, + 0.694 + ], + "angle": 0, + "content": "Keywords: Under-Display Camera \\(\\cdot\\) Image Restoration \\(\\cdot\\) Representation Learning" + }, + { + "type": "title", + "bbox": [ + 0.216, + 0.719, + 0.377, + 0.735 + ], + "angle": 0, + "content": "1 Introduction" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.75, + 0.788, + 0.842 + ], + "angle": 0, + "content": "To hide the camera hole in the frontal side of smartphones, under-display camera (UDC) systems have been developed, enabled by recent advancements in imaging and display panel technologies. With the UDC technology, smartphone users can enjoy a full-screen display without any camera holes or notches on the front side of their device. However, since the camera is hidden under the panel, the light that reaches the camera is altered by the complex layers of the panel, causing" + } + ], + [ + { + "type": "page_number", + "bbox": [ + 0.218, + 0.116, + 0.23, + 0.127 + ], + "angle": 0, + "content": "2" + }, + { + "type": "header", + "bbox": [ + 0.272, + 0.115, + 0.355, + 0.128 + ], + "angle": 0, + "content": "Y. Oh et al." + }, + { + "type": "table_caption", + "bbox": [ + 0.22, + 0.145, + 0.782, + 0.16 + ], + "angle": 0, + "content": "Table 1: Comparison of previous datasets [9,10,30,38] and our commercial dataset." + }, + { + "type": "table", + "bbox": [ + 0.229, + 0.171, + 0.768, + 0.237 + ], + "angle": 0, + "content": "
DatasetZhou et al. [38]Feng et al. [10]Feng et al. [9]Song et al. [30]Ours
FormatRGBRGBRGBRGBRGB/RAW
UDC ModelingReal (prototype)SyntheticRealSyntheticReal
Aligned GT✓✓✗✓✓
Multiple Display✓✗✗✗✓
" + }, + { + "type": "image", + "bbox": [ + 0.319, + 0.255, + 0.393, + 0.312 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.316, + 0.313, + 0.463, + 0.326 + ], + "angle": 0, + "content": "(a) Feng et al. [9], GT" + }, + { + "type": "image", + "bbox": [ + 0.396, + 0.257, + 0.457, + 0.31 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.4, + 0.312, + 0.462, + 0.326 + ], + "angle": 0, + "content": "l. [9], GT" + }, + { + "type": "image", + "bbox": [ + 0.527, + 0.255, + 0.599, + 0.312 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.518, + 0.313, + 0.678, + 0.326 + ], + "angle": 0, + "content": "(b) Feng et al. [9], UDC" + }, + { + "type": "image", + "bbox": [ + 0.603, + 0.257, + 0.665, + 0.31 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.6, + 0.312, + 0.678, + 0.326 + ], + "angle": 0, + "content": "al. [9], UDC" + }, + { + "type": "image", + "bbox": [ + 0.242, + 0.327, + 0.317, + 0.382 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.232, + 0.384, + 0.388, + 0.397 + ], + "angle": 0, + "content": "(d) Feng et al. [10], GT" + }, + { + "type": "image", + "bbox": [ + 0.319, + 0.328, + 0.38, + 0.383 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.319, + 0.384, + 0.388, + 0.397 + ], + "angle": 0, + "content": "l. [10], GT" + }, + { + "type": "image", + "bbox": [ + 0.423, + 0.328, + 0.496, + 0.384 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.413, + 0.384, + 0.579, + 0.397 + ], + "angle": 0, + "content": "(e) Feng et al. [10], UDC" + }, + { + "type": "image", + "bbox": [ + 0.604, + 0.328, + 0.677, + 0.384 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.589, + 0.384, + 0.755, + 0.397 + ], + "angle": 0, + "content": "(f) Song et al. [30], UDC" + }, + { + "type": "image", + "bbox": [ + 0.242, + 0.399, + 0.317, + 0.454 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.227, + 0.456, + 0.355, + 0.47 + ], + "angle": 0, + "content": "(g) Zhou et al. [38]," + }, + { + "type": "image", + "bbox": [ + 0.319, + 0.399, + 0.38, + 0.453 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.388, + 0.456, + 0.436, + 0.469 + ], + "angle": 0, + "content": "(h) Zho" + }, + { + "type": "image", + "bbox": [ + 0.423, + 0.399, + 0.496, + 0.454 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.413, + 0.456, + 0.583, + 0.469 + ], + "angle": 0, + "content": "Zhou et al. [38], T-OLED" + }, + { + "type": "image", + "bbox": [ + 0.604, + 0.399, + 0.678, + 0.454 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.6, + 0.456, + 0.679, + 0.469 + ], + "angle": 0, + "content": ") Zhou et al" + }, + { + "type": "image", + "bbox": [ + 0.681, + 0.399, + 0.743, + 0.453 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.68, + 0.456, + 0.777, + 0.469 + ], + "angle": 0, + "content": "[38], P-OLED" + }, + { + "type": "image", + "bbox": [ + 0.242, + 0.471, + 0.317, + 0.526 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.266, + 0.527, + 0.354, + 0.54 + ], + "angle": 0, + "content": "(j) Ours, GT" + }, + { + "type": "image", + "bbox": [ + 0.319, + 0.471, + 0.38, + 0.526 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.423, + 0.471, + 0.495, + 0.526 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.434, + 0.527, + 0.557, + 0.541 + ], + "angle": 0, + "content": "(k) Ours, Axon 30" + }, + { + "type": "image", + "bbox": [ + 0.604, + 0.47, + 0.677, + 0.526 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.612, + 0.527, + 0.731, + 0.54 + ], + "angle": 0, + "content": "(1) Ours, Z-Fold 3" + }, + { + "type": "image_caption", + "bbox": [ + 0.216, + 0.555, + 0.788, + 0.585 + ], + "angle": 0, + "content": "Fig. 1: Samples and their RGB channel histograms from previous UDC datasets [9, 10,30,38] and our commercial dataset." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.614, + 0.785, + 0.659 + ], + "angle": 0, + "content": "severe degradation in the resulting image. Degradations include blur, diffraction, noise, low-lightness, color shift, haziness, etc., and their spatially-variant nature makes the UDC image restoration very challenging [10, 17, 38]." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.66, + 0.788, + 0.841 + ], + "angle": 0, + "content": "Most UDC image restoration methods used deep neural networks [9-11, 15, 17, 23, 24, 30, 31, 37, 38], providing promising results on public benchmark datasets [9, 10, 30, 38] gathered to facilitate research on UDC image restoration. However, these existing datasets either use prototype display panels to inadequately represent the degradations observed in real UDC images caused by commercial panels [38], lack realistic degradations due to incomplete UDC modeling (e.g., noise and color shift) in the synthesis pipeline [10,30], or mishandles aligning image pairs with occlusion and parallax due to the mechanism of the capturing imaging system [9]. In addition, the potential for enhanced image restoration performance through sensor-level restoration [1, 19, 25, 35] cannot be exploited, because previous datasets predominantly restrict experiments to RGB domain as raw data is unreleased. Tab. 1 shows comparison of the datasets." + } + ], + [ + { + "type": "header", + "bbox": [ + 0.472, + 0.115, + 0.732, + 0.129 + ], + "angle": 0, + "content": "Panel-Specific UDC Image Restoration" + }, + { + "type": "page_number", + "bbox": [ + 0.775, + 0.117, + 0.785, + 0.127 + ], + "angle": 0, + "content": "3" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.147, + 0.784, + 0.313 + ], + "angle": 0, + "content": "In this work, we propose a new UDC dataset that is collected in the raw domain. The proposed dataset contains pairs of real UDC images and clean images that are well-aligned using a monitor-based imaging system with two different panels from commercial UDC smartphones, ZTE Axon 30 5G and Samsung Galaxy Z-Fold 3. In our experiments, we find that training with raw data enhances the performance of UDC image restoration. This raw-domain approach with real data also has the advantage of enabling the use of end-to-end neural ISPs that are being installed in recent smartphones [26], allowing a practical implementation on devices. Additionally, as shown in Fig. 1, real UDC data captured with commercial smartphone panels display noticeable distinct degradations visually and statistically when compared to previous datasets [10,30,38]." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.314, + 0.785, + 0.554 + ], + "angle": 0, + "content": "Furthermore, we develop a two-stage restoration method for UDC images. First, we introduce a framework for learning the implicit representation of degradations in the UDC images. We notice that degradation in a patch of a UDC image is similar to that of other patches of the same UDC image while being distinct from degradations of the same scene taken with a different panel. With this observation, we develop a novel unsupervised learning scheme involving triplet loss [28] to train our encoder. The learning scheme aims to leverage the information of the data collected with different types of panels, ultimately condensing the degradation information into an implicit representation vector that reflects the panel-specific degradation moderately. We also propose an efficient Transformer-based restoration network that incorporates the learned panel-specific degradation representations as a prior by generating channel-wise calibration coefficients for attention. We discover that our strategy yields improved results for UDC image restoration. To the best of our knowledge, this is the first case in UDC image restoration where a dataset collected with one panel is utilized to enhance the performance on a dataset collected with another panel." + }, + { + "type": "text", + "bbox": [ + 0.24, + 0.555, + 0.583, + 0.569 + ], + "angle": 0, + "content": "To summarize, our contributions are as follows:" + }, + { + "type": "text", + "bbox": [ + 0.226, + 0.575, + 0.784, + 0.619 + ], + "angle": 0, + "content": "- We introduce a new real UDC dataset, which is collected with commercial smartphone panels. Our dataset is provided in raw format, which can improve the performance of the UDC restoration task." + }, + { + "type": "text", + "bbox": [ + 0.226, + 0.62, + 0.785, + 0.693 + ], + "angle": 0, + "content": "- We propose an unsupervised learning approach that allows for implicit representations of degradations caused by various display panels in the embedding space. The panel-specific degradation representations are leveraged as priors in the restoration network, which is an efficient Transformer and brings notable improvement." + }, + { + "type": "text", + "bbox": [ + 0.226, + 0.695, + 0.784, + 0.738 + ], + "angle": 0, + "content": "- Extensive experiments demonstrate that our restoration network embedded with learned representations achieves state-of-the-art performance on the proposed dataset and is consistently applicable to previous datasets." + }, + { + "type": "list", + "bbox": [ + 0.226, + 0.575, + 0.785, + 0.738 + ], + "angle": 0, + "content": null + }, + { + "type": "title", + "bbox": [ + 0.216, + 0.759, + 0.396, + 0.775 + ], + "angle": 0, + "content": "2 Related Works" + }, + { + "type": "title", + "bbox": [ + 0.216, + 0.788, + 0.465, + 0.803 + ], + "angle": 0, + "content": "2.1 UDC Image Restoration" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.81, + 0.784, + 0.84 + ], + "angle": 0, + "content": "Since UDC image restoration is a relatively new area, research is still ongoing to provide users with a satisfying experience when UDC is applied in real situa-" + } + ], + [ + { + "type": "page_number", + "bbox": [ + 0.218, + 0.116, + 0.23, + 0.127 + ], + "angle": 0, + "content": "4" + }, + { + "type": "header", + "bbox": [ + 0.272, + 0.115, + 0.355, + 0.128 + ], + "angle": 0, + "content": "Y. Oh et al." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.147, + 0.788, + 0.313 + ], + "angle": 0, + "content": "tions. Research on UDC image restoration was sparked by the pioneering work of [38]. They defined a UDC imaging system and released the first UDC dataset for researchers to experiment with. Synthetic datasets were also made available by simulating the camera pipeline using a pre-calculated point spread function (PSF), which adequately represents the diffraction [10], with further improvements on scattering effects in [30] to account for the haziness of UDC images. More recently, a real UDC dataset generated by pseudo-aligning geometrically misaligned image pairs for a single UDC panel [9] was presented. Although the proposed datasets have some drawbacks, several works [9-11, 15, 17, 23, 24, 30, 31, 37, 38] have investigated UDC image restoration using deep neural networks and have shown promising results." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.314, + 0.788, + 0.51 + ], + "angle": 0, + "content": "Among these deep-learning-based methods, several of them use priors to improve performance. For instance, in [10], PCA is employed on a PSF of the UDC display panel to stretch it as a 'kernel code' and condition the network with it as a prior. In [30], the network is divided into a scattering branch and an image branch, and the former is trained to estimate the scattering parameters, which is prior information that can be used. However, these methods require precise priors; otherwise, the performance is degraded. Our approach differs from previous methods because we do not need strict parameters or ground truth settings. Instead, we aim to learn an acceptable representation of the degradation induced by panel-specific properties in the embedding space rather than a rigorous estimation of the degradation. We only need images taken with different panels because our method is unsupervised, and learn the implicit priors by leveraging the discrepancies of degradation caused by different panels." + }, + { + "type": "title", + "bbox": [ + 0.215, + 0.53, + 0.547, + 0.545 + ], + "angle": 0, + "content": "2.2 Image Restoration with Raw Data" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.554, + 0.788, + 0.842 + ], + "angle": 0, + "content": "Most of the previous research on image restoration relies on RGB images as training data as they are easier to collect. However, some recent studies have shown that using raw sensor data can lead to improved performance. Specifically, [1, 25] are the works on image denoising where they collected DND and SIDD datasets that are available as raw and RGB images, where denoising with raw images usually yields higher PSNR. CycleISP [35] demonstrated that noise is more complex in RGB images and showed that noise modeling is more realistic in the sensor domain. This is because compared to the noise samples in the sensor pixels, which are nearly uncorrelated to each other, the noise samples become highly correlated in RGB images as they pass through the camera ISP pipeline, which involves processes like demosaicing, gamma correction, and white balancing [18, 22]. DeepISP [29] proposed an end-to-end camera ISP modeled with deep neural networks, suggesting that a deep network-based camera ISP is promising. Furthermore, [4] solved the issue of low-light enhancement problem by operating in the sensor domain, and [19] presented a dataset for raw image deblurring and proved that operating on raw sensor data achieves better performance than RGB-based methods. These researches suggest that it is easier to restore raw data than RGB images with deep neural networks, as sensor data contains the linear representation of incoming light intensity of the scene. This ensures that" + } + ], + [ + { + "type": "header", + "bbox": [ + 0.472, + 0.115, + 0.732, + 0.13 + ], + "angle": 0, + "content": "Panel-Specific UDC Image Restoration" + }, + { + "type": "page_number", + "bbox": [ + 0.775, + 0.117, + 0.785, + 0.127 + ], + "angle": 0, + "content": "5" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.147, + 0.784, + 0.178 + ], + "angle": 0, + "content": "the scene's authenticity is maintained more faithfully in raw data than in heavily processed RGB pixel values that are nonlinear with respect to light intensity." + }, + { + "type": "title", + "bbox": [ + 0.215, + 0.2, + 0.427, + 0.218 + ], + "angle": 0, + "content": "3 Proposed Method" + }, + { + "type": "title", + "bbox": [ + 0.215, + 0.23, + 0.522, + 0.245 + ], + "angle": 0, + "content": "3.1 Sensor-level Real UDC Dataset" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.255, + 0.788, + 0.36 + ], + "angle": 0, + "content": "A raw dataset is necessary to achieve end-to-end image restoration using sensor images. However, the available UDC datasets are limited to the RGB domain. Moreover, public UDC datasets do not show the types of degradation that are present in images taken with commercially available smartphones equipped with UDC, limiting their assessment of practical scenarios. Therefore, we capture real raw UDC data by using OLED panels obtained from commercial UDC smartphones, a smartphone camera, and a monitor-based imaging system [38]." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.381, + 0.788, + 0.622 + ], + "angle": 0, + "content": "UDC data acquisition process Capturing a raw UDC image requires that we model the UDC imaging system as similarly as possible. Consequently, our data-capturing process involves rendering an image onto a high-resolution monitor, followed by capturing the displayed image using a smartphone camera sensor located closely behind a UDC panel. The panels used to capture the data are from two commercial UDC smartphones: ZTE Axon 30 5G and Samsung Galaxy Z-Fold 3. They introduce low-light/color shift, scattering, diffraction, and blur. As noise is usually generated from the sensor, we chose Samsung Galaxy S23 Ultra as our camera device to capture the authentic noise of smartphone camera sensors. Raw images are collected using the internal software of the smartphone that automatically saves the minimally processed raw data. The preference for monitor-displayed images [17,38] over real natural scenes is driven by the need to control additional factors beyond the panels' presence. Specifically, we can enforce the target scene to be static as dynamic scenes are prone to motion blur, and therefore lead to misaligned data pairs. We can also eliminate interference from external ambient sources of light other than the monitor's emitted light." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.623, + 0.788, + 0.73 + ], + "angle": 0, + "content": "The monitor, UDC panel, and camera sensor are aligned in sequence, and the principal axis of the camera is aligned perpendicular to the plane of the monitor and the panel. The camera is mounted on a sturdy tripod, and the panel is situated close to the sensor using a firm holding device. The displayed images are then captured remotely with/without the smartphone panels to ensure the alignment of clean/UDC image pairs during capture while keeping the camera settings the same. We provide more details in the Supplementary Material." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.75, + 0.788, + 0.842 + ], + "angle": 0, + "content": "Our UDC dataset The dataset is captured using 400 images from the DIV2K dataset [2], which is comprised of natural scene images from various themes, encompassing diverse objects captured under different lighting conditions. This makes our dataset 1,200 images in total, as there are three different image sets (Axon 30, Z-Fold 3, ground truth) in raw format. The images are stored in 16 bits with a Bayer pattern of G-B-R-G. For the ground truth images, we capture" + } + ], + [ + { + "type": "page_number", + "bbox": [ + 0.218, + 0.116, + 0.23, + 0.127 + ], + "angle": 0, + "content": "6" + }, + { + "type": "header", + "bbox": [ + 0.272, + 0.115, + 0.355, + 0.127 + ], + "angle": 0, + "content": "Y. Oh et al." + }, + { + "type": "image", + "bbox": [ + 0.283, + 0.15, + 0.415, + 0.252 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.283, + 0.253, + 0.415, + 0.355 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.292, + 0.355, + 0.399, + 0.368 + ], + "angle": 0, + "content": "(a) Clean image" + }, + { + "type": "image", + "bbox": [ + 0.436, + 0.15, + 0.569, + 0.252 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.436, + 0.253, + 0.568, + 0.355 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.436, + 0.355, + 0.562, + 0.368 + ], + "angle": 0, + "content": "(b) Axon 30 image" + }, + { + "type": "image", + "bbox": [ + 0.589, + 0.15, + 0.722, + 0.252 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.589, + 0.253, + 0.721, + 0.355 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.593, + 0.355, + 0.716, + 0.368 + ], + "angle": 0, + "content": "(c) Z-Fold 3 image" + }, + { + "type": "image_caption", + "bbox": [ + 0.214, + 0.382, + 0.788, + 0.466 + ], + "angle": 0, + "content": "Fig. 2: Cropped examples from our captured dataset. All images are visualized with the same procedure that includes normalization, demosaicing, channel gain adjustment, and gamma correction. Due to the difference in smartphone display panels, noise and haze are more apparent in Axon 30 UDC images, while the Z-Fold 3 UDC images show blurrier results. The diffraction induces flare artifacts when the image intensity becomes too strong that the sensor saturates." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.494, + 0.785, + 0.54 + ], + "angle": 0, + "content": "bursts of a scene without the display panel and average them. To check for misalignments that can be induced by optical image stabilization in the smartphone hardware [1], we utilize registration and feature matching algorithms [3, 12]." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.54, + 0.788, + 0.751 + ], + "angle": 0, + "content": "Samples from our dataset are displayed in Fig. 2, with each UDC image taken with Axon 30 and Z-Fold 3 display panels exhibiting panel-specific degradations. As such, our dataset can be used to evaluate if a restoration method generally performs well in the restoration of various UDC systems. Also, our real commercial UDC data show different behaviors from previous datasets, as shown in Fig. 1. Specifically, low-light, color shift, blur, and noise are observed, and an object is seen repeated faintly near the main object due to diffraction. We also detect a phenomenon similar to haze due to the scattering effects of the panel [30]. Synthetic datasets that are generated using UDC modeling [10, 30] do not take noise into account. However, including noise in UDC images is vital as the level of noise captured with frontal smartphone camera sensors, which are relatively inferior to DSLR camera sensors [4, 26], is amplified when normalizing sensor data. This amplification is related to the attenuation caused by the panel which reduces light intensity. More samples are available in the Supplementary." + }, + { + "type": "title", + "bbox": [ + 0.215, + 0.772, + 0.588, + 0.786 + ], + "angle": 0, + "content": "3.2 Proposed UDC Restoration Framework" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.795, + 0.787, + 0.84 + ], + "angle": 0, + "content": "In this section, we propose our restoration framework called Degradation Representation Embedded Transformer for UDC image restoration (DREUDC), which consists of a degradation representation encoder and a restoration network. We" + } + ], + [ + { + "type": "header", + "bbox": [ + 0.472, + 0.115, + 0.732, + 0.13 + ], + "angle": 0, + "content": "Panel-Specific UDC Image Restoration" + }, + { + "type": "page_number", + "bbox": [ + 0.775, + 0.117, + 0.785, + 0.127 + ], + "angle": 0, + "content": "7" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.147, + 0.788, + 0.239 + ], + "angle": 0, + "content": "first train the encoder with our proposed unsupervised scheme using triplet loss to implicitly learn the degradations of a display panel. Then, we utilize the representations generated from the encoder by integrating them in the restoration network as a prior. It is worth noting that our approach's novelty lies in how we take advantage of UDC data from one display panel to restore a UDC image from another panel rather than in the restoration network's architecture itself." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.259, + 0.79, + 0.442 + ], + "angle": 0, + "content": "Motivation Representation learning has been demonstrated to be effective in blind super-resolution, as demonstrated in [32]. Inspired by this, we explore the use of representation learning to enhance the restoration of UDC images by extracting useful prior information about the panels. Since each UDC system has a different panel, it yields panel-specific degradations. For example, the UDC images presented in Fig. 2, which are captured using two different panels of UDC smartphones, show differences in the degree of degradations. This implies that the degradations present in an image captured using one panel are similar to each other and that they differ significantly from those of the image with the same content but are captured using another panel. Using this information, our approach intends to utilize these differences and train an encoder to encode them in an unsupervised manner into the embedding space." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.463, + 0.788, + 0.598 + ], + "angle": 0, + "content": "Unsupervised representation encoder training Our proposed unsupervised training scheme is visualized in Fig. 3. As mentioned above, the objective of training the panel-specific degradation representation encoder is to learn the unique properties of the panels and produce implicit representations of the degradations to be used as priors in UDC image restoration. To achieve this, the encoder should learn the degradations using attributes related to a specific panel, such as the level of noise, blur, and color shift. On the other hand, the encoder should be discouraged from learning degradations with features in an image that are panel-agnostic, such as textures and contents of a scene." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.599, + 0.788, + 0.794 + ], + "angle": 0, + "content": "We believe that this condition is analogous to the situation of tasks where triplet loss [28] is exploited to learn favorable representations. Triplet loss is widely applied to various tasks of computer vision such as facial recognition [28], person re-identification [13,34], and domain adaptation [8] for its ability to learn useful embeddings by leveraging the distances between embeddings of similar instances and dissimilar instances. For instance, in person re-identification, networks are guided to identify among different people using features that are unique to a specific person, not with features that are shared among people, by learning suitable representations using triplet loss. In our case, the features unique to a specific person would correspond to properties unique to a specific panel, and the features that are shared among people would correspond to panel-agnostic features. Therefore, we opt to apply triplet loss in the training of our encoder to capture the panel-specific degradation representations." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.796, + 0.788, + 0.841 + ], + "angle": 0, + "content": "Nevertheless, triplet loss is only effective when the network is provided with valid triplets [28, 33]. To collect meaningful triplets, we impose constraints on the training pairs that will be used to train our encoder. We choose a pair of" + } + ], + [ + { + "type": "page_number", + "bbox": [ + 0.218, + 0.116, + 0.23, + 0.127 + ], + "angle": 0, + "content": "8" + }, + { + "type": "header", + "bbox": [ + 0.272, + 0.115, + 0.355, + 0.127 + ], + "angle": 0, + "content": "Y. Oh et al." + }, + { + "type": "image", + "bbox": [ + 0.237, + 0.152, + 0.353, + 0.251 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.289, + 0.253, + 0.302, + 0.261 + ], + "angle": 0, + "content": "\\(x_{i}\\)" + }, + { + "type": "image", + "bbox": [ + 0.237, + 0.264, + 0.352, + 0.361 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.256, + 0.362, + 0.337, + 0.39 + ], + "angle": 0, + "content": "\\(x_{j}\\) Raw UDC images (visualized for figure)" + }, + { + "type": "image", + "bbox": [ + 0.368, + 0.151, + 0.775, + 0.372 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.368, + 0.373, + 0.435, + 0.385 + ], + "angle": 0, + "content": "Training patches" + }, + { + "type": "image_caption", + "bbox": [ + 0.513, + 0.371, + 0.619, + 0.386 + ], + "angle": 0, + "content": "Panel-specific degradation representation" + }, + { + "type": "image_caption", + "bbox": [ + 0.65, + 0.369, + 0.752, + 0.377 + ], + "angle": 0, + "content": "Projected representations" + }, + { + "type": "image_caption", + "bbox": [ + 0.215, + 0.407, + 0.789, + 0.464 + ], + "angle": 0, + "content": "Fig. 3: Our proposed unsupervised panel-specific degradation representation learning scheme. By using triplet loss, representations of the same panel-specific degradations with different image content are pulled closer together, while representations of different panel-specific degradations with the same image content are pushed away." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.49, + 0.789, + 0.581 + ], + "angle": 0, + "content": "raw images \\( x_{i} \\) and \\( x_{j} \\) from our dataset that contains the same content but have been taken with different display panels. Then, we randomly crop two pairs of patches \\( (x_{i}^{1}, x_{i}^{2}) \\) and \\( (x_{j}^{1}, x_{j}^{2}) \\) from each image, where \\( x_{i}^{1} \\) and \\( x_{j}^{1} \\), and \\( x_{i}^{2} \\) and \\( x_{j}^{2} \\), respectively, are cropped from the same location of \\( x_{i} \\) and \\( x_{j} \\). We choose patch \\( x_{i}^{1} \\) as the anchor, and the positive and negative samples are decided to be \\( x_{i}^{2} \\) and \\( x_{j}^{1} \\). The same process is repeated with \\( x_{j}^{1} \\) as the anchor." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.581, + 0.789, + 0.702 + ], + "angle": 0, + "content": "With this triplet selection, the encoder is guided to learn the intrinsic properties of the degradations caused by the display panel by pulling representation of images that show similar degradation while disregarding the features that are unrelated to degradations by pushing the representation of the same contents. The encoder \\( f \\) is trained with triplet loss using the selected triplets to obtain the panel-specific degradation representation \\( y \\in \\mathbb{R}^{256} \\). The representations are further projected to a projection \\( z \\) with two fully connected layers, following [5, 6]. The triplet loss for training the encoder is as follows:" + }, + { + "type": "equation", + "bbox": [ + 0.232, + 0.712, + 0.786, + 0.754 + ], + "angle": 0, + "content": "\\[\n\\begin{array}{l} L = \\max \\left(\\left\\| z _ {i} ^ {1} - z _ {i} ^ {2} \\right\\| _ {2} ^ {2} - \\left\\| z _ {i} ^ {1} - z _ {j} ^ {1} \\right\\| _ {2} ^ {2} + \\alpha , 0\\right) \\\\ + m a x (| | z _ {j} ^ {1} - z _ {j} ^ {2} | | _ {2} ^ {2} - | | z _ {j} ^ {1} - z _ {i} ^ {1} | | _ {2} ^ {2} + \\alpha , 0), \\quad (1) \\\\ \\end{array}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.215, + 0.759, + 0.788, + 0.791 + ], + "angle": 0, + "content": "where \\(\\alpha\\) is the margin of the triplet loss. The margin is set as \\(\\alpha = 1\\). Following [32], the encoder is a simple network with six convolutional layers." + }, + { + "type": "text", + "bbox": [ + 0.215, + 0.81, + 0.788, + 0.84 + ], + "angle": 0, + "content": "Representation embedded restoration network For our restoration network, we design it to be as simple and efficient as possible to emphasize the ef-" + } + ], + [ + { + "type": "header", + "bbox": [ + 0.472, + 0.115, + 0.732, + 0.129 + ], + "angle": 0, + "content": "Panel-Specific UDC Image Restoration" + }, + { + "type": "page_number", + "bbox": [ + 0.775, + 0.116, + 0.787, + 0.127 + ], + "angle": 0, + "content": "9" + }, + { + "type": "image", + "bbox": [ + 0.241, + 0.147, + 0.761, + 0.306 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.344, + 0.307, + 0.61, + 0.321 + ], + "angle": 0, + "content": "(a) Our proposed framework, DREUDC" + }, + { + "type": "image_caption", + "bbox": [ + 0.666, + 0.307, + 0.761, + 0.321 + ], + "angle": 0, + "content": "(b) DREblock" + }, + { + "type": "image_caption", + "bbox": [ + 0.215, + 0.335, + 0.788, + 0.378 + ], + "angle": 0, + "content": "Fig. 4: Our proposed DREUDC. The framework consists of an encoder that produces an implicit representation of the panel-specific degradations of a UDC image and a restoration network that exploits the representations via the proposed DREblocks." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.408, + 0.789, + 0.544 + ], + "angle": 0, + "content": "fectiveness of our contribution of how we leverage the difference in data collected with various panels to enhance the restoration performance rather than focusing on the impact of complex restoration network architectures. Therefore, we adopt a U-shaped [27] Transformer with Frequency domain-based Self-Attention Solver (FSAS) [16] as our restoration architecture. In previous works on UDC image restoration [9, 10, 17, 31], researchers have highlighted the importance of a wide receptive field. As FSAS utilizes the convolutional theorem of Fourier transform to model global contexts in the frequency domain efficiently [16], we choose it as a component of the building block for our restoration network." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.545, + 0.788, + 0.621 + ], + "angle": 0, + "content": "Our restoration network is based on stacking repeated Transformer blocks, which we name DREblock (Degradation Representation Embedded Transformer block, Fig. 4(b)). The proposed DREblock efficiently integrates the degradation representation \\( y \\) of the input UDC image as a condition of our restoration network by calibrating the features in the block in a channel-wise manner, i.e.," + }, + { + "type": "equation", + "bbox": [ + 0.447, + 0.635, + 0.786, + 0.652 + ], + "angle": 0, + "content": "\\[\nF _ {o} = F _ {a t t} \\otimes \\beta , \\tag {2}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.659, + 0.788, + 0.735 + ], + "angle": 0, + "content": "where \\( F_{att} \\) is the feature maps of the estimated attention of FSAS, \\( F_{o} \\) is the feature after the conditioning, \\( \\otimes \\) denotes channel-wise multiplication, and \\( \\beta \\) is the channel conditioning coefficients that are produced by reshaping the panel-specific degradation representation \\( y \\) using a single layer of \\( 1 \\times 1 \\) convolution. The features \\( F_{o} \\) are then processed by the point-wise feed-forward network." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.735, + 0.788, + 0.841 + ], + "angle": 0, + "content": "The entire framework DREUDC is illustrated in Fig. 4. In training, the encoder is first trained with Eq. (1). Then, the encoder is frozen while the restoration network is trained with \\(\\mathcal{L}_1\\) loss. During inference, the encoder takes the input raw UDC image and outputs a panel-specific degradation representation vector that contains prior information on UDC degradation. The representation is seamlessly embedded into the restoration network through DREblocks, finally producing a clean restored RGB image." + } + ], + [ + { + "type": "page_number", + "bbox": [ + 0.218, + 0.116, + 0.236, + 0.127 + ], + "angle": 0, + "content": "10" + }, + { + "type": "header", + "bbox": [ + 0.272, + 0.115, + 0.355, + 0.128 + ], + "angle": 0, + "content": "Y. Oh et al." + }, + { + "type": "title", + "bbox": [ + 0.217, + 0.145, + 0.498, + 0.163 + ], + "angle": 0, + "content": "4 Experiment and Analysis" + }, + { + "type": "title", + "bbox": [ + 0.217, + 0.177, + 0.334, + 0.19 + ], + "angle": 0, + "content": "4.1 Datasets" + }, + { + "type": "text", + "bbox": [ + 0.216, + 0.202, + 0.789, + 0.399 + ], + "angle": 0, + "content": "We evaluate the restoration performance using our captured data with display panels of ZTE Axon 30 5G and Samsung Galaxy Z-Fold 3. The raw images are normalized to have values within the range of [0,1]. To train and evaluate RGB-based methods, we use a simplified ISP algorithm written with MATLAB to visualize the normalized raw images to RGB images. The algorithm contains demosaicing, channel-wise gain control, and gamma correction in sequence. Note that the visualized images are not sRGB images, as actual camera ISPs in practice involve additional steps to render the image in sRGB domain. Please check the Supplementary for more discussion on the usage of simplified ISP. For both display panels, we split the dataset into 300 images for training and 100 images for testing. The training set is cropped into patches of \\(512 \\times 512\\). We also use previous UDC datasets [38], which consists of 300 images of both T-OLED and P-OLED scenes, to verify that our method is applicable to other real datasets." + }, + { + "type": "title", + "bbox": [ + 0.217, + 0.421, + 0.569, + 0.436 + ], + "angle": 0, + "content": "4.2 Training and Implementation Details" + }, + { + "type": "text", + "bbox": [ + 0.216, + 0.445, + 0.789, + 0.613 + ], + "angle": 0, + "content": "We use NVIDIA RTX 3090 GPUs in our experiments. To train the encoder, we use Adam optimizer \\((\\beta_{1} = 0.9, \\beta_{2} = 0.999)\\) for \\(1 \\times 10^{5}\\) iterations with a batch size of 64, and the loss is given as Eq. (1). The restoration networks are trained with \\(\\mathcal{L}_1\\) loss to minimize the distance between the restored image of the network and the ground truth. The initial number of channels \\(C\\) of our restoration network is 32, and the number of blocks in each layer \\(L_1, L_2\\), and \\(L_3\\) are [5,5,6]. We use Adam optimizer \\((\\beta_{1} = 0.9, \\beta_{2} = 0.999)\\) for \\(2 \\times 10^{5}\\) iterations with a batch size of 2. For both networks, the initial learning rate is set as \\(2 \\times 10^{-4}\\) and gradually decayed to \\(1 \\times 10^{-6}\\) with cosine annealing [20]. This setting is applied to all experiments on our dataset, and every previous method reported in this paper is also reproduced by training on our dataset with the same setting for fairness." + }, + { + "type": "title", + "bbox": [ + 0.217, + 0.634, + 0.507, + 0.649 + ], + "angle": 0, + "content": "4.3 Raw UDC Image Restoration" + }, + { + "type": "text", + "bbox": [ + 0.216, + 0.659, + 0.789, + 0.841 + ], + "angle": 0, + "content": "First, we experiment to validate that using raw images can improve the UDC image restoration process. For this purpose, we employ a baseline neural network called DE-UNET [38] and use our captured data. We compare the performance of the network under three different settings: RAW-to-RAW, RAW-to-RGB, and RGB-to-RGB. We evaluate the performance of the network using the PSNR and SSIM metrics. To assess the performance of the RAW-to-RAW setting, we convert the output restored raw images to RGB using the same ISP process mentioned in Sec. 4.1 and measure them in the RGB domain. The results are presented in Tab. 2, showing that using raw sensor images as input (RAW-to-RAW, RAW-to-RGB) yields superior results compared to only using RGB images. Among the two settings that use raw images as input, the latter shows better performance overall, so we choose the method of converting and restoring" + } + ], + [ + { + "type": "header", + "bbox": [ + 0.472, + 0.115, + 0.732, + 0.129 + ], + "angle": 0, + "content": "Panel-Specific UDC Image Restoration" + }, + { + "type": "page_number", + "bbox": [ + 0.768, + 0.116, + 0.784, + 0.127 + ], + "angle": 0, + "content": "11" + }, + { + "type": "table_caption", + "bbox": [ + 0.216, + 0.145, + 0.785, + 0.173 + ], + "angle": 0, + "content": "Table 2: Comparison of different input-to-output settings of DE-UNET [38] on our captured datasets. The best results are bold-faced." + }, + { + "type": "table", + "bbox": [ + 0.218, + 0.185, + 0.784, + 0.258 + ], + "angle": 0, + "content": "
MethodAxon 30 (raw)Z-Fold 3 (raw)Axon 30 (rgb)Z-Fold 3 (rgb)
PSNRSSIMPSNRSSIMPSNRSSIMPSNRSSIM
DE-UNET(RGB-to-RGB)----23.330.910926.900.9390
DE-UNET(RAW-to-RAW)27.360.933529.660.963724.210.917027.180.9388
DE-UNET(RAW-to-RGB)----24.200.919627.290.9419
" + }, + { + "type": "table_caption", + "bbox": [ + 0.216, + 0.271, + 0.785, + 0.311 + ], + "angle": 0, + "content": "Table 3: Results on our captured datasets for ZTE Axon 30 5G and Samsung Galaxy Z-Fold 3 smartphone display panels on various methods of UDC image restoration. The best results are bold-faced." + }, + { + "type": "table", + "bbox": [ + 0.218, + 0.326, + 0.784, + 0.538 + ], + "angle": 0, + "content": "
Input-to-outputMethodParamsMACsAxon 30 (rgb)Z-Fold 3 (rgb)
PSNRSSIMPSNRSSIM
RGB-to-RGBDE-UNET [38]8.93M273.84G23.330.910926.900.9390
DAGF [31]1.09M73.84G24.490.915028.420.9404
DISCNET [10]3.80M596.95G23.110.906326.670.9361
BNUDC [15]4.57M993.38G25.850.928328.540.9458
DWFormer [36]1.45M216.77G21.520.892326.360.9338
SRUDC [30]6.70M371.32G22.670.899827.340.9336
PPM-UNET [9]4.23M409.29G18.210.850223.850.9188
AWNet [7]47.00M1.53T20.270.883925.600.9316
RAW-to-RGBDE-UNET [38]8.94M69.29G24.200.919627.290.9419
PyNet [14]47.56M1.79T22.580.906927.350.9377
AWNet [7]49.07M392.25G20.980.888527.250.9369
DREUDC6.44M200.07G26.660.936829.040.9511
" + }, + { + "type": "text", + "bbox": [ + 0.216, + 0.567, + 0.785, + 0.613 + ], + "angle": 0, + "content": "raw UDC images to a clean RGB image (RAW-to-RGB). This is also the setting of most neural ISPs currently being used in smartphones [26] since a successful network is able to model the image processing pipeline as well." + }, + { + "type": "title", + "bbox": [ + 0.216, + 0.635, + 0.484, + 0.648 + ], + "angle": 0, + "content": "4.4 Evaluation on Our Dataset" + }, + { + "type": "text", + "bbox": [ + 0.216, + 0.659, + 0.787, + 0.749 + ], + "angle": 0, + "content": "To provide experimental results on our dataset and demonstrate the performance of DREUDC on UDC image restoration, we train numerous previous methods using our dataset for comparison. The quality of the restoration performance is evaluated with PSNR and SSIM as performance metrics. Additionally, we assess the efficiency of the methods by comparing the number of parameters and MACs. To calculate the MACs, we use a dummy image with a resolution of \\(1024 \\times 1024\\)." + }, + { + "type": "text", + "bbox": [ + 0.216, + 0.75, + 0.787, + 0.84 + ], + "angle": 0, + "content": "Tab. 3 shows experimental results for raw and RGB image restoration using our dataset. Because there are no neural networks on UDC image restoration that take raw sensor images as input except [38] as previous datasets were provided only in the RGB domain, we choose [7] and [14] for further comparison, which are RAW-to-RGB image conversion models. We observe that DREUDC achieves state-of-the-art performance on both Axon 30 and Z-fold 3 datasets," + } + ], + [ + { + "type": "page_number", + "bbox": [ + 0.218, + 0.116, + 0.236, + 0.127 + ], + "angle": 0, + "content": "12" + }, + { + "type": "header", + "bbox": [ + 0.272, + 0.115, + 0.355, + 0.128 + ], + "angle": 0, + "content": "Y. Oh et al." + }, + { + "type": "image", + "bbox": [ + 0.231, + 0.148, + 0.303, + 0.204 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.24, + 0.204, + 0.298, + 0.214 + ], + "angle": 0, + "content": "Axon 30" + }, + { + "type": "image", + "bbox": [ + 0.308, + 0.148, + 0.383, + 0.204 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.31, + 0.204, + 0.381, + 0.214 + ], + "angle": 0, + "content": "DE-UNET" + }, + { + "type": "image", + "bbox": [ + 0.388, + 0.148, + 0.462, + 0.204 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.404, + 0.204, + 0.446, + 0.214 + ], + "angle": 0, + "content": "DAGF" + }, + { + "type": "image", + "bbox": [ + 0.468, + 0.148, + 0.539, + 0.204 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.469, + 0.204, + 0.538, + 0.214 + ], + "angle": 0, + "content": "DISCNET" + }, + { + "type": "image", + "bbox": [ + 0.545, + 0.148, + 0.618, + 0.204 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.551, + 0.204, + 0.611, + 0.214 + ], + "angle": 0, + "content": "BNUDC" + }, + { + "type": "image", + "bbox": [ + 0.625, + 0.148, + 0.697, + 0.204 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.625, + 0.204, + 0.697, + 0.214 + ], + "angle": 0, + "content": "DWFormer" + }, + { + "type": "image", + "bbox": [ + 0.704, + 0.148, + 0.776, + 0.204 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.712, + 0.204, + 0.767, + 0.213 + ], + "angle": 0, + "content": "SRUDC" + }, + { + "type": "image", + "bbox": [ + 0.231, + 0.214, + 0.303, + 0.268 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.23, + 0.268, + 0.303, + 0.279 + ], + "angle": 0, + "content": "PPM-UNet" + }, + { + "type": "image", + "bbox": [ + 0.309, + 0.214, + 0.383, + 0.268 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.31, + 0.268, + 0.371, + 0.279 + ], + "angle": 0, + "content": "AWNet" + }, + { + "type": "image", + "bbox": [ + 0.388, + 0.214, + 0.462, + 0.268 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.387, + 0.268, + 0.461, + 0.279 + ], + "angle": 0, + "content": "DE-UNET*" + }, + { + "type": "image", + "bbox": [ + 0.468, + 0.214, + 0.539, + 0.268 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.47, + 0.268, + 0.528, + 0.278 + ], + "angle": 0, + "content": "PyNet*" + }, + { + "type": "image", + "bbox": [ + 0.546, + 0.214, + 0.618, + 0.268 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.551, + 0.268, + 0.61, + 0.279 + ], + "angle": 0, + "content": "AWNet*" + }, + { + "type": "image", + "bbox": [ + 0.625, + 0.214, + 0.697, + 0.268 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.625, + 0.268, + 0.697, + 0.279 + ], + "angle": 0, + "content": "DREUDC*" + }, + { + "type": "image", + "bbox": [ + 0.704, + 0.214, + 0.776, + 0.268 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.71, + 0.268, + 0.751, + 0.278 + ], + "angle": 0, + "content": "GT" + }, + { + "type": "image", + "bbox": [ + 0.231, + 0.279, + 0.303, + 0.333 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.24, + 0.333, + 0.298, + 0.342 + ], + "angle": 0, + "content": "Z-Fold 3" + }, + { + "type": "image", + "bbox": [ + 0.309, + 0.279, + 0.383, + 0.333 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.31, + 0.333, + 0.379, + 0.342 + ], + "angle": 0, + "content": "DE-UNET" + }, + { + "type": "image", + "bbox": [ + 0.388, + 0.279, + 0.462, + 0.333 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.399, + 0.333, + 0.446, + 0.342 + ], + "angle": 0, + "content": "DAGE" + }, + { + "type": "image", + "bbox": [ + 0.468, + 0.279, + 0.539, + 0.333 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.469, + 0.333, + 0.537, + 0.342 + ], + "angle": 0, + "content": "DISCNET" + }, + { + "type": "image", + "bbox": [ + 0.546, + 0.279, + 0.618, + 0.333 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.551, + 0.333, + 0.611, + 0.342 + ], + "angle": 0, + "content": "BNUDC" + }, + { + "type": "image", + "bbox": [ + 0.625, + 0.279, + 0.697, + 0.333 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.625, + 0.333, + 0.697, + 0.342 + ], + "angle": 0, + "content": "DWFormer" + }, + { + "type": "image", + "bbox": [ + 0.704, + 0.279, + 0.776, + 0.333 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.71, + 0.333, + 0.767, + 0.342 + ], + "angle": 0, + "content": "SRUDC" + }, + { + "type": "image", + "bbox": [ + 0.231, + 0.342, + 0.303, + 0.397 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.23, + 0.397, + 0.304, + 0.408 + ], + "angle": 0, + "content": "PPM-UNet" + }, + { + "type": "image", + "bbox": [ + 0.309, + 0.342, + 0.383, + 0.397 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.321, + 0.397, + 0.371, + 0.408 + ], + "angle": 0, + "content": "AWNet" + }, + { + "type": "image", + "bbox": [ + 0.388, + 0.342, + 0.462, + 0.397 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.386, + 0.397, + 0.461, + 0.408 + ], + "angle": 0, + "content": "DE-UNET*" + }, + { + "type": "image", + "bbox": [ + 0.468, + 0.342, + 0.539, + 0.397 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.477, + 0.397, + 0.527, + 0.408 + ], + "angle": 0, + "content": "PyNet*" + }, + { + "type": "image", + "bbox": [ + 0.546, + 0.342, + 0.618, + 0.397 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.551, + 0.397, + 0.611, + 0.408 + ], + "angle": 0, + "content": "AWNet*" + }, + { + "type": "image", + "bbox": [ + 0.625, + 0.342, + 0.697, + 0.397 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.624, + 0.397, + 0.696, + 0.408 + ], + "angle": 0, + "content": "DREUDC*" + }, + { + "type": "image", + "bbox": [ + 0.704, + 0.342, + 0.776, + 0.397 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.728, + 0.397, + 0.751, + 0.407 + ], + "angle": 0, + "content": "GT" + }, + { + "type": "image_caption", + "bbox": [ + 0.215, + 0.423, + 0.788, + 0.493 + ], + "angle": 0, + "content": "Fig. 5: Visual comparison of restoration examples for our dataset. Zoom in for a better comparison between the methods. The top two rows are results for Axon 30 images, and the bottom two rows are for Z-Fold 3 images. Methods that take raw input and produce RGB images are annotated with the asterisk *. Our method is capable of removing UDC degradations more clearly than previous UDC image restoration networks." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.522, + 0.788, + 0.629 + ], + "angle": 0, + "content": "surpassing BNUDC [15] by 0.81dB and 0.50dB, demonstrating that our method can successfully restore various UDC degradations. With the exception of a lightweight model [31] and the baseline model using raw data input [38], our method is the most efficient when it comes to computation while showing superior performance. Qualitative results are illustrated in Fig. 5, and more results including restoration of scenes that are not displayed on a monitor are available in the Supplementary Material." + }, + { + "type": "title", + "bbox": [ + 0.215, + 0.65, + 0.717, + 0.667 + ], + "angle": 0, + "content": "4.5 Analysis on Panel-Specific Degradation Representation" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.674, + 0.788, + 0.795 + ], + "angle": 0, + "content": "In this subsection, we analyze the effectiveness of DREUDC and panel-specific degradation representations learned with our proposed method. The strength of DREUDC is reported in Tab. 4 with an ablation study. We observe that our strategy of embedding an implicit representation of degraded UDC images is indeed effective by comparing (a) and (c) of Tab. 4, where Model 1 is our framework with only the restoration network without channel calibration and representation learning, while DREUDC is our entire framework trained with the protocol explained in Sec. 3.2." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.796, + 0.788, + 0.842 + ], + "angle": 0, + "content": "However, with only this comparison, we cannot rule out the possibility that the rise in performance is solely because of increased parameters. Therefore, we further compare (a) and (c) with (b), where Model 2 is our framework trained in" + } + ], + [ + { + "type": "header", + "bbox": [ + 0.472, + 0.115, + 0.732, + 0.129 + ], + "angle": 0, + "content": "Panel-Specific UDC Image Restoration" + }, + { + "type": "page_number", + "bbox": [ + 0.769, + 0.116, + 0.786, + 0.127 + ], + "angle": 0, + "content": "13" + }, + { + "type": "table_caption", + "bbox": [ + 0.351, + 0.145, + 0.651, + 0.159 + ], + "angle": 0, + "content": "Table 4: Results on ablation of our method." + }, + { + "type": "table", + "bbox": [ + 0.223, + 0.171, + 0.778, + 0.244 + ], + "angle": 0, + "content": "
MethodChannel calibrationRepresentation learningParams MACsAxon 30Z-Fold 3
PNSRSSIMPNSRSSIM
(a) Model 1XX4.52M160.57G26.090.933828.66
(b) Model 2✓X6.44M200.07G26.380.934828.71
(c) DREUDC✓✓6.44M200.07G26.660.936829.04
" + }, + { + "type": "image", + "bbox": [ + 0.253, + 0.284, + 0.337, + 0.341 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.288, + 0.341, + 0.304, + 0.352 + ], + "angle": 0, + "content": "(a)" + }, + { + "type": "image", + "bbox": [ + 0.343, + 0.284, + 0.419, + 0.341 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.371, + 0.341, + 0.385, + 0.352 + ], + "angle": 0, + "content": "(b)" + }, + { + "type": "image", + "bbox": [ + 0.424, + 0.285, + 0.501, + 0.341 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.456, + 0.341, + 0.471, + 0.352 + ], + "angle": 0, + "content": "(c)" + }, + { + "type": "image", + "bbox": [ + 0.506, + 0.285, + 0.582, + 0.341 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.538, + 0.341, + 0.553, + 0.352 + ], + "angle": 0, + "content": "(d)" + }, + { + "type": "image", + "bbox": [ + 0.591, + 0.285, + 0.666, + 0.34 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.621, + 0.341, + 0.635, + 0.352 + ], + "angle": 0, + "content": "(e)" + }, + { + "type": "image", + "bbox": [ + 0.671, + 0.285, + 0.748, + 0.34 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.704, + 0.341, + 0.717, + 0.352 + ], + "angle": 0, + "content": "(f)" + }, + { + "type": "image_caption", + "bbox": [ + 0.215, + 0.369, + 0.788, + 0.468 + ], + "angle": 0, + "content": "Fig. 6: T-SNE [21] visualization of representation generated from the encoder. (a) and (b) show the representations generated from encoders of Model 2 in Tab. 4 for Axon 30 and Z-fold 3 images, respectively, which are trained without representation learning. (d) and (e) show the representations generated from encoders of Model 2 in Tab. 5 for T-OLLED and P-OLED images, respectively, which are trained without representation learning. (c) and (f) illustrate the representations that are generated with our unsupervised triplet learning on our dataset and Zhou's dataset [38]." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.51, + 0.788, + 0.661 + ], + "angle": 0, + "content": "an end-to-end manner by training the encoder and the restoration network jointly with \\(\\mathcal{L}_1\\) loss without our proposed representation learning. By comparing Model 1 and Model 2, we see that channel calibration with representations learned from a single dataset without triplet training increases the performance marginally. On the contrary, when given an appropriate representation learned from two datasets by leveraging the difference in degradations with our proposed method, we notice that the restoration model benefits substantially from it. Moreover, it is noteworthy that our proposed embedding scheme has a negligible effect on the computational increase (MACs) of the restoration network, as most of the increase is due to the encoder." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.668, + 0.788, + 0.744 + ], + "angle": 0, + "content": "The validity of our panel-specific degradation representation learning scheme is visualized in Fig. 6. Without representation learning (Fig. 6(a) and (b)), the encoders are unable to distinguish different degradations. Conversely, with representation learning (Fig. 6(c)), the encoder identifies the degradation and discriminates them in clusters, offering advantageous priors for the restoration network." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.75, + 0.788, + 0.841 + ], + "angle": 0, + "content": "This finding suggests an approach to train a restoration network by making good use of datasets other than the one gathered for a specific UDC system. Typically, a specific dataset is used to train a restoration network tailored for a particular UDC system. However, this dataset would be sub-optimal or ineffective when future UDC smartphones have different panels. Our method offers a way to utilize other datasets collected with different panels, mediating this issue." + } + ], + [ + { + "type": "page_number", + "bbox": [ + 0.218, + 0.116, + 0.236, + 0.127 + ], + "angle": 0, + "content": "14" + }, + { + "type": "header", + "bbox": [ + 0.272, + 0.115, + 0.355, + 0.128 + ], + "angle": 0, + "content": "Y. Oh et al." + }, + { + "type": "table_caption", + "bbox": [ + 0.227, + 0.144, + 0.776, + 0.16 + ], + "angle": 0, + "content": "Table 5: Results of our method and ablation on T-OLED/P-OLED datasets [38]." + }, + { + "type": "table", + "bbox": [ + 0.222, + 0.171, + 0.778, + 0.257 + ], + "angle": 0, + "content": "
MethodChannel calibrationRepresentation learningParams MACsT-OLEDP-OLED
PSNR SSIMPSNR SSIM
DE-UNET [38]--8.94M 69.29G36.70 0.974528.58 0.9344
Model 1XX4.52M 160.57G37.37 0.977430.85 0.9510
Model 2✓X6.44M 200.07G37.46 0.977731.01 0.9515
DREUDC✓✓6.44M 200.07G37.55 0.977831.17 0.9521
" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.286, + 0.787, + 0.497 + ], + "angle": 0, + "content": "Application to previous datasets Our restoration method using panel-specific degradation representations is also applied to public UDC datasets to show that it generalizes well to other datasets. Our method requires a real dataset comprised of degraded images taken with more than a single panel from the same capturing environment, so we choose Zhou's P-OLED/T-OLED datasets [38]. DREUDC is designed to take raw images as input, so we make a pseudo-raw version of the dataset by mosaicing the RGB images. This pseudo-raw version of the RGB images that used to train the methods, including a new encoder, is created by reversing the in-camera pipeline mentioned in [38]. We only reversed the demosaicing step as it was the only mentioned process; hence we call it 'pseudo-raw' as the images have not been fully reversed to the original high-bit raw data. We choose DE-UNET [38] to compare with our results because it is a model that can be trained with RAW-to-RGB setting. Additional results on RGB image restoration are provided in the Supplementary Material." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.498, + 0.788, + 0.648 + ], + "angle": 0, + "content": "Tab. 5 shows the results on T-OLED/P-OLED datasets. Our method outperforms DE-UNET on both panels with slightly fewer parameters. Most importantly, our strategy of representation embedding is also effective on [38] when we compare Model 1 to DREUDC, showing an increase of \\(0.18\\mathrm{dB}\\) and \\(0.32\\mathrm{dB}\\) in the restoration of T-OLED and P-OLED images. We also prove that the boost in performance is not only due to increased parameters but also because we enhance the restoration network with effective representations of the degradations learned using our training scheme by comparing Model 2 and DREUDC. Visualization of the representations are provided in Fig. 6(d), (e), and (f). Overall, the results imply that our method is also effective on other UDC datasets." + }, + { + "type": "title", + "bbox": [ + 0.216, + 0.672, + 0.36, + 0.687 + ], + "angle": 0, + "content": "5 Conclusion" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.704, + 0.788, + 0.841 + ], + "angle": 0, + "content": "We have collected a new UDC dataset using commercial UDC-equipped smartphones and demonstrated that UDC image restoration using raw sensor images achieves better performance than using RGB images. We have also proposed a UDC image restoration framework that generates panel-specific degradation representations in the embedding space with a novel unsupervised representation learning scheme. Our restoration network efficiently accommodates the learned representations as a prior with our proposed Transformer blocks, improving the performance. Experiments show that our method achieves state-of-the-art performance on our dataset and is generalizable to previous datasets." + } + ], + [ + { + "type": "header", + "bbox": [ + 0.472, + 0.115, + 0.732, + 0.129 + ], + "angle": 0, + "content": "Panel-Specific UDC Image Restoration" + }, + { + "type": "page_number", + "bbox": [ + 0.769, + 0.117, + 0.785, + 0.127 + ], + "angle": 0, + "content": "15" + }, + { + "type": "title", + "bbox": [ + 0.218, + 0.145, + 0.383, + 0.163 + ], + "angle": 0, + "content": "Acknowledgment" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.175, + 0.788, + 0.283 + ], + "angle": 0, + "content": "This paper was result of the research project supported by SK hynix Inc., including the support of camera equipments and lab environment for collecting the dataset. It was also partially supported by Institute of Information & communications Technology Planning & Evaluation (IITP) grant funded by the Korea government (MSIT) (No.2021-0-01062), and in part by the BK21 FOUR program of the Education and Research Program for Future ICT Pioneers, Seoul National University in 2024." + }, + { + "type": "title", + "bbox": [ + 0.218, + 0.303, + 0.323, + 0.318 + ], + "angle": 0, + "content": "References" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.332, + 0.785, + 0.374 + ], + "angle": 0, + "content": "1. Abdelhamed, A., Lin, S., Brown, M.S.: A high-quality denoising dataset for smartphone cameras. In: Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition. pp. 1692-1700 (2018)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.374, + 0.785, + 0.415 + ], + "angle": 0, + "content": "2. Agustsson, E., Timofte, R.: Ntire 2017 challenge on single image super-resolution: Dataset and study. In: Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition Workshops. pp. 126-135 (2017)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.415, + 0.785, + 0.457 + ], + "angle": 0, + "content": "3. Bay, H., Tuytelaars, T., Van Gool, L.: Surf: Speeded up robust features. In: Computer Vision-ECCV 2006: 9th European Conference on Computer Vision, Graz, Austria, May 7-13, 2006. Proceedings, Part I 9. pp. 404-417. Springer (2006)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.457, + 0.785, + 0.497 + ], + "angle": 0, + "content": "4. Chen, C., Chen, Q., Xu, J., Koltun, V.: Learning to see in the dark. In: Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition. pp. 3291-3300 (2018)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.497, + 0.785, + 0.538 + ], + "angle": 0, + "content": "5. Chen, T., Kornblith, S., Norouzi, M., Hinton, G.: A simple framework for contrastive learning of visual representations. In: International Conference on Machine Learning. pp. 1597-1607. PMLR (2020)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.539, + 0.785, + 0.565 + ], + "angle": 0, + "content": "6. Chen, X., Fan, H., Girshick, R., He, K.: Improved baselines with momentum contrastive learning. arXiv preprint arXiv:2003.04297 (2020)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.566, + 0.785, + 0.607 + ], + "angle": 0, + "content": "7. Dai, L., Liu, X., Li, C., Chen, J.: Attnective wavelet network for image isp. In: Computer Vision-ECCV 2020 Workshops: Glasgow, UK, August 23-28, 2020, Proceedings, Part III 16. pp. 185-201. Springer (2020)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.608, + 0.785, + 0.648 + ], + "angle": 0, + "content": "8. Deng, W., Zheng, L., Sun, Y., Jiao, J.: Rethinking triplet loss for domain adaptation. IEEE Transactions on Circuits and Systems for Video Technology 31(1), 29-37 (2020)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.649, + 0.785, + 0.703 + ], + "angle": 0, + "content": "9. Feng, R., Li, C., Chen, H., Li, S., Gu, J., Loy, C.C.: Generating aligned pseudosupervision from non-aligned data for image restoration in under-display camera. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 5013-5022 (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.22, + 0.703, + 0.785, + 0.758 + ], + "angle": 0, + "content": "10. Feng, R., Li, C., Chen, H., Li, S., Loy, C.C., Gu, J.: Removing diffraction image artifacts in under-display camera via dynamic skip connection network. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 662-671 (2021)" + }, + { + "type": "ref_text", + "bbox": [ + 0.22, + 0.759, + 0.785, + 0.813 + ], + "angle": 0, + "content": "11. Feng, R., Li, C., Zhou, S., Sun, W., Zhu, Q., Jiang, J., Yang, Q., Loy, C.C., Gu, J., Zhu, Y., et al.: Mipi 2022 challenge on under-display camera image restoration: Methods and results. In: European Conference on Computer Vision. pp. 60-77. Springer (2022)" + }, + { + "type": "ref_text", + "bbox": [ + 0.22, + 0.813, + 0.785, + 0.84 + ], + "angle": 0, + "content": "12. Guizar-Sicairos, M., Thurman, S.T., Fienup, J.R.: Efficient subpixel image registration algorithms. Optics letters 33(2), 156-158 (2008)" + }, + { + "type": "list", + "bbox": [ + 0.22, + 0.332, + 0.785, + 0.84 + ], + "angle": 0, + "content": null + } + ], + [ + { + "type": "page_number", + "bbox": [ + 0.218, + 0.116, + 0.236, + 0.127 + ], + "angle": 0, + "content": "16" + }, + { + "type": "header", + "bbox": [ + 0.272, + 0.115, + 0.353, + 0.127 + ], + "angle": 0, + "content": "Y. Oh et al." + }, + { + "type": "ref_text", + "bbox": [ + 0.218, + 0.148, + 0.785, + 0.175 + ], + "angle": 0, + "content": "13. Hermans, A., Beyer, L., Leibe, B.: In defense of the triplet loss for person re-identification. arXiv preprint arXiv:1703.07737 (2017)" + }, + { + "type": "ref_text", + "bbox": [ + 0.218, + 0.177, + 0.785, + 0.204 + ], + "angle": 0, + "content": "14. Ignatov, A., Van Gool, L., Timofte, R.: Replacing mobile camera isp with a single deep learning model. arXiv preprint arXiv:2002.05509 (2020)" + }, + { + "type": "ref_text", + "bbox": [ + 0.219, + 0.206, + 0.785, + 0.247 + ], + "angle": 0, + "content": "15. Koh, J., Lee, J., Yoon, S.: Bnudc: a two-branched deep neural network for restoring images from under-display cameras. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 1950-1959 (2022)" + }, + { + "type": "ref_text", + "bbox": [ + 0.219, + 0.249, + 0.785, + 0.289 + ], + "angle": 0, + "content": "16. Kong, L., Dong, J., Ge, J., Li, M., Pan, J.: Efficient frequency domain-based transformers for high-quality image deblurring. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 5886-5895 (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.219, + 0.291, + 0.785, + 0.345 + ], + "angle": 0, + "content": "17. Kwon, K., Kang, E., Lee, S., Lee, S.J., Lee, H.E., Yoo, B., Han, J.J.: Controllable image restoration for under-display camera in smartphones. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 2073-2082 (2021)" + }, + { + "type": "ref_text", + "bbox": [ + 0.219, + 0.347, + 0.785, + 0.388 + ], + "angle": 0, + "content": "18. Lee, W., Son, S., Lee, K.M.: Ap-bsn: Self-supervised denoising for real-world images via asymmetric pd and blind-spot network. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 17725-17734 (2022)" + }, + { + "type": "ref_text", + "bbox": [ + 0.219, + 0.39, + 0.785, + 0.416 + ], + "angle": 0, + "content": "19. Liang, C.H., Chen, Y.A., Liu, Y.C., Hsu, W.H.: Raw image deblurring. IEEE Transactions on Multimedia 24, 61-72 (2020)" + }, + { + "type": "ref_text", + "bbox": [ + 0.218, + 0.418, + 0.785, + 0.445 + ], + "angle": 0, + "content": "20. Loshchilov, I., Hutter, F.: Sgdr: Stochastic gradient descent with warm restarts. arXiv preprint arXiv:1608.03983 (2016)" + }, + { + "type": "ref_text", + "bbox": [ + 0.218, + 0.447, + 0.785, + 0.473 + ], + "angle": 0, + "content": "21. Van der Maaten, L., Hinton, G.: Visualizing data using t-sne. Journal of machine learning research 9(11) (2008)" + }, + { + "type": "ref_text", + "bbox": [ + 0.218, + 0.476, + 0.785, + 0.529 + ], + "angle": 0, + "content": "22. Nam, S., Hwang, Y., Matsushita, Y., Kim, S.J.: A holistic approach to cross-channel image noise modeling and its application to image denoising. In: Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition. pp. 1683-1691 (2016)" + }, + { + "type": "ref_text", + "bbox": [ + 0.218, + 0.532, + 0.785, + 0.585 + ], + "angle": 0, + "content": "23. Oh, Y., Park, G.Y., Cho, N.I.: Restoration of high-frequency components in under display camera images. In: 2022 Asia-Pacific Signal and Information Processing Association Annual Summit and Conference (APSIPA ASC). pp. 1040-1046. IEEE (2022)" + }, + { + "type": "ref_text", + "bbox": [ + 0.218, + 0.588, + 0.785, + 0.642 + ], + "angle": 0, + "content": "24. Oh, Y., Park, G.Y., Chung, H., Cho, S., Cho, N.I.: Residual dilated u-net with spatially adaptive normalization for the restoration of under display camera images. In: 2021 Asia-Pacific Signal and Information Processing Association Annual Summit and Conference (APSIPA ASC). pp. 151-157. IEEE (2021)" + }, + { + "type": "ref_text", + "bbox": [ + 0.218, + 0.645, + 0.785, + 0.685 + ], + "angle": 0, + "content": "25. Plotz, T., Roth, S.: Benchmarking denoising algorithms with real photographs. In: Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition. pp. 1586-1595 (2017)" + }, + { + "type": "ref_text", + "bbox": [ + 0.218, + 0.687, + 0.785, + 0.741 + ], + "angle": 0, + "content": "26. Punnappurath, A., Abuolaim, A., Abdelhamed, A., Levinshtein, A., Brown, M.S.: Day-to-night image synthesis for training nighttime neural isps. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 10769-10778 (2022)" + }, + { + "type": "ref_text", + "bbox": [ + 0.218, + 0.743, + 0.785, + 0.797 + ], + "angle": 0, + "content": "27. Ronneberger, O., Fischer, P., Brox, T.: U-net: Convolutional networks for biomedical image segmentation. In: Medical Image Computing and Computer-Assisted Intervention-MICCAI 2015: 18th International Conference, Munich, Germany, October 5-9, 2015, Proceedings, Part III 18. pp. 234-241. Springer (2015)" + }, + { + "type": "ref_text", + "bbox": [ + 0.218, + 0.799, + 0.785, + 0.84 + ], + "angle": 0, + "content": "28. Schroff, F., Kalenichenko, D., Philbin, J.: Facenet: A unified embedding for face recognition and clustering. In: Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition. pp. 815-823 (2015)" + }, + { + "type": "list", + "bbox": [ + 0.218, + 0.148, + 0.785, + 0.84 + ], + "angle": 0, + "content": null + } + ], + [ + { + "type": "header", + "bbox": [ + 0.472, + 0.115, + 0.732, + 0.129 + ], + "angle": 0, + "content": "Panel-Specific UDC Image Restoration" + }, + { + "type": "page_number", + "bbox": [ + 0.768, + 0.116, + 0.786, + 0.127 + ], + "angle": 0, + "content": "17" + }, + { + "type": "ref_text", + "bbox": [ + 0.214, + 0.147, + 0.788, + 0.189 + ], + "angle": 0, + "content": "29. Schwartz, E., Giryes, R., Bronstein, A.M.: Deepisp: Toward learning an end-to-end image processing pipeline. IEEE Transactions on Image Processing 28(2), 912-923 (2018)" + }, + { + "type": "ref_text", + "bbox": [ + 0.214, + 0.19, + 0.788, + 0.232 + ], + "angle": 0, + "content": "30. Song, B., Chen, X., Xu, S., Zhou, J.: Under-display camera image restoration with scattering effect. In: Proceedings of the IEEE/CVF International Conference on Computer Vision. pp. 12580-12589 (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.214, + 0.232, + 0.788, + 0.286 + ], + "angle": 0, + "content": "31. Sundar, V., Hegde, S., Kothandaraman, D., Mitra, K.: Deep atrous guided filter for image restoration in under display cameras. In: Computer Vision-ECCV 2020 Workshops: Glasgow, UK, August 23-28, 2020, Proceedings, Part V 16. pp. 379-397. Springer (2020)" + }, + { + "type": "ref_text", + "bbox": [ + 0.214, + 0.286, + 0.788, + 0.342 + ], + "angle": 0, + "content": "32. Wang, L., Wang, Y., Dong, X., Xu, Q., Yang, J., An, W., Guo, Y.: Unsupervised degradation representation learning for blind super-resolution. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 10581-10590 (2021)" + }, + { + "type": "ref_text", + "bbox": [ + 0.214, + 0.342, + 0.788, + 0.384 + ], + "angle": 0, + "content": "33. Yu, B., Liu, T., Gong, M., Ding, C., Tao, D.: Correcting the triplet selection bias for triplet loss. In: Proceedings of the European Conference on Computer Vision (ECCV). pp. 71-87 (2018)" + }, + { + "type": "ref_text", + "bbox": [ + 0.214, + 0.384, + 0.788, + 0.438 + ], + "angle": 0, + "content": "34. Yuan, Y., Chen, W., Yang, Y., Wang, Z.: In defense of the triplet loss again: Learning robust person re-identification with fast approximated triplet loss and label distillation. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition Workshops. pp. 354-355 (2020)" + }, + { + "type": "ref_text", + "bbox": [ + 0.214, + 0.438, + 0.788, + 0.493 + ], + "angle": 0, + "content": "35. Zamir, S.W., Arora, A., Khan, S., Hayat, M., Khan, F.S., Yang, M.H., Shao, L.: Cycleisp: Real image restoration via improved data synthesis. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 2696-2705 (2020)" + }, + { + "type": "ref_text", + "bbox": [ + 0.214, + 0.493, + 0.788, + 0.535 + ], + "angle": 0, + "content": "36. Zhou, Y., Song, Y., Du, X.: Modular degradation simulation and restoration for under-display camera. In: Proceedings of the Asian Conference on Computer Vision. pp. 265-282 (2022)" + }, + { + "type": "ref_text", + "bbox": [ + 0.214, + 0.535, + 0.788, + 0.605 + ], + "angle": 0, + "content": "37. Zhou, Y., Kwan, M., Tolentino, K., Emerton, N., Lim, S., Large, T., Fu, L., Pan, Z., Li, B., Yang, Q., et al.: Udc 2020 challenge on image restoration of under-display camera: Methods and results. In: Computer Vision-ECCV 2020 Workshops: Glasgow, UK, August 23-28, 2020, Proceedings, Part V 16. pp. 337-351. Springer (2020)" + }, + { + "type": "ref_text", + "bbox": [ + 0.214, + 0.605, + 0.788, + 0.647 + ], + "angle": 0, + "content": "38. Zhou, Y., Ren, D., Emerton, N., Lim, S., Large, T.: Image restoration for underdisplay camera. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 9179-9188 (2021)" + }, + { + "type": "list", + "bbox": [ + 0.214, + 0.147, + 0.788, + 0.647 + ], + "angle": 0, + "content": null + } + ] +] \ No newline at end of file diff --git a/2024/Panel-Specific Degradation Representation for Raw Under-Display Camera Image Restoration/c089de88-5e00-46c3-9141-1e16159753ca_origin.pdf b/2024/Panel-Specific Degradation Representation for Raw Under-Display Camera Image Restoration/c089de88-5e00-46c3-9141-1e16159753ca_origin.pdf new file mode 100644 index 0000000000000000000000000000000000000000..96eb22d42a34ccbec0e119ca0477d603371fa577 --- /dev/null +++ b/2024/Panel-Specific Degradation Representation for Raw Under-Display Camera Image Restoration/c089de88-5e00-46c3-9141-1e16159753ca_origin.pdf @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:5cfe29d1e264d60e0401318d2b031ce1c5fd14df9e762ff5e8e5de87db845c6a +size 8176653 diff --git a/2024/Panel-Specific Degradation Representation for Raw Under-Display Camera Image Restoration/full.md b/2024/Panel-Specific Degradation Representation for Raw Under-Display Camera Image Restoration/full.md new file mode 100644 index 0000000000000000000000000000000000000000..e967dea7068e9968db96b3e538ed30ea474d6e27 --- /dev/null +++ b/2024/Panel-Specific Degradation Representation for Raw Under-Display Camera Image Restoration/full.md @@ -0,0 +1,401 @@ +# Panel-Specific Degradation Representation for Raw Under-Display Camera Image Restoration + +Youngjin Oh $^{1}$ , Keuntek Lee $^{1}$ , Jooyoung Lee $^{3}$ , Dae-Hyun Lee $^{3}$ , and Nam Ik Cho $^{1,2}$ + +$^{1}$ Department of ECE & INMC, Seoul National University, Seoul, Korea {yjymoh0211,leekt000,nicho}@snu.ac.kr +$^{2}$ IPAI, Seoul National University, Seoul, Korea +3 SK hynix Inc., Korea {jooyoung2.lee, daehyun1.lee}@sk.com + +Abstract. Under-display camera (UDC) image restoration aims to restore images distorted by the OLED display panel covering the frontal camera on a smartphone. Previous deep learning-based UDC restoration methods focused on restoring the image within the RGB domain with the collection of real or synthetic RGB datasets. However, UDC images in these datasets exhibit domain differences from real commercial smartphone UDC images while inherently constraining the problem and solution within the RGB domain. To address this issue, we collect well-aligned sensor-level real UDC images using panels from two commercial smartphones equipped with UDC. We also propose a new UDC restoration method to exploit the disparities between degradations caused by different panels, considering that UDC degradations are specific to the type of OLED panel. For this purpose, we train an encoder with an unsupervised learning scheme using triplet loss that aims to extract the inherent degradations caused by different panels from degraded UDC images as implicit representations. The learned panel-specific degradation representations are then provided as priors to our restoration network based on an efficient Transformer network. Extensive experiments show that our proposed method achieves state-of-the-art performance on our real raw image dataset and generalizes well to previous datasets. Our dataset and code is available at https://github.com/OBAKSA/DREUDC. + +Keywords: Under-Display Camera $\cdot$ Image Restoration $\cdot$ Representation Learning + +# 1 Introduction + +To hide the camera hole in the frontal side of smartphones, under-display camera (UDC) systems have been developed, enabled by recent advancements in imaging and display panel technologies. With the UDC technology, smartphone users can enjoy a full-screen display without any camera holes or notches on the front side of their device. However, since the camera is hidden under the panel, the light that reaches the camera is altered by the complex layers of the panel, causing + +Table 1: Comparison of previous datasets [9,10,30,38] and our commercial dataset. + +
DatasetZhou et al. [38]Feng et al. [10]Feng et al. [9]Song et al. [30]Ours
FormatRGBRGBRGBRGBRGB/RAW
UDC ModelingReal (prototype)SyntheticRealSyntheticReal
Aligned GT✓✓✗✓✓
Multiple Display✓✗✗✗✓
+ +![](images/d948b0a515a4d41a88c7fbc0d82e2f66ae1c6d20a86b66a90d3aa0bcf36e02d1.jpg) +(a) Feng et al. [9], GT + +![](images/6ffebcb9c500395872be2223822ed775976f904faa0de2179265e107609aa0cf.jpg) + +l. [9], GT + +![](images/c44ea9092d19ab626987bfc963fe0afc1d6cea238d5dd4ba61c644430f212330.jpg) +(b) Feng et al. [9], UDC + +![](images/a07cd45ac42d6e56651124ea01d51ba65f9194188f40d955e0344d4b1a7628b4.jpg) + +al. [9], UDC + +![](images/f6255a2f4b4f8f70c9e59f1bedbc4a12154fac0f34dba31677604939441d38e0.jpg) + +![](images/3f4084fbd01901d35c7e56711e655a381b04282122bdd002b908db7c1fcf519c.jpg) +(d) Feng et al. [10], GT + +l. [10], GT + +![](images/852a895706697611d4b76c13c5b46fb97831956a03d1e9d819fccdffb10d5c7c.jpg) +(e) Feng et al. [10], UDC + +![](images/d0f96c5e388cf4d5cb82d4beea9c506c816e7c1ad43d65e6ee4aa8c7b31b07f6.jpg) +(f) Song et al. [30], UDC + +![](images/656f2a6209f476ac2a4613a836bbb4c45d6e89bf1c78caf87f96dd154cb4d8a9.jpg) + +![](images/aacadb8f0484e70271a5c0dc591bda948a147c3661a8a22a0376a7491093fac0.jpg) + +![](images/b78df607efece1427d9a1c6fb5dab173aa63b7e57b631e540e202f62b710ed46.jpg) +(h) Zho +Zhou et al. [38], T-OLED + +![](images/eca13371b9bbecceb84ed27e6ba1cee8f122782f1df832359620a0647d8f5708.jpg) + +![](images/9e0966a407f913e42c0d6c843a91c52368e6507478027bf8554a990c5da152de.jpg) + +![](images/175e3e054a157fa6399689707e170996e0f102d7afa2ae76bd82a96b99bc5865.jpg) +(g) Zhou et al. [38], +(j) Ours, GT +Fig. 1: Samples and their RGB channel histograms from previous UDC datasets [9, 10,30,38] and our commercial dataset. + +![](images/33eb15ba359e820dc42a731fafa38907df0578b97b4b51c168ed0a9761c3298c.jpg) + +![](images/b79214c6f053dea6614bebde267b73bcbdf0893279ac48494551b4ccadfd0bbc.jpg) +(k) Ours, Axon 30 + +![](images/5cb59ca95b7eb7347ba8642a6a8fcb253962c30ccffa8eb2c6203640970b2d43.jpg) +) Zhou et al +[38], P-OLED +(1) Ours, Z-Fold 3 + +severe degradation in the resulting image. Degradations include blur, diffraction, noise, low-lightness, color shift, haziness, etc., and their spatially-variant nature makes the UDC image restoration very challenging [10, 17, 38]. + +Most UDC image restoration methods used deep neural networks [9-11, 15, 17, 23, 24, 30, 31, 37, 38], providing promising results on public benchmark datasets [9, 10, 30, 38] gathered to facilitate research on UDC image restoration. However, these existing datasets either use prototype display panels to inadequately represent the degradations observed in real UDC images caused by commercial panels [38], lack realistic degradations due to incomplete UDC modeling (e.g., noise and color shift) in the synthesis pipeline [10,30], or mishandles aligning image pairs with occlusion and parallax due to the mechanism of the capturing imaging system [9]. In addition, the potential for enhanced image restoration performance through sensor-level restoration [1, 19, 25, 35] cannot be exploited, because previous datasets predominantly restrict experiments to RGB domain as raw data is unreleased. Tab. 1 shows comparison of the datasets. + +In this work, we propose a new UDC dataset that is collected in the raw domain. The proposed dataset contains pairs of real UDC images and clean images that are well-aligned using a monitor-based imaging system with two different panels from commercial UDC smartphones, ZTE Axon 30 5G and Samsung Galaxy Z-Fold 3. In our experiments, we find that training with raw data enhances the performance of UDC image restoration. This raw-domain approach with real data also has the advantage of enabling the use of end-to-end neural ISPs that are being installed in recent smartphones [26], allowing a practical implementation on devices. Additionally, as shown in Fig. 1, real UDC data captured with commercial smartphone panels display noticeable distinct degradations visually and statistically when compared to previous datasets [10,30,38]. + +Furthermore, we develop a two-stage restoration method for UDC images. First, we introduce a framework for learning the implicit representation of degradations in the UDC images. We notice that degradation in a patch of a UDC image is similar to that of other patches of the same UDC image while being distinct from degradations of the same scene taken with a different panel. With this observation, we develop a novel unsupervised learning scheme involving triplet loss [28] to train our encoder. The learning scheme aims to leverage the information of the data collected with different types of panels, ultimately condensing the degradation information into an implicit representation vector that reflects the panel-specific degradation moderately. We also propose an efficient Transformer-based restoration network that incorporates the learned panel-specific degradation representations as a prior by generating channel-wise calibration coefficients for attention. We discover that our strategy yields improved results for UDC image restoration. To the best of our knowledge, this is the first case in UDC image restoration where a dataset collected with one panel is utilized to enhance the performance on a dataset collected with another panel. + +To summarize, our contributions are as follows: + +- We introduce a new real UDC dataset, which is collected with commercial smartphone panels. Our dataset is provided in raw format, which can improve the performance of the UDC restoration task. +- We propose an unsupervised learning approach that allows for implicit representations of degradations caused by various display panels in the embedding space. The panel-specific degradation representations are leveraged as priors in the restoration network, which is an efficient Transformer and brings notable improvement. +- Extensive experiments demonstrate that our restoration network embedded with learned representations achieves state-of-the-art performance on the proposed dataset and is consistently applicable to previous datasets. + +# 2 Related Works + +# 2.1 UDC Image Restoration + +Since UDC image restoration is a relatively new area, research is still ongoing to provide users with a satisfying experience when UDC is applied in real situa- + +tions. Research on UDC image restoration was sparked by the pioneering work of [38]. They defined a UDC imaging system and released the first UDC dataset for researchers to experiment with. Synthetic datasets were also made available by simulating the camera pipeline using a pre-calculated point spread function (PSF), which adequately represents the diffraction [10], with further improvements on scattering effects in [30] to account for the haziness of UDC images. More recently, a real UDC dataset generated by pseudo-aligning geometrically misaligned image pairs for a single UDC panel [9] was presented. Although the proposed datasets have some drawbacks, several works [9-11, 15, 17, 23, 24, 30, 31, 37, 38] have investigated UDC image restoration using deep neural networks and have shown promising results. + +Among these deep-learning-based methods, several of them use priors to improve performance. For instance, in [10], PCA is employed on a PSF of the UDC display panel to stretch it as a 'kernel code' and condition the network with it as a prior. In [30], the network is divided into a scattering branch and an image branch, and the former is trained to estimate the scattering parameters, which is prior information that can be used. However, these methods require precise priors; otherwise, the performance is degraded. Our approach differs from previous methods because we do not need strict parameters or ground truth settings. Instead, we aim to learn an acceptable representation of the degradation induced by panel-specific properties in the embedding space rather than a rigorous estimation of the degradation. We only need images taken with different panels because our method is unsupervised, and learn the implicit priors by leveraging the discrepancies of degradation caused by different panels. + +# 2.2 Image Restoration with Raw Data + +Most of the previous research on image restoration relies on RGB images as training data as they are easier to collect. However, some recent studies have shown that using raw sensor data can lead to improved performance. Specifically, [1, 25] are the works on image denoising where they collected DND and SIDD datasets that are available as raw and RGB images, where denoising with raw images usually yields higher PSNR. CycleISP [35] demonstrated that noise is more complex in RGB images and showed that noise modeling is more realistic in the sensor domain. This is because compared to the noise samples in the sensor pixels, which are nearly uncorrelated to each other, the noise samples become highly correlated in RGB images as they pass through the camera ISP pipeline, which involves processes like demosaicing, gamma correction, and white balancing [18, 22]. DeepISP [29] proposed an end-to-end camera ISP modeled with deep neural networks, suggesting that a deep network-based camera ISP is promising. Furthermore, [4] solved the issue of low-light enhancement problem by operating in the sensor domain, and [19] presented a dataset for raw image deblurring and proved that operating on raw sensor data achieves better performance than RGB-based methods. These researches suggest that it is easier to restore raw data than RGB images with deep neural networks, as sensor data contains the linear representation of incoming light intensity of the scene. This ensures that + +the scene's authenticity is maintained more faithfully in raw data than in heavily processed RGB pixel values that are nonlinear with respect to light intensity. + +# 3 Proposed Method + +# 3.1 Sensor-level Real UDC Dataset + +A raw dataset is necessary to achieve end-to-end image restoration using sensor images. However, the available UDC datasets are limited to the RGB domain. Moreover, public UDC datasets do not show the types of degradation that are present in images taken with commercially available smartphones equipped with UDC, limiting their assessment of practical scenarios. Therefore, we capture real raw UDC data by using OLED panels obtained from commercial UDC smartphones, a smartphone camera, and a monitor-based imaging system [38]. + +UDC data acquisition process Capturing a raw UDC image requires that we model the UDC imaging system as similarly as possible. Consequently, our data-capturing process involves rendering an image onto a high-resolution monitor, followed by capturing the displayed image using a smartphone camera sensor located closely behind a UDC panel. The panels used to capture the data are from two commercial UDC smartphones: ZTE Axon 30 5G and Samsung Galaxy Z-Fold 3. They introduce low-light/color shift, scattering, diffraction, and blur. As noise is usually generated from the sensor, we chose Samsung Galaxy S23 Ultra as our camera device to capture the authentic noise of smartphone camera sensors. Raw images are collected using the internal software of the smartphone that automatically saves the minimally processed raw data. The preference for monitor-displayed images [17,38] over real natural scenes is driven by the need to control additional factors beyond the panels' presence. Specifically, we can enforce the target scene to be static as dynamic scenes are prone to motion blur, and therefore lead to misaligned data pairs. We can also eliminate interference from external ambient sources of light other than the monitor's emitted light. + +The monitor, UDC panel, and camera sensor are aligned in sequence, and the principal axis of the camera is aligned perpendicular to the plane of the monitor and the panel. The camera is mounted on a sturdy tripod, and the panel is situated close to the sensor using a firm holding device. The displayed images are then captured remotely with/without the smartphone panels to ensure the alignment of clean/UDC image pairs during capture while keeping the camera settings the same. We provide more details in the Supplementary Material. + +Our UDC dataset The dataset is captured using 400 images from the DIV2K dataset [2], which is comprised of natural scene images from various themes, encompassing diverse objects captured under different lighting conditions. This makes our dataset 1,200 images in total, as there are three different image sets (Axon 30, Z-Fold 3, ground truth) in raw format. The images are stored in 16 bits with a Bayer pattern of G-B-R-G. For the ground truth images, we capture + +![](images/8d7838f33a0fc9b19804d0939e843031ad78f3fb51bac4fbeaf68210bf52f226.jpg) + +![](images/86ac0c9bc19cf47552625d9c3439435a1bffb1f18c2061258c88989b772b2d54.jpg) +(a) Clean image +Fig. 2: Cropped examples from our captured dataset. All images are visualized with the same procedure that includes normalization, demosaicing, channel gain adjustment, and gamma correction. Due to the difference in smartphone display panels, noise and haze are more apparent in Axon 30 UDC images, while the Z-Fold 3 UDC images show blurrier results. The diffraction induces flare artifacts when the image intensity becomes too strong that the sensor saturates. + +![](images/f4e1113a491cc030422183607632a0f7bac0a31aaa6b9e9f1806dd50a70b365f.jpg) + +![](images/17fdc077e18fb78a35e80a41fead853513427d56d89a7d63ed5091f8ef20e8fe.jpg) +(b) Axon 30 image + +![](images/ed02638ba7fa4724740b8648c53514befc953003693debced1e1d41f68a3fc53.jpg) + +![](images/dedcff742a8060f4fa0c23ca613882a2884f8fc74f73d5220d5e832fa096a21c.jpg) +(c) Z-Fold 3 image + +bursts of a scene without the display panel and average them. To check for misalignments that can be induced by optical image stabilization in the smartphone hardware [1], we utilize registration and feature matching algorithms [3, 12]. + +Samples from our dataset are displayed in Fig. 2, with each UDC image taken with Axon 30 and Z-Fold 3 display panels exhibiting panel-specific degradations. As such, our dataset can be used to evaluate if a restoration method generally performs well in the restoration of various UDC systems. Also, our real commercial UDC data show different behaviors from previous datasets, as shown in Fig. 1. Specifically, low-light, color shift, blur, and noise are observed, and an object is seen repeated faintly near the main object due to diffraction. We also detect a phenomenon similar to haze due to the scattering effects of the panel [30]. Synthetic datasets that are generated using UDC modeling [10, 30] do not take noise into account. However, including noise in UDC images is vital as the level of noise captured with frontal smartphone camera sensors, which are relatively inferior to DSLR camera sensors [4, 26], is amplified when normalizing sensor data. This amplification is related to the attenuation caused by the panel which reduces light intensity. More samples are available in the Supplementary. + +# 3.2 Proposed UDC Restoration Framework + +In this section, we propose our restoration framework called Degradation Representation Embedded Transformer for UDC image restoration (DREUDC), which consists of a degradation representation encoder and a restoration network. We + +first train the encoder with our proposed unsupervised scheme using triplet loss to implicitly learn the degradations of a display panel. Then, we utilize the representations generated from the encoder by integrating them in the restoration network as a prior. It is worth noting that our approach's novelty lies in how we take advantage of UDC data from one display panel to restore a UDC image from another panel rather than in the restoration network's architecture itself. + +Motivation Representation learning has been demonstrated to be effective in blind super-resolution, as demonstrated in [32]. Inspired by this, we explore the use of representation learning to enhance the restoration of UDC images by extracting useful prior information about the panels. Since each UDC system has a different panel, it yields panel-specific degradations. For example, the UDC images presented in Fig. 2, which are captured using two different panels of UDC smartphones, show differences in the degree of degradations. This implies that the degradations present in an image captured using one panel are similar to each other and that they differ significantly from those of the image with the same content but are captured using another panel. Using this information, our approach intends to utilize these differences and train an encoder to encode them in an unsupervised manner into the embedding space. + +Unsupervised representation encoder training Our proposed unsupervised training scheme is visualized in Fig. 3. As mentioned above, the objective of training the panel-specific degradation representation encoder is to learn the unique properties of the panels and produce implicit representations of the degradations to be used as priors in UDC image restoration. To achieve this, the encoder should learn the degradations using attributes related to a specific panel, such as the level of noise, blur, and color shift. On the other hand, the encoder should be discouraged from learning degradations with features in an image that are panel-agnostic, such as textures and contents of a scene. + +We believe that this condition is analogous to the situation of tasks where triplet loss [28] is exploited to learn favorable representations. Triplet loss is widely applied to various tasks of computer vision such as facial recognition [28], person re-identification [13,34], and domain adaptation [8] for its ability to learn useful embeddings by leveraging the distances between embeddings of similar instances and dissimilar instances. For instance, in person re-identification, networks are guided to identify among different people using features that are unique to a specific person, not with features that are shared among people, by learning suitable representations using triplet loss. In our case, the features unique to a specific person would correspond to properties unique to a specific panel, and the features that are shared among people would correspond to panel-agnostic features. Therefore, we opt to apply triplet loss in the training of our encoder to capture the panel-specific degradation representations. + +Nevertheless, triplet loss is only effective when the network is provided with valid triplets [28, 33]. To collect meaningful triplets, we impose constraints on the training pairs that will be used to train our encoder. We choose a pair of + +![](images/102c7da16a0a2a33a483935078135635c3dfde019631cf67c6a2d2f0a9d21ada.jpg) +$x_{i}$ + +![](images/be880941c16cdc39543dd9a8e4879c36736fe48e2f0fbda52fa472477312ae56.jpg) +$x_{j}$ Raw UDC images (visualized for figure) + +![](images/04272cd6e3cf06c4a0814180c5e155924f18263db7ffbc4c7aa71ef051d4ee0f.jpg) +Training patches +Panel-specific degradation representation +Projected representations +Fig. 3: Our proposed unsupervised panel-specific degradation representation learning scheme. By using triplet loss, representations of the same panel-specific degradations with different image content are pulled closer together, while representations of different panel-specific degradations with the same image content are pushed away. + +raw images $x_{i}$ and $x_{j}$ from our dataset that contains the same content but have been taken with different display panels. Then, we randomly crop two pairs of patches $(x_{i}^{1}, x_{i}^{2})$ and $(x_{j}^{1}, x_{j}^{2})$ from each image, where $x_{i}^{1}$ and $x_{j}^{1}$ , and $x_{i}^{2}$ and $x_{j}^{2}$ , respectively, are cropped from the same location of $x_{i}$ and $x_{j}$ . We choose patch $x_{i}^{1}$ as the anchor, and the positive and negative samples are decided to be $x_{i}^{2}$ and $x_{j}^{1}$ . The same process is repeated with $x_{j}^{1}$ as the anchor. + +With this triplet selection, the encoder is guided to learn the intrinsic properties of the degradations caused by the display panel by pulling representation of images that show similar degradation while disregarding the features that are unrelated to degradations by pushing the representation of the same contents. The encoder $f$ is trained with triplet loss using the selected triplets to obtain the panel-specific degradation representation $y \in \mathbb{R}^{256}$ . The representations are further projected to a projection $z$ with two fully connected layers, following [5, 6]. The triplet loss for training the encoder is as follows: + +$$ +\begin{array}{l} L = \max \left(\left\| z _ {i} ^ {1} - z _ {i} ^ {2} \right\| _ {2} ^ {2} - \left\| z _ {i} ^ {1} - z _ {j} ^ {1} \right\| _ {2} ^ {2} + \alpha , 0\right) \\ + m a x (| | z _ {j} ^ {1} - z _ {j} ^ {2} | | _ {2} ^ {2} - | | z _ {j} ^ {1} - z _ {i} ^ {1} | | _ {2} ^ {2} + \alpha , 0), \quad (1) \\ \end{array} +$$ + +where $\alpha$ is the margin of the triplet loss. The margin is set as $\alpha = 1$ . Following [32], the encoder is a simple network with six convolutional layers. + +Representation embedded restoration network For our restoration network, we design it to be as simple and efficient as possible to emphasize the ef- + +![](images/baca668d5f8b5a85e243d7382e7c330b0b3f5b757a74ed434aec8ba3b75c838f.jpg) +(a) Our proposed framework, DREUDC +(b) DREblock +Fig. 4: Our proposed DREUDC. The framework consists of an encoder that produces an implicit representation of the panel-specific degradations of a UDC image and a restoration network that exploits the representations via the proposed DREblocks. + +fectiveness of our contribution of how we leverage the difference in data collected with various panels to enhance the restoration performance rather than focusing on the impact of complex restoration network architectures. Therefore, we adopt a U-shaped [27] Transformer with Frequency domain-based Self-Attention Solver (FSAS) [16] as our restoration architecture. In previous works on UDC image restoration [9, 10, 17, 31], researchers have highlighted the importance of a wide receptive field. As FSAS utilizes the convolutional theorem of Fourier transform to model global contexts in the frequency domain efficiently [16], we choose it as a component of the building block for our restoration network. + +Our restoration network is based on stacking repeated Transformer blocks, which we name DREblock (Degradation Representation Embedded Transformer block, Fig. 4(b)). The proposed DREblock efficiently integrates the degradation representation $y$ of the input UDC image as a condition of our restoration network by calibrating the features in the block in a channel-wise manner, i.e., + +$$ +F _ {o} = F _ {a t t} \otimes \beta , \tag {2} +$$ + +where $F_{att}$ is the feature maps of the estimated attention of FSAS, $F_{o}$ is the feature after the conditioning, $\otimes$ denotes channel-wise multiplication, and $\beta$ is the channel conditioning coefficients that are produced by reshaping the panel-specific degradation representation $y$ using a single layer of $1 \times 1$ convolution. The features $F_{o}$ are then processed by the point-wise feed-forward network. + +The entire framework DREUDC is illustrated in Fig. 4. In training, the encoder is first trained with Eq. (1). Then, the encoder is frozen while the restoration network is trained with $\mathcal{L}_1$ loss. During inference, the encoder takes the input raw UDC image and outputs a panel-specific degradation representation vector that contains prior information on UDC degradation. The representation is seamlessly embedded into the restoration network through DREblocks, finally producing a clean restored RGB image. + +# 4 Experiment and Analysis + +# 4.1 Datasets + +We evaluate the restoration performance using our captured data with display panels of ZTE Axon 30 5G and Samsung Galaxy Z-Fold 3. The raw images are normalized to have values within the range of [0,1]. To train and evaluate RGB-based methods, we use a simplified ISP algorithm written with MATLAB to visualize the normalized raw images to RGB images. The algorithm contains demosaicing, channel-wise gain control, and gamma correction in sequence. Note that the visualized images are not sRGB images, as actual camera ISPs in practice involve additional steps to render the image in sRGB domain. Please check the Supplementary for more discussion on the usage of simplified ISP. For both display panels, we split the dataset into 300 images for training and 100 images for testing. The training set is cropped into patches of $512 \times 512$ . We also use previous UDC datasets [38], which consists of 300 images of both T-OLED and P-OLED scenes, to verify that our method is applicable to other real datasets. + +# 4.2 Training and Implementation Details + +We use NVIDIA RTX 3090 GPUs in our experiments. To train the encoder, we use Adam optimizer $(\beta_{1} = 0.9, \beta_{2} = 0.999)$ for $1 \times 10^{5}$ iterations with a batch size of 64, and the loss is given as Eq. (1). The restoration networks are trained with $\mathcal{L}_1$ loss to minimize the distance between the restored image of the network and the ground truth. The initial number of channels $C$ of our restoration network is 32, and the number of blocks in each layer $L_1, L_2$ , and $L_3$ are [5,5,6]. We use Adam optimizer $(\beta_{1} = 0.9, \beta_{2} = 0.999)$ for $2 \times 10^{5}$ iterations with a batch size of 2. For both networks, the initial learning rate is set as $2 \times 10^{-4}$ and gradually decayed to $1 \times 10^{-6}$ with cosine annealing [20]. This setting is applied to all experiments on our dataset, and every previous method reported in this paper is also reproduced by training on our dataset with the same setting for fairness. + +# 4.3 Raw UDC Image Restoration + +First, we experiment to validate that using raw images can improve the UDC image restoration process. For this purpose, we employ a baseline neural network called DE-UNET [38] and use our captured data. We compare the performance of the network under three different settings: RAW-to-RAW, RAW-to-RGB, and RGB-to-RGB. We evaluate the performance of the network using the PSNR and SSIM metrics. To assess the performance of the RAW-to-RAW setting, we convert the output restored raw images to RGB using the same ISP process mentioned in Sec. 4.1 and measure them in the RGB domain. The results are presented in Tab. 2, showing that using raw sensor images as input (RAW-to-RAW, RAW-to-RGB) yields superior results compared to only using RGB images. Among the two settings that use raw images as input, the latter shows better performance overall, so we choose the method of converting and restoring + +Table 2: Comparison of different input-to-output settings of DE-UNET [38] on our captured datasets. The best results are bold-faced. + +
MethodAxon 30 (raw)Z-Fold 3 (raw)Axon 30 (rgb)Z-Fold 3 (rgb)
PSNRSSIMPSNRSSIMPSNRSSIMPSNRSSIM
DE-UNET(RGB-to-RGB)----23.330.910926.900.9390
DE-UNET(RAW-to-RAW)27.360.933529.660.963724.210.917027.180.9388
DE-UNET(RAW-to-RGB)----24.200.919627.290.9419
+ +Table 3: Results on our captured datasets for ZTE Axon 30 5G and Samsung Galaxy Z-Fold 3 smartphone display panels on various methods of UDC image restoration. The best results are bold-faced. + +
Input-to-outputMethodParamsMACsAxon 30 (rgb)Z-Fold 3 (rgb)
PSNRSSIMPSNRSSIM
RGB-to-RGBDE-UNET [38]8.93M273.84G23.330.910926.900.9390
DAGF [31]1.09M73.84G24.490.915028.420.9404
DISCNET [10]3.80M596.95G23.110.906326.670.9361
BNUDC [15]4.57M993.38G25.850.928328.540.9458
DWFormer [36]1.45M216.77G21.520.892326.360.9338
SRUDC [30]6.70M371.32G22.670.899827.340.9336
PPM-UNET [9]4.23M409.29G18.210.850223.850.9188
AWNet [7]47.00M1.53T20.270.883925.600.9316
RAW-to-RGBDE-UNET [38]8.94M69.29G24.200.919627.290.9419
PyNet [14]47.56M1.79T22.580.906927.350.9377
AWNet [7]49.07M392.25G20.980.888527.250.9369
DREUDC6.44M200.07G26.660.936829.040.9511
+ +raw UDC images to a clean RGB image (RAW-to-RGB). This is also the setting of most neural ISPs currently being used in smartphones [26] since a successful network is able to model the image processing pipeline as well. + +# 4.4 Evaluation on Our Dataset + +To provide experimental results on our dataset and demonstrate the performance of DREUDC on UDC image restoration, we train numerous previous methods using our dataset for comparison. The quality of the restoration performance is evaluated with PSNR and SSIM as performance metrics. Additionally, we assess the efficiency of the methods by comparing the number of parameters and MACs. To calculate the MACs, we use a dummy image with a resolution of $1024 \times 1024$ . + +Tab. 3 shows experimental results for raw and RGB image restoration using our dataset. Because there are no neural networks on UDC image restoration that take raw sensor images as input except [38] as previous datasets were provided only in the RGB domain, we choose [7] and [14] for further comparison, which are RAW-to-RGB image conversion models. We observe that DREUDC achieves state-of-the-art performance on both Axon 30 and Z-fold 3 datasets, + +![](images/cf58ad7779c04e55c3e89ea859c1eaaca0bbc928c454a5744b52b5ec35cd483f.jpg) +Axon 30 + +![](images/5a1a8cb18d978c4675bdfe9e419cdcd277ec46436c5f66f118596ce7557bcb4c.jpg) +DE-UNET + +![](images/443c069547bae31500f4df171a91cef2cb0d9ece0a885cf0b621b624b1fe3eaa.jpg) +DAGF + +![](images/1caa6d41d56031b6e8ea6abf11ee2cfdfc93a7366a8b91dd6a93cddc0c73735e.jpg) +DISCNET + +![](images/999eaef93d5b29bc7c11ce9721f9aa5977a0822732c42b96c9321c8dd2e228a2.jpg) +BNUDC + +![](images/de2b7ff4b043a6512ebb7b040e478633b48541ee459f2fc7c90449b255b5603e.jpg) +DWFormer + +![](images/6096c77c949deacc39372e624b262af9bf37fc56928e6b2e9e860f9bf079b002.jpg) +SRUDC + +![](images/51089fc1f79f4c8142433ecfc7d382fe91b547626d99c32fd262a77dc8e5b7f6.jpg) +PPM-UNet + +![](images/4ff5e71daf11dd9b66559d9a4f3362ff0ada86b13a0d4b0fafe821c8b084d69d.jpg) +AWNet + +![](images/36927a5073b0582a182a878734401017125dcf564e9b5c2a37e9203de448c778.jpg) +DE-UNET* + +![](images/3bfee9f0cd52fd79fe809ff4a57f8eaf3bc47d042de60068cc678c81f4e468b7.jpg) +PyNet* + +![](images/da2042856f64d2302393156e591606e5692b43e16cc78a33c0837b612ab2e325.jpg) +AWNet* + +![](images/cb53a85ef3ac108165db10d937ea1362965ad65f48b7ce111eef22a1d6ecbeae.jpg) +DREUDC* + +![](images/4f45c0a2f1c303941e6fb978baae763cedccde3d64323f982c5031a416288961.jpg) +GT + +![](images/28de6ae19e42b77bbbc157d038977b00bd492172412b94a6d946b5f1c3bac0f8.jpg) +Z-Fold 3 + +![](images/64fecfff12cf6f2be39cd3bbba86cdc45930ac18b4a453adcd1d90557a229167.jpg) +DE-UNET + +![](images/11f2f6831a9fc818a427e74d9dc74aac14688a42aaee80aad7c7615d7ffe8788.jpg) +DAGE + +![](images/c913bb6114da146be22a155aac2b9b910b8f85fda682cf58f9da7f04660978f3.jpg) +DISCNET + +![](images/e91c89e4cf8ed108ccdc67db7c4c6484fa1f7f3eb632bb30a0af0b8ae862ef5d.jpg) +BNUDC + +![](images/3606fcd252b55853f99f3b3f00923e2e553763b8da48090f01d5f27500e00c16.jpg) +DWFormer + +![](images/fa9928952d2f6f5c2d057093b78ab1d74df365c69a455761edf511511e6b478c.jpg) +SRUDC + +![](images/ab01dbcaa3f52cceddbbb6f3978fec19d414d9599db3f1b4e1bfaf9e4b9508a8.jpg) +PPM-UNet +Fig. 5: Visual comparison of restoration examples for our dataset. Zoom in for a better comparison between the methods. The top two rows are results for Axon 30 images, and the bottom two rows are for Z-Fold 3 images. Methods that take raw input and produce RGB images are annotated with the asterisk *. Our method is capable of removing UDC degradations more clearly than previous UDC image restoration networks. + +![](images/917c1e9d880f2e4772a68b9194a7e0a054e967976951b4c4f3a8799838b2f09e.jpg) +AWNet + +![](images/7581d9d8b76118ca94beef42656713ac9c3892ab949ceeb6eb406fa43a4e9d72.jpg) +DE-UNET* + +![](images/bb7cb2031bd8d94490db9686bc8a0a5393d58c215d8a4192d976d1750caae689.jpg) +PyNet* + +![](images/551417cccfd2aa5f9376bbda18fa2f3829c48d51fba0842194d24b60effe900c.jpg) +AWNet* + +![](images/0eebec71e83df9cc38e8ef93358879ed3008d6c100e0e3f08ccf2defd69afe42.jpg) +DREUDC* + +![](images/720186f975f12819cef15bb5695abc5a413fc91798a8ab844a3c80122a87a805.jpg) +GT + +surpassing BNUDC [15] by 0.81dB and 0.50dB, demonstrating that our method can successfully restore various UDC degradations. With the exception of a lightweight model [31] and the baseline model using raw data input [38], our method is the most efficient when it comes to computation while showing superior performance. Qualitative results are illustrated in Fig. 5, and more results including restoration of scenes that are not displayed on a monitor are available in the Supplementary Material. + +# 4.5 Analysis on Panel-Specific Degradation Representation + +In this subsection, we analyze the effectiveness of DREUDC and panel-specific degradation representations learned with our proposed method. The strength of DREUDC is reported in Tab. 4 with an ablation study. We observe that our strategy of embedding an implicit representation of degraded UDC images is indeed effective by comparing (a) and (c) of Tab. 4, where Model 1 is our framework with only the restoration network without channel calibration and representation learning, while DREUDC is our entire framework trained with the protocol explained in Sec. 3.2. + +However, with only this comparison, we cannot rule out the possibility that the rise in performance is solely because of increased parameters. Therefore, we further compare (a) and (c) with (b), where Model 2 is our framework trained in + +Table 4: Results on ablation of our method. + +
MethodChannel calibrationRepresentation learningParams MACsAxon 30Z-Fold 3
PNSRSSIMPNSRSSIM
(a) Model 1XX4.52M160.57G26.090.933828.66
(b) Model 2✓X6.44M200.07G26.380.934828.71
(c) DREUDC✓✓6.44M200.07G26.660.936829.04
+ +![](images/30a8160d912299f840c2c91ff8310327c227fc8485795b99133bf6aaf5c4acfa.jpg) +(a) +Fig. 6: T-SNE [21] visualization of representation generated from the encoder. (a) and (b) show the representations generated from encoders of Model 2 in Tab. 4 for Axon 30 and Z-fold 3 images, respectively, which are trained without representation learning. (d) and (e) show the representations generated from encoders of Model 2 in Tab. 5 for T-OLLED and P-OLED images, respectively, which are trained without representation learning. (c) and (f) illustrate the representations that are generated with our unsupervised triplet learning on our dataset and Zhou's dataset [38]. + +![](images/ca577c75228f3381d81ff06c21f555e00c31b65ad3bae9bbb1b223ee79c211ab.jpg) +(b) + +![](images/c9c552d6708ed8c7b94a7aa5ae4cf7aec4b7a17cf944c8f105965aca7810b75a.jpg) +(c) + +![](images/985738607ba8d6416cbec90cb2aa1ee6e37818d715dca0d1e4e90c8bbebf302c.jpg) +(d) + +![](images/766f86b111c0330b006cea3df6e848b7d2b79161ad9c9acaa15edfa69e5e8252.jpg) +(e) + +![](images/e2b861872f2ccfa4529898572c04002f18e2d97e8ef9c3044a66afb75ae79f25.jpg) +(f) + +an end-to-end manner by training the encoder and the restoration network jointly with $\mathcal{L}_1$ loss without our proposed representation learning. By comparing Model 1 and Model 2, we see that channel calibration with representations learned from a single dataset without triplet training increases the performance marginally. On the contrary, when given an appropriate representation learned from two datasets by leveraging the difference in degradations with our proposed method, we notice that the restoration model benefits substantially from it. Moreover, it is noteworthy that our proposed embedding scheme has a negligible effect on the computational increase (MACs) of the restoration network, as most of the increase is due to the encoder. + +The validity of our panel-specific degradation representation learning scheme is visualized in Fig. 6. Without representation learning (Fig. 6(a) and (b)), the encoders are unable to distinguish different degradations. Conversely, with representation learning (Fig. 6(c)), the encoder identifies the degradation and discriminates them in clusters, offering advantageous priors for the restoration network. + +This finding suggests an approach to train a restoration network by making good use of datasets other than the one gathered for a specific UDC system. Typically, a specific dataset is used to train a restoration network tailored for a particular UDC system. However, this dataset would be sub-optimal or ineffective when future UDC smartphones have different panels. Our method offers a way to utilize other datasets collected with different panels, mediating this issue. + +Table 5: Results of our method and ablation on T-OLED/P-OLED datasets [38]. + +
MethodChannel calibrationRepresentation learningParams MACsT-OLEDP-OLED
PSNR SSIMPSNR SSIM
DE-UNET [38]--8.94M 69.29G36.70 0.974528.58 0.9344
Model 1XX4.52M 160.57G37.37 0.977430.85 0.9510
Model 2✓X6.44M 200.07G37.46 0.977731.01 0.9515
DREUDC✓✓6.44M 200.07G37.55 0.977831.17 0.9521
+ +Application to previous datasets Our restoration method using panel-specific degradation representations is also applied to public UDC datasets to show that it generalizes well to other datasets. Our method requires a real dataset comprised of degraded images taken with more than a single panel from the same capturing environment, so we choose Zhou's P-OLED/T-OLED datasets [38]. DREUDC is designed to take raw images as input, so we make a pseudo-raw version of the dataset by mosaicing the RGB images. This pseudo-raw version of the RGB images that used to train the methods, including a new encoder, is created by reversing the in-camera pipeline mentioned in [38]. We only reversed the demosaicing step as it was the only mentioned process; hence we call it 'pseudo-raw' as the images have not been fully reversed to the original high-bit raw data. We choose DE-UNET [38] to compare with our results because it is a model that can be trained with RAW-to-RGB setting. Additional results on RGB image restoration are provided in the Supplementary Material. + +Tab. 5 shows the results on T-OLED/P-OLED datasets. Our method outperforms DE-UNET on both panels with slightly fewer parameters. Most importantly, our strategy of representation embedding is also effective on [38] when we compare Model 1 to DREUDC, showing an increase of $0.18\mathrm{dB}$ and $0.32\mathrm{dB}$ in the restoration of T-OLED and P-OLED images. We also prove that the boost in performance is not only due to increased parameters but also because we enhance the restoration network with effective representations of the degradations learned using our training scheme by comparing Model 2 and DREUDC. Visualization of the representations are provided in Fig. 6(d), (e), and (f). Overall, the results imply that our method is also effective on other UDC datasets. + +# 5 Conclusion + +We have collected a new UDC dataset using commercial UDC-equipped smartphones and demonstrated that UDC image restoration using raw sensor images achieves better performance than using RGB images. We have also proposed a UDC image restoration framework that generates panel-specific degradation representations in the embedding space with a novel unsupervised representation learning scheme. Our restoration network efficiently accommodates the learned representations as a prior with our proposed Transformer blocks, improving the performance. Experiments show that our method achieves state-of-the-art performance on our dataset and is generalizable to previous datasets. + +# Acknowledgment + +This paper was result of the research project supported by SK hynix Inc., including the support of camera equipments and lab environment for collecting the dataset. It was also partially supported by Institute of Information & communications Technology Planning & Evaluation (IITP) grant funded by the Korea government (MSIT) (No.2021-0-01062), and in part by the BK21 FOUR program of the Education and Research Program for Future ICT Pioneers, Seoul National University in 2024. + +# References + +1. Abdelhamed, A., Lin, S., Brown, M.S.: A high-quality denoising dataset for smartphone cameras. In: Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition. pp. 1692-1700 (2018) +2. Agustsson, E., Timofte, R.: Ntire 2017 challenge on single image super-resolution: Dataset and study. In: Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition Workshops. pp. 126-135 (2017) +3. Bay, H., Tuytelaars, T., Van Gool, L.: Surf: Speeded up robust features. In: Computer Vision-ECCV 2006: 9th European Conference on Computer Vision, Graz, Austria, May 7-13, 2006. Proceedings, Part I 9. pp. 404-417. Springer (2006) +4. Chen, C., Chen, Q., Xu, J., Koltun, V.: Learning to see in the dark. In: Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition. pp. 3291-3300 (2018) +5. Chen, T., Kornblith, S., Norouzi, M., Hinton, G.: A simple framework for contrastive learning of visual representations. In: International Conference on Machine Learning. pp. 1597-1607. PMLR (2020) +6. Chen, X., Fan, H., Girshick, R., He, K.: Improved baselines with momentum contrastive learning. arXiv preprint arXiv:2003.04297 (2020) +7. Dai, L., Liu, X., Li, C., Chen, J.: Attnective wavelet network for image isp. In: Computer Vision-ECCV 2020 Workshops: Glasgow, UK, August 23-28, 2020, Proceedings, Part III 16. pp. 185-201. Springer (2020) +8. Deng, W., Zheng, L., Sun, Y., Jiao, J.: Rethinking triplet loss for domain adaptation. IEEE Transactions on Circuits and Systems for Video Technology 31(1), 29-37 (2020) +9. Feng, R., Li, C., Chen, H., Li, S., Gu, J., Loy, C.C.: Generating aligned pseudosupervision from non-aligned data for image restoration in under-display camera. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 5013-5022 (2023) +10. Feng, R., Li, C., Chen, H., Li, S., Loy, C.C., Gu, J.: Removing diffraction image artifacts in under-display camera via dynamic skip connection network. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 662-671 (2021) +11. Feng, R., Li, C., Zhou, S., Sun, W., Zhu, Q., Jiang, J., Yang, Q., Loy, C.C., Gu, J., Zhu, Y., et al.: Mipi 2022 challenge on under-display camera image restoration: Methods and results. In: European Conference on Computer Vision. pp. 60-77. Springer (2022) +12. Guizar-Sicairos, M., Thurman, S.T., Fienup, J.R.: Efficient subpixel image registration algorithms. Optics letters 33(2), 156-158 (2008) + +13. Hermans, A., Beyer, L., Leibe, B.: In defense of the triplet loss for person re-identification. arXiv preprint arXiv:1703.07737 (2017) +14. Ignatov, A., Van Gool, L., Timofte, R.: Replacing mobile camera isp with a single deep learning model. arXiv preprint arXiv:2002.05509 (2020) +15. Koh, J., Lee, J., Yoon, S.: Bnudc: a two-branched deep neural network for restoring images from under-display cameras. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 1950-1959 (2022) +16. Kong, L., Dong, J., Ge, J., Li, M., Pan, J.: Efficient frequency domain-based transformers for high-quality image deblurring. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 5886-5895 (2023) +17. Kwon, K., Kang, E., Lee, S., Lee, S.J., Lee, H.E., Yoo, B., Han, J.J.: Controllable image restoration for under-display camera in smartphones. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 2073-2082 (2021) +18. Lee, W., Son, S., Lee, K.M.: Ap-bsn: Self-supervised denoising for real-world images via asymmetric pd and blind-spot network. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 17725-17734 (2022) +19. Liang, C.H., Chen, Y.A., Liu, Y.C., Hsu, W.H.: Raw image deblurring. IEEE Transactions on Multimedia 24, 61-72 (2020) +20. Loshchilov, I., Hutter, F.: Sgdr: Stochastic gradient descent with warm restarts. arXiv preprint arXiv:1608.03983 (2016) +21. Van der Maaten, L., Hinton, G.: Visualizing data using t-sne. Journal of machine learning research 9(11) (2008) +22. Nam, S., Hwang, Y., Matsushita, Y., Kim, S.J.: A holistic approach to cross-channel image noise modeling and its application to image denoising. In: Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition. pp. 1683-1691 (2016) +23. Oh, Y., Park, G.Y., Cho, N.I.: Restoration of high-frequency components in under display camera images. In: 2022 Asia-Pacific Signal and Information Processing Association Annual Summit and Conference (APSIPA ASC). pp. 1040-1046. IEEE (2022) +24. Oh, Y., Park, G.Y., Chung, H., Cho, S., Cho, N.I.: Residual dilated u-net with spatially adaptive normalization for the restoration of under display camera images. In: 2021 Asia-Pacific Signal and Information Processing Association Annual Summit and Conference (APSIPA ASC). pp. 151-157. IEEE (2021) +25. Plotz, T., Roth, S.: Benchmarking denoising algorithms with real photographs. In: Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition. pp. 1586-1595 (2017) +26. Punnappurath, A., Abuolaim, A., Abdelhamed, A., Levinshtein, A., Brown, M.S.: Day-to-night image synthesis for training nighttime neural isps. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 10769-10778 (2022) +27. Ronneberger, O., Fischer, P., Brox, T.: U-net: Convolutional networks for biomedical image segmentation. In: Medical Image Computing and Computer-Assisted Intervention-MICCAI 2015: 18th International Conference, Munich, Germany, October 5-9, 2015, Proceedings, Part III 18. pp. 234-241. Springer (2015) +28. Schroff, F., Kalenichenko, D., Philbin, J.: Facenet: A unified embedding for face recognition and clustering. In: Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition. pp. 815-823 (2015) + +29. Schwartz, E., Giryes, R., Bronstein, A.M.: Deepisp: Toward learning an end-to-end image processing pipeline. IEEE Transactions on Image Processing 28(2), 912-923 (2018) +30. Song, B., Chen, X., Xu, S., Zhou, J.: Under-display camera image restoration with scattering effect. In: Proceedings of the IEEE/CVF International Conference on Computer Vision. pp. 12580-12589 (2023) +31. Sundar, V., Hegde, S., Kothandaraman, D., Mitra, K.: Deep atrous guided filter for image restoration in under display cameras. In: Computer Vision-ECCV 2020 Workshops: Glasgow, UK, August 23-28, 2020, Proceedings, Part V 16. pp. 379-397. Springer (2020) +32. Wang, L., Wang, Y., Dong, X., Xu, Q., Yang, J., An, W., Guo, Y.: Unsupervised degradation representation learning for blind super-resolution. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 10581-10590 (2021) +33. Yu, B., Liu, T., Gong, M., Ding, C., Tao, D.: Correcting the triplet selection bias for triplet loss. In: Proceedings of the European Conference on Computer Vision (ECCV). pp. 71-87 (2018) +34. Yuan, Y., Chen, W., Yang, Y., Wang, Z.: In defense of the triplet loss again: Learning robust person re-identification with fast approximated triplet loss and label distillation. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition Workshops. pp. 354-355 (2020) +35. Zamir, S.W., Arora, A., Khan, S., Hayat, M., Khan, F.S., Yang, M.H., Shao, L.: Cycleisp: Real image restoration via improved data synthesis. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 2696-2705 (2020) +36. Zhou, Y., Song, Y., Du, X.: Modular degradation simulation and restoration for under-display camera. In: Proceedings of the Asian Conference on Computer Vision. pp. 265-282 (2022) +37. Zhou, Y., Kwan, M., Tolentino, K., Emerton, N., Lim, S., Large, T., Fu, L., Pan, Z., Li, B., Yang, Q., et al.: Udc 2020 challenge on image restoration of under-display camera: Methods and results. In: Computer Vision-ECCV 2020 Workshops: Glasgow, UK, August 23-28, 2020, Proceedings, Part V 16. pp. 337-351. Springer (2020) +38. Zhou, Y., Ren, D., Emerton, N., Lim, S., Large, T.: Image restoration for underdisplay camera. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 9179-9188 (2021) \ No newline at end of file diff --git a/2024/Panel-Specific Degradation Representation for Raw Under-Display Camera Image Restoration/images.zip b/2024/Panel-Specific Degradation Representation for Raw Under-Display Camera Image Restoration/images.zip new file mode 100644 index 0000000000000000000000000000000000000000..0c7cd9dd249bc59ccc643eb53398d93d49f48225 --- /dev/null +++ b/2024/Panel-Specific Degradation Representation for Raw Under-Display Camera Image Restoration/images.zip @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:5be1182962a9bec9e960c38d90b30be94aa933e5edb9b5e774c221326efeb3e1 +size 553915 diff --git a/2024/Panel-Specific Degradation Representation for Raw Under-Display Camera Image Restoration/layout.json b/2024/Panel-Specific Degradation Representation for Raw Under-Display Camera Image Restoration/layout.json new file mode 100644 index 0000000000000000000000000000000000000000..4293ecd9aed37bf1332fe6f3f22ceac26d98d17d --- /dev/null +++ b/2024/Panel-Specific Degradation Representation for Raw Under-Display Camera Image Restoration/layout.json @@ -0,0 +1,11387 @@ +{ + "pdf_info": [ + { + "para_blocks": [ + { + "bbox": [ + 141, + 111, + 473, + 148 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 141, + 111, + 473, + 148 + ], + "spans": [ + { + "bbox": [ + 141, + 111, + 473, + 148 + ], + "type": "text", + "content": "Panel-Specific Degradation Representation for Raw Under-Display Camera Image Restoration" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 141, + 167, + 473, + 192 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 141, + 167, + 473, + 192 + ], + "spans": [ + { + "bbox": [ + 141, + 167, + 473, + 192 + ], + "type": "text", + "content": "Youngjin Oh" + }, + { + "bbox": [ + 141, + 167, + 473, + 192 + ], + "type": "inline_equation", + "content": "^{1}" + }, + { + "bbox": [ + 141, + 167, + 473, + 192 + ], + "type": "text", + "content": ", Keuntek Lee" + }, + { + "bbox": [ + 141, + 167, + 473, + 192 + ], + "type": "inline_equation", + "content": "^{1}" + }, + { + "bbox": [ + 141, + 167, + 473, + 192 + ], + "type": "text", + "content": ", Jooyoung Lee" + }, + { + "bbox": [ + 141, + 167, + 473, + 192 + ], + "type": "inline_equation", + "content": "^{3}" + }, + { + "bbox": [ + 141, + 167, + 473, + 192 + ], + "type": "text", + "content": ", Dae-Hyun Lee" + }, + { + "bbox": [ + 141, + 167, + 473, + 192 + ], + "type": "inline_equation", + "content": "^{3}" + }, + { + "bbox": [ + 141, + 167, + 473, + 192 + ], + "type": "text", + "content": ", and Nam Ik Cho" + }, + { + "bbox": [ + 141, + 167, + 473, + 192 + ], + "type": "inline_equation", + "content": "^{1,2}" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 157, + 201, + 454, + 257 + ], + "type": "list", + "angle": 0, + "index": 5, + "blocks": [ + { + "bbox": [ + 157, + 201, + 454, + 223 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 157, + 201, + 454, + 223 + ], + "spans": [ + { + "bbox": [ + 157, + 201, + 454, + 223 + ], + "type": "inline_equation", + "content": "^{1}" + }, + { + "bbox": [ + 157, + 201, + 454, + 223 + ], + "type": "text", + "content": " Department of ECE & INMC, Seoul National University, Seoul, Korea {yjymoh0211,leekt000,nicho}@snu.ac.kr" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 209, + 224, + 404, + 235 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 209, + 224, + 404, + 235 + ], + "spans": [ + { + "bbox": [ + 209, + 224, + 404, + 235 + ], + "type": "inline_equation", + "content": "^{2}" + }, + { + "bbox": [ + 209, + 224, + 404, + 235 + ], + "type": "text", + "content": " IPAI, Seoul National University, Seoul, Korea" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 223, + 235, + 391, + 257 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 223, + 235, + 391, + 257 + ], + "spans": [ + { + "bbox": [ + 223, + 235, + 391, + 257 + ], + "type": "text", + "content": "3 SK hynix Inc., Korea {jooyoung2.lee, daehyun1.lee}@sk.com" + } + ] + } + ], + "index": 4 + } + ], + "sub_type": "text" + }, + { + "bbox": [ + 160, + 285, + 452, + 516 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 160, + 285, + 452, + 516 + ], + "spans": [ + { + "bbox": [ + 160, + 285, + 452, + 516 + ], + "type": "text", + "content": "Abstract. Under-display camera (UDC) image restoration aims to restore images distorted by the OLED display panel covering the frontal camera on a smartphone. Previous deep learning-based UDC restoration methods focused on restoring the image within the RGB domain with the collection of real or synthetic RGB datasets. However, UDC images in these datasets exhibit domain differences from real commercial smartphone UDC images while inherently constraining the problem and solution within the RGB domain. To address this issue, we collect well-aligned sensor-level real UDC images using panels from two commercial smartphones equipped with UDC. We also propose a new UDC restoration method to exploit the disparities between degradations caused by different panels, considering that UDC degradations are specific to the type of OLED panel. For this purpose, we train an encoder with an unsupervised learning scheme using triplet loss that aims to extract the inherent degradations caused by different panels from degraded UDC images as implicit representations. The learned panel-specific degradation representations are then provided as priors to our restoration network based on an efficient Transformer network. Extensive experiments show that our proposed method achieves state-of-the-art performance on our real raw image dataset and generalizes well to previous datasets. Our dataset and code is available at https://github.com/OBAKSA/DREUDC." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 159, + 526, + 453, + 549 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 159, + 526, + 453, + 549 + ], + "spans": [ + { + "bbox": [ + 159, + 526, + 453, + 549 + ], + "type": "text", + "content": "Keywords: Under-Display Camera " + }, + { + "bbox": [ + 159, + 526, + 453, + 549 + ], + "type": "inline_equation", + "content": "\\cdot" + }, + { + "bbox": [ + 159, + 526, + 453, + 549 + ], + "type": "text", + "content": " Image Restoration " + }, + { + "bbox": [ + 159, + 526, + 453, + 549 + ], + "type": "inline_equation", + "content": "\\cdot" + }, + { + "bbox": [ + 159, + 526, + 453, + 549 + ], + "type": "text", + "content": " Representation Learning" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 132, + 569, + 230, + 582 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 569, + 230, + 582 + ], + "spans": [ + { + "bbox": [ + 132, + 569, + 230, + 582 + ], + "type": "text", + "content": "1 Introduction" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 130, + 594, + 482, + 666 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 594, + 482, + 666 + ], + "spans": [ + { + "bbox": [ + 130, + 594, + 482, + 666 + ], + "type": "text", + "content": "To hide the camera hole in the frontal side of smartphones, under-display camera (UDC) systems have been developed, enabled by recent advancements in imaging and display panel technologies. With the UDC technology, smartphone users can enjoy a full-screen display without any camera holes or notches on the front side of their device. However, since the camera is hidden under the panel, the light that reaches the camera is altered by the complex layers of the panel, causing" + } + ] + } + ], + "index": 9 + } + ], + "discarded_blocks": [], + "page_size": [ + 612, + 792 + ], + "page_idx": 0 + }, + { + "para_blocks": [ + { + "type": "table", + "bbox": [ + 140, + 135, + 470, + 187 + ], + "blocks": [ + { + "bbox": [ + 134, + 114, + 478, + 126 + ], + "lines": [ + { + "bbox": [ + 134, + 114, + 478, + 126 + ], + "spans": [ + { + "bbox": [ + 134, + 114, + 478, + 126 + ], + "type": "text", + "content": "Table 1: Comparison of previous datasets [9,10,30,38] and our commercial dataset." + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 140, + 135, + 470, + 187 + ], + "lines": [ + { + "bbox": [ + 140, + 135, + 470, + 187 + ], + "spans": [ + { + "bbox": [ + 140, + 135, + 470, + 187 + ], + "type": "table", + "html": "
DatasetZhou et al. [38]Feng et al. [10]Feng et al. [9]Song et al. [30]Ours
FormatRGBRGBRGBRGBRGB/RAW
UDC ModelingReal (prototype)SyntheticRealSyntheticReal
Aligned GT✓✓✗✓✓
Multiple Display✓✗✗✗✓
", + "image_path": "55f3bc253a196e3460cf0c79f466c984027730b3af48f2349a4d18eed6efe7d9.jpg" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "table_body" + } + ], + "index": 3 + }, + { + "type": "image", + "bbox": [ + 195, + 201, + 240, + 247 + ], + "blocks": [ + { + "bbox": [ + 195, + 201, + 240, + 247 + ], + "lines": [ + { + "bbox": [ + 195, + 201, + 240, + 247 + ], + "spans": [ + { + "bbox": [ + 195, + 201, + 240, + 247 + ], + "type": "image", + "image_path": "d948b0a515a4d41a88c7fbc0d82e2f66ae1c6d20a86b66a90d3aa0bcf36e02d1.jpg" + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 193, + 247, + 283, + 258 + ], + "lines": [ + { + "bbox": [ + 193, + 247, + 283, + 258 + ], + "spans": [ + { + "bbox": [ + 193, + 247, + 283, + 258 + ], + "type": "text", + "content": "(a) Feng et al. [9], GT" + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "image_caption" + } + ], + "index": 4 + }, + { + "type": "image", + "bbox": [ + 242, + 203, + 279, + 245 + ], + "blocks": [ + { + "bbox": [ + 242, + 203, + 279, + 245 + ], + "lines": [ + { + "bbox": [ + 242, + 203, + 279, + 245 + ], + "spans": [ + { + "bbox": [ + 242, + 203, + 279, + 245 + ], + "type": "image", + "image_path": "6ffebcb9c500395872be2223822ed775976f904faa0de2179265e107609aa0cf.jpg" + } + ] + } + ], + "index": 6, + "angle": 0, + "type": "image_body" + } + ], + "index": 6 + }, + { + "bbox": [ + 244, + 247, + 282, + 258 + ], + "angle": 0, + "lines": [ + { + "bbox": [ + 244, + 247, + 282, + 258 + ], + "spans": [ + { + "bbox": [ + 244, + 247, + 282, + 258 + ], + "type": "text", + "content": "l. [9], GT" + } + ] + } + ], + "index": 7, + "type": "text" + }, + { + "type": "image", + "bbox": [ + 322, + 201, + 366, + 247 + ], + "blocks": [ + { + "bbox": [ + 322, + 201, + 366, + 247 + ], + "lines": [ + { + "bbox": [ + 322, + 201, + 366, + 247 + ], + "spans": [ + { + "bbox": [ + 322, + 201, + 366, + 247 + ], + "type": "image", + "image_path": "c44ea9092d19ab626987bfc963fe0afc1d6cea238d5dd4ba61c644430f212330.jpg" + } + ] + } + ], + "index": 8, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 317, + 247, + 414, + 258 + ], + "lines": [ + { + "bbox": [ + 317, + 247, + 414, + 258 + ], + "spans": [ + { + "bbox": [ + 317, + 247, + 414, + 258 + ], + "type": "text", + "content": "(b) Feng et al. [9], UDC" + } + ] + } + ], + "index": 9, + "angle": 0, + "type": "image_caption" + } + ], + "index": 8 + }, + { + "type": "image", + "bbox": [ + 369, + 203, + 406, + 245 + ], + "blocks": [ + { + "bbox": [ + 369, + 203, + 406, + 245 + ], + "lines": [ + { + "bbox": [ + 369, + 203, + 406, + 245 + ], + "spans": [ + { + "bbox": [ + 369, + 203, + 406, + 245 + ], + "type": "image", + "image_path": "a07cd45ac42d6e56651124ea01d51ba65f9194188f40d955e0344d4b1a7628b4.jpg" + } + ] + } + ], + "index": 10, + "angle": 0, + "type": "image_body" + } + ], + "index": 10 + }, + { + "bbox": [ + 367, + 247, + 414, + 258 + ], + "angle": 0, + "lines": [ + { + "bbox": [ + 367, + 247, + 414, + 258 + ], + "spans": [ + { + "bbox": [ + 367, + 247, + 414, + 258 + ], + "type": "text", + "content": "al. [9], UDC" + } + ] + } + ], + "index": 11, + "type": "text" + }, + { + "type": "image", + "bbox": [ + 148, + 258, + 194, + 302 + ], + "blocks": [ + { + "bbox": [ + 148, + 258, + 194, + 302 + ], + "lines": [ + { + "bbox": [ + 148, + 258, + 194, + 302 + ], + "spans": [ + { + "bbox": [ + 148, + 258, + 194, + 302 + ], + "type": "image", + "image_path": "f6255a2f4b4f8f70c9e59f1bedbc4a12154fac0f34dba31677604939441d38e0.jpg" + } + ] + } + ], + "index": 12, + "angle": 0, + "type": "image_body" + } + ], + "index": 12 + }, + { + "type": "image", + "bbox": [ + 195, + 259, + 232, + 303 + ], + "blocks": [ + { + "bbox": [ + 141, + 304, + 237, + 314 + ], + "lines": [ + { + "bbox": [ + 141, + 304, + 237, + 314 + ], + "spans": [ + { + "bbox": [ + 141, + 304, + 237, + 314 + ], + "type": "text", + "content": "(d) Feng et al. [10], GT" + } + ] + } + ], + "index": 13, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 195, + 259, + 232, + 303 + ], + "lines": [ + { + "bbox": [ + 195, + 259, + 232, + 303 + ], + "spans": [ + { + "bbox": [ + 195, + 259, + 232, + 303 + ], + "type": "image", + "image_path": "3f4084fbd01901d35c7e56711e655a381b04282122bdd002b908db7c1fcf519c.jpg" + } + ] + } + ], + "index": 14, + "angle": 0, + "type": "image_body" + } + ], + "index": 14 + }, + { + "bbox": [ + 195, + 304, + 237, + 314 + ], + "angle": 0, + "lines": [ + { + "bbox": [ + 195, + 304, + 237, + 314 + ], + "spans": [ + { + "bbox": [ + 195, + 304, + 237, + 314 + ], + "type": "text", + "content": "l. [10], GT" + } + ] + } + ], + "index": 15, + "type": "text" + }, + { + "type": "image", + "bbox": [ + 258, + 259, + 303, + 304 + ], + "blocks": [ + { + "bbox": [ + 258, + 259, + 303, + 304 + ], + "lines": [ + { + "bbox": [ + 258, + 259, + 303, + 304 + ], + "spans": [ + { + "bbox": [ + 258, + 259, + 303, + 304 + ], + "type": "image", + "image_path": "852a895706697611d4b76c13c5b46fb97831956a03d1e9d819fccdffb10d5c7c.jpg" + } + ] + } + ], + "index": 16, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 252, + 304, + 354, + 314 + ], + "lines": [ + { + "bbox": [ + 252, + 304, + 354, + 314 + ], + "spans": [ + { + "bbox": [ + 252, + 304, + 354, + 314 + ], + "type": "text", + "content": "(e) Feng et al. [10], UDC" + } + ] + } + ], + "index": 17, + "angle": 0, + "type": "image_caption" + } + ], + "index": 16 + }, + { + "type": "image", + "bbox": [ + 369, + 259, + 414, + 304 + ], + "blocks": [ + { + "bbox": [ + 369, + 259, + 414, + 304 + ], + "lines": [ + { + "bbox": [ + 369, + 259, + 414, + 304 + ], + "spans": [ + { + "bbox": [ + 369, + 259, + 414, + 304 + ], + "type": "image", + "image_path": "d0f96c5e388cf4d5cb82d4beea9c506c816e7c1ad43d65e6ee4aa8c7b31b07f6.jpg" + } + ] + } + ], + "index": 18, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 360, + 304, + 462, + 314 + ], + "lines": [ + { + "bbox": [ + 360, + 304, + 462, + 314 + ], + "spans": [ + { + "bbox": [ + 360, + 304, + 462, + 314 + ], + "type": "text", + "content": "(f) Song et al. [30], UDC" + } + ] + } + ], + "index": 19, + "angle": 0, + "type": "image_caption" + } + ], + "index": 18 + }, + { + "type": "image", + "bbox": [ + 148, + 316, + 194, + 359 + ], + "blocks": [ + { + "bbox": [ + 148, + 316, + 194, + 359 + ], + "lines": [ + { + "bbox": [ + 148, + 316, + 194, + 359 + ], + "spans": [ + { + "bbox": [ + 148, + 316, + 194, + 359 + ], + "type": "image", + "image_path": "656f2a6209f476ac2a4613a836bbb4c45d6e89bf1c78caf87f96dd154cb4d8a9.jpg" + } + ] + } + ], + "index": 20, + "angle": 0, + "type": "image_body" + } + ], + "index": 20 + }, + { + "type": "image", + "bbox": [ + 195, + 316, + 232, + 358 + ], + "blocks": [ + { + "bbox": [ + 195, + 316, + 232, + 358 + ], + "lines": [ + { + "bbox": [ + 195, + 316, + 232, + 358 + ], + "spans": [ + { + "bbox": [ + 195, + 316, + 232, + 358 + ], + "type": "image", + "image_path": "aacadb8f0484e70271a5c0dc591bda948a147c3661a8a22a0376a7491093fac0.jpg" + } + ] + } + ], + "index": 22, + "angle": 0, + "type": "image_body" + } + ], + "index": 22 + }, + { + "type": "image", + "bbox": [ + 258, + 316, + 303, + 359 + ], + "blocks": [ + { + "bbox": [ + 237, + 361, + 266, + 371 + ], + "lines": [ + { + "bbox": [ + 237, + 361, + 266, + 371 + ], + "spans": [ + { + "bbox": [ + 237, + 361, + 266, + 371 + ], + "type": "text", + "content": "(h) Zho" + } + ] + } + ], + "index": 23, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 258, + 316, + 303, + 359 + ], + "lines": [ + { + "bbox": [ + 258, + 316, + 303, + 359 + ], + "spans": [ + { + "bbox": [ + 258, + 316, + 303, + 359 + ], + "type": "image", + "image_path": "b78df607efece1427d9a1c6fb5dab173aa63b7e57b631e540e202f62b710ed46.jpg" + } + ] + } + ], + "index": 24, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 252, + 361, + 356, + 371 + ], + "lines": [ + { + "bbox": [ + 252, + 361, + 356, + 371 + ], + "spans": [ + { + "bbox": [ + 252, + 361, + 356, + 371 + ], + "type": "text", + "content": "Zhou et al. [38], T-OLED" + } + ] + } + ], + "index": 25, + "angle": 0, + "type": "image_caption" + } + ], + "index": 24 + }, + { + "type": "image", + "bbox": [ + 369, + 316, + 414, + 359 + ], + "blocks": [ + { + "bbox": [ + 369, + 316, + 414, + 359 + ], + "lines": [ + { + "bbox": [ + 369, + 316, + 414, + 359 + ], + "spans": [ + { + "bbox": [ + 369, + 316, + 414, + 359 + ], + "type": "image", + "image_path": "eca13371b9bbecceb84ed27e6ba1cee8f122782f1df832359620a0647d8f5708.jpg" + } + ] + } + ], + "index": 26, + "angle": 0, + "type": "image_body" + } + ], + "index": 26 + }, + { + "type": "image", + "bbox": [ + 416, + 316, + 454, + 358 + ], + "blocks": [ + { + "bbox": [ + 416, + 316, + 454, + 358 + ], + "lines": [ + { + "bbox": [ + 416, + 316, + 454, + 358 + ], + "spans": [ + { + "bbox": [ + 416, + 316, + 454, + 358 + ], + "type": "image", + "image_path": "9e0966a407f913e42c0d6c843a91c52368e6507478027bf8554a990c5da152de.jpg" + } + ] + } + ], + "index": 28, + "angle": 0, + "type": "image_body" + } + ], + "index": 28 + }, + { + "type": "image", + "bbox": [ + 148, + 373, + 194, + 416 + ], + "blocks": [ + { + "bbox": [ + 138, + 361, + 217, + 372 + ], + "lines": [ + { + "bbox": [ + 138, + 361, + 217, + 372 + ], + "spans": [ + { + "bbox": [ + 138, + 361, + 217, + 372 + ], + "type": "text", + "content": "(g) Zhou et al. [38]," + } + ] + } + ], + "index": 21, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 148, + 373, + 194, + 416 + ], + "lines": [ + { + "bbox": [ + 148, + 373, + 194, + 416 + ], + "spans": [ + { + "bbox": [ + 148, + 373, + 194, + 416 + ], + "type": "image", + "image_path": "175e3e054a157fa6399689707e170996e0f102d7afa2ae76bd82a96b99bc5865.jpg" + } + ] + } + ], + "index": 30, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 162, + 417, + 216, + 427 + ], + "lines": [ + { + "bbox": [ + 162, + 417, + 216, + 427 + ], + "spans": [ + { + "bbox": [ + 162, + 417, + 216, + 427 + ], + "type": "text", + "content": "(j) Ours, GT" + } + ] + } + ], + "index": 31, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 132, + 439, + 482, + 463 + ], + "lines": [ + { + "bbox": [ + 132, + 439, + 482, + 463 + ], + "spans": [ + { + "bbox": [ + 132, + 439, + 482, + 463 + ], + "type": "text", + "content": "Fig. 1: Samples and their RGB channel histograms from previous UDC datasets [9, 10,30,38] and our commercial dataset." + } + ] + } + ], + "index": 37, + "angle": 0, + "type": "image_caption" + } + ], + "index": 30 + }, + { + "type": "image", + "bbox": [ + 195, + 373, + 232, + 416 + ], + "blocks": [ + { + "bbox": [ + 195, + 373, + 232, + 416 + ], + "lines": [ + { + "bbox": [ + 195, + 373, + 232, + 416 + ], + "spans": [ + { + "bbox": [ + 195, + 373, + 232, + 416 + ], + "type": "image", + "image_path": "33eb15ba359e820dc42a731fafa38907df0578b97b4b51c168ed0a9761c3298c.jpg" + } + ] + } + ], + "index": 32, + "angle": 0, + "type": "image_body" + } + ], + "index": 32 + }, + { + "type": "image", + "bbox": [ + 258, + 373, + 302, + 416 + ], + "blocks": [ + { + "bbox": [ + 258, + 373, + 302, + 416 + ], + "lines": [ + { + "bbox": [ + 258, + 373, + 302, + 416 + ], + "spans": [ + { + "bbox": [ + 258, + 373, + 302, + 416 + ], + "type": "image", + "image_path": "b79214c6f053dea6614bebde267b73bcbdf0893279ac48494551b4ccadfd0bbc.jpg" + } + ] + } + ], + "index": 33, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 265, + 417, + 340, + 428 + ], + "lines": [ + { + "bbox": [ + 265, + 417, + 340, + 428 + ], + "spans": [ + { + "bbox": [ + 265, + 417, + 340, + 428 + ], + "type": "text", + "content": "(k) Ours, Axon 30" + } + ] + } + ], + "index": 34, + "angle": 0, + "type": "image_caption" + } + ], + "index": 33 + }, + { + "type": "image", + "bbox": [ + 369, + 372, + 414, + 416 + ], + "blocks": [ + { + "bbox": [ + 367, + 361, + 415, + 371 + ], + "lines": [ + { + "bbox": [ + 367, + 361, + 415, + 371 + ], + "spans": [ + { + "bbox": [ + 367, + 361, + 415, + 371 + ], + "type": "text", + "content": ") Zhou et al" + } + ] + } + ], + "index": 27, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 416, + 361, + 475, + 371 + ], + "lines": [ + { + "bbox": [ + 416, + 361, + 475, + 371 + ], + "spans": [ + { + "bbox": [ + 416, + 361, + 475, + 371 + ], + "type": "text", + "content": "[38], P-OLED" + } + ] + } + ], + "index": 29, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 369, + 372, + 414, + 416 + ], + "lines": [ + { + "bbox": [ + 369, + 372, + 414, + 416 + ], + "spans": [ + { + "bbox": [ + 369, + 372, + 414, + 416 + ], + "type": "image", + "image_path": "5cb59ca95b7eb7347ba8642a6a8fcb253962c30ccffa8eb2c6203640970b2d43.jpg" + } + ] + } + ], + "index": 35, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 374, + 417, + 447, + 427 + ], + "lines": [ + { + "bbox": [ + 374, + 417, + 447, + 427 + ], + "spans": [ + { + "bbox": [ + 374, + 417, + 447, + 427 + ], + "type": "text", + "content": "(1) Ours, Z-Fold 3" + } + ] + } + ], + "index": 36, + "angle": 0, + "type": "image_caption" + } + ], + "index": 35 + }, + { + "bbox": [ + 130, + 486, + 480, + 521 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 486, + 480, + 521 + ], + "spans": [ + { + "bbox": [ + 130, + 486, + 480, + 521 + ], + "type": "text", + "content": "severe degradation in the resulting image. Degradations include blur, diffraction, noise, low-lightness, color shift, haziness, etc., and their spatially-variant nature makes the UDC image restoration very challenging [10, 17, 38]." + } + ] + } + ], + "index": 38 + }, + { + "bbox": [ + 130, + 522, + 482, + 666 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 522, + 482, + 666 + ], + "spans": [ + { + "bbox": [ + 130, + 522, + 482, + 666 + ], + "type": "text", + "content": "Most UDC image restoration methods used deep neural networks [9-11, 15, 17, 23, 24, 30, 31, 37, 38], providing promising results on public benchmark datasets [9, 10, 30, 38] gathered to facilitate research on UDC image restoration. However, these existing datasets either use prototype display panels to inadequately represent the degradations observed in real UDC images caused by commercial panels [38], lack realistic degradations due to incomplete UDC modeling (e.g., noise and color shift) in the synthesis pipeline [10,30], or mishandles aligning image pairs with occlusion and parallax due to the mechanism of the capturing imaging system [9]. In addition, the potential for enhanced image restoration performance through sensor-level restoration [1, 19, 25, 35] cannot be exploited, because previous datasets predominantly restrict experiments to RGB domain as raw data is unreleased. Tab. 1 shows comparison of the datasets." + } + ] + } + ], + "index": 39 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "spans": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "type": "text", + "content": "2" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 166, + 91, + 217, + 101 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 166, + 91, + 217, + 101 + ], + "spans": [ + { + "bbox": [ + 166, + 91, + 217, + 101 + ], + "type": "text", + "content": "Y. Oh et al." + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 1 + }, + { + "para_blocks": [ + { + "bbox": [ + 130, + 116, + 479, + 247 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 116, + 479, + 247 + ], + "spans": [ + { + "bbox": [ + 130, + 116, + 479, + 247 + ], + "type": "text", + "content": "In this work, we propose a new UDC dataset that is collected in the raw domain. The proposed dataset contains pairs of real UDC images and clean images that are well-aligned using a monitor-based imaging system with two different panels from commercial UDC smartphones, ZTE Axon 30 5G and Samsung Galaxy Z-Fold 3. In our experiments, we find that training with raw data enhances the performance of UDC image restoration. This raw-domain approach with real data also has the advantage of enabling the use of end-to-end neural ISPs that are being installed in recent smartphones [26], allowing a practical implementation on devices. Additionally, as shown in Fig. 1, real UDC data captured with commercial smartphone panels display noticeable distinct degradations visually and statistically when compared to previous datasets [10,30,38]." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 130, + 248, + 480, + 438 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 248, + 480, + 438 + ], + "spans": [ + { + "bbox": [ + 130, + 248, + 480, + 438 + ], + "type": "text", + "content": "Furthermore, we develop a two-stage restoration method for UDC images. First, we introduce a framework for learning the implicit representation of degradations in the UDC images. We notice that degradation in a patch of a UDC image is similar to that of other patches of the same UDC image while being distinct from degradations of the same scene taken with a different panel. With this observation, we develop a novel unsupervised learning scheme involving triplet loss [28] to train our encoder. The learning scheme aims to leverage the information of the data collected with different types of panels, ultimately condensing the degradation information into an implicit representation vector that reflects the panel-specific degradation moderately. We also propose an efficient Transformer-based restoration network that incorporates the learned panel-specific degradation representations as a prior by generating channel-wise calibration coefficients for attention. We discover that our strategy yields improved results for UDC image restoration. To the best of our knowledge, this is the first case in UDC image restoration where a dataset collected with one panel is utilized to enhance the performance on a dataset collected with another panel." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 146, + 439, + 356, + 450 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 146, + 439, + 356, + 450 + ], + "spans": [ + { + "bbox": [ + 146, + 439, + 356, + 450 + ], + "type": "text", + "content": "To summarize, our contributions are as follows:" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 138, + 455, + 480, + 584 + ], + "type": "list", + "angle": 0, + "index": 8, + "blocks": [ + { + "bbox": [ + 138, + 455, + 479, + 490 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 455, + 479, + 490 + ], + "spans": [ + { + "bbox": [ + 138, + 455, + 479, + 490 + ], + "type": "text", + "content": "- We introduce a new real UDC dataset, which is collected with commercial smartphone panels. Our dataset is provided in raw format, which can improve the performance of the UDC restoration task." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 138, + 491, + 480, + 548 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 491, + 480, + 548 + ], + "spans": [ + { + "bbox": [ + 138, + 491, + 480, + 548 + ], + "type": "text", + "content": "- We propose an unsupervised learning approach that allows for implicit representations of degradations caused by various display panels in the embedding space. The panel-specific degradation representations are leveraged as priors in the restoration network, which is an efficient Transformer and brings notable improvement." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 138, + 550, + 479, + 584 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 550, + 479, + 584 + ], + "spans": [ + { + "bbox": [ + 138, + 550, + 479, + 584 + ], + "type": "text", + "content": "- Extensive experiments demonstrate that our restoration network embedded with learned representations achieves state-of-the-art performance on the proposed dataset and is consistently applicable to previous datasets." + } + ] + } + ], + "index": 7 + } + ], + "sub_type": "text" + }, + { + "bbox": [ + 132, + 601, + 242, + 613 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 601, + 242, + 613 + ], + "spans": [ + { + "bbox": [ + 132, + 601, + 242, + 613 + ], + "type": "text", + "content": "2 Related Works" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 132, + 624, + 284, + 635 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 624, + 284, + 635 + ], + "spans": [ + { + "bbox": [ + 132, + 624, + 284, + 635 + ], + "type": "text", + "content": "2.1 UDC Image Restoration" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 130, + 641, + 479, + 665 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 641, + 479, + 665 + ], + "spans": [ + { + "bbox": [ + 130, + 641, + 479, + 665 + ], + "type": "text", + "content": "Since UDC image restoration is a relatively new area, research is still ongoing to provide users with a satisfying experience when UDC is applied in real situa-" + } + ] + } + ], + "index": 11 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 288, + 91, + 447, + 102 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 288, + 91, + 447, + 102 + ], + "spans": [ + { + "bbox": [ + 288, + 91, + 447, + 102 + ], + "type": "text", + "content": "Panel-Specific UDC Image Restoration" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 474, + 92, + 480, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 474, + 92, + 480, + 100 + ], + "spans": [ + { + "bbox": [ + 474, + 92, + 480, + 100 + ], + "type": "text", + "content": "3" + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 2 + }, + { + "para_blocks": [ + { + "bbox": [ + 130, + 116, + 482, + 247 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 116, + 482, + 247 + ], + "spans": [ + { + "bbox": [ + 130, + 116, + 482, + 247 + ], + "type": "text", + "content": "tions. Research on UDC image restoration was sparked by the pioneering work of [38]. They defined a UDC imaging system and released the first UDC dataset for researchers to experiment with. Synthetic datasets were also made available by simulating the camera pipeline using a pre-calculated point spread function (PSF), which adequately represents the diffraction [10], with further improvements on scattering effects in [30] to account for the haziness of UDC images. More recently, a real UDC dataset generated by pseudo-aligning geometrically misaligned image pairs for a single UDC panel [9] was presented. Although the proposed datasets have some drawbacks, several works [9-11, 15, 17, 23, 24, 30, 31, 37, 38] have investigated UDC image restoration using deep neural networks and have shown promising results." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 130, + 248, + 482, + 403 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 248, + 482, + 403 + ], + "spans": [ + { + "bbox": [ + 130, + 248, + 482, + 403 + ], + "type": "text", + "content": "Among these deep-learning-based methods, several of them use priors to improve performance. For instance, in [10], PCA is employed on a PSF of the UDC display panel to stretch it as a 'kernel code' and condition the network with it as a prior. In [30], the network is divided into a scattering branch and an image branch, and the former is trained to estimate the scattering parameters, which is prior information that can be used. However, these methods require precise priors; otherwise, the performance is degraded. Our approach differs from previous methods because we do not need strict parameters or ground truth settings. Instead, we aim to learn an acceptable representation of the degradation induced by panel-specific properties in the embedding space rather than a rigorous estimation of the degradation. We only need images taken with different panels because our method is unsupervised, and learn the implicit priors by leveraging the discrepancies of degradation caused by different panels." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 131, + 419, + 334, + 431 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 131, + 419, + 334, + 431 + ], + "spans": [ + { + "bbox": [ + 131, + 419, + 334, + 431 + ], + "type": "text", + "content": "2.2 Image Restoration with Raw Data" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 130, + 438, + 482, + 666 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 438, + 482, + 666 + ], + "spans": [ + { + "bbox": [ + 130, + 438, + 482, + 666 + ], + "type": "text", + "content": "Most of the previous research on image restoration relies on RGB images as training data as they are easier to collect. However, some recent studies have shown that using raw sensor data can lead to improved performance. Specifically, [1, 25] are the works on image denoising where they collected DND and SIDD datasets that are available as raw and RGB images, where denoising with raw images usually yields higher PSNR. CycleISP [35] demonstrated that noise is more complex in RGB images and showed that noise modeling is more realistic in the sensor domain. This is because compared to the noise samples in the sensor pixels, which are nearly uncorrelated to each other, the noise samples become highly correlated in RGB images as they pass through the camera ISP pipeline, which involves processes like demosaicing, gamma correction, and white balancing [18, 22]. DeepISP [29] proposed an end-to-end camera ISP modeled with deep neural networks, suggesting that a deep network-based camera ISP is promising. Furthermore, [4] solved the issue of low-light enhancement problem by operating in the sensor domain, and [19] presented a dataset for raw image deblurring and proved that operating on raw sensor data achieves better performance than RGB-based methods. These researches suggest that it is easier to restore raw data than RGB images with deep neural networks, as sensor data contains the linear representation of incoming light intensity of the scene. This ensures that" + } + ] + } + ], + "index": 5 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "spans": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "type": "text", + "content": "4" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 166, + 91, + 217, + 101 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 166, + 91, + 217, + 101 + ], + "spans": [ + { + "bbox": [ + 166, + 91, + 217, + 101 + ], + "type": "text", + "content": "Y. Oh et al." + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 3 + }, + { + "para_blocks": [ + { + "bbox": [ + 130, + 116, + 479, + 140 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 116, + 479, + 140 + ], + "spans": [ + { + "bbox": [ + 130, + 116, + 479, + 140 + ], + "type": "text", + "content": "the scene's authenticity is maintained more faithfully in raw data than in heavily processed RGB pixel values that are nonlinear with respect to light intensity." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 131, + 158, + 261, + 172 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 131, + 158, + 261, + 172 + ], + "spans": [ + { + "bbox": [ + 131, + 158, + 261, + 172 + ], + "type": "text", + "content": "3 Proposed Method" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 131, + 182, + 319, + 194 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 131, + 182, + 319, + 194 + ], + "spans": [ + { + "bbox": [ + 131, + 182, + 319, + 194 + ], + "type": "text", + "content": "3.1 Sensor-level Real UDC Dataset" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 130, + 201, + 482, + 285 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 201, + 482, + 285 + ], + "spans": [ + { + "bbox": [ + 130, + 201, + 482, + 285 + ], + "type": "text", + "content": "A raw dataset is necessary to achieve end-to-end image restoration using sensor images. However, the available UDC datasets are limited to the RGB domain. Moreover, public UDC datasets do not show the types of degradation that are present in images taken with commercially available smartphones equipped with UDC, limiting their assessment of practical scenarios. Therefore, we capture real raw UDC data by using OLED panels obtained from commercial UDC smartphones, a smartphone camera, and a monitor-based imaging system [38]." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 130, + 301, + 482, + 492 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 301, + 482, + 492 + ], + "spans": [ + { + "bbox": [ + 130, + 301, + 482, + 492 + ], + "type": "text", + "content": "UDC data acquisition process Capturing a raw UDC image requires that we model the UDC imaging system as similarly as possible. Consequently, our data-capturing process involves rendering an image onto a high-resolution monitor, followed by capturing the displayed image using a smartphone camera sensor located closely behind a UDC panel. The panels used to capture the data are from two commercial UDC smartphones: ZTE Axon 30 5G and Samsung Galaxy Z-Fold 3. They introduce low-light/color shift, scattering, diffraction, and blur. As noise is usually generated from the sensor, we chose Samsung Galaxy S23 Ultra as our camera device to capture the authentic noise of smartphone camera sensors. Raw images are collected using the internal software of the smartphone that automatically saves the minimally processed raw data. The preference for monitor-displayed images [17,38] over real natural scenes is driven by the need to control additional factors beyond the panels' presence. Specifically, we can enforce the target scene to be static as dynamic scenes are prone to motion blur, and therefore lead to misaligned data pairs. We can also eliminate interference from external ambient sources of light other than the monitor's emitted light." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 130, + 493, + 482, + 578 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 493, + 482, + 578 + ], + "spans": [ + { + "bbox": [ + 130, + 493, + 482, + 578 + ], + "type": "text", + "content": "The monitor, UDC panel, and camera sensor are aligned in sequence, and the principal axis of the camera is aligned perpendicular to the plane of the monitor and the panel. The camera is mounted on a sturdy tripod, and the panel is situated close to the sensor using a firm holding device. The displayed images are then captured remotely with/without the smartphone panels to ensure the alignment of clean/UDC image pairs during capture while keeping the camera settings the same. We provide more details in the Supplementary Material." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 130, + 594, + 482, + 666 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 594, + 482, + 666 + ], + "spans": [ + { + "bbox": [ + 130, + 594, + 482, + 666 + ], + "type": "text", + "content": "Our UDC dataset The dataset is captured using 400 images from the DIV2K dataset [2], which is comprised of natural scene images from various themes, encompassing diverse objects captured under different lighting conditions. This makes our dataset 1,200 images in total, as there are three different image sets (Axon 30, Z-Fold 3, ground truth) in raw format. The images are stored in 16 bits with a Bayer pattern of G-B-R-G. For the ground truth images, we capture" + } + ] + } + ], + "index": 8 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 288, + 91, + 447, + 102 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 288, + 91, + 447, + 102 + ], + "spans": [ + { + "bbox": [ + 288, + 91, + 447, + 102 + ], + "type": "text", + "content": "Panel-Specific UDC Image Restoration" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 474, + 92, + 480, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 474, + 92, + 480, + 100 + ], + "spans": [ + { + "bbox": [ + 474, + 92, + 480, + 100 + ], + "type": "text", + "content": "5" + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 4 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 173, + 118, + 253, + 199 + ], + "blocks": [ + { + "bbox": [ + 173, + 118, + 253, + 199 + ], + "lines": [ + { + "bbox": [ + 173, + 118, + 253, + 199 + ], + "spans": [ + { + "bbox": [ + 173, + 118, + 253, + 199 + ], + "type": "image", + "image_path": "8d7838f33a0fc9b19804d0939e843031ad78f3fb51bac4fbeaf68210bf52f226.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_body" + } + ], + "index": 2 + }, + { + "type": "image", + "bbox": [ + 173, + 200, + 253, + 281 + ], + "blocks": [ + { + "bbox": [ + 173, + 200, + 253, + 281 + ], + "lines": [ + { + "bbox": [ + 173, + 200, + 253, + 281 + ], + "spans": [ + { + "bbox": [ + 173, + 200, + 253, + 281 + ], + "type": "image", + "image_path": "86ac0c9bc19cf47552625d9c3439435a1bffb1f18c2061258c88989b772b2d54.jpg" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 178, + 281, + 244, + 291 + ], + "lines": [ + { + "bbox": [ + 178, + 281, + 244, + 291 + ], + "spans": [ + { + "bbox": [ + 178, + 281, + 244, + 291 + ], + "type": "text", + "content": "(a) Clean image" + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 130, + 302, + 482, + 369 + ], + "lines": [ + { + "bbox": [ + 130, + 302, + 482, + 369 + ], + "spans": [ + { + "bbox": [ + 130, + 302, + 482, + 369 + ], + "type": "text", + "content": "Fig. 2: Cropped examples from our captured dataset. All images are visualized with the same procedure that includes normalization, demosaicing, channel gain adjustment, and gamma correction. Due to the difference in smartphone display panels, noise and haze are more apparent in Axon 30 UDC images, while the Z-Fold 3 UDC images show blurrier results. The diffraction induces flare artifacts when the image intensity becomes too strong that the sensor saturates." + } + ] + } + ], + "index": 11, + "angle": 0, + "type": "image_caption" + } + ], + "index": 3 + }, + { + "type": "image", + "bbox": [ + 266, + 118, + 348, + 199 + ], + "blocks": [ + { + "bbox": [ + 266, + 118, + 348, + 199 + ], + "lines": [ + { + "bbox": [ + 266, + 118, + 348, + 199 + ], + "spans": [ + { + "bbox": [ + 266, + 118, + 348, + 199 + ], + "type": "image", + "image_path": "f4e1113a491cc030422183607632a0f7bac0a31aaa6b9e9f1806dd50a70b365f.jpg" + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "image_body" + } + ], + "index": 5 + }, + { + "type": "image", + "bbox": [ + 266, + 200, + 347, + 281 + ], + "blocks": [ + { + "bbox": [ + 266, + 200, + 347, + 281 + ], + "lines": [ + { + "bbox": [ + 266, + 200, + 347, + 281 + ], + "spans": [ + { + "bbox": [ + 266, + 200, + 347, + 281 + ], + "type": "image", + "image_path": "17fdc077e18fb78a35e80a41fead853513427d56d89a7d63ed5091f8ef20e8fe.jpg" + } + ] + } + ], + "index": 6, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 266, + 281, + 343, + 291 + ], + "lines": [ + { + "bbox": [ + 266, + 281, + 343, + 291 + ], + "spans": [ + { + "bbox": [ + 266, + 281, + 343, + 291 + ], + "type": "text", + "content": "(b) Axon 30 image" + } + ] + } + ], + "index": 7, + "angle": 0, + "type": "image_caption" + } + ], + "index": 6 + }, + { + "type": "image", + "bbox": [ + 360, + 118, + 441, + 199 + ], + "blocks": [ + { + "bbox": [ + 360, + 118, + 441, + 199 + ], + "lines": [ + { + "bbox": [ + 360, + 118, + 441, + 199 + ], + "spans": [ + { + "bbox": [ + 360, + 118, + 441, + 199 + ], + "type": "image", + "image_path": "ed02638ba7fa4724740b8648c53514befc953003693debced1e1d41f68a3fc53.jpg" + } + ] + } + ], + "index": 8, + "angle": 0, + "type": "image_body" + } + ], + "index": 8 + }, + { + "type": "image", + "bbox": [ + 360, + 200, + 441, + 281 + ], + "blocks": [ + { + "bbox": [ + 360, + 200, + 441, + 281 + ], + "lines": [ + { + "bbox": [ + 360, + 200, + 441, + 281 + ], + "spans": [ + { + "bbox": [ + 360, + 200, + 441, + 281 + ], + "type": "image", + "image_path": "dedcff742a8060f4fa0c23ca613882a2884f8fc74f73d5220d5e832fa096a21c.jpg" + } + ] + } + ], + "index": 9, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 362, + 281, + 438, + 291 + ], + "lines": [ + { + "bbox": [ + 362, + 281, + 438, + 291 + ], + "spans": [ + { + "bbox": [ + 362, + 281, + 438, + 291 + ], + "type": "text", + "content": "(c) Z-Fold 3 image" + } + ] + } + ], + "index": 10, + "angle": 0, + "type": "image_caption" + } + ], + "index": 9 + }, + { + "bbox": [ + 130, + 391, + 480, + 427 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 391, + 480, + 427 + ], + "spans": [ + { + "bbox": [ + 130, + 391, + 480, + 427 + ], + "type": "text", + "content": "bursts of a scene without the display panel and average them. To check for misalignments that can be induced by optical image stabilization in the smartphone hardware [1], we utilize registration and feature matching algorithms [3, 12]." + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 130, + 427, + 482, + 594 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 427, + 482, + 594 + ], + "spans": [ + { + "bbox": [ + 130, + 427, + 482, + 594 + ], + "type": "text", + "content": "Samples from our dataset are displayed in Fig. 2, with each UDC image taken with Axon 30 and Z-Fold 3 display panels exhibiting panel-specific degradations. As such, our dataset can be used to evaluate if a restoration method generally performs well in the restoration of various UDC systems. Also, our real commercial UDC data show different behaviors from previous datasets, as shown in Fig. 1. Specifically, low-light, color shift, blur, and noise are observed, and an object is seen repeated faintly near the main object due to diffraction. We also detect a phenomenon similar to haze due to the scattering effects of the panel [30]. Synthetic datasets that are generated using UDC modeling [10, 30] do not take noise into account. However, including noise in UDC images is vital as the level of noise captured with frontal smartphone camera sensors, which are relatively inferior to DSLR camera sensors [4, 26], is amplified when normalizing sensor data. This amplification is related to the attenuation caused by the panel which reduces light intensity. More samples are available in the Supplementary." + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 131, + 611, + 359, + 622 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 131, + 611, + 359, + 622 + ], + "spans": [ + { + "bbox": [ + 131, + 611, + 359, + 622 + ], + "type": "text", + "content": "3.2 Proposed UDC Restoration Framework" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 130, + 629, + 481, + 665 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 629, + 481, + 665 + ], + "spans": [ + { + "bbox": [ + 130, + 629, + 481, + 665 + ], + "type": "text", + "content": "In this section, we propose our restoration framework called Degradation Representation Embedded Transformer for UDC image restoration (DREUDC), which consists of a degradation representation encoder and a restoration network. We" + } + ] + } + ], + "index": 15 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "spans": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "type": "text", + "content": "6" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 166, + 91, + 217, + 100 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 166, + 91, + 217, + 100 + ], + "spans": [ + { + "bbox": [ + 166, + 91, + 217, + 100 + ], + "type": "text", + "content": "Y. Oh et al." + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 5 + }, + { + "para_blocks": [ + { + "bbox": [ + 130, + 116, + 482, + 189 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 116, + 482, + 189 + ], + "spans": [ + { + "bbox": [ + 130, + 116, + 482, + 189 + ], + "type": "text", + "content": "first train the encoder with our proposed unsupervised scheme using triplet loss to implicitly learn the degradations of a display panel. Then, we utilize the representations generated from the encoder by integrating them in the restoration network as a prior. It is worth noting that our approach's novelty lies in how we take advantage of UDC data from one display panel to restore a UDC image from another panel rather than in the restoration network's architecture itself." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 130, + 205, + 483, + 350 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 205, + 483, + 350 + ], + "spans": [ + { + "bbox": [ + 130, + 205, + 483, + 350 + ], + "type": "text", + "content": "Motivation Representation learning has been demonstrated to be effective in blind super-resolution, as demonstrated in [32]. Inspired by this, we explore the use of representation learning to enhance the restoration of UDC images by extracting useful prior information about the panels. Since each UDC system has a different panel, it yields panel-specific degradations. For example, the UDC images presented in Fig. 2, which are captured using two different panels of UDC smartphones, show differences in the degree of degradations. This implies that the degradations present in an image captured using one panel are similar to each other and that they differ significantly from those of the image with the same content but are captured using another panel. Using this information, our approach intends to utilize these differences and train an encoder to encode them in an unsupervised manner into the embedding space." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 130, + 366, + 482, + 473 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 366, + 482, + 473 + ], + "spans": [ + { + "bbox": [ + 130, + 366, + 482, + 473 + ], + "type": "text", + "content": "Unsupervised representation encoder training Our proposed unsupervised training scheme is visualized in Fig. 3. As mentioned above, the objective of training the panel-specific degradation representation encoder is to learn the unique properties of the panels and produce implicit representations of the degradations to be used as priors in UDC image restoration. To achieve this, the encoder should learn the degradations using attributes related to a specific panel, such as the level of noise, blur, and color shift. On the other hand, the encoder should be discouraged from learning degradations with features in an image that are panel-agnostic, such as textures and contents of a scene." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 130, + 474, + 482, + 628 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 474, + 482, + 628 + ], + "spans": [ + { + "bbox": [ + 130, + 474, + 482, + 628 + ], + "type": "text", + "content": "We believe that this condition is analogous to the situation of tasks where triplet loss [28] is exploited to learn favorable representations. Triplet loss is widely applied to various tasks of computer vision such as facial recognition [28], person re-identification [13,34], and domain adaptation [8] for its ability to learn useful embeddings by leveraging the distances between embeddings of similar instances and dissimilar instances. For instance, in person re-identification, networks are guided to identify among different people using features that are unique to a specific person, not with features that are shared among people, by learning suitable representations using triplet loss. In our case, the features unique to a specific person would correspond to properties unique to a specific panel, and the features that are shared among people would correspond to panel-agnostic features. Therefore, we opt to apply triplet loss in the training of our encoder to capture the panel-specific degradation representations." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 130, + 630, + 482, + 666 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 630, + 482, + 666 + ], + "spans": [ + { + "bbox": [ + 130, + 630, + 482, + 666 + ], + "type": "text", + "content": "Nevertheless, triplet loss is only effective when the network is provided with valid triplets [28, 33]. To collect meaningful triplets, we impose constraints on the training pairs that will be used to train our encoder. We choose a pair of" + } + ] + } + ], + "index": 6 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 288, + 91, + 447, + 102 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 288, + 91, + 447, + 102 + ], + "spans": [ + { + "bbox": [ + 288, + 91, + 447, + 102 + ], + "type": "text", + "content": "Panel-Specific UDC Image Restoration" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 474, + 92, + 480, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 474, + 92, + 480, + 100 + ], + "spans": [ + { + "bbox": [ + 474, + 92, + 480, + 100 + ], + "type": "text", + "content": "7" + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 6 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 145, + 120, + 216, + 198 + ], + "blocks": [ + { + "bbox": [ + 145, + 120, + 216, + 198 + ], + "lines": [ + { + "bbox": [ + 145, + 120, + 216, + 198 + ], + "spans": [ + { + "bbox": [ + 145, + 120, + 216, + 198 + ], + "type": "image", + "image_path": "102c7da16a0a2a33a483935078135635c3dfde019631cf67c6a2d2f0a9d21ada.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 176, + 200, + 184, + 206 + ], + "lines": [ + { + "bbox": [ + 176, + 200, + 184, + 206 + ], + "spans": [ + { + "bbox": [ + 176, + 200, + 184, + 206 + ], + "type": "inline_equation", + "content": "x_{i}" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_caption" + } + ], + "index": 2 + }, + { + "type": "image", + "bbox": [ + 145, + 209, + 215, + 285 + ], + "blocks": [ + { + "bbox": [ + 145, + 209, + 215, + 285 + ], + "lines": [ + { + "bbox": [ + 145, + 209, + 215, + 285 + ], + "spans": [ + { + "bbox": [ + 145, + 209, + 215, + 285 + ], + "type": "image", + "image_path": "be880941c16cdc39543dd9a8e4879c36736fe48e2f0fbda52fa472477312ae56.jpg" + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 156, + 286, + 206, + 308 + ], + "lines": [ + { + "bbox": [ + 156, + 286, + 206, + 308 + ], + "spans": [ + { + "bbox": [ + 156, + 286, + 206, + 308 + ], + "type": "inline_equation", + "content": "x_{j}" + }, + { + "bbox": [ + 156, + 286, + 206, + 308 + ], + "type": "text", + "content": " Raw UDC images (visualized for figure)" + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "image_caption" + } + ], + "index": 4 + }, + { + "type": "image", + "bbox": [ + 225, + 119, + 474, + 294 + ], + "blocks": [ + { + "bbox": [ + 225, + 119, + 474, + 294 + ], + "lines": [ + { + "bbox": [ + 225, + 119, + 474, + 294 + ], + "spans": [ + { + "bbox": [ + 225, + 119, + 474, + 294 + ], + "type": "image", + "image_path": "04272cd6e3cf06c4a0814180c5e155924f18263db7ffbc4c7aa71ef051d4ee0f.jpg" + } + ] + } + ], + "index": 6, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 225, + 295, + 266, + 304 + ], + "lines": [ + { + "bbox": [ + 225, + 295, + 266, + 304 + ], + "spans": [ + { + "bbox": [ + 225, + 295, + 266, + 304 + ], + "type": "text", + "content": "Training patches" + } + ] + } + ], + "index": 7, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 313, + 293, + 378, + 305 + ], + "lines": [ + { + "bbox": [ + 313, + 293, + 378, + 305 + ], + "spans": [ + { + "bbox": [ + 313, + 293, + 378, + 305 + ], + "type": "text", + "content": "Panel-specific degradation representation" + } + ] + } + ], + "index": 8, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 397, + 292, + 460, + 298 + ], + "lines": [ + { + "bbox": [ + 397, + 292, + 460, + 298 + ], + "spans": [ + { + "bbox": [ + 397, + 292, + 460, + 298 + ], + "type": "text", + "content": "Projected representations" + } + ] + } + ], + "index": 9, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 131, + 322, + 482, + 367 + ], + "lines": [ + { + "bbox": [ + 131, + 322, + 482, + 367 + ], + "spans": [ + { + "bbox": [ + 131, + 322, + 482, + 367 + ], + "type": "text", + "content": "Fig. 3: Our proposed unsupervised panel-specific degradation representation learning scheme. By using triplet loss, representations of the same panel-specific degradations with different image content are pulled closer together, while representations of different panel-specific degradations with the same image content are pushed away." + } + ] + } + ], + "index": 10, + "angle": 0, + "type": "image_caption" + } + ], + "index": 6 + }, + { + "bbox": [ + 130, + 388, + 482, + 460 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 388, + 482, + 460 + ], + "spans": [ + { + "bbox": [ + 130, + 388, + 482, + 460 + ], + "type": "text", + "content": "raw images " + }, + { + "bbox": [ + 130, + 388, + 482, + 460 + ], + "type": "inline_equation", + "content": "x_{i}" + }, + { + "bbox": [ + 130, + 388, + 482, + 460 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 130, + 388, + 482, + 460 + ], + "type": "inline_equation", + "content": "x_{j}" + }, + { + "bbox": [ + 130, + 388, + 482, + 460 + ], + "type": "text", + "content": " from our dataset that contains the same content but have been taken with different display panels. Then, we randomly crop two pairs of patches " + }, + { + "bbox": [ + 130, + 388, + 482, + 460 + ], + "type": "inline_equation", + "content": "(x_{i}^{1}, x_{i}^{2})" + }, + { + "bbox": [ + 130, + 388, + 482, + 460 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 130, + 388, + 482, + 460 + ], + "type": "inline_equation", + "content": "(x_{j}^{1}, x_{j}^{2})" + }, + { + "bbox": [ + 130, + 388, + 482, + 460 + ], + "type": "text", + "content": " from each image, where " + }, + { + "bbox": [ + 130, + 388, + 482, + 460 + ], + "type": "inline_equation", + "content": "x_{i}^{1}" + }, + { + "bbox": [ + 130, + 388, + 482, + 460 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 130, + 388, + 482, + 460 + ], + "type": "inline_equation", + "content": "x_{j}^{1}" + }, + { + "bbox": [ + 130, + 388, + 482, + 460 + ], + "type": "text", + "content": ", and " + }, + { + "bbox": [ + 130, + 388, + 482, + 460 + ], + "type": "inline_equation", + "content": "x_{i}^{2}" + }, + { + "bbox": [ + 130, + 388, + 482, + 460 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 130, + 388, + 482, + 460 + ], + "type": "inline_equation", + "content": "x_{j}^{2}" + }, + { + "bbox": [ + 130, + 388, + 482, + 460 + ], + "type": "text", + "content": ", respectively, are cropped from the same location of " + }, + { + "bbox": [ + 130, + 388, + 482, + 460 + ], + "type": "inline_equation", + "content": "x_{i}" + }, + { + "bbox": [ + 130, + 388, + 482, + 460 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 130, + 388, + 482, + 460 + ], + "type": "inline_equation", + "content": "x_{j}" + }, + { + "bbox": [ + 130, + 388, + 482, + 460 + ], + "type": "text", + "content": ". We choose patch " + }, + { + "bbox": [ + 130, + 388, + 482, + 460 + ], + "type": "inline_equation", + "content": "x_{i}^{1}" + }, + { + "bbox": [ + 130, + 388, + 482, + 460 + ], + "type": "text", + "content": " as the anchor, and the positive and negative samples are decided to be " + }, + { + "bbox": [ + 130, + 388, + 482, + 460 + ], + "type": "inline_equation", + "content": "x_{i}^{2}" + }, + { + "bbox": [ + 130, + 388, + 482, + 460 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 130, + 388, + 482, + 460 + ], + "type": "inline_equation", + "content": "x_{j}^{1}" + }, + { + "bbox": [ + 130, + 388, + 482, + 460 + ], + "type": "text", + "content": ". The same process is repeated with " + }, + { + "bbox": [ + 130, + 388, + 482, + 460 + ], + "type": "inline_equation", + "content": "x_{j}^{1}" + }, + { + "bbox": [ + 130, + 388, + 482, + 460 + ], + "type": "text", + "content": " as the anchor." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 130, + 460, + 482, + 555 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 460, + 482, + 555 + ], + "spans": [ + { + "bbox": [ + 130, + 460, + 482, + 555 + ], + "type": "text", + "content": "With this triplet selection, the encoder is guided to learn the intrinsic properties of the degradations caused by the display panel by pulling representation of images that show similar degradation while disregarding the features that are unrelated to degradations by pushing the representation of the same contents. The encoder " + }, + { + "bbox": [ + 130, + 460, + 482, + 555 + ], + "type": "inline_equation", + "content": "f" + }, + { + "bbox": [ + 130, + 460, + 482, + 555 + ], + "type": "text", + "content": " is trained with triplet loss using the selected triplets to obtain the panel-specific degradation representation " + }, + { + "bbox": [ + 130, + 460, + 482, + 555 + ], + "type": "inline_equation", + "content": "y \\in \\mathbb{R}^{256}" + }, + { + "bbox": [ + 130, + 460, + 482, + 555 + ], + "type": "text", + "content": ". The representations are further projected to a projection " + }, + { + "bbox": [ + 130, + 460, + 482, + 555 + ], + "type": "inline_equation", + "content": "z" + }, + { + "bbox": [ + 130, + 460, + 482, + 555 + ], + "type": "text", + "content": " with two fully connected layers, following [5, 6]. The triplet loss for training the encoder is as follows:" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 141, + 563, + 481, + 597 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 141, + 563, + 481, + 597 + ], + "spans": [ + { + "bbox": [ + 141, + 563, + 481, + 597 + ], + "type": "interline_equation", + "content": "\\begin{array}{l} L = \\max \\left(\\left\\| z _ {i} ^ {1} - z _ {i} ^ {2} \\right\\| _ {2} ^ {2} - \\left\\| z _ {i} ^ {1} - z _ {j} ^ {1} \\right\\| _ {2} ^ {2} + \\alpha , 0\\right) \\\\ + m a x (| | z _ {j} ^ {1} - z _ {j} ^ {2} | | _ {2} ^ {2} - | | z _ {j} ^ {1} - z _ {i} ^ {1} | | _ {2} ^ {2} + \\alpha , 0), \\quad (1) \\\\ \\end{array}", + "image_path": "99dfa8ff95befb1d24da4436c4d57d9874a52413591bc7d638a49e473f374b9f.jpg" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 131, + 601, + 482, + 626 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 131, + 601, + 482, + 626 + ], + "spans": [ + { + "bbox": [ + 131, + 601, + 482, + 626 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 131, + 601, + 482, + 626 + ], + "type": "inline_equation", + "content": "\\alpha" + }, + { + "bbox": [ + 131, + 601, + 482, + 626 + ], + "type": "text", + "content": " is the margin of the triplet loss. The margin is set as " + }, + { + "bbox": [ + 131, + 601, + 482, + 626 + ], + "type": "inline_equation", + "content": "\\alpha = 1" + }, + { + "bbox": [ + 131, + 601, + 482, + 626 + ], + "type": "text", + "content": ". Following [32], the encoder is a simple network with six convolutional layers." + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 131, + 641, + 482, + 665 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 131, + 641, + 482, + 665 + ], + "spans": [ + { + "bbox": [ + 131, + 641, + 482, + 665 + ], + "type": "text", + "content": "Representation embedded restoration network For our restoration network, we design it to be as simple and efficient as possible to emphasize the ef-" + } + ] + } + ], + "index": 15 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "spans": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "type": "text", + "content": "8" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 166, + 91, + 217, + 100 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 166, + 91, + 217, + 100 + ], + "spans": [ + { + "bbox": [ + 166, + 91, + 217, + 100 + ], + "type": "text", + "content": "Y. Oh et al." + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 7 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 147, + 116, + 465, + 242 + ], + "blocks": [ + { + "bbox": [ + 147, + 116, + 465, + 242 + ], + "lines": [ + { + "bbox": [ + 147, + 116, + 465, + 242 + ], + "spans": [ + { + "bbox": [ + 147, + 116, + 465, + 242 + ], + "type": "image", + "image_path": "baca668d5f8b5a85e243d7382e7c330b0b3f5b757a74ed434aec8ba3b75c838f.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 210, + 243, + 373, + 254 + ], + "lines": [ + { + "bbox": [ + 210, + 243, + 373, + 254 + ], + "spans": [ + { + "bbox": [ + 210, + 243, + 373, + 254 + ], + "type": "text", + "content": "(a) Our proposed framework, DREUDC" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 407, + 243, + 465, + 254 + ], + "lines": [ + { + "bbox": [ + 407, + 243, + 465, + 254 + ], + "spans": [ + { + "bbox": [ + 407, + 243, + 465, + 254 + ], + "type": "text", + "content": "(b) DREblock" + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 131, + 265, + 482, + 299 + ], + "lines": [ + { + "bbox": [ + 131, + 265, + 482, + 299 + ], + "spans": [ + { + "bbox": [ + 131, + 265, + 482, + 299 + ], + "type": "text", + "content": "Fig. 4: Our proposed DREUDC. The framework consists of an encoder that produces an implicit representation of the panel-specific degradations of a UDC image and a restoration network that exploits the representations via the proposed DREblocks." + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "image_caption" + } + ], + "index": 2 + }, + { + "bbox": [ + 130, + 323, + 482, + 430 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 323, + 482, + 430 + ], + "spans": [ + { + "bbox": [ + 130, + 323, + 482, + 430 + ], + "type": "text", + "content": "fectiveness of our contribution of how we leverage the difference in data collected with various panels to enhance the restoration performance rather than focusing on the impact of complex restoration network architectures. Therefore, we adopt a U-shaped [27] Transformer with Frequency domain-based Self-Attention Solver (FSAS) [16] as our restoration architecture. In previous works on UDC image restoration [9, 10, 17, 31], researchers have highlighted the importance of a wide receptive field. As FSAS utilizes the convolutional theorem of Fourier transform to model global contexts in the frequency domain efficiently [16], we choose it as a component of the building block for our restoration network." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 130, + 431, + 482, + 491 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 431, + 482, + 491 + ], + "spans": [ + { + "bbox": [ + 130, + 431, + 482, + 491 + ], + "type": "text", + "content": "Our restoration network is based on stacking repeated Transformer blocks, which we name DREblock (Degradation Representation Embedded Transformer block, Fig. 4(b)). The proposed DREblock efficiently integrates the degradation representation " + }, + { + "bbox": [ + 130, + 431, + 482, + 491 + ], + "type": "inline_equation", + "content": "y" + }, + { + "bbox": [ + 130, + 431, + 482, + 491 + ], + "type": "text", + "content": " of the input UDC image as a condition of our restoration network by calibrating the features in the block in a channel-wise manner, i.e.," + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 273, + 502, + 481, + 516 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 273, + 502, + 481, + 516 + ], + "spans": [ + { + "bbox": [ + 273, + 502, + 481, + 516 + ], + "type": "interline_equation", + "content": "F _ {o} = F _ {a t t} \\otimes \\beta , \\tag {2}", + "image_path": "354d8397544dd602d5fb8a2c44269f544d080dbea583d85c969b761ce1fa3f8e.jpg" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 130, + 521, + 482, + 582 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 521, + 482, + 582 + ], + "spans": [ + { + "bbox": [ + 130, + 521, + 482, + 582 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 130, + 521, + 482, + 582 + ], + "type": "inline_equation", + "content": "F_{att}" + }, + { + "bbox": [ + 130, + 521, + 482, + 582 + ], + "type": "text", + "content": " is the feature maps of the estimated attention of FSAS, " + }, + { + "bbox": [ + 130, + 521, + 482, + 582 + ], + "type": "inline_equation", + "content": "F_{o}" + }, + { + "bbox": [ + 130, + 521, + 482, + 582 + ], + "type": "text", + "content": " is the feature after the conditioning, " + }, + { + "bbox": [ + 130, + 521, + 482, + 582 + ], + "type": "inline_equation", + "content": "\\otimes" + }, + { + "bbox": [ + 130, + 521, + 482, + 582 + ], + "type": "text", + "content": " denotes channel-wise multiplication, and " + }, + { + "bbox": [ + 130, + 521, + 482, + 582 + ], + "type": "inline_equation", + "content": "\\beta" + }, + { + "bbox": [ + 130, + 521, + 482, + 582 + ], + "type": "text", + "content": " is the channel conditioning coefficients that are produced by reshaping the panel-specific degradation representation " + }, + { + "bbox": [ + 130, + 521, + 482, + 582 + ], + "type": "inline_equation", + "content": "y" + }, + { + "bbox": [ + 130, + 521, + 482, + 582 + ], + "type": "text", + "content": " using a single layer of " + }, + { + "bbox": [ + 130, + 521, + 482, + 582 + ], + "type": "inline_equation", + "content": "1 \\times 1" + }, + { + "bbox": [ + 130, + 521, + 482, + 582 + ], + "type": "text", + "content": " convolution. The features " + }, + { + "bbox": [ + 130, + 521, + 482, + 582 + ], + "type": "inline_equation", + "content": "F_{o}" + }, + { + "bbox": [ + 130, + 521, + 482, + 582 + ], + "type": "text", + "content": " are then processed by the point-wise feed-forward network." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 130, + 582, + 482, + 666 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 582, + 482, + 666 + ], + "spans": [ + { + "bbox": [ + 130, + 582, + 482, + 666 + ], + "type": "text", + "content": "The entire framework DREUDC is illustrated in Fig. 4. In training, the encoder is first trained with Eq. (1). Then, the encoder is frozen while the restoration network is trained with " + }, + { + "bbox": [ + 130, + 582, + 482, + 666 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_1" + }, + { + "bbox": [ + 130, + 582, + 482, + 666 + ], + "type": "text", + "content": " loss. During inference, the encoder takes the input raw UDC image and outputs a panel-specific degradation representation vector that contains prior information on UDC degradation. The representation is seamlessly embedded into the restoration network through DREblocks, finally producing a clean restored RGB image." + } + ] + } + ], + "index": 10 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 288, + 91, + 447, + 102 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 288, + 91, + 447, + 102 + ], + "spans": [ + { + "bbox": [ + 288, + 91, + 447, + 102 + ], + "type": "text", + "content": "Panel-Specific UDC Image Restoration" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 474, + 91, + 481, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 474, + 91, + 481, + 100 + ], + "spans": [ + { + "bbox": [ + 474, + 91, + 481, + 100 + ], + "type": "text", + "content": "9" + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 8 + }, + { + "para_blocks": [ + { + "bbox": [ + 132, + 114, + 304, + 129 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 114, + 304, + 129 + ], + "spans": [ + { + "bbox": [ + 132, + 114, + 304, + 129 + ], + "type": "text", + "content": "4 Experiment and Analysis" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 132, + 140, + 204, + 150 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 140, + 204, + 150 + ], + "spans": [ + { + "bbox": [ + 132, + 140, + 204, + 150 + ], + "type": "text", + "content": "4.1 Datasets" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 132, + 159, + 482, + 316 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 159, + 482, + 316 + ], + "spans": [ + { + "bbox": [ + 132, + 159, + 482, + 316 + ], + "type": "text", + "content": "We evaluate the restoration performance using our captured data with display panels of ZTE Axon 30 5G and Samsung Galaxy Z-Fold 3. The raw images are normalized to have values within the range of [0,1]. To train and evaluate RGB-based methods, we use a simplified ISP algorithm written with MATLAB to visualize the normalized raw images to RGB images. The algorithm contains demosaicing, channel-wise gain control, and gamma correction in sequence. Note that the visualized images are not sRGB images, as actual camera ISPs in practice involve additional steps to render the image in sRGB domain. Please check the Supplementary for more discussion on the usage of simplified ISP. For both display panels, we split the dataset into 300 images for training and 100 images for testing. The training set is cropped into patches of " + }, + { + "bbox": [ + 132, + 159, + 482, + 316 + ], + "type": "inline_equation", + "content": "512 \\times 512" + }, + { + "bbox": [ + 132, + 159, + 482, + 316 + ], + "type": "text", + "content": ". We also use previous UDC datasets [38], which consists of 300 images of both T-OLED and P-OLED scenes, to verify that our method is applicable to other real datasets." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 132, + 333, + 348, + 345 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 333, + 348, + 345 + ], + "spans": [ + { + "bbox": [ + 132, + 333, + 348, + 345 + ], + "type": "text", + "content": "4.2 Training and Implementation Details" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 132, + 352, + 482, + 485 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 352, + 482, + 485 + ], + "spans": [ + { + "bbox": [ + 132, + 352, + 482, + 485 + ], + "type": "text", + "content": "We use NVIDIA RTX 3090 GPUs in our experiments. To train the encoder, we use Adam optimizer " + }, + { + "bbox": [ + 132, + 352, + 482, + 485 + ], + "type": "inline_equation", + "content": "(\\beta_{1} = 0.9, \\beta_{2} = 0.999)" + }, + { + "bbox": [ + 132, + 352, + 482, + 485 + ], + "type": "text", + "content": " for " + }, + { + "bbox": [ + 132, + 352, + 482, + 485 + ], + "type": "inline_equation", + "content": "1 \\times 10^{5}" + }, + { + "bbox": [ + 132, + 352, + 482, + 485 + ], + "type": "text", + "content": " iterations with a batch size of 64, and the loss is given as Eq. (1). The restoration networks are trained with " + }, + { + "bbox": [ + 132, + 352, + 482, + 485 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_1" + }, + { + "bbox": [ + 132, + 352, + 482, + 485 + ], + "type": "text", + "content": " loss to minimize the distance between the restored image of the network and the ground truth. The initial number of channels " + }, + { + "bbox": [ + 132, + 352, + 482, + 485 + ], + "type": "inline_equation", + "content": "C" + }, + { + "bbox": [ + 132, + 352, + 482, + 485 + ], + "type": "text", + "content": " of our restoration network is 32, and the number of blocks in each layer " + }, + { + "bbox": [ + 132, + 352, + 482, + 485 + ], + "type": "inline_equation", + "content": "L_1, L_2" + }, + { + "bbox": [ + 132, + 352, + 482, + 485 + ], + "type": "text", + "content": ", and " + }, + { + "bbox": [ + 132, + 352, + 482, + 485 + ], + "type": "inline_equation", + "content": "L_3" + }, + { + "bbox": [ + 132, + 352, + 482, + 485 + ], + "type": "text", + "content": " are [5,5,6]. We use Adam optimizer " + }, + { + "bbox": [ + 132, + 352, + 482, + 485 + ], + "type": "inline_equation", + "content": "(\\beta_{1} = 0.9, \\beta_{2} = 0.999)" + }, + { + "bbox": [ + 132, + 352, + 482, + 485 + ], + "type": "text", + "content": " for " + }, + { + "bbox": [ + 132, + 352, + 482, + 485 + ], + "type": "inline_equation", + "content": "2 \\times 10^{5}" + }, + { + "bbox": [ + 132, + 352, + 482, + 485 + ], + "type": "text", + "content": " iterations with a batch size of 2. For both networks, the initial learning rate is set as " + }, + { + "bbox": [ + 132, + 352, + 482, + 485 + ], + "type": "inline_equation", + "content": "2 \\times 10^{-4}" + }, + { + "bbox": [ + 132, + 352, + 482, + 485 + ], + "type": "text", + "content": " and gradually decayed to " + }, + { + "bbox": [ + 132, + 352, + 482, + 485 + ], + "type": "inline_equation", + "content": "1 \\times 10^{-6}" + }, + { + "bbox": [ + 132, + 352, + 482, + 485 + ], + "type": "text", + "content": " with cosine annealing [20]. This setting is applied to all experiments on our dataset, and every previous method reported in this paper is also reproduced by training on our dataset with the same setting for fairness." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 132, + 502, + 310, + 514 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 502, + 310, + 514 + ], + "spans": [ + { + "bbox": [ + 132, + 502, + 310, + 514 + ], + "type": "text", + "content": "4.3 Raw UDC Image Restoration" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 132, + 521, + 482, + 666 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 521, + 482, + 666 + ], + "spans": [ + { + "bbox": [ + 132, + 521, + 482, + 666 + ], + "type": "text", + "content": "First, we experiment to validate that using raw images can improve the UDC image restoration process. For this purpose, we employ a baseline neural network called DE-UNET [38] and use our captured data. We compare the performance of the network under three different settings: RAW-to-RAW, RAW-to-RGB, and RGB-to-RGB. We evaluate the performance of the network using the PSNR and SSIM metrics. To assess the performance of the RAW-to-RAW setting, we convert the output restored raw images to RGB using the same ISP process mentioned in Sec. 4.1 and measure them in the RGB domain. The results are presented in Tab. 2, showing that using raw sensor images as input (RAW-to-RAW, RAW-to-RGB) yields superior results compared to only using RGB images. Among the two settings that use raw images as input, the latter shows better performance overall, so we choose the method of converting and restoring" + } + ] + } + ], + "index": 8 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "spans": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "type": "text", + "content": "10" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 166, + 91, + 217, + 101 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 166, + 91, + 217, + 101 + ], + "spans": [ + { + "bbox": [ + 166, + 91, + 217, + 101 + ], + "type": "text", + "content": "Y. Oh et al." + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 9 + }, + { + "para_blocks": [ + { + "type": "table", + "bbox": [ + 133, + 146, + 479, + 204 + ], + "blocks": [ + { + "bbox": [ + 132, + 114, + 480, + 137 + ], + "lines": [ + { + "bbox": [ + 132, + 114, + 480, + 137 + ], + "spans": [ + { + "bbox": [ + 132, + 114, + 480, + 137 + ], + "type": "text", + "content": "Table 2: Comparison of different input-to-output settings of DE-UNET [38] on our captured datasets. The best results are bold-faced." + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 133, + 146, + 479, + 204 + ], + "lines": [ + { + "bbox": [ + 133, + 146, + 479, + 204 + ], + "spans": [ + { + "bbox": [ + 133, + 146, + 479, + 204 + ], + "type": "table", + "html": "
MethodAxon 30 (raw)Z-Fold 3 (raw)Axon 30 (rgb)Z-Fold 3 (rgb)
PSNRSSIMPSNRSSIMPSNRSSIMPSNRSSIM
DE-UNET(RGB-to-RGB)----23.330.910926.900.9390
DE-UNET(RAW-to-RAW)27.360.933529.660.963724.210.917027.180.9388
DE-UNET(RAW-to-RGB)----24.200.919627.290.9419
", + "image_path": "ae2963b391961012ee355c8771a89b6ae6f5f9c660dbdcf2aaef34b175a7db4e.jpg" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "table_body" + } + ], + "index": 3 + }, + { + "type": "table", + "bbox": [ + 133, + 258, + 479, + 426 + ], + "blocks": [ + { + "bbox": [ + 132, + 214, + 480, + 246 + ], + "lines": [ + { + "bbox": [ + 132, + 214, + 480, + 246 + ], + "spans": [ + { + "bbox": [ + 132, + 214, + 480, + 246 + ], + "type": "text", + "content": "Table 3: Results on our captured datasets for ZTE Axon 30 5G and Samsung Galaxy Z-Fold 3 smartphone display panels on various methods of UDC image restoration. The best results are bold-faced." + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 133, + 258, + 479, + 426 + ], + "lines": [ + { + "bbox": [ + 133, + 258, + 479, + 426 + ], + "spans": [ + { + "bbox": [ + 133, + 258, + 479, + 426 + ], + "type": "table", + "html": "
Input-to-outputMethodParamsMACsAxon 30 (rgb)Z-Fold 3 (rgb)
PSNRSSIMPSNRSSIM
RGB-to-RGBDE-UNET [38]8.93M273.84G23.330.910926.900.9390
DAGF [31]1.09M73.84G24.490.915028.420.9404
DISCNET [10]3.80M596.95G23.110.906326.670.9361
BNUDC [15]4.57M993.38G25.850.928328.540.9458
DWFormer [36]1.45M216.77G21.520.892326.360.9338
SRUDC [30]6.70M371.32G22.670.899827.340.9336
PPM-UNET [9]4.23M409.29G18.210.850223.850.9188
AWNet [7]47.00M1.53T20.270.883925.600.9316
RAW-to-RGBDE-UNET [38]8.94M69.29G24.200.919627.290.9419
PyNet [14]47.56M1.79T22.580.906927.350.9377
AWNet [7]49.07M392.25G20.980.888527.250.9369
DREUDC6.44M200.07G26.660.936829.040.9511
", + "image_path": "a3f698db37d5ced7ed48e35fed0ce9601d8c7cb0cb14a15fc37545d3383104f4.jpg" + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "table_body" + } + ], + "index": 5 + }, + { + "bbox": [ + 132, + 449, + 480, + 485 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 449, + 480, + 485 + ], + "spans": [ + { + "bbox": [ + 132, + 449, + 480, + 485 + ], + "type": "text", + "content": "raw UDC images to a clean RGB image (RAW-to-RGB). This is also the setting of most neural ISPs currently being used in smartphones [26] since a successful network is able to model the image processing pipeline as well." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 132, + 502, + 296, + 513 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 502, + 296, + 513 + ], + "spans": [ + { + "bbox": [ + 132, + 502, + 296, + 513 + ], + "type": "text", + "content": "4.4 Evaluation on Our Dataset" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 132, + 521, + 481, + 593 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 521, + 481, + 593 + ], + "spans": [ + { + "bbox": [ + 132, + 521, + 481, + 593 + ], + "type": "text", + "content": "To provide experimental results on our dataset and demonstrate the performance of DREUDC on UDC image restoration, we train numerous previous methods using our dataset for comparison. The quality of the restoration performance is evaluated with PSNR and SSIM as performance metrics. Additionally, we assess the efficiency of the methods by comparing the number of parameters and MACs. To calculate the MACs, we use a dummy image with a resolution of " + }, + { + "bbox": [ + 132, + 521, + 481, + 593 + ], + "type": "inline_equation", + "content": "1024 \\times 1024" + }, + { + "bbox": [ + 132, + 521, + 481, + 593 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 132, + 594, + 481, + 665 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 594, + 481, + 665 + ], + "spans": [ + { + "bbox": [ + 132, + 594, + 481, + 665 + ], + "type": "text", + "content": "Tab. 3 shows experimental results for raw and RGB image restoration using our dataset. Because there are no neural networks on UDC image restoration that take raw sensor images as input except [38] as previous datasets were provided only in the RGB domain, we choose [7] and [14] for further comparison, which are RAW-to-RGB image conversion models. We observe that DREUDC achieves state-of-the-art performance on both Axon 30 and Z-fold 3 datasets," + } + ] + } + ], + "index": 9 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 288, + 91, + 447, + 102 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 288, + 91, + 447, + 102 + ], + "spans": [ + { + "bbox": [ + 288, + 91, + 447, + 102 + ], + "type": "text", + "content": "Panel-Specific UDC Image Restoration" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 470, + 91, + 479, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 470, + 91, + 479, + 100 + ], + "spans": [ + { + "bbox": [ + 470, + 91, + 479, + 100 + ], + "type": "text", + "content": "11" + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 10 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 141, + 117, + 185, + 161 + ], + "blocks": [ + { + "bbox": [ + 141, + 117, + 185, + 161 + ], + "lines": [ + { + "bbox": [ + 141, + 117, + 185, + 161 + ], + "spans": [ + { + "bbox": [ + 141, + 117, + 185, + 161 + ], + "type": "image", + "image_path": "cf58ad7779c04e55c3e89ea859c1eaaca0bbc928c454a5744b52b5ec35cd483f.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 146, + 161, + 182, + 169 + ], + "lines": [ + { + "bbox": [ + 146, + 161, + 182, + 169 + ], + "spans": [ + { + "bbox": [ + 146, + 161, + 182, + 169 + ], + "type": "text", + "content": "Axon 30" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_caption" + } + ], + "index": 2 + }, + { + "type": "image", + "bbox": [ + 188, + 117, + 234, + 161 + ], + "blocks": [ + { + "bbox": [ + 188, + 117, + 234, + 161 + ], + "lines": [ + { + "bbox": [ + 188, + 117, + 234, + 161 + ], + "spans": [ + { + "bbox": [ + 188, + 117, + 234, + 161 + ], + "type": "image", + "image_path": "5a1a8cb18d978c4675bdfe9e419cdcd277ec46436c5f66f118596ce7557bcb4c.jpg" + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 189, + 161, + 233, + 169 + ], + "lines": [ + { + "bbox": [ + 189, + 161, + 233, + 169 + ], + "spans": [ + { + "bbox": [ + 189, + 161, + 233, + 169 + ], + "type": "text", + "content": "DE-UNET" + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "image_caption" + } + ], + "index": 4 + }, + { + "type": "image", + "bbox": [ + 237, + 117, + 282, + 161 + ], + "blocks": [ + { + "bbox": [ + 237, + 117, + 282, + 161 + ], + "lines": [ + { + "bbox": [ + 237, + 117, + 282, + 161 + ], + "spans": [ + { + "bbox": [ + 237, + 117, + 282, + 161 + ], + "type": "image", + "image_path": "443c069547bae31500f4df171a91cef2cb0d9ece0a885cf0b621b624b1fe3eaa.jpg" + } + ] + } + ], + "index": 6, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 247, + 161, + 272, + 169 + ], + "lines": [ + { + "bbox": [ + 247, + 161, + 272, + 169 + ], + "spans": [ + { + "bbox": [ + 247, + 161, + 272, + 169 + ], + "type": "text", + "content": "DAGF" + } + ] + } + ], + "index": 7, + "angle": 0, + "type": "image_caption" + } + ], + "index": 6 + }, + { + "type": "image", + "bbox": [ + 286, + 117, + 329, + 161 + ], + "blocks": [ + { + "bbox": [ + 286, + 117, + 329, + 161 + ], + "lines": [ + { + "bbox": [ + 286, + 117, + 329, + 161 + ], + "spans": [ + { + "bbox": [ + 286, + 117, + 329, + 161 + ], + "type": "image", + "image_path": "1caa6d41d56031b6e8ea6abf11ee2cfdfc93a7366a8b91dd6a93cddc0c73735e.jpg" + } + ] + } + ], + "index": 8, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 287, + 161, + 329, + 169 + ], + "lines": [ + { + "bbox": [ + 287, + 161, + 329, + 169 + ], + "spans": [ + { + "bbox": [ + 287, + 161, + 329, + 169 + ], + "type": "text", + "content": "DISCNET" + } + ] + } + ], + "index": 9, + "angle": 0, + "type": "image_caption" + } + ], + "index": 8 + }, + { + "type": "image", + "bbox": [ + 333, + 117, + 378, + 161 + ], + "blocks": [ + { + "bbox": [ + 333, + 117, + 378, + 161 + ], + "lines": [ + { + "bbox": [ + 333, + 117, + 378, + 161 + ], + "spans": [ + { + "bbox": [ + 333, + 117, + 378, + 161 + ], + "type": "image", + "image_path": "999eaef93d5b29bc7c11ce9721f9aa5977a0822732c42b96c9321c8dd2e228a2.jpg" + } + ] + } + ], + "index": 10, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 337, + 161, + 373, + 169 + ], + "lines": [ + { + "bbox": [ + 337, + 161, + 373, + 169 + ], + "spans": [ + { + "bbox": [ + 337, + 161, + 373, + 169 + ], + "type": "text", + "content": "BNUDC" + } + ] + } + ], + "index": 11, + "angle": 0, + "type": "image_caption" + } + ], + "index": 10 + }, + { + "type": "image", + "bbox": [ + 382, + 117, + 426, + 161 + ], + "blocks": [ + { + "bbox": [ + 382, + 117, + 426, + 161 + ], + "lines": [ + { + "bbox": [ + 382, + 117, + 426, + 161 + ], + "spans": [ + { + "bbox": [ + 382, + 117, + 426, + 161 + ], + "type": "image", + "image_path": "de2b7ff4b043a6512ebb7b040e478633b48541ee459f2fc7c90449b255b5603e.jpg" + } + ] + } + ], + "index": 12, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 382, + 161, + 426, + 169 + ], + "lines": [ + { + "bbox": [ + 382, + 161, + 426, + 169 + ], + "spans": [ + { + "bbox": [ + 382, + 161, + 426, + 169 + ], + "type": "text", + "content": "DWFormer" + } + ] + } + ], + "index": 13, + "angle": 0, + "type": "image_caption" + } + ], + "index": 12 + }, + { + "type": "image", + "bbox": [ + 430, + 117, + 474, + 161 + ], + "blocks": [ + { + "bbox": [ + 430, + 117, + 474, + 161 + ], + "lines": [ + { + "bbox": [ + 430, + 117, + 474, + 161 + ], + "spans": [ + { + "bbox": [ + 430, + 117, + 474, + 161 + ], + "type": "image", + "image_path": "6096c77c949deacc39372e624b262af9bf37fc56928e6b2e9e860f9bf079b002.jpg" + } + ] + } + ], + "index": 14, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 435, + 161, + 469, + 168 + ], + "lines": [ + { + "bbox": [ + 435, + 161, + 469, + 168 + ], + "spans": [ + { + "bbox": [ + 435, + 161, + 469, + 168 + ], + "type": "text", + "content": "SRUDC" + } + ] + } + ], + "index": 15, + "angle": 0, + "type": "image_caption" + } + ], + "index": 14 + }, + { + "type": "image", + "bbox": [ + 141, + 169, + 185, + 212 + ], + "blocks": [ + { + "bbox": [ + 141, + 169, + 185, + 212 + ], + "lines": [ + { + "bbox": [ + 141, + 169, + 185, + 212 + ], + "spans": [ + { + "bbox": [ + 141, + 169, + 185, + 212 + ], + "type": "image", + "image_path": "51089fc1f79f4c8142433ecfc7d382fe91b547626d99c32fd262a77dc8e5b7f6.jpg" + } + ] + } + ], + "index": 16, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 140, + 212, + 185, + 220 + ], + "lines": [ + { + "bbox": [ + 140, + 212, + 185, + 220 + ], + "spans": [ + { + "bbox": [ + 140, + 212, + 185, + 220 + ], + "type": "text", + "content": "PPM-UNet" + } + ] + } + ], + "index": 17, + "angle": 0, + "type": "image_caption" + } + ], + "index": 16 + }, + { + "type": "image", + "bbox": [ + 189, + 169, + 234, + 212 + ], + "blocks": [ + { + "bbox": [ + 189, + 169, + 234, + 212 + ], + "lines": [ + { + "bbox": [ + 189, + 169, + 234, + 212 + ], + "spans": [ + { + "bbox": [ + 189, + 169, + 234, + 212 + ], + "type": "image", + "image_path": "4ff5e71daf11dd9b66559d9a4f3362ff0ada86b13a0d4b0fafe821c8b084d69d.jpg" + } + ] + } + ], + "index": 18, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 189, + 212, + 227, + 220 + ], + "lines": [ + { + "bbox": [ + 189, + 212, + 227, + 220 + ], + "spans": [ + { + "bbox": [ + 189, + 212, + 227, + 220 + ], + "type": "text", + "content": "AWNet" + } + ] + } + ], + "index": 19, + "angle": 0, + "type": "image_caption" + } + ], + "index": 18 + }, + { + "type": "image", + "bbox": [ + 237, + 169, + 282, + 212 + ], + "blocks": [ + { + "bbox": [ + 237, + 169, + 282, + 212 + ], + "lines": [ + { + "bbox": [ + 237, + 169, + 282, + 212 + ], + "spans": [ + { + "bbox": [ + 237, + 169, + 282, + 212 + ], + "type": "image", + "image_path": "36927a5073b0582a182a878734401017125dcf564e9b5c2a37e9203de448c778.jpg" + } + ] + } + ], + "index": 20, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 236, + 212, + 282, + 220 + ], + "lines": [ + { + "bbox": [ + 236, + 212, + 282, + 220 + ], + "spans": [ + { + "bbox": [ + 236, + 212, + 282, + 220 + ], + "type": "text", + "content": "DE-UNET*" + } + ] + } + ], + "index": 21, + "angle": 0, + "type": "image_caption" + } + ], + "index": 20 + }, + { + "type": "image", + "bbox": [ + 286, + 169, + 329, + 212 + ], + "blocks": [ + { + "bbox": [ + 286, + 169, + 329, + 212 + ], + "lines": [ + { + "bbox": [ + 286, + 169, + 329, + 212 + ], + "spans": [ + { + "bbox": [ + 286, + 169, + 329, + 212 + ], + "type": "image", + "image_path": "3bfee9f0cd52fd79fe809ff4a57f8eaf3bc47d042de60068cc678c81f4e468b7.jpg" + } + ] + } + ], + "index": 22, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 287, + 212, + 323, + 220 + ], + "lines": [ + { + "bbox": [ + 287, + 212, + 323, + 220 + ], + "spans": [ + { + "bbox": [ + 287, + 212, + 323, + 220 + ], + "type": "text", + "content": "PyNet*" + } + ] + } + ], + "index": 23, + "angle": 0, + "type": "image_caption" + } + ], + "index": 22 + }, + { + "type": "image", + "bbox": [ + 334, + 169, + 378, + 212 + ], + "blocks": [ + { + "bbox": [ + 334, + 169, + 378, + 212 + ], + "lines": [ + { + "bbox": [ + 334, + 169, + 378, + 212 + ], + "spans": [ + { + "bbox": [ + 334, + 169, + 378, + 212 + ], + "type": "image", + "image_path": "da2042856f64d2302393156e591606e5692b43e16cc78a33c0837b612ab2e325.jpg" + } + ] + } + ], + "index": 24, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 337, + 212, + 373, + 220 + ], + "lines": [ + { + "bbox": [ + 337, + 212, + 373, + 220 + ], + "spans": [ + { + "bbox": [ + 337, + 212, + 373, + 220 + ], + "type": "text", + "content": "AWNet*" + } + ] + } + ], + "index": 25, + "angle": 0, + "type": "image_caption" + } + ], + "index": 24 + }, + { + "type": "image", + "bbox": [ + 382, + 169, + 426, + 212 + ], + "blocks": [ + { + "bbox": [ + 382, + 169, + 426, + 212 + ], + "lines": [ + { + "bbox": [ + 382, + 169, + 426, + 212 + ], + "spans": [ + { + "bbox": [ + 382, + 169, + 426, + 212 + ], + "type": "image", + "image_path": "cb53a85ef3ac108165db10d937ea1362965ad65f48b7ce111eef22a1d6ecbeae.jpg" + } + ] + } + ], + "index": 26, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 382, + 212, + 426, + 220 + ], + "lines": [ + { + "bbox": [ + 382, + 212, + 426, + 220 + ], + "spans": [ + { + "bbox": [ + 382, + 212, + 426, + 220 + ], + "type": "text", + "content": "DREUDC*" + } + ] + } + ], + "index": 27, + "angle": 0, + "type": "image_caption" + } + ], + "index": 26 + }, + { + "type": "image", + "bbox": [ + 430, + 169, + 474, + 212 + ], + "blocks": [ + { + "bbox": [ + 430, + 169, + 474, + 212 + ], + "lines": [ + { + "bbox": [ + 430, + 169, + 474, + 212 + ], + "spans": [ + { + "bbox": [ + 430, + 169, + 474, + 212 + ], + "type": "image", + "image_path": "4f45c0a2f1c303941e6fb978baae763cedccde3d64323f982c5031a416288961.jpg" + } + ] + } + ], + "index": 28, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 434, + 212, + 459, + 220 + ], + "lines": [ + { + "bbox": [ + 434, + 212, + 459, + 220 + ], + "spans": [ + { + "bbox": [ + 434, + 212, + 459, + 220 + ], + "type": "text", + "content": "GT" + } + ] + } + ], + "index": 29, + "angle": 0, + "type": "image_caption" + } + ], + "index": 28 + }, + { + "type": "image", + "bbox": [ + 141, + 220, + 185, + 263 + ], + "blocks": [ + { + "bbox": [ + 141, + 220, + 185, + 263 + ], + "lines": [ + { + "bbox": [ + 141, + 220, + 185, + 263 + ], + "spans": [ + { + "bbox": [ + 141, + 220, + 185, + 263 + ], + "type": "image", + "image_path": "28de6ae19e42b77bbbc157d038977b00bd492172412b94a6d946b5f1c3bac0f8.jpg" + } + ] + } + ], + "index": 30, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 146, + 263, + 182, + 270 + ], + "lines": [ + { + "bbox": [ + 146, + 263, + 182, + 270 + ], + "spans": [ + { + "bbox": [ + 146, + 263, + 182, + 270 + ], + "type": "text", + "content": "Z-Fold 3" + } + ] + } + ], + "index": 31, + "angle": 0, + "type": "image_caption" + } + ], + "index": 30 + }, + { + "type": "image", + "bbox": [ + 189, + 220, + 234, + 263 + ], + "blocks": [ + { + "bbox": [ + 189, + 220, + 234, + 263 + ], + "lines": [ + { + "bbox": [ + 189, + 220, + 234, + 263 + ], + "spans": [ + { + "bbox": [ + 189, + 220, + 234, + 263 + ], + "type": "image", + "image_path": "64fecfff12cf6f2be39cd3bbba86cdc45930ac18b4a453adcd1d90557a229167.jpg" + } + ] + } + ], + "index": 32, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 189, + 263, + 231, + 270 + ], + "lines": [ + { + "bbox": [ + 189, + 263, + 231, + 270 + ], + "spans": [ + { + "bbox": [ + 189, + 263, + 231, + 270 + ], + "type": "text", + "content": "DE-UNET" + } + ] + } + ], + "index": 33, + "angle": 0, + "type": "image_caption" + } + ], + "index": 32 + }, + { + "type": "image", + "bbox": [ + 237, + 220, + 282, + 263 + ], + "blocks": [ + { + "bbox": [ + 237, + 220, + 282, + 263 + ], + "lines": [ + { + "bbox": [ + 237, + 220, + 282, + 263 + ], + "spans": [ + { + "bbox": [ + 237, + 220, + 282, + 263 + ], + "type": "image", + "image_path": "11f2f6831a9fc818a427e74d9dc74aac14688a42aaee80aad7c7615d7ffe8788.jpg" + } + ] + } + ], + "index": 34, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 244, + 263, + 272, + 270 + ], + "lines": [ + { + "bbox": [ + 244, + 263, + 272, + 270 + ], + "spans": [ + { + "bbox": [ + 244, + 263, + 272, + 270 + ], + "type": "text", + "content": "DAGE" + } + ] + } + ], + "index": 35, + "angle": 0, + "type": "image_caption" + } + ], + "index": 34 + }, + { + "type": "image", + "bbox": [ + 286, + 220, + 329, + 263 + ], + "blocks": [ + { + "bbox": [ + 286, + 220, + 329, + 263 + ], + "lines": [ + { + "bbox": [ + 286, + 220, + 329, + 263 + ], + "spans": [ + { + "bbox": [ + 286, + 220, + 329, + 263 + ], + "type": "image", + "image_path": "c913bb6114da146be22a155aac2b9b910b8f85fda682cf58f9da7f04660978f3.jpg" + } + ] + } + ], + "index": 36, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 287, + 263, + 328, + 270 + ], + "lines": [ + { + "bbox": [ + 287, + 263, + 328, + 270 + ], + "spans": [ + { + "bbox": [ + 287, + 263, + 328, + 270 + ], + "type": "text", + "content": "DISCNET" + } + ] + } + ], + "index": 37, + "angle": 0, + "type": "image_caption" + } + ], + "index": 36 + }, + { + "type": "image", + "bbox": [ + 334, + 220, + 378, + 263 + ], + "blocks": [ + { + "bbox": [ + 334, + 220, + 378, + 263 + ], + "lines": [ + { + "bbox": [ + 334, + 220, + 378, + 263 + ], + "spans": [ + { + "bbox": [ + 334, + 220, + 378, + 263 + ], + "type": "image", + "image_path": "e91c89e4cf8ed108ccdc67db7c4c6484fa1f7f3eb632bb30a0af0b8ae862ef5d.jpg" + } + ] + } + ], + "index": 38, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 337, + 263, + 373, + 270 + ], + "lines": [ + { + "bbox": [ + 337, + 263, + 373, + 270 + ], + "spans": [ + { + "bbox": [ + 337, + 263, + 373, + 270 + ], + "type": "text", + "content": "BNUDC" + } + ] + } + ], + "index": 39, + "angle": 0, + "type": "image_caption" + } + ], + "index": 38 + }, + { + "type": "image", + "bbox": [ + 382, + 220, + 426, + 263 + ], + "blocks": [ + { + "bbox": [ + 382, + 220, + 426, + 263 + ], + "lines": [ + { + "bbox": [ + 382, + 220, + 426, + 263 + ], + "spans": [ + { + "bbox": [ + 382, + 220, + 426, + 263 + ], + "type": "image", + "image_path": "3606fcd252b55853f99f3b3f00923e2e553763b8da48090f01d5f27500e00c16.jpg" + } + ] + } + ], + "index": 40, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 382, + 263, + 426, + 270 + ], + "lines": [ + { + "bbox": [ + 382, + 263, + 426, + 270 + ], + "spans": [ + { + "bbox": [ + 382, + 263, + 426, + 270 + ], + "type": "text", + "content": "DWFormer" + } + ] + } + ], + "index": 41, + "angle": 0, + "type": "image_caption" + } + ], + "index": 40 + }, + { + "type": "image", + "bbox": [ + 430, + 220, + 474, + 263 + ], + "blocks": [ + { + "bbox": [ + 430, + 220, + 474, + 263 + ], + "lines": [ + { + "bbox": [ + 430, + 220, + 474, + 263 + ], + "spans": [ + { + "bbox": [ + 430, + 220, + 474, + 263 + ], + "type": "image", + "image_path": "fa9928952d2f6f5c2d057093b78ab1d74df365c69a455761edf511511e6b478c.jpg" + } + ] + } + ], + "index": 42, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 434, + 263, + 469, + 270 + ], + "lines": [ + { + "bbox": [ + 434, + 263, + 469, + 270 + ], + "spans": [ + { + "bbox": [ + 434, + 263, + 469, + 270 + ], + "type": "text", + "content": "SRUDC" + } + ] + } + ], + "index": 43, + "angle": 0, + "type": "image_caption" + } + ], + "index": 42 + }, + { + "type": "image", + "bbox": [ + 141, + 270, + 185, + 314 + ], + "blocks": [ + { + "bbox": [ + 141, + 270, + 185, + 314 + ], + "lines": [ + { + "bbox": [ + 141, + 270, + 185, + 314 + ], + "spans": [ + { + "bbox": [ + 141, + 270, + 185, + 314 + ], + "type": "image", + "image_path": "ab01dbcaa3f52cceddbbb6f3978fec19d414d9599db3f1b4e1bfaf9e4b9508a8.jpg" + } + ] + } + ], + "index": 44, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 140, + 314, + 186, + 323 + ], + "lines": [ + { + "bbox": [ + 140, + 314, + 186, + 323 + ], + "spans": [ + { + "bbox": [ + 140, + 314, + 186, + 323 + ], + "type": "text", + "content": "PPM-UNet" + } + ] + } + ], + "index": 45, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 131, + 335, + 482, + 390 + ], + "lines": [ + { + "bbox": [ + 131, + 335, + 482, + 390 + ], + "spans": [ + { + "bbox": [ + 131, + 335, + 482, + 390 + ], + "type": "text", + "content": "Fig. 5: Visual comparison of restoration examples for our dataset. Zoom in for a better comparison between the methods. The top two rows are results for Axon 30 images, and the bottom two rows are for Z-Fold 3 images. Methods that take raw input and produce RGB images are annotated with the asterisk *. Our method is capable of removing UDC degradations more clearly than previous UDC image restoration networks." + } + ] + } + ], + "index": 58, + "angle": 0, + "type": "image_caption" + } + ], + "index": 44 + }, + { + "type": "image", + "bbox": [ + 189, + 270, + 234, + 314 + ], + "blocks": [ + { + "bbox": [ + 189, + 270, + 234, + 314 + ], + "lines": [ + { + "bbox": [ + 189, + 270, + 234, + 314 + ], + "spans": [ + { + "bbox": [ + 189, + 270, + 234, + 314 + ], + "type": "image", + "image_path": "917c1e9d880f2e4772a68b9194a7e0a054e967976951b4c4f3a8799838b2f09e.jpg" + } + ] + } + ], + "index": 46, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 196, + 314, + 227, + 323 + ], + "lines": [ + { + "bbox": [ + 196, + 314, + 227, + 323 + ], + "spans": [ + { + "bbox": [ + 196, + 314, + 227, + 323 + ], + "type": "text", + "content": "AWNet" + } + ] + } + ], + "index": 47, + "angle": 0, + "type": "image_caption" + } + ], + "index": 46 + }, + { + "type": "image", + "bbox": [ + 237, + 270, + 282, + 314 + ], + "blocks": [ + { + "bbox": [ + 237, + 270, + 282, + 314 + ], + "lines": [ + { + "bbox": [ + 237, + 270, + 282, + 314 + ], + "spans": [ + { + "bbox": [ + 237, + 270, + 282, + 314 + ], + "type": "image", + "image_path": "7581d9d8b76118ca94beef42656713ac9c3892ab949ceeb6eb406fa43a4e9d72.jpg" + } + ] + } + ], + "index": 48, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 236, + 314, + 282, + 323 + ], + "lines": [ + { + "bbox": [ + 236, + 314, + 282, + 323 + ], + "spans": [ + { + "bbox": [ + 236, + 314, + 282, + 323 + ], + "type": "text", + "content": "DE-UNET*" + } + ] + } + ], + "index": 49, + "angle": 0, + "type": "image_caption" + } + ], + "index": 48 + }, + { + "type": "image", + "bbox": [ + 286, + 270, + 329, + 314 + ], + "blocks": [ + { + "bbox": [ + 286, + 270, + 329, + 314 + ], + "lines": [ + { + "bbox": [ + 286, + 270, + 329, + 314 + ], + "spans": [ + { + "bbox": [ + 286, + 270, + 329, + 314 + ], + "type": "image", + "image_path": "bb7cb2031bd8d94490db9686bc8a0a5393d58c215d8a4192d976d1750caae689.jpg" + } + ] + } + ], + "index": 50, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 291, + 314, + 322, + 323 + ], + "lines": [ + { + "bbox": [ + 291, + 314, + 322, + 323 + ], + "spans": [ + { + "bbox": [ + 291, + 314, + 322, + 323 + ], + "type": "text", + "content": "PyNet*" + } + ] + } + ], + "index": 51, + "angle": 0, + "type": "image_caption" + } + ], + "index": 50 + }, + { + "type": "image", + "bbox": [ + 334, + 270, + 378, + 314 + ], + "blocks": [ + { + "bbox": [ + 334, + 270, + 378, + 314 + ], + "lines": [ + { + "bbox": [ + 334, + 270, + 378, + 314 + ], + "spans": [ + { + "bbox": [ + 334, + 270, + 378, + 314 + ], + "type": "image", + "image_path": "551417cccfd2aa5f9376bbda18fa2f3829c48d51fba0842194d24b60effe900c.jpg" + } + ] + } + ], + "index": 52, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 337, + 314, + 373, + 323 + ], + "lines": [ + { + "bbox": [ + 337, + 314, + 373, + 323 + ], + "spans": [ + { + "bbox": [ + 337, + 314, + 373, + 323 + ], + "type": "text", + "content": "AWNet*" + } + ] + } + ], + "index": 53, + "angle": 0, + "type": "image_caption" + } + ], + "index": 52 + }, + { + "type": "image", + "bbox": [ + 382, + 270, + 426, + 314 + ], + "blocks": [ + { + "bbox": [ + 382, + 270, + 426, + 314 + ], + "lines": [ + { + "bbox": [ + 382, + 270, + 426, + 314 + ], + "spans": [ + { + "bbox": [ + 382, + 270, + 426, + 314 + ], + "type": "image", + "image_path": "0eebec71e83df9cc38e8ef93358879ed3008d6c100e0e3f08ccf2defd69afe42.jpg" + } + ] + } + ], + "index": 54, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 381, + 314, + 425, + 323 + ], + "lines": [ + { + "bbox": [ + 381, + 314, + 425, + 323 + ], + "spans": [ + { + "bbox": [ + 381, + 314, + 425, + 323 + ], + "type": "text", + "content": "DREUDC*" + } + ] + } + ], + "index": 55, + "angle": 0, + "type": "image_caption" + } + ], + "index": 54 + }, + { + "type": "image", + "bbox": [ + 430, + 270, + 474, + 314 + ], + "blocks": [ + { + "bbox": [ + 430, + 270, + 474, + 314 + ], + "lines": [ + { + "bbox": [ + 430, + 270, + 474, + 314 + ], + "spans": [ + { + "bbox": [ + 430, + 270, + 474, + 314 + ], + "type": "image", + "image_path": "720186f975f12819cef15bb5695abc5a413fc91798a8ab844a3c80122a87a805.jpg" + } + ] + } + ], + "index": 56, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 445, + 314, + 459, + 322 + ], + "lines": [ + { + "bbox": [ + 445, + 314, + 459, + 322 + ], + "spans": [ + { + "bbox": [ + 445, + 314, + 459, + 322 + ], + "type": "text", + "content": "GT" + } + ] + } + ], + "index": 57, + "angle": 0, + "type": "image_caption" + } + ], + "index": 56 + }, + { + "bbox": [ + 130, + 413, + 482, + 498 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 413, + 482, + 498 + ], + "spans": [ + { + "bbox": [ + 130, + 413, + 482, + 498 + ], + "type": "text", + "content": "surpassing BNUDC [15] by 0.81dB and 0.50dB, demonstrating that our method can successfully restore various UDC degradations. With the exception of a lightweight model [31] and the baseline model using raw data input [38], our method is the most efficient when it comes to computation while showing superior performance. Qualitative results are illustrated in Fig. 5, and more results including restoration of scenes that are not displayed on a monitor are available in the Supplementary Material." + } + ] + } + ], + "index": 59 + }, + { + "bbox": [ + 131, + 514, + 438, + 528 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 131, + 514, + 438, + 528 + ], + "spans": [ + { + "bbox": [ + 131, + 514, + 438, + 528 + ], + "type": "text", + "content": "4.5 Analysis on Panel-Specific Degradation Representation" + } + ] + } + ], + "index": 60 + }, + { + "bbox": [ + 130, + 533, + 482, + 629 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 533, + 482, + 629 + ], + "spans": [ + { + "bbox": [ + 130, + 533, + 482, + 629 + ], + "type": "text", + "content": "In this subsection, we analyze the effectiveness of DREUDC and panel-specific degradation representations learned with our proposed method. The strength of DREUDC is reported in Tab. 4 with an ablation study. We observe that our strategy of embedding an implicit representation of degraded UDC images is indeed effective by comparing (a) and (c) of Tab. 4, where Model 1 is our framework with only the restoration network without channel calibration and representation learning, while DREUDC is our entire framework trained with the protocol explained in Sec. 3.2." + } + ] + } + ], + "index": 61 + }, + { + "bbox": [ + 130, + 630, + 482, + 666 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 630, + 482, + 666 + ], + "spans": [ + { + "bbox": [ + 130, + 630, + 482, + 666 + ], + "type": "text", + "content": "However, with only this comparison, we cannot rule out the possibility that the rise in performance is solely because of increased parameters. Therefore, we further compare (a) and (c) with (b), where Model 2 is our framework trained in" + } + ] + } + ], + "index": 62 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "spans": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "type": "text", + "content": "12" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 166, + 91, + 217, + 101 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 166, + 91, + 217, + 101 + ], + "spans": [ + { + "bbox": [ + 166, + 91, + 217, + 101 + ], + "type": "text", + "content": "Y. Oh et al." + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 11 + }, + { + "para_blocks": [ + { + "type": "table", + "bbox": [ + 136, + 135, + 476, + 193 + ], + "blocks": [ + { + "bbox": [ + 214, + 114, + 398, + 125 + ], + "lines": [ + { + "bbox": [ + 214, + 114, + 398, + 125 + ], + "spans": [ + { + "bbox": [ + 214, + 114, + 398, + 125 + ], + "type": "text", + "content": "Table 4: Results on ablation of our method." + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 136, + 135, + 476, + 193 + ], + "lines": [ + { + "bbox": [ + 136, + 135, + 476, + 193 + ], + "spans": [ + { + "bbox": [ + 136, + 135, + 476, + 193 + ], + "type": "table", + "html": "
MethodChannel calibrationRepresentation learningParams MACsAxon 30Z-Fold 3
PNSRSSIMPNSRSSIM
(a) Model 1XX4.52M160.57G26.090.933828.66
(b) Model 2✓X6.44M200.07G26.380.934828.71
(c) DREUDC✓✓6.44M200.07G26.660.936829.04
", + "image_path": "42873a23ad1479e4f9c967e611ab7056ead45740aa386a48684e9312db625d39.jpg" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "table_body" + } + ], + "index": 3 + }, + { + "type": "image", + "bbox": [ + 154, + 224, + 206, + 270 + ], + "blocks": [ + { + "bbox": [ + 154, + 224, + 206, + 270 + ], + "lines": [ + { + "bbox": [ + 154, + 224, + 206, + 270 + ], + "spans": [ + { + "bbox": [ + 154, + 224, + 206, + 270 + ], + "type": "image", + "image_path": "30a8160d912299f840c2c91ff8310327c227fc8485795b99133bf6aaf5c4acfa.jpg" + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 176, + 270, + 186, + 278 + ], + "lines": [ + { + "bbox": [ + 176, + 270, + 186, + 278 + ], + "spans": [ + { + "bbox": [ + 176, + 270, + 186, + 278 + ], + "type": "text", + "content": "(a)" + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 131, + 292, + 482, + 370 + ], + "lines": [ + { + "bbox": [ + 131, + 292, + 482, + 370 + ], + "spans": [ + { + "bbox": [ + 131, + 292, + 482, + 370 + ], + "type": "text", + "content": "Fig. 6: T-SNE [21] visualization of representation generated from the encoder. (a) and (b) show the representations generated from encoders of Model 2 in Tab. 4 for Axon 30 and Z-fold 3 images, respectively, which are trained without representation learning. (d) and (e) show the representations generated from encoders of Model 2 in Tab. 5 for T-OLLED and P-OLED images, respectively, which are trained without representation learning. (c) and (f) illustrate the representations that are generated with our unsupervised triplet learning on our dataset and Zhou's dataset [38]." + } + ] + } + ], + "index": 16, + "angle": 0, + "type": "image_caption" + } + ], + "index": 4 + }, + { + "type": "image", + "bbox": [ + 209, + 224, + 256, + 270 + ], + "blocks": [ + { + "bbox": [ + 209, + 224, + 256, + 270 + ], + "lines": [ + { + "bbox": [ + 209, + 224, + 256, + 270 + ], + "spans": [ + { + "bbox": [ + 209, + 224, + 256, + 270 + ], + "type": "image", + "image_path": "ca577c75228f3381d81ff06c21f555e00c31b65ad3bae9bbb1b223ee79c211ab.jpg" + } + ] + } + ], + "index": 6, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 227, + 270, + 235, + 278 + ], + "lines": [ + { + "bbox": [ + 227, + 270, + 235, + 278 + ], + "spans": [ + { + "bbox": [ + 227, + 270, + 235, + 278 + ], + "type": "text", + "content": "(b)" + } + ] + } + ], + "index": 7, + "angle": 0, + "type": "image_caption" + } + ], + "index": 6 + }, + { + "type": "image", + "bbox": [ + 259, + 225, + 306, + 270 + ], + "blocks": [ + { + "bbox": [ + 259, + 225, + 306, + 270 + ], + "lines": [ + { + "bbox": [ + 259, + 225, + 306, + 270 + ], + "spans": [ + { + "bbox": [ + 259, + 225, + 306, + 270 + ], + "type": "image", + "image_path": "c9c552d6708ed8c7b94a7aa5ae4cf7aec4b7a17cf944c8f105965aca7810b75a.jpg" + } + ] + } + ], + "index": 8, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 279, + 270, + 288, + 278 + ], + "lines": [ + { + "bbox": [ + 279, + 270, + 288, + 278 + ], + "spans": [ + { + "bbox": [ + 279, + 270, + 288, + 278 + ], + "type": "text", + "content": "(c)" + } + ] + } + ], + "index": 9, + "angle": 0, + "type": "image_caption" + } + ], + "index": 8 + }, + { + "type": "image", + "bbox": [ + 309, + 225, + 356, + 270 + ], + "blocks": [ + { + "bbox": [ + 309, + 225, + 356, + 270 + ], + "lines": [ + { + "bbox": [ + 309, + 225, + 356, + 270 + ], + "spans": [ + { + "bbox": [ + 309, + 225, + 356, + 270 + ], + "type": "image", + "image_path": "985738607ba8d6416cbec90cb2aa1ee6e37818d715dca0d1e4e90c8bbebf302c.jpg" + } + ] + } + ], + "index": 10, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 329, + 270, + 338, + 278 + ], + "lines": [ + { + "bbox": [ + 329, + 270, + 338, + 278 + ], + "spans": [ + { + "bbox": [ + 329, + 270, + 338, + 278 + ], + "type": "text", + "content": "(d)" + } + ] + } + ], + "index": 11, + "angle": 0, + "type": "image_caption" + } + ], + "index": 10 + }, + { + "type": "image", + "bbox": [ + 361, + 225, + 407, + 269 + ], + "blocks": [ + { + "bbox": [ + 361, + 225, + 407, + 269 + ], + "lines": [ + { + "bbox": [ + 361, + 225, + 407, + 269 + ], + "spans": [ + { + "bbox": [ + 361, + 225, + 407, + 269 + ], + "type": "image", + "image_path": "766f86b111c0330b006cea3df6e848b7d2b79161ad9c9acaa15edfa69e5e8252.jpg" + } + ] + } + ], + "index": 12, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 380, + 270, + 388, + 278 + ], + "lines": [ + { + "bbox": [ + 380, + 270, + 388, + 278 + ], + "spans": [ + { + "bbox": [ + 380, + 270, + 388, + 278 + ], + "type": "text", + "content": "(e)" + } + ] + } + ], + "index": 13, + "angle": 0, + "type": "image_caption" + } + ], + "index": 12 + }, + { + "type": "image", + "bbox": [ + 410, + 225, + 457, + 269 + ], + "blocks": [ + { + "bbox": [ + 410, + 225, + 457, + 269 + ], + "lines": [ + { + "bbox": [ + 410, + 225, + 457, + 269 + ], + "spans": [ + { + "bbox": [ + 410, + 225, + 457, + 269 + ], + "type": "image", + "image_path": "e2b861872f2ccfa4529898572c04002f18e2d97e8ef9c3044a66afb75ae79f25.jpg" + } + ] + } + ], + "index": 14, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 430, + 270, + 438, + 278 + ], + "lines": [ + { + "bbox": [ + 430, + 270, + 438, + 278 + ], + "spans": [ + { + "bbox": [ + 430, + 270, + 438, + 278 + ], + "type": "text", + "content": "(f)" + } + ] + } + ], + "index": 15, + "angle": 0, + "type": "image_caption" + } + ], + "index": 14 + }, + { + "bbox": [ + 130, + 403, + 482, + 523 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 403, + 482, + 523 + ], + "spans": [ + { + "bbox": [ + 130, + 403, + 482, + 523 + ], + "type": "text", + "content": "an end-to-end manner by training the encoder and the restoration network jointly with " + }, + { + "bbox": [ + 130, + 403, + 482, + 523 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_1" + }, + { + "bbox": [ + 130, + 403, + 482, + 523 + ], + "type": "text", + "content": " loss without our proposed representation learning. By comparing Model 1 and Model 2, we see that channel calibration with representations learned from a single dataset without triplet training increases the performance marginally. On the contrary, when given an appropriate representation learned from two datasets by leveraging the difference in degradations with our proposed method, we notice that the restoration model benefits substantially from it. Moreover, it is noteworthy that our proposed embedding scheme has a negligible effect on the computational increase (MACs) of the restoration network, as most of the increase is due to the encoder." + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 130, + 529, + 482, + 589 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 529, + 482, + 589 + ], + "spans": [ + { + "bbox": [ + 130, + 529, + 482, + 589 + ], + "type": "text", + "content": "The validity of our panel-specific degradation representation learning scheme is visualized in Fig. 6. Without representation learning (Fig. 6(a) and (b)), the encoders are unable to distinguish different degradations. Conversely, with representation learning (Fig. 6(c)), the encoder identifies the degradation and discriminates them in clusters, offering advantageous priors for the restoration network." + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 130, + 594, + 482, + 666 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 594, + 482, + 666 + ], + "spans": [ + { + "bbox": [ + 130, + 594, + 482, + 666 + ], + "type": "text", + "content": "This finding suggests an approach to train a restoration network by making good use of datasets other than the one gathered for a specific UDC system. Typically, a specific dataset is used to train a restoration network tailored for a particular UDC system. However, this dataset would be sub-optimal or ineffective when future UDC smartphones have different panels. Our method offers a way to utilize other datasets collected with different panels, mediating this issue." + } + ] + } + ], + "index": 19 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 288, + 91, + 447, + 102 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 288, + 91, + 447, + 102 + ], + "spans": [ + { + "bbox": [ + 288, + 91, + 447, + 102 + ], + "type": "text", + "content": "Panel-Specific UDC Image Restoration" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 470, + 91, + 481, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 470, + 91, + 481, + 100 + ], + "spans": [ + { + "bbox": [ + 470, + 91, + 481, + 100 + ], + "type": "text", + "content": "13" + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 12 + }, + { + "para_blocks": [ + { + "type": "table", + "bbox": [ + 135, + 135, + 476, + 203 + ], + "blocks": [ + { + "bbox": [ + 138, + 114, + 474, + 126 + ], + "lines": [ + { + "bbox": [ + 138, + 114, + 474, + 126 + ], + "spans": [ + { + "bbox": [ + 138, + 114, + 474, + 126 + ], + "type": "text", + "content": "Table 5: Results of our method and ablation on T-OLED/P-OLED datasets [38]." + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 135, + 135, + 476, + 203 + ], + "lines": [ + { + "bbox": [ + 135, + 135, + 476, + 203 + ], + "spans": [ + { + "bbox": [ + 135, + 135, + 476, + 203 + ], + "type": "table", + "html": "
MethodChannel calibrationRepresentation learningParams MACsT-OLEDP-OLED
PSNR SSIMPSNR SSIM
DE-UNET [38]--8.94M 69.29G36.70 0.974528.58 0.9344
Model 1XX4.52M 160.57G37.37 0.977430.85 0.9510
Model 2✓X6.44M 200.07G37.46 0.977731.01 0.9515
DREUDC✓✓6.44M 200.07G37.55 0.977831.17 0.9521
", + "image_path": "0ee50de85ce3425557997c1b7b2b04db5dc15488b67f4a483068f3e31c9e619b.jpg" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "table_body" + } + ], + "index": 3 + }, + { + "bbox": [ + 130, + 226, + 481, + 393 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 226, + 481, + 393 + ], + "spans": [ + { + "bbox": [ + 130, + 226, + 481, + 393 + ], + "type": "text", + "content": "Application to previous datasets Our restoration method using panel-specific degradation representations is also applied to public UDC datasets to show that it generalizes well to other datasets. Our method requires a real dataset comprised of degraded images taken with more than a single panel from the same capturing environment, so we choose Zhou's P-OLED/T-OLED datasets [38]. DREUDC is designed to take raw images as input, so we make a pseudo-raw version of the dataset by mosaicing the RGB images. This pseudo-raw version of the RGB images that used to train the methods, including a new encoder, is created by reversing the in-camera pipeline mentioned in [38]. We only reversed the demosaicing step as it was the only mentioned process; hence we call it 'pseudo-raw' as the images have not been fully reversed to the original high-bit raw data. We choose DE-UNET [38] to compare with our results because it is a model that can be trained with RAW-to-RGB setting. Additional results on RGB image restoration are provided in the Supplementary Material." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 130, + 394, + 482, + 513 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 394, + 482, + 513 + ], + "spans": [ + { + "bbox": [ + 130, + 394, + 482, + 513 + ], + "type": "text", + "content": "Tab. 5 shows the results on T-OLED/P-OLED datasets. Our method outperforms DE-UNET on both panels with slightly fewer parameters. Most importantly, our strategy of representation embedding is also effective on [38] when we compare Model 1 to DREUDC, showing an increase of " + }, + { + "bbox": [ + 130, + 394, + 482, + 513 + ], + "type": "inline_equation", + "content": "0.18\\mathrm{dB}" + }, + { + "bbox": [ + 130, + 394, + 482, + 513 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 130, + 394, + 482, + 513 + ], + "type": "inline_equation", + "content": "0.32\\mathrm{dB}" + }, + { + "bbox": [ + 130, + 394, + 482, + 513 + ], + "type": "text", + "content": " in the restoration of T-OLED and P-OLED images. We also prove that the boost in performance is not only due to increased parameters but also because we enhance the restoration network with effective representations of the degradations learned using our training scheme by comparing Model 2 and DREUDC. Visualization of the representations are provided in Fig. 6(d), (e), and (f). Overall, the results imply that our method is also effective on other UDC datasets." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 132, + 532, + 220, + 544 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 532, + 220, + 544 + ], + "spans": [ + { + "bbox": [ + 132, + 532, + 220, + 544 + ], + "type": "text", + "content": "5 Conclusion" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 130, + 557, + 482, + 666 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 557, + 482, + 666 + ], + "spans": [ + { + "bbox": [ + 130, + 557, + 482, + 666 + ], + "type": "text", + "content": "We have collected a new UDC dataset using commercial UDC-equipped smartphones and demonstrated that UDC image restoration using raw sensor images achieves better performance than using RGB images. We have also proposed a UDC image restoration framework that generates panel-specific degradation representations in the embedding space with a novel unsupervised representation learning scheme. Our restoration network efficiently accommodates the learned representations as a prior with our proposed Transformer blocks, improving the performance. Experiments show that our method achieves state-of-the-art performance on our dataset and is generalizable to previous datasets." + } + ] + } + ], + "index": 7 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "spans": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "type": "text", + "content": "14" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 166, + 91, + 217, + 101 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 166, + 91, + 217, + 101 + ], + "spans": [ + { + "bbox": [ + 166, + 91, + 217, + 101 + ], + "type": "text", + "content": "Y. Oh et al." + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 13 + }, + { + "para_blocks": [ + { + "bbox": [ + 133, + 114, + 234, + 129 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 114, + 234, + 129 + ], + "spans": [ + { + "bbox": [ + 133, + 114, + 234, + 129 + ], + "type": "text", + "content": "Acknowledgment" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 130, + 138, + 482, + 224 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 138, + 482, + 224 + ], + "spans": [ + { + "bbox": [ + 130, + 138, + 482, + 224 + ], + "type": "text", + "content": "This paper was result of the research project supported by SK hynix Inc., including the support of camera equipments and lab environment for collecting the dataset. It was also partially supported by Institute of Information & communications Technology Planning & Evaluation (IITP) grant funded by the Korea government (MSIT) (No.2021-0-01062), and in part by the BK21 FOUR program of the Education and Research Program for Future ICT Pioneers, Seoul National University in 2024." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 133, + 239, + 197, + 251 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 239, + 197, + 251 + ], + "spans": [ + { + "bbox": [ + 133, + 239, + 197, + 251 + ], + "type": "text", + "content": "References" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 134, + 262, + 480, + 665 + ], + "type": "list", + "angle": 0, + "index": 17, + "blocks": [ + { + "bbox": [ + 138, + 262, + 480, + 296 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 262, + 480, + 296 + ], + "spans": [ + { + "bbox": [ + 138, + 262, + 480, + 296 + ], + "type": "text", + "content": "1. Abdelhamed, A., Lin, S., Brown, M.S.: A high-quality denoising dataset for smartphone cameras. In: Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition. pp. 1692-1700 (2018)" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 138, + 296, + 480, + 328 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 296, + 480, + 328 + ], + "spans": [ + { + "bbox": [ + 138, + 296, + 480, + 328 + ], + "type": "text", + "content": "2. Agustsson, E., Timofte, R.: Ntire 2017 challenge on single image super-resolution: Dataset and study. In: Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition Workshops. pp. 126-135 (2017)" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 138, + 328, + 480, + 361 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 328, + 480, + 361 + ], + "spans": [ + { + "bbox": [ + 138, + 328, + 480, + 361 + ], + "type": "text", + "content": "3. Bay, H., Tuytelaars, T., Van Gool, L.: Surf: Speeded up robust features. In: Computer Vision-ECCV 2006: 9th European Conference on Computer Vision, Graz, Austria, May 7-13, 2006. Proceedings, Part I 9. pp. 404-417. Springer (2006)" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 138, + 361, + 480, + 393 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 361, + 480, + 393 + ], + "spans": [ + { + "bbox": [ + 138, + 361, + 480, + 393 + ], + "type": "text", + "content": "4. Chen, C., Chen, Q., Xu, J., Koltun, V.: Learning to see in the dark. In: Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition. pp. 3291-3300 (2018)" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 138, + 393, + 480, + 426 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 393, + 480, + 426 + ], + "spans": [ + { + "bbox": [ + 138, + 393, + 480, + 426 + ], + "type": "text", + "content": "5. Chen, T., Kornblith, S., Norouzi, M., Hinton, G.: A simple framework for contrastive learning of visual representations. In: International Conference on Machine Learning. pp. 1597-1607. PMLR (2020)" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 138, + 426, + 480, + 447 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 426, + 480, + 447 + ], + "spans": [ + { + "bbox": [ + 138, + 426, + 480, + 447 + ], + "type": "text", + "content": "6. Chen, X., Fan, H., Girshick, R., He, K.: Improved baselines with momentum contrastive learning. arXiv preprint arXiv:2003.04297 (2020)" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 138, + 448, + 480, + 480 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 448, + 480, + 480 + ], + "spans": [ + { + "bbox": [ + 138, + 448, + 480, + 480 + ], + "type": "text", + "content": "7. Dai, L., Liu, X., Li, C., Chen, J.: Attnective wavelet network for image isp. In: Computer Vision-ECCV 2020 Workshops: Glasgow, UK, August 23-28, 2020, Proceedings, Part III 16. pp. 185-201. Springer (2020)" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 138, + 481, + 480, + 513 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 481, + 480, + 513 + ], + "spans": [ + { + "bbox": [ + 138, + 481, + 480, + 513 + ], + "type": "text", + "content": "8. Deng, W., Zheng, L., Sun, Y., Jiao, J.: Rethinking triplet loss for domain adaptation. IEEE Transactions on Circuits and Systems for Video Technology 31(1), 29-37 (2020)" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 138, + 514, + 480, + 556 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 514, + 480, + 556 + ], + "spans": [ + { + "bbox": [ + 138, + 514, + 480, + 556 + ], + "type": "text", + "content": "9. Feng, R., Li, C., Chen, H., Li, S., Gu, J., Loy, C.C.: Generating aligned pseudosupervision from non-aligned data for image restoration in under-display camera. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 5013-5022 (2023)" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 134, + 556, + 480, + 600 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 134, + 556, + 480, + 600 + ], + "spans": [ + { + "bbox": [ + 134, + 556, + 480, + 600 + ], + "type": "text", + "content": "10. Feng, R., Li, C., Chen, H., Li, S., Loy, C.C., Gu, J.: Removing diffraction image artifacts in under-display camera via dynamic skip connection network. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 662-671 (2021)" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 134, + 601, + 480, + 643 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 134, + 601, + 480, + 643 + ], + "spans": [ + { + "bbox": [ + 134, + 601, + 480, + 643 + ], + "type": "text", + "content": "11. Feng, R., Li, C., Zhou, S., Sun, W., Zhu, Q., Jiang, J., Yang, Q., Loy, C.C., Gu, J., Zhu, Y., et al.: Mipi 2022 challenge on under-display camera image restoration: Methods and results. In: European Conference on Computer Vision. pp. 60-77. Springer (2022)" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 134, + 643, + 480, + 665 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 134, + 643, + 480, + 665 + ], + "spans": [ + { + "bbox": [ + 134, + 643, + 480, + 665 + ], + "type": "text", + "content": "12. Guizar-Sicairos, M., Thurman, S.T., Fienup, J.R.: Efficient subpixel image registration algorithms. Optics letters 33(2), 156-158 (2008)" + } + ] + } + ], + "index": 16 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 288, + 91, + 447, + 102 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 288, + 91, + 447, + 102 + ], + "spans": [ + { + "bbox": [ + 288, + 91, + 447, + 102 + ], + "type": "text", + "content": "Panel-Specific UDC Image Restoration" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 470, + 92, + 480, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 470, + 92, + 480, + 100 + ], + "spans": [ + { + "bbox": [ + 470, + 92, + 480, + 100 + ], + "type": "text", + "content": "15" + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 14 + }, + { + "para_blocks": [ + { + "bbox": [ + 133, + 117, + 480, + 665 + ], + "type": "list", + "angle": 0, + "index": 18, + "blocks": [ + { + "bbox": [ + 133, + 117, + 480, + 138 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 117, + 480, + 138 + ], + "spans": [ + { + "bbox": [ + 133, + 117, + 480, + 138 + ], + "type": "text", + "content": "13. Hermans, A., Beyer, L., Leibe, B.: In defense of the triplet loss for person re-identification. arXiv preprint arXiv:1703.07737 (2017)" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 133, + 140, + 480, + 161 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 140, + 480, + 161 + ], + "spans": [ + { + "bbox": [ + 133, + 140, + 480, + 161 + ], + "type": "text", + "content": "14. Ignatov, A., Van Gool, L., Timofte, R.: Replacing mobile camera isp with a single deep learning model. arXiv preprint arXiv:2002.05509 (2020)" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 134, + 163, + 480, + 195 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 134, + 163, + 480, + 195 + ], + "spans": [ + { + "bbox": [ + 134, + 163, + 480, + 195 + ], + "type": "text", + "content": "15. Koh, J., Lee, J., Yoon, S.: Bnudc: a two-branched deep neural network for restoring images from under-display cameras. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 1950-1959 (2022)" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 134, + 197, + 480, + 228 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 134, + 197, + 480, + 228 + ], + "spans": [ + { + "bbox": [ + 134, + 197, + 480, + 228 + ], + "type": "text", + "content": "16. Kong, L., Dong, J., Ge, J., Li, M., Pan, J.: Efficient frequency domain-based transformers for high-quality image deblurring. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 5886-5895 (2023)" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 134, + 230, + 480, + 273 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 134, + 230, + 480, + 273 + ], + "spans": [ + { + "bbox": [ + 134, + 230, + 480, + 273 + ], + "type": "text", + "content": "17. Kwon, K., Kang, E., Lee, S., Lee, S.J., Lee, H.E., Yoo, B., Han, J.J.: Controllable image restoration for under-display camera in smartphones. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 2073-2082 (2021)" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 134, + 274, + 480, + 307 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 134, + 274, + 480, + 307 + ], + "spans": [ + { + "bbox": [ + 134, + 274, + 480, + 307 + ], + "type": "text", + "content": "18. Lee, W., Son, S., Lee, K.M.: Ap-bsn: Self-supervised denoising for real-world images via asymmetric pd and blind-spot network. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 17725-17734 (2022)" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 134, + 308, + 480, + 329 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 134, + 308, + 480, + 329 + ], + "spans": [ + { + "bbox": [ + 134, + 308, + 480, + 329 + ], + "type": "text", + "content": "19. Liang, C.H., Chen, Y.A., Liu, Y.C., Hsu, W.H.: Raw image deblurring. IEEE Transactions on Multimedia 24, 61-72 (2020)" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 133, + 331, + 480, + 352 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 331, + 480, + 352 + ], + "spans": [ + { + "bbox": [ + 133, + 331, + 480, + 352 + ], + "type": "text", + "content": "20. Loshchilov, I., Hutter, F.: Sgdr: Stochastic gradient descent with warm restarts. arXiv preprint arXiv:1608.03983 (2016)" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 133, + 354, + 480, + 374 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 354, + 480, + 374 + ], + "spans": [ + { + "bbox": [ + 133, + 354, + 480, + 374 + ], + "type": "text", + "content": "21. Van der Maaten, L., Hinton, G.: Visualizing data using t-sne. Journal of machine learning research 9(11) (2008)" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 133, + 376, + 480, + 418 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 376, + 480, + 418 + ], + "spans": [ + { + "bbox": [ + 133, + 376, + 480, + 418 + ], + "type": "text", + "content": "22. Nam, S., Hwang, Y., Matsushita, Y., Kim, S.J.: A holistic approach to cross-channel image noise modeling and its application to image denoising. In: Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition. pp. 1683-1691 (2016)" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 133, + 421, + 480, + 463 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 421, + 480, + 463 + ], + "spans": [ + { + "bbox": [ + 133, + 421, + 480, + 463 + ], + "type": "text", + "content": "23. Oh, Y., Park, G.Y., Cho, N.I.: Restoration of high-frequency components in under display camera images. In: 2022 Asia-Pacific Signal and Information Processing Association Annual Summit and Conference (APSIPA ASC). pp. 1040-1046. IEEE (2022)" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 133, + 465, + 480, + 508 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 465, + 480, + 508 + ], + "spans": [ + { + "bbox": [ + 133, + 465, + 480, + 508 + ], + "type": "text", + "content": "24. Oh, Y., Park, G.Y., Chung, H., Cho, S., Cho, N.I.: Residual dilated u-net with spatially adaptive normalization for the restoration of under display camera images. In: 2021 Asia-Pacific Signal and Information Processing Association Annual Summit and Conference (APSIPA ASC). pp. 151-157. IEEE (2021)" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 133, + 510, + 480, + 542 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 510, + 480, + 542 + ], + "spans": [ + { + "bbox": [ + 133, + 510, + 480, + 542 + ], + "type": "text", + "content": "25. Plotz, T., Roth, S.: Benchmarking denoising algorithms with real photographs. In: Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition. pp. 1586-1595 (2017)" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 133, + 544, + 480, + 586 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 544, + 480, + 586 + ], + "spans": [ + { + "bbox": [ + 133, + 544, + 480, + 586 + ], + "type": "text", + "content": "26. Punnappurath, A., Abuolaim, A., Abdelhamed, A., Levinshtein, A., Brown, M.S.: Day-to-night image synthesis for training nighttime neural isps. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 10769-10778 (2022)" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 133, + 588, + 480, + 631 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 588, + 480, + 631 + ], + "spans": [ + { + "bbox": [ + 133, + 588, + 480, + 631 + ], + "type": "text", + "content": "27. Ronneberger, O., Fischer, P., Brox, T.: U-net: Convolutional networks for biomedical image segmentation. In: Medical Image Computing and Computer-Assisted Intervention-MICCAI 2015: 18th International Conference, Munich, Germany, October 5-9, 2015, Proceedings, Part III 18. pp. 234-241. Springer (2015)" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 133, + 632, + 480, + 665 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 632, + 480, + 665 + ], + "spans": [ + { + "bbox": [ + 133, + 632, + 480, + 665 + ], + "type": "text", + "content": "28. Schroff, F., Kalenichenko, D., Philbin, J.: Facenet: A unified embedding for face recognition and clustering. In: Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition. pp. 815-823 (2015)" + } + ] + } + ], + "index": 17 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "spans": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "type": "text", + "content": "16" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 166, + 91, + 216, + 100 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 166, + 91, + 216, + 100 + ], + "spans": [ + { + "bbox": [ + 166, + 91, + 216, + 100 + ], + "type": "text", + "content": "Y. Oh et al." + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 15 + }, + { + "para_blocks": [ + { + "bbox": [ + 130, + 116, + 482, + 512 + ], + "type": "list", + "angle": 0, + "index": 12, + "blocks": [ + { + "bbox": [ + 130, + 116, + 482, + 149 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 116, + 482, + 149 + ], + "spans": [ + { + "bbox": [ + 130, + 116, + 482, + 149 + ], + "type": "text", + "content": "29. Schwartz, E., Giryes, R., Bronstein, A.M.: Deepisp: Toward learning an end-to-end image processing pipeline. IEEE Transactions on Image Processing 28(2), 912-923 (2018)" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 130, + 150, + 482, + 183 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 150, + 482, + 183 + ], + "spans": [ + { + "bbox": [ + 130, + 150, + 482, + 183 + ], + "type": "text", + "content": "30. Song, B., Chen, X., Xu, S., Zhou, J.: Under-display camera image restoration with scattering effect. In: Proceedings of the IEEE/CVF International Conference on Computer Vision. pp. 12580-12589 (2023)" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 130, + 183, + 482, + 226 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 183, + 482, + 226 + ], + "spans": [ + { + "bbox": [ + 130, + 183, + 482, + 226 + ], + "type": "text", + "content": "31. Sundar, V., Hegde, S., Kothandaraman, D., Mitra, K.: Deep atrous guided filter for image restoration in under display cameras. In: Computer Vision-ECCV 2020 Workshops: Glasgow, UK, August 23-28, 2020, Proceedings, Part V 16. pp. 379-397. Springer (2020)" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 130, + 226, + 482, + 270 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 226, + 482, + 270 + ], + "spans": [ + { + "bbox": [ + 130, + 226, + 482, + 270 + ], + "type": "text", + "content": "32. Wang, L., Wang, Y., Dong, X., Xu, Q., Yang, J., An, W., Guo, Y.: Unsupervised degradation representation learning for blind super-resolution. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 10581-10590 (2021)" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 130, + 270, + 482, + 304 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 270, + 482, + 304 + ], + "spans": [ + { + "bbox": [ + 130, + 270, + 482, + 304 + ], + "type": "text", + "content": "33. Yu, B., Liu, T., Gong, M., Ding, C., Tao, D.: Correcting the triplet selection bias for triplet loss. In: Proceedings of the European Conference on Computer Vision (ECCV). pp. 71-87 (2018)" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 130, + 304, + 482, + 346 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 304, + 482, + 346 + ], + "spans": [ + { + "bbox": [ + 130, + 304, + 482, + 346 + ], + "type": "text", + "content": "34. Yuan, Y., Chen, W., Yang, Y., Wang, Z.: In defense of the triplet loss again: Learning robust person re-identification with fast approximated triplet loss and label distillation. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition Workshops. pp. 354-355 (2020)" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 130, + 346, + 482, + 390 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 346, + 482, + 390 + ], + "spans": [ + { + "bbox": [ + 130, + 346, + 482, + 390 + ], + "type": "text", + "content": "35. Zamir, S.W., Arora, A., Khan, S., Hayat, M., Khan, F.S., Yang, M.H., Shao, L.: Cycleisp: Real image restoration via improved data synthesis. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 2696-2705 (2020)" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 130, + 390, + 482, + 423 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 390, + 482, + 423 + ], + "spans": [ + { + "bbox": [ + 130, + 390, + 482, + 423 + ], + "type": "text", + "content": "36. Zhou, Y., Song, Y., Du, X.: Modular degradation simulation and restoration for under-display camera. In: Proceedings of the Asian Conference on Computer Vision. pp. 265-282 (2022)" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 130, + 423, + 482, + 479 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 423, + 482, + 479 + ], + "spans": [ + { + "bbox": [ + 130, + 423, + 482, + 479 + ], + "type": "text", + "content": "37. Zhou, Y., Kwan, M., Tolentino, K., Emerton, N., Lim, S., Large, T., Fu, L., Pan, Z., Li, B., Yang, Q., et al.: Udc 2020 challenge on image restoration of under-display camera: Methods and results. In: Computer Vision-ECCV 2020 Workshops: Glasgow, UK, August 23-28, 2020, Proceedings, Part V 16. pp. 337-351. Springer (2020)" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 130, + 479, + 482, + 512 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 479, + 482, + 512 + ], + "spans": [ + { + "bbox": [ + 130, + 479, + 482, + 512 + ], + "type": "text", + "content": "38. Zhou, Y., Ren, D., Emerton, N., Lim, S., Large, T.: Image restoration for underdisplay camera. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 9179-9188 (2021)" + } + ] + } + ], + "index": 11 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 288, + 91, + 447, + 102 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 288, + 91, + 447, + 102 + ], + "spans": [ + { + "bbox": [ + 288, + 91, + 447, + 102 + ], + "type": "text", + "content": "Panel-Specific UDC Image Restoration" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 470, + 91, + 481, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 470, + 91, + 481, + 100 + ], + "spans": [ + { + "bbox": [ + 470, + 91, + 481, + 100 + ], + "type": "text", + "content": "17" + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 16 + } + ], + "_backend": "vlm", + "_version_name": "2.6.4" +} \ No newline at end of file diff --git a/2024/PanoFree_ Tuning-Free Holistic Multi-view Image Generation with Cross-view Self-Guidance/cbd590c4-f320-4c08-83b4-3c8681128a51_content_list.json b/2024/PanoFree_ Tuning-Free Holistic Multi-view Image Generation with Cross-view Self-Guidance/cbd590c4-f320-4c08-83b4-3c8681128a51_content_list.json new file mode 100644 index 0000000000000000000000000000000000000000..264dc701e3307effaabf55e972ff900a1410edfe --- /dev/null +++ b/2024/PanoFree_ Tuning-Free Holistic Multi-view Image Generation with Cross-view Self-Guidance/cbd590c4-f320-4c08-83b4-3c8681128a51_content_list.json @@ -0,0 +1,1810 @@ +[ + { + "type": "text", + "text": "PanoFree: Tuning-Free Holistic Multi-view Image Generation with Cross-view Self-Guidance", + "text_level": 1, + "bbox": [ + 218, + 140, + 782, + 185 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Aoming Liu $^{1,2*}$ , Zhong Li $^{1\\dagger}$ , Zhang Chen $^{1\\dagger}$ , Nannan Li $^{2}$ , Yi Xu $^{1}$ , and Bryan A. Plummer $^{2}$", + "bbox": [ + 218, + 210, + 782, + 243 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "$^{1}$ OPPO US Research Center, Palo Alto, CA 94303, USA \n $^{2}$ Boston University, Boston, MA 02215, USA", + "bbox": [ + 305, + 253, + 694, + 282 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Abstract. Immersive scene generation, notably panorama creation, benefits significantly from the adaptation of large pre-trained text-to-image (T2I) models for multi-view image generation. Due to the high cost of acquiring multi-view images, tuning-free generation is preferred. However, existing methods are either limited to simple correspondences or require extensive fine-tuning to capture complex ones. We present PanoFree, a novel method for tuning-free multi-view image generation that supports an extensive array of correspondences. PanoFree sequentially generates multi-view images using iterative warping and inpainting, addressing the key issues of inconsistency and artifacts from error accumulation without the need for fine-tuning. It improves error accumulation by enhancing cross-view awareness and refines the warping and inpainting processes via cross-view guidance, risky area estimation and erasing, and symmetric bidirectional guided generation for loop closure, alongside guidance-based semantic and density control for scene structure preservation. In experiments on Planar, $360^{\\circ}$ , and Full Spherical Panoramas, PanoFree demonstrates significant error reduction, improves global consistency, and boosts image quality without extra fine-tuning. Compared to existing methods, PanoFree is up to 5x more efficient in time and 3x more efficient in GPU memory usage, and maintains superior diversity of results (2x better in our user study). PanoFree offers a viable alternative to costly fine-tuning or the use of additional pre-trained models.", + "bbox": [ + 261, + 320, + 743, + 625 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Keywords: Tuning-free generation; Multi-view Image, Panorama", + "bbox": [ + 261, + 638, + 705, + 652 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "1 Introduction", + "text_level": 1, + "bbox": [ + 215, + 679, + 375, + 695 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Text-to-image (T2I) generation over multiple views for immersive scenes, like panorama generation, is a challenging task requiring coherence and diversity among many generated images (e.g., [3, 8, 12, 16, 17, 21, 36, 37, 39, 48]). Early work using GANs or VAEs (e.g., [4-6, 11, 24, 25, 47, 50]) have been replaced recently with diffusion-based models (e.g., [1, 9, 10, 22, 23, 46, 49, 51, 52, 54]), often leveraging Stable Diffusion [38]. State-of-the-art panorama generation methods", + "bbox": [ + 212, + 710, + 787, + 801 + ], + "page_idx": 0 + }, + { + "type": "page_footnote", + "text": "*Work partly completed during Aoming's internship at OPPO US Research Center.", + "bbox": [ + 236, + 810, + 785, + 824 + ], + "page_idx": 0 + }, + { + "type": "page_footnote", + "text": "†Corresponding Authors.", + "bbox": [ + 240, + 824, + 406, + 839 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/62107d5d06df8c37b569648ab5279266d0f49673f42d9a021305129cce9c2d81.jpg", + "image_caption": [ + "Fig. 1: PanoFree can generate multi-view images according to different types of correspondences without fine-tuning, and a natural application is tuning-free generation for different types of panoramas. We demonstrate this by generating three commonly used panoramas. Top: Planar Panorama; Middle: 360 Panorama; Bottom: Full Panorama." + ], + "image_footnote": [], + "bbox": [ + 274, + 143, + 723, + 378 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "use Joint Diffusion (e.g., [1,22,46]), where parallel diffusion processes to generate multi-view images and enhancing global consistency by fusing latent or attention features based on cross-view correspondences. However, we find these methods can only generate some types of panoramas, limiting their scope (e.g., [1,22]), or require fine-tuning using expensive panorama datasets (e.g., [22,46]).", + "bbox": [ + 212, + 454, + 787, + 531 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "To address these challenges, we propose PanoFree a tuning-free multi-view image generation method using iterative warping and inpainting of perspective images to support diverse correspondences with low costs (see Fig. 1 for example generations). Iterative warping and inpainting of perspective images provide a means to the diverse multi-view correspondences required in panorama generation without additional fine-tuning [4, 11, 18]. However, recent work has overlooked these benefits due to accumulated errors causing suboptimal image quality [1, 22, 46]. We find that most accumulated errors from iterative warping can be attributed to the deficient conditions during generation. Specifically, conditioning solely on the previous image narrows cross-view awareness, leading to inconsistencies. Warping and inpainting can also propagate noise, e.g., truncated objects or jagged edges. Additionally, the given conditions may be incomplete to meet specific requirements, such as ensuring 360-degree consistency for loop closure and maintaining correct spatial relationships for realistic scenes.", + "bbox": [ + 212, + 534, + 787, + 744 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "To address these issues, PanoFree expand cross-view awareness by conditioning the current view on multiple views with guided image synthesis techniques such as SDEdit [29]. Then, PanoFree estimates and erases the risky areas, regions likely containing noise, to reduce the noise introduced by warping and inpainting. In addition, PanoFree adopts a bidirectional generation path with a symmetrical conditioning strategy for loop closure. Lastly, PanoFree further", + "bbox": [ + 212, + 750, + 787, + 840 + ], + "page_idx": 1 + }, + { + "type": "page_number", + "text": "2", + "bbox": [ + 217, + 114, + 228, + 126 + ], + "page_idx": 1 + }, + { + "type": "header", + "text": "A. Liu et al.", + "bbox": [ + 271, + 114, + 354, + 126 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "utilizes pseudo global guidance with region-specific semantic and density control to make scene structure more reasonable.", + "bbox": [ + 217, + 146, + 784, + 175 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "We evaluate PanoFree on three text-to-panorama generation tasks: Planar, $360^{\\circ}$ , and Full Spherical Panoramas. PanoFree effectively alleviating accumulated errors in sequential generation, and significantly improves image quality and global consistency (e.g. $31.6\\%$ better in FID). This enables PanoFree to have better (or at least comparable) results to the state-of-the-art [1,22,46], despite these methods either having narrower applications or requiring fine-tuning datasets. Specifically, PanoFree is up to 5x more efficient in time and 3x more efficient in GPU memory usage, and maintains superior diversity of results (2x better in our user study). Lastly, PanoFree is also highly flexible, enabling it to plug-in-and-play with various pre-trained T2I models and adapters.", + "bbox": [ + 217, + 176, + 785, + 325 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Our contributions can be summarized as follows:", + "bbox": [ + 240, + 329, + 591, + 340 + ], + "page_idx": 2 + }, + { + "type": "list", + "sub_type": "text", + "list_items": [ + "- We introduce PanoFree, a tuning-free multi-view image generation method applicable for various correspondences and pre-trained T2I models. Thus PanoFree can greatly reduce the data and fine-tuning costs for immersive scene generation tasks such as text-to-panorama generation.", + "- We provide a in-depth perspective of accumulated errors and identify the deficient conditions as the main causes. We further effectively rectify deficient conditions and alleviate accumulated errors with the cross-view guidance as well as risky area estimation and erasing in PanoFree.", + "- As far as we know, PanoFree is the first to achieve feasible tuning-free generation for $360^{\\circ}$ Panoramas and Full Spherical Panoramas." + ], + "bbox": [ + 215, + 344, + 792, + 494 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "2 Related Work", + "text_level": 1, + "bbox": [ + 217, + 518, + 385, + 534 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Diffusion Models [15, 19, 40, 42-45] are a popular framework for generative models. Early work required a long trajectory for sampling to produce high-quality samples [7, 45], before being sped up with advanced sampling techniques that also preserved generation quality [20, 26, 27, 41]. Latent Diffusion Models (LDMs) [33, 38] made these models more efficient by training in the latent space.", + "bbox": [ + 217, + 551, + 785, + 628 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "T2I Diffusion and Panorama Generation. Diffusion models are widely adopted for text-to-image (T2I) generation [31, 36, 38, 39]. Many downstream tasks used large pre-trained T2I diffusion models, like Stable Diffusion [38], to boost performance [4, 11, 18], including panorama generation [4, 11, 18]. These methods have largely supplanted GAN and VAE methods [4-6, 11, 24, 25, 32, 47, 50], with most recent work in panorama generation tasks using diffusion models [1, 9, 10, 22, 23, 46, 49, 51, 52, 54]. These diffusion-based panorama generation methods use joint diffusion to fuses multiple diffusion processes through latent or attention manipulation [1, 22, 46]. However, they are either limited to modeling simple correspondences or require extensive fine-tuning to model complex ones.", + "bbox": [ + 217, + 628, + 785, + 777 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Guided Image Synthesis with Diffusion Models. It can be challenging to achieve satisfactory results solely relying on text guidance. Therefore, some prior work [28-30, 53] guide or control the generation results with reference images as fine-grained condition. ControlNet [53] and T2I-Adapter [30] are the most", + "bbox": [ + 217, + 779, + 785, + 839 + ], + "page_idx": 2 + }, + { + "type": "header", + "text": "PanoFree", + "bbox": [ + 666, + 114, + 730, + 126 + ], + "page_idx": 2 + }, + { + "type": "page_number", + "text": "3", + "bbox": [ + 774, + 116, + 784, + 126 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "commonly used methods to incorporate additional image conditions by adding extra image encoders, but they all require few-shot fine-tuning. SDEdit [29] achieves tuning-free guided image synthesis by adding noise to the guide image and then denoising it back to a real image using a pre-trained diffusion model.", + "bbox": [ + 212, + 146, + 787, + 208 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "3 Method", + "text_level": 1, + "bbox": [ + 215, + 228, + 330, + 244 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "PanoFree targets the text-to-panorama generation task, which takes textual descriptions as guidance to create multi-view perspective images that can be stitched into a wide-angle, high-quality panorama. PanoFree generates multiview images through sequential warping and inpainting steps, which typically results in acclimated errors due to deficit conditions (discussed in Sec. 3.1). Each component of Panofree is designed to minimize the effect of these various deficiencies. Specifically, Sec. 3.2 mitigates inconsistency using SDEdit-based cross-view guidance and Sec. 3.3 removes artifact-inducing content by estimating and erasing risky areas. At a higher level, PanoFree employs Bidirectional Generation with Symmetric Guidance for loop closure and error reduction (Sec. 3.4). Additionally, it applies guidance-based semantic and density control for scene structure preservation (Sec. 3.5). See Fig. 2 for an overview of our approach.", + "bbox": [ + 212, + 258, + 787, + 441 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "3.1 Deficient Conditions behind Accumulated Errors", + "text_level": 1, + "bbox": [ + 215, + 459, + 665, + 474 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "In this section, we discuss the causes behind the deficit conditions in the iterative warping and inpainting process and reformulate the problem as conditional generation. Given the text prompt $c_{t}$ , $i$ -th view's image $\\mathbf{x}_i$ , warping function $\\mathcal{W}$ , transformation matrix of the projection from $i$ -th view to the $(i + 1)$ -th view $\\mathbf{P}_i^{i + 1}$ , and pre-trained T2I inpainting model $\\Phi_{inp}$ , the warping and inpainting step to generate the $(i + 1)$ -th view can be denoted as:", + "bbox": [ + 212, + 482, + 787, + 574 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\n\\hat {\\mathbf {x}} _ {i}, \\mathbf {m} _ {i} = \\mathcal {W} \\left(\\mathbf {x} _ {i}, \\mathbf {P} _ {i} ^ {i + 1}\\right); \\quad \\mathbf {x} _ {i + 1} = \\boldsymbol {\\Phi} _ {i n p} \\left(\\hat {\\mathbf {x}} _ {i}, \\mathbf {m} _ {i}, c _ {t}\\right), \\tag {1}\n$$\n", + "text_format": "latex", + "bbox": [ + 326, + 583, + 785, + 599 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "where $\\hat{\\mathbf{x}}_i$ is the image warped from $i$ -th view to $(i + 1)$ -th view and $\\mathbf{m}_i$ is the masking indicating the area to inpaint. And we can simplify the warping and inpainting steps in the following conditional image generation form:", + "bbox": [ + 212, + 608, + 787, + 655 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\n\\mathbf {x} _ {i + 1} \\sim q (\\mathbf {x} | c _ {t}, \\mathbf {x} _ {i}, \\mathbf {P} _ {i} ^ {i + 1}). \\tag {2}\n$$\n", + "text_format": "latex", + "bbox": [ + 406, + 662, + 785, + 680 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "However, during this generation process, we found conditions can become deficient. Major accumulated errors arise from three types of deficient conditions: Biased Conditions, Noisy Conditions, and Partial Conditions. See Sec. A.2 in the supplementary for detailed error illustrations.", + "bbox": [ + 212, + 689, + 787, + 748 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Biased Conditions is the most obvious problem. In the above step, $\\mathbf{x}_{i + 1}$ is solely conditioned on $\\mathbf{x}_i$ , which biases the cross-view awareness heavily to $i$ -th view. If $\\mathbf{x}_i$ has deviated from the desired global distribution in certain aspects, then $\\mathbf{x}_{i + 1}$ is likely to continue deviating in the same direction, resulting in significant inconsistency. We also found that slight style and content shifts often accumulate in this way, leading to significant inconsistency between distant views.", + "bbox": [ + 212, + 750, + 787, + 840 + ], + "page_idx": 3 + }, + { + "type": "page_number", + "text": "4", + "bbox": [ + 217, + 114, + 228, + 126 + ], + "page_idx": 3 + }, + { + "type": "header", + "text": "A. Liu et al.", + "bbox": [ + 271, + 114, + 354, + 127 + ], + "page_idx": 3 + }, + { + "type": "image", + "img_path": "images/f4ba846b44b812434a4f6c4d8a722475914a09d12b41c9f8cc42bacce32643ec.jpg", + "image_caption": [ + "(a) Overview of the generation and guidance framework." + ], + "image_footnote": [], + "bbox": [ + 267, + 143, + 723, + 258 + ], + "page_idx": 4 + }, + { + "type": "image", + "img_path": "images/a6d11eb0e77c32c25a9be5c47e6eb0de6b4e27f38465edc050d87c2e447d6e48.jpg", + "image_caption": [ + "Fig. 2: Overview of our PanoFree method, taking 360 Panorama Generation as an example. (a): At a framework level, PanoFree adopts two generation paths with opposite viewpoint translation or rotation. It enhances consistency by symmetrically selecting views from the other path as guidance to generate a new view (Sec. 3.4). Loop closure is ensured by merging these two paths. (b): In each warping and inpainting step, PanoFree reduces accumulated error by guiding the inpainting process with cross-view images (Sec. 3.2), along with estimating and erasing risky areas (Sec. 3.3)." + ], + "image_footnote": [], + "bbox": [ + 267, + 284, + 723, + 435 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Noisy Conditions mainly refer to $\\mathbf{x}_i$ containing artifact-inducing contents. Existing artifacts in $\\mathbf{x}_i$ could guide inpainting model to generate similar artifacts in $\\mathbf{x}_{i + 1}$ and propagate to every following view. Additionally, disjointed or distorted areas, jagged or sharp content, and objects truncated by edges in $\\mathbf{x}_i$ are also highly risky to introduce artifacts in $\\mathbf{x}_{i + 1}$ .", + "bbox": [ + 212, + 559, + 787, + 633 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Partial Conditions refer to the conditions not containing all the necessary information to meet specific requirements. For example, if we follow Eq. (2) on the final view, we lack information about $\\mathbf{x}_0$ , making it impossible to generate image coherent with $\\mathbf{x}_0$ to ensure loop closure. Additionally, using a single text prompt to generate all views within a full spherical panorama may lead to hallucinations, such as cities floating in the sky or underwater mountains.", + "bbox": [ + 212, + 635, + 787, + 726 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "3.2 Cross-View Guidance", + "text_level": 1, + "bbox": [ + 214, + 750, + 441, + 763 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "To rectify the biased conditions discussed in Sec.3.1 and enlarge cross-view awareness, a natural idea is to let $\\mathbf{x}_{i + 1}$ conditioned on more views,", + "bbox": [ + 214, + 776, + 785, + 808 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\n\\mathbf {x} _ {i + 1} \\sim q (\\mathbf {x} | c _ {t}, \\mathbf {x} _ {i}, \\mathbf {P} _ {i} ^ {i + 1}, \\mathbf {x} _ {1} ^ {g}, \\dots , \\mathbf {x} _ {m} ^ {g}) \\qquad (3)\n$$\n", + "text_format": "latex", + "bbox": [ + 367, + 818, + 785, + 837 + ], + "page_idx": 4 + }, + { + "type": "header", + "text": "PanoFree", + "bbox": [ + 666, + 114, + 730, + 126 + ], + "page_idx": 4 + }, + { + "type": "page_number", + "text": "5", + "bbox": [ + 774, + 116, + 784, + 126 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "where $\\mathbf{x}_1^g, \\ldots, \\mathbf{x}_m^g$ are selected from $\\mathbf{x}_0, \\ldots, \\mathbf{x}_{i-1}$ . This naturally results in a Guided Image Synthesis task form with self-generated images as guidance. Many existing methods can be adapted to implement our design, such as ControlNet [53] and T2I adapter [30]. To avoid relying on fine-tuning and reduce memory and time costs, we choose SDEdit [29], a training-free guided image synthesis approach, with a single guidance image $\\mathbf{x}^g \\in [\\mathbf{x}_0, \\ldots, \\mathbf{x}_{i-1}]$ .", + "bbox": [ + 212, + 146, + 782, + 236 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Guided Image Synthesis using SDEdit. Given $\\mathbf{x}^g$ as guidance, SDEdit establishes a Gaussian distribution using $\\mathbf{x}^g$ as the expectation and the intermediate status at time $t_0$ in the reverse SDE process. The desired data distribution is obtained by gradually removing noise from $\\mathbf{x}^g(t_0)$ :", + "bbox": [ + 212, + 237, + 782, + 297 + ], + "page_idx": 5 + }, + { + "type": "equation", + "text": "\n$$\n\\mathbf {x} ^ {g} \\left(t _ {0}\\right) \\sim \\mathcal {N} \\left(\\mathbf {x} ^ {g}, \\sigma^ {2} \\left(t _ {0}\\right) \\mathbf {I}\\right); \\quad \\mathbf {x} \\sim \\operatorname {S D E d i t} \\left(\\mathbf {x} ^ {g}, t _ {0}, \\boldsymbol {\\Phi}\\right), \\tag {4}\n$$\n", + "text_format": "latex", + "bbox": [ + 328, + 308, + 784, + 325 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "where $\\Phi$ denotes a generative model. In PanoFree, we use inpainting mask $\\mathbf{m}_i$ to paste the guidance image $\\mathbf{x}^g$ to the blank areas in the warped image $\\hat{\\mathbf{x}}_i$ and then use SDEdit in the inpainting process:", + "bbox": [ + 214, + 334, + 782, + 378 + ], + "page_idx": 5 + }, + { + "type": "equation", + "text": "\n$$\n\\hat {\\mathbf {x}} _ {i} ^ {g} = \\mathbf {m} _ {i} \\cdot \\mathbf {x} ^ {g} + (1 - \\mathbf {m} _ {i}) \\cdot \\hat {\\mathbf {x}} _ {i}; \\quad \\mathbf {x} _ {i} \\sim \\operatorname {S D E d i t} \\left(\\hat {\\mathbf {x}} _ {i} ^ {g}, t _ {0}, \\boldsymbol {\\Phi} _ {i n p}\\right). \\tag {5}\n$$\n", + "text_format": "latex", + "bbox": [ + 299, + 390, + 784, + 407 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Since we only want to use additional guidance images to rectify the biased conditions rather than replicate the guidance image, we use $t_0 \\in [0.9, 1.0)$ in practice. Meanwhile, we found that different selection of generation path and guidance image results in different generation qualities, and the optimal choice may vary for different tasks. We introduce a general selection effective for various tasks in Section 3.4 and provide an example of extending this technique to make scenes more realistic in specific scenarios in Section 3.5.", + "bbox": [ + 212, + 416, + 782, + 521 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "3.3 Risky Area Estimation and Erasing", + "text_level": 1, + "bbox": [ + 214, + 544, + 553, + 559 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "To rectify the noisy conditions discussed in Sec.3.1 and eliminate the accumulation of artifacts, a natural idea is to detect and localize the artifact-inducing contents, and erase them. However, precise detection and localization often requires costly training. Thus, we turn to roughly estimate and erase the risky areas that are likely to contain artifact-inducing contents, based on indicators often associated with artifact: distances, color and smoothness. See Sec. A.2 in the supplementary for examples.", + "bbox": [ + 212, + 566, + 782, + 672 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Risk Estimation based on Distances. We consider the distance from the center point of the initial view $\\mathbf{x}_0$ and the distance to the edges. This is based on two priors: 1. The farther from the initial view, the more accumulated errors and the more likely to contain artifact-inducing contents. 2. Areas close to the edges are highly risky because truncated objects are mostly generated around the edges, and areas near the edges are often more severely distorted during warping. We use initial risk $\\mathbf{r}_{init}$ to represent the risk estimated based on the distance from the initial view, and edge risk $\\mathbf{r}_{edge}$ to represent the risk estimated based on the distance from edges. They are derived from the following:", + "bbox": [ + 212, + 672, + 784, + 808 + ], + "page_idx": 5 + }, + { + "type": "equation", + "text": "\n$$\n\\mathbf {r} _ {\\text {i n i t}} \\left(\\mathbf {c} _ {i}\\right) = \\mathcal {R} _ {p} \\left(\\mathcal {D} _ {0} \\left(\\mathbf {c} _ {i}\\right)\\right); \\quad \\mathbf {r} _ {\\text {e d g e}} \\left(\\mathbf {c} _ {i}\\right) = \\mathcal {R} _ {p} \\left(\\mathcal {D} _ {\\mathbf {e}} \\left(\\mathbf {c} _ {i}\\right)\\right), \\tag {6}\n$$\n", + "text_format": "latex", + "bbox": [ + 328, + 820, + 784, + 837 + ], + "page_idx": 5 + }, + { + "type": "page_number", + "text": "6", + "bbox": [ + 217, + 114, + 228, + 126 + ], + "page_idx": 5 + }, + { + "type": "header", + "text": "A. Liu et al.", + "bbox": [ + 271, + 114, + 354, + 126 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "where $\\mathbf{c}_i$ represents the pixel coordinates of $\\mathbf{x}_i$ within the panorama coordinate system, $\\mathcal{D}_0$ measures the distance to the center point of the initial view along the generation path, $\\mathcal{D}_{\\mathbf{e}}$ measures the distance to all edges $\\mathbf{e}$ , and $\\mathcal{R}_p$ is a scaling function. We use weighted euclidean distance for $\\mathcal{D}_0$ , Gaussian filters for $\\mathcal{D}_{\\mathbf{e}}$ , and min-max normalization for $\\mathcal{R}_p$ .", + "bbox": [ + 212, + 146, + 782, + 222 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Risk Estimation based on Color and Smoothness. After generating a view, we can predict the risk based on color and smoothness. This uses two priors: 1. Artifacts are often not smooth or distinct in color. 2. Salient areas with abrupt colors or unevenness are prone to causing artifacts. Color-based risk $\\mathbf{r}_{color}$ and smoothness-based risk $\\mathbf{r}_{smooth}$ are estimated in similar forms:", + "bbox": [ + 212, + 220, + 782, + 297 + ], + "page_idx": 6 + }, + { + "type": "equation", + "text": "\n$$\n\\mathbf {r} _ {\\text {c o l o r}} \\left(\\mathbf {x} _ {i}\\right) = \\mathcal {R} _ {f} \\left(\\mathcal {D} _ {c} \\left(\\mathbf {x} _ {i}\\right)\\right); \\quad \\mathbf {r} _ {\\text {s m o o t h}} \\left(\\mathbf {x} _ {i}\\right) = \\mathcal {R} _ {f} \\left(\\mathcal {D} _ {s} \\left(\\mathbf {x} _ {i}\\right)\\right), \\tag {7}\n$$\n", + "text_format": "latex", + "bbox": [ + 313, + 310, + 784, + 325 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "where $\\mathcal{D}_c$ and $\\mathcal{D}_s$ measures the abruptness of each pixel based on color and smoothness. When implementing them, we choose pixels with the same vertical coordinates across views, and calculate the \"distances\" of each pixel to the mean color and color gradient. Within $\\mathcal{R}_f$ , we applied Gaussian filtering after min-max normalization, as those estimated risks are usually noisy.", + "bbox": [ + 212, + 335, + 782, + 409 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Erasing with Estimated Risks. With the estimated risks, we can erase the risky areas on the image warped to next view and the inpainting mask. Assume that we get inpainting mask for current view $\\mathbf{m}_i$ and risks for previous view $\\mathbf{r}_{i - 1} = [\\mathbf{r}_{i - 1}^i,\\mathbf{r}_{i - 1}^e,\\mathbf{r}_{i - 1}^c,\\mathbf{r}_{i - 1}^s]$ . The risks are combined linearly and new inpainting mask for current view can be obtained with:", + "bbox": [ + 212, + 411, + 784, + 484 + ], + "page_idx": 6 + }, + { + "type": "equation", + "text": "\n$$\n\\mathbf {m} _ {i} ^ {r} = \\mathcal {M} _ {r} \\left(\\mathbf {m} _ {i}, \\mathcal {W} \\left(\\mathbf {r} _ {i - 1} \\cdot \\mathbf {w}, \\mathbf {P} _ {i - 1} ^ {i}\\right)\\right). \\tag {8}\n$$\n", + "text_format": "latex", + "bbox": [ + 380, + 498, + 784, + 513 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "$\\mathcal{M}_r$ is the risk-based remasking function, and $\\mathbf{w}$ are user defined combination weights. We define $\\mathcal{M}_r$ as thresholding the risk within the warped area.", + "bbox": [ + 212, + 525, + 782, + 555 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Smoothing and Anti-aliasing. We note that the inpainting mask from risk-based erasing may not be smooth. Additionally, sharp and jagged edges on the inpainting mask can lead to artifacts. Therefore, we also employ fixed filtering $\\mathcal{M}_f$ , where Gaussian filtering and thresholding are used to smooth the mask and reduce sharp edges, while median filtering is used to reduce jagged edges. Then, we use the final inpainting mask for the combination with guidance, and the risky areas on the warped image are removed and regenerated.", + "bbox": [ + 212, + 555, + 784, + 660 + ], + "page_idx": 6 + }, + { + "type": "equation", + "text": "\n$$\n\\mathbf {m} _ {i} ^ {f} = \\mathcal {M} _ {f} \\left(\\mathbf {m} _ {i} ^ {r}\\right); \\quad \\hat {\\mathbf {x}} _ {i} ^ {g} = \\mathbf {m} _ {i} ^ {f} \\cdot \\mathbf {x} ^ {g} + \\left(1 - \\mathbf {m} _ {i} ^ {f}\\right) \\cdot \\hat {\\mathbf {x}} _ {i}. \\tag {9}\n$$\n", + "text_format": "latex", + "bbox": [ + 333, + 671, + 784, + 688 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "3.4 Bidirectional Generation with Symmetric Guidance", + "text_level": 1, + "bbox": [ + 214, + 709, + 687, + 724 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Bidirectional Generation. We begin by dividing a unidirectional generation path $\\mathbf{x}_0 \\to \\mathbf{x}_1 \\dots \\mathbf{x}_{2n}$ into two bidirectional generation paths $\\mathbf{x}_0 \\to \\mathbf{x}_1 \\to \\dots \\to \\mathbf{x}_n$ and $\\mathbf{x}_{-n} \\gets \\dots \\gets \\mathbf{x}_{-1} \\gets \\mathbf{x}_0$ . Typically, we would make these two generation paths symmetric. And we found this can reduce accumulated errors because the distance to the initial view is reduced in each direction. This consistently reduces artifacts, but may not reduce style and content inconsistency, as there may be different style/content shift in the two directions.", + "bbox": [ + 212, + 733, + 782, + 839 + ], + "page_idx": 6 + }, + { + "type": "header", + "text": "PanoFree", + "bbox": [ + 666, + 114, + 730, + 126 + ], + "page_idx": 6 + }, + { + "type": "page_number", + "text": "7", + "bbox": [ + 774, + 116, + 785, + 126 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Loop Closure. To ensure loop closure, we can add a $(2n + 1)$ -th view as the \"merging view\" to merge the 2 generation paths by warping $\\mathbf{x}_n$ and $\\mathbf{x}_{-n}$ to the $(2n + 1)$ -th view and inpaint it. However, if the differences between the two paths are too large, $\\mathbf{x}_{2n + 1}$ may contain image tearing, failing to ensure loop closure. This is due to the partial conditions on each path: there is no information from the other path before merging. Therefore, we rectify the partial conditions by introducing awareness of the other path.", + "bbox": [ + 212, + 146, + 782, + 252 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Symmetric Guidance. We introduce the awareness of the other path by selecting symmetric guidance images from the other path. Specifically, when generating $\\mathbf{x}_{i + 1}$ , we will select $\\mathbf{x}_{-i}$ as the guidance image. Thus, $\\mathbf{x}_{i + 1}$ will get the awareness of both paths as it is conditioned on $\\mathbf{x}_i$ and $\\mathbf{x}_{-i}$ :", + "bbox": [ + 212, + 252, + 782, + 313 + ], + "page_idx": 7 + }, + { + "type": "equation", + "text": "\n$$\n\\mathbf {x} _ {i + 1} \\sim q (\\mathbf {x} \\mid c _ {t}, \\mathbf {x} _ {i}, \\mathbf {x} _ {- i}, \\mathbf {P} _ {i} ^ {i + 1}) \\tag {10}\n$$\n", + "text_format": "latex", + "bbox": [ + 393, + 320, + 784, + 338 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "We empirically found that bidirectional generation with symmetric guidance is not only effective in ensuring loop closure but also a universally applicable strategy to effectively reduce accumulated errors in various scenarios.", + "bbox": [ + 212, + 345, + 782, + 390 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "3.5 Aligning with Scene Structure Prior", + "text_level": 1, + "bbox": [ + 214, + 410, + 560, + 424 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "When generating full spherical panoramas, we divide a spherical panorama into five parts: first, we generate a 360 panorama as the central part, then we expand upwards and downwards, and finally, we generate two images centered around the top and bottom poles to close up the entire spherical surface. During the expansion and closing stages, models often fail to align with scene structure priors due to partial conditions and generate artifacts.", + "bbox": [ + 212, + 431, + 782, + 518 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Hallucination refers to the artifacts caused by mismatches between partial conditions and scene structure priors. For example, when generating a city scene, using the same prompt during the expansion and closing stages may result in a floating city in the sky or a city underwater. The most direct solution is to input a new prompt, but this would require additional manual effort, which is not ideal. So, we attempt to rectify the partial conditions by extracting scene structure priors and applying semantic and variance control from the initial view.", + "bbox": [ + 212, + 522, + 782, + 626 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Prior Extraction. Although the pretrained T2I model may not align a full panorama with scene structure priors, it can align a single perspective view image with them. Therefore, we extract the scene structure prior from the initial view image $\\mathbf{x}_0$ and incorporate it into the expansion process. For example, when generating the first view image in the upward expansion $\\mathbf{x}_0^{ue}$ , we use upper $1/3$ part of the initial view image as guidance with resizing it to the size of $\\mathbf{x}_0^{ue}$ .", + "bbox": [ + 212, + 628, + 782, + 717 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Semantic and Variance Tuning. When the give text prompt only describe part of the scene, we may want generated semantic contents less conditioned on the partial prompt and more conditioned on the prior images during expansion and closing. We achieve this by reducing guidance scale and widen the field of view. Meanwhile, we adjust the variance of the initial noise to avoid the color blocks caused by low guidance scale. Through experimentation, we've found that a combination of slightly high initial variance and low guidance scale can stably reduce hallucinations and color blocks during the expansion and closing stages.", + "bbox": [ + 212, + 719, + 782, + 837 + ], + "page_idx": 7 + }, + { + "type": "page_number", + "text": "8", + "bbox": [ + 217, + 114, + 228, + 126 + ], + "page_idx": 7 + }, + { + "type": "header", + "text": "A. Liu et al.", + "bbox": [ + 271, + 114, + 354, + 127 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "4 Experiments", + "text_level": 1, + "bbox": [ + 215, + 143, + 375, + 162 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "We evaluate the performance of PanoFree across three generation tasks: Planar Panorama Generation, 360 Panorama Generation, and Full Panorama Generation. However, note that we focus on planar panorama and $360^{\\circ}$ panorama generation, where the comparisons are more precise and consistent.", + "bbox": [ + 212, + 198, + 787, + 260 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "Implementation details. PanoFree is implemented using the publicly available Stable Diffusion code from Diffusers [34] based on the PyTorch framework. For the experiments in the main paper, we utilized the generation and inpainting models of Stable Diffusion (SD) v2.0 [38]. All experiments are conducted on a single NVIDIA RTX A6000 GPU. Further details and specific configurations can be found in the corresponding sections of the main paper and the supplementary.", + "bbox": [ + 212, + 263, + 787, + 354 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "Evaluation metrics. We introduce a more comprehensive set of evaluation metrics than prior work [1, 22, 46] covering five themes: image quality, global consistency, prompt capability, diversity, and resource consumption.", + "bbox": [ + 212, + 358, + 787, + 404 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "text", + "list_items": [ + "- Image Quality is measured with Fréchet Inception Distance (FID) [14], Kernel Inception Distance (KID) [2], which measure fidelity and diversity. FID and KID calculated between the views randomly cropped from the panorama and reference images generated by SD with the same prompts.", + "- Global Consistency is measured with Intra-LPIPS (IL) [55] used by SyncDiffusion [22], which is computed by cropping non-overlapping views from a panorama and computing the averaged LPIPS scores of all view pairs.", + "- Prompt Capability is measured via CLIP Score (CS) [13] by computing the text-image similarity of randomly cropped views of the panorama.", + "- Panorama Diversity is also measured by FID and KID. Additionally, we propose Cross-LPIPS (CS) [55]. Cross-LPIPS is computed across 2 panoramas generated with a same text with different random seeds. We crop non-overlapping views from each panorama, and compute the averaged LPIPS scores of all view pairs where two views come from different panoramas.", + "- Resource Consumption includes time consumption, measured by the cumulative time cost of all diffusion processes to generate a single panorama, and peak GPU memory consumption, measured by the maximum GPU memory consumption during inference." + ], + "bbox": [ + 214, + 407, + 782, + 676 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "Evaluation Settings. Prior work either used arbitrary prompts [1,22] or only focused on a single type of scene [46]. Instead, we consider 3 distinct scene types: indoor, street, and city scenes, and natural scenes. We obtained 100 prompts for each type from ChatGPT [35]. We use 10 random seeds per prompt for planar panorama and 360 panorama generation, and 3 different random seeds per prompt for full panorama generation (see supplementary for details).", + "bbox": [ + 212, + 684, + 787, + 776 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "User Study. For planar panorama generation and $360^{\\circ}$ panorama generation, we conducted four user studies for each task to further evaluate the global consistency, image quality, prompt compatibility, and diversity of the generated panoramas (see supplementary for details).", + "bbox": [ + 212, + 779, + 787, + 840 + ], + "page_idx": 8 + }, + { + "type": "header", + "text": "PanoFree", + "bbox": [ + 666, + 114, + 730, + 126 + ], + "page_idx": 8 + }, + { + "type": "page_number", + "text": "9", + "bbox": [ + 774, + 114, + 785, + 126 + ], + "page_idx": 8 + }, + { + "type": "image", + "img_path": "images/c85aecde535a1bec8d23bcc3582d7bba3a47c6d8ca6e7ac19655d03409efa4c4.jpg", + "image_caption": [ + "Fig. 3: Planar Panorama generation results. Compared to vanilla Sequential Generation and MultiDiffusion (MD) [1], PanoFree achieves superior global consistency and image quality. It is also comparable to SyncDiffusion (SYD) [22] in these aspects." + ], + "image_footnote": [], + "bbox": [ + 259, + 143, + 736, + 276 + ], + "page_idx": 9 + }, + { + "type": "table", + "img_path": "images/4ddebd2515bb9be9314c9e80cf1f229afa487a6274081e53d626cafe55229b4b.jpg", + "table_caption": [ + "Table 1: Comparison of tuning free methods for Planar Panorama generation using Stable Diffusion [38]. We find PanoFree (PF) outperforms the state-of-the-art while having low computational requirements. Note that Cross-LPIPS and Intra-LPIPS are in $10^{-2}$ scale, KID is in $10^{-3}$ scale." + ], + "table_footnote": [], + "table_body": "
MethodIntra-LPIPS↓Cross-LPIPS↑FID↓KID↓CS↑Time (s)↓Memory (GB)↓
SG70.4071.0324.914.3326.68253.2
MD [1]68.4869.9221.163.5027.89955.8
SYD [22]64.4868.0720.563.6227.1812810.0
PF (Ours)65.3469.6817.053.8027.21263.2
", + "bbox": [ + 258, + 402, + 733, + 479 + ], + "page_idx": 9 + }, + { + "type": "text", + "text": "4.1 Planar Panorama Generation", + "text_level": 1, + "bbox": [ + 214, + 503, + 503, + 517 + ], + "page_idx": 9 + }, + { + "type": "text", + "text": "Planar Panorama corresponds to the scene observed with camera translation along the focal plane in reality. This is a relatively simple task, as it only involves extending the image without considering more complex geometric changes.", + "bbox": [ + 212, + 535, + 785, + 580 + ], + "page_idx": 9 + }, + { + "type": "text", + "text": "Baselines. We have chosen 3 tuning-free baselines for comparison, Vanilla Sequential Generation (SG), MultiDiffusion (MD) [1] and SyncDiffusion (SYD) [22]. Additional details are in the supplementary.", + "bbox": [ + 212, + 580, + 789, + 626 + ], + "page_idx": 9 + }, + { + "type": "text", + "text": "Results. The quantitative and qualitative evaluations are shown in Table 1 and Fig. 3, respectively. Below we compare PanoFree to each baseline.", + "bbox": [ + 212, + 627, + 784, + 656 + ], + "page_idx": 9 + }, + { + "type": "list", + "sub_type": "text", + "list_items": [ + "- Compared with vanilla Sequential Generation, PanoFree significantly enhances image quality and global consistency, demonstrating its effectiveness in reducing accumulated errors. Moreover, PanoFree does not compromise diversity or have a significant effect on GPU time and memory overhead.", + "- Compared with MultiDiffusion, PanoFree has significant advantages in image quality and global consistency. Meanwhile, its time and GPU memory overhead is only $26\\%$ and $55\\%$ that of MultiDiffusion, respectively.", + "- Compared with SyncDiffusion, PanoFree achieves comparable performance in global consistency and image quality. Although SyncDiffusion performs better in consistency, it requires introducing additional models for latent optimization. This leads PanoFree's time overhead to be $20\\%$ of SyncDiffusion and GPU memory overhead to be $32\\%$ of SyncDiffusion." + ], + "bbox": [ + 214, + 657, + 784, + 839 + ], + "page_idx": 9 + }, + { + "type": "page_number", + "text": "10", + "bbox": [ + 217, + 114, + 235, + 126 + ], + "page_idx": 9 + }, + { + "type": "header", + "text": "A. Liu et al.", + "bbox": [ + 271, + 114, + 354, + 127 + ], + "page_idx": 9 + }, + { + "type": "image", + "img_path": "images/370e94a9c5a2cde88e8f1ad51d2edc54951b80405d196c6691e9e67acd0b50f2.jpg", + "image_caption": [ + "Fig. 4: Diversity comparison on Planar Panorama Generation task. Each group is generated using the same text with different random seeds. Compared to MultiDiffusion (MD) [1] and SyncDiffusion (SYD) [22], PanoFree achieves superior diversity." + ], + "image_footnote": [], + "bbox": [ + 233, + 145, + 764, + 252 + ], + "page_idx": 10 + }, + { + "type": "table", + "img_path": "images/659b7c10e98555d26fed19f8b302f29766035d789e5295b58257d1d4622ff609.jpg", + "table_caption": [ + "Table 2: User study results of Planar Panorama Generation. 15 questions are used for each evaluation item and answered by 5 Amazon MTurk workers." + ], + "table_footnote": [], + "table_body": "
Consistency (%)Quality (%)Prompt Compatibility (%)Diversity (%)
SYD [22]52.746.041.335.3
PF (ours)47.354.058.764.7
", + "bbox": [ + 225, + 351, + 777, + 409 + ], + "page_idx": 10 + }, + { + "type": "text", + "text": "The Loss of Diversity with Joint Diffusion. When using different random seeds with the same prompt, methods using Joint Diffusion exhibit reduced diversity in their results. In contrast, our PanoFree method can better maintain diversity (see cross-LPIPS scores in Table 1). Additionally, we believe this is the source of PanoFree's gains over MultiDiffusion and SyncDiffusion in FID.", + "bbox": [ + 212, + 430, + 782, + 505 + ], + "page_idx": 10 + }, + { + "type": "text", + "text": "This diversity issue becomes particularly apparent when given some underspecified prompts. Therefore, we generated a \"underspecified set\" consisting of 20 short and blurry prompts to demonstrate this issue. For each prompt, we used 20 different random seeds. We demonstrate the diversity differences qualitatively in Fig. 4. Please refer to supplementary for quantitative analysis.", + "bbox": [ + 212, + 506, + 784, + 580 + ], + "page_idx": 10 + }, + { + "type": "text", + "text": "User Study. The results in Table 2 clearly show that human evaluators believe PanoFree produces more diverse panoramas and demonstrates better compatibility with prompts than SyncDiffusion [22]. Additionally, both methods exhibit similar levels of global consistency and image quality.", + "bbox": [ + 212, + 580, + 784, + 642 + ], + "page_idx": 10 + }, + { + "type": "text", + "text": "4.2 360 Panorama Generation", + "text_level": 1, + "bbox": [ + 214, + 664, + 477, + 678 + ], + "page_idx": 10 + }, + { + "type": "text", + "text": "Due to the distortion caused by equirectangular projection, generating 360-degree panoramas is more challenging than planar panorama generation. Vanilla sequential generation tends to produce many artifacts, significantly decreasing image quality. Moreover, MultiDiffusion [1] and SyncDiffusion [22] cannot be directly used for generating 360-degree panoramas. As far as we know, PanoFree is the first implementation of training-free 360-degree panorama generation.", + "bbox": [ + 212, + 689, + 784, + 777 + ], + "page_idx": 10 + }, + { + "type": "text", + "text": "Baselines. We used 2 baselines for comparison: Vanilla Sequential Generation (SG) and MVDiffusion (MVD). Additional details are in the supplementary.", + "bbox": [ + 214, + 779, + 785, + 809 + ], + "page_idx": 10 + }, + { + "type": "text", + "text": "Results. The quantitative and qualitative evaluations are shown in Table 3 and Fig. 5 respectively. Below we compare PanoFree to each baseline.", + "bbox": [ + 214, + 809, + 785, + 839 + ], + "page_idx": 10 + }, + { + "type": "header", + "text": "PanoFree", + "bbox": [ + 666, + 114, + 730, + 126 + ], + "page_idx": 10 + }, + { + "type": "page_number", + "text": "11", + "bbox": [ + 767, + 114, + 782, + 126 + ], + "page_idx": 10 + }, + { + "type": "table", + "img_path": "images/130833398340003c7cdf6c28488dbe39fc20ec687b3c94599906f7575b10d0c2.jpg", + "table_caption": [ + "Table 3: Comparison of $360^{\\circ}$ Panorama generation methods using Stable Diffusion [38]. We find PanoFree (PF) still outperforms the state-of-the-art while having low computational requirements. Note that Cross-LPIPS and Intra-LPIPS are in $10^{-2}$ scale, KID is in $10^{-3}$ scale." + ], + "table_footnote": [], + "table_body": "
MethodIntra-LPIPS↓Cross-LPIPS↑FID↓KID↓CS↑Time (s)↓Memory (GB)↓
SG70.6273.0632.287.9026.35213.2
MVD [46]67.7170.0737.898.7626.271106.9
PF (ours)68.6272.6725.847.4826.51223.2
", + "bbox": [ + 256, + 212, + 736, + 280 + ], + "page_idx": 11 + }, + { + "type": "image", + "img_path": "images/3043c22643dd25c5abe0f5f7f8e86baec2d046fa46ea0de2d672715d85cfe54a.jpg", + "image_caption": [ + "Fig. 5: $360^{\\circ}$ Panorama generation results. Compared to vanilla Sequential Generation (SG), PanoFree achieves superior global consistency and image quality. It is also comparable to MVDiffusion (MVD) [46] in these aspects." + ], + "image_footnote": [], + "bbox": [ + 243, + 282, + 754, + 421 + ], + "page_idx": 11 + }, + { + "type": "text", + "text": "- Compared with vanilla Sequential Generation, PanoFree significantly enhances image quality and global consistency. Specifically, vanilla sequential generation creates artifacts with complex optical geometry transformations, severely impacting image quality. However, PanoFree effectively recovers image quality by estimating and erasing risky areas, minimizing artifact propagation.", + "bbox": [ + 212, + 479, + 787, + 555 + ], + "page_idx": 11 + }, + { + "type": "text", + "text": "- Compared with MVDiffusion, PanoFree achieves comparability in image quality and global consistency, yet significantly outperforms in terms of time, GPU memory overhead, and diversity. Particularly, MVDiffusion is significantly worse than PanoFree in terms of FID and KID scores, even underperforming vanilla Sequential Generation. This is partly due to the inevitable bias of MVDiffusion's generated results towards the training dataset, resulting in larger discrepancies compared to those produced by Stable Diffusion. Visually, MVDiffusion also exhibits a noticeable lack of generation diversity. As depicted in Fig. 6, given a prompt, results generated with different random seeds show minimal variation in both content and style.", + "bbox": [ + 212, + 555, + 787, + 705 + ], + "page_idx": 11 + }, + { + "type": "text", + "text": "User Study. The user study results in Table 4 show that human evaluators believe PanoFree also produces more diverse $360^{\\circ}$ panoramas compared with MVDiffusion [46]. And PanoFree demonstrates better global consistency. Both methods exhibit similar levels of image quality and prompt comparability.", + "bbox": [ + 212, + 705, + 787, + 767 + ], + "page_idx": 11 + }, + { + "type": "text", + "text": "4.3 Full Spherical Panorama Generation", + "text_level": 1, + "bbox": [ + 214, + 786, + 563, + 801 + ], + "page_idx": 11 + }, + { + "type": "text", + "text": "PanoFree is also the first to achieve feasible tuning-free generation for Full Spherical Panoramas. However, it's hard to conduct meaningful comparisons due the", + "bbox": [ + 212, + 809, + 785, + 839 + ], + "page_idx": 11 + }, + { + "type": "page_number", + "text": "12", + "bbox": [ + 217, + 114, + 235, + 126 + ], + "page_idx": 11 + }, + { + "type": "header", + "text": "A. Liu et al.", + "bbox": [ + 271, + 114, + 354, + 126 + ], + "page_idx": 11 + }, + { + "type": "image", + "img_path": "images/8acd5007191249bdfb9bf3f8c8c8d669c9b8816e04fe224fb779af2ded2f2dcc.jpg", + "image_caption": [ + "\"Charming alpine village nestled among snow-capped peaks\"", + "MVD", + "Fig. 6: Diversity comparison on $360^{\\circ}$ Panorama Generation task. Each group is generated using the same text with different random seeds. Compared to MVDiffusion (MVD) [46], PanoFree achieves superior diversity." + ], + "image_footnote": [], + "bbox": [ + 263, + 157, + 496, + 271 + ], + "page_idx": 12 + }, + { + "type": "image", + "img_path": "images/d9cba1a78fe86f2de9153b9768d86cedb5ba05f56c6435bf11ec496354479830.jpg", + "image_caption": [ + "PF" + ], + "image_footnote": [], + "bbox": [ + 501, + 157, + 733, + 271 + ], + "page_idx": 12 + }, + { + "type": "table", + "img_path": "images/389b4c352725aadf85380b42203f08eb4d61caab079889e4b1a8748370246455.jpg", + "table_caption": [ + "Table 4: User study results of $360^{\\circ}$ Panorama Generation. 15 questions are used for each evaluation item and answered by 5 Amazon MTurk workers." + ], + "table_footnote": [], + "table_body": "
Consistency (%)Quality (%)Prompt Compatibility (%)Diversity (%)
MVD [46]40.748.747.333.3
PF (ours)59.351.352.666.6
", + "bbox": [ + 225, + 386, + 777, + 446 + ], + "page_idx": 12 + }, + { + "type": "text", + "text": "lack of tuning-free generation baseline methods or those with strong out-of-scope generation capabilities for Full Spherical Panoramas generation task. Thus, we conduct qualitative evaluation as well as comparison with vanilla sequential generation by showcasing generated results in Fig. 7. Vanilla sequential generation exhibit more artifacts as distortion increases. Additionally, partial conditioning issue mentioned in Sec. 3.5 causes hallucinations. And PanoFree still could effectively reduce artifacts and hallucinations. Note that we start both methods from $360^{\\circ}$ panoramas generated by PanoFree into full spherical panoramas, otherwise vanilla sequential generation will perform even worse.", + "bbox": [ + 212, + 464, + 787, + 602 + ], + "page_idx": 12 + }, + { + "type": "text", + "text": "4.4 Ablation Study", + "text_level": 1, + "bbox": [ + 215, + 619, + 390, + 633 + ], + "page_idx": 12 + }, + { + "type": "text", + "text": "Tab. 5 contains an ablation study that sequentially integrates each PanoFree component. We evaluate consistency (Intra-LPIPS) and image quality (FID) with $30\\%$ of the prompts from Sec. 4.1 & Sec. 4.2 for both planar and $360^{\\circ}$ panorama. We show that cross-view guidance provides the strongest benefit, followed by distance and edge-based risky area erasing. These components effectively reduce image tearing and visual chaos. Color and smoothness-based erasing have a smaller impact, likely due noise in these low-level features. Qualitative results are in Sec. B of the supplementary.", + "bbox": [ + 212, + 640, + 787, + 763 + ], + "page_idx": 12 + }, + { + "type": "text", + "text": "5 Conclusion", + "text_level": 1, + "bbox": [ + 215, + 781, + 359, + 797 + ], + "page_idx": 12 + }, + { + "type": "text", + "text": "We present PanoFree, a tuning-free multi-view image generation that supports an extensive array of correspondences. PanoFree improves error accumulation", + "bbox": [ + 214, + 809, + 785, + 839 + ], + "page_idx": 12 + }, + { + "type": "header", + "text": "PanoFree", + "bbox": [ + 666, + 114, + 730, + 126 + ], + "page_idx": 12 + }, + { + "type": "page_number", + "text": "13", + "bbox": [ + 767, + 114, + 785, + 126 + ], + "page_idx": 12 + }, + { + "type": "image", + "img_path": "images/e7adb3f70906836ab705afcf8479db7d0491a05beebe9f5200a7e1256913485c.jpg", + "image_caption": [ + "Fig. 7: Full Spherical Panorama generation results. Vanilla Sequential Generation (SG) tends to generate hallucinations due to partial conditions, while PanoFree effectively mitigates this issue." + ], + "image_footnote": [], + "bbox": [ + 218, + 143, + 777, + 334 + ], + "page_idx": 13 + }, + { + "type": "table", + "img_path": "images/d4329e01ed82016b271e65a6f98120d9121ebd9ecb53d8f7ee231803709ed0bf.jpg", + "table_caption": [ + "Table 5: Quantitative ablation of PanoFree components using $30\\%$ of prompts from Sec. 4.1 & 4.2. Intra-LPIPS $(10^{-2})$ and FID show cross-view guidance offers the most benefit, followed by distance and edge-based risky area erasing. Color and smoothness-based erasing have minimal impact." + ], + "table_footnote": [], + "table_body": "
TaskMethodIntra-LPIPS↓FID↓
PlanarNone70.8424.75
+ CG66.6320.63
+ Dist65.8718.21
360°None71.3433.47
+ CG69.2127.38
+ Dist & Edge69.0326.69
+ Color & Smooth68.9426.51
", + "bbox": [ + 457, + 398, + 772, + 534 + ], + "page_idx": 13 + }, + { + "type": "text", + "text": "by enhancing cross-view awareness and refining the warping and inpainting processes through cross-view guidance, risky area estimation and erasing, and symmetric bidirectional guided generation for loop closure, alongside guidance-based semantic and density control for scene structure preservation. PanoFree is evaluated on various panorama types—Planar, $360^{\\circ}$ , and Full Spherical Panoramas. PanoFree demonstrates significant error reduction, improved global consistency, and image quality across different scenarios without extra fine-tuning. Compared to existing methods, PanoFree is up to 5x more efficient in time and 3x more efficient in GPU memory usage, and maintains superior diversity of results (2x better in our user study). Moreover, PanoFree can be extended to texture generation for 3D models. We intend to explore these possibilities in future research.", + "bbox": [ + 212, + 593, + 787, + 760 + ], + "page_idx": 13 + }, + { + "type": "text", + "text": "Limitations. A limitation of our work is that we are unable to generate scenes beyond the capability of the pre-trained T2I model. Therefore, we rely on large pre-trained T2I models to ensure the broad application scope. And when provided with text descriptions beyond the capability range of the pre-trained T2I models, the generated results may not match the text.", + "bbox": [ + 212, + 763, + 787, + 839 + ], + "page_idx": 13 + }, + { + "type": "page_number", + "text": "14", + "bbox": [ + 217, + 114, + 235, + 126 + ], + "page_idx": 13 + }, + { + "type": "header", + "text": "A. Liu et al.", + "bbox": [ + 271, + 114, + 354, + 126 + ], + "page_idx": 13 + }, + { + "type": "text", + "text": "References", + "text_level": 1, + "bbox": [ + 217, + 143, + 321, + 159 + ], + "page_idx": 14 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "1. Bar-Tal, O., Yariv, L., Lipman, Y., Dekel, T.: Multidiffusion: Fusing diffusion paths for controlled image generation (2023)", + "2. Binkowski, M., Sutherland, D.J., Arbel, M., Gretton, A.: Demystifying mmd gans. ArXiv abs/1801.01401 (2018), https://apisemantic scholar.org/CorpusID: 3531856", + "3. Brock, A., Donahue, J., Simonyan, K.: Large scale gan training for high fidelity natural image synthesis. arXiv preprint arXiv:1809.11096 (2018)", + "4. Chen, D.Z., Siddiqui, Y., Lee, H.Y., Tulyakov, S., Nießner, M.: Text2tex: Text-driven texture synthesis via diffusion models. arXiv preprint arXiv:2303.11396 (2023)", + "5. Chen, Z., Wang, G., Liu, Z.: Text2light: Zero-shot text-driven hdr panorama generation. ACM Transactions on Graphics (TOG) 41(6), 1-16 (2022)", + "6. Cheng, Y.C., Lin, C.H., Lee, H.Y., Ren, J., Tulyakov, S., Yang, M.H.: Inout: Diverse image outpainting via gan inversion. 2022 IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR) pp. 11421-11430 (2021), https://apisemantic scholar.org/CorpusID:232478397", + "7. Dhariwal, P., Nichol, A.: Diffusion models beat gans on image synthesis. Advances in Neural Information Processing Systems 34, 8780-8794 (2021)", + "8. Esser, P., Rombach, R., Ommer, B.: Taming transformers for high-resolution image synthesis. In: Proceedings of the IEEE/CVF conference on computer vision and pattern recognition. pp. 12873-12883 (2021)", + "9. Fang, C., Hu, X., Luo, K., Tan, P.: Ctrl-room: Controllable text-to-3d room meshes generation with layout constraints. arXiv preprint arXiv:2310.03602 (2023)", + "0. Feng, M., Liu, J., Cui, M., Xie, X.: Diffusion360: Seamless 360 degree panoramic image generation based on diffusion models. ArXiv abs/2311.13141 (2023), https://api-semanticscholar.org/CorpusID:265351889", + "1. Fridman, R., Abecasis, A., Kasten, Y., Dekel, T.: Scenescape: Text-driven consistent scene generation. arXiv preprint arXiv:2302.01133 (2023)", + "2. Goodfellow, I., Pouget-Abadie, J., Mirza, M., Xu, B., Warde-Farley, D., Ozair, S., Courville, A., Bengio, Y.: Generative adversarial networks. Communications of the ACM 63(11), 139-144 (2020)", + "3. Hessel, J., Holtzman, A., Forbes, M., Bras, R.L., Choi, Y.: Clipscore: A reference-free evaluation metric for image captioning. ArXiv abs/2104.08718 (2021), https://api-semanticscholar.org/CorpusID:233296711", + "4. Heusel, M., Ramsauer, H., Unterthiner, T., Nessler, B., Hochreiter, S.: Gans trained by a two time-scale update rule converge to a local nash equilibrium. In: Neural Information Processing Systems (2017), https://api-semanticscholar.org/CorpusID:326772", + "5. Ho, J., Jain, A., Abbeel, P.: Denoising diffusion probabilistic models. Advances in Neural Information Processing Systems (NeurIPS) (2020)", + "6. Ho, J., Sahara, C., Chan, W., Fleet, D.J., Norouzi, M., Salimans, T.: Cascaded diffusion models for high fidelity image generation. J. Mach. Learn. Res. 23(47), 1-33 (2022)", + "7. Ho, J., Salimans, T.: Classifier-free diffusion guidance. arXiv preprint arXiv:2207.12598 (2022)", + "8. Hollein, L., Cao, A., Owens, A., Johnson, J., Nießner, M.: Text2room: Extracting textured 3d meshes from 2d text-to-image models. arXiv preprint arXiv:2303.11989 (2023)" + ], + "bbox": [ + 225, + 176, + 784, + 839 + ], + "page_idx": 14 + }, + { + "type": "header", + "text": "PanoFree", + "bbox": [ + 666, + 114, + 730, + 126 + ], + "page_idx": 14 + }, + { + "type": "page_number", + "text": "15", + "bbox": [ + 767, + 116, + 784, + 126 + ], + "page_idx": 14 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "19. Karras, T., Aittala, M., Aila, T., Laine, S.: Elucidating the design space of diffusion-based generative models. arXiv preprint arXiv:2206.00364 (2022)", + "20. Karras, T., Aittala, M., Aila, T., Laine, S.: Elucidating the design space of diffusion-based generative models. arXiv preprint arXiv:2206.00364 (2022)", + "21. Karras, T., Aittala, M., Laine, S., Härkönen, E., Hellsten, J., Lehtinen, J., Aila, T.: Alias-free generative adversarial networks. Advances in Neural Information Processing Systems 34, 852-863 (2021)", + "22. Lee, Y., Kim, K., Kim, H., Sung, M.: Syncdiffusion: Coherent montage via synchronized joint diffusions. Advances in Neural Information Processing Systems 36 (2024)", + "23. Li, J., Bansal, M.: Panogen: Text-conditioned panoramic environment generation for vision-and-language navigation. ArXiv abs/2305.19195 (2023), https://api.sementicscholar.org/CorpusID:258967291", + "24. Lin, C.H., Chang, C.C., Chen, Y.S., Juan, D.C., Wei, W., Chen, H.T.: Cocogan: Generation by parts via conditional coordinating. 2019 IEEE/CVF International Conference on Computer Vision (ICCV) pp. 4511-4520 (2019), https://api-semanticscholar.org/CorpusID:90262507", + "25. Lin, C.H., Lee, H.Y., Cheng, Y.C., Tulyakov, S., Yang, M.H.: Infinitygan: Towards infinite-pixel image synthesis. In: International Conference on Learning Representations (2021), https://api-semanticscholar.org/CorpusID:238419701", + "26. Liu, X., Zhang, X., Ma, J., Peng, J., et al.: Instaflow: One step is enough for high-quality diffusion-based text-to-image generation. In: The International Conference on Learning Representations (2023)", + "27. Lu, C., Zhou, Y., Bao, F., Chen, J., Li, C., Zhu, J.: Dpm-solver: A fast ode solver for diffusion probabilistic model sampling in around 10 steps. arXiv preprint arXiv:2206.00927 (2022)", + "28. Lugmayr, A., Danelljan, M., Romero, A., Yu, F., Timofte, R., Gool, L.V.: Repaint: Inpainting using denoising diffusion probabilistic models. 2022 IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR) pp. 11451-11461 (2022), https://api-semanticscholar.org/CorpusID:246240274", + "29. Meng, C., He, Y., Song, Y., Song, J., Wu, J., Zhu, J.Y., Ermon, S.: Sdedit: Guided image synthesis and editing with stochastic differential equations. In: International Conference on Learning Representations (2021), https://apisemantic scholar.org/CorpusID:245704504", + "30. Mou, C., Wang, X., Xie, L., Zhang, J., Qi, Z., Shan, Y., Qie, X.: T2i-adapter: Learning adapters to dig out more controllable ability for text-to-image diffusion models. ArXiv abs/2302.08453 (2023), https://api_semanticscholar.org/CorpusID: 256900833", + "31. Nichol, A., Dhariwal, P., Ramesh, A., Shyam, P., Mishkin, P., McGrew, B., Sutskever, I., Chen, M.: Glide: Towards photorealistic image generation and editing with text-guided diffusion models. arXiv preprint arXiv:2112.10741 (2021)", + "32. Oh, C.H., Cho, W., Park, D., Chae, Y., Wang, L., Yoon, K.J.: Bips: Bi-modal indoor panorama synthesis via residual depth-aided adversarial learning. ArXiv abs/2112.06179 (2021), https://api-semanticscholar.org/CorpusID: 245123664", + "33. Peebles, W., Xie, S.: Scalable diffusion models with transformers. In: International Conference on Computer Vision. pp. 4195-4205 (2023)", + "34. von Platen, P., Patil, S., Lozhkov, A., Cuenca, P., Lambert, N., Rasul, K., Davaadorj, M., Wolf, T.: Diffusers: State-of-the-art diffusion models. https://github.com/huggingface/diffusers (2022)" + ], + "bbox": [ + 215, + 147, + 784, + 840 + ], + "page_idx": 15 + }, + { + "type": "page_number", + "text": "16", + "bbox": [ + 217, + 114, + 235, + 126 + ], + "page_idx": 15 + }, + { + "type": "header", + "text": "A. Liu et al.", + "bbox": [ + 271, + 114, + 354, + 127 + ], + "page_idx": 15 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "35. Radford, A., Narasimhan, K., Salimans, T., Sutskever, I.: Language models are unsupervised multitask learners (2019)", + "36. Ramesh, A., Dhariwal, P., Nichol, A., Chu, C., Chen, M.: Hierarchical text-conditional image generation with clip latents. arXiv preprint arXiv:2204.06125 (2022)", + "37. Razavi, A., Van den Oord, A., Vinyals, O.: Generating diverse high-fidelity images with vq-vae-2. Advances in neural information processing systems 32 (2019)", + "38. Rombach, R., Blattmann, A., Lorenz, D., Esser, P., Ommer, B.: High-resolution image synthesis with latent diffusion models. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 10684-10695 (2022)", + "39. Sahara, C., Chan, W., Saxena, S., Li, L., Whang, J., Denton, E.L., Ghasemipour, K., Gontijo Lopes, R., Karagol Ayan, B., Salimans, T., et al.: Photorealistic text-to-image diffusion models with deep language understanding. Advances in Neural Information Processing Systems (NeurIPS) (2022)", + "40. Sohl-Dickstein, J., Weiss, E., Maheswaranathan, N., Ganguli, S.: Deep unsupervised learning using nonequilibrium thermodynamics. In: International Conference on Machine Learning. PMLR (2015)", + "41. Song, J., Meng, C., Ermon, S.: Denoising diffusion implicit models. arXiv preprint arXiv:2010.02502 (2020)", + "42. Song, Y., Ermon, S.: Generative modeling by estimating gradients of the data distribution. Advances in neural information processing systems (NeurIPS) (2019)", + "43. Song, Y., Ermon, S.: Generative modeling by estimating gradients of the data distribution. Advances in neural information processing systems 32 (2019)", + "44. Song, Y., Ermon, S.: Improved techniques for training score-based generative models. Advances in neural information processing systems 33, 12438-12448 (2020)", + "45. Song, Y., Sohl-Dickstein, J.N., Kingma, D.P., Kumar, A., Ermon, S., Poole, B.: Score-based generative modeling through stochastic differential equations. ICLR (2021)", + "46. Tang, S., Zhang, F., Chen, J., Wang, P., Furukawa, Y.: Mvdiffusion: Enabling holistic multi-view image generation with correspondence-aware diffusion. ArXiv abs/2307.01097 (2023), https://api_semanticscholar.org/CorpusID: 259316427", + "47. Teterwak, P., Sarna, A., Krishnan, D., Maschinot, A., Belanger, D., Liu, C., Freeman, W.T.: Boundless: Generative adversarial networks for image extension. 2019 IEEE/CVF International Conference on Computer Vision (ICCV) pp. 10520-10529 (2019), https://api_semanticscholar.org/CorpusID:201106503", + "48. Van Den Oord, A., Vinyals, O., et al.: Neural discrete representation learning. Advances in neural information processing systems 30 (2017)", + "49. Voynov, A., Hertz, A., Arar, M., Fruchter, S., Cohen-Or, D.: Anylens: A generative diffusion model with any rendering lens (2023)", + "50. Wang, G., Yang, Y., Loy, C.C., Liu, Z.: Stylelight: Hdr panorama generation for lighting estimation and editing. In: European Conference on Computer Vision (2022), https://api_semanticscholar.org/CorpusID:251196614", + "51. Wang, H., Xiang, X., Fan, Y., Xue, J.H.: Customizing 360-degree panoramicas through text-to-image diffusion models. ArXiv abs/2310.18840 (2023), https://api.sementicscholar.org/CorpusID:264590753", + "52. Wu, T., Zheng, C., Cham, T.J.: Panodiffusion: 360-degree panorama outpainting via diffusion (2023), https://api-semanticscholar.org/CorpusID:259360663", + "53. Zhang, L., Rao, A., Agrawala, M.: Adding conditional control to text-to-image diffusion models (2023)" + ], + "bbox": [ + 212, + 146, + 787, + 839 + ], + "page_idx": 16 + }, + { + "type": "header", + "text": "PanoFree", + "bbox": [ + 666, + 114, + 730, + 126 + ], + "page_idx": 16 + }, + { + "type": "page_number", + "text": "17", + "bbox": [ + 767, + 114, + 785, + 126 + ], + "page_idx": 16 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "54. Zhang, Q., Song, J., Huang, X., Chen, Y., Liu, M.Y.: Diffcollage: Parallel generation of large content with diffusion models. 2023 IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR) pp. 10188-10198 (2023), https://api-semanticscholar.org/CorpusID:257834007", + "55. Zhang, R., Isola, P., Efros, A.A., Shechtman, E., Wang, O.: The unreasonable effectiveness of deep features as a perceptual metric. 2018 IEEE/CVF Conference on Computer Vision and Pattern Recognition pp. 586-595 (2018), https://apisemantic scholar.org/CorpusID:4766599" + ], + "bbox": [ + 215, + 146, + 785, + 258 + ], + "page_idx": 17 + }, + { + "type": "page_number", + "text": "18", + "bbox": [ + 217, + 114, + 235, + 126 + ], + "page_idx": 17 + }, + { + "type": "header", + "text": "A. Liu et al.", + "bbox": [ + 271, + 114, + 354, + 127 + ], + "page_idx": 17 + } +] \ No newline at end of file diff --git a/2024/PanoFree_ Tuning-Free Holistic Multi-view Image Generation with Cross-view Self-Guidance/cbd590c4-f320-4c08-83b4-3c8681128a51_model.json b/2024/PanoFree_ Tuning-Free Holistic Multi-view Image Generation with Cross-view Self-Guidance/cbd590c4-f320-4c08-83b4-3c8681128a51_model.json new file mode 100644 index 0000000000000000000000000000000000000000..2f382867932fcd224639bab5e99162876988d252 --- /dev/null +++ b/2024/PanoFree_ Tuning-Free Holistic Multi-view Image Generation with Cross-view Self-Guidance/cbd590c4-f320-4c08-83b4-3c8681128a51_model.json @@ -0,0 +1,2579 @@ +[ + [ + { + "type": "title", + "bbox": [ + 0.22, + 0.141, + 0.783, + 0.186 + ], + "angle": 0, + "content": "PanoFree: Tuning-Free Holistic Multi-view Image Generation with Cross-view Self-Guidance" + }, + { + "type": "text", + "bbox": [ + 0.22, + 0.212, + 0.783, + 0.244 + ], + "angle": 0, + "content": "Aoming Liu\\(^{1,2*}\\), Zhong Li\\(^{1\\dagger}\\), Zhang Chen\\(^{1\\dagger}\\), Nannan Li\\(^{2}\\), Yi Xu\\(^{1}\\), and Bryan A. Plummer\\(^{2}\\)" + }, + { + "type": "text", + "bbox": [ + 0.307, + 0.254, + 0.696, + 0.283 + ], + "angle": 0, + "content": "\\(^{1}\\) OPPO US Research Center, Palo Alto, CA 94303, USA \n\\(^{2}\\) Boston University, Boston, MA 02215, USA" + }, + { + "type": "text", + "bbox": [ + 0.262, + 0.321, + 0.744, + 0.626 + ], + "angle": 0, + "content": "Abstract. Immersive scene generation, notably panorama creation, benefits significantly from the adaptation of large pre-trained text-to-image (T2I) models for multi-view image generation. Due to the high cost of acquiring multi-view images, tuning-free generation is preferred. However, existing methods are either limited to simple correspondences or require extensive fine-tuning to capture complex ones. We present PanoFree, a novel method for tuning-free multi-view image generation that supports an extensive array of correspondences. PanoFree sequentially generates multi-view images using iterative warping and inpainting, addressing the key issues of inconsistency and artifacts from error accumulation without the need for fine-tuning. It improves error accumulation by enhancing cross-view awareness and refines the warping and inpainting processes via cross-view guidance, risky area estimation and erasing, and symmetric bidirectional guided generation for loop closure, alongside guidance-based semantic and density control for scene structure preservation. In experiments on Planar, \\(360^{\\circ}\\), and Full Spherical Panoramas, PanoFree demonstrates significant error reduction, improves global consistency, and boosts image quality without extra fine-tuning. Compared to existing methods, PanoFree is up to 5x more efficient in time and 3x more efficient in GPU memory usage, and maintains superior diversity of results (2x better in our user study). PanoFree offers a viable alternative to costly fine-tuning or the use of additional pre-trained models." + }, + { + "type": "text", + "bbox": [ + 0.262, + 0.639, + 0.706, + 0.654 + ], + "angle": 0, + "content": "Keywords: Tuning-free generation; Multi-view Image, Panorama" + }, + { + "type": "title", + "bbox": [ + 0.216, + 0.68, + 0.377, + 0.696 + ], + "angle": 0, + "content": "1 Introduction" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.711, + 0.788, + 0.803 + ], + "angle": 0, + "content": "Text-to-image (T2I) generation over multiple views for immersive scenes, like panorama generation, is a challenging task requiring coherence and diversity among many generated images (e.g., [3, 8, 12, 16, 17, 21, 36, 37, 39, 48]). Early work using GANs or VAEs (e.g., [4-6, 11, 24, 25, 47, 50]) have been replaced recently with diffusion-based models (e.g., [1, 9, 10, 22, 23, 46, 49, 51, 52, 54]), often leveraging Stable Diffusion [38]. State-of-the-art panorama generation methods" + }, + { + "type": "page_footnote", + "bbox": [ + 0.238, + 0.811, + 0.786, + 0.825 + ], + "angle": 0, + "content": "*Work partly completed during Aoming's internship at OPPO US Research Center." + }, + { + "type": "page_footnote", + "bbox": [ + 0.241, + 0.825, + 0.407, + 0.84 + ], + "angle": 0, + "content": "†Corresponding Authors." + } + ], + [ + { + "type": "page_number", + "bbox": [ + 0.218, + 0.116, + 0.23, + 0.127 + ], + "angle": 0, + "content": "2" + }, + { + "type": "header", + "bbox": [ + 0.272, + 0.115, + 0.356, + 0.127 + ], + "angle": 0, + "content": "A. Liu et al." + }, + { + "type": "image", + "bbox": [ + 0.276, + 0.145, + 0.725, + 0.38 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.214, + 0.385, + 0.788, + 0.442 + ], + "angle": 0, + "content": "Fig. 1: PanoFree can generate multi-view images according to different types of correspondences without fine-tuning, and a natural application is tuning-free generation for different types of panoramas. We demonstrate this by generating three commonly used panoramas. Top: Planar Panorama; Middle: 360 Panorama; Bottom: Full Panorama." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.455, + 0.788, + 0.532 + ], + "angle": 0, + "content": "use Joint Diffusion (e.g., [1,22,46]), where parallel diffusion processes to generate multi-view images and enhancing global consistency by fusing latent or attention features based on cross-view correspondences. However, we find these methods can only generate some types of panoramas, limiting their scope (e.g., [1,22]), or require fine-tuning using expensive panorama datasets (e.g., [22,46])." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.535, + 0.788, + 0.746 + ], + "angle": 0, + "content": "To address these challenges, we propose PanoFree a tuning-free multi-view image generation method using iterative warping and inpainting of perspective images to support diverse correspondences with low costs (see Fig. 1 for example generations). Iterative warping and inpainting of perspective images provide a means to the diverse multi-view correspondences required in panorama generation without additional fine-tuning [4, 11, 18]. However, recent work has overlooked these benefits due to accumulated errors causing suboptimal image quality [1, 22, 46]. We find that most accumulated errors from iterative warping can be attributed to the deficient conditions during generation. Specifically, conditioning solely on the previous image narrows cross-view awareness, leading to inconsistencies. Warping and inpainting can also propagate noise, e.g., truncated objects or jagged edges. Additionally, the given conditions may be incomplete to meet specific requirements, such as ensuring 360-degree consistency for loop closure and maintaining correct spatial relationships for realistic scenes." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.75, + 0.788, + 0.841 + ], + "angle": 0, + "content": "To address these issues, PanoFree expand cross-view awareness by conditioning the current view on multiple views with guided image synthesis techniques such as SDEdit [29]. Then, PanoFree estimates and erases the risky areas, regions likely containing noise, to reduce the noise introduced by warping and inpainting. In addition, PanoFree adopts a bidirectional generation path with a symmetrical conditioning strategy for loop closure. Lastly, PanoFree further" + } + ], + [ + { + "type": "header", + "bbox": [ + 0.668, + 0.115, + 0.732, + 0.127 + ], + "angle": 0, + "content": "PanoFree" + }, + { + "type": "page_number", + "bbox": [ + 0.775, + 0.117, + 0.785, + 0.127 + ], + "angle": 0, + "content": "3" + }, + { + "type": "text", + "bbox": [ + 0.218, + 0.147, + 0.785, + 0.176 + ], + "angle": 0, + "content": "utilizes pseudo global guidance with region-specific semantic and density control to make scene structure more reasonable." + }, + { + "type": "text", + "bbox": [ + 0.218, + 0.178, + 0.786, + 0.327 + ], + "angle": 0, + "content": "We evaluate PanoFree on three text-to-panorama generation tasks: Planar, \\(360^{\\circ}\\), and Full Spherical Panoramas. PanoFree effectively alleviating accumulated errors in sequential generation, and significantly improves image quality and global consistency (e.g. \\(31.6\\%\\) better in FID). This enables PanoFree to have better (or at least comparable) results to the state-of-the-art [1,22,46], despite these methods either having narrower applications or requiring fine-tuning datasets. Specifically, PanoFree is up to 5x more efficient in time and 3x more efficient in GPU memory usage, and maintains superior diversity of results (2x better in our user study). Lastly, PanoFree is also highly flexible, enabling it to plug-in-and-play with various pre-trained T2I models and adapters." + }, + { + "type": "text", + "bbox": [ + 0.241, + 0.33, + 0.593, + 0.342 + ], + "angle": 0, + "content": "Our contributions can be summarized as follows:" + }, + { + "type": "text", + "bbox": [ + 0.217, + 0.345, + 0.794, + 0.404 + ], + "angle": 0, + "content": "- We introduce PanoFree, a tuning-free multi-view image generation method applicable for various correspondences and pre-trained T2I models. Thus PanoFree can greatly reduce the data and fine-tuning costs for immersive scene generation tasks such as text-to-panorama generation." + }, + { + "type": "text", + "bbox": [ + 0.217, + 0.405, + 0.785, + 0.463 + ], + "angle": 0, + "content": "- We provide a in-depth perspective of accumulated errors and identify the deficient conditions as the main causes. We further effectively rectify deficient conditions and alleviate accumulated errors with the cross-view guidance as well as risky area estimation and erasing in PanoFree." + }, + { + "type": "text", + "bbox": [ + 0.217, + 0.465, + 0.785, + 0.495 + ], + "angle": 0, + "content": "- As far as we know, PanoFree is the first to achieve feasible tuning-free generation for \\(360^{\\circ}\\) Panoramas and Full Spherical Panoramas." + }, + { + "type": "list", + "bbox": [ + 0.217, + 0.345, + 0.794, + 0.495 + ], + "angle": 0, + "content": null + }, + { + "type": "title", + "bbox": [ + 0.218, + 0.52, + 0.387, + 0.535 + ], + "angle": 0, + "content": "2 Related Work" + }, + { + "type": "text", + "bbox": [ + 0.218, + 0.553, + 0.786, + 0.629 + ], + "angle": 0, + "content": "Diffusion Models [15, 19, 40, 42-45] are a popular framework for generative models. Early work required a long trajectory for sampling to produce high-quality samples [7, 45], before being sped up with advanced sampling techniques that also preserved generation quality [20, 26, 27, 41]. Latent Diffusion Models (LDMs) [33, 38] made these models more efficient by training in the latent space." + }, + { + "type": "text", + "bbox": [ + 0.218, + 0.63, + 0.786, + 0.779 + ], + "angle": 0, + "content": "T2I Diffusion and Panorama Generation. Diffusion models are widely adopted for text-to-image (T2I) generation [31, 36, 38, 39]. Many downstream tasks used large pre-trained T2I diffusion models, like Stable Diffusion [38], to boost performance [4, 11, 18], including panorama generation [4, 11, 18]. These methods have largely supplanted GAN and VAE methods [4-6, 11, 24, 25, 32, 47, 50], with most recent work in panorama generation tasks using diffusion models [1, 9, 10, 22, 23, 46, 49, 51, 52, 54]. These diffusion-based panorama generation methods use joint diffusion to fuses multiple diffusion processes through latent or attention manipulation [1, 22, 46]. However, they are either limited to modeling simple correspondences or require extensive fine-tuning to model complex ones." + }, + { + "type": "text", + "bbox": [ + 0.218, + 0.78, + 0.786, + 0.84 + ], + "angle": 0, + "content": "Guided Image Synthesis with Diffusion Models. It can be challenging to achieve satisfactory results solely relying on text guidance. Therefore, some prior work [28-30, 53] guide or control the generation results with reference images as fine-grained condition. ControlNet [53] and T2I-Adapter [30] are the most" + } + ], + [ + { + "type": "page_number", + "bbox": [ + 0.218, + 0.116, + 0.23, + 0.127 + ], + "angle": 0, + "content": "4" + }, + { + "type": "header", + "bbox": [ + 0.272, + 0.115, + 0.355, + 0.128 + ], + "angle": 0, + "content": "A. Liu et al." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.147, + 0.788, + 0.209 + ], + "angle": 0, + "content": "commonly used methods to incorporate additional image conditions by adding extra image encoders, but they all require few-shot fine-tuning. SDEdit [29] achieves tuning-free guided image synthesis by adding noise to the guide image and then denoising it back to a real image using a pre-trained diffusion model." + }, + { + "type": "title", + "bbox": [ + 0.216, + 0.229, + 0.331, + 0.245 + ], + "angle": 0, + "content": "3 Method" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.259, + 0.789, + 0.442 + ], + "angle": 0, + "content": "PanoFree targets the text-to-panorama generation task, which takes textual descriptions as guidance to create multi-view perspective images that can be stitched into a wide-angle, high-quality panorama. PanoFree generates multiview images through sequential warping and inpainting steps, which typically results in acclimated errors due to deficit conditions (discussed in Sec. 3.1). Each component of Panofree is designed to minimize the effect of these various deficiencies. Specifically, Sec. 3.2 mitigates inconsistency using SDEdit-based cross-view guidance and Sec. 3.3 removes artifact-inducing content by estimating and erasing risky areas. At a higher level, PanoFree employs Bidirectional Generation with Symmetric Guidance for loop closure and error reduction (Sec. 3.4). Additionally, it applies guidance-based semantic and density control for scene structure preservation (Sec. 3.5). See Fig. 2 for an overview of our approach." + }, + { + "type": "title", + "bbox": [ + 0.216, + 0.46, + 0.666, + 0.475 + ], + "angle": 0, + "content": "3.1 Deficient Conditions behind Accumulated Errors" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.483, + 0.789, + 0.575 + ], + "angle": 0, + "content": "In this section, we discuss the causes behind the deficit conditions in the iterative warping and inpainting process and reformulate the problem as conditional generation. Given the text prompt \\( c_{t} \\), \\( i \\)-th view's image \\( \\mathbf{x}_i \\), warping function \\( \\mathcal{W} \\), transformation matrix of the projection from \\( i \\)-th view to the \\( (i + 1) \\)-th view \\( \\mathbf{P}_i^{i + 1} \\), and pre-trained T2I inpainting model \\( \\Phi_{inp} \\), the warping and inpainting step to generate the \\( (i + 1) \\)-th view can be denoted as:" + }, + { + "type": "equation", + "bbox": [ + 0.327, + 0.584, + 0.787, + 0.601 + ], + "angle": 0, + "content": "\\[\n\\hat {\\mathbf {x}} _ {i}, \\mathbf {m} _ {i} = \\mathcal {W} \\left(\\mathbf {x} _ {i}, \\mathbf {P} _ {i} ^ {i + 1}\\right); \\quad \\mathbf {x} _ {i + 1} = \\boldsymbol {\\Phi} _ {i n p} \\left(\\hat {\\mathbf {x}} _ {i}, \\mathbf {m} _ {i}, c _ {t}\\right), \\tag {1}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.609, + 0.788, + 0.656 + ], + "angle": 0, + "content": "where \\(\\hat{\\mathbf{x}}_i\\) is the image warped from \\(i\\)-th view to \\((i + 1)\\)-th view and \\(\\mathbf{m}_i\\) is the masking indicating the area to inpaint. And we can simplify the warping and inpainting steps in the following conditional image generation form:" + }, + { + "type": "equation", + "bbox": [ + 0.408, + 0.664, + 0.787, + 0.681 + ], + "angle": 0, + "content": "\\[\n\\mathbf {x} _ {i + 1} \\sim q (\\mathbf {x} | c _ {t}, \\mathbf {x} _ {i}, \\mathbf {P} _ {i} ^ {i + 1}). \\tag {2}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.69, + 0.788, + 0.749 + ], + "angle": 0, + "content": "However, during this generation process, we found conditions can become deficient. Major accumulated errors arise from three types of deficient conditions: Biased Conditions, Noisy Conditions, and Partial Conditions. See Sec. A.2 in the supplementary for detailed error illustrations." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.75, + 0.789, + 0.841 + ], + "angle": 0, + "content": "Biased Conditions is the most obvious problem. In the above step, \\(\\mathbf{x}_{i + 1}\\) is solely conditioned on \\(\\mathbf{x}_i\\), which biases the cross-view awareness heavily to \\(i\\)-th view. If \\(\\mathbf{x}_i\\) has deviated from the desired global distribution in certain aspects, then \\(\\mathbf{x}_{i + 1}\\) is likely to continue deviating in the same direction, resulting in significant inconsistency. We also found that slight style and content shifts often accumulate in this way, leading to significant inconsistency between distant views." + } + ], + [ + { + "type": "header", + "bbox": [ + 0.667, + 0.115, + 0.732, + 0.127 + ], + "angle": 0, + "content": "PanoFree" + }, + { + "type": "page_number", + "bbox": [ + 0.775, + 0.117, + 0.785, + 0.127 + ], + "angle": 0, + "content": "5" + }, + { + "type": "image", + "bbox": [ + 0.269, + 0.145, + 0.725, + 0.26 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.336, + 0.267, + 0.662, + 0.279 + ], + "angle": 0, + "content": "(a) Overview of the generation and guidance framework." + }, + { + "type": "image", + "bbox": [ + 0.269, + 0.285, + 0.725, + 0.436 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.214, + 0.446, + 0.788, + 0.543 + ], + "angle": 0, + "content": "Fig. 2: Overview of our PanoFree method, taking 360 Panorama Generation as an example. (a): At a framework level, PanoFree adopts two generation paths with opposite viewpoint translation or rotation. It enhances consistency by symmetrically selecting views from the other path as guidance to generate a new view (Sec. 3.4). Loop closure is ensured by merging these two paths. (b): In each warping and inpainting step, PanoFree reduces accumulated error by guiding the inpainting process with cross-view images (Sec. 3.2), along with estimating and erasing risky areas (Sec. 3.3)." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.56, + 0.788, + 0.635 + ], + "angle": 0, + "content": "Noisy Conditions mainly refer to \\(\\mathbf{x}_i\\) containing artifact-inducing contents. Existing artifacts in \\(\\mathbf{x}_i\\) could guide inpainting model to generate similar artifacts in \\(\\mathbf{x}_{i + 1}\\) and propagate to every following view. Additionally, disjointed or distorted areas, jagged or sharp content, and objects truncated by edges in \\(\\mathbf{x}_i\\) are also highly risky to introduce artifacts in \\(\\mathbf{x}_{i + 1}\\)." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.636, + 0.788, + 0.727 + ], + "angle": 0, + "content": "Partial Conditions refer to the conditions not containing all the necessary information to meet specific requirements. For example, if we follow Eq. (2) on the final view, we lack information about \\(\\mathbf{x}_0\\), making it impossible to generate image coherent with \\(\\mathbf{x}_0\\) to ensure loop closure. Additionally, using a single text prompt to generate all views within a full spherical panorama may lead to hallucinations, such as cities floating in the sky or underwater mountains." + }, + { + "type": "title", + "bbox": [ + 0.215, + 0.751, + 0.442, + 0.765 + ], + "angle": 0, + "content": "3.2 Cross-View Guidance" + }, + { + "type": "text", + "bbox": [ + 0.215, + 0.777, + 0.787, + 0.809 + ], + "angle": 0, + "content": "To rectify the biased conditions discussed in Sec.3.1 and enlarge cross-view awareness, a natural idea is to let \\(\\mathbf{x}_{i + 1}\\) conditioned on more views," + }, + { + "type": "equation", + "bbox": [ + 0.368, + 0.819, + 0.786, + 0.838 + ], + "angle": 0, + "content": "\\[\n\\mathbf {x} _ {i + 1} \\sim q (\\mathbf {x} | c _ {t}, \\mathbf {x} _ {i}, \\mathbf {P} _ {i} ^ {i + 1}, \\mathbf {x} _ {1} ^ {g}, \\dots , \\mathbf {x} _ {m} ^ {g}) \\qquad (3)\n\\]" + } + ], + [ + { + "type": "page_number", + "bbox": [ + 0.218, + 0.116, + 0.23, + 0.127 + ], + "angle": 0, + "content": "6" + }, + { + "type": "header", + "bbox": [ + 0.272, + 0.115, + 0.355, + 0.127 + ], + "angle": 0, + "content": "A. Liu et al." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.147, + 0.784, + 0.237 + ], + "angle": 0, + "content": "where \\(\\mathbf{x}_1^g, \\ldots, \\mathbf{x}_m^g\\) are selected from \\(\\mathbf{x}_0, \\ldots, \\mathbf{x}_{i-1}\\). This naturally results in a Guided Image Synthesis task form with self-generated images as guidance. Many existing methods can be adapted to implement our design, such as ControlNet [53] and T2I adapter [30]. To avoid relying on fine-tuning and reduce memory and time costs, we choose SDEdit [29], a training-free guided image synthesis approach, with a single guidance image \\(\\mathbf{x}^g \\in [\\mathbf{x}_0, \\ldots, \\mathbf{x}_{i-1}]\\)." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.238, + 0.784, + 0.298 + ], + "angle": 0, + "content": "Guided Image Synthesis using SDEdit. Given \\(\\mathbf{x}^g\\) as guidance, SDEdit establishes a Gaussian distribution using \\(\\mathbf{x}^g\\) as the expectation and the intermediate status at time \\(t_0\\) in the reverse SDE process. The desired data distribution is obtained by gradually removing noise from \\(\\mathbf{x}^g(t_0)\\):" + }, + { + "type": "equation", + "bbox": [ + 0.329, + 0.309, + 0.785, + 0.326 + ], + "angle": 0, + "content": "\\[\n\\mathbf {x} ^ {g} \\left(t _ {0}\\right) \\sim \\mathcal {N} \\left(\\mathbf {x} ^ {g}, \\sigma^ {2} \\left(t _ {0}\\right) \\mathbf {I}\\right); \\quad \\mathbf {x} \\sim \\operatorname {S D E d i t} \\left(\\mathbf {x} ^ {g}, t _ {0}, \\boldsymbol {\\Phi}\\right), \\tag {4}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.215, + 0.335, + 0.784, + 0.38 + ], + "angle": 0, + "content": "where \\(\\Phi\\) denotes a generative model. In PanoFree, we use inpainting mask \\(\\mathbf{m}_i\\) to paste the guidance image \\(\\mathbf{x}^g\\) to the blank areas in the warped image \\(\\hat{\\mathbf{x}}_i\\) and then use SDEdit in the inpainting process:" + }, + { + "type": "equation", + "bbox": [ + 0.3, + 0.391, + 0.785, + 0.408 + ], + "angle": 0, + "content": "\\[\n\\hat {\\mathbf {x}} _ {i} ^ {g} = \\mathbf {m} _ {i} \\cdot \\mathbf {x} ^ {g} + (1 - \\mathbf {m} _ {i}) \\cdot \\hat {\\mathbf {x}} _ {i}; \\quad \\mathbf {x} _ {i} \\sim \\operatorname {S D E d i t} \\left(\\hat {\\mathbf {x}} _ {i} ^ {g}, t _ {0}, \\boldsymbol {\\Phi} _ {i n p}\\right). \\tag {5}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.417, + 0.784, + 0.522 + ], + "angle": 0, + "content": "Since we only want to use additional guidance images to rectify the biased conditions rather than replicate the guidance image, we use \\( t_0 \\in [0.9, 1.0) \\) in practice. Meanwhile, we found that different selection of generation path and guidance image results in different generation qualities, and the optimal choice may vary for different tasks. We introduce a general selection effective for various tasks in Section 3.4 and provide an example of extending this technique to make scenes more realistic in specific scenarios in Section 3.5." + }, + { + "type": "title", + "bbox": [ + 0.215, + 0.545, + 0.555, + 0.56 + ], + "angle": 0, + "content": "3.3 Risky Area Estimation and Erasing" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.568, + 0.784, + 0.673 + ], + "angle": 0, + "content": "To rectify the noisy conditions discussed in Sec.3.1 and eliminate the accumulation of artifacts, a natural idea is to detect and localize the artifact-inducing contents, and erase them. However, precise detection and localization often requires costly training. Thus, we turn to roughly estimate and erase the risky areas that are likely to contain artifact-inducing contents, based on indicators often associated with artifact: distances, color and smoothness. See Sec. A.2 in the supplementary for examples." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.674, + 0.785, + 0.809 + ], + "angle": 0, + "content": "Risk Estimation based on Distances. We consider the distance from the center point of the initial view \\(\\mathbf{x}_0\\) and the distance to the edges. This is based on two priors: 1. The farther from the initial view, the more accumulated errors and the more likely to contain artifact-inducing contents. 2. Areas close to the edges are highly risky because truncated objects are mostly generated around the edges, and areas near the edges are often more severely distorted during warping. We use initial risk \\(\\mathbf{r}_{init}\\) to represent the risk estimated based on the distance from the initial view, and edge risk \\(\\mathbf{r}_{edge}\\) to represent the risk estimated based on the distance from edges. They are derived from the following:" + }, + { + "type": "equation", + "bbox": [ + 0.329, + 0.821, + 0.785, + 0.838 + ], + "angle": 0, + "content": "\\[\n\\mathbf {r} _ {\\text {i n i t}} \\left(\\mathbf {c} _ {i}\\right) = \\mathcal {R} _ {p} \\left(\\mathcal {D} _ {0} \\left(\\mathbf {c} _ {i}\\right)\\right); \\quad \\mathbf {r} _ {\\text {e d g e}} \\left(\\mathbf {c} _ {i}\\right) = \\mathcal {R} _ {p} \\left(\\mathcal {D} _ {\\mathbf {e}} \\left(\\mathbf {c} _ {i}\\right)\\right), \\tag {6}\n\\]" + } + ], + [ + { + "type": "header", + "bbox": [ + 0.667, + 0.115, + 0.732, + 0.127 + ], + "angle": 0, + "content": "PanoFree" + }, + { + "type": "page_number", + "bbox": [ + 0.776, + 0.117, + 0.786, + 0.127 + ], + "angle": 0, + "content": "7" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.147, + 0.784, + 0.223 + ], + "angle": 0, + "content": "where \\(\\mathbf{c}_i\\) represents the pixel coordinates of \\(\\mathbf{x}_i\\) within the panorama coordinate system, \\(\\mathcal{D}_0\\) measures the distance to the center point of the initial view along the generation path, \\(\\mathcal{D}_{\\mathbf{e}}\\) measures the distance to all edges \\(\\mathbf{e}\\), and \\(\\mathcal{R}_p\\) is a scaling function. We use weighted euclidean distance for \\(\\mathcal{D}_0\\), Gaussian filters for \\(\\mathcal{D}_{\\mathbf{e}}\\), and min-max normalization for \\(\\mathcal{R}_p\\)." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.222, + 0.784, + 0.299 + ], + "angle": 0, + "content": "Risk Estimation based on Color and Smoothness. After generating a view, we can predict the risk based on color and smoothness. This uses two priors: 1. Artifacts are often not smooth or distinct in color. 2. Salient areas with abrupt colors or unevenness are prone to causing artifacts. Color-based risk \\(\\mathbf{r}_{color}\\) and smoothness-based risk \\(\\mathbf{r}_{smooth}\\) are estimated in similar forms:" + }, + { + "type": "equation", + "bbox": [ + 0.314, + 0.311, + 0.785, + 0.326 + ], + "angle": 0, + "content": "\\[\n\\mathbf {r} _ {\\text {c o l o r}} \\left(\\mathbf {x} _ {i}\\right) = \\mathcal {R} _ {f} \\left(\\mathcal {D} _ {c} \\left(\\mathbf {x} _ {i}\\right)\\right); \\quad \\mathbf {r} _ {\\text {s m o o t h}} \\left(\\mathbf {x} _ {i}\\right) = \\mathcal {R} _ {f} \\left(\\mathcal {D} _ {s} \\left(\\mathbf {x} _ {i}\\right)\\right), \\tag {7}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.336, + 0.784, + 0.41 + ], + "angle": 0, + "content": "where \\(\\mathcal{D}_c\\) and \\(\\mathcal{D}_s\\) measures the abruptness of each pixel based on color and smoothness. When implementing them, we choose pixels with the same vertical coordinates across views, and calculate the \"distances\" of each pixel to the mean color and color gradient. Within \\(\\mathcal{R}_f\\), we applied Gaussian filtering after min-max normalization, as those estimated risks are usually noisy." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.412, + 0.785, + 0.485 + ], + "angle": 0, + "content": "Erasing with Estimated Risks. With the estimated risks, we can erase the risky areas on the image warped to next view and the inpainting mask. Assume that we get inpainting mask for current view \\(\\mathbf{m}_i\\) and risks for previous view \\(\\mathbf{r}_{i - 1} = [\\mathbf{r}_{i - 1}^i,\\mathbf{r}_{i - 1}^e,\\mathbf{r}_{i - 1}^c,\\mathbf{r}_{i - 1}^s]\\). The risks are combined linearly and new inpainting mask for current view can be obtained with:" + }, + { + "type": "equation", + "bbox": [ + 0.381, + 0.499, + 0.785, + 0.515 + ], + "angle": 0, + "content": "\\[\n\\mathbf {m} _ {i} ^ {r} = \\mathcal {M} _ {r} \\left(\\mathbf {m} _ {i}, \\mathcal {W} \\left(\\mathbf {r} _ {i - 1} \\cdot \\mathbf {w}, \\mathbf {P} _ {i - 1} ^ {i}\\right)\\right). \\tag {8}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.526, + 0.784, + 0.556 + ], + "angle": 0, + "content": "\\(\\mathcal{M}_r\\) is the risk-based remasking function, and \\(\\mathbf{w}\\) are user defined combination weights. We define \\(\\mathcal{M}_r\\) as thresholding the risk within the warped area." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.556, + 0.785, + 0.661 + ], + "angle": 0, + "content": "Smoothing and Anti-aliasing. We note that the inpainting mask from risk-based erasing may not be smooth. Additionally, sharp and jagged edges on the inpainting mask can lead to artifacts. Therefore, we also employ fixed filtering \\(\\mathcal{M}_f\\), where Gaussian filtering and thresholding are used to smooth the mask and reduce sharp edges, while median filtering is used to reduce jagged edges. Then, we use the final inpainting mask for the combination with guidance, and the risky areas on the warped image are removed and regenerated." + }, + { + "type": "equation", + "bbox": [ + 0.334, + 0.672, + 0.785, + 0.689 + ], + "angle": 0, + "content": "\\[\n\\mathbf {m} _ {i} ^ {f} = \\mathcal {M} _ {f} \\left(\\mathbf {m} _ {i} ^ {r}\\right); \\quad \\hat {\\mathbf {x}} _ {i} ^ {g} = \\mathbf {m} _ {i} ^ {f} \\cdot \\mathbf {x} ^ {g} + \\left(1 - \\mathbf {m} _ {i} ^ {f}\\right) \\cdot \\hat {\\mathbf {x}} _ {i}. \\tag {9}\n\\]" + }, + { + "type": "title", + "bbox": [ + 0.215, + 0.71, + 0.688, + 0.725 + ], + "angle": 0, + "content": "3.4 Bidirectional Generation with Symmetric Guidance" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.734, + 0.784, + 0.84 + ], + "angle": 0, + "content": "Bidirectional Generation. We begin by dividing a unidirectional generation path \\(\\mathbf{x}_0 \\to \\mathbf{x}_1 \\dots \\mathbf{x}_{2n}\\) into two bidirectional generation paths \\(\\mathbf{x}_0 \\to \\mathbf{x}_1 \\to \\dots \\to \\mathbf{x}_n\\) and \\(\\mathbf{x}_{-n} \\gets \\dots \\gets \\mathbf{x}_{-1} \\gets \\mathbf{x}_0\\). Typically, we would make these two generation paths symmetric. And we found this can reduce accumulated errors because the distance to the initial view is reduced in each direction. This consistently reduces artifacts, but may not reduce style and content inconsistency, as there may be different style/content shift in the two directions." + } + ], + [ + { + "type": "page_number", + "bbox": [ + 0.218, + 0.116, + 0.23, + 0.127 + ], + "angle": 0, + "content": "8" + }, + { + "type": "header", + "bbox": [ + 0.272, + 0.115, + 0.355, + 0.128 + ], + "angle": 0, + "content": "A. Liu et al." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.147, + 0.784, + 0.253 + ], + "angle": 0, + "content": "Loop Closure. To ensure loop closure, we can add a \\((2n + 1)\\)-th view as the \"merging view\" to merge the 2 generation paths by warping \\(\\mathbf{x}_n\\) and \\(\\mathbf{x}_{-n}\\) to the \\((2n + 1)\\)-th view and inpaint it. However, if the differences between the two paths are too large, \\(\\mathbf{x}_{2n + 1}\\) may contain image tearing, failing to ensure loop closure. This is due to the partial conditions on each path: there is no information from the other path before merging. Therefore, we rectify the partial conditions by introducing awareness of the other path." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.253, + 0.784, + 0.314 + ], + "angle": 0, + "content": "Symmetric Guidance. We introduce the awareness of the other path by selecting symmetric guidance images from the other path. Specifically, when generating \\(\\mathbf{x}_{i + 1}\\), we will select \\(\\mathbf{x}_{-i}\\) as the guidance image. Thus, \\(\\mathbf{x}_{i + 1}\\) will get the awareness of both paths as it is conditioned on \\(\\mathbf{x}_i\\) and \\(\\mathbf{x}_{-i}\\):" + }, + { + "type": "equation", + "bbox": [ + 0.394, + 0.321, + 0.785, + 0.339 + ], + "angle": 0, + "content": "\\[\n\\mathbf {x} _ {i + 1} \\sim q (\\mathbf {x} \\mid c _ {t}, \\mathbf {x} _ {i}, \\mathbf {x} _ {- i}, \\mathbf {P} _ {i} ^ {i + 1}) \\tag {10}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.346, + 0.784, + 0.391 + ], + "angle": 0, + "content": "We empirically found that bidirectional generation with symmetric guidance is not only effective in ensuring loop closure but also a universally applicable strategy to effectively reduce accumulated errors in various scenarios." + }, + { + "type": "title", + "bbox": [ + 0.215, + 0.411, + 0.561, + 0.425 + ], + "angle": 0, + "content": "3.5 Aligning with Scene Structure Prior" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.433, + 0.784, + 0.52 + ], + "angle": 0, + "content": "When generating full spherical panoramas, we divide a spherical panorama into five parts: first, we generate a 360 panorama as the central part, then we expand upwards and downwards, and finally, we generate two images centered around the top and bottom poles to close up the entire spherical surface. During the expansion and closing stages, models often fail to align with scene structure priors due to partial conditions and generate artifacts." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.523, + 0.784, + 0.627 + ], + "angle": 0, + "content": "Hallucination refers to the artifacts caused by mismatches between partial conditions and scene structure priors. For example, when generating a city scene, using the same prompt during the expansion and closing stages may result in a floating city in the sky or a city underwater. The most direct solution is to input a new prompt, but this would require additional manual effort, which is not ideal. So, we attempt to rectify the partial conditions by extracting scene structure priors and applying semantic and variance control from the initial view." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.63, + 0.784, + 0.718 + ], + "angle": 0, + "content": "Prior Extraction. Although the pretrained T2I model may not align a full panorama with scene structure priors, it can align a single perspective view image with them. Therefore, we extract the scene structure prior from the initial view image \\(\\mathbf{x}_0\\) and incorporate it into the expansion process. For example, when generating the first view image in the upward expansion \\(\\mathbf{x}_0^{ue}\\), we use upper \\(1/3\\) part of the initial view image as guidance with resizing it to the size of \\(\\mathbf{x}_0^{ue}\\)." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.72, + 0.784, + 0.838 + ], + "angle": 0, + "content": "Semantic and Variance Tuning. When the give text prompt only describe part of the scene, we may want generated semantic contents less conditioned on the partial prompt and more conditioned on the prior images during expansion and closing. We achieve this by reducing guidance scale and widen the field of view. Meanwhile, we adjust the variance of the initial noise to avoid the color blocks caused by low guidance scale. Through experimentation, we've found that a combination of slightly high initial variance and low guidance scale can stably reduce hallucinations and color blocks during the expansion and closing stages." + } + ], + [ + { + "type": "header", + "bbox": [ + 0.667, + 0.115, + 0.732, + 0.127 + ], + "angle": 0, + "content": "PanoFree" + }, + { + "type": "page_number", + "bbox": [ + 0.775, + 0.116, + 0.787, + 0.127 + ], + "angle": 0, + "content": "9" + }, + { + "type": "title", + "bbox": [ + 0.217, + 0.145, + 0.377, + 0.163 + ], + "angle": 0, + "content": "4 Experiments" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.199, + 0.788, + 0.261 + ], + "angle": 0, + "content": "We evaluate the performance of PanoFree across three generation tasks: Planar Panorama Generation, 360 Panorama Generation, and Full Panorama Generation. However, note that we focus on planar panorama and \\(360^{\\circ}\\) panorama generation, where the comparisons are more precise and consistent." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.264, + 0.788, + 0.356 + ], + "angle": 0, + "content": "Implementation details. PanoFree is implemented using the publicly available Stable Diffusion code from Diffusers [34] based on the PyTorch framework. For the experiments in the main paper, we utilized the generation and inpainting models of Stable Diffusion (SD) v2.0 [38]. All experiments are conducted on a single NVIDIA RTX A6000 GPU. Further details and specific configurations can be found in the corresponding sections of the main paper and the supplementary." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.359, + 0.788, + 0.405 + ], + "angle": 0, + "content": "Evaluation metrics. We introduce a more comprehensive set of evaluation metrics than prior work [1, 22, 46] covering five themes: image quality, global consistency, prompt capability, diversity, and resource consumption." + }, + { + "type": "text", + "bbox": [ + 0.215, + 0.408, + 0.784, + 0.467 + ], + "angle": 0, + "content": "- Image Quality is measured with Fréchet Inception Distance (FID) [14], Kernel Inception Distance (KID) [2], which measure fidelity and diversity. FID and KID calculated between the views randomly cropped from the panorama and reference images generated by SD with the same prompts." + }, + { + "type": "text", + "bbox": [ + 0.215, + 0.469, + 0.784, + 0.511 + ], + "angle": 0, + "content": "- Global Consistency is measured with Intra-LPIPS (IL) [55] used by SyncDiffusion [22], which is computed by cropping non-overlapping views from a panorama and computing the averaged LPIPS scores of all view pairs." + }, + { + "type": "text", + "bbox": [ + 0.215, + 0.514, + 0.784, + 0.541 + ], + "angle": 0, + "content": "- Prompt Capability is measured via CLIP Score (CS) [13] by computing the text-image similarity of randomly cropped views of the panorama." + }, + { + "type": "text", + "bbox": [ + 0.215, + 0.545, + 0.784, + 0.617 + ], + "angle": 0, + "content": "- Panorama Diversity is also measured by FID and KID. Additionally, we propose Cross-LPIPS (CS) [55]. Cross-LPIPS is computed across 2 panoramas generated with a same text with different random seeds. We crop non-overlapping views from each panorama, and compute the averaged LPIPS scores of all view pairs where two views come from different panoramas." + }, + { + "type": "text", + "bbox": [ + 0.215, + 0.62, + 0.784, + 0.678 + ], + "angle": 0, + "content": "- Resource Consumption includes time consumption, measured by the cumulative time cost of all diffusion processes to generate a single panorama, and peak GPU memory consumption, measured by the maximum GPU memory consumption during inference." + }, + { + "type": "list", + "bbox": [ + 0.215, + 0.408, + 0.784, + 0.678 + ], + "angle": 0, + "content": null + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.685, + 0.788, + 0.777 + ], + "angle": 0, + "content": "Evaluation Settings. Prior work either used arbitrary prompts [1,22] or only focused on a single type of scene [46]. Instead, we consider 3 distinct scene types: indoor, street, and city scenes, and natural scenes. We obtained 100 prompts for each type from ChatGPT [35]. We use 10 random seeds per prompt for planar panorama and 360 panorama generation, and 3 different random seeds per prompt for full panorama generation (see supplementary for details)." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.78, + 0.788, + 0.841 + ], + "angle": 0, + "content": "User Study. For planar panorama generation and \\(360^{\\circ}\\) panorama generation, we conducted four user studies for each task to further evaluate the global consistency, image quality, prompt compatibility, and diversity of the generated panoramas (see supplementary for details)." + } + ], + [ + { + "type": "page_number", + "bbox": [ + 0.218, + 0.116, + 0.236, + 0.127 + ], + "angle": 0, + "content": "10" + }, + { + "type": "header", + "bbox": [ + 0.272, + 0.115, + 0.356, + 0.128 + ], + "angle": 0, + "content": "A. Liu et al." + }, + { + "type": "image", + "bbox": [ + 0.26, + 0.145, + 0.737, + 0.277 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.214, + 0.283, + 0.785, + 0.325 + ], + "angle": 0, + "content": "Fig. 3: Planar Panorama generation results. Compared to vanilla Sequential Generation and MultiDiffusion (MD) [1], PanoFree achieves superior global consistency and image quality. It is also comparable to SyncDiffusion (SYD) [22] in these aspects." + }, + { + "type": "table_caption", + "bbox": [ + 0.214, + 0.334, + 0.785, + 0.39 + ], + "angle": 0, + "content": "Table 1: Comparison of tuning free methods for Planar Panorama generation using Stable Diffusion [38]. We find PanoFree (PF) outperforms the state-of-the-art while having low computational requirements. Note that Cross-LPIPS and Intra-LPIPS are in \\(10^{-2}\\) scale, KID is in \\(10^{-3}\\) scale." + }, + { + "type": "table", + "bbox": [ + 0.259, + 0.403, + 0.735, + 0.48 + ], + "angle": 0, + "content": "
MethodIntra-LPIPS↓Cross-LPIPS↑FID↓KID↓CS↑Time (s)↓Memory (GB)↓
SG70.4071.0324.914.3326.68253.2
MD [1]68.4869.9221.163.5027.89955.8
SYD [22]64.4868.0720.563.6227.1812810.0
PF (Ours)65.3469.6817.053.8027.21263.2
" + }, + { + "type": "title", + "bbox": [ + 0.215, + 0.505, + 0.504, + 0.518 + ], + "angle": 0, + "content": "4.1 Planar Panorama Generation" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.536, + 0.787, + 0.581 + ], + "angle": 0, + "content": "Planar Panorama corresponds to the scene observed with camera translation along the focal plane in reality. This is a relatively simple task, as it only involves extending the image without considering more complex geometric changes." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.582, + 0.79, + 0.627 + ], + "angle": 0, + "content": "Baselines. We have chosen 3 tuning-free baselines for comparison, Vanilla Sequential Generation (SG), MultiDiffusion (MD) [1] and SyncDiffusion (SYD) [22]. Additional details are in the supplementary." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.628, + 0.785, + 0.657 + ], + "angle": 0, + "content": "Results. The quantitative and qualitative evaluations are shown in Table 1 and Fig. 3, respectively. Below we compare PanoFree to each baseline." + }, + { + "type": "text", + "bbox": [ + 0.215, + 0.659, + 0.785, + 0.719 + ], + "angle": 0, + "content": "- Compared with vanilla Sequential Generation, PanoFree significantly enhances image quality and global consistency, demonstrating its effectiveness in reducing accumulated errors. Moreover, PanoFree does not compromise diversity or have a significant effect on GPU time and memory overhead." + }, + { + "type": "text", + "bbox": [ + 0.215, + 0.72, + 0.785, + 0.763 + ], + "angle": 0, + "content": "- Compared with MultiDiffusion, PanoFree has significant advantages in image quality and global consistency. Meanwhile, its time and GPU memory overhead is only \\(26\\%\\) and \\(55\\%\\) that of MultiDiffusion, respectively." + }, + { + "type": "text", + "bbox": [ + 0.215, + 0.765, + 0.785, + 0.84 + ], + "angle": 0, + "content": "- Compared with SyncDiffusion, PanoFree achieves comparable performance in global consistency and image quality. Although SyncDiffusion performs better in consistency, it requires introducing additional models for latent optimization. This leads PanoFree's time overhead to be \\(20\\%\\) of SyncDiffusion and GPU memory overhead to be \\(32\\%\\) of SyncDiffusion." + }, + { + "type": "list", + "bbox": [ + 0.215, + 0.659, + 0.785, + 0.84 + ], + "angle": 0, + "content": null + } + ], + [ + { + "type": "header", + "bbox": [ + 0.667, + 0.115, + 0.732, + 0.127 + ], + "angle": 0, + "content": "PanoFree" + }, + { + "type": "page_number", + "bbox": [ + 0.769, + 0.116, + 0.784, + 0.127 + ], + "angle": 0, + "content": "11" + }, + { + "type": "image", + "bbox": [ + 0.235, + 0.146, + 0.765, + 0.253 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.215, + 0.262, + 0.784, + 0.304 + ], + "angle": 0, + "content": "Fig. 4: Diversity comparison on Planar Panorama Generation task. Each group is generated using the same text with different random seeds. Compared to MultiDiffusion (MD) [1] and SyncDiffusion (SYD) [22], PanoFree achieves superior diversity." + }, + { + "type": "table_caption", + "bbox": [ + 0.215, + 0.318, + 0.784, + 0.345 + ], + "angle": 0, + "content": "Table 2: User study results of Planar Panorama Generation. 15 questions are used for each evaluation item and answered by 5 Amazon MTurk workers." + }, + { + "type": "table", + "bbox": [ + 0.226, + 0.352, + 0.779, + 0.41 + ], + "angle": 0, + "content": "
Consistency (%)Quality (%)Prompt Compatibility (%)Diversity (%)
SYD [22]52.746.041.335.3
PF (ours)47.354.058.764.7
" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.431, + 0.784, + 0.506 + ], + "angle": 0, + "content": "The Loss of Diversity with Joint Diffusion. When using different random seeds with the same prompt, methods using Joint Diffusion exhibit reduced diversity in their results. In contrast, our PanoFree method can better maintain diversity (see cross-LPIPS scores in Table 1). Additionally, we believe this is the source of PanoFree's gains over MultiDiffusion and SyncDiffusion in FID." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.507, + 0.785, + 0.581 + ], + "angle": 0, + "content": "This diversity issue becomes particularly apparent when given some underspecified prompts. Therefore, we generated a \"underspecified set\" consisting of 20 short and blurry prompts to demonstrate this issue. For each prompt, we used 20 different random seeds. We demonstrate the diversity differences qualitatively in Fig. 4. Please refer to supplementary for quantitative analysis." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.582, + 0.785, + 0.643 + ], + "angle": 0, + "content": "User Study. The results in Table 2 clearly show that human evaluators believe PanoFree produces more diverse panoramas and demonstrates better compatibility with prompts than SyncDiffusion [22]. Additionally, both methods exhibit similar levels of global consistency and image quality." + }, + { + "type": "title", + "bbox": [ + 0.215, + 0.665, + 0.478, + 0.679 + ], + "angle": 0, + "content": "4.2 360 Panorama Generation" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.69, + 0.785, + 0.779 + ], + "angle": 0, + "content": "Due to the distortion caused by equirectangular projection, generating 360-degree panoramas is more challenging than planar panorama generation. Vanilla sequential generation tends to produce many artifacts, significantly decreasing image quality. Moreover, MultiDiffusion [1] and SyncDiffusion [22] cannot be directly used for generating 360-degree panoramas. As far as we know, PanoFree is the first implementation of training-free 360-degree panorama generation." + }, + { + "type": "text", + "bbox": [ + 0.215, + 0.78, + 0.787, + 0.81 + ], + "angle": 0, + "content": "Baselines. We used 2 baselines for comparison: Vanilla Sequential Generation (SG) and MVDiffusion (MVD). Additional details are in the supplementary." + }, + { + "type": "text", + "bbox": [ + 0.215, + 0.81, + 0.787, + 0.84 + ], + "angle": 0, + "content": "Results. The quantitative and qualitative evaluations are shown in Table 3 and Fig. 5 respectively. Below we compare PanoFree to each baseline." + } + ], + [ + { + "type": "page_number", + "bbox": [ + 0.218, + 0.116, + 0.236, + 0.127 + ], + "angle": 0, + "content": "12" + }, + { + "type": "header", + "bbox": [ + 0.272, + 0.115, + 0.355, + 0.127 + ], + "angle": 0, + "content": "A. Liu et al." + }, + { + "type": "table_caption", + "bbox": [ + 0.214, + 0.145, + 0.788, + 0.2 + ], + "angle": 0, + "content": "Table 3: Comparison of \\(360^{\\circ}\\) Panorama generation methods using Stable Diffusion [38]. We find PanoFree (PF) still outperforms the state-of-the-art while having low computational requirements. Note that Cross-LPIPS and Intra-LPIPS are in \\(10^{-2}\\) scale, KID is in \\(10^{-3}\\) scale." + }, + { + "type": "table", + "bbox": [ + 0.258, + 0.213, + 0.737, + 0.281 + ], + "angle": 0, + "content": "
MethodIntra-LPIPS↓Cross-LPIPS↑FID↓KID↓CS↑Time (s)↓Memory (GB)↓
SG70.6273.0632.287.9026.35213.2
MVD [46]67.7170.0737.898.7626.271106.9
PF (ours)68.6272.6725.847.4826.51223.2
" + }, + { + "type": "image", + "bbox": [ + 0.245, + 0.283, + 0.756, + 0.422 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.214, + 0.429, + 0.785, + 0.471 + ], + "angle": 0, + "content": "Fig. 5: \\(360^{\\circ}\\) Panorama generation results. Compared to vanilla Sequential Generation (SG), PanoFree achieves superior global consistency and image quality. It is also comparable to MVDiffusion (MVD) [46] in these aspects." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.481, + 0.788, + 0.556 + ], + "angle": 0, + "content": "- Compared with vanilla Sequential Generation, PanoFree significantly enhances image quality and global consistency. Specifically, vanilla sequential generation creates artifacts with complex optical geometry transformations, severely impacting image quality. However, PanoFree effectively recovers image quality by estimating and erasing risky areas, minimizing artifact propagation." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.556, + 0.788, + 0.707 + ], + "angle": 0, + "content": "- Compared with MVDiffusion, PanoFree achieves comparability in image quality and global consistency, yet significantly outperforms in terms of time, GPU memory overhead, and diversity. Particularly, MVDiffusion is significantly worse than PanoFree in terms of FID and KID scores, even underperforming vanilla Sequential Generation. This is partly due to the inevitable bias of MVDiffusion's generated results towards the training dataset, resulting in larger discrepancies compared to those produced by Stable Diffusion. Visually, MVDiffusion also exhibits a noticeable lack of generation diversity. As depicted in Fig. 6, given a prompt, results generated with different random seeds show minimal variation in both content and style." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.707, + 0.788, + 0.768 + ], + "angle": 0, + "content": "User Study. The user study results in Table 4 show that human evaluators believe PanoFree also produces more diverse \\(360^{\\circ}\\) panoramas compared with MVDiffusion [46]. And PanoFree demonstrates better global consistency. Both methods exhibit similar levels of image quality and prompt comparability." + }, + { + "type": "title", + "bbox": [ + 0.215, + 0.787, + 0.564, + 0.803 + ], + "angle": 0, + "content": "4.3 Full Spherical Panorama Generation" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.81, + 0.787, + 0.84 + ], + "angle": 0, + "content": "PanoFree is also the first to achieve feasible tuning-free generation for Full Spherical Panoramas. However, it's hard to conduct meaningful comparisons due the" + } + ], + [ + { + "type": "header", + "bbox": [ + 0.668, + 0.115, + 0.732, + 0.127 + ], + "angle": 0, + "content": "PanoFree" + }, + { + "type": "page_number", + "bbox": [ + 0.769, + 0.116, + 0.786, + 0.127 + ], + "angle": 0, + "content": "13" + }, + { + "type": "image_caption", + "bbox": [ + 0.316, + 0.147, + 0.672, + 0.158 + ], + "angle": 0, + "content": "\"Charming alpine village nestled among snow-capped peaks\"" + }, + { + "type": "image", + "bbox": [ + 0.264, + 0.159, + 0.498, + 0.272 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.368, + 0.275, + 0.402, + 0.286 + ], + "angle": 0, + "content": "MVD" + }, + { + "type": "image", + "bbox": [ + 0.502, + 0.159, + 0.735, + 0.272 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.605, + 0.275, + 0.624, + 0.286 + ], + "angle": 0, + "content": "PF" + }, + { + "type": "image_caption", + "bbox": [ + 0.215, + 0.293, + 0.785, + 0.335 + ], + "angle": 0, + "content": "Fig. 6: Diversity comparison on \\(360^{\\circ}\\) Panorama Generation task. Each group is generated using the same text with different random seeds. Compared to MVDiffusion (MVD) [46], PanoFree achieves superior diversity." + }, + { + "type": "table_caption", + "bbox": [ + 0.215, + 0.347, + 0.785, + 0.375 + ], + "angle": 0, + "content": "Table 4: User study results of \\(360^{\\circ}\\) Panorama Generation. 15 questions are used for each evaluation item and answered by 5 Amazon MTurk workers." + }, + { + "type": "table", + "bbox": [ + 0.226, + 0.387, + 0.779, + 0.447 + ], + "angle": 0, + "content": "
Consistency (%)Quality (%)Prompt Compatibility (%)Diversity (%)
MVD [46]40.748.747.333.3
PF (ours)59.351.352.666.6
" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.465, + 0.788, + 0.603 + ], + "angle": 0, + "content": "lack of tuning-free generation baseline methods or those with strong out-of-scope generation capabilities for Full Spherical Panoramas generation task. Thus, we conduct qualitative evaluation as well as comparison with vanilla sequential generation by showcasing generated results in Fig. 7. Vanilla sequential generation exhibit more artifacts as distortion increases. Additionally, partial conditioning issue mentioned in Sec. 3.5 causes hallucinations. And PanoFree still could effectively reduce artifacts and hallucinations. Note that we start both methods from \\(360^{\\circ}\\) panoramas generated by PanoFree into full spherical panoramas, otherwise vanilla sequential generation will perform even worse." + }, + { + "type": "title", + "bbox": [ + 0.216, + 0.62, + 0.391, + 0.635 + ], + "angle": 0, + "content": "4.4 Ablation Study" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.641, + 0.788, + 0.764 + ], + "angle": 0, + "content": "Tab. 5 contains an ablation study that sequentially integrates each PanoFree component. We evaluate consistency (Intra-LPIPS) and image quality (FID) with \\(30\\%\\) of the prompts from Sec. 4.1 & Sec. 4.2 for both planar and \\(360^{\\circ}\\) panorama. We show that cross-view guidance provides the strongest benefit, followed by distance and edge-based risky area erasing. These components effectively reduce image tearing and visual chaos. Color and smoothness-based erasing have a smaller impact, likely due noise in these low-level features. Qualitative results are in Sec. B of the supplementary." + }, + { + "type": "title", + "bbox": [ + 0.216, + 0.782, + 0.36, + 0.798 + ], + "angle": 0, + "content": "5 Conclusion" + }, + { + "type": "text", + "bbox": [ + 0.215, + 0.81, + 0.787, + 0.84 + ], + "angle": 0, + "content": "We present PanoFree, a tuning-free multi-view image generation that supports an extensive array of correspondences. PanoFree improves error accumulation" + } + ], + [ + { + "type": "page_number", + "bbox": [ + 0.218, + 0.116, + 0.236, + 0.127 + ], + "angle": 0, + "content": "14" + }, + { + "type": "header", + "bbox": [ + 0.272, + 0.115, + 0.356, + 0.127 + ], + "angle": 0, + "content": "A. Liu et al." + }, + { + "type": "image", + "bbox": [ + 0.219, + 0.145, + 0.778, + 0.335 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.214, + 0.341, + 0.785, + 0.382 + ], + "angle": 0, + "content": "Fig. 7: Full Spherical Panorama generation results. Vanilla Sequential Generation (SG) tends to generate hallucinations due to partial conditions, while PanoFree effectively mitigates this issue." + }, + { + "type": "table_caption", + "bbox": [ + 0.214, + 0.401, + 0.446, + 0.54 + ], + "angle": 0, + "content": "Table 5: Quantitative ablation of PanoFree components using \\(30\\%\\) of prompts from Sec. 4.1 & 4.2. Intra-LPIPS \\((10^{-2})\\) and FID show cross-view guidance offers the most benefit, followed by distance and edge-based risky area erasing. Color and smoothness-based erasing have minimal impact." + }, + { + "type": "table", + "bbox": [ + 0.458, + 0.4, + 0.773, + 0.535 + ], + "angle": 0, + "content": "
TaskMethodIntra-LPIPS↓FID↓
PlanarNone70.8424.75
+ CG66.6320.63
+ Dist65.8718.21
360°None71.3433.47
+ CG69.2127.38
+ Dist & Edge69.0326.69
+ Color & Smooth68.9426.51
" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.594, + 0.788, + 0.761 + ], + "angle": 0, + "content": "by enhancing cross-view awareness and refining the warping and inpainting processes through cross-view guidance, risky area estimation and erasing, and symmetric bidirectional guided generation for loop closure, alongside guidance-based semantic and density control for scene structure preservation. PanoFree is evaluated on various panorama types—Planar, \\(360^{\\circ}\\), and Full Spherical Panoramas. PanoFree demonstrates significant error reduction, improved global consistency, and image quality across different scenarios without extra fine-tuning. Compared to existing methods, PanoFree is up to 5x more efficient in time and 3x more efficient in GPU memory usage, and maintains superior diversity of results (2x better in our user study). Moreover, PanoFree can be extended to texture generation for 3D models. We intend to explore these possibilities in future research." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.765, + 0.788, + 0.84 + ], + "angle": 0, + "content": "Limitations. A limitation of our work is that we are unable to generate scenes beyond the capability of the pre-trained T2I model. Therefore, we rely on large pre-trained T2I models to ensure the broad application scope. And when provided with text descriptions beyond the capability range of the pre-trained T2I models, the generated results may not match the text." + } + ], + [ + { + "type": "header", + "bbox": [ + 0.667, + 0.115, + 0.732, + 0.127 + ], + "angle": 0, + "content": "PanoFree" + }, + { + "type": "page_number", + "bbox": [ + 0.769, + 0.117, + 0.785, + 0.127 + ], + "angle": 0, + "content": "15" + }, + { + "type": "title", + "bbox": [ + 0.218, + 0.145, + 0.323, + 0.16 + ], + "angle": 0, + "content": "References" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.177, + 0.785, + 0.204 + ], + "angle": 0, + "content": "1. Bar-Tal, O., Yariv, L., Lipman, Y., Dekel, T.: Multidiffusion: Fusing diffusion paths for controlled image generation (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.205, + 0.785, + 0.244 + ], + "angle": 0, + "content": "2. Binkowski, M., Sutherland, D.J., Arbel, M., Gretton, A.: Demystifying mmd gans. ArXiv abs/1801.01401 (2018), https://apisemantic scholar.org/CorpusID: 3531856" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.246, + 0.785, + 0.273 + ], + "angle": 0, + "content": "3. Brock, A., Donahue, J., Simonyan, K.: Large scale gan training for high fidelity natural image synthesis. arXiv preprint arXiv:1809.11096 (2018)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.274, + 0.785, + 0.314 + ], + "angle": 0, + "content": "4. Chen, D.Z., Siddiqui, Y., Lee, H.Y., Tulyakov, S., Nießner, M.: Text2tex: Text-driven texture synthesis via diffusion models. arXiv preprint arXiv:2303.11396 (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.315, + 0.785, + 0.342 + ], + "angle": 0, + "content": "5. Chen, Z., Wang, G., Liu, Z.: Text2light: Zero-shot text-driven hdr panorama generation. ACM Transactions on Graphics (TOG) 41(6), 1-16 (2022)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.343, + 0.785, + 0.397 + ], + "angle": 0, + "content": "6. Cheng, Y.C., Lin, C.H., Lee, H.Y., Ren, J., Tulyakov, S., Yang, M.H.: Inout: Diverse image outpainting via gan inversion. 2022 IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR) pp. 11421-11430 (2021), https://apisemantic scholar.org/CorpusID:232478397" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.398, + 0.785, + 0.425 + ], + "angle": 0, + "content": "7. Dhariwal, P., Nichol, A.: Diffusion models beat gans on image synthesis. Advances in Neural Information Processing Systems 34, 8780-8794 (2021)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.426, + 0.785, + 0.467 + ], + "angle": 0, + "content": "8. Esser, P., Rombach, R., Ommer, B.: Taming transformers for high-resolution image synthesis. In: Proceedings of the IEEE/CVF conference on computer vision and pattern recognition. pp. 12873-12883 (2021)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.468, + 0.785, + 0.495 + ], + "angle": 0, + "content": "9. Fang, C., Hu, X., Luo, K., Tan, P.: Ctrl-room: Controllable text-to-3d room meshes generation with layout constraints. arXiv preprint arXiv:2310.03602 (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.496, + 0.785, + 0.536 + ], + "angle": 0, + "content": "0. Feng, M., Liu, J., Cui, M., Xie, X.: Diffusion360: Seamless 360 degree panoramic image generation based on diffusion models. ArXiv abs/2311.13141 (2023), https://api-semanticscholar.org/CorpusID:265351889" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.537, + 0.785, + 0.563 + ], + "angle": 0, + "content": "1. Fridman, R., Abecasis, A., Kasten, Y., Dekel, T.: Scenescape: Text-driven consistent scene generation. arXiv preprint arXiv:2302.01133 (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.565, + 0.785, + 0.605 + ], + "angle": 0, + "content": "2. Goodfellow, I., Pouget-Abadie, J., Mirza, M., Xu, B., Warde-Farley, D., Ozair, S., Courville, A., Bengio, Y.: Generative adversarial networks. Communications of the ACM 63(11), 139-144 (2020)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.606, + 0.785, + 0.646 + ], + "angle": 0, + "content": "3. Hessel, J., Holtzman, A., Forbes, M., Bras, R.L., Choi, Y.: Clipscore: A reference-free evaluation metric for image captioning. ArXiv abs/2104.08718 (2021), https://api-semanticscholar.org/CorpusID:233296711" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.647, + 0.785, + 0.701 + ], + "angle": 0, + "content": "4. Heusel, M., Ramsauer, H., Unterthiner, T., Nessler, B., Hochreiter, S.: Gans trained by a two time-scale update rule converge to a local nash equilibrium. In: Neural Information Processing Systems (2017), https://api-semanticscholar.org/CorpusID:326772" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.702, + 0.785, + 0.73 + ], + "angle": 0, + "content": "5. Ho, J., Jain, A., Abbeel, P.: Denoising diffusion probabilistic models. Advances in Neural Information Processing Systems (NeurIPS) (2020)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.731, + 0.785, + 0.77 + ], + "angle": 0, + "content": "6. Ho, J., Sahara, C., Chan, W., Fleet, D.J., Norouzi, M., Salimans, T.: Cascaded diffusion models for high fidelity image generation. J. Mach. Learn. Res. 23(47), 1-33 (2022)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.772, + 0.785, + 0.798 + ], + "angle": 0, + "content": "7. Ho, J., Salimans, T.: Classifier-free diffusion guidance. arXiv preprint arXiv:2207.12598 (2022)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.799, + 0.785, + 0.84 + ], + "angle": 0, + "content": "8. Hollein, L., Cao, A., Owens, A., Johnson, J., Nießner, M.: Text2room: Extracting textured 3d meshes from 2d text-to-image models. arXiv preprint arXiv:2303.11989 (2023)" + }, + { + "type": "list", + "bbox": [ + 0.226, + 0.177, + 0.785, + 0.84 + ], + "angle": 0, + "content": null + } + ], + [ + { + "type": "page_number", + "bbox": [ + 0.218, + 0.116, + 0.236, + 0.127 + ], + "angle": 0, + "content": "16" + }, + { + "type": "header", + "bbox": [ + 0.272, + 0.115, + 0.356, + 0.128 + ], + "angle": 0, + "content": "A. Liu et al." + }, + { + "type": "ref_text", + "bbox": [ + 0.218, + 0.148, + 0.785, + 0.175 + ], + "angle": 0, + "content": "19. Karras, T., Aittala, M., Aila, T., Laine, S.: Elucidating the design space of diffusion-based generative models. arXiv preprint arXiv:2206.00364 (2022)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.177, + 0.785, + 0.203 + ], + "angle": 0, + "content": "20. Karras, T., Aittala, M., Aila, T., Laine, S.: Elucidating the design space of diffusion-based generative models. arXiv preprint arXiv:2206.00364 (2022)" + }, + { + "type": "ref_text", + "bbox": [ + 0.218, + 0.204, + 0.785, + 0.245 + ], + "angle": 0, + "content": "21. Karras, T., Aittala, M., Laine, S., Härkönen, E., Hellsten, J., Lehtinen, J., Aila, T.: Alias-free generative adversarial networks. Advances in Neural Information Processing Systems 34, 852-863 (2021)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.246, + 0.785, + 0.286 + ], + "angle": 0, + "content": "22. Lee, Y., Kim, K., Kim, H., Sung, M.: Syncdiffusion: Coherent montage via synchronized joint diffusions. Advances in Neural Information Processing Systems 36 (2024)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.287, + 0.785, + 0.327 + ], + "angle": 0, + "content": "23. Li, J., Bansal, M.: Panogen: Text-conditioned panoramic environment generation for vision-and-language navigation. ArXiv abs/2305.19195 (2023), https://api.sementicscholar.org/CorpusID:258967291" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.328, + 0.785, + 0.384 + ], + "angle": 0, + "content": "24. Lin, C.H., Chang, C.C., Chen, Y.S., Juan, D.C., Wei, W., Chen, H.T.: Cocogan: Generation by parts via conditional coordinating. 2019 IEEE/CVF International Conference on Computer Vision (ICCV) pp. 4511-4520 (2019), https://api-semanticscholar.org/CorpusID:90262507" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.384, + 0.785, + 0.424 + ], + "angle": 0, + "content": "25. Lin, C.H., Lee, H.Y., Cheng, Y.C., Tulyakov, S., Yang, M.H.: Infinitygan: Towards infinite-pixel image synthesis. In: International Conference on Learning Representations (2021), https://api-semanticscholar.org/CorpusID:238419701" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.425, + 0.785, + 0.466 + ], + "angle": 0, + "content": "26. Liu, X., Zhang, X., Ma, J., Peng, J., et al.: Instaflow: One step is enough for high-quality diffusion-based text-to-image generation. In: The International Conference on Learning Representations (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.467, + 0.785, + 0.508 + ], + "angle": 0, + "content": "27. Lu, C., Zhou, Y., Bao, F., Chen, J., Li, C., Zhu, J.: Dpm-solver: A fast ode solver for diffusion probabilistic model sampling in around 10 steps. arXiv preprint arXiv:2206.00927 (2022)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.509, + 0.785, + 0.563 + ], + "angle": 0, + "content": "28. Lugmayr, A., Danelljan, M., Romero, A., Yu, F., Timofte, R., Gool, L.V.: Repaint: Inpainting using denoising diffusion probabilistic models. 2022 IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR) pp. 11451-11461 (2022), https://api-semanticscholar.org/CorpusID:246240274" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.564, + 0.785, + 0.618 + ], + "angle": 0, + "content": "29. Meng, C., He, Y., Song, Y., Song, J., Wu, J., Zhu, J.Y., Ermon, S.: Sdedit: Guided image synthesis and editing with stochastic differential equations. In: International Conference on Learning Representations (2021), https://apisemantic scholar.org/CorpusID:245704504" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.619, + 0.785, + 0.672 + ], + "angle": 0, + "content": "30. Mou, C., Wang, X., Xie, L., Zhang, J., Qi, Z., Shan, Y., Qie, X.: T2i-adapter: Learning adapters to dig out more controllable ability for text-to-image diffusion models. ArXiv abs/2302.08453 (2023), https://api_semanticscholar.org/CorpusID: 256900833" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.674, + 0.785, + 0.715 + ], + "angle": 0, + "content": "31. Nichol, A., Dhariwal, P., Ramesh, A., Shyam, P., Mishkin, P., McGrew, B., Sutskever, I., Chen, M.: Glide: Towards photorealistic image generation and editing with text-guided diffusion models. arXiv preprint arXiv:2112.10741 (2021)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.716, + 0.785, + 0.769 + ], + "angle": 0, + "content": "32. Oh, C.H., Cho, W., Park, D., Chae, Y., Wang, L., Yoon, K.J.: Bips: Bi-modal indoor panorama synthesis via residual depth-aided adversarial learning. ArXiv abs/2112.06179 (2021), https://api-semanticscholar.org/CorpusID: 245123664" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.771, + 0.785, + 0.799 + ], + "angle": 0, + "content": "33. Peebles, W., Xie, S.: Scalable diffusion models with transformers. In: International Conference on Computer Vision. pp. 4195-4205 (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.799, + 0.785, + 0.841 + ], + "angle": 0, + "content": "34. von Platen, P., Patil, S., Lozhkov, A., Cuenca, P., Lambert, N., Rasul, K., Davaadorj, M., Wolf, T.: Diffusers: State-of-the-art diffusion models. https://github.com/huggingface/diffusers (2022)" + }, + { + "type": "list", + "bbox": [ + 0.217, + 0.148, + 0.785, + 0.841 + ], + "angle": 0, + "content": null + } + ], + [ + { + "type": "header", + "bbox": [ + 0.667, + 0.115, + 0.732, + 0.127 + ], + "angle": 0, + "content": "PanoFree" + }, + { + "type": "page_number", + "bbox": [ + 0.769, + 0.116, + 0.786, + 0.127 + ], + "angle": 0, + "content": "17" + }, + { + "type": "ref_text", + "bbox": [ + 0.214, + 0.147, + 0.788, + 0.175 + ], + "angle": 0, + "content": "35. Radford, A., Narasimhan, K., Salimans, T., Sutskever, I.: Language models are unsupervised multitask learners (2019)" + }, + { + "type": "ref_text", + "bbox": [ + 0.214, + 0.175, + 0.788, + 0.217 + ], + "angle": 0, + "content": "36. Ramesh, A., Dhariwal, P., Nichol, A., Chu, C., Chen, M.: Hierarchical text-conditional image generation with clip latents. arXiv preprint arXiv:2204.06125 (2022)" + }, + { + "type": "ref_text", + "bbox": [ + 0.216, + 0.218, + 0.787, + 0.245 + ], + "angle": 0, + "content": "37. Razavi, A., Van den Oord, A., Vinyals, O.: Generating diverse high-fidelity images with vq-vae-2. Advances in neural information processing systems 32 (2019)" + }, + { + "type": "ref_text", + "bbox": [ + 0.216, + 0.246, + 0.787, + 0.286 + ], + "angle": 0, + "content": "38. Rombach, R., Blattmann, A., Lorenz, D., Esser, P., Ommer, B.: High-resolution image synthesis with latent diffusion models. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 10684-10695 (2022)" + }, + { + "type": "ref_text", + "bbox": [ + 0.216, + 0.287, + 0.787, + 0.342 + ], + "angle": 0, + "content": "39. Sahara, C., Chan, W., Saxena, S., Li, L., Whang, J., Denton, E.L., Ghasemipour, K., Gontijo Lopes, R., Karagol Ayan, B., Salimans, T., et al.: Photorealistic text-to-image diffusion models with deep language understanding. Advances in Neural Information Processing Systems (NeurIPS) (2022)" + }, + { + "type": "ref_text", + "bbox": [ + 0.216, + 0.343, + 0.787, + 0.383 + ], + "angle": 0, + "content": "40. Sohl-Dickstein, J., Weiss, E., Maheswaranathan, N., Ganguli, S.: Deep unsupervised learning using nonequilibrium thermodynamics. In: International Conference on Machine Learning. PMLR (2015)" + }, + { + "type": "ref_text", + "bbox": [ + 0.216, + 0.384, + 0.787, + 0.411 + ], + "angle": 0, + "content": "41. Song, J., Meng, C., Ermon, S.: Denoising diffusion implicit models. arXiv preprint arXiv:2010.02502 (2020)" + }, + { + "type": "ref_text", + "bbox": [ + 0.216, + 0.412, + 0.787, + 0.438 + ], + "angle": 0, + "content": "42. Song, Y., Ermon, S.: Generative modeling by estimating gradients of the data distribution. Advances in neural information processing systems (NeurIPS) (2019)" + }, + { + "type": "ref_text", + "bbox": [ + 0.216, + 0.439, + 0.787, + 0.466 + ], + "angle": 0, + "content": "43. Song, Y., Ermon, S.: Generative modeling by estimating gradients of the data distribution. Advances in neural information processing systems 32 (2019)" + }, + { + "type": "ref_text", + "bbox": [ + 0.216, + 0.467, + 0.787, + 0.494 + ], + "angle": 0, + "content": "44. Song, Y., Ermon, S.: Improved techniques for training score-based generative models. Advances in neural information processing systems 33, 12438-12448 (2020)" + }, + { + "type": "ref_text", + "bbox": [ + 0.216, + 0.495, + 0.787, + 0.535 + ], + "angle": 0, + "content": "45. Song, Y., Sohl-Dickstein, J.N., Kingma, D.P., Kumar, A., Ermon, S., Poole, B.: Score-based generative modeling through stochastic differential equations. ICLR (2021)" + }, + { + "type": "ref_text", + "bbox": [ + 0.216, + 0.536, + 0.787, + 0.589 + ], + "angle": 0, + "content": "46. Tang, S., Zhang, F., Chen, J., Wang, P., Furukawa, Y.: Mvdiffusion: Enabling holistic multi-view image generation with correspondence-aware diffusion. ArXiv abs/2307.01097 (2023), https://api_semanticscholar.org/CorpusID: 259316427" + }, + { + "type": "ref_text", + "bbox": [ + 0.216, + 0.591, + 0.787, + 0.647 + ], + "angle": 0, + "content": "47. Teterwak, P., Sarna, A., Krishnan, D., Maschinot, A., Belanger, D., Liu, C., Freeman, W.T.: Boundless: Generative adversarial networks for image extension. 2019 IEEE/CVF International Conference on Computer Vision (ICCV) pp. 10520-10529 (2019), https://api_semanticscholar.org/CorpusID:201106503" + }, + { + "type": "ref_text", + "bbox": [ + 0.216, + 0.648, + 0.787, + 0.674 + ], + "angle": 0, + "content": "48. Van Den Oord, A., Vinyals, O., et al.: Neural discrete representation learning. Advances in neural information processing systems 30 (2017)" + }, + { + "type": "ref_text", + "bbox": [ + 0.216, + 0.675, + 0.787, + 0.702 + ], + "angle": 0, + "content": "49. Voynov, A., Hertz, A., Arar, M., Fruchter, S., Cohen-Or, D.: Anylens: A generative diffusion model with any rendering lens (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.216, + 0.702, + 0.787, + 0.743 + ], + "angle": 0, + "content": "50. Wang, G., Yang, Y., Loy, C.C., Liu, Z.: Stylelight: Hdr panorama generation for lighting estimation and editing. In: European Conference on Computer Vision (2022), https://api_semanticscholar.org/CorpusID:251196614" + }, + { + "type": "ref_text", + "bbox": [ + 0.216, + 0.744, + 0.787, + 0.785 + ], + "angle": 0, + "content": "51. Wang, H., Xiang, X., Fan, Y., Xue, J.H.: Customizing 360-degree panoramicas through text-to-image diffusion models. ArXiv abs/2310.18840 (2023), https://api.sementicscholar.org/CorpusID:264590753" + }, + { + "type": "ref_text", + "bbox": [ + 0.216, + 0.786, + 0.787, + 0.813 + ], + "angle": 0, + "content": "52. Wu, T., Zheng, C., Cham, T.J.: Panodiffusion: 360-degree panorama outpainting via diffusion (2023), https://api-semanticscholar.org/CorpusID:259360663" + }, + { + "type": "ref_text", + "bbox": [ + 0.216, + 0.814, + 0.787, + 0.84 + ], + "angle": 0, + "content": "53. Zhang, L., Rao, A., Agrawala, M.: Adding conditional control to text-to-image diffusion models (2023)" + }, + { + "type": "list", + "bbox": [ + 0.214, + 0.147, + 0.788, + 0.84 + ], + "angle": 0, + "content": null + } + ], + [ + { + "type": "page_number", + "bbox": [ + 0.218, + 0.116, + 0.236, + 0.127 + ], + "angle": 0, + "content": "18" + }, + { + "type": "header", + "bbox": [ + 0.272, + 0.115, + 0.356, + 0.128 + ], + "angle": 0, + "content": "A. Liu et al." + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.147, + 0.787, + 0.203 + ], + "angle": 0, + "content": "54. Zhang, Q., Song, J., Huang, X., Chen, Y., Liu, M.Y.: Diffcollage: Parallel generation of large content with diffusion models. 2023 IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR) pp. 10188-10198 (2023), https://api-semanticscholar.org/CorpusID:257834007" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.204, + 0.787, + 0.259 + ], + "angle": 0, + "content": "55. Zhang, R., Isola, P., Efros, A.A., Shechtman, E., Wang, O.: The unreasonable effectiveness of deep features as a perceptual metric. 2018 IEEE/CVF Conference on Computer Vision and Pattern Recognition pp. 586-595 (2018), https://apisemantic scholar.org/CorpusID:4766599" + }, + { + "type": "list", + "bbox": [ + 0.217, + 0.147, + 0.787, + 0.259 + ], + "angle": 0, + "content": null + } + ] +] \ No newline at end of file diff --git a/2024/PanoFree_ Tuning-Free Holistic Multi-view Image Generation with Cross-view Self-Guidance/cbd590c4-f320-4c08-83b4-3c8681128a51_origin.pdf b/2024/PanoFree_ Tuning-Free Holistic Multi-view Image Generation with Cross-view Self-Guidance/cbd590c4-f320-4c08-83b4-3c8681128a51_origin.pdf new file mode 100644 index 0000000000000000000000000000000000000000..80f2825041d28a08881e83b2fc122219cb84956b --- /dev/null +++ b/2024/PanoFree_ Tuning-Free Holistic Multi-view Image Generation with Cross-view Self-Guidance/cbd590c4-f320-4c08-83b4-3c8681128a51_origin.pdf @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:035fa29cc46fd26431704914c83c5704eb31f7eeaf057f4c5924dc67b1a2319c +size 8936239 diff --git a/2024/PanoFree_ Tuning-Free Holistic Multi-view Image Generation with Cross-view Self-Guidance/full.md b/2024/PanoFree_ Tuning-Free Holistic Multi-view Image Generation with Cross-view Self-Guidance/full.md new file mode 100644 index 0000000000000000000000000000000000000000..d8258aa62c95473436d934bf47ae2f2f7021a021 --- /dev/null +++ b/2024/PanoFree_ Tuning-Free Holistic Multi-view Image Generation with Cross-view Self-Guidance/full.md @@ -0,0 +1,326 @@ +# PanoFree: Tuning-Free Holistic Multi-view Image Generation with Cross-view Self-Guidance + +Aoming Liu $^{1,2*}$ , Zhong Li $^{1\dagger}$ , Zhang Chen $^{1\dagger}$ , Nannan Li $^{2}$ , Yi Xu $^{1}$ , and Bryan A. Plummer $^{2}$ + +$^{1}$ OPPO US Research Center, Palo Alto, CA 94303, USA + $^{2}$ Boston University, Boston, MA 02215, USA + +Abstract. Immersive scene generation, notably panorama creation, benefits significantly from the adaptation of large pre-trained text-to-image (T2I) models for multi-view image generation. Due to the high cost of acquiring multi-view images, tuning-free generation is preferred. However, existing methods are either limited to simple correspondences or require extensive fine-tuning to capture complex ones. We present PanoFree, a novel method for tuning-free multi-view image generation that supports an extensive array of correspondences. PanoFree sequentially generates multi-view images using iterative warping and inpainting, addressing the key issues of inconsistency and artifacts from error accumulation without the need for fine-tuning. It improves error accumulation by enhancing cross-view awareness and refines the warping and inpainting processes via cross-view guidance, risky area estimation and erasing, and symmetric bidirectional guided generation for loop closure, alongside guidance-based semantic and density control for scene structure preservation. In experiments on Planar, $360^{\circ}$ , and Full Spherical Panoramas, PanoFree demonstrates significant error reduction, improves global consistency, and boosts image quality without extra fine-tuning. Compared to existing methods, PanoFree is up to 5x more efficient in time and 3x more efficient in GPU memory usage, and maintains superior diversity of results (2x better in our user study). PanoFree offers a viable alternative to costly fine-tuning or the use of additional pre-trained models. + +Keywords: Tuning-free generation; Multi-view Image, Panorama + +# 1 Introduction + +Text-to-image (T2I) generation over multiple views for immersive scenes, like panorama generation, is a challenging task requiring coherence and diversity among many generated images (e.g., [3, 8, 12, 16, 17, 21, 36, 37, 39, 48]). Early work using GANs or VAEs (e.g., [4-6, 11, 24, 25, 47, 50]) have been replaced recently with diffusion-based models (e.g., [1, 9, 10, 22, 23, 46, 49, 51, 52, 54]), often leveraging Stable Diffusion [38]. State-of-the-art panorama generation methods + +![](images/62107d5d06df8c37b569648ab5279266d0f49673f42d9a021305129cce9c2d81.jpg) +Fig. 1: PanoFree can generate multi-view images according to different types of correspondences without fine-tuning, and a natural application is tuning-free generation for different types of panoramas. We demonstrate this by generating three commonly used panoramas. Top: Planar Panorama; Middle: 360 Panorama; Bottom: Full Panorama. + +use Joint Diffusion (e.g., [1,22,46]), where parallel diffusion processes to generate multi-view images and enhancing global consistency by fusing latent or attention features based on cross-view correspondences. However, we find these methods can only generate some types of panoramas, limiting their scope (e.g., [1,22]), or require fine-tuning using expensive panorama datasets (e.g., [22,46]). + +To address these challenges, we propose PanoFree a tuning-free multi-view image generation method using iterative warping and inpainting of perspective images to support diverse correspondences with low costs (see Fig. 1 for example generations). Iterative warping and inpainting of perspective images provide a means to the diverse multi-view correspondences required in panorama generation without additional fine-tuning [4, 11, 18]. However, recent work has overlooked these benefits due to accumulated errors causing suboptimal image quality [1, 22, 46]. We find that most accumulated errors from iterative warping can be attributed to the deficient conditions during generation. Specifically, conditioning solely on the previous image narrows cross-view awareness, leading to inconsistencies. Warping and inpainting can also propagate noise, e.g., truncated objects or jagged edges. Additionally, the given conditions may be incomplete to meet specific requirements, such as ensuring 360-degree consistency for loop closure and maintaining correct spatial relationships for realistic scenes. + +To address these issues, PanoFree expand cross-view awareness by conditioning the current view on multiple views with guided image synthesis techniques such as SDEdit [29]. Then, PanoFree estimates and erases the risky areas, regions likely containing noise, to reduce the noise introduced by warping and inpainting. In addition, PanoFree adopts a bidirectional generation path with a symmetrical conditioning strategy for loop closure. Lastly, PanoFree further + +utilizes pseudo global guidance with region-specific semantic and density control to make scene structure more reasonable. + +We evaluate PanoFree on three text-to-panorama generation tasks: Planar, $360^{\circ}$ , and Full Spherical Panoramas. PanoFree effectively alleviating accumulated errors in sequential generation, and significantly improves image quality and global consistency (e.g. $31.6\%$ better in FID). This enables PanoFree to have better (or at least comparable) results to the state-of-the-art [1,22,46], despite these methods either having narrower applications or requiring fine-tuning datasets. Specifically, PanoFree is up to 5x more efficient in time and 3x more efficient in GPU memory usage, and maintains superior diversity of results (2x better in our user study). Lastly, PanoFree is also highly flexible, enabling it to plug-in-and-play with various pre-trained T2I models and adapters. + +Our contributions can be summarized as follows: + +- We introduce PanoFree, a tuning-free multi-view image generation method applicable for various correspondences and pre-trained T2I models. Thus PanoFree can greatly reduce the data and fine-tuning costs for immersive scene generation tasks such as text-to-panorama generation. +- We provide a in-depth perspective of accumulated errors and identify the deficient conditions as the main causes. We further effectively rectify deficient conditions and alleviate accumulated errors with the cross-view guidance as well as risky area estimation and erasing in PanoFree. +- As far as we know, PanoFree is the first to achieve feasible tuning-free generation for $360^{\circ}$ Panoramas and Full Spherical Panoramas. + +# 2 Related Work + +Diffusion Models [15, 19, 40, 42-45] are a popular framework for generative models. Early work required a long trajectory for sampling to produce high-quality samples [7, 45], before being sped up with advanced sampling techniques that also preserved generation quality [20, 26, 27, 41]. Latent Diffusion Models (LDMs) [33, 38] made these models more efficient by training in the latent space. + +T2I Diffusion and Panorama Generation. Diffusion models are widely adopted for text-to-image (T2I) generation [31, 36, 38, 39]. Many downstream tasks used large pre-trained T2I diffusion models, like Stable Diffusion [38], to boost performance [4, 11, 18], including panorama generation [4, 11, 18]. These methods have largely supplanted GAN and VAE methods [4-6, 11, 24, 25, 32, 47, 50], with most recent work in panorama generation tasks using diffusion models [1, 9, 10, 22, 23, 46, 49, 51, 52, 54]. These diffusion-based panorama generation methods use joint diffusion to fuses multiple diffusion processes through latent or attention manipulation [1, 22, 46]. However, they are either limited to modeling simple correspondences or require extensive fine-tuning to model complex ones. + +Guided Image Synthesis with Diffusion Models. It can be challenging to achieve satisfactory results solely relying on text guidance. Therefore, some prior work [28-30, 53] guide or control the generation results with reference images as fine-grained condition. ControlNet [53] and T2I-Adapter [30] are the most + +commonly used methods to incorporate additional image conditions by adding extra image encoders, but they all require few-shot fine-tuning. SDEdit [29] achieves tuning-free guided image synthesis by adding noise to the guide image and then denoising it back to a real image using a pre-trained diffusion model. + +# 3 Method + +PanoFree targets the text-to-panorama generation task, which takes textual descriptions as guidance to create multi-view perspective images that can be stitched into a wide-angle, high-quality panorama. PanoFree generates multiview images through sequential warping and inpainting steps, which typically results in acclimated errors due to deficit conditions (discussed in Sec. 3.1). Each component of Panofree is designed to minimize the effect of these various deficiencies. Specifically, Sec. 3.2 mitigates inconsistency using SDEdit-based cross-view guidance and Sec. 3.3 removes artifact-inducing content by estimating and erasing risky areas. At a higher level, PanoFree employs Bidirectional Generation with Symmetric Guidance for loop closure and error reduction (Sec. 3.4). Additionally, it applies guidance-based semantic and density control for scene structure preservation (Sec. 3.5). See Fig. 2 for an overview of our approach. + +# 3.1 Deficient Conditions behind Accumulated Errors + +In this section, we discuss the causes behind the deficit conditions in the iterative warping and inpainting process and reformulate the problem as conditional generation. Given the text prompt $c_{t}$ , $i$ -th view's image $\mathbf{x}_i$ , warping function $\mathcal{W}$ , transformation matrix of the projection from $i$ -th view to the $(i + 1)$ -th view $\mathbf{P}_i^{i + 1}$ , and pre-trained T2I inpainting model $\Phi_{inp}$ , the warping and inpainting step to generate the $(i + 1)$ -th view can be denoted as: + +$$ +\hat {\mathbf {x}} _ {i}, \mathbf {m} _ {i} = \mathcal {W} \left(\mathbf {x} _ {i}, \mathbf {P} _ {i} ^ {i + 1}\right); \quad \mathbf {x} _ {i + 1} = \boldsymbol {\Phi} _ {i n p} \left(\hat {\mathbf {x}} _ {i}, \mathbf {m} _ {i}, c _ {t}\right), \tag {1} +$$ + +where $\hat{\mathbf{x}}_i$ is the image warped from $i$ -th view to $(i + 1)$ -th view and $\mathbf{m}_i$ is the masking indicating the area to inpaint. And we can simplify the warping and inpainting steps in the following conditional image generation form: + +$$ +\mathbf {x} _ {i + 1} \sim q (\mathbf {x} | c _ {t}, \mathbf {x} _ {i}, \mathbf {P} _ {i} ^ {i + 1}). \tag {2} +$$ + +However, during this generation process, we found conditions can become deficient. Major accumulated errors arise from three types of deficient conditions: Biased Conditions, Noisy Conditions, and Partial Conditions. See Sec. A.2 in the supplementary for detailed error illustrations. + +Biased Conditions is the most obvious problem. In the above step, $\mathbf{x}_{i + 1}$ is solely conditioned on $\mathbf{x}_i$ , which biases the cross-view awareness heavily to $i$ -th view. If $\mathbf{x}_i$ has deviated from the desired global distribution in certain aspects, then $\mathbf{x}_{i + 1}$ is likely to continue deviating in the same direction, resulting in significant inconsistency. We also found that slight style and content shifts often accumulate in this way, leading to significant inconsistency between distant views. + +![](images/f4ba846b44b812434a4f6c4d8a722475914a09d12b41c9f8cc42bacce32643ec.jpg) +(a) Overview of the generation and guidance framework. + +![](images/a6d11eb0e77c32c25a9be5c47e6eb0de6b4e27f38465edc050d87c2e447d6e48.jpg) +Fig. 2: Overview of our PanoFree method, taking 360 Panorama Generation as an example. (a): At a framework level, PanoFree adopts two generation paths with opposite viewpoint translation or rotation. It enhances consistency by symmetrically selecting views from the other path as guidance to generate a new view (Sec. 3.4). Loop closure is ensured by merging these two paths. (b): In each warping and inpainting step, PanoFree reduces accumulated error by guiding the inpainting process with cross-view images (Sec. 3.2), along with estimating and erasing risky areas (Sec. 3.3). + +Noisy Conditions mainly refer to $\mathbf{x}_i$ containing artifact-inducing contents. Existing artifacts in $\mathbf{x}_i$ could guide inpainting model to generate similar artifacts in $\mathbf{x}_{i + 1}$ and propagate to every following view. Additionally, disjointed or distorted areas, jagged or sharp content, and objects truncated by edges in $\mathbf{x}_i$ are also highly risky to introduce artifacts in $\mathbf{x}_{i + 1}$ . + +Partial Conditions refer to the conditions not containing all the necessary information to meet specific requirements. For example, if we follow Eq. (2) on the final view, we lack information about $\mathbf{x}_0$ , making it impossible to generate image coherent with $\mathbf{x}_0$ to ensure loop closure. Additionally, using a single text prompt to generate all views within a full spherical panorama may lead to hallucinations, such as cities floating in the sky or underwater mountains. + +# 3.2 Cross-View Guidance + +To rectify the biased conditions discussed in Sec.3.1 and enlarge cross-view awareness, a natural idea is to let $\mathbf{x}_{i + 1}$ conditioned on more views, + +$$ +\mathbf {x} _ {i + 1} \sim q (\mathbf {x} | c _ {t}, \mathbf {x} _ {i}, \mathbf {P} _ {i} ^ {i + 1}, \mathbf {x} _ {1} ^ {g}, \dots , \mathbf {x} _ {m} ^ {g}) \qquad (3) +$$ + +where $\mathbf{x}_1^g, \ldots, \mathbf{x}_m^g$ are selected from $\mathbf{x}_0, \ldots, \mathbf{x}_{i-1}$ . This naturally results in a Guided Image Synthesis task form with self-generated images as guidance. Many existing methods can be adapted to implement our design, such as ControlNet [53] and T2I adapter [30]. To avoid relying on fine-tuning and reduce memory and time costs, we choose SDEdit [29], a training-free guided image synthesis approach, with a single guidance image $\mathbf{x}^g \in [\mathbf{x}_0, \ldots, \mathbf{x}_{i-1}]$ . + +Guided Image Synthesis using SDEdit. Given $\mathbf{x}^g$ as guidance, SDEdit establishes a Gaussian distribution using $\mathbf{x}^g$ as the expectation and the intermediate status at time $t_0$ in the reverse SDE process. The desired data distribution is obtained by gradually removing noise from $\mathbf{x}^g(t_0)$ : + +$$ +\mathbf {x} ^ {g} \left(t _ {0}\right) \sim \mathcal {N} \left(\mathbf {x} ^ {g}, \sigma^ {2} \left(t _ {0}\right) \mathbf {I}\right); \quad \mathbf {x} \sim \operatorname {S D E d i t} \left(\mathbf {x} ^ {g}, t _ {0}, \boldsymbol {\Phi}\right), \tag {4} +$$ + +where $\Phi$ denotes a generative model. In PanoFree, we use inpainting mask $\mathbf{m}_i$ to paste the guidance image $\mathbf{x}^g$ to the blank areas in the warped image $\hat{\mathbf{x}}_i$ and then use SDEdit in the inpainting process: + +$$ +\hat {\mathbf {x}} _ {i} ^ {g} = \mathbf {m} _ {i} \cdot \mathbf {x} ^ {g} + (1 - \mathbf {m} _ {i}) \cdot \hat {\mathbf {x}} _ {i}; \quad \mathbf {x} _ {i} \sim \operatorname {S D E d i t} \left(\hat {\mathbf {x}} _ {i} ^ {g}, t _ {0}, \boldsymbol {\Phi} _ {i n p}\right). \tag {5} +$$ + +Since we only want to use additional guidance images to rectify the biased conditions rather than replicate the guidance image, we use $t_0 \in [0.9, 1.0)$ in practice. Meanwhile, we found that different selection of generation path and guidance image results in different generation qualities, and the optimal choice may vary for different tasks. We introduce a general selection effective for various tasks in Section 3.4 and provide an example of extending this technique to make scenes more realistic in specific scenarios in Section 3.5. + +# 3.3 Risky Area Estimation and Erasing + +To rectify the noisy conditions discussed in Sec.3.1 and eliminate the accumulation of artifacts, a natural idea is to detect and localize the artifact-inducing contents, and erase them. However, precise detection and localization often requires costly training. Thus, we turn to roughly estimate and erase the risky areas that are likely to contain artifact-inducing contents, based on indicators often associated with artifact: distances, color and smoothness. See Sec. A.2 in the supplementary for examples. + +Risk Estimation based on Distances. We consider the distance from the center point of the initial view $\mathbf{x}_0$ and the distance to the edges. This is based on two priors: 1. The farther from the initial view, the more accumulated errors and the more likely to contain artifact-inducing contents. 2. Areas close to the edges are highly risky because truncated objects are mostly generated around the edges, and areas near the edges are often more severely distorted during warping. We use initial risk $\mathbf{r}_{init}$ to represent the risk estimated based on the distance from the initial view, and edge risk $\mathbf{r}_{edge}$ to represent the risk estimated based on the distance from edges. They are derived from the following: + +$$ +\mathbf {r} _ {\text {i n i t}} \left(\mathbf {c} _ {i}\right) = \mathcal {R} _ {p} \left(\mathcal {D} _ {0} \left(\mathbf {c} _ {i}\right)\right); \quad \mathbf {r} _ {\text {e d g e}} \left(\mathbf {c} _ {i}\right) = \mathcal {R} _ {p} \left(\mathcal {D} _ {\mathbf {e}} \left(\mathbf {c} _ {i}\right)\right), \tag {6} +$$ + +where $\mathbf{c}_i$ represents the pixel coordinates of $\mathbf{x}_i$ within the panorama coordinate system, $\mathcal{D}_0$ measures the distance to the center point of the initial view along the generation path, $\mathcal{D}_{\mathbf{e}}$ measures the distance to all edges $\mathbf{e}$ , and $\mathcal{R}_p$ is a scaling function. We use weighted euclidean distance for $\mathcal{D}_0$ , Gaussian filters for $\mathcal{D}_{\mathbf{e}}$ , and min-max normalization for $\mathcal{R}_p$ . + +Risk Estimation based on Color and Smoothness. After generating a view, we can predict the risk based on color and smoothness. This uses two priors: 1. Artifacts are often not smooth or distinct in color. 2. Salient areas with abrupt colors or unevenness are prone to causing artifacts. Color-based risk $\mathbf{r}_{color}$ and smoothness-based risk $\mathbf{r}_{smooth}$ are estimated in similar forms: + +$$ +\mathbf {r} _ {\text {c o l o r}} \left(\mathbf {x} _ {i}\right) = \mathcal {R} _ {f} \left(\mathcal {D} _ {c} \left(\mathbf {x} _ {i}\right)\right); \quad \mathbf {r} _ {\text {s m o o t h}} \left(\mathbf {x} _ {i}\right) = \mathcal {R} _ {f} \left(\mathcal {D} _ {s} \left(\mathbf {x} _ {i}\right)\right), \tag {7} +$$ + +where $\mathcal{D}_c$ and $\mathcal{D}_s$ measures the abruptness of each pixel based on color and smoothness. When implementing them, we choose pixels with the same vertical coordinates across views, and calculate the "distances" of each pixel to the mean color and color gradient. Within $\mathcal{R}_f$ , we applied Gaussian filtering after min-max normalization, as those estimated risks are usually noisy. + +Erasing with Estimated Risks. With the estimated risks, we can erase the risky areas on the image warped to next view and the inpainting mask. Assume that we get inpainting mask for current view $\mathbf{m}_i$ and risks for previous view $\mathbf{r}_{i - 1} = [\mathbf{r}_{i - 1}^i,\mathbf{r}_{i - 1}^e,\mathbf{r}_{i - 1}^c,\mathbf{r}_{i - 1}^s]$ . The risks are combined linearly and new inpainting mask for current view can be obtained with: + +$$ +\mathbf {m} _ {i} ^ {r} = \mathcal {M} _ {r} \left(\mathbf {m} _ {i}, \mathcal {W} \left(\mathbf {r} _ {i - 1} \cdot \mathbf {w}, \mathbf {P} _ {i - 1} ^ {i}\right)\right). \tag {8} +$$ + +$\mathcal{M}_r$ is the risk-based remasking function, and $\mathbf{w}$ are user defined combination weights. We define $\mathcal{M}_r$ as thresholding the risk within the warped area. + +Smoothing and Anti-aliasing. We note that the inpainting mask from risk-based erasing may not be smooth. Additionally, sharp and jagged edges on the inpainting mask can lead to artifacts. Therefore, we also employ fixed filtering $\mathcal{M}_f$ , where Gaussian filtering and thresholding are used to smooth the mask and reduce sharp edges, while median filtering is used to reduce jagged edges. Then, we use the final inpainting mask for the combination with guidance, and the risky areas on the warped image are removed and regenerated. + +$$ +\mathbf {m} _ {i} ^ {f} = \mathcal {M} _ {f} \left(\mathbf {m} _ {i} ^ {r}\right); \quad \hat {\mathbf {x}} _ {i} ^ {g} = \mathbf {m} _ {i} ^ {f} \cdot \mathbf {x} ^ {g} + \left(1 - \mathbf {m} _ {i} ^ {f}\right) \cdot \hat {\mathbf {x}} _ {i}. \tag {9} +$$ + +# 3.4 Bidirectional Generation with Symmetric Guidance + +Bidirectional Generation. We begin by dividing a unidirectional generation path $\mathbf{x}_0 \to \mathbf{x}_1 \dots \mathbf{x}_{2n}$ into two bidirectional generation paths $\mathbf{x}_0 \to \mathbf{x}_1 \to \dots \to \mathbf{x}_n$ and $\mathbf{x}_{-n} \gets \dots \gets \mathbf{x}_{-1} \gets \mathbf{x}_0$ . Typically, we would make these two generation paths symmetric. And we found this can reduce accumulated errors because the distance to the initial view is reduced in each direction. This consistently reduces artifacts, but may not reduce style and content inconsistency, as there may be different style/content shift in the two directions. + +Loop Closure. To ensure loop closure, we can add a $(2n + 1)$ -th view as the "merging view" to merge the 2 generation paths by warping $\mathbf{x}_n$ and $\mathbf{x}_{-n}$ to the $(2n + 1)$ -th view and inpaint it. However, if the differences between the two paths are too large, $\mathbf{x}_{2n + 1}$ may contain image tearing, failing to ensure loop closure. This is due to the partial conditions on each path: there is no information from the other path before merging. Therefore, we rectify the partial conditions by introducing awareness of the other path. + +Symmetric Guidance. We introduce the awareness of the other path by selecting symmetric guidance images from the other path. Specifically, when generating $\mathbf{x}_{i + 1}$ , we will select $\mathbf{x}_{-i}$ as the guidance image. Thus, $\mathbf{x}_{i + 1}$ will get the awareness of both paths as it is conditioned on $\mathbf{x}_i$ and $\mathbf{x}_{-i}$ : + +$$ +\mathbf {x} _ {i + 1} \sim q (\mathbf {x} \mid c _ {t}, \mathbf {x} _ {i}, \mathbf {x} _ {- i}, \mathbf {P} _ {i} ^ {i + 1}) \tag {10} +$$ + +We empirically found that bidirectional generation with symmetric guidance is not only effective in ensuring loop closure but also a universally applicable strategy to effectively reduce accumulated errors in various scenarios. + +# 3.5 Aligning with Scene Structure Prior + +When generating full spherical panoramas, we divide a spherical panorama into five parts: first, we generate a 360 panorama as the central part, then we expand upwards and downwards, and finally, we generate two images centered around the top and bottom poles to close up the entire spherical surface. During the expansion and closing stages, models often fail to align with scene structure priors due to partial conditions and generate artifacts. + +Hallucination refers to the artifacts caused by mismatches between partial conditions and scene structure priors. For example, when generating a city scene, using the same prompt during the expansion and closing stages may result in a floating city in the sky or a city underwater. The most direct solution is to input a new prompt, but this would require additional manual effort, which is not ideal. So, we attempt to rectify the partial conditions by extracting scene structure priors and applying semantic and variance control from the initial view. + +Prior Extraction. Although the pretrained T2I model may not align a full panorama with scene structure priors, it can align a single perspective view image with them. Therefore, we extract the scene structure prior from the initial view image $\mathbf{x}_0$ and incorporate it into the expansion process. For example, when generating the first view image in the upward expansion $\mathbf{x}_0^{ue}$ , we use upper $1/3$ part of the initial view image as guidance with resizing it to the size of $\mathbf{x}_0^{ue}$ . + +Semantic and Variance Tuning. When the give text prompt only describe part of the scene, we may want generated semantic contents less conditioned on the partial prompt and more conditioned on the prior images during expansion and closing. We achieve this by reducing guidance scale and widen the field of view. Meanwhile, we adjust the variance of the initial noise to avoid the color blocks caused by low guidance scale. Through experimentation, we've found that a combination of slightly high initial variance and low guidance scale can stably reduce hallucinations and color blocks during the expansion and closing stages. + +# 4 Experiments + +We evaluate the performance of PanoFree across three generation tasks: Planar Panorama Generation, 360 Panorama Generation, and Full Panorama Generation. However, note that we focus on planar panorama and $360^{\circ}$ panorama generation, where the comparisons are more precise and consistent. + +Implementation details. PanoFree is implemented using the publicly available Stable Diffusion code from Diffusers [34] based on the PyTorch framework. For the experiments in the main paper, we utilized the generation and inpainting models of Stable Diffusion (SD) v2.0 [38]. All experiments are conducted on a single NVIDIA RTX A6000 GPU. Further details and specific configurations can be found in the corresponding sections of the main paper and the supplementary. + +Evaluation metrics. We introduce a more comprehensive set of evaluation metrics than prior work [1, 22, 46] covering five themes: image quality, global consistency, prompt capability, diversity, and resource consumption. + +- Image Quality is measured with Fréchet Inception Distance (FID) [14], Kernel Inception Distance (KID) [2], which measure fidelity and diversity. FID and KID calculated between the views randomly cropped from the panorama and reference images generated by SD with the same prompts. +- Global Consistency is measured with Intra-LPIPS (IL) [55] used by SyncDiffusion [22], which is computed by cropping non-overlapping views from a panorama and computing the averaged LPIPS scores of all view pairs. +- Prompt Capability is measured via CLIP Score (CS) [13] by computing the text-image similarity of randomly cropped views of the panorama. +- Panorama Diversity is also measured by FID and KID. Additionally, we propose Cross-LPIPS (CS) [55]. Cross-LPIPS is computed across 2 panoramas generated with a same text with different random seeds. We crop non-overlapping views from each panorama, and compute the averaged LPIPS scores of all view pairs where two views come from different panoramas. +- Resource Consumption includes time consumption, measured by the cumulative time cost of all diffusion processes to generate a single panorama, and peak GPU memory consumption, measured by the maximum GPU memory consumption during inference. + +Evaluation Settings. Prior work either used arbitrary prompts [1,22] or only focused on a single type of scene [46]. Instead, we consider 3 distinct scene types: indoor, street, and city scenes, and natural scenes. We obtained 100 prompts for each type from ChatGPT [35]. We use 10 random seeds per prompt for planar panorama and 360 panorama generation, and 3 different random seeds per prompt for full panorama generation (see supplementary for details). + +User Study. For planar panorama generation and $360^{\circ}$ panorama generation, we conducted four user studies for each task to further evaluate the global consistency, image quality, prompt compatibility, and diversity of the generated panoramas (see supplementary for details). + +![](images/c85aecde535a1bec8d23bcc3582d7bba3a47c6d8ca6e7ac19655d03409efa4c4.jpg) +Fig. 3: Planar Panorama generation results. Compared to vanilla Sequential Generation and MultiDiffusion (MD) [1], PanoFree achieves superior global consistency and image quality. It is also comparable to SyncDiffusion (SYD) [22] in these aspects. + +Table 1: Comparison of tuning free methods for Planar Panorama generation using Stable Diffusion [38]. We find PanoFree (PF) outperforms the state-of-the-art while having low computational requirements. Note that Cross-LPIPS and Intra-LPIPS are in $10^{-2}$ scale, KID is in $10^{-3}$ scale. + +
MethodIntra-LPIPS↓Cross-LPIPS↑FID↓KID↓CS↑Time (s)↓Memory (GB)↓
SG70.4071.0324.914.3326.68253.2
MD [1]68.4869.9221.163.5027.89955.8
SYD [22]64.4868.0720.563.6227.1812810.0
PF (Ours)65.3469.6817.053.8027.21263.2
+ +# 4.1 Planar Panorama Generation + +Planar Panorama corresponds to the scene observed with camera translation along the focal plane in reality. This is a relatively simple task, as it only involves extending the image without considering more complex geometric changes. + +Baselines. We have chosen 3 tuning-free baselines for comparison, Vanilla Sequential Generation (SG), MultiDiffusion (MD) [1] and SyncDiffusion (SYD) [22]. Additional details are in the supplementary. + +Results. The quantitative and qualitative evaluations are shown in Table 1 and Fig. 3, respectively. Below we compare PanoFree to each baseline. + +- Compared with vanilla Sequential Generation, PanoFree significantly enhances image quality and global consistency, demonstrating its effectiveness in reducing accumulated errors. Moreover, PanoFree does not compromise diversity or have a significant effect on GPU time and memory overhead. +- Compared with MultiDiffusion, PanoFree has significant advantages in image quality and global consistency. Meanwhile, its time and GPU memory overhead is only $26\%$ and $55\%$ that of MultiDiffusion, respectively. +- Compared with SyncDiffusion, PanoFree achieves comparable performance in global consistency and image quality. Although SyncDiffusion performs better in consistency, it requires introducing additional models for latent optimization. This leads PanoFree's time overhead to be $20\%$ of SyncDiffusion and GPU memory overhead to be $32\%$ of SyncDiffusion. + +![](images/370e94a9c5a2cde88e8f1ad51d2edc54951b80405d196c6691e9e67acd0b50f2.jpg) +Fig. 4: Diversity comparison on Planar Panorama Generation task. Each group is generated using the same text with different random seeds. Compared to MultiDiffusion (MD) [1] and SyncDiffusion (SYD) [22], PanoFree achieves superior diversity. + +Table 2: User study results of Planar Panorama Generation. 15 questions are used for each evaluation item and answered by 5 Amazon MTurk workers. + +
Consistency (%)Quality (%)Prompt Compatibility (%)Diversity (%)
SYD [22]52.746.041.335.3
PF (ours)47.354.058.764.7
+ +The Loss of Diversity with Joint Diffusion. When using different random seeds with the same prompt, methods using Joint Diffusion exhibit reduced diversity in their results. In contrast, our PanoFree method can better maintain diversity (see cross-LPIPS scores in Table 1). Additionally, we believe this is the source of PanoFree's gains over MultiDiffusion and SyncDiffusion in FID. + +This diversity issue becomes particularly apparent when given some underspecified prompts. Therefore, we generated a "underspecified set" consisting of 20 short and blurry prompts to demonstrate this issue. For each prompt, we used 20 different random seeds. We demonstrate the diversity differences qualitatively in Fig. 4. Please refer to supplementary for quantitative analysis. + +User Study. The results in Table 2 clearly show that human evaluators believe PanoFree produces more diverse panoramas and demonstrates better compatibility with prompts than SyncDiffusion [22]. Additionally, both methods exhibit similar levels of global consistency and image quality. + +# 4.2 360 Panorama Generation + +Due to the distortion caused by equirectangular projection, generating 360-degree panoramas is more challenging than planar panorama generation. Vanilla sequential generation tends to produce many artifacts, significantly decreasing image quality. Moreover, MultiDiffusion [1] and SyncDiffusion [22] cannot be directly used for generating 360-degree panoramas. As far as we know, PanoFree is the first implementation of training-free 360-degree panorama generation. + +Baselines. We used 2 baselines for comparison: Vanilla Sequential Generation (SG) and MVDiffusion (MVD). Additional details are in the supplementary. + +Results. The quantitative and qualitative evaluations are shown in Table 3 and Fig. 5 respectively. Below we compare PanoFree to each baseline. + +Table 3: Comparison of $360^{\circ}$ Panorama generation methods using Stable Diffusion [38]. We find PanoFree (PF) still outperforms the state-of-the-art while having low computational requirements. Note that Cross-LPIPS and Intra-LPIPS are in $10^{-2}$ scale, KID is in $10^{-3}$ scale. + +
MethodIntra-LPIPS↓Cross-LPIPS↑FID↓KID↓CS↑Time (s)↓Memory (GB)↓
SG70.6273.0632.287.9026.35213.2
MVD [46]67.7170.0737.898.7626.271106.9
PF (ours)68.6272.6725.847.4826.51223.2
+ +![](images/3043c22643dd25c5abe0f5f7f8e86baec2d046fa46ea0de2d672715d85cfe54a.jpg) +Fig. 5: $360^{\circ}$ Panorama generation results. Compared to vanilla Sequential Generation (SG), PanoFree achieves superior global consistency and image quality. It is also comparable to MVDiffusion (MVD) [46] in these aspects. + +- Compared with vanilla Sequential Generation, PanoFree significantly enhances image quality and global consistency. Specifically, vanilla sequential generation creates artifacts with complex optical geometry transformations, severely impacting image quality. However, PanoFree effectively recovers image quality by estimating and erasing risky areas, minimizing artifact propagation. + +- Compared with MVDiffusion, PanoFree achieves comparability in image quality and global consistency, yet significantly outperforms in terms of time, GPU memory overhead, and diversity. Particularly, MVDiffusion is significantly worse than PanoFree in terms of FID and KID scores, even underperforming vanilla Sequential Generation. This is partly due to the inevitable bias of MVDiffusion's generated results towards the training dataset, resulting in larger discrepancies compared to those produced by Stable Diffusion. Visually, MVDiffusion also exhibits a noticeable lack of generation diversity. As depicted in Fig. 6, given a prompt, results generated with different random seeds show minimal variation in both content and style. + +User Study. The user study results in Table 4 show that human evaluators believe PanoFree also produces more diverse $360^{\circ}$ panoramas compared with MVDiffusion [46]. And PanoFree demonstrates better global consistency. Both methods exhibit similar levels of image quality and prompt comparability. + +# 4.3 Full Spherical Panorama Generation + +PanoFree is also the first to achieve feasible tuning-free generation for Full Spherical Panoramas. However, it's hard to conduct meaningful comparisons due the + +![](images/8acd5007191249bdfb9bf3f8c8c8d669c9b8816e04fe224fb779af2ded2f2dcc.jpg) +"Charming alpine village nestled among snow-capped peaks" +MVD +Fig. 6: Diversity comparison on $360^{\circ}$ Panorama Generation task. Each group is generated using the same text with different random seeds. Compared to MVDiffusion (MVD) [46], PanoFree achieves superior diversity. + +![](images/d9cba1a78fe86f2de9153b9768d86cedb5ba05f56c6435bf11ec496354479830.jpg) +PF + +Table 4: User study results of $360^{\circ}$ Panorama Generation. 15 questions are used for each evaluation item and answered by 5 Amazon MTurk workers. + +
Consistency (%)Quality (%)Prompt Compatibility (%)Diversity (%)
MVD [46]40.748.747.333.3
PF (ours)59.351.352.666.6
+ +lack of tuning-free generation baseline methods or those with strong out-of-scope generation capabilities for Full Spherical Panoramas generation task. Thus, we conduct qualitative evaluation as well as comparison with vanilla sequential generation by showcasing generated results in Fig. 7. Vanilla sequential generation exhibit more artifacts as distortion increases. Additionally, partial conditioning issue mentioned in Sec. 3.5 causes hallucinations. And PanoFree still could effectively reduce artifacts and hallucinations. Note that we start both methods from $360^{\circ}$ panoramas generated by PanoFree into full spherical panoramas, otherwise vanilla sequential generation will perform even worse. + +# 4.4 Ablation Study + +Tab. 5 contains an ablation study that sequentially integrates each PanoFree component. We evaluate consistency (Intra-LPIPS) and image quality (FID) with $30\%$ of the prompts from Sec. 4.1 & Sec. 4.2 for both planar and $360^{\circ}$ panorama. We show that cross-view guidance provides the strongest benefit, followed by distance and edge-based risky area erasing. These components effectively reduce image tearing and visual chaos. Color and smoothness-based erasing have a smaller impact, likely due noise in these low-level features. Qualitative results are in Sec. B of the supplementary. + +# 5 Conclusion + +We present PanoFree, a tuning-free multi-view image generation that supports an extensive array of correspondences. PanoFree improves error accumulation + +![](images/e7adb3f70906836ab705afcf8479db7d0491a05beebe9f5200a7e1256913485c.jpg) +Fig. 7: Full Spherical Panorama generation results. Vanilla Sequential Generation (SG) tends to generate hallucinations due to partial conditions, while PanoFree effectively mitigates this issue. + +Table 5: Quantitative ablation of PanoFree components using $30\%$ of prompts from Sec. 4.1 & 4.2. Intra-LPIPS $(10^{-2})$ and FID show cross-view guidance offers the most benefit, followed by distance and edge-based risky area erasing. Color and smoothness-based erasing have minimal impact. + +
TaskMethodIntra-LPIPS↓FID↓
PlanarNone70.8424.75
+ CG66.6320.63
+ Dist65.8718.21
360°None71.3433.47
+ CG69.2127.38
+ Dist & Edge69.0326.69
+ Color & Smooth68.9426.51
+ +by enhancing cross-view awareness and refining the warping and inpainting processes through cross-view guidance, risky area estimation and erasing, and symmetric bidirectional guided generation for loop closure, alongside guidance-based semantic and density control for scene structure preservation. PanoFree is evaluated on various panorama types—Planar, $360^{\circ}$ , and Full Spherical Panoramas. PanoFree demonstrates significant error reduction, improved global consistency, and image quality across different scenarios without extra fine-tuning. Compared to existing methods, PanoFree is up to 5x more efficient in time and 3x more efficient in GPU memory usage, and maintains superior diversity of results (2x better in our user study). Moreover, PanoFree can be extended to texture generation for 3D models. We intend to explore these possibilities in future research. + +Limitations. A limitation of our work is that we are unable to generate scenes beyond the capability of the pre-trained T2I model. Therefore, we rely on large pre-trained T2I models to ensure the broad application scope. And when provided with text descriptions beyond the capability range of the pre-trained T2I models, the generated results may not match the text. + +# References + +1. Bar-Tal, O., Yariv, L., Lipman, Y., Dekel, T.: Multidiffusion: Fusing diffusion paths for controlled image generation (2023) +2. Binkowski, M., Sutherland, D.J., Arbel, M., Gretton, A.: Demystifying mmd gans. ArXiv abs/1801.01401 (2018), https://apisemantic scholar.org/CorpusID: 3531856 +3. Brock, A., Donahue, J., Simonyan, K.: Large scale gan training for high fidelity natural image synthesis. arXiv preprint arXiv:1809.11096 (2018) +4. Chen, D.Z., Siddiqui, Y., Lee, H.Y., Tulyakov, S., Nießner, M.: Text2tex: Text-driven texture synthesis via diffusion models. arXiv preprint arXiv:2303.11396 (2023) +5. Chen, Z., Wang, G., Liu, Z.: Text2light: Zero-shot text-driven hdr panorama generation. ACM Transactions on Graphics (TOG) 41(6), 1-16 (2022) +6. Cheng, Y.C., Lin, C.H., Lee, H.Y., Ren, J., Tulyakov, S., Yang, M.H.: Inout: Diverse image outpainting via gan inversion. 2022 IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR) pp. 11421-11430 (2021), https://apisemantic scholar.org/CorpusID:232478397 +7. Dhariwal, P., Nichol, A.: Diffusion models beat gans on image synthesis. Advances in Neural Information Processing Systems 34, 8780-8794 (2021) +8. Esser, P., Rombach, R., Ommer, B.: Taming transformers for high-resolution image synthesis. In: Proceedings of the IEEE/CVF conference on computer vision and pattern recognition. pp. 12873-12883 (2021) +9. Fang, C., Hu, X., Luo, K., Tan, P.: Ctrl-room: Controllable text-to-3d room meshes generation with layout constraints. arXiv preprint arXiv:2310.03602 (2023) +0. Feng, M., Liu, J., Cui, M., Xie, X.: Diffusion360: Seamless 360 degree panoramic image generation based on diffusion models. ArXiv abs/2311.13141 (2023), https://api-semanticscholar.org/CorpusID:265351889 +1. Fridman, R., Abecasis, A., Kasten, Y., Dekel, T.: Scenescape: Text-driven consistent scene generation. arXiv preprint arXiv:2302.01133 (2023) +2. Goodfellow, I., Pouget-Abadie, J., Mirza, M., Xu, B., Warde-Farley, D., Ozair, S., Courville, A., Bengio, Y.: Generative adversarial networks. Communications of the ACM 63(11), 139-144 (2020) +3. Hessel, J., Holtzman, A., Forbes, M., Bras, R.L., Choi, Y.: Clipscore: A reference-free evaluation metric for image captioning. ArXiv abs/2104.08718 (2021), https://api-semanticscholar.org/CorpusID:233296711 +4. Heusel, M., Ramsauer, H., Unterthiner, T., Nessler, B., Hochreiter, S.: Gans trained by a two time-scale update rule converge to a local nash equilibrium. In: Neural Information Processing Systems (2017), https://api-semanticscholar.org/CorpusID:326772 +5. Ho, J., Jain, A., Abbeel, P.: Denoising diffusion probabilistic models. Advances in Neural Information Processing Systems (NeurIPS) (2020) +6. Ho, J., Sahara, C., Chan, W., Fleet, D.J., Norouzi, M., Salimans, T.: Cascaded diffusion models for high fidelity image generation. J. Mach. Learn. Res. 23(47), 1-33 (2022) +7. Ho, J., Salimans, T.: Classifier-free diffusion guidance. arXiv preprint arXiv:2207.12598 (2022) +8. Hollein, L., Cao, A., Owens, A., Johnson, J., Nießner, M.: Text2room: Extracting textured 3d meshes from 2d text-to-image models. arXiv preprint arXiv:2303.11989 (2023) + +19. Karras, T., Aittala, M., Aila, T., Laine, S.: Elucidating the design space of diffusion-based generative models. arXiv preprint arXiv:2206.00364 (2022) +20. Karras, T., Aittala, M., Aila, T., Laine, S.: Elucidating the design space of diffusion-based generative models. arXiv preprint arXiv:2206.00364 (2022) +21. Karras, T., Aittala, M., Laine, S., Härkönen, E., Hellsten, J., Lehtinen, J., Aila, T.: Alias-free generative adversarial networks. Advances in Neural Information Processing Systems 34, 852-863 (2021) +22. Lee, Y., Kim, K., Kim, H., Sung, M.: Syncdiffusion: Coherent montage via synchronized joint diffusions. Advances in Neural Information Processing Systems 36 (2024) +23. Li, J., Bansal, M.: Panogen: Text-conditioned panoramic environment generation for vision-and-language navigation. ArXiv abs/2305.19195 (2023), https://api.sementicscholar.org/CorpusID:258967291 +24. Lin, C.H., Chang, C.C., Chen, Y.S., Juan, D.C., Wei, W., Chen, H.T.: Cocogan: Generation by parts via conditional coordinating. 2019 IEEE/CVF International Conference on Computer Vision (ICCV) pp. 4511-4520 (2019), https://api-semanticscholar.org/CorpusID:90262507 +25. Lin, C.H., Lee, H.Y., Cheng, Y.C., Tulyakov, S., Yang, M.H.: Infinitygan: Towards infinite-pixel image synthesis. In: International Conference on Learning Representations (2021), https://api-semanticscholar.org/CorpusID:238419701 +26. Liu, X., Zhang, X., Ma, J., Peng, J., et al.: Instaflow: One step is enough for high-quality diffusion-based text-to-image generation. In: The International Conference on Learning Representations (2023) +27. Lu, C., Zhou, Y., Bao, F., Chen, J., Li, C., Zhu, J.: Dpm-solver: A fast ode solver for diffusion probabilistic model sampling in around 10 steps. arXiv preprint arXiv:2206.00927 (2022) +28. Lugmayr, A., Danelljan, M., Romero, A., Yu, F., Timofte, R., Gool, L.V.: Repaint: Inpainting using denoising diffusion probabilistic models. 2022 IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR) pp. 11451-11461 (2022), https://api-semanticscholar.org/CorpusID:246240274 +29. Meng, C., He, Y., Song, Y., Song, J., Wu, J., Zhu, J.Y., Ermon, S.: Sdedit: Guided image synthesis and editing with stochastic differential equations. In: International Conference on Learning Representations (2021), https://apisemantic scholar.org/CorpusID:245704504 +30. Mou, C., Wang, X., Xie, L., Zhang, J., Qi, Z., Shan, Y., Qie, X.: T2i-adapter: Learning adapters to dig out more controllable ability for text-to-image diffusion models. ArXiv abs/2302.08453 (2023), https://api_semanticscholar.org/CorpusID: 256900833 +31. Nichol, A., Dhariwal, P., Ramesh, A., Shyam, P., Mishkin, P., McGrew, B., Sutskever, I., Chen, M.: Glide: Towards photorealistic image generation and editing with text-guided diffusion models. arXiv preprint arXiv:2112.10741 (2021) +32. Oh, C.H., Cho, W., Park, D., Chae, Y., Wang, L., Yoon, K.J.: Bips: Bi-modal indoor panorama synthesis via residual depth-aided adversarial learning. ArXiv abs/2112.06179 (2021), https://api-semanticscholar.org/CorpusID: 245123664 +33. Peebles, W., Xie, S.: Scalable diffusion models with transformers. In: International Conference on Computer Vision. pp. 4195-4205 (2023) +34. von Platen, P., Patil, S., Lozhkov, A., Cuenca, P., Lambert, N., Rasul, K., Davaadorj, M., Wolf, T.: Diffusers: State-of-the-art diffusion models. https://github.com/huggingface/diffusers (2022) + +35. Radford, A., Narasimhan, K., Salimans, T., Sutskever, I.: Language models are unsupervised multitask learners (2019) +36. Ramesh, A., Dhariwal, P., Nichol, A., Chu, C., Chen, M.: Hierarchical text-conditional image generation with clip latents. arXiv preprint arXiv:2204.06125 (2022) +37. Razavi, A., Van den Oord, A., Vinyals, O.: Generating diverse high-fidelity images with vq-vae-2. Advances in neural information processing systems 32 (2019) +38. Rombach, R., Blattmann, A., Lorenz, D., Esser, P., Ommer, B.: High-resolution image synthesis with latent diffusion models. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 10684-10695 (2022) +39. Sahara, C., Chan, W., Saxena, S., Li, L., Whang, J., Denton, E.L., Ghasemipour, K., Gontijo Lopes, R., Karagol Ayan, B., Salimans, T., et al.: Photorealistic text-to-image diffusion models with deep language understanding. Advances in Neural Information Processing Systems (NeurIPS) (2022) +40. Sohl-Dickstein, J., Weiss, E., Maheswaranathan, N., Ganguli, S.: Deep unsupervised learning using nonequilibrium thermodynamics. In: International Conference on Machine Learning. PMLR (2015) +41. Song, J., Meng, C., Ermon, S.: Denoising diffusion implicit models. arXiv preprint arXiv:2010.02502 (2020) +42. Song, Y., Ermon, S.: Generative modeling by estimating gradients of the data distribution. Advances in neural information processing systems (NeurIPS) (2019) +43. Song, Y., Ermon, S.: Generative modeling by estimating gradients of the data distribution. Advances in neural information processing systems 32 (2019) +44. Song, Y., Ermon, S.: Improved techniques for training score-based generative models. Advances in neural information processing systems 33, 12438-12448 (2020) +45. Song, Y., Sohl-Dickstein, J.N., Kingma, D.P., Kumar, A., Ermon, S., Poole, B.: Score-based generative modeling through stochastic differential equations. ICLR (2021) +46. Tang, S., Zhang, F., Chen, J., Wang, P., Furukawa, Y.: Mvdiffusion: Enabling holistic multi-view image generation with correspondence-aware diffusion. ArXiv abs/2307.01097 (2023), https://api_semanticscholar.org/CorpusID: 259316427 +47. Teterwak, P., Sarna, A., Krishnan, D., Maschinot, A., Belanger, D., Liu, C., Freeman, W.T.: Boundless: Generative adversarial networks for image extension. 2019 IEEE/CVF International Conference on Computer Vision (ICCV) pp. 10520-10529 (2019), https://api_semanticscholar.org/CorpusID:201106503 +48. Van Den Oord, A., Vinyals, O., et al.: Neural discrete representation learning. Advances in neural information processing systems 30 (2017) +49. Voynov, A., Hertz, A., Arar, M., Fruchter, S., Cohen-Or, D.: Anylens: A generative diffusion model with any rendering lens (2023) +50. Wang, G., Yang, Y., Loy, C.C., Liu, Z.: Stylelight: Hdr panorama generation for lighting estimation and editing. In: European Conference on Computer Vision (2022), https://api_semanticscholar.org/CorpusID:251196614 +51. Wang, H., Xiang, X., Fan, Y., Xue, J.H.: Customizing 360-degree panoramicas through text-to-image diffusion models. ArXiv abs/2310.18840 (2023), https://api.sementicscholar.org/CorpusID:264590753 +52. Wu, T., Zheng, C., Cham, T.J.: Panodiffusion: 360-degree panorama outpainting via diffusion (2023), https://api-semanticscholar.org/CorpusID:259360663 +53. Zhang, L., Rao, A., Agrawala, M.: Adding conditional control to text-to-image diffusion models (2023) + +54. Zhang, Q., Song, J., Huang, X., Chen, Y., Liu, M.Y.: Diffcollage: Parallel generation of large content with diffusion models. 2023 IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR) pp. 10188-10198 (2023), https://api-semanticscholar.org/CorpusID:257834007 +55. Zhang, R., Isola, P., Efros, A.A., Shechtman, E., Wang, O.: The unreasonable effectiveness of deep features as a perceptual metric. 2018 IEEE/CVF Conference on Computer Vision and Pattern Recognition pp. 586-595 (2018), https://apisemantic scholar.org/CorpusID:4766599 \ No newline at end of file diff --git a/2024/PanoFree_ Tuning-Free Holistic Multi-view Image Generation with Cross-view Self-Guidance/images.zip b/2024/PanoFree_ Tuning-Free Holistic Multi-view Image Generation with Cross-view Self-Guidance/images.zip new file mode 100644 index 0000000000000000000000000000000000000000..91993f9183c7163796ad1851f7235739576745fb --- /dev/null +++ b/2024/PanoFree_ Tuning-Free Holistic Multi-view Image Generation with Cross-view Self-Guidance/images.zip @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:452797d0f43c8074ea1a8c24fe30c102f7506b5de40a151a14118933b1d3ab09 +size 729563 diff --git a/2024/PanoFree_ Tuning-Free Holistic Multi-view Image Generation with Cross-view Self-Guidance/layout.json b/2024/PanoFree_ Tuning-Free Holistic Multi-view Image Generation with Cross-view Self-Guidance/layout.json new file mode 100644 index 0000000000000000000000000000000000000000..fcaa491124700141a92e23e0ff47da163a5e9868 --- /dev/null +++ b/2024/PanoFree_ Tuning-Free Holistic Multi-view Image Generation with Cross-view Self-Guidance/layout.json @@ -0,0 +1,10067 @@ +{ + "pdf_info": [ + { + "para_blocks": [ + { + "bbox": [ + 134, + 111, + 479, + 147 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 134, + 111, + 479, + 147 + ], + "spans": [ + { + "bbox": [ + 134, + 111, + 479, + 147 + ], + "type": "text", + "content": "PanoFree: Tuning-Free Holistic Multi-view Image Generation with Cross-view Self-Guidance" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 134, + 167, + 479, + 193 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 134, + 167, + 479, + 193 + ], + "spans": [ + { + "bbox": [ + 134, + 167, + 479, + 193 + ], + "type": "text", + "content": "Aoming Liu" + }, + { + "bbox": [ + 134, + 167, + 479, + 193 + ], + "type": "inline_equation", + "content": "^{1,2*}" + }, + { + "bbox": [ + 134, + 167, + 479, + 193 + ], + "type": "text", + "content": ", Zhong Li" + }, + { + "bbox": [ + 134, + 167, + 479, + 193 + ], + "type": "inline_equation", + "content": "^{1\\dagger}" + }, + { + "bbox": [ + 134, + 167, + 479, + 193 + ], + "type": "text", + "content": ", Zhang Chen" + }, + { + "bbox": [ + 134, + 167, + 479, + 193 + ], + "type": "inline_equation", + "content": "^{1\\dagger}" + }, + { + "bbox": [ + 134, + 167, + 479, + 193 + ], + "type": "text", + "content": ", Nannan Li" + }, + { + "bbox": [ + 134, + 167, + 479, + 193 + ], + "type": "inline_equation", + "content": "^{2}" + }, + { + "bbox": [ + 134, + 167, + 479, + 193 + ], + "type": "text", + "content": ", Yi Xu" + }, + { + "bbox": [ + 134, + 167, + 479, + 193 + ], + "type": "inline_equation", + "content": "^{1}" + }, + { + "bbox": [ + 134, + 167, + 479, + 193 + ], + "type": "text", + "content": ", and Bryan A. Plummer" + }, + { + "bbox": [ + 134, + 167, + 479, + 193 + ], + "type": "inline_equation", + "content": "^{2}" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 187, + 201, + 425, + 224 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 187, + 201, + 425, + 224 + ], + "spans": [ + { + "bbox": [ + 187, + 201, + 425, + 224 + ], + "type": "inline_equation", + "content": "^{1}" + }, + { + "bbox": [ + 187, + 201, + 425, + 224 + ], + "type": "text", + "content": " OPPO US Research Center, Palo Alto, CA 94303, USA \n" + }, + { + "bbox": [ + 187, + 201, + 425, + 224 + ], + "type": "inline_equation", + "content": "^{2}" + }, + { + "bbox": [ + 187, + 201, + 425, + 224 + ], + "type": "text", + "content": " Boston University, Boston, MA 02215, USA" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 160, + 254, + 455, + 495 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 160, + 254, + 455, + 495 + ], + "spans": [ + { + "bbox": [ + 160, + 254, + 455, + 495 + ], + "type": "text", + "content": "Abstract. Immersive scene generation, notably panorama creation, benefits significantly from the adaptation of large pre-trained text-to-image (T2I) models for multi-view image generation. Due to the high cost of acquiring multi-view images, tuning-free generation is preferred. However, existing methods are either limited to simple correspondences or require extensive fine-tuning to capture complex ones. We present PanoFree, a novel method for tuning-free multi-view image generation that supports an extensive array of correspondences. PanoFree sequentially generates multi-view images using iterative warping and inpainting, addressing the key issues of inconsistency and artifacts from error accumulation without the need for fine-tuning. It improves error accumulation by enhancing cross-view awareness and refines the warping and inpainting processes via cross-view guidance, risky area estimation and erasing, and symmetric bidirectional guided generation for loop closure, alongside guidance-based semantic and density control for scene structure preservation. In experiments on Planar, " + }, + { + "bbox": [ + 160, + 254, + 455, + 495 + ], + "type": "inline_equation", + "content": "360^{\\circ}" + }, + { + "bbox": [ + 160, + 254, + 455, + 495 + ], + "type": "text", + "content": ", and Full Spherical Panoramas, PanoFree demonstrates significant error reduction, improves global consistency, and boosts image quality without extra fine-tuning. Compared to existing methods, PanoFree is up to 5x more efficient in time and 3x more efficient in GPU memory usage, and maintains superior diversity of results (2x better in our user study). PanoFree offers a viable alternative to costly fine-tuning or the use of additional pre-trained models." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 160, + 506, + 432, + 517 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 160, + 506, + 432, + 517 + ], + "spans": [ + { + "bbox": [ + 160, + 506, + 432, + 517 + ], + "type": "text", + "content": "Keywords: Tuning-free generation; Multi-view Image, Panorama" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 132, + 538, + 230, + 551 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 538, + 230, + 551 + ], + "spans": [ + { + "bbox": [ + 132, + 538, + 230, + 551 + ], + "type": "text", + "content": "1 Introduction" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 130, + 563, + 482, + 635 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 563, + 482, + 635 + ], + "spans": [ + { + "bbox": [ + 130, + 563, + 482, + 635 + ], + "type": "text", + "content": "Text-to-image (T2I) generation over multiple views for immersive scenes, like panorama generation, is a challenging task requiring coherence and diversity among many generated images (e.g., [3, 8, 12, 16, 17, 21, 36, 37, 39, 48]). Early work using GANs or VAEs (e.g., [4-6, 11, 24, 25, 47, 50]) have been replaced recently with diffusion-based models (e.g., [1, 9, 10, 22, 23, 46, 49, 51, 52, 54]), often leveraging Stable Diffusion [38]. State-of-the-art panorama generation methods" + } + ] + } + ], + "index": 6 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 145, + 642, + 481, + 653 + ], + "type": "page_footnote", + "angle": 0, + "lines": [ + { + "bbox": [ + 145, + 642, + 481, + 653 + ], + "spans": [ + { + "bbox": [ + 145, + 642, + 481, + 653 + ], + "type": "text", + "content": "*Work partly completed during Aoming's internship at OPPO US Research Center." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 147, + 653, + 249, + 665 + ], + "type": "page_footnote", + "angle": 0, + "lines": [ + { + "bbox": [ + 147, + 653, + 249, + 665 + ], + "spans": [ + { + "bbox": [ + 147, + 653, + 249, + 665 + ], + "type": "text", + "content": "†Corresponding Authors." + } + ] + } + ], + "index": 8 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 0 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 168, + 114, + 443, + 300 + ], + "blocks": [ + { + "bbox": [ + 168, + 114, + 443, + 300 + ], + "lines": [ + { + "bbox": [ + 168, + 114, + 443, + 300 + ], + "spans": [ + { + "bbox": [ + 168, + 114, + 443, + 300 + ], + "type": "image", + "image_path": "62107d5d06df8c37b569648ab5279266d0f49673f42d9a021305129cce9c2d81.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 130, + 304, + 482, + 350 + ], + "lines": [ + { + "bbox": [ + 130, + 304, + 482, + 350 + ], + "spans": [ + { + "bbox": [ + 130, + 304, + 482, + 350 + ], + "type": "text", + "content": "Fig. 1: PanoFree can generate multi-view images according to different types of correspondences without fine-tuning, and a natural application is tuning-free generation for different types of panoramas. We demonstrate this by generating three commonly used panoramas. Top: Planar Panorama; Middle: 360 Panorama; Bottom: Full Panorama." + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_caption" + } + ], + "index": 2 + }, + { + "bbox": [ + 130, + 360, + 482, + 421 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 360, + 482, + 421 + ], + "spans": [ + { + "bbox": [ + 130, + 360, + 482, + 421 + ], + "type": "text", + "content": "use Joint Diffusion (e.g., [1,22,46]), where parallel diffusion processes to generate multi-view images and enhancing global consistency by fusing latent or attention features based on cross-view correspondences. However, we find these methods can only generate some types of panoramas, limiting their scope (e.g., [1,22]), or require fine-tuning using expensive panorama datasets (e.g., [22,46])." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 130, + 423, + 482, + 590 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 423, + 482, + 590 + ], + "spans": [ + { + "bbox": [ + 130, + 423, + 482, + 590 + ], + "type": "text", + "content": "To address these challenges, we propose PanoFree a tuning-free multi-view image generation method using iterative warping and inpainting of perspective images to support diverse correspondences with low costs (see Fig. 1 for example generations). Iterative warping and inpainting of perspective images provide a means to the diverse multi-view correspondences required in panorama generation without additional fine-tuning [4, 11, 18]. However, recent work has overlooked these benefits due to accumulated errors causing suboptimal image quality [1, 22, 46]. We find that most accumulated errors from iterative warping can be attributed to the deficient conditions during generation. Specifically, conditioning solely on the previous image narrows cross-view awareness, leading to inconsistencies. Warping and inpainting can also propagate noise, e.g., truncated objects or jagged edges. Additionally, the given conditions may be incomplete to meet specific requirements, such as ensuring 360-degree consistency for loop closure and maintaining correct spatial relationships for realistic scenes." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 130, + 594, + 482, + 666 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 594, + 482, + 666 + ], + "spans": [ + { + "bbox": [ + 130, + 594, + 482, + 666 + ], + "type": "text", + "content": "To address these issues, PanoFree expand cross-view awareness by conditioning the current view on multiple views with guided image synthesis techniques such as SDEdit [29]. Then, PanoFree estimates and erases the risky areas, regions likely containing noise, to reduce the noise introduced by warping and inpainting. In addition, PanoFree adopts a bidirectional generation path with a symmetrical conditioning strategy for loop closure. Lastly, PanoFree further" + } + ] + } + ], + "index": 6 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "spans": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "type": "text", + "content": "2" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 166, + 91, + 217, + 100 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 166, + 91, + 217, + 100 + ], + "spans": [ + { + "bbox": [ + 166, + 91, + 217, + 100 + ], + "type": "text", + "content": "A. Liu et al." + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 1 + }, + { + "para_blocks": [ + { + "bbox": [ + 133, + 116, + 480, + 139 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 116, + 480, + 139 + ], + "spans": [ + { + "bbox": [ + 133, + 116, + 480, + 139 + ], + "type": "text", + "content": "utilizes pseudo global guidance with region-specific semantic and density control to make scene structure more reasonable." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 133, + 140, + 481, + 258 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 140, + 481, + 258 + ], + "spans": [ + { + "bbox": [ + 133, + 140, + 481, + 258 + ], + "type": "text", + "content": "We evaluate PanoFree on three text-to-panorama generation tasks: Planar, " + }, + { + "bbox": [ + 133, + 140, + 481, + 258 + ], + "type": "inline_equation", + "content": "360^{\\circ}" + }, + { + "bbox": [ + 133, + 140, + 481, + 258 + ], + "type": "text", + "content": ", and Full Spherical Panoramas. PanoFree effectively alleviating accumulated errors in sequential generation, and significantly improves image quality and global consistency (e.g. " + }, + { + "bbox": [ + 133, + 140, + 481, + 258 + ], + "type": "inline_equation", + "content": "31.6\\%" + }, + { + "bbox": [ + 133, + 140, + 481, + 258 + ], + "type": "text", + "content": " better in FID). This enables PanoFree to have better (or at least comparable) results to the state-of-the-art [1,22,46], despite these methods either having narrower applications or requiring fine-tuning datasets. Specifically, PanoFree is up to 5x more efficient in time and 3x more efficient in GPU memory usage, and maintains superior diversity of results (2x better in our user study). Lastly, PanoFree is also highly flexible, enabling it to plug-in-and-play with various pre-trained T2I models and adapters." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 147, + 261, + 362, + 270 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 147, + 261, + 362, + 270 + ], + "spans": [ + { + "bbox": [ + 147, + 261, + 362, + 270 + ], + "type": "text", + "content": "Our contributions can be summarized as follows:" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 132, + 273, + 485, + 392 + ], + "type": "list", + "angle": 0, + "index": 8, + "blocks": [ + { + "bbox": [ + 132, + 273, + 485, + 319 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 273, + 485, + 319 + ], + "spans": [ + { + "bbox": [ + 132, + 273, + 485, + 319 + ], + "type": "text", + "content": "- We introduce PanoFree, a tuning-free multi-view image generation method applicable for various correspondences and pre-trained T2I models. Thus PanoFree can greatly reduce the data and fine-tuning costs for immersive scene generation tasks such as text-to-panorama generation." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 132, + 320, + 480, + 366 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 320, + 480, + 366 + ], + "spans": [ + { + "bbox": [ + 132, + 320, + 480, + 366 + ], + "type": "text", + "content": "- We provide a in-depth perspective of accumulated errors and identify the deficient conditions as the main causes. We further effectively rectify deficient conditions and alleviate accumulated errors with the cross-view guidance as well as risky area estimation and erasing in PanoFree." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 132, + 368, + 480, + 392 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 368, + 480, + 392 + ], + "spans": [ + { + "bbox": [ + 132, + 368, + 480, + 392 + ], + "type": "text", + "content": "- As far as we know, PanoFree is the first to achieve feasible tuning-free generation for " + }, + { + "bbox": [ + 132, + 368, + 480, + 392 + ], + "type": "inline_equation", + "content": "360^{\\circ}" + }, + { + "bbox": [ + 132, + 368, + 480, + 392 + ], + "type": "text", + "content": " Panoramas and Full Spherical Panoramas." + } + ] + } + ], + "index": 7 + } + ], + "sub_type": "text" + }, + { + "bbox": [ + 133, + 411, + 236, + 423 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 411, + 236, + 423 + ], + "spans": [ + { + "bbox": [ + 133, + 411, + 236, + 423 + ], + "type": "text", + "content": "2 Related Work" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 133, + 437, + 481, + 498 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 437, + 481, + 498 + ], + "spans": [ + { + "bbox": [ + 133, + 437, + 481, + 498 + ], + "type": "text", + "content": "Diffusion Models [15, 19, 40, 42-45] are a popular framework for generative models. Early work required a long trajectory for sampling to produce high-quality samples [7, 45], before being sped up with advanced sampling techniques that also preserved generation quality [20, 26, 27, 41]. Latent Diffusion Models (LDMs) [33, 38] made these models more efficient by training in the latent space." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 133, + 498, + 481, + 616 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 498, + 481, + 616 + ], + "spans": [ + { + "bbox": [ + 133, + 498, + 481, + 616 + ], + "type": "text", + "content": "T2I Diffusion and Panorama Generation. Diffusion models are widely adopted for text-to-image (T2I) generation [31, 36, 38, 39]. Many downstream tasks used large pre-trained T2I diffusion models, like Stable Diffusion [38], to boost performance [4, 11, 18], including panorama generation [4, 11, 18]. These methods have largely supplanted GAN and VAE methods [4-6, 11, 24, 25, 32, 47, 50], with most recent work in panorama generation tasks using diffusion models [1, 9, 10, 22, 23, 46, 49, 51, 52, 54]. These diffusion-based panorama generation methods use joint diffusion to fuses multiple diffusion processes through latent or attention manipulation [1, 22, 46]. However, they are either limited to modeling simple correspondences or require extensive fine-tuning to model complex ones." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 133, + 617, + 481, + 665 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 617, + 481, + 665 + ], + "spans": [ + { + "bbox": [ + 133, + 617, + 481, + 665 + ], + "type": "text", + "content": "Guided Image Synthesis with Diffusion Models. It can be challenging to achieve satisfactory results solely relying on text guidance. Therefore, some prior work [28-30, 53] guide or control the generation results with reference images as fine-grained condition. ControlNet [53] and T2I-Adapter [30] are the most" + } + ] + } + ], + "index": 12 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 408, + 91, + 447, + 100 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 408, + 91, + 447, + 100 + ], + "spans": [ + { + "bbox": [ + 408, + 91, + 447, + 100 + ], + "type": "text", + "content": "PanoFree" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 474, + 92, + 480, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 474, + 92, + 480, + 100 + ], + "spans": [ + { + "bbox": [ + 474, + 92, + 480, + 100 + ], + "type": "text", + "content": "3" + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 2 + }, + { + "para_blocks": [ + { + "bbox": [ + 130, + 116, + 482, + 165 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 116, + 482, + 165 + ], + "spans": [ + { + "bbox": [ + 130, + 116, + 482, + 165 + ], + "type": "text", + "content": "commonly used methods to incorporate additional image conditions by adding extra image encoders, but they all require few-shot fine-tuning. SDEdit [29] achieves tuning-free guided image synthesis by adding noise to the guide image and then denoising it back to a real image using a pre-trained diffusion model." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 132, + 181, + 202, + 194 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 181, + 202, + 194 + ], + "spans": [ + { + "bbox": [ + 132, + 181, + 202, + 194 + ], + "type": "text", + "content": "3 Method" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 130, + 205, + 482, + 350 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 205, + 482, + 350 + ], + "spans": [ + { + "bbox": [ + 130, + 205, + 482, + 350 + ], + "type": "text", + "content": "PanoFree targets the text-to-panorama generation task, which takes textual descriptions as guidance to create multi-view perspective images that can be stitched into a wide-angle, high-quality panorama. PanoFree generates multiview images through sequential warping and inpainting steps, which typically results in acclimated errors due to deficit conditions (discussed in Sec. 3.1). Each component of Panofree is designed to minimize the effect of these various deficiencies. Specifically, Sec. 3.2 mitigates inconsistency using SDEdit-based cross-view guidance and Sec. 3.3 removes artifact-inducing content by estimating and erasing risky areas. At a higher level, PanoFree employs Bidirectional Generation with Symmetric Guidance for loop closure and error reduction (Sec. 3.4). Additionally, it applies guidance-based semantic and density control for scene structure preservation (Sec. 3.5). See Fig. 2 for an overview of our approach." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 132, + 364, + 407, + 376 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 364, + 407, + 376 + ], + "spans": [ + { + "bbox": [ + 132, + 364, + 407, + 376 + ], + "type": "text", + "content": "3.1 Deficient Conditions behind Accumulated Errors" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 130, + 382, + 482, + 455 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 382, + 482, + 455 + ], + "spans": [ + { + "bbox": [ + 130, + 382, + 482, + 455 + ], + "type": "text", + "content": "In this section, we discuss the causes behind the deficit conditions in the iterative warping and inpainting process and reformulate the problem as conditional generation. Given the text prompt " + }, + { + "bbox": [ + 130, + 382, + 482, + 455 + ], + "type": "inline_equation", + "content": "c_{t}" + }, + { + "bbox": [ + 130, + 382, + 482, + 455 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 130, + 382, + 482, + 455 + ], + "type": "inline_equation", + "content": "i" + }, + { + "bbox": [ + 130, + 382, + 482, + 455 + ], + "type": "text", + "content": "-th view's image " + }, + { + "bbox": [ + 130, + 382, + 482, + 455 + ], + "type": "inline_equation", + "content": "\\mathbf{x}_i" + }, + { + "bbox": [ + 130, + 382, + 482, + 455 + ], + "type": "text", + "content": ", warping function " + }, + { + "bbox": [ + 130, + 382, + 482, + 455 + ], + "type": "inline_equation", + "content": "\\mathcal{W}" + }, + { + "bbox": [ + 130, + 382, + 482, + 455 + ], + "type": "text", + "content": ", transformation matrix of the projection from " + }, + { + "bbox": [ + 130, + 382, + 482, + 455 + ], + "type": "inline_equation", + "content": "i" + }, + { + "bbox": [ + 130, + 382, + 482, + 455 + ], + "type": "text", + "content": "-th view to the " + }, + { + "bbox": [ + 130, + 382, + 482, + 455 + ], + "type": "inline_equation", + "content": "(i + 1)" + }, + { + "bbox": [ + 130, + 382, + 482, + 455 + ], + "type": "text", + "content": "-th view " + }, + { + "bbox": [ + 130, + 382, + 482, + 455 + ], + "type": "inline_equation", + "content": "\\mathbf{P}_i^{i + 1}" + }, + { + "bbox": [ + 130, + 382, + 482, + 455 + ], + "type": "text", + "content": ", and pre-trained T2I inpainting model " + }, + { + "bbox": [ + 130, + 382, + 482, + 455 + ], + "type": "inline_equation", + "content": "\\Phi_{inp}" + }, + { + "bbox": [ + 130, + 382, + 482, + 455 + ], + "type": "text", + "content": ", the warping and inpainting step to generate the " + }, + { + "bbox": [ + 130, + 382, + 482, + 455 + ], + "type": "inline_equation", + "content": "(i + 1)" + }, + { + "bbox": [ + 130, + 382, + 482, + 455 + ], + "type": "text", + "content": "-th view can be denoted as:" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 200, + 462, + 481, + 475 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 200, + 462, + 481, + 475 + ], + "spans": [ + { + "bbox": [ + 200, + 462, + 481, + 475 + ], + "type": "interline_equation", + "content": "\\hat {\\mathbf {x}} _ {i}, \\mathbf {m} _ {i} = \\mathcal {W} \\left(\\mathbf {x} _ {i}, \\mathbf {P} _ {i} ^ {i + 1}\\right); \\quad \\mathbf {x} _ {i + 1} = \\boldsymbol {\\Phi} _ {i n p} \\left(\\hat {\\mathbf {x}} _ {i}, \\mathbf {m} _ {i}, c _ {t}\\right), \\tag {1}", + "image_path": "2ef3171e9afe977fb8189835847330f219f75001ad2fa6f3a572c93982154123.jpg" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 130, + 482, + 482, + 519 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 482, + 482, + 519 + ], + "spans": [ + { + "bbox": [ + 130, + 482, + 482, + 519 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 130, + 482, + 482, + 519 + ], + "type": "inline_equation", + "content": "\\hat{\\mathbf{x}}_i" + }, + { + "bbox": [ + 130, + 482, + 482, + 519 + ], + "type": "text", + "content": " is the image warped from " + }, + { + "bbox": [ + 130, + 482, + 482, + 519 + ], + "type": "inline_equation", + "content": "i" + }, + { + "bbox": [ + 130, + 482, + 482, + 519 + ], + "type": "text", + "content": "-th view to " + }, + { + "bbox": [ + 130, + 482, + 482, + 519 + ], + "type": "inline_equation", + "content": "(i + 1)" + }, + { + "bbox": [ + 130, + 482, + 482, + 519 + ], + "type": "text", + "content": "-th view and " + }, + { + "bbox": [ + 130, + 482, + 482, + 519 + ], + "type": "inline_equation", + "content": "\\mathbf{m}_i" + }, + { + "bbox": [ + 130, + 482, + 482, + 519 + ], + "type": "text", + "content": " is the masking indicating the area to inpaint. And we can simplify the warping and inpainting steps in the following conditional image generation form:" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 249, + 525, + 481, + 539 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 249, + 525, + 481, + 539 + ], + "spans": [ + { + "bbox": [ + 249, + 525, + 481, + 539 + ], + "type": "interline_equation", + "content": "\\mathbf {x} _ {i + 1} \\sim q (\\mathbf {x} | c _ {t}, \\mathbf {x} _ {i}, \\mathbf {P} _ {i} ^ {i + 1}). \\tag {2}", + "image_path": "81bb3b74a168eef6c038b04b1ba85f4d99e46afe4cb07cff3d0f74f1cf134122.jpg" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 130, + 546, + 482, + 593 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 546, + 482, + 593 + ], + "spans": [ + { + "bbox": [ + 130, + 546, + 482, + 593 + ], + "type": "text", + "content": "However, during this generation process, we found conditions can become deficient. Major accumulated errors arise from three types of deficient conditions: Biased Conditions, Noisy Conditions, and Partial Conditions. See Sec. A.2 in the supplementary for detailed error illustrations." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 130, + 594, + 482, + 666 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 594, + 482, + 666 + ], + "spans": [ + { + "bbox": [ + 130, + 594, + 482, + 666 + ], + "type": "text", + "content": "Biased Conditions is the most obvious problem. In the above step, " + }, + { + "bbox": [ + 130, + 594, + 482, + 666 + ], + "type": "inline_equation", + "content": "\\mathbf{x}_{i + 1}" + }, + { + "bbox": [ + 130, + 594, + 482, + 666 + ], + "type": "text", + "content": " is solely conditioned on " + }, + { + "bbox": [ + 130, + 594, + 482, + 666 + ], + "type": "inline_equation", + "content": "\\mathbf{x}_i" + }, + { + "bbox": [ + 130, + 594, + 482, + 666 + ], + "type": "text", + "content": ", which biases the cross-view awareness heavily to " + }, + { + "bbox": [ + 130, + 594, + 482, + 666 + ], + "type": "inline_equation", + "content": "i" + }, + { + "bbox": [ + 130, + 594, + 482, + 666 + ], + "type": "text", + "content": "-th view. If " + }, + { + "bbox": [ + 130, + 594, + 482, + 666 + ], + "type": "inline_equation", + "content": "\\mathbf{x}_i" + }, + { + "bbox": [ + 130, + 594, + 482, + 666 + ], + "type": "text", + "content": " has deviated from the desired global distribution in certain aspects, then " + }, + { + "bbox": [ + 130, + 594, + 482, + 666 + ], + "type": "inline_equation", + "content": "\\mathbf{x}_{i + 1}" + }, + { + "bbox": [ + 130, + 594, + 482, + 666 + ], + "type": "text", + "content": " is likely to continue deviating in the same direction, resulting in significant inconsistency. We also found that slight style and content shifts often accumulate in this way, leading to significant inconsistency between distant views." + } + ] + } + ], + "index": 11 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "spans": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "type": "text", + "content": "4" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 166, + 91, + 217, + 101 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 166, + 91, + 217, + 101 + ], + "spans": [ + { + "bbox": [ + 166, + 91, + 217, + 101 + ], + "type": "text", + "content": "A. Liu et al." + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 3 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 164, + 114, + 443, + 205 + ], + "blocks": [ + { + "bbox": [ + 164, + 114, + 443, + 205 + ], + "lines": [ + { + "bbox": [ + 164, + 114, + 443, + 205 + ], + "spans": [ + { + "bbox": [ + 164, + 114, + 443, + 205 + ], + "type": "image", + "image_path": "f4ba846b44b812434a4f6c4d8a722475914a09d12b41c9f8cc42bacce32643ec.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 205, + 211, + 405, + 220 + ], + "lines": [ + { + "bbox": [ + 205, + 211, + 405, + 220 + ], + "spans": [ + { + "bbox": [ + 205, + 211, + 405, + 220 + ], + "type": "text", + "content": "(a) Overview of the generation and guidance framework." + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_caption" + } + ], + "index": 2 + }, + { + "type": "image", + "bbox": [ + 164, + 225, + 443, + 345 + ], + "blocks": [ + { + "bbox": [ + 164, + 225, + 443, + 345 + ], + "lines": [ + { + "bbox": [ + 164, + 225, + 443, + 345 + ], + "spans": [ + { + "bbox": [ + 164, + 225, + 443, + 345 + ], + "type": "image", + "image_path": "a6d11eb0e77c32c25a9be5c47e6eb0de6b4e27f38465edc050d87c2e447d6e48.jpg" + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 130, + 353, + 482, + 430 + ], + "lines": [ + { + "bbox": [ + 130, + 353, + 482, + 430 + ], + "spans": [ + { + "bbox": [ + 130, + 353, + 482, + 430 + ], + "type": "text", + "content": "Fig. 2: Overview of our PanoFree method, taking 360 Panorama Generation as an example. (a): At a framework level, PanoFree adopts two generation paths with opposite viewpoint translation or rotation. It enhances consistency by symmetrically selecting views from the other path as guidance to generate a new view (Sec. 3.4). Loop closure is ensured by merging these two paths. (b): In each warping and inpainting step, PanoFree reduces accumulated error by guiding the inpainting process with cross-view images (Sec. 3.2), along with estimating and erasing risky areas (Sec. 3.3)." + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "image_caption" + } + ], + "index": 4 + }, + { + "bbox": [ + 130, + 443, + 482, + 502 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 443, + 482, + 502 + ], + "spans": [ + { + "bbox": [ + 130, + 443, + 482, + 502 + ], + "type": "text", + "content": "Noisy Conditions mainly refer to " + }, + { + "bbox": [ + 130, + 443, + 482, + 502 + ], + "type": "inline_equation", + "content": "\\mathbf{x}_i" + }, + { + "bbox": [ + 130, + 443, + 482, + 502 + ], + "type": "text", + "content": " containing artifact-inducing contents. Existing artifacts in " + }, + { + "bbox": [ + 130, + 443, + 482, + 502 + ], + "type": "inline_equation", + "content": "\\mathbf{x}_i" + }, + { + "bbox": [ + 130, + 443, + 482, + 502 + ], + "type": "text", + "content": " could guide inpainting model to generate similar artifacts in " + }, + { + "bbox": [ + 130, + 443, + 482, + 502 + ], + "type": "inline_equation", + "content": "\\mathbf{x}_{i + 1}" + }, + { + "bbox": [ + 130, + 443, + 482, + 502 + ], + "type": "text", + "content": " and propagate to every following view. Additionally, disjointed or distorted areas, jagged or sharp content, and objects truncated by edges in " + }, + { + "bbox": [ + 130, + 443, + 482, + 502 + ], + "type": "inline_equation", + "content": "\\mathbf{x}_i" + }, + { + "bbox": [ + 130, + 443, + 482, + 502 + ], + "type": "text", + "content": " are also highly risky to introduce artifacts in " + }, + { + "bbox": [ + 130, + 443, + 482, + 502 + ], + "type": "inline_equation", + "content": "\\mathbf{x}_{i + 1}" + }, + { + "bbox": [ + 130, + 443, + 482, + 502 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 130, + 503, + 482, + 575 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 503, + 482, + 575 + ], + "spans": [ + { + "bbox": [ + 130, + 503, + 482, + 575 + ], + "type": "text", + "content": "Partial Conditions refer to the conditions not containing all the necessary information to meet specific requirements. For example, if we follow Eq. (2) on the final view, we lack information about " + }, + { + "bbox": [ + 130, + 503, + 482, + 575 + ], + "type": "inline_equation", + "content": "\\mathbf{x}_0" + }, + { + "bbox": [ + 130, + 503, + 482, + 575 + ], + "type": "text", + "content": ", making it impossible to generate image coherent with " + }, + { + "bbox": [ + 130, + 503, + 482, + 575 + ], + "type": "inline_equation", + "content": "\\mathbf{x}_0" + }, + { + "bbox": [ + 130, + 503, + 482, + 575 + ], + "type": "text", + "content": " to ensure loop closure. Additionally, using a single text prompt to generate all views within a full spherical panorama may lead to hallucinations, such as cities floating in the sky or underwater mountains." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 131, + 594, + 270, + 605 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 131, + 594, + 270, + 605 + ], + "spans": [ + { + "bbox": [ + 131, + 594, + 270, + 605 + ], + "type": "text", + "content": "3.2 Cross-View Guidance" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 131, + 615, + 481, + 640 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 131, + 615, + 481, + 640 + ], + "spans": [ + { + "bbox": [ + 131, + 615, + 481, + 640 + ], + "type": "text", + "content": "To rectify the biased conditions discussed in Sec.3.1 and enlarge cross-view awareness, a natural idea is to let " + }, + { + "bbox": [ + 131, + 615, + 481, + 640 + ], + "type": "inline_equation", + "content": "\\mathbf{x}_{i + 1}" + }, + { + "bbox": [ + 131, + 615, + 481, + 640 + ], + "type": "text", + "content": " conditioned on more views," + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 225, + 648, + 481, + 663 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 225, + 648, + 481, + 663 + ], + "spans": [ + { + "bbox": [ + 225, + 648, + 481, + 663 + ], + "type": "interline_equation", + "content": "\\mathbf {x} _ {i + 1} \\sim q (\\mathbf {x} | c _ {t}, \\mathbf {x} _ {i}, \\mathbf {P} _ {i} ^ {i + 1}, \\mathbf {x} _ {1} ^ {g}, \\dots , \\mathbf {x} _ {m} ^ {g}) \\qquad (3)", + "image_path": "f9437b6806c8793baccf76bd160d17ad4e9b4968fbf83f64af1036ef3045fe4b.jpg" + } + ] + } + ], + "index": 10 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 408, + 91, + 447, + 100 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 408, + 91, + 447, + 100 + ], + "spans": [ + { + "bbox": [ + 408, + 91, + 447, + 100 + ], + "type": "text", + "content": "PanoFree" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 474, + 92, + 480, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 474, + 92, + 480, + 100 + ], + "spans": [ + { + "bbox": [ + 474, + 92, + 480, + 100 + ], + "type": "text", + "content": "5" + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 4 + }, + { + "para_blocks": [ + { + "bbox": [ + 130, + 116, + 479, + 187 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 116, + 479, + 187 + ], + "spans": [ + { + "bbox": [ + 130, + 116, + 479, + 187 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 130, + 116, + 479, + 187 + ], + "type": "inline_equation", + "content": "\\mathbf{x}_1^g, \\ldots, \\mathbf{x}_m^g" + }, + { + "bbox": [ + 130, + 116, + 479, + 187 + ], + "type": "text", + "content": " are selected from " + }, + { + "bbox": [ + 130, + 116, + 479, + 187 + ], + "type": "inline_equation", + "content": "\\mathbf{x}_0, \\ldots, \\mathbf{x}_{i-1}" + }, + { + "bbox": [ + 130, + 116, + 479, + 187 + ], + "type": "text", + "content": ". This naturally results in a Guided Image Synthesis task form with self-generated images as guidance. Many existing methods can be adapted to implement our design, such as ControlNet [53] and T2I adapter [30]. To avoid relying on fine-tuning and reduce memory and time costs, we choose SDEdit [29], a training-free guided image synthesis approach, with a single guidance image " + }, + { + "bbox": [ + 130, + 116, + 479, + 187 + ], + "type": "inline_equation", + "content": "\\mathbf{x}^g \\in [\\mathbf{x}_0, \\ldots, \\mathbf{x}_{i-1}]" + }, + { + "bbox": [ + 130, + 116, + 479, + 187 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 130, + 188, + 479, + 236 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 188, + 479, + 236 + ], + "spans": [ + { + "bbox": [ + 130, + 188, + 479, + 236 + ], + "type": "text", + "content": "Guided Image Synthesis using SDEdit. Given " + }, + { + "bbox": [ + 130, + 188, + 479, + 236 + ], + "type": "inline_equation", + "content": "\\mathbf{x}^g" + }, + { + "bbox": [ + 130, + 188, + 479, + 236 + ], + "type": "text", + "content": " as guidance, SDEdit establishes a Gaussian distribution using " + }, + { + "bbox": [ + 130, + 188, + 479, + 236 + ], + "type": "inline_equation", + "content": "\\mathbf{x}^g" + }, + { + "bbox": [ + 130, + 188, + 479, + 236 + ], + "type": "text", + "content": " as the expectation and the intermediate status at time " + }, + { + "bbox": [ + 130, + 188, + 479, + 236 + ], + "type": "inline_equation", + "content": "t_0" + }, + { + "bbox": [ + 130, + 188, + 479, + 236 + ], + "type": "text", + "content": " in the reverse SDE process. The desired data distribution is obtained by gradually removing noise from " + }, + { + "bbox": [ + 130, + 188, + 479, + 236 + ], + "type": "inline_equation", + "content": "\\mathbf{x}^g(t_0)" + }, + { + "bbox": [ + 130, + 188, + 479, + 236 + ], + "type": "text", + "content": ":" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 201, + 244, + 480, + 258 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 201, + 244, + 480, + 258 + ], + "spans": [ + { + "bbox": [ + 201, + 244, + 480, + 258 + ], + "type": "interline_equation", + "content": "\\mathbf {x} ^ {g} \\left(t _ {0}\\right) \\sim \\mathcal {N} \\left(\\mathbf {x} ^ {g}, \\sigma^ {2} \\left(t _ {0}\\right) \\mathbf {I}\\right); \\quad \\mathbf {x} \\sim \\operatorname {S D E d i t} \\left(\\mathbf {x} ^ {g}, t _ {0}, \\boldsymbol {\\Phi}\\right), \\tag {4}", + "image_path": "f9693938ed65ce6a2b5244693f78d93c40d78e34861e1128f7df4e9506404df4.jpg" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 131, + 265, + 479, + 300 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 131, + 265, + 479, + 300 + ], + "spans": [ + { + "bbox": [ + 131, + 265, + 479, + 300 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 131, + 265, + 479, + 300 + ], + "type": "inline_equation", + "content": "\\Phi" + }, + { + "bbox": [ + 131, + 265, + 479, + 300 + ], + "type": "text", + "content": " denotes a generative model. In PanoFree, we use inpainting mask " + }, + { + "bbox": [ + 131, + 265, + 479, + 300 + ], + "type": "inline_equation", + "content": "\\mathbf{m}_i" + }, + { + "bbox": [ + 131, + 265, + 479, + 300 + ], + "type": "text", + "content": " to paste the guidance image " + }, + { + "bbox": [ + 131, + 265, + 479, + 300 + ], + "type": "inline_equation", + "content": "\\mathbf{x}^g" + }, + { + "bbox": [ + 131, + 265, + 479, + 300 + ], + "type": "text", + "content": " to the blank areas in the warped image " + }, + { + "bbox": [ + 131, + 265, + 479, + 300 + ], + "type": "inline_equation", + "content": "\\hat{\\mathbf{x}}_i" + }, + { + "bbox": [ + 131, + 265, + 479, + 300 + ], + "type": "text", + "content": " and then use SDEdit in the inpainting process:" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 183, + 309, + 480, + 323 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 183, + 309, + 480, + 323 + ], + "spans": [ + { + "bbox": [ + 183, + 309, + 480, + 323 + ], + "type": "interline_equation", + "content": "\\hat {\\mathbf {x}} _ {i} ^ {g} = \\mathbf {m} _ {i} \\cdot \\mathbf {x} ^ {g} + (1 - \\mathbf {m} _ {i}) \\cdot \\hat {\\mathbf {x}} _ {i}; \\quad \\mathbf {x} _ {i} \\sim \\operatorname {S D E d i t} \\left(\\hat {\\mathbf {x}} _ {i} ^ {g}, t _ {0}, \\boldsymbol {\\Phi} _ {i n p}\\right). \\tag {5}", + "image_path": "5b23bde529347ccf42f5cd58e84e8e380a7e02d811b8358a85ae5782f02a46af.jpg" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 130, + 330, + 479, + 413 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 330, + 479, + 413 + ], + "spans": [ + { + "bbox": [ + 130, + 330, + 479, + 413 + ], + "type": "text", + "content": "Since we only want to use additional guidance images to rectify the biased conditions rather than replicate the guidance image, we use " + }, + { + "bbox": [ + 130, + 330, + 479, + 413 + ], + "type": "inline_equation", + "content": "t_0 \\in [0.9, 1.0)" + }, + { + "bbox": [ + 130, + 330, + 479, + 413 + ], + "type": "text", + "content": " in practice. Meanwhile, we found that different selection of generation path and guidance image results in different generation qualities, and the optimal choice may vary for different tasks. We introduce a general selection effective for various tasks in Section 3.4 and provide an example of extending this technique to make scenes more realistic in specific scenarios in Section 3.5." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 131, + 431, + 339, + 443 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 131, + 431, + 339, + 443 + ], + "spans": [ + { + "bbox": [ + 131, + 431, + 339, + 443 + ], + "type": "text", + "content": "3.3 Risky Area Estimation and Erasing" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 130, + 449, + 479, + 533 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 449, + 479, + 533 + ], + "spans": [ + { + "bbox": [ + 130, + 449, + 479, + 533 + ], + "type": "text", + "content": "To rectify the noisy conditions discussed in Sec.3.1 and eliminate the accumulation of artifacts, a natural idea is to detect and localize the artifact-inducing contents, and erase them. However, precise detection and localization often requires costly training. Thus, we turn to roughly estimate and erase the risky areas that are likely to contain artifact-inducing contents, based on indicators often associated with artifact: distances, color and smoothness. See Sec. A.2 in the supplementary for examples." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 130, + 533, + 480, + 640 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 533, + 480, + 640 + ], + "spans": [ + { + "bbox": [ + 130, + 533, + 480, + 640 + ], + "type": "text", + "content": "Risk Estimation based on Distances. We consider the distance from the center point of the initial view " + }, + { + "bbox": [ + 130, + 533, + 480, + 640 + ], + "type": "inline_equation", + "content": "\\mathbf{x}_0" + }, + { + "bbox": [ + 130, + 533, + 480, + 640 + ], + "type": "text", + "content": " and the distance to the edges. This is based on two priors: 1. The farther from the initial view, the more accumulated errors and the more likely to contain artifact-inducing contents. 2. Areas close to the edges are highly risky because truncated objects are mostly generated around the edges, and areas near the edges are often more severely distorted during warping. We use initial risk " + }, + { + "bbox": [ + 130, + 533, + 480, + 640 + ], + "type": "inline_equation", + "content": "\\mathbf{r}_{init}" + }, + { + "bbox": [ + 130, + 533, + 480, + 640 + ], + "type": "text", + "content": " to represent the risk estimated based on the distance from the initial view, and edge risk " + }, + { + "bbox": [ + 130, + 533, + 480, + 640 + ], + "type": "inline_equation", + "content": "\\mathbf{r}_{edge}" + }, + { + "bbox": [ + 130, + 533, + 480, + 640 + ], + "type": "text", + "content": " to represent the risk estimated based on the distance from edges. They are derived from the following:" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 201, + 650, + 480, + 663 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 201, + 650, + 480, + 663 + ], + "spans": [ + { + "bbox": [ + 201, + 650, + 480, + 663 + ], + "type": "interline_equation", + "content": "\\mathbf {r} _ {\\text {i n i t}} \\left(\\mathbf {c} _ {i}\\right) = \\mathcal {R} _ {p} \\left(\\mathcal {D} _ {0} \\left(\\mathbf {c} _ {i}\\right)\\right); \\quad \\mathbf {r} _ {\\text {e d g e}} \\left(\\mathbf {c} _ {i}\\right) = \\mathcal {R} _ {p} \\left(\\mathcal {D} _ {\\mathbf {e}} \\left(\\mathbf {c} _ {i}\\right)\\right), \\tag {6}", + "image_path": "8f324e0f97490adb7a32362aee04b9d47ca0828e56399f6de388342e89cd2aff.jpg" + } + ] + } + ], + "index": 11 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "spans": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "type": "text", + "content": "6" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 166, + 91, + 217, + 100 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 166, + 91, + 217, + 100 + ], + "spans": [ + { + "bbox": [ + 166, + 91, + 217, + 100 + ], + "type": "text", + "content": "A. Liu et al." + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 5 + }, + { + "para_blocks": [ + { + "bbox": [ + 130, + 116, + 479, + 176 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 116, + 479, + 176 + ], + "spans": [ + { + "bbox": [ + 130, + 116, + 479, + 176 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 130, + 116, + 479, + 176 + ], + "type": "inline_equation", + "content": "\\mathbf{c}_i" + }, + { + "bbox": [ + 130, + 116, + 479, + 176 + ], + "type": "text", + "content": " represents the pixel coordinates of " + }, + { + "bbox": [ + 130, + 116, + 479, + 176 + ], + "type": "inline_equation", + "content": "\\mathbf{x}_i" + }, + { + "bbox": [ + 130, + 116, + 479, + 176 + ], + "type": "text", + "content": " within the panorama coordinate system, " + }, + { + "bbox": [ + 130, + 116, + 479, + 176 + ], + "type": "inline_equation", + "content": "\\mathcal{D}_0" + }, + { + "bbox": [ + 130, + 116, + 479, + 176 + ], + "type": "text", + "content": " measures the distance to the center point of the initial view along the generation path, " + }, + { + "bbox": [ + 130, + 116, + 479, + 176 + ], + "type": "inline_equation", + "content": "\\mathcal{D}_{\\mathbf{e}}" + }, + { + "bbox": [ + 130, + 116, + 479, + 176 + ], + "type": "text", + "content": " measures the distance to all edges " + }, + { + "bbox": [ + 130, + 116, + 479, + 176 + ], + "type": "inline_equation", + "content": "\\mathbf{e}" + }, + { + "bbox": [ + 130, + 116, + 479, + 176 + ], + "type": "text", + "content": ", and " + }, + { + "bbox": [ + 130, + 116, + 479, + 176 + ], + "type": "inline_equation", + "content": "\\mathcal{R}_p" + }, + { + "bbox": [ + 130, + 116, + 479, + 176 + ], + "type": "text", + "content": " is a scaling function. We use weighted euclidean distance for " + }, + { + "bbox": [ + 130, + 116, + 479, + 176 + ], + "type": "inline_equation", + "content": "\\mathcal{D}_0" + }, + { + "bbox": [ + 130, + 116, + 479, + 176 + ], + "type": "text", + "content": ", Gaussian filters for " + }, + { + "bbox": [ + 130, + 116, + 479, + 176 + ], + "type": "inline_equation", + "content": "\\mathcal{D}_{\\mathbf{e}}" + }, + { + "bbox": [ + 130, + 116, + 479, + 176 + ], + "type": "text", + "content": ", and min-max normalization for " + }, + { + "bbox": [ + 130, + 116, + 479, + 176 + ], + "type": "inline_equation", + "content": "\\mathcal{R}_p" + }, + { + "bbox": [ + 130, + 116, + 479, + 176 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 130, + 175, + 479, + 236 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 175, + 479, + 236 + ], + "spans": [ + { + "bbox": [ + 130, + 175, + 479, + 236 + ], + "type": "text", + "content": "Risk Estimation based on Color and Smoothness. After generating a view, we can predict the risk based on color and smoothness. This uses two priors: 1. Artifacts are often not smooth or distinct in color. 2. Salient areas with abrupt colors or unevenness are prone to causing artifacts. Color-based risk " + }, + { + "bbox": [ + 130, + 175, + 479, + 236 + ], + "type": "inline_equation", + "content": "\\mathbf{r}_{color}" + }, + { + "bbox": [ + 130, + 175, + 479, + 236 + ], + "type": "text", + "content": " and smoothness-based risk " + }, + { + "bbox": [ + 130, + 175, + 479, + 236 + ], + "type": "inline_equation", + "content": "\\mathbf{r}_{smooth}" + }, + { + "bbox": [ + 130, + 175, + 479, + 236 + ], + "type": "text", + "content": " are estimated in similar forms:" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 192, + 246, + 480, + 258 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 192, + 246, + 480, + 258 + ], + "spans": [ + { + "bbox": [ + 192, + 246, + 480, + 258 + ], + "type": "interline_equation", + "content": "\\mathbf {r} _ {\\text {c o l o r}} \\left(\\mathbf {x} _ {i}\\right) = \\mathcal {R} _ {f} \\left(\\mathcal {D} _ {c} \\left(\\mathbf {x} _ {i}\\right)\\right); \\quad \\mathbf {r} _ {\\text {s m o o t h}} \\left(\\mathbf {x} _ {i}\\right) = \\mathcal {R} _ {f} \\left(\\mathcal {D} _ {s} \\left(\\mathbf {x} _ {i}\\right)\\right), \\tag {7}", + "image_path": "1993e2ac57faf690990489a63aab9f0c636cf5f8906423fd011364cbe3fd1cf2.jpg" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 130, + 266, + 479, + 324 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 266, + 479, + 324 + ], + "spans": [ + { + "bbox": [ + 130, + 266, + 479, + 324 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 130, + 266, + 479, + 324 + ], + "type": "inline_equation", + "content": "\\mathcal{D}_c" + }, + { + "bbox": [ + 130, + 266, + 479, + 324 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 130, + 266, + 479, + 324 + ], + "type": "inline_equation", + "content": "\\mathcal{D}_s" + }, + { + "bbox": [ + 130, + 266, + 479, + 324 + ], + "type": "text", + "content": " measures the abruptness of each pixel based on color and smoothness. When implementing them, we choose pixels with the same vertical coordinates across views, and calculate the \"distances\" of each pixel to the mean color and color gradient. Within " + }, + { + "bbox": [ + 130, + 266, + 479, + 324 + ], + "type": "inline_equation", + "content": "\\mathcal{R}_f" + }, + { + "bbox": [ + 130, + 266, + 479, + 324 + ], + "type": "text", + "content": ", we applied Gaussian filtering after min-max normalization, as those estimated risks are usually noisy." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 130, + 326, + 480, + 384 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 326, + 480, + 384 + ], + "spans": [ + { + "bbox": [ + 130, + 326, + 480, + 384 + ], + "type": "text", + "content": "Erasing with Estimated Risks. With the estimated risks, we can erase the risky areas on the image warped to next view and the inpainting mask. Assume that we get inpainting mask for current view " + }, + { + "bbox": [ + 130, + 326, + 480, + 384 + ], + "type": "inline_equation", + "content": "\\mathbf{m}_i" + }, + { + "bbox": [ + 130, + 326, + 480, + 384 + ], + "type": "text", + "content": " and risks for previous view " + }, + { + "bbox": [ + 130, + 326, + 480, + 384 + ], + "type": "inline_equation", + "content": "\\mathbf{r}_{i - 1} = [\\mathbf{r}_{i - 1}^i,\\mathbf{r}_{i - 1}^e,\\mathbf{r}_{i - 1}^c,\\mathbf{r}_{i - 1}^s]" + }, + { + "bbox": [ + 130, + 326, + 480, + 384 + ], + "type": "text", + "content": ". The risks are combined linearly and new inpainting mask for current view can be obtained with:" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 233, + 395, + 480, + 407 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 233, + 395, + 480, + 407 + ], + "spans": [ + { + "bbox": [ + 233, + 395, + 480, + 407 + ], + "type": "interline_equation", + "content": "\\mathbf {m} _ {i} ^ {r} = \\mathcal {M} _ {r} \\left(\\mathbf {m} _ {i}, \\mathcal {W} \\left(\\mathbf {r} _ {i - 1} \\cdot \\mathbf {w}, \\mathbf {P} _ {i - 1} ^ {i}\\right)\\right). \\tag {8}", + "image_path": "9e4a11e777ac39587097e98a5e40beda54fc676b261b3c4f9775576cdfc07eca.jpg" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 130, + 416, + 479, + 440 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 416, + 479, + 440 + ], + "spans": [ + { + "bbox": [ + 130, + 416, + 479, + 440 + ], + "type": "inline_equation", + "content": "\\mathcal{M}_r" + }, + { + "bbox": [ + 130, + 416, + 479, + 440 + ], + "type": "text", + "content": " is the risk-based remasking function, and " + }, + { + "bbox": [ + 130, + 416, + 479, + 440 + ], + "type": "inline_equation", + "content": "\\mathbf{w}" + }, + { + "bbox": [ + 130, + 416, + 479, + 440 + ], + "type": "text", + "content": " are user defined combination weights. We define " + }, + { + "bbox": [ + 130, + 416, + 479, + 440 + ], + "type": "inline_equation", + "content": "\\mathcal{M}_r" + }, + { + "bbox": [ + 130, + 416, + 479, + 440 + ], + "type": "text", + "content": " as thresholding the risk within the warped area." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 130, + 440, + 480, + 523 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 440, + 480, + 523 + ], + "spans": [ + { + "bbox": [ + 130, + 440, + 480, + 523 + ], + "type": "text", + "content": "Smoothing and Anti-aliasing. We note that the inpainting mask from risk-based erasing may not be smooth. Additionally, sharp and jagged edges on the inpainting mask can lead to artifacts. Therefore, we also employ fixed filtering " + }, + { + "bbox": [ + 130, + 440, + 480, + 523 + ], + "type": "inline_equation", + "content": "\\mathcal{M}_f" + }, + { + "bbox": [ + 130, + 440, + 480, + 523 + ], + "type": "text", + "content": ", where Gaussian filtering and thresholding are used to smooth the mask and reduce sharp edges, while median filtering is used to reduce jagged edges. Then, we use the final inpainting mask for the combination with guidance, and the risky areas on the warped image are removed and regenerated." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 204, + 532, + 480, + 545 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 204, + 532, + 480, + 545 + ], + "spans": [ + { + "bbox": [ + 204, + 532, + 480, + 545 + ], + "type": "interline_equation", + "content": "\\mathbf {m} _ {i} ^ {f} = \\mathcal {M} _ {f} \\left(\\mathbf {m} _ {i} ^ {r}\\right); \\quad \\hat {\\mathbf {x}} _ {i} ^ {g} = \\mathbf {m} _ {i} ^ {f} \\cdot \\mathbf {x} ^ {g} + \\left(1 - \\mathbf {m} _ {i} ^ {f}\\right) \\cdot \\hat {\\mathbf {x}} _ {i}. \\tag {9}", + "image_path": "46dc6661159e43b843c2791a4a73bda6945b4a58041989d96ffa021a96604257.jpg" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 131, + 562, + 421, + 574 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 131, + 562, + 421, + 574 + ], + "spans": [ + { + "bbox": [ + 131, + 562, + 421, + 574 + ], + "type": "text", + "content": "3.4 Bidirectional Generation with Symmetric Guidance" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 130, + 581, + 479, + 665 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 581, + 479, + 665 + ], + "spans": [ + { + "bbox": [ + 130, + 581, + 479, + 665 + ], + "type": "text", + "content": "Bidirectional Generation. We begin by dividing a unidirectional generation path " + }, + { + "bbox": [ + 130, + 581, + 479, + 665 + ], + "type": "inline_equation", + "content": "\\mathbf{x}_0 \\to \\mathbf{x}_1 \\dots \\mathbf{x}_{2n}" + }, + { + "bbox": [ + 130, + 581, + 479, + 665 + ], + "type": "text", + "content": " into two bidirectional generation paths " + }, + { + "bbox": [ + 130, + 581, + 479, + 665 + ], + "type": "inline_equation", + "content": "\\mathbf{x}_0 \\to \\mathbf{x}_1 \\to \\dots \\to \\mathbf{x}_n" + }, + { + "bbox": [ + 130, + 581, + 479, + 665 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 130, + 581, + 479, + 665 + ], + "type": "inline_equation", + "content": "\\mathbf{x}_{-n} \\gets \\dots \\gets \\mathbf{x}_{-1} \\gets \\mathbf{x}_0" + }, + { + "bbox": [ + 130, + 581, + 479, + 665 + ], + "type": "text", + "content": ". Typically, we would make these two generation paths symmetric. And we found this can reduce accumulated errors because the distance to the initial view is reduced in each direction. This consistently reduces artifacts, but may not reduce style and content inconsistency, as there may be different style/content shift in the two directions." + } + ] + } + ], + "index": 12 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 408, + 91, + 447, + 100 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 408, + 91, + 447, + 100 + ], + "spans": [ + { + "bbox": [ + 408, + 91, + 447, + 100 + ], + "type": "text", + "content": "PanoFree" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 474, + 92, + 481, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 474, + 92, + 481, + 100 + ], + "spans": [ + { + "bbox": [ + 474, + 92, + 481, + 100 + ], + "type": "text", + "content": "7" + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 6 + }, + { + "para_blocks": [ + { + "bbox": [ + 130, + 116, + 479, + 200 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 116, + 479, + 200 + ], + "spans": [ + { + "bbox": [ + 130, + 116, + 479, + 200 + ], + "type": "text", + "content": "Loop Closure. To ensure loop closure, we can add a " + }, + { + "bbox": [ + 130, + 116, + 479, + 200 + ], + "type": "inline_equation", + "content": "(2n + 1)" + }, + { + "bbox": [ + 130, + 116, + 479, + 200 + ], + "type": "text", + "content": "-th view as the \"merging view\" to merge the 2 generation paths by warping " + }, + { + "bbox": [ + 130, + 116, + 479, + 200 + ], + "type": "inline_equation", + "content": "\\mathbf{x}_n" + }, + { + "bbox": [ + 130, + 116, + 479, + 200 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 130, + 116, + 479, + 200 + ], + "type": "inline_equation", + "content": "\\mathbf{x}_{-n}" + }, + { + "bbox": [ + 130, + 116, + 479, + 200 + ], + "type": "text", + "content": " to the " + }, + { + "bbox": [ + 130, + 116, + 479, + 200 + ], + "type": "inline_equation", + "content": "(2n + 1)" + }, + { + "bbox": [ + 130, + 116, + 479, + 200 + ], + "type": "text", + "content": "-th view and inpaint it. However, if the differences between the two paths are too large, " + }, + { + "bbox": [ + 130, + 116, + 479, + 200 + ], + "type": "inline_equation", + "content": "\\mathbf{x}_{2n + 1}" + }, + { + "bbox": [ + 130, + 116, + 479, + 200 + ], + "type": "text", + "content": " may contain image tearing, failing to ensure loop closure. This is due to the partial conditions on each path: there is no information from the other path before merging. Therefore, we rectify the partial conditions by introducing awareness of the other path." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 130, + 200, + 479, + 248 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 200, + 479, + 248 + ], + "spans": [ + { + "bbox": [ + 130, + 200, + 479, + 248 + ], + "type": "text", + "content": "Symmetric Guidance. We introduce the awareness of the other path by selecting symmetric guidance images from the other path. Specifically, when generating " + }, + { + "bbox": [ + 130, + 200, + 479, + 248 + ], + "type": "inline_equation", + "content": "\\mathbf{x}_{i + 1}" + }, + { + "bbox": [ + 130, + 200, + 479, + 248 + ], + "type": "text", + "content": ", we will select " + }, + { + "bbox": [ + 130, + 200, + 479, + 248 + ], + "type": "inline_equation", + "content": "\\mathbf{x}_{-i}" + }, + { + "bbox": [ + 130, + 200, + 479, + 248 + ], + "type": "text", + "content": " as the guidance image. Thus, " + }, + { + "bbox": [ + 130, + 200, + 479, + 248 + ], + "type": "inline_equation", + "content": "\\mathbf{x}_{i + 1}" + }, + { + "bbox": [ + 130, + 200, + 479, + 248 + ], + "type": "text", + "content": " will get the awareness of both paths as it is conditioned on " + }, + { + "bbox": [ + 130, + 200, + 479, + 248 + ], + "type": "inline_equation", + "content": "\\mathbf{x}_i" + }, + { + "bbox": [ + 130, + 200, + 479, + 248 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 130, + 200, + 479, + 248 + ], + "type": "inline_equation", + "content": "\\mathbf{x}_{-i}" + }, + { + "bbox": [ + 130, + 200, + 479, + 248 + ], + "type": "text", + "content": ":" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 241, + 254, + 480, + 268 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 241, + 254, + 480, + 268 + ], + "spans": [ + { + "bbox": [ + 241, + 254, + 480, + 268 + ], + "type": "interline_equation", + "content": "\\mathbf {x} _ {i + 1} \\sim q (\\mathbf {x} \\mid c _ {t}, \\mathbf {x} _ {i}, \\mathbf {x} _ {- i}, \\mathbf {P} _ {i} ^ {i + 1}) \\tag {10}", + "image_path": "24509d94246343a080fc3cb9546be0de0f1de748bab5167ecf96dbcacdc7ac7c.jpg" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 130, + 274, + 479, + 309 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 274, + 479, + 309 + ], + "spans": [ + { + "bbox": [ + 130, + 274, + 479, + 309 + ], + "type": "text", + "content": "We empirically found that bidirectional generation with symmetric guidance is not only effective in ensuring loop closure but also a universally applicable strategy to effectively reduce accumulated errors in various scenarios." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 131, + 325, + 343, + 336 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 131, + 325, + 343, + 336 + ], + "spans": [ + { + "bbox": [ + 131, + 325, + 343, + 336 + ], + "type": "text", + "content": "3.5 Aligning with Scene Structure Prior" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 130, + 342, + 479, + 411 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 342, + 479, + 411 + ], + "spans": [ + { + "bbox": [ + 130, + 342, + 479, + 411 + ], + "type": "text", + "content": "When generating full spherical panoramas, we divide a spherical panorama into five parts: first, we generate a 360 panorama as the central part, then we expand upwards and downwards, and finally, we generate two images centered around the top and bottom poles to close up the entire spherical surface. During the expansion and closing stages, models often fail to align with scene structure priors due to partial conditions and generate artifacts." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 130, + 414, + 479, + 496 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 414, + 479, + 496 + ], + "spans": [ + { + "bbox": [ + 130, + 414, + 479, + 496 + ], + "type": "text", + "content": "Hallucination refers to the artifacts caused by mismatches between partial conditions and scene structure priors. For example, when generating a city scene, using the same prompt during the expansion and closing stages may result in a floating city in the sky or a city underwater. The most direct solution is to input a new prompt, but this would require additional manual effort, which is not ideal. So, we attempt to rectify the partial conditions by extracting scene structure priors and applying semantic and variance control from the initial view." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 130, + 498, + 479, + 568 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 498, + 479, + 568 + ], + "spans": [ + { + "bbox": [ + 130, + 498, + 479, + 568 + ], + "type": "text", + "content": "Prior Extraction. Although the pretrained T2I model may not align a full panorama with scene structure priors, it can align a single perspective view image with them. Therefore, we extract the scene structure prior from the initial view image " + }, + { + "bbox": [ + 130, + 498, + 479, + 568 + ], + "type": "inline_equation", + "content": "\\mathbf{x}_0" + }, + { + "bbox": [ + 130, + 498, + 479, + 568 + ], + "type": "text", + "content": " and incorporate it into the expansion process. For example, when generating the first view image in the upward expansion " + }, + { + "bbox": [ + 130, + 498, + 479, + 568 + ], + "type": "inline_equation", + "content": "\\mathbf{x}_0^{ue}" + }, + { + "bbox": [ + 130, + 498, + 479, + 568 + ], + "type": "text", + "content": ", we use upper " + }, + { + "bbox": [ + 130, + 498, + 479, + 568 + ], + "type": "inline_equation", + "content": "1/3" + }, + { + "bbox": [ + 130, + 498, + 479, + 568 + ], + "type": "text", + "content": " part of the initial view image as guidance with resizing it to the size of " + }, + { + "bbox": [ + 130, + 498, + 479, + 568 + ], + "type": "inline_equation", + "content": "\\mathbf{x}_0^{ue}" + }, + { + "bbox": [ + 130, + 498, + 479, + 568 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 130, + 570, + 479, + 663 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 570, + 479, + 663 + ], + "spans": [ + { + "bbox": [ + 130, + 570, + 479, + 663 + ], + "type": "text", + "content": "Semantic and Variance Tuning. When the give text prompt only describe part of the scene, we may want generated semantic contents less conditioned on the partial prompt and more conditioned on the prior images during expansion and closing. We achieve this by reducing guidance scale and widen the field of view. Meanwhile, we adjust the variance of the initial noise to avoid the color blocks caused by low guidance scale. Through experimentation, we've found that a combination of slightly high initial variance and low guidance scale can stably reduce hallucinations and color blocks during the expansion and closing stages." + } + ] + } + ], + "index": 10 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "spans": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "type": "text", + "content": "8" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 166, + 91, + 217, + 101 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 166, + 91, + 217, + 101 + ], + "spans": [ + { + "bbox": [ + 166, + 91, + 217, + 101 + ], + "type": "text", + "content": "A. Liu et al." + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 7 + }, + { + "para_blocks": [ + { + "bbox": [ + 132, + 114, + 230, + 129 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 114, + 230, + 129 + ], + "spans": [ + { + "bbox": [ + 132, + 114, + 230, + 129 + ], + "type": "text", + "content": "4 Experiments" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 130, + 157, + 482, + 206 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 157, + 482, + 206 + ], + "spans": [ + { + "bbox": [ + 130, + 157, + 482, + 206 + ], + "type": "text", + "content": "We evaluate the performance of PanoFree across three generation tasks: Planar Panorama Generation, 360 Panorama Generation, and Full Panorama Generation. However, note that we focus on planar panorama and " + }, + { + "bbox": [ + 130, + 157, + 482, + 206 + ], + "type": "inline_equation", + "content": "360^{\\circ}" + }, + { + "bbox": [ + 130, + 157, + 482, + 206 + ], + "type": "text", + "content": " panorama generation, where the comparisons are more precise and consistent." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 130, + 209, + 482, + 281 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 209, + 482, + 281 + ], + "spans": [ + { + "bbox": [ + 130, + 209, + 482, + 281 + ], + "type": "text", + "content": "Implementation details. PanoFree is implemented using the publicly available Stable Diffusion code from Diffusers [34] based on the PyTorch framework. For the experiments in the main paper, we utilized the generation and inpainting models of Stable Diffusion (SD) v2.0 [38]. All experiments are conducted on a single NVIDIA RTX A6000 GPU. Further details and specific configurations can be found in the corresponding sections of the main paper and the supplementary." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 130, + 284, + 482, + 320 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 284, + 482, + 320 + ], + "spans": [ + { + "bbox": [ + 130, + 284, + 482, + 320 + ], + "type": "text", + "content": "Evaluation metrics. We introduce a more comprehensive set of evaluation metrics than prior work [1, 22, 46] covering five themes: image quality, global consistency, prompt capability, diversity, and resource consumption." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 131, + 323, + 479, + 536 + ], + "type": "list", + "angle": 0, + "index": 11, + "blocks": [ + { + "bbox": [ + 131, + 323, + 479, + 369 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 131, + 323, + 479, + 369 + ], + "spans": [ + { + "bbox": [ + 131, + 323, + 479, + 369 + ], + "type": "text", + "content": "- Image Quality is measured with Fréchet Inception Distance (FID) [14], Kernel Inception Distance (KID) [2], which measure fidelity and diversity. FID and KID calculated between the views randomly cropped from the panorama and reference images generated by SD with the same prompts." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 131, + 371, + 479, + 404 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 131, + 371, + 479, + 404 + ], + "spans": [ + { + "bbox": [ + 131, + 371, + 479, + 404 + ], + "type": "text", + "content": "- Global Consistency is measured with Intra-LPIPS (IL) [55] used by SyncDiffusion [22], which is computed by cropping non-overlapping views from a panorama and computing the averaged LPIPS scores of all view pairs." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 131, + 407, + 479, + 428 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 131, + 407, + 479, + 428 + ], + "spans": [ + { + "bbox": [ + 131, + 407, + 479, + 428 + ], + "type": "text", + "content": "- Prompt Capability is measured via CLIP Score (CS) [13] by computing the text-image similarity of randomly cropped views of the panorama." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 131, + 431, + 479, + 488 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 131, + 431, + 479, + 488 + ], + "spans": [ + { + "bbox": [ + 131, + 431, + 479, + 488 + ], + "type": "text", + "content": "- Panorama Diversity is also measured by FID and KID. Additionally, we propose Cross-LPIPS (CS) [55]. Cross-LPIPS is computed across 2 panoramas generated with a same text with different random seeds. We crop non-overlapping views from each panorama, and compute the averaged LPIPS scores of all view pairs where two views come from different panoramas." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 131, + 491, + 479, + 536 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 131, + 491, + 479, + 536 + ], + "spans": [ + { + "bbox": [ + 131, + 491, + 479, + 536 + ], + "type": "text", + "content": "- Resource Consumption includes time consumption, measured by the cumulative time cost of all diffusion processes to generate a single panorama, and peak GPU memory consumption, measured by the maximum GPU memory consumption during inference." + } + ] + } + ], + "index": 10 + } + ], + "sub_type": "text" + }, + { + "bbox": [ + 130, + 542, + 482, + 615 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 542, + 482, + 615 + ], + "spans": [ + { + "bbox": [ + 130, + 542, + 482, + 615 + ], + "type": "text", + "content": "Evaluation Settings. Prior work either used arbitrary prompts [1,22] or only focused on a single type of scene [46]. Instead, we consider 3 distinct scene types: indoor, street, and city scenes, and natural scenes. We obtained 100 prompts for each type from ChatGPT [35]. We use 10 random seeds per prompt for planar panorama and 360 panorama generation, and 3 different random seeds per prompt for full panorama generation (see supplementary for details)." + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 130, + 617, + 482, + 666 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 617, + 482, + 666 + ], + "spans": [ + { + "bbox": [ + 130, + 617, + 482, + 666 + ], + "type": "text", + "content": "User Study. For planar panorama generation and " + }, + { + "bbox": [ + 130, + 617, + 482, + 666 + ], + "type": "inline_equation", + "content": "360^{\\circ}" + }, + { + "bbox": [ + 130, + 617, + 482, + 666 + ], + "type": "text", + "content": " panorama generation, we conducted four user studies for each task to further evaluate the global consistency, image quality, prompt compatibility, and diversity of the generated panoramas (see supplementary for details)." + } + ] + } + ], + "index": 13 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 408, + 91, + 447, + 100 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 408, + 91, + 447, + 100 + ], + "spans": [ + { + "bbox": [ + 408, + 91, + 447, + 100 + ], + "type": "text", + "content": "PanoFree" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 474, + 91, + 481, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 474, + 91, + 481, + 100 + ], + "spans": [ + { + "bbox": [ + 474, + 91, + 481, + 100 + ], + "type": "text", + "content": "9" + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 8 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 159, + 114, + 451, + 219 + ], + "blocks": [ + { + "bbox": [ + 159, + 114, + 451, + 219 + ], + "lines": [ + { + "bbox": [ + 159, + 114, + 451, + 219 + ], + "spans": [ + { + "bbox": [ + 159, + 114, + 451, + 219 + ], + "type": "image", + "image_path": "c85aecde535a1bec8d23bcc3582d7bba3a47c6d8ca6e7ac19655d03409efa4c4.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 130, + 224, + 480, + 257 + ], + "lines": [ + { + "bbox": [ + 130, + 224, + 480, + 257 + ], + "spans": [ + { + "bbox": [ + 130, + 224, + 480, + 257 + ], + "type": "text", + "content": "Fig. 3: Planar Panorama generation results. Compared to vanilla Sequential Generation and MultiDiffusion (MD) [1], PanoFree achieves superior global consistency and image quality. It is also comparable to SyncDiffusion (SYD) [22] in these aspects." + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_caption" + } + ], + "index": 2 + }, + { + "type": "table", + "bbox": [ + 158, + 319, + 449, + 380 + ], + "blocks": [ + { + "bbox": [ + 130, + 264, + 480, + 308 + ], + "lines": [ + { + "bbox": [ + 130, + 264, + 480, + 308 + ], + "spans": [ + { + "bbox": [ + 130, + 264, + 480, + 308 + ], + "type": "text", + "content": "Table 1: Comparison of tuning free methods for Planar Panorama generation using Stable Diffusion [38]. We find PanoFree (PF) outperforms the state-of-the-art while having low computational requirements. Note that Cross-LPIPS and Intra-LPIPS are in " + }, + { + "bbox": [ + 130, + 264, + 480, + 308 + ], + "type": "inline_equation", + "content": "10^{-2}" + }, + { + "bbox": [ + 130, + 264, + 480, + 308 + ], + "type": "text", + "content": " scale, KID is in " + }, + { + "bbox": [ + 130, + 264, + 480, + 308 + ], + "type": "inline_equation", + "content": "10^{-3}" + }, + { + "bbox": [ + 130, + 264, + 480, + 308 + ], + "type": "text", + "content": " scale." + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 158, + 319, + 449, + 380 + ], + "lines": [ + { + "bbox": [ + 158, + 319, + 449, + 380 + ], + "spans": [ + { + "bbox": [ + 158, + 319, + 449, + 380 + ], + "type": "table", + "html": "
MethodIntra-LPIPS↓Cross-LPIPS↑FID↓KID↓CS↑Time (s)↓Memory (GB)↓
SG70.4071.0324.914.3326.68253.2
MD [1]68.4869.9221.163.5027.89955.8
SYD [22]64.4868.0720.563.6227.1812810.0
PF (Ours)65.3469.6817.053.8027.21263.2
", + "image_path": "4ddebd2515bb9be9314c9e80cf1f229afa487a6274081e53d626cafe55229b4b.jpg" + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "table_body" + } + ], + "index": 5 + }, + { + "bbox": [ + 131, + 399, + 308, + 410 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 131, + 399, + 308, + 410 + ], + "spans": [ + { + "bbox": [ + 131, + 399, + 308, + 410 + ], + "type": "text", + "content": "4.1 Planar Panorama Generation" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 130, + 424, + 481, + 460 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 424, + 481, + 460 + ], + "spans": [ + { + "bbox": [ + 130, + 424, + 481, + 460 + ], + "type": "text", + "content": "Planar Panorama corresponds to the scene observed with camera translation along the focal plane in reality. This is a relatively simple task, as it only involves extending the image without considering more complex geometric changes." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 130, + 460, + 483, + 496 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 460, + 483, + 496 + ], + "spans": [ + { + "bbox": [ + 130, + 460, + 483, + 496 + ], + "type": "text", + "content": "Baselines. We have chosen 3 tuning-free baselines for comparison, Vanilla Sequential Generation (SG), MultiDiffusion (MD) [1] and SyncDiffusion (SYD) [22]. Additional details are in the supplementary." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 130, + 497, + 480, + 520 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 497, + 480, + 520 + ], + "spans": [ + { + "bbox": [ + 130, + 497, + 480, + 520 + ], + "type": "text", + "content": "Results. The quantitative and qualitative evaluations are shown in Table 1 and Fig. 3, respectively. Below we compare PanoFree to each baseline." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 131, + 521, + 480, + 665 + ], + "type": "list", + "angle": 0, + "index": 13, + "blocks": [ + { + "bbox": [ + 131, + 521, + 480, + 569 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 131, + 521, + 480, + 569 + ], + "spans": [ + { + "bbox": [ + 131, + 521, + 480, + 569 + ], + "type": "text", + "content": "- Compared with vanilla Sequential Generation, PanoFree significantly enhances image quality and global consistency, demonstrating its effectiveness in reducing accumulated errors. Moreover, PanoFree does not compromise diversity or have a significant effect on GPU time and memory overhead." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 131, + 570, + 480, + 604 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 131, + 570, + 480, + 604 + ], + "spans": [ + { + "bbox": [ + 131, + 570, + 480, + 604 + ], + "type": "text", + "content": "- Compared with MultiDiffusion, PanoFree has significant advantages in image quality and global consistency. Meanwhile, its time and GPU memory overhead is only " + }, + { + "bbox": [ + 131, + 570, + 480, + 604 + ], + "type": "inline_equation", + "content": "26\\%" + }, + { + "bbox": [ + 131, + 570, + 480, + 604 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 131, + 570, + 480, + 604 + ], + "type": "inline_equation", + "content": "55\\%" + }, + { + "bbox": [ + 131, + 570, + 480, + 604 + ], + "type": "text", + "content": " that of MultiDiffusion, respectively." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 131, + 605, + 480, + 665 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 131, + 605, + 480, + 665 + ], + "spans": [ + { + "bbox": [ + 131, + 605, + 480, + 665 + ], + "type": "text", + "content": "- Compared with SyncDiffusion, PanoFree achieves comparable performance in global consistency and image quality. Although SyncDiffusion performs better in consistency, it requires introducing additional models for latent optimization. This leads PanoFree's time overhead to be " + }, + { + "bbox": [ + 131, + 605, + 480, + 665 + ], + "type": "inline_equation", + "content": "20\\%" + }, + { + "bbox": [ + 131, + 605, + 480, + 665 + ], + "type": "text", + "content": " of SyncDiffusion and GPU memory overhead to be " + }, + { + "bbox": [ + 131, + 605, + 480, + 665 + ], + "type": "inline_equation", + "content": "32\\%" + }, + { + "bbox": [ + 131, + 605, + 480, + 665 + ], + "type": "text", + "content": " of SyncDiffusion." + } + ] + } + ], + "index": 12 + } + ], + "sub_type": "text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "spans": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "type": "text", + "content": "10" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 166, + 91, + 217, + 101 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 166, + 91, + 217, + 101 + ], + "spans": [ + { + "bbox": [ + 166, + 91, + 217, + 101 + ], + "type": "text", + "content": "A. Liu et al." + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 9 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 143, + 115, + 468, + 200 + ], + "blocks": [ + { + "bbox": [ + 143, + 115, + 468, + 200 + ], + "lines": [ + { + "bbox": [ + 143, + 115, + 468, + 200 + ], + "spans": [ + { + "bbox": [ + 143, + 115, + 468, + 200 + ], + "type": "image", + "image_path": "370e94a9c5a2cde88e8f1ad51d2edc54951b80405d196c6691e9e67acd0b50f2.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 131, + 207, + 479, + 240 + ], + "lines": [ + { + "bbox": [ + 131, + 207, + 479, + 240 + ], + "spans": [ + { + "bbox": [ + 131, + 207, + 479, + 240 + ], + "type": "text", + "content": "Fig. 4: Diversity comparison on Planar Panorama Generation task. Each group is generated using the same text with different random seeds. Compared to MultiDiffusion (MD) [1] and SyncDiffusion (SYD) [22], PanoFree achieves superior diversity." + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_caption" + } + ], + "index": 2 + }, + { + "type": "table", + "bbox": [ + 138, + 278, + 476, + 324 + ], + "blocks": [ + { + "bbox": [ + 131, + 251, + 479, + 273 + ], + "lines": [ + { + "bbox": [ + 131, + 251, + 479, + 273 + ], + "spans": [ + { + "bbox": [ + 131, + 251, + 479, + 273 + ], + "type": "text", + "content": "Table 2: User study results of Planar Panorama Generation. 15 questions are used for each evaluation item and answered by 5 Amazon MTurk workers." + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 138, + 278, + 476, + 324 + ], + "lines": [ + { + "bbox": [ + 138, + 278, + 476, + 324 + ], + "spans": [ + { + "bbox": [ + 138, + 278, + 476, + 324 + ], + "type": "table", + "html": "
Consistency (%)Quality (%)Prompt Compatibility (%)Diversity (%)
SYD [22]52.746.041.335.3
PF (ours)47.354.058.764.7
", + "image_path": "659b7c10e98555d26fed19f8b302f29766035d789e5295b58257d1d4622ff609.jpg" + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "table_body" + } + ], + "index": 5 + }, + { + "bbox": [ + 130, + 341, + 479, + 400 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 341, + 479, + 400 + ], + "spans": [ + { + "bbox": [ + 130, + 341, + 479, + 400 + ], + "type": "text", + "content": "The Loss of Diversity with Joint Diffusion. When using different random seeds with the same prompt, methods using Joint Diffusion exhibit reduced diversity in their results. In contrast, our PanoFree method can better maintain diversity (see cross-LPIPS scores in Table 1). Additionally, we believe this is the source of PanoFree's gains over MultiDiffusion and SyncDiffusion in FID." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 130, + 401, + 480, + 460 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 401, + 480, + 460 + ], + "spans": [ + { + "bbox": [ + 130, + 401, + 480, + 460 + ], + "type": "text", + "content": "This diversity issue becomes particularly apparent when given some underspecified prompts. Therefore, we generated a \"underspecified set\" consisting of 20 short and blurry prompts to demonstrate this issue. For each prompt, we used 20 different random seeds. We demonstrate the diversity differences qualitatively in Fig. 4. Please refer to supplementary for quantitative analysis." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 130, + 460, + 480, + 509 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 460, + 480, + 509 + ], + "spans": [ + { + "bbox": [ + 130, + 460, + 480, + 509 + ], + "type": "text", + "content": "User Study. The results in Table 2 clearly show that human evaluators believe PanoFree produces more diverse panoramas and demonstrates better compatibility with prompts than SyncDiffusion [22]. Additionally, both methods exhibit similar levels of global consistency and image quality." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 131, + 526, + 292, + 537 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 131, + 526, + 292, + 537 + ], + "spans": [ + { + "bbox": [ + 131, + 526, + 292, + 537 + ], + "type": "text", + "content": "4.2 360 Panorama Generation" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 130, + 546, + 480, + 616 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 546, + 480, + 616 + ], + "spans": [ + { + "bbox": [ + 130, + 546, + 480, + 616 + ], + "type": "text", + "content": "Due to the distortion caused by equirectangular projection, generating 360-degree panoramas is more challenging than planar panorama generation. Vanilla sequential generation tends to produce many artifacts, significantly decreasing image quality. Moreover, MultiDiffusion [1] and SyncDiffusion [22] cannot be directly used for generating 360-degree panoramas. As far as we know, PanoFree is the first implementation of training-free 360-degree panorama generation." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 131, + 617, + 481, + 641 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 131, + 617, + 481, + 641 + ], + "spans": [ + { + "bbox": [ + 131, + 617, + 481, + 641 + ], + "type": "text", + "content": "Baselines. We used 2 baselines for comparison: Vanilla Sequential Generation (SG) and MVDiffusion (MVD). Additional details are in the supplementary." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 131, + 641, + 481, + 665 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 131, + 641, + 481, + 665 + ], + "spans": [ + { + "bbox": [ + 131, + 641, + 481, + 665 + ], + "type": "text", + "content": "Results. The quantitative and qualitative evaluations are shown in Table 3 and Fig. 5 respectively. Below we compare PanoFree to each baseline." + } + ] + } + ], + "index": 12 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 408, + 91, + 447, + 100 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 408, + 91, + 447, + 100 + ], + "spans": [ + { + "bbox": [ + 408, + 91, + 447, + 100 + ], + "type": "text", + "content": "PanoFree" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 470, + 91, + 479, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 470, + 91, + 479, + 100 + ], + "spans": [ + { + "bbox": [ + 470, + 91, + 479, + 100 + ], + "type": "text", + "content": "11" + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 10 + }, + { + "para_blocks": [ + { + "type": "table", + "bbox": [ + 157, + 168, + 451, + 222 + ], + "blocks": [ + { + "bbox": [ + 130, + 114, + 482, + 158 + ], + "lines": [ + { + "bbox": [ + 130, + 114, + 482, + 158 + ], + "spans": [ + { + "bbox": [ + 130, + 114, + 482, + 158 + ], + "type": "text", + "content": "Table 3: Comparison of " + }, + { + "bbox": [ + 130, + 114, + 482, + 158 + ], + "type": "inline_equation", + "content": "360^{\\circ}" + }, + { + "bbox": [ + 130, + 114, + 482, + 158 + ], + "type": "text", + "content": " Panorama generation methods using Stable Diffusion [38]. We find PanoFree (PF) still outperforms the state-of-the-art while having low computational requirements. Note that Cross-LPIPS and Intra-LPIPS are in " + }, + { + "bbox": [ + 130, + 114, + 482, + 158 + ], + "type": "inline_equation", + "content": "10^{-2}" + }, + { + "bbox": [ + 130, + 114, + 482, + 158 + ], + "type": "text", + "content": " scale, KID is in " + }, + { + "bbox": [ + 130, + 114, + 482, + 158 + ], + "type": "inline_equation", + "content": "10^{-3}" + }, + { + "bbox": [ + 130, + 114, + 482, + 158 + ], + "type": "text", + "content": " scale." + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 157, + 168, + 451, + 222 + ], + "lines": [ + { + "bbox": [ + 157, + 168, + 451, + 222 + ], + "spans": [ + { + "bbox": [ + 157, + 168, + 451, + 222 + ], + "type": "table", + "html": "
MethodIntra-LPIPS↓Cross-LPIPS↑FID↓KID↓CS↑Time (s)↓Memory (GB)↓
SG70.6273.0632.287.9026.35213.2
MVD [46]67.7170.0737.898.7626.271106.9
PF (ours)68.6272.6725.847.4826.51223.2
", + "image_path": "130833398340003c7cdf6c28488dbe39fc20ec687b3c94599906f7575b10d0c2.jpg" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "table_body" + } + ], + "index": 3 + }, + { + "type": "image", + "bbox": [ + 149, + 224, + 462, + 334 + ], + "blocks": [ + { + "bbox": [ + 149, + 224, + 462, + 334 + ], + "lines": [ + { + "bbox": [ + 149, + 224, + 462, + 334 + ], + "spans": [ + { + "bbox": [ + 149, + 224, + 462, + 334 + ], + "type": "image", + "image_path": "3043c22643dd25c5abe0f5f7f8e86baec2d046fa46ea0de2d672715d85cfe54a.jpg" + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 130, + 339, + 480, + 373 + ], + "lines": [ + { + "bbox": [ + 130, + 339, + 480, + 373 + ], + "spans": [ + { + "bbox": [ + 130, + 339, + 480, + 373 + ], + "type": "text", + "content": "Fig. 5: " + }, + { + "bbox": [ + 130, + 339, + 480, + 373 + ], + "type": "inline_equation", + "content": "360^{\\circ}" + }, + { + "bbox": [ + 130, + 339, + 480, + 373 + ], + "type": "text", + "content": " Panorama generation results. Compared to vanilla Sequential Generation (SG), PanoFree achieves superior global consistency and image quality. It is also comparable to MVDiffusion (MVD) [46] in these aspects." + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "image_caption" + } + ], + "index": 4 + }, + { + "bbox": [ + 130, + 380, + 482, + 440 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 380, + 482, + 440 + ], + "spans": [ + { + "bbox": [ + 130, + 380, + 482, + 440 + ], + "type": "text", + "content": "- Compared with vanilla Sequential Generation, PanoFree significantly enhances image quality and global consistency. Specifically, vanilla sequential generation creates artifacts with complex optical geometry transformations, severely impacting image quality. However, PanoFree effectively recovers image quality by estimating and erasing risky areas, minimizing artifact propagation." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 130, + 440, + 482, + 559 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 440, + 482, + 559 + ], + "spans": [ + { + "bbox": [ + 130, + 440, + 482, + 559 + ], + "type": "text", + "content": "- Compared with MVDiffusion, PanoFree achieves comparability in image quality and global consistency, yet significantly outperforms in terms of time, GPU memory overhead, and diversity. Particularly, MVDiffusion is significantly worse than PanoFree in terms of FID and KID scores, even underperforming vanilla Sequential Generation. This is partly due to the inevitable bias of MVDiffusion's generated results towards the training dataset, resulting in larger discrepancies compared to those produced by Stable Diffusion. Visually, MVDiffusion also exhibits a noticeable lack of generation diversity. As depicted in Fig. 6, given a prompt, results generated with different random seeds show minimal variation in both content and style." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 130, + 559, + 482, + 608 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 559, + 482, + 608 + ], + "spans": [ + { + "bbox": [ + 130, + 559, + 482, + 608 + ], + "type": "text", + "content": "User Study. The user study results in Table 4 show that human evaluators believe PanoFree also produces more diverse " + }, + { + "bbox": [ + 130, + 559, + 482, + 608 + ], + "type": "inline_equation", + "content": "360^{\\circ}" + }, + { + "bbox": [ + 130, + 559, + 482, + 608 + ], + "type": "text", + "content": " panoramas compared with MVDiffusion [46]. And PanoFree demonstrates better global consistency. Both methods exhibit similar levels of image quality and prompt comparability." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 131, + 623, + 345, + 635 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 131, + 623, + 345, + 635 + ], + "spans": [ + { + "bbox": [ + 131, + 623, + 345, + 635 + ], + "type": "text", + "content": "4.3 Full Spherical Panorama Generation" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 130, + 641, + 481, + 665 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 641, + 481, + 665 + ], + "spans": [ + { + "bbox": [ + 130, + 641, + 481, + 665 + ], + "type": "text", + "content": "PanoFree is also the first to achieve feasible tuning-free generation for Full Spherical Panoramas. However, it's hard to conduct meaningful comparisons due the" + } + ] + } + ], + "index": 10 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "spans": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "type": "text", + "content": "12" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 166, + 91, + 217, + 100 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 166, + 91, + 217, + 100 + ], + "spans": [ + { + "bbox": [ + 166, + 91, + 217, + 100 + ], + "type": "text", + "content": "A. Liu et al." + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 11 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 161, + 125, + 304, + 215 + ], + "blocks": [ + { + "bbox": [ + 193, + 116, + 411, + 125 + ], + "lines": [ + { + "bbox": [ + 193, + 116, + 411, + 125 + ], + "spans": [ + { + "bbox": [ + 193, + 116, + 411, + 125 + ], + "type": "text", + "content": "\"Charming alpine village nestled among snow-capped peaks\"" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 161, + 125, + 304, + 215 + ], + "lines": [ + { + "bbox": [ + 161, + 125, + 304, + 215 + ], + "spans": [ + { + "bbox": [ + 161, + 125, + 304, + 215 + ], + "type": "image", + "image_path": "8acd5007191249bdfb9bf3f8c8c8d669c9b8816e04fe224fb779af2ded2f2dcc.jpg" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 225, + 217, + 246, + 226 + ], + "lines": [ + { + "bbox": [ + 225, + 217, + 246, + 226 + ], + "spans": [ + { + "bbox": [ + 225, + 217, + 246, + 226 + ], + "type": "text", + "content": "MVD" + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 131, + 232, + 480, + 265 + ], + "lines": [ + { + "bbox": [ + 131, + 232, + 480, + 265 + ], + "spans": [ + { + "bbox": [ + 131, + 232, + 480, + 265 + ], + "type": "text", + "content": "Fig. 6: Diversity comparison on " + }, + { + "bbox": [ + 131, + 232, + 480, + 265 + ], + "type": "inline_equation", + "content": "360^{\\circ}" + }, + { + "bbox": [ + 131, + 232, + 480, + 265 + ], + "type": "text", + "content": " Panorama Generation task. Each group is generated using the same text with different random seeds. Compared to MVDiffusion (MVD) [46], PanoFree achieves superior diversity." + } + ] + } + ], + "index": 7, + "angle": 0, + "type": "image_caption" + } + ], + "index": 3 + }, + { + "type": "image", + "bbox": [ + 307, + 125, + 449, + 215 + ], + "blocks": [ + { + "bbox": [ + 307, + 125, + 449, + 215 + ], + "lines": [ + { + "bbox": [ + 307, + 125, + 449, + 215 + ], + "spans": [ + { + "bbox": [ + 307, + 125, + 449, + 215 + ], + "type": "image", + "image_path": "d9cba1a78fe86f2de9153b9768d86cedb5ba05f56c6435bf11ec496354479830.jpg" + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 370, + 217, + 381, + 226 + ], + "lines": [ + { + "bbox": [ + 370, + 217, + 381, + 226 + ], + "spans": [ + { + "bbox": [ + 370, + 217, + 381, + 226 + ], + "type": "text", + "content": "PF" + } + ] + } + ], + "index": 6, + "angle": 0, + "type": "image_caption" + } + ], + "index": 5 + }, + { + "type": "table", + "bbox": [ + 138, + 306, + 476, + 354 + ], + "blocks": [ + { + "bbox": [ + 131, + 274, + 480, + 297 + ], + "lines": [ + { + "bbox": [ + 131, + 274, + 480, + 297 + ], + "spans": [ + { + "bbox": [ + 131, + 274, + 480, + 297 + ], + "type": "text", + "content": "Table 4: User study results of " + }, + { + "bbox": [ + 131, + 274, + 480, + 297 + ], + "type": "inline_equation", + "content": "360^{\\circ}" + }, + { + "bbox": [ + 131, + 274, + 480, + 297 + ], + "type": "text", + "content": " Panorama Generation. 15 questions are used for each evaluation item and answered by 5 Amazon MTurk workers." + } + ] + } + ], + "index": 8, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 138, + 306, + 476, + 354 + ], + "lines": [ + { + "bbox": [ + 138, + 306, + 476, + 354 + ], + "spans": [ + { + "bbox": [ + 138, + 306, + 476, + 354 + ], + "type": "table", + "html": "
Consistency (%)Quality (%)Prompt Compatibility (%)Diversity (%)
MVD [46]40.748.747.333.3
PF (ours)59.351.352.666.6
", + "image_path": "389b4c352725aadf85380b42203f08eb4d61caab079889e4b1a8748370246455.jpg" + } + ] + } + ], + "index": 9, + "angle": 0, + "type": "table_body" + } + ], + "index": 9 + }, + { + "bbox": [ + 130, + 368, + 482, + 477 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 368, + 482, + 477 + ], + "spans": [ + { + "bbox": [ + 130, + 368, + 482, + 477 + ], + "type": "text", + "content": "lack of tuning-free generation baseline methods or those with strong out-of-scope generation capabilities for Full Spherical Panoramas generation task. Thus, we conduct qualitative evaluation as well as comparison with vanilla sequential generation by showcasing generated results in Fig. 7. Vanilla sequential generation exhibit more artifacts as distortion increases. Additionally, partial conditioning issue mentioned in Sec. 3.5 causes hallucinations. And PanoFree still could effectively reduce artifacts and hallucinations. Note that we start both methods from " + }, + { + "bbox": [ + 130, + 368, + 482, + 477 + ], + "type": "inline_equation", + "content": "360^{\\circ}" + }, + { + "bbox": [ + 130, + 368, + 482, + 477 + ], + "type": "text", + "content": " panoramas generated by PanoFree into full spherical panoramas, otherwise vanilla sequential generation will perform even worse." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 132, + 491, + 239, + 502 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 491, + 239, + 502 + ], + "spans": [ + { + "bbox": [ + 132, + 491, + 239, + 502 + ], + "type": "text", + "content": "4.4 Ablation Study" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 130, + 507, + 482, + 605 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 507, + 482, + 605 + ], + "spans": [ + { + "bbox": [ + 130, + 507, + 482, + 605 + ], + "type": "text", + "content": "Tab. 5 contains an ablation study that sequentially integrates each PanoFree component. We evaluate consistency (Intra-LPIPS) and image quality (FID) with " + }, + { + "bbox": [ + 130, + 507, + 482, + 605 + ], + "type": "inline_equation", + "content": "30\\%" + }, + { + "bbox": [ + 130, + 507, + 482, + 605 + ], + "type": "text", + "content": " of the prompts from Sec. 4.1 & Sec. 4.2 for both planar and " + }, + { + "bbox": [ + 130, + 507, + 482, + 605 + ], + "type": "inline_equation", + "content": "360^{\\circ}" + }, + { + "bbox": [ + 130, + 507, + 482, + 605 + ], + "type": "text", + "content": " panorama. We show that cross-view guidance provides the strongest benefit, followed by distance and edge-based risky area erasing. These components effectively reduce image tearing and visual chaos. Color and smoothness-based erasing have a smaller impact, likely due noise in these low-level features. Qualitative results are in Sec. B of the supplementary." + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 132, + 619, + 220, + 632 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 619, + 220, + 632 + ], + "spans": [ + { + "bbox": [ + 132, + 619, + 220, + 632 + ], + "type": "text", + "content": "5 Conclusion" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 131, + 641, + 481, + 665 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 131, + 641, + 481, + 665 + ], + "spans": [ + { + "bbox": [ + 131, + 641, + 481, + 665 + ], + "type": "text", + "content": "We present PanoFree, a tuning-free multi-view image generation that supports an extensive array of correspondences. PanoFree improves error accumulation" + } + ] + } + ], + "index": 14 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 408, + 91, + 447, + 100 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 408, + 91, + 447, + 100 + ], + "spans": [ + { + "bbox": [ + 408, + 91, + 447, + 100 + ], + "type": "text", + "content": "PanoFree" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 470, + 91, + 481, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 470, + 91, + 481, + 100 + ], + "spans": [ + { + "bbox": [ + 470, + 91, + 481, + 100 + ], + "type": "text", + "content": "13" + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 12 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 134, + 114, + 476, + 265 + ], + "blocks": [ + { + "bbox": [ + 134, + 114, + 476, + 265 + ], + "lines": [ + { + "bbox": [ + 134, + 114, + 476, + 265 + ], + "spans": [ + { + "bbox": [ + 134, + 114, + 476, + 265 + ], + "type": "image", + "image_path": "e7adb3f70906836ab705afcf8479db7d0491a05beebe9f5200a7e1256913485c.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 130, + 270, + 480, + 302 + ], + "lines": [ + { + "bbox": [ + 130, + 270, + 480, + 302 + ], + "spans": [ + { + "bbox": [ + 130, + 270, + 480, + 302 + ], + "type": "text", + "content": "Fig. 7: Full Spherical Panorama generation results. Vanilla Sequential Generation (SG) tends to generate hallucinations due to partial conditions, while PanoFree effectively mitigates this issue." + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_caption" + } + ], + "index": 2 + }, + { + "type": "table", + "bbox": [ + 280, + 316, + 473, + 423 + ], + "blocks": [ + { + "bbox": [ + 130, + 317, + 272, + 427 + ], + "lines": [ + { + "bbox": [ + 130, + 317, + 272, + 427 + ], + "spans": [ + { + "bbox": [ + 130, + 317, + 272, + 427 + ], + "type": "text", + "content": "Table 5: Quantitative ablation of PanoFree components using " + }, + { + "bbox": [ + 130, + 317, + 272, + 427 + ], + "type": "inline_equation", + "content": "30\\%" + }, + { + "bbox": [ + 130, + 317, + 272, + 427 + ], + "type": "text", + "content": " of prompts from Sec. 4.1 & 4.2. Intra-LPIPS " + }, + { + "bbox": [ + 130, + 317, + 272, + 427 + ], + "type": "inline_equation", + "content": "(10^{-2})" + }, + { + "bbox": [ + 130, + 317, + 272, + 427 + ], + "type": "text", + "content": " and FID show cross-view guidance offers the most benefit, followed by distance and edge-based risky area erasing. Color and smoothness-based erasing have minimal impact." + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 280, + 316, + 473, + 423 + ], + "lines": [ + { + "bbox": [ + 280, + 316, + 473, + 423 + ], + "spans": [ + { + "bbox": [ + 280, + 316, + 473, + 423 + ], + "type": "table", + "html": "
TaskMethodIntra-LPIPS↓FID↓
PlanarNone70.8424.75
+ CG66.6320.63
+ Dist65.8718.21
360°None71.3433.47
+ CG69.2127.38
+ Dist & Edge69.0326.69
+ Color & Smooth68.9426.51
", + "image_path": "d4329e01ed82016b271e65a6f98120d9121ebd9ecb53d8f7ee231803709ed0bf.jpg" + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "table_body" + } + ], + "index": 5 + }, + { + "bbox": [ + 130, + 470, + 482, + 602 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 470, + 482, + 602 + ], + "spans": [ + { + "bbox": [ + 130, + 470, + 482, + 602 + ], + "type": "text", + "content": "by enhancing cross-view awareness and refining the warping and inpainting processes through cross-view guidance, risky area estimation and erasing, and symmetric bidirectional guided generation for loop closure, alongside guidance-based semantic and density control for scene structure preservation. PanoFree is evaluated on various panorama types—Planar, " + }, + { + "bbox": [ + 130, + 470, + 482, + 602 + ], + "type": "inline_equation", + "content": "360^{\\circ}" + }, + { + "bbox": [ + 130, + 470, + 482, + 602 + ], + "type": "text", + "content": ", and Full Spherical Panoramas. PanoFree demonstrates significant error reduction, improved global consistency, and image quality across different scenarios without extra fine-tuning. Compared to existing methods, PanoFree is up to 5x more efficient in time and 3x more efficient in GPU memory usage, and maintains superior diversity of results (2x better in our user study). Moreover, PanoFree can be extended to texture generation for 3D models. We intend to explore these possibilities in future research." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 130, + 605, + 482, + 665 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 605, + 482, + 665 + ], + "spans": [ + { + "bbox": [ + 130, + 605, + 482, + 665 + ], + "type": "text", + "content": "Limitations. A limitation of our work is that we are unable to generate scenes beyond the capability of the pre-trained T2I model. Therefore, we rely on large pre-trained T2I models to ensure the broad application scope. And when provided with text descriptions beyond the capability range of the pre-trained T2I models, the generated results may not match the text." + } + ] + } + ], + "index": 7 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "spans": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "type": "text", + "content": "14" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 166, + 91, + 217, + 100 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 166, + 91, + 217, + 100 + ], + "spans": [ + { + "bbox": [ + 166, + 91, + 217, + 100 + ], + "type": "text", + "content": "A. Liu et al." + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 13 + }, + { + "para_blocks": [ + { + "bbox": [ + 133, + 114, + 197, + 126 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 114, + 197, + 126 + ], + "spans": [ + { + "bbox": [ + 133, + 114, + 197, + 126 + ], + "type": "text", + "content": "References" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 138, + 140, + 480, + 665 + ], + "type": "list", + "angle": 0, + "index": 21, + "blocks": [ + { + "bbox": [ + 138, + 140, + 480, + 161 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 140, + 480, + 161 + ], + "spans": [ + { + "bbox": [ + 138, + 140, + 480, + 161 + ], + "type": "text", + "content": "1. Bar-Tal, O., Yariv, L., Lipman, Y., Dekel, T.: Multidiffusion: Fusing diffusion paths for controlled image generation (2023)" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 138, + 162, + 480, + 193 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 162, + 480, + 193 + ], + "spans": [ + { + "bbox": [ + 138, + 162, + 480, + 193 + ], + "type": "text", + "content": "2. Binkowski, M., Sutherland, D.J., Arbel, M., Gretton, A.: Demystifying mmd gans. ArXiv abs/1801.01401 (2018), https://apisemantic scholar.org/CorpusID: 3531856" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 138, + 194, + 480, + 216 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 194, + 480, + 216 + ], + "spans": [ + { + "bbox": [ + 138, + 194, + 480, + 216 + ], + "type": "text", + "content": "3. Brock, A., Donahue, J., Simonyan, K.: Large scale gan training for high fidelity natural image synthesis. arXiv preprint arXiv:1809.11096 (2018)" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 138, + 217, + 480, + 248 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 217, + 480, + 248 + ], + "spans": [ + { + "bbox": [ + 138, + 217, + 480, + 248 + ], + "type": "text", + "content": "4. Chen, D.Z., Siddiqui, Y., Lee, H.Y., Tulyakov, S., Nießner, M.: Text2tex: Text-driven texture synthesis via diffusion models. arXiv preprint arXiv:2303.11396 (2023)" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 138, + 249, + 480, + 270 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 249, + 480, + 270 + ], + "spans": [ + { + "bbox": [ + 138, + 249, + 480, + 270 + ], + "type": "text", + "content": "5. Chen, Z., Wang, G., Liu, Z.: Text2light: Zero-shot text-driven hdr panorama generation. ACM Transactions on Graphics (TOG) 41(6), 1-16 (2022)" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 138, + 271, + 480, + 314 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 271, + 480, + 314 + ], + "spans": [ + { + "bbox": [ + 138, + 271, + 480, + 314 + ], + "type": "text", + "content": "6. Cheng, Y.C., Lin, C.H., Lee, H.Y., Ren, J., Tulyakov, S., Yang, M.H.: Inout: Diverse image outpainting via gan inversion. 2022 IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR) pp. 11421-11430 (2021), https://apisemantic scholar.org/CorpusID:232478397" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 138, + 315, + 480, + 336 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 315, + 480, + 336 + ], + "spans": [ + { + "bbox": [ + 138, + 315, + 480, + 336 + ], + "type": "text", + "content": "7. Dhariwal, P., Nichol, A.: Diffusion models beat gans on image synthesis. Advances in Neural Information Processing Systems 34, 8780-8794 (2021)" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 138, + 337, + 480, + 369 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 337, + 480, + 369 + ], + "spans": [ + { + "bbox": [ + 138, + 337, + 480, + 369 + ], + "type": "text", + "content": "8. Esser, P., Rombach, R., Ommer, B.: Taming transformers for high-resolution image synthesis. In: Proceedings of the IEEE/CVF conference on computer vision and pattern recognition. pp. 12873-12883 (2021)" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 138, + 370, + 480, + 392 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 370, + 480, + 392 + ], + "spans": [ + { + "bbox": [ + 138, + 370, + 480, + 392 + ], + "type": "text", + "content": "9. Fang, C., Hu, X., Luo, K., Tan, P.: Ctrl-room: Controllable text-to-3d room meshes generation with layout constraints. arXiv preprint arXiv:2310.03602 (2023)" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 138, + 392, + 480, + 424 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 392, + 480, + 424 + ], + "spans": [ + { + "bbox": [ + 138, + 392, + 480, + 424 + ], + "type": "text", + "content": "0. Feng, M., Liu, J., Cui, M., Xie, X.: Diffusion360: Seamless 360 degree panoramic image generation based on diffusion models. ArXiv abs/2311.13141 (2023), https://api-semanticscholar.org/CorpusID:265351889" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 138, + 425, + 480, + 445 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 425, + 480, + 445 + ], + "spans": [ + { + "bbox": [ + 138, + 425, + 480, + 445 + ], + "type": "text", + "content": "1. Fridman, R., Abecasis, A., Kasten, Y., Dekel, T.: Scenescape: Text-driven consistent scene generation. arXiv preprint arXiv:2302.01133 (2023)" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 138, + 447, + 480, + 479 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 447, + 480, + 479 + ], + "spans": [ + { + "bbox": [ + 138, + 447, + 480, + 479 + ], + "type": "text", + "content": "2. Goodfellow, I., Pouget-Abadie, J., Mirza, M., Xu, B., Warde-Farley, D., Ozair, S., Courville, A., Bengio, Y.: Generative adversarial networks. Communications of the ACM 63(11), 139-144 (2020)" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 138, + 479, + 480, + 511 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 479, + 480, + 511 + ], + "spans": [ + { + "bbox": [ + 138, + 479, + 480, + 511 + ], + "type": "text", + "content": "3. Hessel, J., Holtzman, A., Forbes, M., Bras, R.L., Choi, Y.: Clipscore: A reference-free evaluation metric for image captioning. ArXiv abs/2104.08718 (2021), https://api-semanticscholar.org/CorpusID:233296711" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 138, + 512, + 480, + 555 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 512, + 480, + 555 + ], + "spans": [ + { + "bbox": [ + 138, + 512, + 480, + 555 + ], + "type": "text", + "content": "4. Heusel, M., Ramsauer, H., Unterthiner, T., Nessler, B., Hochreiter, S.: Gans trained by a two time-scale update rule converge to a local nash equilibrium. In: Neural Information Processing Systems (2017), https://api-semanticscholar.org/CorpusID:326772" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 138, + 555, + 480, + 578 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 555, + 480, + 578 + ], + "spans": [ + { + "bbox": [ + 138, + 555, + 480, + 578 + ], + "type": "text", + "content": "5. Ho, J., Jain, A., Abbeel, P.: Denoising diffusion probabilistic models. Advances in Neural Information Processing Systems (NeurIPS) (2020)" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 138, + 578, + 480, + 609 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 578, + 480, + 609 + ], + "spans": [ + { + "bbox": [ + 138, + 578, + 480, + 609 + ], + "type": "text", + "content": "6. Ho, J., Sahara, C., Chan, W., Fleet, D.J., Norouzi, M., Salimans, T.: Cascaded diffusion models for high fidelity image generation. J. Mach. Learn. Res. 23(47), 1-33 (2022)" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 138, + 611, + 480, + 632 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 611, + 480, + 632 + ], + "spans": [ + { + "bbox": [ + 138, + 611, + 480, + 632 + ], + "type": "text", + "content": "7. Ho, J., Salimans, T.: Classifier-free diffusion guidance. arXiv preprint arXiv:2207.12598 (2022)" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 138, + 632, + 480, + 665 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 632, + 480, + 665 + ], + "spans": [ + { + "bbox": [ + 138, + 632, + 480, + 665 + ], + "type": "text", + "content": "8. Hollein, L., Cao, A., Owens, A., Johnson, J., Nießner, M.: Text2room: Extracting textured 3d meshes from 2d text-to-image models. arXiv preprint arXiv:2303.11989 (2023)" + } + ] + } + ], + "index": 20 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 408, + 91, + 447, + 100 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 408, + 91, + 447, + 100 + ], + "spans": [ + { + "bbox": [ + 408, + 91, + 447, + 100 + ], + "type": "text", + "content": "PanoFree" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 470, + 92, + 480, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 470, + 92, + 480, + 100 + ], + "spans": [ + { + "bbox": [ + 470, + 92, + 480, + 100 + ], + "type": "text", + "content": "15" + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 14 + }, + { + "para_blocks": [ + { + "bbox": [ + 132, + 117, + 480, + 666 + ], + "type": "list", + "angle": 0, + "index": 18, + "blocks": [ + { + "bbox": [ + 133, + 117, + 480, + 138 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 117, + 480, + 138 + ], + "spans": [ + { + "bbox": [ + 133, + 117, + 480, + 138 + ], + "type": "text", + "content": "19. Karras, T., Aittala, M., Aila, T., Laine, S.: Elucidating the design space of diffusion-based generative models. arXiv preprint arXiv:2206.00364 (2022)" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 132, + 140, + 480, + 160 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 140, + 480, + 160 + ], + "spans": [ + { + "bbox": [ + 132, + 140, + 480, + 160 + ], + "type": "text", + "content": "20. Karras, T., Aittala, M., Aila, T., Laine, S.: Elucidating the design space of diffusion-based generative models. arXiv preprint arXiv:2206.00364 (2022)" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 133, + 161, + 480, + 194 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 161, + 480, + 194 + ], + "spans": [ + { + "bbox": [ + 133, + 161, + 480, + 194 + ], + "type": "text", + "content": "21. Karras, T., Aittala, M., Laine, S., Härkönen, E., Hellsten, J., Lehtinen, J., Aila, T.: Alias-free generative adversarial networks. Advances in Neural Information Processing Systems 34, 852-863 (2021)" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 132, + 194, + 480, + 226 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 194, + 480, + 226 + ], + "spans": [ + { + "bbox": [ + 132, + 194, + 480, + 226 + ], + "type": "text", + "content": "22. Lee, Y., Kim, K., Kim, H., Sung, M.: Syncdiffusion: Coherent montage via synchronized joint diffusions. Advances in Neural Information Processing Systems 36 (2024)" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 132, + 227, + 480, + 258 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 227, + 480, + 258 + ], + "spans": [ + { + "bbox": [ + 132, + 227, + 480, + 258 + ], + "type": "text", + "content": "23. Li, J., Bansal, M.: Panogen: Text-conditioned panoramic environment generation for vision-and-language navigation. ArXiv abs/2305.19195 (2023), https://api.sementicscholar.org/CorpusID:258967291" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 132, + 259, + 480, + 304 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 259, + 480, + 304 + ], + "spans": [ + { + "bbox": [ + 132, + 259, + 480, + 304 + ], + "type": "text", + "content": "24. Lin, C.H., Chang, C.C., Chen, Y.S., Juan, D.C., Wei, W., Chen, H.T.: Cocogan: Generation by parts via conditional coordinating. 2019 IEEE/CVF International Conference on Computer Vision (ICCV) pp. 4511-4520 (2019), https://api-semanticscholar.org/CorpusID:90262507" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 132, + 304, + 480, + 335 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 304, + 480, + 335 + ], + "spans": [ + { + "bbox": [ + 132, + 304, + 480, + 335 + ], + "type": "text", + "content": "25. Lin, C.H., Lee, H.Y., Cheng, Y.C., Tulyakov, S., Yang, M.H.: Infinitygan: Towards infinite-pixel image synthesis. In: International Conference on Learning Representations (2021), https://api-semanticscholar.org/CorpusID:238419701" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 132, + 336, + 480, + 369 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 336, + 480, + 369 + ], + "spans": [ + { + "bbox": [ + 132, + 336, + 480, + 369 + ], + "type": "text", + "content": "26. Liu, X., Zhang, X., Ma, J., Peng, J., et al.: Instaflow: One step is enough for high-quality diffusion-based text-to-image generation. In: The International Conference on Learning Representations (2023)" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 132, + 369, + 480, + 402 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 369, + 480, + 402 + ], + "spans": [ + { + "bbox": [ + 132, + 369, + 480, + 402 + ], + "type": "text", + "content": "27. Lu, C., Zhou, Y., Bao, F., Chen, J., Li, C., Zhu, J.: Dpm-solver: A fast ode solver for diffusion probabilistic model sampling in around 10 steps. arXiv preprint arXiv:2206.00927 (2022)" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 132, + 403, + 480, + 445 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 403, + 480, + 445 + ], + "spans": [ + { + "bbox": [ + 132, + 403, + 480, + 445 + ], + "type": "text", + "content": "28. Lugmayr, A., Danelljan, M., Romero, A., Yu, F., Timofte, R., Gool, L.V.: Repaint: Inpainting using denoising diffusion probabilistic models. 2022 IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR) pp. 11451-11461 (2022), https://api-semanticscholar.org/CorpusID:246240274" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 132, + 446, + 480, + 489 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 446, + 480, + 489 + ], + "spans": [ + { + "bbox": [ + 132, + 446, + 480, + 489 + ], + "type": "text", + "content": "29. Meng, C., He, Y., Song, Y., Song, J., Wu, J., Zhu, J.Y., Ermon, S.: Sdedit: Guided image synthesis and editing with stochastic differential equations. In: International Conference on Learning Representations (2021), https://apisemantic scholar.org/CorpusID:245704504" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 132, + 490, + 480, + 532 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 490, + 480, + 532 + ], + "spans": [ + { + "bbox": [ + 132, + 490, + 480, + 532 + ], + "type": "text", + "content": "30. Mou, C., Wang, X., Xie, L., Zhang, J., Qi, Z., Shan, Y., Qie, X.: T2i-adapter: Learning adapters to dig out more controllable ability for text-to-image diffusion models. ArXiv abs/2302.08453 (2023), https://api_semanticscholar.org/CorpusID: 256900833" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 132, + 533, + 480, + 566 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 533, + 480, + 566 + ], + "spans": [ + { + "bbox": [ + 132, + 533, + 480, + 566 + ], + "type": "text", + "content": "31. Nichol, A., Dhariwal, P., Ramesh, A., Shyam, P., Mishkin, P., McGrew, B., Sutskever, I., Chen, M.: Glide: Towards photorealistic image generation and editing with text-guided diffusion models. arXiv preprint arXiv:2112.10741 (2021)" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 132, + 567, + 480, + 609 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 567, + 480, + 609 + ], + "spans": [ + { + "bbox": [ + 132, + 567, + 480, + 609 + ], + "type": "text", + "content": "32. Oh, C.H., Cho, W., Park, D., Chae, Y., Wang, L., Yoon, K.J.: Bips: Bi-modal indoor panorama synthesis via residual depth-aided adversarial learning. ArXiv abs/2112.06179 (2021), https://api-semanticscholar.org/CorpusID: 245123664" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 132, + 610, + 480, + 632 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 610, + 480, + 632 + ], + "spans": [ + { + "bbox": [ + 132, + 610, + 480, + 632 + ], + "type": "text", + "content": "33. Peebles, W., Xie, S.: Scalable diffusion models with transformers. In: International Conference on Computer Vision. pp. 4195-4205 (2023)" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 132, + 632, + 480, + 666 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 632, + 480, + 666 + ], + "spans": [ + { + "bbox": [ + 132, + 632, + 480, + 666 + ], + "type": "text", + "content": "34. von Platen, P., Patil, S., Lozhkov, A., Cuenca, P., Lambert, N., Rasul, K., Davaadorj, M., Wolf, T.: Diffusers: State-of-the-art diffusion models. https://github.com/huggingface/diffusers (2022)" + } + ] + } + ], + "index": 17 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "spans": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "type": "text", + "content": "16" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 166, + 91, + 217, + 101 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 166, + 91, + 217, + 101 + ], + "spans": [ + { + "bbox": [ + 166, + 91, + 217, + 101 + ], + "type": "text", + "content": "A. Liu et al." + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 15 + }, + { + "para_blocks": [ + { + "bbox": [ + 130, + 116, + 482, + 665 + ], + "type": "list", + "angle": 0, + "index": 21, + "blocks": [ + { + "bbox": [ + 130, + 116, + 482, + 138 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 116, + 482, + 138 + ], + "spans": [ + { + "bbox": [ + 130, + 116, + 482, + 138 + ], + "type": "text", + "content": "35. Radford, A., Narasimhan, K., Salimans, T., Sutskever, I.: Language models are unsupervised multitask learners (2019)" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 130, + 138, + 482, + 171 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 138, + 482, + 171 + ], + "spans": [ + { + "bbox": [ + 130, + 138, + 482, + 171 + ], + "type": "text", + "content": "36. Ramesh, A., Dhariwal, P., Nichol, A., Chu, C., Chen, M.: Hierarchical text-conditional image generation with clip latents. arXiv preprint arXiv:2204.06125 (2022)" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 132, + 172, + 481, + 194 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 172, + 481, + 194 + ], + "spans": [ + { + "bbox": [ + 132, + 172, + 481, + 194 + ], + "type": "text", + "content": "37. Razavi, A., Van den Oord, A., Vinyals, O.: Generating diverse high-fidelity images with vq-vae-2. Advances in neural information processing systems 32 (2019)" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 132, + 194, + 481, + 226 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 194, + 481, + 226 + ], + "spans": [ + { + "bbox": [ + 132, + 194, + 481, + 226 + ], + "type": "text", + "content": "38. Rombach, R., Blattmann, A., Lorenz, D., Esser, P., Ommer, B.: High-resolution image synthesis with latent diffusion models. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 10684-10695 (2022)" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 132, + 227, + 481, + 270 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 227, + 481, + 270 + ], + "spans": [ + { + "bbox": [ + 132, + 227, + 481, + 270 + ], + "type": "text", + "content": "39. Sahara, C., Chan, W., Saxena, S., Li, L., Whang, J., Denton, E.L., Ghasemipour, K., Gontijo Lopes, R., Karagol Ayan, B., Salimans, T., et al.: Photorealistic text-to-image diffusion models with deep language understanding. Advances in Neural Information Processing Systems (NeurIPS) (2022)" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 132, + 271, + 481, + 303 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 271, + 481, + 303 + ], + "spans": [ + { + "bbox": [ + 132, + 271, + 481, + 303 + ], + "type": "text", + "content": "40. Sohl-Dickstein, J., Weiss, E., Maheswaranathan, N., Ganguli, S.: Deep unsupervised learning using nonequilibrium thermodynamics. In: International Conference on Machine Learning. PMLR (2015)" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 132, + 304, + 481, + 325 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 304, + 481, + 325 + ], + "spans": [ + { + "bbox": [ + 132, + 304, + 481, + 325 + ], + "type": "text", + "content": "41. Song, J., Meng, C., Ermon, S.: Denoising diffusion implicit models. arXiv preprint arXiv:2010.02502 (2020)" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 132, + 326, + 481, + 346 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 326, + 481, + 346 + ], + "spans": [ + { + "bbox": [ + 132, + 326, + 481, + 346 + ], + "type": "text", + "content": "42. Song, Y., Ermon, S.: Generative modeling by estimating gradients of the data distribution. Advances in neural information processing systems (NeurIPS) (2019)" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 132, + 347, + 481, + 369 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 347, + 481, + 369 + ], + "spans": [ + { + "bbox": [ + 132, + 347, + 481, + 369 + ], + "type": "text", + "content": "43. Song, Y., Ermon, S.: Generative modeling by estimating gradients of the data distribution. Advances in neural information processing systems 32 (2019)" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 132, + 369, + 481, + 391 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 369, + 481, + 391 + ], + "spans": [ + { + "bbox": [ + 132, + 369, + 481, + 391 + ], + "type": "text", + "content": "44. Song, Y., Ermon, S.: Improved techniques for training score-based generative models. Advances in neural information processing systems 33, 12438-12448 (2020)" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 132, + 392, + 481, + 423 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 392, + 481, + 423 + ], + "spans": [ + { + "bbox": [ + 132, + 392, + 481, + 423 + ], + "type": "text", + "content": "45. Song, Y., Sohl-Dickstein, J.N., Kingma, D.P., Kumar, A., Ermon, S., Poole, B.: Score-based generative modeling through stochastic differential equations. ICLR (2021)" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 132, + 424, + 481, + 466 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 424, + 481, + 466 + ], + "spans": [ + { + "bbox": [ + 132, + 424, + 481, + 466 + ], + "type": "text", + "content": "46. Tang, S., Zhang, F., Chen, J., Wang, P., Furukawa, Y.: Mvdiffusion: Enabling holistic multi-view image generation with correspondence-aware diffusion. ArXiv abs/2307.01097 (2023), https://api_semanticscholar.org/CorpusID: 259316427" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 132, + 468, + 481, + 512 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 468, + 481, + 512 + ], + "spans": [ + { + "bbox": [ + 132, + 468, + 481, + 512 + ], + "type": "text", + "content": "47. Teterwak, P., Sarna, A., Krishnan, D., Maschinot, A., Belanger, D., Liu, C., Freeman, W.T.: Boundless: Generative adversarial networks for image extension. 2019 IEEE/CVF International Conference on Computer Vision (ICCV) pp. 10520-10529 (2019), https://api_semanticscholar.org/CorpusID:201106503" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 132, + 513, + 481, + 533 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 513, + 481, + 533 + ], + "spans": [ + { + "bbox": [ + 132, + 513, + 481, + 533 + ], + "type": "text", + "content": "48. Van Den Oord, A., Vinyals, O., et al.: Neural discrete representation learning. Advances in neural information processing systems 30 (2017)" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 132, + 534, + 481, + 555 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 534, + 481, + 555 + ], + "spans": [ + { + "bbox": [ + 132, + 534, + 481, + 555 + ], + "type": "text", + "content": "49. Voynov, A., Hertz, A., Arar, M., Fruchter, S., Cohen-Or, D.: Anylens: A generative diffusion model with any rendering lens (2023)" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 132, + 555, + 481, + 588 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 555, + 481, + 588 + ], + "spans": [ + { + "bbox": [ + 132, + 555, + 481, + 588 + ], + "type": "text", + "content": "50. Wang, G., Yang, Y., Loy, C.C., Liu, Z.: Stylelight: Hdr panorama generation for lighting estimation and editing. In: European Conference on Computer Vision (2022), https://api_semanticscholar.org/CorpusID:251196614" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 132, + 589, + 481, + 621 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 589, + 481, + 621 + ], + "spans": [ + { + "bbox": [ + 132, + 589, + 481, + 621 + ], + "type": "text", + "content": "51. Wang, H., Xiang, X., Fan, Y., Xue, J.H.: Customizing 360-degree panoramicas through text-to-image diffusion models. ArXiv abs/2310.18840 (2023), https://api.sementicscholar.org/CorpusID:264590753" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 132, + 622, + 481, + 643 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 622, + 481, + 643 + ], + "spans": [ + { + "bbox": [ + 132, + 622, + 481, + 643 + ], + "type": "text", + "content": "52. Wu, T., Zheng, C., Cham, T.J.: Panodiffusion: 360-degree panorama outpainting via diffusion (2023), https://api-semanticscholar.org/CorpusID:259360663" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 132, + 644, + 481, + 665 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 644, + 481, + 665 + ], + "spans": [ + { + "bbox": [ + 132, + 644, + 481, + 665 + ], + "type": "text", + "content": "53. Zhang, L., Rao, A., Agrawala, M.: Adding conditional control to text-to-image diffusion models (2023)" + } + ] + } + ], + "index": 20 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 408, + 91, + 447, + 100 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 408, + 91, + 447, + 100 + ], + "spans": [ + { + "bbox": [ + 408, + 91, + 447, + 100 + ], + "type": "text", + "content": "PanoFree" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 470, + 91, + 481, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 470, + 91, + 481, + 100 + ], + "spans": [ + { + "bbox": [ + 470, + 91, + 481, + 100 + ], + "type": "text", + "content": "17" + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 16 + }, + { + "para_blocks": [ + { + "bbox": [ + 132, + 116, + 481, + 205 + ], + "type": "list", + "angle": 0, + "index": 4, + "blocks": [ + { + "bbox": [ + 132, + 116, + 481, + 160 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 116, + 481, + 160 + ], + "spans": [ + { + "bbox": [ + 132, + 116, + 481, + 160 + ], + "type": "text", + "content": "54. Zhang, Q., Song, J., Huang, X., Chen, Y., Liu, M.Y.: Diffcollage: Parallel generation of large content with diffusion models. 2023 IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR) pp. 10188-10198 (2023), https://api-semanticscholar.org/CorpusID:257834007" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 132, + 161, + 481, + 205 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 161, + 481, + 205 + ], + "spans": [ + { + "bbox": [ + 132, + 161, + 481, + 205 + ], + "type": "text", + "content": "55. Zhang, R., Isola, P., Efros, A.A., Shechtman, E., Wang, O.: The unreasonable effectiveness of deep features as a perceptual metric. 2018 IEEE/CVF Conference on Computer Vision and Pattern Recognition pp. 586-595 (2018), https://apisemantic scholar.org/CorpusID:4766599" + } + ] + } + ], + "index": 3 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "spans": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "type": "text", + "content": "18" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 166, + 91, + 217, + 101 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 166, + 91, + 217, + 101 + ], + "spans": [ + { + "bbox": [ + 166, + 91, + 217, + 101 + ], + "type": "text", + "content": "A. Liu et al." + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 17 + } + ], + "_backend": "vlm", + "_version_name": "2.6.4" +} \ No newline at end of file diff --git a/2024/PanoVOS_ Bridging Non-panoramic and Panoramic Views with Transformer for Video Segmentation/5c945af1-cf63-4f10-9827-c4fd88e7d8a2_content_list.json b/2024/PanoVOS_ Bridging Non-panoramic and Panoramic Views with Transformer for Video Segmentation/5c945af1-cf63-4f10-9827-c4fd88e7d8a2_content_list.json new file mode 100644 index 0000000000000000000000000000000000000000..0f74382b5cb968daf6dd7cd659f57781985ff8b1 --- /dev/null +++ b/2024/PanoVOS_ Bridging Non-panoramic and Panoramic Views with Transformer for Video Segmentation/5c945af1-cf63-4f10-9827-c4fd88e7d8a2_content_list.json @@ -0,0 +1,1904 @@ +[ + { + "type": "text", + "text": "PanoVOS: Bridging Non-panoramic and Panoramic Views with Transformer for Video Segmentation", + "text_level": 1, + "bbox": [ + 271, + 140, + 733, + 209 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Shilin Yan $^{1}$ , Xiaohao Xu $^{2}$ , Renrui Zhang $^{3}$ , Lingyi Hong $^{1}$ , Wenchao Chen $^{1}$ , Wenqiang Zhang $^{1}$ , and Wei Zhang $^{1\\star}$", + "bbox": [ + 292, + 234, + 709, + 267 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "$^{1}$ Shanghai Key Lab of Intelligent Information Processing, School of Computer Science, Fudan University $^{2}$ University of Michigan, Ann Arbor $^{3}$ MMLab CUHK $^{4}$ tattoo.ysl@mail.com weizh@fudan.edu.cn", + "bbox": [ + 305, + 276, + 696, + 347 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/b65bc70a30bd22450f37c9d5d3d5f285f3e04f37f97eb5930cc855143aea5ca7.jpg", + "image_caption": [ + "Fig.1: Panoramic video object segmentation (PanoVOS). PanoVOS targets tracking and distinguishing the particular instances under content discontinuities (e.g. penguin in the image of $T = 15$ ) and serve distortion (e.g. penguin in the image of $T = 65$ ). We show the sample of (a) frames, (b) segmentation annotations, and (c) area proportion of foreground for the Penguin video in our dataset." + ], + "image_footnote": [], + "bbox": [ + 243, + 359, + 781, + 458 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Abstract. Panoramic videos contain richer spatial information and have attracted tremendous amounts of attention due to their exceptional experience in some fields such as autonomous driving and virtual reality. However, existing datasets for video segmentation only focus on conventional planar images. To address the challenge, in this paper, we present a panoramic video dataset, i.e., PanoVOS. The dataset provides 150 videos with high video resolutions and diverse motions. To quantify the domain gap between 2D planar videos and panoramic videos, we evaluate 15 off-the-shelf video object segmentation (VOS) models on PanoVOS. Through error analysis, we found that all of them fail to tackle pixel-level content discontinues of panoramic videos. Thus, we present a Panoramic Space Consistency Transformer (PSCFormer), which can effectively utilize the semantic boundary information of the previous frame for pixel-level matching with the current frame. Extensive experiments demonstrate that compared with the previous SOTA models, our PSCFormer network exhibits a great advantage in terms of segmentation results under the panoramic setting. Our dataset poses new challenges in panoramic VOS and we hope that", + "bbox": [ + 259, + 565, + 743, + 816 + ], + "page_idx": 0 + }, + { + "type": "page_footnote", + "text": "* Corresponding author.", + "bbox": [ + 218, + 825, + 383, + 839 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "our PanoVOS can advance the development of panoramic segmentation/tracking. The dataset, codes, and pre-train models will be published at https://github.com/shilinyan99/PanoVOS.", + "bbox": [ + 259, + 147, + 743, + 189 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "1 Introduction", + "text_level": 1, + "bbox": [ + 215, + 217, + 375, + 233 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Semi-supervised video object segmentation (VOS) [52], which targets tracking and distinguishing the particular instances across the entire video sequence based on the first frame masks, plays an essential role in video understanding and editing. Conventionally, the images or videos studied in VOS are 2D planar data with a limited Field of View (FoV), which may lead to some ambiguities, especially when objects are out of view. Meanwhile, with the rapid development of VR/AR collection devices [12,22], panoramic videos with a $360^{\\circ} \\times 180^{\\circ}$ FoV are able to collect the entire viewing sphere and richer spatial information [1,21, 27, 60]. To the best of our knowledge, we are the first to attempt to tackle the promising but challenging task of panoramic video object segmentation.", + "bbox": [ + 212, + 250, + 787, + 402 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "To foster the development of panoramic VOS, we propose a new dataset in this work, aiming at panoramic video object segmentation. The dataset contains a wide range of real-world scenarios in which scenes have a large magnitude of motion. The main characteristics of our dataset are three aspects. 1) Panoramic videos bring certain advantages (richer geometric information and wider FoV) in real-world applications as well as challenges (serve distortion and content discontinuities). 2) Compared to all existing VOS datasets, our dataset has longer video clips with an average length of 20 seconds. 3) Nearly half of the video resolutions in our dataset are $4K$ , which may help facilitate broader video tracking/segmentation research under the high-resolution scenario.", + "bbox": [ + 212, + 402, + 787, + 551 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "In the proposed dataset, we annotated 150 videos with 19,145 annotated instance masks, including sports (e.g. parkour, skateboard), animals (e.g. elephant, monkey), and common objects (basketball, hot balloon). Since, annotating a pixel-level intensive task is very time-consuming and expensive, we proposed a semi-supervised human-computer joint annotation strategy. Concretely, we first annotated objects at selected keyframes (1 fps) Then we adopted the state-of-the-art video object segmentation model AOT [58] for mask propagation to the rest frames of videos and we manually refine parts of them.", + "bbox": [ + 212, + 553, + 787, + 672 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Then, we conducted extensive experiments on PanoVOS to evaluate 15 off-the-shelf video object segmentation models. The results suggest that existing approaches can not handle several domain-unique challenges. The first is content discontinuities, which means the foreground object may be separated in the left and the right boundaries of the planar image, such as the case in the image of $T = 15$ in Fig. 1. The second is the severe distortions and deformations, such as the case in the image of $T = 65$ in Fig. 1.", + "bbox": [ + 212, + 672, + 787, + 779 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "To tackle these challenges of panoramic video segmentation, we proposed a PSC-Former model which consists of key component Panoramic Space Consistent (PSC) blocks. The PSC block is designed for constructing spatial-temporal class-agnostic correspondence and propagating the segmentation masks. Each PSC", + "bbox": [ + 214, + 779, + 787, + 840 + ], + "page_idx": 1 + }, + { + "type": "page_number", + "text": "2", + "bbox": [ + 217, + 114, + 228, + 126 + ], + "page_idx": 1 + }, + { + "type": "header", + "text": "Yan et al.", + "bbox": [ + 271, + 114, + 339, + 127 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "block utilizes a cross-attention for matching with references' embeddings and a PSC-attention for modeling the boundary semantic relationship between the previous frame and the query frame. Hence, the network can effectively alleviate the problem that the left and the right boundaries are actually continuous in panoramic videos. Our method outperforms the SOTA models that are re-trained on PanoVOS train set in segmentation quality under the panoramic setting. Our contributions are three-fold.", + "bbox": [ + 212, + 146, + 787, + 252 + ], + "page_idx": 2 + }, + { + "type": "list", + "sub_type": "text", + "list_items": [ + "- We introduce a panoramic video object segmentation dataset (PanoVOS) with 150 videos and 19K annotated instance masks, which fills the gap of long-term instance-level annotated panoramic video segmentation datasets.", + "- Extensive experiments are conducted on 15 off-the-shelf VOS benchmarks and our baseline model on PanoVOS, which reveals that current methods could not tackle content discontinuities in panoramic videos well.", + "- We propose a Panoramic Space Consistency Transformer (PSCFormer) on PanoVOS that successfully resolves the challenges of discontinuity of pixel-level content segmentation." + ], + "bbox": [ + 223, + 263, + 785, + 397 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "2 Related Work", + "text_level": 1, + "bbox": [ + 215, + 421, + 387, + 436 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "2.1 Panoramic Datasets", + "text_level": 1, + "bbox": [ + 215, + 453, + 426, + 467 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "In this paper, panoramic videos refers to complete (360°, full view) panoramic videos, which is different from the definition in [38], which only include wide but partial views of some range-view images collected from multiple cameras.", + "bbox": [ + 212, + 478, + 784, + 521 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Image-based panoramic datasets. Existing popular image-level panoramic segmentation datasets are Stanford2D3D [2] and DensePASS [35]. The former one is mainly focused on indoor spaces including a total of 1,413 panoramic images with instance-level annotations in 13 categories. The latter targets driving scenes in cities. DensePASS [35] provides only 100 labeled panoramic images for testing and 2,000 unlabeled images for cross-domain transfer optimization.", + "bbox": [ + 212, + 523, + 784, + 611 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Video-based panoramic datasets. Video-based benchmarks mainly include SHD360 [62], SOD360 [64] and Wild360 [7]. All of them are used for panoramic video saliency object detection. Specifically, 1) SHD360 only targets human-centric video scenes with little movement. It provides 6,268 object-level pixel-wise masks and 16,238 instance-level pixel-wise masks. 2) SOD360 focuses on the sports-centric scenario with 41 video clips (12 outdoor and 29 indoor). 3) Wild360 concentrates on natural scenes with 85 videos. Note that SOD360 and Wild360 have no object-level or instance-level annotations.", + "bbox": [ + 212, + 613, + 784, + 732 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "We make a comparison with the existing video panoramic datasets in Table 1. Specifically, our PanoVOS dataset contains 150 videos mainly from three different domains: person, animal, and common object, which makes the dataset more general for object-agnostic evaluations. Besides, videos in our dataset have a relatively large range of motion, making our PanoVOS dataset suitable for video tracking and segmentation evaluation tasks under panoramic scenes. Moreover, the average duration of each video in our dataset is $20s$ , which is about 4 times", + "bbox": [ + 212, + 734, + 784, + 838 + ], + "page_idx": 2 + }, + { + "type": "header", + "text": "PanoVOS: Bridging Non-panoramic and Panoramic Views", + "bbox": [ + 338, + 114, + 727, + 128 + ], + "page_idx": 2 + }, + { + "type": "page_number", + "text": "3", + "bbox": [ + 774, + 116, + 784, + 126 + ], + "page_idx": 2 + }, + { + "type": "table", + "img_path": "images/01fc6a7856dfb02c3781ad1dafadffce5543acd5b7cde876f9e4182e20e526de.jpg", + "table_caption": [], + "table_footnote": [], + "table_body": "
DatasetsMotion#Videos#Frames#Total MasksAverage Duration
SHD360 [62]Small416,26816,2385s
SOD360 [64]Large104N/A0N/A
Wild360 [7]Large85N/A0N/A
PanoVOSLarge15013,99519,14520s
", + "bbox": [ + 290, + 143, + 709, + 244 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Table 1: Comparison of panoramic video datasets. Our PanoVOS is the first long-term panoramic video segmentation dataset with instance-level masks. Compared with existing panoramic video datasets [7,62,64] that are used for saliency detection, our panoramic video dataset for video segmentation, i.e., PanoVOS, includes more diverse and larger motion, making it suitable for dense video tracking evaluation.", + "bbox": [ + 212, + 255, + 787, + 325 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "longer than SHD360 [62] (5s per video). By the way, the longer video is highlighted in a recent survey [49]. The longer the video, the more likely it is to introduce more panoramic video characteristics such as distortion and discontinuity, which is more challenging and more practical.", + "bbox": [ + 212, + 354, + 787, + 416 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "2.2 Video Object Segmentation Datasets", + "text_level": 1, + "bbox": [ + 214, + 436, + 568, + 452 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "The establishment of DAVIS [41, 42] and YouTube-VOS [52] datasets pave the way for the boosting development of VOS methods. They are collected by traditional pinhole cameras and the duration of each video clip is very short, only 5s on average. In contrast, the average video length in the proposed PanoVOS dataset is 20s, which is 4 times longer than the existing video datasets. Our dataset includes more challenging scenes (e.g. distortion and discontinuity) that is non-negligible in real-world applications.", + "bbox": [ + 212, + 460, + 787, + 566 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "2.3 Video Object Segmentation Methods", + "text_level": 1, + "bbox": [ + 214, + 588, + 568, + 604 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Existing video object segmentation methods can be roughly classified into three subsets: online-learning-based, propagation-based, and matching-based.", + "bbox": [ + 212, + 613, + 785, + 643 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Online learning-based. Online learning-based approaches [3,36,50], which either train or fine-tune their networks with the first-frame ground truth at test time and are therefore a great waste of resources. OnAVOS [47] achieves promising results by introducing an online adaptation mechanism, but it still requires online fine-tuning. To a certain extent, it restricts networks' efficiency.", + "bbox": [ + 212, + 643, + 785, + 718 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Propagation-based. Propagation-based models [4,8,39] get the target masks in a frame-to-frame prerogation way. Although propagation-based methods improve efficiency, they lack long-term context and therefore are difficult to handle object disappearance and reappearance, severe obscuration, and distortion.", + "bbox": [ + 212, + 719, + 785, + 779 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Matching-based. Matching-based methods [6, 10, 11, 14, 16, 25, 26, 31, 37, 40, 53, 54, 63] aim to learn an embedding space of target objects between query and memory. Recently state-of-the-art methods encode many frames into embeddings and store them as a feature memory bank. The most representative is STM [40],", + "bbox": [ + 212, + 779, + 785, + 840 + ], + "page_idx": 3 + }, + { + "type": "page_number", + "text": "4", + "bbox": [ + 217, + 114, + 228, + 126 + ], + "page_idx": 3 + }, + { + "type": "header", + "text": "Yan et al.", + "bbox": [ + 271, + 114, + 339, + 127 + ], + "page_idx": 3 + }, + { + "type": "image", + "img_path": "images/9f598b6b7a822982a3a0723c6809c280df59c6b5aa85b81de402b40d4a2c42d2.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 220, + 143, + 330, + 191 + ], + "page_idx": 4 + }, + { + "type": "image", + "img_path": "images/ea645df97cdbf5de7919a92a916cc7cf1dc91bfaf9d2b53ce04b066645e454a9.jpg", + "image_caption": [ + "Fig. 2: PanoVOS dataset. We select 10 samples from the dataset involving major scenes. For each video, there are high-quality instance-level pixel-wise masks." + ], + "image_footnote": [], + "bbox": [ + 220, + 195, + 330, + 239 + ], + "page_idx": 4 + }, + { + "type": "image", + "img_path": "images/b1eb488d17e5b0bda1eefbc9bf0f4c163341a3680d52acd78d06cc8e85c60134.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 331, + 143, + 442, + 193 + ], + "page_idx": 4 + }, + { + "type": "image", + "img_path": "images/743cea0236f2b529412fa835efc5a45de1e3e9108d9594a2ae9d859b8413b313.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 333, + 195, + 442, + 239 + ], + "page_idx": 4 + }, + { + "type": "image", + "img_path": "images/001a3c1c54c107ee62c30e8b0967b8c9140dca22575ab534f8cdc0e459341807.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 444, + 143, + 555, + 191 + ], + "page_idx": 4 + }, + { + "type": "image", + "img_path": "images/562a3ed9d72a4ee8e0fccf8445f742d4d3c9b391724fcf312de8fcee7a305d15.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 444, + 195, + 557, + 239 + ], + "page_idx": 4 + }, + { + "type": "image", + "img_path": "images/69f21cea20fdc4cc0ee8e1dd102b1de00304555e80ba4f09cf21677b311671f4.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 557, + 143, + 668, + 191 + ], + "page_idx": 4 + }, + { + "type": "image", + "img_path": "images/05e4a16e9f6a76b30ad78fb01c6b15e3aeff2b93ea389258578199259f564b42.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 558, + 195, + 666, + 239 + ], + "page_idx": 4 + }, + { + "type": "image", + "img_path": "images/d9ee8a8bbe976ef7a549c1d34a45b73112bcc182363e6857d87f06a41f61bcf4.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 669, + 147, + 779, + 191 + ], + "page_idx": 4 + }, + { + "type": "image", + "img_path": "images/3d76807a3ce47afba0c1a90987e61102f800ee9f5bb84839536180a29490442c.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 669, + 195, + 782, + 239 + ], + "page_idx": 4 + }, + { + "type": "table", + "img_path": "images/f06b6935522d4104dad9db018b0fa405448958f3ad3c30bbe0ed977120f68e8b.jpg", + "table_caption": [], + "table_footnote": [], + "table_body": "
SplitsTrainValTest
#Videos8035 (10 unseen)35 (10 unseen)
#Images7,070(50.5%)3,464(24.8%)3,461(24.7%)
#Masks9,585(50.1%)4,957(25.9%)4,603(24.0%)
", + "bbox": [ + 333, + 301, + 669, + 367 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Table 2: Statistics of PanoVOS dataset", + "bbox": [ + 349, + 377, + 650, + 391 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "which has been extended to many works [5, 19, 34, 44, 48, 51]. AOT [58] introduces an identification mechanism by encoding multiple targets into the same embedding space, which can simultaneously segment multiple objects. However, they fail to address the challenges of the tremendous proportion of distortion and discontinuity under panoramic setting.", + "bbox": [ + 212, + 420, + 784, + 497 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "3 PanoVOS Dataset", + "text_level": 1, + "bbox": [ + 214, + 518, + 426, + 535 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "We introduced the proposed PanoVOS dataset in three parts, (1) collection process, (2) statistical summary, and (3) annotation pipeline.", + "bbox": [ + 212, + 550, + 784, + 582 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "3.1 Data collection", + "text_level": 1, + "bbox": [ + 214, + 603, + 385, + 616 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "We built our PanoVOS dataset with the principle of diversity in mind. Moreover, the objects in the video should have a large amplitude of motion or camera movement. Based on the above viewpoint, we collected videos from the YouTube website for further annotation, respectively. The range of the video length is from 3 to 40 seconds. The average sequence length of each video in the dataset is approximately 20 seconds. We followed the settings of YouTube-VOS [52] to sample the frames at 6 fps.", + "bbox": [ + 212, + 626, + 784, + 733 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "3.2 Dataset Statistics", + "text_level": 1, + "bbox": [ + 214, + 753, + 408, + 768 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "PanoVOS contains 150 videos, including 13,995 frames and 19,145 instance annotations from 35 categories. The average length of each video is 20 seconds. We believe that visual categories are representative of common life scenarios, and Fig.2 shows some samples of PanoVOS. To create our PanoVOS, in the", + "bbox": [ + 212, + 779, + 784, + 840 + ], + "page_idx": 4 + }, + { + "type": "header", + "text": "PanoVOS: Bridging Non-panoramic and Panoramic Views", + "bbox": [ + 336, + 114, + 727, + 128 + ], + "page_idx": 4 + }, + { + "type": "page_number", + "text": "5", + "bbox": [ + 774, + 116, + 784, + 126 + ], + "page_idx": 4 + }, + { + "type": "image", + "img_path": "images/66c37404571ed1221551acfb4db3f3ee0b2f465bbc13a548d08f8438d84408fb.jpg", + "image_caption": [ + "Fig. 3: Instance-level distribution of PanoVOS dataset. Our dataset contains three major divisions: person, animals, and common objects with 35 sub-divisions." + ], + "image_footnote": [], + "bbox": [ + 263, + 145, + 743, + 291 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "spirit of the video object segmentation task, we carefully selected videos with relatively large motion amplitudes and chose a set of video categories including person (e.g. parkour, dance, BMX, skateboard), animals (e.g. elephant, monkey, giraffe, rhino, birds) and common objects (e.g. basketball, hot balloon) as shown in Fig 3. PanoVOS dataset consists of 150 videos split into training (80), validation (35), and test (35) sets. Table 2 shows detailed division results. Both the validation and test sets have 35 videos (about $23\\%$ of the frames and the masks). For validation and test sets, we keep some unseen visual categories for generalization ability evaluation.", + "bbox": [ + 212, + 364, + 789, + 503 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "3.3 Annotation Pipeline", + "text_level": 1, + "bbox": [ + 215, + 527, + 429, + 544 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Annotation is very time-consuming and expensive for a pixel-level panoramic segmentation dataset. To obtain accurate large-scale video panoramic segmentation annotations and make the process more efficient, we propose a semi-automatic human-computer joint annotation strategy, as shown in Fig 4. First, keyframes are selected and manually annotated for each video, which are images with a speed of 1 fps. This is followed by a frame-by-frame propagation from the annotated keyframes to those unlabeled intermediate frames with a sophisticated semi-supervised VOS model. Then, to tackle the distortions and discontinuities in panoramic videos, we need to re-calibrate the resulting annotations via human refinement. More details will unfold below.", + "bbox": [ + 212, + 556, + 787, + 708 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Annotation Propagation For the annotation of each video, we first need an expert to browse the current video and note down all objects that have a large amplitude of movement. Then, for each video, the recorded objects in keyframes with a speed of 1 fps are selected for manual annotation. To avoid consistency errors or the problem of objects being labeled as other instances when they disappear and reappear, another expert needs to double-check the annotations of all objects to improve the accuracy of the dataset annotation.", + "bbox": [ + 212, + 734, + 787, + 840 + ], + "page_idx": 5 + }, + { + "type": "page_number", + "text": "6", + "bbox": [ + 217, + 114, + 228, + 126 + ], + "page_idx": 5 + }, + { + "type": "header", + "text": "Yan et al.", + "bbox": [ + 271, + 114, + 339, + 126 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/4dec3a67269029bede7f37571e681b2d68e66e02b8dbbe1df1c13fa14295ec52.jpg", + "image_caption": [ + "Fig. 4: PanoVOS annotation pipeline. Our annotation pipeline includes two phases. (1) The first phase is called Key Frames Select and Annotate. The annotator browses the video and picks out the object to be annotated. Then, instances are manually annotated at 1 fps and corrected by another annotator. (2) The second phase is called All Frames Propagate and Refine. In this phase, we apply a semi-supervised video object segmentation model to help propagate the annotated masks and the generated instances are refined by annotators." + ], + "image_footnote": [], + "bbox": [ + 217, + 143, + 787, + 256 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "We then use the off-the-shelf video labeling method [58] to propagate the instance masks frame by frame from the annotated keyframes to untagged intermediate frames and generate masks at 6 fps.", + "bbox": [ + 212, + 393, + 787, + 439 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Annotation Refinement To present a new Panoramic dataset of high quality. After obtaining masks of the first propagation stage, annotators are asked to check the quality of the masks and refine them. The main amendments are in the following two areas. 1). Since our video resolution is generally relatively high, the propagation method will often fail when encountering complex videos with many small objects in a scene. 2) Due to the huge distortions and discontinuities present in the panoramic video, the quality of the masks obtained is relatively poor. Manual correction of the mask is checked by another annotator until the result is satisfactory before proceeding to the next video annotation.", + "bbox": [ + 212, + 464, + 787, + 602 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "4 Method", + "text_level": 1, + "bbox": [ + 215, + 626, + 330, + 642 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "4.1 Overview", + "text_level": 1, + "bbox": [ + 215, + 660, + 341, + 674 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Video object segmentation targets assigning an instance label to every pixel in the given video sequence based on the first frame mask. Recent works [5,6,13,59] have demonstrated that the attention mechanism can significantly help improve the segmentation performance. However, for the challenge of content discontinuation in panoramic videos, only considering the original attention mechanism will not be able to fully utilize the semantic information on the left and right boundaries (pixel contiguity) in the spatial dimension and will lose valuable contextual information when segmenting objects. Therefore, in this work, our mission is to design an effective network architecture, which can help acquire valuable boundary relationships.", + "bbox": [ + 212, + 688, + 787, + 840 + ], + "page_idx": 6 + }, + { + "type": "header", + "text": "PanoVOS: Bridging Non-panoramic and Panoramic Views", + "bbox": [ + 338, + 114, + 727, + 128 + ], + "page_idx": 6 + }, + { + "type": "page_number", + "text": "7", + "bbox": [ + 774, + 116, + 784, + 126 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/20e411bca2ba711233ee86b4077a06ef30b8f2cf913068acf309d6179ff7e29c.jpg", + "image_caption": [ + "Fig. 5: (a) PSCFormer overview. Given the query frame $\\mathbf{x}_t$ and reference frames $\\{\\mathbf{x}_i | i \\in \\mathcal{R}\\}$ , the goal of VOS is to delineate objects from the background by generating mask $\\mathbf{y}_t$ for query frame $\\mathbf{x}_t$ . References and the query frame are encoded by the memory encoder and query encoder, respectively. Multiple stacking panoramic space consistency (PSC) blocks are used to leverage the correspondence in the panoramic space between references and the query frame. A decoder is used for generating the prediction of the query frame. (b) Panoramic space consistency block architecture details." + ], + "image_footnote": [], + "bbox": [ + 264, + 152, + 738, + 335 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Fig. 5(a) illustrates the overall architecture of the proposed network. Given the query frame $\\mathbf{x}_t$ and references $\\{\\mathbf{x}_i | i \\in \\mathcal{R}\\}$ , the goal of VOS is to delineate objects from the background by generating mask $\\mathbf{y}_t$ for query frame $\\mathbf{x}_t$ . Following [57], our basic setting uses the first and previous frame as references $\\mathcal{R} = \\{1, t - 1\\}$ . The memory encoder and query encoder are responsible for extracting frame-level features. After this, the panoramic space consistency block takes them as input and aggregates the spatial-temporal information between the reference frames and the query frame at the pixel level. Finally, the decoder uses the output of the sequence stacking PSC blocks to predict the mask of the object.", + "bbox": [ + 212, + 472, + 787, + 609 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "4.2 Panoramic Space Consistency Block", + "text_level": 1, + "bbox": [ + 214, + 632, + 558, + 648 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Fig. 5(b) shows the structure of a PSC block. Motivated by the common transformer blocks [46], PSC firstly contains a self-attention layer, which is used to aggregate the target objects' correlation information within the query frame. Then, the middle module is composed of cross-attention and PSC-attention, in which cross-attention is responsible for learning the target objects' information from references $\\mathcal{R}$ and the PSC-attention targets on exploring the boundary relationship between the query frame and previous frame. Finally, PSC employs a two-layer feed-forward MLP with GELU [17] non-linearity activation function.", + "bbox": [ + 212, + 657, + 787, + 777 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Panoramic Space Consistency Attention (PSC-Attn). PSC-Attn is employed to model the spatial-temporal relationship between the query frame and reference frames considering the continuity of pixels of images in the panoramic space. How to establish a connection between the left and right boundaries be", + "bbox": [ + 212, + 779, + 787, + 839 + ], + "page_idx": 7 + }, + { + "type": "page_number", + "text": "8", + "bbox": [ + 217, + 114, + 228, + 126 + ], + "page_idx": 7 + }, + { + "type": "header", + "text": "Yan et al.", + "bbox": [ + 271, + 114, + 339, + 126 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "come especially important? The most intuitive solution would be to directly splice in length, but this would lead to a huge amount of computation. Therefore, we take the approach of moving a portion of the region in the length dimension from the right boundary to the leftmost boundary for stitching. Consequently, we only focus on the left and right boundaries between the query frame and the reference frame. Thus, unlike the original attention, where each query token is counted for attention along with all key tokens in the reference frame, our PSC attention takes care of the key tokens in a fixed window size. In particular, we define the reference frame feature embedding $\\mathbf{f}(\\mathbf{x})\\in \\mathbb{R}^{H\\times W\\times C}$ , which is extracted from the query encoder. $H$ , $W$ , and $C$ represent the height, width, and channel dimensions, respectively. According to the solutions mentioned above, the new feature embedding $\\mathbf{f}(\\mathbf{x})'$ is calculated as follows:", + "bbox": [ + 212, + 146, + 787, + 328 + ], + "page_idx": 8 + }, + { + "type": "equation", + "text": "\n$$\n\\mathbf {f} (\\mathbf {x}) ^ {\\prime} [ 0: W / p ] = \\mathbf {f} (\\mathbf {x}) [ W / p: W ]\n$$\n", + "text_format": "latex", + "bbox": [ + 397, + 343, + 609, + 357 + ], + "page_idx": 8 + }, + { + "type": "equation", + "text": "\n$$\n\\mathbf {f} (\\mathbf {x}) ^ {\\prime} [ W / p: W ] = \\mathbf {f} (\\mathbf {x}) [ 0: W / p ] \\tag {1}\n$$\n", + "text_format": "latex", + "bbox": [ + 388, + 359, + 784, + 376 + ], + "page_idx": 8 + }, + { + "type": "equation", + "text": "\n$$\n\\mathbf {f} (\\mathbf {x}) ^ {\\prime} [ W / p: W - W / p ] = \\mathbf {f} (\\mathbf {x}) [ W / p: W - W / p ],\n$$\n", + "text_format": "latex", + "bbox": [ + 339, + 378, + 663, + 393 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "where $p \\in \\mathbb{Z}^+$ . We define query embedding $Q \\in \\mathbb{R}^{HW \\times C}$ , key embedding $K \\in \\mathbb{R}^{HW \\times C}$ , value embedding $V \\in \\mathbb{R}^{HW \\times C}$ , where $Q$ is from the query frame feature embedding, $K$ and $V$ are from $\\mathbf{f}(\\mathbf{x})'$ by performing dimensional transformations. Mathematically, we define the PSC attention as follows,", + "bbox": [ + 214, + 407, + 787, + 469 + ], + "page_idx": 8 + }, + { + "type": "equation", + "text": "\n$$\n\\operatorname {P S C A t t n} (Q, K, V) = \\operatorname {s o f t m a x} \\left(\\frac {Q K ^ {T} \\mathbf {R}}{\\sqrt {C}}\\right) V, \\tag {2}\n$$\n", + "text_format": "latex", + "bbox": [ + 344, + 481, + 785, + 516 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "where $\\mathbf{R} \\in [0,1]^{HW \\times HW}$ means a window that represents the attention range of each query token. For query $Q_{(x,y)}$ at $(x,y)$ position, we define the $\\mathbf{R}_{(x,y)}$ as:", + "bbox": [ + 214, + 529, + 787, + 565 + ], + "page_idx": 8 + }, + { + "type": "equation", + "text": "\n$$\n\\mathbf {R} _ {x, y} (i, j) = \\left\\{ \\begin{array}{l l} 1 & \\text {i f} (x - i) ^ {2} \\leqslant s ^ {2} \\text {a n d} (y - j) ^ {2} \\leqslant s ^ {2} \\\\ 0 & \\text {o t h e r w i s e} \\end{array} , \\right. \\tag {3}\n$$\n", + "text_format": "latex", + "bbox": [ + 310, + 578, + 785, + 617 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "where $(i,j)$ is the position for each key token, $s$ is the window size. For each query token, it calculates the attention with another key token only if they are spatially limited to a $(2\\times s + 1)$ size window, which significantly reduces the time complexity from $(h\\times w)^{2}$ to $(2\\times s + 1)^{2}$ .", + "bbox": [ + 212, + 631, + 787, + 691 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "Following [46], we implement the representational form of our PSCAttn module with multi-headed attention, defined mathematically as follows,", + "bbox": [ + 212, + 691, + 787, + 722 + ], + "page_idx": 8 + }, + { + "type": "equation", + "text": "\n$$\n\\operatorname {M u l t i H e a d} (Q, K, V) = \\operatorname {C o n c a t} \\left(\\operatorname {h e a d} _ {1}, \\dots , \\operatorname {h e a d} _ {h}\\right) W ^ {O}\n$$\n", + "text_format": "latex", + "bbox": [ + 313, + 736, + 669, + 751 + ], + "page_idx": 8 + }, + { + "type": "equation", + "text": "\n$$\n\\operatorname {h e a d} _ {i} = \\operatorname {P S C A t t n} \\left(Q W _ {i} ^ {Q}, K W _ {i} ^ {K}, V W _ {i} ^ {V}\\right), \\tag {4}\n$$\n", + "text_format": "latex", + "bbox": [ + 410, + 751, + 784, + 779 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "where $W_{i}^{Q}\\in \\mathbb{R}^{C\\times d_{model}}$ , $W_{i}^{K}\\in \\mathbb{R}^{C\\times d_{model}}$ , $W_{i}^{V}\\in \\mathbb{R}^{C\\times d_{model}}$ and $W_{i}^{O}\\in \\mathbb{R}^{C\\times C}$ are the linear projections. As [46], we set the number of heads to $(h = C / d_{model})$ 8, where $d_{model}$ is the projection dimension of each head.", + "bbox": [ + 212, + 792, + 787, + 840 + ], + "page_idx": 8 + }, + { + "type": "header", + "text": "PanoVOS: Bridging Non-panoramic and Panoramic Views", + "bbox": [ + 336, + 114, + 727, + 128 + ], + "page_idx": 8 + }, + { + "type": "page_number", + "text": "9", + "bbox": [ + 774, + 114, + 784, + 126 + ], + "page_idx": 8 + }, + { + "type": "table", + "img_path": "images/e85f1e7940eed0104d2a9cb85eb3fcab1bf3268b330c33eec9214e3eec0380c3.jpg", + "table_caption": [], + "table_footnote": [], + "table_body": "
Methods|MFYouTube-VOSPanoVOS ValidationPanoVOS Test
J&FJ&FJsFsJuFuJ&FJsFsJuFu
AOTT [58]73.753.8↓19.944.458.346.965.743.7↓30.036.349.839.649.2
AOTS [58]74.655.8↓18.849.162.246.465.544.7↓29.932.443.346.556.8
AOTB [58]75.253.7↓21.546.258.146.364.139.5↓35.734.444.435.044.4
AFB-URR [32]✓65.240.1↓25.131.141.535.851.830.4↓34.823.132.728.836.9
STCN [6]✓76.149.9↓26.242.753.445.158.448.0↓28.139.350.246.755.7
XMem [5]✓77.048.6↓28.440.750.144.858.640.2↓36.835.344.936.444.0
AOTL [58]✓74.749.2↓25.543.357.338.957.138.2↓36.532.343.732.744.1
R50 AOTL [58]✓76.550.1↓26.444.558.640.357.241.4↓35.133.745.038.348.4
SwinB_AOTL [58]✓74.444.8↓29.639.152.234.953.036.2↓38.231.142.031.040.6
RDE* [24]✓61.743.1↓18.636.048.435.252.741.3↓20.430.944.641.448.5
STCN* [6]✓56.343.2↓13.141.653.733.244.538.0↓18.332.843.235.540.4
XMem* [5]✓65.855.9↓9.952.264.047.260.049.6↓16.239.252.646.859.9
", + "bbox": [ + 218, + 143, + 782, + 263 + ], + "page_idx": 9 + }, + { + "type": "table", + "img_path": "images/8183adfe9ce53152edf57e6b157b1b396816f23d28c869b45bca93113b29af07.jpg", + "table_caption": [ + "Table 3: Domain transfer result of (static image datasets) $\\rightarrow$ (PanoVOS Validation & Test). Subscript $s$ and $u$ denote scores in seen and unseen categories. $MF$ denotes multiple historical frames as reference. $\\downarrow$ represents the performance of the declining values compared to the YouTube-VOS dataset [52]. * denotes a large-scale external dataset BL30K [6] dataset is used during training." + ], + "table_footnote": [], + "table_body": "
Methods|MF|YouTube-VOS |J&F|PanoVOS ValidationPanoVOS Test
J&FJsFsJuFuJ&FJsFsJuFu
CFBI†[59]81.460.9↓20.553.065.256.369.049.0↓32.449.447.646.252.6
CFBL+†[59]82.857.6↓25.252.167.048.163.453.7↓29.151.659.346.657.5
AOTT[58]80.261.5↓18.755.667.754.668.252.6↓27.644.855.351.558.8
AOTS[58]82.666.7↓15.958.070.562.076.457.3↓25.350.261.054.663.5
AOTB[58]83.570.5↓13.059.271.768.582.760.8↓22.753.064.457.868.2
AFB-URR [32]✓79.655.1↓24.544.755.653.466.752.4↓27.243.654.252.059.9
RDE[24]✓81.954.7↓27.250.363.944.660.155.4↓26.545.559.251.065.9
STCN[6]✓83.061.8↓21.250.363.561.372.153.4↓29.646.258.949.059.9
XMem[5]✓85.766.1↓19.656.668.762.077.262.5↓23.253.165.461.170.4
AOTL[58]✓83.871.9↓11.962.175.367.482.862.1↓21.757.169.056.266.1
R50_AOTL[58]✓84.169.2↓14.956.769.467.583.161.4↓22.757.569.053.365.7
SwinB_AOTL[58]✓84.567.5↓17.060.273.660.376.060.9↓23.653.963.758.767.4
RDE* [24]✓83.360.9↓22.451.464.756.071.655.6↓27.748.160.852.661.0
STCN* [6]✓84.361.7↓22.649.961.859.775.555.8↓28.548.259.852.762.5
XMem* [5]✓86.163.4↓22.753.564.461.574.161.0↓25.153.565.157.568.0
", + "bbox": [ + 220, + 366, + 782, + 510 + ], + "page_idx": 9 + }, + { + "type": "text", + "text": "Table 4: Domain transfer result of (static image datasets & YouTubeVOS) $\\rightarrow$ (PanoVOS Validation & Test). Subscript $s$ and $u$ denote scores in seen and unseen categories. $MF$ denotes multiple historical frames as reference. $\\downarrow$ represents the performance of the declining values compared to the YouTube-VOS dataset [52]. * denotes a large-scale external dataset BL30K [6] dataset is used during training. $\\dagger$ denotes no synthetic data is used during the training stage.", + "bbox": [ + 212, + 521, + 787, + 604 + ], + "page_idx": 9 + }, + { + "type": "text", + "text": "5 Experiment", + "text_level": 1, + "bbox": [ + 215, + 641, + 366, + 657 + ], + "page_idx": 9 + }, + { + "type": "text", + "text": "In this section, we design a series of experiments to answer the following research questions related to how to tackle video object segmentation in panoramic scenes:", + "bbox": [ + 212, + 689, + 785, + 719 + ], + "page_idx": 9 + }, + { + "type": "list", + "sub_type": "text", + "list_items": [ + "RQ1: How well are current VOS methods trained on non-panoramic videos adapted to the panoramic world?", + "RQ2: How well do variations of the foundation model Segment Anything Model [23] adapt to the panoramic world?", + "RQ3: Can the proposed PanoVOS datasets bring about a consistent performance gain to VOS methods?", + "RQ4: How well does Panoramic Space Consistency Attention contribute?", + "RQ5: What are the remained problems for panoramic-related research?" + ], + "bbox": [ + 212, + 719, + 785, + 839 + ], + "page_idx": 9 + }, + { + "type": "page_number", + "text": "10", + "bbox": [ + 217, + 114, + 235, + 126 + ], + "page_idx": 9 + }, + { + "type": "header", + "text": "Yan et al.", + "bbox": [ + 271, + 114, + 338, + 126 + ], + "page_idx": 9 + }, + { + "type": "table", + "img_path": "images/83e7f16b5c8b78c3c972774dcb073f2134db6fc91801566309320c7927274d06.jpg", + "table_caption": [], + "table_footnote": [], + "table_body": "
MethodsPanoVOS ValidationPanoVOS Test
J&FJsFsJuF_uJ&FJsFsJuF_u
PerSAM [61]19.112.319.817.427.119.57.414.923.831.7
SAM-PT [43]47.536.748.646.058.741.031.140.540.252.3
SAM-PT-reinit [43]43.734.344.341.354.943.635.042.743.553.0
", + "bbox": [ + 220, + 143, + 782, + 196 + ], + "page_idx": 10 + }, + { + "type": "table", + "img_path": "images/b7ab8bfaf4c6a5f3cf1ba76fefe205da3c47fdffa58ee928946d39993ce1cff5.jpg", + "table_caption": [ + "Table 5: Quantitative comparison on PanoVOS for variations of foundation model Segment Anything Model [23]. Subscript $s$ and $u$ denote scores in seen and unseen categories." + ], + "table_footnote": [], + "table_body": "
MethodsMFPanoVOS ValidationPanoVOS Test
J&FJsFsJuFuJ&FJsFsJuFu
CFBI† [57]35.834.644.824.239.719.118.226.112.219.8
CFBI+† [59]41.338.047.932.546.930.930.842.721.428.5
AOTT [58]65.659.468.359.775.053.449.361.647.555.1
AOTS [58]67.761.270.062.477.155.953.265.148.657.0
AOTB [58]67.662.372.061.574.855.453.564.247.756.0
Ours-Based74.066.480.466.283.056.849.462.752.462.5
AFB-URR [32]✓34.334.842.824.934.534.228.238.832.936.8
RDE [24]✓50.549.758.439.254.942.536.946.638.548.2
STCN [6]✓52.051.260.841.554.550.843.656.549.353.7
XMem [5]✓55.754.863.345.259.753.549.562.647.154.8
AOTL [58]✓66.661.471.159.474.353.850.060.347.857.1
R50-AOTL [58]✓65.361.971.456.471.654.652.963.247.554.9
SwinB-AOTL [58]✓62.158.966.554.368.853.149.057.849.056.6
Ours-Large✓77.970.585.269.586.459.954.969.253.062.4
RDE* [24]✓54.352.861.644.658.252.244.556.049.359.1
STCN* [6]✓51.751.260.641.353.653.853.758.146.057.3
XMem* [5]✓57.755.664.648.661.957.951.364.553.262.7
", + "bbox": [ + 220, + 263, + 782, + 416 + ], + "page_idx": 10 + }, + { + "type": "text", + "text": "Table 6: Quantitative comparison on PanoVOS for models with pretraining on static image datasets. Subscript $s$ and $u$ denote scores in seen and unseen categories. ${MF}$ denotes multiple historical frames as reference. * denotes a large-scale external dataset BL30K [6] dataset is used during training.", + "bbox": [ + 215, + 426, + 784, + 483 + ], + "page_idx": 10 + }, + { + "type": "text", + "text": "5.1 Implementation Details", + "text_level": 1, + "bbox": [ + 215, + 513, + 454, + 529 + ], + "page_idx": 10 + }, + { + "type": "text", + "text": "Model Architecture. We build two variants of our method with different reference bank sizes $\\mathcal{R}$ for a fair comparison with previous methods. Ours-Base uses only the first frame and the previous frame as reference ( $\\mathcal{R} = \\{1, t - 1\\}$ ), which are for the sake of high inference speed and low memory consumption. OursLarge uses multiple historical frames as reference ( $\\mathcal{R} = \\{1 + 2\\delta, 1 + 2\\delta, 1 + 3\\delta \\ldots\\}$ ), which follows [34,58]. In our work, we set $\\delta$ to 2 and 5 for training and testing respectively. For $p$ and $s$ in PSC block, we set them as 2 and 7.", + "bbox": [ + 212, + 542, + 784, + 648 + ], + "page_idx": 10 + }, + { + "type": "text", + "text": "Evaluation Metrics. Following the standard protocol [41, 42], we adopt the region accuracy $\\mathcal{I}$ and boundary accuracy $\\mathcal{F}$ . $\\mathcal{I}$ means the Jaccard Index/Intersection over Union (IoU), which is the ratio of intersection and the joint area between predicted masks and ground truths. And $\\mathcal{F}$ evaluates the accuracy of the segmentation boundary, which is computed by transforming it into a bipartite graph matching problem with predicted masks and ground truths.", + "bbox": [ + 212, + 648, + 784, + 739 + ], + "page_idx": 10 + }, + { + "type": "text", + "text": "5.2 Domain Transfer Results (RQ1)", + "text_level": 1, + "bbox": [ + 215, + 765, + 522, + 781 + ], + "page_idx": 10 + }, + { + "type": "text", + "text": "We evaluate previous SOTA methods, which are trained on conventional datasets that are captured by pinhole cameras, on PanoVOS datasets to evaluate the domain transfer performance. To quantify the transfer performance of advanced", + "bbox": [ + 212, + 794, + 784, + 839 + ], + "page_idx": 10 + }, + { + "type": "header", + "text": "PanoVOS: Bridging Non-panoramic and Panoramic Views", + "bbox": [ + 339, + 114, + 727, + 128 + ], + "page_idx": 10 + }, + { + "type": "page_number", + "text": "11", + "bbox": [ + 767, + 116, + 782, + 126 + ], + "page_idx": 10 + }, + { + "type": "image", + "img_path": "images/800c148b3ad0066f6c6b8b8d2f5dafc8655f2717768b8d7ef6d72cf91a6cdcc7.jpg", + "image_caption": [ + "Fig.6: Qualitative comparison to the state-of-the-art methods, RDE [24], STCN [6], and XMem [5], on PanoVOS dataset. Our model performs better under the challenge of content discontinuities. Error regions are bounded." + ], + "image_footnote": [], + "bbox": [ + 222, + 146, + 782, + 305 + ], + "page_idx": 11 + }, + { + "type": "table", + "img_path": "images/ab48327d3c85cd9a7b9f3bf082d7bfe4254dfc53c08ea021bbafcf60915eca97.jpg", + "table_caption": [], + "table_footnote": [], + "table_body": "
MethodsValidationTest
PSCAttnJ&FJ&F
Ours-Base✓72.855.4
74.056.8
Ours-Large✓74.859.5
77.959.9
", + "bbox": [ + 361, + 378, + 640, + 441 + ], + "page_idx": 11 + }, + { + "type": "text", + "text": "Table 7: Ablation study of $PSCAttn$ module on PanoVOS.", + "bbox": [ + 292, + 453, + 707, + 468 + ], + "page_idx": 11 + }, + { + "type": "text", + "text": "models trained on planar video datasets, we evaluated 15 off-the-shelf VOS models, including [5, 6, 24, 32, 57-59], and we follow official implementations and training strategies details of them. Table 3 summarizes the domain transfer results of methods that are only trained on synthetic datasets, such as COCO [33] and ECSSD [45], on PanoVOS dataset. Table 4 shows the domain transfer results of state-of-the-art methods, that are trained on synthetic datasets (e.g. COCO [33]) and video datasets (e.g. YouTube-VOS [52]), on our PanoVOS validation and test sets. By analyzing the performance of advanced VOS methods that target conventional planar videos on panoramic videos, we provide the following insights. Firstly, the performance of current sophisticated VOS models will largely degrade when employed to tackle panoramic videos. Secondly, we can observe a trend that training on larger VOS datasets, i.e., YouTube-VOS [52] and BL30K [6] can help mitigate the gap between planar and panoramic videos.", + "bbox": [ + 212, + 502, + 787, + 700 + ], + "page_idx": 11 + }, + { + "type": "text", + "text": "5.3 Results via Visual Foundation Model (RQ2)", + "text_level": 1, + "bbox": [ + 214, + 729, + 632, + 746 + ], + "page_idx": 11 + }, + { + "type": "text", + "text": "To quantify the segmentation performance of different variations of the foundation model Segment Anything Model [23] on PanoVOS, we evaluate the latest top performing models PerSAM [61] and SAM-PT [43], as shown in Table 5. The performance of these models on our challenging PanoVOS dataset is still unsatisfactory, which leaves space for further exploration.", + "bbox": [ + 212, + 763, + 787, + 840 + ], + "page_idx": 11 + }, + { + "type": "page_number", + "text": "12", + "bbox": [ + 217, + 114, + 235, + 126 + ], + "page_idx": 11 + }, + { + "type": "header", + "text": "Yan et al.", + "bbox": [ + 271, + 114, + 339, + 126 + ], + "page_idx": 11 + }, + { + "type": "table", + "img_path": "images/2b5431346cb259ac41a16bf529c49588f68147d776449609e3e382423fa03d64.jpg", + "table_caption": [], + "table_footnote": [], + "table_body": "
MethodsAttention TypeValidation J&FTest J&F
Ours-BaseCrossAttn72.554.8
PSCAttn74.056.8
Ours-LargeCrossAttn76.859.1
PSCAttn77.959.9
", + "bbox": [ + 361, + 143, + 643, + 207 + ], + "page_idx": 12 + }, + { + "type": "table", + "img_path": "images/aa6df0c14ad436ddc92dde623a746a2318c4aa5ce65373bebbc2df67cd15e2c9.jpg", + "table_caption": [ + "Table 8: Comparison between our PSC attention (PSCAttn) and cross attention (CrossAttn) module on PanoVOS dataset." + ], + "table_footnote": [], + "table_body": "
MethodsPanoVOS Validation
J&FJsFsJuFu
w/o73.768.882.663.180.3
p=376.370.485.065.884.1
p=574.265.379.767.584.2
p=1075.968.181.968.685.2
p=1575.066.781.067.085.4
p=2 (Ours)77.970.585.269.586.4
", + "bbox": [ + 361, + 258, + 642, + 340 + ], + "page_idx": 12 + }, + { + "type": "text", + "text": "Table 9: Hyperparameter Analysis of $p$ , which enables the stitching mechanism, in PSCAttn for Ours-Large model.", + "bbox": [ + 215, + 351, + 784, + 378 + ], + "page_idx": 12 + }, + { + "type": "text", + "text": "5.4 Main Results on PanoVOS (RQ3)", + "text_level": 1, + "bbox": [ + 215, + 407, + 540, + 424 + ], + "page_idx": 12 + }, + { + "type": "text", + "text": "To evaluate the performance of previous methods on the proposed panoramic VOS dataset, we re-trained them on the training set of PanoVOS for the sake of fairness. We report the performance in Table 6,", + "bbox": [ + 212, + 431, + 785, + 477 + ], + "page_idx": 12 + }, + { + "type": "text", + "text": "which demonstrates that all the previous VOS models perform worse on PanoVOS than on the traditional VOS benchmarks, e.g., YouTube-VOS. Our model substantially outperforms all these methods and achieves state-of-the-art on all evaluation metrics on PanoVOS, which verifies the effectiveness of our model in tackling panoramic videos. Fig.6 visualizes some qualitative comparisons between our model and previous state-of-the-art methods on PanoVOS dataset, which shows that previous benchmarks fail to cope with content discontinuities while our model tackles them well.", + "bbox": [ + 212, + 478, + 787, + 599 + ], + "page_idx": 12 + }, + { + "type": "text", + "text": "5.5 Ablation Study (RQ4)", + "text_level": 1, + "bbox": [ + 215, + 619, + 446, + 635 + ], + "page_idx": 12 + }, + { + "type": "text", + "text": "In this section, we conduct ablation studies to demonstrate the effectiveness of the main component, i.e., Panoramic Space Consistency Attention (PSCAttn), of our model, with all the experiments performed based on our two model variants, i.e., Ours-Base and Ours-Large. For training, static image datasets are used for pre-training and PanoVOS is used for main training. Table 7 demonstrates the effectiveness of our PSCAttn module. Besides, Fig. 7 illustrates the qualitative comparison between our default model (Ours-Base) and the setting without PSCAttn module. Our model performs better when coping with the pixel discontinuity problem. Moreover, as is shown in Table 8, compared to the conventional cross-attention (CrossAttn) module, PSCAttn also achieves better performance. In Table 9, we analyze the hyperparameter $p$ , which influences the stitching mechanism in PSCAttn, of our model (Ours-Large) on the PanoVOS validation set. Specifically, the highest overall performance ( $\\mathcal{J} \\& \\mathcal{F}$ ) is achieved", + "bbox": [ + 212, + 643, + 787, + 840 + ], + "page_idx": 12 + }, + { + "type": "header", + "text": "PanoVOS: Bridging Non-panoramic and Panoramic Views", + "bbox": [ + 338, + 114, + 727, + 128 + ], + "page_idx": 12 + }, + { + "type": "page_number", + "text": "13", + "bbox": [ + 767, + 114, + 785, + 126 + ], + "page_idx": 12 + }, + { + "type": "image", + "img_path": "images/9aae5cfcf093a8c9b0f1d38aaa5378447d41276ac2ec1c73a8e34f0ccc6510b9.jpg", + "image_caption": [ + "Fig. 7: Qualitative ablation study of PSCAttn module." + ], + "image_footnote": [], + "bbox": [ + 362, + 151, + 643, + 228 + ], + "page_idx": 13 + }, + { + "type": "image", + "img_path": "images/be77e287852906b466b08a095b21268ebaf5e6ed7db608afbae02dfd0b698270.jpg", + "image_caption": [ + "Fig. 8: Challenge. Our model fails to segment some objects with strong distortion." + ], + "image_footnote": [], + "bbox": [ + 357, + 273, + 643, + 313 + ], + "page_idx": 13 + }, + { + "type": "text", + "text": "when setting $p$ as 2. Compared to the setting without using the stitching mechanism $(w / o)$ , our model can achieve much better performance. Specifically, our final model (Ours-Large, $p = 2$ ) achieves more than $4\\%$ gain in $\\mathcal{I} \\& \\mathcal{F}$ .", + "bbox": [ + 212, + 366, + 787, + 412 + ], + "page_idx": 13 + }, + { + "type": "text", + "text": "5.6 Limitation and Future Work (RQ5)", + "text_level": 1, + "bbox": [ + 215, + 431, + 553, + 448 + ], + "page_idx": 13 + }, + { + "type": "text", + "text": "To prompt greater progress of panoramic VOS, we also analyze the limitations of our method. Specifically, our method has no notion of severe distortion challenge since we do not employ a special design (such as deformable convolution [9]) to tackle deformations. That means our model may not segment the objects with large distortions. One such failure case is shown in Fig. 8. Besides, our panoramic dataset can be applied to broader video segmentation and tracking domains, such as referring video object segmentation [28-30, 56], video object tracking [18], video instance segmentation [15], few-shot segmentation [20], and more broader embodied navigation tasks [55]. Also, it would be valuable to investigate the zero-shot segmentation performance of visual foundation models [23] on our challenging panoramic dataset. We hope our work can shed light on efficient adaptation from non-panoramic to panoramic perception.", + "bbox": [ + 212, + 455, + 787, + 638 + ], + "page_idx": 13 + }, + { + "type": "text", + "text": "6 Conclusion", + "text_level": 1, + "bbox": [ + 215, + 657, + 359, + 674 + ], + "page_idx": 13 + }, + { + "type": "text", + "text": "In this paper, we introduce a high-quality dataset, i.e., PanoVOS, for panoramic video object segmentation. Our PanoVOS dataset provides pixel-level instance annotations with diverse scenarios and significant motions. Based on this dataset, we evaluate 15 off-the-shelf VOS models and carefully analyze their limitations. Then, we further present our model, i.e., PSCFormer, which is equipped with the proposed panoramic space consistency transformer block. Our preliminary experiment demonstrates the effectiveness of our proposed model to enhance the segmentation performance and consistency in panoramic scenes. In conclusion, this provides a new challenge for video understanding, and we hope our PanoVOS dataset can attract more researchers to pay attention to panoramic videos.", + "bbox": [ + 212, + 688, + 787, + 840 + ], + "page_idx": 13 + }, + { + "type": "page_number", + "text": "14", + "bbox": [ + 217, + 114, + 235, + 126 + ], + "page_idx": 13 + }, + { + "type": "header", + "text": "Yan et al.", + "bbox": [ + 271, + 114, + 339, + 126 + ], + "page_idx": 13 + }, + { + "type": "text", + "text": "Acknowledgements", + "text_level": 1, + "bbox": [ + 217, + 143, + 401, + 162 + ], + "page_idx": 14 + }, + { + "type": "text", + "text": "This work was supported in part by National Natural Science Foundation of China (No.62072112), and Scientific and Technological Innovation Action Plan of Shanghai Science and Technology Committee (No.22511101502, No.22511102202 and No.21DZ2203300).", + "bbox": [ + 215, + 172, + 787, + 233 + ], + "page_idx": 14 + }, + { + "type": "text", + "text": "References", + "text_level": 1, + "bbox": [ + 217, + 253, + 321, + 270 + ], + "page_idx": 14 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "1. Ai, H., Cao, Z., Zhu, J., Bai, H., Chen, Y., Wang, L.: Deep learning for omnidirectional vision: A survey and new perspectives. arXiv preprint arXiv:2205.10468 (2022)", + "2. Armeni, I., Sax, S., Zamir, A.R., Savarese, S.: Joint 2d-3dsemantic data for indoor scene understanding. arXiv preprint arXiv:1702.01105 (2017)", + "3. Caelles, S., Maninis, K.K., Pont-Tuset, J., Leal-Taixe, L., Cremers, D., Van Gool, L.: One-shot video object segmentation. In: Proceedings of the IEEE conference on computer vision and pattern recognition. pp. 221-230 (2017)", + "4. Chen, X., Li, Z., Yuan, Y., Yu, G., Shen, J., Qi, D.: State-aware tracker for real-time video object segmentation. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 9384-9393 (2020)", + "5. Cheng, H.K., Schwing, A.G.: Xmem: Long-term video object segmentation with an atkinson-shiffrin memory model. In: European Conference on Computer Vision. pp. 640-658. Springer (2022)", + "6. Cheng, H.K., Tai, Y.W., Tang, C.K.: Rethinking space-time networks with improved memory coverage for efficient video object segmentation. Advances in Neural Information Processing Systems 34, 11781-11794 (2021)", + "7. Cheng, H.T., Chao, C.H., Dong, J.D., Wen, H.K., Liu, T.L., Sun, M.: Cube padding for weakly-supervised saliency prediction in 360 videos. In: Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition. pp. 1420-1429 (2018)", + "8. Cheng, J., Tsai, Y.H., Hung, W.C., Wang, S., Yang, M.H.: Fast and accurate online video object segmentation via tracking parts. In: Proceedings of the IEEE conference on computer vision and pattern recognition. pp. 7415-7424 (2018)", + "9. Dai, J., Qi, H., Xiong, Y., Li, Y., Zhang, G., Hu, H., Wei, Y.: Deformable convolutional networks. In: Proceedings of the IEEE international conference on computer vision. pp. 764-773 (2017)", + "10. Dang, J., Zheng, H., Xu, X., Guo, Y.: Unified spatio-temporal dynamic routing for efficient video object segmentation. IEEE Transactions on Intelligent Transportation Systems (2023)", + "11. Dang, J., Zheng, H., Xu, X., Wang, L., Hu, Q., Guo, Y.: Adaptive sparse memory networks for efficient and robust video object segmentation. IEEE Transactions on Neural Networks and Learning Systems (2024)", + "12. Eger Passos, D., Jung, B.: Measuring the accuracy of inside-out tracking in xr devices using a high-precision robotic arm. In: International Conference on Human-Computer Interaction. pp. 19–26. Springer (2020)", + "13. Fang, R., Yan, S., Huang, Z., Zhou, J., Tian, H., Dai, J., Li, H.: Instructseq: Unifying vision tasks with instruction-conditioned multi-modal sequence generation. arXiv preprint arXiv:2311.18835 (2023)", + "14. Guo, P., Hong, L., Zhou, X., Gao, S., Li, W., Li, J., Chen, Z., Li, X., Zhang, W., Zhang, W.: Clickvos: Click video object segmentation. arXiv preprint arXiv:2403.06130 (2024)" + ], + "bbox": [ + 218, + 281, + 784, + 839 + ], + "page_idx": 14 + }, + { + "type": "header", + "text": "PanoVOS: Bridging Non-panoramic and Panoramic Views", + "bbox": [ + 338, + 114, + 727, + 128 + ], + "page_idx": 14 + }, + { + "type": "page_number", + "text": "15", + "bbox": [ + 767, + 116, + 784, + 126 + ], + "page_idx": 14 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "15. Guo, P., Huang, T., He, P., Liu, X., Xiao, T., Chen, Z., Zhang, W.: Openvis: Open-vocabulary video instance segmentation. arXiv preprint arXiv:2305.16835 (2023)", + "16. Guo, P., Zhang, W., Li, X., Zhang, W.: Adaptive online mutual learning bi-decoders for video object segmentation. IEEE Transactions on Image Processing 31, 7063-7077 (2022)", + "17. Hendrycks, D., Gimpel, K.: Gaussian error linear units (gelus). arXiv preprint arXiv:1606.08415 (2016)", + "18. Hong, L., Yan, S., Zhang, R., Li, W., Zhou, X., Guo, P., Jiang, K., Chen, Y., Li, J., Chen, Z., et al.: Onetracker: Unifying visual object tracking with foundation models and efficient tuning. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 19079-19091 (2024)", + "19. Hu, L., Zhang, P., Zhang, B., Pan, P., Xu, Y., Jin, R.: Learning position and target consistency for memory-based video object segmentation. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 4144-4154 (2021)", + "20. Iqbal, E., Safarov, S., Bang, S.: Misanet: Multi-similarity and attention guidance for boosting few-shot segmentation. arXiv preprint arXiv:2206.09667 (2022)", + "21. Jiang, H., Jiang, G., Yu, M., Zhang, Y., Yang, Y., Peng, Z., Chen, F., Zhang, Q.: Cubemap-based perception-driven blind quality assessment for 360-degree images. IEEE Transactions on Image Processing 30, 2364-2377 (2021)", + "22. Jost, T.A., Nelson, B., Rylander, J.: Quantitative analysis of the oculus rifts in controlled movement. Disability and Rehabilitation: Assistive Technology 16(6), 632-636 (2021)", + "23. Kirillov, A., Mintun, E., Ravi, N., Mao, H., Rolland, C., Gustafson, L., Xiao, T., Whitehead, S., Berg, A.C., Lo, W.Y., et al.: Segment anything. arXiv preprint arXiv:2304.02643 (2023)", + "24. Li, M., Hu, L., Xiong, Z., Zhang, B., Pan, P., Liu, D.: Recurrent dynamic embedding for video object segmentation. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 1332-1341 (2022)", + "25. Li, W., Fan, J., Guo, P., Hong, L., Zhang, W.: Hfvos: History-future integrated dynamic memory for video object segmentation. IEEE Transactions on Circuits and Systems for Video Technology (2024)", + "26. Li, W., Guo, P., Zhou, X., Hong, L., He, Y., Zheng, X., Zhang, W., Zhang, W.: Onevos: Unifying video object segmentation with all-in-one transformer framework. arXiv preprint arXiv:2403.08682 (2024)", + "27. Li, X., Cao, H., Zhao, S., Li, J., Zhang, L., Raj, B.: Panoramic video salient object detection with ambisonic audio guidance. arXiv preprint arXiv:2211.14419 (2022)", + "28. Li, X., Wang, J., Xu, X., Li, X., Raj, B., Lu, Y.: Robust referring video object segmentation with cyclic structural consensus. In: Proceedings of the IEEE/CVF International Conference on Computer Vision. pp. 22236-22245 (2023)", + "29. Li, X., Wang, J., Xu, X., Peng, X., Singh, R., Lu, Y., Raj, B.: Qdformer: Towards robust audiovisual segmentation in complex environments with quantization-based semantic decomposition. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 3402-3413 (2024)", + "30. Li, X., Wang, J., Xu, X., Yang, M., Yang, F., Zhao, Y., Singh, R., Raj, B.: Towards noise-tolerant speech-referring video object segmentation: Bridging speech and text. In: Proceedings of the 2023 Conference on Empirical Methods in Natural Language Processing. pp. 2283-2296 (2023)", + "31. Liang, S., Shen, X., Huang, J., Hua, X.S.: Video object segmentation with dynamic memory networks and adaptive object alignment. In: Proceedings of the IEEE/CVF International Conference on Computer Vision. pp. 8065-8074 (2021)" + ], + "bbox": [ + 215, + 146, + 785, + 839 + ], + "page_idx": 15 + }, + { + "type": "page_number", + "text": "16", + "bbox": [ + 217, + 114, + 235, + 126 + ], + "page_idx": 15 + }, + { + "type": "header", + "text": "Yan et al.", + "bbox": [ + 271, + 114, + 336, + 126 + ], + "page_idx": 15 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "32. Liang, Y., Li, X., Jafari, N., Chen, J.: Video object segmentation with adaptive feature bank and uncertain-region refinement. Advances in Neural Information Processing Systems 33, 3430-3441 (2020)", + "33. Lin, T.Y., Maire, M., Belongie, S., Hays, J., Perona, P., Ramanan, D., Dollar, P., Zitnick, C.L.: Microsoft coco: Common objects in context. In: European conference on computer vision. pp. 740-755. Springer (2014)", + "34. Liu, Y., Yu, R., Wang, J., Zhao, X., Wang, Y., Tang, Y., Yang, Y.: Global spectral filter memory network for video object segmentation. In: European Conference on Computer Vision. pp. 648-665. Springer (2022)", + "35. Ma, C., Zhang, J., Yang, K., Roitberg, A., Stiefelhagen, R.: Densepass: Dense panoramic semantic segmentation via unsupervised domain adaptation with attention-augmented context exchange. In: 2021 IEEE International Intelligent Transportation Systems Conference (ITSC). pp. 2766-2772. IEEE (2021)", + "36. Maninis, K.K., Caelles, S., Chen, Y., Pont-Tuset, J., Leal-Taixe, L., Cremers, D., Van Gool, L.: Video object segmentation without temporal information. IEEE transactions on pattern analysis and machine intelligence 41(6), 1515-1530 (2018)", + "37. Mao, Y., Wang, N., Zhou, W., Li, H.: Joint inductive and transductive learning for video object segmentation. In: Proceedings of the IEEE/CVF International Conference on Computer Vision. pp. 9670-9679 (2021)", + "38. Mei, J., Zhu, A.Z., Yan, X., Yan, H., Qiao, S., Chen, L.C., Kretzschmar, H.: Waymo open dataset: Panoramic video panoptic segmentation. In: Computer Vision-ECCV 2022: 17th European Conference, Tel Aviv, Israel, October 23-27, 2022, Proceedings, Part XXIX. pp. 53-72. Springer (2022)", + "39. Oh, S.W., Lee, J.Y., Sunkavalli, K., Kim, S.J.: Fast video object segmentation by reference-guided mask propagation. In: Proceedings of the IEEE conference on computer vision and pattern recognition. pp. 7376-7385 (2018)", + "40. Oh, S.W., Lee, J.Y., Xu, N., Kim, S.J.: Video object segmentation using space-time memory networks. In: Proceedings of the IEEE/CVF International Conference on Computer Vision. pp. 9226-9235 (2019)", + "41. Perazzi, F., Pont-Tuset, J., McWilliams, B., Van Gool, L., Gross, M., Sorkine-Hornung, A.: A benchmark dataset and evaluation methodology for video object segmentation. In: Proceedings of the IEEE conference on computer vision and pattern recognition. pp. 724-732 (2016)", + "42. Pont-Tuset, J., Perazzi, F., Caelles, S., Arbeláez, P., Sorkine-Hornung, A., Van Gool, L.: The 2017 davis challenge on video object segmentation. arXiv preprint arXiv:1704.00675 (2017)", + "43. Rajic, F., Ke, L., Tai, Y.W., Tang, C.K., Danelljan, M., Yu, F.: Segment anything meets point tracking. arXiv preprint arXiv:2307.01197 (2023)", + "44. Seong, H., Hyun, J., Kim, E.: Kernelized memory network for video object segmentation. In: European Conference on Computer Vision. pp. 629-645. Springer (2020)", + "45. Shi, J., Yan, Q., Xu, L., Jia, J.: Hierarchical image saliency detection on extended cssd. IEEE transactions on pattern analysis and machine intelligence 38(4), 717-729 (2015)", + "46. Vaswani, A., Shazeer, N., Parmar, N., Uszkoreit, J., Jones, L., Gomez, A.N., Kaiser, L., Polosukhin, I.: Attention is all you need. Advances in neural information processing systems 30 (2017)", + "47. Voigtlaender, P., Leibe, B.: Online adaptation of convolutional neural networks for video object segmentation. arXiv preprint arXiv:1706.09364 (2017)" + ], + "bbox": [ + 212, + 146, + 787, + 839 + ], + "page_idx": 16 + }, + { + "type": "header", + "text": "PanoVOS: Bridging Non-panoramic and Panoramic Views", + "bbox": [ + 338, + 114, + 727, + 128 + ], + "page_idx": 16 + }, + { + "type": "page_number", + "text": "17", + "bbox": [ + 767, + 114, + 785, + 126 + ], + "page_idx": 16 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "48. Wang, H., Jiang, X., Ren, H., Hu, Y., Bai, S.: Swiftnet: Real-time video object segmentation. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 1296-1305 (2021)", + "49. Wang, W., Zhou, T., Porikli, F., Crandall, D., Van Gool, L.: A survey on deep learning technique for video segmentation. arXiv preprint arXiv:2107.01153 (2021)", + "50. Xiao, H., Feng, J., Lin, G., Liu, Y., Zhang, M.: Monet: Deep motion exploitation for video object segmentation. In: Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition. pp. 1140-1148 (2018)", + "51. Xie, H., Yao, H., Zhou, S., Zhang, S., Sun, W.: Efficient regional memory network for video object segmentation. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 1286-1295 (2021)", + "52. Xu, N., Yang, L., Fan, Y., Yue, D., Liang, Y., Yang, J., Huang, T.: Youtube-vos: A large-scale video object segmentation benchmark. arXiv preprint arXiv:1809.03327 (2018)", + "53. Xu, X., Wang, J., Li, X., Lu, Y.: Reliable propagation-correction modulation for video object segmentation. In: Proceedings of the AAAI Conference on Artificial Intelligence. pp. 2946-2954 (2022)", + "54. Xu, X., Wang, J., Ming, X., Lu, Y.: Towards robust video object segmentation with adaptive object calibration. In: Proceedings of the 30th ACM International Conference on Multimedia. pp. 2709-2718 (2022)", + "55. Xu, X., Zhang, T., Wang, S., Li, X., Chen, Y., Li, Y., Raj, B., Johnson-Roberson, M., Huang, X.: Customizable perturbation synthesis for robust slam benchmarking. arXiv preprint arXiv:2402.08125 (2024)", + "56. Yan, S., Zhang, R., Guo, Z., Chen, W., Zhang, W., Li, H., Qiao, Y., Dong, H., He, Z., Gao, P.: Referred by multi-modality: A unified temporal transformer for video object segmentation. In: Proceedings of the AAAI Conference on Artificial Intelligence. vol. 38, pp. 6449-6457 (2024)", + "57. Yang, Z., Wei, Y., Yang, Y.: Collaborative video object segmentation by foreground-background integration. In: European Conference on Computer Vision. pp. 332-348. Springer (2020)", + "58. Yang, Z., Wei, Y., Yang, Y.: Associating objects with transformers for video object segmentation. Advances in Neural Information Processing Systems 34, 2491-2502 (2021)", + "59. Yang, Z., Wei, Y., Yang, Y.: Collaborative video object segmentation by multiscale foreground-background integration. IEEE Transactions on Pattern Analysis and Machine Intelligence (2021)", + "60. Yuan, M., Richardt, C.: 360 optical flow using tangent images. In: British Machine Vision Conference (BMVC) (2021)", + "61. Zhang, R., Jiang, Z., Guo, Z., Yan, S., Pan, J., Dong, H., Gao, P., Li, H.: Personalize segment anything model with one shot. arXiv preprint arXiv:2305.03048 (2023)", + "62. Zhang, Y., Zhang, L., Wang, K., Hamidouche, W., Deforges, O.: Shd360: A benchmark dataset for salient human detection in 360 videos. arXiv preprint arXiv:2105.11578 (2021)", + "63. Zhang, Y., Wu, Z., Peng, H., Lin, S.: A transductive approach for video object segmentation. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 6949-6958 (2020)", + "64. Zhang, Z., Xu, Y., Yu, J., Gao, S.: Saliency detection in 360 videos. In: Proceedings of the European conference on computer vision (ECCV). pp. 488-503 (2018)" + ], + "bbox": [ + 215, + 147, + 784, + 811 + ], + "page_idx": 17 + }, + { + "type": "page_number", + "text": "18", + "bbox": [ + 217, + 114, + 235, + 126 + ], + "page_idx": 17 + }, + { + "type": "header", + "text": "Yan et al.", + "bbox": [ + 271, + 114, + 336, + 126 + ], + "page_idx": 17 + } +] \ No newline at end of file diff --git a/2024/PanoVOS_ Bridging Non-panoramic and Panoramic Views with Transformer for Video Segmentation/5c945af1-cf63-4f10-9827-c4fd88e7d8a2_model.json b/2024/PanoVOS_ Bridging Non-panoramic and Panoramic Views with Transformer for Video Segmentation/5c945af1-cf63-4f10-9827-c4fd88e7d8a2_model.json new file mode 100644 index 0000000000000000000000000000000000000000..98853bb84189164ccc4020e21ff5afa8e6e95eac --- /dev/null +++ b/2024/PanoVOS_ Bridging Non-panoramic and Panoramic Views with Transformer for Video Segmentation/5c945af1-cf63-4f10-9827-c4fd88e7d8a2_model.json @@ -0,0 +1,2656 @@ +[ + [ + { + "type": "title", + "bbox": [ + 0.272, + 0.141, + 0.734, + 0.21 + ], + "angle": 0, + "content": "PanoVOS: Bridging Non-panoramic and Panoramic Views with Transformer for Video Segmentation" + }, + { + "type": "text", + "bbox": [ + 0.294, + 0.235, + 0.71, + 0.268 + ], + "angle": 0, + "content": "Shilin Yan\\(^{1}\\), Xiaohao Xu\\(^{2}\\), Renrui Zhang\\(^{3}\\), Lingyi Hong\\(^{1}\\), Wenchao Chen\\(^{1}\\), Wenqiang Zhang\\(^{1}\\), and Wei Zhang\\(^{1\\star}\\)" + }, + { + "type": "text", + "bbox": [ + 0.306, + 0.277, + 0.697, + 0.348 + ], + "angle": 0, + "content": "\\(^{1}\\) Shanghai Key Lab of Intelligent Information Processing, School of Computer Science, Fudan University \\(^{2}\\) University of Michigan, Ann Arbor \\(^{3}\\) MMLab CUHK \\(^{4}\\) tattoo.ysl@mail.com weizh@fudan.edu.cn" + }, + { + "type": "image", + "bbox": [ + 0.245, + 0.361, + 0.782, + 0.459 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.214, + 0.47, + 0.79, + 0.541 + ], + "angle": 0, + "content": "Fig.1: Panoramic video object segmentation (PanoVOS). PanoVOS targets tracking and distinguishing the particular instances under content discontinuities (e.g. penguin in the image of \\( T = 15 \\)) and serve distortion (e.g. penguin in the image of \\( T = 65 \\)). We show the sample of (a) frames, (b) segmentation annotations, and (c) area proportion of foreground for the Penguin video in our dataset." + }, + { + "type": "text", + "bbox": [ + 0.261, + 0.566, + 0.744, + 0.817 + ], + "angle": 0, + "content": "Abstract. Panoramic videos contain richer spatial information and have attracted tremendous amounts of attention due to their exceptional experience in some fields such as autonomous driving and virtual reality. However, existing datasets for video segmentation only focus on conventional planar images. To address the challenge, in this paper, we present a panoramic video dataset, i.e., PanoVOS. The dataset provides 150 videos with high video resolutions and diverse motions. To quantify the domain gap between 2D planar videos and panoramic videos, we evaluate 15 off-the-shelf video object segmentation (VOS) models on PanoVOS. Through error analysis, we found that all of them fail to tackle pixel-level content discontinues of panoramic videos. Thus, we present a Panoramic Space Consistency Transformer (PSCFormer), which can effectively utilize the semantic boundary information of the previous frame for pixel-level matching with the current frame. Extensive experiments demonstrate that compared with the previous SOTA models, our PSCFormer network exhibits a great advantage in terms of segmentation results under the panoramic setting. Our dataset poses new challenges in panoramic VOS and we hope that" + }, + { + "type": "page_footnote", + "bbox": [ + 0.22, + 0.826, + 0.385, + 0.84 + ], + "angle": 0, + "content": "* Corresponding author." + } + ], + [ + { + "type": "page_number", + "bbox": [ + 0.218, + 0.116, + 0.23, + 0.127 + ], + "angle": 0, + "content": "2" + }, + { + "type": "header", + "bbox": [ + 0.272, + 0.115, + 0.34, + 0.128 + ], + "angle": 0, + "content": "Yan et al." + }, + { + "type": "text", + "bbox": [ + 0.261, + 0.148, + 0.744, + 0.19 + ], + "angle": 0, + "content": "our PanoVOS can advance the development of panoramic segmentation/tracking. The dataset, codes, and pre-train models will be published at https://github.com/shilinyan99/PanoVOS." + }, + { + "type": "title", + "bbox": [ + 0.217, + 0.218, + 0.377, + 0.234 + ], + "angle": 0, + "content": "1 Introduction" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.251, + 0.788, + 0.403 + ], + "angle": 0, + "content": "Semi-supervised video object segmentation (VOS) [52], which targets tracking and distinguishing the particular instances across the entire video sequence based on the first frame masks, plays an essential role in video understanding and editing. Conventionally, the images or videos studied in VOS are 2D planar data with a limited Field of View (FoV), which may lead to some ambiguities, especially when objects are out of view. Meanwhile, with the rapid development of VR/AR collection devices [12,22], panoramic videos with a \\(360^{\\circ} \\times 180^{\\circ}\\) FoV are able to collect the entire viewing sphere and richer spatial information [1,21, 27, 60]. To the best of our knowledge, we are the first to attempt to tackle the promising but challenging task of panoramic video object segmentation." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.403, + 0.788, + 0.553 + ], + "angle": 0, + "content": "To foster the development of panoramic VOS, we propose a new dataset in this work, aiming at panoramic video object segmentation. The dataset contains a wide range of real-world scenarios in which scenes have a large magnitude of motion. The main characteristics of our dataset are three aspects. 1) Panoramic videos bring certain advantages (richer geometric information and wider FoV) in real-world applications as well as challenges (serve distortion and content discontinuities). 2) Compared to all existing VOS datasets, our dataset has longer video clips with an average length of 20 seconds. 3) Nearly half of the video resolutions in our dataset are \\(4K\\), which may help facilitate broader video tracking/segmentation research under the high-resolution scenario." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.554, + 0.788, + 0.674 + ], + "angle": 0, + "content": "In the proposed dataset, we annotated 150 videos with 19,145 annotated instance masks, including sports (e.g. parkour, skateboard), animals (e.g. elephant, monkey), and common objects (basketball, hot balloon). Since, annotating a pixel-level intensive task is very time-consuming and expensive, we proposed a semi-supervised human-computer joint annotation strategy. Concretely, we first annotated objects at selected keyframes (1 fps) Then we adopted the state-of-the-art video object segmentation model AOT [58] for mask propagation to the rest frames of videos and we manually refine parts of them." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.674, + 0.788, + 0.78 + ], + "angle": 0, + "content": "Then, we conducted extensive experiments on PanoVOS to evaluate 15 off-the-shelf video object segmentation models. The results suggest that existing approaches can not handle several domain-unique challenges. The first is content discontinuities, which means the foreground object may be separated in the left and the right boundaries of the planar image, such as the case in the image of \\( T = 15 \\) in Fig. 1. The second is the severe distortions and deformations, such as the case in the image of \\( T = 65 \\) in Fig. 1." + }, + { + "type": "text", + "bbox": [ + 0.215, + 0.78, + 0.788, + 0.841 + ], + "angle": 0, + "content": "To tackle these challenges of panoramic video segmentation, we proposed a PSC-Former model which consists of key component Panoramic Space Consistent (PSC) blocks. The PSC block is designed for constructing spatial-temporal class-agnostic correspondence and propagating the segmentation masks. Each PSC" + } + ], + [ + { + "type": "header", + "bbox": [ + 0.339, + 0.115, + 0.728, + 0.129 + ], + "angle": 0, + "content": "PanoVOS: Bridging Non-panoramic and Panoramic Views" + }, + { + "type": "page_number", + "bbox": [ + 0.775, + 0.117, + 0.785, + 0.127 + ], + "angle": 0, + "content": "3" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.147, + 0.788, + 0.253 + ], + "angle": 0, + "content": "block utilizes a cross-attention for matching with references' embeddings and a PSC-attention for modeling the boundary semantic relationship between the previous frame and the query frame. Hence, the network can effectively alleviate the problem that the left and the right boundaries are actually continuous in panoramic videos. Our method outperforms the SOTA models that are re-trained on PanoVOS train set in segmentation quality under the panoramic setting. Our contributions are three-fold." + }, + { + "type": "text", + "bbox": [ + 0.226, + 0.264, + 0.786, + 0.308 + ], + "angle": 0, + "content": "- We introduce a panoramic video object segmentation dataset (PanoVOS) with 150 videos and 19K annotated instance masks, which fills the gap of long-term instance-level annotated panoramic video segmentation datasets." + }, + { + "type": "text", + "bbox": [ + 0.225, + 0.31, + 0.785, + 0.353 + ], + "angle": 0, + "content": "- Extensive experiments are conducted on 15 off-the-shelf VOS benchmarks and our baseline model on PanoVOS, which reveals that current methods could not tackle content discontinuities in panoramic videos well." + }, + { + "type": "text", + "bbox": [ + 0.226, + 0.355, + 0.785, + 0.398 + ], + "angle": 0, + "content": "- We propose a Panoramic Space Consistency Transformer (PSCFormer) on PanoVOS that successfully resolves the challenges of discontinuity of pixel-level content segmentation." + }, + { + "type": "list", + "bbox": [ + 0.225, + 0.264, + 0.786, + 0.398 + ], + "angle": 0, + "content": null + }, + { + "type": "title", + "bbox": [ + 0.216, + 0.422, + 0.388, + 0.437 + ], + "angle": 0, + "content": "2 Related Work" + }, + { + "type": "title", + "bbox": [ + 0.216, + 0.454, + 0.427, + 0.468 + ], + "angle": 0, + "content": "2.1 Panoramic Datasets" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.479, + 0.785, + 0.522 + ], + "angle": 0, + "content": "In this paper, panoramic videos refers to complete (360°, full view) panoramic videos, which is different from the definition in [38], which only include wide but partial views of some range-view images collected from multiple cameras." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.524, + 0.785, + 0.612 + ], + "angle": 0, + "content": "Image-based panoramic datasets. Existing popular image-level panoramic segmentation datasets are Stanford2D3D [2] and DensePASS [35]. The former one is mainly focused on indoor spaces including a total of 1,413 panoramic images with instance-level annotations in 13 categories. The latter targets driving scenes in cities. DensePASS [35] provides only 100 labeled panoramic images for testing and 2,000 unlabeled images for cross-domain transfer optimization." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.614, + 0.785, + 0.733 + ], + "angle": 0, + "content": "Video-based panoramic datasets. Video-based benchmarks mainly include SHD360 [62], SOD360 [64] and Wild360 [7]. All of them are used for panoramic video saliency object detection. Specifically, 1) SHD360 only targets human-centric video scenes with little movement. It provides 6,268 object-level pixel-wise masks and 16,238 instance-level pixel-wise masks. 2) SOD360 focuses on the sports-centric scenario with 41 video clips (12 outdoor and 29 indoor). 3) Wild360 concentrates on natural scenes with 85 videos. Note that SOD360 and Wild360 have no object-level or instance-level annotations." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.735, + 0.785, + 0.839 + ], + "angle": 0, + "content": "We make a comparison with the existing video panoramic datasets in Table 1. Specifically, our PanoVOS dataset contains 150 videos mainly from three different domains: person, animal, and common object, which makes the dataset more general for object-agnostic evaluations. Besides, videos in our dataset have a relatively large range of motion, making our PanoVOS dataset suitable for video tracking and segmentation evaluation tasks under panoramic scenes. Moreover, the average duration of each video in our dataset is \\(20s\\), which is about 4 times" + } + ], + [ + { + "type": "page_number", + "bbox": [ + 0.218, + 0.116, + 0.23, + 0.127 + ], + "angle": 0, + "content": "4" + }, + { + "type": "header", + "bbox": [ + 0.272, + 0.115, + 0.34, + 0.128 + ], + "angle": 0, + "content": "Yan et al." + }, + { + "type": "table", + "bbox": [ + 0.292, + 0.144, + 0.71, + 0.245 + ], + "angle": 0, + "content": "
DatasetsMotion#Videos#Frames#Total MasksAverage Duration
SHD360 [62]Small416,26816,2385s
SOD360 [64]Large104N/A0N/A
Wild360 [7]Large85N/A0N/A
PanoVOSLarge15013,99519,14520s
" + }, + { + "type": "table_caption", + "bbox": [ + 0.214, + 0.256, + 0.788, + 0.327 + ], + "angle": 0, + "content": "Table 1: Comparison of panoramic video datasets. Our PanoVOS is the first long-term panoramic video segmentation dataset with instance-level masks. Compared with existing panoramic video datasets [7,62,64] that are used for saliency detection, our panoramic video dataset for video segmentation, i.e., PanoVOS, includes more diverse and larger motion, making it suitable for dense video tracking evaluation." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.355, + 0.788, + 0.417 + ], + "angle": 0, + "content": "longer than SHD360 [62] (5s per video). By the way, the longer video is highlighted in a recent survey [49]. The longer the video, the more likely it is to introduce more panoramic video characteristics such as distortion and discontinuity, which is more challenging and more practical." + }, + { + "type": "title", + "bbox": [ + 0.215, + 0.437, + 0.569, + 0.453 + ], + "angle": 0, + "content": "2.2 Video Object Segmentation Datasets" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.461, + 0.788, + 0.568 + ], + "angle": 0, + "content": "The establishment of DAVIS [41, 42] and YouTube-VOS [52] datasets pave the way for the boosting development of VOS methods. They are collected by traditional pinhole cameras and the duration of each video clip is very short, only 5s on average. In contrast, the average video length in the proposed PanoVOS dataset is 20s, which is 4 times longer than the existing video datasets. Our dataset includes more challenging scenes (e.g. distortion and discontinuity) that is non-negligible in real-world applications." + }, + { + "type": "title", + "bbox": [ + 0.215, + 0.589, + 0.57, + 0.605 + ], + "angle": 0, + "content": "2.3 Video Object Segmentation Methods" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.614, + 0.787, + 0.644 + ], + "angle": 0, + "content": "Existing video object segmentation methods can be roughly classified into three subsets: online-learning-based, propagation-based, and matching-based." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.644, + 0.787, + 0.719 + ], + "angle": 0, + "content": "Online learning-based. Online learning-based approaches [3,36,50], which either train or fine-tune their networks with the first-frame ground truth at test time and are therefore a great waste of resources. OnAVOS [47] achieves promising results by introducing an online adaptation mechanism, but it still requires online fine-tuning. To a certain extent, it restricts networks' efficiency." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.72, + 0.787, + 0.78 + ], + "angle": 0, + "content": "Propagation-based. Propagation-based models [4,8,39] get the target masks in a frame-to-frame prerogation way. Although propagation-based methods improve efficiency, they lack long-term context and therefore are difficult to handle object disappearance and reappearance, severe obscuration, and distortion." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.78, + 0.787, + 0.842 + ], + "angle": 0, + "content": "Matching-based. Matching-based methods [6, 10, 11, 14, 16, 25, 26, 31, 37, 40, 53, 54, 63] aim to learn an embedding space of target objects between query and memory. Recently state-of-the-art methods encode many frames into embeddings and store them as a feature memory bank. The most representative is STM [40]," + } + ], + [ + { + "type": "header", + "bbox": [ + 0.338, + 0.115, + 0.728, + 0.129 + ], + "angle": 0, + "content": "PanoVOS: Bridging Non-panoramic and Panoramic Views" + }, + { + "type": "page_number", + "bbox": [ + 0.775, + 0.117, + 0.785, + 0.127 + ], + "angle": 0, + "content": "5" + }, + { + "type": "image", + "bbox": [ + 0.221, + 0.145, + 0.331, + 0.193 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.221, + 0.196, + 0.331, + 0.241 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.332, + 0.145, + 0.443, + 0.194 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.334, + 0.196, + 0.443, + 0.241 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.445, + 0.145, + 0.556, + 0.193 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.445, + 0.196, + 0.558, + 0.241 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.558, + 0.145, + 0.669, + 0.193 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.559, + 0.196, + 0.668, + 0.241 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.67, + 0.148, + 0.781, + 0.193 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.671, + 0.196, + 0.783, + 0.241 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.214, + 0.259, + 0.785, + 0.288 + ], + "angle": 0, + "content": "Fig. 2: PanoVOS dataset. We select 10 samples from the dataset involving major scenes. For each video, there are high-quality instance-level pixel-wise masks." + }, + { + "type": "table", + "bbox": [ + 0.334, + 0.302, + 0.67, + 0.368 + ], + "angle": 0, + "content": "
SplitsTrainValTest
#Videos8035 (10 unseen)35 (10 unseen)
#Images7,070(50.5%)3,464(24.8%)3,461(24.7%)
#Masks9,585(50.1%)4,957(25.9%)4,603(24.0%)
" + }, + { + "type": "table_caption", + "bbox": [ + 0.351, + 0.378, + 0.651, + 0.392 + ], + "angle": 0, + "content": "Table 2: Statistics of PanoVOS dataset" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.421, + 0.785, + 0.498 + ], + "angle": 0, + "content": "which has been extended to many works [5, 19, 34, 44, 48, 51]. AOT [58] introduces an identification mechanism by encoding multiple targets into the same embedding space, which can simultaneously segment multiple objects. However, they fail to address the challenges of the tremendous proportion of distortion and discontinuity under panoramic setting." + }, + { + "type": "title", + "bbox": [ + 0.215, + 0.52, + 0.428, + 0.536 + ], + "angle": 0, + "content": "3 PanoVOS Dataset" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.551, + 0.785, + 0.583 + ], + "angle": 0, + "content": "We introduced the proposed PanoVOS dataset in three parts, (1) collection process, (2) statistical summary, and (3) annotation pipeline." + }, + { + "type": "title", + "bbox": [ + 0.215, + 0.604, + 0.387, + 0.617 + ], + "angle": 0, + "content": "3.1 Data collection" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.627, + 0.785, + 0.734 + ], + "angle": 0, + "content": "We built our PanoVOS dataset with the principle of diversity in mind. Moreover, the objects in the video should have a large amplitude of motion or camera movement. Based on the above viewpoint, we collected videos from the YouTube website for further annotation, respectively. The range of the video length is from 3 to 40 seconds. The average sequence length of each video in the dataset is approximately 20 seconds. We followed the settings of YouTube-VOS [52] to sample the frames at 6 fps." + }, + { + "type": "title", + "bbox": [ + 0.215, + 0.755, + 0.409, + 0.769 + ], + "angle": 0, + "content": "3.2 Dataset Statistics" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.78, + 0.785, + 0.841 + ], + "angle": 0, + "content": "PanoVOS contains 150 videos, including 13,995 frames and 19,145 instance annotations from 35 categories. The average length of each video is 20 seconds. We believe that visual categories are representative of common life scenarios, and Fig.2 shows some samples of PanoVOS. To create our PanoVOS, in the" + } + ], + [ + { + "type": "page_number", + "bbox": [ + 0.218, + 0.116, + 0.23, + 0.127 + ], + "angle": 0, + "content": "6" + }, + { + "type": "header", + "bbox": [ + 0.272, + 0.115, + 0.34, + 0.127 + ], + "angle": 0, + "content": "Yan et al." + }, + { + "type": "image", + "bbox": [ + 0.264, + 0.146, + 0.744, + 0.292 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.216, + 0.306, + 0.788, + 0.337 + ], + "angle": 0, + "content": "Fig. 3: Instance-level distribution of PanoVOS dataset. Our dataset contains three major divisions: person, animals, and common objects with 35 sub-divisions." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.366, + 0.79, + 0.504 + ], + "angle": 0, + "content": "spirit of the video object segmentation task, we carefully selected videos with relatively large motion amplitudes and chose a set of video categories including person (e.g. parkour, dance, BMX, skateboard), animals (e.g. elephant, monkey, giraffe, rhino, birds) and common objects (e.g. basketball, hot balloon) as shown in Fig 3. PanoVOS dataset consists of 150 videos split into training (80), validation (35), and test (35) sets. Table 2 shows detailed division results. Both the validation and test sets have 35 videos (about \\(23\\%\\) of the frames and the masks). For validation and test sets, we keep some unseen visual categories for generalization ability evaluation." + }, + { + "type": "title", + "bbox": [ + 0.216, + 0.529, + 0.431, + 0.545 + ], + "angle": 0, + "content": "3.3 Annotation Pipeline" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.557, + 0.789, + 0.709 + ], + "angle": 0, + "content": "Annotation is very time-consuming and expensive for a pixel-level panoramic segmentation dataset. To obtain accurate large-scale video panoramic segmentation annotations and make the process more efficient, we propose a semi-automatic human-computer joint annotation strategy, as shown in Fig 4. First, keyframes are selected and manually annotated for each video, which are images with a speed of 1 fps. This is followed by a frame-by-frame propagation from the annotated keyframes to those unlabeled intermediate frames with a sophisticated semi-supervised VOS model. Then, to tackle the distortions and discontinuities in panoramic videos, we need to re-calibrate the resulting annotations via human refinement. More details will unfold below." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.735, + 0.789, + 0.842 + ], + "angle": 0, + "content": "Annotation Propagation For the annotation of each video, we first need an expert to browse the current video and note down all objects that have a large amplitude of movement. Then, for each video, the recorded objects in keyframes with a speed of 1 fps are selected for manual annotation. To avoid consistency errors or the problem of objects being labeled as other instances when they disappear and reappear, another expert needs to double-check the annotations of all objects to improve the accuracy of the dataset annotation." + } + ], + [ + { + "type": "header", + "bbox": [ + 0.339, + 0.115, + 0.728, + 0.129 + ], + "angle": 0, + "content": "PanoVOS: Bridging Non-panoramic and Panoramic Views" + }, + { + "type": "page_number", + "bbox": [ + 0.775, + 0.117, + 0.785, + 0.127 + ], + "angle": 0, + "content": "7" + }, + { + "type": "image", + "bbox": [ + 0.218, + 0.144, + 0.789, + 0.257 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.214, + 0.265, + 0.789, + 0.364 + ], + "angle": 0, + "content": "Fig. 4: PanoVOS annotation pipeline. Our annotation pipeline includes two phases. (1) The first phase is called Key Frames Select and Annotate. The annotator browses the video and picks out the object to be annotated. Then, instances are manually annotated at 1 fps and corrected by another annotator. (2) The second phase is called All Frames Propagate and Refine. In this phase, we apply a semi-supervised video object segmentation model to help propagate the annotated masks and the generated instances are refined by annotators." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.394, + 0.789, + 0.44 + ], + "angle": 0, + "content": "We then use the off-the-shelf video labeling method [58] to propagate the instance masks frame by frame from the annotated keyframes to untagged intermediate frames and generate masks at 6 fps." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.465, + 0.789, + 0.603 + ], + "angle": 0, + "content": "Annotation Refinement To present a new Panoramic dataset of high quality. After obtaining masks of the first propagation stage, annotators are asked to check the quality of the masks and refine them. The main amendments are in the following two areas. 1). Since our video resolution is generally relatively high, the propagation method will often fail when encountering complex videos with many small objects in a scene. 2) Due to the huge distortions and discontinuities present in the panoramic video, the quality of the masks obtained is relatively poor. Manual correction of the mask is checked by another annotator until the result is satisfactory before proceeding to the next video annotation." + }, + { + "type": "title", + "bbox": [ + 0.216, + 0.627, + 0.331, + 0.643 + ], + "angle": 0, + "content": "4 Method" + }, + { + "type": "title", + "bbox": [ + 0.216, + 0.661, + 0.342, + 0.675 + ], + "angle": 0, + "content": "4.1 Overview" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.689, + 0.789, + 0.841 + ], + "angle": 0, + "content": "Video object segmentation targets assigning an instance label to every pixel in the given video sequence based on the first frame mask. Recent works [5,6,13,59] have demonstrated that the attention mechanism can significantly help improve the segmentation performance. However, for the challenge of content discontinuation in panoramic videos, only considering the original attention mechanism will not be able to fully utilize the semantic information on the left and right boundaries (pixel contiguity) in the spatial dimension and will lose valuable contextual information when segmenting objects. Therefore, in this work, our mission is to design an effective network architecture, which can help acquire valuable boundary relationships." + } + ], + [ + { + "type": "page_number", + "bbox": [ + 0.218, + 0.116, + 0.23, + 0.127 + ], + "angle": 0, + "content": "8" + }, + { + "type": "header", + "bbox": [ + 0.272, + 0.115, + 0.34, + 0.127 + ], + "angle": 0, + "content": "Yan et al." + }, + { + "type": "image", + "bbox": [ + 0.266, + 0.153, + 0.74, + 0.337 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.214, + 0.346, + 0.788, + 0.444 + ], + "angle": 0, + "content": "Fig. 5: (a) PSCFormer overview. Given the query frame \\(\\mathbf{x}_t\\) and reference frames \\(\\{\\mathbf{x}_i | i \\in \\mathcal{R}\\}\\), the goal of VOS is to delineate objects from the background by generating mask \\(\\mathbf{y}_t\\) for query frame \\(\\mathbf{x}_t\\). References and the query frame are encoded by the memory encoder and query encoder, respectively. Multiple stacking panoramic space consistency (PSC) blocks are used to leverage the correspondence in the panoramic space between references and the query frame. A decoder is used for generating the prediction of the query frame. (b) Panoramic space consistency block architecture details." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.473, + 0.788, + 0.611 + ], + "angle": 0, + "content": "Fig. 5(a) illustrates the overall architecture of the proposed network. Given the query frame \\(\\mathbf{x}_t\\) and references \\(\\{\\mathbf{x}_i | i \\in \\mathcal{R}\\}\\), the goal of VOS is to delineate objects from the background by generating mask \\(\\mathbf{y}_t\\) for query frame \\(\\mathbf{x}_t\\). Following [57], our basic setting uses the first and previous frame as references \\(\\mathcal{R} = \\{1, t - 1\\}\\). The memory encoder and query encoder are responsible for extracting frame-level features. After this, the panoramic space consistency block takes them as input and aggregates the spatial-temporal information between the reference frames and the query frame at the pixel level. Finally, the decoder uses the output of the sequence stacking PSC blocks to predict the mask of the object." + }, + { + "type": "title", + "bbox": [ + 0.215, + 0.633, + 0.56, + 0.649 + ], + "angle": 0, + "content": "4.2 Panoramic Space Consistency Block" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.659, + 0.788, + 0.779 + ], + "angle": 0, + "content": "Fig. 5(b) shows the structure of a PSC block. Motivated by the common transformer blocks [46], PSC firstly contains a self-attention layer, which is used to aggregate the target objects' correlation information within the query frame. Then, the middle module is composed of cross-attention and PSC-attention, in which cross-attention is responsible for learning the target objects' information from references \\(\\mathcal{R}\\) and the PSC-attention targets on exploring the boundary relationship between the query frame and previous frame. Finally, PSC employs a two-layer feed-forward MLP with GELU [17] non-linearity activation function." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.78, + 0.788, + 0.84 + ], + "angle": 0, + "content": "Panoramic Space Consistency Attention (PSC-Attn). PSC-Attn is employed to model the spatial-temporal relationship between the query frame and reference frames considering the continuity of pixels of images in the panoramic space. How to establish a connection between the left and right boundaries be" + } + ], + [ + { + "type": "header", + "bbox": [ + 0.338, + 0.115, + 0.728, + 0.129 + ], + "angle": 0, + "content": "PanoVOS: Bridging Non-panoramic and Panoramic Views" + }, + { + "type": "page_number", + "bbox": [ + 0.775, + 0.116, + 0.785, + 0.127 + ], + "angle": 0, + "content": "9" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.147, + 0.788, + 0.329 + ], + "angle": 0, + "content": "come especially important? The most intuitive solution would be to directly splice in length, but this would lead to a huge amount of computation. Therefore, we take the approach of moving a portion of the region in the length dimension from the right boundary to the leftmost boundary for stitching. Consequently, we only focus on the left and right boundaries between the query frame and the reference frame. Thus, unlike the original attention, where each query token is counted for attention along with all key tokens in the reference frame, our PSC attention takes care of the key tokens in a fixed window size. In particular, we define the reference frame feature embedding \\(\\mathbf{f}(\\mathbf{x})\\in \\mathbb{R}^{H\\times W\\times C}\\), which is extracted from the query encoder. \\(H\\), \\(W\\), and \\(C\\) represent the height, width, and channel dimensions, respectively. According to the solutions mentioned above, the new feature embedding \\(\\mathbf{f}(\\mathbf{x})'\\) is calculated as follows:" + }, + { + "type": "equation", + "bbox": [ + 0.398, + 0.344, + 0.61, + 0.358 + ], + "angle": 0, + "content": "\\[\n\\mathbf {f} (\\mathbf {x}) ^ {\\prime} [ 0: W / p ] = \\mathbf {f} (\\mathbf {x}) [ W / p: W ]\n\\]" + }, + { + "type": "equation", + "bbox": [ + 0.389, + 0.361, + 0.785, + 0.377 + ], + "angle": 0, + "content": "\\[\n\\mathbf {f} (\\mathbf {x}) ^ {\\prime} [ W / p: W ] = \\mathbf {f} (\\mathbf {x}) [ 0: W / p ] \\tag {1}\n\\]" + }, + { + "type": "equation", + "bbox": [ + 0.34, + 0.379, + 0.664, + 0.394 + ], + "angle": 0, + "content": "\\[\n\\mathbf {f} (\\mathbf {x}) ^ {\\prime} [ W / p: W - W / p ] = \\mathbf {f} (\\mathbf {x}) [ W / p: W - W / p ],\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.215, + 0.408, + 0.789, + 0.47 + ], + "angle": 0, + "content": "where \\(p \\in \\mathbb{Z}^+\\). We define query embedding \\(Q \\in \\mathbb{R}^{HW \\times C}\\), key embedding \\(K \\in \\mathbb{R}^{HW \\times C}\\), value embedding \\(V \\in \\mathbb{R}^{HW \\times C}\\), where \\(Q\\) is from the query frame feature embedding, \\(K\\) and \\(V\\) are from \\(\\mathbf{f}(\\mathbf{x})'\\) by performing dimensional transformations. Mathematically, we define the PSC attention as follows," + }, + { + "type": "equation", + "bbox": [ + 0.345, + 0.482, + 0.786, + 0.517 + ], + "angle": 0, + "content": "\\[\n\\operatorname {P S C A t t n} (Q, K, V) = \\operatorname {s o f t m a x} \\left(\\frac {Q K ^ {T} \\mathbf {R}}{\\sqrt {C}}\\right) V, \\tag {2}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.215, + 0.53, + 0.788, + 0.566 + ], + "angle": 0, + "content": "where \\(\\mathbf{R} \\in [0,1]^{HW \\times HW}\\) means a window that represents the attention range of each query token. For query \\(Q_{(x,y)}\\) at \\((x,y)\\) position, we define the \\(\\mathbf{R}_{(x,y)}\\) as:" + }, + { + "type": "equation", + "bbox": [ + 0.312, + 0.579, + 0.786, + 0.618 + ], + "angle": 0, + "content": "\\[\n\\mathbf {R} _ {x, y} (i, j) = \\left\\{ \\begin{array}{l l} 1 & \\text {i f} (x - i) ^ {2} \\leqslant s ^ {2} \\text {a n d} (y - j) ^ {2} \\leqslant s ^ {2} \\\\ 0 & \\text {o t h e r w i s e} \\end{array} , \\right. \\tag {3}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.632, + 0.788, + 0.692 + ], + "angle": 0, + "content": "where \\((i,j)\\) is the position for each key token, \\(s\\) is the window size. For each query token, it calculates the attention with another key token only if they are spatially limited to a \\((2\\times s + 1)\\) size window, which significantly reduces the time complexity from \\((h\\times w)^{2}\\) to \\((2\\times s + 1)^{2}\\)." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.692, + 0.788, + 0.723 + ], + "angle": 0, + "content": "Following [46], we implement the representational form of our PSCAttn module with multi-headed attention, defined mathematically as follows," + }, + { + "type": "equation", + "bbox": [ + 0.315, + 0.737, + 0.67, + 0.752 + ], + "angle": 0, + "content": "\\[\n\\operatorname {M u l t i H e a d} (Q, K, V) = \\operatorname {C o n c a t} \\left(\\operatorname {h e a d} _ {1}, \\dots , \\operatorname {h e a d} _ {h}\\right) W ^ {O}\n\\]" + }, + { + "type": "equation", + "bbox": [ + 0.411, + 0.752, + 0.785, + 0.78 + ], + "angle": 0, + "content": "\\[\n\\operatorname {h e a d} _ {i} = \\operatorname {P S C A t t n} \\left(Q W _ {i} ^ {Q}, K W _ {i} ^ {K}, V W _ {i} ^ {V}\\right), \\tag {4}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.793, + 0.788, + 0.842 + ], + "angle": 0, + "content": "where \\( W_{i}^{Q}\\in \\mathbb{R}^{C\\times d_{model}} \\), \\( W_{i}^{K}\\in \\mathbb{R}^{C\\times d_{model}} \\), \\( W_{i}^{V}\\in \\mathbb{R}^{C\\times d_{model}} \\) and \\( W_{i}^{O}\\in \\mathbb{R}^{C\\times C} \\) are the linear projections. As [46], we set the number of heads to \\( (h = C / d_{model}) \\) 8, where \\( d_{model} \\) is the projection dimension of each head." + } + ], + [ + { + "type": "page_number", + "bbox": [ + 0.218, + 0.116, + 0.236, + 0.127 + ], + "angle": 0, + "content": "10" + }, + { + "type": "header", + "bbox": [ + 0.272, + 0.115, + 0.339, + 0.127 + ], + "angle": 0, + "content": "Yan et al." + }, + { + "type": "table", + "bbox": [ + 0.22, + 0.144, + 0.784, + 0.265 + ], + "angle": 0, + "content": "
Methods|MFYouTube-VOSPanoVOS ValidationPanoVOS Test
J&FJ&FJsFsJuFuJ&FJsFsJuFu
AOTT [58]73.753.8↓19.944.458.346.965.743.7↓30.036.349.839.649.2
AOTS [58]74.655.8↓18.849.162.246.465.544.7↓29.932.443.346.556.8
AOTB [58]75.253.7↓21.546.258.146.364.139.5↓35.734.444.435.044.4
AFB-URR [32]✓65.240.1↓25.131.141.535.851.830.4↓34.823.132.728.836.9
STCN [6]✓76.149.9↓26.242.753.445.158.448.0↓28.139.350.246.755.7
XMem [5]✓77.048.6↓28.440.750.144.858.640.2↓36.835.344.936.444.0
AOTL [58]✓74.749.2↓25.543.357.338.957.138.2↓36.532.343.732.744.1
R50 AOTL [58]✓76.550.1↓26.444.558.640.357.241.4↓35.133.745.038.348.4
SwinB_AOTL [58]✓74.444.8↓29.639.152.234.953.036.2↓38.231.142.031.040.6
RDE* [24]✓61.743.1↓18.636.048.435.252.741.3↓20.430.944.641.448.5
STCN* [6]✓56.343.2↓13.141.653.733.244.538.0↓18.332.843.235.540.4
XMem* [5]✓65.855.9↓9.952.264.047.260.049.6↓16.239.252.646.859.9
" + }, + { + "type": "table_caption", + "bbox": [ + 0.214, + 0.275, + 0.788, + 0.346 + ], + "angle": 0, + "content": "Table 3: Domain transfer result of (static image datasets) \\(\\rightarrow\\) (PanoVOS Validation & Test). Subscript \\(s\\) and \\(u\\) denote scores in seen and unseen categories. \\(MF\\) denotes multiple historical frames as reference. \\(\\downarrow\\) represents the performance of the declining values compared to the YouTube-VOS dataset [52]. * denotes a large-scale external dataset BL30K [6] dataset is used during training." + }, + { + "type": "table", + "bbox": [ + 0.221, + 0.367, + 0.784, + 0.511 + ], + "angle": 0, + "content": "
Methods|MF|YouTube-VOS |J&F|PanoVOS ValidationPanoVOS Test
J&FJsFsJuFuJ&FJsFsJuFu
CFBI†[59]81.460.9↓20.553.065.256.369.049.0↓32.449.447.646.252.6
CFBL+†[59]82.857.6↓25.252.167.048.163.453.7↓29.151.659.346.657.5
AOTT[58]80.261.5↓18.755.667.754.668.252.6↓27.644.855.351.558.8
AOTS[58]82.666.7↓15.958.070.562.076.457.3↓25.350.261.054.663.5
AOTB[58]83.570.5↓13.059.271.768.582.760.8↓22.753.064.457.868.2
AFB-URR [32]✓79.655.1↓24.544.755.653.466.752.4↓27.243.654.252.059.9
RDE[24]✓81.954.7↓27.250.363.944.660.155.4↓26.545.559.251.065.9
STCN[6]✓83.061.8↓21.250.363.561.372.153.4↓29.646.258.949.059.9
XMem[5]✓85.766.1↓19.656.668.762.077.262.5↓23.253.165.461.170.4
AOTL[58]✓83.871.9↓11.962.175.367.482.862.1↓21.757.169.056.266.1
R50_AOTL[58]✓84.169.2↓14.956.769.467.583.161.4↓22.757.569.053.365.7
SwinB_AOTL[58]✓84.567.5↓17.060.273.660.376.060.9↓23.653.963.758.767.4
RDE* [24]✓83.360.9↓22.451.464.756.071.655.6↓27.748.160.852.661.0
STCN* [6]✓84.361.7↓22.649.961.859.775.555.8↓28.548.259.852.762.5
XMem* [5]✓86.163.4↓22.753.564.461.574.161.0↓25.153.565.157.568.0
" + }, + { + "type": "table_caption", + "bbox": [ + 0.214, + 0.522, + 0.788, + 0.606 + ], + "angle": 0, + "content": "Table 4: Domain transfer result of (static image datasets & YouTubeVOS) \\(\\rightarrow\\) (PanoVOS Validation & Test). Subscript \\(s\\) and \\(u\\) denote scores in seen and unseen categories. \\(MF\\) denotes multiple historical frames as reference. \\(\\downarrow\\) represents the performance of the declining values compared to the YouTube-VOS dataset [52]. * denotes a large-scale external dataset BL30K [6] dataset is used during training. \\(\\dagger\\) denotes no synthetic data is used during the training stage." + }, + { + "type": "title", + "bbox": [ + 0.216, + 0.642, + 0.367, + 0.659 + ], + "angle": 0, + "content": "5 Experiment" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.69, + 0.786, + 0.72 + ], + "angle": 0, + "content": "In this section, we design a series of experiments to answer the following research questions related to how to tackle video object segmentation in panoramic scenes:" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.72, + 0.786, + 0.749 + ], + "angle": 0, + "content": "RQ1: How well are current VOS methods trained on non-panoramic videos adapted to the panoramic world?" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.75, + 0.786, + 0.779 + ], + "angle": 0, + "content": "RQ2: How well do variations of the foundation model Segment Anything Model [23] adapt to the panoramic world?" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.78, + 0.786, + 0.809 + ], + "angle": 0, + "content": "RQ3: Can the proposed PanoVOS datasets bring about a consistent performance gain to VOS methods?" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.81, + 0.745, + 0.825 + ], + "angle": 0, + "content": "RQ4: How well does Panoramic Space Consistency Attention contribute?" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.825, + 0.733, + 0.84 + ], + "angle": 0, + "content": "RQ5: What are the remained problems for panoramic-related research?" + }, + { + "type": "list", + "bbox": [ + 0.214, + 0.72, + 0.786, + 0.84 + ], + "angle": 0, + "content": null + } + ], + [ + { + "type": "header", + "bbox": [ + 0.34, + 0.115, + 0.728, + 0.129 + ], + "angle": 0, + "content": "PanoVOS: Bridging Non-panoramic and Panoramic Views" + }, + { + "type": "page_number", + "bbox": [ + 0.769, + 0.117, + 0.784, + 0.127 + ], + "angle": 0, + "content": "11" + }, + { + "type": "table", + "bbox": [ + 0.221, + 0.144, + 0.784, + 0.198 + ], + "angle": 0, + "content": "
MethodsPanoVOS ValidationPanoVOS Test
J&FJsFsJuF_uJ&FJsFsJuF_u
PerSAM [61]19.112.319.817.427.119.57.414.923.831.7
SAM-PT [43]47.536.748.646.058.741.031.140.540.252.3
SAM-PT-reinit [43]43.734.344.341.354.943.635.042.743.553.0
" + }, + { + "type": "table_caption", + "bbox": [ + 0.216, + 0.208, + 0.785, + 0.249 + ], + "angle": 0, + "content": "Table 5: Quantitative comparison on PanoVOS for variations of foundation model Segment Anything Model [23]. Subscript \\( s \\) and \\( u \\) denote scores in seen and unseen categories." + }, + { + "type": "table", + "bbox": [ + 0.221, + 0.265, + 0.784, + 0.417 + ], + "angle": 0, + "content": "
MethodsMFPanoVOS ValidationPanoVOS Test
J&FJsFsJuFuJ&FJsFsJuFu
CFBI† [57]35.834.644.824.239.719.118.226.112.219.8
CFBI+† [59]41.338.047.932.546.930.930.842.721.428.5
AOTT [58]65.659.468.359.775.053.449.361.647.555.1
AOTS [58]67.761.270.062.477.155.953.265.148.657.0
AOTB [58]67.662.372.061.574.855.453.564.247.756.0
Ours-Based74.066.480.466.283.056.849.462.752.462.5
AFB-URR [32]✓34.334.842.824.934.534.228.238.832.936.8
RDE [24]✓50.549.758.439.254.942.536.946.638.548.2
STCN [6]✓52.051.260.841.554.550.843.656.549.353.7
XMem [5]✓55.754.863.345.259.753.549.562.647.154.8
AOTL [58]✓66.661.471.159.474.353.850.060.347.857.1
R50-AOTL [58]✓65.361.971.456.471.654.652.963.247.554.9
SwinB-AOTL [58]✓62.158.966.554.368.853.149.057.849.056.6
Ours-Large✓77.970.585.269.586.459.954.969.253.062.4
RDE* [24]✓54.352.861.644.658.252.244.556.049.359.1
STCN* [6]✓51.751.260.641.353.653.853.758.146.057.3
XMem* [5]✓57.755.664.648.661.957.951.364.553.262.7
" + }, + { + "type": "table_caption", + "bbox": [ + 0.216, + 0.428, + 0.785, + 0.484 + ], + "angle": 0, + "content": "Table 6: Quantitative comparison on PanoVOS for models with pretraining on static image datasets. Subscript \\( s \\) and \\( u \\) denote scores in seen and unseen categories. \\( {MF} \\) denotes multiple historical frames as reference. * denotes a large-scale external dataset BL30K [6] dataset is used during training." + }, + { + "type": "title", + "bbox": [ + 0.216, + 0.515, + 0.455, + 0.53 + ], + "angle": 0, + "content": "5.1 Implementation Details" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.544, + 0.785, + 0.649 + ], + "angle": 0, + "content": "Model Architecture. We build two variants of our method with different reference bank sizes \\(\\mathcal{R}\\) for a fair comparison with previous methods. Ours-Base uses only the first frame and the previous frame as reference (\\(\\mathcal{R} = \\{1, t - 1\\}\\)), which are for the sake of high inference speed and low memory consumption. OursLarge uses multiple historical frames as reference (\\(\\mathcal{R} = \\{1 + 2\\delta, 1 + 2\\delta, 1 + 3\\delta \\ldots\\}\\)), which follows [34,58]. In our work, we set \\(\\delta\\) to 2 and 5 for training and testing respectively. For \\(p\\) and \\(s\\) in PSC block, we set them as 2 and 7." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.65, + 0.785, + 0.74 + ], + "angle": 0, + "content": "Evaluation Metrics. Following the standard protocol [41, 42], we adopt the region accuracy \\(\\mathcal{I}\\) and boundary accuracy \\(\\mathcal{F}\\). \\(\\mathcal{I}\\) means the Jaccard Index/Intersection over Union (IoU), which is the ratio of intersection and the joint area between predicted masks and ground truths. And \\(\\mathcal{F}\\) evaluates the accuracy of the segmentation boundary, which is computed by transforming it into a bipartite graph matching problem with predicted masks and ground truths." + }, + { + "type": "title", + "bbox": [ + 0.216, + 0.766, + 0.524, + 0.782 + ], + "angle": 0, + "content": "5.2 Domain Transfer Results (RQ1)" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.795, + 0.785, + 0.84 + ], + "angle": 0, + "content": "We evaluate previous SOTA methods, which are trained on conventional datasets that are captured by pinhole cameras, on PanoVOS datasets to evaluate the domain transfer performance. To quantify the transfer performance of advanced" + } + ], + [ + { + "type": "page_number", + "bbox": [ + 0.218, + 0.116, + 0.236, + 0.127 + ], + "angle": 0, + "content": "12" + }, + { + "type": "header", + "bbox": [ + 0.272, + 0.115, + 0.34, + 0.127 + ], + "angle": 0, + "content": "Yan et al." + }, + { + "type": "image", + "bbox": [ + 0.223, + 0.147, + 0.784, + 0.306 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.214, + 0.318, + 0.788, + 0.361 + ], + "angle": 0, + "content": "Fig.6: Qualitative comparison to the state-of-the-art methods, RDE [24], STCN [6], and XMem [5], on PanoVOS dataset. Our model performs better under the challenge of content discontinuities. Error regions are bounded." + }, + { + "type": "table", + "bbox": [ + 0.362, + 0.379, + 0.642, + 0.443 + ], + "angle": 0, + "content": "
MethodsValidationTest
PSCAttnJ&FJ&F
Ours-Base✓72.855.4
74.056.8
Ours-Large✓74.859.5
77.959.9
" + }, + { + "type": "table_caption", + "bbox": [ + 0.293, + 0.454, + 0.709, + 0.469 + ], + "angle": 0, + "content": "Table 7: Ablation study of \\(PSCAttn\\) module on PanoVOS." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.503, + 0.788, + 0.701 + ], + "angle": 0, + "content": "models trained on planar video datasets, we evaluated 15 off-the-shelf VOS models, including [5, 6, 24, 32, 57-59], and we follow official implementations and training strategies details of them. Table 3 summarizes the domain transfer results of methods that are only trained on synthetic datasets, such as COCO [33] and ECSSD [45], on PanoVOS dataset. Table 4 shows the domain transfer results of state-of-the-art methods, that are trained on synthetic datasets (e.g. COCO [33]) and video datasets (e.g. YouTube-VOS [52]), on our PanoVOS validation and test sets. By analyzing the performance of advanced VOS methods that target conventional planar videos on panoramic videos, we provide the following insights. Firstly, the performance of current sophisticated VOS models will largely degrade when employed to tackle panoramic videos. Secondly, we can observe a trend that training on larger VOS datasets, i.e., YouTube-VOS [52] and BL30K [6] can help mitigate the gap between planar and panoramic videos." + }, + { + "type": "title", + "bbox": [ + 0.215, + 0.731, + 0.633, + 0.747 + ], + "angle": 0, + "content": "5.3 Results via Visual Foundation Model (RQ2)" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.765, + 0.788, + 0.841 + ], + "angle": 0, + "content": "To quantify the segmentation performance of different variations of the foundation model Segment Anything Model [23] on PanoVOS, we evaluate the latest top performing models PerSAM [61] and SAM-PT [43], as shown in Table 5. The performance of these models on our challenging PanoVOS dataset is still unsatisfactory, which leaves space for further exploration." + } + ], + [ + { + "type": "header", + "bbox": [ + 0.339, + 0.115, + 0.728, + 0.129 + ], + "angle": 0, + "content": "PanoVOS: Bridging Non-panoramic and Panoramic Views" + }, + { + "type": "page_number", + "bbox": [ + 0.769, + 0.116, + 0.786, + 0.127 + ], + "angle": 0, + "content": "13" + }, + { + "type": "table", + "bbox": [ + 0.362, + 0.144, + 0.645, + 0.208 + ], + "angle": 0, + "content": "
MethodsAttention TypeValidation J&FTest J&F
Ours-BaseCrossAttn72.554.8
PSCAttn74.056.8
Ours-LargeCrossAttn76.859.1
PSCAttn77.959.9
" + }, + { + "type": "table_caption", + "bbox": [ + 0.216, + 0.219, + 0.787, + 0.247 + ], + "angle": 0, + "content": "Table 8: Comparison between our PSC attention (PSCAttn) and cross attention (CrossAttn) module on PanoVOS dataset." + }, + { + "type": "table", + "bbox": [ + 0.362, + 0.26, + 0.643, + 0.342 + ], + "angle": 0, + "content": "
MethodsPanoVOS Validation
J&FJsFsJuFu
w/o73.768.882.663.180.3
p=376.370.485.065.884.1
p=574.265.379.767.584.2
p=1075.968.181.968.685.2
p=1575.066.781.067.085.4
p=2 (Ours)77.970.585.269.586.4
" + }, + { + "type": "table_caption", + "bbox": [ + 0.216, + 0.352, + 0.785, + 0.38 + ], + "angle": 0, + "content": "Table 9: Hyperparameter Analysis of \\( p \\), which enables the stitching mechanism, in PSCAttn for Ours-Large model." + }, + { + "type": "title", + "bbox": [ + 0.216, + 0.408, + 0.542, + 0.425 + ], + "angle": 0, + "content": "5.4 Main Results on PanoVOS (RQ3)" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.433, + 0.786, + 0.478 + ], + "angle": 0, + "content": "To evaluate the performance of previous methods on the proposed panoramic VOS dataset, we re-trained them on the training set of PanoVOS for the sake of fairness. We report the performance in Table 6," + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.479, + 0.788, + 0.6 + ], + "angle": 0, + "content": "which demonstrates that all the previous VOS models perform worse on PanoVOS than on the traditional VOS benchmarks, e.g., YouTube-VOS. Our model substantially outperforms all these methods and achieves state-of-the-art on all evaluation metrics on PanoVOS, which verifies the effectiveness of our model in tackling panoramic videos. Fig.6 visualizes some qualitative comparisons between our model and previous state-of-the-art methods on PanoVOS dataset, which shows that previous benchmarks fail to cope with content discontinuities while our model tackles them well." + }, + { + "type": "title", + "bbox": [ + 0.216, + 0.62, + 0.447, + 0.636 + ], + "angle": 0, + "content": "5.5 Ablation Study (RQ4)" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.644, + 0.788, + 0.842 + ], + "angle": 0, + "content": "In this section, we conduct ablation studies to demonstrate the effectiveness of the main component, i.e., Panoramic Space Consistency Attention (PSCAttn), of our model, with all the experiments performed based on our two model variants, i.e., Ours-Base and Ours-Large. For training, static image datasets are used for pre-training and PanoVOS is used for main training. Table 7 demonstrates the effectiveness of our PSCAttn module. Besides, Fig. 7 illustrates the qualitative comparison between our default model (Ours-Base) and the setting without PSCAttn module. Our model performs better when coping with the pixel discontinuity problem. Moreover, as is shown in Table 8, compared to the conventional cross-attention (CrossAttn) module, PSCAttn also achieves better performance. In Table 9, we analyze the hyperparameter \\( p \\), which influences the stitching mechanism in PSCAttn, of our model (Ours-Large) on the PanoVOS validation set. Specifically, the highest overall performance (\\( \\mathcal{J} \\& \\mathcal{F} \\)) is achieved" + } + ], + [ + { + "type": "page_number", + "bbox": [ + 0.218, + 0.116, + 0.236, + 0.127 + ], + "angle": 0, + "content": "14" + }, + { + "type": "header", + "bbox": [ + 0.272, + 0.115, + 0.34, + 0.127 + ], + "angle": 0, + "content": "Yan et al." + }, + { + "type": "image", + "bbox": [ + 0.364, + 0.152, + 0.645, + 0.229 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.303, + 0.246, + 0.699, + 0.262 + ], + "angle": 0, + "content": "Fig. 7: Qualitative ablation study of PSCAttn module." + }, + { + "type": "image", + "bbox": [ + 0.359, + 0.275, + 0.644, + 0.314 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.219, + 0.324, + 0.782, + 0.34 + ], + "angle": 0, + "content": "Fig. 8: Challenge. Our model fails to segment some objects with strong distortion." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.367, + 0.788, + 0.414 + ], + "angle": 0, + "content": "when setting \\( p \\) as 2. Compared to the setting without using the stitching mechanism \\( (w / o) \\), our model can achieve much better performance. Specifically, our final model (Ours-Large, \\( p = 2 \\)) achieves more than \\( 4\\% \\) gain in \\( \\mathcal{I} \\& \\mathcal{F} \\)." + }, + { + "type": "title", + "bbox": [ + 0.216, + 0.433, + 0.555, + 0.449 + ], + "angle": 0, + "content": "5.6 Limitation and Future Work (RQ5)" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.457, + 0.789, + 0.64 + ], + "angle": 0, + "content": "To prompt greater progress of panoramic VOS, we also analyze the limitations of our method. Specifically, our method has no notion of severe distortion challenge since we do not employ a special design (such as deformable convolution [9]) to tackle deformations. That means our model may not segment the objects with large distortions. One such failure case is shown in Fig. 8. Besides, our panoramic dataset can be applied to broader video segmentation and tracking domains, such as referring video object segmentation [28-30, 56], video object tracking [18], video instance segmentation [15], few-shot segmentation [20], and more broader embodied navigation tasks [55]. Also, it would be valuable to investigate the zero-shot segmentation performance of visual foundation models [23] on our challenging panoramic dataset. We hope our work can shed light on efficient adaptation from non-panoramic to panoramic perception." + }, + { + "type": "title", + "bbox": [ + 0.216, + 0.659, + 0.36, + 0.675 + ], + "angle": 0, + "content": "6 Conclusion" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.689, + 0.789, + 0.841 + ], + "angle": 0, + "content": "In this paper, we introduce a high-quality dataset, i.e., PanoVOS, for panoramic video object segmentation. Our PanoVOS dataset provides pixel-level instance annotations with diverse scenarios and significant motions. Based on this dataset, we evaluate 15 off-the-shelf VOS models and carefully analyze their limitations. Then, we further present our model, i.e., PSCFormer, which is equipped with the proposed panoramic space consistency transformer block. Our preliminary experiment demonstrates the effectiveness of our proposed model to enhance the segmentation performance and consistency in panoramic scenes. In conclusion, this provides a new challenge for video understanding, and we hope our PanoVOS dataset can attract more researchers to pay attention to panoramic videos." + } + ], + [ + { + "type": "header", + "bbox": [ + 0.339, + 0.115, + 0.728, + 0.129 + ], + "angle": 0, + "content": "PanoVOS: Bridging Non-panoramic and Panoramic Views" + }, + { + "type": "page_number", + "bbox": [ + 0.769, + 0.117, + 0.785, + 0.127 + ], + "angle": 0, + "content": "15" + }, + { + "type": "title", + "bbox": [ + 0.218, + 0.145, + 0.403, + 0.163 + ], + "angle": 0, + "content": "Acknowledgements" + }, + { + "type": "text", + "bbox": [ + 0.217, + 0.174, + 0.788, + 0.234 + ], + "angle": 0, + "content": "This work was supported in part by National Natural Science Foundation of China (No.62072112), and Scientific and Technological Innovation Action Plan of Shanghai Science and Technology Committee (No.22511101502, No.22511102202 and No.21DZ2203300)." + }, + { + "type": "title", + "bbox": [ + 0.218, + 0.255, + 0.323, + 0.271 + ], + "angle": 0, + "content": "References" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.282, + 0.785, + 0.324 + ], + "angle": 0, + "content": "1. Ai, H., Cao, Z., Zhu, J., Bai, H., Chen, Y., Wang, L.: Deep learning for omnidirectional vision: A survey and new perspectives. arXiv preprint arXiv:2205.10468 (2022)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.325, + 0.785, + 0.351 + ], + "angle": 0, + "content": "2. Armeni, I., Sax, S., Zamir, A.R., Savarese, S.: Joint 2d-3dsemantic data for indoor scene understanding. arXiv preprint arXiv:1702.01105 (2017)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.352, + 0.785, + 0.392 + ], + "angle": 0, + "content": "3. Caelles, S., Maninis, K.K., Pont-Tuset, J., Leal-Taixe, L., Cremers, D., Van Gool, L.: One-shot video object segmentation. In: Proceedings of the IEEE conference on computer vision and pattern recognition. pp. 221-230 (2017)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.393, + 0.785, + 0.433 + ], + "angle": 0, + "content": "4. Chen, X., Li, Z., Yuan, Y., Yu, G., Shen, J., Qi, D.: State-aware tracker for real-time video object segmentation. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 9384-9393 (2020)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.434, + 0.785, + 0.473 + ], + "angle": 0, + "content": "5. Cheng, H.K., Schwing, A.G.: Xmem: Long-term video object segmentation with an atkinson-shiffrin memory model. In: European Conference on Computer Vision. pp. 640-658. Springer (2022)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.474, + 0.785, + 0.514 + ], + "angle": 0, + "content": "6. Cheng, H.K., Tai, Y.W., Tang, C.K.: Rethinking space-time networks with improved memory coverage for efficient video object segmentation. Advances in Neural Information Processing Systems 34, 11781-11794 (2021)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.515, + 0.785, + 0.555 + ], + "angle": 0, + "content": "7. Cheng, H.T., Chao, C.H., Dong, J.D., Wen, H.K., Liu, T.L., Sun, M.: Cube padding for weakly-supervised saliency prediction in 360 videos. In: Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition. pp. 1420-1429 (2018)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.556, + 0.785, + 0.595 + ], + "angle": 0, + "content": "8. Cheng, J., Tsai, Y.H., Hung, W.C., Wang, S., Yang, M.H.: Fast and accurate online video object segmentation via tracking parts. In: Proceedings of the IEEE conference on computer vision and pattern recognition. pp. 7415-7424 (2018)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.597, + 0.785, + 0.636 + ], + "angle": 0, + "content": "9. Dai, J., Qi, H., Xiong, Y., Li, Y., Zhang, G., Hu, H., Wei, Y.: Deformable convolutional networks. In: Proceedings of the IEEE international conference on computer vision. pp. 764-773 (2017)" + }, + { + "type": "ref_text", + "bbox": [ + 0.22, + 0.637, + 0.785, + 0.677 + ], + "angle": 0, + "content": "10. Dang, J., Zheng, H., Xu, X., Guo, Y.: Unified spatio-temporal dynamic routing for efficient video object segmentation. IEEE Transactions on Intelligent Transportation Systems (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.22, + 0.678, + 0.785, + 0.717 + ], + "angle": 0, + "content": "11. Dang, J., Zheng, H., Xu, X., Wang, L., Hu, Q., Guo, Y.: Adaptive sparse memory networks for efficient and robust video object segmentation. IEEE Transactions on Neural Networks and Learning Systems (2024)" + }, + { + "type": "ref_text", + "bbox": [ + 0.22, + 0.718, + 0.785, + 0.759 + ], + "angle": 0, + "content": "12. Eger Passos, D., Jung, B.: Measuring the accuracy of inside-out tracking in xr devices using a high-precision robotic arm. In: International Conference on Human-Computer Interaction. pp. 19–26. Springer (2020)" + }, + { + "type": "ref_text", + "bbox": [ + 0.22, + 0.76, + 0.785, + 0.799 + ], + "angle": 0, + "content": "13. Fang, R., Yan, S., Huang, Z., Zhou, J., Tian, H., Dai, J., Li, H.: Instructseq: Unifying vision tasks with instruction-conditioned multi-modal sequence generation. arXiv preprint arXiv:2311.18835 (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.22, + 0.8, + 0.785, + 0.84 + ], + "angle": 0, + "content": "14. Guo, P., Hong, L., Zhou, X., Gao, S., Li, W., Li, J., Chen, Z., Li, X., Zhang, W., Zhang, W.: Clickvos: Click video object segmentation. arXiv preprint arXiv:2403.06130 (2024)" + }, + { + "type": "list", + "bbox": [ + 0.22, + 0.282, + 0.785, + 0.84 + ], + "angle": 0, + "content": null + } + ], + [ + { + "type": "page_number", + "bbox": [ + 0.218, + 0.116, + 0.236, + 0.127 + ], + "angle": 0, + "content": "16" + }, + { + "type": "header", + "bbox": [ + 0.272, + 0.115, + 0.338, + 0.127 + ], + "angle": 0, + "content": "Yan et al." + }, + { + "type": "ref_text", + "bbox": [ + 0.218, + 0.147, + 0.785, + 0.175 + ], + "angle": 0, + "content": "15. Guo, P., Huang, T., He, P., Liu, X., Xiao, T., Chen, Z., Zhang, W.: Openvis: Open-vocabulary video instance segmentation. arXiv preprint arXiv:2305.16835 (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.218, + 0.175, + 0.787, + 0.215 + ], + "angle": 0, + "content": "16. Guo, P., Zhang, W., Li, X., Zhang, W.: Adaptive online mutual learning bi-decoders for video object segmentation. IEEE Transactions on Image Processing 31, 7063-7077 (2022)" + }, + { + "type": "ref_text", + "bbox": [ + 0.218, + 0.216, + 0.785, + 0.242 + ], + "angle": 0, + "content": "17. Hendrycks, D., Gimpel, K.: Gaussian error linear units (gelus). arXiv preprint arXiv:1606.08415 (2016)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.243, + 0.785, + 0.297 + ], + "angle": 0, + "content": "18. Hong, L., Yan, S., Zhang, R., Li, W., Zhou, X., Guo, P., Jiang, K., Chen, Y., Li, J., Chen, Z., et al.: Onetracker: Unifying visual object tracking with foundation models and efficient tuning. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 19079-19091 (2024)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.298, + 0.785, + 0.352 + ], + "angle": 0, + "content": "19. Hu, L., Zhang, P., Zhang, B., Pan, P., Xu, Y., Jin, R.: Learning position and target consistency for memory-based video object segmentation. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 4144-4154 (2021)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.353, + 0.785, + 0.379 + ], + "angle": 0, + "content": "20. Iqbal, E., Safarov, S., Bang, S.: Misanet: Multi-similarity and attention guidance for boosting few-shot segmentation. arXiv preprint arXiv:2206.09667 (2022)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.379, + 0.785, + 0.42 + ], + "angle": 0, + "content": "21. Jiang, H., Jiang, G., Yu, M., Zhang, Y., Yang, Y., Peng, Z., Chen, F., Zhang, Q.: Cubemap-based perception-driven blind quality assessment for 360-degree images. IEEE Transactions on Image Processing 30, 2364-2377 (2021)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.42, + 0.785, + 0.46 + ], + "angle": 0, + "content": "22. Jost, T.A., Nelson, B., Rylander, J.: Quantitative analysis of the oculus rifts in controlled movement. Disability and Rehabilitation: Assistive Technology 16(6), 632-636 (2021)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.46, + 0.785, + 0.5 + ], + "angle": 0, + "content": "23. Kirillov, A., Mintun, E., Ravi, N., Mao, H., Rolland, C., Gustafson, L., Xiao, T., Whitehead, S., Berg, A.C., Lo, W.Y., et al.: Segment anything. arXiv preprint arXiv:2304.02643 (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.501, + 0.785, + 0.542 + ], + "angle": 0, + "content": "24. Li, M., Hu, L., Xiong, Z., Zhang, B., Pan, P., Liu, D.: Recurrent dynamic embedding for video object segmentation. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 1332-1341 (2022)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.542, + 0.785, + 0.582 + ], + "angle": 0, + "content": "25. Li, W., Fan, J., Guo, P., Hong, L., Zhang, W.: Hfvos: History-future integrated dynamic memory for video object segmentation. IEEE Transactions on Circuits and Systems for Video Technology (2024)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.582, + 0.785, + 0.623 + ], + "angle": 0, + "content": "26. Li, W., Guo, P., Zhou, X., Hong, L., He, Y., Zheng, X., Zhang, W., Zhang, W.: Onevos: Unifying video object segmentation with all-in-one transformer framework. arXiv preprint arXiv:2403.08682 (2024)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.623, + 0.785, + 0.649 + ], + "angle": 0, + "content": "27. Li, X., Cao, H., Zhao, S., Li, J., Zhang, L., Raj, B.: Panoramic video salient object detection with ambisonic audio guidance. arXiv preprint arXiv:2211.14419 (2022)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.649, + 0.785, + 0.69 + ], + "angle": 0, + "content": "28. Li, X., Wang, J., Xu, X., Li, X., Raj, B., Lu, Y.: Robust referring video object segmentation with cyclic structural consensus. In: Proceedings of the IEEE/CVF International Conference on Computer Vision. pp. 22236-22245 (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.691, + 0.785, + 0.745 + ], + "angle": 0, + "content": "29. Li, X., Wang, J., Xu, X., Peng, X., Singh, R., Lu, Y., Raj, B.: Qdformer: Towards robust audiovisual segmentation in complex environments with quantization-based semantic decomposition. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 3402-3413 (2024)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.745, + 0.785, + 0.8 + ], + "angle": 0, + "content": "30. Li, X., Wang, J., Xu, X., Yang, M., Yang, F., Zhao, Y., Singh, R., Raj, B.: Towards noise-tolerant speech-referring video object segmentation: Bridging speech and text. In: Proceedings of the 2023 Conference on Empirical Methods in Natural Language Processing. pp. 2283-2296 (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.8, + 0.785, + 0.84 + ], + "angle": 0, + "content": "31. Liang, S., Shen, X., Huang, J., Hua, X.S.: Video object segmentation with dynamic memory networks and adaptive object alignment. In: Proceedings of the IEEE/CVF International Conference on Computer Vision. pp. 8065-8074 (2021)" + }, + { + "type": "list", + "bbox": [ + 0.217, + 0.147, + 0.787, + 0.84 + ], + "angle": 0, + "content": null + } + ], + [ + { + "type": "header", + "bbox": [ + 0.339, + 0.115, + 0.728, + 0.129 + ], + "angle": 0, + "content": "PanoVOS: Bridging Non-panoramic and Panoramic Views" + }, + { + "type": "page_number", + "bbox": [ + 0.768, + 0.116, + 0.786, + 0.127 + ], + "angle": 0, + "content": "17" + }, + { + "type": "ref_text", + "bbox": [ + 0.214, + 0.147, + 0.788, + 0.189 + ], + "angle": 0, + "content": "32. Liang, Y., Li, X., Jafari, N., Chen, J.: Video object segmentation with adaptive feature bank and uncertain-region refinement. Advances in Neural Information Processing Systems 33, 3430-3441 (2020)" + }, + { + "type": "ref_text", + "bbox": [ + 0.214, + 0.19, + 0.787, + 0.233 + ], + "angle": 0, + "content": "33. Lin, T.Y., Maire, M., Belongie, S., Hays, J., Perona, P., Ramanan, D., Dollar, P., Zitnick, C.L.: Microsoft coco: Common objects in context. In: European conference on computer vision. pp. 740-755. Springer (2014)" + }, + { + "type": "ref_text", + "bbox": [ + 0.214, + 0.234, + 0.787, + 0.275 + ], + "angle": 0, + "content": "34. Liu, Y., Yu, R., Wang, J., Zhao, X., Wang, Y., Tang, Y., Yang, Y.: Global spectral filter memory network for video object segmentation. In: European Conference on Computer Vision. pp. 648-665. Springer (2022)" + }, + { + "type": "ref_text", + "bbox": [ + 0.214, + 0.276, + 0.787, + 0.331 + ], + "angle": 0, + "content": "35. Ma, C., Zhang, J., Yang, K., Roitberg, A., Stiefelhagen, R.: Densepass: Dense panoramic semantic segmentation via unsupervised domain adaptation with attention-augmented context exchange. In: 2021 IEEE International Intelligent Transportation Systems Conference (ITSC). pp. 2766-2772. IEEE (2021)" + }, + { + "type": "ref_text", + "bbox": [ + 0.214, + 0.332, + 0.787, + 0.374 + ], + "angle": 0, + "content": "36. Maninis, K.K., Caelles, S., Chen, Y., Pont-Tuset, J., Leal-Taixe, L., Cremers, D., Van Gool, L.: Video object segmentation without temporal information. IEEE transactions on pattern analysis and machine intelligence 41(6), 1515-1530 (2018)" + }, + { + "type": "ref_text", + "bbox": [ + 0.214, + 0.375, + 0.787, + 0.416 + ], + "angle": 0, + "content": "37. Mao, Y., Wang, N., Zhou, W., Li, H.: Joint inductive and transductive learning for video object segmentation. In: Proceedings of the IEEE/CVF International Conference on Computer Vision. pp. 9670-9679 (2021)" + }, + { + "type": "ref_text", + "bbox": [ + 0.214, + 0.417, + 0.787, + 0.472 + ], + "angle": 0, + "content": "38. Mei, J., Zhu, A.Z., Yan, X., Yan, H., Qiao, S., Chen, L.C., Kretzschmar, H.: Waymo open dataset: Panoramic video panoptic segmentation. In: Computer Vision-ECCV 2022: 17th European Conference, Tel Aviv, Israel, October 23-27, 2022, Proceedings, Part XXIX. pp. 53-72. Springer (2022)" + }, + { + "type": "ref_text", + "bbox": [ + 0.214, + 0.473, + 0.787, + 0.514 + ], + "angle": 0, + "content": "39. Oh, S.W., Lee, J.Y., Sunkavalli, K., Kim, S.J.: Fast video object segmentation by reference-guided mask propagation. In: Proceedings of the IEEE conference on computer vision and pattern recognition. pp. 7376-7385 (2018)" + }, + { + "type": "ref_text", + "bbox": [ + 0.214, + 0.515, + 0.787, + 0.557 + ], + "angle": 0, + "content": "40. Oh, S.W., Lee, J.Y., Xu, N., Kim, S.J.: Video object segmentation using space-time memory networks. In: Proceedings of the IEEE/CVF International Conference on Computer Vision. pp. 9226-9235 (2019)" + }, + { + "type": "ref_text", + "bbox": [ + 0.214, + 0.558, + 0.787, + 0.613 + ], + "angle": 0, + "content": "41. Perazzi, F., Pont-Tuset, J., McWilliams, B., Van Gool, L., Gross, M., Sorkine-Hornung, A.: A benchmark dataset and evaluation methodology for video object segmentation. In: Proceedings of the IEEE conference on computer vision and pattern recognition. pp. 724-732 (2016)" + }, + { + "type": "ref_text", + "bbox": [ + 0.214, + 0.614, + 0.787, + 0.655 + ], + "angle": 0, + "content": "42. Pont-Tuset, J., Perazzi, F., Caelles, S., Arbeláez, P., Sorkine-Hornung, A., Van Gool, L.: The 2017 davis challenge on video object segmentation. arXiv preprint arXiv:1704.00675 (2017)" + }, + { + "type": "ref_text", + "bbox": [ + 0.214, + 0.657, + 0.787, + 0.684 + ], + "angle": 0, + "content": "43. Rajic, F., Ke, L., Tai, Y.W., Tang, C.K., Danelljan, M., Yu, F.: Segment anything meets point tracking. arXiv preprint arXiv:2307.01197 (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.214, + 0.685, + 0.787, + 0.727 + ], + "angle": 0, + "content": "44. Seong, H., Hyun, J., Kim, E.: Kernelized memory network for video object segmentation. In: European Conference on Computer Vision. pp. 629-645. Springer (2020)" + }, + { + "type": "ref_text", + "bbox": [ + 0.214, + 0.728, + 0.787, + 0.769 + ], + "angle": 0, + "content": "45. Shi, J., Yan, Q., Xu, L., Jia, J.: Hierarchical image saliency detection on extended cssd. IEEE transactions on pattern analysis and machine intelligence 38(4), 717-729 (2015)" + }, + { + "type": "ref_text", + "bbox": [ + 0.214, + 0.77, + 0.787, + 0.812 + ], + "angle": 0, + "content": "46. Vaswani, A., Shazeer, N., Parmar, N., Uszkoreit, J., Jones, L., Gomez, A.N., Kaiser, L., Polosukhin, I.: Attention is all you need. Advances in neural information processing systems 30 (2017)" + }, + { + "type": "ref_text", + "bbox": [ + 0.214, + 0.813, + 0.787, + 0.84 + ], + "angle": 0, + "content": "47. Voigtlaender, P., Leibe, B.: Online adaptation of convolutional neural networks for video object segmentation. arXiv preprint arXiv:1706.09364 (2017)" + }, + { + "type": "list", + "bbox": [ + 0.214, + 0.147, + 0.788, + 0.84 + ], + "angle": 0, + "content": null + } + ], + [ + { + "type": "page_number", + "bbox": [ + 0.218, + 0.116, + 0.236, + 0.127 + ], + "angle": 0, + "content": "18" + }, + { + "type": "header", + "bbox": [ + 0.272, + 0.115, + 0.338, + 0.127 + ], + "angle": 0, + "content": "Yan et al." + }, + { + "type": "ref_text", + "bbox": [ + 0.218, + 0.148, + 0.785, + 0.189 + ], + "angle": 0, + "content": "48. Wang, H., Jiang, X., Ren, H., Hu, Y., Bai, S.: Swiftnet: Real-time video object segmentation. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 1296-1305 (2021)" + }, + { + "type": "ref_text", + "bbox": [ + 0.218, + 0.19, + 0.785, + 0.218 + ], + "angle": 0, + "content": "49. Wang, W., Zhou, T., Porikli, F., Crandall, D., Van Gool, L.: A survey on deep learning technique for video segmentation. arXiv preprint arXiv:2107.01153 (2021)" + }, + { + "type": "ref_text", + "bbox": [ + 0.218, + 0.218, + 0.785, + 0.259 + ], + "angle": 0, + "content": "50. Xiao, H., Feng, J., Lin, G., Liu, Y., Zhang, M.: Monet: Deep motion exploitation for video object segmentation. In: Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition. pp. 1140-1148 (2018)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.259, + 0.785, + 0.3 + ], + "angle": 0, + "content": "51. Xie, H., Yao, H., Zhou, S., Zhang, S., Sun, W.: Efficient regional memory network for video object segmentation. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 1286-1295 (2021)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.3, + 0.785, + 0.342 + ], + "angle": 0, + "content": "52. Xu, N., Yang, L., Fan, Y., Yue, D., Liang, Y., Yang, J., Huang, T.: Youtube-vos: A large-scale video object segmentation benchmark. arXiv preprint arXiv:1809.03327 (2018)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.342, + 0.785, + 0.384 + ], + "angle": 0, + "content": "53. Xu, X., Wang, J., Li, X., Lu, Y.: Reliable propagation-correction modulation for video object segmentation. In: Proceedings of the AAAI Conference on Artificial Intelligence. pp. 2946-2954 (2022)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.384, + 0.785, + 0.425 + ], + "angle": 0, + "content": "54. Xu, X., Wang, J., Ming, X., Lu, Y.: Towards robust video object segmentation with adaptive object calibration. In: Proceedings of the 30th ACM International Conference on Multimedia. pp. 2709-2718 (2022)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.425, + 0.785, + 0.467 + ], + "angle": 0, + "content": "55. Xu, X., Zhang, T., Wang, S., Li, X., Chen, Y., Li, Y., Raj, B., Johnson-Roberson, M., Huang, X.: Customizable perturbation synthesis for robust slam benchmarking. arXiv preprint arXiv:2402.08125 (2024)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.467, + 0.785, + 0.522 + ], + "angle": 0, + "content": "56. Yan, S., Zhang, R., Guo, Z., Chen, W., Zhang, W., Li, H., Qiao, Y., Dong, H., He, Z., Gao, P.: Referred by multi-modality: A unified temporal transformer for video object segmentation. In: Proceedings of the AAAI Conference on Artificial Intelligence. vol. 38, pp. 6449-6457 (2024)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.522, + 0.785, + 0.564 + ], + "angle": 0, + "content": "57. Yang, Z., Wei, Y., Yang, Y.: Collaborative video object segmentation by foreground-background integration. In: European Conference on Computer Vision. pp. 332-348. Springer (2020)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.564, + 0.785, + 0.605 + ], + "angle": 0, + "content": "58. Yang, Z., Wei, Y., Yang, Y.: Associating objects with transformers for video object segmentation. Advances in Neural Information Processing Systems 34, 2491-2502 (2021)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.605, + 0.785, + 0.647 + ], + "angle": 0, + "content": "59. Yang, Z., Wei, Y., Yang, Y.: Collaborative video object segmentation by multiscale foreground-background integration. IEEE Transactions on Pattern Analysis and Machine Intelligence (2021)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.647, + 0.785, + 0.674 + ], + "angle": 0, + "content": "60. Yuan, M., Richardt, C.: 360 optical flow using tangent images. In: British Machine Vision Conference (BMVC) (2021)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.674, + 0.785, + 0.702 + ], + "angle": 0, + "content": "61. Zhang, R., Jiang, Z., Guo, Z., Yan, S., Pan, J., Dong, H., Gao, P., Li, H.: Personalize segment anything model with one shot. arXiv preprint arXiv:2305.03048 (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.702, + 0.785, + 0.743 + ], + "angle": 0, + "content": "62. Zhang, Y., Zhang, L., Wang, K., Hamidouche, W., Deforges, O.: Shd360: A benchmark dataset for salient human detection in 360 videos. arXiv preprint arXiv:2105.11578 (2021)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.743, + 0.785, + 0.785 + ], + "angle": 0, + "content": "63. Zhang, Y., Wu, Z., Peng, H., Lin, S.: A transductive approach for video object segmentation. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 6949-6958 (2020)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.785, + 0.785, + 0.813 + ], + "angle": 0, + "content": "64. Zhang, Z., Xu, Y., Yu, J., Gao, S.: Saliency detection in 360 videos. In: Proceedings of the European conference on computer vision (ECCV). pp. 488-503 (2018)" + }, + { + "type": "list", + "bbox": [ + 0.217, + 0.148, + 0.785, + 0.813 + ], + "angle": 0, + "content": null + } + ] +] \ No newline at end of file diff --git a/2024/PanoVOS_ Bridging Non-panoramic and Panoramic Views with Transformer for Video Segmentation/5c945af1-cf63-4f10-9827-c4fd88e7d8a2_origin.pdf b/2024/PanoVOS_ Bridging Non-panoramic and Panoramic Views with Transformer for Video Segmentation/5c945af1-cf63-4f10-9827-c4fd88e7d8a2_origin.pdf new file mode 100644 index 0000000000000000000000000000000000000000..b5bf99e6f3057d61eddffdf1c3f8cdc438634efe --- /dev/null +++ b/2024/PanoVOS_ Bridging Non-panoramic and Panoramic Views with Transformer for Video Segmentation/5c945af1-cf63-4f10-9827-c4fd88e7d8a2_origin.pdf @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:79227e5b05b0275cb32bbb3b8a710c3a078e1620648054427b40a8df50479fdb +size 16009173 diff --git a/2024/PanoVOS_ Bridging Non-panoramic and Panoramic Views with Transformer for Video Segmentation/full.md b/2024/PanoVOS_ Bridging Non-panoramic and Panoramic Views with Transformer for Video Segmentation/full.md new file mode 100644 index 0000000000000000000000000000000000000000..634c63abd40ff94f4f1815d0a846b7ea3da0790c --- /dev/null +++ b/2024/PanoVOS_ Bridging Non-panoramic and Panoramic Views with Transformer for Video Segmentation/full.md @@ -0,0 +1,333 @@ +# PanoVOS: Bridging Non-panoramic and Panoramic Views with Transformer for Video Segmentation + +Shilin Yan $^{1}$ , Xiaohao Xu $^{2}$ , Renrui Zhang $^{3}$ , Lingyi Hong $^{1}$ , Wenchao Chen $^{1}$ , Wenqiang Zhang $^{1}$ , and Wei Zhang $^{1\star}$ + +$^{1}$ Shanghai Key Lab of Intelligent Information Processing, School of Computer Science, Fudan University $^{2}$ University of Michigan, Ann Arbor $^{3}$ MMLab CUHK $^{4}$ tattoo.ysl@mail.com weizh@fudan.edu.cn + +![](images/b65bc70a30bd22450f37c9d5d3d5f285f3e04f37f97eb5930cc855143aea5ca7.jpg) +Fig.1: Panoramic video object segmentation (PanoVOS). PanoVOS targets tracking and distinguishing the particular instances under content discontinuities (e.g. penguin in the image of $T = 15$ ) and serve distortion (e.g. penguin in the image of $T = 65$ ). We show the sample of (a) frames, (b) segmentation annotations, and (c) area proportion of foreground for the Penguin video in our dataset. + +Abstract. Panoramic videos contain richer spatial information and have attracted tremendous amounts of attention due to their exceptional experience in some fields such as autonomous driving and virtual reality. However, existing datasets for video segmentation only focus on conventional planar images. To address the challenge, in this paper, we present a panoramic video dataset, i.e., PanoVOS. The dataset provides 150 videos with high video resolutions and diverse motions. To quantify the domain gap between 2D planar videos and panoramic videos, we evaluate 15 off-the-shelf video object segmentation (VOS) models on PanoVOS. Through error analysis, we found that all of them fail to tackle pixel-level content discontinues of panoramic videos. Thus, we present a Panoramic Space Consistency Transformer (PSCFormer), which can effectively utilize the semantic boundary information of the previous frame for pixel-level matching with the current frame. Extensive experiments demonstrate that compared with the previous SOTA models, our PSCFormer network exhibits a great advantage in terms of segmentation results under the panoramic setting. Our dataset poses new challenges in panoramic VOS and we hope that + +our PanoVOS can advance the development of panoramic segmentation/tracking. The dataset, codes, and pre-train models will be published at https://github.com/shilinyan99/PanoVOS. + +# 1 Introduction + +Semi-supervised video object segmentation (VOS) [52], which targets tracking and distinguishing the particular instances across the entire video sequence based on the first frame masks, plays an essential role in video understanding and editing. Conventionally, the images or videos studied in VOS are 2D planar data with a limited Field of View (FoV), which may lead to some ambiguities, especially when objects are out of view. Meanwhile, with the rapid development of VR/AR collection devices [12,22], panoramic videos with a $360^{\circ} \times 180^{\circ}$ FoV are able to collect the entire viewing sphere and richer spatial information [1,21, 27, 60]. To the best of our knowledge, we are the first to attempt to tackle the promising but challenging task of panoramic video object segmentation. + +To foster the development of panoramic VOS, we propose a new dataset in this work, aiming at panoramic video object segmentation. The dataset contains a wide range of real-world scenarios in which scenes have a large magnitude of motion. The main characteristics of our dataset are three aspects. 1) Panoramic videos bring certain advantages (richer geometric information and wider FoV) in real-world applications as well as challenges (serve distortion and content discontinuities). 2) Compared to all existing VOS datasets, our dataset has longer video clips with an average length of 20 seconds. 3) Nearly half of the video resolutions in our dataset are $4K$ , which may help facilitate broader video tracking/segmentation research under the high-resolution scenario. + +In the proposed dataset, we annotated 150 videos with 19,145 annotated instance masks, including sports (e.g. parkour, skateboard), animals (e.g. elephant, monkey), and common objects (basketball, hot balloon). Since, annotating a pixel-level intensive task is very time-consuming and expensive, we proposed a semi-supervised human-computer joint annotation strategy. Concretely, we first annotated objects at selected keyframes (1 fps) Then we adopted the state-of-the-art video object segmentation model AOT [58] for mask propagation to the rest frames of videos and we manually refine parts of them. + +Then, we conducted extensive experiments on PanoVOS to evaluate 15 off-the-shelf video object segmentation models. The results suggest that existing approaches can not handle several domain-unique challenges. The first is content discontinuities, which means the foreground object may be separated in the left and the right boundaries of the planar image, such as the case in the image of $T = 15$ in Fig. 1. The second is the severe distortions and deformations, such as the case in the image of $T = 65$ in Fig. 1. + +To tackle these challenges of panoramic video segmentation, we proposed a PSC-Former model which consists of key component Panoramic Space Consistent (PSC) blocks. The PSC block is designed for constructing spatial-temporal class-agnostic correspondence and propagating the segmentation masks. Each PSC + +block utilizes a cross-attention for matching with references' embeddings and a PSC-attention for modeling the boundary semantic relationship between the previous frame and the query frame. Hence, the network can effectively alleviate the problem that the left and the right boundaries are actually continuous in panoramic videos. Our method outperforms the SOTA models that are re-trained on PanoVOS train set in segmentation quality under the panoramic setting. Our contributions are three-fold. + +- We introduce a panoramic video object segmentation dataset (PanoVOS) with 150 videos and 19K annotated instance masks, which fills the gap of long-term instance-level annotated panoramic video segmentation datasets. +- Extensive experiments are conducted on 15 off-the-shelf VOS benchmarks and our baseline model on PanoVOS, which reveals that current methods could not tackle content discontinuities in panoramic videos well. +- We propose a Panoramic Space Consistency Transformer (PSCFormer) on PanoVOS that successfully resolves the challenges of discontinuity of pixel-level content segmentation. + +# 2 Related Work + +# 2.1 Panoramic Datasets + +In this paper, panoramic videos refers to complete (360°, full view) panoramic videos, which is different from the definition in [38], which only include wide but partial views of some range-view images collected from multiple cameras. + +Image-based panoramic datasets. Existing popular image-level panoramic segmentation datasets are Stanford2D3D [2] and DensePASS [35]. The former one is mainly focused on indoor spaces including a total of 1,413 panoramic images with instance-level annotations in 13 categories. The latter targets driving scenes in cities. DensePASS [35] provides only 100 labeled panoramic images for testing and 2,000 unlabeled images for cross-domain transfer optimization. + +Video-based panoramic datasets. Video-based benchmarks mainly include SHD360 [62], SOD360 [64] and Wild360 [7]. All of them are used for panoramic video saliency object detection. Specifically, 1) SHD360 only targets human-centric video scenes with little movement. It provides 6,268 object-level pixel-wise masks and 16,238 instance-level pixel-wise masks. 2) SOD360 focuses on the sports-centric scenario with 41 video clips (12 outdoor and 29 indoor). 3) Wild360 concentrates on natural scenes with 85 videos. Note that SOD360 and Wild360 have no object-level or instance-level annotations. + +We make a comparison with the existing video panoramic datasets in Table 1. Specifically, our PanoVOS dataset contains 150 videos mainly from three different domains: person, animal, and common object, which makes the dataset more general for object-agnostic evaluations. Besides, videos in our dataset have a relatively large range of motion, making our PanoVOS dataset suitable for video tracking and segmentation evaluation tasks under panoramic scenes. Moreover, the average duration of each video in our dataset is $20s$ , which is about 4 times + +
DatasetsMotion#Videos#Frames#Total MasksAverage Duration
SHD360 [62]Small416,26816,2385s
SOD360 [64]Large104N/A0N/A
Wild360 [7]Large85N/A0N/A
PanoVOSLarge15013,99519,14520s
+ +Table 1: Comparison of panoramic video datasets. Our PanoVOS is the first long-term panoramic video segmentation dataset with instance-level masks. Compared with existing panoramic video datasets [7,62,64] that are used for saliency detection, our panoramic video dataset for video segmentation, i.e., PanoVOS, includes more diverse and larger motion, making it suitable for dense video tracking evaluation. + +longer than SHD360 [62] (5s per video). By the way, the longer video is highlighted in a recent survey [49]. The longer the video, the more likely it is to introduce more panoramic video characteristics such as distortion and discontinuity, which is more challenging and more practical. + +# 2.2 Video Object Segmentation Datasets + +The establishment of DAVIS [41, 42] and YouTube-VOS [52] datasets pave the way for the boosting development of VOS methods. They are collected by traditional pinhole cameras and the duration of each video clip is very short, only 5s on average. In contrast, the average video length in the proposed PanoVOS dataset is 20s, which is 4 times longer than the existing video datasets. Our dataset includes more challenging scenes (e.g. distortion and discontinuity) that is non-negligible in real-world applications. + +# 2.3 Video Object Segmentation Methods + +Existing video object segmentation methods can be roughly classified into three subsets: online-learning-based, propagation-based, and matching-based. + +Online learning-based. Online learning-based approaches [3,36,50], which either train or fine-tune their networks with the first-frame ground truth at test time and are therefore a great waste of resources. OnAVOS [47] achieves promising results by introducing an online adaptation mechanism, but it still requires online fine-tuning. To a certain extent, it restricts networks' efficiency. + +Propagation-based. Propagation-based models [4,8,39] get the target masks in a frame-to-frame prerogation way. Although propagation-based methods improve efficiency, they lack long-term context and therefore are difficult to handle object disappearance and reappearance, severe obscuration, and distortion. + +Matching-based. Matching-based methods [6, 10, 11, 14, 16, 25, 26, 31, 37, 40, 53, 54, 63] aim to learn an embedding space of target objects between query and memory. Recently state-of-the-art methods encode many frames into embeddings and store them as a feature memory bank. The most representative is STM [40], + +![](images/9f598b6b7a822982a3a0723c6809c280df59c6b5aa85b81de402b40d4a2c42d2.jpg) + +![](images/ea645df97cdbf5de7919a92a916cc7cf1dc91bfaf9d2b53ce04b066645e454a9.jpg) +Fig. 2: PanoVOS dataset. We select 10 samples from the dataset involving major scenes. For each video, there are high-quality instance-level pixel-wise masks. + +![](images/b1eb488d17e5b0bda1eefbc9bf0f4c163341a3680d52acd78d06cc8e85c60134.jpg) + +![](images/743cea0236f2b529412fa835efc5a45de1e3e9108d9594a2ae9d859b8413b313.jpg) + +![](images/001a3c1c54c107ee62c30e8b0967b8c9140dca22575ab534f8cdc0e459341807.jpg) + +![](images/562a3ed9d72a4ee8e0fccf8445f742d4d3c9b391724fcf312de8fcee7a305d15.jpg) + +![](images/69f21cea20fdc4cc0ee8e1dd102b1de00304555e80ba4f09cf21677b311671f4.jpg) + +![](images/05e4a16e9f6a76b30ad78fb01c6b15e3aeff2b93ea389258578199259f564b42.jpg) + +![](images/d9ee8a8bbe976ef7a549c1d34a45b73112bcc182363e6857d87f06a41f61bcf4.jpg) + +![](images/3d76807a3ce47afba0c1a90987e61102f800ee9f5bb84839536180a29490442c.jpg) + +
SplitsTrainValTest
#Videos8035 (10 unseen)35 (10 unseen)
#Images7,070(50.5%)3,464(24.8%)3,461(24.7%)
#Masks9,585(50.1%)4,957(25.9%)4,603(24.0%)
+ +Table 2: Statistics of PanoVOS dataset + +which has been extended to many works [5, 19, 34, 44, 48, 51]. AOT [58] introduces an identification mechanism by encoding multiple targets into the same embedding space, which can simultaneously segment multiple objects. However, they fail to address the challenges of the tremendous proportion of distortion and discontinuity under panoramic setting. + +# 3 PanoVOS Dataset + +We introduced the proposed PanoVOS dataset in three parts, (1) collection process, (2) statistical summary, and (3) annotation pipeline. + +# 3.1 Data collection + +We built our PanoVOS dataset with the principle of diversity in mind. Moreover, the objects in the video should have a large amplitude of motion or camera movement. Based on the above viewpoint, we collected videos from the YouTube website for further annotation, respectively. The range of the video length is from 3 to 40 seconds. The average sequence length of each video in the dataset is approximately 20 seconds. We followed the settings of YouTube-VOS [52] to sample the frames at 6 fps. + +# 3.2 Dataset Statistics + +PanoVOS contains 150 videos, including 13,995 frames and 19,145 instance annotations from 35 categories. The average length of each video is 20 seconds. We believe that visual categories are representative of common life scenarios, and Fig.2 shows some samples of PanoVOS. To create our PanoVOS, in the + +![](images/66c37404571ed1221551acfb4db3f3ee0b2f465bbc13a548d08f8438d84408fb.jpg) +Fig. 3: Instance-level distribution of PanoVOS dataset. Our dataset contains three major divisions: person, animals, and common objects with 35 sub-divisions. + +spirit of the video object segmentation task, we carefully selected videos with relatively large motion amplitudes and chose a set of video categories including person (e.g. parkour, dance, BMX, skateboard), animals (e.g. elephant, monkey, giraffe, rhino, birds) and common objects (e.g. basketball, hot balloon) as shown in Fig 3. PanoVOS dataset consists of 150 videos split into training (80), validation (35), and test (35) sets. Table 2 shows detailed division results. Both the validation and test sets have 35 videos (about $23\%$ of the frames and the masks). For validation and test sets, we keep some unseen visual categories for generalization ability evaluation. + +# 3.3 Annotation Pipeline + +Annotation is very time-consuming and expensive for a pixel-level panoramic segmentation dataset. To obtain accurate large-scale video panoramic segmentation annotations and make the process more efficient, we propose a semi-automatic human-computer joint annotation strategy, as shown in Fig 4. First, keyframes are selected and manually annotated for each video, which are images with a speed of 1 fps. This is followed by a frame-by-frame propagation from the annotated keyframes to those unlabeled intermediate frames with a sophisticated semi-supervised VOS model. Then, to tackle the distortions and discontinuities in panoramic videos, we need to re-calibrate the resulting annotations via human refinement. More details will unfold below. + +Annotation Propagation For the annotation of each video, we first need an expert to browse the current video and note down all objects that have a large amplitude of movement. Then, for each video, the recorded objects in keyframes with a speed of 1 fps are selected for manual annotation. To avoid consistency errors or the problem of objects being labeled as other instances when they disappear and reappear, another expert needs to double-check the annotations of all objects to improve the accuracy of the dataset annotation. + +![](images/4dec3a67269029bede7f37571e681b2d68e66e02b8dbbe1df1c13fa14295ec52.jpg) +Fig. 4: PanoVOS annotation pipeline. Our annotation pipeline includes two phases. (1) The first phase is called Key Frames Select and Annotate. The annotator browses the video and picks out the object to be annotated. Then, instances are manually annotated at 1 fps and corrected by another annotator. (2) The second phase is called All Frames Propagate and Refine. In this phase, we apply a semi-supervised video object segmentation model to help propagate the annotated masks and the generated instances are refined by annotators. + +We then use the off-the-shelf video labeling method [58] to propagate the instance masks frame by frame from the annotated keyframes to untagged intermediate frames and generate masks at 6 fps. + +Annotation Refinement To present a new Panoramic dataset of high quality. After obtaining masks of the first propagation stage, annotators are asked to check the quality of the masks and refine them. The main amendments are in the following two areas. 1). Since our video resolution is generally relatively high, the propagation method will often fail when encountering complex videos with many small objects in a scene. 2) Due to the huge distortions and discontinuities present in the panoramic video, the quality of the masks obtained is relatively poor. Manual correction of the mask is checked by another annotator until the result is satisfactory before proceeding to the next video annotation. + +# 4 Method + +# 4.1 Overview + +Video object segmentation targets assigning an instance label to every pixel in the given video sequence based on the first frame mask. Recent works [5,6,13,59] have demonstrated that the attention mechanism can significantly help improve the segmentation performance. However, for the challenge of content discontinuation in panoramic videos, only considering the original attention mechanism will not be able to fully utilize the semantic information on the left and right boundaries (pixel contiguity) in the spatial dimension and will lose valuable contextual information when segmenting objects. Therefore, in this work, our mission is to design an effective network architecture, which can help acquire valuable boundary relationships. + +![](images/20e411bca2ba711233ee86b4077a06ef30b8f2cf913068acf309d6179ff7e29c.jpg) +Fig. 5: (a) PSCFormer overview. Given the query frame $\mathbf{x}_t$ and reference frames $\{\mathbf{x}_i | i \in \mathcal{R}\}$ , the goal of VOS is to delineate objects from the background by generating mask $\mathbf{y}_t$ for query frame $\mathbf{x}_t$ . References and the query frame are encoded by the memory encoder and query encoder, respectively. Multiple stacking panoramic space consistency (PSC) blocks are used to leverage the correspondence in the panoramic space between references and the query frame. A decoder is used for generating the prediction of the query frame. (b) Panoramic space consistency block architecture details. + +Fig. 5(a) illustrates the overall architecture of the proposed network. Given the query frame $\mathbf{x}_t$ and references $\{\mathbf{x}_i | i \in \mathcal{R}\}$ , the goal of VOS is to delineate objects from the background by generating mask $\mathbf{y}_t$ for query frame $\mathbf{x}_t$ . Following [57], our basic setting uses the first and previous frame as references $\mathcal{R} = \{1, t - 1\}$ . The memory encoder and query encoder are responsible for extracting frame-level features. After this, the panoramic space consistency block takes them as input and aggregates the spatial-temporal information between the reference frames and the query frame at the pixel level. Finally, the decoder uses the output of the sequence stacking PSC blocks to predict the mask of the object. + +# 4.2 Panoramic Space Consistency Block + +Fig. 5(b) shows the structure of a PSC block. Motivated by the common transformer blocks [46], PSC firstly contains a self-attention layer, which is used to aggregate the target objects' correlation information within the query frame. Then, the middle module is composed of cross-attention and PSC-attention, in which cross-attention is responsible for learning the target objects' information from references $\mathcal{R}$ and the PSC-attention targets on exploring the boundary relationship between the query frame and previous frame. Finally, PSC employs a two-layer feed-forward MLP with GELU [17] non-linearity activation function. + +Panoramic Space Consistency Attention (PSC-Attn). PSC-Attn is employed to model the spatial-temporal relationship between the query frame and reference frames considering the continuity of pixels of images in the panoramic space. How to establish a connection between the left and right boundaries be + +come especially important? The most intuitive solution would be to directly splice in length, but this would lead to a huge amount of computation. Therefore, we take the approach of moving a portion of the region in the length dimension from the right boundary to the leftmost boundary for stitching. Consequently, we only focus on the left and right boundaries between the query frame and the reference frame. Thus, unlike the original attention, where each query token is counted for attention along with all key tokens in the reference frame, our PSC attention takes care of the key tokens in a fixed window size. In particular, we define the reference frame feature embedding $\mathbf{f}(\mathbf{x})\in \mathbb{R}^{H\times W\times C}$ , which is extracted from the query encoder. $H$ , $W$ , and $C$ represent the height, width, and channel dimensions, respectively. According to the solutions mentioned above, the new feature embedding $\mathbf{f}(\mathbf{x})'$ is calculated as follows: + +$$ +\mathbf {f} (\mathbf {x}) ^ {\prime} [ 0: W / p ] = \mathbf {f} (\mathbf {x}) [ W / p: W ] +$$ + +$$ +\mathbf {f} (\mathbf {x}) ^ {\prime} [ W / p: W ] = \mathbf {f} (\mathbf {x}) [ 0: W / p ] \tag {1} +$$ + +$$ +\mathbf {f} (\mathbf {x}) ^ {\prime} [ W / p: W - W / p ] = \mathbf {f} (\mathbf {x}) [ W / p: W - W / p ], +$$ + +where $p \in \mathbb{Z}^+$ . We define query embedding $Q \in \mathbb{R}^{HW \times C}$ , key embedding $K \in \mathbb{R}^{HW \times C}$ , value embedding $V \in \mathbb{R}^{HW \times C}$ , where $Q$ is from the query frame feature embedding, $K$ and $V$ are from $\mathbf{f}(\mathbf{x})'$ by performing dimensional transformations. Mathematically, we define the PSC attention as follows, + +$$ +\operatorname {P S C A t t n} (Q, K, V) = \operatorname {s o f t m a x} \left(\frac {Q K ^ {T} \mathbf {R}}{\sqrt {C}}\right) V, \tag {2} +$$ + +where $\mathbf{R} \in [0,1]^{HW \times HW}$ means a window that represents the attention range of each query token. For query $Q_{(x,y)}$ at $(x,y)$ position, we define the $\mathbf{R}_{(x,y)}$ as: + +$$ +\mathbf {R} _ {x, y} (i, j) = \left\{ \begin{array}{l l} 1 & \text {i f} (x - i) ^ {2} \leqslant s ^ {2} \text {a n d} (y - j) ^ {2} \leqslant s ^ {2} \\ 0 & \text {o t h e r w i s e} \end{array} , \right. \tag {3} +$$ + +where $(i,j)$ is the position for each key token, $s$ is the window size. For each query token, it calculates the attention with another key token only if they are spatially limited to a $(2\times s + 1)$ size window, which significantly reduces the time complexity from $(h\times w)^{2}$ to $(2\times s + 1)^{2}$ . + +Following [46], we implement the representational form of our PSCAttn module with multi-headed attention, defined mathematically as follows, + +$$ +\operatorname {M u l t i H e a d} (Q, K, V) = \operatorname {C o n c a t} \left(\operatorname {h e a d} _ {1}, \dots , \operatorname {h e a d} _ {h}\right) W ^ {O} +$$ + +$$ +\operatorname {h e a d} _ {i} = \operatorname {P S C A t t n} \left(Q W _ {i} ^ {Q}, K W _ {i} ^ {K}, V W _ {i} ^ {V}\right), \tag {4} +$$ + +where $W_{i}^{Q}\in \mathbb{R}^{C\times d_{model}}$ , $W_{i}^{K}\in \mathbb{R}^{C\times d_{model}}$ , $W_{i}^{V}\in \mathbb{R}^{C\times d_{model}}$ and $W_{i}^{O}\in \mathbb{R}^{C\times C}$ are the linear projections. As [46], we set the number of heads to $(h = C / d_{model})$ 8, where $d_{model}$ is the projection dimension of each head. + +
Methods|MFYouTube-VOSPanoVOS ValidationPanoVOS Test
J&FJ&FJsFsJuFuJ&FJsFsJuFu
AOTT [58]73.753.8↓19.944.458.346.965.743.7↓30.036.349.839.649.2
AOTS [58]74.655.8↓18.849.162.246.465.544.7↓29.932.443.346.556.8
AOTB [58]75.253.7↓21.546.258.146.364.139.5↓35.734.444.435.044.4
AFB-URR [32]✓65.240.1↓25.131.141.535.851.830.4↓34.823.132.728.836.9
STCN [6]✓76.149.9↓26.242.753.445.158.448.0↓28.139.350.246.755.7
XMem [5]✓77.048.6↓28.440.750.144.858.640.2↓36.835.344.936.444.0
AOTL [58]✓74.749.2↓25.543.357.338.957.138.2↓36.532.343.732.744.1
R50 AOTL [58]✓76.550.1↓26.444.558.640.357.241.4↓35.133.745.038.348.4
SwinB_AOTL [58]✓74.444.8↓29.639.152.234.953.036.2↓38.231.142.031.040.6
RDE* [24]✓61.743.1↓18.636.048.435.252.741.3↓20.430.944.641.448.5
STCN* [6]✓56.343.2↓13.141.653.733.244.538.0↓18.332.843.235.540.4
XMem* [5]✓65.855.9↓9.952.264.047.260.049.6↓16.239.252.646.859.9
+ +Table 3: Domain transfer result of (static image datasets) $\rightarrow$ (PanoVOS Validation & Test). Subscript $s$ and $u$ denote scores in seen and unseen categories. $MF$ denotes multiple historical frames as reference. $\downarrow$ represents the performance of the declining values compared to the YouTube-VOS dataset [52]. * denotes a large-scale external dataset BL30K [6] dataset is used during training. + +
Methods|MF|YouTube-VOS |J&F|PanoVOS ValidationPanoVOS Test
J&FJsFsJuFuJ&FJsFsJuFu
CFBI†[59]81.460.9↓20.553.065.256.369.049.0↓32.449.447.646.252.6
CFBL+†[59]82.857.6↓25.252.167.048.163.453.7↓29.151.659.346.657.5
AOTT[58]80.261.5↓18.755.667.754.668.252.6↓27.644.855.351.558.8
AOTS[58]82.666.7↓15.958.070.562.076.457.3↓25.350.261.054.663.5
AOTB[58]83.570.5↓13.059.271.768.582.760.8↓22.753.064.457.868.2
AFB-URR [32]✓79.655.1↓24.544.755.653.466.752.4↓27.243.654.252.059.9
RDE[24]✓81.954.7↓27.250.363.944.660.155.4↓26.545.559.251.065.9
STCN[6]✓83.061.8↓21.250.363.561.372.153.4↓29.646.258.949.059.9
XMem[5]✓85.766.1↓19.656.668.762.077.262.5↓23.253.165.461.170.4
AOTL[58]✓83.871.9↓11.962.175.367.482.862.1↓21.757.169.056.266.1
R50_AOTL[58]✓84.169.2↓14.956.769.467.583.161.4↓22.757.569.053.365.7
SwinB_AOTL[58]✓84.567.5↓17.060.273.660.376.060.9↓23.653.963.758.767.4
RDE* [24]✓83.360.9↓22.451.464.756.071.655.6↓27.748.160.852.661.0
STCN* [6]✓84.361.7↓22.649.961.859.775.555.8↓28.548.259.852.762.5
XMem* [5]✓86.163.4↓22.753.564.461.574.161.0↓25.153.565.157.568.0
+ +Table 4: Domain transfer result of (static image datasets & YouTubeVOS) $\rightarrow$ (PanoVOS Validation & Test). Subscript $s$ and $u$ denote scores in seen and unseen categories. $MF$ denotes multiple historical frames as reference. $\downarrow$ represents the performance of the declining values compared to the YouTube-VOS dataset [52]. * denotes a large-scale external dataset BL30K [6] dataset is used during training. $\dagger$ denotes no synthetic data is used during the training stage. + +# 5 Experiment + +In this section, we design a series of experiments to answer the following research questions related to how to tackle video object segmentation in panoramic scenes: + +RQ1: How well are current VOS methods trained on non-panoramic videos adapted to the panoramic world? +RQ2: How well do variations of the foundation model Segment Anything Model [23] adapt to the panoramic world? +RQ3: Can the proposed PanoVOS datasets bring about a consistent performance gain to VOS methods? +RQ4: How well does Panoramic Space Consistency Attention contribute? +RQ5: What are the remained problems for panoramic-related research? + +
MethodsPanoVOS ValidationPanoVOS Test
J&FJsFsJuF_uJ&FJsFsJuF_u
PerSAM [61]19.112.319.817.427.119.57.414.923.831.7
SAM-PT [43]47.536.748.646.058.741.031.140.540.252.3
SAM-PT-reinit [43]43.734.344.341.354.943.635.042.743.553.0
+ +Table 5: Quantitative comparison on PanoVOS for variations of foundation model Segment Anything Model [23]. Subscript $s$ and $u$ denote scores in seen and unseen categories. + +
MethodsMFPanoVOS ValidationPanoVOS Test
J&FJsFsJuFuJ&FJsFsJuFu
CFBI† [57]35.834.644.824.239.719.118.226.112.219.8
CFBI+† [59]41.338.047.932.546.930.930.842.721.428.5
AOTT [58]65.659.468.359.775.053.449.361.647.555.1
AOTS [58]67.761.270.062.477.155.953.265.148.657.0
AOTB [58]67.662.372.061.574.855.453.564.247.756.0
Ours-Based74.066.480.466.283.056.849.462.752.462.5
AFB-URR [32]✓34.334.842.824.934.534.228.238.832.936.8
RDE [24]✓50.549.758.439.254.942.536.946.638.548.2
STCN [6]✓52.051.260.841.554.550.843.656.549.353.7
XMem [5]✓55.754.863.345.259.753.549.562.647.154.8
AOTL [58]✓66.661.471.159.474.353.850.060.347.857.1
R50-AOTL [58]✓65.361.971.456.471.654.652.963.247.554.9
SwinB-AOTL [58]✓62.158.966.554.368.853.149.057.849.056.6
Ours-Large✓77.970.585.269.586.459.954.969.253.062.4
RDE* [24]✓54.352.861.644.658.252.244.556.049.359.1
STCN* [6]✓51.751.260.641.353.653.853.758.146.057.3
XMem* [5]✓57.755.664.648.661.957.951.364.553.262.7
+ +Table 6: Quantitative comparison on PanoVOS for models with pretraining on static image datasets. Subscript $s$ and $u$ denote scores in seen and unseen categories. ${MF}$ denotes multiple historical frames as reference. * denotes a large-scale external dataset BL30K [6] dataset is used during training. + +# 5.1 Implementation Details + +Model Architecture. We build two variants of our method with different reference bank sizes $\mathcal{R}$ for a fair comparison with previous methods. Ours-Base uses only the first frame and the previous frame as reference ( $\mathcal{R} = \{1, t - 1\}$ ), which are for the sake of high inference speed and low memory consumption. OursLarge uses multiple historical frames as reference ( $\mathcal{R} = \{1 + 2\delta, 1 + 2\delta, 1 + 3\delta \ldots\}$ ), which follows [34,58]. In our work, we set $\delta$ to 2 and 5 for training and testing respectively. For $p$ and $s$ in PSC block, we set them as 2 and 7. + +Evaluation Metrics. Following the standard protocol [41, 42], we adopt the region accuracy $\mathcal{I}$ and boundary accuracy $\mathcal{F}$ . $\mathcal{I}$ means the Jaccard Index/Intersection over Union (IoU), which is the ratio of intersection and the joint area between predicted masks and ground truths. And $\mathcal{F}$ evaluates the accuracy of the segmentation boundary, which is computed by transforming it into a bipartite graph matching problem with predicted masks and ground truths. + +# 5.2 Domain Transfer Results (RQ1) + +We evaluate previous SOTA methods, which are trained on conventional datasets that are captured by pinhole cameras, on PanoVOS datasets to evaluate the domain transfer performance. To quantify the transfer performance of advanced + +![](images/800c148b3ad0066f6c6b8b8d2f5dafc8655f2717768b8d7ef6d72cf91a6cdcc7.jpg) +Fig.6: Qualitative comparison to the state-of-the-art methods, RDE [24], STCN [6], and XMem [5], on PanoVOS dataset. Our model performs better under the challenge of content discontinuities. Error regions are bounded. + +
MethodsValidationTest
PSCAttnJ&FJ&F
Ours-Base✓72.855.4
74.056.8
Ours-Large✓74.859.5
77.959.9
+ +Table 7: Ablation study of $PSCAttn$ module on PanoVOS. + +models trained on planar video datasets, we evaluated 15 off-the-shelf VOS models, including [5, 6, 24, 32, 57-59], and we follow official implementations and training strategies details of them. Table 3 summarizes the domain transfer results of methods that are only trained on synthetic datasets, such as COCO [33] and ECSSD [45], on PanoVOS dataset. Table 4 shows the domain transfer results of state-of-the-art methods, that are trained on synthetic datasets (e.g. COCO [33]) and video datasets (e.g. YouTube-VOS [52]), on our PanoVOS validation and test sets. By analyzing the performance of advanced VOS methods that target conventional planar videos on panoramic videos, we provide the following insights. Firstly, the performance of current sophisticated VOS models will largely degrade when employed to tackle panoramic videos. Secondly, we can observe a trend that training on larger VOS datasets, i.e., YouTube-VOS [52] and BL30K [6] can help mitigate the gap between planar and panoramic videos. + +# 5.3 Results via Visual Foundation Model (RQ2) + +To quantify the segmentation performance of different variations of the foundation model Segment Anything Model [23] on PanoVOS, we evaluate the latest top performing models PerSAM [61] and SAM-PT [43], as shown in Table 5. The performance of these models on our challenging PanoVOS dataset is still unsatisfactory, which leaves space for further exploration. + +
MethodsAttention TypeValidation J&FTest J&F
Ours-BaseCrossAttn72.554.8
PSCAttn74.056.8
Ours-LargeCrossAttn76.859.1
PSCAttn77.959.9
+ +Table 8: Comparison between our PSC attention (PSCAttn) and cross attention (CrossAttn) module on PanoVOS dataset. + +
MethodsPanoVOS Validation
J&FJsFsJuFu
w/o73.768.882.663.180.3
p=376.370.485.065.884.1
p=574.265.379.767.584.2
p=1075.968.181.968.685.2
p=1575.066.781.067.085.4
p=2 (Ours)77.970.585.269.586.4
+ +Table 9: Hyperparameter Analysis of $p$ , which enables the stitching mechanism, in PSCAttn for Ours-Large model. + +# 5.4 Main Results on PanoVOS (RQ3) + +To evaluate the performance of previous methods on the proposed panoramic VOS dataset, we re-trained them on the training set of PanoVOS for the sake of fairness. We report the performance in Table 6, + +which demonstrates that all the previous VOS models perform worse on PanoVOS than on the traditional VOS benchmarks, e.g., YouTube-VOS. Our model substantially outperforms all these methods and achieves state-of-the-art on all evaluation metrics on PanoVOS, which verifies the effectiveness of our model in tackling panoramic videos. Fig.6 visualizes some qualitative comparisons between our model and previous state-of-the-art methods on PanoVOS dataset, which shows that previous benchmarks fail to cope with content discontinuities while our model tackles them well. + +# 5.5 Ablation Study (RQ4) + +In this section, we conduct ablation studies to demonstrate the effectiveness of the main component, i.e., Panoramic Space Consistency Attention (PSCAttn), of our model, with all the experiments performed based on our two model variants, i.e., Ours-Base and Ours-Large. For training, static image datasets are used for pre-training and PanoVOS is used for main training. Table 7 demonstrates the effectiveness of our PSCAttn module. Besides, Fig. 7 illustrates the qualitative comparison between our default model (Ours-Base) and the setting without PSCAttn module. Our model performs better when coping with the pixel discontinuity problem. Moreover, as is shown in Table 8, compared to the conventional cross-attention (CrossAttn) module, PSCAttn also achieves better performance. In Table 9, we analyze the hyperparameter $p$ , which influences the stitching mechanism in PSCAttn, of our model (Ours-Large) on the PanoVOS validation set. Specifically, the highest overall performance ( $\mathcal{J} \& \mathcal{F}$ ) is achieved + +![](images/9aae5cfcf093a8c9b0f1d38aaa5378447d41276ac2ec1c73a8e34f0ccc6510b9.jpg) +Fig. 7: Qualitative ablation study of PSCAttn module. + +![](images/be77e287852906b466b08a095b21268ebaf5e6ed7db608afbae02dfd0b698270.jpg) +Fig. 8: Challenge. Our model fails to segment some objects with strong distortion. + +when setting $p$ as 2. Compared to the setting without using the stitching mechanism $(w / o)$ , our model can achieve much better performance. Specifically, our final model (Ours-Large, $p = 2$ ) achieves more than $4\%$ gain in $\mathcal{I} \& \mathcal{F}$ . + +# 5.6 Limitation and Future Work (RQ5) + +To prompt greater progress of panoramic VOS, we also analyze the limitations of our method. Specifically, our method has no notion of severe distortion challenge since we do not employ a special design (such as deformable convolution [9]) to tackle deformations. That means our model may not segment the objects with large distortions. One such failure case is shown in Fig. 8. Besides, our panoramic dataset can be applied to broader video segmentation and tracking domains, such as referring video object segmentation [28-30, 56], video object tracking [18], video instance segmentation [15], few-shot segmentation [20], and more broader embodied navigation tasks [55]. Also, it would be valuable to investigate the zero-shot segmentation performance of visual foundation models [23] on our challenging panoramic dataset. We hope our work can shed light on efficient adaptation from non-panoramic to panoramic perception. + +# 6 Conclusion + +In this paper, we introduce a high-quality dataset, i.e., PanoVOS, for panoramic video object segmentation. Our PanoVOS dataset provides pixel-level instance annotations with diverse scenarios and significant motions. Based on this dataset, we evaluate 15 off-the-shelf VOS models and carefully analyze their limitations. Then, we further present our model, i.e., PSCFormer, which is equipped with the proposed panoramic space consistency transformer block. Our preliminary experiment demonstrates the effectiveness of our proposed model to enhance the segmentation performance and consistency in panoramic scenes. In conclusion, this provides a new challenge for video understanding, and we hope our PanoVOS dataset can attract more researchers to pay attention to panoramic videos. + +# Acknowledgements + +This work was supported in part by National Natural Science Foundation of China (No.62072112), and Scientific and Technological Innovation Action Plan of Shanghai Science and Technology Committee (No.22511101502, No.22511102202 and No.21DZ2203300). + +# References + +1. Ai, H., Cao, Z., Zhu, J., Bai, H., Chen, Y., Wang, L.: Deep learning for omnidirectional vision: A survey and new perspectives. arXiv preprint arXiv:2205.10468 (2022) +2. Armeni, I., Sax, S., Zamir, A.R., Savarese, S.: Joint 2d-3dsemantic data for indoor scene understanding. arXiv preprint arXiv:1702.01105 (2017) +3. Caelles, S., Maninis, K.K., Pont-Tuset, J., Leal-Taixe, L., Cremers, D., Van Gool, L.: One-shot video object segmentation. In: Proceedings of the IEEE conference on computer vision and pattern recognition. pp. 221-230 (2017) +4. Chen, X., Li, Z., Yuan, Y., Yu, G., Shen, J., Qi, D.: State-aware tracker for real-time video object segmentation. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 9384-9393 (2020) +5. Cheng, H.K., Schwing, A.G.: Xmem: Long-term video object segmentation with an atkinson-shiffrin memory model. In: European Conference on Computer Vision. pp. 640-658. Springer (2022) +6. Cheng, H.K., Tai, Y.W., Tang, C.K.: Rethinking space-time networks with improved memory coverage for efficient video object segmentation. Advances in Neural Information Processing Systems 34, 11781-11794 (2021) +7. Cheng, H.T., Chao, C.H., Dong, J.D., Wen, H.K., Liu, T.L., Sun, M.: Cube padding for weakly-supervised saliency prediction in 360 videos. In: Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition. pp. 1420-1429 (2018) +8. Cheng, J., Tsai, Y.H., Hung, W.C., Wang, S., Yang, M.H.: Fast and accurate online video object segmentation via tracking parts. In: Proceedings of the IEEE conference on computer vision and pattern recognition. pp. 7415-7424 (2018) +9. Dai, J., Qi, H., Xiong, Y., Li, Y., Zhang, G., Hu, H., Wei, Y.: Deformable convolutional networks. In: Proceedings of the IEEE international conference on computer vision. pp. 764-773 (2017) +10. Dang, J., Zheng, H., Xu, X., Guo, Y.: Unified spatio-temporal dynamic routing for efficient video object segmentation. IEEE Transactions on Intelligent Transportation Systems (2023) +11. Dang, J., Zheng, H., Xu, X., Wang, L., Hu, Q., Guo, Y.: Adaptive sparse memory networks for efficient and robust video object segmentation. IEEE Transactions on Neural Networks and Learning Systems (2024) +12. Eger Passos, D., Jung, B.: Measuring the accuracy of inside-out tracking in xr devices using a high-precision robotic arm. In: International Conference on Human-Computer Interaction. pp. 19–26. Springer (2020) +13. Fang, R., Yan, S., Huang, Z., Zhou, J., Tian, H., Dai, J., Li, H.: Instructseq: Unifying vision tasks with instruction-conditioned multi-modal sequence generation. arXiv preprint arXiv:2311.18835 (2023) +14. Guo, P., Hong, L., Zhou, X., Gao, S., Li, W., Li, J., Chen, Z., Li, X., Zhang, W., Zhang, W.: Clickvos: Click video object segmentation. arXiv preprint arXiv:2403.06130 (2024) + +15. Guo, P., Huang, T., He, P., Liu, X., Xiao, T., Chen, Z., Zhang, W.: Openvis: Open-vocabulary video instance segmentation. arXiv preprint arXiv:2305.16835 (2023) +16. Guo, P., Zhang, W., Li, X., Zhang, W.: Adaptive online mutual learning bi-decoders for video object segmentation. IEEE Transactions on Image Processing 31, 7063-7077 (2022) +17. Hendrycks, D., Gimpel, K.: Gaussian error linear units (gelus). arXiv preprint arXiv:1606.08415 (2016) +18. Hong, L., Yan, S., Zhang, R., Li, W., Zhou, X., Guo, P., Jiang, K., Chen, Y., Li, J., Chen, Z., et al.: Onetracker: Unifying visual object tracking with foundation models and efficient tuning. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 19079-19091 (2024) +19. Hu, L., Zhang, P., Zhang, B., Pan, P., Xu, Y., Jin, R.: Learning position and target consistency for memory-based video object segmentation. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 4144-4154 (2021) +20. Iqbal, E., Safarov, S., Bang, S.: Misanet: Multi-similarity and attention guidance for boosting few-shot segmentation. arXiv preprint arXiv:2206.09667 (2022) +21. Jiang, H., Jiang, G., Yu, M., Zhang, Y., Yang, Y., Peng, Z., Chen, F., Zhang, Q.: Cubemap-based perception-driven blind quality assessment for 360-degree images. IEEE Transactions on Image Processing 30, 2364-2377 (2021) +22. Jost, T.A., Nelson, B., Rylander, J.: Quantitative analysis of the oculus rifts in controlled movement. Disability and Rehabilitation: Assistive Technology 16(6), 632-636 (2021) +23. Kirillov, A., Mintun, E., Ravi, N., Mao, H., Rolland, C., Gustafson, L., Xiao, T., Whitehead, S., Berg, A.C., Lo, W.Y., et al.: Segment anything. arXiv preprint arXiv:2304.02643 (2023) +24. Li, M., Hu, L., Xiong, Z., Zhang, B., Pan, P., Liu, D.: Recurrent dynamic embedding for video object segmentation. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 1332-1341 (2022) +25. Li, W., Fan, J., Guo, P., Hong, L., Zhang, W.: Hfvos: History-future integrated dynamic memory for video object segmentation. IEEE Transactions on Circuits and Systems for Video Technology (2024) +26. Li, W., Guo, P., Zhou, X., Hong, L., He, Y., Zheng, X., Zhang, W., Zhang, W.: Onevos: Unifying video object segmentation with all-in-one transformer framework. arXiv preprint arXiv:2403.08682 (2024) +27. Li, X., Cao, H., Zhao, S., Li, J., Zhang, L., Raj, B.: Panoramic video salient object detection with ambisonic audio guidance. arXiv preprint arXiv:2211.14419 (2022) +28. Li, X., Wang, J., Xu, X., Li, X., Raj, B., Lu, Y.: Robust referring video object segmentation with cyclic structural consensus. In: Proceedings of the IEEE/CVF International Conference on Computer Vision. pp. 22236-22245 (2023) +29. Li, X., Wang, J., Xu, X., Peng, X., Singh, R., Lu, Y., Raj, B.: Qdformer: Towards robust audiovisual segmentation in complex environments with quantization-based semantic decomposition. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 3402-3413 (2024) +30. Li, X., Wang, J., Xu, X., Yang, M., Yang, F., Zhao, Y., Singh, R., Raj, B.: Towards noise-tolerant speech-referring video object segmentation: Bridging speech and text. In: Proceedings of the 2023 Conference on Empirical Methods in Natural Language Processing. pp. 2283-2296 (2023) +31. Liang, S., Shen, X., Huang, J., Hua, X.S.: Video object segmentation with dynamic memory networks and adaptive object alignment. In: Proceedings of the IEEE/CVF International Conference on Computer Vision. pp. 8065-8074 (2021) + +32. Liang, Y., Li, X., Jafari, N., Chen, J.: Video object segmentation with adaptive feature bank and uncertain-region refinement. Advances in Neural Information Processing Systems 33, 3430-3441 (2020) +33. Lin, T.Y., Maire, M., Belongie, S., Hays, J., Perona, P., Ramanan, D., Dollar, P., Zitnick, C.L.: Microsoft coco: Common objects in context. In: European conference on computer vision. pp. 740-755. Springer (2014) +34. Liu, Y., Yu, R., Wang, J., Zhao, X., Wang, Y., Tang, Y., Yang, Y.: Global spectral filter memory network for video object segmentation. In: European Conference on Computer Vision. pp. 648-665. Springer (2022) +35. Ma, C., Zhang, J., Yang, K., Roitberg, A., Stiefelhagen, R.: Densepass: Dense panoramic semantic segmentation via unsupervised domain adaptation with attention-augmented context exchange. In: 2021 IEEE International Intelligent Transportation Systems Conference (ITSC). pp. 2766-2772. IEEE (2021) +36. Maninis, K.K., Caelles, S., Chen, Y., Pont-Tuset, J., Leal-Taixe, L., Cremers, D., Van Gool, L.: Video object segmentation without temporal information. IEEE transactions on pattern analysis and machine intelligence 41(6), 1515-1530 (2018) +37. Mao, Y., Wang, N., Zhou, W., Li, H.: Joint inductive and transductive learning for video object segmentation. In: Proceedings of the IEEE/CVF International Conference on Computer Vision. pp. 9670-9679 (2021) +38. Mei, J., Zhu, A.Z., Yan, X., Yan, H., Qiao, S., Chen, L.C., Kretzschmar, H.: Waymo open dataset: Panoramic video panoptic segmentation. In: Computer Vision-ECCV 2022: 17th European Conference, Tel Aviv, Israel, October 23-27, 2022, Proceedings, Part XXIX. pp. 53-72. Springer (2022) +39. Oh, S.W., Lee, J.Y., Sunkavalli, K., Kim, S.J.: Fast video object segmentation by reference-guided mask propagation. In: Proceedings of the IEEE conference on computer vision and pattern recognition. pp. 7376-7385 (2018) +40. Oh, S.W., Lee, J.Y., Xu, N., Kim, S.J.: Video object segmentation using space-time memory networks. In: Proceedings of the IEEE/CVF International Conference on Computer Vision. pp. 9226-9235 (2019) +41. Perazzi, F., Pont-Tuset, J., McWilliams, B., Van Gool, L., Gross, M., Sorkine-Hornung, A.: A benchmark dataset and evaluation methodology for video object segmentation. In: Proceedings of the IEEE conference on computer vision and pattern recognition. pp. 724-732 (2016) +42. Pont-Tuset, J., Perazzi, F., Caelles, S., Arbeláez, P., Sorkine-Hornung, A., Van Gool, L.: The 2017 davis challenge on video object segmentation. arXiv preprint arXiv:1704.00675 (2017) +43. Rajic, F., Ke, L., Tai, Y.W., Tang, C.K., Danelljan, M., Yu, F.: Segment anything meets point tracking. arXiv preprint arXiv:2307.01197 (2023) +44. Seong, H., Hyun, J., Kim, E.: Kernelized memory network for video object segmentation. In: European Conference on Computer Vision. pp. 629-645. Springer (2020) +45. Shi, J., Yan, Q., Xu, L., Jia, J.: Hierarchical image saliency detection on extended cssd. IEEE transactions on pattern analysis and machine intelligence 38(4), 717-729 (2015) +46. Vaswani, A., Shazeer, N., Parmar, N., Uszkoreit, J., Jones, L., Gomez, A.N., Kaiser, L., Polosukhin, I.: Attention is all you need. Advances in neural information processing systems 30 (2017) +47. Voigtlaender, P., Leibe, B.: Online adaptation of convolutional neural networks for video object segmentation. arXiv preprint arXiv:1706.09364 (2017) + +48. Wang, H., Jiang, X., Ren, H., Hu, Y., Bai, S.: Swiftnet: Real-time video object segmentation. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 1296-1305 (2021) +49. Wang, W., Zhou, T., Porikli, F., Crandall, D., Van Gool, L.: A survey on deep learning technique for video segmentation. arXiv preprint arXiv:2107.01153 (2021) +50. Xiao, H., Feng, J., Lin, G., Liu, Y., Zhang, M.: Monet: Deep motion exploitation for video object segmentation. In: Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition. pp. 1140-1148 (2018) +51. Xie, H., Yao, H., Zhou, S., Zhang, S., Sun, W.: Efficient regional memory network for video object segmentation. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 1286-1295 (2021) +52. Xu, N., Yang, L., Fan, Y., Yue, D., Liang, Y., Yang, J., Huang, T.: Youtube-vos: A large-scale video object segmentation benchmark. arXiv preprint arXiv:1809.03327 (2018) +53. Xu, X., Wang, J., Li, X., Lu, Y.: Reliable propagation-correction modulation for video object segmentation. In: Proceedings of the AAAI Conference on Artificial Intelligence. pp. 2946-2954 (2022) +54. Xu, X., Wang, J., Ming, X., Lu, Y.: Towards robust video object segmentation with adaptive object calibration. In: Proceedings of the 30th ACM International Conference on Multimedia. pp. 2709-2718 (2022) +55. Xu, X., Zhang, T., Wang, S., Li, X., Chen, Y., Li, Y., Raj, B., Johnson-Roberson, M., Huang, X.: Customizable perturbation synthesis for robust slam benchmarking. arXiv preprint arXiv:2402.08125 (2024) +56. Yan, S., Zhang, R., Guo, Z., Chen, W., Zhang, W., Li, H., Qiao, Y., Dong, H., He, Z., Gao, P.: Referred by multi-modality: A unified temporal transformer for video object segmentation. In: Proceedings of the AAAI Conference on Artificial Intelligence. vol. 38, pp. 6449-6457 (2024) +57. Yang, Z., Wei, Y., Yang, Y.: Collaborative video object segmentation by foreground-background integration. In: European Conference on Computer Vision. pp. 332-348. Springer (2020) +58. Yang, Z., Wei, Y., Yang, Y.: Associating objects with transformers for video object segmentation. Advances in Neural Information Processing Systems 34, 2491-2502 (2021) +59. Yang, Z., Wei, Y., Yang, Y.: Collaborative video object segmentation by multiscale foreground-background integration. IEEE Transactions on Pattern Analysis and Machine Intelligence (2021) +60. Yuan, M., Richardt, C.: 360 optical flow using tangent images. In: British Machine Vision Conference (BMVC) (2021) +61. Zhang, R., Jiang, Z., Guo, Z., Yan, S., Pan, J., Dong, H., Gao, P., Li, H.: Personalize segment anything model with one shot. arXiv preprint arXiv:2305.03048 (2023) +62. Zhang, Y., Zhang, L., Wang, K., Hamidouche, W., Deforges, O.: Shd360: A benchmark dataset for salient human detection in 360 videos. arXiv preprint arXiv:2105.11578 (2021) +63. Zhang, Y., Wu, Z., Peng, H., Lin, S.: A transductive approach for video object segmentation. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 6949-6958 (2020) +64. Zhang, Z., Xu, Y., Yu, J., Gao, S.: Saliency detection in 360 videos. In: Proceedings of the European conference on computer vision (ECCV). pp. 488-503 (2018) \ No newline at end of file diff --git a/2024/PanoVOS_ Bridging Non-panoramic and Panoramic Views with Transformer for Video Segmentation/images.zip b/2024/PanoVOS_ Bridging Non-panoramic and Panoramic Views with Transformer for Video Segmentation/images.zip new file mode 100644 index 0000000000000000000000000000000000000000..a56e047b1d45da8b3a8327cd2a703dc9c0df8213 --- /dev/null +++ b/2024/PanoVOS_ Bridging Non-panoramic and Panoramic Views with Transformer for Video Segmentation/images.zip @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:d697cc2ac2eed178e399589d17e8023023395999cebfae601a3e9bcf514cb771 +size 666841 diff --git a/2024/PanoVOS_ Bridging Non-panoramic and Panoramic Views with Transformer for Video Segmentation/layout.json b/2024/PanoVOS_ Bridging Non-panoramic and Panoramic Views with Transformer for Video Segmentation/layout.json new file mode 100644 index 0000000000000000000000000000000000000000..a44b262234b2f3fa779bac3d8f16083be09ea646 --- /dev/null +++ b/2024/PanoVOS_ Bridging Non-panoramic and Panoramic Views with Transformer for Video Segmentation/layout.json @@ -0,0 +1,10092 @@ +{ + "pdf_info": [ + { + "para_blocks": [ + { + "bbox": [ + 166, + 111, + 449, + 166 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 166, + 111, + 449, + 166 + ], + "spans": [ + { + "bbox": [ + 166, + 111, + 449, + 166 + ], + "type": "text", + "content": "PanoVOS: Bridging Non-panoramic and Panoramic Views with Transformer for Video Segmentation" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 179, + 186, + 434, + 212 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 179, + 186, + 434, + 212 + ], + "spans": [ + { + "bbox": [ + 179, + 186, + 434, + 212 + ], + "type": "text", + "content": "Shilin Yan" + }, + { + "bbox": [ + 179, + 186, + 434, + 212 + ], + "type": "inline_equation", + "content": "^{1}" + }, + { + "bbox": [ + 179, + 186, + 434, + 212 + ], + "type": "text", + "content": ", Xiaohao Xu" + }, + { + "bbox": [ + 179, + 186, + 434, + 212 + ], + "type": "inline_equation", + "content": "^{2}" + }, + { + "bbox": [ + 179, + 186, + 434, + 212 + ], + "type": "text", + "content": ", Renrui Zhang" + }, + { + "bbox": [ + 179, + 186, + 434, + 212 + ], + "type": "inline_equation", + "content": "^{3}" + }, + { + "bbox": [ + 179, + 186, + 434, + 212 + ], + "type": "text", + "content": ", Lingyi Hong" + }, + { + "bbox": [ + 179, + 186, + 434, + 212 + ], + "type": "inline_equation", + "content": "^{1}" + }, + { + "bbox": [ + 179, + 186, + 434, + 212 + ], + "type": "text", + "content": ", Wenchao Chen" + }, + { + "bbox": [ + 179, + 186, + 434, + 212 + ], + "type": "inline_equation", + "content": "^{1}" + }, + { + "bbox": [ + 179, + 186, + 434, + 212 + ], + "type": "text", + "content": ", Wenqiang Zhang" + }, + { + "bbox": [ + 179, + 186, + 434, + 212 + ], + "type": "inline_equation", + "content": "^{1}" + }, + { + "bbox": [ + 179, + 186, + 434, + 212 + ], + "type": "text", + "content": ", and Wei Zhang" + }, + { + "bbox": [ + 179, + 186, + 434, + 212 + ], + "type": "inline_equation", + "content": "^{1\\star}" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 187, + 219, + 426, + 275 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 187, + 219, + 426, + 275 + ], + "spans": [ + { + "bbox": [ + 187, + 219, + 426, + 275 + ], + "type": "inline_equation", + "content": "^{1}" + }, + { + "bbox": [ + 187, + 219, + 426, + 275 + ], + "type": "text", + "content": " Shanghai Key Lab of Intelligent Information Processing, School of Computer Science, Fudan University " + }, + { + "bbox": [ + 187, + 219, + 426, + 275 + ], + "type": "inline_equation", + "content": "^{2}" + }, + { + "bbox": [ + 187, + 219, + 426, + 275 + ], + "type": "text", + "content": " University of Michigan, Ann Arbor " + }, + { + "bbox": [ + 187, + 219, + 426, + 275 + ], + "type": "inline_equation", + "content": "^{3}" + }, + { + "bbox": [ + 187, + 219, + 426, + 275 + ], + "type": "text", + "content": " MMLab CUHK " + }, + { + "bbox": [ + 187, + 219, + 426, + 275 + ], + "type": "inline_equation", + "content": "^{4}" + }, + { + "bbox": [ + 187, + 219, + 426, + 275 + ], + "type": "text", + "content": " tattoo.ysl@mail.com weizh@fudan.edu.cn" + } + ] + } + ], + "index": 2 + }, + { + "type": "image", + "bbox": [ + 149, + 285, + 478, + 363 + ], + "blocks": [ + { + "bbox": [ + 149, + 285, + 478, + 363 + ], + "lines": [ + { + "bbox": [ + 149, + 285, + 478, + 363 + ], + "spans": [ + { + "bbox": [ + 149, + 285, + 478, + 363 + ], + "type": "image", + "image_path": "b65bc70a30bd22450f37c9d5d3d5f285f3e04f37f97eb5930cc855143aea5ca7.jpg" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 130, + 372, + 483, + 428 + ], + "lines": [ + { + "bbox": [ + 130, + 372, + 483, + 428 + ], + "spans": [ + { + "bbox": [ + 130, + 372, + 483, + 428 + ], + "type": "text", + "content": "Fig.1: Panoramic video object segmentation (PanoVOS). PanoVOS targets tracking and distinguishing the particular instances under content discontinuities (e.g. penguin in the image of " + }, + { + "bbox": [ + 130, + 372, + 483, + 428 + ], + "type": "inline_equation", + "content": "T = 15" + }, + { + "bbox": [ + 130, + 372, + 483, + 428 + ], + "type": "text", + "content": ") and serve distortion (e.g. penguin in the image of " + }, + { + "bbox": [ + 130, + 372, + 483, + 428 + ], + "type": "inline_equation", + "content": "T = 65" + }, + { + "bbox": [ + 130, + 372, + 483, + 428 + ], + "type": "text", + "content": "). We show the sample of (a) frames, (b) segmentation annotations, and (c) area proportion of foreground for the Penguin video in our dataset." + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "image_caption" + } + ], + "index": 3 + }, + { + "bbox": [ + 159, + 448, + 455, + 647 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 159, + 448, + 455, + 647 + ], + "spans": [ + { + "bbox": [ + 159, + 448, + 455, + 647 + ], + "type": "text", + "content": "Abstract. Panoramic videos contain richer spatial information and have attracted tremendous amounts of attention due to their exceptional experience in some fields such as autonomous driving and virtual reality. However, existing datasets for video segmentation only focus on conventional planar images. To address the challenge, in this paper, we present a panoramic video dataset, i.e., PanoVOS. The dataset provides 150 videos with high video resolutions and diverse motions. To quantify the domain gap between 2D planar videos and panoramic videos, we evaluate 15 off-the-shelf video object segmentation (VOS) models on PanoVOS. Through error analysis, we found that all of them fail to tackle pixel-level content discontinues of panoramic videos. Thus, we present a Panoramic Space Consistency Transformer (PSCFormer), which can effectively utilize the semantic boundary information of the previous frame for pixel-level matching with the current frame. Extensive experiments demonstrate that compared with the previous SOTA models, our PSCFormer network exhibits a great advantage in terms of segmentation results under the panoramic setting. Our dataset poses new challenges in panoramic VOS and we hope that" + } + ] + } + ], + "index": 5 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 134, + 654, + 235, + 665 + ], + "type": "page_footnote", + "angle": 0, + "lines": [ + { + "bbox": [ + 134, + 654, + 235, + 665 + ], + "spans": [ + { + "bbox": [ + 134, + 654, + 235, + 665 + ], + "type": "text", + "content": "* Corresponding author." + } + ] + } + ], + "index": 6 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 0 + }, + { + "para_blocks": [ + { + "bbox": [ + 159, + 117, + 455, + 150 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 159, + 117, + 455, + 150 + ], + "spans": [ + { + "bbox": [ + 159, + 117, + 455, + 150 + ], + "type": "text", + "content": "our PanoVOS can advance the development of panoramic segmentation/tracking. The dataset, codes, and pre-train models will be published at https://github.com/shilinyan99/PanoVOS." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 132, + 172, + 230, + 185 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 172, + 230, + 185 + ], + "spans": [ + { + "bbox": [ + 132, + 172, + 230, + 185 + ], + "type": "text", + "content": "1 Introduction" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 130, + 198, + 482, + 319 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 198, + 482, + 319 + ], + "spans": [ + { + "bbox": [ + 130, + 198, + 482, + 319 + ], + "type": "text", + "content": "Semi-supervised video object segmentation (VOS) [52], which targets tracking and distinguishing the particular instances across the entire video sequence based on the first frame masks, plays an essential role in video understanding and editing. Conventionally, the images or videos studied in VOS are 2D planar data with a limited Field of View (FoV), which may lead to some ambiguities, especially when objects are out of view. Meanwhile, with the rapid development of VR/AR collection devices [12,22], panoramic videos with a " + }, + { + "bbox": [ + 130, + 198, + 482, + 319 + ], + "type": "inline_equation", + "content": "360^{\\circ} \\times 180^{\\circ}" + }, + { + "bbox": [ + 130, + 198, + 482, + 319 + ], + "type": "text", + "content": " FoV are able to collect the entire viewing sphere and richer spatial information [1,21, 27, 60]. To the best of our knowledge, we are the first to attempt to tackle the promising but challenging task of panoramic video object segmentation." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 130, + 319, + 482, + 437 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 319, + 482, + 437 + ], + "spans": [ + { + "bbox": [ + 130, + 319, + 482, + 437 + ], + "type": "text", + "content": "To foster the development of panoramic VOS, we propose a new dataset in this work, aiming at panoramic video object segmentation. The dataset contains a wide range of real-world scenarios in which scenes have a large magnitude of motion. The main characteristics of our dataset are three aspects. 1) Panoramic videos bring certain advantages (richer geometric information and wider FoV) in real-world applications as well as challenges (serve distortion and content discontinuities). 2) Compared to all existing VOS datasets, our dataset has longer video clips with an average length of 20 seconds. 3) Nearly half of the video resolutions in our dataset are " + }, + { + "bbox": [ + 130, + 319, + 482, + 437 + ], + "type": "inline_equation", + "content": "4K" + }, + { + "bbox": [ + 130, + 319, + 482, + 437 + ], + "type": "text", + "content": ", which may help facilitate broader video tracking/segmentation research under the high-resolution scenario." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 130, + 438, + 482, + 533 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 438, + 482, + 533 + ], + "spans": [ + { + "bbox": [ + 130, + 438, + 482, + 533 + ], + "type": "text", + "content": "In the proposed dataset, we annotated 150 videos with 19,145 annotated instance masks, including sports (e.g. parkour, skateboard), animals (e.g. elephant, monkey), and common objects (basketball, hot balloon). Since, annotating a pixel-level intensive task is very time-consuming and expensive, we proposed a semi-supervised human-computer joint annotation strategy. Concretely, we first annotated objects at selected keyframes (1 fps) Then we adopted the state-of-the-art video object segmentation model AOT [58] for mask propagation to the rest frames of videos and we manually refine parts of them." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 130, + 533, + 482, + 617 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 533, + 482, + 617 + ], + "spans": [ + { + "bbox": [ + 130, + 533, + 482, + 617 + ], + "type": "text", + "content": "Then, we conducted extensive experiments on PanoVOS to evaluate 15 off-the-shelf video object segmentation models. The results suggest that existing approaches can not handle several domain-unique challenges. The first is content discontinuities, which means the foreground object may be separated in the left and the right boundaries of the planar image, such as the case in the image of " + }, + { + "bbox": [ + 130, + 533, + 482, + 617 + ], + "type": "inline_equation", + "content": "T = 15" + }, + { + "bbox": [ + 130, + 533, + 482, + 617 + ], + "type": "text", + "content": " in Fig. 1. The second is the severe distortions and deformations, such as the case in the image of " + }, + { + "bbox": [ + 130, + 533, + 482, + 617 + ], + "type": "inline_equation", + "content": "T = 65" + }, + { + "bbox": [ + 130, + 533, + 482, + 617 + ], + "type": "text", + "content": " in Fig. 1." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 131, + 617, + 482, + 666 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 131, + 617, + 482, + 666 + ], + "spans": [ + { + "bbox": [ + 131, + 617, + 482, + 666 + ], + "type": "text", + "content": "To tackle these challenges of panoramic video segmentation, we proposed a PSC-Former model which consists of key component Panoramic Space Consistent (PSC) blocks. The PSC block is designed for constructing spatial-temporal class-agnostic correspondence and propagating the segmentation masks. Each PSC" + } + ] + } + ], + "index": 8 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "spans": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "type": "text", + "content": "2" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 166, + 91, + 208, + 101 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 166, + 91, + 208, + 101 + ], + "spans": [ + { + "bbox": [ + 166, + 91, + 208, + 101 + ], + "type": "text", + "content": "Yan et al." + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 1 + }, + { + "para_blocks": [ + { + "bbox": [ + 130, + 116, + 482, + 200 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 116, + 482, + 200 + ], + "spans": [ + { + "bbox": [ + 130, + 116, + 482, + 200 + ], + "type": "text", + "content": "block utilizes a cross-attention for matching with references' embeddings and a PSC-attention for modeling the boundary semantic relationship between the previous frame and the query frame. Hence, the network can effectively alleviate the problem that the left and the right boundaries are actually continuous in panoramic videos. Our method outperforms the SOTA models that are re-trained on PanoVOS train set in segmentation quality under the panoramic setting. Our contributions are three-fold." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 137, + 209, + 481, + 315 + ], + "type": "list", + "angle": 0, + "index": 6, + "blocks": [ + { + "bbox": [ + 138, + 209, + 481, + 243 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 209, + 481, + 243 + ], + "spans": [ + { + "bbox": [ + 138, + 209, + 481, + 243 + ], + "type": "text", + "content": "- We introduce a panoramic video object segmentation dataset (PanoVOS) with 150 videos and 19K annotated instance masks, which fills the gap of long-term instance-level annotated panoramic video segmentation datasets." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 137, + 245, + 480, + 279 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 137, + 245, + 480, + 279 + ], + "spans": [ + { + "bbox": [ + 137, + 245, + 480, + 279 + ], + "type": "text", + "content": "- Extensive experiments are conducted on 15 off-the-shelf VOS benchmarks and our baseline model on PanoVOS, which reveals that current methods could not tackle content discontinuities in panoramic videos well." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 138, + 281, + 480, + 315 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 281, + 480, + 315 + ], + "spans": [ + { + "bbox": [ + 138, + 281, + 480, + 315 + ], + "type": "text", + "content": "- We propose a Panoramic Space Consistency Transformer (PSCFormer) on PanoVOS that successfully resolves the challenges of discontinuity of pixel-level content segmentation." + } + ] + } + ], + "index": 5 + } + ], + "sub_type": "text" + }, + { + "bbox": [ + 132, + 334, + 237, + 346 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 334, + 237, + 346 + ], + "spans": [ + { + "bbox": [ + 132, + 334, + 237, + 346 + ], + "type": "text", + "content": "2 Related Work" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 132, + 359, + 261, + 370 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 359, + 261, + 370 + ], + "spans": [ + { + "bbox": [ + 132, + 359, + 261, + 370 + ], + "type": "text", + "content": "2.1 Panoramic Datasets" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 130, + 379, + 480, + 413 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 379, + 480, + 413 + ], + "spans": [ + { + "bbox": [ + 130, + 379, + 480, + 413 + ], + "type": "text", + "content": "In this paper, panoramic videos refers to complete (360°, full view) panoramic videos, which is different from the definition in [38], which only include wide but partial views of some range-view images collected from multiple cameras." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 130, + 415, + 480, + 484 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 415, + 480, + 484 + ], + "spans": [ + { + "bbox": [ + 130, + 415, + 480, + 484 + ], + "type": "text", + "content": "Image-based panoramic datasets. Existing popular image-level panoramic segmentation datasets are Stanford2D3D [2] and DensePASS [35]. The former one is mainly focused on indoor spaces including a total of 1,413 panoramic images with instance-level annotations in 13 categories. The latter targets driving scenes in cities. DensePASS [35] provides only 100 labeled panoramic images for testing and 2,000 unlabeled images for cross-domain transfer optimization." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 130, + 486, + 480, + 580 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 486, + 480, + 580 + ], + "spans": [ + { + "bbox": [ + 130, + 486, + 480, + 580 + ], + "type": "text", + "content": "Video-based panoramic datasets. Video-based benchmarks mainly include SHD360 [62], SOD360 [64] and Wild360 [7]. All of them are used for panoramic video saliency object detection. Specifically, 1) SHD360 only targets human-centric video scenes with little movement. It provides 6,268 object-level pixel-wise masks and 16,238 instance-level pixel-wise masks. 2) SOD360 focuses on the sports-centric scenario with 41 video clips (12 outdoor and 29 indoor). 3) Wild360 concentrates on natural scenes with 85 videos. Note that SOD360 and Wild360 have no object-level or instance-level annotations." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 130, + 582, + 480, + 664 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 582, + 480, + 664 + ], + "spans": [ + { + "bbox": [ + 130, + 582, + 480, + 664 + ], + "type": "text", + "content": "We make a comparison with the existing video panoramic datasets in Table 1. Specifically, our PanoVOS dataset contains 150 videos mainly from three different domains: person, animal, and common object, which makes the dataset more general for object-agnostic evaluations. Besides, videos in our dataset have a relatively large range of motion, making our PanoVOS dataset suitable for video tracking and segmentation evaluation tasks under panoramic scenes. Moreover, the average duration of each video in our dataset is " + }, + { + "bbox": [ + 130, + 582, + 480, + 664 + ], + "type": "inline_equation", + "content": "20s" + }, + { + "bbox": [ + 130, + 582, + 480, + 664 + ], + "type": "text", + "content": ", which is about 4 times" + } + ] + } + ], + "index": 12 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 207, + 91, + 445, + 102 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 207, + 91, + 445, + 102 + ], + "spans": [ + { + "bbox": [ + 207, + 91, + 445, + 102 + ], + "type": "text", + "content": "PanoVOS: Bridging Non-panoramic and Panoramic Views" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 474, + 92, + 480, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 474, + 92, + 480, + 100 + ], + "spans": [ + { + "bbox": [ + 474, + 92, + 480, + 100 + ], + "type": "text", + "content": "3" + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 2 + }, + { + "para_blocks": [ + { + "type": "table", + "bbox": [ + 178, + 114, + 434, + 194 + ], + "blocks": [ + { + "bbox": [ + 178, + 114, + 434, + 194 + ], + "lines": [ + { + "bbox": [ + 178, + 114, + 434, + 194 + ], + "spans": [ + { + "bbox": [ + 178, + 114, + 434, + 194 + ], + "type": "table", + "html": "
DatasetsMotion#Videos#Frames#Total MasksAverage Duration
SHD360 [62]Small416,26816,2385s
SOD360 [64]Large104N/A0N/A
Wild360 [7]Large85N/A0N/A
PanoVOSLarge15013,99519,14520s
", + "image_path": "01fc6a7856dfb02c3781ad1dafadffce5543acd5b7cde876f9e4182e20e526de.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "table_body" + } + ], + "index": 2 + }, + { + "bbox": [ + 130, + 202, + 482, + 258 + ], + "lines": [ + { + "bbox": [ + 130, + 202, + 482, + 258 + ], + "spans": [ + { + "bbox": [ + 130, + 202, + 482, + 258 + ], + "type": "text", + "content": "Table 1: Comparison of panoramic video datasets. Our PanoVOS is the first long-term panoramic video segmentation dataset with instance-level masks. Compared with existing panoramic video datasets [7,62,64] that are used for saliency detection, our panoramic video dataset for video segmentation, i.e., PanoVOS, includes more diverse and larger motion, making it suitable for dense video tracking evaluation." + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "text" + }, + { + "bbox": [ + 130, + 281, + 482, + 330 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 281, + 482, + 330 + ], + "spans": [ + { + "bbox": [ + 130, + 281, + 482, + 330 + ], + "type": "text", + "content": "longer than SHD360 [62] (5s per video). By the way, the longer video is highlighted in a recent survey [49]. The longer the video, the more likely it is to introduce more panoramic video characteristics such as distortion and discontinuity, which is more challenging and more practical." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 131, + 346, + 348, + 358 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 131, + 346, + 348, + 358 + ], + "spans": [ + { + "bbox": [ + 131, + 346, + 348, + 358 + ], + "type": "text", + "content": "2.2 Video Object Segmentation Datasets" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 130, + 365, + 482, + 449 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 365, + 482, + 449 + ], + "spans": [ + { + "bbox": [ + 130, + 365, + 482, + 449 + ], + "type": "text", + "content": "The establishment of DAVIS [41, 42] and YouTube-VOS [52] datasets pave the way for the boosting development of VOS methods. They are collected by traditional pinhole cameras and the duration of each video clip is very short, only 5s on average. In contrast, the average video length in the proposed PanoVOS dataset is 20s, which is 4 times longer than the existing video datasets. Our dataset includes more challenging scenes (e.g. distortion and discontinuity) that is non-negligible in real-world applications." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 131, + 466, + 348, + 479 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 131, + 466, + 348, + 479 + ], + "spans": [ + { + "bbox": [ + 131, + 466, + 348, + 479 + ], + "type": "text", + "content": "2.3 Video Object Segmentation Methods" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 130, + 486, + 481, + 510 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 486, + 481, + 510 + ], + "spans": [ + { + "bbox": [ + 130, + 486, + 481, + 510 + ], + "type": "text", + "content": "Existing video object segmentation methods can be roughly classified into three subsets: online-learning-based, propagation-based, and matching-based." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 130, + 510, + 481, + 569 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 510, + 481, + 569 + ], + "spans": [ + { + "bbox": [ + 130, + 510, + 481, + 569 + ], + "type": "text", + "content": "Online learning-based. Online learning-based approaches [3,36,50], which either train or fine-tune their networks with the first-frame ground truth at test time and are therefore a great waste of resources. OnAVOS [47] achieves promising results by introducing an online adaptation mechanism, but it still requires online fine-tuning. To a certain extent, it restricts networks' efficiency." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 130, + 570, + 481, + 617 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 570, + 481, + 617 + ], + "spans": [ + { + "bbox": [ + 130, + 570, + 481, + 617 + ], + "type": "text", + "content": "Propagation-based. Propagation-based models [4,8,39] get the target masks in a frame-to-frame prerogation way. Although propagation-based methods improve efficiency, they lack long-term context and therefore are difficult to handle object disappearance and reappearance, severe obscuration, and distortion." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 130, + 617, + 481, + 666 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 617, + 481, + 666 + ], + "spans": [ + { + "bbox": [ + 130, + 617, + 481, + 666 + ], + "type": "text", + "content": "Matching-based. Matching-based methods [6, 10, 11, 14, 16, 25, 26, 31, 37, 40, 53, 54, 63] aim to learn an embedding space of target objects between query and memory. Recently state-of-the-art methods encode many frames into embeddings and store them as a feature memory bank. The most representative is STM [40]," + } + ] + } + ], + "index": 11 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "spans": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "type": "text", + "content": "4" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 166, + 91, + 208, + 101 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 166, + 91, + 208, + 101 + ], + "spans": [ + { + "bbox": [ + 166, + 91, + 208, + 101 + ], + "type": "text", + "content": "Yan et al." + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 3 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 135, + 114, + 202, + 152 + ], + "blocks": [ + { + "bbox": [ + 135, + 114, + 202, + 152 + ], + "lines": [ + { + "bbox": [ + 135, + 114, + 202, + 152 + ], + "spans": [ + { + "bbox": [ + 135, + 114, + 202, + 152 + ], + "type": "image", + "image_path": "9f598b6b7a822982a3a0723c6809c280df59c6b5aa85b81de402b40d4a2c42d2.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_body" + } + ], + "index": 2 + }, + { + "type": "image", + "bbox": [ + 135, + 155, + 202, + 190 + ], + "blocks": [ + { + "bbox": [ + 135, + 155, + 202, + 190 + ], + "lines": [ + { + "bbox": [ + 135, + 155, + 202, + 190 + ], + "spans": [ + { + "bbox": [ + 135, + 155, + 202, + 190 + ], + "type": "image", + "image_path": "ea645df97cdbf5de7919a92a916cc7cf1dc91bfaf9d2b53ce04b066645e454a9.jpg" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 130, + 205, + 480, + 228 + ], + "lines": [ + { + "bbox": [ + 130, + 205, + 480, + 228 + ], + "spans": [ + { + "bbox": [ + 130, + 205, + 480, + 228 + ], + "type": "text", + "content": "Fig. 2: PanoVOS dataset. We select 10 samples from the dataset involving major scenes. For each video, there are high-quality instance-level pixel-wise masks." + } + ] + } + ], + "index": 12, + "angle": 0, + "type": "image_caption" + } + ], + "index": 3 + }, + { + "type": "image", + "bbox": [ + 203, + 114, + 271, + 153 + ], + "blocks": [ + { + "bbox": [ + 203, + 114, + 271, + 153 + ], + "lines": [ + { + "bbox": [ + 203, + 114, + 271, + 153 + ], + "spans": [ + { + "bbox": [ + 203, + 114, + 271, + 153 + ], + "type": "image", + "image_path": "b1eb488d17e5b0bda1eefbc9bf0f4c163341a3680d52acd78d06cc8e85c60134.jpg" + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "image_body" + } + ], + "index": 4 + }, + { + "type": "image", + "bbox": [ + 204, + 155, + 271, + 190 + ], + "blocks": [ + { + "bbox": [ + 204, + 155, + 271, + 190 + ], + "lines": [ + { + "bbox": [ + 204, + 155, + 271, + 190 + ], + "spans": [ + { + "bbox": [ + 204, + 155, + 271, + 190 + ], + "type": "image", + "image_path": "743cea0236f2b529412fa835efc5a45de1e3e9108d9594a2ae9d859b8413b313.jpg" + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "image_body" + } + ], + "index": 5 + }, + { + "type": "image", + "bbox": [ + 272, + 114, + 340, + 152 + ], + "blocks": [ + { + "bbox": [ + 272, + 114, + 340, + 152 + ], + "lines": [ + { + "bbox": [ + 272, + 114, + 340, + 152 + ], + "spans": [ + { + "bbox": [ + 272, + 114, + 340, + 152 + ], + "type": "image", + "image_path": "001a3c1c54c107ee62c30e8b0967b8c9140dca22575ab534f8cdc0e459341807.jpg" + } + ] + } + ], + "index": 6, + "angle": 0, + "type": "image_body" + } + ], + "index": 6 + }, + { + "type": "image", + "bbox": [ + 272, + 155, + 341, + 190 + ], + "blocks": [ + { + "bbox": [ + 272, + 155, + 341, + 190 + ], + "lines": [ + { + "bbox": [ + 272, + 155, + 341, + 190 + ], + "spans": [ + { + "bbox": [ + 272, + 155, + 341, + 190 + ], + "type": "image", + "image_path": "562a3ed9d72a4ee8e0fccf8445f742d4d3c9b391724fcf312de8fcee7a305d15.jpg" + } + ] + } + ], + "index": 7, + "angle": 0, + "type": "image_body" + } + ], + "index": 7 + }, + { + "type": "image", + "bbox": [ + 341, + 114, + 409, + 152 + ], + "blocks": [ + { + "bbox": [ + 341, + 114, + 409, + 152 + ], + "lines": [ + { + "bbox": [ + 341, + 114, + 409, + 152 + ], + "spans": [ + { + "bbox": [ + 341, + 114, + 409, + 152 + ], + "type": "image", + "image_path": "69f21cea20fdc4cc0ee8e1dd102b1de00304555e80ba4f09cf21677b311671f4.jpg" + } + ] + } + ], + "index": 8, + "angle": 0, + "type": "image_body" + } + ], + "index": 8 + }, + { + "type": "image", + "bbox": [ + 342, + 155, + 408, + 190 + ], + "blocks": [ + { + "bbox": [ + 342, + 155, + 408, + 190 + ], + "lines": [ + { + "bbox": [ + 342, + 155, + 408, + 190 + ], + "spans": [ + { + "bbox": [ + 342, + 155, + 408, + 190 + ], + "type": "image", + "image_path": "05e4a16e9f6a76b30ad78fb01c6b15e3aeff2b93ea389258578199259f564b42.jpg" + } + ] + } + ], + "index": 9, + "angle": 0, + "type": "image_body" + } + ], + "index": 9 + }, + { + "type": "image", + "bbox": [ + 410, + 117, + 477, + 152 + ], + "blocks": [ + { + "bbox": [ + 410, + 117, + 477, + 152 + ], + "lines": [ + { + "bbox": [ + 410, + 117, + 477, + 152 + ], + "spans": [ + { + "bbox": [ + 410, + 117, + 477, + 152 + ], + "type": "image", + "image_path": "d9ee8a8bbe976ef7a549c1d34a45b73112bcc182363e6857d87f06a41f61bcf4.jpg" + } + ] + } + ], + "index": 10, + "angle": 0, + "type": "image_body" + } + ], + "index": 10 + }, + { + "type": "image", + "bbox": [ + 410, + 155, + 479, + 190 + ], + "blocks": [ + { + "bbox": [ + 410, + 155, + 479, + 190 + ], + "lines": [ + { + "bbox": [ + 410, + 155, + 479, + 190 + ], + "spans": [ + { + "bbox": [ + 410, + 155, + 479, + 190 + ], + "type": "image", + "image_path": "3d76807a3ce47afba0c1a90987e61102f800ee9f5bb84839536180a29490442c.jpg" + } + ] + } + ], + "index": 11, + "angle": 0, + "type": "image_body" + } + ], + "index": 11 + }, + { + "type": "table", + "bbox": [ + 204, + 239, + 410, + 291 + ], + "blocks": [ + { + "bbox": [ + 204, + 239, + 410, + 291 + ], + "lines": [ + { + "bbox": [ + 204, + 239, + 410, + 291 + ], + "spans": [ + { + "bbox": [ + 204, + 239, + 410, + 291 + ], + "type": "table", + "html": "
SplitsTrainValTest
#Videos8035 (10 unseen)35 (10 unseen)
#Images7,070(50.5%)3,464(24.8%)3,461(24.7%)
#Masks9,585(50.1%)4,957(25.9%)4,603(24.0%)
", + "image_path": "f06b6935522d4104dad9db018b0fa405448958f3ad3c30bbe0ed977120f68e8b.jpg" + } + ] + } + ], + "index": 13, + "angle": 0, + "type": "table_body" + } + ], + "index": 13 + }, + { + "bbox": [ + 214, + 299, + 398, + 310 + ], + "lines": [ + { + "bbox": [ + 214, + 299, + 398, + 310 + ], + "spans": [ + { + "bbox": [ + 214, + 299, + 398, + 310 + ], + "type": "text", + "content": "Table 2: Statistics of PanoVOS dataset" + } + ] + } + ], + "index": 14, + "angle": 0, + "type": "text" + }, + { + "bbox": [ + 130, + 333, + 480, + 394 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 333, + 480, + 394 + ], + "spans": [ + { + "bbox": [ + 130, + 333, + 480, + 394 + ], + "type": "text", + "content": "which has been extended to many works [5, 19, 34, 44, 48, 51]. AOT [58] introduces an identification mechanism by encoding multiple targets into the same embedding space, which can simultaneously segment multiple objects. However, they fail to address the challenges of the tremendous proportion of distortion and discontinuity under panoramic setting." + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 131, + 411, + 261, + 424 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 131, + 411, + 261, + 424 + ], + "spans": [ + { + "bbox": [ + 131, + 411, + 261, + 424 + ], + "type": "text", + "content": "3 PanoVOS Dataset" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 130, + 436, + 480, + 461 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 436, + 480, + 461 + ], + "spans": [ + { + "bbox": [ + 130, + 436, + 480, + 461 + ], + "type": "text", + "content": "We introduced the proposed PanoVOS dataset in three parts, (1) collection process, (2) statistical summary, and (3) annotation pipeline." + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 131, + 478, + 236, + 488 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 131, + 478, + 236, + 488 + ], + "spans": [ + { + "bbox": [ + 131, + 478, + 236, + 488 + ], + "type": "text", + "content": "3.1 Data collection" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 130, + 496, + 480, + 581 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 496, + 480, + 581 + ], + "spans": [ + { + "bbox": [ + 130, + 496, + 480, + 581 + ], + "type": "text", + "content": "We built our PanoVOS dataset with the principle of diversity in mind. Moreover, the objects in the video should have a large amplitude of motion or camera movement. Based on the above viewpoint, we collected videos from the YouTube website for further annotation, respectively. The range of the video length is from 3 to 40 seconds. The average sequence length of each video in the dataset is approximately 20 seconds. We followed the settings of YouTube-VOS [52] to sample the frames at 6 fps." + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 131, + 597, + 250, + 609 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 131, + 597, + 250, + 609 + ], + "spans": [ + { + "bbox": [ + 131, + 597, + 250, + 609 + ], + "type": "text", + "content": "3.2 Dataset Statistics" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 130, + 617, + 480, + 666 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 617, + 480, + 666 + ], + "spans": [ + { + "bbox": [ + 130, + 617, + 480, + 666 + ], + "type": "text", + "content": "PanoVOS contains 150 videos, including 13,995 frames and 19,145 instance annotations from 35 categories. The average length of each video is 20 seconds. We believe that visual categories are representative of common life scenarios, and Fig.2 shows some samples of PanoVOS. To create our PanoVOS, in the" + } + ] + } + ], + "index": 21 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 206, + 91, + 445, + 102 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 206, + 91, + 445, + 102 + ], + "spans": [ + { + "bbox": [ + 206, + 91, + 445, + 102 + ], + "type": "text", + "content": "PanoVOS: Bridging Non-panoramic and Panoramic Views" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 474, + 92, + 480, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 474, + 92, + 480, + 100 + ], + "spans": [ + { + "bbox": [ + 474, + 92, + 480, + 100 + ], + "type": "text", + "content": "5" + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 4 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 161, + 115, + 455, + 231 + ], + "blocks": [ + { + "bbox": [ + 161, + 115, + 455, + 231 + ], + "lines": [ + { + "bbox": [ + 161, + 115, + 455, + 231 + ], + "spans": [ + { + "bbox": [ + 161, + 115, + 455, + 231 + ], + "type": "image", + "image_path": "66c37404571ed1221551acfb4db3f3ee0b2f465bbc13a548d08f8438d84408fb.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 132, + 242, + 482, + 266 + ], + "lines": [ + { + "bbox": [ + 132, + 242, + 482, + 266 + ], + "spans": [ + { + "bbox": [ + 132, + 242, + 482, + 266 + ], + "type": "text", + "content": "Fig. 3: Instance-level distribution of PanoVOS dataset. Our dataset contains three major divisions: person, animals, and common objects with 35 sub-divisions." + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_caption" + } + ], + "index": 2 + }, + { + "bbox": [ + 130, + 289, + 483, + 399 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 289, + 483, + 399 + ], + "spans": [ + { + "bbox": [ + 130, + 289, + 483, + 399 + ], + "type": "text", + "content": "spirit of the video object segmentation task, we carefully selected videos with relatively large motion amplitudes and chose a set of video categories including person (e.g. parkour, dance, BMX, skateboard), animals (e.g. elephant, monkey, giraffe, rhino, birds) and common objects (e.g. basketball, hot balloon) as shown in Fig 3. PanoVOS dataset consists of 150 videos split into training (80), validation (35), and test (35) sets. Table 2 shows detailed division results. Both the validation and test sets have 35 videos (about " + }, + { + "bbox": [ + 130, + 289, + 483, + 399 + ], + "type": "inline_equation", + "content": "23\\%" + }, + { + "bbox": [ + 130, + 289, + 483, + 399 + ], + "type": "text", + "content": " of the frames and the masks). For validation and test sets, we keep some unseen visual categories for generalization ability evaluation." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 132, + 418, + 263, + 431 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 418, + 263, + 431 + ], + "spans": [ + { + "bbox": [ + 132, + 418, + 263, + 431 + ], + "type": "text", + "content": "3.3 Annotation Pipeline" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 130, + 441, + 482, + 561 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 441, + 482, + 561 + ], + "spans": [ + { + "bbox": [ + 130, + 441, + 482, + 561 + ], + "type": "text", + "content": "Annotation is very time-consuming and expensive for a pixel-level panoramic segmentation dataset. To obtain accurate large-scale video panoramic segmentation annotations and make the process more efficient, we propose a semi-automatic human-computer joint annotation strategy, as shown in Fig 4. First, keyframes are selected and manually annotated for each video, which are images with a speed of 1 fps. This is followed by a frame-by-frame propagation from the annotated keyframes to those unlabeled intermediate frames with a sophisticated semi-supervised VOS model. Then, to tackle the distortions and discontinuities in panoramic videos, we need to re-calibrate the resulting annotations via human refinement. More details will unfold below." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 130, + 582, + 482, + 666 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 582, + 482, + 666 + ], + "spans": [ + { + "bbox": [ + 130, + 582, + 482, + 666 + ], + "type": "text", + "content": "Annotation Propagation For the annotation of each video, we first need an expert to browse the current video and note down all objects that have a large amplitude of movement. Then, for each video, the recorded objects in keyframes with a speed of 1 fps are selected for manual annotation. To avoid consistency errors or the problem of objects being labeled as other instances when they disappear and reappear, another expert needs to double-check the annotations of all objects to improve the accuracy of the dataset annotation." + } + ] + } + ], + "index": 7 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "spans": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "type": "text", + "content": "6" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 166, + 91, + 208, + 100 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 166, + 91, + 208, + 100 + ], + "spans": [ + { + "bbox": [ + 166, + 91, + 208, + 100 + ], + "type": "text", + "content": "Yan et al." + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 5 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 133, + 114, + 482, + 203 + ], + "blocks": [ + { + "bbox": [ + 133, + 114, + 482, + 203 + ], + "lines": [ + { + "bbox": [ + 133, + 114, + 482, + 203 + ], + "spans": [ + { + "bbox": [ + 133, + 114, + 482, + 203 + ], + "type": "image", + "image_path": "4dec3a67269029bede7f37571e681b2d68e66e02b8dbbe1df1c13fa14295ec52.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 130, + 209, + 482, + 288 + ], + "lines": [ + { + "bbox": [ + 130, + 209, + 482, + 288 + ], + "spans": [ + { + "bbox": [ + 130, + 209, + 482, + 288 + ], + "type": "text", + "content": "Fig. 4: PanoVOS annotation pipeline. Our annotation pipeline includes two phases. (1) The first phase is called Key Frames Select and Annotate. The annotator browses the video and picks out the object to be annotated. Then, instances are manually annotated at 1 fps and corrected by another annotator. (2) The second phase is called All Frames Propagate and Refine. In this phase, we apply a semi-supervised video object segmentation model to help propagate the annotated masks and the generated instances are refined by annotators." + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_caption" + } + ], + "index": 2 + }, + { + "bbox": [ + 130, + 312, + 482, + 348 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 312, + 482, + 348 + ], + "spans": [ + { + "bbox": [ + 130, + 312, + 482, + 348 + ], + "type": "text", + "content": "We then use the off-the-shelf video labeling method [58] to propagate the instance masks frame by frame from the annotated keyframes to untagged intermediate frames and generate masks at 6 fps." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 130, + 368, + 482, + 477 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 368, + 482, + 477 + ], + "spans": [ + { + "bbox": [ + 130, + 368, + 482, + 477 + ], + "type": "text", + "content": "Annotation Refinement To present a new Panoramic dataset of high quality. After obtaining masks of the first propagation stage, annotators are asked to check the quality of the masks and refine them. The main amendments are in the following two areas. 1). Since our video resolution is generally relatively high, the propagation method will often fail when encountering complex videos with many small objects in a scene. 2) Due to the huge distortions and discontinuities present in the panoramic video, the quality of the masks obtained is relatively poor. Manual correction of the mask is checked by another annotator until the result is satisfactory before proceeding to the next video annotation." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 132, + 496, + 202, + 509 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 496, + 202, + 509 + ], + "spans": [ + { + "bbox": [ + 132, + 496, + 202, + 509 + ], + "type": "text", + "content": "4 Method" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 132, + 523, + 209, + 534 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 523, + 209, + 534 + ], + "spans": [ + { + "bbox": [ + 132, + 523, + 209, + 534 + ], + "type": "text", + "content": "4.1 Overview" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 130, + 545, + 482, + 666 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 545, + 482, + 666 + ], + "spans": [ + { + "bbox": [ + 130, + 545, + 482, + 666 + ], + "type": "text", + "content": "Video object segmentation targets assigning an instance label to every pixel in the given video sequence based on the first frame mask. Recent works [5,6,13,59] have demonstrated that the attention mechanism can significantly help improve the segmentation performance. However, for the challenge of content discontinuation in panoramic videos, only considering the original attention mechanism will not be able to fully utilize the semantic information on the left and right boundaries (pixel contiguity) in the spatial dimension and will lose valuable contextual information when segmenting objects. Therefore, in this work, our mission is to design an effective network architecture, which can help acquire valuable boundary relationships." + } + ] + } + ], + "index": 8 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 207, + 91, + 445, + 102 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 207, + 91, + 445, + 102 + ], + "spans": [ + { + "bbox": [ + 207, + 91, + 445, + 102 + ], + "type": "text", + "content": "PanoVOS: Bridging Non-panoramic and Panoramic Views" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 474, + 92, + 480, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 474, + 92, + 480, + 100 + ], + "spans": [ + { + "bbox": [ + 474, + 92, + 480, + 100 + ], + "type": "text", + "content": "7" + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 6 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 162, + 121, + 452, + 266 + ], + "blocks": [ + { + "bbox": [ + 162, + 121, + 452, + 266 + ], + "lines": [ + { + "bbox": [ + 162, + 121, + 452, + 266 + ], + "spans": [ + { + "bbox": [ + 162, + 121, + 452, + 266 + ], + "type": "image", + "image_path": "20e411bca2ba711233ee86b4077a06ef30b8f2cf913068acf309d6179ff7e29c.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 130, + 274, + 482, + 351 + ], + "lines": [ + { + "bbox": [ + 130, + 274, + 482, + 351 + ], + "spans": [ + { + "bbox": [ + 130, + 274, + 482, + 351 + ], + "type": "text", + "content": "Fig. 5: (a) PSCFormer overview. Given the query frame " + }, + { + "bbox": [ + 130, + 274, + 482, + 351 + ], + "type": "inline_equation", + "content": "\\mathbf{x}_t" + }, + { + "bbox": [ + 130, + 274, + 482, + 351 + ], + "type": "text", + "content": " and reference frames " + }, + { + "bbox": [ + 130, + 274, + 482, + 351 + ], + "type": "inline_equation", + "content": "\\{\\mathbf{x}_i | i \\in \\mathcal{R}\\}" + }, + { + "bbox": [ + 130, + 274, + 482, + 351 + ], + "type": "text", + "content": ", the goal of VOS is to delineate objects from the background by generating mask " + }, + { + "bbox": [ + 130, + 274, + 482, + 351 + ], + "type": "inline_equation", + "content": "\\mathbf{y}_t" + }, + { + "bbox": [ + 130, + 274, + 482, + 351 + ], + "type": "text", + "content": " for query frame " + }, + { + "bbox": [ + 130, + 274, + 482, + 351 + ], + "type": "inline_equation", + "content": "\\mathbf{x}_t" + }, + { + "bbox": [ + 130, + 274, + 482, + 351 + ], + "type": "text", + "content": ". References and the query frame are encoded by the memory encoder and query encoder, respectively. Multiple stacking panoramic space consistency (PSC) blocks are used to leverage the correspondence in the panoramic space between references and the query frame. A decoder is used for generating the prediction of the query frame. (b) Panoramic space consistency block architecture details." + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_caption" + } + ], + "index": 2 + }, + { + "bbox": [ + 130, + 374, + 482, + 483 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 374, + 482, + 483 + ], + "spans": [ + { + "bbox": [ + 130, + 374, + 482, + 483 + ], + "type": "text", + "content": "Fig. 5(a) illustrates the overall architecture of the proposed network. Given the query frame " + }, + { + "bbox": [ + 130, + 374, + 482, + 483 + ], + "type": "inline_equation", + "content": "\\mathbf{x}_t" + }, + { + "bbox": [ + 130, + 374, + 482, + 483 + ], + "type": "text", + "content": " and references " + }, + { + "bbox": [ + 130, + 374, + 482, + 483 + ], + "type": "inline_equation", + "content": "\\{\\mathbf{x}_i | i \\in \\mathcal{R}\\}" + }, + { + "bbox": [ + 130, + 374, + 482, + 483 + ], + "type": "text", + "content": ", the goal of VOS is to delineate objects from the background by generating mask " + }, + { + "bbox": [ + 130, + 374, + 482, + 483 + ], + "type": "inline_equation", + "content": "\\mathbf{y}_t" + }, + { + "bbox": [ + 130, + 374, + 482, + 483 + ], + "type": "text", + "content": " for query frame " + }, + { + "bbox": [ + 130, + 374, + 482, + 483 + ], + "type": "inline_equation", + "content": "\\mathbf{x}_t" + }, + { + "bbox": [ + 130, + 374, + 482, + 483 + ], + "type": "text", + "content": ". Following [57], our basic setting uses the first and previous frame as references " + }, + { + "bbox": [ + 130, + 374, + 482, + 483 + ], + "type": "inline_equation", + "content": "\\mathcal{R} = \\{1, t - 1\\}" + }, + { + "bbox": [ + 130, + 374, + 482, + 483 + ], + "type": "text", + "content": ". The memory encoder and query encoder are responsible for extracting frame-level features. After this, the panoramic space consistency block takes them as input and aggregates the spatial-temporal information between the reference frames and the query frame at the pixel level. Finally, the decoder uses the output of the sequence stacking PSC blocks to predict the mask of the object." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 131, + 501, + 342, + 514 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 131, + 501, + 342, + 514 + ], + "spans": [ + { + "bbox": [ + 131, + 501, + 342, + 514 + ], + "type": "text", + "content": "4.2 Panoramic Space Consistency Block" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 130, + 521, + 482, + 616 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 521, + 482, + 616 + ], + "spans": [ + { + "bbox": [ + 130, + 521, + 482, + 616 + ], + "type": "text", + "content": "Fig. 5(b) shows the structure of a PSC block. Motivated by the common transformer blocks [46], PSC firstly contains a self-attention layer, which is used to aggregate the target objects' correlation information within the query frame. Then, the middle module is composed of cross-attention and PSC-attention, in which cross-attention is responsible for learning the target objects' information from references " + }, + { + "bbox": [ + 130, + 521, + 482, + 616 + ], + "type": "inline_equation", + "content": "\\mathcal{R}" + }, + { + "bbox": [ + 130, + 521, + 482, + 616 + ], + "type": "text", + "content": " and the PSC-attention targets on exploring the boundary relationship between the query frame and previous frame. Finally, PSC employs a two-layer feed-forward MLP with GELU [17] non-linearity activation function." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 130, + 617, + 482, + 665 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 617, + 482, + 665 + ], + "spans": [ + { + "bbox": [ + 130, + 617, + 482, + 665 + ], + "type": "text", + "content": "Panoramic Space Consistency Attention (PSC-Attn). PSC-Attn is employed to model the spatial-temporal relationship between the query frame and reference frames considering the continuity of pixels of images in the panoramic space. How to establish a connection between the left and right boundaries be" + } + ] + } + ], + "index": 7 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "spans": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "type": "text", + "content": "8" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 166, + 91, + 208, + 100 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 166, + 91, + 208, + 100 + ], + "spans": [ + { + "bbox": [ + 166, + 91, + 208, + 100 + ], + "type": "text", + "content": "Yan et al." + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 7 + }, + { + "para_blocks": [ + { + "bbox": [ + 130, + 116, + 482, + 260 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 116, + 482, + 260 + ], + "spans": [ + { + "bbox": [ + 130, + 116, + 482, + 260 + ], + "type": "text", + "content": "come especially important? The most intuitive solution would be to directly splice in length, but this would lead to a huge amount of computation. Therefore, we take the approach of moving a portion of the region in the length dimension from the right boundary to the leftmost boundary for stitching. Consequently, we only focus on the left and right boundaries between the query frame and the reference frame. Thus, unlike the original attention, where each query token is counted for attention along with all key tokens in the reference frame, our PSC attention takes care of the key tokens in a fixed window size. In particular, we define the reference frame feature embedding " + }, + { + "bbox": [ + 130, + 116, + 482, + 260 + ], + "type": "inline_equation", + "content": "\\mathbf{f}(\\mathbf{x})\\in \\mathbb{R}^{H\\times W\\times C}" + }, + { + "bbox": [ + 130, + 116, + 482, + 260 + ], + "type": "text", + "content": ", which is extracted from the query encoder. " + }, + { + "bbox": [ + 130, + 116, + 482, + 260 + ], + "type": "inline_equation", + "content": "H" + }, + { + "bbox": [ + 130, + 116, + 482, + 260 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 130, + 116, + 482, + 260 + ], + "type": "inline_equation", + "content": "W" + }, + { + "bbox": [ + 130, + 116, + 482, + 260 + ], + "type": "text", + "content": ", and " + }, + { + "bbox": [ + 130, + 116, + 482, + 260 + ], + "type": "inline_equation", + "content": "C" + }, + { + "bbox": [ + 130, + 116, + 482, + 260 + ], + "type": "text", + "content": " represent the height, width, and channel dimensions, respectively. According to the solutions mentioned above, the new feature embedding " + }, + { + "bbox": [ + 130, + 116, + 482, + 260 + ], + "type": "inline_equation", + "content": "\\mathbf{f}(\\mathbf{x})'" + }, + { + "bbox": [ + 130, + 116, + 482, + 260 + ], + "type": "text", + "content": " is calculated as follows:" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 243, + 272, + 373, + 283 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 243, + 272, + 373, + 283 + ], + "spans": [ + { + "bbox": [ + 243, + 272, + 373, + 283 + ], + "type": "interline_equation", + "content": "\\mathbf {f} (\\mathbf {x}) ^ {\\prime} [ 0: W / p ] = \\mathbf {f} (\\mathbf {x}) [ W / p: W ]", + "image_path": "41a04b5eddb409d0b37d16a443e63a32cf761582415befac9a1fac6508363a89.jpg" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 238, + 285, + 480, + 298 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 238, + 285, + 480, + 298 + ], + "spans": [ + { + "bbox": [ + 238, + 285, + 480, + 298 + ], + "type": "interline_equation", + "content": "\\mathbf {f} (\\mathbf {x}) ^ {\\prime} [ W / p: W ] = \\mathbf {f} (\\mathbf {x}) [ 0: W / p ] \\tag {1}", + "image_path": "6d8106c7b38762c814bdff4a0c83eda389d70cdbb678d021d31b6b8e9ada66b8.jpg" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 208, + 300, + 406, + 312 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 208, + 300, + 406, + 312 + ], + "spans": [ + { + "bbox": [ + 208, + 300, + 406, + 312 + ], + "type": "interline_equation", + "content": "\\mathbf {f} (\\mathbf {x}) ^ {\\prime} [ W / p: W - W / p ] = \\mathbf {f} (\\mathbf {x}) [ W / p: W - W / p ],", + "image_path": "707b749173b5506f794d2aa68493240ae8538921e963f5a361ab3ebbb6192799.jpg" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 131, + 323, + 482, + 372 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 131, + 323, + 482, + 372 + ], + "spans": [ + { + "bbox": [ + 131, + 323, + 482, + 372 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 131, + 323, + 482, + 372 + ], + "type": "inline_equation", + "content": "p \\in \\mathbb{Z}^+" + }, + { + "bbox": [ + 131, + 323, + 482, + 372 + ], + "type": "text", + "content": ". We define query embedding " + }, + { + "bbox": [ + 131, + 323, + 482, + 372 + ], + "type": "inline_equation", + "content": "Q \\in \\mathbb{R}^{HW \\times C}" + }, + { + "bbox": [ + 131, + 323, + 482, + 372 + ], + "type": "text", + "content": ", key embedding " + }, + { + "bbox": [ + 131, + 323, + 482, + 372 + ], + "type": "inline_equation", + "content": "K \\in \\mathbb{R}^{HW \\times C}" + }, + { + "bbox": [ + 131, + 323, + 482, + 372 + ], + "type": "text", + "content": ", value embedding " + }, + { + "bbox": [ + 131, + 323, + 482, + 372 + ], + "type": "inline_equation", + "content": "V \\in \\mathbb{R}^{HW \\times C}" + }, + { + "bbox": [ + 131, + 323, + 482, + 372 + ], + "type": "text", + "content": ", where " + }, + { + "bbox": [ + 131, + 323, + 482, + 372 + ], + "type": "inline_equation", + "content": "Q" + }, + { + "bbox": [ + 131, + 323, + 482, + 372 + ], + "type": "text", + "content": " is from the query frame feature embedding, " + }, + { + "bbox": [ + 131, + 323, + 482, + 372 + ], + "type": "inline_equation", + "content": "K" + }, + { + "bbox": [ + 131, + 323, + 482, + 372 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 131, + 323, + 482, + 372 + ], + "type": "inline_equation", + "content": "V" + }, + { + "bbox": [ + 131, + 323, + 482, + 372 + ], + "type": "text", + "content": " are from " + }, + { + "bbox": [ + 131, + 323, + 482, + 372 + ], + "type": "inline_equation", + "content": "\\mathbf{f}(\\mathbf{x})'" + }, + { + "bbox": [ + 131, + 323, + 482, + 372 + ], + "type": "text", + "content": " by performing dimensional transformations. Mathematically, we define the PSC attention as follows," + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 211, + 381, + 481, + 409 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 211, + 381, + 481, + 409 + ], + "spans": [ + { + "bbox": [ + 211, + 381, + 481, + 409 + ], + "type": "interline_equation", + "content": "\\operatorname {P S C A t t n} (Q, K, V) = \\operatorname {s o f t m a x} \\left(\\frac {Q K ^ {T} \\mathbf {R}}{\\sqrt {C}}\\right) V, \\tag {2}", + "image_path": "01d5e77b369179d12882cd40be672cc78b9425967a5350b24a7ead48801e1dd3.jpg" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 131, + 419, + 482, + 448 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 131, + 419, + 482, + 448 + ], + "spans": [ + { + "bbox": [ + 131, + 419, + 482, + 448 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 131, + 419, + 482, + 448 + ], + "type": "inline_equation", + "content": "\\mathbf{R} \\in [0,1]^{HW \\times HW}" + }, + { + "bbox": [ + 131, + 419, + 482, + 448 + ], + "type": "text", + "content": " means a window that represents the attention range of each query token. For query " + }, + { + "bbox": [ + 131, + 419, + 482, + 448 + ], + "type": "inline_equation", + "content": "Q_{(x,y)}" + }, + { + "bbox": [ + 131, + 419, + 482, + 448 + ], + "type": "text", + "content": " at " + }, + { + "bbox": [ + 131, + 419, + 482, + 448 + ], + "type": "inline_equation", + "content": "(x,y)" + }, + { + "bbox": [ + 131, + 419, + 482, + 448 + ], + "type": "text", + "content": " position, we define the " + }, + { + "bbox": [ + 131, + 419, + 482, + 448 + ], + "type": "inline_equation", + "content": "\\mathbf{R}_{(x,y)}" + }, + { + "bbox": [ + 131, + 419, + 482, + 448 + ], + "type": "text", + "content": " as:" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 190, + 458, + 481, + 489 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 190, + 458, + 481, + 489 + ], + "spans": [ + { + "bbox": [ + 190, + 458, + 481, + 489 + ], + "type": "interline_equation", + "content": "\\mathbf {R} _ {x, y} (i, j) = \\left\\{ \\begin{array}{l l} 1 & \\text {i f} (x - i) ^ {2} \\leqslant s ^ {2} \\text {a n d} (y - j) ^ {2} \\leqslant s ^ {2} \\\\ 0 & \\text {o t h e r w i s e} \\end{array} , \\right. \\tag {3}", + "image_path": "7b512ec400194a92c6f9705741554f68e57f43f0f2a7923971ed9051d8850785.jpg" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 130, + 500, + 482, + 548 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 500, + 482, + 548 + ], + "spans": [ + { + "bbox": [ + 130, + 500, + 482, + 548 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 130, + 500, + 482, + 548 + ], + "type": "inline_equation", + "content": "(i,j)" + }, + { + "bbox": [ + 130, + 500, + 482, + 548 + ], + "type": "text", + "content": " is the position for each key token, " + }, + { + "bbox": [ + 130, + 500, + 482, + 548 + ], + "type": "inline_equation", + "content": "s" + }, + { + "bbox": [ + 130, + 500, + 482, + 548 + ], + "type": "text", + "content": " is the window size. For each query token, it calculates the attention with another key token only if they are spatially limited to a " + }, + { + "bbox": [ + 130, + 500, + 482, + 548 + ], + "type": "inline_equation", + "content": "(2\\times s + 1)" + }, + { + "bbox": [ + 130, + 500, + 482, + 548 + ], + "type": "text", + "content": " size window, which significantly reduces the time complexity from " + }, + { + "bbox": [ + 130, + 500, + 482, + 548 + ], + "type": "inline_equation", + "content": "(h\\times w)^{2}" + }, + { + "bbox": [ + 130, + 500, + 482, + 548 + ], + "type": "text", + "content": " to " + }, + { + "bbox": [ + 130, + 500, + 482, + 548 + ], + "type": "inline_equation", + "content": "(2\\times s + 1)^{2}" + }, + { + "bbox": [ + 130, + 500, + 482, + 548 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 130, + 548, + 482, + 572 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 548, + 482, + 572 + ], + "spans": [ + { + "bbox": [ + 130, + 548, + 482, + 572 + ], + "type": "text", + "content": "Following [46], we implement the representational form of our PSCAttn module with multi-headed attention, defined mathematically as follows," + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 192, + 583, + 410, + 595 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 192, + 583, + 410, + 595 + ], + "spans": [ + { + "bbox": [ + 192, + 583, + 410, + 595 + ], + "type": "interline_equation", + "content": "\\operatorname {M u l t i H e a d} (Q, K, V) = \\operatorname {C o n c a t} \\left(\\operatorname {h e a d} _ {1}, \\dots , \\operatorname {h e a d} _ {h}\\right) W ^ {O}", + "image_path": "2d655d605f17a4c28e48275cbb7d5b23de7e98fbdbb38846d82470b40dab19fc.jpg" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 251, + 595, + 480, + 617 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 251, + 595, + 480, + 617 + ], + "spans": [ + { + "bbox": [ + 251, + 595, + 480, + 617 + ], + "type": "interline_equation", + "content": "\\operatorname {h e a d} _ {i} = \\operatorname {P S C A t t n} \\left(Q W _ {i} ^ {Q}, K W _ {i} ^ {K}, V W _ {i} ^ {V}\\right), \\tag {4}", + "image_path": "08b5f0e7368c9504d1b081fca341fb595011501cdc6e200a3b43e90e6dec8f08.jpg" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 130, + 628, + 482, + 666 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 628, + 482, + 666 + ], + "spans": [ + { + "bbox": [ + 130, + 628, + 482, + 666 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 130, + 628, + 482, + 666 + ], + "type": "inline_equation", + "content": "W_{i}^{Q}\\in \\mathbb{R}^{C\\times d_{model}}" + }, + { + "bbox": [ + 130, + 628, + 482, + 666 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 130, + 628, + 482, + 666 + ], + "type": "inline_equation", + "content": "W_{i}^{K}\\in \\mathbb{R}^{C\\times d_{model}}" + }, + { + "bbox": [ + 130, + 628, + 482, + 666 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 130, + 628, + 482, + 666 + ], + "type": "inline_equation", + "content": "W_{i}^{V}\\in \\mathbb{R}^{C\\times d_{model}}" + }, + { + "bbox": [ + 130, + 628, + 482, + 666 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 130, + 628, + 482, + 666 + ], + "type": "inline_equation", + "content": "W_{i}^{O}\\in \\mathbb{R}^{C\\times C}" + }, + { + "bbox": [ + 130, + 628, + 482, + 666 + ], + "type": "text", + "content": " are the linear projections. As [46], we set the number of heads to " + }, + { + "bbox": [ + 130, + 628, + 482, + 666 + ], + "type": "inline_equation", + "content": "(h = C / d_{model})" + }, + { + "bbox": [ + 130, + 628, + 482, + 666 + ], + "type": "text", + "content": " 8, where " + }, + { + "bbox": [ + 130, + 628, + 482, + 666 + ], + "type": "inline_equation", + "content": "d_{model}" + }, + { + "bbox": [ + 130, + 628, + 482, + 666 + ], + "type": "text", + "content": " is the projection dimension of each head." + } + ] + } + ], + "index": 14 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 206, + 91, + 445, + 102 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 206, + 91, + 445, + 102 + ], + "spans": [ + { + "bbox": [ + 206, + 91, + 445, + 102 + ], + "type": "text", + "content": "PanoVOS: Bridging Non-panoramic and Panoramic Views" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 474, + 91, + 480, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 474, + 91, + 480, + 100 + ], + "spans": [ + { + "bbox": [ + 474, + 91, + 480, + 100 + ], + "type": "text", + "content": "9" + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 8 + }, + { + "para_blocks": [ + { + "type": "table", + "bbox": [ + 134, + 114, + 479, + 209 + ], + "blocks": [ + { + "bbox": [ + 134, + 114, + 479, + 209 + ], + "lines": [ + { + "bbox": [ + 134, + 114, + 479, + 209 + ], + "spans": [ + { + "bbox": [ + 134, + 114, + 479, + 209 + ], + "type": "table", + "html": "
Methods|MFYouTube-VOSPanoVOS ValidationPanoVOS Test
J&FJ&FJsFsJuFuJ&FJsFsJuFu
AOTT [58]73.753.8↓19.944.458.346.965.743.7↓30.036.349.839.649.2
AOTS [58]74.655.8↓18.849.162.246.465.544.7↓29.932.443.346.556.8
AOTB [58]75.253.7↓21.546.258.146.364.139.5↓35.734.444.435.044.4
AFB-URR [32]✓65.240.1↓25.131.141.535.851.830.4↓34.823.132.728.836.9
STCN [6]✓76.149.9↓26.242.753.445.158.448.0↓28.139.350.246.755.7
XMem [5]✓77.048.6↓28.440.750.144.858.640.2↓36.835.344.936.444.0
AOTL [58]✓74.749.2↓25.543.357.338.957.138.2↓36.532.343.732.744.1
R50 AOTL [58]✓76.550.1↓26.444.558.640.357.241.4↓35.133.745.038.348.4
SwinB_AOTL [58]✓74.444.8↓29.639.152.234.953.036.2↓38.231.142.031.040.6
RDE* [24]✓61.743.1↓18.636.048.435.252.741.3↓20.430.944.641.448.5
STCN* [6]✓56.343.2↓13.141.653.733.244.538.0↓18.332.843.235.540.4
XMem* [5]✓65.855.9↓9.952.264.047.260.049.6↓16.239.252.646.859.9
", + "image_path": "e85f1e7940eed0104d2a9cb85eb3fcab1bf3268b330c33eec9214e3eec0380c3.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "table_body" + } + ], + "index": 2 + }, + { + "type": "table", + "bbox": [ + 135, + 290, + 479, + 404 + ], + "blocks": [ + { + "bbox": [ + 130, + 217, + 482, + 274 + ], + "lines": [ + { + "bbox": [ + 130, + 217, + 482, + 274 + ], + "spans": [ + { + "bbox": [ + 130, + 217, + 482, + 274 + ], + "type": "text", + "content": "Table 3: Domain transfer result of (static image datasets) " + }, + { + "bbox": [ + 130, + 217, + 482, + 274 + ], + "type": "inline_equation", + "content": "\\rightarrow" + }, + { + "bbox": [ + 130, + 217, + 482, + 274 + ], + "type": "text", + "content": " (PanoVOS Validation & Test). Subscript " + }, + { + "bbox": [ + 130, + 217, + 482, + 274 + ], + "type": "inline_equation", + "content": "s" + }, + { + "bbox": [ + 130, + 217, + 482, + 274 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 130, + 217, + 482, + 274 + ], + "type": "inline_equation", + "content": "u" + }, + { + "bbox": [ + 130, + 217, + 482, + 274 + ], + "type": "text", + "content": " denote scores in seen and unseen categories. " + }, + { + "bbox": [ + 130, + 217, + 482, + 274 + ], + "type": "inline_equation", + "content": "MF" + }, + { + "bbox": [ + 130, + 217, + 482, + 274 + ], + "type": "text", + "content": " denotes multiple historical frames as reference. " + }, + { + "bbox": [ + 130, + 217, + 482, + 274 + ], + "type": "inline_equation", + "content": "\\downarrow" + }, + { + "bbox": [ + 130, + 217, + 482, + 274 + ], + "type": "text", + "content": " represents the performance of the declining values compared to the YouTube-VOS dataset [52]. * denotes a large-scale external dataset BL30K [6] dataset is used during training." + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 135, + 290, + 479, + 404 + ], + "lines": [ + { + "bbox": [ + 135, + 290, + 479, + 404 + ], + "spans": [ + { + "bbox": [ + 135, + 290, + 479, + 404 + ], + "type": "table", + "html": "
Methods|MF|YouTube-VOS |J&F|PanoVOS ValidationPanoVOS Test
J&FJsFsJuFuJ&FJsFsJuFu
CFBI†[59]81.460.9↓20.553.065.256.369.049.0↓32.449.447.646.252.6
CFBL+†[59]82.857.6↓25.252.167.048.163.453.7↓29.151.659.346.657.5
AOTT[58]80.261.5↓18.755.667.754.668.252.6↓27.644.855.351.558.8
AOTS[58]82.666.7↓15.958.070.562.076.457.3↓25.350.261.054.663.5
AOTB[58]83.570.5↓13.059.271.768.582.760.8↓22.753.064.457.868.2
AFB-URR [32]✓79.655.1↓24.544.755.653.466.752.4↓27.243.654.252.059.9
RDE[24]✓81.954.7↓27.250.363.944.660.155.4↓26.545.559.251.065.9
STCN[6]✓83.061.8↓21.250.363.561.372.153.4↓29.646.258.949.059.9
XMem[5]✓85.766.1↓19.656.668.762.077.262.5↓23.253.165.461.170.4
AOTL[58]✓83.871.9↓11.962.175.367.482.862.1↓21.757.169.056.266.1
R50_AOTL[58]✓84.169.2↓14.956.769.467.583.161.4↓22.757.569.053.365.7
SwinB_AOTL[58]✓84.567.5↓17.060.273.660.376.060.9↓23.653.963.758.767.4
RDE* [24]✓83.360.9↓22.451.464.756.071.655.6↓27.748.160.852.661.0
STCN* [6]✓84.361.7↓22.649.961.859.775.555.8↓28.548.259.852.762.5
XMem* [5]✓86.163.4↓22.753.564.461.574.161.0↓25.153.565.157.568.0
", + "image_path": "8183adfe9ce53152edf57e6b157b1b396816f23d28c869b45bca93113b29af07.jpg" + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "table_body" + } + ], + "index": 4 + }, + { + "bbox": [ + 130, + 413, + 482, + 479 + ], + "lines": [ + { + "bbox": [ + 130, + 413, + 482, + 479 + ], + "spans": [ + { + "bbox": [ + 130, + 413, + 482, + 479 + ], + "type": "text", + "content": "Table 4: Domain transfer result of (static image datasets & YouTubeVOS) " + }, + { + "bbox": [ + 130, + 413, + 482, + 479 + ], + "type": "inline_equation", + "content": "\\rightarrow" + }, + { + "bbox": [ + 130, + 413, + 482, + 479 + ], + "type": "text", + "content": " (PanoVOS Validation & Test). Subscript " + }, + { + "bbox": [ + 130, + 413, + 482, + 479 + ], + "type": "inline_equation", + "content": "s" + }, + { + "bbox": [ + 130, + 413, + 482, + 479 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 130, + 413, + 482, + 479 + ], + "type": "inline_equation", + "content": "u" + }, + { + "bbox": [ + 130, + 413, + 482, + 479 + ], + "type": "text", + "content": " denote scores in seen and unseen categories. " + }, + { + "bbox": [ + 130, + 413, + 482, + 479 + ], + "type": "inline_equation", + "content": "MF" + }, + { + "bbox": [ + 130, + 413, + 482, + 479 + ], + "type": "text", + "content": " denotes multiple historical frames as reference. " + }, + { + "bbox": [ + 130, + 413, + 482, + 479 + ], + "type": "inline_equation", + "content": "\\downarrow" + }, + { + "bbox": [ + 130, + 413, + 482, + 479 + ], + "type": "text", + "content": " represents the performance of the declining values compared to the YouTube-VOS dataset [52]. * denotes a large-scale external dataset BL30K [6] dataset is used during training. " + }, + { + "bbox": [ + 130, + 413, + 482, + 479 + ], + "type": "inline_equation", + "content": "\\dagger" + }, + { + "bbox": [ + 130, + 413, + 482, + 479 + ], + "type": "text", + "content": " denotes no synthetic data is used during the training stage." + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "text" + }, + { + "bbox": [ + 132, + 508, + 224, + 521 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 508, + 224, + 521 + ], + "spans": [ + { + "bbox": [ + 132, + 508, + 224, + 521 + ], + "type": "text", + "content": "5 Experiment" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 130, + 546, + 481, + 570 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 546, + 481, + 570 + ], + "spans": [ + { + "bbox": [ + 130, + 546, + 481, + 570 + ], + "type": "text", + "content": "In this section, we design a series of experiments to answer the following research questions related to how to tackle video object segmentation in panoramic scenes:" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 130, + 570, + 481, + 665 + ], + "type": "list", + "angle": 0, + "index": 13, + "blocks": [ + { + "bbox": [ + 130, + 570, + 481, + 593 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 570, + 481, + 593 + ], + "spans": [ + { + "bbox": [ + 130, + 570, + 481, + 593 + ], + "type": "text", + "content": "RQ1: How well are current VOS methods trained on non-panoramic videos adapted to the panoramic world?" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 130, + 594, + 481, + 616 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 594, + 481, + 616 + ], + "spans": [ + { + "bbox": [ + 130, + 594, + 481, + 616 + ], + "type": "text", + "content": "RQ2: How well do variations of the foundation model Segment Anything Model [23] adapt to the panoramic world?" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 130, + 617, + 481, + 640 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 617, + 481, + 640 + ], + "spans": [ + { + "bbox": [ + 130, + 617, + 481, + 640 + ], + "type": "text", + "content": "RQ3: Can the proposed PanoVOS datasets bring about a consistent performance gain to VOS methods?" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 130, + 641, + 455, + 653 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 641, + 455, + 653 + ], + "spans": [ + { + "bbox": [ + 130, + 641, + 455, + 653 + ], + "type": "text", + "content": "RQ4: How well does Panoramic Space Consistency Attention contribute?" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 130, + 653, + 448, + 665 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 653, + 448, + 665 + ], + "spans": [ + { + "bbox": [ + 130, + 653, + 448, + 665 + ], + "type": "text", + "content": "RQ5: What are the remained problems for panoramic-related research?" + } + ] + } + ], + "index": 12 + } + ], + "sub_type": "text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "spans": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "type": "text", + "content": "10" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 166, + 91, + 207, + 100 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 166, + 91, + 207, + 100 + ], + "spans": [ + { + "bbox": [ + 166, + 91, + 207, + 100 + ], + "type": "text", + "content": "Yan et al." + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 9 + }, + { + "para_blocks": [ + { + "type": "table", + "bbox": [ + 135, + 114, + 479, + 156 + ], + "blocks": [ + { + "bbox": [ + 135, + 114, + 479, + 156 + ], + "lines": [ + { + "bbox": [ + 135, + 114, + 479, + 156 + ], + "spans": [ + { + "bbox": [ + 135, + 114, + 479, + 156 + ], + "type": "table", + "html": "
MethodsPanoVOS ValidationPanoVOS Test
J&FJsFsJuF_uJ&FJsFsJuF_u
PerSAM [61]19.112.319.817.427.119.57.414.923.831.7
SAM-PT [43]47.536.748.646.058.741.031.140.540.252.3
SAM-PT-reinit [43]43.734.344.341.354.943.635.042.743.553.0
", + "image_path": "83e7f16b5c8b78c3c972774dcb073f2134db6fc91801566309320c7927274d06.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "table_body" + } + ], + "index": 2 + }, + { + "type": "table", + "bbox": [ + 135, + 209, + 479, + 330 + ], + "blocks": [ + { + "bbox": [ + 132, + 164, + 480, + 197 + ], + "lines": [ + { + "bbox": [ + 132, + 164, + 480, + 197 + ], + "spans": [ + { + "bbox": [ + 132, + 164, + 480, + 197 + ], + "type": "text", + "content": "Table 5: Quantitative comparison on PanoVOS for variations of foundation model Segment Anything Model [23]. Subscript " + }, + { + "bbox": [ + 132, + 164, + 480, + 197 + ], + "type": "inline_equation", + "content": "s" + }, + { + "bbox": [ + 132, + 164, + 480, + 197 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 132, + 164, + 480, + 197 + ], + "type": "inline_equation", + "content": "u" + }, + { + "bbox": [ + 132, + 164, + 480, + 197 + ], + "type": "text", + "content": " denote scores in seen and unseen categories." + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 135, + 209, + 479, + 330 + ], + "lines": [ + { + "bbox": [ + 135, + 209, + 479, + 330 + ], + "spans": [ + { + "bbox": [ + 135, + 209, + 479, + 330 + ], + "type": "table", + "html": "
MethodsMFPanoVOS ValidationPanoVOS Test
J&FJsFsJuFuJ&FJsFsJuFu
CFBI† [57]35.834.644.824.239.719.118.226.112.219.8
CFBI+† [59]41.338.047.932.546.930.930.842.721.428.5
AOTT [58]65.659.468.359.775.053.449.361.647.555.1
AOTS [58]67.761.270.062.477.155.953.265.148.657.0
AOTB [58]67.662.372.061.574.855.453.564.247.756.0
Ours-Based74.066.480.466.283.056.849.462.752.462.5
AFB-URR [32]✓34.334.842.824.934.534.228.238.832.936.8
RDE [24]✓50.549.758.439.254.942.536.946.638.548.2
STCN [6]✓52.051.260.841.554.550.843.656.549.353.7
XMem [5]✓55.754.863.345.259.753.549.562.647.154.8
AOTL [58]✓66.661.471.159.474.353.850.060.347.857.1
R50-AOTL [58]✓65.361.971.456.471.654.652.963.247.554.9
SwinB-AOTL [58]✓62.158.966.554.368.853.149.057.849.056.6
Ours-Large✓77.970.585.269.586.459.954.969.253.062.4
RDE* [24]✓54.352.861.644.658.252.244.556.049.359.1
STCN* [6]✓51.751.260.641.353.653.853.758.146.057.3
XMem* [5]✓57.755.664.648.661.957.951.364.553.262.7
", + "image_path": "b7ab8bfaf4c6a5f3cf1ba76fefe205da3c47fdffa58ee928946d39993ce1cff5.jpg" + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "table_body" + } + ], + "index": 4 + }, + { + "bbox": [ + 132, + 338, + 480, + 383 + ], + "lines": [ + { + "bbox": [ + 132, + 338, + 480, + 383 + ], + "spans": [ + { + "bbox": [ + 132, + 338, + 480, + 383 + ], + "type": "text", + "content": "Table 6: Quantitative comparison on PanoVOS for models with pretraining on static image datasets. Subscript " + }, + { + "bbox": [ + 132, + 338, + 480, + 383 + ], + "type": "inline_equation", + "content": "s" + }, + { + "bbox": [ + 132, + 338, + 480, + 383 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 132, + 338, + 480, + 383 + ], + "type": "inline_equation", + "content": "u" + }, + { + "bbox": [ + 132, + 338, + 480, + 383 + ], + "type": "text", + "content": " denote scores in seen and unseen categories. " + }, + { + "bbox": [ + 132, + 338, + 480, + 383 + ], + "type": "inline_equation", + "content": "{MF}" + }, + { + "bbox": [ + 132, + 338, + 480, + 383 + ], + "type": "text", + "content": " denotes multiple historical frames as reference. * denotes a large-scale external dataset BL30K [6] dataset is used during training." + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "text" + }, + { + "bbox": [ + 132, + 407, + 278, + 419 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 407, + 278, + 419 + ], + "spans": [ + { + "bbox": [ + 132, + 407, + 278, + 419 + ], + "type": "text", + "content": "5.1 Implementation Details" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 130, + 430, + 480, + 514 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 430, + 480, + 514 + ], + "spans": [ + { + "bbox": [ + 130, + 430, + 480, + 514 + ], + "type": "text", + "content": "Model Architecture. We build two variants of our method with different reference bank sizes " + }, + { + "bbox": [ + 130, + 430, + 480, + 514 + ], + "type": "inline_equation", + "content": "\\mathcal{R}" + }, + { + "bbox": [ + 130, + 430, + 480, + 514 + ], + "type": "text", + "content": " for a fair comparison with previous methods. Ours-Base uses only the first frame and the previous frame as reference (" + }, + { + "bbox": [ + 130, + 430, + 480, + 514 + ], + "type": "inline_equation", + "content": "\\mathcal{R} = \\{1, t - 1\\}" + }, + { + "bbox": [ + 130, + 430, + 480, + 514 + ], + "type": "text", + "content": "), which are for the sake of high inference speed and low memory consumption. OursLarge uses multiple historical frames as reference (" + }, + { + "bbox": [ + 130, + 430, + 480, + 514 + ], + "type": "inline_equation", + "content": "\\mathcal{R} = \\{1 + 2\\delta, 1 + 2\\delta, 1 + 3\\delta \\ldots\\}" + }, + { + "bbox": [ + 130, + 430, + 480, + 514 + ], + "type": "text", + "content": "), which follows [34,58]. In our work, we set " + }, + { + "bbox": [ + 130, + 430, + 480, + 514 + ], + "type": "inline_equation", + "content": "\\delta" + }, + { + "bbox": [ + 130, + 430, + 480, + 514 + ], + "type": "text", + "content": " to 2 and 5 for training and testing respectively. For " + }, + { + "bbox": [ + 130, + 430, + 480, + 514 + ], + "type": "inline_equation", + "content": "p" + }, + { + "bbox": [ + 130, + 430, + 480, + 514 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 130, + 430, + 480, + 514 + ], + "type": "inline_equation", + "content": "s" + }, + { + "bbox": [ + 130, + 430, + 480, + 514 + ], + "type": "text", + "content": " in PSC block, we set them as 2 and 7." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 130, + 514, + 480, + 586 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 514, + 480, + 586 + ], + "spans": [ + { + "bbox": [ + 130, + 514, + 480, + 586 + ], + "type": "text", + "content": "Evaluation Metrics. Following the standard protocol [41, 42], we adopt the region accuracy " + }, + { + "bbox": [ + 130, + 514, + 480, + 586 + ], + "type": "inline_equation", + "content": "\\mathcal{I}" + }, + { + "bbox": [ + 130, + 514, + 480, + 586 + ], + "type": "text", + "content": " and boundary accuracy " + }, + { + "bbox": [ + 130, + 514, + 480, + 586 + ], + "type": "inline_equation", + "content": "\\mathcal{F}" + }, + { + "bbox": [ + 130, + 514, + 480, + 586 + ], + "type": "text", + "content": ". " + }, + { + "bbox": [ + 130, + 514, + 480, + 586 + ], + "type": "inline_equation", + "content": "\\mathcal{I}" + }, + { + "bbox": [ + 130, + 514, + 480, + 586 + ], + "type": "text", + "content": " means the Jaccard Index/Intersection over Union (IoU), which is the ratio of intersection and the joint area between predicted masks and ground truths. And " + }, + { + "bbox": [ + 130, + 514, + 480, + 586 + ], + "type": "inline_equation", + "content": "\\mathcal{F}" + }, + { + "bbox": [ + 130, + 514, + 480, + 586 + ], + "type": "text", + "content": " evaluates the accuracy of the segmentation boundary, which is computed by transforming it into a bipartite graph matching problem with predicted masks and ground truths." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 132, + 606, + 320, + 619 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 606, + 320, + 619 + ], + "spans": [ + { + "bbox": [ + 132, + 606, + 320, + 619 + ], + "type": "text", + "content": "5.2 Domain Transfer Results (RQ1)" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 130, + 629, + 480, + 665 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 629, + 480, + 665 + ], + "spans": [ + { + "bbox": [ + 130, + 629, + 480, + 665 + ], + "type": "text", + "content": "We evaluate previous SOTA methods, which are trained on conventional datasets that are captured by pinhole cameras, on PanoVOS datasets to evaluate the domain transfer performance. To quantify the transfer performance of advanced" + } + ] + } + ], + "index": 10 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 208, + 91, + 445, + 102 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 208, + 91, + 445, + 102 + ], + "spans": [ + { + "bbox": [ + 208, + 91, + 445, + 102 + ], + "type": "text", + "content": "PanoVOS: Bridging Non-panoramic and Panoramic Views" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 470, + 92, + 479, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 470, + 92, + 479, + 100 + ], + "spans": [ + { + "bbox": [ + 470, + 92, + 479, + 100 + ], + "type": "text", + "content": "11" + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 10 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 136, + 116, + 479, + 242 + ], + "blocks": [ + { + "bbox": [ + 136, + 116, + 479, + 242 + ], + "lines": [ + { + "bbox": [ + 136, + 116, + 479, + 242 + ], + "spans": [ + { + "bbox": [ + 136, + 116, + 479, + 242 + ], + "type": "image", + "image_path": "800c148b3ad0066f6c6b8b8d2f5dafc8655f2717768b8d7ef6d72cf91a6cdcc7.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 130, + 251, + 482, + 285 + ], + "lines": [ + { + "bbox": [ + 130, + 251, + 482, + 285 + ], + "spans": [ + { + "bbox": [ + 130, + 251, + 482, + 285 + ], + "type": "text", + "content": "Fig.6: Qualitative comparison to the state-of-the-art methods, RDE [24], STCN [6], and XMem [5], on PanoVOS dataset. Our model performs better under the challenge of content discontinuities. Error regions are bounded." + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_caption" + } + ], + "index": 2 + }, + { + "type": "table", + "bbox": [ + 221, + 300, + 392, + 350 + ], + "blocks": [ + { + "bbox": [ + 221, + 300, + 392, + 350 + ], + "lines": [ + { + "bbox": [ + 221, + 300, + 392, + 350 + ], + "spans": [ + { + "bbox": [ + 221, + 300, + 392, + 350 + ], + "type": "table", + "html": "
MethodsValidationTest
PSCAttnJ&FJ&F
Ours-Base✓72.855.4
74.056.8
Ours-Large✓74.859.5
77.959.9
", + "image_path": "ab48327d3c85cd9a7b9f3bf082d7bfe4254dfc53c08ea021bbafcf60915eca97.jpg" + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "table_body" + } + ], + "index": 4 + }, + { + "bbox": [ + 179, + 359, + 433, + 371 + ], + "lines": [ + { + "bbox": [ + 179, + 359, + 433, + 371 + ], + "spans": [ + { + "bbox": [ + 179, + 359, + 433, + 371 + ], + "type": "text", + "content": "Table 7: Ablation study of " + }, + { + "bbox": [ + 179, + 359, + 433, + 371 + ], + "type": "inline_equation", + "content": "PSCAttn" + }, + { + "bbox": [ + 179, + 359, + 433, + 371 + ], + "type": "text", + "content": " module on PanoVOS." + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "text" + }, + { + "bbox": [ + 130, + 398, + 482, + 555 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 398, + 482, + 555 + ], + "spans": [ + { + "bbox": [ + 130, + 398, + 482, + 555 + ], + "type": "text", + "content": "models trained on planar video datasets, we evaluated 15 off-the-shelf VOS models, including [5, 6, 24, 32, 57-59], and we follow official implementations and training strategies details of them. Table 3 summarizes the domain transfer results of methods that are only trained on synthetic datasets, such as COCO [33] and ECSSD [45], on PanoVOS dataset. Table 4 shows the domain transfer results of state-of-the-art methods, that are trained on synthetic datasets (e.g. COCO [33]) and video datasets (e.g. YouTube-VOS [52]), on our PanoVOS validation and test sets. By analyzing the performance of advanced VOS methods that target conventional planar videos on panoramic videos, we provide the following insights. Firstly, the performance of current sophisticated VOS models will largely degrade when employed to tackle panoramic videos. Secondly, we can observe a trend that training on larger VOS datasets, i.e., YouTube-VOS [52] and BL30K [6] can help mitigate the gap between planar and panoramic videos." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 131, + 578, + 387, + 591 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 131, + 578, + 387, + 591 + ], + "spans": [ + { + "bbox": [ + 131, + 578, + 387, + 591 + ], + "type": "text", + "content": "5.3 Results via Visual Foundation Model (RQ2)" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 130, + 605, + 482, + 666 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 605, + 482, + 666 + ], + "spans": [ + { + "bbox": [ + 130, + 605, + 482, + 666 + ], + "type": "text", + "content": "To quantify the segmentation performance of different variations of the foundation model Segment Anything Model [23] on PanoVOS, we evaluate the latest top performing models PerSAM [61] and SAM-PT [43], as shown in Table 5. The performance of these models on our challenging PanoVOS dataset is still unsatisfactory, which leaves space for further exploration." + } + ] + } + ], + "index": 8 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "spans": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "type": "text", + "content": "12" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 166, + 91, + 208, + 100 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 166, + 91, + 208, + 100 + ], + "spans": [ + { + "bbox": [ + 166, + 91, + 208, + 100 + ], + "type": "text", + "content": "Yan et al." + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 11 + }, + { + "para_blocks": [ + { + "type": "table", + "bbox": [ + 221, + 114, + 394, + 164 + ], + "blocks": [ + { + "bbox": [ + 221, + 114, + 394, + 164 + ], + "lines": [ + { + "bbox": [ + 221, + 114, + 394, + 164 + ], + "spans": [ + { + "bbox": [ + 221, + 114, + 394, + 164 + ], + "type": "table", + "html": "
MethodsAttention TypeValidation J&FTest J&F
Ours-BaseCrossAttn72.554.8
PSCAttn74.056.8
Ours-LargeCrossAttn76.859.1
PSCAttn77.959.9
", + "image_path": "2b5431346cb259ac41a16bf529c49588f68147d776449609e3e382423fa03d64.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "table_body" + } + ], + "index": 2 + }, + { + "type": "table", + "bbox": [ + 221, + 205, + 393, + 270 + ], + "blocks": [ + { + "bbox": [ + 132, + 173, + 481, + 195 + ], + "lines": [ + { + "bbox": [ + 132, + 173, + 481, + 195 + ], + "spans": [ + { + "bbox": [ + 132, + 173, + 481, + 195 + ], + "type": "text", + "content": "Table 8: Comparison between our PSC attention (PSCAttn) and cross attention (CrossAttn) module on PanoVOS dataset." + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 221, + 205, + 393, + 270 + ], + "lines": [ + { + "bbox": [ + 221, + 205, + 393, + 270 + ], + "spans": [ + { + "bbox": [ + 221, + 205, + 393, + 270 + ], + "type": "table", + "html": "
MethodsPanoVOS Validation
J&FJsFsJuFu
w/o73.768.882.663.180.3
p=376.370.485.065.884.1
p=574.265.379.767.584.2
p=1075.968.181.968.685.2
p=1575.066.781.067.085.4
p=2 (Ours)77.970.585.269.586.4
", + "image_path": "aa6df0c14ad436ddc92dde623a746a2318c4aa5ce65373bebbc2df67cd15e2c9.jpg" + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "table_body" + } + ], + "index": 4 + }, + { + "bbox": [ + 132, + 278, + 480, + 300 + ], + "lines": [ + { + "bbox": [ + 132, + 278, + 480, + 300 + ], + "spans": [ + { + "bbox": [ + 132, + 278, + 480, + 300 + ], + "type": "text", + "content": "Table 9: Hyperparameter Analysis of " + }, + { + "bbox": [ + 132, + 278, + 480, + 300 + ], + "type": "inline_equation", + "content": "p" + }, + { + "bbox": [ + 132, + 278, + 480, + 300 + ], + "type": "text", + "content": ", which enables the stitching mechanism, in PSCAttn for Ours-Large model." + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "text" + }, + { + "bbox": [ + 132, + 323, + 331, + 336 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 323, + 331, + 336 + ], + "spans": [ + { + "bbox": [ + 132, + 323, + 331, + 336 + ], + "type": "text", + "content": "5.4 Main Results on PanoVOS (RQ3)" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 130, + 342, + 481, + 378 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 342, + 481, + 378 + ], + "spans": [ + { + "bbox": [ + 130, + 342, + 481, + 378 + ], + "type": "text", + "content": "To evaluate the performance of previous methods on the proposed panoramic VOS dataset, we re-trained them on the training set of PanoVOS for the sake of fairness. We report the performance in Table 6," + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 130, + 379, + 482, + 475 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 379, + 482, + 475 + ], + "spans": [ + { + "bbox": [ + 130, + 379, + 482, + 475 + ], + "type": "text", + "content": "which demonstrates that all the previous VOS models perform worse on PanoVOS than on the traditional VOS benchmarks, e.g., YouTube-VOS. Our model substantially outperforms all these methods and achieves state-of-the-art on all evaluation metrics on PanoVOS, which verifies the effectiveness of our model in tackling panoramic videos. Fig.6 visualizes some qualitative comparisons between our model and previous state-of-the-art methods on PanoVOS dataset, which shows that previous benchmarks fail to cope with content discontinuities while our model tackles them well." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 132, + 491, + 273, + 503 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 491, + 273, + 503 + ], + "spans": [ + { + "bbox": [ + 132, + 491, + 273, + 503 + ], + "type": "text", + "content": "5.5 Ablation Study (RQ4)" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 130, + 510, + 482, + 666 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 510, + 482, + 666 + ], + "spans": [ + { + "bbox": [ + 130, + 510, + 482, + 666 + ], + "type": "text", + "content": "In this section, we conduct ablation studies to demonstrate the effectiveness of the main component, i.e., Panoramic Space Consistency Attention (PSCAttn), of our model, with all the experiments performed based on our two model variants, i.e., Ours-Base and Ours-Large. For training, static image datasets are used for pre-training and PanoVOS is used for main training. Table 7 demonstrates the effectiveness of our PSCAttn module. Besides, Fig. 7 illustrates the qualitative comparison between our default model (Ours-Base) and the setting without PSCAttn module. Our model performs better when coping with the pixel discontinuity problem. Moreover, as is shown in Table 8, compared to the conventional cross-attention (CrossAttn) module, PSCAttn also achieves better performance. In Table 9, we analyze the hyperparameter " + }, + { + "bbox": [ + 130, + 510, + 482, + 666 + ], + "type": "inline_equation", + "content": "p" + }, + { + "bbox": [ + 130, + 510, + 482, + 666 + ], + "type": "text", + "content": ", which influences the stitching mechanism in PSCAttn, of our model (Ours-Large) on the PanoVOS validation set. Specifically, the highest overall performance (" + }, + { + "bbox": [ + 130, + 510, + 482, + 666 + ], + "type": "inline_equation", + "content": "\\mathcal{J} \\& \\mathcal{F}" + }, + { + "bbox": [ + 130, + 510, + 482, + 666 + ], + "type": "text", + "content": ") is achieved" + } + ] + } + ], + "index": 10 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 207, + 91, + 445, + 102 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 207, + 91, + 445, + 102 + ], + "spans": [ + { + "bbox": [ + 207, + 91, + 445, + 102 + ], + "type": "text", + "content": "PanoVOS: Bridging Non-panoramic and Panoramic Views" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 470, + 91, + 481, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 470, + 91, + 481, + 100 + ], + "spans": [ + { + "bbox": [ + 470, + 91, + 481, + 100 + ], + "type": "text", + "content": "13" + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 12 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 222, + 120, + 394, + 181 + ], + "blocks": [ + { + "bbox": [ + 222, + 120, + 394, + 181 + ], + "lines": [ + { + "bbox": [ + 222, + 120, + 394, + 181 + ], + "spans": [ + { + "bbox": [ + 222, + 120, + 394, + 181 + ], + "type": "image", + "image_path": "9aae5cfcf093a8c9b0f1d38aaa5378447d41276ac2ec1c73a8e34f0ccc6510b9.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 185, + 194, + 427, + 207 + ], + "lines": [ + { + "bbox": [ + 185, + 194, + 427, + 207 + ], + "spans": [ + { + "bbox": [ + 185, + 194, + 427, + 207 + ], + "type": "text", + "content": "Fig. 7: Qualitative ablation study of PSCAttn module." + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_caption" + } + ], + "index": 2 + }, + { + "type": "image", + "bbox": [ + 219, + 217, + 394, + 248 + ], + "blocks": [ + { + "bbox": [ + 219, + 217, + 394, + 248 + ], + "lines": [ + { + "bbox": [ + 219, + 217, + 394, + 248 + ], + "spans": [ + { + "bbox": [ + 219, + 217, + 394, + 248 + ], + "type": "image", + "image_path": "be77e287852906b466b08a095b21268ebaf5e6ed7db608afbae02dfd0b698270.jpg" + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 134, + 256, + 478, + 269 + ], + "lines": [ + { + "bbox": [ + 134, + 256, + 478, + 269 + ], + "spans": [ + { + "bbox": [ + 134, + 256, + 478, + 269 + ], + "type": "text", + "content": "Fig. 8: Challenge. Our model fails to segment some objects with strong distortion." + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "image_caption" + } + ], + "index": 4 + }, + { + "bbox": [ + 130, + 290, + 482, + 327 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 290, + 482, + 327 + ], + "spans": [ + { + "bbox": [ + 130, + 290, + 482, + 327 + ], + "type": "text", + "content": "when setting " + }, + { + "bbox": [ + 130, + 290, + 482, + 327 + ], + "type": "inline_equation", + "content": "p" + }, + { + "bbox": [ + 130, + 290, + 482, + 327 + ], + "type": "text", + "content": " as 2. Compared to the setting without using the stitching mechanism " + }, + { + "bbox": [ + 130, + 290, + 482, + 327 + ], + "type": "inline_equation", + "content": "(w / o)" + }, + { + "bbox": [ + 130, + 290, + 482, + 327 + ], + "type": "text", + "content": ", our model can achieve much better performance. Specifically, our final model (Ours-Large, " + }, + { + "bbox": [ + 130, + 290, + 482, + 327 + ], + "type": "inline_equation", + "content": "p = 2" + }, + { + "bbox": [ + 130, + 290, + 482, + 327 + ], + "type": "text", + "content": ") achieves more than " + }, + { + "bbox": [ + 130, + 290, + 482, + 327 + ], + "type": "inline_equation", + "content": "4\\%" + }, + { + "bbox": [ + 130, + 290, + 482, + 327 + ], + "type": "text", + "content": " gain in " + }, + { + "bbox": [ + 130, + 290, + 482, + 327 + ], + "type": "inline_equation", + "content": "\\mathcal{I} \\& \\mathcal{F}" + }, + { + "bbox": [ + 130, + 290, + 482, + 327 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 132, + 342, + 339, + 355 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 342, + 339, + 355 + ], + "spans": [ + { + "bbox": [ + 132, + 342, + 339, + 355 + ], + "type": "text", + "content": "5.6 Limitation and Future Work (RQ5)" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 130, + 361, + 482, + 506 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 361, + 482, + 506 + ], + "spans": [ + { + "bbox": [ + 130, + 361, + 482, + 506 + ], + "type": "text", + "content": "To prompt greater progress of panoramic VOS, we also analyze the limitations of our method. Specifically, our method has no notion of severe distortion challenge since we do not employ a special design (such as deformable convolution [9]) to tackle deformations. That means our model may not segment the objects with large distortions. One such failure case is shown in Fig. 8. Besides, our panoramic dataset can be applied to broader video segmentation and tracking domains, such as referring video object segmentation [28-30, 56], video object tracking [18], video instance segmentation [15], few-shot segmentation [20], and more broader embodied navigation tasks [55]. Also, it would be valuable to investigate the zero-shot segmentation performance of visual foundation models [23] on our challenging panoramic dataset. We hope our work can shed light on efficient adaptation from non-panoramic to panoramic perception." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 132, + 521, + 220, + 534 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 521, + 220, + 534 + ], + "spans": [ + { + "bbox": [ + 132, + 521, + 220, + 534 + ], + "type": "text", + "content": "6 Conclusion" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 130, + 545, + 482, + 666 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 545, + 482, + 666 + ], + "spans": [ + { + "bbox": [ + 130, + 545, + 482, + 666 + ], + "type": "text", + "content": "In this paper, we introduce a high-quality dataset, i.e., PanoVOS, for panoramic video object segmentation. Our PanoVOS dataset provides pixel-level instance annotations with diverse scenarios and significant motions. Based on this dataset, we evaluate 15 off-the-shelf VOS models and carefully analyze their limitations. Then, we further present our model, i.e., PSCFormer, which is equipped with the proposed panoramic space consistency transformer block. Our preliminary experiment demonstrates the effectiveness of our proposed model to enhance the segmentation performance and consistency in panoramic scenes. In conclusion, this provides a new challenge for video understanding, and we hope our PanoVOS dataset can attract more researchers to pay attention to panoramic videos." + } + ] + } + ], + "index": 10 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "spans": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "type": "text", + "content": "14" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 166, + 91, + 208, + 100 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 166, + 91, + 208, + 100 + ], + "spans": [ + { + "bbox": [ + 166, + 91, + 208, + 100 + ], + "type": "text", + "content": "Yan et al." + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 13 + }, + { + "para_blocks": [ + { + "bbox": [ + 133, + 114, + 246, + 129 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 114, + 246, + 129 + ], + "spans": [ + { + "bbox": [ + 133, + 114, + 246, + 129 + ], + "type": "text", + "content": "Acknowledgements" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 132, + 137, + 482, + 185 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 137, + 482, + 185 + ], + "spans": [ + { + "bbox": [ + 132, + 137, + 482, + 185 + ], + "type": "text", + "content": "This work was supported in part by National Natural Science Foundation of China (No.62072112), and Scientific and Technological Innovation Action Plan of Shanghai Science and Technology Committee (No.22511101502, No.22511102202 and No.21DZ2203300)." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 133, + 201, + 197, + 214 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 201, + 197, + 214 + ], + "spans": [ + { + "bbox": [ + 133, + 201, + 197, + 214 + ], + "type": "text", + "content": "References" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 134, + 223, + 480, + 665 + ], + "type": "list", + "angle": 0, + "index": 19, + "blocks": [ + { + "bbox": [ + 138, + 223, + 480, + 256 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 223, + 480, + 256 + ], + "spans": [ + { + "bbox": [ + 138, + 223, + 480, + 256 + ], + "type": "text", + "content": "1. Ai, H., Cao, Z., Zhu, J., Bai, H., Chen, Y., Wang, L.: Deep learning for omnidirectional vision: A survey and new perspectives. arXiv preprint arXiv:2205.10468 (2022)" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 138, + 257, + 480, + 277 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 257, + 480, + 277 + ], + "spans": [ + { + "bbox": [ + 138, + 257, + 480, + 277 + ], + "type": "text", + "content": "2. Armeni, I., Sax, S., Zamir, A.R., Savarese, S.: Joint 2d-3dsemantic data for indoor scene understanding. arXiv preprint arXiv:1702.01105 (2017)" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 138, + 278, + 480, + 310 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 278, + 480, + 310 + ], + "spans": [ + { + "bbox": [ + 138, + 278, + 480, + 310 + ], + "type": "text", + "content": "3. Caelles, S., Maninis, K.K., Pont-Tuset, J., Leal-Taixe, L., Cremers, D., Van Gool, L.: One-shot video object segmentation. In: Proceedings of the IEEE conference on computer vision and pattern recognition. pp. 221-230 (2017)" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 138, + 311, + 480, + 342 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 311, + 480, + 342 + ], + "spans": [ + { + "bbox": [ + 138, + 311, + 480, + 342 + ], + "type": "text", + "content": "4. Chen, X., Li, Z., Yuan, Y., Yu, G., Shen, J., Qi, D.: State-aware tracker for real-time video object segmentation. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 9384-9393 (2020)" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 138, + 343, + 480, + 374 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 343, + 480, + 374 + ], + "spans": [ + { + "bbox": [ + 138, + 343, + 480, + 374 + ], + "type": "text", + "content": "5. Cheng, H.K., Schwing, A.G.: Xmem: Long-term video object segmentation with an atkinson-shiffrin memory model. In: European Conference on Computer Vision. pp. 640-658. Springer (2022)" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 138, + 375, + 480, + 407 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 375, + 480, + 407 + ], + "spans": [ + { + "bbox": [ + 138, + 375, + 480, + 407 + ], + "type": "text", + "content": "6. Cheng, H.K., Tai, Y.W., Tang, C.K.: Rethinking space-time networks with improved memory coverage for efficient video object segmentation. Advances in Neural Information Processing Systems 34, 11781-11794 (2021)" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 138, + 407, + 480, + 439 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 407, + 480, + 439 + ], + "spans": [ + { + "bbox": [ + 138, + 407, + 480, + 439 + ], + "type": "text", + "content": "7. Cheng, H.T., Chao, C.H., Dong, J.D., Wen, H.K., Liu, T.L., Sun, M.: Cube padding for weakly-supervised saliency prediction in 360 videos. In: Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition. pp. 1420-1429 (2018)" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 138, + 440, + 480, + 471 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 440, + 480, + 471 + ], + "spans": [ + { + "bbox": [ + 138, + 440, + 480, + 471 + ], + "type": "text", + "content": "8. Cheng, J., Tsai, Y.H., Hung, W.C., Wang, S., Yang, M.H.: Fast and accurate online video object segmentation via tracking parts. In: Proceedings of the IEEE conference on computer vision and pattern recognition. pp. 7415-7424 (2018)" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 138, + 472, + 480, + 503 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 472, + 480, + 503 + ], + "spans": [ + { + "bbox": [ + 138, + 472, + 480, + 503 + ], + "type": "text", + "content": "9. Dai, J., Qi, H., Xiong, Y., Li, Y., Zhang, G., Hu, H., Wei, Y.: Deformable convolutional networks. In: Proceedings of the IEEE international conference on computer vision. pp. 764-773 (2017)" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 134, + 504, + 480, + 536 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 134, + 504, + 480, + 536 + ], + "spans": [ + { + "bbox": [ + 134, + 504, + 480, + 536 + ], + "type": "text", + "content": "10. Dang, J., Zheng, H., Xu, X., Guo, Y.: Unified spatio-temporal dynamic routing for efficient video object segmentation. IEEE Transactions on Intelligent Transportation Systems (2023)" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 134, + 536, + 480, + 567 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 134, + 536, + 480, + 567 + ], + "spans": [ + { + "bbox": [ + 134, + 536, + 480, + 567 + ], + "type": "text", + "content": "11. Dang, J., Zheng, H., Xu, X., Wang, L., Hu, Q., Guo, Y.: Adaptive sparse memory networks for efficient and robust video object segmentation. IEEE Transactions on Neural Networks and Learning Systems (2024)" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 134, + 568, + 480, + 601 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 134, + 568, + 480, + 601 + ], + "spans": [ + { + "bbox": [ + 134, + 568, + 480, + 601 + ], + "type": "text", + "content": "12. Eger Passos, D., Jung, B.: Measuring the accuracy of inside-out tracking in xr devices using a high-precision robotic arm. In: International Conference on Human-Computer Interaction. pp. 19–26. Springer (2020)" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 134, + 601, + 480, + 632 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 134, + 601, + 480, + 632 + ], + "spans": [ + { + "bbox": [ + 134, + 601, + 480, + 632 + ], + "type": "text", + "content": "13. Fang, R., Yan, S., Huang, Z., Zhou, J., Tian, H., Dai, J., Li, H.: Instructseq: Unifying vision tasks with instruction-conditioned multi-modal sequence generation. arXiv preprint arXiv:2311.18835 (2023)" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 134, + 633, + 480, + 665 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 134, + 633, + 480, + 665 + ], + "spans": [ + { + "bbox": [ + 134, + 633, + 480, + 665 + ], + "type": "text", + "content": "14. Guo, P., Hong, L., Zhou, X., Gao, S., Li, W., Li, J., Chen, Z., Li, X., Zhang, W., Zhang, W.: Clickvos: Click video object segmentation. arXiv preprint arXiv:2403.06130 (2024)" + } + ] + } + ], + "index": 18 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 207, + 91, + 445, + 102 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 207, + 91, + 445, + 102 + ], + "spans": [ + { + "bbox": [ + 207, + 91, + 445, + 102 + ], + "type": "text", + "content": "PanoVOS: Bridging Non-panoramic and Panoramic Views" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 470, + 92, + 480, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 470, + 92, + 480, + 100 + ], + "spans": [ + { + "bbox": [ + 470, + 92, + 480, + 100 + ], + "type": "text", + "content": "15" + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 14 + }, + { + "para_blocks": [ + { + "bbox": [ + 132, + 116, + 481, + 665 + ], + "type": "list", + "angle": 0, + "index": 19, + "blocks": [ + { + "bbox": [ + 133, + 116, + 480, + 138 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 116, + 480, + 138 + ], + "spans": [ + { + "bbox": [ + 133, + 116, + 480, + 138 + ], + "type": "text", + "content": "15. Guo, P., Huang, T., He, P., Liu, X., Xiao, T., Chen, Z., Zhang, W.: Openvis: Open-vocabulary video instance segmentation. arXiv preprint arXiv:2305.16835 (2023)" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 133, + 138, + 481, + 170 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 138, + 481, + 170 + ], + "spans": [ + { + "bbox": [ + 133, + 138, + 481, + 170 + ], + "type": "text", + "content": "16. Guo, P., Zhang, W., Li, X., Zhang, W.: Adaptive online mutual learning bi-decoders for video object segmentation. IEEE Transactions on Image Processing 31, 7063-7077 (2022)" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 133, + 171, + 480, + 191 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 171, + 480, + 191 + ], + "spans": [ + { + "bbox": [ + 133, + 171, + 480, + 191 + ], + "type": "text", + "content": "17. Hendrycks, D., Gimpel, K.: Gaussian error linear units (gelus). arXiv preprint arXiv:1606.08415 (2016)" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 132, + 192, + 480, + 235 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 192, + 480, + 235 + ], + "spans": [ + { + "bbox": [ + 132, + 192, + 480, + 235 + ], + "type": "text", + "content": "18. Hong, L., Yan, S., Zhang, R., Li, W., Zhou, X., Guo, P., Jiang, K., Chen, Y., Li, J., Chen, Z., et al.: Onetracker: Unifying visual object tracking with foundation models and efficient tuning. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 19079-19091 (2024)" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 132, + 236, + 480, + 278 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 236, + 480, + 278 + ], + "spans": [ + { + "bbox": [ + 132, + 236, + 480, + 278 + ], + "type": "text", + "content": "19. Hu, L., Zhang, P., Zhang, B., Pan, P., Xu, Y., Jin, R.: Learning position and target consistency for memory-based video object segmentation. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 4144-4154 (2021)" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 132, + 279, + 480, + 300 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 279, + 480, + 300 + ], + "spans": [ + { + "bbox": [ + 132, + 279, + 480, + 300 + ], + "type": "text", + "content": "20. Iqbal, E., Safarov, S., Bang, S.: Misanet: Multi-similarity and attention guidance for boosting few-shot segmentation. arXiv preprint arXiv:2206.09667 (2022)" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 132, + 300, + 480, + 332 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 300, + 480, + 332 + ], + "spans": [ + { + "bbox": [ + 132, + 300, + 480, + 332 + ], + "type": "text", + "content": "21. Jiang, H., Jiang, G., Yu, M., Zhang, Y., Yang, Y., Peng, Z., Chen, F., Zhang, Q.: Cubemap-based perception-driven blind quality assessment for 360-degree images. IEEE Transactions on Image Processing 30, 2364-2377 (2021)" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 132, + 332, + 480, + 364 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 332, + 480, + 364 + ], + "spans": [ + { + "bbox": [ + 132, + 332, + 480, + 364 + ], + "type": "text", + "content": "22. Jost, T.A., Nelson, B., Rylander, J.: Quantitative analysis of the oculus rifts in controlled movement. Disability and Rehabilitation: Assistive Technology 16(6), 632-636 (2021)" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 132, + 364, + 480, + 396 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 364, + 480, + 396 + ], + "spans": [ + { + "bbox": [ + 132, + 364, + 480, + 396 + ], + "type": "text", + "content": "23. Kirillov, A., Mintun, E., Ravi, N., Mao, H., Rolland, C., Gustafson, L., Xiao, T., Whitehead, S., Berg, A.C., Lo, W.Y., et al.: Segment anything. arXiv preprint arXiv:2304.02643 (2023)" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 132, + 396, + 480, + 429 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 396, + 480, + 429 + ], + "spans": [ + { + "bbox": [ + 132, + 396, + 480, + 429 + ], + "type": "text", + "content": "24. Li, M., Hu, L., Xiong, Z., Zhang, B., Pan, P., Liu, D.: Recurrent dynamic embedding for video object segmentation. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 1332-1341 (2022)" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 132, + 429, + 480, + 460 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 429, + 480, + 460 + ], + "spans": [ + { + "bbox": [ + 132, + 429, + 480, + 460 + ], + "type": "text", + "content": "25. Li, W., Fan, J., Guo, P., Hong, L., Zhang, W.: Hfvos: History-future integrated dynamic memory for video object segmentation. IEEE Transactions on Circuits and Systems for Video Technology (2024)" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 132, + 460, + 480, + 493 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 460, + 480, + 493 + ], + "spans": [ + { + "bbox": [ + 132, + 460, + 480, + 493 + ], + "type": "text", + "content": "26. Li, W., Guo, P., Zhou, X., Hong, L., He, Y., Zheng, X., Zhang, W., Zhang, W.: Onevos: Unifying video object segmentation with all-in-one transformer framework. arXiv preprint arXiv:2403.08682 (2024)" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 132, + 493, + 480, + 514 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 493, + 480, + 514 + ], + "spans": [ + { + "bbox": [ + 132, + 493, + 480, + 514 + ], + "type": "text", + "content": "27. Li, X., Cao, H., Zhao, S., Li, J., Zhang, L., Raj, B.: Panoramic video salient object detection with ambisonic audio guidance. arXiv preprint arXiv:2211.14419 (2022)" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 132, + 514, + 480, + 546 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 514, + 480, + 546 + ], + "spans": [ + { + "bbox": [ + 132, + 514, + 480, + 546 + ], + "type": "text", + "content": "28. Li, X., Wang, J., Xu, X., Li, X., Raj, B., Lu, Y.: Robust referring video object segmentation with cyclic structural consensus. In: Proceedings of the IEEE/CVF International Conference on Computer Vision. pp. 22236-22245 (2023)" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 132, + 547, + 480, + 590 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 547, + 480, + 590 + ], + "spans": [ + { + "bbox": [ + 132, + 547, + 480, + 590 + ], + "type": "text", + "content": "29. Li, X., Wang, J., Xu, X., Peng, X., Singh, R., Lu, Y., Raj, B.: Qdformer: Towards robust audiovisual segmentation in complex environments with quantization-based semantic decomposition. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 3402-3413 (2024)" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 132, + 590, + 480, + 633 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 590, + 480, + 633 + ], + "spans": [ + { + "bbox": [ + 132, + 590, + 480, + 633 + ], + "type": "text", + "content": "30. Li, X., Wang, J., Xu, X., Yang, M., Yang, F., Zhao, Y., Singh, R., Raj, B.: Towards noise-tolerant speech-referring video object segmentation: Bridging speech and text. In: Proceedings of the 2023 Conference on Empirical Methods in Natural Language Processing. pp. 2283-2296 (2023)" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 132, + 633, + 480, + 665 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 633, + 480, + 665 + ], + "spans": [ + { + "bbox": [ + 132, + 633, + 480, + 665 + ], + "type": "text", + "content": "31. Liang, S., Shen, X., Huang, J., Hua, X.S.: Video object segmentation with dynamic memory networks and adaptive object alignment. In: Proceedings of the IEEE/CVF International Conference on Computer Vision. pp. 8065-8074 (2021)" + } + ] + } + ], + "index": 18 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "spans": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "type": "text", + "content": "16" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 166, + 91, + 206, + 100 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 166, + 91, + 206, + 100 + ], + "spans": [ + { + "bbox": [ + 166, + 91, + 206, + 100 + ], + "type": "text", + "content": "Yan et al." + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 15 + }, + { + "para_blocks": [ + { + "bbox": [ + 130, + 116, + 482, + 665 + ], + "type": "list", + "angle": 0, + "index": 18, + "blocks": [ + { + "bbox": [ + 130, + 116, + 482, + 149 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 116, + 482, + 149 + ], + "spans": [ + { + "bbox": [ + 130, + 116, + 482, + 149 + ], + "type": "text", + "content": "32. Liang, Y., Li, X., Jafari, N., Chen, J.: Video object segmentation with adaptive feature bank and uncertain-region refinement. Advances in Neural Information Processing Systems 33, 3430-3441 (2020)" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 130, + 150, + 481, + 184 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 150, + 481, + 184 + ], + "spans": [ + { + "bbox": [ + 130, + 150, + 481, + 184 + ], + "type": "text", + "content": "33. Lin, T.Y., Maire, M., Belongie, S., Hays, J., Perona, P., Ramanan, D., Dollar, P., Zitnick, C.L.: Microsoft coco: Common objects in context. In: European conference on computer vision. pp. 740-755. Springer (2014)" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 130, + 185, + 481, + 217 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 185, + 481, + 217 + ], + "spans": [ + { + "bbox": [ + 130, + 185, + 481, + 217 + ], + "type": "text", + "content": "34. Liu, Y., Yu, R., Wang, J., Zhao, X., Wang, Y., Tang, Y., Yang, Y.: Global spectral filter memory network for video object segmentation. In: European Conference on Computer Vision. pp. 648-665. Springer (2022)" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 130, + 218, + 481, + 262 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 218, + 481, + 262 + ], + "spans": [ + { + "bbox": [ + 130, + 218, + 481, + 262 + ], + "type": "text", + "content": "35. Ma, C., Zhang, J., Yang, K., Roitberg, A., Stiefelhagen, R.: Densepass: Dense panoramic semantic segmentation via unsupervised domain adaptation with attention-augmented context exchange. In: 2021 IEEE International Intelligent Transportation Systems Conference (ITSC). pp. 2766-2772. IEEE (2021)" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 130, + 262, + 481, + 296 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 262, + 481, + 296 + ], + "spans": [ + { + "bbox": [ + 130, + 262, + 481, + 296 + ], + "type": "text", + "content": "36. Maninis, K.K., Caelles, S., Chen, Y., Pont-Tuset, J., Leal-Taixe, L., Cremers, D., Van Gool, L.: Video object segmentation without temporal information. IEEE transactions on pattern analysis and machine intelligence 41(6), 1515-1530 (2018)" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 130, + 297, + 481, + 329 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 297, + 481, + 329 + ], + "spans": [ + { + "bbox": [ + 130, + 297, + 481, + 329 + ], + "type": "text", + "content": "37. Mao, Y., Wang, N., Zhou, W., Li, H.: Joint inductive and transductive learning for video object segmentation. In: Proceedings of the IEEE/CVF International Conference on Computer Vision. pp. 9670-9679 (2021)" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 130, + 330, + 481, + 373 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 330, + 481, + 373 + ], + "spans": [ + { + "bbox": [ + 130, + 330, + 481, + 373 + ], + "type": "text", + "content": "38. Mei, J., Zhu, A.Z., Yan, X., Yan, H., Qiao, S., Chen, L.C., Kretzschmar, H.: Waymo open dataset: Panoramic video panoptic segmentation. In: Computer Vision-ECCV 2022: 17th European Conference, Tel Aviv, Israel, October 23-27, 2022, Proceedings, Part XXIX. pp. 53-72. Springer (2022)" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 130, + 374, + 481, + 407 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 374, + 481, + 407 + ], + "spans": [ + { + "bbox": [ + 130, + 374, + 481, + 407 + ], + "type": "text", + "content": "39. Oh, S.W., Lee, J.Y., Sunkavalli, K., Kim, S.J.: Fast video object segmentation by reference-guided mask propagation. In: Proceedings of the IEEE conference on computer vision and pattern recognition. pp. 7376-7385 (2018)" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 130, + 407, + 481, + 441 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 407, + 481, + 441 + ], + "spans": [ + { + "bbox": [ + 130, + 407, + 481, + 441 + ], + "type": "text", + "content": "40. Oh, S.W., Lee, J.Y., Xu, N., Kim, S.J.: Video object segmentation using space-time memory networks. In: Proceedings of the IEEE/CVF International Conference on Computer Vision. pp. 9226-9235 (2019)" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 130, + 441, + 481, + 485 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 441, + 481, + 485 + ], + "spans": [ + { + "bbox": [ + 130, + 441, + 481, + 485 + ], + "type": "text", + "content": "41. Perazzi, F., Pont-Tuset, J., McWilliams, B., Van Gool, L., Gross, M., Sorkine-Hornung, A.: A benchmark dataset and evaluation methodology for video object segmentation. In: Proceedings of the IEEE conference on computer vision and pattern recognition. pp. 724-732 (2016)" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 130, + 486, + 481, + 518 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 486, + 481, + 518 + ], + "spans": [ + { + "bbox": [ + 130, + 486, + 481, + 518 + ], + "type": "text", + "content": "42. Pont-Tuset, J., Perazzi, F., Caelles, S., Arbeláez, P., Sorkine-Hornung, A., Van Gool, L.: The 2017 davis challenge on video object segmentation. arXiv preprint arXiv:1704.00675 (2017)" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 130, + 520, + 481, + 541 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 520, + 481, + 541 + ], + "spans": [ + { + "bbox": [ + 130, + 520, + 481, + 541 + ], + "type": "text", + "content": "43. Rajic, F., Ke, L., Tai, Y.W., Tang, C.K., Danelljan, M., Yu, F.: Segment anything meets point tracking. arXiv preprint arXiv:2307.01197 (2023)" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 130, + 542, + 481, + 575 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 542, + 481, + 575 + ], + "spans": [ + { + "bbox": [ + 130, + 542, + 481, + 575 + ], + "type": "text", + "content": "44. Seong, H., Hyun, J., Kim, E.: Kernelized memory network for video object segmentation. In: European Conference on Computer Vision. pp. 629-645. Springer (2020)" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 130, + 576, + 481, + 609 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 576, + 481, + 609 + ], + "spans": [ + { + "bbox": [ + 130, + 576, + 481, + 609 + ], + "type": "text", + "content": "45. Shi, J., Yan, Q., Xu, L., Jia, J.: Hierarchical image saliency detection on extended cssd. IEEE transactions on pattern analysis and machine intelligence 38(4), 717-729 (2015)" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 130, + 609, + 481, + 643 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 609, + 481, + 643 + ], + "spans": [ + { + "bbox": [ + 130, + 609, + 481, + 643 + ], + "type": "text", + "content": "46. Vaswani, A., Shazeer, N., Parmar, N., Uszkoreit, J., Jones, L., Gomez, A.N., Kaiser, L., Polosukhin, I.: Attention is all you need. Advances in neural information processing systems 30 (2017)" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 130, + 643, + 481, + 665 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 643, + 481, + 665 + ], + "spans": [ + { + "bbox": [ + 130, + 643, + 481, + 665 + ], + "type": "text", + "content": "47. Voigtlaender, P., Leibe, B.: Online adaptation of convolutional neural networks for video object segmentation. arXiv preprint arXiv:1706.09364 (2017)" + } + ] + } + ], + "index": 17 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 207, + 91, + 445, + 102 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 207, + 91, + 445, + 102 + ], + "spans": [ + { + "bbox": [ + 207, + 91, + 445, + 102 + ], + "type": "text", + "content": "PanoVOS: Bridging Non-panoramic and Panoramic Views" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 470, + 91, + 481, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 470, + 91, + 481, + 100 + ], + "spans": [ + { + "bbox": [ + 470, + 91, + 481, + 100 + ], + "type": "text", + "content": "17" + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 16 + }, + { + "para_blocks": [ + { + "bbox": [ + 132, + 117, + 480, + 643 + ], + "type": "list", + "angle": 0, + "index": 19, + "blocks": [ + { + "bbox": [ + 133, + 117, + 480, + 149 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 117, + 480, + 149 + ], + "spans": [ + { + "bbox": [ + 133, + 117, + 480, + 149 + ], + "type": "text", + "content": "48. Wang, H., Jiang, X., Ren, H., Hu, Y., Bai, S.: Swiftnet: Real-time video object segmentation. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 1296-1305 (2021)" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 133, + 150, + 480, + 172 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 150, + 480, + 172 + ], + "spans": [ + { + "bbox": [ + 133, + 150, + 480, + 172 + ], + "type": "text", + "content": "49. Wang, W., Zhou, T., Porikli, F., Crandall, D., Van Gool, L.: A survey on deep learning technique for video segmentation. arXiv preprint arXiv:2107.01153 (2021)" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 133, + 172, + 480, + 205 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 172, + 480, + 205 + ], + "spans": [ + { + "bbox": [ + 133, + 172, + 480, + 205 + ], + "type": "text", + "content": "50. Xiao, H., Feng, J., Lin, G., Liu, Y., Zhang, M.: Monet: Deep motion exploitation for video object segmentation. In: Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition. pp. 1140-1148 (2018)" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 132, + 205, + 480, + 237 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 205, + 480, + 237 + ], + "spans": [ + { + "bbox": [ + 132, + 205, + 480, + 237 + ], + "type": "text", + "content": "51. Xie, H., Yao, H., Zhou, S., Zhang, S., Sun, W.: Efficient regional memory network for video object segmentation. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 1286-1295 (2021)" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 132, + 237, + 480, + 270 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 237, + 480, + 270 + ], + "spans": [ + { + "bbox": [ + 132, + 237, + 480, + 270 + ], + "type": "text", + "content": "52. Xu, N., Yang, L., Fan, Y., Yue, D., Liang, Y., Yang, J., Huang, T.: Youtube-vos: A large-scale video object segmentation benchmark. arXiv preprint arXiv:1809.03327 (2018)" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 132, + 270, + 480, + 304 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 270, + 480, + 304 + ], + "spans": [ + { + "bbox": [ + 132, + 270, + 480, + 304 + ], + "type": "text", + "content": "53. Xu, X., Wang, J., Li, X., Lu, Y.: Reliable propagation-correction modulation for video object segmentation. In: Proceedings of the AAAI Conference on Artificial Intelligence. pp. 2946-2954 (2022)" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 132, + 304, + 480, + 336 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 304, + 480, + 336 + ], + "spans": [ + { + "bbox": [ + 132, + 304, + 480, + 336 + ], + "type": "text", + "content": "54. Xu, X., Wang, J., Ming, X., Lu, Y.: Towards robust video object segmentation with adaptive object calibration. In: Proceedings of the 30th ACM International Conference on Multimedia. pp. 2709-2718 (2022)" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 132, + 336, + 480, + 369 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 336, + 480, + 369 + ], + "spans": [ + { + "bbox": [ + 132, + 336, + 480, + 369 + ], + "type": "text", + "content": "55. Xu, X., Zhang, T., Wang, S., Li, X., Chen, Y., Li, Y., Raj, B., Johnson-Roberson, M., Huang, X.: Customizable perturbation synthesis for robust slam benchmarking. arXiv preprint arXiv:2402.08125 (2024)" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 132, + 369, + 480, + 413 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 369, + 480, + 413 + ], + "spans": [ + { + "bbox": [ + 132, + 369, + 480, + 413 + ], + "type": "text", + "content": "56. Yan, S., Zhang, R., Guo, Z., Chen, W., Zhang, W., Li, H., Qiao, Y., Dong, H., He, Z., Gao, P.: Referred by multi-modality: A unified temporal transformer for video object segmentation. In: Proceedings of the AAAI Conference on Artificial Intelligence. vol. 38, pp. 6449-6457 (2024)" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 132, + 413, + 480, + 446 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 413, + 480, + 446 + ], + "spans": [ + { + "bbox": [ + 132, + 413, + 480, + 446 + ], + "type": "text", + "content": "57. Yang, Z., Wei, Y., Yang, Y.: Collaborative video object segmentation by foreground-background integration. In: European Conference on Computer Vision. pp. 332-348. Springer (2020)" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 132, + 446, + 480, + 479 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 446, + 480, + 479 + ], + "spans": [ + { + "bbox": [ + 132, + 446, + 480, + 479 + ], + "type": "text", + "content": "58. Yang, Z., Wei, Y., Yang, Y.: Associating objects with transformers for video object segmentation. Advances in Neural Information Processing Systems 34, 2491-2502 (2021)" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 132, + 479, + 480, + 512 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 479, + 480, + 512 + ], + "spans": [ + { + "bbox": [ + 132, + 479, + 480, + 512 + ], + "type": "text", + "content": "59. Yang, Z., Wei, Y., Yang, Y.: Collaborative video object segmentation by multiscale foreground-background integration. IEEE Transactions on Pattern Analysis and Machine Intelligence (2021)" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 132, + 512, + 480, + 533 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 512, + 480, + 533 + ], + "spans": [ + { + "bbox": [ + 132, + 512, + 480, + 533 + ], + "type": "text", + "content": "60. Yuan, M., Richardt, C.: 360 optical flow using tangent images. In: British Machine Vision Conference (BMVC) (2021)" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 132, + 533, + 480, + 555 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 533, + 480, + 555 + ], + "spans": [ + { + "bbox": [ + 132, + 533, + 480, + 555 + ], + "type": "text", + "content": "61. Zhang, R., Jiang, Z., Guo, Z., Yan, S., Pan, J., Dong, H., Gao, P., Li, H.: Personalize segment anything model with one shot. arXiv preprint arXiv:2305.03048 (2023)" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 132, + 555, + 480, + 588 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 555, + 480, + 588 + ], + "spans": [ + { + "bbox": [ + 132, + 555, + 480, + 588 + ], + "type": "text", + "content": "62. Zhang, Y., Zhang, L., Wang, K., Hamidouche, W., Deforges, O.: Shd360: A benchmark dataset for salient human detection in 360 videos. arXiv preprint arXiv:2105.11578 (2021)" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 132, + 588, + 480, + 621 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 588, + 480, + 621 + ], + "spans": [ + { + "bbox": [ + 132, + 588, + 480, + 621 + ], + "type": "text", + "content": "63. Zhang, Y., Wu, Z., Peng, H., Lin, S.: A transductive approach for video object segmentation. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 6949-6958 (2020)" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 132, + 621, + 480, + 643 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 621, + 480, + 643 + ], + "spans": [ + { + "bbox": [ + 132, + 621, + 480, + 643 + ], + "type": "text", + "content": "64. Zhang, Z., Xu, Y., Yu, J., Gao, S.: Saliency detection in 360 videos. In: Proceedings of the European conference on computer vision (ECCV). pp. 488-503 (2018)" + } + ] + } + ], + "index": 18 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "spans": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "type": "text", + "content": "18" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 166, + 91, + 206, + 100 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 166, + 91, + 206, + 100 + ], + "spans": [ + { + "bbox": [ + 166, + 91, + 206, + 100 + ], + "type": "text", + "content": "Yan et al." + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 17 + } + ], + "_backend": "vlm", + "_version_name": "2.6.4" +} \ No newline at end of file diff --git a/2024/PapMOT_ Exploring Adversarial Patch Attack against Multiple Object Tracking/e6b4ef0f-e80b-4da1-8f00-4210586c7699_content_list.json b/2024/PapMOT_ Exploring Adversarial Patch Attack against Multiple Object Tracking/e6b4ef0f-e80b-4da1-8f00-4210586c7699_content_list.json new file mode 100644 index 0000000000000000000000000000000000000000..3969ee08a0112f877ebc852d23155789df3b518f --- /dev/null +++ b/2024/PapMOT_ Exploring Adversarial Patch Attack against Multiple Object Tracking/e6b4ef0f-e80b-4da1-8f00-4210586c7699_content_list.json @@ -0,0 +1,1535 @@ +[ + { + "type": "text", + "text": "PapMOT: Exploring Adversarial Patch Attack against Multiple Object Tracking", + "text_level": 1, + "bbox": [ + 236, + 140, + 767, + 186 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Jiahuan Long $^{1,2}$ , Tingsong Jiang $^{2}$ , Wen Yao $^{2\\star}$ , Shuai Jia $^{1}$ , Weijia Zhang $^{1}$ , Weien Zhou $^{2}$ , Chao Ma $^{1\\star}$ , and Xiaojian Chen $^{2}$", + "bbox": [ + 233, + 212, + 767, + 243 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "1 MoE Key Lab of Artificial Intelligence, AI Institute, Shanghai Jiao Tong University 2 Defense Innovation Institute, Chinese Academy of Military Science jiahuanlong@sjtu.edu.cn, tingsong@pku.edu.cn, wendy0782@126.com, jiashuai@sjtu.edu.cn, weijia.zhang@sjtu.edu.cn, weienzhou@outlook.com, chaoma@sjtu.edu.cn, chenxiaoqian@nudt.edu.cn", + "bbox": [ + 215, + 253, + 785, + 325 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Abstract. Tracking multiple objects in a continuous video stream is crucial for many computer vision tasks. It involves detecting and associating objects with their respective identities across successive frames. Despite significant progress made in multiple object tracking (MOT), recent studies have revealed the vulnerability of existing MOT methods to adversarial attacks. Nevertheless, all of these attacks belong to digital attacks that inject pixel-level noise into input images, and are therefore ineffective in physical scenarios. To fill this gap, we propose PapMOT, which can generate physical adversarial patches against MOT for both digital and physical scenarios. Besides attacking the detection mechanism, PapMOT also optimizes a printable patch that can be detected as new targets to mislead the identity association process. Moreover, we introduce a patch enhancement strategy to further degrade the temporal consistency of tracking results across video frames, resulting in more aggressive attacks. We further develop new evaluation metrics to assess the robustness of MOT against such attacks. Extensive evaluations on multiple datasets demonstrate that our PapMOT can successfully attack various architectures of MOT trackers in digital scenarios. We also validate the effectiveness of PapMOT for physical attacks by deploying printed adversarial patches in the real world.", + "bbox": [ + 261, + 368, + 743, + 645 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Keywords: Physical adversarial patches $\\cdot$ Multiple object tracking $\\cdot$ Evaluation metrics", + "bbox": [ + 259, + 659, + 740, + 686 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "1 Introduction", + "text_level": 1, + "bbox": [ + 215, + 717, + 375, + 734 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Multiple object tracking (MOT) becomes increasingly widespread with numerous applications including autonomous driving [5] and intelligent surveillance systems [22]. Meanwhile, recent studies [17, 20] have shown that deep multiple object tracking algorithms are susceptible to attacks by adversarial examples.", + "bbox": [ + 212, + 752, + 787, + 814 + ], + "page_idx": 0 + }, + { + "type": "page_footnote", + "text": "* Corresponding authors.", + "bbox": [ + 217, + 824, + 392, + 839 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/fb3c2c34a465e45fbc9ca0a0f4fd3374fc6853478302effb70d2538ad8d55eaa.jpg", + "image_caption": [ + "(a) Digital attack", + "(b) Previous patch attack", + "(c) Our patch attack", + "Fig. 1: A comparison of various adversarial attack methods against MOT. (a) shows a digital attack [20] against MOT, where pixel-level noise is added to each frame and the identities of two people are switched within ten frames. (b) shows a patch attack [28] that fools only the detectors of MOT, generating new fake identities. (c) demonstrates that our patch attack simultaneously fools both the detectors and trackers of MOT, which creates new identities and alters the identities of persons as well." + ], + "image_footnote": [], + "bbox": [ + 222, + 143, + 782, + 246 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "These examples are designed to fool MOT models by adding imperceptible perturbations to video sequences. If an attack is successful, the identities of detected objects are no longer associated with their original identities. Existing adversarial attacks [17,20] against MOT have mainly focused on digital attacks, manipulating the pixels of the entire image to fool the trackers. Although these digital attacks can craft imperceptible perturbations, they are generally ineffective in physical scenarios. For instance, Figure 1(a) illustrates an adversarial instance [20] against MOT that cannot be physically implemented. Compared to digital attacks, launching attacks against MOT in the physical world is more challenging. It requires strong perturbations that can withstand diverse environmental conditions, such as illumination and geometric alterations. Also, physical attacks are inherently more challenging to detect and defend against, making them more significant threats to detectors and trackers. Therefore, developing a physical attack against MOT is of great importance to the disclosure and understanding of MOT systems' vulnerabilities in real-world applications.", + "bbox": [ + 212, + 395, + 787, + 623 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Physical adversarial patches are the most common form of physical attacks in the existing literature. They are widely utilized to gauge the reliability and robustness of computer vision systems, such as for face recognition [12,35] and object detection [9,13]. However, applying these methods to MOT systems presents new technical challenges. A major requirement of successful MOT systems is to maintain object identity consistency across frames. This requires the detector to identify targets in each frame and the tracker to associate these targets frame-by-frame. Simply applying adversarial patches generated by detector-based attacks to the MOT system does not sufficiently disrupt data associations across frames. As shown in Figure 1(b), applying the patch attack [28] can create new fake identities but fails to alter the identity of the real attack object with \"ID = 1\". Moreover, the existing evaluation metrics of MOT attacks are not enough to reflect the performance of patch attacks. Existing metrics neglect the importance of the detection phase, which is a critical component of MOT algorithms.", + "bbox": [ + 212, + 628, + 787, + 840 + ], + "page_idx": 1 + }, + { + "type": "page_number", + "text": "2", + "bbox": [ + 217, + 114, + 228, + 126 + ], + "page_idx": 1 + }, + { + "type": "header", + "text": "J. Long et al.", + "bbox": [ + 271, + 114, + 362, + 127 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Also, these metrics only focus on the effects on one or a few targets, ignoring how attacks on these targets could affect the whole system.", + "bbox": [ + 212, + 146, + 782, + 176 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "To address these problems, we propose PapMOT, a physical adversarial patch scheme against MOT systems. PapMOT generates printable adversarial patches that create new identities in the tracking phase, thereby disrupting the existing identity associations. As depicted in Figure 1(c), PapMOT deceives high-performance detectors into outputting false detection boxes (i.e., blue detection boxes), overlapping with true detection boxes. This overlap presents a significant threat to the tracking system, causing it to mistakenly alter the identities of targets in a few frames. To this end, we introduce three adversarial losses that respectively consider the size of bounding boxes, scores of detection boxes, and the smoothness of color transitions. Besides, we propose a patch enhancement strategy that crops targets over time to obtain an extra dataset. The joint training of the original training dataset and the clip dataset will increase the attack success rate on specific targets. To better evaluate the effectiveness of the attack, we construct maximum-attack-threshold evaluation metrics that integrate detection and tracking to comprehensively evaluate the performance of MOT attacks. We summarize our contributions as follows.", + "bbox": [ + 212, + 176, + 787, + 417 + ], + "page_idx": 2 + }, + { + "type": "list", + "sub_type": "text", + "list_items": [ + "- To the best of our knowledge, this is the first work to explore physical adversarial attack against MOT. The proposed PapMOT attack is able to produce printable patches that successfully disrupt the data association process in tracking.", + "- Upon delving into the association algorithm in MOT, we propose a novel bounding box restriction loss that generates a wrong detection box with a high IoU value, leading to incorrect identity assignment.", + "- We propose a new set of metrics to evaluate attacks against MOT that is more comprehensive than previous MOT attacks.", + "- We evaluate PapMOT in both digital and physical scenarios and demonstrate its favorable attack performance compared to state-of-the-art approaches across multiple datasets." + ], + "bbox": [ + 225, + 428, + 784, + 612 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "2 Related Work", + "text_level": 1, + "bbox": [ + 214, + 633, + 387, + 648 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "2.1 Multiple Object Tracking", + "text_level": 1, + "bbox": [ + 214, + 665, + 472, + 681 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Existing MOT methods can be classified into two main categories: Tracking-by-Detection [4, 22, 31, 38] and Joint Detection and Tracking [39, 40]. In the former paradigm, object detectors are used to detect targets in each frame, and data association algorithms are then employed to link the detection results across video frames, including ByteTrack [38], DeepSORT [31], and OCSort [4]. ByteTrack [38] categorizes all detections into high and low scores and associates them with predicted boxes using the Kalman filter. This improves tracking performance in cases of occlusion and motion blur problems. In the latter paradigm, object detection and data association tasks are integrated into a single framework. FairMOT [39] introduces an anchor-free detection head, which allows its", + "bbox": [ + 212, + 688, + 787, + 840 + ], + "page_idx": 2 + }, + { + "type": "header", + "text": "PapMOT: Exploring Adversarial Patch Attack against MOT", + "bbox": [ + 326, + 114, + 732, + 128 + ], + "page_idx": 2 + }, + { + "type": "page_number", + "text": "3", + "bbox": [ + 774, + 116, + 784, + 126 + ], + "page_idx": 2 + }, + { + "type": "table", + "img_path": "images/18cc9621b959504f7a1f1e6983cbc3d82e29cc6de0517a2622ec2322e4d9c03d.jpg", + "table_caption": [ + "Table 1: A comparison of representative adversarial attack methods." + ], + "table_footnote": [], + "table_body": "
MethodsNatPatch [12]ObjPatch [28]AdvTshirt [34]Trasw [20]Hijack [17]PapMOT (ours)
Advanced detectors✓
Detectors & trackers✓✓✓
Diverse datasets✓✓✓✓
Physical attack✓✓✓✓
New evaluation metrics✓
", + "bbox": [ + 218, + 183, + 784, + 297 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "network to predict the position and size of the targets and generate a unique ID feature to achieve data association. In this work, we use ByteTrack [38] and FairMOT [39] as representative trackers to validate the effectiveness and generalization of the proposed adversarial patch attack.", + "bbox": [ + 212, + 340, + 787, + 402 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "2.2 Adversarial Patch Attacks", + "text_level": 1, + "bbox": [ + 214, + 430, + 478, + 443 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Previous digital attacks [14-16, 33] are difficult to implement in physical scenarios as the perturbations are imperceptible for visual systems. To achieve physical attack, Thys et al. [28] generate physically feasible adversarial patches for deceiving person detectors. Expanding upon this approach, Hu et al. [13] develop AdvTexture by printing the patch on a large cloth and making different T-shirts, skirts, and dresses to realize physical adversarial attacks on pedestrian detection. These two schemes enhance the robustness of patches in the real world by involving patch transformation [1]. However, they mainly conduct attacks on lower-precision detectors such as YOLOv2 [24] and YOLOv3 [25]. For attacks on consecutive frames, Jia et al. [17] propose a tracker hijacking attack scheme, which allows adversarial samples to successfully add or remove objects within a single frame, potentially posing a safety hazard for autonomous vehicles. Lin et al. [20] present a Tracklet-Switch method, which introduces noise into certain frames, causing an ID-switch when pedestrians cross paths with each other. Yet, to our best knowledge, there has been no successful physical attack against MOT systems.", + "bbox": [ + 212, + 460, + 787, + 700 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "To fill this research gap, we propose PapMOT, the first physically realizable attack method against MOT, alongside a set of newly proposed evaluation metrics. Different from all existing works, PapMOT is capable of not only attacking more advanced detectors (i.e., YOLOX), but also crippling both the detector and the tracker of an MOT system simultaneously and across a diversity of datasets and scenarios (i.e., MOT15, MOT17, MOT20, and BDD100K). Our contributions are expected to bring new insights into existing research on the digital and physical robustness of MOT systems. A comparison of the features and contributions of our method against prior works is provided in Table 1.", + "bbox": [ + 212, + 704, + 787, + 840 + ], + "page_idx": 3 + }, + { + "type": "page_number", + "text": "4", + "bbox": [ + 217, + 114, + 228, + 126 + ], + "page_idx": 3 + }, + { + "type": "header", + "text": "J. Long et al.", + "bbox": [ + 271, + 114, + 362, + 128 + ], + "page_idx": 3 + }, + { + "type": "image", + "img_path": "images/de94d365eee5c5fe0ffb9dedfd66400ddbcb17d3b66df115f1356b457714e15e.jpg", + "image_caption": [ + "Fig. 2: The overall framework of PapMOT. (a) In the patch training phase, the objective is to optimize an adversarial patch to achieve effective attacks in both digital and physical domains. During this phase, Expectation over Transformation (EOT) is employed as data augmentation for patches. (b) Subsequently, the well-trained adversarial patch is deployed in the patch attack phase to disrupt object detection and data association in MOT." + ], + "image_footnote": [], + "bbox": [ + 217, + 145, + 785, + 344 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "3 Methodology", + "text_level": 1, + "bbox": [ + 215, + 467, + 380, + 484 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "3.1 Overall Framework of PapMOT", + "text_level": 1, + "bbox": [ + 215, + 500, + 524, + 515 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "An MOT system typically consists of a detection module and an association module. The association module generally employs IoU matching to maintain consistent IDs of objects across consecutive frames, which is at the core of high-performance MOT. Provided high IoU values of a detection box across adjacent frames, this box is likely to be assigned the same ID across these frames. As such, the core objective of our PapMOT is to train printable patches that can induce false detection boxes with a high IoU value with a correctly detected pedestrian box, leading to incorrect identity assignment of this very detection. The entire pipeline of the proposed PapMOT attack is depicted in Figure 2.", + "bbox": [ + 212, + 526, + 787, + 662 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Patch Training Phase. The patch training phase is illustrated in Figure 2(a). First, we randomly initialize a patch and employ a data enhancement strategy (i.e., EOT [1]) to enhance its robustness. EOT incorporates potential patch transformations such as rotation, brightness, blurring, and scaling that commonly exist in the physical world into patch training. Such transformations make our patch attacks more feasible in real-world situations. Next, we embed the EOT-enhanced patch into clean videos to obtain an adversarial dataset, and crop out pedestrian images based on their bounding boxes to produce an adversarial clip dataset. Finally, we combine these two adversarial datasets to optimize the physical adversarial patch using three losses. We use total variation loss $(\\mathcal{L}_{tv})$ to ensure patch smoothness and printability, bounding box restric", + "bbox": [ + 212, + 672, + 787, + 840 + ], + "page_idx": 4 + }, + { + "type": "header", + "text": "PapMOT: Exploring Adversarial Patch Attack against MOT", + "bbox": [ + 326, + 114, + 732, + 128 + ], + "page_idx": 4 + }, + { + "type": "page_number", + "text": "5", + "bbox": [ + 774, + 116, + 784, + 126 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "tion loss $(\\mathcal{L}_{bbr})$ to minimize the discrepancy between the detection boxes of the patch and the object, and average precision loss $(\\mathcal{L}_{ap})$ to decrease the detection confidence of targets with the patch.", + "bbox": [ + 212, + 146, + 787, + 191 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Patch Attack Phase. Our patch can deceive detectors into outputting false detection boxes, thereby disrupting the identity association in tracking. This specific attack process is diagrammed in Figure 2(b). The adversarial patch, already well-trained during the patch training phase, is embedded into a video to create an attack video. Next, the attack video is processed frame-by-frame throughYOLOX, which outputs detection boxes for each frame. To clarify the disruption of data association, we consider frames $t - 1$ and $t$ as illustrative examples. Assuming that there is a pedestrian with \"id=1\" in frame $t - 1$ , the Kalman filter [30] outputs a predicted detection box (i.e., blue dashed-line box) by inputting the current motion states of the pedestrian. In the next frame $t$ , two types of detection boxes appear, with one being a false detection box of the patch (i.e., green solid-line box) and the other being the detection box of the pedestrian (i.e., blue solid-line box). These two detection boxes do Intersection over Union (IoU) calculation with the predicted detection box from frame $t - 1$ . Following this, the Hungarian algorithm [18] uses these IoU values to assign identities to the detection boxes, leading the disruption of data association. (i.e., The false detection box of the patch in frame $t$ is matched with the identity of pedestrians from frame $t - 1$ ).", + "bbox": [ + 212, + 202, + 789, + 474 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "3.2 Adversarial Loss for Patch Attack", + "text_level": 1, + "bbox": [ + 214, + 497, + 542, + 512 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "For our loss formulation, we adopt the widely used average precision loss and total variation loss [12, 27, 29] to reduce the confidence score of detection while encouraging visual smoothness of the printable patches. However, only using these two losses cannot disrupt the identity association in tracking, as illustrated Figure 1(b). This motivates us to propose a novel bounding box restriction loss, which generates a wrong detection box with a high IoU value, leading to incorrect identity assignment and successfully disrupted data association.", + "bbox": [ + 212, + 523, + 787, + 630 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Bounding Box Restriction Loss $(\\mathcal{L}_{bbr})$ . To effectively disrupt identity association across frames, it is essential to ensure that the sizes of the target detection box and the patch detection box are similar. To this end, a bounding box restriction loss $\\mathcal{L}_{bbr}$ is formulated to restrict the height and width of the detection box, while minimizing the discrepancy between the detection boxes of the patch and the object. $\\mathcal{L}_{bbr}$ is defined as:", + "bbox": [ + 212, + 640, + 787, + 734 + ], + "page_idx": 5 + }, + { + "type": "equation", + "text": "\n$$\n\\mathcal {L} _ {b b r} = \\frac {1}{N} \\sum_ {i = 1} ^ {N} \\left(w _ {i} / I _ {w i d t h} + h _ {i} / I _ {h e i g h t}\\right) + \\left(1 - \\frac {1}{N} \\sum_ {i = 1} ^ {N} I o U \\left(P _ {i}, T _ {i}\\right)\\right), \\tag {1}\n$$\n", + "text_format": "latex", + "bbox": [ + 269, + 742, + 785, + 785 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "where $N$ is the total number of attacked objects; $w_{i}$ and $h_{i}$ are the width and height of predicted bounding boxes of the $i$ -th target; and $I_{width}$ and $I_{height}$ are the width and height of video frames that are used for normalization; detection", + "bbox": [ + 212, + 794, + 787, + 840 + ], + "page_idx": 5 + }, + { + "type": "page_number", + "text": "6", + "bbox": [ + 217, + 114, + 228, + 126 + ], + "page_idx": 5 + }, + { + "type": "header", + "text": "J. Long et al.", + "bbox": [ + 271, + 114, + 362, + 128 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "boxes of the patch and the target are denoted as $P_{i}$ and $T_{i}$ , respectively. It is worth noting that the first term seems to make $T_{i}$ have zero width and height, but we empirically find that it do not completely vanish and even create more new detection boxes due to the balance of IoU losses.", + "bbox": [ + 214, + 146, + 787, + 205 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Total Variation Loss $(\\mathcal{L}_{tv})$ . The aim of $\\mathcal{L}_{tv}$ is to ensure the generation of printable patches with smooth color transitions, which helps to reduce the presence of visual noise. $\\mathcal{L}_{tv}$ is formulated as:", + "bbox": [ + 214, + 218, + 785, + 263 + ], + "page_idx": 6 + }, + { + "type": "equation", + "text": "\n$$\n\\mathcal {L} _ {t v} = \\sum_ {i, j} \\sqrt {\\left(\\left(I _ {i , j} - I _ {i + 1 , j}\\right) ^ {2} + \\left(I _ {i , j} - I _ {i , j + 1}\\right) ^ {2}\\right)}, \\tag {2}\n$$\n", + "text_format": "latex", + "bbox": [ + 336, + 276, + 785, + 314 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "where $I_{i,j}$ denotes the pixel value at position $(i,j)$ in the patch.", + "bbox": [ + 214, + 325, + 673, + 343 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Average Precision Loss $(\\mathcal{L}_{ap})$ . It is essential to lower the confidence of the original detection box in order to have it swapped to the one induced by the patch. $\\mathcal{L}_{ap}$ does so exactly by reducing the average detection score of targets that have been attached to an adversarial patch. Given a set of $N$ specific detection boxes, where the detection score for the $i$ -th box is $s_i$ , $\\mathcal{L}_{ap}$ is calculated as the sum of detection scores divided by the total number of detection boxes as follows:", + "bbox": [ + 214, + 352, + 787, + 444 + ], + "page_idx": 6 + }, + { + "type": "equation", + "text": "\n$$\n\\mathcal {L} _ {a p} = \\frac {1}{N} \\sum_ {i = 1} ^ {N} s _ {i}. \\tag {3}\n$$\n", + "text_format": "latex", + "bbox": [ + 444, + 457, + 785, + 496 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Finally, the overall adversarial loss $L_{adv}$ is a sum of the above three loss terms weighted by hyperparameters $\\beta$ , $\\gamma$ , and $\\delta$ :", + "bbox": [ + 214, + 503, + 785, + 532 + ], + "page_idx": 6 + }, + { + "type": "equation", + "text": "\n$$\n\\mathcal {L} _ {a d v} = \\beta \\mathcal {L} _ {b b r} + \\gamma \\mathcal {L} _ {t v} + \\delta \\mathcal {L} _ {a p}, \\tag {4}\n$$\n", + "text_format": "latex", + "bbox": [ + 398, + 547, + 785, + 564 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "3.3 Evaluation Metrics for MOT Attack", + "text_level": 1, + "bbox": [ + 214, + 585, + 560, + 599 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Current attack methods against MOT [17, 20] have limitations in their evaluation since they utilize identity-based evaluation metrics. These metrics have two main drawbacks. Firstly, they are primarily designed for gauging tracking performance, thereby largely overlooking the quality of detection - the other core ingredient in successful MOT. Secondly, they tend to focus only on a few specific targets, disregarding a comprehensive evaluation of the entire MOT system.", + "bbox": [ + 212, + 612, + 785, + 703 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "As depicted in Figure 3(a), identity-based evaluation metrics [17,20] evaluate the identity of the victim in every frame, which considers tracking. However, existing MOT algorithms typically include both tracking and detection modules. Therefore, while evaluating attacks on MOT, it is essential to calculate the performance of both tracking and detection modules simultaneously. Furthermore, identity-based evaluation metrics usually adopt a strategy that aims to attack one or a few specific targets. This approach overlooks the cumulative impact of attacking these specific targets on the entire system. For instance, as shown in Figure 3(b), an attack on the victim (in the red detection box) might also affect", + "bbox": [ + 214, + 704, + 787, + 840 + ], + "page_idx": 6 + }, + { + "type": "header", + "text": "PapMOT: Exploring Adversarial Patch Attack against MOT", + "bbox": [ + 326, + 114, + 732, + 130 + ], + "page_idx": 6 + }, + { + "type": "page_number", + "text": "7", + "bbox": [ + 774, + 114, + 784, + 126 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/84ed199bf3cf9c7b6a41af4164bee7d98d3abe0959b8081e55ce9fc9f5f68071.jpg", + "image_caption": [ + "Fig. 3: A comparison of existing and our proposed evaluation metrics for MOT attacks." + ], + "image_footnote": [ + "(a) Previous evaluation metrics are identity-based and only consider ID association. \n(b) Our proposed metrics emphasize the attack evaluation of both detection and ID association." + ], + "bbox": [ + 274, + 142, + 727, + 281 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "the detection of other pedestrians (in yellow boxes), leading to switches of their identities or a decrease in detection accuracy.", + "bbox": [ + 212, + 380, + 784, + 410 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "To overcome these limitations, we construct a maximum-attack-threshold evaluation metric that integrates detection and tracking. This type of evaluation metric is based on the classic evaluation metrics for MOT [3]. It provides a comprehensive evaluation of the robustness of MOT systems.", + "bbox": [ + 212, + 411, + 785, + 472 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Tracking Attack Success Rate (TASR). To assess the impact of specific targets during attacks on the overall system, TASR determines the change in MOTA and IDF1 before and after attacks and then divides it by the percentage of bounding boxes that are attacked $(R_{bbox})$ . A high TASR indicates that the MOT system is highly susceptible to particular attacks, enabling an attacker to degrade its performance by compromising fewer targets. A low TASR implies that the attacker must compromise a greater number of targets to inflict considerable damage. The formula for TASR is: $TASR = (MOTA + IDF1)_{\\text{decline}} / 2R_{bbox}$ , where $R_{bbox}$ is the ratio of the number of bounding boxes under attack to the total bounding boxes.", + "bbox": [ + 212, + 483, + 787, + 636 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Identity Obfuscation Rate (IOR). To measure the effectiveness of adversarial attacks to cause identity obfuscation during data association in MOT, we define the Identity Obfuscation Rate as: $IOR = D_s / P_t$ , where $P_t = N / T$ , where $T$ represents the number of frames that are attacked consecutively, $N$ is the total number of frames of the video, $P_t$ denotes the maximum number of identity switches in the video, and $D_s$ is the number of identity switches throughout the video sequences.", + "bbox": [ + 212, + 646, + 787, + 752 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Specific Target Attack Success Rate (STASR). To evaluate attacks on specific targets in terms of both false positive rate (FP) and the number of identity switches, STASR focuses not only on how the attack interferes with the detection of the targets but also on how the attack disrupts the tracking of the targets. STASR is defined as: $STASR = (FP + ID_{switch})_{increase} / A_{max}$ , where", + "bbox": [ + 212, + 763, + 787, + 842 + ], + "page_idx": 7 + }, + { + "type": "page_number", + "text": "8", + "bbox": [ + 217, + 114, + 228, + 126 + ], + "page_idx": 7 + }, + { + "type": "header", + "text": "J. Long et al.", + "bbox": [ + 271, + 114, + 362, + 127 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "$A_{max} = P_t + P_n$ , $P_t$ represents the maximum number of identity switches in the video, and $P_n$ denotes the number of detection boxes that are attacked.", + "bbox": [ + 215, + 146, + 784, + 176 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "4 Experiments", + "text_level": 1, + "bbox": [ + 215, + 200, + 374, + 218 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "In this section, we perform extensive empirical analysis of the proposed PapMOT attack on MOT in both digital and physical environments. We also conduct ablation studies on the adversarial loss and variants of adversarial patches.", + "bbox": [ + 215, + 233, + 784, + 279 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "4.1 Experimental Setup", + "text_level": 1, + "bbox": [ + 215, + 301, + 426, + 316 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "Datasets and Metrics. We evaluate our attack on four popular MOT benchmarks: MOT15 [19], MOT17 [23], MOT20 [7], and BDD100K [36]. The target objects of the first three datasets are pedestrian-only, while BDD100K further includes vehicle objects. We adopt our proposed evaluation metrics, namely TASR, SATSR, and IOR, to assess the performance of our PapMOT against state-of-the-art methods on the MOT task. More comparisons are found in Appendix.", + "bbox": [ + 215, + 339, + 784, + 429 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "Baselines. Existing MOT attack methods can be divided into three categories: 1) methods that attack detectors in the MOT system by injecting pixel-level noise, e.g., RAnoise [21] and TOG [6]. 2) methods that attack detectors in the MOT system by placing optimized patches, e.g., AdvTexture [13] and ObjPatch [28]. 3) methods that attack both detectors and trackers in the MOT system by injecting pixel-level noise, e.g., Hijack [17] and Trasw [20]. Our PapMOT introduces a physically realizable adversarial patch into the MOT system to attack both detectors and trackers.", + "bbox": [ + 215, + 440, + 784, + 560 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "Victim Models. We choose two representative MOT methods as victim models in our experiments: ByteTrack [38] that represents the tracking-by-detection paradigm and FairMOT [39] that represents the joint detection and tracking paradigm. Besides, ByteTrack is equipped with YOLOX [11], a more advanced version of theYOLO detector family trained on CrowdHuman [26], MOT17 [23], CityPerson [37], and ETHZ [10] datasets. The pre-trained model of FairMOT is provided by PaddleDetection [2], and is trained on Caltech Pedestrian [8], CityPersons [37], CUHK-SYSU [32], PRW [41], ETHZ [10], and MOT17 [23] datasets.", + "bbox": [ + 215, + 571, + 784, + 707 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "Implementation Details. In our digital experiments, to strike a balance between the stealthiness and effectiveness of attacks, we set the noise standard deviation of RAnoise to 10, and limit the maximum disturbance range in TOG to 8/255. The patch sizes for both AdvTexture and ObjPatch are set to one-third of the targeted detection box's size. We define an object as \"attacked\" if it overlaps with another object with an IOU greater than 0.2. In our physical experiments, we print the patches on foam boards with a fixed width and height of $50\\mathrm{cm}$ . The size ratio between the patch and an adult who is $1.6\\mathrm{m}$ tall is kept at", + "bbox": [ + 215, + 719, + 784, + 839 + ], + "page_idx": 8 + }, + { + "type": "header", + "text": "PapMOT: Exploring Adversarial Patch Attack against MOT", + "bbox": [ + 326, + 114, + 732, + 128 + ], + "page_idx": 8 + }, + { + "type": "page_number", + "text": "9", + "bbox": [ + 774, + 116, + 784, + 126 + ], + "page_idx": 8 + }, + { + "type": "table", + "img_path": "images/31c514562d68e6a946822c75631390b5b2161e9504e0a91fcb61593613a98f57.jpg", + "table_caption": [ + "Table 2: Comparison with other representative adversarial attack methods on the MOT15 and MOT17 datasets." + ], + "table_footnote": [], + "table_body": "
DatasetMOT15MOT17
VictimFairMOTByteTrackFairMOTByteTrack
MetricsTASRSATSRIORTASRSATSRIORTASRSATSRIORTASRSATSRIOR
RAnoise [21]0.70.6-0.60.3-0.30.5--0.2-
TOG [6]13.77.04.111.46.33.89.14.82.54.73.91.2
AdvTexture [13]36.435.64.638.143.75.845.432.15.142.247.76.0
ObjPatch [28]45.240.66.343.759.38.261.349.56.261.476.77.4
Hijack [17]14.218.539.311.415.929.010.019.226.58.415.425.4
Trasw [20]40.645.866.231.439.255.422.128.944.821.026.842.3
Ours76.485.062.385.379.265.976.893.046.294.689.679.5
", + "bbox": [ + 220, + 184, + 772, + 343 + ], + "page_idx": 9 + }, + { + "type": "text", + "text": "around $30\\%$ , ensuring that essential areas such as the face are not obstructed to maintain physical feasibility. For hyperparameters in the overall adversarial loss (i.e., Eq. 4), we set $\\beta = 1$ , $\\gamma = 2.5$ , and $\\delta = 2$ . Datasets used in our patch generation consist of ETHZ, MOT17, and MOT17 clip datasets.", + "bbox": [ + 212, + 369, + 787, + 433 + ], + "page_idx": 9 + }, + { + "type": "text", + "text": "4.2 Digital Domain Performance", + "text_level": 1, + "bbox": [ + 215, + 452, + 501, + 469 + ], + "page_idx": 9 + }, + { + "type": "text", + "text": "Table 2 and Table 3 compare the performance of various attacks on the MOT15, MOT17, MOT20, and BDD100K datasets. As can be seen, PapMOT achieves the highest attack success rate across all four datasets with a $50\\%$ bounding box attack, manifesting its superiority over existing methods (i.e., RAnoise [21], TOG [6], ADVTexure [13], ObjPatch [28], Hijack [17], Trasw [20]) against MOT in both pedestrian (i.e., MOT15, MOT17, and MOT20) and autonomous driving scenarios (i.e., BDD100K). We notice that the IOR of PapMOT is marginally below that of Trasw on the MOT15 and MOT20 datasets, which is possibly due to Trasw's sole emphasis on attacking data association. Nonetheless, PapMOT surpasses Trasw by a huge margin in terms of TASR and SATSR. Overall, Trasw as a digital method is hard to be implemented in the physical world, and its attack is strictly triggered only when an object overlaps with the attack object. Thus, PapMOT outperforms Trasw in terms of functionality and overall attack performance.", + "bbox": [ + 212, + 477, + 787, + 688 + ], + "page_idx": 9 + }, + { + "type": "text", + "text": "The superior performance of PapMOT is most evident on the MOT17 dataset, where it consistently surpasses all competing methods across all three metrics when applied against both FairMOT and ByteTrack. The success of PapMOT is largely due to incorporating the MOT17 clip dataset and MOT17 into its patch training dataset (more details about the clip dataset are found in Appendix). It is noteworthy that TASR and SATSR both incorporate the evaluation of detection box performance, whereas IOR predominantly assesses identity association across frames. As such, it is observed that TASR and SATSR are typically higher for patch-based attacks and digital attacks compared to IOR (e.g., when attacking ByteTrack in MOT17, the IOR of AdvTexture is 6.0, while TASR and", + "bbox": [ + 212, + 688, + 789, + 840 + ], + "page_idx": 9 + }, + { + "type": "page_number", + "text": "10", + "bbox": [ + 217, + 114, + 235, + 126 + ], + "page_idx": 9 + }, + { + "type": "header", + "text": "J. Long et al.", + "bbox": [ + 271, + 114, + 361, + 128 + ], + "page_idx": 9 + }, + { + "type": "table", + "img_path": "images/7223f4e49f05891bd7a3a762545c2fad5f6214401cbb6763ffe5f0bed3fc64c8.jpg", + "table_caption": [ + "Table 3: Comparison with other representative adversarial attack methods on the MOT20 and BDD100K datasets." + ], + "table_footnote": [], + "table_body": "
DatasetMOT20BDD100K
VictimFairMOTByteTrackFairMOTByteTrack
MetricsTASRSATSRIORTASRSATSRIORTASRSATSRIORTASRSATSRIOR
RAnoise [21]2.60.90.44.11.50.42.20.7-1.90.7-
TOG [6]8.62.50.67.62.30.618.16.72.413.54.82.2
AdvTexture [13]23.518.52.932.026.73.221.112.64.418.319.04.8
ObjPatch [28]42.321.44.441.327.56.826.319.45.220.523.37.6
Hijack [17]17.918.436.815.934.727.07.19.810.57.48.213.5
Trasw [20]31.233.567.026.628.564.212.217.423.311.013.119.3
Ours61.949.556.658.541.653.946.454.734.163.461.238.8
", + "bbox": [ + 220, + 184, + 772, + 344 + ], + "page_idx": 10 + }, + { + "type": "text", + "text": "SATSR in ADvTexture are 42.2 and 47.7, respectively). This observation can be accounted for by the fact that these attacks primarily target detectors, which significantly affect the performance of detection boxes, but minimally impact identity association across frames.", + "bbox": [ + 212, + 372, + 784, + 434 + ], + "page_idx": 10 + }, + { + "type": "text", + "text": "4.3 Physical Domain Performance", + "text_level": 1, + "bbox": [ + 215, + 455, + 509, + 470 + ], + "page_idx": 10 + }, + { + "type": "text", + "text": "While our patches have shown outstanding performance in the digital domain, translating this success to the physical world is challenging due to the ever-changing of physical environments. In this section, we qualitatively evaluate the effectiveness of our adversarial patches, and further probe their limits under diverse physical environments.", + "bbox": [ + 212, + 481, + 784, + 556 + ], + "page_idx": 10 + }, + { + "type": "text", + "text": "Physical Validation. We train three patches with different datasets (i.e., ETHZ, MOT17, MOT20) for FairMOT and evaluate them against the ByteTrack detector. Figure 4 showcases the demonstration of these adversarial patches for physical-world attacks under diverse scenes (refer to Supplementary Materials for a video demo).", + "bbox": [ + 212, + 566, + 784, + 642 + ], + "page_idx": 10 + }, + { + "type": "text", + "text": "In Figure 4 (a), an adversarial patch is attached to a pedestrian, targeting a single object. In the absence of our attack, the target's ID detected by the tracker is consistently 1, and the detected bounding box is accurate. When the attack is imposed, the detector produces an additional false detection box (in blue), and the accuracy of the initially accurate detection box deteriorates significantly. Moreover, the tracker is quickly deprived of the ability to identity associations across frames, and an ID switch takes place within 30 frames. These outcomes validate the effectiveness of our attack in single-object scenarios.", + "bbox": [ + 212, + 643, + 784, + 763 + ], + "page_idx": 10 + }, + { + "type": "text", + "text": "Figure 4 (b) illustrates a multi-object scenario, where the adversarial patch is attached to one of the two pedestrians in the scene. When the pedestrian in white with the patch (denoted as \"target A\") passes by in front of the other subject in black (\"target B\"), their detected identities are successfully maintained in the absence of the attack. When attacked, however, the tracker mistakenly", + "bbox": [ + 212, + 763, + 785, + 839 + ], + "page_idx": 10 + }, + { + "type": "header", + "text": "PapMOT: Exploring Adversarial Patch Attack against MOT", + "bbox": [ + 326, + 114, + 730, + 128 + ], + "page_idx": 10 + }, + { + "type": "page_number", + "text": "11", + "bbox": [ + 767, + 116, + 782, + 126 + ], + "page_idx": 10 + }, + { + "type": "image", + "img_path": "images/95c589e9cf0162945ddab5276d582e17531acbe28c5ae6f4bb626f69e7fcaec6.jpg", + "image_caption": [ + "Fig. 4: Black-box transfer attack of different PapMOT patches in the real world. It illustrates the effective attack for the scenarios of a single pedestrian (a), two pedestrians overlap (b), and a pedestrian walking past from behind a fixed patch (c)." + ], + "image_footnote": [], + "bbox": [ + 228, + 143, + 777, + 329 + ], + "page_idx": 11 + }, + { + "type": "text", + "text": "predicts two identities (i.e., 1 and 3) for target A. Also, an ID switch occurs and the accuracy of both boxes drops as target A passes by. This example shows our proposed attack works in multiple-object scenarios and its effect applies beyond the object with the attached patch.", + "bbox": [ + 212, + 416, + 784, + 477 + ], + "page_idx": 11 + }, + { + "type": "text", + "text": "Figure 4 (c) shows the scenario where a patch is kept stationary and is misidentified as a pedestrian by the tracker. When attacked, a pedestrian passing behind it immediately causes an ID switch to occur (i.e., ID changes from 2 to 1). This example suggests that our adversarial patch is also suitable for static placement and is capable of affecting pedestrians as they pass by.", + "bbox": [ + 212, + 478, + 785, + 554 + ], + "page_idx": 11 + }, + { + "type": "text", + "text": "Attack Boundary. We examine three key physical conditions that may affect the effectiveness of patch attacks: (a) illumination, (b) distance, and (c) angle, shown in Fig. 5. For illumination, we sweep over the entire spectrum from total darkness to total brightness by adjusting the ISO sensitivity of the camera. We divide the illumination levels into 10 segments between 0 and 1. An illumination value of 0.5 denotes the normal exposure under overcast outdoor conditions with the ISO set to 200; 0 and 1 correspond to full darkness and full brightness, respectively. For distance, we vary the range between the patch and the camera from 0.5 to 10 meters to obtain a better view. For angle, we control the angle of the patch relative to the camera by adjusting the deflection angle of the person and using 5-degree intervals to present the final attack results. The quantitative evaluation results of the attack boundary effectiveness for three different patches are presented in Table 4.", + "bbox": [ + 212, + 566, + 787, + 762 + ], + "page_idx": 11 + }, + { + "type": "text", + "text": "Based on the experimental results, several important conclusions can be drawn. Firstly, different patches can achieve varying attack effects depending on the physical conditions. Secondly, the distance between the patches and the camera must be maintained within a specific range (0.8m-4.1m) to achieve an effective attack. Thirdly, patch attacks work better in relatively dark environ", + "bbox": [ + 212, + 763, + 787, + 839 + ], + "page_idx": 11 + }, + { + "type": "page_number", + "text": "12", + "bbox": [ + 217, + 114, + 235, + 126 + ], + "page_idx": 11 + }, + { + "type": "header", + "text": "J. Long et al.", + "bbox": [ + 271, + 114, + 361, + 127 + ], + "page_idx": 11 + }, + { + "type": "image", + "img_path": "images/2f5cc22b3002e2ef1c281fd3995402af56d8a40580f62298d8b61b6d787c4efb.jpg", + "image_caption": [ + "Fig. 5: Studies on patch effectiveness at varying illuminations, distances, and angles." + ], + "image_footnote": [], + "bbox": [ + 225, + 145, + 777, + 256 + ], + "page_idx": 12 + }, + { + "type": "table", + "img_path": "images/9ae93028cd95e837ee2afc069f40eaacde18aa613f3d486f2db5200e6c517d35.jpg", + "table_caption": [ + "Table 4: Studies on the effectiveness range of three adversarial patches." + ], + "table_footnote": [], + "table_body": "
DistanceIlluminationPatch SizeAngle
Patch 10.8m-3.1m0.4-0.740cm-100cm0°-25°
Patch 21.2m-4.1m0.5-0.840cm-100cm0°-15°
Patch 31.1m-3.6m0.4-0.840cm-100cm0°-30°
", + "bbox": [ + 313, + 323, + 683, + 395 + ], + "page_idx": 12 + }, + { + "type": "text", + "text": "ments, resulting in more false detection boxes. Moreover, a larger patch size has a stronger attack ability. However, to ensure the stealthiness and effectiveness of the patch, the patch size in PapMOT is set to approximately $30\\%$ of the target size. Lastly, to ensure effective attack outcomes, the angle of the patch relative to the camera generally needs to be limited to a small range (i.e., within 15 degrees). It may be considered to add rotation transformation in the patch training to enhance the robustness of the patches against rotation.", + "bbox": [ + 212, + 424, + 784, + 529 + ], + "page_idx": 12 + }, + { + "type": "text", + "text": "4.4 Ablation Studies", + "text_level": 1, + "bbox": [ + 215, + 551, + 400, + 565 + ], + "page_idx": 12 + }, + { + "type": "text", + "text": "We analyze the effectiveness of each adversarial loss term in our proposed training objective. We also conduct real-world experiments to evaluate the robustness against unoptimized patches, including real images and uniform color patches.", + "bbox": [ + 212, + 577, + 784, + 622 + ], + "page_idx": 12 + }, + { + "type": "text", + "text": "Loss Functions. Table 5 shows how different adversarial losses affect the performance of our proposed on against MOT17. The results indicate that PapMOT without $\\mathcal{L}_{bbr}$ performs significantly worse than PapMOT without $\\mathcal{L}_{ap}$ in terms of IOR, scoring 7.2 and 52.6, respectively, when attacking ByteTrack. This suggests that the $\\mathcal{L}_{bbr}$ loss is critical in disrupting the consistency of data association. Conversely, when attacking FairMOT, PapMOT without $\\mathcal{L}_{ap}$ scores substantially lower in TASR and SATSR than PapMOT without $\\mathcal{L}_{bbr}$ , with scores of 11.0 and 13.1, respectively, compared to 50.5 and 43.3. This underscores the significant impact of the $\\mathcal{L}_{ap}$ loss on detector performance.", + "bbox": [ + 212, + 632, + 787, + 768 + ], + "page_idx": 12 + }, + { + "type": "text", + "text": "Various Patches. We evaluate the effectiveness of attacking the MOT system using unoptimized patches (e.g., full-white patches, full-black patches, and real image patches). As shown in Figure 6, although the target is partially occluded by (a) the full-white patch or (b) the full-black patch, it is tracked steadily", + "bbox": [ + 212, + 779, + 787, + 840 + ], + "page_idx": 12 + }, + { + "type": "header", + "text": "PapMOT: Exploring Adversarial Patch Attack against MOT", + "bbox": [ + 326, + 114, + 732, + 128 + ], + "page_idx": 12 + }, + { + "type": "page_number", + "text": "13", + "bbox": [ + 767, + 114, + 784, + 126 + ], + "page_idx": 12 + }, + { + "type": "image", + "img_path": "images/24c0579cb70f138ff9246131c7fcdef2593151fa31c4c96e193d2f5c619bec25.jpg", + "image_caption": [ + "(a) Full-white patch" + ], + "image_footnote": [], + "bbox": [ + 236, + 142, + 411, + 231 + ], + "page_idx": 13 + }, + { + "type": "image", + "img_path": "images/07bcf3370924edbbfb8cf79914ed27088acbe16670c23e7bbd4db125b06eeba6.jpg", + "image_caption": [ + "(b) Full-black patch", + "Fig. 6: The attack testing using real image, full-white, and full-black patches on the MOT system. These patches have the same size of $50 \\times 50\\mathrm{cm}$ as ours." + ], + "image_footnote": [], + "bbox": [ + 416, + 143, + 586, + 231 + ], + "page_idx": 13 + }, + { + "type": "image", + "img_path": "images/2b9a8610d269ad49cc294f3110cc21bf1c8810ff176494313f689dee3cc0e81c.jpg", + "image_caption": [ + "(c) Real image patch" + ], + "image_footnote": [], + "bbox": [ + 591, + 143, + 767, + 231 + ], + "page_idx": 13 + }, + { + "type": "table", + "img_path": "images/fff639955aab891474a854513ed557c5c1c0a3e47fca16f7869b82d32d57a602.jpg", + "table_caption": [ + "Table 5: The effectiveness analysis of the adversarial losses." + ], + "table_footnote": [], + "table_body": "
VictimAttackerMetrics
TASRSATSRIOR
ByteTrack [38]PapMOT w/o Lbbr61.364.77.2
PapMOT w/o Lap21.926.852.6
PapMOT94.689.679.5
FairMOT [39]PapMOT w/o Lbbr50.543.35.6
PapMOT w/o Lap11.013.129.3
PapMOT76.893.046.2
", + "bbox": [ + 295, + 330, + 705, + 470 + ], + "page_idx": 13 + }, + { + "type": "text", + "text": "and the ID remains unchanged throughout the whole video sequence. Similarly, although the target is partially occluded by (c) the real-image patch, it is tracked steadily and the ID is maintained at 1 throughout the entire video sequence. The experiments indicate that it is difficult to attack MOT systems with adversarial patches that are not optimized.", + "bbox": [ + 212, + 502, + 787, + 579 + ], + "page_idx": 13 + }, + { + "type": "text", + "text": "5 Conclusion", + "text_level": 1, + "bbox": [ + 215, + 606, + 359, + 622 + ], + "page_idx": 13 + }, + { + "type": "text", + "text": "In this paper, we introduce a novel physical adversarial attack method named PapMOT against multiple object tracking systems. Unlike all previous attacks that inject noise into input video frames in the digital domain, our approach optimizes printable adversarial patches that effectively disrupt the data association of MOT trackers across frames in the actual physical world. We demonstrate the feasibility and effectiveness of PapMOT in both digital and physical scenarios through extensive experiments on multiple datasets. In future work, we plan to investigate the creation of more natural-looking adversarial patches, such as transforming these patches into clothing to deceive the MOT system. Additionally, we will expand physical adversarial MOT attacks from pedestrian scenarios to autonomous driving scenarios.", + "bbox": [ + 212, + 642, + 787, + 808 + ], + "page_idx": 13 + }, + { + "type": "text", + "text": "Acknowledgements. This work was supported in part by NSFC (62322113, 62376156).", + "bbox": [ + 214, + 809, + 785, + 840 + ], + "page_idx": 13 + }, + { + "type": "page_number", + "text": "14", + "bbox": [ + 217, + 114, + 235, + 126 + ], + "page_idx": 13 + }, + { + "type": "header", + "text": "J. Long et al.", + "bbox": [ + 271, + 114, + 362, + 128 + ], + "page_idx": 13 + }, + { + "type": "text", + "text": "References", + "text_level": 1, + "bbox": [ + 215, + 143, + 321, + 159 + ], + "page_idx": 14 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "1. Athalye, A., Engstrom, L., Ilyas, A., Kwok, K.: Synthesizing robust adversarial examples. In: International conference on machine learning (2018) 4, 5", + "2. Authors, P.: Paddledetection, object detection and instance segmentation toolkit based on paddlepaddle. https://github.com/PaddlePaddle/PaddleDetection (2019) 9", + "3. Bernardin, K., Stiefelhagen, R.: Evaluating multiple object tracking performance: the clear mot metrics. EURASIP Journal on Image and Video Processing (2008) 8", + "4. Cao, J., Weng, X., Khirodkar, R., Pang, J., Kitani, K.: Observation-centric sort: Rethinking sort for robust multi-object tracking. arXiv preprint arXiv:2203.14360 (2022) 3", + "5. Chiu, H.K., Li, J., Ambrus, R., Bohg, J.: Probabilistic 3d multi-modal, multi-object tracking for autonomous driving. In: IEEE International Conference on Robotics and Automation (2021) 1", + "6. Chow, K.H., Liu, L., Gursoy, M.E., Truex, S., Wei, W., Wu, Y.: Understanding object detection through an adversarial lens. In: European Symposium on Research in Computer Security (2020) 9, 10, 11", + "7. Dendorfer, P., Rezatofighi, H., Milan, A., Shi, J., Cremers, D., Reid, I., Roth, S., Schindler, K., Leal-Taixe, L.: Mot20: A benchmark for multi object tracking in crowded scenes. arXiv preprint arXiv:2003.09003 (2020) 9", + "8. Dollar, P., Wojek, C., Schiele, B., Perona, P.: Pedestrian detection: A benchmark. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (2009) 9", + "9. Du, A., Chen, B., Chin, T.J., Law, Y.W., Sasdelli, M., Rajasegaran, R., Campbell, D.: Physical adversarial attacks on an aerial imagery object detector. In: Proceedings of the IEEE/CVF Winter Conference on Applications of Computer Vision (2022) 2", + "10. Ess, A., Leibe, B., Schindler, K., van Gool, L.: A mobile vision system for robust multi-person tracking. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (2008) 9", + "1. Ge, Z., Liu, S., Wang, F., Li, Z., Sun, J.: Yolox: Exceeding yolo series in 2021. arXiv preprint arXiv:2107.08430 (2021) 9", + "2. Hu, Y.C.T., Kung, B.H., Tan, D.S., Chen, J.C., Hua, K.L., Cheng, W.H.: Naturalistic physical adversarial patch for object detectors. In: Proceedings of the IEEE/CVF International Conference on Computer Vision (2021) 2, 4, 6", + "3. Hu, Z., Huang, S., Zhu, X., Sun, F., Zhang, B., Hu, X.: Adversarial texture for fooling person detectors in the physical world. In: Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition (2022) 2, 4, 9, 10, 11", + "4. Jia, S., Ma, C., Song, Y., Yang, X.: Robust tracking against adversarial attacks. In: European Conference on Computer Vision (2020) 4", + "5. Jia, S., Ma, C., Yao, T., Yin, B., Ding, S., Yang, X.: Exploring frequency adversarial attacks for face forgery detection. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (2022) 4", + "6. Jia, S., Song, Y., Ma, C., Yang, X.: Iou attack: Towards temporally coherent blackbox adversarial attack for visual object tracking. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (2021) 4", + "7. Jia, Y., Lu, Y., Shen, J., Chen, Q.A., Zhong, Z., Wei, T.: Fooling detection alone is not enough: Adversarial attack against multiple object tracking. In: International Conference on Learning Representations (2020) 1, 2, 4, 7, 9, 10, 11" + ], + "bbox": [ + 225, + 172, + 784, + 839 + ], + "page_idx": 14 + }, + { + "type": "header", + "text": "PapMOT: Exploring Adversarial Patch Attack against MOT", + "bbox": [ + 326, + 114, + 730, + 128 + ], + "page_idx": 14 + }, + { + "type": "page_number", + "text": "15", + "bbox": [ + 767, + 116, + 784, + 126 + ], + "page_idx": 14 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "18. Kuhn, H.W.: The hungarian method for the assignment problem. Naval research logistics quarterly 2(1-2), 83-97 (1955) 6", + "19. Leal-Taixe, L., Milan, A., Reid, I., Roth, S., Schindler, K.: Motchallenge 2015: Towards a benchmark for multi-target tracking. arXiv preprint arXiv:1504.01942 (2015) 9", + "20. Lin, D., Chen, Q., Zhou, C., He, K.: Trasw: Tracklet-switch adversarial attacks against multi-object tracking. arXiv preprint arXiv:2111.08954 (2023) 1, 2, 4, 7, 9, 10, 11", + "21. Luisier, F., Blu, T., Unser, M.: Image denoising in mixed poisson-gaussian noise. IEEE Transactions on image processing (2010) 9, 10, 11", + "22. Luo, W., Xing, J., Milan, A., Zhang, X., Liu, W., Kim, T.K.: Multiple object tracking: A literature review. Artificial intelligence (2021) 1, 3", + "23. Milan, A., Leal-Taixe, L., Reid, I., Roth, S., Schindler, K.: Mot16: A benchmark for multi-object tracking. arXiv preprint arXiv:1603.00831 (2016) 9", + "24. Redmon, J., Farhadi, A.: Yolo9000: Better, faster, stronger. In: IEEE Conference on Computer Vision and Pattern Recognition (2017) 4", + "25. Redmon, J., Farhadi, A.: Yolov3: An incremental improvement. arXiv preprint arXiv:1804.02767 (2018) 4", + "26. Shao, S., Zhao, Z., Li, B., Xiao, T., Yu, G., Zhang, X., Sun, J.: Crowdhuman: A benchmark for detecting human in a crowd. arXiv preprint arXiv:1805.00123 (2018) 9", + "27. Sharif, M., Bhagavatula, S., Bauer, L., Reiter, M.K.: Accessorize to a crime. In: Proceedings of the 2016 ACM SIGSAC Conference on Computer and Communications Security (2016) 6", + "28. Thys, S., Van Ranst, W., Goedemé, T.: Fooling automated surveillance cameras: adversarial patches to attack person detection. In: Proceedings of the IEEE/CVF conference on computer vision and pattern recognition workshops (2019) 2, 4, 9, 10, 11", + "29. Wang, D., Jiang, T., Sun, J., Zhou, W., Gong, Z., Zhang, X., Yao, W., Chen, X.: Fca: Learning a 3d full-coverage vehicle camouflage for multi-view physical adversarial attack. In: Proceedings of the AAAI conference on artificial intelligence (2022) 6", + "30. Welch, G., Bishop, G., et al.: An introduction to the kalman filter (1995) 6", + "31. Wojke, N., Bewley, A., Paulus, D.: Simple online and realtime tracking with a deep association metric. In: IEEE International Conference on Image Processing (2017) 3", + "32. Xiao, T., Li, S., Wang, B., Lin, L., Wang, X.: Joint detection and identification feature learning for person search. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (2017) 9", + "33. Xu, H., Ma, Y., Liu, H.C., Deb, D., Liu, H., Tang, J.L., Jain, A.K.: Adversarial attacks and defenses in images, graphs and text: A review. International Journal of Automation and Computing (2020) 4", + "34. Xu, K., Zhang, G., Liu, S., Fan, Q., Sun, M., Chen, H., Chen, P.Y., Wang, Y., Lin, X.: Adversarial t-shirt! evading person detectors in a physical world. In: European Conference on Computer Vision (2020) 4", + "35. Yang, X., Wei, F., Zhang, H., Zhu, J.: Design and interpretation of universal adversarial patches in face detection. In: European Conference on Computer Vision (2020) 2", + "36. Yu, F., Chen, H., Wang, X., Xian, W., Chen, Y., Liu, F., Madhavan, V., Darrell, T.: Bdd100k: A diverse driving dataset for heterogeneous multitask learning. In: IEEE/CVF Conference on Computer Vision and Pattern Recognition (2020) 9" + ], + "bbox": [ + 215, + 146, + 784, + 840 + ], + "page_idx": 15 + }, + { + "type": "page_number", + "text": "16", + "bbox": [ + 217, + 114, + 235, + 126 + ], + "page_idx": 15 + }, + { + "type": "header", + "text": "J. Long et al.", + "bbox": [ + 271, + 114, + 361, + 127 + ], + "page_idx": 15 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "37. Zhang, S., Benenson, R., Schiele, B.: Citypersons: A diverse dataset for pedestrian detection. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (2017) 9", + "38. Zhang, Y., Sun, P., Jiang, Y., Yu, D., Weng, F., Yuan, Z., Luo, P., Liu, W., Wang, X.: Bytetrack: Multi-object tracking by associating every detection box. In: European Conference on Computer Vision (2022) 3, 4, 9, 14", + "39. Zhang, Y., Wang, C., Wang, X., Zeng, W., Liu, W.: Fairmot: On the fairness of detection and re-identification in multiple object tracking. International Journal of Computer Vision (2021) 3, 4, 9, 14", + "40. Zhao, Z., Wu, Z., Zhuang, Y., Li, B., Jia, J.: Tracking objects as pixel-wise distributions. In: European Conference on Computer Vision (2022) 3", + "41. Zheng, L., Zhang, H., Sun, S., Chandraker, M., Yang, Y., Tian, Q.: Person re-identification in the wild. In: Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition (2017) 9" + ], + "bbox": [ + 212, + 146, + 787, + 340 + ], + "page_idx": 16 + }, + { + "type": "header", + "text": "PapMOT: Exploring Adversarial Patch Attack against MOT", + "bbox": [ + 326, + 114, + 730, + 128 + ], + "page_idx": 16 + }, + { + "type": "page_number", + "text": "17", + "bbox": [ + 767, + 116, + 784, + 126 + ], + "page_idx": 16 + } +] \ No newline at end of file diff --git a/2024/PapMOT_ Exploring Adversarial Patch Attack against Multiple Object Tracking/e6b4ef0f-e80b-4da1-8f00-4210586c7699_model.json b/2024/PapMOT_ Exploring Adversarial Patch Attack against Multiple Object Tracking/e6b4ef0f-e80b-4da1-8f00-4210586c7699_model.json new file mode 100644 index 0000000000000000000000000000000000000000..0836b3cf325b1897d0d247a7dd5074cafb7883c9 --- /dev/null +++ b/2024/PapMOT_ Exploring Adversarial Patch Attack against Multiple Object Tracking/e6b4ef0f-e80b-4da1-8f00-4210586c7699_model.json @@ -0,0 +1,2126 @@ +[ + [ + { + "type": "title", + "bbox": [ + 0.237, + 0.141, + 0.768, + 0.187 + ], + "angle": 0, + "content": "PapMOT: Exploring Adversarial Patch Attack against Multiple Object Tracking" + }, + { + "type": "text", + "bbox": [ + 0.235, + 0.213, + 0.768, + 0.244 + ], + "angle": 0, + "content": "Jiahuan Long\\(^{1,2}\\), Tingsong Jiang\\(^{2}\\), Wen Yao\\(^{2\\star}\\), Shuai Jia\\(^{1}\\), Weijia Zhang\\(^{1}\\), Weien Zhou\\(^{2}\\), Chao Ma\\(^{1\\star}\\), and Xiaojian Chen\\(^{2}\\)" + }, + { + "type": "text", + "bbox": [ + 0.216, + 0.255, + 0.786, + 0.327 + ], + "angle": 0, + "content": "1 MoE Key Lab of Artificial Intelligence, AI Institute, Shanghai Jiao Tong University 2 Defense Innovation Institute, Chinese Academy of Military Science jiahuanlong@sjtu.edu.cn, tingsong@pku.edu.cn, wendy0782@126.com, jiashuai@sjtu.edu.cn, weijia.zhang@sjtu.edu.cn, weienzhou@outlook.com, chaoma@sjtu.edu.cn, chenxiaoqian@nudt.edu.cn" + }, + { + "type": "text", + "bbox": [ + 0.262, + 0.369, + 0.744, + 0.646 + ], + "angle": 0, + "content": "Abstract. Tracking multiple objects in a continuous video stream is crucial for many computer vision tasks. It involves detecting and associating objects with their respective identities across successive frames. Despite significant progress made in multiple object tracking (MOT), recent studies have revealed the vulnerability of existing MOT methods to adversarial attacks. Nevertheless, all of these attacks belong to digital attacks that inject pixel-level noise into input images, and are therefore ineffective in physical scenarios. To fill this gap, we propose PapMOT, which can generate physical adversarial patches against MOT for both digital and physical scenarios. Besides attacking the detection mechanism, PapMOT also optimizes a printable patch that can be detected as new targets to mislead the identity association process. Moreover, we introduce a patch enhancement strategy to further degrade the temporal consistency of tracking results across video frames, resulting in more aggressive attacks. We further develop new evaluation metrics to assess the robustness of MOT against such attacks. Extensive evaluations on multiple datasets demonstrate that our PapMOT can successfully attack various architectures of MOT trackers in digital scenarios. We also validate the effectiveness of PapMOT for physical attacks by deploying printed adversarial patches in the real world." + }, + { + "type": "text", + "bbox": [ + 0.261, + 0.66, + 0.741, + 0.688 + ], + "angle": 0, + "content": "Keywords: Physical adversarial patches \\(\\cdot\\) Multiple object tracking \\(\\cdot\\) Evaluation metrics" + }, + { + "type": "title", + "bbox": [ + 0.216, + 0.718, + 0.377, + 0.735 + ], + "angle": 0, + "content": "1 Introduction" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.753, + 0.788, + 0.815 + ], + "angle": 0, + "content": "Multiple object tracking (MOT) becomes increasingly widespread with numerous applications including autonomous driving [5] and intelligent surveillance systems [22]. Meanwhile, recent studies [17, 20] have shown that deep multiple object tracking algorithms are susceptible to attacks by adversarial examples." + }, + { + "type": "page_footnote", + "bbox": [ + 0.218, + 0.825, + 0.393, + 0.84 + ], + "angle": 0, + "content": "* Corresponding authors." + } + ], + [ + { + "type": "page_number", + "bbox": [ + 0.218, + 0.116, + 0.23, + 0.127 + ], + "angle": 0, + "content": "2" + }, + { + "type": "header", + "bbox": [ + 0.272, + 0.116, + 0.364, + 0.128 + ], + "angle": 0, + "content": "J. Long et al." + }, + { + "type": "image", + "bbox": [ + 0.223, + 0.145, + 0.784, + 0.247 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.344, + 0.248, + 0.425, + 0.258 + ], + "angle": 0, + "content": "(a) Digital attack" + }, + { + "type": "image_caption", + "bbox": [ + 0.482, + 0.248, + 0.604, + 0.259 + ], + "angle": 0, + "content": "(b) Previous patch attack" + }, + { + "type": "image_caption", + "bbox": [ + 0.656, + 0.248, + 0.752, + 0.259 + ], + "angle": 0, + "content": "(c) Our patch attack" + }, + { + "type": "image_caption", + "bbox": [ + 0.215, + 0.27, + 0.788, + 0.354 + ], + "angle": 0, + "content": "Fig. 1: A comparison of various adversarial attack methods against MOT. (a) shows a digital attack [20] against MOT, where pixel-level noise is added to each frame and the identities of two people are switched within ten frames. (b) shows a patch attack [28] that fools only the detectors of MOT, generating new fake identities. (c) demonstrates that our patch attack simultaneously fools both the detectors and trackers of MOT, which creates new identities and alters the identities of persons as well." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.396, + 0.789, + 0.624 + ], + "angle": 0, + "content": "These examples are designed to fool MOT models by adding imperceptible perturbations to video sequences. If an attack is successful, the identities of detected objects are no longer associated with their original identities. Existing adversarial attacks [17,20] against MOT have mainly focused on digital attacks, manipulating the pixels of the entire image to fool the trackers. Although these digital attacks can craft imperceptible perturbations, they are generally ineffective in physical scenarios. For instance, Figure 1(a) illustrates an adversarial instance [20] against MOT that cannot be physically implemented. Compared to digital attacks, launching attacks against MOT in the physical world is more challenging. It requires strong perturbations that can withstand diverse environmental conditions, such as illumination and geometric alterations. Also, physical attacks are inherently more challenging to detect and defend against, making them more significant threats to detectors and trackers. Therefore, developing a physical attack against MOT is of great importance to the disclosure and understanding of MOT systems' vulnerabilities in real-world applications." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.629, + 0.789, + 0.842 + ], + "angle": 0, + "content": "Physical adversarial patches are the most common form of physical attacks in the existing literature. They are widely utilized to gauge the reliability and robustness of computer vision systems, such as for face recognition [12,35] and object detection [9,13]. However, applying these methods to MOT systems presents new technical challenges. A major requirement of successful MOT systems is to maintain object identity consistency across frames. This requires the detector to identify targets in each frame and the tracker to associate these targets frame-by-frame. Simply applying adversarial patches generated by detector-based attacks to the MOT system does not sufficiently disrupt data associations across frames. As shown in Figure 1(b), applying the patch attack [28] can create new fake identities but fails to alter the identity of the real attack object with \"ID = 1\". Moreover, the existing evaluation metrics of MOT attacks are not enough to reflect the performance of patch attacks. Existing metrics neglect the importance of the detection phase, which is a critical component of MOT algorithms." + } + ], + [ + { + "type": "header", + "bbox": [ + 0.327, + 0.115, + 0.733, + 0.129 + ], + "angle": 0, + "content": "PapMOT: Exploring Adversarial Patch Attack against MOT" + }, + { + "type": "page_number", + "bbox": [ + 0.775, + 0.117, + 0.785, + 0.127 + ], + "angle": 0, + "content": "3" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.147, + 0.784, + 0.177 + ], + "angle": 0, + "content": "Also, these metrics only focus on the effects on one or a few targets, ignoring how attacks on these targets could affect the whole system." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.178, + 0.788, + 0.419 + ], + "angle": 0, + "content": "To address these problems, we propose PapMOT, a physical adversarial patch scheme against MOT systems. PapMOT generates printable adversarial patches that create new identities in the tracking phase, thereby disrupting the existing identity associations. As depicted in Figure 1(c), PapMOT deceives high-performance detectors into outputting false detection boxes (i.e., blue detection boxes), overlapping with true detection boxes. This overlap presents a significant threat to the tracking system, causing it to mistakenly alter the identities of targets in a few frames. To this end, we introduce three adversarial losses that respectively consider the size of bounding boxes, scores of detection boxes, and the smoothness of color transitions. Besides, we propose a patch enhancement strategy that crops targets over time to obtain an extra dataset. The joint training of the original training dataset and the clip dataset will increase the attack success rate on specific targets. To better evaluate the effectiveness of the attack, we construct maximum-attack-threshold evaluation metrics that integrate detection and tracking to comprehensively evaluate the performance of MOT attacks. We summarize our contributions as follows." + }, + { + "type": "text", + "bbox": [ + 0.226, + 0.429, + 0.784, + 0.488 + ], + "angle": 0, + "content": "- To the best of our knowledge, this is the first work to explore physical adversarial attack against MOT. The proposed PapMOT attack is able to produce printable patches that successfully disrupt the data association process in tracking." + }, + { + "type": "text", + "bbox": [ + 0.226, + 0.49, + 0.785, + 0.535 + ], + "angle": 0, + "content": "- Upon delving into the association algorithm in MOT, we propose a novel bounding box restriction loss that generates a wrong detection box with a high IoU value, leading to incorrect identity assignment." + }, + { + "type": "text", + "bbox": [ + 0.226, + 0.537, + 0.784, + 0.566 + ], + "angle": 0, + "content": "- We propose a new set of metrics to evaluate attacks against MOT that is more comprehensive than previous MOT attacks." + }, + { + "type": "text", + "bbox": [ + 0.226, + 0.568, + 0.784, + 0.613 + ], + "angle": 0, + "content": "- We evaluate PapMOT in both digital and physical scenarios and demonstrate its favorable attack performance compared to state-of-the-art approaches across multiple datasets." + }, + { + "type": "list", + "bbox": [ + 0.226, + 0.429, + 0.785, + 0.613 + ], + "angle": 0, + "content": null + }, + { + "type": "title", + "bbox": [ + 0.215, + 0.635, + 0.388, + 0.65 + ], + "angle": 0, + "content": "2 Related Work" + }, + { + "type": "title", + "bbox": [ + 0.215, + 0.666, + 0.473, + 0.682 + ], + "angle": 0, + "content": "2.1 Multiple Object Tracking" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.689, + 0.788, + 0.841 + ], + "angle": 0, + "content": "Existing MOT methods can be classified into two main categories: Tracking-by-Detection [4, 22, 31, 38] and Joint Detection and Tracking [39, 40]. In the former paradigm, object detectors are used to detect targets in each frame, and data association algorithms are then employed to link the detection results across video frames, including ByteTrack [38], DeepSORT [31], and OCSort [4]. ByteTrack [38] categorizes all detections into high and low scores and associates them with predicted boxes using the Kalman filter. This improves tracking performance in cases of occlusion and motion blur problems. In the latter paradigm, object detection and data association tasks are integrated into a single framework. FairMOT [39] introduces an anchor-free detection head, which allows its" + } + ], + [ + { + "type": "page_number", + "bbox": [ + 0.218, + 0.116, + 0.23, + 0.127 + ], + "angle": 0, + "content": "4" + }, + { + "type": "header", + "bbox": [ + 0.272, + 0.115, + 0.364, + 0.129 + ], + "angle": 0, + "content": "J. Long et al." + }, + { + "type": "table_caption", + "bbox": [ + 0.27, + 0.145, + 0.733, + 0.16 + ], + "angle": 0, + "content": "Table 1: A comparison of representative adversarial attack methods." + }, + { + "type": "table", + "bbox": [ + 0.22, + 0.184, + 0.785, + 0.299 + ], + "angle": 0, + "content": "
MethodsNatPatch [12]ObjPatch [28]AdvTshirt [34]Trasw [20]Hijack [17]PapMOT (ours)
Advanced detectors✓
Detectors & trackers✓✓✓
Diverse datasets✓✓✓✓
Physical attack✓✓✓✓
New evaluation metrics✓
" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.342, + 0.788, + 0.404 + ], + "angle": 0, + "content": "network to predict the position and size of the targets and generate a unique ID feature to achieve data association. In this work, we use ByteTrack [38] and FairMOT [39] as representative trackers to validate the effectiveness and generalization of the proposed adversarial patch attack." + }, + { + "type": "title", + "bbox": [ + 0.215, + 0.431, + 0.48, + 0.444 + ], + "angle": 0, + "content": "2.2 Adversarial Patch Attacks" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.462, + 0.788, + 0.702 + ], + "angle": 0, + "content": "Previous digital attacks [14-16, 33] are difficult to implement in physical scenarios as the perturbations are imperceptible for visual systems. To achieve physical attack, Thys et al. [28] generate physically feasible adversarial patches for deceiving person detectors. Expanding upon this approach, Hu et al. [13] develop AdvTexture by printing the patch on a large cloth and making different T-shirts, skirts, and dresses to realize physical adversarial attacks on pedestrian detection. These two schemes enhance the robustness of patches in the real world by involving patch transformation [1]. However, they mainly conduct attacks on lower-precision detectors such as YOLOv2 [24] and YOLOv3 [25]. For attacks on consecutive frames, Jia et al. [17] propose a tracker hijacking attack scheme, which allows adversarial samples to successfully add or remove objects within a single frame, potentially posing a safety hazard for autonomous vehicles. Lin et al. [20] present a Tracklet-Switch method, which introduces noise into certain frames, causing an ID-switch when pedestrians cross paths with each other. Yet, to our best knowledge, there has been no successful physical attack against MOT systems." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.705, + 0.789, + 0.841 + ], + "angle": 0, + "content": "To fill this research gap, we propose PapMOT, the first physically realizable attack method against MOT, alongside a set of newly proposed evaluation metrics. Different from all existing works, PapMOT is capable of not only attacking more advanced detectors (i.e., YOLOX), but also crippling both the detector and the tracker of an MOT system simultaneously and across a diversity of datasets and scenarios (i.e., MOT15, MOT17, MOT20, and BDD100K). Our contributions are expected to bring new insights into existing research on the digital and physical robustness of MOT systems. A comparison of the features and contributions of our method against prior works is provided in Table 1." + } + ], + [ + { + "type": "header", + "bbox": [ + 0.327, + 0.115, + 0.733, + 0.129 + ], + "angle": 0, + "content": "PapMOT: Exploring Adversarial Patch Attack against MOT" + }, + { + "type": "page_number", + "bbox": [ + 0.775, + 0.117, + 0.785, + 0.127 + ], + "angle": 0, + "content": "5" + }, + { + "type": "image", + "bbox": [ + 0.218, + 0.146, + 0.787, + 0.345 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.214, + 0.355, + 0.788, + 0.438 + ], + "angle": 0, + "content": "Fig. 2: The overall framework of PapMOT. (a) In the patch training phase, the objective is to optimize an adversarial patch to achieve effective attacks in both digital and physical domains. During this phase, Expectation over Transformation (EOT) is employed as data augmentation for patches. (b) Subsequently, the well-trained adversarial patch is deployed in the patch attack phase to disrupt object detection and data association in MOT." + }, + { + "type": "title", + "bbox": [ + 0.216, + 0.468, + 0.381, + 0.485 + ], + "angle": 0, + "content": "3 Methodology" + }, + { + "type": "title", + "bbox": [ + 0.216, + 0.501, + 0.525, + 0.516 + ], + "angle": 0, + "content": "3.1 Overall Framework of PapMOT" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.527, + 0.788, + 0.664 + ], + "angle": 0, + "content": "An MOT system typically consists of a detection module and an association module. The association module generally employs IoU matching to maintain consistent IDs of objects across consecutive frames, which is at the core of high-performance MOT. Provided high IoU values of a detection box across adjacent frames, this box is likely to be assigned the same ID across these frames. As such, the core objective of our PapMOT is to train printable patches that can induce false detection boxes with a high IoU value with a correctly detected pedestrian box, leading to incorrect identity assignment of this very detection. The entire pipeline of the proposed PapMOT attack is depicted in Figure 2." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.674, + 0.789, + 0.842 + ], + "angle": 0, + "content": "Patch Training Phase. The patch training phase is illustrated in Figure 2(a). First, we randomly initialize a patch and employ a data enhancement strategy (i.e., EOT [1]) to enhance its robustness. EOT incorporates potential patch transformations such as rotation, brightness, blurring, and scaling that commonly exist in the physical world into patch training. Such transformations make our patch attacks more feasible in real-world situations. Next, we embed the EOT-enhanced patch into clean videos to obtain an adversarial dataset, and crop out pedestrian images based on their bounding boxes to produce an adversarial clip dataset. Finally, we combine these two adversarial datasets to optimize the physical adversarial patch using three losses. We use total variation loss \\((\\mathcal{L}_{tv})\\) to ensure patch smoothness and printability, bounding box restric" + } + ], + [ + { + "type": "page_number", + "bbox": [ + 0.218, + 0.116, + 0.23, + 0.127 + ], + "angle": 0, + "content": "6" + }, + { + "type": "header", + "bbox": [ + 0.272, + 0.115, + 0.364, + 0.129 + ], + "angle": 0, + "content": "J. Long et al." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.147, + 0.788, + 0.193 + ], + "angle": 0, + "content": "tion loss \\((\\mathcal{L}_{bbr})\\) to minimize the discrepancy between the detection boxes of the patch and the object, and average precision loss \\((\\mathcal{L}_{ap})\\) to decrease the detection confidence of targets with the patch." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.203, + 0.79, + 0.476 + ], + "angle": 0, + "content": "Patch Attack Phase. Our patch can deceive detectors into outputting false detection boxes, thereby disrupting the identity association in tracking. This specific attack process is diagrammed in Figure 2(b). The adversarial patch, already well-trained during the patch training phase, is embedded into a video to create an attack video. Next, the attack video is processed frame-by-frame throughYOLOX, which outputs detection boxes for each frame. To clarify the disruption of data association, we consider frames \\( t - 1 \\) and \\( t \\) as illustrative examples. Assuming that there is a pedestrian with \"id=1\" in frame \\( t - 1 \\), the Kalman filter [30] outputs a predicted detection box (i.e., blue dashed-line box) by inputting the current motion states of the pedestrian. In the next frame \\( t \\), two types of detection boxes appear, with one being a false detection box of the patch (i.e., green solid-line box) and the other being the detection box of the pedestrian (i.e., blue solid-line box). These two detection boxes do Intersection over Union (IoU) calculation with the predicted detection box from frame \\( t - 1 \\). Following this, the Hungarian algorithm [18] uses these IoU values to assign identities to the detection boxes, leading the disruption of data association. (i.e., The false detection box of the patch in frame \\( t \\) is matched with the identity of pedestrians from frame \\( t - 1 \\))." + }, + { + "type": "title", + "bbox": [ + 0.215, + 0.498, + 0.544, + 0.513 + ], + "angle": 0, + "content": "3.2 Adversarial Loss for Patch Attack" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.524, + 0.788, + 0.631 + ], + "angle": 0, + "content": "For our loss formulation, we adopt the widely used average precision loss and total variation loss [12, 27, 29] to reduce the confidence score of detection while encouraging visual smoothness of the printable patches. However, only using these two losses cannot disrupt the identity association in tracking, as illustrated Figure 1(b). This motivates us to propose a novel bounding box restriction loss, which generates a wrong detection box with a high IoU value, leading to incorrect identity assignment and successfully disrupted data association." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.641, + 0.788, + 0.735 + ], + "angle": 0, + "content": "Bounding Box Restriction Loss \\((\\mathcal{L}_{bbr})\\). To effectively disrupt identity association across frames, it is essential to ensure that the sizes of the target detection box and the patch detection box are similar. To this end, a bounding box restriction loss \\(\\mathcal{L}_{bbr}\\) is formulated to restrict the height and width of the detection box, while minimizing the discrepancy between the detection boxes of the patch and the object. \\(\\mathcal{L}_{bbr}\\) is defined as:" + }, + { + "type": "equation", + "bbox": [ + 0.271, + 0.743, + 0.787, + 0.786 + ], + "angle": 0, + "content": "\\[\n\\mathcal {L} _ {b b r} = \\frac {1}{N} \\sum_ {i = 1} ^ {N} \\left(w _ {i} / I _ {w i d t h} + h _ {i} / I _ {h e i g h t}\\right) + \\left(1 - \\frac {1}{N} \\sum_ {i = 1} ^ {N} I o U \\left(P _ {i}, T _ {i}\\right)\\right), \\tag {1}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.795, + 0.788, + 0.842 + ], + "angle": 0, + "content": "where \\( N \\) is the total number of attacked objects; \\( w_{i} \\) and \\( h_{i} \\) are the width and height of predicted bounding boxes of the \\( i \\)-th target; and \\( I_{width} \\) and \\( I_{height} \\) are the width and height of video frames that are used for normalization; detection" + } + ], + [ + { + "type": "header", + "bbox": [ + 0.327, + 0.115, + 0.733, + 0.131 + ], + "angle": 0, + "content": "PapMOT: Exploring Adversarial Patch Attack against MOT" + }, + { + "type": "page_number", + "bbox": [ + 0.775, + 0.116, + 0.785, + 0.127 + ], + "angle": 0, + "content": "7" + }, + { + "type": "text", + "bbox": [ + 0.215, + 0.147, + 0.788, + 0.207 + ], + "angle": 0, + "content": "boxes of the patch and the target are denoted as \\( P_{i} \\) and \\( T_{i} \\), respectively. It is worth noting that the first term seems to make \\( T_{i} \\) have zero width and height, but we empirically find that it do not completely vanish and even create more new detection boxes due to the balance of IoU losses." + }, + { + "type": "text", + "bbox": [ + 0.215, + 0.219, + 0.787, + 0.264 + ], + "angle": 0, + "content": "Total Variation Loss \\((\\mathcal{L}_{tv})\\). The aim of \\(\\mathcal{L}_{tv}\\) is to ensure the generation of printable patches with smooth color transitions, which helps to reduce the presence of visual noise. \\(\\mathcal{L}_{tv}\\) is formulated as:" + }, + { + "type": "equation", + "bbox": [ + 0.338, + 0.277, + 0.786, + 0.315 + ], + "angle": 0, + "content": "\\[\n\\mathcal {L} _ {t v} = \\sum_ {i, j} \\sqrt {\\left(\\left(I _ {i , j} - I _ {i + 1 , j}\\right) ^ {2} + \\left(I _ {i , j} - I _ {i , j + 1}\\right) ^ {2}\\right)}, \\tag {2}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.215, + 0.327, + 0.674, + 0.344 + ], + "angle": 0, + "content": "where \\(I_{i,j}\\) denotes the pixel value at position \\((i,j)\\) in the patch." + }, + { + "type": "text", + "bbox": [ + 0.215, + 0.353, + 0.788, + 0.445 + ], + "angle": 0, + "content": "Average Precision Loss \\((\\mathcal{L}_{ap})\\). It is essential to lower the confidence of the original detection box in order to have it swapped to the one induced by the patch. \\(\\mathcal{L}_{ap}\\) does so exactly by reducing the average detection score of targets that have been attached to an adversarial patch. Given a set of \\(N\\) specific detection boxes, where the detection score for the \\(i\\)-th box is \\(s_i\\), \\(\\mathcal{L}_{ap}\\) is calculated as the sum of detection scores divided by the total number of detection boxes as follows:" + }, + { + "type": "equation", + "bbox": [ + 0.446, + 0.458, + 0.786, + 0.497 + ], + "angle": 0, + "content": "\\[\n\\mathcal {L} _ {a p} = \\frac {1}{N} \\sum_ {i = 1} ^ {N} s _ {i}. \\tag {3}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.215, + 0.504, + 0.787, + 0.534 + ], + "angle": 0, + "content": "Finally, the overall adversarial loss \\( L_{adv} \\) is a sum of the above three loss terms weighted by hyperparameters \\( \\beta \\), \\( \\gamma \\), and \\( \\delta \\):" + }, + { + "type": "equation", + "bbox": [ + 0.399, + 0.549, + 0.786, + 0.565 + ], + "angle": 0, + "content": "\\[\n\\mathcal {L} _ {a d v} = \\beta \\mathcal {L} _ {b b r} + \\gamma \\mathcal {L} _ {t v} + \\delta \\mathcal {L} _ {a p}, \\tag {4}\n\\]" + }, + { + "type": "title", + "bbox": [ + 0.215, + 0.587, + 0.562, + 0.601 + ], + "angle": 0, + "content": "3.3 Evaluation Metrics for MOT Attack" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.613, + 0.787, + 0.704 + ], + "angle": 0, + "content": "Current attack methods against MOT [17, 20] have limitations in their evaluation since they utilize identity-based evaluation metrics. These metrics have two main drawbacks. Firstly, they are primarily designed for gauging tracking performance, thereby largely overlooking the quality of detection - the other core ingredient in successful MOT. Secondly, they tend to focus only on a few specific targets, disregarding a comprehensive evaluation of the entire MOT system." + }, + { + "type": "text", + "bbox": [ + 0.215, + 0.705, + 0.788, + 0.841 + ], + "angle": 0, + "content": "As depicted in Figure 3(a), identity-based evaluation metrics [17,20] evaluate the identity of the victim in every frame, which considers tracking. However, existing MOT algorithms typically include both tracking and detection modules. Therefore, while evaluating attacks on MOT, it is essential to calculate the performance of both tracking and detection modules simultaneously. Furthermore, identity-based evaluation metrics usually adopt a strategy that aims to attack one or a few specific targets. This approach overlooks the cumulative impact of attacking these specific targets on the entire system. For instance, as shown in Figure 3(b), an attack on the victim (in the red detection box) might also affect" + } + ], + [ + { + "type": "page_number", + "bbox": [ + 0.218, + 0.116, + 0.23, + 0.127 + ], + "angle": 0, + "content": "8" + }, + { + "type": "header", + "bbox": [ + 0.272, + 0.115, + 0.364, + 0.128 + ], + "angle": 0, + "content": "J. Long et al." + }, + { + "type": "image", + "bbox": [ + 0.276, + 0.143, + 0.728, + 0.282 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.215, + 0.293, + 0.788, + 0.307 + ], + "angle": 0, + "content": "Fig. 3: A comparison of existing and our proposed evaluation metrics for MOT attacks." + }, + { + "type": "image_footnote", + "bbox": [ + 0.217, + 0.307, + 0.786, + 0.348 + ], + "angle": 0, + "content": "(a) Previous evaluation metrics are identity-based and only consider ID association. \n(b) Our proposed metrics emphasize the attack evaluation of both detection and ID association." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.381, + 0.785, + 0.411 + ], + "angle": 0, + "content": "the detection of other pedestrians (in yellow boxes), leading to switches of their identities or a decrease in detection accuracy." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.412, + 0.787, + 0.473 + ], + "angle": 0, + "content": "To overcome these limitations, we construct a maximum-attack-threshold evaluation metric that integrates detection and tracking. This type of evaluation metric is based on the classic evaluation metrics for MOT [3]. It provides a comprehensive evaluation of the robustness of MOT systems." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.484, + 0.789, + 0.637 + ], + "angle": 0, + "content": "Tracking Attack Success Rate (TASR). To assess the impact of specific targets during attacks on the overall system, TASR determines the change in MOTA and IDF1 before and after attacks and then divides it by the percentage of bounding boxes that are attacked \\((R_{bbox})\\). A high TASR indicates that the MOT system is highly susceptible to particular attacks, enabling an attacker to degrade its performance by compromising fewer targets. A low TASR implies that the attacker must compromise a greater number of targets to inflict considerable damage. The formula for TASR is: \\( TASR = (MOTA + IDF1)_{\\text{decline}} / 2R_{bbox} \\), where \\( R_{bbox} \\) is the ratio of the number of bounding boxes under attack to the total bounding boxes." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.647, + 0.788, + 0.753 + ], + "angle": 0, + "content": "Identity Obfuscation Rate (IOR). To measure the effectiveness of adversarial attacks to cause identity obfuscation during data association in MOT, we define the Identity Obfuscation Rate as: \\( IOR = D_s / P_t \\), where \\( P_t = N / T \\), where \\( T \\) represents the number of frames that are attacked consecutively, \\( N \\) is the total number of frames of the video, \\( P_t \\) denotes the maximum number of identity switches in the video, and \\( D_s \\) is the number of identity switches throughout the video sequences." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.765, + 0.789, + 0.843 + ], + "angle": 0, + "content": "Specific Target Attack Success Rate (STASR). To evaluate attacks on specific targets in terms of both false positive rate (FP) and the number of identity switches, STASR focuses not only on how the attack interferes with the detection of the targets but also on how the attack disrupts the tracking of the targets. STASR is defined as: \\( STASR = (FP + ID_{switch})_{increase} / A_{max} \\), where" + } + ], + [ + { + "type": "header", + "bbox": [ + 0.327, + 0.115, + 0.733, + 0.129 + ], + "angle": 0, + "content": "PapMOT: Exploring Adversarial Patch Attack against MOT" + }, + { + "type": "page_number", + "bbox": [ + 0.775, + 0.117, + 0.785, + 0.127 + ], + "angle": 0, + "content": "9" + }, + { + "type": "text", + "bbox": [ + 0.217, + 0.147, + 0.785, + 0.177 + ], + "angle": 0, + "content": "\\(A_{max} = P_t + P_n\\), \\(P_t\\) represents the maximum number of identity switches in the video, and \\(P_n\\) denotes the number of detection boxes that are attacked." + }, + { + "type": "title", + "bbox": [ + 0.217, + 0.202, + 0.375, + 0.219 + ], + "angle": 0, + "content": "4 Experiments" + }, + { + "type": "text", + "bbox": [ + 0.217, + 0.234, + 0.785, + 0.28 + ], + "angle": 0, + "content": "In this section, we perform extensive empirical analysis of the proposed PapMOT attack on MOT in both digital and physical environments. We also conduct ablation studies on the adversarial loss and variants of adversarial patches." + }, + { + "type": "title", + "bbox": [ + 0.217, + 0.303, + 0.427, + 0.318 + ], + "angle": 0, + "content": "4.1 Experimental Setup" + }, + { + "type": "text", + "bbox": [ + 0.217, + 0.34, + 0.785, + 0.43 + ], + "angle": 0, + "content": "Datasets and Metrics. We evaluate our attack on four popular MOT benchmarks: MOT15 [19], MOT17 [23], MOT20 [7], and BDD100K [36]. The target objects of the first three datasets are pedestrian-only, while BDD100K further includes vehicle objects. We adopt our proposed evaluation metrics, namely TASR, SATSR, and IOR, to assess the performance of our PapMOT against state-of-the-art methods on the MOT task. More comparisons are found in Appendix." + }, + { + "type": "text", + "bbox": [ + 0.217, + 0.441, + 0.785, + 0.561 + ], + "angle": 0, + "content": "Baselines. Existing MOT attack methods can be divided into three categories: 1) methods that attack detectors in the MOT system by injecting pixel-level noise, e.g., RAnoise [21] and TOG [6]. 2) methods that attack detectors in the MOT system by placing optimized patches, e.g., AdvTexture [13] and ObjPatch [28]. 3) methods that attack both detectors and trackers in the MOT system by injecting pixel-level noise, e.g., Hijack [17] and Trasw [20]. Our PapMOT introduces a physically realizable adversarial patch into the MOT system to attack both detectors and trackers." + }, + { + "type": "text", + "bbox": [ + 0.217, + 0.573, + 0.785, + 0.708 + ], + "angle": 0, + "content": "Victim Models. We choose two representative MOT methods as victim models in our experiments: ByteTrack [38] that represents the tracking-by-detection paradigm and FairMOT [39] that represents the joint detection and tracking paradigm. Besides, ByteTrack is equipped with YOLOX [11], a more advanced version of theYOLO detector family trained on CrowdHuman [26], MOT17 [23], CityPerson [37], and ETHZ [10] datasets. The pre-trained model of FairMOT is provided by PaddleDetection [2], and is trained on Caltech Pedestrian [8], CityPersons [37], CUHK-SYSU [32], PRW [41], ETHZ [10], and MOT17 [23] datasets." + }, + { + "type": "text", + "bbox": [ + 0.217, + 0.72, + 0.785, + 0.84 + ], + "angle": 0, + "content": "Implementation Details. In our digital experiments, to strike a balance between the stealthiness and effectiveness of attacks, we set the noise standard deviation of RAnoise to 10, and limit the maximum disturbance range in TOG to 8/255. The patch sizes for both AdvTexture and ObjPatch are set to one-third of the targeted detection box's size. We define an object as \"attacked\" if it overlaps with another object with an IOU greater than 0.2. In our physical experiments, we print the patches on foam boards with a fixed width and height of \\(50\\mathrm{cm}\\). The size ratio between the patch and an adult who is \\(1.6\\mathrm{m}\\) tall is kept at" + } + ], + [ + { + "type": "page_number", + "bbox": [ + 0.218, + 0.116, + 0.236, + 0.127 + ], + "angle": 0, + "content": "10" + }, + { + "type": "header", + "bbox": [ + 0.272, + 0.115, + 0.362, + 0.129 + ], + "angle": 0, + "content": "J. Long et al." + }, + { + "type": "table_caption", + "bbox": [ + 0.216, + 0.145, + 0.788, + 0.173 + ], + "angle": 0, + "content": "Table 2: Comparison with other representative adversarial attack methods on the MOT15 and MOT17 datasets." + }, + { + "type": "table", + "bbox": [ + 0.221, + 0.185, + 0.773, + 0.344 + ], + "angle": 0, + "content": "
DatasetMOT15MOT17
VictimFairMOTByteTrackFairMOTByteTrack
MetricsTASRSATSRIORTASRSATSRIORTASRSATSRIORTASRSATSRIOR
RAnoise [21]0.70.6-0.60.3-0.30.5--0.2-
TOG [6]13.77.04.111.46.33.89.14.82.54.73.91.2
AdvTexture [13]36.435.64.638.143.75.845.432.15.142.247.76.0
ObjPatch [28]45.240.66.343.759.38.261.349.56.261.476.77.4
Hijack [17]14.218.539.311.415.929.010.019.226.58.415.425.4
Trasw [20]40.645.866.231.439.255.422.128.944.821.026.842.3
Ours76.485.062.385.379.265.976.893.046.294.689.679.5
" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.371, + 0.788, + 0.434 + ], + "angle": 0, + "content": "around \\(30\\%\\), ensuring that essential areas such as the face are not obstructed to maintain physical feasibility. For hyperparameters in the overall adversarial loss (i.e., Eq. 4), we set \\(\\beta = 1\\), \\(\\gamma = 2.5\\), and \\(\\delta = 2\\). Datasets used in our patch generation consist of ETHZ, MOT17, and MOT17 clip datasets." + }, + { + "type": "title", + "bbox": [ + 0.216, + 0.453, + 0.502, + 0.47 + ], + "angle": 0, + "content": "4.2 Digital Domain Performance" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.478, + 0.789, + 0.689 + ], + "angle": 0, + "content": "Table 2 and Table 3 compare the performance of various attacks on the MOT15, MOT17, MOT20, and BDD100K datasets. As can be seen, PapMOT achieves the highest attack success rate across all four datasets with a \\(50\\%\\) bounding box attack, manifesting its superiority over existing methods (i.e., RAnoise [21], TOG [6], ADVTexure [13], ObjPatch [28], Hijack [17], Trasw [20]) against MOT in both pedestrian (i.e., MOT15, MOT17, and MOT20) and autonomous driving scenarios (i.e., BDD100K). We notice that the IOR of PapMOT is marginally below that of Trasw on the MOT15 and MOT20 datasets, which is possibly due to Trasw's sole emphasis on attacking data association. Nonetheless, PapMOT surpasses Trasw by a huge margin in terms of TASR and SATSR. Overall, Trasw as a digital method is hard to be implemented in the physical world, and its attack is strictly triggered only when an object overlaps with the attack object. Thus, PapMOT outperforms Trasw in terms of functionality and overall attack performance." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.689, + 0.79, + 0.841 + ], + "angle": 0, + "content": "The superior performance of PapMOT is most evident on the MOT17 dataset, where it consistently surpasses all competing methods across all three metrics when applied against both FairMOT and ByteTrack. The success of PapMOT is largely due to incorporating the MOT17 clip dataset and MOT17 into its patch training dataset (more details about the clip dataset are found in Appendix). It is noteworthy that TASR and SATSR both incorporate the evaluation of detection box performance, whereas IOR predominantly assesses identity association across frames. As such, it is observed that TASR and SATSR are typically higher for patch-based attacks and digital attacks compared to IOR (e.g., when attacking ByteTrack in MOT17, the IOR of AdvTexture is 6.0, while TASR and" + } + ], + [ + { + "type": "header", + "bbox": [ + 0.327, + 0.115, + 0.732, + 0.129 + ], + "angle": 0, + "content": "PapMOT: Exploring Adversarial Patch Attack against MOT" + }, + { + "type": "page_number", + "bbox": [ + 0.769, + 0.117, + 0.784, + 0.127 + ], + "angle": 0, + "content": "11" + }, + { + "type": "table_caption", + "bbox": [ + 0.217, + 0.145, + 0.785, + 0.173 + ], + "angle": 0, + "content": "Table 3: Comparison with other representative adversarial attack methods on the MOT20 and BDD100K datasets." + }, + { + "type": "table", + "bbox": [ + 0.221, + 0.185, + 0.773, + 0.345 + ], + "angle": 0, + "content": "
DatasetMOT20BDD100K
VictimFairMOTByteTrackFairMOTByteTrack
MetricsTASRSATSRIORTASRSATSRIORTASRSATSRIORTASRSATSRIOR
RAnoise [21]2.60.90.44.11.50.42.20.7-1.90.7-
TOG [6]8.62.50.67.62.30.618.16.72.413.54.82.2
AdvTexture [13]23.518.52.932.026.73.221.112.64.418.319.04.8
ObjPatch [28]42.321.44.441.327.56.826.319.45.220.523.37.6
Hijack [17]17.918.436.815.934.727.07.19.810.57.48.213.5
Trasw [20]31.233.567.026.628.564.212.217.423.311.013.119.3
Ours61.949.556.658.541.653.946.454.734.163.461.238.8
" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.373, + 0.785, + 0.435 + ], + "angle": 0, + "content": "SATSR in ADvTexture are 42.2 and 47.7, respectively). This observation can be accounted for by the fact that these attacks primarily target detectors, which significantly affect the performance of detection boxes, but minimally impact identity association across frames." + }, + { + "type": "title", + "bbox": [ + 0.216, + 0.457, + 0.51, + 0.472 + ], + "angle": 0, + "content": "4.3 Physical Domain Performance" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.482, + 0.785, + 0.558 + ], + "angle": 0, + "content": "While our patches have shown outstanding performance in the digital domain, translating this success to the physical world is challenging due to the ever-changing of physical environments. In this section, we qualitatively evaluate the effectiveness of our adversarial patches, and further probe their limits under diverse physical environments." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.568, + 0.785, + 0.643 + ], + "angle": 0, + "content": "Physical Validation. We train three patches with different datasets (i.e., ETHZ, MOT17, MOT20) for FairMOT and evaluate them against the ByteTrack detector. Figure 4 showcases the demonstration of these adversarial patches for physical-world attacks under diverse scenes (refer to Supplementary Materials for a video demo)." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.644, + 0.785, + 0.764 + ], + "angle": 0, + "content": "In Figure 4 (a), an adversarial patch is attached to a pedestrian, targeting a single object. In the absence of our attack, the target's ID detected by the tracker is consistently 1, and the detected bounding box is accurate. When the attack is imposed, the detector produces an additional false detection box (in blue), and the accuracy of the initially accurate detection box deteriorates significantly. Moreover, the tracker is quickly deprived of the ability to identity associations across frames, and an ID switch takes place within 30 frames. These outcomes validate the effectiveness of our attack in single-object scenarios." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.765, + 0.787, + 0.84 + ], + "angle": 0, + "content": "Figure 4 (b) illustrates a multi-object scenario, where the adversarial patch is attached to one of the two pedestrians in the scene. When the pedestrian in white with the patch (denoted as \"target A\") passes by in front of the other subject in black (\"target B\"), their detected identities are successfully maintained in the absence of the attack. When attacked, however, the tracker mistakenly" + } + ], + [ + { + "type": "page_number", + "bbox": [ + 0.218, + 0.116, + 0.236, + 0.127 + ], + "angle": 0, + "content": "12" + }, + { + "type": "header", + "bbox": [ + 0.272, + 0.115, + 0.362, + 0.128 + ], + "angle": 0, + "content": "J. Long et al." + }, + { + "type": "image", + "bbox": [ + 0.23, + 0.144, + 0.778, + 0.33 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.214, + 0.341, + 0.785, + 0.384 + ], + "angle": 0, + "content": "Fig. 4: Black-box transfer attack of different PapMOT patches in the real world. It illustrates the effective attack for the scenarios of a single pedestrian (a), two pedestrians overlap (b), and a pedestrian walking past from behind a fixed patch (c)." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.417, + 0.785, + 0.478 + ], + "angle": 0, + "content": "predicts two identities (i.e., 1 and 3) for target A. Also, an ID switch occurs and the accuracy of both boxes drops as target A passes by. This example shows our proposed attack works in multiple-object scenarios and its effect applies beyond the object with the attached patch." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.479, + 0.787, + 0.555 + ], + "angle": 0, + "content": "Figure 4 (c) shows the scenario where a patch is kept stationary and is misidentified as a pedestrian by the tracker. When attacked, a pedestrian passing behind it immediately causes an ID switch to occur (i.e., ID changes from 2 to 1). This example suggests that our adversarial patch is also suitable for static placement and is capable of affecting pedestrians as they pass by." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.567, + 0.788, + 0.763 + ], + "angle": 0, + "content": "Attack Boundary. We examine three key physical conditions that may affect the effectiveness of patch attacks: (a) illumination, (b) distance, and (c) angle, shown in Fig. 5. For illumination, we sweep over the entire spectrum from total darkness to total brightness by adjusting the ISO sensitivity of the camera. We divide the illumination levels into 10 segments between 0 and 1. An illumination value of 0.5 denotes the normal exposure under overcast outdoor conditions with the ISO set to 200; 0 and 1 correspond to full darkness and full brightness, respectively. For distance, we vary the range between the patch and the camera from 0.5 to 10 meters to obtain a better view. For angle, we control the angle of the patch relative to the camera by adjusting the deflection angle of the person and using 5-degree intervals to present the final attack results. The quantitative evaluation results of the attack boundary effectiveness for three different patches are presented in Table 4." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.765, + 0.788, + 0.84 + ], + "angle": 0, + "content": "Based on the experimental results, several important conclusions can be drawn. Firstly, different patches can achieve varying attack effects depending on the physical conditions. Secondly, the distance between the patches and the camera must be maintained within a specific range (0.8m-4.1m) to achieve an effective attack. Thirdly, patch attacks work better in relatively dark environ" + } + ], + [ + { + "type": "header", + "bbox": [ + 0.327, + 0.115, + 0.733, + 0.129 + ], + "angle": 0, + "content": "PapMOT: Exploring Adversarial Patch Attack against MOT" + }, + { + "type": "page_number", + "bbox": [ + 0.769, + 0.116, + 0.785, + 0.127 + ], + "angle": 0, + "content": "13" + }, + { + "type": "image", + "bbox": [ + 0.227, + 0.146, + 0.778, + 0.257 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.218, + 0.268, + 0.782, + 0.283 + ], + "angle": 0, + "content": "Fig. 5: Studies on patch effectiveness at varying illuminations, distances, and angles." + }, + { + "type": "table_caption", + "bbox": [ + 0.26, + 0.297, + 0.741, + 0.312 + ], + "angle": 0, + "content": "Table 4: Studies on the effectiveness range of three adversarial patches." + }, + { + "type": "table", + "bbox": [ + 0.314, + 0.324, + 0.684, + 0.396 + ], + "angle": 0, + "content": "
DistanceIlluminationPatch SizeAngle
Patch 10.8m-3.1m0.4-0.740cm-100cm0°-25°
Patch 21.2m-4.1m0.5-0.840cm-100cm0°-15°
Patch 31.1m-3.6m0.4-0.840cm-100cm0°-30°
" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.425, + 0.785, + 0.53 + ], + "angle": 0, + "content": "ments, resulting in more false detection boxes. Moreover, a larger patch size has a stronger attack ability. However, to ensure the stealthiness and effectiveness of the patch, the patch size in PapMOT is set to approximately \\(30\\%\\) of the target size. Lastly, to ensure effective attack outcomes, the angle of the patch relative to the camera generally needs to be limited to a small range (i.e., within 15 degrees). It may be considered to add rotation transformation in the patch training to enhance the robustness of the patches against rotation." + }, + { + "type": "title", + "bbox": [ + 0.216, + 0.552, + 0.401, + 0.566 + ], + "angle": 0, + "content": "4.4 Ablation Studies" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.578, + 0.785, + 0.623 + ], + "angle": 0, + "content": "We analyze the effectiveness of each adversarial loss term in our proposed training objective. We also conduct real-world experiments to evaluate the robustness against unoptimized patches, including real images and uniform color patches." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.633, + 0.788, + 0.77 + ], + "angle": 0, + "content": "Loss Functions. Table 5 shows how different adversarial losses affect the performance of our proposed on against MOT17. The results indicate that PapMOT without \\(\\mathcal{L}_{bbr}\\) performs significantly worse than PapMOT without \\(\\mathcal{L}_{ap}\\) in terms of IOR, scoring 7.2 and 52.6, respectively, when attacking ByteTrack. This suggests that the \\(\\mathcal{L}_{bbr}\\) loss is critical in disrupting the consistency of data association. Conversely, when attacking FairMOT, PapMOT without \\(\\mathcal{L}_{ap}\\) scores substantially lower in TASR and SATSR than PapMOT without \\(\\mathcal{L}_{bbr}\\), with scores of 11.0 and 13.1, respectively, compared to 50.5 and 43.3. This underscores the significant impact of the \\(\\mathcal{L}_{ap}\\) loss on detector performance." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.78, + 0.788, + 0.841 + ], + "angle": 0, + "content": "Various Patches. We evaluate the effectiveness of attacking the MOT system using unoptimized patches (e.g., full-white patches, full-black patches, and real image patches). As shown in Figure 6, although the target is partially occluded by (a) the full-white patch or (b) the full-black patch, it is tracked steadily" + } + ], + [ + { + "type": "page_number", + "bbox": [ + 0.218, + 0.116, + 0.236, + 0.127 + ], + "angle": 0, + "content": "14" + }, + { + "type": "header", + "bbox": [ + 0.272, + 0.115, + 0.364, + 0.129 + ], + "angle": 0, + "content": "J. Long et al." + }, + { + "type": "image", + "bbox": [ + 0.237, + 0.143, + 0.413, + 0.232 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.264, + 0.234, + 0.4, + 0.248 + ], + "angle": 0, + "content": "(a) Full-white patch" + }, + { + "type": "image", + "bbox": [ + 0.417, + 0.144, + 0.588, + 0.232 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.441, + 0.234, + 0.578, + 0.248 + ], + "angle": 0, + "content": "(b) Full-black patch" + }, + { + "type": "image", + "bbox": [ + 0.593, + 0.144, + 0.768, + 0.232 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.61, + 0.234, + 0.754, + 0.247 + ], + "angle": 0, + "content": "(c) Real image patch" + }, + { + "type": "image_caption", + "bbox": [ + 0.215, + 0.26, + 0.788, + 0.289 + ], + "angle": 0, + "content": "Fig. 6: The attack testing using real image, full-white, and full-black patches on the MOT system. These patches have the same size of \\(50 \\times 50\\mathrm{cm}\\) as ours." + }, + { + "type": "table_caption", + "bbox": [ + 0.299, + 0.305, + 0.704, + 0.32 + ], + "angle": 0, + "content": "Table 5: The effectiveness analysis of the adversarial losses." + }, + { + "type": "table", + "bbox": [ + 0.297, + 0.332, + 0.706, + 0.472 + ], + "angle": 0, + "content": "
VictimAttackerMetrics
TASRSATSRIOR
ByteTrack [38]PapMOT w/o Lbbr61.364.77.2
PapMOT w/o Lap21.926.852.6
PapMOT94.689.679.5
FairMOT [39]PapMOT w/o Lbbr50.543.35.6
PapMOT w/o Lap11.013.129.3
PapMOT76.893.046.2
" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.503, + 0.788, + 0.58 + ], + "angle": 0, + "content": "and the ID remains unchanged throughout the whole video sequence. Similarly, although the target is partially occluded by (c) the real-image patch, it is tracked steadily and the ID is maintained at 1 throughout the entire video sequence. The experiments indicate that it is difficult to attack MOT systems with adversarial patches that are not optimized." + }, + { + "type": "title", + "bbox": [ + 0.216, + 0.607, + 0.36, + 0.623 + ], + "angle": 0, + "content": "5 Conclusion" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.643, + 0.788, + 0.809 + ], + "angle": 0, + "content": "In this paper, we introduce a novel physical adversarial attack method named PapMOT against multiple object tracking systems. Unlike all previous attacks that inject noise into input video frames in the digital domain, our approach optimizes printable adversarial patches that effectively disrupt the data association of MOT trackers across frames in the actual physical world. We demonstrate the feasibility and effectiveness of PapMOT in both digital and physical scenarios through extensive experiments on multiple datasets. In future work, we plan to investigate the creation of more natural-looking adversarial patches, such as transforming these patches into clothing to deceive the MOT system. Additionally, we will expand physical adversarial MOT attacks from pedestrian scenarios to autonomous driving scenarios." + }, + { + "type": "text", + "bbox": [ + 0.215, + 0.81, + 0.787, + 0.841 + ], + "angle": 0, + "content": "Acknowledgements. This work was supported in part by NSFC (62322113, 62376156)." + } + ], + [ + { + "type": "header", + "bbox": [ + 0.327, + 0.115, + 0.732, + 0.129 + ], + "angle": 0, + "content": "PapMOT: Exploring Adversarial Patch Attack against MOT" + }, + { + "type": "page_number", + "bbox": [ + 0.769, + 0.117, + 0.785, + 0.127 + ], + "angle": 0, + "content": "15" + }, + { + "type": "title", + "bbox": [ + 0.217, + 0.145, + 0.323, + 0.16 + ], + "angle": 0, + "content": "References" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.173, + 0.785, + 0.2 + ], + "angle": 0, + "content": "1. Athalye, A., Engstrom, L., Ilyas, A., Kwok, K.: Synthesizing robust adversarial examples. In: International conference on machine learning (2018) 4, 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.201, + 0.785, + 0.241 + ], + "angle": 0, + "content": "2. Authors, P.: Paddledetection, object detection and instance segmentation toolkit based on paddlepaddle. https://github.com/PaddlePaddle/PaddleDetection (2019) 9" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.242, + 0.785, + 0.28 + ], + "angle": 0, + "content": "3. Bernardin, K., Stiefelhagen, R.: Evaluating multiple object tracking performance: the clear mot metrics. EURASIP Journal on Image and Video Processing (2008) 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.282, + 0.785, + 0.322 + ], + "angle": 0, + "content": "4. Cao, J., Weng, X., Khirodkar, R., Pang, J., Kitani, K.: Observation-centric sort: Rethinking sort for robust multi-object tracking. arXiv preprint arXiv:2203.14360 (2022) 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.323, + 0.785, + 0.363 + ], + "angle": 0, + "content": "5. Chiu, H.K., Li, J., Ambrus, R., Bohg, J.: Probabilistic 3d multi-modal, multi-object tracking for autonomous driving. In: IEEE International Conference on Robotics and Automation (2021) 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.364, + 0.785, + 0.404 + ], + "angle": 0, + "content": "6. Chow, K.H., Liu, L., Gursoy, M.E., Truex, S., Wei, W., Wu, Y.: Understanding object detection through an adversarial lens. In: European Symposium on Research in Computer Security (2020) 9, 10, 11" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.405, + 0.785, + 0.445 + ], + "angle": 0, + "content": "7. Dendorfer, P., Rezatofighi, H., Milan, A., Shi, J., Cremers, D., Reid, I., Roth, S., Schindler, K., Leal-Taixe, L.: Mot20: A benchmark for multi object tracking in crowded scenes. arXiv preprint arXiv:2003.09003 (2020) 9" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.446, + 0.785, + 0.486 + ], + "angle": 0, + "content": "8. Dollar, P., Wojek, C., Schiele, B., Perona, P.: Pedestrian detection: A benchmark. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (2009) 9" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.486, + 0.785, + 0.54 + ], + "angle": 0, + "content": "9. Du, A., Chen, B., Chin, T.J., Law, Y.W., Sasdelli, M., Rajasegaran, R., Campbell, D.: Physical adversarial attacks on an aerial imagery object detector. In: Proceedings of the IEEE/CVF Winter Conference on Applications of Computer Vision (2022) 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.541, + 0.785, + 0.581 + ], + "angle": 0, + "content": "10. Ess, A., Leibe, B., Schindler, K., van Gool, L.: A mobile vision system for robust multi-person tracking. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (2008) 9" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.582, + 0.785, + 0.609 + ], + "angle": 0, + "content": "1. Ge, Z., Liu, S., Wang, F., Li, Z., Sun, J.: Yolox: Exceeding yolo series in 2021. arXiv preprint arXiv:2107.08430 (2021) 9" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.61, + 0.785, + 0.649 + ], + "angle": 0, + "content": "2. Hu, Y.C.T., Kung, B.H., Tan, D.S., Chen, J.C., Hua, K.L., Cheng, W.H.: Naturalistic physical adversarial patch for object detectors. In: Proceedings of the IEEE/CVF International Conference on Computer Vision (2021) 2, 4, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.65, + 0.785, + 0.69 + ], + "angle": 0, + "content": "3. Hu, Z., Huang, S., Zhu, X., Sun, F., Zhang, B., Hu, X.: Adversarial texture for fooling person detectors in the physical world. In: Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition (2022) 2, 4, 9, 10, 11" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.691, + 0.785, + 0.718 + ], + "angle": 0, + "content": "4. Jia, S., Ma, C., Song, Y., Yang, X.: Robust tracking against adversarial attacks. In: European Conference on Computer Vision (2020) 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.718, + 0.785, + 0.758 + ], + "angle": 0, + "content": "5. Jia, S., Ma, C., Yao, T., Yin, B., Ding, S., Yang, X.: Exploring frequency adversarial attacks for face forgery detection. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (2022) 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.759, + 0.785, + 0.799 + ], + "angle": 0, + "content": "6. Jia, S., Song, Y., Ma, C., Yang, X.: Iou attack: Towards temporally coherent blackbox adversarial attack for visual object tracking. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (2021) 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.8, + 0.785, + 0.84 + ], + "angle": 0, + "content": "7. Jia, Y., Lu, Y., Shen, J., Chen, Q.A., Zhong, Z., Wei, T.: Fooling detection alone is not enough: Adversarial attack against multiple object tracking. In: International Conference on Learning Representations (2020) 1, 2, 4, 7, 9, 10, 11" + }, + { + "type": "list", + "bbox": [ + 0.226, + 0.173, + 0.785, + 0.84 + ], + "angle": 0, + "content": null + } + ], + [ + { + "type": "page_number", + "bbox": [ + 0.218, + 0.116, + 0.236, + 0.127 + ], + "angle": 0, + "content": "16" + }, + { + "type": "header", + "bbox": [ + 0.272, + 0.115, + 0.362, + 0.128 + ], + "angle": 0, + "content": "J. Long et al." + }, + { + "type": "ref_text", + "bbox": [ + 0.218, + 0.147, + 0.785, + 0.175 + ], + "angle": 0, + "content": "18. Kuhn, H.W.: The hungarian method for the assignment problem. Naval research logistics quarterly 2(1-2), 83-97 (1955) 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.218, + 0.175, + 0.785, + 0.217 + ], + "angle": 0, + "content": "19. Leal-Taixe, L., Milan, A., Reid, I., Roth, S., Schindler, K.: Motchallenge 2015: Towards a benchmark for multi-target tracking. arXiv preprint arXiv:1504.01942 (2015) 9" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.217, + 0.785, + 0.257 + ], + "angle": 0, + "content": "20. Lin, D., Chen, Q., Zhou, C., He, K.: Trasw: Tracklet-switch adversarial attacks against multi-object tracking. arXiv preprint arXiv:2111.08954 (2023) 1, 2, 4, 7, 9, 10, 11" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.257, + 0.785, + 0.284 + ], + "angle": 0, + "content": "21. Luisier, F., Blu, T., Unser, M.: Image denoising in mixed poisson-gaussian noise. IEEE Transactions on image processing (2010) 9, 10, 11" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.284, + 0.785, + 0.311 + ], + "angle": 0, + "content": "22. Luo, W., Xing, J., Milan, A., Zhang, X., Liu, W., Kim, T.K.: Multiple object tracking: A literature review. Artificial intelligence (2021) 1, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.311, + 0.785, + 0.338 + ], + "angle": 0, + "content": "23. Milan, A., Leal-Taixe, L., Reid, I., Roth, S., Schindler, K.: Mot16: A benchmark for multi-object tracking. arXiv preprint arXiv:1603.00831 (2016) 9" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.338, + 0.785, + 0.365 + ], + "angle": 0, + "content": "24. Redmon, J., Farhadi, A.: Yolo9000: Better, faster, stronger. In: IEEE Conference on Computer Vision and Pattern Recognition (2017) 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.365, + 0.785, + 0.392 + ], + "angle": 0, + "content": "25. Redmon, J., Farhadi, A.: Yolov3: An incremental improvement. arXiv preprint arXiv:1804.02767 (2018) 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.392, + 0.785, + 0.432 + ], + "angle": 0, + "content": "26. Shao, S., Zhao, Z., Li, B., Xiao, T., Yu, G., Zhang, X., Sun, J.: Crowdhuman: A benchmark for detecting human in a crowd. arXiv preprint arXiv:1805.00123 (2018) 9" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.432, + 0.785, + 0.473 + ], + "angle": 0, + "content": "27. Sharif, M., Bhagavatula, S., Bauer, L., Reiter, M.K.: Accessorize to a crime. In: Proceedings of the 2016 ACM SIGSAC Conference on Computer and Communications Security (2016) 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.473, + 0.785, + 0.527 + ], + "angle": 0, + "content": "28. Thys, S., Van Ranst, W., Goedemé, T.: Fooling automated surveillance cameras: adversarial patches to attack person detection. In: Proceedings of the IEEE/CVF conference on computer vision and pattern recognition workshops (2019) 2, 4, 9, 10, 11" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.527, + 0.785, + 0.582 + ], + "angle": 0, + "content": "29. Wang, D., Jiang, T., Sun, J., Zhou, W., Gong, Z., Zhang, X., Yao, W., Chen, X.: Fca: Learning a 3d full-coverage vehicle camouflage for multi-view physical adversarial attack. In: Proceedings of the AAAI conference on artificial intelligence (2022) 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.582, + 0.744, + 0.596 + ], + "angle": 0, + "content": "30. Welch, G., Bishop, G., et al.: An introduction to the kalman filter (1995) 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.596, + 0.785, + 0.634 + ], + "angle": 0, + "content": "31. Wojke, N., Bewley, A., Paulus, D.: Simple online and realtime tracking with a deep association metric. In: IEEE International Conference on Image Processing (2017) 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.635, + 0.785, + 0.677 + ], + "angle": 0, + "content": "32. Xiao, T., Li, S., Wang, B., Lin, L., Wang, X.: Joint detection and identification feature learning for person search. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (2017) 9" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.677, + 0.785, + 0.717 + ], + "angle": 0, + "content": "33. Xu, H., Ma, Y., Liu, H.C., Deb, D., Liu, H., Tang, J.L., Jain, A.K.: Adversarial attacks and defenses in images, graphs and text: A review. International Journal of Automation and Computing (2020) 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.717, + 0.785, + 0.759 + ], + "angle": 0, + "content": "34. Xu, K., Zhang, G., Liu, S., Fan, Q., Sun, M., Chen, H., Chen, P.Y., Wang, Y., Lin, X.: Adversarial t-shirt! evading person detectors in a physical world. In: European Conference on Computer Vision (2020) 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.759, + 0.785, + 0.8 + ], + "angle": 0, + "content": "35. Yang, X., Wei, F., Zhang, H., Zhu, J.: Design and interpretation of universal adversarial patches in face detection. In: European Conference on Computer Vision (2020) 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.8, + 0.785, + 0.841 + ], + "angle": 0, + "content": "36. Yu, F., Chen, H., Wang, X., Xian, W., Chen, Y., Liu, F., Madhavan, V., Darrell, T.: Bdd100k: A diverse driving dataset for heterogeneous multitask learning. In: IEEE/CVF Conference on Computer Vision and Pattern Recognition (2020) 9" + }, + { + "type": "list", + "bbox": [ + 0.217, + 0.147, + 0.785, + 0.841 + ], + "angle": 0, + "content": null + } + ], + [ + { + "type": "header", + "bbox": [ + 0.327, + 0.115, + 0.732, + 0.129 + ], + "angle": 0, + "content": "PapMOT: Exploring Adversarial Patch Attack against MOT" + }, + { + "type": "page_number", + "bbox": [ + 0.769, + 0.117, + 0.785, + 0.127 + ], + "angle": 0, + "content": "17" + }, + { + "type": "ref_text", + "bbox": [ + 0.214, + 0.147, + 0.788, + 0.189 + ], + "angle": 0, + "content": "37. Zhang, S., Benenson, R., Schiele, B.: Citypersons: A diverse dataset for pedestrian detection. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (2017) 9" + }, + { + "type": "ref_text", + "bbox": [ + 0.214, + 0.19, + 0.788, + 0.232 + ], + "angle": 0, + "content": "38. Zhang, Y., Sun, P., Jiang, Y., Yu, D., Weng, F., Yuan, Z., Luo, P., Liu, W., Wang, X.: Bytetrack: Multi-object tracking by associating every detection box. In: European Conference on Computer Vision (2022) 3, 4, 9, 14" + }, + { + "type": "ref_text", + "bbox": [ + 0.214, + 0.232, + 0.788, + 0.272 + ], + "angle": 0, + "content": "39. Zhang, Y., Wang, C., Wang, X., Zeng, W., Liu, W.: Fairmot: On the fairness of detection and re-identification in multiple object tracking. International Journal of Computer Vision (2021) 3, 4, 9, 14" + }, + { + "type": "ref_text", + "bbox": [ + 0.214, + 0.273, + 0.788, + 0.3 + ], + "angle": 0, + "content": "40. Zhao, Z., Wu, Z., Zhuang, Y., Li, B., Jia, J.: Tracking objects as pixel-wise distributions. In: European Conference on Computer Vision (2022) 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.214, + 0.3, + 0.788, + 0.342 + ], + "angle": 0, + "content": "41. Zheng, L., Zhang, H., Sun, S., Chandraker, M., Yang, Y., Tian, Q.: Person re-identification in the wild. In: Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition (2017) 9" + }, + { + "type": "list", + "bbox": [ + 0.214, + 0.147, + 0.788, + 0.342 + ], + "angle": 0, + "content": null + } + ] +] \ No newline at end of file diff --git a/2024/PapMOT_ Exploring Adversarial Patch Attack against Multiple Object Tracking/e6b4ef0f-e80b-4da1-8f00-4210586c7699_origin.pdf b/2024/PapMOT_ Exploring Adversarial Patch Attack against Multiple Object Tracking/e6b4ef0f-e80b-4da1-8f00-4210586c7699_origin.pdf new file mode 100644 index 0000000000000000000000000000000000000000..5e285fa07dfa0c2985c0e313222919d16aa2018c --- /dev/null +++ b/2024/PapMOT_ Exploring Adversarial Patch Attack against Multiple Object Tracking/e6b4ef0f-e80b-4da1-8f00-4210586c7699_origin.pdf @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:7f62d16f787da98fe55548c9f8c7ad4e3764a3e6495074c1a0c0653fe9c2fa9e +size 14876099 diff --git a/2024/PapMOT_ Exploring Adversarial Patch Attack against Multiple Object Tracking/full.md b/2024/PapMOT_ Exploring Adversarial Patch Attack against Multiple Object Tracking/full.md new file mode 100644 index 0000000000000000000000000000000000000000..b6b1cc67a96bfe5fb030630ad832da489a04b6bc --- /dev/null +++ b/2024/PapMOT_ Exploring Adversarial Patch Attack against Multiple Object Tracking/full.md @@ -0,0 +1,260 @@ +# PapMOT: Exploring Adversarial Patch Attack against Multiple Object Tracking + +Jiahuan Long $^{1,2}$ , Tingsong Jiang $^{2}$ , Wen Yao $^{2\star}$ , Shuai Jia $^{1}$ , Weijia Zhang $^{1}$ , Weien Zhou $^{2}$ , Chao Ma $^{1\star}$ , and Xiaojian Chen $^{2}$ + +1 MoE Key Lab of Artificial Intelligence, AI Institute, Shanghai Jiao Tong University 2 Defense Innovation Institute, Chinese Academy of Military Science jiahuanlong@sjtu.edu.cn, tingsong@pku.edu.cn, wendy0782@126.com, jiashuai@sjtu.edu.cn, weijia.zhang@sjtu.edu.cn, weienzhou@outlook.com, chaoma@sjtu.edu.cn, chenxiaoqian@nudt.edu.cn + +Abstract. Tracking multiple objects in a continuous video stream is crucial for many computer vision tasks. It involves detecting and associating objects with their respective identities across successive frames. Despite significant progress made in multiple object tracking (MOT), recent studies have revealed the vulnerability of existing MOT methods to adversarial attacks. Nevertheless, all of these attacks belong to digital attacks that inject pixel-level noise into input images, and are therefore ineffective in physical scenarios. To fill this gap, we propose PapMOT, which can generate physical adversarial patches against MOT for both digital and physical scenarios. Besides attacking the detection mechanism, PapMOT also optimizes a printable patch that can be detected as new targets to mislead the identity association process. Moreover, we introduce a patch enhancement strategy to further degrade the temporal consistency of tracking results across video frames, resulting in more aggressive attacks. We further develop new evaluation metrics to assess the robustness of MOT against such attacks. Extensive evaluations on multiple datasets demonstrate that our PapMOT can successfully attack various architectures of MOT trackers in digital scenarios. We also validate the effectiveness of PapMOT for physical attacks by deploying printed adversarial patches in the real world. + +Keywords: Physical adversarial patches $\cdot$ Multiple object tracking $\cdot$ Evaluation metrics + +# 1 Introduction + +Multiple object tracking (MOT) becomes increasingly widespread with numerous applications including autonomous driving [5] and intelligent surveillance systems [22]. Meanwhile, recent studies [17, 20] have shown that deep multiple object tracking algorithms are susceptible to attacks by adversarial examples. + +![](images/fb3c2c34a465e45fbc9ca0a0f4fd3374fc6853478302effb70d2538ad8d55eaa.jpg) +(a) Digital attack +(b) Previous patch attack +(c) Our patch attack +Fig. 1: A comparison of various adversarial attack methods against MOT. (a) shows a digital attack [20] against MOT, where pixel-level noise is added to each frame and the identities of two people are switched within ten frames. (b) shows a patch attack [28] that fools only the detectors of MOT, generating new fake identities. (c) demonstrates that our patch attack simultaneously fools both the detectors and trackers of MOT, which creates new identities and alters the identities of persons as well. + +These examples are designed to fool MOT models by adding imperceptible perturbations to video sequences. If an attack is successful, the identities of detected objects are no longer associated with their original identities. Existing adversarial attacks [17,20] against MOT have mainly focused on digital attacks, manipulating the pixels of the entire image to fool the trackers. Although these digital attacks can craft imperceptible perturbations, they are generally ineffective in physical scenarios. For instance, Figure 1(a) illustrates an adversarial instance [20] against MOT that cannot be physically implemented. Compared to digital attacks, launching attacks against MOT in the physical world is more challenging. It requires strong perturbations that can withstand diverse environmental conditions, such as illumination and geometric alterations. Also, physical attacks are inherently more challenging to detect and defend against, making them more significant threats to detectors and trackers. Therefore, developing a physical attack against MOT is of great importance to the disclosure and understanding of MOT systems' vulnerabilities in real-world applications. + +Physical adversarial patches are the most common form of physical attacks in the existing literature. They are widely utilized to gauge the reliability and robustness of computer vision systems, such as for face recognition [12,35] and object detection [9,13]. However, applying these methods to MOT systems presents new technical challenges. A major requirement of successful MOT systems is to maintain object identity consistency across frames. This requires the detector to identify targets in each frame and the tracker to associate these targets frame-by-frame. Simply applying adversarial patches generated by detector-based attacks to the MOT system does not sufficiently disrupt data associations across frames. As shown in Figure 1(b), applying the patch attack [28] can create new fake identities but fails to alter the identity of the real attack object with "ID = 1". Moreover, the existing evaluation metrics of MOT attacks are not enough to reflect the performance of patch attacks. Existing metrics neglect the importance of the detection phase, which is a critical component of MOT algorithms. + +Also, these metrics only focus on the effects on one or a few targets, ignoring how attacks on these targets could affect the whole system. + +To address these problems, we propose PapMOT, a physical adversarial patch scheme against MOT systems. PapMOT generates printable adversarial patches that create new identities in the tracking phase, thereby disrupting the existing identity associations. As depicted in Figure 1(c), PapMOT deceives high-performance detectors into outputting false detection boxes (i.e., blue detection boxes), overlapping with true detection boxes. This overlap presents a significant threat to the tracking system, causing it to mistakenly alter the identities of targets in a few frames. To this end, we introduce three adversarial losses that respectively consider the size of bounding boxes, scores of detection boxes, and the smoothness of color transitions. Besides, we propose a patch enhancement strategy that crops targets over time to obtain an extra dataset. The joint training of the original training dataset and the clip dataset will increase the attack success rate on specific targets. To better evaluate the effectiveness of the attack, we construct maximum-attack-threshold evaluation metrics that integrate detection and tracking to comprehensively evaluate the performance of MOT attacks. We summarize our contributions as follows. + +- To the best of our knowledge, this is the first work to explore physical adversarial attack against MOT. The proposed PapMOT attack is able to produce printable patches that successfully disrupt the data association process in tracking. +- Upon delving into the association algorithm in MOT, we propose a novel bounding box restriction loss that generates a wrong detection box with a high IoU value, leading to incorrect identity assignment. +- We propose a new set of metrics to evaluate attacks against MOT that is more comprehensive than previous MOT attacks. +- We evaluate PapMOT in both digital and physical scenarios and demonstrate its favorable attack performance compared to state-of-the-art approaches across multiple datasets. + +# 2 Related Work + +# 2.1 Multiple Object Tracking + +Existing MOT methods can be classified into two main categories: Tracking-by-Detection [4, 22, 31, 38] and Joint Detection and Tracking [39, 40]. In the former paradigm, object detectors are used to detect targets in each frame, and data association algorithms are then employed to link the detection results across video frames, including ByteTrack [38], DeepSORT [31], and OCSort [4]. ByteTrack [38] categorizes all detections into high and low scores and associates them with predicted boxes using the Kalman filter. This improves tracking performance in cases of occlusion and motion blur problems. In the latter paradigm, object detection and data association tasks are integrated into a single framework. FairMOT [39] introduces an anchor-free detection head, which allows its + +Table 1: A comparison of representative adversarial attack methods. + +
MethodsNatPatch [12]ObjPatch [28]AdvTshirt [34]Trasw [20]Hijack [17]PapMOT (ours)
Advanced detectors✓
Detectors & trackers✓✓✓
Diverse datasets✓✓✓✓
Physical attack✓✓✓✓
New evaluation metrics✓
+ +network to predict the position and size of the targets and generate a unique ID feature to achieve data association. In this work, we use ByteTrack [38] and FairMOT [39] as representative trackers to validate the effectiveness and generalization of the proposed adversarial patch attack. + +# 2.2 Adversarial Patch Attacks + +Previous digital attacks [14-16, 33] are difficult to implement in physical scenarios as the perturbations are imperceptible for visual systems. To achieve physical attack, Thys et al. [28] generate physically feasible adversarial patches for deceiving person detectors. Expanding upon this approach, Hu et al. [13] develop AdvTexture by printing the patch on a large cloth and making different T-shirts, skirts, and dresses to realize physical adversarial attacks on pedestrian detection. These two schemes enhance the robustness of patches in the real world by involving patch transformation [1]. However, they mainly conduct attacks on lower-precision detectors such as YOLOv2 [24] and YOLOv3 [25]. For attacks on consecutive frames, Jia et al. [17] propose a tracker hijacking attack scheme, which allows adversarial samples to successfully add or remove objects within a single frame, potentially posing a safety hazard for autonomous vehicles. Lin et al. [20] present a Tracklet-Switch method, which introduces noise into certain frames, causing an ID-switch when pedestrians cross paths with each other. Yet, to our best knowledge, there has been no successful physical attack against MOT systems. + +To fill this research gap, we propose PapMOT, the first physically realizable attack method against MOT, alongside a set of newly proposed evaluation metrics. Different from all existing works, PapMOT is capable of not only attacking more advanced detectors (i.e., YOLOX), but also crippling both the detector and the tracker of an MOT system simultaneously and across a diversity of datasets and scenarios (i.e., MOT15, MOT17, MOT20, and BDD100K). Our contributions are expected to bring new insights into existing research on the digital and physical robustness of MOT systems. A comparison of the features and contributions of our method against prior works is provided in Table 1. + +![](images/de94d365eee5c5fe0ffb9dedfd66400ddbcb17d3b66df115f1356b457714e15e.jpg) +Fig. 2: The overall framework of PapMOT. (a) In the patch training phase, the objective is to optimize an adversarial patch to achieve effective attacks in both digital and physical domains. During this phase, Expectation over Transformation (EOT) is employed as data augmentation for patches. (b) Subsequently, the well-trained adversarial patch is deployed in the patch attack phase to disrupt object detection and data association in MOT. + +# 3 Methodology + +# 3.1 Overall Framework of PapMOT + +An MOT system typically consists of a detection module and an association module. The association module generally employs IoU matching to maintain consistent IDs of objects across consecutive frames, which is at the core of high-performance MOT. Provided high IoU values of a detection box across adjacent frames, this box is likely to be assigned the same ID across these frames. As such, the core objective of our PapMOT is to train printable patches that can induce false detection boxes with a high IoU value with a correctly detected pedestrian box, leading to incorrect identity assignment of this very detection. The entire pipeline of the proposed PapMOT attack is depicted in Figure 2. + +Patch Training Phase. The patch training phase is illustrated in Figure 2(a). First, we randomly initialize a patch and employ a data enhancement strategy (i.e., EOT [1]) to enhance its robustness. EOT incorporates potential patch transformations such as rotation, brightness, blurring, and scaling that commonly exist in the physical world into patch training. Such transformations make our patch attacks more feasible in real-world situations. Next, we embed the EOT-enhanced patch into clean videos to obtain an adversarial dataset, and crop out pedestrian images based on their bounding boxes to produce an adversarial clip dataset. Finally, we combine these two adversarial datasets to optimize the physical adversarial patch using three losses. We use total variation loss $(\mathcal{L}_{tv})$ to ensure patch smoothness and printability, bounding box restric + +tion loss $(\mathcal{L}_{bbr})$ to minimize the discrepancy between the detection boxes of the patch and the object, and average precision loss $(\mathcal{L}_{ap})$ to decrease the detection confidence of targets with the patch. + +Patch Attack Phase. Our patch can deceive detectors into outputting false detection boxes, thereby disrupting the identity association in tracking. This specific attack process is diagrammed in Figure 2(b). The adversarial patch, already well-trained during the patch training phase, is embedded into a video to create an attack video. Next, the attack video is processed frame-by-frame throughYOLOX, which outputs detection boxes for each frame. To clarify the disruption of data association, we consider frames $t - 1$ and $t$ as illustrative examples. Assuming that there is a pedestrian with "id=1" in frame $t - 1$ , the Kalman filter [30] outputs a predicted detection box (i.e., blue dashed-line box) by inputting the current motion states of the pedestrian. In the next frame $t$ , two types of detection boxes appear, with one being a false detection box of the patch (i.e., green solid-line box) and the other being the detection box of the pedestrian (i.e., blue solid-line box). These two detection boxes do Intersection over Union (IoU) calculation with the predicted detection box from frame $t - 1$ . Following this, the Hungarian algorithm [18] uses these IoU values to assign identities to the detection boxes, leading the disruption of data association. (i.e., The false detection box of the patch in frame $t$ is matched with the identity of pedestrians from frame $t - 1$ ). + +# 3.2 Adversarial Loss for Patch Attack + +For our loss formulation, we adopt the widely used average precision loss and total variation loss [12, 27, 29] to reduce the confidence score of detection while encouraging visual smoothness of the printable patches. However, only using these two losses cannot disrupt the identity association in tracking, as illustrated Figure 1(b). This motivates us to propose a novel bounding box restriction loss, which generates a wrong detection box with a high IoU value, leading to incorrect identity assignment and successfully disrupted data association. + +Bounding Box Restriction Loss $(\mathcal{L}_{bbr})$ . To effectively disrupt identity association across frames, it is essential to ensure that the sizes of the target detection box and the patch detection box are similar. To this end, a bounding box restriction loss $\mathcal{L}_{bbr}$ is formulated to restrict the height and width of the detection box, while minimizing the discrepancy between the detection boxes of the patch and the object. $\mathcal{L}_{bbr}$ is defined as: + +$$ +\mathcal {L} _ {b b r} = \frac {1}{N} \sum_ {i = 1} ^ {N} \left(w _ {i} / I _ {w i d t h} + h _ {i} / I _ {h e i g h t}\right) + \left(1 - \frac {1}{N} \sum_ {i = 1} ^ {N} I o U \left(P _ {i}, T _ {i}\right)\right), \tag {1} +$$ + +where $N$ is the total number of attacked objects; $w_{i}$ and $h_{i}$ are the width and height of predicted bounding boxes of the $i$ -th target; and $I_{width}$ and $I_{height}$ are the width and height of video frames that are used for normalization; detection + +boxes of the patch and the target are denoted as $P_{i}$ and $T_{i}$ , respectively. It is worth noting that the first term seems to make $T_{i}$ have zero width and height, but we empirically find that it do not completely vanish and even create more new detection boxes due to the balance of IoU losses. + +Total Variation Loss $(\mathcal{L}_{tv})$ . The aim of $\mathcal{L}_{tv}$ is to ensure the generation of printable patches with smooth color transitions, which helps to reduce the presence of visual noise. $\mathcal{L}_{tv}$ is formulated as: + +$$ +\mathcal {L} _ {t v} = \sum_ {i, j} \sqrt {\left(\left(I _ {i , j} - I _ {i + 1 , j}\right) ^ {2} + \left(I _ {i , j} - I _ {i , j + 1}\right) ^ {2}\right)}, \tag {2} +$$ + +where $I_{i,j}$ denotes the pixel value at position $(i,j)$ in the patch. + +Average Precision Loss $(\mathcal{L}_{ap})$ . It is essential to lower the confidence of the original detection box in order to have it swapped to the one induced by the patch. $\mathcal{L}_{ap}$ does so exactly by reducing the average detection score of targets that have been attached to an adversarial patch. Given a set of $N$ specific detection boxes, where the detection score for the $i$ -th box is $s_i$ , $\mathcal{L}_{ap}$ is calculated as the sum of detection scores divided by the total number of detection boxes as follows: + +$$ +\mathcal {L} _ {a p} = \frac {1}{N} \sum_ {i = 1} ^ {N} s _ {i}. \tag {3} +$$ + +Finally, the overall adversarial loss $L_{adv}$ is a sum of the above three loss terms weighted by hyperparameters $\beta$ , $\gamma$ , and $\delta$ : + +$$ +\mathcal {L} _ {a d v} = \beta \mathcal {L} _ {b b r} + \gamma \mathcal {L} _ {t v} + \delta \mathcal {L} _ {a p}, \tag {4} +$$ + +# 3.3 Evaluation Metrics for MOT Attack + +Current attack methods against MOT [17, 20] have limitations in their evaluation since they utilize identity-based evaluation metrics. These metrics have two main drawbacks. Firstly, they are primarily designed for gauging tracking performance, thereby largely overlooking the quality of detection - the other core ingredient in successful MOT. Secondly, they tend to focus only on a few specific targets, disregarding a comprehensive evaluation of the entire MOT system. + +As depicted in Figure 3(a), identity-based evaluation metrics [17,20] evaluate the identity of the victim in every frame, which considers tracking. However, existing MOT algorithms typically include both tracking and detection modules. Therefore, while evaluating attacks on MOT, it is essential to calculate the performance of both tracking and detection modules simultaneously. Furthermore, identity-based evaluation metrics usually adopt a strategy that aims to attack one or a few specific targets. This approach overlooks the cumulative impact of attacking these specific targets on the entire system. For instance, as shown in Figure 3(b), an attack on the victim (in the red detection box) might also affect + +Fig. 3: A comparison of existing and our proposed evaluation metrics for MOT attacks. +![](images/84ed199bf3cf9c7b6a41af4164bee7d98d3abe0959b8081e55ce9fc9f5f68071.jpg) +(a) Previous evaluation metrics are identity-based and only consider ID association. +(b) Our proposed metrics emphasize the attack evaluation of both detection and ID association. + +the detection of other pedestrians (in yellow boxes), leading to switches of their identities or a decrease in detection accuracy. + +To overcome these limitations, we construct a maximum-attack-threshold evaluation metric that integrates detection and tracking. This type of evaluation metric is based on the classic evaluation metrics for MOT [3]. It provides a comprehensive evaluation of the robustness of MOT systems. + +Tracking Attack Success Rate (TASR). To assess the impact of specific targets during attacks on the overall system, TASR determines the change in MOTA and IDF1 before and after attacks and then divides it by the percentage of bounding boxes that are attacked $(R_{bbox})$ . A high TASR indicates that the MOT system is highly susceptible to particular attacks, enabling an attacker to degrade its performance by compromising fewer targets. A low TASR implies that the attacker must compromise a greater number of targets to inflict considerable damage. The formula for TASR is: $TASR = (MOTA + IDF1)_{\text{decline}} / 2R_{bbox}$ , where $R_{bbox}$ is the ratio of the number of bounding boxes under attack to the total bounding boxes. + +Identity Obfuscation Rate (IOR). To measure the effectiveness of adversarial attacks to cause identity obfuscation during data association in MOT, we define the Identity Obfuscation Rate as: $IOR = D_s / P_t$ , where $P_t = N / T$ , where $T$ represents the number of frames that are attacked consecutively, $N$ is the total number of frames of the video, $P_t$ denotes the maximum number of identity switches in the video, and $D_s$ is the number of identity switches throughout the video sequences. + +Specific Target Attack Success Rate (STASR). To evaluate attacks on specific targets in terms of both false positive rate (FP) and the number of identity switches, STASR focuses not only on how the attack interferes with the detection of the targets but also on how the attack disrupts the tracking of the targets. STASR is defined as: $STASR = (FP + ID_{switch})_{increase} / A_{max}$ , where + +$A_{max} = P_t + P_n$ , $P_t$ represents the maximum number of identity switches in the video, and $P_n$ denotes the number of detection boxes that are attacked. + +# 4 Experiments + +In this section, we perform extensive empirical analysis of the proposed PapMOT attack on MOT in both digital and physical environments. We also conduct ablation studies on the adversarial loss and variants of adversarial patches. + +# 4.1 Experimental Setup + +Datasets and Metrics. We evaluate our attack on four popular MOT benchmarks: MOT15 [19], MOT17 [23], MOT20 [7], and BDD100K [36]. The target objects of the first three datasets are pedestrian-only, while BDD100K further includes vehicle objects. We adopt our proposed evaluation metrics, namely TASR, SATSR, and IOR, to assess the performance of our PapMOT against state-of-the-art methods on the MOT task. More comparisons are found in Appendix. + +Baselines. Existing MOT attack methods can be divided into three categories: 1) methods that attack detectors in the MOT system by injecting pixel-level noise, e.g., RAnoise [21] and TOG [6]. 2) methods that attack detectors in the MOT system by placing optimized patches, e.g., AdvTexture [13] and ObjPatch [28]. 3) methods that attack both detectors and trackers in the MOT system by injecting pixel-level noise, e.g., Hijack [17] and Trasw [20]. Our PapMOT introduces a physically realizable adversarial patch into the MOT system to attack both detectors and trackers. + +Victim Models. We choose two representative MOT methods as victim models in our experiments: ByteTrack [38] that represents the tracking-by-detection paradigm and FairMOT [39] that represents the joint detection and tracking paradigm. Besides, ByteTrack is equipped with YOLOX [11], a more advanced version of theYOLO detector family trained on CrowdHuman [26], MOT17 [23], CityPerson [37], and ETHZ [10] datasets. The pre-trained model of FairMOT is provided by PaddleDetection [2], and is trained on Caltech Pedestrian [8], CityPersons [37], CUHK-SYSU [32], PRW [41], ETHZ [10], and MOT17 [23] datasets. + +Implementation Details. In our digital experiments, to strike a balance between the stealthiness and effectiveness of attacks, we set the noise standard deviation of RAnoise to 10, and limit the maximum disturbance range in TOG to 8/255. The patch sizes for both AdvTexture and ObjPatch are set to one-third of the targeted detection box's size. We define an object as "attacked" if it overlaps with another object with an IOU greater than 0.2. In our physical experiments, we print the patches on foam boards with a fixed width and height of $50\mathrm{cm}$ . The size ratio between the patch and an adult who is $1.6\mathrm{m}$ tall is kept at + +Table 2: Comparison with other representative adversarial attack methods on the MOT15 and MOT17 datasets. + +
DatasetMOT15MOT17
VictimFairMOTByteTrackFairMOTByteTrack
MetricsTASRSATSRIORTASRSATSRIORTASRSATSRIORTASRSATSRIOR
RAnoise [21]0.70.6-0.60.3-0.30.5--0.2-
TOG [6]13.77.04.111.46.33.89.14.82.54.73.91.2
AdvTexture [13]36.435.64.638.143.75.845.432.15.142.247.76.0
ObjPatch [28]45.240.66.343.759.38.261.349.56.261.476.77.4
Hijack [17]14.218.539.311.415.929.010.019.226.58.415.425.4
Trasw [20]40.645.866.231.439.255.422.128.944.821.026.842.3
Ours76.485.062.385.379.265.976.893.046.294.689.679.5
+ +around $30\%$ , ensuring that essential areas such as the face are not obstructed to maintain physical feasibility. For hyperparameters in the overall adversarial loss (i.e., Eq. 4), we set $\beta = 1$ , $\gamma = 2.5$ , and $\delta = 2$ . Datasets used in our patch generation consist of ETHZ, MOT17, and MOT17 clip datasets. + +# 4.2 Digital Domain Performance + +Table 2 and Table 3 compare the performance of various attacks on the MOT15, MOT17, MOT20, and BDD100K datasets. As can be seen, PapMOT achieves the highest attack success rate across all four datasets with a $50\%$ bounding box attack, manifesting its superiority over existing methods (i.e., RAnoise [21], TOG [6], ADVTexure [13], ObjPatch [28], Hijack [17], Trasw [20]) against MOT in both pedestrian (i.e., MOT15, MOT17, and MOT20) and autonomous driving scenarios (i.e., BDD100K). We notice that the IOR of PapMOT is marginally below that of Trasw on the MOT15 and MOT20 datasets, which is possibly due to Trasw's sole emphasis on attacking data association. Nonetheless, PapMOT surpasses Trasw by a huge margin in terms of TASR and SATSR. Overall, Trasw as a digital method is hard to be implemented in the physical world, and its attack is strictly triggered only when an object overlaps with the attack object. Thus, PapMOT outperforms Trasw in terms of functionality and overall attack performance. + +The superior performance of PapMOT is most evident on the MOT17 dataset, where it consistently surpasses all competing methods across all three metrics when applied against both FairMOT and ByteTrack. The success of PapMOT is largely due to incorporating the MOT17 clip dataset and MOT17 into its patch training dataset (more details about the clip dataset are found in Appendix). It is noteworthy that TASR and SATSR both incorporate the evaluation of detection box performance, whereas IOR predominantly assesses identity association across frames. As such, it is observed that TASR and SATSR are typically higher for patch-based attacks and digital attacks compared to IOR (e.g., when attacking ByteTrack in MOT17, the IOR of AdvTexture is 6.0, while TASR and + +Table 3: Comparison with other representative adversarial attack methods on the MOT20 and BDD100K datasets. + +
DatasetMOT20BDD100K
VictimFairMOTByteTrackFairMOTByteTrack
MetricsTASRSATSRIORTASRSATSRIORTASRSATSRIORTASRSATSRIOR
RAnoise [21]2.60.90.44.11.50.42.20.7-1.90.7-
TOG [6]8.62.50.67.62.30.618.16.72.413.54.82.2
AdvTexture [13]23.518.52.932.026.73.221.112.64.418.319.04.8
ObjPatch [28]42.321.44.441.327.56.826.319.45.220.523.37.6
Hijack [17]17.918.436.815.934.727.07.19.810.57.48.213.5
Trasw [20]31.233.567.026.628.564.212.217.423.311.013.119.3
Ours61.949.556.658.541.653.946.454.734.163.461.238.8
+ +SATSR in ADvTexture are 42.2 and 47.7, respectively). This observation can be accounted for by the fact that these attacks primarily target detectors, which significantly affect the performance of detection boxes, but minimally impact identity association across frames. + +# 4.3 Physical Domain Performance + +While our patches have shown outstanding performance in the digital domain, translating this success to the physical world is challenging due to the ever-changing of physical environments. In this section, we qualitatively evaluate the effectiveness of our adversarial patches, and further probe their limits under diverse physical environments. + +Physical Validation. We train three patches with different datasets (i.e., ETHZ, MOT17, MOT20) for FairMOT and evaluate them against the ByteTrack detector. Figure 4 showcases the demonstration of these adversarial patches for physical-world attacks under diverse scenes (refer to Supplementary Materials for a video demo). + +In Figure 4 (a), an adversarial patch is attached to a pedestrian, targeting a single object. In the absence of our attack, the target's ID detected by the tracker is consistently 1, and the detected bounding box is accurate. When the attack is imposed, the detector produces an additional false detection box (in blue), and the accuracy of the initially accurate detection box deteriorates significantly. Moreover, the tracker is quickly deprived of the ability to identity associations across frames, and an ID switch takes place within 30 frames. These outcomes validate the effectiveness of our attack in single-object scenarios. + +Figure 4 (b) illustrates a multi-object scenario, where the adversarial patch is attached to one of the two pedestrians in the scene. When the pedestrian in white with the patch (denoted as "target A") passes by in front of the other subject in black ("target B"), their detected identities are successfully maintained in the absence of the attack. When attacked, however, the tracker mistakenly + +![](images/95c589e9cf0162945ddab5276d582e17531acbe28c5ae6f4bb626f69e7fcaec6.jpg) +Fig. 4: Black-box transfer attack of different PapMOT patches in the real world. It illustrates the effective attack for the scenarios of a single pedestrian (a), two pedestrians overlap (b), and a pedestrian walking past from behind a fixed patch (c). + +predicts two identities (i.e., 1 and 3) for target A. Also, an ID switch occurs and the accuracy of both boxes drops as target A passes by. This example shows our proposed attack works in multiple-object scenarios and its effect applies beyond the object with the attached patch. + +Figure 4 (c) shows the scenario where a patch is kept stationary and is misidentified as a pedestrian by the tracker. When attacked, a pedestrian passing behind it immediately causes an ID switch to occur (i.e., ID changes from 2 to 1). This example suggests that our adversarial patch is also suitable for static placement and is capable of affecting pedestrians as they pass by. + +Attack Boundary. We examine three key physical conditions that may affect the effectiveness of patch attacks: (a) illumination, (b) distance, and (c) angle, shown in Fig. 5. For illumination, we sweep over the entire spectrum from total darkness to total brightness by adjusting the ISO sensitivity of the camera. We divide the illumination levels into 10 segments between 0 and 1. An illumination value of 0.5 denotes the normal exposure under overcast outdoor conditions with the ISO set to 200; 0 and 1 correspond to full darkness and full brightness, respectively. For distance, we vary the range between the patch and the camera from 0.5 to 10 meters to obtain a better view. For angle, we control the angle of the patch relative to the camera by adjusting the deflection angle of the person and using 5-degree intervals to present the final attack results. The quantitative evaluation results of the attack boundary effectiveness for three different patches are presented in Table 4. + +Based on the experimental results, several important conclusions can be drawn. Firstly, different patches can achieve varying attack effects depending on the physical conditions. Secondly, the distance between the patches and the camera must be maintained within a specific range (0.8m-4.1m) to achieve an effective attack. Thirdly, patch attacks work better in relatively dark environ + +![](images/2f5cc22b3002e2ef1c281fd3995402af56d8a40580f62298d8b61b6d787c4efb.jpg) +Fig. 5: Studies on patch effectiveness at varying illuminations, distances, and angles. + +Table 4: Studies on the effectiveness range of three adversarial patches. + +
DistanceIlluminationPatch SizeAngle
Patch 10.8m-3.1m0.4-0.740cm-100cm0°-25°
Patch 21.2m-4.1m0.5-0.840cm-100cm0°-15°
Patch 31.1m-3.6m0.4-0.840cm-100cm0°-30°
+ +ments, resulting in more false detection boxes. Moreover, a larger patch size has a stronger attack ability. However, to ensure the stealthiness and effectiveness of the patch, the patch size in PapMOT is set to approximately $30\%$ of the target size. Lastly, to ensure effective attack outcomes, the angle of the patch relative to the camera generally needs to be limited to a small range (i.e., within 15 degrees). It may be considered to add rotation transformation in the patch training to enhance the robustness of the patches against rotation. + +# 4.4 Ablation Studies + +We analyze the effectiveness of each adversarial loss term in our proposed training objective. We also conduct real-world experiments to evaluate the robustness against unoptimized patches, including real images and uniform color patches. + +Loss Functions. Table 5 shows how different adversarial losses affect the performance of our proposed on against MOT17. The results indicate that PapMOT without $\mathcal{L}_{bbr}$ performs significantly worse than PapMOT without $\mathcal{L}_{ap}$ in terms of IOR, scoring 7.2 and 52.6, respectively, when attacking ByteTrack. This suggests that the $\mathcal{L}_{bbr}$ loss is critical in disrupting the consistency of data association. Conversely, when attacking FairMOT, PapMOT without $\mathcal{L}_{ap}$ scores substantially lower in TASR and SATSR than PapMOT without $\mathcal{L}_{bbr}$ , with scores of 11.0 and 13.1, respectively, compared to 50.5 and 43.3. This underscores the significant impact of the $\mathcal{L}_{ap}$ loss on detector performance. + +Various Patches. We evaluate the effectiveness of attacking the MOT system using unoptimized patches (e.g., full-white patches, full-black patches, and real image patches). As shown in Figure 6, although the target is partially occluded by (a) the full-white patch or (b) the full-black patch, it is tracked steadily + +![](images/24c0579cb70f138ff9246131c7fcdef2593151fa31c4c96e193d2f5c619bec25.jpg) +(a) Full-white patch + +![](images/07bcf3370924edbbfb8cf79914ed27088acbe16670c23e7bbd4db125b06eeba6.jpg) +(b) Full-black patch +Fig. 6: The attack testing using real image, full-white, and full-black patches on the MOT system. These patches have the same size of $50 \times 50\mathrm{cm}$ as ours. + +![](images/2b9a8610d269ad49cc294f3110cc21bf1c8810ff176494313f689dee3cc0e81c.jpg) +(c) Real image patch + +Table 5: The effectiveness analysis of the adversarial losses. + +
VictimAttackerMetrics
TASRSATSRIOR
ByteTrack [38]PapMOT w/o Lbbr61.364.77.2
PapMOT w/o Lap21.926.852.6
PapMOT94.689.679.5
FairMOT [39]PapMOT w/o Lbbr50.543.35.6
PapMOT w/o Lap11.013.129.3
PapMOT76.893.046.2
+ +and the ID remains unchanged throughout the whole video sequence. Similarly, although the target is partially occluded by (c) the real-image patch, it is tracked steadily and the ID is maintained at 1 throughout the entire video sequence. The experiments indicate that it is difficult to attack MOT systems with adversarial patches that are not optimized. + +# 5 Conclusion + +In this paper, we introduce a novel physical adversarial attack method named PapMOT against multiple object tracking systems. Unlike all previous attacks that inject noise into input video frames in the digital domain, our approach optimizes printable adversarial patches that effectively disrupt the data association of MOT trackers across frames in the actual physical world. We demonstrate the feasibility and effectiveness of PapMOT in both digital and physical scenarios through extensive experiments on multiple datasets. In future work, we plan to investigate the creation of more natural-looking adversarial patches, such as transforming these patches into clothing to deceive the MOT system. Additionally, we will expand physical adversarial MOT attacks from pedestrian scenarios to autonomous driving scenarios. + +Acknowledgements. This work was supported in part by NSFC (62322113, 62376156). + +# References + +1. Athalye, A., Engstrom, L., Ilyas, A., Kwok, K.: Synthesizing robust adversarial examples. In: International conference on machine learning (2018) 4, 5 +2. Authors, P.: Paddledetection, object detection and instance segmentation toolkit based on paddlepaddle. https://github.com/PaddlePaddle/PaddleDetection (2019) 9 +3. Bernardin, K., Stiefelhagen, R.: Evaluating multiple object tracking performance: the clear mot metrics. EURASIP Journal on Image and Video Processing (2008) 8 +4. Cao, J., Weng, X., Khirodkar, R., Pang, J., Kitani, K.: Observation-centric sort: Rethinking sort for robust multi-object tracking. arXiv preprint arXiv:2203.14360 (2022) 3 +5. Chiu, H.K., Li, J., Ambrus, R., Bohg, J.: Probabilistic 3d multi-modal, multi-object tracking for autonomous driving. In: IEEE International Conference on Robotics and Automation (2021) 1 +6. Chow, K.H., Liu, L., Gursoy, M.E., Truex, S., Wei, W., Wu, Y.: Understanding object detection through an adversarial lens. In: European Symposium on Research in Computer Security (2020) 9, 10, 11 +7. Dendorfer, P., Rezatofighi, H., Milan, A., Shi, J., Cremers, D., Reid, I., Roth, S., Schindler, K., Leal-Taixe, L.: Mot20: A benchmark for multi object tracking in crowded scenes. arXiv preprint arXiv:2003.09003 (2020) 9 +8. Dollar, P., Wojek, C., Schiele, B., Perona, P.: Pedestrian detection: A benchmark. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (2009) 9 +9. Du, A., Chen, B., Chin, T.J., Law, Y.W., Sasdelli, M., Rajasegaran, R., Campbell, D.: Physical adversarial attacks on an aerial imagery object detector. In: Proceedings of the IEEE/CVF Winter Conference on Applications of Computer Vision (2022) 2 +10. Ess, A., Leibe, B., Schindler, K., van Gool, L.: A mobile vision system for robust multi-person tracking. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (2008) 9 +1. Ge, Z., Liu, S., Wang, F., Li, Z., Sun, J.: Yolox: Exceeding yolo series in 2021. arXiv preprint arXiv:2107.08430 (2021) 9 +2. Hu, Y.C.T., Kung, B.H., Tan, D.S., Chen, J.C., Hua, K.L., Cheng, W.H.: Naturalistic physical adversarial patch for object detectors. In: Proceedings of the IEEE/CVF International Conference on Computer Vision (2021) 2, 4, 6 +3. Hu, Z., Huang, S., Zhu, X., Sun, F., Zhang, B., Hu, X.: Adversarial texture for fooling person detectors in the physical world. In: Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition (2022) 2, 4, 9, 10, 11 +4. Jia, S., Ma, C., Song, Y., Yang, X.: Robust tracking against adversarial attacks. In: European Conference on Computer Vision (2020) 4 +5. Jia, S., Ma, C., Yao, T., Yin, B., Ding, S., Yang, X.: Exploring frequency adversarial attacks for face forgery detection. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (2022) 4 +6. Jia, S., Song, Y., Ma, C., Yang, X.: Iou attack: Towards temporally coherent blackbox adversarial attack for visual object tracking. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (2021) 4 +7. Jia, Y., Lu, Y., Shen, J., Chen, Q.A., Zhong, Z., Wei, T.: Fooling detection alone is not enough: Adversarial attack against multiple object tracking. In: International Conference on Learning Representations (2020) 1, 2, 4, 7, 9, 10, 11 + +18. Kuhn, H.W.: The hungarian method for the assignment problem. Naval research logistics quarterly 2(1-2), 83-97 (1955) 6 +19. Leal-Taixe, L., Milan, A., Reid, I., Roth, S., Schindler, K.: Motchallenge 2015: Towards a benchmark for multi-target tracking. arXiv preprint arXiv:1504.01942 (2015) 9 +20. Lin, D., Chen, Q., Zhou, C., He, K.: Trasw: Tracklet-switch adversarial attacks against multi-object tracking. arXiv preprint arXiv:2111.08954 (2023) 1, 2, 4, 7, 9, 10, 11 +21. Luisier, F., Blu, T., Unser, M.: Image denoising in mixed poisson-gaussian noise. IEEE Transactions on image processing (2010) 9, 10, 11 +22. Luo, W., Xing, J., Milan, A., Zhang, X., Liu, W., Kim, T.K.: Multiple object tracking: A literature review. Artificial intelligence (2021) 1, 3 +23. Milan, A., Leal-Taixe, L., Reid, I., Roth, S., Schindler, K.: Mot16: A benchmark for multi-object tracking. arXiv preprint arXiv:1603.00831 (2016) 9 +24. Redmon, J., Farhadi, A.: Yolo9000: Better, faster, stronger. In: IEEE Conference on Computer Vision and Pattern Recognition (2017) 4 +25. Redmon, J., Farhadi, A.: Yolov3: An incremental improvement. arXiv preprint arXiv:1804.02767 (2018) 4 +26. Shao, S., Zhao, Z., Li, B., Xiao, T., Yu, G., Zhang, X., Sun, J.: Crowdhuman: A benchmark for detecting human in a crowd. arXiv preprint arXiv:1805.00123 (2018) 9 +27. Sharif, M., Bhagavatula, S., Bauer, L., Reiter, M.K.: Accessorize to a crime. In: Proceedings of the 2016 ACM SIGSAC Conference on Computer and Communications Security (2016) 6 +28. Thys, S., Van Ranst, W., Goedemé, T.: Fooling automated surveillance cameras: adversarial patches to attack person detection. In: Proceedings of the IEEE/CVF conference on computer vision and pattern recognition workshops (2019) 2, 4, 9, 10, 11 +29. Wang, D., Jiang, T., Sun, J., Zhou, W., Gong, Z., Zhang, X., Yao, W., Chen, X.: Fca: Learning a 3d full-coverage vehicle camouflage for multi-view physical adversarial attack. In: Proceedings of the AAAI conference on artificial intelligence (2022) 6 +30. Welch, G., Bishop, G., et al.: An introduction to the kalman filter (1995) 6 +31. Wojke, N., Bewley, A., Paulus, D.: Simple online and realtime tracking with a deep association metric. In: IEEE International Conference on Image Processing (2017) 3 +32. Xiao, T., Li, S., Wang, B., Lin, L., Wang, X.: Joint detection and identification feature learning for person search. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (2017) 9 +33. Xu, H., Ma, Y., Liu, H.C., Deb, D., Liu, H., Tang, J.L., Jain, A.K.: Adversarial attacks and defenses in images, graphs and text: A review. International Journal of Automation and Computing (2020) 4 +34. Xu, K., Zhang, G., Liu, S., Fan, Q., Sun, M., Chen, H., Chen, P.Y., Wang, Y., Lin, X.: Adversarial t-shirt! evading person detectors in a physical world. In: European Conference on Computer Vision (2020) 4 +35. Yang, X., Wei, F., Zhang, H., Zhu, J.: Design and interpretation of universal adversarial patches in face detection. In: European Conference on Computer Vision (2020) 2 +36. Yu, F., Chen, H., Wang, X., Xian, W., Chen, Y., Liu, F., Madhavan, V., Darrell, T.: Bdd100k: A diverse driving dataset for heterogeneous multitask learning. In: IEEE/CVF Conference on Computer Vision and Pattern Recognition (2020) 9 + +37. Zhang, S., Benenson, R., Schiele, B.: Citypersons: A diverse dataset for pedestrian detection. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (2017) 9 +38. Zhang, Y., Sun, P., Jiang, Y., Yu, D., Weng, F., Yuan, Z., Luo, P., Liu, W., Wang, X.: Bytetrack: Multi-object tracking by associating every detection box. In: European Conference on Computer Vision (2022) 3, 4, 9, 14 +39. Zhang, Y., Wang, C., Wang, X., Zeng, W., Liu, W.: Fairmot: On the fairness of detection and re-identification in multiple object tracking. International Journal of Computer Vision (2021) 3, 4, 9, 14 +40. Zhao, Z., Wu, Z., Zhuang, Y., Li, B., Jia, J.: Tracking objects as pixel-wise distributions. In: European Conference on Computer Vision (2022) 3 +41. Zheng, L., Zhang, H., Sun, S., Chandraker, M., Yang, Y., Tian, Q.: Person re-identification in the wild. In: Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition (2017) 9 \ No newline at end of file diff --git a/2024/PapMOT_ Exploring Adversarial Patch Attack against Multiple Object Tracking/images.zip b/2024/PapMOT_ Exploring Adversarial Patch Attack against Multiple Object Tracking/images.zip new file mode 100644 index 0000000000000000000000000000000000000000..83e5294e0ca3c6fb53f607d6babab20e53ed3af5 --- /dev/null +++ b/2024/PapMOT_ Exploring Adversarial Patch Attack against Multiple Object Tracking/images.zip @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:09ee8cc56257a98e07b14b6a6ff198ddd815f9a6476d443277dd162b23cc2ee4 +size 589810 diff --git a/2024/PapMOT_ Exploring Adversarial Patch Attack against Multiple Object Tracking/layout.json b/2024/PapMOT_ Exploring Adversarial Patch Attack against Multiple Object Tracking/layout.json new file mode 100644 index 0000000000000000000000000000000000000000..94c126b1cdc0a7ad4ca554c093594faabfe81a34 --- /dev/null +++ b/2024/PapMOT_ Exploring Adversarial Patch Attack against Multiple Object Tracking/layout.json @@ -0,0 +1,8032 @@ +{ + "pdf_info": [ + { + "para_blocks": [ + { + "bbox": [ + 145, + 111, + 470, + 148 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 145, + 111, + 470, + 148 + ], + "spans": [ + { + "bbox": [ + 145, + 111, + 470, + 148 + ], + "type": "text", + "content": "PapMOT: Exploring Adversarial Patch Attack against Multiple Object Tracking" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 143, + 168, + 470, + 193 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 143, + 168, + 470, + 193 + ], + "spans": [ + { + "bbox": [ + 143, + 168, + 470, + 193 + ], + "type": "text", + "content": "Jiahuan Long" + }, + { + "bbox": [ + 143, + 168, + 470, + 193 + ], + "type": "inline_equation", + "content": "^{1,2}" + }, + { + "bbox": [ + 143, + 168, + 470, + 193 + ], + "type": "text", + "content": ", Tingsong Jiang" + }, + { + "bbox": [ + 143, + 168, + 470, + 193 + ], + "type": "inline_equation", + "content": "^{2}" + }, + { + "bbox": [ + 143, + 168, + 470, + 193 + ], + "type": "text", + "content": ", Wen Yao" + }, + { + "bbox": [ + 143, + 168, + 470, + 193 + ], + "type": "inline_equation", + "content": "^{2\\star}" + }, + { + "bbox": [ + 143, + 168, + 470, + 193 + ], + "type": "text", + "content": ", Shuai Jia" + }, + { + "bbox": [ + 143, + 168, + 470, + 193 + ], + "type": "inline_equation", + "content": "^{1}" + }, + { + "bbox": [ + 143, + 168, + 470, + 193 + ], + "type": "text", + "content": ", Weijia Zhang" + }, + { + "bbox": [ + 143, + 168, + 470, + 193 + ], + "type": "inline_equation", + "content": "^{1}" + }, + { + "bbox": [ + 143, + 168, + 470, + 193 + ], + "type": "text", + "content": ", Weien Zhou" + }, + { + "bbox": [ + 143, + 168, + 470, + 193 + ], + "type": "inline_equation", + "content": "^{2}" + }, + { + "bbox": [ + 143, + 168, + 470, + 193 + ], + "type": "text", + "content": ", Chao Ma" + }, + { + "bbox": [ + 143, + 168, + 470, + 193 + ], + "type": "inline_equation", + "content": "^{1\\star}" + }, + { + "bbox": [ + 143, + 168, + 470, + 193 + ], + "type": "text", + "content": ", and Xiaojian Chen" + }, + { + "bbox": [ + 143, + 168, + 470, + 193 + ], + "type": "inline_equation", + "content": "^{2}" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 132, + 201, + 481, + 258 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 201, + 481, + 258 + ], + "spans": [ + { + "bbox": [ + 132, + 201, + 481, + 258 + ], + "type": "text", + "content": "1 MoE Key Lab of Artificial Intelligence, AI Institute, Shanghai Jiao Tong University 2 Defense Innovation Institute, Chinese Academy of Military Science jiahuanlong@sjtu.edu.cn, tingsong@pku.edu.cn, wendy0782@126.com, jiashuai@sjtu.edu.cn, weijia.zhang@sjtu.edu.cn, weienzhou@outlook.com, chaoma@sjtu.edu.cn, chenxiaoqian@nudt.edu.cn" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 160, + 292, + 455, + 511 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 160, + 292, + 455, + 511 + ], + "spans": [ + { + "bbox": [ + 160, + 292, + 455, + 511 + ], + "type": "text", + "content": "Abstract. Tracking multiple objects in a continuous video stream is crucial for many computer vision tasks. It involves detecting and associating objects with their respective identities across successive frames. Despite significant progress made in multiple object tracking (MOT), recent studies have revealed the vulnerability of existing MOT methods to adversarial attacks. Nevertheless, all of these attacks belong to digital attacks that inject pixel-level noise into input images, and are therefore ineffective in physical scenarios. To fill this gap, we propose PapMOT, which can generate physical adversarial patches against MOT for both digital and physical scenarios. Besides attacking the detection mechanism, PapMOT also optimizes a printable patch that can be detected as new targets to mislead the identity association process. Moreover, we introduce a patch enhancement strategy to further degrade the temporal consistency of tracking results across video frames, resulting in more aggressive attacks. We further develop new evaluation metrics to assess the robustness of MOT against such attacks. Extensive evaluations on multiple datasets demonstrate that our PapMOT can successfully attack various architectures of MOT trackers in digital scenarios. We also validate the effectiveness of PapMOT for physical attacks by deploying printed adversarial patches in the real world." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 159, + 522, + 453, + 544 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 159, + 522, + 453, + 544 + ], + "spans": [ + { + "bbox": [ + 159, + 522, + 453, + 544 + ], + "type": "text", + "content": "Keywords: Physical adversarial patches " + }, + { + "bbox": [ + 159, + 522, + 453, + 544 + ], + "type": "inline_equation", + "content": "\\cdot" + }, + { + "bbox": [ + 159, + 522, + 453, + 544 + ], + "type": "text", + "content": " Multiple object tracking " + }, + { + "bbox": [ + 159, + 522, + 453, + 544 + ], + "type": "inline_equation", + "content": "\\cdot" + }, + { + "bbox": [ + 159, + 522, + 453, + 544 + ], + "type": "text", + "content": " Evaluation metrics" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 132, + 568, + 230, + 582 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 568, + 230, + 582 + ], + "spans": [ + { + "bbox": [ + 132, + 568, + 230, + 582 + ], + "type": "text", + "content": "1 Introduction" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 130, + 596, + 482, + 645 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 596, + 482, + 645 + ], + "spans": [ + { + "bbox": [ + 130, + 596, + 482, + 645 + ], + "type": "text", + "content": "Multiple object tracking (MOT) becomes increasingly widespread with numerous applications including autonomous driving [5] and intelligent surveillance systems [22]. Meanwhile, recent studies [17, 20] have shown that deep multiple object tracking algorithms are susceptible to attacks by adversarial examples." + } + ] + } + ], + "index": 6 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 133, + 653, + 240, + 665 + ], + "type": "page_footnote", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 653, + 240, + 665 + ], + "spans": [ + { + "bbox": [ + 133, + 653, + 240, + 665 + ], + "type": "text", + "content": "* Corresponding authors." + } + ] + } + ], + "index": 7 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 0 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 136, + 114, + 479, + 195 + ], + "blocks": [ + { + "bbox": [ + 136, + 114, + 479, + 195 + ], + "lines": [ + { + "bbox": [ + 136, + 114, + 479, + 195 + ], + "spans": [ + { + "bbox": [ + 136, + 114, + 479, + 195 + ], + "type": "image", + "image_path": "fb3c2c34a465e45fbc9ca0a0f4fd3374fc6853478302effb70d2538ad8d55eaa.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 210, + 196, + 260, + 204 + ], + "lines": [ + { + "bbox": [ + 210, + 196, + 260, + 204 + ], + "spans": [ + { + "bbox": [ + 210, + 196, + 260, + 204 + ], + "type": "text", + "content": "(a) Digital attack" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 294, + 196, + 369, + 205 + ], + "lines": [ + { + "bbox": [ + 294, + 196, + 369, + 205 + ], + "spans": [ + { + "bbox": [ + 294, + 196, + 369, + 205 + ], + "type": "text", + "content": "(b) Previous patch attack" + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 401, + 196, + 460, + 205 + ], + "lines": [ + { + "bbox": [ + 401, + 196, + 460, + 205 + ], + "spans": [ + { + "bbox": [ + 401, + 196, + 460, + 205 + ], + "type": "text", + "content": "(c) Our patch attack" + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 131, + 213, + 482, + 280 + ], + "lines": [ + { + "bbox": [ + 131, + 213, + 482, + 280 + ], + "spans": [ + { + "bbox": [ + 131, + 213, + 482, + 280 + ], + "type": "text", + "content": "Fig. 1: A comparison of various adversarial attack methods against MOT. (a) shows a digital attack [20] against MOT, where pixel-level noise is added to each frame and the identities of two people are switched within ten frames. (b) shows a patch attack [28] that fools only the detectors of MOT, generating new fake identities. (c) demonstrates that our patch attack simultaneously fools both the detectors and trackers of MOT, which creates new identities and alters the identities of persons as well." + } + ] + } + ], + "index": 6, + "angle": 0, + "type": "image_caption" + } + ], + "index": 2 + }, + { + "bbox": [ + 130, + 313, + 482, + 494 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 313, + 482, + 494 + ], + "spans": [ + { + "bbox": [ + 130, + 313, + 482, + 494 + ], + "type": "text", + "content": "These examples are designed to fool MOT models by adding imperceptible perturbations to video sequences. If an attack is successful, the identities of detected objects are no longer associated with their original identities. Existing adversarial attacks [17,20] against MOT have mainly focused on digital attacks, manipulating the pixels of the entire image to fool the trackers. Although these digital attacks can craft imperceptible perturbations, they are generally ineffective in physical scenarios. For instance, Figure 1(a) illustrates an adversarial instance [20] against MOT that cannot be physically implemented. Compared to digital attacks, launching attacks against MOT in the physical world is more challenging. It requires strong perturbations that can withstand diverse environmental conditions, such as illumination and geometric alterations. Also, physical attacks are inherently more challenging to detect and defend against, making them more significant threats to detectors and trackers. Therefore, developing a physical attack against MOT is of great importance to the disclosure and understanding of MOT systems' vulnerabilities in real-world applications." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 130, + 498, + 482, + 666 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 498, + 482, + 666 + ], + "spans": [ + { + "bbox": [ + 130, + 498, + 482, + 666 + ], + "type": "text", + "content": "Physical adversarial patches are the most common form of physical attacks in the existing literature. They are widely utilized to gauge the reliability and robustness of computer vision systems, such as for face recognition [12,35] and object detection [9,13]. However, applying these methods to MOT systems presents new technical challenges. A major requirement of successful MOT systems is to maintain object identity consistency across frames. This requires the detector to identify targets in each frame and the tracker to associate these targets frame-by-frame. Simply applying adversarial patches generated by detector-based attacks to the MOT system does not sufficiently disrupt data associations across frames. As shown in Figure 1(b), applying the patch attack [28] can create new fake identities but fails to alter the identity of the real attack object with \"ID = 1\". Moreover, the existing evaluation metrics of MOT attacks are not enough to reflect the performance of patch attacks. Existing metrics neglect the importance of the detection phase, which is a critical component of MOT algorithms." + } + ] + } + ], + "index": 8 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "spans": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "type": "text", + "content": "2" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 166, + 91, + 222, + 101 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 166, + 91, + 222, + 101 + ], + "spans": [ + { + "bbox": [ + 166, + 91, + 222, + 101 + ], + "type": "text", + "content": "J. Long et al." + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 1 + }, + { + "para_blocks": [ + { + "bbox": [ + 130, + 116, + 479, + 140 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 116, + 479, + 140 + ], + "spans": [ + { + "bbox": [ + 130, + 116, + 479, + 140 + ], + "type": "text", + "content": "Also, these metrics only focus on the effects on one or a few targets, ignoring how attacks on these targets could affect the whole system." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 130, + 140, + 482, + 331 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 140, + 482, + 331 + ], + "spans": [ + { + "bbox": [ + 130, + 140, + 482, + 331 + ], + "type": "text", + "content": "To address these problems, we propose PapMOT, a physical adversarial patch scheme against MOT systems. PapMOT generates printable adversarial patches that create new identities in the tracking phase, thereby disrupting the existing identity associations. As depicted in Figure 1(c), PapMOT deceives high-performance detectors into outputting false detection boxes (i.e., blue detection boxes), overlapping with true detection boxes. This overlap presents a significant threat to the tracking system, causing it to mistakenly alter the identities of targets in a few frames. To this end, we introduce three adversarial losses that respectively consider the size of bounding boxes, scores of detection boxes, and the smoothness of color transitions. Besides, we propose a patch enhancement strategy that crops targets over time to obtain an extra dataset. The joint training of the original training dataset and the clip dataset will increase the attack success rate on specific targets. To better evaluate the effectiveness of the attack, we construct maximum-attack-threshold evaluation metrics that integrate detection and tracking to comprehensively evaluate the performance of MOT attacks. We summarize our contributions as follows." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 138, + 339, + 480, + 485 + ], + "type": "list", + "angle": 0, + "index": 8, + "blocks": [ + { + "bbox": [ + 138, + 339, + 479, + 386 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 339, + 479, + 386 + ], + "spans": [ + { + "bbox": [ + 138, + 339, + 479, + 386 + ], + "type": "text", + "content": "- To the best of our knowledge, this is the first work to explore physical adversarial attack against MOT. The proposed PapMOT attack is able to produce printable patches that successfully disrupt the data association process in tracking." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 138, + 388, + 480, + 423 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 388, + 480, + 423 + ], + "spans": [ + { + "bbox": [ + 138, + 388, + 480, + 423 + ], + "type": "text", + "content": "- Upon delving into the association algorithm in MOT, we propose a novel bounding box restriction loss that generates a wrong detection box with a high IoU value, leading to incorrect identity assignment." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 138, + 425, + 479, + 448 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 425, + 479, + 448 + ], + "spans": [ + { + "bbox": [ + 138, + 425, + 479, + 448 + ], + "type": "text", + "content": "- We propose a new set of metrics to evaluate attacks against MOT that is more comprehensive than previous MOT attacks." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 138, + 449, + 479, + 485 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 449, + 479, + 485 + ], + "spans": [ + { + "bbox": [ + 138, + 449, + 479, + 485 + ], + "type": "text", + "content": "- We evaluate PapMOT in both digital and physical scenarios and demonstrate its favorable attack performance compared to state-of-the-art approaches across multiple datasets." + } + ] + } + ], + "index": 7 + } + ], + "sub_type": "text" + }, + { + "bbox": [ + 131, + 502, + 237, + 514 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 131, + 502, + 237, + 514 + ], + "spans": [ + { + "bbox": [ + 131, + 502, + 237, + 514 + ], + "type": "text", + "content": "2 Related Work" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 131, + 527, + 289, + 540 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 131, + 527, + 289, + 540 + ], + "spans": [ + { + "bbox": [ + 131, + 527, + 289, + 540 + ], + "type": "text", + "content": "2.1 Multiple Object Tracking" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 130, + 545, + 482, + 666 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 545, + 482, + 666 + ], + "spans": [ + { + "bbox": [ + 130, + 545, + 482, + 666 + ], + "type": "text", + "content": "Existing MOT methods can be classified into two main categories: Tracking-by-Detection [4, 22, 31, 38] and Joint Detection and Tracking [39, 40]. In the former paradigm, object detectors are used to detect targets in each frame, and data association algorithms are then employed to link the detection results across video frames, including ByteTrack [38], DeepSORT [31], and OCSort [4]. ByteTrack [38] categorizes all detections into high and low scores and associates them with predicted boxes using the Kalman filter. This improves tracking performance in cases of occlusion and motion blur problems. In the latter paradigm, object detection and data association tasks are integrated into a single framework. FairMOT [39] introduces an anchor-free detection head, which allows its" + } + ] + } + ], + "index": 11 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 200, + 91, + 448, + 102 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 200, + 91, + 448, + 102 + ], + "spans": [ + { + "bbox": [ + 200, + 91, + 448, + 102 + ], + "type": "text", + "content": "PapMOT: Exploring Adversarial Patch Attack against MOT" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 474, + 92, + 480, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 474, + 92, + 480, + 100 + ], + "spans": [ + { + "bbox": [ + 474, + 92, + 480, + 100 + ], + "type": "text", + "content": "3" + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 2 + }, + { + "para_blocks": [ + { + "type": "table", + "bbox": [ + 134, + 145, + 480, + 236 + ], + "blocks": [ + { + "bbox": [ + 165, + 114, + 448, + 126 + ], + "lines": [ + { + "bbox": [ + 165, + 114, + 448, + 126 + ], + "spans": [ + { + "bbox": [ + 165, + 114, + 448, + 126 + ], + "type": "text", + "content": "Table 1: A comparison of representative adversarial attack methods." + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 134, + 145, + 480, + 236 + ], + "lines": [ + { + "bbox": [ + 134, + 145, + 480, + 236 + ], + "spans": [ + { + "bbox": [ + 134, + 145, + 480, + 236 + ], + "type": "table", + "html": "
MethodsNatPatch [12]ObjPatch [28]AdvTshirt [34]Trasw [20]Hijack [17]PapMOT (ours)
Advanced detectors✓
Detectors & trackers✓✓✓
Diverse datasets✓✓✓✓
Physical attack✓✓✓✓
New evaluation metrics✓
", + "image_path": "18cc9621b959504f7a1f1e6983cbc3d82e29cc6de0517a2622ec2322e4d9c03d.jpg" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "table_body" + } + ], + "index": 3 + }, + { + "bbox": [ + 130, + 270, + 482, + 319 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 270, + 482, + 319 + ], + "spans": [ + { + "bbox": [ + 130, + 270, + 482, + 319 + ], + "type": "text", + "content": "network to predict the position and size of the targets and generate a unique ID feature to achieve data association. In this work, we use ByteTrack [38] and FairMOT [39] as representative trackers to validate the effectiveness and generalization of the proposed adversarial patch attack." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 131, + 341, + 293, + 351 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 131, + 341, + 293, + 351 + ], + "spans": [ + { + "bbox": [ + 131, + 341, + 293, + 351 + ], + "type": "text", + "content": "2.2 Adversarial Patch Attacks" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 130, + 365, + 482, + 555 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 365, + 482, + 555 + ], + "spans": [ + { + "bbox": [ + 130, + 365, + 482, + 555 + ], + "type": "text", + "content": "Previous digital attacks [14-16, 33] are difficult to implement in physical scenarios as the perturbations are imperceptible for visual systems. To achieve physical attack, Thys et al. [28] generate physically feasible adversarial patches for deceiving person detectors. Expanding upon this approach, Hu et al. [13] develop AdvTexture by printing the patch on a large cloth and making different T-shirts, skirts, and dresses to realize physical adversarial attacks on pedestrian detection. These two schemes enhance the robustness of patches in the real world by involving patch transformation [1]. However, they mainly conduct attacks on lower-precision detectors such as YOLOv2 [24] and YOLOv3 [25]. For attacks on consecutive frames, Jia et al. [17] propose a tracker hijacking attack scheme, which allows adversarial samples to successfully add or remove objects within a single frame, potentially posing a safety hazard for autonomous vehicles. Lin et al. [20] present a Tracklet-Switch method, which introduces noise into certain frames, causing an ID-switch when pedestrians cross paths with each other. Yet, to our best knowledge, there has been no successful physical attack against MOT systems." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 130, + 558, + 482, + 666 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 558, + 482, + 666 + ], + "spans": [ + { + "bbox": [ + 130, + 558, + 482, + 666 + ], + "type": "text", + "content": "To fill this research gap, we propose PapMOT, the first physically realizable attack method against MOT, alongside a set of newly proposed evaluation metrics. Different from all existing works, PapMOT is capable of not only attacking more advanced detectors (i.e., YOLOX), but also crippling both the detector and the tracker of an MOT system simultaneously and across a diversity of datasets and scenarios (i.e., MOT15, MOT17, MOT20, and BDD100K). Our contributions are expected to bring new insights into existing research on the digital and physical robustness of MOT systems. A comparison of the features and contributions of our method against prior works is provided in Table 1." + } + ] + } + ], + "index": 7 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "spans": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "type": "text", + "content": "4" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 166, + 91, + 222, + 102 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 166, + 91, + 222, + 102 + ], + "spans": [ + { + "bbox": [ + 166, + 91, + 222, + 102 + ], + "type": "text", + "content": "J. Long et al." + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 3 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 133, + 115, + 481, + 273 + ], + "blocks": [ + { + "bbox": [ + 133, + 115, + 481, + 273 + ], + "lines": [ + { + "bbox": [ + 133, + 115, + 481, + 273 + ], + "spans": [ + { + "bbox": [ + 133, + 115, + 481, + 273 + ], + "type": "image", + "image_path": "de94d365eee5c5fe0ffb9dedfd66400ddbcb17d3b66df115f1356b457714e15e.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 130, + 281, + 482, + 346 + ], + "lines": [ + { + "bbox": [ + 130, + 281, + 482, + 346 + ], + "spans": [ + { + "bbox": [ + 130, + 281, + 482, + 346 + ], + "type": "text", + "content": "Fig. 2: The overall framework of PapMOT. (a) In the patch training phase, the objective is to optimize an adversarial patch to achieve effective attacks in both digital and physical domains. During this phase, Expectation over Transformation (EOT) is employed as data augmentation for patches. (b) Subsequently, the well-trained adversarial patch is deployed in the patch attack phase to disrupt object detection and data association in MOT." + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_caption" + } + ], + "index": 2 + }, + { + "bbox": [ + 132, + 370, + 233, + 384 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 370, + 233, + 384 + ], + "spans": [ + { + "bbox": [ + 132, + 370, + 233, + 384 + ], + "type": "text", + "content": "3 Methodology" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 132, + 396, + 321, + 408 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 396, + 321, + 408 + ], + "spans": [ + { + "bbox": [ + 132, + 396, + 321, + 408 + ], + "type": "text", + "content": "3.1 Overall Framework of PapMOT" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 130, + 417, + 482, + 525 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 417, + 482, + 525 + ], + "spans": [ + { + "bbox": [ + 130, + 417, + 482, + 525 + ], + "type": "text", + "content": "An MOT system typically consists of a detection module and an association module. The association module generally employs IoU matching to maintain consistent IDs of objects across consecutive frames, which is at the core of high-performance MOT. Provided high IoU values of a detection box across adjacent frames, this box is likely to be assigned the same ID across these frames. As such, the core objective of our PapMOT is to train printable patches that can induce false detection boxes with a high IoU value with a correctly detected pedestrian box, leading to incorrect identity assignment of this very detection. The entire pipeline of the proposed PapMOT attack is depicted in Figure 2." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 130, + 533, + 482, + 666 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 533, + 482, + 666 + ], + "spans": [ + { + "bbox": [ + 130, + 533, + 482, + 666 + ], + "type": "text", + "content": "Patch Training Phase. The patch training phase is illustrated in Figure 2(a). First, we randomly initialize a patch and employ a data enhancement strategy (i.e., EOT [1]) to enhance its robustness. EOT incorporates potential patch transformations such as rotation, brightness, blurring, and scaling that commonly exist in the physical world into patch training. Such transformations make our patch attacks more feasible in real-world situations. Next, we embed the EOT-enhanced patch into clean videos to obtain an adversarial dataset, and crop out pedestrian images based on their bounding boxes to produce an adversarial clip dataset. Finally, we combine these two adversarial datasets to optimize the physical adversarial patch using three losses. We use total variation loss " + }, + { + "bbox": [ + 130, + 533, + 482, + 666 + ], + "type": "inline_equation", + "content": "(\\mathcal{L}_{tv})" + }, + { + "bbox": [ + 130, + 533, + 482, + 666 + ], + "type": "text", + "content": " to ensure patch smoothness and printability, bounding box restric" + } + ] + } + ], + "index": 7 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 200, + 91, + 448, + 102 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 200, + 91, + 448, + 102 + ], + "spans": [ + { + "bbox": [ + 200, + 91, + 448, + 102 + ], + "type": "text", + "content": "PapMOT: Exploring Adversarial Patch Attack against MOT" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 474, + 92, + 480, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 474, + 92, + 480, + 100 + ], + "spans": [ + { + "bbox": [ + 474, + 92, + 480, + 100 + ], + "type": "text", + "content": "5" + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 4 + }, + { + "para_blocks": [ + { + "bbox": [ + 130, + 116, + 482, + 152 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 116, + 482, + 152 + ], + "spans": [ + { + "bbox": [ + 130, + 116, + 482, + 152 + ], + "type": "text", + "content": "tion loss " + }, + { + "bbox": [ + 130, + 116, + 482, + 152 + ], + "type": "inline_equation", + "content": "(\\mathcal{L}_{bbr})" + }, + { + "bbox": [ + 130, + 116, + 482, + 152 + ], + "type": "text", + "content": " to minimize the discrepancy between the detection boxes of the patch and the object, and average precision loss " + }, + { + "bbox": [ + 130, + 116, + 482, + 152 + ], + "type": "inline_equation", + "content": "(\\mathcal{L}_{ap})" + }, + { + "bbox": [ + 130, + 116, + 482, + 152 + ], + "type": "text", + "content": " to decrease the detection confidence of targets with the patch." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 130, + 160, + 483, + 376 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 160, + 483, + 376 + ], + "spans": [ + { + "bbox": [ + 130, + 160, + 483, + 376 + ], + "type": "text", + "content": "Patch Attack Phase. Our patch can deceive detectors into outputting false detection boxes, thereby disrupting the identity association in tracking. This specific attack process is diagrammed in Figure 2(b). The adversarial patch, already well-trained during the patch training phase, is embedded into a video to create an attack video. Next, the attack video is processed frame-by-frame throughYOLOX, which outputs detection boxes for each frame. To clarify the disruption of data association, we consider frames " + }, + { + "bbox": [ + 130, + 160, + 483, + 376 + ], + "type": "inline_equation", + "content": "t - 1" + }, + { + "bbox": [ + 130, + 160, + 483, + 376 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 130, + 160, + 483, + 376 + ], + "type": "inline_equation", + "content": "t" + }, + { + "bbox": [ + 130, + 160, + 483, + 376 + ], + "type": "text", + "content": " as illustrative examples. Assuming that there is a pedestrian with \"id=1\" in frame " + }, + { + "bbox": [ + 130, + 160, + 483, + 376 + ], + "type": "inline_equation", + "content": "t - 1" + }, + { + "bbox": [ + 130, + 160, + 483, + 376 + ], + "type": "text", + "content": ", the Kalman filter [30] outputs a predicted detection box (i.e., blue dashed-line box) by inputting the current motion states of the pedestrian. In the next frame " + }, + { + "bbox": [ + 130, + 160, + 483, + 376 + ], + "type": "inline_equation", + "content": "t" + }, + { + "bbox": [ + 130, + 160, + 483, + 376 + ], + "type": "text", + "content": ", two types of detection boxes appear, with one being a false detection box of the patch (i.e., green solid-line box) and the other being the detection box of the pedestrian (i.e., blue solid-line box). These two detection boxes do Intersection over Union (IoU) calculation with the predicted detection box from frame " + }, + { + "bbox": [ + 130, + 160, + 483, + 376 + ], + "type": "inline_equation", + "content": "t - 1" + }, + { + "bbox": [ + 130, + 160, + 483, + 376 + ], + "type": "text", + "content": ". Following this, the Hungarian algorithm [18] uses these IoU values to assign identities to the detection boxes, leading the disruption of data association. (i.e., The false detection box of the patch in frame " + }, + { + "bbox": [ + 130, + 160, + 483, + 376 + ], + "type": "inline_equation", + "content": "t" + }, + { + "bbox": [ + 130, + 160, + 483, + 376 + ], + "type": "text", + "content": " is matched with the identity of pedestrians from frame " + }, + { + "bbox": [ + 130, + 160, + 483, + 376 + ], + "type": "inline_equation", + "content": "t - 1" + }, + { + "bbox": [ + 130, + 160, + 483, + 376 + ], + "type": "text", + "content": ")." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 131, + 394, + 332, + 406 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 131, + 394, + 332, + 406 + ], + "spans": [ + { + "bbox": [ + 131, + 394, + 332, + 406 + ], + "type": "text", + "content": "3.2 Adversarial Loss for Patch Attack" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 130, + 415, + 482, + 499 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 415, + 482, + 499 + ], + "spans": [ + { + "bbox": [ + 130, + 415, + 482, + 499 + ], + "type": "text", + "content": "For our loss formulation, we adopt the widely used average precision loss and total variation loss [12, 27, 29] to reduce the confidence score of detection while encouraging visual smoothness of the printable patches. However, only using these two losses cannot disrupt the identity association in tracking, as illustrated Figure 1(b). This motivates us to propose a novel bounding box restriction loss, which generates a wrong detection box with a high IoU value, leading to incorrect identity assignment and successfully disrupted data association." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 130, + 507, + 482, + 582 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 507, + 482, + 582 + ], + "spans": [ + { + "bbox": [ + 130, + 507, + 482, + 582 + ], + "type": "text", + "content": "Bounding Box Restriction Loss " + }, + { + "bbox": [ + 130, + 507, + 482, + 582 + ], + "type": "inline_equation", + "content": "(\\mathcal{L}_{bbr})" + }, + { + "bbox": [ + 130, + 507, + 482, + 582 + ], + "type": "text", + "content": ". To effectively disrupt identity association across frames, it is essential to ensure that the sizes of the target detection box and the patch detection box are similar. To this end, a bounding box restriction loss " + }, + { + "bbox": [ + 130, + 507, + 482, + 582 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_{bbr}" + }, + { + "bbox": [ + 130, + 507, + 482, + 582 + ], + "type": "text", + "content": " is formulated to restrict the height and width of the detection box, while minimizing the discrepancy between the detection boxes of the patch and the object. " + }, + { + "bbox": [ + 130, + 507, + 482, + 582 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_{bbr}" + }, + { + "bbox": [ + 130, + 507, + 482, + 582 + ], + "type": "text", + "content": " is defined as:" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 165, + 588, + 481, + 622 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 165, + 588, + 481, + 622 + ], + "spans": [ + { + "bbox": [ + 165, + 588, + 481, + 622 + ], + "type": "interline_equation", + "content": "\\mathcal {L} _ {b b r} = \\frac {1}{N} \\sum_ {i = 1} ^ {N} \\left(w _ {i} / I _ {w i d t h} + h _ {i} / I _ {h e i g h t}\\right) + \\left(1 - \\frac {1}{N} \\sum_ {i = 1} ^ {N} I o U \\left(P _ {i}, T _ {i}\\right)\\right), \\tag {1}", + "image_path": "50ffa9c5325324cf8c29d276dc7452451b95bdc9bd31b6d34f206b3f6d729832.jpg" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 130, + 629, + 482, + 666 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 629, + 482, + 666 + ], + "spans": [ + { + "bbox": [ + 130, + 629, + 482, + 666 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 130, + 629, + 482, + 666 + ], + "type": "inline_equation", + "content": "N" + }, + { + "bbox": [ + 130, + 629, + 482, + 666 + ], + "type": "text", + "content": " is the total number of attacked objects; " + }, + { + "bbox": [ + 130, + 629, + 482, + 666 + ], + "type": "inline_equation", + "content": "w_{i}" + }, + { + "bbox": [ + 130, + 629, + 482, + 666 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 130, + 629, + 482, + 666 + ], + "type": "inline_equation", + "content": "h_{i}" + }, + { + "bbox": [ + 130, + 629, + 482, + 666 + ], + "type": "text", + "content": " are the width and height of predicted bounding boxes of the " + }, + { + "bbox": [ + 130, + 629, + 482, + 666 + ], + "type": "inline_equation", + "content": "i" + }, + { + "bbox": [ + 130, + 629, + 482, + 666 + ], + "type": "text", + "content": "-th target; and " + }, + { + "bbox": [ + 130, + 629, + 482, + 666 + ], + "type": "inline_equation", + "content": "I_{width}" + }, + { + "bbox": [ + 130, + 629, + 482, + 666 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 130, + 629, + 482, + 666 + ], + "type": "inline_equation", + "content": "I_{height}" + }, + { + "bbox": [ + 130, + 629, + 482, + 666 + ], + "type": "text", + "content": " are the width and height of video frames that are used for normalization; detection" + } + ] + } + ], + "index": 8 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "spans": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "type": "text", + "content": "6" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 166, + 91, + 222, + 102 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 166, + 91, + 222, + 102 + ], + "spans": [ + { + "bbox": [ + 166, + 91, + 222, + 102 + ], + "type": "text", + "content": "J. Long et al." + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 5 + }, + { + "para_blocks": [ + { + "bbox": [ + 131, + 116, + 482, + 163 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 131, + 116, + 482, + 163 + ], + "spans": [ + { + "bbox": [ + 131, + 116, + 482, + 163 + ], + "type": "text", + "content": "boxes of the patch and the target are denoted as " + }, + { + "bbox": [ + 131, + 116, + 482, + 163 + ], + "type": "inline_equation", + "content": "P_{i}" + }, + { + "bbox": [ + 131, + 116, + 482, + 163 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 131, + 116, + 482, + 163 + ], + "type": "inline_equation", + "content": "T_{i}" + }, + { + "bbox": [ + 131, + 116, + 482, + 163 + ], + "type": "text", + "content": ", respectively. It is worth noting that the first term seems to make " + }, + { + "bbox": [ + 131, + 116, + 482, + 163 + ], + "type": "inline_equation", + "content": "T_{i}" + }, + { + "bbox": [ + 131, + 116, + 482, + 163 + ], + "type": "text", + "content": " have zero width and height, but we empirically find that it do not completely vanish and even create more new detection boxes due to the balance of IoU losses." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 131, + 173, + 481, + 209 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 131, + 173, + 481, + 209 + ], + "spans": [ + { + "bbox": [ + 131, + 173, + 481, + 209 + ], + "type": "text", + "content": "Total Variation Loss " + }, + { + "bbox": [ + 131, + 173, + 481, + 209 + ], + "type": "inline_equation", + "content": "(\\mathcal{L}_{tv})" + }, + { + "bbox": [ + 131, + 173, + 481, + 209 + ], + "type": "text", + "content": ". The aim of " + }, + { + "bbox": [ + 131, + 173, + 481, + 209 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_{tv}" + }, + { + "bbox": [ + 131, + 173, + 481, + 209 + ], + "type": "text", + "content": " is to ensure the generation of printable patches with smooth color transitions, which helps to reduce the presence of visual noise. " + }, + { + "bbox": [ + 131, + 173, + 481, + 209 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_{tv}" + }, + { + "bbox": [ + 131, + 173, + 481, + 209 + ], + "type": "text", + "content": " is formulated as:" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 206, + 219, + 481, + 249 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 206, + 219, + 481, + 249 + ], + "spans": [ + { + "bbox": [ + 206, + 219, + 481, + 249 + ], + "type": "interline_equation", + "content": "\\mathcal {L} _ {t v} = \\sum_ {i, j} \\sqrt {\\left(\\left(I _ {i , j} - I _ {i + 1 , j}\\right) ^ {2} + \\left(I _ {i , j} - I _ {i , j + 1}\\right) ^ {2}\\right)}, \\tag {2}", + "image_path": "c2755fb942b7f56eec60fe62de2fd4dcd98bb93b6a1ca36ec4a464bf81533c74.jpg" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 131, + 258, + 412, + 272 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 131, + 258, + 412, + 272 + ], + "spans": [ + { + "bbox": [ + 131, + 258, + 412, + 272 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 131, + 258, + 412, + 272 + ], + "type": "inline_equation", + "content": "I_{i,j}" + }, + { + "bbox": [ + 131, + 258, + 412, + 272 + ], + "type": "text", + "content": " denotes the pixel value at position " + }, + { + "bbox": [ + 131, + 258, + 412, + 272 + ], + "type": "inline_equation", + "content": "(i,j)" + }, + { + "bbox": [ + 131, + 258, + 412, + 272 + ], + "type": "text", + "content": " in the patch." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 131, + 279, + 482, + 352 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 131, + 279, + 482, + 352 + ], + "spans": [ + { + "bbox": [ + 131, + 279, + 482, + 352 + ], + "type": "text", + "content": "Average Precision Loss " + }, + { + "bbox": [ + 131, + 279, + 482, + 352 + ], + "type": "inline_equation", + "content": "(\\mathcal{L}_{ap})" + }, + { + "bbox": [ + 131, + 279, + 482, + 352 + ], + "type": "text", + "content": ". It is essential to lower the confidence of the original detection box in order to have it swapped to the one induced by the patch. " + }, + { + "bbox": [ + 131, + 279, + 482, + 352 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_{ap}" + }, + { + "bbox": [ + 131, + 279, + 482, + 352 + ], + "type": "text", + "content": " does so exactly by reducing the average detection score of targets that have been attached to an adversarial patch. Given a set of " + }, + { + "bbox": [ + 131, + 279, + 482, + 352 + ], + "type": "inline_equation", + "content": "N" + }, + { + "bbox": [ + 131, + 279, + 482, + 352 + ], + "type": "text", + "content": " specific detection boxes, where the detection score for the " + }, + { + "bbox": [ + 131, + 279, + 482, + 352 + ], + "type": "inline_equation", + "content": "i" + }, + { + "bbox": [ + 131, + 279, + 482, + 352 + ], + "type": "text", + "content": "-th box is " + }, + { + "bbox": [ + 131, + 279, + 482, + 352 + ], + "type": "inline_equation", + "content": "s_i" + }, + { + "bbox": [ + 131, + 279, + 482, + 352 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 131, + 279, + 482, + 352 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_{ap}" + }, + { + "bbox": [ + 131, + 279, + 482, + 352 + ], + "type": "text", + "content": " is calculated as the sum of detection scores divided by the total number of detection boxes as follows:" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 272, + 362, + 481, + 393 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 272, + 362, + 481, + 393 + ], + "spans": [ + { + "bbox": [ + 272, + 362, + 481, + 393 + ], + "type": "interline_equation", + "content": "\\mathcal {L} _ {a p} = \\frac {1}{N} \\sum_ {i = 1} ^ {N} s _ {i}. \\tag {3}", + "image_path": "a6da87bf5a8072778e53beae1e044f6a202caa1ec7c4cf06dcf82bee2c5ff26d.jpg" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 131, + 399, + 481, + 422 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 131, + 399, + 481, + 422 + ], + "spans": [ + { + "bbox": [ + 131, + 399, + 481, + 422 + ], + "type": "text", + "content": "Finally, the overall adversarial loss " + }, + { + "bbox": [ + 131, + 399, + 481, + 422 + ], + "type": "inline_equation", + "content": "L_{adv}" + }, + { + "bbox": [ + 131, + 399, + 481, + 422 + ], + "type": "text", + "content": " is a sum of the above three loss terms weighted by hyperparameters " + }, + { + "bbox": [ + 131, + 399, + 481, + 422 + ], + "type": "inline_equation", + "content": "\\beta" + }, + { + "bbox": [ + 131, + 399, + 481, + 422 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 131, + 399, + 481, + 422 + ], + "type": "inline_equation", + "content": "\\gamma" + }, + { + "bbox": [ + 131, + 399, + 481, + 422 + ], + "type": "text", + "content": ", and " + }, + { + "bbox": [ + 131, + 399, + 481, + 422 + ], + "type": "inline_equation", + "content": "\\delta" + }, + { + "bbox": [ + 131, + 399, + 481, + 422 + ], + "type": "text", + "content": ":" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 244, + 434, + 481, + 447 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 244, + 434, + 481, + 447 + ], + "spans": [ + { + "bbox": [ + 244, + 434, + 481, + 447 + ], + "type": "interline_equation", + "content": "\\mathcal {L} _ {a d v} = \\beta \\mathcal {L} _ {b b r} + \\gamma \\mathcal {L} _ {t v} + \\delta \\mathcal {L} _ {a p}, \\tag {4}", + "image_path": "e5c910a0be904c54e2c2042b7257d09939039c344eda71a3069c0b747853422e.jpg" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 131, + 464, + 343, + 475 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 131, + 464, + 343, + 475 + ], + "spans": [ + { + "bbox": [ + 131, + 464, + 343, + 475 + ], + "type": "text", + "content": "3.3 Evaluation Metrics for MOT Attack" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 130, + 485, + 481, + 557 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 485, + 481, + 557 + ], + "spans": [ + { + "bbox": [ + 130, + 485, + 481, + 557 + ], + "type": "text", + "content": "Current attack methods against MOT [17, 20] have limitations in their evaluation since they utilize identity-based evaluation metrics. These metrics have two main drawbacks. Firstly, they are primarily designed for gauging tracking performance, thereby largely overlooking the quality of detection - the other core ingredient in successful MOT. Secondly, they tend to focus only on a few specific targets, disregarding a comprehensive evaluation of the entire MOT system." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 131, + 558, + 482, + 666 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 131, + 558, + 482, + 666 + ], + "spans": [ + { + "bbox": [ + 131, + 558, + 482, + 666 + ], + "type": "text", + "content": "As depicted in Figure 3(a), identity-based evaluation metrics [17,20] evaluate the identity of the victim in every frame, which considers tracking. However, existing MOT algorithms typically include both tracking and detection modules. Therefore, while evaluating attacks on MOT, it is essential to calculate the performance of both tracking and detection modules simultaneously. Furthermore, identity-based evaluation metrics usually adopt a strategy that aims to attack one or a few specific targets. This approach overlooks the cumulative impact of attacking these specific targets on the entire system. For instance, as shown in Figure 3(b), an attack on the victim (in the red detection box) might also affect" + } + ] + } + ], + "index": 12 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 200, + 91, + 448, + 103 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 200, + 91, + 448, + 103 + ], + "spans": [ + { + "bbox": [ + 200, + 91, + 448, + 103 + ], + "type": "text", + "content": "PapMOT: Exploring Adversarial Patch Attack against MOT" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 474, + 91, + 480, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 474, + 91, + 480, + 100 + ], + "spans": [ + { + "bbox": [ + 474, + 91, + 480, + 100 + ], + "type": "text", + "content": "7" + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 6 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 168, + 113, + 445, + 223 + ], + "blocks": [ + { + "bbox": [ + 168, + 113, + 445, + 223 + ], + "lines": [ + { + "bbox": [ + 168, + 113, + 445, + 223 + ], + "spans": [ + { + "bbox": [ + 168, + 113, + 445, + 223 + ], + "type": "image", + "image_path": "84ed199bf3cf9c7b6a41af4164bee7d98d3abe0959b8081e55ce9fc9f5f68071.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 131, + 232, + 482, + 243 + ], + "lines": [ + { + "bbox": [ + 131, + 232, + 482, + 243 + ], + "spans": [ + { + "bbox": [ + 131, + 232, + 482, + 243 + ], + "type": "text", + "content": "Fig. 3: A comparison of existing and our proposed evaluation metrics for MOT attacks." + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 132, + 243, + 481, + 275 + ], + "lines": [ + { + "bbox": [ + 132, + 243, + 481, + 275 + ], + "spans": [ + { + "bbox": [ + 132, + 243, + 481, + 275 + ], + "type": "text", + "content": "(a) Previous evaluation metrics are identity-based and only consider ID association. \n(b) Our proposed metrics emphasize the attack evaluation of both detection and ID association." + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "image_footnote" + } + ], + "index": 2 + }, + { + "bbox": [ + 130, + 301, + 480, + 325 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 301, + 480, + 325 + ], + "spans": [ + { + "bbox": [ + 130, + 301, + 480, + 325 + ], + "type": "text", + "content": "the detection of other pedestrians (in yellow boxes), leading to switches of their identities or a decrease in detection accuracy." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 130, + 326, + 481, + 374 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 326, + 481, + 374 + ], + "spans": [ + { + "bbox": [ + 130, + 326, + 481, + 374 + ], + "type": "text", + "content": "To overcome these limitations, we construct a maximum-attack-threshold evaluation metric that integrates detection and tracking. This type of evaluation metric is based on the classic evaluation metrics for MOT [3]. It provides a comprehensive evaluation of the robustness of MOT systems." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 130, + 383, + 482, + 504 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 383, + 482, + 504 + ], + "spans": [ + { + "bbox": [ + 130, + 383, + 482, + 504 + ], + "type": "text", + "content": "Tracking Attack Success Rate (TASR). To assess the impact of specific targets during attacks on the overall system, TASR determines the change in MOTA and IDF1 before and after attacks and then divides it by the percentage of bounding boxes that are attacked " + }, + { + "bbox": [ + 130, + 383, + 482, + 504 + ], + "type": "inline_equation", + "content": "(R_{bbox})" + }, + { + "bbox": [ + 130, + 383, + 482, + 504 + ], + "type": "text", + "content": ". A high TASR indicates that the MOT system is highly susceptible to particular attacks, enabling an attacker to degrade its performance by compromising fewer targets. A low TASR implies that the attacker must compromise a greater number of targets to inflict considerable damage. The formula for TASR is: " + }, + { + "bbox": [ + 130, + 383, + 482, + 504 + ], + "type": "inline_equation", + "content": "TASR = (MOTA + IDF1)_{\\text{decline}} / 2R_{bbox}" + }, + { + "bbox": [ + 130, + 383, + 482, + 504 + ], + "type": "text", + "content": ", where " + }, + { + "bbox": [ + 130, + 383, + 482, + 504 + ], + "type": "inline_equation", + "content": "R_{bbox}" + }, + { + "bbox": [ + 130, + 383, + 482, + 504 + ], + "type": "text", + "content": " is the ratio of the number of bounding boxes under attack to the total bounding boxes." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 130, + 512, + 482, + 596 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 512, + 482, + 596 + ], + "spans": [ + { + "bbox": [ + 130, + 512, + 482, + 596 + ], + "type": "text", + "content": "Identity Obfuscation Rate (IOR). To measure the effectiveness of adversarial attacks to cause identity obfuscation during data association in MOT, we define the Identity Obfuscation Rate as: " + }, + { + "bbox": [ + 130, + 512, + 482, + 596 + ], + "type": "inline_equation", + "content": "IOR = D_s / P_t" + }, + { + "bbox": [ + 130, + 512, + 482, + 596 + ], + "type": "text", + "content": ", where " + }, + { + "bbox": [ + 130, + 512, + 482, + 596 + ], + "type": "inline_equation", + "content": "P_t = N / T" + }, + { + "bbox": [ + 130, + 512, + 482, + 596 + ], + "type": "text", + "content": ", where " + }, + { + "bbox": [ + 130, + 512, + 482, + 596 + ], + "type": "inline_equation", + "content": "T" + }, + { + "bbox": [ + 130, + 512, + 482, + 596 + ], + "type": "text", + "content": " represents the number of frames that are attacked consecutively, " + }, + { + "bbox": [ + 130, + 512, + 482, + 596 + ], + "type": "inline_equation", + "content": "N" + }, + { + "bbox": [ + 130, + 512, + 482, + 596 + ], + "type": "text", + "content": " is the total number of frames of the video, " + }, + { + "bbox": [ + 130, + 512, + 482, + 596 + ], + "type": "inline_equation", + "content": "P_t" + }, + { + "bbox": [ + 130, + 512, + 482, + 596 + ], + "type": "text", + "content": " denotes the maximum number of identity switches in the video, and " + }, + { + "bbox": [ + 130, + 512, + 482, + 596 + ], + "type": "inline_equation", + "content": "D_s" + }, + { + "bbox": [ + 130, + 512, + 482, + 596 + ], + "type": "text", + "content": " is the number of identity switches throughout the video sequences." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 130, + 605, + 482, + 667 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 605, + 482, + 667 + ], + "spans": [ + { + "bbox": [ + 130, + 605, + 482, + 667 + ], + "type": "text", + "content": "Specific Target Attack Success Rate (STASR). To evaluate attacks on specific targets in terms of both false positive rate (FP) and the number of identity switches, STASR focuses not only on how the attack interferes with the detection of the targets but also on how the attack disrupts the tracking of the targets. STASR is defined as: " + }, + { + "bbox": [ + 130, + 605, + 482, + 667 + ], + "type": "inline_equation", + "content": "STASR = (FP + ID_{switch})_{increase} / A_{max}" + }, + { + "bbox": [ + 130, + 605, + 482, + 667 + ], + "type": "text", + "content": ", where" + } + ] + } + ], + "index": 9 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "spans": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "type": "text", + "content": "8" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 166, + 91, + 222, + 101 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 166, + 91, + 222, + 101 + ], + "spans": [ + { + "bbox": [ + 166, + 91, + 222, + 101 + ], + "type": "text", + "content": "J. Long et al." + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 7 + }, + { + "para_blocks": [ + { + "bbox": [ + 132, + 116, + 480, + 140 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 116, + 480, + 140 + ], + "spans": [ + { + "bbox": [ + 132, + 116, + 480, + 140 + ], + "type": "inline_equation", + "content": "A_{max} = P_t + P_n" + }, + { + "bbox": [ + 132, + 116, + 480, + 140 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 132, + 116, + 480, + 140 + ], + "type": "inline_equation", + "content": "P_t" + }, + { + "bbox": [ + 132, + 116, + 480, + 140 + ], + "type": "text", + "content": " represents the maximum number of identity switches in the video, and " + }, + { + "bbox": [ + 132, + 116, + 480, + 140 + ], + "type": "inline_equation", + "content": "P_n" + }, + { + "bbox": [ + 132, + 116, + 480, + 140 + ], + "type": "text", + "content": " denotes the number of detection boxes that are attacked." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 132, + 159, + 229, + 173 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 159, + 229, + 173 + ], + "spans": [ + { + "bbox": [ + 132, + 159, + 229, + 173 + ], + "type": "text", + "content": "4 Experiments" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 132, + 185, + 480, + 221 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 185, + 480, + 221 + ], + "spans": [ + { + "bbox": [ + 132, + 185, + 480, + 221 + ], + "type": "text", + "content": "In this section, we perform extensive empirical analysis of the proposed PapMOT attack on MOT in both digital and physical environments. We also conduct ablation studies on the adversarial loss and variants of adversarial patches." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 132, + 239, + 261, + 251 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 239, + 261, + 251 + ], + "spans": [ + { + "bbox": [ + 132, + 239, + 261, + 251 + ], + "type": "text", + "content": "4.1 Experimental Setup" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 132, + 269, + 480, + 340 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 269, + 480, + 340 + ], + "spans": [ + { + "bbox": [ + 132, + 269, + 480, + 340 + ], + "type": "text", + "content": "Datasets and Metrics. We evaluate our attack on four popular MOT benchmarks: MOT15 [19], MOT17 [23], MOT20 [7], and BDD100K [36]. The target objects of the first three datasets are pedestrian-only, while BDD100K further includes vehicle objects. We adopt our proposed evaluation metrics, namely TASR, SATSR, and IOR, to assess the performance of our PapMOT against state-of-the-art methods on the MOT task. More comparisons are found in Appendix." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 132, + 349, + 480, + 444 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 349, + 480, + 444 + ], + "spans": [ + { + "bbox": [ + 132, + 349, + 480, + 444 + ], + "type": "text", + "content": "Baselines. Existing MOT attack methods can be divided into three categories: 1) methods that attack detectors in the MOT system by injecting pixel-level noise, e.g., RAnoise [21] and TOG [6]. 2) methods that attack detectors in the MOT system by placing optimized patches, e.g., AdvTexture [13] and ObjPatch [28]. 3) methods that attack both detectors and trackers in the MOT system by injecting pixel-level noise, e.g., Hijack [17] and Trasw [20]. Our PapMOT introduces a physically realizable adversarial patch into the MOT system to attack both detectors and trackers." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 132, + 453, + 480, + 560 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 453, + 480, + 560 + ], + "spans": [ + { + "bbox": [ + 132, + 453, + 480, + 560 + ], + "type": "text", + "content": "Victim Models. We choose two representative MOT methods as victim models in our experiments: ByteTrack [38] that represents the tracking-by-detection paradigm and FairMOT [39] that represents the joint detection and tracking paradigm. Besides, ByteTrack is equipped with YOLOX [11], a more advanced version of theYOLO detector family trained on CrowdHuman [26], MOT17 [23], CityPerson [37], and ETHZ [10] datasets. The pre-trained model of FairMOT is provided by PaddleDetection [2], and is trained on Caltech Pedestrian [8], CityPersons [37], CUHK-SYSU [32], PRW [41], ETHZ [10], and MOT17 [23] datasets." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 132, + 570, + 480, + 665 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 570, + 480, + 665 + ], + "spans": [ + { + "bbox": [ + 132, + 570, + 480, + 665 + ], + "type": "text", + "content": "Implementation Details. In our digital experiments, to strike a balance between the stealthiness and effectiveness of attacks, we set the noise standard deviation of RAnoise to 10, and limit the maximum disturbance range in TOG to 8/255. The patch sizes for both AdvTexture and ObjPatch are set to one-third of the targeted detection box's size. We define an object as \"attacked\" if it overlaps with another object with an IOU greater than 0.2. In our physical experiments, we print the patches on foam boards with a fixed width and height of " + }, + { + "bbox": [ + 132, + 570, + 480, + 665 + ], + "type": "inline_equation", + "content": "50\\mathrm{cm}" + }, + { + "bbox": [ + 132, + 570, + 480, + 665 + ], + "type": "text", + "content": ". The size ratio between the patch and an adult who is " + }, + { + "bbox": [ + 132, + 570, + 480, + 665 + ], + "type": "inline_equation", + "content": "1.6\\mathrm{m}" + }, + { + "bbox": [ + 132, + 570, + 480, + 665 + ], + "type": "text", + "content": " tall is kept at" + } + ] + } + ], + "index": 9 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 200, + 91, + 448, + 102 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 200, + 91, + 448, + 102 + ], + "spans": [ + { + "bbox": [ + 200, + 91, + 448, + 102 + ], + "type": "text", + "content": "PapMOT: Exploring Adversarial Patch Attack against MOT" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 474, + 92, + 480, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 474, + 92, + 480, + 100 + ], + "spans": [ + { + "bbox": [ + 474, + 92, + 480, + 100 + ], + "type": "text", + "content": "9" + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 8 + }, + { + "para_blocks": [ + { + "type": "table", + "bbox": [ + 135, + 146, + 473, + 272 + ], + "blocks": [ + { + "bbox": [ + 132, + 114, + 482, + 137 + ], + "lines": [ + { + "bbox": [ + 132, + 114, + 482, + 137 + ], + "spans": [ + { + "bbox": [ + 132, + 114, + 482, + 137 + ], + "type": "text", + "content": "Table 2: Comparison with other representative adversarial attack methods on the MOT15 and MOT17 datasets." + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 135, + 146, + 473, + 272 + ], + "lines": [ + { + "bbox": [ + 135, + 146, + 473, + 272 + ], + "spans": [ + { + "bbox": [ + 135, + 146, + 473, + 272 + ], + "type": "table", + "html": "
DatasetMOT15MOT17
VictimFairMOTByteTrackFairMOTByteTrack
MetricsTASRSATSRIORTASRSATSRIORTASRSATSRIORTASRSATSRIOR
RAnoise [21]0.70.6-0.60.3-0.30.5--0.2-
TOG [6]13.77.04.111.46.33.89.14.82.54.73.91.2
AdvTexture [13]36.435.64.638.143.75.845.432.15.142.247.76.0
ObjPatch [28]45.240.66.343.759.38.261.349.56.261.476.77.4
Hijack [17]14.218.539.311.415.929.010.019.226.58.415.425.4
Trasw [20]40.645.866.231.439.255.422.128.944.821.026.842.3
Ours76.485.062.385.379.265.976.893.046.294.689.679.5
", + "image_path": "31c514562d68e6a946822c75631390b5b2161e9504e0a91fcb61593613a98f57.jpg" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "table_body" + } + ], + "index": 3 + }, + { + "bbox": [ + 130, + 293, + 482, + 343 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 293, + 482, + 343 + ], + "spans": [ + { + "bbox": [ + 130, + 293, + 482, + 343 + ], + "type": "text", + "content": "around " + }, + { + "bbox": [ + 130, + 293, + 482, + 343 + ], + "type": "inline_equation", + "content": "30\\%" + }, + { + "bbox": [ + 130, + 293, + 482, + 343 + ], + "type": "text", + "content": ", ensuring that essential areas such as the face are not obstructed to maintain physical feasibility. For hyperparameters in the overall adversarial loss (i.e., Eq. 4), we set " + }, + { + "bbox": [ + 130, + 293, + 482, + 343 + ], + "type": "inline_equation", + "content": "\\beta = 1" + }, + { + "bbox": [ + 130, + 293, + 482, + 343 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 130, + 293, + 482, + 343 + ], + "type": "inline_equation", + "content": "\\gamma = 2.5" + }, + { + "bbox": [ + 130, + 293, + 482, + 343 + ], + "type": "text", + "content": ", and " + }, + { + "bbox": [ + 130, + 293, + 482, + 343 + ], + "type": "inline_equation", + "content": "\\delta = 2" + }, + { + "bbox": [ + 130, + 293, + 482, + 343 + ], + "type": "text", + "content": ". Datasets used in our patch generation consist of ETHZ, MOT17, and MOT17 clip datasets." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 132, + 358, + 307, + 372 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 358, + 307, + 372 + ], + "spans": [ + { + "bbox": [ + 132, + 358, + 307, + 372 + ], + "type": "text", + "content": "4.2 Digital Domain Performance" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 130, + 378, + 482, + 545 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 378, + 482, + 545 + ], + "spans": [ + { + "bbox": [ + 130, + 378, + 482, + 545 + ], + "type": "text", + "content": "Table 2 and Table 3 compare the performance of various attacks on the MOT15, MOT17, MOT20, and BDD100K datasets. As can be seen, PapMOT achieves the highest attack success rate across all four datasets with a " + }, + { + "bbox": [ + 130, + 378, + 482, + 545 + ], + "type": "inline_equation", + "content": "50\\%" + }, + { + "bbox": [ + 130, + 378, + 482, + 545 + ], + "type": "text", + "content": " bounding box attack, manifesting its superiority over existing methods (i.e., RAnoise [21], TOG [6], ADVTexure [13], ObjPatch [28], Hijack [17], Trasw [20]) against MOT in both pedestrian (i.e., MOT15, MOT17, and MOT20) and autonomous driving scenarios (i.e., BDD100K). We notice that the IOR of PapMOT is marginally below that of Trasw on the MOT15 and MOT20 datasets, which is possibly due to Trasw's sole emphasis on attacking data association. Nonetheless, PapMOT surpasses Trasw by a huge margin in terms of TASR and SATSR. Overall, Trasw as a digital method is hard to be implemented in the physical world, and its attack is strictly triggered only when an object overlaps with the attack object. Thus, PapMOT outperforms Trasw in terms of functionality and overall attack performance." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 130, + 545, + 483, + 666 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 545, + 483, + 666 + ], + "spans": [ + { + "bbox": [ + 130, + 545, + 483, + 666 + ], + "type": "text", + "content": "The superior performance of PapMOT is most evident on the MOT17 dataset, where it consistently surpasses all competing methods across all three metrics when applied against both FairMOT and ByteTrack. The success of PapMOT is largely due to incorporating the MOT17 clip dataset and MOT17 into its patch training dataset (more details about the clip dataset are found in Appendix). It is noteworthy that TASR and SATSR both incorporate the evaluation of detection box performance, whereas IOR predominantly assesses identity association across frames. As such, it is observed that TASR and SATSR are typically higher for patch-based attacks and digital attacks compared to IOR (e.g., when attacking ByteTrack in MOT17, the IOR of AdvTexture is 6.0, while TASR and" + } + ] + } + ], + "index": 7 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "spans": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "type": "text", + "content": "10" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 166, + 91, + 221, + 102 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 166, + 91, + 221, + 102 + ], + "spans": [ + { + "bbox": [ + 166, + 91, + 221, + 102 + ], + "type": "text", + "content": "J. Long et al." + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 9 + }, + { + "para_blocks": [ + { + "type": "table", + "bbox": [ + 135, + 146, + 473, + 273 + ], + "blocks": [ + { + "bbox": [ + 132, + 114, + 480, + 137 + ], + "lines": [ + { + "bbox": [ + 132, + 114, + 480, + 137 + ], + "spans": [ + { + "bbox": [ + 132, + 114, + 480, + 137 + ], + "type": "text", + "content": "Table 3: Comparison with other representative adversarial attack methods on the MOT20 and BDD100K datasets." + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 135, + 146, + 473, + 273 + ], + "lines": [ + { + "bbox": [ + 135, + 146, + 473, + 273 + ], + "spans": [ + { + "bbox": [ + 135, + 146, + 473, + 273 + ], + "type": "table", + "html": "
DatasetMOT20BDD100K
VictimFairMOTByteTrackFairMOTByteTrack
MetricsTASRSATSRIORTASRSATSRIORTASRSATSRIORTASRSATSRIOR
RAnoise [21]2.60.90.44.11.50.42.20.7-1.90.7-
TOG [6]8.62.50.67.62.30.618.16.72.413.54.82.2
AdvTexture [13]23.518.52.932.026.73.221.112.64.418.319.04.8
ObjPatch [28]42.321.44.441.327.56.826.319.45.220.523.37.6
Hijack [17]17.918.436.815.934.727.07.19.810.57.48.213.5
Trasw [20]31.233.567.026.628.564.212.217.423.311.013.119.3
Ours61.949.556.658.541.653.946.454.734.163.461.238.8
", + "image_path": "7223f4e49f05891bd7a3a762545c2fad5f6214401cbb6763ffe5f0bed3fc64c8.jpg" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "table_body" + } + ], + "index": 3 + }, + { + "bbox": [ + 130, + 295, + 480, + 344 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 295, + 480, + 344 + ], + "spans": [ + { + "bbox": [ + 130, + 295, + 480, + 344 + ], + "type": "text", + "content": "SATSR in ADvTexture are 42.2 and 47.7, respectively). This observation can be accounted for by the fact that these attacks primarily target detectors, which significantly affect the performance of detection boxes, but minimally impact identity association across frames." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 132, + 361, + 312, + 373 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 361, + 312, + 373 + ], + "spans": [ + { + "bbox": [ + 132, + 361, + 312, + 373 + ], + "type": "text", + "content": "4.3 Physical Domain Performance" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 130, + 381, + 480, + 441 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 381, + 480, + 441 + ], + "spans": [ + { + "bbox": [ + 130, + 381, + 480, + 441 + ], + "type": "text", + "content": "While our patches have shown outstanding performance in the digital domain, translating this success to the physical world is challenging due to the ever-changing of physical environments. In this section, we qualitatively evaluate the effectiveness of our adversarial patches, and further probe their limits under diverse physical environments." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 130, + 449, + 480, + 509 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 449, + 480, + 509 + ], + "spans": [ + { + "bbox": [ + 130, + 449, + 480, + 509 + ], + "type": "text", + "content": "Physical Validation. We train three patches with different datasets (i.e., ETHZ, MOT17, MOT20) for FairMOT and evaluate them against the ByteTrack detector. Figure 4 showcases the demonstration of these adversarial patches for physical-world attacks under diverse scenes (refer to Supplementary Materials for a video demo)." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 130, + 510, + 480, + 605 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 510, + 480, + 605 + ], + "spans": [ + { + "bbox": [ + 130, + 510, + 480, + 605 + ], + "type": "text", + "content": "In Figure 4 (a), an adversarial patch is attached to a pedestrian, targeting a single object. In the absence of our attack, the target's ID detected by the tracker is consistently 1, and the detected bounding box is accurate. When the attack is imposed, the detector produces an additional false detection box (in blue), and the accuracy of the initially accurate detection box deteriorates significantly. Moreover, the tracker is quickly deprived of the ability to identity associations across frames, and an ID switch takes place within 30 frames. These outcomes validate the effectiveness of our attack in single-object scenarios." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 130, + 605, + 481, + 665 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 605, + 481, + 665 + ], + "spans": [ + { + "bbox": [ + 130, + 605, + 481, + 665 + ], + "type": "text", + "content": "Figure 4 (b) illustrates a multi-object scenario, where the adversarial patch is attached to one of the two pedestrians in the scene. When the pedestrian in white with the patch (denoted as \"target A\") passes by in front of the other subject in black (\"target B\"), their detected identities are successfully maintained in the absence of the attack. When attacked, however, the tracker mistakenly" + } + ] + } + ], + "index": 9 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 200, + 91, + 447, + 102 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 200, + 91, + 447, + 102 + ], + "spans": [ + { + "bbox": [ + 200, + 91, + 447, + 102 + ], + "type": "text", + "content": "PapMOT: Exploring Adversarial Patch Attack against MOT" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 470, + 92, + 479, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 470, + 92, + 479, + 100 + ], + "spans": [ + { + "bbox": [ + 470, + 92, + 479, + 100 + ], + "type": "text", + "content": "11" + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 10 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 140, + 114, + 476, + 261 + ], + "blocks": [ + { + "bbox": [ + 140, + 114, + 476, + 261 + ], + "lines": [ + { + "bbox": [ + 140, + 114, + 476, + 261 + ], + "spans": [ + { + "bbox": [ + 140, + 114, + 476, + 261 + ], + "type": "image", + "image_path": "95c589e9cf0162945ddab5276d582e17531acbe28c5ae6f4bb626f69e7fcaec6.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 130, + 270, + 480, + 304 + ], + "lines": [ + { + "bbox": [ + 130, + 270, + 480, + 304 + ], + "spans": [ + { + "bbox": [ + 130, + 270, + 480, + 304 + ], + "type": "text", + "content": "Fig. 4: Black-box transfer attack of different PapMOT patches in the real world. It illustrates the effective attack for the scenarios of a single pedestrian (a), two pedestrians overlap (b), and a pedestrian walking past from behind a fixed patch (c)." + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_caption" + } + ], + "index": 2 + }, + { + "bbox": [ + 130, + 330, + 480, + 378 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 330, + 480, + 378 + ], + "spans": [ + { + "bbox": [ + 130, + 330, + 480, + 378 + ], + "type": "text", + "content": "predicts two identities (i.e., 1 and 3) for target A. Also, an ID switch occurs and the accuracy of both boxes drops as target A passes by. This example shows our proposed attack works in multiple-object scenarios and its effect applies beyond the object with the attached patch." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 130, + 379, + 481, + 439 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 379, + 481, + 439 + ], + "spans": [ + { + "bbox": [ + 130, + 379, + 481, + 439 + ], + "type": "text", + "content": "Figure 4 (c) shows the scenario where a patch is kept stationary and is misidentified as a pedestrian by the tracker. When attacked, a pedestrian passing behind it immediately causes an ID switch to occur (i.e., ID changes from 2 to 1). This example suggests that our adversarial patch is also suitable for static placement and is capable of affecting pedestrians as they pass by." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 130, + 449, + 482, + 604 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 449, + 482, + 604 + ], + "spans": [ + { + "bbox": [ + 130, + 449, + 482, + 604 + ], + "type": "text", + "content": "Attack Boundary. We examine three key physical conditions that may affect the effectiveness of patch attacks: (a) illumination, (b) distance, and (c) angle, shown in Fig. 5. For illumination, we sweep over the entire spectrum from total darkness to total brightness by adjusting the ISO sensitivity of the camera. We divide the illumination levels into 10 segments between 0 and 1. An illumination value of 0.5 denotes the normal exposure under overcast outdoor conditions with the ISO set to 200; 0 and 1 correspond to full darkness and full brightness, respectively. For distance, we vary the range between the patch and the camera from 0.5 to 10 meters to obtain a better view. For angle, we control the angle of the patch relative to the camera by adjusting the deflection angle of the person and using 5-degree intervals to present the final attack results. The quantitative evaluation results of the attack boundary effectiveness for three different patches are presented in Table 4." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 130, + 605, + 482, + 665 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 605, + 482, + 665 + ], + "spans": [ + { + "bbox": [ + 130, + 605, + 482, + 665 + ], + "type": "text", + "content": "Based on the experimental results, several important conclusions can be drawn. Firstly, different patches can achieve varying attack effects depending on the physical conditions. Secondly, the distance between the patches and the camera must be maintained within a specific range (0.8m-4.1m) to achieve an effective attack. Thirdly, patch attacks work better in relatively dark environ" + } + ] + } + ], + "index": 7 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "spans": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "type": "text", + "content": "12" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 166, + 91, + 221, + 101 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 166, + 91, + 221, + 101 + ], + "spans": [ + { + "bbox": [ + 166, + 91, + 221, + 101 + ], + "type": "text", + "content": "J. Long et al." + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 11 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 138, + 115, + 476, + 203 + ], + "blocks": [ + { + "bbox": [ + 138, + 115, + 476, + 203 + ], + "lines": [ + { + "bbox": [ + 138, + 115, + 476, + 203 + ], + "spans": [ + { + "bbox": [ + 138, + 115, + 476, + 203 + ], + "type": "image", + "image_path": "2f5cc22b3002e2ef1c281fd3995402af56d8a40580f62298d8b61b6d787c4efb.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 133, + 212, + 478, + 224 + ], + "lines": [ + { + "bbox": [ + 133, + 212, + 478, + 224 + ], + "spans": [ + { + "bbox": [ + 133, + 212, + 478, + 224 + ], + "type": "text", + "content": "Fig. 5: Studies on patch effectiveness at varying illuminations, distances, and angles." + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_caption" + } + ], + "index": 2 + }, + { + "type": "table", + "bbox": [ + 192, + 256, + 418, + 313 + ], + "blocks": [ + { + "bbox": [ + 159, + 235, + 453, + 247 + ], + "lines": [ + { + "bbox": [ + 159, + 235, + 453, + 247 + ], + "spans": [ + { + "bbox": [ + 159, + 235, + 453, + 247 + ], + "type": "text", + "content": "Table 4: Studies on the effectiveness range of three adversarial patches." + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 192, + 256, + 418, + 313 + ], + "lines": [ + { + "bbox": [ + 192, + 256, + 418, + 313 + ], + "spans": [ + { + "bbox": [ + 192, + 256, + 418, + 313 + ], + "type": "table", + "html": "
DistanceIlluminationPatch SizeAngle
Patch 10.8m-3.1m0.4-0.740cm-100cm0°-25°
Patch 21.2m-4.1m0.5-0.840cm-100cm0°-15°
Patch 31.1m-3.6m0.4-0.840cm-100cm0°-30°
", + "image_path": "9ae93028cd95e837ee2afc069f40eaacde18aa613f3d486f2db5200e6c517d35.jpg" + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "table_body" + } + ], + "index": 5 + }, + { + "bbox": [ + 130, + 336, + 480, + 419 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 336, + 480, + 419 + ], + "spans": [ + { + "bbox": [ + 130, + 336, + 480, + 419 + ], + "type": "text", + "content": "ments, resulting in more false detection boxes. Moreover, a larger patch size has a stronger attack ability. However, to ensure the stealthiness and effectiveness of the patch, the patch size in PapMOT is set to approximately " + }, + { + "bbox": [ + 130, + 336, + 480, + 419 + ], + "type": "inline_equation", + "content": "30\\%" + }, + { + "bbox": [ + 130, + 336, + 480, + 419 + ], + "type": "text", + "content": " of the target size. Lastly, to ensure effective attack outcomes, the angle of the patch relative to the camera generally needs to be limited to a small range (i.e., within 15 degrees). It may be considered to add rotation transformation in the patch training to enhance the robustness of the patches against rotation." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 132, + 437, + 245, + 448 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 437, + 245, + 448 + ], + "spans": [ + { + "bbox": [ + 132, + 437, + 245, + 448 + ], + "type": "text", + "content": "4.4 Ablation Studies" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 130, + 457, + 480, + 493 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 457, + 480, + 493 + ], + "spans": [ + { + "bbox": [ + 130, + 457, + 480, + 493 + ], + "type": "text", + "content": "We analyze the effectiveness of each adversarial loss term in our proposed training objective. We also conduct real-world experiments to evaluate the robustness against unoptimized patches, including real images and uniform color patches." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 130, + 501, + 482, + 609 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 501, + 482, + 609 + ], + "spans": [ + { + "bbox": [ + 130, + 501, + 482, + 609 + ], + "type": "text", + "content": "Loss Functions. Table 5 shows how different adversarial losses affect the performance of our proposed on against MOT17. The results indicate that PapMOT without " + }, + { + "bbox": [ + 130, + 501, + 482, + 609 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_{bbr}" + }, + { + "bbox": [ + 130, + 501, + 482, + 609 + ], + "type": "text", + "content": " performs significantly worse than PapMOT without " + }, + { + "bbox": [ + 130, + 501, + 482, + 609 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_{ap}" + }, + { + "bbox": [ + 130, + 501, + 482, + 609 + ], + "type": "text", + "content": " in terms of IOR, scoring 7.2 and 52.6, respectively, when attacking ByteTrack. This suggests that the " + }, + { + "bbox": [ + 130, + 501, + 482, + 609 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_{bbr}" + }, + { + "bbox": [ + 130, + 501, + 482, + 609 + ], + "type": "text", + "content": " loss is critical in disrupting the consistency of data association. Conversely, when attacking FairMOT, PapMOT without " + }, + { + "bbox": [ + 130, + 501, + 482, + 609 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_{ap}" + }, + { + "bbox": [ + 130, + 501, + 482, + 609 + ], + "type": "text", + "content": " scores substantially lower in TASR and SATSR than PapMOT without " + }, + { + "bbox": [ + 130, + 501, + 482, + 609 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_{bbr}" + }, + { + "bbox": [ + 130, + 501, + 482, + 609 + ], + "type": "text", + "content": ", with scores of 11.0 and 13.1, respectively, compared to 50.5 and 43.3. This underscores the significant impact of the " + }, + { + "bbox": [ + 130, + 501, + 482, + 609 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_{ap}" + }, + { + "bbox": [ + 130, + 501, + 482, + 609 + ], + "type": "text", + "content": " loss on detector performance." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 130, + 617, + 482, + 666 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 617, + 482, + 666 + ], + "spans": [ + { + "bbox": [ + 130, + 617, + 482, + 666 + ], + "type": "text", + "content": "Various Patches. We evaluate the effectiveness of attacking the MOT system using unoptimized patches (e.g., full-white patches, full-black patches, and real image patches). As shown in Figure 6, although the target is partially occluded by (a) the full-white patch or (b) the full-black patch, it is tracked steadily" + } + ] + } + ], + "index": 10 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 200, + 91, + 448, + 102 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 200, + 91, + 448, + 102 + ], + "spans": [ + { + "bbox": [ + 200, + 91, + 448, + 102 + ], + "type": "text", + "content": "PapMOT: Exploring Adversarial Patch Attack against MOT" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 470, + 91, + 480, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 470, + 91, + 480, + 100 + ], + "spans": [ + { + "bbox": [ + 470, + 91, + 480, + 100 + ], + "type": "text", + "content": "13" + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 12 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 145, + 113, + 252, + 183 + ], + "blocks": [ + { + "bbox": [ + 145, + 113, + 252, + 183 + ], + "lines": [ + { + "bbox": [ + 145, + 113, + 252, + 183 + ], + "spans": [ + { + "bbox": [ + 145, + 113, + 252, + 183 + ], + "type": "image", + "image_path": "24c0579cb70f138ff9246131c7fcdef2593151fa31c4c96e193d2f5c619bec25.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 161, + 185, + 244, + 196 + ], + "lines": [ + { + "bbox": [ + 161, + 185, + 244, + 196 + ], + "spans": [ + { + "bbox": [ + 161, + 185, + 244, + 196 + ], + "type": "text", + "content": "(a) Full-white patch" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_caption" + } + ], + "index": 2 + }, + { + "type": "image", + "bbox": [ + 255, + 114, + 359, + 183 + ], + "blocks": [ + { + "bbox": [ + 255, + 114, + 359, + 183 + ], + "lines": [ + { + "bbox": [ + 255, + 114, + 359, + 183 + ], + "spans": [ + { + "bbox": [ + 255, + 114, + 359, + 183 + ], + "type": "image", + "image_path": "07bcf3370924edbbfb8cf79914ed27088acbe16670c23e7bbd4db125b06eeba6.jpg" + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 269, + 185, + 353, + 196 + ], + "lines": [ + { + "bbox": [ + 269, + 185, + 353, + 196 + ], + "spans": [ + { + "bbox": [ + 269, + 185, + 353, + 196 + ], + "type": "text", + "content": "(b) Full-black patch" + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 131, + 205, + 482, + 228 + ], + "lines": [ + { + "bbox": [ + 131, + 205, + 482, + 228 + ], + "spans": [ + { + "bbox": [ + 131, + 205, + 482, + 228 + ], + "type": "text", + "content": "Fig. 6: The attack testing using real image, full-white, and full-black patches on the MOT system. These patches have the same size of " + }, + { + "bbox": [ + 131, + 205, + 482, + 228 + ], + "type": "inline_equation", + "content": "50 \\times 50\\mathrm{cm}" + }, + { + "bbox": [ + 131, + 205, + 482, + 228 + ], + "type": "text", + "content": " as ours." + } + ] + } + ], + "index": 8, + "angle": 0, + "type": "image_caption" + } + ], + "index": 4 + }, + { + "type": "image", + "bbox": [ + 362, + 114, + 470, + 183 + ], + "blocks": [ + { + "bbox": [ + 362, + 114, + 470, + 183 + ], + "lines": [ + { + "bbox": [ + 362, + 114, + 470, + 183 + ], + "spans": [ + { + "bbox": [ + 362, + 114, + 470, + 183 + ], + "type": "image", + "image_path": "2b9a8610d269ad49cc294f3110cc21bf1c8810ff176494313f689dee3cc0e81c.jpg" + } + ] + } + ], + "index": 6, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 373, + 185, + 461, + 195 + ], + "lines": [ + { + "bbox": [ + 373, + 185, + 461, + 195 + ], + "spans": [ + { + "bbox": [ + 373, + 185, + 461, + 195 + ], + "type": "text", + "content": "(c) Real image patch" + } + ] + } + ], + "index": 7, + "angle": 0, + "type": "image_caption" + } + ], + "index": 6 + }, + { + "type": "table", + "bbox": [ + 181, + 262, + 432, + 373 + ], + "blocks": [ + { + "bbox": [ + 182, + 241, + 430, + 253 + ], + "lines": [ + { + "bbox": [ + 182, + 241, + 430, + 253 + ], + "spans": [ + { + "bbox": [ + 182, + 241, + 430, + 253 + ], + "type": "text", + "content": "Table 5: The effectiveness analysis of the adversarial losses." + } + ] + } + ], + "index": 9, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 181, + 262, + 432, + 373 + ], + "lines": [ + { + "bbox": [ + 181, + 262, + 432, + 373 + ], + "spans": [ + { + "bbox": [ + 181, + 262, + 432, + 373 + ], + "type": "table", + "html": "
VictimAttackerMetrics
TASRSATSRIOR
ByteTrack [38]PapMOT w/o Lbbr61.364.77.2
PapMOT w/o Lap21.926.852.6
PapMOT94.689.679.5
FairMOT [39]PapMOT w/o Lbbr50.543.35.6
PapMOT w/o Lap11.013.129.3
PapMOT76.893.046.2
", + "image_path": "fff639955aab891474a854513ed557c5c1c0a3e47fca16f7869b82d32d57a602.jpg" + } + ] + } + ], + "index": 10, + "angle": 0, + "type": "table_body" + } + ], + "index": 10 + }, + { + "bbox": [ + 130, + 398, + 482, + 459 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 398, + 482, + 459 + ], + "spans": [ + { + "bbox": [ + 130, + 398, + 482, + 459 + ], + "type": "text", + "content": "and the ID remains unchanged throughout the whole video sequence. Similarly, although the target is partially occluded by (c) the real-image patch, it is tracked steadily and the ID is maintained at 1 throughout the entire video sequence. The experiments indicate that it is difficult to attack MOT systems with adversarial patches that are not optimized." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 132, + 480, + 220, + 493 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 480, + 220, + 493 + ], + "spans": [ + { + "bbox": [ + 132, + 480, + 220, + 493 + ], + "type": "text", + "content": "5 Conclusion" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 130, + 509, + 482, + 640 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 509, + 482, + 640 + ], + "spans": [ + { + "bbox": [ + 130, + 509, + 482, + 640 + ], + "type": "text", + "content": "In this paper, we introduce a novel physical adversarial attack method named PapMOT against multiple object tracking systems. Unlike all previous attacks that inject noise into input video frames in the digital domain, our approach optimizes printable adversarial patches that effectively disrupt the data association of MOT trackers across frames in the actual physical world. We demonstrate the feasibility and effectiveness of PapMOT in both digital and physical scenarios through extensive experiments on multiple datasets. In future work, we plan to investigate the creation of more natural-looking adversarial patches, such as transforming these patches into clothing to deceive the MOT system. Additionally, we will expand physical adversarial MOT attacks from pedestrian scenarios to autonomous driving scenarios." + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 131, + 641, + 481, + 666 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 131, + 641, + 481, + 666 + ], + "spans": [ + { + "bbox": [ + 131, + 641, + 481, + 666 + ], + "type": "text", + "content": "Acknowledgements. This work was supported in part by NSFC (62322113, 62376156)." + } + ] + } + ], + "index": 14 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "spans": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "type": "text", + "content": "14" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 166, + 91, + 222, + 102 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 166, + 91, + 222, + 102 + ], + "spans": [ + { + "bbox": [ + 166, + 91, + 222, + 102 + ], + "type": "text", + "content": "J. Long et al." + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 13 + }, + { + "para_blocks": [ + { + "bbox": [ + 132, + 114, + 197, + 126 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 114, + 197, + 126 + ], + "spans": [ + { + "bbox": [ + 132, + 114, + 197, + 126 + ], + "type": "text", + "content": "References" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 138, + 137, + 480, + 665 + ], + "type": "list", + "angle": 0, + "index": 20, + "blocks": [ + { + "bbox": [ + 138, + 137, + 480, + 158 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 137, + 480, + 158 + ], + "spans": [ + { + "bbox": [ + 138, + 137, + 480, + 158 + ], + "type": "text", + "content": "1. Athalye, A., Engstrom, L., Ilyas, A., Kwok, K.: Synthesizing robust adversarial examples. In: International conference on machine learning (2018) 4, 5" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 138, + 159, + 480, + 190 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 159, + 480, + 190 + ], + "spans": [ + { + "bbox": [ + 138, + 159, + 480, + 190 + ], + "type": "text", + "content": "2. Authors, P.: Paddledetection, object detection and instance segmentation toolkit based on paddlepaddle. https://github.com/PaddlePaddle/PaddleDetection (2019) 9" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 138, + 191, + 480, + 221 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 191, + 480, + 221 + ], + "spans": [ + { + "bbox": [ + 138, + 191, + 480, + 221 + ], + "type": "text", + "content": "3. Bernardin, K., Stiefelhagen, R.: Evaluating multiple object tracking performance: the clear mot metrics. EURASIP Journal on Image and Video Processing (2008) 8" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 138, + 223, + 480, + 255 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 223, + 480, + 255 + ], + "spans": [ + { + "bbox": [ + 138, + 223, + 480, + 255 + ], + "type": "text", + "content": "4. Cao, J., Weng, X., Khirodkar, R., Pang, J., Kitani, K.: Observation-centric sort: Rethinking sort for robust multi-object tracking. arXiv preprint arXiv:2203.14360 (2022) 3" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 138, + 255, + 480, + 287 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 255, + 480, + 287 + ], + "spans": [ + { + "bbox": [ + 138, + 255, + 480, + 287 + ], + "type": "text", + "content": "5. Chiu, H.K., Li, J., Ambrus, R., Bohg, J.: Probabilistic 3d multi-modal, multi-object tracking for autonomous driving. In: IEEE International Conference on Robotics and Automation (2021) 1" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 138, + 288, + 480, + 319 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 288, + 480, + 319 + ], + "spans": [ + { + "bbox": [ + 138, + 288, + 480, + 319 + ], + "type": "text", + "content": "6. Chow, K.H., Liu, L., Gursoy, M.E., Truex, S., Wei, W., Wu, Y.: Understanding object detection through an adversarial lens. In: European Symposium on Research in Computer Security (2020) 9, 10, 11" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 138, + 320, + 480, + 352 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 320, + 480, + 352 + ], + "spans": [ + { + "bbox": [ + 138, + 320, + 480, + 352 + ], + "type": "text", + "content": "7. Dendorfer, P., Rezatofighi, H., Milan, A., Shi, J., Cremers, D., Reid, I., Roth, S., Schindler, K., Leal-Taixe, L.: Mot20: A benchmark for multi object tracking in crowded scenes. arXiv preprint arXiv:2003.09003 (2020) 9" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 138, + 353, + 480, + 384 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 353, + 480, + 384 + ], + "spans": [ + { + "bbox": [ + 138, + 353, + 480, + 384 + ], + "type": "text", + "content": "8. Dollar, P., Wojek, C., Schiele, B., Perona, P.: Pedestrian detection: A benchmark. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (2009) 9" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 138, + 384, + 480, + 427 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 384, + 480, + 427 + ], + "spans": [ + { + "bbox": [ + 138, + 384, + 480, + 427 + ], + "type": "text", + "content": "9. Du, A., Chen, B., Chin, T.J., Law, Y.W., Sasdelli, M., Rajasegaran, R., Campbell, D.: Physical adversarial attacks on an aerial imagery object detector. In: Proceedings of the IEEE/CVF Winter Conference on Applications of Computer Vision (2022) 2" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 138, + 428, + 480, + 460 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 428, + 480, + 460 + ], + "spans": [ + { + "bbox": [ + 138, + 428, + 480, + 460 + ], + "type": "text", + "content": "10. Ess, A., Leibe, B., Schindler, K., van Gool, L.: A mobile vision system for robust multi-person tracking. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (2008) 9" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 138, + 460, + 480, + 482 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 460, + 480, + 482 + ], + "spans": [ + { + "bbox": [ + 138, + 460, + 480, + 482 + ], + "type": "text", + "content": "1. Ge, Z., Liu, S., Wang, F., Li, Z., Sun, J.: Yolox: Exceeding yolo series in 2021. arXiv preprint arXiv:2107.08430 (2021) 9" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 138, + 483, + 480, + 514 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 483, + 480, + 514 + ], + "spans": [ + { + "bbox": [ + 138, + 483, + 480, + 514 + ], + "type": "text", + "content": "2. Hu, Y.C.T., Kung, B.H., Tan, D.S., Chen, J.C., Hua, K.L., Cheng, W.H.: Naturalistic physical adversarial patch for object detectors. In: Proceedings of the IEEE/CVF International Conference on Computer Vision (2021) 2, 4, 6" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 138, + 514, + 480, + 546 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 514, + 480, + 546 + ], + "spans": [ + { + "bbox": [ + 138, + 514, + 480, + 546 + ], + "type": "text", + "content": "3. Hu, Z., Huang, S., Zhu, X., Sun, F., Zhang, B., Hu, X.: Adversarial texture for fooling person detectors in the physical world. In: Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition (2022) 2, 4, 9, 10, 11" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 138, + 547, + 480, + 568 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 547, + 480, + 568 + ], + "spans": [ + { + "bbox": [ + 138, + 547, + 480, + 568 + ], + "type": "text", + "content": "4. Jia, S., Ma, C., Song, Y., Yang, X.: Robust tracking against adversarial attacks. In: European Conference on Computer Vision (2020) 4" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 138, + 568, + 480, + 600 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 568, + 480, + 600 + ], + "spans": [ + { + "bbox": [ + 138, + 568, + 480, + 600 + ], + "type": "text", + "content": "5. Jia, S., Ma, C., Yao, T., Yin, B., Ding, S., Yang, X.: Exploring frequency adversarial attacks for face forgery detection. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (2022) 4" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 138, + 601, + 480, + 632 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 601, + 480, + 632 + ], + "spans": [ + { + "bbox": [ + 138, + 601, + 480, + 632 + ], + "type": "text", + "content": "6. Jia, S., Song, Y., Ma, C., Yang, X.: Iou attack: Towards temporally coherent blackbox adversarial attack for visual object tracking. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (2021) 4" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 138, + 633, + 480, + 665 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 633, + 480, + 665 + ], + "spans": [ + { + "bbox": [ + 138, + 633, + 480, + 665 + ], + "type": "text", + "content": "7. Jia, Y., Lu, Y., Shen, J., Chen, Q.A., Zhong, Z., Wei, T.: Fooling detection alone is not enough: Adversarial attack against multiple object tracking. In: International Conference on Learning Representations (2020) 1, 2, 4, 7, 9, 10, 11" + } + ] + } + ], + "index": 19 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 200, + 91, + 447, + 102 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 200, + 91, + 447, + 102 + ], + "spans": [ + { + "bbox": [ + 200, + 91, + 447, + 102 + ], + "type": "text", + "content": "PapMOT: Exploring Adversarial Patch Attack against MOT" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 470, + 92, + 480, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 470, + 92, + 480, + 100 + ], + "spans": [ + { + "bbox": [ + 470, + 92, + 480, + 100 + ], + "type": "text", + "content": "15" + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 14 + }, + { + "para_blocks": [ + { + "bbox": [ + 132, + 116, + 480, + 666 + ], + "type": "list", + "angle": 0, + "index": 21, + "blocks": [ + { + "bbox": [ + 133, + 116, + 480, + 138 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 116, + 480, + 138 + ], + "spans": [ + { + "bbox": [ + 133, + 116, + 480, + 138 + ], + "type": "text", + "content": "18. Kuhn, H.W.: The hungarian method for the assignment problem. Naval research logistics quarterly 2(1-2), 83-97 (1955) 6" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 133, + 138, + 480, + 171 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 138, + 480, + 171 + ], + "spans": [ + { + "bbox": [ + 133, + 138, + 480, + 171 + ], + "type": "text", + "content": "19. Leal-Taixe, L., Milan, A., Reid, I., Roth, S., Schindler, K.: Motchallenge 2015: Towards a benchmark for multi-target tracking. arXiv preprint arXiv:1504.01942 (2015) 9" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 132, + 171, + 480, + 203 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 171, + 480, + 203 + ], + "spans": [ + { + "bbox": [ + 132, + 171, + 480, + 203 + ], + "type": "text", + "content": "20. Lin, D., Chen, Q., Zhou, C., He, K.: Trasw: Tracklet-switch adversarial attacks against multi-object tracking. arXiv preprint arXiv:2111.08954 (2023) 1, 2, 4, 7, 9, 10, 11" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 132, + 203, + 480, + 224 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 203, + 480, + 224 + ], + "spans": [ + { + "bbox": [ + 132, + 203, + 480, + 224 + ], + "type": "text", + "content": "21. Luisier, F., Blu, T., Unser, M.: Image denoising in mixed poisson-gaussian noise. IEEE Transactions on image processing (2010) 9, 10, 11" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 132, + 224, + 480, + 246 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 224, + 480, + 246 + ], + "spans": [ + { + "bbox": [ + 132, + 224, + 480, + 246 + ], + "type": "text", + "content": "22. Luo, W., Xing, J., Milan, A., Zhang, X., Liu, W., Kim, T.K.: Multiple object tracking: A literature review. Artificial intelligence (2021) 1, 3" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 132, + 246, + 480, + 267 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 246, + 480, + 267 + ], + "spans": [ + { + "bbox": [ + 132, + 246, + 480, + 267 + ], + "type": "text", + "content": "23. Milan, A., Leal-Taixe, L., Reid, I., Roth, S., Schindler, K.: Mot16: A benchmark for multi-object tracking. arXiv preprint arXiv:1603.00831 (2016) 9" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 132, + 267, + 480, + 289 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 267, + 480, + 289 + ], + "spans": [ + { + "bbox": [ + 132, + 267, + 480, + 289 + ], + "type": "text", + "content": "24. Redmon, J., Farhadi, A.: Yolo9000: Better, faster, stronger. In: IEEE Conference on Computer Vision and Pattern Recognition (2017) 4" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 132, + 289, + 480, + 310 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 289, + 480, + 310 + ], + "spans": [ + { + "bbox": [ + 132, + 289, + 480, + 310 + ], + "type": "text", + "content": "25. Redmon, J., Farhadi, A.: Yolov3: An incremental improvement. arXiv preprint arXiv:1804.02767 (2018) 4" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 132, + 310, + 480, + 342 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 310, + 480, + 342 + ], + "spans": [ + { + "bbox": [ + 132, + 310, + 480, + 342 + ], + "type": "text", + "content": "26. Shao, S., Zhao, Z., Li, B., Xiao, T., Yu, G., Zhang, X., Sun, J.: Crowdhuman: A benchmark for detecting human in a crowd. arXiv preprint arXiv:1805.00123 (2018) 9" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 132, + 342, + 480, + 374 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 342, + 480, + 374 + ], + "spans": [ + { + "bbox": [ + 132, + 342, + 480, + 374 + ], + "type": "text", + "content": "27. Sharif, M., Bhagavatula, S., Bauer, L., Reiter, M.K.: Accessorize to a crime. In: Proceedings of the 2016 ACM SIGSAC Conference on Computer and Communications Security (2016) 6" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 132, + 374, + 480, + 417 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 374, + 480, + 417 + ], + "spans": [ + { + "bbox": [ + 132, + 374, + 480, + 417 + ], + "type": "text", + "content": "28. Thys, S., Van Ranst, W., Goedemé, T.: Fooling automated surveillance cameras: adversarial patches to attack person detection. In: Proceedings of the IEEE/CVF conference on computer vision and pattern recognition workshops (2019) 2, 4, 9, 10, 11" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 132, + 417, + 480, + 460 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 417, + 480, + 460 + ], + "spans": [ + { + "bbox": [ + 132, + 417, + 480, + 460 + ], + "type": "text", + "content": "29. Wang, D., Jiang, T., Sun, J., Zhou, W., Gong, Z., Zhang, X., Yao, W., Chen, X.: Fca: Learning a 3d full-coverage vehicle camouflage for multi-view physical adversarial attack. In: Proceedings of the AAAI conference on artificial intelligence (2022) 6" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 132, + 460, + 455, + 472 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 460, + 455, + 472 + ], + "spans": [ + { + "bbox": [ + 132, + 460, + 455, + 472 + ], + "type": "text", + "content": "30. Welch, G., Bishop, G., et al.: An introduction to the kalman filter (1995) 6" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 132, + 472, + 480, + 502 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 472, + 480, + 502 + ], + "spans": [ + { + "bbox": [ + 132, + 472, + 480, + 502 + ], + "type": "text", + "content": "31. Wojke, N., Bewley, A., Paulus, D.: Simple online and realtime tracking with a deep association metric. In: IEEE International Conference on Image Processing (2017) 3" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 132, + 502, + 480, + 536 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 502, + 480, + 536 + ], + "spans": [ + { + "bbox": [ + 132, + 502, + 480, + 536 + ], + "type": "text", + "content": "32. Xiao, T., Li, S., Wang, B., Lin, L., Wang, X.: Joint detection and identification feature learning for person search. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (2017) 9" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 132, + 536, + 480, + 567 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 536, + 480, + 567 + ], + "spans": [ + { + "bbox": [ + 132, + 536, + 480, + 567 + ], + "type": "text", + "content": "33. Xu, H., Ma, Y., Liu, H.C., Deb, D., Liu, H., Tang, J.L., Jain, A.K.: Adversarial attacks and defenses in images, graphs and text: A review. International Journal of Automation and Computing (2020) 4" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 132, + 567, + 480, + 601 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 567, + 480, + 601 + ], + "spans": [ + { + "bbox": [ + 132, + 567, + 480, + 601 + ], + "type": "text", + "content": "34. Xu, K., Zhang, G., Liu, S., Fan, Q., Sun, M., Chen, H., Chen, P.Y., Wang, Y., Lin, X.: Adversarial t-shirt! evading person detectors in a physical world. In: European Conference on Computer Vision (2020) 4" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 132, + 601, + 480, + 633 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 601, + 480, + 633 + ], + "spans": [ + { + "bbox": [ + 132, + 601, + 480, + 633 + ], + "type": "text", + "content": "35. Yang, X., Wei, F., Zhang, H., Zhu, J.: Design and interpretation of universal adversarial patches in face detection. In: European Conference on Computer Vision (2020) 2" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 132, + 633, + 480, + 666 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 633, + 480, + 666 + ], + "spans": [ + { + "bbox": [ + 132, + 633, + 480, + 666 + ], + "type": "text", + "content": "36. Yu, F., Chen, H., Wang, X., Xian, W., Chen, Y., Liu, F., Madhavan, V., Darrell, T.: Bdd100k: A diverse driving dataset for heterogeneous multitask learning. In: IEEE/CVF Conference on Computer Vision and Pattern Recognition (2020) 9" + } + ] + } + ], + "index": 20 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "spans": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "type": "text", + "content": "16" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 166, + 91, + 221, + 101 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 166, + 91, + 221, + 101 + ], + "spans": [ + { + "bbox": [ + 166, + 91, + 221, + 101 + ], + "type": "text", + "content": "J. Long et al." + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 15 + }, + { + "para_blocks": [ + { + "bbox": [ + 130, + 116, + 482, + 270 + ], + "type": "list", + "angle": 0, + "index": 7, + "blocks": [ + { + "bbox": [ + 130, + 116, + 482, + 149 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 116, + 482, + 149 + ], + "spans": [ + { + "bbox": [ + 130, + 116, + 482, + 149 + ], + "type": "text", + "content": "37. Zhang, S., Benenson, R., Schiele, B.: Citypersons: A diverse dataset for pedestrian detection. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (2017) 9" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 130, + 150, + 482, + 183 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 150, + 482, + 183 + ], + "spans": [ + { + "bbox": [ + 130, + 150, + 482, + 183 + ], + "type": "text", + "content": "38. Zhang, Y., Sun, P., Jiang, Y., Yu, D., Weng, F., Yuan, Z., Luo, P., Liu, W., Wang, X.: Bytetrack: Multi-object tracking by associating every detection box. In: European Conference on Computer Vision (2022) 3, 4, 9, 14" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 130, + 183, + 482, + 215 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 183, + 482, + 215 + ], + "spans": [ + { + "bbox": [ + 130, + 183, + 482, + 215 + ], + "type": "text", + "content": "39. Zhang, Y., Wang, C., Wang, X., Zeng, W., Liu, W.: Fairmot: On the fairness of detection and re-identification in multiple object tracking. International Journal of Computer Vision (2021) 3, 4, 9, 14" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 130, + 216, + 482, + 237 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 216, + 482, + 237 + ], + "spans": [ + { + "bbox": [ + 130, + 216, + 482, + 237 + ], + "type": "text", + "content": "40. Zhao, Z., Wu, Z., Zhuang, Y., Li, B., Jia, J.: Tracking objects as pixel-wise distributions. In: European Conference on Computer Vision (2022) 3" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 130, + 237, + 482, + 270 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 237, + 482, + 270 + ], + "spans": [ + { + "bbox": [ + 130, + 237, + 482, + 270 + ], + "type": "text", + "content": "41. Zheng, L., Zhang, H., Sun, S., Chandraker, M., Yang, Y., Tian, Q.: Person re-identification in the wild. In: Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition (2017) 9" + } + ] + } + ], + "index": 6 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 200, + 91, + 447, + 102 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 200, + 91, + 447, + 102 + ], + "spans": [ + { + "bbox": [ + 200, + 91, + 447, + 102 + ], + "type": "text", + "content": "PapMOT: Exploring Adversarial Patch Attack against MOT" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 470, + 92, + 480, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 470, + 92, + 480, + 100 + ], + "spans": [ + { + "bbox": [ + 470, + 92, + 480, + 100 + ], + "type": "text", + "content": "17" + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 16 + } + ], + "_backend": "vlm", + "_version_name": "2.6.4" +} \ No newline at end of file diff --git a/2024/ParCo_ Part-Coordinating Text-to-Motion Synthesis/5d0214df-d42e-4bdf-a60a-2b013c23d49a_content_list.json b/2024/ParCo_ Part-Coordinating Text-to-Motion Synthesis/5d0214df-d42e-4bdf-a60a-2b013c23d49a_content_list.json new file mode 100644 index 0000000000000000000000000000000000000000..3f8b89f08d8f7bd0ad99d6cdd1a5df48832834f9 --- /dev/null +++ b/2024/ParCo_ Part-Coordinating Text-to-Motion Synthesis/5d0214df-d42e-4bdf-a60a-2b013c23d49a_content_list.json @@ -0,0 +1,1660 @@ +[ + { + "type": "text", + "text": "ParCo: Part-Coordinating Text-to-Motion Synthesis", + "text_level": 1, + "bbox": [ + 259, + 140, + 743, + 186 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Qiran Zou\\*, Shangyuan Yuan\\*1, Shian Du', Yu Wang?, Chang Liu?, Yi Xu?, Jie Chen?, and Xiangyang Ji!", + "bbox": [ + 232, + 210, + 772, + 244 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "$^{1}$ Tsinghua University, China \n $^{2}$ Peking University Shenzhen Graduate School, China", + "bbox": [ + 316, + 253, + 684, + 282 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "$^{3}$ Dalian University of Technology, China qiranzou@gmail.com, {yuansy21, dsa23}@mails.tsinghua.edu.cn, 2201212856@stu.pku.edu.cn, liuchang2022@tsinghua.edu.cn, yxu@dlut.edu.cn, chenj@pcl.ac.cn, xyji@tsinghua.edu.cn", + "bbox": [ + 225, + 282, + 774, + 338 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Abstract. We study a challenging task: text-to-motion synthesis, aiming to generate motions that align with textual descriptions and exhibit coordinated movements. Currently, the part-based methods introduce part partition into the motion synthesis process to achieve finer-grained generation. However, these methods encounter challenges such as the lack of coordination between different part motions and difficulties for networks to understand part concepts. Moreover, introducing finer-grained part concepts poses computational complexity challenges. In this paper, we propose Part-Coordinating Text-to-Motion Synthesis (ParCo), endowed with enhanced capabilities for understanding part motions and communication among different part motion generators, ensuring a coordinated and fined-grained motion synthesis. Specifically, we discretize whole-body motion into multiple part motions to establish the prior concept of different parts. Afterward, we employ multiple lightweight generators designed to synthesize different part motions and coordinate them through our part coordination module. Our approach demonstrates superior performance on common benchmarks with economic computations, including HumanML3D and KIT-ML, providing substantial evidence of its effectiveness. Code is available at: https://github.com/qrzou/ParCo.", + "bbox": [ + 259, + 377, + 743, + 654 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Keywords: Motion synthesis $\\cdot$ Part coordination $\\cdot$ Text-to-motion", + "bbox": [ + 261, + 667, + 712, + 681 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "1 Introduction", + "text_level": 1, + "bbox": [ + 215, + 708, + 377, + 724 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Text-to-motion synthesis aims to generate motion that aligns with textual descriptions and exhibits coordinated movements. It facilitates obtaining desired motion through textual descriptions which benefits numerous applications in industrial scenarios such as animation [23], AR/VR applications, video games [37, 65], autonomous driving [10], and robotics [3,26,27].", + "bbox": [ + 212, + 739, + 787, + 816 + ], + "page_idx": 0 + }, + { + "type": "page_footnote", + "text": "* Equal contribution.", + "bbox": [ + 217, + 824, + 364, + 839 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "\"am is raising his left hand while kicking something with his right foot\"", + "bbox": [ + 346, + 161, + 658, + 172 + ], + "page_idx": 1 + }, + { + "type": "image", + "img_path": "images/e6ad0c4221961cca5a2376237fc46ca9d144a908a77881126e96cda2c07efcd1.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 228, + 172, + 776, + 244 + ], + "page_idx": 1 + }, + { + "type": "image", + "img_path": "images/90d386a8f15cfc164f3f8952b2cc9b611fe6d7ee49dde270eb4231c1d2c7f6fb.jpg", + "image_caption": [ + "Fig. 1: Our ParCo is capable of coordinating the motion of various body parts to produce realistic and accurate motion." + ], + "image_footnote": [], + "bbox": [ + 228, + 247, + 776, + 321 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Recent advancements leveraging powerful generation capabilities of transformers [33,42,62] and diffusion models [22,41,51,53,54] have yielded impressive results in generating realistic and smooth motions. Despite this progress, existing methods often struggle to generate semantically matched and coherently coordinated motions, especially when faced with commands involving multiple coordinated body parts. We attribute this challenge to the intricate alignment problem between the modalities of text and motion, where a single text/motion can correspond to multiple possible motion/text, posing a challenge in learning the complex relationship between the two.", + "bbox": [ + 212, + 371, + 784, + 507 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "In the realm of Text-to-Motion Generation, part-based methods aim to achieve a higher level of sophistication in motion generation. These approaches can be broadly classified into two categories: single generator with part-level motion embeddings (Fig. 2 (a)) and independent upper and lower body motion generators (Fig. 2 (b)). The former one constructs whole-body motion embeddings by amalgamating multiple part motion embeddings, explicitly introducing the concept of parts [57], or implicitly integrating part concepts with part-level attention [71]. However, utilizing a single generator to generate whole-body motion embeddings presents a challenge for the generator to understand the concept of parts. The latter one segregates whole-body motion into upper and lower body motions, deploying two independent generators to produce motions separately [13]. Although this design enhances the generator's comprehension of upper and lower body motions, the absence of information exchange between them leads to a lack of coordination in the resulting upper and lower body motions. To increase the granularity of part division, an intuitive solution is to employ additional generators for generating finer-grained part motions, such as the left leg. However, this approach introduces computational complexity challenges.", + "bbox": [ + 212, + 507, + 785, + 763 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Neuroscientific discoveries reveal that discrete regions within the human brain manifest unique functions, and these regions engage in communication to coordinate diverse activities [19, 55]. This design, where different low-level subsystems communicate to form a higher-level system, is prevalent in the natural and artificial world, offering a robust, perceptually strong structural design. In adher-", + "bbox": [ + 212, + 763, + 785, + 839 + ], + "page_idx": 1 + }, + { + "type": "page_number", + "text": "2", + "bbox": [ + 217, + 114, + 228, + 126 + ], + "page_idx": 1 + }, + { + "type": "header", + "text": "Q. Zou et al.", + "bbox": [ + 271, + 114, + 359, + 128 + ], + "page_idx": 1 + }, + { + "type": "image", + "img_path": "images/e59bce9596372b2b546709d57639cb47f437108783eda0ae41831c3d5d1439c5.jpg", + "image_caption": [ + "Fig. 2: Conceptual comparison of three part-based synthesis methods. (a): One generator synthesizes the whole-body embedding, which contains information about different parts internally. (b): Two separate generators synthesize the upper and lower body's motions independently, without information exchange between them. (c): Our ParCo employs multiple lightweight generators designed to synthesize different part motions, which are coordinated by the Part Coordination module." + ], + "image_footnote": [], + "bbox": [ + 218, + 143, + 398, + 286 + ], + "page_idx": 2 + }, + { + "type": "image", + "img_path": "images/0a24dc73775f6e81ea18b8f13d7c54a90b4dbf191702fe192e17eb9a2bc2322c.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 408, + 143, + 589, + 286 + ], + "page_idx": 2 + }, + { + "type": "image", + "img_path": "images/7d86fb4db43edb9f2676c60a35762d2e95e2b9278e6c329aa7d3bb6cc605389b.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 604, + 143, + 787, + 286 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "ence to these principles, we present Part-Coordinating Text-to-Motion Synthesis (ParCo). It comprises six small generators, which are tasked with various part motions, and accompanied by a Part Coordination module facilitating communication among the generators. This communication enables the generation of coordinated whole-body motions while comprehending distinct parts (Fig. 2 (c)). Thanks to the design of small generators, our method demonstrates a lower parameter count, reduced computational complexity, and shorter generation time in comparison to baseline and state-of-the-art methods.", + "bbox": [ + 212, + 412, + 787, + 532 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Specifically, our approach consists of two stages. In the first stage, we discretize whole-body motion into multiple part motions and quantize them using VQ-VAEs, providing prior knowledge of \"what is part\" for the next stage. In the second stage, we use multiple Part-Coordinated Transformers, which are capable of communicating with each other, to generate coordinated motions of different parts. These part motions are integrated into whole-body motion subsequently. Extensive experiments on HumanML3D [14] and KIT-ML [47] demonstrate that our method can generate realistic and coordinated motions that align with the semantic descriptions.", + "bbox": [ + 212, + 534, + 787, + 669 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "We delineate our contributions as follows:", + "bbox": [ + 238, + 670, + 540, + 684 + ], + "page_idx": 2 + }, + { + "type": "list", + "sub_type": "text", + "list_items": [ + "- We propose ParCo, which enables the generators to better understand finer-grained parts and coordinate the generated part motions, ultimately achieving fine-grained and coordinated motion synthesis.", + "- Our approach is computationally efficient, although employing multiple part generators, and maintains excellent motion generation performance.", + "- On the HumanML3D and KIT-ML datasets, our method significantly outperforms methods that do not rely on GT motion length as our ParCo, while demonstrating comparable performance to methods that do rely on GT motion length." + ], + "bbox": [ + 233, + 691, + 785, + 839 + ], + "page_idx": 2 + }, + { + "type": "header", + "text": "ParCo: Part-Coordinating Text-to-Motion Synthesis", + "bbox": [ + 383, + 114, + 732, + 128 + ], + "page_idx": 2 + }, + { + "type": "page_number", + "text": "3", + "bbox": [ + 774, + 116, + 784, + 126 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "2 Related Work", + "text_level": 1, + "bbox": [ + 217, + 143, + 387, + 160 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Human Motion Synthesis. Tasks in the domain of human motion synthesis fall within two distinct categories: unconditional motion generation and conditional motion generation. The categorization of these tasks is based on the input signals employed. Unconditional motion generation [49, 64, 69, 70], such as VPoser [43] and ACTOR [45], is a comprehensive task involving the modeling of the entire motion space, utilizing solely motion data for training and prediction. Human motion prediction, a highly dynamic field, seeks to forecast future movements based on observed motion. Another significant domain pertains to the generation of \"in-betweening\" motions, which fill the gaps between past and future poses [11, 17, 18, 24, 58]. Unconditional motion generation commonly utilizes models well-suited for processing sequential data, including recursive [7, 12, 40, 44], generative adversarial [5, 20], graph convolutional [39], and attention [38] approaches. This enables the efficient generation of diverse motions by concurrently processing spatial and temporal signals. Conditional motion generation involves various multimodal data types, including text [2, 14, 15, 25, 46, 61], occluded pose sequences [11, 18, 61], images [9, 52], and sound [30-32]. Due to the rapid advancements in NLP, text-driven human motion generation has sustained a notably active status.", + "bbox": [ + 217, + 179, + 785, + 450 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Text-driven Human Motion Generation. Text-to-motion aims to generate human motion based on input textual descriptions. In earlier studies, joint-latent models [2, 46] were employed, which integrate a text encoder and a motion encoder. Text2Action [1] employs a recursive model, generating motion from short texts. TEMOS [46] adopts a similar approach, using self-encoding structures for both text and motion constrained by KL divergence [28]. T2M-GPT [66], TM2T [15] replace recursive encoders with transformers and GRU structure, achieving promising results. MotionCLIP [60] directly introduces the powerful zero-shot-capable CLIP [50] text encoder, following the alignment of text and pose as in Language2Pose [2], and additionally renders images with CLIP image encoder for auxiliary supervision. Subsequently, solutions based on diffusion models [8,56,67] emerge. MDM [61], MotionDiffuse [67], ReMoDiffuse [68] introduce diffusion models based on probability mapping, enabling the generation of human motion sequences from textual descriptions. Although current research enables the convenient generation of human motions based on text, challenges persist, especially in handling complex textual descriptions involving different parts. These approaches often treat human motion as a whole, lacking an understanding of different body parts, exhibiting limited ability of aligning text and motion. Furthermore, other methods incorporate the concept of parts into the model to generate more granular motions [13,57,71]. However, these approaches encounter challenges, including the lack of coordination among different part motions and difficulties for networks to comprehend part concepts. In contrast, our ParCo demonstrates a superior understanding of parts and effectively coordinates part motions, with lower computational complexity.", + "bbox": [ + 217, + 477, + 785, + 839 + ], + "page_idx": 3 + }, + { + "type": "page_number", + "text": "4", + "bbox": [ + 217, + 114, + 228, + 126 + ], + "page_idx": 3 + }, + { + "type": "header", + "text": "Q. Zou et al.", + "bbox": [ + 271, + 114, + 357, + 128 + ], + "page_idx": 3 + }, + { + "type": "image", + "img_path": "images/cad4d6d83d8ed3ccd48cd341f4bb922398834c8076bb8c320c23c8199f195183.jpg", + "image_caption": [ + "Fig. 3: Pipeline of ParCo. ParCo consists of two stages: (a) The whole-body motion is discretized into 6 part motions, and encoded into 6 quantized code index sequences by 6 VQ-VAEs (encoder and quantizer). This process provides a priori about the concept of part motions for the second stage. (b) We use the quantized index sequences and corresponding textual description to train 6 transformers for part motion generation. At the same time, these generators are coordinated by our Part Coordination module. The generated part motion codes are decoded by VQ-VAE (decoder) to reconstruct the 6 part motions, which will be integrated into the final whole-body motion." + ], + "image_footnote": [], + "bbox": [ + 215, + 143, + 491, + 272 + ], + "page_idx": 4 + }, + { + "type": "image", + "img_path": "images/c75e5c2b492baa1cddeaaf27d33b143669dc267341769b66affca63c7df4b321.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 511, + 146, + 787, + 272 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "3 Method", + "text_level": 1, + "bbox": [ + 215, + 426, + 328, + 441 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Our method consists of two stages to generate motion with an understanding of part motions. In the first stage, we discretize the whole-body motion into multiple part motions to provide prior knowledge of \"what is part\" for the second stage. In the second stage, the objective is to enable the model to learn the concept of part and achieve mutual coordination among multiple part motion generators. With this design, our method can handle textual inputs involving different parts and generate human motion that aligns with the semantic descriptions in the text.", + "bbox": [ + 212, + 459, + 787, + 580 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "3.1 Part-Aware Motion Discretization", + "text_level": 1, + "bbox": [ + 214, + 603, + 542, + 617 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "In stage 1, our method partitions the whole-body motion into multiple part motions and independently encodes each of these part motions using a VQ-VAE. This ensures that each part motion possesses an independent representation space (encoding space), providing prior knowledge about the concept of part motion for the next stage.", + "bbox": [ + 212, + 628, + 785, + 703 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "The 3D human body models (e.g. SMPL, MMM) typically use Kinematic Trees to model human skeleton as 5 chains (limbs and backbone) for motion modeling. We inherit this division and add the Root part to represent trajectories. Therefore, we divide the whole-body motion into six parts motions: R.Arm, L.Arm, R.Leg, L.Leg, Backbone, and Root. As illustrated in Fig. 3 (a), the first four represent the right/left arm and right/left leg, while the backbone denotes the spine and skull, and the root represents the pelvis joint's movement information. The commonly used human motion datasets, HumanML3D [14] and KIT-ML [47], utilize different body skeleton models, SMPL [34] and MMM [59]", + "bbox": [ + 212, + 704, + 787, + 840 + ], + "page_idx": 4 + }, + { + "type": "header", + "text": "ParCo: Part-Coordinating Text-to-Motion Synthesis", + "bbox": [ + 383, + 114, + 732, + 128 + ], + "page_idx": 4 + }, + { + "type": "page_number", + "text": "5", + "bbox": [ + 774, + 116, + 784, + 126 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "respectively. We provide a detailed explanation of how we perform our six-part partitioning for these two skeleton models in the supplementary material.", + "bbox": [ + 212, + 146, + 782, + 176 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "The aforementioned partitioning process can be formalized as: given a motion sequence $M = [m_1, \\dots, m_C]$ , where $C$ is the number of frames, we separate it into part motions $\\{P^i = [p_1^i, \\dots, p_C^i]\\}$ , $i \\in [1, \\dots, S]$ , where $S$ is the number of parts and $m_* = [p_*^1, \\dots, p_*^S]$ .", + "bbox": [ + 212, + 176, + 784, + 237 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "After separating the whole-body motion into part motions, we further discretize the part motions into code sequences using VQ-VAE. We utilize this discretized representation in the next stage's generation process, as it exhibits better generalization capabilities and contributes to the improvement of training and inference efficiency.", + "bbox": [ + 212, + 238, + 782, + 311 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Firstly, we use $Encoder^i$ to get the $i$ -th part motion's encoding that $E^i = Encoder^i(P^i) = [e_1^i, \\dots, e_l^i, \\dots, e_L^i]$ , where $L = \\frac{C}{r}$ and $r$ is the downsampling rate of encoder. Then, we discretize the $E^i$ into $Q^i = [v_{k_1^i}^i, \\dots, v_{k_l^i}^i, \\dots, v_{k_L^i}^i]$ according to a learnable codebook $V^i = \\{v_j^i\\}, j = 1, \\dots, J$ , where $J$ is the number of codes in the codebook. The index $k_l$ is obtained by finding the most similar code:", + "bbox": [ + 214, + 313, + 784, + 393 + ], + "page_idx": 5 + }, + { + "type": "equation", + "text": "\n$$\nk _ {l} ^ {i} = \\underset {j \\in \\{1, \\dots , J \\}} {\\arg \\min } \\left\\| e _ {l} ^ {i} - v _ {j} ^ {i} \\right\\|. \\tag {1}\n$$\n", + "text_format": "latex", + "bbox": [ + 406, + 405, + 784, + 433 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "By this way, we discretize the motion $M$ into $S$ discretized part motion representations $\\{Q^i\\}$ , $i = 1, \\dots, S$ .", + "bbox": [ + 212, + 444, + 782, + 473 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "For training the VQ-VAE, We use a decoder to reconstruct the $i$ -th part motion $\\hat{P}^i = \\text{Decoder}^i(Q^i)$ with reconstruction loss $\\mathcal{L}_r^i = \\| \\hat{P}^i - P^i\\|$ . The optimization objective of the $i$ -th part's VQ-VAE is:", + "bbox": [ + 212, + 474, + 784, + 521 + ], + "page_idx": 5 + }, + { + "type": "equation", + "text": "\n$$\n\\mathcal {L} ^ {i} = \\mathcal {L} _ {r} ^ {i} + \\| s g (E ^ {i}) - Q ^ {i} \\| + \\beta \\| E ^ {i} - s g (Q ^ {i}) \\|, \\tag {2}\n$$\n", + "text_format": "latex", + "bbox": [ + 344, + 534, + 784, + 551 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "where $sg$ represents the stop-gradient operation. The first term is the reconstruction loss function, ensuring that the VQ-VAE can reconstruct the original part motion from the encoding. The second term is the codebook loss function. And the third term is the commitment loss, aiming to make the representation $e_l^i$ output by the encoder as close as possible to the code $v_{k_l}^i$ contained in the codebook. The weight of this loss is controlled by the hyperparameter $\\beta$ .", + "bbox": [ + 212, + 561, + 784, + 652 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "3.2 Text-Driven Part Coordination", + "text_level": 1, + "bbox": [ + 214, + 676, + 517, + 691 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "In this stage, we employ transformers as generators to achieve text-to-motion generation. Diverging from the approach of using a large transformer for the entire whole-body motion, we utilize multiple small transformers to generate each part motion code sequence obtained from the previous stage. This allows small transformers to be aware of the conception of \"part motion\" constructed before. However, relying solely on these separate part motion generators can lead to an inability to collectively generate whole-body motions, due to a lack of knowledge about the motions of other parts. To address this, we introduce the Part Coordination module, facilitating communication among all part transformer", + "bbox": [ + 212, + 704, + 784, + 839 + ], + "page_idx": 5 + }, + { + "type": "page_number", + "text": "6", + "bbox": [ + 217, + 114, + 228, + 126 + ], + "page_idx": 5 + }, + { + "type": "header", + "text": "Q. Zou et al.", + "bbox": [ + 271, + 114, + 359, + 128 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/4bb813629d5341c3a151e237d89aeecfe3de8849cf4f0a68dcf05eabf5c7a15a.jpg", + "image_caption": [ + "Fig. 4: The architecture of our Part-Coordinated Transformer." + ], + "image_footnote": [], + "bbox": [ + 300, + 142, + 702, + 327 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "generators to collaboratively generate whole-body motion. This capability enables our ParCo to handle textual inputs involving different parts, generating human motion that aligns with semantic descriptions in the text.", + "bbox": [ + 212, + 390, + 782, + 436 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "As depicted in Fig. 3 (b), we employ six transformers to generate code index sequences of part motions based on the input text. At the same time, these six transformers collaborate through our Part Coordination module to coordinate with each other. Subsequently, these generated index sequences are decoded into the original representation of part motions by the decoder of the VQ-VAE. These part motions are integrated to form whole-body motion.", + "bbox": [ + 212, + 440, + 784, + 531 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Specifically, we model the whole text-to-motion generation process as estimating the distribution $p(M|t)$ of motion $M$ given the text $t$ . Since we have discretized the whole-body motion into part motions, we can model the entire motion distribution $p(M|t)$ through the estimation of conditional distributions $p(K^i |t)$ for each part, where $\\{K^i = [k_1^i,\\dots,k_L^i]\\} , i\\in [1,\\dots,S]$ is the code indices obtained in the first stage.", + "bbox": [ + 212, + 536, + 784, + 626 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "To model $i$ -th part motion's distribution, we propose an autoregressive distribution, which allows parts to coordinate with each other,", + "bbox": [ + 212, + 630, + 782, + 661 + ], + "page_idx": 6 + }, + { + "type": "equation", + "text": "\n$$\n\\begin{array}{l} p \\left(K ^ {i} \\mid t\\right) = \\prod_ {h = 1} ^ {L} p \\left(k _ {h} ^ {i} \\mid k _ {1} ^ {i}, o _ {1} ^ {i}; \\dots ; k _ {h - 1} ^ {i}, o _ {h - 1} ^ {i}; t\\right), \\tag {3} \\\\ o _ {*} ^ {i} = \\left\\{k _ {*} ^ {j} \\right\\}, j \\neq i, j \\in [ 1,..., S ]. \\\\ \\end{array}\n$$\n", + "text_format": "latex", + "bbox": [ + 341, + 681, + 784, + 744 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "When predicting token $k_h^i$ for the $i$ -th part, the prediction not only relies on all tokens predicted by itself from time 1 to $h - 1$ but is also conditioned by predictions from all other parts during the same time span. Also, after the $i$ -th part transformer predicts the token $K_h^i$ , the prediction will be used by all part generators to predict the next token.", + "bbox": [ + 212, + 763, + 784, + 840 + ], + "page_idx": 6 + }, + { + "type": "header", + "text": "ParCo: Part-Coordinating Text-to-Motion Synthesis", + "bbox": [ + 383, + 114, + 732, + 130 + ], + "page_idx": 6 + }, + { + "type": "page_number", + "text": "7", + "bbox": [ + 774, + 114, + 784, + 126 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Finally, we learn the entire body motion by estimating the distributions of all part motions,", + "bbox": [ + 212, + 146, + 782, + 175 + ], + "page_idx": 7 + }, + { + "type": "equation", + "text": "\n$$\n\\begin{array}{l} \\mathcal {L} = \\mathbb {E} _ {M, t \\sim p (M, t)} [ - \\log p (M | t) ] \\\\ = \\mathbb {E} _ {M, t \\sim p (M, t)} \\left[ - \\sum_ {i = 1} ^ {S} \\log p \\left(K ^ {i} | t\\right) \\right]. \\tag {4} \\\\ \\end{array}\n$$\n", + "text_format": "latex", + "bbox": [ + 379, + 176, + 785, + 237 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "We further propose Part-Coordinated Transformer, a transformer capable of coordinating with other transformers, to approximate $p(K^i |t)$ . As shown in Fig. 4, we insert a Part Coordination Layer before each transformer layer (except for the first transformer layer). For each token $x^i$ output by the previous transformer layer, it passes through our ParCo Block. The ParCo Block coordinates with other part motion generators, fusing current token $x^i$ with tokens from other part transformers,", + "bbox": [ + 212, + 241, + 784, + 348 + ], + "page_idx": 7 + }, + { + "type": "equation", + "text": "\n$$\n\\begin{array}{l} x _ {\\text {c o o r d}} ^ {i} = L N \\left(x ^ {i} + M L P ^ {i} (y)\\right), \\tag {5} \\\\ y = \\left\\{x ^ {j} \\right\\}, j \\neq i, j \\in [ 1, \\dots , S ], \\\\ \\end{array}\n$$\n", + "text_format": "latex", + "bbox": [ + 377, + 359, + 785, + 398 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "where $LN$ denotes the LayerNorm operation, and $y$ represents the tokens from other transformers' layers. The fused token $x_{coord}^{i}$ is then input into the subsequent transformer layer.", + "bbox": [ + 212, + 407, + 782, + 453 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "3.3 Discussion", + "text_level": 1, + "bbox": [ + 215, + 476, + 349, + 489 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Text-to-motion synthesis intrinsically consists of two stages, i.e., body part representation and generation, to fulfill actions described by texts. To precisely ground each word to the corresponding body part, Balando [57] partitions the body into upper and lower parts for motion quantization while reconstructing through a shared decoder. SCA [13] steps further to equip independent generators for flexible correspondence. However, coarse-grained sub-body level modeling yields sub-optimal results.", + "bbox": [ + 212, + 501, + 784, + 606 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "AttT2M [71] introduces a global-local attention mechanism to learn hierarchical body-part semantics for accurate motion synthesis. SINC [4] adopts a simple additive composition of part motion for GPT-guided synthetic training data creation. In contrast, we advocate explicitly discretizing body parts as individual action atoms and synthesis motion with decentralized generators and a centralized part coordinating module. With a series of computation-economic designs, we keep the relative independence yet close coordination relationships of body parts and report superior results.", + "bbox": [ + 212, + 607, + 784, + 728 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "4 Experiment", + "text_level": 1, + "bbox": [ + 215, + 752, + 366, + 768 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "4.1 Settings", + "text_level": 1, + "bbox": [ + 215, + 784, + 330, + 799 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Datasets. We utilized two widely used text-to-motion datasets, KIT-ML and HumanML3D, for training and testing our method. Comparative evaluations", + "bbox": [ + 212, + 809, + 782, + 839 + ], + "page_idx": 7 + }, + { + "type": "page_number", + "text": "8", + "bbox": [ + 217, + 114, + 228, + 126 + ], + "page_idx": 7 + }, + { + "type": "header", + "text": "Q. Zou et al.", + "bbox": [ + 271, + 114, + 359, + 128 + ], + "page_idx": 7 + }, + { + "type": "table", + "img_path": "images/ef99595b7da82840c78fd880f20ea8cd69ae250fe7877652205328b0d2a621a9.jpg", + "table_caption": [ + "Table 1: Comparisons to current state-of-the-art methods on HumanML3D test set. “↑” denotes that higher is better. “↓” denotes that lower is better. “→” denotes that results are better if the metric is closer to the real motion. Bold and underlined indicate the best and second-best results, respectively. § reports results using ground-truth motion length. The results of ReMoDiffuse* are obtained from official checkpoints and employ uniform random sampling of motion lengths as input." + ], + "table_footnote": [], + "table_body": "
MethodsR-Precision ↑FID ↓MM-Dist ↓Diversity → MModality ↑
Top-1Top-2Top-3
Real motion0.511±.0030.703±.0030.797±.0020.002±.0002.974±.0089.503±.065-
MDM§ [61]0.320±.0050.498±.0040.611±.0070.544±.0445.566±.0279.559±.0862.799±.072
MLD§ [8]0.481±.0030.673±.0030.772±.0020.473±.0133.196±.0109.724±.0822.413±.079
MotionDiffuse§ [67]0.491±.0010.681±.0010.782±.0010.630±.0013.113±.0019.410±.0491.553±.042
ReMoDiffuse§ [68]0.510±.0050.698±.0060.795±.0040.103±.0042.974±.0169.018±.0751.795±.043
ReMoDiffuse*0.450±.0030.638±.0020.743±.0030.281±.0103.271±.0089.236±.085-
Text2Gesture [6]0.165±.0010.267±.0020.345±.0027.664±.0306.030±.0086.409±.071-
Seq2Seq [48]0.180±.0020.300±.0020.396±.00211.75±.0355.529±.0076.223±.061-
Language2Pose [2]0.246±.0010.387±.0020.486±.00211.02±.0465.296±.0087.676±.058-
Hier [13]0.301±.0020.425±.0020.552±.0046.532±.0245.012±.0188.332±.042-
TEMOS [46]0.424±.0020.612±.0020.722±.0023.734±.0283.703±.0088.973±.0710.368±.018
TM2T [15]0.424±.0030.618±.0030.729±.0021.501±.0173.467±.0118.589±.0762.424±.093
T2M [14]0.457±.0020.639±.0030.740±.0031.067±.0023.340±.0089.188±.0022.090±.083
T2M-GPT [66]0.492±.0030.679±.0020.775±.0020.141±.0053.121±.0099.722±.0821.831±.048
Fg-T2M [63]0.492±.0020.683±.0030.783±.0020.243±.0193.109±.0079.278±.0721.614±.049
AttT2M [71]0.499±.0030.690±.0020.786±.0020.112±.0063.038±.0079.700±.0902.452±.051
ParCo (Ours)0.515±.0030.706±.0030.801±.0020.109±.0052.927±.0089.576±.0881.382±.060
", + "bbox": [ + 217, + 239, + 787, + 484 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "were conducted on these datasets against other existing methods. Processed from KIT [47] and CMU [29] datasets, KIT-ML Dataset [47] comprises 3,911 sequences of human body motions with 6,278 text annotations. Each motion is annotated with 1 to 4 text descriptions, averaging approximately 8 words each. KIT-ML employs the MMM [59] skeletal model which has 21 joints, and we detail in the supplementary materials how we partition its joints into 6 parts. We follow the train, validation, and test set divisions as outlined in [14] and report our ParCo's performance on the test set. HumanML3D [14] Dataset is the largest dataset with 14,616 3D human body motion sequences and 44,970 corresponding textual descriptions, sourced from AMASS [36] and HumanAct12 [16]. Each motion includes a minimum of 3 text descriptions, with an average length of 12 words. HumanML3D adopts the SMPL [34] skeletal model which has 22 joints, and we detail in the supplementary materials how we partition its joints into 6 parts. Similar to KIT-ML, we follow the division into train, validation, and test sets as specified in [14] and report our ParCo's performance on the test set.", + "bbox": [ + 212, + 526, + 787, + 752 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "Evaluation Metrics. Following prior text-to-motion work, we leverage pretrained text and motion feature extractors [14] to measure cross-modal alignment and semantic similarity between motions, rather than joint coordinate position. Based on the extracted features, we employ the following evaluation metrics: (i)", + "bbox": [ + 212, + 779, + 787, + 840 + ], + "page_idx": 8 + }, + { + "type": "header", + "text": "ParCo: Part-Coordinating Text-to-Motion Synthesis", + "bbox": [ + 383, + 114, + 730, + 128 + ], + "page_idx": 8 + }, + { + "type": "page_number", + "text": "9", + "bbox": [ + 774, + 116, + 784, + 126 + ], + "page_idx": 8 + }, + { + "type": "table", + "img_path": "images/cec3130bfd6a896ad8d638e15dc3c9e6bb6243d6123f4b2c2b07843933587db0.jpg", + "table_caption": [ + "Table 2: Comparisons to current state-of-the-art methods on KIT-ML test set." + ], + "table_footnote": [], + "table_body": "
MethodsR-Precision ↑FID ↓MM-Dist ↓Diversity → MModality ↑
Top-1Top-2Top-3
Real motion0.424±.0050.649±.0060.779±.0060.031±.0042.788±.01211.08±.097-
MDM§ [61]0.164±.0040.291±.0040.396±.0040.497±.0219.191±.02210.85±.1091.907±.214
MLD§ [8]0.390±.0080.609±.0080.734±.0070.404±.0273.204±.02710.80±.1172.192±.071
MotionDiffuse§ [67]0.417±.0040.621±.0040.739±.0041.954±.0622.958±.00511.10±.1430.730±.013
ReMoDiffuse§ [68]0.427±.0140.641±.0040.765±.0550.155±.0062.814±.01210.80±.1051.239±.028
ReMoDiffuse*0.382±.0050.586±.0070.706±.0060.589±.0223.324±.03010.31±.065-
Seq2Seq [48]0.103±.0030.178±.0050.241±.00624.86±.3487.960±.0316.744±.106-
Text2Gesture [6]0.156±.0040.255±.0040.338±.00512.12±.1836.946±.0299.334±.079-
Language2Pose [2]0.221±.0050.373±.0040.483±.0056.545±.0725.147±.0309.073±.100-
Hier [13]0.255±.0060.432±.0070.531±.0075.203±.1074.986±.0279.563±.0722.090±.083
TM2T [15]0.280±.0050.463±.0060.587±.0053.599±.1534.591±.0269.473±.1173.292±.081
TEMOS [46]0.353±.0060.561±.0070.687±.0053.717±.0513.417±.01910.84±.1000.532±.034
T2M [14]0.370±.0050.569±.0070.693±.0072.770±.1093.401±.00810.91±.1191.482±.065
AttT2M [71]0.413±.0060.632±.0060.751±.0060.870±.0393.039±.02110.96±.1232.281±.047
T2M-GPT [66]0.416±.0060.627±.0060.745±.0060.514±.0293.007±.02310.92±.1081.570±.039
Fg-T2M [63]0.418±.0050.626±.0040.745±.0040.571±.0473.114±.01510.93±.0831.019±.029
ParCo (Ours)0.430±.0040.649±.0070.772±.0060.453±.0272.820±.02810.95±.0941.245±.022
", + "bbox": [ + 218, + 171, + 787, + 416 + ], + "page_idx": 9 + }, + { + "type": "text", + "text": "R-Precision: Reflects the accurate semantic matching between text and motion. We calculate Top-1, Top-2, and Top-3 accuracy based on the Euclidean distance between a given motion sequence and 32 text descriptions (1 ground truth and 31 randomly selected non-matching). (ii) FID: We use FID [21] to quantify the distributional disparity between generated and real motions based on extracted motion features. It's crucial to highlight that FID does not assess the alignment between textual descriptions and generated motions. (iii) MM-Dist: Measures the Euclidean distance between feature vectors of text and motion, reflecting the semantic similarity. (iv) Diversity: Indicates the variance in generated motions. We randomly sample two equal-sized subsets from all motions, calculating the average Euclidean distance between the subsets. Closer diversity values between generated and real motions signify a better match. (v) MModality: Reflects the diversity of generated motions for a given text. We generate 10 pairs of motions for each text, compute feature vectors' distance between each pair, and take the average. Given that incorrectly generated motions lead to a high MModality value, it fails to reflect the alignment between motions and texts. In addition, following [61], we run each evaluation 20 times (except MModality for 5 times) and report the average with a $95\\%$ confidence interval.", + "bbox": [ + 212, + 455, + 787, + 728 + ], + "page_idx": 9 + }, + { + "type": "text", + "text": "Implementation Details. Our ParCo employs 6 small VQ-VAEs for discretizing part motions and 6 small transformers with Part Coordination modules for text-to-motion generation. For VQ-VAEs, all codebooks contain 512 codes, and all parts have 128 code dimensions, except for the Root part, which has a code dimension of 64. The encoder's downsampling rate is set to $r = 4$ . For transformers, each has 14 layers, and token dimension is 256. And we insert a Part Coordi", + "bbox": [ + 212, + 748, + 787, + 840 + ], + "page_idx": 9 + }, + { + "type": "page_number", + "text": "10", + "bbox": [ + 217, + 114, + 235, + 126 + ], + "page_idx": 9 + }, + { + "type": "header", + "text": "Q. Zou et al.", + "bbox": [ + 271, + 114, + 357, + 128 + ], + "page_idx": 9 + }, + { + "type": "table", + "img_path": "images/e4a37b2aec97ffbee3636a9b047d94e2a790a09e57455d4ed8687b236e2a4c9e.jpg", + "table_caption": [], + "table_footnote": [], + "table_body": "
"a man steps back, picks something up and put it to his head and then puts it back.""a person walks forward one foot in front of the other until he loses his balance and tilts far to the left then stumbles."
MDMX no steps backX pick upX put to headX no put it backX no walks forward one foot in front of the otherX dance instead of stumble
Motion DiffuseX no steps backX raise hand instead of pickX put to headX puts it backX no walks forward one foot in front of the otherX lose balance
T2M-GPTX no steps backX swing arm instead of pickX put to headX puts it backX tilts to the right instead of leftX no stumble
OursX steps backX picks upX put to headX puts it backX walks forward one foot in front of the otherX lose balance
X tilts to the right instead of leftX no stumble
X walks forward one foot in front of the otherX lose balance
X tilts far to the leftX stumble
", + "bbox": [ + 217, + 143, + 785, + 395 + ], + "page_idx": 10 + }, + { + "type": "text", + "text": "Fig.5: Qualitative comparison with existing methods. Green indicates the motion is consistent with the text description. Red indicates the text description lacks the corresponding motion or got the wrong motion.", + "bbox": [ + 212, + 406, + 787, + 448 + ], + "page_idx": 10 + }, + { + "type": "text", + "text": "nation Layer before all remaining layers except for the first transformer layer. ParCo Blocks in the same layer of the same transformer share parameter weights. We set the number of MLP layers in ParCo Block to 3. For training VQ-VAE, we use a learning rate of 2e-4 before 200K and 1e-5 after 200K, AdamW [35] optimizer with $beta_{1} = 0.9$ and $beta_{2} = 0.99$ , and batch size of 256. The commitment loss weight $\\beta$ is set to 1.0. For training transformer, we use a learning rate of 1e-4 before 150K and 5e-6 after 150K, AdamW optimizer with $beta_{1} = 0.5$ and $beta_{2} = 0.99$ , and batch size of 128.", + "bbox": [ + 212, + 477, + 787, + 599 + ], + "page_idx": 10 + }, + { + "type": "text", + "text": "4.2 Comparisons to State-of-the-art", + "text_level": 1, + "bbox": [ + 215, + 619, + 524, + 635 + ], + "page_idx": 10 + }, + { + "type": "text", + "text": "We compare our ParCo with other methods (including the methods using ground-truth motion length) on HumanML3D test set (Table. 1) and KIT-ML test set (Table. 2). Our method demonstrates superior performance compared to previous state-of-the-art methods on R-Precision and MM-Dist, and comparable results on FID, indicating ParCo's superiority. As for Top-1, Top-2, and Top-3 of R-Precision, our ParCo surpasses previous SOTA, ReMoDiffuse [68] (using GT motion length), with 0.005, 0.008, and 0.006 on HumanML3D, 0.003, 0.008, and 0.007 on KIT-ML. And for MM-Dist, our ParCo exceeds ReMoDiffuse's performance with 0.047 on HumanML3D and has comparable result on KIT-ML. In terms of FID, our result is on par with ReMoDiffuse on HumanML3D, and achieve the SOTA performance on HumanML3D and KIT-ML compared to the methods not relying GT information. Furthermore, our ParCo yields a lower MModality value compared to the previous state-of-the-art. This may be", + "bbox": [ + 212, + 643, + 787, + 840 + ], + "page_idx": 10 + }, + { + "type": "header", + "text": "ParCo: Part-Coordinating Text-to-Motion Synthesis", + "bbox": [ + 383, + 114, + 732, + 128 + ], + "page_idx": 10 + }, + { + "type": "page_number", + "text": "11", + "bbox": [ + 767, + 116, + 782, + 126 + ], + "page_idx": 10 + }, + { + "type": "text", + "text": "attributed to our ParCo generating motions that align more accurately with the text, thereby reducing the occurrence of irrelevant or incorrect motions. Qualitative results presented in Fig. 5 demonstrate that our generated motions are more realistic, coordinated, and aligned with textual descriptions for texts involving multiple body parts. Details such as \"steps back,\" \"tilts far to the left,\" and \"stumbles\" are accurately captured by our ParCo, while other methods either ignore or incorrectly generate these nuanced actions. Besides, as illustrated in Fig. 1 and Fig. 5, ParCo is endowed the part-level composition superiority across action-level spatial and temporal compositions respectively. More details are provided in the supplementary materials.", + "bbox": [ + 212, + 146, + 787, + 297 + ], + "page_idx": 11 + }, + { + "type": "text", + "text": "GT Leakage of Diffusion-based Methods Previous studies [8, 61, 67, 68] based on the Diffusion model utilize the ground truth motion length as an input for synthesizing motions during evaluation, contributing to their remarkable FID scores. However, this approach is impractical for real-world applications. For clarity, we replace the GT motion lengths with random lengths sampled uniformly from dataset's motion length range to evaluate ReMoDiff. On HumanML3D (KIT-ML), the ReMoDiff's Top-1, Top-2, and Top-3 R-precision decrease by 0.041 (0.053), 0.040 (0.068), and 0.033 (0.075) respectively, while the FID increases by 0.147 (0.426), justifying our speculation. As an auto-regressive approach, our ParCo demonstrates competitive R-Precision and FID without requiring pre-defined motion lengths Table 1. Given its lower computational and parameter consumption (Table. 4), the superiority of ParCo is convincing.", + "bbox": [ + 212, + 321, + 787, + 505 + ], + "page_idx": 11 + }, + { + "type": "text", + "text": "4.3 Analysis", + "text_level": 1, + "bbox": [ + 215, + 529, + 333, + 544 + ], + "page_idx": 11 + }, + { + "type": "text", + "text": "Performance on Text Inputs of Different Lengths. In order to investigate the synthetic performance given textual descriptions of different lengths, we sort the HumanML3D test set based on the length of textual descriptions, and split it into four subsets with approximately equal numbers of text-motion pairs: 0-25%, 25-50%, 50-75%, and 75-100%, from short to long. The details of these splits are available in the supplementary materials. We conducted evaluations of real motion, T2M-GPT [66], and our method on these four splits. As illustrated in Fig. 6, our method exhibits comprehensive improvements in all subsets when contrasted with the baseline [66]. It demonstrates that our design is beneficial for text inputs of different lengths without compromise.", + "bbox": [ + 212, + 556, + 787, + 709 + ], + "page_idx": 11 + }, + { + "type": "text", + "text": "Ablation Study. We investigate different discretizations for whole-body motion and conduct ablation of our Part Coordination module. We compare our 6-part partition and upper-lower body partition proposed by SCA [13]. It is noteworthy that, despite SCA dividing whole-body into upper and lower body for motion generation, its generations of upper body motion and lower body motion are entirely independent and lack coordination compared with our method. As demonstrated in Table. 3, the comparison between our ParCo (D) and the", + "bbox": [ + 212, + 733, + 787, + 842 + ], + "page_idx": 11 + }, + { + "type": "page_number", + "text": "12", + "bbox": [ + 217, + 114, + 235, + 126 + ], + "page_idx": 11 + }, + { + "type": "header", + "text": "Q. Zou et al.", + "bbox": [ + 271, + 114, + 357, + 128 + ], + "page_idx": 11 + }, + { + "type": "image", + "img_path": "images/6e7c5d6c4ea560f9da8b49f829c29c248037bf744019bf988effa215c1a7c755.jpg", + "image_caption": [ + "Fig.6: Comparison on 4 HumanML3D test subsets divided based on text length. $0-25\\%$ , $25-50\\%$ , $50-75\\%$ , and $75-100\\%$ respectively represent four subsets based on text length from short to long." + ], + "image_footnote": [], + "bbox": [ + 218, + 146, + 393, + 227 + ], + "page_idx": 12 + }, + { + "type": "image", + "img_path": "images/186b49529ed66f5fe5eac65f3731e5a63e04ddd574eadaf66a951d4499c2a3c8.jpg", + "image_caption": [ + "(b) Performance gap relative to Real" + ], + "image_footnote": [], + "bbox": [ + 395, + 147, + 542, + 220 + ], + "page_idx": 12 + }, + { + "type": "image", + "img_path": "images/233a93902faa6074809a7d02602e9afcb9401fde6871336ff6c95e8aa426397c.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 542, + 148, + 660, + 220 + ], + "page_idx": 12 + }, + { + "type": "image", + "img_path": "images/2e5abcbfb74c6907abf0c3a1c29831975e6eb7086aac0a7cf57ec2fe9b221dc8.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 660, + 148, + 784, + 220 + ], + "page_idx": 12 + }, + { + "type": "table", + "img_path": "images/0ac9672bef32bd99899cbf33a936f23d273c7e32d52c2bb982aff5b5076c846f.jpg", + "table_caption": [ + "Table 3: Ablations of body discretization and part coordination module. * denotes our ParCo." + ], + "table_footnote": [], + "table_body": "
Discretization Part Coord.Top-1 ↑FID ↓MM-Dist ↓
AUp&LowBody×0.444±.0030.497±.0153.367±.009
B✓0.491±.0030.172±.0073.071±.008
C6 Parts×0.375±.0033.652±.0304.012±.012
*D✓0.515±.0030.109±.0052.927±.008
", + "bbox": [ + 218, + 342, + 537, + 412 + ], + "page_idx": 12 + }, + { + "type": "table", + "img_path": "images/da710e85a3f9248282f97ebda54a4b421bf67977ab103bc069fcd3b68605b006.jpg", + "table_caption": [ + "Table 4: Computational complexity analysis." + ], + "table_footnote": [], + "table_body": "
MethodParam(M)FLOPs(G)InferTime(s)
ReMoDiff198.2481.00.091
T2M-GPT237.6292.30.544
ParCo168.4211.70.036
", + "bbox": [ + 558, + 340, + 785, + 411 + ], + "page_idx": 12 + }, + { + "type": "text", + "text": "\"a man picks something with his [part description], shakes it, then puts it back.\"", + "bbox": [ + 215, + 443, + 784, + 459 + ], + "page_idx": 12 + }, + { + "type": "image", + "img_path": "images/ef918190c3df28eacbb9360a9a6e3da3e40fca8d11754530646b6549053f25ce.jpg", + "image_caption": [ + "Fig. 7: Qualitative result of left-right exchange experiment on our ParCo." + ], + "image_footnote": [], + "bbox": [ + 218, + 460, + 493, + 521 + ], + "page_idx": 12 + }, + { + "type": "image", + "img_path": "images/b95f506c8cf2bb9c7a0b9499a7a80e52579b5c2a7941516f7f6efeca96b350d6.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 519, + 460, + 785, + 521 + ], + "page_idx": 12 + }, + { + "type": "text", + "text": "re-implemented SCA (A) indicates that our method's generated motions exhibit significantly higher performance in Top-1, MM-Dist, and FID than SCA. Furthermore, the contrasts between results (A) and (B), as well as results (C) and (D), underscore the necessity of our ParCo facilitating communication and coordination among different part motions during the generation process. In addition, the comparison between results (B) and (D) validates the effectiveness of our 6-part partitioning, which makes the model aware of the concept of parts and contributes to text-to-motion synthesis.", + "bbox": [ + 212, + 580, + 787, + 702 + ], + "page_idx": 12 + }, + { + "type": "text", + "text": "Computational complexity analysis. With lightweight architectures and parallel computing support, our ParCo shows superior efficiency from parameters, flops, and inference time, justifying the efficacy of our part discretization and coordination designs, Table. 4.", + "bbox": [ + 212, + 724, + 787, + 786 + ], + "page_idx": 12 + }, + { + "type": "text", + "text": "Precise part control. To compare the control of different methods over the movement of human parts, we conducted a left-right exchange experiment on", + "bbox": [ + 214, + 809, + 785, + 839 + ], + "page_idx": 12 + }, + { + "type": "header", + "text": "ParCo: Part-Coordinating Text-to-Motion Synthesis", + "bbox": [ + 383, + 114, + 732, + 128 + ], + "page_idx": 12 + }, + { + "type": "page_number", + "text": "13", + "bbox": [ + 767, + 114, + 785, + 126 + ], + "page_idx": 12 + }, + { + "type": "table", + "img_path": "images/621732f790ba00691afd685f7749ce6a989f42e43e9b7428e40e5ae5c8b901df.jpg", + "table_caption": [ + "Table 5: Evaluation of real motion data on train, val, and test set of HumanML3D." + ], + "table_footnote": [], + "table_body": "
SplitTop-1Top-2Top-3MM-DistDiversity
Train0.628±.0010.810±.0010.888±.0012.388±.0039.685±.083
Val0.513±.0040.703±.0040.800±.0032.911±.0109.575±.081
Test0.512±.0030.703±.0020.797±.0022.973±.0079.495±.079
", + "bbox": [ + 287, + 171, + 715, + 250 + ], + "page_idx": 13 + }, + { + "type": "text", + "text": "ten sentences. We examined several methods for generating actions and accomplishing the left-right swap task, the success rates are listed as follows: $70\\%$ (ParCo), $50\\%$ (T2M-GPT), $30\\%$ (ReMoDiff), $20\\%$ (MDM), $0\\%$ (MoDiff). Notably, our ParCo exhibited the highest accuracy, underscoring its proficiency in the awareness of human parts. The qualitative results are showcased in Fig. 7.", + "bbox": [ + 212, + 292, + 785, + 372 + ], + "page_idx": 13 + }, + { + "type": "text", + "text": "Discussion. As illustrated in Table 1 and Table 2, our performance surpasses that of real motion due to the evaluation protocol [14], which uses pretrained feature extractors trained only on the train split. These extractors are more effective at extracting features from motions similar to those in the train split. Table 5 shows that evaluations on real motions from the Train, Val, and Test sets reveal significantly higher Top-1/2/3 and MM-Dist metrics on the Train split. This indicates better extractor understanding of the Train split data. Our model, trained on the train split, generates motions resembling the train data, making feature extraction and matching easier, leading to superior performance on the test set. Due to the lack of a comprehensive metric for text-to-motion semantic alignment, motion generation fidelity, and diversity, we advocate for a more holistic evaluation within the community.", + "bbox": [ + 212, + 391, + 787, + 574 + ], + "page_idx": 13 + }, + { + "type": "text", + "text": "5 Conclusion", + "text_level": 1, + "bbox": [ + 215, + 595, + 359, + 612 + ], + "page_idx": 13 + }, + { + "type": "text", + "text": "In this study, we focus on enhancing the text-to-motion generation model's ability to comprehend part concepts and facilitate communication between different part motion generators, ultimately yielding the synthesis of coordinated and fined-grained motion. Specifically, we discretize whole-body motion into multiple-part motions to establish the prior concept of parts. Afterward, we employ multiple lightweight generators designed to synthesize different part motions and coordinate them through our part coordination module. Extensive experiments showcase that our method achieves higher consistency between generated motions and textual descriptions compared to previous SOTA methods. Furthermore, in-depth analytical results suggest that our approach excels in achieving more precise part control and has lower computational complexity. More encouragingly, our method exhibits adaptability to various part separation schemes and holds the potential for further refinement toward hierarchical part motion. We anticipate that this will have a far-reaching impact on the community.", + "bbox": [ + 212, + 628, + 787, + 840 + ], + "page_idx": 13 + }, + { + "type": "page_number", + "text": "14", + "bbox": [ + 217, + 114, + 235, + 126 + ], + "page_idx": 13 + }, + { + "type": "header", + "text": "Q. Zou et al.", + "bbox": [ + 271, + 114, + 359, + 128 + ], + "page_idx": 13 + }, + { + "type": "text", + "text": "Acknowledgements", + "text_level": 1, + "bbox": [ + 217, + 143, + 401, + 162 + ], + "page_idx": 14 + }, + { + "type": "text", + "text": "This work was supported by the National Key R&D Program of China under Grant 2018AAA0102801.", + "bbox": [ + 215, + 176, + 784, + 208 + ], + "page_idx": 14 + }, + { + "type": "text", + "text": "References", + "text_level": 1, + "bbox": [ + 217, + 233, + 321, + 250 + ], + "page_idx": 14 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "1. Ahn, H., Ha, T., Choi, Y., Yoo, H., Oh, S.: Text2action: Generative adversarial synthesis from language to action. In: ICRA (2018)", + "2. Ahuja, C., Morency, L.P.: Language2pose: Natural language grounded pose forecasting. In: 3DV (2019)", + "3. Antakli, A., Hermann, E., Zinnikus, I., Du, H., Fischer, K.: Intelligent distributed human motion simulation in human-robot collaboration environments. In: ACM IVA (2018)", + "4. Athanasiou, N., Petrovich, M., Black, M.J., Varol, G.: Sinc: Spatial composition of 3d human motions for simultaneous action generation. arXiv preprint arXiv:2304.10417 (2023)", + "5. Barsoum, E., Kender, J., Liu, Z.: Hp-gan: Probabilistic 3d human motion prediction via gan. In: CVPRW (2018)", + "6. Bhattacharya, U., Rewkowski, N., Banerjee, A., Guhan, P., Bera, A., Manocha, D.: Text2gestures: A transformer-based network for generating emotive body gestures for virtual agents. In: 2021 IEEE virtual reality and 3D user interfaces (VR) (2021)", + "7. Butepage, J., Black, M.J., Kragic, D., Kjellstrom, H.: Deep representation learning for human motion prediction and classification. In: CVPR (2017)", + "8. Chen, X., Jiang, B., Liu, W., Huang, Z., Fu, B., Chen, T., Yu, G.: Executing your commands via motion diffusion in latent space. In: CVPR (2023)", + "9. Chen, X., Su, Z., Yang, L., Cheng, P., Xu, L., Fu, B., Yu, G.: Learning variational motion prior for video-based motion capture. arXiv preprint arXiv:2210.15134 (2022)", + "0. Djuric, N., Radosavljevic, V., Cui, H., Nguyen, T., Chou, F.C., Lin, T.H., Singh, N., Schneider, J.: Uncertainty-aware short-term motion prediction of traffic actors for autonomous driving. In: WACV (2020)", + "1. Duan, Y., Shi, T., Zou, Z., Lin, Y., Qian, Z., Zhang, B., Yuan, Y.: Single-shot motion completion with transformer. arXiv preprint arXiv:2103.00776 (2021)", + "2. Fragkiadaki, K., Levine, S., Felsen, P., Malik, J.: Recurrent network models for human dynamics. In: ICCV (2015)", + "3. Ghosh, A., Cheema, N., Oguz, C., Theobalt, C., Slusallek, P.: Synthesis of compositional animations from textual descriptions. In: ICCV (2021)", + "4. Guo, C., Zou, S., Zuo, X., Wang, S., Ji, W., Li, X., Cheng, L.: Generating diverse and natural 3d human motions from text. In: CVPR (2022)", + "5. Guo, C., Zuo, X., Wang, S., Cheng, L.: Tm2t: Stochastic and tokenized modeling for the reciprocal generation of 3d human motions and texts. In: ECCV (2022)", + "6. Guo, C., Zuo, X., Wang, S., Zou, S., Sun, Q., Deng, A., Gong, M., Cheng, L.: Action2motion: Conditioned generation of 3d human motions. In: ACM MM (2020)", + "7. Harvey, F.G., Pal, C.: Recurrent transition networks for character locomotion. In: SIGGRAPH (2018)", + "8. Harvey, F.G., Yurick, M., Nowrouzezahrai, D., Pal, C.: Robust motion in betweening. ACM TOG (2020)" + ], + "bbox": [ + 225, + 266, + 784, + 840 + ], + "page_idx": 14 + }, + { + "type": "header", + "text": "ParCo: Part-Coordinating Text-to-Motion Synthesis", + "bbox": [ + 383, + 114, + 730, + 128 + ], + "page_idx": 14 + }, + { + "type": "page_number", + "text": "15", + "bbox": [ + 767, + 116, + 784, + 126 + ], + "page_idx": 14 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "19. Herbet, G., Duffau, H.: Revisiting the functional anatomy of the human brain: toward a meta-networking theory of cerebral functions. Physiological Reviews (2020)", + "20. Hernandez, A., Gall, J., Moreno-Noguer, F.: Human motion prediction via spatiotemporal inpainting. In: ICCV (2019)", + "21. Heusel, M., Ramsauer, H., Unterthiner, T., Nessler, B., Hochreiter, S.: Gans trained by a two time-scale update rule converge to a local nash equilibrium. NeurIPS (2017)", + "22. Ho, J., Jain, A., Abbeel, P.: Denoising diffusion probabilistic models. NeurIPS (2020)", + "23. Kappel, M., Golyanik, V., Elgharib, M., Henningson, J.O., Seidel, H.P., Castillo, S., Theobalt, C., Magnor, M.: High-fidelity neural human motion transfer from monocular video. In: CVPR (2021)", + "24. Kaufmann, M., Aksan, E., Song, J., Pece, F., Ziegler, R., Hilliges, O.: Convolutional autoencoders for human motion infilling. In: 3DV (2020)", + "25. Kim, J., Kim, J., Choi, S.: Flame: Free-form language-based motion synthesis & editing. In: AAAI (2023)", + "26. Koppula, H., Saxena, A.: Learning spatio-temporal structure from rgb-d videos for human activity detection and anticipation. In: ICML (2013)", + "27. Koppula, H.S., Saxena, A.: Anticipating human activities using object affordances for reactive robotic response. IEEE TPAMI (2015)", + "28. Kullback, S.: Information theory and statistics (1997)", + "29. Lab, C.G.: Cmu graphics lab motion capture database (2016)", + "30. Lee, H.Y., Yang, X., Liu, M.Y., Wang, T.C., Lu, Y.D., Yang, M.H., Kautz, J.: Dancing to music. NeurIPS (2019)", + "31. Li, B., Zhao, Y., Zhelun, S., Sheng, L.: Danceformer: Music conditioned 3d dance generation with parametric motion transformer. In: AAAI (2022)", + "32. Li, R., Yang, S., Ross, D.A., Kanazawa, A.: Ai choreographer: Music conditioned 3d dance generation with aist++. In: ICCV (2021)", + "33. Liu, Z., Lin, Y., Cao, Y., Hu, H., Wei, Y., Zhang, Z., Lin, S., Guo, B.: Swin transformer: Hierarchical vision transformer using shifted windows. In: ICCV (2021)", + "34. Loper, M., Mahmood, N., Romero, J., Pons-Moll, G., Black, M.J.: Spl: A skinned multi-person linear model. In: Seminal Graphics Papers: Pushing the Boundaries, Volume 2 (2023)", + "35. Loshchilov, I., Hutter, F.: Decoupled weight decay regularization. arXiv preprint arXiv:1711.05101 (2017)", + "36. Mahmood, N., Ghorbani, N., Troje, N.F., Pons-Moll, G., Black, M.J.: Amass: Archive of motion capture as surface shapes. In: ICCV (2019)", + "37. Majoe, D., Widmer, L., Gutknecht, J.: Enhanced motion interaction for multimedia applications. In: Proceedings of the 7th International Conference on Advances in Mobile Computing and Multimedia (2009)", + "38. Mao, W., Liu, M., Salzmann, M.: History repeats itself: Human motion prediction via motion attention. In: ECCV (2020)", + "39. Mao, W., Liu, M., Salzmann, M., Li, H.: Learning trajectory dependencies for human motion prediction. In: ICCV (2019)", + "40. Martinez, J., Black, M.J., Romero, J.: On human motion prediction using recurrent neural networks. In: CVPR (2017)", + "41. Nichol, A., Dhariwal, P., Ramesh, A., Shyam, P., Mishkin, P., McGrew, B., Sutskever, I., Chen, M.: Glide: Towards photorealistic image generation and editing with text-guided diffusion models. arXiv preprint arXiv:2112.10741 (2021)", + "42. Parmar, N., Vaswani, A., Uszkoreit, J., Kaiser, L., Shazeer, N., Ku, A., Tran, D.: Image transformer. In: ICML (2018)" + ], + "bbox": [ + 215, + 147, + 784, + 839 + ], + "page_idx": 15 + }, + { + "type": "page_number", + "text": "16", + "bbox": [ + 217, + 114, + 235, + 126 + ], + "page_idx": 15 + }, + { + "type": "header", + "text": "Q. Zou et al.", + "bbox": [ + 271, + 114, + 357, + 128 + ], + "page_idx": 15 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "43. Pavlakos, G., Choutas, V., Ghorbani, N., Bolkart, T., Osman, A.A., Tzionas, D., Black, M.J.: Expressive body capture: 3d hands, face, and body from a single image. In: CVPR (2019)", + "44. Pavllo, D., Grangier, D., Auli, M.: Quaternet: A quaternion-based recurrent model for human motion. arXiv preprint arXiv:1805.06485 (2018)", + "45. Petrovich, M., Black, M.J., Varol, G.: Action-conditioned 3d human motion synthesis with transformer vae. In: ICCV (2021)", + "46. Petrovich, M., Black, M.J., Varol, G.: Temos: Generating diverse human motions from textual descriptions. In: ECCV (2022)", + "47. Plappert, M., Mandery, C., Asfour, T.: The kit motion-language dataset. Big data (2016)", + "48. Plappert, M., Mandery, C., Asfour, T.: Learning a bidirectional mapping between human whole-body motion and natural language using deep recurrent neural networks. Robotics and Autonomous Systems (2018)", + "49. Raab, S., Leibovitch, I., Li, P., Aberman, K., Sorkine-Hornung, O., Cohen-Or, D.: Modi: Unconditional motion synthesis from diverse data. In: CVPR (2023)", + "50. Radford, A., Kim, J.W., Hallacy, C., Ramesh, A., Goh, G., Agarwal, S., Sastry, G., Askell, A., Mishkin, P., Clark, J., et al.: Learning transferable visual models from natural language supervision. In: ICML (2021)", + "51. Ramesh, A., Dhariwal, P., Nichol, A., Chu, C., Chen, M.: Hierarchical text-conditional image generation with clip latents. arXiv preprint arXiv:2204.06125 (2022)", + "52. Rempe, D., Birdal, T., Hertzmann, A., Yang, J., Sridhar, S., Guibas, L.J.: Humor: 3d human motion model for robust pose estimation. In: ICCV (2021)", + "53. Rombach, R., Blattmann, A., Lorenz, D., Esser, P., Ommer, B.: High-resolution image synthesis with latent diffusion models. In: CVPR (2022)", + "54. Sahara, C., Chan, W., Saxena, S., Li, L., Whang, J., Denton, E.L., Ghasemipour, K., Gontijo Lopes, R., Karagol Ayan, B., Salimans, T., et al.: Photorealistic text-to-image diffusion models with deep language understanding. NeurIPS (2022)", + "55. Thiebaut de Schotten, M., Forkel, S.J.: The emergent properties of the connected brain. Science (2022)", + "56. Shafir, Y., Tevet, G., Kapon, R., Bermano, A.H.: Human motion diffusion as a generative prior. arXiv preprint arXiv:2303.01418 (2023)", + "57. Siyao, L., Yu, W., Gu, T., Lin, C., Wang, Q., Qian, C., Loy, C.C., Liu, Z.: Bailando: 3d dance generation by actor-critic gpt with choreographic memory. In: CVPR (2022)", + "58. Tang, X., Wang, H., Hu, B., Gong, X., Yi, R., Kou, Q., Jin, X.: Real-time controllable motion transition for characters. ACM TOG (2022)", + "59. Terlemez, Ö., Ulbrich, S., Mandery, C., Do, M., Vahrenkamp, N., Asfour, T.: Master motor map (mmm)—framework and toolkit for capturing, representing, and reproducing human motion on humanoid robots. In: 2014 IEEE-RAS International Conference on Humanoid Robots (2014)", + "60. Tevet, G., Gordon, B., Hertz, A., Bermano, A.H., Cohen-Or, D.: Motionclip: Exposing human motion generation to clip space. In: ECCV (2022)", + "61. Tevet, G., Raab, S., Gordon, B., Shafir, Y., Cohen-Or, D., Bermano, A.H.: Human motion diffusion model. arXiv preprint arXiv:2209.14916 (2022)", + "62. Vaswani, A., Shazeer, N., Parmar, N., Uszkoreit, J., Jones, L., Gomez, A.N., Kaiser, L., Polosukhin, I.: Attention is all you need. NeurIPS (2017)", + "63. Wang, Y., Leng, Z., Li, F.W., Wu, S.C., Liang, X.: Fg-t2m: Fine-grained text-driven human motion generation via diffusion model. In: ICCV (2023)" + ], + "bbox": [ + 215, + 146, + 785, + 839 + ], + "page_idx": 16 + }, + { + "type": "header", + "text": "ParCo: Part-Coordinating Text-to-Motion Synthesis", + "bbox": [ + 383, + 114, + 730, + 128 + ], + "page_idx": 16 + }, + { + "type": "page_number", + "text": "17", + "bbox": [ + 767, + 114, + 785, + 126 + ], + "page_idx": 16 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "64. Yan, S., Li, Z., Xiong, Y., Yan, H., Lin, D.: Convolutional sequence generation for skeleton-based action synthesis. In: ICCV (2019)", + "65. Yeasin, M., Polat, E., Sharma, R.: A multiobject tracking framework for interactive multimedia applications. IEEE TMM (2004)", + "66. Zhang, J., Zhang, Y., Cun, X., Huang, S., Zhang, Y., Zhao, H., Lu, H., Shen, X.: T2m-gpt: Generating human motion from textual descriptions with discrete representations. arXiv preprint arXiv:2301.06052 (2023)", + "67. Zhang, M., Cai, Z., Pan, L., Hong, F., Guo, X., Yang, L., Liu, Z.: Motiondiffuse: Text-driven human motion generation with diffusion model. arXiv preprint arXiv:2208.15001 (2022)", + "68. Zhang, M., Guo, X., Pan, L., Cai, Z., Hong, F., Li, H., Yang, L., Liu, Z.: Remodif- fuse: Retrieval-augmented motion diffusion model. arXiv preprint arXiv:2304.01116 (2023)", + "69. Zhang, Y., Black, M.J., Tang, S.: Perpetual motion: Generating unbounded human motion. arXiv preprint arXiv:2007.13886 (2020)", + "70. Zhao, R., Su, H., Ji, Q.: Bayesian adversarial human motion synthesis. In: CVPR (2020)", + "71. Zhong, C., Hu, L., Zhang, Z., Xia, S.: Attt2m: Text-driven human motion generation with multi-perspective attention mechanism. In: ICCV (2023)" + ], + "bbox": [ + 215, + 146, + 784, + 410 + ], + "page_idx": 17 + }, + { + "type": "page_number", + "text": "18", + "bbox": [ + 217, + 114, + 235, + 126 + ], + "page_idx": 17 + }, + { + "type": "header", + "text": "Q. Zou et al.", + "bbox": [ + 271, + 114, + 357, + 128 + ], + "page_idx": 17 + } +] \ No newline at end of file diff --git a/2024/ParCo_ Part-Coordinating Text-to-Motion Synthesis/5d0214df-d42e-4bdf-a60a-2b013c23d49a_model.json b/2024/ParCo_ Part-Coordinating Text-to-Motion Synthesis/5d0214df-d42e-4bdf-a60a-2b013c23d49a_model.json new file mode 100644 index 0000000000000000000000000000000000000000..18a721b1652cf21df7ea7b111a19ab7d0bed1917 --- /dev/null +++ b/2024/ParCo_ Part-Coordinating Text-to-Motion Synthesis/5d0214df-d42e-4bdf-a60a-2b013c23d49a_model.json @@ -0,0 +1,2469 @@ +[ + [ + { + "type": "title", + "bbox": [ + 0.261, + 0.141, + 0.744, + 0.187 + ], + "angle": 0, + "content": "ParCo: Part-Coordinating Text-to-Motion Synthesis" + }, + { + "type": "text", + "bbox": [ + 0.233, + 0.212, + 0.773, + 0.245 + ], + "angle": 0, + "content": "Qiran Zou\\*, Shangyuan Yuan\\*1, Shian Du', Yu Wang?, Chang Liu?, Yi Xu?, Jie Chen?, and Xiangyang Ji!" + }, + { + "type": "text", + "bbox": [ + 0.318, + 0.255, + 0.686, + 0.283 + ], + "angle": 0, + "content": "\\(^{1}\\) Tsinghua University, China \n\\(^{2}\\) Peking University Shenzhen Graduate School, China" + }, + { + "type": "text", + "bbox": [ + 0.226, + 0.283, + 0.776, + 0.339 + ], + "angle": 0, + "content": "\\(^{3}\\) Dalian University of Technology, China qiranzou@gmail.com, {yuansy21, dsa23}@mails.tsinghua.edu.cn, 2201212856@stu.pku.edu.cn, liuchang2022@tsinghua.edu.cn, yxu@dlut.edu.cn, chenj@pcl.ac.cn, xyji@tsinghua.edu.cn" + }, + { + "type": "text", + "bbox": [ + 0.261, + 0.378, + 0.744, + 0.655 + ], + "angle": 0, + "content": "Abstract. We study a challenging task: text-to-motion synthesis, aiming to generate motions that align with textual descriptions and exhibit coordinated movements. Currently, the part-based methods introduce part partition into the motion synthesis process to achieve finer-grained generation. However, these methods encounter challenges such as the lack of coordination between different part motions and difficulties for networks to understand part concepts. Moreover, introducing finer-grained part concepts poses computational complexity challenges. In this paper, we propose Part-Coordinating Text-to-Motion Synthesis (ParCo), endowed with enhanced capabilities for understanding part motions and communication among different part motion generators, ensuring a coordinated and fined-grained motion synthesis. Specifically, we discretize whole-body motion into multiple part motions to establish the prior concept of different parts. Afterward, we employ multiple lightweight generators designed to synthesize different part motions and coordinate them through our part coordination module. Our approach demonstrates superior performance on common benchmarks with economic computations, including HumanML3D and KIT-ML, providing substantial evidence of its effectiveness. Code is available at: https://github.com/qrzou/ParCo." + }, + { + "type": "text", + "bbox": [ + 0.262, + 0.668, + 0.713, + 0.683 + ], + "angle": 0, + "content": "Keywords: Motion synthesis \\(\\cdot\\) Part coordination \\(\\cdot\\) Text-to-motion" + }, + { + "type": "title", + "bbox": [ + 0.216, + 0.709, + 0.378, + 0.725 + ], + "angle": 0, + "content": "1 Introduction" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.74, + 0.788, + 0.817 + ], + "angle": 0, + "content": "Text-to-motion synthesis aims to generate motion that aligns with textual descriptions and exhibits coordinated movements. It facilitates obtaining desired motion through textual descriptions which benefits numerous applications in industrial scenarios such as animation [23], AR/VR applications, video games [37, 65], autonomous driving [10], and robotics [3,26,27]." + }, + { + "type": "page_footnote", + "bbox": [ + 0.218, + 0.825, + 0.365, + 0.84 + ], + "angle": 0, + "content": "* Equal contribution." + } + ], + [ + { + "type": "page_number", + "bbox": [ + 0.218, + 0.116, + 0.23, + 0.127 + ], + "angle": 0, + "content": "2" + }, + { + "type": "header", + "bbox": [ + 0.272, + 0.115, + 0.361, + 0.129 + ], + "angle": 0, + "content": "Q. Zou et al." + }, + { + "type": "text", + "bbox": [ + 0.348, + 0.162, + 0.66, + 0.173 + ], + "angle": 0, + "content": "\"am is raising his left hand while kicking something with his right foot\"" + }, + { + "type": "image", + "bbox": [ + 0.229, + 0.174, + 0.777, + 0.245 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.229, + 0.248, + 0.777, + 0.323 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.214, + 0.338, + 0.785, + 0.366 + ], + "angle": 0, + "content": "Fig. 1: Our ParCo is capable of coordinating the motion of various body parts to produce realistic and accurate motion." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.372, + 0.785, + 0.508 + ], + "angle": 0, + "content": "Recent advancements leveraging powerful generation capabilities of transformers [33,42,62] and diffusion models [22,41,51,53,54] have yielded impressive results in generating realistic and smooth motions. Despite this progress, existing methods often struggle to generate semantically matched and coherently coordinated motions, especially when faced with commands involving multiple coordinated body parts. We attribute this challenge to the intricate alignment problem between the modalities of text and motion, where a single text/motion can correspond to multiple possible motion/text, posing a challenge in learning the complex relationship between the two." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.508, + 0.787, + 0.765 + ], + "angle": 0, + "content": "In the realm of Text-to-Motion Generation, part-based methods aim to achieve a higher level of sophistication in motion generation. These approaches can be broadly classified into two categories: single generator with part-level motion embeddings (Fig. 2 (a)) and independent upper and lower body motion generators (Fig. 2 (b)). The former one constructs whole-body motion embeddings by amalgamating multiple part motion embeddings, explicitly introducing the concept of parts [57], or implicitly integrating part concepts with part-level attention [71]. However, utilizing a single generator to generate whole-body motion embeddings presents a challenge for the generator to understand the concept of parts. The latter one segregates whole-body motion into upper and lower body motions, deploying two independent generators to produce motions separately [13]. Although this design enhances the generator's comprehension of upper and lower body motions, the absence of information exchange between them leads to a lack of coordination in the resulting upper and lower body motions. To increase the granularity of part division, an intuitive solution is to employ additional generators for generating finer-grained part motions, such as the left leg. However, this approach introduces computational complexity challenges." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.765, + 0.787, + 0.84 + ], + "angle": 0, + "content": "Neuroscientific discoveries reveal that discrete regions within the human brain manifest unique functions, and these regions engage in communication to coordinate diverse activities [19, 55]. This design, where different low-level subsystems communicate to form a higher-level system, is prevalent in the natural and artificial world, offering a robust, perceptually strong structural design. In adher-" + } + ], + [ + { + "type": "header", + "bbox": [ + 0.384, + 0.115, + 0.733, + 0.13 + ], + "angle": 0, + "content": "ParCo: Part-Coordinating Text-to-Motion Synthesis" + }, + { + "type": "page_number", + "bbox": [ + 0.775, + 0.117, + 0.785, + 0.127 + ], + "angle": 0, + "content": "3" + }, + { + "type": "image", + "bbox": [ + 0.219, + 0.145, + 0.399, + 0.287 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.41, + 0.144, + 0.591, + 0.287 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.605, + 0.144, + 0.788, + 0.287 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.214, + 0.302, + 0.788, + 0.387 + ], + "angle": 0, + "content": "Fig. 2: Conceptual comparison of three part-based synthesis methods. (a): One generator synthesizes the whole-body embedding, which contains information about different parts internally. (b): Two separate generators synthesize the upper and lower body's motions independently, without information exchange between them. (c): Our ParCo employs multiple lightweight generators designed to synthesize different part motions, which are coordinated by the Part Coordination module." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.414, + 0.788, + 0.534 + ], + "angle": 0, + "content": "ence to these principles, we present Part-Coordinating Text-to-Motion Synthesis (ParCo). It comprises six small generators, which are tasked with various part motions, and accompanied by a Part Coordination module facilitating communication among the generators. This communication enables the generation of coordinated whole-body motions while comprehending distinct parts (Fig. 2 (c)). Thanks to the design of small generators, our method demonstrates a lower parameter count, reduced computational complexity, and shorter generation time in comparison to baseline and state-of-the-art methods." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.535, + 0.788, + 0.67 + ], + "angle": 0, + "content": "Specifically, our approach consists of two stages. In the first stage, we discretize whole-body motion into multiple part motions and quantize them using VQ-VAEs, providing prior knowledge of \"what is part\" for the next stage. In the second stage, we use multiple Part-Coordinated Transformers, which are capable of communicating with each other, to generate coordinated motions of different parts. These part motions are integrated into whole-body motion subsequently. Extensive experiments on HumanML3D [14] and KIT-ML [47] demonstrate that our method can generate realistic and coordinated motions that align with the semantic descriptions." + }, + { + "type": "text", + "bbox": [ + 0.24, + 0.671, + 0.542, + 0.685 + ], + "angle": 0, + "content": "We delineate our contributions as follows:" + }, + { + "type": "text", + "bbox": [ + 0.234, + 0.693, + 0.785, + 0.739 + ], + "angle": 0, + "content": "- We propose ParCo, which enables the generators to better understand finer-grained parts and coordinate the generated part motions, ultimately achieving fine-grained and coordinated motion synthesis." + }, + { + "type": "text", + "bbox": [ + 0.234, + 0.744, + 0.785, + 0.775 + ], + "angle": 0, + "content": "- Our approach is computationally efficient, although employing multiple part generators, and maintains excellent motion generation performance." + }, + { + "type": "text", + "bbox": [ + 0.234, + 0.78, + 0.787, + 0.84 + ], + "angle": 0, + "content": "- On the HumanML3D and KIT-ML datasets, our method significantly outperforms methods that do not rely on GT motion length as our ParCo, while demonstrating comparable performance to methods that do rely on GT motion length." + }, + { + "type": "list", + "bbox": [ + 0.234, + 0.693, + 0.787, + 0.84 + ], + "angle": 0, + "content": null + } + ], + [ + { + "type": "page_number", + "bbox": [ + 0.218, + 0.116, + 0.23, + 0.127 + ], + "angle": 0, + "content": "4" + }, + { + "type": "header", + "bbox": [ + 0.272, + 0.115, + 0.359, + 0.129 + ], + "angle": 0, + "content": "Q. Zou et al." + }, + { + "type": "title", + "bbox": [ + 0.218, + 0.144, + 0.388, + 0.161 + ], + "angle": 0, + "content": "2 Related Work" + }, + { + "type": "text", + "bbox": [ + 0.218, + 0.18, + 0.786, + 0.452 + ], + "angle": 0, + "content": "Human Motion Synthesis. Tasks in the domain of human motion synthesis fall within two distinct categories: unconditional motion generation and conditional motion generation. The categorization of these tasks is based on the input signals employed. Unconditional motion generation [49, 64, 69, 70], such as VPoser [43] and ACTOR [45], is a comprehensive task involving the modeling of the entire motion space, utilizing solely motion data for training and prediction. Human motion prediction, a highly dynamic field, seeks to forecast future movements based on observed motion. Another significant domain pertains to the generation of \"in-betweening\" motions, which fill the gaps between past and future poses [11, 17, 18, 24, 58]. Unconditional motion generation commonly utilizes models well-suited for processing sequential data, including recursive [7, 12, 40, 44], generative adversarial [5, 20], graph convolutional [39], and attention [38] approaches. This enables the efficient generation of diverse motions by concurrently processing spatial and temporal signals. Conditional motion generation involves various multimodal data types, including text [2, 14, 15, 25, 46, 61], occluded pose sequences [11, 18, 61], images [9, 52], and sound [30-32]. Due to the rapid advancements in NLP, text-driven human motion generation has sustained a notably active status." + }, + { + "type": "text", + "bbox": [ + 0.218, + 0.478, + 0.786, + 0.84 + ], + "angle": 0, + "content": "Text-driven Human Motion Generation. Text-to-motion aims to generate human motion based on input textual descriptions. In earlier studies, joint-latent models [2, 46] were employed, which integrate a text encoder and a motion encoder. Text2Action [1] employs a recursive model, generating motion from short texts. TEMOS [46] adopts a similar approach, using self-encoding structures for both text and motion constrained by KL divergence [28]. T2M-GPT [66], TM2T [15] replace recursive encoders with transformers and GRU structure, achieving promising results. MotionCLIP [60] directly introduces the powerful zero-shot-capable CLIP [50] text encoder, following the alignment of text and pose as in Language2Pose [2], and additionally renders images with CLIP image encoder for auxiliary supervision. Subsequently, solutions based on diffusion models [8,56,67] emerge. MDM [61], MotionDiffuse [67], ReMoDiffuse [68] introduce diffusion models based on probability mapping, enabling the generation of human motion sequences from textual descriptions. Although current research enables the convenient generation of human motions based on text, challenges persist, especially in handling complex textual descriptions involving different parts. These approaches often treat human motion as a whole, lacking an understanding of different body parts, exhibiting limited ability of aligning text and motion. Furthermore, other methods incorporate the concept of parts into the model to generate more granular motions [13,57,71]. However, these approaches encounter challenges, including the lack of coordination among different part motions and difficulties for networks to comprehend part concepts. In contrast, our ParCo demonstrates a superior understanding of parts and effectively coordinates part motions, with lower computational complexity." + } + ], + [ + { + "type": "header", + "bbox": [ + 0.384, + 0.115, + 0.733, + 0.13 + ], + "angle": 0, + "content": "ParCo: Part-Coordinating Text-to-Motion Synthesis" + }, + { + "type": "page_number", + "bbox": [ + 0.775, + 0.117, + 0.785, + 0.127 + ], + "angle": 0, + "content": "5" + }, + { + "type": "image", + "bbox": [ + 0.217, + 0.145, + 0.493, + 0.273 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.512, + 0.147, + 0.788, + 0.273 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.214, + 0.287, + 0.788, + 0.4 + ], + "angle": 0, + "content": "Fig. 3: Pipeline of ParCo. ParCo consists of two stages: (a) The whole-body motion is discretized into 6 part motions, and encoded into 6 quantized code index sequences by 6 VQ-VAEs (encoder and quantizer). This process provides a priori about the concept of part motions for the second stage. (b) We use the quantized index sequences and corresponding textual description to train 6 transformers for part motion generation. At the same time, these generators are coordinated by our Part Coordination module. The generated part motion codes are decoded by VQ-VAE (decoder) to reconstruct the 6 part motions, which will be integrated into the final whole-body motion." + }, + { + "type": "title", + "bbox": [ + 0.216, + 0.427, + 0.33, + 0.442 + ], + "angle": 0, + "content": "3 Method" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.46, + 0.788, + 0.581 + ], + "angle": 0, + "content": "Our method consists of two stages to generate motion with an understanding of part motions. In the first stage, we discretize the whole-body motion into multiple part motions to provide prior knowledge of \"what is part\" for the second stage. In the second stage, the objective is to enable the model to learn the concept of part and achieve mutual coordination among multiple part motion generators. With this design, our method can handle textual inputs involving different parts and generate human motion that aligns with the semantic descriptions in the text." + }, + { + "type": "title", + "bbox": [ + 0.215, + 0.604, + 0.544, + 0.618 + ], + "angle": 0, + "content": "3.1 Part-Aware Motion Discretization" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.629, + 0.787, + 0.704 + ], + "angle": 0, + "content": "In stage 1, our method partitions the whole-body motion into multiple part motions and independently encodes each of these part motions using a VQ-VAE. This ensures that each part motion possesses an independent representation space (encoding space), providing prior knowledge about the concept of part motion for the next stage." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.705, + 0.788, + 0.842 + ], + "angle": 0, + "content": "The 3D human body models (e.g. SMPL, MMM) typically use Kinematic Trees to model human skeleton as 5 chains (limbs and backbone) for motion modeling. We inherit this division and add the Root part to represent trajectories. Therefore, we divide the whole-body motion into six parts motions: R.Arm, L.Arm, R.Leg, L.Leg, Backbone, and Root. As illustrated in Fig. 3 (a), the first four represent the right/left arm and right/left leg, while the backbone denotes the spine and skull, and the root represents the pelvis joint's movement information. The commonly used human motion datasets, HumanML3D [14] and KIT-ML [47], utilize different body skeleton models, SMPL [34] and MMM [59]" + } + ], + [ + { + "type": "page_number", + "bbox": [ + 0.218, + 0.116, + 0.23, + 0.127 + ], + "angle": 0, + "content": "6" + }, + { + "type": "header", + "bbox": [ + 0.272, + 0.115, + 0.36, + 0.129 + ], + "angle": 0, + "content": "Q. Zou et al." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.147, + 0.784, + 0.177 + ], + "angle": 0, + "content": "respectively. We provide a detailed explanation of how we perform our six-part partitioning for these two skeleton models in the supplementary material." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.178, + 0.785, + 0.238 + ], + "angle": 0, + "content": "The aforementioned partitioning process can be formalized as: given a motion sequence \\( M = [m_1, \\dots, m_C] \\), where \\( C \\) is the number of frames, we separate it into part motions \\( \\{P^i = [p_1^i, \\dots, p_C^i]\\} \\), \\( i \\in [1, \\dots, S] \\), where \\( S \\) is the number of parts and \\( m_* = [p_*^1, \\dots, p_*^S] \\)." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.239, + 0.784, + 0.313 + ], + "angle": 0, + "content": "After separating the whole-body motion into part motions, we further discretize the part motions into code sequences using VQ-VAE. We utilize this discretized representation in the next stage's generation process, as it exhibits better generalization capabilities and contributes to the improvement of training and inference efficiency." + }, + { + "type": "text", + "bbox": [ + 0.215, + 0.314, + 0.785, + 0.395 + ], + "angle": 0, + "content": "Firstly, we use \\(Encoder^i\\) to get the \\(i\\)-th part motion's encoding that \\(E^i = Encoder^i(P^i) = [e_1^i, \\dots, e_l^i, \\dots, e_L^i]\\), where \\(L = \\frac{C}{r}\\) and \\(r\\) is the downsampling rate of encoder. Then, we discretize the \\(E^i\\) into \\(Q^i = [v_{k_1^i}^i, \\dots, v_{k_l^i}^i, \\dots, v_{k_L^i}^i]\\) according to a learnable codebook \\(V^i = \\{v_j^i\\}, j = 1, \\dots, J\\), where \\(J\\) is the number of codes in the codebook. The index \\(k_l\\) is obtained by finding the most similar code:" + }, + { + "type": "equation", + "bbox": [ + 0.407, + 0.406, + 0.785, + 0.434 + ], + "angle": 0, + "content": "\\[\nk _ {l} ^ {i} = \\underset {j \\in \\{1, \\dots , J \\}} {\\arg \\min } \\left\\| e _ {l} ^ {i} - v _ {j} ^ {i} \\right\\|. \\tag {1}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.445, + 0.784, + 0.474 + ], + "angle": 0, + "content": "By this way, we discretize the motion \\( M \\) into \\( S \\) discretized part motion representations \\( \\{Q^i\\} \\), \\( i = 1, \\dots, S \\)." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.476, + 0.785, + 0.522 + ], + "angle": 0, + "content": "For training the VQ-VAE, We use a decoder to reconstruct the \\(i\\)-th part motion \\(\\hat{P}^i = \\text{Decoder}^i(Q^i)\\) with reconstruction loss \\(\\mathcal{L}_r^i = \\| \\hat{P}^i - P^i\\|\\). The optimization objective of the \\(i\\)-th part's VQ-VAE is:" + }, + { + "type": "equation", + "bbox": [ + 0.345, + 0.535, + 0.785, + 0.552 + ], + "angle": 0, + "content": "\\[\n\\mathcal {L} ^ {i} = \\mathcal {L} _ {r} ^ {i} + \\| s g (E ^ {i}) - Q ^ {i} \\| + \\beta \\| E ^ {i} - s g (Q ^ {i}) \\|, \\tag {2}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.563, + 0.785, + 0.653 + ], + "angle": 0, + "content": "where \\( sg \\) represents the stop-gradient operation. The first term is the reconstruction loss function, ensuring that the VQ-VAE can reconstruct the original part motion from the encoding. The second term is the codebook loss function. And the third term is the commitment loss, aiming to make the representation \\( e_l^i \\) output by the encoder as close as possible to the code \\( v_{k_l}^i \\) contained in the codebook. The weight of this loss is controlled by the hyperparameter \\( \\beta \\)." + }, + { + "type": "title", + "bbox": [ + 0.215, + 0.678, + 0.519, + 0.692 + ], + "angle": 0, + "content": "3.2 Text-Driven Part Coordination" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.705, + 0.785, + 0.84 + ], + "angle": 0, + "content": "In this stage, we employ transformers as generators to achieve text-to-motion generation. Diverging from the approach of using a large transformer for the entire whole-body motion, we utilize multiple small transformers to generate each part motion code sequence obtained from the previous stage. This allows small transformers to be aware of the conception of \"part motion\" constructed before. However, relying solely on these separate part motion generators can lead to an inability to collectively generate whole-body motions, due to a lack of knowledge about the motions of other parts. To address this, we introduce the Part Coordination module, facilitating communication among all part transformer" + } + ], + [ + { + "type": "header", + "bbox": [ + 0.384, + 0.115, + 0.733, + 0.131 + ], + "angle": 0, + "content": "ParCo: Part-Coordinating Text-to-Motion Synthesis" + }, + { + "type": "page_number", + "bbox": [ + 0.775, + 0.116, + 0.785, + 0.127 + ], + "angle": 0, + "content": "7" + }, + { + "type": "image", + "bbox": [ + 0.302, + 0.143, + 0.703, + 0.328 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.291, + 0.337, + 0.71, + 0.352 + ], + "angle": 0, + "content": "Fig. 4: The architecture of our Part-Coordinated Transformer." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.391, + 0.784, + 0.437 + ], + "angle": 0, + "content": "generators to collaboratively generate whole-body motion. This capability enables our ParCo to handle textual inputs involving different parts, generating human motion that aligns with semantic descriptions in the text." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.441, + 0.785, + 0.532 + ], + "angle": 0, + "content": "As depicted in Fig. 3 (b), we employ six transformers to generate code index sequences of part motions based on the input text. At the same time, these six transformers collaborate through our Part Coordination module to coordinate with each other. Subsequently, these generated index sequences are decoded into the original representation of part motions by the decoder of the VQ-VAE. These part motions are integrated to form whole-body motion." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.537, + 0.785, + 0.627 + ], + "angle": 0, + "content": "Specifically, we model the whole text-to-motion generation process as estimating the distribution \\( p(M|t) \\) of motion \\( M \\) given the text \\( t \\). Since we have discretized the whole-body motion into part motions, we can model the entire motion distribution \\( p(M|t) \\) through the estimation of conditional distributions \\( p(K^i |t) \\) for each part, where \\( \\{K^i = [k_1^i,\\dots,k_L^i]\\} , i\\in [1,\\dots,S] \\) is the code indices obtained in the first stage." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.631, + 0.784, + 0.662 + ], + "angle": 0, + "content": "To model \\(i\\)-th part motion's distribution, we propose an autoregressive distribution, which allows parts to coordinate with each other," + }, + { + "type": "equation", + "bbox": [ + 0.342, + 0.682, + 0.785, + 0.745 + ], + "angle": 0, + "content": "\\[\n\\begin{array}{l} p \\left(K ^ {i} \\mid t\\right) = \\prod_ {h = 1} ^ {L} p \\left(k _ {h} ^ {i} \\mid k _ {1} ^ {i}, o _ {1} ^ {i}; \\dots ; k _ {h - 1} ^ {i}, o _ {h - 1} ^ {i}; t\\right), \\tag {3} \\\\ o _ {*} ^ {i} = \\left\\{k _ {*} ^ {j} \\right\\}, j \\neq i, j \\in [ 1,..., S ]. \\\\ \\end{array}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.765, + 0.785, + 0.841 + ], + "angle": 0, + "content": "When predicting token \\( k_h^i \\) for the \\( i \\)-th part, the prediction not only relies on all tokens predicted by itself from time 1 to \\( h - 1 \\) but is also conditioned by predictions from all other parts during the same time span. Also, after the \\( i \\)-th part transformer predicts the token \\( K_h^i \\), the prediction will be used by all part generators to predict the next token." + } + ], + [ + { + "type": "page_number", + "bbox": [ + 0.218, + 0.116, + 0.23, + 0.127 + ], + "angle": 0, + "content": "8" + }, + { + "type": "header", + "bbox": [ + 0.272, + 0.115, + 0.36, + 0.129 + ], + "angle": 0, + "content": "Q. Zou et al." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.147, + 0.784, + 0.176 + ], + "angle": 0, + "content": "Finally, we learn the entire body motion by estimating the distributions of all part motions," + }, + { + "type": "equation", + "bbox": [ + 0.38, + 0.177, + 0.786, + 0.238 + ], + "angle": 0, + "content": "\\[\n\\begin{array}{l} \\mathcal {L} = \\mathbb {E} _ {M, t \\sim p (M, t)} [ - \\log p (M | t) ] \\\\ = \\mathbb {E} _ {M, t \\sim p (M, t)} \\left[ - \\sum_ {i = 1} ^ {S} \\log p \\left(K ^ {i} | t\\right) \\right]. \\tag {4} \\\\ \\end{array}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.242, + 0.785, + 0.349 + ], + "angle": 0, + "content": "We further propose Part-Coordinated Transformer, a transformer capable of coordinating with other transformers, to approximate \\( p(K^i |t) \\). As shown in Fig. 4, we insert a Part Coordination Layer before each transformer layer (except for the first transformer layer). For each token \\( x^i \\) output by the previous transformer layer, it passes through our ParCo Block. The ParCo Block coordinates with other part motion generators, fusing current token \\( x^i \\) with tokens from other part transformers," + }, + { + "type": "equation", + "bbox": [ + 0.379, + 0.361, + 0.786, + 0.399 + ], + "angle": 0, + "content": "\\[\n\\begin{array}{l} x _ {\\text {c o o r d}} ^ {i} = L N \\left(x ^ {i} + M L P ^ {i} (y)\\right), \\tag {5} \\\\ y = \\left\\{x ^ {j} \\right\\}, j \\neq i, j \\in [ 1, \\dots , S ], \\\\ \\end{array}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.409, + 0.784, + 0.454 + ], + "angle": 0, + "content": "where \\( LN \\) denotes the LayerNorm operation, and \\( y \\) represents the tokens from other transformers' layers. The fused token \\( x_{coord}^{i} \\) is then input into the subsequent transformer layer." + }, + { + "type": "title", + "bbox": [ + 0.216, + 0.477, + 0.35, + 0.491 + ], + "angle": 0, + "content": "3.3 Discussion" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.502, + 0.785, + 0.607 + ], + "angle": 0, + "content": "Text-to-motion synthesis intrinsically consists of two stages, i.e., body part representation and generation, to fulfill actions described by texts. To precisely ground each word to the corresponding body part, Balando [57] partitions the body into upper and lower parts for motion quantization while reconstructing through a shared decoder. SCA [13] steps further to equip independent generators for flexible correspondence. However, coarse-grained sub-body level modeling yields sub-optimal results." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.608, + 0.785, + 0.729 + ], + "angle": 0, + "content": "AttT2M [71] introduces a global-local attention mechanism to learn hierarchical body-part semantics for accurate motion synthesis. SINC [4] adopts a simple additive composition of part motion for GPT-guided synthetic training data creation. In contrast, we advocate explicitly discretizing body parts as individual action atoms and synthesis motion with decentralized generators and a centralized part coordinating module. With a series of computation-economic designs, we keep the relative independence yet close coordination relationships of body parts and report superior results." + }, + { + "type": "title", + "bbox": [ + 0.216, + 0.753, + 0.367, + 0.77 + ], + "angle": 0, + "content": "4 Experiment" + }, + { + "type": "title", + "bbox": [ + 0.216, + 0.785, + 0.331, + 0.8 + ], + "angle": 0, + "content": "4.1 Settings" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.81, + 0.784, + 0.84 + ], + "angle": 0, + "content": "Datasets. We utilized two widely used text-to-motion datasets, KIT-ML and HumanML3D, for training and testing our method. Comparative evaluations" + } + ], + [ + { + "type": "header", + "bbox": [ + 0.384, + 0.115, + 0.732, + 0.13 + ], + "angle": 0, + "content": "ParCo: Part-Coordinating Text-to-Motion Synthesis" + }, + { + "type": "page_number", + "bbox": [ + 0.775, + 0.117, + 0.785, + 0.127 + ], + "angle": 0, + "content": "9" + }, + { + "type": "table_caption", + "bbox": [ + 0.214, + 0.145, + 0.788, + 0.229 + ], + "angle": 0, + "content": "Table 1: Comparisons to current state-of-the-art methods on HumanML3D test set. “↑” denotes that higher is better. “↓” denotes that lower is better. “→” denotes that results are better if the metric is closer to the real motion. Bold and underlined indicate the best and second-best results, respectively. § reports results using ground-truth motion length. The results of ReMoDiffuse* are obtained from official checkpoints and employ uniform random sampling of motion lengths as input." + }, + { + "type": "table", + "bbox": [ + 0.218, + 0.24, + 0.789, + 0.486 + ], + "angle": 0, + "content": "
MethodsR-Precision ↑FID ↓MM-Dist ↓Diversity → MModality ↑
Top-1Top-2Top-3
Real motion0.511±.0030.703±.0030.797±.0020.002±.0002.974±.0089.503±.065-
MDM§ [61]0.320±.0050.498±.0040.611±.0070.544±.0445.566±.0279.559±.0862.799±.072
MLD§ [8]0.481±.0030.673±.0030.772±.0020.473±.0133.196±.0109.724±.0822.413±.079
MotionDiffuse§ [67]0.491±.0010.681±.0010.782±.0010.630±.0013.113±.0019.410±.0491.553±.042
ReMoDiffuse§ [68]0.510±.0050.698±.0060.795±.0040.103±.0042.974±.0169.018±.0751.795±.043
ReMoDiffuse*0.450±.0030.638±.0020.743±.0030.281±.0103.271±.0089.236±.085-
Text2Gesture [6]0.165±.0010.267±.0020.345±.0027.664±.0306.030±.0086.409±.071-
Seq2Seq [48]0.180±.0020.300±.0020.396±.00211.75±.0355.529±.0076.223±.061-
Language2Pose [2]0.246±.0010.387±.0020.486±.00211.02±.0465.296±.0087.676±.058-
Hier [13]0.301±.0020.425±.0020.552±.0046.532±.0245.012±.0188.332±.042-
TEMOS [46]0.424±.0020.612±.0020.722±.0023.734±.0283.703±.0088.973±.0710.368±.018
TM2T [15]0.424±.0030.618±.0030.729±.0021.501±.0173.467±.0118.589±.0762.424±.093
T2M [14]0.457±.0020.639±.0030.740±.0031.067±.0023.340±.0089.188±.0022.090±.083
T2M-GPT [66]0.492±.0030.679±.0020.775±.0020.141±.0053.121±.0099.722±.0821.831±.048
Fg-T2M [63]0.492±.0020.683±.0030.783±.0020.243±.0193.109±.0079.278±.0721.614±.049
AttT2M [71]0.499±.0030.690±.0020.786±.0020.112±.0063.038±.0079.700±.0902.452±.051
ParCo (Ours)0.515±.0030.706±.0030.801±.0020.109±.0052.927±.0089.576±.0881.382±.060
" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.527, + 0.789, + 0.753 + ], + "angle": 0, + "content": "were conducted on these datasets against other existing methods. Processed from KIT [47] and CMU [29] datasets, KIT-ML Dataset [47] comprises 3,911 sequences of human body motions with 6,278 text annotations. Each motion is annotated with 1 to 4 text descriptions, averaging approximately 8 words each. KIT-ML employs the MMM [59] skeletal model which has 21 joints, and we detail in the supplementary materials how we partition its joints into 6 parts. We follow the train, validation, and test set divisions as outlined in [14] and report our ParCo's performance on the test set. HumanML3D [14] Dataset is the largest dataset with 14,616 3D human body motion sequences and 44,970 corresponding textual descriptions, sourced from AMASS [36] and HumanAct12 [16]. Each motion includes a minimum of 3 text descriptions, with an average length of 12 words. HumanML3D adopts the SMPL [34] skeletal model which has 22 joints, and we detail in the supplementary materials how we partition its joints into 6 parts. Similar to KIT-ML, we follow the division into train, validation, and test sets as specified in [14] and report our ParCo's performance on the test set." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.78, + 0.788, + 0.842 + ], + "angle": 0, + "content": "Evaluation Metrics. Following prior text-to-motion work, we leverage pretrained text and motion feature extractors [14] to measure cross-modal alignment and semantic similarity between motions, rather than joint coordinate position. Based on the extracted features, we employ the following evaluation metrics: (i)" + } + ], + [ + { + "type": "page_number", + "bbox": [ + 0.218, + 0.116, + 0.236, + 0.127 + ], + "angle": 0, + "content": "10" + }, + { + "type": "header", + "bbox": [ + 0.272, + 0.115, + 0.359, + 0.129 + ], + "angle": 0, + "content": "Q. Zou et al." + }, + { + "type": "table_caption", + "bbox": [ + 0.236, + 0.145, + 0.766, + 0.159 + ], + "angle": 0, + "content": "Table 2: Comparisons to current state-of-the-art methods on KIT-ML test set." + }, + { + "type": "table", + "bbox": [ + 0.22, + 0.172, + 0.788, + 0.417 + ], + "angle": 0, + "content": "
MethodsR-Precision ↑FID ↓MM-Dist ↓Diversity → MModality ↑
Top-1Top-2Top-3
Real motion0.424±.0050.649±.0060.779±.0060.031±.0042.788±.01211.08±.097-
MDM§ [61]0.164±.0040.291±.0040.396±.0040.497±.0219.191±.02210.85±.1091.907±.214
MLD§ [8]0.390±.0080.609±.0080.734±.0070.404±.0273.204±.02710.80±.1172.192±.071
MotionDiffuse§ [67]0.417±.0040.621±.0040.739±.0041.954±.0622.958±.00511.10±.1430.730±.013
ReMoDiffuse§ [68]0.427±.0140.641±.0040.765±.0550.155±.0062.814±.01210.80±.1051.239±.028
ReMoDiffuse*0.382±.0050.586±.0070.706±.0060.589±.0223.324±.03010.31±.065-
Seq2Seq [48]0.103±.0030.178±.0050.241±.00624.86±.3487.960±.0316.744±.106-
Text2Gesture [6]0.156±.0040.255±.0040.338±.00512.12±.1836.946±.0299.334±.079-
Language2Pose [2]0.221±.0050.373±.0040.483±.0056.545±.0725.147±.0309.073±.100-
Hier [13]0.255±.0060.432±.0070.531±.0075.203±.1074.986±.0279.563±.0722.090±.083
TM2T [15]0.280±.0050.463±.0060.587±.0053.599±.1534.591±.0269.473±.1173.292±.081
TEMOS [46]0.353±.0060.561±.0070.687±.0053.717±.0513.417±.01910.84±.1000.532±.034
T2M [14]0.370±.0050.569±.0070.693±.0072.770±.1093.401±.00810.91±.1191.482±.065
AttT2M [71]0.413±.0060.632±.0060.751±.0060.870±.0393.039±.02110.96±.1232.281±.047
T2M-GPT [66]0.416±.0060.627±.0060.745±.0060.514±.0293.007±.02310.92±.1081.570±.039
Fg-T2M [63]0.418±.0050.626±.0040.745±.0040.571±.0473.114±.01510.93±.0831.019±.029
ParCo (Ours)0.430±.0040.649±.0070.772±.0060.453±.0272.820±.02810.95±.0941.245±.022
" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.456, + 0.788, + 0.729 + ], + "angle": 0, + "content": "R-Precision: Reflects the accurate semantic matching between text and motion. We calculate Top-1, Top-2, and Top-3 accuracy based on the Euclidean distance between a given motion sequence and 32 text descriptions (1 ground truth and 31 randomly selected non-matching). (ii) FID: We use FID [21] to quantify the distributional disparity between generated and real motions based on extracted motion features. It's crucial to highlight that FID does not assess the alignment between textual descriptions and generated motions. (iii) MM-Dist: Measures the Euclidean distance between feature vectors of text and motion, reflecting the semantic similarity. (iv) Diversity: Indicates the variance in generated motions. We randomly sample two equal-sized subsets from all motions, calculating the average Euclidean distance between the subsets. Closer diversity values between generated and real motions signify a better match. (v) MModality: Reflects the diversity of generated motions for a given text. We generate 10 pairs of motions for each text, compute feature vectors' distance between each pair, and take the average. Given that incorrectly generated motions lead to a high MModality value, it fails to reflect the alignment between motions and texts. In addition, following [61], we run each evaluation 20 times (except MModality for 5 times) and report the average with a \\(95\\%\\) confidence interval." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.749, + 0.788, + 0.841 + ], + "angle": 0, + "content": "Implementation Details. Our ParCo employs 6 small VQ-VAEs for discretizing part motions and 6 small transformers with Part Coordination modules for text-to-motion generation. For VQ-VAEs, all codebooks contain 512 codes, and all parts have 128 code dimensions, except for the Root part, which has a code dimension of 64. The encoder's downsampling rate is set to \\( r = 4 \\). For transformers, each has 14 layers, and token dimension is 256. And we insert a Part Coordi" + } + ], + [ + { + "type": "header", + "bbox": [ + 0.384, + 0.115, + 0.733, + 0.129 + ], + "angle": 0, + "content": "ParCo: Part-Coordinating Text-to-Motion Synthesis" + }, + { + "type": "page_number", + "bbox": [ + 0.769, + 0.117, + 0.784, + 0.127 + ], + "angle": 0, + "content": "11" + }, + { + "type": "table", + "bbox": [ + 0.218, + 0.144, + 0.787, + 0.396 + ], + "angle": 0, + "content": "
"a man steps back, picks something up and put it to his head and then puts it back.""a person walks forward one foot in front of the other until he loses his balance and tilts far to the left then stumbles."
MDMX no steps backX pick upX put to headX no put it backX no walks forward one foot in front of the otherX dance instead of stumble
Motion DiffuseX no steps backX raise hand instead of pickX put to headX puts it backX no walks forward one foot in front of the otherX lose balance
T2M-GPTX no steps backX swing arm instead of pickX put to headX puts it backX tilts to the right instead of leftX no stumble
OursX steps backX picks upX put to headX puts it backX walks forward one foot in front of the otherX lose balance
X tilts to the right instead of leftX no stumble
X walks forward one foot in front of the otherX lose balance
X tilts far to the leftX stumble
" + }, + { + "type": "image_caption", + "bbox": [ + 0.214, + 0.407, + 0.788, + 0.449 + ], + "angle": 0, + "content": "Fig.5: Qualitative comparison with existing methods. Green indicates the motion is consistent with the text description. Red indicates the text description lacks the corresponding motion or got the wrong motion." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.478, + 0.788, + 0.6 + ], + "angle": 0, + "content": "nation Layer before all remaining layers except for the first transformer layer. ParCo Blocks in the same layer of the same transformer share parameter weights. We set the number of MLP layers in ParCo Block to 3. For training VQ-VAE, we use a learning rate of 2e-4 before 200K and 1e-5 after 200K, AdamW [35] optimizer with \\(beta_{1} = 0.9\\) and \\(beta_{2} = 0.99\\), and batch size of 256. The commitment loss weight \\(\\beta\\) is set to 1.0. For training transformer, we use a learning rate of 1e-4 before 150K and 5e-6 after 150K, AdamW optimizer with \\(beta_{1} = 0.5\\) and \\(beta_{2} = 0.99\\), and batch size of 128." + }, + { + "type": "title", + "bbox": [ + 0.216, + 0.62, + 0.525, + 0.636 + ], + "angle": 0, + "content": "4.2 Comparisons to State-of-the-art" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.644, + 0.789, + 0.842 + ], + "angle": 0, + "content": "We compare our ParCo with other methods (including the methods using ground-truth motion length) on HumanML3D test set (Table. 1) and KIT-ML test set (Table. 2). Our method demonstrates superior performance compared to previous state-of-the-art methods on R-Precision and MM-Dist, and comparable results on FID, indicating ParCo's superiority. As for Top-1, Top-2, and Top-3 of R-Precision, our ParCo surpasses previous SOTA, ReMoDiffuse [68] (using GT motion length), with 0.005, 0.008, and 0.006 on HumanML3D, 0.003, 0.008, and 0.007 on KIT-ML. And for MM-Dist, our ParCo exceeds ReMoDiffuse's performance with 0.047 on HumanML3D and has comparable result on KIT-ML. In terms of FID, our result is on par with ReMoDiffuse on HumanML3D, and achieve the SOTA performance on HumanML3D and KIT-ML compared to the methods not relying GT information. Furthermore, our ParCo yields a lower MModality value compared to the previous state-of-the-art. This may be" + } + ], + [ + { + "type": "page_number", + "bbox": [ + 0.218, + 0.116, + 0.236, + 0.127 + ], + "angle": 0, + "content": "12" + }, + { + "type": "header", + "bbox": [ + 0.272, + 0.115, + 0.359, + 0.129 + ], + "angle": 0, + "content": "Q. Zou et al." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.147, + 0.788, + 0.299 + ], + "angle": 0, + "content": "attributed to our ParCo generating motions that align more accurately with the text, thereby reducing the occurrence of irrelevant or incorrect motions. Qualitative results presented in Fig. 5 demonstrate that our generated motions are more realistic, coordinated, and aligned with textual descriptions for texts involving multiple body parts. Details such as \"steps back,\" \"tilts far to the left,\" and \"stumbles\" are accurately captured by our ParCo, while other methods either ignore or incorrectly generate these nuanced actions. Besides, as illustrated in Fig. 1 and Fig. 5, ParCo is endowed the part-level composition superiority across action-level spatial and temporal compositions respectively. More details are provided in the supplementary materials." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.323, + 0.788, + 0.506 + ], + "angle": 0, + "content": "GT Leakage of Diffusion-based Methods Previous studies [8, 61, 67, 68] based on the Diffusion model utilize the ground truth motion length as an input for synthesizing motions during evaluation, contributing to their remarkable FID scores. However, this approach is impractical for real-world applications. For clarity, we replace the GT motion lengths with random lengths sampled uniformly from dataset's motion length range to evaluate ReMoDiff. On HumanML3D (KIT-ML), the ReMoDiff's Top-1, Top-2, and Top-3 R-precision decrease by 0.041 (0.053), 0.040 (0.068), and 0.033 (0.075) respectively, while the FID increases by 0.147 (0.426), justifying our speculation. As an auto-regressive approach, our ParCo demonstrates competitive R-Precision and FID without requiring pre-defined motion lengths Table 1. Given its lower computational and parameter consumption (Table. 4), the superiority of ParCo is convincing." + }, + { + "type": "title", + "bbox": [ + 0.216, + 0.53, + 0.334, + 0.545 + ], + "angle": 0, + "content": "4.3 Analysis" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.558, + 0.788, + 0.71 + ], + "angle": 0, + "content": "Performance on Text Inputs of Different Lengths. In order to investigate the synthetic performance given textual descriptions of different lengths, we sort the HumanML3D test set based on the length of textual descriptions, and split it into four subsets with approximately equal numbers of text-motion pairs: 0-25%, 25-50%, 50-75%, and 75-100%, from short to long. The details of these splits are available in the supplementary materials. We conducted evaluations of real motion, T2M-GPT [66], and our method on these four splits. As illustrated in Fig. 6, our method exhibits comprehensive improvements in all subsets when contrasted with the baseline [66]. It demonstrates that our design is beneficial for text inputs of different lengths without compromise." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.734, + 0.788, + 0.843 + ], + "angle": 0, + "content": "Ablation Study. We investigate different discretizations for whole-body motion and conduct ablation of our Part Coordination module. We compare our 6-part partition and upper-lower body partition proposed by SCA [13]. It is noteworthy that, despite SCA dividing whole-body into upper and lower body for motion generation, its generations of upper body motion and lower body motion are entirely independent and lack coordination compared with our method. As demonstrated in Table. 3, the comparison between our ParCo (D) and the" + } + ], + [ + { + "type": "header", + "bbox": [ + 0.384, + 0.115, + 0.733, + 0.13 + ], + "angle": 0, + "content": "ParCo: Part-Coordinating Text-to-Motion Synthesis" + }, + { + "type": "page_number", + "bbox": [ + 0.768, + 0.116, + 0.786, + 0.127 + ], + "angle": 0, + "content": "13" + }, + { + "type": "image", + "bbox": [ + 0.22, + 0.147, + 0.394, + 0.228 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.396, + 0.148, + 0.543, + 0.221 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.543, + 0.149, + 0.661, + 0.221 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.661, + 0.149, + 0.785, + 0.221 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.54, + 0.22, + 0.672, + 0.229 + ], + "angle": 0, + "content": "(b) Performance gap relative to Real" + }, + { + "type": "image_caption", + "bbox": [ + 0.215, + 0.243, + 0.784, + 0.285 + ], + "angle": 0, + "content": "Fig.6: Comparison on 4 HumanML3D test subsets divided based on text length. \\(0-25\\%\\), \\(25-50\\%\\), \\(50-75\\%\\), and \\(75-100\\%\\) respectively represent four subsets based on text length from short to long." + }, + { + "type": "table_caption", + "bbox": [ + 0.216, + 0.303, + 0.538, + 0.331 + ], + "angle": 0, + "content": "Table 3: Ablations of body discretization and part coordination module. * denotes our ParCo." + }, + { + "type": "table", + "bbox": [ + 0.22, + 0.343, + 0.538, + 0.414 + ], + "angle": 0, + "content": "
Discretization Part Coord.Top-1 ↑FID ↓MM-Dist ↓
AUp&LowBody×0.444±.0030.497±.0153.367±.009
B✓0.491±.0030.172±.0073.071±.008
C6 Parts×0.375±.0033.652±.0304.012±.012
*D✓0.515±.0030.109±.0052.927±.008
" + }, + { + "type": "table_caption", + "bbox": [ + 0.555, + 0.301, + 0.785, + 0.329 + ], + "angle": 0, + "content": "Table 4: Computational complexity analysis." + }, + { + "type": "table", + "bbox": [ + 0.559, + 0.342, + 0.787, + 0.412 + ], + "angle": 0, + "content": "
MethodParam(M)FLOPs(G)InferTime(s)
ReMoDiff198.2481.00.091
T2M-GPT237.6292.30.544
ParCo168.4211.70.036
" + }, + { + "type": "text", + "bbox": [ + 0.216, + 0.444, + 0.785, + 0.46 + ], + "angle": 0, + "content": "\"a man picks something with his [part description], shakes it, then puts it back.\"" + }, + { + "type": "image", + "bbox": [ + 0.22, + 0.462, + 0.495, + 0.522 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.52, + 0.462, + 0.787, + 0.522 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.254, + 0.536, + 0.747, + 0.551 + ], + "angle": 0, + "content": "Fig. 7: Qualitative result of left-right exchange experiment on our ParCo." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.581, + 0.788, + 0.703 + ], + "angle": 0, + "content": "re-implemented SCA (A) indicates that our method's generated motions exhibit significantly higher performance in Top-1, MM-Dist, and FID than SCA. Furthermore, the contrasts between results (A) and (B), as well as results (C) and (D), underscore the necessity of our ParCo facilitating communication and coordination among different part motions during the generation process. In addition, the comparison between results (B) and (D) validates the effectiveness of our 6-part partitioning, which makes the model aware of the concept of parts and contributes to text-to-motion synthesis." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.726, + 0.788, + 0.787 + ], + "angle": 0, + "content": "Computational complexity analysis. With lightweight architectures and parallel computing support, our ParCo shows superior efficiency from parameters, flops, and inference time, justifying the efficacy of our part discretization and coordination designs, Table. 4." + }, + { + "type": "text", + "bbox": [ + 0.215, + 0.81, + 0.787, + 0.84 + ], + "angle": 0, + "content": "Precise part control. To compare the control of different methods over the movement of human parts, we conducted a left-right exchange experiment on" + } + ], + [ + { + "type": "page_number", + "bbox": [ + 0.218, + 0.116, + 0.236, + 0.127 + ], + "angle": 0, + "content": "14" + }, + { + "type": "header", + "bbox": [ + 0.272, + 0.115, + 0.361, + 0.129 + ], + "angle": 0, + "content": "Q. Zou et al." + }, + { + "type": "table_caption", + "bbox": [ + 0.22, + 0.144, + 0.782, + 0.159 + ], + "angle": 0, + "content": "Table 5: Evaluation of real motion data on train, val, and test set of HumanML3D." + }, + { + "type": "table", + "bbox": [ + 0.288, + 0.172, + 0.716, + 0.251 + ], + "angle": 0, + "content": "
SplitTop-1Top-2Top-3MM-DistDiversity
Train0.628±.0010.810±.0010.888±.0012.388±.0039.685±.083
Val0.513±.0040.703±.0040.800±.0032.911±.0109.575±.081
Test0.512±.0030.703±.0020.797±.0022.973±.0079.495±.079
" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.294, + 0.787, + 0.373 + ], + "angle": 0, + "content": "ten sentences. We examined several methods for generating actions and accomplishing the left-right swap task, the success rates are listed as follows: \\(70\\%\\) (ParCo), \\(50\\%\\) (T2M-GPT), \\(30\\%\\) (ReMoDiff), \\(20\\%\\) (MDM), \\(0\\%\\) (MoDiff). Notably, our ParCo exhibited the highest accuracy, underscoring its proficiency in the awareness of human parts. The qualitative results are showcased in Fig. 7." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.392, + 0.788, + 0.575 + ], + "angle": 0, + "content": "Discussion. As illustrated in Table 1 and Table 2, our performance surpasses that of real motion due to the evaluation protocol [14], which uses pretrained feature extractors trained only on the train split. These extractors are more effective at extracting features from motions similar to those in the train split. Table 5 shows that evaluations on real motions from the Train, Val, and Test sets reveal significantly higher Top-1/2/3 and MM-Dist metrics on the Train split. This indicates better extractor understanding of the Train split data. Our model, trained on the train split, generates motions resembling the train data, making feature extraction and matching easier, leading to superior performance on the test set. Due to the lack of a comprehensive metric for text-to-motion semantic alignment, motion generation fidelity, and diversity, we advocate for a more holistic evaluation within the community." + }, + { + "type": "title", + "bbox": [ + 0.216, + 0.597, + 0.36, + 0.613 + ], + "angle": 0, + "content": "5 Conclusion" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.629, + 0.788, + 0.842 + ], + "angle": 0, + "content": "In this study, we focus on enhancing the text-to-motion generation model's ability to comprehend part concepts and facilitate communication between different part motion generators, ultimately yielding the synthesis of coordinated and fined-grained motion. Specifically, we discretize whole-body motion into multiple-part motions to establish the prior concept of parts. Afterward, we employ multiple lightweight generators designed to synthesize different part motions and coordinate them through our part coordination module. Extensive experiments showcase that our method achieves higher consistency between generated motions and textual descriptions compared to previous SOTA methods. Furthermore, in-depth analytical results suggest that our approach excels in achieving more precise part control and has lower computational complexity. More encouragingly, our method exhibits adaptability to various part separation schemes and holds the potential for further refinement toward hierarchical part motion. We anticipate that this will have a far-reaching impact on the community." + } + ], + [ + { + "type": "header", + "bbox": [ + 0.384, + 0.115, + 0.732, + 0.129 + ], + "angle": 0, + "content": "ParCo: Part-Coordinating Text-to-Motion Synthesis" + }, + { + "type": "page_number", + "bbox": [ + 0.769, + 0.117, + 0.785, + 0.127 + ], + "angle": 0, + "content": "15" + }, + { + "type": "title", + "bbox": [ + 0.218, + 0.145, + 0.403, + 0.163 + ], + "angle": 0, + "content": "Acknowledgements" + }, + { + "type": "text", + "bbox": [ + 0.217, + 0.178, + 0.785, + 0.209 + ], + "angle": 0, + "content": "This work was supported by the National Key R&D Program of China under Grant 2018AAA0102801." + }, + { + "type": "title", + "bbox": [ + 0.218, + 0.234, + 0.323, + 0.25 + ], + "angle": 0, + "content": "References" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.267, + 0.785, + 0.295 + ], + "angle": 0, + "content": "1. Ahn, H., Ha, T., Choi, Y., Yoo, H., Oh, S.: Text2action: Generative adversarial synthesis from language to action. In: ICRA (2018)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.296, + 0.785, + 0.323 + ], + "angle": 0, + "content": "2. Ahuja, C., Morency, L.P.: Language2pose: Natural language grounded pose forecasting. In: 3DV (2019)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.324, + 0.785, + 0.365 + ], + "angle": 0, + "content": "3. Antakli, A., Hermann, E., Zinnikus, I., Du, H., Fischer, K.: Intelligent distributed human motion simulation in human-robot collaboration environments. In: ACM IVA (2018)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.365, + 0.785, + 0.407 + ], + "angle": 0, + "content": "4. Athanasiou, N., Petrovich, M., Black, M.J., Varol, G.: Sinc: Spatial composition of 3d human motions for simultaneous action generation. arXiv preprint arXiv:2304.10417 (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.407, + 0.785, + 0.435 + ], + "angle": 0, + "content": "5. Barsoum, E., Kender, J., Liu, Z.: Hp-gan: Probabilistic 3d human motion prediction via gan. In: CVPRW (2018)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.435, + 0.785, + 0.478 + ], + "angle": 0, + "content": "6. Bhattacharya, U., Rewkowski, N., Banerjee, A., Guhan, P., Bera, A., Manocha, D.: Text2gestures: A transformer-based network for generating emotive body gestures for virtual agents. In: 2021 IEEE virtual reality and 3D user interfaces (VR) (2021)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.478, + 0.785, + 0.505 + ], + "angle": 0, + "content": "7. Butepage, J., Black, M.J., Kragic, D., Kjellstrom, H.: Deep representation learning for human motion prediction and classification. In: CVPR (2017)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.505, + 0.785, + 0.533 + ], + "angle": 0, + "content": "8. Chen, X., Jiang, B., Liu, W., Huang, Z., Fu, B., Chen, T., Yu, G.: Executing your commands via motion diffusion in latent space. In: CVPR (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.533, + 0.785, + 0.575 + ], + "angle": 0, + "content": "9. Chen, X., Su, Z., Yang, L., Cheng, P., Xu, L., Fu, B., Yu, G.: Learning variational motion prior for video-based motion capture. arXiv preprint arXiv:2210.15134 (2022)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.575, + 0.785, + 0.617 + ], + "angle": 0, + "content": "0. Djuric, N., Radosavljevic, V., Cui, H., Nguyen, T., Chou, F.C., Lin, T.H., Singh, N., Schneider, J.: Uncertainty-aware short-term motion prediction of traffic actors for autonomous driving. In: WACV (2020)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.617, + 0.785, + 0.644 + ], + "angle": 0, + "content": "1. Duan, Y., Shi, T., Zou, Z., Lin, Y., Qian, Z., Zhang, B., Yuan, Y.: Single-shot motion completion with transformer. arXiv preprint arXiv:2103.00776 (2021)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.644, + 0.785, + 0.673 + ], + "angle": 0, + "content": "2. Fragkiadaki, K., Levine, S., Felsen, P., Malik, J.: Recurrent network models for human dynamics. In: ICCV (2015)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.673, + 0.785, + 0.701 + ], + "angle": 0, + "content": "3. Ghosh, A., Cheema, N., Oguz, C., Theobalt, C., Slusallek, P.: Synthesis of compositional animations from textual descriptions. In: ICCV (2021)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.701, + 0.785, + 0.728 + ], + "angle": 0, + "content": "4. Guo, C., Zou, S., Zuo, X., Wang, S., Ji, W., Li, X., Cheng, L.: Generating diverse and natural 3d human motions from text. In: CVPR (2022)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.729, + 0.785, + 0.757 + ], + "angle": 0, + "content": "5. Guo, C., Zuo, X., Wang, S., Cheng, L.: Tm2t: Stochastic and tokenized modeling for the reciprocal generation of 3d human motions and texts. In: ECCV (2022)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.757, + 0.785, + 0.785 + ], + "angle": 0, + "content": "6. Guo, C., Zuo, X., Wang, S., Zou, S., Sun, Q., Deng, A., Gong, M., Cheng, L.: Action2motion: Conditioned generation of 3d human motions. In: ACM MM (2020)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.785, + 0.785, + 0.813 + ], + "angle": 0, + "content": "7. Harvey, F.G., Pal, C.: Recurrent transition networks for character locomotion. In: SIGGRAPH (2018)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.813, + 0.785, + 0.841 + ], + "angle": 0, + "content": "8. Harvey, F.G., Yurick, M., Nowrouzezahrai, D., Pal, C.: Robust motion in betweening. ACM TOG (2020)" + }, + { + "type": "list", + "bbox": [ + 0.226, + 0.267, + 0.785, + 0.841 + ], + "angle": 0, + "content": null + } + ], + [ + { + "type": "page_number", + "bbox": [ + 0.218, + 0.116, + 0.236, + 0.127 + ], + "angle": 0, + "content": "16" + }, + { + "type": "header", + "bbox": [ + 0.272, + 0.115, + 0.359, + 0.129 + ], + "angle": 0, + "content": "Q. Zou et al." + }, + { + "type": "ref_text", + "bbox": [ + 0.218, + 0.148, + 0.785, + 0.175 + ], + "angle": 0, + "content": "19. Herbet, G., Duffau, H.: Revisiting the functional anatomy of the human brain: toward a meta-networking theory of cerebral functions. Physiological Reviews (2020)" + }, + { + "type": "ref_text", + "bbox": [ + 0.218, + 0.176, + 0.785, + 0.202 + ], + "angle": 0, + "content": "20. Hernandez, A., Gall, J., Moreno-Noguer, F.: Human motion prediction via spatiotemporal inpainting. In: ICCV (2019)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.203, + 0.785, + 0.242 + ], + "angle": 0, + "content": "21. Heusel, M., Ramsauer, H., Unterthiner, T., Nessler, B., Hochreiter, S.: Gans trained by a two time-scale update rule converge to a local nash equilibrium. NeurIPS (2017)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.243, + 0.785, + 0.271 + ], + "angle": 0, + "content": "22. Ho, J., Jain, A., Abbeel, P.: Denoising diffusion probabilistic models. NeurIPS (2020)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.271, + 0.785, + 0.311 + ], + "angle": 0, + "content": "23. Kappel, M., Golyanik, V., Elgharib, M., Henningson, J.O., Seidel, H.P., Castillo, S., Theobalt, C., Magnor, M.: High-fidelity neural human motion transfer from monocular video. In: CVPR (2021)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.312, + 0.785, + 0.338 + ], + "angle": 0, + "content": "24. Kaufmann, M., Aksan, E., Song, J., Pece, F., Ziegler, R., Hilliges, O.: Convolutional autoencoders for human motion infilling. In: 3DV (2020)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.339, + 0.785, + 0.366 + ], + "angle": 0, + "content": "25. Kim, J., Kim, J., Choi, S.: Flame: Free-form language-based motion synthesis & editing. In: AAAI (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.366, + 0.785, + 0.393 + ], + "angle": 0, + "content": "26. Koppula, H., Saxena, A.: Learning spatio-temporal structure from rgb-d videos for human activity detection and anticipation. In: ICML (2013)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.394, + 0.785, + 0.42 + ], + "angle": 0, + "content": "27. Koppula, H.S., Saxena, A.: Anticipating human activities using object affordances for reactive robotic response. IEEE TPAMI (2015)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.421, + 0.6, + 0.433 + ], + "angle": 0, + "content": "28. Kullback, S.: Information theory and statistics (1997)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.434, + 0.652, + 0.446 + ], + "angle": 0, + "content": "29. Lab, C.G.: Cmu graphics lab motion capture database (2016)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.447, + 0.785, + 0.473 + ], + "angle": 0, + "content": "30. Lee, H.Y., Yang, X., Liu, M.Y., Wang, T.C., Lu, Y.D., Yang, M.H., Kautz, J.: Dancing to music. NeurIPS (2019)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.474, + 0.785, + 0.5 + ], + "angle": 0, + "content": "31. Li, B., Zhao, Y., Zhelun, S., Sheng, L.: Danceformer: Music conditioned 3d dance generation with parametric motion transformer. In: AAAI (2022)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.501, + 0.785, + 0.527 + ], + "angle": 0, + "content": "32. Li, R., Yang, S., Ross, D.A., Kanazawa, A.: Ai choreographer: Music conditioned 3d dance generation with aist++. In: ICCV (2021)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.528, + 0.785, + 0.554 + ], + "angle": 0, + "content": "33. Liu, Z., Lin, Y., Cao, Y., Hu, H., Wei, Y., Zhang, Z., Lin, S., Guo, B.: Swin transformer: Hierarchical vision transformer using shifted windows. In: ICCV (2021)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.555, + 0.785, + 0.595 + ], + "angle": 0, + "content": "34. Loper, M., Mahmood, N., Romero, J., Pons-Moll, G., Black, M.J.: Spl: A skinned multi-person linear model. In: Seminal Graphics Papers: Pushing the Boundaries, Volume 2 (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.596, + 0.785, + 0.622 + ], + "angle": 0, + "content": "35. Loshchilov, I., Hutter, F.: Decoupled weight decay regularization. arXiv preprint arXiv:1711.05101 (2017)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.623, + 0.785, + 0.65 + ], + "angle": 0, + "content": "36. Mahmood, N., Ghorbani, N., Troje, N.F., Pons-Moll, G., Black, M.J.: Amass: Archive of motion capture as surface shapes. In: ICCV (2019)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.65, + 0.785, + 0.69 + ], + "angle": 0, + "content": "37. Majoe, D., Widmer, L., Gutknecht, J.: Enhanced motion interaction for multimedia applications. In: Proceedings of the 7th International Conference on Advances in Mobile Computing and Multimedia (2009)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.691, + 0.785, + 0.718 + ], + "angle": 0, + "content": "38. Mao, W., Liu, M., Salzmann, M.: History repeats itself: Human motion prediction via motion attention. In: ECCV (2020)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.719, + 0.785, + 0.744 + ], + "angle": 0, + "content": "39. Mao, W., Liu, M., Salzmann, M., Li, H.: Learning trajectory dependencies for human motion prediction. In: ICCV (2019)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.745, + 0.785, + 0.772 + ], + "angle": 0, + "content": "40. Martinez, J., Black, M.J., Romero, J.: On human motion prediction using recurrent neural networks. In: CVPR (2017)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.773, + 0.785, + 0.813 + ], + "angle": 0, + "content": "41. Nichol, A., Dhariwal, P., Ramesh, A., Shyam, P., Mishkin, P., McGrew, B., Sutskever, I., Chen, M.: Glide: Towards photorealistic image generation and editing with text-guided diffusion models. arXiv preprint arXiv:2112.10741 (2021)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.814, + 0.785, + 0.84 + ], + "angle": 0, + "content": "42. Parmar, N., Vaswani, A., Uszkoreit, J., Kaiser, L., Shazeer, N., Ku, A., Tran, D.: Image transformer. In: ICML (2018)" + }, + { + "type": "list", + "bbox": [ + 0.217, + 0.148, + 0.785, + 0.84 + ], + "angle": 0, + "content": null + } + ], + [ + { + "type": "header", + "bbox": [ + 0.384, + 0.115, + 0.732, + 0.129 + ], + "angle": 0, + "content": "ParCo: Part-Coordinating Text-to-Motion Synthesis" + }, + { + "type": "page_number", + "bbox": [ + 0.768, + 0.116, + 0.786, + 0.127 + ], + "angle": 0, + "content": "17" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.147, + 0.786, + 0.189 + ], + "angle": 0, + "content": "43. Pavlakos, G., Choutas, V., Ghorbani, N., Bolkart, T., Osman, A.A., Tzionas, D., Black, M.J.: Expressive body capture: 3d hands, face, and body from a single image. In: CVPR (2019)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.19, + 0.787, + 0.218 + ], + "angle": 0, + "content": "44. Pavllo, D., Grangier, D., Auli, M.: Quaternet: A quaternion-based recurrent model for human motion. arXiv preprint arXiv:1805.06485 (2018)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.218, + 0.787, + 0.245 + ], + "angle": 0, + "content": "45. Petrovich, M., Black, M.J., Varol, G.: Action-conditioned 3d human motion synthesis with transformer vae. In: ICCV (2021)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.245, + 0.787, + 0.273 + ], + "angle": 0, + "content": "46. Petrovich, M., Black, M.J., Varol, G.: Temos: Generating diverse human motions from textual descriptions. In: ECCV (2022)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.273, + 0.787, + 0.3 + ], + "angle": 0, + "content": "47. Plappert, M., Mandery, C., Asfour, T.: The kit motion-language dataset. Big data (2016)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.3, + 0.787, + 0.342 + ], + "angle": 0, + "content": "48. Plappert, M., Mandery, C., Asfour, T.: Learning a bidirectional mapping between human whole-body motion and natural language using deep recurrent neural networks. Robotics and Autonomous Systems (2018)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.342, + 0.787, + 0.37 + ], + "angle": 0, + "content": "49. Raab, S., Leibovitch, I., Li, P., Aberman, K., Sorkine-Hornung, O., Cohen-Or, D.: Modi: Unconditional motion synthesis from diverse data. In: CVPR (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.37, + 0.787, + 0.412 + ], + "angle": 0, + "content": "50. Radford, A., Kim, J.W., Hallacy, C., Ramesh, A., Goh, G., Agarwal, S., Sastry, G., Askell, A., Mishkin, P., Clark, J., et al.: Learning transferable visual models from natural language supervision. In: ICML (2021)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.412, + 0.787, + 0.453 + ], + "angle": 0, + "content": "51. Ramesh, A., Dhariwal, P., Nichol, A., Chu, C., Chen, M.: Hierarchical text-conditional image generation with clip latents. arXiv preprint arXiv:2204.06125 (2022)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.453, + 0.787, + 0.481 + ], + "angle": 0, + "content": "52. Rempe, D., Birdal, T., Hertzmann, A., Yang, J., Sridhar, S., Guibas, L.J.: Humor: 3d human motion model for robust pose estimation. In: ICCV (2021)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.481, + 0.787, + 0.508 + ], + "angle": 0, + "content": "53. Rombach, R., Blattmann, A., Lorenz, D., Esser, P., Ommer, B.: High-resolution image synthesis with latent diffusion models. In: CVPR (2022)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.509, + 0.787, + 0.55 + ], + "angle": 0, + "content": "54. Sahara, C., Chan, W., Saxena, S., Li, L., Whang, J., Denton, E.L., Ghasemipour, K., Gontijo Lopes, R., Karagol Ayan, B., Salimans, T., et al.: Photorealistic text-to-image diffusion models with deep language understanding. NeurIPS (2022)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.55, + 0.787, + 0.577 + ], + "angle": 0, + "content": "55. Thiebaut de Schotten, M., Forkel, S.J.: The emergent properties of the connected brain. Science (2022)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.578, + 0.787, + 0.605 + ], + "angle": 0, + "content": "56. Shafir, Y., Tevet, G., Kapon, R., Bermano, A.H.: Human motion diffusion as a generative prior. arXiv preprint arXiv:2303.01418 (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.606, + 0.787, + 0.646 + ], + "angle": 0, + "content": "57. Siyao, L., Yu, W., Gu, T., Lin, C., Wang, Q., Qian, C., Loy, C.C., Liu, Z.: Bailando: 3d dance generation by actor-critic gpt with choreographic memory. In: CVPR (2022)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.646, + 0.787, + 0.674 + ], + "angle": 0, + "content": "58. Tang, X., Wang, H., Hu, B., Gong, X., Yi, R., Kou, Q., Jin, X.: Real-time controllable motion transition for characters. ACM TOG (2022)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.674, + 0.787, + 0.729 + ], + "angle": 0, + "content": "59. Terlemez, Ö., Ulbrich, S., Mandery, C., Do, M., Vahrenkamp, N., Asfour, T.: Master motor map (mmm)—framework and toolkit for capturing, representing, and reproducing human motion on humanoid robots. In: 2014 IEEE-RAS International Conference on Humanoid Robots (2014)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.729, + 0.787, + 0.758 + ], + "angle": 0, + "content": "60. Tevet, G., Gordon, B., Hertz, A., Bermano, A.H., Cohen-Or, D.: Motionclip: Exposing human motion generation to clip space. In: ECCV (2022)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.758, + 0.787, + 0.786 + ], + "angle": 0, + "content": "61. Tevet, G., Raab, S., Gordon, B., Shafir, Y., Cohen-Or, D., Bermano, A.H.: Human motion diffusion model. arXiv preprint arXiv:2209.14916 (2022)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.786, + 0.787, + 0.813 + ], + "angle": 0, + "content": "62. Vaswani, A., Shazeer, N., Parmar, N., Uszkoreit, J., Jones, L., Gomez, A.N., Kaiser, L., Polosukhin, I.: Attention is all you need. NeurIPS (2017)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.813, + 0.787, + 0.84 + ], + "angle": 0, + "content": "63. Wang, Y., Leng, Z., Li, F.W., Wu, S.C., Liang, X.: Fg-t2m: Fine-grained text-driven human motion generation via diffusion model. In: ICCV (2023)" + }, + { + "type": "list", + "bbox": [ + 0.217, + 0.147, + 0.787, + 0.84 + ], + "angle": 0, + "content": null + } + ], + [ + { + "type": "page_number", + "bbox": [ + 0.218, + 0.116, + 0.236, + 0.127 + ], + "angle": 0, + "content": "18" + }, + { + "type": "header", + "bbox": [ + 0.272, + 0.115, + 0.359, + 0.129 + ], + "angle": 0, + "content": "Q. Zou et al." + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.147, + 0.785, + 0.175 + ], + "angle": 0, + "content": "64. Yan, S., Li, Z., Xiong, Y., Yan, H., Lin, D.: Convolutional sequence generation for skeleton-based action synthesis. In: ICCV (2019)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.175, + 0.785, + 0.203 + ], + "angle": 0, + "content": "65. Yeasin, M., Polat, E., Sharma, R.: A multiobject tracking framework for interactive multimedia applications. IEEE TMM (2004)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.203, + 0.785, + 0.245 + ], + "angle": 0, + "content": "66. Zhang, J., Zhang, Y., Cun, X., Huang, S., Zhang, Y., Zhao, H., Lu, H., Shen, X.: T2m-gpt: Generating human motion from textual descriptions with discrete representations. arXiv preprint arXiv:2301.06052 (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.245, + 0.785, + 0.285 + ], + "angle": 0, + "content": "67. Zhang, M., Cai, Z., Pan, L., Hong, F., Guo, X., Yang, L., Liu, Z.: Motiondiffuse: Text-driven human motion generation with diffusion model. arXiv preprint arXiv:2208.15001 (2022)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.285, + 0.785, + 0.327 + ], + "angle": 0, + "content": "68. Zhang, M., Guo, X., Pan, L., Cai, Z., Hong, F., Li, H., Yang, L., Liu, Z.: Remodif- fuse: Retrieval-augmented motion diffusion model. arXiv preprint arXiv:2304.01116 (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.327, + 0.785, + 0.356 + ], + "angle": 0, + "content": "69. Zhang, Y., Black, M.J., Tang, S.: Perpetual motion: Generating unbounded human motion. arXiv preprint arXiv:2007.13886 (2020)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.356, + 0.785, + 0.383 + ], + "angle": 0, + "content": "70. Zhao, R., Su, H., Ji, Q.: Bayesian adversarial human motion synthesis. In: CVPR (2020)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.383, + 0.785, + 0.411 + ], + "angle": 0, + "content": "71. Zhong, C., Hu, L., Zhang, Z., Xia, S.: Attt2m: Text-driven human motion generation with multi-perspective attention mechanism. In: ICCV (2023)" + }, + { + "type": "list", + "bbox": [ + 0.217, + 0.147, + 0.785, + 0.411 + ], + "angle": 0, + "content": null + } + ] +] \ No newline at end of file diff --git a/2024/ParCo_ Part-Coordinating Text-to-Motion Synthesis/5d0214df-d42e-4bdf-a60a-2b013c23d49a_origin.pdf b/2024/ParCo_ Part-Coordinating Text-to-Motion Synthesis/5d0214df-d42e-4bdf-a60a-2b013c23d49a_origin.pdf new file mode 100644 index 0000000000000000000000000000000000000000..a84ea557708486b10fcde2086500116cd2d4258b --- /dev/null +++ b/2024/ParCo_ Part-Coordinating Text-to-Motion Synthesis/5d0214df-d42e-4bdf-a60a-2b013c23d49a_origin.pdf @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:d955530d6ba166bbd835ee0b370da8ee35a58c25e89b0fec450dc1b3bad01302 +size 7082156 diff --git a/2024/ParCo_ Part-Coordinating Text-to-Motion Synthesis/full.md b/2024/ParCo_ Part-Coordinating Text-to-Motion Synthesis/full.md new file mode 100644 index 0000000000000000000000000000000000000000..fdd4f9bcb3fae9594f907345315a0d7429a0f5ec --- /dev/null +++ b/2024/ParCo_ Part-Coordinating Text-to-Motion Synthesis/full.md @@ -0,0 +1,296 @@ +# ParCo: Part-Coordinating Text-to-Motion Synthesis + +Qiran Zou\*, Shangyuan Yuan\*1, Shian Du', Yu Wang?, Chang Liu?, Yi Xu?, Jie Chen?, and Xiangyang Ji! + +$^{1}$ Tsinghua University, China + $^{2}$ Peking University Shenzhen Graduate School, China + +$^{3}$ Dalian University of Technology, China qiranzou@gmail.com, {yuansy21, dsa23}@mails.tsinghua.edu.cn, 2201212856@stu.pku.edu.cn, liuchang2022@tsinghua.edu.cn, yxu@dlut.edu.cn, chenj@pcl.ac.cn, xyji@tsinghua.edu.cn + +Abstract. We study a challenging task: text-to-motion synthesis, aiming to generate motions that align with textual descriptions and exhibit coordinated movements. Currently, the part-based methods introduce part partition into the motion synthesis process to achieve finer-grained generation. However, these methods encounter challenges such as the lack of coordination between different part motions and difficulties for networks to understand part concepts. Moreover, introducing finer-grained part concepts poses computational complexity challenges. In this paper, we propose Part-Coordinating Text-to-Motion Synthesis (ParCo), endowed with enhanced capabilities for understanding part motions and communication among different part motion generators, ensuring a coordinated and fined-grained motion synthesis. Specifically, we discretize whole-body motion into multiple part motions to establish the prior concept of different parts. Afterward, we employ multiple lightweight generators designed to synthesize different part motions and coordinate them through our part coordination module. Our approach demonstrates superior performance on common benchmarks with economic computations, including HumanML3D and KIT-ML, providing substantial evidence of its effectiveness. Code is available at: https://github.com/qrzou/ParCo. + +Keywords: Motion synthesis $\cdot$ Part coordination $\cdot$ Text-to-motion + +# 1 Introduction + +Text-to-motion synthesis aims to generate motion that aligns with textual descriptions and exhibits coordinated movements. It facilitates obtaining desired motion through textual descriptions which benefits numerous applications in industrial scenarios such as animation [23], AR/VR applications, video games [37, 65], autonomous driving [10], and robotics [3,26,27]. + +"am is raising his left hand while kicking something with his right foot" + +![](images/e6ad0c4221961cca5a2376237fc46ca9d144a908a77881126e96cda2c07efcd1.jpg) + +![](images/90d386a8f15cfc164f3f8952b2cc9b611fe6d7ee49dde270eb4231c1d2c7f6fb.jpg) +Fig. 1: Our ParCo is capable of coordinating the motion of various body parts to produce realistic and accurate motion. + +Recent advancements leveraging powerful generation capabilities of transformers [33,42,62] and diffusion models [22,41,51,53,54] have yielded impressive results in generating realistic and smooth motions. Despite this progress, existing methods often struggle to generate semantically matched and coherently coordinated motions, especially when faced with commands involving multiple coordinated body parts. We attribute this challenge to the intricate alignment problem between the modalities of text and motion, where a single text/motion can correspond to multiple possible motion/text, posing a challenge in learning the complex relationship between the two. + +In the realm of Text-to-Motion Generation, part-based methods aim to achieve a higher level of sophistication in motion generation. These approaches can be broadly classified into two categories: single generator with part-level motion embeddings (Fig. 2 (a)) and independent upper and lower body motion generators (Fig. 2 (b)). The former one constructs whole-body motion embeddings by amalgamating multiple part motion embeddings, explicitly introducing the concept of parts [57], or implicitly integrating part concepts with part-level attention [71]. However, utilizing a single generator to generate whole-body motion embeddings presents a challenge for the generator to understand the concept of parts. The latter one segregates whole-body motion into upper and lower body motions, deploying two independent generators to produce motions separately [13]. Although this design enhances the generator's comprehension of upper and lower body motions, the absence of information exchange between them leads to a lack of coordination in the resulting upper and lower body motions. To increase the granularity of part division, an intuitive solution is to employ additional generators for generating finer-grained part motions, such as the left leg. However, this approach introduces computational complexity challenges. + +Neuroscientific discoveries reveal that discrete regions within the human brain manifest unique functions, and these regions engage in communication to coordinate diverse activities [19, 55]. This design, where different low-level subsystems communicate to form a higher-level system, is prevalent in the natural and artificial world, offering a robust, perceptually strong structural design. In adher- + +![](images/e59bce9596372b2b546709d57639cb47f437108783eda0ae41831c3d5d1439c5.jpg) +Fig. 2: Conceptual comparison of three part-based synthesis methods. (a): One generator synthesizes the whole-body embedding, which contains information about different parts internally. (b): Two separate generators synthesize the upper and lower body's motions independently, without information exchange between them. (c): Our ParCo employs multiple lightweight generators designed to synthesize different part motions, which are coordinated by the Part Coordination module. + +![](images/0a24dc73775f6e81ea18b8f13d7c54a90b4dbf191702fe192e17eb9a2bc2322c.jpg) + +![](images/7d86fb4db43edb9f2676c60a35762d2e95e2b9278e6c329aa7d3bb6cc605389b.jpg) + +ence to these principles, we present Part-Coordinating Text-to-Motion Synthesis (ParCo). It comprises six small generators, which are tasked with various part motions, and accompanied by a Part Coordination module facilitating communication among the generators. This communication enables the generation of coordinated whole-body motions while comprehending distinct parts (Fig. 2 (c)). Thanks to the design of small generators, our method demonstrates a lower parameter count, reduced computational complexity, and shorter generation time in comparison to baseline and state-of-the-art methods. + +Specifically, our approach consists of two stages. In the first stage, we discretize whole-body motion into multiple part motions and quantize them using VQ-VAEs, providing prior knowledge of "what is part" for the next stage. In the second stage, we use multiple Part-Coordinated Transformers, which are capable of communicating with each other, to generate coordinated motions of different parts. These part motions are integrated into whole-body motion subsequently. Extensive experiments on HumanML3D [14] and KIT-ML [47] demonstrate that our method can generate realistic and coordinated motions that align with the semantic descriptions. + +We delineate our contributions as follows: + +- We propose ParCo, which enables the generators to better understand finer-grained parts and coordinate the generated part motions, ultimately achieving fine-grained and coordinated motion synthesis. +- Our approach is computationally efficient, although employing multiple part generators, and maintains excellent motion generation performance. +- On the HumanML3D and KIT-ML datasets, our method significantly outperforms methods that do not rely on GT motion length as our ParCo, while demonstrating comparable performance to methods that do rely on GT motion length. + +# 2 Related Work + +Human Motion Synthesis. Tasks in the domain of human motion synthesis fall within two distinct categories: unconditional motion generation and conditional motion generation. The categorization of these tasks is based on the input signals employed. Unconditional motion generation [49, 64, 69, 70], such as VPoser [43] and ACTOR [45], is a comprehensive task involving the modeling of the entire motion space, utilizing solely motion data for training and prediction. Human motion prediction, a highly dynamic field, seeks to forecast future movements based on observed motion. Another significant domain pertains to the generation of "in-betweening" motions, which fill the gaps between past and future poses [11, 17, 18, 24, 58]. Unconditional motion generation commonly utilizes models well-suited for processing sequential data, including recursive [7, 12, 40, 44], generative adversarial [5, 20], graph convolutional [39], and attention [38] approaches. This enables the efficient generation of diverse motions by concurrently processing spatial and temporal signals. Conditional motion generation involves various multimodal data types, including text [2, 14, 15, 25, 46, 61], occluded pose sequences [11, 18, 61], images [9, 52], and sound [30-32]. Due to the rapid advancements in NLP, text-driven human motion generation has sustained a notably active status. + +Text-driven Human Motion Generation. Text-to-motion aims to generate human motion based on input textual descriptions. In earlier studies, joint-latent models [2, 46] were employed, which integrate a text encoder and a motion encoder. Text2Action [1] employs a recursive model, generating motion from short texts. TEMOS [46] adopts a similar approach, using self-encoding structures for both text and motion constrained by KL divergence [28]. T2M-GPT [66], TM2T [15] replace recursive encoders with transformers and GRU structure, achieving promising results. MotionCLIP [60] directly introduces the powerful zero-shot-capable CLIP [50] text encoder, following the alignment of text and pose as in Language2Pose [2], and additionally renders images with CLIP image encoder for auxiliary supervision. Subsequently, solutions based on diffusion models [8,56,67] emerge. MDM [61], MotionDiffuse [67], ReMoDiffuse [68] introduce diffusion models based on probability mapping, enabling the generation of human motion sequences from textual descriptions. Although current research enables the convenient generation of human motions based on text, challenges persist, especially in handling complex textual descriptions involving different parts. These approaches often treat human motion as a whole, lacking an understanding of different body parts, exhibiting limited ability of aligning text and motion. Furthermore, other methods incorporate the concept of parts into the model to generate more granular motions [13,57,71]. However, these approaches encounter challenges, including the lack of coordination among different part motions and difficulties for networks to comprehend part concepts. In contrast, our ParCo demonstrates a superior understanding of parts and effectively coordinates part motions, with lower computational complexity. + +![](images/cad4d6d83d8ed3ccd48cd341f4bb922398834c8076bb8c320c23c8199f195183.jpg) +Fig. 3: Pipeline of ParCo. ParCo consists of two stages: (a) The whole-body motion is discretized into 6 part motions, and encoded into 6 quantized code index sequences by 6 VQ-VAEs (encoder and quantizer). This process provides a priori about the concept of part motions for the second stage. (b) We use the quantized index sequences and corresponding textual description to train 6 transformers for part motion generation. At the same time, these generators are coordinated by our Part Coordination module. The generated part motion codes are decoded by VQ-VAE (decoder) to reconstruct the 6 part motions, which will be integrated into the final whole-body motion. + +![](images/c75e5c2b492baa1cddeaaf27d33b143669dc267341769b66affca63c7df4b321.jpg) + +# 3 Method + +Our method consists of two stages to generate motion with an understanding of part motions. In the first stage, we discretize the whole-body motion into multiple part motions to provide prior knowledge of "what is part" for the second stage. In the second stage, the objective is to enable the model to learn the concept of part and achieve mutual coordination among multiple part motion generators. With this design, our method can handle textual inputs involving different parts and generate human motion that aligns with the semantic descriptions in the text. + +# 3.1 Part-Aware Motion Discretization + +In stage 1, our method partitions the whole-body motion into multiple part motions and independently encodes each of these part motions using a VQ-VAE. This ensures that each part motion possesses an independent representation space (encoding space), providing prior knowledge about the concept of part motion for the next stage. + +The 3D human body models (e.g. SMPL, MMM) typically use Kinematic Trees to model human skeleton as 5 chains (limbs and backbone) for motion modeling. We inherit this division and add the Root part to represent trajectories. Therefore, we divide the whole-body motion into six parts motions: R.Arm, L.Arm, R.Leg, L.Leg, Backbone, and Root. As illustrated in Fig. 3 (a), the first four represent the right/left arm and right/left leg, while the backbone denotes the spine and skull, and the root represents the pelvis joint's movement information. The commonly used human motion datasets, HumanML3D [14] and KIT-ML [47], utilize different body skeleton models, SMPL [34] and MMM [59] + +respectively. We provide a detailed explanation of how we perform our six-part partitioning for these two skeleton models in the supplementary material. + +The aforementioned partitioning process can be formalized as: given a motion sequence $M = [m_1, \dots, m_C]$ , where $C$ is the number of frames, we separate it into part motions $\{P^i = [p_1^i, \dots, p_C^i]\}$ , $i \in [1, \dots, S]$ , where $S$ is the number of parts and $m_* = [p_*^1, \dots, p_*^S]$ . + +After separating the whole-body motion into part motions, we further discretize the part motions into code sequences using VQ-VAE. We utilize this discretized representation in the next stage's generation process, as it exhibits better generalization capabilities and contributes to the improvement of training and inference efficiency. + +Firstly, we use $Encoder^i$ to get the $i$ -th part motion's encoding that $E^i = Encoder^i(P^i) = [e_1^i, \dots, e_l^i, \dots, e_L^i]$ , where $L = \frac{C}{r}$ and $r$ is the downsampling rate of encoder. Then, we discretize the $E^i$ into $Q^i = [v_{k_1^i}^i, \dots, v_{k_l^i}^i, \dots, v_{k_L^i}^i]$ according to a learnable codebook $V^i = \{v_j^i\}, j = 1, \dots, J$ , where $J$ is the number of codes in the codebook. The index $k_l$ is obtained by finding the most similar code: + +$$ +k _ {l} ^ {i} = \underset {j \in \{1, \dots , J \}} {\arg \min } \left\| e _ {l} ^ {i} - v _ {j} ^ {i} \right\|. \tag {1} +$$ + +By this way, we discretize the motion $M$ into $S$ discretized part motion representations $\{Q^i\}$ , $i = 1, \dots, S$ . + +For training the VQ-VAE, We use a decoder to reconstruct the $i$ -th part motion $\hat{P}^i = \text{Decoder}^i(Q^i)$ with reconstruction loss $\mathcal{L}_r^i = \| \hat{P}^i - P^i\|$ . The optimization objective of the $i$ -th part's VQ-VAE is: + +$$ +\mathcal {L} ^ {i} = \mathcal {L} _ {r} ^ {i} + \| s g (E ^ {i}) - Q ^ {i} \| + \beta \| E ^ {i} - s g (Q ^ {i}) \|, \tag {2} +$$ + +where $sg$ represents the stop-gradient operation. The first term is the reconstruction loss function, ensuring that the VQ-VAE can reconstruct the original part motion from the encoding. The second term is the codebook loss function. And the third term is the commitment loss, aiming to make the representation $e_l^i$ output by the encoder as close as possible to the code $v_{k_l}^i$ contained in the codebook. The weight of this loss is controlled by the hyperparameter $\beta$ . + +# 3.2 Text-Driven Part Coordination + +In this stage, we employ transformers as generators to achieve text-to-motion generation. Diverging from the approach of using a large transformer for the entire whole-body motion, we utilize multiple small transformers to generate each part motion code sequence obtained from the previous stage. This allows small transformers to be aware of the conception of "part motion" constructed before. However, relying solely on these separate part motion generators can lead to an inability to collectively generate whole-body motions, due to a lack of knowledge about the motions of other parts. To address this, we introduce the Part Coordination module, facilitating communication among all part transformer + +![](images/4bb813629d5341c3a151e237d89aeecfe3de8849cf4f0a68dcf05eabf5c7a15a.jpg) +Fig. 4: The architecture of our Part-Coordinated Transformer. + +generators to collaboratively generate whole-body motion. This capability enables our ParCo to handle textual inputs involving different parts, generating human motion that aligns with semantic descriptions in the text. + +As depicted in Fig. 3 (b), we employ six transformers to generate code index sequences of part motions based on the input text. At the same time, these six transformers collaborate through our Part Coordination module to coordinate with each other. Subsequently, these generated index sequences are decoded into the original representation of part motions by the decoder of the VQ-VAE. These part motions are integrated to form whole-body motion. + +Specifically, we model the whole text-to-motion generation process as estimating the distribution $p(M|t)$ of motion $M$ given the text $t$ . Since we have discretized the whole-body motion into part motions, we can model the entire motion distribution $p(M|t)$ through the estimation of conditional distributions $p(K^i |t)$ for each part, where $\{K^i = [k_1^i,\dots,k_L^i]\} , i\in [1,\dots,S]$ is the code indices obtained in the first stage. + +To model $i$ -th part motion's distribution, we propose an autoregressive distribution, which allows parts to coordinate with each other, + +$$ +\begin{array}{l} p \left(K ^ {i} \mid t\right) = \prod_ {h = 1} ^ {L} p \left(k _ {h} ^ {i} \mid k _ {1} ^ {i}, o _ {1} ^ {i}; \dots ; k _ {h - 1} ^ {i}, o _ {h - 1} ^ {i}; t\right), \tag {3} \\ o _ {*} ^ {i} = \left\{k _ {*} ^ {j} \right\}, j \neq i, j \in [ 1,..., S ]. \\ \end{array} +$$ + +When predicting token $k_h^i$ for the $i$ -th part, the prediction not only relies on all tokens predicted by itself from time 1 to $h - 1$ but is also conditioned by predictions from all other parts during the same time span. Also, after the $i$ -th part transformer predicts the token $K_h^i$ , the prediction will be used by all part generators to predict the next token. + +Finally, we learn the entire body motion by estimating the distributions of all part motions, + +$$ +\begin{array}{l} \mathcal {L} = \mathbb {E} _ {M, t \sim p (M, t)} [ - \log p (M | t) ] \\ = \mathbb {E} _ {M, t \sim p (M, t)} \left[ - \sum_ {i = 1} ^ {S} \log p \left(K ^ {i} | t\right) \right]. \tag {4} \\ \end{array} +$$ + +We further propose Part-Coordinated Transformer, a transformer capable of coordinating with other transformers, to approximate $p(K^i |t)$ . As shown in Fig. 4, we insert a Part Coordination Layer before each transformer layer (except for the first transformer layer). For each token $x^i$ output by the previous transformer layer, it passes through our ParCo Block. The ParCo Block coordinates with other part motion generators, fusing current token $x^i$ with tokens from other part transformers, + +$$ +\begin{array}{l} x _ {\text {c o o r d}} ^ {i} = L N \left(x ^ {i} + M L P ^ {i} (y)\right), \tag {5} \\ y = \left\{x ^ {j} \right\}, j \neq i, j \in [ 1, \dots , S ], \\ \end{array} +$$ + +where $LN$ denotes the LayerNorm operation, and $y$ represents the tokens from other transformers' layers. The fused token $x_{coord}^{i}$ is then input into the subsequent transformer layer. + +# 3.3 Discussion + +Text-to-motion synthesis intrinsically consists of two stages, i.e., body part representation and generation, to fulfill actions described by texts. To precisely ground each word to the corresponding body part, Balando [57] partitions the body into upper and lower parts for motion quantization while reconstructing through a shared decoder. SCA [13] steps further to equip independent generators for flexible correspondence. However, coarse-grained sub-body level modeling yields sub-optimal results. + +AttT2M [71] introduces a global-local attention mechanism to learn hierarchical body-part semantics for accurate motion synthesis. SINC [4] adopts a simple additive composition of part motion for GPT-guided synthetic training data creation. In contrast, we advocate explicitly discretizing body parts as individual action atoms and synthesis motion with decentralized generators and a centralized part coordinating module. With a series of computation-economic designs, we keep the relative independence yet close coordination relationships of body parts and report superior results. + +# 4 Experiment + +# 4.1 Settings + +Datasets. We utilized two widely used text-to-motion datasets, KIT-ML and HumanML3D, for training and testing our method. Comparative evaluations + +Table 1: Comparisons to current state-of-the-art methods on HumanML3D test set. “↑” denotes that higher is better. “↓” denotes that lower is better. “→” denotes that results are better if the metric is closer to the real motion. Bold and underlined indicate the best and second-best results, respectively. § reports results using ground-truth motion length. The results of ReMoDiffuse* are obtained from official checkpoints and employ uniform random sampling of motion lengths as input. + +
MethodsR-Precision ↑FID ↓MM-Dist ↓Diversity → MModality ↑
Top-1Top-2Top-3
Real motion0.511±.0030.703±.0030.797±.0020.002±.0002.974±.0089.503±.065-
MDM§ [61]0.320±.0050.498±.0040.611±.0070.544±.0445.566±.0279.559±.0862.799±.072
MLD§ [8]0.481±.0030.673±.0030.772±.0020.473±.0133.196±.0109.724±.0822.413±.079
MotionDiffuse§ [67]0.491±.0010.681±.0010.782±.0010.630±.0013.113±.0019.410±.0491.553±.042
ReMoDiffuse§ [68]0.510±.0050.698±.0060.795±.0040.103±.0042.974±.0169.018±.0751.795±.043
ReMoDiffuse*0.450±.0030.638±.0020.743±.0030.281±.0103.271±.0089.236±.085-
Text2Gesture [6]0.165±.0010.267±.0020.345±.0027.664±.0306.030±.0086.409±.071-
Seq2Seq [48]0.180±.0020.300±.0020.396±.00211.75±.0355.529±.0076.223±.061-
Language2Pose [2]0.246±.0010.387±.0020.486±.00211.02±.0465.296±.0087.676±.058-
Hier [13]0.301±.0020.425±.0020.552±.0046.532±.0245.012±.0188.332±.042-
TEMOS [46]0.424±.0020.612±.0020.722±.0023.734±.0283.703±.0088.973±.0710.368±.018
TM2T [15]0.424±.0030.618±.0030.729±.0021.501±.0173.467±.0118.589±.0762.424±.093
T2M [14]0.457±.0020.639±.0030.740±.0031.067±.0023.340±.0089.188±.0022.090±.083
T2M-GPT [66]0.492±.0030.679±.0020.775±.0020.141±.0053.121±.0099.722±.0821.831±.048
Fg-T2M [63]0.492±.0020.683±.0030.783±.0020.243±.0193.109±.0079.278±.0721.614±.049
AttT2M [71]0.499±.0030.690±.0020.786±.0020.112±.0063.038±.0079.700±.0902.452±.051
ParCo (Ours)0.515±.0030.706±.0030.801±.0020.109±.0052.927±.0089.576±.0881.382±.060
+ +were conducted on these datasets against other existing methods. Processed from KIT [47] and CMU [29] datasets, KIT-ML Dataset [47] comprises 3,911 sequences of human body motions with 6,278 text annotations. Each motion is annotated with 1 to 4 text descriptions, averaging approximately 8 words each. KIT-ML employs the MMM [59] skeletal model which has 21 joints, and we detail in the supplementary materials how we partition its joints into 6 parts. We follow the train, validation, and test set divisions as outlined in [14] and report our ParCo's performance on the test set. HumanML3D [14] Dataset is the largest dataset with 14,616 3D human body motion sequences and 44,970 corresponding textual descriptions, sourced from AMASS [36] and HumanAct12 [16]. Each motion includes a minimum of 3 text descriptions, with an average length of 12 words. HumanML3D adopts the SMPL [34] skeletal model which has 22 joints, and we detail in the supplementary materials how we partition its joints into 6 parts. Similar to KIT-ML, we follow the division into train, validation, and test sets as specified in [14] and report our ParCo's performance on the test set. + +Evaluation Metrics. Following prior text-to-motion work, we leverage pretrained text and motion feature extractors [14] to measure cross-modal alignment and semantic similarity between motions, rather than joint coordinate position. Based on the extracted features, we employ the following evaluation metrics: (i) + +Table 2: Comparisons to current state-of-the-art methods on KIT-ML test set. + +
MethodsR-Precision ↑FID ↓MM-Dist ↓Diversity → MModality ↑
Top-1Top-2Top-3
Real motion0.424±.0050.649±.0060.779±.0060.031±.0042.788±.01211.08±.097-
MDM§ [61]0.164±.0040.291±.0040.396±.0040.497±.0219.191±.02210.85±.1091.907±.214
MLD§ [8]0.390±.0080.609±.0080.734±.0070.404±.0273.204±.02710.80±.1172.192±.071
MotionDiffuse§ [67]0.417±.0040.621±.0040.739±.0041.954±.0622.958±.00511.10±.1430.730±.013
ReMoDiffuse§ [68]0.427±.0140.641±.0040.765±.0550.155±.0062.814±.01210.80±.1051.239±.028
ReMoDiffuse*0.382±.0050.586±.0070.706±.0060.589±.0223.324±.03010.31±.065-
Seq2Seq [48]0.103±.0030.178±.0050.241±.00624.86±.3487.960±.0316.744±.106-
Text2Gesture [6]0.156±.0040.255±.0040.338±.00512.12±.1836.946±.0299.334±.079-
Language2Pose [2]0.221±.0050.373±.0040.483±.0056.545±.0725.147±.0309.073±.100-
Hier [13]0.255±.0060.432±.0070.531±.0075.203±.1074.986±.0279.563±.0722.090±.083
TM2T [15]0.280±.0050.463±.0060.587±.0053.599±.1534.591±.0269.473±.1173.292±.081
TEMOS [46]0.353±.0060.561±.0070.687±.0053.717±.0513.417±.01910.84±.1000.532±.034
T2M [14]0.370±.0050.569±.0070.693±.0072.770±.1093.401±.00810.91±.1191.482±.065
AttT2M [71]0.413±.0060.632±.0060.751±.0060.870±.0393.039±.02110.96±.1232.281±.047
T2M-GPT [66]0.416±.0060.627±.0060.745±.0060.514±.0293.007±.02310.92±.1081.570±.039
Fg-T2M [63]0.418±.0050.626±.0040.745±.0040.571±.0473.114±.01510.93±.0831.019±.029
ParCo (Ours)0.430±.0040.649±.0070.772±.0060.453±.0272.820±.02810.95±.0941.245±.022
+ +R-Precision: Reflects the accurate semantic matching between text and motion. We calculate Top-1, Top-2, and Top-3 accuracy based on the Euclidean distance between a given motion sequence and 32 text descriptions (1 ground truth and 31 randomly selected non-matching). (ii) FID: We use FID [21] to quantify the distributional disparity between generated and real motions based on extracted motion features. It's crucial to highlight that FID does not assess the alignment between textual descriptions and generated motions. (iii) MM-Dist: Measures the Euclidean distance between feature vectors of text and motion, reflecting the semantic similarity. (iv) Diversity: Indicates the variance in generated motions. We randomly sample two equal-sized subsets from all motions, calculating the average Euclidean distance between the subsets. Closer diversity values between generated and real motions signify a better match. (v) MModality: Reflects the diversity of generated motions for a given text. We generate 10 pairs of motions for each text, compute feature vectors' distance between each pair, and take the average. Given that incorrectly generated motions lead to a high MModality value, it fails to reflect the alignment between motions and texts. In addition, following [61], we run each evaluation 20 times (except MModality for 5 times) and report the average with a $95\%$ confidence interval. + +Implementation Details. Our ParCo employs 6 small VQ-VAEs for discretizing part motions and 6 small transformers with Part Coordination modules for text-to-motion generation. For VQ-VAEs, all codebooks contain 512 codes, and all parts have 128 code dimensions, except for the Root part, which has a code dimension of 64. The encoder's downsampling rate is set to $r = 4$ . For transformers, each has 14 layers, and token dimension is 256. And we insert a Part Coordi + +
"a man steps back, picks something up and put it to his head and then puts it back.""a person walks forward one foot in front of the other until he loses his balance and tilts far to the left then stumbles."
MDMX no steps backX pick upX put to headX no put it backX no walks forward one foot in front of the otherX dance instead of stumble
Motion DiffuseX no steps backX raise hand instead of pickX put to headX puts it backX no walks forward one foot in front of the otherX lose balance
T2M-GPTX no steps backX swing arm instead of pickX put to headX puts it backX tilts to the right instead of leftX no stumble
OursX steps backX picks upX put to headX puts it backX walks forward one foot in front of the otherX lose balance
X tilts to the right instead of leftX no stumble
X walks forward one foot in front of the otherX lose balance
X tilts far to the leftX stumble
+ +Fig.5: Qualitative comparison with existing methods. Green indicates the motion is consistent with the text description. Red indicates the text description lacks the corresponding motion or got the wrong motion. + +nation Layer before all remaining layers except for the first transformer layer. ParCo Blocks in the same layer of the same transformer share parameter weights. We set the number of MLP layers in ParCo Block to 3. For training VQ-VAE, we use a learning rate of 2e-4 before 200K and 1e-5 after 200K, AdamW [35] optimizer with $beta_{1} = 0.9$ and $beta_{2} = 0.99$ , and batch size of 256. The commitment loss weight $\beta$ is set to 1.0. For training transformer, we use a learning rate of 1e-4 before 150K and 5e-6 after 150K, AdamW optimizer with $beta_{1} = 0.5$ and $beta_{2} = 0.99$ , and batch size of 128. + +# 4.2 Comparisons to State-of-the-art + +We compare our ParCo with other methods (including the methods using ground-truth motion length) on HumanML3D test set (Table. 1) and KIT-ML test set (Table. 2). Our method demonstrates superior performance compared to previous state-of-the-art methods on R-Precision and MM-Dist, and comparable results on FID, indicating ParCo's superiority. As for Top-1, Top-2, and Top-3 of R-Precision, our ParCo surpasses previous SOTA, ReMoDiffuse [68] (using GT motion length), with 0.005, 0.008, and 0.006 on HumanML3D, 0.003, 0.008, and 0.007 on KIT-ML. And for MM-Dist, our ParCo exceeds ReMoDiffuse's performance with 0.047 on HumanML3D and has comparable result on KIT-ML. In terms of FID, our result is on par with ReMoDiffuse on HumanML3D, and achieve the SOTA performance on HumanML3D and KIT-ML compared to the methods not relying GT information. Furthermore, our ParCo yields a lower MModality value compared to the previous state-of-the-art. This may be + +attributed to our ParCo generating motions that align more accurately with the text, thereby reducing the occurrence of irrelevant or incorrect motions. Qualitative results presented in Fig. 5 demonstrate that our generated motions are more realistic, coordinated, and aligned with textual descriptions for texts involving multiple body parts. Details such as "steps back," "tilts far to the left," and "stumbles" are accurately captured by our ParCo, while other methods either ignore or incorrectly generate these nuanced actions. Besides, as illustrated in Fig. 1 and Fig. 5, ParCo is endowed the part-level composition superiority across action-level spatial and temporal compositions respectively. More details are provided in the supplementary materials. + +GT Leakage of Diffusion-based Methods Previous studies [8, 61, 67, 68] based on the Diffusion model utilize the ground truth motion length as an input for synthesizing motions during evaluation, contributing to their remarkable FID scores. However, this approach is impractical for real-world applications. For clarity, we replace the GT motion lengths with random lengths sampled uniformly from dataset's motion length range to evaluate ReMoDiff. On HumanML3D (KIT-ML), the ReMoDiff's Top-1, Top-2, and Top-3 R-precision decrease by 0.041 (0.053), 0.040 (0.068), and 0.033 (0.075) respectively, while the FID increases by 0.147 (0.426), justifying our speculation. As an auto-regressive approach, our ParCo demonstrates competitive R-Precision and FID without requiring pre-defined motion lengths Table 1. Given its lower computational and parameter consumption (Table. 4), the superiority of ParCo is convincing. + +# 4.3 Analysis + +Performance on Text Inputs of Different Lengths. In order to investigate the synthetic performance given textual descriptions of different lengths, we sort the HumanML3D test set based on the length of textual descriptions, and split it into four subsets with approximately equal numbers of text-motion pairs: 0-25%, 25-50%, 50-75%, and 75-100%, from short to long. The details of these splits are available in the supplementary materials. We conducted evaluations of real motion, T2M-GPT [66], and our method on these four splits. As illustrated in Fig. 6, our method exhibits comprehensive improvements in all subsets when contrasted with the baseline [66]. It demonstrates that our design is beneficial for text inputs of different lengths without compromise. + +Ablation Study. We investigate different discretizations for whole-body motion and conduct ablation of our Part Coordination module. We compare our 6-part partition and upper-lower body partition proposed by SCA [13]. It is noteworthy that, despite SCA dividing whole-body into upper and lower body for motion generation, its generations of upper body motion and lower body motion are entirely independent and lack coordination compared with our method. As demonstrated in Table. 3, the comparison between our ParCo (D) and the + +![](images/6e7c5d6c4ea560f9da8b49f829c29c248037bf744019bf988effa215c1a7c755.jpg) +Fig.6: Comparison on 4 HumanML3D test subsets divided based on text length. $0-25\%$ , $25-50\%$ , $50-75\%$ , and $75-100\%$ respectively represent four subsets based on text length from short to long. + +![](images/186b49529ed66f5fe5eac65f3731e5a63e04ddd574eadaf66a951d4499c2a3c8.jpg) +(b) Performance gap relative to Real + +![](images/233a93902faa6074809a7d02602e9afcb9401fde6871336ff6c95e8aa426397c.jpg) + +![](images/2e5abcbfb74c6907abf0c3a1c29831975e6eb7086aac0a7cf57ec2fe9b221dc8.jpg) + +Table 3: Ablations of body discretization and part coordination module. * denotes our ParCo. + +
Discretization Part Coord.Top-1 ↑FID ↓MM-Dist ↓
AUp&LowBody×0.444±.0030.497±.0153.367±.009
B✓0.491±.0030.172±.0073.071±.008
C6 Parts×0.375±.0033.652±.0304.012±.012
*D✓0.515±.0030.109±.0052.927±.008
+ +Table 4: Computational complexity analysis. + +
MethodParam(M)FLOPs(G)InferTime(s)
ReMoDiff198.2481.00.091
T2M-GPT237.6292.30.544
ParCo168.4211.70.036
+ +"a man picks something with his [part description], shakes it, then puts it back." + +![](images/ef918190c3df28eacbb9360a9a6e3da3e40fca8d11754530646b6549053f25ce.jpg) +Fig. 7: Qualitative result of left-right exchange experiment on our ParCo. + +![](images/b95f506c8cf2bb9c7a0b9499a7a80e52579b5c2a7941516f7f6efeca96b350d6.jpg) + +re-implemented SCA (A) indicates that our method's generated motions exhibit significantly higher performance in Top-1, MM-Dist, and FID than SCA. Furthermore, the contrasts between results (A) and (B), as well as results (C) and (D), underscore the necessity of our ParCo facilitating communication and coordination among different part motions during the generation process. In addition, the comparison between results (B) and (D) validates the effectiveness of our 6-part partitioning, which makes the model aware of the concept of parts and contributes to text-to-motion synthesis. + +Computational complexity analysis. With lightweight architectures and parallel computing support, our ParCo shows superior efficiency from parameters, flops, and inference time, justifying the efficacy of our part discretization and coordination designs, Table. 4. + +Precise part control. To compare the control of different methods over the movement of human parts, we conducted a left-right exchange experiment on + +Table 5: Evaluation of real motion data on train, val, and test set of HumanML3D. + +
SplitTop-1Top-2Top-3MM-DistDiversity
Train0.628±.0010.810±.0010.888±.0012.388±.0039.685±.083
Val0.513±.0040.703±.0040.800±.0032.911±.0109.575±.081
Test0.512±.0030.703±.0020.797±.0022.973±.0079.495±.079
+ +ten sentences. We examined several methods for generating actions and accomplishing the left-right swap task, the success rates are listed as follows: $70\%$ (ParCo), $50\%$ (T2M-GPT), $30\%$ (ReMoDiff), $20\%$ (MDM), $0\%$ (MoDiff). Notably, our ParCo exhibited the highest accuracy, underscoring its proficiency in the awareness of human parts. The qualitative results are showcased in Fig. 7. + +Discussion. As illustrated in Table 1 and Table 2, our performance surpasses that of real motion due to the evaluation protocol [14], which uses pretrained feature extractors trained only on the train split. These extractors are more effective at extracting features from motions similar to those in the train split. Table 5 shows that evaluations on real motions from the Train, Val, and Test sets reveal significantly higher Top-1/2/3 and MM-Dist metrics on the Train split. This indicates better extractor understanding of the Train split data. Our model, trained on the train split, generates motions resembling the train data, making feature extraction and matching easier, leading to superior performance on the test set. Due to the lack of a comprehensive metric for text-to-motion semantic alignment, motion generation fidelity, and diversity, we advocate for a more holistic evaluation within the community. + +# 5 Conclusion + +In this study, we focus on enhancing the text-to-motion generation model's ability to comprehend part concepts and facilitate communication between different part motion generators, ultimately yielding the synthesis of coordinated and fined-grained motion. Specifically, we discretize whole-body motion into multiple-part motions to establish the prior concept of parts. Afterward, we employ multiple lightweight generators designed to synthesize different part motions and coordinate them through our part coordination module. Extensive experiments showcase that our method achieves higher consistency between generated motions and textual descriptions compared to previous SOTA methods. Furthermore, in-depth analytical results suggest that our approach excels in achieving more precise part control and has lower computational complexity. More encouragingly, our method exhibits adaptability to various part separation schemes and holds the potential for further refinement toward hierarchical part motion. We anticipate that this will have a far-reaching impact on the community. + +# Acknowledgements + +This work was supported by the National Key R&D Program of China under Grant 2018AAA0102801. + +# References + +1. Ahn, H., Ha, T., Choi, Y., Yoo, H., Oh, S.: Text2action: Generative adversarial synthesis from language to action. In: ICRA (2018) +2. Ahuja, C., Morency, L.P.: Language2pose: Natural language grounded pose forecasting. In: 3DV (2019) +3. Antakli, A., Hermann, E., Zinnikus, I., Du, H., Fischer, K.: Intelligent distributed human motion simulation in human-robot collaboration environments. In: ACM IVA (2018) +4. Athanasiou, N., Petrovich, M., Black, M.J., Varol, G.: Sinc: Spatial composition of 3d human motions for simultaneous action generation. arXiv preprint arXiv:2304.10417 (2023) +5. Barsoum, E., Kender, J., Liu, Z.: Hp-gan: Probabilistic 3d human motion prediction via gan. In: CVPRW (2018) +6. Bhattacharya, U., Rewkowski, N., Banerjee, A., Guhan, P., Bera, A., Manocha, D.: Text2gestures: A transformer-based network for generating emotive body gestures for virtual agents. In: 2021 IEEE virtual reality and 3D user interfaces (VR) (2021) +7. Butepage, J., Black, M.J., Kragic, D., Kjellstrom, H.: Deep representation learning for human motion prediction and classification. In: CVPR (2017) +8. Chen, X., Jiang, B., Liu, W., Huang, Z., Fu, B., Chen, T., Yu, G.: Executing your commands via motion diffusion in latent space. In: CVPR (2023) +9. Chen, X., Su, Z., Yang, L., Cheng, P., Xu, L., Fu, B., Yu, G.: Learning variational motion prior for video-based motion capture. arXiv preprint arXiv:2210.15134 (2022) +0. Djuric, N., Radosavljevic, V., Cui, H., Nguyen, T., Chou, F.C., Lin, T.H., Singh, N., Schneider, J.: Uncertainty-aware short-term motion prediction of traffic actors for autonomous driving. In: WACV (2020) +1. Duan, Y., Shi, T., Zou, Z., Lin, Y., Qian, Z., Zhang, B., Yuan, Y.: Single-shot motion completion with transformer. arXiv preprint arXiv:2103.00776 (2021) +2. Fragkiadaki, K., Levine, S., Felsen, P., Malik, J.: Recurrent network models for human dynamics. In: ICCV (2015) +3. Ghosh, A., Cheema, N., Oguz, C., Theobalt, C., Slusallek, P.: Synthesis of compositional animations from textual descriptions. In: ICCV (2021) +4. Guo, C., Zou, S., Zuo, X., Wang, S., Ji, W., Li, X., Cheng, L.: Generating diverse and natural 3d human motions from text. In: CVPR (2022) +5. Guo, C., Zuo, X., Wang, S., Cheng, L.: Tm2t: Stochastic and tokenized modeling for the reciprocal generation of 3d human motions and texts. In: ECCV (2022) +6. Guo, C., Zuo, X., Wang, S., Zou, S., Sun, Q., Deng, A., Gong, M., Cheng, L.: Action2motion: Conditioned generation of 3d human motions. In: ACM MM (2020) +7. Harvey, F.G., Pal, C.: Recurrent transition networks for character locomotion. In: SIGGRAPH (2018) +8. Harvey, F.G., Yurick, M., Nowrouzezahrai, D., Pal, C.: Robust motion in betweening. ACM TOG (2020) + +19. Herbet, G., Duffau, H.: Revisiting the functional anatomy of the human brain: toward a meta-networking theory of cerebral functions. Physiological Reviews (2020) +20. Hernandez, A., Gall, J., Moreno-Noguer, F.: Human motion prediction via spatiotemporal inpainting. In: ICCV (2019) +21. Heusel, M., Ramsauer, H., Unterthiner, T., Nessler, B., Hochreiter, S.: Gans trained by a two time-scale update rule converge to a local nash equilibrium. NeurIPS (2017) +22. Ho, J., Jain, A., Abbeel, P.: Denoising diffusion probabilistic models. NeurIPS (2020) +23. Kappel, M., Golyanik, V., Elgharib, M., Henningson, J.O., Seidel, H.P., Castillo, S., Theobalt, C., Magnor, M.: High-fidelity neural human motion transfer from monocular video. In: CVPR (2021) +24. Kaufmann, M., Aksan, E., Song, J., Pece, F., Ziegler, R., Hilliges, O.: Convolutional autoencoders for human motion infilling. In: 3DV (2020) +25. Kim, J., Kim, J., Choi, S.: Flame: Free-form language-based motion synthesis & editing. In: AAAI (2023) +26. Koppula, H., Saxena, A.: Learning spatio-temporal structure from rgb-d videos for human activity detection and anticipation. In: ICML (2013) +27. Koppula, H.S., Saxena, A.: Anticipating human activities using object affordances for reactive robotic response. IEEE TPAMI (2015) +28. Kullback, S.: Information theory and statistics (1997) +29. Lab, C.G.: Cmu graphics lab motion capture database (2016) +30. Lee, H.Y., Yang, X., Liu, M.Y., Wang, T.C., Lu, Y.D., Yang, M.H., Kautz, J.: Dancing to music. NeurIPS (2019) +31. Li, B., Zhao, Y., Zhelun, S., Sheng, L.: Danceformer: Music conditioned 3d dance generation with parametric motion transformer. In: AAAI (2022) +32. Li, R., Yang, S., Ross, D.A., Kanazawa, A.: Ai choreographer: Music conditioned 3d dance generation with aist++. In: ICCV (2021) +33. Liu, Z., Lin, Y., Cao, Y., Hu, H., Wei, Y., Zhang, Z., Lin, S., Guo, B.: Swin transformer: Hierarchical vision transformer using shifted windows. In: ICCV (2021) +34. Loper, M., Mahmood, N., Romero, J., Pons-Moll, G., Black, M.J.: Spl: A skinned multi-person linear model. In: Seminal Graphics Papers: Pushing the Boundaries, Volume 2 (2023) +35. Loshchilov, I., Hutter, F.: Decoupled weight decay regularization. arXiv preprint arXiv:1711.05101 (2017) +36. Mahmood, N., Ghorbani, N., Troje, N.F., Pons-Moll, G., Black, M.J.: Amass: Archive of motion capture as surface shapes. In: ICCV (2019) +37. Majoe, D., Widmer, L., Gutknecht, J.: Enhanced motion interaction for multimedia applications. In: Proceedings of the 7th International Conference on Advances in Mobile Computing and Multimedia (2009) +38. Mao, W., Liu, M., Salzmann, M.: History repeats itself: Human motion prediction via motion attention. In: ECCV (2020) +39. Mao, W., Liu, M., Salzmann, M., Li, H.: Learning trajectory dependencies for human motion prediction. In: ICCV (2019) +40. Martinez, J., Black, M.J., Romero, J.: On human motion prediction using recurrent neural networks. In: CVPR (2017) +41. Nichol, A., Dhariwal, P., Ramesh, A., Shyam, P., Mishkin, P., McGrew, B., Sutskever, I., Chen, M.: Glide: Towards photorealistic image generation and editing with text-guided diffusion models. arXiv preprint arXiv:2112.10741 (2021) +42. Parmar, N., Vaswani, A., Uszkoreit, J., Kaiser, L., Shazeer, N., Ku, A., Tran, D.: Image transformer. In: ICML (2018) + +43. Pavlakos, G., Choutas, V., Ghorbani, N., Bolkart, T., Osman, A.A., Tzionas, D., Black, M.J.: Expressive body capture: 3d hands, face, and body from a single image. In: CVPR (2019) +44. Pavllo, D., Grangier, D., Auli, M.: Quaternet: A quaternion-based recurrent model for human motion. arXiv preprint arXiv:1805.06485 (2018) +45. Petrovich, M., Black, M.J., Varol, G.: Action-conditioned 3d human motion synthesis with transformer vae. In: ICCV (2021) +46. Petrovich, M., Black, M.J., Varol, G.: Temos: Generating diverse human motions from textual descriptions. In: ECCV (2022) +47. Plappert, M., Mandery, C., Asfour, T.: The kit motion-language dataset. Big data (2016) +48. Plappert, M., Mandery, C., Asfour, T.: Learning a bidirectional mapping between human whole-body motion and natural language using deep recurrent neural networks. Robotics and Autonomous Systems (2018) +49. Raab, S., Leibovitch, I., Li, P., Aberman, K., Sorkine-Hornung, O., Cohen-Or, D.: Modi: Unconditional motion synthesis from diverse data. In: CVPR (2023) +50. Radford, A., Kim, J.W., Hallacy, C., Ramesh, A., Goh, G., Agarwal, S., Sastry, G., Askell, A., Mishkin, P., Clark, J., et al.: Learning transferable visual models from natural language supervision. In: ICML (2021) +51. Ramesh, A., Dhariwal, P., Nichol, A., Chu, C., Chen, M.: Hierarchical text-conditional image generation with clip latents. arXiv preprint arXiv:2204.06125 (2022) +52. Rempe, D., Birdal, T., Hertzmann, A., Yang, J., Sridhar, S., Guibas, L.J.: Humor: 3d human motion model for robust pose estimation. In: ICCV (2021) +53. Rombach, R., Blattmann, A., Lorenz, D., Esser, P., Ommer, B.: High-resolution image synthesis with latent diffusion models. In: CVPR (2022) +54. Sahara, C., Chan, W., Saxena, S., Li, L., Whang, J., Denton, E.L., Ghasemipour, K., Gontijo Lopes, R., Karagol Ayan, B., Salimans, T., et al.: Photorealistic text-to-image diffusion models with deep language understanding. NeurIPS (2022) +55. Thiebaut de Schotten, M., Forkel, S.J.: The emergent properties of the connected brain. Science (2022) +56. Shafir, Y., Tevet, G., Kapon, R., Bermano, A.H.: Human motion diffusion as a generative prior. arXiv preprint arXiv:2303.01418 (2023) +57. Siyao, L., Yu, W., Gu, T., Lin, C., Wang, Q., Qian, C., Loy, C.C., Liu, Z.: Bailando: 3d dance generation by actor-critic gpt with choreographic memory. In: CVPR (2022) +58. Tang, X., Wang, H., Hu, B., Gong, X., Yi, R., Kou, Q., Jin, X.: Real-time controllable motion transition for characters. ACM TOG (2022) +59. Terlemez, Ö., Ulbrich, S., Mandery, C., Do, M., Vahrenkamp, N., Asfour, T.: Master motor map (mmm)—framework and toolkit for capturing, representing, and reproducing human motion on humanoid robots. In: 2014 IEEE-RAS International Conference on Humanoid Robots (2014) +60. Tevet, G., Gordon, B., Hertz, A., Bermano, A.H., Cohen-Or, D.: Motionclip: Exposing human motion generation to clip space. In: ECCV (2022) +61. Tevet, G., Raab, S., Gordon, B., Shafir, Y., Cohen-Or, D., Bermano, A.H.: Human motion diffusion model. arXiv preprint arXiv:2209.14916 (2022) +62. Vaswani, A., Shazeer, N., Parmar, N., Uszkoreit, J., Jones, L., Gomez, A.N., Kaiser, L., Polosukhin, I.: Attention is all you need. NeurIPS (2017) +63. Wang, Y., Leng, Z., Li, F.W., Wu, S.C., Liang, X.: Fg-t2m: Fine-grained text-driven human motion generation via diffusion model. In: ICCV (2023) + +64. Yan, S., Li, Z., Xiong, Y., Yan, H., Lin, D.: Convolutional sequence generation for skeleton-based action synthesis. In: ICCV (2019) +65. Yeasin, M., Polat, E., Sharma, R.: A multiobject tracking framework for interactive multimedia applications. IEEE TMM (2004) +66. Zhang, J., Zhang, Y., Cun, X., Huang, S., Zhang, Y., Zhao, H., Lu, H., Shen, X.: T2m-gpt: Generating human motion from textual descriptions with discrete representations. arXiv preprint arXiv:2301.06052 (2023) +67. Zhang, M., Cai, Z., Pan, L., Hong, F., Guo, X., Yang, L., Liu, Z.: Motiondiffuse: Text-driven human motion generation with diffusion model. arXiv preprint arXiv:2208.15001 (2022) +68. Zhang, M., Guo, X., Pan, L., Cai, Z., Hong, F., Li, H., Yang, L., Liu, Z.: Remodif- fuse: Retrieval-augmented motion diffusion model. arXiv preprint arXiv:2304.01116 (2023) +69. Zhang, Y., Black, M.J., Tang, S.: Perpetual motion: Generating unbounded human motion. arXiv preprint arXiv:2007.13886 (2020) +70. Zhao, R., Su, H., Ji, Q.: Bayesian adversarial human motion synthesis. In: CVPR (2020) +71. Zhong, C., Hu, L., Zhang, Z., Xia, S.: Attt2m: Text-driven human motion generation with multi-perspective attention mechanism. In: ICCV (2023) \ No newline at end of file diff --git a/2024/ParCo_ Part-Coordinating Text-to-Motion Synthesis/images.zip b/2024/ParCo_ Part-Coordinating Text-to-Motion Synthesis/images.zip new file mode 100644 index 0000000000000000000000000000000000000000..06e5b2111962ef7bb9d48a403cba465751801b96 --- /dev/null +++ b/2024/ParCo_ Part-Coordinating Text-to-Motion Synthesis/images.zip @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:d7aec16b8d2ebca8c45c32c3add998c0f6b53c0108c422142b501193c1a896f4 +size 650182 diff --git a/2024/ParCo_ Part-Coordinating Text-to-Motion Synthesis/layout.json b/2024/ParCo_ Part-Coordinating Text-to-Motion Synthesis/layout.json new file mode 100644 index 0000000000000000000000000000000000000000..85fbc7b34c76d758a507c0fa20d1dcf5ee5537c3 --- /dev/null +++ b/2024/ParCo_ Part-Coordinating Text-to-Motion Synthesis/layout.json @@ -0,0 +1,8973 @@ +{ + "pdf_info": [ + { + "para_blocks": [ + { + "bbox": [ + 159, + 111, + 455, + 148 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 159, + 111, + 455, + 148 + ], + "spans": [ + { + "bbox": [ + 159, + 111, + 455, + 148 + ], + "type": "text", + "content": "ParCo: Part-Coordinating Text-to-Motion Synthesis" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 142, + 167, + 473, + 194 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 142, + 167, + 473, + 194 + ], + "spans": [ + { + "bbox": [ + 142, + 167, + 473, + 194 + ], + "type": "text", + "content": "Qiran Zou\\*, Shangyuan Yuan\\*1, Shian Du', Yu Wang?, Chang Liu?, Yi Xu?, Jie Chen?, and Xiangyang Ji!" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 194, + 201, + 419, + 224 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 194, + 201, + 419, + 224 + ], + "spans": [ + { + "bbox": [ + 194, + 201, + 419, + 224 + ], + "type": "inline_equation", + "content": "^{1}" + }, + { + "bbox": [ + 194, + 201, + 419, + 224 + ], + "type": "text", + "content": " Tsinghua University, China \n" + }, + { + "bbox": [ + 194, + 201, + 419, + 224 + ], + "type": "inline_equation", + "content": "^{2}" + }, + { + "bbox": [ + 194, + 201, + 419, + 224 + ], + "type": "text", + "content": " Peking University Shenzhen Graduate School, China" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 138, + 224, + 474, + 268 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 224, + 474, + 268 + ], + "spans": [ + { + "bbox": [ + 138, + 224, + 474, + 268 + ], + "type": "inline_equation", + "content": "^{3}" + }, + { + "bbox": [ + 138, + 224, + 474, + 268 + ], + "type": "text", + "content": " Dalian University of Technology, China qiranzou@gmail.com, {yuansy21, dsa23}@mails.tsinghua.edu.cn, 2201212856@stu.pku.edu.cn, liuchang2022@tsinghua.edu.cn, yxu@dlut.edu.cn, chenj@pcl.ac.cn, xyji@tsinghua.edu.cn" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 159, + 299, + 455, + 518 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 159, + 299, + 455, + 518 + ], + "spans": [ + { + "bbox": [ + 159, + 299, + 455, + 518 + ], + "type": "text", + "content": "Abstract. We study a challenging task: text-to-motion synthesis, aiming to generate motions that align with textual descriptions and exhibit coordinated movements. Currently, the part-based methods introduce part partition into the motion synthesis process to achieve finer-grained generation. However, these methods encounter challenges such as the lack of coordination between different part motions and difficulties for networks to understand part concepts. Moreover, introducing finer-grained part concepts poses computational complexity challenges. In this paper, we propose Part-Coordinating Text-to-Motion Synthesis (ParCo), endowed with enhanced capabilities for understanding part motions and communication among different part motion generators, ensuring a coordinated and fined-grained motion synthesis. Specifically, we discretize whole-body motion into multiple part motions to establish the prior concept of different parts. Afterward, we employ multiple lightweight generators designed to synthesize different part motions and coordinate them through our part coordination module. Our approach demonstrates superior performance on common benchmarks with economic computations, including HumanML3D and KIT-ML, providing substantial evidence of its effectiveness. Code is available at: https://github.com/qrzou/ParCo." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 160, + 529, + 436, + 540 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 160, + 529, + 436, + 540 + ], + "spans": [ + { + "bbox": [ + 160, + 529, + 436, + 540 + ], + "type": "text", + "content": "Keywords: Motion synthesis " + }, + { + "bbox": [ + 160, + 529, + 436, + 540 + ], + "type": "inline_equation", + "content": "\\cdot" + }, + { + "bbox": [ + 160, + 529, + 436, + 540 + ], + "type": "text", + "content": " Part coordination " + }, + { + "bbox": [ + 160, + 529, + 436, + 540 + ], + "type": "inline_equation", + "content": "\\cdot" + }, + { + "bbox": [ + 160, + 529, + 436, + 540 + ], + "type": "text", + "content": " Text-to-motion" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 132, + 561, + 231, + 574 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 561, + 231, + 574 + ], + "spans": [ + { + "bbox": [ + 132, + 561, + 231, + 574 + ], + "type": "text", + "content": "1 Introduction" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 130, + 586, + 482, + 647 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 586, + 482, + 647 + ], + "spans": [ + { + "bbox": [ + 130, + 586, + 482, + 647 + ], + "type": "text", + "content": "Text-to-motion synthesis aims to generate motion that aligns with textual descriptions and exhibits coordinated movements. It facilitates obtaining desired motion through textual descriptions which benefits numerous applications in industrial scenarios such as animation [23], AR/VR applications, video games [37, 65], autonomous driving [10], and robotics [3,26,27]." + } + ] + } + ], + "index": 7 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 133, + 653, + 223, + 665 + ], + "type": "page_footnote", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 653, + 223, + 665 + ], + "spans": [ + { + "bbox": [ + 133, + 653, + 223, + 665 + ], + "type": "text", + "content": "* Equal contribution." + } + ] + } + ], + "index": 8 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 0 + }, + { + "para_blocks": [ + { + "bbox": [ + 212, + 128, + 403, + 137 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 212, + 128, + 403, + 137 + ], + "spans": [ + { + "bbox": [ + 212, + 128, + 403, + 137 + ], + "type": "text", + "content": "\"am is raising his left hand while kicking something with his right foot\"" + } + ] + } + ], + "index": 2 + }, + { + "type": "image", + "bbox": [ + 140, + 137, + 475, + 194 + ], + "blocks": [ + { + "bbox": [ + 140, + 137, + 475, + 194 + ], + "lines": [ + { + "bbox": [ + 140, + 137, + 475, + 194 + ], + "spans": [ + { + "bbox": [ + 140, + 137, + 475, + 194 + ], + "type": "image", + "image_path": "e6ad0c4221961cca5a2376237fc46ca9d144a908a77881126e96cda2c07efcd1.jpg" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_body" + } + ], + "index": 3 + }, + { + "type": "image", + "bbox": [ + 140, + 196, + 475, + 255 + ], + "blocks": [ + { + "bbox": [ + 140, + 196, + 475, + 255 + ], + "lines": [ + { + "bbox": [ + 140, + 196, + 475, + 255 + ], + "spans": [ + { + "bbox": [ + 140, + 196, + 475, + 255 + ], + "type": "image", + "image_path": "90d386a8f15cfc164f3f8952b2cc9b611fe6d7ee49dde270eb4231c1d2c7f6fb.jpg" + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 130, + 267, + 480, + 289 + ], + "lines": [ + { + "bbox": [ + 130, + 267, + 480, + 289 + ], + "spans": [ + { + "bbox": [ + 130, + 267, + 480, + 289 + ], + "type": "text", + "content": "Fig. 1: Our ParCo is capable of coordinating the motion of various body parts to produce realistic and accurate motion." + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "image_caption" + } + ], + "index": 4 + }, + { + "bbox": [ + 130, + 294, + 480, + 402 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 294, + 480, + 402 + ], + "spans": [ + { + "bbox": [ + 130, + 294, + 480, + 402 + ], + "type": "text", + "content": "Recent advancements leveraging powerful generation capabilities of transformers [33,42,62] and diffusion models [22,41,51,53,54] have yielded impressive results in generating realistic and smooth motions. Despite this progress, existing methods often struggle to generate semantically matched and coherently coordinated motions, especially when faced with commands involving multiple coordinated body parts. We attribute this challenge to the intricate alignment problem between the modalities of text and motion, where a single text/motion can correspond to multiple possible motion/text, posing a challenge in learning the complex relationship between the two." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 130, + 402, + 481, + 605 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 402, + 481, + 605 + ], + "spans": [ + { + "bbox": [ + 130, + 402, + 481, + 605 + ], + "type": "text", + "content": "In the realm of Text-to-Motion Generation, part-based methods aim to achieve a higher level of sophistication in motion generation. These approaches can be broadly classified into two categories: single generator with part-level motion embeddings (Fig. 2 (a)) and independent upper and lower body motion generators (Fig. 2 (b)). The former one constructs whole-body motion embeddings by amalgamating multiple part motion embeddings, explicitly introducing the concept of parts [57], or implicitly integrating part concepts with part-level attention [71]. However, utilizing a single generator to generate whole-body motion embeddings presents a challenge for the generator to understand the concept of parts. The latter one segregates whole-body motion into upper and lower body motions, deploying two independent generators to produce motions separately [13]. Although this design enhances the generator's comprehension of upper and lower body motions, the absence of information exchange between them leads to a lack of coordination in the resulting upper and lower body motions. To increase the granularity of part division, an intuitive solution is to employ additional generators for generating finer-grained part motions, such as the left leg. However, this approach introduces computational complexity challenges." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 130, + 605, + 481, + 665 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 605, + 481, + 665 + ], + "spans": [ + { + "bbox": [ + 130, + 605, + 481, + 665 + ], + "type": "text", + "content": "Neuroscientific discoveries reveal that discrete regions within the human brain manifest unique functions, and these regions engage in communication to coordinate diverse activities [19, 55]. This design, where different low-level subsystems communicate to form a higher-level system, is prevalent in the natural and artificial world, offering a robust, perceptually strong structural design. In adher-" + } + ] + } + ], + "index": 8 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "spans": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "type": "text", + "content": "2" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 166, + 91, + 220, + 102 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 166, + 91, + 220, + 102 + ], + "spans": [ + { + "bbox": [ + 166, + 91, + 220, + 102 + ], + "type": "text", + "content": "Q. Zou et al." + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 1 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 134, + 114, + 244, + 227 + ], + "blocks": [ + { + "bbox": [ + 134, + 114, + 244, + 227 + ], + "lines": [ + { + "bbox": [ + 134, + 114, + 244, + 227 + ], + "spans": [ + { + "bbox": [ + 134, + 114, + 244, + 227 + ], + "type": "image", + "image_path": "e59bce9596372b2b546709d57639cb47f437108783eda0ae41831c3d5d1439c5.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 130, + 239, + 482, + 306 + ], + "lines": [ + { + "bbox": [ + 130, + 239, + 482, + 306 + ], + "spans": [ + { + "bbox": [ + 130, + 239, + 482, + 306 + ], + "type": "text", + "content": "Fig. 2: Conceptual comparison of three part-based synthesis methods. (a): One generator synthesizes the whole-body embedding, which contains information about different parts internally. (b): Two separate generators synthesize the upper and lower body's motions independently, without information exchange between them. (c): Our ParCo employs multiple lightweight generators designed to synthesize different part motions, which are coordinated by the Part Coordination module." + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "image_caption" + } + ], + "index": 2 + }, + { + "type": "image", + "bbox": [ + 250, + 114, + 361, + 227 + ], + "blocks": [ + { + "bbox": [ + 250, + 114, + 361, + 227 + ], + "lines": [ + { + "bbox": [ + 250, + 114, + 361, + 227 + ], + "spans": [ + { + "bbox": [ + 250, + 114, + 361, + 227 + ], + "type": "image", + "image_path": "0a24dc73775f6e81ea18b8f13d7c54a90b4dbf191702fe192e17eb9a2bc2322c.jpg" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_body" + } + ], + "index": 3 + }, + { + "type": "image", + "bbox": [ + 370, + 114, + 482, + 227 + ], + "blocks": [ + { + "bbox": [ + 370, + 114, + 482, + 227 + ], + "lines": [ + { + "bbox": [ + 370, + 114, + 482, + 227 + ], + "spans": [ + { + "bbox": [ + 370, + 114, + 482, + 227 + ], + "type": "image", + "image_path": "7d86fb4db43edb9f2676c60a35762d2e95e2b9278e6c329aa7d3bb6cc605389b.jpg" + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "image_body" + } + ], + "index": 4 + }, + { + "bbox": [ + 130, + 327, + 482, + 422 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 327, + 482, + 422 + ], + "spans": [ + { + "bbox": [ + 130, + 327, + 482, + 422 + ], + "type": "text", + "content": "ence to these principles, we present Part-Coordinating Text-to-Motion Synthesis (ParCo). It comprises six small generators, which are tasked with various part motions, and accompanied by a Part Coordination module facilitating communication among the generators. This communication enables the generation of coordinated whole-body motions while comprehending distinct parts (Fig. 2 (c)). Thanks to the design of small generators, our method demonstrates a lower parameter count, reduced computational complexity, and shorter generation time in comparison to baseline and state-of-the-art methods." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 130, + 423, + 482, + 530 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 423, + 482, + 530 + ], + "spans": [ + { + "bbox": [ + 130, + 423, + 482, + 530 + ], + "type": "text", + "content": "Specifically, our approach consists of two stages. In the first stage, we discretize whole-body motion into multiple part motions and quantize them using VQ-VAEs, providing prior knowledge of \"what is part\" for the next stage. In the second stage, we use multiple Part-Coordinated Transformers, which are capable of communicating with each other, to generate coordinated motions of different parts. These part motions are integrated into whole-body motion subsequently. Extensive experiments on HumanML3D [14] and KIT-ML [47] demonstrate that our method can generate realistic and coordinated motions that align with the semantic descriptions." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 146, + 531, + 331, + 542 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 146, + 531, + 331, + 542 + ], + "spans": [ + { + "bbox": [ + 146, + 531, + 331, + 542 + ], + "type": "text", + "content": "We delineate our contributions as follows:" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 143, + 548, + 481, + 665 + ], + "type": "list", + "angle": 0, + "index": 12, + "blocks": [ + { + "bbox": [ + 143, + 548, + 480, + 585 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 143, + 548, + 480, + 585 + ], + "spans": [ + { + "bbox": [ + 143, + 548, + 480, + 585 + ], + "type": "text", + "content": "- We propose ParCo, which enables the generators to better understand finer-grained parts and coordinate the generated part motions, ultimately achieving fine-grained and coordinated motion synthesis." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 143, + 589, + 480, + 613 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 143, + 589, + 480, + 613 + ], + "spans": [ + { + "bbox": [ + 143, + 589, + 480, + 613 + ], + "type": "text", + "content": "- Our approach is computationally efficient, although employing multiple part generators, and maintains excellent motion generation performance." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 143, + 617, + 481, + 665 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 143, + 617, + 481, + 665 + ], + "spans": [ + { + "bbox": [ + 143, + 617, + 481, + 665 + ], + "type": "text", + "content": "- On the HumanML3D and KIT-ML datasets, our method significantly outperforms methods that do not rely on GT motion length as our ParCo, while demonstrating comparable performance to methods that do rely on GT motion length." + } + ] + } + ], + "index": 11 + } + ], + "sub_type": "text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 235, + 91, + 448, + 102 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 235, + 91, + 448, + 102 + ], + "spans": [ + { + "bbox": [ + 235, + 91, + 448, + 102 + ], + "type": "text", + "content": "ParCo: Part-Coordinating Text-to-Motion Synthesis" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 474, + 92, + 480, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 474, + 92, + 480, + 100 + ], + "spans": [ + { + "bbox": [ + 474, + 92, + 480, + 100 + ], + "type": "text", + "content": "3" + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 2 + }, + { + "para_blocks": [ + { + "bbox": [ + 133, + 114, + 237, + 127 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 114, + 237, + 127 + ], + "spans": [ + { + "bbox": [ + 133, + 114, + 237, + 127 + ], + "type": "text", + "content": "2 Related Work" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 133, + 142, + 481, + 357 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 142, + 481, + 357 + ], + "spans": [ + { + "bbox": [ + 133, + 142, + 481, + 357 + ], + "type": "text", + "content": "Human Motion Synthesis. Tasks in the domain of human motion synthesis fall within two distinct categories: unconditional motion generation and conditional motion generation. The categorization of these tasks is based on the input signals employed. Unconditional motion generation [49, 64, 69, 70], such as VPoser [43] and ACTOR [45], is a comprehensive task involving the modeling of the entire motion space, utilizing solely motion data for training and prediction. Human motion prediction, a highly dynamic field, seeks to forecast future movements based on observed motion. Another significant domain pertains to the generation of \"in-betweening\" motions, which fill the gaps between past and future poses [11, 17, 18, 24, 58]. Unconditional motion generation commonly utilizes models well-suited for processing sequential data, including recursive [7, 12, 40, 44], generative adversarial [5, 20], graph convolutional [39], and attention [38] approaches. This enables the efficient generation of diverse motions by concurrently processing spatial and temporal signals. Conditional motion generation involves various multimodal data types, including text [2, 14, 15, 25, 46, 61], occluded pose sequences [11, 18, 61], images [9, 52], and sound [30-32]. Due to the rapid advancements in NLP, text-driven human motion generation has sustained a notably active status." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 133, + 378, + 481, + 665 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 378, + 481, + 665 + ], + "spans": [ + { + "bbox": [ + 133, + 378, + 481, + 665 + ], + "type": "text", + "content": "Text-driven Human Motion Generation. Text-to-motion aims to generate human motion based on input textual descriptions. In earlier studies, joint-latent models [2, 46] were employed, which integrate a text encoder and a motion encoder. Text2Action [1] employs a recursive model, generating motion from short texts. TEMOS [46] adopts a similar approach, using self-encoding structures for both text and motion constrained by KL divergence [28]. T2M-GPT [66], TM2T [15] replace recursive encoders with transformers and GRU structure, achieving promising results. MotionCLIP [60] directly introduces the powerful zero-shot-capable CLIP [50] text encoder, following the alignment of text and pose as in Language2Pose [2], and additionally renders images with CLIP image encoder for auxiliary supervision. Subsequently, solutions based on diffusion models [8,56,67] emerge. MDM [61], MotionDiffuse [67], ReMoDiffuse [68] introduce diffusion models based on probability mapping, enabling the generation of human motion sequences from textual descriptions. Although current research enables the convenient generation of human motions based on text, challenges persist, especially in handling complex textual descriptions involving different parts. These approaches often treat human motion as a whole, lacking an understanding of different body parts, exhibiting limited ability of aligning text and motion. Furthermore, other methods incorporate the concept of parts into the model to generate more granular motions [13,57,71]. However, these approaches encounter challenges, including the lack of coordination among different part motions and difficulties for networks to comprehend part concepts. In contrast, our ParCo demonstrates a superior understanding of parts and effectively coordinates part motions, with lower computational complexity." + } + ] + } + ], + "index": 4 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "spans": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "type": "text", + "content": "4" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 166, + 91, + 219, + 102 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 166, + 91, + 219, + 102 + ], + "spans": [ + { + "bbox": [ + 166, + 91, + 219, + 102 + ], + "type": "text", + "content": "Q. Zou et al." + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 3 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 132, + 114, + 301, + 216 + ], + "blocks": [ + { + "bbox": [ + 132, + 114, + 301, + 216 + ], + "lines": [ + { + "bbox": [ + 132, + 114, + 301, + 216 + ], + "spans": [ + { + "bbox": [ + 132, + 114, + 301, + 216 + ], + "type": "image", + "image_path": "cad4d6d83d8ed3ccd48cd341f4bb922398834c8076bb8c320c23c8199f195183.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 130, + 227, + 482, + 316 + ], + "lines": [ + { + "bbox": [ + 130, + 227, + 482, + 316 + ], + "spans": [ + { + "bbox": [ + 130, + 227, + 482, + 316 + ], + "type": "text", + "content": "Fig. 3: Pipeline of ParCo. ParCo consists of two stages: (a) The whole-body motion is discretized into 6 part motions, and encoded into 6 quantized code index sequences by 6 VQ-VAEs (encoder and quantizer). This process provides a priori about the concept of part motions for the second stage. (b) We use the quantized index sequences and corresponding textual description to train 6 transformers for part motion generation. At the same time, these generators are coordinated by our Part Coordination module. The generated part motion codes are decoded by VQ-VAE (decoder) to reconstruct the 6 part motions, which will be integrated into the final whole-body motion." + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "image_caption" + } + ], + "index": 2 + }, + { + "type": "image", + "bbox": [ + 313, + 116, + 482, + 216 + ], + "blocks": [ + { + "bbox": [ + 313, + 116, + 482, + 216 + ], + "lines": [ + { + "bbox": [ + 313, + 116, + 482, + 216 + ], + "spans": [ + { + "bbox": [ + 313, + 116, + 482, + 216 + ], + "type": "image", + "image_path": "c75e5c2b492baa1cddeaaf27d33b143669dc267341769b66affca63c7df4b321.jpg" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_body" + } + ], + "index": 3 + }, + { + "bbox": [ + 132, + 338, + 201, + 350 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 338, + 201, + 350 + ], + "spans": [ + { + "bbox": [ + 132, + 338, + 201, + 350 + ], + "type": "text", + "content": "3 Method" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 130, + 364, + 482, + 460 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 364, + 482, + 460 + ], + "spans": [ + { + "bbox": [ + 130, + 364, + 482, + 460 + ], + "type": "text", + "content": "Our method consists of two stages to generate motion with an understanding of part motions. In the first stage, we discretize the whole-body motion into multiple part motions to provide prior knowledge of \"what is part\" for the second stage. In the second stage, the objective is to enable the model to learn the concept of part and achieve mutual coordination among multiple part motion generators. With this design, our method can handle textual inputs involving different parts and generate human motion that aligns with the semantic descriptions in the text." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 131, + 478, + 332, + 489 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 131, + 478, + 332, + 489 + ], + "spans": [ + { + "bbox": [ + 131, + 478, + 332, + 489 + ], + "type": "text", + "content": "3.1 Part-Aware Motion Discretization" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 130, + 498, + 481, + 557 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 498, + 481, + 557 + ], + "spans": [ + { + "bbox": [ + 130, + 498, + 481, + 557 + ], + "type": "text", + "content": "In stage 1, our method partitions the whole-body motion into multiple part motions and independently encodes each of these part motions using a VQ-VAE. This ensures that each part motion possesses an independent representation space (encoding space), providing prior knowledge about the concept of part motion for the next stage." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 130, + 558, + 482, + 666 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 558, + 482, + 666 + ], + "spans": [ + { + "bbox": [ + 130, + 558, + 482, + 666 + ], + "type": "text", + "content": "The 3D human body models (e.g. SMPL, MMM) typically use Kinematic Trees to model human skeleton as 5 chains (limbs and backbone) for motion modeling. We inherit this division and add the Root part to represent trajectories. Therefore, we divide the whole-body motion into six parts motions: R.Arm, L.Arm, R.Leg, L.Leg, Backbone, and Root. As illustrated in Fig. 3 (a), the first four represent the right/left arm and right/left leg, while the backbone denotes the spine and skull, and the root represents the pelvis joint's movement information. The commonly used human motion datasets, HumanML3D [14] and KIT-ML [47], utilize different body skeleton models, SMPL [34] and MMM [59]" + } + ] + } + ], + "index": 9 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 235, + 91, + 448, + 102 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 235, + 91, + 448, + 102 + ], + "spans": [ + { + "bbox": [ + 235, + 91, + 448, + 102 + ], + "type": "text", + "content": "ParCo: Part-Coordinating Text-to-Motion Synthesis" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 474, + 92, + 480, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 474, + 92, + 480, + 100 + ], + "spans": [ + { + "bbox": [ + 474, + 92, + 480, + 100 + ], + "type": "text", + "content": "5" + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 4 + }, + { + "para_blocks": [ + { + "bbox": [ + 130, + 116, + 479, + 140 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 116, + 479, + 140 + ], + "spans": [ + { + "bbox": [ + 130, + 116, + 479, + 140 + ], + "type": "text", + "content": "respectively. We provide a detailed explanation of how we perform our six-part partitioning for these two skeleton models in the supplementary material." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 130, + 140, + 480, + 188 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 140, + 480, + 188 + ], + "spans": [ + { + "bbox": [ + 130, + 140, + 480, + 188 + ], + "type": "text", + "content": "The aforementioned partitioning process can be formalized as: given a motion sequence " + }, + { + "bbox": [ + 130, + 140, + 480, + 188 + ], + "type": "inline_equation", + "content": "M = [m_1, \\dots, m_C]" + }, + { + "bbox": [ + 130, + 140, + 480, + 188 + ], + "type": "text", + "content": ", where " + }, + { + "bbox": [ + 130, + 140, + 480, + 188 + ], + "type": "inline_equation", + "content": "C" + }, + { + "bbox": [ + 130, + 140, + 480, + 188 + ], + "type": "text", + "content": " is the number of frames, we separate it into part motions " + }, + { + "bbox": [ + 130, + 140, + 480, + 188 + ], + "type": "inline_equation", + "content": "\\{P^i = [p_1^i, \\dots, p_C^i]\\}" + }, + { + "bbox": [ + 130, + 140, + 480, + 188 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 130, + 140, + 480, + 188 + ], + "type": "inline_equation", + "content": "i \\in [1, \\dots, S]" + }, + { + "bbox": [ + 130, + 140, + 480, + 188 + ], + "type": "text", + "content": ", where " + }, + { + "bbox": [ + 130, + 140, + 480, + 188 + ], + "type": "inline_equation", + "content": "S" + }, + { + "bbox": [ + 130, + 140, + 480, + 188 + ], + "type": "text", + "content": " is the number of parts and " + }, + { + "bbox": [ + 130, + 140, + 480, + 188 + ], + "type": "inline_equation", + "content": "m_* = [p_*^1, \\dots, p_*^S]" + }, + { + "bbox": [ + 130, + 140, + 480, + 188 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 130, + 189, + 479, + 247 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 189, + 479, + 247 + ], + "spans": [ + { + "bbox": [ + 130, + 189, + 479, + 247 + ], + "type": "text", + "content": "After separating the whole-body motion into part motions, we further discretize the part motions into code sequences using VQ-VAE. We utilize this discretized representation in the next stage's generation process, as it exhibits better generalization capabilities and contributes to the improvement of training and inference efficiency." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 131, + 248, + 480, + 312 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 131, + 248, + 480, + 312 + ], + "spans": [ + { + "bbox": [ + 131, + 248, + 480, + 312 + ], + "type": "text", + "content": "Firstly, we use " + }, + { + "bbox": [ + 131, + 248, + 480, + 312 + ], + "type": "inline_equation", + "content": "Encoder^i" + }, + { + "bbox": [ + 131, + 248, + 480, + 312 + ], + "type": "text", + "content": " to get the " + }, + { + "bbox": [ + 131, + 248, + 480, + 312 + ], + "type": "inline_equation", + "content": "i" + }, + { + "bbox": [ + 131, + 248, + 480, + 312 + ], + "type": "text", + "content": "-th part motion's encoding that " + }, + { + "bbox": [ + 131, + 248, + 480, + 312 + ], + "type": "inline_equation", + "content": "E^i = Encoder^i(P^i) = [e_1^i, \\dots, e_l^i, \\dots, e_L^i]" + }, + { + "bbox": [ + 131, + 248, + 480, + 312 + ], + "type": "text", + "content": ", where " + }, + { + "bbox": [ + 131, + 248, + 480, + 312 + ], + "type": "inline_equation", + "content": "L = \\frac{C}{r}" + }, + { + "bbox": [ + 131, + 248, + 480, + 312 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 131, + 248, + 480, + 312 + ], + "type": "inline_equation", + "content": "r" + }, + { + "bbox": [ + 131, + 248, + 480, + 312 + ], + "type": "text", + "content": " is the downsampling rate of encoder. Then, we discretize the " + }, + { + "bbox": [ + 131, + 248, + 480, + 312 + ], + "type": "inline_equation", + "content": "E^i" + }, + { + "bbox": [ + 131, + 248, + 480, + 312 + ], + "type": "text", + "content": " into " + }, + { + "bbox": [ + 131, + 248, + 480, + 312 + ], + "type": "inline_equation", + "content": "Q^i = [v_{k_1^i}^i, \\dots, v_{k_l^i}^i, \\dots, v_{k_L^i}^i]" + }, + { + "bbox": [ + 131, + 248, + 480, + 312 + ], + "type": "text", + "content": " according to a learnable codebook " + }, + { + "bbox": [ + 131, + 248, + 480, + 312 + ], + "type": "inline_equation", + "content": "V^i = \\{v_j^i\\}, j = 1, \\dots, J" + }, + { + "bbox": [ + 131, + 248, + 480, + 312 + ], + "type": "text", + "content": ", where " + }, + { + "bbox": [ + 131, + 248, + 480, + 312 + ], + "type": "inline_equation", + "content": "J" + }, + { + "bbox": [ + 131, + 248, + 480, + 312 + ], + "type": "text", + "content": " is the number of codes in the codebook. The index " + }, + { + "bbox": [ + 131, + 248, + 480, + 312 + ], + "type": "inline_equation", + "content": "k_l" + }, + { + "bbox": [ + 131, + 248, + 480, + 312 + ], + "type": "text", + "content": " is obtained by finding the most similar code:" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 249, + 321, + 480, + 343 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 249, + 321, + 480, + 343 + ], + "spans": [ + { + "bbox": [ + 249, + 321, + 480, + 343 + ], + "type": "interline_equation", + "content": "k _ {l} ^ {i} = \\underset {j \\in \\{1, \\dots , J \\}} {\\arg \\min } \\left\\| e _ {l} ^ {i} - v _ {j} ^ {i} \\right\\|. \\tag {1}", + "image_path": "6c4d0b9775624bf7a1a6b9a5ef25db06770bbe71feb83e35b9bf17ef8ff3739c.jpg" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 130, + 352, + 479, + 375 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 352, + 479, + 375 + ], + "spans": [ + { + "bbox": [ + 130, + 352, + 479, + 375 + ], + "type": "text", + "content": "By this way, we discretize the motion " + }, + { + "bbox": [ + 130, + 352, + 479, + 375 + ], + "type": "inline_equation", + "content": "M" + }, + { + "bbox": [ + 130, + 352, + 479, + 375 + ], + "type": "text", + "content": " into " + }, + { + "bbox": [ + 130, + 352, + 479, + 375 + ], + "type": "inline_equation", + "content": "S" + }, + { + "bbox": [ + 130, + 352, + 479, + 375 + ], + "type": "text", + "content": " discretized part motion representations " + }, + { + "bbox": [ + 130, + 352, + 479, + 375 + ], + "type": "inline_equation", + "content": "\\{Q^i\\}" + }, + { + "bbox": [ + 130, + 352, + 479, + 375 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 130, + 352, + 479, + 375 + ], + "type": "inline_equation", + "content": "i = 1, \\dots, S" + }, + { + "bbox": [ + 130, + 352, + 479, + 375 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 130, + 376, + 480, + 413 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 376, + 480, + 413 + ], + "spans": [ + { + "bbox": [ + 130, + 376, + 480, + 413 + ], + "type": "text", + "content": "For training the VQ-VAE, We use a decoder to reconstruct the " + }, + { + "bbox": [ + 130, + 376, + 480, + 413 + ], + "type": "inline_equation", + "content": "i" + }, + { + "bbox": [ + 130, + 376, + 480, + 413 + ], + "type": "text", + "content": "-th part motion " + }, + { + "bbox": [ + 130, + 376, + 480, + 413 + ], + "type": "inline_equation", + "content": "\\hat{P}^i = \\text{Decoder}^i(Q^i)" + }, + { + "bbox": [ + 130, + 376, + 480, + 413 + ], + "type": "text", + "content": " with reconstruction loss " + }, + { + "bbox": [ + 130, + 376, + 480, + 413 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_r^i = \\| \\hat{P}^i - P^i\\|" + }, + { + "bbox": [ + 130, + 376, + 480, + 413 + ], + "type": "text", + "content": ". The optimization objective of the " + }, + { + "bbox": [ + 130, + 376, + 480, + 413 + ], + "type": "inline_equation", + "content": "i" + }, + { + "bbox": [ + 130, + 376, + 480, + 413 + ], + "type": "text", + "content": "-th part's VQ-VAE is:" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 211, + 423, + 480, + 437 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 211, + 423, + 480, + 437 + ], + "spans": [ + { + "bbox": [ + 211, + 423, + 480, + 437 + ], + "type": "interline_equation", + "content": "\\mathcal {L} ^ {i} = \\mathcal {L} _ {r} ^ {i} + \\| s g (E ^ {i}) - Q ^ {i} \\| + \\beta \\| E ^ {i} - s g (Q ^ {i}) \\|, \\tag {2}", + "image_path": "65d212ffc95912c110a6dab0cf50ad5b8c7531d71b59dc0c6c50495c146efd3d.jpg" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 130, + 445, + 480, + 517 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 445, + 480, + 517 + ], + "spans": [ + { + "bbox": [ + 130, + 445, + 480, + 517 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 130, + 445, + 480, + 517 + ], + "type": "inline_equation", + "content": "sg" + }, + { + "bbox": [ + 130, + 445, + 480, + 517 + ], + "type": "text", + "content": " represents the stop-gradient operation. The first term is the reconstruction loss function, ensuring that the VQ-VAE can reconstruct the original part motion from the encoding. The second term is the codebook loss function. And the third term is the commitment loss, aiming to make the representation " + }, + { + "bbox": [ + 130, + 445, + 480, + 517 + ], + "type": "inline_equation", + "content": "e_l^i" + }, + { + "bbox": [ + 130, + 445, + 480, + 517 + ], + "type": "text", + "content": " output by the encoder as close as possible to the code " + }, + { + "bbox": [ + 130, + 445, + 480, + 517 + ], + "type": "inline_equation", + "content": "v_{k_l}^i" + }, + { + "bbox": [ + 130, + 445, + 480, + 517 + ], + "type": "text", + "content": " contained in the codebook. The weight of this loss is controlled by the hyperparameter " + }, + { + "bbox": [ + 130, + 445, + 480, + 517 + ], + "type": "inline_equation", + "content": "\\beta" + }, + { + "bbox": [ + 130, + 445, + 480, + 517 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 131, + 536, + 317, + 548 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 131, + 536, + 317, + 548 + ], + "spans": [ + { + "bbox": [ + 131, + 536, + 317, + 548 + ], + "type": "text", + "content": "3.2 Text-Driven Part Coordination" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 130, + 558, + 480, + 665 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 558, + 480, + 665 + ], + "spans": [ + { + "bbox": [ + 130, + 558, + 480, + 665 + ], + "type": "text", + "content": "In this stage, we employ transformers as generators to achieve text-to-motion generation. Diverging from the approach of using a large transformer for the entire whole-body motion, we utilize multiple small transformers to generate each part motion code sequence obtained from the previous stage. This allows small transformers to be aware of the conception of \"part motion\" constructed before. However, relying solely on these separate part motion generators can lead to an inability to collectively generate whole-body motions, due to a lack of knowledge about the motions of other parts. To address this, we introduce the Part Coordination module, facilitating communication among all part transformer" + } + ] + } + ], + "index": 12 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "spans": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "type": "text", + "content": "6" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 166, + 91, + 220, + 102 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 166, + 91, + 220, + 102 + ], + "spans": [ + { + "bbox": [ + 166, + 91, + 220, + 102 + ], + "type": "text", + "content": "Q. Zou et al." + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 5 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 184, + 113, + 430, + 259 + ], + "blocks": [ + { + "bbox": [ + 184, + 113, + 430, + 259 + ], + "lines": [ + { + "bbox": [ + 184, + 113, + 430, + 259 + ], + "spans": [ + { + "bbox": [ + 184, + 113, + 430, + 259 + ], + "type": "image", + "image_path": "4bb813629d5341c3a151e237d89aeecfe3de8849cf4f0a68dcf05eabf5c7a15a.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 178, + 266, + 434, + 278 + ], + "lines": [ + { + "bbox": [ + 178, + 266, + 434, + 278 + ], + "spans": [ + { + "bbox": [ + 178, + 266, + 434, + 278 + ], + "type": "text", + "content": "Fig. 4: The architecture of our Part-Coordinated Transformer." + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_caption" + } + ], + "index": 2 + }, + { + "bbox": [ + 130, + 309, + 479, + 346 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 309, + 479, + 346 + ], + "spans": [ + { + "bbox": [ + 130, + 309, + 479, + 346 + ], + "type": "text", + "content": "generators to collaboratively generate whole-body motion. This capability enables our ParCo to handle textual inputs involving different parts, generating human motion that aligns with semantic descriptions in the text." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 130, + 349, + 480, + 421 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 349, + 480, + 421 + ], + "spans": [ + { + "bbox": [ + 130, + 349, + 480, + 421 + ], + "type": "text", + "content": "As depicted in Fig. 3 (b), we employ six transformers to generate code index sequences of part motions based on the input text. At the same time, these six transformers collaborate through our Part Coordination module to coordinate with each other. Subsequently, these generated index sequences are decoded into the original representation of part motions by the decoder of the VQ-VAE. These part motions are integrated to form whole-body motion." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 130, + 425, + 480, + 496 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 425, + 480, + 496 + ], + "spans": [ + { + "bbox": [ + 130, + 425, + 480, + 496 + ], + "type": "text", + "content": "Specifically, we model the whole text-to-motion generation process as estimating the distribution " + }, + { + "bbox": [ + 130, + 425, + 480, + 496 + ], + "type": "inline_equation", + "content": "p(M|t)" + }, + { + "bbox": [ + 130, + 425, + 480, + 496 + ], + "type": "text", + "content": " of motion " + }, + { + "bbox": [ + 130, + 425, + 480, + 496 + ], + "type": "inline_equation", + "content": "M" + }, + { + "bbox": [ + 130, + 425, + 480, + 496 + ], + "type": "text", + "content": " given the text " + }, + { + "bbox": [ + 130, + 425, + 480, + 496 + ], + "type": "inline_equation", + "content": "t" + }, + { + "bbox": [ + 130, + 425, + 480, + 496 + ], + "type": "text", + "content": ". Since we have discretized the whole-body motion into part motions, we can model the entire motion distribution " + }, + { + "bbox": [ + 130, + 425, + 480, + 496 + ], + "type": "inline_equation", + "content": "p(M|t)" + }, + { + "bbox": [ + 130, + 425, + 480, + 496 + ], + "type": "text", + "content": " through the estimation of conditional distributions " + }, + { + "bbox": [ + 130, + 425, + 480, + 496 + ], + "type": "inline_equation", + "content": "p(K^i |t)" + }, + { + "bbox": [ + 130, + 425, + 480, + 496 + ], + "type": "text", + "content": " for each part, where " + }, + { + "bbox": [ + 130, + 425, + 480, + 496 + ], + "type": "inline_equation", + "content": "\\{K^i = [k_1^i,\\dots,k_L^i]\\} , i\\in [1,\\dots,S]" + }, + { + "bbox": [ + 130, + 425, + 480, + 496 + ], + "type": "text", + "content": " is the code indices obtained in the first stage." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 130, + 499, + 479, + 524 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 499, + 479, + 524 + ], + "spans": [ + { + "bbox": [ + 130, + 499, + 479, + 524 + ], + "type": "text", + "content": "To model " + }, + { + "bbox": [ + 130, + 499, + 479, + 524 + ], + "type": "inline_equation", + "content": "i" + }, + { + "bbox": [ + 130, + 499, + 479, + 524 + ], + "type": "text", + "content": "-th part motion's distribution, we propose an autoregressive distribution, which allows parts to coordinate with each other," + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 209, + 540, + 480, + 590 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 209, + 540, + 480, + 590 + ], + "spans": [ + { + "bbox": [ + 209, + 540, + 480, + 590 + ], + "type": "interline_equation", + "content": "\\begin{array}{l} p \\left(K ^ {i} \\mid t\\right) = \\prod_ {h = 1} ^ {L} p \\left(k _ {h} ^ {i} \\mid k _ {1} ^ {i}, o _ {1} ^ {i}; \\dots ; k _ {h - 1} ^ {i}, o _ {h - 1} ^ {i}; t\\right), \\tag {3} \\\\ o _ {*} ^ {i} = \\left\\{k _ {*} ^ {j} \\right\\}, j \\neq i, j \\in [ 1,..., S ]. \\\\ \\end{array}", + "image_path": "fbc4b5895e72f9983d23d02fde1282f4f5001844172a0cd3c11d2c058456c671.jpg" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 130, + 605, + 480, + 666 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 605, + 480, + 666 + ], + "spans": [ + { + "bbox": [ + 130, + 605, + 480, + 666 + ], + "type": "text", + "content": "When predicting token " + }, + { + "bbox": [ + 130, + 605, + 480, + 666 + ], + "type": "inline_equation", + "content": "k_h^i" + }, + { + "bbox": [ + 130, + 605, + 480, + 666 + ], + "type": "text", + "content": " for the " + }, + { + "bbox": [ + 130, + 605, + 480, + 666 + ], + "type": "inline_equation", + "content": "i" + }, + { + "bbox": [ + 130, + 605, + 480, + 666 + ], + "type": "text", + "content": "-th part, the prediction not only relies on all tokens predicted by itself from time 1 to " + }, + { + "bbox": [ + 130, + 605, + 480, + 666 + ], + "type": "inline_equation", + "content": "h - 1" + }, + { + "bbox": [ + 130, + 605, + 480, + 666 + ], + "type": "text", + "content": " but is also conditioned by predictions from all other parts during the same time span. Also, after the " + }, + { + "bbox": [ + 130, + 605, + 480, + 666 + ], + "type": "inline_equation", + "content": "i" + }, + { + "bbox": [ + 130, + 605, + 480, + 666 + ], + "type": "text", + "content": "-th part transformer predicts the token " + }, + { + "bbox": [ + 130, + 605, + 480, + 666 + ], + "type": "inline_equation", + "content": "K_h^i" + }, + { + "bbox": [ + 130, + 605, + 480, + 666 + ], + "type": "text", + "content": ", the prediction will be used by all part generators to predict the next token." + } + ] + } + ], + "index": 9 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 235, + 91, + 448, + 103 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 235, + 91, + 448, + 103 + ], + "spans": [ + { + "bbox": [ + 235, + 91, + 448, + 103 + ], + "type": "text", + "content": "ParCo: Part-Coordinating Text-to-Motion Synthesis" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 474, + 91, + 480, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 474, + 91, + 480, + 100 + ], + "spans": [ + { + "bbox": [ + 474, + 91, + 480, + 100 + ], + "type": "text", + "content": "7" + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 6 + }, + { + "para_blocks": [ + { + "bbox": [ + 130, + 116, + 479, + 139 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 116, + 479, + 139 + ], + "spans": [ + { + "bbox": [ + 130, + 116, + 479, + 139 + ], + "type": "text", + "content": "Finally, we learn the entire body motion by estimating the distributions of all part motions," + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 232, + 140, + 481, + 188 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 232, + 140, + 481, + 188 + ], + "spans": [ + { + "bbox": [ + 232, + 140, + 481, + 188 + ], + "type": "interline_equation", + "content": "\\begin{array}{l} \\mathcal {L} = \\mathbb {E} _ {M, t \\sim p (M, t)} [ - \\log p (M | t) ] \\\\ = \\mathbb {E} _ {M, t \\sim p (M, t)} \\left[ - \\sum_ {i = 1} ^ {S} \\log p \\left(K ^ {i} | t\\right) \\right]. \\tag {4} \\\\ \\end{array}", + "image_path": "be7f621777fb9fdc48e7de665542898cae60b9ed50736f6aca20f23c88089106.jpg" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 130, + 191, + 480, + 276 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 191, + 480, + 276 + ], + "spans": [ + { + "bbox": [ + 130, + 191, + 480, + 276 + ], + "type": "text", + "content": "We further propose Part-Coordinated Transformer, a transformer capable of coordinating with other transformers, to approximate " + }, + { + "bbox": [ + 130, + 191, + 480, + 276 + ], + "type": "inline_equation", + "content": "p(K^i |t)" + }, + { + "bbox": [ + 130, + 191, + 480, + 276 + ], + "type": "text", + "content": ". As shown in Fig. 4, we insert a Part Coordination Layer before each transformer layer (except for the first transformer layer). For each token " + }, + { + "bbox": [ + 130, + 191, + 480, + 276 + ], + "type": "inline_equation", + "content": "x^i" + }, + { + "bbox": [ + 130, + 191, + 480, + 276 + ], + "type": "text", + "content": " output by the previous transformer layer, it passes through our ParCo Block. The ParCo Block coordinates with other part motion generators, fusing current token " + }, + { + "bbox": [ + 130, + 191, + 480, + 276 + ], + "type": "inline_equation", + "content": "x^i" + }, + { + "bbox": [ + 130, + 191, + 480, + 276 + ], + "type": "text", + "content": " with tokens from other part transformers," + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 231, + 285, + 481, + 316 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 231, + 285, + 481, + 316 + ], + "spans": [ + { + "bbox": [ + 231, + 285, + 481, + 316 + ], + "type": "interline_equation", + "content": "\\begin{array}{l} x _ {\\text {c o o r d}} ^ {i} = L N \\left(x ^ {i} + M L P ^ {i} (y)\\right), \\tag {5} \\\\ y = \\left\\{x ^ {j} \\right\\}, j \\neq i, j \\in [ 1, \\dots , S ], \\\\ \\end{array}", + "image_path": "85b6533ca2f84180e3e19dbf1328306c10f2a14c8cbefb9da194440f41e00ced.jpg" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 130, + 323, + 479, + 359 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 323, + 479, + 359 + ], + "spans": [ + { + "bbox": [ + 130, + 323, + 479, + 359 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 130, + 323, + 479, + 359 + ], + "type": "inline_equation", + "content": "LN" + }, + { + "bbox": [ + 130, + 323, + 479, + 359 + ], + "type": "text", + "content": " denotes the LayerNorm operation, and " + }, + { + "bbox": [ + 130, + 323, + 479, + 359 + ], + "type": "inline_equation", + "content": "y" + }, + { + "bbox": [ + 130, + 323, + 479, + 359 + ], + "type": "text", + "content": " represents the tokens from other transformers' layers. The fused token " + }, + { + "bbox": [ + 130, + 323, + 479, + 359 + ], + "type": "inline_equation", + "content": "x_{coord}^{i}" + }, + { + "bbox": [ + 130, + 323, + 479, + 359 + ], + "type": "text", + "content": " is then input into the subsequent transformer layer." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 132, + 377, + 214, + 388 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 377, + 214, + 388 + ], + "spans": [ + { + "bbox": [ + 132, + 377, + 214, + 388 + ], + "type": "text", + "content": "3.3 Discussion" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 130, + 397, + 480, + 480 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 397, + 480, + 480 + ], + "spans": [ + { + "bbox": [ + 130, + 397, + 480, + 480 + ], + "type": "text", + "content": "Text-to-motion synthesis intrinsically consists of two stages, i.e., body part representation and generation, to fulfill actions described by texts. To precisely ground each word to the corresponding body part, Balando [57] partitions the body into upper and lower parts for motion quantization while reconstructing through a shared decoder. SCA [13] steps further to equip independent generators for flexible correspondence. However, coarse-grained sub-body level modeling yields sub-optimal results." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 130, + 481, + 480, + 577 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 481, + 480, + 577 + ], + "spans": [ + { + "bbox": [ + 130, + 481, + 480, + 577 + ], + "type": "text", + "content": "AttT2M [71] introduces a global-local attention mechanism to learn hierarchical body-part semantics for accurate motion synthesis. SINC [4] adopts a simple additive composition of part motion for GPT-guided synthetic training data creation. In contrast, we advocate explicitly discretizing body parts as individual action atoms and synthesis motion with decentralized generators and a centralized part coordinating module. With a series of computation-economic designs, we keep the relative independence yet close coordination relationships of body parts and report superior results." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 132, + 596, + 224, + 609 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 596, + 224, + 609 + ], + "spans": [ + { + "bbox": [ + 132, + 596, + 224, + 609 + ], + "type": "text", + "content": "4 Experiment" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 132, + 621, + 202, + 633 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 621, + 202, + 633 + ], + "spans": [ + { + "bbox": [ + 132, + 621, + 202, + 633 + ], + "type": "text", + "content": "4.1 Settings" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 130, + 641, + 479, + 665 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 641, + 479, + 665 + ], + "spans": [ + { + "bbox": [ + 130, + 641, + 479, + 665 + ], + "type": "text", + "content": "Datasets. We utilized two widely used text-to-motion datasets, KIT-ML and HumanML3D, for training and testing our method. Comparative evaluations" + } + ] + } + ], + "index": 12 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "spans": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "type": "text", + "content": "8" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 166, + 91, + 220, + 102 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 166, + 91, + 220, + 102 + ], + "spans": [ + { + "bbox": [ + 166, + 91, + 220, + 102 + ], + "type": "text", + "content": "Q. Zou et al." + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 7 + }, + { + "para_blocks": [ + { + "type": "table", + "bbox": [ + 133, + 190, + 482, + 384 + ], + "blocks": [ + { + "bbox": [ + 130, + 114, + 482, + 181 + ], + "lines": [ + { + "bbox": [ + 130, + 114, + 482, + 181 + ], + "spans": [ + { + "bbox": [ + 130, + 114, + 482, + 181 + ], + "type": "text", + "content": "Table 1: Comparisons to current state-of-the-art methods on HumanML3D test set. “↑” denotes that higher is better. “↓” denotes that lower is better. “→” denotes that results are better if the metric is closer to the real motion. Bold and underlined indicate the best and second-best results, respectively. § reports results using ground-truth motion length. The results of ReMoDiffuse* are obtained from official checkpoints and employ uniform random sampling of motion lengths as input." + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 133, + 190, + 482, + 384 + ], + "lines": [ + { + "bbox": [ + 133, + 190, + 482, + 384 + ], + "spans": [ + { + "bbox": [ + 133, + 190, + 482, + 384 + ], + "type": "table", + "html": "
MethodsR-Precision ↑FID ↓MM-Dist ↓Diversity → MModality ↑
Top-1Top-2Top-3
Real motion0.511±.0030.703±.0030.797±.0020.002±.0002.974±.0089.503±.065-
MDM§ [61]0.320±.0050.498±.0040.611±.0070.544±.0445.566±.0279.559±.0862.799±.072
MLD§ [8]0.481±.0030.673±.0030.772±.0020.473±.0133.196±.0109.724±.0822.413±.079
MotionDiffuse§ [67]0.491±.0010.681±.0010.782±.0010.630±.0013.113±.0019.410±.0491.553±.042
ReMoDiffuse§ [68]0.510±.0050.698±.0060.795±.0040.103±.0042.974±.0169.018±.0751.795±.043
ReMoDiffuse*0.450±.0030.638±.0020.743±.0030.281±.0103.271±.0089.236±.085-
Text2Gesture [6]0.165±.0010.267±.0020.345±.0027.664±.0306.030±.0086.409±.071-
Seq2Seq [48]0.180±.0020.300±.0020.396±.00211.75±.0355.529±.0076.223±.061-
Language2Pose [2]0.246±.0010.387±.0020.486±.00211.02±.0465.296±.0087.676±.058-
Hier [13]0.301±.0020.425±.0020.552±.0046.532±.0245.012±.0188.332±.042-
TEMOS [46]0.424±.0020.612±.0020.722±.0023.734±.0283.703±.0088.973±.0710.368±.018
TM2T [15]0.424±.0030.618±.0030.729±.0021.501±.0173.467±.0118.589±.0762.424±.093
T2M [14]0.457±.0020.639±.0030.740±.0031.067±.0023.340±.0089.188±.0022.090±.083
T2M-GPT [66]0.492±.0030.679±.0020.775±.0020.141±.0053.121±.0099.722±.0821.831±.048
Fg-T2M [63]0.492±.0020.683±.0030.783±.0020.243±.0193.109±.0079.278±.0721.614±.049
AttT2M [71]0.499±.0030.690±.0020.786±.0020.112±.0063.038±.0079.700±.0902.452±.051
ParCo (Ours)0.515±.0030.706±.0030.801±.0020.109±.0052.927±.0089.576±.0881.382±.060
", + "image_path": "ef99595b7da82840c78fd880f20ea8cd69ae250fe7877652205328b0d2a621a9.jpg" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "table_body" + } + ], + "index": 3 + }, + { + "bbox": [ + 130, + 417, + 482, + 596 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 417, + 482, + 596 + ], + "spans": [ + { + "bbox": [ + 130, + 417, + 482, + 596 + ], + "type": "text", + "content": "were conducted on these datasets against other existing methods. Processed from KIT [47] and CMU [29] datasets, KIT-ML Dataset [47] comprises 3,911 sequences of human body motions with 6,278 text annotations. Each motion is annotated with 1 to 4 text descriptions, averaging approximately 8 words each. KIT-ML employs the MMM [59] skeletal model which has 21 joints, and we detail in the supplementary materials how we partition its joints into 6 parts. We follow the train, validation, and test set divisions as outlined in [14] and report our ParCo's performance on the test set. HumanML3D [14] Dataset is the largest dataset with 14,616 3D human body motion sequences and 44,970 corresponding textual descriptions, sourced from AMASS [36] and HumanAct12 [16]. Each motion includes a minimum of 3 text descriptions, with an average length of 12 words. HumanML3D adopts the SMPL [34] skeletal model which has 22 joints, and we detail in the supplementary materials how we partition its joints into 6 parts. Similar to KIT-ML, we follow the division into train, validation, and test sets as specified in [14] and report our ParCo's performance on the test set." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 130, + 617, + 482, + 666 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 617, + 482, + 666 + ], + "spans": [ + { + "bbox": [ + 130, + 617, + 482, + 666 + ], + "type": "text", + "content": "Evaluation Metrics. Following prior text-to-motion work, we leverage pretrained text and motion feature extractors [14] to measure cross-modal alignment and semantic similarity between motions, rather than joint coordinate position. Based on the extracted features, we employ the following evaluation metrics: (i)" + } + ] + } + ], + "index": 5 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 235, + 91, + 447, + 102 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 235, + 91, + 447, + 102 + ], + "spans": [ + { + "bbox": [ + 235, + 91, + 447, + 102 + ], + "type": "text", + "content": "ParCo: Part-Coordinating Text-to-Motion Synthesis" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 474, + 92, + 480, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 474, + 92, + 480, + 100 + ], + "spans": [ + { + "bbox": [ + 474, + 92, + 480, + 100 + ], + "type": "text", + "content": "9" + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 8 + }, + { + "para_blocks": [ + { + "type": "table", + "bbox": [ + 134, + 136, + 482, + 330 + ], + "blocks": [ + { + "bbox": [ + 144, + 114, + 468, + 125 + ], + "lines": [ + { + "bbox": [ + 144, + 114, + 468, + 125 + ], + "spans": [ + { + "bbox": [ + 144, + 114, + 468, + 125 + ], + "type": "text", + "content": "Table 2: Comparisons to current state-of-the-art methods on KIT-ML test set." + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 134, + 136, + 482, + 330 + ], + "lines": [ + { + "bbox": [ + 134, + 136, + 482, + 330 + ], + "spans": [ + { + "bbox": [ + 134, + 136, + 482, + 330 + ], + "type": "table", + "html": "
MethodsR-Precision ↑FID ↓MM-Dist ↓Diversity → MModality ↑
Top-1Top-2Top-3
Real motion0.424±.0050.649±.0060.779±.0060.031±.0042.788±.01211.08±.097-
MDM§ [61]0.164±.0040.291±.0040.396±.0040.497±.0219.191±.02210.85±.1091.907±.214
MLD§ [8]0.390±.0080.609±.0080.734±.0070.404±.0273.204±.02710.80±.1172.192±.071
MotionDiffuse§ [67]0.417±.0040.621±.0040.739±.0041.954±.0622.958±.00511.10±.1430.730±.013
ReMoDiffuse§ [68]0.427±.0140.641±.0040.765±.0550.155±.0062.814±.01210.80±.1051.239±.028
ReMoDiffuse*0.382±.0050.586±.0070.706±.0060.589±.0223.324±.03010.31±.065-
Seq2Seq [48]0.103±.0030.178±.0050.241±.00624.86±.3487.960±.0316.744±.106-
Text2Gesture [6]0.156±.0040.255±.0040.338±.00512.12±.1836.946±.0299.334±.079-
Language2Pose [2]0.221±.0050.373±.0040.483±.0056.545±.0725.147±.0309.073±.100-
Hier [13]0.255±.0060.432±.0070.531±.0075.203±.1074.986±.0279.563±.0722.090±.083
TM2T [15]0.280±.0050.463±.0060.587±.0053.599±.1534.591±.0269.473±.1173.292±.081
TEMOS [46]0.353±.0060.561±.0070.687±.0053.717±.0513.417±.01910.84±.1000.532±.034
T2M [14]0.370±.0050.569±.0070.693±.0072.770±.1093.401±.00810.91±.1191.482±.065
AttT2M [71]0.413±.0060.632±.0060.751±.0060.870±.0393.039±.02110.96±.1232.281±.047
T2M-GPT [66]0.416±.0060.627±.0060.745±.0060.514±.0293.007±.02310.92±.1081.570±.039
Fg-T2M [63]0.418±.0050.626±.0040.745±.0040.571±.0473.114±.01510.93±.0831.019±.029
ParCo (Ours)0.430±.0040.649±.0070.772±.0060.453±.0272.820±.02810.95±.0941.245±.022
", + "image_path": "cec3130bfd6a896ad8d638e15dc3c9e6bb6243d6123f4b2c2b07843933587db0.jpg" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "table_body" + } + ], + "index": 3 + }, + { + "bbox": [ + 130, + 361, + 482, + 577 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 361, + 482, + 577 + ], + "spans": [ + { + "bbox": [ + 130, + 361, + 482, + 577 + ], + "type": "text", + "content": "R-Precision: Reflects the accurate semantic matching between text and motion. We calculate Top-1, Top-2, and Top-3 accuracy based on the Euclidean distance between a given motion sequence and 32 text descriptions (1 ground truth and 31 randomly selected non-matching). (ii) FID: We use FID [21] to quantify the distributional disparity between generated and real motions based on extracted motion features. It's crucial to highlight that FID does not assess the alignment between textual descriptions and generated motions. (iii) MM-Dist: Measures the Euclidean distance between feature vectors of text and motion, reflecting the semantic similarity. (iv) Diversity: Indicates the variance in generated motions. We randomly sample two equal-sized subsets from all motions, calculating the average Euclidean distance between the subsets. Closer diversity values between generated and real motions signify a better match. (v) MModality: Reflects the diversity of generated motions for a given text. We generate 10 pairs of motions for each text, compute feature vectors' distance between each pair, and take the average. Given that incorrectly generated motions lead to a high MModality value, it fails to reflect the alignment between motions and texts. In addition, following [61], we run each evaluation 20 times (except MModality for 5 times) and report the average with a " + }, + { + "bbox": [ + 130, + 361, + 482, + 577 + ], + "type": "inline_equation", + "content": "95\\%" + }, + { + "bbox": [ + 130, + 361, + 482, + 577 + ], + "type": "text", + "content": " confidence interval." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 130, + 593, + 482, + 666 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 593, + 482, + 666 + ], + "spans": [ + { + "bbox": [ + 130, + 593, + 482, + 666 + ], + "type": "text", + "content": "Implementation Details. Our ParCo employs 6 small VQ-VAEs for discretizing part motions and 6 small transformers with Part Coordination modules for text-to-motion generation. For VQ-VAEs, all codebooks contain 512 codes, and all parts have 128 code dimensions, except for the Root part, which has a code dimension of 64. The encoder's downsampling rate is set to " + }, + { + "bbox": [ + 130, + 593, + 482, + 666 + ], + "type": "inline_equation", + "content": "r = 4" + }, + { + "bbox": [ + 130, + 593, + 482, + 666 + ], + "type": "text", + "content": ". For transformers, each has 14 layers, and token dimension is 256. And we insert a Part Coordi" + } + ] + } + ], + "index": 5 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "spans": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "type": "text", + "content": "10" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 166, + 91, + 219, + 102 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 166, + 91, + 219, + 102 + ], + "spans": [ + { + "bbox": [ + 166, + 91, + 219, + 102 + ], + "type": "text", + "content": "Q. Zou et al." + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 9 + }, + { + "para_blocks": [ + { + "type": "table", + "bbox": [ + 133, + 114, + 481, + 313 + ], + "blocks": [ + { + "bbox": [ + 133, + 114, + 481, + 313 + ], + "lines": [ + { + "bbox": [ + 133, + 114, + 481, + 313 + ], + "spans": [ + { + "bbox": [ + 133, + 114, + 481, + 313 + ], + "type": "table", + "html": "
"a man steps back, picks something up and put it to his head and then puts it back.""a person walks forward one foot in front of the other until he loses his balance and tilts far to the left then stumbles."
MDMX no steps backX pick upX put to headX no put it backX no walks forward one foot in front of the otherX dance instead of stumble
Motion DiffuseX no steps backX raise hand instead of pickX put to headX puts it backX no walks forward one foot in front of the otherX lose balance
T2M-GPTX no steps backX swing arm instead of pickX put to headX puts it backX tilts to the right instead of leftX no stumble
OursX steps backX picks upX put to headX puts it backX walks forward one foot in front of the otherX lose balance
X tilts to the right instead of leftX no stumble
X walks forward one foot in front of the otherX lose balance
X tilts far to the leftX stumble
", + "image_path": "e4a37b2aec97ffbee3636a9b047d94e2a790a09e57455d4ed8687b236e2a4c9e.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "table_body" + } + ], + "index": 2 + }, + { + "bbox": [ + 130, + 322, + 482, + 355 + ], + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 322, + 482, + 355 + ], + "spans": [ + { + "bbox": [ + 130, + 322, + 482, + 355 + ], + "type": "text", + "content": "Fig.5: Qualitative comparison with existing methods. Green indicates the motion is consistent with the text description. Red indicates the text description lacks the corresponding motion or got the wrong motion." + } + ] + } + ], + "index": 3, + "type": "text" + }, + { + "bbox": [ + 130, + 378, + 482, + 475 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 378, + 482, + 475 + ], + "spans": [ + { + "bbox": [ + 130, + 378, + 482, + 475 + ], + "type": "text", + "content": "nation Layer before all remaining layers except for the first transformer layer. ParCo Blocks in the same layer of the same transformer share parameter weights. We set the number of MLP layers in ParCo Block to 3. For training VQ-VAE, we use a learning rate of 2e-4 before 200K and 1e-5 after 200K, AdamW [35] optimizer with " + }, + { + "bbox": [ + 130, + 378, + 482, + 475 + ], + "type": "inline_equation", + "content": "beta_{1} = 0.9" + }, + { + "bbox": [ + 130, + 378, + 482, + 475 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 130, + 378, + 482, + 475 + ], + "type": "inline_equation", + "content": "beta_{2} = 0.99" + }, + { + "bbox": [ + 130, + 378, + 482, + 475 + ], + "type": "text", + "content": ", and batch size of 256. The commitment loss weight " + }, + { + "bbox": [ + 130, + 378, + 482, + 475 + ], + "type": "inline_equation", + "content": "\\beta" + }, + { + "bbox": [ + 130, + 378, + 482, + 475 + ], + "type": "text", + "content": " is set to 1.0. For training transformer, we use a learning rate of 1e-4 before 150K and 5e-6 after 150K, AdamW optimizer with " + }, + { + "bbox": [ + 130, + 378, + 482, + 475 + ], + "type": "inline_equation", + "content": "beta_{1} = 0.5" + }, + { + "bbox": [ + 130, + 378, + 482, + 475 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 130, + 378, + 482, + 475 + ], + "type": "inline_equation", + "content": "beta_{2} = 0.99" + }, + { + "bbox": [ + 130, + 378, + 482, + 475 + ], + "type": "text", + "content": ", and batch size of 128." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 132, + 491, + 321, + 503 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 491, + 321, + 503 + ], + "spans": [ + { + "bbox": [ + 132, + 491, + 321, + 503 + ], + "type": "text", + "content": "4.2 Comparisons to State-of-the-art" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 130, + 510, + 482, + 666 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 510, + 482, + 666 + ], + "spans": [ + { + "bbox": [ + 130, + 510, + 482, + 666 + ], + "type": "text", + "content": "We compare our ParCo with other methods (including the methods using ground-truth motion length) on HumanML3D test set (Table. 1) and KIT-ML test set (Table. 2). Our method demonstrates superior performance compared to previous state-of-the-art methods on R-Precision and MM-Dist, and comparable results on FID, indicating ParCo's superiority. As for Top-1, Top-2, and Top-3 of R-Precision, our ParCo surpasses previous SOTA, ReMoDiffuse [68] (using GT motion length), with 0.005, 0.008, and 0.006 on HumanML3D, 0.003, 0.008, and 0.007 on KIT-ML. And for MM-Dist, our ParCo exceeds ReMoDiffuse's performance with 0.047 on HumanML3D and has comparable result on KIT-ML. In terms of FID, our result is on par with ReMoDiffuse on HumanML3D, and achieve the SOTA performance on HumanML3D and KIT-ML compared to the methods not relying GT information. Furthermore, our ParCo yields a lower MModality value compared to the previous state-of-the-art. This may be" + } + ] + } + ], + "index": 6 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 235, + 91, + 448, + 102 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 235, + 91, + 448, + 102 + ], + "spans": [ + { + "bbox": [ + 235, + 91, + 448, + 102 + ], + "type": "text", + "content": "ParCo: Part-Coordinating Text-to-Motion Synthesis" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 470, + 92, + 479, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 470, + 92, + 479, + 100 + ], + "spans": [ + { + "bbox": [ + 470, + 92, + 479, + 100 + ], + "type": "text", + "content": "11" + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 10 + }, + { + "para_blocks": [ + { + "bbox": [ + 130, + 116, + 482, + 236 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 116, + 482, + 236 + ], + "spans": [ + { + "bbox": [ + 130, + 116, + 482, + 236 + ], + "type": "text", + "content": "attributed to our ParCo generating motions that align more accurately with the text, thereby reducing the occurrence of irrelevant or incorrect motions. Qualitative results presented in Fig. 5 demonstrate that our generated motions are more realistic, coordinated, and aligned with textual descriptions for texts involving multiple body parts. Details such as \"steps back,\" \"tilts far to the left,\" and \"stumbles\" are accurately captured by our ParCo, while other methods either ignore or incorrectly generate these nuanced actions. Besides, as illustrated in Fig. 1 and Fig. 5, ParCo is endowed the part-level composition superiority across action-level spatial and temporal compositions respectively. More details are provided in the supplementary materials." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 130, + 255, + 482, + 400 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 255, + 482, + 400 + ], + "spans": [ + { + "bbox": [ + 130, + 255, + 482, + 400 + ], + "type": "text", + "content": "GT Leakage of Diffusion-based Methods Previous studies [8, 61, 67, 68] based on the Diffusion model utilize the ground truth motion length as an input for synthesizing motions during evaluation, contributing to their remarkable FID scores. However, this approach is impractical for real-world applications. For clarity, we replace the GT motion lengths with random lengths sampled uniformly from dataset's motion length range to evaluate ReMoDiff. On HumanML3D (KIT-ML), the ReMoDiff's Top-1, Top-2, and Top-3 R-precision decrease by 0.041 (0.053), 0.040 (0.068), and 0.033 (0.075) respectively, while the FID increases by 0.147 (0.426), justifying our speculation. As an auto-regressive approach, our ParCo demonstrates competitive R-Precision and FID without requiring pre-defined motion lengths Table 1. Given its lower computational and parameter consumption (Table. 4), the superiority of ParCo is convincing." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 132, + 419, + 204, + 431 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 419, + 204, + 431 + ], + "spans": [ + { + "bbox": [ + 132, + 419, + 204, + 431 + ], + "type": "text", + "content": "4.3 Analysis" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 130, + 441, + 482, + 562 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 441, + 482, + 562 + ], + "spans": [ + { + "bbox": [ + 130, + 441, + 482, + 562 + ], + "type": "text", + "content": "Performance on Text Inputs of Different Lengths. In order to investigate the synthetic performance given textual descriptions of different lengths, we sort the HumanML3D test set based on the length of textual descriptions, and split it into four subsets with approximately equal numbers of text-motion pairs: 0-25%, 25-50%, 50-75%, and 75-100%, from short to long. The details of these splits are available in the supplementary materials. We conducted evaluations of real motion, T2M-GPT [66], and our method on these four splits. As illustrated in Fig. 6, our method exhibits comprehensive improvements in all subsets when contrasted with the baseline [66]. It demonstrates that our design is beneficial for text inputs of different lengths without compromise." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 130, + 581, + 482, + 667 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 581, + 482, + 667 + ], + "spans": [ + { + "bbox": [ + 130, + 581, + 482, + 667 + ], + "type": "text", + "content": "Ablation Study. We investigate different discretizations for whole-body motion and conduct ablation of our Part Coordination module. We compare our 6-part partition and upper-lower body partition proposed by SCA [13]. It is noteworthy that, despite SCA dividing whole-body into upper and lower body for motion generation, its generations of upper body motion and lower body motion are entirely independent and lack coordination compared with our method. As demonstrated in Table. 3, the comparison between our ParCo (D) and the" + } + ] + } + ], + "index": 6 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "spans": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "type": "text", + "content": "12" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 166, + 91, + 219, + 102 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 166, + 91, + 219, + 102 + ], + "spans": [ + { + "bbox": [ + 166, + 91, + 219, + 102 + ], + "type": "text", + "content": "Q. Zou et al." + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 11 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 134, + 116, + 241, + 180 + ], + "blocks": [ + { + "bbox": [ + 134, + 116, + 241, + 180 + ], + "lines": [ + { + "bbox": [ + 134, + 116, + 241, + 180 + ], + "spans": [ + { + "bbox": [ + 134, + 116, + 241, + 180 + ], + "type": "image", + "image_path": "6e7c5d6c4ea560f9da8b49f829c29c248037bf744019bf988effa215c1a7c755.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 131, + 192, + 479, + 225 + ], + "lines": [ + { + "bbox": [ + 131, + 192, + 479, + 225 + ], + "spans": [ + { + "bbox": [ + 131, + 192, + 479, + 225 + ], + "type": "text", + "content": "Fig.6: Comparison on 4 HumanML3D test subsets divided based on text length. " + }, + { + "bbox": [ + 131, + 192, + 479, + 225 + ], + "type": "inline_equation", + "content": "0-25\\%" + }, + { + "bbox": [ + 131, + 192, + 479, + 225 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 131, + 192, + 479, + 225 + ], + "type": "inline_equation", + "content": "25-50\\%" + }, + { + "bbox": [ + 131, + 192, + 479, + 225 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 131, + 192, + 479, + 225 + ], + "type": "inline_equation", + "content": "50-75\\%" + }, + { + "bbox": [ + 131, + 192, + 479, + 225 + ], + "type": "text", + "content": ", and " + }, + { + "bbox": [ + 131, + 192, + 479, + 225 + ], + "type": "inline_equation", + "content": "75-100\\%" + }, + { + "bbox": [ + 131, + 192, + 479, + 225 + ], + "type": "text", + "content": " respectively represent four subsets based on text length from short to long." + } + ] + } + ], + "index": 7, + "angle": 0, + "type": "image_caption" + } + ], + "index": 2 + }, + { + "type": "image", + "bbox": [ + 242, + 117, + 332, + 175 + ], + "blocks": [ + { + "bbox": [ + 242, + 117, + 332, + 175 + ], + "lines": [ + { + "bbox": [ + 242, + 117, + 332, + 175 + ], + "spans": [ + { + "bbox": [ + 242, + 117, + 332, + 175 + ], + "type": "image", + "image_path": "186b49529ed66f5fe5eac65f3731e5a63e04ddd574eadaf66a951d4499c2a3c8.jpg" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 330, + 174, + 411, + 181 + ], + "lines": [ + { + "bbox": [ + 330, + 174, + 411, + 181 + ], + "spans": [ + { + "bbox": [ + 330, + 174, + 411, + 181 + ], + "type": "text", + "content": "(b) Performance gap relative to Real" + } + ] + } + ], + "index": 6, + "angle": 0, + "type": "image_caption" + } + ], + "index": 3 + }, + { + "type": "image", + "bbox": [ + 332, + 118, + 404, + 175 + ], + "blocks": [ + { + "bbox": [ + 332, + 118, + 404, + 175 + ], + "lines": [ + { + "bbox": [ + 332, + 118, + 404, + 175 + ], + "spans": [ + { + "bbox": [ + 332, + 118, + 404, + 175 + ], + "type": "image", + "image_path": "233a93902faa6074809a7d02602e9afcb9401fde6871336ff6c95e8aa426397c.jpg" + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "image_body" + } + ], + "index": 4 + }, + { + "type": "image", + "bbox": [ + 404, + 118, + 480, + 175 + ], + "blocks": [ + { + "bbox": [ + 404, + 118, + 480, + 175 + ], + "lines": [ + { + "bbox": [ + 404, + 118, + 480, + 175 + ], + "spans": [ + { + "bbox": [ + 404, + 118, + 480, + 175 + ], + "type": "image", + "image_path": "2e5abcbfb74c6907abf0c3a1c29831975e6eb7086aac0a7cf57ec2fe9b221dc8.jpg" + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "image_body" + } + ], + "index": 5 + }, + { + "type": "table", + "bbox": [ + 134, + 271, + 329, + 327 + ], + "blocks": [ + { + "bbox": [ + 132, + 239, + 329, + 262 + ], + "lines": [ + { + "bbox": [ + 132, + 239, + 329, + 262 + ], + "spans": [ + { + "bbox": [ + 132, + 239, + 329, + 262 + ], + "type": "text", + "content": "Table 3: Ablations of body discretization and part coordination module. * denotes our ParCo." + } + ] + } + ], + "index": 8, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 134, + 271, + 329, + 327 + ], + "lines": [ + { + "bbox": [ + 134, + 271, + 329, + 327 + ], + "spans": [ + { + "bbox": [ + 134, + 271, + 329, + 327 + ], + "type": "table", + "html": "
Discretization Part Coord.Top-1 ↑FID ↓MM-Dist ↓
AUp&LowBody×0.444±.0030.497±.0153.367±.009
B✓0.491±.0030.172±.0073.071±.008
C6 Parts×0.375±.0033.652±.0304.012±.012
*D✓0.515±.0030.109±.0052.927±.008
", + "image_path": "0ac9672bef32bd99899cbf33a936f23d273c7e32d52c2bb982aff5b5076c846f.jpg" + } + ] + } + ], + "index": 9, + "angle": 0, + "type": "table_body" + } + ], + "index": 9 + }, + { + "type": "table", + "bbox": [ + 342, + 270, + 481, + 326 + ], + "blocks": [ + { + "bbox": [ + 339, + 238, + 480, + 260 + ], + "lines": [ + { + "bbox": [ + 339, + 238, + 480, + 260 + ], + "spans": [ + { + "bbox": [ + 339, + 238, + 480, + 260 + ], + "type": "text", + "content": "Table 4: Computational complexity analysis." + } + ] + } + ], + "index": 10, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 342, + 270, + 481, + 326 + ], + "lines": [ + { + "bbox": [ + 342, + 270, + 481, + 326 + ], + "spans": [ + { + "bbox": [ + 342, + 270, + 481, + 326 + ], + "type": "table", + "html": "
MethodParam(M)FLOPs(G)InferTime(s)
ReMoDiff198.2481.00.091
T2M-GPT237.6292.30.544
ParCo168.4211.70.036
", + "image_path": "da710e85a3f9248282f97ebda54a4b421bf67977ab103bc069fcd3b68605b006.jpg" + } + ] + } + ], + "index": 11, + "angle": 0, + "type": "table_body" + } + ], + "index": 11 + }, + { + "bbox": [ + 132, + 351, + 480, + 364 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 351, + 480, + 364 + ], + "spans": [ + { + "bbox": [ + 132, + 351, + 480, + 364 + ], + "type": "text", + "content": "\"a man picks something with his [part description], shakes it, then puts it back.\"" + } + ] + } + ], + "index": 12 + }, + { + "type": "image", + "bbox": [ + 134, + 365, + 302, + 413 + ], + "blocks": [ + { + "bbox": [ + 134, + 365, + 302, + 413 + ], + "lines": [ + { + "bbox": [ + 134, + 365, + 302, + 413 + ], + "spans": [ + { + "bbox": [ + 134, + 365, + 302, + 413 + ], + "type": "image", + "image_path": "ef918190c3df28eacbb9360a9a6e3da3e40fca8d11754530646b6549053f25ce.jpg" + } + ] + } + ], + "index": 13, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 155, + 424, + 457, + 436 + ], + "lines": [ + { + "bbox": [ + 155, + 424, + 457, + 436 + ], + "spans": [ + { + "bbox": [ + 155, + 424, + 457, + 436 + ], + "type": "text", + "content": "Fig. 7: Qualitative result of left-right exchange experiment on our ParCo." + } + ] + } + ], + "index": 15, + "angle": 0, + "type": "image_caption" + } + ], + "index": 13 + }, + { + "type": "image", + "bbox": [ + 318, + 365, + 481, + 413 + ], + "blocks": [ + { + "bbox": [ + 318, + 365, + 481, + 413 + ], + "lines": [ + { + "bbox": [ + 318, + 365, + 481, + 413 + ], + "spans": [ + { + "bbox": [ + 318, + 365, + 481, + 413 + ], + "type": "image", + "image_path": "b95f506c8cf2bb9c7a0b9499a7a80e52579b5c2a7941516f7f6efeca96b350d6.jpg" + } + ] + } + ], + "index": 14, + "angle": 0, + "type": "image_body" + } + ], + "index": 14 + }, + { + "bbox": [ + 130, + 460, + 482, + 556 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 460, + 482, + 556 + ], + "spans": [ + { + "bbox": [ + 130, + 460, + 482, + 556 + ], + "type": "text", + "content": "re-implemented SCA (A) indicates that our method's generated motions exhibit significantly higher performance in Top-1, MM-Dist, and FID than SCA. Furthermore, the contrasts between results (A) and (B), as well as results (C) and (D), underscore the necessity of our ParCo facilitating communication and coordination among different part motions during the generation process. In addition, the comparison between results (B) and (D) validates the effectiveness of our 6-part partitioning, which makes the model aware of the concept of parts and contributes to text-to-motion synthesis." + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 130, + 574, + 482, + 623 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 574, + 482, + 623 + ], + "spans": [ + { + "bbox": [ + 130, + 574, + 482, + 623 + ], + "type": "text", + "content": "Computational complexity analysis. With lightweight architectures and parallel computing support, our ParCo shows superior efficiency from parameters, flops, and inference time, justifying the efficacy of our part discretization and coordination designs, Table. 4." + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 131, + 641, + 481, + 665 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 131, + 641, + 481, + 665 + ], + "spans": [ + { + "bbox": [ + 131, + 641, + 481, + 665 + ], + "type": "text", + "content": "Precise part control. To compare the control of different methods over the movement of human parts, we conducted a left-right exchange experiment on" + } + ] + } + ], + "index": 18 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 235, + 91, + 448, + 102 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 235, + 91, + 448, + 102 + ], + "spans": [ + { + "bbox": [ + 235, + 91, + 448, + 102 + ], + "type": "text", + "content": "ParCo: Part-Coordinating Text-to-Motion Synthesis" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 470, + 91, + 481, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 470, + 91, + 481, + 100 + ], + "spans": [ + { + "bbox": [ + 470, + 91, + 481, + 100 + ], + "type": "text", + "content": "13" + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 12 + }, + { + "para_blocks": [ + { + "type": "table", + "bbox": [ + 176, + 136, + 438, + 198 + ], + "blocks": [ + { + "bbox": [ + 134, + 114, + 478, + 125 + ], + "lines": [ + { + "bbox": [ + 134, + 114, + 478, + 125 + ], + "spans": [ + { + "bbox": [ + 134, + 114, + 478, + 125 + ], + "type": "text", + "content": "Table 5: Evaluation of real motion data on train, val, and test set of HumanML3D." + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 176, + 136, + 438, + 198 + ], + "lines": [ + { + "bbox": [ + 176, + 136, + 438, + 198 + ], + "spans": [ + { + "bbox": [ + 176, + 136, + 438, + 198 + ], + "type": "table", + "html": "
SplitTop-1Top-2Top-3MM-DistDiversity
Train0.628±.0010.810±.0010.888±.0012.388±.0039.685±.083
Val0.513±.0040.703±.0040.800±.0032.911±.0109.575±.081
Test0.512±.0030.703±.0020.797±.0022.973±.0079.495±.079
", + "image_path": "621732f790ba00691afd685f7749ce6a989f42e43e9b7428e40e5ae5c8b901df.jpg" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "table_body" + } + ], + "index": 3 + }, + { + "bbox": [ + 130, + 232, + 481, + 295 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 232, + 481, + 295 + ], + "spans": [ + { + "bbox": [ + 130, + 232, + 481, + 295 + ], + "type": "text", + "content": "ten sentences. We examined several methods for generating actions and accomplishing the left-right swap task, the success rates are listed as follows: " + }, + { + "bbox": [ + 130, + 232, + 481, + 295 + ], + "type": "inline_equation", + "content": "70\\%" + }, + { + "bbox": [ + 130, + 232, + 481, + 295 + ], + "type": "text", + "content": " (ParCo), " + }, + { + "bbox": [ + 130, + 232, + 481, + 295 + ], + "type": "inline_equation", + "content": "50\\%" + }, + { + "bbox": [ + 130, + 232, + 481, + 295 + ], + "type": "text", + "content": " (T2M-GPT), " + }, + { + "bbox": [ + 130, + 232, + 481, + 295 + ], + "type": "inline_equation", + "content": "30\\%" + }, + { + "bbox": [ + 130, + 232, + 481, + 295 + ], + "type": "text", + "content": " (ReMoDiff), " + }, + { + "bbox": [ + 130, + 232, + 481, + 295 + ], + "type": "inline_equation", + "content": "20\\%" + }, + { + "bbox": [ + 130, + 232, + 481, + 295 + ], + "type": "text", + "content": " (MDM), " + }, + { + "bbox": [ + 130, + 232, + 481, + 295 + ], + "type": "inline_equation", + "content": "0\\%" + }, + { + "bbox": [ + 130, + 232, + 481, + 295 + ], + "type": "text", + "content": " (MoDiff). Notably, our ParCo exhibited the highest accuracy, underscoring its proficiency in the awareness of human parts. The qualitative results are showcased in Fig. 7." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 130, + 310, + 482, + 455 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 310, + 482, + 455 + ], + "spans": [ + { + "bbox": [ + 130, + 310, + 482, + 455 + ], + "type": "text", + "content": "Discussion. As illustrated in Table 1 and Table 2, our performance surpasses that of real motion due to the evaluation protocol [14], which uses pretrained feature extractors trained only on the train split. These extractors are more effective at extracting features from motions similar to those in the train split. Table 5 shows that evaluations on real motions from the Train, Val, and Test sets reveal significantly higher Top-1/2/3 and MM-Dist metrics on the Train split. This indicates better extractor understanding of the Train split data. Our model, trained on the train split, generates motions resembling the train data, making feature extraction and matching easier, leading to superior performance on the test set. Due to the lack of a comprehensive metric for text-to-motion semantic alignment, motion generation fidelity, and diversity, we advocate for a more holistic evaluation within the community." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 132, + 472, + 220, + 485 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 472, + 220, + 485 + ], + "spans": [ + { + "bbox": [ + 132, + 472, + 220, + 485 + ], + "type": "text", + "content": "5 Conclusion" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 130, + 498, + 482, + 666 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 498, + 482, + 666 + ], + "spans": [ + { + "bbox": [ + 130, + 498, + 482, + 666 + ], + "type": "text", + "content": "In this study, we focus on enhancing the text-to-motion generation model's ability to comprehend part concepts and facilitate communication between different part motion generators, ultimately yielding the synthesis of coordinated and fined-grained motion. Specifically, we discretize whole-body motion into multiple-part motions to establish the prior concept of parts. Afterward, we employ multiple lightweight generators designed to synthesize different part motions and coordinate them through our part coordination module. Extensive experiments showcase that our method achieves higher consistency between generated motions and textual descriptions compared to previous SOTA methods. Furthermore, in-depth analytical results suggest that our approach excels in achieving more precise part control and has lower computational complexity. More encouragingly, our method exhibits adaptability to various part separation schemes and holds the potential for further refinement toward hierarchical part motion. We anticipate that this will have a far-reaching impact on the community." + } + ] + } + ], + "index": 7 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "spans": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "type": "text", + "content": "14" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 166, + 91, + 220, + 102 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 166, + 91, + 220, + 102 + ], + "spans": [ + { + "bbox": [ + 166, + 91, + 220, + 102 + ], + "type": "text", + "content": "Q. Zou et al." + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 13 + }, + { + "para_blocks": [ + { + "bbox": [ + 133, + 114, + 246, + 129 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 114, + 246, + 129 + ], + "spans": [ + { + "bbox": [ + 133, + 114, + 246, + 129 + ], + "type": "text", + "content": "Acknowledgements" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 132, + 140, + 480, + 165 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 140, + 480, + 165 + ], + "spans": [ + { + "bbox": [ + 132, + 140, + 480, + 165 + ], + "type": "text", + "content": "This work was supported by the National Key R&D Program of China under Grant 2018AAA0102801." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 133, + 185, + 197, + 198 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 185, + 197, + 198 + ], + "spans": [ + { + "bbox": [ + 133, + 185, + 197, + 198 + ], + "type": "text", + "content": "References" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 138, + 211, + 480, + 666 + ], + "type": "list", + "angle": 0, + "index": 23, + "blocks": [ + { + "bbox": [ + 138, + 211, + 480, + 233 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 211, + 480, + 233 + ], + "spans": [ + { + "bbox": [ + 138, + 211, + 480, + 233 + ], + "type": "text", + "content": "1. Ahn, H., Ha, T., Choi, Y., Yoo, H., Oh, S.: Text2action: Generative adversarial synthesis from language to action. In: ICRA (2018)" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 138, + 234, + 480, + 255 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 234, + 480, + 255 + ], + "spans": [ + { + "bbox": [ + 138, + 234, + 480, + 255 + ], + "type": "text", + "content": "2. Ahuja, C., Morency, L.P.: Language2pose: Natural language grounded pose forecasting. In: 3DV (2019)" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 138, + 256, + 480, + 289 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 256, + 480, + 289 + ], + "spans": [ + { + "bbox": [ + 138, + 256, + 480, + 289 + ], + "type": "text", + "content": "3. Antakli, A., Hermann, E., Zinnikus, I., Du, H., Fischer, K.: Intelligent distributed human motion simulation in human-robot collaboration environments. In: ACM IVA (2018)" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 138, + 289, + 480, + 322 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 289, + 480, + 322 + ], + "spans": [ + { + "bbox": [ + 138, + 289, + 480, + 322 + ], + "type": "text", + "content": "4. Athanasiou, N., Petrovich, M., Black, M.J., Varol, G.: Sinc: Spatial composition of 3d human motions for simultaneous action generation. arXiv preprint arXiv:2304.10417 (2023)" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 138, + 322, + 480, + 344 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 322, + 480, + 344 + ], + "spans": [ + { + "bbox": [ + 138, + 322, + 480, + 344 + ], + "type": "text", + "content": "5. Barsoum, E., Kender, J., Liu, Z.: Hp-gan: Probabilistic 3d human motion prediction via gan. In: CVPRW (2018)" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 138, + 344, + 480, + 378 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 344, + 480, + 378 + ], + "spans": [ + { + "bbox": [ + 138, + 344, + 480, + 378 + ], + "type": "text", + "content": "6. Bhattacharya, U., Rewkowski, N., Banerjee, A., Guhan, P., Bera, A., Manocha, D.: Text2gestures: A transformer-based network for generating emotive body gestures for virtual agents. In: 2021 IEEE virtual reality and 3D user interfaces (VR) (2021)" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 138, + 378, + 480, + 399 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 378, + 480, + 399 + ], + "spans": [ + { + "bbox": [ + 138, + 378, + 480, + 399 + ], + "type": "text", + "content": "7. Butepage, J., Black, M.J., Kragic, D., Kjellstrom, H.: Deep representation learning for human motion prediction and classification. In: CVPR (2017)" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 138, + 399, + 480, + 422 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 399, + 480, + 422 + ], + "spans": [ + { + "bbox": [ + 138, + 399, + 480, + 422 + ], + "type": "text", + "content": "8. Chen, X., Jiang, B., Liu, W., Huang, Z., Fu, B., Chen, T., Yu, G.: Executing your commands via motion diffusion in latent space. In: CVPR (2023)" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 138, + 422, + 480, + 455 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 422, + 480, + 455 + ], + "spans": [ + { + "bbox": [ + 138, + 422, + 480, + 455 + ], + "type": "text", + "content": "9. Chen, X., Su, Z., Yang, L., Cheng, P., Xu, L., Fu, B., Yu, G.: Learning variational motion prior for video-based motion capture. arXiv preprint arXiv:2210.15134 (2022)" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 138, + 455, + 480, + 488 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 455, + 480, + 488 + ], + "spans": [ + { + "bbox": [ + 138, + 455, + 480, + 488 + ], + "type": "text", + "content": "0. Djuric, N., Radosavljevic, V., Cui, H., Nguyen, T., Chou, F.C., Lin, T.H., Singh, N., Schneider, J.: Uncertainty-aware short-term motion prediction of traffic actors for autonomous driving. In: WACV (2020)" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 138, + 488, + 480, + 510 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 488, + 480, + 510 + ], + "spans": [ + { + "bbox": [ + 138, + 488, + 480, + 510 + ], + "type": "text", + "content": "1. Duan, Y., Shi, T., Zou, Z., Lin, Y., Qian, Z., Zhang, B., Yuan, Y.: Single-shot motion completion with transformer. arXiv preprint arXiv:2103.00776 (2021)" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 138, + 510, + 480, + 533 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 510, + 480, + 533 + ], + "spans": [ + { + "bbox": [ + 138, + 510, + 480, + 533 + ], + "type": "text", + "content": "2. Fragkiadaki, K., Levine, S., Felsen, P., Malik, J.: Recurrent network models for human dynamics. In: ICCV (2015)" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 138, + 533, + 480, + 555 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 533, + 480, + 555 + ], + "spans": [ + { + "bbox": [ + 138, + 533, + 480, + 555 + ], + "type": "text", + "content": "3. Ghosh, A., Cheema, N., Oguz, C., Theobalt, C., Slusallek, P.: Synthesis of compositional animations from textual descriptions. In: ICCV (2021)" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 138, + 555, + 480, + 576 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 555, + 480, + 576 + ], + "spans": [ + { + "bbox": [ + 138, + 555, + 480, + 576 + ], + "type": "text", + "content": "4. Guo, C., Zou, S., Zuo, X., Wang, S., Ji, W., Li, X., Cheng, L.: Generating diverse and natural 3d human motions from text. In: CVPR (2022)" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 138, + 577, + 480, + 599 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 577, + 480, + 599 + ], + "spans": [ + { + "bbox": [ + 138, + 577, + 480, + 599 + ], + "type": "text", + "content": "5. Guo, C., Zuo, X., Wang, S., Cheng, L.: Tm2t: Stochastic and tokenized modeling for the reciprocal generation of 3d human motions and texts. In: ECCV (2022)" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 138, + 599, + 480, + 621 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 599, + 480, + 621 + ], + "spans": [ + { + "bbox": [ + 138, + 599, + 480, + 621 + ], + "type": "text", + "content": "6. Guo, C., Zuo, X., Wang, S., Zou, S., Sun, Q., Deng, A., Gong, M., Cheng, L.: Action2motion: Conditioned generation of 3d human motions. In: ACM MM (2020)" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 138, + 621, + 480, + 643 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 621, + 480, + 643 + ], + "spans": [ + { + "bbox": [ + 138, + 621, + 480, + 643 + ], + "type": "text", + "content": "7. Harvey, F.G., Pal, C.: Recurrent transition networks for character locomotion. In: SIGGRAPH (2018)" + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 138, + 643, + 480, + 666 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 643, + 480, + 666 + ], + "spans": [ + { + "bbox": [ + 138, + 643, + 480, + 666 + ], + "type": "text", + "content": "8. Harvey, F.G., Yurick, M., Nowrouzezahrai, D., Pal, C.: Robust motion in betweening. ACM TOG (2020)" + } + ] + } + ], + "index": 22 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 235, + 91, + 447, + 102 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 235, + 91, + 447, + 102 + ], + "spans": [ + { + "bbox": [ + 235, + 91, + 447, + 102 + ], + "type": "text", + "content": "ParCo: Part-Coordinating Text-to-Motion Synthesis" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 470, + 92, + 480, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 470, + 92, + 480, + 100 + ], + "spans": [ + { + "bbox": [ + 470, + 92, + 480, + 100 + ], + "type": "text", + "content": "15" + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 14 + }, + { + "para_blocks": [ + { + "bbox": [ + 132, + 117, + 480, + 665 + ], + "type": "list", + "angle": 0, + "index": 26, + "blocks": [ + { + "bbox": [ + 133, + 117, + 480, + 138 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 117, + 480, + 138 + ], + "spans": [ + { + "bbox": [ + 133, + 117, + 480, + 138 + ], + "type": "text", + "content": "19. Herbet, G., Duffau, H.: Revisiting the functional anatomy of the human brain: toward a meta-networking theory of cerebral functions. Physiological Reviews (2020)" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 133, + 139, + 480, + 159 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 139, + 480, + 159 + ], + "spans": [ + { + "bbox": [ + 133, + 139, + 480, + 159 + ], + "type": "text", + "content": "20. Hernandez, A., Gall, J., Moreno-Noguer, F.: Human motion prediction via spatiotemporal inpainting. In: ICCV (2019)" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 132, + 160, + 480, + 191 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 160, + 480, + 191 + ], + "spans": [ + { + "bbox": [ + 132, + 160, + 480, + 191 + ], + "type": "text", + "content": "21. Heusel, M., Ramsauer, H., Unterthiner, T., Nessler, B., Hochreiter, S.: Gans trained by a two time-scale update rule converge to a local nash equilibrium. NeurIPS (2017)" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 132, + 192, + 480, + 214 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 192, + 480, + 214 + ], + "spans": [ + { + "bbox": [ + 132, + 192, + 480, + 214 + ], + "type": "text", + "content": "22. Ho, J., Jain, A., Abbeel, P.: Denoising diffusion probabilistic models. NeurIPS (2020)" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 132, + 214, + 480, + 246 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 214, + 480, + 246 + ], + "spans": [ + { + "bbox": [ + 132, + 214, + 480, + 246 + ], + "type": "text", + "content": "23. Kappel, M., Golyanik, V., Elgharib, M., Henningson, J.O., Seidel, H.P., Castillo, S., Theobalt, C., Magnor, M.: High-fidelity neural human motion transfer from monocular video. In: CVPR (2021)" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 132, + 247, + 480, + 267 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 247, + 480, + 267 + ], + "spans": [ + { + "bbox": [ + 132, + 247, + 480, + 267 + ], + "type": "text", + "content": "24. Kaufmann, M., Aksan, E., Song, J., Pece, F., Ziegler, R., Hilliges, O.: Convolutional autoencoders for human motion infilling. In: 3DV (2020)" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 132, + 268, + 480, + 289 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 268, + 480, + 289 + ], + "spans": [ + { + "bbox": [ + 132, + 268, + 480, + 289 + ], + "type": "text", + "content": "25. Kim, J., Kim, J., Choi, S.: Flame: Free-form language-based motion synthesis & editing. In: AAAI (2023)" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 132, + 289, + 480, + 311 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 289, + 480, + 311 + ], + "spans": [ + { + "bbox": [ + 132, + 289, + 480, + 311 + ], + "type": "text", + "content": "26. Koppula, H., Saxena, A.: Learning spatio-temporal structure from rgb-d videos for human activity detection and anticipation. In: ICML (2013)" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 132, + 312, + 480, + 332 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 312, + 480, + 332 + ], + "spans": [ + { + "bbox": [ + 132, + 312, + 480, + 332 + ], + "type": "text", + "content": "27. Koppula, H.S., Saxena, A.: Anticipating human activities using object affordances for reactive robotic response. IEEE TPAMI (2015)" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 132, + 333, + 367, + 342 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 333, + 367, + 342 + ], + "spans": [ + { + "bbox": [ + 132, + 333, + 367, + 342 + ], + "type": "text", + "content": "28. Kullback, S.: Information theory and statistics (1997)" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 132, + 343, + 399, + 353 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 343, + 399, + 353 + ], + "spans": [ + { + "bbox": [ + 132, + 343, + 399, + 353 + ], + "type": "text", + "content": "29. Lab, C.G.: Cmu graphics lab motion capture database (2016)" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 132, + 354, + 480, + 374 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 354, + 480, + 374 + ], + "spans": [ + { + "bbox": [ + 132, + 354, + 480, + 374 + ], + "type": "text", + "content": "30. Lee, H.Y., Yang, X., Liu, M.Y., Wang, T.C., Lu, Y.D., Yang, M.H., Kautz, J.: Dancing to music. NeurIPS (2019)" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 132, + 375, + 480, + 396 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 375, + 480, + 396 + ], + "spans": [ + { + "bbox": [ + 132, + 375, + 480, + 396 + ], + "type": "text", + "content": "31. Li, B., Zhao, Y., Zhelun, S., Sheng, L.: Danceformer: Music conditioned 3d dance generation with parametric motion transformer. In: AAAI (2022)" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 132, + 396, + 480, + 417 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 396, + 480, + 417 + ], + "spans": [ + { + "bbox": [ + 132, + 396, + 480, + 417 + ], + "type": "text", + "content": "32. Li, R., Yang, S., Ross, D.A., Kanazawa, A.: Ai choreographer: Music conditioned 3d dance generation with aist++. In: ICCV (2021)" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 132, + 418, + 480, + 438 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 418, + 480, + 438 + ], + "spans": [ + { + "bbox": [ + 132, + 418, + 480, + 438 + ], + "type": "text", + "content": "33. Liu, Z., Lin, Y., Cao, Y., Hu, H., Wei, Y., Zhang, Z., Lin, S., Guo, B.: Swin transformer: Hierarchical vision transformer using shifted windows. In: ICCV (2021)" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 132, + 439, + 480, + 471 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 439, + 480, + 471 + ], + "spans": [ + { + "bbox": [ + 132, + 439, + 480, + 471 + ], + "type": "text", + "content": "34. Loper, M., Mahmood, N., Romero, J., Pons-Moll, G., Black, M.J.: Spl: A skinned multi-person linear model. In: Seminal Graphics Papers: Pushing the Boundaries, Volume 2 (2023)" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 132, + 472, + 480, + 492 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 472, + 480, + 492 + ], + "spans": [ + { + "bbox": [ + 132, + 472, + 480, + 492 + ], + "type": "text", + "content": "35. Loshchilov, I., Hutter, F.: Decoupled weight decay regularization. arXiv preprint arXiv:1711.05101 (2017)" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 132, + 493, + 480, + 514 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 493, + 480, + 514 + ], + "spans": [ + { + "bbox": [ + 132, + 493, + 480, + 514 + ], + "type": "text", + "content": "36. Mahmood, N., Ghorbani, N., Troje, N.F., Pons-Moll, G., Black, M.J.: Amass: Archive of motion capture as surface shapes. In: ICCV (2019)" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 132, + 514, + 480, + 546 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 514, + 480, + 546 + ], + "spans": [ + { + "bbox": [ + 132, + 514, + 480, + 546 + ], + "type": "text", + "content": "37. Majoe, D., Widmer, L., Gutknecht, J.: Enhanced motion interaction for multimedia applications. In: Proceedings of the 7th International Conference on Advances in Mobile Computing and Multimedia (2009)" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 132, + 547, + 480, + 568 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 547, + 480, + 568 + ], + "spans": [ + { + "bbox": [ + 132, + 547, + 480, + 568 + ], + "type": "text", + "content": "38. Mao, W., Liu, M., Salzmann, M.: History repeats itself: Human motion prediction via motion attention. In: ECCV (2020)" + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 132, + 569, + 480, + 589 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 569, + 480, + 589 + ], + "spans": [ + { + "bbox": [ + 132, + 569, + 480, + 589 + ], + "type": "text", + "content": "39. Mao, W., Liu, M., Salzmann, M., Li, H.: Learning trajectory dependencies for human motion prediction. In: ICCV (2019)" + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 132, + 590, + 480, + 611 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 590, + 480, + 611 + ], + "spans": [ + { + "bbox": [ + 132, + 590, + 480, + 611 + ], + "type": "text", + "content": "40. Martinez, J., Black, M.J., Romero, J.: On human motion prediction using recurrent neural networks. In: CVPR (2017)" + } + ] + } + ], + "index": 23 + }, + { + "bbox": [ + 132, + 612, + 480, + 643 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 612, + 480, + 643 + ], + "spans": [ + { + "bbox": [ + 132, + 612, + 480, + 643 + ], + "type": "text", + "content": "41. Nichol, A., Dhariwal, P., Ramesh, A., Shyam, P., Mishkin, P., McGrew, B., Sutskever, I., Chen, M.: Glide: Towards photorealistic image generation and editing with text-guided diffusion models. arXiv preprint arXiv:2112.10741 (2021)" + } + ] + } + ], + "index": 24 + }, + { + "bbox": [ + 132, + 644, + 480, + 665 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 644, + 480, + 665 + ], + "spans": [ + { + "bbox": [ + 132, + 644, + 480, + 665 + ], + "type": "text", + "content": "42. Parmar, N., Vaswani, A., Uszkoreit, J., Kaiser, L., Shazeer, N., Ku, A., Tran, D.: Image transformer. In: ICML (2018)" + } + ] + } + ], + "index": 25 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "spans": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "type": "text", + "content": "16" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 166, + 91, + 219, + 102 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 166, + 91, + 219, + 102 + ], + "spans": [ + { + "bbox": [ + 166, + 91, + 219, + 102 + ], + "type": "text", + "content": "Q. Zou et al." + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 15 + }, + { + "para_blocks": [ + { + "bbox": [ + 132, + 116, + 481, + 665 + ], + "type": "list", + "angle": 0, + "index": 23, + "blocks": [ + { + "bbox": [ + 132, + 116, + 481, + 149 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 116, + 481, + 149 + ], + "spans": [ + { + "bbox": [ + 132, + 116, + 481, + 149 + ], + "type": "text", + "content": "43. Pavlakos, G., Choutas, V., Ghorbani, N., Bolkart, T., Osman, A.A., Tzionas, D., Black, M.J.: Expressive body capture: 3d hands, face, and body from a single image. In: CVPR (2019)" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 132, + 150, + 481, + 172 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 150, + 481, + 172 + ], + "spans": [ + { + "bbox": [ + 132, + 150, + 481, + 172 + ], + "type": "text", + "content": "44. Pavllo, D., Grangier, D., Auli, M.: Quaternet: A quaternion-based recurrent model for human motion. arXiv preprint arXiv:1805.06485 (2018)" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 132, + 172, + 481, + 194 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 172, + 481, + 194 + ], + "spans": [ + { + "bbox": [ + 132, + 172, + 481, + 194 + ], + "type": "text", + "content": "45. Petrovich, M., Black, M.J., Varol, G.: Action-conditioned 3d human motion synthesis with transformer vae. In: ICCV (2021)" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 132, + 194, + 481, + 216 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 194, + 481, + 216 + ], + "spans": [ + { + "bbox": [ + 132, + 194, + 481, + 216 + ], + "type": "text", + "content": "46. Petrovich, M., Black, M.J., Varol, G.: Temos: Generating diverse human motions from textual descriptions. In: ECCV (2022)" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 132, + 216, + 481, + 237 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 216, + 481, + 237 + ], + "spans": [ + { + "bbox": [ + 132, + 216, + 481, + 237 + ], + "type": "text", + "content": "47. Plappert, M., Mandery, C., Asfour, T.: The kit motion-language dataset. Big data (2016)" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 132, + 237, + 481, + 270 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 237, + 481, + 270 + ], + "spans": [ + { + "bbox": [ + 132, + 237, + 481, + 270 + ], + "type": "text", + "content": "48. Plappert, M., Mandery, C., Asfour, T.: Learning a bidirectional mapping between human whole-body motion and natural language using deep recurrent neural networks. Robotics and Autonomous Systems (2018)" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 132, + 270, + 481, + 293 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 270, + 481, + 293 + ], + "spans": [ + { + "bbox": [ + 132, + 270, + 481, + 293 + ], + "type": "text", + "content": "49. Raab, S., Leibovitch, I., Li, P., Aberman, K., Sorkine-Hornung, O., Cohen-Or, D.: Modi: Unconditional motion synthesis from diverse data. In: CVPR (2023)" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 132, + 293, + 481, + 326 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 293, + 481, + 326 + ], + "spans": [ + { + "bbox": [ + 132, + 293, + 481, + 326 + ], + "type": "text", + "content": "50. Radford, A., Kim, J.W., Hallacy, C., Ramesh, A., Goh, G., Agarwal, S., Sastry, G., Askell, A., Mishkin, P., Clark, J., et al.: Learning transferable visual models from natural language supervision. In: ICML (2021)" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 132, + 326, + 481, + 358 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 326, + 481, + 358 + ], + "spans": [ + { + "bbox": [ + 132, + 326, + 481, + 358 + ], + "type": "text", + "content": "51. Ramesh, A., Dhariwal, P., Nichol, A., Chu, C., Chen, M.: Hierarchical text-conditional image generation with clip latents. arXiv preprint arXiv:2204.06125 (2022)" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 132, + 358, + 481, + 380 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 358, + 481, + 380 + ], + "spans": [ + { + "bbox": [ + 132, + 358, + 481, + 380 + ], + "type": "text", + "content": "52. Rempe, D., Birdal, T., Hertzmann, A., Yang, J., Sridhar, S., Guibas, L.J.: Humor: 3d human motion model for robust pose estimation. In: ICCV (2021)" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 132, + 380, + 481, + 402 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 380, + 481, + 402 + ], + "spans": [ + { + "bbox": [ + 132, + 380, + 481, + 402 + ], + "type": "text", + "content": "53. Rombach, R., Blattmann, A., Lorenz, D., Esser, P., Ommer, B.: High-resolution image synthesis with latent diffusion models. In: CVPR (2022)" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 132, + 403, + 481, + 435 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 403, + 481, + 435 + ], + "spans": [ + { + "bbox": [ + 132, + 403, + 481, + 435 + ], + "type": "text", + "content": "54. Sahara, C., Chan, W., Saxena, S., Li, L., Whang, J., Denton, E.L., Ghasemipour, K., Gontijo Lopes, R., Karagol Ayan, B., Salimans, T., et al.: Photorealistic text-to-image diffusion models with deep language understanding. NeurIPS (2022)" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 132, + 435, + 481, + 456 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 435, + 481, + 456 + ], + "spans": [ + { + "bbox": [ + 132, + 435, + 481, + 456 + ], + "type": "text", + "content": "55. Thiebaut de Schotten, M., Forkel, S.J.: The emergent properties of the connected brain. Science (2022)" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 132, + 457, + 481, + 479 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 457, + 481, + 479 + ], + "spans": [ + { + "bbox": [ + 132, + 457, + 481, + 479 + ], + "type": "text", + "content": "56. Shafir, Y., Tevet, G., Kapon, R., Bermano, A.H.: Human motion diffusion as a generative prior. arXiv preprint arXiv:2303.01418 (2023)" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 132, + 479, + 481, + 511 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 479, + 481, + 511 + ], + "spans": [ + { + "bbox": [ + 132, + 479, + 481, + 511 + ], + "type": "text", + "content": "57. Siyao, L., Yu, W., Gu, T., Lin, C., Wang, Q., Qian, C., Loy, C.C., Liu, Z.: Bailando: 3d dance generation by actor-critic gpt with choreographic memory. In: CVPR (2022)" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 132, + 511, + 481, + 533 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 511, + 481, + 533 + ], + "spans": [ + { + "bbox": [ + 132, + 511, + 481, + 533 + ], + "type": "text", + "content": "58. Tang, X., Wang, H., Hu, B., Gong, X., Yi, R., Kou, Q., Jin, X.: Real-time controllable motion transition for characters. ACM TOG (2022)" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 132, + 533, + 481, + 577 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 533, + 481, + 577 + ], + "spans": [ + { + "bbox": [ + 132, + 533, + 481, + 577 + ], + "type": "text", + "content": "59. Terlemez, Ö., Ulbrich, S., Mandery, C., Do, M., Vahrenkamp, N., Asfour, T.: Master motor map (mmm)—framework and toolkit for capturing, representing, and reproducing human motion on humanoid robots. In: 2014 IEEE-RAS International Conference on Humanoid Robots (2014)" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 132, + 577, + 481, + 600 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 577, + 481, + 600 + ], + "spans": [ + { + "bbox": [ + 132, + 577, + 481, + 600 + ], + "type": "text", + "content": "60. Tevet, G., Gordon, B., Hertz, A., Bermano, A.H., Cohen-Or, D.: Motionclip: Exposing human motion generation to clip space. In: ECCV (2022)" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 132, + 600, + 481, + 622 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 600, + 481, + 622 + ], + "spans": [ + { + "bbox": [ + 132, + 600, + 481, + 622 + ], + "type": "text", + "content": "61. Tevet, G., Raab, S., Gordon, B., Shafir, Y., Cohen-Or, D., Bermano, A.H.: Human motion diffusion model. arXiv preprint arXiv:2209.14916 (2022)" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 132, + 622, + 481, + 643 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 622, + 481, + 643 + ], + "spans": [ + { + "bbox": [ + 132, + 622, + 481, + 643 + ], + "type": "text", + "content": "62. Vaswani, A., Shazeer, N., Parmar, N., Uszkoreit, J., Jones, L., Gomez, A.N., Kaiser, L., Polosukhin, I.: Attention is all you need. NeurIPS (2017)" + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 132, + 643, + 481, + 665 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 643, + 481, + 665 + ], + "spans": [ + { + "bbox": [ + 132, + 643, + 481, + 665 + ], + "type": "text", + "content": "63. Wang, Y., Leng, Z., Li, F.W., Wu, S.C., Liang, X.: Fg-t2m: Fine-grained text-driven human motion generation via diffusion model. In: ICCV (2023)" + } + ] + } + ], + "index": 22 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 235, + 91, + 447, + 102 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 235, + 91, + 447, + 102 + ], + "spans": [ + { + "bbox": [ + 235, + 91, + 447, + 102 + ], + "type": "text", + "content": "ParCo: Part-Coordinating Text-to-Motion Synthesis" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 470, + 91, + 481, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 470, + 91, + 481, + 100 + ], + "spans": [ + { + "bbox": [ + 470, + 91, + 481, + 100 + ], + "type": "text", + "content": "17" + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 16 + }, + { + "para_blocks": [ + { + "bbox": [ + 132, + 116, + 480, + 325 + ], + "type": "list", + "angle": 0, + "index": 10, + "blocks": [ + { + "bbox": [ + 132, + 116, + 480, + 138 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 116, + 480, + 138 + ], + "spans": [ + { + "bbox": [ + 132, + 116, + 480, + 138 + ], + "type": "text", + "content": "64. Yan, S., Li, Z., Xiong, Y., Yan, H., Lin, D.: Convolutional sequence generation for skeleton-based action synthesis. In: ICCV (2019)" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 132, + 138, + 480, + 160 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 138, + 480, + 160 + ], + "spans": [ + { + "bbox": [ + 132, + 138, + 480, + 160 + ], + "type": "text", + "content": "65. Yeasin, M., Polat, E., Sharma, R.: A multiobject tracking framework for interactive multimedia applications. IEEE TMM (2004)" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 132, + 160, + 480, + 194 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 160, + 480, + 194 + ], + "spans": [ + { + "bbox": [ + 132, + 160, + 480, + 194 + ], + "type": "text", + "content": "66. Zhang, J., Zhang, Y., Cun, X., Huang, S., Zhang, Y., Zhao, H., Lu, H., Shen, X.: T2m-gpt: Generating human motion from textual descriptions with discrete representations. arXiv preprint arXiv:2301.06052 (2023)" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 132, + 194, + 480, + 225 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 194, + 480, + 225 + ], + "spans": [ + { + "bbox": [ + 132, + 194, + 480, + 225 + ], + "type": "text", + "content": "67. Zhang, M., Cai, Z., Pan, L., Hong, F., Guo, X., Yang, L., Liu, Z.: Motiondiffuse: Text-driven human motion generation with diffusion model. arXiv preprint arXiv:2208.15001 (2022)" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 132, + 225, + 480, + 258 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 225, + 480, + 258 + ], + "spans": [ + { + "bbox": [ + 132, + 225, + 480, + 258 + ], + "type": "text", + "content": "68. Zhang, M., Guo, X., Pan, L., Cai, Z., Hong, F., Li, H., Yang, L., Liu, Z.: Remodif- fuse: Retrieval-augmented motion diffusion model. arXiv preprint arXiv:2304.01116 (2023)" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 132, + 258, + 480, + 281 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 258, + 480, + 281 + ], + "spans": [ + { + "bbox": [ + 132, + 258, + 480, + 281 + ], + "type": "text", + "content": "69. Zhang, Y., Black, M.J., Tang, S.: Perpetual motion: Generating unbounded human motion. arXiv preprint arXiv:2007.13886 (2020)" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 132, + 281, + 480, + 303 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 281, + 480, + 303 + ], + "spans": [ + { + "bbox": [ + 132, + 281, + 480, + 303 + ], + "type": "text", + "content": "70. Zhao, R., Su, H., Ji, Q.: Bayesian adversarial human motion synthesis. In: CVPR (2020)" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 132, + 303, + 480, + 325 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 303, + 480, + 325 + ], + "spans": [ + { + "bbox": [ + 132, + 303, + 480, + 325 + ], + "type": "text", + "content": "71. Zhong, C., Hu, L., Zhang, Z., Xia, S.: Attt2m: Text-driven human motion generation with multi-perspective attention mechanism. In: ICCV (2023)" + } + ] + } + ], + "index": 9 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "spans": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "type": "text", + "content": "18" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 166, + 91, + 219, + 102 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 166, + 91, + 219, + 102 + ], + "spans": [ + { + "bbox": [ + 166, + 91, + 219, + 102 + ], + "type": "text", + "content": "Q. Zou et al." + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 17 + } + ], + "_backend": "vlm", + "_version_name": "2.6.4" +} \ No newline at end of file diff --git a/2024/Parameter-Efficient and Memory-Efficient Tuning for Vision Transformer_ A Disentangled Approach/e724c6a7-eae9-4271-a988-9b07cf4a65fc_content_list.json b/2024/Parameter-Efficient and Memory-Efficient Tuning for Vision Transformer_ A Disentangled Approach/e724c6a7-eae9-4271-a988-9b07cf4a65fc_content_list.json new file mode 100644 index 0000000000000000000000000000000000000000..1583bd0f1038b6eeb35ad2e89cb21cec990f1d35 --- /dev/null +++ b/2024/Parameter-Efficient and Memory-Efficient Tuning for Vision Transformer_ A Disentangled Approach/e724c6a7-eae9-4271-a988-9b07cf4a65fc_content_list.json @@ -0,0 +1,1891 @@ +[ + { + "type": "text", + "text": "Parameter-Efficient and Memory-Efficient Tuning for Vision Transformer: A Disentangled Approach", + "text_level": 1, + "bbox": [ + 215, + 140, + 785, + 186 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Taolin Zhang $^{1*}$ , Jiawang Bai $^{2,3*}$ , Zhihe Lu $^{3}$ , Dongze Lian $^{3}$ , Genping Wang $^{4,\\text{凹}}$ , Xinchao Wang $^{3,\\text{凹}}$ , and Shu-Tao Xia $^{1,5}$", + "bbox": [ + 217, + 210, + 784, + 244 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "$^{1}$ Tsinghua Shenzhen International Graduate School, Tsinghua University $^{2}$ Tencent $^{3}$ National University of Singapore $^{4}$ Shenzhen Qiji Technology Co., Ltd.", + "bbox": [ + 256, + 253, + 753, + 297 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "$^{5}$ Research Center of Artificial Intelligence, Peng Cheng Laboratory ztl23@mails.tsinghua.edu.cn; jiawangbai@tencent.com; zhihelu.academic@gmail.com; dzlianx@gmail.com; genpingwang@163.com; xinchao@nus.edu.sg; xiast@sz.tsinghua.edu.cn", + "bbox": [ + 250, + 297, + 753, + 353 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Abstract. Recent works on parameter-efficient transfer learning (PETL) show the potential to adapt a pre-trained Vision Transformer to downstream recognition tasks with only a few learnable parameters. However, since they usually insert new structures into the pre-trained model, entire intermediate features of that model are changed and thus need to be stored to be involved in back-propagation, resulting in memory-heavy training. We solve this problem from a novel disentangled perspective, i.e., dividing PETL into two aspects: task-specific learning and pretrained knowledge utilization. Specifically, we synthesize the task-specific query with a learnable and lightweight module, which is independent of the pre-trained model. The synthesized query equipped with task-specific knowledge serves to extract the useful features for downstream tasks from the intermediate representations of the pre-trained model in a query-only manner. Built upon these features, a customized classification head is proposed to make the prediction for the input sample. Given that our method employs an extremely lightweight architecture and avoids the use of heavy intermediate features for running gradient descent, it demonstrates limited memory usage in training. Extensive experiments manifest that our method achieves state-of-the-art performance under memory constraints, showcasing its applicability in real-world situations. The code is available at: https://synqt.github.io/.", + "bbox": [ + 259, + 391, + 743, + 681 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "1 Introduction", + "text_level": 1, + "bbox": [ + 215, + 709, + 375, + 724 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Motivated by the success of data-driven solutions, there has been an explosive growth in the size of vision models [4,10,13,20,31,49,62] and the scale of accessible datasets like ImageNet [9]. To unleash the power of large models, it becomes a tendency to first pre-train on a general and large-scale dataset, and then adapt", + "bbox": [ + 212, + 741, + 787, + 801 + ], + "page_idx": 0 + }, + { + "type": "page_footnote", + "text": "* Equal contribution. \n $\\boxtimes$ Corresponding authors.", + "bbox": [ + 230, + 811, + 410, + 840 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "them to downstream tasks [21, 41, 48, 52]. The most obvious adaption strategy is to fully fine-tune all parameters of the pre-trained model. However, it necessitates that the device saves a dedicated set of model parameters for each single task, leading to substantial storage overhead [57].", + "bbox": [ + 212, + 146, + 782, + 207 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "To improve the storage efficiency, parameter-efficient transfer learning (PETL) methods [8,16-19,59] have been explored by tuning only a small set of the parameters in adapting the large pre-trained models to downstream tasks. Typically, these methods insert small structures into the frozen pre-trained model and specify the newly added parameters as the only learnable ones. For instance, Visual Prompt Tuning (VPT) [17] prepends learnable tokens to the patch embeddings; AdaptFormer [8] inserts a bottleneck-structured fully connected layers parallel to the feed-forward network (FFN) of the original ViT. These studies demonstrate that PETL is a promising way to save storage resources.", + "bbox": [ + 212, + 209, + 787, + 345 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "However, even though updating a small set of parameters, existing PETL methods still require a very high consumption of memory in training. The direct reason is that they need a full back-propagation throughout the entire backbone. Given the fact that the model size is exponentially increasing, it becomes infeasible to apply existing PETL methods to the memory-constraint situations. Therefore, it is desirable to investigate memory-efficient PETL while maintaining performance on downstream tasks.", + "bbox": [ + 212, + 348, + 787, + 454 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "In this work, we attribute the high memory cost of existing PETL methods in training to their intrinsic regime, i.e., entangling task-specific learning and pre-trained knowledge utilization together. It is usually achieved by integrating trainable structures into the pre-trained ViT and changing all intermediate features. Consequently, a great number of intermediate features need to be stored in training to run gradient descent. The aforementioned analysis motivates us to design a memory-efficient PETL method from a disentangled perspective.", + "bbox": [ + 212, + 457, + 787, + 561 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Accordingly, we propose Synthesized Query Tuning (SynQT), a memory-efficient PETL method for adapting pre-trained ViTs to downstream recognition tasks. Overall, as shown in Figure 1, SynQT takes the intermediate representations of the original pre-trained model as input and combines them with task-specific knowledge using stacked blocks. We propose the Query Synthesis Module (QSM) and Knowledge Extraction Module (KEM) to accomplish the disentanglement. To capture task-specific knowledge, SynQT employs a lightweight QSM, which is independent of the pre-trained model. In QSM, an attention layer and an FFN are trained on the target downstream task to synthesize task-specific query. Then, to utilize the pre-trained knowledge, the synthesized query interacts with the intermediate representations of the original pre-trained model in a KEM. It outputs useful features for downstream recognition, where all parameters are inherited from the pre-trained model and frozen. Relying on rich hierarchical features from KEMs, we propose a classification head, which adaptively aggregates these features conditioned on the input sample for prediction.", + "bbox": [ + 212, + 565, + 789, + 792 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "The success of SynQT in saving memory in training is derived from two technical aspects. Firstly, we keep intermediate representations of the original pretrained model intact and directly exploit them in a query-only manner, avoiding", + "bbox": [ + 212, + 795, + 787, + 839 + ], + "page_idx": 1 + }, + { + "type": "page_number", + "text": "2", + "bbox": [ + 217, + 114, + 228, + 126 + ], + "page_idx": 1 + }, + { + "type": "header", + "text": "Taolin Zhang, Jiawang Bai et al.", + "bbox": [ + 271, + 114, + 488, + 128 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "the use of heavy intermediate features for running gradient descent. Besides, our design including QSM and KEM is extremely lightweight. For instance, we adopt bottleneck structures for Q-K-V projection [46] and a small number of tokens (lower than 4 in most cases) in QSM.", + "bbox": [ + 212, + 146, + 782, + 205 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "We validate SynQT on VTAB-1K [58] benchmark, including 19 tasks from diverse domains. We also evaluate our method in few-shot learning using 5 fine-grained visual classification datasets. Extensive experiments show that SynQT can significantly reduce the memory footprint in training with promising accuracy. Compared to VPT-Deep, a powerful version of VPT that inserts prompts in each layer of ViT, SynQT saves $56\\%$ memory consumption and achieves higher Top-1 accuracy on average. The contributions of our work are as follows:", + "bbox": [ + 212, + 207, + 784, + 311 + ], + "page_idx": 2 + }, + { + "type": "list", + "sub_type": "text", + "list_items": [ + "1. We provide an insightful analysis of the high memory cost of existing PETL methods, i.e., they entangle task-specific learning and pre-trained knowledge utilization together.", + "2. Inspired by the disentangled idea, we propose SynQT, which synthesizes task-specific query and then utilizes the pre-trained knowledge in a query-only manner.", + "3. Extensive experiments show that SynQT is highly memory-efficient and effective, notably outperforming existing baselines under memory constraints." + ], + "bbox": [ + 222, + 325, + 784, + 446 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "2 Related Work", + "text_level": 1, + "bbox": [ + 215, + 472, + 387, + 489 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Parameter-Efficient Transfer Learning (PETL). The idea of PETL initially shows great success in the field of natural language processing (NLP) [6,16,28,40]. It fine-tunes a small number of trainable parameters to transfer large pre-trained models to downstream tasks, significantly reducing storage costs by avoiding a copy of model parameters for each new task. Due to the increased size of pre-trained vision models, especially for ViT [10], PETL has been introduced to computer vision and attracted much attention [3,8,12,17,26,27,34,39,44, 45,50,55,56,60,61]. To reduce the number of fine-tuned parameters, BitFit [54] only fine-tunes the bias terms and freezes most of the network, resulting in comparable performance to full tuning. VPT [17] introduces learnable tokens (i.e., prompts [24,25,29,30]) to each layer of ViT. SSF [27] fine-tunes scale and shift factors in each layer to adapt original features to the target domain. Some other works [8,15,16] propose to insert additional trainable structures into attention blocks of the pre-trained Transformer. For instance, LoRA [16] observes that the learned over-parametrized models reside on a low-rank dimension and further injects trainable rank decomposition parameters into each layer of the Transformer; AdaptFormer [8] adds parallel lightweight module to FFN. Furthermore, for extreme parameter efficiency, a tensorization-decomposition framework [18] is exploited to store the weight increments, and the low-bit adapter [19] is proposed to reduce the precision redundancy. However, how to improve memory efficiency in training is still an under-explored problem in PETL, as discussed below.", + "bbox": [ + 212, + 507, + 785, + 839 + ], + "page_idx": 2 + }, + { + "type": "header", + "text": "Parameter-Efficient and Memory-Efficient Tuning for Vision Transformer", + "bbox": [ + 243, + 114, + 730, + 128 + ], + "page_idx": 2 + }, + { + "type": "page_number", + "text": "3", + "bbox": [ + 774, + 116, + 784, + 126 + ], + "page_idx": 2 + }, + { + "type": "image", + "img_path": "images/6b9e702486b0bd3db86ad0a5cac53de8656c0e72bda71ddf5c90b382220bf160.jpg", + "image_caption": [ + "Fig. 1: Overall architecture of SynQT. Each block in SynQT contains two key designs, including Query Synthesis Module (QSM) and Knowledge Extractor Module (KEM). QSM is a lightweight learnable module to capture task-specific knowledge and KEM reuses frozen weights in the pre-trained model for feature extraction. The rich and hierarchical features from KEMs are fed into the classification head." + ], + "image_footnote": [], + "bbox": [ + 233, + 143, + 772, + 300 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Memory-Efficient PETL. Although existing PETL methods have shown promising performance in transfer learning while saving most trainable parameters, they still suffer from issues of high memory consumption in training. The full back-propagation throughout the entire backbone prevents traditional PETL methods from being applied in memory-constraint situations. Recently, some recent works [11,43,45] have made several attempts to address this issue, whose key idea is to leverage the power of intermediate representations of the pre-trained model and thus avoid full back-propagation. In NLP, LST [43] proposes a trainable lightweight structure in the sidewalk for memory-saving during training. To adapt vision models, Head2Toe [11] selects features from all layers of the source model to feed into a classification head. VQT [45] inserts trainable tokens in the query space to capture useful information from the pre-trained model. Due to the information redundancy in the intermediate representations, both Head2Toe and VQT further adopt group lasso [1,53] for feature selection.", + "bbox": [ + 212, + 411, + 787, + 622 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "In contrast to previous works, we provide an insightful analysis that entangled combination of task-specific learning and pre-trained knowledge utilization results in high memory consumption of traditional PETL methods. Therefore, we propose SynQT to disentangle PETL into two distinct modules, with the aim of capturing task-specific knowledge [51] and extracting useful features for downstream tasks, respectively. Such a disentangled paradigm enjoys memory efficiency in transfer learning compared with traditional PETL methods.", + "bbox": [ + 212, + 623, + 787, + 729 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "3 Method", + "text_level": 1, + "bbox": [ + 215, + 752, + 330, + 767 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "3.1 Overall Framework", + "text_level": 1, + "bbox": [ + 215, + 784, + 419, + 797 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "As can be seen from Figures 1 and 2, SynQT makes the prediction relying on the features produced by stacked blocks, which take the intermediate representations", + "bbox": [ + 212, + 809, + 785, + 839 + ], + "page_idx": 3 + }, + { + "type": "page_number", + "text": "4", + "bbox": [ + 217, + 116, + 228, + 126 + ], + "page_idx": 3 + }, + { + "type": "header", + "text": "Taolin Zhang, Jiawang Bai et al.", + "bbox": [ + 271, + 114, + 488, + 128 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "of the original pre-trained model as external input. In SynQT, each block mainly includes two parts: Query Synthesis Module (QSM) and Knowledge Extraction Module (KEM). QSM is a lightweight module, which is used to capture task-specific knowledge, connecting the pre-trained model with the downstream task. KEM uses the tokens generated by QSM as queries, and the intermediate features in the pre-trained ViT as keys and values. It reuses the original model's weights to extract features that are beneficial to downstream tasks. Finally, the rich hierarchical features extracted by the KEMs are fed into the classification head, as shown in Figure 2. We will elaborate our designs in the following sections.", + "bbox": [ + 212, + 146, + 787, + 282 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "3.2 Query Synthesis Module", + "text_level": 1, + "bbox": [ + 214, + 305, + 465, + 320 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "PETL integrates the task-specific knowledge into the adapted model through training on the downstream dataset. Typically, existing methods [8, 15-17] insert additional trainable structures into the pre-trained model. However, this is the root of high memory consumption, since it changes all original features, and all intermediate features are modified and stored to be involved in calculating gradients. From the disentangled perspective, we solve this problem with a task-specific learning module, named Query Synthesis Module (QSM), which is trainable, lightweight, and structurally independent from the pre-trained model.", + "bbox": [ + 212, + 329, + 787, + 450 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "The purpose of QSM is to synthesize query tokens, interacting with intermediate representations of the original pre-trained model. Besides outputs from the previous block, we introduce a trainable prompt into the input space of each QSM. These per-block additional prompts increase the capacity of QSM to capture task-specific knowledge. Inspired by the Transformer, QSM includes an attention layer and an FFN. Specifically, the input tokens interact with each other through a self-attention layer and are further fed into the feed-forward network for producing the query tokens. Besides, to keep the parameter efficiency, another design principle is to extensively use bottleneck-structured fully connected layers to reduce the number of parameters.", + "bbox": [ + 212, + 450, + 787, + 601 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Let $n$ and $d$ denote the number of tokens and the hidden size in QSM, respectively. Note that $n$ is very small, e.g., $n = 4$ , much lower than the number of tokens in the original model and the value of $d$ corresponds to that of the pre-trained model. Before the self-attention layer, the output of the last block $\\hat{H}_{i - 1} \\in \\mathbb{R}^{n \\times d}$ are added with trainable prompt $P_{i}$ and then fed into bottleneck projection layers. It is formally formulated as follows,", + "bbox": [ + 212, + 601, + 787, + 691 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\nZ _ {i} ^ {\\prime} = \\left(\\left(H _ {i} + P _ {i}\\right) \\cdot \\mathbf {W} _ {\\text {d o w n}}\\right) \\cdot \\mathbf {W} _ {\\text {u p}}, \\tag {1}\n$$\n", + "text_format": "latex", + "bbox": [ + 385, + 705, + 784, + 720 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "where $\\mathbf{W}_{down} \\in \\mathbb{R}^{d \\times \\hat{d}}$ and $\\mathbf{W}_{up} \\in \\mathbb{R}^{\\hat{d} \\times d}$ denote the down-projection and up-projection layer respectively, with $\\hat{d} \\ll d$ .", + "bbox": [ + 212, + 733, + 785, + 766 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "In the attention layer, $Z_{i}^{\\prime}$ is projected by three bottleneck modules to produce the query $Q_{i}^{qs}$ , key $K_{i}^{qs}$ , and value $V_{i}^{qs}$ . Then the hidden embedding after attention layers $Z_{i}^{\\prime \\prime}$ can be calculated as follows, controlled by a scale factor $s^{\\prime}$ :", + "bbox": [ + 212, + 767, + 785, + 811 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\nZ _ {i} ^ {\\prime \\prime} = s ^ {\\prime} \\cdot \\operatorname {A t t n} \\left(Q _ {i} ^ {q s}, K _ {i} ^ {q s}, V _ {i} ^ {q s}\\right) + Z _ {i} ^ {\\prime}. \\tag {2}\n$$\n", + "text_format": "latex", + "bbox": [ + 377, + 824, + 784, + 840 + ], + "page_idx": 4 + }, + { + "type": "header", + "text": "Parameter-Efficient and Memory-Efficient Tuning for Vision Transformer", + "bbox": [ + 243, + 114, + 732, + 128 + ], + "page_idx": 4 + }, + { + "type": "page_number", + "text": "5", + "bbox": [ + 774, + 116, + 784, + 126 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Then we can further obtain the synthesized query $\\hat{H}_i$ after the FFN as the input of KEM, controlled by another scale factor $s''$ :", + "bbox": [ + 212, + 145, + 782, + 176 + ], + "page_idx": 5 + }, + { + "type": "equation", + "text": "\n$$\n\\hat {H} _ {i} = s ^ {\\prime \\prime} \\cdot \\operatorname {F F N} \\left(\\operatorname {L N} \\left(Z _ {i} ^ {\\prime \\prime}\\right)\\right) + Z _ {i} ^ {\\prime \\prime}, \\tag {3}\n$$\n", + "text_format": "latex", + "bbox": [ + 393, + 186, + 785, + 203 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "where FFN also consists of bottleneck-structured fully connected layers and $\\mathrm{LN}(\\cdot)$ represents the layer normalization [2]. $s^{\\prime}$ and $s''$ are hyper-parameters to modulate the effects of the attention and FFN on the synthesized query. Since QSM is trained on the downstream dataset from scratch, we believe that its output $\\hat{H}_i$ captures the task-specific knowledge. We use $\\hat{H}_i$ in a query-only manner as illustrated in the next section.", + "bbox": [ + 212, + 212, + 782, + 301 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "3.3 Knowledge Extraction Module", + "text_level": 1, + "bbox": [ + 215, + 323, + 514, + 338 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "With the task-specific query obtained from QSM, we propose Knowledge Extraction Module (KEM) to exploit intermediate features of the original pre-trained model. Before that, we obtain original features $\\{X_{1}, X_{2}, \\dots, X_{l}\\}$ of the input sample from forward-propagation, where $l$ is the number of transformer blocks of the pre-train model. Note that this process is efficient due to no gradient calculation.", + "bbox": [ + 212, + 347, + 782, + 421 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "KEM takes the synthesized query $\\hat{H}_i$ and original features $\\{X_1, X_2, \\dots, X_l\\}$ as inputs and outputs useful features for the downstream task. Since $\\hat{H}_i$ has a small number of tokens and original features are only queried by $\\hat{H}_i$ , there are no heavy features in the middle of KEM. Besides, we find that it is effective to reuse and freeze the pre-trained parameters for feature interaction in KEM. These designs ensure the parameter- and memory-efficiency of our method.", + "bbox": [ + 212, + 422, + 784, + 512 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Specifically, with the projection weight in the pre-trained ViT blocks, we first obtain the query $Q_{i}^{ke}$ , key $K_{i}^{ke}$ , and value $V_{i}^{ke}$ . Then follow-up attention calculation is performed by:", + "bbox": [ + 212, + 512, + 782, + 556 + ], + "page_idx": 5 + }, + { + "type": "equation", + "text": "\n$$\nE _ {i} = \\operatorname {A t t n} \\left(Q _ {i} ^ {k e}, K _ {i} ^ {k e}, V _ {i} ^ {k e}\\right) + \\hat {H} _ {i}. \\tag {4}\n$$\n", + "text_format": "latex", + "bbox": [ + 388, + 566, + 784, + 584 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "We also reuse the weights of FFN to perform the same space transformation as the pre-trained model:", + "bbox": [ + 212, + 593, + 782, + 622 + ], + "page_idx": 5 + }, + { + "type": "equation", + "text": "\n$$\nH _ {i} = \\operatorname {F F N} \\left(\\operatorname {N o r m} \\left(E _ {i}\\right)\\right) + E _ {i}, \\tag {5}\n$$\n", + "text_format": "latex", + "bbox": [ + 401, + 633, + 784, + 650 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "where $\\mathrm{Norm}(\\cdot)$ denotes the normalization used in the pre-trained model.", + "bbox": [ + 212, + 659, + 735, + 672 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "To maximize the potential of KEM, apart from the final output $H_{i}$ , we additionally leverage features $F_{i}^{att}$ and $F_{i}^{ffn}$ after the attention module and the FFN for classification. They are obtained as follows:", + "bbox": [ + 212, + 674, + 782, + 718 + ], + "page_idx": 5 + }, + { + "type": "equation", + "text": "\n$$\n\\left\\{ \\begin{array}{l} F _ {i} ^ {a t t} = \\operatorname {A t t n} \\left(Q _ {i} ^ {k e}, K _ {i} ^ {k e}, V _ {i} ^ {k e}\\right) \\\\ F _ {i} ^ {f f n} = \\operatorname {F F N} \\left(\\operatorname {N o r m} \\left(E _ {i}\\right)\\right) \\end{array} . \\right. \\tag {6}\n$$\n", + "text_format": "latex", + "bbox": [ + 390, + 728, + 784, + 768 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "3.4 Classification Head", + "text_level": 1, + "bbox": [ + 215, + 786, + 419, + 800 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "The classification head makes a prediction of the input sample relying on rich and hierarchical features from KEMs, $\\{F_1^{att}, F_1^{ffn}, H_1, \\dots, F_l^{att}, F_l^{ffn}, H_l\\}$ , where $l$", + "bbox": [ + 212, + 809, + 782, + 840 + ], + "page_idx": 5 + }, + { + "type": "page_number", + "text": "6", + "bbox": [ + 217, + 114, + 228, + 126 + ], + "page_idx": 5 + }, + { + "type": "header", + "text": "Taolin Zhang, Jiawang Bai et al.", + "bbox": [ + 271, + 114, + 488, + 128 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "is the number of blocks of the pre-trained ViT. Instead of directly applying an MLP classifier on top of these features, we propose several techniques for better performance. Figure 2 shows the overall architecture of the classification head.", + "bbox": [ + 212, + 146, + 782, + 191 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Feature projection. Since these features are generated across different layers from KEMs, they may be misaligned in the representation space. Therefore, before aggregating all features, we align all features with projection layers. We use bottleneck-structured fully connected layers and share the projection layers across all features to improve the parameter efficiency.", + "bbox": [ + 212, + 191, + 531, + 327 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Conditional feature aggregation. These features contain a large amount of information but also some redundancy, perhaps hurting the performance of downstream tasks. Therefore, it is crucial to perform feature selection, aiming to extract the truly beneficial information. We propose a strategy of conditional feature aggregation, which generates weights conditioned on $H_{l}$ and performs a weighted-sum over all features. Our strategy enables SynQT to adaptively rely on different features to classify different samples. Despite only a few parameters, the conditional feature aggregation results in better performance compared to a simple average or weighted sum with learnable but fixed weights, as shown in our experiments.", + "bbox": [ + 212, + 329, + 532, + 584 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "DropFeat in training. Due to rich features fed into the classification head, overfitting likely happens and degrades the performance. To solve this issue, inspired by the idea of dropout [23, 42, 47], we propose DropFeat, a simple regularization technique to prevent overfitting. DropFeat randomly removes some input features for the classification head during training, while all features are leveraged during inference. DropFeat ensures that the model does not overly rely on a specific feature, which significantly improves the model's performance.", + "bbox": [ + 212, + 585, + 787, + 691 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/e2c9082112eeda8075dbba1853437621577731d2ee22dc2c4b3a2831bdbde3c7.jpg", + "image_caption": [ + "Fig.2: Classification head in SynQT, corresponding to the head in Figure 1. We utilize a shared projection to align the hierarchical features and apply feature aggregation conditioned on the output of the last block $H_{l}$ . We also adopt DropFeat to prevent overfitting, which randomly drops some features during training." + ], + "image_footnote": [], + "bbox": [ + 555, + 223, + 785, + 428 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "4 Experiment", + "text_level": 1, + "bbox": [ + 214, + 717, + 366, + 734 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "4.1 Setup", + "text_level": 1, + "bbox": [ + 214, + 751, + 310, + 767 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Datasets. We conduct experiments on the VTAB-1K benchmark and the fine-grained visual classification (FGVC) datasets. VTAB-1K consists of 19 different classification tasks categorized into three groups: Natural, Specialized, and Structured. Each task of VTAB-1K contains 1000 training samples. We evaluate the", + "bbox": [ + 212, + 779, + 787, + 840 + ], + "page_idx": 6 + }, + { + "type": "header", + "text": "Parameter-Efficient and Memory-Efficient Tuning for Vision Transformer", + "bbox": [ + 243, + 113, + 732, + 128 + ], + "page_idx": 6 + }, + { + "type": "page_number", + "text": "7", + "bbox": [ + 774, + 114, + 784, + 126 + ], + "page_idx": 6 + }, + { + "type": "table", + "img_path": "images/b56fe2d14702c227b54a844826dfbd203c99d46e8d9673019f9ce1994c535321.jpg", + "table_caption": [ + "Table 1: Full results on VTAB-1K benchmark with pre-trained ViT-B/16 on ImageNet21K. “# Memory” specifies memory cost during training with 64 as batch size on CIFAR100. We report the Top-1 accuracy and the results are averaged over group-wise mean values." + ], + "table_footnote": [], + "table_body": "
# Memory (GB)NaturalSpecializedStructured
CIFAR100Caltech101DTDFlower102PetsSVHNSun397CamelyonEuroSATRess45RetinopathyClever-CountClever-DistDMLabKITTL-DistdSpr-LocdSpr-OrisNORB-AzimsNORB-EleAverage
Traditional Fine-Tuning
Linear Probing2.550.685.661.479.586.540.838.079.791.571.765.541.434.434.155.418.126.416.524.852.7
Full Fine-Tuning11.144.384.554.184.774.787.226.985.395.076.070.471.560.546.972.974.538.728.523.863.2
PETL Methods
BitFit [54]7.872.887.059.297.585.359.951.478.791.672.969.861.555.632.455.966.640.015.725.165.2
VPT-Shallow [17]8.177.786.962.697.587.374.551.278.292.075.672.950.558.640.567.168.736.120.234.167.8
VPT-Deep [17]8.178.890.865.898.088.378.149.681.896.183.468.468.560.046.572.873.647.932.937.872.0
Memory-Efficient PETL Methods
VQT [45]3.266.389.967.897.984.779.945.579.095.280.974.746.761.645.163.662.932.130.028.868.3
SynQT (Ours)3.470.989.768.898.589.677.850.682.396.783.575.271.862.748.575.474.149.031.736.172.9
", + "bbox": [ + 218, + 212, + 784, + 382 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "few-shot performance of SynQT on 5 FGVC datasets: FGVC-Aircraft [36], OxfordPets [38], Food-101 [5], Stanford Cars [22], and Oxford-Flowers102 [37]. We follow the split in [59] and report the results in 1, 2, 4, 8, and 16-shot settings. The accuracy is averaged over three trials with different seeds.", + "bbox": [ + 212, + 410, + 784, + 469 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Implementation details. We utilize a pre-trained ViT-B/16 [10] on the ImageNet-21K dataset as the foundational model. We also conduct experiments on the ImageNet-1K supervised pre-trained backbone, as well as on the self-supervised checkpoints (DINO [7] and MAE [14]). For a lightweight design of QSM, we use an extremely small number of tokens, with the number of tokens $n$ from $\\{1, 4, 16\\}$ . This is significantly fewer than the number of tokens in VPT, which can reach up to 200. Additionally, the hyper-parameters $s'$ and $s''$ are chosen from $\\{0.1, 1\\}$ . The hidden size of the bottleneck projection is set to 48, except attention projection of the query, key, and value in QSM, which is set to 8.", + "bbox": [ + 212, + 470, + 805, + 607 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Due to the limited number of validation samples in VTAB-1K, we implement 5-fold cross-validation on the training sets to determine the optimal hyperparameters. For few-shot learning, we use the default splits of FGVC datasets to search for the best hyperparameters. The Top-1 accuracy is reported on the test sets in all our experiments. Following common practice [18], we adopt AdamW [33] as the optimizer and a cosine learning rate scheduler with a constant batch size of 64. For data augmentation, we adjust the image size to $224 \\times 224$ directly for VTAB-1K, and for few-shot learning, we resize the image to $256 \\times 256$ before applying a center crop.", + "bbox": [ + 212, + 608, + 787, + 744 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "4.2 Main Results", + "text_level": 1, + "bbox": [ + 215, + 767, + 372, + 781 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Besides the traditional fine-tuning methods (linear probing and full fine-tuning), we compare our method with PETL methods including BitFit [54], VPT-Shallow [17], and VPT-Deep [17], as well as a memory-efficient one, VQT [45], on VTAB-", + "bbox": [ + 212, + 794, + 785, + 842 + ], + "page_idx": 7 + }, + { + "type": "page_number", + "text": "8", + "bbox": [ + 217, + 114, + 228, + 126 + ], + "page_idx": 7 + }, + { + "type": "header", + "text": "Taolin Zhang, Jiawang Bai et al.", + "bbox": [ + 271, + 114, + 488, + 128 + ], + "page_idx": 7 + }, + { + "type": "table", + "img_path": "images/65fba2f418828f031e49bd7b49d8c618445fd72c013581f94e36172076d6d161.jpg", + "table_caption": [ + "Table 2: Classification results on MAE, DINO, and Supervised ImageNet-1K pre-trained backbones. We report the Top-1 accuracy and the averaged results." + ], + "table_footnote": [], + "table_body": "
MethodsNaturalSpecializedStructuredAverage
MAE backbone
BitFit [54]63.580.657.267.1
VPT-Shallow [17]38.372.025.045.1
VPT-Deep [17]50.876.437.354.8
VQT [45]56.678.643.459.5
SynQT (Ours)66.082.658.268.9
DINO backbone
BitFit [54]76.883.849.269.9
VPT-Shallow [17]69.882.638.563.6
VPT-Deep [17]76.783.847.369.3
VQT [45]71.083.637.263.9
SynQT (Ours)75.184.856.972.3
Supervised ImageNet-1K backbone
BitFit [54]74.683.449.469.1
VPT-Shallow [17]63.780.340.261.4
VPT-Deep [17]74.982.953.970.6
VQT [45]72.784.549.368.8
SynQT (Ours)73.284.756.271.4
", + "bbox": [ + 263, + 184, + 743, + 426 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "1K. As can be seen from Table 1, the most striking observation emerging from the comparison is that SynQT outperforms other methods while saving up to $50\\%$ of GPU memory. Given the same batch size, previous PETL methods rely on the full backward propagation of the original model and consume $7 \\sim 8$ GB of memory, whereas SynQT disentangles transfer learning and reduces the memory cost to $3.4$ GB. Notably, our method performs better than VPT-Deep on 14 of 19 tasks. We notice that although the memory cost of SynQT is slightly higher, it achieves a significant performance improvement of $4.6\\%$ compared to VQT.", + "bbox": [ + 212, + 450, + 787, + 571 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "In order to further validate the robustness of our method, we compare SynQT with baseline models across various checkpoints which are derived from different learning paradigms, including supervised learning (pretrained over ImageNet-21K [9]), and self-supervised learning (DINO", + "bbox": [ + 212, + 571, + 787, + 631 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "[7] and MAE [14]). The results in Table 2 indicate that SynQT shows strong compatibility and outperforms PETL methods by a substantial margin over different backbones. For instance, compared to the strongest baselines, our method achieves improvements of $1.8\\%$ , $3.3\\%$ , and $0.8\\%$ on average with MAE, DINO, and supervised ImageNet-1K backbones, respectively.", + "bbox": [ + 212, + 631, + 540, + 768 + ], + "page_idx": 8 + }, + { + "type": "image", + "img_path": "images/ca231b864699c3f170917ef9f73482e8e8b38cb1d2960eea8e28b4cca27a12ec.jpg", + "image_caption": [ + "Fig.3: Performance comparison under memory constraints. SynQT outperforms other methods with different memory costs." + ], + "image_footnote": [], + "bbox": [ + 552, + 648, + 782, + 785 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "4.3 Memory-Efficiency Evaluation", + "text_level": 1, + "bbox": [ + 215, + 787, + 511, + 803 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "Previous PETL methods [8, 15, 16, 18] focus on reducing the number of parameters when", + "bbox": [ + 215, + 808, + 537, + 839 + ], + "page_idx": 8 + }, + { + "type": "header", + "text": "Parameter-Efficient and Memory-Efficient Tuning for Vision Transformer", + "bbox": [ + 245, + 114, + 732, + 128 + ], + "page_idx": 8 + }, + { + "type": "page_number", + "text": "9", + "bbox": [ + 774, + 114, + 784, + 126 + ], + "page_idx": 8 + }, + { + "type": "table", + "img_path": "images/c211f0350bfab81e39e15cd92d6edc2b3c32c551748568e3d7de60610bcbf66a.jpg", + "table_caption": [ + "Table 3: Full results on VTAB-1K benchmark under memory constraints. The memory is constrained within 5GB and “# Memory” specifies the memory cost during training with 64 as batch size on CIFAR100. We report the Top-1 accuracy and the results are averaged over group-wise mean values." + ], + "table_footnote": [], + "table_body": "
# Memory (GB)NaturalSpecializedStructured
CIFAR100Caltech101DTDFlower102PetsSVHNSun397CanelyonEuroSATResics45RetinopathyClevr-CountClevr-DistDMLabKITTI-DistdSpr-LocdSpr-OrisNORB-AzimsNORB-EteAverage
FacT-TT [18]4.340.484.259.896.886.653.445.879.492.273.374.459.250.242.568.848.225.919.828.763.2
Adapter [15]4.764.986.062.897.386.646.844.479.893.976.774.663.340.238.968.533.132.319.830.564.0
LoRA [16]4.862.582.160.497.586.658.645.380.695.378.872.864.845.141.368.549.628.124.630.565.5
AdaptFormer [8]4.665.686.162.096.986.235.545.682.696.277.674.466.247.741.372.355.629.422.633.965.8
SynQT(Ours)3.470.989.768.898.589.677.850.682.396.783.575.271.862.748.575.474.149.031.736.172.9
", + "bbox": [ + 218, + 203, + 784, + 318 + ], + "page_idx": 9 + }, + { + "type": "text", + "text": "fine-tuning the model. While decreasing the cost of model storage, these methods are not widely applicable in practical training because they still require a significant amount of GPU memory. Therefore, we propose SynQT to overcome this problem. In this section, we highlight our improvements in terms of memory costs by comparing our method with some SOTA PETL methods under a given GPU memory constraint.", + "bbox": [ + 212, + 347, + 782, + 436 + ], + "page_idx": 9 + }, + { + "type": "text", + "text": "For baseline methods, we discard some specific trainable structures to meet the memory constraint. Specifically, we remove these additional structures near the input layer, which means that the training only requires a partial backpropagation with a lower memory usage. We progressively remove the structure nearest to the input layer until it meets the given memory constraint. Besides, for a fair comparison, we reduce the memory usage of SynQT in the same manner.", + "bbox": [ + 212, + 438, + 784, + 527 + ], + "page_idx": 9 + }, + { + "type": "text", + "text": "We compare SynQT with SOTA baselines including FacT-TT [18], Adapter [15], LoRA [16] and AdaptFormer [8] under such a constraint and report the results in Table 3. While previous methods can achieve promising performance with few trainable parameters, they unfortunately suffer from significant performance drops when such memory constraints are imposed. In contrast, SynQT excels in this scenario, achieving state-of-the-art performance and surpassing AdaptFormer by $7.1\\%$ and VQT by $4.6\\%$ .", + "bbox": [ + 212, + 529, + 785, + 635 + ], + "page_idx": 9 + }, + { + "type": "text", + "text": "To have a better understanding of how memory constraints affect model performance, we estimate the results of the methods mentioned above across varying memory costs. As depicted in Figure 3, SynQT continues to outperform the baselines across different memory costs. Notably, when the constraint is very strict, i.e., 3 GB, most of the existing PETL methods perform poorly, with a gap of almost $7\\%$ compared to SynQT. Moreover, it is noteworthy that relaxing the memory constraints improves the performance of all the methods, especially for a significant improvement observed for our SynQT.", + "bbox": [ + 212, + 636, + 787, + 757 + ], + "page_idx": 9 + }, + { + "type": "text", + "text": "4.4 Few-Shot Learning on FGVC", + "text_level": 1, + "bbox": [ + 215, + 781, + 503, + 797 + ], + "page_idx": 9 + }, + { + "type": "text", + "text": "To further evaluate SynQT's ability to transfer knowledge from the pre-trained model using few samples, we compare SynQT with BitFit [54], VPT-Shallow [17],", + "bbox": [ + 212, + 809, + 785, + 842 + ], + "page_idx": 9 + }, + { + "type": "page_number", + "text": "10", + "bbox": [ + 217, + 114, + 235, + 126 + ], + "page_idx": 9 + }, + { + "type": "header", + "text": "Taolin Zhang, Jiawang Bai et al.", + "bbox": [ + 271, + 114, + 488, + 128 + ], + "page_idx": 9 + }, + { + "type": "image", + "img_path": "images/cf53836e9aaebae18596c30f1248303246867fe49cef0b84359871c698c159f5.jpg", + "image_caption": [ + "Fig. 4: Performance comparison in few-shot learning. Overall, SynQT outperforms other baselines in the extreme low-data regime." + ], + "image_footnote": [], + "bbox": [ + 222, + 143, + 781, + 272 + ], + "page_idx": 10 + }, + { + "type": "table", + "img_path": "images/9285a73bfda58c5a1c317b169a689e926b8434311a170e6a647e08bea4cb512e.jpg", + "table_caption": [ + "Table 4: Applying SynQT to Swin-B and ConvNeXt-B. We report the Top-1 accuracy and the averaged results." + ], + "table_footnote": [], + "table_body": "
MethodSwin-BConvNeXt-B
NaturalSpecializedStructuredAverageNaturalSpecializedStructuredAverage
Linear73.580.833.562.674.581.534.863.6
VPT-Deep76.884.553.471.678.583.044.668.7
SynQT77.586.057.873.878.182.148.269.5
", + "bbox": [ + 220, + 369, + 782, + 434 + ], + "page_idx": 10 + }, + { + "type": "text", + "text": "VPT-Deep [17], and VQT [45] in few-shot setting on 5 FGVC datasets. The results are reported in Figure 4.", + "bbox": [ + 212, + 460, + 784, + 491 + ], + "page_idx": 10 + }, + { + "type": "text", + "text": "Overall, SynQT significantly outperforms all other baselines in 4, 6, 8, and 16-shot scenarios, and its performance is comparable in the 1-shot scenario. The gaps between VQT and SynQT indicate that the lightweight, task-specific QSM enables SynQT to capture beneficial information from samples in the extremely low-data regime. Simply incorporating a trainable prompt as a query, as in VQT, is not sufficient to handle such challenging few-shot tasks. Interestingly, VPT-Shallow outperforms VPT-Deep in the low-data regime, due to the fact that the larger number of prompts in VPT-Deep are not well-trained on the limited data. This suggests that merely increasing the number of prompt numbers could potentially lead to a performance drop.", + "bbox": [ + 212, + 492, + 785, + 643 + ], + "page_idx": 10 + }, + { + "type": "text", + "text": "4.5 Versatility of SynQT", + "text_level": 1, + "bbox": [ + 214, + 664, + 433, + 680 + ], + "page_idx": 10 + }, + { + "type": "text", + "text": "To demonstrate the versatility of SynQT, we apply SynQT to different model architectures and combine SynQT with existing PETL methods.", + "bbox": [ + 212, + 688, + 784, + 718 + ], + "page_idx": 10 + }, + { + "type": "text", + "text": "Applying SynQT to different model architectures. Here, we apply SynQT to an advanced Transformer-like architecture (Swin [31]) and a convolutional neural network (ConvNeXt [32]). Following [17], we introduce prompts for each window in Swin and incorporate learnable prompt pixels for the feature map in ConvNeXt. Specifically, due to the varying hidden sizes and number of blocks in different layers, we extract features from the third layer of Swin and the last layer of ConvNext to ensure sufficient extraction of beneficial information. We find that both of them significantly benefit from SynQT, as shown in Table 4.", + "bbox": [ + 212, + 719, + 785, + 839 + ], + "page_idx": 10 + }, + { + "type": "header", + "text": "Parameter-Efficient and Memory-Efficient Tuning for Vision Transformer", + "bbox": [ + 245, + 114, + 732, + 128 + ], + "page_idx": 10 + }, + { + "type": "page_number", + "text": "11", + "bbox": [ + 767, + 114, + 782, + 126 + ], + "page_idx": 10 + }, + { + "type": "table", + "img_path": "images/222a00b7ca35462df68e3db11e2636134ae84f9b0738da944aafb436e95321a0.jpg", + "table_caption": [ + "Table 5: Combining different PETL methods. AF is short for AdaptFormer." + ], + "table_footnote": [], + "table_body": "
MethodsNaturalSpecializedStructuredAverage
AF80.182.350.370.9
AF+VQT79.684.353.072.3
AF+SynQT82.685.457.975.3
VPT79.184.654.472.7
VPT+VQT78.983.754.672.4
VPT+SynQT81.184.957.774.1
", + "bbox": [ + 218, + 186, + 509, + 281 + ], + "page_idx": 11 + }, + { + "type": "table", + "img_path": "images/478c2ad1893b38ee7237375105b08dd0fbe9ef71c85a3c02bbb491b7a06f746b.jpg", + "table_caption": [ + "Table 6: Necessity of disentangle-ment (KEM and QSM)." + ], + "table_footnote": [], + "table_body": "
MethodNatural SpecializedStructuredAverage
Linear63.277.131.457.2
+KEM71.880.132.461.4
+QSM77.883.655.072.1
+Head (SynQT)78.088.456.272.9
", + "bbox": [ + 517, + 203, + 782, + 263 + ], + "page_idx": 11 + }, + { + "type": "text", + "text": "On average, SynQT outperforms VPT-Deep by $2.2\\%$ and $1.2\\%$ for Swin and ConvNeXt, respectively.", + "bbox": [ + 212, + 310, + 782, + 339 + ], + "page_idx": 11 + }, + { + "type": "text", + "text": "Combining SynQT with existing PETL methods. We further jointly utilize SynQT and existing PETL methods to adapt pre-trained models, as shown in Table 5. The results demonstrate that both VPT and AdaptFormer benefit from SynQT on VTAB-1K, with improvements of $4.4\\%$ and $1.4\\%$ on average, respectively. We can see that for AdaptFormer and VPT, VQT may result in a performance drop in some cases (e.g., the Natural group), whereas equipping SynQT leads to significant performance improvements across all groups.", + "bbox": [ + 212, + 340, + 784, + 446 + ], + "page_idx": 11 + }, + { + "type": "text", + "text": "4.6 Ablation Study", + "text_level": 1, + "bbox": [ + 215, + 468, + 388, + 484 + ], + "page_idx": 11 + }, + { + "type": "text", + "text": "We ablate our SynQT with different settings in Tables 6 and 7 for a better understanding of SynQT.", + "bbox": [ + 212, + 492, + 782, + 521 + ], + "page_idx": 11 + }, + { + "type": "text", + "text": "Disentanglement. To demonstrate the necessity of disentanglement, we perform an ablation in Table 6. We set the baseline as linear probing since KEM exploits intermediate features. To gain deeper insights into the roles of QSM and Head, we employ random prompts to query in KEM as the model without QSM, and we utilize a simple average for aggregation as the model without Head. The results demonstrate that our idea of disentangling (QSM and KEM) is the key to achieving superior performance.", + "bbox": [ + 212, + 523, + 784, + 628 + ], + "page_idx": 11 + }, + { + "type": "text", + "text": "Hidden size in QSM. We vary the hidden size of the bottleneck module in QSM and report the results in Table 7a. With a larger hidden size, the performance improves due to the increasing capacity. However, when the hidden size exceeds 48, the benefit of introducing new parameters diminishes and overfitting may happen in datasets with few training samples. The results indicate that redundant parameters might not enhance the model's performance, and a hidden size of 48 is a reasonable choice.", + "bbox": [ + 212, + 628, + 784, + 733 + ], + "page_idx": 11 + }, + { + "type": "text", + "text": "Input of QSM. Table 7c studies the influence of different inputs of QSM. We can see that both the trainable prompts and the output from the last block are indispensable for QSM to generate high-quality task-specific query. The trainable prompts improve the capacity of QSM, and the results from the previous block ensure that knowledge is retained during the forward process.", + "bbox": [ + 212, + 734, + 784, + 808 + ], + "page_idx": 11 + }, + { + "type": "text", + "text": "Hierarchical features. In SynQT, we extract three different types of features from KSMs for classification: final output of KSM $H_{i}$ , attention features $F_{i}^{attn}$", + "bbox": [ + 212, + 809, + 784, + 840 + ], + "page_idx": 11 + }, + { + "type": "page_number", + "text": "12", + "bbox": [ + 217, + 114, + 235, + 126 + ], + "page_idx": 11 + }, + { + "type": "header", + "text": "Taolin Zhang, Jiawang Bai et al.", + "bbox": [ + 271, + 114, + 488, + 128 + ], + "page_idx": 11 + }, + { + "type": "table", + "img_path": "images/4e9befeff2f130fb258709d23110ffbe4da992c8a7571abb0c86f2b4a6a1320b.jpg", + "table_caption": [ + "Table 7: SynQT ablation experiments with ViT-B/16 with ImageNet-21K backbone. Default settings are marked in gray." + ], + "table_footnote": [ + "(a) Hidden size in QSM. 48 as the width is a trade-off between parameters and performance." + ], + "table_body": "
DimParamNat.Spe.Str.Avg.
162.14 M77.783.855.872.4
322.43 M77.783.955.372.3
482.73 M78.084.456.272.9
643.02 M77.684.456.172.7
963.61 M77.784.555.972.7
", + "bbox": [ + 238, + 188, + 493, + 268 + ], + "page_idx": 12 + }, + { + "type": "table", + "img_path": "images/8512d5d0e5e489cb3be34631f18b5114f60ba4ee5b35bb97a7613d3e99d8c0b5.jpg", + "table_caption": [], + "table_footnote": [ + "(b) Feature selection. Hierarchical features contains rich information that benefit to downstream tasks." + ], + "table_body": "
SynQTNat.Spe.Str.Avg.
w/o Hi78.384.053.872.0
w/o Fiattn78.383.555.472.4
w/o Fiuffn77.884.755.472.6
Full78.084.456.272.9
", + "bbox": [ + 517, + 191, + 774, + 258 + ], + "page_idx": 12 + }, + { + "type": "table", + "img_path": "images/2f785034f7e0390cf87005066c0654e9c8de16dfe0992096bfec8926699b9cdb.jpg", + "table_caption": [], + "table_footnote": [ + "(c) Input of QSM. Both the trainable prompt and the output from last block are important for generating task-specific query." + ], + "table_body": "
SynQTNat.Spe.Str.Avg.
w/o Prompt77.282.951.970.7
w/o Last Output77.584.452.371.4
Full78.084.456.272.9
", + "bbox": [ + 245, + 319, + 500, + 369 + ], + "page_idx": 12 + }, + { + "type": "table", + "img_path": "images/764bbe78b3bcda56925adc4aa2b5fe97cb03c573bc2efea4a529ca30ab6191fb.jpg", + "table_caption": [], + "table_footnote": [ + "(d) Projection in the classification head. Shared projection works the best for downstream tasks." + ], + "table_body": "
ProjectionNat.Spe.Str.Avg.
w/o Proj.78.483.855.072.4
Shared78.084.456.272.9
Independent77.984.455.572.6
", + "bbox": [ + 524, + 316, + 779, + 369 + ], + "page_idx": 12 + }, + { + "type": "table", + "img_path": "images/90071a74be65f7733bd1d1558b5d755ec36649d59f6274e9a803ccc1e0da1dde.jpg", + "table_caption": [], + "table_footnote": [ + "(e) Feature aggregation. Conditional feature aggregation is more effective than mean and learnable but fixed weight." + ], + "table_body": "
SynQTNat.Spe.Str.Avg.
Simply Avg.77.984.655.672.7
Fixed Weights77.984.755.572.7
Conditional78.084.456.272.9
", + "bbox": [ + 238, + 417, + 493, + 470 + ], + "page_idx": 12 + }, + { + "type": "table", + "img_path": "images/58e0203110fc1e969abde82639d021f5aeaae51845775de81c66cb23be7ea670.jpg", + "table_caption": [], + "table_footnote": [ + "(f) DropFeat in the classification head. The DropFeat module in the conditional head makes model more robust." + ], + "table_body": "
SynQTNat.Spe.Str.Avg.
w/o DropFeat77.784.355.272.4
w/ DropFeat78.084.456.272.9
", + "bbox": [ + 517, + 419, + 772, + 459 + ], + "page_idx": 12 + }, + { + "type": "text", + "text": "and FFN features $F_{i}^{ffn}$ , where $i \\in \\{1,2,\\dots,l\\}$ and $l$ is the number of layers. All these hierarchical features contain rich information and contribute to the performance boost of SynQT, as demonstrated in Table 7b. In particular, $H_{i}$ plays a more important role than other types of features.", + "bbox": [ + 212, + 547, + 784, + 609 + ], + "page_idx": 12 + }, + { + "type": "text", + "text": "Projection in the classification head. We vary the projection architecture in the classification head and investigate its performance in Table 7d. A projection of the input is necessary for the classification head, as hierarchical features across various layers may be misaligned in the representation space. However, to our surprise, a shared bottleneck structure is enough to handle these hierarchical features. The reason may be that independent projection for distinct features leads to overfitting, subsequently causing a slight drop in performance.", + "bbox": [ + 212, + 611, + 784, + 717 + ], + "page_idx": 12 + }, + { + "type": "text", + "text": "Conditional feature aggregation. In the classification head, we aggregate features using the weights conditioned on $H_{l}$ . We further conduct experiments over variants of SynQT with different aggregation methods including a simple average and weighted-sum with learnable but fixed weights. The results in Table 7e show that conditional aggregation improves the performance of SynQT especially for the Structured group.", + "bbox": [ + 212, + 718, + 784, + 808 + ], + "page_idx": 12 + }, + { + "type": "text", + "text": "DropFeat in the classification head. We study the effect of DropFeat, which is proposed to prevent overfitting by randomly dropping features in the classi-", + "bbox": [ + 212, + 809, + 784, + 839 + ], + "page_idx": 12 + }, + { + "type": "header", + "text": "Parameter-Efficient and Memory-Efficient Tuning for Vision Transformer", + "bbox": [ + 243, + 114, + 732, + 128 + ], + "page_idx": 12 + }, + { + "type": "page_number", + "text": "13", + "bbox": [ + 767, + 114, + 785, + 126 + ], + "page_idx": 12 + }, + { + "type": "text", + "text": "fication head. As can be observed from Table 7f, SynQT performs worse without DropFeat, indicating that DropFeat enhances the model's generalization", + "bbox": [ + 212, + 146, + 782, + 176 + ], + "page_idx": 13 + }, + { + "type": "text", + "text": "in handling rich hierarchical feature input. DropFeat ensures that the model does not overly rely on a specific feature during training and encourages the model to focus on a general way to select beneficial information from these features for downstream tasks.", + "bbox": [ + 212, + 176, + 486, + 296 + ], + "page_idx": 13 + }, + { + "type": "text", + "text": "4.7 t-SNE Visualization", + "text_level": 1, + "bbox": [ + 215, + 316, + 426, + 330 + ], + "page_idx": 13 + }, + { + "type": "text", + "text": "We present t-SNE visualization [35] of the CLS Token in pre-trained ViT, VQT, and our SynQT on EuroSAT, SVHN, and Clevr-Dist in Figure 5. The visualization shows that our SynQT achieves better feature clustering results compared to the", + "bbox": [ + 212, + 338, + 486, + 443 + ], + "page_idx": 13 + }, + { + "type": "text", + "text": "original CLS Token and VQT. We provide more visualization in Appendix.", + "bbox": [ + 212, + 443, + 754, + 459 + ], + "page_idx": 13 + }, + { + "type": "image", + "img_path": "images/a43dc0aa692ea145d2920a1b59c03454706cd6cde26a532bb8f6a9c86182c108.jpg", + "image_caption": [ + "Fig. 5: t-SNE visualization on SVHN, EuroSAT, and Clevr-Dist, one from each category of VTAB-1K. The features of our SynQT are more separable." + ], + "image_footnote": [], + "bbox": [ + 509, + 205, + 781, + 377 + ], + "page_idx": 13 + }, + { + "type": "text", + "text": "5 Conclusion, Limitation, and Future Work", + "text_level": 1, + "bbox": [ + 215, + 479, + 653, + 494 + ], + "page_idx": 13 + }, + { + "type": "text", + "text": "In this work, we present an insightful analysis of the high memory cost of existing PETL methods from a unique perspective, i.e., they entangle task-specific learning and pre-trained knowledge utilization together. Inspired by the idea of disentanglement, we propose a new paradigm, namely SynQT, consisting of two key structures: Query Synthesis Module (QSM) and Knowledge Extraction Module (KEM). QSM synthesizes task-specific query and then KEM extracts pre-trained knowledge in a query-only manner. Finally, a classification head is proposed to handle rich hierarchical features from KEMs and make predictions. Extensive experiments on VTAB-1K and FGVC demonstrate the effectiveness of SynQT under memory constraints and in an extremely low-data regime.", + "bbox": [ + 212, + 507, + 787, + 657 + ], + "page_idx": 13 + }, + { + "type": "text", + "text": "Despite the promising performance of our SynQT, it also has some limitations. One limitation is that SynQT requires slightly more trainable parameters than previous PETL methods due to a few projection layers and FFN operations, which is analyzed in Appendix. However, we would like to emphasize that as shown in their original papers, for VPT-Deep [17] and VQT [45], straightforwardly increasing the number of tokens to use more trainable parameters can not consistently result in better performance.", + "bbox": [ + 212, + 659, + 787, + 763 + ], + "page_idx": 13 + }, + { + "type": "text", + "text": "We believe that our work will inspire the community to explore memory-efficient PETL along with our disentangled perspective. Besides, to overcome the aforementioned limitation, one future direction is to develop PETL methods that achieve high accuracy and meanwhile have extreme memory efficiency and parameter efficiency.", + "bbox": [ + 212, + 765, + 787, + 839 + ], + "page_idx": 13 + }, + { + "type": "page_number", + "text": "14", + "bbox": [ + 217, + 114, + 235, + 126 + ], + "page_idx": 13 + }, + { + "type": "header", + "text": "Taolin Zhang, Jiawang Bai et al.", + "bbox": [ + 271, + 114, + 488, + 128 + ], + "page_idx": 13 + }, + { + "type": "text", + "text": "Acknowledgments", + "text_level": 1, + "bbox": [ + 217, + 145, + 392, + 162 + ], + "page_idx": 14 + }, + { + "type": "text", + "text": "This work is supported in part by the National Natural Science Foundation of China under Grant 62171248, Shenzhen Science and Technology Program (JCYJ20220818101012025), the PCNL KEY project (PCL2023AS6-1), and the National Research Foundation, Singapore, under its AI Singapore Programme (AISG Award No: AISG2-RP-2021-023).", + "bbox": [ + 212, + 175, + 787, + 251 + ], + "page_idx": 14 + }, + { + "type": "text", + "text": "References", + "text_level": 1, + "bbox": [ + 217, + 273, + 321, + 287 + ], + "page_idx": 14 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "1. Argyriou, A., Evgeniou, T., Pontil, M.: Multi-task feature learning. Advances in neural information processing systems 19 (2006)", + "2. Ba, J.L., Kiros, J.R., Hinton, G.E.: Layer normalization. arXiv preprint arXiv:1607.06450 (2016)", + "3. Bai, J., Gao, K., Min, S., Xia, S.T., Li, Z., Liu, W.: Badclip: Trigger-aware prompt learning for backdoor attacks on clip. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 24239-24250 (2024)", + "4. Bai, J., Yuan, L., Xia, S.T., Yan, S., Li, Z., Liu, W.: Improving vision transformers by revisiting high-frequency components. In: European Conference on Computer Vision. pp. 1-18. Springer (2022)", + "5. Bossard, L., Guillaumeun, M., Van Gool, L.: Food-101-mining discriminative components with random forests. In: Computer Vision-ECCV 2014: 13th European Conference, Zurich, Switzerland, September 6-12, 2014, Proceedings, Part VI 13. pp. 446-461. Springer (2014)", + "6. Brown, T., Mann, B., Ryder, N., Subbiah, M., Kaplan, J.D., Dhariwal, P., Neelakantan, A., Shyam, P., Sastry, G., Askell, A., et al.: Language models are few-shot learners. NeurIPS (2020)", + "7. Caron, M., Touvron, H., Misra, I., Jégou, H., Mairal, J., Bojanowski, P., Joulin, A.: Emerging properties in self-supervised vision transformers. In: Proceedings of the IEEE/CVF international conference on computer vision. pp. 9650-9660 (2021)", + "8. Chen, S., Ge, C., Tong, Z., Wang, J., Song, Y., Wang, J., Luo, P.: Adaptformer: Adapting vision transformers for scalable visual recognition. Advances in Neural Information Processing Systems 35, 16664-16678 (2022)", + "9. Deng, J., Dong, W., Socher, R., Li, L.J., Li, K., Fei-Fei, L.: Imagenet: A large-scale hierarchical image database. In: 2009 IEEE conference on computer vision and pattern recognition. pp. 248-255. IEEE (2009)", + "10. Dosovitskiy, A., Beyer, L., Kolesnikov, A., Weissenborn, D., Zhai, X., Unterthiner, T., Dehghani, M., Minderer, M., Heigold, G., Gelly, S., et al.: An image is worth 16x16 words: Transformers for image recognition at scale. arXiv preprint arXiv:2010.11929 (2020)", + "11. Evci, U., Dumoulin, V., Larochelle, H., Mozer, M.C.: Head2toe: Utilizing intermediate representations for better transfer learning. In: International Conference on Machine Learning. pp. 6009-6033. PMLR (2022)", + "12. Guo, H., Dai, T., Bai, Y., Chen, B., Xia, S.T., Zhu, Z.: Adaptir: Parameter efficient multi-task adaptation for pre-trained image restoration models. arXiv preprint arXiv:2312.08881 (2023)", + "13. Guo, H., Li, J., Dai, T., Ouyang, Z., Ren, X., Xia, S.T.: Mambair: A simple baseline for image restoration with state-space model. arXiv preprint arXiv:2402.15648 (2024)" + ], + "bbox": [ + 218, + 301, + 784, + 839 + ], + "page_idx": 14 + }, + { + "type": "header", + "text": "Parameter-Efficient and Memory-Efficient Tuning for Vision Transformer", + "bbox": [ + 243, + 114, + 730, + 128 + ], + "page_idx": 14 + }, + { + "type": "page_number", + "text": "15", + "bbox": [ + 767, + 116, + 784, + 126 + ], + "page_idx": 14 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "14. He, K., Chen, X., Xie, S., Li, Y., Dollár, P., Girshick, R.: Masked autoencoders are scalable vision learners. In: Proceedings of the IEEE/CVF conference on computer vision and pattern recognition. pp. 16000-16009 (2022)", + "15. Houlsby, N., Giurgiu, A., Jastrzebski, S., Morrone, B., De Laroussilhe, Q., Gesmundo, A., Attariyan, M., Gelly, S.: Parameter-efficient transfer learning for nlp. In: International Conference on Machine Learning. pp. 2790-2799. PMLR (2019)", + "16. Hu, E.J., Wallis, P., Allen-Zhu, Z., Li, Y., Wang, S., Wang, L., Chen, W., et al.: Lora: Low-rank adaptation of large language models. In: International Conference on Learning Representations (2021)", + "17. Jia, M., Tang, L., Chen, B.C., Cardie, C., Belongie, S., Hariharan, B., Lim, S.N.: Visual prompt tuning. In: European Conference on Computer Vision. pp. 709-727. Springer (2022)", + "18. Jie, S., Deng, Z.H.: Fact: Factor-tuning for lightweight adaptation on vision transformer. In: Proceedings of the AAAI Conference on Artificial Intelligence. vol. 37, pp. 1060-1068 (2023)", + "19. Jie, S., Wang, H., Deng, Z.H.: Revisiting the parameter efficiency of adapters from the perspective of precision redundancy. In: Proceedings of the IEEE/CVF International Conference on Computer Vision. pp. 17217-17226 (2023)", + "20. Kirillov, A., Mintun, E., Ravi, N., Mao, H., Rolland, C., Gustafson, L., Xiao, T., Whitehead, S., Berg, A.C., Lo, W.Y., et al.: Segment anything. arXiv preprint arXiv:2304.02643 (2023)", + "21. Kornblith, S., Shlens, J., Le, Q.V.: Do better imagenet models transfer better? In: Proceedings of the IEEE/CVF conference on computer vision and pattern recognition. pp. 2661-2671 (2019)", + "22. Krause, J., Stark, M., Deng, J., Fei-Fei, L.: 3d object representations for fine-grained categorization. In: Proceedings of the IEEE international conference on computer vision workshops. pp. 554-561 (2013)", + "23. Larsson, G., Maire, M., Shakhnarovich, G.: Fractalnet: Ultra-deep neural networks without residuals. In: ICLR (2017)", + "24. Lester, B., Al-Rfou, R., Constant, N.: The power of scale for parameter-efficient prompt tuning. arXiv preprint arXiv:2104.08691 (2021)", + "25. Li, X.L., Liang, P.: Prefix-tuning: Optimizing continuous prompts for generation. arXiv preprint arXiv:2101.00190 (2021)", + "26. Li, X., Lian, D., Lu, Z., Bai, J., Chen, Z., Wang, X.: Graphadapter: Tuning vision-language models with dual knowledge graph. Advances in Neural Information Processing Systems 36 (2024)", + "27. Lian, D., Zhou, D., Feng, J., Wang, X.: Scaling & shifting your features: A new baseline for efficient model tuning. Advances in Neural Information Processing Systems 35, 109-123 (2022)", + "28. Liu, H., Tam, D., Muqeeth, M., Mohta, J., Huang, T., Bansal, M., Raffel, C.A.: Few-shot parameter-efficient fine-tuning is better and cheaper than in-context learning. NeurIPS (2022)", + "29. Liu, P., Yuan, W., Fu, J., Jiang, Z., Hayashi, H., Neubig, G.: Pre-train, prompt, and predict: A systematic survey of prompting methods in natural language processing. ACM Computing Surveys 55(9), 1-35 (2023)", + "30. Liu, X., Ji, K., Fu, Y., Tam, W.L., Du, Z., Yang, Z., Tang, J.: P-tuning v2: Prompt tuning can be comparable to fine-tuning universally across scales and tasks. arXiv preprint arXiv:2110.07602 (2021)", + "31. Liu, Z., Lin, Y., Cao, Y., Hu, H., Wei, Y., Zhang, Z., Lin, S., Guo, B.: Swin transformer: Hierarchical vision transformer using shifted windows. In: Proceedings" + ], + "bbox": [ + 215, + 146, + 785, + 839 + ], + "page_idx": 15 + }, + { + "type": "page_number", + "text": "16", + "bbox": [ + 217, + 114, + 235, + 126 + ], + "page_idx": 15 + }, + { + "type": "header", + "text": "Taolin Zhang, Jiawang Bai et al.", + "bbox": [ + 271, + 114, + 488, + 128 + ], + "page_idx": 15 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "of the IEEE/CVF international conference on computer vision. pp. 10012-10022 (2021)", + "32. Liu, Z., Mao, H., Wu, C.Y., Feichtenhofer, C., Darrell, T., Xie, S.: A convnet for the 2020s. In: Proceedings of the IEEE/CVF conference on computer vision and pattern recognition. pp. 11976-11986 (2022)", + "33. Loshchilov, I., Hutter, F.: Decoupled weight decay regularization. arXiv preprint arXiv:1711.05101 (2017)", + "34. Lu, Z., Bai, J., Li, X., Xiao, Z., Wang, X.: Beyond sole strength: Customized ensembles for generalized vision-language models. In: Forty-first International Conference on Machine Learning (2024)", + "35. Van der Maaten, L., Hinton, G.: Visualizing data using t-sne. Journal of machine learning research 9(11) (2008)", + "36. Maji, S., Rahtu, E., Kannala, J., Blaschko, M., Vedaldi, A.: Fine-grained visual classification of aircraft. arXiv preprint arXiv:1306.5151 (2013)", + "37. Nilsback, M.E., Zisserman, A.: A visual vocabulary for flower classification. In: 2006 IEEE computer society conference on computer vision and pattern recognition (CVPR'06). vol. 2, pp. 1447-1454. IEEE (2006)", + "38. Parkhi, O.M., Vedaldi, A., Zisserman, A., Jawahar, C.: Cats and dogs. In: 2012 IEEE conference on computer vision and pattern recognition. pp. 3498-3505. IEEE (2012)", + "39. Pfeiffer, J., Kamath, A., Rückle, A., Cho, K., Gurevych, I.: Adapterfusion: Nondestructive task composition for transfer learning. arXiv preprint arXiv:2005.00247 (2020)", + "40. Rebuffi, S.A., Bilen, H., Vedaldi, A.: Learning multiple visual domains with residual adapters. Advances in neural information processing systems 30 (2017)", + "41. Shao, L., Zhu, F., Li, X.: Transfer learning for visual categorization: A survey. IEEE transactions on neural networks and learning systems 26(5), 1019-1034 (2014)", + "42. Srivastava, N., Hinton, G., Krizhevsky, A., Sutskever, I., Salakhutdinov, R.: Dropout: a simple way to prevent neural networks from overfitting. The journal of machine learning research 15(1), 1929-1958 (2014)", + "43. Sung, Y.L., Cho, J., Bansal, M.: Lst: Ladder side-tuning for parameter and memory efficient transfer learning. Advances in Neural Information Processing Systems 35, 12991-13005 (2022)", + "44. Sung, Y.L., Cho, J., Bansal, M.: Vl-adapter: Parameter-efficient transfer learning for vision-and-language tasks. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 5227-5237 (2022)", + "45. Tu, C.H., Mai, Z., Chao, W.L.: Visual query tuning: Towards effective usage of intermediate representations for parameter and memory efficient transfer learning. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 7725-7735 (2023)", + "46. Vaswani, A., Shazeer, N., Parmar, N., Uszkoreit, J., Jones, L., Gomez, A.N., Kaiser, L., Polosukhin, I.: Attention is all you need. Advances in neural information processing systems 30 (2017)", + "47. Wan, L., Zeiler, M., Zhang, S., Le Cun, Y., Fergus, R.: Regularization of neural networks using dropconnect. In: International conference on machine learning. pp. 1058-1066. PMLR (2013)", + "48. Wang, M., Deng, W.: Deep visual domain adaptation: A survey. Neurocomputing 312, 135-153 (2018)", + "49. Wang, W., Bao, H., Dong, L., Bjorck, J., Peng, Z., Liu, Q., Aggarwal, K., Mohammed, O.K., Singhal, S., Som, S., et al.: Image as a foreign language: Beit pre" + ], + "bbox": [ + 215, + 146, + 785, + 839 + ], + "page_idx": 16 + }, + { + "type": "header", + "text": "Parameter-Efficient and Memory-Efficient Tuning for Vision Transformer", + "bbox": [ + 245, + 114, + 730, + 128 + ], + "page_idx": 16 + }, + { + "type": "page_number", + "text": "17", + "bbox": [ + 767, + 114, + 785, + 126 + ], + "page_idx": 16 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "training for all vision and vision-language tasks. arXiv preprint arXiv:2208.10442 (2022)", + "50. Yang, S., Bai, J., Gao, K., Yang, Y., Li, Y., Xia, S.T.: Not all prompts are secure: A switchable backdoor attack against pre-trained vision transformers. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 24431-24441 (2024)", + "51. Yang, X., Ye, J., Wang, X.: Factorizing knowledge in neural networks. In: European Conference on Computer Vision. pp. 73-91. Springer (2022)", + "52. Yang, X., Zhou, D., Liu, S., Ye, J., Wang, X.: Deep model reassembly. Advances in neural information processing systems 35, 25739-25753 (2022)", + "53. Yuan, M., Lin, Y.: Model selection and estimation in regression with grouped variables. Journal of the Royal Statistical Society Series B: Statistical Methodology 68(1), 49-67 (2006)", + "54. Zaken, E.B., Ravfogel, S., Goldberg, Y.: Bitfit: Simple parameter-efficient fine-tuning for transformer-based masked language-models. arXiv preprint arXiv:2106.10199 (2021)", + "55. Zang, Y., Li, W., Zhou, K., Huang, C., Loy, C.C.: Unified vision and language prompt learning. arXiv preprint arXiv:2210.07225 (2022)", + "56. Zha, Y., Wang, J., Dai, T., Chen, B., Wang, Z., Xia, S.T.: Instance-aware dynamic prompt tuning for pre-trained point cloud models. In: Proceedings of the IEEE/CVF International Conference on Computer Vision. pp. 14161-14170 (2023)", + "57. Zhai, X., Kolesnikov, A., Houlsby, N., Beyer, L.: Scaling vision transformers. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 12104-12113 (2022)", + "58. Zhai, X., Puigcerver, J., Kolesnikov, A., Ruyssen, P., Riquelme, C., Lucic, M., Djolonga, J., Pinto, A.S., Neumann, M., Dosovitskiy, A., et al.: The visual task adaptation benchmark (2019)", + "59. Zhang, Y., Zhou, K., Liu, Z.: Neural prompt search. arXiv (2022)", + "60. Zhou, K., Yang, J., Loy, C.C., Liu, Z.: Conditional prompt learning for vision-language models. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 16816-16825 (2022)", + "61. Zhou, K., Yang, J., Loy, C.C., Liu, Z.: Learning to prompt for vision-language models. International Journal of Computer Vision 130(9), 2337-2348 (2022)", + "62. Zhu, L., Liao, B., Zhang, Q., Wang, X., Liu, W., Wang, X.: Vision mamba: Efficient visual representation learning with bidirectional state space model. arXiv preprint arXiv:2401.09417 (2024)" + ], + "bbox": [ + 215, + 146, + 785, + 646 + ], + "page_idx": 17 + }, + { + "type": "page_number", + "text": "18", + "bbox": [ + 217, + 114, + 235, + 126 + ], + "page_idx": 17 + }, + { + "type": "header", + "text": "Taolin Zhang, Jiawang Bai et al.", + "bbox": [ + 271, + 114, + 488, + 128 + ], + "page_idx": 17 + } +] \ No newline at end of file diff --git a/2024/Parameter-Efficient and Memory-Efficient Tuning for Vision Transformer_ A Disentangled Approach/e724c6a7-eae9-4271-a988-9b07cf4a65fc_model.json b/2024/Parameter-Efficient and Memory-Efficient Tuning for Vision Transformer_ A Disentangled Approach/e724c6a7-eae9-4271-a988-9b07cf4a65fc_model.json new file mode 100644 index 0000000000000000000000000000000000000000..297c4ff280b59bb8780278b6afc4eb1378d8d312 --- /dev/null +++ b/2024/Parameter-Efficient and Memory-Efficient Tuning for Vision Transformer_ A Disentangled Approach/e724c6a7-eae9-4271-a988-9b07cf4a65fc_model.json @@ -0,0 +1,2678 @@ +[ + [ + { + "type": "title", + "bbox": [ + 0.217, + 0.141, + 0.787, + 0.187 + ], + "angle": 0, + "content": "Parameter-Efficient and Memory-Efficient Tuning for Vision Transformer: A Disentangled Approach" + }, + { + "type": "text", + "bbox": [ + 0.218, + 0.212, + 0.785, + 0.245 + ], + "angle": 0, + "content": "Taolin Zhang\\(^{1*}\\), Jiawang Bai\\(^{2,3*}\\), Zhihe Lu\\(^{3}\\), Dongze Lian\\(^{3}\\), Genping Wang\\(^{4,\\text{凹}}\\), Xinchao Wang\\(^{3,\\text{凹}}\\), and Shu-Tao Xia\\(^{1,5}\\)" + }, + { + "type": "text", + "bbox": [ + 0.257, + 0.255, + 0.754, + 0.298 + ], + "angle": 0, + "content": "\\(^{1}\\)Tsinghua Shenzhen International Graduate School, Tsinghua University \\(^{2}\\)Tencent \\(^{3}\\)National University of Singapore \\(^{4}\\)Shenzhen Qiji Technology Co., Ltd." + }, + { + "type": "text", + "bbox": [ + 0.25, + 0.298, + 0.754, + 0.354 + ], + "angle": 0, + "content": "\\(^{5}\\)Research Center of Artificial Intelligence, Peng Cheng Laboratory ztl23@mails.tsinghua.edu.cn; jiawangbai@tencent.com; zhihelu.academic@gmail.com; dzlianx@gmail.com; genpingwang@163.com; xinchao@nus.edu.sg; xiast@sz.tsinghua.edu.cn" + }, + { + "type": "text", + "bbox": [ + 0.261, + 0.392, + 0.744, + 0.683 + ], + "angle": 0, + "content": "Abstract. Recent works on parameter-efficient transfer learning (PETL) show the potential to adapt a pre-trained Vision Transformer to downstream recognition tasks with only a few learnable parameters. However, since they usually insert new structures into the pre-trained model, entire intermediate features of that model are changed and thus need to be stored to be involved in back-propagation, resulting in memory-heavy training. We solve this problem from a novel disentangled perspective, i.e., dividing PETL into two aspects: task-specific learning and pretrained knowledge utilization. Specifically, we synthesize the task-specific query with a learnable and lightweight module, which is independent of the pre-trained model. The synthesized query equipped with task-specific knowledge serves to extract the useful features for downstream tasks from the intermediate representations of the pre-trained model in a query-only manner. Built upon these features, a customized classification head is proposed to make the prediction for the input sample. Given that our method employs an extremely lightweight architecture and avoids the use of heavy intermediate features for running gradient descent, it demonstrates limited memory usage in training. Extensive experiments manifest that our method achieves state-of-the-art performance under memory constraints, showcasing its applicability in real-world situations. The code is available at: https://synqt.github.io/." + }, + { + "type": "title", + "bbox": [ + 0.217, + 0.71, + 0.377, + 0.725 + ], + "angle": 0, + "content": "1 Introduction" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.742, + 0.788, + 0.803 + ], + "angle": 0, + "content": "Motivated by the success of data-driven solutions, there has been an explosive growth in the size of vision models [4,10,13,20,31,49,62] and the scale of accessible datasets like ImageNet [9]. To unleash the power of large models, it becomes a tendency to first pre-train on a general and large-scale dataset, and then adapt" + }, + { + "type": "page_footnote", + "bbox": [ + 0.232, + 0.812, + 0.411, + 0.841 + ], + "angle": 0, + "content": "* Equal contribution. \n\\(\\boxtimes\\) Corresponding authors." + } + ], + [ + { + "type": "page_number", + "bbox": [ + 0.218, + 0.116, + 0.23, + 0.127 + ], + "angle": 0, + "content": "2" + }, + { + "type": "header", + "bbox": [ + 0.272, + 0.115, + 0.49, + 0.129 + ], + "angle": 0, + "content": "Taolin Zhang, Jiawang Bai et al." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.147, + 0.784, + 0.208 + ], + "angle": 0, + "content": "them to downstream tasks [21, 41, 48, 52]. The most obvious adaption strategy is to fully fine-tune all parameters of the pre-trained model. However, it necessitates that the device saves a dedicated set of model parameters for each single task, leading to substantial storage overhead [57]." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.21, + 0.788, + 0.346 + ], + "angle": 0, + "content": "To improve the storage efficiency, parameter-efficient transfer learning (PETL) methods [8,16-19,59] have been explored by tuning only a small set of the parameters in adapting the large pre-trained models to downstream tasks. Typically, these methods insert small structures into the frozen pre-trained model and specify the newly added parameters as the only learnable ones. For instance, Visual Prompt Tuning (VPT) [17] prepends learnable tokens to the patch embeddings; AdaptFormer [8] inserts a bottleneck-structured fully connected layers parallel to the feed-forward network (FFN) of the original ViT. These studies demonstrate that PETL is a promising way to save storage resources." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.349, + 0.788, + 0.455 + ], + "angle": 0, + "content": "However, even though updating a small set of parameters, existing PETL methods still require a very high consumption of memory in training. The direct reason is that they need a full back-propagation throughout the entire backbone. Given the fact that the model size is exponentially increasing, it becomes infeasible to apply existing PETL methods to the memory-constraint situations. Therefore, it is desirable to investigate memory-efficient PETL while maintaining performance on downstream tasks." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.458, + 0.788, + 0.563 + ], + "angle": 0, + "content": "In this work, we attribute the high memory cost of existing PETL methods in training to their intrinsic regime, i.e., entangling task-specific learning and pre-trained knowledge utilization together. It is usually achieved by integrating trainable structures into the pre-trained ViT and changing all intermediate features. Consequently, a great number of intermediate features need to be stored in training to run gradient descent. The aforementioned analysis motivates us to design a memory-efficient PETL method from a disentangled perspective." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.566, + 0.79, + 0.793 + ], + "angle": 0, + "content": "Accordingly, we propose Synthesized Query Tuning (SynQT), a memory-efficient PETL method for adapting pre-trained ViTs to downstream recognition tasks. Overall, as shown in Figure 1, SynQT takes the intermediate representations of the original pre-trained model as input and combines them with task-specific knowledge using stacked blocks. We propose the Query Synthesis Module (QSM) and Knowledge Extraction Module (KEM) to accomplish the disentanglement. To capture task-specific knowledge, SynQT employs a lightweight QSM, which is independent of the pre-trained model. In QSM, an attention layer and an FFN are trained on the target downstream task to synthesize task-specific query. Then, to utilize the pre-trained knowledge, the synthesized query interacts with the intermediate representations of the original pre-trained model in a KEM. It outputs useful features for downstream recognition, where all parameters are inherited from the pre-trained model and frozen. Relying on rich hierarchical features from KEMs, we propose a classification head, which adaptively aggregates these features conditioned on the input sample for prediction." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.796, + 0.788, + 0.84 + ], + "angle": 0, + "content": "The success of SynQT in saving memory in training is derived from two technical aspects. Firstly, we keep intermediate representations of the original pretrained model intact and directly exploit them in a query-only manner, avoiding" + } + ], + [ + { + "type": "header", + "bbox": [ + 0.245, + 0.115, + 0.732, + 0.129 + ], + "angle": 0, + "content": "Parameter-Efficient and Memory-Efficient Tuning for Vision Transformer" + }, + { + "type": "page_number", + "bbox": [ + 0.775, + 0.117, + 0.785, + 0.127 + ], + "angle": 0, + "content": "3" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.147, + 0.784, + 0.207 + ], + "angle": 0, + "content": "the use of heavy intermediate features for running gradient descent. Besides, our design including QSM and KEM is extremely lightweight. For instance, we adopt bottleneck structures for Q-K-V projection [46] and a small number of tokens (lower than 4 in most cases) in QSM." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.208, + 0.785, + 0.313 + ], + "angle": 0, + "content": "We validate SynQT on VTAB-1K [58] benchmark, including 19 tasks from diverse domains. We also evaluate our method in few-shot learning using 5 fine-grained visual classification datasets. Extensive experiments show that SynQT can significantly reduce the memory footprint in training with promising accuracy. Compared to VPT-Deep, a powerful version of VPT that inserts prompts in each layer of ViT, SynQT saves \\(56\\%\\) memory consumption and achieves higher Top-1 accuracy on average. The contributions of our work are as follows:" + }, + { + "type": "text", + "bbox": [ + 0.223, + 0.326, + 0.784, + 0.37 + ], + "angle": 0, + "content": "1. We provide an insightful analysis of the high memory cost of existing PETL methods, i.e., they entangle task-specific learning and pre-trained knowledge utilization together." + }, + { + "type": "text", + "bbox": [ + 0.223, + 0.372, + 0.785, + 0.416 + ], + "angle": 0, + "content": "2. Inspired by the disentangled idea, we propose SynQT, which synthesizes task-specific query and then utilizes the pre-trained knowledge in a query-only manner." + }, + { + "type": "text", + "bbox": [ + 0.223, + 0.418, + 0.784, + 0.448 + ], + "angle": 0, + "content": "3. Extensive experiments show that SynQT is highly memory-efficient and effective, notably outperforming existing baselines under memory constraints." + }, + { + "type": "list", + "bbox": [ + 0.223, + 0.326, + 0.785, + 0.448 + ], + "angle": 0, + "content": null + }, + { + "type": "title", + "bbox": [ + 0.216, + 0.473, + 0.388, + 0.49 + ], + "angle": 0, + "content": "2 Related Work" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.508, + 0.787, + 0.84 + ], + "angle": 0, + "content": "Parameter-Efficient Transfer Learning (PETL). The idea of PETL initially shows great success in the field of natural language processing (NLP) [6,16,28,40]. It fine-tunes a small number of trainable parameters to transfer large pre-trained models to downstream tasks, significantly reducing storage costs by avoiding a copy of model parameters for each new task. Due to the increased size of pre-trained vision models, especially for ViT [10], PETL has been introduced to computer vision and attracted much attention [3,8,12,17,26,27,34,39,44, 45,50,55,56,60,61]. To reduce the number of fine-tuned parameters, BitFit [54] only fine-tunes the bias terms and freezes most of the network, resulting in comparable performance to full tuning. VPT [17] introduces learnable tokens (i.e., prompts [24,25,29,30]) to each layer of ViT. SSF [27] fine-tunes scale and shift factors in each layer to adapt original features to the target domain. Some other works [8,15,16] propose to insert additional trainable structures into attention blocks of the pre-trained Transformer. For instance, LoRA [16] observes that the learned over-parametrized models reside on a low-rank dimension and further injects trainable rank decomposition parameters into each layer of the Transformer; AdaptFormer [8] adds parallel lightweight module to FFN. Furthermore, for extreme parameter efficiency, a tensorization-decomposition framework [18] is exploited to store the weight increments, and the low-bit adapter [19] is proposed to reduce the precision redundancy. However, how to improve memory efficiency in training is still an under-explored problem in PETL, as discussed below." + } + ], + [ + { + "type": "page_number", + "bbox": [ + 0.218, + 0.117, + 0.23, + 0.127 + ], + "angle": 0, + "content": "4" + }, + { + "type": "header", + "bbox": [ + 0.272, + 0.115, + 0.49, + 0.129 + ], + "angle": 0, + "content": "Taolin Zhang, Jiawang Bai et al." + }, + { + "type": "image", + "bbox": [ + 0.234, + 0.145, + 0.773, + 0.301 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.214, + 0.311, + 0.788, + 0.382 + ], + "angle": 0, + "content": "Fig. 1: Overall architecture of SynQT. Each block in SynQT contains two key designs, including Query Synthesis Module (QSM) and Knowledge Extractor Module (KEM). QSM is a lightweight learnable module to capture task-specific knowledge and KEM reuses frozen weights in the pre-trained model for feature extraction. The rich and hierarchical features from KEMs are fed into the classification head." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.412, + 0.789, + 0.623 + ], + "angle": 0, + "content": "Memory-Efficient PETL. Although existing PETL methods have shown promising performance in transfer learning while saving most trainable parameters, they still suffer from issues of high memory consumption in training. The full back-propagation throughout the entire backbone prevents traditional PETL methods from being applied in memory-constraint situations. Recently, some recent works [11,43,45] have made several attempts to address this issue, whose key idea is to leverage the power of intermediate representations of the pre-trained model and thus avoid full back-propagation. In NLP, LST [43] proposes a trainable lightweight structure in the sidewalk for memory-saving during training. To adapt vision models, Head2Toe [11] selects features from all layers of the source model to feed into a classification head. VQT [45] inserts trainable tokens in the query space to capture useful information from the pre-trained model. Due to the information redundancy in the intermediate representations, both Head2Toe and VQT further adopt group lasso [1,53] for feature selection." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.624, + 0.788, + 0.73 + ], + "angle": 0, + "content": "In contrast to previous works, we provide an insightful analysis that entangled combination of task-specific learning and pre-trained knowledge utilization results in high memory consumption of traditional PETL methods. Therefore, we propose SynQT to disentangle PETL into two distinct modules, with the aim of capturing task-specific knowledge [51] and extracting useful features for downstream tasks, respectively. Such a disentangled paradigm enjoys memory efficiency in transfer learning compared with traditional PETL methods." + }, + { + "type": "title", + "bbox": [ + 0.216, + 0.753, + 0.331, + 0.768 + ], + "angle": 0, + "content": "3 Method" + }, + { + "type": "title", + "bbox": [ + 0.216, + 0.785, + 0.421, + 0.799 + ], + "angle": 0, + "content": "3.1 Overall Framework" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.81, + 0.787, + 0.84 + ], + "angle": 0, + "content": "As can be seen from Figures 1 and 2, SynQT makes the prediction relying on the features produced by stacked blocks, which take the intermediate representations" + } + ], + [ + { + "type": "header", + "bbox": [ + 0.245, + 0.115, + 0.733, + 0.129 + ], + "angle": 0, + "content": "Parameter-Efficient and Memory-Efficient Tuning for Vision Transformer" + }, + { + "type": "page_number", + "bbox": [ + 0.775, + 0.117, + 0.785, + 0.127 + ], + "angle": 0, + "content": "5" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.147, + 0.788, + 0.284 + ], + "angle": 0, + "content": "of the original pre-trained model as external input. In SynQT, each block mainly includes two parts: Query Synthesis Module (QSM) and Knowledge Extraction Module (KEM). QSM is a lightweight module, which is used to capture task-specific knowledge, connecting the pre-trained model with the downstream task. KEM uses the tokens generated by QSM as queries, and the intermediate features in the pre-trained ViT as keys and values. It reuses the original model's weights to extract features that are beneficial to downstream tasks. Finally, the rich hierarchical features extracted by the KEMs are fed into the classification head, as shown in Figure 2. We will elaborate our designs in the following sections." + }, + { + "type": "title", + "bbox": [ + 0.215, + 0.306, + 0.466, + 0.321 + ], + "angle": 0, + "content": "3.2 Query Synthesis Module" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.33, + 0.788, + 0.451 + ], + "angle": 0, + "content": "PETL integrates the task-specific knowledge into the adapted model through training on the downstream dataset. Typically, existing methods [8, 15-17] insert additional trainable structures into the pre-trained model. However, this is the root of high memory consumption, since it changes all original features, and all intermediate features are modified and stored to be involved in calculating gradients. From the disentangled perspective, we solve this problem with a task-specific learning module, named Query Synthesis Module (QSM), which is trainable, lightweight, and structurally independent from the pre-trained model." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.452, + 0.788, + 0.602 + ], + "angle": 0, + "content": "The purpose of QSM is to synthesize query tokens, interacting with intermediate representations of the original pre-trained model. Besides outputs from the previous block, we introduce a trainable prompt into the input space of each QSM. These per-block additional prompts increase the capacity of QSM to capture task-specific knowledge. Inspired by the Transformer, QSM includes an attention layer and an FFN. Specifically, the input tokens interact with each other through a self-attention layer and are further fed into the feed-forward network for producing the query tokens. Besides, to keep the parameter efficiency, another design principle is to extensively use bottleneck-structured fully connected layers to reduce the number of parameters." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.602, + 0.788, + 0.693 + ], + "angle": 0, + "content": "Let \\( n \\) and \\( d \\) denote the number of tokens and the hidden size in QSM, respectively. Note that \\( n \\) is very small, e.g., \\( n = 4 \\), much lower than the number of tokens in the original model and the value of \\( d \\) corresponds to that of the pre-trained model. Before the self-attention layer, the output of the last block \\( \\hat{H}_{i - 1} \\in \\mathbb{R}^{n \\times d} \\) are added with trainable prompt \\( P_{i} \\) and then fed into bottleneck projection layers. It is formally formulated as follows," + }, + { + "type": "equation", + "bbox": [ + 0.387, + 0.706, + 0.785, + 0.722 + ], + "angle": 0, + "content": "\\[\nZ _ {i} ^ {\\prime} = \\left(\\left(H _ {i} + P _ {i}\\right) \\cdot \\mathbf {W} _ {\\text {d o w n}}\\right) \\cdot \\mathbf {W} _ {\\text {u p}}, \\tag {1}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.734, + 0.787, + 0.767 + ], + "angle": 0, + "content": "where \\(\\mathbf{W}_{down} \\in \\mathbb{R}^{d \\times \\hat{d}}\\) and \\(\\mathbf{W}_{up} \\in \\mathbb{R}^{\\hat{d} \\times d}\\) denote the down-projection and up-projection layer respectively, with \\(\\hat{d} \\ll d\\)." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.768, + 0.787, + 0.813 + ], + "angle": 0, + "content": "In the attention layer, \\( Z_{i}^{\\prime} \\) is projected by three bottleneck modules to produce the query \\( Q_{i}^{qs} \\), key \\( K_{i}^{qs} \\), and value \\( V_{i}^{qs} \\). Then the hidden embedding after attention layers \\( Z_{i}^{\\prime \\prime} \\) can be calculated as follows, controlled by a scale factor \\( s^{\\prime} \\):" + }, + { + "type": "equation", + "bbox": [ + 0.378, + 0.825, + 0.785, + 0.842 + ], + "angle": 0, + "content": "\\[\nZ _ {i} ^ {\\prime \\prime} = s ^ {\\prime} \\cdot \\operatorname {A t t n} \\left(Q _ {i} ^ {q s}, K _ {i} ^ {q s}, V _ {i} ^ {q s}\\right) + Z _ {i} ^ {\\prime}. \\tag {2}\n\\]" + } + ], + [ + { + "type": "page_number", + "bbox": [ + 0.218, + 0.116, + 0.23, + 0.127 + ], + "angle": 0, + "content": "6" + }, + { + "type": "header", + "bbox": [ + 0.272, + 0.115, + 0.49, + 0.129 + ], + "angle": 0, + "content": "Taolin Zhang, Jiawang Bai et al." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.146, + 0.784, + 0.178 + ], + "angle": 0, + "content": "Then we can further obtain the synthesized query \\(\\hat{H}_i\\) after the FFN as the input of KEM, controlled by another scale factor \\(s''\\):" + }, + { + "type": "equation", + "bbox": [ + 0.395, + 0.187, + 0.786, + 0.204 + ], + "angle": 0, + "content": "\\[\n\\hat {H} _ {i} = s ^ {\\prime \\prime} \\cdot \\operatorname {F F N} \\left(\\operatorname {L N} \\left(Z _ {i} ^ {\\prime \\prime}\\right)\\right) + Z _ {i} ^ {\\prime \\prime}, \\tag {3}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.213, + 0.784, + 0.302 + ], + "angle": 0, + "content": "where FFN also consists of bottleneck-structured fully connected layers and \\(\\mathrm{LN}(\\cdot)\\) represents the layer normalization [2]. \\(s^{\\prime}\\) and \\(s''\\) are hyper-parameters to modulate the effects of the attention and FFN on the synthesized query. Since QSM is trained on the downstream dataset from scratch, we believe that its output \\(\\hat{H}_i\\) captures the task-specific knowledge. We use \\(\\hat{H}_i\\) in a query-only manner as illustrated in the next section." + }, + { + "type": "title", + "bbox": [ + 0.216, + 0.324, + 0.515, + 0.339 + ], + "angle": 0, + "content": "3.3 Knowledge Extraction Module" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.348, + 0.784, + 0.422 + ], + "angle": 0, + "content": "With the task-specific query obtained from QSM, we propose Knowledge Extraction Module (KEM) to exploit intermediate features of the original pre-trained model. Before that, we obtain original features \\(\\{X_{1}, X_{2}, \\dots, X_{l}\\}\\) of the input sample from forward-propagation, where \\(l\\) is the number of transformer blocks of the pre-train model. Note that this process is efficient due to no gradient calculation." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.423, + 0.785, + 0.513 + ], + "angle": 0, + "content": "KEM takes the synthesized query \\(\\hat{H}_i\\) and original features \\(\\{X_1, X_2, \\dots, X_l\\}\\) as inputs and outputs useful features for the downstream task. Since \\(\\hat{H}_i\\) has a small number of tokens and original features are only queried by \\(\\hat{H}_i\\), there are no heavy features in the middle of KEM. Besides, we find that it is effective to reuse and freeze the pre-trained parameters for feature interaction in KEM. These designs ensure the parameter- and memory-efficiency of our method." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.513, + 0.784, + 0.558 + ], + "angle": 0, + "content": "Specifically, with the projection weight in the pre-trained ViT blocks, we first obtain the query \\( Q_{i}^{ke} \\), key \\( K_{i}^{ke} \\), and value \\( V_{i}^{ke} \\). Then follow-up attention calculation is performed by:" + }, + { + "type": "equation", + "bbox": [ + 0.389, + 0.568, + 0.785, + 0.585 + ], + "angle": 0, + "content": "\\[\nE _ {i} = \\operatorname {A t t n} \\left(Q _ {i} ^ {k e}, K _ {i} ^ {k e}, V _ {i} ^ {k e}\\right) + \\hat {H} _ {i}. \\tag {4}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.594, + 0.784, + 0.623 + ], + "angle": 0, + "content": "We also reuse the weights of FFN to perform the same space transformation as the pre-trained model:" + }, + { + "type": "equation", + "bbox": [ + 0.403, + 0.635, + 0.785, + 0.651 + ], + "angle": 0, + "content": "\\[\nH _ {i} = \\operatorname {F F N} \\left(\\operatorname {N o r m} \\left(E _ {i}\\right)\\right) + E _ {i}, \\tag {5}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.66, + 0.736, + 0.674 + ], + "angle": 0, + "content": "where \\(\\mathrm{Norm}(\\cdot)\\) denotes the normalization used in the pre-trained model." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.675, + 0.784, + 0.719 + ], + "angle": 0, + "content": "To maximize the potential of KEM, apart from the final output \\( H_{i} \\), we additionally leverage features \\( F_{i}^{att} \\) and \\( F_{i}^{ffn} \\) after the attention module and the FFN for classification. They are obtained as follows:" + }, + { + "type": "equation", + "bbox": [ + 0.392, + 0.729, + 0.785, + 0.769 + ], + "angle": 0, + "content": "\\[\n\\left\\{ \\begin{array}{l} F _ {i} ^ {a t t} = \\operatorname {A t t n} \\left(Q _ {i} ^ {k e}, K _ {i} ^ {k e}, V _ {i} ^ {k e}\\right) \\\\ F _ {i} ^ {f f n} = \\operatorname {F F N} \\left(\\operatorname {N o r m} \\left(E _ {i}\\right)\\right) \\end{array} . \\right. \\tag {6}\n\\]" + }, + { + "type": "title", + "bbox": [ + 0.216, + 0.787, + 0.421, + 0.801 + ], + "angle": 0, + "content": "3.4 Classification Head" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.81, + 0.784, + 0.841 + ], + "angle": 0, + "content": "The classification head makes a prediction of the input sample relying on rich and hierarchical features from KEMs, \\(\\{F_1^{att}, F_1^{ffn}, H_1, \\dots, F_l^{att}, F_l^{ffn}, H_l\\}\\), where \\(l\\)" + } + ], + [ + { + "type": "header", + "bbox": [ + 0.245, + 0.114, + 0.733, + 0.129 + ], + "angle": 0, + "content": "Parameter-Efficient and Memory-Efficient Tuning for Vision Transformer" + }, + { + "type": "page_number", + "bbox": [ + 0.775, + 0.116, + 0.785, + 0.127 + ], + "angle": 0, + "content": "7" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.147, + 0.784, + 0.192 + ], + "angle": 0, + "content": "is the number of blocks of the pre-trained ViT. Instead of directly applying an MLP classifier on top of these features, we propose several techniques for better performance. Figure 2 shows the overall architecture of the classification head." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.193, + 0.532, + 0.328 + ], + "angle": 0, + "content": "Feature projection. Since these features are generated across different layers from KEMs, they may be misaligned in the representation space. Therefore, before aggregating all features, we align all features with projection layers. We use bottleneck-structured fully connected layers and share the projection layers across all features to improve the parameter efficiency." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.33, + 0.534, + 0.585 + ], + "angle": 0, + "content": "Conditional feature aggregation. These features contain a large amount of information but also some redundancy, perhaps hurting the performance of downstream tasks. Therefore, it is crucial to perform feature selection, aiming to extract the truly beneficial information. We propose a strategy of conditional feature aggregation, which generates weights conditioned on \\( H_{l} \\) and performs a weighted-sum over all features. Our strategy enables SynQT to adaptively rely on different features to classify different samples. Despite only a few parameters, the conditional feature aggregation results in better performance compared to a simple average or weighted sum with learnable but fixed weights, as shown in our experiments." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.587, + 0.788, + 0.693 + ], + "angle": 0, + "content": "DropFeat in training. Due to rich features fed into the classification head, overfitting likely happens and degrades the performance. To solve this issue, inspired by the idea of dropout [23, 42, 47], we propose DropFeat, a simple regularization technique to prevent overfitting. DropFeat randomly removes some input features for the classification head during training, while all features are leveraged during inference. DropFeat ensures that the model does not overly rely on a specific feature, which significantly improves the model's performance." + }, + { + "type": "image", + "bbox": [ + 0.557, + 0.224, + 0.787, + 0.429 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.542, + 0.44, + 0.788, + 0.567 + ], + "angle": 0, + "content": "Fig.2: Classification head in SynQT, corresponding to the head in Figure 1. We utilize a shared projection to align the hierarchical features and apply feature aggregation conditioned on the output of the last block \\( H_{l} \\). We also adopt DropFeat to prevent overfitting, which randomly drops some features during training." + }, + { + "type": "title", + "bbox": [ + 0.215, + 0.718, + 0.367, + 0.736 + ], + "angle": 0, + "content": "4 Experiment" + }, + { + "type": "title", + "bbox": [ + 0.215, + 0.752, + 0.312, + 0.768 + ], + "angle": 0, + "content": "4.1 Setup" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.78, + 0.788, + 0.842 + ], + "angle": 0, + "content": "Datasets. We conduct experiments on the VTAB-1K benchmark and the fine-grained visual classification (FGVC) datasets. VTAB-1K consists of 19 different classification tasks categorized into three groups: Natural, Specialized, and Structured. Each task of VTAB-1K contains 1000 training samples. We evaluate the" + } + ], + [ + { + "type": "page_number", + "bbox": [ + 0.218, + 0.116, + 0.23, + 0.127 + ], + "angle": 0, + "content": "8" + }, + { + "type": "header", + "bbox": [ + 0.272, + 0.115, + 0.49, + 0.13 + ], + "angle": 0, + "content": "Taolin Zhang, Jiawang Bai et al." + }, + { + "type": "table_caption", + "bbox": [ + 0.214, + 0.145, + 0.788, + 0.201 + ], + "angle": 0, + "content": "Table 1: Full results on VTAB-1K benchmark with pre-trained ViT-B/16 on ImageNet21K. “# Memory” specifies memory cost during training with 64 as batch size on CIFAR100. We report the Top-1 accuracy and the results are averaged over group-wise mean values." + }, + { + "type": "table", + "bbox": [ + 0.219, + 0.213, + 0.785, + 0.383 + ], + "angle": 0, + "content": "
# Memory (GB)NaturalSpecializedStructured
CIFAR100Caltech101DTDFlower102PetsSVHNSun397CamelyonEuroSATRess45RetinopathyClever-CountClever-DistDMLabKITTL-DistdSpr-LocdSpr-OrisNORB-AzimsNORB-EleAverage
Traditional Fine-Tuning
Linear Probing2.550.685.661.479.586.540.838.079.791.571.765.541.434.434.155.418.126.416.524.852.7
Full Fine-Tuning11.144.384.554.184.774.787.226.985.395.076.070.471.560.546.972.974.538.728.523.863.2
PETL Methods
BitFit [54]7.872.887.059.297.585.359.951.478.791.672.969.861.555.632.455.966.640.015.725.165.2
VPT-Shallow [17]8.177.786.962.697.587.374.551.278.292.075.672.950.558.640.567.168.736.120.234.167.8
VPT-Deep [17]8.178.890.865.898.088.378.149.681.896.183.468.468.560.046.572.873.647.932.937.872.0
Memory-Efficient PETL Methods
VQT [45]3.266.389.967.897.984.779.945.579.095.280.974.746.761.645.163.662.932.130.028.868.3
SynQT (Ours)3.470.989.768.898.589.677.850.682.396.783.575.271.862.748.575.474.149.031.736.172.9
" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.411, + 0.785, + 0.47 + ], + "angle": 0, + "content": "few-shot performance of SynQT on 5 FGVC datasets: FGVC-Aircraft [36], OxfordPets [38], Food-101 [5], Stanford Cars [22], and Oxford-Flowers102 [37]. We follow the split in [59] and report the results in 1, 2, 4, 8, and 16-shot settings. The accuracy is averaged over three trials with different seeds." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.472, + 0.806, + 0.608 + ], + "angle": 0, + "content": "Implementation details. We utilize a pre-trained ViT-B/16 [10] on the ImageNet-21K dataset as the foundational model. We also conduct experiments on the ImageNet-1K supervised pre-trained backbone, as well as on the self-supervised checkpoints (DINO [7] and MAE [14]). For a lightweight design of QSM, we use an extremely small number of tokens, with the number of tokens \\( n \\) from \\(\\{1, 4, 16\\}\\). This is significantly fewer than the number of tokens in VPT, which can reach up to 200. Additionally, the hyper-parameters \\( s' \\) and \\( s'' \\) are chosen from \\(\\{0.1, 1\\}\\). The hidden size of the bottleneck projection is set to 48, except attention projection of the query, key, and value in QSM, which is set to 8." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.609, + 0.788, + 0.745 + ], + "angle": 0, + "content": "Due to the limited number of validation samples in VTAB-1K, we implement 5-fold cross-validation on the training sets to determine the optimal hyperparameters. For few-shot learning, we use the default splits of FGVC datasets to search for the best hyperparameters. The Top-1 accuracy is reported on the test sets in all our experiments. Following common practice [18], we adopt AdamW [33] as the optimizer and a cosine learning rate scheduler with a constant batch size of 64. For data augmentation, we adjust the image size to \\(224 \\times 224\\) directly for VTAB-1K, and for few-shot learning, we resize the image to \\(256 \\times 256\\) before applying a center crop." + }, + { + "type": "title", + "bbox": [ + 0.216, + 0.768, + 0.373, + 0.782 + ], + "angle": 0, + "content": "4.2 Main Results" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.795, + 0.787, + 0.843 + ], + "angle": 0, + "content": "Besides the traditional fine-tuning methods (linear probing and full fine-tuning), we compare our method with PETL methods including BitFit [54], VPT-Shallow [17], and VPT-Deep [17], as well as a memory-efficient one, VQT [45], on VTAB-" + } + ], + [ + { + "type": "header", + "bbox": [ + 0.246, + 0.115, + 0.733, + 0.129 + ], + "angle": 0, + "content": "Parameter-Efficient and Memory-Efficient Tuning for Vision Transformer" + }, + { + "type": "page_number", + "bbox": [ + 0.775, + 0.116, + 0.785, + 0.127 + ], + "angle": 0, + "content": "9" + }, + { + "type": "table_caption", + "bbox": [ + 0.216, + 0.145, + 0.785, + 0.174 + ], + "angle": 0, + "content": "Table 2: Classification results on MAE, DINO, and Supervised ImageNet-1K pre-trained backbones. We report the Top-1 accuracy and the averaged results." + }, + { + "type": "table", + "bbox": [ + 0.264, + 0.185, + 0.744, + 0.427 + ], + "angle": 0, + "content": "
MethodsNaturalSpecializedStructuredAverage
MAE backbone
BitFit [54]63.580.657.267.1
VPT-Shallow [17]38.372.025.045.1
VPT-Deep [17]50.876.437.354.8
VQT [45]56.678.643.459.5
SynQT (Ours)66.082.658.268.9
DINO backbone
BitFit [54]76.883.849.269.9
VPT-Shallow [17]69.882.638.563.6
VPT-Deep [17]76.783.847.369.3
VQT [45]71.083.637.263.9
SynQT (Ours)75.184.856.972.3
Supervised ImageNet-1K backbone
BitFit [54]74.683.449.469.1
VPT-Shallow [17]63.780.340.261.4
VPT-Deep [17]74.982.953.970.6
VQT [45]72.784.549.368.8
SynQT (Ours)73.284.756.271.4
" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.451, + 0.788, + 0.572 + ], + "angle": 0, + "content": "1K. As can be seen from Table 1, the most striking observation emerging from the comparison is that SynQT outperforms other methods while saving up to \\(50\\%\\) of GPU memory. Given the same batch size, previous PETL methods rely on the full backward propagation of the original model and consume \\(7 \\sim 8\\) GB of memory, whereas SynQT disentangles transfer learning and reduces the memory cost to \\(3.4\\) GB. Notably, our method performs better than VPT-Deep on 14 of 19 tasks. We notice that although the memory cost of SynQT is slightly higher, it achieves a significant performance improvement of \\(4.6\\%\\) compared to VQT." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.572, + 0.788, + 0.632 + ], + "angle": 0, + "content": "In order to further validate the robustness of our method, we compare SynQT with baseline models across various checkpoints which are derived from different learning paradigms, including supervised learning (pretrained over ImageNet-21K [9]), and self-supervised learning (DINO" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.632, + 0.541, + 0.769 + ], + "angle": 0, + "content": "[7] and MAE [14]). The results in Table 2 indicate that SynQT shows strong compatibility and outperforms PETL methods by a substantial margin over different backbones. For instance, compared to the strongest baselines, our method achieves improvements of \\(1.8\\%\\), \\(3.3\\%\\), and \\(0.8\\%\\) on average with MAE, DINO, and supervised ImageNet-1K backbones, respectively." + }, + { + "type": "image", + "bbox": [ + 0.553, + 0.65, + 0.784, + 0.786 + ], + "angle": 0, + "content": null + }, + { + "type": "title", + "bbox": [ + 0.216, + 0.788, + 0.512, + 0.804 + ], + "angle": 0, + "content": "4.3 Memory-Efficiency Evaluation" + }, + { + "type": "text", + "bbox": [ + 0.216, + 0.809, + 0.538, + 0.84 + ], + "angle": 0, + "content": "Previous PETL methods [8, 15, 16, 18] focus on reducing the number of parameters when" + }, + { + "type": "image_caption", + "bbox": [ + 0.549, + 0.799, + 0.788, + 0.856 + ], + "angle": 0, + "content": "Fig.3: Performance comparison under memory constraints. SynQT outperforms other methods with different memory costs." + } + ], + [ + { + "type": "page_number", + "bbox": [ + 0.218, + 0.116, + 0.236, + 0.127 + ], + "angle": 0, + "content": "10" + }, + { + "type": "header", + "bbox": [ + 0.272, + 0.115, + 0.49, + 0.129 + ], + "angle": 0, + "content": "Taolin Zhang, Jiawang Bai et al." + }, + { + "type": "table_caption", + "bbox": [ + 0.216, + 0.136, + 0.787, + 0.193 + ], + "angle": 0, + "content": "Table 3: Full results on VTAB-1K benchmark under memory constraints. The memory is constrained within 5GB and “# Memory” specifies the memory cost during training with 64 as batch size on CIFAR100. We report the Top-1 accuracy and the results are averaged over group-wise mean values." + }, + { + "type": "table", + "bbox": [ + 0.219, + 0.204, + 0.785, + 0.319 + ], + "angle": 0, + "content": "
# Memory (GB)NaturalSpecializedStructured
CIFAR100Caltech101DTDFlower102PetsSVHNSun397CanelyonEuroSATResics45RetinopathyClevr-CountClevr-DistDMLabKITTI-DistdSpr-LocdSpr-OrisNORB-AzimsNORB-EteAverage
FacT-TT [18]4.340.484.259.896.886.653.445.879.492.273.374.459.250.242.568.848.225.919.828.763.2
Adapter [15]4.764.986.062.897.386.646.844.479.893.976.774.663.340.238.968.533.132.319.830.564.0
LoRA [16]4.862.582.160.497.586.658.645.380.695.378.872.864.845.141.368.549.628.124.630.565.5
AdaptFormer [8]4.665.686.162.096.986.235.545.682.696.277.674.466.247.741.372.355.629.422.633.965.8
SynQT(Ours)3.470.989.768.898.589.677.850.682.396.783.575.271.862.748.575.474.149.031.736.172.9
" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.348, + 0.784, + 0.438 + ], + "angle": 0, + "content": "fine-tuning the model. While decreasing the cost of model storage, these methods are not widely applicable in practical training because they still require a significant amount of GPU memory. Therefore, we propose SynQT to overcome this problem. In this section, we highlight our improvements in terms of memory costs by comparing our method with some SOTA PETL methods under a given GPU memory constraint." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.439, + 0.785, + 0.529 + ], + "angle": 0, + "content": "For baseline methods, we discard some specific trainable structures to meet the memory constraint. Specifically, we remove these additional structures near the input layer, which means that the training only requires a partial backpropagation with a lower memory usage. We progressively remove the structure nearest to the input layer until it meets the given memory constraint. Besides, for a fair comparison, we reduce the memory usage of SynQT in the same manner." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.53, + 0.787, + 0.636 + ], + "angle": 0, + "content": "We compare SynQT with SOTA baselines including FacT-TT [18], Adapter [15], LoRA [16] and AdaptFormer [8] under such a constraint and report the results in Table 3. While previous methods can achieve promising performance with few trainable parameters, they unfortunately suffer from significant performance drops when such memory constraints are imposed. In contrast, SynQT excels in this scenario, achieving state-of-the-art performance and surpassing AdaptFormer by \\(7.1\\%\\) and VQT by \\(4.6\\%\\)." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.637, + 0.788, + 0.758 + ], + "angle": 0, + "content": "To have a better understanding of how memory constraints affect model performance, we estimate the results of the methods mentioned above across varying memory costs. As depicted in Figure 3, SynQT continues to outperform the baselines across different memory costs. Notably, when the constraint is very strict, i.e., 3 GB, most of the existing PETL methods perform poorly, with a gap of almost \\(7\\%\\) compared to SynQT. Moreover, it is noteworthy that relaxing the memory constraints improves the performance of all the methods, especially for a significant improvement observed for our SynQT." + }, + { + "type": "title", + "bbox": [ + 0.216, + 0.782, + 0.504, + 0.798 + ], + "angle": 0, + "content": "4.4 Few-Shot Learning on FGVC" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.81, + 0.787, + 0.843 + ], + "angle": 0, + "content": "To further evaluate SynQT's ability to transfer knowledge from the pre-trained model using few samples, we compare SynQT with BitFit [54], VPT-Shallow [17]," + } + ], + [ + { + "type": "header", + "bbox": [ + 0.246, + 0.115, + 0.733, + 0.129 + ], + "angle": 0, + "content": "Parameter-Efficient and Memory-Efficient Tuning for Vision Transformer" + }, + { + "type": "page_number", + "bbox": [ + 0.768, + 0.116, + 0.784, + 0.127 + ], + "angle": 0, + "content": "11" + }, + { + "type": "image", + "bbox": [ + 0.223, + 0.145, + 0.782, + 0.273 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.215, + 0.286, + 0.785, + 0.315 + ], + "angle": 0, + "content": "Fig. 4: Performance comparison in few-shot learning. Overall, SynQT outperforms other baselines in the extreme low-data regime." + }, + { + "type": "table_caption", + "bbox": [ + 0.215, + 0.328, + 0.784, + 0.357 + ], + "angle": 0, + "content": "Table 4: Applying SynQT to Swin-B and ConvNeXt-B. We report the Top-1 accuracy and the averaged results." + }, + { + "type": "table", + "bbox": [ + 0.221, + 0.37, + 0.784, + 0.435 + ], + "angle": 0, + "content": "
MethodSwin-BConvNeXt-B
NaturalSpecializedStructuredAverageNaturalSpecializedStructuredAverage
Linear73.580.833.562.674.581.534.863.6
VPT-Deep76.884.553.471.678.583.044.668.7
SynQT77.586.057.873.878.182.148.269.5
" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.462, + 0.785, + 0.492 + ], + "angle": 0, + "content": "VPT-Deep [17], and VQT [45] in few-shot setting on 5 FGVC datasets. The results are reported in Figure 4." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.493, + 0.787, + 0.645 + ], + "angle": 0, + "content": "Overall, SynQT significantly outperforms all other baselines in 4, 6, 8, and 16-shot scenarios, and its performance is comparable in the 1-shot scenario. The gaps between VQT and SynQT indicate that the lightweight, task-specific QSM enables SynQT to capture beneficial information from samples in the extremely low-data regime. Simply incorporating a trainable prompt as a query, as in VQT, is not sufficient to handle such challenging few-shot tasks. Interestingly, VPT-Shallow outperforms VPT-Deep in the low-data regime, due to the fact that the larger number of prompts in VPT-Deep are not well-trained on the limited data. This suggests that merely increasing the number of prompt numbers could potentially lead to a performance drop." + }, + { + "type": "title", + "bbox": [ + 0.215, + 0.665, + 0.434, + 0.681 + ], + "angle": 0, + "content": "4.5 Versatility of SynQT" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.689, + 0.785, + 0.719 + ], + "angle": 0, + "content": "To demonstrate the versatility of SynQT, we apply SynQT to different model architectures and combine SynQT with existing PETL methods." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.72, + 0.787, + 0.84 + ], + "angle": 0, + "content": "Applying SynQT to different model architectures. Here, we apply SynQT to an advanced Transformer-like architecture (Swin [31]) and a convolutional neural network (ConvNeXt [32]). Following [17], we introduce prompts for each window in Swin and incorporate learnable prompt pixels for the feature map in ConvNeXt. Specifically, due to the varying hidden sizes and number of blocks in different layers, we extract features from the third layer of Swin and the last layer of ConvNext to ensure sufficient extraction of beneficial information. We find that both of them significantly benefit from SynQT, as shown in Table 4." + } + ], + [ + { + "type": "page_number", + "bbox": [ + 0.218, + 0.116, + 0.236, + 0.127 + ], + "angle": 0, + "content": "12" + }, + { + "type": "header", + "bbox": [ + 0.272, + 0.115, + 0.49, + 0.129 + ], + "angle": 0, + "content": "Taolin Zhang, Jiawang Bai et al." + }, + { + "type": "table_caption", + "bbox": [ + 0.217, + 0.158, + 0.514, + 0.185 + ], + "angle": 0, + "content": "Table 5: Combining different PETL methods. AF is short for AdaptFormer." + }, + { + "type": "table", + "bbox": [ + 0.22, + 0.187, + 0.51, + 0.282 + ], + "angle": 0, + "content": "
MethodsNaturalSpecializedStructuredAverage
AF80.182.350.370.9
AF+VQT79.684.353.072.3
AF+SynQT82.685.457.975.3
VPT79.184.654.472.7
VPT+VQT78.983.754.672.4
VPT+SynQT81.184.957.774.1
" + }, + { + "type": "table_caption", + "bbox": [ + 0.516, + 0.175, + 0.785, + 0.203 + ], + "angle": 0, + "content": "Table 6: Necessity of disentangle-ment (KEM and QSM)." + }, + { + "type": "table", + "bbox": [ + 0.518, + 0.204, + 0.784, + 0.265 + ], + "angle": 0, + "content": "
MethodNatural SpecializedStructuredAverage
Linear63.277.131.457.2
+KEM71.880.132.461.4
+QSM77.883.655.072.1
+Head (SynQT)78.088.456.272.9
" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.311, + 0.784, + 0.34 + ], + "angle": 0, + "content": "On average, SynQT outperforms VPT-Deep by \\(2.2\\%\\) and \\(1.2\\%\\) for Swin and ConvNeXt, respectively." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.342, + 0.785, + 0.448 + ], + "angle": 0, + "content": "Combining SynQT with existing PETL methods. We further jointly utilize SynQT and existing PETL methods to adapt pre-trained models, as shown in Table 5. The results demonstrate that both VPT and AdaptFormer benefit from SynQT on VTAB-1K, with improvements of \\(4.4\\%\\) and \\(1.4\\%\\) on average, respectively. We can see that for AdaptFormer and VPT, VQT may result in a performance drop in some cases (e.g., the Natural group), whereas equipping SynQT leads to significant performance improvements across all groups." + }, + { + "type": "title", + "bbox": [ + 0.216, + 0.469, + 0.389, + 0.485 + ], + "angle": 0, + "content": "4.6 Ablation Study" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.493, + 0.784, + 0.522 + ], + "angle": 0, + "content": "We ablate our SynQT with different settings in Tables 6 and 7 for a better understanding of SynQT." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.524, + 0.785, + 0.629 + ], + "angle": 0, + "content": "Disentanglement. To demonstrate the necessity of disentanglement, we perform an ablation in Table 6. We set the baseline as linear probing since KEM exploits intermediate features. To gain deeper insights into the roles of QSM and Head, we employ random prompts to query in KEM as the model without QSM, and we utilize a simple average for aggregation as the model without Head. The results demonstrate that our idea of disentangling (QSM and KEM) is the key to achieving superior performance." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.63, + 0.785, + 0.734 + ], + "angle": 0, + "content": "Hidden size in QSM. We vary the hidden size of the bottleneck module in QSM and report the results in Table 7a. With a larger hidden size, the performance improves due to the increasing capacity. However, when the hidden size exceeds 48, the benefit of introducing new parameters diminishes and overfitting may happen in datasets with few training samples. The results indicate that redundant parameters might not enhance the model's performance, and a hidden size of 48 is a reasonable choice." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.735, + 0.785, + 0.809 + ], + "angle": 0, + "content": "Input of QSM. Table 7c studies the influence of different inputs of QSM. We can see that both the trainable prompts and the output from the last block are indispensable for QSM to generate high-quality task-specific query. The trainable prompts improve the capacity of QSM, and the results from the previous block ensure that knowledge is retained during the forward process." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.81, + 0.785, + 0.842 + ], + "angle": 0, + "content": "Hierarchical features. In SynQT, we extract three different types of features from KSMs for classification: final output of KSM \\( H_{i} \\), attention features \\( F_{i}^{attn} \\)" + } + ], + [ + { + "type": "header", + "bbox": [ + 0.245, + 0.115, + 0.733, + 0.129 + ], + "angle": 0, + "content": "Parameter-Efficient and Memory-Efficient Tuning for Vision Transformer" + }, + { + "type": "page_number", + "bbox": [ + 0.768, + 0.116, + 0.786, + 0.127 + ], + "angle": 0, + "content": "13" + }, + { + "type": "table_caption", + "bbox": [ + 0.216, + 0.145, + 0.785, + 0.175 + ], + "angle": 0, + "content": "Table 7: SynQT ablation experiments with ViT-B/16 with ImageNet-21K backbone. Default settings are marked in gray." + }, + { + "type": "table", + "bbox": [ + 0.24, + 0.189, + 0.495, + 0.269 + ], + "angle": 0, + "content": "
DimParamNat.Spe.Str.Avg.
162.14 M77.783.855.872.4
322.43 M77.783.955.372.3
482.73 M78.084.456.272.9
643.02 M77.684.456.172.7
963.61 M77.784.555.972.7
" + }, + { + "type": "table_footnote", + "bbox": [ + 0.228, + 0.271, + 0.497, + 0.294 + ], + "angle": 0, + "content": "(a) Hidden size in QSM. 48 as the width is a trade-off between parameters and performance." + }, + { + "type": "table", + "bbox": [ + 0.518, + 0.192, + 0.775, + 0.259 + ], + "angle": 0, + "content": "
SynQTNat.Spe.Str.Avg.
w/o Hi78.384.053.872.0
w/o Fiattn78.383.555.472.4
w/o Fiuffn77.884.755.472.6
Full78.084.456.272.9
" + }, + { + "type": "table_footnote", + "bbox": [ + 0.507, + 0.271, + 0.776, + 0.303 + ], + "angle": 0, + "content": "(b) Feature selection. Hierarchical features contains rich information that benefit to downstream tasks." + }, + { + "type": "table", + "bbox": [ + 0.246, + 0.32, + 0.501, + 0.37 + ], + "angle": 0, + "content": "
SynQTNat.Spe.Str.Avg.
w/o Prompt77.282.951.970.7
w/o Last Output77.584.452.371.4
Full78.084.456.272.9
" + }, + { + "type": "table_footnote", + "bbox": [ + 0.235, + 0.373, + 0.503, + 0.406 + ], + "angle": 0, + "content": "(c) Input of QSM. Both the trainable prompt and the output from last block are important for generating task-specific query." + }, + { + "type": "table", + "bbox": [ + 0.526, + 0.318, + 0.78, + 0.37 + ], + "angle": 0, + "content": "
ProjectionNat.Spe.Str.Avg.
w/o Proj.78.483.855.072.4
Shared78.084.456.272.9
Independent77.984.455.572.6
" + }, + { + "type": "table_footnote", + "bbox": [ + 0.514, + 0.373, + 0.783, + 0.405 + ], + "angle": 0, + "content": "(d) Projection in the classification head. Shared projection works the best for downstream tasks." + }, + { + "type": "table", + "bbox": [ + 0.24, + 0.419, + 0.495, + 0.471 + ], + "angle": 0, + "content": "
SynQTNat.Spe.Str.Avg.
Simply Avg.77.984.655.672.7
Fixed Weights77.984.755.572.7
Conditional78.084.456.272.9
" + }, + { + "type": "table_footnote", + "bbox": [ + 0.228, + 0.474, + 0.496, + 0.506 + ], + "angle": 0, + "content": "(e) Feature aggregation. Conditional feature aggregation is more effective than mean and learnable but fixed weight." + }, + { + "type": "table", + "bbox": [ + 0.518, + 0.42, + 0.774, + 0.46 + ], + "angle": 0, + "content": "
SynQTNat.Spe.Str.Avg.
w/o DropFeat77.784.355.272.4
w/ DropFeat78.084.456.272.9
" + }, + { + "type": "table_footnote", + "bbox": [ + 0.507, + 0.473, + 0.776, + 0.505 + ], + "angle": 0, + "content": "(f) DropFeat in the classification head. The DropFeat module in the conditional head makes model more robust." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.549, + 0.785, + 0.611 + ], + "angle": 0, + "content": "and FFN features \\( F_{i}^{ffn} \\), where \\( i \\in \\{1,2,\\dots,l\\} \\) and \\( l \\) is the number of layers. All these hierarchical features contain rich information and contribute to the performance boost of SynQT, as demonstrated in Table 7b. In particular, \\( H_{i} \\) plays a more important role than other types of features." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.612, + 0.785, + 0.718 + ], + "angle": 0, + "content": "Projection in the classification head. We vary the projection architecture in the classification head and investigate its performance in Table 7d. A projection of the input is necessary for the classification head, as hierarchical features across various layers may be misaligned in the representation space. However, to our surprise, a shared bottleneck structure is enough to handle these hierarchical features. The reason may be that independent projection for distinct features leads to overfitting, subsequently causing a slight drop in performance." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.719, + 0.785, + 0.809 + ], + "angle": 0, + "content": "Conditional feature aggregation. In the classification head, we aggregate features using the weights conditioned on \\( H_{l} \\). We further conduct experiments over variants of SynQT with different aggregation methods including a simple average and weighted-sum with learnable but fixed weights. The results in Table 7e show that conditional aggregation improves the performance of SynQT especially for the Structured group." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.81, + 0.785, + 0.84 + ], + "angle": 0, + "content": "DropFeat in the classification head. We study the effect of DropFeat, which is proposed to prevent overfitting by randomly dropping features in the classi-" + } + ], + [ + { + "type": "page_number", + "bbox": [ + 0.218, + 0.116, + 0.236, + 0.127 + ], + "angle": 0, + "content": "14" + }, + { + "type": "header", + "bbox": [ + 0.272, + 0.115, + 0.49, + 0.129 + ], + "angle": 0, + "content": "Taolin Zhang, Jiawang Bai et al." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.147, + 0.784, + 0.177 + ], + "angle": 0, + "content": "fication head. As can be observed from Table 7f, SynQT performs worse without DropFeat, indicating that DropFeat enhances the model's generalization" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.178, + 0.488, + 0.297 + ], + "angle": 0, + "content": "in handling rich hierarchical feature input. DropFeat ensures that the model does not overly rely on a specific feature during training and encourages the model to focus on a general way to select beneficial information from these features for downstream tasks." + }, + { + "type": "title", + "bbox": [ + 0.216, + 0.317, + 0.427, + 0.331 + ], + "angle": 0, + "content": "4.7 t-SNE Visualization" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.339, + 0.487, + 0.444 + ], + "angle": 0, + "content": "We present t-SNE visualization [35] of the CLS Token in pre-trained ViT, VQT, and our SynQT on EuroSAT, SVHN, and Clevr-Dist in Figure 5. The visualization shows that our SynQT achieves better feature clustering results compared to the" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.444, + 0.755, + 0.46 + ], + "angle": 0, + "content": "original CLS Token and VQT. We provide more visualization in Appendix." + }, + { + "type": "image", + "bbox": [ + 0.51, + 0.206, + 0.782, + 0.378 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.498, + 0.39, + 0.788, + 0.446 + ], + "angle": 0, + "content": "Fig. 5: t-SNE visualization on SVHN, EuroSAT, and Clevr-Dist, one from each category of VTAB-1K. The features of our SynQT are more separable." + }, + { + "type": "title", + "bbox": [ + 0.216, + 0.48, + 0.654, + 0.496 + ], + "angle": 0, + "content": "5 Conclusion, Limitation, and Future Work" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.508, + 0.788, + 0.659 + ], + "angle": 0, + "content": "In this work, we present an insightful analysis of the high memory cost of existing PETL methods from a unique perspective, i.e., they entangle task-specific learning and pre-trained knowledge utilization together. Inspired by the idea of disentanglement, we propose a new paradigm, namely SynQT, consisting of two key structures: Query Synthesis Module (QSM) and Knowledge Extraction Module (KEM). QSM synthesizes task-specific query and then KEM extracts pre-trained knowledge in a query-only manner. Finally, a classification head is proposed to handle rich hierarchical features from KEMs and make predictions. Extensive experiments on VTAB-1K and FGVC demonstrate the effectiveness of SynQT under memory constraints and in an extremely low-data regime." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.66, + 0.788, + 0.765 + ], + "angle": 0, + "content": "Despite the promising performance of our SynQT, it also has some limitations. One limitation is that SynQT requires slightly more trainable parameters than previous PETL methods due to a few projection layers and FFN operations, which is analyzed in Appendix. However, we would like to emphasize that as shown in their original papers, for VPT-Deep [17] and VQT [45], straightforwardly increasing the number of tokens to use more trainable parameters can not consistently result in better performance." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.766, + 0.788, + 0.84 + ], + "angle": 0, + "content": "We believe that our work will inspire the community to explore memory-efficient PETL along with our disentangled perspective. Besides, to overcome the aforementioned limitation, one future direction is to develop PETL methods that achieve high accuracy and meanwhile have extreme memory efficiency and parameter efficiency." + } + ], + [ + { + "type": "header", + "bbox": [ + 0.245, + 0.115, + 0.732, + 0.129 + ], + "angle": 0, + "content": "Parameter-Efficient and Memory-Efficient Tuning for Vision Transformer" + }, + { + "type": "page_number", + "bbox": [ + 0.768, + 0.117, + 0.785, + 0.127 + ], + "angle": 0, + "content": "15" + }, + { + "type": "title", + "bbox": [ + 0.218, + 0.146, + 0.393, + 0.163 + ], + "angle": 0, + "content": "Acknowledgments" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.176, + 0.788, + 0.252 + ], + "angle": 0, + "content": "This work is supported in part by the National Natural Science Foundation of China under Grant 62171248, Shenzhen Science and Technology Program (JCYJ20220818101012025), the PCNL KEY project (PCL2023AS6-1), and the National Research Foundation, Singapore, under its AI Singapore Programme (AISG Award No: AISG2-RP-2021-023)." + }, + { + "type": "title", + "bbox": [ + 0.218, + 0.274, + 0.323, + 0.289 + ], + "angle": 0, + "content": "References" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.303, + 0.785, + 0.331 + ], + "angle": 0, + "content": "1. Argyriou, A., Evgeniou, T., Pontil, M.: Multi-task feature learning. Advances in neural information processing systems 19 (2006)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.333, + 0.785, + 0.358 + ], + "angle": 0, + "content": "2. Ba, J.L., Kiros, J.R., Hinton, G.E.: Layer normalization. arXiv preprint arXiv:1607.06450 (2016)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.36, + 0.785, + 0.4 + ], + "angle": 0, + "content": "3. Bai, J., Gao, K., Min, S., Xia, S.T., Li, Z., Liu, W.: Badclip: Trigger-aware prompt learning for backdoor attacks on clip. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 24239-24250 (2024)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.402, + 0.785, + 0.441 + ], + "angle": 0, + "content": "4. Bai, J., Yuan, L., Xia, S.T., Yan, S., Li, Z., Liu, W.: Improving vision transformers by revisiting high-frequency components. In: European Conference on Computer Vision. pp. 1-18. Springer (2022)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.443, + 0.785, + 0.496 + ], + "angle": 0, + "content": "5. Bossard, L., Guillaumeun, M., Van Gool, L.: Food-101-mining discriminative components with random forests. In: Computer Vision-ECCV 2014: 13th European Conference, Zurich, Switzerland, September 6-12, 2014, Proceedings, Part VI 13. pp. 446-461. Springer (2014)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.498, + 0.785, + 0.537 + ], + "angle": 0, + "content": "6. Brown, T., Mann, B., Ryder, N., Subbiah, M., Kaplan, J.D., Dhariwal, P., Neelakantan, A., Shyam, P., Sastry, G., Askell, A., et al.: Language models are few-shot learners. NeurIPS (2020)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.539, + 0.785, + 0.579 + ], + "angle": 0, + "content": "7. Caron, M., Touvron, H., Misra, I., Jégou, H., Mairal, J., Bojanowski, P., Joulin, A.: Emerging properties in self-supervised vision transformers. In: Proceedings of the IEEE/CVF international conference on computer vision. pp. 9650-9660 (2021)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.581, + 0.785, + 0.62 + ], + "angle": 0, + "content": "8. Chen, S., Ge, C., Tong, Z., Wang, J., Song, Y., Wang, J., Luo, P.: Adaptformer: Adapting vision transformers for scalable visual recognition. Advances in Neural Information Processing Systems 35, 16664-16678 (2022)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.621, + 0.785, + 0.661 + ], + "angle": 0, + "content": "9. Deng, J., Dong, W., Socher, R., Li, L.J., Li, K., Fei-Fei, L.: Imagenet: A large-scale hierarchical image database. In: 2009 IEEE conference on computer vision and pattern recognition. pp. 248-255. IEEE (2009)" + }, + { + "type": "ref_text", + "bbox": [ + 0.22, + 0.663, + 0.785, + 0.716 + ], + "angle": 0, + "content": "10. Dosovitskiy, A., Beyer, L., Kolesnikov, A., Weissenborn, D., Zhai, X., Unterthiner, T., Dehghani, M., Minderer, M., Heigold, G., Gelly, S., et al.: An image is worth 16x16 words: Transformers for image recognition at scale. arXiv preprint arXiv:2010.11929 (2020)" + }, + { + "type": "ref_text", + "bbox": [ + 0.22, + 0.718, + 0.785, + 0.757 + ], + "angle": 0, + "content": "11. Evci, U., Dumoulin, V., Larochelle, H., Mozer, M.C.: Head2toe: Utilizing intermediate representations for better transfer learning. In: International Conference on Machine Learning. pp. 6009-6033. PMLR (2022)" + }, + { + "type": "ref_text", + "bbox": [ + 0.22, + 0.759, + 0.785, + 0.799 + ], + "angle": 0, + "content": "12. Guo, H., Dai, T., Bai, Y., Chen, B., Xia, S.T., Zhu, Z.: Adaptir: Parameter efficient multi-task adaptation for pre-trained image restoration models. arXiv preprint arXiv:2312.08881 (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.22, + 0.8, + 0.785, + 0.84 + ], + "angle": 0, + "content": "13. Guo, H., Li, J., Dai, T., Ouyang, Z., Ren, X., Xia, S.T.: Mambair: A simple baseline for image restoration with state-space model. arXiv preprint arXiv:2402.15648 (2024)" + }, + { + "type": "list", + "bbox": [ + 0.22, + 0.303, + 0.785, + 0.84 + ], + "angle": 0, + "content": null + } + ], + [ + { + "type": "page_number", + "bbox": [ + 0.218, + 0.116, + 0.236, + 0.127 + ], + "angle": 0, + "content": "16" + }, + { + "type": "header", + "bbox": [ + 0.272, + 0.115, + 0.49, + 0.129 + ], + "angle": 0, + "content": "Taolin Zhang, Jiawang Bai et al." + }, + { + "type": "ref_text", + "bbox": [ + 0.218, + 0.147, + 0.787, + 0.189 + ], + "angle": 0, + "content": "14. He, K., Chen, X., Xie, S., Li, Y., Dollár, P., Girshick, R.: Masked autoencoders are scalable vision learners. In: Proceedings of the IEEE/CVF conference on computer vision and pattern recognition. pp. 16000-16009 (2022)" + }, + { + "type": "ref_text", + "bbox": [ + 0.218, + 0.19, + 0.787, + 0.232 + ], + "angle": 0, + "content": "15. Houlsby, N., Giurgiu, A., Jastrzebski, S., Morrone, B., De Laroussilhe, Q., Gesmundo, A., Attariyan, M., Gelly, S.: Parameter-efficient transfer learning for nlp. In: International Conference on Machine Learning. pp. 2790-2799. PMLR (2019)" + }, + { + "type": "ref_text", + "bbox": [ + 0.218, + 0.232, + 0.787, + 0.272 + ], + "angle": 0, + "content": "16. Hu, E.J., Wallis, P., Allen-Zhu, Z., Li, Y., Wang, S., Wang, L., Chen, W., et al.: Lora: Low-rank adaptation of large language models. In: International Conference on Learning Representations (2021)" + }, + { + "type": "ref_text", + "bbox": [ + 0.218, + 0.273, + 0.787, + 0.314 + ], + "angle": 0, + "content": "17. Jia, M., Tang, L., Chen, B.C., Cardie, C., Belongie, S., Hariharan, B., Lim, S.N.: Visual prompt tuning. In: European Conference on Computer Vision. pp. 709-727. Springer (2022)" + }, + { + "type": "ref_text", + "bbox": [ + 0.218, + 0.315, + 0.787, + 0.355 + ], + "angle": 0, + "content": "18. Jie, S., Deng, Z.H.: Fact: Factor-tuning for lightweight adaptation on vision transformer. In: Proceedings of the AAAI Conference on Artificial Intelligence. vol. 37, pp. 1060-1068 (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.218, + 0.356, + 0.787, + 0.398 + ], + "angle": 0, + "content": "19. Jie, S., Wang, H., Deng, Z.H.: Revisiting the parameter efficiency of adapters from the perspective of precision redundancy. In: Proceedings of the IEEE/CVF International Conference on Computer Vision. pp. 17217-17226 (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.398, + 0.787, + 0.438 + ], + "angle": 0, + "content": "20. Kirillov, A., Mintun, E., Ravi, N., Mao, H., Rolland, C., Gustafson, L., Xiao, T., Whitehead, S., Berg, A.C., Lo, W.Y., et al.: Segment anything. arXiv preprint arXiv:2304.02643 (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.439, + 0.787, + 0.481 + ], + "angle": 0, + "content": "21. Kornblith, S., Shlens, J., Le, Q.V.: Do better imagenet models transfer better? In: Proceedings of the IEEE/CVF conference on computer vision and pattern recognition. pp. 2661-2671 (2019)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.481, + 0.787, + 0.522 + ], + "angle": 0, + "content": "22. Krause, J., Stark, M., Deng, J., Fei-Fei, L.: 3d object representations for fine-grained categorization. In: Proceedings of the IEEE international conference on computer vision workshops. pp. 554-561 (2013)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.522, + 0.787, + 0.549 + ], + "angle": 0, + "content": "23. Larsson, G., Maire, M., Shakhnarovich, G.: Fractalnet: Ultra-deep neural networks without residuals. In: ICLR (2017)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.549, + 0.787, + 0.577 + ], + "angle": 0, + "content": "24. Lester, B., Al-Rfou, R., Constant, N.: The power of scale for parameter-efficient prompt tuning. arXiv preprint arXiv:2104.08691 (2021)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.577, + 0.787, + 0.604 + ], + "angle": 0, + "content": "25. Li, X.L., Liang, P.: Prefix-tuning: Optimizing continuous prompts for generation. arXiv preprint arXiv:2101.00190 (2021)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.605, + 0.787, + 0.646 + ], + "angle": 0, + "content": "26. Li, X., Lian, D., Lu, Z., Bai, J., Chen, Z., Wang, X.: Graphadapter: Tuning vision-language models with dual knowledge graph. Advances in Neural Information Processing Systems 36 (2024)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.646, + 0.787, + 0.688 + ], + "angle": 0, + "content": "27. Lian, D., Zhou, D., Feng, J., Wang, X.: Scaling & shifting your features: A new baseline for efficient model tuning. Advances in Neural Information Processing Systems 35, 109-123 (2022)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.688, + 0.787, + 0.729 + ], + "angle": 0, + "content": "28. Liu, H., Tam, D., Muqeeth, M., Mohta, J., Huang, T., Bansal, M., Raffel, C.A.: Few-shot parameter-efficient fine-tuning is better and cheaper than in-context learning. NeurIPS (2022)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.729, + 0.787, + 0.771 + ], + "angle": 0, + "content": "29. Liu, P., Yuan, W., Fu, J., Jiang, Z., Hayashi, H., Neubig, G.: Pre-train, prompt, and predict: A systematic survey of prompting methods in natural language processing. ACM Computing Surveys 55(9), 1-35 (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.771, + 0.787, + 0.812 + ], + "angle": 0, + "content": "30. Liu, X., Ji, K., Fu, Y., Tam, W.L., Du, Z., Yang, Z., Tang, J.: P-tuning v2: Prompt tuning can be comparable to fine-tuning universally across scales and tasks. arXiv preprint arXiv:2110.07602 (2021)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.812, + 0.787, + 0.84 + ], + "angle": 0, + "content": "31. Liu, Z., Lin, Y., Cao, Y., Hu, H., Wei, Y., Zhang, Z., Lin, S., Guo, B.: Swin transformer: Hierarchical vision transformer using shifted windows. In: Proceedings" + }, + { + "type": "list", + "bbox": [ + 0.217, + 0.147, + 0.787, + 0.84 + ], + "angle": 0, + "content": null + } + ], + [ + { + "type": "header", + "bbox": [ + 0.246, + 0.115, + 0.732, + 0.129 + ], + "angle": 0, + "content": "Parameter-Efficient and Memory-Efficient Tuning for Vision Transformer" + }, + { + "type": "page_number", + "bbox": [ + 0.768, + 0.116, + 0.786, + 0.127 + ], + "angle": 0, + "content": "17" + }, + { + "type": "ref_text", + "bbox": [ + 0.245, + 0.147, + 0.785, + 0.175 + ], + "angle": 0, + "content": "of the IEEE/CVF international conference on computer vision. pp. 10012-10022 (2021)" + }, + { + "type": "ref_text", + "bbox": [ + 0.218, + 0.176, + 0.786, + 0.217 + ], + "angle": 0, + "content": "32. Liu, Z., Mao, H., Wu, C.Y., Feichtenhofer, C., Darrell, T., Xie, S.: A convnet for the 2020s. In: Proceedings of the IEEE/CVF conference on computer vision and pattern recognition. pp. 11976-11986 (2022)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.218, + 0.786, + 0.244 + ], + "angle": 0, + "content": "33. Loshchilov, I., Hutter, F.: Decoupled weight decay regularization. arXiv preprint arXiv:1711.05101 (2017)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.245, + 0.786, + 0.286 + ], + "angle": 0, + "content": "34. Lu, Z., Bai, J., Li, X., Xiao, Z., Wang, X.: Beyond sole strength: Customized ensembles for generalized vision-language models. In: Forty-first International Conference on Machine Learning (2024)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.287, + 0.786, + 0.314 + ], + "angle": 0, + "content": "35. Van der Maaten, L., Hinton, G.: Visualizing data using t-sne. Journal of machine learning research 9(11) (2008)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.315, + 0.786, + 0.342 + ], + "angle": 0, + "content": "36. Maji, S., Rahtu, E., Kannala, J., Blaschko, M., Vedaldi, A.: Fine-grained visual classification of aircraft. arXiv preprint arXiv:1306.5151 (2013)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.343, + 0.786, + 0.383 + ], + "angle": 0, + "content": "37. Nilsback, M.E., Zisserman, A.: A visual vocabulary for flower classification. In: 2006 IEEE computer society conference on computer vision and pattern recognition (CVPR'06). vol. 2, pp. 1447-1454. IEEE (2006)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.384, + 0.786, + 0.424 + ], + "angle": 0, + "content": "38. Parkhi, O.M., Vedaldi, A., Zisserman, A., Jawahar, C.: Cats and dogs. In: 2012 IEEE conference on computer vision and pattern recognition. pp. 3498-3505. IEEE (2012)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.425, + 0.786, + 0.466 + ], + "angle": 0, + "content": "39. Pfeiffer, J., Kamath, A., Rückle, A., Cho, K., Gurevych, I.: Adapterfusion: Nondestructive task composition for transfer learning. arXiv preprint arXiv:2005.00247 (2020)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.467, + 0.786, + 0.494 + ], + "angle": 0, + "content": "40. Rebuffi, S.A., Bilen, H., Vedaldi, A.: Learning multiple visual domains with residual adapters. Advances in neural information processing systems 30 (2017)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.495, + 0.786, + 0.522 + ], + "angle": 0, + "content": "41. Shao, L., Zhu, F., Li, X.: Transfer learning for visual categorization: A survey. IEEE transactions on neural networks and learning systems 26(5), 1019-1034 (2014)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.523, + 0.786, + 0.563 + ], + "angle": 0, + "content": "42. Srivastava, N., Hinton, G., Krizhevsky, A., Sutskever, I., Salakhutdinov, R.: Dropout: a simple way to prevent neural networks from overfitting. The journal of machine learning research 15(1), 1929-1958 (2014)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.565, + 0.786, + 0.604 + ], + "angle": 0, + "content": "43. Sung, Y.L., Cho, J., Bansal, M.: Lst: Ladder side-tuning for parameter and memory efficient transfer learning. Advances in Neural Information Processing Systems 35, 12991-13005 (2022)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.605, + 0.786, + 0.646 + ], + "angle": 0, + "content": "44. Sung, Y.L., Cho, J., Bansal, M.: Vl-adapter: Parameter-efficient transfer learning for vision-and-language tasks. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 5227-5237 (2022)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.647, + 0.786, + 0.701 + ], + "angle": 0, + "content": "45. Tu, C.H., Mai, Z., Chao, W.L.: Visual query tuning: Towards effective usage of intermediate representations for parameter and memory efficient transfer learning. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 7725-7735 (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.702, + 0.786, + 0.742 + ], + "angle": 0, + "content": "46. Vaswani, A., Shazeer, N., Parmar, N., Uszkoreit, J., Jones, L., Gomez, A.N., Kaiser, L., Polosukhin, I.: Attention is all you need. Advances in neural information processing systems 30 (2017)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.743, + 0.786, + 0.785 + ], + "angle": 0, + "content": "47. Wan, L., Zeiler, M., Zhang, S., Le Cun, Y., Fergus, R.: Regularization of neural networks using dropconnect. In: International conference on machine learning. pp. 1058-1066. PMLR (2013)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.786, + 0.786, + 0.812 + ], + "angle": 0, + "content": "48. Wang, M., Deng, W.: Deep visual domain adaptation: A survey. Neurocomputing 312, 135-153 (2018)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.813, + 0.786, + 0.84 + ], + "angle": 0, + "content": "49. Wang, W., Bao, H., Dong, L., Bjorck, J., Peng, Z., Liu, Q., Aggarwal, K., Mohammed, O.K., Singhal, S., Som, S., et al.: Image as a foreign language: Beit pre" + }, + { + "type": "list", + "bbox": [ + 0.217, + 0.147, + 0.786, + 0.84 + ], + "angle": 0, + "content": null + } + ], + [ + { + "type": "page_number", + "bbox": [ + 0.218, + 0.116, + 0.236, + 0.127 + ], + "angle": 0, + "content": "18" + }, + { + "type": "header", + "bbox": [ + 0.272, + 0.115, + 0.49, + 0.129 + ], + "angle": 0, + "content": "Taolin Zhang, Jiawang Bai et al." + }, + { + "type": "ref_text", + "bbox": [ + 0.241, + 0.147, + 0.786, + 0.175 + ], + "angle": 0, + "content": "training for all vision and vision-language tasks. arXiv preprint arXiv:2208.10442 (2022)" + }, + { + "type": "ref_text", + "bbox": [ + 0.218, + 0.176, + 0.787, + 0.231 + ], + "angle": 0, + "content": "50. Yang, S., Bai, J., Gao, K., Yang, Y., Li, Y., Xia, S.T.: Not all prompts are secure: A switchable backdoor attack against pre-trained vision transformers. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 24431-24441 (2024)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.232, + 0.786, + 0.259 + ], + "angle": 0, + "content": "51. Yang, X., Ye, J., Wang, X.: Factorizing knowledge in neural networks. In: European Conference on Computer Vision. pp. 73-91. Springer (2022)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.26, + 0.786, + 0.286 + ], + "angle": 0, + "content": "52. Yang, X., Zhou, D., Liu, S., Ye, J., Wang, X.: Deep model reassembly. Advances in neural information processing systems 35, 25739-25753 (2022)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.288, + 0.786, + 0.327 + ], + "angle": 0, + "content": "53. Yuan, M., Lin, Y.: Model selection and estimation in regression with grouped variables. Journal of the Royal Statistical Society Series B: Statistical Methodology 68(1), 49-67 (2006)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.328, + 0.786, + 0.37 + ], + "angle": 0, + "content": "54. Zaken, E.B., Ravfogel, S., Goldberg, Y.: Bitfit: Simple parameter-efficient fine-tuning for transformer-based masked language-models. arXiv preprint arXiv:2106.10199 (2021)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.371, + 0.786, + 0.397 + ], + "angle": 0, + "content": "55. Zang, Y., Li, W., Zhou, K., Huang, C., Loy, C.C.: Unified vision and language prompt learning. arXiv preprint arXiv:2210.07225 (2022)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.398, + 0.786, + 0.439 + ], + "angle": 0, + "content": "56. Zha, Y., Wang, J., Dai, T., Chen, B., Wang, Z., Xia, S.T.: Instance-aware dynamic prompt tuning for pre-trained point cloud models. In: Proceedings of the IEEE/CVF International Conference on Computer Vision. pp. 14161-14170 (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.439, + 0.786, + 0.481 + ], + "angle": 0, + "content": "57. Zhai, X., Kolesnikov, A., Houlsby, N., Beyer, L.: Scaling vision transformers. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 12104-12113 (2022)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.481, + 0.786, + 0.522 + ], + "angle": 0, + "content": "58. Zhai, X., Puigcerver, J., Kolesnikov, A., Ruyssen, P., Riquelme, C., Lucic, M., Djolonga, J., Pinto, A.S., Neumann, M., Dosovitskiy, A., et al.: The visual task adaptation benchmark (2019)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.522, + 0.681, + 0.536 + ], + "angle": 0, + "content": "59. Zhang, Y., Zhou, K., Liu, Z.: Neural prompt search. arXiv (2022)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.537, + 0.786, + 0.577 + ], + "angle": 0, + "content": "60. Zhou, K., Yang, J., Loy, C.C., Liu, Z.: Conditional prompt learning for vision-language models. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 16816-16825 (2022)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.578, + 0.786, + 0.605 + ], + "angle": 0, + "content": "61. Zhou, K., Yang, J., Loy, C.C., Liu, Z.: Learning to prompt for vision-language models. International Journal of Computer Vision 130(9), 2337-2348 (2022)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.606, + 0.786, + 0.647 + ], + "angle": 0, + "content": "62. Zhu, L., Liao, B., Zhang, Q., Wang, X., Liu, W., Wang, X.: Vision mamba: Efficient visual representation learning with bidirectional state space model. arXiv preprint arXiv:2401.09417 (2024)" + }, + { + "type": "list", + "bbox": [ + 0.217, + 0.147, + 0.787, + 0.647 + ], + "angle": 0, + "content": null + } + ] +] \ No newline at end of file diff --git a/2024/Parameter-Efficient and Memory-Efficient Tuning for Vision Transformer_ A Disentangled Approach/e724c6a7-eae9-4271-a988-9b07cf4a65fc_origin.pdf b/2024/Parameter-Efficient and Memory-Efficient Tuning for Vision Transformer_ A Disentangled Approach/e724c6a7-eae9-4271-a988-9b07cf4a65fc_origin.pdf new file mode 100644 index 0000000000000000000000000000000000000000..cf1ae9f43b2b3470492985c0a867230893ccf85b --- /dev/null +++ b/2024/Parameter-Efficient and Memory-Efficient Tuning for Vision Transformer_ A Disentangled Approach/e724c6a7-eae9-4271-a988-9b07cf4a65fc_origin.pdf @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:b61745498c1744ac4285b2348bacca393c4671d278d5227d761bf65173a69e80 +size 2120183 diff --git a/2024/Parameter-Efficient and Memory-Efficient Tuning for Vision Transformer_ A Disentangled Approach/full.md b/2024/Parameter-Efficient and Memory-Efficient Tuning for Vision Transformer_ A Disentangled Approach/full.md new file mode 100644 index 0000000000000000000000000000000000000000..457148579e81dd5602fb73c53dc6abf23b497627 --- /dev/null +++ b/2024/Parameter-Efficient and Memory-Efficient Tuning for Vision Transformer_ A Disentangled Approach/full.md @@ -0,0 +1,344 @@ +# Parameter-Efficient and Memory-Efficient Tuning for Vision Transformer: A Disentangled Approach + +Taolin Zhang $^{1*}$ , Jiawang Bai $^{2,3*}$ , Zhihe Lu $^{3}$ , Dongze Lian $^{3}$ , Genping Wang $^{4,\text{凹}}$ , Xinchao Wang $^{3,\text{凹}}$ , and Shu-Tao Xia $^{1,5}$ + +$^{1}$ Tsinghua Shenzhen International Graduate School, Tsinghua University $^{2}$ Tencent $^{3}$ National University of Singapore $^{4}$ Shenzhen Qiji Technology Co., Ltd. + +$^{5}$ Research Center of Artificial Intelligence, Peng Cheng Laboratory ztl23@mails.tsinghua.edu.cn; jiawangbai@tencent.com; zhihelu.academic@gmail.com; dzlianx@gmail.com; genpingwang@163.com; xinchao@nus.edu.sg; xiast@sz.tsinghua.edu.cn + +Abstract. Recent works on parameter-efficient transfer learning (PETL) show the potential to adapt a pre-trained Vision Transformer to downstream recognition tasks with only a few learnable parameters. However, since they usually insert new structures into the pre-trained model, entire intermediate features of that model are changed and thus need to be stored to be involved in back-propagation, resulting in memory-heavy training. We solve this problem from a novel disentangled perspective, i.e., dividing PETL into two aspects: task-specific learning and pretrained knowledge utilization. Specifically, we synthesize the task-specific query with a learnable and lightweight module, which is independent of the pre-trained model. The synthesized query equipped with task-specific knowledge serves to extract the useful features for downstream tasks from the intermediate representations of the pre-trained model in a query-only manner. Built upon these features, a customized classification head is proposed to make the prediction for the input sample. Given that our method employs an extremely lightweight architecture and avoids the use of heavy intermediate features for running gradient descent, it demonstrates limited memory usage in training. Extensive experiments manifest that our method achieves state-of-the-art performance under memory constraints, showcasing its applicability in real-world situations. The code is available at: https://synqt.github.io/. + +# 1 Introduction + +Motivated by the success of data-driven solutions, there has been an explosive growth in the size of vision models [4,10,13,20,31,49,62] and the scale of accessible datasets like ImageNet [9]. To unleash the power of large models, it becomes a tendency to first pre-train on a general and large-scale dataset, and then adapt + +them to downstream tasks [21, 41, 48, 52]. The most obvious adaption strategy is to fully fine-tune all parameters of the pre-trained model. However, it necessitates that the device saves a dedicated set of model parameters for each single task, leading to substantial storage overhead [57]. + +To improve the storage efficiency, parameter-efficient transfer learning (PETL) methods [8,16-19,59] have been explored by tuning only a small set of the parameters in adapting the large pre-trained models to downstream tasks. Typically, these methods insert small structures into the frozen pre-trained model and specify the newly added parameters as the only learnable ones. For instance, Visual Prompt Tuning (VPT) [17] prepends learnable tokens to the patch embeddings; AdaptFormer [8] inserts a bottleneck-structured fully connected layers parallel to the feed-forward network (FFN) of the original ViT. These studies demonstrate that PETL is a promising way to save storage resources. + +However, even though updating a small set of parameters, existing PETL methods still require a very high consumption of memory in training. The direct reason is that they need a full back-propagation throughout the entire backbone. Given the fact that the model size is exponentially increasing, it becomes infeasible to apply existing PETL methods to the memory-constraint situations. Therefore, it is desirable to investigate memory-efficient PETL while maintaining performance on downstream tasks. + +In this work, we attribute the high memory cost of existing PETL methods in training to their intrinsic regime, i.e., entangling task-specific learning and pre-trained knowledge utilization together. It is usually achieved by integrating trainable structures into the pre-trained ViT and changing all intermediate features. Consequently, a great number of intermediate features need to be stored in training to run gradient descent. The aforementioned analysis motivates us to design a memory-efficient PETL method from a disentangled perspective. + +Accordingly, we propose Synthesized Query Tuning (SynQT), a memory-efficient PETL method for adapting pre-trained ViTs to downstream recognition tasks. Overall, as shown in Figure 1, SynQT takes the intermediate representations of the original pre-trained model as input and combines them with task-specific knowledge using stacked blocks. We propose the Query Synthesis Module (QSM) and Knowledge Extraction Module (KEM) to accomplish the disentanglement. To capture task-specific knowledge, SynQT employs a lightweight QSM, which is independent of the pre-trained model. In QSM, an attention layer and an FFN are trained on the target downstream task to synthesize task-specific query. Then, to utilize the pre-trained knowledge, the synthesized query interacts with the intermediate representations of the original pre-trained model in a KEM. It outputs useful features for downstream recognition, where all parameters are inherited from the pre-trained model and frozen. Relying on rich hierarchical features from KEMs, we propose a classification head, which adaptively aggregates these features conditioned on the input sample for prediction. + +The success of SynQT in saving memory in training is derived from two technical aspects. Firstly, we keep intermediate representations of the original pretrained model intact and directly exploit them in a query-only manner, avoiding + +the use of heavy intermediate features for running gradient descent. Besides, our design including QSM and KEM is extremely lightweight. For instance, we adopt bottleneck structures for Q-K-V projection [46] and a small number of tokens (lower than 4 in most cases) in QSM. + +We validate SynQT on VTAB-1K [58] benchmark, including 19 tasks from diverse domains. We also evaluate our method in few-shot learning using 5 fine-grained visual classification datasets. Extensive experiments show that SynQT can significantly reduce the memory footprint in training with promising accuracy. Compared to VPT-Deep, a powerful version of VPT that inserts prompts in each layer of ViT, SynQT saves $56\%$ memory consumption and achieves higher Top-1 accuracy on average. The contributions of our work are as follows: + +1. We provide an insightful analysis of the high memory cost of existing PETL methods, i.e., they entangle task-specific learning and pre-trained knowledge utilization together. +2. Inspired by the disentangled idea, we propose SynQT, which synthesizes task-specific query and then utilizes the pre-trained knowledge in a query-only manner. +3. Extensive experiments show that SynQT is highly memory-efficient and effective, notably outperforming existing baselines under memory constraints. + +# 2 Related Work + +Parameter-Efficient Transfer Learning (PETL). The idea of PETL initially shows great success in the field of natural language processing (NLP) [6,16,28,40]. It fine-tunes a small number of trainable parameters to transfer large pre-trained models to downstream tasks, significantly reducing storage costs by avoiding a copy of model parameters for each new task. Due to the increased size of pre-trained vision models, especially for ViT [10], PETL has been introduced to computer vision and attracted much attention [3,8,12,17,26,27,34,39,44, 45,50,55,56,60,61]. To reduce the number of fine-tuned parameters, BitFit [54] only fine-tunes the bias terms and freezes most of the network, resulting in comparable performance to full tuning. VPT [17] introduces learnable tokens (i.e., prompts [24,25,29,30]) to each layer of ViT. SSF [27] fine-tunes scale and shift factors in each layer to adapt original features to the target domain. Some other works [8,15,16] propose to insert additional trainable structures into attention blocks of the pre-trained Transformer. For instance, LoRA [16] observes that the learned over-parametrized models reside on a low-rank dimension and further injects trainable rank decomposition parameters into each layer of the Transformer; AdaptFormer [8] adds parallel lightweight module to FFN. Furthermore, for extreme parameter efficiency, a tensorization-decomposition framework [18] is exploited to store the weight increments, and the low-bit adapter [19] is proposed to reduce the precision redundancy. However, how to improve memory efficiency in training is still an under-explored problem in PETL, as discussed below. + +![](images/6b9e702486b0bd3db86ad0a5cac53de8656c0e72bda71ddf5c90b382220bf160.jpg) +Fig. 1: Overall architecture of SynQT. Each block in SynQT contains two key designs, including Query Synthesis Module (QSM) and Knowledge Extractor Module (KEM). QSM is a lightweight learnable module to capture task-specific knowledge and KEM reuses frozen weights in the pre-trained model for feature extraction. The rich and hierarchical features from KEMs are fed into the classification head. + +Memory-Efficient PETL. Although existing PETL methods have shown promising performance in transfer learning while saving most trainable parameters, they still suffer from issues of high memory consumption in training. The full back-propagation throughout the entire backbone prevents traditional PETL methods from being applied in memory-constraint situations. Recently, some recent works [11,43,45] have made several attempts to address this issue, whose key idea is to leverage the power of intermediate representations of the pre-trained model and thus avoid full back-propagation. In NLP, LST [43] proposes a trainable lightweight structure in the sidewalk for memory-saving during training. To adapt vision models, Head2Toe [11] selects features from all layers of the source model to feed into a classification head. VQT [45] inserts trainable tokens in the query space to capture useful information from the pre-trained model. Due to the information redundancy in the intermediate representations, both Head2Toe and VQT further adopt group lasso [1,53] for feature selection. + +In contrast to previous works, we provide an insightful analysis that entangled combination of task-specific learning and pre-trained knowledge utilization results in high memory consumption of traditional PETL methods. Therefore, we propose SynQT to disentangle PETL into two distinct modules, with the aim of capturing task-specific knowledge [51] and extracting useful features for downstream tasks, respectively. Such a disentangled paradigm enjoys memory efficiency in transfer learning compared with traditional PETL methods. + +# 3 Method + +# 3.1 Overall Framework + +As can be seen from Figures 1 and 2, SynQT makes the prediction relying on the features produced by stacked blocks, which take the intermediate representations + +of the original pre-trained model as external input. In SynQT, each block mainly includes two parts: Query Synthesis Module (QSM) and Knowledge Extraction Module (KEM). QSM is a lightweight module, which is used to capture task-specific knowledge, connecting the pre-trained model with the downstream task. KEM uses the tokens generated by QSM as queries, and the intermediate features in the pre-trained ViT as keys and values. It reuses the original model's weights to extract features that are beneficial to downstream tasks. Finally, the rich hierarchical features extracted by the KEMs are fed into the classification head, as shown in Figure 2. We will elaborate our designs in the following sections. + +# 3.2 Query Synthesis Module + +PETL integrates the task-specific knowledge into the adapted model through training on the downstream dataset. Typically, existing methods [8, 15-17] insert additional trainable structures into the pre-trained model. However, this is the root of high memory consumption, since it changes all original features, and all intermediate features are modified and stored to be involved in calculating gradients. From the disentangled perspective, we solve this problem with a task-specific learning module, named Query Synthesis Module (QSM), which is trainable, lightweight, and structurally independent from the pre-trained model. + +The purpose of QSM is to synthesize query tokens, interacting with intermediate representations of the original pre-trained model. Besides outputs from the previous block, we introduce a trainable prompt into the input space of each QSM. These per-block additional prompts increase the capacity of QSM to capture task-specific knowledge. Inspired by the Transformer, QSM includes an attention layer and an FFN. Specifically, the input tokens interact with each other through a self-attention layer and are further fed into the feed-forward network for producing the query tokens. Besides, to keep the parameter efficiency, another design principle is to extensively use bottleneck-structured fully connected layers to reduce the number of parameters. + +Let $n$ and $d$ denote the number of tokens and the hidden size in QSM, respectively. Note that $n$ is very small, e.g., $n = 4$ , much lower than the number of tokens in the original model and the value of $d$ corresponds to that of the pre-trained model. Before the self-attention layer, the output of the last block $\hat{H}_{i - 1} \in \mathbb{R}^{n \times d}$ are added with trainable prompt $P_{i}$ and then fed into bottleneck projection layers. It is formally formulated as follows, + +$$ +Z _ {i} ^ {\prime} = \left(\left(H _ {i} + P _ {i}\right) \cdot \mathbf {W} _ {\text {d o w n}}\right) \cdot \mathbf {W} _ {\text {u p}}, \tag {1} +$$ + +where $\mathbf{W}_{down} \in \mathbb{R}^{d \times \hat{d}}$ and $\mathbf{W}_{up} \in \mathbb{R}^{\hat{d} \times d}$ denote the down-projection and up-projection layer respectively, with $\hat{d} \ll d$ . + +In the attention layer, $Z_{i}^{\prime}$ is projected by three bottleneck modules to produce the query $Q_{i}^{qs}$ , key $K_{i}^{qs}$ , and value $V_{i}^{qs}$ . Then the hidden embedding after attention layers $Z_{i}^{\prime \prime}$ can be calculated as follows, controlled by a scale factor $s^{\prime}$ : + +$$ +Z _ {i} ^ {\prime \prime} = s ^ {\prime} \cdot \operatorname {A t t n} \left(Q _ {i} ^ {q s}, K _ {i} ^ {q s}, V _ {i} ^ {q s}\right) + Z _ {i} ^ {\prime}. \tag {2} +$$ + +Then we can further obtain the synthesized query $\hat{H}_i$ after the FFN as the input of KEM, controlled by another scale factor $s''$ : + +$$ +\hat {H} _ {i} = s ^ {\prime \prime} \cdot \operatorname {F F N} \left(\operatorname {L N} \left(Z _ {i} ^ {\prime \prime}\right)\right) + Z _ {i} ^ {\prime \prime}, \tag {3} +$$ + +where FFN also consists of bottleneck-structured fully connected layers and $\mathrm{LN}(\cdot)$ represents the layer normalization [2]. $s^{\prime}$ and $s''$ are hyper-parameters to modulate the effects of the attention and FFN on the synthesized query. Since QSM is trained on the downstream dataset from scratch, we believe that its output $\hat{H}_i$ captures the task-specific knowledge. We use $\hat{H}_i$ in a query-only manner as illustrated in the next section. + +# 3.3 Knowledge Extraction Module + +With the task-specific query obtained from QSM, we propose Knowledge Extraction Module (KEM) to exploit intermediate features of the original pre-trained model. Before that, we obtain original features $\{X_{1}, X_{2}, \dots, X_{l}\}$ of the input sample from forward-propagation, where $l$ is the number of transformer blocks of the pre-train model. Note that this process is efficient due to no gradient calculation. + +KEM takes the synthesized query $\hat{H}_i$ and original features $\{X_1, X_2, \dots, X_l\}$ as inputs and outputs useful features for the downstream task. Since $\hat{H}_i$ has a small number of tokens and original features are only queried by $\hat{H}_i$ , there are no heavy features in the middle of KEM. Besides, we find that it is effective to reuse and freeze the pre-trained parameters for feature interaction in KEM. These designs ensure the parameter- and memory-efficiency of our method. + +Specifically, with the projection weight in the pre-trained ViT blocks, we first obtain the query $Q_{i}^{ke}$ , key $K_{i}^{ke}$ , and value $V_{i}^{ke}$ . Then follow-up attention calculation is performed by: + +$$ +E _ {i} = \operatorname {A t t n} \left(Q _ {i} ^ {k e}, K _ {i} ^ {k e}, V _ {i} ^ {k e}\right) + \hat {H} _ {i}. \tag {4} +$$ + +We also reuse the weights of FFN to perform the same space transformation as the pre-trained model: + +$$ +H _ {i} = \operatorname {F F N} \left(\operatorname {N o r m} \left(E _ {i}\right)\right) + E _ {i}, \tag {5} +$$ + +where $\mathrm{Norm}(\cdot)$ denotes the normalization used in the pre-trained model. + +To maximize the potential of KEM, apart from the final output $H_{i}$ , we additionally leverage features $F_{i}^{att}$ and $F_{i}^{ffn}$ after the attention module and the FFN for classification. They are obtained as follows: + +$$ +\left\{ \begin{array}{l} F _ {i} ^ {a t t} = \operatorname {A t t n} \left(Q _ {i} ^ {k e}, K _ {i} ^ {k e}, V _ {i} ^ {k e}\right) \\ F _ {i} ^ {f f n} = \operatorname {F F N} \left(\operatorname {N o r m} \left(E _ {i}\right)\right) \end{array} . \right. \tag {6} +$$ + +# 3.4 Classification Head + +The classification head makes a prediction of the input sample relying on rich and hierarchical features from KEMs, $\{F_1^{att}, F_1^{ffn}, H_1, \dots, F_l^{att}, F_l^{ffn}, H_l\}$ , where $l$ + +is the number of blocks of the pre-trained ViT. Instead of directly applying an MLP classifier on top of these features, we propose several techniques for better performance. Figure 2 shows the overall architecture of the classification head. + +Feature projection. Since these features are generated across different layers from KEMs, they may be misaligned in the representation space. Therefore, before aggregating all features, we align all features with projection layers. We use bottleneck-structured fully connected layers and share the projection layers across all features to improve the parameter efficiency. + +Conditional feature aggregation. These features contain a large amount of information but also some redundancy, perhaps hurting the performance of downstream tasks. Therefore, it is crucial to perform feature selection, aiming to extract the truly beneficial information. We propose a strategy of conditional feature aggregation, which generates weights conditioned on $H_{l}$ and performs a weighted-sum over all features. Our strategy enables SynQT to adaptively rely on different features to classify different samples. Despite only a few parameters, the conditional feature aggregation results in better performance compared to a simple average or weighted sum with learnable but fixed weights, as shown in our experiments. + +DropFeat in training. Due to rich features fed into the classification head, overfitting likely happens and degrades the performance. To solve this issue, inspired by the idea of dropout [23, 42, 47], we propose DropFeat, a simple regularization technique to prevent overfitting. DropFeat randomly removes some input features for the classification head during training, while all features are leveraged during inference. DropFeat ensures that the model does not overly rely on a specific feature, which significantly improves the model's performance. + +![](images/e2c9082112eeda8075dbba1853437621577731d2ee22dc2c4b3a2831bdbde3c7.jpg) +Fig.2: Classification head in SynQT, corresponding to the head in Figure 1. We utilize a shared projection to align the hierarchical features and apply feature aggregation conditioned on the output of the last block $H_{l}$ . We also adopt DropFeat to prevent overfitting, which randomly drops some features during training. + +# 4 Experiment + +# 4.1 Setup + +Datasets. We conduct experiments on the VTAB-1K benchmark and the fine-grained visual classification (FGVC) datasets. VTAB-1K consists of 19 different classification tasks categorized into three groups: Natural, Specialized, and Structured. Each task of VTAB-1K contains 1000 training samples. We evaluate the + +Table 1: Full results on VTAB-1K benchmark with pre-trained ViT-B/16 on ImageNet21K. “# Memory” specifies memory cost during training with 64 as batch size on CIFAR100. We report the Top-1 accuracy and the results are averaged over group-wise mean values. + +
# Memory (GB)NaturalSpecializedStructured
CIFAR100Caltech101DTDFlower102PetsSVHNSun397CamelyonEuroSATRess45RetinopathyClever-CountClever-DistDMLabKITTL-DistdSpr-LocdSpr-OrisNORB-AzimsNORB-EleAverage
Traditional Fine-Tuning
Linear Probing2.550.685.661.479.586.540.838.079.791.571.765.541.434.434.155.418.126.416.524.852.7
Full Fine-Tuning11.144.384.554.184.774.787.226.985.395.076.070.471.560.546.972.974.538.728.523.863.2
PETL Methods
BitFit [54]7.872.887.059.297.585.359.951.478.791.672.969.861.555.632.455.966.640.015.725.165.2
VPT-Shallow [17]8.177.786.962.697.587.374.551.278.292.075.672.950.558.640.567.168.736.120.234.167.8
VPT-Deep [17]8.178.890.865.898.088.378.149.681.896.183.468.468.560.046.572.873.647.932.937.872.0
Memory-Efficient PETL Methods
VQT [45]3.266.389.967.897.984.779.945.579.095.280.974.746.761.645.163.662.932.130.028.868.3
SynQT (Ours)3.470.989.768.898.589.677.850.682.396.783.575.271.862.748.575.474.149.031.736.172.9
+ +few-shot performance of SynQT on 5 FGVC datasets: FGVC-Aircraft [36], OxfordPets [38], Food-101 [5], Stanford Cars [22], and Oxford-Flowers102 [37]. We follow the split in [59] and report the results in 1, 2, 4, 8, and 16-shot settings. The accuracy is averaged over three trials with different seeds. + +Implementation details. We utilize a pre-trained ViT-B/16 [10] on the ImageNet-21K dataset as the foundational model. We also conduct experiments on the ImageNet-1K supervised pre-trained backbone, as well as on the self-supervised checkpoints (DINO [7] and MAE [14]). For a lightweight design of QSM, we use an extremely small number of tokens, with the number of tokens $n$ from $\{1, 4, 16\}$ . This is significantly fewer than the number of tokens in VPT, which can reach up to 200. Additionally, the hyper-parameters $s'$ and $s''$ are chosen from $\{0.1, 1\}$ . The hidden size of the bottleneck projection is set to 48, except attention projection of the query, key, and value in QSM, which is set to 8. + +Due to the limited number of validation samples in VTAB-1K, we implement 5-fold cross-validation on the training sets to determine the optimal hyperparameters. For few-shot learning, we use the default splits of FGVC datasets to search for the best hyperparameters. The Top-1 accuracy is reported on the test sets in all our experiments. Following common practice [18], we adopt AdamW [33] as the optimizer and a cosine learning rate scheduler with a constant batch size of 64. For data augmentation, we adjust the image size to $224 \times 224$ directly for VTAB-1K, and for few-shot learning, we resize the image to $256 \times 256$ before applying a center crop. + +# 4.2 Main Results + +Besides the traditional fine-tuning methods (linear probing and full fine-tuning), we compare our method with PETL methods including BitFit [54], VPT-Shallow [17], and VPT-Deep [17], as well as a memory-efficient one, VQT [45], on VTAB- + +Table 2: Classification results on MAE, DINO, and Supervised ImageNet-1K pre-trained backbones. We report the Top-1 accuracy and the averaged results. + +
MethodsNaturalSpecializedStructuredAverage
MAE backbone
BitFit [54]63.580.657.267.1
VPT-Shallow [17]38.372.025.045.1
VPT-Deep [17]50.876.437.354.8
VQT [45]56.678.643.459.5
SynQT (Ours)66.082.658.268.9
DINO backbone
BitFit [54]76.883.849.269.9
VPT-Shallow [17]69.882.638.563.6
VPT-Deep [17]76.783.847.369.3
VQT [45]71.083.637.263.9
SynQT (Ours)75.184.856.972.3
Supervised ImageNet-1K backbone
BitFit [54]74.683.449.469.1
VPT-Shallow [17]63.780.340.261.4
VPT-Deep [17]74.982.953.970.6
VQT [45]72.784.549.368.8
SynQT (Ours)73.284.756.271.4
+ +1K. As can be seen from Table 1, the most striking observation emerging from the comparison is that SynQT outperforms other methods while saving up to $50\%$ of GPU memory. Given the same batch size, previous PETL methods rely on the full backward propagation of the original model and consume $7 \sim 8$ GB of memory, whereas SynQT disentangles transfer learning and reduces the memory cost to $3.4$ GB. Notably, our method performs better than VPT-Deep on 14 of 19 tasks. We notice that although the memory cost of SynQT is slightly higher, it achieves a significant performance improvement of $4.6\%$ compared to VQT. + +In order to further validate the robustness of our method, we compare SynQT with baseline models across various checkpoints which are derived from different learning paradigms, including supervised learning (pretrained over ImageNet-21K [9]), and self-supervised learning (DINO + +[7] and MAE [14]). The results in Table 2 indicate that SynQT shows strong compatibility and outperforms PETL methods by a substantial margin over different backbones. For instance, compared to the strongest baselines, our method achieves improvements of $1.8\%$ , $3.3\%$ , and $0.8\%$ on average with MAE, DINO, and supervised ImageNet-1K backbones, respectively. + +![](images/ca231b864699c3f170917ef9f73482e8e8b38cb1d2960eea8e28b4cca27a12ec.jpg) +Fig.3: Performance comparison under memory constraints. SynQT outperforms other methods with different memory costs. + +# 4.3 Memory-Efficiency Evaluation + +Previous PETL methods [8, 15, 16, 18] focus on reducing the number of parameters when + +Table 3: Full results on VTAB-1K benchmark under memory constraints. The memory is constrained within 5GB and “# Memory” specifies the memory cost during training with 64 as batch size on CIFAR100. We report the Top-1 accuracy and the results are averaged over group-wise mean values. + +
# Memory (GB)NaturalSpecializedStructured
CIFAR100Caltech101DTDFlower102PetsSVHNSun397CanelyonEuroSATResics45RetinopathyClevr-CountClevr-DistDMLabKITTI-DistdSpr-LocdSpr-OrisNORB-AzimsNORB-EteAverage
FacT-TT [18]4.340.484.259.896.886.653.445.879.492.273.374.459.250.242.568.848.225.919.828.763.2
Adapter [15]4.764.986.062.897.386.646.844.479.893.976.774.663.340.238.968.533.132.319.830.564.0
LoRA [16]4.862.582.160.497.586.658.645.380.695.378.872.864.845.141.368.549.628.124.630.565.5
AdaptFormer [8]4.665.686.162.096.986.235.545.682.696.277.674.466.247.741.372.355.629.422.633.965.8
SynQT(Ours)3.470.989.768.898.589.677.850.682.396.783.575.271.862.748.575.474.149.031.736.172.9
+ +fine-tuning the model. While decreasing the cost of model storage, these methods are not widely applicable in practical training because they still require a significant amount of GPU memory. Therefore, we propose SynQT to overcome this problem. In this section, we highlight our improvements in terms of memory costs by comparing our method with some SOTA PETL methods under a given GPU memory constraint. + +For baseline methods, we discard some specific trainable structures to meet the memory constraint. Specifically, we remove these additional structures near the input layer, which means that the training only requires a partial backpropagation with a lower memory usage. We progressively remove the structure nearest to the input layer until it meets the given memory constraint. Besides, for a fair comparison, we reduce the memory usage of SynQT in the same manner. + +We compare SynQT with SOTA baselines including FacT-TT [18], Adapter [15], LoRA [16] and AdaptFormer [8] under such a constraint and report the results in Table 3. While previous methods can achieve promising performance with few trainable parameters, they unfortunately suffer from significant performance drops when such memory constraints are imposed. In contrast, SynQT excels in this scenario, achieving state-of-the-art performance and surpassing AdaptFormer by $7.1\%$ and VQT by $4.6\%$ . + +To have a better understanding of how memory constraints affect model performance, we estimate the results of the methods mentioned above across varying memory costs. As depicted in Figure 3, SynQT continues to outperform the baselines across different memory costs. Notably, when the constraint is very strict, i.e., 3 GB, most of the existing PETL methods perform poorly, with a gap of almost $7\%$ compared to SynQT. Moreover, it is noteworthy that relaxing the memory constraints improves the performance of all the methods, especially for a significant improvement observed for our SynQT. + +# 4.4 Few-Shot Learning on FGVC + +To further evaluate SynQT's ability to transfer knowledge from the pre-trained model using few samples, we compare SynQT with BitFit [54], VPT-Shallow [17], + +![](images/cf53836e9aaebae18596c30f1248303246867fe49cef0b84359871c698c159f5.jpg) +Fig. 4: Performance comparison in few-shot learning. Overall, SynQT outperforms other baselines in the extreme low-data regime. + +Table 4: Applying SynQT to Swin-B and ConvNeXt-B. We report the Top-1 accuracy and the averaged results. + +
MethodSwin-BConvNeXt-B
NaturalSpecializedStructuredAverageNaturalSpecializedStructuredAverage
Linear73.580.833.562.674.581.534.863.6
VPT-Deep76.884.553.471.678.583.044.668.7
SynQT77.586.057.873.878.182.148.269.5
+ +VPT-Deep [17], and VQT [45] in few-shot setting on 5 FGVC datasets. The results are reported in Figure 4. + +Overall, SynQT significantly outperforms all other baselines in 4, 6, 8, and 16-shot scenarios, and its performance is comparable in the 1-shot scenario. The gaps between VQT and SynQT indicate that the lightweight, task-specific QSM enables SynQT to capture beneficial information from samples in the extremely low-data regime. Simply incorporating a trainable prompt as a query, as in VQT, is not sufficient to handle such challenging few-shot tasks. Interestingly, VPT-Shallow outperforms VPT-Deep in the low-data regime, due to the fact that the larger number of prompts in VPT-Deep are not well-trained on the limited data. This suggests that merely increasing the number of prompt numbers could potentially lead to a performance drop. + +# 4.5 Versatility of SynQT + +To demonstrate the versatility of SynQT, we apply SynQT to different model architectures and combine SynQT with existing PETL methods. + +Applying SynQT to different model architectures. Here, we apply SynQT to an advanced Transformer-like architecture (Swin [31]) and a convolutional neural network (ConvNeXt [32]). Following [17], we introduce prompts for each window in Swin and incorporate learnable prompt pixels for the feature map in ConvNeXt. Specifically, due to the varying hidden sizes and number of blocks in different layers, we extract features from the third layer of Swin and the last layer of ConvNext to ensure sufficient extraction of beneficial information. We find that both of them significantly benefit from SynQT, as shown in Table 4. + +Table 5: Combining different PETL methods. AF is short for AdaptFormer. + +
MethodsNaturalSpecializedStructuredAverage
AF80.182.350.370.9
AF+VQT79.684.353.072.3
AF+SynQT82.685.457.975.3
VPT79.184.654.472.7
VPT+VQT78.983.754.672.4
VPT+SynQT81.184.957.774.1
+ +Table 6: Necessity of disentangle-ment (KEM and QSM). + +
MethodNatural SpecializedStructuredAverage
Linear63.277.131.457.2
+KEM71.880.132.461.4
+QSM77.883.655.072.1
+Head (SynQT)78.088.456.272.9
+ +On average, SynQT outperforms VPT-Deep by $2.2\%$ and $1.2\%$ for Swin and ConvNeXt, respectively. + +Combining SynQT with existing PETL methods. We further jointly utilize SynQT and existing PETL methods to adapt pre-trained models, as shown in Table 5. The results demonstrate that both VPT and AdaptFormer benefit from SynQT on VTAB-1K, with improvements of $4.4\%$ and $1.4\%$ on average, respectively. We can see that for AdaptFormer and VPT, VQT may result in a performance drop in some cases (e.g., the Natural group), whereas equipping SynQT leads to significant performance improvements across all groups. + +# 4.6 Ablation Study + +We ablate our SynQT with different settings in Tables 6 and 7 for a better understanding of SynQT. + +Disentanglement. To demonstrate the necessity of disentanglement, we perform an ablation in Table 6. We set the baseline as linear probing since KEM exploits intermediate features. To gain deeper insights into the roles of QSM and Head, we employ random prompts to query in KEM as the model without QSM, and we utilize a simple average for aggregation as the model without Head. The results demonstrate that our idea of disentangling (QSM and KEM) is the key to achieving superior performance. + +Hidden size in QSM. We vary the hidden size of the bottleneck module in QSM and report the results in Table 7a. With a larger hidden size, the performance improves due to the increasing capacity. However, when the hidden size exceeds 48, the benefit of introducing new parameters diminishes and overfitting may happen in datasets with few training samples. The results indicate that redundant parameters might not enhance the model's performance, and a hidden size of 48 is a reasonable choice. + +Input of QSM. Table 7c studies the influence of different inputs of QSM. We can see that both the trainable prompts and the output from the last block are indispensable for QSM to generate high-quality task-specific query. The trainable prompts improve the capacity of QSM, and the results from the previous block ensure that knowledge is retained during the forward process. + +Hierarchical features. In SynQT, we extract three different types of features from KSMs for classification: final output of KSM $H_{i}$ , attention features $F_{i}^{attn}$ + +Table 7: SynQT ablation experiments with ViT-B/16 with ImageNet-21K backbone. Default settings are marked in gray. + +
DimParamNat.Spe.Str.Avg.
162.14 M77.783.855.872.4
322.43 M77.783.955.372.3
482.73 M78.084.456.272.9
643.02 M77.684.456.172.7
963.61 M77.784.555.972.7
+ +(a) Hidden size in QSM. 48 as the width is a trade-off between parameters and performance. + +
SynQTNat.Spe.Str.Avg.
w/o Hi78.384.053.872.0
w/o Fiattn78.383.555.472.4
w/o Fiuffn77.884.755.472.6
Full78.084.456.272.9
+ +(b) Feature selection. Hierarchical features contains rich information that benefit to downstream tasks. + +
SynQTNat.Spe.Str.Avg.
w/o Prompt77.282.951.970.7
w/o Last Output77.584.452.371.4
Full78.084.456.272.9
+ +(c) Input of QSM. Both the trainable prompt and the output from last block are important for generating task-specific query. + +
ProjectionNat.Spe.Str.Avg.
w/o Proj.78.483.855.072.4
Shared78.084.456.272.9
Independent77.984.455.572.6
+ +(d) Projection in the classification head. Shared projection works the best for downstream tasks. + +
SynQTNat.Spe.Str.Avg.
Simply Avg.77.984.655.672.7
Fixed Weights77.984.755.572.7
Conditional78.084.456.272.9
+ +(e) Feature aggregation. Conditional feature aggregation is more effective than mean and learnable but fixed weight. + +
SynQTNat.Spe.Str.Avg.
w/o DropFeat77.784.355.272.4
w/ DropFeat78.084.456.272.9
+ +(f) DropFeat in the classification head. The DropFeat module in the conditional head makes model more robust. + +and FFN features $F_{i}^{ffn}$ , where $i \in \{1,2,\dots,l\}$ and $l$ is the number of layers. All these hierarchical features contain rich information and contribute to the performance boost of SynQT, as demonstrated in Table 7b. In particular, $H_{i}$ plays a more important role than other types of features. + +Projection in the classification head. We vary the projection architecture in the classification head and investigate its performance in Table 7d. A projection of the input is necessary for the classification head, as hierarchical features across various layers may be misaligned in the representation space. However, to our surprise, a shared bottleneck structure is enough to handle these hierarchical features. The reason may be that independent projection for distinct features leads to overfitting, subsequently causing a slight drop in performance. + +Conditional feature aggregation. In the classification head, we aggregate features using the weights conditioned on $H_{l}$ . We further conduct experiments over variants of SynQT with different aggregation methods including a simple average and weighted-sum with learnable but fixed weights. The results in Table 7e show that conditional aggregation improves the performance of SynQT especially for the Structured group. + +DropFeat in the classification head. We study the effect of DropFeat, which is proposed to prevent overfitting by randomly dropping features in the classi- + +fication head. As can be observed from Table 7f, SynQT performs worse without DropFeat, indicating that DropFeat enhances the model's generalization + +in handling rich hierarchical feature input. DropFeat ensures that the model does not overly rely on a specific feature during training and encourages the model to focus on a general way to select beneficial information from these features for downstream tasks. + +# 4.7 t-SNE Visualization + +We present t-SNE visualization [35] of the CLS Token in pre-trained ViT, VQT, and our SynQT on EuroSAT, SVHN, and Clevr-Dist in Figure 5. The visualization shows that our SynQT achieves better feature clustering results compared to the + +original CLS Token and VQT. We provide more visualization in Appendix. + +![](images/a43dc0aa692ea145d2920a1b59c03454706cd6cde26a532bb8f6a9c86182c108.jpg) +Fig. 5: t-SNE visualization on SVHN, EuroSAT, and Clevr-Dist, one from each category of VTAB-1K. The features of our SynQT are more separable. + +# 5 Conclusion, Limitation, and Future Work + +In this work, we present an insightful analysis of the high memory cost of existing PETL methods from a unique perspective, i.e., they entangle task-specific learning and pre-trained knowledge utilization together. Inspired by the idea of disentanglement, we propose a new paradigm, namely SynQT, consisting of two key structures: Query Synthesis Module (QSM) and Knowledge Extraction Module (KEM). QSM synthesizes task-specific query and then KEM extracts pre-trained knowledge in a query-only manner. Finally, a classification head is proposed to handle rich hierarchical features from KEMs and make predictions. Extensive experiments on VTAB-1K and FGVC demonstrate the effectiveness of SynQT under memory constraints and in an extremely low-data regime. + +Despite the promising performance of our SynQT, it also has some limitations. One limitation is that SynQT requires slightly more trainable parameters than previous PETL methods due to a few projection layers and FFN operations, which is analyzed in Appendix. However, we would like to emphasize that as shown in their original papers, for VPT-Deep [17] and VQT [45], straightforwardly increasing the number of tokens to use more trainable parameters can not consistently result in better performance. + +We believe that our work will inspire the community to explore memory-efficient PETL along with our disentangled perspective. Besides, to overcome the aforementioned limitation, one future direction is to develop PETL methods that achieve high accuracy and meanwhile have extreme memory efficiency and parameter efficiency. + +# Acknowledgments + +This work is supported in part by the National Natural Science Foundation of China under Grant 62171248, Shenzhen Science and Technology Program (JCYJ20220818101012025), the PCNL KEY project (PCL2023AS6-1), and the National Research Foundation, Singapore, under its AI Singapore Programme (AISG Award No: AISG2-RP-2021-023). + +# References + +1. Argyriou, A., Evgeniou, T., Pontil, M.: Multi-task feature learning. Advances in neural information processing systems 19 (2006) +2. Ba, J.L., Kiros, J.R., Hinton, G.E.: Layer normalization. arXiv preprint arXiv:1607.06450 (2016) +3. Bai, J., Gao, K., Min, S., Xia, S.T., Li, Z., Liu, W.: Badclip: Trigger-aware prompt learning for backdoor attacks on clip. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 24239-24250 (2024) +4. Bai, J., Yuan, L., Xia, S.T., Yan, S., Li, Z., Liu, W.: Improving vision transformers by revisiting high-frequency components. In: European Conference on Computer Vision. pp. 1-18. Springer (2022) +5. Bossard, L., Guillaumeun, M., Van Gool, L.: Food-101-mining discriminative components with random forests. In: Computer Vision-ECCV 2014: 13th European Conference, Zurich, Switzerland, September 6-12, 2014, Proceedings, Part VI 13. pp. 446-461. Springer (2014) +6. Brown, T., Mann, B., Ryder, N., Subbiah, M., Kaplan, J.D., Dhariwal, P., Neelakantan, A., Shyam, P., Sastry, G., Askell, A., et al.: Language models are few-shot learners. NeurIPS (2020) +7. Caron, M., Touvron, H., Misra, I., Jégou, H., Mairal, J., Bojanowski, P., Joulin, A.: Emerging properties in self-supervised vision transformers. In: Proceedings of the IEEE/CVF international conference on computer vision. pp. 9650-9660 (2021) +8. Chen, S., Ge, C., Tong, Z., Wang, J., Song, Y., Wang, J., Luo, P.: Adaptformer: Adapting vision transformers for scalable visual recognition. Advances in Neural Information Processing Systems 35, 16664-16678 (2022) +9. Deng, J., Dong, W., Socher, R., Li, L.J., Li, K., Fei-Fei, L.: Imagenet: A large-scale hierarchical image database. In: 2009 IEEE conference on computer vision and pattern recognition. pp. 248-255. IEEE (2009) +10. Dosovitskiy, A., Beyer, L., Kolesnikov, A., Weissenborn, D., Zhai, X., Unterthiner, T., Dehghani, M., Minderer, M., Heigold, G., Gelly, S., et al.: An image is worth 16x16 words: Transformers for image recognition at scale. arXiv preprint arXiv:2010.11929 (2020) +11. Evci, U., Dumoulin, V., Larochelle, H., Mozer, M.C.: Head2toe: Utilizing intermediate representations for better transfer learning. In: International Conference on Machine Learning. pp. 6009-6033. PMLR (2022) +12. Guo, H., Dai, T., Bai, Y., Chen, B., Xia, S.T., Zhu, Z.: Adaptir: Parameter efficient multi-task adaptation for pre-trained image restoration models. arXiv preprint arXiv:2312.08881 (2023) +13. Guo, H., Li, J., Dai, T., Ouyang, Z., Ren, X., Xia, S.T.: Mambair: A simple baseline for image restoration with state-space model. arXiv preprint arXiv:2402.15648 (2024) + +14. He, K., Chen, X., Xie, S., Li, Y., Dollár, P., Girshick, R.: Masked autoencoders are scalable vision learners. In: Proceedings of the IEEE/CVF conference on computer vision and pattern recognition. pp. 16000-16009 (2022) +15. Houlsby, N., Giurgiu, A., Jastrzebski, S., Morrone, B., De Laroussilhe, Q., Gesmundo, A., Attariyan, M., Gelly, S.: Parameter-efficient transfer learning for nlp. In: International Conference on Machine Learning. pp. 2790-2799. PMLR (2019) +16. Hu, E.J., Wallis, P., Allen-Zhu, Z., Li, Y., Wang, S., Wang, L., Chen, W., et al.: Lora: Low-rank adaptation of large language models. In: International Conference on Learning Representations (2021) +17. Jia, M., Tang, L., Chen, B.C., Cardie, C., Belongie, S., Hariharan, B., Lim, S.N.: Visual prompt tuning. In: European Conference on Computer Vision. pp. 709-727. Springer (2022) +18. Jie, S., Deng, Z.H.: Fact: Factor-tuning for lightweight adaptation on vision transformer. In: Proceedings of the AAAI Conference on Artificial Intelligence. vol. 37, pp. 1060-1068 (2023) +19. Jie, S., Wang, H., Deng, Z.H.: Revisiting the parameter efficiency of adapters from the perspective of precision redundancy. In: Proceedings of the IEEE/CVF International Conference on Computer Vision. pp. 17217-17226 (2023) +20. Kirillov, A., Mintun, E., Ravi, N., Mao, H., Rolland, C., Gustafson, L., Xiao, T., Whitehead, S., Berg, A.C., Lo, W.Y., et al.: Segment anything. arXiv preprint arXiv:2304.02643 (2023) +21. Kornblith, S., Shlens, J., Le, Q.V.: Do better imagenet models transfer better? In: Proceedings of the IEEE/CVF conference on computer vision and pattern recognition. pp. 2661-2671 (2019) +22. Krause, J., Stark, M., Deng, J., Fei-Fei, L.: 3d object representations for fine-grained categorization. In: Proceedings of the IEEE international conference on computer vision workshops. pp. 554-561 (2013) +23. Larsson, G., Maire, M., Shakhnarovich, G.: Fractalnet: Ultra-deep neural networks without residuals. In: ICLR (2017) +24. Lester, B., Al-Rfou, R., Constant, N.: The power of scale for parameter-efficient prompt tuning. arXiv preprint arXiv:2104.08691 (2021) +25. Li, X.L., Liang, P.: Prefix-tuning: Optimizing continuous prompts for generation. arXiv preprint arXiv:2101.00190 (2021) +26. Li, X., Lian, D., Lu, Z., Bai, J., Chen, Z., Wang, X.: Graphadapter: Tuning vision-language models with dual knowledge graph. Advances in Neural Information Processing Systems 36 (2024) +27. Lian, D., Zhou, D., Feng, J., Wang, X.: Scaling & shifting your features: A new baseline for efficient model tuning. Advances in Neural Information Processing Systems 35, 109-123 (2022) +28. Liu, H., Tam, D., Muqeeth, M., Mohta, J., Huang, T., Bansal, M., Raffel, C.A.: Few-shot parameter-efficient fine-tuning is better and cheaper than in-context learning. NeurIPS (2022) +29. Liu, P., Yuan, W., Fu, J., Jiang, Z., Hayashi, H., Neubig, G.: Pre-train, prompt, and predict: A systematic survey of prompting methods in natural language processing. ACM Computing Surveys 55(9), 1-35 (2023) +30. Liu, X., Ji, K., Fu, Y., Tam, W.L., Du, Z., Yang, Z., Tang, J.: P-tuning v2: Prompt tuning can be comparable to fine-tuning universally across scales and tasks. arXiv preprint arXiv:2110.07602 (2021) +31. Liu, Z., Lin, Y., Cao, Y., Hu, H., Wei, Y., Zhang, Z., Lin, S., Guo, B.: Swin transformer: Hierarchical vision transformer using shifted windows. In: Proceedings + +of the IEEE/CVF international conference on computer vision. pp. 10012-10022 (2021) +32. Liu, Z., Mao, H., Wu, C.Y., Feichtenhofer, C., Darrell, T., Xie, S.: A convnet for the 2020s. In: Proceedings of the IEEE/CVF conference on computer vision and pattern recognition. pp. 11976-11986 (2022) +33. Loshchilov, I., Hutter, F.: Decoupled weight decay regularization. arXiv preprint arXiv:1711.05101 (2017) +34. Lu, Z., Bai, J., Li, X., Xiao, Z., Wang, X.: Beyond sole strength: Customized ensembles for generalized vision-language models. In: Forty-first International Conference on Machine Learning (2024) +35. Van der Maaten, L., Hinton, G.: Visualizing data using t-sne. Journal of machine learning research 9(11) (2008) +36. Maji, S., Rahtu, E., Kannala, J., Blaschko, M., Vedaldi, A.: Fine-grained visual classification of aircraft. arXiv preprint arXiv:1306.5151 (2013) +37. Nilsback, M.E., Zisserman, A.: A visual vocabulary for flower classification. In: 2006 IEEE computer society conference on computer vision and pattern recognition (CVPR'06). vol. 2, pp. 1447-1454. IEEE (2006) +38. Parkhi, O.M., Vedaldi, A., Zisserman, A., Jawahar, C.: Cats and dogs. In: 2012 IEEE conference on computer vision and pattern recognition. pp. 3498-3505. IEEE (2012) +39. Pfeiffer, J., Kamath, A., Rückle, A., Cho, K., Gurevych, I.: Adapterfusion: Nondestructive task composition for transfer learning. arXiv preprint arXiv:2005.00247 (2020) +40. Rebuffi, S.A., Bilen, H., Vedaldi, A.: Learning multiple visual domains with residual adapters. Advances in neural information processing systems 30 (2017) +41. Shao, L., Zhu, F., Li, X.: Transfer learning for visual categorization: A survey. IEEE transactions on neural networks and learning systems 26(5), 1019-1034 (2014) +42. Srivastava, N., Hinton, G., Krizhevsky, A., Sutskever, I., Salakhutdinov, R.: Dropout: a simple way to prevent neural networks from overfitting. The journal of machine learning research 15(1), 1929-1958 (2014) +43. Sung, Y.L., Cho, J., Bansal, M.: Lst: Ladder side-tuning for parameter and memory efficient transfer learning. Advances in Neural Information Processing Systems 35, 12991-13005 (2022) +44. Sung, Y.L., Cho, J., Bansal, M.: Vl-adapter: Parameter-efficient transfer learning for vision-and-language tasks. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 5227-5237 (2022) +45. Tu, C.H., Mai, Z., Chao, W.L.: Visual query tuning: Towards effective usage of intermediate representations for parameter and memory efficient transfer learning. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 7725-7735 (2023) +46. Vaswani, A., Shazeer, N., Parmar, N., Uszkoreit, J., Jones, L., Gomez, A.N., Kaiser, L., Polosukhin, I.: Attention is all you need. Advances in neural information processing systems 30 (2017) +47. Wan, L., Zeiler, M., Zhang, S., Le Cun, Y., Fergus, R.: Regularization of neural networks using dropconnect. In: International conference on machine learning. pp. 1058-1066. PMLR (2013) +48. Wang, M., Deng, W.: Deep visual domain adaptation: A survey. Neurocomputing 312, 135-153 (2018) +49. Wang, W., Bao, H., Dong, L., Bjorck, J., Peng, Z., Liu, Q., Aggarwal, K., Mohammed, O.K., Singhal, S., Som, S., et al.: Image as a foreign language: Beit pre + +training for all vision and vision-language tasks. arXiv preprint arXiv:2208.10442 (2022) +50. Yang, S., Bai, J., Gao, K., Yang, Y., Li, Y., Xia, S.T.: Not all prompts are secure: A switchable backdoor attack against pre-trained vision transformers. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 24431-24441 (2024) +51. Yang, X., Ye, J., Wang, X.: Factorizing knowledge in neural networks. In: European Conference on Computer Vision. pp. 73-91. Springer (2022) +52. Yang, X., Zhou, D., Liu, S., Ye, J., Wang, X.: Deep model reassembly. Advances in neural information processing systems 35, 25739-25753 (2022) +53. Yuan, M., Lin, Y.: Model selection and estimation in regression with grouped variables. Journal of the Royal Statistical Society Series B: Statistical Methodology 68(1), 49-67 (2006) +54. Zaken, E.B., Ravfogel, S., Goldberg, Y.: Bitfit: Simple parameter-efficient fine-tuning for transformer-based masked language-models. arXiv preprint arXiv:2106.10199 (2021) +55. Zang, Y., Li, W., Zhou, K., Huang, C., Loy, C.C.: Unified vision and language prompt learning. arXiv preprint arXiv:2210.07225 (2022) +56. Zha, Y., Wang, J., Dai, T., Chen, B., Wang, Z., Xia, S.T.: Instance-aware dynamic prompt tuning for pre-trained point cloud models. In: Proceedings of the IEEE/CVF International Conference on Computer Vision. pp. 14161-14170 (2023) +57. Zhai, X., Kolesnikov, A., Houlsby, N., Beyer, L.: Scaling vision transformers. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 12104-12113 (2022) +58. Zhai, X., Puigcerver, J., Kolesnikov, A., Ruyssen, P., Riquelme, C., Lucic, M., Djolonga, J., Pinto, A.S., Neumann, M., Dosovitskiy, A., et al.: The visual task adaptation benchmark (2019) +59. Zhang, Y., Zhou, K., Liu, Z.: Neural prompt search. arXiv (2022) +60. Zhou, K., Yang, J., Loy, C.C., Liu, Z.: Conditional prompt learning for vision-language models. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 16816-16825 (2022) +61. Zhou, K., Yang, J., Loy, C.C., Liu, Z.: Learning to prompt for vision-language models. International Journal of Computer Vision 130(9), 2337-2348 (2022) +62. Zhu, L., Liao, B., Zhang, Q., Wang, X., Liu, W., Wang, X.: Vision mamba: Efficient visual representation learning with bidirectional state space model. arXiv preprint arXiv:2401.09417 (2024) \ No newline at end of file diff --git a/2024/Parameter-Efficient and Memory-Efficient Tuning for Vision Transformer_ A Disentangled Approach/images.zip b/2024/Parameter-Efficient and Memory-Efficient Tuning for Vision Transformer_ A Disentangled Approach/images.zip new file mode 100644 index 0000000000000000000000000000000000000000..fb7133fc5e5bb24bccb564bb3264e2f4e8e3f58d --- /dev/null +++ b/2024/Parameter-Efficient and Memory-Efficient Tuning for Vision Transformer_ A Disentangled Approach/images.zip @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:415fa41490138a5e3d0dfa21b34b631149dc00145fe1d385e2174a975031582c +size 486258 diff --git a/2024/Parameter-Efficient and Memory-Efficient Tuning for Vision Transformer_ A Disentangled Approach/layout.json b/2024/Parameter-Efficient and Memory-Efficient Tuning for Vision Transformer_ A Disentangled Approach/layout.json new file mode 100644 index 0000000000000000000000000000000000000000..4021535f9bb338cb294a8707ff67fd787f402c71 --- /dev/null +++ b/2024/Parameter-Efficient and Memory-Efficient Tuning for Vision Transformer_ A Disentangled Approach/layout.json @@ -0,0 +1,9841 @@ +{ + "pdf_info": [ + { + "para_blocks": [ + { + "bbox": [ + 132, + 111, + 481, + 148 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 111, + 481, + 148 + ], + "spans": [ + { + "bbox": [ + 132, + 111, + 481, + 148 + ], + "type": "text", + "content": "Parameter-Efficient and Memory-Efficient Tuning for Vision Transformer: A Disentangled Approach" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 133, + 167, + 480, + 194 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 167, + 480, + 194 + ], + "spans": [ + { + "bbox": [ + 133, + 167, + 480, + 194 + ], + "type": "text", + "content": "Taolin Zhang" + }, + { + "bbox": [ + 133, + 167, + 480, + 194 + ], + "type": "inline_equation", + "content": "^{1*}" + }, + { + "bbox": [ + 133, + 167, + 480, + 194 + ], + "type": "text", + "content": ", Jiawang Bai" + }, + { + "bbox": [ + 133, + 167, + 480, + 194 + ], + "type": "inline_equation", + "content": "^{2,3*}" + }, + { + "bbox": [ + 133, + 167, + 480, + 194 + ], + "type": "text", + "content": ", Zhihe Lu" + }, + { + "bbox": [ + 133, + 167, + 480, + 194 + ], + "type": "inline_equation", + "content": "^{3}" + }, + { + "bbox": [ + 133, + 167, + 480, + 194 + ], + "type": "text", + "content": ", Dongze Lian" + }, + { + "bbox": [ + 133, + 167, + 480, + 194 + ], + "type": "inline_equation", + "content": "^{3}" + }, + { + "bbox": [ + 133, + 167, + 480, + 194 + ], + "type": "text", + "content": ", Genping Wang" + }, + { + "bbox": [ + 133, + 167, + 480, + 194 + ], + "type": "inline_equation", + "content": "^{4,\\text{凹}}" + }, + { + "bbox": [ + 133, + 167, + 480, + 194 + ], + "type": "text", + "content": ", Xinchao Wang" + }, + { + "bbox": [ + 133, + 167, + 480, + 194 + ], + "type": "inline_equation", + "content": "^{3,\\text{凹}}" + }, + { + "bbox": [ + 133, + 167, + 480, + 194 + ], + "type": "text", + "content": ", and Shu-Tao Xia" + }, + { + "bbox": [ + 133, + 167, + 480, + 194 + ], + "type": "inline_equation", + "content": "^{1,5}" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 157, + 201, + 461, + 236 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 157, + 201, + 461, + 236 + ], + "spans": [ + { + "bbox": [ + 157, + 201, + 461, + 236 + ], + "type": "inline_equation", + "content": "^{1}" + }, + { + "bbox": [ + 157, + 201, + 461, + 236 + ], + "type": "text", + "content": "Tsinghua Shenzhen International Graduate School, Tsinghua University " + }, + { + "bbox": [ + 157, + 201, + 461, + 236 + ], + "type": "inline_equation", + "content": "^{2}" + }, + { + "bbox": [ + 157, + 201, + 461, + 236 + ], + "type": "text", + "content": "Tencent " + }, + { + "bbox": [ + 157, + 201, + 461, + 236 + ], + "type": "inline_equation", + "content": "^{3}" + }, + { + "bbox": [ + 157, + 201, + 461, + 236 + ], + "type": "text", + "content": "National University of Singapore " + }, + { + "bbox": [ + 157, + 201, + 461, + 236 + ], + "type": "inline_equation", + "content": "^{4}" + }, + { + "bbox": [ + 157, + 201, + 461, + 236 + ], + "type": "text", + "content": "Shenzhen Qiji Technology Co., Ltd." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 153, + 236, + 461, + 280 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 153, + 236, + 461, + 280 + ], + "spans": [ + { + "bbox": [ + 153, + 236, + 461, + 280 + ], + "type": "inline_equation", + "content": "^{5}" + }, + { + "bbox": [ + 153, + 236, + 461, + 280 + ], + "type": "text", + "content": "Research Center of Artificial Intelligence, Peng Cheng Laboratory ztl23@mails.tsinghua.edu.cn; jiawangbai@tencent.com; zhihelu.academic@gmail.com; dzlianx@gmail.com; genpingwang@163.com; xinchao@nus.edu.sg; xiast@sz.tsinghua.edu.cn" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 159, + 310, + 455, + 540 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 159, + 310, + 455, + 540 + ], + "spans": [ + { + "bbox": [ + 159, + 310, + 455, + 540 + ], + "type": "text", + "content": "Abstract. Recent works on parameter-efficient transfer learning (PETL) show the potential to adapt a pre-trained Vision Transformer to downstream recognition tasks with only a few learnable parameters. However, since they usually insert new structures into the pre-trained model, entire intermediate features of that model are changed and thus need to be stored to be involved in back-propagation, resulting in memory-heavy training. We solve this problem from a novel disentangled perspective, i.e., dividing PETL into two aspects: task-specific learning and pretrained knowledge utilization. Specifically, we synthesize the task-specific query with a learnable and lightweight module, which is independent of the pre-trained model. The synthesized query equipped with task-specific knowledge serves to extract the useful features for downstream tasks from the intermediate representations of the pre-trained model in a query-only manner. Built upon these features, a customized classification head is proposed to make the prediction for the input sample. Given that our method employs an extremely lightweight architecture and avoids the use of heavy intermediate features for running gradient descent, it demonstrates limited memory usage in training. Extensive experiments manifest that our method achieves state-of-the-art performance under memory constraints, showcasing its applicability in real-world situations. The code is available at: https://synqt.github.io/." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 132, + 562, + 230, + 574 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 562, + 230, + 574 + ], + "spans": [ + { + "bbox": [ + 132, + 562, + 230, + 574 + ], + "type": "text", + "content": "1 Introduction" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 130, + 587, + 482, + 635 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 587, + 482, + 635 + ], + "spans": [ + { + "bbox": [ + 130, + 587, + 482, + 635 + ], + "type": "text", + "content": "Motivated by the success of data-driven solutions, there has been an explosive growth in the size of vision models [4,10,13,20,31,49,62] and the scale of accessible datasets like ImageNet [9]. To unleash the power of large models, it becomes a tendency to first pre-train on a general and large-scale dataset, and then adapt" + } + ] + } + ], + "index": 6 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 141, + 643, + 251, + 666 + ], + "type": "page_footnote", + "angle": 0, + "lines": [ + { + "bbox": [ + 141, + 643, + 251, + 666 + ], + "spans": [ + { + "bbox": [ + 141, + 643, + 251, + 666 + ], + "type": "text", + "content": "* Equal contribution. \n" + }, + { + "bbox": [ + 141, + 643, + 251, + 666 + ], + "type": "inline_equation", + "content": "\\boxtimes" + }, + { + "bbox": [ + 141, + 643, + 251, + 666 + ], + "type": "text", + "content": " Corresponding authors." + } + ] + } + ], + "index": 7 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 0 + }, + { + "para_blocks": [ + { + "bbox": [ + 130, + 116, + 479, + 164 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 116, + 479, + 164 + ], + "spans": [ + { + "bbox": [ + 130, + 116, + 479, + 164 + ], + "type": "text", + "content": "them to downstream tasks [21, 41, 48, 52]. The most obvious adaption strategy is to fully fine-tune all parameters of the pre-trained model. However, it necessitates that the device saves a dedicated set of model parameters for each single task, leading to substantial storage overhead [57]." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 130, + 166, + 482, + 274 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 166, + 482, + 274 + ], + "spans": [ + { + "bbox": [ + 130, + 166, + 482, + 274 + ], + "type": "text", + "content": "To improve the storage efficiency, parameter-efficient transfer learning (PETL) methods [8,16-19,59] have been explored by tuning only a small set of the parameters in adapting the large pre-trained models to downstream tasks. Typically, these methods insert small structures into the frozen pre-trained model and specify the newly added parameters as the only learnable ones. For instance, Visual Prompt Tuning (VPT) [17] prepends learnable tokens to the patch embeddings; AdaptFormer [8] inserts a bottleneck-structured fully connected layers parallel to the feed-forward network (FFN) of the original ViT. These studies demonstrate that PETL is a promising way to save storage resources." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 130, + 276, + 482, + 360 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 276, + 482, + 360 + ], + "spans": [ + { + "bbox": [ + 130, + 276, + 482, + 360 + ], + "type": "text", + "content": "However, even though updating a small set of parameters, existing PETL methods still require a very high consumption of memory in training. The direct reason is that they need a full back-propagation throughout the entire backbone. Given the fact that the model size is exponentially increasing, it becomes infeasible to apply existing PETL methods to the memory-constraint situations. Therefore, it is desirable to investigate memory-efficient PETL while maintaining performance on downstream tasks." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 130, + 362, + 482, + 445 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 362, + 482, + 445 + ], + "spans": [ + { + "bbox": [ + 130, + 362, + 482, + 445 + ], + "type": "text", + "content": "In this work, we attribute the high memory cost of existing PETL methods in training to their intrinsic regime, i.e., entangling task-specific learning and pre-trained knowledge utilization together. It is usually achieved by integrating trainable structures into the pre-trained ViT and changing all intermediate features. Consequently, a great number of intermediate features need to be stored in training to run gradient descent. The aforementioned analysis motivates us to design a memory-efficient PETL method from a disentangled perspective." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 130, + 448, + 483, + 628 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 448, + 483, + 628 + ], + "spans": [ + { + "bbox": [ + 130, + 448, + 483, + 628 + ], + "type": "text", + "content": "Accordingly, we propose Synthesized Query Tuning (SynQT), a memory-efficient PETL method for adapting pre-trained ViTs to downstream recognition tasks. Overall, as shown in Figure 1, SynQT takes the intermediate representations of the original pre-trained model as input and combines them with task-specific knowledge using stacked blocks. We propose the Query Synthesis Module (QSM) and Knowledge Extraction Module (KEM) to accomplish the disentanglement. To capture task-specific knowledge, SynQT employs a lightweight QSM, which is independent of the pre-trained model. In QSM, an attention layer and an FFN are trained on the target downstream task to synthesize task-specific query. Then, to utilize the pre-trained knowledge, the synthesized query interacts with the intermediate representations of the original pre-trained model in a KEM. It outputs useful features for downstream recognition, where all parameters are inherited from the pre-trained model and frozen. Relying on rich hierarchical features from KEMs, we propose a classification head, which adaptively aggregates these features conditioned on the input sample for prediction." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 130, + 630, + 482, + 665 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 630, + 482, + 665 + ], + "spans": [ + { + "bbox": [ + 130, + 630, + 482, + 665 + ], + "type": "text", + "content": "The success of SynQT in saving memory in training is derived from two technical aspects. Firstly, we keep intermediate representations of the original pretrained model intact and directly exploit them in a query-only manner, avoiding" + } + ] + } + ], + "index": 7 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "spans": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "type": "text", + "content": "2" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 166, + 91, + 299, + 102 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 166, + 91, + 299, + 102 + ], + "spans": [ + { + "bbox": [ + 166, + 91, + 299, + 102 + ], + "type": "text", + "content": "Taolin Zhang, Jiawang Bai et al." + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 1 + }, + { + "para_blocks": [ + { + "bbox": [ + 130, + 116, + 479, + 163 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 116, + 479, + 163 + ], + "spans": [ + { + "bbox": [ + 130, + 116, + 479, + 163 + ], + "type": "text", + "content": "the use of heavy intermediate features for running gradient descent. Besides, our design including QSM and KEM is extremely lightweight. For instance, we adopt bottleneck structures for Q-K-V projection [46] and a small number of tokens (lower than 4 in most cases) in QSM." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 130, + 164, + 480, + 247 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 164, + 480, + 247 + ], + "spans": [ + { + "bbox": [ + 130, + 164, + 480, + 247 + ], + "type": "text", + "content": "We validate SynQT on VTAB-1K [58] benchmark, including 19 tasks from diverse domains. We also evaluate our method in few-shot learning using 5 fine-grained visual classification datasets. Extensive experiments show that SynQT can significantly reduce the memory footprint in training with promising accuracy. Compared to VPT-Deep, a powerful version of VPT that inserts prompts in each layer of ViT, SynQT saves " + }, + { + "bbox": [ + 130, + 164, + 480, + 247 + ], + "type": "inline_equation", + "content": "56\\%" + }, + { + "bbox": [ + 130, + 164, + 480, + 247 + ], + "type": "text", + "content": " memory consumption and achieves higher Top-1 accuracy on average. The contributions of our work are as follows:" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 136, + 258, + 480, + 354 + ], + "type": "list", + "angle": 0, + "index": 7, + "blocks": [ + { + "bbox": [ + 136, + 258, + 479, + 293 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 136, + 258, + 479, + 293 + ], + "spans": [ + { + "bbox": [ + 136, + 258, + 479, + 293 + ], + "type": "text", + "content": "1. We provide an insightful analysis of the high memory cost of existing PETL methods, i.e., they entangle task-specific learning and pre-trained knowledge utilization together." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 136, + 294, + 480, + 329 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 136, + 294, + 480, + 329 + ], + "spans": [ + { + "bbox": [ + 136, + 294, + 480, + 329 + ], + "type": "text", + "content": "2. Inspired by the disentangled idea, we propose SynQT, which synthesizes task-specific query and then utilizes the pre-trained knowledge in a query-only manner." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 136, + 331, + 479, + 354 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 136, + 331, + 479, + 354 + ], + "spans": [ + { + "bbox": [ + 136, + 331, + 479, + 354 + ], + "type": "text", + "content": "3. Extensive experiments show that SynQT is highly memory-efficient and effective, notably outperforming existing baselines under memory constraints." + } + ] + } + ], + "index": 6 + } + ], + "sub_type": "text" + }, + { + "bbox": [ + 132, + 374, + 237, + 388 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 374, + 237, + 388 + ], + "spans": [ + { + "bbox": [ + 132, + 374, + 237, + 388 + ], + "type": "text", + "content": "2 Related Work" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 130, + 402, + 481, + 665 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 402, + 481, + 665 + ], + "spans": [ + { + "bbox": [ + 130, + 402, + 481, + 665 + ], + "type": "text", + "content": "Parameter-Efficient Transfer Learning (PETL). The idea of PETL initially shows great success in the field of natural language processing (NLP) [6,16,28,40]. It fine-tunes a small number of trainable parameters to transfer large pre-trained models to downstream tasks, significantly reducing storage costs by avoiding a copy of model parameters for each new task. Due to the increased size of pre-trained vision models, especially for ViT [10], PETL has been introduced to computer vision and attracted much attention [3,8,12,17,26,27,34,39,44, 45,50,55,56,60,61]. To reduce the number of fine-tuned parameters, BitFit [54] only fine-tunes the bias terms and freezes most of the network, resulting in comparable performance to full tuning. VPT [17] introduces learnable tokens (i.e., prompts [24,25,29,30]) to each layer of ViT. SSF [27] fine-tunes scale and shift factors in each layer to adapt original features to the target domain. Some other works [8,15,16] propose to insert additional trainable structures into attention blocks of the pre-trained Transformer. For instance, LoRA [16] observes that the learned over-parametrized models reside on a low-rank dimension and further injects trainable rank decomposition parameters into each layer of the Transformer; AdaptFormer [8] adds parallel lightweight module to FFN. Furthermore, for extreme parameter efficiency, a tensorization-decomposition framework [18] is exploited to store the weight increments, and the low-bit adapter [19] is proposed to reduce the precision redundancy. However, how to improve memory efficiency in training is still an under-explored problem in PETL, as discussed below." + } + ] + } + ], + "index": 9 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 149, + 91, + 447, + 102 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 149, + 91, + 447, + 102 + ], + "spans": [ + { + "bbox": [ + 149, + 91, + 447, + 102 + ], + "type": "text", + "content": "Parameter-Efficient and Memory-Efficient Tuning for Vision Transformer" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 474, + 92, + 480, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 474, + 92, + 480, + 100 + ], + "spans": [ + { + "bbox": [ + 474, + 92, + 480, + 100 + ], + "type": "text", + "content": "3" + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 2 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 143, + 114, + 473, + 238 + ], + "blocks": [ + { + "bbox": [ + 143, + 114, + 473, + 238 + ], + "lines": [ + { + "bbox": [ + 143, + 114, + 473, + 238 + ], + "spans": [ + { + "bbox": [ + 143, + 114, + 473, + 238 + ], + "type": "image", + "image_path": "6b9e702486b0bd3db86ad0a5cac53de8656c0e72bda71ddf5c90b382220bf160.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 130, + 246, + 482, + 302 + ], + "lines": [ + { + "bbox": [ + 130, + 246, + 482, + 302 + ], + "spans": [ + { + "bbox": [ + 130, + 246, + 482, + 302 + ], + "type": "text", + "content": "Fig. 1: Overall architecture of SynQT. Each block in SynQT contains two key designs, including Query Synthesis Module (QSM) and Knowledge Extractor Module (KEM). QSM is a lightweight learnable module to capture task-specific knowledge and KEM reuses frozen weights in the pre-trained model for feature extraction. The rich and hierarchical features from KEMs are fed into the classification head." + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_caption" + } + ], + "index": 2 + }, + { + "bbox": [ + 130, + 326, + 482, + 493 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 326, + 482, + 493 + ], + "spans": [ + { + "bbox": [ + 130, + 326, + 482, + 493 + ], + "type": "text", + "content": "Memory-Efficient PETL. Although existing PETL methods have shown promising performance in transfer learning while saving most trainable parameters, they still suffer from issues of high memory consumption in training. The full back-propagation throughout the entire backbone prevents traditional PETL methods from being applied in memory-constraint situations. Recently, some recent works [11,43,45] have made several attempts to address this issue, whose key idea is to leverage the power of intermediate representations of the pre-trained model and thus avoid full back-propagation. In NLP, LST [43] proposes a trainable lightweight structure in the sidewalk for memory-saving during training. To adapt vision models, Head2Toe [11] selects features from all layers of the source model to feed into a classification head. VQT [45] inserts trainable tokens in the query space to capture useful information from the pre-trained model. Due to the information redundancy in the intermediate representations, both Head2Toe and VQT further adopt group lasso [1,53] for feature selection." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 130, + 494, + 482, + 578 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 494, + 482, + 578 + ], + "spans": [ + { + "bbox": [ + 130, + 494, + 482, + 578 + ], + "type": "text", + "content": "In contrast to previous works, we provide an insightful analysis that entangled combination of task-specific learning and pre-trained knowledge utilization results in high memory consumption of traditional PETL methods. Therefore, we propose SynQT to disentangle PETL into two distinct modules, with the aim of capturing task-specific knowledge [51] and extracting useful features for downstream tasks, respectively. Such a disentangled paradigm enjoys memory efficiency in transfer learning compared with traditional PETL methods." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 132, + 596, + 202, + 608 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 596, + 202, + 608 + ], + "spans": [ + { + "bbox": [ + 132, + 596, + 202, + 608 + ], + "type": "text", + "content": "3 Method" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 132, + 621, + 257, + 632 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 621, + 257, + 632 + ], + "spans": [ + { + "bbox": [ + 132, + 621, + 257, + 632 + ], + "type": "text", + "content": "3.1 Overall Framework" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 130, + 641, + 481, + 665 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 641, + 481, + 665 + ], + "spans": [ + { + "bbox": [ + 130, + 641, + 481, + 665 + ], + "type": "text", + "content": "As can be seen from Figures 1 and 2, SynQT makes the prediction relying on the features produced by stacked blocks, which take the intermediate representations" + } + ] + } + ], + "index": 8 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 133, + 92, + 140, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 92, + 140, + 100 + ], + "spans": [ + { + "bbox": [ + 133, + 92, + 140, + 100 + ], + "type": "text", + "content": "4" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 166, + 91, + 299, + 102 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 166, + 91, + 299, + 102 + ], + "spans": [ + { + "bbox": [ + 166, + 91, + 299, + 102 + ], + "type": "text", + "content": "Taolin Zhang, Jiawang Bai et al." + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 3 + }, + { + "para_blocks": [ + { + "bbox": [ + 130, + 116, + 482, + 224 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 116, + 482, + 224 + ], + "spans": [ + { + "bbox": [ + 130, + 116, + 482, + 224 + ], + "type": "text", + "content": "of the original pre-trained model as external input. In SynQT, each block mainly includes two parts: Query Synthesis Module (QSM) and Knowledge Extraction Module (KEM). QSM is a lightweight module, which is used to capture task-specific knowledge, connecting the pre-trained model with the downstream task. KEM uses the tokens generated by QSM as queries, and the intermediate features in the pre-trained ViT as keys and values. It reuses the original model's weights to extract features that are beneficial to downstream tasks. Finally, the rich hierarchical features extracted by the KEMs are fed into the classification head, as shown in Figure 2. We will elaborate our designs in the following sections." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 131, + 242, + 285, + 254 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 131, + 242, + 285, + 254 + ], + "spans": [ + { + "bbox": [ + 131, + 242, + 285, + 254 + ], + "type": "text", + "content": "3.2 Query Synthesis Module" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 130, + 261, + 482, + 357 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 261, + 482, + 357 + ], + "spans": [ + { + "bbox": [ + 130, + 261, + 482, + 357 + ], + "type": "text", + "content": "PETL integrates the task-specific knowledge into the adapted model through training on the downstream dataset. Typically, existing methods [8, 15-17] insert additional trainable structures into the pre-trained model. However, this is the root of high memory consumption, since it changes all original features, and all intermediate features are modified and stored to be involved in calculating gradients. From the disentangled perspective, we solve this problem with a task-specific learning module, named Query Synthesis Module (QSM), which is trainable, lightweight, and structurally independent from the pre-trained model." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 130, + 357, + 482, + 476 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 357, + 482, + 476 + ], + "spans": [ + { + "bbox": [ + 130, + 357, + 482, + 476 + ], + "type": "text", + "content": "The purpose of QSM is to synthesize query tokens, interacting with intermediate representations of the original pre-trained model. Besides outputs from the previous block, we introduce a trainable prompt into the input space of each QSM. These per-block additional prompts increase the capacity of QSM to capture task-specific knowledge. Inspired by the Transformer, QSM includes an attention layer and an FFN. Specifically, the input tokens interact with each other through a self-attention layer and are further fed into the feed-forward network for producing the query tokens. Besides, to keep the parameter efficiency, another design principle is to extensively use bottleneck-structured fully connected layers to reduce the number of parameters." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 130, + 476, + 482, + 548 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 476, + 482, + 548 + ], + "spans": [ + { + "bbox": [ + 130, + 476, + 482, + 548 + ], + "type": "text", + "content": "Let " + }, + { + "bbox": [ + 130, + 476, + 482, + 548 + ], + "type": "inline_equation", + "content": "n" + }, + { + "bbox": [ + 130, + 476, + 482, + 548 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 130, + 476, + 482, + 548 + ], + "type": "inline_equation", + "content": "d" + }, + { + "bbox": [ + 130, + 476, + 482, + 548 + ], + "type": "text", + "content": " denote the number of tokens and the hidden size in QSM, respectively. Note that " + }, + { + "bbox": [ + 130, + 476, + 482, + 548 + ], + "type": "inline_equation", + "content": "n" + }, + { + "bbox": [ + 130, + 476, + 482, + 548 + ], + "type": "text", + "content": " is very small, e.g., " + }, + { + "bbox": [ + 130, + 476, + 482, + 548 + ], + "type": "inline_equation", + "content": "n = 4" + }, + { + "bbox": [ + 130, + 476, + 482, + 548 + ], + "type": "text", + "content": ", much lower than the number of tokens in the original model and the value of " + }, + { + "bbox": [ + 130, + 476, + 482, + 548 + ], + "type": "inline_equation", + "content": "d" + }, + { + "bbox": [ + 130, + 476, + 482, + 548 + ], + "type": "text", + "content": " corresponds to that of the pre-trained model. Before the self-attention layer, the output of the last block " + }, + { + "bbox": [ + 130, + 476, + 482, + 548 + ], + "type": "inline_equation", + "content": "\\hat{H}_{i - 1} \\in \\mathbb{R}^{n \\times d}" + }, + { + "bbox": [ + 130, + 476, + 482, + 548 + ], + "type": "text", + "content": " are added with trainable prompt " + }, + { + "bbox": [ + 130, + 476, + 482, + 548 + ], + "type": "inline_equation", + "content": "P_{i}" + }, + { + "bbox": [ + 130, + 476, + 482, + 548 + ], + "type": "text", + "content": " and then fed into bottleneck projection layers. It is formally formulated as follows," + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 236, + 559, + 480, + 571 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 236, + 559, + 480, + 571 + ], + "spans": [ + { + "bbox": [ + 236, + 559, + 480, + 571 + ], + "type": "interline_equation", + "content": "Z _ {i} ^ {\\prime} = \\left(\\left(H _ {i} + P _ {i}\\right) \\cdot \\mathbf {W} _ {\\text {d o w n}}\\right) \\cdot \\mathbf {W} _ {\\text {u p}}, \\tag {1}", + "image_path": "44e5ffb4b8828e68a301412b350df05dee6e3bebc6a95753adc2eebc700b1964.jpg" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 130, + 581, + 481, + 607 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 581, + 481, + 607 + ], + "spans": [ + { + "bbox": [ + 130, + 581, + 481, + 607 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 130, + 581, + 481, + 607 + ], + "type": "inline_equation", + "content": "\\mathbf{W}_{down} \\in \\mathbb{R}^{d \\times \\hat{d}}" + }, + { + "bbox": [ + 130, + 581, + 481, + 607 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 130, + 581, + 481, + 607 + ], + "type": "inline_equation", + "content": "\\mathbf{W}_{up} \\in \\mathbb{R}^{\\hat{d} \\times d}" + }, + { + "bbox": [ + 130, + 581, + 481, + 607 + ], + "type": "text", + "content": " denote the down-projection and up-projection layer respectively, with " + }, + { + "bbox": [ + 130, + 581, + 481, + 607 + ], + "type": "inline_equation", + "content": "\\hat{d} \\ll d" + }, + { + "bbox": [ + 130, + 581, + 481, + 607 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 130, + 608, + 481, + 643 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 608, + 481, + 643 + ], + "spans": [ + { + "bbox": [ + 130, + 608, + 481, + 643 + ], + "type": "text", + "content": "In the attention layer, " + }, + { + "bbox": [ + 130, + 608, + 481, + 643 + ], + "type": "inline_equation", + "content": "Z_{i}^{\\prime}" + }, + { + "bbox": [ + 130, + 608, + 481, + 643 + ], + "type": "text", + "content": " is projected by three bottleneck modules to produce the query " + }, + { + "bbox": [ + 130, + 608, + 481, + 643 + ], + "type": "inline_equation", + "content": "Q_{i}^{qs}" + }, + { + "bbox": [ + 130, + 608, + 481, + 643 + ], + "type": "text", + "content": ", key " + }, + { + "bbox": [ + 130, + 608, + 481, + 643 + ], + "type": "inline_equation", + "content": "K_{i}^{qs}" + }, + { + "bbox": [ + 130, + 608, + 481, + 643 + ], + "type": "text", + "content": ", and value " + }, + { + "bbox": [ + 130, + 608, + 481, + 643 + ], + "type": "inline_equation", + "content": "V_{i}^{qs}" + }, + { + "bbox": [ + 130, + 608, + 481, + 643 + ], + "type": "text", + "content": ". Then the hidden embedding after attention layers " + }, + { + "bbox": [ + 130, + 608, + 481, + 643 + ], + "type": "inline_equation", + "content": "Z_{i}^{\\prime \\prime}" + }, + { + "bbox": [ + 130, + 608, + 481, + 643 + ], + "type": "text", + "content": " can be calculated as follows, controlled by a scale factor " + }, + { + "bbox": [ + 130, + 608, + 481, + 643 + ], + "type": "inline_equation", + "content": "s^{\\prime}" + }, + { + "bbox": [ + 130, + 608, + 481, + 643 + ], + "type": "text", + "content": ":" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 231, + 653, + 480, + 666 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 231, + 653, + 480, + 666 + ], + "spans": [ + { + "bbox": [ + 231, + 653, + 480, + 666 + ], + "type": "interline_equation", + "content": "Z _ {i} ^ {\\prime \\prime} = s ^ {\\prime} \\cdot \\operatorname {A t t n} \\left(Q _ {i} ^ {q s}, K _ {i} ^ {q s}, V _ {i} ^ {q s}\\right) + Z _ {i} ^ {\\prime}. \\tag {2}", + "image_path": "edf1cdc671df52e8c473e2cfc7707ac47fe044e791bae7026b4ed88fd0060974.jpg" + } + ] + } + ], + "index": 10 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 149, + 91, + 448, + 102 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 149, + 91, + 448, + 102 + ], + "spans": [ + { + "bbox": [ + 149, + 91, + 448, + 102 + ], + "type": "text", + "content": "Parameter-Efficient and Memory-Efficient Tuning for Vision Transformer" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 474, + 92, + 480, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 474, + 92, + 480, + 100 + ], + "spans": [ + { + "bbox": [ + 474, + 92, + 480, + 100 + ], + "type": "text", + "content": "5" + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 4 + }, + { + "para_blocks": [ + { + "bbox": [ + 130, + 115, + 479, + 140 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 115, + 479, + 140 + ], + "spans": [ + { + "bbox": [ + 130, + 115, + 479, + 140 + ], + "type": "text", + "content": "Then we can further obtain the synthesized query " + }, + { + "bbox": [ + 130, + 115, + 479, + 140 + ], + "type": "inline_equation", + "content": "\\hat{H}_i" + }, + { + "bbox": [ + 130, + 115, + 479, + 140 + ], + "type": "text", + "content": " after the FFN as the input of KEM, controlled by another scale factor " + }, + { + "bbox": [ + 130, + 115, + 479, + 140 + ], + "type": "inline_equation", + "content": "s''" + }, + { + "bbox": [ + 130, + 115, + 479, + 140 + ], + "type": "text", + "content": ":" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 241, + 148, + 481, + 161 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 241, + 148, + 481, + 161 + ], + "spans": [ + { + "bbox": [ + 241, + 148, + 481, + 161 + ], + "type": "interline_equation", + "content": "\\hat {H} _ {i} = s ^ {\\prime \\prime} \\cdot \\operatorname {F F N} \\left(\\operatorname {L N} \\left(Z _ {i} ^ {\\prime \\prime}\\right)\\right) + Z _ {i} ^ {\\prime \\prime}, \\tag {3}", + "image_path": "b152b2f10b26aed38e28911216f5b36d3c1352e2c590cfe989fa096284b21b0b.jpg" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 130, + 168, + 479, + 239 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 168, + 479, + 239 + ], + "spans": [ + { + "bbox": [ + 130, + 168, + 479, + 239 + ], + "type": "text", + "content": "where FFN also consists of bottleneck-structured fully connected layers and " + }, + { + "bbox": [ + 130, + 168, + 479, + 239 + ], + "type": "inline_equation", + "content": "\\mathrm{LN}(\\cdot)" + }, + { + "bbox": [ + 130, + 168, + 479, + 239 + ], + "type": "text", + "content": " represents the layer normalization [2]. " + }, + { + "bbox": [ + 130, + 168, + 479, + 239 + ], + "type": "inline_equation", + "content": "s^{\\prime}" + }, + { + "bbox": [ + 130, + 168, + 479, + 239 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 130, + 168, + 479, + 239 + ], + "type": "inline_equation", + "content": "s''" + }, + { + "bbox": [ + 130, + 168, + 479, + 239 + ], + "type": "text", + "content": " are hyper-parameters to modulate the effects of the attention and FFN on the synthesized query. Since QSM is trained on the downstream dataset from scratch, we believe that its output " + }, + { + "bbox": [ + 130, + 168, + 479, + 239 + ], + "type": "inline_equation", + "content": "\\hat{H}_i" + }, + { + "bbox": [ + 130, + 168, + 479, + 239 + ], + "type": "text", + "content": " captures the task-specific knowledge. We use " + }, + { + "bbox": [ + 130, + 168, + 479, + 239 + ], + "type": "inline_equation", + "content": "\\hat{H}_i" + }, + { + "bbox": [ + 130, + 168, + 479, + 239 + ], + "type": "text", + "content": " in a query-only manner as illustrated in the next section." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 132, + 256, + 315, + 268 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 256, + 315, + 268 + ], + "spans": [ + { + "bbox": [ + 132, + 256, + 315, + 268 + ], + "type": "text", + "content": "3.3 Knowledge Extraction Module" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 130, + 275, + 479, + 334 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 275, + 479, + 334 + ], + "spans": [ + { + "bbox": [ + 130, + 275, + 479, + 334 + ], + "type": "text", + "content": "With the task-specific query obtained from QSM, we propose Knowledge Extraction Module (KEM) to exploit intermediate features of the original pre-trained model. Before that, we obtain original features " + }, + { + "bbox": [ + 130, + 275, + 479, + 334 + ], + "type": "inline_equation", + "content": "\\{X_{1}, X_{2}, \\dots, X_{l}\\}" + }, + { + "bbox": [ + 130, + 275, + 479, + 334 + ], + "type": "text", + "content": " of the input sample from forward-propagation, where " + }, + { + "bbox": [ + 130, + 275, + 479, + 334 + ], + "type": "inline_equation", + "content": "l" + }, + { + "bbox": [ + 130, + 275, + 479, + 334 + ], + "type": "text", + "content": " is the number of transformer blocks of the pre-train model. Note that this process is efficient due to no gradient calculation." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 130, + 335, + 480, + 406 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 335, + 480, + 406 + ], + "spans": [ + { + "bbox": [ + 130, + 335, + 480, + 406 + ], + "type": "text", + "content": "KEM takes the synthesized query " + }, + { + "bbox": [ + 130, + 335, + 480, + 406 + ], + "type": "inline_equation", + "content": "\\hat{H}_i" + }, + { + "bbox": [ + 130, + 335, + 480, + 406 + ], + "type": "text", + "content": " and original features " + }, + { + "bbox": [ + 130, + 335, + 480, + 406 + ], + "type": "inline_equation", + "content": "\\{X_1, X_2, \\dots, X_l\\}" + }, + { + "bbox": [ + 130, + 335, + 480, + 406 + ], + "type": "text", + "content": " as inputs and outputs useful features for the downstream task. Since " + }, + { + "bbox": [ + 130, + 335, + 480, + 406 + ], + "type": "inline_equation", + "content": "\\hat{H}_i" + }, + { + "bbox": [ + 130, + 335, + 480, + 406 + ], + "type": "text", + "content": " has a small number of tokens and original features are only queried by " + }, + { + "bbox": [ + 130, + 335, + 480, + 406 + ], + "type": "inline_equation", + "content": "\\hat{H}_i" + }, + { + "bbox": [ + 130, + 335, + 480, + 406 + ], + "type": "text", + "content": ", there are no heavy features in the middle of KEM. Besides, we find that it is effective to reuse and freeze the pre-trained parameters for feature interaction in KEM. These designs ensure the parameter- and memory-efficiency of our method." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 130, + 406, + 479, + 441 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 406, + 479, + 441 + ], + "spans": [ + { + "bbox": [ + 130, + 406, + 479, + 441 + ], + "type": "text", + "content": "Specifically, with the projection weight in the pre-trained ViT blocks, we first obtain the query " + }, + { + "bbox": [ + 130, + 406, + 479, + 441 + ], + "type": "inline_equation", + "content": "Q_{i}^{ke}" + }, + { + "bbox": [ + 130, + 406, + 479, + 441 + ], + "type": "text", + "content": ", key " + }, + { + "bbox": [ + 130, + 406, + 479, + 441 + ], + "type": "inline_equation", + "content": "K_{i}^{ke}" + }, + { + "bbox": [ + 130, + 406, + 479, + 441 + ], + "type": "text", + "content": ", and value " + }, + { + "bbox": [ + 130, + 406, + 479, + 441 + ], + "type": "inline_equation", + "content": "V_{i}^{ke}" + }, + { + "bbox": [ + 130, + 406, + 479, + 441 + ], + "type": "text", + "content": ". Then follow-up attention calculation is performed by:" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 238, + 449, + 480, + 463 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 238, + 449, + 480, + 463 + ], + "spans": [ + { + "bbox": [ + 238, + 449, + 480, + 463 + ], + "type": "interline_equation", + "content": "E _ {i} = \\operatorname {A t t n} \\left(Q _ {i} ^ {k e}, K _ {i} ^ {k e}, V _ {i} ^ {k e}\\right) + \\hat {H} _ {i}. \\tag {4}", + "image_path": "0c71766eef3301d5a1f5c74915b92430a303db759065d242ed76ee5a5f296bac.jpg" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 130, + 470, + 479, + 493 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 470, + 479, + 493 + ], + "spans": [ + { + "bbox": [ + 130, + 470, + 479, + 493 + ], + "type": "text", + "content": "We also reuse the weights of FFN to perform the same space transformation as the pre-trained model:" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 246, + 502, + 480, + 515 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 246, + 502, + 480, + 515 + ], + "spans": [ + { + "bbox": [ + 246, + 502, + 480, + 515 + ], + "type": "interline_equation", + "content": "H _ {i} = \\operatorname {F F N} \\left(\\operatorname {N o r m} \\left(E _ {i}\\right)\\right) + E _ {i}, \\tag {5}", + "image_path": "874af01d470d0f3f1428cea6f81029d75196c403da4cb8f7504b6fce857198e7.jpg" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 130, + 522, + 450, + 533 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 522, + 450, + 533 + ], + "spans": [ + { + "bbox": [ + 130, + 522, + 450, + 533 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 130, + 522, + 450, + 533 + ], + "type": "inline_equation", + "content": "\\mathrm{Norm}(\\cdot)" + }, + { + "bbox": [ + 130, + 522, + 450, + 533 + ], + "type": "text", + "content": " denotes the normalization used in the pre-trained model." + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 130, + 534, + 479, + 569 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 534, + 479, + 569 + ], + "spans": [ + { + "bbox": [ + 130, + 534, + 479, + 569 + ], + "type": "text", + "content": "To maximize the potential of KEM, apart from the final output " + }, + { + "bbox": [ + 130, + 534, + 479, + 569 + ], + "type": "inline_equation", + "content": "H_{i}" + }, + { + "bbox": [ + 130, + 534, + 479, + 569 + ], + "type": "text", + "content": ", we additionally leverage features " + }, + { + "bbox": [ + 130, + 534, + 479, + 569 + ], + "type": "inline_equation", + "content": "F_{i}^{att}" + }, + { + "bbox": [ + 130, + 534, + 479, + 569 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 130, + 534, + 479, + 569 + ], + "type": "inline_equation", + "content": "F_{i}^{ffn}" + }, + { + "bbox": [ + 130, + 534, + 479, + 569 + ], + "type": "text", + "content": " after the attention module and the FFN for classification. They are obtained as follows:" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 239, + 577, + 480, + 609 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 239, + 577, + 480, + 609 + ], + "spans": [ + { + "bbox": [ + 239, + 577, + 480, + 609 + ], + "type": "interline_equation", + "content": "\\left\\{ \\begin{array}{l} F _ {i} ^ {a t t} = \\operatorname {A t t n} \\left(Q _ {i} ^ {k e}, K _ {i} ^ {k e}, V _ {i} ^ {k e}\\right) \\\\ F _ {i} ^ {f f n} = \\operatorname {F F N} \\left(\\operatorname {N o r m} \\left(E _ {i}\\right)\\right) \\end{array} . \\right. \\tag {6}", + "image_path": "9233e41a92fef8d04c8d526a2ae3f7f16144b79620296fe6195f9b14a4fb448c.jpg" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 132, + 623, + 257, + 634 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 623, + 257, + 634 + ], + "spans": [ + { + "bbox": [ + 132, + 623, + 257, + 634 + ], + "type": "text", + "content": "3.4 Classification Head" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 130, + 641, + 479, + 666 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 641, + 479, + 666 + ], + "spans": [ + { + "bbox": [ + 130, + 641, + 479, + 666 + ], + "type": "text", + "content": "The classification head makes a prediction of the input sample relying on rich and hierarchical features from KEMs, " + }, + { + "bbox": [ + 130, + 641, + 479, + 666 + ], + "type": "inline_equation", + "content": "\\{F_1^{att}, F_1^{ffn}, H_1, \\dots, F_l^{att}, F_l^{ffn}, H_l\\}" + }, + { + "bbox": [ + 130, + 641, + 479, + 666 + ], + "type": "text", + "content": ", where " + }, + { + "bbox": [ + 130, + 641, + 479, + 666 + ], + "type": "inline_equation", + "content": "l" + } + ] + } + ], + "index": 16 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "spans": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "type": "text", + "content": "6" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 166, + 91, + 299, + 102 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 166, + 91, + 299, + 102 + ], + "spans": [ + { + "bbox": [ + 166, + 91, + 299, + 102 + ], + "type": "text", + "content": "Taolin Zhang, Jiawang Bai et al." + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 5 + }, + { + "para_blocks": [ + { + "bbox": [ + 130, + 116, + 479, + 152 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 116, + 479, + 152 + ], + "spans": [ + { + "bbox": [ + 130, + 116, + 479, + 152 + ], + "type": "text", + "content": "is the number of blocks of the pre-trained ViT. Instead of directly applying an MLP classifier on top of these features, we propose several techniques for better performance. Figure 2 shows the overall architecture of the classification head." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 130, + 152, + 325, + 259 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 152, + 325, + 259 + ], + "spans": [ + { + "bbox": [ + 130, + 152, + 325, + 259 + ], + "type": "text", + "content": "Feature projection. Since these features are generated across different layers from KEMs, they may be misaligned in the representation space. Therefore, before aggregating all features, we align all features with projection layers. We use bottleneck-structured fully connected layers and share the projection layers across all features to improve the parameter efficiency." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 130, + 261, + 326, + 463 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 261, + 326, + 463 + ], + "spans": [ + { + "bbox": [ + 130, + 261, + 326, + 463 + ], + "type": "text", + "content": "Conditional feature aggregation. These features contain a large amount of information but also some redundancy, perhaps hurting the performance of downstream tasks. Therefore, it is crucial to perform feature selection, aiming to extract the truly beneficial information. We propose a strategy of conditional feature aggregation, which generates weights conditioned on " + }, + { + "bbox": [ + 130, + 261, + 326, + 463 + ], + "type": "inline_equation", + "content": "H_{l}" + }, + { + "bbox": [ + 130, + 261, + 326, + 463 + ], + "type": "text", + "content": " and performs a weighted-sum over all features. Our strategy enables SynQT to adaptively rely on different features to classify different samples. Despite only a few parameters, the conditional feature aggregation results in better performance compared to a simple average or weighted sum with learnable but fixed weights, as shown in our experiments." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 130, + 464, + 482, + 548 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 464, + 482, + 548 + ], + "spans": [ + { + "bbox": [ + 130, + 464, + 482, + 548 + ], + "type": "text", + "content": "DropFeat in training. Due to rich features fed into the classification head, overfitting likely happens and degrades the performance. To solve this issue, inspired by the idea of dropout [23, 42, 47], we propose DropFeat, a simple regularization technique to prevent overfitting. DropFeat randomly removes some input features for the classification head during training, while all features are leveraged during inference. DropFeat ensures that the model does not overly rely on a specific feature, which significantly improves the model's performance." + } + ] + } + ], + "index": 5 + }, + { + "type": "image", + "bbox": [ + 340, + 177, + 481, + 339 + ], + "blocks": [ + { + "bbox": [ + 340, + 177, + 481, + 339 + ], + "lines": [ + { + "bbox": [ + 340, + 177, + 481, + 339 + ], + "spans": [ + { + "bbox": [ + 340, + 177, + 481, + 339 + ], + "type": "image", + "image_path": "e2c9082112eeda8075dbba1853437621577731d2ee22dc2c4b3a2831bdbde3c7.jpg" + } + ] + } + ], + "index": 6, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 331, + 348, + 482, + 449 + ], + "lines": [ + { + "bbox": [ + 331, + 348, + 482, + 449 + ], + "spans": [ + { + "bbox": [ + 331, + 348, + 482, + 449 + ], + "type": "text", + "content": "Fig.2: Classification head in SynQT, corresponding to the head in Figure 1. We utilize a shared projection to align the hierarchical features and apply feature aggregation conditioned on the output of the last block " + }, + { + "bbox": [ + 331, + 348, + 482, + 449 + ], + "type": "inline_equation", + "content": "H_{l}" + }, + { + "bbox": [ + 331, + 348, + 482, + 449 + ], + "type": "text", + "content": ". We also adopt DropFeat to prevent overfitting, which randomly drops some features during training." + } + ] + } + ], + "index": 7, + "angle": 0, + "type": "image_caption" + } + ], + "index": 6 + }, + { + "bbox": [ + 131, + 568, + 224, + 582 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 131, + 568, + 224, + 582 + ], + "spans": [ + { + "bbox": [ + 131, + 568, + 224, + 582 + ], + "type": "text", + "content": "4 Experiment" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 131, + 595, + 190, + 608 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 131, + 595, + 190, + 608 + ], + "spans": [ + { + "bbox": [ + 131, + 595, + 190, + 608 + ], + "type": "text", + "content": "4.1 Setup" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 130, + 617, + 482, + 666 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 617, + 482, + 666 + ], + "spans": [ + { + "bbox": [ + 130, + 617, + 482, + 666 + ], + "type": "text", + "content": "Datasets. We conduct experiments on the VTAB-1K benchmark and the fine-grained visual classification (FGVC) datasets. VTAB-1K consists of 19 different classification tasks categorized into three groups: Natural, Specialized, and Structured. Each task of VTAB-1K contains 1000 training samples. We evaluate the" + } + ] + } + ], + "index": 10 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 149, + 90, + 448, + 102 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 149, + 90, + 448, + 102 + ], + "spans": [ + { + "bbox": [ + 149, + 90, + 448, + 102 + ], + "type": "text", + "content": "Parameter-Efficient and Memory-Efficient Tuning for Vision Transformer" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 474, + 91, + 480, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 474, + 91, + 480, + 100 + ], + "spans": [ + { + "bbox": [ + 474, + 91, + 480, + 100 + ], + "type": "text", + "content": "7" + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 6 + }, + { + "para_blocks": [ + { + "type": "table", + "bbox": [ + 134, + 168, + 480, + 303 + ], + "blocks": [ + { + "bbox": [ + 130, + 114, + 482, + 159 + ], + "lines": [ + { + "bbox": [ + 130, + 114, + 482, + 159 + ], + "spans": [ + { + "bbox": [ + 130, + 114, + 482, + 159 + ], + "type": "text", + "content": "Table 1: Full results on VTAB-1K benchmark with pre-trained ViT-B/16 on ImageNet21K. “# Memory” specifies memory cost during training with 64 as batch size on CIFAR100. We report the Top-1 accuracy and the results are averaged over group-wise mean values." + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 134, + 168, + 480, + 303 + ], + "lines": [ + { + "bbox": [ + 134, + 168, + 480, + 303 + ], + "spans": [ + { + "bbox": [ + 134, + 168, + 480, + 303 + ], + "type": "table", + "html": "
# Memory (GB)NaturalSpecializedStructured
CIFAR100Caltech101DTDFlower102PetsSVHNSun397CamelyonEuroSATRess45RetinopathyClever-CountClever-DistDMLabKITTL-DistdSpr-LocdSpr-OrisNORB-AzimsNORB-EleAverage
Traditional Fine-Tuning
Linear Probing2.550.685.661.479.586.540.838.079.791.571.765.541.434.434.155.418.126.416.524.852.7
Full Fine-Tuning11.144.384.554.184.774.787.226.985.395.076.070.471.560.546.972.974.538.728.523.863.2
PETL Methods
BitFit [54]7.872.887.059.297.585.359.951.478.791.672.969.861.555.632.455.966.640.015.725.165.2
VPT-Shallow [17]8.177.786.962.697.587.374.551.278.292.075.672.950.558.640.567.168.736.120.234.167.8
VPT-Deep [17]8.178.890.865.898.088.378.149.681.896.183.468.468.560.046.572.873.647.932.937.872.0
Memory-Efficient PETL Methods
VQT [45]3.266.389.967.897.984.779.945.579.095.280.974.746.761.645.163.662.932.130.028.868.3
SynQT (Ours)3.470.989.768.898.589.677.850.682.396.783.575.271.862.748.575.474.149.031.736.172.9
", + "image_path": "b56fe2d14702c227b54a844826dfbd203c99d46e8d9673019f9ce1994c535321.jpg" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "table_body" + } + ], + "index": 3 + }, + { + "bbox": [ + 130, + 325, + 480, + 372 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 325, + 480, + 372 + ], + "spans": [ + { + "bbox": [ + 130, + 325, + 480, + 372 + ], + "type": "text", + "content": "few-shot performance of SynQT on 5 FGVC datasets: FGVC-Aircraft [36], OxfordPets [38], Food-101 [5], Stanford Cars [22], and Oxford-Flowers102 [37]. We follow the split in [59] and report the results in 1, 2, 4, 8, and 16-shot settings. The accuracy is averaged over three trials with different seeds." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 130, + 373, + 493, + 481 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 373, + 493, + 481 + ], + "spans": [ + { + "bbox": [ + 130, + 373, + 493, + 481 + ], + "type": "text", + "content": "Implementation details. We utilize a pre-trained ViT-B/16 [10] on the ImageNet-21K dataset as the foundational model. We also conduct experiments on the ImageNet-1K supervised pre-trained backbone, as well as on the self-supervised checkpoints (DINO [7] and MAE [14]). For a lightweight design of QSM, we use an extremely small number of tokens, with the number of tokens " + }, + { + "bbox": [ + 130, + 373, + 493, + 481 + ], + "type": "inline_equation", + "content": "n" + }, + { + "bbox": [ + 130, + 373, + 493, + 481 + ], + "type": "text", + "content": " from " + }, + { + "bbox": [ + 130, + 373, + 493, + 481 + ], + "type": "inline_equation", + "content": "\\{1, 4, 16\\}" + }, + { + "bbox": [ + 130, + 373, + 493, + 481 + ], + "type": "text", + "content": ". This is significantly fewer than the number of tokens in VPT, which can reach up to 200. Additionally, the hyper-parameters " + }, + { + "bbox": [ + 130, + 373, + 493, + 481 + ], + "type": "inline_equation", + "content": "s'" + }, + { + "bbox": [ + 130, + 373, + 493, + 481 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 130, + 373, + 493, + 481 + ], + "type": "inline_equation", + "content": "s''" + }, + { + "bbox": [ + 130, + 373, + 493, + 481 + ], + "type": "text", + "content": " are chosen from " + }, + { + "bbox": [ + 130, + 373, + 493, + 481 + ], + "type": "inline_equation", + "content": "\\{0.1, 1\\}" + }, + { + "bbox": [ + 130, + 373, + 493, + 481 + ], + "type": "text", + "content": ". The hidden size of the bottleneck projection is set to 48, except attention projection of the query, key, and value in QSM, which is set to 8." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 130, + 482, + 482, + 590 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 482, + 482, + 590 + ], + "spans": [ + { + "bbox": [ + 130, + 482, + 482, + 590 + ], + "type": "text", + "content": "Due to the limited number of validation samples in VTAB-1K, we implement 5-fold cross-validation on the training sets to determine the optimal hyperparameters. For few-shot learning, we use the default splits of FGVC datasets to search for the best hyperparameters. The Top-1 accuracy is reported on the test sets in all our experiments. Following common practice [18], we adopt AdamW [33] as the optimizer and a cosine learning rate scheduler with a constant batch size of 64. For data augmentation, we adjust the image size to " + }, + { + "bbox": [ + 130, + 482, + 482, + 590 + ], + "type": "inline_equation", + "content": "224 \\times 224" + }, + { + "bbox": [ + 130, + 482, + 482, + 590 + ], + "type": "text", + "content": " directly for VTAB-1K, and for few-shot learning, we resize the image to " + }, + { + "bbox": [ + 130, + 482, + 482, + 590 + ], + "type": "inline_equation", + "content": "256 \\times 256" + }, + { + "bbox": [ + 130, + 482, + 482, + 590 + ], + "type": "text", + "content": " before applying a center crop." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 132, + 608, + 228, + 619 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 608, + 228, + 619 + ], + "spans": [ + { + "bbox": [ + 132, + 608, + 228, + 619 + ], + "type": "text", + "content": "4.2 Main Results" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 130, + 629, + 481, + 667 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 629, + 481, + 667 + ], + "spans": [ + { + "bbox": [ + 130, + 629, + 481, + 667 + ], + "type": "text", + "content": "Besides the traditional fine-tuning methods (linear probing and full fine-tuning), we compare our method with PETL methods including BitFit [54], VPT-Shallow [17], and VPT-Deep [17], as well as a memory-efficient one, VQT [45], on VTAB-" + } + ] + } + ], + "index": 8 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "spans": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "type": "text", + "content": "8" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 166, + 91, + 299, + 102 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 166, + 91, + 299, + 102 + ], + "spans": [ + { + "bbox": [ + 166, + 91, + 299, + 102 + ], + "type": "text", + "content": "Taolin Zhang, Jiawang Bai et al." + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 7 + }, + { + "para_blocks": [ + { + "type": "table", + "bbox": [ + 161, + 146, + 455, + 338 + ], + "blocks": [ + { + "bbox": [ + 132, + 114, + 480, + 137 + ], + "lines": [ + { + "bbox": [ + 132, + 114, + 480, + 137 + ], + "spans": [ + { + "bbox": [ + 132, + 114, + 480, + 137 + ], + "type": "text", + "content": "Table 2: Classification results on MAE, DINO, and Supervised ImageNet-1K pre-trained backbones. We report the Top-1 accuracy and the averaged results." + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 161, + 146, + 455, + 338 + ], + "lines": [ + { + "bbox": [ + 161, + 146, + 455, + 338 + ], + "spans": [ + { + "bbox": [ + 161, + 146, + 455, + 338 + ], + "type": "table", + "html": "
MethodsNaturalSpecializedStructuredAverage
MAE backbone
BitFit [54]63.580.657.267.1
VPT-Shallow [17]38.372.025.045.1
VPT-Deep [17]50.876.437.354.8
VQT [45]56.678.643.459.5
SynQT (Ours)66.082.658.268.9
DINO backbone
BitFit [54]76.883.849.269.9
VPT-Shallow [17]69.882.638.563.6
VPT-Deep [17]76.783.847.369.3
VQT [45]71.083.637.263.9
SynQT (Ours)75.184.856.972.3
Supervised ImageNet-1K backbone
BitFit [54]74.683.449.469.1
VPT-Shallow [17]63.780.340.261.4
VPT-Deep [17]74.982.953.970.6
VQT [45]72.784.549.368.8
SynQT (Ours)73.284.756.271.4
", + "image_path": "65fba2f418828f031e49bd7b49d8c618445fd72c013581f94e36172076d6d161.jpg" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "table_body" + } + ], + "index": 3 + }, + { + "bbox": [ + 130, + 357, + 482, + 453 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 357, + 482, + 453 + ], + "spans": [ + { + "bbox": [ + 130, + 357, + 482, + 453 + ], + "type": "text", + "content": "1K. As can be seen from Table 1, the most striking observation emerging from the comparison is that SynQT outperforms other methods while saving up to " + }, + { + "bbox": [ + 130, + 357, + 482, + 453 + ], + "type": "inline_equation", + "content": "50\\%" + }, + { + "bbox": [ + 130, + 357, + 482, + 453 + ], + "type": "text", + "content": " of GPU memory. Given the same batch size, previous PETL methods rely on the full backward propagation of the original model and consume " + }, + { + "bbox": [ + 130, + 357, + 482, + 453 + ], + "type": "inline_equation", + "content": "7 \\sim 8" + }, + { + "bbox": [ + 130, + 357, + 482, + 453 + ], + "type": "text", + "content": " GB of memory, whereas SynQT disentangles transfer learning and reduces the memory cost to " + }, + { + "bbox": [ + 130, + 357, + 482, + 453 + ], + "type": "inline_equation", + "content": "3.4" + }, + { + "bbox": [ + 130, + 357, + 482, + 453 + ], + "type": "text", + "content": " GB. Notably, our method performs better than VPT-Deep on 14 of 19 tasks. We notice that although the memory cost of SynQT is slightly higher, it achieves a significant performance improvement of " + }, + { + "bbox": [ + 130, + 357, + 482, + 453 + ], + "type": "inline_equation", + "content": "4.6\\%" + }, + { + "bbox": [ + 130, + 357, + 482, + 453 + ], + "type": "text", + "content": " compared to VQT." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 130, + 453, + 482, + 500 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 453, + 482, + 500 + ], + "spans": [ + { + "bbox": [ + 130, + 453, + 482, + 500 + ], + "type": "text", + "content": "In order to further validate the robustness of our method, we compare SynQT with baseline models across various checkpoints which are derived from different learning paradigms, including supervised learning (pretrained over ImageNet-21K [9]), and self-supervised learning (DINO" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 130, + 500, + 331, + 609 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 500, + 331, + 609 + ], + "spans": [ + { + "bbox": [ + 130, + 500, + 331, + 609 + ], + "type": "text", + "content": "[7] and MAE [14]). The results in Table 2 indicate that SynQT shows strong compatibility and outperforms PETL methods by a substantial margin over different backbones. For instance, compared to the strongest baselines, our method achieves improvements of " + }, + { + "bbox": [ + 130, + 500, + 331, + 609 + ], + "type": "inline_equation", + "content": "1.8\\%" + }, + { + "bbox": [ + 130, + 500, + 331, + 609 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 130, + 500, + 331, + 609 + ], + "type": "inline_equation", + "content": "3.3\\%" + }, + { + "bbox": [ + 130, + 500, + 331, + 609 + ], + "type": "text", + "content": ", and " + }, + { + "bbox": [ + 130, + 500, + 331, + 609 + ], + "type": "inline_equation", + "content": "0.8\\%" + }, + { + "bbox": [ + 130, + 500, + 331, + 609 + ], + "type": "text", + "content": " on average with MAE, DINO, and supervised ImageNet-1K backbones, respectively." + } + ] + } + ], + "index": 6 + }, + { + "type": "image", + "bbox": [ + 338, + 514, + 479, + 622 + ], + "blocks": [ + { + "bbox": [ + 338, + 514, + 479, + 622 + ], + "lines": [ + { + "bbox": [ + 338, + 514, + 479, + 622 + ], + "spans": [ + { + "bbox": [ + 338, + 514, + 479, + 622 + ], + "type": "image", + "image_path": "ca231b864699c3f170917ef9f73482e8e8b38cb1d2960eea8e28b4cca27a12ec.jpg" + } + ] + } + ], + "index": 7, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 335, + 632, + 482, + 677 + ], + "lines": [ + { + "bbox": [ + 335, + 632, + 482, + 677 + ], + "spans": [ + { + "bbox": [ + 335, + 632, + 482, + 677 + ], + "type": "text", + "content": "Fig.3: Performance comparison under memory constraints. SynQT outperforms other methods with different memory costs." + } + ] + } + ], + "index": 10, + "angle": 0, + "type": "image_caption" + } + ], + "index": 7 + }, + { + "bbox": [ + 132, + 624, + 313, + 636 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 624, + 313, + 636 + ], + "spans": [ + { + "bbox": [ + 132, + 624, + 313, + 636 + ], + "type": "text", + "content": "4.3 Memory-Efficiency Evaluation" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 132, + 640, + 329, + 665 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 640, + 329, + 665 + ], + "spans": [ + { + "bbox": [ + 132, + 640, + 329, + 665 + ], + "type": "text", + "content": "Previous PETL methods [8, 15, 16, 18] focus on reducing the number of parameters when" + } + ] + } + ], + "index": 9 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 150, + 91, + 448, + 102 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 150, + 91, + 448, + 102 + ], + "spans": [ + { + "bbox": [ + 150, + 91, + 448, + 102 + ], + "type": "text", + "content": "Parameter-Efficient and Memory-Efficient Tuning for Vision Transformer" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 474, + 91, + 480, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 474, + 91, + 480, + 100 + ], + "spans": [ + { + "bbox": [ + 474, + 91, + 480, + 100 + ], + "type": "text", + "content": "9" + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 8 + }, + { + "para_blocks": [ + { + "type": "table", + "bbox": [ + 134, + 161, + 480, + 252 + ], + "blocks": [ + { + "bbox": [ + 132, + 107, + 481, + 152 + ], + "lines": [ + { + "bbox": [ + 132, + 107, + 481, + 152 + ], + "spans": [ + { + "bbox": [ + 132, + 107, + 481, + 152 + ], + "type": "text", + "content": "Table 3: Full results on VTAB-1K benchmark under memory constraints. The memory is constrained within 5GB and “# Memory” specifies the memory cost during training with 64 as batch size on CIFAR100. We report the Top-1 accuracy and the results are averaged over group-wise mean values." + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 134, + 161, + 480, + 252 + ], + "lines": [ + { + "bbox": [ + 134, + 161, + 480, + 252 + ], + "spans": [ + { + "bbox": [ + 134, + 161, + 480, + 252 + ], + "type": "table", + "html": "
# Memory (GB)NaturalSpecializedStructured
CIFAR100Caltech101DTDFlower102PetsSVHNSun397CanelyonEuroSATResics45RetinopathyClevr-CountClevr-DistDMLabKITTI-DistdSpr-LocdSpr-OrisNORB-AzimsNORB-EteAverage
FacT-TT [18]4.340.484.259.896.886.653.445.879.492.273.374.459.250.242.568.848.225.919.828.763.2
Adapter [15]4.764.986.062.897.386.646.844.479.893.976.774.663.340.238.968.533.132.319.830.564.0
LoRA [16]4.862.582.160.497.586.658.645.380.695.378.872.864.845.141.368.549.628.124.630.565.5
AdaptFormer [8]4.665.686.162.096.986.235.545.682.696.277.674.466.247.741.372.355.629.422.633.965.8
SynQT(Ours)3.470.989.768.898.589.677.850.682.396.783.575.271.862.748.575.474.149.031.736.172.9
", + "image_path": "c211f0350bfab81e39e15cd92d6edc2b3c32c551748568e3d7de60610bcbf66a.jpg" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "table_body" + } + ], + "index": 3 + }, + { + "bbox": [ + 130, + 275, + 479, + 346 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 275, + 479, + 346 + ], + "spans": [ + { + "bbox": [ + 130, + 275, + 479, + 346 + ], + "type": "text", + "content": "fine-tuning the model. While decreasing the cost of model storage, these methods are not widely applicable in practical training because they still require a significant amount of GPU memory. Therefore, we propose SynQT to overcome this problem. In this section, we highlight our improvements in terms of memory costs by comparing our method with some SOTA PETL methods under a given GPU memory constraint." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 130, + 347, + 480, + 418 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 347, + 480, + 418 + ], + "spans": [ + { + "bbox": [ + 130, + 347, + 480, + 418 + ], + "type": "text", + "content": "For baseline methods, we discard some specific trainable structures to meet the memory constraint. Specifically, we remove these additional structures near the input layer, which means that the training only requires a partial backpropagation with a lower memory usage. We progressively remove the structure nearest to the input layer until it meets the given memory constraint. Besides, for a fair comparison, we reduce the memory usage of SynQT in the same manner." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 130, + 419, + 481, + 503 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 419, + 481, + 503 + ], + "spans": [ + { + "bbox": [ + 130, + 419, + 481, + 503 + ], + "type": "text", + "content": "We compare SynQT with SOTA baselines including FacT-TT [18], Adapter [15], LoRA [16] and AdaptFormer [8] under such a constraint and report the results in Table 3. While previous methods can achieve promising performance with few trainable parameters, they unfortunately suffer from significant performance drops when such memory constraints are imposed. In contrast, SynQT excels in this scenario, achieving state-of-the-art performance and surpassing AdaptFormer by " + }, + { + "bbox": [ + 130, + 419, + 481, + 503 + ], + "type": "inline_equation", + "content": "7.1\\%" + }, + { + "bbox": [ + 130, + 419, + 481, + 503 + ], + "type": "text", + "content": " and VQT by " + }, + { + "bbox": [ + 130, + 419, + 481, + 503 + ], + "type": "inline_equation", + "content": "4.6\\%" + }, + { + "bbox": [ + 130, + 419, + 481, + 503 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 130, + 504, + 482, + 600 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 504, + 482, + 600 + ], + "spans": [ + { + "bbox": [ + 130, + 504, + 482, + 600 + ], + "type": "text", + "content": "To have a better understanding of how memory constraints affect model performance, we estimate the results of the methods mentioned above across varying memory costs. As depicted in Figure 3, SynQT continues to outperform the baselines across different memory costs. Notably, when the constraint is very strict, i.e., 3 GB, most of the existing PETL methods perform poorly, with a gap of almost " + }, + { + "bbox": [ + 130, + 504, + 482, + 600 + ], + "type": "inline_equation", + "content": "7\\%" + }, + { + "bbox": [ + 130, + 504, + 482, + 600 + ], + "type": "text", + "content": " compared to SynQT. Moreover, it is noteworthy that relaxing the memory constraints improves the performance of all the methods, especially for a significant improvement observed for our SynQT." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 132, + 619, + 308, + 632 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 619, + 308, + 632 + ], + "spans": [ + { + "bbox": [ + 132, + 619, + 308, + 632 + ], + "type": "text", + "content": "4.4 Few-Shot Learning on FGVC" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 130, + 641, + 481, + 667 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 641, + 481, + 667 + ], + "spans": [ + { + "bbox": [ + 130, + 641, + 481, + 667 + ], + "type": "text", + "content": "To further evaluate SynQT's ability to transfer knowledge from the pre-trained model using few samples, we compare SynQT with BitFit [54], VPT-Shallow [17]," + } + ] + } + ], + "index": 9 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "spans": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "type": "text", + "content": "10" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 166, + 91, + 299, + 102 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 166, + 91, + 299, + 102 + ], + "spans": [ + { + "bbox": [ + 166, + 91, + 299, + 102 + ], + "type": "text", + "content": "Taolin Zhang, Jiawang Bai et al." + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 9 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 136, + 114, + 478, + 216 + ], + "blocks": [ + { + "bbox": [ + 136, + 114, + 478, + 216 + ], + "lines": [ + { + "bbox": [ + 136, + 114, + 478, + 216 + ], + "spans": [ + { + "bbox": [ + 136, + 114, + 478, + 216 + ], + "type": "image", + "image_path": "cf53836e9aaebae18596c30f1248303246867fe49cef0b84359871c698c159f5.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 131, + 226, + 480, + 249 + ], + "lines": [ + { + "bbox": [ + 131, + 226, + 480, + 249 + ], + "spans": [ + { + "bbox": [ + 131, + 226, + 480, + 249 + ], + "type": "text", + "content": "Fig. 4: Performance comparison in few-shot learning. Overall, SynQT outperforms other baselines in the extreme low-data regime." + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_caption" + } + ], + "index": 2 + }, + { + "type": "table", + "bbox": [ + 135, + 293, + 479, + 344 + ], + "blocks": [ + { + "bbox": [ + 131, + 259, + 479, + 282 + ], + "lines": [ + { + "bbox": [ + 131, + 259, + 479, + 282 + ], + "spans": [ + { + "bbox": [ + 131, + 259, + 479, + 282 + ], + "type": "text", + "content": "Table 4: Applying SynQT to Swin-B and ConvNeXt-B. We report the Top-1 accuracy and the averaged results." + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 135, + 293, + 479, + 344 + ], + "lines": [ + { + "bbox": [ + 135, + 293, + 479, + 344 + ], + "spans": [ + { + "bbox": [ + 135, + 293, + 479, + 344 + ], + "type": "table", + "html": "
MethodSwin-BConvNeXt-B
NaturalSpecializedStructuredAverageNaturalSpecializedStructuredAverage
Linear73.580.833.562.674.581.534.863.6
VPT-Deep76.884.553.471.678.583.044.668.7
SynQT77.586.057.873.878.182.148.269.5
", + "image_path": "9285a73bfda58c5a1c317b169a689e926b8434311a170e6a647e08bea4cb512e.jpg" + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "table_body" + } + ], + "index": 5 + }, + { + "bbox": [ + 130, + 365, + 480, + 389 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 365, + 480, + 389 + ], + "spans": [ + { + "bbox": [ + 130, + 365, + 480, + 389 + ], + "type": "text", + "content": "VPT-Deep [17], and VQT [45] in few-shot setting on 5 FGVC datasets. The results are reported in Figure 4." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 130, + 390, + 481, + 510 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 390, + 481, + 510 + ], + "spans": [ + { + "bbox": [ + 130, + 390, + 481, + 510 + ], + "type": "text", + "content": "Overall, SynQT significantly outperforms all other baselines in 4, 6, 8, and 16-shot scenarios, and its performance is comparable in the 1-shot scenario. The gaps between VQT and SynQT indicate that the lightweight, task-specific QSM enables SynQT to capture beneficial information from samples in the extremely low-data regime. Simply incorporating a trainable prompt as a query, as in VQT, is not sufficient to handle such challenging few-shot tasks. Interestingly, VPT-Shallow outperforms VPT-Deep in the low-data regime, due to the fact that the larger number of prompts in VPT-Deep are not well-trained on the limited data. This suggests that merely increasing the number of prompt numbers could potentially lead to a performance drop." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 131, + 526, + 265, + 539 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 131, + 526, + 265, + 539 + ], + "spans": [ + { + "bbox": [ + 131, + 526, + 265, + 539 + ], + "type": "text", + "content": "4.5 Versatility of SynQT" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 130, + 545, + 480, + 569 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 545, + 480, + 569 + ], + "spans": [ + { + "bbox": [ + 130, + 545, + 480, + 569 + ], + "type": "text", + "content": "To demonstrate the versatility of SynQT, we apply SynQT to different model architectures and combine SynQT with existing PETL methods." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 130, + 570, + 481, + 665 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 570, + 481, + 665 + ], + "spans": [ + { + "bbox": [ + 130, + 570, + 481, + 665 + ], + "type": "text", + "content": "Applying SynQT to different model architectures. Here, we apply SynQT to an advanced Transformer-like architecture (Swin [31]) and a convolutional neural network (ConvNeXt [32]). Following [17], we introduce prompts for each window in Swin and incorporate learnable prompt pixels for the feature map in ConvNeXt. Specifically, due to the varying hidden sizes and number of blocks in different layers, we extract features from the third layer of Swin and the last layer of ConvNext to ensure sufficient extraction of beneficial information. We find that both of them significantly benefit from SynQT, as shown in Table 4." + } + ] + } + ], + "index": 10 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 150, + 91, + 448, + 102 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 150, + 91, + 448, + 102 + ], + "spans": [ + { + "bbox": [ + 150, + 91, + 448, + 102 + ], + "type": "text", + "content": "Parameter-Efficient and Memory-Efficient Tuning for Vision Transformer" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 470, + 91, + 479, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 470, + 91, + 479, + 100 + ], + "spans": [ + { + "bbox": [ + 470, + 91, + 479, + 100 + ], + "type": "text", + "content": "11" + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 10 + }, + { + "para_blocks": [ + { + "type": "table", + "bbox": [ + 134, + 148, + 312, + 223 + ], + "blocks": [ + { + "bbox": [ + 132, + 125, + 314, + 146 + ], + "lines": [ + { + "bbox": [ + 132, + 125, + 314, + 146 + ], + "spans": [ + { + "bbox": [ + 132, + 125, + 314, + 146 + ], + "type": "text", + "content": "Table 5: Combining different PETL methods. AF is short for AdaptFormer." + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 134, + 148, + 312, + 223 + ], + "lines": [ + { + "bbox": [ + 134, + 148, + 312, + 223 + ], + "spans": [ + { + "bbox": [ + 134, + 148, + 312, + 223 + ], + "type": "table", + "html": "
MethodsNaturalSpecializedStructuredAverage
AF80.182.350.370.9
AF+VQT79.684.353.072.3
AF+SynQT82.685.457.975.3
VPT79.184.654.472.7
VPT+VQT78.983.754.672.4
VPT+SynQT81.184.957.774.1
", + "image_path": "222a00b7ca35462df68e3db11e2636134ae84f9b0738da944aafb436e95321a0.jpg" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "table_body" + } + ], + "index": 3 + }, + { + "type": "table", + "bbox": [ + 317, + 161, + 479, + 209 + ], + "blocks": [ + { + "bbox": [ + 315, + 138, + 480, + 160 + ], + "lines": [ + { + "bbox": [ + 315, + 138, + 480, + 160 + ], + "spans": [ + { + "bbox": [ + 315, + 138, + 480, + 160 + ], + "type": "text", + "content": "Table 6: Necessity of disentangle-ment (KEM and QSM)." + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 317, + 161, + 479, + 209 + ], + "lines": [ + { + "bbox": [ + 317, + 161, + 479, + 209 + ], + "spans": [ + { + "bbox": [ + 317, + 161, + 479, + 209 + ], + "type": "table", + "html": "
MethodNatural SpecializedStructuredAverage
Linear63.277.131.457.2
+KEM71.880.132.461.4
+QSM77.883.655.072.1
+Head (SynQT)78.088.456.272.9
", + "image_path": "478c2ad1893b38ee7237375105b08dd0fbe9ef71c85a3c02bbb491b7a06f746b.jpg" + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "table_body" + } + ], + "index": 5 + }, + { + "bbox": [ + 130, + 246, + 479, + 269 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 246, + 479, + 269 + ], + "spans": [ + { + "bbox": [ + 130, + 246, + 479, + 269 + ], + "type": "text", + "content": "On average, SynQT outperforms VPT-Deep by " + }, + { + "bbox": [ + 130, + 246, + 479, + 269 + ], + "type": "inline_equation", + "content": "2.2\\%" + }, + { + "bbox": [ + 130, + 246, + 479, + 269 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 130, + 246, + 479, + 269 + ], + "type": "inline_equation", + "content": "1.2\\%" + }, + { + "bbox": [ + 130, + 246, + 479, + 269 + ], + "type": "text", + "content": " for Swin and ConvNeXt, respectively." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 130, + 270, + 480, + 354 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 270, + 480, + 354 + ], + "spans": [ + { + "bbox": [ + 130, + 270, + 480, + 354 + ], + "type": "text", + "content": "Combining SynQT with existing PETL methods. We further jointly utilize SynQT and existing PETL methods to adapt pre-trained models, as shown in Table 5. The results demonstrate that both VPT and AdaptFormer benefit from SynQT on VTAB-1K, with improvements of " + }, + { + "bbox": [ + 130, + 270, + 480, + 354 + ], + "type": "inline_equation", + "content": "4.4\\%" + }, + { + "bbox": [ + 130, + 270, + 480, + 354 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 130, + 270, + 480, + 354 + ], + "type": "inline_equation", + "content": "1.4\\%" + }, + { + "bbox": [ + 130, + 270, + 480, + 354 + ], + "type": "text", + "content": " on average, respectively. We can see that for AdaptFormer and VPT, VQT may result in a performance drop in some cases (e.g., the Natural group), whereas equipping SynQT leads to significant performance improvements across all groups." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 132, + 371, + 238, + 384 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 371, + 238, + 384 + ], + "spans": [ + { + "bbox": [ + 132, + 371, + 238, + 384 + ], + "type": "text", + "content": "4.6 Ablation Study" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 130, + 390, + 479, + 413 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 390, + 479, + 413 + ], + "spans": [ + { + "bbox": [ + 130, + 390, + 479, + 413 + ], + "type": "text", + "content": "We ablate our SynQT with different settings in Tables 6 and 7 for a better understanding of SynQT." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 130, + 415, + 480, + 498 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 415, + 480, + 498 + ], + "spans": [ + { + "bbox": [ + 130, + 415, + 480, + 498 + ], + "type": "text", + "content": "Disentanglement. To demonstrate the necessity of disentanglement, we perform an ablation in Table 6. We set the baseline as linear probing since KEM exploits intermediate features. To gain deeper insights into the roles of QSM and Head, we employ random prompts to query in KEM as the model without QSM, and we utilize a simple average for aggregation as the model without Head. The results demonstrate that our idea of disentangling (QSM and KEM) is the key to achieving superior performance." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 130, + 498, + 480, + 581 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 498, + 480, + 581 + ], + "spans": [ + { + "bbox": [ + 130, + 498, + 480, + 581 + ], + "type": "text", + "content": "Hidden size in QSM. We vary the hidden size of the bottleneck module in QSM and report the results in Table 7a. With a larger hidden size, the performance improves due to the increasing capacity. However, when the hidden size exceeds 48, the benefit of introducing new parameters diminishes and overfitting may happen in datasets with few training samples. The results indicate that redundant parameters might not enhance the model's performance, and a hidden size of 48 is a reasonable choice." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 130, + 582, + 480, + 640 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 582, + 480, + 640 + ], + "spans": [ + { + "bbox": [ + 130, + 582, + 480, + 640 + ], + "type": "text", + "content": "Input of QSM. Table 7c studies the influence of different inputs of QSM. We can see that both the trainable prompts and the output from the last block are indispensable for QSM to generate high-quality task-specific query. The trainable prompts improve the capacity of QSM, and the results from the previous block ensure that knowledge is retained during the forward process." + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 130, + 641, + 480, + 666 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 641, + 480, + 666 + ], + "spans": [ + { + "bbox": [ + 130, + 641, + 480, + 666 + ], + "type": "text", + "content": "Hierarchical features. In SynQT, we extract three different types of features from KSMs for classification: final output of KSM " + }, + { + "bbox": [ + 130, + 641, + 480, + 666 + ], + "type": "inline_equation", + "content": "H_{i}" + }, + { + "bbox": [ + 130, + 641, + 480, + 666 + ], + "type": "text", + "content": ", attention features " + }, + { + "bbox": [ + 130, + 641, + 480, + 666 + ], + "type": "inline_equation", + "content": "F_{i}^{attn}" + } + ] + } + ], + "index": 13 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "spans": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "type": "text", + "content": "12" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 166, + 91, + 299, + 102 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 166, + 91, + 299, + 102 + ], + "spans": [ + { + "bbox": [ + 166, + 91, + 299, + 102 + ], + "type": "text", + "content": "Taolin Zhang, Jiawang Bai et al." + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 11 + }, + { + "para_blocks": [ + { + "type": "table", + "bbox": [ + 146, + 149, + 302, + 213 + ], + "blocks": [ + { + "bbox": [ + 132, + 114, + 480, + 138 + ], + "lines": [ + { + "bbox": [ + 132, + 114, + 480, + 138 + ], + "spans": [ + { + "bbox": [ + 132, + 114, + 480, + 138 + ], + "type": "text", + "content": "Table 7: SynQT ablation experiments with ViT-B/16 with ImageNet-21K backbone. Default settings are marked in gray." + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 146, + 149, + 302, + 213 + ], + "lines": [ + { + "bbox": [ + 146, + 149, + 302, + 213 + ], + "spans": [ + { + "bbox": [ + 146, + 149, + 302, + 213 + ], + "type": "table", + "html": "
DimParamNat.Spe.Str.Avg.
162.14 M77.783.855.872.4
322.43 M77.783.955.372.3
482.73 M78.084.456.272.9
643.02 M77.684.456.172.7
963.61 M77.784.555.972.7
", + "image_path": "4e9befeff2f130fb258709d23110ffbe4da992c8a7571abb0c86f2b4a6a1320b.jpg" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "table_body" + }, + { + "bbox": [ + 139, + 214, + 304, + 232 + ], + "lines": [ + { + "bbox": [ + 139, + 214, + 304, + 232 + ], + "spans": [ + { + "bbox": [ + 139, + 214, + 304, + 232 + ], + "type": "text", + "content": "(a) Hidden size in QSM. 48 as the width is a trade-off between parameters and performance." + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "table_footnote" + } + ], + "index": 3 + }, + { + "type": "table", + "bbox": [ + 317, + 152, + 474, + 205 + ], + "blocks": [ + { + "bbox": [ + 317, + 152, + 474, + 205 + ], + "lines": [ + { + "bbox": [ + 317, + 152, + 474, + 205 + ], + "spans": [ + { + "bbox": [ + 317, + 152, + 474, + 205 + ], + "type": "table", + "html": "
SynQTNat.Spe.Str.Avg.
w/o Hi78.384.053.872.0
w/o Fiattn78.383.555.472.4
w/o Fiuffn77.884.755.472.6
Full78.084.456.272.9
", + "image_path": "8512d5d0e5e489cb3be34631f18b5114f60ba4ee5b35bb97a7613d3e99d8c0b5.jpg" + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "table_body" + }, + { + "bbox": [ + 310, + 214, + 474, + 239 + ], + "lines": [ + { + "bbox": [ + 310, + 214, + 474, + 239 + ], + "spans": [ + { + "bbox": [ + 310, + 214, + 474, + 239 + ], + "type": "text", + "content": "(b) Feature selection. Hierarchical features contains rich information that benefit to downstream tasks." + } + ] + } + ], + "index": 6, + "angle": 0, + "type": "table_footnote" + } + ], + "index": 5 + }, + { + "type": "table", + "bbox": [ + 150, + 253, + 306, + 293 + ], + "blocks": [ + { + "bbox": [ + 150, + 253, + 306, + 293 + ], + "lines": [ + { + "bbox": [ + 150, + 253, + 306, + 293 + ], + "spans": [ + { + "bbox": [ + 150, + 253, + 306, + 293 + ], + "type": "table", + "html": "
SynQTNat.Spe.Str.Avg.
w/o Prompt77.282.951.970.7
w/o Last Output77.584.452.371.4
Full78.084.456.272.9
", + "image_path": "2f785034f7e0390cf87005066c0654e9c8de16dfe0992096bfec8926699b9cdb.jpg" + } + ] + } + ], + "index": 7, + "angle": 0, + "type": "table_body" + }, + { + "bbox": [ + 143, + 295, + 307, + 321 + ], + "lines": [ + { + "bbox": [ + 143, + 295, + 307, + 321 + ], + "spans": [ + { + "bbox": [ + 143, + 295, + 307, + 321 + ], + "type": "text", + "content": "(c) Input of QSM. Both the trainable prompt and the output from last block are important for generating task-specific query." + } + ] + } + ], + "index": 8, + "angle": 0, + "type": "table_footnote" + } + ], + "index": 7 + }, + { + "type": "table", + "bbox": [ + 321, + 251, + 477, + 293 + ], + "blocks": [ + { + "bbox": [ + 321, + 251, + 477, + 293 + ], + "lines": [ + { + "bbox": [ + 321, + 251, + 477, + 293 + ], + "spans": [ + { + "bbox": [ + 321, + 251, + 477, + 293 + ], + "type": "table", + "html": "
ProjectionNat.Spe.Str.Avg.
w/o Proj.78.483.855.072.4
Shared78.084.456.272.9
Independent77.984.455.572.6
", + "image_path": "764bbe78b3bcda56925adc4aa2b5fe97cb03c573bc2efea4a529ca30ab6191fb.jpg" + } + ] + } + ], + "index": 9, + "angle": 0, + "type": "table_body" + }, + { + "bbox": [ + 314, + 295, + 479, + 320 + ], + "lines": [ + { + "bbox": [ + 314, + 295, + 479, + 320 + ], + "spans": [ + { + "bbox": [ + 314, + 295, + 479, + 320 + ], + "type": "text", + "content": "(d) Projection in the classification head. Shared projection works the best for downstream tasks." + } + ] + } + ], + "index": 10, + "angle": 0, + "type": "table_footnote" + } + ], + "index": 9 + }, + { + "type": "table", + "bbox": [ + 146, + 331, + 302, + 373 + ], + "blocks": [ + { + "bbox": [ + 146, + 331, + 302, + 373 + ], + "lines": [ + { + "bbox": [ + 146, + 331, + 302, + 373 + ], + "spans": [ + { + "bbox": [ + 146, + 331, + 302, + 373 + ], + "type": "table", + "html": "
SynQTNat.Spe.Str.Avg.
Simply Avg.77.984.655.672.7
Fixed Weights77.984.755.572.7
Conditional78.084.456.272.9
", + "image_path": "90071a74be65f7733bd1d1558b5d755ec36649d59f6274e9a803ccc1e0da1dde.jpg" + } + ] + } + ], + "index": 11, + "angle": 0, + "type": "table_body" + }, + { + "bbox": [ + 139, + 375, + 303, + 400 + ], + "lines": [ + { + "bbox": [ + 139, + 375, + 303, + 400 + ], + "spans": [ + { + "bbox": [ + 139, + 375, + 303, + 400 + ], + "type": "text", + "content": "(e) Feature aggregation. Conditional feature aggregation is more effective than mean and learnable but fixed weight." + } + ] + } + ], + "index": 12, + "angle": 0, + "type": "table_footnote" + } + ], + "index": 11 + }, + { + "type": "table", + "bbox": [ + 317, + 332, + 473, + 364 + ], + "blocks": [ + { + "bbox": [ + 317, + 332, + 473, + 364 + ], + "lines": [ + { + "bbox": [ + 317, + 332, + 473, + 364 + ], + "spans": [ + { + "bbox": [ + 317, + 332, + 473, + 364 + ], + "type": "table", + "html": "
SynQTNat.Spe.Str.Avg.
w/o DropFeat77.784.355.272.4
w/ DropFeat78.084.456.272.9
", + "image_path": "58e0203110fc1e969abde82639d021f5aeaae51845775de81c66cb23be7ea670.jpg" + } + ] + } + ], + "index": 13, + "angle": 0, + "type": "table_body" + }, + { + "bbox": [ + 310, + 374, + 474, + 399 + ], + "lines": [ + { + "bbox": [ + 310, + 374, + 474, + 399 + ], + "spans": [ + { + "bbox": [ + 310, + 374, + 474, + 399 + ], + "type": "text", + "content": "(f) DropFeat in the classification head. The DropFeat module in the conditional head makes model more robust." + } + ] + } + ], + "index": 14, + "angle": 0, + "type": "table_footnote" + } + ], + "index": 13 + }, + { + "bbox": [ + 130, + 434, + 480, + 483 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 434, + 480, + 483 + ], + "spans": [ + { + "bbox": [ + 130, + 434, + 480, + 483 + ], + "type": "text", + "content": "and FFN features " + }, + { + "bbox": [ + 130, + 434, + 480, + 483 + ], + "type": "inline_equation", + "content": "F_{i}^{ffn}" + }, + { + "bbox": [ + 130, + 434, + 480, + 483 + ], + "type": "text", + "content": ", where " + }, + { + "bbox": [ + 130, + 434, + 480, + 483 + ], + "type": "inline_equation", + "content": "i \\in \\{1,2,\\dots,l\\}" + }, + { + "bbox": [ + 130, + 434, + 480, + 483 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 130, + 434, + 480, + 483 + ], + "type": "inline_equation", + "content": "l" + }, + { + "bbox": [ + 130, + 434, + 480, + 483 + ], + "type": "text", + "content": " is the number of layers. All these hierarchical features contain rich information and contribute to the performance boost of SynQT, as demonstrated in Table 7b. In particular, " + }, + { + "bbox": [ + 130, + 434, + 480, + 483 + ], + "type": "inline_equation", + "content": "H_{i}" + }, + { + "bbox": [ + 130, + 434, + 480, + 483 + ], + "type": "text", + "content": " plays a more important role than other types of features." + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 130, + 484, + 480, + 568 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 484, + 480, + 568 + ], + "spans": [ + { + "bbox": [ + 130, + 484, + 480, + 568 + ], + "type": "text", + "content": "Projection in the classification head. We vary the projection architecture in the classification head and investigate its performance in Table 7d. A projection of the input is necessary for the classification head, as hierarchical features across various layers may be misaligned in the representation space. However, to our surprise, a shared bottleneck structure is enough to handle these hierarchical features. The reason may be that independent projection for distinct features leads to overfitting, subsequently causing a slight drop in performance." + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 130, + 569, + 480, + 640 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 569, + 480, + 640 + ], + "spans": [ + { + "bbox": [ + 130, + 569, + 480, + 640 + ], + "type": "text", + "content": "Conditional feature aggregation. In the classification head, we aggregate features using the weights conditioned on " + }, + { + "bbox": [ + 130, + 569, + 480, + 640 + ], + "type": "inline_equation", + "content": "H_{l}" + }, + { + "bbox": [ + 130, + 569, + 480, + 640 + ], + "type": "text", + "content": ". We further conduct experiments over variants of SynQT with different aggregation methods including a simple average and weighted-sum with learnable but fixed weights. The results in Table 7e show that conditional aggregation improves the performance of SynQT especially for the Structured group." + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 130, + 641, + 480, + 665 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 641, + 480, + 665 + ], + "spans": [ + { + "bbox": [ + 130, + 641, + 480, + 665 + ], + "type": "text", + "content": "DropFeat in the classification head. We study the effect of DropFeat, which is proposed to prevent overfitting by randomly dropping features in the classi-" + } + ] + } + ], + "index": 18 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 149, + 91, + 448, + 102 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 149, + 91, + 448, + 102 + ], + "spans": [ + { + "bbox": [ + 149, + 91, + 448, + 102 + ], + "type": "text", + "content": "Parameter-Efficient and Memory-Efficient Tuning for Vision Transformer" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 470, + 91, + 481, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 470, + 91, + 481, + 100 + ], + "spans": [ + { + "bbox": [ + 470, + 91, + 481, + 100 + ], + "type": "text", + "content": "13" + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 12 + }, + { + "para_blocks": [ + { + "bbox": [ + 130, + 116, + 479, + 140 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 116, + 479, + 140 + ], + "spans": [ + { + "bbox": [ + 130, + 116, + 479, + 140 + ], + "type": "text", + "content": "fication head. As can be observed from Table 7f, SynQT performs worse without DropFeat, indicating that DropFeat enhances the model's generalization" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 130, + 140, + 298, + 235 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 140, + 298, + 235 + ], + "spans": [ + { + "bbox": [ + 130, + 140, + 298, + 235 + ], + "type": "text", + "content": "in handling rich hierarchical feature input. DropFeat ensures that the model does not overly rely on a specific feature during training and encourages the model to focus on a general way to select beneficial information from these features for downstream tasks." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 132, + 251, + 261, + 262 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 251, + 261, + 262 + ], + "spans": [ + { + "bbox": [ + 132, + 251, + 261, + 262 + ], + "type": "text", + "content": "4.7 t-SNE Visualization" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 130, + 268, + 298, + 351 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 268, + 298, + 351 + ], + "spans": [ + { + "bbox": [ + 130, + 268, + 298, + 351 + ], + "type": "text", + "content": "We present t-SNE visualization [35] of the CLS Token in pre-trained ViT, VQT, and our SynQT on EuroSAT, SVHN, and Clevr-Dist in Figure 5. The visualization shows that our SynQT achieves better feature clustering results compared to the" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 130, + 351, + 462, + 364 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 351, + 462, + 364 + ], + "spans": [ + { + "bbox": [ + 130, + 351, + 462, + 364 + ], + "type": "text", + "content": "original CLS Token and VQT. We provide more visualization in Appendix." + } + ] + } + ], + "index": 6 + }, + { + "type": "image", + "bbox": [ + 312, + 163, + 478, + 299 + ], + "blocks": [ + { + "bbox": [ + 312, + 163, + 478, + 299 + ], + "lines": [ + { + "bbox": [ + 312, + 163, + 478, + 299 + ], + "spans": [ + { + "bbox": [ + 312, + 163, + 478, + 299 + ], + "type": "image", + "image_path": "a43dc0aa692ea145d2920a1b59c03454706cd6cde26a532bb8f6a9c86182c108.jpg" + } + ] + } + ], + "index": 7, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 304, + 308, + 482, + 353 + ], + "lines": [ + { + "bbox": [ + 304, + 308, + 482, + 353 + ], + "spans": [ + { + "bbox": [ + 304, + 308, + 482, + 353 + ], + "type": "text", + "content": "Fig. 5: t-SNE visualization on SVHN, EuroSAT, and Clevr-Dist, one from each category of VTAB-1K. The features of our SynQT are more separable." + } + ] + } + ], + "index": 8, + "angle": 0, + "type": "image_caption" + } + ], + "index": 7 + }, + { + "bbox": [ + 132, + 380, + 400, + 392 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 380, + 400, + 392 + ], + "spans": [ + { + "bbox": [ + 132, + 380, + 400, + 392 + ], + "type": "text", + "content": "5 Conclusion, Limitation, and Future Work" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 130, + 402, + 482, + 521 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 402, + 482, + 521 + ], + "spans": [ + { + "bbox": [ + 130, + 402, + 482, + 521 + ], + "type": "text", + "content": "In this work, we present an insightful analysis of the high memory cost of existing PETL methods from a unique perspective, i.e., they entangle task-specific learning and pre-trained knowledge utilization together. Inspired by the idea of disentanglement, we propose a new paradigm, namely SynQT, consisting of two key structures: Query Synthesis Module (QSM) and Knowledge Extraction Module (KEM). QSM synthesizes task-specific query and then KEM extracts pre-trained knowledge in a query-only manner. Finally, a classification head is proposed to handle rich hierarchical features from KEMs and make predictions. Extensive experiments on VTAB-1K and FGVC demonstrate the effectiveness of SynQT under memory constraints and in an extremely low-data regime." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 130, + 522, + 482, + 605 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 522, + 482, + 605 + ], + "spans": [ + { + "bbox": [ + 130, + 522, + 482, + 605 + ], + "type": "text", + "content": "Despite the promising performance of our SynQT, it also has some limitations. One limitation is that SynQT requires slightly more trainable parameters than previous PETL methods due to a few projection layers and FFN operations, which is analyzed in Appendix. However, we would like to emphasize that as shown in their original papers, for VPT-Deep [17] and VQT [45], straightforwardly increasing the number of tokens to use more trainable parameters can not consistently result in better performance." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 130, + 606, + 482, + 665 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 606, + 482, + 665 + ], + "spans": [ + { + "bbox": [ + 130, + 606, + 482, + 665 + ], + "type": "text", + "content": "We believe that our work will inspire the community to explore memory-efficient PETL along with our disentangled perspective. Besides, to overcome the aforementioned limitation, one future direction is to develop PETL methods that achieve high accuracy and meanwhile have extreme memory efficiency and parameter efficiency." + } + ] + } + ], + "index": 12 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "spans": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "type": "text", + "content": "14" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 166, + 91, + 299, + 102 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 166, + 91, + 299, + 102 + ], + "spans": [ + { + "bbox": [ + 166, + 91, + 299, + 102 + ], + "type": "text", + "content": "Taolin Zhang, Jiawang Bai et al." + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 13 + }, + { + "para_blocks": [ + { + "bbox": [ + 133, + 115, + 240, + 129 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 115, + 240, + 129 + ], + "spans": [ + { + "bbox": [ + 133, + 115, + 240, + 129 + ], + "type": "text", + "content": "Acknowledgments" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 130, + 139, + 482, + 199 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 139, + 482, + 199 + ], + "spans": [ + { + "bbox": [ + 130, + 139, + 482, + 199 + ], + "type": "text", + "content": "This work is supported in part by the National Natural Science Foundation of China under Grant 62171248, Shenzhen Science and Technology Program (JCYJ20220818101012025), the PCNL KEY project (PCL2023AS6-1), and the National Research Foundation, Singapore, under its AI Singapore Programme (AISG Award No: AISG2-RP-2021-023)." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 133, + 217, + 197, + 228 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 217, + 197, + 228 + ], + "spans": [ + { + "bbox": [ + 133, + 217, + 197, + 228 + ], + "type": "text", + "content": "References" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 134, + 239, + 480, + 665 + ], + "type": "list", + "angle": 0, + "index": 18, + "blocks": [ + { + "bbox": [ + 138, + 239, + 480, + 262 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 239, + 480, + 262 + ], + "spans": [ + { + "bbox": [ + 138, + 239, + 480, + 262 + ], + "type": "text", + "content": "1. Argyriou, A., Evgeniou, T., Pontil, M.: Multi-task feature learning. Advances in neural information processing systems 19 (2006)" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 138, + 263, + 480, + 283 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 263, + 480, + 283 + ], + "spans": [ + { + "bbox": [ + 138, + 263, + 480, + 283 + ], + "type": "text", + "content": "2. Ba, J.L., Kiros, J.R., Hinton, G.E.: Layer normalization. arXiv preprint arXiv:1607.06450 (2016)" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 138, + 285, + 480, + 316 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 285, + 480, + 316 + ], + "spans": [ + { + "bbox": [ + 138, + 285, + 480, + 316 + ], + "type": "text", + "content": "3. Bai, J., Gao, K., Min, S., Xia, S.T., Li, Z., Liu, W.: Badclip: Trigger-aware prompt learning for backdoor attacks on clip. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 24239-24250 (2024)" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 138, + 318, + 480, + 349 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 318, + 480, + 349 + ], + "spans": [ + { + "bbox": [ + 138, + 318, + 480, + 349 + ], + "type": "text", + "content": "4. Bai, J., Yuan, L., Xia, S.T., Yan, S., Li, Z., Liu, W.: Improving vision transformers by revisiting high-frequency components. In: European Conference on Computer Vision. pp. 1-18. Springer (2022)" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 138, + 350, + 480, + 392 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 350, + 480, + 392 + ], + "spans": [ + { + "bbox": [ + 138, + 350, + 480, + 392 + ], + "type": "text", + "content": "5. Bossard, L., Guillaumeun, M., Van Gool, L.: Food-101-mining discriminative components with random forests. In: Computer Vision-ECCV 2014: 13th European Conference, Zurich, Switzerland, September 6-12, 2014, Proceedings, Part VI 13. pp. 446-461. Springer (2014)" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 138, + 394, + 480, + 425 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 394, + 480, + 425 + ], + "spans": [ + { + "bbox": [ + 138, + 394, + 480, + 425 + ], + "type": "text", + "content": "6. Brown, T., Mann, B., Ryder, N., Subbiah, M., Kaplan, J.D., Dhariwal, P., Neelakantan, A., Shyam, P., Sastry, G., Askell, A., et al.: Language models are few-shot learners. NeurIPS (2020)" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 138, + 426, + 480, + 458 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 426, + 480, + 458 + ], + "spans": [ + { + "bbox": [ + 138, + 426, + 480, + 458 + ], + "type": "text", + "content": "7. Caron, M., Touvron, H., Misra, I., Jégou, H., Mairal, J., Bojanowski, P., Joulin, A.: Emerging properties in self-supervised vision transformers. In: Proceedings of the IEEE/CVF international conference on computer vision. pp. 9650-9660 (2021)" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 138, + 460, + 480, + 491 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 460, + 480, + 491 + ], + "spans": [ + { + "bbox": [ + 138, + 460, + 480, + 491 + ], + "type": "text", + "content": "8. Chen, S., Ge, C., Tong, Z., Wang, J., Song, Y., Wang, J., Luo, P.: Adaptformer: Adapting vision transformers for scalable visual recognition. Advances in Neural Information Processing Systems 35, 16664-16678 (2022)" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 138, + 491, + 480, + 523 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 491, + 480, + 523 + ], + "spans": [ + { + "bbox": [ + 138, + 491, + 480, + 523 + ], + "type": "text", + "content": "9. Deng, J., Dong, W., Socher, R., Li, L.J., Li, K., Fei-Fei, L.: Imagenet: A large-scale hierarchical image database. In: 2009 IEEE conference on computer vision and pattern recognition. pp. 248-255. IEEE (2009)" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 134, + 525, + 480, + 567 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 134, + 525, + 480, + 567 + ], + "spans": [ + { + "bbox": [ + 134, + 525, + 480, + 567 + ], + "type": "text", + "content": "10. Dosovitskiy, A., Beyer, L., Kolesnikov, A., Weissenborn, D., Zhai, X., Unterthiner, T., Dehghani, M., Minderer, M., Heigold, G., Gelly, S., et al.: An image is worth 16x16 words: Transformers for image recognition at scale. arXiv preprint arXiv:2010.11929 (2020)" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 134, + 568, + 480, + 599 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 134, + 568, + 480, + 599 + ], + "spans": [ + { + "bbox": [ + 134, + 568, + 480, + 599 + ], + "type": "text", + "content": "11. Evci, U., Dumoulin, V., Larochelle, H., Mozer, M.C.: Head2toe: Utilizing intermediate representations for better transfer learning. In: International Conference on Machine Learning. pp. 6009-6033. PMLR (2022)" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 134, + 601, + 480, + 632 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 134, + 601, + 480, + 632 + ], + "spans": [ + { + "bbox": [ + 134, + 601, + 480, + 632 + ], + "type": "text", + "content": "12. Guo, H., Dai, T., Bai, Y., Chen, B., Xia, S.T., Zhu, Z.: Adaptir: Parameter efficient multi-task adaptation for pre-trained image restoration models. arXiv preprint arXiv:2312.08881 (2023)" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 134, + 633, + 480, + 665 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 134, + 633, + 480, + 665 + ], + "spans": [ + { + "bbox": [ + 134, + 633, + 480, + 665 + ], + "type": "text", + "content": "13. Guo, H., Li, J., Dai, T., Ouyang, Z., Ren, X., Xia, S.T.: Mambair: A simple baseline for image restoration with state-space model. arXiv preprint arXiv:2402.15648 (2024)" + } + ] + } + ], + "index": 17 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 149, + 91, + 447, + 102 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 149, + 91, + 447, + 102 + ], + "spans": [ + { + "bbox": [ + 149, + 91, + 447, + 102 + ], + "type": "text", + "content": "Parameter-Efficient and Memory-Efficient Tuning for Vision Transformer" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 470, + 92, + 480, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 470, + 92, + 480, + 100 + ], + "spans": [ + { + "bbox": [ + 470, + 92, + 480, + 100 + ], + "type": "text", + "content": "15" + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 14 + }, + { + "para_blocks": [ + { + "bbox": [ + 132, + 116, + 481, + 665 + ], + "type": "list", + "angle": 0, + "index": 20, + "blocks": [ + { + "bbox": [ + 133, + 116, + 481, + 149 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 116, + 481, + 149 + ], + "spans": [ + { + "bbox": [ + 133, + 116, + 481, + 149 + ], + "type": "text", + "content": "14. He, K., Chen, X., Xie, S., Li, Y., Dollár, P., Girshick, R.: Masked autoencoders are scalable vision learners. In: Proceedings of the IEEE/CVF conference on computer vision and pattern recognition. pp. 16000-16009 (2022)" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 133, + 150, + 481, + 183 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 150, + 481, + 183 + ], + "spans": [ + { + "bbox": [ + 133, + 150, + 481, + 183 + ], + "type": "text", + "content": "15. Houlsby, N., Giurgiu, A., Jastrzebski, S., Morrone, B., De Laroussilhe, Q., Gesmundo, A., Attariyan, M., Gelly, S.: Parameter-efficient transfer learning for nlp. In: International Conference on Machine Learning. pp. 2790-2799. PMLR (2019)" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 133, + 183, + 481, + 215 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 183, + 481, + 215 + ], + "spans": [ + { + "bbox": [ + 133, + 183, + 481, + 215 + ], + "type": "text", + "content": "16. Hu, E.J., Wallis, P., Allen-Zhu, Z., Li, Y., Wang, S., Wang, L., Chen, W., et al.: Lora: Low-rank adaptation of large language models. In: International Conference on Learning Representations (2021)" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 133, + 216, + 481, + 248 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 216, + 481, + 248 + ], + "spans": [ + { + "bbox": [ + 133, + 216, + 481, + 248 + ], + "type": "text", + "content": "17. Jia, M., Tang, L., Chen, B.C., Cardie, C., Belongie, S., Hariharan, B., Lim, S.N.: Visual prompt tuning. In: European Conference on Computer Vision. pp. 709-727. Springer (2022)" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 133, + 249, + 481, + 281 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 249, + 481, + 281 + ], + "spans": [ + { + "bbox": [ + 133, + 249, + 481, + 281 + ], + "type": "text", + "content": "18. Jie, S., Deng, Z.H.: Fact: Factor-tuning for lightweight adaptation on vision transformer. In: Proceedings of the AAAI Conference on Artificial Intelligence. vol. 37, pp. 1060-1068 (2023)" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 133, + 281, + 481, + 315 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 281, + 481, + 315 + ], + "spans": [ + { + "bbox": [ + 133, + 281, + 481, + 315 + ], + "type": "text", + "content": "19. Jie, S., Wang, H., Deng, Z.H.: Revisiting the parameter efficiency of adapters from the perspective of precision redundancy. In: Proceedings of the IEEE/CVF International Conference on Computer Vision. pp. 17217-17226 (2023)" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 132, + 315, + 481, + 346 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 315, + 481, + 346 + ], + "spans": [ + { + "bbox": [ + 132, + 315, + 481, + 346 + ], + "type": "text", + "content": "20. Kirillov, A., Mintun, E., Ravi, N., Mao, H., Rolland, C., Gustafson, L., Xiao, T., Whitehead, S., Berg, A.C., Lo, W.Y., et al.: Segment anything. arXiv preprint arXiv:2304.02643 (2023)" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 132, + 347, + 481, + 380 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 347, + 481, + 380 + ], + "spans": [ + { + "bbox": [ + 132, + 347, + 481, + 380 + ], + "type": "text", + "content": "21. Kornblith, S., Shlens, J., Le, Q.V.: Do better imagenet models transfer better? In: Proceedings of the IEEE/CVF conference on computer vision and pattern recognition. pp. 2661-2671 (2019)" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 132, + 380, + 481, + 413 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 380, + 481, + 413 + ], + "spans": [ + { + "bbox": [ + 132, + 380, + 481, + 413 + ], + "type": "text", + "content": "22. Krause, J., Stark, M., Deng, J., Fei-Fei, L.: 3d object representations for fine-grained categorization. In: Proceedings of the IEEE international conference on computer vision workshops. pp. 554-561 (2013)" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 132, + 413, + 481, + 434 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 413, + 481, + 434 + ], + "spans": [ + { + "bbox": [ + 132, + 413, + 481, + 434 + ], + "type": "text", + "content": "23. Larsson, G., Maire, M., Shakhnarovich, G.: Fractalnet: Ultra-deep neural networks without residuals. In: ICLR (2017)" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 132, + 434, + 481, + 456 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 434, + 481, + 456 + ], + "spans": [ + { + "bbox": [ + 132, + 434, + 481, + 456 + ], + "type": "text", + "content": "24. Lester, B., Al-Rfou, R., Constant, N.: The power of scale for parameter-efficient prompt tuning. arXiv preprint arXiv:2104.08691 (2021)" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 132, + 456, + 481, + 478 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 456, + 481, + 478 + ], + "spans": [ + { + "bbox": [ + 132, + 456, + 481, + 478 + ], + "type": "text", + "content": "25. Li, X.L., Liang, P.: Prefix-tuning: Optimizing continuous prompts for generation. arXiv preprint arXiv:2101.00190 (2021)" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 132, + 479, + 481, + 511 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 479, + 481, + 511 + ], + "spans": [ + { + "bbox": [ + 132, + 479, + 481, + 511 + ], + "type": "text", + "content": "26. Li, X., Lian, D., Lu, Z., Bai, J., Chen, Z., Wang, X.: Graphadapter: Tuning vision-language models with dual knowledge graph. Advances in Neural Information Processing Systems 36 (2024)" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 132, + 511, + 481, + 544 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 511, + 481, + 544 + ], + "spans": [ + { + "bbox": [ + 132, + 511, + 481, + 544 + ], + "type": "text", + "content": "27. Lian, D., Zhou, D., Feng, J., Wang, X.: Scaling & shifting your features: A new baseline for efficient model tuning. Advances in Neural Information Processing Systems 35, 109-123 (2022)" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 132, + 544, + 481, + 577 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 544, + 481, + 577 + ], + "spans": [ + { + "bbox": [ + 132, + 544, + 481, + 577 + ], + "type": "text", + "content": "28. Liu, H., Tam, D., Muqeeth, M., Mohta, J., Huang, T., Bansal, M., Raffel, C.A.: Few-shot parameter-efficient fine-tuning is better and cheaper than in-context learning. NeurIPS (2022)" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 132, + 577, + 481, + 610 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 577, + 481, + 610 + ], + "spans": [ + { + "bbox": [ + 132, + 577, + 481, + 610 + ], + "type": "text", + "content": "29. Liu, P., Yuan, W., Fu, J., Jiang, Z., Hayashi, H., Neubig, G.: Pre-train, prompt, and predict: A systematic survey of prompting methods in natural language processing. ACM Computing Surveys 55(9), 1-35 (2023)" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 132, + 610, + 481, + 643 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 610, + 481, + 643 + ], + "spans": [ + { + "bbox": [ + 132, + 610, + 481, + 643 + ], + "type": "text", + "content": "30. Liu, X., Ji, K., Fu, Y., Tam, W.L., Du, Z., Yang, Z., Tang, J.: P-tuning v2: Prompt tuning can be comparable to fine-tuning universally across scales and tasks. arXiv preprint arXiv:2110.07602 (2021)" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 132, + 643, + 481, + 665 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 643, + 481, + 665 + ], + "spans": [ + { + "bbox": [ + 132, + 643, + 481, + 665 + ], + "type": "text", + "content": "31. Liu, Z., Lin, Y., Cao, Y., Hu, H., Wei, Y., Zhang, Z., Lin, S., Guo, B.: Swin transformer: Hierarchical vision transformer using shifted windows. In: Proceedings" + } + ] + } + ], + "index": 19 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "spans": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "type": "text", + "content": "16" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 166, + 91, + 299, + 102 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 166, + 91, + 299, + 102 + ], + "spans": [ + { + "bbox": [ + 166, + 91, + 299, + 102 + ], + "type": "text", + "content": "Taolin Zhang, Jiawang Bai et al." + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 15 + }, + { + "para_blocks": [ + { + "bbox": [ + 132, + 116, + 481, + 665 + ], + "type": "list", + "angle": 0, + "index": 21, + "blocks": [ + { + "bbox": [ + 149, + 116, + 480, + 138 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 149, + 116, + 480, + 138 + ], + "spans": [ + { + "bbox": [ + 149, + 116, + 480, + 138 + ], + "type": "text", + "content": "of the IEEE/CVF international conference on computer vision. pp. 10012-10022 (2021)" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 133, + 139, + 481, + 171 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 139, + 481, + 171 + ], + "spans": [ + { + "bbox": [ + 133, + 139, + 481, + 171 + ], + "type": "text", + "content": "32. Liu, Z., Mao, H., Wu, C.Y., Feichtenhofer, C., Darrell, T., Xie, S.: A convnet for the 2020s. In: Proceedings of the IEEE/CVF conference on computer vision and pattern recognition. pp. 11976-11986 (2022)" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 132, + 172, + 481, + 193 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 172, + 481, + 193 + ], + "spans": [ + { + "bbox": [ + 132, + 172, + 481, + 193 + ], + "type": "text", + "content": "33. Loshchilov, I., Hutter, F.: Decoupled weight decay regularization. arXiv preprint arXiv:1711.05101 (2017)" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 132, + 194, + 481, + 226 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 194, + 481, + 226 + ], + "spans": [ + { + "bbox": [ + 132, + 194, + 481, + 226 + ], + "type": "text", + "content": "34. Lu, Z., Bai, J., Li, X., Xiao, Z., Wang, X.: Beyond sole strength: Customized ensembles for generalized vision-language models. In: Forty-first International Conference on Machine Learning (2024)" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 132, + 227, + 481, + 248 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 227, + 481, + 248 + ], + "spans": [ + { + "bbox": [ + 132, + 227, + 481, + 248 + ], + "type": "text", + "content": "35. Van der Maaten, L., Hinton, G.: Visualizing data using t-sne. Journal of machine learning research 9(11) (2008)" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 132, + 249, + 481, + 270 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 249, + 481, + 270 + ], + "spans": [ + { + "bbox": [ + 132, + 249, + 481, + 270 + ], + "type": "text", + "content": "36. Maji, S., Rahtu, E., Kannala, J., Blaschko, M., Vedaldi, A.: Fine-grained visual classification of aircraft. arXiv preprint arXiv:1306.5151 (2013)" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 132, + 271, + 481, + 303 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 271, + 481, + 303 + ], + "spans": [ + { + "bbox": [ + 132, + 271, + 481, + 303 + ], + "type": "text", + "content": "37. Nilsback, M.E., Zisserman, A.: A visual vocabulary for flower classification. In: 2006 IEEE computer society conference on computer vision and pattern recognition (CVPR'06). vol. 2, pp. 1447-1454. IEEE (2006)" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 132, + 304, + 481, + 335 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 304, + 481, + 335 + ], + "spans": [ + { + "bbox": [ + 132, + 304, + 481, + 335 + ], + "type": "text", + "content": "38. Parkhi, O.M., Vedaldi, A., Zisserman, A., Jawahar, C.: Cats and dogs. In: 2012 IEEE conference on computer vision and pattern recognition. pp. 3498-3505. IEEE (2012)" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 132, + 336, + 481, + 369 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 336, + 481, + 369 + ], + "spans": [ + { + "bbox": [ + 132, + 336, + 481, + 369 + ], + "type": "text", + "content": "39. Pfeiffer, J., Kamath, A., Rückle, A., Cho, K., Gurevych, I.: Adapterfusion: Nondestructive task composition for transfer learning. arXiv preprint arXiv:2005.00247 (2020)" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 132, + 369, + 481, + 391 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 369, + 481, + 391 + ], + "spans": [ + { + "bbox": [ + 132, + 369, + 481, + 391 + ], + "type": "text", + "content": "40. Rebuffi, S.A., Bilen, H., Vedaldi, A.: Learning multiple visual domains with residual adapters. Advances in neural information processing systems 30 (2017)" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 132, + 392, + 481, + 413 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 392, + 481, + 413 + ], + "spans": [ + { + "bbox": [ + 132, + 392, + 481, + 413 + ], + "type": "text", + "content": "41. Shao, L., Zhu, F., Li, X.: Transfer learning for visual categorization: A survey. IEEE transactions on neural networks and learning systems 26(5), 1019-1034 (2014)" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 132, + 414, + 481, + 445 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 414, + 481, + 445 + ], + "spans": [ + { + "bbox": [ + 132, + 414, + 481, + 445 + ], + "type": "text", + "content": "42. Srivastava, N., Hinton, G., Krizhevsky, A., Sutskever, I., Salakhutdinov, R.: Dropout: a simple way to prevent neural networks from overfitting. The journal of machine learning research 15(1), 1929-1958 (2014)" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 132, + 447, + 481, + 478 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 447, + 481, + 478 + ], + "spans": [ + { + "bbox": [ + 132, + 447, + 481, + 478 + ], + "type": "text", + "content": "43. Sung, Y.L., Cho, J., Bansal, M.: Lst: Ladder side-tuning for parameter and memory efficient transfer learning. Advances in Neural Information Processing Systems 35, 12991-13005 (2022)" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 132, + 479, + 481, + 511 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 479, + 481, + 511 + ], + "spans": [ + { + "bbox": [ + 132, + 479, + 481, + 511 + ], + "type": "text", + "content": "44. Sung, Y.L., Cho, J., Bansal, M.: Vl-adapter: Parameter-efficient transfer learning for vision-and-language tasks. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 5227-5237 (2022)" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 132, + 512, + 481, + 555 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 512, + 481, + 555 + ], + "spans": [ + { + "bbox": [ + 132, + 512, + 481, + 555 + ], + "type": "text", + "content": "45. Tu, C.H., Mai, Z., Chao, W.L.: Visual query tuning: Towards effective usage of intermediate representations for parameter and memory efficient transfer learning. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 7725-7735 (2023)" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 132, + 555, + 481, + 587 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 555, + 481, + 587 + ], + "spans": [ + { + "bbox": [ + 132, + 555, + 481, + 587 + ], + "type": "text", + "content": "46. Vaswani, A., Shazeer, N., Parmar, N., Uszkoreit, J., Jones, L., Gomez, A.N., Kaiser, L., Polosukhin, I.: Attention is all you need. Advances in neural information processing systems 30 (2017)" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 132, + 588, + 481, + 621 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 588, + 481, + 621 + ], + "spans": [ + { + "bbox": [ + 132, + 588, + 481, + 621 + ], + "type": "text", + "content": "47. Wan, L., Zeiler, M., Zhang, S., Le Cun, Y., Fergus, R.: Regularization of neural networks using dropconnect. In: International conference on machine learning. pp. 1058-1066. PMLR (2013)" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 132, + 622, + 481, + 643 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 622, + 481, + 643 + ], + "spans": [ + { + "bbox": [ + 132, + 622, + 481, + 643 + ], + "type": "text", + "content": "48. Wang, M., Deng, W.: Deep visual domain adaptation: A survey. Neurocomputing 312, 135-153 (2018)" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 132, + 643, + 481, + 665 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 643, + 481, + 665 + ], + "spans": [ + { + "bbox": [ + 132, + 643, + 481, + 665 + ], + "type": "text", + "content": "49. Wang, W., Bao, H., Dong, L., Bjorck, J., Peng, Z., Liu, Q., Aggarwal, K., Mohammed, O.K., Singhal, S., Som, S., et al.: Image as a foreign language: Beit pre" + } + ] + } + ], + "index": 20 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 150, + 91, + 447, + 102 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 150, + 91, + 447, + 102 + ], + "spans": [ + { + "bbox": [ + 150, + 91, + 447, + 102 + ], + "type": "text", + "content": "Parameter-Efficient and Memory-Efficient Tuning for Vision Transformer" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 470, + 91, + 481, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 470, + 91, + 481, + 100 + ], + "spans": [ + { + "bbox": [ + 470, + 91, + 481, + 100 + ], + "type": "text", + "content": "17" + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 16 + }, + { + "para_blocks": [ + { + "bbox": [ + 132, + 116, + 481, + 512 + ], + "type": "list", + "angle": 0, + "index": 16, + "blocks": [ + { + "bbox": [ + 147, + 116, + 481, + 138 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 147, + 116, + 481, + 138 + ], + "spans": [ + { + "bbox": [ + 147, + 116, + 481, + 138 + ], + "type": "text", + "content": "training for all vision and vision-language tasks. arXiv preprint arXiv:2208.10442 (2022)" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 133, + 139, + 481, + 182 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 139, + 481, + 182 + ], + "spans": [ + { + "bbox": [ + 133, + 139, + 481, + 182 + ], + "type": "text", + "content": "50. Yang, S., Bai, J., Gao, K., Yang, Y., Li, Y., Xia, S.T.: Not all prompts are secure: A switchable backdoor attack against pre-trained vision transformers. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 24431-24441 (2024)" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 132, + 183, + 481, + 205 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 183, + 481, + 205 + ], + "spans": [ + { + "bbox": [ + 132, + 183, + 481, + 205 + ], + "type": "text", + "content": "51. Yang, X., Ye, J., Wang, X.: Factorizing knowledge in neural networks. In: European Conference on Computer Vision. pp. 73-91. Springer (2022)" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 132, + 205, + 481, + 226 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 205, + 481, + 226 + ], + "spans": [ + { + "bbox": [ + 132, + 205, + 481, + 226 + ], + "type": "text", + "content": "52. Yang, X., Zhou, D., Liu, S., Ye, J., Wang, X.: Deep model reassembly. Advances in neural information processing systems 35, 25739-25753 (2022)" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 132, + 228, + 481, + 258 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 228, + 481, + 258 + ], + "spans": [ + { + "bbox": [ + 132, + 228, + 481, + 258 + ], + "type": "text", + "content": "53. Yuan, M., Lin, Y.: Model selection and estimation in regression with grouped variables. Journal of the Royal Statistical Society Series B: Statistical Methodology 68(1), 49-67 (2006)" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 132, + 259, + 481, + 293 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 259, + 481, + 293 + ], + "spans": [ + { + "bbox": [ + 132, + 259, + 481, + 293 + ], + "type": "text", + "content": "54. Zaken, E.B., Ravfogel, S., Goldberg, Y.: Bitfit: Simple parameter-efficient fine-tuning for transformer-based masked language-models. arXiv preprint arXiv:2106.10199 (2021)" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 132, + 293, + 481, + 314 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 293, + 481, + 314 + ], + "spans": [ + { + "bbox": [ + 132, + 293, + 481, + 314 + ], + "type": "text", + "content": "55. Zang, Y., Li, W., Zhou, K., Huang, C., Loy, C.C.: Unified vision and language prompt learning. arXiv preprint arXiv:2210.07225 (2022)" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 132, + 315, + 481, + 347 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 315, + 481, + 347 + ], + "spans": [ + { + "bbox": [ + 132, + 315, + 481, + 347 + ], + "type": "text", + "content": "56. Zha, Y., Wang, J., Dai, T., Chen, B., Wang, Z., Xia, S.T.: Instance-aware dynamic prompt tuning for pre-trained point cloud models. In: Proceedings of the IEEE/CVF International Conference on Computer Vision. pp. 14161-14170 (2023)" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 132, + 347, + 481, + 380 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 347, + 481, + 380 + ], + "spans": [ + { + "bbox": [ + 132, + 347, + 481, + 380 + ], + "type": "text", + "content": "57. Zhai, X., Kolesnikov, A., Houlsby, N., Beyer, L.: Scaling vision transformers. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 12104-12113 (2022)" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 132, + 380, + 481, + 413 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 380, + 481, + 413 + ], + "spans": [ + { + "bbox": [ + 132, + 380, + 481, + 413 + ], + "type": "text", + "content": "58. Zhai, X., Puigcerver, J., Kolesnikov, A., Ruyssen, P., Riquelme, C., Lucic, M., Djolonga, J., Pinto, A.S., Neumann, M., Dosovitskiy, A., et al.: The visual task adaptation benchmark (2019)" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 132, + 413, + 416, + 424 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 413, + 416, + 424 + ], + "spans": [ + { + "bbox": [ + 132, + 413, + 416, + 424 + ], + "type": "text", + "content": "59. Zhang, Y., Zhou, K., Liu, Z.: Neural prompt search. arXiv (2022)" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 132, + 425, + 481, + 456 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 425, + 481, + 456 + ], + "spans": [ + { + "bbox": [ + 132, + 425, + 481, + 456 + ], + "type": "text", + "content": "60. Zhou, K., Yang, J., Loy, C.C., Liu, Z.: Conditional prompt learning for vision-language models. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 16816-16825 (2022)" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 132, + 457, + 481, + 479 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 457, + 481, + 479 + ], + "spans": [ + { + "bbox": [ + 132, + 457, + 481, + 479 + ], + "type": "text", + "content": "61. Zhou, K., Yang, J., Loy, C.C., Liu, Z.: Learning to prompt for vision-language models. International Journal of Computer Vision 130(9), 2337-2348 (2022)" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 132, + 479, + 481, + 512 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 479, + 481, + 512 + ], + "spans": [ + { + "bbox": [ + 132, + 479, + 481, + 512 + ], + "type": "text", + "content": "62. Zhu, L., Liao, B., Zhang, Q., Wang, X., Liu, W., Wang, X.: Vision mamba: Efficient visual representation learning with bidirectional state space model. arXiv preprint arXiv:2401.09417 (2024)" + } + ] + } + ], + "index": 15 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "spans": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "type": "text", + "content": "18" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 166, + 91, + 299, + 102 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 166, + 91, + 299, + 102 + ], + "spans": [ + { + "bbox": [ + 166, + 91, + 299, + 102 + ], + "type": "text", + "content": "Taolin Zhang, Jiawang Bai et al." + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 17 + } + ], + "_backend": "vlm", + "_version_name": "2.6.4" +} \ No newline at end of file diff --git a/2024/Parameterization-driven Neural Surface Reconstruction for Object-oriented Editing in Neural Rendering/20a75bf0-0a6a-4ea4-b1e9-a52604587383_content_list.json b/2024/Parameterization-driven Neural Surface Reconstruction for Object-oriented Editing in Neural Rendering/20a75bf0-0a6a-4ea4-b1e9-a52604587383_content_list.json new file mode 100644 index 0000000000000000000000000000000000000000..2d56278f6e6cc95fe7156bb17dd31102e1f73e9d --- /dev/null +++ b/2024/Parameterization-driven Neural Surface Reconstruction for Object-oriented Editing in Neural Rendering/20a75bf0-0a6a-4ea4-b1e9-a52604587383_content_list.json @@ -0,0 +1,2215 @@ +[ + { + "type": "text", + "text": "Parameterization-driven Neural Surface Reconstruction for Object-oriented Editing in Neural Rendering", + "text_level": 1, + "bbox": [ + 241, + 140, + 759, + 209 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Baixin $\\mathrm{Xu}^{1}$ , Jiangbei $\\mathrm{Hu}^{2,3}$ , Fei $\\mathrm{Hou}^{4,5}$ , Kwan-Yee $\\mathrm{Lin}^{6}$ , Wayne $\\mathrm{Wu}^{7}$ , Chen Qian $^{7}$ , and Ying $\\mathrm{He}^{1,2}$", + "bbox": [ + 222, + 234, + 777, + 265 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "$^{1}$ S-Lab, Nanyang Technological University, $^{2}$ College of Computing and Data Science, Nanyang Technological University, $^{3}$ Dalian University of Technology, $^{4}$ Key Laboratory of System Software (Chinese Academy of Sciences) and State Key Laboratory of Computer Science, Institute of Software, Chinese Academy of Sciences, $^{5}$ University of Chinese Academy of Sciences, $^{6}$ Shanghai AI Lab, $^{7}$ SenseTime Research", + "bbox": [ + 215, + 277, + 784, + 347 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Abstract. The advancements in neural rendering have increased the need for techniques that enable intuitive editing of 3D objects represented as neural implicit surfaces. This paper introduces a novel neural algorithm for parameterizing neural implicit surfaces to simple parametric domains like spheres and polycubes. Our method allows users to specify the number of cubes in the parametric domain, learning a configuration that closely resembles the target 3D object's geometry. It computes bidirectional deformation between the object and the domain using a forward mapping from the object's zero level set and an inverse deformation for backward mapping. We ensure nearly bijective mapping with a cycle loss and optimize deformation smoothness. The parameterization quality, assessed by angle and area distortions, is guaranteed using a Laplacian regularizer and an optimized learned parametric domain. Our framework integrates with existing neural rendering pipelines, using multi-view images of a single object or multiple objects of similar geometries to reconstruct 3D geometry and compute texture maps automatically, eliminating the need for any prior information. We demonstrate the method's effectiveness on images of human heads and man-made objects. The source code is available at https://xubaixinxbx.github.io/neuparam.", + "bbox": [ + 261, + 387, + 738, + 650 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Keywords: Polycube parameterization $\\cdot$ Neural surface reconstruction and editing $\\cdot$ Neural rendering", + "bbox": [ + 259, + 664, + 740, + 693 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "1 Introduction", + "text_level": 1, + "bbox": [ + 215, + 720, + 375, + 734 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Neural radiance fields (NeRF) [27] have garnered remarkable success in both the computer vision and computer graphics communities, redefining benchmarks for high-quality renderings and novel view synthesis. Building upon NeRF, a variety of methods leveraging implicit neural representations [30, 46, 58] have emerged,", + "bbox": [ + 212, + 753, + 787, + 815 + ], + "page_idx": 0 + }, + { + "type": "page_footnote", + "text": "Corresponding author: Ying He (email: yhe@ntu.edu.sg).", + "bbox": [ + 230, + 825, + 611, + 840 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "delivering high-fidelity 3D reconstructions. The demand for intuitive 3D object editing techniques has increased with the growing capabilities of neural rendering, especially for neural implicit surfaces. Extensive explorations have been made in editing within the Neural Rendering framework to achieve diverse visual effects. By manipulating the latent codes of the network, one can implicitly alter the shape and color of 3D objects [23,53,61]. In the work of [18,44], palette-based methods have been employed to facilitate transformations in the color and style of rendered scenes. However, these methods suffer from limited controllability and disallow local pixel editing. Seal3D [47] proposed a teacher-student training strategy to edit the interactive scene at the pixel level. Yet, this method is tricky to handle shading, and each subsequent edit required a new training iteration. By employing neural networks to formulate a differentiable surface parameterization, it can be facilitated to edit textures or shapes within the neural rendering pipeline [25,52]. Existing neural parameterization techniques still face significant challenges, such as the need for appropriate distortion constraints to ensure accurate detail reconstruction or reducing the reliance on prior information from tracked mesh and UV mapping. The final quality of edited visual effects is significantly impacted by these issues.", + "bbox": [ + 212, + 145, + 787, + 419 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "This paper aims to develop an intuitive and easy-to-use tool for appearance editing in neural implicit surfaces within the neural rendering framework, utilizing only 2D images as input. Towards this goal, we propose a novel method to parameterize neural implicit surfaces to parametric domains where the radiance field is represented as a texture map. Distinguished from previous work that using 2D texture mapping, our neural parameterization algorithm supports the selection of different 3D parametric domains according to geometric shapes to avoid seams and reduce distortion, such as spheres, and polycubes. Technically, our parameterization algorithm utilizes learnable parameters to fit a parametric domain first and a neural network to learn a bi-directional deformation between 3D objects and the chosen parametric domains. This involves a forward deformation that maps points from the zero level set of the neural implicit surface to the parametric domain, followed by an inverse deformation, mapping points backward. Notably, we do not require any explicit prior information in learning both deformations. We employ a cycle loss to ensure the smoothness of the bi-directional deformation, and a Laplacian regularization to effectively control angle distortion. With the parameterization, 3D radiance field is mapped to a simple 3D parametric domain which can be further disassembled into a 2D domain, facilitating visualization and various editing tasks. Furthermore, we decompose the 2D radiance field into two components: view-independent material and view-dependent shading, streamlining both texture and shading editing. Our neural parameterization algorithm is fully compatible with existing neural rendering pipelines, allowing 3D reconstruction from multi-view images as input and the creation of texture maps simultaneously. It allows for the immediate rendering of edited textures through volume rendering, without network re-training. Moreover, it supports co-parameterization of objects of similar geometry and enables texture transfer between them. We validate the effectiveness of", + "bbox": [ + 212, + 431, + 787, + 843 + ], + "page_idx": 1 + }, + { + "type": "page_number", + "text": "2", + "bbox": [ + 217, + 114, + 228, + 126 + ], + "page_idx": 1 + }, + { + "type": "header", + "text": "B. Xu et al.", + "bbox": [ + 271, + 114, + 352, + 127 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "our method on scenes of human heads and man-made objects. Our contributions are summarized as follows:", + "bbox": [ + 215, + 146, + 785, + 175 + ], + "page_idx": 2 + }, + { + "type": "list", + "sub_type": "text", + "list_items": [ + "- We present a neural parameterization framework that automatically learns a parametric domain and computes bi-directional deformation between 3D objects and their parametric domains, eliminating the need for any prior information from tracked mesh or UV mapping. Our approach utilizes Laplacian loss to minimize angle distortion and allows the user to specify the domain complexity to control area distortion.", + "- We introduce a simple yet effective technique for intrinsic radiance decomposition, facilitating both view-independent texture editing and view-dependent shading editing.", + "- Our neural parameterization framework is end-to-end and fully compatible with existing neural rendering architectures, accepting multi-view images as input to reconstruct 3D geometry and generate UV maps as output. Additionally, it enables the direct rendering of modified textures using the volume rendering pipeline.", + "- Our approach supports co-parameterization of multiple objects and allows for texture transfer between different objects." + ], + "bbox": [ + 225, + 188, + 787, + 431 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "2 Related Work", + "text_level": 1, + "bbox": [ + 215, + 457, + 387, + 472 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Parameterization. Surface parameterization [6, 40] aims at computing a bijective mapping between the 3D surface and a suitable parametric domain, usually a 2D region or a simple 3D object, such as spheres [8] and polycubes [7, 43]. Serving as an important computational tool in computer graphics and digital geometry processing [40], surface parameterization facilitates various applications, including texture editing [5], surface painting [42], details transfer [2], remeshing [33], among others. Neural networks are now commonly used in digital geometry processing, resulting in the creation of parameterization algorithms based on deep learning techniques [1, 11, 14, 50]. Unlike classical methods that yield global and seamless parametrization, most neural parameterization methods are computing local parametrization [11, 63, 64]. These methods typically partition the 3D model into multiple patches and parameterize each to a 2D region, possibly followed by a post-processing step to stitch the patches together. Although these methods are efficient and can handle surfaces of arbitrary geometry and topology, they frequently encounter problems including non-bijectivity, a lack of smoothness, the presence of seams and overlaps, and large distortions. There are also works on neural parametric surfaces, defined in rectangular domains [24] and $n$ -sided patches [57]. They can model complex surface geometries with high precision; however, such representations are not applicable to neural implicit surfaces and thereby incompatible with neural rendering.", + "bbox": [ + 215, + 491, + 787, + 792 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Neural Implicit Functions. Recent years have seen a surge in the successful application of neural implicit functions in 3D deep learning. Compared to explicit representations such as point clouds [34], voxels [35], and meshes [45],", + "bbox": [ + 215, + 795, + 787, + 842 + ], + "page_idx": 2 + }, + { + "type": "header", + "text": "NeuParam", + "bbox": [ + 658, + 114, + 730, + 127 + ], + "page_idx": 2 + }, + { + "type": "page_number", + "text": "3", + "bbox": [ + 774, + 116, + 784, + 126 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "neural implicit representations offer unique advantages including flexibility, continuity, and robustness [26,31,39]. When integrated within the neural rendering pipeline, neural implicit surfaces exhibit exceptional quality in 3D reconstructions from multi-view images [19,20,28,37,48,49,53]. However, as geometries and radiance fields are encoded as network parameters, editing them becomes complex and non-intuitive, compared to explicit representations. This paper aims at tackling this challenge by explicitly representing radiance fields as texture mapping that encompasses both view-dependent and view-independent components.", + "bbox": [ + 212, + 146, + 787, + 267 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "NeRF Editing. Numerous studies have focused on editing neural radiance fields, including relighting [21, 41], composition [32, 55], content generation [29], and shape editing [22, 62]. Editing tools operating at the scene- or object-level, such as [16, 55, 61], utilize latent codes to stylize the appearance of an object or entire scenes. Conversely, pixel-level editing tools, such as NeuMesh [54] and Seal3D [47], utilize training-based approaches to edit fine-grained details. However, NeuMesh relies on explicit meshes which lack efficiency and robustness. Although Seal3D achieves interactive editing, it still suffers from poor geometry reconstruction and cannot change the illumination scene. Parameterization-based methods, such as NeuTex [52], ISO-UVField [38] and NeP [25], integrate the learning of UV mapping into neural rendering, enabling intuitive editing within 2D domains. However, these methods often suffer from large distortions in the parametrization. Furthermore, they require 3D prior information from tracked meshes and UV mapping - conditions that may not be readily met in practical, real-world scenarios. IntrinsicNeRF [60] enhances standard neural radiance fields by generating additional outputs, including reflectance, shading, and a residual term. Incorporating a semantic branch, it facilitates real-time scene editing. However, it often results in aggregated colors, akin to palette-based methods [18, 44], hence cannot offer fine-grained editing results at the pixel level. We decompose the radiance field into view-dependent shadings and view-independent materials. Both components are represented as textures, enabling intuitive material modifications while preserving consistent shading.", + "bbox": [ + 212, + 268, + 787, + 599 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "3 Method", + "text_level": 1, + "bbox": [ + 215, + 625, + 328, + 641 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "The input of our method is a collection of RGB images, denoted $\\mathcal{I} = \\{\\mathbf{I}_i\\}$ , representing multiple objects of similar geometry or a single object, captured from different viewpoints. To differentiate among objects, we assign a latent shape code $\\mathbf{z}_s$ and an appearance code $\\mathbf{z}_a$ to each object. This facilitates the co-parameterization of multiple objects and enables the transfer of material and shading among them. As illustrated in Fig. 1, our method reconstructs 3D geometry as a neural implicit surface and encodes the 3D radiance field in a simple parametric domain, such as spheres and polycubes, which we learn from an initial phase in our training (Sec. 3.1). For each input image, we randomly select a set of pixels. Then, for each pixel, we shoot a ray originating from the location of camera and traversing through the pixel. If the ray intersects with the surface, at a point, denoted as $\\mathbf{p}$ , we compute its corresponding position in the", + "bbox": [ + 212, + 657, + 787, + 840 + ], + "page_idx": 3 + }, + { + "type": "page_number", + "text": "4", + "bbox": [ + 217, + 114, + 228, + 126 + ], + "page_idx": 3 + }, + { + "type": "header", + "text": "B. Xu et al.", + "bbox": [ + 271, + 114, + 352, + 127 + ], + "page_idx": 3 + }, + { + "type": "image", + "img_path": "images/963da5eeb0bf9c1d84a74bb4a661b1492b3045f853141d1807f00fcb01a79d64.jpg", + "image_caption": [ + "Fig. 1: Algorithmic pipeline. Initially, we learn the parametric domain $F_{\\mathrm{sdf}}$ using a coarse SDF $G_{\\mathrm{sdf}}$ of the target surface $\\mathcal{S}$ during the early phase of neural rendering (in the left grey box). Once the parametric domain $\\mathcal{D}$ is determined, our pipeline subsequently utilizes the bi-directional deformation $(F_{\\mathrm{def}}$ and $F_{\\mathrm{inv-def}})$ , the parametric domain SDF $(F_{\\mathrm{sdf}})$ and the radiance decomposition $(F_{\\mathrm{mat}}$ and $F_{\\mathrm{shd}})$ . The output includes the reconstructed 3D surface $\\mathcal{S}$ , the decomposed radiance fields (featuring both view-dependent and view-independent components) and the map between $\\mathcal{S}$ and $\\mathcal{D}$ . It is worth mentioning that while our framework involves two SDFs - one for the 3D surface $\\mathcal{S}$ and the other for the parametric domain $\\mathcal{D}$ - our network design only requires one of them. This is because one SDF can be derived from the other via either the forward or the backward deformation. To reduce the network complexity, we only adopt one geometry sub-network $F_{\\mathrm{sdf}}$ , which is used to represent the parametric domain $\\mathcal{D}$ ." + ], + "image_footnote": [], + "bbox": [ + 217, + 143, + 781, + 262 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "parametric domain $\\mathcal{D}$ , denoted as $\\mathbf{p}'$ , via a bi-directional deformation (Sec. 3.2): the forward deformation (depicted in orange) generates a displacement vector so that the updated position $\\mathbf{p}'$ is in the parametric domain; subsequently, the inverse deformation (depicted in green) maps $\\mathbf{p}'$ back to a point $\\mathbf{p}''$ , which is anticipated to be in close proximity to the surface. We incorporate a cycle loss to penalize occurrences where $\\mathbf{p}''$ deviates from $\\mathbf{p}$ . In addition, we employ a Laplacian loss (Sec. 3.3) to effectively reduce the angle distortion between the parametric domain and the original surface. Furthermore, we decompose the radiance into a view-independent material field and a view-dependent shading field (Sec. 3.4), each of which can be edited independently, thus augmenting the editing capability of our framework. See Fig. 1 for the algorithmic pipeline.", + "bbox": [ + 212, + 472, + 787, + 640 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "3.1 Parametric Domain Learning", + "text_level": 1, + "bbox": [ + 215, + 669, + 503, + 686 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "In our framework, we consider two types of parametric domains: spheres and polycubes, both represented as signed distance fields. Spheres are characterized by four learnable parameters: the center position and radius. Polycubes, a generalization of cubes, are simple structures yet offer sufficient degrees of freedom to capture the salient geometric features of target surfaces. We define a learnable polycube with $k$ elements as follows: For the $i$ -th element, which is an axis-aligned box, its SDF $\\phi_i$ is given by", + "bbox": [ + 212, + 703, + 787, + 809 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\n\\phi_ {i} = \\max \\left(\\left| x - x _ {i} \\right| - a _ {i}, \\left| y - y _ {i} \\right| - b _ {i}, \\left| z - z _ {i} \\right| - c _ {i}\\right), \\tag {1}\n$$\n", + "text_format": "latex", + "bbox": [ + 320, + 824, + 785, + 840 + ], + "page_idx": 4 + }, + { + "type": "header", + "text": "NeuParam", + "bbox": [ + 658, + 114, + 730, + 126 + ], + "page_idx": 4 + }, + { + "type": "page_number", + "text": "5", + "bbox": [ + 774, + 116, + 784, + 126 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "where $(x_{i},y_{i},z_{i})$ represents the box center, and $(a_i,b_i,c_i)$ is its dimensions. By combining the $k$ boxes, the SDF of the polycube is obtained as", + "bbox": [ + 212, + 146, + 782, + 176 + ], + "page_idx": 5 + }, + { + "type": "equation", + "text": "\n$$\nf _ {\\mathrm {p c}} = \\max (- \\phi_ {1}, - \\phi_ {2}, \\dots , - \\phi_ {k}). \\tag {2}\n$$\n", + "text_format": "latex", + "bbox": [ + 388, + 191, + 785, + 210 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "To ensure the differentiability of Eq. (2), we approximate it using Kreisselmeier-Steinhauser (KS) function [17] as", + "bbox": [ + 212, + 223, + 782, + 255 + ], + "page_idx": 5 + }, + { + "type": "equation", + "text": "\n$$\nf _ {\\mathrm {p c}} \\approx - 1 / \\lambda * \\log \\left(\\sum_ {i} \\exp (- \\lambda * \\phi_ {i})\\right),\n$$\n", + "text_format": "latex", + "bbox": [ + 361, + 268, + 638, + 310 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "where $\\lambda$ is a typically large value, set to 100 in our experiments.", + "bbox": [ + 214, + 323, + 676, + 338 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Since determining the configuration of boxes does not require a highly accurate geometry of the target object $S$ , we design the parametric domain learning as a one-time, intermediate step in the neural rendering pipeline. Specifically, after an initial training phase of approximately 100 epochs, the coarse geometry of the surface $S$ is obtained. Subsequently, we activate the domain learning module, which takes the SDF $G_{\\mathrm{sdf}}$ of the coarse surface $S$ and the user-specified number of boxes $k$ as input. It computes the SDF $F_{\\mathrm{sdf}}$ of the parametric domain $\\mathcal{D}$ via minimizing the loss function:", + "bbox": [ + 212, + 340, + 784, + 460 + ], + "page_idx": 5 + }, + { + "type": "equation", + "text": "\n$$\n\\mathcal {L} _ {\\text {p a r a m}} = \\mathcal {L} _ {\\text {s d f}} + \\lambda_ {s} \\mathcal {L} _ {\\text {l a p s d f}}, \\tag {3}\n$$\n", + "text_format": "latex", + "bbox": [ + 405, + 476, + 785, + 494 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "where", + "bbox": [ + 214, + 508, + 263, + 521 + ], + "page_idx": 5 + }, + { + "type": "equation", + "text": "\n$$\n\\mathcal {L} _ {\\mathrm {s d f}} = \\left| F _ {\\mathrm {s d f}} (p) - G _ {\\mathrm {s d f}} (p) \\right|, \\mathcal {L} _ {\\mathrm {l a p s d f}} = \\left| \\frac {\\partial^ {2} F _ {\\mathrm {s d f}}}{\\partial x ^ {2}} + \\frac {\\partial^ {2} F _ {\\mathrm {s d f}}}{\\partial y ^ {2}} + \\frac {\\partial^ {2} F _ {\\mathrm {s d f}}}{\\partial z ^ {2}} \\right|, \\tag {4}\n$$\n", + "text_format": "latex", + "bbox": [ + 267, + 532, + 784, + 568 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "and the weight $\\lambda_{s}$ is set to 0.01. The term $\\mathcal{L}_{\\mathrm{lapsdf}}$ , which is the Laplacian of the signed distance field of the polycube $\\mathcal{D}$ , plays a critical role in controlling its geometric complexity.", + "bbox": [ + 212, + 580, + 782, + 626 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "After obtaining the SDF $F_{\\mathrm{sdf}}$ of the polycube $\\mathcal{D}$ , we deprecate the SDF network $G_{\\mathrm{sdf}}$ and instead use the deformation network $F_{\\mathrm{def}}$ and parametric domain SDF $F_{\\mathrm{sdf}}$ network to represent the source surface $\\mathcal{S}$ . The right inset showcases a polycube learning example for a human head model. Given a user-specified parameter $k = 3$ , our method initializes three randomly placed boxes. It proceeds to minimize the loss function $\\mathcal{L}_{\\mathrm{param}}$ , iteratively refining the size and positioning of each box to ensure the polycube closely approximates the surface $\\mathcal{S}$ . This parametric domain learning phase typically converges within 30 to 60 iterations, taking under 10 minutes to complete. For more details, see the supplementary video.", + "bbox": [ + 212, + 628, + 602, + 840 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/f169aa05f57a731bc608eb88a19c769316c937284fa4aa4ffcc2584ddaee1890.jpg", + "image_caption": [ + "S" + ], + "image_footnote": [], + "bbox": [ + 617, + 660, + 666, + 708 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/c3bc912a01a6b50b0d4032b7e6d298d9bb203835e0752bba878fa696dcaf815d.jpg", + "image_caption": [ + "Iter 0" + ], + "image_footnote": [], + "bbox": [ + 673, + 660, + 722, + 709 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/8f82e0edccb21e8f205ff02045a024ab208980f0bfc1820bc264a8d1aab95d62.jpg", + "image_caption": [ + "Iter 5" + ], + "image_footnote": [], + "bbox": [ + 735, + 662, + 776, + 708 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/5ed87422f6da40ad1ddb0af8f30248c9032fd45f4d9ac7bb9ff6af0fe0f055c6.jpg", + "image_caption": [ + "Iter 10" + ], + "image_footnote": [], + "bbox": [ + 620, + 729, + 666, + 775 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/b309029130d53c4913353dd82b6a113f5e71a7b5a035e104d5a46c1329e0a5f9.jpg", + "image_caption": [ + "Iter 45" + ], + "image_footnote": [], + "bbox": [ + 674, + 728, + 722, + 773 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/89aa5be24bf93e0694c1392ef9c152514fcd43115506f3be6644341795aa6e39.jpg", + "image_caption": [ + "Iter 55" + ], + "image_footnote": [], + "bbox": [ + 735, + 728, + 779, + 775 + ], + "page_idx": 5 + }, + { + "type": "page_number", + "text": "6", + "bbox": [ + 217, + 114, + 228, + 126 + ], + "page_idx": 5 + }, + { + "type": "header", + "text": "B. Xu et al.", + "bbox": [ + 271, + 114, + 352, + 127 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "3.2 Bi-directional Deformation", + "text_level": 1, + "bbox": [ + 215, + 146, + 483, + 159 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "The bi-directional deformation module serves as the pivotal component within our neural parameterization framework. As illustrated in Fig. 1, this module consists of two sub-networks: $F_{\\mathrm{def}}$ , which is responsible for the forward mapping from the original model to the parameter domain, and $F_{\\mathrm{inv-def}}$ , which establishes the inverse mapping.", + "bbox": [ + 212, + 167, + 782, + 242 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Let $S$ be the 3D surface and $\\mathcal{D}$ be the parametric domain; both are represented by the zero level-set of neural signed distance fields. For an arbitrary point $\\mathbf{p} \\in S$ , the forward deformation $F_{\\mathrm{def}}(\\mathbf{p}, \\mathbf{z}_s)$ computes a displacement vector, mapping the point $\\mathbf{p}$ to a corresponding point $\\mathbf{p}' \\in \\mathcal{D}$ in the parametric domain as follows:", + "bbox": [ + 212, + 243, + 782, + 316 + ], + "page_idx": 6 + }, + { + "type": "equation", + "text": "\n$$\n\\mathbf {p} ^ {\\prime} = \\mathbf {p} + F _ {\\mathrm {d e f}} (\\mathbf {p}, \\mathbf {z} _ {s}). \\tag {5}\n$$\n", + "text_format": "latex", + "bbox": [ + 426, + 319, + 784, + 334 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Conversely, we employ $F_{\\mathrm{inv - def}}$ to map points inversely from the parametric domain $\\mathcal{D}$ back to the original shape $S$ :", + "bbox": [ + 214, + 338, + 782, + 368 + ], + "page_idx": 6 + }, + { + "type": "equation", + "text": "\n$$\n\\mathbf {p} ^ {\\prime \\prime} = \\mathbf {p} ^ {\\prime} + F _ {\\text {i n v - d e f}} \\left(\\mathbf {p} ^ {\\prime}, \\mathbf {z} _ {s}\\right), \\tag {6}\n$$\n", + "text_format": "latex", + "bbox": [ + 410, + 376, + 784, + 392 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "where the point $\\mathbf{p}''$ is expected to be in close proximity to $p$ . To encourage the smoothness and stability of the deformations, we impose an $L_{2}$ loss on the magnitude of the displacement vectors:", + "bbox": [ + 214, + 398, + 782, + 443 + ], + "page_idx": 6 + }, + { + "type": "equation", + "text": "\n$$\n\\mathcal {L} _ {\\text {s m o o t h}} = \\left\\| F _ {\\text {d e f}} \\left(\\mathbf {p}, \\mathbf {z} _ {s}\\right) \\right\\| _ {2} + \\left\\| F _ {\\text {i n v - d e f}} \\left(\\mathbf {p} ^ {\\prime}, \\mathbf {z} _ {s}\\right) \\right\\| _ {2}. \\tag {7}\n$$\n", + "text_format": "latex", + "bbox": [ + 339, + 452, + 784, + 467 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "This loss helps minimize abrupt and irregular deformations, leading to more coherent and gradual changes between $\\mathcal{S}$ and $\\mathcal{D}$ .", + "bbox": [ + 214, + 474, + 782, + 503 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "In training the networks $F_{\\mathrm{def}}$ and $F_{\\mathrm{inv - def}}$ , our objective is to achieve a bijective mapping between the 3D surface $S$ and the parametric domain $\\mathcal{D}$ . To achieve this goal, we adopt a cycle loss, which is similar to NeuTex [52]:", + "bbox": [ + 212, + 503, + 782, + 550 + ], + "page_idx": 6 + }, + { + "type": "equation", + "text": "\n$$\n\\mathcal {L} _ {\\text {c y c l e}} = \\sum_ {\\mathbf {p} \\in S} \\lambda (\\mathbf {p}) \\| \\mathbf {p} - \\mathbf {p} ^ {\\prime \\prime} \\| _ {2}, \\tag {8}\n$$\n", + "text_format": "latex", + "bbox": [ + 388, + 556, + 784, + 580 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "where the weight $\\lambda$ characterizes the importance of sample $\\mathbf{p}$ to the loss. The reason that we do not consider all sample points equally is that a sample point may not be exactly on the zero level-set of the SDF. To tolerate such an inaccuracy, we define $\\lambda (\\mathbf{p}) = T(\\mathbf{p})(1 - \\exp (-\\sigma (s)\\delta (\\mathbf{p})))$ as the color weight used in volume rendering [27]. Here, $T(\\mathbf{p})$ is the transparency of sample $\\mathbf{p}$ in the viewing direction, $\\delta$ is the length of the sample interval along the ray, and $s$ is the signed distance value of the point $\\mathbf{p}$ . Clearly, when $\\mathbf{p}$ is away from the surface, the coefficient $\\lambda (\\mathbf{p})$ has few effects on this loss. We refer the readers to [27] for details about volume rendering and the computation of transparency $T$ .", + "bbox": [ + 212, + 585, + 784, + 720 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "By explicitly minimizing the distance between $\\mathbf{p}$ and $\\mathbf{p}''$ , the cycle loss effectively prevents scenarios where two distinct points $\\mathbf{p}_1, \\mathbf{p}_2 \\in S$ map to an identical point $\\mathbf{p}' \\in \\mathcal{D}$ . In such cases, the inverse deformation $F_{\\mathrm{inv-def}}$ would be unable to map the single point $\\mathbf{p}'$ back to the respective distinct points $\\mathbf{p}_1''$ and $\\mathbf{p}_2''$ .", + "bbox": [ + 212, + 722, + 782, + 799 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "3.3 Laplacian Regularization", + "text_level": 1, + "bbox": [ + 214, + 816, + 465, + 830 + ], + "page_idx": 6 + }, + { + "type": "header", + "text": "NeuParam", + "bbox": [ + 658, + 114, + 730, + 127 + ], + "page_idx": 6 + }, + { + "type": "page_number", + "text": "7", + "bbox": [ + 774, + 116, + 785, + 126 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Mitigating angle distortion is crucial in ensuring the quality of parameterization. One approach to reduce angle distortion involves minimizing the ratio of the two singular values of the Jacobian matrix associated with the parameterization [6]. However, considering the representation of the surface $S$ , the parametric domain $\\mathcal{D}$ , and the parameterization $F_{\\mathrm{def}}$ as neural implicit functions in our approach, directly computing the Jacobian matrix is a non-trivial task. To circum", + "bbox": [ + 212, + 146, + 602, + 281 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/aaf6a8fe310706b9a07514d27adc58af13eb6ad0f7d8997ac3b934d57e018a0d.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 616, + 148, + 782, + 252 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "vent this, we employ a Laplacian regularizer in our implementation to control angle distortion. Achieving the minimizer of the Laplacian of the map results in a harmonic map. For genus-0 surfaces, harmonic maps are also conformal, which preserves angles [12]. Although for surfaces of higher genus, a harmonic map is not necessarily conformal, it still serves an effective tool for reducing angle distortion [13]. Specifically, for a sample point $\\mathbf{p} \\in S$ and its neighboring points $\\mathbf{q}_j \\in S$ , the forward deformation maps them to $\\mathbf{p}' \\in \\mathcal{D}$ and $\\mathbf{q}_j' \\in \\mathcal{D}$ , respectively. We compute the Laplacian of the sample $\\mathbf{p}$ as follows:", + "bbox": [ + 212, + 282, + 787, + 404 + ], + "page_idx": 7 + }, + { + "type": "equation", + "text": "\n$$\n\\Delta (\\mathbf {p}) = \\sum_ {j = 1} ^ {m} \\omega_ {j} \\left(\\mathbf {p} ^ {\\prime} - \\mathbf {q} _ {j} ^ {\\prime}\\right). \\tag {9}\n$$\n", + "text_format": "latex", + "bbox": [ + 401, + 412, + 785, + 438 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Here, the weights $w_{j}$ are defined using the metric from the original surface $S$ as:", + "bbox": [ + 215, + 446, + 785, + 463 + ], + "page_idx": 7 + }, + { + "type": "equation", + "text": "\n$$\n\\omega_ {j} = \\exp (- \\| \\mathbf {p} - \\mathbf {q} _ {j} \\| _ {2} / l), \\tag {10}\n$$\n", + "text_format": "latex", + "bbox": [ + 405, + 477, + 785, + 494 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "where $l$ is the average distance from point $\\mathbf{p}$ to its neighboring points. Subsequently, the weights are normalized $\\omega_{j} = \\omega_{j} / \\sum_{k = 1}^{m}\\omega_{k}$ . Notice that $\\varDelta(\\mathbf{p})\\in\\mathbb{R}^3$ is a 3D vector, which may not necessarily lie on the tangent plane. Therefore, we define the normal component of the Laplacian as:", + "bbox": [ + 212, + 500, + 787, + 561 + ], + "page_idx": 7 + }, + { + "type": "equation", + "text": "\n$$\n\\Delta (\\mathbf {p}) _ {\\perp} = \\langle \\Delta (\\mathbf {p}), \\mathbf {n} \\rangle \\mathbf {n}, \\tag {11}\n$$\n", + "text_format": "latex", + "bbox": [ + 421, + 571, + 785, + 589 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "where $\\mathbf{n}$ is the normal at $\\mathbf{p}'$ , and $\\langle, \\rangle$ represents the inner product. Summing the tangential components yields the Laplacian loss:", + "bbox": [ + 212, + 598, + 785, + 628 + ], + "page_idx": 7 + }, + { + "type": "equation", + "text": "\n$$\n\\mathcal {L} _ {\\mathrm {L a p}} = \\sum_ {\\mathbf {p} \\in \\mathcal {S}} \\| \\Delta (\\mathbf {p}) - \\Delta (\\mathbf {p}) _ {\\perp} \\| _ {2}. \\tag {12}\n$$\n", + "text_format": "latex", + "bbox": [ + 377, + 638, + 785, + 664 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "In our implementation, we set $m = 6$ neighboring points for each sample $\\mathbf{p}$ .", + "bbox": [ + 214, + 672, + 758, + 688 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "3.4 Appearance Decomposition", + "text_level": 1, + "bbox": [ + 214, + 710, + 488, + 724 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Radiance represents the directional emission of color, illustrating how the appearance of an object varies in response to changes in the observational viewpoint [27]. Drawing inspiration from previous works [25, 60] and applying the assumptions of Lambertian and grayscale shading [4], we decompose the radiance field on the parametric domain into two components: a view-independent material field and a view-dependent shading field (see Fig. 6(a)). This decomposition allows for independent editing of material and shading. More precisely, for", + "bbox": [ + 212, + 733, + 787, + 840 + ], + "page_idx": 7 + }, + { + "type": "page_number", + "text": "8", + "bbox": [ + 217, + 114, + 228, + 126 + ], + "page_idx": 7 + }, + { + "type": "header", + "text": "B. Xu et al.", + "bbox": [ + 271, + 114, + 352, + 127 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "In a sample point $\\mathbf{p} \\in \\mathcal{S}$ , we employ two sub-networks, $F_{\\mathrm{mat}}$ and $F_{\\mathrm{shd}}$ , to decompose the radiance $\\mathbf{r} \\in \\mathbb{R}^3$ at point $\\mathbf{p}$ as follows:", + "bbox": [ + 214, + 146, + 782, + 176 + ], + "page_idx": 8 + }, + { + "type": "equation", + "text": "\n$$\n\\mathbf {r} (\\mathbf {p}) = F _ {\\mathrm {m a t}} \\left(\\mathbf {p} ^ {\\prime}, \\mathbf {n}, \\mathbf {z} _ {a}\\right) \\cdot \\exp \\left(F _ {\\mathrm {s h d}} \\left(\\mathbf {p} ^ {\\prime}, \\mathbf {n}, \\mathbf {v}, \\mathbf {z} _ {a}\\right)\\right), \\tag {13}\n$$\n", + "text_format": "latex", + "bbox": [ + 336, + 191, + 784, + 207 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "where $\\mathbf{p}' \\in \\mathcal{D}$ is the image of $\\mathbf{p}$ and $\\mathbf{n}$ is the unit normal of $\\mathbf{p}$ on the original surface. The material network $F_{\\mathrm{mat}}$ takes the mapped point $\\mathbf{p}'$ , the surface normal $\\mathbf{n}$ , and the appearance latent code $\\mathbf{z}_a$ as input, generating the view-independent material field in response. The shading network $F_{\\mathrm{shd}}$ , which is designed for computing the view-dependent shading field, takes an additional input: the viewpoint $\\mathbf{v}$ . To ensure the shading sub-network $F_{\\mathrm{shd}}$ concentrates on local shading features, we employ an $L_1$ loss to the output of $F_{\\mathrm{shd}}$ , encouraging sparsity,", + "bbox": [ + 214, + 219, + 782, + 325 + ], + "page_idx": 8 + }, + { + "type": "equation", + "text": "\n$$\n\\mathcal {L} _ {\\mathrm {s h d}} = \\left\\| F _ {\\mathrm {s h d}} \\left(\\mathbf {p} ^ {\\prime}, \\mathbf {n}, \\mathbf {v}, \\mathbf {z} _ {a}\\right) \\right\\| _ {1}. \\tag {14}\n$$\n", + "text_format": "latex", + "bbox": [ + 401, + 340, + 784, + 354 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "To compute the color, we need the signed distance values for the samples $\\mathbf{p}$ . As mentioned above, our network adopts only one SDF network $F_{\\mathrm{sdf}}$ , which represents the geometry of the parametric domain $\\mathcal{D}$ . We use the forward deformation $F_{\\mathrm{def}}$ to compute $s(\\mathbf{p})$ as", + "bbox": [ + 214, + 369, + 782, + 429 + ], + "page_idx": 8 + }, + { + "type": "equation", + "text": "\n$$\ns (\\mathbf {p}) = F _ {\\mathrm {s d f}} \\left(\\mathbf {p} + F _ {\\mathrm {d e f}} (\\mathbf {p}, \\mathbf {z} _ {s})\\right). \\tag {15}\n$$\n", + "text_format": "latex", + "bbox": [ + 398, + 444, + 784, + 459 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "Then we employ the SDF-based volume rendering [58] to integrate the radiances along the ray to compute color $\\mathbf{c}_{\\mathrm{pred}} = \\sum_{i}\\lambda_{i}\\mathbf{r}_{i}$ , where $\\mathbf{r}_i$ is the radiance of a sample point $\\mathbf{p}_i$ and $\\lambda_{i}$ is the same color weight as used in Eq. (8).", + "bbox": [ + 214, + 472, + 782, + 518 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "3.5 Training Losses", + "text_level": 1, + "bbox": [ + 215, + 544, + 388, + 559 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "In contrast to NeP [25], which necessitate prior information such as tracked mesh and UV mapping for supervision, our method is capable of reconstructing 3D geometry and computing the corresponding texture map directly from multiview images. This process is supervised by the image loss", + "bbox": [ + 214, + 571, + 782, + 633 + ], + "page_idx": 8 + }, + { + "type": "equation", + "text": "\n$$\n\\mathcal {L} _ {\\mathrm {r g b}} = \\left\\| \\mathbf {c} _ {\\mathrm {g t}} - \\mathbf {c} _ {\\mathrm {p r e d}} \\right\\| _ {1}, \\tag {16}\n$$\n", + "text_format": "latex", + "bbox": [ + 423, + 647, + 784, + 662 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "where $\\mathbf{c}_{\\mathrm{gt}}$ and $\\mathbf{c}_{\\mathrm{pred}}$ are the ground-truth and predicted colors, respectively. To ensure the shape and appearance latent codes, $\\mathbf{z}_s$ and $\\mathbf{z}_a$ , conform to Gaussian distributions, we incorporate a regularization term, expressed as $\\mathcal{L}_{\\mathrm{code}} = \\| \\mathbf{z}_s\\|_2 + \\| \\mathbf{z}_a\\|_2$ [31]. Additionally, with geometry represented as signed distance fields, we employ the Eikonal loss [10, 59], defined as $\\mathcal{L}_{\\mathrm{Eik}} = (\\| \\nabla s\\|_2 - 1)^2$ . Putting it all together, we define the total loss function as follows:", + "bbox": [ + 214, + 676, + 782, + 766 + ], + "page_idx": 8 + }, + { + "type": "equation", + "text": "\n$$\n\\mathcal {L} = \\mathcal {L} _ {\\mathrm {r g b}} + \\lambda_ {1} \\mathcal {L} _ {\\text {E i k}} + \\lambda_ {2} \\mathcal {L} _ {\\text {c y c l e}} + \\lambda_ {3} \\mathcal {L} _ {\\text {s m o o t h}} + \\lambda_ {4} \\mathcal {L} _ {\\text {L a p}} + \\lambda_ {5} \\mathcal {L} _ {\\text {s h d}} + \\lambda_ {6} \\mathcal {L} _ {\\text {c o d e}}. \\tag {17}\n$$\n", + "text_format": "latex", + "bbox": [ + 228, + 781, + 784, + 796 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "In our implementation, we empirically set the coefficients as $\\lambda_1 = \\lambda_2 = \\lambda_5 = \\lambda_6 = 0.01$ and $\\lambda_3 = \\lambda_4 = 0.001$ .", + "bbox": [ + 214, + 809, + 782, + 839 + ], + "page_idx": 8 + }, + { + "type": "header", + "text": "NeuParam", + "bbox": [ + 658, + 114, + 730, + 127 + ], + "page_idx": 8 + }, + { + "type": "page_number", + "text": "9", + "bbox": [ + 774, + 116, + 785, + 126 + ], + "page_idx": 8 + }, + { + "type": "image", + "img_path": "images/5b09bf6d523c14c80d6f045f8073405bb68530e58d45c28c39ab9050e5b41531.jpg", + "image_caption": [ + "Fig. 2: Parameterization results. Human heads are co-parameterized to a sphere owing to the clear geometric resemblance. Given the diverse geometry of cars, each is parameterized to a polycube domain (bottom three rows). From top to bottom: the input image, the normal map of the reconstructed geometry $S$ , the parametric domain $\\mathcal{D}$ , and the reconstructed surface $S$ . See supplementary material for additional results." + ], + "image_footnote": [], + "bbox": [ + 271, + 142, + 728, + 441 + ], + "page_idx": 9 + }, + { + "type": "text", + "text": "4 Experiments", + "text_level": 1, + "bbox": [ + 215, + 549, + 375, + 566 + ], + "page_idx": 9 + }, + { + "type": "text", + "text": "Datasets. For our experiments, we used a customized dataset named FS-Syn, derived from the Facescape dataset [56], selecting 10 human head models and configuring 30 fixed viewpoints and lighting conditions to obtain synthesized multi-view images. We also evaluated our method on several man-made objects from the OmniObject dataset [51], and the H3DS dataset [36].", + "bbox": [ + 212, + 580, + 787, + 656 + ], + "page_idx": 9 + }, + { + "type": "text", + "text": "Training details and metrics. All the sub-networks $F_{\\mathrm{def}}$ , $F_{\\mathrm{inv-def}}$ , $F_{\\mathrm{sdf}}$ , $F_{\\mathrm{mat}}$ , and $F_{\\mathrm{shd}}$ are MLPs. See the supplementary material for other detailed network hyperparameters. Once the parametric domain is chosen, we fixed $F_{\\mathrm{sdf}}$ and trained other sub-networks for 2,000 epochs driven by the loss function as Eq. (17). We applied the standard Marching Cubes algorithm to extract triangle meshes for the zero level-set of the source surface $S$ and then used the forward map to obtain the mesh representing the parametric surface. With triangle meshes, we measured the quality of the computed parameterization by the angle distortion $\\mathcal{E}_{\\mathrm{angle}}$ and area distortion [3,15] $\\mathcal{E}_{\\mathrm{area}}$ as follows:", + "bbox": [ + 212, + 656, + 787, + 792 + ], + "page_idx": 9 + }, + { + "type": "equation", + "text": "\n$$\n\\mathcal {E} _ {\\mathrm {a n g l e}} = \\frac {\\cot \\alpha | a | ^ {2} + \\cot \\beta | b | ^ {2} + \\cot \\gamma | c | ^ {2}}{4 \\mathrm {a r e a} (\\varDelta_ {i})}, \\quad \\mathcal {E} _ {\\mathrm {a r e a}} = \\frac {1}{2} \\left(\\frac {\\mathrm {a r e a} (\\varDelta_ {i})}{\\mathrm {a r e a} (\\hat {\\varDelta} _ {i})} + \\frac {\\mathrm {a r e a} (\\hat {\\varDelta} _ {i})}{\\mathrm {a r e a} (\\varDelta_ {i})}\\right),\n$$\n", + "text_format": "latex", + "bbox": [ + 217, + 801, + 782, + 845 + ], + "page_idx": 9 + }, + { + "type": "page_number", + "text": "10", + "bbox": [ + 217, + 114, + 235, + 126 + ], + "page_idx": 9 + }, + { + "type": "header", + "text": "B. Xu et al.", + "bbox": [ + 271, + 114, + 352, + 126 + ], + "page_idx": 9 + }, + { + "type": "text", + "text": "where $a, b, c$ represent the side lengths of the triangle $\\Delta_i$ of the original surface. $\\alpha, \\beta, \\gamma$ are angles of the corresponding triangle $\\hat{\\Delta}_i$ in parametric domain. In the case of conformal mapping, $\\mathcal{E}_{\\mathrm{angle}}$ is equal to 1, and $\\mathcal{E}_{\\mathrm{area}}$ is also equal to 1 for an isometric map.", + "bbox": [ + 212, + 146, + 787, + 205 + ], + "page_idx": 10 + }, + { + "type": "image", + "img_path": "images/6a00a32da747038db7cde00c08403bd135ea08c15804652627fab46f77b186cc.jpg", + "image_caption": [ + "1.545/1.396" + ], + "image_footnote": [], + "bbox": [ + 251, + 237, + 346, + 320 + ], + "page_idx": 10 + }, + { + "type": "image", + "img_path": "images/750fc2cbafbd9ed1e04663849ce79353cd61ae5efbcd23d35bb76c168343d6ab.jpg", + "image_caption": [ + "1.755/1.446" + ], + "image_footnote": [], + "bbox": [ + 351, + 237, + 447, + 320 + ], + "page_idx": 10 + }, + { + "type": "image", + "img_path": "images/7681f59da7455f21547f6e81443fd0eb5233e083fb0bb35df31878d6bcd3aeb8.jpg", + "image_caption": [ + "1.394/1.302", + "Fig. 3: Results of cross-polycube parameterization on the H3DS dataset [36]. We learned a common polycube domain for all human heads with the parameter $k = 3$ (i.e., the polycube is formed by 3 boxes) and then computed the polycube parameterization for all the human heads simultaneously. Since human heads share similar geometries, the computed polycube parameterizations are also highly consistent. The values below each figure represent the angle distortion and the area distortion, respectively." + ], + "image_footnote": [], + "bbox": [ + 450, + 238, + 547, + 320 + ], + "page_idx": 10 + }, + { + "type": "image", + "img_path": "images/28bb4180d730adcae61ab5233b9287a83231d102fdcaf4bbb40f3bdcebd5e06b.jpg", + "image_caption": [ + "1.271/1.216" + ], + "image_footnote": [], + "bbox": [ + 553, + 239, + 650, + 320 + ], + "page_idx": 10 + }, + { + "type": "image", + "img_path": "images/6a0347dd2df0b4c34042a16e62a8d674254377381dd8afdf0a7ee0a3359085cc.jpg", + "image_caption": [ + "1.351/1.275" + ], + "image_footnote": [], + "bbox": [ + 653, + 238, + 751, + 320 + ], + "page_idx": 10 + }, + { + "type": "text", + "text": "Results. Our framework constructs a nearly bijective map between objects and multiple parametric domains such as spheres and polycubes, as shown in Fig. 2 and Fig. 3. It supports co-parameterization of multiple objects and facilitates material and shading transfer. Fully compatible with existing neural rendering pipelines, our neural parameterization algorithm uses multi-view images to simultaneously reconstruct 3D geometry and create texture maps. As shown in Fig. 2, we achieve high-fidelity results without prior information.", + "bbox": [ + 212, + 463, + 787, + 570 + ], + "page_idx": 10 + }, + { + "type": "image", + "img_path": "images/88c65bc948910005a4ba6f345e0bf59c409d1d687eab48214fb3b3446b2ba4e6.jpg", + "image_caption": [ + "UV prior", + "Fig. 4: Parametrization results of NeP [25] in comparison with our method. NeP heavily relies on the initial UV prior, leading to poor texture maps when such prior information is lacking. Our method learns texture maps from a set of multi-view images without any 3D tracked mesh or UV prior as input." + ], + "image_footnote": [], + "bbox": [ + 217, + 597, + 321, + 679 + ], + "page_idx": 10 + }, + { + "type": "image", + "img_path": "images/f07aa17abbfb2eceb87234b69ea04416b4dd9b4d51a3200fc398044d43d2a2e7.jpg", + "image_caption": [ + "NeP w/o prior" + ], + "image_footnote": [], + "bbox": [ + 336, + 597, + 439, + 679 + ], + "page_idx": 10 + }, + { + "type": "image", + "img_path": "images/d3309025c2dccb88f6336bda646ae931e72565ea5f444e48adc1669865705e22.jpg", + "image_caption": [ + "w/ prior" + ], + "image_footnote": [], + "bbox": [ + 455, + 597, + 558, + 679 + ], + "page_idx": 10 + }, + { + "type": "image", + "img_path": "images/a8af08f8d760abec17f6d44470ba8b68364aec39d1053fe19ab60b085d341025.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 576, + 599, + 655, + 679 + ], + "page_idx": 10 + }, + { + "type": "image", + "img_path": "images/2d8095154c1dde858c8ae7f093ffbc348f641faac0abcd6cf0d5f2729f4b33e2.jpg", + "image_caption": [ + "Ours" + ], + "image_footnote": [], + "bbox": [ + 666, + 609, + 785, + 679 + ], + "page_idx": 10 + }, + { + "type": "text", + "text": "Compared with NeP [25], which is the latest neural surface editing work using neural parameterization, our method eliminates the need for prior information from UV mapping. Unlike our algorithm, NeP learns a map between", + "bbox": [ + 212, + 794, + 787, + 840 + ], + "page_idx": 10 + }, + { + "type": "header", + "text": "NeuParam", + "bbox": [ + 656, + 114, + 730, + 127 + ], + "page_idx": 10 + }, + { + "type": "page_number", + "text": "11", + "bbox": [ + 767, + 114, + 784, + 126 + ], + "page_idx": 10 + }, + { + "type": "text", + "text": "surfaces and the UV plane. However, it heavily relies on an existing UV mapping as initialization, and yields rather poor texture mapping results without the UV mapping as prior (see Fig.4). Moreover, the results from NeP still exhibit significant distortions even with UV prior.", + "bbox": [ + 212, + 146, + 787, + 207 + ], + "page_idx": 11 + }, + { + "type": "image", + "img_path": "images/f61ee14219f8b893bfb539e15bef8f0cd09e3369cd3907e8b11a2abf140e7f5a.jpg", + "image_caption": [ + "Fig. 5: Representing the materials as textures defined in a simple parametric domain enables easy editing. Also, our co-parameterization naturally supports texture transfer between different objects." + ], + "image_footnote": [], + "bbox": [ + 232, + 233, + 769, + 362 + ], + "page_idx": 11 + }, + { + "type": "text", + "text": "Texture editing. Our parameterization results enable flexible and pixel-level editing as shown in Fig. 5. We re-paint the extracted 2D texture map using cubemaps [9] and retrieve the modified color using the established mapping when volume rendering novel images after each editing. Both NeuMesh [54] and NeP [25] can edit the object's texture at a pixel level. We compare the editing and rendering results with NeuMesh and NeP, as shown in Fig. 5. NeuMesh requires a mask of the editing region. It also re-trains the network after each edit, which is time-consuming and cannot guarantee the stability of the training results. Our method outperforms NeuMesh in terms of the visual quality. NeP, which is also neural parameterization-based method, requires a UV prior for initialization. It often yields poor parameterization results without such prior. Even with the UV prior, its parameterization results exhibit large distortion, making the edit in the 2D domain challenging. Our method is free of prior, and can yield parameterization with low distortion, thereby facilitating pixel-level editing and material transferring without network re-training.", + "bbox": [ + 212, + 446, + 787, + 672 + ], + "page_idx": 11 + }, + { + "type": "text", + "text": "Shading editing. We decompose the 3D radiance field into view-independent material and view-dependent shading defined in the parametric domain. Taking advantage of this decomposition and the embedding into a simple domain, we can edit both the material and shading of the 3D surfaces in an intuitive manner. As illustrated in Fig. 6(b), we modify the shading of one model by transferring the shading from another model under different lighting conditions, while keeping the material unchanged.", + "bbox": [ + 212, + 674, + 787, + 779 + ], + "page_idx": 11 + }, + { + "type": "text", + "text": "Inverse deformation. To validate the cycle loss and illustrate inverse deformation, we sample points on the polycube parametric surface $\\mathcal{D}$ . Subsequently, we inversely deform these points back to the surface $\\mathcal{S}$ . The point cloud is visualized in the right inset, and we quantify the CD $(\\times 10^{-3})$ with respect to", + "bbox": [ + 212, + 779, + 787, + 840 + ], + "page_idx": 11 + }, + { + "type": "page_number", + "text": "12", + "bbox": [ + 217, + 114, + 235, + 126 + ], + "page_idx": 11 + }, + { + "type": "header", + "text": "B. Xu et al.", + "bbox": [ + 271, + 114, + 352, + 126 + ], + "page_idx": 11 + }, + { + "type": "image", + "img_path": "images/b334bf3e51c1828f978a9f690c8dfdf1028f4f0381218519f94e8c5e0e0bfd5f.jpg", + "image_caption": [ + "(a) We decompose the radiance into view-dependent shadings and view-independent materials. The shading map is visualized from the front view." + ], + "image_footnote": [], + "bbox": [ + 246, + 141, + 759, + 220 + ], + "page_idx": 12 + }, + { + "type": "image", + "img_path": "images/a9133be996c0193c69560d7e14614ce69f391e4cca377547fa373c443f3b6b27.jpg", + "image_caption": [ + "(b) We transfer the decomposed shading from one model to another. See also the accompanying video.", + "Fig. 6: We transfer the decomposed shading from one model to another. For further illustration, please refer to the accompanying video." + ], + "image_footnote": [], + "bbox": [ + 232, + 247, + 769, + 383 + ], + "page_idx": 12 + }, + { + "type": "text", + "text": "the ground truth. The obtained low CD results confirm the efficacy of inverse deformation and cycle loss.", + "bbox": [ + 212, + 481, + 787, + 508 + ], + "page_idx": 12 + }, + { + "type": "text", + "text": "Ablation studies. As shown in Fig. 7 (left), the term $\\mathcal{L}_{\\mathrm{smooth}}$ helps to smooth the process of the two deformations, while $\\mathcal{L}_{\\mathrm{Lap}}$ further reduces the angle distortion. Our method can effectively reduce angle distortion by minimizing Laplacian loss and reduce area distortion by choosing an appropriate parametric domain. As demonstrated in Fig. 7 (right), we calculate the angle distortion $\\mathcal{E}_{\\mathrm{angle}}$ of parameterization as described in [3, 15]. We can observe that the Laplacian loss can effectively reduce angle distortion.", + "bbox": [ + 212, + 512, + 787, + 617 + ], + "page_idx": 12 + }, + { + "type": "text", + "text": "The choice of parametric domain. The different parametric domain influences the quality of parameterization as shown in Fig. 8 (a), which is measured by area distortions. Generally, increasing the number of cubic primitives for better resembling the geometry of the target surface effectively reduces distortions but increases the complexity of the parametric domain. Finding an optimal balance between domain complexity and parameterization quality is challenging.", + "bbox": [ + 212, + 618, + 787, + 709 + ], + "page_idx": 12 + }, + { + "type": "text", + "text": "5 Conclusion", + "text_level": 1, + "bbox": [ + 215, + 747, + 359, + 763 + ], + "page_idx": 12 + }, + { + "type": "text", + "text": "We introduce a novel neural algorithm for parameterizing 3D surfaces represented by neural implicit functions to a user-defined parametric domain, such as spheres and polycubes. This parametric domain is learned given some user-specified hyperparameters. Utilizing bi-directional deformation, our method is", + "bbox": [ + 212, + 779, + 787, + 840 + ], + "page_idx": 12 + }, + { + "type": "header", + "text": "NeuParam", + "bbox": [ + 658, + 114, + 730, + 126 + ], + "page_idx": 12 + }, + { + "type": "page_number", + "text": "13", + "bbox": [ + 767, + 114, + 785, + 126 + ], + "page_idx": 12 + }, + { + "type": "image", + "img_path": "images/9a8a79ca679bf74a26ce4837890ada73372ce0d612bd80b450abff0695bf7d4e.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 232, + 143, + 300, + 202 + ], + "page_idx": 13 + }, + { + "type": "image", + "img_path": "images/9350fc57010a9ffe32c19e532f83c38ddc9769cbe4c421a1c822f1b291a6c606.jpg", + "image_caption": [ + "VolSDF 4.65" + ], + "image_footnote": [], + "bbox": [ + 310, + 146, + 375, + 200 + ], + "page_idx": 13 + }, + { + "type": "image", + "img_path": "images/a39ec8569b33187b8e52116a3b2a0d67eb05cd497e7d06eb6a2aa0e8a0270501.jpg", + "image_caption": [ + "Ours 4.87" + ], + "image_footnote": [], + "bbox": [ + 388, + 146, + 452, + 200 + ], + "page_idx": 13 + }, + { + "type": "image", + "img_path": "images/3c31f91daa790caed9c7085953902d81c1300895a374af3a9709e3a744093d2d.jpg", + "image_caption": [ + "Inverse 4.89" + ], + "image_footnote": [], + "bbox": [ + 464, + 148, + 529, + 200 + ], + "page_idx": 13 + }, + { + "type": "image", + "img_path": "images/e9ea5a3ac843f9e224dd2479efdfaa287aba6c31cff943a2d987f65ca045e5e1.jpg", + "image_caption": [ + "GT", + "Input" + ], + "image_footnote": [], + "bbox": [ + 228, + 218, + 303, + 260 + ], + "page_idx": 13 + }, + { + "type": "image", + "img_path": "images/fb117f9fd62ebcfb351490d013e8ecdb45130a87494e121accb2a93fc82dbb93.jpg", + "image_caption": [ + "w/o $\\mathcal{L}_{\\mathrm{smooth}}\\& \\mathcal{L}_{\\mathrm{Lap}} + \\mathcal{L}_{\\mathrm{smooth}}$" + ], + "image_footnote": [], + "bbox": [ + 310, + 226, + 380, + 260 + ], + "page_idx": 13 + }, + { + "type": "image", + "img_path": "images/f7aa11f928aa7ae118782a7d42896ee4386646e259d227c73604894ec405ac82.jpg", + "image_caption": [ + "Fig. 7: Ablation studies. On the left, we calculate the CD $(\\times 10^{-3})$ based on the point cloud extracted from the inverse deformation module and present the ablation studies for $\\mathcal{L}_{\\mathrm{smooth}}$ and $\\mathcal{L}_{\\mathrm{Lap}}$ . On the right, utilization of the Laplacian loss has proven effective in reducing angle distortion $(\\mathcal{E}_{\\mathrm{angle}})$ , especially in regions with high curvature." + ], + "image_footnote": [], + "bbox": [ + 398, + 231, + 457, + 261 + ], + "page_idx": 13 + }, + { + "type": "image", + "img_path": "images/1ca98573884c0f11a5df2f798dd8cdfdd53001a0186d47812f92c83625035d5a.jpg", + "image_caption": [ + "+Lap" + ], + "image_footnote": [], + "bbox": [ + 485, + 231, + 532, + 261 + ], + "page_idx": 13 + }, + { + "type": "image", + "img_path": "images/4edd16536b809a1e699e0cfef5e3ff8d76596289f56217a5ac6f23bee6d0441e.jpg", + "image_caption": [], + "image_footnote": [ + "$\\mathcal{E}_{\\mathrm{angle}} = 1.147$", + "$\\mathcal{E}_{\\mathrm{angle}} = 1.092$", + "w/o $\\mathcal{L}_{\\mathrm{Lap}}$", + "W/Lap" + ], + "bbox": [ + 545, + 148, + 754, + 246 + ], + "page_idx": 13 + }, + { + "type": "text", + "text": "capable of learning a nearly bijective mapping without relying on any prior knowledge, while controlling angle distortion through the usage of a Laplacian regularizer. Furthermore, our approach seamlessly integrates with the neural rendering pipeline, enabling the reconstruction of 3D objects from multi-view images and facilitating efficient volume rendering of modified textures and shadings — eliminating the necessity for network re-training. We demonstrated the efficacy of our method on human heads and man-made objects.", + "bbox": [ + 212, + 375, + 784, + 479 + ], + "page_idx": 13 + }, + { + "type": "text", + "text": "Our method has a few limitations that require further improvement in the future. First, quantitative comparisons on the FS-Syn dataset show that the reconstruction quality of our method is slightly worse than the standard SDF-based neural rendering algorithms, such as VolSDF [58] and NeuS [46]. For example, as illustrated in Fig. 8 (b) and, We report PSNR and CD compared with VolSDF [58] on Fs-Syn dataset [56]. This decline in visual quality arises from the incorporation of additional modules, such as bi-directional deformation and radiance decomposition, resulting in the cessation of updates to the SDF network within our neural parameterization pipeline. Therefore, computing a parameterization without compromising the reconstruction quality is highly desired. Second, our radiance decomposition operates under the assumptions of Lambertian reflectance and grayscale shading, which might hinder its effectiveness when dealing with intricate materials or sophisticated shadings.", + "bbox": [ + 212, + 479, + 787, + 676 + ], + "page_idx": 13 + }, + { + "type": "image", + "img_path": "images/8d4289f81d14eb1c9ce65cadae2ab9b437dbd150d0a4bed6e090a50fb327053c.jpg", + "image_caption": [ + "S", + "$\\mathcal{E}_{\\mathrm{area}} = 2.505$ 6.094", + "Fig. 8: We present area distortion $\\mathcal{E}_{\\mathrm{area}}$ based on the different choices of the parametric domain in (a) and geometry reconstruction results of VolSDF [58] and our method in (b). Our method is capable of editing through parameterization, accompanied by an acceptable compromise in reconstruction quality." + ], + "image_footnote": [], + "bbox": [ + 227, + 704, + 287, + 741 + ], + "page_idx": 13 + }, + { + "type": "image", + "img_path": "images/6b5e247eb1b1ef116e7444d83f3ffe4f73855ff430a8ffe7d3a2cf17f0a23e71.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 297, + 705, + 343, + 741 + ], + "page_idx": 13 + }, + { + "type": "image", + "img_path": "images/babc350263d4387380837fde4a940c8d855b6b527ed7a4ad8e1d8ab8d1f65b4d.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 354, + 705, + 413, + 741 + ], + "page_idx": 13 + }, + { + "type": "image", + "img_path": "images/86a7b75151500e7d032cd56115263caea9918a4822ef798c1076b4c1bd5b8205.jpg", + "image_caption": [ + "1.633" + ], + "image_footnote": [], + "bbox": [ + 431, + 703, + 478, + 741 + ], + "page_idx": 13 + }, + { + "type": "image", + "img_path": "images/256ec429d93fffdd5d89afb162e25fd169930e7e1483ae18b86645bd1bf6c960.jpg", + "image_caption": [ + "GT" + ], + "image_footnote": [], + "bbox": [ + 490, + 704, + 519, + 739 + ], + "page_idx": 13 + }, + { + "type": "image", + "img_path": "images/6ced61452f26507a3bc11e4290547712db9c5517aebeea9afd80d1f17f459ffb.jpg", + "image_caption": [ + "VolSDF" + ], + "image_footnote": [], + "bbox": [ + 524, + 705, + 552, + 739 + ], + "page_idx": 13 + }, + { + "type": "image", + "img_path": "images/384f501987360754e48a3213a7bc6e07ed05d784b086a25116648b59fdcf736c.jpg", + "image_caption": [ + "Ours", + "(b)" + ], + "image_footnote": [], + "bbox": [ + 593, + 705, + 620, + 741 + ], + "page_idx": 13 + }, + { + "type": "image", + "img_path": "images/a83df78bc300a819ef6d625e81ee15521b238974721cba3f29a13da97eec0f34.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 627, + 708, + 656, + 739 + ], + "page_idx": 13 + }, + { + "type": "text", + "text": "Method PSNR CD(10-3)", + "bbox": [ + 663, + 715, + 769, + 724 + ], + "page_idx": 13 + }, + { + "type": "text", + "text": "VolSDF 31.15 3.270", + "bbox": [ + 665, + 728, + 759, + 738 + ], + "page_idx": 13 + }, + { + "type": "text", + "text": "Ours 30.87 3.423", + "bbox": [ + 669, + 741, + 761, + 752 + ], + "page_idx": 13 + }, + { + "type": "page_number", + "text": "14", + "bbox": [ + 217, + 114, + 235, + 126 + ], + "page_idx": 13 + }, + { + "type": "header", + "text": "B. Xu et al.", + "bbox": [ + 271, + 114, + 352, + 126 + ], + "page_idx": 13 + }, + { + "type": "text", + "text": "Acknowledgment", + "text_level": 1, + "bbox": [ + 217, + 143, + 382, + 162 + ], + "page_idx": 14 + }, + { + "type": "text", + "text": "This study is supported under the RIE2020 Industry Alignment Fund - Industry Collaboration Projects (IAF-ICP) Funding Initiative, as well as cash and in-kind contribution from the industry partner(s). This project is also partially supported by the Ministry of Education, Singapore, under its Academic Research Fund Grants (MOE-T2EP20220-0005 & RT19/22).", + "bbox": [ + 212, + 176, + 787, + 253 + ], + "page_idx": 14 + }, + { + "type": "text", + "text": "References", + "text_level": 1, + "bbox": [ + 217, + 277, + 321, + 292 + ], + "page_idx": 14 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "1. Bednarik, J., Parashar, S., Gundogdu, E., Salzmann, M., Fua, P.: Shape reconstruction by learning differentiable surface representations. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 4716-4725 (2020) 3", + "2. Biermann, H., Martin, I., Bernardini, F., Zorin, D.: Cut-and-paste editing of multiresolution surfaces. ACM transactions on graphics (TOG) 21(3), 312-321 (2002) 3", + "3. Degener, P., Meseth, J., Klein, R.: An adaptable surface parameterization method. IMR 3, 201-213 (2003) 10, 13", + "4. Fan, Q., Yang, J., Hua, G., Chen, B., Wipf, D.: Revisiting deep intrinsic image decompositions. In: Proceedings of the IEEE conference on computer vision and pattern recognition. pp. 8944-8952 (2018) 8", + "5. Fang, H., Hart, J.C.: Textureshop: texture synthesis as a photograph editing tool. ACM Transactions on Graphics (TOG) 23(3), 354-359 (2004) 3", + "6. Floater, M.S., Hormann, K.: Surface parameterization: a tutorial and survey. Advances in multiresolution for geometric modelling pp. 157-186 (2005) 3, 8", + "7. García, I., Xia, J., He, Y., Xin, S., Patow, G.: Interactive applications for sketch-based editable polycube map. IEEE Trans. Vis. Comput. Graph. 19(7), 1158-1171 (2013) 3", + "8. Gotsman, C., Gu, X., Sheffer, A.: Fundamentals of spherical parameterization for 3d meshes. ACM Trans. Graph. 22(3), 358-363 (2003) 3", + "9. Greene, N.: Environment mapping and other applications of world projections. IEEE computer graphics and Applications 6(11), 21-29 (1986) 12", + "0. Gropp, A., Yariv, L., Haim, N., Atzmon, M., Lipman, Y.: Implicit geometric regularization for learning shapes. In: International Conference on Machine Learning. pp. 3789-3799. PMLR (2020) 9", + "1. Groueix, T., Fisher, M., Kim, V.G., Russell, B.C., Aubry, M.: A papier-mâché approach to learning 3d surface generation. In: Proceedings of the IEEE conference on computer vision and pattern recognition. pp. 216-224 (2018) 3", + "2. Gu, X., Wang, Y., Chan, T.F., Thompson, P.M., Yau, S.: Genus zero surface conformal mapping and its application to brain surface mapping. IEEE Trans. Medical Imaging 23(8), 949-958 (2004) 8", + "3. Gu, X., Yau, S.: Global conformal parameterization. In: Kobbelt, L., Schröder, P., Hoppe, H. (eds.) First Eurographics Symposium on Geometry Processing. vol. 43, pp. 127-137 (2003) 8", + "4. Guo, H., Liu, S., Pan, H., Liu, Y., Tong, X., Guo, B.: Complexgen: Cad reconstruction by b-rep chain complex generation. ACM Transactions on Graphics (TOG) 41(4), 1-18 (2022) 3" + ], + "bbox": [ + 225, + 310, + 785, + 839 + ], + "page_idx": 14 + }, + { + "type": "header", + "text": "NeuParam", + "bbox": [ + 658, + 114, + 730, + 126 + ], + "page_idx": 14 + }, + { + "type": "page_number", + "text": "15", + "bbox": [ + 767, + 114, + 785, + 126 + ], + "page_idx": 14 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "15. He, Y., Wang, H., Fu, C.W., Qin, H.: A divide-and-conquer approach for automatic polycube map construction. Computers & Graphics 33(3), 369-380 (2009) 10, 13", + "16. Huang, Y.H., He, Y., Yuan, Y.J., Lai, Y.K., Gao, L.: Stylizednerf: Consistent 3d scene stylization as stylized nerf via 2d-3d mutual learning. In: Computer Vision and Pattern Recognition (CVPR) (2022) 4", + "17. Kreisselmeier, G., Steinhauser, R.: Systematic control design by optimizing a vector performance index. In: Computer aided design of control systems, pp. 113-117. Elsevier (1980) 6", + "18. Kuang, Z., Luan, F., Bi, S., Shu, Z., Wetzstein, G., Sunkavalli, K.: Palettenerf: Palette-based appearance editing of neural radiance fields. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 20691-20700 (2023) 2, 4", + "19. Li, M., Zhang, H.: D2im-net: Learning detail disentangled implicit fields from single images. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 10246-10255 (2021) 4", + "20. Li, Z., Müller, T., Evans, A., Taylor, R.H., Unberath, M., Liu, M.Y., Lin, C.H.: Neuralangelo: High-fidelity neural surface reconstruction. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 8456-8465 (2023) 4", + "21. Li, Z., Shi, J., Bi, S., Zhu, R., Sunkavalli, K., Hasan, M., Xu, Z., Ramamoorthi, R., Chandraker, M.: Physically-based editing of indoor scene lighting from a single image. In: European Conference on Computer Vision. pp. 555-572. Springer (2022) 4", + "22. Lin, C., Mitra, N., Wetzstein, G., Guibas, L.J., Guerrero, P.: Neuform: Adaptive overfitting for neural shape editing. Advances in Neural Information Processing Systems 35, 15217-15229 (2022) 4", + "23. Liu, S., Zhang, X., Zhang, Z., Zhang, R., Zhu, J.Y., Russell, B.: Editing conditional radiance fields. In: Proceedings of the IEEE/CVF international conference on computer vision. pp. 5773-5783 (2021) 2", + "24. Low, W.F., Lee, G.H.: Minimal neural atlas: Parameterizing complex surfaces with minimal charts and distortion. In: European Conference on Computer Vision. pp. 465-481. Springer (2022) 3", + "25. Ma, L., Li, X., Liao, J., Wang, X., Zhang, Q., Wang, J., Sander, P.V.: Neural parameterization for dynamic human head editing. ACM Transactions on Graphics (TOG) 41(6), 1-15 (2022) 2, 4, 8, 9, 11, 12", + "26. Mescheder, L., Oechsle, M., Niemeyer, M., Nowozin, S., Geiger, A.: Occupancy networks: Learning 3d reconstruction in function space. In: Proceedings of the IEEE/CVF conference on computer vision and pattern recognition. pp. 4460-4470 (2019) 4", + "27. Mildenhall, B., Srinivasan, P.P., Tancik, M., Barron, J.T., Ramamoorthi, R., Ng, R.: Nerf: Representing scenes as neural radiance fields for view synthesis. In: ECCV (2020) 1, 7, 8", + "28. Müller, T., Evans, A., Schied, C., Keller, A.: Instant neural graphics primitives with a multiresolution hash encoding. ACM Transactions on Graphics (ToG) 41(4), 1-15 (2022) 4", + "29. Niemeyer, M., Geiger, A.: Giraffe: Representing scenes as compositional generative neural feature fields. In: Proc. IEEE Conf. on Computer Vision and Pattern Recognition (CVPR) (2021) 4", + "30. Oechsle, M., Peng, S., Geiger, A.: Unisurf: Unifying neural implicit surfaces and radiance fields for multi-view reconstruction. In: Proceedings of the IEEE/CVF International Conference on Computer Vision. pp. 5589-5599 (2021) 1" + ], + "bbox": [ + 215, + 146, + 785, + 840 + ], + "page_idx": 15 + }, + { + "type": "page_number", + "text": "16", + "bbox": [ + 217, + 114, + 235, + 126 + ], + "page_idx": 15 + }, + { + "type": "header", + "text": "B. Xu et al.", + "bbox": [ + 271, + 114, + 351, + 126 + ], + "page_idx": 15 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "31. Park, J.J., Florence, P., Straub, J., Newcombe, R., Lovegrove, S.: Deepsdf: Learning continuous signed distance functions for shape representation. In: Proceedings of the IEEE/CVF conference on computer vision and pattern recognition. pp. 165-174 (2019) 4, 9", + "32. Pérez, P., Gangnet, M., Blake, A.: Poisson image editing. In: Seminal Graphics Papers: Pushing the Boundaries, Volume 2, pp. 577-582 (2023) 4", + "33. Praun, E., Hoppe, H.: Spherical parametrization and remeshing. ACM transactions on graphics (TOG) 22(3), 340-349 (2003) 3", + "34. Qi, C.R., Su, H., Mo, K., Guibas, L.J.: Pointnet: Deep learning on point sets for 3d classification and segmentation. In: Proceedings of the IEEE conference on computer vision and pattern recognition. pp. 652-660 (2017) 3", + "35. Qi, C.R., Su, H., Nießner, M., Dai, A., Yan, M., Guibas, L.J.: Volumetric and multi-view cnns for object classification on 3d data. In: Proceedings of the IEEE conference on computer vision and pattern recognition. pp. 5648-5656 (2016) 3", + "36. Ramon, E., Triginer, G., Escur, J., Pumarola, A., Garcia, J., Giro-i Nieto, X., Moreno-Noguer, F.: H3d-net: Few-shot high-fidelity 3d head reconstruction. In: Proceedings of the IEEE/CVF International Conference on Computer Vision. pp. 5620-5629 (2021) 10, 11", + "37. Rosu, R.A., Behnke, S.: Permutosdf: Fast multi-view reconstruction with implicit surfaces using permutohedral lattices. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 8466-8475 (2023) 4", + "38. Sagnik Das, Ke Ma, Z.S., Samaras, D.: Learning an isometric surface parameterization for texture unwrapping. In: European Conference of Computer Vision 2022, ECCV 2022, Tel Aviv, Israel, October 23-27, 2022. Springer (2022) 4", + "39. Saito, S., Huang, Z., Natsume, R., Morishima, S., Kanazawa, A., Li, H.: Pifu: Pixel-aligned implicit function for high-resolution clothed human digitization. In: Proceedings of the IEEE/CVF international conference on computer vision. pp. 2304-2314 (2019) 4", + "40. Sheffer, A., Praun, E., Rose, K., et al.: Mesh parameterization methods and their applications. Foundations and Trends in Computer Graphics and Vision 2(2), 105-171 (2007) 3", + "41. Srinivasan, P.P., Deng, B., Zhang, X., Tancik, M., Mildenhall, B., Barron, J.T.: Nerv: Neural reflectance and visibility fields for relighting and view synthesis. In: CVPR (2021) 4", + "42. Sun, Q., Zhang, L., Zhang, M., Ying, X., Xin, S., Xia, J., He, Y.: Texture brush: an interactive surface texturing interface. pp. 153-160 (2013) 3", + "43. Tarini, M., Hormann, K., Cignoni, P., Montani, C.: Polycube-maps. ACM transactions on graphics (TOG) 23(3), 853-860 (2004) 3", + "44. Tojo, K., Umetani, N.: Recolorable posterization of volumetric radiance fields using visibility-weighted palette extraction. In: Computer Graphics Forum. vol. 41, pp. 149-160. Wiley Online Library (2022) 2, 4", + "45. Wang, N., Zhang, Y., Li, Z., Fu, Y., Liu, W., Jiang, Y.G.: Pixel2mesh: Generating 3d mesh models from single rgb images. In: Proceedings of the European conference on computer vision (ECCV). pp. 52-67 (2018) 3", + "46. Wang, P., Liu, L., Liu, Y., Theobalt, C., Komura, T., Wang, W.: Neus: Learning neural implicit surfaces by volume rendering for multi-view reconstruction. NeurIPS (2021) 1, 14", + "47. Wang, X., Zhu, J., Ye, Q., Huo, Y., Ran, Y., Zhong, Z., Chen, J.: Seal-3d: Interactive pixel-level editing for neural radiance fields (2023) 2, 4" + ], + "bbox": [ + 212, + 146, + 787, + 839 + ], + "page_idx": 16 + }, + { + "type": "header", + "text": "NeuParam", + "bbox": [ + 658, + 114, + 730, + 126 + ], + "page_idx": 16 + }, + { + "type": "page_number", + "text": "17", + "bbox": [ + 767, + 114, + 785, + 126 + ], + "page_idx": 16 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "48. Wang, Y., Rahmann, L., Sorkine-Hornung, O.: Geometry-consistent neural shape representation with implicit displacement fields. In: The Tenth International Conference on Learning Representations. OpenReview (2022) 4", + "49. Wang, Y., Skorokhodov, I., Wonka, P.: Hf-neus: Improved surface reconstruction using high-frequency details. Advances in Neural Information Processing Systems 35, 1966-1978 (2022) 4", + "50. Williams, F., Schneider, T., Silva, C., Zorin, D., Bruna, J., Panozzo, D.: Deep geometric prior for surface reconstruction. In: Proceedings of the IEEE/CVF conference on computer vision and pattern recognition. pp. 10130-10139 (2019) 3", + "51. Wu, T., Zhang, J., Fu, X., Wang, Y., Ren, J., Pan, L., Wu, W., Yang, L., Wang, J., Qian, C., Lin, D., Liu, Z.: Omniobject3d: Large-vocabulary 3d object dataset for realistic perception, reconstruction and generation (2023) 10", + "52. Xiang, F., Xu, Z., Hasan, M., Hold-Geoffroy, Y., Sunkavalli, K., Su, H.: NeuTex: Neural Texture Mapping for Volumetric Neural Rendering. In: The IEEE Conference on Computer Vision and Pattern Recognition (CVPR) (2021) 2, 4, 7", + "53. Xu, B., Zhang, J., Lin, K.Y., Qian, C., He, Y.: Deformable model driven neural rendering for high-fidelity 3d reconstruction of human heads under low-view settings. In: Proceedings of the IEEE/CVF International Conference on Computer Vision (2023) 2, 4", + "54. Yang, B., Bao, C., Zeng, J., Bao, H., Zhang, Y., Cui, Z., Zhang, G.: Neumesh: Learning disentangled neural mesh-based implicit field for geometry and texture editing. In: European Conference on Computer Vision. pp. 597-614. Springer (2022) 4, 12", + "55. Yang, B., Zhang, Y., Xu, Y., Li, Y., Zhou, H., Bao, H., Zhang, G., Cui, Z.: Learning object-compositional neural radiance field for editable scene rendering. In: International Conference on Computer Vision (ICCV) (October 2021) 4", + "56. Yang, H., Zhu, H., Wang, Y., Huang, M., Shen, Q., Yang, R., Cao, X.: Facescape: a large-scale high quality 3d face dataset and detailed riggable 3d face prediction. In: Proceedings of the IEEE/cvf conference on computer vision and pattern recognition. pp. 601-610 (2020) 10, 14", + "57. Yang, L., Liang, Y., Li, X., Zhang, C., Lin, G., Sheffer, A., Schaefer, S., Keyser, J., Wang, W.: Neural parametric surfaces for shape modeling. arXiv preprint arXiv:2309.09911 (2023) 3", + "58. Yariv, L., Gu, J., Kasten, Y., Lipman, Y.: Volume rendering of neural implicit surfaces. Advances in Neural Information Processing Systems 34, 4805-4815 (2021) 1, 9, 14", + "59. Yariv, L., Kasten, Y., Moran, D., Galun, M., Atzmon, M., Ronen, B., Lipman, Y.: Multiview neural surface reconstruction by disentangling geometry and appearance. Advances in Neural Information Processing Systems 33 (2020) 9", + "60. Ye, W., Chen, S., Bao, C., Bao, H., Pollefeys, M., Cui, Z., Zhang, G.: IntrinsicNeRF: Learning Intrinsic Neural Radiance Fields for Editable Novel View Synthesis. In: Proceedings of the IEEE/CVF International Conference on Computer Vision (2023) 4, 8", + "61. Yenamandra, T., Tewari, A., Bernard, F., Seidel, H.P., Elgharib, M., Cremers, D., Theobalt, C.: i3dmm: Deep implicit 3d morphable model of human heads. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 12803-12813 (2021) 2, 4", + "62. Yuan, Y.J., Sun, Y.T., Lai, Y.K., Ma, Y., Jia, R., Kobbelt, L., Gao, L.: Interactive nef geometry editing with shape priors. IEEE Transactions on Pattern Analysis and Machine Intelligence (2023) 4" + ], + "bbox": [ + 215, + 147, + 785, + 840 + ], + "page_idx": 17 + }, + { + "type": "page_number", + "text": "18", + "bbox": [ + 217, + 114, + 235, + 126 + ], + "page_idx": 17 + }, + { + "type": "header", + "text": "B. Xu et al.", + "bbox": [ + 271, + 114, + 351, + 126 + ], + "page_idx": 17 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "63. Zhang, Q., Hou, J., Qian, Y., Chan, A.B., Zhang, J., He, Y.: Reggeonet: Learning regular representations for large-scale 3d point clouds. Int. J. Comput. Vis. 130(12), 3100-3122 (2022) 3", + "64. Zhang, Q., Hou, J., Qian, Y., Zeng, Y., Zhang, J., He, Y.: Flattening-net: Deep regular 2d representation for 3d point cloud analysis. IEEE Trans. Pattern Anal. Mach. Intell. 45(8), 9726-9742 (2023) 3" + ], + "bbox": [ + 215, + 146, + 787, + 231 + ], + "page_idx": 18 + }, + { + "type": "header", + "text": "NeuParam", + "bbox": [ + 658, + 114, + 730, + 126 + ], + "page_idx": 18 + }, + { + "type": "page_number", + "text": "19", + "bbox": [ + 767, + 114, + 785, + 126 + ], + "page_idx": 18 + } +] \ No newline at end of file diff --git a/2024/Parameterization-driven Neural Surface Reconstruction for Object-oriented Editing in Neural Rendering/20a75bf0-0a6a-4ea4-b1e9-a52604587383_model.json b/2024/Parameterization-driven Neural Surface Reconstruction for Object-oriented Editing in Neural Rendering/20a75bf0-0a6a-4ea4-b1e9-a52604587383_model.json new file mode 100644 index 0000000000000000000000000000000000000000..003e20c61cb8c4c5c0f0715442528747f47af9db --- /dev/null +++ b/2024/Parameterization-driven Neural Surface Reconstruction for Object-oriented Editing in Neural Rendering/20a75bf0-0a6a-4ea4-b1e9-a52604587383_model.json @@ -0,0 +1,3208 @@ +[ + [ + { + "type": "title", + "bbox": [ + 0.242, + 0.141, + 0.761, + 0.21 + ], + "angle": 0, + "content": "Parameterization-driven Neural Surface Reconstruction for Object-oriented Editing in Neural Rendering" + }, + { + "type": "text", + "bbox": [ + 0.223, + 0.235, + 0.778, + 0.266 + ], + "angle": 0, + "content": "Baixin \\(\\mathrm{Xu}^{1}\\), Jiangbei \\(\\mathrm{Hu}^{2,3}\\), Fei \\(\\mathrm{Hou}^{4,5}\\), Kwan-Yee \\(\\mathrm{Lin}^{6}\\), Wayne \\(\\mathrm{Wu}^{7}\\), Chen Qian\\(^{7}\\), and Ying \\(\\mathrm{He}^{1,2}\\)" + }, + { + "type": "text", + "bbox": [ + 0.216, + 0.278, + 0.785, + 0.348 + ], + "angle": 0, + "content": "\\(^{1}\\)S-Lab, Nanyang Technological University, \\(^{2}\\)College of Computing and Data Science, Nanyang Technological University, \\(^{3}\\)Dalian University of Technology, \\(^{4}\\)Key Laboratory of System Software (Chinese Academy of Sciences) and State Key Laboratory of Computer Science, Institute of Software, Chinese Academy of Sciences, \\(^{5}\\)University of Chinese Academy of Sciences, \\(^{6}\\)Shanghai AI Lab, \\(^{7}\\)SenseTime Research" + }, + { + "type": "text", + "bbox": [ + 0.262, + 0.388, + 0.74, + 0.651 + ], + "angle": 0, + "content": "Abstract. The advancements in neural rendering have increased the need for techniques that enable intuitive editing of 3D objects represented as neural implicit surfaces. This paper introduces a novel neural algorithm for parameterizing neural implicit surfaces to simple parametric domains like spheres and polycubes. Our method allows users to specify the number of cubes in the parametric domain, learning a configuration that closely resembles the target 3D object's geometry. It computes bidirectional deformation between the object and the domain using a forward mapping from the object's zero level set and an inverse deformation for backward mapping. We ensure nearly bijective mapping with a cycle loss and optimize deformation smoothness. The parameterization quality, assessed by angle and area distortions, is guaranteed using a Laplacian regularizer and an optimized learned parametric domain. Our framework integrates with existing neural rendering pipelines, using multi-view images of a single object or multiple objects of similar geometries to reconstruct 3D geometry and compute texture maps automatically, eliminating the need for any prior information. We demonstrate the method's effectiveness on images of human heads and man-made objects. The source code is available at https://xubaixinxbx.github.io/neuparam." + }, + { + "type": "text", + "bbox": [ + 0.261, + 0.665, + 0.741, + 0.694 + ], + "angle": 0, + "content": "Keywords: Polycube parameterization \\(\\cdot\\) Neural surface reconstruction and editing \\(\\cdot\\) Neural rendering" + }, + { + "type": "title", + "bbox": [ + 0.216, + 0.721, + 0.377, + 0.736 + ], + "angle": 0, + "content": "1 Introduction" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.754, + 0.788, + 0.816 + ], + "angle": 0, + "content": "Neural radiance fields (NeRF) [27] have garnered remarkable success in both the computer vision and computer graphics communities, redefining benchmarks for high-quality renderings and novel view synthesis. Building upon NeRF, a variety of methods leveraging implicit neural representations [30, 46, 58] have emerged," + }, + { + "type": "page_footnote", + "bbox": [ + 0.232, + 0.826, + 0.612, + 0.841 + ], + "angle": 0, + "content": "Corresponding author: Ying He (email: yhe@ntu.edu.sg)." + } + ], + [ + { + "type": "page_number", + "bbox": [ + 0.218, + 0.116, + 0.23, + 0.127 + ], + "angle": 0, + "content": "2" + }, + { + "type": "header", + "bbox": [ + 0.272, + 0.115, + 0.353, + 0.128 + ], + "angle": 0, + "content": "B. Xu et al." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.146, + 0.788, + 0.42 + ], + "angle": 0, + "content": "delivering high-fidelity 3D reconstructions. The demand for intuitive 3D object editing techniques has increased with the growing capabilities of neural rendering, especially for neural implicit surfaces. Extensive explorations have been made in editing within the Neural Rendering framework to achieve diverse visual effects. By manipulating the latent codes of the network, one can implicitly alter the shape and color of 3D objects [23,53,61]. In the work of [18,44], palette-based methods have been employed to facilitate transformations in the color and style of rendered scenes. However, these methods suffer from limited controllability and disallow local pixel editing. Seal3D [47] proposed a teacher-student training strategy to edit the interactive scene at the pixel level. Yet, this method is tricky to handle shading, and each subsequent edit required a new training iteration. By employing neural networks to formulate a differentiable surface parameterization, it can be facilitated to edit textures or shapes within the neural rendering pipeline [25,52]. Existing neural parameterization techniques still face significant challenges, such as the need for appropriate distortion constraints to ensure accurate detail reconstruction or reducing the reliance on prior information from tracked mesh and UV mapping. The final quality of edited visual effects is significantly impacted by these issues." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.433, + 0.789, + 0.844 + ], + "angle": 0, + "content": "This paper aims to develop an intuitive and easy-to-use tool for appearance editing in neural implicit surfaces within the neural rendering framework, utilizing only 2D images as input. Towards this goal, we propose a novel method to parameterize neural implicit surfaces to parametric domains where the radiance field is represented as a texture map. Distinguished from previous work that using 2D texture mapping, our neural parameterization algorithm supports the selection of different 3D parametric domains according to geometric shapes to avoid seams and reduce distortion, such as spheres, and polycubes. Technically, our parameterization algorithm utilizes learnable parameters to fit a parametric domain first and a neural network to learn a bi-directional deformation between 3D objects and the chosen parametric domains. This involves a forward deformation that maps points from the zero level set of the neural implicit surface to the parametric domain, followed by an inverse deformation, mapping points backward. Notably, we do not require any explicit prior information in learning both deformations. We employ a cycle loss to ensure the smoothness of the bi-directional deformation, and a Laplacian regularization to effectively control angle distortion. With the parameterization, 3D radiance field is mapped to a simple 3D parametric domain which can be further disassembled into a 2D domain, facilitating visualization and various editing tasks. Furthermore, we decompose the 2D radiance field into two components: view-independent material and view-dependent shading, streamlining both texture and shading editing. Our neural parameterization algorithm is fully compatible with existing neural rendering pipelines, allowing 3D reconstruction from multi-view images as input and the creation of texture maps simultaneously. It allows for the immediate rendering of edited textures through volume rendering, without network re-training. Moreover, it supports co-parameterization of objects of similar geometry and enables texture transfer between them. We validate the effectiveness of" + } + ], + [ + { + "type": "header", + "bbox": [ + 0.66, + 0.115, + 0.732, + 0.128 + ], + "angle": 0, + "content": "NeuParam" + }, + { + "type": "page_number", + "bbox": [ + 0.775, + 0.117, + 0.785, + 0.127 + ], + "angle": 0, + "content": "3" + }, + { + "type": "text", + "bbox": [ + 0.217, + 0.147, + 0.786, + 0.176 + ], + "angle": 0, + "content": "our method on scenes of human heads and man-made objects. Our contributions are summarized as follows:" + }, + { + "type": "text", + "bbox": [ + 0.226, + 0.189, + 0.788, + 0.279 + ], + "angle": 0, + "content": "- We present a neural parameterization framework that automatically learns a parametric domain and computes bi-directional deformation between 3D objects and their parametric domains, eliminating the need for any prior information from tracked mesh or UV mapping. Our approach utilizes Laplacian loss to minimize angle distortion and allows the user to specify the domain complexity to control area distortion." + }, + { + "type": "text", + "bbox": [ + 0.226, + 0.281, + 0.788, + 0.326 + ], + "angle": 0, + "content": "- We introduce a simple yet effective technique for intrinsic radiance decomposition, facilitating both view-independent texture editing and view-dependent shading editing." + }, + { + "type": "text", + "bbox": [ + 0.226, + 0.327, + 0.788, + 0.401 + ], + "angle": 0, + "content": "- Our neural parameterization framework is end-to-end and fully compatible with existing neural rendering architectures, accepting multi-view images as input to reconstruct 3D geometry and generate UV maps as output. Additionally, it enables the direct rendering of modified textures using the volume rendering pipeline." + }, + { + "type": "text", + "bbox": [ + 0.226, + 0.403, + 0.788, + 0.433 + ], + "angle": 0, + "content": "- Our approach supports co-parameterization of multiple objects and allows for texture transfer between different objects." + }, + { + "type": "list", + "bbox": [ + 0.226, + 0.189, + 0.788, + 0.433 + ], + "angle": 0, + "content": null + }, + { + "type": "title", + "bbox": [ + 0.217, + 0.458, + 0.388, + 0.473 + ], + "angle": 0, + "content": "2 Related Work" + }, + { + "type": "text", + "bbox": [ + 0.216, + 0.492, + 0.788, + 0.794 + ], + "angle": 0, + "content": "Parameterization. Surface parameterization [6, 40] aims at computing a bijective mapping between the 3D surface and a suitable parametric domain, usually a 2D region or a simple 3D object, such as spheres [8] and polycubes [7, 43]. Serving as an important computational tool in computer graphics and digital geometry processing [40], surface parameterization facilitates various applications, including texture editing [5], surface painting [42], details transfer [2], remeshing [33], among others. Neural networks are now commonly used in digital geometry processing, resulting in the creation of parameterization algorithms based on deep learning techniques [1, 11, 14, 50]. Unlike classical methods that yield global and seamless parametrization, most neural parameterization methods are computing local parametrization [11, 63, 64]. These methods typically partition the 3D model into multiple patches and parameterize each to a 2D region, possibly followed by a post-processing step to stitch the patches together. Although these methods are efficient and can handle surfaces of arbitrary geometry and topology, they frequently encounter problems including non-bijectivity, a lack of smoothness, the presence of seams and overlaps, and large distortions. There are also works on neural parametric surfaces, defined in rectangular domains [24] and \\(n\\)-sided patches [57]. They can model complex surface geometries with high precision; however, such representations are not applicable to neural implicit surfaces and thereby incompatible with neural rendering." + }, + { + "type": "text", + "bbox": [ + 0.216, + 0.796, + 0.788, + 0.843 + ], + "angle": 0, + "content": "Neural Implicit Functions. Recent years have seen a surge in the successful application of neural implicit functions in 3D deep learning. Compared to explicit representations such as point clouds [34], voxels [35], and meshes [45]," + } + ], + [ + { + "type": "page_number", + "bbox": [ + 0.218, + 0.116, + 0.23, + 0.127 + ], + "angle": 0, + "content": "4" + }, + { + "type": "header", + "bbox": [ + 0.272, + 0.115, + 0.353, + 0.128 + ], + "angle": 0, + "content": "B. Xu et al." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.147, + 0.788, + 0.268 + ], + "angle": 0, + "content": "neural implicit representations offer unique advantages including flexibility, continuity, and robustness [26,31,39]. When integrated within the neural rendering pipeline, neural implicit surfaces exhibit exceptional quality in 3D reconstructions from multi-view images [19,20,28,37,48,49,53]. However, as geometries and radiance fields are encoded as network parameters, editing them becomes complex and non-intuitive, compared to explicit representations. This paper aims at tackling this challenge by explicitly representing radiance fields as texture mapping that encompasses both view-dependent and view-independent components." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.269, + 0.789, + 0.601 + ], + "angle": 0, + "content": "NeRF Editing. Numerous studies have focused on editing neural radiance fields, including relighting [21, 41], composition [32, 55], content generation [29], and shape editing [22, 62]. Editing tools operating at the scene- or object-level, such as [16, 55, 61], utilize latent codes to stylize the appearance of an object or entire scenes. Conversely, pixel-level editing tools, such as NeuMesh [54] and Seal3D [47], utilize training-based approaches to edit fine-grained details. However, NeuMesh relies on explicit meshes which lack efficiency and robustness. Although Seal3D achieves interactive editing, it still suffers from poor geometry reconstruction and cannot change the illumination scene. Parameterization-based methods, such as NeuTex [52], ISO-UVField [38] and NeP [25], integrate the learning of UV mapping into neural rendering, enabling intuitive editing within 2D domains. However, these methods often suffer from large distortions in the parametrization. Furthermore, they require 3D prior information from tracked meshes and UV mapping - conditions that may not be readily met in practical, real-world scenarios. IntrinsicNeRF [60] enhances standard neural radiance fields by generating additional outputs, including reflectance, shading, and a residual term. Incorporating a semantic branch, it facilitates real-time scene editing. However, it often results in aggregated colors, akin to palette-based methods [18, 44], hence cannot offer fine-grained editing results at the pixel level. We decompose the radiance field into view-dependent shadings and view-independent materials. Both components are represented as textures, enabling intuitive material modifications while preserving consistent shading." + }, + { + "type": "title", + "bbox": [ + 0.216, + 0.626, + 0.33, + 0.642 + ], + "angle": 0, + "content": "3 Method" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.659, + 0.789, + 0.842 + ], + "angle": 0, + "content": "The input of our method is a collection of RGB images, denoted \\(\\mathcal{I} = \\{\\mathbf{I}_i\\}\\), representing multiple objects of similar geometry or a single object, captured from different viewpoints. To differentiate among objects, we assign a latent shape code \\(\\mathbf{z}_s\\) and an appearance code \\(\\mathbf{z}_a\\) to each object. This facilitates the co-parameterization of multiple objects and enables the transfer of material and shading among them. As illustrated in Fig. 1, our method reconstructs 3D geometry as a neural implicit surface and encodes the 3D radiance field in a simple parametric domain, such as spheres and polycubes, which we learn from an initial phase in our training (Sec. 3.1). For each input image, we randomly select a set of pixels. Then, for each pixel, we shoot a ray originating from the location of camera and traversing through the pixel. If the ray intersects with the surface, at a point, denoted as \\(\\mathbf{p}\\), we compute its corresponding position in the" + } + ], + [ + { + "type": "header", + "bbox": [ + 0.66, + 0.115, + 0.732, + 0.127 + ], + "angle": 0, + "content": "NeuParam" + }, + { + "type": "page_number", + "bbox": [ + 0.775, + 0.117, + 0.785, + 0.127 + ], + "angle": 0, + "content": "5" + }, + { + "type": "image", + "bbox": [ + 0.218, + 0.144, + 0.782, + 0.263 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.214, + 0.273, + 0.788, + 0.442 + ], + "angle": 0, + "content": "Fig. 1: Algorithmic pipeline. Initially, we learn the parametric domain \\( F_{\\mathrm{sdf}} \\) using a coarse SDF \\( G_{\\mathrm{sdf}} \\) of the target surface \\( \\mathcal{S} \\) during the early phase of neural rendering (in the left grey box). Once the parametric domain \\( \\mathcal{D} \\) is determined, our pipeline subsequently utilizes the bi-directional deformation \\( (F_{\\mathrm{def}} \\) and \\( F_{\\mathrm{inv-def}}) \\), the parametric domain SDF \\( (F_{\\mathrm{sdf}}) \\) and the radiance decomposition \\( (F_{\\mathrm{mat}} \\) and \\( F_{\\mathrm{shd}}) \\). The output includes the reconstructed 3D surface \\( \\mathcal{S} \\), the decomposed radiance fields (featuring both view-dependent and view-independent components) and the map between \\( \\mathcal{S} \\) and \\( \\mathcal{D} \\). It is worth mentioning that while our framework involves two SDFs - one for the 3D surface \\( \\mathcal{S} \\) and the other for the parametric domain \\( \\mathcal{D} \\) - our network design only requires one of them. This is because one SDF can be derived from the other via either the forward or the backward deformation. To reduce the network complexity, we only adopt one geometry sub-network \\( F_{\\mathrm{sdf}} \\), which is used to represent the parametric domain \\( \\mathcal{D} \\)." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.473, + 0.789, + 0.641 + ], + "angle": 0, + "content": "parametric domain \\(\\mathcal{D}\\), denoted as \\(\\mathbf{p}'\\), via a bi-directional deformation (Sec. 3.2): the forward deformation (depicted in orange) generates a displacement vector so that the updated position \\(\\mathbf{p}'\\) is in the parametric domain; subsequently, the inverse deformation (depicted in green) maps \\(\\mathbf{p}'\\) back to a point \\(\\mathbf{p}''\\), which is anticipated to be in close proximity to the surface. We incorporate a cycle loss to penalize occurrences where \\(\\mathbf{p}''\\) deviates from \\(\\mathbf{p}\\). In addition, we employ a Laplacian loss (Sec. 3.3) to effectively reduce the angle distortion between the parametric domain and the original surface. Furthermore, we decompose the radiance into a view-independent material field and a view-dependent shading field (Sec. 3.4), each of which can be edited independently, thus augmenting the editing capability of our framework. See Fig. 1 for the algorithmic pipeline." + }, + { + "type": "title", + "bbox": [ + 0.216, + 0.67, + 0.504, + 0.687 + ], + "angle": 0, + "content": "3.1 Parametric Domain Learning" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.704, + 0.788, + 0.81 + ], + "angle": 0, + "content": "In our framework, we consider two types of parametric domains: spheres and polycubes, both represented as signed distance fields. Spheres are characterized by four learnable parameters: the center position and radius. Polycubes, a generalization of cubes, are simple structures yet offer sufficient degrees of freedom to capture the salient geometric features of target surfaces. We define a learnable polycube with \\( k \\) elements as follows: For the \\( i \\)-th element, which is an axis-aligned box, its SDF \\( \\phi_i \\) is given by" + }, + { + "type": "equation", + "bbox": [ + 0.321, + 0.825, + 0.787, + 0.842 + ], + "angle": 0, + "content": "\\[\n\\phi_ {i} = \\max \\left(\\left| x - x _ {i} \\right| - a _ {i}, \\left| y - y _ {i} \\right| - b _ {i}, \\left| z - z _ {i} \\right| - c _ {i}\\right), \\tag {1}\n\\]" + } + ], + [ + { + "type": "page_number", + "bbox": [ + 0.218, + 0.116, + 0.23, + 0.127 + ], + "angle": 0, + "content": "6" + }, + { + "type": "header", + "bbox": [ + 0.272, + 0.115, + 0.353, + 0.128 + ], + "angle": 0, + "content": "B. Xu et al." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.147, + 0.784, + 0.178 + ], + "angle": 0, + "content": "where \\((x_{i},y_{i},z_{i})\\) represents the box center, and \\((a_i,b_i,c_i)\\) is its dimensions. By combining the \\(k\\) boxes, the SDF of the polycube is obtained as" + }, + { + "type": "equation", + "bbox": [ + 0.39, + 0.193, + 0.786, + 0.211 + ], + "angle": 0, + "content": "\\[\nf _ {\\mathrm {p c}} = \\max (- \\phi_ {1}, - \\phi_ {2}, \\dots , - \\phi_ {k}). \\tag {2}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.224, + 0.784, + 0.256 + ], + "angle": 0, + "content": "To ensure the differentiability of Eq. (2), we approximate it using Kreisselmeier-Steinhauser (KS) function [17] as" + }, + { + "type": "equation", + "bbox": [ + 0.362, + 0.269, + 0.64, + 0.311 + ], + "angle": 0, + "content": "\\[\nf _ {\\mathrm {p c}} \\approx - 1 / \\lambda * \\log \\left(\\sum_ {i} \\exp (- \\lambda * \\phi_ {i})\\right),\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.215, + 0.324, + 0.678, + 0.339 + ], + "angle": 0, + "content": "where \\(\\lambda\\) is a typically large value, set to 100 in our experiments." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.341, + 0.785, + 0.461 + ], + "angle": 0, + "content": "Since determining the configuration of boxes does not require a highly accurate geometry of the target object \\(S\\), we design the parametric domain learning as a one-time, intermediate step in the neural rendering pipeline. Specifically, after an initial training phase of approximately 100 epochs, the coarse geometry of the surface \\(S\\) is obtained. Subsequently, we activate the domain learning module, which takes the SDF \\(G_{\\mathrm{sdf}}\\) of the coarse surface \\(S\\) and the user-specified number of boxes \\(k\\) as input. It computes the SDF \\(F_{\\mathrm{sdf}}\\) of the parametric domain \\(\\mathcal{D}\\) via minimizing the loss function:" + }, + { + "type": "equation", + "bbox": [ + 0.406, + 0.477, + 0.786, + 0.495 + ], + "angle": 0, + "content": "\\[\n\\mathcal {L} _ {\\text {p a r a m}} = \\mathcal {L} _ {\\text {s d f}} + \\lambda_ {s} \\mathcal {L} _ {\\text {l a p s d f}}, \\tag {3}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.215, + 0.509, + 0.264, + 0.522 + ], + "angle": 0, + "content": "where" + }, + { + "type": "equation", + "bbox": [ + 0.268, + 0.534, + 0.785, + 0.569 + ], + "angle": 0, + "content": "\\[\n\\mathcal {L} _ {\\mathrm {s d f}} = \\left| F _ {\\mathrm {s d f}} (p) - G _ {\\mathrm {s d f}} (p) \\right|, \\mathcal {L} _ {\\mathrm {l a p s d f}} = \\left| \\frac {\\partial^ {2} F _ {\\mathrm {s d f}}}{\\partial x ^ {2}} + \\frac {\\partial^ {2} F _ {\\mathrm {s d f}}}{\\partial y ^ {2}} + \\frac {\\partial^ {2} F _ {\\mathrm {s d f}}}{\\partial z ^ {2}} \\right|, \\tag {4}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.582, + 0.784, + 0.627 + ], + "angle": 0, + "content": "and the weight \\(\\lambda_{s}\\) is set to 0.01. The term \\(\\mathcal{L}_{\\mathrm{lapsdf}}\\), which is the Laplacian of the signed distance field of the polycube \\(\\mathcal{D}\\), plays a critical role in controlling its geometric complexity." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.63, + 0.603, + 0.841 + ], + "angle": 0, + "content": "After obtaining the SDF \\( F_{\\mathrm{sdf}} \\) of the polycube \\( \\mathcal{D} \\), we deprecate the SDF network \\( G_{\\mathrm{sdf}} \\) and instead use the deformation network \\( F_{\\mathrm{def}} \\) and parametric domain SDF \\( F_{\\mathrm{sdf}} \\) network to represent the source surface \\( \\mathcal{S} \\). The right inset showcases a polycube learning example for a human head model. Given a user-specified parameter \\( k = 3 \\), our method initializes three randomly placed boxes. It proceeds to minimize the loss function \\( \\mathcal{L}_{\\mathrm{param}} \\), iteratively refining the size and positioning of each box to ensure the polycube closely approximates the surface \\( \\mathcal{S} \\). This parametric domain learning phase typically converges within 30 to 60 iterations, taking under 10 minutes to complete. For more details, see the supplementary video." + }, + { + "type": "image", + "bbox": [ + 0.619, + 0.661, + 0.667, + 0.709 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.636, + 0.714, + 0.65, + 0.726 + ], + "angle": 0, + "content": "S" + }, + { + "type": "image", + "bbox": [ + 0.674, + 0.661, + 0.723, + 0.71 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.68, + 0.714, + 0.72, + 0.726 + ], + "angle": 0, + "content": "Iter 0" + }, + { + "type": "image", + "bbox": [ + 0.736, + 0.663, + 0.777, + 0.709 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.737, + 0.714, + 0.776, + 0.726 + ], + "angle": 0, + "content": "Iter 5" + }, + { + "type": "image", + "bbox": [ + 0.622, + 0.73, + 0.667, + 0.776 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.676, + 0.729, + 0.723, + 0.775 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.621, + 0.781, + 0.667, + 0.793 + ], + "angle": 0, + "content": "Iter 10" + }, + { + "type": "image_caption", + "bbox": [ + 0.677, + 0.781, + 0.723, + 0.793 + ], + "angle": 0, + "content": "Iter 45" + }, + { + "type": "image", + "bbox": [ + 0.736, + 0.729, + 0.78, + 0.776 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.734, + 0.781, + 0.78, + 0.793 + ], + "angle": 0, + "content": "Iter 55" + } + ], + [ + { + "type": "header", + "bbox": [ + 0.659, + 0.115, + 0.732, + 0.128 + ], + "angle": 0, + "content": "NeuParam" + }, + { + "type": "page_number", + "bbox": [ + 0.776, + 0.117, + 0.786, + 0.127 + ], + "angle": 0, + "content": "7" + }, + { + "type": "title", + "bbox": [ + 0.216, + 0.147, + 0.484, + 0.16 + ], + "angle": 0, + "content": "3.2 Bi-directional Deformation" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.168, + 0.784, + 0.243 + ], + "angle": 0, + "content": "The bi-directional deformation module serves as the pivotal component within our neural parameterization framework. As illustrated in Fig. 1, this module consists of two sub-networks: \\( F_{\\mathrm{def}} \\), which is responsible for the forward mapping from the original model to the parameter domain, and \\( F_{\\mathrm{inv-def}} \\), which establishes the inverse mapping." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.244, + 0.784, + 0.317 + ], + "angle": 0, + "content": "Let \\( S \\) be the 3D surface and \\( \\mathcal{D} \\) be the parametric domain; both are represented by the zero level-set of neural signed distance fields. For an arbitrary point \\( \\mathbf{p} \\in S \\), the forward deformation \\( F_{\\mathrm{def}}(\\mathbf{p}, \\mathbf{z}_s) \\) computes a displacement vector, mapping the point \\( \\mathbf{p} \\) to a corresponding point \\( \\mathbf{p}' \\in \\mathcal{D} \\) in the parametric domain as follows:" + }, + { + "type": "equation", + "bbox": [ + 0.428, + 0.32, + 0.785, + 0.335 + ], + "angle": 0, + "content": "\\[\n\\mathbf {p} ^ {\\prime} = \\mathbf {p} + F _ {\\mathrm {d e f}} (\\mathbf {p}, \\mathbf {z} _ {s}). \\tag {5}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.215, + 0.339, + 0.784, + 0.369 + ], + "angle": 0, + "content": "Conversely, we employ \\( F_{\\mathrm{inv - def}} \\) to map points inversely from the parametric domain \\( \\mathcal{D} \\) back to the original shape \\( S \\):" + }, + { + "type": "equation", + "bbox": [ + 0.411, + 0.377, + 0.785, + 0.393 + ], + "angle": 0, + "content": "\\[\n\\mathbf {p} ^ {\\prime \\prime} = \\mathbf {p} ^ {\\prime} + F _ {\\text {i n v - d e f}} \\left(\\mathbf {p} ^ {\\prime}, \\mathbf {z} _ {s}\\right), \\tag {6}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.215, + 0.399, + 0.784, + 0.444 + ], + "angle": 0, + "content": "where the point \\(\\mathbf{p}''\\) is expected to be in close proximity to \\(p\\). To encourage the smoothness and stability of the deformations, we impose an \\(L_{2}\\) loss on the magnitude of the displacement vectors:" + }, + { + "type": "equation", + "bbox": [ + 0.341, + 0.453, + 0.785, + 0.468 + ], + "angle": 0, + "content": "\\[\n\\mathcal {L} _ {\\text {s m o o t h}} = \\left\\| F _ {\\text {d e f}} \\left(\\mathbf {p}, \\mathbf {z} _ {s}\\right) \\right\\| _ {2} + \\left\\| F _ {\\text {i n v - d e f}} \\left(\\mathbf {p} ^ {\\prime}, \\mathbf {z} _ {s}\\right) \\right\\| _ {2}. \\tag {7}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.215, + 0.475, + 0.784, + 0.504 + ], + "angle": 0, + "content": "This loss helps minimize abrupt and irregular deformations, leading to more coherent and gradual changes between \\(\\mathcal{S}\\) and \\(\\mathcal{D}\\)." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.505, + 0.784, + 0.551 + ], + "angle": 0, + "content": "In training the networks \\( F_{\\mathrm{def}} \\) and \\( F_{\\mathrm{inv - def}} \\), our objective is to achieve a bijective mapping between the 3D surface \\( S \\) and the parametric domain \\( \\mathcal{D} \\). To achieve this goal, we adopt a cycle loss, which is similar to NeuTex [52]:" + }, + { + "type": "equation", + "bbox": [ + 0.389, + 0.558, + 0.785, + 0.582 + ], + "angle": 0, + "content": "\\[\n\\mathcal {L} _ {\\text {c y c l e}} = \\sum_ {\\mathbf {p} \\in S} \\lambda (\\mathbf {p}) \\| \\mathbf {p} - \\mathbf {p} ^ {\\prime \\prime} \\| _ {2}, \\tag {8}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.587, + 0.785, + 0.722 + ], + "angle": 0, + "content": "where the weight \\(\\lambda\\) characterizes the importance of sample \\(\\mathbf{p}\\) to the loss. The reason that we do not consider all sample points equally is that a sample point may not be exactly on the zero level-set of the SDF. To tolerate such an inaccuracy, we define \\(\\lambda (\\mathbf{p}) = T(\\mathbf{p})(1 - \\exp (-\\sigma (s)\\delta (\\mathbf{p})))\\) as the color weight used in volume rendering [27]. Here, \\(T(\\mathbf{p})\\) is the transparency of sample \\(\\mathbf{p}\\) in the viewing direction, \\(\\delta\\) is the length of the sample interval along the ray, and \\(s\\) is the signed distance value of the point \\(\\mathbf{p}\\). Clearly, when \\(\\mathbf{p}\\) is away from the surface, the coefficient \\(\\lambda (\\mathbf{p})\\) has few effects on this loss. We refer the readers to [27] for details about volume rendering and the computation of transparency \\(T\\)." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.723, + 0.784, + 0.8 + ], + "angle": 0, + "content": "By explicitly minimizing the distance between \\(\\mathbf{p}\\) and \\(\\mathbf{p}''\\), the cycle loss effectively prevents scenarios where two distinct points \\(\\mathbf{p}_1, \\mathbf{p}_2 \\in S\\) map to an identical point \\(\\mathbf{p}' \\in \\mathcal{D}\\). In such cases, the inverse deformation \\(F_{\\mathrm{inv-def}}\\) would be unable to map the single point \\(\\mathbf{p}'\\) back to the respective distinct points \\(\\mathbf{p}_1''\\) and \\(\\mathbf{p}_2''\\)." + }, + { + "type": "title", + "bbox": [ + 0.215, + 0.817, + 0.467, + 0.832 + ], + "angle": 0, + "content": "3.3 Laplacian Regularization" + } + ], + [ + { + "type": "page_number", + "bbox": [ + 0.218, + 0.116, + 0.23, + 0.127 + ], + "angle": 0, + "content": "8" + }, + { + "type": "header", + "bbox": [ + 0.272, + 0.115, + 0.353, + 0.128 + ], + "angle": 0, + "content": "B. Xu et al." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.147, + 0.603, + 0.282 + ], + "angle": 0, + "content": "Mitigating angle distortion is crucial in ensuring the quality of parameterization. One approach to reduce angle distortion involves minimizing the ratio of the two singular values of the Jacobian matrix associated with the parameterization [6]. However, considering the representation of the surface \\(S\\), the parametric domain \\(\\mathcal{D}\\), and the parameterization \\(F_{\\mathrm{def}}\\) as neural implicit functions in our approach, directly computing the Jacobian matrix is a non-trivial task. To circum" + }, + { + "type": "image", + "bbox": [ + 0.617, + 0.149, + 0.784, + 0.253 + ], + "angle": 0, + "content": null + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.283, + 0.788, + 0.405 + ], + "angle": 0, + "content": "vent this, we employ a Laplacian regularizer in our implementation to control angle distortion. Achieving the minimizer of the Laplacian of the map results in a harmonic map. For genus-0 surfaces, harmonic maps are also conformal, which preserves angles [12]. Although for surfaces of higher genus, a harmonic map is not necessarily conformal, it still serves an effective tool for reducing angle distortion [13]. Specifically, for a sample point \\(\\mathbf{p} \\in S\\) and its neighboring points \\(\\mathbf{q}_j \\in S\\), the forward deformation maps them to \\(\\mathbf{p}' \\in \\mathcal{D}\\) and \\(\\mathbf{q}_j' \\in \\mathcal{D}\\), respectively. We compute the Laplacian of the sample \\(\\mathbf{p}\\) as follows:" + }, + { + "type": "equation", + "bbox": [ + 0.402, + 0.413, + 0.786, + 0.439 + ], + "angle": 0, + "content": "\\[\n\\Delta (\\mathbf {p}) = \\sum_ {j = 1} ^ {m} \\omega_ {j} \\left(\\mathbf {p} ^ {\\prime} - \\mathbf {q} _ {j} ^ {\\prime}\\right). \\tag {9}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.216, + 0.448, + 0.786, + 0.464 + ], + "angle": 0, + "content": "Here, the weights \\( w_{j} \\) are defined using the metric from the original surface \\( S \\) as:" + }, + { + "type": "equation", + "bbox": [ + 0.406, + 0.478, + 0.786, + 0.496 + ], + "angle": 0, + "content": "\\[\n\\omega_ {j} = \\exp (- \\| \\mathbf {p} - \\mathbf {q} _ {j} \\| _ {2} / l), \\tag {10}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.501, + 0.788, + 0.562 + ], + "angle": 0, + "content": "where \\(l\\) is the average distance from point \\(\\mathbf{p}\\) to its neighboring points. Subsequently, the weights are normalized \\(\\omega_{j} = \\omega_{j} / \\sum_{k = 1}^{m}\\omega_{k}\\). Notice that \\(\\varDelta(\\mathbf{p})\\in\\mathbb{R}^3\\) is a 3D vector, which may not necessarily lie on the tangent plane. Therefore, we define the normal component of the Laplacian as:" + }, + { + "type": "equation", + "bbox": [ + 0.423, + 0.573, + 0.786, + 0.59 + ], + "angle": 0, + "content": "\\[\n\\Delta (\\mathbf {p}) _ {\\perp} = \\langle \\Delta (\\mathbf {p}), \\mathbf {n} \\rangle \\mathbf {n}, \\tag {11}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.599, + 0.787, + 0.63 + ], + "angle": 0, + "content": "where \\(\\mathbf{n}\\) is the normal at \\(\\mathbf{p}'\\), and \\(\\langle, \\rangle\\) represents the inner product. Summing the tangential components yields the Laplacian loss:" + }, + { + "type": "equation", + "bbox": [ + 0.378, + 0.64, + 0.786, + 0.665 + ], + "angle": 0, + "content": "\\[\n\\mathcal {L} _ {\\mathrm {L a p}} = \\sum_ {\\mathbf {p} \\in \\mathcal {S}} \\| \\Delta (\\mathbf {p}) - \\Delta (\\mathbf {p}) _ {\\perp} \\| _ {2}. \\tag {12}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.215, + 0.673, + 0.759, + 0.689 + ], + "angle": 0, + "content": "In our implementation, we set \\( m = 6 \\) neighboring points for each sample \\( \\mathbf{p} \\)." + }, + { + "type": "title", + "bbox": [ + 0.215, + 0.711, + 0.49, + 0.726 + ], + "angle": 0, + "content": "3.4 Appearance Decomposition" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.734, + 0.788, + 0.842 + ], + "angle": 0, + "content": "Radiance represents the directional emission of color, illustrating how the appearance of an object varies in response to changes in the observational viewpoint [27]. Drawing inspiration from previous works [25, 60] and applying the assumptions of Lambertian and grayscale shading [4], we decompose the radiance field on the parametric domain into two components: a view-independent material field and a view-dependent shading field (see Fig. 6(a)). This decomposition allows for independent editing of material and shading. More precisely, for" + } + ], + [ + { + "type": "header", + "bbox": [ + 0.659, + 0.115, + 0.732, + 0.128 + ], + "angle": 0, + "content": "NeuParam" + }, + { + "type": "page_number", + "bbox": [ + 0.776, + 0.117, + 0.786, + 0.127 + ], + "angle": 0, + "content": "9" + }, + { + "type": "text", + "bbox": [ + 0.215, + 0.147, + 0.784, + 0.178 + ], + "angle": 0, + "content": "In a sample point \\(\\mathbf{p} \\in \\mathcal{S}\\), we employ two sub-networks, \\(F_{\\mathrm{mat}}\\) and \\(F_{\\mathrm{shd}}\\), to decompose the radiance \\(\\mathbf{r} \\in \\mathbb{R}^3\\) at point \\(\\mathbf{p}\\) as follows:" + }, + { + "type": "equation", + "bbox": [ + 0.338, + 0.192, + 0.785, + 0.208 + ], + "angle": 0, + "content": "\\[\n\\mathbf {r} (\\mathbf {p}) = F _ {\\mathrm {m a t}} \\left(\\mathbf {p} ^ {\\prime}, \\mathbf {n}, \\mathbf {z} _ {a}\\right) \\cdot \\exp \\left(F _ {\\mathrm {s h d}} \\left(\\mathbf {p} ^ {\\prime}, \\mathbf {n}, \\mathbf {v}, \\mathbf {z} _ {a}\\right)\\right), \\tag {13}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.215, + 0.22, + 0.784, + 0.327 + ], + "angle": 0, + "content": "where \\(\\mathbf{p}' \\in \\mathcal{D}\\) is the image of \\(\\mathbf{p}\\) and \\(\\mathbf{n}\\) is the unit normal of \\(\\mathbf{p}\\) on the original surface. The material network \\(F_{\\mathrm{mat}}\\) takes the mapped point \\(\\mathbf{p}'\\), the surface normal \\(\\mathbf{n}\\), and the appearance latent code \\(\\mathbf{z}_a\\) as input, generating the view-independent material field in response. The shading network \\(F_{\\mathrm{shd}}\\), which is designed for computing the view-dependent shading field, takes an additional input: the viewpoint \\(\\mathbf{v}\\). To ensure the shading sub-network \\(F_{\\mathrm{shd}}\\) concentrates on local shading features, we employ an \\(L_1\\) loss to the output of \\(F_{\\mathrm{shd}}\\), encouraging sparsity," + }, + { + "type": "equation", + "bbox": [ + 0.403, + 0.341, + 0.785, + 0.356 + ], + "angle": 0, + "content": "\\[\n\\mathcal {L} _ {\\mathrm {s h d}} = \\left\\| F _ {\\mathrm {s h d}} \\left(\\mathbf {p} ^ {\\prime}, \\mathbf {n}, \\mathbf {v}, \\mathbf {z} _ {a}\\right) \\right\\| _ {1}. \\tag {14}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.215, + 0.37, + 0.784, + 0.43 + ], + "angle": 0, + "content": "To compute the color, we need the signed distance values for the samples \\(\\mathbf{p}\\). As mentioned above, our network adopts only one SDF network \\(F_{\\mathrm{sdf}}\\), which represents the geometry of the parametric domain \\(\\mathcal{D}\\). We use the forward deformation \\(F_{\\mathrm{def}}\\) to compute \\(s(\\mathbf{p})\\) as" + }, + { + "type": "equation", + "bbox": [ + 0.399, + 0.445, + 0.785, + 0.46 + ], + "angle": 0, + "content": "\\[\ns (\\mathbf {p}) = F _ {\\mathrm {s d f}} \\left(\\mathbf {p} + F _ {\\mathrm {d e f}} (\\mathbf {p}, \\mathbf {z} _ {s})\\right). \\tag {15}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.215, + 0.473, + 0.784, + 0.519 + ], + "angle": 0, + "content": "Then we employ the SDF-based volume rendering [58] to integrate the radiances along the ray to compute color \\(\\mathbf{c}_{\\mathrm{pred}} = \\sum_{i}\\lambda_{i}\\mathbf{r}_{i}\\), where \\(\\mathbf{r}_i\\) is the radiance of a sample point \\(\\mathbf{p}_i\\) and \\(\\lambda_{i}\\) is the same color weight as used in Eq. (8)." + }, + { + "type": "title", + "bbox": [ + 0.216, + 0.545, + 0.39, + 0.56 + ], + "angle": 0, + "content": "3.5 Training Losses" + }, + { + "type": "text", + "bbox": [ + 0.215, + 0.573, + 0.784, + 0.634 + ], + "angle": 0, + "content": "In contrast to NeP [25], which necessitate prior information such as tracked mesh and UV mapping for supervision, our method is capable of reconstructing 3D geometry and computing the corresponding texture map directly from multiview images. This process is supervised by the image loss" + }, + { + "type": "equation", + "bbox": [ + 0.424, + 0.648, + 0.785, + 0.664 + ], + "angle": 0, + "content": "\\[\n\\mathcal {L} _ {\\mathrm {r g b}} = \\left\\| \\mathbf {c} _ {\\mathrm {g t}} - \\mathbf {c} _ {\\mathrm {p r e d}} \\right\\| _ {1}, \\tag {16}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.215, + 0.677, + 0.784, + 0.767 + ], + "angle": 0, + "content": "where \\(\\mathbf{c}_{\\mathrm{gt}}\\) and \\(\\mathbf{c}_{\\mathrm{pred}}\\) are the ground-truth and predicted colors, respectively. To ensure the shape and appearance latent codes, \\(\\mathbf{z}_s\\) and \\(\\mathbf{z}_a\\), conform to Gaussian distributions, we incorporate a regularization term, expressed as \\(\\mathcal{L}_{\\mathrm{code}} = \\| \\mathbf{z}_s\\|_2 + \\| \\mathbf{z}_a\\|_2\\) [31]. Additionally, with geometry represented as signed distance fields, we employ the Eikonal loss [10, 59], defined as \\(\\mathcal{L}_{\\mathrm{Eik}} = (\\| \\nabla s\\|_2 - 1)^2\\). Putting it all together, we define the total loss function as follows:" + }, + { + "type": "equation", + "bbox": [ + 0.229, + 0.782, + 0.785, + 0.797 + ], + "angle": 0, + "content": "\\[\n\\mathcal {L} = \\mathcal {L} _ {\\mathrm {r g b}} + \\lambda_ {1} \\mathcal {L} _ {\\text {E i k}} + \\lambda_ {2} \\mathcal {L} _ {\\text {c y c l e}} + \\lambda_ {3} \\mathcal {L} _ {\\text {s m o o t h}} + \\lambda_ {4} \\mathcal {L} _ {\\text {L a p}} + \\lambda_ {5} \\mathcal {L} _ {\\text {s h d}} + \\lambda_ {6} \\mathcal {L} _ {\\text {c o d e}}. \\tag {17}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.215, + 0.81, + 0.784, + 0.84 + ], + "angle": 0, + "content": "In our implementation, we empirically set the coefficients as \\(\\lambda_1 = \\lambda_2 = \\lambda_5 = \\lambda_6 = 0.01\\) and \\(\\lambda_3 = \\lambda_4 = 0.001\\)." + } + ], + [ + { + "type": "page_number", + "bbox": [ + 0.218, + 0.116, + 0.236, + 0.127 + ], + "angle": 0, + "content": "10" + }, + { + "type": "header", + "bbox": [ + 0.272, + 0.115, + 0.353, + 0.127 + ], + "angle": 0, + "content": "B. Xu et al." + }, + { + "type": "image", + "bbox": [ + 0.272, + 0.143, + 0.729, + 0.442 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.214, + 0.453, + 0.788, + 0.523 + ], + "angle": 0, + "content": "Fig. 2: Parameterization results. Human heads are co-parameterized to a sphere owing to the clear geometric resemblance. Given the diverse geometry of cars, each is parameterized to a polycube domain (bottom three rows). From top to bottom: the input image, the normal map of the reconstructed geometry \\( S \\), the parametric domain \\( \\mathcal{D} \\), and the reconstructed surface \\( S \\). See supplementary material for additional results." + }, + { + "type": "title", + "bbox": [ + 0.216, + 0.55, + 0.376, + 0.568 + ], + "angle": 0, + "content": "4 Experiments" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.581, + 0.788, + 0.657 + ], + "angle": 0, + "content": "Datasets. For our experiments, we used a customized dataset named FS-Syn, derived from the Facescape dataset [56], selecting 10 human head models and configuring 30 fixed viewpoints and lighting conditions to obtain synthesized multi-view images. We also evaluated our method on several man-made objects from the OmniObject dataset [51], and the H3DS dataset [36]." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.657, + 0.788, + 0.794 + ], + "angle": 0, + "content": "Training details and metrics. All the sub-networks \\( F_{\\mathrm{def}} \\), \\( F_{\\mathrm{inv-def}} \\), \\( F_{\\mathrm{sdf}} \\), \\( F_{\\mathrm{mat}} \\), and \\( F_{\\mathrm{shd}} \\) are MLPs. See the supplementary material for other detailed network hyperparameters. Once the parametric domain is chosen, we fixed \\( F_{\\mathrm{sdf}} \\) and trained other sub-networks for 2,000 epochs driven by the loss function as Eq. (17). We applied the standard Marching Cubes algorithm to extract triangle meshes for the zero level-set of the source surface \\( S \\) and then used the forward map to obtain the mesh representing the parametric surface. With triangle meshes, we measured the quality of the computed parameterization by the angle distortion \\( \\mathcal{E}_{\\mathrm{angle}} \\) and area distortion [3,15] \\( \\mathcal{E}_{\\mathrm{area}} \\) as follows:" + }, + { + "type": "equation", + "bbox": [ + 0.218, + 0.803, + 0.784, + 0.846 + ], + "angle": 0, + "content": "\\[\n\\mathcal {E} _ {\\mathrm {a n g l e}} = \\frac {\\cot \\alpha | a | ^ {2} + \\cot \\beta | b | ^ {2} + \\cot \\gamma | c | ^ {2}}{4 \\mathrm {a r e a} (\\varDelta_ {i})}, \\quad \\mathcal {E} _ {\\mathrm {a r e a}} = \\frac {1}{2} \\left(\\frac {\\mathrm {a r e a} (\\varDelta_ {i})}{\\mathrm {a r e a} (\\hat {\\varDelta} _ {i})} + \\frac {\\mathrm {a r e a} (\\hat {\\varDelta} _ {i})}{\\mathrm {a r e a} (\\varDelta_ {i})}\\right),\n\\]" + } + ], + [ + { + "type": "header", + "bbox": [ + 0.658, + 0.115, + 0.732, + 0.128 + ], + "angle": 0, + "content": "NeuParam" + }, + { + "type": "page_number", + "bbox": [ + 0.769, + 0.116, + 0.785, + 0.127 + ], + "angle": 0, + "content": "11" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.147, + 0.788, + 0.207 + ], + "angle": 0, + "content": "where \\(a, b, c\\) represent the side lengths of the triangle \\(\\Delta_i\\) of the original surface. \\(\\alpha, \\beta, \\gamma\\) are angles of the corresponding triangle \\(\\hat{\\Delta}_i\\) in parametric domain. In the case of conformal mapping, \\(\\mathcal{E}_{\\mathrm{angle}}\\) is equal to 1, and \\(\\mathcal{E}_{\\mathrm{area}}\\) is also equal to 1 for an isometric map." + }, + { + "type": "image", + "bbox": [ + 0.252, + 0.238, + 0.348, + 0.321 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.261, + 0.323, + 0.34, + 0.336 + ], + "angle": 0, + "content": "1.545/1.396" + }, + { + "type": "image", + "bbox": [ + 0.352, + 0.238, + 0.448, + 0.321 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.362, + 0.324, + 0.441, + 0.336 + ], + "angle": 0, + "content": "1.755/1.446" + }, + { + "type": "image", + "bbox": [ + 0.452, + 0.239, + 0.548, + 0.321 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.462, + 0.323, + 0.542, + 0.336 + ], + "angle": 0, + "content": "1.394/1.302" + }, + { + "type": "image", + "bbox": [ + 0.555, + 0.24, + 0.651, + 0.321 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.565, + 0.323, + 0.643, + 0.336 + ], + "angle": 0, + "content": "1.271/1.216" + }, + { + "type": "image", + "bbox": [ + 0.654, + 0.239, + 0.752, + 0.321 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.665, + 0.323, + 0.744, + 0.336 + ], + "angle": 0, + "content": "1.351/1.275" + }, + { + "type": "image_caption", + "bbox": [ + 0.214, + 0.348, + 0.788, + 0.433 + ], + "angle": 0, + "content": "Fig. 3: Results of cross-polycube parameterization on the H3DS dataset [36]. We learned a common polycube domain for all human heads with the parameter \\( k = 3 \\) (i.e., the polycube is formed by 3 boxes) and then computed the polycube parameterization for all the human heads simultaneously. Since human heads share similar geometries, the computed polycube parameterizations are also highly consistent. The values below each figure represent the angle distortion and the area distortion, respectively." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.464, + 0.788, + 0.571 + ], + "angle": 0, + "content": "Results. Our framework constructs a nearly bijective map between objects and multiple parametric domains such as spheres and polycubes, as shown in Fig. 2 and Fig. 3. It supports co-parameterization of multiple objects and facilitates material and shading transfer. Fully compatible with existing neural rendering pipelines, our neural parameterization algorithm uses multi-view images to simultaneously reconstruct 3D geometry and create texture maps. As shown in Fig. 2, we achieve high-fidelity results without prior information." + }, + { + "type": "image", + "bbox": [ + 0.218, + 0.598, + 0.323, + 0.68 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.237, + 0.682, + 0.3, + 0.696 + ], + "angle": 0, + "content": "UV prior" + }, + { + "type": "image", + "bbox": [ + 0.338, + 0.598, + 0.441, + 0.68 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.348, + 0.682, + 0.446, + 0.696 + ], + "angle": 0, + "content": "NeP w/o prior" + }, + { + "type": "image", + "bbox": [ + 0.456, + 0.598, + 0.56, + 0.68 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.496, + 0.682, + 0.554, + 0.696 + ], + "angle": 0, + "content": "w/ prior" + }, + { + "type": "image", + "bbox": [ + 0.578, + 0.6, + 0.656, + 0.68 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.667, + 0.611, + 0.786, + 0.68 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.666, + 0.682, + 0.7, + 0.695 + ], + "angle": 0, + "content": "Ours" + }, + { + "type": "image_caption", + "bbox": [ + 0.214, + 0.707, + 0.788, + 0.763 + ], + "angle": 0, + "content": "Fig. 4: Parametrization results of NeP [25] in comparison with our method. NeP heavily relies on the initial UV prior, leading to poor texture maps when such prior information is lacking. Our method learns texture maps from a set of multi-view images without any 3D tracked mesh or UV prior as input." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.795, + 0.788, + 0.841 + ], + "angle": 0, + "content": "Compared with NeP [25], which is the latest neural surface editing work using neural parameterization, our method eliminates the need for prior information from UV mapping. Unlike our algorithm, NeP learns a map between" + } + ], + [ + { + "type": "page_number", + "bbox": [ + 0.218, + 0.116, + 0.236, + 0.127 + ], + "angle": 0, + "content": "12" + }, + { + "type": "header", + "bbox": [ + 0.272, + 0.115, + 0.353, + 0.127 + ], + "angle": 0, + "content": "B. Xu et al." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.147, + 0.788, + 0.208 + ], + "angle": 0, + "content": "surfaces and the UV plane. However, it heavily relies on an existing UV mapping as initialization, and yields rather poor texture mapping results without the UV mapping as prior (see Fig.4). Moreover, the results from NeP still exhibit significant distortions even with UV prior." + }, + { + "type": "image", + "bbox": [ + 0.233, + 0.234, + 0.77, + 0.363 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.214, + 0.376, + 0.788, + 0.418 + ], + "angle": 0, + "content": "Fig. 5: Representing the materials as textures defined in a simple parametric domain enables easy editing. Also, our co-parameterization naturally supports texture transfer between different objects." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.448, + 0.788, + 0.674 + ], + "angle": 0, + "content": "Texture editing. Our parameterization results enable flexible and pixel-level editing as shown in Fig. 5. We re-paint the extracted 2D texture map using cubemaps [9] and retrieve the modified color using the established mapping when volume rendering novel images after each editing. Both NeuMesh [54] and NeP [25] can edit the object's texture at a pixel level. We compare the editing and rendering results with NeuMesh and NeP, as shown in Fig. 5. NeuMesh requires a mask of the editing region. It also re-trains the network after each edit, which is time-consuming and cannot guarantee the stability of the training results. Our method outperforms NeuMesh in terms of the visual quality. NeP, which is also neural parameterization-based method, requires a UV prior for initialization. It often yields poor parameterization results without such prior. Even with the UV prior, its parameterization results exhibit large distortion, making the edit in the 2D domain challenging. Our method is free of prior, and can yield parameterization with low distortion, thereby facilitating pixel-level editing and material transferring without network re-training." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.675, + 0.788, + 0.78 + ], + "angle": 0, + "content": "Shading editing. We decompose the 3D radiance field into view-independent material and view-dependent shading defined in the parametric domain. Taking advantage of this decomposition and the embedding into a simple domain, we can edit both the material and shading of the 3D surfaces in an intuitive manner. As illustrated in Fig. 6(b), we modify the shading of one model by transferring the shading from another model under different lighting conditions, while keeping the material unchanged." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.78, + 0.788, + 0.841 + ], + "angle": 0, + "content": "Inverse deformation. To validate the cycle loss and illustrate inverse deformation, we sample points on the polycube parametric surface \\(\\mathcal{D}\\). Subsequently, we inversely deform these points back to the surface \\(\\mathcal{S}\\). The point cloud is visualized in the right inset, and we quantify the CD \\((\\times 10^{-3})\\) with respect to" + } + ], + [ + { + "type": "header", + "bbox": [ + 0.66, + 0.115, + 0.732, + 0.127 + ], + "angle": 0, + "content": "NeuParam" + }, + { + "type": "page_number", + "bbox": [ + 0.769, + 0.116, + 0.786, + 0.127 + ], + "angle": 0, + "content": "13" + }, + { + "type": "image", + "bbox": [ + 0.248, + 0.142, + 0.761, + 0.221 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.223, + 0.221, + 0.782, + 0.247 + ], + "angle": 0, + "content": "(a) We decompose the radiance into view-dependent shadings and view-independent materials. The shading map is visualized from the front view." + }, + { + "type": "image", + "bbox": [ + 0.233, + 0.248, + 0.77, + 0.385 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.232, + 0.385, + 0.773, + 0.414 + ], + "angle": 0, + "content": "(b) We transfer the decomposed shading from one model to another. See also the accompanying video." + }, + { + "type": "image_caption", + "bbox": [ + 0.215, + 0.424, + 0.788, + 0.455 + ], + "angle": 0, + "content": "Fig. 6: We transfer the decomposed shading from one model to another. For further illustration, please refer to the accompanying video." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.482, + 0.789, + 0.51 + ], + "angle": 0, + "content": "the ground truth. The obtained low CD results confirm the efficacy of inverse deformation and cycle loss." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.513, + 0.789, + 0.618 + ], + "angle": 0, + "content": "Ablation studies. As shown in Fig. 7 (left), the term \\(\\mathcal{L}_{\\mathrm{smooth}}\\) helps to smooth the process of the two deformations, while \\(\\mathcal{L}_{\\mathrm{Lap}}\\) further reduces the angle distortion. Our method can effectively reduce angle distortion by minimizing Laplacian loss and reduce area distortion by choosing an appropriate parametric domain. As demonstrated in Fig. 7 (right), we calculate the angle distortion \\(\\mathcal{E}_{\\mathrm{angle}}\\) of parameterization as described in [3, 15]. We can observe that the Laplacian loss can effectively reduce angle distortion." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.619, + 0.789, + 0.71 + ], + "angle": 0, + "content": "The choice of parametric domain. The different parametric domain influences the quality of parameterization as shown in Fig. 8 (a), which is measured by area distortions. Generally, increasing the number of cubic primitives for better resembling the geometry of the target surface effectively reduces distortions but increases the complexity of the parametric domain. Finding an optimal balance between domain complexity and parameterization quality is challenging." + }, + { + "type": "title", + "bbox": [ + 0.216, + 0.748, + 0.36, + 0.765 + ], + "angle": 0, + "content": "5 Conclusion" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.78, + 0.789, + 0.842 + ], + "angle": 0, + "content": "We introduce a novel neural algorithm for parameterizing 3D surfaces represented by neural implicit functions to a user-defined parametric domain, such as spheres and polycubes. This parametric domain is learned given some user-specified hyperparameters. Utilizing bi-directional deformation, our method is" + } + ], + [ + { + "type": "page_number", + "bbox": [ + 0.218, + 0.116, + 0.236, + 0.127 + ], + "angle": 0, + "content": "14" + }, + { + "type": "header", + "bbox": [ + 0.272, + 0.115, + 0.353, + 0.127 + ], + "angle": 0, + "content": "B. Xu et al." + }, + { + "type": "image", + "bbox": [ + 0.233, + 0.144, + 0.301, + 0.203 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.312, + 0.147, + 0.377, + 0.202 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.389, + 0.147, + 0.453, + 0.202 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.465, + 0.149, + 0.531, + 0.202 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.256, + 0.21, + 0.274, + 0.218 + ], + "angle": 0, + "content": "GT" + }, + { + "type": "image_caption", + "bbox": [ + 0.313, + 0.21, + 0.374, + 0.218 + ], + "angle": 0, + "content": "VolSDF 4.65" + }, + { + "type": "image_caption", + "bbox": [ + 0.396, + 0.21, + 0.445, + 0.218 + ], + "angle": 0, + "content": "Ours 4.87" + }, + { + "type": "image_caption", + "bbox": [ + 0.466, + 0.21, + 0.526, + 0.218 + ], + "angle": 0, + "content": "Inverse 4.89" + }, + { + "type": "image", + "bbox": [ + 0.229, + 0.219, + 0.304, + 0.261 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.25, + 0.266, + 0.284, + 0.277 + ], + "angle": 0, + "content": "Input" + }, + { + "type": "image", + "bbox": [ + 0.312, + 0.227, + 0.382, + 0.261 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.296, + 0.266, + 0.458, + 0.278 + ], + "angle": 0, + "content": "w/o \\(\\mathcal{L}_{\\mathrm{smooth}}\\& \\mathcal{L}_{\\mathrm{Lap}} + \\mathcal{L}_{\\mathrm{smooth}}\\)" + }, + { + "type": "image", + "bbox": [ + 0.4, + 0.232, + 0.459, + 0.262 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.486, + 0.232, + 0.534, + 0.262 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.487, + 0.266, + 0.525, + 0.278 + ], + "angle": 0, + "content": "+Lap" + }, + { + "type": "image", + "bbox": [ + 0.546, + 0.15, + 0.756, + 0.247 + ], + "angle": 0, + "content": null + }, + { + "type": "image_footnote", + "bbox": [ + 0.55, + 0.251, + 0.633, + 0.263 + ], + "angle": 0, + "content": "\\(\\mathcal{E}_{\\mathrm{angle}} = 1.147\\)" + }, + { + "type": "image_footnote", + "bbox": [ + 0.661, + 0.251, + 0.747, + 0.263 + ], + "angle": 0, + "content": "\\(\\mathcal{E}_{\\mathrm{angle}} = 1.092\\)" + }, + { + "type": "image_footnote", + "bbox": [ + 0.561, + 0.264, + 0.623, + 0.278 + ], + "angle": 0, + "content": "w/o \\(\\mathcal{L}_{\\mathrm{Lap}}\\)" + }, + { + "type": "image_footnote", + "bbox": [ + 0.676, + 0.264, + 0.732, + 0.278 + ], + "angle": 0, + "content": "W/Lap" + }, + { + "type": "image_caption", + "bbox": [ + 0.215, + 0.288, + 0.787, + 0.346 + ], + "angle": 0, + "content": "Fig. 7: Ablation studies. On the left, we calculate the CD \\((\\times 10^{-3})\\) based on the point cloud extracted from the inverse deformation module and present the ablation studies for \\(\\mathcal{L}_{\\mathrm{smooth}}\\) and \\(\\mathcal{L}_{\\mathrm{Lap}}\\). On the right, utilization of the Laplacian loss has proven effective in reducing angle distortion \\((\\mathcal{E}_{\\mathrm{angle}})\\), especially in regions with high curvature." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.375, + 0.785, + 0.48 + ], + "angle": 0, + "content": "capable of learning a nearly bijective mapping without relying on any prior knowledge, while controlling angle distortion through the usage of a Laplacian regularizer. Furthermore, our approach seamlessly integrates with the neural rendering pipeline, enabling the reconstruction of 3D objects from multi-view images and facilitating efficient volume rendering of modified textures and shadings — eliminating the necessity for network re-training. We demonstrated the efficacy of our method on human heads and man-made objects." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.481, + 0.788, + 0.678 + ], + "angle": 0, + "content": "Our method has a few limitations that require further improvement in the future. First, quantitative comparisons on the FS-Syn dataset show that the reconstruction quality of our method is slightly worse than the standard SDF-based neural rendering algorithms, such as VolSDF [58] and NeuS [46]. For example, as illustrated in Fig. 8 (b) and, We report PSNR and CD compared with VolSDF [58] on Fs-Syn dataset [56]. This decline in visual quality arises from the incorporation of additional modules, such as bi-directional deformation and radiance decomposition, resulting in the cessation of updates to the SDF network within our neural parameterization pipeline. Therefore, computing a parameterization without compromising the reconstruction quality is highly desired. Second, our radiance decomposition operates under the assumptions of Lambertian reflectance and grayscale shading, which might hinder its effectiveness when dealing with intricate materials or sophisticated shadings." + }, + { + "type": "image", + "bbox": [ + 0.228, + 0.705, + 0.288, + 0.742 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.252, + 0.746, + 0.264, + 0.756 + ], + "angle": 0, + "content": "S" + }, + { + "type": "image", + "bbox": [ + 0.299, + 0.706, + 0.344, + 0.742 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.282, + 0.747, + 0.401, + 0.757 + ], + "angle": 0, + "content": "\\(\\mathcal{E}_{\\mathrm{area}} = 2.505\\) 6.094" + }, + { + "type": "image", + "bbox": [ + 0.355, + 0.706, + 0.415, + 0.742 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.432, + 0.704, + 0.479, + 0.742 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.432, + 0.746, + 0.464, + 0.756 + ], + "angle": 0, + "content": "1.633" + }, + { + "type": "image", + "bbox": [ + 0.491, + 0.705, + 0.52, + 0.741 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.492, + 0.744, + 0.516, + 0.755 + ], + "angle": 0, + "content": "GT" + }, + { + "type": "image", + "bbox": [ + 0.525, + 0.707, + 0.553, + 0.741 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.527, + 0.744, + 0.58, + 0.755 + ], + "angle": 0, + "content": "VolSDF" + }, + { + "type": "image", + "bbox": [ + 0.594, + 0.707, + 0.622, + 0.742 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.602, + 0.745, + 0.637, + 0.755 + ], + "angle": 0, + "content": "Ours" + }, + { + "type": "image", + "bbox": [ + 0.628, + 0.709, + 0.657, + 0.741 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.606, + 0.756, + 0.627, + 0.769 + ], + "angle": 0, + "content": "(b)" + }, + { + "type": "text", + "bbox": [ + 0.665, + 0.716, + 0.77, + 0.726 + ], + "angle": 0, + "content": "Method PSNR CD(10-3)" + }, + { + "type": "text", + "bbox": [ + 0.666, + 0.729, + 0.761, + 0.739 + ], + "angle": 0, + "content": "VolSDF 31.15 3.270" + }, + { + "type": "text", + "bbox": [ + 0.67, + 0.742, + 0.762, + 0.753 + ], + "angle": 0, + "content": "Ours 30.87 3.423" + }, + { + "type": "image_caption", + "bbox": [ + 0.215, + 0.78, + 0.787, + 0.837 + ], + "angle": 0, + "content": "Fig. 8: We present area distortion \\(\\mathcal{E}_{\\mathrm{area}}\\) based on the different choices of the parametric domain in (a) and geometry reconstruction results of VolSDF [58] and our method in (b). Our method is capable of editing through parameterization, accompanied by an acceptable compromise in reconstruction quality." + } + ], + [ + { + "type": "header", + "bbox": [ + 0.659, + 0.115, + 0.732, + 0.127 + ], + "angle": 0, + "content": "NeuParam" + }, + { + "type": "page_number", + "bbox": [ + 0.769, + 0.116, + 0.786, + 0.127 + ], + "angle": 0, + "content": "15" + }, + { + "type": "title", + "bbox": [ + 0.218, + 0.145, + 0.383, + 0.163 + ], + "angle": 0, + "content": "Acknowledgment" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.178, + 0.788, + 0.254 + ], + "angle": 0, + "content": "This study is supported under the RIE2020 Industry Alignment Fund - Industry Collaboration Projects (IAF-ICP) Funding Initiative, as well as cash and in-kind contribution from the industry partner(s). This project is also partially supported by the Ministry of Education, Singapore, under its Academic Research Fund Grants (MOE-T2EP20220-0005 & RT19/22)." + }, + { + "type": "title", + "bbox": [ + 0.218, + 0.278, + 0.323, + 0.294 + ], + "angle": 0, + "content": "References" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.311, + 0.787, + 0.367 + ], + "angle": 0, + "content": "1. Bednarik, J., Parashar, S., Gundogdu, E., Salzmann, M., Fua, P.: Shape reconstruction by learning differentiable surface representations. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 4716-4725 (2020) 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.368, + 0.787, + 0.407 + ], + "angle": 0, + "content": "2. Biermann, H., Martin, I., Bernardini, F., Zorin, D.: Cut-and-paste editing of multiresolution surfaces. ACM transactions on graphics (TOG) 21(3), 312-321 (2002) 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.409, + 0.787, + 0.435 + ], + "angle": 0, + "content": "3. Degener, P., Meseth, J., Klein, R.: An adaptable surface parameterization method. IMR 3, 201-213 (2003) 10, 13" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.437, + 0.787, + 0.478 + ], + "angle": 0, + "content": "4. Fan, Q., Yang, J., Hua, G., Chen, B., Wipf, D.: Revisiting deep intrinsic image decompositions. In: Proceedings of the IEEE conference on computer vision and pattern recognition. pp. 8944-8952 (2018) 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.479, + 0.787, + 0.506 + ], + "angle": 0, + "content": "5. Fang, H., Hart, J.C.: Textureshop: texture synthesis as a photograph editing tool. ACM Transactions on Graphics (TOG) 23(3), 354-359 (2004) 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.507, + 0.787, + 0.534 + ], + "angle": 0, + "content": "6. Floater, M.S., Hormann, K.: Surface parameterization: a tutorial and survey. Advances in multiresolution for geometric modelling pp. 157-186 (2005) 3, 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.535, + 0.787, + 0.576 + ], + "angle": 0, + "content": "7. García, I., Xia, J., He, Y., Xin, S., Patow, G.: Interactive applications for sketch-based editable polycube map. IEEE Trans. Vis. Comput. Graph. 19(7), 1158-1171 (2013) 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.577, + 0.787, + 0.604 + ], + "angle": 0, + "content": "8. Gotsman, C., Gu, X., Sheffer, A.: Fundamentals of spherical parameterization for 3d meshes. ACM Trans. Graph. 22(3), 358-363 (2003) 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.605, + 0.787, + 0.632 + ], + "angle": 0, + "content": "9. Greene, N.: Environment mapping and other applications of world projections. IEEE computer graphics and Applications 6(11), 21-29 (1986) 12" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.633, + 0.787, + 0.673 + ], + "angle": 0, + "content": "0. Gropp, A., Yariv, L., Haim, N., Atzmon, M., Lipman, Y.: Implicit geometric regularization for learning shapes. In: International Conference on Machine Learning. pp. 3789-3799. PMLR (2020) 9" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.674, + 0.787, + 0.715 + ], + "angle": 0, + "content": "1. Groueix, T., Fisher, M., Kim, V.G., Russell, B.C., Aubry, M.: A papier-mâché approach to learning 3d surface generation. In: Proceedings of the IEEE conference on computer vision and pattern recognition. pp. 216-224 (2018) 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.716, + 0.787, + 0.757 + ], + "angle": 0, + "content": "2. Gu, X., Wang, Y., Chan, T.F., Thompson, P.M., Yau, S.: Genus zero surface conformal mapping and its application to brain surface mapping. IEEE Trans. Medical Imaging 23(8), 949-958 (2004) 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.758, + 0.787, + 0.799 + ], + "angle": 0, + "content": "3. Gu, X., Yau, S.: Global conformal parameterization. In: Kobbelt, L., Schröder, P., Hoppe, H. (eds.) First Eurographics Symposium on Geometry Processing. vol. 43, pp. 127-137 (2003) 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.8, + 0.787, + 0.84 + ], + "angle": 0, + "content": "4. Guo, H., Liu, S., Pan, H., Liu, Y., Tong, X., Guo, B.: Complexgen: Cad reconstruction by b-rep chain complex generation. ACM Transactions on Graphics (TOG) 41(4), 1-18 (2022) 3" + }, + { + "type": "list", + "bbox": [ + 0.226, + 0.311, + 0.787, + 0.84 + ], + "angle": 0, + "content": null + } + ], + [ + { + "type": "page_number", + "bbox": [ + 0.218, + 0.116, + 0.236, + 0.127 + ], + "angle": 0, + "content": "16" + }, + { + "type": "header", + "bbox": [ + 0.272, + 0.115, + 0.352, + 0.127 + ], + "angle": 0, + "content": "B. Xu et al." + }, + { + "type": "ref_text", + "bbox": [ + 0.218, + 0.147, + 0.787, + 0.175 + ], + "angle": 0, + "content": "15. He, Y., Wang, H., Fu, C.W., Qin, H.: A divide-and-conquer approach for automatic polycube map construction. Computers & Graphics 33(3), 369-380 (2009) 10, 13" + }, + { + "type": "ref_text", + "bbox": [ + 0.218, + 0.175, + 0.787, + 0.215 + ], + "angle": 0, + "content": "16. Huang, Y.H., He, Y., Yuan, Y.J., Lai, Y.K., Gao, L.: Stylizednerf: Consistent 3d scene stylization as stylized nerf via 2d-3d mutual learning. In: Computer Vision and Pattern Recognition (CVPR) (2022) 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.218, + 0.215, + 0.787, + 0.257 + ], + "angle": 0, + "content": "17. Kreisselmeier, G., Steinhauser, R.: Systematic control design by optimizing a vector performance index. In: Computer aided design of control systems, pp. 113-117. Elsevier (1980) 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.257, + 0.787, + 0.311 + ], + "angle": 0, + "content": "18. Kuang, Z., Luan, F., Bi, S., Shu, Z., Wetzstein, G., Sunkavalli, K.: Palettenerf: Palette-based appearance editing of neural radiance fields. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 20691-20700 (2023) 2, 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.311, + 0.787, + 0.352 + ], + "angle": 0, + "content": "19. Li, M., Zhang, H.: D2im-net: Learning detail disentangled implicit fields from single images. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 10246-10255 (2021) 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.352, + 0.787, + 0.405 + ], + "angle": 0, + "content": "20. Li, Z., Müller, T., Evans, A., Taylor, R.H., Unberath, M., Liu, M.Y., Lin, C.H.: Neuralangelo: High-fidelity neural surface reconstruction. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 8456-8465 (2023) 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.405, + 0.787, + 0.459 + ], + "angle": 0, + "content": "21. Li, Z., Shi, J., Bi, S., Zhu, R., Sunkavalli, K., Hasan, M., Xu, Z., Ramamoorthi, R., Chandraker, M.: Physically-based editing of indoor scene lighting from a single image. In: European Conference on Computer Vision. pp. 555-572. Springer (2022) 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.459, + 0.787, + 0.5 + ], + "angle": 0, + "content": "22. Lin, C., Mitra, N., Wetzstein, G., Guibas, L.J., Guerrero, P.: Neuform: Adaptive overfitting for neural shape editing. Advances in Neural Information Processing Systems 35, 15217-15229 (2022) 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.501, + 0.787, + 0.542 + ], + "angle": 0, + "content": "23. Liu, S., Zhang, X., Zhang, Z., Zhang, R., Zhu, J.Y., Russell, B.: Editing conditional radiance fields. In: Proceedings of the IEEE/CVF international conference on computer vision. pp. 5773-5783 (2021) 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.542, + 0.787, + 0.582 + ], + "angle": 0, + "content": "24. Low, W.F., Lee, G.H.: Minimal neural atlas: Parameterizing complex surfaces with minimal charts and distortion. In: European Conference on Computer Vision. pp. 465-481. Springer (2022) 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.582, + 0.787, + 0.623 + ], + "angle": 0, + "content": "25. Ma, L., Li, X., Liao, J., Wang, X., Zhang, Q., Wang, J., Sander, P.V.: Neural parameterization for dynamic human head editing. ACM Transactions on Graphics (TOG) 41(6), 1-15 (2022) 2, 4, 8, 9, 11, 12" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.623, + 0.787, + 0.678 + ], + "angle": 0, + "content": "26. Mescheder, L., Oechsle, M., Niemeyer, M., Nowozin, S., Geiger, A.: Occupancy networks: Learning 3d reconstruction in function space. In: Proceedings of the IEEE/CVF conference on computer vision and pattern recognition. pp. 4460-4470 (2019) 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.678, + 0.787, + 0.718 + ], + "angle": 0, + "content": "27. Mildenhall, B., Srinivasan, P.P., Tancik, M., Barron, J.T., Ramamoorthi, R., Ng, R.: Nerf: Representing scenes as neural radiance fields for view synthesis. In: ECCV (2020) 1, 7, 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.718, + 0.787, + 0.759 + ], + "angle": 0, + "content": "28. Müller, T., Evans, A., Schied, C., Keller, A.: Instant neural graphics primitives with a multiresolution hash encoding. ACM Transactions on Graphics (ToG) 41(4), 1-15 (2022) 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.759, + 0.787, + 0.8 + ], + "angle": 0, + "content": "29. Niemeyer, M., Geiger, A.: Giraffe: Representing scenes as compositional generative neural feature fields. In: Proc. IEEE Conf. on Computer Vision and Pattern Recognition (CVPR) (2021) 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.8, + 0.787, + 0.841 + ], + "angle": 0, + "content": "30. Oechsle, M., Peng, S., Geiger, A.: Unisurf: Unifying neural implicit surfaces and radiance fields for multi-view reconstruction. In: Proceedings of the IEEE/CVF International Conference on Computer Vision. pp. 5589-5599 (2021) 1" + }, + { + "type": "list", + "bbox": [ + 0.217, + 0.147, + 0.787, + 0.841 + ], + "angle": 0, + "content": null + } + ], + [ + { + "type": "header", + "bbox": [ + 0.659, + 0.115, + 0.732, + 0.127 + ], + "angle": 0, + "content": "NeuParam" + }, + { + "type": "page_number", + "bbox": [ + 0.769, + 0.116, + 0.786, + 0.127 + ], + "angle": 0, + "content": "17" + }, + { + "type": "ref_text", + "bbox": [ + 0.214, + 0.147, + 0.788, + 0.204 + ], + "angle": 0, + "content": "31. Park, J.J., Florence, P., Straub, J., Newcombe, R., Lovegrove, S.: Deepsdf: Learning continuous signed distance functions for shape representation. In: Proceedings of the IEEE/CVF conference on computer vision and pattern recognition. pp. 165-174 (2019) 4, 9" + }, + { + "type": "ref_text", + "bbox": [ + 0.214, + 0.205, + 0.787, + 0.233 + ], + "angle": 0, + "content": "32. Pérez, P., Gangnet, M., Blake, A.: Poisson image editing. In: Seminal Graphics Papers: Pushing the Boundaries, Volume 2, pp. 577-582 (2023) 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.216, + 0.234, + 0.786, + 0.261 + ], + "angle": 0, + "content": "33. Praun, E., Hoppe, H.: Spherical parametrization and remeshing. ACM transactions on graphics (TOG) 22(3), 340-349 (2003) 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.262, + 0.787, + 0.303 + ], + "angle": 0, + "content": "34. Qi, C.R., Su, H., Mo, K., Guibas, L.J.: Pointnet: Deep learning on point sets for 3d classification and segmentation. In: Proceedings of the IEEE conference on computer vision and pattern recognition. pp. 652-660 (2017) 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.304, + 0.787, + 0.346 + ], + "angle": 0, + "content": "35. Qi, C.R., Su, H., Nießner, M., Dai, A., Yan, M., Guibas, L.J.: Volumetric and multi-view cnns for object classification on 3d data. In: Proceedings of the IEEE conference on computer vision and pattern recognition. pp. 5648-5656 (2016) 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.346, + 0.787, + 0.401 + ], + "angle": 0, + "content": "36. Ramon, E., Triginer, G., Escur, J., Pumarola, A., Garcia, J., Giro-i Nieto, X., Moreno-Noguer, F.: H3d-net: Few-shot high-fidelity 3d head reconstruction. In: Proceedings of the IEEE/CVF International Conference on Computer Vision. pp. 5620-5629 (2021) 10, 11" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.403, + 0.787, + 0.445 + ], + "angle": 0, + "content": "37. Rosu, R.A., Behnke, S.: Permutosdf: Fast multi-view reconstruction with implicit surfaces using permutohedral lattices. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 8466-8475 (2023) 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.445, + 0.787, + 0.487 + ], + "angle": 0, + "content": "38. Sagnik Das, Ke Ma, Z.S., Samaras, D.: Learning an isometric surface parameterization for texture unwrapping. In: European Conference of Computer Vision 2022, ECCV 2022, Tel Aviv, Israel, October 23-27, 2022. Springer (2022) 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.488, + 0.787, + 0.543 + ], + "angle": 0, + "content": "39. Saito, S., Huang, Z., Natsume, R., Morishima, S., Kanazawa, A., Li, H.: Pifu: Pixel-aligned implicit function for high-resolution clothed human digitization. In: Proceedings of the IEEE/CVF international conference on computer vision. pp. 2304-2314 (2019) 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.544, + 0.787, + 0.585 + ], + "angle": 0, + "content": "40. Sheffer, A., Praun, E., Rose, K., et al.: Mesh parameterization methods and their applications. Foundations and Trends in Computer Graphics and Vision 2(2), 105-171 (2007) 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.586, + 0.787, + 0.627 + ], + "angle": 0, + "content": "41. Srinivasan, P.P., Deng, B., Zhang, X., Tancik, M., Mildenhall, B., Barron, J.T.: Nerv: Neural reflectance and visibility fields for relighting and view synthesis. In: CVPR (2021) 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.628, + 0.787, + 0.656 + ], + "angle": 0, + "content": "42. Sun, Q., Zhang, L., Zhang, M., Ying, X., Xin, S., Xia, J., He, Y.: Texture brush: an interactive surface texturing interface. pp. 153-160 (2013) 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.657, + 0.787, + 0.685 + ], + "angle": 0, + "content": "43. Tarini, M., Hormann, K., Cignoni, P., Montani, C.: Polycube-maps. ACM transactions on graphics (TOG) 23(3), 853-860 (2004) 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.686, + 0.787, + 0.727 + ], + "angle": 0, + "content": "44. Tojo, K., Umetani, N.: Recolorable posterization of volumetric radiance fields using visibility-weighted palette extraction. In: Computer Graphics Forum. vol. 41, pp. 149-160. Wiley Online Library (2022) 2, 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.728, + 0.787, + 0.77 + ], + "angle": 0, + "content": "45. Wang, N., Zhang, Y., Li, Z., Fu, Y., Liu, W., Jiang, Y.G.: Pixel2mesh: Generating 3d mesh models from single rgb images. In: Proceedings of the European conference on computer vision (ECCV). pp. 52-67 (2018) 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.771, + 0.787, + 0.812 + ], + "angle": 0, + "content": "46. Wang, P., Liu, L., Liu, Y., Theobalt, C., Komura, T., Wang, W.: Neus: Learning neural implicit surfaces by volume rendering for multi-view reconstruction. NeurIPS (2021) 1, 14" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.813, + 0.787, + 0.84 + ], + "angle": 0, + "content": "47. Wang, X., Zhu, J., Ye, Q., Huo, Y., Ran, Y., Zhong, Z., Chen, J.: Seal-3d: Interactive pixel-level editing for neural radiance fields (2023) 2, 4" + }, + { + "type": "list", + "bbox": [ + 0.214, + 0.147, + 0.788, + 0.84 + ], + "angle": 0, + "content": null + } + ], + [ + { + "type": "page_number", + "bbox": [ + 0.218, + 0.116, + 0.236, + 0.127 + ], + "angle": 0, + "content": "18" + }, + { + "type": "header", + "bbox": [ + 0.272, + 0.115, + 0.352, + 0.127 + ], + "angle": 0, + "content": "B. Xu et al." + }, + { + "type": "ref_text", + "bbox": [ + 0.218, + 0.148, + 0.787, + 0.189 + ], + "angle": 0, + "content": "48. Wang, Y., Rahmann, L., Sorkine-Hornung, O.: Geometry-consistent neural shape representation with implicit displacement fields. In: The Tenth International Conference on Learning Representations. OpenReview (2022) 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.218, + 0.19, + 0.787, + 0.231 + ], + "angle": 0, + "content": "49. Wang, Y., Skorokhodov, I., Wonka, P.: Hf-neus: Improved surface reconstruction using high-frequency details. Advances in Neural Information Processing Systems 35, 1966-1978 (2022) 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.231, + 0.787, + 0.273 + ], + "angle": 0, + "content": "50. Williams, F., Schneider, T., Silva, C., Zorin, D., Bruna, J., Panozzo, D.: Deep geometric prior for surface reconstruction. In: Proceedings of the IEEE/CVF conference on computer vision and pattern recognition. pp. 10130-10139 (2019) 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.273, + 0.787, + 0.314 + ], + "angle": 0, + "content": "51. Wu, T., Zhang, J., Fu, X., Wang, Y., Ren, J., Pan, L., Wu, W., Yang, L., Wang, J., Qian, C., Lin, D., Liu, Z.: Omniobject3d: Large-vocabulary 3d object dataset for realistic perception, reconstruction and generation (2023) 10" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.315, + 0.787, + 0.356 + ], + "angle": 0, + "content": "52. Xiang, F., Xu, Z., Hasan, M., Hold-Geoffroy, Y., Sunkavalli, K., Su, H.: NeuTex: Neural Texture Mapping for Volumetric Neural Rendering. In: The IEEE Conference on Computer Vision and Pattern Recognition (CVPR) (2021) 2, 4, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.356, + 0.787, + 0.412 + ], + "angle": 0, + "content": "53. Xu, B., Zhang, J., Lin, K.Y., Qian, C., He, Y.: Deformable model driven neural rendering for high-fidelity 3d reconstruction of human heads under low-view settings. In: Proceedings of the IEEE/CVF International Conference on Computer Vision (2023) 2, 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.412, + 0.787, + 0.467 + ], + "angle": 0, + "content": "54. Yang, B., Bao, C., Zeng, J., Bao, H., Zhang, Y., Cui, Z., Zhang, G.: Neumesh: Learning disentangled neural mesh-based implicit field for geometry and texture editing. In: European Conference on Computer Vision. pp. 597-614. Springer (2022) 4, 12" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.467, + 0.787, + 0.509 + ], + "angle": 0, + "content": "55. Yang, B., Zhang, Y., Xu, Y., Li, Y., Zhou, H., Bao, H., Zhang, G., Cui, Z.: Learning object-compositional neural radiance field for editable scene rendering. In: International Conference on Computer Vision (ICCV) (October 2021) 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.509, + 0.787, + 0.563 + ], + "angle": 0, + "content": "56. Yang, H., Zhu, H., Wang, Y., Huang, M., Shen, Q., Yang, R., Cao, X.: Facescape: a large-scale high quality 3d face dataset and detailed riggable 3d face prediction. In: Proceedings of the IEEE/cvf conference on computer vision and pattern recognition. pp. 601-610 (2020) 10, 14" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.564, + 0.787, + 0.605 + ], + "angle": 0, + "content": "57. Yang, L., Liang, Y., Li, X., Zhang, C., Lin, G., Sheffer, A., Schaefer, S., Keyser, J., Wang, W.: Neural parametric surfaces for shape modeling. arXiv preprint arXiv:2309.09911 (2023) 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.606, + 0.787, + 0.646 + ], + "angle": 0, + "content": "58. Yariv, L., Gu, J., Kasten, Y., Lipman, Y.: Volume rendering of neural implicit surfaces. Advances in Neural Information Processing Systems 34, 4805-4815 (2021) 1, 9, 14" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.647, + 0.787, + 0.688 + ], + "angle": 0, + "content": "59. Yariv, L., Kasten, Y., Moran, D., Galun, M., Atzmon, M., Ronen, B., Lipman, Y.: Multiview neural surface reconstruction by disentangling geometry and appearance. Advances in Neural Information Processing Systems 33 (2020) 9" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.689, + 0.787, + 0.743 + ], + "angle": 0, + "content": "60. Ye, W., Chen, S., Bao, C., Bao, H., Pollefeys, M., Cui, Z., Zhang, G.: IntrinsicNeRF: Learning Intrinsic Neural Radiance Fields for Editable Novel View Synthesis. In: Proceedings of the IEEE/CVF International Conference on Computer Vision (2023) 4, 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.744, + 0.787, + 0.799 + ], + "angle": 0, + "content": "61. Yenamandra, T., Tewari, A., Bernard, F., Seidel, H.P., Elgharib, M., Cremers, D., Theobalt, C.: i3dmm: Deep implicit 3d morphable model of human heads. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 12803-12813 (2021) 2, 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.799, + 0.787, + 0.841 + ], + "angle": 0, + "content": "62. Yuan, Y.J., Sun, Y.T., Lai, Y.K., Ma, Y., Jia, R., Kobbelt, L., Gao, L.: Interactive nef geometry editing with shape priors. IEEE Transactions on Pattern Analysis and Machine Intelligence (2023) 4" + }, + { + "type": "list", + "bbox": [ + 0.217, + 0.148, + 0.787, + 0.841 + ], + "angle": 0, + "content": null + } + ], + [ + { + "type": "header", + "bbox": [ + 0.659, + 0.115, + 0.732, + 0.127 + ], + "angle": 0, + "content": "NeuParam" + }, + { + "type": "page_number", + "bbox": [ + 0.769, + 0.116, + 0.786, + 0.127 + ], + "angle": 0, + "content": "19" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.147, + 0.788, + 0.189 + ], + "angle": 0, + "content": "63. Zhang, Q., Hou, J., Qian, Y., Chan, A.B., Zhang, J., He, Y.: Reggeonet: Learning regular representations for large-scale 3d point clouds. Int. J. Comput. Vis. 130(12), 3100-3122 (2022) 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.19, + 0.788, + 0.232 + ], + "angle": 0, + "content": "64. Zhang, Q., Hou, J., Qian, Y., Zeng, Y., Zhang, J., He, Y.: Flattening-net: Deep regular 2d representation for 3d point cloud analysis. IEEE Trans. Pattern Anal. Mach. Intell. 45(8), 9726-9742 (2023) 3" + }, + { + "type": "list", + "bbox": [ + 0.217, + 0.147, + 0.788, + 0.232 + ], + "angle": 0, + "content": null + } + ] +] \ No newline at end of file diff --git a/2024/Parameterization-driven Neural Surface Reconstruction for Object-oriented Editing in Neural Rendering/20a75bf0-0a6a-4ea4-b1e9-a52604587383_origin.pdf b/2024/Parameterization-driven Neural Surface Reconstruction for Object-oriented Editing in Neural Rendering/20a75bf0-0a6a-4ea4-b1e9-a52604587383_origin.pdf new file mode 100644 index 0000000000000000000000000000000000000000..909e1564777876161c6ce8b34ec14516fc6060ac --- /dev/null +++ b/2024/Parameterization-driven Neural Surface Reconstruction for Object-oriented Editing in Neural Rendering/20a75bf0-0a6a-4ea4-b1e9-a52604587383_origin.pdf @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:d2634102018a12de369c62ffd5dfd0c59adb29328a6b0765ee8010073e2fb514 +size 9855325 diff --git a/2024/Parameterization-driven Neural Surface Reconstruction for Object-oriented Editing in Neural Rendering/full.md b/2024/Parameterization-driven Neural Surface Reconstruction for Object-oriented Editing in Neural Rendering/full.md new file mode 100644 index 0000000000000000000000000000000000000000..cb08e039701c30303464fe9b2989f3bc10a18a2d --- /dev/null +++ b/2024/Parameterization-driven Neural Surface Reconstruction for Object-oriented Editing in Neural Rendering/full.md @@ -0,0 +1,422 @@ +# Parameterization-driven Neural Surface Reconstruction for Object-oriented Editing in Neural Rendering + +Baixin $\mathrm{Xu}^{1}$ , Jiangbei $\mathrm{Hu}^{2,3}$ , Fei $\mathrm{Hou}^{4,5}$ , Kwan-Yee $\mathrm{Lin}^{6}$ , Wayne $\mathrm{Wu}^{7}$ , Chen Qian $^{7}$ , and Ying $\mathrm{He}^{1,2}$ + +$^{1}$ S-Lab, Nanyang Technological University, $^{2}$ College of Computing and Data Science, Nanyang Technological University, $^{3}$ Dalian University of Technology, $^{4}$ Key Laboratory of System Software (Chinese Academy of Sciences) and State Key Laboratory of Computer Science, Institute of Software, Chinese Academy of Sciences, $^{5}$ University of Chinese Academy of Sciences, $^{6}$ Shanghai AI Lab, $^{7}$ SenseTime Research + +Abstract. The advancements in neural rendering have increased the need for techniques that enable intuitive editing of 3D objects represented as neural implicit surfaces. This paper introduces a novel neural algorithm for parameterizing neural implicit surfaces to simple parametric domains like spheres and polycubes. Our method allows users to specify the number of cubes in the parametric domain, learning a configuration that closely resembles the target 3D object's geometry. It computes bidirectional deformation between the object and the domain using a forward mapping from the object's zero level set and an inverse deformation for backward mapping. We ensure nearly bijective mapping with a cycle loss and optimize deformation smoothness. The parameterization quality, assessed by angle and area distortions, is guaranteed using a Laplacian regularizer and an optimized learned parametric domain. Our framework integrates with existing neural rendering pipelines, using multi-view images of a single object or multiple objects of similar geometries to reconstruct 3D geometry and compute texture maps automatically, eliminating the need for any prior information. We demonstrate the method's effectiveness on images of human heads and man-made objects. The source code is available at https://xubaixinxbx.github.io/neuparam. + +Keywords: Polycube parameterization $\cdot$ Neural surface reconstruction and editing $\cdot$ Neural rendering + +# 1 Introduction + +Neural radiance fields (NeRF) [27] have garnered remarkable success in both the computer vision and computer graphics communities, redefining benchmarks for high-quality renderings and novel view synthesis. Building upon NeRF, a variety of methods leveraging implicit neural representations [30, 46, 58] have emerged, + +delivering high-fidelity 3D reconstructions. The demand for intuitive 3D object editing techniques has increased with the growing capabilities of neural rendering, especially for neural implicit surfaces. Extensive explorations have been made in editing within the Neural Rendering framework to achieve diverse visual effects. By manipulating the latent codes of the network, one can implicitly alter the shape and color of 3D objects [23,53,61]. In the work of [18,44], palette-based methods have been employed to facilitate transformations in the color and style of rendered scenes. However, these methods suffer from limited controllability and disallow local pixel editing. Seal3D [47] proposed a teacher-student training strategy to edit the interactive scene at the pixel level. Yet, this method is tricky to handle shading, and each subsequent edit required a new training iteration. By employing neural networks to formulate a differentiable surface parameterization, it can be facilitated to edit textures or shapes within the neural rendering pipeline [25,52]. Existing neural parameterization techniques still face significant challenges, such as the need for appropriate distortion constraints to ensure accurate detail reconstruction or reducing the reliance on prior information from tracked mesh and UV mapping. The final quality of edited visual effects is significantly impacted by these issues. + +This paper aims to develop an intuitive and easy-to-use tool for appearance editing in neural implicit surfaces within the neural rendering framework, utilizing only 2D images as input. Towards this goal, we propose a novel method to parameterize neural implicit surfaces to parametric domains where the radiance field is represented as a texture map. Distinguished from previous work that using 2D texture mapping, our neural parameterization algorithm supports the selection of different 3D parametric domains according to geometric shapes to avoid seams and reduce distortion, such as spheres, and polycubes. Technically, our parameterization algorithm utilizes learnable parameters to fit a parametric domain first and a neural network to learn a bi-directional deformation between 3D objects and the chosen parametric domains. This involves a forward deformation that maps points from the zero level set of the neural implicit surface to the parametric domain, followed by an inverse deformation, mapping points backward. Notably, we do not require any explicit prior information in learning both deformations. We employ a cycle loss to ensure the smoothness of the bi-directional deformation, and a Laplacian regularization to effectively control angle distortion. With the parameterization, 3D radiance field is mapped to a simple 3D parametric domain which can be further disassembled into a 2D domain, facilitating visualization and various editing tasks. Furthermore, we decompose the 2D radiance field into two components: view-independent material and view-dependent shading, streamlining both texture and shading editing. Our neural parameterization algorithm is fully compatible with existing neural rendering pipelines, allowing 3D reconstruction from multi-view images as input and the creation of texture maps simultaneously. It allows for the immediate rendering of edited textures through volume rendering, without network re-training. Moreover, it supports co-parameterization of objects of similar geometry and enables texture transfer between them. We validate the effectiveness of + +our method on scenes of human heads and man-made objects. Our contributions are summarized as follows: + +- We present a neural parameterization framework that automatically learns a parametric domain and computes bi-directional deformation between 3D objects and their parametric domains, eliminating the need for any prior information from tracked mesh or UV mapping. Our approach utilizes Laplacian loss to minimize angle distortion and allows the user to specify the domain complexity to control area distortion. +- We introduce a simple yet effective technique for intrinsic radiance decomposition, facilitating both view-independent texture editing and view-dependent shading editing. +- Our neural parameterization framework is end-to-end and fully compatible with existing neural rendering architectures, accepting multi-view images as input to reconstruct 3D geometry and generate UV maps as output. Additionally, it enables the direct rendering of modified textures using the volume rendering pipeline. +- Our approach supports co-parameterization of multiple objects and allows for texture transfer between different objects. + +# 2 Related Work + +Parameterization. Surface parameterization [6, 40] aims at computing a bijective mapping between the 3D surface and a suitable parametric domain, usually a 2D region or a simple 3D object, such as spheres [8] and polycubes [7, 43]. Serving as an important computational tool in computer graphics and digital geometry processing [40], surface parameterization facilitates various applications, including texture editing [5], surface painting [42], details transfer [2], remeshing [33], among others. Neural networks are now commonly used in digital geometry processing, resulting in the creation of parameterization algorithms based on deep learning techniques [1, 11, 14, 50]. Unlike classical methods that yield global and seamless parametrization, most neural parameterization methods are computing local parametrization [11, 63, 64]. These methods typically partition the 3D model into multiple patches and parameterize each to a 2D region, possibly followed by a post-processing step to stitch the patches together. Although these methods are efficient and can handle surfaces of arbitrary geometry and topology, they frequently encounter problems including non-bijectivity, a lack of smoothness, the presence of seams and overlaps, and large distortions. There are also works on neural parametric surfaces, defined in rectangular domains [24] and $n$ -sided patches [57]. They can model complex surface geometries with high precision; however, such representations are not applicable to neural implicit surfaces and thereby incompatible with neural rendering. + +Neural Implicit Functions. Recent years have seen a surge in the successful application of neural implicit functions in 3D deep learning. Compared to explicit representations such as point clouds [34], voxels [35], and meshes [45], + +neural implicit representations offer unique advantages including flexibility, continuity, and robustness [26,31,39]. When integrated within the neural rendering pipeline, neural implicit surfaces exhibit exceptional quality in 3D reconstructions from multi-view images [19,20,28,37,48,49,53]. However, as geometries and radiance fields are encoded as network parameters, editing them becomes complex and non-intuitive, compared to explicit representations. This paper aims at tackling this challenge by explicitly representing radiance fields as texture mapping that encompasses both view-dependent and view-independent components. + +NeRF Editing. Numerous studies have focused on editing neural radiance fields, including relighting [21, 41], composition [32, 55], content generation [29], and shape editing [22, 62]. Editing tools operating at the scene- or object-level, such as [16, 55, 61], utilize latent codes to stylize the appearance of an object or entire scenes. Conversely, pixel-level editing tools, such as NeuMesh [54] and Seal3D [47], utilize training-based approaches to edit fine-grained details. However, NeuMesh relies on explicit meshes which lack efficiency and robustness. Although Seal3D achieves interactive editing, it still suffers from poor geometry reconstruction and cannot change the illumination scene. Parameterization-based methods, such as NeuTex [52], ISO-UVField [38] and NeP [25], integrate the learning of UV mapping into neural rendering, enabling intuitive editing within 2D domains. However, these methods often suffer from large distortions in the parametrization. Furthermore, they require 3D prior information from tracked meshes and UV mapping - conditions that may not be readily met in practical, real-world scenarios. IntrinsicNeRF [60] enhances standard neural radiance fields by generating additional outputs, including reflectance, shading, and a residual term. Incorporating a semantic branch, it facilitates real-time scene editing. However, it often results in aggregated colors, akin to palette-based methods [18, 44], hence cannot offer fine-grained editing results at the pixel level. We decompose the radiance field into view-dependent shadings and view-independent materials. Both components are represented as textures, enabling intuitive material modifications while preserving consistent shading. + +# 3 Method + +The input of our method is a collection of RGB images, denoted $\mathcal{I} = \{\mathbf{I}_i\}$ , representing multiple objects of similar geometry or a single object, captured from different viewpoints. To differentiate among objects, we assign a latent shape code $\mathbf{z}_s$ and an appearance code $\mathbf{z}_a$ to each object. This facilitates the co-parameterization of multiple objects and enables the transfer of material and shading among them. As illustrated in Fig. 1, our method reconstructs 3D geometry as a neural implicit surface and encodes the 3D radiance field in a simple parametric domain, such as spheres and polycubes, which we learn from an initial phase in our training (Sec. 3.1). For each input image, we randomly select a set of pixels. Then, for each pixel, we shoot a ray originating from the location of camera and traversing through the pixel. If the ray intersects with the surface, at a point, denoted as $\mathbf{p}$ , we compute its corresponding position in the + +![](images/963da5eeb0bf9c1d84a74bb4a661b1492b3045f853141d1807f00fcb01a79d64.jpg) +Fig. 1: Algorithmic pipeline. Initially, we learn the parametric domain $F_{\mathrm{sdf}}$ using a coarse SDF $G_{\mathrm{sdf}}$ of the target surface $\mathcal{S}$ during the early phase of neural rendering (in the left grey box). Once the parametric domain $\mathcal{D}$ is determined, our pipeline subsequently utilizes the bi-directional deformation $(F_{\mathrm{def}}$ and $F_{\mathrm{inv-def}})$ , the parametric domain SDF $(F_{\mathrm{sdf}})$ and the radiance decomposition $(F_{\mathrm{mat}}$ and $F_{\mathrm{shd}})$ . The output includes the reconstructed 3D surface $\mathcal{S}$ , the decomposed radiance fields (featuring both view-dependent and view-independent components) and the map between $\mathcal{S}$ and $\mathcal{D}$ . It is worth mentioning that while our framework involves two SDFs - one for the 3D surface $\mathcal{S}$ and the other for the parametric domain $\mathcal{D}$ - our network design only requires one of them. This is because one SDF can be derived from the other via either the forward or the backward deformation. To reduce the network complexity, we only adopt one geometry sub-network $F_{\mathrm{sdf}}$ , which is used to represent the parametric domain $\mathcal{D}$ . + +parametric domain $\mathcal{D}$ , denoted as $\mathbf{p}'$ , via a bi-directional deformation (Sec. 3.2): the forward deformation (depicted in orange) generates a displacement vector so that the updated position $\mathbf{p}'$ is in the parametric domain; subsequently, the inverse deformation (depicted in green) maps $\mathbf{p}'$ back to a point $\mathbf{p}''$ , which is anticipated to be in close proximity to the surface. We incorporate a cycle loss to penalize occurrences where $\mathbf{p}''$ deviates from $\mathbf{p}$ . In addition, we employ a Laplacian loss (Sec. 3.3) to effectively reduce the angle distortion between the parametric domain and the original surface. Furthermore, we decompose the radiance into a view-independent material field and a view-dependent shading field (Sec. 3.4), each of which can be edited independently, thus augmenting the editing capability of our framework. See Fig. 1 for the algorithmic pipeline. + +# 3.1 Parametric Domain Learning + +In our framework, we consider two types of parametric domains: spheres and polycubes, both represented as signed distance fields. Spheres are characterized by four learnable parameters: the center position and radius. Polycubes, a generalization of cubes, are simple structures yet offer sufficient degrees of freedom to capture the salient geometric features of target surfaces. We define a learnable polycube with $k$ elements as follows: For the $i$ -th element, which is an axis-aligned box, its SDF $\phi_i$ is given by + +$$ +\phi_ {i} = \max \left(\left| x - x _ {i} \right| - a _ {i}, \left| y - y _ {i} \right| - b _ {i}, \left| z - z _ {i} \right| - c _ {i}\right), \tag {1} +$$ + +where $(x_{i},y_{i},z_{i})$ represents the box center, and $(a_i,b_i,c_i)$ is its dimensions. By combining the $k$ boxes, the SDF of the polycube is obtained as + +$$ +f _ {\mathrm {p c}} = \max (- \phi_ {1}, - \phi_ {2}, \dots , - \phi_ {k}). \tag {2} +$$ + +To ensure the differentiability of Eq. (2), we approximate it using Kreisselmeier-Steinhauser (KS) function [17] as + +$$ +f _ {\mathrm {p c}} \approx - 1 / \lambda * \log \left(\sum_ {i} \exp (- \lambda * \phi_ {i})\right), +$$ + +where $\lambda$ is a typically large value, set to 100 in our experiments. + +Since determining the configuration of boxes does not require a highly accurate geometry of the target object $S$ , we design the parametric domain learning as a one-time, intermediate step in the neural rendering pipeline. Specifically, after an initial training phase of approximately 100 epochs, the coarse geometry of the surface $S$ is obtained. Subsequently, we activate the domain learning module, which takes the SDF $G_{\mathrm{sdf}}$ of the coarse surface $S$ and the user-specified number of boxes $k$ as input. It computes the SDF $F_{\mathrm{sdf}}$ of the parametric domain $\mathcal{D}$ via minimizing the loss function: + +$$ +\mathcal {L} _ {\text {p a r a m}} = \mathcal {L} _ {\text {s d f}} + \lambda_ {s} \mathcal {L} _ {\text {l a p s d f}}, \tag {3} +$$ + +where + +$$ +\mathcal {L} _ {\mathrm {s d f}} = \left| F _ {\mathrm {s d f}} (p) - G _ {\mathrm {s d f}} (p) \right|, \mathcal {L} _ {\mathrm {l a p s d f}} = \left| \frac {\partial^ {2} F _ {\mathrm {s d f}}}{\partial x ^ {2}} + \frac {\partial^ {2} F _ {\mathrm {s d f}}}{\partial y ^ {2}} + \frac {\partial^ {2} F _ {\mathrm {s d f}}}{\partial z ^ {2}} \right|, \tag {4} +$$ + +and the weight $\lambda_{s}$ is set to 0.01. The term $\mathcal{L}_{\mathrm{lapsdf}}$ , which is the Laplacian of the signed distance field of the polycube $\mathcal{D}$ , plays a critical role in controlling its geometric complexity. + +After obtaining the SDF $F_{\mathrm{sdf}}$ of the polycube $\mathcal{D}$ , we deprecate the SDF network $G_{\mathrm{sdf}}$ and instead use the deformation network $F_{\mathrm{def}}$ and parametric domain SDF $F_{\mathrm{sdf}}$ network to represent the source surface $\mathcal{S}$ . The right inset showcases a polycube learning example for a human head model. Given a user-specified parameter $k = 3$ , our method initializes three randomly placed boxes. It proceeds to minimize the loss function $\mathcal{L}_{\mathrm{param}}$ , iteratively refining the size and positioning of each box to ensure the polycube closely approximates the surface $\mathcal{S}$ . This parametric domain learning phase typically converges within 30 to 60 iterations, taking under 10 minutes to complete. For more details, see the supplementary video. + +![](images/f169aa05f57a731bc608eb88a19c769316c937284fa4aa4ffcc2584ddaee1890.jpg) +S + +![](images/c3bc912a01a6b50b0d4032b7e6d298d9bb203835e0752bba878fa696dcaf815d.jpg) +Iter 0 + +![](images/8f82e0edccb21e8f205ff02045a024ab208980f0bfc1820bc264a8d1aab95d62.jpg) +Iter 5 + +![](images/5ed87422f6da40ad1ddb0af8f30248c9032fd45f4d9ac7bb9ff6af0fe0f055c6.jpg) +Iter 10 + +![](images/b309029130d53c4913353dd82b6a113f5e71a7b5a035e104d5a46c1329e0a5f9.jpg) +Iter 45 + +![](images/89aa5be24bf93e0694c1392ef9c152514fcd43115506f3be6644341795aa6e39.jpg) +Iter 55 + +# 3.2 Bi-directional Deformation + +The bi-directional deformation module serves as the pivotal component within our neural parameterization framework. As illustrated in Fig. 1, this module consists of two sub-networks: $F_{\mathrm{def}}$ , which is responsible for the forward mapping from the original model to the parameter domain, and $F_{\mathrm{inv-def}}$ , which establishes the inverse mapping. + +Let $S$ be the 3D surface and $\mathcal{D}$ be the parametric domain; both are represented by the zero level-set of neural signed distance fields. For an arbitrary point $\mathbf{p} \in S$ , the forward deformation $F_{\mathrm{def}}(\mathbf{p}, \mathbf{z}_s)$ computes a displacement vector, mapping the point $\mathbf{p}$ to a corresponding point $\mathbf{p}' \in \mathcal{D}$ in the parametric domain as follows: + +$$ +\mathbf {p} ^ {\prime} = \mathbf {p} + F _ {\mathrm {d e f}} (\mathbf {p}, \mathbf {z} _ {s}). \tag {5} +$$ + +Conversely, we employ $F_{\mathrm{inv - def}}$ to map points inversely from the parametric domain $\mathcal{D}$ back to the original shape $S$ : + +$$ +\mathbf {p} ^ {\prime \prime} = \mathbf {p} ^ {\prime} + F _ {\text {i n v - d e f}} \left(\mathbf {p} ^ {\prime}, \mathbf {z} _ {s}\right), \tag {6} +$$ + +where the point $\mathbf{p}''$ is expected to be in close proximity to $p$ . To encourage the smoothness and stability of the deformations, we impose an $L_{2}$ loss on the magnitude of the displacement vectors: + +$$ +\mathcal {L} _ {\text {s m o o t h}} = \left\| F _ {\text {d e f}} \left(\mathbf {p}, \mathbf {z} _ {s}\right) \right\| _ {2} + \left\| F _ {\text {i n v - d e f}} \left(\mathbf {p} ^ {\prime}, \mathbf {z} _ {s}\right) \right\| _ {2}. \tag {7} +$$ + +This loss helps minimize abrupt and irregular deformations, leading to more coherent and gradual changes between $\mathcal{S}$ and $\mathcal{D}$ . + +In training the networks $F_{\mathrm{def}}$ and $F_{\mathrm{inv - def}}$ , our objective is to achieve a bijective mapping between the 3D surface $S$ and the parametric domain $\mathcal{D}$ . To achieve this goal, we adopt a cycle loss, which is similar to NeuTex [52]: + +$$ +\mathcal {L} _ {\text {c y c l e}} = \sum_ {\mathbf {p} \in S} \lambda (\mathbf {p}) \| \mathbf {p} - \mathbf {p} ^ {\prime \prime} \| _ {2}, \tag {8} +$$ + +where the weight $\lambda$ characterizes the importance of sample $\mathbf{p}$ to the loss. The reason that we do not consider all sample points equally is that a sample point may not be exactly on the zero level-set of the SDF. To tolerate such an inaccuracy, we define $\lambda (\mathbf{p}) = T(\mathbf{p})(1 - \exp (-\sigma (s)\delta (\mathbf{p})))$ as the color weight used in volume rendering [27]. Here, $T(\mathbf{p})$ is the transparency of sample $\mathbf{p}$ in the viewing direction, $\delta$ is the length of the sample interval along the ray, and $s$ is the signed distance value of the point $\mathbf{p}$ . Clearly, when $\mathbf{p}$ is away from the surface, the coefficient $\lambda (\mathbf{p})$ has few effects on this loss. We refer the readers to [27] for details about volume rendering and the computation of transparency $T$ . + +By explicitly minimizing the distance between $\mathbf{p}$ and $\mathbf{p}''$ , the cycle loss effectively prevents scenarios where two distinct points $\mathbf{p}_1, \mathbf{p}_2 \in S$ map to an identical point $\mathbf{p}' \in \mathcal{D}$ . In such cases, the inverse deformation $F_{\mathrm{inv-def}}$ would be unable to map the single point $\mathbf{p}'$ back to the respective distinct points $\mathbf{p}_1''$ and $\mathbf{p}_2''$ . + +# 3.3 Laplacian Regularization + +Mitigating angle distortion is crucial in ensuring the quality of parameterization. One approach to reduce angle distortion involves minimizing the ratio of the two singular values of the Jacobian matrix associated with the parameterization [6]. However, considering the representation of the surface $S$ , the parametric domain $\mathcal{D}$ , and the parameterization $F_{\mathrm{def}}$ as neural implicit functions in our approach, directly computing the Jacobian matrix is a non-trivial task. To circum + +![](images/aaf6a8fe310706b9a07514d27adc58af13eb6ad0f7d8997ac3b934d57e018a0d.jpg) + +vent this, we employ a Laplacian regularizer in our implementation to control angle distortion. Achieving the minimizer of the Laplacian of the map results in a harmonic map. For genus-0 surfaces, harmonic maps are also conformal, which preserves angles [12]. Although for surfaces of higher genus, a harmonic map is not necessarily conformal, it still serves an effective tool for reducing angle distortion [13]. Specifically, for a sample point $\mathbf{p} \in S$ and its neighboring points $\mathbf{q}_j \in S$ , the forward deformation maps them to $\mathbf{p}' \in \mathcal{D}$ and $\mathbf{q}_j' \in \mathcal{D}$ , respectively. We compute the Laplacian of the sample $\mathbf{p}$ as follows: + +$$ +\Delta (\mathbf {p}) = \sum_ {j = 1} ^ {m} \omega_ {j} \left(\mathbf {p} ^ {\prime} - \mathbf {q} _ {j} ^ {\prime}\right). \tag {9} +$$ + +Here, the weights $w_{j}$ are defined using the metric from the original surface $S$ as: + +$$ +\omega_ {j} = \exp (- \| \mathbf {p} - \mathbf {q} _ {j} \| _ {2} / l), \tag {10} +$$ + +where $l$ is the average distance from point $\mathbf{p}$ to its neighboring points. Subsequently, the weights are normalized $\omega_{j} = \omega_{j} / \sum_{k = 1}^{m}\omega_{k}$ . Notice that $\varDelta(\mathbf{p})\in\mathbb{R}^3$ is a 3D vector, which may not necessarily lie on the tangent plane. Therefore, we define the normal component of the Laplacian as: + +$$ +\Delta (\mathbf {p}) _ {\perp} = \langle \Delta (\mathbf {p}), \mathbf {n} \rangle \mathbf {n}, \tag {11} +$$ + +where $\mathbf{n}$ is the normal at $\mathbf{p}'$ , and $\langle, \rangle$ represents the inner product. Summing the tangential components yields the Laplacian loss: + +$$ +\mathcal {L} _ {\mathrm {L a p}} = \sum_ {\mathbf {p} \in \mathcal {S}} \| \Delta (\mathbf {p}) - \Delta (\mathbf {p}) _ {\perp} \| _ {2}. \tag {12} +$$ + +In our implementation, we set $m = 6$ neighboring points for each sample $\mathbf{p}$ . + +# 3.4 Appearance Decomposition + +Radiance represents the directional emission of color, illustrating how the appearance of an object varies in response to changes in the observational viewpoint [27]. Drawing inspiration from previous works [25, 60] and applying the assumptions of Lambertian and grayscale shading [4], we decompose the radiance field on the parametric domain into two components: a view-independent material field and a view-dependent shading field (see Fig. 6(a)). This decomposition allows for independent editing of material and shading. More precisely, for + +In a sample point $\mathbf{p} \in \mathcal{S}$ , we employ two sub-networks, $F_{\mathrm{mat}}$ and $F_{\mathrm{shd}}$ , to decompose the radiance $\mathbf{r} \in \mathbb{R}^3$ at point $\mathbf{p}$ as follows: + +$$ +\mathbf {r} (\mathbf {p}) = F _ {\mathrm {m a t}} \left(\mathbf {p} ^ {\prime}, \mathbf {n}, \mathbf {z} _ {a}\right) \cdot \exp \left(F _ {\mathrm {s h d}} \left(\mathbf {p} ^ {\prime}, \mathbf {n}, \mathbf {v}, \mathbf {z} _ {a}\right)\right), \tag {13} +$$ + +where $\mathbf{p}' \in \mathcal{D}$ is the image of $\mathbf{p}$ and $\mathbf{n}$ is the unit normal of $\mathbf{p}$ on the original surface. The material network $F_{\mathrm{mat}}$ takes the mapped point $\mathbf{p}'$ , the surface normal $\mathbf{n}$ , and the appearance latent code $\mathbf{z}_a$ as input, generating the view-independent material field in response. The shading network $F_{\mathrm{shd}}$ , which is designed for computing the view-dependent shading field, takes an additional input: the viewpoint $\mathbf{v}$ . To ensure the shading sub-network $F_{\mathrm{shd}}$ concentrates on local shading features, we employ an $L_1$ loss to the output of $F_{\mathrm{shd}}$ , encouraging sparsity, + +$$ +\mathcal {L} _ {\mathrm {s h d}} = \left\| F _ {\mathrm {s h d}} \left(\mathbf {p} ^ {\prime}, \mathbf {n}, \mathbf {v}, \mathbf {z} _ {a}\right) \right\| _ {1}. \tag {14} +$$ + +To compute the color, we need the signed distance values for the samples $\mathbf{p}$ . As mentioned above, our network adopts only one SDF network $F_{\mathrm{sdf}}$ , which represents the geometry of the parametric domain $\mathcal{D}$ . We use the forward deformation $F_{\mathrm{def}}$ to compute $s(\mathbf{p})$ as + +$$ +s (\mathbf {p}) = F _ {\mathrm {s d f}} \left(\mathbf {p} + F _ {\mathrm {d e f}} (\mathbf {p}, \mathbf {z} _ {s})\right). \tag {15} +$$ + +Then we employ the SDF-based volume rendering [58] to integrate the radiances along the ray to compute color $\mathbf{c}_{\mathrm{pred}} = \sum_{i}\lambda_{i}\mathbf{r}_{i}$ , where $\mathbf{r}_i$ is the radiance of a sample point $\mathbf{p}_i$ and $\lambda_{i}$ is the same color weight as used in Eq. (8). + +# 3.5 Training Losses + +In contrast to NeP [25], which necessitate prior information such as tracked mesh and UV mapping for supervision, our method is capable of reconstructing 3D geometry and computing the corresponding texture map directly from multiview images. This process is supervised by the image loss + +$$ +\mathcal {L} _ {\mathrm {r g b}} = \left\| \mathbf {c} _ {\mathrm {g t}} - \mathbf {c} _ {\mathrm {p r e d}} \right\| _ {1}, \tag {16} +$$ + +where $\mathbf{c}_{\mathrm{gt}}$ and $\mathbf{c}_{\mathrm{pred}}$ are the ground-truth and predicted colors, respectively. To ensure the shape and appearance latent codes, $\mathbf{z}_s$ and $\mathbf{z}_a$ , conform to Gaussian distributions, we incorporate a regularization term, expressed as $\mathcal{L}_{\mathrm{code}} = \| \mathbf{z}_s\|_2 + \| \mathbf{z}_a\|_2$ [31]. Additionally, with geometry represented as signed distance fields, we employ the Eikonal loss [10, 59], defined as $\mathcal{L}_{\mathrm{Eik}} = (\| \nabla s\|_2 - 1)^2$ . Putting it all together, we define the total loss function as follows: + +$$ +\mathcal {L} = \mathcal {L} _ {\mathrm {r g b}} + \lambda_ {1} \mathcal {L} _ {\text {E i k}} + \lambda_ {2} \mathcal {L} _ {\text {c y c l e}} + \lambda_ {3} \mathcal {L} _ {\text {s m o o t h}} + \lambda_ {4} \mathcal {L} _ {\text {L a p}} + \lambda_ {5} \mathcal {L} _ {\text {s h d}} + \lambda_ {6} \mathcal {L} _ {\text {c o d e}}. \tag {17} +$$ + +In our implementation, we empirically set the coefficients as $\lambda_1 = \lambda_2 = \lambda_5 = \lambda_6 = 0.01$ and $\lambda_3 = \lambda_4 = 0.001$ . + +![](images/5b09bf6d523c14c80d6f045f8073405bb68530e58d45c28c39ab9050e5b41531.jpg) +Fig. 2: Parameterization results. Human heads are co-parameterized to a sphere owing to the clear geometric resemblance. Given the diverse geometry of cars, each is parameterized to a polycube domain (bottom three rows). From top to bottom: the input image, the normal map of the reconstructed geometry $S$ , the parametric domain $\mathcal{D}$ , and the reconstructed surface $S$ . See supplementary material for additional results. + +# 4 Experiments + +Datasets. For our experiments, we used a customized dataset named FS-Syn, derived from the Facescape dataset [56], selecting 10 human head models and configuring 30 fixed viewpoints and lighting conditions to obtain synthesized multi-view images. We also evaluated our method on several man-made objects from the OmniObject dataset [51], and the H3DS dataset [36]. + +Training details and metrics. All the sub-networks $F_{\mathrm{def}}$ , $F_{\mathrm{inv-def}}$ , $F_{\mathrm{sdf}}$ , $F_{\mathrm{mat}}$ , and $F_{\mathrm{shd}}$ are MLPs. See the supplementary material for other detailed network hyperparameters. Once the parametric domain is chosen, we fixed $F_{\mathrm{sdf}}$ and trained other sub-networks for 2,000 epochs driven by the loss function as Eq. (17). We applied the standard Marching Cubes algorithm to extract triangle meshes for the zero level-set of the source surface $S$ and then used the forward map to obtain the mesh representing the parametric surface. With triangle meshes, we measured the quality of the computed parameterization by the angle distortion $\mathcal{E}_{\mathrm{angle}}$ and area distortion [3,15] $\mathcal{E}_{\mathrm{area}}$ as follows: + +$$ +\mathcal {E} _ {\mathrm {a n g l e}} = \frac {\cot \alpha | a | ^ {2} + \cot \beta | b | ^ {2} + \cot \gamma | c | ^ {2}}{4 \mathrm {a r e a} (\varDelta_ {i})}, \quad \mathcal {E} _ {\mathrm {a r e a}} = \frac {1}{2} \left(\frac {\mathrm {a r e a} (\varDelta_ {i})}{\mathrm {a r e a} (\hat {\varDelta} _ {i})} + \frac {\mathrm {a r e a} (\hat {\varDelta} _ {i})}{\mathrm {a r e a} (\varDelta_ {i})}\right), +$$ + +where $a, b, c$ represent the side lengths of the triangle $\Delta_i$ of the original surface. $\alpha, \beta, \gamma$ are angles of the corresponding triangle $\hat{\Delta}_i$ in parametric domain. In the case of conformal mapping, $\mathcal{E}_{\mathrm{angle}}$ is equal to 1, and $\mathcal{E}_{\mathrm{area}}$ is also equal to 1 for an isometric map. + +![](images/6a00a32da747038db7cde00c08403bd135ea08c15804652627fab46f77b186cc.jpg) +1.545/1.396 + +![](images/750fc2cbafbd9ed1e04663849ce79353cd61ae5efbcd23d35bb76c168343d6ab.jpg) +1.755/1.446 + +![](images/7681f59da7455f21547f6e81443fd0eb5233e083fb0bb35df31878d6bcd3aeb8.jpg) +1.394/1.302 +Fig. 3: Results of cross-polycube parameterization on the H3DS dataset [36]. We learned a common polycube domain for all human heads with the parameter $k = 3$ (i.e., the polycube is formed by 3 boxes) and then computed the polycube parameterization for all the human heads simultaneously. Since human heads share similar geometries, the computed polycube parameterizations are also highly consistent. The values below each figure represent the angle distortion and the area distortion, respectively. + +![](images/28bb4180d730adcae61ab5233b9287a83231d102fdcaf4bbb40f3bdcebd5e06b.jpg) +1.271/1.216 + +![](images/6a0347dd2df0b4c34042a16e62a8d674254377381dd8afdf0a7ee0a3359085cc.jpg) +1.351/1.275 + +Results. Our framework constructs a nearly bijective map between objects and multiple parametric domains such as spheres and polycubes, as shown in Fig. 2 and Fig. 3. It supports co-parameterization of multiple objects and facilitates material and shading transfer. Fully compatible with existing neural rendering pipelines, our neural parameterization algorithm uses multi-view images to simultaneously reconstruct 3D geometry and create texture maps. As shown in Fig. 2, we achieve high-fidelity results without prior information. + +![](images/88c65bc948910005a4ba6f345e0bf59c409d1d687eab48214fb3b3446b2ba4e6.jpg) +UV prior +Fig. 4: Parametrization results of NeP [25] in comparison with our method. NeP heavily relies on the initial UV prior, leading to poor texture maps when such prior information is lacking. Our method learns texture maps from a set of multi-view images without any 3D tracked mesh or UV prior as input. + +![](images/f07aa17abbfb2eceb87234b69ea04416b4dd9b4d51a3200fc398044d43d2a2e7.jpg) +NeP w/o prior + +![](images/d3309025c2dccb88f6336bda646ae931e72565ea5f444e48adc1669865705e22.jpg) +w/ prior + +![](images/a8af08f8d760abec17f6d44470ba8b68364aec39d1053fe19ab60b085d341025.jpg) + +![](images/2d8095154c1dde858c8ae7f093ffbc348f641faac0abcd6cf0d5f2729f4b33e2.jpg) +Ours + +Compared with NeP [25], which is the latest neural surface editing work using neural parameterization, our method eliminates the need for prior information from UV mapping. Unlike our algorithm, NeP learns a map between + +surfaces and the UV plane. However, it heavily relies on an existing UV mapping as initialization, and yields rather poor texture mapping results without the UV mapping as prior (see Fig.4). Moreover, the results from NeP still exhibit significant distortions even with UV prior. + +![](images/f61ee14219f8b893bfb539e15bef8f0cd09e3369cd3907e8b11a2abf140e7f5a.jpg) +Fig. 5: Representing the materials as textures defined in a simple parametric domain enables easy editing. Also, our co-parameterization naturally supports texture transfer between different objects. + +Texture editing. Our parameterization results enable flexible and pixel-level editing as shown in Fig. 5. We re-paint the extracted 2D texture map using cubemaps [9] and retrieve the modified color using the established mapping when volume rendering novel images after each editing. Both NeuMesh [54] and NeP [25] can edit the object's texture at a pixel level. We compare the editing and rendering results with NeuMesh and NeP, as shown in Fig. 5. NeuMesh requires a mask of the editing region. It also re-trains the network after each edit, which is time-consuming and cannot guarantee the stability of the training results. Our method outperforms NeuMesh in terms of the visual quality. NeP, which is also neural parameterization-based method, requires a UV prior for initialization. It often yields poor parameterization results without such prior. Even with the UV prior, its parameterization results exhibit large distortion, making the edit in the 2D domain challenging. Our method is free of prior, and can yield parameterization with low distortion, thereby facilitating pixel-level editing and material transferring without network re-training. + +Shading editing. We decompose the 3D radiance field into view-independent material and view-dependent shading defined in the parametric domain. Taking advantage of this decomposition and the embedding into a simple domain, we can edit both the material and shading of the 3D surfaces in an intuitive manner. As illustrated in Fig. 6(b), we modify the shading of one model by transferring the shading from another model under different lighting conditions, while keeping the material unchanged. + +Inverse deformation. To validate the cycle loss and illustrate inverse deformation, we sample points on the polycube parametric surface $\mathcal{D}$ . Subsequently, we inversely deform these points back to the surface $\mathcal{S}$ . The point cloud is visualized in the right inset, and we quantify the CD $(\times 10^{-3})$ with respect to + +![](images/b334bf3e51c1828f978a9f690c8dfdf1028f4f0381218519f94e8c5e0e0bfd5f.jpg) +(a) We decompose the radiance into view-dependent shadings and view-independent materials. The shading map is visualized from the front view. + +![](images/a9133be996c0193c69560d7e14614ce69f391e4cca377547fa373c443f3b6b27.jpg) +(b) We transfer the decomposed shading from one model to another. See also the accompanying video. +Fig. 6: We transfer the decomposed shading from one model to another. For further illustration, please refer to the accompanying video. + +the ground truth. The obtained low CD results confirm the efficacy of inverse deformation and cycle loss. + +Ablation studies. As shown in Fig. 7 (left), the term $\mathcal{L}_{\mathrm{smooth}}$ helps to smooth the process of the two deformations, while $\mathcal{L}_{\mathrm{Lap}}$ further reduces the angle distortion. Our method can effectively reduce angle distortion by minimizing Laplacian loss and reduce area distortion by choosing an appropriate parametric domain. As demonstrated in Fig. 7 (right), we calculate the angle distortion $\mathcal{E}_{\mathrm{angle}}$ of parameterization as described in [3, 15]. We can observe that the Laplacian loss can effectively reduce angle distortion. + +The choice of parametric domain. The different parametric domain influences the quality of parameterization as shown in Fig. 8 (a), which is measured by area distortions. Generally, increasing the number of cubic primitives for better resembling the geometry of the target surface effectively reduces distortions but increases the complexity of the parametric domain. Finding an optimal balance between domain complexity and parameterization quality is challenging. + +# 5 Conclusion + +We introduce a novel neural algorithm for parameterizing 3D surfaces represented by neural implicit functions to a user-defined parametric domain, such as spheres and polycubes. This parametric domain is learned given some user-specified hyperparameters. Utilizing bi-directional deformation, our method is + +![](images/9a8a79ca679bf74a26ce4837890ada73372ce0d612bd80b450abff0695bf7d4e.jpg) + +![](images/9350fc57010a9ffe32c19e532f83c38ddc9769cbe4c421a1c822f1b291a6c606.jpg) +VolSDF 4.65 + +![](images/a39ec8569b33187b8e52116a3b2a0d67eb05cd497e7d06eb6a2aa0e8a0270501.jpg) +Ours 4.87 + +![](images/3c31f91daa790caed9c7085953902d81c1300895a374af3a9709e3a744093d2d.jpg) +Inverse 4.89 + +![](images/e9ea5a3ac843f9e224dd2479efdfaa287aba6c31cff943a2d987f65ca045e5e1.jpg) +GT +Input + +![](images/fb117f9fd62ebcfb351490d013e8ecdb45130a87494e121accb2a93fc82dbb93.jpg) +w/o $\mathcal{L}_{\mathrm{smooth}}\& \mathcal{L}_{\mathrm{Lap}} + \mathcal{L}_{\mathrm{smooth}}$ + +![](images/f7aa11f928aa7ae118782a7d42896ee4386646e259d227c73604894ec405ac82.jpg) +Fig. 7: Ablation studies. On the left, we calculate the CD $(\times 10^{-3})$ based on the point cloud extracted from the inverse deformation module and present the ablation studies for $\mathcal{L}_{\mathrm{smooth}}$ and $\mathcal{L}_{\mathrm{Lap}}$ . On the right, utilization of the Laplacian loss has proven effective in reducing angle distortion $(\mathcal{E}_{\mathrm{angle}})$ , especially in regions with high curvature. + +![](images/1ca98573884c0f11a5df2f798dd8cdfdd53001a0186d47812f92c83625035d5a.jpg) ++Lap + +![](images/4edd16536b809a1e699e0cfef5e3ff8d76596289f56217a5ac6f23bee6d0441e.jpg) +$\mathcal{E}_{\mathrm{angle}} = 1.147$ +$\mathcal{E}_{\mathrm{angle}} = 1.092$ +w/o $\mathcal{L}_{\mathrm{Lap}}$ +W/Lap + +capable of learning a nearly bijective mapping without relying on any prior knowledge, while controlling angle distortion through the usage of a Laplacian regularizer. Furthermore, our approach seamlessly integrates with the neural rendering pipeline, enabling the reconstruction of 3D objects from multi-view images and facilitating efficient volume rendering of modified textures and shadings — eliminating the necessity for network re-training. We demonstrated the efficacy of our method on human heads and man-made objects. + +Our method has a few limitations that require further improvement in the future. First, quantitative comparisons on the FS-Syn dataset show that the reconstruction quality of our method is slightly worse than the standard SDF-based neural rendering algorithms, such as VolSDF [58] and NeuS [46]. For example, as illustrated in Fig. 8 (b) and, We report PSNR and CD compared with VolSDF [58] on Fs-Syn dataset [56]. This decline in visual quality arises from the incorporation of additional modules, such as bi-directional deformation and radiance decomposition, resulting in the cessation of updates to the SDF network within our neural parameterization pipeline. Therefore, computing a parameterization without compromising the reconstruction quality is highly desired. Second, our radiance decomposition operates under the assumptions of Lambertian reflectance and grayscale shading, which might hinder its effectiveness when dealing with intricate materials or sophisticated shadings. + +![](images/8d4289f81d14eb1c9ce65cadae2ab9b437dbd150d0a4bed6e090a50fb327053c.jpg) +S +$\mathcal{E}_{\mathrm{area}} = 2.505$ 6.094 +Fig. 8: We present area distortion $\mathcal{E}_{\mathrm{area}}$ based on the different choices of the parametric domain in (a) and geometry reconstruction results of VolSDF [58] and our method in (b). Our method is capable of editing through parameterization, accompanied by an acceptable compromise in reconstruction quality. + +![](images/6b5e247eb1b1ef116e7444d83f3ffe4f73855ff430a8ffe7d3a2cf17f0a23e71.jpg) + +![](images/babc350263d4387380837fde4a940c8d855b6b527ed7a4ad8e1d8ab8d1f65b4d.jpg) + +![](images/86a7b75151500e7d032cd56115263caea9918a4822ef798c1076b4c1bd5b8205.jpg) +1.633 + +![](images/256ec429d93fffdd5d89afb162e25fd169930e7e1483ae18b86645bd1bf6c960.jpg) +GT + +![](images/6ced61452f26507a3bc11e4290547712db9c5517aebeea9afd80d1f17f459ffb.jpg) +VolSDF + +![](images/384f501987360754e48a3213a7bc6e07ed05d784b086a25116648b59fdcf736c.jpg) +Ours +(b) + +![](images/a83df78bc300a819ef6d625e81ee15521b238974721cba3f29a13da97eec0f34.jpg) + +Method PSNR CD(10-3) + +VolSDF 31.15 3.270 + +Ours 30.87 3.423 + +# Acknowledgment + +This study is supported under the RIE2020 Industry Alignment Fund - Industry Collaboration Projects (IAF-ICP) Funding Initiative, as well as cash and in-kind contribution from the industry partner(s). This project is also partially supported by the Ministry of Education, Singapore, under its Academic Research Fund Grants (MOE-T2EP20220-0005 & RT19/22). + +# References + +1. Bednarik, J., Parashar, S., Gundogdu, E., Salzmann, M., Fua, P.: Shape reconstruction by learning differentiable surface representations. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 4716-4725 (2020) 3 +2. Biermann, H., Martin, I., Bernardini, F., Zorin, D.: Cut-and-paste editing of multiresolution surfaces. ACM transactions on graphics (TOG) 21(3), 312-321 (2002) 3 +3. Degener, P., Meseth, J., Klein, R.: An adaptable surface parameterization method. IMR 3, 201-213 (2003) 10, 13 +4. Fan, Q., Yang, J., Hua, G., Chen, B., Wipf, D.: Revisiting deep intrinsic image decompositions. In: Proceedings of the IEEE conference on computer vision and pattern recognition. pp. 8944-8952 (2018) 8 +5. Fang, H., Hart, J.C.: Textureshop: texture synthesis as a photograph editing tool. ACM Transactions on Graphics (TOG) 23(3), 354-359 (2004) 3 +6. Floater, M.S., Hormann, K.: Surface parameterization: a tutorial and survey. Advances in multiresolution for geometric modelling pp. 157-186 (2005) 3, 8 +7. García, I., Xia, J., He, Y., Xin, S., Patow, G.: Interactive applications for sketch-based editable polycube map. IEEE Trans. Vis. Comput. Graph. 19(7), 1158-1171 (2013) 3 +8. Gotsman, C., Gu, X., Sheffer, A.: Fundamentals of spherical parameterization for 3d meshes. ACM Trans. Graph. 22(3), 358-363 (2003) 3 +9. Greene, N.: Environment mapping and other applications of world projections. IEEE computer graphics and Applications 6(11), 21-29 (1986) 12 +0. Gropp, A., Yariv, L., Haim, N., Atzmon, M., Lipman, Y.: Implicit geometric regularization for learning shapes. In: International Conference on Machine Learning. pp. 3789-3799. PMLR (2020) 9 +1. Groueix, T., Fisher, M., Kim, V.G., Russell, B.C., Aubry, M.: A papier-mâché approach to learning 3d surface generation. In: Proceedings of the IEEE conference on computer vision and pattern recognition. pp. 216-224 (2018) 3 +2. Gu, X., Wang, Y., Chan, T.F., Thompson, P.M., Yau, S.: Genus zero surface conformal mapping and its application to brain surface mapping. IEEE Trans. Medical Imaging 23(8), 949-958 (2004) 8 +3. Gu, X., Yau, S.: Global conformal parameterization. In: Kobbelt, L., Schröder, P., Hoppe, H. (eds.) First Eurographics Symposium on Geometry Processing. vol. 43, pp. 127-137 (2003) 8 +4. Guo, H., Liu, S., Pan, H., Liu, Y., Tong, X., Guo, B.: Complexgen: Cad reconstruction by b-rep chain complex generation. ACM Transactions on Graphics (TOG) 41(4), 1-18 (2022) 3 + +15. He, Y., Wang, H., Fu, C.W., Qin, H.: A divide-and-conquer approach for automatic polycube map construction. Computers & Graphics 33(3), 369-380 (2009) 10, 13 +16. Huang, Y.H., He, Y., Yuan, Y.J., Lai, Y.K., Gao, L.: Stylizednerf: Consistent 3d scene stylization as stylized nerf via 2d-3d mutual learning. In: Computer Vision and Pattern Recognition (CVPR) (2022) 4 +17. Kreisselmeier, G., Steinhauser, R.: Systematic control design by optimizing a vector performance index. In: Computer aided design of control systems, pp. 113-117. Elsevier (1980) 6 +18. Kuang, Z., Luan, F., Bi, S., Shu, Z., Wetzstein, G., Sunkavalli, K.: Palettenerf: Palette-based appearance editing of neural radiance fields. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 20691-20700 (2023) 2, 4 +19. Li, M., Zhang, H.: D2im-net: Learning detail disentangled implicit fields from single images. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 10246-10255 (2021) 4 +20. Li, Z., Müller, T., Evans, A., Taylor, R.H., Unberath, M., Liu, M.Y., Lin, C.H.: Neuralangelo: High-fidelity neural surface reconstruction. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 8456-8465 (2023) 4 +21. Li, Z., Shi, J., Bi, S., Zhu, R., Sunkavalli, K., Hasan, M., Xu, Z., Ramamoorthi, R., Chandraker, M.: Physically-based editing of indoor scene lighting from a single image. In: European Conference on Computer Vision. pp. 555-572. Springer (2022) 4 +22. Lin, C., Mitra, N., Wetzstein, G., Guibas, L.J., Guerrero, P.: Neuform: Adaptive overfitting for neural shape editing. Advances in Neural Information Processing Systems 35, 15217-15229 (2022) 4 +23. Liu, S., Zhang, X., Zhang, Z., Zhang, R., Zhu, J.Y., Russell, B.: Editing conditional radiance fields. In: Proceedings of the IEEE/CVF international conference on computer vision. pp. 5773-5783 (2021) 2 +24. Low, W.F., Lee, G.H.: Minimal neural atlas: Parameterizing complex surfaces with minimal charts and distortion. In: European Conference on Computer Vision. pp. 465-481. Springer (2022) 3 +25. Ma, L., Li, X., Liao, J., Wang, X., Zhang, Q., Wang, J., Sander, P.V.: Neural parameterization for dynamic human head editing. ACM Transactions on Graphics (TOG) 41(6), 1-15 (2022) 2, 4, 8, 9, 11, 12 +26. Mescheder, L., Oechsle, M., Niemeyer, M., Nowozin, S., Geiger, A.: Occupancy networks: Learning 3d reconstruction in function space. In: Proceedings of the IEEE/CVF conference on computer vision and pattern recognition. pp. 4460-4470 (2019) 4 +27. Mildenhall, B., Srinivasan, P.P., Tancik, M., Barron, J.T., Ramamoorthi, R., Ng, R.: Nerf: Representing scenes as neural radiance fields for view synthesis. In: ECCV (2020) 1, 7, 8 +28. Müller, T., Evans, A., Schied, C., Keller, A.: Instant neural graphics primitives with a multiresolution hash encoding. ACM Transactions on Graphics (ToG) 41(4), 1-15 (2022) 4 +29. Niemeyer, M., Geiger, A.: Giraffe: Representing scenes as compositional generative neural feature fields. In: Proc. IEEE Conf. on Computer Vision and Pattern Recognition (CVPR) (2021) 4 +30. Oechsle, M., Peng, S., Geiger, A.: Unisurf: Unifying neural implicit surfaces and radiance fields for multi-view reconstruction. In: Proceedings of the IEEE/CVF International Conference on Computer Vision. pp. 5589-5599 (2021) 1 + +31. Park, J.J., Florence, P., Straub, J., Newcombe, R., Lovegrove, S.: Deepsdf: Learning continuous signed distance functions for shape representation. In: Proceedings of the IEEE/CVF conference on computer vision and pattern recognition. pp. 165-174 (2019) 4, 9 +32. Pérez, P., Gangnet, M., Blake, A.: Poisson image editing. In: Seminal Graphics Papers: Pushing the Boundaries, Volume 2, pp. 577-582 (2023) 4 +33. Praun, E., Hoppe, H.: Spherical parametrization and remeshing. ACM transactions on graphics (TOG) 22(3), 340-349 (2003) 3 +34. Qi, C.R., Su, H., Mo, K., Guibas, L.J.: Pointnet: Deep learning on point sets for 3d classification and segmentation. In: Proceedings of the IEEE conference on computer vision and pattern recognition. pp. 652-660 (2017) 3 +35. Qi, C.R., Su, H., Nießner, M., Dai, A., Yan, M., Guibas, L.J.: Volumetric and multi-view cnns for object classification on 3d data. In: Proceedings of the IEEE conference on computer vision and pattern recognition. pp. 5648-5656 (2016) 3 +36. Ramon, E., Triginer, G., Escur, J., Pumarola, A., Garcia, J., Giro-i Nieto, X., Moreno-Noguer, F.: H3d-net: Few-shot high-fidelity 3d head reconstruction. In: Proceedings of the IEEE/CVF International Conference on Computer Vision. pp. 5620-5629 (2021) 10, 11 +37. Rosu, R.A., Behnke, S.: Permutosdf: Fast multi-view reconstruction with implicit surfaces using permutohedral lattices. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 8466-8475 (2023) 4 +38. Sagnik Das, Ke Ma, Z.S., Samaras, D.: Learning an isometric surface parameterization for texture unwrapping. In: European Conference of Computer Vision 2022, ECCV 2022, Tel Aviv, Israel, October 23-27, 2022. Springer (2022) 4 +39. Saito, S., Huang, Z., Natsume, R., Morishima, S., Kanazawa, A., Li, H.: Pifu: Pixel-aligned implicit function for high-resolution clothed human digitization. In: Proceedings of the IEEE/CVF international conference on computer vision. pp. 2304-2314 (2019) 4 +40. Sheffer, A., Praun, E., Rose, K., et al.: Mesh parameterization methods and their applications. Foundations and Trends in Computer Graphics and Vision 2(2), 105-171 (2007) 3 +41. Srinivasan, P.P., Deng, B., Zhang, X., Tancik, M., Mildenhall, B., Barron, J.T.: Nerv: Neural reflectance and visibility fields for relighting and view synthesis. In: CVPR (2021) 4 +42. Sun, Q., Zhang, L., Zhang, M., Ying, X., Xin, S., Xia, J., He, Y.: Texture brush: an interactive surface texturing interface. pp. 153-160 (2013) 3 +43. Tarini, M., Hormann, K., Cignoni, P., Montani, C.: Polycube-maps. ACM transactions on graphics (TOG) 23(3), 853-860 (2004) 3 +44. Tojo, K., Umetani, N.: Recolorable posterization of volumetric radiance fields using visibility-weighted palette extraction. In: Computer Graphics Forum. vol. 41, pp. 149-160. Wiley Online Library (2022) 2, 4 +45. Wang, N., Zhang, Y., Li, Z., Fu, Y., Liu, W., Jiang, Y.G.: Pixel2mesh: Generating 3d mesh models from single rgb images. In: Proceedings of the European conference on computer vision (ECCV). pp. 52-67 (2018) 3 +46. Wang, P., Liu, L., Liu, Y., Theobalt, C., Komura, T., Wang, W.: Neus: Learning neural implicit surfaces by volume rendering for multi-view reconstruction. NeurIPS (2021) 1, 14 +47. Wang, X., Zhu, J., Ye, Q., Huo, Y., Ran, Y., Zhong, Z., Chen, J.: Seal-3d: Interactive pixel-level editing for neural radiance fields (2023) 2, 4 + +48. Wang, Y., Rahmann, L., Sorkine-Hornung, O.: Geometry-consistent neural shape representation with implicit displacement fields. In: The Tenth International Conference on Learning Representations. OpenReview (2022) 4 +49. Wang, Y., Skorokhodov, I., Wonka, P.: Hf-neus: Improved surface reconstruction using high-frequency details. Advances in Neural Information Processing Systems 35, 1966-1978 (2022) 4 +50. Williams, F., Schneider, T., Silva, C., Zorin, D., Bruna, J., Panozzo, D.: Deep geometric prior for surface reconstruction. In: Proceedings of the IEEE/CVF conference on computer vision and pattern recognition. pp. 10130-10139 (2019) 3 +51. Wu, T., Zhang, J., Fu, X., Wang, Y., Ren, J., Pan, L., Wu, W., Yang, L., Wang, J., Qian, C., Lin, D., Liu, Z.: Omniobject3d: Large-vocabulary 3d object dataset for realistic perception, reconstruction and generation (2023) 10 +52. Xiang, F., Xu, Z., Hasan, M., Hold-Geoffroy, Y., Sunkavalli, K., Su, H.: NeuTex: Neural Texture Mapping for Volumetric Neural Rendering. In: The IEEE Conference on Computer Vision and Pattern Recognition (CVPR) (2021) 2, 4, 7 +53. Xu, B., Zhang, J., Lin, K.Y., Qian, C., He, Y.: Deformable model driven neural rendering for high-fidelity 3d reconstruction of human heads under low-view settings. In: Proceedings of the IEEE/CVF International Conference on Computer Vision (2023) 2, 4 +54. Yang, B., Bao, C., Zeng, J., Bao, H., Zhang, Y., Cui, Z., Zhang, G.: Neumesh: Learning disentangled neural mesh-based implicit field for geometry and texture editing. In: European Conference on Computer Vision. pp. 597-614. Springer (2022) 4, 12 +55. Yang, B., Zhang, Y., Xu, Y., Li, Y., Zhou, H., Bao, H., Zhang, G., Cui, Z.: Learning object-compositional neural radiance field for editable scene rendering. In: International Conference on Computer Vision (ICCV) (October 2021) 4 +56. Yang, H., Zhu, H., Wang, Y., Huang, M., Shen, Q., Yang, R., Cao, X.: Facescape: a large-scale high quality 3d face dataset and detailed riggable 3d face prediction. In: Proceedings of the IEEE/cvf conference on computer vision and pattern recognition. pp. 601-610 (2020) 10, 14 +57. Yang, L., Liang, Y., Li, X., Zhang, C., Lin, G., Sheffer, A., Schaefer, S., Keyser, J., Wang, W.: Neural parametric surfaces for shape modeling. arXiv preprint arXiv:2309.09911 (2023) 3 +58. Yariv, L., Gu, J., Kasten, Y., Lipman, Y.: Volume rendering of neural implicit surfaces. Advances in Neural Information Processing Systems 34, 4805-4815 (2021) 1, 9, 14 +59. Yariv, L., Kasten, Y., Moran, D., Galun, M., Atzmon, M., Ronen, B., Lipman, Y.: Multiview neural surface reconstruction by disentangling geometry and appearance. Advances in Neural Information Processing Systems 33 (2020) 9 +60. Ye, W., Chen, S., Bao, C., Bao, H., Pollefeys, M., Cui, Z., Zhang, G.: IntrinsicNeRF: Learning Intrinsic Neural Radiance Fields for Editable Novel View Synthesis. In: Proceedings of the IEEE/CVF International Conference on Computer Vision (2023) 4, 8 +61. Yenamandra, T., Tewari, A., Bernard, F., Seidel, H.P., Elgharib, M., Cremers, D., Theobalt, C.: i3dmm: Deep implicit 3d morphable model of human heads. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 12803-12813 (2021) 2, 4 +62. Yuan, Y.J., Sun, Y.T., Lai, Y.K., Ma, Y., Jia, R., Kobbelt, L., Gao, L.: Interactive nef geometry editing with shape priors. IEEE Transactions on Pattern Analysis and Machine Intelligence (2023) 4 + +63. Zhang, Q., Hou, J., Qian, Y., Chan, A.B., Zhang, J., He, Y.: Reggeonet: Learning regular representations for large-scale 3d point clouds. Int. J. Comput. Vis. 130(12), 3100-3122 (2022) 3 +64. Zhang, Q., Hou, J., Qian, Y., Zeng, Y., Zhang, J., He, Y.: Flattening-net: Deep regular 2d representation for 3d point cloud analysis. IEEE Trans. Pattern Anal. Mach. Intell. 45(8), 9726-9742 (2023) 3 \ No newline at end of file diff --git a/2024/Parameterization-driven Neural Surface Reconstruction for Object-oriented Editing in Neural Rendering/images.zip b/2024/Parameterization-driven Neural Surface Reconstruction for Object-oriented Editing in Neural Rendering/images.zip new file mode 100644 index 0000000000000000000000000000000000000000..c90f7177546375e4d3be91236060b7b5dd4f5cd5 --- /dev/null +++ b/2024/Parameterization-driven Neural Surface Reconstruction for Object-oriented Editing in Neural Rendering/images.zip @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:a814bfc48429f4f17678662d1e952fd1937e863482a22cacc92ee121d623fbf6 +size 428273 diff --git a/2024/Parameterization-driven Neural Surface Reconstruction for Object-oriented Editing in Neural Rendering/layout.json b/2024/Parameterization-driven Neural Surface Reconstruction for Object-oriented Editing in Neural Rendering/layout.json new file mode 100644 index 0000000000000000000000000000000000000000..3733714bb5d0a3c8136f47302f76bfa0c98cf3df --- /dev/null +++ b/2024/Parameterization-driven Neural Surface Reconstruction for Object-oriented Editing in Neural Rendering/layout.json @@ -0,0 +1,13962 @@ +{ + "pdf_info": [ + { + "para_blocks": [ + { + "bbox": [ + 148, + 111, + 465, + 166 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 148, + 111, + 465, + 166 + ], + "spans": [ + { + "bbox": [ + 148, + 111, + 465, + 166 + ], + "type": "text", + "content": "Parameterization-driven Neural Surface Reconstruction for Object-oriented Editing in Neural Rendering" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 136, + 186, + 476, + 210 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 136, + 186, + 476, + 210 + ], + "spans": [ + { + "bbox": [ + 136, + 186, + 476, + 210 + ], + "type": "text", + "content": "Baixin " + }, + { + "bbox": [ + 136, + 186, + 476, + 210 + ], + "type": "inline_equation", + "content": "\\mathrm{Xu}^{1}" + }, + { + "bbox": [ + 136, + 186, + 476, + 210 + ], + "type": "text", + "content": ", Jiangbei " + }, + { + "bbox": [ + 136, + 186, + 476, + 210 + ], + "type": "inline_equation", + "content": "\\mathrm{Hu}^{2,3}" + }, + { + "bbox": [ + 136, + 186, + 476, + 210 + ], + "type": "text", + "content": ", Fei " + }, + { + "bbox": [ + 136, + 186, + 476, + 210 + ], + "type": "inline_equation", + "content": "\\mathrm{Hou}^{4,5}" + }, + { + "bbox": [ + 136, + 186, + 476, + 210 + ], + "type": "text", + "content": ", Kwan-Yee " + }, + { + "bbox": [ + 136, + 186, + 476, + 210 + ], + "type": "inline_equation", + "content": "\\mathrm{Lin}^{6}" + }, + { + "bbox": [ + 136, + 186, + 476, + 210 + ], + "type": "text", + "content": ", Wayne " + }, + { + "bbox": [ + 136, + 186, + 476, + 210 + ], + "type": "inline_equation", + "content": "\\mathrm{Wu}^{7}" + }, + { + "bbox": [ + 136, + 186, + 476, + 210 + ], + "type": "text", + "content": ", Chen Qian" + }, + { + "bbox": [ + 136, + 186, + 476, + 210 + ], + "type": "inline_equation", + "content": "^{7}" + }, + { + "bbox": [ + 136, + 186, + 476, + 210 + ], + "type": "text", + "content": ", and Ying " + }, + { + "bbox": [ + 136, + 186, + 476, + 210 + ], + "type": "inline_equation", + "content": "\\mathrm{He}^{1,2}" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 132, + 220, + 480, + 275 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 220, + 480, + 275 + ], + "spans": [ + { + "bbox": [ + 132, + 220, + 480, + 275 + ], + "type": "inline_equation", + "content": "^{1}" + }, + { + "bbox": [ + 132, + 220, + 480, + 275 + ], + "type": "text", + "content": "S-Lab, Nanyang Technological University, " + }, + { + "bbox": [ + 132, + 220, + 480, + 275 + ], + "type": "inline_equation", + "content": "^{2}" + }, + { + "bbox": [ + 132, + 220, + 480, + 275 + ], + "type": "text", + "content": "College of Computing and Data Science, Nanyang Technological University, " + }, + { + "bbox": [ + 132, + 220, + 480, + 275 + ], + "type": "inline_equation", + "content": "^{3}" + }, + { + "bbox": [ + 132, + 220, + 480, + 275 + ], + "type": "text", + "content": "Dalian University of Technology, " + }, + { + "bbox": [ + 132, + 220, + 480, + 275 + ], + "type": "inline_equation", + "content": "^{4}" + }, + { + "bbox": [ + 132, + 220, + 480, + 275 + ], + "type": "text", + "content": "Key Laboratory of System Software (Chinese Academy of Sciences) and State Key Laboratory of Computer Science, Institute of Software, Chinese Academy of Sciences, " + }, + { + "bbox": [ + 132, + 220, + 480, + 275 + ], + "type": "inline_equation", + "content": "^{5}" + }, + { + "bbox": [ + 132, + 220, + 480, + 275 + ], + "type": "text", + "content": "University of Chinese Academy of Sciences, " + }, + { + "bbox": [ + 132, + 220, + 480, + 275 + ], + "type": "inline_equation", + "content": "^{6}" + }, + { + "bbox": [ + 132, + 220, + 480, + 275 + ], + "type": "text", + "content": "Shanghai AI Lab, " + }, + { + "bbox": [ + 132, + 220, + 480, + 275 + ], + "type": "inline_equation", + "content": "^{7}" + }, + { + "bbox": [ + 132, + 220, + 480, + 275 + ], + "type": "text", + "content": "SenseTime Research" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 160, + 307, + 452, + 515 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 160, + 307, + 452, + 515 + ], + "spans": [ + { + "bbox": [ + 160, + 307, + 452, + 515 + ], + "type": "text", + "content": "Abstract. The advancements in neural rendering have increased the need for techniques that enable intuitive editing of 3D objects represented as neural implicit surfaces. This paper introduces a novel neural algorithm for parameterizing neural implicit surfaces to simple parametric domains like spheres and polycubes. Our method allows users to specify the number of cubes in the parametric domain, learning a configuration that closely resembles the target 3D object's geometry. It computes bidirectional deformation between the object and the domain using a forward mapping from the object's zero level set and an inverse deformation for backward mapping. We ensure nearly bijective mapping with a cycle loss and optimize deformation smoothness. The parameterization quality, assessed by angle and area distortions, is guaranteed using a Laplacian regularizer and an optimized learned parametric domain. Our framework integrates with existing neural rendering pipelines, using multi-view images of a single object or multiple objects of similar geometries to reconstruct 3D geometry and compute texture maps automatically, eliminating the need for any prior information. We demonstrate the method's effectiveness on images of human heads and man-made objects. The source code is available at https://xubaixinxbx.github.io/neuparam." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 159, + 526, + 453, + 549 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 159, + 526, + 453, + 549 + ], + "spans": [ + { + "bbox": [ + 159, + 526, + 453, + 549 + ], + "type": "text", + "content": "Keywords: Polycube parameterization " + }, + { + "bbox": [ + 159, + 526, + 453, + 549 + ], + "type": "inline_equation", + "content": "\\cdot" + }, + { + "bbox": [ + 159, + 526, + 453, + 549 + ], + "type": "text", + "content": " Neural surface reconstruction and editing " + }, + { + "bbox": [ + 159, + 526, + 453, + 549 + ], + "type": "inline_equation", + "content": "\\cdot" + }, + { + "bbox": [ + 159, + 526, + 453, + 549 + ], + "type": "text", + "content": " Neural rendering" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 132, + 571, + 230, + 582 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 571, + 230, + 582 + ], + "spans": [ + { + "bbox": [ + 132, + 571, + 230, + 582 + ], + "type": "text", + "content": "1 Introduction" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 130, + 597, + 482, + 646 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 597, + 482, + 646 + ], + "spans": [ + { + "bbox": [ + 130, + 597, + 482, + 646 + ], + "type": "text", + "content": "Neural radiance fields (NeRF) [27] have garnered remarkable success in both the computer vision and computer graphics communities, redefining benchmarks for high-quality renderings and novel view synthesis. Building upon NeRF, a variety of methods leveraging implicit neural representations [30, 46, 58] have emerged," + } + ] + } + ], + "index": 6 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 141, + 654, + 374, + 666 + ], + "type": "page_footnote", + "angle": 0, + "lines": [ + { + "bbox": [ + 141, + 654, + 374, + 666 + ], + "spans": [ + { + "bbox": [ + 141, + 654, + 374, + 666 + ], + "type": "text", + "content": "Corresponding author: Ying He (email: yhe@ntu.edu.sg)." + } + ] + } + ], + "index": 7 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 0 + }, + { + "para_blocks": [ + { + "bbox": [ + 130, + 115, + 482, + 332 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 115, + 482, + 332 + ], + "spans": [ + { + "bbox": [ + 130, + 115, + 482, + 332 + ], + "type": "text", + "content": "delivering high-fidelity 3D reconstructions. The demand for intuitive 3D object editing techniques has increased with the growing capabilities of neural rendering, especially for neural implicit surfaces. Extensive explorations have been made in editing within the Neural Rendering framework to achieve diverse visual effects. By manipulating the latent codes of the network, one can implicitly alter the shape and color of 3D objects [23,53,61]. In the work of [18,44], palette-based methods have been employed to facilitate transformations in the color and style of rendered scenes. However, these methods suffer from limited controllability and disallow local pixel editing. Seal3D [47] proposed a teacher-student training strategy to edit the interactive scene at the pixel level. Yet, this method is tricky to handle shading, and each subsequent edit required a new training iteration. By employing neural networks to formulate a differentiable surface parameterization, it can be facilitated to edit textures or shapes within the neural rendering pipeline [25,52]. Existing neural parameterization techniques still face significant challenges, such as the need for appropriate distortion constraints to ensure accurate detail reconstruction or reducing the reliance on prior information from tracked mesh and UV mapping. The final quality of edited visual effects is significantly impacted by these issues." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 130, + 342, + 482, + 668 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 342, + 482, + 668 + ], + "spans": [ + { + "bbox": [ + 130, + 342, + 482, + 668 + ], + "type": "text", + "content": "This paper aims to develop an intuitive and easy-to-use tool for appearance editing in neural implicit surfaces within the neural rendering framework, utilizing only 2D images as input. Towards this goal, we propose a novel method to parameterize neural implicit surfaces to parametric domains where the radiance field is represented as a texture map. Distinguished from previous work that using 2D texture mapping, our neural parameterization algorithm supports the selection of different 3D parametric domains according to geometric shapes to avoid seams and reduce distortion, such as spheres, and polycubes. Technically, our parameterization algorithm utilizes learnable parameters to fit a parametric domain first and a neural network to learn a bi-directional deformation between 3D objects and the chosen parametric domains. This involves a forward deformation that maps points from the zero level set of the neural implicit surface to the parametric domain, followed by an inverse deformation, mapping points backward. Notably, we do not require any explicit prior information in learning both deformations. We employ a cycle loss to ensure the smoothness of the bi-directional deformation, and a Laplacian regularization to effectively control angle distortion. With the parameterization, 3D radiance field is mapped to a simple 3D parametric domain which can be further disassembled into a 2D domain, facilitating visualization and various editing tasks. Furthermore, we decompose the 2D radiance field into two components: view-independent material and view-dependent shading, streamlining both texture and shading editing. Our neural parameterization algorithm is fully compatible with existing neural rendering pipelines, allowing 3D reconstruction from multi-view images as input and the creation of texture maps simultaneously. It allows for the immediate rendering of edited textures through volume rendering, without network re-training. Moreover, it supports co-parameterization of objects of similar geometry and enables texture transfer between them. We validate the effectiveness of" + } + ] + } + ], + "index": 3 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "spans": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "type": "text", + "content": "2" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 166, + 91, + 216, + 101 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 166, + 91, + 216, + 101 + ], + "spans": [ + { + "bbox": [ + 166, + 91, + 216, + 101 + ], + "type": "text", + "content": "B. Xu et al." + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 1 + }, + { + "para_blocks": [ + { + "bbox": [ + 132, + 116, + 481, + 139 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 116, + 481, + 139 + ], + "spans": [ + { + "bbox": [ + 132, + 116, + 481, + 139 + ], + "type": "text", + "content": "our method on scenes of human heads and man-made objects. Our contributions are summarized as follows:" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 138, + 149, + 482, + 342 + ], + "type": "list", + "angle": 0, + "index": 7, + "blocks": [ + { + "bbox": [ + 138, + 149, + 482, + 220 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 149, + 482, + 220 + ], + "spans": [ + { + "bbox": [ + 138, + 149, + 482, + 220 + ], + "type": "text", + "content": "- We present a neural parameterization framework that automatically learns a parametric domain and computes bi-directional deformation between 3D objects and their parametric domains, eliminating the need for any prior information from tracked mesh or UV mapping. Our approach utilizes Laplacian loss to minimize angle distortion and allows the user to specify the domain complexity to control area distortion." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 138, + 222, + 482, + 258 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 222, + 482, + 258 + ], + "spans": [ + { + "bbox": [ + 138, + 222, + 482, + 258 + ], + "type": "text", + "content": "- We introduce a simple yet effective technique for intrinsic radiance decomposition, facilitating both view-independent texture editing and view-dependent shading editing." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 138, + 258, + 482, + 317 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 258, + 482, + 317 + ], + "spans": [ + { + "bbox": [ + 138, + 258, + 482, + 317 + ], + "type": "text", + "content": "- Our neural parameterization framework is end-to-end and fully compatible with existing neural rendering architectures, accepting multi-view images as input to reconstruct 3D geometry and generate UV maps as output. Additionally, it enables the direct rendering of modified textures using the volume rendering pipeline." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 138, + 319, + 482, + 342 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 319, + 482, + 342 + ], + "spans": [ + { + "bbox": [ + 138, + 319, + 482, + 342 + ], + "type": "text", + "content": "- Our approach supports co-parameterization of multiple objects and allows for texture transfer between different objects." + } + ] + } + ], + "index": 6 + } + ], + "sub_type": "text" + }, + { + "bbox": [ + 132, + 362, + 237, + 374 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 362, + 237, + 374 + ], + "spans": [ + { + "bbox": [ + 132, + 362, + 237, + 374 + ], + "type": "text", + "content": "2 Related Work" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 132, + 389, + 482, + 628 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 389, + 482, + 628 + ], + "spans": [ + { + "bbox": [ + 132, + 389, + 482, + 628 + ], + "type": "text", + "content": "Parameterization. Surface parameterization [6, 40] aims at computing a bijective mapping between the 3D surface and a suitable parametric domain, usually a 2D region or a simple 3D object, such as spheres [8] and polycubes [7, 43]. Serving as an important computational tool in computer graphics and digital geometry processing [40], surface parameterization facilitates various applications, including texture editing [5], surface painting [42], details transfer [2], remeshing [33], among others. Neural networks are now commonly used in digital geometry processing, resulting in the creation of parameterization algorithms based on deep learning techniques [1, 11, 14, 50]. Unlike classical methods that yield global and seamless parametrization, most neural parameterization methods are computing local parametrization [11, 63, 64]. These methods typically partition the 3D model into multiple patches and parameterize each to a 2D region, possibly followed by a post-processing step to stitch the patches together. Although these methods are efficient and can handle surfaces of arbitrary geometry and topology, they frequently encounter problems including non-bijectivity, a lack of smoothness, the presence of seams and overlaps, and large distortions. There are also works on neural parametric surfaces, defined in rectangular domains [24] and " + }, + { + "bbox": [ + 132, + 389, + 482, + 628 + ], + "type": "inline_equation", + "content": "n" + }, + { + "bbox": [ + 132, + 389, + 482, + 628 + ], + "type": "text", + "content": "-sided patches [57]. They can model complex surface geometries with high precision; however, such representations are not applicable to neural implicit surfaces and thereby incompatible with neural rendering." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 132, + 630, + 482, + 667 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 630, + 482, + 667 + ], + "spans": [ + { + "bbox": [ + 132, + 630, + 482, + 667 + ], + "type": "text", + "content": "Neural Implicit Functions. Recent years have seen a surge in the successful application of neural implicit functions in 3D deep learning. Compared to explicit representations such as point clouds [34], voxels [35], and meshes [45]," + } + ] + } + ], + "index": 10 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 403, + 91, + 447, + 101 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 403, + 91, + 447, + 101 + ], + "spans": [ + { + "bbox": [ + 403, + 91, + 447, + 101 + ], + "type": "text", + "content": "NeuParam" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 474, + 92, + 480, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 474, + 92, + 480, + 100 + ], + "spans": [ + { + "bbox": [ + 474, + 92, + 480, + 100 + ], + "type": "text", + "content": "3" + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 2 + }, + { + "para_blocks": [ + { + "bbox": [ + 130, + 116, + 482, + 212 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 116, + 482, + 212 + ], + "spans": [ + { + "bbox": [ + 130, + 116, + 482, + 212 + ], + "type": "text", + "content": "neural implicit representations offer unique advantages including flexibility, continuity, and robustness [26,31,39]. When integrated within the neural rendering pipeline, neural implicit surfaces exhibit exceptional quality in 3D reconstructions from multi-view images [19,20,28,37,48,49,53]. However, as geometries and radiance fields are encoded as network parameters, editing them becomes complex and non-intuitive, compared to explicit representations. This paper aims at tackling this challenge by explicitly representing radiance fields as texture mapping that encompasses both view-dependent and view-independent components." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 130, + 213, + 482, + 475 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 213, + 482, + 475 + ], + "spans": [ + { + "bbox": [ + 130, + 213, + 482, + 475 + ], + "type": "text", + "content": "NeRF Editing. Numerous studies have focused on editing neural radiance fields, including relighting [21, 41], composition [32, 55], content generation [29], and shape editing [22, 62]. Editing tools operating at the scene- or object-level, such as [16, 55, 61], utilize latent codes to stylize the appearance of an object or entire scenes. Conversely, pixel-level editing tools, such as NeuMesh [54] and Seal3D [47], utilize training-based approaches to edit fine-grained details. However, NeuMesh relies on explicit meshes which lack efficiency and robustness. Although Seal3D achieves interactive editing, it still suffers from poor geometry reconstruction and cannot change the illumination scene. Parameterization-based methods, such as NeuTex [52], ISO-UVField [38] and NeP [25], integrate the learning of UV mapping into neural rendering, enabling intuitive editing within 2D domains. However, these methods often suffer from large distortions in the parametrization. Furthermore, they require 3D prior information from tracked meshes and UV mapping - conditions that may not be readily met in practical, real-world scenarios. IntrinsicNeRF [60] enhances standard neural radiance fields by generating additional outputs, including reflectance, shading, and a residual term. Incorporating a semantic branch, it facilitates real-time scene editing. However, it often results in aggregated colors, akin to palette-based methods [18, 44], hence cannot offer fine-grained editing results at the pixel level. We decompose the radiance field into view-dependent shadings and view-independent materials. Both components are represented as textures, enabling intuitive material modifications while preserving consistent shading." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 132, + 495, + 201, + 508 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 495, + 201, + 508 + ], + "spans": [ + { + "bbox": [ + 132, + 495, + 201, + 508 + ], + "type": "text", + "content": "3 Method" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 130, + 521, + 482, + 666 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 521, + 482, + 666 + ], + "spans": [ + { + "bbox": [ + 130, + 521, + 482, + 666 + ], + "type": "text", + "content": "The input of our method is a collection of RGB images, denoted " + }, + { + "bbox": [ + 130, + 521, + 482, + 666 + ], + "type": "inline_equation", + "content": "\\mathcal{I} = \\{\\mathbf{I}_i\\}" + }, + { + "bbox": [ + 130, + 521, + 482, + 666 + ], + "type": "text", + "content": ", representing multiple objects of similar geometry or a single object, captured from different viewpoints. To differentiate among objects, we assign a latent shape code " + }, + { + "bbox": [ + 130, + 521, + 482, + 666 + ], + "type": "inline_equation", + "content": "\\mathbf{z}_s" + }, + { + "bbox": [ + 130, + 521, + 482, + 666 + ], + "type": "text", + "content": " and an appearance code " + }, + { + "bbox": [ + 130, + 521, + 482, + 666 + ], + "type": "inline_equation", + "content": "\\mathbf{z}_a" + }, + { + "bbox": [ + 130, + 521, + 482, + 666 + ], + "type": "text", + "content": " to each object. This facilitates the co-parameterization of multiple objects and enables the transfer of material and shading among them. As illustrated in Fig. 1, our method reconstructs 3D geometry as a neural implicit surface and encodes the 3D radiance field in a simple parametric domain, such as spheres and polycubes, which we learn from an initial phase in our training (Sec. 3.1). For each input image, we randomly select a set of pixels. Then, for each pixel, we shoot a ray originating from the location of camera and traversing through the pixel. If the ray intersects with the surface, at a point, denoted as " + }, + { + "bbox": [ + 130, + 521, + 482, + 666 + ], + "type": "inline_equation", + "content": "\\mathbf{p}" + }, + { + "bbox": [ + 130, + 521, + 482, + 666 + ], + "type": "text", + "content": ", we compute its corresponding position in the" + } + ] + } + ], + "index": 5 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "spans": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "type": "text", + "content": "4" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 166, + 91, + 216, + 101 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 166, + 91, + 216, + 101 + ], + "spans": [ + { + "bbox": [ + 166, + 91, + 216, + 101 + ], + "type": "text", + "content": "B. Xu et al." + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 3 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 133, + 114, + 478, + 208 + ], + "blocks": [ + { + "bbox": [ + 133, + 114, + 478, + 208 + ], + "lines": [ + { + "bbox": [ + 133, + 114, + 478, + 208 + ], + "spans": [ + { + "bbox": [ + 133, + 114, + 478, + 208 + ], + "type": "image", + "image_path": "963da5eeb0bf9c1d84a74bb4a661b1492b3045f853141d1807f00fcb01a79d64.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 130, + 216, + 482, + 350 + ], + "lines": [ + { + "bbox": [ + 130, + 216, + 482, + 350 + ], + "spans": [ + { + "bbox": [ + 130, + 216, + 482, + 350 + ], + "type": "text", + "content": "Fig. 1: Algorithmic pipeline. Initially, we learn the parametric domain " + }, + { + "bbox": [ + 130, + 216, + 482, + 350 + ], + "type": "inline_equation", + "content": "F_{\\mathrm{sdf}}" + }, + { + "bbox": [ + 130, + 216, + 482, + 350 + ], + "type": "text", + "content": " using a coarse SDF " + }, + { + "bbox": [ + 130, + 216, + 482, + 350 + ], + "type": "inline_equation", + "content": "G_{\\mathrm{sdf}}" + }, + { + "bbox": [ + 130, + 216, + 482, + 350 + ], + "type": "text", + "content": " of the target surface " + }, + { + "bbox": [ + 130, + 216, + 482, + 350 + ], + "type": "inline_equation", + "content": "\\mathcal{S}" + }, + { + "bbox": [ + 130, + 216, + 482, + 350 + ], + "type": "text", + "content": " during the early phase of neural rendering (in the left grey box). Once the parametric domain " + }, + { + "bbox": [ + 130, + 216, + 482, + 350 + ], + "type": "inline_equation", + "content": "\\mathcal{D}" + }, + { + "bbox": [ + 130, + 216, + 482, + 350 + ], + "type": "text", + "content": " is determined, our pipeline subsequently utilizes the bi-directional deformation " + }, + { + "bbox": [ + 130, + 216, + 482, + 350 + ], + "type": "inline_equation", + "content": "(F_{\\mathrm{def}}" + }, + { + "bbox": [ + 130, + 216, + 482, + 350 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 130, + 216, + 482, + 350 + ], + "type": "inline_equation", + "content": "F_{\\mathrm{inv-def}})" + }, + { + "bbox": [ + 130, + 216, + 482, + 350 + ], + "type": "text", + "content": ", the parametric domain SDF " + }, + { + "bbox": [ + 130, + 216, + 482, + 350 + ], + "type": "inline_equation", + "content": "(F_{\\mathrm{sdf}})" + }, + { + "bbox": [ + 130, + 216, + 482, + 350 + ], + "type": "text", + "content": " and the radiance decomposition " + }, + { + "bbox": [ + 130, + 216, + 482, + 350 + ], + "type": "inline_equation", + "content": "(F_{\\mathrm{mat}}" + }, + { + "bbox": [ + 130, + 216, + 482, + 350 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 130, + 216, + 482, + 350 + ], + "type": "inline_equation", + "content": "F_{\\mathrm{shd}})" + }, + { + "bbox": [ + 130, + 216, + 482, + 350 + ], + "type": "text", + "content": ". The output includes the reconstructed 3D surface " + }, + { + "bbox": [ + 130, + 216, + 482, + 350 + ], + "type": "inline_equation", + "content": "\\mathcal{S}" + }, + { + "bbox": [ + 130, + 216, + 482, + 350 + ], + "type": "text", + "content": ", the decomposed radiance fields (featuring both view-dependent and view-independent components) and the map between " + }, + { + "bbox": [ + 130, + 216, + 482, + 350 + ], + "type": "inline_equation", + "content": "\\mathcal{S}" + }, + { + "bbox": [ + 130, + 216, + 482, + 350 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 130, + 216, + 482, + 350 + ], + "type": "inline_equation", + "content": "\\mathcal{D}" + }, + { + "bbox": [ + 130, + 216, + 482, + 350 + ], + "type": "text", + "content": ". It is worth mentioning that while our framework involves two SDFs - one for the 3D surface " + }, + { + "bbox": [ + 130, + 216, + 482, + 350 + ], + "type": "inline_equation", + "content": "\\mathcal{S}" + }, + { + "bbox": [ + 130, + 216, + 482, + 350 + ], + "type": "text", + "content": " and the other for the parametric domain " + }, + { + "bbox": [ + 130, + 216, + 482, + 350 + ], + "type": "inline_equation", + "content": "\\mathcal{D}" + }, + { + "bbox": [ + 130, + 216, + 482, + 350 + ], + "type": "text", + "content": " - our network design only requires one of them. This is because one SDF can be derived from the other via either the forward or the backward deformation. To reduce the network complexity, we only adopt one geometry sub-network " + }, + { + "bbox": [ + 130, + 216, + 482, + 350 + ], + "type": "inline_equation", + "content": "F_{\\mathrm{sdf}}" + }, + { + "bbox": [ + 130, + 216, + 482, + 350 + ], + "type": "text", + "content": ", which is used to represent the parametric domain " + }, + { + "bbox": [ + 130, + 216, + 482, + 350 + ], + "type": "inline_equation", + "content": "\\mathcal{D}" + }, + { + "bbox": [ + 130, + 216, + 482, + 350 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_caption" + } + ], + "index": 2 + }, + { + "bbox": [ + 130, + 374, + 482, + 507 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 374, + 482, + 507 + ], + "spans": [ + { + "bbox": [ + 130, + 374, + 482, + 507 + ], + "type": "text", + "content": "parametric domain " + }, + { + "bbox": [ + 130, + 374, + 482, + 507 + ], + "type": "inline_equation", + "content": "\\mathcal{D}" + }, + { + "bbox": [ + 130, + 374, + 482, + 507 + ], + "type": "text", + "content": ", denoted as " + }, + { + "bbox": [ + 130, + 374, + 482, + 507 + ], + "type": "inline_equation", + "content": "\\mathbf{p}'" + }, + { + "bbox": [ + 130, + 374, + 482, + 507 + ], + "type": "text", + "content": ", via a bi-directional deformation (Sec. 3.2): the forward deformation (depicted in orange) generates a displacement vector so that the updated position " + }, + { + "bbox": [ + 130, + 374, + 482, + 507 + ], + "type": "inline_equation", + "content": "\\mathbf{p}'" + }, + { + "bbox": [ + 130, + 374, + 482, + 507 + ], + "type": "text", + "content": " is in the parametric domain; subsequently, the inverse deformation (depicted in green) maps " + }, + { + "bbox": [ + 130, + 374, + 482, + 507 + ], + "type": "inline_equation", + "content": "\\mathbf{p}'" + }, + { + "bbox": [ + 130, + 374, + 482, + 507 + ], + "type": "text", + "content": " back to a point " + }, + { + "bbox": [ + 130, + 374, + 482, + 507 + ], + "type": "inline_equation", + "content": "\\mathbf{p}''" + }, + { + "bbox": [ + 130, + 374, + 482, + 507 + ], + "type": "text", + "content": ", which is anticipated to be in close proximity to the surface. We incorporate a cycle loss to penalize occurrences where " + }, + { + "bbox": [ + 130, + 374, + 482, + 507 + ], + "type": "inline_equation", + "content": "\\mathbf{p}''" + }, + { + "bbox": [ + 130, + 374, + 482, + 507 + ], + "type": "text", + "content": " deviates from " + }, + { + "bbox": [ + 130, + 374, + 482, + 507 + ], + "type": "inline_equation", + "content": "\\mathbf{p}" + }, + { + "bbox": [ + 130, + 374, + 482, + 507 + ], + "type": "text", + "content": ". In addition, we employ a Laplacian loss (Sec. 3.3) to effectively reduce the angle distortion between the parametric domain and the original surface. Furthermore, we decompose the radiance into a view-independent material field and a view-dependent shading field (Sec. 3.4), each of which can be edited independently, thus augmenting the editing capability of our framework. See Fig. 1 for the algorithmic pipeline." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 132, + 530, + 308, + 544 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 530, + 308, + 544 + ], + "spans": [ + { + "bbox": [ + 132, + 530, + 308, + 544 + ], + "type": "text", + "content": "3.1 Parametric Domain Learning" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 130, + 557, + 482, + 641 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 557, + 482, + 641 + ], + "spans": [ + { + "bbox": [ + 130, + 557, + 482, + 641 + ], + "type": "text", + "content": "In our framework, we consider two types of parametric domains: spheres and polycubes, both represented as signed distance fields. Spheres are characterized by four learnable parameters: the center position and radius. Polycubes, a generalization of cubes, are simple structures yet offer sufficient degrees of freedom to capture the salient geometric features of target surfaces. We define a learnable polycube with " + }, + { + "bbox": [ + 130, + 557, + 482, + 641 + ], + "type": "inline_equation", + "content": "k" + }, + { + "bbox": [ + 130, + 557, + 482, + 641 + ], + "type": "text", + "content": " elements as follows: For the " + }, + { + "bbox": [ + 130, + 557, + 482, + 641 + ], + "type": "inline_equation", + "content": "i" + }, + { + "bbox": [ + 130, + 557, + 482, + 641 + ], + "type": "text", + "content": "-th element, which is an axis-aligned box, its SDF " + }, + { + "bbox": [ + 130, + 557, + 482, + 641 + ], + "type": "inline_equation", + "content": "\\phi_i" + }, + { + "bbox": [ + 130, + 557, + 482, + 641 + ], + "type": "text", + "content": " is given by" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 196, + 653, + 481, + 666 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 196, + 653, + 481, + 666 + ], + "spans": [ + { + "bbox": [ + 196, + 653, + 481, + 666 + ], + "type": "interline_equation", + "content": "\\phi_ {i} = \\max \\left(\\left| x - x _ {i} \\right| - a _ {i}, \\left| y - y _ {i} \\right| - b _ {i}, \\left| z - z _ {i} \\right| - c _ {i}\\right), \\tag {1}", + "image_path": "c2a15d8edd0bdd4e5fd94fc081eb9596f91de7f1dadc1ad3060e53109a97cdf9.jpg" + } + ] + } + ], + "index": 7 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 403, + 91, + 447, + 100 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 403, + 91, + 447, + 100 + ], + "spans": [ + { + "bbox": [ + 403, + 91, + 447, + 100 + ], + "type": "text", + "content": "NeuParam" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 474, + 92, + 480, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 474, + 92, + 480, + 100 + ], + "spans": [ + { + "bbox": [ + 474, + 92, + 480, + 100 + ], + "type": "text", + "content": "5" + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 4 + }, + { + "para_blocks": [ + { + "bbox": [ + 130, + 116, + 479, + 140 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 116, + 479, + 140 + ], + "spans": [ + { + "bbox": [ + 130, + 116, + 479, + 140 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 130, + 116, + 479, + 140 + ], + "type": "inline_equation", + "content": "(x_{i},y_{i},z_{i})" + }, + { + "bbox": [ + 130, + 116, + 479, + 140 + ], + "type": "text", + "content": " represents the box center, and " + }, + { + "bbox": [ + 130, + 116, + 479, + 140 + ], + "type": "inline_equation", + "content": "(a_i,b_i,c_i)" + }, + { + "bbox": [ + 130, + 116, + 479, + 140 + ], + "type": "text", + "content": " is its dimensions. By combining the " + }, + { + "bbox": [ + 130, + 116, + 479, + 140 + ], + "type": "inline_equation", + "content": "k" + }, + { + "bbox": [ + 130, + 116, + 479, + 140 + ], + "type": "text", + "content": " boxes, the SDF of the polycube is obtained as" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 238, + 152, + 481, + 167 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 238, + 152, + 481, + 167 + ], + "spans": [ + { + "bbox": [ + 238, + 152, + 481, + 167 + ], + "type": "interline_equation", + "content": "f _ {\\mathrm {p c}} = \\max (- \\phi_ {1}, - \\phi_ {2}, \\dots , - \\phi_ {k}). \\tag {2}", + "image_path": "1bc38447b2338f59374cdb8553cbb902e2ad52eed679cecf689c87d3931b2aa2.jpg" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 130, + 177, + 479, + 202 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 177, + 479, + 202 + ], + "spans": [ + { + "bbox": [ + 130, + 177, + 479, + 202 + ], + "type": "text", + "content": "To ensure the differentiability of Eq. (2), we approximate it using Kreisselmeier-Steinhauser (KS) function [17] as" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 221, + 213, + 391, + 246 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 221, + 213, + 391, + 246 + ], + "spans": [ + { + "bbox": [ + 221, + 213, + 391, + 246 + ], + "type": "interline_equation", + "content": "f _ {\\mathrm {p c}} \\approx - 1 / \\lambda * \\log \\left(\\sum_ {i} \\exp (- \\lambda * \\phi_ {i})\\right),", + "image_path": "fc656113a5b3ead510a4530ab797d11c091caea7d295678b225be1a5df8b8845.jpg" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 131, + 256, + 414, + 268 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 131, + 256, + 414, + 268 + ], + "spans": [ + { + "bbox": [ + 131, + 256, + 414, + 268 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 131, + 256, + 414, + 268 + ], + "type": "inline_equation", + "content": "\\lambda" + }, + { + "bbox": [ + 131, + 256, + 414, + 268 + ], + "type": "text", + "content": " is a typically large value, set to 100 in our experiments." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 130, + 270, + 480, + 365 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 270, + 480, + 365 + ], + "spans": [ + { + "bbox": [ + 130, + 270, + 480, + 365 + ], + "type": "text", + "content": "Since determining the configuration of boxes does not require a highly accurate geometry of the target object " + }, + { + "bbox": [ + 130, + 270, + 480, + 365 + ], + "type": "inline_equation", + "content": "S" + }, + { + "bbox": [ + 130, + 270, + 480, + 365 + ], + "type": "text", + "content": ", we design the parametric domain learning as a one-time, intermediate step in the neural rendering pipeline. Specifically, after an initial training phase of approximately 100 epochs, the coarse geometry of the surface " + }, + { + "bbox": [ + 130, + 270, + 480, + 365 + ], + "type": "inline_equation", + "content": "S" + }, + { + "bbox": [ + 130, + 270, + 480, + 365 + ], + "type": "text", + "content": " is obtained. Subsequently, we activate the domain learning module, which takes the SDF " + }, + { + "bbox": [ + 130, + 270, + 480, + 365 + ], + "type": "inline_equation", + "content": "G_{\\mathrm{sdf}}" + }, + { + "bbox": [ + 130, + 270, + 480, + 365 + ], + "type": "text", + "content": " of the coarse surface " + }, + { + "bbox": [ + 130, + 270, + 480, + 365 + ], + "type": "inline_equation", + "content": "S" + }, + { + "bbox": [ + 130, + 270, + 480, + 365 + ], + "type": "text", + "content": " and the user-specified number of boxes " + }, + { + "bbox": [ + 130, + 270, + 480, + 365 + ], + "type": "inline_equation", + "content": "k" + }, + { + "bbox": [ + 130, + 270, + 480, + 365 + ], + "type": "text", + "content": " as input. It computes the SDF " + }, + { + "bbox": [ + 130, + 270, + 480, + 365 + ], + "type": "inline_equation", + "content": "F_{\\mathrm{sdf}}" + }, + { + "bbox": [ + 130, + 270, + 480, + 365 + ], + "type": "text", + "content": " of the parametric domain " + }, + { + "bbox": [ + 130, + 270, + 480, + 365 + ], + "type": "inline_equation", + "content": "\\mathcal{D}" + }, + { + "bbox": [ + 130, + 270, + 480, + 365 + ], + "type": "text", + "content": " via minimizing the loss function:" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 248, + 377, + 481, + 392 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 248, + 377, + 481, + 392 + ], + "spans": [ + { + "bbox": [ + 248, + 377, + 481, + 392 + ], + "type": "interline_equation", + "content": "\\mathcal {L} _ {\\text {p a r a m}} = \\mathcal {L} _ {\\text {s d f}} + \\lambda_ {s} \\mathcal {L} _ {\\text {l a p s d f}}, \\tag {3}", + "image_path": "50d9754cc166e4cfed3ff8fddb6426e202018fcb2d15ff13d78c2a9e58f936e0.jpg" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 131, + 403, + 161, + 413 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 131, + 403, + 161, + 413 + ], + "spans": [ + { + "bbox": [ + 131, + 403, + 161, + 413 + ], + "type": "text", + "content": "where" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 164, + 422, + 480, + 450 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 164, + 422, + 480, + 450 + ], + "spans": [ + { + "bbox": [ + 164, + 422, + 480, + 450 + ], + "type": "interline_equation", + "content": "\\mathcal {L} _ {\\mathrm {s d f}} = \\left| F _ {\\mathrm {s d f}} (p) - G _ {\\mathrm {s d f}} (p) \\right|, \\mathcal {L} _ {\\mathrm {l a p s d f}} = \\left| \\frac {\\partial^ {2} F _ {\\mathrm {s d f}}}{\\partial x ^ {2}} + \\frac {\\partial^ {2} F _ {\\mathrm {s d f}}}{\\partial y ^ {2}} + \\frac {\\partial^ {2} F _ {\\mathrm {s d f}}}{\\partial z ^ {2}} \\right|, \\tag {4}", + "image_path": "7d0e7b06356f53da119d39d9602bef8126270035926decf38d2d11b757063c54.jpg" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 130, + 460, + 479, + 496 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 460, + 479, + 496 + ], + "spans": [ + { + "bbox": [ + 130, + 460, + 479, + 496 + ], + "type": "text", + "content": "and the weight " + }, + { + "bbox": [ + 130, + 460, + 479, + 496 + ], + "type": "inline_equation", + "content": "\\lambda_{s}" + }, + { + "bbox": [ + 130, + 460, + 479, + 496 + ], + "type": "text", + "content": " is set to 0.01. The term " + }, + { + "bbox": [ + 130, + 460, + 479, + 496 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_{\\mathrm{lapsdf}}" + }, + { + "bbox": [ + 130, + 460, + 479, + 496 + ], + "type": "text", + "content": ", which is the Laplacian of the signed distance field of the polycube " + }, + { + "bbox": [ + 130, + 460, + 479, + 496 + ], + "type": "inline_equation", + "content": "\\mathcal{D}" + }, + { + "bbox": [ + 130, + 460, + 479, + 496 + ], + "type": "text", + "content": ", plays a critical role in controlling its geometric complexity." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 130, + 498, + 369, + 666 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 498, + 369, + 666 + ], + "spans": [ + { + "bbox": [ + 130, + 498, + 369, + 666 + ], + "type": "text", + "content": "After obtaining the SDF " + }, + { + "bbox": [ + 130, + 498, + 369, + 666 + ], + "type": "inline_equation", + "content": "F_{\\mathrm{sdf}}" + }, + { + "bbox": [ + 130, + 498, + 369, + 666 + ], + "type": "text", + "content": " of the polycube " + }, + { + "bbox": [ + 130, + 498, + 369, + 666 + ], + "type": "inline_equation", + "content": "\\mathcal{D}" + }, + { + "bbox": [ + 130, + 498, + 369, + 666 + ], + "type": "text", + "content": ", we deprecate the SDF network " + }, + { + "bbox": [ + 130, + 498, + 369, + 666 + ], + "type": "inline_equation", + "content": "G_{\\mathrm{sdf}}" + }, + { + "bbox": [ + 130, + 498, + 369, + 666 + ], + "type": "text", + "content": " and instead use the deformation network " + }, + { + "bbox": [ + 130, + 498, + 369, + 666 + ], + "type": "inline_equation", + "content": "F_{\\mathrm{def}}" + }, + { + "bbox": [ + 130, + 498, + 369, + 666 + ], + "type": "text", + "content": " and parametric domain SDF " + }, + { + "bbox": [ + 130, + 498, + 369, + 666 + ], + "type": "inline_equation", + "content": "F_{\\mathrm{sdf}}" + }, + { + "bbox": [ + 130, + 498, + 369, + 666 + ], + "type": "text", + "content": " network to represent the source surface " + }, + { + "bbox": [ + 130, + 498, + 369, + 666 + ], + "type": "inline_equation", + "content": "\\mathcal{S}" + }, + { + "bbox": [ + 130, + 498, + 369, + 666 + ], + "type": "text", + "content": ". The right inset showcases a polycube learning example for a human head model. Given a user-specified parameter " + }, + { + "bbox": [ + 130, + 498, + 369, + 666 + ], + "type": "inline_equation", + "content": "k = 3" + }, + { + "bbox": [ + 130, + 498, + 369, + 666 + ], + "type": "text", + "content": ", our method initializes three randomly placed boxes. It proceeds to minimize the loss function " + }, + { + "bbox": [ + 130, + 498, + 369, + 666 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_{\\mathrm{param}}" + }, + { + "bbox": [ + 130, + 498, + 369, + 666 + ], + "type": "text", + "content": ", iteratively refining the size and positioning of each box to ensure the polycube closely approximates the surface " + }, + { + "bbox": [ + 130, + 498, + 369, + 666 + ], + "type": "inline_equation", + "content": "\\mathcal{S}" + }, + { + "bbox": [ + 130, + 498, + 369, + 666 + ], + "type": "text", + "content": ". This parametric domain learning phase typically converges within 30 to 60 iterations, taking under 10 minutes to complete. For more details, see the supplementary video." + } + ] + } + ], + "index": 12 + }, + { + "type": "image", + "bbox": [ + 378, + 523, + 408, + 561 + ], + "blocks": [ + { + "bbox": [ + 378, + 523, + 408, + 561 + ], + "lines": [ + { + "bbox": [ + 378, + 523, + 408, + 561 + ], + "spans": [ + { + "bbox": [ + 378, + 523, + 408, + 561 + ], + "type": "image", + "image_path": "f169aa05f57a731bc608eb88a19c769316c937284fa4aa4ffcc2584ddaee1890.jpg" + } + ] + } + ], + "index": 13, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 389, + 565, + 397, + 574 + ], + "lines": [ + { + "bbox": [ + 389, + 565, + 397, + 574 + ], + "spans": [ + { + "bbox": [ + 389, + 565, + 397, + 574 + ], + "type": "text", + "content": "S" + } + ] + } + ], + "index": 14, + "angle": 0, + "type": "image_caption" + } + ], + "index": 13 + }, + { + "type": "image", + "bbox": [ + 412, + 523, + 442, + 562 + ], + "blocks": [ + { + "bbox": [ + 412, + 523, + 442, + 562 + ], + "lines": [ + { + "bbox": [ + 412, + 523, + 442, + 562 + ], + "spans": [ + { + "bbox": [ + 412, + 523, + 442, + 562 + ], + "type": "image", + "image_path": "c3bc912a01a6b50b0d4032b7e6d298d9bb203835e0752bba878fa696dcaf815d.jpg" + } + ] + } + ], + "index": 15, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 416, + 565, + 440, + 574 + ], + "lines": [ + { + "bbox": [ + 416, + 565, + 440, + 574 + ], + "spans": [ + { + "bbox": [ + 416, + 565, + 440, + 574 + ], + "type": "text", + "content": "Iter 0" + } + ] + } + ], + "index": 16, + "angle": 0, + "type": "image_caption" + } + ], + "index": 15 + }, + { + "type": "image", + "bbox": [ + 450, + 525, + 475, + 561 + ], + "blocks": [ + { + "bbox": [ + 450, + 525, + 475, + 561 + ], + "lines": [ + { + "bbox": [ + 450, + 525, + 475, + 561 + ], + "spans": [ + { + "bbox": [ + 450, + 525, + 475, + 561 + ], + "type": "image", + "image_path": "8f82e0edccb21e8f205ff02045a024ab208980f0bfc1820bc264a8d1aab95d62.jpg" + } + ] + } + ], + "index": 17, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 451, + 565, + 474, + 574 + ], + "lines": [ + { + "bbox": [ + 451, + 565, + 474, + 574 + ], + "spans": [ + { + "bbox": [ + 451, + 565, + 474, + 574 + ], + "type": "text", + "content": "Iter 5" + } + ] + } + ], + "index": 18, + "angle": 0, + "type": "image_caption" + } + ], + "index": 17 + }, + { + "type": "image", + "bbox": [ + 380, + 578, + 408, + 614 + ], + "blocks": [ + { + "bbox": [ + 380, + 578, + 408, + 614 + ], + "lines": [ + { + "bbox": [ + 380, + 578, + 408, + 614 + ], + "spans": [ + { + "bbox": [ + 380, + 578, + 408, + 614 + ], + "type": "image", + "image_path": "5ed87422f6da40ad1ddb0af8f30248c9032fd45f4d9ac7bb9ff6af0fe0f055c6.jpg" + } + ] + } + ], + "index": 19, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 380, + 618, + 408, + 628 + ], + "lines": [ + { + "bbox": [ + 380, + 618, + 408, + 628 + ], + "spans": [ + { + "bbox": [ + 380, + 618, + 408, + 628 + ], + "type": "text", + "content": "Iter 10" + } + ] + } + ], + "index": 21, + "angle": 0, + "type": "image_caption" + } + ], + "index": 19 + }, + { + "type": "image", + "bbox": [ + 413, + 577, + 442, + 613 + ], + "blocks": [ + { + "bbox": [ + 413, + 577, + 442, + 613 + ], + "lines": [ + { + "bbox": [ + 413, + 577, + 442, + 613 + ], + "spans": [ + { + "bbox": [ + 413, + 577, + 442, + 613 + ], + "type": "image", + "image_path": "b309029130d53c4913353dd82b6a113f5e71a7b5a035e104d5a46c1329e0a5f9.jpg" + } + ] + } + ], + "index": 20, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 414, + 618, + 442, + 628 + ], + "lines": [ + { + "bbox": [ + 414, + 618, + 442, + 628 + ], + "spans": [ + { + "bbox": [ + 414, + 618, + 442, + 628 + ], + "type": "text", + "content": "Iter 45" + } + ] + } + ], + "index": 22, + "angle": 0, + "type": "image_caption" + } + ], + "index": 20 + }, + { + "type": "image", + "bbox": [ + 450, + 577, + 477, + 614 + ], + "blocks": [ + { + "bbox": [ + 450, + 577, + 477, + 614 + ], + "lines": [ + { + "bbox": [ + 450, + 577, + 477, + 614 + ], + "spans": [ + { + "bbox": [ + 450, + 577, + 477, + 614 + ], + "type": "image", + "image_path": "89aa5be24bf93e0694c1392ef9c152514fcd43115506f3be6644341795aa6e39.jpg" + } + ] + } + ], + "index": 23, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 449, + 618, + 477, + 628 + ], + "lines": [ + { + "bbox": [ + 449, + 618, + 477, + 628 + ], + "spans": [ + { + "bbox": [ + 449, + 618, + 477, + 628 + ], + "type": "text", + "content": "Iter 55" + } + ] + } + ], + "index": 24, + "angle": 0, + "type": "image_caption" + } + ], + "index": 23 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "spans": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "type": "text", + "content": "6" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 166, + 91, + 216, + 101 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 166, + 91, + 216, + 101 + ], + "spans": [ + { + "bbox": [ + 166, + 91, + 216, + 101 + ], + "type": "text", + "content": "B. Xu et al." + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 5 + }, + { + "para_blocks": [ + { + "bbox": [ + 132, + 116, + 296, + 126 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 116, + 296, + 126 + ], + "spans": [ + { + "bbox": [ + 132, + 116, + 296, + 126 + ], + "type": "text", + "content": "3.2 Bi-directional Deformation" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 130, + 133, + 479, + 192 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 133, + 479, + 192 + ], + "spans": [ + { + "bbox": [ + 130, + 133, + 479, + 192 + ], + "type": "text", + "content": "The bi-directional deformation module serves as the pivotal component within our neural parameterization framework. As illustrated in Fig. 1, this module consists of two sub-networks: " + }, + { + "bbox": [ + 130, + 133, + 479, + 192 + ], + "type": "inline_equation", + "content": "F_{\\mathrm{def}}" + }, + { + "bbox": [ + 130, + 133, + 479, + 192 + ], + "type": "text", + "content": ", which is responsible for the forward mapping from the original model to the parameter domain, and " + }, + { + "bbox": [ + 130, + 133, + 479, + 192 + ], + "type": "inline_equation", + "content": "F_{\\mathrm{inv-def}}" + }, + { + "bbox": [ + 130, + 133, + 479, + 192 + ], + "type": "text", + "content": ", which establishes the inverse mapping." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 130, + 193, + 479, + 251 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 193, + 479, + 251 + ], + "spans": [ + { + "bbox": [ + 130, + 193, + 479, + 251 + ], + "type": "text", + "content": "Let " + }, + { + "bbox": [ + 130, + 193, + 479, + 251 + ], + "type": "inline_equation", + "content": "S" + }, + { + "bbox": [ + 130, + 193, + 479, + 251 + ], + "type": "text", + "content": " be the 3D surface and " + }, + { + "bbox": [ + 130, + 193, + 479, + 251 + ], + "type": "inline_equation", + "content": "\\mathcal{D}" + }, + { + "bbox": [ + 130, + 193, + 479, + 251 + ], + "type": "text", + "content": " be the parametric domain; both are represented by the zero level-set of neural signed distance fields. For an arbitrary point " + }, + { + "bbox": [ + 130, + 193, + 479, + 251 + ], + "type": "inline_equation", + "content": "\\mathbf{p} \\in S" + }, + { + "bbox": [ + 130, + 193, + 479, + 251 + ], + "type": "text", + "content": ", the forward deformation " + }, + { + "bbox": [ + 130, + 193, + 479, + 251 + ], + "type": "inline_equation", + "content": "F_{\\mathrm{def}}(\\mathbf{p}, \\mathbf{z}_s)" + }, + { + "bbox": [ + 130, + 193, + 479, + 251 + ], + "type": "text", + "content": " computes a displacement vector, mapping the point " + }, + { + "bbox": [ + 130, + 193, + 479, + 251 + ], + "type": "inline_equation", + "content": "\\mathbf{p}" + }, + { + "bbox": [ + 130, + 193, + 479, + 251 + ], + "type": "text", + "content": " to a corresponding point " + }, + { + "bbox": [ + 130, + 193, + 479, + 251 + ], + "type": "inline_equation", + "content": "\\mathbf{p}' \\in \\mathcal{D}" + }, + { + "bbox": [ + 130, + 193, + 479, + 251 + ], + "type": "text", + "content": " in the parametric domain as follows:" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 261, + 253, + 480, + 265 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 261, + 253, + 480, + 265 + ], + "spans": [ + { + "bbox": [ + 261, + 253, + 480, + 265 + ], + "type": "interline_equation", + "content": "\\mathbf {p} ^ {\\prime} = \\mathbf {p} + F _ {\\mathrm {d e f}} (\\mathbf {p}, \\mathbf {z} _ {s}). \\tag {5}", + "image_path": "e2a318e6a27d7e6b878090fa047947308ddfc9c5ed27193067b0fd526573ec99.jpg" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 131, + 268, + 479, + 292 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 131, + 268, + 479, + 292 + ], + "spans": [ + { + "bbox": [ + 131, + 268, + 479, + 292 + ], + "type": "text", + "content": "Conversely, we employ " + }, + { + "bbox": [ + 131, + 268, + 479, + 292 + ], + "type": "inline_equation", + "content": "F_{\\mathrm{inv - def}}" + }, + { + "bbox": [ + 131, + 268, + 479, + 292 + ], + "type": "text", + "content": " to map points inversely from the parametric domain " + }, + { + "bbox": [ + 131, + 268, + 479, + 292 + ], + "type": "inline_equation", + "content": "\\mathcal{D}" + }, + { + "bbox": [ + 131, + 268, + 479, + 292 + ], + "type": "text", + "content": " back to the original shape " + }, + { + "bbox": [ + 131, + 268, + 479, + 292 + ], + "type": "inline_equation", + "content": "S" + }, + { + "bbox": [ + 131, + 268, + 479, + 292 + ], + "type": "text", + "content": ":" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 251, + 298, + 480, + 311 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 251, + 298, + 480, + 311 + ], + "spans": [ + { + "bbox": [ + 251, + 298, + 480, + 311 + ], + "type": "interline_equation", + "content": "\\mathbf {p} ^ {\\prime \\prime} = \\mathbf {p} ^ {\\prime} + F _ {\\text {i n v - d e f}} \\left(\\mathbf {p} ^ {\\prime}, \\mathbf {z} _ {s}\\right), \\tag {6}", + "image_path": "7a75e18a270519be22e36906445ba6a5063b96dcae65da93ab9988f771550653.jpg" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 131, + 316, + 479, + 351 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 131, + 316, + 479, + 351 + ], + "spans": [ + { + "bbox": [ + 131, + 316, + 479, + 351 + ], + "type": "text", + "content": "where the point " + }, + { + "bbox": [ + 131, + 316, + 479, + 351 + ], + "type": "inline_equation", + "content": "\\mathbf{p}''" + }, + { + "bbox": [ + 131, + 316, + 479, + 351 + ], + "type": "text", + "content": " is expected to be in close proximity to " + }, + { + "bbox": [ + 131, + 316, + 479, + 351 + ], + "type": "inline_equation", + "content": "p" + }, + { + "bbox": [ + 131, + 316, + 479, + 351 + ], + "type": "text", + "content": ". To encourage the smoothness and stability of the deformations, we impose an " + }, + { + "bbox": [ + 131, + 316, + 479, + 351 + ], + "type": "inline_equation", + "content": "L_{2}" + }, + { + "bbox": [ + 131, + 316, + 479, + 351 + ], + "type": "text", + "content": " loss on the magnitude of the displacement vectors:" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 208, + 358, + 480, + 370 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 208, + 358, + 480, + 370 + ], + "spans": [ + { + "bbox": [ + 208, + 358, + 480, + 370 + ], + "type": "interline_equation", + "content": "\\mathcal {L} _ {\\text {s m o o t h}} = \\left\\| F _ {\\text {d e f}} \\left(\\mathbf {p}, \\mathbf {z} _ {s}\\right) \\right\\| _ {2} + \\left\\| F _ {\\text {i n v - d e f}} \\left(\\mathbf {p} ^ {\\prime}, \\mathbf {z} _ {s}\\right) \\right\\| _ {2}. \\tag {7}", + "image_path": "2b5523746bea13606c27e2609d50095faa2d3ac53662aeffc421a397564dee84.jpg" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 131, + 376, + 479, + 399 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 131, + 376, + 479, + 399 + ], + "spans": [ + { + "bbox": [ + 131, + 376, + 479, + 399 + ], + "type": "text", + "content": "This loss helps minimize abrupt and irregular deformations, leading to more coherent and gradual changes between " + }, + { + "bbox": [ + 131, + 376, + 479, + 399 + ], + "type": "inline_equation", + "content": "\\mathcal{S}" + }, + { + "bbox": [ + 131, + 376, + 479, + 399 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 131, + 376, + 479, + 399 + ], + "type": "inline_equation", + "content": "\\mathcal{D}" + }, + { + "bbox": [ + 131, + 376, + 479, + 399 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 130, + 399, + 479, + 436 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 399, + 479, + 436 + ], + "spans": [ + { + "bbox": [ + 130, + 399, + 479, + 436 + ], + "type": "text", + "content": "In training the networks " + }, + { + "bbox": [ + 130, + 399, + 479, + 436 + ], + "type": "inline_equation", + "content": "F_{\\mathrm{def}}" + }, + { + "bbox": [ + 130, + 399, + 479, + 436 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 130, + 399, + 479, + 436 + ], + "type": "inline_equation", + "content": "F_{\\mathrm{inv - def}}" + }, + { + "bbox": [ + 130, + 399, + 479, + 436 + ], + "type": "text", + "content": ", our objective is to achieve a bijective mapping between the 3D surface " + }, + { + "bbox": [ + 130, + 399, + 479, + 436 + ], + "type": "inline_equation", + "content": "S" + }, + { + "bbox": [ + 130, + 399, + 479, + 436 + ], + "type": "text", + "content": " and the parametric domain " + }, + { + "bbox": [ + 130, + 399, + 479, + 436 + ], + "type": "inline_equation", + "content": "\\mathcal{D}" + }, + { + "bbox": [ + 130, + 399, + 479, + 436 + ], + "type": "text", + "content": ". To achieve this goal, we adopt a cycle loss, which is similar to NeuTex [52]:" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 238, + 441, + 480, + 460 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 238, + 441, + 480, + 460 + ], + "spans": [ + { + "bbox": [ + 238, + 441, + 480, + 460 + ], + "type": "interline_equation", + "content": "\\mathcal {L} _ {\\text {c y c l e}} = \\sum_ {\\mathbf {p} \\in S} \\lambda (\\mathbf {p}) \\| \\mathbf {p} - \\mathbf {p} ^ {\\prime \\prime} \\| _ {2}, \\tag {8}", + "image_path": "c0d8ffb07ddf496d5a531d3f9f25796ee4384058480436a5348bab1f5f46e801.jpg" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 130, + 464, + 480, + 571 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 464, + 480, + 571 + ], + "spans": [ + { + "bbox": [ + 130, + 464, + 480, + 571 + ], + "type": "text", + "content": "where the weight " + }, + { + "bbox": [ + 130, + 464, + 480, + 571 + ], + "type": "inline_equation", + "content": "\\lambda" + }, + { + "bbox": [ + 130, + 464, + 480, + 571 + ], + "type": "text", + "content": " characterizes the importance of sample " + }, + { + "bbox": [ + 130, + 464, + 480, + 571 + ], + "type": "inline_equation", + "content": "\\mathbf{p}" + }, + { + "bbox": [ + 130, + 464, + 480, + 571 + ], + "type": "text", + "content": " to the loss. The reason that we do not consider all sample points equally is that a sample point may not be exactly on the zero level-set of the SDF. To tolerate such an inaccuracy, we define " + }, + { + "bbox": [ + 130, + 464, + 480, + 571 + ], + "type": "inline_equation", + "content": "\\lambda (\\mathbf{p}) = T(\\mathbf{p})(1 - \\exp (-\\sigma (s)\\delta (\\mathbf{p})))" + }, + { + "bbox": [ + 130, + 464, + 480, + 571 + ], + "type": "text", + "content": " as the color weight used in volume rendering [27]. Here, " + }, + { + "bbox": [ + 130, + 464, + 480, + 571 + ], + "type": "inline_equation", + "content": "T(\\mathbf{p})" + }, + { + "bbox": [ + 130, + 464, + 480, + 571 + ], + "type": "text", + "content": " is the transparency of sample " + }, + { + "bbox": [ + 130, + 464, + 480, + 571 + ], + "type": "inline_equation", + "content": "\\mathbf{p}" + }, + { + "bbox": [ + 130, + 464, + 480, + 571 + ], + "type": "text", + "content": " in the viewing direction, " + }, + { + "bbox": [ + 130, + 464, + 480, + 571 + ], + "type": "inline_equation", + "content": "\\delta" + }, + { + "bbox": [ + 130, + 464, + 480, + 571 + ], + "type": "text", + "content": " is the length of the sample interval along the ray, and " + }, + { + "bbox": [ + 130, + 464, + 480, + 571 + ], + "type": "inline_equation", + "content": "s" + }, + { + "bbox": [ + 130, + 464, + 480, + 571 + ], + "type": "text", + "content": " is the signed distance value of the point " + }, + { + "bbox": [ + 130, + 464, + 480, + 571 + ], + "type": "inline_equation", + "content": "\\mathbf{p}" + }, + { + "bbox": [ + 130, + 464, + 480, + 571 + ], + "type": "text", + "content": ". Clearly, when " + }, + { + "bbox": [ + 130, + 464, + 480, + 571 + ], + "type": "inline_equation", + "content": "\\mathbf{p}" + }, + { + "bbox": [ + 130, + 464, + 480, + 571 + ], + "type": "text", + "content": " is away from the surface, the coefficient " + }, + { + "bbox": [ + 130, + 464, + 480, + 571 + ], + "type": "inline_equation", + "content": "\\lambda (\\mathbf{p})" + }, + { + "bbox": [ + 130, + 464, + 480, + 571 + ], + "type": "text", + "content": " has few effects on this loss. We refer the readers to [27] for details about volume rendering and the computation of transparency " + }, + { + "bbox": [ + 130, + 464, + 480, + 571 + ], + "type": "inline_equation", + "content": "T" + }, + { + "bbox": [ + 130, + 464, + 480, + 571 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 130, + 572, + 479, + 633 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 572, + 479, + 633 + ], + "spans": [ + { + "bbox": [ + 130, + 572, + 479, + 633 + ], + "type": "text", + "content": "By explicitly minimizing the distance between " + }, + { + "bbox": [ + 130, + 572, + 479, + 633 + ], + "type": "inline_equation", + "content": "\\mathbf{p}" + }, + { + "bbox": [ + 130, + 572, + 479, + 633 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 130, + 572, + 479, + 633 + ], + "type": "inline_equation", + "content": "\\mathbf{p}''" + }, + { + "bbox": [ + 130, + 572, + 479, + 633 + ], + "type": "text", + "content": ", the cycle loss effectively prevents scenarios where two distinct points " + }, + { + "bbox": [ + 130, + 572, + 479, + 633 + ], + "type": "inline_equation", + "content": "\\mathbf{p}_1, \\mathbf{p}_2 \\in S" + }, + { + "bbox": [ + 130, + 572, + 479, + 633 + ], + "type": "text", + "content": " map to an identical point " + }, + { + "bbox": [ + 130, + 572, + 479, + 633 + ], + "type": "inline_equation", + "content": "\\mathbf{p}' \\in \\mathcal{D}" + }, + { + "bbox": [ + 130, + 572, + 479, + 633 + ], + "type": "text", + "content": ". In such cases, the inverse deformation " + }, + { + "bbox": [ + 130, + 572, + 479, + 633 + ], + "type": "inline_equation", + "content": "F_{\\mathrm{inv-def}}" + }, + { + "bbox": [ + 130, + 572, + 479, + 633 + ], + "type": "text", + "content": " would be unable to map the single point " + }, + { + "bbox": [ + 130, + 572, + 479, + 633 + ], + "type": "inline_equation", + "content": "\\mathbf{p}'" + }, + { + "bbox": [ + 130, + 572, + 479, + 633 + ], + "type": "text", + "content": " back to the respective distinct points " + }, + { + "bbox": [ + 130, + 572, + 479, + 633 + ], + "type": "inline_equation", + "content": "\\mathbf{p}_1''" + }, + { + "bbox": [ + 130, + 572, + 479, + 633 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 130, + 572, + 479, + 633 + ], + "type": "inline_equation", + "content": "\\mathbf{p}_2''" + }, + { + "bbox": [ + 130, + 572, + 479, + 633 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 131, + 647, + 285, + 658 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 131, + 647, + 285, + 658 + ], + "spans": [ + { + "bbox": [ + 131, + 647, + 285, + 658 + ], + "type": "text", + "content": "3.3 Laplacian Regularization" + } + ] + } + ], + "index": 15 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 403, + 91, + 447, + 101 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 403, + 91, + 447, + 101 + ], + "spans": [ + { + "bbox": [ + 403, + 91, + 447, + 101 + ], + "type": "text", + "content": "NeuParam" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 474, + 92, + 481, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 474, + 92, + 481, + 100 + ], + "spans": [ + { + "bbox": [ + 474, + 92, + 481, + 100 + ], + "type": "text", + "content": "7" + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 6 + }, + { + "para_blocks": [ + { + "bbox": [ + 130, + 116, + 369, + 223 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 116, + 369, + 223 + ], + "spans": [ + { + "bbox": [ + 130, + 116, + 369, + 223 + ], + "type": "text", + "content": "Mitigating angle distortion is crucial in ensuring the quality of parameterization. One approach to reduce angle distortion involves minimizing the ratio of the two singular values of the Jacobian matrix associated with the parameterization [6]. However, considering the representation of the surface " + }, + { + "bbox": [ + 130, + 116, + 369, + 223 + ], + "type": "inline_equation", + "content": "S" + }, + { + "bbox": [ + 130, + 116, + 369, + 223 + ], + "type": "text", + "content": ", the parametric domain " + }, + { + "bbox": [ + 130, + 116, + 369, + 223 + ], + "type": "inline_equation", + "content": "\\mathcal{D}" + }, + { + "bbox": [ + 130, + 116, + 369, + 223 + ], + "type": "text", + "content": ", and the parameterization " + }, + { + "bbox": [ + 130, + 116, + 369, + 223 + ], + "type": "inline_equation", + "content": "F_{\\mathrm{def}}" + }, + { + "bbox": [ + 130, + 116, + 369, + 223 + ], + "type": "text", + "content": " as neural implicit functions in our approach, directly computing the Jacobian matrix is a non-trivial task. To circum" + } + ] + } + ], + "index": 2 + }, + { + "type": "image", + "bbox": [ + 377, + 118, + 479, + 200 + ], + "blocks": [ + { + "bbox": [ + 377, + 118, + 479, + 200 + ], + "lines": [ + { + "bbox": [ + 377, + 118, + 479, + 200 + ], + "spans": [ + { + "bbox": [ + 377, + 118, + 479, + 200 + ], + "type": "image", + "image_path": "aaf6a8fe310706b9a07514d27adc58af13eb6ad0f7d8997ac3b934d57e018a0d.jpg" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_body" + } + ], + "index": 3 + }, + { + "bbox": [ + 130, + 224, + 482, + 320 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 224, + 482, + 320 + ], + "spans": [ + { + "bbox": [ + 130, + 224, + 482, + 320 + ], + "type": "text", + "content": "vent this, we employ a Laplacian regularizer in our implementation to control angle distortion. Achieving the minimizer of the Laplacian of the map results in a harmonic map. For genus-0 surfaces, harmonic maps are also conformal, which preserves angles [12]. Although for surfaces of higher genus, a harmonic map is not necessarily conformal, it still serves an effective tool for reducing angle distortion [13]. Specifically, for a sample point " + }, + { + "bbox": [ + 130, + 224, + 482, + 320 + ], + "type": "inline_equation", + "content": "\\mathbf{p} \\in S" + }, + { + "bbox": [ + 130, + 224, + 482, + 320 + ], + "type": "text", + "content": " and its neighboring points " + }, + { + "bbox": [ + 130, + 224, + 482, + 320 + ], + "type": "inline_equation", + "content": "\\mathbf{q}_j \\in S" + }, + { + "bbox": [ + 130, + 224, + 482, + 320 + ], + "type": "text", + "content": ", the forward deformation maps them to " + }, + { + "bbox": [ + 130, + 224, + 482, + 320 + ], + "type": "inline_equation", + "content": "\\mathbf{p}' \\in \\mathcal{D}" + }, + { + "bbox": [ + 130, + 224, + 482, + 320 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 130, + 224, + 482, + 320 + ], + "type": "inline_equation", + "content": "\\mathbf{q}_j' \\in \\mathcal{D}" + }, + { + "bbox": [ + 130, + 224, + 482, + 320 + ], + "type": "text", + "content": ", respectively. We compute the Laplacian of the sample " + }, + { + "bbox": [ + 130, + 224, + 482, + 320 + ], + "type": "inline_equation", + "content": "\\mathbf{p}" + }, + { + "bbox": [ + 130, + 224, + 482, + 320 + ], + "type": "text", + "content": " as follows:" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 246, + 327, + 481, + 347 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 246, + 327, + 481, + 347 + ], + "spans": [ + { + "bbox": [ + 246, + 327, + 481, + 347 + ], + "type": "interline_equation", + "content": "\\Delta (\\mathbf {p}) = \\sum_ {j = 1} ^ {m} \\omega_ {j} \\left(\\mathbf {p} ^ {\\prime} - \\mathbf {q} _ {j} ^ {\\prime}\\right). \\tag {9}", + "image_path": "db6adfb8ffa8aade4dc8ec853cf8b4e1b3cbc8230374bae708ba149bc612629b.jpg" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 132, + 354, + 481, + 367 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 354, + 481, + 367 + ], + "spans": [ + { + "bbox": [ + 132, + 354, + 481, + 367 + ], + "type": "text", + "content": "Here, the weights " + }, + { + "bbox": [ + 132, + 354, + 481, + 367 + ], + "type": "inline_equation", + "content": "w_{j}" + }, + { + "bbox": [ + 132, + 354, + 481, + 367 + ], + "type": "text", + "content": " are defined using the metric from the original surface " + }, + { + "bbox": [ + 132, + 354, + 481, + 367 + ], + "type": "inline_equation", + "content": "S" + }, + { + "bbox": [ + 132, + 354, + 481, + 367 + ], + "type": "text", + "content": " as:" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 248, + 378, + 481, + 392 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 248, + 378, + 481, + 392 + ], + "spans": [ + { + "bbox": [ + 248, + 378, + 481, + 392 + ], + "type": "interline_equation", + "content": "\\omega_ {j} = \\exp (- \\| \\mathbf {p} - \\mathbf {q} _ {j} \\| _ {2} / l), \\tag {10}", + "image_path": "4dc1a4674c5afbdaabedaaf51377818ba1403eaf7f2376233ed2b7a04e46e0bd.jpg" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 130, + 396, + 482, + 445 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 396, + 482, + 445 + ], + "spans": [ + { + "bbox": [ + 130, + 396, + 482, + 445 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 130, + 396, + 482, + 445 + ], + "type": "inline_equation", + "content": "l" + }, + { + "bbox": [ + 130, + 396, + 482, + 445 + ], + "type": "text", + "content": " is the average distance from point " + }, + { + "bbox": [ + 130, + 396, + 482, + 445 + ], + "type": "inline_equation", + "content": "\\mathbf{p}" + }, + { + "bbox": [ + 130, + 396, + 482, + 445 + ], + "type": "text", + "content": " to its neighboring points. Subsequently, the weights are normalized " + }, + { + "bbox": [ + 130, + 396, + 482, + 445 + ], + "type": "inline_equation", + "content": "\\omega_{j} = \\omega_{j} / \\sum_{k = 1}^{m}\\omega_{k}" + }, + { + "bbox": [ + 130, + 396, + 482, + 445 + ], + "type": "text", + "content": ". Notice that " + }, + { + "bbox": [ + 130, + 396, + 482, + 445 + ], + "type": "inline_equation", + "content": "\\varDelta(\\mathbf{p})\\in\\mathbb{R}^3" + }, + { + "bbox": [ + 130, + 396, + 482, + 445 + ], + "type": "text", + "content": " is a 3D vector, which may not necessarily lie on the tangent plane. Therefore, we define the normal component of the Laplacian as:" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 258, + 453, + 481, + 467 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 258, + 453, + 481, + 467 + ], + "spans": [ + { + "bbox": [ + 258, + 453, + 481, + 467 + ], + "type": "interline_equation", + "content": "\\Delta (\\mathbf {p}) _ {\\perp} = \\langle \\Delta (\\mathbf {p}), \\mathbf {n} \\rangle \\mathbf {n}, \\tag {11}", + "image_path": "0512c9a8ede760542136e678541ce2286342d96d5b79954756ad5551620cd555.jpg" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 130, + 474, + 481, + 498 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 474, + 481, + 498 + ], + "spans": [ + { + "bbox": [ + 130, + 474, + 481, + 498 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 130, + 474, + 481, + 498 + ], + "type": "inline_equation", + "content": "\\mathbf{n}" + }, + { + "bbox": [ + 130, + 474, + 481, + 498 + ], + "type": "text", + "content": " is the normal at " + }, + { + "bbox": [ + 130, + 474, + 481, + 498 + ], + "type": "inline_equation", + "content": "\\mathbf{p}'" + }, + { + "bbox": [ + 130, + 474, + 481, + 498 + ], + "type": "text", + "content": ", and " + }, + { + "bbox": [ + 130, + 474, + 481, + 498 + ], + "type": "inline_equation", + "content": "\\langle, \\rangle" + }, + { + "bbox": [ + 130, + 474, + 481, + 498 + ], + "type": "text", + "content": " represents the inner product. Summing the tangential components yields the Laplacian loss:" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 231, + 506, + 481, + 526 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 231, + 506, + 481, + 526 + ], + "spans": [ + { + "bbox": [ + 231, + 506, + 481, + 526 + ], + "type": "interline_equation", + "content": "\\mathcal {L} _ {\\mathrm {L a p}} = \\sum_ {\\mathbf {p} \\in \\mathcal {S}} \\| \\Delta (\\mathbf {p}) - \\Delta (\\mathbf {p}) _ {\\perp} \\| _ {2}. \\tag {12}", + "image_path": "2ebf30dd1b9e848ffe9bddb2c38b19b54773fc23f3ceac07134cd6ee525187d7.jpg" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 131, + 533, + 464, + 545 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 131, + 533, + 464, + 545 + ], + "spans": [ + { + "bbox": [ + 131, + 533, + 464, + 545 + ], + "type": "text", + "content": "In our implementation, we set " + }, + { + "bbox": [ + 131, + 533, + 464, + 545 + ], + "type": "inline_equation", + "content": "m = 6" + }, + { + "bbox": [ + 131, + 533, + 464, + 545 + ], + "type": "text", + "content": " neighboring points for each sample " + }, + { + "bbox": [ + 131, + 533, + 464, + 545 + ], + "type": "inline_equation", + "content": "\\mathbf{p}" + }, + { + "bbox": [ + 131, + 533, + 464, + 545 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 131, + 563, + 299, + 574 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 131, + 563, + 299, + 574 + ], + "spans": [ + { + "bbox": [ + 131, + 563, + 299, + 574 + ], + "type": "text", + "content": "3.4 Appearance Decomposition" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 130, + 581, + 482, + 666 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 581, + 482, + 666 + ], + "spans": [ + { + "bbox": [ + 130, + 581, + 482, + 666 + ], + "type": "text", + "content": "Radiance represents the directional emission of color, illustrating how the appearance of an object varies in response to changes in the observational viewpoint [27]. Drawing inspiration from previous works [25, 60] and applying the assumptions of Lambertian and grayscale shading [4], we decompose the radiance field on the parametric domain into two components: a view-independent material field and a view-dependent shading field (see Fig. 6(a)). This decomposition allows for independent editing of material and shading. More precisely, for" + } + ] + } + ], + "index": 14 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "spans": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "type": "text", + "content": "8" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 166, + 91, + 216, + 101 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 166, + 91, + 216, + 101 + ], + "spans": [ + { + "bbox": [ + 166, + 91, + 216, + 101 + ], + "type": "text", + "content": "B. Xu et al." + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 7 + }, + { + "para_blocks": [ + { + "bbox": [ + 131, + 116, + 479, + 140 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 131, + 116, + 479, + 140 + ], + "spans": [ + { + "bbox": [ + 131, + 116, + 479, + 140 + ], + "type": "text", + "content": "In a sample point " + }, + { + "bbox": [ + 131, + 116, + 479, + 140 + ], + "type": "inline_equation", + "content": "\\mathbf{p} \\in \\mathcal{S}" + }, + { + "bbox": [ + 131, + 116, + 479, + 140 + ], + "type": "text", + "content": ", we employ two sub-networks, " + }, + { + "bbox": [ + 131, + 116, + 479, + 140 + ], + "type": "inline_equation", + "content": "F_{\\mathrm{mat}}" + }, + { + "bbox": [ + 131, + 116, + 479, + 140 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 131, + 116, + 479, + 140 + ], + "type": "inline_equation", + "content": "F_{\\mathrm{shd}}" + }, + { + "bbox": [ + 131, + 116, + 479, + 140 + ], + "type": "text", + "content": ", to decompose the radiance " + }, + { + "bbox": [ + 131, + 116, + 479, + 140 + ], + "type": "inline_equation", + "content": "\\mathbf{r} \\in \\mathbb{R}^3" + }, + { + "bbox": [ + 131, + 116, + 479, + 140 + ], + "type": "text", + "content": " at point " + }, + { + "bbox": [ + 131, + 116, + 479, + 140 + ], + "type": "inline_equation", + "content": "\\mathbf{p}" + }, + { + "bbox": [ + 131, + 116, + 479, + 140 + ], + "type": "text", + "content": " as follows:" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 206, + 152, + 480, + 164 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 206, + 152, + 480, + 164 + ], + "spans": [ + { + "bbox": [ + 206, + 152, + 480, + 164 + ], + "type": "interline_equation", + "content": "\\mathbf {r} (\\mathbf {p}) = F _ {\\mathrm {m a t}} \\left(\\mathbf {p} ^ {\\prime}, \\mathbf {n}, \\mathbf {z} _ {a}\\right) \\cdot \\exp \\left(F _ {\\mathrm {s h d}} \\left(\\mathbf {p} ^ {\\prime}, \\mathbf {n}, \\mathbf {v}, \\mathbf {z} _ {a}\\right)\\right), \\tag {13}", + "image_path": "0c3e29dce7afad1a78187bc5195684b74c16db955daff7f94ac3c1af01454ecb.jpg" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 131, + 174, + 479, + 258 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 131, + 174, + 479, + 258 + ], + "spans": [ + { + "bbox": [ + 131, + 174, + 479, + 258 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 131, + 174, + 479, + 258 + ], + "type": "inline_equation", + "content": "\\mathbf{p}' \\in \\mathcal{D}" + }, + { + "bbox": [ + 131, + 174, + 479, + 258 + ], + "type": "text", + "content": " is the image of " + }, + { + "bbox": [ + 131, + 174, + 479, + 258 + ], + "type": "inline_equation", + "content": "\\mathbf{p}" + }, + { + "bbox": [ + 131, + 174, + 479, + 258 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 131, + 174, + 479, + 258 + ], + "type": "inline_equation", + "content": "\\mathbf{n}" + }, + { + "bbox": [ + 131, + 174, + 479, + 258 + ], + "type": "text", + "content": " is the unit normal of " + }, + { + "bbox": [ + 131, + 174, + 479, + 258 + ], + "type": "inline_equation", + "content": "\\mathbf{p}" + }, + { + "bbox": [ + 131, + 174, + 479, + 258 + ], + "type": "text", + "content": " on the original surface. The material network " + }, + { + "bbox": [ + 131, + 174, + 479, + 258 + ], + "type": "inline_equation", + "content": "F_{\\mathrm{mat}}" + }, + { + "bbox": [ + 131, + 174, + 479, + 258 + ], + "type": "text", + "content": " takes the mapped point " + }, + { + "bbox": [ + 131, + 174, + 479, + 258 + ], + "type": "inline_equation", + "content": "\\mathbf{p}'" + }, + { + "bbox": [ + 131, + 174, + 479, + 258 + ], + "type": "text", + "content": ", the surface normal " + }, + { + "bbox": [ + 131, + 174, + 479, + 258 + ], + "type": "inline_equation", + "content": "\\mathbf{n}" + }, + { + "bbox": [ + 131, + 174, + 479, + 258 + ], + "type": "text", + "content": ", and the appearance latent code " + }, + { + "bbox": [ + 131, + 174, + 479, + 258 + ], + "type": "inline_equation", + "content": "\\mathbf{z}_a" + }, + { + "bbox": [ + 131, + 174, + 479, + 258 + ], + "type": "text", + "content": " as input, generating the view-independent material field in response. The shading network " + }, + { + "bbox": [ + 131, + 174, + 479, + 258 + ], + "type": "inline_equation", + "content": "F_{\\mathrm{shd}}" + }, + { + "bbox": [ + 131, + 174, + 479, + 258 + ], + "type": "text", + "content": ", which is designed for computing the view-dependent shading field, takes an additional input: the viewpoint " + }, + { + "bbox": [ + 131, + 174, + 479, + 258 + ], + "type": "inline_equation", + "content": "\\mathbf{v}" + }, + { + "bbox": [ + 131, + 174, + 479, + 258 + ], + "type": "text", + "content": ". To ensure the shading sub-network " + }, + { + "bbox": [ + 131, + 174, + 479, + 258 + ], + "type": "inline_equation", + "content": "F_{\\mathrm{shd}}" + }, + { + "bbox": [ + 131, + 174, + 479, + 258 + ], + "type": "text", + "content": " concentrates on local shading features, we employ an " + }, + { + "bbox": [ + 131, + 174, + 479, + 258 + ], + "type": "inline_equation", + "content": "L_1" + }, + { + "bbox": [ + 131, + 174, + 479, + 258 + ], + "type": "text", + "content": " loss to the output of " + }, + { + "bbox": [ + 131, + 174, + 479, + 258 + ], + "type": "inline_equation", + "content": "F_{\\mathrm{shd}}" + }, + { + "bbox": [ + 131, + 174, + 479, + 258 + ], + "type": "text", + "content": ", encouraging sparsity," + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 246, + 270, + 480, + 281 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 246, + 270, + 480, + 281 + ], + "spans": [ + { + "bbox": [ + 246, + 270, + 480, + 281 + ], + "type": "interline_equation", + "content": "\\mathcal {L} _ {\\mathrm {s h d}} = \\left\\| F _ {\\mathrm {s h d}} \\left(\\mathbf {p} ^ {\\prime}, \\mathbf {n}, \\mathbf {v}, \\mathbf {z} _ {a}\\right) \\right\\| _ {1}. \\tag {14}", + "image_path": "4d4b511fdd00c13d71f01fd6e16ba8969be24de8b4b7a38cd6aa6a3e095dc226.jpg" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 131, + 293, + 479, + 340 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 131, + 293, + 479, + 340 + ], + "spans": [ + { + "bbox": [ + 131, + 293, + 479, + 340 + ], + "type": "text", + "content": "To compute the color, we need the signed distance values for the samples " + }, + { + "bbox": [ + 131, + 293, + 479, + 340 + ], + "type": "inline_equation", + "content": "\\mathbf{p}" + }, + { + "bbox": [ + 131, + 293, + 479, + 340 + ], + "type": "text", + "content": ". As mentioned above, our network adopts only one SDF network " + }, + { + "bbox": [ + 131, + 293, + 479, + 340 + ], + "type": "inline_equation", + "content": "F_{\\mathrm{sdf}}" + }, + { + "bbox": [ + 131, + 293, + 479, + 340 + ], + "type": "text", + "content": ", which represents the geometry of the parametric domain " + }, + { + "bbox": [ + 131, + 293, + 479, + 340 + ], + "type": "inline_equation", + "content": "\\mathcal{D}" + }, + { + "bbox": [ + 131, + 293, + 479, + 340 + ], + "type": "text", + "content": ". We use the forward deformation " + }, + { + "bbox": [ + 131, + 293, + 479, + 340 + ], + "type": "inline_equation", + "content": "F_{\\mathrm{def}}" + }, + { + "bbox": [ + 131, + 293, + 479, + 340 + ], + "type": "text", + "content": " to compute " + }, + { + "bbox": [ + 131, + 293, + 479, + 340 + ], + "type": "inline_equation", + "content": "s(\\mathbf{p})" + }, + { + "bbox": [ + 131, + 293, + 479, + 340 + ], + "type": "text", + "content": " as" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 244, + 352, + 480, + 364 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 244, + 352, + 480, + 364 + ], + "spans": [ + { + "bbox": [ + 244, + 352, + 480, + 364 + ], + "type": "interline_equation", + "content": "s (\\mathbf {p}) = F _ {\\mathrm {s d f}} \\left(\\mathbf {p} + F _ {\\mathrm {d e f}} (\\mathbf {p}, \\mathbf {z} _ {s})\\right). \\tag {15}", + "image_path": "f5fda1a190d0594eaadf35975772dd2fbc725fbc642e0676bee8f795dde22da8.jpg" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 131, + 374, + 479, + 411 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 131, + 374, + 479, + 411 + ], + "spans": [ + { + "bbox": [ + 131, + 374, + 479, + 411 + ], + "type": "text", + "content": "Then we employ the SDF-based volume rendering [58] to integrate the radiances along the ray to compute color " + }, + { + "bbox": [ + 131, + 374, + 479, + 411 + ], + "type": "inline_equation", + "content": "\\mathbf{c}_{\\mathrm{pred}} = \\sum_{i}\\lambda_{i}\\mathbf{r}_{i}" + }, + { + "bbox": [ + 131, + 374, + 479, + 411 + ], + "type": "text", + "content": ", where " + }, + { + "bbox": [ + 131, + 374, + 479, + 411 + ], + "type": "inline_equation", + "content": "\\mathbf{r}_i" + }, + { + "bbox": [ + 131, + 374, + 479, + 411 + ], + "type": "text", + "content": " is the radiance of a sample point " + }, + { + "bbox": [ + 131, + 374, + 479, + 411 + ], + "type": "inline_equation", + "content": "\\mathbf{p}_i" + }, + { + "bbox": [ + 131, + 374, + 479, + 411 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 131, + 374, + 479, + 411 + ], + "type": "inline_equation", + "content": "\\lambda_{i}" + }, + { + "bbox": [ + 131, + 374, + 479, + 411 + ], + "type": "text", + "content": " is the same color weight as used in Eq. (8)." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 132, + 431, + 238, + 443 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 431, + 238, + 443 + ], + "spans": [ + { + "bbox": [ + 132, + 431, + 238, + 443 + ], + "type": "text", + "content": "3.5 Training Losses" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 131, + 453, + 479, + 502 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 131, + 453, + 479, + 502 + ], + "spans": [ + { + "bbox": [ + 131, + 453, + 479, + 502 + ], + "type": "text", + "content": "In contrast to NeP [25], which necessitate prior information such as tracked mesh and UV mapping for supervision, our method is capable of reconstructing 3D geometry and computing the corresponding texture map directly from multiview images. This process is supervised by the image loss" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 259, + 513, + 480, + 525 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 259, + 513, + 480, + 525 + ], + "spans": [ + { + "bbox": [ + 259, + 513, + 480, + 525 + ], + "type": "interline_equation", + "content": "\\mathcal {L} _ {\\mathrm {r g b}} = \\left\\| \\mathbf {c} _ {\\mathrm {g t}} - \\mathbf {c} _ {\\mathrm {p r e d}} \\right\\| _ {1}, \\tag {16}", + "image_path": "f5dc1b9bbd2c5e0c7395545cbfb609a35610263b8323589c2b25c4fa19802e76.jpg" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 131, + 536, + 479, + 607 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 131, + 536, + 479, + 607 + ], + "spans": [ + { + "bbox": [ + 131, + 536, + 479, + 607 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 131, + 536, + 479, + 607 + ], + "type": "inline_equation", + "content": "\\mathbf{c}_{\\mathrm{gt}}" + }, + { + "bbox": [ + 131, + 536, + 479, + 607 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 131, + 536, + 479, + 607 + ], + "type": "inline_equation", + "content": "\\mathbf{c}_{\\mathrm{pred}}" + }, + { + "bbox": [ + 131, + 536, + 479, + 607 + ], + "type": "text", + "content": " are the ground-truth and predicted colors, respectively. To ensure the shape and appearance latent codes, " + }, + { + "bbox": [ + 131, + 536, + 479, + 607 + ], + "type": "inline_equation", + "content": "\\mathbf{z}_s" + }, + { + "bbox": [ + 131, + 536, + 479, + 607 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 131, + 536, + 479, + 607 + ], + "type": "inline_equation", + "content": "\\mathbf{z}_a" + }, + { + "bbox": [ + 131, + 536, + 479, + 607 + ], + "type": "text", + "content": ", conform to Gaussian distributions, we incorporate a regularization term, expressed as " + }, + { + "bbox": [ + 131, + 536, + 479, + 607 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_{\\mathrm{code}} = \\| \\mathbf{z}_s\\|_2 + \\| \\mathbf{z}_a\\|_2" + }, + { + "bbox": [ + 131, + 536, + 479, + 607 + ], + "type": "text", + "content": " [31]. Additionally, with geometry represented as signed distance fields, we employ the Eikonal loss [10, 59], defined as " + }, + { + "bbox": [ + 131, + 536, + 479, + 607 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_{\\mathrm{Eik}} = (\\| \\nabla s\\|_2 - 1)^2" + }, + { + "bbox": [ + 131, + 536, + 479, + 607 + ], + "type": "text", + "content": ". Putting it all together, we define the total loss function as follows:" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 140, + 619, + 480, + 631 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 140, + 619, + 480, + 631 + ], + "spans": [ + { + "bbox": [ + 140, + 619, + 480, + 631 + ], + "type": "interline_equation", + "content": "\\mathcal {L} = \\mathcal {L} _ {\\mathrm {r g b}} + \\lambda_ {1} \\mathcal {L} _ {\\text {E i k}} + \\lambda_ {2} \\mathcal {L} _ {\\text {c y c l e}} + \\lambda_ {3} \\mathcal {L} _ {\\text {s m o o t h}} + \\lambda_ {4} \\mathcal {L} _ {\\text {L a p}} + \\lambda_ {5} \\mathcal {L} _ {\\text {s h d}} + \\lambda_ {6} \\mathcal {L} _ {\\text {c o d e}}. \\tag {17}", + "image_path": "5b0c51e23a7f1898e57f7b9826b089d59ab10b21f2dc46604c7ca082967d5414.jpg" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 131, + 641, + 479, + 665 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 131, + 641, + 479, + 665 + ], + "spans": [ + { + "bbox": [ + 131, + 641, + 479, + 665 + ], + "type": "text", + "content": "In our implementation, we empirically set the coefficients as " + }, + { + "bbox": [ + 131, + 641, + 479, + 665 + ], + "type": "inline_equation", + "content": "\\lambda_1 = \\lambda_2 = \\lambda_5 = \\lambda_6 = 0.01" + }, + { + "bbox": [ + 131, + 641, + 479, + 665 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 131, + 641, + 479, + 665 + ], + "type": "inline_equation", + "content": "\\lambda_3 = \\lambda_4 = 0.001" + }, + { + "bbox": [ + 131, + 641, + 479, + 665 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 14 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 403, + 91, + 447, + 101 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 403, + 91, + 447, + 101 + ], + "spans": [ + { + "bbox": [ + 403, + 91, + 447, + 101 + ], + "type": "text", + "content": "NeuParam" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 474, + 92, + 481, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 474, + 92, + 481, + 100 + ], + "spans": [ + { + "bbox": [ + 474, + 92, + 481, + 100 + ], + "type": "text", + "content": "9" + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 8 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 166, + 113, + 446, + 350 + ], + "blocks": [ + { + "bbox": [ + 166, + 113, + 446, + 350 + ], + "lines": [ + { + "bbox": [ + 166, + 113, + 446, + 350 + ], + "spans": [ + { + "bbox": [ + 166, + 113, + 446, + 350 + ], + "type": "image", + "image_path": "5b09bf6d523c14c80d6f045f8073405bb68530e58d45c28c39ab9050e5b41531.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 130, + 358, + 482, + 414 + ], + "lines": [ + { + "bbox": [ + 130, + 358, + 482, + 414 + ], + "spans": [ + { + "bbox": [ + 130, + 358, + 482, + 414 + ], + "type": "text", + "content": "Fig. 2: Parameterization results. Human heads are co-parameterized to a sphere owing to the clear geometric resemblance. Given the diverse geometry of cars, each is parameterized to a polycube domain (bottom three rows). From top to bottom: the input image, the normal map of the reconstructed geometry " + }, + { + "bbox": [ + 130, + 358, + 482, + 414 + ], + "type": "inline_equation", + "content": "S" + }, + { + "bbox": [ + 130, + 358, + 482, + 414 + ], + "type": "text", + "content": ", the parametric domain " + }, + { + "bbox": [ + 130, + 358, + 482, + 414 + ], + "type": "inline_equation", + "content": "\\mathcal{D}" + }, + { + "bbox": [ + 130, + 358, + 482, + 414 + ], + "type": "text", + "content": ", and the reconstructed surface " + }, + { + "bbox": [ + 130, + 358, + 482, + 414 + ], + "type": "inline_equation", + "content": "S" + }, + { + "bbox": [ + 130, + 358, + 482, + 414 + ], + "type": "text", + "content": ". See supplementary material for additional results." + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_caption" + } + ], + "index": 2 + }, + { + "bbox": [ + 132, + 435, + 230, + 449 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 435, + 230, + 449 + ], + "spans": [ + { + "bbox": [ + 132, + 435, + 230, + 449 + ], + "type": "text", + "content": "4 Experiments" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 130, + 460, + 482, + 520 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 460, + 482, + 520 + ], + "spans": [ + { + "bbox": [ + 130, + 460, + 482, + 520 + ], + "type": "text", + "content": "Datasets. For our experiments, we used a customized dataset named FS-Syn, derived from the Facescape dataset [56], selecting 10 human head models and configuring 30 fixed viewpoints and lighting conditions to obtain synthesized multi-view images. We also evaluated our method on several man-made objects from the OmniObject dataset [51], and the H3DS dataset [36]." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 130, + 520, + 482, + 628 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 520, + 482, + 628 + ], + "spans": [ + { + "bbox": [ + 130, + 520, + 482, + 628 + ], + "type": "text", + "content": "Training details and metrics. All the sub-networks " + }, + { + "bbox": [ + 130, + 520, + 482, + 628 + ], + "type": "inline_equation", + "content": "F_{\\mathrm{def}}" + }, + { + "bbox": [ + 130, + 520, + 482, + 628 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 130, + 520, + 482, + 628 + ], + "type": "inline_equation", + "content": "F_{\\mathrm{inv-def}}" + }, + { + "bbox": [ + 130, + 520, + 482, + 628 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 130, + 520, + 482, + 628 + ], + "type": "inline_equation", + "content": "F_{\\mathrm{sdf}}" + }, + { + "bbox": [ + 130, + 520, + 482, + 628 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 130, + 520, + 482, + 628 + ], + "type": "inline_equation", + "content": "F_{\\mathrm{mat}}" + }, + { + "bbox": [ + 130, + 520, + 482, + 628 + ], + "type": "text", + "content": ", and " + }, + { + "bbox": [ + 130, + 520, + 482, + 628 + ], + "type": "inline_equation", + "content": "F_{\\mathrm{shd}}" + }, + { + "bbox": [ + 130, + 520, + 482, + 628 + ], + "type": "text", + "content": " are MLPs. See the supplementary material for other detailed network hyperparameters. Once the parametric domain is chosen, we fixed " + }, + { + "bbox": [ + 130, + 520, + 482, + 628 + ], + "type": "inline_equation", + "content": "F_{\\mathrm{sdf}}" + }, + { + "bbox": [ + 130, + 520, + 482, + 628 + ], + "type": "text", + "content": " and trained other sub-networks for 2,000 epochs driven by the loss function as Eq. (17). We applied the standard Marching Cubes algorithm to extract triangle meshes for the zero level-set of the source surface " + }, + { + "bbox": [ + 130, + 520, + 482, + 628 + ], + "type": "inline_equation", + "content": "S" + }, + { + "bbox": [ + 130, + 520, + 482, + 628 + ], + "type": "text", + "content": " and then used the forward map to obtain the mesh representing the parametric surface. With triangle meshes, we measured the quality of the computed parameterization by the angle distortion " + }, + { + "bbox": [ + 130, + 520, + 482, + 628 + ], + "type": "inline_equation", + "content": "\\mathcal{E}_{\\mathrm{angle}}" + }, + { + "bbox": [ + 130, + 520, + 482, + 628 + ], + "type": "text", + "content": " and area distortion [3,15] " + }, + { + "bbox": [ + 130, + 520, + 482, + 628 + ], + "type": "inline_equation", + "content": "\\mathcal{E}_{\\mathrm{area}}" + }, + { + "bbox": [ + 130, + 520, + 482, + 628 + ], + "type": "text", + "content": " as follows:" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 133, + 635, + 479, + 670 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 635, + 479, + 670 + ], + "spans": [ + { + "bbox": [ + 133, + 635, + 479, + 670 + ], + "type": "interline_equation", + "content": "\\mathcal {E} _ {\\mathrm {a n g l e}} = \\frac {\\cot \\alpha | a | ^ {2} + \\cot \\beta | b | ^ {2} + \\cot \\gamma | c | ^ {2}}{4 \\mathrm {a r e a} (\\varDelta_ {i})}, \\quad \\mathcal {E} _ {\\mathrm {a r e a}} = \\frac {1}{2} \\left(\\frac {\\mathrm {a r e a} (\\varDelta_ {i})}{\\mathrm {a r e a} (\\hat {\\varDelta} _ {i})} + \\frac {\\mathrm {a r e a} (\\hat {\\varDelta} _ {i})}{\\mathrm {a r e a} (\\varDelta_ {i})}\\right),", + "image_path": "d6de1b9296f47b52ac0ba0ac352b187641dd48e5fff0dc4d94699ca1b3c29344.jpg" + } + ] + } + ], + "index": 7 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "spans": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "type": "text", + "content": "10" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 166, + 91, + 216, + 100 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 166, + 91, + 216, + 100 + ], + "spans": [ + { + "bbox": [ + 166, + 91, + 216, + 100 + ], + "type": "text", + "content": "B. Xu et al." + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 9 + }, + { + "para_blocks": [ + { + "bbox": [ + 130, + 116, + 482, + 163 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 116, + 482, + 163 + ], + "spans": [ + { + "bbox": [ + 130, + 116, + 482, + 163 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 130, + 116, + 482, + 163 + ], + "type": "inline_equation", + "content": "a, b, c" + }, + { + "bbox": [ + 130, + 116, + 482, + 163 + ], + "type": "text", + "content": " represent the side lengths of the triangle " + }, + { + "bbox": [ + 130, + 116, + 482, + 163 + ], + "type": "inline_equation", + "content": "\\Delta_i" + }, + { + "bbox": [ + 130, + 116, + 482, + 163 + ], + "type": "text", + "content": " of the original surface. " + }, + { + "bbox": [ + 130, + 116, + 482, + 163 + ], + "type": "inline_equation", + "content": "\\alpha, \\beta, \\gamma" + }, + { + "bbox": [ + 130, + 116, + 482, + 163 + ], + "type": "text", + "content": " are angles of the corresponding triangle " + }, + { + "bbox": [ + 130, + 116, + 482, + 163 + ], + "type": "inline_equation", + "content": "\\hat{\\Delta}_i" + }, + { + "bbox": [ + 130, + 116, + 482, + 163 + ], + "type": "text", + "content": " in parametric domain. In the case of conformal mapping, " + }, + { + "bbox": [ + 130, + 116, + 482, + 163 + ], + "type": "inline_equation", + "content": "\\mathcal{E}_{\\mathrm{angle}}" + }, + { + "bbox": [ + 130, + 116, + 482, + 163 + ], + "type": "text", + "content": " is equal to 1, and " + }, + { + "bbox": [ + 130, + 116, + 482, + 163 + ], + "type": "inline_equation", + "content": "\\mathcal{E}_{\\mathrm{area}}" + }, + { + "bbox": [ + 130, + 116, + 482, + 163 + ], + "type": "text", + "content": " is also equal to 1 for an isometric map." + } + ] + } + ], + "index": 2 + }, + { + "type": "image", + "bbox": [ + 154, + 188, + 212, + 254 + ], + "blocks": [ + { + "bbox": [ + 154, + 188, + 212, + 254 + ], + "lines": [ + { + "bbox": [ + 154, + 188, + 212, + 254 + ], + "spans": [ + { + "bbox": [ + 154, + 188, + 212, + 254 + ], + "type": "image", + "image_path": "6a00a32da747038db7cde00c08403bd135ea08c15804652627fab46f77b186cc.jpg" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 159, + 255, + 208, + 266 + ], + "lines": [ + { + "bbox": [ + 159, + 255, + 208, + 266 + ], + "spans": [ + { + "bbox": [ + 159, + 255, + 208, + 266 + ], + "type": "text", + "content": "1.545/1.396" + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "image_caption" + } + ], + "index": 3 + }, + { + "type": "image", + "bbox": [ + 215, + 188, + 274, + 254 + ], + "blocks": [ + { + "bbox": [ + 215, + 188, + 274, + 254 + ], + "lines": [ + { + "bbox": [ + 215, + 188, + 274, + 254 + ], + "spans": [ + { + "bbox": [ + 215, + 188, + 274, + 254 + ], + "type": "image", + "image_path": "750fc2cbafbd9ed1e04663849ce79353cd61ae5efbcd23d35bb76c168343d6ab.jpg" + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 221, + 256, + 269, + 266 + ], + "lines": [ + { + "bbox": [ + 221, + 256, + 269, + 266 + ], + "spans": [ + { + "bbox": [ + 221, + 256, + 269, + 266 + ], + "type": "text", + "content": "1.755/1.446" + } + ] + } + ], + "index": 6, + "angle": 0, + "type": "image_caption" + } + ], + "index": 5 + }, + { + "type": "image", + "bbox": [ + 276, + 189, + 335, + 254 + ], + "blocks": [ + { + "bbox": [ + 276, + 189, + 335, + 254 + ], + "lines": [ + { + "bbox": [ + 276, + 189, + 335, + 254 + ], + "spans": [ + { + "bbox": [ + 276, + 189, + 335, + 254 + ], + "type": "image", + "image_path": "7681f59da7455f21547f6e81443fd0eb5233e083fb0bb35df31878d6bcd3aeb8.jpg" + } + ] + } + ], + "index": 7, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 282, + 255, + 331, + 266 + ], + "lines": [ + { + "bbox": [ + 282, + 255, + 331, + 266 + ], + "spans": [ + { + "bbox": [ + 282, + 255, + 331, + 266 + ], + "type": "text", + "content": "1.394/1.302" + } + ] + } + ], + "index": 8, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 130, + 275, + 482, + 342 + ], + "lines": [ + { + "bbox": [ + 130, + 275, + 482, + 342 + ], + "spans": [ + { + "bbox": [ + 130, + 275, + 482, + 342 + ], + "type": "text", + "content": "Fig. 3: Results of cross-polycube parameterization on the H3DS dataset [36]. We learned a common polycube domain for all human heads with the parameter " + }, + { + "bbox": [ + 130, + 275, + 482, + 342 + ], + "type": "inline_equation", + "content": "k = 3" + }, + { + "bbox": [ + 130, + 275, + 482, + 342 + ], + "type": "text", + "content": " (i.e., the polycube is formed by 3 boxes) and then computed the polycube parameterization for all the human heads simultaneously. Since human heads share similar geometries, the computed polycube parameterizations are also highly consistent. The values below each figure represent the angle distortion and the area distortion, respectively." + } + ] + } + ], + "index": 13, + "angle": 0, + "type": "image_caption" + } + ], + "index": 7 + }, + { + "type": "image", + "bbox": [ + 339, + 190, + 398, + 254 + ], + "blocks": [ + { + "bbox": [ + 339, + 190, + 398, + 254 + ], + "lines": [ + { + "bbox": [ + 339, + 190, + 398, + 254 + ], + "spans": [ + { + "bbox": [ + 339, + 190, + 398, + 254 + ], + "type": "image", + "image_path": "28bb4180d730adcae61ab5233b9287a83231d102fdcaf4bbb40f3bdcebd5e06b.jpg" + } + ] + } + ], + "index": 9, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 345, + 255, + 393, + 266 + ], + "lines": [ + { + "bbox": [ + 345, + 255, + 393, + 266 + ], + "spans": [ + { + "bbox": [ + 345, + 255, + 393, + 266 + ], + "type": "text", + "content": "1.271/1.216" + } + ] + } + ], + "index": 10, + "angle": 0, + "type": "image_caption" + } + ], + "index": 9 + }, + { + "type": "image", + "bbox": [ + 400, + 189, + 460, + 254 + ], + "blocks": [ + { + "bbox": [ + 400, + 189, + 460, + 254 + ], + "lines": [ + { + "bbox": [ + 400, + 189, + 460, + 254 + ], + "spans": [ + { + "bbox": [ + 400, + 189, + 460, + 254 + ], + "type": "image", + "image_path": "6a0347dd2df0b4c34042a16e62a8d674254377381dd8afdf0a7ee0a3359085cc.jpg" + } + ] + } + ], + "index": 11, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 406, + 255, + 455, + 266 + ], + "lines": [ + { + "bbox": [ + 406, + 255, + 455, + 266 + ], + "spans": [ + { + "bbox": [ + 406, + 255, + 455, + 266 + ], + "type": "text", + "content": "1.351/1.275" + } + ] + } + ], + "index": 12, + "angle": 0, + "type": "image_caption" + } + ], + "index": 11 + }, + { + "bbox": [ + 130, + 367, + 482, + 452 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 367, + 482, + 452 + ], + "spans": [ + { + "bbox": [ + 130, + 367, + 482, + 452 + ], + "type": "text", + "content": "Results. Our framework constructs a nearly bijective map between objects and multiple parametric domains such as spheres and polycubes, as shown in Fig. 2 and Fig. 3. It supports co-parameterization of multiple objects and facilitates material and shading transfer. Fully compatible with existing neural rendering pipelines, our neural parameterization algorithm uses multi-view images to simultaneously reconstruct 3D geometry and create texture maps. As shown in Fig. 2, we achieve high-fidelity results without prior information." + } + ] + } + ], + "index": 14 + }, + { + "type": "image", + "bbox": [ + 133, + 473, + 197, + 538 + ], + "blocks": [ + { + "bbox": [ + 133, + 473, + 197, + 538 + ], + "lines": [ + { + "bbox": [ + 133, + 473, + 197, + 538 + ], + "spans": [ + { + "bbox": [ + 133, + 473, + 197, + 538 + ], + "type": "image", + "image_path": "88c65bc948910005a4ba6f345e0bf59c409d1d687eab48214fb3b3446b2ba4e6.jpg" + } + ] + } + ], + "index": 15, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 145, + 540, + 183, + 551 + ], + "lines": [ + { + "bbox": [ + 145, + 540, + 183, + 551 + ], + "spans": [ + { + "bbox": [ + 145, + 540, + 183, + 551 + ], + "type": "text", + "content": "UV prior" + } + ] + } + ], + "index": 16, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 130, + 559, + 482, + 604 + ], + "lines": [ + { + "bbox": [ + 130, + 559, + 482, + 604 + ], + "spans": [ + { + "bbox": [ + 130, + 559, + 482, + 604 + ], + "type": "text", + "content": "Fig. 4: Parametrization results of NeP [25] in comparison with our method. NeP heavily relies on the initial UV prior, leading to poor texture maps when such prior information is lacking. Our method learns texture maps from a set of multi-view images without any 3D tracked mesh or UV prior as input." + } + ] + } + ], + "index": 24, + "angle": 0, + "type": "image_caption" + } + ], + "index": 15 + }, + { + "type": "image", + "bbox": [ + 206, + 473, + 269, + 538 + ], + "blocks": [ + { + "bbox": [ + 206, + 473, + 269, + 538 + ], + "lines": [ + { + "bbox": [ + 206, + 473, + 269, + 538 + ], + "spans": [ + { + "bbox": [ + 206, + 473, + 269, + 538 + ], + "type": "image", + "image_path": "f07aa17abbfb2eceb87234b69ea04416b4dd9b4d51a3200fc398044d43d2a2e7.jpg" + } + ] + } + ], + "index": 17, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 212, + 540, + 272, + 551 + ], + "lines": [ + { + "bbox": [ + 212, + 540, + 272, + 551 + ], + "spans": [ + { + "bbox": [ + 212, + 540, + 272, + 551 + ], + "type": "text", + "content": "NeP w/o prior" + } + ] + } + ], + "index": 18, + "angle": 0, + "type": "image_caption" + } + ], + "index": 17 + }, + { + "type": "image", + "bbox": [ + 279, + 473, + 342, + 538 + ], + "blocks": [ + { + "bbox": [ + 279, + 473, + 342, + 538 + ], + "lines": [ + { + "bbox": [ + 279, + 473, + 342, + 538 + ], + "spans": [ + { + "bbox": [ + 279, + 473, + 342, + 538 + ], + "type": "image", + "image_path": "d3309025c2dccb88f6336bda646ae931e72565ea5f444e48adc1669865705e22.jpg" + } + ] + } + ], + "index": 19, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 303, + 540, + 339, + 551 + ], + "lines": [ + { + "bbox": [ + 303, + 540, + 339, + 551 + ], + "spans": [ + { + "bbox": [ + 303, + 540, + 339, + 551 + ], + "type": "text", + "content": "w/ prior" + } + ] + } + ], + "index": 20, + "angle": 0, + "type": "image_caption" + } + ], + "index": 19 + }, + { + "type": "image", + "bbox": [ + 353, + 475, + 401, + 538 + ], + "blocks": [ + { + "bbox": [ + 353, + 475, + 401, + 538 + ], + "lines": [ + { + "bbox": [ + 353, + 475, + 401, + 538 + ], + "spans": [ + { + "bbox": [ + 353, + 475, + 401, + 538 + ], + "type": "image", + "image_path": "a8af08f8d760abec17f6d44470ba8b68364aec39d1053fe19ab60b085d341025.jpg" + } + ] + } + ], + "index": 21, + "angle": 0, + "type": "image_body" + } + ], + "index": 21 + }, + { + "type": "image", + "bbox": [ + 408, + 483, + 481, + 538 + ], + "blocks": [ + { + "bbox": [ + 408, + 483, + 481, + 538 + ], + "lines": [ + { + "bbox": [ + 408, + 483, + 481, + 538 + ], + "spans": [ + { + "bbox": [ + 408, + 483, + 481, + 538 + ], + "type": "image", + "image_path": "2d8095154c1dde858c8ae7f093ffbc348f641faac0abcd6cf0d5f2729f4b33e2.jpg" + } + ] + } + ], + "index": 22, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 407, + 540, + 428, + 550 + ], + "lines": [ + { + "bbox": [ + 407, + 540, + 428, + 550 + ], + "spans": [ + { + "bbox": [ + 407, + 540, + 428, + 550 + ], + "type": "text", + "content": "Ours" + } + ] + } + ], + "index": 23, + "angle": 0, + "type": "image_caption" + } + ], + "index": 22 + }, + { + "bbox": [ + 130, + 629, + 482, + 666 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 629, + 482, + 666 + ], + "spans": [ + { + "bbox": [ + 130, + 629, + 482, + 666 + ], + "type": "text", + "content": "Compared with NeP [25], which is the latest neural surface editing work using neural parameterization, our method eliminates the need for prior information from UV mapping. Unlike our algorithm, NeP learns a map between" + } + ] + } + ], + "index": 25 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 402, + 91, + 447, + 101 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 402, + 91, + 447, + 101 + ], + "spans": [ + { + "bbox": [ + 402, + 91, + 447, + 101 + ], + "type": "text", + "content": "NeuParam" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 470, + 91, + 480, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 470, + 91, + 480, + 100 + ], + "spans": [ + { + "bbox": [ + 470, + 91, + 480, + 100 + ], + "type": "text", + "content": "11" + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 10 + }, + { + "para_blocks": [ + { + "bbox": [ + 130, + 116, + 482, + 164 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 116, + 482, + 164 + ], + "spans": [ + { + "bbox": [ + 130, + 116, + 482, + 164 + ], + "type": "text", + "content": "surfaces and the UV plane. However, it heavily relies on an existing UV mapping as initialization, and yields rather poor texture mapping results without the UV mapping as prior (see Fig.4). Moreover, the results from NeP still exhibit significant distortions even with UV prior." + } + ] + } + ], + "index": 2 + }, + { + "type": "image", + "bbox": [ + 142, + 185, + 471, + 287 + ], + "blocks": [ + { + "bbox": [ + 142, + 185, + 471, + 287 + ], + "lines": [ + { + "bbox": [ + 142, + 185, + 471, + 287 + ], + "spans": [ + { + "bbox": [ + 142, + 185, + 471, + 287 + ], + "type": "image", + "image_path": "f61ee14219f8b893bfb539e15bef8f0cd09e3369cd3907e8b11a2abf140e7f5a.jpg" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 130, + 297, + 482, + 331 + ], + "lines": [ + { + "bbox": [ + 130, + 297, + 482, + 331 + ], + "spans": [ + { + "bbox": [ + 130, + 297, + 482, + 331 + ], + "type": "text", + "content": "Fig. 5: Representing the materials as textures defined in a simple parametric domain enables easy editing. Also, our co-parameterization naturally supports texture transfer between different objects." + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "image_caption" + } + ], + "index": 3 + }, + { + "bbox": [ + 130, + 354, + 482, + 533 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 354, + 482, + 533 + ], + "spans": [ + { + "bbox": [ + 130, + 354, + 482, + 533 + ], + "type": "text", + "content": "Texture editing. Our parameterization results enable flexible and pixel-level editing as shown in Fig. 5. We re-paint the extracted 2D texture map using cubemaps [9] and retrieve the modified color using the established mapping when volume rendering novel images after each editing. Both NeuMesh [54] and NeP [25] can edit the object's texture at a pixel level. We compare the editing and rendering results with NeuMesh and NeP, as shown in Fig. 5. NeuMesh requires a mask of the editing region. It also re-trains the network after each edit, which is time-consuming and cannot guarantee the stability of the training results. Our method outperforms NeuMesh in terms of the visual quality. NeP, which is also neural parameterization-based method, requires a UV prior for initialization. It often yields poor parameterization results without such prior. Even with the UV prior, its parameterization results exhibit large distortion, making the edit in the 2D domain challenging. Our method is free of prior, and can yield parameterization with low distortion, thereby facilitating pixel-level editing and material transferring without network re-training." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 130, + 534, + 482, + 617 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 534, + 482, + 617 + ], + "spans": [ + { + "bbox": [ + 130, + 534, + 482, + 617 + ], + "type": "text", + "content": "Shading editing. We decompose the 3D radiance field into view-independent material and view-dependent shading defined in the parametric domain. Taking advantage of this decomposition and the embedding into a simple domain, we can edit both the material and shading of the 3D surfaces in an intuitive manner. As illustrated in Fig. 6(b), we modify the shading of one model by transferring the shading from another model under different lighting conditions, while keeping the material unchanged." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 130, + 617, + 482, + 666 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 617, + 482, + 666 + ], + "spans": [ + { + "bbox": [ + 130, + 617, + 482, + 666 + ], + "type": "text", + "content": "Inverse deformation. To validate the cycle loss and illustrate inverse deformation, we sample points on the polycube parametric surface " + }, + { + "bbox": [ + 130, + 617, + 482, + 666 + ], + "type": "inline_equation", + "content": "\\mathcal{D}" + }, + { + "bbox": [ + 130, + 617, + 482, + 666 + ], + "type": "text", + "content": ". Subsequently, we inversely deform these points back to the surface " + }, + { + "bbox": [ + 130, + 617, + 482, + 666 + ], + "type": "inline_equation", + "content": "\\mathcal{S}" + }, + { + "bbox": [ + 130, + 617, + 482, + 666 + ], + "type": "text", + "content": ". The point cloud is visualized in the right inset, and we quantify the CD " + }, + { + "bbox": [ + 130, + 617, + 482, + 666 + ], + "type": "inline_equation", + "content": "(\\times 10^{-3})" + }, + { + "bbox": [ + 130, + 617, + 482, + 666 + ], + "type": "text", + "content": " with respect to" + } + ] + } + ], + "index": 7 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "spans": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "type": "text", + "content": "12" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 166, + 91, + 216, + 100 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 166, + 91, + 216, + 100 + ], + "spans": [ + { + "bbox": [ + 166, + 91, + 216, + 100 + ], + "type": "text", + "content": "B. Xu et al." + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 11 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 151, + 112, + 465, + 175 + ], + "blocks": [ + { + "bbox": [ + 151, + 112, + 465, + 175 + ], + "lines": [ + { + "bbox": [ + 151, + 112, + 465, + 175 + ], + "spans": [ + { + "bbox": [ + 151, + 112, + 465, + 175 + ], + "type": "image", + "image_path": "b334bf3e51c1828f978a9f690c8dfdf1028f4f0381218519f94e8c5e0e0bfd5f.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 136, + 175, + 478, + 195 + ], + "lines": [ + { + "bbox": [ + 136, + 175, + 478, + 195 + ], + "spans": [ + { + "bbox": [ + 136, + 175, + 478, + 195 + ], + "type": "text", + "content": "(a) We decompose the radiance into view-dependent shadings and view-independent materials. The shading map is visualized from the front view." + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_caption" + } + ], + "index": 2 + }, + { + "type": "image", + "bbox": [ + 142, + 196, + 471, + 304 + ], + "blocks": [ + { + "bbox": [ + 142, + 196, + 471, + 304 + ], + "lines": [ + { + "bbox": [ + 142, + 196, + 471, + 304 + ], + "spans": [ + { + "bbox": [ + 142, + 196, + 471, + 304 + ], + "type": "image", + "image_path": "a9133be996c0193c69560d7e14614ce69f391e4cca377547fa373c443f3b6b27.jpg" + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 141, + 304, + 473, + 327 + ], + "lines": [ + { + "bbox": [ + 141, + 304, + 473, + 327 + ], + "spans": [ + { + "bbox": [ + 141, + 304, + 473, + 327 + ], + "type": "text", + "content": "(b) We transfer the decomposed shading from one model to another. See also the accompanying video." + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 131, + 335, + 482, + 360 + ], + "lines": [ + { + "bbox": [ + 131, + 335, + 482, + 360 + ], + "spans": [ + { + "bbox": [ + 131, + 335, + 482, + 360 + ], + "type": "text", + "content": "Fig. 6: We transfer the decomposed shading from one model to another. For further illustration, please refer to the accompanying video." + } + ] + } + ], + "index": 6, + "angle": 0, + "type": "image_caption" + } + ], + "index": 4 + }, + { + "bbox": [ + 130, + 381, + 482, + 403 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 381, + 482, + 403 + ], + "spans": [ + { + "bbox": [ + 130, + 381, + 482, + 403 + ], + "type": "text", + "content": "the ground truth. The obtained low CD results confirm the efficacy of inverse deformation and cycle loss." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 130, + 406, + 482, + 489 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 406, + 482, + 489 + ], + "spans": [ + { + "bbox": [ + 130, + 406, + 482, + 489 + ], + "type": "text", + "content": "Ablation studies. As shown in Fig. 7 (left), the term " + }, + { + "bbox": [ + 130, + 406, + 482, + 489 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_{\\mathrm{smooth}}" + }, + { + "bbox": [ + 130, + 406, + 482, + 489 + ], + "type": "text", + "content": " helps to smooth the process of the two deformations, while " + }, + { + "bbox": [ + 130, + 406, + 482, + 489 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_{\\mathrm{Lap}}" + }, + { + "bbox": [ + 130, + 406, + 482, + 489 + ], + "type": "text", + "content": " further reduces the angle distortion. Our method can effectively reduce angle distortion by minimizing Laplacian loss and reduce area distortion by choosing an appropriate parametric domain. As demonstrated in Fig. 7 (right), we calculate the angle distortion " + }, + { + "bbox": [ + 130, + 406, + 482, + 489 + ], + "type": "inline_equation", + "content": "\\mathcal{E}_{\\mathrm{angle}}" + }, + { + "bbox": [ + 130, + 406, + 482, + 489 + ], + "type": "text", + "content": " of parameterization as described in [3, 15]. We can observe that the Laplacian loss can effectively reduce angle distortion." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 130, + 490, + 482, + 562 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 490, + 482, + 562 + ], + "spans": [ + { + "bbox": [ + 130, + 490, + 482, + 562 + ], + "type": "text", + "content": "The choice of parametric domain. The different parametric domain influences the quality of parameterization as shown in Fig. 8 (a), which is measured by area distortions. Generally, increasing the number of cubic primitives for better resembling the geometry of the target surface effectively reduces distortions but increases the complexity of the parametric domain. Finding an optimal balance between domain complexity and parameterization quality is challenging." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 132, + 592, + 220, + 605 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 592, + 220, + 605 + ], + "spans": [ + { + "bbox": [ + 132, + 592, + 220, + 605 + ], + "type": "text", + "content": "5 Conclusion" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 130, + 617, + 482, + 666 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 617, + 482, + 666 + ], + "spans": [ + { + "bbox": [ + 130, + 617, + 482, + 666 + ], + "type": "text", + "content": "We introduce a novel neural algorithm for parameterizing 3D surfaces represented by neural implicit functions to a user-defined parametric domain, such as spheres and polycubes. This parametric domain is learned given some user-specified hyperparameters. Utilizing bi-directional deformation, our method is" + } + ] + } + ], + "index": 11 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 403, + 91, + 447, + 100 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 403, + 91, + 447, + 100 + ], + "spans": [ + { + "bbox": [ + 403, + 91, + 447, + 100 + ], + "type": "text", + "content": "NeuParam" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 470, + 91, + 481, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 470, + 91, + 481, + 100 + ], + "spans": [ + { + "bbox": [ + 470, + 91, + 481, + 100 + ], + "type": "text", + "content": "13" + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 12 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 142, + 114, + 184, + 160 + ], + "blocks": [ + { + "bbox": [ + 142, + 114, + 184, + 160 + ], + "lines": [ + { + "bbox": [ + 142, + 114, + 184, + 160 + ], + "spans": [ + { + "bbox": [ + 142, + 114, + 184, + 160 + ], + "type": "image", + "image_path": "9a8a79ca679bf74a26ce4837890ada73372ce0d612bd80b450abff0695bf7d4e.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_body" + } + ], + "index": 2 + }, + { + "type": "image", + "bbox": [ + 190, + 116, + 230, + 159 + ], + "blocks": [ + { + "bbox": [ + 190, + 116, + 230, + 159 + ], + "lines": [ + { + "bbox": [ + 190, + 116, + 230, + 159 + ], + "spans": [ + { + "bbox": [ + 190, + 116, + 230, + 159 + ], + "type": "image", + "image_path": "9350fc57010a9ffe32c19e532f83c38ddc9769cbe4c421a1c822f1b291a6c606.jpg" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 191, + 166, + 228, + 172 + ], + "lines": [ + { + "bbox": [ + 191, + 166, + 228, + 172 + ], + "spans": [ + { + "bbox": [ + 191, + 166, + 228, + 172 + ], + "type": "text", + "content": "VolSDF 4.65" + } + ] + } + ], + "index": 7, + "angle": 0, + "type": "image_caption" + } + ], + "index": 3 + }, + { + "type": "image", + "bbox": [ + 238, + 116, + 277, + 159 + ], + "blocks": [ + { + "bbox": [ + 238, + 116, + 277, + 159 + ], + "lines": [ + { + "bbox": [ + 238, + 116, + 277, + 159 + ], + "spans": [ + { + "bbox": [ + 238, + 116, + 277, + 159 + ], + "type": "image", + "image_path": "a39ec8569b33187b8e52116a3b2a0d67eb05cd497e7d06eb6a2aa0e8a0270501.jpg" + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 242, + 166, + 272, + 172 + ], + "lines": [ + { + "bbox": [ + 242, + 166, + 272, + 172 + ], + "spans": [ + { + "bbox": [ + 242, + 166, + 272, + 172 + ], + "type": "text", + "content": "Ours 4.87" + } + ] + } + ], + "index": 8, + "angle": 0, + "type": "image_caption" + } + ], + "index": 4 + }, + { + "type": "image", + "bbox": [ + 284, + 118, + 324, + 159 + ], + "blocks": [ + { + "bbox": [ + 284, + 118, + 324, + 159 + ], + "lines": [ + { + "bbox": [ + 284, + 118, + 324, + 159 + ], + "spans": [ + { + "bbox": [ + 284, + 118, + 324, + 159 + ], + "type": "image", + "image_path": "3c31f91daa790caed9c7085953902d81c1300895a374af3a9709e3a744093d2d.jpg" + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 285, + 166, + 321, + 172 + ], + "lines": [ + { + "bbox": [ + 285, + 166, + 321, + 172 + ], + "spans": [ + { + "bbox": [ + 285, + 166, + 321, + 172 + ], + "type": "text", + "content": "Inverse 4.89" + } + ] + } + ], + "index": 9, + "angle": 0, + "type": "image_caption" + } + ], + "index": 5 + }, + { + "type": "image", + "bbox": [ + 140, + 173, + 186, + 206 + ], + "blocks": [ + { + "bbox": [ + 156, + 166, + 167, + 172 + ], + "lines": [ + { + "bbox": [ + 156, + 166, + 167, + 172 + ], + "spans": [ + { + "bbox": [ + 156, + 166, + 167, + 172 + ], + "type": "text", + "content": "GT" + } + ] + } + ], + "index": 6, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 140, + 173, + 186, + 206 + ], + "lines": [ + { + "bbox": [ + 140, + 173, + 186, + 206 + ], + "spans": [ + { + "bbox": [ + 140, + 173, + 186, + 206 + ], + "type": "image", + "image_path": "e9ea5a3ac843f9e224dd2479efdfaa287aba6c31cff943a2d987f65ca045e5e1.jpg" + } + ] + } + ], + "index": 10, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 153, + 210, + 173, + 219 + ], + "lines": [ + { + "bbox": [ + 153, + 210, + 173, + 219 + ], + "spans": [ + { + "bbox": [ + 153, + 210, + 173, + 219 + ], + "type": "text", + "content": "Input" + } + ] + } + ], + "index": 11, + "angle": 0, + "type": "image_caption" + } + ], + "index": 10 + }, + { + "type": "image", + "bbox": [ + 190, + 179, + 233, + 206 + ], + "blocks": [ + { + "bbox": [ + 190, + 179, + 233, + 206 + ], + "lines": [ + { + "bbox": [ + 190, + 179, + 233, + 206 + ], + "spans": [ + { + "bbox": [ + 190, + 179, + 233, + 206 + ], + "type": "image", + "image_path": "fb117f9fd62ebcfb351490d013e8ecdb45130a87494e121accb2a93fc82dbb93.jpg" + } + ] + } + ], + "index": 12, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 181, + 210, + 280, + 220 + ], + "lines": [ + { + "bbox": [ + 181, + 210, + 280, + 220 + ], + "spans": [ + { + "bbox": [ + 181, + 210, + 280, + 220 + ], + "type": "text", + "content": "w/o " + }, + { + "bbox": [ + 181, + 210, + 280, + 220 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_{\\mathrm{smooth}}\\& \\mathcal{L}_{\\mathrm{Lap}} + \\mathcal{L}_{\\mathrm{smooth}}" + } + ] + } + ], + "index": 13, + "angle": 0, + "type": "image_caption" + } + ], + "index": 12 + }, + { + "type": "image", + "bbox": [ + 244, + 183, + 280, + 207 + ], + "blocks": [ + { + "bbox": [ + 244, + 183, + 280, + 207 + ], + "lines": [ + { + "bbox": [ + 244, + 183, + 280, + 207 + ], + "spans": [ + { + "bbox": [ + 244, + 183, + 280, + 207 + ], + "type": "image", + "image_path": "f7aa11f928aa7ae118782a7d42896ee4386646e259d227c73604894ec405ac82.jpg" + } + ] + } + ], + "index": 14, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 131, + 228, + 481, + 274 + ], + "lines": [ + { + "bbox": [ + 131, + 228, + 481, + 274 + ], + "spans": [ + { + "bbox": [ + 131, + 228, + 481, + 274 + ], + "type": "text", + "content": "Fig. 7: Ablation studies. On the left, we calculate the CD " + }, + { + "bbox": [ + 131, + 228, + 481, + 274 + ], + "type": "inline_equation", + "content": "(\\times 10^{-3})" + }, + { + "bbox": [ + 131, + 228, + 481, + 274 + ], + "type": "text", + "content": " based on the point cloud extracted from the inverse deformation module and present the ablation studies for " + }, + { + "bbox": [ + 131, + 228, + 481, + 274 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_{\\mathrm{smooth}}" + }, + { + "bbox": [ + 131, + 228, + 481, + 274 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 131, + 228, + 481, + 274 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_{\\mathrm{Lap}}" + }, + { + "bbox": [ + 131, + 228, + 481, + 274 + ], + "type": "text", + "content": ". On the right, utilization of the Laplacian loss has proven effective in reducing angle distortion " + }, + { + "bbox": [ + 131, + 228, + 481, + 274 + ], + "type": "inline_equation", + "content": "(\\mathcal{E}_{\\mathrm{angle}})" + }, + { + "bbox": [ + 131, + 228, + 481, + 274 + ], + "type": "text", + "content": ", especially in regions with high curvature." + } + ] + } + ], + "index": 22, + "angle": 0, + "type": "image_caption" + } + ], + "index": 14 + }, + { + "type": "image", + "bbox": [ + 297, + 183, + 326, + 207 + ], + "blocks": [ + { + "bbox": [ + 297, + 183, + 326, + 207 + ], + "lines": [ + { + "bbox": [ + 297, + 183, + 326, + 207 + ], + "spans": [ + { + "bbox": [ + 297, + 183, + 326, + 207 + ], + "type": "image", + "image_path": "1ca98573884c0f11a5df2f798dd8cdfdd53001a0186d47812f92c83625035d5a.jpg" + } + ] + } + ], + "index": 15, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 298, + 210, + 321, + 220 + ], + "lines": [ + { + "bbox": [ + 298, + 210, + 321, + 220 + ], + "spans": [ + { + "bbox": [ + 298, + 210, + 321, + 220 + ], + "type": "text", + "content": "+Lap" + } + ] + } + ], + "index": 16, + "angle": 0, + "type": "image_caption" + } + ], + "index": 15 + }, + { + "type": "image", + "bbox": [ + 334, + 118, + 462, + 195 + ], + "blocks": [ + { + "bbox": [ + 334, + 118, + 462, + 195 + ], + "lines": [ + { + "bbox": [ + 334, + 118, + 462, + 195 + ], + "spans": [ + { + "bbox": [ + 334, + 118, + 462, + 195 + ], + "type": "image", + "image_path": "4edd16536b809a1e699e0cfef5e3ff8d76596289f56217a5ac6f23bee6d0441e.jpg" + } + ] + } + ], + "index": 17, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 336, + 198, + 387, + 208 + ], + "lines": [ + { + "bbox": [ + 336, + 198, + 387, + 208 + ], + "spans": [ + { + "bbox": [ + 336, + 198, + 387, + 208 + ], + "type": "inline_equation", + "content": "\\mathcal{E}_{\\mathrm{angle}} = 1.147" + } + ] + } + ], + "index": 18, + "angle": 0, + "type": "image_footnote" + }, + { + "bbox": [ + 404, + 198, + 457, + 208 + ], + "lines": [ + { + "bbox": [ + 404, + 198, + 457, + 208 + ], + "spans": [ + { + "bbox": [ + 404, + 198, + 457, + 208 + ], + "type": "inline_equation", + "content": "\\mathcal{E}_{\\mathrm{angle}} = 1.092" + } + ] + } + ], + "index": 19, + "angle": 0, + "type": "image_footnote" + }, + { + "bbox": [ + 343, + 209, + 381, + 220 + ], + "lines": [ + { + "bbox": [ + 343, + 209, + 381, + 220 + ], + "spans": [ + { + "bbox": [ + 343, + 209, + 381, + 220 + ], + "type": "text", + "content": "w/o " + }, + { + "bbox": [ + 343, + 209, + 381, + 220 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_{\\mathrm{Lap}}" + } + ] + } + ], + "index": 20, + "angle": 0, + "type": "image_footnote" + }, + { + "bbox": [ + 413, + 209, + 447, + 220 + ], + "lines": [ + { + "bbox": [ + 413, + 209, + 447, + 220 + ], + "spans": [ + { + "bbox": [ + 413, + 209, + 447, + 220 + ], + "type": "text", + "content": "W/Lap" + } + ] + } + ], + "index": 21, + "angle": 0, + "type": "image_footnote" + } + ], + "index": 17 + }, + { + "bbox": [ + 130, + 297, + 480, + 380 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 297, + 480, + 380 + ], + "spans": [ + { + "bbox": [ + 130, + 297, + 480, + 380 + ], + "type": "text", + "content": "capable of learning a nearly bijective mapping without relying on any prior knowledge, while controlling angle distortion through the usage of a Laplacian regularizer. Furthermore, our approach seamlessly integrates with the neural rendering pipeline, enabling the reconstruction of 3D objects from multi-view images and facilitating efficient volume rendering of modified textures and shadings — eliminating the necessity for network re-training. We demonstrated the efficacy of our method on human heads and man-made objects." + } + ] + } + ], + "index": 23 + }, + { + "bbox": [ + 130, + 380, + 482, + 536 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 380, + 482, + 536 + ], + "spans": [ + { + "bbox": [ + 130, + 380, + 482, + 536 + ], + "type": "text", + "content": "Our method has a few limitations that require further improvement in the future. First, quantitative comparisons on the FS-Syn dataset show that the reconstruction quality of our method is slightly worse than the standard SDF-based neural rendering algorithms, such as VolSDF [58] and NeuS [46]. For example, as illustrated in Fig. 8 (b) and, We report PSNR and CD compared with VolSDF [58] on Fs-Syn dataset [56]. This decline in visual quality arises from the incorporation of additional modules, such as bi-directional deformation and radiance decomposition, resulting in the cessation of updates to the SDF network within our neural parameterization pipeline. Therefore, computing a parameterization without compromising the reconstruction quality is highly desired. Second, our radiance decomposition operates under the assumptions of Lambertian reflectance and grayscale shading, which might hinder its effectiveness when dealing with intricate materials or sophisticated shadings." + } + ] + } + ], + "index": 24 + }, + { + "type": "image", + "bbox": [ + 139, + 558, + 176, + 587 + ], + "blocks": [ + { + "bbox": [ + 139, + 558, + 176, + 587 + ], + "lines": [ + { + "bbox": [ + 139, + 558, + 176, + 587 + ], + "spans": [ + { + "bbox": [ + 139, + 558, + 176, + 587 + ], + "type": "image", + "image_path": "8d4289f81d14eb1c9ce65cadae2ab9b437dbd150d0a4bed6e090a50fb327053c.jpg" + } + ] + } + ], + "index": 25, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 154, + 590, + 161, + 598 + ], + "lines": [ + { + "bbox": [ + 154, + 590, + 161, + 598 + ], + "spans": [ + { + "bbox": [ + 154, + 590, + 161, + 598 + ], + "type": "text", + "content": "S" + } + ] + } + ], + "index": 26, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 172, + 591, + 245, + 599 + ], + "lines": [ + { + "bbox": [ + 172, + 591, + 245, + 599 + ], + "spans": [ + { + "bbox": [ + 172, + 591, + 245, + 599 + ], + "type": "inline_equation", + "content": "\\mathcal{E}_{\\mathrm{area}} = 2.505" + }, + { + "bbox": [ + 172, + 591, + 245, + 599 + ], + "type": "text", + "content": " 6.094" + } + ] + } + ], + "index": 28, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 131, + 617, + 481, + 662 + ], + "lines": [ + { + "bbox": [ + 131, + 617, + 481, + 662 + ], + "spans": [ + { + "bbox": [ + 131, + 617, + 481, + 662 + ], + "type": "text", + "content": "Fig. 8: We present area distortion " + }, + { + "bbox": [ + 131, + 617, + 481, + 662 + ], + "type": "inline_equation", + "content": "\\mathcal{E}_{\\mathrm{area}}" + }, + { + "bbox": [ + 131, + 617, + 481, + 662 + ], + "type": "text", + "content": " based on the different choices of the parametric domain in (a) and geometry reconstruction results of VolSDF [58] and our method in (b). Our method is capable of editing through parameterization, accompanied by an acceptable compromise in reconstruction quality." + } + ] + } + ], + "index": 43, + "angle": 0, + "type": "image_caption" + } + ], + "index": 25 + }, + { + "type": "image", + "bbox": [ + 182, + 559, + 210, + 587 + ], + "blocks": [ + { + "bbox": [ + 182, + 559, + 210, + 587 + ], + "lines": [ + { + "bbox": [ + 182, + 559, + 210, + 587 + ], + "spans": [ + { + "bbox": [ + 182, + 559, + 210, + 587 + ], + "type": "image", + "image_path": "6b5e247eb1b1ef116e7444d83f3ffe4f73855ff430a8ffe7d3a2cf17f0a23e71.jpg" + } + ] + } + ], + "index": 27, + "angle": 0, + "type": "image_body" + } + ], + "index": 27 + }, + { + "type": "image", + "bbox": [ + 217, + 559, + 253, + 587 + ], + "blocks": [ + { + "bbox": [ + 217, + 559, + 253, + 587 + ], + "lines": [ + { + "bbox": [ + 217, + 559, + 253, + 587 + ], + "spans": [ + { + "bbox": [ + 217, + 559, + 253, + 587 + ], + "type": "image", + "image_path": "babc350263d4387380837fde4a940c8d855b6b527ed7a4ad8e1d8ab8d1f65b4d.jpg" + } + ] + } + ], + "index": 29, + "angle": 0, + "type": "image_body" + } + ], + "index": 29 + }, + { + "type": "image", + "bbox": [ + 264, + 557, + 293, + 587 + ], + "blocks": [ + { + "bbox": [ + 264, + 557, + 293, + 587 + ], + "lines": [ + { + "bbox": [ + 264, + 557, + 293, + 587 + ], + "spans": [ + { + "bbox": [ + 264, + 557, + 293, + 587 + ], + "type": "image", + "image_path": "86a7b75151500e7d032cd56115263caea9918a4822ef798c1076b4c1bd5b8205.jpg" + } + ] + } + ], + "index": 30, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 264, + 590, + 283, + 598 + ], + "lines": [ + { + "bbox": [ + 264, + 590, + 283, + 598 + ], + "spans": [ + { + "bbox": [ + 264, + 590, + 283, + 598 + ], + "type": "text", + "content": "1.633" + } + ] + } + ], + "index": 31, + "angle": 0, + "type": "image_caption" + } + ], + "index": 30 + }, + { + "type": "image", + "bbox": [ + 300, + 558, + 318, + 586 + ], + "blocks": [ + { + "bbox": [ + 300, + 558, + 318, + 586 + ], + "lines": [ + { + "bbox": [ + 300, + 558, + 318, + 586 + ], + "spans": [ + { + "bbox": [ + 300, + 558, + 318, + 586 + ], + "type": "image", + "image_path": "256ec429d93fffdd5d89afb162e25fd169930e7e1483ae18b86645bd1bf6c960.jpg" + } + ] + } + ], + "index": 32, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 301, + 589, + 315, + 597 + ], + "lines": [ + { + "bbox": [ + 301, + 589, + 315, + 597 + ], + "spans": [ + { + "bbox": [ + 301, + 589, + 315, + 597 + ], + "type": "text", + "content": "GT" + } + ] + } + ], + "index": 33, + "angle": 0, + "type": "image_caption" + } + ], + "index": 32 + }, + { + "type": "image", + "bbox": [ + 321, + 559, + 338, + 586 + ], + "blocks": [ + { + "bbox": [ + 321, + 559, + 338, + 586 + ], + "lines": [ + { + "bbox": [ + 321, + 559, + 338, + 586 + ], + "spans": [ + { + "bbox": [ + 321, + 559, + 338, + 586 + ], + "type": "image", + "image_path": "6ced61452f26507a3bc11e4290547712db9c5517aebeea9afd80d1f17f459ffb.jpg" + } + ] + } + ], + "index": 34, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 322, + 589, + 354, + 597 + ], + "lines": [ + { + "bbox": [ + 322, + 589, + 354, + 597 + ], + "spans": [ + { + "bbox": [ + 322, + 589, + 354, + 597 + ], + "type": "text", + "content": "VolSDF" + } + ] + } + ], + "index": 35, + "angle": 0, + "type": "image_caption" + } + ], + "index": 34 + }, + { + "type": "image", + "bbox": [ + 363, + 559, + 380, + 587 + ], + "blocks": [ + { + "bbox": [ + 363, + 559, + 380, + 587 + ], + "lines": [ + { + "bbox": [ + 363, + 559, + 380, + 587 + ], + "spans": [ + { + "bbox": [ + 363, + 559, + 380, + 587 + ], + "type": "image", + "image_path": "384f501987360754e48a3213a7bc6e07ed05d784b086a25116648b59fdcf736c.jpg" + } + ] + } + ], + "index": 36, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 368, + 590, + 389, + 597 + ], + "lines": [ + { + "bbox": [ + 368, + 590, + 389, + 597 + ], + "spans": [ + { + "bbox": [ + 368, + 590, + 389, + 597 + ], + "type": "text", + "content": "Ours" + } + ] + } + ], + "index": 37, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 370, + 598, + 383, + 609 + ], + "lines": [ + { + "bbox": [ + 370, + 598, + 383, + 609 + ], + "spans": [ + { + "bbox": [ + 370, + 598, + 383, + 609 + ], + "type": "text", + "content": "(b)" + } + ] + } + ], + "index": 39, + "angle": 0, + "type": "image_caption" + } + ], + "index": 36 + }, + { + "type": "image", + "bbox": [ + 384, + 561, + 402, + 586 + ], + "blocks": [ + { + "bbox": [ + 384, + 561, + 402, + 586 + ], + "lines": [ + { + "bbox": [ + 384, + 561, + 402, + 586 + ], + "spans": [ + { + "bbox": [ + 384, + 561, + 402, + 586 + ], + "type": "image", + "image_path": "a83df78bc300a819ef6d625e81ee15521b238974721cba3f29a13da97eec0f34.jpg" + } + ] + } + ], + "index": 38, + "angle": 0, + "type": "image_body" + } + ], + "index": 38 + }, + { + "bbox": [ + 406, + 567, + 471, + 574 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 406, + 567, + 471, + 574 + ], + "spans": [ + { + "bbox": [ + 406, + 567, + 471, + 574 + ], + "type": "text", + "content": "Method PSNR CD(10-3)" + } + ] + } + ], + "index": 40 + }, + { + "bbox": [ + 407, + 577, + 465, + 585 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 407, + 577, + 465, + 585 + ], + "spans": [ + { + "bbox": [ + 407, + 577, + 465, + 585 + ], + "type": "text", + "content": "VolSDF 31.15 3.270" + } + ] + } + ], + "index": 41 + }, + { + "bbox": [ + 410, + 587, + 466, + 596 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 410, + 587, + 466, + 596 + ], + "spans": [ + { + "bbox": [ + 410, + 587, + 466, + 596 + ], + "type": "text", + "content": "Ours 30.87 3.423" + } + ] + } + ], + "index": 42 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "spans": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "type": "text", + "content": "14" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 166, + 91, + 216, + 100 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 166, + 91, + 216, + 100 + ], + "spans": [ + { + "bbox": [ + 166, + 91, + 216, + 100 + ], + "type": "text", + "content": "B. Xu et al." + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 13 + }, + { + "para_blocks": [ + { + "bbox": [ + 133, + 114, + 234, + 129 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 114, + 234, + 129 + ], + "spans": [ + { + "bbox": [ + 133, + 114, + 234, + 129 + ], + "type": "text", + "content": "Acknowledgment" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 130, + 140, + 482, + 201 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 140, + 482, + 201 + ], + "spans": [ + { + "bbox": [ + 130, + 140, + 482, + 201 + ], + "type": "text", + "content": "This study is supported under the RIE2020 Industry Alignment Fund - Industry Collaboration Projects (IAF-ICP) Funding Initiative, as well as cash and in-kind contribution from the industry partner(s). This project is also partially supported by the Ministry of Education, Singapore, under its Academic Research Fund Grants (MOE-T2EP20220-0005 & RT19/22)." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 133, + 220, + 197, + 232 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 220, + 197, + 232 + ], + "spans": [ + { + "bbox": [ + 133, + 220, + 197, + 232 + ], + "type": "text", + "content": "References" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 138, + 246, + 481, + 665 + ], + "type": "list", + "angle": 0, + "index": 19, + "blocks": [ + { + "bbox": [ + 138, + 246, + 481, + 290 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 246, + 481, + 290 + ], + "spans": [ + { + "bbox": [ + 138, + 246, + 481, + 290 + ], + "type": "text", + "content": "1. Bednarik, J., Parashar, S., Gundogdu, E., Salzmann, M., Fua, P.: Shape reconstruction by learning differentiable surface representations. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 4716-4725 (2020) 3" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 138, + 291, + 481, + 322 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 291, + 481, + 322 + ], + "spans": [ + { + "bbox": [ + 138, + 291, + 481, + 322 + ], + "type": "text", + "content": "2. Biermann, H., Martin, I., Bernardini, F., Zorin, D.: Cut-and-paste editing of multiresolution surfaces. ACM transactions on graphics (TOG) 21(3), 312-321 (2002) 3" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 138, + 323, + 481, + 344 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 323, + 481, + 344 + ], + "spans": [ + { + "bbox": [ + 138, + 323, + 481, + 344 + ], + "type": "text", + "content": "3. Degener, P., Meseth, J., Klein, R.: An adaptable surface parameterization method. IMR 3, 201-213 (2003) 10, 13" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 138, + 346, + 481, + 378 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 346, + 481, + 378 + ], + "spans": [ + { + "bbox": [ + 138, + 346, + 481, + 378 + ], + "type": "text", + "content": "4. Fan, Q., Yang, J., Hua, G., Chen, B., Wipf, D.: Revisiting deep intrinsic image decompositions. In: Proceedings of the IEEE conference on computer vision and pattern recognition. pp. 8944-8952 (2018) 8" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 138, + 379, + 481, + 400 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 379, + 481, + 400 + ], + "spans": [ + { + "bbox": [ + 138, + 379, + 481, + 400 + ], + "type": "text", + "content": "5. Fang, H., Hart, J.C.: Textureshop: texture synthesis as a photograph editing tool. ACM Transactions on Graphics (TOG) 23(3), 354-359 (2004) 3" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 138, + 401, + 481, + 422 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 401, + 481, + 422 + ], + "spans": [ + { + "bbox": [ + 138, + 401, + 481, + 422 + ], + "type": "text", + "content": "6. Floater, M.S., Hormann, K.: Surface parameterization: a tutorial and survey. Advances in multiresolution for geometric modelling pp. 157-186 (2005) 3, 8" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 138, + 423, + 481, + 456 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 423, + 481, + 456 + ], + "spans": [ + { + "bbox": [ + 138, + 423, + 481, + 456 + ], + "type": "text", + "content": "7. García, I., Xia, J., He, Y., Xin, S., Patow, G.: Interactive applications for sketch-based editable polycube map. IEEE Trans. Vis. Comput. Graph. 19(7), 1158-1171 (2013) 3" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 138, + 456, + 481, + 478 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 456, + 481, + 478 + ], + "spans": [ + { + "bbox": [ + 138, + 456, + 481, + 478 + ], + "type": "text", + "content": "8. Gotsman, C., Gu, X., Sheffer, A.: Fundamentals of spherical parameterization for 3d meshes. ACM Trans. Graph. 22(3), 358-363 (2003) 3" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 138, + 479, + 481, + 500 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 479, + 481, + 500 + ], + "spans": [ + { + "bbox": [ + 138, + 479, + 481, + 500 + ], + "type": "text", + "content": "9. Greene, N.: Environment mapping and other applications of world projections. IEEE computer graphics and Applications 6(11), 21-29 (1986) 12" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 138, + 501, + 481, + 533 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 501, + 481, + 533 + ], + "spans": [ + { + "bbox": [ + 138, + 501, + 481, + 533 + ], + "type": "text", + "content": "0. Gropp, A., Yariv, L., Haim, N., Atzmon, M., Lipman, Y.: Implicit geometric regularization for learning shapes. In: International Conference on Machine Learning. pp. 3789-3799. PMLR (2020) 9" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 138, + 533, + 481, + 566 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 533, + 481, + 566 + ], + "spans": [ + { + "bbox": [ + 138, + 533, + 481, + 566 + ], + "type": "text", + "content": "1. Groueix, T., Fisher, M., Kim, V.G., Russell, B.C., Aubry, M.: A papier-mâché approach to learning 3d surface generation. In: Proceedings of the IEEE conference on computer vision and pattern recognition. pp. 216-224 (2018) 3" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 138, + 567, + 481, + 599 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 567, + 481, + 599 + ], + "spans": [ + { + "bbox": [ + 138, + 567, + 481, + 599 + ], + "type": "text", + "content": "2. Gu, X., Wang, Y., Chan, T.F., Thompson, P.M., Yau, S.: Genus zero surface conformal mapping and its application to brain surface mapping. IEEE Trans. Medical Imaging 23(8), 949-958 (2004) 8" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 138, + 600, + 481, + 632 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 600, + 481, + 632 + ], + "spans": [ + { + "bbox": [ + 138, + 600, + 481, + 632 + ], + "type": "text", + "content": "3. Gu, X., Yau, S.: Global conformal parameterization. In: Kobbelt, L., Schröder, P., Hoppe, H. (eds.) First Eurographics Symposium on Geometry Processing. vol. 43, pp. 127-137 (2003) 8" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 138, + 633, + 481, + 665 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 633, + 481, + 665 + ], + "spans": [ + { + "bbox": [ + 138, + 633, + 481, + 665 + ], + "type": "text", + "content": "4. Guo, H., Liu, S., Pan, H., Liu, Y., Tong, X., Guo, B.: Complexgen: Cad reconstruction by b-rep chain complex generation. ACM Transactions on Graphics (TOG) 41(4), 1-18 (2022) 3" + } + ] + } + ], + "index": 18 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 403, + 91, + 447, + 100 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 403, + 91, + 447, + 100 + ], + "spans": [ + { + "bbox": [ + 403, + 91, + 447, + 100 + ], + "type": "text", + "content": "NeuParam" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 470, + 91, + 481, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 470, + 91, + 481, + 100 + ], + "spans": [ + { + "bbox": [ + 470, + 91, + 481, + 100 + ], + "type": "text", + "content": "15" + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 14 + }, + { + "para_blocks": [ + { + "bbox": [ + 132, + 116, + 481, + 666 + ], + "type": "list", + "angle": 0, + "index": 18, + "blocks": [ + { + "bbox": [ + 133, + 116, + 481, + 138 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 116, + 481, + 138 + ], + "spans": [ + { + "bbox": [ + 133, + 116, + 481, + 138 + ], + "type": "text", + "content": "15. He, Y., Wang, H., Fu, C.W., Qin, H.: A divide-and-conquer approach for automatic polycube map construction. Computers & Graphics 33(3), 369-380 (2009) 10, 13" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 133, + 138, + 481, + 170 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 138, + 481, + 170 + ], + "spans": [ + { + "bbox": [ + 133, + 138, + 481, + 170 + ], + "type": "text", + "content": "16. Huang, Y.H., He, Y., Yuan, Y.J., Lai, Y.K., Gao, L.: Stylizednerf: Consistent 3d scene stylization as stylized nerf via 2d-3d mutual learning. In: Computer Vision and Pattern Recognition (CVPR) (2022) 4" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 133, + 170, + 481, + 203 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 170, + 481, + 203 + ], + "spans": [ + { + "bbox": [ + 133, + 170, + 481, + 203 + ], + "type": "text", + "content": "17. Kreisselmeier, G., Steinhauser, R.: Systematic control design by optimizing a vector performance index. In: Computer aided design of control systems, pp. 113-117. Elsevier (1980) 6" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 132, + 203, + 481, + 246 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 203, + 481, + 246 + ], + "spans": [ + { + "bbox": [ + 132, + 203, + 481, + 246 + ], + "type": "text", + "content": "18. Kuang, Z., Luan, F., Bi, S., Shu, Z., Wetzstein, G., Sunkavalli, K.: Palettenerf: Palette-based appearance editing of neural radiance fields. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 20691-20700 (2023) 2, 4" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 132, + 246, + 481, + 278 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 246, + 481, + 278 + ], + "spans": [ + { + "bbox": [ + 132, + 246, + 481, + 278 + ], + "type": "text", + "content": "19. Li, M., Zhang, H.: D2im-net: Learning detail disentangled implicit fields from single images. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 10246-10255 (2021) 4" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 132, + 278, + 481, + 320 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 278, + 481, + 320 + ], + "spans": [ + { + "bbox": [ + 132, + 278, + 481, + 320 + ], + "type": "text", + "content": "20. Li, Z., Müller, T., Evans, A., Taylor, R.H., Unberath, M., Liu, M.Y., Lin, C.H.: Neuralangelo: High-fidelity neural surface reconstruction. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 8456-8465 (2023) 4" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 132, + 320, + 481, + 363 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 320, + 481, + 363 + ], + "spans": [ + { + "bbox": [ + 132, + 320, + 481, + 363 + ], + "type": "text", + "content": "21. Li, Z., Shi, J., Bi, S., Zhu, R., Sunkavalli, K., Hasan, M., Xu, Z., Ramamoorthi, R., Chandraker, M.: Physically-based editing of indoor scene lighting from a single image. In: European Conference on Computer Vision. pp. 555-572. Springer (2022) 4" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 132, + 363, + 481, + 396 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 363, + 481, + 396 + ], + "spans": [ + { + "bbox": [ + 132, + 363, + 481, + 396 + ], + "type": "text", + "content": "22. Lin, C., Mitra, N., Wetzstein, G., Guibas, L.J., Guerrero, P.: Neuform: Adaptive overfitting for neural shape editing. Advances in Neural Information Processing Systems 35, 15217-15229 (2022) 4" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 132, + 396, + 481, + 429 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 396, + 481, + 429 + ], + "spans": [ + { + "bbox": [ + 132, + 396, + 481, + 429 + ], + "type": "text", + "content": "23. Liu, S., Zhang, X., Zhang, Z., Zhang, R., Zhu, J.Y., Russell, B.: Editing conditional radiance fields. In: Proceedings of the IEEE/CVF international conference on computer vision. pp. 5773-5783 (2021) 2" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 132, + 429, + 481, + 460 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 429, + 481, + 460 + ], + "spans": [ + { + "bbox": [ + 132, + 429, + 481, + 460 + ], + "type": "text", + "content": "24. Low, W.F., Lee, G.H.: Minimal neural atlas: Parameterizing complex surfaces with minimal charts and distortion. In: European Conference on Computer Vision. pp. 465-481. Springer (2022) 3" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 132, + 460, + 481, + 493 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 460, + 481, + 493 + ], + "spans": [ + { + "bbox": [ + 132, + 460, + 481, + 493 + ], + "type": "text", + "content": "25. Ma, L., Li, X., Liao, J., Wang, X., Zhang, Q., Wang, J., Sander, P.V.: Neural parameterization for dynamic human head editing. ACM Transactions on Graphics (TOG) 41(6), 1-15 (2022) 2, 4, 8, 9, 11, 12" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 132, + 493, + 481, + 536 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 493, + 481, + 536 + ], + "spans": [ + { + "bbox": [ + 132, + 493, + 481, + 536 + ], + "type": "text", + "content": "26. Mescheder, L., Oechsle, M., Niemeyer, M., Nowozin, S., Geiger, A.: Occupancy networks: Learning 3d reconstruction in function space. In: Proceedings of the IEEE/CVF conference on computer vision and pattern recognition. pp. 4460-4470 (2019) 4" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 132, + 536, + 481, + 568 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 536, + 481, + 568 + ], + "spans": [ + { + "bbox": [ + 132, + 536, + 481, + 568 + ], + "type": "text", + "content": "27. Mildenhall, B., Srinivasan, P.P., Tancik, M., Barron, J.T., Ramamoorthi, R., Ng, R.: Nerf: Representing scenes as neural radiance fields for view synthesis. In: ECCV (2020) 1, 7, 8" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 132, + 568, + 481, + 601 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 568, + 481, + 601 + ], + "spans": [ + { + "bbox": [ + 132, + 568, + 481, + 601 + ], + "type": "text", + "content": "28. Müller, T., Evans, A., Schied, C., Keller, A.: Instant neural graphics primitives with a multiresolution hash encoding. ACM Transactions on Graphics (ToG) 41(4), 1-15 (2022) 4" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 132, + 601, + 481, + 633 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 601, + 481, + 633 + ], + "spans": [ + { + "bbox": [ + 132, + 601, + 481, + 633 + ], + "type": "text", + "content": "29. Niemeyer, M., Geiger, A.: Giraffe: Representing scenes as compositional generative neural feature fields. In: Proc. IEEE Conf. on Computer Vision and Pattern Recognition (CVPR) (2021) 4" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 132, + 633, + 481, + 666 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 633, + 481, + 666 + ], + "spans": [ + { + "bbox": [ + 132, + 633, + 481, + 666 + ], + "type": "text", + "content": "30. Oechsle, M., Peng, S., Geiger, A.: Unisurf: Unifying neural implicit surfaces and radiance fields for multi-view reconstruction. In: Proceedings of the IEEE/CVF International Conference on Computer Vision. pp. 5589-5599 (2021) 1" + } + ] + } + ], + "index": 17 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "spans": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "type": "text", + "content": "16" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 166, + 91, + 215, + 100 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 166, + 91, + 215, + 100 + ], + "spans": [ + { + "bbox": [ + 166, + 91, + 215, + 100 + ], + "type": "text", + "content": "B. Xu et al." + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 15 + }, + { + "para_blocks": [ + { + "bbox": [ + 130, + 116, + 482, + 665 + ], + "type": "list", + "angle": 0, + "index": 19, + "blocks": [ + { + "bbox": [ + 130, + 116, + 482, + 161 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 116, + 482, + 161 + ], + "spans": [ + { + "bbox": [ + 130, + 116, + 482, + 161 + ], + "type": "text", + "content": "31. Park, J.J., Florence, P., Straub, J., Newcombe, R., Lovegrove, S.: Deepsdf: Learning continuous signed distance functions for shape representation. In: Proceedings of the IEEE/CVF conference on computer vision and pattern recognition. pp. 165-174 (2019) 4, 9" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 130, + 162, + 481, + 184 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 162, + 481, + 184 + ], + "spans": [ + { + "bbox": [ + 130, + 162, + 481, + 184 + ], + "type": "text", + "content": "32. Pérez, P., Gangnet, M., Blake, A.: Poisson image editing. In: Seminal Graphics Papers: Pushing the Boundaries, Volume 2, pp. 577-582 (2023) 4" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 132, + 185, + 481, + 206 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 185, + 481, + 206 + ], + "spans": [ + { + "bbox": [ + 132, + 185, + 481, + 206 + ], + "type": "text", + "content": "33. Praun, E., Hoppe, H.: Spherical parametrization and remeshing. ACM transactions on graphics (TOG) 22(3), 340-349 (2003) 3" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 132, + 207, + 481, + 239 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 207, + 481, + 239 + ], + "spans": [ + { + "bbox": [ + 132, + 207, + 481, + 239 + ], + "type": "text", + "content": "34. Qi, C.R., Su, H., Mo, K., Guibas, L.J.: Pointnet: Deep learning on point sets for 3d classification and segmentation. In: Proceedings of the IEEE conference on computer vision and pattern recognition. pp. 652-660 (2017) 3" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 132, + 240, + 481, + 274 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 240, + 481, + 274 + ], + "spans": [ + { + "bbox": [ + 132, + 240, + 481, + 274 + ], + "type": "text", + "content": "35. Qi, C.R., Su, H., Nießner, M., Dai, A., Yan, M., Guibas, L.J.: Volumetric and multi-view cnns for object classification on 3d data. In: Proceedings of the IEEE conference on computer vision and pattern recognition. pp. 5648-5656 (2016) 3" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 132, + 274, + 481, + 317 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 274, + 481, + 317 + ], + "spans": [ + { + "bbox": [ + 132, + 274, + 481, + 317 + ], + "type": "text", + "content": "36. Ramon, E., Triginer, G., Escur, J., Pumarola, A., Garcia, J., Giro-i Nieto, X., Moreno-Noguer, F.: H3d-net: Few-shot high-fidelity 3d head reconstruction. In: Proceedings of the IEEE/CVF International Conference on Computer Vision. pp. 5620-5629 (2021) 10, 11" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 132, + 319, + 481, + 352 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 319, + 481, + 352 + ], + "spans": [ + { + "bbox": [ + 132, + 319, + 481, + 352 + ], + "type": "text", + "content": "37. Rosu, R.A., Behnke, S.: Permutosdf: Fast multi-view reconstruction with implicit surfaces using permutohedral lattices. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 8466-8475 (2023) 4" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 132, + 352, + 481, + 385 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 352, + 481, + 385 + ], + "spans": [ + { + "bbox": [ + 132, + 352, + 481, + 385 + ], + "type": "text", + "content": "38. Sagnik Das, Ke Ma, Z.S., Samaras, D.: Learning an isometric surface parameterization for texture unwrapping. In: European Conference of Computer Vision 2022, ECCV 2022, Tel Aviv, Israel, October 23-27, 2022. Springer (2022) 4" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 132, + 386, + 481, + 430 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 386, + 481, + 430 + ], + "spans": [ + { + "bbox": [ + 132, + 386, + 481, + 430 + ], + "type": "text", + "content": "39. Saito, S., Huang, Z., Natsume, R., Morishima, S., Kanazawa, A., Li, H.: Pifu: Pixel-aligned implicit function for high-resolution clothed human digitization. In: Proceedings of the IEEE/CVF international conference on computer vision. pp. 2304-2314 (2019) 4" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 132, + 430, + 481, + 463 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 430, + 481, + 463 + ], + "spans": [ + { + "bbox": [ + 132, + 430, + 481, + 463 + ], + "type": "text", + "content": "40. Sheffer, A., Praun, E., Rose, K., et al.: Mesh parameterization methods and their applications. Foundations and Trends in Computer Graphics and Vision 2(2), 105-171 (2007) 3" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 132, + 464, + 481, + 496 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 464, + 481, + 496 + ], + "spans": [ + { + "bbox": [ + 132, + 464, + 481, + 496 + ], + "type": "text", + "content": "41. Srinivasan, P.P., Deng, B., Zhang, X., Tancik, M., Mildenhall, B., Barron, J.T.: Nerv: Neural reflectance and visibility fields for relighting and view synthesis. In: CVPR (2021) 4" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 132, + 497, + 481, + 519 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 497, + 481, + 519 + ], + "spans": [ + { + "bbox": [ + 132, + 497, + 481, + 519 + ], + "type": "text", + "content": "42. Sun, Q., Zhang, L., Zhang, M., Ying, X., Xin, S., Xia, J., He, Y.: Texture brush: an interactive surface texturing interface. pp. 153-160 (2013) 3" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 132, + 520, + 481, + 542 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 520, + 481, + 542 + ], + "spans": [ + { + "bbox": [ + 132, + 520, + 481, + 542 + ], + "type": "text", + "content": "43. Tarini, M., Hormann, K., Cignoni, P., Montani, C.: Polycube-maps. ACM transactions on graphics (TOG) 23(3), 853-860 (2004) 3" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 132, + 543, + 481, + 575 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 543, + 481, + 575 + ], + "spans": [ + { + "bbox": [ + 132, + 543, + 481, + 575 + ], + "type": "text", + "content": "44. Tojo, K., Umetani, N.: Recolorable posterization of volumetric radiance fields using visibility-weighted palette extraction. In: Computer Graphics Forum. vol. 41, pp. 149-160. Wiley Online Library (2022) 2, 4" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 132, + 576, + 481, + 609 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 576, + 481, + 609 + ], + "spans": [ + { + "bbox": [ + 132, + 576, + 481, + 609 + ], + "type": "text", + "content": "45. Wang, N., Zhang, Y., Li, Z., Fu, Y., Liu, W., Jiang, Y.G.: Pixel2mesh: Generating 3d mesh models from single rgb images. In: Proceedings of the European conference on computer vision (ECCV). pp. 52-67 (2018) 3" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 132, + 610, + 481, + 643 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 610, + 481, + 643 + ], + "spans": [ + { + "bbox": [ + 132, + 610, + 481, + 643 + ], + "type": "text", + "content": "46. Wang, P., Liu, L., Liu, Y., Theobalt, C., Komura, T., Wang, W.: Neus: Learning neural implicit surfaces by volume rendering for multi-view reconstruction. NeurIPS (2021) 1, 14" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 132, + 643, + 481, + 665 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 643, + 481, + 665 + ], + "spans": [ + { + "bbox": [ + 132, + 643, + 481, + 665 + ], + "type": "text", + "content": "47. Wang, X., Zhu, J., Ye, Q., Huo, Y., Ran, Y., Zhong, Z., Chen, J.: Seal-3d: Interactive pixel-level editing for neural radiance fields (2023) 2, 4" + } + ] + } + ], + "index": 18 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 403, + 91, + 447, + 100 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 403, + 91, + 447, + 100 + ], + "spans": [ + { + "bbox": [ + 403, + 91, + 447, + 100 + ], + "type": "text", + "content": "NeuParam" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 470, + 91, + 481, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 470, + 91, + 481, + 100 + ], + "spans": [ + { + "bbox": [ + 470, + 91, + 481, + 100 + ], + "type": "text", + "content": "17" + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 16 + }, + { + "para_blocks": [ + { + "bbox": [ + 132, + 117, + 481, + 666 + ], + "type": "list", + "angle": 0, + "index": 17, + "blocks": [ + { + "bbox": [ + 133, + 117, + 481, + 149 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 117, + 481, + 149 + ], + "spans": [ + { + "bbox": [ + 133, + 117, + 481, + 149 + ], + "type": "text", + "content": "48. Wang, Y., Rahmann, L., Sorkine-Hornung, O.: Geometry-consistent neural shape representation with implicit displacement fields. In: The Tenth International Conference on Learning Representations. OpenReview (2022) 4" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 133, + 150, + 481, + 182 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 150, + 481, + 182 + ], + "spans": [ + { + "bbox": [ + 133, + 150, + 481, + 182 + ], + "type": "text", + "content": "49. Wang, Y., Skorokhodov, I., Wonka, P.: Hf-neus: Improved surface reconstruction using high-frequency details. Advances in Neural Information Processing Systems 35, 1966-1978 (2022) 4" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 132, + 182, + 481, + 216 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 182, + 481, + 216 + ], + "spans": [ + { + "bbox": [ + 132, + 182, + 481, + 216 + ], + "type": "text", + "content": "50. Williams, F., Schneider, T., Silva, C., Zorin, D., Bruna, J., Panozzo, D.: Deep geometric prior for surface reconstruction. In: Proceedings of the IEEE/CVF conference on computer vision and pattern recognition. pp. 10130-10139 (2019) 3" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 132, + 216, + 481, + 248 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 216, + 481, + 248 + ], + "spans": [ + { + "bbox": [ + 132, + 216, + 481, + 248 + ], + "type": "text", + "content": "51. Wu, T., Zhang, J., Fu, X., Wang, Y., Ren, J., Pan, L., Wu, W., Yang, L., Wang, J., Qian, C., Lin, D., Liu, Z.: Omniobject3d: Large-vocabulary 3d object dataset for realistic perception, reconstruction and generation (2023) 10" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 132, + 249, + 481, + 281 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 249, + 481, + 281 + ], + "spans": [ + { + "bbox": [ + 132, + 249, + 481, + 281 + ], + "type": "text", + "content": "52. Xiang, F., Xu, Z., Hasan, M., Hold-Geoffroy, Y., Sunkavalli, K., Su, H.: NeuTex: Neural Texture Mapping for Volumetric Neural Rendering. In: The IEEE Conference on Computer Vision and Pattern Recognition (CVPR) (2021) 2, 4, 7" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 132, + 281, + 481, + 326 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 281, + 481, + 326 + ], + "spans": [ + { + "bbox": [ + 132, + 281, + 481, + 326 + ], + "type": "text", + "content": "53. Xu, B., Zhang, J., Lin, K.Y., Qian, C., He, Y.: Deformable model driven neural rendering for high-fidelity 3d reconstruction of human heads under low-view settings. In: Proceedings of the IEEE/CVF International Conference on Computer Vision (2023) 2, 4" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 132, + 326, + 481, + 369 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 326, + 481, + 369 + ], + "spans": [ + { + "bbox": [ + 132, + 326, + 481, + 369 + ], + "type": "text", + "content": "54. Yang, B., Bao, C., Zeng, J., Bao, H., Zhang, Y., Cui, Z., Zhang, G.: Neumesh: Learning disentangled neural mesh-based implicit field for geometry and texture editing. In: European Conference on Computer Vision. pp. 597-614. Springer (2022) 4, 12" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 132, + 369, + 481, + 403 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 369, + 481, + 403 + ], + "spans": [ + { + "bbox": [ + 132, + 369, + 481, + 403 + ], + "type": "text", + "content": "55. Yang, B., Zhang, Y., Xu, Y., Li, Y., Zhou, H., Bao, H., Zhang, G., Cui, Z.: Learning object-compositional neural radiance field for editable scene rendering. In: International Conference on Computer Vision (ICCV) (October 2021) 4" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 132, + 403, + 481, + 445 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 403, + 481, + 445 + ], + "spans": [ + { + "bbox": [ + 132, + 403, + 481, + 445 + ], + "type": "text", + "content": "56. Yang, H., Zhu, H., Wang, Y., Huang, M., Shen, Q., Yang, R., Cao, X.: Facescape: a large-scale high quality 3d face dataset and detailed riggable 3d face prediction. In: Proceedings of the IEEE/cvf conference on computer vision and pattern recognition. pp. 601-610 (2020) 10, 14" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 132, + 446, + 481, + 479 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 446, + 481, + 479 + ], + "spans": [ + { + "bbox": [ + 132, + 446, + 481, + 479 + ], + "type": "text", + "content": "57. Yang, L., Liang, Y., Li, X., Zhang, C., Lin, G., Sheffer, A., Schaefer, S., Keyser, J., Wang, W.: Neural parametric surfaces for shape modeling. arXiv preprint arXiv:2309.09911 (2023) 3" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 132, + 479, + 481, + 511 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 479, + 481, + 511 + ], + "spans": [ + { + "bbox": [ + 132, + 479, + 481, + 511 + ], + "type": "text", + "content": "58. Yariv, L., Gu, J., Kasten, Y., Lipman, Y.: Volume rendering of neural implicit surfaces. Advances in Neural Information Processing Systems 34, 4805-4815 (2021) 1, 9, 14" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 132, + 512, + 481, + 544 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 512, + 481, + 544 + ], + "spans": [ + { + "bbox": [ + 132, + 512, + 481, + 544 + ], + "type": "text", + "content": "59. Yariv, L., Kasten, Y., Moran, D., Galun, M., Atzmon, M., Ronen, B., Lipman, Y.: Multiview neural surface reconstruction by disentangling geometry and appearance. Advances in Neural Information Processing Systems 33 (2020) 9" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 132, + 545, + 481, + 588 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 545, + 481, + 588 + ], + "spans": [ + { + "bbox": [ + 132, + 545, + 481, + 588 + ], + "type": "text", + "content": "60. Ye, W., Chen, S., Bao, C., Bao, H., Pollefeys, M., Cui, Z., Zhang, G.: IntrinsicNeRF: Learning Intrinsic Neural Radiance Fields for Editable Novel View Synthesis. In: Proceedings of the IEEE/CVF International Conference on Computer Vision (2023) 4, 8" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 132, + 589, + 481, + 632 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 589, + 481, + 632 + ], + "spans": [ + { + "bbox": [ + 132, + 589, + 481, + 632 + ], + "type": "text", + "content": "61. Yenamandra, T., Tewari, A., Bernard, F., Seidel, H.P., Elgharib, M., Cremers, D., Theobalt, C.: i3dmm: Deep implicit 3d morphable model of human heads. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 12803-12813 (2021) 2, 4" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 132, + 632, + 481, + 666 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 632, + 481, + 666 + ], + "spans": [ + { + "bbox": [ + 132, + 632, + 481, + 666 + ], + "type": "text", + "content": "62. Yuan, Y.J., Sun, Y.T., Lai, Y.K., Ma, Y., Jia, R., Kobbelt, L., Gao, L.: Interactive nef geometry editing with shape priors. IEEE Transactions on Pattern Analysis and Machine Intelligence (2023) 4" + } + ] + } + ], + "index": 16 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "spans": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "type": "text", + "content": "18" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 166, + 91, + 215, + 100 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 166, + 91, + 215, + 100 + ], + "spans": [ + { + "bbox": [ + 166, + 91, + 215, + 100 + ], + "type": "text", + "content": "B. Xu et al." + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 17 + }, + { + "para_blocks": [ + { + "bbox": [ + 132, + 116, + 482, + 183 + ], + "type": "list", + "angle": 0, + "index": 4, + "blocks": [ + { + "bbox": [ + 132, + 116, + 482, + 149 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 116, + 482, + 149 + ], + "spans": [ + { + "bbox": [ + 132, + 116, + 482, + 149 + ], + "type": "text", + "content": "63. Zhang, Q., Hou, J., Qian, Y., Chan, A.B., Zhang, J., He, Y.: Reggeonet: Learning regular representations for large-scale 3d point clouds. Int. J. Comput. Vis. 130(12), 3100-3122 (2022) 3" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 132, + 150, + 482, + 183 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 150, + 482, + 183 + ], + "spans": [ + { + "bbox": [ + 132, + 150, + 482, + 183 + ], + "type": "text", + "content": "64. Zhang, Q., Hou, J., Qian, Y., Zeng, Y., Zhang, J., He, Y.: Flattening-net: Deep regular 2d representation for 3d point cloud analysis. IEEE Trans. Pattern Anal. Mach. Intell. 45(8), 9726-9742 (2023) 3" + } + ] + } + ], + "index": 3 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 403, + 91, + 447, + 100 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 403, + 91, + 447, + 100 + ], + "spans": [ + { + "bbox": [ + 403, + 91, + 447, + 100 + ], + "type": "text", + "content": "NeuParam" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 470, + 91, + 481, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 470, + 91, + 481, + 100 + ], + "spans": [ + { + "bbox": [ + 470, + 91, + 481, + 100 + ], + "type": "text", + "content": "19" + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 18 + } + ], + "_backend": "vlm", + "_version_name": "2.6.4" +} \ No newline at end of file diff --git a/2024/Parameterized Quasi-Physical Simulators for Dexterous Manipulations Transfer/ae282149-332d-459a-aebe-a693d1ac89fd_content_list.json b/2024/Parameterized Quasi-Physical Simulators for Dexterous Manipulations Transfer/ae282149-332d-459a-aebe-a693d1ac89fd_content_list.json new file mode 100644 index 0000000000000000000000000000000000000000..7b4a17c49a30bab3251dfd1539088bad644fd3a0 --- /dev/null +++ b/2024/Parameterized Quasi-Physical Simulators for Dexterous Manipulations Transfer/ae282149-332d-459a-aebe-a693d1ac89fd_content_list.json @@ -0,0 +1,1890 @@ +[ + { + "type": "text", + "text": "Parameterized Quasi-Physical Simulators for Dexterous Manipulations Transfer", + "text_level": 1, + "bbox": [ + 245, + 140, + 756, + 186 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Xueyi Liu $^{1,3}$ , Kangbo Lyu $^{1}$ , Jieqiong Zhang $^{1}$ , Tao Du $^{1,2,3}$ , and Li Yi $^{1,2,3}$", + "bbox": [ + 227, + 212, + 774, + 229 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "$^{1}$ Tsinghua University $^{2}$ Shanghai AI Laboratory $^{3}$ Shanghai Qi Zhi Institute https://meowu7.github.io/QuasiSim", + "bbox": [ + 236, + 239, + 764, + 270 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/9ad5b625ad6097a9287a6f84a453606d61ed514863586c2d9313a99e550120d8.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 246, + 316, + 312, + 367 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/d46ebd66664a8e9824b4bc4a38ff0319c5a628c34a96148c2b8581e84cac049c.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 248, + 368, + 310, + 416 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/944ad22b523cc633929f6e4f51084547ad0b17b00bec9e06fb856b317d937ab6.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 312, + 316, + 372, + 367 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/fd6cc2d5e7a115c6c7fdfee25e073dbf68c4df1eb5093be9e7230c23ea8d04d2.jpg", + "image_caption": [ + "(e) Result of a failed baseline" + ], + "image_footnote": [], + "bbox": [ + 312, + 368, + 372, + 417 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/0edadcc3e9758ed79a8aad322965386e9e91d087ce242cec0c73fc29b704884b.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 374, + 316, + 434, + 367 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/75270e327eb3aba02f2eda6843c5fa10cf480ab83aa757109c9587df0210a50e.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 374, + 368, + 434, + 417 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/ba69399690ce426d9287bc1f75cc59f3aa880cec7b32d13d6b10e1ebbbdad237.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 434, + 316, + 496, + 367 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/8180f3da0aa5fa07e0eb5556a305879fd99051828dee4246b4f7c99c53a86c40.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 434, + 368, + 496, + 417 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/30e7418e3123a22bbac62a8bd8aa308e845bd78284e41a20639982258f594a57.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 500, + 316, + 562, + 367 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/6abae0fc3a34145058f86b59fa3c5e861daf85c2165afab851543ddc36323164.jpg", + "image_caption": [ + "(f) Improved result via optimizing through the curriculum" + ], + "image_footnote": [], + "bbox": [ + 500, + 368, + 562, + 417 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/85bb9f79b34813de8c619a59a0b4a518693688115132bf53a6a90353a905cb84.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 563, + 316, + 625, + 367 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/5437d53ea46bd32cc5ebeb033b96f4ee8e0d47af725c3fe7e2638def7f06ee5e.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 563, + 368, + 625, + 417 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/743909a9a222135261106888ea15fd109ce2bb572622c6a49cc5bcbd31597076.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 625, + 316, + 687, + 367 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/8e7d30b22609c28a925427059d560e03a1b28788eb4f3e5c6770973167b182c7.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 625, + 368, + 687, + 417 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/e6b018b2d8a9175a37b4e174c9e694f9100858b74971cbcc8f106f8846623b26.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 687, + 316, + 748, + 367 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/201868126a183c1936cb9a37a5bdb7a6a4a612fca5d8c4b0dfd96842d278fbad.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 687, + 368, + 748, + 417 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/4e2e798039f97c7bfd5069933f0bd79e6bd9944792e9835348a69fa3084f0c34.jpg", + "image_caption": [ + "Timestamp", + "Fig. 1: By optimizing through a quasi-physical simulator curriculum, we successfully transfer human demonstrations to dexterous robot hand simulations. We enable accurate tracking of complex manipulations with changing contacts (Fig. (a)), nontrivial object motions (Fig. (b)) and intricate tool-using (Fig. (c,d)). Besides, our physics curriculum can substantially improve a failed baseline (Fig. (e,f))." + ], + "image_footnote": [], + "bbox": [ + 250, + 426, + 310, + 474 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/bf61530f9eee1b1889c739226cb4cedc69b104d8e13e63757eeac42440152d35.jpg", + "image_caption": [ + "__________" + ], + "image_footnote": [], + "bbox": [ + 312, + 426, + 372, + 474 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/258c6c475945a5eb1669ade573c7825ed880287a852285635fb55a5b7e0b85d7.jpg", + "image_caption": [ + "$\\left( {x - {2x}}\\right) t - x{y}^{2} = \\left( {x - {2x}}\\right) {f}^{\\prime }t$", + "#" + ], + "image_footnote": [], + "bbox": [ + 374, + 426, + 434, + 474 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/5acaa4e87ba70f059827886562ffbcb1bbe5eae44aaea88c37a916554423f4aa.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 434, + 426, + 496, + 474 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/4ee93e74bf06383554440472a7d95bc98360bdb5445b376563f56691ecc6a397.jpg", + "image_caption": [ + "Timestamp" + ], + "image_footnote": [], + "bbox": [ + 501, + 426, + 562, + 474 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/8663ea3d05dff9f3f1b53eec11297eb590d152876e3d9acaa097d72e2e58f7f1.jpg", + "image_caption": [ + "#", + "" + ], + "image_footnote": [], + "bbox": [ + 563, + 426, + 625, + 474 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/f723652b1dcf649b48ab6399b00dee90e13af0aafe50b90311e6f3373a0cee3c.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 625, + 426, + 687, + 474 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/1f6e7ad2fc61d7db57ae195837c09a233a1e5982d99a20eddbffc80d18f77904.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 687, + 426, + 748, + 474 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Abstract. We explore the dexterous manipulation transfer problem by designing simulators. The task wishes to transfer human manipulations to dexterous robot hand simulations and is inherently difficult due to its intricate, highly-constrained, and discontinuous dynamics and the need to control a dexterous hand with a DoF to accurately replicate human manipulations. Previous approaches that optimize in high-fidelity black-box simulators or a modified one with relaxed constraints only demonstrate limited capabilities or are restricted by insufficient simulation fidelity. We introduce parameterized quasi-physical simulators and a physics curriculum to overcome these limitations. The key ideas are 1) balancing between fidelity and optimizability of the simulation via a curriculum of parameterized simulators, and 2) solving the problem in each of the simulators from the curriculum, with properties ranging from high task optimizability to high fidelity. We successfully enable a dexterous hand to track complex and diverse manipulations in high-fidelity simulated environments, boosting the success rate by $11\\%+$ from the best-performed baseline. The project website is available at QuasiSim.", + "bbox": [ + 259, + 604, + 740, + 840 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "1 Introduction", + "text_level": 1, + "bbox": [ + 217, + 143, + 374, + 160 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Advancing an embodied agent's capacity to interact with the world represents a significant stride toward achieving general artificial intelligence. Due to the substantial costs and potential hazards of setting up real robots to do trial and error, the standard approach for developing embodied algorithms involves learning in physical simulators [9, 15, 23, 25, 33, 56, 59] before transitioning to real-world deployment. In most cases, physical simulators are treated as black boxes, and extensive efforts have been devoted to developing learning and optimization methods for embodied skills within these black boxes. Despite the considerable progress [2, 6-8, 16, 20, 21, 31, 36, 39, 43, 46, 60, 62, 66, 68], the question like whether the simulators used are the most suitable ones is rarely discussed. In this work, we investigate this issue and illustrate how optimizing the simulator concurrently with skill acquisition can benefit a popular yet challenging task in robot manipulation - dexterous manipulation transfer.", + "bbox": [ + 212, + 175, + 785, + 372 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "The task aims at transferring human-object manipulations to a dexterous robot hand, enabling it to physically track the reference motion of both the hand and the object (see Fig. 1). It is challenged by 1) the complex, highly constrained, non-smooth, and discontinuous dynamics with frequent contact establishment and breaking involved in the robot manipulation, 2) the requirement of precisely controlling a dexterous hand with a high DoF to densely track the manipulation at each frame, and 3) the morphology difference. Some existing works rely on high-fidelity black-box simulators, where a small difference in robot control can result in dramatically different manipulation outcomes due to abrupt contact changes, making the tracking objective highly non-smooth and hard to optimize [4,6,8,43,46]. In this way, their tasks are restricted to relatively simple goal-driven manipulations such as pouring and re-locating [8,43,46,68], in-hand re-orientation, flipping and spinning [4,6] with a fixed-root robot hand, or manipulating objects with simple geometry such as balls [36]. Other approaches attempt to improve optimization by relaxing physical constraints, with a primary focus on smoothing out contact responses [3,24,38,55,56]. However, their dynamics models may significantly deviate from real physics [38], hindering skill deployment. Consequently, we ask how to address the optimization challenge while preserving the high fidelity of the simulator.", + "bbox": [ + 212, + 372, + 785, + 657 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Our key insight is that a single simulator can hardly provide both high fidelity and excellent optimizability for contact-rich dexterous manipulations. Inspired by the line of homotopy methods [12, 28, 29, 61], we propose a curriculum of simulators to realize this. We start by utilizing a quasi-physical simulator to initially relax physical constraints and warm up the optimization. Subsequently, we transfer the optimization outcomes to simulators with gradually tightened physical constraints. Finally, we transition to a physically realistic simulator for skill deployment in realistic dynamics.", + "bbox": [ + 212, + 657, + 785, + 779 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "To realize this vision, we propose a family of parameterized quasi-physical simulators for contact-rich dexterous manipulation tasks. These simulators can be customized to enhance task optimizability while can also be tailored to approximate realistic physics. The parameterized simulator represents", + "bbox": [ + 212, + 779, + 785, + 839 + ], + "page_idx": 1 + }, + { + "type": "page_number", + "text": "2", + "bbox": [ + 217, + 114, + 228, + 126 + ], + "page_idx": 1 + }, + { + "type": "header", + "text": "X. Liu et al.", + "bbox": [ + 271, + 114, + 354, + 126 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "an articulated multi rigid body as a parameterized point set, models contact using an unconstrained parameterized spring-damper, and compensates for unmodeled effects via parameterized residual physics. Specifically, the articulated multi-body dynamics model is relaxed as the point set dynamics model. An articulated object is relaxed into a set of points, sampled from the ambient space surrounding each body's surface mesh. The resulting dynamics model combines the original articulated dynamics with the mass-point dynamics of each individual point. Parameters are introduced to control the point set construction and the dynamics model. The contact model is softened as a parameterized spring-damper model [3, 19, 35, 38, 51] with parameters introduced to control when to calculate contacts and contact spring stiffness. The residual physics network compensate for unmodeled effects from the analytical modeling [22]. The parameterized simulator can be programmed for high optimizability by relaxing constraints in the analytical model and can be tailored to approximate realistic physics by learning excellent residual physics. We demonstrate that the challenging dexterous manipulation transfer task can be effectively addressed through curriculum optimization using a series of parameterized physical simulators. Initially, both articulated rigid constraints and the contact model stiffness are relaxed in the simulator. It may not reflect physical realism but provides a good environment where the manipulation transfer problem can be solved easily. Subsequently, the articulated rigid constraints and the contact model are gradually tightened. Task-solving proceeds iteratively within each simulator in the curriculum. Finally, the parameterized simulator is optimized to approximate realistic physics. Task optimization continues, yielding a dexterous hand trajectory capable of executing the manipulation in environments with realistic physics.", + "bbox": [ + 217, + 147, + 785, + 537 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "We demonstrate the superiority of our method and compare it with previous model-free and model-based methods on challenging manipulation sequences from three datasets, describing single-hand or bimanual manipulations with daily objects or using tools. We conduct dexterous manipulation transfer on two widely used simulators, namely Bullet [9] and Isaac Gym [33] to demonstrate the generality and the efficacy of our method and the capability of our quasi-physical simulator to approximate the unknown black-box physics model in the contact-rich manipulation scenario (Fig. 1). We can track complex manipulations involving non-trivial object motions such as large rotations and complicated tool-using such as using a spoon to bring the water back and forth. Our approach successfully surpasses the previous best-performed method both quantitatively and qualitatively, achieving more than $11\\%$ success rate than the previous best-performed method. Besides, optimizing through the physics curriculum can significantly enhance the performance of previously under-performed RL-based methods, almost completing the tracking problem from failure, as demonstrated in Fig. 1. This indicates the universality of our approach to embodied AI through optimization via a physics curriculum. Thorough ablations are conducted to validate the efficacy of our designs.", + "bbox": [ + 217, + 550, + 785, + 821 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Our contributions are three-fold:", + "bbox": [ + 240, + 833, + 477, + 847 + ], + "page_idx": 2 + }, + { + "type": "header", + "text": "Quasi-Physical Simulators for Dexterous Manipulations Transfer", + "bbox": [ + 302, + 114, + 730, + 128 + ], + "page_idx": 2 + }, + { + "type": "page_number", + "text": "3", + "bbox": [ + 774, + 116, + 784, + 126 + ], + "page_idx": 2 + }, + { + "type": "list", + "sub_type": "text", + "list_items": [ + "- We introduce a family of parameterized quasi-physical simulators that can be configured to relax various physical constraints, facilitating skill optimization, and can also be tailored to achieve high simulation fidelity.", + "- We present a quasi-physics curriculum along with a corresponding optimization method to address the challenging dexterous manipulation transfer problem.", + "- Extensive experiments demonstrate the effectiveness of our method in transferring complex manipulations, including non-trivial object motions and changing contacts, to a dexterous robot hand in simulation." + ], + "bbox": [ + 223, + 146, + 784, + 281 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "2 Related Works", + "text_level": 1, + "bbox": [ + 215, + 297, + 395, + 314 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Dexterous manipulation transfer. Transferring human manipulations to dexterous robot-hand simulations is an important topic in robot skill acquisition [8,21,31,43,60,62,68,70]. Most approaches treat the simulator as black-box physics models and try to learn skills directly from that [4,6,8,43,46]. However, their demonstrated capabilities are restricted to relatively simple tasks. Another trend of work tries to relax the physics model [37,38] to create a better environment for task optimization. However, due to the disparity between their modeling approach and realistic physics, successful trials are typically demonstrated only in their simulators, which can hardly complete the task under physically realistic dynamics. In this work, we introduce various parameterized analytical relaxations to improve the task optimizability while compensating for unmodeled effects via residual physics networks so the fidelity would not be sacrificed.", + "bbox": [ + 212, + 320, + 785, + 500 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Learning for simulation. Analytical methods can hardly approximate an extremely realistic physical world despite lots of smart and tremendous efforts made in developing numerical algorithms [19, 23, 26, 27]. Recently, data-driven approaches have attracted lots of interest for their high efficiency and strong approximation ability [10, 11, 22, 40, 41, 50, 63]. Special network designs are proposed to learn the contact behaviour [22, 41]. We in this work propose to leverage an analytical-neural hybrid approach and carefully design network modules for approximating residual contact forces in the contact-rich manipulation scenario.", + "bbox": [ + 212, + 502, + 785, + 621 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Sim-to-Sim and Sim-to-Real transfer. The field of robot manipulation continues to face challenges in the areas of Sim2Sim and Sim2Real transferability [71]. Considering the modeling gaps, the optimal strategy learned in a specific simulator is difficult to transfer to a different simulator or the real world. Therefore, many techniques for solving the problem have been proposed, including imitation learning [34,42,43,45,46,48], transfer learning [72], distillation [47,57], residual physics [17,67], and efforts on bridging the gap from the dynamics model aspect [22,69]. Our parameterized simulators learn residual physics involved in contact-rich robot manipulations. By combining an analytical base with residual networks, we showcase their ability to approximate realistic physics.", + "bbox": [ + 212, + 622, + 785, + 773 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "3 Method", + "text_level": 1, + "bbox": [ + 215, + 787, + 330, + 801 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Given a human manipulation demonstration, composed of a human hand mesh trajectory and an object pose trajectory $\\{\\mathcal{H},\\mathcal{O}\\}$ , the goal is transferring the", + "bbox": [ + 212, + 809, + 785, + 840 + ], + "page_idx": 3 + }, + { + "type": "page_number", + "text": "4", + "bbox": [ + 217, + 114, + 228, + 126 + ], + "page_idx": 3 + }, + { + "type": "header", + "text": "X. Liu et al.", + "bbox": [ + 271, + 114, + 354, + 126 + ], + "page_idx": 3 + }, + { + "type": "image", + "img_path": "images/99664072a887cf47fdcf35e54f57509bda3392ca2b5ecd914f4b592804383add.jpg", + "image_caption": [ + "Parameterized Quasi-Physical Simulator", + "Fig. 2: The parameterized quasi-physical simulator relaxes the articulated multi rigid body dynamics as the parameterized point set dynamics, controls the contact behavior via an unconstrained parameterized spring-damper contact model, and compensates for unmodeled effects via parameterized residual physics networks. We tackle the difficult dexterous manipulation transfer problem via a physics curriculum." + ], + "image_footnote": [], + "bbox": [ + 218, + 143, + 781, + 265 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "demonstration to a dexterous robot hand in simulation. Formally, we aim to optimize a control trajectory $\\mathcal{A}$ that drives the dexterous hand to manipulate the object in a realistic simulated environment so that the resulting hand trajectory $\\hat{\\mathcal{H}}$ and the object trajectory $\\hat{\\mathcal{O}}$ are close to the reference motion $\\{\\mathcal{H},\\mathcal{O}\\}$ . The problem is challenged by difficulties from the highly constrained, discontinuous, and non-smooth dynamics, the requirement of controlling a high DoF dexterous hand for tracking, and the morphology difference.", + "bbox": [ + 212, + 352, + 787, + 458 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Our method comprises two key designs to tackle the challenges: 1) a family of parameterized quasi-physical simulators, which can be programmed to enhance the optimizability of contact-rich dexterous manipulation tasks and can also be tailored to approximate realistic physics (Section 3.1), and 2) a physics curriculum that carefully adjusts the parameters of a line of quasi-physical simulators and a strategy that solves the difficult dexterous manipulation transfer task by addressing it within each simulator in the curriculum (Section 3.2).", + "bbox": [ + 212, + 458, + 787, + 564 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "3.1 Parameterized Quasi-Physical Simulators", + "text_level": 1, + "bbox": [ + 214, + 587, + 602, + 602 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Our quasi-physical simulator represents an articulated multi-body, i.e., the robotic dexterous hand, as a point set. The object is represented as a signed distance field. The base of the simulator is in an analytical form leveraging an unconstrained spring-damper contact model. Parameters are introduced to control the analytical relaxations on the articulated rigid constraints and the softness of the contact model. Additionally, neural networks are introduced to compensate for unmodeled effects beyond the analytical framework. We will elaborate on each of these design aspects below.", + "bbox": [ + 212, + 613, + 792, + 733 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Parameterized point set dynamics. Articulated multi-body represented in the reduced coordinate system [19,59] may require a large change in joint states to achieve a small adjustment in the Euclidean space. Moving the end effector from one point to a nearby point may require adjusting all joint states (Fig. 3). Besides, transferring the hand trajectory to a morphologically different hand requires correspondences to make the resulting trajectory close to the original one. Defining correspondences in the reduced coordinate or via sparse correspon", + "bbox": [ + 212, + 734, + 787, + 840 + ], + "page_idx": 4 + }, + { + "type": "header", + "text": "Quasi-Physical Simulators for Dexterous Manipulations Transfer", + "bbox": [ + 300, + 114, + 732, + 128 + ], + "page_idx": 4 + }, + { + "type": "page_number", + "text": "5", + "bbox": [ + 774, + 116, + 784, + 126 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "dences will make the result suffer from noise in the data, leading to unwanted results finally (Fig. 3). Hence, we propose relaxing an articulated multi-rigid body into a mass-point set sampled from the ambient space surrounding each body. Each point is considered attached to the body from which it is sampled and is capable of both self-actuation and actuation via joint motors. We introduce a parameter $\\alpha$ to control the point set construction and the dynamics. This representation allows an articulated rigid object to behave similarly to a deformable object, providing a larger action space to adjust its state and thereby easing the control optimization problem.", + "bbox": [ + 212, + 146, + 787, + 282 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Specifically, for each body of the articulated object, we sample a set of points from the ambient space near the body mesh. The point set $\\mathcal{Q}$ is constructed by concatenating all sampled points together. Each point $\\mathbf{p}_i\\in \\mathcal{Q}$ is treated as a mass point with a finite mass $\\mathbf{m}_i$ and infinitesimal volume. The dynamics of the point set consist of articulated multi-body dynamics [14, 30], as well as the mass point dynamics of each point $\\mathbf{p}_i$ . For each $\\mathbf{p}_i$ , we have:", + "bbox": [ + 212, + 282, + 589, + 434 + ], + "page_idx": 5 + }, + { + "type": "equation", + "text": "\n$$\nm _ {i} \\ddot {\\mathbf {x}} _ {i} = \\mathbf {J} _ {i} \\mathbf {u} + \\alpha \\mathbf {f} _ {i} + \\alpha \\mathbf {a} _ {i}, \\tag {1}\n$$\n", + "text_format": "latex", + "bbox": [ + 313, + 446, + 589, + 464 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "where $\\mathbf{J}_i$ represents the Jacobian mapping from the generalized velocity to the point velocity $\\dot{\\mathbf{x}}_i$ , $\\mathbf{u}$ denotes the generalized joint force, $\\mathbf{f}_i$ accounts for ex", + "bbox": [ + 212, + 476, + 589, + 523 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "ternal forces acting on $\\mathbf{p}_i$ , and $\\mathbf{a}_i \\in \\mathbb{R}^3$ represents the actuation force applied to the point $\\mathbf{p}_i$ . Consequently, the point set is controlled by a shared control in the reduced coordinate space $\\mathbf{u}$ and per-point actuation force $\\mathbf{a}_i$ .", + "bbox": [ + 212, + 523, + 785, + 566 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Parameterized spring-damper contact modeling. To ease the optimization challenges posed by contact-rich manipulations, which arise from contact constraints such as the non-penetration requirement and Coulomb friction law [3,5], as well as discontinuous dynamics involving frequent contact establishment and breaking, we propose a parameterized contact model for relaxing constraints and controlling the contact behavior. Specifically, we leverage a classical unconstrained spring-damper model [19,35,51,59,64] to model the contacts. This model allows us to flexibly adjust the contact behavior by tuning the contact threshold and the spring stiffness coefficients. Intuitively, a contact model with a high threshold and low spring stiffness presents \"soft\" behaviors, resulting in a continuous and smooth optimization space. This makes optimization through such a contact model relatively easy. Conversely, a model with a low threshold and large stiffness coefficients will produce \"stiff\" behaviors, increasing the discontinuity of the optimization space due to frequent contact establishment and breaking. However, it also becomes more physically realistic, meaning contact forces are calculated only when two objects collide, and a large force is applied to separate them if penetrations are observed, thus better satisfying the non-penetration condition. Therefore, by adjusting the contact distance threshold", + "bbox": [ + 212, + 568, + 787, + 843 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/dd5c4074daa60bddaf492a4d97d3815fc002598ecae4e5ba4392d05d6e671f52.jpg", + "image_caption": [ + "Fig.3: Point Set can flexibly adjust its states, avoid overfitting to data noise, and ease the difficulty brought by the morphology difference." + ], + "image_footnote": [], + "bbox": [ + 611, + 314, + 785, + 406 + ], + "page_idx": 5 + }, + { + "type": "page_number", + "text": "6", + "bbox": [ + 217, + 114, + 228, + 126 + ], + "page_idx": 5 + }, + { + "type": "header", + "text": "X. Liu et al.", + "bbox": [ + 271, + 114, + 354, + 127 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "and spring stiffness coefficients, we can modulate the optimizability and fidelity of the contact model. The parameter set of the contact model comprises a distance threshold $d^{c}$ and spring stiffness coefficients. Next, we will delve into the details of the contact establishment, breaking, and force calculations processes.", + "bbox": [ + 212, + 146, + 782, + 205 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Contacts are established between points in the manipulator's point set $\\mathcal{Q}$ and the object. A point $\\mathbf{p} \\in \\mathcal{Q}$ is considered to be in \"contact\" with the object if its signed distance to the object $\\operatorname{sd}(\\mathbf{p})$ is smaller than the contact distance threshold $d^{c}$ . Subsequently, the object surface point nearest to $\\mathbf{p}$ is identified as the corresponding contact point on the object, denoted as $\\mathbf{p}^{o}$ . The normal direction of the object point $\\mathbf{p}^{o}$ is then determined as the contact normal direction, denoted as $\\mathbf{n}^{o}$ . The contact force $\\mathbf{f}^{c}$ applied from the manipulator point $\\mathbf{p}$ to $\\mathbf{p}^{o}$ is calculated as follows:", + "bbox": [ + 212, + 207, + 784, + 325 + ], + "page_idx": 6 + }, + { + "type": "equation", + "text": "\n$$\n\\mathbf {f} ^ {c} = - \\left(k ^ {n} d - k ^ {d} d \\dot {d}\\right) \\mathbf {n} ^ {o}, \\tag {2}\n$$\n", + "text_format": "latex", + "bbox": [ + 419, + 327, + 784, + 343 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "where, $k^n$ represents the spring stiffness coefficient, $k^d$ denotes the damping coefficient, and $d = d^{c} - \\mathrm{sd}(\\mathbf{p})$ is always positive. To enhance the continuity of $\\mathbf{f}^c$ [64], $k^d\\ddot{d}\\dot{d}$ is used as the magnitude of the damping force, rather than $k^d\\dot{d}$ .", + "bbox": [ + 212, + 348, + 785, + 393 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Friction forces are modeled as penalty-based spring forces [3,65]. Once a point $\\mathbf{p}$ is identified as in contact with the object, with the object contact point denoted as $\\mathbf{p}^o$ , the contact pair is stored. Contact forces between them are continually calculated until the contact breaking conditions are met. In more detail, the static friction force from $\\mathbf{p}$ to $\\mathbf{p}^o$ is calculated using a spring model:", + "bbox": [ + 212, + 395, + 784, + 470 + ], + "page_idx": 6 + }, + { + "type": "equation", + "text": "\n$$\n\\mathbf {f} _ {s} ^ {f} = k ^ {f} \\mathbf {T} _ {n} (\\mathbf {p} - \\mathbf {p} ^ {o}), \\tag {3}\n$$\n", + "text_format": "latex", + "bbox": [ + 429, + 479, + 784, + 497 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "where $k^f$ is the friction spring stiffness coefficient, $\\mathbf{T}_n = \\mathbf{I} - \\mathbf{n}^o\\mathbf{n}^{oT}$ is a tangential projection operator. When the static friction satisfies $\\| \\mathbf{f}_s^f\\| \\leq \\mu \\| \\mathbf{f}^c\\|$ , $\\mathbf{f}_s^f$ is applied to the object point $\\mathbf{p}^o$ . Otherwise, the dynamic friction force is applied, and the contact breaks:", + "bbox": [ + 212, + 503, + 782, + 563 + ], + "page_idx": 6 + }, + { + "type": "equation", + "text": "\n$$\n\\mathbf {f} _ {d} ^ {f} = - \\mu \\| \\mathbf {f} _ {s} ^ {f} \\| \\frac {\\mathbf {T} _ {n} \\mathbf {v} _ {\\mathbf {p} \\leftarrow \\mathbf {p} ^ {o}}}{\\| \\mathbf {T} _ {n} \\mathbf {v} _ {\\mathbf {p} \\leftarrow \\mathbf {p} ^ {o}} \\|}, \\tag {4}\n$$\n", + "text_format": "latex", + "bbox": [ + 406, + 563, + 784, + 595 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "where $\\mathbf{v}_{\\mathbf{p}^{\\leftarrow}\\mathbf{p}^{o}}$ is the relative velocity between $\\mathbf{p}$ and $\\mathbf{p}^o$ .", + "bbox": [ + 212, + 598, + 614, + 613 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Parameterized residual physics. The analytical designs facilitate relaxation but may limit the use of highly sophisticated and realistic dynamics models, deviating from real physics. To address this, the final component of our quasi-physical simulator is a flexible neural residual physics model [1,22,41].", + "bbox": [ + 212, + 614, + 782, + 672 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Specifically, we propose to employ neural networks to learn and predict residual contact forces and friction forces based on contact-related information. For detailed residual contact force prediction, we introduce a local contact network $f_{\\psi_{\\mathrm{local}}}$ that utilizes contact information identified in the parameterized contact model and predicts residual forces between each contact pair. To address discrepancies in contact region identification between the parameterized contact model and real contact region, we also incorporate a global residual network $f_{\\psi_{\\mathrm{global}}}$ that predicts residual forces and torques applied directly to the object's center of mass. In more detail, for a given contact pair $(\\mathbf{p}, \\mathbf{p}^o)$ , the local contact network utilizes contact-related features from the local contact region, comprising geometry, per-point velocity, and per-object point normal. It then maps these", + "bbox": [ + 212, + 674, + 785, + 840 + ], + "page_idx": 6 + }, + { + "type": "header", + "text": "Quasi-Physical Simulators for Dexterous Manipulations Transfer", + "bbox": [ + 300, + 114, + 732, + 130 + ], + "page_idx": 6 + }, + { + "type": "page_number", + "text": "7", + "bbox": [ + 774, + 116, + 784, + 126 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "features to predict the residual contact force and residual friction force between the two points in the contact pair. Additionally, the global residual network incorporates contact-related information from the global contact region, including geometry, per-point velocity, and per-object point normal, as input. It then predicts a residual force and residual torque to be applied to the object's center of mass. Details such as contact region identification and network architectures are deferred to the Supp. We denote the estimizable parameters in the residual physics network as $\\psi = (\\psi_{\\mathrm{global}},\\psi_{\\mathrm{local}})$ . Through optimization of the residual physics network, we unlock the possibility of introducing highly non-linear dynamics to align our parametrized quasi-physical simulator with any realistic black-box physical simulator.", + "bbox": [ + 212, + 146, + 787, + 311 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Semi-implicit time-stepping is leveraged to make the simulation auto differentiable and easy to combine with neural networks [22].", + "bbox": [ + 212, + 313, + 785, + 344 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "3.2 Dexterous Manipulation Transfer via a Physics Curriculum", + "text_level": 1, + "bbox": [ + 214, + 357, + 750, + 372 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Building upon the family of parameterized quasi-physical simulators, we present a solution to the challenging dexterous manipulation transfer problem through a physics curriculum. This curriculum consists of a sequence of parameterized simulators, ranging from those with minimal constraints and the softest contact behavior to increasingly realistic simulators. We address the problem by transferring the manipulation demonstration to the dexterous hand within each simulator across the curriculum progressively. To elaborate further, the optimization process begins within the parameterized simulator where articulated rigid constraints are removed and the contact model is tuned to its softest level. Additionally, the residual physics networks are deactivated. This initial simulator configuration offers a friendly environment for optimization. Subsequently, the physics constraints are gradually tightened as we progress through each simulator within the curriculum. The task is solved iteratively within each simulator. After reaching the most tightened analytical model, the analytical part is fixed and residual networks are activated. The simulator is gradually optimized to approximate the dynamics in a realistic physical environment. Concurrently, the control trajectory $\\mathcal{A}$ continues to be refined in the quasi-physical simulator. Finally, we arrive at a simulator optimized to be with high fidelity and a trajectory $\\mathcal{A}$ capable of guiding the dexterous hand to accurately track the demonstration within a realistically simulated physical environment. Additionally, since object properties as well as system parameters are unknown from the kinematics-only demonstration, we set them estimizable and identify them (denoted $S$ ) together with optimizing the hand control trajectory. Next we'll illustrate this in detail.", + "bbox": [ + 212, + 386, + 787, + 734 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Transferring human demonstration via point set dynamics. To robustly transfer the human demonstration to a morphologically different dexterous robot hand in simulation and to overcome noise in the kinematic trajectory, we initially relax the articulated rigid constraints and transfer the kinematics human demonstration to the control trajectory of the point set. Specifically, the point set representation with the relaxation parameter $\\alpha$ for the dynamic human hand [8] is constructed. The shared control trajectory $\\mathcal{A}$ and per-point per-frame actions", + "bbox": [ + 212, + 734, + 787, + 839 + ], + "page_idx": 7 + }, + { + "type": "page_number", + "text": "8", + "bbox": [ + 217, + 114, + 228, + 126 + ], + "page_idx": 7 + }, + { + "type": "header", + "text": "X. Liu et al.", + "bbox": [ + 271, + 114, + 354, + 126 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "are optimized so that the resulting trajectory of the point set can manipulate the object according to the demonstration. After that, a point set with the same parameter $\\alpha$ is constructed to represent the dexterous robot hand. Subsequently, the shared control trajectory $\\mathcal{A}$ and per-point per-frame actions are optimized to track the manipulation accordingly.", + "bbox": [ + 212, + 146, + 782, + 222 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "Transferring through a contact model curriculum. After that, the articulated rigid constraint is tightened by freezing the point set parameter $\\alpha$ to zero. The following optimization starts from a parameterized simulator with the softest contact model. We then gradually tighten the contact model by adjusting its distance threshold, contact force spring stiffness, etc. By curriculum optimizing the trajectory $\\mathcal{A}$ and parameters $\\mathcal{S}$ in each of the quasi-physical simulators, we finally arrive at the control trajectory that can drive a dexterous hand to accomplish the tracking task in the parameterized simulator with the most tightened analytical model.", + "bbox": [ + 212, + 222, + 784, + 356 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "Optimizing towards a realistic physical environment. Subsequently, the residual physics network is activated and the parameterized simulator is optimized to approximate the dynamics in a realistic physical environment. We continue to optimize the hand trajectory in the quasi-physical simulator. Specifically, we leverage the successful trial in model-based human tracking literature [16,66] and iteratively optimize the control trajectory $\\mathcal{A}$ and the parameterized simulator. In more detail, the following two subproblems are iteratively solved: 1) optimizing the quasi-physical simulator to approximate the realistic dynamics, and 2) optimizing the control trajectory $\\mathcal{A}$ to complete the manipulation in the quasi-physical simulator. Gradient-based optimization is leveraged taking advantage of the differentiability of the parameterized simulator.", + "bbox": [ + 212, + 358, + 785, + 523 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "After completing the optimization, the final control trajectory is yielded by model predictive control (MPC) [18] based on the optimized parameterized simulator and the hand trajectory $\\mathcal{A}$ . Specifically, in each step, the current and the following controls in several subsequent frames are optimized to reduce the tracking error. More details are deferred to the Supp.", + "bbox": [ + 212, + 523, + 785, + 602 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "4 Experiments", + "text_level": 1, + "bbox": [ + 214, + 609, + 375, + 626 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "We conduct extensive experiments to demonstrate the effectiveness of our method. The evaluation dataset is constructed from three HOI datasets with both single-hand and bimanual manipulations (with rigid objects), with complex manipulations with non-trivial object movements, and rich and changing contacts involved (see Section 4.1). We use Shadow hand [49] and test in two simulators widely used in the embodied AI community: Bullet [9] and Isaac Gym [33]. We compare our method with both model-free approaches and model-based strategies and demonstrate the superiority of our method both quantitatively and qualitatively. We can track complex contact-rich manipulations with large object rotations, back-and-forth object movements, and changing contacts successfully in both of the two simulators, while the best-performed baseline fails (see Section 4.2, Fig. 4). On average, we boost the tracking success rate by $11\\%+$ from the previous best-performed (see Section 4.2). We make further analysis and discussions and show that the core philosophy of our work, optimizing through", + "bbox": [ + 212, + 628, + 787, + 840 + ], + "page_idx": 8 + }, + { + "type": "header", + "text": "Quasi-Physical Simulators for Dexterous Manipulations Transfer", + "bbox": [ + 300, + 114, + 730, + 130 + ], + "page_idx": 8 + }, + { + "type": "page_number", + "text": "9", + "bbox": [ + 774, + 116, + 784, + 126 + ], + "page_idx": 8 + }, + { + "type": "image", + "img_path": "images/1643c1ef6225e2e5da463e90dde2d591e5feb855311fe0a7f69da54f8a7c43d3.jpg", + "image_caption": [ + "Fig. 4: Qualitative comparisons. Please refer to our website and the accompanying video for animated results." + ], + "image_footnote": [], + "bbox": [ + 220, + 132, + 517, + 507 + ], + "page_idx": 9 + }, + { + "type": "image", + "img_path": "images/a18c830fb7a8710b8a9fb490b5e13cf1260fdd493357208fc08afa10d732cde9.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 522, + 132, + 784, + 507 + ], + "page_idx": 9 + }, + { + "type": "text", + "text": "a quasi-physics curriculum, is potentially general and can help improve the performance of a model-free baseline (see Section 4.3).", + "bbox": [ + 212, + 554, + 784, + 584 + ], + "page_idx": 9 + }, + { + "type": "text", + "text": "4.1 Experimental Settings", + "text_level": 1, + "bbox": [ + 214, + 604, + 444, + 619 + ], + "page_idx": 9 + }, + { + "type": "text", + "text": "Datasets. Our evaluation dataset is compiled from three distinct sources, namely GRAB [53], containing single-hand interactions with daily objects, TACO [32], containing humans manipulating tools, and ARCTIC [13] with bimanual manipulations. For GRAB, we randomly sample a manipulation trajectory for each object. If its manipulation is extremely simple, we additionally sample one trajectory for it. The object is not considered if its corresponding manipulation is bimanual such as binoculars, involves other body parts such as bowl, or with detailed part movements such as the game controller. The number of manipulation sequences from GRAB is 27. For TACO [32], we acquire data by contacting authors. We randomly select one sequence for each right-hand tool object. Sequences with very low quality like erroneous object motions are excluded. 14 trajectories in total are selected finally. For ARCTIC [13], we randomly select one sequence for each object from its available manipulation trajectories, resulting in 10 sequences in total. More details are deferred to the Supp.", + "bbox": [ + 212, + 628, + 787, + 840 + ], + "page_idx": 9 + }, + { + "type": "page_number", + "text": "10", + "bbox": [ + 217, + 114, + 235, + 126 + ], + "page_idx": 9 + }, + { + "type": "header", + "text": "X. Liu et al.", + "bbox": [ + 271, + 114, + 354, + 126 + ], + "page_idx": 9 + }, + { + "type": "text", + "text": "Metrics. We introduce three distinct metrics to assess the quality of object tracking, the accuracy of hand tracking, and the overall success of the tracking task: 1) Per-frame average object rotation error: $R_{\\mathrm{err}} = \\frac{1}{N}\\sum_{n=1}^{N}(1 - (\\mathbf{q}_n \\cdot \\hat{\\mathbf{q}}_n))$ , where $\\mathbf{q}_n$ is the ground-truth orientation and $\\hat{\\mathbf{q}}_n$ is the tracked result, represented in quaternion. 2) Per-frame average object translation error: $T_{\\mathrm{err}} = \\frac{1}{N}\\sum_{n=1}^{N}\\|\\mathbf{t}_n - \\hat{\\mathbf{t}}_n\\|$ , where $\\mathbf{t}$ and $\\mathbf{t}_n$ are ground-truth and tracked translations respectively. 3) Mean Per-Joint Position Error (MPJPE) = $\\frac{1}{N}\\sum_{n=1}^{N}\\|\\mathbf{J}_n - \\hat{\\mathbf{J}}_n\\|$ [20, 44, 58], where $\\mathbf{J}_n$ and $\\hat{\\mathbf{J}}_n$ are keypoints of GT human hand and the simulated robot hand respectively. We manually define the keypoints and the correspondences to the human hand keypoints for the Shadow hand. 4) Per-frame average hand Chamfer Distance: $\\mathrm{CD} = \\frac{1}{N}\\sum_{n=1}^{N}\\mathrm{Chamfer-Distance}(\\mathbf{H}_n - \\hat{\\mathbf{H}}_n)$ , for evaluating whether the Shadow hand can \"densely\" track the demonstration. 5) Success rate: a tracking is regarded as successful if the object rotation error $R_{\\mathrm{err}}$ , object translation error $T_{\\mathrm{err}}$ , and the hand tracking error MPJPE are smaller than their corresponding threshold. Three success rates are calculated using three different thresholds, namely $10^{\\circ} - 10cm - 10cm$ , $15^{\\circ} - 15cm - 15cm$ .", + "bbox": [ + 212, + 145, + 787, + 397 + ], + "page_idx": 10 + }, + { + "type": "text", + "text": "Baselines. We compare with two trends of baselines. For model-free approaches, since there is no prior work with exactly the same problem setting as us, we try to modify and improve a goal-driven rigid object manipulation method DGrasp [8] into two methods for tracking: 1) DGrasp-Base, where the method is almost kept with same with the original DGrasp. We use the first frame where the hand and the object are in contact with each other as the reference frame. Then the policy is trained to grasp the object according to the reference hand and object goal at first. After that, only the root is guided to complete the task. 2) DGraspTracking, where we divide the whole sequence into several subsequences, each of which has 10 frames, and define the end frame of the subsequence as the reference frame. Then the grasping policy is used to guide the hand and gradually track the object according to the hand and the object pose of each reference frame. We improve the DGrasp-Tracking by optimizing the policy through the quasi-physical curriculum and creating \"DGrasp-Tracking (w/ Curriculum)\" trying to improve its performance. For model-based methods, we compare with Control-VAE [66] and traditional MPC approaches. For Control-VAE, we modify its implementation for the manipulation tracking task. We additionally consider three differentiable physics models to conduct model-predictive control for solving the task. Taking the analytical model with the most tightened contact model as the base model (\"MPC (w/ base sim.)\"), we further augment it with a general state-of-the-art contact smoothing for robot manipulation [52] and create \"MPC (w/ base sim. w/ soften)\". Details of baseline models are deferred to the Supp.", + "bbox": [ + 212, + 398, + 787, + 733 + ], + "page_idx": 10 + }, + { + "type": "text", + "text": "Training and evaluation settings. The physics curriculum is composed of three stages. In the first stage, the parameter $\\alpha$ varies from 0.1 to 0.0 and the contact model stiffness is relaxed to the softest level. In the second stage, $\\alpha$ is fixed and the contact model stiffness varies from the softest version to the most tightened level gradually through eight stages. Details w.r.t. parameter settings are deferred to the Supp. In the first two stages, we alternately optimize the trajectory $\\mathcal{A}$ and parameters $S$ . In each optimization iteration, the $\\mathcal{A}$ is optimized", + "bbox": [ + 212, + 734, + 787, + 840 + ], + "page_idx": 10 + }, + { + "type": "header", + "text": "Quasi-Physical Simulators for Dexterous Manipulations Transfer", + "bbox": [ + 300, + 113, + 730, + 128 + ], + "page_idx": 10 + }, + { + "type": "page_number", + "text": "11", + "bbox": [ + 767, + 114, + 782, + 126 + ], + "page_idx": 10 + }, + { + "type": "table", + "img_path": "images/a697834157e3e46fab2fc4995163e9266548dcbd52b4c04adc09a8ec774c8151.jpg", + "table_caption": [ + "Table 1: Quantitative evaluations and comparisons to baselines. Bold red numbers for best values and italic blue values for the second best-performed ones." + ], + "table_footnote": [], + "table_body": "
SimulatorMethod\\( {R}_{\\text{err }}\\left( {{}^{ \\circ }, \\downarrow }\\right) \\)\\( {T}_{\\text{err }}\\left( {{cm}, \\downarrow }\\right) \\)MPJPE (mm,↓)CD (mm,↓)Success Rate (%,↑)
BulletModel FreeDGrasp-Base44.245.8240.5516.370/13.73/15.69
DGrasp-Tracking44.455.0437.5614.720/15.69/15.69
DGrasp-Tracking (w/ curric.)33.864.6030.4713.537.84/23.53/37.25
Model BasedControl-VAE42.452.7325.2110.940/15.68/23.53
MPC (w/ base sim.)32.563.6724.6210.800/15.68/31.37
MPC (w/ base sim. w/ soften)31.893.6328.2611.310/21.57/37.25
Ours24.211.9724.409.8527.45/37.25/58.82
Isaac GymModel FreeDGrasp-Base36.414.5650.9718.780/7.84/7.84
DGrasp-Tracking44.715.5741.5316.720/0/7.84
DGrasp-Tracking (w/ curric.)38.755.1340.0916.260/23.53/31.37
Model BasedControl-VAE35.404.6127.6313.170/13.73/29.41
MPC (w/ base sim.)37.234.7323.199.750/15.69/31.37
MPC (w/ base sim. w/ soften)36.404.4623.2710.340/9.80/23.53
Ours25.972.0825.3310.3121.57/43.14/56.86
", + "bbox": [ + 217, + 172, + 782, + 344 + ], + "page_idx": 11 + }, + { + "type": "text", + "text": "for 100 steps while $S$ is optimized for 1000 steps. In the third stage, $\\mathcal{A}$ and $\\psi$ are optimized for 256 steps in each iteration. For time-stepping, $\\mathrm{dt}$ is set to $5 \\times 10^{-4}$ in the parameterized and the target simulators. The articulated multi-body is controlled by joint motors and root velocities in the parameterized quasi-physical simulator while PD control [54] is leveraged in the target simulators.", + "bbox": [ + 212, + 361, + 784, + 436 + ], + "page_idx": 11 + }, + { + "type": "text", + "text": "4.2 Dexterous Manipulating Tracking", + "text_level": 1, + "bbox": [ + 214, + 446, + 539, + 462 + ], + "page_idx": 11 + }, + { + "type": "text", + "text": "We conducted thorough experiments in two widely used simulators [9, 33]. We treat them as realistic simulated physical environments with high fidelity and wish to track the manipulation in them. In summary, we can control a dexterous hand to complete a wide range of the manipulation tracking tasks with nontrivial object movements and changing contacts. As presented in Table 1, we can achieve significantly higher success rates calculated under three thresholds than the best-performed baseline in both tested simulators. Fig. 4 showcases qualitative examples and comparisons. Please check out our website and video for animated results.", + "bbox": [ + 212, + 464, + 784, + 598 + ], + "page_idx": 11 + }, + { + "type": "text", + "text": "Complex manipulations. For examples shown in Fig. 4, we can complete the tracking task on examples with large object re-orientations and complicated tool-using (Fig. $(a,b,c)$ ). However, DGrasp-Tracking fails to establish sufficient contact for correctly manipulating the object. In more detail, in Fig. 4(b), the bunny gradually bounced out from its hand in Bullet, while our method does not suffer from this difficulty. In Fig. 4(c), the spoon can be successfully picked up and waved back-and-forth in our method, while DGrasp-Tracking loses the track right from the start.", + "bbox": [ + 212, + 599, + 784, + 719 + ], + "page_idx": 11 + }, + { + "type": "text", + "text": "Bimanual manipulations. We are also capable of tracking bimanual manipulations. As shown in the example in Fig. 4(d), where two hands collaborate to relocate the object, DGrasp-Tracking fails to accurately track the object, while our method significantly outperforms it.", + "bbox": [ + 212, + 720, + 784, + 781 + ], + "page_idx": 11 + }, + { + "type": "text", + "text": "4.3 Further Analysis and Discussions", + "text_level": 1, + "bbox": [ + 214, + 784, + 534, + 799 + ], + "page_idx": 11 + }, + { + "type": "text", + "text": "Could model-free methods benefit from the physics curriculum? In addition to the demonstrated merits of our quasi-physical simulators, we further", + "bbox": [ + 212, + 809, + 784, + 839 + ], + "page_idx": 11 + }, + { + "type": "page_number", + "text": "12", + "bbox": [ + 217, + 114, + 235, + 126 + ], + "page_idx": 11 + }, + { + "type": "header", + "text": "X. Liu et al.", + "bbox": [ + 271, + 114, + 354, + 126 + ], + "page_idx": 11 + }, + { + "type": "table", + "img_path": "images/3b3aba97c5dd7aa347311ff5b8567d5184c9907db1ce256bac4affb1bcf0c09c.jpg", + "table_caption": [ + "Table 2: Ablation studies. Bold red numbers for best values and italic blue values for the second best-performed ones. The simulation environment is Bullet." + ], + "table_footnote": [], + "table_body": "
MethodRerr(°,↓)Terr(cm,↓)MPJPE(mm,↓)CD(mm,↓)Success Rate(%,↑)
Ours w/o Analytical Sim.44.274.3929.8412.910/13.73/25.49
Ours w/o Residual Physics33.693.8126.5710.345.88/23.53/41.18
Ours w/o Local Force NN35.982.9032.8712.440/19.61/35.29
Ours w/o Curriculum42.404.8732.6113.370/17.64/29.41
Ours w/ Curriculum II29.582.3331.6110.2911.76/27.45/50.98
Ours24.211.9724.409.8527.45/37.25/58.82
", + "bbox": [ + 215, + 172, + 782, + 268 + ], + "page_idx": 12 + }, + { + "type": "text", + "text": "explore whether model-free strategies can benefit from them. We introduce the \"DGrasp-Tracking (w/ Curriculum)\" method and compare its performance with the original DGrasp-Tracking model. As shown in Table 1 and the visual comparisons in Fig. 6, the DGrasp-Tracking model indeed benefits from a well-designed physics curriculum. For example, as illustrated in Fig. 6, the curriculum can significantly improve its performance, enabling it to nearly complete challenging tracking tasks where the original version struggles.", + "bbox": [ + 212, + 270, + 787, + 377 + ], + "page_idx": 12 + }, + { + "type": "text", + "text": "5 Ablation Study", + "text_level": 1, + "bbox": [ + 215, + 383, + 401, + 402 + ], + "page_idx": 12 + }, + { + "type": "image", + "img_path": "images/74859ef9bb99bc24c0df62b177b3f4806a4684985c9b0d55fdbe24355cb4c1c7.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 302, + 417, + 537, + 608 + ], + "page_idx": 12 + }, + { + "type": "image", + "img_path": "images/ec93451ebcec063eefb956eefd5a67143f2152306343e6a9965a064f065ccbb5.jpg", + "image_caption": [ + "Fig. 5: (a) Qualitative comparisons between our full method and the ablated models; (b) Training loss curve comparisons; (c) Tracking loss curve comparisons." + ], + "image_footnote": [], + "bbox": [ + 542, + 428, + 694, + 612 + ], + "page_idx": 12 + }, + { + "type": "image", + "img_path": "images/57fbeae52f7988b590d585056da1c3efa289c1cd94a717ddec72ad5be60d0e90.jpg", + "image_caption": [ + "Fig. 6: Visual evidence on boosting DGrasp-Tracking's performance via optimizing it through a physics curriculum." + ], + "image_footnote": [], + "bbox": [ + 222, + 640, + 787, + 763 + ], + "page_idx": 12 + }, + { + "type": "text", + "text": "We conduct a wide range of ablation studies to validate the effectiveness of some of our crucial designs, including the parameterized analytical physics", + "bbox": [ + 212, + 809, + 787, + 840 + ], + "page_idx": 12 + }, + { + "type": "header", + "text": "Quasi-Physical Simulators for Dexterous Manipulations Transfer", + "bbox": [ + 300, + 114, + 732, + 128 + ], + "page_idx": 12 + }, + { + "type": "page_number", + "text": "13", + "bbox": [ + 767, + 114, + 785, + 126 + ], + "page_idx": 12 + }, + { + "type": "text", + "text": "model, the parameterized residual physics, the role of the local force network, the necessity of introducing a physics curriculum into the optimization, and how the design on the curriculum stages affects the result.", + "bbox": [ + 215, + 146, + 784, + 191 + ], + "page_idx": 13 + }, + { + "type": "text", + "text": "Parameterized analytical model. The skeleton of the quasi-physical simulator is an analytical physics model. The intuition is that the parameterized simulator with such physical bias can be optimized towards a realistic simulator more easily than training pure neural networks for approximating. To validate this, we ablate the analytical model and use neural networks to approximate physics in Bullet directly (denoted as \"Ours w/o Analytical Sim\"). The quantitative (Table 2) and qualitative (Fig. 5) results indicate that the physical biases brought by the analytical model could help the parameterized simulator to learn better physics in the contact-rich scenario. For instance, in the example demonstrated in Fig. 5, the ablated version fails to guide the robot hand to successfully pinch the object in the second figure.", + "bbox": [ + 215, + 191, + 785, + 359 + ], + "page_idx": 13 + }, + { + "type": "text", + "text": "Parameterized residual physics. To validate the necessity of introducing residual force networks to close the gap between the physics modeled in the parameterized analytical simulator and that of a realistic simulator, we ablate the parameterized force network and create a version named \"Ours w/o Residual Physics\". Table 2 demonstrated its role in enabling the parameterized simulator to approximate realistic physics models.", + "bbox": [ + 215, + 361, + 785, + 450 + ], + "page_idx": 13 + }, + { + "type": "text", + "text": "Local residual force network. To adequately leverage state and contact-related information for predicting residual contact forces, we propose to use two types of networks: 1) a local force network for per contact pair residual forces and 2) a global network for additionally compensating. The local network is introduced for fine-grained approximation. We ablate this design and compare the result with our full model to validate this (see Fig. 5 and Table 1).", + "bbox": [ + 215, + 453, + 785, + 542 + ], + "page_idx": 13 + }, + { + "type": "text", + "text": "Optimizing through an analytical physics curriculum. We further investigate the effectiveness of the analytical curriculum design and how its design influences the result. Specifically, we create two ablated versions: 1) \"Ours w/o Curriculum\", where the optimization starts directly from the parameterized analytical model with articulated rigid constraints tightened and the stiffest contact model, and 2) \"Ours w/ Curriculum II\", where we move some stages out from the original curriculum. Table 2 and Fig. 5 demonstrate that both the curriculum and the optimization path will affect the model's performance.", + "bbox": [ + 215, + 545, + 785, + 666 + ], + "page_idx": 13 + }, + { + "type": "text", + "text": "6 Conclusion and Limitations", + "text_level": 1, + "bbox": [ + 215, + 675, + 517, + 691 + ], + "page_idx": 13 + }, + { + "type": "text", + "text": "In this work, we investigate creating better simulators for solving complex robotic tasks involving complicated dynamics where the previous best-performed optimization strategy fails. We present a family of parameterized quasi-physical simulators that can be both programmed to relax various constraints for task optimization and can be tailored to approximate realistic physics. We tackle the difficult manipulation transfer task via a physics curriculum.", + "bbox": [ + 215, + 702, + 785, + 792 + ], + "page_idx": 13 + }, + { + "type": "text", + "text": "Limitations. The method is limited by the relatively simple spring-damper model for contact constraint relaxation. Introducing delicate analytical contact models to parameterized simulators is an interesting research direction.", + "bbox": [ + 215, + 794, + 785, + 839 + ], + "page_idx": 13 + }, + { + "type": "page_number", + "text": "14", + "bbox": [ + 217, + 114, + 235, + 126 + ], + "page_idx": 13 + }, + { + "type": "header", + "text": "X. Liu et al.", + "bbox": [ + 271, + 114, + 354, + 126 + ], + "page_idx": 13 + }, + { + "type": "text", + "text": "Acknowledgement", + "text_level": 1, + "bbox": [ + 217, + 143, + 392, + 162 + ], + "page_idx": 14 + }, + { + "type": "text", + "text": "We thank Rui Chen, Yiheng Guan, and Taoran Jiang for their valuable support in providing the Allegro hardware and setting up the environment setup during the rebuttal period.", + "bbox": [ + 212, + 176, + 785, + 223 + ], + "page_idx": 14 + }, + { + "type": "text", + "text": "References", + "text_level": 1, + "bbox": [ + 217, + 248, + 321, + 263 + ], + "page_idx": 14 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "1. Ajay, A., Wu, J., Fazeli, N., Bauza, M., Kaelbling, L.P., Tenenbaum, J.B., Rodriguez, A.: Augmenting physical simulators with stochastic neural networks: Case study of planar pushing and bouncing. In: 2018 IEEE/RSJ International Conference on Intelligent Robots and Systems (IROS). pp. 3066-3073. IEEE (2018)", + "2. Akkaya, I., Andrychowicz, M., Chociej, M., Litwin, M., McGrew, B., Petron, A., Paino, A., Plappert, M., Powell, G., Ribas, R., et al.: Solving rubik's cube with a robot hand. arXiv preprint arXiv:1910.07113 (2019)", + "3. Andrews, S., Erleben, K., Ferguson, Z.: Contact and friction simulation for computer graphics. In: ACM SIGGRAPH 2022 Courses, pp. 1-172 (2022)", + "4. Andrychowicz, O.M., Baker, B., Chociej, M., Jozefowicz, R., McGrew, B., Pachocki, J., Petron, A., Plappert, M., Powell, G., Ray, A., et al.: Learning dexterous in-hand manipulation. The International Journal of Robotics Research 39(1), 3-20 (2020)", + "5. Baraff, D.: An introduction to physically based modeling: rigid body simulation ii—nonpenetration constraints. SIGGRAPH course notes pp. D31–D68 (1997)", + "6. Chen, T., Tippur, M., Wu, S., Kumar, V., Adelson, E., Agrawal, P.: Visual dexterity: In-hand reorientation of novel and complex object shapes. Science Robotics 8(84), eadc9244 (2023). https://doi.org/10.1126/scirobotics.adc9244, https://www.science.org/doi/abs/10.1126/scirobotics.adc9244", + "7. Chen, T., Xu, J., Agrawal, P.: A system for general in-hand object re-orientation. Conference on Robot Learning (2021)", + "8. Christen, S., Kocabas, M., Aksan, E., Hwangbo, J., Song, J., Hilliges, O.: D-grasp: Physically plausible dynamic grasp synthesis for hand-object interactions. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 20577-20586 (2022)", + "9. Coumans, E., Bai, Y.: Pybullet, a python module for physics simulation for games, robotics and machine learning (2016)", + "0. Deng, Y., Yu, H.X., Wu, J., Zhu, B.: Learning vortex dynamics for fluid inference and prediction. arXiv preprint arXiv:2301.11494 (2023)", + "1. Du, T.: Deep learning for physics simulation. In: ACM SIGGRAPH 2023 Courses, pp. 1-25 (2023)", + "2. Dunlavy, D.M., O'Leary, D.P.: Homotopy optimization methods for global optimization. Tech. rep., Sandia National Laboratories (SNL), Albuquerque, NM, and Livermore, CA ... (2005)", + "13. Fan, Z., Taheri, O., Tzionas, D., Kocabas, M., Kaufmann, M., Black, M.J., Hilliges, O.: ARCTIC: A dataset for dexterous bimanual hand-object manipulation. In: Proceedings IEEE Conference on Computer Vision and Pattern Recognition (CVPR) (2023)", + "14. Featherstone, R.: Rigid body dynamics algorithms (2007), https://apisemantic scholar.org/CorpusID:58437819" + ], + "bbox": [ + 218, + 281, + 785, + 839 + ], + "page_idx": 14 + }, + { + "type": "header", + "text": "Quasi-Physical Simulators for Dexterous Manipulations Transfer", + "bbox": [ + 300, + 114, + 730, + 128 + ], + "page_idx": 14 + }, + { + "type": "page_number", + "text": "15", + "bbox": [ + 767, + 116, + 784, + 126 + ], + "page_idx": 14 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "15. Freeman, C.D., Frey, E., Raichuk, A., Girgin, S., Mordatch, I., Bachem, O.: Brax-a differentiable physics engine for large scale rigid body simulation. arXiv preprint arXiv:2106.13281 (2021)", + "16. Fussell, L., Bergamin, K., Holden, D.: Supertrack: Motion tracking for physically simulated characters using supervised learning. ACM Transactions on Graphics (TOG) 40(6), 1-13 (2021)", + "17. Gao, J., Michelis, M.Y., Spielberg, A., Katzschmann, R.K.: Sim-to-real of soft robots with learned residual physics. arXiv preprint arXiv:2402.01086 (2024)", + "18. Garcia, C.E., Prett, D.M., Morari, M.: Model predictive control: Theory and practice—a survey. Automatica 25(3), 335-348 (1989)", + "19. Geilinger, M., Hahn, D., Zehnder, J., Bächer, M., Thomaszewski, B., Coros, S.: Add: Analytically differentiable dynamics for multi-body systems with frictional contact. ACM Transactions on Graphics (TOG) 39(6), 1-15 (2020)", + "20. Grandia, R., Farshidian, F., Knoop, E., Schumacher, C., Hutter, M., Bächer, M.: Doc: Differentiable optimal control for retargeting motions onto legged robots. ACM Transactions on Graphics (TOG) 42(4), 1-14 (2023)", + "21. Gupta, A., Eppner, C., Levine, S., Abbeel, P.: Learning dexterous manipulation for a soft robotic hand from human demonstrations. In: 2016 IEEE/RSJ International Conference on Intelligent Robots and Systems (IROS). pp. 3786-3793. IEEE (2016)", + "22. Heiden, E., Millard, D., Coumans, E., Sheng, Y., Sukhatme, G.S.: Neuralsim: Augmenting differentiable simulators with neural networks. In: 2021 IEEE International Conference on Robotics and Automation (ICRA). pp. 9474-9481. IEEE (2021)", + "23. Howell, T.A., Le Cleac'h, S., Kolter, J.Z., Schwager, M., Manchester, Z.: Dojo: A differentiable simulator for robotics. arXiv preprint arXiv:2203.00806 9 (2022)", + "24. Howell, T.A., Le Cleac'h, S., Singh, S., Florence, P., Manchester, Z., Sindhwani, V.: Trajectory optimization with optimization-based dynamics. IEEE Robotics and Automation Letters 7(3), 6750-6757 (2022)", + "25. Hwangbo, J., Lee, J., Hutter, M.: Per-contact iteration method for solving contact dynamics. IEEE Robotics and Automation Letters 3(2), 895-902 (2018)", + "26. Lan, L., Yang, Y., Kaufman, D., Yao, J., Li, M., Jiang, C.: Medial ipc: accelerated incremental potential contact with medial elastics. ACM Transactions on Graphics 40(4) (2021)", + "27. Li, M., Ferguson, Z., Schneider, T., Langlois, T.R., Zorin, D., Panozzo, D., Jiang, C., Kaufman, D.M.: Incremental potential contact: intersection-and inversion-free, large-deformation dynamics. ACM Trans. Graph. 39(4), 49 (2020)", + "28. Liao, S.: On the homotopy analysis method for nonlinear problems. Applied mathematics and computation 147(2), 499-513 (2004)", + "29. Lin, X., Yang, Z., Zhang, X., Zhang, Q.: Continuation path learning for homotopy optimization (2023)", + "30. Liu, C.K., Jain, S.: A quick tutorial on multibody dynamics. Online tutorial, June p. 7 (2012)", + "31. Liu, X., Pathak, D., Kitani, K.M.: Herd: Continuous human-to-robot evolution for learning from human demonstration. arXiv preprint arXiv:2212.04359 (2022)", + "32. Liu, Y., Yang, H., Si, X., Liu, L., Li, Z., Zhang, Y., Liu, Y., Yi, L.: Taco: Benchmarking generalizable bimanual tool-action-object understanding. arXiv preprint arXiv:2401.08399 (2024)", + "33. Makoviychuk, V., Wawrzyniak, L., Guo, Y., Lu, M., Storey, K., Macklin, M., Hoeller, D., Rudin, N., Allshire, A., Handa, A., et al.: Isaac gym: High performancegpu-based physics simulation for robot learning. arXiv preprint arXiv:2108.10470 (2021)" + ], + "bbox": [ + 215, + 146, + 785, + 839 + ], + "page_idx": 15 + }, + { + "type": "page_number", + "text": "16", + "bbox": [ + 217, + 114, + 235, + 126 + ], + "page_idx": 15 + }, + { + "type": "header", + "text": "X. Liu et al.", + "bbox": [ + 271, + 114, + 354, + 126 + ], + "page_idx": 15 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "34. Mandikal, P., Grauman, K.: Dexvip: Learning dexterous grasping with human hand pose priors from video. In: Conference on Robot Learning. pp. 651-661. PMLR (2022)", + "35. Marcucci, T., Gabiccini, M., Artoni, A.: A two-stage trajectory optimization strategy for articulated bodies with unscheduled contact sequences. IEEE Robotics and Automation Letters 2(1), 104-111 (2016)", + "36. Mordatch, I., Popovic, Z., Todorov, E.: Contact-invariant optimization for hand manipulation. In: Proceedings of the ACM SIGGRAPH/Eurographics symposium on computer animation. pp. 137-144 (2012)", + "37. Pang, T., Suh, H.T., Yang, L., Tedrake, R.: Global planning for contact-rich manipulation via local smoothing of quasi-dynamic contact models. IEEE Transactions on Robotics (2023)", + "38. Pang, T., Tedrake, R.: A convex quasistatic time-stepping scheme for rigid multi-body systems with contact and friction. In: 2021 IEEE International Conference on Robotics and Automation (ICRA). pp. 6614-6620. IEEE (2021)", + "39. Peng, X.B., Abbeel, P., Levine, S., Van de Panne, M.: Deepmimic: Example-guided deep reinforcement learning of physics-based character skills. ACM Transactions On Graphics (TOG) 37(4), 1-14 (2018)", + "40. Pfaff, T., Fortunato, M., Sanchez-Gonzalez, A., Battaglia, P.W.: Learning mesh-based simulation with graph networks. arXiv preprint arXiv:2010.03409 (2020)", + "41. Pfrommer, S., Halm, M., Posa, M.: Contactnets: Learning discontinuous contact dynamics with smooth, implicit representations. In: Conference on Robot Learning. pp. 2279-2291. PMLR (2021)", + "42. Qin, Y., Su, H., Wang, X.: From one hand to multiple hands: Imitation learning for dexterous manipulation from single-camera teleoperation. IEEE Robotics and Automation Letters 7(4), 10873-10881 (2022)", + "43. Qin, Y., Wu, Y.H., Liu, S., Jiang, H., Yang, R., Fu, Y., Wang, X.: Dexamv: Imitation learning for dexterous manipulation from human videos. In: European Conference on Computer Vision. pp. 570-587. Springer (2022)", + "44. Qin, Y., Yang, W., Huang, B., Van Wyk, K., Su, H., Wang, X., Chao, Y.W., Fox, D.: Anyteleop: A general vision-based dexterous robot arm-hand teleoperation system. arXiv preprint arXiv:2307.04577 (2023)", + "45. Radosavovic, I., Wang, X., Pinto, L., Malik, J.: State-only imitation learning for dexterous manipulation. In: 2021 IEEE/RSJ International Conference on Intelligent Robots and Systems (IROS). pp. 7865-7871. IEEE (2021)", + "46. Rajeswaran, A., Kumar, V., Gupta, A., Vezzani, G., Schulman, J., Todorov, E., Levine, S.: Learning complex dexterous manipulation with deep reinforcement learning and demonstrations. arXiv preprint arXiv:1709.10087 (2017)", + "47. Rusu, A.A., Colmenarejo, S.G., Gulcehre, C., Desjardins, G., Kirkpatrick, J., Pascanu, R., Mnih, V., Kavukcuoglu, K., Hadsell, R.: Policy distillation. arXiv preprint arXiv:1511.06295 (2015)", + "48. Schmeckpeper, K., Rybkin, O., Daniilidis, K., Levine, S., Finn, C.: Reinforcement learning with videos: Combining offline observations with interaction. arXiv preprint arXiv:2011.06507 (2020)", + "49. ShadowRobot: Shadowrobot dexterous hand (2005), https://www.shadowrobot.com/dexterous-hand-series/", + "50. Siahkoohi, A., Louboutin, M., Herrmann, F.J.: Neural network augmented wave-equation simulation. arXiv preprint arXiv:1910.00925 (2019)", + "51. Suh, H., Wang, Y.: Comparing effectiveness of relaxation methods for warm starting trajectory optimization through soft contact (2019)" + ], + "bbox": [ + 212, + 146, + 787, + 839 + ], + "page_idx": 16 + }, + { + "type": "header", + "text": "Quasi-Physical Simulators for Dexterous Manipulations Transfer", + "bbox": [ + 300, + 114, + 730, + 128 + ], + "page_idx": 16 + }, + { + "type": "page_number", + "text": "17", + "bbox": [ + 767, + 116, + 785, + 126 + ], + "page_idx": 16 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "52. Suh, H.J.T., Pang, T., Tedrake, R.: Bundled gradients through contact via randomized smoothing. IEEE Robotics and Automation Letters 7(2), 4000-4007 (2022)", + "53. Taheri, O., Ghorbani, N., Black, M.J., Tzionas, D.: Grab: A dataset of whole-body human grasping of objects. In: Computer Vision-ECCV 2020: 16th European Conference, Glasgow, UK, August 23-28, 2020, Proceedings, Part IV 16. pp. 581-600. Springer (2020)", + "54. Tan, K.K., Wang, Q.G., Hang, C.C.: Advances in PID control. Springer Science & Business Media (2012)", + "55. Tedrake, R., the Drake Development Team: Drake: Model-based design and verification for robotics (2019), https://drake.mit.edu", + "56. Todorov, E., Erez, T., Tassa, Y.: Mujoco: A physics engine for model-based control. In: 2012 IEEE/RSJ international conference on intelligent robots and systems. pp. 5026-5033. IEEE (2012)", + "57. Traore, R., Caseles-Dupré, H., Lesort, T., Sun, T., Díaz-Rodriguez, N., Filliat, D.: Continual reinforcement learning deployed in real-life using policy distillation and sim2real transfer. arXiv preprint arXiv:1906.04452 (2019)", + "58. Villegas, R., Ceylan, D., Hertzmann, A., Yang, J., Saito, J.: Contact-aware retargeting of skinned motion. In: Proceedings of the IEEE/CVF International Conference on Computer Vision. pp. 9720-9729 (2021)", + "59. Wang, Y., Weidner, N.J., Baxter, M.A., Hwang, Y., Kaufman, D.M., Sueda, S.: Redmax: Efficient & flexible approach for articulated dynamics. ACM Transactions on Graphics (TOG) 38(4), 1-10 (2019)", + "60. Wang, Y., Lin, J., Zeng, A., Luo, Z., Zhang, J., Zhang, L.: Physhoi: Physics-based imitation of dynamic human-object interaction. arXiv preprint arXiv:2312.04393 (2023)", + "61. Watson, L.T., Haftka, R.T.: Modern homotopy methods in optimization. Computer Methods in Applied Mechanics and Engineering 74(3), 289-305 (1989)", + "62. Wu, Y.H., Wang, J., Wang, X.: Learning generalizable dexterous manipulation from human grasp affordance. In: Conference on Robot Learning. pp. 618-629. PMLR (2023)", + "63. Xiong, S., He, X., Tong, Y., Deng, Y., Zhu, B.: Neural vortex method: from finite lagrangian particles to infinite dimensional eulerian dynamics. Computers & Fluids 258, 105811 (2023)", + "64. Xu, J., Chen, T., Zlokapa, L., Foshey, M., Matusik, W., Sueda, S., Agrawal, P.: An end-to-end differentiable framework for contact-aware robot design. arXiv preprint arXiv:2107.07501 (2021)", + "65. Yamane, K., Nakamura, Y.: Stable penalty-based model of frictional contacts. In: Proceedings 2006 IEEE International Conference on Robotics and Automation, 2006. ICRA 2006. pp. 1904-1909. IEEE (2006)", + "66. Yao, H., Song, Z., Chen, B., Liu, L.: Controlvae: Model-based learning of generative controllers for physics-based characters. ACM Transactions on Graphics (TOG) 41(6), 1-16 (2022)", + "67. Zeng, A., Song, S., Lee, J., Rodriguez, A., Funkhouser, T.: Tossingbot: Learning to throw arbitrary objects with residual physics. IEEE Transactions on Robotics 36(4), 1307-1319 (2020)", + "68. Zhang, H., Christen, S., Fan, Z., Zheng, L., Hwangbo, J., Song, J., Hilliges, O.: Artigrasp: Physically plausible synthesis of bi-manual dexterous grasping and articulation. arXiv preprint arXiv:2309.03891 (2023)", + "69. Zhang, S., Liu, B., Wang, Z., Zhao, T.: Model-based reparameterization policy gradient methods: Theory and practical algorithms. Advances in Neural Information Processing Systems 36 (2024)" + ], + "bbox": [ + 215, + 147, + 785, + 839 + ], + "page_idx": 17 + }, + { + "type": "page_number", + "text": "18", + "bbox": [ + 217, + 114, + 235, + 126 + ], + "page_idx": 17 + }, + { + "type": "header", + "text": "X. Liu et al.", + "bbox": [ + 271, + 114, + 354, + 127 + ], + "page_idx": 17 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "70. Zhang, Y., Clegg, A., Ha, S., Turk, G., Ye, Y.: Learning to transfer in-hand manipulations using a greedy shape curriculum. In: Computer Graphics Forum. vol. 42, pp. 25-36. Wiley Online Library (2023)", + "71. Zhao, W., Queralta, J.P., Westerlund, T.: Sim-to-real transfer in deep reinforcement learning for robotics: a survey. In: 2020 IEEE symposium series on computational intelligence (SSCI). pp. 737-744. IEEE (2020)", + "72. Zhuang, F., Qi, Z., Duan, K., Xi, D., Zhu, Y., Zhu, H., Xiong, H., He, Q.: A comprehensive survey on transfer learning. Proceedings of the IEEE 109(1), 43-76 (2020)" + ], + "bbox": [ + 215, + 146, + 787, + 272 + ], + "page_idx": 18 + }, + { + "type": "header", + "text": "Quasi-Physical Simulators for Dexterous Manipulations Transfer", + "bbox": [ + 300, + 114, + 730, + 128 + ], + "page_idx": 18 + }, + { + "type": "page_number", + "text": "19", + "bbox": [ + 767, + 114, + 785, + 126 + ], + "page_idx": 18 + } +] \ No newline at end of file diff --git a/2024/Parameterized Quasi-Physical Simulators for Dexterous Manipulations Transfer/ae282149-332d-459a-aebe-a693d1ac89fd_model.json b/2024/Parameterized Quasi-Physical Simulators for Dexterous Manipulations Transfer/ae282149-332d-459a-aebe-a693d1ac89fd_model.json new file mode 100644 index 0000000000000000000000000000000000000000..b3936911d395605b0c641495b1486e659d2223c9 --- /dev/null +++ b/2024/Parameterized Quasi-Physical Simulators for Dexterous Manipulations Transfer/ae282149-332d-459a-aebe-a693d1ac89fd_model.json @@ -0,0 +1,2746 @@ +[ + [ + { + "type": "title", + "bbox": [ + 0.246, + 0.141, + 0.758, + 0.187 + ], + "angle": 0, + "content": "Parameterized Quasi-Physical Simulators for Dexterous Manipulations Transfer" + }, + { + "type": "text", + "bbox": [ + 0.228, + 0.213, + 0.776, + 0.231 + ], + "angle": 0, + "content": "Xueyi Liu\\(^{1,3}\\), Kangbo Lyu\\(^{1}\\), Jieqiong Zhang\\(^{1}\\), Tao Du\\(^{1,2,3}\\), and Li Yi\\(^{1,2,3}\\)" + }, + { + "type": "text", + "bbox": [ + 0.238, + 0.241, + 0.765, + 0.271 + ], + "angle": 0, + "content": "\\(^{1}\\) Tsinghua University \\(^{2}\\) Shanghai AI Laboratory \\(^{3}\\) Shanghai Qi Zhi Institute https://meowu7.github.io/QuasiSim" + }, + { + "type": "image", + "bbox": [ + 0.248, + 0.317, + 0.313, + 0.368 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.249, + 0.369, + 0.311, + 0.417 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.313, + 0.318, + 0.373, + 0.368 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.313, + 0.369, + 0.373, + 0.418 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.375, + 0.318, + 0.436, + 0.368 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.375, + 0.369, + 0.436, + 0.418 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.436, + 0.318, + 0.497, + 0.368 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.436, + 0.369, + 0.497, + 0.418 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.501, + 0.318, + 0.563, + 0.368 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.501, + 0.369, + 0.563, + 0.418 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.565, + 0.318, + 0.627, + 0.368 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.565, + 0.369, + 0.627, + 0.418 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.627, + 0.318, + 0.688, + 0.368 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.627, + 0.369, + 0.688, + 0.418 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.688, + 0.318, + 0.749, + 0.368 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.688, + 0.369, + 0.749, + 0.418 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.25, + 0.419, + 0.375, + 0.427 + ], + "angle": 0, + "content": "(e) Result of a failed baseline" + }, + { + "type": "image", + "bbox": [ + 0.25, + 0.428, + 0.312, + 0.476 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.25, + 0.477, + 0.288, + 0.485 + ], + "angle": 0, + "content": "Timestamp" + }, + { + "type": "image", + "bbox": [ + 0.313, + 0.428, + 0.373, + 0.476 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.313, + 0.477, + 0.373, + 0.479 + ], + "angle": 0, + "content": "__________" + }, + { + "type": "image", + "bbox": [ + 0.375, + 0.428, + 0.436, + 0.476 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.375, + 0.477, + 0.436, + 0.479 + ], + "angle": 0, + "content": "\\( \\left( {x - {2x}}\\right) t - x{y}^{2} = \\left( {x - {2x}}\\right) {f}^{\\prime }t \\)" + }, + { + "type": "image", + "bbox": [ + 0.436, + 0.428, + 0.497, + 0.476 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.436, + 0.477, + 0.497, + 0.485 + ], + "angle": 0, + "content": "#" + }, + { + "type": "image_caption", + "bbox": [ + 0.503, + 0.419, + 0.746, + 0.427 + ], + "angle": 0, + "content": "(f) Improved result via optimizing through the curriculum" + }, + { + "type": "image", + "bbox": [ + 0.502, + 0.428, + 0.563, + 0.476 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.503, + 0.477, + 0.542, + 0.485 + ], + "angle": 0, + "content": "Timestamp" + }, + { + "type": "image", + "bbox": [ + 0.565, + 0.428, + 0.627, + 0.476 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.565, + 0.477, + 0.595, + 0.484 + ], + "angle": 0, + "content": "#" + }, + { + "type": "image", + "bbox": [ + 0.627, + 0.428, + 0.688, + 0.476 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.627, + 0.477, + 0.688, + 0.479 + ], + "angle": 0, + "content": "" + }, + { + "type": "image", + "bbox": [ + 0.688, + 0.428, + 0.749, + 0.476 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.214, + 0.486, + 0.788, + 0.555 + ], + "angle": 0, + "content": "Fig. 1: By optimizing through a quasi-physical simulator curriculum, we successfully transfer human demonstrations to dexterous robot hand simulations. We enable accurate tracking of complex manipulations with changing contacts (Fig. (a)), nontrivial object motions (Fig. (b)) and intricate tool-using (Fig. (c,d)). Besides, our physics curriculum can substantially improve a failed baseline (Fig. (e,f))." + }, + { + "type": "text", + "bbox": [ + 0.261, + 0.605, + 0.741, + 0.841 + ], + "angle": 0, + "content": "Abstract. We explore the dexterous manipulation transfer problem by designing simulators. The task wishes to transfer human manipulations to dexterous robot hand simulations and is inherently difficult due to its intricate, highly-constrained, and discontinuous dynamics and the need to control a dexterous hand with a DoF to accurately replicate human manipulations. Previous approaches that optimize in high-fidelity black-box simulators or a modified one with relaxed constraints only demonstrate limited capabilities or are restricted by insufficient simulation fidelity. We introduce parameterized quasi-physical simulators and a physics curriculum to overcome these limitations. The key ideas are 1) balancing between fidelity and optimizability of the simulation via a curriculum of parameterized simulators, and 2) solving the problem in each of the simulators from the curriculum, with properties ranging from high task optimizability to high fidelity. We successfully enable a dexterous hand to track complex and diverse manipulations in high-fidelity simulated environments, boosting the success rate by \\(11\\%+\\) from the best-performed baseline. The project website is available at QuasiSim." + } + ], + [ + { + "type": "page_number", + "bbox": [ + 0.218, + 0.116, + 0.23, + 0.127 + ], + "angle": 0, + "content": "2" + }, + { + "type": "header", + "bbox": [ + 0.272, + 0.115, + 0.355, + 0.127 + ], + "angle": 0, + "content": "X. Liu et al." + }, + { + "type": "title", + "bbox": [ + 0.218, + 0.145, + 0.375, + 0.161 + ], + "angle": 0, + "content": "1 Introduction" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.176, + 0.787, + 0.373 + ], + "angle": 0, + "content": "Advancing an embodied agent's capacity to interact with the world represents a significant stride toward achieving general artificial intelligence. Due to the substantial costs and potential hazards of setting up real robots to do trial and error, the standard approach for developing embodied algorithms involves learning in physical simulators [9, 15, 23, 25, 33, 56, 59] before transitioning to real-world deployment. In most cases, physical simulators are treated as black boxes, and extensive efforts have been devoted to developing learning and optimization methods for embodied skills within these black boxes. Despite the considerable progress [2, 6-8, 16, 20, 21, 31, 36, 39, 43, 46, 60, 62, 66, 68], the question like whether the simulators used are the most suitable ones is rarely discussed. In this work, we investigate this issue and illustrate how optimizing the simulator concurrently with skill acquisition can benefit a popular yet challenging task in robot manipulation - dexterous manipulation transfer." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.373, + 0.787, + 0.658 + ], + "angle": 0, + "content": "The task aims at transferring human-object manipulations to a dexterous robot hand, enabling it to physically track the reference motion of both the hand and the object (see Fig. 1). It is challenged by 1) the complex, highly constrained, non-smooth, and discontinuous dynamics with frequent contact establishment and breaking involved in the robot manipulation, 2) the requirement of precisely controlling a dexterous hand with a high DoF to densely track the manipulation at each frame, and 3) the morphology difference. Some existing works rely on high-fidelity black-box simulators, where a small difference in robot control can result in dramatically different manipulation outcomes due to abrupt contact changes, making the tracking objective highly non-smooth and hard to optimize [4,6,8,43,46]. In this way, their tasks are restricted to relatively simple goal-driven manipulations such as pouring and re-locating [8,43,46,68], in-hand re-orientation, flipping and spinning [4,6] with a fixed-root robot hand, or manipulating objects with simple geometry such as balls [36]. Other approaches attempt to improve optimization by relaxing physical constraints, with a primary focus on smoothing out contact responses [3,24,38,55,56]. However, their dynamics models may significantly deviate from real physics [38], hindering skill deployment. Consequently, we ask how to address the optimization challenge while preserving the high fidelity of the simulator." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.659, + 0.787, + 0.78 + ], + "angle": 0, + "content": "Our key insight is that a single simulator can hardly provide both high fidelity and excellent optimizability for contact-rich dexterous manipulations. Inspired by the line of homotopy methods [12, 28, 29, 61], we propose a curriculum of simulators to realize this. We start by utilizing a quasi-physical simulator to initially relax physical constraints and warm up the optimization. Subsequently, we transfer the optimization outcomes to simulators with gradually tightened physical constraints. Finally, we transition to a physically realistic simulator for skill deployment in realistic dynamics." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.78, + 0.787, + 0.84 + ], + "angle": 0, + "content": "To realize this vision, we propose a family of parameterized quasi-physical simulators for contact-rich dexterous manipulation tasks. These simulators can be customized to enhance task optimizability while can also be tailored to approximate realistic physics. The parameterized simulator represents" + } + ], + [ + { + "type": "header", + "bbox": [ + 0.303, + 0.115, + 0.731, + 0.129 + ], + "angle": 0, + "content": "Quasi-Physical Simulators for Dexterous Manipulations Transfer" + }, + { + "type": "page_number", + "bbox": [ + 0.775, + 0.117, + 0.785, + 0.127 + ], + "angle": 0, + "content": "3" + }, + { + "type": "text", + "bbox": [ + 0.218, + 0.148, + 0.786, + 0.539 + ], + "angle": 0, + "content": "an articulated multi rigid body as a parameterized point set, models contact using an unconstrained parameterized spring-damper, and compensates for unmodeled effects via parameterized residual physics. Specifically, the articulated multi-body dynamics model is relaxed as the point set dynamics model. An articulated object is relaxed into a set of points, sampled from the ambient space surrounding each body's surface mesh. The resulting dynamics model combines the original articulated dynamics with the mass-point dynamics of each individual point. Parameters are introduced to control the point set construction and the dynamics model. The contact model is softened as a parameterized spring-damper model [3, 19, 35, 38, 51] with parameters introduced to control when to calculate contacts and contact spring stiffness. The residual physics network compensate for unmodeled effects from the analytical modeling [22]. The parameterized simulator can be programmed for high optimizability by relaxing constraints in the analytical model and can be tailored to approximate realistic physics by learning excellent residual physics. We demonstrate that the challenging dexterous manipulation transfer task can be effectively addressed through curriculum optimization using a series of parameterized physical simulators. Initially, both articulated rigid constraints and the contact model stiffness are relaxed in the simulator. It may not reflect physical realism but provides a good environment where the manipulation transfer problem can be solved easily. Subsequently, the articulated rigid constraints and the contact model are gradually tightened. Task-solving proceeds iteratively within each simulator in the curriculum. Finally, the parameterized simulator is optimized to approximate realistic physics. Task optimization continues, yielding a dexterous hand trajectory capable of executing the manipulation in environments with realistic physics." + }, + { + "type": "text", + "bbox": [ + 0.218, + 0.551, + 0.786, + 0.822 + ], + "angle": 0, + "content": "We demonstrate the superiority of our method and compare it with previous model-free and model-based methods on challenging manipulation sequences from three datasets, describing single-hand or bimanual manipulations with daily objects or using tools. We conduct dexterous manipulation transfer on two widely used simulators, namely Bullet [9] and Isaac Gym [33] to demonstrate the generality and the efficacy of our method and the capability of our quasi-physical simulator to approximate the unknown black-box physics model in the contact-rich manipulation scenario (Fig. 1). We can track complex manipulations involving non-trivial object motions such as large rotations and complicated tool-using such as using a spoon to bring the water back and forth. Our approach successfully surpasses the previous best-performed method both quantitatively and qualitatively, achieving more than \\(11\\%\\) success rate than the previous best-performed method. Besides, optimizing through the physics curriculum can significantly enhance the performance of previously under-performed RL-based methods, almost completing the tracking problem from failure, as demonstrated in Fig. 1. This indicates the universality of our approach to embodied AI through optimization via a physics curriculum. Thorough ablations are conducted to validate the efficacy of our designs." + }, + { + "type": "text", + "bbox": [ + 0.241, + 0.834, + 0.478, + 0.848 + ], + "angle": 0, + "content": "Our contributions are three-fold:" + } + ], + [ + { + "type": "page_number", + "bbox": [ + 0.218, + 0.116, + 0.23, + 0.127 + ], + "angle": 0, + "content": "4" + }, + { + "type": "header", + "bbox": [ + 0.272, + 0.115, + 0.355, + 0.127 + ], + "angle": 0, + "content": "X. Liu et al." + }, + { + "type": "text", + "bbox": [ + 0.225, + 0.147, + 0.784, + 0.191 + ], + "angle": 0, + "content": "- We introduce a family of parameterized quasi-physical simulators that can be configured to relax various physical constraints, facilitating skill optimization, and can also be tailored to achieve high simulation fidelity." + }, + { + "type": "text", + "bbox": [ + 0.226, + 0.192, + 0.785, + 0.236 + ], + "angle": 0, + "content": "- We present a quasi-physics curriculum along with a corresponding optimization method to address the challenging dexterous manipulation transfer problem." + }, + { + "type": "text", + "bbox": [ + 0.226, + 0.238, + 0.785, + 0.282 + ], + "angle": 0, + "content": "- Extensive experiments demonstrate the effectiveness of our method in transferring complex manipulations, including non-trivial object motions and changing contacts, to a dexterous robot hand in simulation." + }, + { + "type": "list", + "bbox": [ + 0.225, + 0.147, + 0.785, + 0.282 + ], + "angle": 0, + "content": null + }, + { + "type": "title", + "bbox": [ + 0.217, + 0.299, + 0.396, + 0.315 + ], + "angle": 0, + "content": "2 Related Works" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.321, + 0.786, + 0.501 + ], + "angle": 0, + "content": "Dexterous manipulation transfer. Transferring human manipulations to dexterous robot-hand simulations is an important topic in robot skill acquisition [8,21,31,43,60,62,68,70]. Most approaches treat the simulator as black-box physics models and try to learn skills directly from that [4,6,8,43,46]. However, their demonstrated capabilities are restricted to relatively simple tasks. Another trend of work tries to relax the physics model [37,38] to create a better environment for task optimization. However, due to the disparity between their modeling approach and realistic physics, successful trials are typically demonstrated only in their simulators, which can hardly complete the task under physically realistic dynamics. In this work, we introduce various parameterized analytical relaxations to improve the task optimizability while compensating for unmodeled effects via residual physics networks so the fidelity would not be sacrificed." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.503, + 0.787, + 0.622 + ], + "angle": 0, + "content": "Learning for simulation. Analytical methods can hardly approximate an extremely realistic physical world despite lots of smart and tremendous efforts made in developing numerical algorithms [19, 23, 26, 27]. Recently, data-driven approaches have attracted lots of interest for their high efficiency and strong approximation ability [10, 11, 22, 40, 41, 50, 63]. Special network designs are proposed to learn the contact behaviour [22, 41]. We in this work propose to leverage an analytical-neural hybrid approach and carefully design network modules for approximating residual contact forces in the contact-rich manipulation scenario." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.623, + 0.786, + 0.775 + ], + "angle": 0, + "content": "Sim-to-Sim and Sim-to-Real transfer. The field of robot manipulation continues to face challenges in the areas of Sim2Sim and Sim2Real transferability [71]. Considering the modeling gaps, the optimal strategy learned in a specific simulator is difficult to transfer to a different simulator or the real world. Therefore, many techniques for solving the problem have been proposed, including imitation learning [34,42,43,45,46,48], transfer learning [72], distillation [47,57], residual physics [17,67], and efforts on bridging the gap from the dynamics model aspect [22,69]. Our parameterized simulators learn residual physics involved in contact-rich robot manipulations. By combining an analytical base with residual networks, we showcase their ability to approximate realistic physics." + }, + { + "type": "title", + "bbox": [ + 0.216, + 0.788, + 0.331, + 0.803 + ], + "angle": 0, + "content": "3 Method" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.81, + 0.787, + 0.842 + ], + "angle": 0, + "content": "Given a human manipulation demonstration, composed of a human hand mesh trajectory and an object pose trajectory \\(\\{\\mathcal{H},\\mathcal{O}\\}\\), the goal is transferring the" + } + ], + [ + { + "type": "header", + "bbox": [ + 0.302, + 0.115, + 0.733, + 0.129 + ], + "angle": 0, + "content": "Quasi-Physical Simulators for Dexterous Manipulations Transfer" + }, + { + "type": "page_number", + "bbox": [ + 0.775, + 0.117, + 0.785, + 0.127 + ], + "angle": 0, + "content": "5" + }, + { + "type": "image_caption", + "bbox": [ + 0.223, + 0.135, + 0.391, + 0.145 + ], + "angle": 0, + "content": "Parameterized Quasi-Physical Simulator" + }, + { + "type": "image", + "bbox": [ + 0.22, + 0.145, + 0.782, + 0.266 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.214, + 0.274, + 0.788, + 0.345 + ], + "angle": 0, + "content": "Fig. 2: The parameterized quasi-physical simulator relaxes the articulated multi rigid body dynamics as the parameterized point set dynamics, controls the contact behavior via an unconstrained parameterized spring-damper contact model, and compensates for unmodeled effects via parameterized residual physics networks. We tackle the difficult dexterous manipulation transfer problem via a physics curriculum." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.353, + 0.788, + 0.459 + ], + "angle": 0, + "content": "demonstration to a dexterous robot hand in simulation. Formally, we aim to optimize a control trajectory \\(\\mathcal{A}\\) that drives the dexterous hand to manipulate the object in a realistic simulated environment so that the resulting hand trajectory \\(\\hat{\\mathcal{H}}\\) and the object trajectory \\(\\hat{\\mathcal{O}}\\) are close to the reference motion \\(\\{\\mathcal{H},\\mathcal{O}\\}\\). The problem is challenged by difficulties from the highly constrained, discontinuous, and non-smooth dynamics, the requirement of controlling a high DoF dexterous hand for tracking, and the morphology difference." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.459, + 0.788, + 0.565 + ], + "angle": 0, + "content": "Our method comprises two key designs to tackle the challenges: 1) a family of parameterized quasi-physical simulators, which can be programmed to enhance the optimizability of contact-rich dexterous manipulation tasks and can also be tailored to approximate realistic physics (Section 3.1), and 2) a physics curriculum that carefully adjusts the parameters of a line of quasi-physical simulators and a strategy that solves the difficult dexterous manipulation transfer task by addressing it within each simulator in the curriculum (Section 3.2)." + }, + { + "type": "title", + "bbox": [ + 0.215, + 0.588, + 0.603, + 0.603 + ], + "angle": 0, + "content": "3.1 Parameterized Quasi-Physical Simulators" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.614, + 0.794, + 0.734 + ], + "angle": 0, + "content": "Our quasi-physical simulator represents an articulated multi-body, i.e., the robotic dexterous hand, as a point set. The object is represented as a signed distance field. The base of the simulator is in an analytical form leveraging an unconstrained spring-damper contact model. Parameters are introduced to control the analytical relaxations on the articulated rigid constraints and the softness of the contact model. Additionally, neural networks are introduced to compensate for unmodeled effects beyond the analytical framework. We will elaborate on each of these design aspects below." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.735, + 0.788, + 0.841 + ], + "angle": 0, + "content": "Parameterized point set dynamics. Articulated multi-body represented in the reduced coordinate system [19,59] may require a large change in joint states to achieve a small adjustment in the Euclidean space. Moving the end effector from one point to a nearby point may require adjusting all joint states (Fig. 3). Besides, transferring the hand trajectory to a morphologically different hand requires correspondences to make the resulting trajectory close to the original one. Defining correspondences in the reduced coordinate or via sparse correspon" + } + ], + [ + { + "type": "page_number", + "bbox": [ + 0.218, + 0.116, + 0.23, + 0.127 + ], + "angle": 0, + "content": "6" + }, + { + "type": "header", + "bbox": [ + 0.272, + 0.115, + 0.356, + 0.128 + ], + "angle": 0, + "content": "X. Liu et al." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.147, + 0.788, + 0.283 + ], + "angle": 0, + "content": "dences will make the result suffer from noise in the data, leading to unwanted results finally (Fig. 3). Hence, we propose relaxing an articulated multi-rigid body into a mass-point set sampled from the ambient space surrounding each body. Each point is considered attached to the body from which it is sampled and is capable of both self-actuation and actuation via joint motors. We introduce a parameter \\(\\alpha\\) to control the point set construction and the dynamics. This representation allows an articulated rigid object to behave similarly to a deformable object, providing a larger action space to adjust its state and thereby easing the control optimization problem." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.284, + 0.591, + 0.435 + ], + "angle": 0, + "content": "Specifically, for each body of the articulated object, we sample a set of points from the ambient space near the body mesh. The point set \\(\\mathcal{Q}\\) is constructed by concatenating all sampled points together. Each point \\(\\mathbf{p}_i\\in \\mathcal{Q}\\) is treated as a mass point with a finite mass \\(\\mathbf{m}_i\\) and infinitesimal volume. The dynamics of the point set consist of articulated multi-body dynamics [14, 30], as well as the mass point dynamics of each point \\(\\mathbf{p}_i\\). For each \\(\\mathbf{p}_i\\), we have:" + }, + { + "type": "equation", + "bbox": [ + 0.314, + 0.448, + 0.59, + 0.465 + ], + "angle": 0, + "content": "\\[\nm _ {i} \\ddot {\\mathbf {x}} _ {i} = \\mathbf {J} _ {i} \\mathbf {u} + \\alpha \\mathbf {f} _ {i} + \\alpha \\mathbf {a} _ {i}, \\tag {1}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.477, + 0.591, + 0.524 + ], + "angle": 0, + "content": "where \\(\\mathbf{J}_i\\) represents the Jacobian mapping from the generalized velocity to the point velocity \\(\\dot{\\mathbf{x}}_i\\), \\(\\mathbf{u}\\) denotes the generalized joint force, \\(\\mathbf{f}_i\\) accounts for ex" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.524, + 0.787, + 0.568 + ], + "angle": 0, + "content": "ternal forces acting on \\(\\mathbf{p}_i\\), and \\(\\mathbf{a}_i \\in \\mathbb{R}^3\\) represents the actuation force applied to the point \\(\\mathbf{p}_i\\). Consequently, the point set is controlled by a shared control in the reduced coordinate space \\(\\mathbf{u}\\) and per-point actuation force \\(\\mathbf{a}_i\\)." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.569, + 0.789, + 0.844 + ], + "angle": 0, + "content": "Parameterized spring-damper contact modeling. To ease the optimization challenges posed by contact-rich manipulations, which arise from contact constraints such as the non-penetration requirement and Coulomb friction law [3,5], as well as discontinuous dynamics involving frequent contact establishment and breaking, we propose a parameterized contact model for relaxing constraints and controlling the contact behavior. Specifically, we leverage a classical unconstrained spring-damper model [19,35,51,59,64] to model the contacts. This model allows us to flexibly adjust the contact behavior by tuning the contact threshold and the spring stiffness coefficients. Intuitively, a contact model with a high threshold and low spring stiffness presents \"soft\" behaviors, resulting in a continuous and smooth optimization space. This makes optimization through such a contact model relatively easy. Conversely, a model with a low threshold and large stiffness coefficients will produce \"stiff\" behaviors, increasing the discontinuity of the optimization space due to frequent contact establishment and breaking. However, it also becomes more physically realistic, meaning contact forces are calculated only when two objects collide, and a large force is applied to separate them if penetrations are observed, thus better satisfying the non-penetration condition. Therefore, by adjusting the contact distance threshold" + }, + { + "type": "image", + "bbox": [ + 0.612, + 0.315, + 0.787, + 0.407 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.6, + 0.419, + 0.788, + 0.502 + ], + "angle": 0, + "content": "Fig.3: Point Set can flexibly adjust its states, avoid overfitting to data noise, and ease the difficulty brought by the morphology difference." + } + ], + [ + { + "type": "header", + "bbox": [ + 0.302, + 0.115, + 0.733, + 0.131 + ], + "angle": 0, + "content": "Quasi-Physical Simulators for Dexterous Manipulations Transfer" + }, + { + "type": "page_number", + "bbox": [ + 0.775, + 0.117, + 0.785, + 0.127 + ], + "angle": 0, + "content": "7" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.147, + 0.784, + 0.207 + ], + "angle": 0, + "content": "and spring stiffness coefficients, we can modulate the optimizability and fidelity of the contact model. The parameter set of the contact model comprises a distance threshold \\( d^{c} \\) and spring stiffness coefficients. Next, we will delve into the details of the contact establishment, breaking, and force calculations processes." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.208, + 0.785, + 0.326 + ], + "angle": 0, + "content": "Contacts are established between points in the manipulator's point set \\(\\mathcal{Q}\\) and the object. A point \\(\\mathbf{p} \\in \\mathcal{Q}\\) is considered to be in \"contact\" with the object if its signed distance to the object \\(\\operatorname{sd}(\\mathbf{p})\\) is smaller than the contact distance threshold \\(d^{c}\\). Subsequently, the object surface point nearest to \\(\\mathbf{p}\\) is identified as the corresponding contact point on the object, denoted as \\(\\mathbf{p}^{o}\\). The normal direction of the object point \\(\\mathbf{p}^{o}\\) is then determined as the contact normal direction, denoted as \\(\\mathbf{n}^{o}\\). The contact force \\(\\mathbf{f}^{c}\\) applied from the manipulator point \\(\\mathbf{p}\\) to \\(\\mathbf{p}^{o}\\) is calculated as follows:" + }, + { + "type": "equation", + "bbox": [ + 0.42, + 0.328, + 0.785, + 0.344 + ], + "angle": 0, + "content": "\\[\n\\mathbf {f} ^ {c} = - \\left(k ^ {n} d - k ^ {d} d \\dot {d}\\right) \\mathbf {n} ^ {o}, \\tag {2}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.349, + 0.787, + 0.395 + ], + "angle": 0, + "content": "where, \\( k^n \\) represents the spring stiffness coefficient, \\( k^d \\) denotes the damping coefficient, and \\( d = d^{c} - \\mathrm{sd}(\\mathbf{p}) \\) is always positive. To enhance the continuity of \\( \\mathbf{f}^c \\) [64], \\( k^d\\ddot{d}\\dot{d} \\) is used as the magnitude of the damping force, rather than \\( k^d\\dot{d} \\)." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.396, + 0.785, + 0.471 + ], + "angle": 0, + "content": "Friction forces are modeled as penalty-based spring forces [3,65]. Once a point \\(\\mathbf{p}\\) is identified as in contact with the object, with the object contact point denoted as \\(\\mathbf{p}^o\\), the contact pair is stored. Contact forces between them are continually calculated until the contact breaking conditions are met. In more detail, the static friction force from \\(\\mathbf{p}\\) to \\(\\mathbf{p}^o\\) is calculated using a spring model:" + }, + { + "type": "equation", + "bbox": [ + 0.43, + 0.48, + 0.785, + 0.498 + ], + "angle": 0, + "content": "\\[\n\\mathbf {f} _ {s} ^ {f} = k ^ {f} \\mathbf {T} _ {n} (\\mathbf {p} - \\mathbf {p} ^ {o}), \\tag {3}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.505, + 0.784, + 0.564 + ], + "angle": 0, + "content": "where \\( k^f \\) is the friction spring stiffness coefficient, \\( \\mathbf{T}_n = \\mathbf{I} - \\mathbf{n}^o\\mathbf{n}^{oT} \\) is a tangential projection operator. When the static friction satisfies \\( \\| \\mathbf{f}_s^f\\| \\leq \\mu \\| \\mathbf{f}^c\\| \\), \\( \\mathbf{f}_s^f \\) is applied to the object point \\( \\mathbf{p}^o \\). Otherwise, the dynamic friction force is applied, and the contact breaks:" + }, + { + "type": "equation", + "bbox": [ + 0.408, + 0.564, + 0.785, + 0.597 + ], + "angle": 0, + "content": "\\[\n\\mathbf {f} _ {d} ^ {f} = - \\mu \\| \\mathbf {f} _ {s} ^ {f} \\| \\frac {\\mathbf {T} _ {n} \\mathbf {v} _ {\\mathbf {p} \\leftarrow \\mathbf {p} ^ {o}}}{\\| \\mathbf {T} _ {n} \\mathbf {v} _ {\\mathbf {p} \\leftarrow \\mathbf {p} ^ {o}} \\|}, \\tag {4}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.599, + 0.615, + 0.614 + ], + "angle": 0, + "content": "where \\(\\mathbf{v}_{\\mathbf{p}^{\\leftarrow}\\mathbf{p}^{o}}\\) is the relative velocity between \\(\\mathbf{p}\\) and \\(\\mathbf{p}^o\\)." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.615, + 0.784, + 0.674 + ], + "angle": 0, + "content": "Parameterized residual physics. The analytical designs facilitate relaxation but may limit the use of highly sophisticated and realistic dynamics models, deviating from real physics. To address this, the final component of our quasi-physical simulator is a flexible neural residual physics model [1,22,41]." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.675, + 0.786, + 0.841 + ], + "angle": 0, + "content": "Specifically, we propose to employ neural networks to learn and predict residual contact forces and friction forces based on contact-related information. For detailed residual contact force prediction, we introduce a local contact network \\( f_{\\psi_{\\mathrm{local}}} \\) that utilizes contact information identified in the parameterized contact model and predicts residual forces between each contact pair. To address discrepancies in contact region identification between the parameterized contact model and real contact region, we also incorporate a global residual network \\( f_{\\psi_{\\mathrm{global}}} \\) that predicts residual forces and torques applied directly to the object's center of mass. In more detail, for a given contact pair \\( (\\mathbf{p}, \\mathbf{p}^o) \\), the local contact network utilizes contact-related features from the local contact region, comprising geometry, per-point velocity, and per-object point normal. It then maps these" + } + ], + [ + { + "type": "page_number", + "bbox": [ + 0.218, + 0.116, + 0.23, + 0.127 + ], + "angle": 0, + "content": "8" + }, + { + "type": "header", + "bbox": [ + 0.272, + 0.115, + 0.355, + 0.127 + ], + "angle": 0, + "content": "X. Liu et al." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.147, + 0.788, + 0.312 + ], + "angle": 0, + "content": "features to predict the residual contact force and residual friction force between the two points in the contact pair. Additionally, the global residual network incorporates contact-related information from the global contact region, including geometry, per-point velocity, and per-object point normal, as input. It then predicts a residual force and residual torque to be applied to the object's center of mass. Details such as contact region identification and network architectures are deferred to the Supp. We denote the estimizable parameters in the residual physics network as \\(\\psi = (\\psi_{\\mathrm{global}},\\psi_{\\mathrm{local}})\\). Through optimization of the residual physics network, we unlock the possibility of introducing highly non-linear dynamics to align our parametrized quasi-physical simulator with any realistic black-box physical simulator." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.314, + 0.786, + 0.345 + ], + "angle": 0, + "content": "Semi-implicit time-stepping is leveraged to make the simulation auto differentiable and easy to combine with neural networks [22]." + }, + { + "type": "title", + "bbox": [ + 0.215, + 0.358, + 0.75, + 0.373 + ], + "angle": 0, + "content": "3.2 Dexterous Manipulation Transfer via a Physics Curriculum" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.387, + 0.789, + 0.735 + ], + "angle": 0, + "content": "Building upon the family of parameterized quasi-physical simulators, we present a solution to the challenging dexterous manipulation transfer problem through a physics curriculum. This curriculum consists of a sequence of parameterized simulators, ranging from those with minimal constraints and the softest contact behavior to increasingly realistic simulators. We address the problem by transferring the manipulation demonstration to the dexterous hand within each simulator across the curriculum progressively. To elaborate further, the optimization process begins within the parameterized simulator where articulated rigid constraints are removed and the contact model is tuned to its softest level. Additionally, the residual physics networks are deactivated. This initial simulator configuration offers a friendly environment for optimization. Subsequently, the physics constraints are gradually tightened as we progress through each simulator within the curriculum. The task is solved iteratively within each simulator. After reaching the most tightened analytical model, the analytical part is fixed and residual networks are activated. The simulator is gradually optimized to approximate the dynamics in a realistic physical environment. Concurrently, the control trajectory \\(\\mathcal{A}\\) continues to be refined in the quasi-physical simulator. Finally, we arrive at a simulator optimized to be with high fidelity and a trajectory \\(\\mathcal{A}\\) capable of guiding the dexterous hand to accurately track the demonstration within a realistically simulated physical environment. Additionally, since object properties as well as system parameters are unknown from the kinematics-only demonstration, we set them estimizable and identify them (denoted \\(S\\)) together with optimizing the hand control trajectory. Next we'll illustrate this in detail." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.735, + 0.788, + 0.84 + ], + "angle": 0, + "content": "Transferring human demonstration via point set dynamics. To robustly transfer the human demonstration to a morphologically different dexterous robot hand in simulation and to overcome noise in the kinematic trajectory, we initially relax the articulated rigid constraints and transfer the kinematics human demonstration to the control trajectory of the point set. Specifically, the point set representation with the relaxation parameter \\(\\alpha\\) for the dynamic human hand [8] is constructed. The shared control trajectory \\(\\mathcal{A}\\) and per-point per-frame actions" + } + ], + [ + { + "type": "header", + "bbox": [ + 0.302, + 0.115, + 0.732, + 0.131 + ], + "angle": 0, + "content": "Quasi-Physical Simulators for Dexterous Manipulations Transfer" + }, + { + "type": "page_number", + "bbox": [ + 0.775, + 0.117, + 0.785, + 0.127 + ], + "angle": 0, + "content": "9" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.147, + 0.784, + 0.223 + ], + "angle": 0, + "content": "are optimized so that the resulting trajectory of the point set can manipulate the object according to the demonstration. After that, a point set with the same parameter \\(\\alpha\\) is constructed to represent the dexterous robot hand. Subsequently, the shared control trajectory \\(\\mathcal{A}\\) and per-point per-frame actions are optimized to track the manipulation accordingly." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.223, + 0.785, + 0.357 + ], + "angle": 0, + "content": "Transferring through a contact model curriculum. After that, the articulated rigid constraint is tightened by freezing the point set parameter \\(\\alpha\\) to zero. The following optimization starts from a parameterized simulator with the softest contact model. We then gradually tighten the contact model by adjusting its distance threshold, contact force spring stiffness, etc. By curriculum optimizing the trajectory \\(\\mathcal{A}\\) and parameters \\(\\mathcal{S}\\) in each of the quasi-physical simulators, we finally arrive at the control trajectory that can drive a dexterous hand to accomplish the tracking task in the parameterized simulator with the most tightened analytical model." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.359, + 0.787, + 0.525 + ], + "angle": 0, + "content": "Optimizing towards a realistic physical environment. Subsequently, the residual physics network is activated and the parameterized simulator is optimized to approximate the dynamics in a realistic physical environment. We continue to optimize the hand trajectory in the quasi-physical simulator. Specifically, we leverage the successful trial in model-based human tracking literature [16,66] and iteratively optimize the control trajectory \\(\\mathcal{A}\\) and the parameterized simulator. In more detail, the following two subproblems are iteratively solved: 1) optimizing the quasi-physical simulator to approximate the realistic dynamics, and 2) optimizing the control trajectory \\(\\mathcal{A}\\) to complete the manipulation in the quasi-physical simulator. Gradient-based optimization is leveraged taking advantage of the differentiability of the parameterized simulator." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.525, + 0.787, + 0.603 + ], + "angle": 0, + "content": "After completing the optimization, the final control trajectory is yielded by model predictive control (MPC) [18] based on the optimized parameterized simulator and the hand trajectory \\(\\mathcal{A}\\). Specifically, in each step, the current and the following controls in several subsequent frames are optimized to reduce the tracking error. More details are deferred to the Supp." + }, + { + "type": "title", + "bbox": [ + 0.215, + 0.611, + 0.377, + 0.627 + ], + "angle": 0, + "content": "4 Experiments" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.629, + 0.788, + 0.842 + ], + "angle": 0, + "content": "We conduct extensive experiments to demonstrate the effectiveness of our method. The evaluation dataset is constructed from three HOI datasets with both single-hand and bimanual manipulations (with rigid objects), with complex manipulations with non-trivial object movements, and rich and changing contacts involved (see Section 4.1). We use Shadow hand [49] and test in two simulators widely used in the embodied AI community: Bullet [9] and Isaac Gym [33]. We compare our method with both model-free approaches and model-based strategies and demonstrate the superiority of our method both quantitatively and qualitatively. We can track complex contact-rich manipulations with large object rotations, back-and-forth object movements, and changing contacts successfully in both of the two simulators, while the best-performed baseline fails (see Section 4.2, Fig. 4). On average, we boost the tracking success rate by \\(11\\%+\\) from the previous best-performed (see Section 4.2). We make further analysis and discussions and show that the core philosophy of our work, optimizing through" + } + ], + [ + { + "type": "page_number", + "bbox": [ + 0.218, + 0.116, + 0.236, + 0.127 + ], + "angle": 0, + "content": "10" + }, + { + "type": "header", + "bbox": [ + 0.272, + 0.115, + 0.356, + 0.127 + ], + "angle": 0, + "content": "X. Liu et al." + }, + { + "type": "image", + "bbox": [ + 0.222, + 0.133, + 0.518, + 0.508 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.523, + 0.133, + 0.785, + 0.508 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.214, + 0.511, + 0.785, + 0.54 + ], + "angle": 0, + "content": "Fig. 4: Qualitative comparisons. Please refer to our website and the accompanying video for animated results." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.555, + 0.785, + 0.585 + ], + "angle": 0, + "content": "a quasi-physics curriculum, is potentially general and can help improve the performance of a model-free baseline (see Section 4.3)." + }, + { + "type": "title", + "bbox": [ + 0.215, + 0.606, + 0.446, + 0.621 + ], + "angle": 0, + "content": "4.1 Experimental Settings" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.629, + 0.788, + 0.842 + ], + "angle": 0, + "content": "Datasets. Our evaluation dataset is compiled from three distinct sources, namely GRAB [53], containing single-hand interactions with daily objects, TACO [32], containing humans manipulating tools, and ARCTIC [13] with bimanual manipulations. For GRAB, we randomly sample a manipulation trajectory for each object. If its manipulation is extremely simple, we additionally sample one trajectory for it. The object is not considered if its corresponding manipulation is bimanual such as binoculars, involves other body parts such as bowl, or with detailed part movements such as the game controller. The number of manipulation sequences from GRAB is 27. For TACO [32], we acquire data by contacting authors. We randomly select one sequence for each right-hand tool object. Sequences with very low quality like erroneous object motions are excluded. 14 trajectories in total are selected finally. For ARCTIC [13], we randomly select one sequence for each object from its available manipulation trajectories, resulting in 10 sequences in total. More details are deferred to the Supp." + } + ], + [ + { + "type": "header", + "bbox": [ + 0.302, + 0.114, + 0.732, + 0.129 + ], + "angle": 0, + "content": "Quasi-Physical Simulators for Dexterous Manipulations Transfer" + }, + { + "type": "page_number", + "bbox": [ + 0.768, + 0.116, + 0.784, + 0.127 + ], + "angle": 0, + "content": "11" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.146, + 0.788, + 0.398 + ], + "angle": 0, + "content": "Metrics. We introduce three distinct metrics to assess the quality of object tracking, the accuracy of hand tracking, and the overall success of the tracking task: 1) Per-frame average object rotation error: \\( R_{\\mathrm{err}} = \\frac{1}{N}\\sum_{n=1}^{N}(1 - (\\mathbf{q}_n \\cdot \\hat{\\mathbf{q}}_n)) \\), where \\( \\mathbf{q}_n \\) is the ground-truth orientation and \\( \\hat{\\mathbf{q}}_n \\) is the tracked result, represented in quaternion. 2) Per-frame average object translation error: \\( T_{\\mathrm{err}} = \\frac{1}{N}\\sum_{n=1}^{N}\\|\\mathbf{t}_n - \\hat{\\mathbf{t}}_n\\| \\), where \\( \\mathbf{t} \\) and \\( \\mathbf{t}_n \\) are ground-truth and tracked translations respectively. 3) Mean Per-Joint Position Error (MPJPE) = \\( \\frac{1}{N}\\sum_{n=1}^{N}\\|\\mathbf{J}_n - \\hat{\\mathbf{J}}_n\\| \\) [20, 44, 58], where \\( \\mathbf{J}_n \\) and \\( \\hat{\\mathbf{J}}_n \\) are keypoints of GT human hand and the simulated robot hand respectively. We manually define the keypoints and the correspondences to the human hand keypoints for the Shadow hand. 4) Per-frame average hand Chamfer Distance: \\( \\mathrm{CD} = \\frac{1}{N}\\sum_{n=1}^{N}\\mathrm{Chamfer-Distance}(\\mathbf{H}_n - \\hat{\\mathbf{H}}_n) \\), for evaluating whether the Shadow hand can \"densely\" track the demonstration. 5) Success rate: a tracking is regarded as successful if the object rotation error \\( R_{\\mathrm{err}} \\), object translation error \\( T_{\\mathrm{err}} \\), and the hand tracking error MPJPE are smaller than their corresponding threshold. Three success rates are calculated using three different thresholds, namely \\( 10^{\\circ} - 10cm - 10cm \\), \\( 15^{\\circ} - 15cm - 15cm \\)." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.4, + 0.788, + 0.734 + ], + "angle": 0, + "content": "Baselines. We compare with two trends of baselines. For model-free approaches, since there is no prior work with exactly the same problem setting as us, we try to modify and improve a goal-driven rigid object manipulation method DGrasp [8] into two methods for tracking: 1) DGrasp-Base, where the method is almost kept with same with the original DGrasp. We use the first frame where the hand and the object are in contact with each other as the reference frame. Then the policy is trained to grasp the object according to the reference hand and object goal at first. After that, only the root is guided to complete the task. 2) DGraspTracking, where we divide the whole sequence into several subsequences, each of which has 10 frames, and define the end frame of the subsequence as the reference frame. Then the grasping policy is used to guide the hand and gradually track the object according to the hand and the object pose of each reference frame. We improve the DGrasp-Tracking by optimizing the policy through the quasi-physical curriculum and creating \"DGrasp-Tracking (w/ Curriculum)\" trying to improve its performance. For model-based methods, we compare with Control-VAE [66] and traditional MPC approaches. For Control-VAE, we modify its implementation for the manipulation tracking task. We additionally consider three differentiable physics models to conduct model-predictive control for solving the task. Taking the analytical model with the most tightened contact model as the base model (\"MPC (w/ base sim.)\"), we further augment it with a general state-of-the-art contact smoothing for robot manipulation [52] and create \"MPC (w/ base sim. w/ soften)\". Details of baseline models are deferred to the Supp." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.735, + 0.788, + 0.842 + ], + "angle": 0, + "content": "Training and evaluation settings. The physics curriculum is composed of three stages. In the first stage, the parameter \\(\\alpha\\) varies from 0.1 to 0.0 and the contact model stiffness is relaxed to the softest level. In the second stage, \\(\\alpha\\) is fixed and the contact model stiffness varies from the softest version to the most tightened level gradually through eight stages. Details w.r.t. parameter settings are deferred to the Supp. In the first two stages, we alternately optimize the trajectory \\(\\mathcal{A}\\) and parameters \\(S\\). In each optimization iteration, the \\(\\mathcal{A}\\) is optimized" + } + ], + [ + { + "type": "page_number", + "bbox": [ + 0.218, + 0.116, + 0.236, + 0.127 + ], + "angle": 0, + "content": "12" + }, + { + "type": "header", + "bbox": [ + 0.272, + 0.115, + 0.355, + 0.127 + ], + "angle": 0, + "content": "X. Liu et al." + }, + { + "type": "table_caption", + "bbox": [ + 0.216, + 0.145, + 0.784, + 0.172 + ], + "angle": 0, + "content": "Table 1: Quantitative evaluations and comparisons to baselines. Bold red numbers for best values and italic blue values for the second best-performed ones." + }, + { + "type": "table", + "bbox": [ + 0.218, + 0.173, + 0.784, + 0.345 + ], + "angle": 0, + "content": "
SimulatorMethod\\( {R}_{\\text{err }}\\left( {{}^{ \\circ }, \\downarrow }\\right) \\)\\( {T}_{\\text{err }}\\left( {{cm}, \\downarrow }\\right) \\)MPJPE (mm,↓)CD (mm,↓)Success Rate (%,↑)
BulletModel FreeDGrasp-Base44.245.8240.5516.370/13.73/15.69
DGrasp-Tracking44.455.0437.5614.720/15.69/15.69
DGrasp-Tracking (w/ curric.)33.864.6030.4713.537.84/23.53/37.25
Model BasedControl-VAE42.452.7325.2110.940/15.68/23.53
MPC (w/ base sim.)32.563.6724.6210.800/15.68/31.37
MPC (w/ base sim. w/ soften)31.893.6328.2611.310/21.57/37.25
Ours24.211.9724.409.8527.45/37.25/58.82
Isaac GymModel FreeDGrasp-Base36.414.5650.9718.780/7.84/7.84
DGrasp-Tracking44.715.5741.5316.720/0/7.84
DGrasp-Tracking (w/ curric.)38.755.1340.0916.260/23.53/31.37
Model BasedControl-VAE35.404.6127.6313.170/13.73/29.41
MPC (w/ base sim.)37.234.7323.199.750/15.69/31.37
MPC (w/ base sim. w/ soften)36.404.4623.2710.340/9.80/23.53
Ours25.972.0825.3310.3121.57/43.14/56.86
" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.362, + 0.785, + 0.437 + ], + "angle": 0, + "content": "for 100 steps while \\( S \\) is optimized for 1000 steps. In the third stage, \\( \\mathcal{A} \\) and \\( \\psi \\) are optimized for 256 steps in each iteration. For time-stepping, \\( \\mathrm{dt} \\) is set to \\( 5 \\times 10^{-4} \\) in the parameterized and the target simulators. The articulated multi-body is controlled by joint motors and root velocities in the parameterized quasi-physical simulator while PD control [54] is leveraged in the target simulators." + }, + { + "type": "title", + "bbox": [ + 0.215, + 0.448, + 0.54, + 0.463 + ], + "angle": 0, + "content": "4.2 Dexterous Manipulating Tracking" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.465, + 0.785, + 0.599 + ], + "angle": 0, + "content": "We conducted thorough experiments in two widely used simulators [9, 33]. We treat them as realistic simulated physical environments with high fidelity and wish to track the manipulation in them. In summary, we can control a dexterous hand to complete a wide range of the manipulation tracking tasks with nontrivial object movements and changing contacts. As presented in Table 1, we can achieve significantly higher success rates calculated under three thresholds than the best-performed baseline in both tested simulators. Fig. 4 showcases qualitative examples and comparisons. Please check out our website and video for animated results." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.601, + 0.785, + 0.72 + ], + "angle": 0, + "content": "Complex manipulations. For examples shown in Fig. 4, we can complete the tracking task on examples with large object re-orientations and complicated tool-using (Fig. \\((a,b,c)\\)). However, DGrasp-Tracking fails to establish sufficient contact for correctly manipulating the object. In more detail, in Fig. 4(b), the bunny gradually bounced out from its hand in Bullet, while our method does not suffer from this difficulty. In Fig. 4(c), the spoon can be successfully picked up and waved back-and-forth in our method, while DGrasp-Tracking loses the track right from the start." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.722, + 0.785, + 0.782 + ], + "angle": 0, + "content": "Bimanual manipulations. We are also capable of tracking bimanual manipulations. As shown in the example in Fig. 4(d), where two hands collaborate to relocate the object, DGrasp-Tracking fails to accurately track the object, while our method significantly outperforms it." + }, + { + "type": "title", + "bbox": [ + 0.215, + 0.785, + 0.535, + 0.8 + ], + "angle": 0, + "content": "4.3 Further Analysis and Discussions" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.81, + 0.785, + 0.84 + ], + "angle": 0, + "content": "Could model-free methods benefit from the physics curriculum? In addition to the demonstrated merits of our quasi-physical simulators, we further" + } + ], + [ + { + "type": "header", + "bbox": [ + 0.302, + 0.115, + 0.733, + 0.13 + ], + "angle": 0, + "content": "Quasi-Physical Simulators for Dexterous Manipulations Transfer" + }, + { + "type": "page_number", + "bbox": [ + 0.768, + 0.116, + 0.786, + 0.127 + ], + "angle": 0, + "content": "13" + }, + { + "type": "table_caption", + "bbox": [ + 0.215, + 0.145, + 0.788, + 0.172 + ], + "angle": 0, + "content": "Table 2: Ablation studies. Bold red numbers for best values and italic blue values for the second best-performed ones. The simulation environment is Bullet." + }, + { + "type": "table", + "bbox": [ + 0.217, + 0.173, + 0.784, + 0.269 + ], + "angle": 0, + "content": "
MethodRerr(°,↓)Terr(cm,↓)MPJPE(mm,↓)CD(mm,↓)Success Rate(%,↑)
Ours w/o Analytical Sim.44.274.3929.8412.910/13.73/25.49
Ours w/o Residual Physics33.693.8126.5710.345.88/23.53/41.18
Ours w/o Local Force NN35.982.9032.8712.440/19.61/35.29
Ours w/o Curriculum42.404.8732.6113.370/17.64/29.41
Ours w/ Curriculum II29.582.3331.6110.2911.76/27.45/50.98
Ours24.211.9724.409.8527.45/37.25/58.82
" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.271, + 0.788, + 0.378 + ], + "angle": 0, + "content": "explore whether model-free strategies can benefit from them. We introduce the \"DGrasp-Tracking (w/ Curriculum)\" method and compare its performance with the original DGrasp-Tracking model. As shown in Table 1 and the visual comparisons in Fig. 6, the DGrasp-Tracking model indeed benefits from a well-designed physics curriculum. For example, as illustrated in Fig. 6, the curriculum can significantly improve its performance, enabling it to nearly complete challenging tracking tasks where the original version struggles." + }, + { + "type": "title", + "bbox": [ + 0.216, + 0.384, + 0.402, + 0.403 + ], + "angle": 0, + "content": "5 Ablation Study" + }, + { + "type": "image", + "bbox": [ + 0.303, + 0.419, + 0.538, + 0.609 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.544, + 0.429, + 0.695, + 0.613 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.215, + 0.61, + 0.785, + 0.639 + ], + "angle": 0, + "content": "Fig. 5: (a) Qualitative comparisons between our full method and the ablated models; (b) Training loss curve comparisons; (c) Tracking loss curve comparisons." + }, + { + "type": "image", + "bbox": [ + 0.223, + 0.641, + 0.788, + 0.765 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.215, + 0.773, + 0.788, + 0.803 + ], + "angle": 0, + "content": "Fig. 6: Visual evidence on boosting DGrasp-Tracking's performance via optimizing it through a physics curriculum." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.81, + 0.788, + 0.842 + ], + "angle": 0, + "content": "We conduct a wide range of ablation studies to validate the effectiveness of some of our crucial designs, including the parameterized analytical physics" + } + ], + [ + { + "type": "page_number", + "bbox": [ + 0.218, + 0.116, + 0.236, + 0.127 + ], + "angle": 0, + "content": "14" + }, + { + "type": "header", + "bbox": [ + 0.272, + 0.115, + 0.355, + 0.127 + ], + "angle": 0, + "content": "X. Liu et al." + }, + { + "type": "text", + "bbox": [ + 0.217, + 0.147, + 0.785, + 0.192 + ], + "angle": 0, + "content": "model, the parameterized residual physics, the role of the local force network, the necessity of introducing a physics curriculum into the optimization, and how the design on the curriculum stages affects the result." + }, + { + "type": "text", + "bbox": [ + 0.216, + 0.193, + 0.787, + 0.36 + ], + "angle": 0, + "content": "Parameterized analytical model. The skeleton of the quasi-physical simulator is an analytical physics model. The intuition is that the parameterized simulator with such physical bias can be optimized towards a realistic simulator more easily than training pure neural networks for approximating. To validate this, we ablate the analytical model and use neural networks to approximate physics in Bullet directly (denoted as \"Ours w/o Analytical Sim\"). The quantitative (Table 2) and qualitative (Fig. 5) results indicate that the physical biases brought by the analytical model could help the parameterized simulator to learn better physics in the contact-rich scenario. For instance, in the example demonstrated in Fig. 5, the ablated version fails to guide the robot hand to successfully pinch the object in the second figure." + }, + { + "type": "text", + "bbox": [ + 0.216, + 0.362, + 0.787, + 0.452 + ], + "angle": 0, + "content": "Parameterized residual physics. To validate the necessity of introducing residual force networks to close the gap between the physics modeled in the parameterized analytical simulator and that of a realistic simulator, we ablate the parameterized force network and create a version named \"Ours w/o Residual Physics\". Table 2 demonstrated its role in enabling the parameterized simulator to approximate realistic physics models." + }, + { + "type": "text", + "bbox": [ + 0.216, + 0.454, + 0.787, + 0.544 + ], + "angle": 0, + "content": "Local residual force network. To adequately leverage state and contact-related information for predicting residual contact forces, we propose to use two types of networks: 1) a local force network for per contact pair residual forces and 2) a global network for additionally compensating. The local network is introduced for fine-grained approximation. We ablate this design and compare the result with our full model to validate this (see Fig. 5 and Table 1)." + }, + { + "type": "text", + "bbox": [ + 0.216, + 0.546, + 0.787, + 0.667 + ], + "angle": 0, + "content": "Optimizing through an analytical physics curriculum. We further investigate the effectiveness of the analytical curriculum design and how its design influences the result. Specifically, we create two ablated versions: 1) \"Ours w/o Curriculum\", where the optimization starts directly from the parameterized analytical model with articulated rigid constraints tightened and the stiffest contact model, and 2) \"Ours w/ Curriculum II\", where we move some stages out from the original curriculum. Table 2 and Fig. 5 demonstrate that both the curriculum and the optimization path will affect the model's performance." + }, + { + "type": "title", + "bbox": [ + 0.217, + 0.676, + 0.518, + 0.693 + ], + "angle": 0, + "content": "6 Conclusion and Limitations" + }, + { + "type": "text", + "bbox": [ + 0.216, + 0.703, + 0.787, + 0.793 + ], + "angle": 0, + "content": "In this work, we investigate creating better simulators for solving complex robotic tasks involving complicated dynamics where the previous best-performed optimization strategy fails. We present a family of parameterized quasi-physical simulators that can be both programmed to relax various constraints for task optimization and can be tailored to approximate realistic physics. We tackle the difficult manipulation transfer task via a physics curriculum." + }, + { + "type": "text", + "bbox": [ + 0.216, + 0.795, + 0.787, + 0.84 + ], + "angle": 0, + "content": "Limitations. The method is limited by the relatively simple spring-damper model for contact constraint relaxation. Introducing delicate analytical contact models to parameterized simulators is an interesting research direction." + } + ], + [ + { + "type": "header", + "bbox": [ + 0.302, + 0.115, + 0.732, + 0.129 + ], + "angle": 0, + "content": "Quasi-Physical Simulators for Dexterous Manipulations Transfer" + }, + { + "type": "page_number", + "bbox": [ + 0.768, + 0.117, + 0.785, + 0.127 + ], + "angle": 0, + "content": "15" + }, + { + "type": "title", + "bbox": [ + 0.218, + 0.145, + 0.393, + 0.163 + ], + "angle": 0, + "content": "Acknowledgement" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.178, + 0.787, + 0.224 + ], + "angle": 0, + "content": "We thank Rui Chen, Yiheng Guan, and Taoran Jiang for their valuable support in providing the Allegro hardware and setting up the environment setup during the rebuttal period." + }, + { + "type": "title", + "bbox": [ + 0.218, + 0.249, + 0.323, + 0.265 + ], + "angle": 0, + "content": "References" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.282, + 0.786, + 0.338 + ], + "angle": 0, + "content": "1. Ajay, A., Wu, J., Fazeli, N., Bauza, M., Kaelbling, L.P., Tenenbaum, J.B., Rodriguez, A.: Augmenting physical simulators with stochastic neural networks: Case study of planar pushing and bouncing. In: 2018 IEEE/RSJ International Conference on Intelligent Robots and Systems (IROS). pp. 3066-3073. IEEE (2018)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.339, + 0.787, + 0.38 + ], + "angle": 0, + "content": "2. Akkaya, I., Andrychowicz, M., Chociej, M., Litwin, M., McGrew, B., Petron, A., Paino, A., Plappert, M., Powell, G., Ribas, R., et al.: Solving rubik's cube with a robot hand. arXiv preprint arXiv:1910.07113 (2019)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.381, + 0.786, + 0.408 + ], + "angle": 0, + "content": "3. Andrews, S., Erleben, K., Ferguson, Z.: Contact and friction simulation for computer graphics. In: ACM SIGGRAPH 2022 Courses, pp. 1-172 (2022)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.409, + 0.786, + 0.463 + ], + "angle": 0, + "content": "4. Andrychowicz, O.M., Baker, B., Chociej, M., Jozefowicz, R., McGrew, B., Pachocki, J., Petron, A., Plappert, M., Powell, G., Ray, A., et al.: Learning dexterous in-hand manipulation. The International Journal of Robotics Research 39(1), 3-20 (2020)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.464, + 0.786, + 0.492 + ], + "angle": 0, + "content": "5. Baraff, D.: An introduction to physically based modeling: rigid body simulation ii—nonpenetration constraints. SIGGRAPH course notes pp. D31–D68 (1997)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.492, + 0.786, + 0.547 + ], + "angle": 0, + "content": "6. Chen, T., Tippur, M., Wu, S., Kumar, V., Adelson, E., Agrawal, P.: Visual dexterity: In-hand reorientation of novel and complex object shapes. Science Robotics 8(84), eadc9244 (2023). https://doi.org/10.1126/scirobotics.adc9244, https://www.science.org/doi/abs/10.1126/scirobotics.adc9244" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.548, + 0.786, + 0.575 + ], + "angle": 0, + "content": "7. Chen, T., Xu, J., Agrawal, P.: A system for general in-hand object re-orientation. Conference on Robot Learning (2021)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.576, + 0.786, + 0.631 + ], + "angle": 0, + "content": "8. Christen, S., Kocabas, M., Aksan, E., Hwangbo, J., Song, J., Hilliges, O.: D-grasp: Physically plausible dynamic grasp synthesis for hand-object interactions. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 20577-20586 (2022)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.632, + 0.786, + 0.659 + ], + "angle": 0, + "content": "9. Coumans, E., Bai, Y.: Pybullet, a python module for physics simulation for games, robotics and machine learning (2016)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.66, + 0.786, + 0.687 + ], + "angle": 0, + "content": "0. Deng, Y., Yu, H.X., Wu, J., Zhu, B.: Learning vortex dynamics for fluid inference and prediction. arXiv preprint arXiv:2301.11494 (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.688, + 0.786, + 0.715 + ], + "angle": 0, + "content": "1. Du, T.: Deep learning for physics simulation. In: ACM SIGGRAPH 2023 Courses, pp. 1-25 (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.716, + 0.786, + 0.757 + ], + "angle": 0, + "content": "2. Dunlavy, D.M., O'Leary, D.P.: Homotopy optimization methods for global optimization. Tech. rep., Sandia National Laboratories (SNL), Albuquerque, NM, and Livermore, CA ... (2005)" + }, + { + "type": "ref_text", + "bbox": [ + 0.22, + 0.758, + 0.786, + 0.813 + ], + "angle": 0, + "content": "13. Fan, Z., Taheri, O., Tzionas, D., Kocabas, M., Kaufmann, M., Black, M.J., Hilliges, O.: ARCTIC: A dataset for dexterous bimanual hand-object manipulation. In: Proceedings IEEE Conference on Computer Vision and Pattern Recognition (CVPR) (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.22, + 0.814, + 0.786, + 0.84 + ], + "angle": 0, + "content": "14. Featherstone, R.: Rigid body dynamics algorithms (2007), https://apisemantic scholar.org/CorpusID:58437819" + }, + { + "type": "list", + "bbox": [ + 0.22, + 0.282, + 0.787, + 0.84 + ], + "angle": 0, + "content": null + } + ], + [ + { + "type": "page_number", + "bbox": [ + 0.218, + 0.116, + 0.236, + 0.127 + ], + "angle": 0, + "content": "16" + }, + { + "type": "header", + "bbox": [ + 0.272, + 0.115, + 0.356, + 0.127 + ], + "angle": 0, + "content": "X. Liu et al." + }, + { + "type": "ref_text", + "bbox": [ + 0.218, + 0.147, + 0.787, + 0.189 + ], + "angle": 0, + "content": "15. Freeman, C.D., Frey, E., Raichuk, A., Girgin, S., Mordatch, I., Bachem, O.: Brax-a differentiable physics engine for large scale rigid body simulation. arXiv preprint arXiv:2106.13281 (2021)" + }, + { + "type": "ref_text", + "bbox": [ + 0.218, + 0.189, + 0.787, + 0.23 + ], + "angle": 0, + "content": "16. Fussell, L., Bergamin, K., Holden, D.: Supertrack: Motion tracking for physically simulated characters using supervised learning. ACM Transactions on Graphics (TOG) 40(6), 1-13 (2021)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.23, + 0.787, + 0.257 + ], + "angle": 0, + "content": "17. Gao, J., Michelis, M.Y., Spielberg, A., Katzschmann, R.K.: Sim-to-real of soft robots with learned residual physics. arXiv preprint arXiv:2402.01086 (2024)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.257, + 0.787, + 0.284 + ], + "angle": 0, + "content": "18. Garcia, C.E., Prett, D.M., Morari, M.: Model predictive control: Theory and practice—a survey. Automatica 25(3), 335-348 (1989)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.284, + 0.787, + 0.325 + ], + "angle": 0, + "content": "19. Geilinger, M., Hahn, D., Zehnder, J., Bächer, M., Thomaszewski, B., Coros, S.: Add: Analytically differentiable dynamics for multi-body systems with frictional contact. ACM Transactions on Graphics (TOG) 39(6), 1-15 (2020)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.325, + 0.787, + 0.366 + ], + "angle": 0, + "content": "20. Grandia, R., Farshidian, F., Knoop, E., Schumacher, C., Hutter, M., Bächer, M.: Doc: Differentiable optimal control for retargeting motions onto legged robots. ACM Transactions on Graphics (TOG) 42(4), 1-14 (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.366, + 0.787, + 0.407 + ], + "angle": 0, + "content": "21. Gupta, A., Eppner, C., Levine, S., Abbeel, P.: Learning dexterous manipulation for a soft robotic hand from human demonstrations. In: 2016 IEEE/RSJ International Conference on Intelligent Robots and Systems (IROS). pp. 3786-3793. IEEE (2016)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.407, + 0.787, + 0.461 + ], + "angle": 0, + "content": "22. Heiden, E., Millard, D., Coumans, E., Sheng, Y., Sukhatme, G.S.: Neuralsim: Augmenting differentiable simulators with neural networks. In: 2021 IEEE International Conference on Robotics and Automation (ICRA). pp. 9474-9481. IEEE (2021)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.461, + 0.787, + 0.488 + ], + "angle": 0, + "content": "23. Howell, T.A., Le Cleac'h, S., Kolter, J.Z., Schwager, M., Manchester, Z.: Dojo: A differentiable simulator for robotics. arXiv preprint arXiv:2203.00806 9 (2022)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.488, + 0.787, + 0.529 + ], + "angle": 0, + "content": "24. Howell, T.A., Le Cleac'h, S., Singh, S., Florence, P., Manchester, Z., Sindhwani, V.: Trajectory optimization with optimization-based dynamics. IEEE Robotics and Automation Letters 7(3), 6750-6757 (2022)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.529, + 0.787, + 0.556 + ], + "angle": 0, + "content": "25. Hwangbo, J., Lee, J., Hutter, M.: Per-contact iteration method for solving contact dynamics. IEEE Robotics and Automation Letters 3(2), 895-902 (2018)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.556, + 0.787, + 0.597 + ], + "angle": 0, + "content": "26. Lan, L., Yang, Y., Kaufman, D., Yao, J., Li, M., Jiang, C.: Medial ipc: accelerated incremental potential contact with medial elastics. ACM Transactions on Graphics 40(4) (2021)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.597, + 0.787, + 0.638 + ], + "angle": 0, + "content": "27. Li, M., Ferguson, Z., Schneider, T., Langlois, T.R., Zorin, D., Panozzo, D., Jiang, C., Kaufman, D.M.: Incremental potential contact: intersection-and inversion-free, large-deformation dynamics. ACM Trans. Graph. 39(4), 49 (2020)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.638, + 0.787, + 0.664 + ], + "angle": 0, + "content": "28. Liao, S.: On the homotopy analysis method for nonlinear problems. Applied mathematics and computation 147(2), 499-513 (2004)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.664, + 0.787, + 0.691 + ], + "angle": 0, + "content": "29. Lin, X., Yang, Z., Zhang, X., Zhang, Q.: Continuation path learning for homotopy optimization (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.691, + 0.787, + 0.718 + ], + "angle": 0, + "content": "30. Liu, C.K., Jain, S.: A quick tutorial on multibody dynamics. Online tutorial, June p. 7 (2012)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.718, + 0.787, + 0.744 + ], + "angle": 0, + "content": "31. Liu, X., Pathak, D., Kitani, K.M.: Herd: Continuous human-to-robot evolution for learning from human demonstration. arXiv preprint arXiv:2212.04359 (2022)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.744, + 0.787, + 0.785 + ], + "angle": 0, + "content": "32. Liu, Y., Yang, H., Si, X., Liu, L., Li, Z., Zhang, Y., Liu, Y., Yi, L.: Taco: Benchmarking generalizable bimanual tool-action-object understanding. arXiv preprint arXiv:2401.08399 (2024)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.785, + 0.787, + 0.84 + ], + "angle": 0, + "content": "33. Makoviychuk, V., Wawrzyniak, L., Guo, Y., Lu, M., Storey, K., Macklin, M., Hoeller, D., Rudin, N., Allshire, A., Handa, A., et al.: Isaac gym: High performancegpu-based physics simulation for robot learning. arXiv preprint arXiv:2108.10470 (2021)" + }, + { + "type": "list", + "bbox": [ + 0.217, + 0.147, + 0.787, + 0.84 + ], + "angle": 0, + "content": null + } + ], + [ + { + "type": "header", + "bbox": [ + 0.302, + 0.115, + 0.732, + 0.129 + ], + "angle": 0, + "content": "Quasi-Physical Simulators for Dexterous Manipulations Transfer" + }, + { + "type": "page_number", + "bbox": [ + 0.768, + 0.117, + 0.786, + 0.127 + ], + "angle": 0, + "content": "17" + }, + { + "type": "ref_text", + "bbox": [ + 0.214, + 0.147, + 0.788, + 0.189 + ], + "angle": 0, + "content": "34. Mandikal, P., Grauman, K.: Dexvip: Learning dexterous grasping with human hand pose priors from video. In: Conference on Robot Learning. pp. 651-661. PMLR (2022)" + }, + { + "type": "ref_text", + "bbox": [ + 0.214, + 0.19, + 0.787, + 0.232 + ], + "angle": 0, + "content": "35. Marcucci, T., Gabiccini, M., Artoni, A.: A two-stage trajectory optimization strategy for articulated bodies with unscheduled contact sequences. IEEE Robotics and Automation Letters 2(1), 104-111 (2016)" + }, + { + "type": "ref_text", + "bbox": [ + 0.214, + 0.232, + 0.787, + 0.273 + ], + "angle": 0, + "content": "36. Mordatch, I., Popovic, Z., Todorov, E.: Contact-invariant optimization for hand manipulation. In: Proceedings of the ACM SIGGRAPH/Eurographics symposium on computer animation. pp. 137-144 (2012)" + }, + { + "type": "ref_text", + "bbox": [ + 0.214, + 0.273, + 0.787, + 0.314 + ], + "angle": 0, + "content": "37. Pang, T., Suh, H.T., Yang, L., Tedrake, R.: Global planning for contact-rich manipulation via local smoothing of quasi-dynamic contact models. IEEE Transactions on Robotics (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.214, + 0.315, + 0.787, + 0.357 + ], + "angle": 0, + "content": "38. Pang, T., Tedrake, R.: A convex quasistatic time-stepping scheme for rigid multi-body systems with contact and friction. In: 2021 IEEE International Conference on Robotics and Automation (ICRA). pp. 6614-6620. IEEE (2021)" + }, + { + "type": "ref_text", + "bbox": [ + 0.214, + 0.357, + 0.787, + 0.398 + ], + "angle": 0, + "content": "39. Peng, X.B., Abbeel, P., Levine, S., Van de Panne, M.: Deepmimic: Example-guided deep reinforcement learning of physics-based character skills. ACM Transactions On Graphics (TOG) 37(4), 1-14 (2018)" + }, + { + "type": "ref_text", + "bbox": [ + 0.214, + 0.398, + 0.787, + 0.424 + ], + "angle": 0, + "content": "40. Pfaff, T., Fortunato, M., Sanchez-Gonzalez, A., Battaglia, P.W.: Learning mesh-based simulation with graph networks. arXiv preprint arXiv:2010.03409 (2020)" + }, + { + "type": "ref_text", + "bbox": [ + 0.214, + 0.425, + 0.787, + 0.467 + ], + "angle": 0, + "content": "41. Pfrommer, S., Halm, M., Posa, M.: Contactnets: Learning discontinuous contact dynamics with smooth, implicit representations. In: Conference on Robot Learning. pp. 2279-2291. PMLR (2021)" + }, + { + "type": "ref_text", + "bbox": [ + 0.214, + 0.467, + 0.787, + 0.509 + ], + "angle": 0, + "content": "42. Qin, Y., Su, H., Wang, X.: From one hand to multiple hands: Imitation learning for dexterous manipulation from single-camera teleoperation. IEEE Robotics and Automation Letters 7(4), 10873-10881 (2022)" + }, + { + "type": "ref_text", + "bbox": [ + 0.214, + 0.509, + 0.787, + 0.55 + ], + "angle": 0, + "content": "43. Qin, Y., Wu, Y.H., Liu, S., Jiang, H., Yang, R., Fu, Y., Wang, X.: Dexamv: Imitation learning for dexterous manipulation from human videos. In: European Conference on Computer Vision. pp. 570-587. Springer (2022)" + }, + { + "type": "ref_text", + "bbox": [ + 0.214, + 0.55, + 0.787, + 0.592 + ], + "angle": 0, + "content": "44. Qin, Y., Yang, W., Huang, B., Van Wyk, K., Su, H., Wang, X., Chao, Y.W., Fox, D.: Anyteleop: A general vision-based dexterous robot arm-hand teleoperation system. arXiv preprint arXiv:2307.04577 (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.214, + 0.592, + 0.787, + 0.633 + ], + "angle": 0, + "content": "45. Radosavovic, I., Wang, X., Pinto, L., Malik, J.: State-only imitation learning for dexterous manipulation. In: 2021 IEEE/RSJ International Conference on Intelligent Robots and Systems (IROS). pp. 7865-7871. IEEE (2021)" + }, + { + "type": "ref_text", + "bbox": [ + 0.214, + 0.633, + 0.787, + 0.675 + ], + "angle": 0, + "content": "46. Rajeswaran, A., Kumar, V., Gupta, A., Vezzani, G., Schulman, J., Todorov, E., Levine, S.: Learning complex dexterous manipulation with deep reinforcement learning and demonstrations. arXiv preprint arXiv:1709.10087 (2017)" + }, + { + "type": "ref_text", + "bbox": [ + 0.214, + 0.675, + 0.787, + 0.716 + ], + "angle": 0, + "content": "47. Rusu, A.A., Colmenarejo, S.G., Gulcehre, C., Desjardins, G., Kirkpatrick, J., Pascanu, R., Mnih, V., Kavukcuoglu, K., Hadsell, R.: Policy distillation. arXiv preprint arXiv:1511.06295 (2015)" + }, + { + "type": "ref_text", + "bbox": [ + 0.214, + 0.716, + 0.787, + 0.758 + ], + "angle": 0, + "content": "48. Schmeckpeper, K., Rybkin, O., Daniilidis, K., Levine, S., Finn, C.: Reinforcement learning with videos: Combining offline observations with interaction. arXiv preprint arXiv:2011.06507 (2020)" + }, + { + "type": "ref_text", + "bbox": [ + 0.214, + 0.758, + 0.787, + 0.784 + ], + "angle": 0, + "content": "49. ShadowRobot: Shadowrobot dexterous hand (2005), https://www.shadowrobot.com/dexterous-hand-series/" + }, + { + "type": "ref_text", + "bbox": [ + 0.214, + 0.785, + 0.787, + 0.813 + ], + "angle": 0, + "content": "50. Siahkoohi, A., Louboutin, M., Herrmann, F.J.: Neural network augmented wave-equation simulation. arXiv preprint arXiv:1910.00925 (2019)" + }, + { + "type": "ref_text", + "bbox": [ + 0.214, + 0.813, + 0.787, + 0.84 + ], + "angle": 0, + "content": "51. Suh, H., Wang, Y.: Comparing effectiveness of relaxation methods for warm starting trajectory optimization through soft contact (2019)" + }, + { + "type": "list", + "bbox": [ + 0.214, + 0.147, + 0.788, + 0.84 + ], + "angle": 0, + "content": null + } + ], + [ + { + "type": "page_number", + "bbox": [ + 0.218, + 0.116, + 0.236, + 0.127 + ], + "angle": 0, + "content": "18" + }, + { + "type": "header", + "bbox": [ + 0.272, + 0.115, + 0.356, + 0.128 + ], + "angle": 0, + "content": "X. Liu et al." + }, + { + "type": "ref_text", + "bbox": [ + 0.218, + 0.148, + 0.785, + 0.175 + ], + "angle": 0, + "content": "52. Suh, H.J.T., Pang, T., Tedrake, R.: Bundled gradients through contact via randomized smoothing. IEEE Robotics and Automation Letters 7(2), 4000-4007 (2022)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.177, + 0.787, + 0.23 + ], + "angle": 0, + "content": "53. Taheri, O., Ghorbani, N., Black, M.J., Tzionas, D.: Grab: A dataset of whole-body human grasping of objects. In: Computer Vision-ECCV 2020: 16th European Conference, Glasgow, UK, August 23-28, 2020, Proceedings, Part IV 16. pp. 581-600. Springer (2020)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.231, + 0.785, + 0.256 + ], + "angle": 0, + "content": "54. Tan, K.K., Wang, Q.G., Hang, C.C.: Advances in PID control. Springer Science & Business Media (2012)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.257, + 0.785, + 0.283 + ], + "angle": 0, + "content": "55. Tedrake, R., the Drake Development Team: Drake: Model-based design and verification for robotics (2019), https://drake.mit.edu" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.284, + 0.785, + 0.324 + ], + "angle": 0, + "content": "56. Todorov, E., Erez, T., Tassa, Y.: Mujoco: A physics engine for model-based control. In: 2012 IEEE/RSJ international conference on intelligent robots and systems. pp. 5026-5033. IEEE (2012)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.325, + 0.785, + 0.365 + ], + "angle": 0, + "content": "57. Traore, R., Caseles-Dupré, H., Lesort, T., Sun, T., Díaz-Rodriguez, N., Filliat, D.: Continual reinforcement learning deployed in real-life using policy distillation and sim2real transfer. arXiv preprint arXiv:1906.04452 (2019)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.366, + 0.785, + 0.405 + ], + "angle": 0, + "content": "58. Villegas, R., Ceylan, D., Hertzmann, A., Yang, J., Saito, J.: Contact-aware retargeting of skinned motion. In: Proceedings of the IEEE/CVF International Conference on Computer Vision. pp. 9720-9729 (2021)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.406, + 0.785, + 0.446 + ], + "angle": 0, + "content": "59. Wang, Y., Weidner, N.J., Baxter, M.A., Hwang, Y., Kaufman, D.M., Sueda, S.: Redmax: Efficient & flexible approach for articulated dynamics. ACM Transactions on Graphics (TOG) 38(4), 1-10 (2019)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.447, + 0.785, + 0.487 + ], + "angle": 0, + "content": "60. Wang, Y., Lin, J., Zeng, A., Luo, Z., Zhang, J., Zhang, L.: Physhoi: Physics-based imitation of dynamic human-object interaction. arXiv preprint arXiv:2312.04393 (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.488, + 0.785, + 0.514 + ], + "angle": 0, + "content": "61. Watson, L.T., Haftka, R.T.: Modern homotopy methods in optimization. Computer Methods in Applied Mechanics and Engineering 74(3), 289-305 (1989)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.515, + 0.785, + 0.554 + ], + "angle": 0, + "content": "62. Wu, Y.H., Wang, J., Wang, X.: Learning generalizable dexterous manipulation from human grasp affordance. In: Conference on Robot Learning. pp. 618-629. PMLR (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.555, + 0.785, + 0.595 + ], + "angle": 0, + "content": "63. Xiong, S., He, X., Tong, Y., Deng, Y., Zhu, B.: Neural vortex method: from finite lagrangian particles to infinite dimensional eulerian dynamics. Computers & Fluids 258, 105811 (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.596, + 0.785, + 0.636 + ], + "angle": 0, + "content": "64. Xu, J., Chen, T., Zlokapa, L., Foshey, M., Matusik, W., Sueda, S., Agrawal, P.: An end-to-end differentiable framework for contact-aware robot design. arXiv preprint arXiv:2107.07501 (2021)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.637, + 0.785, + 0.677 + ], + "angle": 0, + "content": "65. Yamane, K., Nakamura, Y.: Stable penalty-based model of frictional contacts. In: Proceedings 2006 IEEE International Conference on Robotics and Automation, 2006. ICRA 2006. pp. 1904-1909. IEEE (2006)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.678, + 0.785, + 0.717 + ], + "angle": 0, + "content": "66. Yao, H., Song, Z., Chen, B., Liu, L.: Controlvae: Model-based learning of generative controllers for physics-based characters. ACM Transactions on Graphics (TOG) 41(6), 1-16 (2022)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.718, + 0.785, + 0.758 + ], + "angle": 0, + "content": "67. Zeng, A., Song, S., Lee, J., Rodriguez, A., Funkhouser, T.: Tossingbot: Learning to throw arbitrary objects with residual physics. IEEE Transactions on Robotics 36(4), 1307-1319 (2020)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.759, + 0.785, + 0.799 + ], + "angle": 0, + "content": "68. Zhang, H., Christen, S., Fan, Z., Zheng, L., Hwangbo, J., Song, J., Hilliges, O.: Artigrasp: Physically plausible synthesis of bi-manual dexterous grasping and articulation. arXiv preprint arXiv:2309.03891 (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.8, + 0.785, + 0.84 + ], + "angle": 0, + "content": "69. Zhang, S., Liu, B., Wang, Z., Zhao, T.: Model-based reparameterization policy gradient methods: Theory and practical algorithms. Advances in Neural Information Processing Systems 36 (2024)" + }, + { + "type": "list", + "bbox": [ + 0.217, + 0.148, + 0.787, + 0.84 + ], + "angle": 0, + "content": null + } + ], + [ + { + "type": "header", + "bbox": [ + 0.302, + 0.115, + 0.732, + 0.129 + ], + "angle": 0, + "content": "Quasi-Physical Simulators for Dexterous Manipulations Transfer" + }, + { + "type": "page_number", + "bbox": [ + 0.768, + 0.116, + 0.786, + 0.127 + ], + "angle": 0, + "content": "19" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.147, + 0.788, + 0.19 + ], + "angle": 0, + "content": "70. Zhang, Y., Clegg, A., Ha, S., Turk, G., Ye, Y.: Learning to transfer in-hand manipulations using a greedy shape curriculum. In: Computer Graphics Forum. vol. 42, pp. 25-36. Wiley Online Library (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.19, + 0.788, + 0.232 + ], + "angle": 0, + "content": "71. Zhao, W., Queralta, J.P., Westerlund, T.: Sim-to-real transfer in deep reinforcement learning for robotics: a survey. In: 2020 IEEE symposium series on computational intelligence (SSCI). pp. 737-744. IEEE (2020)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.232, + 0.788, + 0.273 + ], + "angle": 0, + "content": "72. Zhuang, F., Qi, Z., Duan, K., Xi, D., Zhu, Y., Zhu, H., Xiong, H., He, Q.: A comprehensive survey on transfer learning. Proceedings of the IEEE 109(1), 43-76 (2020)" + }, + { + "type": "list", + "bbox": [ + 0.217, + 0.147, + 0.788, + 0.273 + ], + "angle": 0, + "content": null + } + ] +] \ No newline at end of file diff --git a/2024/Parameterized Quasi-Physical Simulators for Dexterous Manipulations Transfer/ae282149-332d-459a-aebe-a693d1ac89fd_origin.pdf b/2024/Parameterized Quasi-Physical Simulators for Dexterous Manipulations Transfer/ae282149-332d-459a-aebe-a693d1ac89fd_origin.pdf new file mode 100644 index 0000000000000000000000000000000000000000..80e8a7405240d5321ac297dfd96ebff2a86ef595 --- /dev/null +++ b/2024/Parameterized Quasi-Physical Simulators for Dexterous Manipulations Transfer/ae282149-332d-459a-aebe-a693d1ac89fd_origin.pdf @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:56b5fb1dd6971c114ec5f94199131f409ec005a4eae3c8d0bf2bbc6f64c5451d +size 5321867 diff --git a/2024/Parameterized Quasi-Physical Simulators for Dexterous Manipulations Transfer/full.md b/2024/Parameterized Quasi-Physical Simulators for Dexterous Manipulations Transfer/full.md new file mode 100644 index 0000000000000000000000000000000000000000..40ff6042d483636d5dc30e55268e3e4bac716bcb --- /dev/null +++ b/2024/Parameterized Quasi-Physical Simulators for Dexterous Manipulations Transfer/full.md @@ -0,0 +1,329 @@ +# Parameterized Quasi-Physical Simulators for Dexterous Manipulations Transfer + +Xueyi Liu $^{1,3}$ , Kangbo Lyu $^{1}$ , Jieqiong Zhang $^{1}$ , Tao Du $^{1,2,3}$ , and Li Yi $^{1,2,3}$ + +$^{1}$ Tsinghua University $^{2}$ Shanghai AI Laboratory $^{3}$ Shanghai Qi Zhi Institute https://meowu7.github.io/QuasiSim + +![](images/9ad5b625ad6097a9287a6f84a453606d61ed514863586c2d9313a99e550120d8.jpg) + +![](images/d46ebd66664a8e9824b4bc4a38ff0319c5a628c34a96148c2b8581e84cac049c.jpg) + +![](images/944ad22b523cc633929f6e4f51084547ad0b17b00bec9e06fb856b317d937ab6.jpg) + +![](images/fd6cc2d5e7a115c6c7fdfee25e073dbf68c4df1eb5093be9e7230c23ea8d04d2.jpg) +(e) Result of a failed baseline + +![](images/0edadcc3e9758ed79a8aad322965386e9e91d087ce242cec0c73fc29b704884b.jpg) + +![](images/75270e327eb3aba02f2eda6843c5fa10cf480ab83aa757109c9587df0210a50e.jpg) + +![](images/ba69399690ce426d9287bc1f75cc59f3aa880cec7b32d13d6b10e1ebbbdad237.jpg) + +![](images/8180f3da0aa5fa07e0eb5556a305879fd99051828dee4246b4f7c99c53a86c40.jpg) + +![](images/30e7418e3123a22bbac62a8bd8aa308e845bd78284e41a20639982258f594a57.jpg) + +![](images/6abae0fc3a34145058f86b59fa3c5e861daf85c2165afab851543ddc36323164.jpg) +(f) Improved result via optimizing through the curriculum + +![](images/85bb9f79b34813de8c619a59a0b4a518693688115132bf53a6a90353a905cb84.jpg) + +![](images/5437d53ea46bd32cc5ebeb033b96f4ee8e0d47af725c3fe7e2638def7f06ee5e.jpg) + +![](images/743909a9a222135261106888ea15fd109ce2bb572622c6a49cc5bcbd31597076.jpg) + +![](images/8e7d30b22609c28a925427059d560e03a1b28788eb4f3e5c6770973167b182c7.jpg) + +![](images/e6b018b2d8a9175a37b4e174c9e694f9100858b74971cbcc8f106f8846623b26.jpg) + +![](images/201868126a183c1936cb9a37a5bdb7a6a4a612fca5d8c4b0dfd96842d278fbad.jpg) + +![](images/4e2e798039f97c7bfd5069933f0bd79e6bd9944792e9835348a69fa3084f0c34.jpg) +Timestamp +Fig. 1: By optimizing through a quasi-physical simulator curriculum, we successfully transfer human demonstrations to dexterous robot hand simulations. We enable accurate tracking of complex manipulations with changing contacts (Fig. (a)), nontrivial object motions (Fig. (b)) and intricate tool-using (Fig. (c,d)). Besides, our physics curriculum can substantially improve a failed baseline (Fig. (e,f)). + +![](images/bf61530f9eee1b1889c739226cb4cedc69b104d8e13e63757eeac42440152d35.jpg) +__________ + +![](images/258c6c475945a5eb1669ade573c7825ed880287a852285635fb55a5b7e0b85d7.jpg) +$\left( {x - {2x}}\right) t - x{y}^{2} = \left( {x - {2x}}\right) {f}^{\prime }t$ +# + +![](images/5acaa4e87ba70f059827886562ffbcb1bbe5eae44aaea88c37a916554423f4aa.jpg) + +![](images/4ee93e74bf06383554440472a7d95bc98360bdb5445b376563f56691ecc6a397.jpg) +Timestamp + +![](images/8663ea3d05dff9f3f1b53eec11297eb590d152876e3d9acaa097d72e2e58f7f1.jpg) +# + +![](images/f723652b1dcf649b48ab6399b00dee90e13af0aafe50b90311e6f3373a0cee3c.jpg) + +![](images/1f6e7ad2fc61d7db57ae195837c09a233a1e5982d99a20eddbffc80d18f77904.jpg) + +Abstract. We explore the dexterous manipulation transfer problem by designing simulators. The task wishes to transfer human manipulations to dexterous robot hand simulations and is inherently difficult due to its intricate, highly-constrained, and discontinuous dynamics and the need to control a dexterous hand with a DoF to accurately replicate human manipulations. Previous approaches that optimize in high-fidelity black-box simulators or a modified one with relaxed constraints only demonstrate limited capabilities or are restricted by insufficient simulation fidelity. We introduce parameterized quasi-physical simulators and a physics curriculum to overcome these limitations. The key ideas are 1) balancing between fidelity and optimizability of the simulation via a curriculum of parameterized simulators, and 2) solving the problem in each of the simulators from the curriculum, with properties ranging from high task optimizability to high fidelity. We successfully enable a dexterous hand to track complex and diverse manipulations in high-fidelity simulated environments, boosting the success rate by $11\%+$ from the best-performed baseline. The project website is available at QuasiSim. + +# 1 Introduction + +Advancing an embodied agent's capacity to interact with the world represents a significant stride toward achieving general artificial intelligence. Due to the substantial costs and potential hazards of setting up real robots to do trial and error, the standard approach for developing embodied algorithms involves learning in physical simulators [9, 15, 23, 25, 33, 56, 59] before transitioning to real-world deployment. In most cases, physical simulators are treated as black boxes, and extensive efforts have been devoted to developing learning and optimization methods for embodied skills within these black boxes. Despite the considerable progress [2, 6-8, 16, 20, 21, 31, 36, 39, 43, 46, 60, 62, 66, 68], the question like whether the simulators used are the most suitable ones is rarely discussed. In this work, we investigate this issue and illustrate how optimizing the simulator concurrently with skill acquisition can benefit a popular yet challenging task in robot manipulation - dexterous manipulation transfer. + +The task aims at transferring human-object manipulations to a dexterous robot hand, enabling it to physically track the reference motion of both the hand and the object (see Fig. 1). It is challenged by 1) the complex, highly constrained, non-smooth, and discontinuous dynamics with frequent contact establishment and breaking involved in the robot manipulation, 2) the requirement of precisely controlling a dexterous hand with a high DoF to densely track the manipulation at each frame, and 3) the morphology difference. Some existing works rely on high-fidelity black-box simulators, where a small difference in robot control can result in dramatically different manipulation outcomes due to abrupt contact changes, making the tracking objective highly non-smooth and hard to optimize [4,6,8,43,46]. In this way, their tasks are restricted to relatively simple goal-driven manipulations such as pouring and re-locating [8,43,46,68], in-hand re-orientation, flipping and spinning [4,6] with a fixed-root robot hand, or manipulating objects with simple geometry such as balls [36]. Other approaches attempt to improve optimization by relaxing physical constraints, with a primary focus on smoothing out contact responses [3,24,38,55,56]. However, their dynamics models may significantly deviate from real physics [38], hindering skill deployment. Consequently, we ask how to address the optimization challenge while preserving the high fidelity of the simulator. + +Our key insight is that a single simulator can hardly provide both high fidelity and excellent optimizability for contact-rich dexterous manipulations. Inspired by the line of homotopy methods [12, 28, 29, 61], we propose a curriculum of simulators to realize this. We start by utilizing a quasi-physical simulator to initially relax physical constraints and warm up the optimization. Subsequently, we transfer the optimization outcomes to simulators with gradually tightened physical constraints. Finally, we transition to a physically realistic simulator for skill deployment in realistic dynamics. + +To realize this vision, we propose a family of parameterized quasi-physical simulators for contact-rich dexterous manipulation tasks. These simulators can be customized to enhance task optimizability while can also be tailored to approximate realistic physics. The parameterized simulator represents + +an articulated multi rigid body as a parameterized point set, models contact using an unconstrained parameterized spring-damper, and compensates for unmodeled effects via parameterized residual physics. Specifically, the articulated multi-body dynamics model is relaxed as the point set dynamics model. An articulated object is relaxed into a set of points, sampled from the ambient space surrounding each body's surface mesh. The resulting dynamics model combines the original articulated dynamics with the mass-point dynamics of each individual point. Parameters are introduced to control the point set construction and the dynamics model. The contact model is softened as a parameterized spring-damper model [3, 19, 35, 38, 51] with parameters introduced to control when to calculate contacts and contact spring stiffness. The residual physics network compensate for unmodeled effects from the analytical modeling [22]. The parameterized simulator can be programmed for high optimizability by relaxing constraints in the analytical model and can be tailored to approximate realistic physics by learning excellent residual physics. We demonstrate that the challenging dexterous manipulation transfer task can be effectively addressed through curriculum optimization using a series of parameterized physical simulators. Initially, both articulated rigid constraints and the contact model stiffness are relaxed in the simulator. It may not reflect physical realism but provides a good environment where the manipulation transfer problem can be solved easily. Subsequently, the articulated rigid constraints and the contact model are gradually tightened. Task-solving proceeds iteratively within each simulator in the curriculum. Finally, the parameterized simulator is optimized to approximate realistic physics. Task optimization continues, yielding a dexterous hand trajectory capable of executing the manipulation in environments with realistic physics. + +We demonstrate the superiority of our method and compare it with previous model-free and model-based methods on challenging manipulation sequences from three datasets, describing single-hand or bimanual manipulations with daily objects or using tools. We conduct dexterous manipulation transfer on two widely used simulators, namely Bullet [9] and Isaac Gym [33] to demonstrate the generality and the efficacy of our method and the capability of our quasi-physical simulator to approximate the unknown black-box physics model in the contact-rich manipulation scenario (Fig. 1). We can track complex manipulations involving non-trivial object motions such as large rotations and complicated tool-using such as using a spoon to bring the water back and forth. Our approach successfully surpasses the previous best-performed method both quantitatively and qualitatively, achieving more than $11\%$ success rate than the previous best-performed method. Besides, optimizing through the physics curriculum can significantly enhance the performance of previously under-performed RL-based methods, almost completing the tracking problem from failure, as demonstrated in Fig. 1. This indicates the universality of our approach to embodied AI through optimization via a physics curriculum. Thorough ablations are conducted to validate the efficacy of our designs. + +Our contributions are three-fold: + +- We introduce a family of parameterized quasi-physical simulators that can be configured to relax various physical constraints, facilitating skill optimization, and can also be tailored to achieve high simulation fidelity. +- We present a quasi-physics curriculum along with a corresponding optimization method to address the challenging dexterous manipulation transfer problem. +- Extensive experiments demonstrate the effectiveness of our method in transferring complex manipulations, including non-trivial object motions and changing contacts, to a dexterous robot hand in simulation. + +# 2 Related Works + +Dexterous manipulation transfer. Transferring human manipulations to dexterous robot-hand simulations is an important topic in robot skill acquisition [8,21,31,43,60,62,68,70]. Most approaches treat the simulator as black-box physics models and try to learn skills directly from that [4,6,8,43,46]. However, their demonstrated capabilities are restricted to relatively simple tasks. Another trend of work tries to relax the physics model [37,38] to create a better environment for task optimization. However, due to the disparity between their modeling approach and realistic physics, successful trials are typically demonstrated only in their simulators, which can hardly complete the task under physically realistic dynamics. In this work, we introduce various parameterized analytical relaxations to improve the task optimizability while compensating for unmodeled effects via residual physics networks so the fidelity would not be sacrificed. + +Learning for simulation. Analytical methods can hardly approximate an extremely realistic physical world despite lots of smart and tremendous efforts made in developing numerical algorithms [19, 23, 26, 27]. Recently, data-driven approaches have attracted lots of interest for their high efficiency and strong approximation ability [10, 11, 22, 40, 41, 50, 63]. Special network designs are proposed to learn the contact behaviour [22, 41]. We in this work propose to leverage an analytical-neural hybrid approach and carefully design network modules for approximating residual contact forces in the contact-rich manipulation scenario. + +Sim-to-Sim and Sim-to-Real transfer. The field of robot manipulation continues to face challenges in the areas of Sim2Sim and Sim2Real transferability [71]. Considering the modeling gaps, the optimal strategy learned in a specific simulator is difficult to transfer to a different simulator or the real world. Therefore, many techniques for solving the problem have been proposed, including imitation learning [34,42,43,45,46,48], transfer learning [72], distillation [47,57], residual physics [17,67], and efforts on bridging the gap from the dynamics model aspect [22,69]. Our parameterized simulators learn residual physics involved in contact-rich robot manipulations. By combining an analytical base with residual networks, we showcase their ability to approximate realistic physics. + +# 3 Method + +Given a human manipulation demonstration, composed of a human hand mesh trajectory and an object pose trajectory $\{\mathcal{H},\mathcal{O}\}$ , the goal is transferring the + +![](images/99664072a887cf47fdcf35e54f57509bda3392ca2b5ecd914f4b592804383add.jpg) +Parameterized Quasi-Physical Simulator +Fig. 2: The parameterized quasi-physical simulator relaxes the articulated multi rigid body dynamics as the parameterized point set dynamics, controls the contact behavior via an unconstrained parameterized spring-damper contact model, and compensates for unmodeled effects via parameterized residual physics networks. We tackle the difficult dexterous manipulation transfer problem via a physics curriculum. + +demonstration to a dexterous robot hand in simulation. Formally, we aim to optimize a control trajectory $\mathcal{A}$ that drives the dexterous hand to manipulate the object in a realistic simulated environment so that the resulting hand trajectory $\hat{\mathcal{H}}$ and the object trajectory $\hat{\mathcal{O}}$ are close to the reference motion $\{\mathcal{H},\mathcal{O}\}$ . The problem is challenged by difficulties from the highly constrained, discontinuous, and non-smooth dynamics, the requirement of controlling a high DoF dexterous hand for tracking, and the morphology difference. + +Our method comprises two key designs to tackle the challenges: 1) a family of parameterized quasi-physical simulators, which can be programmed to enhance the optimizability of contact-rich dexterous manipulation tasks and can also be tailored to approximate realistic physics (Section 3.1), and 2) a physics curriculum that carefully adjusts the parameters of a line of quasi-physical simulators and a strategy that solves the difficult dexterous manipulation transfer task by addressing it within each simulator in the curriculum (Section 3.2). + +# 3.1 Parameterized Quasi-Physical Simulators + +Our quasi-physical simulator represents an articulated multi-body, i.e., the robotic dexterous hand, as a point set. The object is represented as a signed distance field. The base of the simulator is in an analytical form leveraging an unconstrained spring-damper contact model. Parameters are introduced to control the analytical relaxations on the articulated rigid constraints and the softness of the contact model. Additionally, neural networks are introduced to compensate for unmodeled effects beyond the analytical framework. We will elaborate on each of these design aspects below. + +Parameterized point set dynamics. Articulated multi-body represented in the reduced coordinate system [19,59] may require a large change in joint states to achieve a small adjustment in the Euclidean space. Moving the end effector from one point to a nearby point may require adjusting all joint states (Fig. 3). Besides, transferring the hand trajectory to a morphologically different hand requires correspondences to make the resulting trajectory close to the original one. Defining correspondences in the reduced coordinate or via sparse correspon + +dences will make the result suffer from noise in the data, leading to unwanted results finally (Fig. 3). Hence, we propose relaxing an articulated multi-rigid body into a mass-point set sampled from the ambient space surrounding each body. Each point is considered attached to the body from which it is sampled and is capable of both self-actuation and actuation via joint motors. We introduce a parameter $\alpha$ to control the point set construction and the dynamics. This representation allows an articulated rigid object to behave similarly to a deformable object, providing a larger action space to adjust its state and thereby easing the control optimization problem. + +Specifically, for each body of the articulated object, we sample a set of points from the ambient space near the body mesh. The point set $\mathcal{Q}$ is constructed by concatenating all sampled points together. Each point $\mathbf{p}_i\in \mathcal{Q}$ is treated as a mass point with a finite mass $\mathbf{m}_i$ and infinitesimal volume. The dynamics of the point set consist of articulated multi-body dynamics [14, 30], as well as the mass point dynamics of each point $\mathbf{p}_i$ . For each $\mathbf{p}_i$ , we have: + +$$ +m _ {i} \ddot {\mathbf {x}} _ {i} = \mathbf {J} _ {i} \mathbf {u} + \alpha \mathbf {f} _ {i} + \alpha \mathbf {a} _ {i}, \tag {1} +$$ + +where $\mathbf{J}_i$ represents the Jacobian mapping from the generalized velocity to the point velocity $\dot{\mathbf{x}}_i$ , $\mathbf{u}$ denotes the generalized joint force, $\mathbf{f}_i$ accounts for ex + +ternal forces acting on $\mathbf{p}_i$ , and $\mathbf{a}_i \in \mathbb{R}^3$ represents the actuation force applied to the point $\mathbf{p}_i$ . Consequently, the point set is controlled by a shared control in the reduced coordinate space $\mathbf{u}$ and per-point actuation force $\mathbf{a}_i$ . + +Parameterized spring-damper contact modeling. To ease the optimization challenges posed by contact-rich manipulations, which arise from contact constraints such as the non-penetration requirement and Coulomb friction law [3,5], as well as discontinuous dynamics involving frequent contact establishment and breaking, we propose a parameterized contact model for relaxing constraints and controlling the contact behavior. Specifically, we leverage a classical unconstrained spring-damper model [19,35,51,59,64] to model the contacts. This model allows us to flexibly adjust the contact behavior by tuning the contact threshold and the spring stiffness coefficients. Intuitively, a contact model with a high threshold and low spring stiffness presents "soft" behaviors, resulting in a continuous and smooth optimization space. This makes optimization through such a contact model relatively easy. Conversely, a model with a low threshold and large stiffness coefficients will produce "stiff" behaviors, increasing the discontinuity of the optimization space due to frequent contact establishment and breaking. However, it also becomes more physically realistic, meaning contact forces are calculated only when two objects collide, and a large force is applied to separate them if penetrations are observed, thus better satisfying the non-penetration condition. Therefore, by adjusting the contact distance threshold + +![](images/dd5c4074daa60bddaf492a4d97d3815fc002598ecae4e5ba4392d05d6e671f52.jpg) +Fig.3: Point Set can flexibly adjust its states, avoid overfitting to data noise, and ease the difficulty brought by the morphology difference. + +and spring stiffness coefficients, we can modulate the optimizability and fidelity of the contact model. The parameter set of the contact model comprises a distance threshold $d^{c}$ and spring stiffness coefficients. Next, we will delve into the details of the contact establishment, breaking, and force calculations processes. + +Contacts are established between points in the manipulator's point set $\mathcal{Q}$ and the object. A point $\mathbf{p} \in \mathcal{Q}$ is considered to be in "contact" with the object if its signed distance to the object $\operatorname{sd}(\mathbf{p})$ is smaller than the contact distance threshold $d^{c}$ . Subsequently, the object surface point nearest to $\mathbf{p}$ is identified as the corresponding contact point on the object, denoted as $\mathbf{p}^{o}$ . The normal direction of the object point $\mathbf{p}^{o}$ is then determined as the contact normal direction, denoted as $\mathbf{n}^{o}$ . The contact force $\mathbf{f}^{c}$ applied from the manipulator point $\mathbf{p}$ to $\mathbf{p}^{o}$ is calculated as follows: + +$$ +\mathbf {f} ^ {c} = - \left(k ^ {n} d - k ^ {d} d \dot {d}\right) \mathbf {n} ^ {o}, \tag {2} +$$ + +where, $k^n$ represents the spring stiffness coefficient, $k^d$ denotes the damping coefficient, and $d = d^{c} - \mathrm{sd}(\mathbf{p})$ is always positive. To enhance the continuity of $\mathbf{f}^c$ [64], $k^d\ddot{d}\dot{d}$ is used as the magnitude of the damping force, rather than $k^d\dot{d}$ . + +Friction forces are modeled as penalty-based spring forces [3,65]. Once a point $\mathbf{p}$ is identified as in contact with the object, with the object contact point denoted as $\mathbf{p}^o$ , the contact pair is stored. Contact forces between them are continually calculated until the contact breaking conditions are met. In more detail, the static friction force from $\mathbf{p}$ to $\mathbf{p}^o$ is calculated using a spring model: + +$$ +\mathbf {f} _ {s} ^ {f} = k ^ {f} \mathbf {T} _ {n} (\mathbf {p} - \mathbf {p} ^ {o}), \tag {3} +$$ + +where $k^f$ is the friction spring stiffness coefficient, $\mathbf{T}_n = \mathbf{I} - \mathbf{n}^o\mathbf{n}^{oT}$ is a tangential projection operator. When the static friction satisfies $\| \mathbf{f}_s^f\| \leq \mu \| \mathbf{f}^c\|$ , $\mathbf{f}_s^f$ is applied to the object point $\mathbf{p}^o$ . Otherwise, the dynamic friction force is applied, and the contact breaks: + +$$ +\mathbf {f} _ {d} ^ {f} = - \mu \| \mathbf {f} _ {s} ^ {f} \| \frac {\mathbf {T} _ {n} \mathbf {v} _ {\mathbf {p} \leftarrow \mathbf {p} ^ {o}}}{\| \mathbf {T} _ {n} \mathbf {v} _ {\mathbf {p} \leftarrow \mathbf {p} ^ {o}} \|}, \tag {4} +$$ + +where $\mathbf{v}_{\mathbf{p}^{\leftarrow}\mathbf{p}^{o}}$ is the relative velocity between $\mathbf{p}$ and $\mathbf{p}^o$ . + +Parameterized residual physics. The analytical designs facilitate relaxation but may limit the use of highly sophisticated and realistic dynamics models, deviating from real physics. To address this, the final component of our quasi-physical simulator is a flexible neural residual physics model [1,22,41]. + +Specifically, we propose to employ neural networks to learn and predict residual contact forces and friction forces based on contact-related information. For detailed residual contact force prediction, we introduce a local contact network $f_{\psi_{\mathrm{local}}}$ that utilizes contact information identified in the parameterized contact model and predicts residual forces between each contact pair. To address discrepancies in contact region identification between the parameterized contact model and real contact region, we also incorporate a global residual network $f_{\psi_{\mathrm{global}}}$ that predicts residual forces and torques applied directly to the object's center of mass. In more detail, for a given contact pair $(\mathbf{p}, \mathbf{p}^o)$ , the local contact network utilizes contact-related features from the local contact region, comprising geometry, per-point velocity, and per-object point normal. It then maps these + +features to predict the residual contact force and residual friction force between the two points in the contact pair. Additionally, the global residual network incorporates contact-related information from the global contact region, including geometry, per-point velocity, and per-object point normal, as input. It then predicts a residual force and residual torque to be applied to the object's center of mass. Details such as contact region identification and network architectures are deferred to the Supp. We denote the estimizable parameters in the residual physics network as $\psi = (\psi_{\mathrm{global}},\psi_{\mathrm{local}})$ . Through optimization of the residual physics network, we unlock the possibility of introducing highly non-linear dynamics to align our parametrized quasi-physical simulator with any realistic black-box physical simulator. + +Semi-implicit time-stepping is leveraged to make the simulation auto differentiable and easy to combine with neural networks [22]. + +# 3.2 Dexterous Manipulation Transfer via a Physics Curriculum + +Building upon the family of parameterized quasi-physical simulators, we present a solution to the challenging dexterous manipulation transfer problem through a physics curriculum. This curriculum consists of a sequence of parameterized simulators, ranging from those with minimal constraints and the softest contact behavior to increasingly realistic simulators. We address the problem by transferring the manipulation demonstration to the dexterous hand within each simulator across the curriculum progressively. To elaborate further, the optimization process begins within the parameterized simulator where articulated rigid constraints are removed and the contact model is tuned to its softest level. Additionally, the residual physics networks are deactivated. This initial simulator configuration offers a friendly environment for optimization. Subsequently, the physics constraints are gradually tightened as we progress through each simulator within the curriculum. The task is solved iteratively within each simulator. After reaching the most tightened analytical model, the analytical part is fixed and residual networks are activated. The simulator is gradually optimized to approximate the dynamics in a realistic physical environment. Concurrently, the control trajectory $\mathcal{A}$ continues to be refined in the quasi-physical simulator. Finally, we arrive at a simulator optimized to be with high fidelity and a trajectory $\mathcal{A}$ capable of guiding the dexterous hand to accurately track the demonstration within a realistically simulated physical environment. Additionally, since object properties as well as system parameters are unknown from the kinematics-only demonstration, we set them estimizable and identify them (denoted $S$ ) together with optimizing the hand control trajectory. Next we'll illustrate this in detail. + +Transferring human demonstration via point set dynamics. To robustly transfer the human demonstration to a morphologically different dexterous robot hand in simulation and to overcome noise in the kinematic trajectory, we initially relax the articulated rigid constraints and transfer the kinematics human demonstration to the control trajectory of the point set. Specifically, the point set representation with the relaxation parameter $\alpha$ for the dynamic human hand [8] is constructed. The shared control trajectory $\mathcal{A}$ and per-point per-frame actions + +are optimized so that the resulting trajectory of the point set can manipulate the object according to the demonstration. After that, a point set with the same parameter $\alpha$ is constructed to represent the dexterous robot hand. Subsequently, the shared control trajectory $\mathcal{A}$ and per-point per-frame actions are optimized to track the manipulation accordingly. + +Transferring through a contact model curriculum. After that, the articulated rigid constraint is tightened by freezing the point set parameter $\alpha$ to zero. The following optimization starts from a parameterized simulator with the softest contact model. We then gradually tighten the contact model by adjusting its distance threshold, contact force spring stiffness, etc. By curriculum optimizing the trajectory $\mathcal{A}$ and parameters $\mathcal{S}$ in each of the quasi-physical simulators, we finally arrive at the control trajectory that can drive a dexterous hand to accomplish the tracking task in the parameterized simulator with the most tightened analytical model. + +Optimizing towards a realistic physical environment. Subsequently, the residual physics network is activated and the parameterized simulator is optimized to approximate the dynamics in a realistic physical environment. We continue to optimize the hand trajectory in the quasi-physical simulator. Specifically, we leverage the successful trial in model-based human tracking literature [16,66] and iteratively optimize the control trajectory $\mathcal{A}$ and the parameterized simulator. In more detail, the following two subproblems are iteratively solved: 1) optimizing the quasi-physical simulator to approximate the realistic dynamics, and 2) optimizing the control trajectory $\mathcal{A}$ to complete the manipulation in the quasi-physical simulator. Gradient-based optimization is leveraged taking advantage of the differentiability of the parameterized simulator. + +After completing the optimization, the final control trajectory is yielded by model predictive control (MPC) [18] based on the optimized parameterized simulator and the hand trajectory $\mathcal{A}$ . Specifically, in each step, the current and the following controls in several subsequent frames are optimized to reduce the tracking error. More details are deferred to the Supp. + +# 4 Experiments + +We conduct extensive experiments to demonstrate the effectiveness of our method. The evaluation dataset is constructed from three HOI datasets with both single-hand and bimanual manipulations (with rigid objects), with complex manipulations with non-trivial object movements, and rich and changing contacts involved (see Section 4.1). We use Shadow hand [49] and test in two simulators widely used in the embodied AI community: Bullet [9] and Isaac Gym [33]. We compare our method with both model-free approaches and model-based strategies and demonstrate the superiority of our method both quantitatively and qualitatively. We can track complex contact-rich manipulations with large object rotations, back-and-forth object movements, and changing contacts successfully in both of the two simulators, while the best-performed baseline fails (see Section 4.2, Fig. 4). On average, we boost the tracking success rate by $11\%+$ from the previous best-performed (see Section 4.2). We make further analysis and discussions and show that the core philosophy of our work, optimizing through + +![](images/1643c1ef6225e2e5da463e90dde2d591e5feb855311fe0a7f69da54f8a7c43d3.jpg) +Fig. 4: Qualitative comparisons. Please refer to our website and the accompanying video for animated results. + +![](images/a18c830fb7a8710b8a9fb490b5e13cf1260fdd493357208fc08afa10d732cde9.jpg) + +a quasi-physics curriculum, is potentially general and can help improve the performance of a model-free baseline (see Section 4.3). + +# 4.1 Experimental Settings + +Datasets. Our evaluation dataset is compiled from three distinct sources, namely GRAB [53], containing single-hand interactions with daily objects, TACO [32], containing humans manipulating tools, and ARCTIC [13] with bimanual manipulations. For GRAB, we randomly sample a manipulation trajectory for each object. If its manipulation is extremely simple, we additionally sample one trajectory for it. The object is not considered if its corresponding manipulation is bimanual such as binoculars, involves other body parts such as bowl, or with detailed part movements such as the game controller. The number of manipulation sequences from GRAB is 27. For TACO [32], we acquire data by contacting authors. We randomly select one sequence for each right-hand tool object. Sequences with very low quality like erroneous object motions are excluded. 14 trajectories in total are selected finally. For ARCTIC [13], we randomly select one sequence for each object from its available manipulation trajectories, resulting in 10 sequences in total. More details are deferred to the Supp. + +Metrics. We introduce three distinct metrics to assess the quality of object tracking, the accuracy of hand tracking, and the overall success of the tracking task: 1) Per-frame average object rotation error: $R_{\mathrm{err}} = \frac{1}{N}\sum_{n=1}^{N}(1 - (\mathbf{q}_n \cdot \hat{\mathbf{q}}_n))$ , where $\mathbf{q}_n$ is the ground-truth orientation and $\hat{\mathbf{q}}_n$ is the tracked result, represented in quaternion. 2) Per-frame average object translation error: $T_{\mathrm{err}} = \frac{1}{N}\sum_{n=1}^{N}\|\mathbf{t}_n - \hat{\mathbf{t}}_n\|$ , where $\mathbf{t}$ and $\mathbf{t}_n$ are ground-truth and tracked translations respectively. 3) Mean Per-Joint Position Error (MPJPE) = $\frac{1}{N}\sum_{n=1}^{N}\|\mathbf{J}_n - \hat{\mathbf{J}}_n\|$ [20, 44, 58], where $\mathbf{J}_n$ and $\hat{\mathbf{J}}_n$ are keypoints of GT human hand and the simulated robot hand respectively. We manually define the keypoints and the correspondences to the human hand keypoints for the Shadow hand. 4) Per-frame average hand Chamfer Distance: $\mathrm{CD} = \frac{1}{N}\sum_{n=1}^{N}\mathrm{Chamfer-Distance}(\mathbf{H}_n - \hat{\mathbf{H}}_n)$ , for evaluating whether the Shadow hand can "densely" track the demonstration. 5) Success rate: a tracking is regarded as successful if the object rotation error $R_{\mathrm{err}}$ , object translation error $T_{\mathrm{err}}$ , and the hand tracking error MPJPE are smaller than their corresponding threshold. Three success rates are calculated using three different thresholds, namely $10^{\circ} - 10cm - 10cm$ , $15^{\circ} - 15cm - 15cm$ . + +Baselines. We compare with two trends of baselines. For model-free approaches, since there is no prior work with exactly the same problem setting as us, we try to modify and improve a goal-driven rigid object manipulation method DGrasp [8] into two methods for tracking: 1) DGrasp-Base, where the method is almost kept with same with the original DGrasp. We use the first frame where the hand and the object are in contact with each other as the reference frame. Then the policy is trained to grasp the object according to the reference hand and object goal at first. After that, only the root is guided to complete the task. 2) DGraspTracking, where we divide the whole sequence into several subsequences, each of which has 10 frames, and define the end frame of the subsequence as the reference frame. Then the grasping policy is used to guide the hand and gradually track the object according to the hand and the object pose of each reference frame. We improve the DGrasp-Tracking by optimizing the policy through the quasi-physical curriculum and creating "DGrasp-Tracking (w/ Curriculum)" trying to improve its performance. For model-based methods, we compare with Control-VAE [66] and traditional MPC approaches. For Control-VAE, we modify its implementation for the manipulation tracking task. We additionally consider three differentiable physics models to conduct model-predictive control for solving the task. Taking the analytical model with the most tightened contact model as the base model ("MPC (w/ base sim.)"), we further augment it with a general state-of-the-art contact smoothing for robot manipulation [52] and create "MPC (w/ base sim. w/ soften)". Details of baseline models are deferred to the Supp. + +Training and evaluation settings. The physics curriculum is composed of three stages. In the first stage, the parameter $\alpha$ varies from 0.1 to 0.0 and the contact model stiffness is relaxed to the softest level. In the second stage, $\alpha$ is fixed and the contact model stiffness varies from the softest version to the most tightened level gradually through eight stages. Details w.r.t. parameter settings are deferred to the Supp. In the first two stages, we alternately optimize the trajectory $\mathcal{A}$ and parameters $S$ . In each optimization iteration, the $\mathcal{A}$ is optimized + +Table 1: Quantitative evaluations and comparisons to baselines. Bold red numbers for best values and italic blue values for the second best-performed ones. + +
SimulatorMethod\( {R}_{\text{err }}\left( {{}^{ \circ }, \downarrow }\right) \)\( {T}_{\text{err }}\left( {{cm}, \downarrow }\right) \)MPJPE (mm,↓)CD (mm,↓)Success Rate (%,↑)
BulletModel FreeDGrasp-Base44.245.8240.5516.370/13.73/15.69
DGrasp-Tracking44.455.0437.5614.720/15.69/15.69
DGrasp-Tracking (w/ curric.)33.864.6030.4713.537.84/23.53/37.25
Model BasedControl-VAE42.452.7325.2110.940/15.68/23.53
MPC (w/ base sim.)32.563.6724.6210.800/15.68/31.37
MPC (w/ base sim. w/ soften)31.893.6328.2611.310/21.57/37.25
Ours24.211.9724.409.8527.45/37.25/58.82
Isaac GymModel FreeDGrasp-Base36.414.5650.9718.780/7.84/7.84
DGrasp-Tracking44.715.5741.5316.720/0/7.84
DGrasp-Tracking (w/ curric.)38.755.1340.0916.260/23.53/31.37
Model BasedControl-VAE35.404.6127.6313.170/13.73/29.41
MPC (w/ base sim.)37.234.7323.199.750/15.69/31.37
MPC (w/ base sim. w/ soften)36.404.4623.2710.340/9.80/23.53
Ours25.972.0825.3310.3121.57/43.14/56.86
+ +for 100 steps while $S$ is optimized for 1000 steps. In the third stage, $\mathcal{A}$ and $\psi$ are optimized for 256 steps in each iteration. For time-stepping, $\mathrm{dt}$ is set to $5 \times 10^{-4}$ in the parameterized and the target simulators. The articulated multi-body is controlled by joint motors and root velocities in the parameterized quasi-physical simulator while PD control [54] is leveraged in the target simulators. + +# 4.2 Dexterous Manipulating Tracking + +We conducted thorough experiments in two widely used simulators [9, 33]. We treat them as realistic simulated physical environments with high fidelity and wish to track the manipulation in them. In summary, we can control a dexterous hand to complete a wide range of the manipulation tracking tasks with nontrivial object movements and changing contacts. As presented in Table 1, we can achieve significantly higher success rates calculated under three thresholds than the best-performed baseline in both tested simulators. Fig. 4 showcases qualitative examples and comparisons. Please check out our website and video for animated results. + +Complex manipulations. For examples shown in Fig. 4, we can complete the tracking task on examples with large object re-orientations and complicated tool-using (Fig. $(a,b,c)$ ). However, DGrasp-Tracking fails to establish sufficient contact for correctly manipulating the object. In more detail, in Fig. 4(b), the bunny gradually bounced out from its hand in Bullet, while our method does not suffer from this difficulty. In Fig. 4(c), the spoon can be successfully picked up and waved back-and-forth in our method, while DGrasp-Tracking loses the track right from the start. + +Bimanual manipulations. We are also capable of tracking bimanual manipulations. As shown in the example in Fig. 4(d), where two hands collaborate to relocate the object, DGrasp-Tracking fails to accurately track the object, while our method significantly outperforms it. + +# 4.3 Further Analysis and Discussions + +Could model-free methods benefit from the physics curriculum? In addition to the demonstrated merits of our quasi-physical simulators, we further + +Table 2: Ablation studies. Bold red numbers for best values and italic blue values for the second best-performed ones. The simulation environment is Bullet. + +
MethodRerr(°,↓)Terr(cm,↓)MPJPE(mm,↓)CD(mm,↓)Success Rate(%,↑)
Ours w/o Analytical Sim.44.274.3929.8412.910/13.73/25.49
Ours w/o Residual Physics33.693.8126.5710.345.88/23.53/41.18
Ours w/o Local Force NN35.982.9032.8712.440/19.61/35.29
Ours w/o Curriculum42.404.8732.6113.370/17.64/29.41
Ours w/ Curriculum II29.582.3331.6110.2911.76/27.45/50.98
Ours24.211.9724.409.8527.45/37.25/58.82
+ +explore whether model-free strategies can benefit from them. We introduce the "DGrasp-Tracking (w/ Curriculum)" method and compare its performance with the original DGrasp-Tracking model. As shown in Table 1 and the visual comparisons in Fig. 6, the DGrasp-Tracking model indeed benefits from a well-designed physics curriculum. For example, as illustrated in Fig. 6, the curriculum can significantly improve its performance, enabling it to nearly complete challenging tracking tasks where the original version struggles. + +# 5 Ablation Study + +![](images/74859ef9bb99bc24c0df62b177b3f4806a4684985c9b0d55fdbe24355cb4c1c7.jpg) + +![](images/ec93451ebcec063eefb956eefd5a67143f2152306343e6a9965a064f065ccbb5.jpg) +Fig. 5: (a) Qualitative comparisons between our full method and the ablated models; (b) Training loss curve comparisons; (c) Tracking loss curve comparisons. + +![](images/57fbeae52f7988b590d585056da1c3efa289c1cd94a717ddec72ad5be60d0e90.jpg) +Fig. 6: Visual evidence on boosting DGrasp-Tracking's performance via optimizing it through a physics curriculum. + +We conduct a wide range of ablation studies to validate the effectiveness of some of our crucial designs, including the parameterized analytical physics + +model, the parameterized residual physics, the role of the local force network, the necessity of introducing a physics curriculum into the optimization, and how the design on the curriculum stages affects the result. + +Parameterized analytical model. The skeleton of the quasi-physical simulator is an analytical physics model. The intuition is that the parameterized simulator with such physical bias can be optimized towards a realistic simulator more easily than training pure neural networks for approximating. To validate this, we ablate the analytical model and use neural networks to approximate physics in Bullet directly (denoted as "Ours w/o Analytical Sim"). The quantitative (Table 2) and qualitative (Fig. 5) results indicate that the physical biases brought by the analytical model could help the parameterized simulator to learn better physics in the contact-rich scenario. For instance, in the example demonstrated in Fig. 5, the ablated version fails to guide the robot hand to successfully pinch the object in the second figure. + +Parameterized residual physics. To validate the necessity of introducing residual force networks to close the gap between the physics modeled in the parameterized analytical simulator and that of a realistic simulator, we ablate the parameterized force network and create a version named "Ours w/o Residual Physics". Table 2 demonstrated its role in enabling the parameterized simulator to approximate realistic physics models. + +Local residual force network. To adequately leverage state and contact-related information for predicting residual contact forces, we propose to use two types of networks: 1) a local force network for per contact pair residual forces and 2) a global network for additionally compensating. The local network is introduced for fine-grained approximation. We ablate this design and compare the result with our full model to validate this (see Fig. 5 and Table 1). + +Optimizing through an analytical physics curriculum. We further investigate the effectiveness of the analytical curriculum design and how its design influences the result. Specifically, we create two ablated versions: 1) "Ours w/o Curriculum", where the optimization starts directly from the parameterized analytical model with articulated rigid constraints tightened and the stiffest contact model, and 2) "Ours w/ Curriculum II", where we move some stages out from the original curriculum. Table 2 and Fig. 5 demonstrate that both the curriculum and the optimization path will affect the model's performance. + +# 6 Conclusion and Limitations + +In this work, we investigate creating better simulators for solving complex robotic tasks involving complicated dynamics where the previous best-performed optimization strategy fails. We present a family of parameterized quasi-physical simulators that can be both programmed to relax various constraints for task optimization and can be tailored to approximate realistic physics. We tackle the difficult manipulation transfer task via a physics curriculum. + +Limitations. The method is limited by the relatively simple spring-damper model for contact constraint relaxation. Introducing delicate analytical contact models to parameterized simulators is an interesting research direction. + +# Acknowledgement + +We thank Rui Chen, Yiheng Guan, and Taoran Jiang for their valuable support in providing the Allegro hardware and setting up the environment setup during the rebuttal period. + +# References + +1. Ajay, A., Wu, J., Fazeli, N., Bauza, M., Kaelbling, L.P., Tenenbaum, J.B., Rodriguez, A.: Augmenting physical simulators with stochastic neural networks: Case study of planar pushing and bouncing. In: 2018 IEEE/RSJ International Conference on Intelligent Robots and Systems (IROS). pp. 3066-3073. IEEE (2018) +2. Akkaya, I., Andrychowicz, M., Chociej, M., Litwin, M., McGrew, B., Petron, A., Paino, A., Plappert, M., Powell, G., Ribas, R., et al.: Solving rubik's cube with a robot hand. arXiv preprint arXiv:1910.07113 (2019) +3. Andrews, S., Erleben, K., Ferguson, Z.: Contact and friction simulation for computer graphics. In: ACM SIGGRAPH 2022 Courses, pp. 1-172 (2022) +4. Andrychowicz, O.M., Baker, B., Chociej, M., Jozefowicz, R., McGrew, B., Pachocki, J., Petron, A., Plappert, M., Powell, G., Ray, A., et al.: Learning dexterous in-hand manipulation. The International Journal of Robotics Research 39(1), 3-20 (2020) +5. Baraff, D.: An introduction to physically based modeling: rigid body simulation ii—nonpenetration constraints. SIGGRAPH course notes pp. D31–D68 (1997) +6. Chen, T., Tippur, M., Wu, S., Kumar, V., Adelson, E., Agrawal, P.: Visual dexterity: In-hand reorientation of novel and complex object shapes. Science Robotics 8(84), eadc9244 (2023). https://doi.org/10.1126/scirobotics.adc9244, https://www.science.org/doi/abs/10.1126/scirobotics.adc9244 +7. Chen, T., Xu, J., Agrawal, P.: A system for general in-hand object re-orientation. Conference on Robot Learning (2021) +8. Christen, S., Kocabas, M., Aksan, E., Hwangbo, J., Song, J., Hilliges, O.: D-grasp: Physically plausible dynamic grasp synthesis for hand-object interactions. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 20577-20586 (2022) +9. Coumans, E., Bai, Y.: Pybullet, a python module for physics simulation for games, robotics and machine learning (2016) +0. Deng, Y., Yu, H.X., Wu, J., Zhu, B.: Learning vortex dynamics for fluid inference and prediction. arXiv preprint arXiv:2301.11494 (2023) +1. Du, T.: Deep learning for physics simulation. In: ACM SIGGRAPH 2023 Courses, pp. 1-25 (2023) +2. Dunlavy, D.M., O'Leary, D.P.: Homotopy optimization methods for global optimization. Tech. rep., Sandia National Laboratories (SNL), Albuquerque, NM, and Livermore, CA ... (2005) +13. Fan, Z., Taheri, O., Tzionas, D., Kocabas, M., Kaufmann, M., Black, M.J., Hilliges, O.: ARCTIC: A dataset for dexterous bimanual hand-object manipulation. In: Proceedings IEEE Conference on Computer Vision and Pattern Recognition (CVPR) (2023) +14. Featherstone, R.: Rigid body dynamics algorithms (2007), https://apisemantic scholar.org/CorpusID:58437819 + +15. Freeman, C.D., Frey, E., Raichuk, A., Girgin, S., Mordatch, I., Bachem, O.: Brax-a differentiable physics engine for large scale rigid body simulation. arXiv preprint arXiv:2106.13281 (2021) +16. Fussell, L., Bergamin, K., Holden, D.: Supertrack: Motion tracking for physically simulated characters using supervised learning. ACM Transactions on Graphics (TOG) 40(6), 1-13 (2021) +17. Gao, J., Michelis, M.Y., Spielberg, A., Katzschmann, R.K.: Sim-to-real of soft robots with learned residual physics. arXiv preprint arXiv:2402.01086 (2024) +18. Garcia, C.E., Prett, D.M., Morari, M.: Model predictive control: Theory and practice—a survey. Automatica 25(3), 335-348 (1989) +19. Geilinger, M., Hahn, D., Zehnder, J., Bächer, M., Thomaszewski, B., Coros, S.: Add: Analytically differentiable dynamics for multi-body systems with frictional contact. ACM Transactions on Graphics (TOG) 39(6), 1-15 (2020) +20. Grandia, R., Farshidian, F., Knoop, E., Schumacher, C., Hutter, M., Bächer, M.: Doc: Differentiable optimal control for retargeting motions onto legged robots. ACM Transactions on Graphics (TOG) 42(4), 1-14 (2023) +21. Gupta, A., Eppner, C., Levine, S., Abbeel, P.: Learning dexterous manipulation for a soft robotic hand from human demonstrations. In: 2016 IEEE/RSJ International Conference on Intelligent Robots and Systems (IROS). pp. 3786-3793. IEEE (2016) +22. Heiden, E., Millard, D., Coumans, E., Sheng, Y., Sukhatme, G.S.: Neuralsim: Augmenting differentiable simulators with neural networks. In: 2021 IEEE International Conference on Robotics and Automation (ICRA). pp. 9474-9481. IEEE (2021) +23. Howell, T.A., Le Cleac'h, S., Kolter, J.Z., Schwager, M., Manchester, Z.: Dojo: A differentiable simulator for robotics. arXiv preprint arXiv:2203.00806 9 (2022) +24. Howell, T.A., Le Cleac'h, S., Singh, S., Florence, P., Manchester, Z., Sindhwani, V.: Trajectory optimization with optimization-based dynamics. IEEE Robotics and Automation Letters 7(3), 6750-6757 (2022) +25. Hwangbo, J., Lee, J., Hutter, M.: Per-contact iteration method for solving contact dynamics. IEEE Robotics and Automation Letters 3(2), 895-902 (2018) +26. Lan, L., Yang, Y., Kaufman, D., Yao, J., Li, M., Jiang, C.: Medial ipc: accelerated incremental potential contact with medial elastics. ACM Transactions on Graphics 40(4) (2021) +27. Li, M., Ferguson, Z., Schneider, T., Langlois, T.R., Zorin, D., Panozzo, D., Jiang, C., Kaufman, D.M.: Incremental potential contact: intersection-and inversion-free, large-deformation dynamics. ACM Trans. Graph. 39(4), 49 (2020) +28. Liao, S.: On the homotopy analysis method for nonlinear problems. Applied mathematics and computation 147(2), 499-513 (2004) +29. Lin, X., Yang, Z., Zhang, X., Zhang, Q.: Continuation path learning for homotopy optimization (2023) +30. Liu, C.K., Jain, S.: A quick tutorial on multibody dynamics. Online tutorial, June p. 7 (2012) +31. Liu, X., Pathak, D., Kitani, K.M.: Herd: Continuous human-to-robot evolution for learning from human demonstration. arXiv preprint arXiv:2212.04359 (2022) +32. Liu, Y., Yang, H., Si, X., Liu, L., Li, Z., Zhang, Y., Liu, Y., Yi, L.: Taco: Benchmarking generalizable bimanual tool-action-object understanding. arXiv preprint arXiv:2401.08399 (2024) +33. Makoviychuk, V., Wawrzyniak, L., Guo, Y., Lu, M., Storey, K., Macklin, M., Hoeller, D., Rudin, N., Allshire, A., Handa, A., et al.: Isaac gym: High performancegpu-based physics simulation for robot learning. arXiv preprint arXiv:2108.10470 (2021) + +34. Mandikal, P., Grauman, K.: Dexvip: Learning dexterous grasping with human hand pose priors from video. In: Conference on Robot Learning. pp. 651-661. PMLR (2022) +35. Marcucci, T., Gabiccini, M., Artoni, A.: A two-stage trajectory optimization strategy for articulated bodies with unscheduled contact sequences. IEEE Robotics and Automation Letters 2(1), 104-111 (2016) +36. Mordatch, I., Popovic, Z., Todorov, E.: Contact-invariant optimization for hand manipulation. In: Proceedings of the ACM SIGGRAPH/Eurographics symposium on computer animation. pp. 137-144 (2012) +37. Pang, T., Suh, H.T., Yang, L., Tedrake, R.: Global planning for contact-rich manipulation via local smoothing of quasi-dynamic contact models. IEEE Transactions on Robotics (2023) +38. Pang, T., Tedrake, R.: A convex quasistatic time-stepping scheme for rigid multi-body systems with contact and friction. In: 2021 IEEE International Conference on Robotics and Automation (ICRA). pp. 6614-6620. IEEE (2021) +39. Peng, X.B., Abbeel, P., Levine, S., Van de Panne, M.: Deepmimic: Example-guided deep reinforcement learning of physics-based character skills. ACM Transactions On Graphics (TOG) 37(4), 1-14 (2018) +40. Pfaff, T., Fortunato, M., Sanchez-Gonzalez, A., Battaglia, P.W.: Learning mesh-based simulation with graph networks. arXiv preprint arXiv:2010.03409 (2020) +41. Pfrommer, S., Halm, M., Posa, M.: Contactnets: Learning discontinuous contact dynamics with smooth, implicit representations. In: Conference on Robot Learning. pp. 2279-2291. PMLR (2021) +42. Qin, Y., Su, H., Wang, X.: From one hand to multiple hands: Imitation learning for dexterous manipulation from single-camera teleoperation. IEEE Robotics and Automation Letters 7(4), 10873-10881 (2022) +43. Qin, Y., Wu, Y.H., Liu, S., Jiang, H., Yang, R., Fu, Y., Wang, X.: Dexamv: Imitation learning for dexterous manipulation from human videos. In: European Conference on Computer Vision. pp. 570-587. Springer (2022) +44. Qin, Y., Yang, W., Huang, B., Van Wyk, K., Su, H., Wang, X., Chao, Y.W., Fox, D.: Anyteleop: A general vision-based dexterous robot arm-hand teleoperation system. arXiv preprint arXiv:2307.04577 (2023) +45. Radosavovic, I., Wang, X., Pinto, L., Malik, J.: State-only imitation learning for dexterous manipulation. In: 2021 IEEE/RSJ International Conference on Intelligent Robots and Systems (IROS). pp. 7865-7871. IEEE (2021) +46. Rajeswaran, A., Kumar, V., Gupta, A., Vezzani, G., Schulman, J., Todorov, E., Levine, S.: Learning complex dexterous manipulation with deep reinforcement learning and demonstrations. arXiv preprint arXiv:1709.10087 (2017) +47. Rusu, A.A., Colmenarejo, S.G., Gulcehre, C., Desjardins, G., Kirkpatrick, J., Pascanu, R., Mnih, V., Kavukcuoglu, K., Hadsell, R.: Policy distillation. arXiv preprint arXiv:1511.06295 (2015) +48. Schmeckpeper, K., Rybkin, O., Daniilidis, K., Levine, S., Finn, C.: Reinforcement learning with videos: Combining offline observations with interaction. arXiv preprint arXiv:2011.06507 (2020) +49. ShadowRobot: Shadowrobot dexterous hand (2005), https://www.shadowrobot.com/dexterous-hand-series/ +50. Siahkoohi, A., Louboutin, M., Herrmann, F.J.: Neural network augmented wave-equation simulation. arXiv preprint arXiv:1910.00925 (2019) +51. Suh, H., Wang, Y.: Comparing effectiveness of relaxation methods for warm starting trajectory optimization through soft contact (2019) + +52. Suh, H.J.T., Pang, T., Tedrake, R.: Bundled gradients through contact via randomized smoothing. IEEE Robotics and Automation Letters 7(2), 4000-4007 (2022) +53. Taheri, O., Ghorbani, N., Black, M.J., Tzionas, D.: Grab: A dataset of whole-body human grasping of objects. In: Computer Vision-ECCV 2020: 16th European Conference, Glasgow, UK, August 23-28, 2020, Proceedings, Part IV 16. pp. 581-600. Springer (2020) +54. Tan, K.K., Wang, Q.G., Hang, C.C.: Advances in PID control. Springer Science & Business Media (2012) +55. Tedrake, R., the Drake Development Team: Drake: Model-based design and verification for robotics (2019), https://drake.mit.edu +56. Todorov, E., Erez, T., Tassa, Y.: Mujoco: A physics engine for model-based control. In: 2012 IEEE/RSJ international conference on intelligent robots and systems. pp. 5026-5033. IEEE (2012) +57. Traore, R., Caseles-Dupré, H., Lesort, T., Sun, T., Díaz-Rodriguez, N., Filliat, D.: Continual reinforcement learning deployed in real-life using policy distillation and sim2real transfer. arXiv preprint arXiv:1906.04452 (2019) +58. Villegas, R., Ceylan, D., Hertzmann, A., Yang, J., Saito, J.: Contact-aware retargeting of skinned motion. In: Proceedings of the IEEE/CVF International Conference on Computer Vision. pp. 9720-9729 (2021) +59. Wang, Y., Weidner, N.J., Baxter, M.A., Hwang, Y., Kaufman, D.M., Sueda, S.: Redmax: Efficient & flexible approach for articulated dynamics. ACM Transactions on Graphics (TOG) 38(4), 1-10 (2019) +60. Wang, Y., Lin, J., Zeng, A., Luo, Z., Zhang, J., Zhang, L.: Physhoi: Physics-based imitation of dynamic human-object interaction. arXiv preprint arXiv:2312.04393 (2023) +61. Watson, L.T., Haftka, R.T.: Modern homotopy methods in optimization. Computer Methods in Applied Mechanics and Engineering 74(3), 289-305 (1989) +62. Wu, Y.H., Wang, J., Wang, X.: Learning generalizable dexterous manipulation from human grasp affordance. In: Conference on Robot Learning. pp. 618-629. PMLR (2023) +63. Xiong, S., He, X., Tong, Y., Deng, Y., Zhu, B.: Neural vortex method: from finite lagrangian particles to infinite dimensional eulerian dynamics. Computers & Fluids 258, 105811 (2023) +64. Xu, J., Chen, T., Zlokapa, L., Foshey, M., Matusik, W., Sueda, S., Agrawal, P.: An end-to-end differentiable framework for contact-aware robot design. arXiv preprint arXiv:2107.07501 (2021) +65. Yamane, K., Nakamura, Y.: Stable penalty-based model of frictional contacts. In: Proceedings 2006 IEEE International Conference on Robotics and Automation, 2006. ICRA 2006. pp. 1904-1909. IEEE (2006) +66. Yao, H., Song, Z., Chen, B., Liu, L.: Controlvae: Model-based learning of generative controllers for physics-based characters. ACM Transactions on Graphics (TOG) 41(6), 1-16 (2022) +67. Zeng, A., Song, S., Lee, J., Rodriguez, A., Funkhouser, T.: Tossingbot: Learning to throw arbitrary objects with residual physics. IEEE Transactions on Robotics 36(4), 1307-1319 (2020) +68. Zhang, H., Christen, S., Fan, Z., Zheng, L., Hwangbo, J., Song, J., Hilliges, O.: Artigrasp: Physically plausible synthesis of bi-manual dexterous grasping and articulation. arXiv preprint arXiv:2309.03891 (2023) +69. Zhang, S., Liu, B., Wang, Z., Zhao, T.: Model-based reparameterization policy gradient methods: Theory and practical algorithms. Advances in Neural Information Processing Systems 36 (2024) + +70. Zhang, Y., Clegg, A., Ha, S., Turk, G., Ye, Y.: Learning to transfer in-hand manipulations using a greedy shape curriculum. In: Computer Graphics Forum. vol. 42, pp. 25-36. Wiley Online Library (2023) +71. Zhao, W., Queralta, J.P., Westerlund, T.: Sim-to-real transfer in deep reinforcement learning for robotics: a survey. In: 2020 IEEE symposium series on computational intelligence (SSCI). pp. 737-744. IEEE (2020) +72. Zhuang, F., Qi, Z., Duan, K., Xi, D., Zhu, Y., Zhu, H., Xiong, H., He, Q.: A comprehensive survey on transfer learning. Proceedings of the IEEE 109(1), 43-76 (2020) \ No newline at end of file diff --git a/2024/Parameterized Quasi-Physical Simulators for Dexterous Manipulations Transfer/images.zip b/2024/Parameterized Quasi-Physical Simulators for Dexterous Manipulations Transfer/images.zip new file mode 100644 index 0000000000000000000000000000000000000000..ae8547d6378b10a3f0e9acb2fc787a9e5487bb8d --- /dev/null +++ b/2024/Parameterized Quasi-Physical Simulators for Dexterous Manipulations Transfer/images.zip @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:bce0c2314bc50737951c4cfff19610035be3472e80c152cb3eeea14cff8260ab +size 438147 diff --git a/2024/Parameterized Quasi-Physical Simulators for Dexterous Manipulations Transfer/layout.json b/2024/Parameterized Quasi-Physical Simulators for Dexterous Manipulations Transfer/layout.json new file mode 100644 index 0000000000000000000000000000000000000000..98406575f0f6fe14712a8b1f7c6892a25b1dee5b --- /dev/null +++ b/2024/Parameterized Quasi-Physical Simulators for Dexterous Manipulations Transfer/layout.json @@ -0,0 +1,10660 @@ +{ + "pdf_info": [ + { + "para_blocks": [ + { + "bbox": [ + 150, + 111, + 463, + 148 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 150, + 111, + 463, + 148 + ], + "spans": [ + { + "bbox": [ + 150, + 111, + 463, + 148 + ], + "type": "text", + "content": "Parameterized Quasi-Physical Simulators for Dexterous Manipulations Transfer" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 139, + 168, + 474, + 182 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 139, + 168, + 474, + 182 + ], + "spans": [ + { + "bbox": [ + 139, + 168, + 474, + 182 + ], + "type": "text", + "content": "Xueyi Liu" + }, + { + "bbox": [ + 139, + 168, + 474, + 182 + ], + "type": "inline_equation", + "content": "^{1,3}" + }, + { + "bbox": [ + 139, + 168, + 474, + 182 + ], + "type": "text", + "content": ", Kangbo Lyu" + }, + { + "bbox": [ + 139, + 168, + 474, + 182 + ], + "type": "inline_equation", + "content": "^{1}" + }, + { + "bbox": [ + 139, + 168, + 474, + 182 + ], + "type": "text", + "content": ", Jieqiong Zhang" + }, + { + "bbox": [ + 139, + 168, + 474, + 182 + ], + "type": "inline_equation", + "content": "^{1}" + }, + { + "bbox": [ + 139, + 168, + 474, + 182 + ], + "type": "text", + "content": ", Tao Du" + }, + { + "bbox": [ + 139, + 168, + 474, + 182 + ], + "type": "inline_equation", + "content": "^{1,2,3}" + }, + { + "bbox": [ + 139, + 168, + 474, + 182 + ], + "type": "text", + "content": ", and Li Yi" + }, + { + "bbox": [ + 139, + 168, + 474, + 182 + ], + "type": "inline_equation", + "content": "^{1,2,3}" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 145, + 190, + 468, + 214 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 145, + 190, + 468, + 214 + ], + "spans": [ + { + "bbox": [ + 145, + 190, + 468, + 214 + ], + "type": "inline_equation", + "content": "^{1}" + }, + { + "bbox": [ + 145, + 190, + 468, + 214 + ], + "type": "text", + "content": " Tsinghua University " + }, + { + "bbox": [ + 145, + 190, + 468, + 214 + ], + "type": "inline_equation", + "content": "^{2}" + }, + { + "bbox": [ + 145, + 190, + 468, + 214 + ], + "type": "text", + "content": " Shanghai AI Laboratory " + }, + { + "bbox": [ + 145, + 190, + 468, + 214 + ], + "type": "inline_equation", + "content": "^{3}" + }, + { + "bbox": [ + 145, + 190, + 468, + 214 + ], + "type": "text", + "content": " Shanghai Qi Zhi Institute https://meowu7.github.io/QuasiSim" + } + ] + } + ], + "index": 2 + }, + { + "type": "image", + "bbox": [ + 151, + 251, + 191, + 291 + ], + "blocks": [ + { + "bbox": [ + 151, + 251, + 191, + 291 + ], + "lines": [ + { + "bbox": [ + 151, + 251, + 191, + 291 + ], + "spans": [ + { + "bbox": [ + 151, + 251, + 191, + 291 + ], + "type": "image", + "image_path": "9ad5b625ad6097a9287a6f84a453606d61ed514863586c2d9313a99e550120d8.jpg" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_body" + } + ], + "index": 3 + }, + { + "type": "image", + "bbox": [ + 152, + 292, + 190, + 330 + ], + "blocks": [ + { + "bbox": [ + 152, + 292, + 190, + 330 + ], + "lines": [ + { + "bbox": [ + 152, + 292, + 190, + 330 + ], + "spans": [ + { + "bbox": [ + 152, + 292, + 190, + 330 + ], + "type": "image", + "image_path": "d46ebd66664a8e9824b4bc4a38ff0319c5a628c34a96148c2b8581e84cac049c.jpg" + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "image_body" + } + ], + "index": 4 + }, + { + "type": "image", + "bbox": [ + 191, + 251, + 228, + 291 + ], + "blocks": [ + { + "bbox": [ + 191, + 251, + 228, + 291 + ], + "lines": [ + { + "bbox": [ + 191, + 251, + 228, + 291 + ], + "spans": [ + { + "bbox": [ + 191, + 251, + 228, + 291 + ], + "type": "image", + "image_path": "944ad22b523cc633929f6e4f51084547ad0b17b00bec9e06fb856b317d937ab6.jpg" + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "image_body" + } + ], + "index": 5 + }, + { + "type": "image", + "bbox": [ + 191, + 292, + 228, + 331 + ], + "blocks": [ + { + "bbox": [ + 191, + 292, + 228, + 331 + ], + "lines": [ + { + "bbox": [ + 191, + 292, + 228, + 331 + ], + "spans": [ + { + "bbox": [ + 191, + 292, + 228, + 331 + ], + "type": "image", + "image_path": "fd6cc2d5e7a115c6c7fdfee25e073dbf68c4df1eb5093be9e7230c23ea8d04d2.jpg" + } + ] + } + ], + "index": 6, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 153, + 331, + 229, + 338 + ], + "lines": [ + { + "bbox": [ + 153, + 331, + 229, + 338 + ], + "spans": [ + { + "bbox": [ + 153, + 331, + 229, + 338 + ], + "type": "text", + "content": "(e) Result of a failed baseline" + } + ] + } + ], + "index": 19, + "angle": 0, + "type": "image_caption" + } + ], + "index": 6 + }, + { + "type": "image", + "bbox": [ + 229, + 251, + 266, + 291 + ], + "blocks": [ + { + "bbox": [ + 229, + 251, + 266, + 291 + ], + "lines": [ + { + "bbox": [ + 229, + 251, + 266, + 291 + ], + "spans": [ + { + "bbox": [ + 229, + 251, + 266, + 291 + ], + "type": "image", + "image_path": "0edadcc3e9758ed79a8aad322965386e9e91d087ce242cec0c73fc29b704884b.jpg" + } + ] + } + ], + "index": 7, + "angle": 0, + "type": "image_body" + } + ], + "index": 7 + }, + { + "type": "image", + "bbox": [ + 229, + 292, + 266, + 331 + ], + "blocks": [ + { + "bbox": [ + 229, + 292, + 266, + 331 + ], + "lines": [ + { + "bbox": [ + 229, + 292, + 266, + 331 + ], + "spans": [ + { + "bbox": [ + 229, + 292, + 266, + 331 + ], + "type": "image", + "image_path": "75270e327eb3aba02f2eda6843c5fa10cf480ab83aa757109c9587df0210a50e.jpg" + } + ] + } + ], + "index": 8, + "angle": 0, + "type": "image_body" + } + ], + "index": 8 + }, + { + "type": "image", + "bbox": [ + 266, + 251, + 304, + 291 + ], + "blocks": [ + { + "bbox": [ + 266, + 251, + 304, + 291 + ], + "lines": [ + { + "bbox": [ + 266, + 251, + 304, + 291 + ], + "spans": [ + { + "bbox": [ + 266, + 251, + 304, + 291 + ], + "type": "image", + "image_path": "ba69399690ce426d9287bc1f75cc59f3aa880cec7b32d13d6b10e1ebbbdad237.jpg" + } + ] + } + ], + "index": 9, + "angle": 0, + "type": "image_body" + } + ], + "index": 9 + }, + { + "type": "image", + "bbox": [ + 266, + 292, + 304, + 331 + ], + "blocks": [ + { + "bbox": [ + 266, + 292, + 304, + 331 + ], + "lines": [ + { + "bbox": [ + 266, + 292, + 304, + 331 + ], + "spans": [ + { + "bbox": [ + 266, + 292, + 304, + 331 + ], + "type": "image", + "image_path": "8180f3da0aa5fa07e0eb5556a305879fd99051828dee4246b4f7c99c53a86c40.jpg" + } + ] + } + ], + "index": 10, + "angle": 0, + "type": "image_body" + } + ], + "index": 10 + }, + { + "type": "image", + "bbox": [ + 306, + 251, + 344, + 291 + ], + "blocks": [ + { + "bbox": [ + 306, + 251, + 344, + 291 + ], + "lines": [ + { + "bbox": [ + 306, + 251, + 344, + 291 + ], + "spans": [ + { + "bbox": [ + 306, + 251, + 344, + 291 + ], + "type": "image", + "image_path": "30e7418e3123a22bbac62a8bd8aa308e845bd78284e41a20639982258f594a57.jpg" + } + ] + } + ], + "index": 11, + "angle": 0, + "type": "image_body" + } + ], + "index": 11 + }, + { + "type": "image", + "bbox": [ + 306, + 292, + 344, + 331 + ], + "blocks": [ + { + "bbox": [ + 306, + 292, + 344, + 331 + ], + "lines": [ + { + "bbox": [ + 306, + 292, + 344, + 331 + ], + "spans": [ + { + "bbox": [ + 306, + 292, + 344, + 331 + ], + "type": "image", + "image_path": "6abae0fc3a34145058f86b59fa3c5e861daf85c2165afab851543ddc36323164.jpg" + } + ] + } + ], + "index": 12, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 307, + 331, + 456, + 338 + ], + "lines": [ + { + "bbox": [ + 307, + 331, + 456, + 338 + ], + "spans": [ + { + "bbox": [ + 307, + 331, + 456, + 338 + ], + "type": "text", + "content": "(f) Improved result via optimizing through the curriculum" + } + ] + } + ], + "index": 28, + "angle": 0, + "type": "image_caption" + } + ], + "index": 12 + }, + { + "type": "image", + "bbox": [ + 345, + 251, + 383, + 291 + ], + "blocks": [ + { + "bbox": [ + 345, + 251, + 383, + 291 + ], + "lines": [ + { + "bbox": [ + 345, + 251, + 383, + 291 + ], + "spans": [ + { + "bbox": [ + 345, + 251, + 383, + 291 + ], + "type": "image", + "image_path": "85bb9f79b34813de8c619a59a0b4a518693688115132bf53a6a90353a905cb84.jpg" + } + ] + } + ], + "index": 13, + "angle": 0, + "type": "image_body" + } + ], + "index": 13 + }, + { + "type": "image", + "bbox": [ + 345, + 292, + 383, + 331 + ], + "blocks": [ + { + "bbox": [ + 345, + 292, + 383, + 331 + ], + "lines": [ + { + "bbox": [ + 345, + 292, + 383, + 331 + ], + "spans": [ + { + "bbox": [ + 345, + 292, + 383, + 331 + ], + "type": "image", + "image_path": "5437d53ea46bd32cc5ebeb033b96f4ee8e0d47af725c3fe7e2638def7f06ee5e.jpg" + } + ] + } + ], + "index": 14, + "angle": 0, + "type": "image_body" + } + ], + "index": 14 + }, + { + "type": "image", + "bbox": [ + 383, + 251, + 421, + 291 + ], + "blocks": [ + { + "bbox": [ + 383, + 251, + 421, + 291 + ], + "lines": [ + { + "bbox": [ + 383, + 251, + 421, + 291 + ], + "spans": [ + { + "bbox": [ + 383, + 251, + 421, + 291 + ], + "type": "image", + "image_path": "743909a9a222135261106888ea15fd109ce2bb572622c6a49cc5bcbd31597076.jpg" + } + ] + } + ], + "index": 15, + "angle": 0, + "type": "image_body" + } + ], + "index": 15 + }, + { + "type": "image", + "bbox": [ + 383, + 292, + 421, + 331 + ], + "blocks": [ + { + "bbox": [ + 383, + 292, + 421, + 331 + ], + "lines": [ + { + "bbox": [ + 383, + 292, + 421, + 331 + ], + "spans": [ + { + "bbox": [ + 383, + 292, + 421, + 331 + ], + "type": "image", + "image_path": "8e7d30b22609c28a925427059d560e03a1b28788eb4f3e5c6770973167b182c7.jpg" + } + ] + } + ], + "index": 16, + "angle": 0, + "type": "image_body" + } + ], + "index": 16 + }, + { + "type": "image", + "bbox": [ + 421, + 251, + 458, + 291 + ], + "blocks": [ + { + "bbox": [ + 421, + 251, + 458, + 291 + ], + "lines": [ + { + "bbox": [ + 421, + 251, + 458, + 291 + ], + "spans": [ + { + "bbox": [ + 421, + 251, + 458, + 291 + ], + "type": "image", + "image_path": "e6b018b2d8a9175a37b4e174c9e694f9100858b74971cbcc8f106f8846623b26.jpg" + } + ] + } + ], + "index": 17, + "angle": 0, + "type": "image_body" + } + ], + "index": 17 + }, + { + "type": "image", + "bbox": [ + 421, + 292, + 458, + 331 + ], + "blocks": [ + { + "bbox": [ + 421, + 292, + 458, + 331 + ], + "lines": [ + { + "bbox": [ + 421, + 292, + 458, + 331 + ], + "spans": [ + { + "bbox": [ + 421, + 292, + 458, + 331 + ], + "type": "image", + "image_path": "201868126a183c1936cb9a37a5bdb7a6a4a612fca5d8c4b0dfd96842d278fbad.jpg" + } + ] + } + ], + "index": 18, + "angle": 0, + "type": "image_body" + } + ], + "index": 18 + }, + { + "type": "image", + "bbox": [ + 153, + 338, + 190, + 376 + ], + "blocks": [ + { + "bbox": [ + 153, + 338, + 190, + 376 + ], + "lines": [ + { + "bbox": [ + 153, + 338, + 190, + 376 + ], + "spans": [ + { + "bbox": [ + 153, + 338, + 190, + 376 + ], + "type": "image", + "image_path": "4e2e798039f97c7bfd5069933f0bd79e6bd9944792e9835348a69fa3084f0c34.jpg" + } + ] + } + ], + "index": 20, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 153, + 377, + 176, + 384 + ], + "lines": [ + { + "bbox": [ + 153, + 377, + 176, + 384 + ], + "spans": [ + { + "bbox": [ + 153, + 377, + 176, + 384 + ], + "type": "text", + "content": "Timestamp" + } + ] + } + ], + "index": 21, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 130, + 384, + 482, + 439 + ], + "lines": [ + { + "bbox": [ + 130, + 384, + 482, + 439 + ], + "spans": [ + { + "bbox": [ + 130, + 384, + 482, + 439 + ], + "type": "text", + "content": "Fig. 1: By optimizing through a quasi-physical simulator curriculum, we successfully transfer human demonstrations to dexterous robot hand simulations. We enable accurate tracking of complex manipulations with changing contacts (Fig. (a)), nontrivial object motions (Fig. (b)) and intricate tool-using (Fig. (c,d)). Besides, our physics curriculum can substantially improve a failed baseline (Fig. (e,f))." + } + ] + } + ], + "index": 36, + "angle": 0, + "type": "image_caption" + } + ], + "index": 20 + }, + { + "type": "image", + "bbox": [ + 191, + 338, + 228, + 376 + ], + "blocks": [ + { + "bbox": [ + 191, + 338, + 228, + 376 + ], + "lines": [ + { + "bbox": [ + 191, + 338, + 228, + 376 + ], + "spans": [ + { + "bbox": [ + 191, + 338, + 228, + 376 + ], + "type": "image", + "image_path": "bf61530f9eee1b1889c739226cb4cedc69b104d8e13e63757eeac42440152d35.jpg" + } + ] + } + ], + "index": 22, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 191, + 377, + 228, + 379 + ], + "lines": [ + { + "bbox": [ + 191, + 377, + 228, + 379 + ], + "spans": [ + { + "bbox": [ + 191, + 377, + 228, + 379 + ], + "type": "text", + "content": "__________" + } + ] + } + ], + "index": 23, + "angle": 0, + "type": "image_caption" + } + ], + "index": 22 + }, + { + "type": "image", + "bbox": [ + 229, + 338, + 266, + 376 + ], + "blocks": [ + { + "bbox": [ + 229, + 338, + 266, + 376 + ], + "lines": [ + { + "bbox": [ + 229, + 338, + 266, + 376 + ], + "spans": [ + { + "bbox": [ + 229, + 338, + 266, + 376 + ], + "type": "image", + "image_path": "258c6c475945a5eb1669ade573c7825ed880287a852285635fb55a5b7e0b85d7.jpg" + } + ] + } + ], + "index": 24, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 229, + 377, + 266, + 379 + ], + "lines": [ + { + "bbox": [ + 229, + 377, + 266, + 379 + ], + "spans": [ + { + "bbox": [ + 229, + 377, + 266, + 379 + ], + "type": "inline_equation", + "content": "\\left( {x - {2x}}\\right) t - x{y}^{2} = \\left( {x - {2x}}\\right) {f}^{\\prime }t" + } + ] + } + ], + "index": 25, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 266, + 377, + 304, + 384 + ], + "lines": [ + { + "bbox": [ + 266, + 377, + 304, + 384 + ], + "spans": [ + { + "bbox": [ + 266, + 377, + 304, + 384 + ], + "type": "text", + "content": "#" + } + ] + } + ], + "index": 27, + "angle": 0, + "type": "image_caption" + } + ], + "index": 24 + }, + { + "type": "image", + "bbox": [ + 266, + 338, + 304, + 376 + ], + "blocks": [ + { + "bbox": [ + 266, + 338, + 304, + 376 + ], + "lines": [ + { + "bbox": [ + 266, + 338, + 304, + 376 + ], + "spans": [ + { + "bbox": [ + 266, + 338, + 304, + 376 + ], + "type": "image", + "image_path": "5acaa4e87ba70f059827886562ffbcb1bbe5eae44aaea88c37a916554423f4aa.jpg" + } + ] + } + ], + "index": 26, + "angle": 0, + "type": "image_body" + } + ], + "index": 26 + }, + { + "type": "image", + "bbox": [ + 307, + 338, + 344, + 376 + ], + "blocks": [ + { + "bbox": [ + 307, + 338, + 344, + 376 + ], + "lines": [ + { + "bbox": [ + 307, + 338, + 344, + 376 + ], + "spans": [ + { + "bbox": [ + 307, + 338, + 344, + 376 + ], + "type": "image", + "image_path": "4ee93e74bf06383554440472a7d95bc98360bdb5445b376563f56691ecc6a397.jpg" + } + ] + } + ], + "index": 29, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 307, + 377, + 331, + 384 + ], + "lines": [ + { + "bbox": [ + 307, + 377, + 331, + 384 + ], + "spans": [ + { + "bbox": [ + 307, + 377, + 331, + 384 + ], + "type": "text", + "content": "Timestamp" + } + ] + } + ], + "index": 30, + "angle": 0, + "type": "image_caption" + } + ], + "index": 29 + }, + { + "type": "image", + "bbox": [ + 345, + 338, + 383, + 376 + ], + "blocks": [ + { + "bbox": [ + 345, + 338, + 383, + 376 + ], + "lines": [ + { + "bbox": [ + 345, + 338, + 383, + 376 + ], + "spans": [ + { + "bbox": [ + 345, + 338, + 383, + 376 + ], + "type": "image", + "image_path": "8663ea3d05dff9f3f1b53eec11297eb590d152876e3d9acaa097d72e2e58f7f1.jpg" + } + ] + } + ], + "index": 31, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 345, + 377, + 364, + 383 + ], + "lines": [ + { + "bbox": [ + 345, + 377, + 364, + 383 + ], + "spans": [ + { + "bbox": [ + 345, + 377, + 364, + 383 + ], + "type": "text", + "content": "#" + } + ] + } + ], + "index": 32, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 383, + 377, + 421, + 379 + ], + "lines": [ + { + "bbox": [ + 383, + 377, + 421, + 379 + ], + "spans": [ + { + "bbox": [ + 383, + 377, + 421, + 379 + ], + "type": "text", + "content": "" + } + ] + } + ], + "index": 34, + "angle": 0, + "type": "image_caption" + } + ], + "index": 31 + }, + { + "type": "image", + "bbox": [ + 383, + 338, + 421, + 376 + ], + "blocks": [ + { + "bbox": [ + 383, + 338, + 421, + 376 + ], + "lines": [ + { + "bbox": [ + 383, + 338, + 421, + 376 + ], + "spans": [ + { + "bbox": [ + 383, + 338, + 421, + 376 + ], + "type": "image", + "image_path": "f723652b1dcf649b48ab6399b00dee90e13af0aafe50b90311e6f3373a0cee3c.jpg" + } + ] + } + ], + "index": 33, + "angle": 0, + "type": "image_body" + } + ], + "index": 33 + }, + { + "type": "image", + "bbox": [ + 421, + 338, + 458, + 376 + ], + "blocks": [ + { + "bbox": [ + 421, + 338, + 458, + 376 + ], + "lines": [ + { + "bbox": [ + 421, + 338, + 458, + 376 + ], + "spans": [ + { + "bbox": [ + 421, + 338, + 458, + 376 + ], + "type": "image", + "image_path": "1f6e7ad2fc61d7db57ae195837c09a233a1e5982d99a20eddbffc80d18f77904.jpg" + } + ] + } + ], + "index": 35, + "angle": 0, + "type": "image_body" + } + ], + "index": 35 + }, + { + "bbox": [ + 159, + 479, + 453, + 666 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 159, + 479, + 453, + 666 + ], + "spans": [ + { + "bbox": [ + 159, + 479, + 453, + 666 + ], + "type": "text", + "content": "Abstract. We explore the dexterous manipulation transfer problem by designing simulators. The task wishes to transfer human manipulations to dexterous robot hand simulations and is inherently difficult due to its intricate, highly-constrained, and discontinuous dynamics and the need to control a dexterous hand with a DoF to accurately replicate human manipulations. Previous approaches that optimize in high-fidelity black-box simulators or a modified one with relaxed constraints only demonstrate limited capabilities or are restricted by insufficient simulation fidelity. We introduce parameterized quasi-physical simulators and a physics curriculum to overcome these limitations. The key ideas are 1) balancing between fidelity and optimizability of the simulation via a curriculum of parameterized simulators, and 2) solving the problem in each of the simulators from the curriculum, with properties ranging from high task optimizability to high fidelity. We successfully enable a dexterous hand to track complex and diverse manipulations in high-fidelity simulated environments, boosting the success rate by " + }, + { + "bbox": [ + 159, + 479, + 453, + 666 + ], + "type": "inline_equation", + "content": "11\\%+" + }, + { + "bbox": [ + 159, + 479, + 453, + 666 + ], + "type": "text", + "content": " from the best-performed baseline. The project website is available at QuasiSim." + } + ] + } + ], + "index": 37 + } + ], + "discarded_blocks": [], + "page_size": [ + 612, + 792 + ], + "page_idx": 0 + }, + { + "para_blocks": [ + { + "bbox": [ + 133, + 114, + 229, + 127 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 114, + 229, + 127 + ], + "spans": [ + { + "bbox": [ + 133, + 114, + 229, + 127 + ], + "type": "text", + "content": "1 Introduction" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 130, + 139, + 481, + 295 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 139, + 481, + 295 + ], + "spans": [ + { + "bbox": [ + 130, + 139, + 481, + 295 + ], + "type": "text", + "content": "Advancing an embodied agent's capacity to interact with the world represents a significant stride toward achieving general artificial intelligence. Due to the substantial costs and potential hazards of setting up real robots to do trial and error, the standard approach for developing embodied algorithms involves learning in physical simulators [9, 15, 23, 25, 33, 56, 59] before transitioning to real-world deployment. In most cases, physical simulators are treated as black boxes, and extensive efforts have been devoted to developing learning and optimization methods for embodied skills within these black boxes. Despite the considerable progress [2, 6-8, 16, 20, 21, 31, 36, 39, 43, 46, 60, 62, 66, 68], the question like whether the simulators used are the most suitable ones is rarely discussed. In this work, we investigate this issue and illustrate how optimizing the simulator concurrently with skill acquisition can benefit a popular yet challenging task in robot manipulation - dexterous manipulation transfer." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 130, + 295, + 481, + 521 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 295, + 481, + 521 + ], + "spans": [ + { + "bbox": [ + 130, + 295, + 481, + 521 + ], + "type": "text", + "content": "The task aims at transferring human-object manipulations to a dexterous robot hand, enabling it to physically track the reference motion of both the hand and the object (see Fig. 1). It is challenged by 1) the complex, highly constrained, non-smooth, and discontinuous dynamics with frequent contact establishment and breaking involved in the robot manipulation, 2) the requirement of precisely controlling a dexterous hand with a high DoF to densely track the manipulation at each frame, and 3) the morphology difference. Some existing works rely on high-fidelity black-box simulators, where a small difference in robot control can result in dramatically different manipulation outcomes due to abrupt contact changes, making the tracking objective highly non-smooth and hard to optimize [4,6,8,43,46]. In this way, their tasks are restricted to relatively simple goal-driven manipulations such as pouring and re-locating [8,43,46,68], in-hand re-orientation, flipping and spinning [4,6] with a fixed-root robot hand, or manipulating objects with simple geometry such as balls [36]. Other approaches attempt to improve optimization by relaxing physical constraints, with a primary focus on smoothing out contact responses [3,24,38,55,56]. However, their dynamics models may significantly deviate from real physics [38], hindering skill deployment. Consequently, we ask how to address the optimization challenge while preserving the high fidelity of the simulator." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 130, + 521, + 481, + 617 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 521, + 481, + 617 + ], + "spans": [ + { + "bbox": [ + 130, + 521, + 481, + 617 + ], + "type": "text", + "content": "Our key insight is that a single simulator can hardly provide both high fidelity and excellent optimizability for contact-rich dexterous manipulations. Inspired by the line of homotopy methods [12, 28, 29, 61], we propose a curriculum of simulators to realize this. We start by utilizing a quasi-physical simulator to initially relax physical constraints and warm up the optimization. Subsequently, we transfer the optimization outcomes to simulators with gradually tightened physical constraints. Finally, we transition to a physically realistic simulator for skill deployment in realistic dynamics." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 130, + 617, + 481, + 665 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 617, + 481, + 665 + ], + "spans": [ + { + "bbox": [ + 130, + 617, + 481, + 665 + ], + "type": "text", + "content": "To realize this vision, we propose a family of parameterized quasi-physical simulators for contact-rich dexterous manipulation tasks. These simulators can be customized to enhance task optimizability while can also be tailored to approximate realistic physics. The parameterized simulator represents" + } + ] + } + ], + "index": 6 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "spans": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "type": "text", + "content": "2" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 166, + 91, + 217, + 100 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 166, + 91, + 217, + 100 + ], + "spans": [ + { + "bbox": [ + 166, + 91, + 217, + 100 + ], + "type": "text", + "content": "X. Liu et al." + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 1 + }, + { + "para_blocks": [ + { + "bbox": [ + 133, + 117, + 481, + 426 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 117, + 481, + 426 + ], + "spans": [ + { + "bbox": [ + 133, + 117, + 481, + 426 + ], + "type": "text", + "content": "an articulated multi rigid body as a parameterized point set, models contact using an unconstrained parameterized spring-damper, and compensates for unmodeled effects via parameterized residual physics. Specifically, the articulated multi-body dynamics model is relaxed as the point set dynamics model. An articulated object is relaxed into a set of points, sampled from the ambient space surrounding each body's surface mesh. The resulting dynamics model combines the original articulated dynamics with the mass-point dynamics of each individual point. Parameters are introduced to control the point set construction and the dynamics model. The contact model is softened as a parameterized spring-damper model [3, 19, 35, 38, 51] with parameters introduced to control when to calculate contacts and contact spring stiffness. The residual physics network compensate for unmodeled effects from the analytical modeling [22]. The parameterized simulator can be programmed for high optimizability by relaxing constraints in the analytical model and can be tailored to approximate realistic physics by learning excellent residual physics. We demonstrate that the challenging dexterous manipulation transfer task can be effectively addressed through curriculum optimization using a series of parameterized physical simulators. Initially, both articulated rigid constraints and the contact model stiffness are relaxed in the simulator. It may not reflect physical realism but provides a good environment where the manipulation transfer problem can be solved easily. Subsequently, the articulated rigid constraints and the contact model are gradually tightened. Task-solving proceeds iteratively within each simulator in the curriculum. Finally, the parameterized simulator is optimized to approximate realistic physics. Task optimization continues, yielding a dexterous hand trajectory capable of executing the manipulation in environments with realistic physics." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 133, + 436, + 481, + 651 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 436, + 481, + 651 + ], + "spans": [ + { + "bbox": [ + 133, + 436, + 481, + 651 + ], + "type": "text", + "content": "We demonstrate the superiority of our method and compare it with previous model-free and model-based methods on challenging manipulation sequences from three datasets, describing single-hand or bimanual manipulations with daily objects or using tools. We conduct dexterous manipulation transfer on two widely used simulators, namely Bullet [9] and Isaac Gym [33] to demonstrate the generality and the efficacy of our method and the capability of our quasi-physical simulator to approximate the unknown black-box physics model in the contact-rich manipulation scenario (Fig. 1). We can track complex manipulations involving non-trivial object motions such as large rotations and complicated tool-using such as using a spoon to bring the water back and forth. Our approach successfully surpasses the previous best-performed method both quantitatively and qualitatively, achieving more than " + }, + { + "bbox": [ + 133, + 436, + 481, + 651 + ], + "type": "inline_equation", + "content": "11\\%" + }, + { + "bbox": [ + 133, + 436, + 481, + 651 + ], + "type": "text", + "content": " success rate than the previous best-performed method. Besides, optimizing through the physics curriculum can significantly enhance the performance of previously under-performed RL-based methods, almost completing the tracking problem from failure, as demonstrated in Fig. 1. This indicates the universality of our approach to embodied AI through optimization via a physics curriculum. Thorough ablations are conducted to validate the efficacy of our designs." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 147, + 660, + 292, + 671 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 147, + 660, + 292, + 671 + ], + "spans": [ + { + "bbox": [ + 147, + 660, + 292, + 671 + ], + "type": "text", + "content": "Our contributions are three-fold:" + } + ] + } + ], + "index": 4 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 185, + 91, + 447, + 102 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 185, + 91, + 447, + 102 + ], + "spans": [ + { + "bbox": [ + 185, + 91, + 447, + 102 + ], + "type": "text", + "content": "Quasi-Physical Simulators for Dexterous Manipulations Transfer" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 474, + 92, + 480, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 474, + 92, + 480, + 100 + ], + "spans": [ + { + "bbox": [ + 474, + 92, + 480, + 100 + ], + "type": "text", + "content": "3" + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 2 + }, + { + "para_blocks": [ + { + "bbox": [ + 137, + 116, + 480, + 223 + ], + "type": "list", + "angle": 0, + "index": 5, + "blocks": [ + { + "bbox": [ + 137, + 116, + 479, + 151 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 137, + 116, + 479, + 151 + ], + "spans": [ + { + "bbox": [ + 137, + 116, + 479, + 151 + ], + "type": "text", + "content": "- We introduce a family of parameterized quasi-physical simulators that can be configured to relax various physical constraints, facilitating skill optimization, and can also be tailored to achieve high simulation fidelity." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 138, + 152, + 480, + 186 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 152, + 480, + 186 + ], + "spans": [ + { + "bbox": [ + 138, + 152, + 480, + 186 + ], + "type": "text", + "content": "- We present a quasi-physics curriculum along with a corresponding optimization method to address the challenging dexterous manipulation transfer problem." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 138, + 188, + 480, + 223 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 188, + 480, + 223 + ], + "spans": [ + { + "bbox": [ + 138, + 188, + 480, + 223 + ], + "type": "text", + "content": "- Extensive experiments demonstrate the effectiveness of our method in transferring complex manipulations, including non-trivial object motions and changing contacts, to a dexterous robot hand in simulation." + } + ] + } + ], + "index": 4 + } + ], + "sub_type": "text" + }, + { + "bbox": [ + 132, + 236, + 242, + 249 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 236, + 242, + 249 + ], + "spans": [ + { + "bbox": [ + 132, + 236, + 242, + 249 + ], + "type": "text", + "content": "2 Related Works" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 130, + 254, + 481, + 396 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 254, + 481, + 396 + ], + "spans": [ + { + "bbox": [ + 130, + 254, + 481, + 396 + ], + "type": "text", + "content": "Dexterous manipulation transfer. Transferring human manipulations to dexterous robot-hand simulations is an important topic in robot skill acquisition [8,21,31,43,60,62,68,70]. Most approaches treat the simulator as black-box physics models and try to learn skills directly from that [4,6,8,43,46]. However, their demonstrated capabilities are restricted to relatively simple tasks. Another trend of work tries to relax the physics model [37,38] to create a better environment for task optimization. However, due to the disparity between their modeling approach and realistic physics, successful trials are typically demonstrated only in their simulators, which can hardly complete the task under physically realistic dynamics. In this work, we introduce various parameterized analytical relaxations to improve the task optimizability while compensating for unmodeled effects via residual physics networks so the fidelity would not be sacrificed." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 130, + 398, + 481, + 492 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 398, + 481, + 492 + ], + "spans": [ + { + "bbox": [ + 130, + 398, + 481, + 492 + ], + "type": "text", + "content": "Learning for simulation. Analytical methods can hardly approximate an extremely realistic physical world despite lots of smart and tremendous efforts made in developing numerical algorithms [19, 23, 26, 27]. Recently, data-driven approaches have attracted lots of interest for their high efficiency and strong approximation ability [10, 11, 22, 40, 41, 50, 63]. Special network designs are proposed to learn the contact behaviour [22, 41]. We in this work propose to leverage an analytical-neural hybrid approach and carefully design network modules for approximating residual contact forces in the contact-rich manipulation scenario." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 130, + 493, + 481, + 613 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 493, + 481, + 613 + ], + "spans": [ + { + "bbox": [ + 130, + 493, + 481, + 613 + ], + "type": "text", + "content": "Sim-to-Sim and Sim-to-Real transfer. The field of robot manipulation continues to face challenges in the areas of Sim2Sim and Sim2Real transferability [71]. Considering the modeling gaps, the optimal strategy learned in a specific simulator is difficult to transfer to a different simulator or the real world. Therefore, many techniques for solving the problem have been proposed, including imitation learning [34,42,43,45,46,48], transfer learning [72], distillation [47,57], residual physics [17,67], and efforts on bridging the gap from the dynamics model aspect [22,69]. Our parameterized simulators learn residual physics involved in contact-rich robot manipulations. By combining an analytical base with residual networks, we showcase their ability to approximate realistic physics." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 132, + 624, + 202, + 635 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 624, + 202, + 635 + ], + "spans": [ + { + "bbox": [ + 132, + 624, + 202, + 635 + ], + "type": "text", + "content": "3 Method" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 130, + 641, + 481, + 666 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 641, + 481, + 666 + ], + "spans": [ + { + "bbox": [ + 130, + 641, + 481, + 666 + ], + "type": "text", + "content": "Given a human manipulation demonstration, composed of a human hand mesh trajectory and an object pose trajectory " + }, + { + "bbox": [ + 130, + 641, + 481, + 666 + ], + "type": "inline_equation", + "content": "\\{\\mathcal{H},\\mathcal{O}\\}" + }, + { + "bbox": [ + 130, + 641, + 481, + 666 + ], + "type": "text", + "content": ", the goal is transferring the" + } + ] + } + ], + "index": 11 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "spans": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "type": "text", + "content": "4" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 166, + 91, + 217, + 100 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 166, + 91, + 217, + 100 + ], + "spans": [ + { + "bbox": [ + 166, + 91, + 217, + 100 + ], + "type": "text", + "content": "X. Liu et al." + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 3 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 134, + 114, + 478, + 210 + ], + "blocks": [ + { + "bbox": [ + 136, + 106, + 239, + 114 + ], + "lines": [ + { + "bbox": [ + 136, + 106, + 239, + 114 + ], + "spans": [ + { + "bbox": [ + 136, + 106, + 239, + 114 + ], + "type": "text", + "content": "Parameterized Quasi-Physical Simulator" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 134, + 114, + 478, + 210 + ], + "lines": [ + { + "bbox": [ + 134, + 114, + 478, + 210 + ], + "spans": [ + { + "bbox": [ + 134, + 114, + 478, + 210 + ], + "type": "image", + "image_path": "99664072a887cf47fdcf35e54f57509bda3392ca2b5ecd914f4b592804383add.jpg" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 130, + 217, + 482, + 273 + ], + "lines": [ + { + "bbox": [ + 130, + 217, + 482, + 273 + ], + "spans": [ + { + "bbox": [ + 130, + 217, + 482, + 273 + ], + "type": "text", + "content": "Fig. 2: The parameterized quasi-physical simulator relaxes the articulated multi rigid body dynamics as the parameterized point set dynamics, controls the contact behavior via an unconstrained parameterized spring-damper contact model, and compensates for unmodeled effects via parameterized residual physics networks. We tackle the difficult dexterous manipulation transfer problem via a physics curriculum." + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "image_caption" + } + ], + "index": 3 + }, + { + "bbox": [ + 130, + 279, + 482, + 363 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 279, + 482, + 363 + ], + "spans": [ + { + "bbox": [ + 130, + 279, + 482, + 363 + ], + "type": "text", + "content": "demonstration to a dexterous robot hand in simulation. Formally, we aim to optimize a control trajectory " + }, + { + "bbox": [ + 130, + 279, + 482, + 363 + ], + "type": "inline_equation", + "content": "\\mathcal{A}" + }, + { + "bbox": [ + 130, + 279, + 482, + 363 + ], + "type": "text", + "content": " that drives the dexterous hand to manipulate the object in a realistic simulated environment so that the resulting hand trajectory " + }, + { + "bbox": [ + 130, + 279, + 482, + 363 + ], + "type": "inline_equation", + "content": "\\hat{\\mathcal{H}}" + }, + { + "bbox": [ + 130, + 279, + 482, + 363 + ], + "type": "text", + "content": " and the object trajectory " + }, + { + "bbox": [ + 130, + 279, + 482, + 363 + ], + "type": "inline_equation", + "content": "\\hat{\\mathcal{O}}" + }, + { + "bbox": [ + 130, + 279, + 482, + 363 + ], + "type": "text", + "content": " are close to the reference motion " + }, + { + "bbox": [ + 130, + 279, + 482, + 363 + ], + "type": "inline_equation", + "content": "\\{\\mathcal{H},\\mathcal{O}\\}" + }, + { + "bbox": [ + 130, + 279, + 482, + 363 + ], + "type": "text", + "content": ". The problem is challenged by difficulties from the highly constrained, discontinuous, and non-smooth dynamics, the requirement of controlling a high DoF dexterous hand for tracking, and the morphology difference." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 130, + 363, + 482, + 447 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 363, + 482, + 447 + ], + "spans": [ + { + "bbox": [ + 130, + 363, + 482, + 447 + ], + "type": "text", + "content": "Our method comprises two key designs to tackle the challenges: 1) a family of parameterized quasi-physical simulators, which can be programmed to enhance the optimizability of contact-rich dexterous manipulation tasks and can also be tailored to approximate realistic physics (Section 3.1), and 2) a physics curriculum that carefully adjusts the parameters of a line of quasi-physical simulators and a strategy that solves the difficult dexterous manipulation transfer task by addressing it within each simulator in the curriculum (Section 3.2)." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 131, + 465, + 369, + 477 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 131, + 465, + 369, + 477 + ], + "spans": [ + { + "bbox": [ + 131, + 465, + 369, + 477 + ], + "type": "text", + "content": "3.1 Parameterized Quasi-Physical Simulators" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 130, + 486, + 485, + 581 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 486, + 485, + 581 + ], + "spans": [ + { + "bbox": [ + 130, + 486, + 485, + 581 + ], + "type": "text", + "content": "Our quasi-physical simulator represents an articulated multi-body, i.e., the robotic dexterous hand, as a point set. The object is represented as a signed distance field. The base of the simulator is in an analytical form leveraging an unconstrained spring-damper contact model. Parameters are introduced to control the analytical relaxations on the articulated rigid constraints and the softness of the contact model. Additionally, neural networks are introduced to compensate for unmodeled effects beyond the analytical framework. We will elaborate on each of these design aspects below." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 130, + 582, + 482, + 666 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 582, + 482, + 666 + ], + "spans": [ + { + "bbox": [ + 130, + 582, + 482, + 666 + ], + "type": "text", + "content": "Parameterized point set dynamics. Articulated multi-body represented in the reduced coordinate system [19,59] may require a large change in joint states to achieve a small adjustment in the Euclidean space. Moving the end effector from one point to a nearby point may require adjusting all joint states (Fig. 3). Besides, transferring the hand trajectory to a morphologically different hand requires correspondences to make the resulting trajectory close to the original one. Defining correspondences in the reduced coordinate or via sparse correspon" + } + ] + } + ], + "index": 9 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 184, + 91, + 448, + 102 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 184, + 91, + 448, + 102 + ], + "spans": [ + { + "bbox": [ + 184, + 91, + 448, + 102 + ], + "type": "text", + "content": "Quasi-Physical Simulators for Dexterous Manipulations Transfer" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 474, + 92, + 480, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 474, + 92, + 480, + 100 + ], + "spans": [ + { + "bbox": [ + 474, + 92, + 480, + 100 + ], + "type": "text", + "content": "5" + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 4 + }, + { + "para_blocks": [ + { + "bbox": [ + 130, + 116, + 482, + 224 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 116, + 482, + 224 + ], + "spans": [ + { + "bbox": [ + 130, + 116, + 482, + 224 + ], + "type": "text", + "content": "dences will make the result suffer from noise in the data, leading to unwanted results finally (Fig. 3). Hence, we propose relaxing an articulated multi-rigid body into a mass-point set sampled from the ambient space surrounding each body. Each point is considered attached to the body from which it is sampled and is capable of both self-actuation and actuation via joint motors. We introduce a parameter " + }, + { + "bbox": [ + 130, + 116, + 482, + 224 + ], + "type": "inline_equation", + "content": "\\alpha" + }, + { + "bbox": [ + 130, + 116, + 482, + 224 + ], + "type": "text", + "content": " to control the point set construction and the dynamics. This representation allows an articulated rigid object to behave similarly to a deformable object, providing a larger action space to adjust its state and thereby easing the control optimization problem." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 130, + 224, + 361, + 344 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 224, + 361, + 344 + ], + "spans": [ + { + "bbox": [ + 130, + 224, + 361, + 344 + ], + "type": "text", + "content": "Specifically, for each body of the articulated object, we sample a set of points from the ambient space near the body mesh. The point set " + }, + { + "bbox": [ + 130, + 224, + 361, + 344 + ], + "type": "inline_equation", + "content": "\\mathcal{Q}" + }, + { + "bbox": [ + 130, + 224, + 361, + 344 + ], + "type": "text", + "content": " is constructed by concatenating all sampled points together. Each point " + }, + { + "bbox": [ + 130, + 224, + 361, + 344 + ], + "type": "inline_equation", + "content": "\\mathbf{p}_i\\in \\mathcal{Q}" + }, + { + "bbox": [ + 130, + 224, + 361, + 344 + ], + "type": "text", + "content": " is treated as a mass point with a finite mass " + }, + { + "bbox": [ + 130, + 224, + 361, + 344 + ], + "type": "inline_equation", + "content": "\\mathbf{m}_i" + }, + { + "bbox": [ + 130, + 224, + 361, + 344 + ], + "type": "text", + "content": " and infinitesimal volume. The dynamics of the point set consist of articulated multi-body dynamics [14, 30], as well as the mass point dynamics of each point " + }, + { + "bbox": [ + 130, + 224, + 361, + 344 + ], + "type": "inline_equation", + "content": "\\mathbf{p}_i" + }, + { + "bbox": [ + 130, + 224, + 361, + 344 + ], + "type": "text", + "content": ". For each " + }, + { + "bbox": [ + 130, + 224, + 361, + 344 + ], + "type": "inline_equation", + "content": "\\mathbf{p}_i" + }, + { + "bbox": [ + 130, + 224, + 361, + 344 + ], + "type": "text", + "content": ", we have:" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 192, + 354, + 361, + 368 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 192, + 354, + 361, + 368 + ], + "spans": [ + { + "bbox": [ + 192, + 354, + 361, + 368 + ], + "type": "interline_equation", + "content": "m _ {i} \\ddot {\\mathbf {x}} _ {i} = \\mathbf {J} _ {i} \\mathbf {u} + \\alpha \\mathbf {f} _ {i} + \\alpha \\mathbf {a} _ {i}, \\tag {1}", + "image_path": "6fc4a7631598eb2cd363a1526507e2ee5172577c1675bcc9fee8daaa9d42bce8.jpg" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 130, + 377, + 361, + 415 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 377, + 361, + 415 + ], + "spans": [ + { + "bbox": [ + 130, + 377, + 361, + 415 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 130, + 377, + 361, + 415 + ], + "type": "inline_equation", + "content": "\\mathbf{J}_i" + }, + { + "bbox": [ + 130, + 377, + 361, + 415 + ], + "type": "text", + "content": " represents the Jacobian mapping from the generalized velocity to the point velocity " + }, + { + "bbox": [ + 130, + 377, + 361, + 415 + ], + "type": "inline_equation", + "content": "\\dot{\\mathbf{x}}_i" + }, + { + "bbox": [ + 130, + 377, + 361, + 415 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 130, + 377, + 361, + 415 + ], + "type": "inline_equation", + "content": "\\mathbf{u}" + }, + { + "bbox": [ + 130, + 377, + 361, + 415 + ], + "type": "text", + "content": " denotes the generalized joint force, " + }, + { + "bbox": [ + 130, + 377, + 361, + 415 + ], + "type": "inline_equation", + "content": "\\mathbf{f}_i" + }, + { + "bbox": [ + 130, + 377, + 361, + 415 + ], + "type": "text", + "content": " accounts for ex" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 130, + 415, + 481, + 449 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 415, + 481, + 449 + ], + "spans": [ + { + "bbox": [ + 130, + 415, + 481, + 449 + ], + "type": "text", + "content": "ternal forces acting on " + }, + { + "bbox": [ + 130, + 415, + 481, + 449 + ], + "type": "inline_equation", + "content": "\\mathbf{p}_i" + }, + { + "bbox": [ + 130, + 415, + 481, + 449 + ], + "type": "text", + "content": ", and " + }, + { + "bbox": [ + 130, + 415, + 481, + 449 + ], + "type": "inline_equation", + "content": "\\mathbf{a}_i \\in \\mathbb{R}^3" + }, + { + "bbox": [ + 130, + 415, + 481, + 449 + ], + "type": "text", + "content": " represents the actuation force applied to the point " + }, + { + "bbox": [ + 130, + 415, + 481, + 449 + ], + "type": "inline_equation", + "content": "\\mathbf{p}_i" + }, + { + "bbox": [ + 130, + 415, + 481, + 449 + ], + "type": "text", + "content": ". Consequently, the point set is controlled by a shared control in the reduced coordinate space " + }, + { + "bbox": [ + 130, + 415, + 481, + 449 + ], + "type": "inline_equation", + "content": "\\mathbf{u}" + }, + { + "bbox": [ + 130, + 415, + 481, + 449 + ], + "type": "text", + "content": " and per-point actuation force " + }, + { + "bbox": [ + 130, + 415, + 481, + 449 + ], + "type": "inline_equation", + "content": "\\mathbf{a}_i" + }, + { + "bbox": [ + 130, + 415, + 481, + 449 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 130, + 450, + 482, + 668 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 450, + 482, + 668 + ], + "spans": [ + { + "bbox": [ + 130, + 450, + 482, + 668 + ], + "type": "text", + "content": "Parameterized spring-damper contact modeling. To ease the optimization challenges posed by contact-rich manipulations, which arise from contact constraints such as the non-penetration requirement and Coulomb friction law [3,5], as well as discontinuous dynamics involving frequent contact establishment and breaking, we propose a parameterized contact model for relaxing constraints and controlling the contact behavior. Specifically, we leverage a classical unconstrained spring-damper model [19,35,51,59,64] to model the contacts. This model allows us to flexibly adjust the contact behavior by tuning the contact threshold and the spring stiffness coefficients. Intuitively, a contact model with a high threshold and low spring stiffness presents \"soft\" behaviors, resulting in a continuous and smooth optimization space. This makes optimization through such a contact model relatively easy. Conversely, a model with a low threshold and large stiffness coefficients will produce \"stiff\" behaviors, increasing the discontinuity of the optimization space due to frequent contact establishment and breaking. However, it also becomes more physically realistic, meaning contact forces are calculated only when two objects collide, and a large force is applied to separate them if penetrations are observed, thus better satisfying the non-penetration condition. Therefore, by adjusting the contact distance threshold" + } + ] + } + ], + "index": 7 + }, + { + "type": "image", + "bbox": [ + 374, + 249, + 481, + 322 + ], + "blocks": [ + { + "bbox": [ + 374, + 249, + 481, + 322 + ], + "lines": [ + { + "bbox": [ + 374, + 249, + 481, + 322 + ], + "spans": [ + { + "bbox": [ + 374, + 249, + 481, + 322 + ], + "type": "image", + "image_path": "dd5c4074daa60bddaf492a4d97d3815fc002598ecae4e5ba4392d05d6e671f52.jpg" + } + ] + } + ], + "index": 8, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 367, + 331, + 482, + 397 + ], + "lines": [ + { + "bbox": [ + 367, + 331, + 482, + 397 + ], + "spans": [ + { + "bbox": [ + 367, + 331, + 482, + 397 + ], + "type": "text", + "content": "Fig.3: Point Set can flexibly adjust its states, avoid overfitting to data noise, and ease the difficulty brought by the morphology difference." + } + ] + } + ], + "index": 9, + "angle": 0, + "type": "image_caption" + } + ], + "index": 8 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "spans": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "type": "text", + "content": "6" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 166, + 91, + 217, + 101 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 166, + 91, + 217, + 101 + ], + "spans": [ + { + "bbox": [ + 166, + 91, + 217, + 101 + ], + "type": "text", + "content": "X. Liu et al." + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 5 + }, + { + "para_blocks": [ + { + "bbox": [ + 130, + 116, + 479, + 163 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 116, + 479, + 163 + ], + "spans": [ + { + "bbox": [ + 130, + 116, + 479, + 163 + ], + "type": "text", + "content": "and spring stiffness coefficients, we can modulate the optimizability and fidelity of the contact model. The parameter set of the contact model comprises a distance threshold " + }, + { + "bbox": [ + 130, + 116, + 479, + 163 + ], + "type": "inline_equation", + "content": "d^{c}" + }, + { + "bbox": [ + 130, + 116, + 479, + 163 + ], + "type": "text", + "content": " and spring stiffness coefficients. Next, we will delve into the details of the contact establishment, breaking, and force calculations processes." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 130, + 164, + 480, + 258 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 164, + 480, + 258 + ], + "spans": [ + { + "bbox": [ + 130, + 164, + 480, + 258 + ], + "type": "text", + "content": "Contacts are established between points in the manipulator's point set " + }, + { + "bbox": [ + 130, + 164, + 480, + 258 + ], + "type": "inline_equation", + "content": "\\mathcal{Q}" + }, + { + "bbox": [ + 130, + 164, + 480, + 258 + ], + "type": "text", + "content": " and the object. A point " + }, + { + "bbox": [ + 130, + 164, + 480, + 258 + ], + "type": "inline_equation", + "content": "\\mathbf{p} \\in \\mathcal{Q}" + }, + { + "bbox": [ + 130, + 164, + 480, + 258 + ], + "type": "text", + "content": " is considered to be in \"contact\" with the object if its signed distance to the object " + }, + { + "bbox": [ + 130, + 164, + 480, + 258 + ], + "type": "inline_equation", + "content": "\\operatorname{sd}(\\mathbf{p})" + }, + { + "bbox": [ + 130, + 164, + 480, + 258 + ], + "type": "text", + "content": " is smaller than the contact distance threshold " + }, + { + "bbox": [ + 130, + 164, + 480, + 258 + ], + "type": "inline_equation", + "content": "d^{c}" + }, + { + "bbox": [ + 130, + 164, + 480, + 258 + ], + "type": "text", + "content": ". Subsequently, the object surface point nearest to " + }, + { + "bbox": [ + 130, + 164, + 480, + 258 + ], + "type": "inline_equation", + "content": "\\mathbf{p}" + }, + { + "bbox": [ + 130, + 164, + 480, + 258 + ], + "type": "text", + "content": " is identified as the corresponding contact point on the object, denoted as " + }, + { + "bbox": [ + 130, + 164, + 480, + 258 + ], + "type": "inline_equation", + "content": "\\mathbf{p}^{o}" + }, + { + "bbox": [ + 130, + 164, + 480, + 258 + ], + "type": "text", + "content": ". The normal direction of the object point " + }, + { + "bbox": [ + 130, + 164, + 480, + 258 + ], + "type": "inline_equation", + "content": "\\mathbf{p}^{o}" + }, + { + "bbox": [ + 130, + 164, + 480, + 258 + ], + "type": "text", + "content": " is then determined as the contact normal direction, denoted as " + }, + { + "bbox": [ + 130, + 164, + 480, + 258 + ], + "type": "inline_equation", + "content": "\\mathbf{n}^{o}" + }, + { + "bbox": [ + 130, + 164, + 480, + 258 + ], + "type": "text", + "content": ". The contact force " + }, + { + "bbox": [ + 130, + 164, + 480, + 258 + ], + "type": "inline_equation", + "content": "\\mathbf{f}^{c}" + }, + { + "bbox": [ + 130, + 164, + 480, + 258 + ], + "type": "text", + "content": " applied from the manipulator point " + }, + { + "bbox": [ + 130, + 164, + 480, + 258 + ], + "type": "inline_equation", + "content": "\\mathbf{p}" + }, + { + "bbox": [ + 130, + 164, + 480, + 258 + ], + "type": "text", + "content": " to " + }, + { + "bbox": [ + 130, + 164, + 480, + 258 + ], + "type": "inline_equation", + "content": "\\mathbf{p}^{o}" + }, + { + "bbox": [ + 130, + 164, + 480, + 258 + ], + "type": "text", + "content": " is calculated as follows:" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 257, + 259, + 480, + 272 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 257, + 259, + 480, + 272 + ], + "spans": [ + { + "bbox": [ + 257, + 259, + 480, + 272 + ], + "type": "interline_equation", + "content": "\\mathbf {f} ^ {c} = - \\left(k ^ {n} d - k ^ {d} d \\dot {d}\\right) \\mathbf {n} ^ {o}, \\tag {2}", + "image_path": "a641d7c9782c2202b06bbdd66d5b5fee225aecad4340f0c2a2cb7be6d709da72.jpg" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 130, + 276, + 481, + 312 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 276, + 481, + 312 + ], + "spans": [ + { + "bbox": [ + 130, + 276, + 481, + 312 + ], + "type": "text", + "content": "where, " + }, + { + "bbox": [ + 130, + 276, + 481, + 312 + ], + "type": "inline_equation", + "content": "k^n" + }, + { + "bbox": [ + 130, + 276, + 481, + 312 + ], + "type": "text", + "content": " represents the spring stiffness coefficient, " + }, + { + "bbox": [ + 130, + 276, + 481, + 312 + ], + "type": "inline_equation", + "content": "k^d" + }, + { + "bbox": [ + 130, + 276, + 481, + 312 + ], + "type": "text", + "content": " denotes the damping coefficient, and " + }, + { + "bbox": [ + 130, + 276, + 481, + 312 + ], + "type": "inline_equation", + "content": "d = d^{c} - \\mathrm{sd}(\\mathbf{p})" + }, + { + "bbox": [ + 130, + 276, + 481, + 312 + ], + "type": "text", + "content": " is always positive. To enhance the continuity of " + }, + { + "bbox": [ + 130, + 276, + 481, + 312 + ], + "type": "inline_equation", + "content": "\\mathbf{f}^c" + }, + { + "bbox": [ + 130, + 276, + 481, + 312 + ], + "type": "text", + "content": " [64], " + }, + { + "bbox": [ + 130, + 276, + 481, + 312 + ], + "type": "inline_equation", + "content": "k^d\\ddot{d}\\dot{d}" + }, + { + "bbox": [ + 130, + 276, + 481, + 312 + ], + "type": "text", + "content": " is used as the magnitude of the damping force, rather than " + }, + { + "bbox": [ + 130, + 276, + 481, + 312 + ], + "type": "inline_equation", + "content": "k^d\\dot{d}" + }, + { + "bbox": [ + 130, + 276, + 481, + 312 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 130, + 313, + 480, + 373 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 313, + 480, + 373 + ], + "spans": [ + { + "bbox": [ + 130, + 313, + 480, + 373 + ], + "type": "text", + "content": "Friction forces are modeled as penalty-based spring forces [3,65]. Once a point " + }, + { + "bbox": [ + 130, + 313, + 480, + 373 + ], + "type": "inline_equation", + "content": "\\mathbf{p}" + }, + { + "bbox": [ + 130, + 313, + 480, + 373 + ], + "type": "text", + "content": " is identified as in contact with the object, with the object contact point denoted as " + }, + { + "bbox": [ + 130, + 313, + 480, + 373 + ], + "type": "inline_equation", + "content": "\\mathbf{p}^o" + }, + { + "bbox": [ + 130, + 313, + 480, + 373 + ], + "type": "text", + "content": ", the contact pair is stored. Contact forces between them are continually calculated until the contact breaking conditions are met. In more detail, the static friction force from " + }, + { + "bbox": [ + 130, + 313, + 480, + 373 + ], + "type": "inline_equation", + "content": "\\mathbf{p}" + }, + { + "bbox": [ + 130, + 313, + 480, + 373 + ], + "type": "text", + "content": " to " + }, + { + "bbox": [ + 130, + 313, + 480, + 373 + ], + "type": "inline_equation", + "content": "\\mathbf{p}^o" + }, + { + "bbox": [ + 130, + 313, + 480, + 373 + ], + "type": "text", + "content": " is calculated using a spring model:" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 263, + 380, + 480, + 394 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 263, + 380, + 480, + 394 + ], + "spans": [ + { + "bbox": [ + 263, + 380, + 480, + 394 + ], + "type": "interline_equation", + "content": "\\mathbf {f} _ {s} ^ {f} = k ^ {f} \\mathbf {T} _ {n} (\\mathbf {p} - \\mathbf {p} ^ {o}), \\tag {3}", + "image_path": "3a1b389629f110c267eb331fedc2babc1ee4fb14b414dbcd87123623e3121ebf.jpg" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 130, + 399, + 479, + 446 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 399, + 479, + 446 + ], + "spans": [ + { + "bbox": [ + 130, + 399, + 479, + 446 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 130, + 399, + 479, + 446 + ], + "type": "inline_equation", + "content": "k^f" + }, + { + "bbox": [ + 130, + 399, + 479, + 446 + ], + "type": "text", + "content": " is the friction spring stiffness coefficient, " + }, + { + "bbox": [ + 130, + 399, + 479, + 446 + ], + "type": "inline_equation", + "content": "\\mathbf{T}_n = \\mathbf{I} - \\mathbf{n}^o\\mathbf{n}^{oT}" + }, + { + "bbox": [ + 130, + 399, + 479, + 446 + ], + "type": "text", + "content": " is a tangential projection operator. When the static friction satisfies " + }, + { + "bbox": [ + 130, + 399, + 479, + 446 + ], + "type": "inline_equation", + "content": "\\| \\mathbf{f}_s^f\\| \\leq \\mu \\| \\mathbf{f}^c\\|" + }, + { + "bbox": [ + 130, + 399, + 479, + 446 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 130, + 399, + 479, + 446 + ], + "type": "inline_equation", + "content": "\\mathbf{f}_s^f" + }, + { + "bbox": [ + 130, + 399, + 479, + 446 + ], + "type": "text", + "content": " is applied to the object point " + }, + { + "bbox": [ + 130, + 399, + 479, + 446 + ], + "type": "inline_equation", + "content": "\\mathbf{p}^o" + }, + { + "bbox": [ + 130, + 399, + 479, + 446 + ], + "type": "text", + "content": ". Otherwise, the dynamic friction force is applied, and the contact breaks:" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 249, + 446, + 480, + 472 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 249, + 446, + 480, + 472 + ], + "spans": [ + { + "bbox": [ + 249, + 446, + 480, + 472 + ], + "type": "interline_equation", + "content": "\\mathbf {f} _ {d} ^ {f} = - \\mu \\| \\mathbf {f} _ {s} ^ {f} \\| \\frac {\\mathbf {T} _ {n} \\mathbf {v} _ {\\mathbf {p} \\leftarrow \\mathbf {p} ^ {o}}}{\\| \\mathbf {T} _ {n} \\mathbf {v} _ {\\mathbf {p} \\leftarrow \\mathbf {p} ^ {o}} \\|}, \\tag {4}", + "image_path": "63bb425d5f1fc82e0b5989de36778b1e225944a5dd2f824e780e49b7670c2c7e.jpg" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 130, + 474, + 376, + 486 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 474, + 376, + 486 + ], + "spans": [ + { + "bbox": [ + 130, + 474, + 376, + 486 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 130, + 474, + 376, + 486 + ], + "type": "inline_equation", + "content": "\\mathbf{v}_{\\mathbf{p}^{\\leftarrow}\\mathbf{p}^{o}}" + }, + { + "bbox": [ + 130, + 474, + 376, + 486 + ], + "type": "text", + "content": " is the relative velocity between " + }, + { + "bbox": [ + 130, + 474, + 376, + 486 + ], + "type": "inline_equation", + "content": "\\mathbf{p}" + }, + { + "bbox": [ + 130, + 474, + 376, + 486 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 130, + 474, + 376, + 486 + ], + "type": "inline_equation", + "content": "\\mathbf{p}^o" + }, + { + "bbox": [ + 130, + 474, + 376, + 486 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 130, + 487, + 479, + 533 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 487, + 479, + 533 + ], + "spans": [ + { + "bbox": [ + 130, + 487, + 479, + 533 + ], + "type": "text", + "content": "Parameterized residual physics. The analytical designs facilitate relaxation but may limit the use of highly sophisticated and realistic dynamics models, deviating from real physics. To address this, the final component of our quasi-physical simulator is a flexible neural residual physics model [1,22,41]." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 130, + 534, + 481, + 666 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 534, + 481, + 666 + ], + "spans": [ + { + "bbox": [ + 130, + 534, + 481, + 666 + ], + "type": "text", + "content": "Specifically, we propose to employ neural networks to learn and predict residual contact forces and friction forces based on contact-related information. For detailed residual contact force prediction, we introduce a local contact network " + }, + { + "bbox": [ + 130, + 534, + 481, + 666 + ], + "type": "inline_equation", + "content": "f_{\\psi_{\\mathrm{local}}}" + }, + { + "bbox": [ + 130, + 534, + 481, + 666 + ], + "type": "text", + "content": " that utilizes contact information identified in the parameterized contact model and predicts residual forces between each contact pair. To address discrepancies in contact region identification between the parameterized contact model and real contact region, we also incorporate a global residual network " + }, + { + "bbox": [ + 130, + 534, + 481, + 666 + ], + "type": "inline_equation", + "content": "f_{\\psi_{\\mathrm{global}}}" + }, + { + "bbox": [ + 130, + 534, + 481, + 666 + ], + "type": "text", + "content": " that predicts residual forces and torques applied directly to the object's center of mass. In more detail, for a given contact pair " + }, + { + "bbox": [ + 130, + 534, + 481, + 666 + ], + "type": "inline_equation", + "content": "(\\mathbf{p}, \\mathbf{p}^o)" + }, + { + "bbox": [ + 130, + 534, + 481, + 666 + ], + "type": "text", + "content": ", the local contact network utilizes contact-related features from the local contact region, comprising geometry, per-point velocity, and per-object point normal. It then maps these" + } + ] + } + ], + "index": 12 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 184, + 91, + 448, + 103 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 184, + 91, + 448, + 103 + ], + "spans": [ + { + "bbox": [ + 184, + 91, + 448, + 103 + ], + "type": "text", + "content": "Quasi-Physical Simulators for Dexterous Manipulations Transfer" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 474, + 92, + 480, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 474, + 92, + 480, + 100 + ], + "spans": [ + { + "bbox": [ + 474, + 92, + 480, + 100 + ], + "type": "text", + "content": "7" + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 6 + }, + { + "para_blocks": [ + { + "bbox": [ + 130, + 116, + 482, + 247 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 116, + 482, + 247 + ], + "spans": [ + { + "bbox": [ + 130, + 116, + 482, + 247 + ], + "type": "text", + "content": "features to predict the residual contact force and residual friction force between the two points in the contact pair. Additionally, the global residual network incorporates contact-related information from the global contact region, including geometry, per-point velocity, and per-object point normal, as input. It then predicts a residual force and residual torque to be applied to the object's center of mass. Details such as contact region identification and network architectures are deferred to the Supp. We denote the estimizable parameters in the residual physics network as " + }, + { + "bbox": [ + 130, + 116, + 482, + 247 + ], + "type": "inline_equation", + "content": "\\psi = (\\psi_{\\mathrm{global}},\\psi_{\\mathrm{local}})" + }, + { + "bbox": [ + 130, + 116, + 482, + 247 + ], + "type": "text", + "content": ". Through optimization of the residual physics network, we unlock the possibility of introducing highly non-linear dynamics to align our parametrized quasi-physical simulator with any realistic black-box physical simulator." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 130, + 248, + 481, + 273 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 248, + 481, + 273 + ], + "spans": [ + { + "bbox": [ + 130, + 248, + 481, + 273 + ], + "type": "text", + "content": "Semi-implicit time-stepping is leveraged to make the simulation auto differentiable and easy to combine with neural networks [22]." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 131, + 283, + 459, + 295 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 131, + 283, + 459, + 295 + ], + "spans": [ + { + "bbox": [ + 131, + 283, + 459, + 295 + ], + "type": "text", + "content": "3.2 Dexterous Manipulation Transfer via a Physics Curriculum" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 130, + 306, + 482, + 582 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 306, + 482, + 582 + ], + "spans": [ + { + "bbox": [ + 130, + 306, + 482, + 582 + ], + "type": "text", + "content": "Building upon the family of parameterized quasi-physical simulators, we present a solution to the challenging dexterous manipulation transfer problem through a physics curriculum. This curriculum consists of a sequence of parameterized simulators, ranging from those with minimal constraints and the softest contact behavior to increasingly realistic simulators. We address the problem by transferring the manipulation demonstration to the dexterous hand within each simulator across the curriculum progressively. To elaborate further, the optimization process begins within the parameterized simulator where articulated rigid constraints are removed and the contact model is tuned to its softest level. Additionally, the residual physics networks are deactivated. This initial simulator configuration offers a friendly environment for optimization. Subsequently, the physics constraints are gradually tightened as we progress through each simulator within the curriculum. The task is solved iteratively within each simulator. After reaching the most tightened analytical model, the analytical part is fixed and residual networks are activated. The simulator is gradually optimized to approximate the dynamics in a realistic physical environment. Concurrently, the control trajectory " + }, + { + "bbox": [ + 130, + 306, + 482, + 582 + ], + "type": "inline_equation", + "content": "\\mathcal{A}" + }, + { + "bbox": [ + 130, + 306, + 482, + 582 + ], + "type": "text", + "content": " continues to be refined in the quasi-physical simulator. Finally, we arrive at a simulator optimized to be with high fidelity and a trajectory " + }, + { + "bbox": [ + 130, + 306, + 482, + 582 + ], + "type": "inline_equation", + "content": "\\mathcal{A}" + }, + { + "bbox": [ + 130, + 306, + 482, + 582 + ], + "type": "text", + "content": " capable of guiding the dexterous hand to accurately track the demonstration within a realistically simulated physical environment. Additionally, since object properties as well as system parameters are unknown from the kinematics-only demonstration, we set them estimizable and identify them (denoted " + }, + { + "bbox": [ + 130, + 306, + 482, + 582 + ], + "type": "inline_equation", + "content": "S" + }, + { + "bbox": [ + 130, + 306, + 482, + 582 + ], + "type": "text", + "content": ") together with optimizing the hand control trajectory. Next we'll illustrate this in detail." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 130, + 582, + 482, + 665 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 582, + 482, + 665 + ], + "spans": [ + { + "bbox": [ + 130, + 582, + 482, + 665 + ], + "type": "text", + "content": "Transferring human demonstration via point set dynamics. To robustly transfer the human demonstration to a morphologically different dexterous robot hand in simulation and to overcome noise in the kinematic trajectory, we initially relax the articulated rigid constraints and transfer the kinematics human demonstration to the control trajectory of the point set. Specifically, the point set representation with the relaxation parameter " + }, + { + "bbox": [ + 130, + 582, + 482, + 665 + ], + "type": "inline_equation", + "content": "\\alpha" + }, + { + "bbox": [ + 130, + 582, + 482, + 665 + ], + "type": "text", + "content": " for the dynamic human hand [8] is constructed. The shared control trajectory " + }, + { + "bbox": [ + 130, + 582, + 482, + 665 + ], + "type": "inline_equation", + "content": "\\mathcal{A}" + }, + { + "bbox": [ + 130, + 582, + 482, + 665 + ], + "type": "text", + "content": " and per-point per-frame actions" + } + ] + } + ], + "index": 6 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "spans": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "type": "text", + "content": "8" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 166, + 91, + 217, + 100 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 166, + 91, + 217, + 100 + ], + "spans": [ + { + "bbox": [ + 166, + 91, + 217, + 100 + ], + "type": "text", + "content": "X. Liu et al." + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 7 + }, + { + "para_blocks": [ + { + "bbox": [ + 130, + 116, + 479, + 176 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 116, + 479, + 176 + ], + "spans": [ + { + "bbox": [ + 130, + 116, + 479, + 176 + ], + "type": "text", + "content": "are optimized so that the resulting trajectory of the point set can manipulate the object according to the demonstration. After that, a point set with the same parameter " + }, + { + "bbox": [ + 130, + 116, + 479, + 176 + ], + "type": "inline_equation", + "content": "\\alpha" + }, + { + "bbox": [ + 130, + 116, + 479, + 176 + ], + "type": "text", + "content": " is constructed to represent the dexterous robot hand. Subsequently, the shared control trajectory " + }, + { + "bbox": [ + 130, + 116, + 479, + 176 + ], + "type": "inline_equation", + "content": "\\mathcal{A}" + }, + { + "bbox": [ + 130, + 116, + 479, + 176 + ], + "type": "text", + "content": " and per-point per-frame actions are optimized to track the manipulation accordingly." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 130, + 176, + 480, + 282 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 176, + 480, + 282 + ], + "spans": [ + { + "bbox": [ + 130, + 176, + 480, + 282 + ], + "type": "text", + "content": "Transferring through a contact model curriculum. After that, the articulated rigid constraint is tightened by freezing the point set parameter " + }, + { + "bbox": [ + 130, + 176, + 480, + 282 + ], + "type": "inline_equation", + "content": "\\alpha" + }, + { + "bbox": [ + 130, + 176, + 480, + 282 + ], + "type": "text", + "content": " to zero. The following optimization starts from a parameterized simulator with the softest contact model. We then gradually tighten the contact model by adjusting its distance threshold, contact force spring stiffness, etc. By curriculum optimizing the trajectory " + }, + { + "bbox": [ + 130, + 176, + 480, + 282 + ], + "type": "inline_equation", + "content": "\\mathcal{A}" + }, + { + "bbox": [ + 130, + 176, + 480, + 282 + ], + "type": "text", + "content": " and parameters " + }, + { + "bbox": [ + 130, + 176, + 480, + 282 + ], + "type": "inline_equation", + "content": "\\mathcal{S}" + }, + { + "bbox": [ + 130, + 176, + 480, + 282 + ], + "type": "text", + "content": " in each of the quasi-physical simulators, we finally arrive at the control trajectory that can drive a dexterous hand to accomplish the tracking task in the parameterized simulator with the most tightened analytical model." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 130, + 284, + 481, + 415 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 284, + 481, + 415 + ], + "spans": [ + { + "bbox": [ + 130, + 284, + 481, + 415 + ], + "type": "text", + "content": "Optimizing towards a realistic physical environment. Subsequently, the residual physics network is activated and the parameterized simulator is optimized to approximate the dynamics in a realistic physical environment. We continue to optimize the hand trajectory in the quasi-physical simulator. Specifically, we leverage the successful trial in model-based human tracking literature [16,66] and iteratively optimize the control trajectory " + }, + { + "bbox": [ + 130, + 284, + 481, + 415 + ], + "type": "inline_equation", + "content": "\\mathcal{A}" + }, + { + "bbox": [ + 130, + 284, + 481, + 415 + ], + "type": "text", + "content": " and the parameterized simulator. In more detail, the following two subproblems are iteratively solved: 1) optimizing the quasi-physical simulator to approximate the realistic dynamics, and 2) optimizing the control trajectory " + }, + { + "bbox": [ + 130, + 284, + 481, + 415 + ], + "type": "inline_equation", + "content": "\\mathcal{A}" + }, + { + "bbox": [ + 130, + 284, + 481, + 415 + ], + "type": "text", + "content": " to complete the manipulation in the quasi-physical simulator. Gradient-based optimization is leveraged taking advantage of the differentiability of the parameterized simulator." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 130, + 415, + 481, + 477 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 415, + 481, + 477 + ], + "spans": [ + { + "bbox": [ + 130, + 415, + 481, + 477 + ], + "type": "text", + "content": "After completing the optimization, the final control trajectory is yielded by model predictive control (MPC) [18] based on the optimized parameterized simulator and the hand trajectory " + }, + { + "bbox": [ + 130, + 415, + 481, + 477 + ], + "type": "inline_equation", + "content": "\\mathcal{A}" + }, + { + "bbox": [ + 130, + 415, + 481, + 477 + ], + "type": "text", + "content": ". Specifically, in each step, the current and the following controls in several subsequent frames are optimized to reduce the tracking error. More details are deferred to the Supp." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 131, + 483, + 230, + 496 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 131, + 483, + 230, + 496 + ], + "spans": [ + { + "bbox": [ + 131, + 483, + 230, + 496 + ], + "type": "text", + "content": "4 Experiments" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 130, + 498, + 482, + 666 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 498, + 482, + 666 + ], + "spans": [ + { + "bbox": [ + 130, + 498, + 482, + 666 + ], + "type": "text", + "content": "We conduct extensive experiments to demonstrate the effectiveness of our method. The evaluation dataset is constructed from three HOI datasets with both single-hand and bimanual manipulations (with rigid objects), with complex manipulations with non-trivial object movements, and rich and changing contacts involved (see Section 4.1). We use Shadow hand [49] and test in two simulators widely used in the embodied AI community: Bullet [9] and Isaac Gym [33]. We compare our method with both model-free approaches and model-based strategies and demonstrate the superiority of our method both quantitatively and qualitatively. We can track complex contact-rich manipulations with large object rotations, back-and-forth object movements, and changing contacts successfully in both of the two simulators, while the best-performed baseline fails (see Section 4.2, Fig. 4). On average, we boost the tracking success rate by " + }, + { + "bbox": [ + 130, + 498, + 482, + 666 + ], + "type": "inline_equation", + "content": "11\\%+" + }, + { + "bbox": [ + 130, + 498, + 482, + 666 + ], + "type": "text", + "content": " from the previous best-performed (see Section 4.2). We make further analysis and discussions and show that the core philosophy of our work, optimizing through" + } + ] + } + ], + "index": 7 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 184, + 91, + 447, + 103 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 184, + 91, + 447, + 103 + ], + "spans": [ + { + "bbox": [ + 184, + 91, + 447, + 103 + ], + "type": "text", + "content": "Quasi-Physical Simulators for Dexterous Manipulations Transfer" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 474, + 92, + 480, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 474, + 92, + 480, + 100 + ], + "spans": [ + { + "bbox": [ + 474, + 92, + 480, + 100 + ], + "type": "text", + "content": "9" + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 8 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 135, + 105, + 317, + 402 + ], + "blocks": [ + { + "bbox": [ + 135, + 105, + 317, + 402 + ], + "lines": [ + { + "bbox": [ + 135, + 105, + 317, + 402 + ], + "spans": [ + { + "bbox": [ + 135, + 105, + 317, + 402 + ], + "type": "image", + "image_path": "1643c1ef6225e2e5da463e90dde2d591e5feb855311fe0a7f69da54f8a7c43d3.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 130, + 404, + 480, + 427 + ], + "lines": [ + { + "bbox": [ + 130, + 404, + 480, + 427 + ], + "spans": [ + { + "bbox": [ + 130, + 404, + 480, + 427 + ], + "type": "text", + "content": "Fig. 4: Qualitative comparisons. Please refer to our website and the accompanying video for animated results." + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "image_caption" + } + ], + "index": 2 + }, + { + "type": "image", + "bbox": [ + 320, + 105, + 480, + 402 + ], + "blocks": [ + { + "bbox": [ + 320, + 105, + 480, + 402 + ], + "lines": [ + { + "bbox": [ + 320, + 105, + 480, + 402 + ], + "spans": [ + { + "bbox": [ + 320, + 105, + 480, + 402 + ], + "type": "image", + "image_path": "a18c830fb7a8710b8a9fb490b5e13cf1260fdd493357208fc08afa10d732cde9.jpg" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_body" + } + ], + "index": 3 + }, + { + "bbox": [ + 130, + 439, + 480, + 463 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 439, + 480, + 463 + ], + "spans": [ + { + "bbox": [ + 130, + 439, + 480, + 463 + ], + "type": "text", + "content": "a quasi-physics curriculum, is potentially general and can help improve the performance of a model-free baseline (see Section 4.3)." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 131, + 479, + 272, + 491 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 131, + 479, + 272, + 491 + ], + "spans": [ + { + "bbox": [ + 131, + 479, + 272, + 491 + ], + "type": "text", + "content": "4.1 Experimental Settings" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 130, + 498, + 482, + 666 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 498, + 482, + 666 + ], + "spans": [ + { + "bbox": [ + 130, + 498, + 482, + 666 + ], + "type": "text", + "content": "Datasets. Our evaluation dataset is compiled from three distinct sources, namely GRAB [53], containing single-hand interactions with daily objects, TACO [32], containing humans manipulating tools, and ARCTIC [13] with bimanual manipulations. For GRAB, we randomly sample a manipulation trajectory for each object. If its manipulation is extremely simple, we additionally sample one trajectory for it. The object is not considered if its corresponding manipulation is bimanual such as binoculars, involves other body parts such as bowl, or with detailed part movements such as the game controller. The number of manipulation sequences from GRAB is 27. For TACO [32], we acquire data by contacting authors. We randomly select one sequence for each right-hand tool object. Sequences with very low quality like erroneous object motions are excluded. 14 trajectories in total are selected finally. For ARCTIC [13], we randomly select one sequence for each object from its available manipulation trajectories, resulting in 10 sequences in total. More details are deferred to the Supp." + } + ] + } + ], + "index": 7 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "spans": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "type": "text", + "content": "10" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 166, + 91, + 217, + 100 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 166, + 91, + 217, + 100 + ], + "spans": [ + { + "bbox": [ + 166, + 91, + 217, + 100 + ], + "type": "text", + "content": "X. Liu et al." + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 9 + }, + { + "para_blocks": [ + { + "bbox": [ + 130, + 115, + 482, + 315 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 115, + 482, + 315 + ], + "spans": [ + { + "bbox": [ + 130, + 115, + 482, + 315 + ], + "type": "text", + "content": "Metrics. We introduce three distinct metrics to assess the quality of object tracking, the accuracy of hand tracking, and the overall success of the tracking task: 1) Per-frame average object rotation error: " + }, + { + "bbox": [ + 130, + 115, + 482, + 315 + ], + "type": "inline_equation", + "content": "R_{\\mathrm{err}} = \\frac{1}{N}\\sum_{n=1}^{N}(1 - (\\mathbf{q}_n \\cdot \\hat{\\mathbf{q}}_n))" + }, + { + "bbox": [ + 130, + 115, + 482, + 315 + ], + "type": "text", + "content": ", where " + }, + { + "bbox": [ + 130, + 115, + 482, + 315 + ], + "type": "inline_equation", + "content": "\\mathbf{q}_n" + }, + { + "bbox": [ + 130, + 115, + 482, + 315 + ], + "type": "text", + "content": " is the ground-truth orientation and " + }, + { + "bbox": [ + 130, + 115, + 482, + 315 + ], + "type": "inline_equation", + "content": "\\hat{\\mathbf{q}}_n" + }, + { + "bbox": [ + 130, + 115, + 482, + 315 + ], + "type": "text", + "content": " is the tracked result, represented in quaternion. 2) Per-frame average object translation error: " + }, + { + "bbox": [ + 130, + 115, + 482, + 315 + ], + "type": "inline_equation", + "content": "T_{\\mathrm{err}} = \\frac{1}{N}\\sum_{n=1}^{N}\\|\\mathbf{t}_n - \\hat{\\mathbf{t}}_n\\|" + }, + { + "bbox": [ + 130, + 115, + 482, + 315 + ], + "type": "text", + "content": ", where " + }, + { + "bbox": [ + 130, + 115, + 482, + 315 + ], + "type": "inline_equation", + "content": "\\mathbf{t}" + }, + { + "bbox": [ + 130, + 115, + 482, + 315 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 130, + 115, + 482, + 315 + ], + "type": "inline_equation", + "content": "\\mathbf{t}_n" + }, + { + "bbox": [ + 130, + 115, + 482, + 315 + ], + "type": "text", + "content": " are ground-truth and tracked translations respectively. 3) Mean Per-Joint Position Error (MPJPE) = " + }, + { + "bbox": [ + 130, + 115, + 482, + 315 + ], + "type": "inline_equation", + "content": "\\frac{1}{N}\\sum_{n=1}^{N}\\|\\mathbf{J}_n - \\hat{\\mathbf{J}}_n\\|" + }, + { + "bbox": [ + 130, + 115, + 482, + 315 + ], + "type": "text", + "content": " [20, 44, 58], where " + }, + { + "bbox": [ + 130, + 115, + 482, + 315 + ], + "type": "inline_equation", + "content": "\\mathbf{J}_n" + }, + { + "bbox": [ + 130, + 115, + 482, + 315 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 130, + 115, + 482, + 315 + ], + "type": "inline_equation", + "content": "\\hat{\\mathbf{J}}_n" + }, + { + "bbox": [ + 130, + 115, + 482, + 315 + ], + "type": "text", + "content": " are keypoints of GT human hand and the simulated robot hand respectively. We manually define the keypoints and the correspondences to the human hand keypoints for the Shadow hand. 4) Per-frame average hand Chamfer Distance: " + }, + { + "bbox": [ + 130, + 115, + 482, + 315 + ], + "type": "inline_equation", + "content": "\\mathrm{CD} = \\frac{1}{N}\\sum_{n=1}^{N}\\mathrm{Chamfer-Distance}(\\mathbf{H}_n - \\hat{\\mathbf{H}}_n)" + }, + { + "bbox": [ + 130, + 115, + 482, + 315 + ], + "type": "text", + "content": ", for evaluating whether the Shadow hand can \"densely\" track the demonstration. 5) Success rate: a tracking is regarded as successful if the object rotation error " + }, + { + "bbox": [ + 130, + 115, + 482, + 315 + ], + "type": "inline_equation", + "content": "R_{\\mathrm{err}}" + }, + { + "bbox": [ + 130, + 115, + 482, + 315 + ], + "type": "text", + "content": ", object translation error " + }, + { + "bbox": [ + 130, + 115, + 482, + 315 + ], + "type": "inline_equation", + "content": "T_{\\mathrm{err}}" + }, + { + "bbox": [ + 130, + 115, + 482, + 315 + ], + "type": "text", + "content": ", and the hand tracking error MPJPE are smaller than their corresponding threshold. Three success rates are calculated using three different thresholds, namely " + }, + { + "bbox": [ + 130, + 115, + 482, + 315 + ], + "type": "inline_equation", + "content": "10^{\\circ} - 10cm - 10cm" + }, + { + "bbox": [ + 130, + 115, + 482, + 315 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 130, + 115, + 482, + 315 + ], + "type": "inline_equation", + "content": "15^{\\circ} - 15cm - 15cm" + }, + { + "bbox": [ + 130, + 115, + 482, + 315 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 130, + 316, + 482, + 581 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 316, + 482, + 581 + ], + "spans": [ + { + "bbox": [ + 130, + 316, + 482, + 581 + ], + "type": "text", + "content": "Baselines. We compare with two trends of baselines. For model-free approaches, since there is no prior work with exactly the same problem setting as us, we try to modify and improve a goal-driven rigid object manipulation method DGrasp [8] into two methods for tracking: 1) DGrasp-Base, where the method is almost kept with same with the original DGrasp. We use the first frame where the hand and the object are in contact with each other as the reference frame. Then the policy is trained to grasp the object according to the reference hand and object goal at first. After that, only the root is guided to complete the task. 2) DGraspTracking, where we divide the whole sequence into several subsequences, each of which has 10 frames, and define the end frame of the subsequence as the reference frame. Then the grasping policy is used to guide the hand and gradually track the object according to the hand and the object pose of each reference frame. We improve the DGrasp-Tracking by optimizing the policy through the quasi-physical curriculum and creating \"DGrasp-Tracking (w/ Curriculum)\" trying to improve its performance. For model-based methods, we compare with Control-VAE [66] and traditional MPC approaches. For Control-VAE, we modify its implementation for the manipulation tracking task. We additionally consider three differentiable physics models to conduct model-predictive control for solving the task. Taking the analytical model with the most tightened contact model as the base model (\"MPC (w/ base sim.)\"), we further augment it with a general state-of-the-art contact smoothing for robot manipulation [52] and create \"MPC (w/ base sim. w/ soften)\". Details of baseline models are deferred to the Supp." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 130, + 582, + 482, + 666 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 582, + 482, + 666 + ], + "spans": [ + { + "bbox": [ + 130, + 582, + 482, + 666 + ], + "type": "text", + "content": "Training and evaluation settings. The physics curriculum is composed of three stages. In the first stage, the parameter " + }, + { + "bbox": [ + 130, + 582, + 482, + 666 + ], + "type": "inline_equation", + "content": "\\alpha" + }, + { + "bbox": [ + 130, + 582, + 482, + 666 + ], + "type": "text", + "content": " varies from 0.1 to 0.0 and the contact model stiffness is relaxed to the softest level. In the second stage, " + }, + { + "bbox": [ + 130, + 582, + 482, + 666 + ], + "type": "inline_equation", + "content": "\\alpha" + }, + { + "bbox": [ + 130, + 582, + 482, + 666 + ], + "type": "text", + "content": " is fixed and the contact model stiffness varies from the softest version to the most tightened level gradually through eight stages. Details w.r.t. parameter settings are deferred to the Supp. In the first two stages, we alternately optimize the trajectory " + }, + { + "bbox": [ + 130, + 582, + 482, + 666 + ], + "type": "inline_equation", + "content": "\\mathcal{A}" + }, + { + "bbox": [ + 130, + 582, + 482, + 666 + ], + "type": "text", + "content": " and parameters " + }, + { + "bbox": [ + 130, + 582, + 482, + 666 + ], + "type": "inline_equation", + "content": "S" + }, + { + "bbox": [ + 130, + 582, + 482, + 666 + ], + "type": "text", + "content": ". In each optimization iteration, the " + }, + { + "bbox": [ + 130, + 582, + 482, + 666 + ], + "type": "inline_equation", + "content": "\\mathcal{A}" + }, + { + "bbox": [ + 130, + 582, + 482, + 666 + ], + "type": "text", + "content": " is optimized" + } + ] + } + ], + "index": 4 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 184, + 90, + 447, + 102 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 184, + 90, + 447, + 102 + ], + "spans": [ + { + "bbox": [ + 184, + 90, + 447, + 102 + ], + "type": "text", + "content": "Quasi-Physical Simulators for Dexterous Manipulations Transfer" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 470, + 91, + 479, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 470, + 91, + 479, + 100 + ], + "spans": [ + { + "bbox": [ + 470, + 91, + 479, + 100 + ], + "type": "text", + "content": "11" + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 10 + }, + { + "para_blocks": [ + { + "type": "table", + "bbox": [ + 133, + 137, + 479, + 273 + ], + "blocks": [ + { + "bbox": [ + 132, + 114, + 479, + 136 + ], + "lines": [ + { + "bbox": [ + 132, + 114, + 479, + 136 + ], + "spans": [ + { + "bbox": [ + 132, + 114, + 479, + 136 + ], + "type": "text", + "content": "Table 1: Quantitative evaluations and comparisons to baselines. Bold red numbers for best values and italic blue values for the second best-performed ones." + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 133, + 137, + 479, + 273 + ], + "lines": [ + { + "bbox": [ + 133, + 137, + 479, + 273 + ], + "spans": [ + { + "bbox": [ + 133, + 137, + 479, + 273 + ], + "type": "table", + "html": "
SimulatorMethod\\( {R}_{\\text{err }}\\left( {{}^{ \\circ }, \\downarrow }\\right) \\)\\( {T}_{\\text{err }}\\left( {{cm}, \\downarrow }\\right) \\)MPJPE (mm,↓)CD (mm,↓)Success Rate (%,↑)
BulletModel FreeDGrasp-Base44.245.8240.5516.370/13.73/15.69
DGrasp-Tracking44.455.0437.5614.720/15.69/15.69
DGrasp-Tracking (w/ curric.)33.864.6030.4713.537.84/23.53/37.25
Model BasedControl-VAE42.452.7325.2110.940/15.68/23.53
MPC (w/ base sim.)32.563.6724.6210.800/15.68/31.37
MPC (w/ base sim. w/ soften)31.893.6328.2611.310/21.57/37.25
Ours24.211.9724.409.8527.45/37.25/58.82
Isaac GymModel FreeDGrasp-Base36.414.5650.9718.780/7.84/7.84
DGrasp-Tracking44.715.5741.5316.720/0/7.84
DGrasp-Tracking (w/ curric.)38.755.1340.0916.260/23.53/31.37
Model BasedControl-VAE35.404.6127.6313.170/13.73/29.41
MPC (w/ base sim.)37.234.7323.199.750/15.69/31.37
MPC (w/ base sim. w/ soften)36.404.4623.2710.340/9.80/23.53
Ours25.972.0825.3310.3121.57/43.14/56.86
", + "image_path": "a697834157e3e46fab2fc4995163e9266548dcbd52b4c04adc09a8ec774c8151.jpg" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "table_body" + } + ], + "index": 3 + }, + { + "bbox": [ + 130, + 286, + 480, + 346 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 286, + 480, + 346 + ], + "spans": [ + { + "bbox": [ + 130, + 286, + 480, + 346 + ], + "type": "text", + "content": "for 100 steps while " + }, + { + "bbox": [ + 130, + 286, + 480, + 346 + ], + "type": "inline_equation", + "content": "S" + }, + { + "bbox": [ + 130, + 286, + 480, + 346 + ], + "type": "text", + "content": " is optimized for 1000 steps. In the third stage, " + }, + { + "bbox": [ + 130, + 286, + 480, + 346 + ], + "type": "inline_equation", + "content": "\\mathcal{A}" + }, + { + "bbox": [ + 130, + 286, + 480, + 346 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 130, + 286, + 480, + 346 + ], + "type": "inline_equation", + "content": "\\psi" + }, + { + "bbox": [ + 130, + 286, + 480, + 346 + ], + "type": "text", + "content": " are optimized for 256 steps in each iteration. For time-stepping, " + }, + { + "bbox": [ + 130, + 286, + 480, + 346 + ], + "type": "inline_equation", + "content": "\\mathrm{dt}" + }, + { + "bbox": [ + 130, + 286, + 480, + 346 + ], + "type": "text", + "content": " is set to " + }, + { + "bbox": [ + 130, + 286, + 480, + 346 + ], + "type": "inline_equation", + "content": "5 \\times 10^{-4}" + }, + { + "bbox": [ + 130, + 286, + 480, + 346 + ], + "type": "text", + "content": " in the parameterized and the target simulators. The articulated multi-body is controlled by joint motors and root velocities in the parameterized quasi-physical simulator while PD control [54] is leveraged in the target simulators." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 131, + 354, + 330, + 366 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 131, + 354, + 330, + 366 + ], + "spans": [ + { + "bbox": [ + 131, + 354, + 330, + 366 + ], + "type": "text", + "content": "4.2 Dexterous Manipulating Tracking" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 130, + 368, + 480, + 474 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 368, + 480, + 474 + ], + "spans": [ + { + "bbox": [ + 130, + 368, + 480, + 474 + ], + "type": "text", + "content": "We conducted thorough experiments in two widely used simulators [9, 33]. We treat them as realistic simulated physical environments with high fidelity and wish to track the manipulation in them. In summary, we can control a dexterous hand to complete a wide range of the manipulation tracking tasks with nontrivial object movements and changing contacts. As presented in Table 1, we can achieve significantly higher success rates calculated under three thresholds than the best-performed baseline in both tested simulators. Fig. 4 showcases qualitative examples and comparisons. Please check out our website and video for animated results." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 130, + 475, + 480, + 570 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 475, + 480, + 570 + ], + "spans": [ + { + "bbox": [ + 130, + 475, + 480, + 570 + ], + "type": "text", + "content": "Complex manipulations. For examples shown in Fig. 4, we can complete the tracking task on examples with large object re-orientations and complicated tool-using (Fig. " + }, + { + "bbox": [ + 130, + 475, + 480, + 570 + ], + "type": "inline_equation", + "content": "(a,b,c)" + }, + { + "bbox": [ + 130, + 475, + 480, + 570 + ], + "type": "text", + "content": "). However, DGrasp-Tracking fails to establish sufficient contact for correctly manipulating the object. In more detail, in Fig. 4(b), the bunny gradually bounced out from its hand in Bullet, while our method does not suffer from this difficulty. In Fig. 4(c), the spoon can be successfully picked up and waved back-and-forth in our method, while DGrasp-Tracking loses the track right from the start." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 130, + 571, + 480, + 619 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 571, + 480, + 619 + ], + "spans": [ + { + "bbox": [ + 130, + 571, + 480, + 619 + ], + "type": "text", + "content": "Bimanual manipulations. We are also capable of tracking bimanual manipulations. As shown in the example in Fig. 4(d), where two hands collaborate to relocate the object, DGrasp-Tracking fails to accurately track the object, while our method significantly outperforms it." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 131, + 621, + 327, + 633 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 131, + 621, + 327, + 633 + ], + "spans": [ + { + "bbox": [ + 131, + 621, + 327, + 633 + ], + "type": "text", + "content": "4.3 Further Analysis and Discussions" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 130, + 641, + 480, + 665 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 641, + 480, + 665 + ], + "spans": [ + { + "bbox": [ + 130, + 641, + 480, + 665 + ], + "type": "text", + "content": "Could model-free methods benefit from the physics curriculum? In addition to the demonstrated merits of our quasi-physical simulators, we further" + } + ] + } + ], + "index": 10 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "spans": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "type": "text", + "content": "12" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 166, + 91, + 217, + 100 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 166, + 91, + 217, + 100 + ], + "spans": [ + { + "bbox": [ + 166, + 91, + 217, + 100 + ], + "type": "text", + "content": "X. Liu et al." + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 11 + }, + { + "para_blocks": [ + { + "type": "table", + "bbox": [ + 132, + 137, + 479, + 213 + ], + "blocks": [ + { + "bbox": [ + 131, + 114, + 482, + 136 + ], + "lines": [ + { + "bbox": [ + 131, + 114, + 482, + 136 + ], + "spans": [ + { + "bbox": [ + 131, + 114, + 482, + 136 + ], + "type": "text", + "content": "Table 2: Ablation studies. Bold red numbers for best values and italic blue values for the second best-performed ones. The simulation environment is Bullet." + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 132, + 137, + 479, + 213 + ], + "lines": [ + { + "bbox": [ + 132, + 137, + 479, + 213 + ], + "spans": [ + { + "bbox": [ + 132, + 137, + 479, + 213 + ], + "type": "table", + "html": "
MethodRerr(°,↓)Terr(cm,↓)MPJPE(mm,↓)CD(mm,↓)Success Rate(%,↑)
Ours w/o Analytical Sim.44.274.3929.8412.910/13.73/25.49
Ours w/o Residual Physics33.693.8126.5710.345.88/23.53/41.18
Ours w/o Local Force NN35.982.9032.8712.440/19.61/35.29
Ours w/o Curriculum42.404.8732.6113.370/17.64/29.41
Ours w/ Curriculum II29.582.3331.6110.2911.76/27.45/50.98
Ours24.211.9724.409.8527.45/37.25/58.82
", + "image_path": "3b3aba97c5dd7aa347311ff5b8567d5184c9907db1ce256bac4affb1bcf0c09c.jpg" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "table_body" + } + ], + "index": 3 + }, + { + "bbox": [ + 130, + 214, + 482, + 299 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 214, + 482, + 299 + ], + "spans": [ + { + "bbox": [ + 130, + 214, + 482, + 299 + ], + "type": "text", + "content": "explore whether model-free strategies can benefit from them. We introduce the \"DGrasp-Tracking (w/ Curriculum)\" method and compare its performance with the original DGrasp-Tracking model. As shown in Table 1 and the visual comparisons in Fig. 6, the DGrasp-Tracking model indeed benefits from a well-designed physics curriculum. For example, as illustrated in Fig. 6, the curriculum can significantly improve its performance, enabling it to nearly complete challenging tracking tasks where the original version struggles." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 132, + 304, + 246, + 319 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 304, + 246, + 319 + ], + "spans": [ + { + "bbox": [ + 132, + 304, + 246, + 319 + ], + "type": "text", + "content": "5 Ablation Study" + } + ] + } + ], + "index": 5 + }, + { + "type": "image", + "bbox": [ + 185, + 331, + 329, + 482 + ], + "blocks": [ + { + "bbox": [ + 185, + 331, + 329, + 482 + ], + "lines": [ + { + "bbox": [ + 185, + 331, + 329, + 482 + ], + "spans": [ + { + "bbox": [ + 185, + 331, + 329, + 482 + ], + "type": "image", + "image_path": "74859ef9bb99bc24c0df62b177b3f4806a4684985c9b0d55fdbe24355cb4c1c7.jpg" + } + ] + } + ], + "index": 6, + "angle": 0, + "type": "image_body" + } + ], + "index": 6 + }, + { + "type": "image", + "bbox": [ + 332, + 339, + 425, + 485 + ], + "blocks": [ + { + "bbox": [ + 332, + 339, + 425, + 485 + ], + "lines": [ + { + "bbox": [ + 332, + 339, + 425, + 485 + ], + "spans": [ + { + "bbox": [ + 332, + 339, + 425, + 485 + ], + "type": "image", + "image_path": "ec93451ebcec063eefb956eefd5a67143f2152306343e6a9965a064f065ccbb5.jpg" + } + ] + } + ], + "index": 7, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 131, + 483, + 480, + 506 + ], + "lines": [ + { + "bbox": [ + 131, + 483, + 480, + 506 + ], + "spans": [ + { + "bbox": [ + 131, + 483, + 480, + 506 + ], + "type": "text", + "content": "Fig. 5: (a) Qualitative comparisons between our full method and the ablated models; (b) Training loss curve comparisons; (c) Tracking loss curve comparisons." + } + ] + } + ], + "index": 8, + "angle": 0, + "type": "image_caption" + } + ], + "index": 7 + }, + { + "type": "image", + "bbox": [ + 136, + 507, + 482, + 605 + ], + "blocks": [ + { + "bbox": [ + 136, + 507, + 482, + 605 + ], + "lines": [ + { + "bbox": [ + 136, + 507, + 482, + 605 + ], + "spans": [ + { + "bbox": [ + 136, + 507, + 482, + 605 + ], + "type": "image", + "image_path": "57fbeae52f7988b590d585056da1c3efa289c1cd94a717ddec72ad5be60d0e90.jpg" + } + ] + } + ], + "index": 9, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 131, + 612, + 482, + 635 + ], + "lines": [ + { + "bbox": [ + 131, + 612, + 482, + 635 + ], + "spans": [ + { + "bbox": [ + 131, + 612, + 482, + 635 + ], + "type": "text", + "content": "Fig. 6: Visual evidence on boosting DGrasp-Tracking's performance via optimizing it through a physics curriculum." + } + ] + } + ], + "index": 10, + "angle": 0, + "type": "image_caption" + } + ], + "index": 9 + }, + { + "bbox": [ + 130, + 641, + 482, + 666 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 641, + 482, + 666 + ], + "spans": [ + { + "bbox": [ + 130, + 641, + 482, + 666 + ], + "type": "text", + "content": "We conduct a wide range of ablation studies to validate the effectiveness of some of our crucial designs, including the parameterized analytical physics" + } + ] + } + ], + "index": 11 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 184, + 91, + 448, + 102 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 184, + 91, + 448, + 102 + ], + "spans": [ + { + "bbox": [ + 184, + 91, + 448, + 102 + ], + "type": "text", + "content": "Quasi-Physical Simulators for Dexterous Manipulations Transfer" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 470, + 91, + 481, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 470, + 91, + 481, + 100 + ], + "spans": [ + { + "bbox": [ + 470, + 91, + 481, + 100 + ], + "type": "text", + "content": "13" + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 12 + }, + { + "para_blocks": [ + { + "bbox": [ + 132, + 116, + 480, + 152 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 116, + 480, + 152 + ], + "spans": [ + { + "bbox": [ + 132, + 116, + 480, + 152 + ], + "type": "text", + "content": "model, the parameterized residual physics, the role of the local force network, the necessity of introducing a physics curriculum into the optimization, and how the design on the curriculum stages affects the result." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 132, + 152, + 481, + 285 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 152, + 481, + 285 + ], + "spans": [ + { + "bbox": [ + 132, + 152, + 481, + 285 + ], + "type": "text", + "content": "Parameterized analytical model. The skeleton of the quasi-physical simulator is an analytical physics model. The intuition is that the parameterized simulator with such physical bias can be optimized towards a realistic simulator more easily than training pure neural networks for approximating. To validate this, we ablate the analytical model and use neural networks to approximate physics in Bullet directly (denoted as \"Ours w/o Analytical Sim\"). The quantitative (Table 2) and qualitative (Fig. 5) results indicate that the physical biases brought by the analytical model could help the parameterized simulator to learn better physics in the contact-rich scenario. For instance, in the example demonstrated in Fig. 5, the ablated version fails to guide the robot hand to successfully pinch the object in the second figure." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 132, + 286, + 481, + 357 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 286, + 481, + 357 + ], + "spans": [ + { + "bbox": [ + 132, + 286, + 481, + 357 + ], + "type": "text", + "content": "Parameterized residual physics. To validate the necessity of introducing residual force networks to close the gap between the physics modeled in the parameterized analytical simulator and that of a realistic simulator, we ablate the parameterized force network and create a version named \"Ours w/o Residual Physics\". Table 2 demonstrated its role in enabling the parameterized simulator to approximate realistic physics models." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 132, + 359, + 481, + 430 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 359, + 481, + 430 + ], + "spans": [ + { + "bbox": [ + 132, + 359, + 481, + 430 + ], + "type": "text", + "content": "Local residual force network. To adequately leverage state and contact-related information for predicting residual contact forces, we propose to use two types of networks: 1) a local force network for per contact pair residual forces and 2) a global network for additionally compensating. The local network is introduced for fine-grained approximation. We ablate this design and compare the result with our full model to validate this (see Fig. 5 and Table 1)." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 132, + 432, + 481, + 528 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 432, + 481, + 528 + ], + "spans": [ + { + "bbox": [ + 132, + 432, + 481, + 528 + ], + "type": "text", + "content": "Optimizing through an analytical physics curriculum. We further investigate the effectiveness of the analytical curriculum design and how its design influences the result. Specifically, we create two ablated versions: 1) \"Ours w/o Curriculum\", where the optimization starts directly from the parameterized analytical model with articulated rigid constraints tightened and the stiffest contact model, and 2) \"Ours w/ Curriculum II\", where we move some stages out from the original curriculum. Table 2 and Fig. 5 demonstrate that both the curriculum and the optimization path will affect the model's performance." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 132, + 535, + 317, + 548 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 535, + 317, + 548 + ], + "spans": [ + { + "bbox": [ + 132, + 535, + 317, + 548 + ], + "type": "text", + "content": "6 Conclusion and Limitations" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 132, + 556, + 481, + 628 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 556, + 481, + 628 + ], + "spans": [ + { + "bbox": [ + 132, + 556, + 481, + 628 + ], + "type": "text", + "content": "In this work, we investigate creating better simulators for solving complex robotic tasks involving complicated dynamics where the previous best-performed optimization strategy fails. We present a family of parameterized quasi-physical simulators that can be both programmed to relax various constraints for task optimization and can be tailored to approximate realistic physics. We tackle the difficult manipulation transfer task via a physics curriculum." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 132, + 629, + 481, + 665 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 629, + 481, + 665 + ], + "spans": [ + { + "bbox": [ + 132, + 629, + 481, + 665 + ], + "type": "text", + "content": "Limitations. The method is limited by the relatively simple spring-damper model for contact constraint relaxation. Introducing delicate analytical contact models to parameterized simulators is an interesting research direction." + } + ] + } + ], + "index": 9 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "spans": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "type": "text", + "content": "14" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 166, + 91, + 217, + 100 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 166, + 91, + 217, + 100 + ], + "spans": [ + { + "bbox": [ + 166, + 91, + 217, + 100 + ], + "type": "text", + "content": "X. Liu et al." + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 13 + }, + { + "para_blocks": [ + { + "bbox": [ + 133, + 114, + 240, + 129 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 114, + 240, + 129 + ], + "spans": [ + { + "bbox": [ + 133, + 114, + 240, + 129 + ], + "type": "text", + "content": "Acknowledgement" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 130, + 140, + 481, + 177 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 140, + 481, + 177 + ], + "spans": [ + { + "bbox": [ + 130, + 140, + 481, + 177 + ], + "type": "text", + "content": "We thank Rui Chen, Yiheng Guan, and Taoran Jiang for their valuable support in providing the Allegro hardware and setting up the environment setup during the rebuttal period." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 133, + 197, + 197, + 209 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 197, + 197, + 209 + ], + "spans": [ + { + "bbox": [ + 133, + 197, + 197, + 209 + ], + "type": "text", + "content": "References" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 134, + 223, + 481, + 665 + ], + "type": "list", + "angle": 0, + "index": 19, + "blocks": [ + { + "bbox": [ + 138, + 223, + 481, + 267 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 223, + 481, + 267 + ], + "spans": [ + { + "bbox": [ + 138, + 223, + 481, + 267 + ], + "type": "text", + "content": "1. Ajay, A., Wu, J., Fazeli, N., Bauza, M., Kaelbling, L.P., Tenenbaum, J.B., Rodriguez, A.: Augmenting physical simulators with stochastic neural networks: Case study of planar pushing and bouncing. In: 2018 IEEE/RSJ International Conference on Intelligent Robots and Systems (IROS). pp. 3066-3073. IEEE (2018)" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 138, + 268, + 481, + 300 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 268, + 481, + 300 + ], + "spans": [ + { + "bbox": [ + 138, + 268, + 481, + 300 + ], + "type": "text", + "content": "2. Akkaya, I., Andrychowicz, M., Chociej, M., Litwin, M., McGrew, B., Petron, A., Paino, A., Plappert, M., Powell, G., Ribas, R., et al.: Solving rubik's cube with a robot hand. arXiv preprint arXiv:1910.07113 (2019)" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 138, + 301, + 481, + 323 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 301, + 481, + 323 + ], + "spans": [ + { + "bbox": [ + 138, + 301, + 481, + 323 + ], + "type": "text", + "content": "3. Andrews, S., Erleben, K., Ferguson, Z.: Contact and friction simulation for computer graphics. In: ACM SIGGRAPH 2022 Courses, pp. 1-172 (2022)" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 138, + 323, + 481, + 366 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 323, + 481, + 366 + ], + "spans": [ + { + "bbox": [ + 138, + 323, + 481, + 366 + ], + "type": "text", + "content": "4. Andrychowicz, O.M., Baker, B., Chociej, M., Jozefowicz, R., McGrew, B., Pachocki, J., Petron, A., Plappert, M., Powell, G., Ray, A., et al.: Learning dexterous in-hand manipulation. The International Journal of Robotics Research 39(1), 3-20 (2020)" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 138, + 367, + 481, + 389 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 367, + 481, + 389 + ], + "spans": [ + { + "bbox": [ + 138, + 367, + 481, + 389 + ], + "type": "text", + "content": "5. Baraff, D.: An introduction to physically based modeling: rigid body simulation ii—nonpenetration constraints. SIGGRAPH course notes pp. D31–D68 (1997)" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 138, + 389, + 481, + 433 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 389, + 481, + 433 + ], + "spans": [ + { + "bbox": [ + 138, + 389, + 481, + 433 + ], + "type": "text", + "content": "6. Chen, T., Tippur, M., Wu, S., Kumar, V., Adelson, E., Agrawal, P.: Visual dexterity: In-hand reorientation of novel and complex object shapes. Science Robotics 8(84), eadc9244 (2023). https://doi.org/10.1126/scirobotics.adc9244, https://www.science.org/doi/abs/10.1126/scirobotics.adc9244" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 138, + 434, + 481, + 455 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 434, + 481, + 455 + ], + "spans": [ + { + "bbox": [ + 138, + 434, + 481, + 455 + ], + "type": "text", + "content": "7. Chen, T., Xu, J., Agrawal, P.: A system for general in-hand object re-orientation. Conference on Robot Learning (2021)" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 138, + 456, + 481, + 499 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 456, + 481, + 499 + ], + "spans": [ + { + "bbox": [ + 138, + 456, + 481, + 499 + ], + "type": "text", + "content": "8. Christen, S., Kocabas, M., Aksan, E., Hwangbo, J., Song, J., Hilliges, O.: D-grasp: Physically plausible dynamic grasp synthesis for hand-object interactions. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 20577-20586 (2022)" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 138, + 500, + 481, + 521 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 500, + 481, + 521 + ], + "spans": [ + { + "bbox": [ + 138, + 500, + 481, + 521 + ], + "type": "text", + "content": "9. Coumans, E., Bai, Y.: Pybullet, a python module for physics simulation for games, robotics and machine learning (2016)" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 138, + 522, + 481, + 544 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 522, + 481, + 544 + ], + "spans": [ + { + "bbox": [ + 138, + 522, + 481, + 544 + ], + "type": "text", + "content": "0. Deng, Y., Yu, H.X., Wu, J., Zhu, B.: Learning vortex dynamics for fluid inference and prediction. arXiv preprint arXiv:2301.11494 (2023)" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 138, + 544, + 481, + 566 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 544, + 481, + 566 + ], + "spans": [ + { + "bbox": [ + 138, + 544, + 481, + 566 + ], + "type": "text", + "content": "1. Du, T.: Deep learning for physics simulation. In: ACM SIGGRAPH 2023 Courses, pp. 1-25 (2023)" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 138, + 567, + 481, + 599 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 567, + 481, + 599 + ], + "spans": [ + { + "bbox": [ + 138, + 567, + 481, + 599 + ], + "type": "text", + "content": "2. Dunlavy, D.M., O'Leary, D.P.: Homotopy optimization methods for global optimization. Tech. rep., Sandia National Laboratories (SNL), Albuquerque, NM, and Livermore, CA ... (2005)" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 134, + 600, + 481, + 643 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 134, + 600, + 481, + 643 + ], + "spans": [ + { + "bbox": [ + 134, + 600, + 481, + 643 + ], + "type": "text", + "content": "13. Fan, Z., Taheri, O., Tzionas, D., Kocabas, M., Kaufmann, M., Black, M.J., Hilliges, O.: ARCTIC: A dataset for dexterous bimanual hand-object manipulation. In: Proceedings IEEE Conference on Computer Vision and Pattern Recognition (CVPR) (2023)" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 134, + 644, + 481, + 665 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 134, + 644, + 481, + 665 + ], + "spans": [ + { + "bbox": [ + 134, + 644, + 481, + 665 + ], + "type": "text", + "content": "14. Featherstone, R.: Rigid body dynamics algorithms (2007), https://apisemantic scholar.org/CorpusID:58437819" + } + ] + } + ], + "index": 18 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 184, + 91, + 447, + 102 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 184, + 91, + 447, + 102 + ], + "spans": [ + { + "bbox": [ + 184, + 91, + 447, + 102 + ], + "type": "text", + "content": "Quasi-Physical Simulators for Dexterous Manipulations Transfer" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 470, + 92, + 480, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 470, + 92, + 480, + 100 + ], + "spans": [ + { + "bbox": [ + 470, + 92, + 480, + 100 + ], + "type": "text", + "content": "15" + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 14 + }, + { + "para_blocks": [ + { + "bbox": [ + 132, + 116, + 481, + 665 + ], + "type": "list", + "angle": 0, + "index": 21, + "blocks": [ + { + "bbox": [ + 133, + 116, + 481, + 149 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 116, + 481, + 149 + ], + "spans": [ + { + "bbox": [ + 133, + 116, + 481, + 149 + ], + "type": "text", + "content": "15. Freeman, C.D., Frey, E., Raichuk, A., Girgin, S., Mordatch, I., Bachem, O.: Brax-a differentiable physics engine for large scale rigid body simulation. arXiv preprint arXiv:2106.13281 (2021)" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 133, + 149, + 481, + 182 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 149, + 481, + 182 + ], + "spans": [ + { + "bbox": [ + 133, + 149, + 481, + 182 + ], + "type": "text", + "content": "16. Fussell, L., Bergamin, K., Holden, D.: Supertrack: Motion tracking for physically simulated characters using supervised learning. ACM Transactions on Graphics (TOG) 40(6), 1-13 (2021)" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 132, + 182, + 481, + 203 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 182, + 481, + 203 + ], + "spans": [ + { + "bbox": [ + 132, + 182, + 481, + 203 + ], + "type": "text", + "content": "17. Gao, J., Michelis, M.Y., Spielberg, A., Katzschmann, R.K.: Sim-to-real of soft robots with learned residual physics. arXiv preprint arXiv:2402.01086 (2024)" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 132, + 203, + 481, + 224 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 203, + 481, + 224 + ], + "spans": [ + { + "bbox": [ + 132, + 203, + 481, + 224 + ], + "type": "text", + "content": "18. Garcia, C.E., Prett, D.M., Morari, M.: Model predictive control: Theory and practice—a survey. Automatica 25(3), 335-348 (1989)" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 132, + 224, + 481, + 257 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 224, + 481, + 257 + ], + "spans": [ + { + "bbox": [ + 132, + 224, + 481, + 257 + ], + "type": "text", + "content": "19. Geilinger, M., Hahn, D., Zehnder, J., Bächer, M., Thomaszewski, B., Coros, S.: Add: Analytically differentiable dynamics for multi-body systems with frictional contact. ACM Transactions on Graphics (TOG) 39(6), 1-15 (2020)" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 132, + 257, + 481, + 289 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 257, + 481, + 289 + ], + "spans": [ + { + "bbox": [ + 132, + 257, + 481, + 289 + ], + "type": "text", + "content": "20. Grandia, R., Farshidian, F., Knoop, E., Schumacher, C., Hutter, M., Bächer, M.: Doc: Differentiable optimal control for retargeting motions onto legged robots. ACM Transactions on Graphics (TOG) 42(4), 1-14 (2023)" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 132, + 289, + 481, + 322 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 289, + 481, + 322 + ], + "spans": [ + { + "bbox": [ + 132, + 289, + 481, + 322 + ], + "type": "text", + "content": "21. Gupta, A., Eppner, C., Levine, S., Abbeel, P.: Learning dexterous manipulation for a soft robotic hand from human demonstrations. In: 2016 IEEE/RSJ International Conference on Intelligent Robots and Systems (IROS). pp. 3786-3793. IEEE (2016)" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 132, + 322, + 481, + 365 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 322, + 481, + 365 + ], + "spans": [ + { + "bbox": [ + 132, + 322, + 481, + 365 + ], + "type": "text", + "content": "22. Heiden, E., Millard, D., Coumans, E., Sheng, Y., Sukhatme, G.S.: Neuralsim: Augmenting differentiable simulators with neural networks. In: 2021 IEEE International Conference on Robotics and Automation (ICRA). pp. 9474-9481. IEEE (2021)" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 132, + 365, + 481, + 386 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 365, + 481, + 386 + ], + "spans": [ + { + "bbox": [ + 132, + 365, + 481, + 386 + ], + "type": "text", + "content": "23. Howell, T.A., Le Cleac'h, S., Kolter, J.Z., Schwager, M., Manchester, Z.: Dojo: A differentiable simulator for robotics. arXiv preprint arXiv:2203.00806 9 (2022)" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 132, + 386, + 481, + 418 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 386, + 481, + 418 + ], + "spans": [ + { + "bbox": [ + 132, + 386, + 481, + 418 + ], + "type": "text", + "content": "24. Howell, T.A., Le Cleac'h, S., Singh, S., Florence, P., Manchester, Z., Sindhwani, V.: Trajectory optimization with optimization-based dynamics. IEEE Robotics and Automation Letters 7(3), 6750-6757 (2022)" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 132, + 418, + 481, + 440 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 418, + 481, + 440 + ], + "spans": [ + { + "bbox": [ + 132, + 418, + 481, + 440 + ], + "type": "text", + "content": "25. Hwangbo, J., Lee, J., Hutter, M.: Per-contact iteration method for solving contact dynamics. IEEE Robotics and Automation Letters 3(2), 895-902 (2018)" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 132, + 440, + 481, + 472 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 440, + 481, + 472 + ], + "spans": [ + { + "bbox": [ + 132, + 440, + 481, + 472 + ], + "type": "text", + "content": "26. Lan, L., Yang, Y., Kaufman, D., Yao, J., Li, M., Jiang, C.: Medial ipc: accelerated incremental potential contact with medial elastics. ACM Transactions on Graphics 40(4) (2021)" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 132, + 472, + 481, + 505 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 472, + 481, + 505 + ], + "spans": [ + { + "bbox": [ + 132, + 472, + 481, + 505 + ], + "type": "text", + "content": "27. Li, M., Ferguson, Z., Schneider, T., Langlois, T.R., Zorin, D., Panozzo, D., Jiang, C., Kaufman, D.M.: Incremental potential contact: intersection-and inversion-free, large-deformation dynamics. ACM Trans. Graph. 39(4), 49 (2020)" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 132, + 505, + 481, + 525 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 505, + 481, + 525 + ], + "spans": [ + { + "bbox": [ + 132, + 505, + 481, + 525 + ], + "type": "text", + "content": "28. Liao, S.: On the homotopy analysis method for nonlinear problems. Applied mathematics and computation 147(2), 499-513 (2004)" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 132, + 525, + 481, + 547 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 525, + 481, + 547 + ], + "spans": [ + { + "bbox": [ + 132, + 525, + 481, + 547 + ], + "type": "text", + "content": "29. Lin, X., Yang, Z., Zhang, X., Zhang, Q.: Continuation path learning for homotopy optimization (2023)" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 132, + 547, + 481, + 568 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 547, + 481, + 568 + ], + "spans": [ + { + "bbox": [ + 132, + 547, + 481, + 568 + ], + "type": "text", + "content": "30. Liu, C.K., Jain, S.: A quick tutorial on multibody dynamics. Online tutorial, June p. 7 (2012)" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 132, + 568, + 481, + 589 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 568, + 481, + 589 + ], + "spans": [ + { + "bbox": [ + 132, + 568, + 481, + 589 + ], + "type": "text", + "content": "31. Liu, X., Pathak, D., Kitani, K.M.: Herd: Continuous human-to-robot evolution for learning from human demonstration. arXiv preprint arXiv:2212.04359 (2022)" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 132, + 589, + 481, + 621 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 589, + 481, + 621 + ], + "spans": [ + { + "bbox": [ + 132, + 589, + 481, + 621 + ], + "type": "text", + "content": "32. Liu, Y., Yang, H., Si, X., Liu, L., Li, Z., Zhang, Y., Liu, Y., Yi, L.: Taco: Benchmarking generalizable bimanual tool-action-object understanding. arXiv preprint arXiv:2401.08399 (2024)" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 132, + 621, + 481, + 665 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 621, + 481, + 665 + ], + "spans": [ + { + "bbox": [ + 132, + 621, + 481, + 665 + ], + "type": "text", + "content": "33. Makoviychuk, V., Wawrzyniak, L., Guo, Y., Lu, M., Storey, K., Macklin, M., Hoeller, D., Rudin, N., Allshire, A., Handa, A., et al.: Isaac gym: High performancegpu-based physics simulation for robot learning. arXiv preprint arXiv:2108.10470 (2021)" + } + ] + } + ], + "index": 20 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "spans": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "type": "text", + "content": "16" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 166, + 91, + 217, + 100 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 166, + 91, + 217, + 100 + ], + "spans": [ + { + "bbox": [ + 166, + 91, + 217, + 100 + ], + "type": "text", + "content": "X. Liu et al." + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 15 + }, + { + "para_blocks": [ + { + "bbox": [ + 130, + 116, + 482, + 665 + ], + "type": "list", + "angle": 0, + "index": 20, + "blocks": [ + { + "bbox": [ + 130, + 116, + 482, + 149 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 116, + 482, + 149 + ], + "spans": [ + { + "bbox": [ + 130, + 116, + 482, + 149 + ], + "type": "text", + "content": "34. Mandikal, P., Grauman, K.: Dexvip: Learning dexterous grasping with human hand pose priors from video. In: Conference on Robot Learning. pp. 651-661. PMLR (2022)" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 130, + 150, + 481, + 183 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 150, + 481, + 183 + ], + "spans": [ + { + "bbox": [ + 130, + 150, + 481, + 183 + ], + "type": "text", + "content": "35. Marcucci, T., Gabiccini, M., Artoni, A.: A two-stage trajectory optimization strategy for articulated bodies with unscheduled contact sequences. IEEE Robotics and Automation Letters 2(1), 104-111 (2016)" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 130, + 183, + 481, + 216 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 183, + 481, + 216 + ], + "spans": [ + { + "bbox": [ + 130, + 183, + 481, + 216 + ], + "type": "text", + "content": "36. Mordatch, I., Popovic, Z., Todorov, E.: Contact-invariant optimization for hand manipulation. In: Proceedings of the ACM SIGGRAPH/Eurographics symposium on computer animation. pp. 137-144 (2012)" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 130, + 216, + 481, + 248 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 216, + 481, + 248 + ], + "spans": [ + { + "bbox": [ + 130, + 216, + 481, + 248 + ], + "type": "text", + "content": "37. Pang, T., Suh, H.T., Yang, L., Tedrake, R.: Global planning for contact-rich manipulation via local smoothing of quasi-dynamic contact models. IEEE Transactions on Robotics (2023)" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 130, + 249, + 481, + 282 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 249, + 481, + 282 + ], + "spans": [ + { + "bbox": [ + 130, + 249, + 481, + 282 + ], + "type": "text", + "content": "38. Pang, T., Tedrake, R.: A convex quasistatic time-stepping scheme for rigid multi-body systems with contact and friction. In: 2021 IEEE International Conference on Robotics and Automation (ICRA). pp. 6614-6620. IEEE (2021)" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 130, + 282, + 481, + 315 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 282, + 481, + 315 + ], + "spans": [ + { + "bbox": [ + 130, + 282, + 481, + 315 + ], + "type": "text", + "content": "39. Peng, X.B., Abbeel, P., Levine, S., Van de Panne, M.: Deepmimic: Example-guided deep reinforcement learning of physics-based character skills. ACM Transactions On Graphics (TOG) 37(4), 1-14 (2018)" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 130, + 315, + 481, + 335 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 315, + 481, + 335 + ], + "spans": [ + { + "bbox": [ + 130, + 315, + 481, + 335 + ], + "type": "text", + "content": "40. Pfaff, T., Fortunato, M., Sanchez-Gonzalez, A., Battaglia, P.W.: Learning mesh-based simulation with graph networks. arXiv preprint arXiv:2010.03409 (2020)" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 130, + 336, + 481, + 369 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 336, + 481, + 369 + ], + "spans": [ + { + "bbox": [ + 130, + 336, + 481, + 369 + ], + "type": "text", + "content": "41. Pfrommer, S., Halm, M., Posa, M.: Contactnets: Learning discontinuous contact dynamics with smooth, implicit representations. In: Conference on Robot Learning. pp. 2279-2291. PMLR (2021)" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 130, + 369, + 481, + 403 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 369, + 481, + 403 + ], + "spans": [ + { + "bbox": [ + 130, + 369, + 481, + 403 + ], + "type": "text", + "content": "42. Qin, Y., Su, H., Wang, X.: From one hand to multiple hands: Imitation learning for dexterous manipulation from single-camera teleoperation. IEEE Robotics and Automation Letters 7(4), 10873-10881 (2022)" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 130, + 403, + 481, + 435 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 403, + 481, + 435 + ], + "spans": [ + { + "bbox": [ + 130, + 403, + 481, + 435 + ], + "type": "text", + "content": "43. Qin, Y., Wu, Y.H., Liu, S., Jiang, H., Yang, R., Fu, Y., Wang, X.: Dexamv: Imitation learning for dexterous manipulation from human videos. In: European Conference on Computer Vision. pp. 570-587. Springer (2022)" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 130, + 435, + 481, + 468 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 435, + 481, + 468 + ], + "spans": [ + { + "bbox": [ + 130, + 435, + 481, + 468 + ], + "type": "text", + "content": "44. Qin, Y., Yang, W., Huang, B., Van Wyk, K., Su, H., Wang, X., Chao, Y.W., Fox, D.: Anyteleop: A general vision-based dexterous robot arm-hand teleoperation system. arXiv preprint arXiv:2307.04577 (2023)" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 130, + 468, + 481, + 501 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 468, + 481, + 501 + ], + "spans": [ + { + "bbox": [ + 130, + 468, + 481, + 501 + ], + "type": "text", + "content": "45. Radosavovic, I., Wang, X., Pinto, L., Malik, J.: State-only imitation learning for dexterous manipulation. In: 2021 IEEE/RSJ International Conference on Intelligent Robots and Systems (IROS). pp. 7865-7871. IEEE (2021)" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 130, + 501, + 481, + 534 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 501, + 481, + 534 + ], + "spans": [ + { + "bbox": [ + 130, + 501, + 481, + 534 + ], + "type": "text", + "content": "46. Rajeswaran, A., Kumar, V., Gupta, A., Vezzani, G., Schulman, J., Todorov, E., Levine, S.: Learning complex dexterous manipulation with deep reinforcement learning and demonstrations. arXiv preprint arXiv:1709.10087 (2017)" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 130, + 534, + 481, + 567 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 534, + 481, + 567 + ], + "spans": [ + { + "bbox": [ + 130, + 534, + 481, + 567 + ], + "type": "text", + "content": "47. Rusu, A.A., Colmenarejo, S.G., Gulcehre, C., Desjardins, G., Kirkpatrick, J., Pascanu, R., Mnih, V., Kavukcuoglu, K., Hadsell, R.: Policy distillation. arXiv preprint arXiv:1511.06295 (2015)" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 130, + 567, + 481, + 600 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 567, + 481, + 600 + ], + "spans": [ + { + "bbox": [ + 130, + 567, + 481, + 600 + ], + "type": "text", + "content": "48. Schmeckpeper, K., Rybkin, O., Daniilidis, K., Levine, S., Finn, C.: Reinforcement learning with videos: Combining offline observations with interaction. arXiv preprint arXiv:2011.06507 (2020)" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 130, + 600, + 481, + 620 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 600, + 481, + 620 + ], + "spans": [ + { + "bbox": [ + 130, + 600, + 481, + 620 + ], + "type": "text", + "content": "49. ShadowRobot: Shadowrobot dexterous hand (2005), https://www.shadowrobot.com/dexterous-hand-series/" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 130, + 621, + 481, + 643 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 621, + 481, + 643 + ], + "spans": [ + { + "bbox": [ + 130, + 621, + 481, + 643 + ], + "type": "text", + "content": "50. Siahkoohi, A., Louboutin, M., Herrmann, F.J.: Neural network augmented wave-equation simulation. arXiv preprint arXiv:1910.00925 (2019)" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 130, + 643, + 481, + 665 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 643, + 481, + 665 + ], + "spans": [ + { + "bbox": [ + 130, + 643, + 481, + 665 + ], + "type": "text", + "content": "51. Suh, H., Wang, Y.: Comparing effectiveness of relaxation methods for warm starting trajectory optimization through soft contact (2019)" + } + ] + } + ], + "index": 19 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 184, + 91, + 447, + 102 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 184, + 91, + 447, + 102 + ], + "spans": [ + { + "bbox": [ + 184, + 91, + 447, + 102 + ], + "type": "text", + "content": "Quasi-Physical Simulators for Dexterous Manipulations Transfer" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 470, + 92, + 481, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 470, + 92, + 481, + 100 + ], + "spans": [ + { + "bbox": [ + 470, + 92, + 481, + 100 + ], + "type": "text", + "content": "17" + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 16 + }, + { + "para_blocks": [ + { + "bbox": [ + 132, + 117, + 481, + 665 + ], + "type": "list", + "angle": 0, + "index": 20, + "blocks": [ + { + "bbox": [ + 133, + 117, + 480, + 138 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 117, + 480, + 138 + ], + "spans": [ + { + "bbox": [ + 133, + 117, + 480, + 138 + ], + "type": "text", + "content": "52. Suh, H.J.T., Pang, T., Tedrake, R.: Bundled gradients through contact via randomized smoothing. IEEE Robotics and Automation Letters 7(2), 4000-4007 (2022)" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 132, + 140, + 481, + 182 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 140, + 481, + 182 + ], + "spans": [ + { + "bbox": [ + 132, + 140, + 481, + 182 + ], + "type": "text", + "content": "53. Taheri, O., Ghorbani, N., Black, M.J., Tzionas, D.: Grab: A dataset of whole-body human grasping of objects. In: Computer Vision-ECCV 2020: 16th European Conference, Glasgow, UK, August 23-28, 2020, Proceedings, Part IV 16. pp. 581-600. Springer (2020)" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 132, + 182, + 480, + 202 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 182, + 480, + 202 + ], + "spans": [ + { + "bbox": [ + 132, + 182, + 480, + 202 + ], + "type": "text", + "content": "54. Tan, K.K., Wang, Q.G., Hang, C.C.: Advances in PID control. Springer Science & Business Media (2012)" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 132, + 203, + 480, + 224 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 203, + 480, + 224 + ], + "spans": [ + { + "bbox": [ + 132, + 203, + 480, + 224 + ], + "type": "text", + "content": "55. Tedrake, R., the Drake Development Team: Drake: Model-based design and verification for robotics (2019), https://drake.mit.edu" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 132, + 224, + 480, + 256 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 224, + 480, + 256 + ], + "spans": [ + { + "bbox": [ + 132, + 224, + 480, + 256 + ], + "type": "text", + "content": "56. Todorov, E., Erez, T., Tassa, Y.: Mujoco: A physics engine for model-based control. In: 2012 IEEE/RSJ international conference on intelligent robots and systems. pp. 5026-5033. IEEE (2012)" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 132, + 257, + 480, + 289 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 257, + 480, + 289 + ], + "spans": [ + { + "bbox": [ + 132, + 257, + 480, + 289 + ], + "type": "text", + "content": "57. Traore, R., Caseles-Dupré, H., Lesort, T., Sun, T., Díaz-Rodriguez, N., Filliat, D.: Continual reinforcement learning deployed in real-life using policy distillation and sim2real transfer. arXiv preprint arXiv:1906.04452 (2019)" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 132, + 289, + 480, + 320 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 289, + 480, + 320 + ], + "spans": [ + { + "bbox": [ + 132, + 289, + 480, + 320 + ], + "type": "text", + "content": "58. Villegas, R., Ceylan, D., Hertzmann, A., Yang, J., Saito, J.: Contact-aware retargeting of skinned motion. In: Proceedings of the IEEE/CVF International Conference on Computer Vision. pp. 9720-9729 (2021)" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 132, + 321, + 480, + 353 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 321, + 480, + 353 + ], + "spans": [ + { + "bbox": [ + 132, + 321, + 480, + 353 + ], + "type": "text", + "content": "59. Wang, Y., Weidner, N.J., Baxter, M.A., Hwang, Y., Kaufman, D.M., Sueda, S.: Redmax: Efficient & flexible approach for articulated dynamics. ACM Transactions on Graphics (TOG) 38(4), 1-10 (2019)" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 132, + 354, + 480, + 385 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 354, + 480, + 385 + ], + "spans": [ + { + "bbox": [ + 132, + 354, + 480, + 385 + ], + "type": "text", + "content": "60. Wang, Y., Lin, J., Zeng, A., Luo, Z., Zhang, J., Zhang, L.: Physhoi: Physics-based imitation of dynamic human-object interaction. arXiv preprint arXiv:2312.04393 (2023)" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 132, + 386, + 480, + 407 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 386, + 480, + 407 + ], + "spans": [ + { + "bbox": [ + 132, + 386, + 480, + 407 + ], + "type": "text", + "content": "61. Watson, L.T., Haftka, R.T.: Modern homotopy methods in optimization. Computer Methods in Applied Mechanics and Engineering 74(3), 289-305 (1989)" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 132, + 407, + 480, + 438 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 407, + 480, + 438 + ], + "spans": [ + { + "bbox": [ + 132, + 407, + 480, + 438 + ], + "type": "text", + "content": "62. Wu, Y.H., Wang, J., Wang, X.: Learning generalizable dexterous manipulation from human grasp affordance. In: Conference on Robot Learning. pp. 618-629. PMLR (2023)" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 132, + 439, + 480, + 471 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 439, + 480, + 471 + ], + "spans": [ + { + "bbox": [ + 132, + 439, + 480, + 471 + ], + "type": "text", + "content": "63. Xiong, S., He, X., Tong, Y., Deng, Y., Zhu, B.: Neural vortex method: from finite lagrangian particles to infinite dimensional eulerian dynamics. Computers & Fluids 258, 105811 (2023)" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 132, + 472, + 480, + 503 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 472, + 480, + 503 + ], + "spans": [ + { + "bbox": [ + 132, + 472, + 480, + 503 + ], + "type": "text", + "content": "64. Xu, J., Chen, T., Zlokapa, L., Foshey, M., Matusik, W., Sueda, S., Agrawal, P.: An end-to-end differentiable framework for contact-aware robot design. arXiv preprint arXiv:2107.07501 (2021)" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 132, + 504, + 480, + 536 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 504, + 480, + 536 + ], + "spans": [ + { + "bbox": [ + 132, + 504, + 480, + 536 + ], + "type": "text", + "content": "65. Yamane, K., Nakamura, Y.: Stable penalty-based model of frictional contacts. In: Proceedings 2006 IEEE International Conference on Robotics and Automation, 2006. ICRA 2006. pp. 1904-1909. IEEE (2006)" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 132, + 536, + 480, + 567 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 536, + 480, + 567 + ], + "spans": [ + { + "bbox": [ + 132, + 536, + 480, + 567 + ], + "type": "text", + "content": "66. Yao, H., Song, Z., Chen, B., Liu, L.: Controlvae: Model-based learning of generative controllers for physics-based characters. ACM Transactions on Graphics (TOG) 41(6), 1-16 (2022)" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 132, + 568, + 480, + 600 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 568, + 480, + 600 + ], + "spans": [ + { + "bbox": [ + 132, + 568, + 480, + 600 + ], + "type": "text", + "content": "67. Zeng, A., Song, S., Lee, J., Rodriguez, A., Funkhouser, T.: Tossingbot: Learning to throw arbitrary objects with residual physics. IEEE Transactions on Robotics 36(4), 1307-1319 (2020)" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 132, + 601, + 480, + 632 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 601, + 480, + 632 + ], + "spans": [ + { + "bbox": [ + 132, + 601, + 480, + 632 + ], + "type": "text", + "content": "68. Zhang, H., Christen, S., Fan, Z., Zheng, L., Hwangbo, J., Song, J., Hilliges, O.: Artigrasp: Physically plausible synthesis of bi-manual dexterous grasping and articulation. arXiv preprint arXiv:2309.03891 (2023)" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 132, + 633, + 480, + 665 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 633, + 480, + 665 + ], + "spans": [ + { + "bbox": [ + 132, + 633, + 480, + 665 + ], + "type": "text", + "content": "69. Zhang, S., Liu, B., Wang, Z., Zhao, T.: Model-based reparameterization policy gradient methods: Theory and practical algorithms. Advances in Neural Information Processing Systems 36 (2024)" + } + ] + } + ], + "index": 19 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "spans": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "type": "text", + "content": "18" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 166, + 91, + 217, + 101 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 166, + 91, + 217, + 101 + ], + "spans": [ + { + "bbox": [ + 166, + 91, + 217, + 101 + ], + "type": "text", + "content": "X. Liu et al." + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 17 + }, + { + "para_blocks": [ + { + "bbox": [ + 132, + 116, + 482, + 216 + ], + "type": "list", + "angle": 0, + "index": 5, + "blocks": [ + { + "bbox": [ + 132, + 116, + 482, + 150 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 116, + 482, + 150 + ], + "spans": [ + { + "bbox": [ + 132, + 116, + 482, + 150 + ], + "type": "text", + "content": "70. Zhang, Y., Clegg, A., Ha, S., Turk, G., Ye, Y.: Learning to transfer in-hand manipulations using a greedy shape curriculum. In: Computer Graphics Forum. vol. 42, pp. 25-36. Wiley Online Library (2023)" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 132, + 150, + 482, + 183 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 150, + 482, + 183 + ], + "spans": [ + { + "bbox": [ + 132, + 150, + 482, + 183 + ], + "type": "text", + "content": "71. Zhao, W., Queralta, J.P., Westerlund, T.: Sim-to-real transfer in deep reinforcement learning for robotics: a survey. In: 2020 IEEE symposium series on computational intelligence (SSCI). pp. 737-744. IEEE (2020)" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 132, + 183, + 482, + 216 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 183, + 482, + 216 + ], + "spans": [ + { + "bbox": [ + 132, + 183, + 482, + 216 + ], + "type": "text", + "content": "72. Zhuang, F., Qi, Z., Duan, K., Xi, D., Zhu, Y., Zhu, H., Xiong, H., He, Q.: A comprehensive survey on transfer learning. Proceedings of the IEEE 109(1), 43-76 (2020)" + } + ] + } + ], + "index": 4 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 184, + 91, + 447, + 102 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 184, + 91, + 447, + 102 + ], + "spans": [ + { + "bbox": [ + 184, + 91, + 447, + 102 + ], + "type": "text", + "content": "Quasi-Physical Simulators for Dexterous Manipulations Transfer" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 470, + 91, + 481, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 470, + 91, + 481, + 100 + ], + "spans": [ + { + "bbox": [ + 470, + 91, + 481, + 100 + ], + "type": "text", + "content": "19" + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 18 + } + ], + "_backend": "vlm", + "_version_name": "2.6.4" +} \ No newline at end of file diff --git a/2024/Parrot Captions Teach CLIP to Spot Text/b85a274a-d711-419a-aeba-49fd9715a806_content_list.json b/2024/Parrot Captions Teach CLIP to Spot Text/b85a274a-d711-419a-aeba-49fd9715a806_content_list.json new file mode 100644 index 0000000000000000000000000000000000000000..1341079dd5cd7fe630516bc40ba1d212538fdd64 --- /dev/null +++ b/2024/Parrot Captions Teach CLIP to Spot Text/b85a274a-d711-419a-aeba-49fd9715a806_content_list.json @@ -0,0 +1,1737 @@ +[ + { + "type": "text", + "text": "Parrot Captions Teach CLIP to Spot Text", + "text_level": 1, + "bbox": [ + 259, + 141, + 741, + 162 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Yiqi Lin $^{1,2*}$ , Conghui He $^{1*}$ , Alex Jinpeng Wang $^{2*}$ , Bin Wang $^{1*}$ , Weijia Li $^{3}$ , and Mike Zheng Shou $^{2}$", + "bbox": [ + 246, + 188, + 754, + 220 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "1 Shanghai Artificial Intelligence Laboratory", + "bbox": [ + 346, + 231, + 651, + 244 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "2 Show Lab, National University of Singapore", + "bbox": [ + 346, + 246, + 653, + 258 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "3 Sun Yat-Sen University", + "bbox": [ + 416, + 260, + 586, + 273 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Abstract. Despite CLIP [29] being the foundation model in numerous vision-language applications, CLIP suffers from a severe text spotting bias. Such bias causes CLIP models to 'Parrot' the visual text embedded within images while disregarding the authentic visual semantics. We uncover that in the most popular image-text dataset LAION-2B [31], the captions also densely parrot (spell) the text embedded in images. Our analysis shows that around $50\\%$ of images are embedded with visual text content and around $30\\%$ of captions words are concurrently embedded in the visual content. Based on such observation, we thoroughly inspect the different released versions of CLIP models and verify that the visual text is a dominant factor in measuring the LAION-style image-text similarity for these models. To examine whether these parrot captions shape the text spotting bias, we train a series of CLIP models with LAION subsets curated by different parrot-caption-oriented criteria. We show that training with parrot captions easily shapes such bias but harms the expected visual-language representation learning in CLIP models across various vision-language downstream tasks. This suggests that it is urgent to revisit either the design of CLIP-like models or the existing image-text dataset curation pipeline built on CLIP score filtering. Project page: https://tinyq17.github.io/CLIP-Parrot-Bias/", + "bbox": [ + 261, + 306, + 743, + 584 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Keywords: Image-Text Dataset $\\cdot$ Text Spotting Bias", + "bbox": [ + 261, + 597, + 625, + 611 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "1 Introduction", + "text_level": 1, + "bbox": [ + 215, + 635, + 375, + 651 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Recently, contrastive learning models [17, 29, 31] pre-trained with large-scale image-text pair data has led to numerous vision-language modeling task breakthroughs. Due to its efficiency and simplicity, the pioneering work CLIP [29] now serves as a foundation model in various applications [20, 26, 30, 49]. However, several works [4,13] have shown that CLIP models have perpetuating biases towards visual text [19,25], color [34,44], gender [41], etc. In this paper, we focus on probing the visual text bias in CLIP, i.e., the capacity of spotting text in images. Most of the previous cues [25, 29, 34] attribute the sources of biases to the noisy pre-training data. Therefore, we begin by taking a close look at the most popular dataset, LAION-2B [31].", + "bbox": [ + 212, + 665, + 787, + 816 + ], + "page_idx": 0 + }, + { + "type": "page_footnote", + "text": "* Equal contribution. ⌒ Corresponding author (Email: heconghui@pjlab.org.cn).", + "bbox": [ + 217, + 824, + 756, + 840 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/66c472be5cef6c6f9d3fbde23da1f2ee370ac6db3009142b22eee8358e6f11e4.jpg", + "image_caption": [ + "(b) Image Paired Captions", + "Fig. 1: In LAION-2B [31], image-text pairs with the Top-5% highest similarity score are most dominant by visual text! These samples have dense concurrent text appearing in captions and images (text form in pixels). We refer to their captions as Parrot Captions as they raise a question: Dose CLIP Simply Parroting Text in Images for Vision-Language Alignment? The concurrent text is spotted by the OCR model and highlighted with color in image-text pairs. (Best view in color)" + ], + "image_footnote": [], + "bbox": [ + 217, + 142, + 785, + 400 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Considering the massive scale of the image-text data, it is non-trivial to assess the bias simply with a rough estimation. To this end, we first do image clustering on the whole dataset and rank each cluster by CLIP scores to analyze the most preferred types of image-text pairs under CLIP score measurement. As shown in Fig. 1, we surprisingly observe that a decent number of samples with top CLIP scores have dense concurrent text appearing in the captions and the images in the form of pixels. These samples break the assumption that CLIP models leverage text supervision to align the visual and language concepts. We refer to these captions as Parrot Captions as they provide another shortcut to achieve the same goal by teaching CLIP to do text spotting even without perceiving the actual visual concepts. To understand the underlying impact, we analyze the parrot captions from three perspectives: dataset, widely used released models, and model training. The results lead to three key findings:", + "bbox": [ + 212, + 537, + 787, + 733 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Firstly, captions in LAION-2B have a significant bias towards describing visual text content embedded in the images. We provide thorough profiling using off-the-self text spotting models on the LAION-2B dataset and show that over $50\\%$ of the images are embedded with visual text content. Moreover, by examining the spotted text content and the paired caption in the subset that images embedded with text, we find that over $60\\%$ of the captions at least have one concurrent word and reach at least around $30\\%$ words overlap", + "bbox": [ + 212, + 734, + 787, + 840 + ], + "page_idx": 1 + }, + { + "type": "page_number", + "text": "2", + "bbox": [ + 217, + 114, + 228, + 126 + ], + "page_idx": 1 + }, + { + "type": "header", + "text": "Y. Lin et al.", + "bbox": [ + 271, + 114, + 354, + 126 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "between the caption and spotted text from images. This finding suggests that the basic assumption of image-text semantic alignment in CLIP does not fully stand its ground when training with LAION-style data.", + "bbox": [ + 212, + 146, + 784, + 191 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Secondly, released CLIP models have strong text spotting bias almost in every style of web images, resulting in CLIP-filtering datasets inherently biased towards visual text-dominant data. We investigate OpenAI released CLIP model's behaviors in the LAION-2B dataset by examining the difference between alignment scores before and after text removal. The results show that CLIP model predictions densely correlate the visual text embedded in images with their parrot captions. Next, we further study the preference of the text spotting capacity on text content in CLIP and OpenCLIP models. Note that CLIP is trained on WIT-400M, while OpenCLIP uses the LAION-2B dataset. Therefore, we use synthetic images embedded with specific rendered text to avoid overfitting in OpenCLIP models. Our analysis shows that OpenCLIP is more biased toward text spotting than CLIP. We believe that the parrot caption plays a lurking role in training these released CLIP models and is the source of text spotting capacity instead of emergence behavior [42].", + "bbox": [ + 212, + 191, + 787, + 402 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Thirdly, CLIP models can easily learn text spotting capacity from parrot captions while failing to connect the vision-language semantics, just like a text spotting parrot. We sample different LAION-2B subsets curated by text-oriented criteria, including the embedded text ratio, the concurrent word ratios, and the relative CLIP score from text removal to train CLIP models under the same setting. The results show that using parrot captions data, CLIP model can learn strong text spotting capacity but lose most of the zero-shot generalization ability on image-text downstream tasks. Moreover, we also observe similar behaviors on various downstream tasks such VQA, captioning, and retrieval on BLIP [21] models. Lastly, we argue that the existing data curation pipeline built on CLIP score and the contrastive fashion urgently needs to be re-examined by considering such hidden parrot captions.", + "bbox": [ + 212, + 402, + 787, + 584 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "2 Related Work", + "text_level": 1, + "bbox": [ + 214, + 604, + 387, + 619 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "2.1 Contrastive Vision-Language Pre-training", + "text_level": 1, + "bbox": [ + 214, + 635, + 604, + 650 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Modeling vision and language by aligning the embedding similarity between paired image-text data [17, 29, 31] has shown great potential for transferable to downstream vision-language tasks. The pre-training techniques mainly contain the vision encoder [10, 15] for image embedding encoding, text encoder [9] for text embedding modeling, and cross-modal contrastive learning [17, 22, 29, 47] for learning a joint embedding space of vision and language. The pioneering work CLIP [29] leverages 400 million noisy image-text pairs to learn transferable visual representation from text supervision and show impressive zero-shot performance for various vision-language tasks. Following CLIP, several vision-language models such as ALIGN [17], BASIC [27], and Open-CLIP [31] are proposed, and CLIP models have been replicated on various datasets including WIT [29], LAION [31], COYO [6], and DataComp [11]. We mainly profile", + "bbox": [ + 212, + 657, + 787, + 840 + ], + "page_idx": 2 + }, + { + "type": "header", + "text": "Parrot Captions Teach CLIP to Spot Text", + "bbox": [ + 447, + 114, + 730, + 128 + ], + "page_idx": 2 + }, + { + "type": "page_number", + "text": "3", + "bbox": [ + 774, + 116, + 784, + 126 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "the LAION-2B [31] dataset due to its large scale and wide usage [26, 30] and two versions of pre-trained models, CLIP and OpenCLIP. Note that the 2 billion image-text pairs in the LAION-2B dataset are filtered by OpenAI released CLIP models, making OpenCLIP connect to CLIP closely.", + "bbox": [ + 212, + 146, + 787, + 207 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "2.2 Studying of CLIP Behaviors", + "text_level": 1, + "bbox": [ + 214, + 229, + 496, + 244 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Despite the strong zero-shot and transferable performance of CLIP, the perpetuating biases [1, 3, 13, 19, 41, 46] in CLIP are still not well investigated due to its large-scale noisy training data. Much research [2, 25, 34, 37, 40, 43, 44] focuses on revealing or enhancing the downstream performance with discovered bias in CLIP. For example, colorful masks [44] or red circles [34] applied to images can improve the zero-shot performance on visual localization tasks. In studying visual text content bias, [13] shows the multimodal neurons of CLIP not only respond to visual content and the visual text embedded in the image. Another work [19] shows that image recognition in CLIP can be strongly dominated by the visual text embedded in the image. To disentangle such bias, [25] attempts to separate the text spotting representation in pre-trained CLIP by training representation projection. Meanwhile, LoGoPrompt [33] enhances the classification performance by utilizing the visual text content as auxiliary prompts as input. Also, CLIPPO [39] shows that directly aligning the image and synthetic images print with the captions can perform similarly to CLIP without a text encoder.", + "bbox": [ + 212, + 253, + 787, + 482 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "2.3 Data Curation with Text Removal", + "text_level": 1, + "bbox": [ + 214, + 503, + 545, + 517 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Due to the successful practice of data curation in LAION datasets [31, 32] on scaling up the image-text datasets, searching advanced selection strategy to improve the data quality from common crawl data pool gains a growing interest [11]. Recently, several works [7, 24, 28] suggest that introducing text-related filtering methods improves the pre-training dataset quality. In DiHT [28], the data curation steps include filtering out the image-text pairs with high OCR confidence and matching text ratio. Moreover, [7, 24] mainly focus on studying the importance of filtering out the text-dominate images utilizing OCR models to improve pre-training dataset quality. Maini et al. [24] also draw the observation that $40\\%$ of LAION's image text is highly correlated with the caption, but only performing a small pilot study on 500 samples with manual judgment. Differently, this paper makes the first attempt to systemically reveal the source of text spotting capacity in CLIP is the data bias and the consequences of such bias in existing widely used datasets and pre-trained models.", + "bbox": [ + 212, + 527, + 787, + 739 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "3 Terminology", + "text_level": 1, + "bbox": [ + 214, + 762, + 375, + 780 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "The data processing on images in the following sections mainly covers clustering, text spotting (OCR), and text inpainting. Firstly, we cluster all images based on feature similarity. For each image-text pair, we then use the pre-trained text", + "bbox": [ + 212, + 794, + 785, + 840 + ], + "page_idx": 3 + }, + { + "type": "page_number", + "text": "4", + "bbox": [ + 217, + 114, + 230, + 126 + ], + "page_idx": 3 + }, + { + "type": "header", + "text": "Y. Lin et al.", + "bbox": [ + 271, + 114, + 354, + 127 + ], + "page_idx": 3 + }, + { + "type": "image", + "img_path": "images/ec43366ccf9e360bbbeaf331d6a94f2609b1b3c7650bc29f046b94e643520bcc.jpg", + "image_caption": [ + "Text Removal" + ], + "image_footnote": [], + "bbox": [ + 289, + 143, + 372, + 208 + ], + "page_idx": 4 + }, + { + "type": "image", + "img_path": "images/c0206ad0f5983bda345a6031654922588a01b6e22fdca1f55da534e9e29fad10.jpg", + "image_caption": [ + "All-Emb" + ], + "image_footnote": [], + "bbox": [ + 374, + 143, + 457, + 208 + ], + "page_idx": 4 + }, + { + "type": "image", + "img_path": "images/22adb419c388a8258eee75c2678728ca559f30dec4bdfa1c39b9bda2173a3117.jpg", + "image_caption": [ + "Co-Emb" + ], + "image_footnote": [], + "bbox": [ + 459, + 143, + 542, + 208 + ], + "page_idx": 4 + }, + { + "type": "image", + "img_path": "images/0229e13685709ee8331372d45a1963ae8fec27b5590c2be688da4129a8e0944e.jpg", + "image_caption": [ + "Random" + ], + "image_footnote": [], + "bbox": [ + 544, + 143, + 627, + 208 + ], + "page_idx": 4 + }, + { + "type": "image", + "img_path": "images/7ecc161b44339da337e15ee9bd347b579cb0936d5353013cc1d2e3b02e25297a.jpg", + "image_caption": [ + "Syn-Emb." + ], + "image_footnote": [], + "bbox": [ + 630, + 143, + 712, + 208 + ], + "page_idx": 4 + }, + { + "type": "image", + "img_path": "images/d576b18a2f5bfaf7df2e147fd0db4c0e19f22a0dc0032f50c36fd2cb769e031d.jpg", + "image_caption": [ + "Caption" + ], + "image_footnote": [], + "bbox": [ + 290, + 228, + 372, + 291 + ], + "page_idx": 4 + }, + { + "type": "image", + "img_path": "images/f8c8a090e256890a569b602e432cd85ab2f52a2c877899f1c3f867bf434d292f.jpg", + "image_caption": [ + "All-Emb.", + "Text Mask", + "Fig. 2: Visualization of defined terminologies. Co-Emb. Text is marked in the caption." + ], + "image_footnote": [], + "bbox": [ + 374, + 228, + 457, + 292 + ], + "page_idx": 4 + }, + { + "type": "image", + "img_path": "images/a25504c95347df076b368e630762b307f8a83fcd1ebca7e8cd7975ba69227ef6.jpg", + "image_caption": [ + "Text Removal", + "Co-Emb.", + "Text Mask" + ], + "image_footnote": [], + "bbox": [ + 459, + 228, + 542, + 292 + ], + "page_idx": 4 + }, + { + "type": "image", + "img_path": "images/73bedcf722a50c5c75f344f0cb05da4444d8294f2faa31b525a9648847fe68e4.jpg", + "image_caption": [ + "Inpainted", + "Random", + "Sampled Mask" + ], + "image_footnote": [], + "bbox": [ + 544, + 228, + 627, + 292 + ], + "page_idx": 4 + }, + { + "type": "image", + "img_path": "images/c8d01481022bbad859c05e185888241888398dd5b3f84e528b9f563ccafb5dba.jpg", + "image_caption": [ + "Text Template", + "Syn-Emb.", + "Text Template" + ], + "image_footnote": [], + "bbox": [ + 638, + 228, + 712, + 292 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Algorithm 1 Pseudocode of Detecting Co-Emb. Text (Rate)", + "text_level": 1, + "bbox": [ + 217, + 354, + 658, + 369 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "caption: captions from LAION-2B dataset.", + "bbox": [ + 215, + 375, + 472, + 383 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "ocr_text: text spotted by OCR model", + "bbox": [ + 220, + 383, + 442, + 392 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "cap_words,OCR_words $=$ set(capTION.split()，set(ocr_text.split()))", + "bbox": [ + 217, + 393, + 617, + 402 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "co_emb_text = intersection(cap_words,OCR_words)", + "bbox": [ + 217, + 402, + 508, + 411 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "co_emb_text_rate = len(co_emb_text) / len(cap_words)", + "bbox": [ + 217, + 411, + 532, + 420 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "spotting model to detect and recognize the text print in image pixels. The mask images in Fig. 2 are the spotted text area. Next, we match the spotted text with the caption using Algorithm 1 to obtain the concurrent words and their ratio in captions. Lastly, we use inpainting to remove the text from the image for CLIPs' pattern ablation. To avoid confusion, we define these concepts as follows,", + "bbox": [ + 212, + 455, + 785, + 531 + ], + "page_idx": 4 + }, + { + "type": "list", + "sub_type": "text", + "list_items": [ + "- Embedded Text: text spotted by OCR models from the images. To study the correlation of embedded text with captions, we define different kinds of embedded text as, 1) All-Emb. Text: all the text is spotted from an image; 2) Co-Emb. Text: spotted text concurrently appears in the image's corresponding captions; 3) Syn-Emb. Text: synthetic text rendered in an image with a fixed font and a blank background. Fig. 2 shows examples of spotted embedded text by binary mask and the rendering synthetic text.", + "- Co-Emb. Text Rate (CoTR): the word set IoU of Co-Emb. text and captions (See Algorithm 1).", + "- Image w/ or w/o Embedded Text: spotted text results of a given image are none-empty or empty.", + "- Text Removal Image: do inpainting in the specific spotted text area (All-Emb., Co-Emb., or Random). Random is implemented by sampling other image's text areas. For the different inpainting results, see Fig. 2.", + "- Relative Scores (RSA/RSC): the difference in CLIP score between images modified by different inpainting operations while keeping the same captions. RSA and RSC are the short for the relative scores before and after removing All-Emb. text and Co-Emb. text.", + "- Image Clusters: image partitions based on K-Means." + ], + "bbox": [ + 223, + 547, + 787, + 839 + ], + "page_idx": 4 + }, + { + "type": "header", + "text": "Parrot Captions Teach CLIP to Spot Text", + "bbox": [ + 447, + 114, + 730, + 128 + ], + "page_idx": 4 + }, + { + "type": "page_number", + "text": "5", + "bbox": [ + 774, + 116, + 784, + 126 + ], + "page_idx": 4 + }, + { + "type": "table", + "img_path": "images/f332ac68efc00bd8d5adaf404d4c4f4e8889d2cf2a81fe6b25ada343c5f83bbd.jpg", + "table_caption": [ + "Table 1: Overall parrot captions statistic. More than $50\\%$ of images are embedded with text, and $30\\%$ of caption words are printed in images!" + ], + "table_footnote": [], + "table_body": "
Number of Total Images1,985,284,122
Number of Images w/ Emb. Text1,083,896,427
Number of Images w/ Co-Emb. Text663,600,432
Co-Emb. Text Rate (in Total)15.42%
Co-Emb. Text Rate (in Images w/ Emb. Text)28.24%
Fuzzy Co-Emb. Text Rate (in Total)30.46%
Fuzzy Co-Emb. Text Rate (in Images w/ Emb. Text)55.79%
", + "bbox": [ + 328, + 185, + 671, + 267 + ], + "page_idx": 5 + }, + { + "type": "list", + "sub_type": "text", + "list_items": [ + "- CLIP and OpenCLIP: CLIP models are trained on WIT-400M [29] and LAION-2B [31] dataset.", + "- N-gram Vocabulary (Vocab): the set of all contiguous N word sequences extracted from a text corpus, such as the collection of all captions." + ], + "bbox": [ + 223, + 290, + 784, + 351 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "4 Profiling LAION-2B Data", + "text_level": 1, + "bbox": [ + 215, + 369, + 503, + 386 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "To better profile the image-text pair data on a billion scale, we first cluster all the images based on CLIP features into 4,000 clusters and sort each cluster with CLIP scores. After obtaining all the cluster labels, we use the SOTA text spotting model [45] to get the visual text content on all the collected images. Finally, we aggregate all the model-predicted results and compare them with their corresponding captions to bring out our observations.", + "bbox": [ + 212, + 398, + 784, + 488 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "4.1 Implementation Details", + "text_level": 1, + "bbox": [ + 215, + 507, + 455, + 522 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Clustering with CLIP Features: We first train K-Means (implemented by Faiss [18]) on the LAION-400M [32] subset using ViT-B-32 [10] CLIP features to speed up the clustering process. Due to the large memory consumption, we reduce the feature dimensions from 512 to 256 using PCA. Then, we partition the whole dataset using trained K-Means with the same feature extraction pipeline.", + "bbox": [ + 212, + 527, + 785, + 603 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Text Spotting and Matching: To detect and recognize text across various scenes, we adopt DeepSolo [45] as our text spotting model and use the pre-trained checkpoints with the ViTAEv2-S [48] backbone in default setting. The output format of the text spotting model is a sequence of polygons of text location and their recognized characters. Despite its strong performance, we empirically find that DeepSolo can not handle the crowd scenes well (with more than 100 separate words) which is only a small proportion of the dataset ( $\\sim 2\\%$ ). To identify the correlation between the spotted text and captions, we use Algorithm 1 to calculate the Co-Emb. text rate in each image-text pair. Considering the imperfect text spotting predictions might miss or misspell words, we also use Levenshtein distance to calculate the fuzzing similarity and reported in Tab. 1.", + "bbox": [ + 212, + 604, + 785, + 771 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "4.2 Statistic and Observations from LAION-2B", + "text_level": 1, + "bbox": [ + 215, + 787, + 619, + 801 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "The overall statistics of the 2 billion image-text pairs are reported in Tab. 1. In summary, the images embedded with visual text content reach a surprisingly", + "bbox": [ + 212, + 809, + 784, + 839 + ], + "page_idx": 5 + }, + { + "type": "page_number", + "text": "6", + "bbox": [ + 217, + 114, + 228, + 126 + ], + "page_idx": 5 + }, + { + "type": "header", + "text": "Y. Lin et al.", + "bbox": [ + 271, + 114, + 354, + 127 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/2736101e6b6790673d162463d80f3b3fdd36658ea8255251ff60763cdfce8256.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 256, + 143, + 743, + 261 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/fe2ca1d48b0669c1f8be8c553c27eb3b8e8658a10422d0f40e5ef845c3115eed.jpg", + "image_caption": [ + "(a) The ratio of different OCR-oriented data types in LAION-2B clusters.", + "(b) Top CLIP score samples visualization from 50 clusters with ratio over $80\\%$ .", + "Fig. 3: (a): Based on the OCR prediction results, the image-text pairs are divided into three types: image without visual embedded text content; the spotted text from the image has no concurrent text with the caption; the spotted text at least share one concurrent word with the caption. The clusters are merged from 4000 into 100 for a better view. (b): In the clusters with high ratio, the top CLIP score samples contain various text sources, such as posters, book covers, advertisements, and even slides." + ], + "image_footnote": [], + "bbox": [ + 259, + 273, + 743, + 351 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "high proportion of $54.60\\%$ in the investigated data. Around $15\\%$ of words in the dataset captions are Co-Emb. text, and the proportion of Co-Emb. text can further reach $30\\%$ when considering the fuzzy matching results of the spotted text and captions. This suggests that CLIP models trained on these data might lead to a high bias toward text spotting. To better visualize the data distribution, we provide cluster-specific statics results and top CLIP score samples of text-dominated clusters in Fig. 3. We divide all images into 100 clusters based on visual similarity and visualize them according to the OCR results. Every cluster contains more or less images embedded with text. Combined with sample visualization, we observe that in the LAION collected data, the parrot captions cover various scenes. In the subsets of images embedded with text, around $60\\%$ of captions at least precisely parrot one concurrent word (Co-Emb. Text Rate $>0$ ) appearing in the image. It suggests that the data collection pipeline of LAION [31] has a strong bias to introduce parrot captions from web data.", + "bbox": [ + 212, + 491, + 787, + 702 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "To better understand Co-Emb. Text, we provide a more thorough analysis of the word counting and text size of parrot captions. As shown in Fig. 4a, the results show that a large proportion of the Co-Emb. Text only takes a few words. However, we also find a large number of captions that are almost full parrot captions (see areas around the heatmap diagonal). Next, in Fig. 4b and Fig. 4c, we investigate the correlation between the size of concurrent words box in the image and CLIP score. The results show that the large text size does not usually lead to a higher score; meanwhile, the small text size can also dominate the score as the score can be significantly different after removing them. (Details", + "bbox": [ + 212, + 704, + 787, + 840 + ], + "page_idx": 6 + }, + { + "type": "header", + "text": "Parrot Captions Teach CLIP to Spot Text", + "bbox": [ + 447, + 114, + 730, + 128 + ], + "page_idx": 6 + }, + { + "type": "page_number", + "text": "7", + "bbox": [ + 774, + 116, + 784, + 126 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/d848e7a6b526e361e58fd20eacc31821b0c9601384979e11031caf73e76bad60.jpg", + "image_caption": [ + "Fig. 4: (a): The number of caption words and associated spotted concurrent words based on precise word matching. (b): Distribution of total box area of concurrent words in the image and its CLIP score. (c): Distribution of total box area of concurrent words and its relative CLIP score before and after removing them from the image. (d): Distribution of text size of the single concurrent word and other spotted word." + ], + "image_footnote": [], + "bbox": [ + 256, + 143, + 743, + 262 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "of text removal described in Sec. 5.1). One possible reason is the text content and input resolution may matter more for CLIP. Moreover, we discover that the larger text is more likely to be parroted in the captions, as shown in Fig. 4d.", + "bbox": [ + 212, + 369, + 787, + 417 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "5 Inspecting Pre-Trained CLIP Models", + "text_level": 1, + "bbox": [ + 214, + 439, + 612, + 458 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "It is important to note that the LAION-2B dataset collection pipeline uses CLIP score from OpenAI's model to filter out the image-text pair below 0.28. Therefore, we inspect these two released CLIP models [29,31] to answer better why LAION data contains such a high proportion of parrot captions. Specifically, OpenAI's CLIP model is trained on the WIT dataset (out-of-domain model), and OpenCLIP is trained on LAION-2B (in-domain model). We first study whether the embedded text is the key factor in CLIP filtering by ablating the embedded text using text inpainting. Moreover, we further investigate whether the text spotting capacity prefers specific text content by examining synthetic images with Syn-Emb. text.", + "bbox": [ + 212, + 472, + 787, + 625 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "5.1 Ablation of Embedded Text Removal", + "text_level": 1, + "bbox": [ + 214, + 647, + 570, + 662 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Text Removal via Inpainting: Given the OCR predicted results, we use the fast marching method [38] to inpaint the area of the spotted text polygons. We generate two versions of text removal results for each image with embedded text, i.e., All-Emb. text removal and Co-Emb. text removal, as the parrot caption prediction is imperfect due to the limitation of OCR models. We also generate random inpainting images with randomly sampled spotted text polygons from other images to ablate the distribution shift caused by image inpainting. Examples of the spotted text masks and inpainting results are shown in Fig. 2.", + "bbox": [ + 212, + 672, + 787, + 794 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Results: Based on the OCR predicted results and text inpainting operations, we can obtain six types of LAION images, including $\\bullet$ ): images without any embedded text (OCR results are empty); $\\bullet$ ): images with any embedded text", + "bbox": [ + 212, + 795, + 787, + 840 + ], + "page_idx": 7 + }, + { + "type": "page_number", + "text": "8", + "bbox": [ + 217, + 114, + 228, + 126 + ], + "page_idx": 7 + }, + { + "type": "header", + "text": "Y. Lin et al.", + "bbox": [ + 271, + 114, + 354, + 126 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/a29d6bf2527dbbbbe5d619be17b238ed6c0779b885764f5faf754d369489b5dc.jpg", + "image_caption": [ + "Fig. 5: Left: Mean CLIP scores of image-text pairs with different text removal operations depicted in Sec. 5.1, and grouped by cluster the same as Fig. 3. Right: Overall relative CLIP score distribution by comparing different text removal operations." + ], + "image_footnote": [], + "bbox": [ + 217, + 145, + 782, + 281 + ], + "page_idx": 8 + }, + { + "type": "table", + "img_path": "images/061577db60f59d0a41587b5e8aab0d1a1cb1f737956cdb8bb4ae5dd6c27838e6.jpg", + "table_caption": [ + "Table 2: Mean CLIP score of different setups of text removal." + ], + "table_footnote": [], + "table_body": "
SetupAverage CLIP Score
Raw Images w/o Emb. Text0.3223 ± 0.0078
Raw Images w/ Emb. Text0.3358 ± 0.0094
× All-Emb. Random Inpainted0.3260 ± 0.0057
× All-Emb. Text Removal0.2974 ± 0.0197
Co-Emb. Random Inpainted0.3341 ± 0.0051
Co-Emb. Text Removal0.2993 ± 0.0146
", + "bbox": [ + 362, + 376, + 643, + 463 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "(OCR results are none-empty); $\\times$ ): images with random inpainting by other image's All-Emb. text area; $\\times$ ): images removed All-Emb. text (Inpaint all the areas of OCR predicted text); ■): images randomly inpainted by other image's Co-Emb. text area, and ■): images removed Co-Emb. text (Inpaint the areas of concurrent text in OCR predicted text and captions). Then, we calculate CLIP scores of all the groups of images and their paired captions using OpenAI released CLIP model (ViT-B-32). Fig. 5 reports the mean scores of different types of images in each cluster and raises four observations as follows: I). The images embedded with text achieve higher CLIP scores in most clusters than those without embedded text; II). CLIP scores significantly drop once we remove the text from the images compared to its random inpainting baseline. It indicates that the parrot captions correlate highly with CLIP score measurement; III). Not all the samples are dominated by the embedded text, as some samples achieve higher scores after removing text, indicating the embedded text can also be a distractor; IV). Most of the relative CLIP scores ( $S(\\square) - S(\\times)$ ) between images removed Co-Emb. text and All-Emb. text are positive, as shown in the right of Fig. 5. The straightforward reason is the images lose more visual information due to the larger in-painting area. Another possible reason is the imperfect text spotting prediction or the corner cases in the matching algorithm leaking parts of the concurrent text in images.", + "bbox": [ + 212, + 492, + 787, + 794 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "Discussion: Due to the text removal, the image distribution may shift from the CLIP training set. Therefore, we provide two random inpainting baselines to examine the effect of distribution shift. In Tab. 2, we report the mean scores", + "bbox": [ + 212, + 794, + 787, + 839 + ], + "page_idx": 8 + }, + { + "type": "header", + "text": "Parrot Captions Teach CLIP to Spot Text", + "bbox": [ + 447, + 114, + 730, + 128 + ], + "page_idx": 8 + }, + { + "type": "page_number", + "text": "9", + "bbox": [ + 774, + 116, + 784, + 126 + ], + "page_idx": 8 + }, + { + "type": "image", + "img_path": "images/b2bf5fc65d8a33143afa94208fd31f2580b7e440c2201ac5e439a8d6211cc792.jpg", + "image_caption": [ + "(a) Statistic of 1-gram vocabularies" + ], + "image_footnote": [], + "bbox": [ + 225, + 143, + 354, + 266 + ], + "page_idx": 9 + }, + { + "type": "image", + "img_path": "images/c51adc3c2e4cc81578b7a958b2f93d739506d8eb4c6238d7eba7dcd6e3abab14.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 354, + 143, + 485, + 266 + ], + "page_idx": 9 + }, + { + "type": "image", + "img_path": "images/eb6efd676f718ce07bca9268be2594cf7fe0ba293c6127178a6a55beab9d8a57.jpg", + "image_caption": [ + "(c) Grouped by Caption 1-gram length" + ], + "image_footnote": [], + "bbox": [ + 225, + 280, + 500, + 396 + ], + "page_idx": 9 + }, + { + "type": "image", + "img_path": "images/3bd8ffa9f9641b7a2f3363fedd8e97c6c8b90db62725fb5f5f55c10dba180ed4.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 500, + 148, + 779, + 267 + ], + "page_idx": 9 + }, + { + "type": "image", + "img_path": "images/6b01f1eb543c6ae006cbca61d13aee87e3471d2c1b3c78d85340378d0c7bc211.jpg", + "image_caption": [ + "(b) Grouped by 1-gram frequency intervals", + "(d) Grouped by Caption 2-gram length", + "Fig.6: OpenCLIP more bias towards text spotting than CLIP model. Grouped score distributions of prompting CLIP and OpenCLIP models with N-gram Syn-Emb. text and synthetic images for model preference investigation." + ], + "image_footnote": [], + "bbox": [ + 500, + 279, + 777, + 396 + ], + "page_idx": 9 + }, + { + "type": "text", + "text": "of different setups. Results show that the random baselines are very close to the raw image baseline, indicating that CLIP model is robust to the distribution shift caused by information loss in inpainted regions.", + "bbox": [ + 212, + 492, + 787, + 539 + ], + "page_idx": 9 + }, + { + "type": "text", + "text": "5.2 Prompting with Syn-Emb. Text", + "text_level": 1, + "bbox": [ + 214, + 559, + 524, + 574 + ], + "page_idx": 9 + }, + { + "type": "text", + "text": "Generating Synthetic Images from N-gram Vocabulary: To investigate CLIP models' text spotting preference, we adopt a similar strategy in [25] to use synthetic images to embed specific text content by rendering text in a blank background. For each text, we use four fore-background style rendering templates (black-white, black-grey, white-grey, and white-black), as shown in Fig. 2. Different from the uniformly sampling letters in [25], we generate the text content from the N-gram vocabulary built from captions and Co-Emb. text to study the text spotting pattern. We only select the top frequent 400,000 grams for each vocabulary. The statistics of 1-gram vocabulary are shown in Fig. 6a, which is a long-tail distribution. Next, we calculate the synthetic images and the rendered text similarity on released ViT-B-32 CLIP and OpenCLIP models.", + "bbox": [ + 212, + 583, + 787, + 748 + ], + "page_idx": 9 + }, + { + "type": "text", + "text": "Results: Firstly, we examine whether CLIP models prefer recognizing more commonly seen words (with high frequency in vocabulary). Therefore, we group the 1-gram results based on their frequency interval in the whole vocabulary, as shown in Fig. 6b. The results show that OpenCLIP model clearly has a stronger text spotting capacity than CLIP, i.e., more biased towards text spotting. We also observe that all CLIP models are more sensitive to the vocabulary built", + "bbox": [ + 212, + 750, + 787, + 839 + ], + "page_idx": 9 + }, + { + "type": "page_number", + "text": "10", + "bbox": [ + 217, + 114, + 235, + 126 + ], + "page_idx": 9 + }, + { + "type": "header", + "text": "Y. Lin et al.", + "bbox": [ + 271, + 114, + 354, + 126 + ], + "page_idx": 9 + }, + { + "type": "text", + "text": "from the concurrent words. Interestingly, both CLIP and OpenCLIP models have slightly higher scores on the less frequent grams. Secondly, considering the long-tail grams might contain more characters, we further group the 1-gram and 2-gram results based on their text length in Fig. 6c and Fig. 6d. Note that the Co-Emb. text is not regularly arranged in the images, making it hard to extract continuous word sequences. Results show that all the models are better at spotting the longer words, possibly due to the tokenizer used in the text encoder, making them more discriminative. Meanwhile, in the groups of 2-gram samples, the scores gradually drop when spotting the highly long text, indicating that the spotting capacity of CLIP models is possibly built on word-by-word.", + "bbox": [ + 212, + 146, + 787, + 297 + ], + "page_idx": 10 + }, + { + "type": "text", + "text": "6 Training on Emb. Text Curated Data", + "text_level": 1, + "bbox": [ + 214, + 316, + 612, + 333 + ], + "page_idx": 10 + }, + { + "type": "text", + "text": "Next, we dive deeper into training CLIP and BLIP [21] models on different Emb. text curated subsets and studying various downstream task behaviors.", + "bbox": [ + 212, + 345, + 782, + 375 + ], + "page_idx": 10 + }, + { + "type": "text", + "text": "Implementation Details: For CLIP, we use the open-source software OpenCLIP [16] for all CLIP model training. Our experiments are conducted on both ViT-B [10] and RN50 [15]. We use 4,096 batch size for 3M and 8,192 for 12M scale subsets. Other settings remain the same as [31]. For BLIP, we mainly conduct on 3M scale subsets with ViT-B [10]. The BLIP models are pre-trained for 10 epochs with an AdamW [23] optimizer. For downstream tasks, we finetune 10 epochs for VQA and 5 epochs for captioning and retrieval.", + "bbox": [ + 212, + 376, + 787, + 479 + ], + "page_idx": 10 + }, + { + "type": "text", + "text": "Evaluation: For CLIP, we follow the DataComp benchmark [11] using 38 zero-shot classification and retrieval tasks as evaluation. We report the average performance (Avg.) of the DataComp benchmark and two subset track performances, ImageNet (IN) and Retrieval (Ret.). To evaluate the text spotting capacity, we use the synthetic benchmark illustrated in Sec. 5.2 and a real-world benchmark sampled from LAION-2B as the validation set. In the synthetic benchmark, we calculate the similarity of all the 1-gram synthetic image-text pairs from caption vocabulary and report all the trained model results in Fig. 7. For the real-world benchmark, we sample 1M image-text pairs without any embedded text and 1M samples dominated by the parrot caption (RSC $\\geq 0.2$ ). Fig. 8 aggregates the mean scores of the 2M evaluation set and also reports the mean scores of applying text removal on the 2M evaluation set results. For BLIP, inspired by [12], we further evaluate the model behavior on downstream tasks requiring reading text, including Visual Question Answering (VQA), Image Captioning, and Text-Image Retrieval. Specifically, for the text-oriented tasks, we use Text VQA [36] and ST-VQA [5] for VQA, and TextCaps [35] for captioning and retrieval. Moreover, we also provide the same tasks on the datasets that only require the model to see, i.e., the natural image dataset. Similarly, we use VQAv2 [14] for VQA and COCO [8] for captioning and retrieval.", + "bbox": [ + 212, + 482, + 787, + 768 + ], + "page_idx": 10 + }, + { + "type": "text", + "text": "6.1 Ablation Study of Data Curation on CLIP", + "text_level": 1, + "bbox": [ + 214, + 787, + 612, + 801 + ], + "page_idx": 10 + }, + { + "type": "text", + "text": "Curation I: Embedded Text in Images. To study the impact of embedded text on overall pre-train data quality, we sample three subsets: random baseline,", + "bbox": [ + 212, + 809, + 785, + 839 + ], + "page_idx": 10 + }, + { + "type": "header", + "text": "Parrot Captions Teach CLIP to Spot Text", + "bbox": [ + 447, + 114, + 730, + 128 + ], + "page_idx": 10 + }, + { + "type": "page_number", + "text": "11", + "bbox": [ + 767, + 116, + 782, + 126 + ], + "page_idx": 10 + }, + { + "type": "table", + "img_path": "images/4f003cbb7dc6fbf71a9d6596437895b94cfb1f7052f74b0b57b8fc27088416fa.jpg", + "table_caption": [ + "Table 3: Ablation of images embedded with or without text." + ], + "table_footnote": [], + "table_body": "
DataModelINRet.Avg.
3M RandomRN500.2040.2220.256
3M w/o Emb. TextRN500.2280.2320.282
3M w/ Emb. Text OnlyRN500.0710.1390.164
3M RandomViT-B0.1310.1480.210
3M w/o Emb. TextViT-B0.1620.1640.234
3M w/ Emb. Text OnlyViT-B0.0520.1110.153
12M RandomRN500.3600.3540.354
12M w/o Emb. TextRN500.4090.3610.372
12M w/ Emb. Text OnlyRN500.1290.1920.218
12M RandomViT-B0.3140.2990.351
12M w/o Emb. TextViT-B0.3700.3180.364
12M w/ Emb. Text OnlyViT-B0.1290.1720.225
", + "bbox": [ + 228, + 185, + 457, + 308 + ], + "page_idx": 11 + }, + { + "type": "table", + "img_path": "images/5474333052371dfaba6663fb90058a51f8cb5d2836591f51df47042e4c816e4d.jpg", + "table_caption": [ + "Table 5: Ablation of models trained on subsets sampled by different RSA." + ], + "table_footnote": [], + "table_body": "
Data (3M)ModelAvg.S(●)INRet.Avg.
RSA < 0.0RN500.3190.1810.2200.239
RSA ≥ 0.0RN500.3390.1260.1800.215
RSA ≥ 0.1RN500.3510.0410.1230.148
RSA ≥ 0.2RN500.3600.0170.0940.109
RSA ≥ 0.3RN500.3760.0090.0750.097
RSA < 0.0ViT-B0.3190.1230.1590.198
RSA ≥ 0.0ViT-B0.3390.0790.1290.185
RSA ≥ 0.1ViT-B0.3510.0310.1030.134
RSA ≥ 0.2ViT-B0.3600.0120.0800.103
RSA ≥ 0.3ViT-B0.3760.0060.0700.096
", + "bbox": [ + 236, + 351, + 450, + 457 + ], + "page_idx": 11 + }, + { + "type": "table", + "img_path": "images/dd92dbe4cc1891e4dddbf8e44c34a9bab7b976826668ab59761dfd3df12cb2f5.jpg", + "table_caption": [ + "Table 4: Ablation of different CoEmb. Text Rate (CoTR)." + ], + "table_footnote": [], + "table_body": "
Data (3M)ModelINRet.Avg.
CoTR = 0.0RN500.1930.2290.247
CoTR ≥ 0.3RN500.0310.1100.137
CoTR ≥ 0.5RN500.0210.0990.124
CoTR ≥ 0.8RN500.0120.0820.096
CoTR = 1.0RN500.0120.0740.102
CoTR = 0.0ViT-B0.1320.1640.206
CoTR ≥ 0.3ViT-B0.0290.0840.130
CoTR ≥ 0.5ViT-B0.0210.0820.119
CoTR ≥ 0.8ViT-B0.0120.0760.104
CoTR = 1.0ViT-B0.0130.0760.103
", + "bbox": [ + 563, + 189, + 753, + 304 + ], + "page_idx": 11 + }, + { + "type": "table", + "img_path": "images/c652c95b3895d074f0dbd2f46f7ca21b700af7f6ea40c82e261f9145b0b77110.jpg", + "table_caption": [ + "Table 6: Ablation of models trained on subsets sampled by different RSC." + ], + "table_footnote": [], + "table_body": "
Data (3M)ModelAvg.S(●)INRet.Avg.
RSC < 0.0RN500.3260.1250.1710.209
RSC ≥ 0.0RN500.3450.0620.1290.168
RSC ≥ 0.1RN500.3540.0140.0910.106
RSC ≥ 0.2RN500.3640.0080.0840.104
RSC ≥ 0.3RN500.3800.0050.0580.084
RSC < 0.0ViT-B0.3260.0790.1290.174
RSC ≥ 0.0ViT-B0.3450.0450.1190.149
RSC ≥ 0.1ViT-B0.3540.0180.0910.116
RSC ≥ 0.2ViT-B0.3640.0080.0760.106
RSC ≥ 0.3ViT-B0.3800.0040.0590.091
", + "bbox": [ + 547, + 351, + 759, + 457 + ], + "page_idx": 11 + }, + { + "type": "text", + "text": "images without any embedded text, and images all embedded with text from LAION-2B. The subsets include 3M and 12M scales. The results in Tab. 3 show that images embedded with text generally reduce the pre-training dataset quality as all performance tracks significantly decrease. Meanwhile, in Fig. 7, the model trained with the images embedded with text achieves the strongest text spotting capacity compared to the random and images without text baselines.", + "bbox": [ + 212, + 489, + 784, + 580 + ], + "page_idx": 11 + }, + { + "type": "text", + "text": "Curation II: Co-Emb. Text Rate (CoTR). Tab. 3 reports CLIP models trained on parrot captions with different CoTR. We first select all the images with embedded text and then sample images based on the CoTR depicted at Algorithm 1 with different thresholds. With increasing CoTR, all the zero-shot benchmark performance drops significantly. Despite the images in the subset $(\\mathrm{CoTR} = 0)$ all embedded with text, the pre-trained model performs similarly to the random baseline in Tab. 3. It indicates that the parrot caption is more crucial than embedded text in reducing the pre-trained data quality. For the text spotting capacity, Fig. 7 and Fig. 8 show that the increasing CoTR does not lead to stronger text spotting capacity, possibly due to the average length of captions decreasing in higher CoTR data.", + "bbox": [ + 212, + 580, + 785, + 747 + ], + "page_idx": 11 + }, + { + "type": "text", + "text": "Curation III: Relative Score from Text Removal (RSA & RSC). Given the observations in Sec. 5.1, we further select a series of subsets based on the relative score of images before and after text removal. The higher relative scores mean the masked text is more dominant in CLIP score measurement. In Tab. 5 and Tab. 6, we report the zero-shot performance of models trained on subsets with different relative score thresholds. CLIP models pre-trained with higher", + "bbox": [ + 212, + 748, + 785, + 840 + ], + "page_idx": 11 + }, + { + "type": "page_number", + "text": "12", + "bbox": [ + 217, + 114, + 235, + 126 + ], + "page_idx": 11 + }, + { + "type": "header", + "text": "Y. Lin et al.", + "bbox": [ + 271, + 114, + 354, + 126 + ], + "page_idx": 11 + }, + { + "type": "image", + "img_path": "images/64665d5bd07426a917b953a91bf6c10ea5b1dc10604bd1eb648f2e6076f8ea24.jpg", + "image_caption": [ + "Fig. 7: CLIP models learn text spotting well from parrot captions. Benchmarking text spotting capacity of CLIP models with 1-gram caption vocabulary synthetic images dataset as the same as Sec. 5.2." + ], + "image_footnote": [], + "bbox": [ + 243, + 143, + 754, + 258 + ], + "page_idx": 12 + }, + { + "type": "image", + "img_path": "images/a21f85ab1ad31fb86cc26962da01e2a2fbf9ca466049c21a88c2a0e29ac1974f.jpg", + "image_caption": [ + "Fig. 8: Text spotting capacity validation on real images. Models trained with more parrot captions are better at aligning the image with parrot captions but perform worse at aligning images without embedded text." + ], + "image_footnote": [], + "bbox": [ + 341, + 325, + 661, + 459 + ], + "page_idx": 12 + }, + { + "type": "text", + "text": "RSA or RSC both get worse downstream performance. Notably, the average raw CLIP scores $\\mathrm{S}(\\bullet)$ of these subsets have a positive correlation with RSA or RSC, indicating using CLIP scores from a biased pre-trained model as the data filtering strategy can be unreliable. When comparing the RSA and RSC, the results show that the samples dominated by the latter, i.e., parrot captions, are less informative for CLIP training. Meanwhile, Fig. 7 and Fig. 8 show that the text spotting capacity of CLIP can be further improved by training on the samples using relative scores as data curation criteria against CoTR.", + "bbox": [ + 212, + 542, + 787, + 662 + ], + "page_idx": 12 + }, + { + "type": "text", + "text": "6.2 More Investigation on Text-Oriented Tasks", + "text_level": 1, + "bbox": [ + 214, + 689, + 617, + 704 + ], + "page_idx": 12 + }, + { + "type": "text", + "text": "Inspired by [12], we further train BLIP [21] models on the curated subsets to better understand the impact of parrot caption on various vision-language tasks, especially the tasks that require models to read the text. We chose BLIP for the ablation study instead of CLIP as it can be directly applied to all these tasks. As shown in Tab. 7, training BLIPs to spot text can boost their performance on the downstream tasks requiring the model to read but impede the performance of downstream tasks only requiring the model to see, which are consistent with the observation on classification tasks. Nevertheless, when BLIPs mainly focus", + "bbox": [ + 212, + 719, + 787, + 839 + ], + "page_idx": 12 + }, + { + "type": "header", + "text": "Parrot Captions Teach CLIP to Spot Text", + "bbox": [ + 447, + 114, + 730, + 128 + ], + "page_idx": 12 + }, + { + "type": "page_number", + "text": "13", + "bbox": [ + 767, + 114, + 784, + 126 + ], + "page_idx": 12 + }, + { + "type": "table", + "img_path": "images/ca4a6e51c0d7557d7987b0094c3886ffe647f5e9516e01ab77120afec38ed653.jpg", + "table_caption": [ + "Table 7: BLIP downstream tasks performance of pre-training on different curated 3M subsets. The gray color represents tasks requiring the model to read the text from images, i.e., spotting text from images." + ], + "table_footnote": [], + "table_body": "
BLIP Data (3M)Visual Question Answering (Acc)Image Captioning (CIDEr)Text-to-Image Retrieval (R@1)Image-to-Text Retrieval (R@1)
VQAv2TextVQAST-VQACOCOTextCapsCOCOTextCapsCOCOTextCaps
Rand71.0715.3610.48115.653.748.9156.3465.4672.45
w/ Emb. Text68.9419.0512.65108.992.142.8970.158.581.42
w/o Emb. Text71.2213.659.29116.241.549.9631.8366.548.7
CoTR = 0.071.1113.979.75116.344.649.5538.0566.0854.57
CoTR ≥ 0.367.419.2811.81104.996.937.7867.2851.9878.2
CoTR ≥ 0.567.0219.6412.38102.794.135.9465.2450.3276.94
CoTR ≥ 0.866.3818.5012.00100.991.634.1362.6546.973.56
CoTR = 1.066.1818.4712.80101.291.333.5561.8346.6273.05
RSA < 0.070.7914.169.64115.744.948.2536.8564.7254.7
RSA ≥ 0.070.0318.7611.81111.984.546.2568.6162.9281.23
RSA ≥ 0.168.1419.4813.33105.696.139.9668.1354.6479.37
RSA ≥ 0.266.0121.0611.8598.794.433.0364.1747.1275.33
RSA ≥ 0.364.2018.4412.0495.2691.126.6460.1137.370.24
RSC < 0.070.1315.1910.74112.246.746.841.9563.2458.05
RSC ≥ 0.068.8620.1213.75107.893.542.069.7857.4280.92
RSC ≥ 0.167.3520.5412.84103.496.936.466.6951.0277.79
RSC ≥ 0.262.6220.3213.1498.792.830.0861.3842.9671.98
RSC ≥ 0.363.7518.9413.0392.988.724.2358.3534.7268.95
", + "bbox": [ + 248, + 199, + 756, + 412 + ], + "page_idx": 13 + }, + { + "type": "text", + "text": "on reading, e.g. (RSA $\\geq 0.3$ ), their text-oriented and natural downstream performance also decreases. In other words, the parrot captions can benefit the text-orient downstream tasks while requiring careful data mixing trade-off. We believe that understanding parrot captions is essential for revealing the underlying mechanisms of existing large vision-language systems.", + "bbox": [ + 215, + 438, + 784, + 513 + ], + "page_idx": 13 + }, + { + "type": "text", + "text": "7 Conclusion and Discussion", + "text_level": 1, + "bbox": [ + 217, + 534, + 506, + 549 + ], + "page_idx": 13 + }, + { + "type": "text", + "text": "The popularity of vision-language contrastive loss stems from its efficiency and simplicity. However, the analysis and experiments we presented show that the embedded text in images and their parrot captions plant significant text spotting bias due to such contrastive fashions. Firstly, almost $30\\%$ of the captions in the widely used LAION-2B dataset are biased towards parroting the embedded text in images. Secondly, the pre-trained CLIP models have strong preferences for the image-text pair with parrot captions, which achieve higher similarity scores than those without. Finally, using data biasing to parrot captions, we can easily train a CLIP model with a strong text spotting bias. Our work demonstrates the emergency of reviewing the impact of parrot captions in the entire ecosystem of CLIP models. Our future endeavor involves building a bias-aware data curation pipeline and a robust training function to mitigate such issues.", + "bbox": [ + 215, + 563, + 785, + 744 + ], + "page_idx": 13 + }, + { + "type": "text", + "text": "Acknowledgements", + "text_level": 1, + "bbox": [ + 217, + 765, + 401, + 782 + ], + "page_idx": 13 + }, + { + "type": "text", + "text": "This project was supported by the National Key R&D Program of China (No. 2022ZD0160101) and the Shanghai Artificial Intelligence Laboratory. Mike Shou does not receive any funding for this work.", + "bbox": [ + 215, + 794, + 784, + 839 + ], + "page_idx": 13 + }, + { + "type": "page_number", + "text": "14", + "bbox": [ + 217, + 114, + 235, + 126 + ], + "page_idx": 13 + }, + { + "type": "header", + "text": "Y. Lin et al.", + "bbox": [ + 271, + 114, + 354, + 126 + ], + "page_idx": 13 + }, + { + "type": "text", + "text": "References", + "text_level": 1, + "bbox": [ + 217, + 143, + 321, + 159 + ], + "page_idx": 14 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "1. Agarwal, S., Krueger, G., Clark, J., Radford, A., Kim, J.W., Brundage, M.: Evaluating clip: towards characterization of broader capabilities and downstream implications. arXiv preprint arXiv:2108.02818 (2021)", + "2. Alabdulmohsin, I., Wang, X., Steiner, A.P., Goyal, P., D'Amour, A., Zhai, X.: Clip the bias: How useful is balancing data in multimodal learning? In: ICLR (2024)", + "3. Ali, J., Kleindessner, M., Wenzel, F., Budhathoki, K., Cevher, V., Russell, C.: Evaluating the fairness of discriminative foundation models in computer vision. In: AIES. pp. 809-833 (2023)", + "4. Berg, H., Hall, S.M., Bhalgat, Y., Yang, W., Kirk, H.R., Shtedritski, A., Bain, M.: A prompt array keeps the bias away: Debiasing vision-language models with adversarial learning. arXiv preprint arXiv:2203.11933 (2022)", + "5. Biten, A.F., Tito, R., Mafla, A., Gomez, L., Rusinol, M., Valveny, E., Jawahar, C., Karatzas, D.: Scene text visual question answering. In: ICCV. pp. 4291-4301 (2019)", + "6. Byeon, M., Park, B., Kim, H., Lee, S., Baek, W., Kim, S.: Coyo-700m: Image-text pair dataset. https://github.com/kakaobrain/coyo-dataset (2022)", + "7. Cao, L., Zhang, B., Chen, C., Yang, Y., Du, X., Zhang, W., Lu, Z., Zheng, Y.: Less is more: Removing text-regions improves clip training efficiency and robustness. arXiv preprint arXiv:2305.05095 (2023)", + "8. Chen, X., Fang, H., Lin, T.Y., Vedantam, R., Gupta, S., Dolkar, P., Zitnick, C.L.: Microsoft coco captions: Data collection and evaluation server. arXiv preprint arXiv:1504.00325 (2015)", + "9. Devlin, J., Chang, M.W., Lee, K., Toutanova, K.: Bert: Pre-training of deep bidirectional transformers for language understanding. arXiv preprint arXiv:1810.04805 (2018)", + "10. Dosovitskiy, A., Beyer, L., Kolesnikov, A., Weissenborn, D., Zhai, X., Unterthiner, T., Dehghani, M., Minderer, M., Heigold, G., Gelly, S., et al.: An image is worth 16x16 words: Transformers for image recognition at scale. arXiv preprint arXiv:2010.11929 (2020)", + "1. Gadre, S.Y., Ilharco, G., Fang, A., Hayase, J., Smyrnis, G., Nguyen, T., Marten, R., Wortsman, M., Ghosh, D., Zhang, J., et al.: Datacomp: In search of the next generation of multimodal datasets. arXiv preprint arXiv:2304.14108 (2023)", + "2. Ganz, R., Nuriel, O., Aberdam, A., Kittenplon, Y., Mazor, S., Litman, R.: Towards models that can see and read. arXiv preprint arXiv:2301.07389 (2023)", + "3. Goh, G., Cammarata, N., Voss, C., Carter, S., Petrov, M., Schubert, L., Radford, A., Olah, C.: Multimodal neurons in artificial neural networks. Distill 6(3), e30 (2021)", + "4. Goyal, Y., Khot, T., Summers-Stay, D., Batra, D., Parikh, D.: Making the v in vqa matter: Elevating the role of image understanding in visual question answering. In: CVPR. pp. 6904-6913 (2017)", + "5. He, K., Zhang, X., Ren, S., Sun, J.: Deep residual learning for image recognition. In: CVPR. pp. 770-778 (2016)", + "6. Ilharco, G., Wortsman, M., Wightman, R., Gordon, C., Carlini, N., Taori, R., Dave, A., Shankar, V., Namkoong, H., Miller, J., Hajishirzi, H., Farhadi, A., Schmidt, L.: Openclip (Jul 2021), https://doi.org/10.5281/zenodo.5143773", + "7. Jia, C., Yang, Y., Xia, Y., Chen, Y.T., Parekh, Z., Pham, H., Le, Q., Sung, Y.H., Li, Z., Duerig, T.: Scaling up visual and vision-language representation learning with noisy text supervision. In: ICML. pp. 4904-4916. PMLR (2021)" + ], + "bbox": [ + 225, + 176, + 784, + 839 + ], + "page_idx": 14 + }, + { + "type": "header", + "text": "Parrot Captions Teach CLIP to Spot Text", + "bbox": [ + 447, + 114, + 730, + 128 + ], + "page_idx": 14 + }, + { + "type": "page_number", + "text": "15", + "bbox": [ + 767, + 116, + 784, + 126 + ], + "page_idx": 14 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "18. Johnson, J., Douze, M., Jégou, H.: Billion-scale similarity search with GPUs. IEEE Transactions on Big Data 7(3), 535-547 (2019)", + "19. Lemesle, Y., Sawayama, M., Valle-Perez, G., Adolphe, M., Sauzeon, H., Oudeyer, P.Y.: Language-biased image classification: evaluation based on semantic representations. arXiv preprint arXiv:2201.11014 (2022)", + "20. Li, B., Weinberger, K.Q., Belongie, S., Koltun, V., Ranftl, R.: Language-driven semantic segmentation. In: ICLR (2022)", + "21. Li, J., Li, D., Xiong, C., Hoi, S.: Blip: Bootstrapping language-image pre-training for unified vision-language understanding and generation. In: ICML. pp. 12888-12900. PMLR (2022)", + "22. Li, J., Selvaraju, R., Gotmare, A., Joty, S., Xiong, C., Hoi, S.C.H.: Align before fuse: Vision and language representation learning with momentum distillation. NeurIPS 34, 9694-9705 (2021)", + "23. Loshchilov, I., Hutter, F.: Decoupled weight decay regularization. arXiv preprint arXiv:1711.05101 (2017)", + "24. Maini, P., Goyal, S., Lipton, Z.C., Kolter, J.Z., Raghunathan, A.: T-mars: Improving visual representations by circumventing text feature learning. arXiv preprint arXiv:2307.03132 (2023)", + "25. Materzyńska, J., Torralba, A., Bau, D.: Disentangling visual and written concepts in clip. In: CVPR. pp. 16410-16419 (2022)", + "26. Nichol, A., Dhariwal, P., Ramesh, A., Shyam, P., Mishkin, P., McGrew, B., Sutskever, I., Chen, M.: Glide: Towards photorealistic image generation and editing with text-guided diffusion models. arXiv preprint arXiv:2112.10741 (2021)", + "27. Pham, H., Dai, Z., Ghiasi, G., Kawaguchi, K., Liu, H., Yu, A.W., Yu, J., Chen, Y.T., Luong, M.T., Wu, Y., et al.: Combined scaling for zero-shot transfer learning. Neurocomputing 555, 126658 (2023)", + "28. Radenovic, F., Dubey, A., Kadian, A., Mihaylov, T., Vandenhende, S., Patel, Y., Wen, Y., Ramanathan, V., Mahajan, D.: Filtering, distillation, and hard negatives for vision-language pre-training. In: CVPR. pp. 6967-6977 (2023)", + "29. Radford, A., Kim, J.W., Hallacy, C., Ramesh, A., Goh, G., Agarwal, S., Sastry, G., Askell, A., Mishkin, P., Clark, J., et al.: Learning transferable visual models from natural language supervision. In: ICML. pp. 8748-8763. PMLR (2021)", + "30. Rombach, R., Blattmann, A., Lorenz, D., Esser, P., Ommer, B.: High-resolution image synthesis with latent diffusion models. In: CVPR. pp. 10684-10695 (2022)", + "31. Schuhmann, C., Beaumont, R., Vencu, R., Gordon, C., Wightman, R., Cherti, M., Coombes, T., Katta, A., Mullis, C., Wortsman, M., et al.: Laion-5b: An open large-scale dataset for training next generation image-text models. NeurIPS 35, 25278-25294 (2022)", + "32. Schuhmann, C., Vencu, R., Beaumont, R., Kaczmarczyk, R., Mullis, C., Katta, A., Coombes, T., Jitsev, J., Komatsuzaki, A.: Laion-400m: Open dataset of clip-filtered 400 million image-text pairs. arXiv preprint arXiv:2111.02114 (2021)", + "33. Shi, C., Yang, S.: Logoprompt: Synthetic text images can be good visual prompts for vision-language models. In: ICCV. pp. 2932-2941 (2023)", + "34. Shtedritski, A., Rupprecht, C., Vedaldi, A.: What does clip know about a red circle? visual prompt engineering for vlms. arXiv preprint arXiv:2304.06712 (2023)", + "35. Sidorov, O., Hu, R., Rohrbach, M., Singh, A.: Textcaps: a dataset for image captioning with reading comprehension. In: ECCV. pp. 742-758. Springer (2020)", + "36. Singh, A., Natarajan, V., Shah, M., Jiang, Y., Chen, X., Batra, D., Parikh, D., Rohrbach, M.: Towards vqa models that can read. In: CVPR. pp. 8317-8326 (2019)", + "37. Tanjin, M.M., Singh, K.K., Kafle, K., Sinha, R., Cottrell, G.W.: Discovering and mitigating biases in clip-based image editing. In: WACV. pp. 2984-2993 (2024)" + ], + "bbox": [ + 215, + 146, + 784, + 839 + ], + "page_idx": 15 + }, + { + "type": "page_number", + "text": "16", + "bbox": [ + 217, + 114, + 235, + 126 + ], + "page_idx": 15 + }, + { + "type": "header", + "text": "Y. Lin et al.", + "bbox": [ + 271, + 114, + 354, + 126 + ], + "page_idx": 15 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "38. Telea, A.: An image inpainting technique based on the fast marching method. Journal of graphics tools 9(1), 23-34 (2004)", + "39. Tschannen, M., Mustafa, B., Houlsby, N.: Clippo: Image-and-language understanding from pixels only. In: CVPR. pp. 11006-11017 (2023)", + "40. Wang, H., Zhan, Y., Liu, L., Ding, L., Yu, J.: Balanced similarity with auxiliary prompts: Towards alleviating text-to-image retrieval bias for clip in zero-shot learning. arXiv preprint arXiv:2402.18400 (2024)", + "41. Wang, J., Liu, Y., Wang, X.E.: Are gender-neutral queries really gender-neutral? mitigating gender bias in image search. arXiv preprint arXiv:2109.05433 (2021)", + "42. Wei, J., Tay, Y., Bommasani, R., Raffel, C., Zoph, B., Borgeaud, S., Yogatama, D., Bosma, M., Zhou, D., Metzler, D., et al.: Emergent abilities of large language models. arXiv preprint arXiv:2206.07682 (2022)", + "43. Xu, Y., Xu, Z., Chai, W., Zhao, Z., Song, E., Wang, G.: Devil in the number: Towards robust multi-modality data filter. arXiv preprint arXiv:2309.13770 (2023)", + "44. Yao, Y., Zhang, A., Zhang, Z., Liu, Z., Chua, T.S., Sun, M.: Cpt: Colorful prompt tuning for pre-trained vision-language models. arXiv preprint arXiv:2109.11797 (2021)", + "45. Ye, M., Zhang, J., Zhao, S., Liu, J., Liu, T., Du, B., Tao, D.: Deepsolo: Let transformer decoder with explicit points solo for text spotting. In: CVPR. pp. 19348-19357 (2023)", + "46. Yuksekgonul, M., Bianchi, F., Kalluri, P., Jurafsky, D., Zou, J.: When and why vision-language models behave like bags-of-words, and what to do about it? In: ICLR (2022)", + "47. Zhai, X., Wang, X., Mustafa, B., Steiner, A., Keysers, D., Kolesnikov, A., Beyer, L.: Lit: Zero-shot transfer with locked-image text tuning. In: CVPR. pp. 18123-18133 (2022)", + "48. Zhang, Q., Xu, Y., Zhang, J., Tao, D.: Vitaev2: Vision transformer advanced by exploring inductive bias for image recognition and beyond. IJCV pp. 1-22 (2023)", + "49. Zhou, K., Yang, J., Loy, C.C., Liu, Z.: Learning to prompt for vision-language models. IJCV 130(9), 2337-2348 (2022)" + ], + "bbox": [ + 212, + 146, + 785, + 563 + ], + "page_idx": 16 + }, + { + "type": "header", + "text": "Parrot Captions Teach CLIP to Spot Text", + "bbox": [ + 447, + 114, + 730, + 128 + ], + "page_idx": 16 + }, + { + "type": "page_number", + "text": "17", + "bbox": [ + 767, + 114, + 785, + 126 + ], + "page_idx": 16 + } +] \ No newline at end of file diff --git a/2024/Parrot Captions Teach CLIP to Spot Text/b85a274a-d711-419a-aeba-49fd9715a806_model.json b/2024/Parrot Captions Teach CLIP to Spot Text/b85a274a-d711-419a-aeba-49fd9715a806_model.json new file mode 100644 index 0000000000000000000000000000000000000000..fd84d3472d26e5f24afabd5254cb9e656e967357 --- /dev/null +++ b/2024/Parrot Captions Teach CLIP to Spot Text/b85a274a-d711-419a-aeba-49fd9715a806_model.json @@ -0,0 +1,2610 @@ +[ + [ + { + "type": "title", + "bbox": [ + 0.26, + 0.142, + 0.743, + 0.163 + ], + "angle": 0, + "content": "Parrot Captions Teach CLIP to Spot Text" + }, + { + "type": "text", + "bbox": [ + 0.247, + 0.189, + 0.756, + 0.221 + ], + "angle": 0, + "content": "Yiqi Lin\\(^{1,2*}\\), Conghui He\\(^{1*}\\), Alex Jinpeng Wang\\(^{2*}\\), Bin Wang\\(^{1*}\\), Weijia Li\\(^{3}\\), and Mike Zheng Shou\\(^{2}\\)" + }, + { + "type": "text", + "bbox": [ + 0.348, + 0.232, + 0.652, + 0.246 + ], + "angle": 0, + "content": "1 Shanghai Artificial Intelligence Laboratory" + }, + { + "type": "text", + "bbox": [ + 0.348, + 0.247, + 0.655, + 0.26 + ], + "angle": 0, + "content": "2 Show Lab, National University of Singapore" + }, + { + "type": "text", + "bbox": [ + 0.418, + 0.261, + 0.588, + 0.274 + ], + "angle": 0, + "content": "3 Sun Yat-Sen University" + }, + { + "type": "text", + "bbox": [ + 0.262, + 0.308, + 0.744, + 0.585 + ], + "angle": 0, + "content": "Abstract. Despite CLIP [29] being the foundation model in numerous vision-language applications, CLIP suffers from a severe text spotting bias. Such bias causes CLIP models to 'Parrot' the visual text embedded within images while disregarding the authentic visual semantics. We uncover that in the most popular image-text dataset LAION-2B [31], the captions also densely parrot (spell) the text embedded in images. Our analysis shows that around \\(50\\%\\) of images are embedded with visual text content and around \\(30\\%\\) of captions words are concurrently embedded in the visual content. Based on such observation, we thoroughly inspect the different released versions of CLIP models and verify that the visual text is a dominant factor in measuring the LAION-style image-text similarity for these models. To examine whether these parrot captions shape the text spotting bias, we train a series of CLIP models with LAION subsets curated by different parrot-caption-oriented criteria. We show that training with parrot captions easily shapes such bias but harms the expected visual-language representation learning in CLIP models across various vision-language downstream tasks. This suggests that it is urgent to revisit either the design of CLIP-like models or the existing image-text dataset curation pipeline built on CLIP score filtering. Project page: https://tinyq17.github.io/CLIP-Parrot-Bias/" + }, + { + "type": "text", + "bbox": [ + 0.262, + 0.598, + 0.627, + 0.612 + ], + "angle": 0, + "content": "Keywords: Image-Text Dataset \\(\\cdot\\) Text Spotting Bias" + }, + { + "type": "title", + "bbox": [ + 0.216, + 0.636, + 0.377, + 0.652 + ], + "angle": 0, + "content": "1 Introduction" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.666, + 0.788, + 0.818 + ], + "angle": 0, + "content": "Recently, contrastive learning models [17, 29, 31] pre-trained with large-scale image-text pair data has led to numerous vision-language modeling task breakthroughs. Due to its efficiency and simplicity, the pioneering work CLIP [29] now serves as a foundation model in various applications [20, 26, 30, 49]. However, several works [4,13] have shown that CLIP models have perpetuating biases towards visual text [19,25], color [34,44], gender [41], etc. In this paper, we focus on probing the visual text bias in CLIP, i.e., the capacity of spotting text in images. Most of the previous cues [25, 29, 34] attribute the sources of biases to the noisy pre-training data. Therefore, we begin by taking a close look at the most popular dataset, LAION-2B [31]." + }, + { + "type": "page_footnote", + "bbox": [ + 0.218, + 0.825, + 0.757, + 0.841 + ], + "angle": 0, + "content": "* Equal contribution. ⌒ Corresponding author (Email: heconghui@pjlab.org.cn)." + } + ], + [ + { + "type": "page_number", + "bbox": [ + 0.218, + 0.116, + 0.23, + 0.127 + ], + "angle": 0, + "content": "2" + }, + { + "type": "header", + "bbox": [ + 0.272, + 0.115, + 0.356, + 0.127 + ], + "angle": 0, + "content": "Y. Lin et al." + }, + { + "type": "image", + "bbox": [ + 0.218, + 0.143, + 0.787, + 0.401 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.44, + 0.4, + 0.553, + 0.41 + ], + "angle": 0, + "content": "(b) Image Paired Captions" + }, + { + "type": "image_caption", + "bbox": [ + 0.216, + 0.421, + 0.788, + 0.505 + ], + "angle": 0, + "content": "Fig. 1: In LAION-2B [31], image-text pairs with the Top-5% highest similarity score are most dominant by visual text! These samples have dense concurrent text appearing in captions and images (text form in pixels). We refer to their captions as Parrot Captions as they raise a question: Dose CLIP Simply Parroting Text in Images for Vision-Language Alignment? The concurrent text is spotted by the OCR model and highlighted with color in image-text pairs. (Best view in color)" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.538, + 0.788, + 0.734 + ], + "angle": 0, + "content": "Considering the massive scale of the image-text data, it is non-trivial to assess the bias simply with a rough estimation. To this end, we first do image clustering on the whole dataset and rank each cluster by CLIP scores to analyze the most preferred types of image-text pairs under CLIP score measurement. As shown in Fig. 1, we surprisingly observe that a decent number of samples with top CLIP scores have dense concurrent text appearing in the captions and the images in the form of pixels. These samples break the assumption that CLIP models leverage text supervision to align the visual and language concepts. We refer to these captions as Parrot Captions as they provide another shortcut to achieve the same goal by teaching CLIP to do text spotting even without perceiving the actual visual concepts. To understand the underlying impact, we analyze the parrot captions from three perspectives: dataset, widely used released models, and model training. The results lead to three key findings:" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.735, + 0.788, + 0.841 + ], + "angle": 0, + "content": "Firstly, captions in LAION-2B have a significant bias towards describing visual text content embedded in the images. We provide thorough profiling using off-the-self text spotting models on the LAION-2B dataset and show that over \\(50\\%\\) of the images are embedded with visual text content. Moreover, by examining the spotted text content and the paired caption in the subset that images embedded with text, we find that over \\(60\\%\\) of the captions at least have one concurrent word and reach at least around \\(30\\%\\) words overlap" + } + ], + [ + { + "type": "header", + "bbox": [ + 0.448, + 0.115, + 0.732, + 0.129 + ], + "angle": 0, + "content": "Parrot Captions Teach CLIP to Spot Text" + }, + { + "type": "page_number", + "bbox": [ + 0.775, + 0.117, + 0.785, + 0.127 + ], + "angle": 0, + "content": "3" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.147, + 0.785, + 0.192 + ], + "angle": 0, + "content": "between the caption and spotted text from images. This finding suggests that the basic assumption of image-text semantic alignment in CLIP does not fully stand its ground when training with LAION-style data." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.192, + 0.788, + 0.403 + ], + "angle": 0, + "content": "Secondly, released CLIP models have strong text spotting bias almost in every style of web images, resulting in CLIP-filtering datasets inherently biased towards visual text-dominant data. We investigate OpenAI released CLIP model's behaviors in the LAION-2B dataset by examining the difference between alignment scores before and after text removal. The results show that CLIP model predictions densely correlate the visual text embedded in images with their parrot captions. Next, we further study the preference of the text spotting capacity on text content in CLIP and OpenCLIP models. Note that CLIP is trained on WIT-400M, while OpenCLIP uses the LAION-2B dataset. Therefore, we use synthetic images embedded with specific rendered text to avoid overfitting in OpenCLIP models. Our analysis shows that OpenCLIP is more biased toward text spotting than CLIP. We believe that the parrot caption plays a lurking role in training these released CLIP models and is the source of text spotting capacity instead of emergence behavior [42]." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.404, + 0.788, + 0.585 + ], + "angle": 0, + "content": "Thirdly, CLIP models can easily learn text spotting capacity from parrot captions while failing to connect the vision-language semantics, just like a text spotting parrot. We sample different LAION-2B subsets curated by text-oriented criteria, including the embedded text ratio, the concurrent word ratios, and the relative CLIP score from text removal to train CLIP models under the same setting. The results show that using parrot captions data, CLIP model can learn strong text spotting capacity but lose most of the zero-shot generalization ability on image-text downstream tasks. Moreover, we also observe similar behaviors on various downstream tasks such VQA, captioning, and retrieval on BLIP [21] models. Lastly, we argue that the existing data curation pipeline built on CLIP score and the contrastive fashion urgently needs to be re-examined by considering such hidden parrot captions." + }, + { + "type": "title", + "bbox": [ + 0.215, + 0.606, + 0.388, + 0.621 + ], + "angle": 0, + "content": "2 Related Work" + }, + { + "type": "title", + "bbox": [ + 0.215, + 0.636, + 0.606, + 0.651 + ], + "angle": 0, + "content": "2.1 Contrastive Vision-Language Pre-training" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.659, + 0.788, + 0.842 + ], + "angle": 0, + "content": "Modeling vision and language by aligning the embedding similarity between paired image-text data [17, 29, 31] has shown great potential for transferable to downstream vision-language tasks. The pre-training techniques mainly contain the vision encoder [10, 15] for image embedding encoding, text encoder [9] for text embedding modeling, and cross-modal contrastive learning [17, 22, 29, 47] for learning a joint embedding space of vision and language. The pioneering work CLIP [29] leverages 400 million noisy image-text pairs to learn transferable visual representation from text supervision and show impressive zero-shot performance for various vision-language tasks. Following CLIP, several vision-language models such as ALIGN [17], BASIC [27], and Open-CLIP [31] are proposed, and CLIP models have been replicated on various datasets including WIT [29], LAION [31], COYO [6], and DataComp [11]. We mainly profile" + } + ], + [ + { + "type": "page_number", + "bbox": [ + 0.218, + 0.116, + 0.232, + 0.127 + ], + "angle": 0, + "content": "4" + }, + { + "type": "header", + "bbox": [ + 0.272, + 0.115, + 0.356, + 0.128 + ], + "angle": 0, + "content": "Y. Lin et al." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.147, + 0.788, + 0.208 + ], + "angle": 0, + "content": "the LAION-2B [31] dataset due to its large scale and wide usage [26, 30] and two versions of pre-trained models, CLIP and OpenCLIP. Note that the 2 billion image-text pairs in the LAION-2B dataset are filtered by OpenAI released CLIP models, making OpenCLIP connect to CLIP closely." + }, + { + "type": "title", + "bbox": [ + 0.215, + 0.23, + 0.498, + 0.246 + ], + "angle": 0, + "content": "2.2 Studying of CLIP Behaviors" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.255, + 0.789, + 0.483 + ], + "angle": 0, + "content": "Despite the strong zero-shot and transferable performance of CLIP, the perpetuating biases [1, 3, 13, 19, 41, 46] in CLIP are still not well investigated due to its large-scale noisy training data. Much research [2, 25, 34, 37, 40, 43, 44] focuses on revealing or enhancing the downstream performance with discovered bias in CLIP. For example, colorful masks [44] or red circles [34] applied to images can improve the zero-shot performance on visual localization tasks. In studying visual text content bias, [13] shows the multimodal neurons of CLIP not only respond to visual content and the visual text embedded in the image. Another work [19] shows that image recognition in CLIP can be strongly dominated by the visual text embedded in the image. To disentangle such bias, [25] attempts to separate the text spotting representation in pre-trained CLIP by training representation projection. Meanwhile, LoGoPrompt [33] enhances the classification performance by utilizing the visual text content as auxiliary prompts as input. Also, CLIPPO [39] shows that directly aligning the image and synthetic images print with the captions can perform similarly to CLIP without a text encoder." + }, + { + "type": "title", + "bbox": [ + 0.215, + 0.504, + 0.547, + 0.518 + ], + "angle": 0, + "content": "2.3 Data Curation with Text Removal" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.529, + 0.789, + 0.741 + ], + "angle": 0, + "content": "Due to the successful practice of data curation in LAION datasets [31, 32] on scaling up the image-text datasets, searching advanced selection strategy to improve the data quality from common crawl data pool gains a growing interest [11]. Recently, several works [7, 24, 28] suggest that introducing text-related filtering methods improves the pre-training dataset quality. In DiHT [28], the data curation steps include filtering out the image-text pairs with high OCR confidence and matching text ratio. Moreover, [7, 24] mainly focus on studying the importance of filtering out the text-dominate images utilizing OCR models to improve pre-training dataset quality. Maini et al. [24] also draw the observation that \\(40\\%\\) of LAION's image text is highly correlated with the caption, but only performing a small pilot study on 500 samples with manual judgment. Differently, this paper makes the first attempt to systemically reveal the source of text spotting capacity in CLIP is the data bias and the consequences of such bias in existing widely used datasets and pre-trained models." + }, + { + "type": "title", + "bbox": [ + 0.215, + 0.763, + 0.377, + 0.781 + ], + "angle": 0, + "content": "3 Terminology" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.795, + 0.787, + 0.842 + ], + "angle": 0, + "content": "The data processing on images in the following sections mainly covers clustering, text spotting (OCR), and text inpainting. Firstly, we cluster all images based on feature similarity. For each image-text pair, we then use the pre-trained text" + } + ], + [ + { + "type": "header", + "bbox": [ + 0.448, + 0.115, + 0.732, + 0.129 + ], + "angle": 0, + "content": "Parrot Captions Teach CLIP to Spot Text" + }, + { + "type": "page_number", + "bbox": [ + 0.775, + 0.117, + 0.785, + 0.127 + ], + "angle": 0, + "content": "5" + }, + { + "type": "image", + "bbox": [ + 0.29, + 0.145, + 0.373, + 0.209 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.375, + 0.145, + 0.458, + 0.209 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.395, + 0.21, + 0.434, + 0.218 + ], + "angle": 0, + "content": "All-Emb" + }, + { + "type": "image_caption", + "bbox": [ + 0.375, + 0.219, + 0.448, + 0.226 + ], + "angle": 0, + "content": "Text Removal" + }, + { + "type": "image", + "bbox": [ + 0.46, + 0.145, + 0.543, + 0.209 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.478, + 0.21, + 0.52, + 0.218 + ], + "angle": 0, + "content": "Co-Emb" + }, + { + "type": "image_caption", + "bbox": [ + 0.468, + 0.219, + 0.534, + 0.226 + ], + "angle": 0, + "content": "Text Removal" + }, + { + "type": "image", + "bbox": [ + 0.545, + 0.145, + 0.628, + 0.209 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.565, + 0.21, + 0.607, + 0.218 + ], + "angle": 0, + "content": "Random" + }, + { + "type": "image_caption", + "bbox": [ + 0.565, + 0.218, + 0.609, + 0.226 + ], + "angle": 0, + "content": "Inpainted" + }, + { + "type": "image", + "bbox": [ + 0.632, + 0.145, + 0.714, + 0.209 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.645, + 0.21, + 0.694, + 0.218 + ], + "angle": 0, + "content": "Syn-Emb." + }, + { + "type": "image_caption", + "bbox": [ + 0.638, + 0.218, + 0.706, + 0.226 + ], + "angle": 0, + "content": "Text Template" + }, + { + "type": "image", + "bbox": [ + 0.291, + 0.229, + 0.373, + 0.292 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.313, + 0.292, + 0.351, + 0.302 + ], + "angle": 0, + "content": "Caption" + }, + { + "type": "image", + "bbox": [ + 0.375, + 0.229, + 0.458, + 0.293 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.388, + 0.293, + 0.437, + 0.302 + ], + "angle": 0, + "content": "All-Emb." + }, + { + "type": "image_caption", + "bbox": [ + 0.388, + 0.302, + 0.44, + 0.31 + ], + "angle": 0, + "content": "Text Mask" + }, + { + "type": "image", + "bbox": [ + 0.46, + 0.229, + 0.543, + 0.293 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.477, + 0.293, + 0.521, + 0.302 + ], + "angle": 0, + "content": "Co-Emb." + }, + { + "type": "image_caption", + "bbox": [ + 0.475, + 0.302, + 0.526, + 0.31 + ], + "angle": 0, + "content": "Text Mask" + }, + { + "type": "image", + "bbox": [ + 0.545, + 0.229, + 0.628, + 0.293 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.566, + 0.293, + 0.607, + 0.301 + ], + "angle": 0, + "content": "Random" + }, + { + "type": "image_caption", + "bbox": [ + 0.551, + 0.302, + 0.621, + 0.311 + ], + "angle": 0, + "content": "Sampled Mask" + }, + { + "type": "image", + "bbox": [ + 0.639, + 0.229, + 0.714, + 0.293 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.646, + 0.293, + 0.695, + 0.301 + ], + "angle": 0, + "content": "Syn-Emb." + }, + { + "type": "image_caption", + "bbox": [ + 0.638, + 0.302, + 0.706, + 0.311 + ], + "angle": 0, + "content": "Text Template" + }, + { + "type": "image_caption", + "bbox": [ + 0.216, + 0.323, + 0.784, + 0.338 + ], + "angle": 0, + "content": "Fig. 2: Visualization of defined terminologies. Co-Emb. Text is marked in the caption." + }, + { + "type": "title", + "bbox": [ + 0.218, + 0.356, + 0.66, + 0.371 + ], + "angle": 0, + "content": "Algorithm 1 Pseudocode of Detecting Co-Emb. Text (Rate)" + }, + { + "type": "text", + "bbox": [ + 0.216, + 0.375, + 0.473, + 0.385 + ], + "angle": 0, + "content": "caption: captions from LAION-2B dataset." + }, + { + "type": "text", + "bbox": [ + 0.221, + 0.385, + 0.444, + 0.393 + ], + "angle": 0, + "content": "ocr_text: text spotted by OCR model" + }, + { + "type": "text", + "bbox": [ + 0.218, + 0.394, + 0.619, + 0.403 + ], + "angle": 0, + "content": "cap_words,OCR_words \\(=\\) set(capTION.split()，set(ocr_text.split()))" + }, + { + "type": "text", + "bbox": [ + 0.218, + 0.403, + 0.509, + 0.412 + ], + "angle": 0, + "content": "co_emb_text = intersection(cap_words,OCR_words)" + }, + { + "type": "text", + "bbox": [ + 0.218, + 0.412, + 0.534, + 0.421 + ], + "angle": 0, + "content": "co_emb_text_rate = len(co_emb_text) / len(cap_words)" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.457, + 0.787, + 0.532 + ], + "angle": 0, + "content": "spotting model to detect and recognize the text print in image pixels. The mask images in Fig. 2 are the spotted text area. Next, we match the spotted text with the caption using Algorithm 1 to obtain the concurrent words and their ratio in captions. Lastly, we use inpainting to remove the text from the image for CLIPs' pattern ablation. To avoid confusion, we define these concepts as follows," + }, + { + "type": "text", + "bbox": [ + 0.225, + 0.549, + 0.788, + 0.655 + ], + "angle": 0, + "content": "- Embedded Text: text spotted by OCR models from the images. To study the correlation of embedded text with captions, we define different kinds of embedded text as, 1) All-Emb. Text: all the text is spotted from an image; 2) Co-Emb. Text: spotted text concurrently appears in the image's corresponding captions; 3) Syn-Emb. Text: synthetic text rendered in an image with a fixed font and a blank background. Fig. 2 shows examples of spotted embedded text by binary mask and the rendering synthetic text." + }, + { + "type": "text", + "bbox": [ + 0.225, + 0.656, + 0.788, + 0.685 + ], + "angle": 0, + "content": "- Co-Emb. Text Rate (CoTR): the word set IoU of Co-Emb. text and captions (See Algorithm 1)." + }, + { + "type": "text", + "bbox": [ + 0.225, + 0.687, + 0.788, + 0.717 + ], + "angle": 0, + "content": "- Image w/ or w/o Embedded Text: spotted text results of a given image are none-empty or empty." + }, + { + "type": "text", + "bbox": [ + 0.225, + 0.718, + 0.788, + 0.763 + ], + "angle": 0, + "content": "- Text Removal Image: do inpainting in the specific spotted text area (All-Emb., Co-Emb., or Random). Random is implemented by sampling other image's text areas. For the different inpainting results, see Fig. 2." + }, + { + "type": "text", + "bbox": [ + 0.225, + 0.764, + 0.788, + 0.824 + ], + "angle": 0, + "content": "- Relative Scores (RSA/RSC): the difference in CLIP score between images modified by different inpainting operations while keeping the same captions. RSA and RSC are the short for the relative scores before and after removing All-Emb. text and Co-Emb. text." + }, + { + "type": "text", + "bbox": [ + 0.225, + 0.825, + 0.641, + 0.84 + ], + "angle": 0, + "content": "- Image Clusters: image partitions based on K-Means." + }, + { + "type": "list", + "bbox": [ + 0.225, + 0.549, + 0.788, + 0.84 + ], + "angle": 0, + "content": null + } + ], + [ + { + "type": "page_number", + "bbox": [ + 0.218, + 0.116, + 0.23, + 0.127 + ], + "angle": 0, + "content": "6" + }, + { + "type": "header", + "bbox": [ + 0.272, + 0.115, + 0.356, + 0.128 + ], + "angle": 0, + "content": "Y. Lin et al." + }, + { + "type": "table_caption", + "bbox": [ + 0.216, + 0.145, + 0.787, + 0.174 + ], + "angle": 0, + "content": "Table 1: Overall parrot captions statistic. More than \\(50\\%\\) of images are embedded with text, and \\(30\\%\\) of caption words are printed in images!" + }, + { + "type": "table", + "bbox": [ + 0.33, + 0.186, + 0.673, + 0.268 + ], + "angle": 0, + "content": "
Number of Total Images1,985,284,122
Number of Images w/ Emb. Text1,083,896,427
Number of Images w/ Co-Emb. Text663,600,432
Co-Emb. Text Rate (in Total)15.42%
Co-Emb. Text Rate (in Images w/ Emb. Text)28.24%
Fuzzy Co-Emb. Text Rate (in Total)30.46%
Fuzzy Co-Emb. Text Rate (in Images w/ Emb. Text)55.79%
" + }, + { + "type": "text", + "bbox": [ + 0.225, + 0.291, + 0.784, + 0.321 + ], + "angle": 0, + "content": "- CLIP and OpenCLIP: CLIP models are trained on WIT-400M [29] and LAION-2B [31] dataset." + }, + { + "type": "text", + "bbox": [ + 0.226, + 0.321, + 0.785, + 0.352 + ], + "angle": 0, + "content": "- N-gram Vocabulary (Vocab): the set of all contiguous N word sequences extracted from a text corpus, such as the collection of all captions." + }, + { + "type": "list", + "bbox": [ + 0.225, + 0.291, + 0.785, + 0.352 + ], + "angle": 0, + "content": null + }, + { + "type": "title", + "bbox": [ + 0.216, + 0.37, + 0.504, + 0.387 + ], + "angle": 0, + "content": "4 Profiling LAION-2B Data" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.399, + 0.785, + 0.489 + ], + "angle": 0, + "content": "To better profile the image-text pair data on a billion scale, we first cluster all the images based on CLIP features into 4,000 clusters and sort each cluster with CLIP scores. After obtaining all the cluster labels, we use the SOTA text spotting model [45] to get the visual text content on all the collected images. Finally, we aggregate all the model-predicted results and compare them with their corresponding captions to bring out our observations." + }, + { + "type": "title", + "bbox": [ + 0.216, + 0.508, + 0.457, + 0.523 + ], + "angle": 0, + "content": "4.1 Implementation Details" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.529, + 0.787, + 0.604 + ], + "angle": 0, + "content": "Clustering with CLIP Features: We first train K-Means (implemented by Faiss [18]) on the LAION-400M [32] subset using ViT-B-32 [10] CLIP features to speed up the clustering process. Due to the large memory consumption, we reduce the feature dimensions from 512 to 256 using PCA. Then, we partition the whole dataset using trained K-Means with the same feature extraction pipeline." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.605, + 0.787, + 0.772 + ], + "angle": 0, + "content": "Text Spotting and Matching: To detect and recognize text across various scenes, we adopt DeepSolo [45] as our text spotting model and use the pre-trained checkpoints with the ViTAEv2-S [48] backbone in default setting. The output format of the text spotting model is a sequence of polygons of text location and their recognized characters. Despite its strong performance, we empirically find that DeepSolo can not handle the crowd scenes well (with more than 100 separate words) which is only a small proportion of the dataset (\\(\\sim 2\\%\\)). To identify the correlation between the spotted text and captions, we use Algorithm 1 to calculate the Co-Emb. text rate in each image-text pair. Considering the imperfect text spotting predictions might miss or misspell words, we also use Levenshtein distance to calculate the fuzzing similarity and reported in Tab. 1." + }, + { + "type": "title", + "bbox": [ + 0.216, + 0.789, + 0.62, + 0.803 + ], + "angle": 0, + "content": "4.2 Statistic and Observations from LAION-2B" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.81, + 0.785, + 0.84 + ], + "angle": 0, + "content": "The overall statistics of the 2 billion image-text pairs are reported in Tab. 1. In summary, the images embedded with visual text content reach a surprisingly" + } + ], + [ + { + "type": "header", + "bbox": [ + 0.448, + 0.115, + 0.731, + 0.129 + ], + "angle": 0, + "content": "Parrot Captions Teach CLIP to Spot Text" + }, + { + "type": "page_number", + "bbox": [ + 0.775, + 0.117, + 0.785, + 0.127 + ], + "angle": 0, + "content": "7" + }, + { + "type": "image", + "bbox": [ + 0.258, + 0.144, + 0.744, + 0.262 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.29, + 0.264, + 0.714, + 0.274 + ], + "angle": 0, + "content": "(a) The ratio of different OCR-oriented data types in LAION-2B clusters." + }, + { + "type": "image", + "bbox": [ + 0.261, + 0.275, + 0.744, + 0.352 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.265, + 0.353, + 0.739, + 0.365 + ], + "angle": 0, + "content": "(b) Top CLIP score samples visualization from 50 clusters with ratio over \\(80\\%\\)." + }, + { + "type": "image_caption", + "bbox": [ + 0.215, + 0.376, + 0.788, + 0.46 + ], + "angle": 0, + "content": "Fig. 3: (a): Based on the OCR prediction results, the image-text pairs are divided into three types: image without visual embedded text content; the spotted text from the image has no concurrent text with the caption; the spotted text at least share one concurrent word with the caption. The clusters are merged from 4000 into 100 for a better view. (b): In the clusters with high ratio, the top CLIP score samples contain various text sources, such as posters, book covers, advertisements, and even slides." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.492, + 0.789, + 0.703 + ], + "angle": 0, + "content": "high proportion of \\(54.60\\%\\) in the investigated data. Around \\(15\\%\\) of words in the dataset captions are Co-Emb. text, and the proportion of Co-Emb. text can further reach \\(30\\%\\) when considering the fuzzy matching results of the spotted text and captions. This suggests that CLIP models trained on these data might lead to a high bias toward text spotting. To better visualize the data distribution, we provide cluster-specific statics results and top CLIP score samples of text-dominated clusters in Fig. 3. We divide all images into 100 clusters based on visual similarity and visualize them according to the OCR results. Every cluster contains more or less images embedded with text. Combined with sample visualization, we observe that in the LAION collected data, the parrot captions cover various scenes. In the subsets of images embedded with text, around \\(60\\%\\) of captions at least precisely parrot one concurrent word (Co-Emb. Text Rate \\(>0\\)) appearing in the image. It suggests that the data collection pipeline of LAION [31] has a strong bias to introduce parrot captions from web data." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.705, + 0.789, + 0.841 + ], + "angle": 0, + "content": "To better understand Co-Emb. Text, we provide a more thorough analysis of the word counting and text size of parrot captions. As shown in Fig. 4a, the results show that a large proportion of the Co-Emb. Text only takes a few words. However, we also find a large number of captions that are almost full parrot captions (see areas around the heatmap diagonal). Next, in Fig. 4b and Fig. 4c, we investigate the correlation between the size of concurrent words box in the image and CLIP score. The results show that the large text size does not usually lead to a higher score; meanwhile, the small text size can also dominate the score as the score can be significantly different after removing them. (Details" + } + ], + [ + { + "type": "page_number", + "bbox": [ + 0.218, + 0.116, + 0.23, + 0.127 + ], + "angle": 0, + "content": "8" + }, + { + "type": "header", + "bbox": [ + 0.272, + 0.115, + 0.356, + 0.127 + ], + "angle": 0, + "content": "Y. Lin et al." + }, + { + "type": "image", + "bbox": [ + 0.258, + 0.145, + 0.744, + 0.263 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.215, + 0.271, + 0.788, + 0.342 + ], + "angle": 0, + "content": "Fig. 4: (a): The number of caption words and associated spotted concurrent words based on precise word matching. (b): Distribution of total box area of concurrent words in the image and its CLIP score. (c): Distribution of total box area of concurrent words and its relative CLIP score before and after removing them from the image. (d): Distribution of text size of the single concurrent word and other spotted word." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.371, + 0.788, + 0.418 + ], + "angle": 0, + "content": "of text removal described in Sec. 5.1). One possible reason is the text content and input resolution may matter more for CLIP. Moreover, we discover that the larger text is more likely to be parroted in the captions, as shown in Fig. 4d." + }, + { + "type": "title", + "bbox": [ + 0.215, + 0.44, + 0.613, + 0.459 + ], + "angle": 0, + "content": "5 Inspecting Pre-Trained CLIP Models" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.473, + 0.789, + 0.625 + ], + "angle": 0, + "content": "It is important to note that the LAION-2B dataset collection pipeline uses CLIP score from OpenAI's model to filter out the image-text pair below 0.28. Therefore, we inspect these two released CLIP models [29,31] to answer better why LAION data contains such a high proportion of parrot captions. Specifically, OpenAI's CLIP model is trained on the WIT dataset (out-of-domain model), and OpenCLIP is trained on LAION-2B (in-domain model). We first study whether the embedded text is the key factor in CLIP filtering by ablating the embedded text using text inpainting. Moreover, we further investigate whether the text spotting capacity prefers specific text content by examining synthetic images with Syn-Emb. text." + }, + { + "type": "title", + "bbox": [ + 0.215, + 0.648, + 0.571, + 0.663 + ], + "angle": 0, + "content": "5.1 Ablation of Embedded Text Removal" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.674, + 0.788, + 0.795 + ], + "angle": 0, + "content": "Text Removal via Inpainting: Given the OCR predicted results, we use the fast marching method [38] to inpaint the area of the spotted text polygons. We generate two versions of text removal results for each image with embedded text, i.e., All-Emb. text removal and Co-Emb. text removal, as the parrot caption prediction is imperfect due to the limitation of OCR models. We also generate random inpainting images with randomly sampled spotted text polygons from other images to ablate the distribution shift caused by image inpainting. Examples of the spotted text masks and inpainting results are shown in Fig. 2." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.796, + 0.788, + 0.842 + ], + "angle": 0, + "content": "Results: Based on the OCR predicted results and text inpainting operations, we can obtain six types of LAION images, including \\(\\bullet\\)): images without any embedded text (OCR results are empty); \\(\\bullet\\)): images with any embedded text" + } + ], + [ + { + "type": "header", + "bbox": [ + 0.448, + 0.115, + 0.732, + 0.129 + ], + "angle": 0, + "content": "Parrot Captions Teach CLIP to Spot Text" + }, + { + "type": "page_number", + "bbox": [ + 0.775, + 0.117, + 0.785, + 0.127 + ], + "angle": 0, + "content": "9" + }, + { + "type": "image", + "bbox": [ + 0.218, + 0.146, + 0.784, + 0.282 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.214, + 0.292, + 0.788, + 0.337 + ], + "angle": 0, + "content": "Fig. 5: Left: Mean CLIP scores of image-text pairs with different text removal operations depicted in Sec. 5.1, and grouped by cluster the same as Fig. 3. Right: Overall relative CLIP score distribution by comparing different text removal operations." + }, + { + "type": "table_caption", + "bbox": [ + 0.264, + 0.349, + 0.737, + 0.364 + ], + "angle": 0, + "content": "Table 2: Mean CLIP score of different setups of text removal." + }, + { + "type": "table", + "bbox": [ + 0.364, + 0.377, + 0.645, + 0.464 + ], + "angle": 0, + "content": "
SetupAverage CLIP Score
Raw Images w/o Emb. Text0.3223 ± 0.0078
Raw Images w/ Emb. Text0.3358 ± 0.0094
× All-Emb. Random Inpainted0.3260 ± 0.0057
× All-Emb. Text Removal0.2974 ± 0.0197
Co-Emb. Random Inpainted0.3341 ± 0.0051
Co-Emb. Text Removal0.2993 ± 0.0146
" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.493, + 0.788, + 0.795 + ], + "angle": 0, + "content": "(OCR results are none-empty); \\(\\times\\)): images with random inpainting by other image's All-Emb. text area; \\(\\times\\)): images removed All-Emb. text (Inpaint all the areas of OCR predicted text); ■): images randomly inpainted by other image's Co-Emb. text area, and ■): images removed Co-Emb. text (Inpaint the areas of concurrent text in OCR predicted text and captions). Then, we calculate CLIP scores of all the groups of images and their paired captions using OpenAI released CLIP model (ViT-B-32). Fig. 5 reports the mean scores of different types of images in each cluster and raises four observations as follows: I). The images embedded with text achieve higher CLIP scores in most clusters than those without embedded text; II). CLIP scores significantly drop once we remove the text from the images compared to its random inpainting baseline. It indicates that the parrot captions correlate highly with CLIP score measurement; III). Not all the samples are dominated by the embedded text, as some samples achieve higher scores after removing text, indicating the embedded text can also be a distractor; IV). Most of the relative CLIP scores (\\(S(\\square) - S(\\times)\\)) between images removed Co-Emb. text and All-Emb. text are positive, as shown in the right of Fig. 5. The straightforward reason is the images lose more visual information due to the larger in-painting area. Another possible reason is the imperfect text spotting prediction or the corner cases in the matching algorithm leaking parts of the concurrent text in images." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.795, + 0.788, + 0.84 + ], + "angle": 0, + "content": "Discussion: Due to the text removal, the image distribution may shift from the CLIP training set. Therefore, we provide two random inpainting baselines to examine the effect of distribution shift. In Tab. 2, we report the mean scores" + } + ], + [ + { + "type": "page_number", + "bbox": [ + 0.218, + 0.116, + 0.236, + 0.127 + ], + "angle": 0, + "content": "10" + }, + { + "type": "header", + "bbox": [ + 0.272, + 0.115, + 0.356, + 0.127 + ], + "angle": 0, + "content": "Y. Lin et al." + }, + { + "type": "image", + "bbox": [ + 0.226, + 0.144, + 0.356, + 0.267 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.356, + 0.144, + 0.486, + 0.267 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.261, + 0.269, + 0.468, + 0.28 + ], + "angle": 0, + "content": "(a) Statistic of 1-gram vocabularies" + }, + { + "type": "image", + "bbox": [ + 0.226, + 0.281, + 0.501, + 0.397 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.251, + 0.398, + 0.477, + 0.41 + ], + "angle": 0, + "content": "(c) Grouped by Caption 1-gram length" + }, + { + "type": "image", + "bbox": [ + 0.501, + 0.149, + 0.78, + 0.268 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.515, + 0.269, + 0.765, + 0.28 + ], + "angle": 0, + "content": "(b) Grouped by 1-gram frequency intervals" + }, + { + "type": "image", + "bbox": [ + 0.501, + 0.28, + 0.778, + 0.397 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.526, + 0.398, + 0.754, + 0.41 + ], + "angle": 0, + "content": "(d) Grouped by Caption 2-gram length" + }, + { + "type": "image_caption", + "bbox": [ + 0.214, + 0.422, + 0.788, + 0.464 + ], + "angle": 0, + "content": "Fig.6: OpenCLIP more bias towards text spotting than CLIP model. Grouped score distributions of prompting CLIP and OpenCLIP models with N-gram Syn-Emb. text and synthetic images for model preference investigation." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.493, + 0.788, + 0.54 + ], + "angle": 0, + "content": "of different setups. Results show that the random baselines are very close to the raw image baseline, indicating that CLIP model is robust to the distribution shift caused by information loss in inpainted regions." + }, + { + "type": "title", + "bbox": [ + 0.215, + 0.56, + 0.525, + 0.575 + ], + "angle": 0, + "content": "5.2 Prompting with Syn-Emb. Text" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.584, + 0.788, + 0.749 + ], + "angle": 0, + "content": "Generating Synthetic Images from N-gram Vocabulary: To investigate CLIP models' text spotting preference, we adopt a similar strategy in [25] to use synthetic images to embed specific text content by rendering text in a blank background. For each text, we use four fore-background style rendering templates (black-white, black-grey, white-grey, and white-black), as shown in Fig. 2. Different from the uniformly sampling letters in [25], we generate the text content from the N-gram vocabulary built from captions and Co-Emb. text to study the text spotting pattern. We only select the top frequent 400,000 grams for each vocabulary. The statistics of 1-gram vocabulary are shown in Fig. 6a, which is a long-tail distribution. Next, we calculate the synthetic images and the rendered text similarity on released ViT-B-32 CLIP and OpenCLIP models." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.75, + 0.788, + 0.84 + ], + "angle": 0, + "content": "Results: Firstly, we examine whether CLIP models prefer recognizing more commonly seen words (with high frequency in vocabulary). Therefore, we group the 1-gram results based on their frequency interval in the whole vocabulary, as shown in Fig. 6b. The results show that OpenCLIP model clearly has a stronger text spotting capacity than CLIP, i.e., more biased towards text spotting. We also observe that all CLIP models are more sensitive to the vocabulary built" + } + ], + [ + { + "type": "header", + "bbox": [ + 0.448, + 0.115, + 0.732, + 0.129 + ], + "angle": 0, + "content": "Parrot Captions Teach CLIP to Spot Text" + }, + { + "type": "page_number", + "bbox": [ + 0.769, + 0.117, + 0.784, + 0.127 + ], + "angle": 0, + "content": "11" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.147, + 0.788, + 0.298 + ], + "angle": 0, + "content": "from the concurrent words. Interestingly, both CLIP and OpenCLIP models have slightly higher scores on the less frequent grams. Secondly, considering the long-tail grams might contain more characters, we further group the 1-gram and 2-gram results based on their text length in Fig. 6c and Fig. 6d. Note that the Co-Emb. text is not regularly arranged in the images, making it hard to extract continuous word sequences. Results show that all the models are better at spotting the longer words, possibly due to the tokenizer used in the text encoder, making them more discriminative. Meanwhile, in the groups of 2-gram samples, the scores gradually drop when spotting the highly long text, indicating that the spotting capacity of CLIP models is possibly built on word-by-word." + }, + { + "type": "title", + "bbox": [ + 0.215, + 0.318, + 0.613, + 0.334 + ], + "angle": 0, + "content": "6 Training on Emb. Text Curated Data" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.346, + 0.784, + 0.375 + ], + "angle": 0, + "content": "Next, we dive deeper into training CLIP and BLIP [21] models on different Emb. text curated subsets and studying various downstream task behaviors." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.377, + 0.788, + 0.481 + ], + "angle": 0, + "content": "Implementation Details: For CLIP, we use the open-source software OpenCLIP [16] for all CLIP model training. Our experiments are conducted on both ViT-B [10] and RN50 [15]. We use 4,096 batch size for 3M and 8,192 for 12M scale subsets. Other settings remain the same as [31]. For BLIP, we mainly conduct on 3M scale subsets with ViT-B [10]. The BLIP models are pre-trained for 10 epochs with an AdamW [23] optimizer. For downstream tasks, we finetune 10 epochs for VQA and 5 epochs for captioning and retrieval." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.483, + 0.788, + 0.77 + ], + "angle": 0, + "content": "Evaluation: For CLIP, we follow the DataComp benchmark [11] using 38 zero-shot classification and retrieval tasks as evaluation. We report the average performance (Avg.) of the DataComp benchmark and two subset track performances, ImageNet (IN) and Retrieval (Ret.). To evaluate the text spotting capacity, we use the synthetic benchmark illustrated in Sec. 5.2 and a real-world benchmark sampled from LAION-2B as the validation set. In the synthetic benchmark, we calculate the similarity of all the 1-gram synthetic image-text pairs from caption vocabulary and report all the trained model results in Fig. 7. For the real-world benchmark, we sample 1M image-text pairs without any embedded text and 1M samples dominated by the parrot caption (RSC \\(\\geq 0.2\\)). Fig. 8 aggregates the mean scores of the 2M evaluation set and also reports the mean scores of applying text removal on the 2M evaluation set results. For BLIP, inspired by [12], we further evaluate the model behavior on downstream tasks requiring reading text, including Visual Question Answering (VQA), Image Captioning, and Text-Image Retrieval. Specifically, for the text-oriented tasks, we use Text VQA [36] and ST-VQA [5] for VQA, and TextCaps [35] for captioning and retrieval. Moreover, we also provide the same tasks on the datasets that only require the model to see, i.e., the natural image dataset. Similarly, we use VQAv2 [14] for VQA and COCO [8] for captioning and retrieval." + }, + { + "type": "title", + "bbox": [ + 0.215, + 0.788, + 0.613, + 0.803 + ], + "angle": 0, + "content": "6.1 Ablation Study of Data Curation on CLIP" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.81, + 0.787, + 0.84 + ], + "angle": 0, + "content": "Curation I: Embedded Text in Images. To study the impact of embedded text on overall pre-train data quality, we sample three subsets: random baseline," + } + ], + [ + { + "type": "page_number", + "bbox": [ + 0.218, + 0.116, + 0.236, + 0.127 + ], + "angle": 0, + "content": "12" + }, + { + "type": "header", + "bbox": [ + 0.272, + 0.115, + 0.355, + 0.127 + ], + "angle": 0, + "content": "Y. Lin et al." + }, + { + "type": "table_caption", + "bbox": [ + 0.217, + 0.145, + 0.475, + 0.172 + ], + "angle": 0, + "content": "Table 3: Ablation of images embedded with or without text." + }, + { + "type": "table", + "bbox": [ + 0.229, + 0.186, + 0.459, + 0.309 + ], + "angle": 0, + "content": "
DataModelINRet.Avg.
3M RandomRN500.2040.2220.256
3M w/o Emb. TextRN500.2280.2320.282
3M w/ Emb. Text OnlyRN500.0710.1390.164
3M RandomViT-B0.1310.1480.210
3M w/o Emb. TextViT-B0.1620.1640.234
3M w/ Emb. Text OnlyViT-B0.0520.1110.153
12M RandomRN500.3600.3540.354
12M w/o Emb. TextRN500.4090.3610.372
12M w/ Emb. Text OnlyRN500.1290.1920.218
12M RandomViT-B0.3140.2990.351
12M w/o Emb. TextViT-B0.3700.3180.364
12M w/ Emb. Text OnlyViT-B0.1290.1720.225
" + }, + { + "type": "table_caption", + "bbox": [ + 0.217, + 0.311, + 0.475, + 0.339 + ], + "angle": 0, + "content": "Table 5: Ablation of models trained on subsets sampled by different RSA." + }, + { + "type": "table", + "bbox": [ + 0.238, + 0.352, + 0.451, + 0.458 + ], + "angle": 0, + "content": "
Data (3M)ModelAvg.S(●)INRet.Avg.
RSA < 0.0RN500.3190.1810.2200.239
RSA ≥ 0.0RN500.3390.1260.1800.215
RSA ≥ 0.1RN500.3510.0410.1230.148
RSA ≥ 0.2RN500.3600.0170.0940.109
RSA ≥ 0.3RN500.3760.0090.0750.097
RSA < 0.0ViT-B0.3190.1230.1590.198
RSA ≥ 0.0ViT-B0.3390.0790.1290.185
RSA ≥ 0.1ViT-B0.3510.0310.1030.134
RSA ≥ 0.2ViT-B0.3600.0120.0800.103
RSA ≥ 0.3ViT-B0.3760.0060.0700.096
" + }, + { + "type": "table_caption", + "bbox": [ + 0.527, + 0.15, + 0.785, + 0.178 + ], + "angle": 0, + "content": "Table 4: Ablation of different CoEmb. Text Rate (CoTR)." + }, + { + "type": "table", + "bbox": [ + 0.564, + 0.19, + 0.754, + 0.305 + ], + "angle": 0, + "content": "
Data (3M)ModelINRet.Avg.
CoTR = 0.0RN500.1930.2290.247
CoTR ≥ 0.3RN500.0310.1100.137
CoTR ≥ 0.5RN500.0210.0990.124
CoTR ≥ 0.8RN500.0120.0820.096
CoTR = 1.0RN500.0120.0740.102
CoTR = 0.0ViT-B0.1320.1640.206
CoTR ≥ 0.3ViT-B0.0290.0840.130
CoTR ≥ 0.5ViT-B0.0210.0820.119
CoTR ≥ 0.8ViT-B0.0120.0760.104
CoTR = 1.0ViT-B0.0130.0760.103
" + }, + { + "type": "table_caption", + "bbox": [ + 0.527, + 0.311, + 0.785, + 0.339 + ], + "angle": 0, + "content": "Table 6: Ablation of models trained on subsets sampled by different RSC." + }, + { + "type": "table", + "bbox": [ + 0.549, + 0.352, + 0.761, + 0.458 + ], + "angle": 0, + "content": "
Data (3M)ModelAvg.S(●)INRet.Avg.
RSC < 0.0RN500.3260.1250.1710.209
RSC ≥ 0.0RN500.3450.0620.1290.168
RSC ≥ 0.1RN500.3540.0140.0910.106
RSC ≥ 0.2RN500.3640.0080.0840.104
RSC ≥ 0.3RN500.3800.0050.0580.084
RSC < 0.0ViT-B0.3260.0790.1290.174
RSC ≥ 0.0ViT-B0.3450.0450.1190.149
RSC ≥ 0.1ViT-B0.3540.0180.0910.116
RSC ≥ 0.2ViT-B0.3640.0080.0760.106
RSC ≥ 0.3ViT-B0.3800.0040.0590.091
" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.49, + 0.785, + 0.581 + ], + "angle": 0, + "content": "images without any embedded text, and images all embedded with text from LAION-2B. The subsets include 3M and 12M scales. The results in Tab. 3 show that images embedded with text generally reduce the pre-training dataset quality as all performance tracks significantly decrease. Meanwhile, in Fig. 7, the model trained with the images embedded with text achieves the strongest text spotting capacity compared to the random and images without text baselines." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.582, + 0.787, + 0.748 + ], + "angle": 0, + "content": "Curation II: Co-Emb. Text Rate (CoTR). Tab. 3 reports CLIP models trained on parrot captions with different CoTR. We first select all the images with embedded text and then sample images based on the CoTR depicted at Algorithm 1 with different thresholds. With increasing CoTR, all the zero-shot benchmark performance drops significantly. Despite the images in the subset \\((\\mathrm{CoTR} = 0)\\) all embedded with text, the pre-trained model performs similarly to the random baseline in Tab. 3. It indicates that the parrot caption is more crucial than embedded text in reducing the pre-trained data quality. For the text spotting capacity, Fig. 7 and Fig. 8 show that the increasing CoTR does not lead to stronger text spotting capacity, possibly due to the average length of captions decreasing in higher CoTR data." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.749, + 0.787, + 0.841 + ], + "angle": 0, + "content": "Curation III: Relative Score from Text Removal (RSA & RSC). Given the observations in Sec. 5.1, we further select a series of subsets based on the relative score of images before and after text removal. The higher relative scores mean the masked text is more dominant in CLIP score measurement. In Tab. 5 and Tab. 6, we report the zero-shot performance of models trained on subsets with different relative score thresholds. CLIP models pre-trained with higher" + } + ], + [ + { + "type": "header", + "bbox": [ + 0.448, + 0.115, + 0.732, + 0.129 + ], + "angle": 0, + "content": "Parrot Captions Teach CLIP to Spot Text" + }, + { + "type": "page_number", + "bbox": [ + 0.769, + 0.116, + 0.785, + 0.127 + ], + "angle": 0, + "content": "13" + }, + { + "type": "image", + "bbox": [ + 0.245, + 0.144, + 0.756, + 0.26 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.215, + 0.27, + 0.788, + 0.312 + ], + "angle": 0, + "content": "Fig. 7: CLIP models learn text spotting well from parrot captions. Benchmarking text spotting capacity of CLIP models with 1-gram caption vocabulary synthetic images dataset as the same as Sec. 5.2." + }, + { + "type": "image", + "bbox": [ + 0.342, + 0.327, + 0.662, + 0.46 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.215, + 0.469, + 0.788, + 0.511 + ], + "angle": 0, + "content": "Fig. 8: Text spotting capacity validation on real images. Models trained with more parrot captions are better at aligning the image with parrot captions but perform worse at aligning images without embedded text." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.543, + 0.788, + 0.664 + ], + "angle": 0, + "content": "RSA or RSC both get worse downstream performance. Notably, the average raw CLIP scores \\(\\mathrm{S}(\\bullet)\\) of these subsets have a positive correlation with RSA or RSC, indicating using CLIP scores from a biased pre-trained model as the data filtering strategy can be unreliable. When comparing the RSA and RSC, the results show that the samples dominated by the latter, i.e., parrot captions, are less informative for CLIP training. Meanwhile, Fig. 7 and Fig. 8 show that the text spotting capacity of CLIP can be further improved by training on the samples using relative scores as data curation criteria against CoTR." + }, + { + "type": "title", + "bbox": [ + 0.215, + 0.69, + 0.618, + 0.705 + ], + "angle": 0, + "content": "6.2 More Investigation on Text-Oriented Tasks" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.72, + 0.788, + 0.84 + ], + "angle": 0, + "content": "Inspired by [12], we further train BLIP [21] models on the curated subsets to better understand the impact of parrot caption on various vision-language tasks, especially the tasks that require models to read the text. We chose BLIP for the ablation study instead of CLIP as it can be directly applied to all these tasks. As shown in Tab. 7, training BLIPs to spot text can boost their performance on the downstream tasks requiring the model to read but impede the performance of downstream tasks only requiring the model to see, which are consistent with the observation on classification tasks. Nevertheless, when BLIPs mainly focus" + } + ], + [ + { + "type": "page_number", + "bbox": [ + 0.218, + 0.116, + 0.236, + 0.127 + ], + "angle": 0, + "content": "14" + }, + { + "type": "header", + "bbox": [ + 0.272, + 0.115, + 0.355, + 0.127 + ], + "angle": 0, + "content": "Y. Lin et al." + }, + { + "type": "table_caption", + "bbox": [ + 0.217, + 0.145, + 0.785, + 0.187 + ], + "angle": 0, + "content": "Table 7: BLIP downstream tasks performance of pre-training on different curated 3M subsets. The gray color represents tasks requiring the model to read the text from images, i.e., spotting text from images." + }, + { + "type": "table", + "bbox": [ + 0.249, + 0.2, + 0.758, + 0.413 + ], + "angle": 0, + "content": "
BLIP Data (3M)Visual Question Answering (Acc)Image Captioning (CIDEr)Text-to-Image Retrieval (R@1)Image-to-Text Retrieval (R@1)
VQAv2TextVQAST-VQACOCOTextCapsCOCOTextCapsCOCOTextCaps
Rand71.0715.3610.48115.653.748.9156.3465.4672.45
w/ Emb. Text68.9419.0512.65108.992.142.8970.158.581.42
w/o Emb. Text71.2213.659.29116.241.549.9631.8366.548.7
CoTR = 0.071.1113.979.75116.344.649.5538.0566.0854.57
CoTR ≥ 0.367.419.2811.81104.996.937.7867.2851.9878.2
CoTR ≥ 0.567.0219.6412.38102.794.135.9465.2450.3276.94
CoTR ≥ 0.866.3818.5012.00100.991.634.1362.6546.973.56
CoTR = 1.066.1818.4712.80101.291.333.5561.8346.6273.05
RSA < 0.070.7914.169.64115.744.948.2536.8564.7254.7
RSA ≥ 0.070.0318.7611.81111.984.546.2568.6162.9281.23
RSA ≥ 0.168.1419.4813.33105.696.139.9668.1354.6479.37
RSA ≥ 0.266.0121.0611.8598.794.433.0364.1747.1275.33
RSA ≥ 0.364.2018.4412.0495.2691.126.6460.1137.370.24
RSC < 0.070.1315.1910.74112.246.746.841.9563.2458.05
RSC ≥ 0.068.8620.1213.75107.893.542.069.7857.4280.92
RSC ≥ 0.167.3520.5412.84103.496.936.466.6951.0277.79
RSC ≥ 0.262.6220.3213.1498.792.830.0861.3842.9671.98
RSC ≥ 0.363.7518.9413.0392.988.724.2358.3534.7268.95
" + }, + { + "type": "text", + "bbox": [ + 0.217, + 0.439, + 0.785, + 0.514 + ], + "angle": 0, + "content": "on reading, e.g. (RSA \\(\\geq 0.3\\)), their text-oriented and natural downstream performance also decreases. In other words, the parrot captions can benefit the text-orient downstream tasks while requiring careful data mixing trade-off. We believe that understanding parrot captions is essential for revealing the underlying mechanisms of existing large vision-language systems." + }, + { + "type": "title", + "bbox": [ + 0.218, + 0.535, + 0.507, + 0.55 + ], + "angle": 0, + "content": "7 Conclusion and Discussion" + }, + { + "type": "text", + "bbox": [ + 0.217, + 0.564, + 0.786, + 0.745 + ], + "angle": 0, + "content": "The popularity of vision-language contrastive loss stems from its efficiency and simplicity. However, the analysis and experiments we presented show that the embedded text in images and their parrot captions plant significant text spotting bias due to such contrastive fashions. Firstly, almost \\(30\\%\\) of the captions in the widely used LAION-2B dataset are biased towards parroting the embedded text in images. Secondly, the pre-trained CLIP models have strong preferences for the image-text pair with parrot captions, which achieve higher similarity scores than those without. Finally, using data biasing to parrot captions, we can easily train a CLIP model with a strong text spotting bias. Our work demonstrates the emergency of reviewing the impact of parrot captions in the entire ecosystem of CLIP models. Our future endeavor involves building a bias-aware data curation pipeline and a robust training function to mitigate such issues." + }, + { + "type": "title", + "bbox": [ + 0.218, + 0.766, + 0.402, + 0.784 + ], + "angle": 0, + "content": "Acknowledgements" + }, + { + "type": "text", + "bbox": [ + 0.217, + 0.795, + 0.785, + 0.84 + ], + "angle": 0, + "content": "This project was supported by the National Key R&D Program of China (No. 2022ZD0160101) and the Shanghai Artificial Intelligence Laboratory. Mike Shou does not receive any funding for this work." + } + ], + [ + { + "type": "header", + "bbox": [ + 0.448, + 0.115, + 0.732, + 0.129 + ], + "angle": 0, + "content": "Parrot Captions Teach CLIP to Spot Text" + }, + { + "type": "page_number", + "bbox": [ + 0.769, + 0.117, + 0.785, + 0.127 + ], + "angle": 0, + "content": "15" + }, + { + "type": "title", + "bbox": [ + 0.218, + 0.145, + 0.323, + 0.16 + ], + "angle": 0, + "content": "References" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.177, + 0.785, + 0.218 + ], + "angle": 0, + "content": "1. Agarwal, S., Krueger, G., Clark, J., Radford, A., Kim, J.W., Brundage, M.: Evaluating clip: towards characterization of broader capabilities and downstream implications. arXiv preprint arXiv:2108.02818 (2021)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.219, + 0.785, + 0.245 + ], + "angle": 0, + "content": "2. Alabdulmohsin, I., Wang, X., Steiner, A.P., Goyal, P., D'Amour, A., Zhai, X.: Clip the bias: How useful is balancing data in multimodal learning? In: ICLR (2024)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.247, + 0.785, + 0.286 + ], + "angle": 0, + "content": "3. Ali, J., Kleindessner, M., Wenzel, F., Budhathoki, K., Cevher, V., Russell, C.: Evaluating the fairness of discriminative foundation models in computer vision. In: AIES. pp. 809-833 (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.288, + 0.785, + 0.328 + ], + "angle": 0, + "content": "4. Berg, H., Hall, S.M., Bhalgat, Y., Yang, W., Kirk, H.R., Shtedritski, A., Bain, M.: A prompt array keeps the bias away: Debiasing vision-language models with adversarial learning. arXiv preprint arXiv:2203.11933 (2022)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.33, + 0.785, + 0.37 + ], + "angle": 0, + "content": "5. Biten, A.F., Tito, R., Mafla, A., Gomez, L., Rusinol, M., Valveny, E., Jawahar, C., Karatzas, D.: Scene text visual question answering. In: ICCV. pp. 4291-4301 (2019)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.371, + 0.785, + 0.398 + ], + "angle": 0, + "content": "6. Byeon, M., Park, B., Kim, H., Lee, S., Baek, W., Kim, S.: Coyo-700m: Image-text pair dataset. https://github.com/kakaobrain/coyo-dataset (2022)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.399, + 0.785, + 0.439 + ], + "angle": 0, + "content": "7. Cao, L., Zhang, B., Chen, C., Yang, Y., Du, X., Zhang, W., Lu, Z., Zheng, Y.: Less is more: Removing text-regions improves clip training efficiency and robustness. arXiv preprint arXiv:2305.05095 (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.44, + 0.785, + 0.481 + ], + "angle": 0, + "content": "8. Chen, X., Fang, H., Lin, T.Y., Vedantam, R., Gupta, S., Dolkar, P., Zitnick, C.L.: Microsoft coco captions: Data collection and evaluation server. arXiv preprint arXiv:1504.00325 (2015)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.482, + 0.785, + 0.522 + ], + "angle": 0, + "content": "9. Devlin, J., Chang, M.W., Lee, K., Toutanova, K.: Bert: Pre-training of deep bidirectional transformers for language understanding. arXiv preprint arXiv:1810.04805 (2018)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.523, + 0.785, + 0.577 + ], + "angle": 0, + "content": "10. Dosovitskiy, A., Beyer, L., Kolesnikov, A., Weissenborn, D., Zhai, X., Unterthiner, T., Dehghani, M., Minderer, M., Heigold, G., Gelly, S., et al.: An image is worth 16x16 words: Transformers for image recognition at scale. arXiv preprint arXiv:2010.11929 (2020)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.578, + 0.785, + 0.619 + ], + "angle": 0, + "content": "1. Gadre, S.Y., Ilharco, G., Fang, A., Hayase, J., Smyrnis, G., Nguyen, T., Marten, R., Wortsman, M., Ghosh, D., Zhang, J., et al.: Datacomp: In search of the next generation of multimodal datasets. arXiv preprint arXiv:2304.14108 (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.62, + 0.785, + 0.646 + ], + "angle": 0, + "content": "2. Ganz, R., Nuriel, O., Aberdam, A., Kittenplon, Y., Mazor, S., Litman, R.: Towards models that can see and read. arXiv preprint arXiv:2301.07389 (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.647, + 0.785, + 0.688 + ], + "angle": 0, + "content": "3. Goh, G., Cammarata, N., Voss, C., Carter, S., Petrov, M., Schubert, L., Radford, A., Olah, C.: Multimodal neurons in artificial neural networks. Distill 6(3), e30 (2021)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.689, + 0.785, + 0.73 + ], + "angle": 0, + "content": "4. Goyal, Y., Khot, T., Summers-Stay, D., Batra, D., Parikh, D.: Making the v in vqa matter: Elevating the role of image understanding in visual question answering. In: CVPR. pp. 6904-6913 (2017)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.731, + 0.785, + 0.757 + ], + "angle": 0, + "content": "5. He, K., Zhang, X., Ren, S., Sun, J.: Deep residual learning for image recognition. In: CVPR. pp. 770-778 (2016)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.758, + 0.785, + 0.799 + ], + "angle": 0, + "content": "6. Ilharco, G., Wortsman, M., Wightman, R., Gordon, C., Carlini, N., Taori, R., Dave, A., Shankar, V., Namkoong, H., Miller, J., Hajishirzi, H., Farhadi, A., Schmidt, L.: Openclip (Jul 2021), https://doi.org/10.5281/zenodo.5143773" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.8, + 0.785, + 0.84 + ], + "angle": 0, + "content": "7. Jia, C., Yang, Y., Xia, Y., Chen, Y.T., Parekh, Z., Pham, H., Le, Q., Sung, Y.H., Li, Z., Duerig, T.: Scaling up visual and vision-language representation learning with noisy text supervision. In: ICML. pp. 4904-4916. PMLR (2021)" + }, + { + "type": "list", + "bbox": [ + 0.226, + 0.177, + 0.785, + 0.84 + ], + "angle": 0, + "content": null + } + ], + [ + { + "type": "page_number", + "bbox": [ + 0.218, + 0.116, + 0.236, + 0.127 + ], + "angle": 0, + "content": "16" + }, + { + "type": "header", + "bbox": [ + 0.272, + 0.115, + 0.356, + 0.127 + ], + "angle": 0, + "content": "Y. Lin et al." + }, + { + "type": "ref_text", + "bbox": [ + 0.218, + 0.147, + 0.785, + 0.175 + ], + "angle": 0, + "content": "18. Johnson, J., Douze, M., Jégou, H.: Billion-scale similarity search with GPUs. IEEE Transactions on Big Data 7(3), 535-547 (2019)" + }, + { + "type": "ref_text", + "bbox": [ + 0.218, + 0.176, + 0.785, + 0.217 + ], + "angle": 0, + "content": "19. Lemesle, Y., Sawayama, M., Valle-Perez, G., Adolphe, M., Sauzeon, H., Oudeyer, P.Y.: Language-biased image classification: evaluation based on semantic representations. arXiv preprint arXiv:2201.11014 (2022)" + }, + { + "type": "ref_text", + "bbox": [ + 0.218, + 0.217, + 0.785, + 0.242 + ], + "angle": 0, + "content": "20. Li, B., Weinberger, K.Q., Belongie, S., Koltun, V., Ranftl, R.: Language-driven semantic segmentation. In: ICLR (2022)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.243, + 0.785, + 0.284 + ], + "angle": 0, + "content": "21. Li, J., Li, D., Xiong, C., Hoi, S.: Blip: Bootstrapping language-image pre-training for unified vision-language understanding and generation. In: ICML. pp. 12888-12900. PMLR (2022)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.285, + 0.785, + 0.324 + ], + "angle": 0, + "content": "22. Li, J., Selvaraju, R., Gotmare, A., Joty, S., Xiong, C., Hoi, S.C.H.: Align before fuse: Vision and language representation learning with momentum distillation. NeurIPS 34, 9694-9705 (2021)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.325, + 0.785, + 0.352 + ], + "angle": 0, + "content": "23. Loshchilov, I., Hutter, F.: Decoupled weight decay regularization. arXiv preprint arXiv:1711.05101 (2017)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.352, + 0.785, + 0.392 + ], + "angle": 0, + "content": "24. Maini, P., Goyal, S., Lipton, Z.C., Kolter, J.Z., Raghunathan, A.: T-mars: Improving visual representations by circumventing text feature learning. arXiv preprint arXiv:2307.03132 (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.393, + 0.785, + 0.419 + ], + "angle": 0, + "content": "25. Materzyńska, J., Torralba, A., Bau, D.: Disentangling visual and written concepts in clip. In: CVPR. pp. 16410-16419 (2022)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.42, + 0.785, + 0.46 + ], + "angle": 0, + "content": "26. Nichol, A., Dhariwal, P., Ramesh, A., Shyam, P., Mishkin, P., McGrew, B., Sutskever, I., Chen, M.: Glide: Towards photorealistic image generation and editing with text-guided diffusion models. arXiv preprint arXiv:2112.10741 (2021)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.46, + 0.785, + 0.5 + ], + "angle": 0, + "content": "27. Pham, H., Dai, Z., Ghiasi, G., Kawaguchi, K., Liu, H., Yu, A.W., Yu, J., Chen, Y.T., Luong, M.T., Wu, Y., et al.: Combined scaling for zero-shot transfer learning. Neurocomputing 555, 126658 (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.501, + 0.785, + 0.542 + ], + "angle": 0, + "content": "28. Radenovic, F., Dubey, A., Kadian, A., Mihaylov, T., Vandenhende, S., Patel, Y., Wen, Y., Ramanathan, V., Mahajan, D.: Filtering, distillation, and hard negatives for vision-language pre-training. In: CVPR. pp. 6967-6977 (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.543, + 0.785, + 0.583 + ], + "angle": 0, + "content": "29. Radford, A., Kim, J.W., Hallacy, C., Ramesh, A., Goh, G., Agarwal, S., Sastry, G., Askell, A., Mishkin, P., Clark, J., et al.: Learning transferable visual models from natural language supervision. In: ICML. pp. 8748-8763. PMLR (2021)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.583, + 0.785, + 0.609 + ], + "angle": 0, + "content": "30. Rombach, R., Blattmann, A., Lorenz, D., Esser, P., Ommer, B.: High-resolution image synthesis with latent diffusion models. In: CVPR. pp. 10684-10695 (2022)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.61, + 0.785, + 0.664 + ], + "angle": 0, + "content": "31. Schuhmann, C., Beaumont, R., Vencu, R., Gordon, C., Wightman, R., Cherti, M., Coombes, T., Katta, A., Mullis, C., Wortsman, M., et al.: Laion-5b: An open large-scale dataset for training next generation image-text models. NeurIPS 35, 25278-25294 (2022)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.664, + 0.785, + 0.705 + ], + "angle": 0, + "content": "32. Schuhmann, C., Vencu, R., Beaumont, R., Kaczmarczyk, R., Mullis, C., Katta, A., Coombes, T., Jitsev, J., Komatsuzaki, A.: Laion-400m: Open dataset of clip-filtered 400 million image-text pairs. arXiv preprint arXiv:2111.02114 (2021)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.706, + 0.785, + 0.732 + ], + "angle": 0, + "content": "33. Shi, C., Yang, S.: Logoprompt: Synthetic text images can be good visual prompts for vision-language models. In: ICCV. pp. 2932-2941 (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.733, + 0.785, + 0.759 + ], + "angle": 0, + "content": "34. Shtedritski, A., Rupprecht, C., Vedaldi, A.: What does clip know about a red circle? visual prompt engineering for vlms. arXiv preprint arXiv:2304.06712 (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.759, + 0.785, + 0.786 + ], + "angle": 0, + "content": "35. Sidorov, O., Hu, R., Rohrbach, M., Singh, A.: Textcaps: a dataset for image captioning with reading comprehension. In: ECCV. pp. 742-758. Springer (2020)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.786, + 0.785, + 0.813 + ], + "angle": 0, + "content": "36. Singh, A., Natarajan, V., Shah, M., Jiang, Y., Chen, X., Batra, D., Parikh, D., Rohrbach, M.: Towards vqa models that can read. In: CVPR. pp. 8317-8326 (2019)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.814, + 0.785, + 0.84 + ], + "angle": 0, + "content": "37. Tanjin, M.M., Singh, K.K., Kafle, K., Sinha, R., Cottrell, G.W.: Discovering and mitigating biases in clip-based image editing. In: WACV. pp. 2984-2993 (2024)" + }, + { + "type": "list", + "bbox": [ + 0.217, + 0.147, + 0.785, + 0.84 + ], + "angle": 0, + "content": null + } + ], + [ + { + "type": "header", + "bbox": [ + 0.448, + 0.115, + 0.732, + 0.129 + ], + "angle": 0, + "content": "Parrot Captions Teach CLIP to Spot Text" + }, + { + "type": "page_number", + "bbox": [ + 0.768, + 0.116, + 0.786, + 0.127 + ], + "angle": 0, + "content": "17" + }, + { + "type": "ref_text", + "bbox": [ + 0.214, + 0.147, + 0.785, + 0.175 + ], + "angle": 0, + "content": "38. Telea, A.: An image inpainting technique based on the fast marching method. Journal of graphics tools 9(1), 23-34 (2004)" + }, + { + "type": "ref_text", + "bbox": [ + 0.214, + 0.175, + 0.785, + 0.203 + ], + "angle": 0, + "content": "39. Tschannen, M., Mustafa, B., Houlsby, N.: Clippo: Image-and-language understanding from pixels only. In: CVPR. pp. 11006-11017 (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.214, + 0.203, + 0.787, + 0.245 + ], + "angle": 0, + "content": "40. Wang, H., Zhan, Y., Liu, L., Ding, L., Yu, J.: Balanced similarity with auxiliary prompts: Towards alleviating text-to-image retrieval bias for clip in zero-shot learning. arXiv preprint arXiv:2402.18400 (2024)" + }, + { + "type": "ref_text", + "bbox": [ + 0.214, + 0.245, + 0.785, + 0.273 + ], + "angle": 0, + "content": "41. Wang, J., Liu, Y., Wang, X.E.: Are gender-neutral queries really gender-neutral? mitigating gender bias in image search. arXiv preprint arXiv:2109.05433 (2021)" + }, + { + "type": "ref_text", + "bbox": [ + 0.214, + 0.273, + 0.785, + 0.313 + ], + "angle": 0, + "content": "42. Wei, J., Tay, Y., Bommasani, R., Raffel, C., Zoph, B., Borgeaud, S., Yogatama, D., Bosma, M., Zhou, D., Metzler, D., et al.: Emergent abilities of large language models. arXiv preprint arXiv:2206.07682 (2022)" + }, + { + "type": "ref_text", + "bbox": [ + 0.214, + 0.314, + 0.785, + 0.342 + ], + "angle": 0, + "content": "43. Xu, Y., Xu, Z., Chai, W., Zhao, Z., Song, E., Wang, G.: Devil in the number: Towards robust multi-modality data filter. arXiv preprint arXiv:2309.13770 (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.214, + 0.342, + 0.785, + 0.383 + ], + "angle": 0, + "content": "44. Yao, Y., Zhang, A., Zhang, Z., Liu, Z., Chua, T.S., Sun, M.: Cpt: Colorful prompt tuning for pre-trained vision-language models. arXiv preprint arXiv:2109.11797 (2021)" + }, + { + "type": "ref_text", + "bbox": [ + 0.214, + 0.383, + 0.785, + 0.424 + ], + "angle": 0, + "content": "45. Ye, M., Zhang, J., Zhao, S., Liu, J., Liu, T., Du, B., Tao, D.: Deepsolo: Let transformer decoder with explicit points solo for text spotting. In: CVPR. pp. 19348-19357 (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.214, + 0.424, + 0.785, + 0.467 + ], + "angle": 0, + "content": "46. Yuksekgonul, M., Bianchi, F., Kalluri, P., Jurafsky, D., Zou, J.: When and why vision-language models behave like bags-of-words, and what to do about it? In: ICLR (2022)" + }, + { + "type": "ref_text", + "bbox": [ + 0.214, + 0.467, + 0.785, + 0.508 + ], + "angle": 0, + "content": "47. Zhai, X., Wang, X., Mustafa, B., Steiner, A., Keysers, D., Kolesnikov, A., Beyer, L.: Lit: Zero-shot transfer with locked-image text tuning. In: CVPR. pp. 18123-18133 (2022)" + }, + { + "type": "ref_text", + "bbox": [ + 0.214, + 0.508, + 0.785, + 0.536 + ], + "angle": 0, + "content": "48. Zhang, Q., Xu, Y., Zhang, J., Tao, D.: Vitaev2: Vision transformer advanced by exploring inductive bias for image recognition and beyond. IJCV pp. 1-22 (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.214, + 0.536, + 0.785, + 0.564 + ], + "angle": 0, + "content": "49. Zhou, K., Yang, J., Loy, C.C., Liu, Z.: Learning to prompt for vision-language models. IJCV 130(9), 2337-2348 (2022)" + }, + { + "type": "list", + "bbox": [ + 0.214, + 0.147, + 0.787, + 0.564 + ], + "angle": 0, + "content": null + } + ] +] \ No newline at end of file diff --git a/2024/Parrot Captions Teach CLIP to Spot Text/b85a274a-d711-419a-aeba-49fd9715a806_origin.pdf b/2024/Parrot Captions Teach CLIP to Spot Text/b85a274a-d711-419a-aeba-49fd9715a806_origin.pdf new file mode 100644 index 0000000000000000000000000000000000000000..51a9ddb07a619a391c33b83a25c079b37adac136 --- /dev/null +++ b/2024/Parrot Captions Teach CLIP to Spot Text/b85a274a-d711-419a-aeba-49fd9715a806_origin.pdf @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:28535a24950309fa3c9f55a5d1cdf162deddbdc8a9d65a90325e33e0abb8edb0 +size 7340947 diff --git a/2024/Parrot Captions Teach CLIP to Spot Text/full.md b/2024/Parrot Captions Teach CLIP to Spot Text/full.md new file mode 100644 index 0000000000000000000000000000000000000000..a208d863980782f67084d84273399a34084d4184 --- /dev/null +++ b/2024/Parrot Captions Teach CLIP to Spot Text/full.md @@ -0,0 +1,307 @@ +# Parrot Captions Teach CLIP to Spot Text + +Yiqi Lin $^{1,2*}$ , Conghui He $^{1*}$ , Alex Jinpeng Wang $^{2*}$ , Bin Wang $^{1*}$ , Weijia Li $^{3}$ , and Mike Zheng Shou $^{2}$ + +1 Shanghai Artificial Intelligence Laboratory + +2 Show Lab, National University of Singapore + +3 Sun Yat-Sen University + +Abstract. Despite CLIP [29] being the foundation model in numerous vision-language applications, CLIP suffers from a severe text spotting bias. Such bias causes CLIP models to 'Parrot' the visual text embedded within images while disregarding the authentic visual semantics. We uncover that in the most popular image-text dataset LAION-2B [31], the captions also densely parrot (spell) the text embedded in images. Our analysis shows that around $50\%$ of images are embedded with visual text content and around $30\%$ of captions words are concurrently embedded in the visual content. Based on such observation, we thoroughly inspect the different released versions of CLIP models and verify that the visual text is a dominant factor in measuring the LAION-style image-text similarity for these models. To examine whether these parrot captions shape the text spotting bias, we train a series of CLIP models with LAION subsets curated by different parrot-caption-oriented criteria. We show that training with parrot captions easily shapes such bias but harms the expected visual-language representation learning in CLIP models across various vision-language downstream tasks. This suggests that it is urgent to revisit either the design of CLIP-like models or the existing image-text dataset curation pipeline built on CLIP score filtering. Project page: https://tinyq17.github.io/CLIP-Parrot-Bias/ + +Keywords: Image-Text Dataset $\cdot$ Text Spotting Bias + +# 1 Introduction + +Recently, contrastive learning models [17, 29, 31] pre-trained with large-scale image-text pair data has led to numerous vision-language modeling task breakthroughs. Due to its efficiency and simplicity, the pioneering work CLIP [29] now serves as a foundation model in various applications [20, 26, 30, 49]. However, several works [4,13] have shown that CLIP models have perpetuating biases towards visual text [19,25], color [34,44], gender [41], etc. In this paper, we focus on probing the visual text bias in CLIP, i.e., the capacity of spotting text in images. Most of the previous cues [25, 29, 34] attribute the sources of biases to the noisy pre-training data. Therefore, we begin by taking a close look at the most popular dataset, LAION-2B [31]. + +![](images/66c472be5cef6c6f9d3fbde23da1f2ee370ac6db3009142b22eee8358e6f11e4.jpg) +(b) Image Paired Captions +Fig. 1: In LAION-2B [31], image-text pairs with the Top-5% highest similarity score are most dominant by visual text! These samples have dense concurrent text appearing in captions and images (text form in pixels). We refer to their captions as Parrot Captions as they raise a question: Dose CLIP Simply Parroting Text in Images for Vision-Language Alignment? The concurrent text is spotted by the OCR model and highlighted with color in image-text pairs. (Best view in color) + +Considering the massive scale of the image-text data, it is non-trivial to assess the bias simply with a rough estimation. To this end, we first do image clustering on the whole dataset and rank each cluster by CLIP scores to analyze the most preferred types of image-text pairs under CLIP score measurement. As shown in Fig. 1, we surprisingly observe that a decent number of samples with top CLIP scores have dense concurrent text appearing in the captions and the images in the form of pixels. These samples break the assumption that CLIP models leverage text supervision to align the visual and language concepts. We refer to these captions as Parrot Captions as they provide another shortcut to achieve the same goal by teaching CLIP to do text spotting even without perceiving the actual visual concepts. To understand the underlying impact, we analyze the parrot captions from three perspectives: dataset, widely used released models, and model training. The results lead to three key findings: + +Firstly, captions in LAION-2B have a significant bias towards describing visual text content embedded in the images. We provide thorough profiling using off-the-self text spotting models on the LAION-2B dataset and show that over $50\%$ of the images are embedded with visual text content. Moreover, by examining the spotted text content and the paired caption in the subset that images embedded with text, we find that over $60\%$ of the captions at least have one concurrent word and reach at least around $30\%$ words overlap + +between the caption and spotted text from images. This finding suggests that the basic assumption of image-text semantic alignment in CLIP does not fully stand its ground when training with LAION-style data. + +Secondly, released CLIP models have strong text spotting bias almost in every style of web images, resulting in CLIP-filtering datasets inherently biased towards visual text-dominant data. We investigate OpenAI released CLIP model's behaviors in the LAION-2B dataset by examining the difference between alignment scores before and after text removal. The results show that CLIP model predictions densely correlate the visual text embedded in images with their parrot captions. Next, we further study the preference of the text spotting capacity on text content in CLIP and OpenCLIP models. Note that CLIP is trained on WIT-400M, while OpenCLIP uses the LAION-2B dataset. Therefore, we use synthetic images embedded with specific rendered text to avoid overfitting in OpenCLIP models. Our analysis shows that OpenCLIP is more biased toward text spotting than CLIP. We believe that the parrot caption plays a lurking role in training these released CLIP models and is the source of text spotting capacity instead of emergence behavior [42]. + +Thirdly, CLIP models can easily learn text spotting capacity from parrot captions while failing to connect the vision-language semantics, just like a text spotting parrot. We sample different LAION-2B subsets curated by text-oriented criteria, including the embedded text ratio, the concurrent word ratios, and the relative CLIP score from text removal to train CLIP models under the same setting. The results show that using parrot captions data, CLIP model can learn strong text spotting capacity but lose most of the zero-shot generalization ability on image-text downstream tasks. Moreover, we also observe similar behaviors on various downstream tasks such VQA, captioning, and retrieval on BLIP [21] models. Lastly, we argue that the existing data curation pipeline built on CLIP score and the contrastive fashion urgently needs to be re-examined by considering such hidden parrot captions. + +# 2 Related Work + +# 2.1 Contrastive Vision-Language Pre-training + +Modeling vision and language by aligning the embedding similarity between paired image-text data [17, 29, 31] has shown great potential for transferable to downstream vision-language tasks. The pre-training techniques mainly contain the vision encoder [10, 15] for image embedding encoding, text encoder [9] for text embedding modeling, and cross-modal contrastive learning [17, 22, 29, 47] for learning a joint embedding space of vision and language. The pioneering work CLIP [29] leverages 400 million noisy image-text pairs to learn transferable visual representation from text supervision and show impressive zero-shot performance for various vision-language tasks. Following CLIP, several vision-language models such as ALIGN [17], BASIC [27], and Open-CLIP [31] are proposed, and CLIP models have been replicated on various datasets including WIT [29], LAION [31], COYO [6], and DataComp [11]. We mainly profile + +the LAION-2B [31] dataset due to its large scale and wide usage [26, 30] and two versions of pre-trained models, CLIP and OpenCLIP. Note that the 2 billion image-text pairs in the LAION-2B dataset are filtered by OpenAI released CLIP models, making OpenCLIP connect to CLIP closely. + +# 2.2 Studying of CLIP Behaviors + +Despite the strong zero-shot and transferable performance of CLIP, the perpetuating biases [1, 3, 13, 19, 41, 46] in CLIP are still not well investigated due to its large-scale noisy training data. Much research [2, 25, 34, 37, 40, 43, 44] focuses on revealing or enhancing the downstream performance with discovered bias in CLIP. For example, colorful masks [44] or red circles [34] applied to images can improve the zero-shot performance on visual localization tasks. In studying visual text content bias, [13] shows the multimodal neurons of CLIP not only respond to visual content and the visual text embedded in the image. Another work [19] shows that image recognition in CLIP can be strongly dominated by the visual text embedded in the image. To disentangle such bias, [25] attempts to separate the text spotting representation in pre-trained CLIP by training representation projection. Meanwhile, LoGoPrompt [33] enhances the classification performance by utilizing the visual text content as auxiliary prompts as input. Also, CLIPPO [39] shows that directly aligning the image and synthetic images print with the captions can perform similarly to CLIP without a text encoder. + +# 2.3 Data Curation with Text Removal + +Due to the successful practice of data curation in LAION datasets [31, 32] on scaling up the image-text datasets, searching advanced selection strategy to improve the data quality from common crawl data pool gains a growing interest [11]. Recently, several works [7, 24, 28] suggest that introducing text-related filtering methods improves the pre-training dataset quality. In DiHT [28], the data curation steps include filtering out the image-text pairs with high OCR confidence and matching text ratio. Moreover, [7, 24] mainly focus on studying the importance of filtering out the text-dominate images utilizing OCR models to improve pre-training dataset quality. Maini et al. [24] also draw the observation that $40\%$ of LAION's image text is highly correlated with the caption, but only performing a small pilot study on 500 samples with manual judgment. Differently, this paper makes the first attempt to systemically reveal the source of text spotting capacity in CLIP is the data bias and the consequences of such bias in existing widely used datasets and pre-trained models. + +# 3 Terminology + +The data processing on images in the following sections mainly covers clustering, text spotting (OCR), and text inpainting. Firstly, we cluster all images based on feature similarity. For each image-text pair, we then use the pre-trained text + +![](images/ec43366ccf9e360bbbeaf331d6a94f2609b1b3c7650bc29f046b94e643520bcc.jpg) +Text Removal + +![](images/c0206ad0f5983bda345a6031654922588a01b6e22fdca1f55da534e9e29fad10.jpg) +All-Emb + +![](images/22adb419c388a8258eee75c2678728ca559f30dec4bdfa1c39b9bda2173a3117.jpg) +Co-Emb + +![](images/0229e13685709ee8331372d45a1963ae8fec27b5590c2be688da4129a8e0944e.jpg) +Random + +![](images/7ecc161b44339da337e15ee9bd347b579cb0936d5353013cc1d2e3b02e25297a.jpg) +Syn-Emb. + +![](images/d576b18a2f5bfaf7df2e147fd0db4c0e19f22a0dc0032f50c36fd2cb769e031d.jpg) +Caption + +![](images/f8c8a090e256890a569b602e432cd85ab2f52a2c877899f1c3f867bf434d292f.jpg) +All-Emb. +Text Mask +Fig. 2: Visualization of defined terminologies. Co-Emb. Text is marked in the caption. + +![](images/a25504c95347df076b368e630762b307f8a83fcd1ebca7e8cd7975ba69227ef6.jpg) +Text Removal +Co-Emb. +Text Mask + +![](images/73bedcf722a50c5c75f344f0cb05da4444d8294f2faa31b525a9648847fe68e4.jpg) +Inpainted +Random +Sampled Mask + +![](images/c8d01481022bbad859c05e185888241888398dd5b3f84e528b9f563ccafb5dba.jpg) +Text Template +Syn-Emb. +Text Template + +# Algorithm 1 Pseudocode of Detecting Co-Emb. Text (Rate) + +caption: captions from LAION-2B dataset. + +ocr_text: text spotted by OCR model + +cap_words,OCR_words $=$ set(capTION.split()，set(ocr_text.split())) + +co_emb_text = intersection(cap_words,OCR_words) + +co_emb_text_rate = len(co_emb_text) / len(cap_words) + +spotting model to detect and recognize the text print in image pixels. The mask images in Fig. 2 are the spotted text area. Next, we match the spotted text with the caption using Algorithm 1 to obtain the concurrent words and their ratio in captions. Lastly, we use inpainting to remove the text from the image for CLIPs' pattern ablation. To avoid confusion, we define these concepts as follows, + +- Embedded Text: text spotted by OCR models from the images. To study the correlation of embedded text with captions, we define different kinds of embedded text as, 1) All-Emb. Text: all the text is spotted from an image; 2) Co-Emb. Text: spotted text concurrently appears in the image's corresponding captions; 3) Syn-Emb. Text: synthetic text rendered in an image with a fixed font and a blank background. Fig. 2 shows examples of spotted embedded text by binary mask and the rendering synthetic text. +- Co-Emb. Text Rate (CoTR): the word set IoU of Co-Emb. text and captions (See Algorithm 1). +- Image w/ or w/o Embedded Text: spotted text results of a given image are none-empty or empty. +- Text Removal Image: do inpainting in the specific spotted text area (All-Emb., Co-Emb., or Random). Random is implemented by sampling other image's text areas. For the different inpainting results, see Fig. 2. +- Relative Scores (RSA/RSC): the difference in CLIP score between images modified by different inpainting operations while keeping the same captions. RSA and RSC are the short for the relative scores before and after removing All-Emb. text and Co-Emb. text. +- Image Clusters: image partitions based on K-Means. + +Table 1: Overall parrot captions statistic. More than $50\%$ of images are embedded with text, and $30\%$ of caption words are printed in images! + +
Number of Total Images1,985,284,122
Number of Images w/ Emb. Text1,083,896,427
Number of Images w/ Co-Emb. Text663,600,432
Co-Emb. Text Rate (in Total)15.42%
Co-Emb. Text Rate (in Images w/ Emb. Text)28.24%
Fuzzy Co-Emb. Text Rate (in Total)30.46%
Fuzzy Co-Emb. Text Rate (in Images w/ Emb. Text)55.79%
+ +- CLIP and OpenCLIP: CLIP models are trained on WIT-400M [29] and LAION-2B [31] dataset. +- N-gram Vocabulary (Vocab): the set of all contiguous N word sequences extracted from a text corpus, such as the collection of all captions. + +# 4 Profiling LAION-2B Data + +To better profile the image-text pair data on a billion scale, we first cluster all the images based on CLIP features into 4,000 clusters and sort each cluster with CLIP scores. After obtaining all the cluster labels, we use the SOTA text spotting model [45] to get the visual text content on all the collected images. Finally, we aggregate all the model-predicted results and compare them with their corresponding captions to bring out our observations. + +# 4.1 Implementation Details + +Clustering with CLIP Features: We first train K-Means (implemented by Faiss [18]) on the LAION-400M [32] subset using ViT-B-32 [10] CLIP features to speed up the clustering process. Due to the large memory consumption, we reduce the feature dimensions from 512 to 256 using PCA. Then, we partition the whole dataset using trained K-Means with the same feature extraction pipeline. + +Text Spotting and Matching: To detect and recognize text across various scenes, we adopt DeepSolo [45] as our text spotting model and use the pre-trained checkpoints with the ViTAEv2-S [48] backbone in default setting. The output format of the text spotting model is a sequence of polygons of text location and their recognized characters. Despite its strong performance, we empirically find that DeepSolo can not handle the crowd scenes well (with more than 100 separate words) which is only a small proportion of the dataset ( $\sim 2\%$ ). To identify the correlation between the spotted text and captions, we use Algorithm 1 to calculate the Co-Emb. text rate in each image-text pair. Considering the imperfect text spotting predictions might miss or misspell words, we also use Levenshtein distance to calculate the fuzzing similarity and reported in Tab. 1. + +# 4.2 Statistic and Observations from LAION-2B + +The overall statistics of the 2 billion image-text pairs are reported in Tab. 1. In summary, the images embedded with visual text content reach a surprisingly + +![](images/2736101e6b6790673d162463d80f3b3fdd36658ea8255251ff60763cdfce8256.jpg) + +![](images/fe2ca1d48b0669c1f8be8c553c27eb3b8e8658a10422d0f40e5ef845c3115eed.jpg) +(a) The ratio of different OCR-oriented data types in LAION-2B clusters. +(b) Top CLIP score samples visualization from 50 clusters with ratio over $80\%$ . +Fig. 3: (a): Based on the OCR prediction results, the image-text pairs are divided into three types: image without visual embedded text content; the spotted text from the image has no concurrent text with the caption; the spotted text at least share one concurrent word with the caption. The clusters are merged from 4000 into 100 for a better view. (b): In the clusters with high ratio, the top CLIP score samples contain various text sources, such as posters, book covers, advertisements, and even slides. + +high proportion of $54.60\%$ in the investigated data. Around $15\%$ of words in the dataset captions are Co-Emb. text, and the proportion of Co-Emb. text can further reach $30\%$ when considering the fuzzy matching results of the spotted text and captions. This suggests that CLIP models trained on these data might lead to a high bias toward text spotting. To better visualize the data distribution, we provide cluster-specific statics results and top CLIP score samples of text-dominated clusters in Fig. 3. We divide all images into 100 clusters based on visual similarity and visualize them according to the OCR results. Every cluster contains more or less images embedded with text. Combined with sample visualization, we observe that in the LAION collected data, the parrot captions cover various scenes. In the subsets of images embedded with text, around $60\%$ of captions at least precisely parrot one concurrent word (Co-Emb. Text Rate $>0$ ) appearing in the image. It suggests that the data collection pipeline of LAION [31] has a strong bias to introduce parrot captions from web data. + +To better understand Co-Emb. Text, we provide a more thorough analysis of the word counting and text size of parrot captions. As shown in Fig. 4a, the results show that a large proportion of the Co-Emb. Text only takes a few words. However, we also find a large number of captions that are almost full parrot captions (see areas around the heatmap diagonal). Next, in Fig. 4b and Fig. 4c, we investigate the correlation between the size of concurrent words box in the image and CLIP score. The results show that the large text size does not usually lead to a higher score; meanwhile, the small text size can also dominate the score as the score can be significantly different after removing them. (Details + +![](images/d848e7a6b526e361e58fd20eacc31821b0c9601384979e11031caf73e76bad60.jpg) +Fig. 4: (a): The number of caption words and associated spotted concurrent words based on precise word matching. (b): Distribution of total box area of concurrent words in the image and its CLIP score. (c): Distribution of total box area of concurrent words and its relative CLIP score before and after removing them from the image. (d): Distribution of text size of the single concurrent word and other spotted word. + +of text removal described in Sec. 5.1). One possible reason is the text content and input resolution may matter more for CLIP. Moreover, we discover that the larger text is more likely to be parroted in the captions, as shown in Fig. 4d. + +# 5 Inspecting Pre-Trained CLIP Models + +It is important to note that the LAION-2B dataset collection pipeline uses CLIP score from OpenAI's model to filter out the image-text pair below 0.28. Therefore, we inspect these two released CLIP models [29,31] to answer better why LAION data contains such a high proportion of parrot captions. Specifically, OpenAI's CLIP model is trained on the WIT dataset (out-of-domain model), and OpenCLIP is trained on LAION-2B (in-domain model). We first study whether the embedded text is the key factor in CLIP filtering by ablating the embedded text using text inpainting. Moreover, we further investigate whether the text spotting capacity prefers specific text content by examining synthetic images with Syn-Emb. text. + +# 5.1 Ablation of Embedded Text Removal + +Text Removal via Inpainting: Given the OCR predicted results, we use the fast marching method [38] to inpaint the area of the spotted text polygons. We generate two versions of text removal results for each image with embedded text, i.e., All-Emb. text removal and Co-Emb. text removal, as the parrot caption prediction is imperfect due to the limitation of OCR models. We also generate random inpainting images with randomly sampled spotted text polygons from other images to ablate the distribution shift caused by image inpainting. Examples of the spotted text masks and inpainting results are shown in Fig. 2. + +Results: Based on the OCR predicted results and text inpainting operations, we can obtain six types of LAION images, including $\bullet$ ): images without any embedded text (OCR results are empty); $\bullet$ ): images with any embedded text + +![](images/a29d6bf2527dbbbbe5d619be17b238ed6c0779b885764f5faf754d369489b5dc.jpg) +Fig. 5: Left: Mean CLIP scores of image-text pairs with different text removal operations depicted in Sec. 5.1, and grouped by cluster the same as Fig. 3. Right: Overall relative CLIP score distribution by comparing different text removal operations. + +Table 2: Mean CLIP score of different setups of text removal. + +
SetupAverage CLIP Score
Raw Images w/o Emb. Text0.3223 ± 0.0078
Raw Images w/ Emb. Text0.3358 ± 0.0094
× All-Emb. Random Inpainted0.3260 ± 0.0057
× All-Emb. Text Removal0.2974 ± 0.0197
Co-Emb. Random Inpainted0.3341 ± 0.0051
Co-Emb. Text Removal0.2993 ± 0.0146
+ +(OCR results are none-empty); $\times$ ): images with random inpainting by other image's All-Emb. text area; $\times$ ): images removed All-Emb. text (Inpaint all the areas of OCR predicted text); ■): images randomly inpainted by other image's Co-Emb. text area, and ■): images removed Co-Emb. text (Inpaint the areas of concurrent text in OCR predicted text and captions). Then, we calculate CLIP scores of all the groups of images and their paired captions using OpenAI released CLIP model (ViT-B-32). Fig. 5 reports the mean scores of different types of images in each cluster and raises four observations as follows: I). The images embedded with text achieve higher CLIP scores in most clusters than those without embedded text; II). CLIP scores significantly drop once we remove the text from the images compared to its random inpainting baseline. It indicates that the parrot captions correlate highly with CLIP score measurement; III). Not all the samples are dominated by the embedded text, as some samples achieve higher scores after removing text, indicating the embedded text can also be a distractor; IV). Most of the relative CLIP scores ( $S(\square) - S(\times)$ ) between images removed Co-Emb. text and All-Emb. text are positive, as shown in the right of Fig. 5. The straightforward reason is the images lose more visual information due to the larger in-painting area. Another possible reason is the imperfect text spotting prediction or the corner cases in the matching algorithm leaking parts of the concurrent text in images. + +Discussion: Due to the text removal, the image distribution may shift from the CLIP training set. Therefore, we provide two random inpainting baselines to examine the effect of distribution shift. In Tab. 2, we report the mean scores + +![](images/b2bf5fc65d8a33143afa94208fd31f2580b7e440c2201ac5e439a8d6211cc792.jpg) +(a) Statistic of 1-gram vocabularies + +![](images/c51adc3c2e4cc81578b7a958b2f93d739506d8eb4c6238d7eba7dcd6e3abab14.jpg) + +![](images/eb6efd676f718ce07bca9268be2594cf7fe0ba293c6127178a6a55beab9d8a57.jpg) +(c) Grouped by Caption 1-gram length + +![](images/3bd8ffa9f9641b7a2f3363fedd8e97c6c8b90db62725fb5f5f55c10dba180ed4.jpg) + +![](images/6b01f1eb543c6ae006cbca61d13aee87e3471d2c1b3c78d85340378d0c7bc211.jpg) +(b) Grouped by 1-gram frequency intervals +(d) Grouped by Caption 2-gram length +Fig.6: OpenCLIP more bias towards text spotting than CLIP model. Grouped score distributions of prompting CLIP and OpenCLIP models with N-gram Syn-Emb. text and synthetic images for model preference investigation. + +of different setups. Results show that the random baselines are very close to the raw image baseline, indicating that CLIP model is robust to the distribution shift caused by information loss in inpainted regions. + +# 5.2 Prompting with Syn-Emb. Text + +Generating Synthetic Images from N-gram Vocabulary: To investigate CLIP models' text spotting preference, we adopt a similar strategy in [25] to use synthetic images to embed specific text content by rendering text in a blank background. For each text, we use four fore-background style rendering templates (black-white, black-grey, white-grey, and white-black), as shown in Fig. 2. Different from the uniformly sampling letters in [25], we generate the text content from the N-gram vocabulary built from captions and Co-Emb. text to study the text spotting pattern. We only select the top frequent 400,000 grams for each vocabulary. The statistics of 1-gram vocabulary are shown in Fig. 6a, which is a long-tail distribution. Next, we calculate the synthetic images and the rendered text similarity on released ViT-B-32 CLIP and OpenCLIP models. + +Results: Firstly, we examine whether CLIP models prefer recognizing more commonly seen words (with high frequency in vocabulary). Therefore, we group the 1-gram results based on their frequency interval in the whole vocabulary, as shown in Fig. 6b. The results show that OpenCLIP model clearly has a stronger text spotting capacity than CLIP, i.e., more biased towards text spotting. We also observe that all CLIP models are more sensitive to the vocabulary built + +from the concurrent words. Interestingly, both CLIP and OpenCLIP models have slightly higher scores on the less frequent grams. Secondly, considering the long-tail grams might contain more characters, we further group the 1-gram and 2-gram results based on their text length in Fig. 6c and Fig. 6d. Note that the Co-Emb. text is not regularly arranged in the images, making it hard to extract continuous word sequences. Results show that all the models are better at spotting the longer words, possibly due to the tokenizer used in the text encoder, making them more discriminative. Meanwhile, in the groups of 2-gram samples, the scores gradually drop when spotting the highly long text, indicating that the spotting capacity of CLIP models is possibly built on word-by-word. + +# 6 Training on Emb. Text Curated Data + +Next, we dive deeper into training CLIP and BLIP [21] models on different Emb. text curated subsets and studying various downstream task behaviors. + +Implementation Details: For CLIP, we use the open-source software OpenCLIP [16] for all CLIP model training. Our experiments are conducted on both ViT-B [10] and RN50 [15]. We use 4,096 batch size for 3M and 8,192 for 12M scale subsets. Other settings remain the same as [31]. For BLIP, we mainly conduct on 3M scale subsets with ViT-B [10]. The BLIP models are pre-trained for 10 epochs with an AdamW [23] optimizer. For downstream tasks, we finetune 10 epochs for VQA and 5 epochs for captioning and retrieval. + +Evaluation: For CLIP, we follow the DataComp benchmark [11] using 38 zero-shot classification and retrieval tasks as evaluation. We report the average performance (Avg.) of the DataComp benchmark and two subset track performances, ImageNet (IN) and Retrieval (Ret.). To evaluate the text spotting capacity, we use the synthetic benchmark illustrated in Sec. 5.2 and a real-world benchmark sampled from LAION-2B as the validation set. In the synthetic benchmark, we calculate the similarity of all the 1-gram synthetic image-text pairs from caption vocabulary and report all the trained model results in Fig. 7. For the real-world benchmark, we sample 1M image-text pairs without any embedded text and 1M samples dominated by the parrot caption (RSC $\geq 0.2$ ). Fig. 8 aggregates the mean scores of the 2M evaluation set and also reports the mean scores of applying text removal on the 2M evaluation set results. For BLIP, inspired by [12], we further evaluate the model behavior on downstream tasks requiring reading text, including Visual Question Answering (VQA), Image Captioning, and Text-Image Retrieval. Specifically, for the text-oriented tasks, we use Text VQA [36] and ST-VQA [5] for VQA, and TextCaps [35] for captioning and retrieval. Moreover, we also provide the same tasks on the datasets that only require the model to see, i.e., the natural image dataset. Similarly, we use VQAv2 [14] for VQA and COCO [8] for captioning and retrieval. + +# 6.1 Ablation Study of Data Curation on CLIP + +Curation I: Embedded Text in Images. To study the impact of embedded text on overall pre-train data quality, we sample three subsets: random baseline, + +Table 3: Ablation of images embedded with or without text. + +
DataModelINRet.Avg.
3M RandomRN500.2040.2220.256
3M w/o Emb. TextRN500.2280.2320.282
3M w/ Emb. Text OnlyRN500.0710.1390.164
3M RandomViT-B0.1310.1480.210
3M w/o Emb. TextViT-B0.1620.1640.234
3M w/ Emb. Text OnlyViT-B0.0520.1110.153
12M RandomRN500.3600.3540.354
12M w/o Emb. TextRN500.4090.3610.372
12M w/ Emb. Text OnlyRN500.1290.1920.218
12M RandomViT-B0.3140.2990.351
12M w/o Emb. TextViT-B0.3700.3180.364
12M w/ Emb. Text OnlyViT-B0.1290.1720.225
+ +Table 5: Ablation of models trained on subsets sampled by different RSA. + +
Data (3M)ModelAvg.S(●)INRet.Avg.
RSA < 0.0RN500.3190.1810.2200.239
RSA ≥ 0.0RN500.3390.1260.1800.215
RSA ≥ 0.1RN500.3510.0410.1230.148
RSA ≥ 0.2RN500.3600.0170.0940.109
RSA ≥ 0.3RN500.3760.0090.0750.097
RSA < 0.0ViT-B0.3190.1230.1590.198
RSA ≥ 0.0ViT-B0.3390.0790.1290.185
RSA ≥ 0.1ViT-B0.3510.0310.1030.134
RSA ≥ 0.2ViT-B0.3600.0120.0800.103
RSA ≥ 0.3ViT-B0.3760.0060.0700.096
+ +Table 4: Ablation of different CoEmb. Text Rate (CoTR). + +
Data (3M)ModelINRet.Avg.
CoTR = 0.0RN500.1930.2290.247
CoTR ≥ 0.3RN500.0310.1100.137
CoTR ≥ 0.5RN500.0210.0990.124
CoTR ≥ 0.8RN500.0120.0820.096
CoTR = 1.0RN500.0120.0740.102
CoTR = 0.0ViT-B0.1320.1640.206
CoTR ≥ 0.3ViT-B0.0290.0840.130
CoTR ≥ 0.5ViT-B0.0210.0820.119
CoTR ≥ 0.8ViT-B0.0120.0760.104
CoTR = 1.0ViT-B0.0130.0760.103
+ +Table 6: Ablation of models trained on subsets sampled by different RSC. + +
Data (3M)ModelAvg.S(●)INRet.Avg.
RSC < 0.0RN500.3260.1250.1710.209
RSC ≥ 0.0RN500.3450.0620.1290.168
RSC ≥ 0.1RN500.3540.0140.0910.106
RSC ≥ 0.2RN500.3640.0080.0840.104
RSC ≥ 0.3RN500.3800.0050.0580.084
RSC < 0.0ViT-B0.3260.0790.1290.174
RSC ≥ 0.0ViT-B0.3450.0450.1190.149
RSC ≥ 0.1ViT-B0.3540.0180.0910.116
RSC ≥ 0.2ViT-B0.3640.0080.0760.106
RSC ≥ 0.3ViT-B0.3800.0040.0590.091
+ +images without any embedded text, and images all embedded with text from LAION-2B. The subsets include 3M and 12M scales. The results in Tab. 3 show that images embedded with text generally reduce the pre-training dataset quality as all performance tracks significantly decrease. Meanwhile, in Fig. 7, the model trained with the images embedded with text achieves the strongest text spotting capacity compared to the random and images without text baselines. + +Curation II: Co-Emb. Text Rate (CoTR). Tab. 3 reports CLIP models trained on parrot captions with different CoTR. We first select all the images with embedded text and then sample images based on the CoTR depicted at Algorithm 1 with different thresholds. With increasing CoTR, all the zero-shot benchmark performance drops significantly. Despite the images in the subset $(\mathrm{CoTR} = 0)$ all embedded with text, the pre-trained model performs similarly to the random baseline in Tab. 3. It indicates that the parrot caption is more crucial than embedded text in reducing the pre-trained data quality. For the text spotting capacity, Fig. 7 and Fig. 8 show that the increasing CoTR does not lead to stronger text spotting capacity, possibly due to the average length of captions decreasing in higher CoTR data. + +Curation III: Relative Score from Text Removal (RSA & RSC). Given the observations in Sec. 5.1, we further select a series of subsets based on the relative score of images before and after text removal. The higher relative scores mean the masked text is more dominant in CLIP score measurement. In Tab. 5 and Tab. 6, we report the zero-shot performance of models trained on subsets with different relative score thresholds. CLIP models pre-trained with higher + +![](images/64665d5bd07426a917b953a91bf6c10ea5b1dc10604bd1eb648f2e6076f8ea24.jpg) +Fig. 7: CLIP models learn text spotting well from parrot captions. Benchmarking text spotting capacity of CLIP models with 1-gram caption vocabulary synthetic images dataset as the same as Sec. 5.2. + +![](images/a21f85ab1ad31fb86cc26962da01e2a2fbf9ca466049c21a88c2a0e29ac1974f.jpg) +Fig. 8: Text spotting capacity validation on real images. Models trained with more parrot captions are better at aligning the image with parrot captions but perform worse at aligning images without embedded text. + +RSA or RSC both get worse downstream performance. Notably, the average raw CLIP scores $\mathrm{S}(\bullet)$ of these subsets have a positive correlation with RSA or RSC, indicating using CLIP scores from a biased pre-trained model as the data filtering strategy can be unreliable. When comparing the RSA and RSC, the results show that the samples dominated by the latter, i.e., parrot captions, are less informative for CLIP training. Meanwhile, Fig. 7 and Fig. 8 show that the text spotting capacity of CLIP can be further improved by training on the samples using relative scores as data curation criteria against CoTR. + +# 6.2 More Investigation on Text-Oriented Tasks + +Inspired by [12], we further train BLIP [21] models on the curated subsets to better understand the impact of parrot caption on various vision-language tasks, especially the tasks that require models to read the text. We chose BLIP for the ablation study instead of CLIP as it can be directly applied to all these tasks. As shown in Tab. 7, training BLIPs to spot text can boost their performance on the downstream tasks requiring the model to read but impede the performance of downstream tasks only requiring the model to see, which are consistent with the observation on classification tasks. Nevertheless, when BLIPs mainly focus + +Table 7: BLIP downstream tasks performance of pre-training on different curated 3M subsets. The gray color represents tasks requiring the model to read the text from images, i.e., spotting text from images. + +
BLIP Data (3M)Visual Question Answering (Acc)Image Captioning (CIDEr)Text-to-Image Retrieval (R@1)Image-to-Text Retrieval (R@1)
VQAv2TextVQAST-VQACOCOTextCapsCOCOTextCapsCOCOTextCaps
Rand71.0715.3610.48115.653.748.9156.3465.4672.45
w/ Emb. Text68.9419.0512.65108.992.142.8970.158.581.42
w/o Emb. Text71.2213.659.29116.241.549.9631.8366.548.7
CoTR = 0.071.1113.979.75116.344.649.5538.0566.0854.57
CoTR ≥ 0.367.419.2811.81104.996.937.7867.2851.9878.2
CoTR ≥ 0.567.0219.6412.38102.794.135.9465.2450.3276.94
CoTR ≥ 0.866.3818.5012.00100.991.634.1362.6546.973.56
CoTR = 1.066.1818.4712.80101.291.333.5561.8346.6273.05
RSA < 0.070.7914.169.64115.744.948.2536.8564.7254.7
RSA ≥ 0.070.0318.7611.81111.984.546.2568.6162.9281.23
RSA ≥ 0.168.1419.4813.33105.696.139.9668.1354.6479.37
RSA ≥ 0.266.0121.0611.8598.794.433.0364.1747.1275.33
RSA ≥ 0.364.2018.4412.0495.2691.126.6460.1137.370.24
RSC < 0.070.1315.1910.74112.246.746.841.9563.2458.05
RSC ≥ 0.068.8620.1213.75107.893.542.069.7857.4280.92
RSC ≥ 0.167.3520.5412.84103.496.936.466.6951.0277.79
RSC ≥ 0.262.6220.3213.1498.792.830.0861.3842.9671.98
RSC ≥ 0.363.7518.9413.0392.988.724.2358.3534.7268.95
+ +on reading, e.g. (RSA $\geq 0.3$ ), their text-oriented and natural downstream performance also decreases. In other words, the parrot captions can benefit the text-orient downstream tasks while requiring careful data mixing trade-off. We believe that understanding parrot captions is essential for revealing the underlying mechanisms of existing large vision-language systems. + +# 7 Conclusion and Discussion + +The popularity of vision-language contrastive loss stems from its efficiency and simplicity. However, the analysis and experiments we presented show that the embedded text in images and their parrot captions plant significant text spotting bias due to such contrastive fashions. Firstly, almost $30\%$ of the captions in the widely used LAION-2B dataset are biased towards parroting the embedded text in images. Secondly, the pre-trained CLIP models have strong preferences for the image-text pair with parrot captions, which achieve higher similarity scores than those without. Finally, using data biasing to parrot captions, we can easily train a CLIP model with a strong text spotting bias. Our work demonstrates the emergency of reviewing the impact of parrot captions in the entire ecosystem of CLIP models. Our future endeavor involves building a bias-aware data curation pipeline and a robust training function to mitigate such issues. + +# Acknowledgements + +This project was supported by the National Key R&D Program of China (No. 2022ZD0160101) and the Shanghai Artificial Intelligence Laboratory. Mike Shou does not receive any funding for this work. + +# References + +1. Agarwal, S., Krueger, G., Clark, J., Radford, A., Kim, J.W., Brundage, M.: Evaluating clip: towards characterization of broader capabilities and downstream implications. arXiv preprint arXiv:2108.02818 (2021) +2. Alabdulmohsin, I., Wang, X., Steiner, A.P., Goyal, P., D'Amour, A., Zhai, X.: Clip the bias: How useful is balancing data in multimodal learning? In: ICLR (2024) +3. Ali, J., Kleindessner, M., Wenzel, F., Budhathoki, K., Cevher, V., Russell, C.: Evaluating the fairness of discriminative foundation models in computer vision. In: AIES. pp. 809-833 (2023) +4. Berg, H., Hall, S.M., Bhalgat, Y., Yang, W., Kirk, H.R., Shtedritski, A., Bain, M.: A prompt array keeps the bias away: Debiasing vision-language models with adversarial learning. arXiv preprint arXiv:2203.11933 (2022) +5. Biten, A.F., Tito, R., Mafla, A., Gomez, L., Rusinol, M., Valveny, E., Jawahar, C., Karatzas, D.: Scene text visual question answering. In: ICCV. pp. 4291-4301 (2019) +6. Byeon, M., Park, B., Kim, H., Lee, S., Baek, W., Kim, S.: Coyo-700m: Image-text pair dataset. https://github.com/kakaobrain/coyo-dataset (2022) +7. Cao, L., Zhang, B., Chen, C., Yang, Y., Du, X., Zhang, W., Lu, Z., Zheng, Y.: Less is more: Removing text-regions improves clip training efficiency and robustness. arXiv preprint arXiv:2305.05095 (2023) +8. Chen, X., Fang, H., Lin, T.Y., Vedantam, R., Gupta, S., Dolkar, P., Zitnick, C.L.: Microsoft coco captions: Data collection and evaluation server. arXiv preprint arXiv:1504.00325 (2015) +9. Devlin, J., Chang, M.W., Lee, K., Toutanova, K.: Bert: Pre-training of deep bidirectional transformers for language understanding. arXiv preprint arXiv:1810.04805 (2018) +10. Dosovitskiy, A., Beyer, L., Kolesnikov, A., Weissenborn, D., Zhai, X., Unterthiner, T., Dehghani, M., Minderer, M., Heigold, G., Gelly, S., et al.: An image is worth 16x16 words: Transformers for image recognition at scale. arXiv preprint arXiv:2010.11929 (2020) +1. Gadre, S.Y., Ilharco, G., Fang, A., Hayase, J., Smyrnis, G., Nguyen, T., Marten, R., Wortsman, M., Ghosh, D., Zhang, J., et al.: Datacomp: In search of the next generation of multimodal datasets. arXiv preprint arXiv:2304.14108 (2023) +2. Ganz, R., Nuriel, O., Aberdam, A., Kittenplon, Y., Mazor, S., Litman, R.: Towards models that can see and read. arXiv preprint arXiv:2301.07389 (2023) +3. Goh, G., Cammarata, N., Voss, C., Carter, S., Petrov, M., Schubert, L., Radford, A., Olah, C.: Multimodal neurons in artificial neural networks. Distill 6(3), e30 (2021) +4. Goyal, Y., Khot, T., Summers-Stay, D., Batra, D., Parikh, D.: Making the v in vqa matter: Elevating the role of image understanding in visual question answering. In: CVPR. pp. 6904-6913 (2017) +5. He, K., Zhang, X., Ren, S., Sun, J.: Deep residual learning for image recognition. In: CVPR. pp. 770-778 (2016) +6. Ilharco, G., Wortsman, M., Wightman, R., Gordon, C., Carlini, N., Taori, R., Dave, A., Shankar, V., Namkoong, H., Miller, J., Hajishirzi, H., Farhadi, A., Schmidt, L.: Openclip (Jul 2021), https://doi.org/10.5281/zenodo.5143773 +7. Jia, C., Yang, Y., Xia, Y., Chen, Y.T., Parekh, Z., Pham, H., Le, Q., Sung, Y.H., Li, Z., Duerig, T.: Scaling up visual and vision-language representation learning with noisy text supervision. In: ICML. pp. 4904-4916. PMLR (2021) + +18. Johnson, J., Douze, M., Jégou, H.: Billion-scale similarity search with GPUs. IEEE Transactions on Big Data 7(3), 535-547 (2019) +19. Lemesle, Y., Sawayama, M., Valle-Perez, G., Adolphe, M., Sauzeon, H., Oudeyer, P.Y.: Language-biased image classification: evaluation based on semantic representations. arXiv preprint arXiv:2201.11014 (2022) +20. Li, B., Weinberger, K.Q., Belongie, S., Koltun, V., Ranftl, R.: Language-driven semantic segmentation. In: ICLR (2022) +21. Li, J., Li, D., Xiong, C., Hoi, S.: Blip: Bootstrapping language-image pre-training for unified vision-language understanding and generation. In: ICML. pp. 12888-12900. PMLR (2022) +22. Li, J., Selvaraju, R., Gotmare, A., Joty, S., Xiong, C., Hoi, S.C.H.: Align before fuse: Vision and language representation learning with momentum distillation. NeurIPS 34, 9694-9705 (2021) +23. Loshchilov, I., Hutter, F.: Decoupled weight decay regularization. arXiv preprint arXiv:1711.05101 (2017) +24. Maini, P., Goyal, S., Lipton, Z.C., Kolter, J.Z., Raghunathan, A.: T-mars: Improving visual representations by circumventing text feature learning. arXiv preprint arXiv:2307.03132 (2023) +25. Materzyńska, J., Torralba, A., Bau, D.: Disentangling visual and written concepts in clip. In: CVPR. pp. 16410-16419 (2022) +26. Nichol, A., Dhariwal, P., Ramesh, A., Shyam, P., Mishkin, P., McGrew, B., Sutskever, I., Chen, M.: Glide: Towards photorealistic image generation and editing with text-guided diffusion models. arXiv preprint arXiv:2112.10741 (2021) +27. Pham, H., Dai, Z., Ghiasi, G., Kawaguchi, K., Liu, H., Yu, A.W., Yu, J., Chen, Y.T., Luong, M.T., Wu, Y., et al.: Combined scaling for zero-shot transfer learning. Neurocomputing 555, 126658 (2023) +28. Radenovic, F., Dubey, A., Kadian, A., Mihaylov, T., Vandenhende, S., Patel, Y., Wen, Y., Ramanathan, V., Mahajan, D.: Filtering, distillation, and hard negatives for vision-language pre-training. In: CVPR. pp. 6967-6977 (2023) +29. Radford, A., Kim, J.W., Hallacy, C., Ramesh, A., Goh, G., Agarwal, S., Sastry, G., Askell, A., Mishkin, P., Clark, J., et al.: Learning transferable visual models from natural language supervision. In: ICML. pp. 8748-8763. PMLR (2021) +30. Rombach, R., Blattmann, A., Lorenz, D., Esser, P., Ommer, B.: High-resolution image synthesis with latent diffusion models. In: CVPR. pp. 10684-10695 (2022) +31. Schuhmann, C., Beaumont, R., Vencu, R., Gordon, C., Wightman, R., Cherti, M., Coombes, T., Katta, A., Mullis, C., Wortsman, M., et al.: Laion-5b: An open large-scale dataset for training next generation image-text models. NeurIPS 35, 25278-25294 (2022) +32. Schuhmann, C., Vencu, R., Beaumont, R., Kaczmarczyk, R., Mullis, C., Katta, A., Coombes, T., Jitsev, J., Komatsuzaki, A.: Laion-400m: Open dataset of clip-filtered 400 million image-text pairs. arXiv preprint arXiv:2111.02114 (2021) +33. Shi, C., Yang, S.: Logoprompt: Synthetic text images can be good visual prompts for vision-language models. In: ICCV. pp. 2932-2941 (2023) +34. Shtedritski, A., Rupprecht, C., Vedaldi, A.: What does clip know about a red circle? visual prompt engineering for vlms. arXiv preprint arXiv:2304.06712 (2023) +35. Sidorov, O., Hu, R., Rohrbach, M., Singh, A.: Textcaps: a dataset for image captioning with reading comprehension. In: ECCV. pp. 742-758. Springer (2020) +36. Singh, A., Natarajan, V., Shah, M., Jiang, Y., Chen, X., Batra, D., Parikh, D., Rohrbach, M.: Towards vqa models that can read. In: CVPR. pp. 8317-8326 (2019) +37. Tanjin, M.M., Singh, K.K., Kafle, K., Sinha, R., Cottrell, G.W.: Discovering and mitigating biases in clip-based image editing. In: WACV. pp. 2984-2993 (2024) + +38. Telea, A.: An image inpainting technique based on the fast marching method. Journal of graphics tools 9(1), 23-34 (2004) +39. Tschannen, M., Mustafa, B., Houlsby, N.: Clippo: Image-and-language understanding from pixels only. In: CVPR. pp. 11006-11017 (2023) +40. Wang, H., Zhan, Y., Liu, L., Ding, L., Yu, J.: Balanced similarity with auxiliary prompts: Towards alleviating text-to-image retrieval bias for clip in zero-shot learning. arXiv preprint arXiv:2402.18400 (2024) +41. Wang, J., Liu, Y., Wang, X.E.: Are gender-neutral queries really gender-neutral? mitigating gender bias in image search. arXiv preprint arXiv:2109.05433 (2021) +42. Wei, J., Tay, Y., Bommasani, R., Raffel, C., Zoph, B., Borgeaud, S., Yogatama, D., Bosma, M., Zhou, D., Metzler, D., et al.: Emergent abilities of large language models. arXiv preprint arXiv:2206.07682 (2022) +43. Xu, Y., Xu, Z., Chai, W., Zhao, Z., Song, E., Wang, G.: Devil in the number: Towards robust multi-modality data filter. arXiv preprint arXiv:2309.13770 (2023) +44. Yao, Y., Zhang, A., Zhang, Z., Liu, Z., Chua, T.S., Sun, M.: Cpt: Colorful prompt tuning for pre-trained vision-language models. arXiv preprint arXiv:2109.11797 (2021) +45. Ye, M., Zhang, J., Zhao, S., Liu, J., Liu, T., Du, B., Tao, D.: Deepsolo: Let transformer decoder with explicit points solo for text spotting. In: CVPR. pp. 19348-19357 (2023) +46. Yuksekgonul, M., Bianchi, F., Kalluri, P., Jurafsky, D., Zou, J.: When and why vision-language models behave like bags-of-words, and what to do about it? In: ICLR (2022) +47. Zhai, X., Wang, X., Mustafa, B., Steiner, A., Keysers, D., Kolesnikov, A., Beyer, L.: Lit: Zero-shot transfer with locked-image text tuning. In: CVPR. pp. 18123-18133 (2022) +48. Zhang, Q., Xu, Y., Zhang, J., Tao, D.: Vitaev2: Vision transformer advanced by exploring inductive bias for image recognition and beyond. IJCV pp. 1-22 (2023) +49. Zhou, K., Yang, J., Loy, C.C., Liu, Z.: Learning to prompt for vision-language models. IJCV 130(9), 2337-2348 (2022) \ No newline at end of file diff --git a/2024/Parrot Captions Teach CLIP to Spot Text/images.zip b/2024/Parrot Captions Teach CLIP to Spot Text/images.zip new file mode 100644 index 0000000000000000000000000000000000000000..7abeac02c2eb896722e460f2bc236a137ed4c906 --- /dev/null +++ b/2024/Parrot Captions Teach CLIP to Spot Text/images.zip @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:4d83dac9dfe1e82c21b551d16dcd848ee1bfbf67c030981fbdd63a2924905cf3 +size 754645 diff --git a/2024/Parrot Captions Teach CLIP to Spot Text/layout.json b/2024/Parrot Captions Teach CLIP to Spot Text/layout.json new file mode 100644 index 0000000000000000000000000000000000000000..61583f425906edca7b62327817b1385d7c525fd3 --- /dev/null +++ b/2024/Parrot Captions Teach CLIP to Spot Text/layout.json @@ -0,0 +1,8825 @@ +{ + "pdf_info": [ + { + "para_blocks": [ + { + "bbox": [ + 159, + 112, + 454, + 129 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 159, + 112, + 454, + 129 + ], + "spans": [ + { + "bbox": [ + 159, + 112, + 454, + 129 + ], + "type": "text", + "content": "Parrot Captions Teach CLIP to Spot Text" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 151, + 149, + 462, + 175 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 151, + 149, + 462, + 175 + ], + "spans": [ + { + "bbox": [ + 151, + 149, + 462, + 175 + ], + "type": "text", + "content": "Yiqi Lin" + }, + { + "bbox": [ + 151, + 149, + 462, + 175 + ], + "type": "inline_equation", + "content": "^{1,2*}" + }, + { + "bbox": [ + 151, + 149, + 462, + 175 + ], + "type": "text", + "content": ", Conghui He" + }, + { + "bbox": [ + 151, + 149, + 462, + 175 + ], + "type": "inline_equation", + "content": "^{1*}" + }, + { + "bbox": [ + 151, + 149, + 462, + 175 + ], + "type": "text", + "content": ", Alex Jinpeng Wang" + }, + { + "bbox": [ + 151, + 149, + 462, + 175 + ], + "type": "inline_equation", + "content": "^{2*}" + }, + { + "bbox": [ + 151, + 149, + 462, + 175 + ], + "type": "text", + "content": ", Bin Wang" + }, + { + "bbox": [ + 151, + 149, + 462, + 175 + ], + "type": "inline_equation", + "content": "^{1*}" + }, + { + "bbox": [ + 151, + 149, + 462, + 175 + ], + "type": "text", + "content": ", Weijia Li" + }, + { + "bbox": [ + 151, + 149, + 462, + 175 + ], + "type": "inline_equation", + "content": "^{3}" + }, + { + "bbox": [ + 151, + 149, + 462, + 175 + ], + "type": "text", + "content": ", and Mike Zheng Shou" + }, + { + "bbox": [ + 151, + 149, + 462, + 175 + ], + "type": "inline_equation", + "content": "^{2}" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 212, + 183, + 399, + 194 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 212, + 183, + 399, + 194 + ], + "spans": [ + { + "bbox": [ + 212, + 183, + 399, + 194 + ], + "type": "text", + "content": "1 Shanghai Artificial Intelligence Laboratory" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 212, + 195, + 400, + 205 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 212, + 195, + 400, + 205 + ], + "spans": [ + { + "bbox": [ + 212, + 195, + 400, + 205 + ], + "type": "text", + "content": "2 Show Lab, National University of Singapore" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 255, + 206, + 359, + 217 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 255, + 206, + 359, + 217 + ], + "spans": [ + { + "bbox": [ + 255, + 206, + 359, + 217 + ], + "type": "text", + "content": "3 Sun Yat-Sen University" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 160, + 243, + 455, + 463 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 160, + 243, + 455, + 463 + ], + "spans": [ + { + "bbox": [ + 160, + 243, + 455, + 463 + ], + "type": "text", + "content": "Abstract. Despite CLIP [29] being the foundation model in numerous vision-language applications, CLIP suffers from a severe text spotting bias. Such bias causes CLIP models to 'Parrot' the visual text embedded within images while disregarding the authentic visual semantics. We uncover that in the most popular image-text dataset LAION-2B [31], the captions also densely parrot (spell) the text embedded in images. Our analysis shows that around " + }, + { + "bbox": [ + 160, + 243, + 455, + 463 + ], + "type": "inline_equation", + "content": "50\\%" + }, + { + "bbox": [ + 160, + 243, + 455, + 463 + ], + "type": "text", + "content": " of images are embedded with visual text content and around " + }, + { + "bbox": [ + 160, + 243, + 455, + 463 + ], + "type": "inline_equation", + "content": "30\\%" + }, + { + "bbox": [ + 160, + 243, + 455, + 463 + ], + "type": "text", + "content": " of captions words are concurrently embedded in the visual content. Based on such observation, we thoroughly inspect the different released versions of CLIP models and verify that the visual text is a dominant factor in measuring the LAION-style image-text similarity for these models. To examine whether these parrot captions shape the text spotting bias, we train a series of CLIP models with LAION subsets curated by different parrot-caption-oriented criteria. We show that training with parrot captions easily shapes such bias but harms the expected visual-language representation learning in CLIP models across various vision-language downstream tasks. This suggests that it is urgent to revisit either the design of CLIP-like models or the existing image-text dataset curation pipeline built on CLIP score filtering. Project page: https://tinyq17.github.io/CLIP-Parrot-Bias/" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 160, + 473, + 383, + 484 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 160, + 473, + 383, + 484 + ], + "spans": [ + { + "bbox": [ + 160, + 473, + 383, + 484 + ], + "type": "text", + "content": "Keywords: Image-Text Dataset " + }, + { + "bbox": [ + 160, + 473, + 383, + 484 + ], + "type": "inline_equation", + "content": "\\cdot" + }, + { + "bbox": [ + 160, + 473, + 383, + 484 + ], + "type": "text", + "content": " Text Spotting Bias" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 132, + 503, + 230, + 516 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 503, + 230, + 516 + ], + "spans": [ + { + "bbox": [ + 132, + 503, + 230, + 516 + ], + "type": "text", + "content": "1 Introduction" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 130, + 527, + 482, + 647 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 527, + 482, + 647 + ], + "spans": [ + { + "bbox": [ + 130, + 527, + 482, + 647 + ], + "type": "text", + "content": "Recently, contrastive learning models [17, 29, 31] pre-trained with large-scale image-text pair data has led to numerous vision-language modeling task breakthroughs. Due to its efficiency and simplicity, the pioneering work CLIP [29] now serves as a foundation model in various applications [20, 26, 30, 49]. However, several works [4,13] have shown that CLIP models have perpetuating biases towards visual text [19,25], color [34,44], gender [41], etc. In this paper, we focus on probing the visual text bias in CLIP, i.e., the capacity of spotting text in images. Most of the previous cues [25, 29, 34] attribute the sources of biases to the noisy pre-training data. Therefore, we begin by taking a close look at the most popular dataset, LAION-2B [31]." + } + ] + } + ], + "index": 8 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 133, + 653, + 463, + 666 + ], + "type": "page_footnote", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 653, + 463, + 666 + ], + "spans": [ + { + "bbox": [ + 133, + 653, + 463, + 666 + ], + "type": "text", + "content": "* Equal contribution. ⌒ Corresponding author (Email: heconghui@pjlab.org.cn)." + } + ] + } + ], + "index": 9 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 0 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 133, + 113, + 481, + 317 + ], + "blocks": [ + { + "bbox": [ + 133, + 113, + 481, + 317 + ], + "lines": [ + { + "bbox": [ + 133, + 113, + 481, + 317 + ], + "spans": [ + { + "bbox": [ + 133, + 113, + 481, + 317 + ], + "type": "image", + "image_path": "66c472be5cef6c6f9d3fbde23da1f2ee370ac6db3009142b22eee8358e6f11e4.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 269, + 316, + 338, + 324 + ], + "lines": [ + { + "bbox": [ + 269, + 316, + 338, + 324 + ], + "spans": [ + { + "bbox": [ + 269, + 316, + 338, + 324 + ], + "type": "text", + "content": "(b) Image Paired Captions" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 132, + 333, + 482, + 399 + ], + "lines": [ + { + "bbox": [ + 132, + 333, + 482, + 399 + ], + "spans": [ + { + "bbox": [ + 132, + 333, + 482, + 399 + ], + "type": "text", + "content": "Fig. 1: In LAION-2B [31], image-text pairs with the Top-5% highest similarity score are most dominant by visual text! These samples have dense concurrent text appearing in captions and images (text form in pixels). We refer to their captions as Parrot Captions as they raise a question: Dose CLIP Simply Parroting Text in Images for Vision-Language Alignment? The concurrent text is spotted by the OCR model and highlighted with color in image-text pairs. (Best view in color)" + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "image_caption" + } + ], + "index": 2 + }, + { + "bbox": [ + 130, + 426, + 482, + 581 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 426, + 482, + 581 + ], + "spans": [ + { + "bbox": [ + 130, + 426, + 482, + 581 + ], + "type": "text", + "content": "Considering the massive scale of the image-text data, it is non-trivial to assess the bias simply with a rough estimation. To this end, we first do image clustering on the whole dataset and rank each cluster by CLIP scores to analyze the most preferred types of image-text pairs under CLIP score measurement. As shown in Fig. 1, we surprisingly observe that a decent number of samples with top CLIP scores have dense concurrent text appearing in the captions and the images in the form of pixels. These samples break the assumption that CLIP models leverage text supervision to align the visual and language concepts. We refer to these captions as Parrot Captions as they provide another shortcut to achieve the same goal by teaching CLIP to do text spotting even without perceiving the actual visual concepts. To understand the underlying impact, we analyze the parrot captions from three perspectives: dataset, widely used released models, and model training. The results lead to three key findings:" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 130, + 582, + 482, + 666 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 582, + 482, + 666 + ], + "spans": [ + { + "bbox": [ + 130, + 582, + 482, + 666 + ], + "type": "text", + "content": "Firstly, captions in LAION-2B have a significant bias towards describing visual text content embedded in the images. We provide thorough profiling using off-the-self text spotting models on the LAION-2B dataset and show that over " + }, + { + "bbox": [ + 130, + 582, + 482, + 666 + ], + "type": "inline_equation", + "content": "50\\%" + }, + { + "bbox": [ + 130, + 582, + 482, + 666 + ], + "type": "text", + "content": " of the images are embedded with visual text content. Moreover, by examining the spotted text content and the paired caption in the subset that images embedded with text, we find that over " + }, + { + "bbox": [ + 130, + 582, + 482, + 666 + ], + "type": "inline_equation", + "content": "60\\%" + }, + { + "bbox": [ + 130, + 582, + 482, + 666 + ], + "type": "text", + "content": " of the captions at least have one concurrent word and reach at least around " + }, + { + "bbox": [ + 130, + 582, + 482, + 666 + ], + "type": "inline_equation", + "content": "30\\%" + }, + { + "bbox": [ + 130, + 582, + 482, + 666 + ], + "type": "text", + "content": " words overlap" + } + ] + } + ], + "index": 6 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "spans": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "type": "text", + "content": "2" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 166, + 91, + 217, + 100 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 166, + 91, + 217, + 100 + ], + "spans": [ + { + "bbox": [ + 166, + 91, + 217, + 100 + ], + "type": "text", + "content": "Y. Lin et al." + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 1 + }, + { + "para_blocks": [ + { + "bbox": [ + 130, + 116, + 480, + 152 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 116, + 480, + 152 + ], + "spans": [ + { + "bbox": [ + 130, + 116, + 480, + 152 + ], + "type": "text", + "content": "between the caption and spotted text from images. This finding suggests that the basic assumption of image-text semantic alignment in CLIP does not fully stand its ground when training with LAION-style data." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 130, + 152, + 482, + 319 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 152, + 482, + 319 + ], + "spans": [ + { + "bbox": [ + 130, + 152, + 482, + 319 + ], + "type": "text", + "content": "Secondly, released CLIP models have strong text spotting bias almost in every style of web images, resulting in CLIP-filtering datasets inherently biased towards visual text-dominant data. We investigate OpenAI released CLIP model's behaviors in the LAION-2B dataset by examining the difference between alignment scores before and after text removal. The results show that CLIP model predictions densely correlate the visual text embedded in images with their parrot captions. Next, we further study the preference of the text spotting capacity on text content in CLIP and OpenCLIP models. Note that CLIP is trained on WIT-400M, while OpenCLIP uses the LAION-2B dataset. Therefore, we use synthetic images embedded with specific rendered text to avoid overfitting in OpenCLIP models. Our analysis shows that OpenCLIP is more biased toward text spotting than CLIP. We believe that the parrot caption plays a lurking role in training these released CLIP models and is the source of text spotting capacity instead of emergence behavior [42]." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 130, + 319, + 482, + 463 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 319, + 482, + 463 + ], + "spans": [ + { + "bbox": [ + 130, + 319, + 482, + 463 + ], + "type": "text", + "content": "Thirdly, CLIP models can easily learn text spotting capacity from parrot captions while failing to connect the vision-language semantics, just like a text spotting parrot. We sample different LAION-2B subsets curated by text-oriented criteria, including the embedded text ratio, the concurrent word ratios, and the relative CLIP score from text removal to train CLIP models under the same setting. The results show that using parrot captions data, CLIP model can learn strong text spotting capacity but lose most of the zero-shot generalization ability on image-text downstream tasks. Moreover, we also observe similar behaviors on various downstream tasks such VQA, captioning, and retrieval on BLIP [21] models. Lastly, we argue that the existing data curation pipeline built on CLIP score and the contrastive fashion urgently needs to be re-examined by considering such hidden parrot captions." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 131, + 479, + 237, + 491 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 131, + 479, + 237, + 491 + ], + "spans": [ + { + "bbox": [ + 131, + 479, + 237, + 491 + ], + "type": "text", + "content": "2 Related Work" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 131, + 503, + 370, + 515 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 131, + 503, + 370, + 515 + ], + "spans": [ + { + "bbox": [ + 131, + 503, + 370, + 515 + ], + "type": "text", + "content": "2.1 Contrastive Vision-Language Pre-training" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 130, + 521, + 482, + 666 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 521, + 482, + 666 + ], + "spans": [ + { + "bbox": [ + 130, + 521, + 482, + 666 + ], + "type": "text", + "content": "Modeling vision and language by aligning the embedding similarity between paired image-text data [17, 29, 31] has shown great potential for transferable to downstream vision-language tasks. The pre-training techniques mainly contain the vision encoder [10, 15] for image embedding encoding, text encoder [9] for text embedding modeling, and cross-modal contrastive learning [17, 22, 29, 47] for learning a joint embedding space of vision and language. The pioneering work CLIP [29] leverages 400 million noisy image-text pairs to learn transferable visual representation from text supervision and show impressive zero-shot performance for various vision-language tasks. Following CLIP, several vision-language models such as ALIGN [17], BASIC [27], and Open-CLIP [31] are proposed, and CLIP models have been replicated on various datasets including WIT [29], LAION [31], COYO [6], and DataComp [11]. We mainly profile" + } + ] + } + ], + "index": 7 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 274, + 91, + 447, + 102 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 274, + 91, + 447, + 102 + ], + "spans": [ + { + "bbox": [ + 274, + 91, + 447, + 102 + ], + "type": "text", + "content": "Parrot Captions Teach CLIP to Spot Text" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 474, + 92, + 480, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 474, + 92, + 480, + 100 + ], + "spans": [ + { + "bbox": [ + 474, + 92, + 480, + 100 + ], + "type": "text", + "content": "3" + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 2 + }, + { + "para_blocks": [ + { + "bbox": [ + 130, + 116, + 482, + 164 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 116, + 482, + 164 + ], + "spans": [ + { + "bbox": [ + 130, + 116, + 482, + 164 + ], + "type": "text", + "content": "the LAION-2B [31] dataset due to its large scale and wide usage [26, 30] and two versions of pre-trained models, CLIP and OpenCLIP. Note that the 2 billion image-text pairs in the LAION-2B dataset are filtered by OpenAI released CLIP models, making OpenCLIP connect to CLIP closely." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 131, + 182, + 304, + 194 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 131, + 182, + 304, + 194 + ], + "spans": [ + { + "bbox": [ + 131, + 182, + 304, + 194 + ], + "type": "text", + "content": "2.2 Studying of CLIP Behaviors" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 130, + 201, + 482, + 382 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 201, + 482, + 382 + ], + "spans": [ + { + "bbox": [ + 130, + 201, + 482, + 382 + ], + "type": "text", + "content": "Despite the strong zero-shot and transferable performance of CLIP, the perpetuating biases [1, 3, 13, 19, 41, 46] in CLIP are still not well investigated due to its large-scale noisy training data. Much research [2, 25, 34, 37, 40, 43, 44] focuses on revealing or enhancing the downstream performance with discovered bias in CLIP. For example, colorful masks [44] or red circles [34] applied to images can improve the zero-shot performance on visual localization tasks. In studying visual text content bias, [13] shows the multimodal neurons of CLIP not only respond to visual content and the visual text embedded in the image. Another work [19] shows that image recognition in CLIP can be strongly dominated by the visual text embedded in the image. To disentangle such bias, [25] attempts to separate the text spotting representation in pre-trained CLIP by training representation projection. Meanwhile, LoGoPrompt [33] enhances the classification performance by utilizing the visual text content as auxiliary prompts as input. Also, CLIPPO [39] shows that directly aligning the image and synthetic images print with the captions can perform similarly to CLIP without a text encoder." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 131, + 399, + 334, + 410 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 131, + 399, + 334, + 410 + ], + "spans": [ + { + "bbox": [ + 131, + 399, + 334, + 410 + ], + "type": "text", + "content": "2.3 Data Curation with Text Removal" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 130, + 418, + 482, + 586 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 418, + 482, + 586 + ], + "spans": [ + { + "bbox": [ + 130, + 418, + 482, + 586 + ], + "type": "text", + "content": "Due to the successful practice of data curation in LAION datasets [31, 32] on scaling up the image-text datasets, searching advanced selection strategy to improve the data quality from common crawl data pool gains a growing interest [11]. Recently, several works [7, 24, 28] suggest that introducing text-related filtering methods improves the pre-training dataset quality. In DiHT [28], the data curation steps include filtering out the image-text pairs with high OCR confidence and matching text ratio. Moreover, [7, 24] mainly focus on studying the importance of filtering out the text-dominate images utilizing OCR models to improve pre-training dataset quality. Maini et al. [24] also draw the observation that " + }, + { + "bbox": [ + 130, + 418, + 482, + 586 + ], + "type": "inline_equation", + "content": "40\\%" + }, + { + "bbox": [ + 130, + 418, + 482, + 586 + ], + "type": "text", + "content": " of LAION's image text is highly correlated with the caption, but only performing a small pilot study on 500 samples with manual judgment. Differently, this paper makes the first attempt to systemically reveal the source of text spotting capacity in CLIP is the data bias and the consequences of such bias in existing widely used datasets and pre-trained models." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 131, + 604, + 230, + 618 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 131, + 604, + 230, + 618 + ], + "spans": [ + { + "bbox": [ + 131, + 604, + 230, + 618 + ], + "type": "text", + "content": "3 Terminology" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 130, + 629, + 481, + 666 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 629, + 481, + 666 + ], + "spans": [ + { + "bbox": [ + 130, + 629, + 481, + 666 + ], + "type": "text", + "content": "The data processing on images in the following sections mainly covers clustering, text spotting (OCR), and text inpainting. Firstly, we cluster all images based on feature similarity. For each image-text pair, we then use the pre-trained text" + } + ] + } + ], + "index": 8 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 133, + 91, + 141, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 91, + 141, + 100 + ], + "spans": [ + { + "bbox": [ + 133, + 91, + 141, + 100 + ], + "type": "text", + "content": "4" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 166, + 91, + 217, + 101 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 166, + 91, + 217, + 101 + ], + "spans": [ + { + "bbox": [ + 166, + 91, + 217, + 101 + ], + "type": "text", + "content": "Y. Lin et al." + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 3 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 177, + 114, + 228, + 165 + ], + "blocks": [ + { + "bbox": [ + 177, + 114, + 228, + 165 + ], + "lines": [ + { + "bbox": [ + 177, + 114, + 228, + 165 + ], + "spans": [ + { + "bbox": [ + 177, + 114, + 228, + 165 + ], + "type": "image", + "image_path": "ec43366ccf9e360bbbeaf331d6a94f2609b1b3c7650bc29f046b94e643520bcc.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 229, + 173, + 274, + 178 + ], + "lines": [ + { + "bbox": [ + 229, + 173, + 274, + 178 + ], + "spans": [ + { + "bbox": [ + 229, + 173, + 274, + 178 + ], + "type": "text", + "content": "Text Removal" + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "image_caption" + } + ], + "index": 2 + }, + { + "type": "image", + "bbox": [ + 229, + 114, + 280, + 165 + ], + "blocks": [ + { + "bbox": [ + 229, + 114, + 280, + 165 + ], + "lines": [ + { + "bbox": [ + 229, + 114, + 280, + 165 + ], + "spans": [ + { + "bbox": [ + 229, + 114, + 280, + 165 + ], + "type": "image", + "image_path": "c0206ad0f5983bda345a6031654922588a01b6e22fdca1f55da534e9e29fad10.jpg" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 241, + 166, + 265, + 172 + ], + "lines": [ + { + "bbox": [ + 241, + 166, + 265, + 172 + ], + "spans": [ + { + "bbox": [ + 241, + 166, + 265, + 172 + ], + "type": "text", + "content": "All-Emb" + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "image_caption" + } + ], + "index": 3 + }, + { + "type": "image", + "bbox": [ + 281, + 114, + 332, + 165 + ], + "blocks": [ + { + "bbox": [ + 281, + 114, + 332, + 165 + ], + "lines": [ + { + "bbox": [ + 281, + 114, + 332, + 165 + ], + "spans": [ + { + "bbox": [ + 281, + 114, + 332, + 165 + ], + "type": "image", + "image_path": "22adb419c388a8258eee75c2678728ca559f30dec4bdfa1c39b9bda2173a3117.jpg" + } + ] + } + ], + "index": 6, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 292, + 166, + 318, + 172 + ], + "lines": [ + { + "bbox": [ + 292, + 166, + 318, + 172 + ], + "spans": [ + { + "bbox": [ + 292, + 166, + 318, + 172 + ], + "type": "text", + "content": "Co-Emb" + } + ] + } + ], + "index": 7, + "angle": 0, + "type": "image_caption" + } + ], + "index": 6 + }, + { + "type": "image", + "bbox": [ + 333, + 114, + 384, + 165 + ], + "blocks": [ + { + "bbox": [ + 333, + 114, + 384, + 165 + ], + "lines": [ + { + "bbox": [ + 333, + 114, + 384, + 165 + ], + "spans": [ + { + "bbox": [ + 333, + 114, + 384, + 165 + ], + "type": "image", + "image_path": "0229e13685709ee8331372d45a1963ae8fec27b5590c2be688da4129a8e0944e.jpg" + } + ] + } + ], + "index": 9, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 345, + 166, + 371, + 172 + ], + "lines": [ + { + "bbox": [ + 345, + 166, + 371, + 172 + ], + "spans": [ + { + "bbox": [ + 345, + 166, + 371, + 172 + ], + "type": "text", + "content": "Random" + } + ] + } + ], + "index": 10, + "angle": 0, + "type": "image_caption" + } + ], + "index": 9 + }, + { + "type": "image", + "bbox": [ + 386, + 114, + 436, + 165 + ], + "blocks": [ + { + "bbox": [ + 386, + 114, + 436, + 165 + ], + "lines": [ + { + "bbox": [ + 386, + 114, + 436, + 165 + ], + "spans": [ + { + "bbox": [ + 386, + 114, + 436, + 165 + ], + "type": "image", + "image_path": "7ecc161b44339da337e15ee9bd347b579cb0936d5353013cc1d2e3b02e25297a.jpg" + } + ] + } + ], + "index": 12, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 394, + 166, + 424, + 172 + ], + "lines": [ + { + "bbox": [ + 394, + 166, + 424, + 172 + ], + "spans": [ + { + "bbox": [ + 394, + 166, + 424, + 172 + ], + "type": "text", + "content": "Syn-Emb." + } + ] + } + ], + "index": 13, + "angle": 0, + "type": "image_caption" + } + ], + "index": 12 + }, + { + "type": "image", + "bbox": [ + 178, + 181, + 228, + 231 + ], + "blocks": [ + { + "bbox": [ + 178, + 181, + 228, + 231 + ], + "lines": [ + { + "bbox": [ + 178, + 181, + 228, + 231 + ], + "spans": [ + { + "bbox": [ + 178, + 181, + 228, + 231 + ], + "type": "image", + "image_path": "d576b18a2f5bfaf7df2e147fd0db4c0e19f22a0dc0032f50c36fd2cb769e031d.jpg" + } + ] + } + ], + "index": 15, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 191, + 231, + 214, + 239 + ], + "lines": [ + { + "bbox": [ + 191, + 231, + 214, + 239 + ], + "spans": [ + { + "bbox": [ + 191, + 231, + 214, + 239 + ], + "type": "text", + "content": "Caption" + } + ] + } + ], + "index": 16, + "angle": 0, + "type": "image_caption" + } + ], + "index": 15 + }, + { + "type": "image", + "bbox": [ + 229, + 181, + 280, + 232 + ], + "blocks": [ + { + "bbox": [ + 229, + 181, + 280, + 232 + ], + "lines": [ + { + "bbox": [ + 229, + 181, + 280, + 232 + ], + "spans": [ + { + "bbox": [ + 229, + 181, + 280, + 232 + ], + "type": "image", + "image_path": "f8c8a090e256890a569b602e432cd85ab2f52a2c877899f1c3f867bf434d292f.jpg" + } + ] + } + ], + "index": 17, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 237, + 232, + 267, + 239 + ], + "lines": [ + { + "bbox": [ + 237, + 232, + 267, + 239 + ], + "spans": [ + { + "bbox": [ + 237, + 232, + 267, + 239 + ], + "type": "text", + "content": "All-Emb." + } + ] + } + ], + "index": 18, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 237, + 239, + 269, + 245 + ], + "lines": [ + { + "bbox": [ + 237, + 239, + 269, + 245 + ], + "spans": [ + { + "bbox": [ + 237, + 239, + 269, + 245 + ], + "type": "text", + "content": "Text Mask" + } + ] + } + ], + "index": 19, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 132, + 255, + 479, + 267 + ], + "lines": [ + { + "bbox": [ + 132, + 255, + 479, + 267 + ], + "spans": [ + { + "bbox": [ + 132, + 255, + 479, + 267 + ], + "type": "text", + "content": "Fig. 2: Visualization of defined terminologies. Co-Emb. Text is marked in the caption." + } + ] + } + ], + "index": 29, + "angle": 0, + "type": "image_caption" + } + ], + "index": 17 + }, + { + "type": "image", + "bbox": [ + 281, + 181, + 332, + 232 + ], + "blocks": [ + { + "bbox": [ + 286, + 173, + 326, + 178 + ], + "lines": [ + { + "bbox": [ + 286, + 173, + 326, + 178 + ], + "spans": [ + { + "bbox": [ + 286, + 173, + 326, + 178 + ], + "type": "text", + "content": "Text Removal" + } + ] + } + ], + "index": 8, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 281, + 181, + 332, + 232 + ], + "lines": [ + { + "bbox": [ + 281, + 181, + 332, + 232 + ], + "spans": [ + { + "bbox": [ + 281, + 181, + 332, + 232 + ], + "type": "image", + "image_path": "a25504c95347df076b368e630762b307f8a83fcd1ebca7e8cd7975ba69227ef6.jpg" + } + ] + } + ], + "index": 20, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 291, + 232, + 318, + 239 + ], + "lines": [ + { + "bbox": [ + 291, + 232, + 318, + 239 + ], + "spans": [ + { + "bbox": [ + 291, + 232, + 318, + 239 + ], + "type": "text", + "content": "Co-Emb." + } + ] + } + ], + "index": 21, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 290, + 239, + 321, + 245 + ], + "lines": [ + { + "bbox": [ + 290, + 239, + 321, + 245 + ], + "spans": [ + { + "bbox": [ + 290, + 239, + 321, + 245 + ], + "type": "text", + "content": "Text Mask" + } + ] + } + ], + "index": 22, + "angle": 0, + "type": "image_caption" + } + ], + "index": 20 + }, + { + "type": "image", + "bbox": [ + 333, + 181, + 384, + 232 + ], + "blocks": [ + { + "bbox": [ + 345, + 172, + 372, + 178 + ], + "lines": [ + { + "bbox": [ + 345, + 172, + 372, + 178 + ], + "spans": [ + { + "bbox": [ + 345, + 172, + 372, + 178 + ], + "type": "text", + "content": "Inpainted" + } + ] + } + ], + "index": 11, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 333, + 181, + 384, + 232 + ], + "lines": [ + { + "bbox": [ + 333, + 181, + 384, + 232 + ], + "spans": [ + { + "bbox": [ + 333, + 181, + 384, + 232 + ], + "type": "image", + "image_path": "73bedcf722a50c5c75f344f0cb05da4444d8294f2faa31b525a9648847fe68e4.jpg" + } + ] + } + ], + "index": 23, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 346, + 232, + 371, + 238 + ], + "lines": [ + { + "bbox": [ + 346, + 232, + 371, + 238 + ], + "spans": [ + { + "bbox": [ + 346, + 232, + 371, + 238 + ], + "type": "text", + "content": "Random" + } + ] + } + ], + "index": 24, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 337, + 239, + 380, + 246 + ], + "lines": [ + { + "bbox": [ + 337, + 239, + 380, + 246 + ], + "spans": [ + { + "bbox": [ + 337, + 239, + 380, + 246 + ], + "type": "text", + "content": "Sampled Mask" + } + ] + } + ], + "index": 25, + "angle": 0, + "type": "image_caption" + } + ], + "index": 23 + }, + { + "type": "image", + "bbox": [ + 391, + 181, + 436, + 232 + ], + "blocks": [ + { + "bbox": [ + 390, + 172, + 432, + 178 + ], + "lines": [ + { + "bbox": [ + 390, + 172, + 432, + 178 + ], + "spans": [ + { + "bbox": [ + 390, + 172, + 432, + 178 + ], + "type": "text", + "content": "Text Template" + } + ] + } + ], + "index": 14, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 391, + 181, + 436, + 232 + ], + "lines": [ + { + "bbox": [ + 391, + 181, + 436, + 232 + ], + "spans": [ + { + "bbox": [ + 391, + 181, + 436, + 232 + ], + "type": "image", + "image_path": "c8d01481022bbad859c05e185888241888398dd5b3f84e528b9f563ccafb5dba.jpg" + } + ] + } + ], + "index": 26, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 395, + 232, + 425, + 238 + ], + "lines": [ + { + "bbox": [ + 395, + 232, + 425, + 238 + ], + "spans": [ + { + "bbox": [ + 395, + 232, + 425, + 238 + ], + "type": "text", + "content": "Syn-Emb." + } + ] + } + ], + "index": 27, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 390, + 239, + 432, + 246 + ], + "lines": [ + { + "bbox": [ + 390, + 239, + 432, + 246 + ], + "spans": [ + { + "bbox": [ + 390, + 239, + 432, + 246 + ], + "type": "text", + "content": "Text Template" + } + ] + } + ], + "index": 28, + "angle": 0, + "type": "image_caption" + } + ], + "index": 26 + }, + { + "bbox": [ + 133, + 281, + 403, + 293 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 281, + 403, + 293 + ], + "spans": [ + { + "bbox": [ + 133, + 281, + 403, + 293 + ], + "type": "text", + "content": "Algorithm 1 Pseudocode of Detecting Co-Emb. Text (Rate)" + } + ] + } + ], + "index": 30 + }, + { + "bbox": [ + 132, + 297, + 289, + 304 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 297, + 289, + 304 + ], + "spans": [ + { + "bbox": [ + 132, + 297, + 289, + 304 + ], + "type": "text", + "content": "caption: captions from LAION-2B dataset." + } + ] + } + ], + "index": 31 + }, + { + "bbox": [ + 135, + 304, + 271, + 311 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 135, + 304, + 271, + 311 + ], + "spans": [ + { + "bbox": [ + 135, + 304, + 271, + 311 + ], + "type": "text", + "content": "ocr_text: text spotted by OCR model" + } + ] + } + ], + "index": 32 + }, + { + "bbox": [ + 133, + 312, + 378, + 319 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 312, + 378, + 319 + ], + "spans": [ + { + "bbox": [ + 133, + 312, + 378, + 319 + ], + "type": "text", + "content": "cap_words,OCR_words " + }, + { + "bbox": [ + 133, + 312, + 378, + 319 + ], + "type": "inline_equation", + "content": "=" + }, + { + "bbox": [ + 133, + 312, + 378, + 319 + ], + "type": "text", + "content": " set(capTION.split()，set(ocr_text.split()))" + } + ] + } + ], + "index": 33 + }, + { + "bbox": [ + 133, + 319, + 311, + 326 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 319, + 311, + 326 + ], + "spans": [ + { + "bbox": [ + 133, + 319, + 311, + 326 + ], + "type": "text", + "content": "co_emb_text = intersection(cap_words,OCR_words)" + } + ] + } + ], + "index": 34 + }, + { + "bbox": [ + 133, + 326, + 326, + 333 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 326, + 326, + 333 + ], + "spans": [ + { + "bbox": [ + 133, + 326, + 326, + 333 + ], + "type": "text", + "content": "co_emb_text_rate = len(co_emb_text) / len(cap_words)" + } + ] + } + ], + "index": 35 + }, + { + "bbox": [ + 130, + 361, + 481, + 421 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 361, + 481, + 421 + ], + "spans": [ + { + "bbox": [ + 130, + 361, + 481, + 421 + ], + "type": "text", + "content": "spotting model to detect and recognize the text print in image pixels. The mask images in Fig. 2 are the spotted text area. Next, we match the spotted text with the caption using Algorithm 1 to obtain the concurrent words and their ratio in captions. Lastly, we use inpainting to remove the text from the image for CLIPs' pattern ablation. To avoid confusion, we define these concepts as follows," + } + ] + } + ], + "index": 36 + }, + { + "bbox": [ + 137, + 434, + 482, + 665 + ], + "type": "list", + "angle": 0, + "index": 43, + "blocks": [ + { + "bbox": [ + 137, + 434, + 482, + 518 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 137, + 434, + 482, + 518 + ], + "spans": [ + { + "bbox": [ + 137, + 434, + 482, + 518 + ], + "type": "text", + "content": "- Embedded Text: text spotted by OCR models from the images. To study the correlation of embedded text with captions, we define different kinds of embedded text as, 1) All-Emb. Text: all the text is spotted from an image; 2) Co-Emb. Text: spotted text concurrently appears in the image's corresponding captions; 3) Syn-Emb. Text: synthetic text rendered in an image with a fixed font and a blank background. Fig. 2 shows examples of spotted embedded text by binary mask and the rendering synthetic text." + } + ] + } + ], + "index": 37 + }, + { + "bbox": [ + 137, + 519, + 482, + 542 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 137, + 519, + 482, + 542 + ], + "spans": [ + { + "bbox": [ + 137, + 519, + 482, + 542 + ], + "type": "text", + "content": "- Co-Emb. Text Rate (CoTR): the word set IoU of Co-Emb. text and captions (See Algorithm 1)." + } + ] + } + ], + "index": 38 + }, + { + "bbox": [ + 137, + 544, + 482, + 567 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 137, + 544, + 482, + 567 + ], + "spans": [ + { + "bbox": [ + 137, + 544, + 482, + 567 + ], + "type": "text", + "content": "- Image w/ or w/o Embedded Text: spotted text results of a given image are none-empty or empty." + } + ] + } + ], + "index": 39 + }, + { + "bbox": [ + 137, + 568, + 482, + 604 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 137, + 568, + 482, + 604 + ], + "spans": [ + { + "bbox": [ + 137, + 568, + 482, + 604 + ], + "type": "text", + "content": "- Text Removal Image: do inpainting in the specific spotted text area (All-Emb., Co-Emb., or Random). Random is implemented by sampling other image's text areas. For the different inpainting results, see Fig. 2." + } + ] + } + ], + "index": 40 + }, + { + "bbox": [ + 137, + 605, + 482, + 652 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 137, + 605, + 482, + 652 + ], + "spans": [ + { + "bbox": [ + 137, + 605, + 482, + 652 + ], + "type": "text", + "content": "- Relative Scores (RSA/RSC): the difference in CLIP score between images modified by different inpainting operations while keeping the same captions. RSA and RSC are the short for the relative scores before and after removing All-Emb. text and Co-Emb. text." + } + ] + } + ], + "index": 41 + }, + { + "bbox": [ + 137, + 653, + 392, + 665 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 137, + 653, + 392, + 665 + ], + "spans": [ + { + "bbox": [ + 137, + 653, + 392, + 665 + ], + "type": "text", + "content": "- Image Clusters: image partitions based on K-Means." + } + ] + } + ], + "index": 42 + } + ], + "sub_type": "text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 274, + 91, + 447, + 102 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 274, + 91, + 447, + 102 + ], + "spans": [ + { + "bbox": [ + 274, + 91, + 447, + 102 + ], + "type": "text", + "content": "Parrot Captions Teach CLIP to Spot Text" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 474, + 92, + 480, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 474, + 92, + 480, + 100 + ], + "spans": [ + { + "bbox": [ + 474, + 92, + 480, + 100 + ], + "type": "text", + "content": "5" + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 4 + }, + { + "para_blocks": [ + { + "type": "table", + "bbox": [ + 201, + 147, + 411, + 212 + ], + "blocks": [ + { + "bbox": [ + 132, + 114, + 481, + 137 + ], + "lines": [ + { + "bbox": [ + 132, + 114, + 481, + 137 + ], + "spans": [ + { + "bbox": [ + 132, + 114, + 481, + 137 + ], + "type": "text", + "content": "Table 1: Overall parrot captions statistic. More than " + }, + { + "bbox": [ + 132, + 114, + 481, + 137 + ], + "type": "inline_equation", + "content": "50\\%" + }, + { + "bbox": [ + 132, + 114, + 481, + 137 + ], + "type": "text", + "content": " of images are embedded with text, and " + }, + { + "bbox": [ + 132, + 114, + 481, + 137 + ], + "type": "inline_equation", + "content": "30\\%" + }, + { + "bbox": [ + 132, + 114, + 481, + 137 + ], + "type": "text", + "content": " of caption words are printed in images!" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 201, + 147, + 411, + 212 + ], + "lines": [ + { + "bbox": [ + 201, + 147, + 411, + 212 + ], + "spans": [ + { + "bbox": [ + 201, + 147, + 411, + 212 + ], + "type": "table", + "html": "
Number of Total Images1,985,284,122
Number of Images w/ Emb. Text1,083,896,427
Number of Images w/ Co-Emb. Text663,600,432
Co-Emb. Text Rate (in Total)15.42%
Co-Emb. Text Rate (in Images w/ Emb. Text)28.24%
Fuzzy Co-Emb. Text Rate (in Total)30.46%
Fuzzy Co-Emb. Text Rate (in Images w/ Emb. Text)55.79%
", + "image_path": "f332ac68efc00bd8d5adaf404d4c4f4e8889d2cf2a81fe6b25ada343c5f83bbd.jpg" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "table_body" + } + ], + "index": 3 + }, + { + "bbox": [ + 137, + 230, + 480, + 278 + ], + "type": "list", + "angle": 0, + "index": 6, + "blocks": [ + { + "bbox": [ + 137, + 230, + 479, + 254 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 137, + 230, + 479, + 254 + ], + "spans": [ + { + "bbox": [ + 137, + 230, + 479, + 254 + ], + "type": "text", + "content": "- CLIP and OpenCLIP: CLIP models are trained on WIT-400M [29] and LAION-2B [31] dataset." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 138, + 254, + 480, + 278 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 254, + 480, + 278 + ], + "spans": [ + { + "bbox": [ + 138, + 254, + 480, + 278 + ], + "type": "text", + "content": "- N-gram Vocabulary (Vocab): the set of all contiguous N word sequences extracted from a text corpus, such as the collection of all captions." + } + ] + } + ], + "index": 5 + } + ], + "sub_type": "text" + }, + { + "bbox": [ + 132, + 293, + 308, + 306 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 293, + 308, + 306 + ], + "spans": [ + { + "bbox": [ + 132, + 293, + 308, + 306 + ], + "type": "text", + "content": "4 Profiling LAION-2B Data" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 130, + 316, + 480, + 387 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 316, + 480, + 387 + ], + "spans": [ + { + "bbox": [ + 130, + 316, + 480, + 387 + ], + "type": "text", + "content": "To better profile the image-text pair data on a billion scale, we first cluster all the images based on CLIP features into 4,000 clusters and sort each cluster with CLIP scores. After obtaining all the cluster labels, we use the SOTA text spotting model [45] to get the visual text content on all the collected images. Finally, we aggregate all the model-predicted results and compare them with their corresponding captions to bring out our observations." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 132, + 402, + 279, + 414 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 402, + 279, + 414 + ], + "spans": [ + { + "bbox": [ + 132, + 402, + 279, + 414 + ], + "type": "text", + "content": "4.1 Implementation Details" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 130, + 418, + 481, + 478 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 418, + 481, + 478 + ], + "spans": [ + { + "bbox": [ + 130, + 418, + 481, + 478 + ], + "type": "text", + "content": "Clustering with CLIP Features: We first train K-Means (implemented by Faiss [18]) on the LAION-400M [32] subset using ViT-B-32 [10] CLIP features to speed up the clustering process. Due to the large memory consumption, we reduce the feature dimensions from 512 to 256 using PCA. Then, we partition the whole dataset using trained K-Means with the same feature extraction pipeline." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 130, + 479, + 481, + 611 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 479, + 481, + 611 + ], + "spans": [ + { + "bbox": [ + 130, + 479, + 481, + 611 + ], + "type": "text", + "content": "Text Spotting and Matching: To detect and recognize text across various scenes, we adopt DeepSolo [45] as our text spotting model and use the pre-trained checkpoints with the ViTAEv2-S [48] backbone in default setting. The output format of the text spotting model is a sequence of polygons of text location and their recognized characters. Despite its strong performance, we empirically find that DeepSolo can not handle the crowd scenes well (with more than 100 separate words) which is only a small proportion of the dataset (" + }, + { + "bbox": [ + 130, + 479, + 481, + 611 + ], + "type": "inline_equation", + "content": "\\sim 2\\%" + }, + { + "bbox": [ + 130, + 479, + 481, + 611 + ], + "type": "text", + "content": "). To identify the correlation between the spotted text and captions, we use Algorithm 1 to calculate the Co-Emb. text rate in each image-text pair. Considering the imperfect text spotting predictions might miss or misspell words, we also use Levenshtein distance to calculate the fuzzing similarity and reported in Tab. 1." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 132, + 624, + 379, + 635 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 624, + 379, + 635 + ], + "spans": [ + { + "bbox": [ + 132, + 624, + 379, + 635 + ], + "type": "text", + "content": "4.2 Statistic and Observations from LAION-2B" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 130, + 641, + 480, + 665 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 641, + 480, + 665 + ], + "spans": [ + { + "bbox": [ + 130, + 641, + 480, + 665 + ], + "type": "text", + "content": "The overall statistics of the 2 billion image-text pairs are reported in Tab. 1. In summary, the images embedded with visual text content reach a surprisingly" + } + ] + } + ], + "index": 13 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "spans": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "type": "text", + "content": "6" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 166, + 91, + 217, + 101 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 166, + 91, + 217, + 101 + ], + "spans": [ + { + "bbox": [ + 166, + 91, + 217, + 101 + ], + "type": "text", + "content": "Y. Lin et al." + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 5 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 157, + 114, + 455, + 207 + ], + "blocks": [ + { + "bbox": [ + 157, + 114, + 455, + 207 + ], + "lines": [ + { + "bbox": [ + 157, + 114, + 455, + 207 + ], + "spans": [ + { + "bbox": [ + 157, + 114, + 455, + 207 + ], + "type": "image", + "image_path": "2736101e6b6790673d162463d80f3b3fdd36658ea8255251ff60763cdfce8256.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_body" + } + ], + "index": 2 + }, + { + "type": "image", + "bbox": [ + 159, + 217, + 455, + 278 + ], + "blocks": [ + { + "bbox": [ + 177, + 209, + 436, + 217 + ], + "lines": [ + { + "bbox": [ + 177, + 209, + 436, + 217 + ], + "spans": [ + { + "bbox": [ + 177, + 209, + 436, + 217 + ], + "type": "text", + "content": "(a) The ratio of different OCR-oriented data types in LAION-2B clusters." + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 159, + 217, + 455, + 278 + ], + "lines": [ + { + "bbox": [ + 159, + 217, + 455, + 278 + ], + "spans": [ + { + "bbox": [ + 159, + 217, + 455, + 278 + ], + "type": "image", + "image_path": "fe2ca1d48b0669c1f8be8c553c27eb3b8e8658a10422d0f40e5ef845c3115eed.jpg" + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 162, + 279, + 452, + 289 + ], + "lines": [ + { + "bbox": [ + 162, + 279, + 452, + 289 + ], + "spans": [ + { + "bbox": [ + 162, + 279, + 452, + 289 + ], + "type": "text", + "content": "(b) Top CLIP score samples visualization from 50 clusters with ratio over " + }, + { + "bbox": [ + 162, + 279, + 452, + 289 + ], + "type": "inline_equation", + "content": "80\\%" + }, + { + "bbox": [ + 162, + 279, + 452, + 289 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 131, + 297, + 482, + 364 + ], + "lines": [ + { + "bbox": [ + 131, + 297, + 482, + 364 + ], + "spans": [ + { + "bbox": [ + 131, + 297, + 482, + 364 + ], + "type": "text", + "content": "Fig. 3: (a): Based on the OCR prediction results, the image-text pairs are divided into three types: image without visual embedded text content; the spotted text from the image has no concurrent text with the caption; the spotted text at least share one concurrent word with the caption. The clusters are merged from 4000 into 100 for a better view. (b): In the clusters with high ratio, the top CLIP score samples contain various text sources, such as posters, book covers, advertisements, and even slides." + } + ] + } + ], + "index": 6, + "angle": 0, + "type": "image_caption" + } + ], + "index": 4 + }, + { + "bbox": [ + 130, + 389, + 482, + 556 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 389, + 482, + 556 + ], + "spans": [ + { + "bbox": [ + 130, + 389, + 482, + 556 + ], + "type": "text", + "content": "high proportion of " + }, + { + "bbox": [ + 130, + 389, + 482, + 556 + ], + "type": "inline_equation", + "content": "54.60\\%" + }, + { + "bbox": [ + 130, + 389, + 482, + 556 + ], + "type": "text", + "content": " in the investigated data. Around " + }, + { + "bbox": [ + 130, + 389, + 482, + 556 + ], + "type": "inline_equation", + "content": "15\\%" + }, + { + "bbox": [ + 130, + 389, + 482, + 556 + ], + "type": "text", + "content": " of words in the dataset captions are Co-Emb. text, and the proportion of Co-Emb. text can further reach " + }, + { + "bbox": [ + 130, + 389, + 482, + 556 + ], + "type": "inline_equation", + "content": "30\\%" + }, + { + "bbox": [ + 130, + 389, + 482, + 556 + ], + "type": "text", + "content": " when considering the fuzzy matching results of the spotted text and captions. This suggests that CLIP models trained on these data might lead to a high bias toward text spotting. To better visualize the data distribution, we provide cluster-specific statics results and top CLIP score samples of text-dominated clusters in Fig. 3. We divide all images into 100 clusters based on visual similarity and visualize them according to the OCR results. Every cluster contains more or less images embedded with text. Combined with sample visualization, we observe that in the LAION collected data, the parrot captions cover various scenes. In the subsets of images embedded with text, around " + }, + { + "bbox": [ + 130, + 389, + 482, + 556 + ], + "type": "inline_equation", + "content": "60\\%" + }, + { + "bbox": [ + 130, + 389, + 482, + 556 + ], + "type": "text", + "content": " of captions at least precisely parrot one concurrent word (Co-Emb. Text Rate " + }, + { + "bbox": [ + 130, + 389, + 482, + 556 + ], + "type": "inline_equation", + "content": ">0" + }, + { + "bbox": [ + 130, + 389, + 482, + 556 + ], + "type": "text", + "content": ") appearing in the image. It suggests that the data collection pipeline of LAION [31] has a strong bias to introduce parrot captions from web data." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 130, + 558, + 482, + 666 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 558, + 482, + 666 + ], + "spans": [ + { + "bbox": [ + 130, + 558, + 482, + 666 + ], + "type": "text", + "content": "To better understand Co-Emb. Text, we provide a more thorough analysis of the word counting and text size of parrot captions. As shown in Fig. 4a, the results show that a large proportion of the Co-Emb. Text only takes a few words. However, we also find a large number of captions that are almost full parrot captions (see areas around the heatmap diagonal). Next, in Fig. 4b and Fig. 4c, we investigate the correlation between the size of concurrent words box in the image and CLIP score. The results show that the large text size does not usually lead to a higher score; meanwhile, the small text size can also dominate the score as the score can be significantly different after removing them. (Details" + } + ] + } + ], + "index": 8 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 274, + 91, + 447, + 102 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 274, + 91, + 447, + 102 + ], + "spans": [ + { + "bbox": [ + 274, + 91, + 447, + 102 + ], + "type": "text", + "content": "Parrot Captions Teach CLIP to Spot Text" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 474, + 92, + 480, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 474, + 92, + 480, + 100 + ], + "spans": [ + { + "bbox": [ + 474, + 92, + 480, + 100 + ], + "type": "text", + "content": "7" + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 6 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 157, + 114, + 455, + 208 + ], + "blocks": [ + { + "bbox": [ + 157, + 114, + 455, + 208 + ], + "lines": [ + { + "bbox": [ + 157, + 114, + 455, + 208 + ], + "spans": [ + { + "bbox": [ + 157, + 114, + 455, + 208 + ], + "type": "image", + "image_path": "d848e7a6b526e361e58fd20eacc31821b0c9601384979e11031caf73e76bad60.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 131, + 214, + 482, + 270 + ], + "lines": [ + { + "bbox": [ + 131, + 214, + 482, + 270 + ], + "spans": [ + { + "bbox": [ + 131, + 214, + 482, + 270 + ], + "type": "text", + "content": "Fig. 4: (a): The number of caption words and associated spotted concurrent words based on precise word matching. (b): Distribution of total box area of concurrent words in the image and its CLIP score. (c): Distribution of total box area of concurrent words and its relative CLIP score before and after removing them from the image. (d): Distribution of text size of the single concurrent word and other spotted word." + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_caption" + } + ], + "index": 2 + }, + { + "bbox": [ + 130, + 293, + 482, + 331 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 293, + 482, + 331 + ], + "spans": [ + { + "bbox": [ + 130, + 293, + 482, + 331 + ], + "type": "text", + "content": "of text removal described in Sec. 5.1). One possible reason is the text content and input resolution may matter more for CLIP. Moreover, we discover that the larger text is more likely to be parroted in the captions, as shown in Fig. 4d." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 131, + 348, + 375, + 363 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 131, + 348, + 375, + 363 + ], + "spans": [ + { + "bbox": [ + 131, + 348, + 375, + 363 + ], + "type": "text", + "content": "5 Inspecting Pre-Trained CLIP Models" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 130, + 374, + 482, + 495 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 374, + 482, + 495 + ], + "spans": [ + { + "bbox": [ + 130, + 374, + 482, + 495 + ], + "type": "text", + "content": "It is important to note that the LAION-2B dataset collection pipeline uses CLIP score from OpenAI's model to filter out the image-text pair below 0.28. Therefore, we inspect these two released CLIP models [29,31] to answer better why LAION data contains such a high proportion of parrot captions. Specifically, OpenAI's CLIP model is trained on the WIT dataset (out-of-domain model), and OpenCLIP is trained on LAION-2B (in-domain model). We first study whether the embedded text is the key factor in CLIP filtering by ablating the embedded text using text inpainting. Moreover, we further investigate whether the text spotting capacity prefers specific text content by examining synthetic images with Syn-Emb. text." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 131, + 513, + 349, + 525 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 131, + 513, + 349, + 525 + ], + "spans": [ + { + "bbox": [ + 131, + 513, + 349, + 525 + ], + "type": "text", + "content": "5.1 Ablation of Embedded Text Removal" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 130, + 533, + 482, + 629 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 533, + 482, + 629 + ], + "spans": [ + { + "bbox": [ + 130, + 533, + 482, + 629 + ], + "type": "text", + "content": "Text Removal via Inpainting: Given the OCR predicted results, we use the fast marching method [38] to inpaint the area of the spotted text polygons. We generate two versions of text removal results for each image with embedded text, i.e., All-Emb. text removal and Co-Emb. text removal, as the parrot caption prediction is imperfect due to the limitation of OCR models. We also generate random inpainting images with randomly sampled spotted text polygons from other images to ablate the distribution shift caused by image inpainting. Examples of the spotted text masks and inpainting results are shown in Fig. 2." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 130, + 630, + 482, + 666 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 630, + 482, + 666 + ], + "spans": [ + { + "bbox": [ + 130, + 630, + 482, + 666 + ], + "type": "text", + "content": "Results: Based on the OCR predicted results and text inpainting operations, we can obtain six types of LAION images, including " + }, + { + "bbox": [ + 130, + 630, + 482, + 666 + ], + "type": "inline_equation", + "content": "\\bullet" + }, + { + "bbox": [ + 130, + 630, + 482, + 666 + ], + "type": "text", + "content": "): images without any embedded text (OCR results are empty); " + }, + { + "bbox": [ + 130, + 630, + 482, + 666 + ], + "type": "inline_equation", + "content": "\\bullet" + }, + { + "bbox": [ + 130, + 630, + 482, + 666 + ], + "type": "text", + "content": "): images with any embedded text" + } + ] + } + ], + "index": 9 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "spans": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "type": "text", + "content": "8" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 166, + 91, + 217, + 100 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 166, + 91, + 217, + 100 + ], + "spans": [ + { + "bbox": [ + 166, + 91, + 217, + 100 + ], + "type": "text", + "content": "Y. Lin et al." + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 7 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 133, + 115, + 479, + 223 + ], + "blocks": [ + { + "bbox": [ + 133, + 115, + 479, + 223 + ], + "lines": [ + { + "bbox": [ + 133, + 115, + 479, + 223 + ], + "spans": [ + { + "bbox": [ + 133, + 115, + 479, + 223 + ], + "type": "image", + "image_path": "a29d6bf2527dbbbbe5d619be17b238ed6c0779b885764f5faf754d369489b5dc.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 130, + 231, + 482, + 266 + ], + "lines": [ + { + "bbox": [ + 130, + 231, + 482, + 266 + ], + "spans": [ + { + "bbox": [ + 130, + 231, + 482, + 266 + ], + "type": "text", + "content": "Fig. 5: Left: Mean CLIP scores of image-text pairs with different text removal operations depicted in Sec. 5.1, and grouped by cluster the same as Fig. 3. Right: Overall relative CLIP score distribution by comparing different text removal operations." + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_caption" + } + ], + "index": 2 + }, + { + "type": "table", + "bbox": [ + 222, + 298, + 394, + 367 + ], + "blocks": [ + { + "bbox": [ + 161, + 276, + 451, + 288 + ], + "lines": [ + { + "bbox": [ + 161, + 276, + 451, + 288 + ], + "spans": [ + { + "bbox": [ + 161, + 276, + 451, + 288 + ], + "type": "text", + "content": "Table 2: Mean CLIP score of different setups of text removal." + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 222, + 298, + 394, + 367 + ], + "lines": [ + { + "bbox": [ + 222, + 298, + 394, + 367 + ], + "spans": [ + { + "bbox": [ + 222, + 298, + 394, + 367 + ], + "type": "table", + "html": "
SetupAverage CLIP Score
Raw Images w/o Emb. Text0.3223 ± 0.0078
Raw Images w/ Emb. Text0.3358 ± 0.0094
× All-Emb. Random Inpainted0.3260 ± 0.0057
× All-Emb. Text Removal0.2974 ± 0.0197
Co-Emb. Random Inpainted0.3341 ± 0.0051
Co-Emb. Text Removal0.2993 ± 0.0146
", + "image_path": "061577db60f59d0a41587b5e8aab0d1a1cb1f737956cdb8bb4ae5dd6c27838e6.jpg" + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "table_body" + } + ], + "index": 5 + }, + { + "bbox": [ + 130, + 390, + 482, + 629 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 390, + 482, + 629 + ], + "spans": [ + { + "bbox": [ + 130, + 390, + 482, + 629 + ], + "type": "text", + "content": "(OCR results are none-empty); " + }, + { + "bbox": [ + 130, + 390, + 482, + 629 + ], + "type": "inline_equation", + "content": "\\times" + }, + { + "bbox": [ + 130, + 390, + 482, + 629 + ], + "type": "text", + "content": "): images with random inpainting by other image's All-Emb. text area; " + }, + { + "bbox": [ + 130, + 390, + 482, + 629 + ], + "type": "inline_equation", + "content": "\\times" + }, + { + "bbox": [ + 130, + 390, + 482, + 629 + ], + "type": "text", + "content": "): images removed All-Emb. text (Inpaint all the areas of OCR predicted text); ■): images randomly inpainted by other image's Co-Emb. text area, and ■): images removed Co-Emb. text (Inpaint the areas of concurrent text in OCR predicted text and captions). Then, we calculate CLIP scores of all the groups of images and their paired captions using OpenAI released CLIP model (ViT-B-32). Fig. 5 reports the mean scores of different types of images in each cluster and raises four observations as follows: I). The images embedded with text achieve higher CLIP scores in most clusters than those without embedded text; II). CLIP scores significantly drop once we remove the text from the images compared to its random inpainting baseline. It indicates that the parrot captions correlate highly with CLIP score measurement; III). Not all the samples are dominated by the embedded text, as some samples achieve higher scores after removing text, indicating the embedded text can also be a distractor; IV). Most of the relative CLIP scores (" + }, + { + "bbox": [ + 130, + 390, + 482, + 629 + ], + "type": "inline_equation", + "content": "S(\\square) - S(\\times)" + }, + { + "bbox": [ + 130, + 390, + 482, + 629 + ], + "type": "text", + "content": ") between images removed Co-Emb. text and All-Emb. text are positive, as shown in the right of Fig. 5. The straightforward reason is the images lose more visual information due to the larger in-painting area. Another possible reason is the imperfect text spotting prediction or the corner cases in the matching algorithm leaking parts of the concurrent text in images." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 130, + 629, + 482, + 665 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 629, + 482, + 665 + ], + "spans": [ + { + "bbox": [ + 130, + 629, + 482, + 665 + ], + "type": "text", + "content": "Discussion: Due to the text removal, the image distribution may shift from the CLIP training set. Therefore, we provide two random inpainting baselines to examine the effect of distribution shift. In Tab. 2, we report the mean scores" + } + ] + } + ], + "index": 7 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 274, + 91, + 447, + 102 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 274, + 91, + 447, + 102 + ], + "spans": [ + { + "bbox": [ + 274, + 91, + 447, + 102 + ], + "type": "text", + "content": "Parrot Captions Teach CLIP to Spot Text" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 474, + 92, + 480, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 474, + 92, + 480, + 100 + ], + "spans": [ + { + "bbox": [ + 474, + 92, + 480, + 100 + ], + "type": "text", + "content": "9" + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 8 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 138, + 114, + 217, + 211 + ], + "blocks": [ + { + "bbox": [ + 138, + 114, + 217, + 211 + ], + "lines": [ + { + "bbox": [ + 138, + 114, + 217, + 211 + ], + "spans": [ + { + "bbox": [ + 138, + 114, + 217, + 211 + ], + "type": "image", + "image_path": "b2bf5fc65d8a33143afa94208fd31f2580b7e440c2201ac5e439a8d6211cc792.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 159, + 213, + 286, + 221 + ], + "lines": [ + { + "bbox": [ + 159, + 213, + 286, + 221 + ], + "spans": [ + { + "bbox": [ + 159, + 213, + 286, + 221 + ], + "type": "text", + "content": "(a) Statistic of 1-gram vocabularies" + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "image_caption" + } + ], + "index": 2 + }, + { + "type": "image", + "bbox": [ + 217, + 114, + 297, + 211 + ], + "blocks": [ + { + "bbox": [ + 217, + 114, + 297, + 211 + ], + "lines": [ + { + "bbox": [ + 217, + 114, + 297, + 211 + ], + "spans": [ + { + "bbox": [ + 217, + 114, + 297, + 211 + ], + "type": "image", + "image_path": "c51adc3c2e4cc81578b7a958b2f93d739506d8eb4c6238d7eba7dcd6e3abab14.jpg" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_body" + } + ], + "index": 3 + }, + { + "type": "image", + "bbox": [ + 138, + 222, + 306, + 314 + ], + "blocks": [ + { + "bbox": [ + 138, + 222, + 306, + 314 + ], + "lines": [ + { + "bbox": [ + 138, + 222, + 306, + 314 + ], + "spans": [ + { + "bbox": [ + 138, + 222, + 306, + 314 + ], + "type": "image", + "image_path": "eb6efd676f718ce07bca9268be2594cf7fe0ba293c6127178a6a55beab9d8a57.jpg" + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 153, + 315, + 291, + 324 + ], + "lines": [ + { + "bbox": [ + 153, + 315, + 291, + 324 + ], + "spans": [ + { + "bbox": [ + 153, + 315, + 291, + 324 + ], + "type": "text", + "content": "(c) Grouped by Caption 1-gram length" + } + ] + } + ], + "index": 6, + "angle": 0, + "type": "image_caption" + } + ], + "index": 5 + }, + { + "type": "image", + "bbox": [ + 306, + 118, + 477, + 212 + ], + "blocks": [ + { + "bbox": [ + 306, + 118, + 477, + 212 + ], + "lines": [ + { + "bbox": [ + 306, + 118, + 477, + 212 + ], + "spans": [ + { + "bbox": [ + 306, + 118, + 477, + 212 + ], + "type": "image", + "image_path": "3bd8ffa9f9641b7a2f3363fedd8e97c6c8b90db62725fb5f5f55c10dba180ed4.jpg" + } + ] + } + ], + "index": 7, + "angle": 0, + "type": "image_body" + } + ], + "index": 7 + }, + { + "type": "image", + "bbox": [ + 306, + 221, + 476, + 314 + ], + "blocks": [ + { + "bbox": [ + 315, + 213, + 468, + 221 + ], + "lines": [ + { + "bbox": [ + 315, + 213, + 468, + 221 + ], + "spans": [ + { + "bbox": [ + 315, + 213, + 468, + 221 + ], + "type": "text", + "content": "(b) Grouped by 1-gram frequency intervals" + } + ] + } + ], + "index": 8, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 306, + 221, + 476, + 314 + ], + "lines": [ + { + "bbox": [ + 306, + 221, + 476, + 314 + ], + "spans": [ + { + "bbox": [ + 306, + 221, + 476, + 314 + ], + "type": "image", + "image_path": "6b01f1eb543c6ae006cbca61d13aee87e3471d2c1b3c78d85340378d0c7bc211.jpg" + } + ] + } + ], + "index": 9, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 321, + 315, + 461, + 324 + ], + "lines": [ + { + "bbox": [ + 321, + 315, + 461, + 324 + ], + "spans": [ + { + "bbox": [ + 321, + 315, + 461, + 324 + ], + "type": "text", + "content": "(d) Grouped by Caption 2-gram length" + } + ] + } + ], + "index": 10, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 130, + 334, + 482, + 367 + ], + "lines": [ + { + "bbox": [ + 130, + 334, + 482, + 367 + ], + "spans": [ + { + "bbox": [ + 130, + 334, + 482, + 367 + ], + "type": "text", + "content": "Fig.6: OpenCLIP more bias towards text spotting than CLIP model. Grouped score distributions of prompting CLIP and OpenCLIP models with N-gram Syn-Emb. text and synthetic images for model preference investigation." + } + ] + } + ], + "index": 11, + "angle": 0, + "type": "image_caption" + } + ], + "index": 9 + }, + { + "bbox": [ + 130, + 390, + 482, + 427 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 390, + 482, + 427 + ], + "spans": [ + { + "bbox": [ + 130, + 390, + 482, + 427 + ], + "type": "text", + "content": "of different setups. Results show that the random baselines are very close to the raw image baseline, indicating that CLIP model is robust to the distribution shift caused by information loss in inpainted regions." + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 131, + 443, + 321, + 455 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 131, + 443, + 321, + 455 + ], + "spans": [ + { + "bbox": [ + 131, + 443, + 321, + 455 + ], + "type": "text", + "content": "5.2 Prompting with Syn-Emb. Text" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 130, + 462, + 482, + 593 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 462, + 482, + 593 + ], + "spans": [ + { + "bbox": [ + 130, + 462, + 482, + 593 + ], + "type": "text", + "content": "Generating Synthetic Images from N-gram Vocabulary: To investigate CLIP models' text spotting preference, we adopt a similar strategy in [25] to use synthetic images to embed specific text content by rendering text in a blank background. For each text, we use four fore-background style rendering templates (black-white, black-grey, white-grey, and white-black), as shown in Fig. 2. Different from the uniformly sampling letters in [25], we generate the text content from the N-gram vocabulary built from captions and Co-Emb. text to study the text spotting pattern. We only select the top frequent 400,000 grams for each vocabulary. The statistics of 1-gram vocabulary are shown in Fig. 6a, which is a long-tail distribution. Next, we calculate the synthetic images and the rendered text similarity on released ViT-B-32 CLIP and OpenCLIP models." + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 130, + 594, + 482, + 665 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 594, + 482, + 665 + ], + "spans": [ + { + "bbox": [ + 130, + 594, + 482, + 665 + ], + "type": "text", + "content": "Results: Firstly, we examine whether CLIP models prefer recognizing more commonly seen words (with high frequency in vocabulary). Therefore, we group the 1-gram results based on their frequency interval in the whole vocabulary, as shown in Fig. 6b. The results show that OpenCLIP model clearly has a stronger text spotting capacity than CLIP, i.e., more biased towards text spotting. We also observe that all CLIP models are more sensitive to the vocabulary built" + } + ] + } + ], + "index": 15 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "spans": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "type": "text", + "content": "10" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 166, + 91, + 217, + 100 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 166, + 91, + 217, + 100 + ], + "spans": [ + { + "bbox": [ + 166, + 91, + 217, + 100 + ], + "type": "text", + "content": "Y. Lin et al." + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 9 + }, + { + "para_blocks": [ + { + "bbox": [ + 130, + 116, + 482, + 236 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 116, + 482, + 236 + ], + "spans": [ + { + "bbox": [ + 130, + 116, + 482, + 236 + ], + "type": "text", + "content": "from the concurrent words. Interestingly, both CLIP and OpenCLIP models have slightly higher scores on the less frequent grams. Secondly, considering the long-tail grams might contain more characters, we further group the 1-gram and 2-gram results based on their text length in Fig. 6c and Fig. 6d. Note that the Co-Emb. text is not regularly arranged in the images, making it hard to extract continuous word sequences. Results show that all the models are better at spotting the longer words, possibly due to the tokenizer used in the text encoder, making them more discriminative. Meanwhile, in the groups of 2-gram samples, the scores gradually drop when spotting the highly long text, indicating that the spotting capacity of CLIP models is possibly built on word-by-word." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 131, + 251, + 375, + 264 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 131, + 251, + 375, + 264 + ], + "spans": [ + { + "bbox": [ + 131, + 251, + 375, + 264 + ], + "type": "text", + "content": "6 Training on Emb. Text Curated Data" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 130, + 274, + 479, + 297 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 274, + 479, + 297 + ], + "spans": [ + { + "bbox": [ + 130, + 274, + 479, + 297 + ], + "type": "text", + "content": "Next, we dive deeper into training CLIP and BLIP [21] models on different Emb. text curated subsets and studying various downstream task behaviors." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 130, + 298, + 482, + 380 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 298, + 482, + 380 + ], + "spans": [ + { + "bbox": [ + 130, + 298, + 482, + 380 + ], + "type": "text", + "content": "Implementation Details: For CLIP, we use the open-source software OpenCLIP [16] for all CLIP model training. Our experiments are conducted on both ViT-B [10] and RN50 [15]. We use 4,096 batch size for 3M and 8,192 for 12M scale subsets. Other settings remain the same as [31]. For BLIP, we mainly conduct on 3M scale subsets with ViT-B [10]. The BLIP models are pre-trained for 10 epochs with an AdamW [23] optimizer. For downstream tasks, we finetune 10 epochs for VQA and 5 epochs for captioning and retrieval." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 130, + 382, + 482, + 609 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 382, + 482, + 609 + ], + "spans": [ + { + "bbox": [ + 130, + 382, + 482, + 609 + ], + "type": "text", + "content": "Evaluation: For CLIP, we follow the DataComp benchmark [11] using 38 zero-shot classification and retrieval tasks as evaluation. We report the average performance (Avg.) of the DataComp benchmark and two subset track performances, ImageNet (IN) and Retrieval (Ret.). To evaluate the text spotting capacity, we use the synthetic benchmark illustrated in Sec. 5.2 and a real-world benchmark sampled from LAION-2B as the validation set. In the synthetic benchmark, we calculate the similarity of all the 1-gram synthetic image-text pairs from caption vocabulary and report all the trained model results in Fig. 7. For the real-world benchmark, we sample 1M image-text pairs without any embedded text and 1M samples dominated by the parrot caption (RSC " + }, + { + "bbox": [ + 130, + 382, + 482, + 609 + ], + "type": "inline_equation", + "content": "\\geq 0.2" + }, + { + "bbox": [ + 130, + 382, + 482, + 609 + ], + "type": "text", + "content": "). Fig. 8 aggregates the mean scores of the 2M evaluation set and also reports the mean scores of applying text removal on the 2M evaluation set results. For BLIP, inspired by [12], we further evaluate the model behavior on downstream tasks requiring reading text, including Visual Question Answering (VQA), Image Captioning, and Text-Image Retrieval. Specifically, for the text-oriented tasks, we use Text VQA [36] and ST-VQA [5] for VQA, and TextCaps [35] for captioning and retrieval. Moreover, we also provide the same tasks on the datasets that only require the model to see, i.e., the natural image dataset. Similarly, we use VQAv2 [14] for VQA and COCO [8] for captioning and retrieval." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 131, + 624, + 375, + 635 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 131, + 624, + 375, + 635 + ], + "spans": [ + { + "bbox": [ + 131, + 624, + 375, + 635 + ], + "type": "text", + "content": "6.1 Ablation Study of Data Curation on CLIP" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 130, + 641, + 481, + 665 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 641, + 481, + 665 + ], + "spans": [ + { + "bbox": [ + 130, + 641, + 481, + 665 + ], + "type": "text", + "content": "Curation I: Embedded Text in Images. To study the impact of embedded text on overall pre-train data quality, we sample three subsets: random baseline," + } + ] + } + ], + "index": 8 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 274, + 91, + 447, + 102 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 274, + 91, + 447, + 102 + ], + "spans": [ + { + "bbox": [ + 274, + 91, + 447, + 102 + ], + "type": "text", + "content": "Parrot Captions Teach CLIP to Spot Text" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 470, + 92, + 479, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 470, + 92, + 479, + 100 + ], + "spans": [ + { + "bbox": [ + 470, + 92, + 479, + 100 + ], + "type": "text", + "content": "11" + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 10 + }, + { + "para_blocks": [ + { + "type": "table", + "bbox": [ + 140, + 147, + 280, + 244 + ], + "blocks": [ + { + "bbox": [ + 132, + 114, + 290, + 136 + ], + "lines": [ + { + "bbox": [ + 132, + 114, + 290, + 136 + ], + "spans": [ + { + "bbox": [ + 132, + 114, + 290, + 136 + ], + "type": "text", + "content": "Table 3: Ablation of images embedded with or without text." + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 140, + 147, + 280, + 244 + ], + "lines": [ + { + "bbox": [ + 140, + 147, + 280, + 244 + ], + "spans": [ + { + "bbox": [ + 140, + 147, + 280, + 244 + ], + "type": "table", + "html": "
DataModelINRet.Avg.
3M RandomRN500.2040.2220.256
3M w/o Emb. TextRN500.2280.2320.282
3M w/ Emb. Text OnlyRN500.0710.1390.164
3M RandomViT-B0.1310.1480.210
3M w/o Emb. TextViT-B0.1620.1640.234
3M w/ Emb. Text OnlyViT-B0.0520.1110.153
12M RandomRN500.3600.3540.354
12M w/o Emb. TextRN500.4090.3610.372
12M w/ Emb. Text OnlyRN500.1290.1920.218
12M RandomViT-B0.3140.2990.351
12M w/o Emb. TextViT-B0.3700.3180.364
12M w/ Emb. Text OnlyViT-B0.1290.1720.225
", + "image_path": "4f003cbb7dc6fbf71a9d6596437895b94cfb1f7052f74b0b57b8fc27088416fa.jpg" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "table_body" + } + ], + "index": 3 + }, + { + "type": "table", + "bbox": [ + 145, + 278, + 276, + 362 + ], + "blocks": [ + { + "bbox": [ + 132, + 246, + 290, + 268 + ], + "lines": [ + { + "bbox": [ + 132, + 246, + 290, + 268 + ], + "spans": [ + { + "bbox": [ + 132, + 246, + 290, + 268 + ], + "type": "text", + "content": "Table 5: Ablation of models trained on subsets sampled by different RSA." + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 145, + 278, + 276, + 362 + ], + "lines": [ + { + "bbox": [ + 145, + 278, + 276, + 362 + ], + "spans": [ + { + "bbox": [ + 145, + 278, + 276, + 362 + ], + "type": "table", + "html": "
Data (3M)ModelAvg.S(●)INRet.Avg.
RSA < 0.0RN500.3190.1810.2200.239
RSA ≥ 0.0RN500.3390.1260.1800.215
RSA ≥ 0.1RN500.3510.0410.1230.148
RSA ≥ 0.2RN500.3600.0170.0940.109
RSA ≥ 0.3RN500.3760.0090.0750.097
RSA < 0.0ViT-B0.3190.1230.1590.198
RSA ≥ 0.0ViT-B0.3390.0790.1290.185
RSA ≥ 0.1ViT-B0.3510.0310.1030.134
RSA ≥ 0.2ViT-B0.3600.0120.0800.103
RSA ≥ 0.3ViT-B0.3760.0060.0700.096
", + "image_path": "5474333052371dfaba6663fb90058a51f8cb5d2836591f51df47042e4c816e4d.jpg" + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "table_body" + } + ], + "index": 5 + }, + { + "type": "table", + "bbox": [ + 345, + 150, + 461, + 241 + ], + "blocks": [ + { + "bbox": [ + 322, + 118, + 480, + 140 + ], + "lines": [ + { + "bbox": [ + 322, + 118, + 480, + 140 + ], + "spans": [ + { + "bbox": [ + 322, + 118, + 480, + 140 + ], + "type": "text", + "content": "Table 4: Ablation of different CoEmb. Text Rate (CoTR)." + } + ] + } + ], + "index": 6, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 345, + 150, + 461, + 241 + ], + "lines": [ + { + "bbox": [ + 345, + 150, + 461, + 241 + ], + "spans": [ + { + "bbox": [ + 345, + 150, + 461, + 241 + ], + "type": "table", + "html": "
Data (3M)ModelINRet.Avg.
CoTR = 0.0RN500.1930.2290.247
CoTR ≥ 0.3RN500.0310.1100.137
CoTR ≥ 0.5RN500.0210.0990.124
CoTR ≥ 0.8RN500.0120.0820.096
CoTR = 1.0RN500.0120.0740.102
CoTR = 0.0ViT-B0.1320.1640.206
CoTR ≥ 0.3ViT-B0.0290.0840.130
CoTR ≥ 0.5ViT-B0.0210.0820.119
CoTR ≥ 0.8ViT-B0.0120.0760.104
CoTR = 1.0ViT-B0.0130.0760.103
", + "image_path": "dd92dbe4cc1891e4dddbf8e44c34a9bab7b976826668ab59761dfd3df12cb2f5.jpg" + } + ] + } + ], + "index": 7, + "angle": 0, + "type": "table_body" + } + ], + "index": 7 + }, + { + "type": "table", + "bbox": [ + 335, + 278, + 465, + 362 + ], + "blocks": [ + { + "bbox": [ + 322, + 246, + 480, + 268 + ], + "lines": [ + { + "bbox": [ + 322, + 246, + 480, + 268 + ], + "spans": [ + { + "bbox": [ + 322, + 246, + 480, + 268 + ], + "type": "text", + "content": "Table 6: Ablation of models trained on subsets sampled by different RSC." + } + ] + } + ], + "index": 8, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 335, + 278, + 465, + 362 + ], + "lines": [ + { + "bbox": [ + 335, + 278, + 465, + 362 + ], + "spans": [ + { + "bbox": [ + 335, + 278, + 465, + 362 + ], + "type": "table", + "html": "
Data (3M)ModelAvg.S(●)INRet.Avg.
RSC < 0.0RN500.3260.1250.1710.209
RSC ≥ 0.0RN500.3450.0620.1290.168
RSC ≥ 0.1RN500.3540.0140.0910.106
RSC ≥ 0.2RN500.3640.0080.0840.104
RSC ≥ 0.3RN500.3800.0050.0580.084
RSC < 0.0ViT-B0.3260.0790.1290.174
RSC ≥ 0.0ViT-B0.3450.0450.1190.149
RSC ≥ 0.1ViT-B0.3540.0180.0910.116
RSC ≥ 0.2ViT-B0.3640.0080.0760.106
RSC ≥ 0.3ViT-B0.3800.0040.0590.091
", + "image_path": "c652c95b3895d074f0dbd2f46f7ca21b700af7f6ea40c82e261f9145b0b77110.jpg" + } + ] + } + ], + "index": 9, + "angle": 0, + "type": "table_body" + } + ], + "index": 9 + }, + { + "bbox": [ + 130, + 388, + 480, + 460 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 388, + 480, + 460 + ], + "spans": [ + { + "bbox": [ + 130, + 388, + 480, + 460 + ], + "type": "text", + "content": "images without any embedded text, and images all embedded with text from LAION-2B. The subsets include 3M and 12M scales. The results in Tab. 3 show that images embedded with text generally reduce the pre-training dataset quality as all performance tracks significantly decrease. Meanwhile, in Fig. 7, the model trained with the images embedded with text achieves the strongest text spotting capacity compared to the random and images without text baselines." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 130, + 460, + 481, + 592 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 460, + 481, + 592 + ], + "spans": [ + { + "bbox": [ + 130, + 460, + 481, + 592 + ], + "type": "text", + "content": "Curation II: Co-Emb. Text Rate (CoTR). Tab. 3 reports CLIP models trained on parrot captions with different CoTR. We first select all the images with embedded text and then sample images based on the CoTR depicted at Algorithm 1 with different thresholds. With increasing CoTR, all the zero-shot benchmark performance drops significantly. Despite the images in the subset " + }, + { + "bbox": [ + 130, + 460, + 481, + 592 + ], + "type": "inline_equation", + "content": "(\\mathrm{CoTR} = 0)" + }, + { + "bbox": [ + 130, + 460, + 481, + 592 + ], + "type": "text", + "content": " all embedded with text, the pre-trained model performs similarly to the random baseline in Tab. 3. It indicates that the parrot caption is more crucial than embedded text in reducing the pre-trained data quality. For the text spotting capacity, Fig. 7 and Fig. 8 show that the increasing CoTR does not lead to stronger text spotting capacity, possibly due to the average length of captions decreasing in higher CoTR data." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 130, + 593, + 481, + 666 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 593, + 481, + 666 + ], + "spans": [ + { + "bbox": [ + 130, + 593, + 481, + 666 + ], + "type": "text", + "content": "Curation III: Relative Score from Text Removal (RSA & RSC). Given the observations in Sec. 5.1, we further select a series of subsets based on the relative score of images before and after text removal. The higher relative scores mean the masked text is more dominant in CLIP score measurement. In Tab. 5 and Tab. 6, we report the zero-shot performance of models trained on subsets with different relative score thresholds. CLIP models pre-trained with higher" + } + ] + } + ], + "index": 12 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "spans": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "type": "text", + "content": "12" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 166, + 91, + 217, + 100 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 166, + 91, + 217, + 100 + ], + "spans": [ + { + "bbox": [ + 166, + 91, + 217, + 100 + ], + "type": "text", + "content": "Y. Lin et al." + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 11 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 149, + 114, + 462, + 205 + ], + "blocks": [ + { + "bbox": [ + 149, + 114, + 462, + 205 + ], + "lines": [ + { + "bbox": [ + 149, + 114, + 462, + 205 + ], + "spans": [ + { + "bbox": [ + 149, + 114, + 462, + 205 + ], + "type": "image", + "image_path": "64665d5bd07426a917b953a91bf6c10ea5b1dc10604bd1eb648f2e6076f8ea24.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 131, + 213, + 482, + 247 + ], + "lines": [ + { + "bbox": [ + 131, + 213, + 482, + 247 + ], + "spans": [ + { + "bbox": [ + 131, + 213, + 482, + 247 + ], + "type": "text", + "content": "Fig. 7: CLIP models learn text spotting well from parrot captions. Benchmarking text spotting capacity of CLIP models with 1-gram caption vocabulary synthetic images dataset as the same as Sec. 5.2." + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_caption" + } + ], + "index": 2 + }, + { + "type": "image", + "bbox": [ + 209, + 258, + 405, + 364 + ], + "blocks": [ + { + "bbox": [ + 209, + 258, + 405, + 364 + ], + "lines": [ + { + "bbox": [ + 209, + 258, + 405, + 364 + ], + "spans": [ + { + "bbox": [ + 209, + 258, + 405, + 364 + ], + "type": "image", + "image_path": "a21f85ab1ad31fb86cc26962da01e2a2fbf9ca466049c21a88c2a0e29ac1974f.jpg" + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 131, + 371, + 482, + 404 + ], + "lines": [ + { + "bbox": [ + 131, + 371, + 482, + 404 + ], + "spans": [ + { + "bbox": [ + 131, + 371, + 482, + 404 + ], + "type": "text", + "content": "Fig. 8: Text spotting capacity validation on real images. Models trained with more parrot captions are better at aligning the image with parrot captions but perform worse at aligning images without embedded text." + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "image_caption" + } + ], + "index": 4 + }, + { + "bbox": [ + 130, + 430, + 482, + 525 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 430, + 482, + 525 + ], + "spans": [ + { + "bbox": [ + 130, + 430, + 482, + 525 + ], + "type": "text", + "content": "RSA or RSC both get worse downstream performance. Notably, the average raw CLIP scores " + }, + { + "bbox": [ + 130, + 430, + 482, + 525 + ], + "type": "inline_equation", + "content": "\\mathrm{S}(\\bullet)" + }, + { + "bbox": [ + 130, + 430, + 482, + 525 + ], + "type": "text", + "content": " of these subsets have a positive correlation with RSA or RSC, indicating using CLIP scores from a biased pre-trained model as the data filtering strategy can be unreliable. When comparing the RSA and RSC, the results show that the samples dominated by the latter, i.e., parrot captions, are less informative for CLIP training. Meanwhile, Fig. 7 and Fig. 8 show that the text spotting capacity of CLIP can be further improved by training on the samples using relative scores as data curation criteria against CoTR." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 131, + 546, + 378, + 558 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 131, + 546, + 378, + 558 + ], + "spans": [ + { + "bbox": [ + 131, + 546, + 378, + 558 + ], + "type": "text", + "content": "6.2 More Investigation on Text-Oriented Tasks" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 130, + 570, + 482, + 665 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 570, + 482, + 665 + ], + "spans": [ + { + "bbox": [ + 130, + 570, + 482, + 665 + ], + "type": "text", + "content": "Inspired by [12], we further train BLIP [21] models on the curated subsets to better understand the impact of parrot caption on various vision-language tasks, especially the tasks that require models to read the text. We chose BLIP for the ablation study instead of CLIP as it can be directly applied to all these tasks. As shown in Tab. 7, training BLIPs to spot text can boost their performance on the downstream tasks requiring the model to read but impede the performance of downstream tasks only requiring the model to see, which are consistent with the observation on classification tasks. Nevertheless, when BLIPs mainly focus" + } + ] + } + ], + "index": 8 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 274, + 91, + 447, + 102 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 274, + 91, + 447, + 102 + ], + "spans": [ + { + "bbox": [ + 274, + 91, + 447, + 102 + ], + "type": "text", + "content": "Parrot Captions Teach CLIP to Spot Text" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 470, + 91, + 480, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 470, + 91, + 480, + 100 + ], + "spans": [ + { + "bbox": [ + 470, + 91, + 480, + 100 + ], + "type": "text", + "content": "13" + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 12 + }, + { + "para_blocks": [ + { + "type": "table", + "bbox": [ + 152, + 158, + 463, + 327 + ], + "blocks": [ + { + "bbox": [ + 132, + 114, + 480, + 148 + ], + "lines": [ + { + "bbox": [ + 132, + 114, + 480, + 148 + ], + "spans": [ + { + "bbox": [ + 132, + 114, + 480, + 148 + ], + "type": "text", + "content": "Table 7: BLIP downstream tasks performance of pre-training on different curated 3M subsets. The gray color represents tasks requiring the model to read the text from images, i.e., spotting text from images." + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 152, + 158, + 463, + 327 + ], + "lines": [ + { + "bbox": [ + 152, + 158, + 463, + 327 + ], + "spans": [ + { + "bbox": [ + 152, + 158, + 463, + 327 + ], + "type": "table", + "html": "
BLIP Data (3M)Visual Question Answering (Acc)Image Captioning (CIDEr)Text-to-Image Retrieval (R@1)Image-to-Text Retrieval (R@1)
VQAv2TextVQAST-VQACOCOTextCapsCOCOTextCapsCOCOTextCaps
Rand71.0715.3610.48115.653.748.9156.3465.4672.45
w/ Emb. Text68.9419.0512.65108.992.142.8970.158.581.42
w/o Emb. Text71.2213.659.29116.241.549.9631.8366.548.7
CoTR = 0.071.1113.979.75116.344.649.5538.0566.0854.57
CoTR ≥ 0.367.419.2811.81104.996.937.7867.2851.9878.2
CoTR ≥ 0.567.0219.6412.38102.794.135.9465.2450.3276.94
CoTR ≥ 0.866.3818.5012.00100.991.634.1362.6546.973.56
CoTR = 1.066.1818.4712.80101.291.333.5561.8346.6273.05
RSA < 0.070.7914.169.64115.744.948.2536.8564.7254.7
RSA ≥ 0.070.0318.7611.81111.984.546.2568.6162.9281.23
RSA ≥ 0.168.1419.4813.33105.696.139.9668.1354.6479.37
RSA ≥ 0.266.0121.0611.8598.794.433.0364.1747.1275.33
RSA ≥ 0.364.2018.4412.0495.2691.126.6460.1137.370.24
RSC < 0.070.1315.1910.74112.246.746.841.9563.2458.05
RSC ≥ 0.068.8620.1213.75107.893.542.069.7857.4280.92
RSC ≥ 0.167.3520.5412.84103.496.936.466.6951.0277.79
RSC ≥ 0.262.6220.3213.1498.792.830.0861.3842.9671.98
RSC ≥ 0.363.7518.9413.0392.988.724.2358.3534.7268.95
", + "image_path": "ca4a6e51c0d7557d7987b0094c3886ffe647f5e9516e01ab77120afec38ed653.jpg" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "table_body" + } + ], + "index": 3 + }, + { + "bbox": [ + 132, + 347, + 480, + 407 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 347, + 480, + 407 + ], + "spans": [ + { + "bbox": [ + 132, + 347, + 480, + 407 + ], + "type": "text", + "content": "on reading, e.g. (RSA " + }, + { + "bbox": [ + 132, + 347, + 480, + 407 + ], + "type": "inline_equation", + "content": "\\geq 0.3" + }, + { + "bbox": [ + 132, + 347, + 480, + 407 + ], + "type": "text", + "content": "), their text-oriented and natural downstream performance also decreases. In other words, the parrot captions can benefit the text-orient downstream tasks while requiring careful data mixing trade-off. We believe that understanding parrot captions is essential for revealing the underlying mechanisms of existing large vision-language systems." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 133, + 423, + 310, + 435 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 423, + 310, + 435 + ], + "spans": [ + { + "bbox": [ + 133, + 423, + 310, + 435 + ], + "type": "text", + "content": "7 Conclusion and Discussion" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 132, + 446, + 481, + 590 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 446, + 481, + 590 + ], + "spans": [ + { + "bbox": [ + 132, + 446, + 481, + 590 + ], + "type": "text", + "content": "The popularity of vision-language contrastive loss stems from its efficiency and simplicity. However, the analysis and experiments we presented show that the embedded text in images and their parrot captions plant significant text spotting bias due to such contrastive fashions. Firstly, almost " + }, + { + "bbox": [ + 132, + 446, + 481, + 590 + ], + "type": "inline_equation", + "content": "30\\%" + }, + { + "bbox": [ + 132, + 446, + 481, + 590 + ], + "type": "text", + "content": " of the captions in the widely used LAION-2B dataset are biased towards parroting the embedded text in images. Secondly, the pre-trained CLIP models have strong preferences for the image-text pair with parrot captions, which achieve higher similarity scores than those without. Finally, using data biasing to parrot captions, we can easily train a CLIP model with a strong text spotting bias. Our work demonstrates the emergency of reviewing the impact of parrot captions in the entire ecosystem of CLIP models. Our future endeavor involves building a bias-aware data curation pipeline and a robust training function to mitigate such issues." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 133, + 606, + 246, + 620 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 606, + 246, + 620 + ], + "spans": [ + { + "bbox": [ + 133, + 606, + 246, + 620 + ], + "type": "text", + "content": "Acknowledgements" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 132, + 629, + 480, + 665 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 629, + 480, + 665 + ], + "spans": [ + { + "bbox": [ + 132, + 629, + 480, + 665 + ], + "type": "text", + "content": "This project was supported by the National Key R&D Program of China (No. 2022ZD0160101) and the Shanghai Artificial Intelligence Laboratory. Mike Shou does not receive any funding for this work." + } + ] + } + ], + "index": 8 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "spans": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "type": "text", + "content": "14" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 166, + 91, + 217, + 100 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 166, + 91, + 217, + 100 + ], + "spans": [ + { + "bbox": [ + 166, + 91, + 217, + 100 + ], + "type": "text", + "content": "Y. Lin et al." + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 13 + }, + { + "para_blocks": [ + { + "bbox": [ + 133, + 114, + 197, + 126 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 114, + 197, + 126 + ], + "spans": [ + { + "bbox": [ + 133, + 114, + 197, + 126 + ], + "type": "text", + "content": "References" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 138, + 140, + 480, + 665 + ], + "type": "list", + "angle": 0, + "index": 20, + "blocks": [ + { + "bbox": [ + 138, + 140, + 480, + 172 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 140, + 480, + 172 + ], + "spans": [ + { + "bbox": [ + 138, + 140, + 480, + 172 + ], + "type": "text", + "content": "1. Agarwal, S., Krueger, G., Clark, J., Radford, A., Kim, J.W., Brundage, M.: Evaluating clip: towards characterization of broader capabilities and downstream implications. arXiv preprint arXiv:2108.02818 (2021)" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 138, + 173, + 480, + 194 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 173, + 480, + 194 + ], + "spans": [ + { + "bbox": [ + 138, + 173, + 480, + 194 + ], + "type": "text", + "content": "2. Alabdulmohsin, I., Wang, X., Steiner, A.P., Goyal, P., D'Amour, A., Zhai, X.: Clip the bias: How useful is balancing data in multimodal learning? In: ICLR (2024)" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 138, + 195, + 480, + 226 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 195, + 480, + 226 + ], + "spans": [ + { + "bbox": [ + 138, + 195, + 480, + 226 + ], + "type": "text", + "content": "3. Ali, J., Kleindessner, M., Wenzel, F., Budhathoki, K., Cevher, V., Russell, C.: Evaluating the fairness of discriminative foundation models in computer vision. In: AIES. pp. 809-833 (2023)" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 138, + 228, + 480, + 259 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 228, + 480, + 259 + ], + "spans": [ + { + "bbox": [ + 138, + 228, + 480, + 259 + ], + "type": "text", + "content": "4. Berg, H., Hall, S.M., Bhalgat, Y., Yang, W., Kirk, H.R., Shtedritski, A., Bain, M.: A prompt array keeps the bias away: Debiasing vision-language models with adversarial learning. arXiv preprint arXiv:2203.11933 (2022)" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 138, + 261, + 480, + 293 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 261, + 480, + 293 + ], + "spans": [ + { + "bbox": [ + 138, + 261, + 480, + 293 + ], + "type": "text", + "content": "5. Biten, A.F., Tito, R., Mafla, A., Gomez, L., Rusinol, M., Valveny, E., Jawahar, C., Karatzas, D.: Scene text visual question answering. In: ICCV. pp. 4291-4301 (2019)" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 138, + 293, + 480, + 315 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 293, + 480, + 315 + ], + "spans": [ + { + "bbox": [ + 138, + 293, + 480, + 315 + ], + "type": "text", + "content": "6. Byeon, M., Park, B., Kim, H., Lee, S., Baek, W., Kim, S.: Coyo-700m: Image-text pair dataset. https://github.com/kakaobrain/coyo-dataset (2022)" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 138, + 316, + 480, + 347 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 316, + 480, + 347 + ], + "spans": [ + { + "bbox": [ + 138, + 316, + 480, + 347 + ], + "type": "text", + "content": "7. Cao, L., Zhang, B., Chen, C., Yang, Y., Du, X., Zhang, W., Lu, Z., Zheng, Y.: Less is more: Removing text-regions improves clip training efficiency and robustness. arXiv preprint arXiv:2305.05095 (2023)" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 138, + 348, + 480, + 380 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 348, + 480, + 380 + ], + "spans": [ + { + "bbox": [ + 138, + 348, + 480, + 380 + ], + "type": "text", + "content": "8. Chen, X., Fang, H., Lin, T.Y., Vedantam, R., Gupta, S., Dolkar, P., Zitnick, C.L.: Microsoft coco captions: Data collection and evaluation server. arXiv preprint arXiv:1504.00325 (2015)" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 138, + 381, + 480, + 413 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 381, + 480, + 413 + ], + "spans": [ + { + "bbox": [ + 138, + 381, + 480, + 413 + ], + "type": "text", + "content": "9. Devlin, J., Chang, M.W., Lee, K., Toutanova, K.: Bert: Pre-training of deep bidirectional transformers for language understanding. arXiv preprint arXiv:1810.04805 (2018)" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 138, + 414, + 480, + 456 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 414, + 480, + 456 + ], + "spans": [ + { + "bbox": [ + 138, + 414, + 480, + 456 + ], + "type": "text", + "content": "10. Dosovitskiy, A., Beyer, L., Kolesnikov, A., Weissenborn, D., Zhai, X., Unterthiner, T., Dehghani, M., Minderer, M., Heigold, G., Gelly, S., et al.: An image is worth 16x16 words: Transformers for image recognition at scale. arXiv preprint arXiv:2010.11929 (2020)" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 138, + 457, + 480, + 490 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 457, + 480, + 490 + ], + "spans": [ + { + "bbox": [ + 138, + 457, + 480, + 490 + ], + "type": "text", + "content": "1. Gadre, S.Y., Ilharco, G., Fang, A., Hayase, J., Smyrnis, G., Nguyen, T., Marten, R., Wortsman, M., Ghosh, D., Zhang, J., et al.: Datacomp: In search of the next generation of multimodal datasets. arXiv preprint arXiv:2304.14108 (2023)" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 138, + 491, + 480, + 511 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 491, + 480, + 511 + ], + "spans": [ + { + "bbox": [ + 138, + 491, + 480, + 511 + ], + "type": "text", + "content": "2. Ganz, R., Nuriel, O., Aberdam, A., Kittenplon, Y., Mazor, S., Litman, R.: Towards models that can see and read. arXiv preprint arXiv:2301.07389 (2023)" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 138, + 512, + 480, + 544 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 512, + 480, + 544 + ], + "spans": [ + { + "bbox": [ + 138, + 512, + 480, + 544 + ], + "type": "text", + "content": "3. Goh, G., Cammarata, N., Voss, C., Carter, S., Petrov, M., Schubert, L., Radford, A., Olah, C.: Multimodal neurons in artificial neural networks. Distill 6(3), e30 (2021)" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 138, + 545, + 480, + 578 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 545, + 480, + 578 + ], + "spans": [ + { + "bbox": [ + 138, + 545, + 480, + 578 + ], + "type": "text", + "content": "4. Goyal, Y., Khot, T., Summers-Stay, D., Batra, D., Parikh, D.: Making the v in vqa matter: Elevating the role of image understanding in visual question answering. In: CVPR. pp. 6904-6913 (2017)" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 138, + 578, + 480, + 599 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 578, + 480, + 599 + ], + "spans": [ + { + "bbox": [ + 138, + 578, + 480, + 599 + ], + "type": "text", + "content": "5. He, K., Zhang, X., Ren, S., Sun, J.: Deep residual learning for image recognition. In: CVPR. pp. 770-778 (2016)" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 138, + 600, + 480, + 632 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 600, + 480, + 632 + ], + "spans": [ + { + "bbox": [ + 138, + 600, + 480, + 632 + ], + "type": "text", + "content": "6. Ilharco, G., Wortsman, M., Wightman, R., Gordon, C., Carlini, N., Taori, R., Dave, A., Shankar, V., Namkoong, H., Miller, J., Hajishirzi, H., Farhadi, A., Schmidt, L.: Openclip (Jul 2021), https://doi.org/10.5281/zenodo.5143773" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 138, + 633, + 480, + 665 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 633, + 480, + 665 + ], + "spans": [ + { + "bbox": [ + 138, + 633, + 480, + 665 + ], + "type": "text", + "content": "7. Jia, C., Yang, Y., Xia, Y., Chen, Y.T., Parekh, Z., Pham, H., Le, Q., Sung, Y.H., Li, Z., Duerig, T.: Scaling up visual and vision-language representation learning with noisy text supervision. In: ICML. pp. 4904-4916. PMLR (2021)" + } + ] + } + ], + "index": 19 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 274, + 91, + 447, + 102 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 274, + 91, + 447, + 102 + ], + "spans": [ + { + "bbox": [ + 274, + 91, + 447, + 102 + ], + "type": "text", + "content": "Parrot Captions Teach CLIP to Spot Text" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 470, + 92, + 480, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 470, + 92, + 480, + 100 + ], + "spans": [ + { + "bbox": [ + 470, + 92, + 480, + 100 + ], + "type": "text", + "content": "15" + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 14 + }, + { + "para_blocks": [ + { + "bbox": [ + 132, + 116, + 480, + 665 + ], + "type": "list", + "angle": 0, + "index": 22, + "blocks": [ + { + "bbox": [ + 133, + 116, + 480, + 138 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 116, + 480, + 138 + ], + "spans": [ + { + "bbox": [ + 133, + 116, + 480, + 138 + ], + "type": "text", + "content": "18. Johnson, J., Douze, M., Jégou, H.: Billion-scale similarity search with GPUs. IEEE Transactions on Big Data 7(3), 535-547 (2019)" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 133, + 139, + 480, + 171 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 139, + 480, + 171 + ], + "spans": [ + { + "bbox": [ + 133, + 139, + 480, + 171 + ], + "type": "text", + "content": "19. Lemesle, Y., Sawayama, M., Valle-Perez, G., Adolphe, M., Sauzeon, H., Oudeyer, P.Y.: Language-biased image classification: evaluation based on semantic representations. arXiv preprint arXiv:2201.11014 (2022)" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 133, + 171, + 480, + 191 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 171, + 480, + 191 + ], + "spans": [ + { + "bbox": [ + 133, + 171, + 480, + 191 + ], + "type": "text", + "content": "20. Li, B., Weinberger, K.Q., Belongie, S., Koltun, V., Ranftl, R.: Language-driven semantic segmentation. In: ICLR (2022)" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 132, + 192, + 480, + 224 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 192, + 480, + 224 + ], + "spans": [ + { + "bbox": [ + 132, + 192, + 480, + 224 + ], + "type": "text", + "content": "21. Li, J., Li, D., Xiong, C., Hoi, S.: Blip: Bootstrapping language-image pre-training for unified vision-language understanding and generation. In: ICML. pp. 12888-12900. PMLR (2022)" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 132, + 225, + 480, + 256 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 225, + 480, + 256 + ], + "spans": [ + { + "bbox": [ + 132, + 225, + 480, + 256 + ], + "type": "text", + "content": "22. Li, J., Selvaraju, R., Gotmare, A., Joty, S., Xiong, C., Hoi, S.C.H.: Align before fuse: Vision and language representation learning with momentum distillation. NeurIPS 34, 9694-9705 (2021)" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 132, + 257, + 480, + 278 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 257, + 480, + 278 + ], + "spans": [ + { + "bbox": [ + 132, + 257, + 480, + 278 + ], + "type": "text", + "content": "23. Loshchilov, I., Hutter, F.: Decoupled weight decay regularization. arXiv preprint arXiv:1711.05101 (2017)" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 132, + 278, + 480, + 310 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 278, + 480, + 310 + ], + "spans": [ + { + "bbox": [ + 132, + 278, + 480, + 310 + ], + "type": "text", + "content": "24. Maini, P., Goyal, S., Lipton, Z.C., Kolter, J.Z., Raghunathan, A.: T-mars: Improving visual representations by circumventing text feature learning. arXiv preprint arXiv:2307.03132 (2023)" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 132, + 311, + 480, + 331 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 311, + 480, + 331 + ], + "spans": [ + { + "bbox": [ + 132, + 311, + 480, + 331 + ], + "type": "text", + "content": "25. Materzyńska, J., Torralba, A., Bau, D.: Disentangling visual and written concepts in clip. In: CVPR. pp. 16410-16419 (2022)" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 132, + 332, + 480, + 364 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 332, + 480, + 364 + ], + "spans": [ + { + "bbox": [ + 132, + 332, + 480, + 364 + ], + "type": "text", + "content": "26. Nichol, A., Dhariwal, P., Ramesh, A., Shyam, P., Mishkin, P., McGrew, B., Sutskever, I., Chen, M.: Glide: Towards photorealistic image generation and editing with text-guided diffusion models. arXiv preprint arXiv:2112.10741 (2021)" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 132, + 364, + 480, + 396 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 364, + 480, + 396 + ], + "spans": [ + { + "bbox": [ + 132, + 364, + 480, + 396 + ], + "type": "text", + "content": "27. Pham, H., Dai, Z., Ghiasi, G., Kawaguchi, K., Liu, H., Yu, A.W., Yu, J., Chen, Y.T., Luong, M.T., Wu, Y., et al.: Combined scaling for zero-shot transfer learning. Neurocomputing 555, 126658 (2023)" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 132, + 396, + 480, + 429 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 396, + 480, + 429 + ], + "spans": [ + { + "bbox": [ + 132, + 396, + 480, + 429 + ], + "type": "text", + "content": "28. Radenovic, F., Dubey, A., Kadian, A., Mihaylov, T., Vandenhende, S., Patel, Y., Wen, Y., Ramanathan, V., Mahajan, D.: Filtering, distillation, and hard negatives for vision-language pre-training. In: CVPR. pp. 6967-6977 (2023)" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 132, + 430, + 480, + 461 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 430, + 480, + 461 + ], + "spans": [ + { + "bbox": [ + 132, + 430, + 480, + 461 + ], + "type": "text", + "content": "29. Radford, A., Kim, J.W., Hallacy, C., Ramesh, A., Goh, G., Agarwal, S., Sastry, G., Askell, A., Mishkin, P., Clark, J., et al.: Learning transferable visual models from natural language supervision. In: ICML. pp. 8748-8763. PMLR (2021)" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 132, + 461, + 480, + 482 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 461, + 480, + 482 + ], + "spans": [ + { + "bbox": [ + 132, + 461, + 480, + 482 + ], + "type": "text", + "content": "30. Rombach, R., Blattmann, A., Lorenz, D., Esser, P., Ommer, B.: High-resolution image synthesis with latent diffusion models. In: CVPR. pp. 10684-10695 (2022)" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 132, + 483, + 480, + 525 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 483, + 480, + 525 + ], + "spans": [ + { + "bbox": [ + 132, + 483, + 480, + 525 + ], + "type": "text", + "content": "31. Schuhmann, C., Beaumont, R., Vencu, R., Gordon, C., Wightman, R., Cherti, M., Coombes, T., Katta, A., Mullis, C., Wortsman, M., et al.: Laion-5b: An open large-scale dataset for training next generation image-text models. NeurIPS 35, 25278-25294 (2022)" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 132, + 525, + 480, + 558 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 525, + 480, + 558 + ], + "spans": [ + { + "bbox": [ + 132, + 525, + 480, + 558 + ], + "type": "text", + "content": "32. Schuhmann, C., Vencu, R., Beaumont, R., Kaczmarczyk, R., Mullis, C., Katta, A., Coombes, T., Jitsev, J., Komatsuzaki, A.: Laion-400m: Open dataset of clip-filtered 400 million image-text pairs. arXiv preprint arXiv:2111.02114 (2021)" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 132, + 559, + 480, + 579 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 559, + 480, + 579 + ], + "spans": [ + { + "bbox": [ + 132, + 559, + 480, + 579 + ], + "type": "text", + "content": "33. Shi, C., Yang, S.: Logoprompt: Synthetic text images can be good visual prompts for vision-language models. In: ICCV. pp. 2932-2941 (2023)" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 132, + 580, + 480, + 601 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 580, + 480, + 601 + ], + "spans": [ + { + "bbox": [ + 132, + 580, + 480, + 601 + ], + "type": "text", + "content": "34. Shtedritski, A., Rupprecht, C., Vedaldi, A.: What does clip know about a red circle? visual prompt engineering for vlms. arXiv preprint arXiv:2304.06712 (2023)" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 132, + 601, + 480, + 622 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 601, + 480, + 622 + ], + "spans": [ + { + "bbox": [ + 132, + 601, + 480, + 622 + ], + "type": "text", + "content": "35. Sidorov, O., Hu, R., Rohrbach, M., Singh, A.: Textcaps: a dataset for image captioning with reading comprehension. In: ECCV. pp. 742-758. Springer (2020)" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 132, + 622, + 480, + 643 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 622, + 480, + 643 + ], + "spans": [ + { + "bbox": [ + 132, + 622, + 480, + 643 + ], + "type": "text", + "content": "36. Singh, A., Natarajan, V., Shah, M., Jiang, Y., Chen, X., Batra, D., Parikh, D., Rohrbach, M.: Towards vqa models that can read. In: CVPR. pp. 8317-8326 (2019)" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 132, + 644, + 480, + 665 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 644, + 480, + 665 + ], + "spans": [ + { + "bbox": [ + 132, + 644, + 480, + 665 + ], + "type": "text", + "content": "37. Tanjin, M.M., Singh, K.K., Kafle, K., Sinha, R., Cottrell, G.W.: Discovering and mitigating biases in clip-based image editing. In: WACV. pp. 2984-2993 (2024)" + } + ] + } + ], + "index": 21 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "spans": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "type": "text", + "content": "16" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 166, + 91, + 217, + 100 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 166, + 91, + 217, + 100 + ], + "spans": [ + { + "bbox": [ + 166, + 91, + 217, + 100 + ], + "type": "text", + "content": "Y. Lin et al." + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 15 + }, + { + "para_blocks": [ + { + "bbox": [ + 130, + 116, + 481, + 446 + ], + "type": "list", + "angle": 0, + "index": 14, + "blocks": [ + { + "bbox": [ + 130, + 116, + 480, + 138 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 116, + 480, + 138 + ], + "spans": [ + { + "bbox": [ + 130, + 116, + 480, + 138 + ], + "type": "text", + "content": "38. Telea, A.: An image inpainting technique based on the fast marching method. Journal of graphics tools 9(1), 23-34 (2004)" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 130, + 138, + 480, + 160 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 138, + 480, + 160 + ], + "spans": [ + { + "bbox": [ + 130, + 138, + 480, + 160 + ], + "type": "text", + "content": "39. Tschannen, M., Mustafa, B., Houlsby, N.: Clippo: Image-and-language understanding from pixels only. In: CVPR. pp. 11006-11017 (2023)" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 130, + 160, + 481, + 194 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 160, + 481, + 194 + ], + "spans": [ + { + "bbox": [ + 130, + 160, + 481, + 194 + ], + "type": "text", + "content": "40. Wang, H., Zhan, Y., Liu, L., Ding, L., Yu, J.: Balanced similarity with auxiliary prompts: Towards alleviating text-to-image retrieval bias for clip in zero-shot learning. arXiv preprint arXiv:2402.18400 (2024)" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 130, + 194, + 480, + 216 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 194, + 480, + 216 + ], + "spans": [ + { + "bbox": [ + 130, + 194, + 480, + 216 + ], + "type": "text", + "content": "41. Wang, J., Liu, Y., Wang, X.E.: Are gender-neutral queries really gender-neutral? mitigating gender bias in image search. arXiv preprint arXiv:2109.05433 (2021)" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 130, + 216, + 480, + 247 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 216, + 480, + 247 + ], + "spans": [ + { + "bbox": [ + 130, + 216, + 480, + 247 + ], + "type": "text", + "content": "42. Wei, J., Tay, Y., Bommasani, R., Raffel, C., Zoph, B., Borgeaud, S., Yogatama, D., Bosma, M., Zhou, D., Metzler, D., et al.: Emergent abilities of large language models. arXiv preprint arXiv:2206.07682 (2022)" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 130, + 248, + 480, + 270 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 248, + 480, + 270 + ], + "spans": [ + { + "bbox": [ + 130, + 248, + 480, + 270 + ], + "type": "text", + "content": "43. Xu, Y., Xu, Z., Chai, W., Zhao, Z., Song, E., Wang, G.: Devil in the number: Towards robust multi-modality data filter. arXiv preprint arXiv:2309.13770 (2023)" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 130, + 270, + 480, + 303 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 270, + 480, + 303 + ], + "spans": [ + { + "bbox": [ + 130, + 270, + 480, + 303 + ], + "type": "text", + "content": "44. Yao, Y., Zhang, A., Zhang, Z., Liu, Z., Chua, T.S., Sun, M.: Cpt: Colorful prompt tuning for pre-trained vision-language models. arXiv preprint arXiv:2109.11797 (2021)" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 130, + 303, + 480, + 335 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 303, + 480, + 335 + ], + "spans": [ + { + "bbox": [ + 130, + 303, + 480, + 335 + ], + "type": "text", + "content": "45. Ye, M., Zhang, J., Zhao, S., Liu, J., Liu, T., Du, B., Tao, D.: Deepsolo: Let transformer decoder with explicit points solo for text spotting. In: CVPR. pp. 19348-19357 (2023)" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 130, + 335, + 480, + 369 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 335, + 480, + 369 + ], + "spans": [ + { + "bbox": [ + 130, + 335, + 480, + 369 + ], + "type": "text", + "content": "46. Yuksekgonul, M., Bianchi, F., Kalluri, P., Jurafsky, D., Zou, J.: When and why vision-language models behave like bags-of-words, and what to do about it? In: ICLR (2022)" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 130, + 369, + 480, + 402 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 369, + 480, + 402 + ], + "spans": [ + { + "bbox": [ + 130, + 369, + 480, + 402 + ], + "type": "text", + "content": "47. Zhai, X., Wang, X., Mustafa, B., Steiner, A., Keysers, D., Kolesnikov, A., Beyer, L.: Lit: Zero-shot transfer with locked-image text tuning. In: CVPR. pp. 18123-18133 (2022)" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 130, + 402, + 480, + 424 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 402, + 480, + 424 + ], + "spans": [ + { + "bbox": [ + 130, + 402, + 480, + 424 + ], + "type": "text", + "content": "48. Zhang, Q., Xu, Y., Zhang, J., Tao, D.: Vitaev2: Vision transformer advanced by exploring inductive bias for image recognition and beyond. IJCV pp. 1-22 (2023)" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 130, + 424, + 480, + 446 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 424, + 480, + 446 + ], + "spans": [ + { + "bbox": [ + 130, + 424, + 480, + 446 + ], + "type": "text", + "content": "49. Zhou, K., Yang, J., Loy, C.C., Liu, Z.: Learning to prompt for vision-language models. IJCV 130(9), 2337-2348 (2022)" + } + ] + } + ], + "index": 13 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 274, + 91, + 447, + 102 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 274, + 91, + 447, + 102 + ], + "spans": [ + { + "bbox": [ + 274, + 91, + 447, + 102 + ], + "type": "text", + "content": "Parrot Captions Teach CLIP to Spot Text" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 470, + 91, + 481, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 470, + 91, + 481, + 100 + ], + "spans": [ + { + "bbox": [ + 470, + 91, + 481, + 100 + ], + "type": "text", + "content": "17" + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 16 + } + ], + "_backend": "vlm", + "_version_name": "2.6.4" +} \ No newline at end of file diff --git a/2024/Parrot_ Pareto-optimal Multi-Reward Reinforcement Learning Framework for Text-to-Image Generation/938566d1-b778-4bb0-b908-cbc4a31c6445_content_list.json b/2024/Parrot_ Pareto-optimal Multi-Reward Reinforcement Learning Framework for Text-to-Image Generation/938566d1-b778-4bb0-b908-cbc4a31c6445_content_list.json new file mode 100644 index 0000000000000000000000000000000000000000..a0b403e8deebaace89773a8adb820a73df32b3ae --- /dev/null +++ b/2024/Parrot_ Pareto-optimal Multi-Reward Reinforcement Learning Framework for Text-to-Image Generation/938566d1-b778-4bb0-b908-cbc4a31c6445_content_list.json @@ -0,0 +1,1994 @@ +[ + { + "type": "text", + "text": "Parrot: Pareto-optimal Multi-Reward Reinforcement Learning Framework for Text-to-Image Generation", + "text_level": 1, + "bbox": [ + 277, + 140, + 727, + 209 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Seung Hyun Lee $^{1,6*}$ , Yinxiao Li $^{1}$ , Junjie Ke $^{1}$ , Innfarn Yoo $^{2}$ , Han Zhang $^{3}$ , Jiahui Yu $^{4\\dagger}$ , Qifei Wang $^{2}$ , Fei Deng $^{2,5*}$ , Glenn Entis $^{1}$ , Junfeng He $^{1}$ , Gang Li $^{1}$ , Sangpil Kim $^{6}$ , Irfan Essa $^{1}$ , Feng Yang $^{1}$", + "bbox": [ + 228, + 233, + 772, + 282 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Google Research1, Google2, Google DeepMind3, OpenAI4, Rutgers University5, Korea University6", + "bbox": [ + 305, + 291, + 696, + 321 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/4c4a4ecb41a55053004f374235ae1011858d93e51976d24ef4603162331103c4.jpg", + "image_caption": [ + "Fig. 1: Parrot visual examples. Parrot consistently improves the quality of generated images across multiple criteria: aesthetics, human preference, text-image alignment, and image sentiment. Each column shows generated images using the same seed." + ], + "image_footnote": [], + "bbox": [ + 217, + 324, + 787, + 571 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Abstract. Recent works have demonstrated that using reinforcement learning (RL) with multiple quality rewards can improve the quality of generated images in text-to-image (T2I) generation. However, manually adjusting reward weights poses challenges and may cause over-optimization in certain metrics. To solve this, we propose Parrot, which addresses the issue through multi-objective optimization and introduces an effective multi-reward optimization strategy to approximate Pareto optimal. Utilizing batch-wise Pareto optimal selection, Parrot automatically identifies the optimal trade-off among different rewards. We use the novel multi-reward optimization algorithm to jointly optimize the T2I model and a prompt expansion network, resulting in significant improvement of image quality and also allow to control the trade-off of", + "bbox": [ + 259, + 640, + 743, + 806 + ], + "page_idx": 0 + }, + { + "type": "page_footnote", + "text": "* This work was done during an internship at Google.", + "bbox": [ + 230, + 816, + 544, + 828 + ], + "page_idx": 0 + }, + { + "type": "page_footnote", + "text": "† This work was done during working at Google.", + "bbox": [ + 230, + 828, + 513, + 839 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "different rewards using a reward related prompt during inference. Furthermore, we introduce original prompt-centered guidance at inference time, ensuring fidelity to user input after prompt expansion. Extensive experiments and a user study validate the superiority of Parrot over several baselines across various quality criteria, including aesthetics, human preference, text-image alignment, and image sentiment.", + "bbox": [ + 259, + 147, + 743, + 232 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "1 Introduction", + "text_level": 1, + "bbox": [ + 215, + 258, + 375, + 276 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Despite significant advancements in text-to-image (T2I) generation [17, 29, 38, 45, 55], recent work like Imagen [42] and Stable Diffusion [39], still struggle to produce high quality images. Images in the first row in Fig. 1 illustrates such quality issues in Stable Diffusion [39], including poor composition (e.g. bad cropping), misalignment with input prompts (e.g. missing objects), or overall lack of aesthetic appeal. Assessing the quality of generated images can involve various metrics such as aesthetics [23, 24, 33], human preference [26], text-image alignment [35], and emotional appeal [44]. Enhancing T2I generation across multiple quality metrics remains a challenging task.", + "bbox": [ + 212, + 294, + 787, + 430 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Recent works [4,12,15,27] have demonstrated that incorporating quality signals as reward functions in fine-tuning T2I with reinforcement learning (RL) can improve image quality. For instance, Promptist [15] fine-tunes prompt expansion model using RL with the sum of aesthetics and text-image alignment scores. However, the simple weighted sum approach may not effectively handle trade-offs among multiple quality metrics. As the number of rewards increases, manually adjusting reward weights becomes impractical. Moreover, optimizing one quality metric may inadvertently compromise others, as the model might prioritize aesthetics over relevance to the input prompt. Additionally, the trade-off for different reward is not controllable after training.", + "bbox": [ + 212, + 431, + 789, + 582 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "To address these challenges, we propose Parrot, a Pareto-optimal multi-reward reinforcement learning algorithm to improve text-to-image generation. Unlike previous approaches that treat T2I reward optimization as a single objective optimization problem, Parrot tackles this challenge through multi-objective optimization and introduces an effective multi-reward optimization strategy to achieve Pareto optimal approximation. Intuitively, each generated sample in a batch embodies a distinctive trade-off among various quality rewards, with some samples exhibiting superior trade-offs compared to others. Instead of updating gradients using all batch samples, Parrot uses non-dominated points [32] which have better trade-offs. Consequently, Parrot automatically learns from the optimal trade-off among different rewards. Moreover, Parrot learns reward-specific preference prompts, which can be utilized individually or in combination to control the trade-off among different rewards during inference time. Unlike prior work, which either solely fine-tunes the T2I model [12] or only tunes the prompt expansion network while freezing the T2I model [15], we employ the Parrot multi-reward optimization algorithm to jointly optimize both the T2I model and the prompt expansion network. This collaborative optimization unlocks the", + "bbox": [ + 212, + 583, + 789, + 840 + ], + "page_idx": 1 + }, + { + "type": "page_number", + "text": "2", + "bbox": [ + 217, + 114, + 228, + 126 + ], + "page_idx": 1 + }, + { + "type": "header", + "text": "Lee, Seung Hyun et al.", + "bbox": [ + 271, + 114, + 424, + 128 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "full potential of Parrot by encouraging both more details from added context from the prompt expansion model, and the overall quality improvement on the T2I generation. During inference, we further introduce original prompt-centered guidance to ensure the output image is relevant to input prompts after prompt expansion.", + "bbox": [ + 212, + 146, + 787, + 222 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "In summary, our contributions can be listed as follows:", + "bbox": [ + 238, + 223, + 635, + 238 + ], + "page_idx": 2 + }, + { + "type": "list", + "sub_type": "text", + "list_items": [ + "- We propose Parrot, a novel multi-reward optimization algorithm for T2I RL fine-tuning. Leveraging batch-wise Pareto-optimal selection, it effectively optimizes multiple T2I rewards, enabling collaborative improvement in aesthetics, human preference, image sentiment, and text-image alignment and also allowing to control the trade-off of different rewards using a reward related prompt during inference.", + "- We show the advantage of jointly optimizing both the prompt expansion network and the T2I model, which has never been explored before.", + "- We introduce original prompt-centered guidance during inference time after prompt expansion, ensuring better alignment with the original prompt while enriching image details.", + "- Extensive results and a user study validate that Parrot outperforms several baseline methods across various quality criteria." + ], + "bbox": [ + 223, + 239, + 782, + 434 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "2 Related Work", + "text_level": 1, + "bbox": [ + 215, + 467, + 387, + 483 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "T2I Generation: The goal of T2I generation is to create an image given an input text prompt. Several T2I generative models have been proposed and have demonstrated promising results [5, 7, 14, 21, 22, 28, 37, 41, 42, 54]. Stable Diffusion [39] shows impressive generation performance in T2I generation, leveraging latent text representations from LLMs. Despite substantial progress, the images generated by those models still exhibit quality issues, such as bad cropping or misalignment with the input texts.", + "bbox": [ + 212, + 513, + 787, + 619 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "RL for T2I Fine-tuning: Starting by Fan et al. [11] to explore RL fine-tuning for T2I models, following works [4, 8, 9, 12, 16] have explored RL fine-tuning technique for T2I diffusion model, showcasing superior performance for human preference learning. DPOK [12] improves quality through RL using ImageReward [51] score as a reward with a few prompts. In addition to fine-tuning the T2I model directly using RL, Promptist [15] fine-tunes the prompt expansion model by using a simple sum of aesthetic and text-image alignment scores as reward. DRaFT [6] proposed not only differentiable rewards for efficient fine-tuning but also effectiveness of using linear summation of multi-rewards. These methods treat T2I RL as a single-objective optimization problem, while Parrot employs multi-objective optimization. Additionally, prior approaches either fine-tune the T2I model or the prompt expansion model while freezing the other. In contrast, Parrot proposes joint optimization of the prompt expansion model and the T2I model using multi-reward RL to foster better collaboration.", + "bbox": [ + 212, + 628, + 787, + 839 + ], + "page_idx": 2 + }, + { + "type": "header", + "text": "Pareto-optimal Multi-Reward RL Framework for T2I Generation", + "bbox": [ + 297, + 114, + 730, + 128 + ], + "page_idx": 2 + }, + { + "type": "page_number", + "text": "3", + "bbox": [ + 774, + 116, + 784, + 126 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Multi-objective Optimization: Multi-objective optimization problem involves optimizing multiple objective functions simultaneously. The scalarization technique [31, 47] formulates multi-objective problem into single-objective problems with the weighted sum of each score, which requires pre-defined weights for each objective. Rame et al. [36] proposed weighted averaging method to find Pareto frontier, leveraging multiple fine-tuned models. Lin et al. [30] proposes to learn a set model to map trade-off preference vectors to their corresponding Pareto solutions. Inspired by this, Parrot introduces a language-based preference vector constructed from task identifiers for each reward, then encoded by the text-encoder. In the context of multi-reward RL for T2I diffusion models, Promptist [15] uses a simple weighted sum of two reward scores. This approach requires manual tuning of the weights, which makes it time-consuming and hard to scale when the number of rewards increases.", + "bbox": [ + 212, + 146, + 787, + 342 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Generated Image Quality: The quality assessment of images generated by T2I models involves multiple dimensions, and various metrics have been proposed. In this paper, we consider using four types of quality metrics as rewards: aesthetics, human preference, text-image alignment, and image sentiment. Aesthetics captures the overall visual appealingness of the image, and it is learned using human ratings for aesthetics in real images [13, 19, 23, 24, 33, 48, 52]. Human preferences, rooted the concept of learning from human feedback [3, 34, 50], involves gathering preferences at scale by having raters to compare generated images [26, 51]. Text-image alignment measures the extent to which the generated image aligns with the input prompt, CLIP [35] score is often employed, measuring the cosine distance of between contrastive image embedding and text embedding. Image sentiment is important for ensuring the generated image evokes positive emotions in the viewer. Serra et al. [44] predict average polarity of sentiments an image elicits and learn estimates for positive, neutral, and negative scores. In Parrot, we use its positive score as a reward for positive emotions.", + "bbox": [ + 212, + 349, + 787, + 578 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "3 Preliminary", + "text_level": 1, + "bbox": [ + 215, + 603, + 369, + 622 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Diffusion Probabilistic Models: Diffusion probabilistic models [17] generate the image by gradually denoising a noisy image. Specifically, given a real image $\\mathbf{x}_0$ from the data distribution $\\mathbf{x}_0 \\sim q(\\mathbf{x}_0)$ , the forward process $q(\\mathbf{x}_t | \\mathbf{x}_0, c)$ of diffusion probabilistic models produce a noisy image $\\mathbf{x}_t$ , which induces a distribution $p(\\mathbf{x}_0, c)$ conditioned on text prompt $c$ . In classifier-free guidance [18], denoising model predicts noise $\\bar{\\epsilon}_{\\theta}$ with a linear combination of the unconditional score estimates $\\epsilon_{\\theta}(\\mathbf{x}_t, t)$ and the conditional score estimates $\\epsilon_{\\theta}(\\mathbf{x}_t, t, c)$ as follows:", + "bbox": [ + 212, + 645, + 787, + 752 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\n\\bar {\\epsilon} _ {\\theta} = w \\cdot \\epsilon_ {\\theta} (\\mathbf {x} _ {t}, t, c) + (1 - w) \\cdot \\epsilon_ {\\theta} (\\mathbf {x} _ {t}, t, \\mathrm {n u l l}), \\tag {1}\n$$\n", + "text_format": "latex", + "bbox": [ + 346, + 766, + 785, + 782 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "where $t$ denotes diffusion time step, the null indicates a null text and $w$ represents the guidance scale of classifier-free guidance where $w \\geq 1$ . Note that $\\epsilon_{\\theta}$ is typically parameterized by the UNet [40].", + "bbox": [ + 212, + 794, + 787, + 842 + ], + "page_idx": 3 + }, + { + "type": "page_number", + "text": "4", + "bbox": [ + 217, + 114, + 228, + 126 + ], + "page_idx": 3 + }, + { + "type": "header", + "text": "Lee, Seung Hyun et al.", + "bbox": [ + 271, + 114, + 424, + 128 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "RL-based T2I Diffusion Model Fine-tuning: Given a reward signal from generated images, the goal of RL-tuning for T2I diffusion models is to optimize the policy defined as one denoising step of T2I diffusion models. In particular, Black et al. [4] apply policy gradient algorithm, which regards the denoising process of diffusion models as a Markov decision process (MDP) by performing multiple denoising steps iteratively. Subsequently, a black box reward model $r(\\cdot ,\\cdot)$ predicts a single scalar value from sampled image $\\mathbf{x}_0$ . Given text condition $c\\sim p(c)$ and image $\\mathbf{x}_0$ , objective function $\\mathcal{I}$ can be defined to maximize the expected reward as follows:", + "bbox": [ + 212, + 146, + 782, + 282 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\n\\mathcal {J} _ {\\theta} = \\mathbb {E} _ {p (c)} \\mathbb {E} _ {p _ {\\theta} (\\mathbf {x} _ {0} | c)} [ r (\\mathbf {x} _ {0}, c) ], \\tag {2}\n$$\n", + "text_format": "latex", + "bbox": [ + 401, + 291, + 784, + 308 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "where the pre-trained diffusion model $p_{\\theta}$ produces a sample distribution $p_{\\theta}(\\mathbf{x}_0|c)$ using text condition $c$ . Modifying this equation, Fan et al. [12] demonstrate that the gradient of objective function $\\nabla \\mathcal{I}_{\\theta}$ can be calculated through gradient ascent algorithm without using the gradient of reward model as follows:", + "bbox": [ + 212, + 313, + 782, + 375 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\n\\nabla \\mathcal {J} _ {\\theta} = \\mathbb {E} \\left[ r \\left(\\mathbf {x} _ {0}, c\\right) \\sum_ {t = 1} ^ {T} \\nabla_ {\\theta} \\log p _ {\\theta} \\left(\\mathbf {x} _ {t - 1} \\mid c, t, \\mathbf {x} _ {t}\\right) \\right], \\tag {3}\n$$\n", + "text_format": "latex", + "bbox": [ + 344, + 381, + 784, + 421 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "where $T$ denotes the total time step of the diffusion sampling. With parameters $\\theta$ , the expectation value can be taken over the trajectories of diffusion sampling.", + "bbox": [ + 214, + 426, + 782, + 458 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "4 Method", + "text_level": 1, + "bbox": [ + 214, + 477, + 330, + 493 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "4.1 Parrot Overview", + "text_level": 1, + "bbox": [ + 214, + 506, + 400, + 520 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Fig. 2 shows the overview of Parrot, which consists of the prompt expansion network (PEN) $p_{\\phi}$ and the T2I diffusion model $p_{\\theta}$ . The PEN is first initialized from a supervised fine-tuning checkpoint on demonstrations of prompt expansion pairs, and the T2I model is initialized from pretrained diffusion model. Given the original prompt $c$ , the PEN generates an expanded prompt $\\hat{c}$ , and the T2I model generates images based on this expanded prompt. During the multi-reward RL fine-tuning, a batch of $N$ images is sampled, and multiple quality rewards are calculated for each image, encompassing aspects like text-image alignment, aesthetics, human preference, and image sentiment. Based on these reward scores, Parrot identifies the batch-wise Pareto-optimal set using a non-dominated sorting algorithm. This optimal set of images is then used for joint optimization of the PEN and T2I model parameters through RL policy gradient update. During inference, Parrot leverages both the original prompt and its expansion, striking a balance between maintaining faithfulness to the original prompt and incorporating additional details for higher quality.", + "bbox": [ + 212, + 527, + 787, + 755 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "4.2 Batch-wise Pareto-optimal Selection", + "text_level": 1, + "bbox": [ + 214, + 773, + 562, + 787 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Li et al. [30] has demonstrated that using batchwise Pareto-set learning and selecting good samples in a batch can approximate Pareto-optimality across multiple objectives. Backed up by this theory, we propose to select non-dominated", + "bbox": [ + 212, + 794, + 785, + 840 + ], + "page_idx": 4 + }, + { + "type": "header", + "text": "Pareto-optimal Multi-Reward RL Framework for T2I Generation", + "bbox": [ + 297, + 114, + 732, + 128 + ], + "page_idx": 4 + }, + { + "type": "page_number", + "text": "5", + "bbox": [ + 774, + 116, + 784, + 126 + ], + "page_idx": 4 + }, + { + "type": "image", + "img_path": "images/108ff289898dda4a663cb1763b479d7e4373fcb644164b567d72c8c0a3ddbd91.jpg", + "image_caption": [ + "Fig. 2: Overview of Parrot. During the training, $N$ images are sampled from the T2I model using the expanded prompt from the prompt expansion network. Multiple quality rewards are calculated for each image, and the Pareto-optimal set is identified using the non-dominated sorting algorithm. These optimal images are then used to perform policy gradient update of the parameters of T2I model and prompt expansion network jointly. During the inference, both the original prompt and the expanded prompt are provided to the T2I model, enabling better faithfulness while adding detail." + ], + "image_footnote": [], + "bbox": [ + 218, + 143, + 787, + 284 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "points in a batch for policy gradient update in RL to achieve Pareto-optimality. Algorithm 1 outlines the procedure of Parrot. Rather than updating the gradients using all images, Parrot focuses on high-quality samples, considering multiple quality rewards in each mini-batch. In the multi-reward RL, each sample generated by the T2I model presents distinct trade-offs for different rewards. Among these samples, a subset with varied optimal trade-offs across multiple objectives, also known as the Pareto set, exists. For a Pareto-optimal sample, none of its objective values can be further improved without damaging others. In other words, the Pareto-optimal set is not dominated by any data points, also known as the non-dominated set. To achieve a Pareto-optimal solution with text-to-image generation diffusion model, Parrot selectively uses data points from the non-dominated set using non-dominated sorting algorithm. This naturally encourages the T2I model to produce Pareto-optimal samples with respect to the multi-reward objectives.", + "bbox": [ + 212, + 415, + 787, + 628 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Reward-specific Preference: Inspired by the use of preference information in multi-objective optimization [30], Parrot incorporates the preference information through reward-specific identifiers. This enables Parrot to automatically determine the importance for each reward objective. Concretely, we enrich the expanded prompt $\\hat{c}$ by preponding reward-specific identifier “$ ” for $k$ -th reward. Based on this reward-specific prompt, $N$ images are generated and are used for maximizing the corresponding $k$ -th reward model during gradient update. At inference time, a concatenation of all the reward identifiers “$ , ..., $ ” is used for image generation.", + "bbox": [ + 212, + 635, + 789, + 772 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Non-dominated Sorting: Parrot constructs Pareto set with non-dominated points based on trade-offs among multiple rewards. These non-dominated points are superior to the remaining solutions and are not dominated by each other. Formally, the dominance relationship is defined as follows: the image $\\mathbf{x}_0^a$ domi", + "bbox": [ + 212, + 779, + 787, + 842 + ], + "page_idx": 5 + }, + { + "type": "page_number", + "text": "6", + "bbox": [ + 217, + 114, + 228, + 126 + ], + "page_idx": 5 + }, + { + "type": "header", + "text": "Lee, Seung Hyun et al.", + "bbox": [ + 271, + 114, + 424, + 128 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Algorithm Parrot: Pareto-optimal Multi-Reward RL", + "text_level": 1, + "bbox": [ + 215, + 145, + 524, + 157 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Input: Prompt $c$ , Batch size $N$ , Total iteration $E$ , the number of rewards: $K$ , Prompt expansion network $p_{\\phi}$ , T2I diffusion model: $p_{\\theta}$ , Total diffusion time step $T$ , Non-dominated set: $\\mathcal{P}$", + "bbox": [ + 215, + 159, + 782, + 180 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "for $e = 1$ to $E$ do", + "bbox": [ + 228, + 180, + 336, + 190 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Sample text prompt $c \\sim p(c)$", + "bbox": [ + 250, + 190, + 416, + 200 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "for $k = 1$ to $K$ do", + "bbox": [ + 251, + 200, + 357, + 209 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Expand text prompt $\\hat{c} \\sim p_{\\phi}(\\hat{c} | c)$", + "bbox": [ + 269, + 210, + 455, + 220 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Prepend reward-specific tokens $<$ reward $k>$ to $\\hat{c}$", + "bbox": [ + 269, + 220, + 558, + 229 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Sample a set of images $\\{\\mathbf{x}_0^1,\\dots ,\\mathbf{x}_0^N\\} \\sim p_\\theta (x_0|\\hat{c})$", + "bbox": [ + 269, + 229, + 540, + 241 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "A set of reward vector $\\mathcal{R} = \\{R_1,\\dots,R_N\\}$", + "bbox": [ + 269, + 241, + 504, + 251 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "$\\mathcal{P}\\gets \\mathrm{NDSET}(\\{\\mathbf{x}_0^1,\\dots,\\mathbf{x}_0^N\\})$", + "bbox": [ + 269, + 251, + 424, + 263 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "$\\nabla \\mathcal{J}_{\\phi} + = -r_k(\\mathbf{x}_0^j,\\hat{c})\\times \\nabla \\log p_{\\phi}(\\hat{c} |c)$", + "bbox": [ + 269, + 263, + 473, + 276 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Update the gradient $p_{\\theta}$ from Eq. 4", + "bbox": [ + 250, + 276, + 449, + 287 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Update the gradient $p_{\\phi}$", + "bbox": [ + 250, + 287, + 385, + 297 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Output: Fine-tuned diffusion model $p_{\\theta}$ , prompt expansion network $p_{\\phi}$", + "bbox": [ + 217, + 297, + 624, + 308 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "function $\\mathrm{NDSET}(\\{\\mathbf{x}_0^1,\\dots,\\mathbf{x}_0^N\\})$", + "bbox": [ + 570, + 186, + 751, + 199 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "$\\mathcal{P}\\gets \\emptyset$", + "bbox": [ + 589, + 199, + 629, + 209 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "for $i = 1$ to $N$ do", + "bbox": [ + 589, + 210, + 694, + 220 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "dominance $\\leftarrow$ True", + "bbox": [ + 609, + 220, + 718, + 229 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "for $j = 1$ to $N$ do", + "bbox": [ + 607, + 229, + 715, + 239 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "if $\\mathbf{x}_0^j$ dominates $\\mathbf{x}_0^i$ then", + "bbox": [ + 627, + 239, + 772, + 251 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "dominance $\\leftarrow$ False", + "bbox": [ + 627, + 251, + 759, + 260 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "if dominance is True then", + "bbox": [ + 607, + 261, + 761, + 271 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Add $i$ to $\\mathcal{P}$", + "bbox": [ + 627, + 271, + 694, + 280 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "return $\\mathcal{P}$", + "bbox": [ + 589, + 281, + 647, + 290 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "nates the image $\\mathbf{x}_0^b$ , denoted as $\\mathbf{x}_0^b < \\mathbf{x}_0^a$ , if and only if $R_{i}(\\mathbf{x}_{0}^{b})\\leq R_{i}(\\mathbf{x}_{0}^{a})$ for all $i\\in 1,2,\\ldots ,m$ , and there exists $j\\in 1,2,\\dots,m$ such that $R_{j}(\\mathbf{x}_{0}^{b}) < R_{j}(\\mathbf{x}_{0}^{a})$ . For example, given the $i$ -th generated image $\\mathbf{x}_0^i$ in a mini-batch, when no point in the mini-batch dominates $\\mathbf{x}_0^i$ , it is referred to as a non-dominated point.", + "bbox": [ + 215, + 338, + 782, + 400 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Policy Gradient Update: We assign a reward value of zero to the data points not included in non-dominated sets and only update the gradient of these non-dominated data points as follows:", + "bbox": [ + 214, + 407, + 782, + 452 + ], + "page_idx": 6 + }, + { + "type": "equation", + "text": "\n$$\n\\nabla \\mathcal {J} _ {\\theta} = \\sum_ {k = 1} ^ {K} \\frac {1}{n (\\mathcal {P})} \\sum_ {i = 1, \\mathbf {x} _ {0} ^ {i} \\in \\mathcal {P}} ^ {N} \\sum_ {t = 1} ^ {T} r _ {k} \\left(\\mathbf {x} _ {0} ^ {i}, c _ {k}\\right) \\times \\nabla_ {\\theta} \\log p _ {\\theta} \\left(\\mathbf {x} _ {t - 1} ^ {i} \\mid c _ {k}, t, \\mathbf {x} _ {t} ^ {i}\\right), \\tag {4}\n$$\n", + "text_format": "latex", + "bbox": [ + 264, + 460, + 784, + 506 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "where $i$ indicates the index of images in mini-batches, and $\\mathcal{P}$ denotes batch-wise a set of non-dominated points. $K$ and $T$ are the total number of reward models and total diffusion time steps, respectively. The same text prompt is used when updating the diffusion model in each batch.", + "bbox": [ + 214, + 515, + 782, + 575 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "4.3 Original Prompt Centered Guidance", + "text_level": 1, + "bbox": [ + 215, + 597, + 562, + 613 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "While prompt expansion enhances details and often improves generation quality, there is a concern that the added context may dilute the main content of the original input. To mitigate this during the inference, we introduce original prompt-centered guidance. When sampling conditioned on the original prompt, the diffusion model $\\epsilon_{\\theta}$ typically predicts noises by combining the unconditioned score estimate and the prompt-conditioned estimate. Instead of relying solely on the expanded prompt from PEN, we propose using a linear combination of two. guidances for T2I generation: one from the user input and the other from the expanded prompt. The strength of the original prompt is controlled by guidance scales $w_{1}$ and $w_{2}$ . The noise $\\bar{\\epsilon}_{\\theta}$ is estimated, derived from Eq. 1, as follows:", + "bbox": [ + 212, + 621, + 784, + 772 + ], + "page_idx": 6 + }, + { + "type": "equation", + "text": "\n$$\n\\bar {\\epsilon} _ {\\theta} = w _ {1} \\cdot \\epsilon_ {\\theta} (\\mathbf {x} _ {t}, t, c) + (1 - w _ {1} - w _ {2}) \\cdot \\epsilon_ {\\theta} (\\mathbf {x} _ {t}, t, \\text {n u l l}) + w _ {2} \\cdot \\epsilon_ {\\theta} (\\mathbf {x} _ {t}, t, \\hat {c}), \\tag {5}\n$$\n", + "text_format": "latex", + "bbox": [ + 256, + 797, + 784, + 814 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "where null denotes a null text.", + "bbox": [ + 215, + 824, + 434, + 838 + ], + "page_idx": 6 + }, + { + "type": "header", + "text": "Pareto-optimal Multi-Reward RL Framework for T2I Generation", + "bbox": [ + 297, + 114, + 730, + 127 + ], + "page_idx": 6 + }, + { + "type": "page_number", + "text": "7", + "bbox": [ + 774, + 116, + 784, + 126 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "5 Experiments", + "text_level": 1, + "bbox": [ + 215, + 143, + 375, + 162 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "5.1 Experiment Setting", + "text_level": 1, + "bbox": [ + 215, + 174, + 426, + 191 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Dataset: The PEN is first supervised fine-tuned on a large-scale text dataset named the Promptist [15], which has 360K constructed prompt pairs for original prompt and prompt expansion demonstration. The original instruction \"Rephrase\" is included per pair in Promptist. We modify the instruction into \"Input: \". This is a text input for image generation. Expand prompt for improving image quality. Output:\". Subsequently, we use the RL tuning prompts (1200K) from Promptist for RL training of the PEN and T2I model.", + "bbox": [ + 212, + 205, + 792, + 311 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "T2I Model: Our T2I model is based on the JAX version of Stable Diffusion 1.5 [39] pre-trained with the LAION-5B [43] dataset. We conduct experiments on a machine equipped with 16 NVIDIA RTX A100 GPUs. DDIM [46] with 50 denoising steps is used, and the classifier-free guidance weight is set to 5.0 with the resolution $512 \\times 512$ . Instead of updating all layers, we specifically update the cross-attention layer in the Denoising U-Net. For optimization, we employ the Adam [25] optimizer with a learning rate of $1 \\times 10^{-5}$ .", + "bbox": [ + 212, + 319, + 789, + 426 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Prompt Expansion Network: For prompt expansion, we use PaLM 2-L-IT [2], one of the PaLM2 variations, which is a multi-layer Transformer [49] decoder with casual language modeling. We optimize LoRA [20] weights for RL-based fine-tuning. The output token length of the PEN is set to 77 to match the maximum number of token length for Stable Diffusion. For original prompt-centered guidance, we set both $w_{1}$ and $w_{2}$ to 5 in Eq. 5.", + "bbox": [ + 212, + 431, + 789, + 525 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Reward Models: We incorporate four quality signals as rewards: Aesthetics, Human preference, Text-Image Alignment, Image Sentiment. For aesthetics, we use the VILA-R [24] pre-trained with the AVA [33] dataset. For human preference, we train a ViT-B/16 [10] using the Pick-a-Pic [26] dataset, which contains 500K examples for human feedback in T2I generation. The ViT-B/16 image encoder consists of 12 transformer layers, and the image resolution is $224 \\times 224$ with a patch size of $16 \\times 16$ . For text-image alignment, we use CLIP [35] with the image encoder ViT-B/32. For image sentiment, we use the pre-trained model from [44], which outputs three labels: positive, neutral, negative. We use the positive score ranging from 0 to 1 as the sentiment reward.", + "bbox": [ + 212, + 530, + 789, + 683 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "5.2 Qualitative Analysis", + "text_level": 1, + "bbox": [ + 215, + 703, + 429, + 719 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Comparison with Baselines: Fig. 3 shows the visual comparison of Parrot and multiple baselines. We include results from Stable Diffusion 1.5, DPOK [12] with a weighted sum of rewards, Promptist [15], and Parrot. DPOK exclusively fine-tunes the T2I model, while Promptist focuses on fine-tuning only the prompt expansion network. Parrot shows visually better images, particularly in aspects like color combination, cropping, perspective, and fine details in the image. This improvement can be attributed to Parrot's T2I model being fine-tuned together", + "bbox": [ + 212, + 733, + 789, + 842 + ], + "page_idx": 7 + }, + { + "type": "page_number", + "text": "8", + "bbox": [ + 217, + 114, + 228, + 126 + ], + "page_idx": 7 + }, + { + "type": "header", + "text": "Lee, Seung Hyun et al.", + "bbox": [ + 271, + 114, + 424, + 128 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/ae6232071aac53bcf03124b48b9d00f41b179c8709e99c0bc68d1720258677f9.jpg", + "image_caption": [ + "SD 1.5", + "A kitchen" + ], + "image_footnote": [], + "bbox": [ + 223, + 162, + 361, + 268 + ], + "page_idx": 8 + }, + { + "type": "image", + "img_path": "images/d4748c79abd0214fde03e30f2695b294f6fff77162411f74f2cde462ce33828e.jpg", + "image_caption": [ + "DPOK", + "A kitchen" + ], + "image_footnote": [], + "bbox": [ + 361, + 162, + 500, + 268 + ], + "page_idx": 8 + }, + { + "type": "image", + "img_path": "images/f3f59b1b91f145bfff3becf93703298da87351ead285bf567944c61b0abd7f1c.jpg", + "image_caption": [ + "Promptist" + ], + "image_footnote": [], + "bbox": [ + 500, + 162, + 635, + 268 + ], + "page_idx": 8 + }, + { + "type": "image", + "img_path": "images/be9ba214bada9817db1a1920ade4c7d433e15213e9106f291c204b80436d1da7.jpg", + "image_caption": [ + "Parrot", + "A kitchen with a lot of natural light and a beautiful view of the mountains" + ], + "image_footnote": [], + "bbox": [ + 637, + 162, + 776, + 268 + ], + "page_idx": 8 + }, + { + "type": "image", + "img_path": "images/3a0fb3345062f16eafbb92de81c52a302a4e796892c3fce8e74314e884b0d422.jpg", + "image_caption": [ + "AF1" + ], + "image_footnote": [], + "bbox": [ + 223, + 311, + 359, + 417 + ], + "page_idx": 8 + }, + { + "type": "image", + "img_path": "images/08b52d6d94c1cebccf191ab1c03df4c01b1de851b4e31987e8cd89d362b5a3b1.jpg", + "image_caption": [ + "AF1" + ], + "image_footnote": [], + "bbox": [ + 361, + 311, + 500, + 417 + ], + "page_idx": 8 + }, + { + "type": "image", + "img_path": "images/74c59804e71369577864594382c6d779f194095db69ae193d868c6b76f026fec.jpg", + "image_caption": [ + "A kitchen, a fantasy digital painting, \ntrending on Artstation, \nhighly detailed", + "F1, concept art oil painting by Jama Jurabaev, extremely detailed, brush hard, artstation" + ], + "image_footnote": [], + "bbox": [ + 500, + 311, + 637, + 417 + ], + "page_idx": 8 + }, + { + "type": "image", + "img_path": "images/ba4568e76894545df4c02602960a8c6246e3f4bf7b50cb1604941114854948cd.jpg", + "image_caption": [ + "3d rendering of a f1 car in dark and moody environment with a red and orange color scheme." + ], + "image_footnote": [], + "bbox": [ + 638, + 311, + 776, + 420 + ], + "page_idx": 8 + }, + { + "type": "image", + "img_path": "images/dff8bd248f9e50e0e809226043dba6e2436c567deb216e1c7e028e945b2657cf.jpg", + "image_caption": [ + "A woman is looking out on a winter archipelago." + ], + "image_footnote": [], + "bbox": [ + 223, + 469, + 359, + 577 + ], + "page_idx": 8 + }, + { + "type": "image", + "img_path": "images/14677736b9eff726b97b475dd96c1c6eda8c0ddfa029e8d7b85473d0bef3a859.jpg", + "image_caption": [ + "A woman is looking out on a winter archipelago." + ], + "image_footnote": [], + "bbox": [ + 361, + 469, + 500, + 577 + ], + "page_idx": 8 + }, + { + "type": "image", + "img_path": "images/4ec405cfaf3fb554d0a36f2eef601d46a48c7d2a86dcda595b57b56131c74d64.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 500, + 469, + 635, + 577 + ], + "page_idx": 8 + }, + { + "type": "image", + "img_path": "images/442e140b2561d2d17d592bbb4ec5673671caa514cce76334205728dbd355721a.jpg", + "image_caption": [ + "A woman is looking out on winter archipelago. The image is very peaceful and serene." + ], + "image_footnote": [], + "bbox": [ + 637, + 469, + 776, + 577 + ], + "page_idx": 8 + }, + { + "type": "image", + "img_path": "images/a10b619707dca6b91163d7f9d93c02b5aa64594cf03a6c32941b916d65399e68.jpg", + "image_caption": [ + "A Japanese Porcelain Imari vase", + "Fig. 3: Comparison of Parrot and diffusion-based RL baselines. From left to right, we provide results of Stable diffusion 1.5 [39] (1st column), DPOK [12] (2nd column) with the weighted sum, Promptist [15] (3rd column), and Parrot (4th column)." + ], + "image_footnote": [], + "bbox": [ + 223, + 627, + 357, + 734 + ], + "page_idx": 8 + }, + { + "type": "image", + "img_path": "images/9d758abedc5dc52c4dcbdb8e4fb459a0f71a232438ea3ae857653039757e8b5c.jpg", + "image_caption": [ + "A Japanese Porcelain Imari vase" + ], + "image_footnote": [], + "bbox": [ + 361, + 627, + 500, + 734 + ], + "page_idx": 8 + }, + { + "type": "image", + "img_path": "images/d3990557d2afa7becae0faaa66fece9a3117ce00d334735a486402b3ccc58874.jpg", + "image_caption": [ + "Painting of a woman looking out on a winter archipelago, ultra realistic, concept art, intricate details, eerie", + "A Japanese Porcelain Imari vase, intricate, elegant, highly detailed, digital painting, artstation, concept art, smooth, sharp focus, illustration." + ], + "image_footnote": [], + "bbox": [ + 500, + 627, + 637, + 734 + ], + "page_idx": 8 + }, + { + "type": "image", + "img_path": "images/a7d7fcd3bfa8c03c77d3b056b2bd5e9ef2d2d94c54a50e0008938bbd07015da0.jpg", + "image_caption": [ + "A photo of a Japanese Porcelain Imari vase on the ground with a beautiful floral pattern on the front" + ], + "image_footnote": [], + "bbox": [ + 638, + 627, + 777, + 734 + ], + "page_idx": 8 + }, + { + "type": "header", + "text": "Pareto-optimal Multi-Reward RL Framework for T2I Generation", + "bbox": [ + 297, + 114, + 730, + 128 + ], + "page_idx": 8 + }, + { + "type": "page_number", + "text": "9", + "bbox": [ + 774, + 116, + 784, + 126 + ], + "page_idx": 8 + }, + { + "type": "image", + "img_path": "images/af97f6e4f7cef419bb1ec58ec2b9a03f8d699915b8b980f61b4a77bfbafc55fb.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 218, + 142, + 364, + 251 + ], + "page_idx": 9 + }, + { + "type": "image", + "img_path": "images/98afd077dd00f4121c3ecb6c6bbb0f054c14642bf77463f176a2a616a1f3063c.jpg", + "image_caption": [ + "Fig.4: Training curve for fine-tuning on weighted sum and Parrot. For weighted sum, WS1 denotes $\\{0.7, 0.1, 0.1, 0.1\\}$ and WS2 denotes $\\{0.25, 0.25, 0.25, 0.25\\}$ for aesthetics, human preference, text-image alignment and image sentiment. Using weighted sum leads to decrease in human preference score and image sentiment score despite an improvement in the aesthetic score. In contrast, Parrot exhibits stable increases across all metrics." + ], + "image_footnote": [], + "bbox": [ + 372, + 142, + 501, + 253 + ], + "page_idx": 9 + }, + { + "type": "image", + "img_path": "images/3f71f1be3076a9147b65313e7a6f37d0f4143330c01fb0134da47b1d0a3d73a2.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 503, + 142, + 643, + 253 + ], + "page_idx": 9 + }, + { + "type": "image", + "img_path": "images/8ab51a1ac6692c195c2948ef192e3f0f1a8f5a552fd98abea6a32dc546e2ad23.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 643, + 142, + 782, + 252 + ], + "page_idx": 9 + }, + { + "type": "table", + "img_path": "images/d400cd7619e318489381f63fe1bf3444b61a9b562550151e7a47e3cc0b71fbce.jpg", + "table_caption": [], + "table_footnote": [], + "table_body": "
ModelQuality Metrics
TIA (↑)Aesth. (↑)HP (↑)Sent. (↑)Average (↑)
SD 1.5 [39]0.23220.57550.19300.30100.3254
DPOK [12] (WS)0.23370.58130.19320.30130.3273 (+0.58%)
Parrot w/o PE0.23550.60340.20090.30180.3354 (+3.07%)
Parrot T2I Model Tuning Only0.25090.70730.33370.30520.3992 (+22.6%)
Promptist [15]0.14490.67830.27590.25180.3377 (+3.77 %)
Parrot with HP Only0.15430.59610.35280.25620.3398 (+4.42 %)
Parrot PEN Tuning Only0.16590.64920.26170.31310.3474 (+6.76 %)
Parrot w/o Joint Optimization0.16610.63080.25660.30840.3404 (+4.60 %)
Parrot w/o ori prompt guidance0.16230.71560.34250.31300.3833 (+17.8 %)
Parrot0.16670.73960.34110.31320.3901 (+19.8 %)
", + "bbox": [ + 215, + 381, + 787, + 529 + ], + "page_idx": 9 + }, + { + "type": "text", + "text": "Table 1: Quantitative comparison between Parrot and alternatives on the Parti dataset [53]. Abbreviations: WS - Weighted Sum; PE - Prompt Expansion; TIA - Text-Image Alignment; Aesth. - Aesthetics; HP - Human Preference; Sent. - Image Sentiment. TIA score is measured against the original prompt without expansion.", + "bbox": [ + 212, + 530, + 787, + 585 + ], + "page_idx": 9 + }, + { + "type": "text", + "text": "with the prompt expansion model that incorporates aesthetic keywords during training. Parrot generates results that are more closely aligned with the input prompt, as well as more visually pleasing.", + "bbox": [ + 212, + 636, + 787, + 683 + ], + "page_idx": 9 + }, + { + "type": "text", + "text": "Weighted sum vs. Parrot: Fig. 4 shows the training curve comparison of Parrot and using a linear combination of the reward scores. Each subgraph represents a reward. WS1 and WS2 denote two different weights with multiple reward scores. WS1 places greater emphasis on the aesthetic score, while WS2 adopts balanced weights across aesthetics, human preference, text-image alignment, and image sentiment. Employing the weighted sum of multiple rewards leads to a decrease in the image sentiment score, despite notable enhancements in aesthetics and human preference. In contrast, Parrot consistently exhibits improvement across all metrics.", + "bbox": [ + 212, + 703, + 787, + 840 + ], + "page_idx": 9 + }, + { + "type": "page_number", + "text": "10", + "bbox": [ + 217, + 114, + 235, + 126 + ], + "page_idx": 9 + }, + { + "type": "header", + "text": "Lee, Seung Hyun et al.", + "bbox": [ + 271, + 114, + 424, + 128 + ], + "page_idx": 9 + }, + { + "type": "text", + "text": "5.3 Quantitative Evaluation", + "text_level": 1, + "bbox": [ + 215, + 146, + 459, + 161 + ], + "page_idx": 10 + }, + { + "type": "text", + "text": "Comparison with Baselines: Table 1 presents our results of the quality score across four quality rewards: text-image alignment score, aesthetic score, human preference score, and emotion score. The first group shows methods without prompt expansion, and the second group compares methods with expansion. The prompt expansion and T2I generation are performed on the PartiPrompts [53]. Using a set of 1632 prompts, we generate 32 images for each text input and calculate the average for each metric.", + "bbox": [ + 212, + 178, + 787, + 282 + ], + "page_idx": 10 + }, + { + "type": "text", + "text": "$w / o$ PE indicates the generation of images solely based on original prompt without expansion. Note that $w / o$ PE is not the same as Parrot T2I Model Tuning Only. The former takes a model trained with PEN and removes it during inference, which results in a notable disparity between prompts used in training and testing. The latter trains with the multi-objective RL using only the T2I model, and it shows substantial improvement upon DPOK [12] (weighted sum) with balanced weights of $\\{0.25, 0.25, 0.25, 0.25\\}$ . This shows that the proposed RL method is indeed effective in multi-objective optimization.", + "bbox": [ + 212, + 284, + 787, + 404 + ], + "page_idx": 10 + }, + { + "type": "text", + "text": "For Promptist [15], we generate prompt expansion from their model. Parrot with HP Only shows solely using a human preference reward leads to decline in others. Parrot PEN Tuning Only shows suboptimal aesthetics and human preference. Parrot w/o Joint Optimization shows suboptimal results than Parrot which demonstrates the necessity of jointly optimizing PEN and T2I models.", + "bbox": [ + 212, + 404, + 787, + 479 + ], + "page_idx": 10 + }, + { + "type": "text", + "text": "Our method outperforms both compared methods in aesthetics, human preference and sentiment scores. The text-image alignment score is measured with the original prompt before expansion for fair comparison. As a result, the group without prompt expansion generally shows a higher text-image alignment score. Parrot shows better text-image alignment in each subgroup.", + "bbox": [ + 212, + 479, + 787, + 556 + ], + "page_idx": 10 + }, + { + "type": "table", + "img_path": "images/ab58a733a019a0ddbb54e1d279b97fd87bca774fdccae52907d98e4ef7f72144.jpg", + "table_caption": [], + "table_footnote": [], + "table_body": "
AreaQuestion
Aesthetics“Which image shows better aesthetics without blurry texture, unnatural focusing, and poor color combination?”
Human Preference“Which generated image do you prefer?”
Text-Image Alignment“Which image is well aligned with the text?”
Image Sentiment“Which image is closer to amusement, excitement, and contentment?”
", + "bbox": [ + 215, + 580, + 787, + 656 + ], + "page_idx": 10 + }, + { + "type": "text", + "text": "Table 2: Questions for user study. For performing user study, we carefully design questions suitable to each quality.", + "bbox": [ + 215, + 657, + 787, + 684 + ], + "page_idx": 10 + }, + { + "type": "text", + "text": "User Study: We conduct a user study using MTurk [1] with generated images from 100 random prompts in the PartiPrompts [53]. Five models are compared: Stable Diffusion v1.5, DPOK [12] with an equal weighted sum, Promptist [15], Parrot without prompt expansion, and Parrot. Each rater is presented with the original prompt (before expansion) and a set of five generated images, with the image order being randomized. Raters are then tasked with selecting the best image from the group, guided by questions outlined in Table 2. Each question", + "bbox": [ + 212, + 733, + 787, + 840 + ], + "page_idx": 10 + }, + { + "type": "header", + "text": "Pareto-optimal Multi-Reward RL Framework for T2I Generation", + "bbox": [ + 297, + 113, + 730, + 128 + ], + "page_idx": 10 + }, + { + "type": "page_number", + "text": "11", + "bbox": [ + 767, + 114, + 782, + 126 + ], + "page_idx": 10 + }, + { + "type": "image", + "img_path": "images/50643f67ee5cbb0900b727734672b7bff184750fe4d7005f80ec79720216fc4e.jpg", + "image_caption": [ + "Fig.5: User study results on PartiPrompts [53]. Parrot outperforms baselines across all metrics." + ], + "image_footnote": [], + "bbox": [ + 227, + 148, + 477, + 292 + ], + "page_idx": 11 + }, + { + "type": "image", + "img_path": "images/60c7bce9ff211310d4c2458526c643ff71c292892d6d409969b66f5973f634a7.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 491, + 143, + 776, + 284 + ], + "page_idx": 11 + }, + { + "type": "image", + "img_path": "images/511c367a10f90faea4dd8a326a1f5ec0de1aa1f9241e579495cf800156934a4a.jpg", + "image_caption": [ + "Fig. 6: Ablation study. We perform an ablation study by removing one of quality signals. We observe that each quality signal affects their improvement of (a) aesthetics, (b) human preference score, (c) text-image alignment score, (d) image sentiment score.", + "Fig.7: Comparing Parrot with/without Pareto Optimal. By pretending different reward-specific preferences, Parrot can change the tradeoff between rewards. Specifically, “”, “”, “”, “” indicates aesthetics, human preference, text-image alignment, and image sentiment respectively. With Pareto-optimal selection, each reward increases based on reward-specific preference." + ], + "image_footnote": [], + "bbox": [ + 217, + 369, + 784, + 525 + ], + "page_idx": 11 + }, + { + "type": "text", + "text": "pertains to a specific quality aspect: aesthetics, human preference, text-image alignment, and image sentiment. For each prompt, 20 rounds of random sampling are conducted and sent to different raters. The user study results, illustrated in Fig. 5, show that Parrot outperforms other baselines across all dimensions.", + "bbox": [ + 212, + 626, + 787, + 686 + ], + "page_idx": 11 + }, + { + "type": "text", + "text": "Proportion of non-dominated points: Using batch size of 256, we observe that in a batch around $20\\%$ to $30\\%$ are non-dominated points and the proportion of non-dominated points in single batch slightly increase as training proceeds.", + "bbox": [ + 212, + 694, + 787, + 739 + ], + "page_idx": 11 + }, + { + "type": "text", + "text": "5.4 Ablations", + "text_level": 1, + "bbox": [ + 214, + 762, + 341, + 776 + ], + "page_idx": 11 + }, + { + "type": "text", + "text": "Effect of Pareto-optimal Multi-reward RL: To show the efficacy of Pareto-optimal Multi-reward RL, we conduct an ablation study by removing one reward model at a time. Fig. 6 shows quantitative results using one hundred random text", + "bbox": [ + 212, + 794, + 785, + 839 + ], + "page_idx": 11 + }, + { + "type": "page_number", + "text": "12", + "bbox": [ + 217, + 114, + 235, + 126 + ], + "page_idx": 11 + }, + { + "type": "header", + "text": "Lee, Seung Hyun et al.", + "bbox": [ + 271, + 114, + 424, + 128 + ], + "page_idx": 11 + }, + { + "type": "image", + "img_path": "images/4945e6cd6998b9566e3a9fb25757a5fe57230c5b9429f8c1cbf73e047211a388.jpg", + "image_caption": [ + "Fig. 8: The comparisons of the diffusion fine-tuning between Pareto-optimal multi-reward RL and single reward RL. We show results with same seed from various methods: Stable Diffusion 1.5 [39] (1st column), T2I model fine-tuned with the aesthetic model (2nd column), the human preference model (3rd column), text-image alignment (4th column), image sentiment (5th column), Parrot without Pareto-optimal selection (6th column) and Parrot (7th column). Parrot is effective to generate acceptable images without sacrificing one of quality signals. For example, T2I model fine-tuned with a single quality signal such as aesthetics, human preference and image sentiment results in text-image misalignment, while our method achieves a balanced visual outcome across multiple criteria." + ], + "image_footnote": [], + "bbox": [ + 218, + 147, + 784, + 335 + ], + "page_idx": 12 + }, + { + "type": "text", + "text": "prompts from the Promptist [15]. We observe that our training scheme improves multiple target objectives.", + "bbox": [ + 212, + 511, + 784, + 541 + ], + "page_idx": 12 + }, + { + "type": "text", + "text": "To verify whether Parrot achieved better trade-off for different reward scores, we generate 1000 images from common animal dataset [4], where each text prompt consists of the name of a common animal. As shown in Fig. 7, using only text-image alignment with reward-specific preference $\\langle reward3\\rangle$ generates images with higher text-image alignment score, while using only aesthetic model with reward-specific preference $\\langle reward1\\rangle$ yields images with higher aesthetic score. In the case of using two reward-specific preferences $\\langle reward1\\rangle$ , $\\langle reward3\\rangle$ , we observe that scores are balanced and show that better results across multiple rewards than Parrot without Pareto optimal selection.", + "bbox": [ + 212, + 542, + 787, + 678 + ], + "page_idx": 12 + }, + { + "type": "text", + "text": "Fig 8 shows the visual comparison between Parrot, Parrot with a single reward, and Parrot without selecting the batch-wise Pareto-optimal solution. Using a single reward model tends to result in degradation of other rewards, especially text-image alignment. In the third column, results of the first row miss the text $a$ tophat in input prompt, even though the Stable Diffusion result includes that attribute. On the other hand, Parrot results capture all prompts, improving other quality signals, such as aesthetics, image sentiment and human preference.", + "bbox": [ + 212, + 680, + 787, + 787 + ], + "page_idx": 12 + }, + { + "type": "text", + "text": "Effect of Original Prompt Centered Guidance: Fig 9 shows the effect of the proposed original prompt-centered guidance. As evident from the figure, using only the expanded prompt as input often results in the main content be", + "bbox": [ + 212, + 794, + 787, + 840 + ], + "page_idx": 12 + }, + { + "type": "header", + "text": "Pareto-optimal Multi-Reward RL Framework for T2I Generation", + "bbox": [ + 297, + 114, + 730, + 127 + ], + "page_idx": 12 + }, + { + "type": "page_number", + "text": "13", + "bbox": [ + 767, + 116, + 784, + 126 + ], + "page_idx": 12 + }, + { + "type": "text", + "text": "ing overwhelmed by the added context. For instance, given the original prompt \"A shiba inu\", the result from the expanded prompt shows a zoomed-out image and the intended main subject (shiba inu) becomes small. The proposed original prompt-centered guidance effectively addresses this issue, generating an image that faithfully captures the input prompt while incorporating visually more pleasing details.", + "bbox": [ + 212, + 146, + 787, + 238 + ], + "page_idx": 13 + }, + { + "type": "image", + "img_path": "images/6a19c507fef9901d104d5d9638edfb87354d7390120f50624207b296ddd439f6.jpg", + "image_caption": [ + "Fig. 9: Results of original prompt centered guidance. As we expand the prompt, the content in the generated image often fades away. This guidance is helpful for keeping the main content of the original prompt." + ], + "image_footnote": [], + "bbox": [ + 217, + 266, + 787, + 375 + ], + "page_idx": 13 + }, + { + "type": "text", + "text": "6 Conclusion and Limitation", + "text_level": 1, + "bbox": [ + 215, + 484, + 509, + 501 + ], + "page_idx": 13 + }, + { + "type": "text", + "text": "We propose Parrot, a novel multi-reward optimization algorithm aimed to improve text-to-image generation by effectively optimizing multiple quality rewards using RL. With batch-wise Pareto-optimal selection, Parrot adaptively balance the optimization of multiple quality rewards. By applying Parrot to jointly fine-tune both the T2I model and the prompt expansion model, we achieve the generation of higher-quality images with richer details. Additionally, our original prompt centered guidance technique ensures that the generated image maintains fidelity to the user prompt after prompt expansion during inference. Results from the user study indicate that Parrot significantly improves the quality of generated images across multiple criteria, including text-image alignment, human preference, aesthetics, and image sentiment. While Parrot has shown effectiveness in enhancing generated image quality, its efficacy is limited by the quality metrics it relies on. Therefore, advancements of the generated image quality metrics will directly enhance the capabilities of Parrot. Additionally, Parrot is adaptable to a broader range of rewards that quantify generated image quality.", + "bbox": [ + 212, + 527, + 787, + 755 + ], + "page_idx": 13 + }, + { + "type": "text", + "text": "Societal Impact: Parrot could potentially raise ethical concerns related to the generation of immoral content. This concern stems from the user's ability to influence T2I generation, allowing for the creation of visual content that may be deemed inappropriate. The risk may be tied to the potential biases in reward models inherited from various datasets.", + "bbox": [ + 212, + 763, + 787, + 839 + ], + "page_idx": 13 + }, + { + "type": "page_number", + "text": "14", + "bbox": [ + 217, + 114, + 235, + 126 + ], + "page_idx": 13 + }, + { + "type": "header", + "text": "Lee, Seung Hyun et al.", + "bbox": [ + 271, + 114, + 424, + 128 + ], + "page_idx": 13 + }, + { + "type": "text", + "text": "References", + "text_level": 1, + "bbox": [ + 215, + 143, + 323, + 160 + ], + "page_idx": 14 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "1. Amazon mechanical turk. https://www.mturk.com/ (2005)", + "2. Anil, R., Dai, A.M., First, O., Johnson, M., Lepikhin, D., Passos, A., Shakeri, S., Taropa, E., Bailey, P., Chen, Z., et al.: Palm 2 technical report. arXiv preprint arXiv:2305.10403 (2023)", + "3. Bai, Y., Jones, A., Ndousse, K., Askell, A., Chen, A., DasSarma, N., Drain, D., Fort, S., Ganguli, D., Henighan, T., et al.: Training a helpful and harmless assistant with reinforcement learning from human feedback. arXiv preprint arXiv:2204.05862 (2022)", + "4. Black, K., Janner, M., Du, Y., Kostrikov, I., Levine, S.: Training diffusion models with reinforcement learning. arXiv preprint arXiv:2305.13301 (2023)", + "5. Chang, H., Zhang, H., Barber, J., Maschinot, A., Lezama, J., Jiang, L., Yang, M.H., Murphy, K., Freeman, W.T., Rubinstein, M., et al.: Muse: Text-to-image generation via masked generative transformers. arXiv preprint arXiv:2301.00704 (2023)", + "6. Clark, K., Vicol, P., Swersky, K., Fleet, D.J.: Directly fine-tuning diffusion models on differentiable rewards. ICLR (2024)", + "7. Dai, X., Hou, J., Ma, C.Y., Tsai, S., Wang, J., Wang, R., Zhang, P., Vandenhende, S., Wang, X., Dubey, A., et al.: Emu: Enhancing image generation models using photogenic needles in a haystack. arXiv preprint arXiv:2309.15807 (2023)", + "8. Deng, F., Wang, Q., Wei, W., Grundmann, M., Hou, T.: Prdp: Proximal reward difference prediction for large-scale reward finetuning of diffusion models. CVPR (2024)", + "9. Dong, H., Xiong, W., Goyal, D., Pan, R., Dao, S., Zhang, J., Shum, K., Zhang, T.: Raft: Reward ranked finetuning for generative foundation model alignment. arXiv preprint arXiv:2304.06767 (2023)", + "10. Dosovitskiy, A., Beyer, L., Kolesnikov, A., Weissenborn, D., Zhai, X., Unterthiner, T., Dehghani, M., Minderer, M., Heigold, G., Gelly, S., et al.: An image is worth 16x16 words: Transformers for image recognition at scale. arXiv preprint arXiv:2010.11929 (2020)", + "11. Fan, Y., Lee, K.: Optimizing ddpm sampling with shortcut fine-tuning. ICML (2023)", + "12. Fan, Y., Watkins, O., Du, Y., Liu, H., Ryu, M., Boutilier, C., Abbeel, P., Ghavamzadeh, M., Lee, K., Lee, K.: Dpok: Reinforcement learning for fine-tuning text-to-image diffusion models. NeurIPS (2023)", + "13. Fang, Y., Zhu, H., Zeng, Y., Ma, K., Wang, Z.: Perceptual quality assessment of smartphone photography. In: CVPR (2020)", + "14. Han, L., Li, Y., Zhang, H., Milanfar, P., Metaxas, D., Yang, F.: Svdiff: Compact parameter space for diffusion fine-tuning. CVPR (2023)", + "15. Hao, Y., Chi, Z., Dong, L., Wei, F.: Optimizing prompts for text-to-image generation. arXiv preprint arXiv:2212.09611 (2022)", + "16. He, H., Wang, T., Yang, H., Fu, J., Yuan, N.J., Yin, J., Chao, H., Zhang, Q.: Learning profitable nft image diffusions via multiple visual-policy guided reinforcement learning. arXiv preprint arXiv:2306.11731 (2023)", + "17. Ho, J., Jain, A., Abbeel, P.: Denoising diffusion probabilistic models. NeurIPS (2020)", + "18. Ho, J., Salimans, T.: Classifier-free diffusion guidance. arXiv preprint arXiv:2207.12598 (2022)" + ], + "bbox": [ + 218, + 178, + 785, + 839 + ], + "page_idx": 14 + }, + { + "type": "header", + "text": "Pareto-optimal Multi-Reward RL Framework for T2I Generation", + "bbox": [ + 297, + 114, + 730, + 128 + ], + "page_idx": 14 + }, + { + "type": "page_number", + "text": "15", + "bbox": [ + 767, + 116, + 785, + 126 + ], + "page_idx": 14 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "19. Hosu, V., Lin, H., Sziranyi, T., Saupe, D.: Koniq-10k: An ecologically valid database for deep learning of blind image quality assessment. TIP 29, 4041-4056 (2020)", + "20. Hu, E.J., Shen, Y., Wallis, P., Allen-Zhu, Z., Li, Y., Wang, S., Wang, L., Chen, W.: Lora: Low-rank adaptation of large language models. arXiv preprint arXiv:2106.09685 (2021)", + "21. Jeong, Y., Ryoo, W., Lee, S., Seo, D., Byeon, W., Kim, S., Kim, J.: The power of sound (tpos): Audio reactive video generation with stable diffusion. In: ICCV (2023)", + "22. Kawar, B., Zada, S., Lang, O., Tov, O., Chang, H., Dekel, T., Mosseri, I., Irani, M.: Imagic: Text-based real image editing with diffusion models. In: CVPR (2023)", + "23. Ke, J., Wang, Q., Wang, Y., Milanfar, P., Yang, F.: Musiq: Multi-scale image quality transformer. In: ICCV (2021)", + "24. Ke, J., Ye, K., Yu, J., Wu, Y., Milanfar, P., Yang, F.: Vila: Learning image aesthetics from user comments with vision-language pretraining. In: CVPR (2023)", + "25. Kingma, D.P., Ba, J.: Adam: A method for stochastic optimization. ICLR (2015)", + "26. Kirstain, Y., Polyak, A., Singer, U., Matiana, S., Penna, J., Levy, O.: Pick-a-pic: An open dataset of user preferences for text-to-image generation. arXiv preprint arXiv:2305.01569 (2023)", + "27. Lee, K., Liu, H., Ryu, M., Watkins, O., Du, Y., Boutilier, C., Abbeel, P., Ghavamzadeh, M., Gu, S.S.: Aligning text-to-image models using human feedback. arXiv preprint arXiv:2302.12192 (2023)", + "28. Lee, S.H., Kim, S., Yoo, I., Yang, F., Cho, D., Kim, Y., Chang, H., Kim, J., Kim, S.: Soundini: Sound-guided diffusion for natural video editing. arXiv preprint arXiv:2304.06818 (2023)", + "29. Li, Y., Liu, H., Wu, Q., Mu, F., Yang, J., Gao, J., Li, C., Lee, Y.J.: Gligen: Open-set grounded text-to-image generation. In: CVPR (2023)", + "30. Lin, X., Yang, Z., Zhang, X., Zhang, Q.: Pareto set learning for expensive multi-objective optimization. NeurIPS (2022)", + "31. Mannor, S., Shimkin, N.: The steering approach for multi-criteria reinforcement learning. NeurIPS (2001)", + "32. Miettinen, K.: Nonlinear multiobjective optimization, vol. 12. Springer Science & Business Media (1999)", + "33. Murray, N., Marchesotti, L., Perronnin, F.: Ava: A large-scale database for aesthetic visual analysis. In: CVPR (2012)", + "34. Ouyang, L., Wu, J., Jiang, X., Almeida, D., Wainwright, C., Mishkin, P., Zhang, C., Agarwal, S., Slama, K., Ray, A., et al.: Training language models to follow instructions with human feedback. NeurIPS (2022)", + "35. Radford, A., Kim, J.W., Hallacy, C., Ramesh, A., Goh, G., Agarwal, S., Sastry, G., Askell, A., Mishkin, P., Clark, J., et al.: Learning transferable visual models from natural language supervision. In: ICML (2021)", + "36. Rame, A., Couairon, G., Shukor, M., Dancette, C., Gaya, J.B., Soulier, L., Cord, M.: Rewarded soups: towards pareto-optimal alignment by interpolating weights fine-tuned on diverse rewards. NeurIPS (2023)", + "37. Ramesh, A., Dhariwal, P., Nichol, A., Chu, C., Chen, M.: Hierarchical text-conditional image generation with clip latents. arXiv preprint arXiv:2204.06125 (2022)", + "38. Richardson, E., Goldberg, K., Alaluf, Y., Cohen-Or, D.: Conceptlab: Creative generation using diffusion prior constraints. arXiv preprint arXiv:2308.02669 (2023)", + "39. Rombach, R., Blattmann, A., Lorenz, D., Esser, P., Ommer, B.: High-resolution image synthesis with latent diffusion models. In: CVPR (2022)" + ], + "bbox": [ + 215, + 146, + 785, + 839 + ], + "page_idx": 15 + }, + { + "type": "page_number", + "text": "16", + "bbox": [ + 217, + 114, + 235, + 126 + ], + "page_idx": 15 + }, + { + "type": "header", + "text": "Lee, Seung Hyun et al.", + "bbox": [ + 271, + 114, + 424, + 127 + ], + "page_idx": 15 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "40. Ronneberger, O., Fischer, P., Brox, T.: U-net: Convolutional networks for biomedical image segmentation. In: MICCAI (2015)", + "41. Sahara, C., Chan, W., Chang, H., Lee, C., Ho, J., Salimans, T., Fleet, D., Norouzi, M.: Palette: Image-to-image diffusion models. In: SIGGRAPH (2022)", + "42. Sahara, C., Chan, W., Saxena, S., Li, L., Whang, J., Denton, E.L., Ghasemipour, K., Gontijo Lopes, R., Karagol Ayan, B., Salimans, T., et al.: Photorealistic text-to-image diffusion models with deep language understanding. NeurIPS (2022)", + "43. Schuhmann, C., Beaumont, R., Vencu, R., Gordon, C., Wightman, R., Cherti, M., Coombes, T., Katta, A., Mullis, C., Wortsman, M., et al.: Laion-5b: An open large-scale dataset for training next generation image-text models. NeurIPS (2022)", + "44. Serra, A., Carrara, F., Tesconi, M., Falchi, F.: The emotions of the crowd: Learning image sentiment from tweets via cross-modal distillation. ECAI (2023)", + "45. Sohl-Dickstein, J., Weiss, E., Maheswaranathan, N., Ganguli, S.: Deep unsupervised learning using nonequilibrium thermodynamics. In: ICML (2015)", + "46. Song, J., Meng, C., Ermon, S.: Denoising diffusion implicit models. arXiv preprint arXiv:2010.02502 (2020)", + "47. Tesauro, G., Das, R., Chan, H., Kephart, J., Levine, D., Rawson, F., Lefury, C.: Managing power consumption and performance of computing systems using reinforcement learning. NeurIPS (2007)", + "48. Tu, Z., Talebi, H., Zhang, H., Yang, F., Milanfar, P., Bovik, A., Li, Y.: Maxvit: Multi-axis vision transformer. In: ECCV (2022)", + "49. Vaswani, A., Shazeer, N., Parmar, N., Uszkoreit, J., Jones, L., Gomez, A.N., Kaiser, L., Polosukhin, I.: Attention is all you need. NeurIPS (2017)", + "50. Wu, X., Sun, K., Zhu, F., Zhao, R., Li, H.: Human preference score: Better aligning text-to-image models with human preference. In: ICCV (2023)", + "51. Xu, J., Liu, X., Wu, Y., Tong, Y., Li, Q., Ding, M., Tang, J., Dong, Y.: Imagereward: Learning and evaluating human preferences for text-to-image generation. arXiv preprint arXiv:2304.05977 (2023)", + "52. Ying, Z., Niu, H., Gupta, P., Mahajan, D., Ghadiyaram, D., Bovik, A.: From patches to pictures (paq-2-piq): Mapping the perceptual space of picture quality. In: CVPR (2020)", + "53. Yu, J., Xu, Y., Koh, J.Y., Luong, T., Baid, G., Wang, Z., Vasudevan, V., Ku, A., Yang, Y., Ayan, B.K., et al.: Scaling autoregressive models for content-rich text-to-image generation. arXiv preprint arXiv:2206.10789 (2022)", + "54. Yu, L., Cheng, Y., Sohn, K., Lezama, J., Zhang, H., Chang, H., Hauptmann, A.G., Yang, M.H., Hao, Y., Essa, I., et al.: Magvit: Masked generative video transformer. In: CVPR (2023)", + "55. Zhou, Y., Liu, B., Zhu, Y., Yang, X., Chen, C., Xu, J.: Shifted diffusion for text-to-image generation. In: CVPR (2023)" + ], + "bbox": [ + 215, + 146, + 784, + 686 + ], + "page_idx": 16 + }, + { + "type": "header", + "text": "Pareto-optimal Multi-Reward RL Framework for T2I Generation", + "bbox": [ + 297, + 114, + 730, + 128 + ], + "page_idx": 16 + }, + { + "type": "page_number", + "text": "17", + "bbox": [ + 767, + 114, + 785, + 126 + ], + "page_idx": 16 + } +] \ No newline at end of file diff --git a/2024/Parrot_ Pareto-optimal Multi-Reward Reinforcement Learning Framework for Text-to-Image Generation/938566d1-b778-4bb0-b908-cbc4a31c6445_model.json b/2024/Parrot_ Pareto-optimal Multi-Reward Reinforcement Learning Framework for Text-to-Image Generation/938566d1-b778-4bb0-b908-cbc4a31c6445_model.json new file mode 100644 index 0000000000000000000000000000000000000000..4087527b20a105c9fe3964b0e109ef6239f3ed75 --- /dev/null +++ b/2024/Parrot_ Pareto-optimal Multi-Reward Reinforcement Learning Framework for Text-to-Image Generation/938566d1-b778-4bb0-b908-cbc4a31c6445_model.json @@ -0,0 +1,2808 @@ +[ + [ + { + "type": "title", + "bbox": [ + 0.278, + 0.141, + 0.728, + 0.21 + ], + "angle": 0, + "content": "Parrot: Pareto-optimal Multi-Reward Reinforcement Learning Framework for Text-to-Image Generation" + }, + { + "type": "text", + "bbox": [ + 0.23, + 0.234, + 0.773, + 0.283 + ], + "angle": 0, + "content": "Seung Hyun Lee\\(^{1,6*}\\), Yinxiao Li\\(^{1}\\), Junjie Ke\\(^{1}\\), Innfarn Yoo\\(^{2}\\), Han Zhang\\(^{3}\\), Jiahui Yu\\(^{4\\dagger}\\), Qifei Wang\\(^{2}\\), Fei Deng\\(^{2,5*}\\), Glenn Entis\\(^{1}\\), Junfeng He\\(^{1}\\), Gang Li\\(^{1}\\), Sangpil Kim\\(^{6}\\), Irfan Essa\\(^{1}\\), Feng Yang\\(^{1}\\)" + }, + { + "type": "text", + "bbox": [ + 0.307, + 0.292, + 0.697, + 0.322 + ], + "angle": 0, + "content": "Google Research1, Google2, Google DeepMind3, OpenAI4, Rutgers University5, Korea University6" + }, + { + "type": "image", + "bbox": [ + 0.218, + 0.325, + 0.789, + 0.572 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.216, + 0.575, + 0.788, + 0.618 + ], + "angle": 0, + "content": "Fig. 1: Parrot visual examples. Parrot consistently improves the quality of generated images across multiple criteria: aesthetics, human preference, text-image alignment, and image sentiment. Each column shows generated images using the same seed." + }, + { + "type": "text", + "bbox": [ + 0.261, + 0.641, + 0.744, + 0.807 + ], + "angle": 0, + "content": "Abstract. Recent works have demonstrated that using reinforcement learning (RL) with multiple quality rewards can improve the quality of generated images in text-to-image (T2I) generation. However, manually adjusting reward weights poses challenges and may cause over-optimization in certain metrics. To solve this, we propose Parrot, which addresses the issue through multi-objective optimization and introduces an effective multi-reward optimization strategy to approximate Pareto optimal. Utilizing batch-wise Pareto optimal selection, Parrot automatically identifies the optimal trade-off among different rewards. We use the novel multi-reward optimization algorithm to jointly optimize the T2I model and a prompt expansion network, resulting in significant improvement of image quality and also allow to control the trade-off of" + }, + { + "type": "page_footnote", + "bbox": [ + 0.232, + 0.817, + 0.545, + 0.829 + ], + "angle": 0, + "content": "* This work was done during an internship at Google." + }, + { + "type": "page_footnote", + "bbox": [ + 0.232, + 0.829, + 0.514, + 0.84 + ], + "angle": 0, + "content": "† This work was done during working at Google." + }, + { + "type": "list", + "bbox": [ + 0.232, + 0.817, + 0.545, + 0.84 + ], + "angle": 0, + "content": null + } + ], + [ + { + "type": "page_number", + "bbox": [ + 0.218, + 0.116, + 0.23, + 0.127 + ], + "angle": 0, + "content": "2" + }, + { + "type": "header", + "bbox": [ + 0.272, + 0.115, + 0.426, + 0.129 + ], + "angle": 0, + "content": "Lee, Seung Hyun et al." + }, + { + "type": "text", + "bbox": [ + 0.261, + 0.148, + 0.744, + 0.233 + ], + "angle": 0, + "content": "different rewards using a reward related prompt during inference. Furthermore, we introduce original prompt-centered guidance at inference time, ensuring fidelity to user input after prompt expansion. Extensive experiments and a user study validate the superiority of Parrot over several baselines across various quality criteria, including aesthetics, human preference, text-image alignment, and image sentiment." + }, + { + "type": "title", + "bbox": [ + 0.217, + 0.26, + 0.377, + 0.277 + ], + "angle": 0, + "content": "1 Introduction" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.295, + 0.789, + 0.431 + ], + "angle": 0, + "content": "Despite significant advancements in text-to-image (T2I) generation [17, 29, 38, 45, 55], recent work like Imagen [42] and Stable Diffusion [39], still struggle to produce high quality images. Images in the first row in Fig. 1 illustrates such quality issues in Stable Diffusion [39], including poor composition (e.g. bad cropping), misalignment with input prompts (e.g. missing objects), or overall lack of aesthetic appeal. Assessing the quality of generated images can involve various metrics such as aesthetics [23, 24, 33], human preference [26], text-image alignment [35], and emotional appeal [44]. Enhancing T2I generation across multiple quality metrics remains a challenging task." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.432, + 0.79, + 0.583 + ], + "angle": 0, + "content": "Recent works [4,12,15,27] have demonstrated that incorporating quality signals as reward functions in fine-tuning T2I with reinforcement learning (RL) can improve image quality. For instance, Promptist [15] fine-tunes prompt expansion model using RL with the sum of aesthetics and text-image alignment scores. However, the simple weighted sum approach may not effectively handle trade-offs among multiple quality metrics. As the number of rewards increases, manually adjusting reward weights becomes impractical. Moreover, optimizing one quality metric may inadvertently compromise others, as the model might prioritize aesthetics over relevance to the input prompt. Additionally, the trade-off for different reward is not controllable after training." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.584, + 0.79, + 0.842 + ], + "angle": 0, + "content": "To address these challenges, we propose Parrot, a Pareto-optimal multi-reward reinforcement learning algorithm to improve text-to-image generation. Unlike previous approaches that treat T2I reward optimization as a single objective optimization problem, Parrot tackles this challenge through multi-objective optimization and introduces an effective multi-reward optimization strategy to achieve Pareto optimal approximation. Intuitively, each generated sample in a batch embodies a distinctive trade-off among various quality rewards, with some samples exhibiting superior trade-offs compared to others. Instead of updating gradients using all batch samples, Parrot uses non-dominated points [32] which have better trade-offs. Consequently, Parrot automatically learns from the optimal trade-off among different rewards. Moreover, Parrot learns reward-specific preference prompts, which can be utilized individually or in combination to control the trade-off among different rewards during inference time. Unlike prior work, which either solely fine-tunes the T2I model [12] or only tunes the prompt expansion network while freezing the T2I model [15], we employ the Parrot multi-reward optimization algorithm to jointly optimize both the T2I model and the prompt expansion network. This collaborative optimization unlocks the" + } + ], + [ + { + "type": "header", + "bbox": [ + 0.299, + 0.115, + 0.732, + 0.129 + ], + "angle": 0, + "content": "Pareto-optimal Multi-Reward RL Framework for T2I Generation" + }, + { + "type": "page_number", + "bbox": [ + 0.775, + 0.117, + 0.785, + 0.127 + ], + "angle": 0, + "content": "3" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.147, + 0.788, + 0.223 + ], + "angle": 0, + "content": "full potential of Parrot by encouraging both more details from added context from the prompt expansion model, and the overall quality improvement on the T2I generation. During inference, we further introduce original prompt-centered guidance to ensure the output image is relevant to input prompts after prompt expansion." + }, + { + "type": "text", + "bbox": [ + 0.239, + 0.224, + 0.637, + 0.239 + ], + "angle": 0, + "content": "In summary, our contributions can be listed as follows:" + }, + { + "type": "text", + "bbox": [ + 0.226, + 0.241, + 0.784, + 0.329 + ], + "angle": 0, + "content": "- We propose Parrot, a novel multi-reward optimization algorithm for T2I RL fine-tuning. Leveraging batch-wise Pareto-optimal selection, it effectively optimizes multiple T2I rewards, enabling collaborative improvement in aesthetics, human preference, image sentiment, and text-image alignment and also allowing to control the trade-off of different rewards using a reward related prompt during inference." + }, + { + "type": "text", + "bbox": [ + 0.225, + 0.331, + 0.784, + 0.36 + ], + "angle": 0, + "content": "- We show the advantage of jointly optimizing both the prompt expansion network and the T2I model, which has never been explored before." + }, + { + "type": "text", + "bbox": [ + 0.225, + 0.362, + 0.784, + 0.405 + ], + "angle": 0, + "content": "- We introduce original prompt-centered guidance during inference time after prompt expansion, ensuring better alignment with the original prompt while enriching image details." + }, + { + "type": "text", + "bbox": [ + 0.225, + 0.407, + 0.784, + 0.435 + ], + "angle": 0, + "content": "- Extensive results and a user study validate that Parrot outperforms several baseline methods across various quality criteria." + }, + { + "type": "list", + "bbox": [ + 0.225, + 0.241, + 0.784, + 0.435 + ], + "angle": 0, + "content": null + }, + { + "type": "title", + "bbox": [ + 0.216, + 0.468, + 0.388, + 0.484 + ], + "angle": 0, + "content": "2 Related Work" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.514, + 0.788, + 0.621 + ], + "angle": 0, + "content": "T2I Generation: The goal of T2I generation is to create an image given an input text prompt. Several T2I generative models have been proposed and have demonstrated promising results [5, 7, 14, 21, 22, 28, 37, 41, 42, 54]. Stable Diffusion [39] shows impressive generation performance in T2I generation, leveraging latent text representations from LLMs. Despite substantial progress, the images generated by those models still exhibit quality issues, such as bad cropping or misalignment with the input texts." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.629, + 0.788, + 0.84 + ], + "angle": 0, + "content": "RL for T2I Fine-tuning: Starting by Fan et al. [11] to explore RL fine-tuning for T2I models, following works [4, 8, 9, 12, 16] have explored RL fine-tuning technique for T2I diffusion model, showcasing superior performance for human preference learning. DPOK [12] improves quality through RL using ImageReward [51] score as a reward with a few prompts. In addition to fine-tuning the T2I model directly using RL, Promptist [15] fine-tunes the prompt expansion model by using a simple sum of aesthetic and text-image alignment scores as reward. DRaFT [6] proposed not only differentiable rewards for efficient fine-tuning but also effectiveness of using linear summation of multi-rewards. These methods treat T2I RL as a single-objective optimization problem, while Parrot employs multi-objective optimization. Additionally, prior approaches either fine-tune the T2I model or the prompt expansion model while freezing the other. In contrast, Parrot proposes joint optimization of the prompt expansion model and the T2I model using multi-reward RL to foster better collaboration." + } + ], + [ + { + "type": "page_number", + "bbox": [ + 0.218, + 0.116, + 0.23, + 0.127 + ], + "angle": 0, + "content": "4" + }, + { + "type": "header", + "bbox": [ + 0.272, + 0.115, + 0.426, + 0.129 + ], + "angle": 0, + "content": "Lee, Seung Hyun et al." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.147, + 0.788, + 0.343 + ], + "angle": 0, + "content": "Multi-objective Optimization: Multi-objective optimization problem involves optimizing multiple objective functions simultaneously. The scalarization technique [31, 47] formulates multi-objective problem into single-objective problems with the weighted sum of each score, which requires pre-defined weights for each objective. Rame et al. [36] proposed weighted averaging method to find Pareto frontier, leveraging multiple fine-tuned models. Lin et al. [30] proposes to learn a set model to map trade-off preference vectors to their corresponding Pareto solutions. Inspired by this, Parrot introduces a language-based preference vector constructed from task identifiers for each reward, then encoded by the text-encoder. In the context of multi-reward RL for T2I diffusion models, Promptist [15] uses a simple weighted sum of two reward scores. This approach requires manual tuning of the weights, which makes it time-consuming and hard to scale when the number of rewards increases." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.351, + 0.788, + 0.579 + ], + "angle": 0, + "content": "Generated Image Quality: The quality assessment of images generated by T2I models involves multiple dimensions, and various metrics have been proposed. In this paper, we consider using four types of quality metrics as rewards: aesthetics, human preference, text-image alignment, and image sentiment. Aesthetics captures the overall visual appealingness of the image, and it is learned using human ratings for aesthetics in real images [13, 19, 23, 24, 33, 48, 52]. Human preferences, rooted the concept of learning from human feedback [3, 34, 50], involves gathering preferences at scale by having raters to compare generated images [26, 51]. Text-image alignment measures the extent to which the generated image aligns with the input prompt, CLIP [35] score is often employed, measuring the cosine distance of between contrastive image embedding and text embedding. Image sentiment is important for ensuring the generated image evokes positive emotions in the viewer. Serra et al. [44] predict average polarity of sentiments an image elicits and learn estimates for positive, neutral, and negative scores. In Parrot, we use its positive score as a reward for positive emotions." + }, + { + "type": "title", + "bbox": [ + 0.216, + 0.604, + 0.37, + 0.623 + ], + "angle": 0, + "content": "3 Preliminary" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.646, + 0.788, + 0.753 + ], + "angle": 0, + "content": "Diffusion Probabilistic Models: Diffusion probabilistic models [17] generate the image by gradually denoising a noisy image. Specifically, given a real image \\(\\mathbf{x}_0\\) from the data distribution \\(\\mathbf{x}_0 \\sim q(\\mathbf{x}_0)\\), the forward process \\(q(\\mathbf{x}_t | \\mathbf{x}_0, c)\\) of diffusion probabilistic models produce a noisy image \\(\\mathbf{x}_t\\), which induces a distribution \\(p(\\mathbf{x}_0, c)\\) conditioned on text prompt \\(c\\). In classifier-free guidance [18], denoising model predicts noise \\(\\bar{\\epsilon}_{\\theta}\\) with a linear combination of the unconditional score estimates \\(\\epsilon_{\\theta}(\\mathbf{x}_t, t)\\) and the conditional score estimates \\(\\epsilon_{\\theta}(\\mathbf{x}_t, t, c)\\) as follows:" + }, + { + "type": "equation", + "bbox": [ + 0.347, + 0.767, + 0.786, + 0.784 + ], + "angle": 0, + "content": "\\[\n\\bar {\\epsilon} _ {\\theta} = w \\cdot \\epsilon_ {\\theta} (\\mathbf {x} _ {t}, t, c) + (1 - w) \\cdot \\epsilon_ {\\theta} (\\mathbf {x} _ {t}, t, \\mathrm {n u l l}), \\tag {1}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.795, + 0.788, + 0.843 + ], + "angle": 0, + "content": "where \\( t \\) denotes diffusion time step, the null indicates a null text and \\( w \\) represents the guidance scale of classifier-free guidance where \\( w \\geq 1 \\). Note that \\( \\epsilon_{\\theta} \\) is typically parameterized by the UNet [40]." + } + ], + [ + { + "type": "header", + "bbox": [ + 0.299, + 0.115, + 0.733, + 0.129 + ], + "angle": 0, + "content": "Pareto-optimal Multi-Reward RL Framework for T2I Generation" + }, + { + "type": "page_number", + "bbox": [ + 0.775, + 0.117, + 0.785, + 0.127 + ], + "angle": 0, + "content": "5" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.147, + 0.784, + 0.283 + ], + "angle": 0, + "content": "RL-based T2I Diffusion Model Fine-tuning: Given a reward signal from generated images, the goal of RL-tuning for T2I diffusion models is to optimize the policy defined as one denoising step of T2I diffusion models. In particular, Black et al. [4] apply policy gradient algorithm, which regards the denoising process of diffusion models as a Markov decision process (MDP) by performing multiple denoising steps iteratively. Subsequently, a black box reward model \\( r(\\cdot ,\\cdot) \\) predicts a single scalar value from sampled image \\( \\mathbf{x}_0 \\). Given text condition \\( c\\sim p(c) \\) and image \\( \\mathbf{x}_0 \\), objective function \\( \\mathcal{I} \\) can be defined to maximize the expected reward as follows:" + }, + { + "type": "equation", + "bbox": [ + 0.403, + 0.292, + 0.785, + 0.309 + ], + "angle": 0, + "content": "\\[\n\\mathcal {J} _ {\\theta} = \\mathbb {E} _ {p (c)} \\mathbb {E} _ {p _ {\\theta} (\\mathbf {x} _ {0} | c)} [ r (\\mathbf {x} _ {0}, c) ], \\tag {2}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.314, + 0.784, + 0.375 + ], + "angle": 0, + "content": "where the pre-trained diffusion model \\(p_{\\theta}\\) produces a sample distribution \\(p_{\\theta}(\\mathbf{x}_0|c)\\) using text condition \\(c\\). Modifying this equation, Fan et al. [12] demonstrate that the gradient of objective function \\(\\nabla \\mathcal{I}_{\\theta}\\) can be calculated through gradient ascent algorithm without using the gradient of reward model as follows:" + }, + { + "type": "equation", + "bbox": [ + 0.345, + 0.382, + 0.785, + 0.422 + ], + "angle": 0, + "content": "\\[\n\\nabla \\mathcal {J} _ {\\theta} = \\mathbb {E} \\left[ r \\left(\\mathbf {x} _ {0}, c\\right) \\sum_ {t = 1} ^ {T} \\nabla_ {\\theta} \\log p _ {\\theta} \\left(\\mathbf {x} _ {t - 1} \\mid c, t, \\mathbf {x} _ {t}\\right) \\right], \\tag {3}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.215, + 0.428, + 0.784, + 0.459 + ], + "angle": 0, + "content": "where \\( T \\) denotes the total time step of the diffusion sampling. With parameters \\( \\theta \\), the expectation value can be taken over the trajectories of diffusion sampling." + }, + { + "type": "title", + "bbox": [ + 0.215, + 0.478, + 0.331, + 0.494 + ], + "angle": 0, + "content": "4 Method" + }, + { + "type": "title", + "bbox": [ + 0.215, + 0.507, + 0.401, + 0.521 + ], + "angle": 0, + "content": "4.1 Parrot Overview" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.528, + 0.788, + 0.756 + ], + "angle": 0, + "content": "Fig. 2 shows the overview of Parrot, which consists of the prompt expansion network (PEN) \\( p_{\\phi} \\) and the T2I diffusion model \\( p_{\\theta} \\). The PEN is first initialized from a supervised fine-tuning checkpoint on demonstrations of prompt expansion pairs, and the T2I model is initialized from pretrained diffusion model. Given the original prompt \\( c \\), the PEN generates an expanded prompt \\( \\hat{c} \\), and the T2I model generates images based on this expanded prompt. During the multi-reward RL fine-tuning, a batch of \\( N \\) images is sampled, and multiple quality rewards are calculated for each image, encompassing aspects like text-image alignment, aesthetics, human preference, and image sentiment. Based on these reward scores, Parrot identifies the batch-wise Pareto-optimal set using a non-dominated sorting algorithm. This optimal set of images is then used for joint optimization of the PEN and T2I model parameters through RL policy gradient update. During inference, Parrot leverages both the original prompt and its expansion, striking a balance between maintaining faithfulness to the original prompt and incorporating additional details for higher quality." + }, + { + "type": "title", + "bbox": [ + 0.215, + 0.774, + 0.563, + 0.789 + ], + "angle": 0, + "content": "4.2 Batch-wise Pareto-optimal Selection" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.795, + 0.787, + 0.842 + ], + "angle": 0, + "content": "Li et al. [30] has demonstrated that using batchwise Pareto-set learning and selecting good samples in a batch can approximate Pareto-optimality across multiple objectives. Backed up by this theory, we propose to select non-dominated" + } + ], + [ + { + "type": "page_number", + "bbox": [ + 0.218, + 0.116, + 0.23, + 0.127 + ], + "angle": 0, + "content": "6" + }, + { + "type": "header", + "bbox": [ + 0.272, + 0.115, + 0.426, + 0.129 + ], + "angle": 0, + "content": "Lee, Seung Hyun et al." + }, + { + "type": "image", + "bbox": [ + 0.22, + 0.144, + 0.788, + 0.285 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.214, + 0.288, + 0.789, + 0.387 + ], + "angle": 0, + "content": "Fig. 2: Overview of Parrot. During the training, \\(N\\) images are sampled from the T2I model using the expanded prompt from the prompt expansion network. Multiple quality rewards are calculated for each image, and the Pareto-optimal set is identified using the non-dominated sorting algorithm. These optimal images are then used to perform policy gradient update of the parameters of T2I model and prompt expansion network jointly. During the inference, both the original prompt and the expanded prompt are provided to the T2I model, enabling better faithfulness while adding detail." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.416, + 0.789, + 0.629 + ], + "angle": 0, + "content": "points in a batch for policy gradient update in RL to achieve Pareto-optimality. Algorithm 1 outlines the procedure of Parrot. Rather than updating the gradients using all images, Parrot focuses on high-quality samples, considering multiple quality rewards in each mini-batch. In the multi-reward RL, each sample generated by the T2I model presents distinct trade-offs for different rewards. Among these samples, a subset with varied optimal trade-offs across multiple objectives, also known as the Pareto set, exists. For a Pareto-optimal sample, none of its objective values can be further improved without damaging others. In other words, the Pareto-optimal set is not dominated by any data points, also known as the non-dominated set. To achieve a Pareto-optimal solution with text-to-image generation diffusion model, Parrot selectively uses data points from the non-dominated set using non-dominated sorting algorithm. This naturally encourages the T2I model to produce Pareto-optimal samples with respect to the multi-reward objectives." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.636, + 0.79, + 0.773 + ], + "angle": 0, + "content": "Reward-specific Preference: Inspired by the use of preference information in multi-objective optimization [30], Parrot incorporates the preference information through reward-specific identifiers. This enables Parrot to automatically determine the importance for each reward objective. Concretely, we enrich the expanded prompt \\(\\hat{c}\\) by preponding reward-specific identifier “\\)” for \\(k\\)-th reward. Based on this reward-specific prompt, \\(N\\) images are generated and are used for maximizing the corresponding \\(k\\)-th reward model during gradient update. At inference time, a concatenation of all the reward identifiers “\\), ..., \\)” is used for image generation." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.78, + 0.789, + 0.843 + ], + "angle": 0, + "content": "Non-dominated Sorting: Parrot constructs Pareto set with non-dominated points based on trade-offs among multiple rewards. These non-dominated points are superior to the remaining solutions and are not dominated by each other. Formally, the dominance relationship is defined as follows: the image \\(\\mathbf{x}_0^a\\) domi" + } + ], + [ + { + "type": "header", + "bbox": [ + 0.299, + 0.115, + 0.731, + 0.128 + ], + "angle": 0, + "content": "Pareto-optimal Multi-Reward RL Framework for T2I Generation" + }, + { + "type": "page_number", + "bbox": [ + 0.775, + 0.117, + 0.785, + 0.127 + ], + "angle": 0, + "content": "7" + }, + { + "type": "title", + "bbox": [ + 0.217, + 0.146, + 0.526, + 0.158 + ], + "angle": 0, + "content": "Algorithm Parrot: Pareto-optimal Multi-Reward RL" + }, + { + "type": "text", + "bbox": [ + 0.217, + 0.16, + 0.784, + 0.181 + ], + "angle": 0, + "content": "Input: Prompt \\( c \\), Batch size \\( N \\), Total iteration \\( E \\), the number of rewards: \\( K \\), Prompt expansion network \\( p_{\\phi} \\), T2I diffusion model: \\( p_{\\theta} \\), Total diffusion time step \\( T \\), Non-dominated set: \\( \\mathcal{P} \\)" + }, + { + "type": "text", + "bbox": [ + 0.23, + 0.181, + 0.338, + 0.191 + ], + "angle": 0, + "content": "for \\(e = 1\\) to \\(E\\) do" + }, + { + "type": "text", + "bbox": [ + 0.251, + 0.191, + 0.417, + 0.202 + ], + "angle": 0, + "content": "Sample text prompt \\( c \\sim p(c) \\)" + }, + { + "type": "text", + "bbox": [ + 0.252, + 0.202, + 0.359, + 0.21 + ], + "angle": 0, + "content": "for \\(k = 1\\) to \\(K\\) do" + }, + { + "type": "text", + "bbox": [ + 0.27, + 0.211, + 0.457, + 0.222 + ], + "angle": 0, + "content": "Expand text prompt \\(\\hat{c} \\sim p_{\\phi}(\\hat{c} | c)\\)" + }, + { + "type": "text", + "bbox": [ + 0.27, + 0.222, + 0.56, + 0.231 + ], + "angle": 0, + "content": "Prepend reward-specific tokens \\(<\\) reward \\(k>\\) to \\(\\hat{c}\\)" + }, + { + "type": "text", + "bbox": [ + 0.27, + 0.231, + 0.541, + 0.242 + ], + "angle": 0, + "content": "Sample a set of images \\(\\{\\mathbf{x}_0^1,\\dots ,\\mathbf{x}_0^N\\} \\sim p_\\theta (x_0|\\hat{c})\\)" + }, + { + "type": "text", + "bbox": [ + 0.27, + 0.242, + 0.505, + 0.252 + ], + "angle": 0, + "content": "A set of reward vector \\(\\mathcal{R} = \\{R_1,\\dots,R_N\\}\\)" + }, + { + "type": "text", + "bbox": [ + 0.27, + 0.252, + 0.425, + 0.265 + ], + "angle": 0, + "content": "\\(\\mathcal{P}\\gets \\mathrm{NDSET}(\\{\\mathbf{x}_0^1,\\dots,\\mathbf{x}_0^N\\})\\)" + }, + { + "type": "text", + "bbox": [ + 0.27, + 0.265, + 0.474, + 0.277 + ], + "angle": 0, + "content": "\\(\\nabla \\mathcal{J}_{\\phi} + = -r_k(\\mathbf{x}_0^j,\\hat{c})\\times \\nabla \\log p_{\\phi}(\\hat{c} |c)\\)" + }, + { + "type": "text", + "bbox": [ + 0.251, + 0.277, + 0.45, + 0.289 + ], + "angle": 0, + "content": "Update the gradient \\(p_{\\theta}\\) from Eq. 4" + }, + { + "type": "text", + "bbox": [ + 0.251, + 0.289, + 0.387, + 0.299 + ], + "angle": 0, + "content": "Update the gradient \\(p_{\\phi}\\)" + }, + { + "type": "text", + "bbox": [ + 0.218, + 0.298, + 0.625, + 0.309 + ], + "angle": 0, + "content": "Output: Fine-tuned diffusion model \\( p_{\\theta} \\), prompt expansion network \\( p_{\\phi} \\)" + }, + { + "type": "text", + "bbox": [ + 0.571, + 0.188, + 0.752, + 0.2 + ], + "angle": 0, + "content": "function \\(\\mathrm{NDSET}(\\{\\mathbf{x}_0^1,\\dots,\\mathbf{x}_0^N\\})\\)" + }, + { + "type": "text", + "bbox": [ + 0.59, + 0.2, + 0.63, + 0.21 + ], + "angle": 0, + "content": "\\(\\mathcal{P}\\gets \\emptyset\\)" + }, + { + "type": "text", + "bbox": [ + 0.59, + 0.211, + 0.695, + 0.221 + ], + "angle": 0, + "content": "for \\(i = 1\\) to \\(N\\) do" + }, + { + "type": "text", + "bbox": [ + 0.61, + 0.221, + 0.72, + 0.23 + ], + "angle": 0, + "content": "dominance \\(\\leftarrow\\) True" + }, + { + "type": "text", + "bbox": [ + 0.609, + 0.23, + 0.716, + 0.241 + ], + "angle": 0, + "content": "for \\(j = 1\\) to \\(N\\) do" + }, + { + "type": "text", + "bbox": [ + 0.629, + 0.241, + 0.773, + 0.252 + ], + "angle": 0, + "content": "if \\(\\mathbf{x}_0^j\\) dominates \\(\\mathbf{x}_0^i\\) then" + }, + { + "type": "text", + "bbox": [ + 0.629, + 0.252, + 0.761, + 0.261 + ], + "angle": 0, + "content": "dominance \\(\\leftarrow\\) False" + }, + { + "type": "text", + "bbox": [ + 0.609, + 0.262, + 0.763, + 0.272 + ], + "angle": 0, + "content": "if dominance is True then" + }, + { + "type": "text", + "bbox": [ + 0.629, + 0.272, + 0.695, + 0.281 + ], + "angle": 0, + "content": "Add \\(i\\) to \\(\\mathcal{P}\\)" + }, + { + "type": "text", + "bbox": [ + 0.59, + 0.282, + 0.648, + 0.291 + ], + "angle": 0, + "content": "return \\(\\mathcal{P}\\)" + }, + { + "type": "text", + "bbox": [ + 0.216, + 0.339, + 0.784, + 0.401 + ], + "angle": 0, + "content": "nates the image \\(\\mathbf{x}_0^b\\), denoted as \\(\\mathbf{x}_0^b < \\mathbf{x}_0^a\\), if and only if \\(R_{i}(\\mathbf{x}_{0}^{b})\\leq R_{i}(\\mathbf{x}_{0}^{a})\\) for all \\(i\\in 1,2,\\ldots ,m\\), and there exists \\(j\\in 1,2,\\dots,m\\) such that \\(R_{j}(\\mathbf{x}_{0}^{b}) < R_{j}(\\mathbf{x}_{0}^{a})\\). For example, given the \\(i\\)-th generated image \\(\\mathbf{x}_0^i\\) in a mini-batch, when no point in the mini-batch dominates \\(\\mathbf{x}_0^i\\), it is referred to as a non-dominated point." + }, + { + "type": "text", + "bbox": [ + 0.215, + 0.408, + 0.784, + 0.453 + ], + "angle": 0, + "content": "Policy Gradient Update: We assign a reward value of zero to the data points not included in non-dominated sets and only update the gradient of these non-dominated data points as follows:" + }, + { + "type": "equation", + "bbox": [ + 0.266, + 0.462, + 0.785, + 0.507 + ], + "angle": 0, + "content": "\\[\n\\nabla \\mathcal {J} _ {\\theta} = \\sum_ {k = 1} ^ {K} \\frac {1}{n (\\mathcal {P})} \\sum_ {i = 1, \\mathbf {x} _ {0} ^ {i} \\in \\mathcal {P}} ^ {N} \\sum_ {t = 1} ^ {T} r _ {k} \\left(\\mathbf {x} _ {0} ^ {i}, c _ {k}\\right) \\times \\nabla_ {\\theta} \\log p _ {\\theta} \\left(\\mathbf {x} _ {t - 1} ^ {i} \\mid c _ {k}, t, \\mathbf {x} _ {t} ^ {i}\\right), \\tag {4}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.215, + 0.516, + 0.784, + 0.577 + ], + "angle": 0, + "content": "where \\( i \\) indicates the index of images in mini-batches, and \\( \\mathcal{P} \\) denotes batch-wise a set of non-dominated points. \\( K \\) and \\( T \\) are the total number of reward models and total diffusion time steps, respectively. The same text prompt is used when updating the diffusion model in each batch." + }, + { + "type": "title", + "bbox": [ + 0.216, + 0.598, + 0.563, + 0.614 + ], + "angle": 0, + "content": "4.3 Original Prompt Centered Guidance" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.622, + 0.785, + 0.773 + ], + "angle": 0, + "content": "While prompt expansion enhances details and often improves generation quality, there is a concern that the added context may dilute the main content of the original input. To mitigate this during the inference, we introduce original prompt-centered guidance. When sampling conditioned on the original prompt, the diffusion model \\(\\epsilon_{\\theta}\\) typically predicts noises by combining the unconditioned score estimate and the prompt-conditioned estimate. Instead of relying solely on the expanded prompt from PEN, we propose using a linear combination of two. guidances for T2I generation: one from the user input and the other from the expanded prompt. The strength of the original prompt is controlled by guidance scales \\(w_{1}\\) and \\(w_{2}\\). The noise \\(\\bar{\\epsilon}_{\\theta}\\) is estimated, derived from Eq. 1, as follows:" + }, + { + "type": "equation", + "bbox": [ + 0.257, + 0.799, + 0.785, + 0.815 + ], + "angle": 0, + "content": "\\[\n\\bar {\\epsilon} _ {\\theta} = w _ {1} \\cdot \\epsilon_ {\\theta} (\\mathbf {x} _ {t}, t, c) + (1 - w _ {1} - w _ {2}) \\cdot \\epsilon_ {\\theta} (\\mathbf {x} _ {t}, t, \\text {n u l l}) + w _ {2} \\cdot \\epsilon_ {\\theta} (\\mathbf {x} _ {t}, t, \\hat {c}), \\tag {5}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.216, + 0.825, + 0.436, + 0.839 + ], + "angle": 0, + "content": "where null denotes a null text." + } + ], + [ + { + "type": "page_number", + "bbox": [ + 0.218, + 0.116, + 0.23, + 0.127 + ], + "angle": 0, + "content": "8" + }, + { + "type": "header", + "bbox": [ + 0.272, + 0.115, + 0.426, + 0.129 + ], + "angle": 0, + "content": "Lee, Seung Hyun et al." + }, + { + "type": "title", + "bbox": [ + 0.217, + 0.145, + 0.377, + 0.163 + ], + "angle": 0, + "content": "5 Experiments" + }, + { + "type": "title", + "bbox": [ + 0.217, + 0.175, + 0.427, + 0.192 + ], + "angle": 0, + "content": "5.1 Experiment Setting" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.207, + 0.794, + 0.313 + ], + "angle": 0, + "content": "Dataset: The PEN is first supervised fine-tuned on a large-scale text dataset named the Promptist [15], which has 360K constructed prompt pairs for original prompt and prompt expansion demonstration. The original instruction \"Rephrase\" is included per pair in Promptist. We modify the instruction into \"Input: \". This is a text input for image generation. Expand prompt for improving image quality. Output:\". Subsequently, we use the RL tuning prompts (1200K) from Promptist for RL training of the PEN and T2I model." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.32, + 0.79, + 0.427 + ], + "angle": 0, + "content": "T2I Model: Our T2I model is based on the JAX version of Stable Diffusion 1.5 [39] pre-trained with the LAION-5B [43] dataset. We conduct experiments on a machine equipped with 16 NVIDIA RTX A100 GPUs. DDIM [46] with 50 denoising steps is used, and the classifier-free guidance weight is set to 5.0 with the resolution \\(512 \\times 512\\). Instead of updating all layers, we specifically update the cross-attention layer in the Denoising U-Net. For optimization, we employ the Adam [25] optimizer with a learning rate of \\(1 \\times 10^{-5}\\)." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.433, + 0.79, + 0.526 + ], + "angle": 0, + "content": "Prompt Expansion Network: For prompt expansion, we use PaLM 2-L-IT [2], one of the PaLM2 variations, which is a multi-layer Transformer [49] decoder with casual language modeling. We optimize LoRA [20] weights for RL-based fine-tuning. The output token length of the PEN is set to 77 to match the maximum number of token length for Stable Diffusion. For original prompt-centered guidance, we set both \\( w_{1} \\) and \\( w_{2} \\) to 5 in Eq. 5." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.531, + 0.79, + 0.684 + ], + "angle": 0, + "content": "Reward Models: We incorporate four quality signals as rewards: Aesthetics, Human preference, Text-Image Alignment, Image Sentiment. For aesthetics, we use the VILA-R [24] pre-trained with the AVA [33] dataset. For human preference, we train a ViT-B/16 [10] using the Pick-a-Pic [26] dataset, which contains 500K examples for human feedback in T2I generation. The ViT-B/16 image encoder consists of 12 transformer layers, and the image resolution is \\(224 \\times 224\\) with a patch size of \\(16 \\times 16\\). For text-image alignment, we use CLIP [35] with the image encoder ViT-B/32. For image sentiment, we use the pre-trained model from [44], which outputs three labels: positive, neutral, negative. We use the positive score ranging from 0 to 1 as the sentiment reward." + }, + { + "type": "title", + "bbox": [ + 0.217, + 0.704, + 0.431, + 0.72 + ], + "angle": 0, + "content": "5.2 Qualitative Analysis" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.734, + 0.79, + 0.843 + ], + "angle": 0, + "content": "Comparison with Baselines: Fig. 3 shows the visual comparison of Parrot and multiple baselines. We include results from Stable Diffusion 1.5, DPOK [12] with a weighted sum of rewards, Promptist [15], and Parrot. DPOK exclusively fine-tunes the T2I model, while Promptist focuses on fine-tuning only the prompt expansion network. Parrot shows visually better images, particularly in aspects like color combination, cropping, perspective, and fine details in the image. This improvement can be attributed to Parrot's T2I model being fine-tuned together" + } + ], + [ + { + "type": "header", + "bbox": [ + 0.299, + 0.115, + 0.731, + 0.129 + ], + "angle": 0, + "content": "Pareto-optimal Multi-Reward RL Framework for T2I Generation" + }, + { + "type": "page_number", + "bbox": [ + 0.775, + 0.117, + 0.785, + 0.127 + ], + "angle": 0, + "content": "9" + }, + { + "type": "image_caption", + "bbox": [ + 0.269, + 0.148, + 0.315, + 0.16 + ], + "angle": 0, + "content": "SD 1.5" + }, + { + "type": "image", + "bbox": [ + 0.225, + 0.163, + 0.362, + 0.27 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.229, + 0.272, + 0.279, + 0.281 + ], + "angle": 0, + "content": "A kitchen" + }, + { + "type": "image_caption", + "bbox": [ + 0.41, + 0.148, + 0.454, + 0.16 + ], + "angle": 0, + "content": "DPOK" + }, + { + "type": "image", + "bbox": [ + 0.362, + 0.163, + 0.501, + 0.27 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.364, + 0.272, + 0.413, + 0.281 + ], + "angle": 0, + "content": "A kitchen" + }, + { + "type": "image_caption", + "bbox": [ + 0.539, + 0.148, + 0.601, + 0.16 + ], + "angle": 0, + "content": "Promptist" + }, + { + "type": "image", + "bbox": [ + 0.501, + 0.163, + 0.637, + 0.27 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.507, + 0.272, + 0.626, + 0.311 + ], + "angle": 0, + "content": "A kitchen, a fantasy digital painting, \ntrending on Artstation, \nhighly detailed" + }, + { + "type": "image_caption", + "bbox": [ + 0.687, + 0.148, + 0.726, + 0.159 + ], + "angle": 0, + "content": "Parrot" + }, + { + "type": "image", + "bbox": [ + 0.638, + 0.163, + 0.777, + 0.27 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.645, + 0.272, + 0.759, + 0.307 + ], + "angle": 0, + "content": "A kitchen with a lot of natural light and a beautiful view of the mountains" + }, + { + "type": "image", + "bbox": [ + 0.224, + 0.313, + 0.361, + 0.419 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.229, + 0.423, + 0.253, + 0.431 + ], + "angle": 0, + "content": "AF1" + }, + { + "type": "image", + "bbox": [ + 0.362, + 0.313, + 0.501, + 0.419 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.367, + 0.423, + 0.391, + 0.431 + ], + "angle": 0, + "content": "AF1" + }, + { + "type": "image", + "bbox": [ + 0.501, + 0.313, + 0.638, + 0.419 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.506, + 0.422, + 0.612, + 0.466 + ], + "angle": 0, + "content": "F1, concept art oil painting by Jama Jurabaev, extremely detailed, brush hard, artstation" + }, + { + "type": "image", + "bbox": [ + 0.639, + 0.313, + 0.777, + 0.421 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.645, + 0.422, + 0.756, + 0.466 + ], + "angle": 0, + "content": "3d rendering of a f1 car in dark and moody environment with a red and orange color scheme." + }, + { + "type": "image", + "bbox": [ + 0.224, + 0.47, + 0.36, + 0.578 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.231, + 0.581, + 0.334, + 0.608 + ], + "angle": 0, + "content": "A woman is looking out on a winter archipelago." + }, + { + "type": "image", + "bbox": [ + 0.362, + 0.47, + 0.501, + 0.578 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.365, + 0.581, + 0.467, + 0.608 + ], + "angle": 0, + "content": "A woman is looking out on a winter archipelago." + }, + { + "type": "image", + "bbox": [ + 0.501, + 0.47, + 0.637, + 0.578 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.507, + 0.581, + 0.622, + 0.625 + ], + "angle": 0, + "content": "Painting of a woman looking out on a winter archipelago, ultra realistic, concept art, intricate details, eerie" + }, + { + "type": "image", + "bbox": [ + 0.638, + 0.47, + 0.777, + 0.578 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.641, + 0.581, + 0.756, + 0.616 + ], + "angle": 0, + "content": "A woman is looking out on winter archipelago. The image is very peaceful and serene." + }, + { + "type": "image", + "bbox": [ + 0.224, + 0.628, + 0.359, + 0.736 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.229, + 0.741, + 0.331, + 0.76 + ], + "angle": 0, + "content": "A Japanese Porcelain Imari vase" + }, + { + "type": "image", + "bbox": [ + 0.362, + 0.628, + 0.5, + 0.736 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.367, + 0.741, + 0.469, + 0.76 + ], + "angle": 0, + "content": "A Japanese Porcelain Imari vase" + }, + { + "type": "image", + "bbox": [ + 0.501, + 0.628, + 0.638, + 0.736 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.503, + 0.738, + 0.638, + 0.792 + ], + "angle": 0, + "content": "A Japanese Porcelain Imari vase, intricate, elegant, highly detailed, digital painting, artstation, concept art, smooth, sharp focus, illustration." + }, + { + "type": "image", + "bbox": [ + 0.639, + 0.628, + 0.778, + 0.736 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.641, + 0.738, + 0.763, + 0.783 + ], + "angle": 0, + "content": "A photo of a Japanese Porcelain Imari vase on the ground with a beautiful floral pattern on the front" + }, + { + "type": "image_caption", + "bbox": [ + 0.216, + 0.806, + 0.787, + 0.849 + ], + "angle": 0, + "content": "Fig. 3: Comparison of Parrot and diffusion-based RL baselines. From left to right, we provide results of Stable diffusion 1.5 [39] (1st column), DPOK [12] (2nd column) with the weighted sum, Promptist [15] (3rd column), and Parrot (4th column)." + } + ], + [ + { + "type": "page_number", + "bbox": [ + 0.218, + 0.116, + 0.236, + 0.127 + ], + "angle": 0, + "content": "10" + }, + { + "type": "header", + "bbox": [ + 0.272, + 0.115, + 0.425, + 0.129 + ], + "angle": 0, + "content": "Lee, Seung Hyun et al." + }, + { + "type": "image", + "bbox": [ + 0.22, + 0.143, + 0.365, + 0.252 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.373, + 0.143, + 0.503, + 0.254 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.504, + 0.143, + 0.644, + 0.254 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.645, + 0.143, + 0.784, + 0.253 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.214, + 0.256, + 0.788, + 0.34 + ], + "angle": 0, + "content": "Fig.4: Training curve for fine-tuning on weighted sum and Parrot. For weighted sum, WS1 denotes \\(\\{0.7, 0.1, 0.1, 0.1\\}\\) and WS2 denotes \\(\\{0.25, 0.25, 0.25, 0.25\\}\\) for aesthetics, human preference, text-image alignment and image sentiment. Using weighted sum leads to decrease in human preference score and image sentiment score despite an improvement in the aesthetic score. In contrast, Parrot exhibits stable increases across all metrics." + }, + { + "type": "table", + "bbox": [ + 0.216, + 0.382, + 0.789, + 0.53 + ], + "angle": 0, + "content": "
ModelQuality Metrics
TIA (↑)Aesth. (↑)HP (↑)Sent. (↑)Average (↑)
SD 1.5 [39]0.23220.57550.19300.30100.3254
DPOK [12] (WS)0.23370.58130.19320.30130.3273 (+0.58%)
Parrot w/o PE0.23550.60340.20090.30180.3354 (+3.07%)
Parrot T2I Model Tuning Only0.25090.70730.33370.30520.3992 (+22.6%)
Promptist [15]0.14490.67830.27590.25180.3377 (+3.77 %)
Parrot with HP Only0.15430.59610.35280.25620.3398 (+4.42 %)
Parrot PEN Tuning Only0.16590.64920.26170.31310.3474 (+6.76 %)
Parrot w/o Joint Optimization0.16610.63080.25660.30840.3404 (+4.60 %)
Parrot w/o ori prompt guidance0.16230.71560.34250.31300.3833 (+17.8 %)
Parrot0.16670.73960.34110.31320.3901 (+19.8 %)
" + }, + { + "type": "table_caption", + "bbox": [ + 0.214, + 0.531, + 0.788, + 0.587 + ], + "angle": 0, + "content": "Table 1: Quantitative comparison between Parrot and alternatives on the Parti dataset [53]. Abbreviations: WS - Weighted Sum; PE - Prompt Expansion; TIA - Text-Image Alignment; Aesth. - Aesthetics; HP - Human Preference; Sent. - Image Sentiment. TIA score is measured against the original prompt without expansion." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.637, + 0.788, + 0.684 + ], + "angle": 0, + "content": "with the prompt expansion model that incorporates aesthetic keywords during training. Parrot generates results that are more closely aligned with the input prompt, as well as more visually pleasing." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.704, + 0.789, + 0.841 + ], + "angle": 0, + "content": "Weighted sum vs. Parrot: Fig. 4 shows the training curve comparison of Parrot and using a linear combination of the reward scores. Each subgraph represents a reward. WS1 and WS2 denote two different weights with multiple reward scores. WS1 places greater emphasis on the aesthetic score, while WS2 adopts balanced weights across aesthetics, human preference, text-image alignment, and image sentiment. Employing the weighted sum of multiple rewards leads to a decrease in the image sentiment score, despite notable enhancements in aesthetics and human preference. In contrast, Parrot consistently exhibits improvement across all metrics." + } + ], + [ + { + "type": "header", + "bbox": [ + 0.299, + 0.114, + 0.732, + 0.129 + ], + "angle": 0, + "content": "Pareto-optimal Multi-Reward RL Framework for T2I Generation" + }, + { + "type": "page_number", + "bbox": [ + 0.768, + 0.116, + 0.784, + 0.127 + ], + "angle": 0, + "content": "11" + }, + { + "type": "title", + "bbox": [ + 0.216, + 0.147, + 0.46, + 0.162 + ], + "angle": 0, + "content": "5.3 Quantitative Evaluation" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.179, + 0.788, + 0.284 + ], + "angle": 0, + "content": "Comparison with Baselines: Table 1 presents our results of the quality score across four quality rewards: text-image alignment score, aesthetic score, human preference score, and emotion score. The first group shows methods without prompt expansion, and the second group compares methods with expansion. The prompt expansion and T2I generation are performed on the PartiPrompts [53]. Using a set of 1632 prompts, we generate 32 images for each text input and calculate the average for each metric." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.285, + 0.788, + 0.405 + ], + "angle": 0, + "content": "\\(w / o\\) PE indicates the generation of images solely based on original prompt without expansion. Note that \\(w / o\\) PE is not the same as Parrot T2I Model Tuning Only. The former takes a model trained with PEN and removes it during inference, which results in a notable disparity between prompts used in training and testing. The latter trains with the multi-objective RL using only the T2I model, and it shows substantial improvement upon DPOK [12] (weighted sum) with balanced weights of \\(\\{0.25, 0.25, 0.25, 0.25\\}\\). This shows that the proposed RL method is indeed effective in multi-objective optimization." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.405, + 0.788, + 0.481 + ], + "angle": 0, + "content": "For Promptist [15], we generate prompt expansion from their model. Parrot with HP Only shows solely using a human preference reward leads to decline in others. Parrot PEN Tuning Only shows suboptimal aesthetics and human preference. Parrot w/o Joint Optimization shows suboptimal results than Parrot which demonstrates the necessity of jointly optimizing PEN and T2I models." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.481, + 0.788, + 0.557 + ], + "angle": 0, + "content": "Our method outperforms both compared methods in aesthetics, human preference and sentiment scores. The text-image alignment score is measured with the original prompt before expansion for fair comparison. As a result, the group without prompt expansion generally shows a higher text-image alignment score. Parrot shows better text-image alignment in each subgroup." + }, + { + "type": "table", + "bbox": [ + 0.216, + 0.581, + 0.788, + 0.657 + ], + "angle": 0, + "content": "
AreaQuestion
Aesthetics“Which image shows better aesthetics without blurry texture, unnatural focusing, and poor color combination?”
Human Preference“Which generated image do you prefer?”
Text-Image Alignment“Which image is well aligned with the text?”
Image Sentiment“Which image is closer to amusement, excitement, and contentment?”
" + }, + { + "type": "table_caption", + "bbox": [ + 0.216, + 0.658, + 0.788, + 0.685 + ], + "angle": 0, + "content": "Table 2: Questions for user study. For performing user study, we carefully design questions suitable to each quality." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.734, + 0.788, + 0.841 + ], + "angle": 0, + "content": "User Study: We conduct a user study using MTurk [1] with generated images from 100 random prompts in the PartiPrompts [53]. Five models are compared: Stable Diffusion v1.5, DPOK [12] with an equal weighted sum, Promptist [15], Parrot without prompt expansion, and Parrot. Each rater is presented with the original prompt (before expansion) and a set of five generated images, with the image order being randomized. Raters are then tasked with selecting the best image from the group, guided by questions outlined in Table 2. Each question" + } + ], + [ + { + "type": "page_number", + "bbox": [ + 0.218, + 0.116, + 0.236, + 0.127 + ], + "angle": 0, + "content": "12" + }, + { + "type": "header", + "bbox": [ + 0.272, + 0.115, + 0.425, + 0.129 + ], + "angle": 0, + "content": "Lee, Seung Hyun et al." + }, + { + "type": "image", + "bbox": [ + 0.228, + 0.149, + 0.478, + 0.293 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.224, + 0.302, + 0.484, + 0.343 + ], + "angle": 0, + "content": "Fig.5: User study results on PartiPrompts [53]. Parrot outperforms baselines across all metrics." + }, + { + "type": "image", + "bbox": [ + 0.492, + 0.145, + 0.777, + 0.285 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.489, + 0.287, + 0.779, + 0.37 + ], + "angle": 0, + "content": "Fig. 6: Ablation study. We perform an ablation study by removing one of quality signals. We observe that each quality signal affects their improvement of (a) aesthetics, (b) human preference score, (c) text-image alignment score, (d) image sentiment score." + }, + { + "type": "image", + "bbox": [ + 0.218, + 0.371, + 0.785, + 0.526 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.214, + 0.528, + 0.788, + 0.598 + ], + "angle": 0, + "content": "Fig.7: Comparing Parrot with/without Pareto Optimal. By pretending different reward-specific preferences, Parrot can change the tradeoff between rewards. Specifically, “”, “”, “”, “” indicates aesthetics, human preference, text-image alignment, and image sentiment respectively. With Pareto-optimal selection, each reward increases based on reward-specific preference." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.627, + 0.788, + 0.687 + ], + "angle": 0, + "content": "pertains to a specific quality aspect: aesthetics, human preference, text-image alignment, and image sentiment. For each prompt, 20 rounds of random sampling are conducted and sent to different raters. The user study results, illustrated in Fig. 5, show that Parrot outperforms other baselines across all dimensions." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.695, + 0.788, + 0.741 + ], + "angle": 0, + "content": "Proportion of non-dominated points: Using batch size of 256, we observe that in a batch around \\(20\\%\\) to \\(30\\%\\) are non-dominated points and the proportion of non-dominated points in single batch slightly increase as training proceeds." + }, + { + "type": "title", + "bbox": [ + 0.215, + 0.763, + 0.343, + 0.777 + ], + "angle": 0, + "content": "5.4 Ablations" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.795, + 0.787, + 0.84 + ], + "angle": 0, + "content": "Effect of Pareto-optimal Multi-reward RL: To show the efficacy of Pareto-optimal Multi-reward RL, we conduct an ablation study by removing one reward model at a time. Fig. 6 shows quantitative results using one hundred random text" + } + ], + [ + { + "type": "header", + "bbox": [ + 0.299, + 0.115, + 0.731, + 0.128 + ], + "angle": 0, + "content": "Pareto-optimal Multi-Reward RL Framework for T2I Generation" + }, + { + "type": "page_number", + "bbox": [ + 0.769, + 0.117, + 0.785, + 0.127 + ], + "angle": 0, + "content": "13" + }, + { + "type": "image", + "bbox": [ + 0.22, + 0.148, + 0.785, + 0.337 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.214, + 0.341, + 0.788, + 0.48 + ], + "angle": 0, + "content": "Fig. 8: The comparisons of the diffusion fine-tuning between Pareto-optimal multi-reward RL and single reward RL. We show results with same seed from various methods: Stable Diffusion 1.5 [39] (1st column), T2I model fine-tuned with the aesthetic model (2nd column), the human preference model (3rd column), text-image alignment (4th column), image sentiment (5th column), Parrot without Pareto-optimal selection (6th column) and Parrot (7th column). Parrot is effective to generate acceptable images without sacrificing one of quality signals. For example, T2I model fine-tuned with a single quality signal such as aesthetics, human preference and image sentiment results in text-image misalignment, while our method achieves a balanced visual outcome across multiple criteria." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.512, + 0.785, + 0.542 + ], + "angle": 0, + "content": "prompts from the Promptist [15]. We observe that our training scheme improves multiple target objectives." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.544, + 0.788, + 0.679 + ], + "angle": 0, + "content": "To verify whether Parrot achieved better trade-off for different reward scores, we generate 1000 images from common animal dataset [4], where each text prompt consists of the name of a common animal. As shown in Fig. 7, using only text-image alignment with reward-specific preference \\( \\langle reward3\\rangle \\) generates images with higher text-image alignment score, while using only aesthetic model with reward-specific preference \\( \\langle reward1\\rangle \\) yields images with higher aesthetic score. In the case of using two reward-specific preferences \\( \\langle reward1\\rangle \\), \\( \\langle reward3\\rangle \\), we observe that scores are balanced and show that better results across multiple rewards than Parrot without Pareto optimal selection." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.681, + 0.788, + 0.788 + ], + "angle": 0, + "content": "Fig 8 shows the visual comparison between Parrot, Parrot with a single reward, and Parrot without selecting the batch-wise Pareto-optimal solution. Using a single reward model tends to result in degradation of other rewards, especially text-image alignment. In the third column, results of the first row miss the text \\(a\\) tophat in input prompt, even though the Stable Diffusion result includes that attribute. On the other hand, Parrot results capture all prompts, improving other quality signals, such as aesthetics, image sentiment and human preference." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.795, + 0.788, + 0.841 + ], + "angle": 0, + "content": "Effect of Original Prompt Centered Guidance: Fig 9 shows the effect of the proposed original prompt-centered guidance. As evident from the figure, using only the expanded prompt as input often results in the main content be" + } + ], + [ + { + "type": "page_number", + "bbox": [ + 0.218, + 0.116, + 0.236, + 0.127 + ], + "angle": 0, + "content": "14" + }, + { + "type": "header", + "bbox": [ + 0.272, + 0.115, + 0.425, + 0.129 + ], + "angle": 0, + "content": "Lee, Seung Hyun et al." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.147, + 0.788, + 0.239 + ], + "angle": 0, + "content": "ing overwhelmed by the added context. For instance, given the original prompt \"A shiba inu\", the result from the expanded prompt shows a zoomed-out image and the intended main subject (shiba inu) becomes small. The proposed original prompt-centered guidance effectively addresses this issue, generating an image that faithfully captures the input prompt while incorporating visually more pleasing details." + }, + { + "type": "image", + "bbox": [ + 0.218, + 0.267, + 0.789, + 0.375 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.214, + 0.377, + 0.789, + 0.421 + ], + "angle": 0, + "content": "Fig. 9: Results of original prompt centered guidance. As we expand the prompt, the content in the generated image often fades away. This guidance is helpful for keeping the main content of the original prompt." + }, + { + "type": "title", + "bbox": [ + 0.216, + 0.486, + 0.51, + 0.502 + ], + "angle": 0, + "content": "6 Conclusion and Limitation" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.529, + 0.789, + 0.756 + ], + "angle": 0, + "content": "We propose Parrot, a novel multi-reward optimization algorithm aimed to improve text-to-image generation by effectively optimizing multiple quality rewards using RL. With batch-wise Pareto-optimal selection, Parrot adaptively balance the optimization of multiple quality rewards. By applying Parrot to jointly fine-tune both the T2I model and the prompt expansion model, we achieve the generation of higher-quality images with richer details. Additionally, our original prompt centered guidance technique ensures that the generated image maintains fidelity to the user prompt after prompt expansion during inference. Results from the user study indicate that Parrot significantly improves the quality of generated images across multiple criteria, including text-image alignment, human preference, aesthetics, and image sentiment. While Parrot has shown effectiveness in enhancing generated image quality, its efficacy is limited by the quality metrics it relies on. Therefore, advancements of the generated image quality metrics will directly enhance the capabilities of Parrot. Additionally, Parrot is adaptable to a broader range of rewards that quantify generated image quality." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.765, + 0.789, + 0.84 + ], + "angle": 0, + "content": "Societal Impact: Parrot could potentially raise ethical concerns related to the generation of immoral content. This concern stems from the user's ability to influence T2I generation, allowing for the creation of visual content that may be deemed inappropriate. The risk may be tied to the potential biases in reward models inherited from various datasets." + } + ], + [ + { + "type": "header", + "bbox": [ + 0.299, + 0.115, + 0.732, + 0.129 + ], + "angle": 0, + "content": "Pareto-optimal Multi-Reward RL Framework for T2I Generation" + }, + { + "type": "page_number", + "bbox": [ + 0.768, + 0.117, + 0.786, + 0.127 + ], + "angle": 0, + "content": "15" + }, + { + "type": "title", + "bbox": [ + 0.217, + 0.145, + 0.324, + 0.161 + ], + "angle": 0, + "content": "References" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.179, + 0.639, + 0.193 + ], + "angle": 0, + "content": "1. Amazon mechanical turk. https://www.mturk.com/ (2005)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.194, + 0.787, + 0.235 + ], + "angle": 0, + "content": "2. Anil, R., Dai, A.M., First, O., Johnson, M., Lepikhin, D., Passos, A., Shakeri, S., Taropa, E., Bailey, P., Chen, Z., et al.: Palm 2 technical report. arXiv preprint arXiv:2305.10403 (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.236, + 0.787, + 0.292 + ], + "angle": 0, + "content": "3. Bai, Y., Jones, A., Ndousse, K., Askell, A., Chen, A., DasSarma, N., Drain, D., Fort, S., Ganguli, D., Henighan, T., et al.: Training a helpful and harmless assistant with reinforcement learning from human feedback. arXiv preprint arXiv:2204.05862 (2022)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.293, + 0.787, + 0.32 + ], + "angle": 0, + "content": "4. Black, K., Janner, M., Du, Y., Kostrikov, I., Levine, S.: Training diffusion models with reinforcement learning. arXiv preprint arXiv:2305.13301 (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.321, + 0.787, + 0.376 + ], + "angle": 0, + "content": "5. Chang, H., Zhang, H., Barber, J., Maschinot, A., Lezama, J., Jiang, L., Yang, M.H., Murphy, K., Freeman, W.T., Rubinstein, M., et al.: Muse: Text-to-image generation via masked generative transformers. arXiv preprint arXiv:2301.00704 (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.377, + 0.787, + 0.404 + ], + "angle": 0, + "content": "6. Clark, K., Vicol, P., Swersky, K., Fleet, D.J.: Directly fine-tuning diffusion models on differentiable rewards. ICLR (2024)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.405, + 0.787, + 0.447 + ], + "angle": 0, + "content": "7. Dai, X., Hou, J., Ma, C.Y., Tsai, S., Wang, J., Wang, R., Zhang, P., Vandenhende, S., Wang, X., Dubey, A., et al.: Emu: Enhancing image generation models using photogenic needles in a haystack. arXiv preprint arXiv:2309.15807 (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.448, + 0.787, + 0.489 + ], + "angle": 0, + "content": "8. Deng, F., Wang, Q., Wei, W., Grundmann, M., Hou, T.: Prdp: Proximal reward difference prediction for large-scale reward finetuning of diffusion models. CVPR (2024)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.489, + 0.787, + 0.531 + ], + "angle": 0, + "content": "9. Dong, H., Xiong, W., Goyal, D., Pan, R., Dao, S., Zhang, J., Shum, K., Zhang, T.: Raft: Reward ranked finetuning for generative foundation model alignment. arXiv preprint arXiv:2304.06767 (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.22, + 0.531, + 0.787, + 0.586 + ], + "angle": 0, + "content": "10. Dosovitskiy, A., Beyer, L., Kolesnikov, A., Weissenborn, D., Zhai, X., Unterthiner, T., Dehghani, M., Minderer, M., Heigold, G., Gelly, S., et al.: An image is worth 16x16 words: Transformers for image recognition at scale. arXiv preprint arXiv:2010.11929 (2020)" + }, + { + "type": "ref_text", + "bbox": [ + 0.219, + 0.587, + 0.787, + 0.614 + ], + "angle": 0, + "content": "11. Fan, Y., Lee, K.: Optimizing ddpm sampling with shortcut fine-tuning. ICML (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.219, + 0.615, + 0.787, + 0.657 + ], + "angle": 0, + "content": "12. Fan, Y., Watkins, O., Du, Y., Liu, H., Ryu, M., Boutilier, C., Abbeel, P., Ghavamzadeh, M., Lee, K., Lee, K.: Dpok: Reinforcement learning for fine-tuning text-to-image diffusion models. NeurIPS (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.219, + 0.658, + 0.787, + 0.685 + ], + "angle": 0, + "content": "13. Fang, Y., Zhu, H., Zeng, Y., Ma, K., Wang, Z.: Perceptual quality assessment of smartphone photography. In: CVPR (2020)" + }, + { + "type": "ref_text", + "bbox": [ + 0.219, + 0.686, + 0.787, + 0.713 + ], + "angle": 0, + "content": "14. Han, L., Li, Y., Zhang, H., Milanfar, P., Metaxas, D., Yang, F.: Svdiff: Compact parameter space for diffusion fine-tuning. CVPR (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.219, + 0.714, + 0.787, + 0.742 + ], + "angle": 0, + "content": "15. Hao, Y., Chi, Z., Dong, L., Wei, F.: Optimizing prompts for text-to-image generation. arXiv preprint arXiv:2212.09611 (2022)" + }, + { + "type": "ref_text", + "bbox": [ + 0.219, + 0.742, + 0.787, + 0.784 + ], + "angle": 0, + "content": "16. He, H., Wang, T., Yang, H., Fu, J., Yuan, N.J., Yin, J., Chao, H., Zhang, Q.: Learning profitable nft image diffusions via multiple visual-policy guided reinforcement learning. arXiv preprint arXiv:2306.11731 (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.219, + 0.785, + 0.787, + 0.812 + ], + "angle": 0, + "content": "17. Ho, J., Jain, A., Abbeel, P.: Denoising diffusion probabilistic models. NeurIPS (2020)" + }, + { + "type": "ref_text", + "bbox": [ + 0.219, + 0.813, + 0.787, + 0.84 + ], + "angle": 0, + "content": "18. Ho, J., Salimans, T.: Classifier-free diffusion guidance. arXiv preprint arXiv:2207.12598 (2022)" + }, + { + "type": "list", + "bbox": [ + 0.219, + 0.179, + 0.787, + 0.84 + ], + "angle": 0, + "content": null + } + ], + [ + { + "type": "page_number", + "bbox": [ + 0.218, + 0.116, + 0.236, + 0.127 + ], + "angle": 0, + "content": "16" + }, + { + "type": "header", + "bbox": [ + 0.272, + 0.115, + 0.425, + 0.128 + ], + "angle": 0, + "content": "Lee, Seung Hyun et al." + }, + { + "type": "ref_text", + "bbox": [ + 0.218, + 0.147, + 0.787, + 0.188 + ], + "angle": 0, + "content": "19. Hosu, V., Lin, H., Sziranyi, T., Saupe, D.: Koniq-10k: An ecologically valid database for deep learning of blind image quality assessment. TIP 29, 4041-4056 (2020)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.189, + 0.786, + 0.23 + ], + "angle": 0, + "content": "20. Hu, E.J., Shen, Y., Wallis, P., Allen-Zhu, Z., Li, Y., Wang, S., Wang, L., Chen, W.: Lora: Low-rank adaptation of large language models. arXiv preprint arXiv:2106.09685 (2021)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.23, + 0.787, + 0.271 + ], + "angle": 0, + "content": "21. Jeong, Y., Ryoo, W., Lee, S., Seo, D., Byeon, W., Kim, S., Kim, J.: The power of sound (tpos): Audio reactive video generation with stable diffusion. In: ICCV (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.271, + 0.786, + 0.298 + ], + "angle": 0, + "content": "22. Kawar, B., Zada, S., Lang, O., Tov, O., Chang, H., Dekel, T., Mosseri, I., Irani, M.: Imagic: Text-based real image editing with diffusion models. In: CVPR (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.298, + 0.786, + 0.325 + ], + "angle": 0, + "content": "23. Ke, J., Wang, Q., Wang, Y., Milanfar, P., Yang, F.: Musiq: Multi-scale image quality transformer. In: ICCV (2021)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.325, + 0.786, + 0.352 + ], + "angle": 0, + "content": "24. Ke, J., Ye, K., Yu, J., Wu, Y., Milanfar, P., Yang, F.: Vila: Learning image aesthetics from user comments with vision-language pretraining. In: CVPR (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.352, + 0.786, + 0.365 + ], + "angle": 0, + "content": "25. Kingma, D.P., Ba, J.: Adam: A method for stochastic optimization. ICLR (2015)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.365, + 0.786, + 0.405 + ], + "angle": 0, + "content": "26. Kirstain, Y., Polyak, A., Singer, U., Matiana, S., Penna, J., Levy, O.: Pick-a-pic: An open dataset of user preferences for text-to-image generation. arXiv preprint arXiv:2305.01569 (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.405, + 0.786, + 0.447 + ], + "angle": 0, + "content": "27. Lee, K., Liu, H., Ryu, M., Watkins, O., Du, Y., Boutilier, C., Abbeel, P., Ghavamzadeh, M., Gu, S.S.: Aligning text-to-image models using human feedback. arXiv preprint arXiv:2302.12192 (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.447, + 0.786, + 0.488 + ], + "angle": 0, + "content": "28. Lee, S.H., Kim, S., Yoo, I., Yang, F., Cho, D., Kim, Y., Chang, H., Kim, J., Kim, S.: Soundini: Sound-guided diffusion for natural video editing. arXiv preprint arXiv:2304.06818 (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.488, + 0.786, + 0.515 + ], + "angle": 0, + "content": "29. Li, Y., Liu, H., Wu, Q., Mu, F., Yang, J., Gao, J., Li, C., Lee, Y.J.: Gligen: Open-set grounded text-to-image generation. In: CVPR (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.515, + 0.786, + 0.542 + ], + "angle": 0, + "content": "30. Lin, X., Yang, Z., Zhang, X., Zhang, Q.: Pareto set learning for expensive multi-objective optimization. NeurIPS (2022)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.542, + 0.786, + 0.569 + ], + "angle": 0, + "content": "31. Mannor, S., Shimkin, N.: The steering approach for multi-criteria reinforcement learning. NeurIPS (2001)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.569, + 0.786, + 0.595 + ], + "angle": 0, + "content": "32. Miettinen, K.: Nonlinear multiobjective optimization, vol. 12. Springer Science & Business Media (1999)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.595, + 0.786, + 0.622 + ], + "angle": 0, + "content": "33. Murray, N., Marchesotti, L., Perronnin, F.: Ava: A large-scale database for aesthetic visual analysis. In: CVPR (2012)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.622, + 0.786, + 0.663 + ], + "angle": 0, + "content": "34. Ouyang, L., Wu, J., Jiang, X., Almeida, D., Wainwright, C., Mishkin, P., Zhang, C., Agarwal, S., Slama, K., Ray, A., et al.: Training language models to follow instructions with human feedback. NeurIPS (2022)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.663, + 0.786, + 0.704 + ], + "angle": 0, + "content": "35. Radford, A., Kim, J.W., Hallacy, C., Ramesh, A., Goh, G., Agarwal, S., Sastry, G., Askell, A., Mishkin, P., Clark, J., et al.: Learning transferable visual models from natural language supervision. In: ICML (2021)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.704, + 0.786, + 0.745 + ], + "angle": 0, + "content": "36. Rame, A., Couairon, G., Shukor, M., Dancette, C., Gaya, J.B., Soulier, L., Cord, M.: Rewarded soups: towards pareto-optimal alignment by interpolating weights fine-tuned on diverse rewards. NeurIPS (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.745, + 0.786, + 0.786 + ], + "angle": 0, + "content": "37. Ramesh, A., Dhariwal, P., Nichol, A., Chu, C., Chen, M.: Hierarchical text-conditional image generation with clip latents. arXiv preprint arXiv:2204.06125 (2022)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.786, + 0.786, + 0.813 + ], + "angle": 0, + "content": "38. Richardson, E., Goldberg, K., Alaluf, Y., Cohen-Or, D.: Conceptlab: Creative generation using diffusion prior constraints. arXiv preprint arXiv:2308.02669 (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.813, + 0.786, + 0.84 + ], + "angle": 0, + "content": "39. Rombach, R., Blattmann, A., Lorenz, D., Esser, P., Ommer, B.: High-resolution image synthesis with latent diffusion models. In: CVPR (2022)" + }, + { + "type": "list", + "bbox": [ + 0.217, + 0.147, + 0.787, + 0.84 + ], + "angle": 0, + "content": null + } + ], + [ + { + "type": "header", + "bbox": [ + 0.299, + 0.115, + 0.732, + 0.129 + ], + "angle": 0, + "content": "Pareto-optimal Multi-Reward RL Framework for T2I Generation" + }, + { + "type": "page_number", + "bbox": [ + 0.768, + 0.116, + 0.786, + 0.127 + ], + "angle": 0, + "content": "17" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.147, + 0.785, + 0.175 + ], + "angle": 0, + "content": "40. Ronneberger, O., Fischer, P., Brox, T.: U-net: Convolutional networks for biomedical image segmentation. In: MICCAI (2015)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.176, + 0.785, + 0.203 + ], + "angle": 0, + "content": "41. Sahara, C., Chan, W., Chang, H., Lee, C., Ho, J., Salimans, T., Fleet, D., Norouzi, M.: Palette: Image-to-image diffusion models. In: SIGGRAPH (2022)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.204, + 0.785, + 0.245 + ], + "angle": 0, + "content": "42. Sahara, C., Chan, W., Saxena, S., Li, L., Whang, J., Denton, E.L., Ghasemipour, K., Gontijo Lopes, R., Karagol Ayan, B., Salimans, T., et al.: Photorealistic text-to-image diffusion models with deep language understanding. NeurIPS (2022)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.246, + 0.785, + 0.286 + ], + "angle": 0, + "content": "43. Schuhmann, C., Beaumont, R., Vencu, R., Gordon, C., Wightman, R., Cherti, M., Coombes, T., Katta, A., Mullis, C., Wortsman, M., et al.: Laion-5b: An open large-scale dataset for training next generation image-text models. NeurIPS (2022)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.287, + 0.785, + 0.314 + ], + "angle": 0, + "content": "44. Serra, A., Carrara, F., Tesconi, M., Falchi, F.: The emotions of the crowd: Learning image sentiment from tweets via cross-modal distillation. ECAI (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.315, + 0.785, + 0.342 + ], + "angle": 0, + "content": "45. Sohl-Dickstein, J., Weiss, E., Maheswaranathan, N., Ganguli, S.: Deep unsupervised learning using nonequilibrium thermodynamics. In: ICML (2015)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.343, + 0.785, + 0.369 + ], + "angle": 0, + "content": "46. Song, J., Meng, C., Ermon, S.: Denoising diffusion implicit models. arXiv preprint arXiv:2010.02502 (2020)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.37, + 0.785, + 0.411 + ], + "angle": 0, + "content": "47. Tesauro, G., Das, R., Chan, H., Kephart, J., Levine, D., Rawson, F., Lefury, C.: Managing power consumption and performance of computing systems using reinforcement learning. NeurIPS (2007)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.412, + 0.785, + 0.438 + ], + "angle": 0, + "content": "48. Tu, Z., Talebi, H., Zhang, H., Yang, F., Milanfar, P., Bovik, A., Li, Y.: Maxvit: Multi-axis vision transformer. In: ECCV (2022)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.439, + 0.785, + 0.466 + ], + "angle": 0, + "content": "49. Vaswani, A., Shazeer, N., Parmar, N., Uszkoreit, J., Jones, L., Gomez, A.N., Kaiser, L., Polosukhin, I.: Attention is all you need. NeurIPS (2017)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.467, + 0.785, + 0.494 + ], + "angle": 0, + "content": "50. Wu, X., Sun, K., Zhu, F., Zhao, R., Li, H.: Human preference score: Better aligning text-to-image models with human preference. In: ICCV (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.495, + 0.785, + 0.535 + ], + "angle": 0, + "content": "51. Xu, J., Liu, X., Wu, Y., Tong, Y., Li, Q., Ding, M., Tang, J., Dong, Y.: Imagereward: Learning and evaluating human preferences for text-to-image generation. arXiv preprint arXiv:2304.05977 (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.536, + 0.785, + 0.577 + ], + "angle": 0, + "content": "52. Ying, Z., Niu, H., Gupta, P., Mahajan, D., Ghadiyaram, D., Bovik, A.: From patches to pictures (paq-2-piq): Mapping the perceptual space of picture quality. In: CVPR (2020)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.578, + 0.785, + 0.619 + ], + "angle": 0, + "content": "53. Yu, J., Xu, Y., Koh, J.Y., Luong, T., Baid, G., Wang, Z., Vasudevan, V., Ku, A., Yang, Y., Ayan, B.K., et al.: Scaling autoregressive models for content-rich text-to-image generation. arXiv preprint arXiv:2206.10789 (2022)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.62, + 0.785, + 0.66 + ], + "angle": 0, + "content": "54. Yu, L., Cheng, Y., Sohn, K., Lezama, J., Zhang, H., Chang, H., Hauptmann, A.G., Yang, M.H., Hao, Y., Essa, I., et al.: Magvit: Masked generative video transformer. In: CVPR (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.661, + 0.785, + 0.688 + ], + "angle": 0, + "content": "55. Zhou, Y., Liu, B., Zhu, Y., Yang, X., Chen, C., Xu, J.: Shifted diffusion for text-to-image generation. In: CVPR (2023)" + }, + { + "type": "list", + "bbox": [ + 0.217, + 0.147, + 0.785, + 0.688 + ], + "angle": 0, + "content": null + } + ] +] \ No newline at end of file diff --git a/2024/Parrot_ Pareto-optimal Multi-Reward Reinforcement Learning Framework for Text-to-Image Generation/938566d1-b778-4bb0-b908-cbc4a31c6445_origin.pdf b/2024/Parrot_ Pareto-optimal Multi-Reward Reinforcement Learning Framework for Text-to-Image Generation/938566d1-b778-4bb0-b908-cbc4a31c6445_origin.pdf new file mode 100644 index 0000000000000000000000000000000000000000..53a604b3a26164c13a06fb74043e3a1ea42074f2 --- /dev/null +++ b/2024/Parrot_ Pareto-optimal Multi-Reward Reinforcement Learning Framework for Text-to-Image Generation/938566d1-b778-4bb0-b908-cbc4a31c6445_origin.pdf @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:abb788e271df4afca9540c9889ec39bd9a3425819c5be895c086fc5fe12a7b48 +size 6252397 diff --git a/2024/Parrot_ Pareto-optimal Multi-Reward Reinforcement Learning Framework for Text-to-Image Generation/full.md b/2024/Parrot_ Pareto-optimal Multi-Reward Reinforcement Learning Framework for Text-to-Image Generation/full.md new file mode 100644 index 0000000000000000000000000000000000000000..ac3bd9b9f2e188428a3a58b81abbb3a4bcb0ab18 --- /dev/null +++ b/2024/Parrot_ Pareto-optimal Multi-Reward Reinforcement Learning Framework for Text-to-Image Generation/full.md @@ -0,0 +1,353 @@ +# Parrot: Pareto-optimal Multi-Reward Reinforcement Learning Framework for Text-to-Image Generation + +Seung Hyun Lee $^{1,6*}$ , Yinxiao Li $^{1}$ , Junjie Ke $^{1}$ , Innfarn Yoo $^{2}$ , Han Zhang $^{3}$ , Jiahui Yu $^{4\dagger}$ , Qifei Wang $^{2}$ , Fei Deng $^{2,5*}$ , Glenn Entis $^{1}$ , Junfeng He $^{1}$ , Gang Li $^{1}$ , Sangpil Kim $^{6}$ , Irfan Essa $^{1}$ , Feng Yang $^{1}$ + +Google Research1, Google2, Google DeepMind3, OpenAI4, Rutgers University5, Korea University6 + +![](images/4c4a4ecb41a55053004f374235ae1011858d93e51976d24ef4603162331103c4.jpg) +Fig. 1: Parrot visual examples. Parrot consistently improves the quality of generated images across multiple criteria: aesthetics, human preference, text-image alignment, and image sentiment. Each column shows generated images using the same seed. + +Abstract. Recent works have demonstrated that using reinforcement learning (RL) with multiple quality rewards can improve the quality of generated images in text-to-image (T2I) generation. However, manually adjusting reward weights poses challenges and may cause over-optimization in certain metrics. To solve this, we propose Parrot, which addresses the issue through multi-objective optimization and introduces an effective multi-reward optimization strategy to approximate Pareto optimal. Utilizing batch-wise Pareto optimal selection, Parrot automatically identifies the optimal trade-off among different rewards. We use the novel multi-reward optimization algorithm to jointly optimize the T2I model and a prompt expansion network, resulting in significant improvement of image quality and also allow to control the trade-off of + +different rewards using a reward related prompt during inference. Furthermore, we introduce original prompt-centered guidance at inference time, ensuring fidelity to user input after prompt expansion. Extensive experiments and a user study validate the superiority of Parrot over several baselines across various quality criteria, including aesthetics, human preference, text-image alignment, and image sentiment. + +# 1 Introduction + +Despite significant advancements in text-to-image (T2I) generation [17, 29, 38, 45, 55], recent work like Imagen [42] and Stable Diffusion [39], still struggle to produce high quality images. Images in the first row in Fig. 1 illustrates such quality issues in Stable Diffusion [39], including poor composition (e.g. bad cropping), misalignment with input prompts (e.g. missing objects), or overall lack of aesthetic appeal. Assessing the quality of generated images can involve various metrics such as aesthetics [23, 24, 33], human preference [26], text-image alignment [35], and emotional appeal [44]. Enhancing T2I generation across multiple quality metrics remains a challenging task. + +Recent works [4,12,15,27] have demonstrated that incorporating quality signals as reward functions in fine-tuning T2I with reinforcement learning (RL) can improve image quality. For instance, Promptist [15] fine-tunes prompt expansion model using RL with the sum of aesthetics and text-image alignment scores. However, the simple weighted sum approach may not effectively handle trade-offs among multiple quality metrics. As the number of rewards increases, manually adjusting reward weights becomes impractical. Moreover, optimizing one quality metric may inadvertently compromise others, as the model might prioritize aesthetics over relevance to the input prompt. Additionally, the trade-off for different reward is not controllable after training. + +To address these challenges, we propose Parrot, a Pareto-optimal multi-reward reinforcement learning algorithm to improve text-to-image generation. Unlike previous approaches that treat T2I reward optimization as a single objective optimization problem, Parrot tackles this challenge through multi-objective optimization and introduces an effective multi-reward optimization strategy to achieve Pareto optimal approximation. Intuitively, each generated sample in a batch embodies a distinctive trade-off among various quality rewards, with some samples exhibiting superior trade-offs compared to others. Instead of updating gradients using all batch samples, Parrot uses non-dominated points [32] which have better trade-offs. Consequently, Parrot automatically learns from the optimal trade-off among different rewards. Moreover, Parrot learns reward-specific preference prompts, which can be utilized individually or in combination to control the trade-off among different rewards during inference time. Unlike prior work, which either solely fine-tunes the T2I model [12] or only tunes the prompt expansion network while freezing the T2I model [15], we employ the Parrot multi-reward optimization algorithm to jointly optimize both the T2I model and the prompt expansion network. This collaborative optimization unlocks the + +full potential of Parrot by encouraging both more details from added context from the prompt expansion model, and the overall quality improvement on the T2I generation. During inference, we further introduce original prompt-centered guidance to ensure the output image is relevant to input prompts after prompt expansion. + +In summary, our contributions can be listed as follows: + +- We propose Parrot, a novel multi-reward optimization algorithm for T2I RL fine-tuning. Leveraging batch-wise Pareto-optimal selection, it effectively optimizes multiple T2I rewards, enabling collaborative improvement in aesthetics, human preference, image sentiment, and text-image alignment and also allowing to control the trade-off of different rewards using a reward related prompt during inference. +- We show the advantage of jointly optimizing both the prompt expansion network and the T2I model, which has never been explored before. +- We introduce original prompt-centered guidance during inference time after prompt expansion, ensuring better alignment with the original prompt while enriching image details. +- Extensive results and a user study validate that Parrot outperforms several baseline methods across various quality criteria. + +# 2 Related Work + +T2I Generation: The goal of T2I generation is to create an image given an input text prompt. Several T2I generative models have been proposed and have demonstrated promising results [5, 7, 14, 21, 22, 28, 37, 41, 42, 54]. Stable Diffusion [39] shows impressive generation performance in T2I generation, leveraging latent text representations from LLMs. Despite substantial progress, the images generated by those models still exhibit quality issues, such as bad cropping or misalignment with the input texts. + +RL for T2I Fine-tuning: Starting by Fan et al. [11] to explore RL fine-tuning for T2I models, following works [4, 8, 9, 12, 16] have explored RL fine-tuning technique for T2I diffusion model, showcasing superior performance for human preference learning. DPOK [12] improves quality through RL using ImageReward [51] score as a reward with a few prompts. In addition to fine-tuning the T2I model directly using RL, Promptist [15] fine-tunes the prompt expansion model by using a simple sum of aesthetic and text-image alignment scores as reward. DRaFT [6] proposed not only differentiable rewards for efficient fine-tuning but also effectiveness of using linear summation of multi-rewards. These methods treat T2I RL as a single-objective optimization problem, while Parrot employs multi-objective optimization. Additionally, prior approaches either fine-tune the T2I model or the prompt expansion model while freezing the other. In contrast, Parrot proposes joint optimization of the prompt expansion model and the T2I model using multi-reward RL to foster better collaboration. + +Multi-objective Optimization: Multi-objective optimization problem involves optimizing multiple objective functions simultaneously. The scalarization technique [31, 47] formulates multi-objective problem into single-objective problems with the weighted sum of each score, which requires pre-defined weights for each objective. Rame et al. [36] proposed weighted averaging method to find Pareto frontier, leveraging multiple fine-tuned models. Lin et al. [30] proposes to learn a set model to map trade-off preference vectors to their corresponding Pareto solutions. Inspired by this, Parrot introduces a language-based preference vector constructed from task identifiers for each reward, then encoded by the text-encoder. In the context of multi-reward RL for T2I diffusion models, Promptist [15] uses a simple weighted sum of two reward scores. This approach requires manual tuning of the weights, which makes it time-consuming and hard to scale when the number of rewards increases. + +Generated Image Quality: The quality assessment of images generated by T2I models involves multiple dimensions, and various metrics have been proposed. In this paper, we consider using four types of quality metrics as rewards: aesthetics, human preference, text-image alignment, and image sentiment. Aesthetics captures the overall visual appealingness of the image, and it is learned using human ratings for aesthetics in real images [13, 19, 23, 24, 33, 48, 52]. Human preferences, rooted the concept of learning from human feedback [3, 34, 50], involves gathering preferences at scale by having raters to compare generated images [26, 51]. Text-image alignment measures the extent to which the generated image aligns with the input prompt, CLIP [35] score is often employed, measuring the cosine distance of between contrastive image embedding and text embedding. Image sentiment is important for ensuring the generated image evokes positive emotions in the viewer. Serra et al. [44] predict average polarity of sentiments an image elicits and learn estimates for positive, neutral, and negative scores. In Parrot, we use its positive score as a reward for positive emotions. + +# 3 Preliminary + +Diffusion Probabilistic Models: Diffusion probabilistic models [17] generate the image by gradually denoising a noisy image. Specifically, given a real image $\mathbf{x}_0$ from the data distribution $\mathbf{x}_0 \sim q(\mathbf{x}_0)$ , the forward process $q(\mathbf{x}_t | \mathbf{x}_0, c)$ of diffusion probabilistic models produce a noisy image $\mathbf{x}_t$ , which induces a distribution $p(\mathbf{x}_0, c)$ conditioned on text prompt $c$ . In classifier-free guidance [18], denoising model predicts noise $\bar{\epsilon}_{\theta}$ with a linear combination of the unconditional score estimates $\epsilon_{\theta}(\mathbf{x}_t, t)$ and the conditional score estimates $\epsilon_{\theta}(\mathbf{x}_t, t, c)$ as follows: + +$$ +\bar {\epsilon} _ {\theta} = w \cdot \epsilon_ {\theta} (\mathbf {x} _ {t}, t, c) + (1 - w) \cdot \epsilon_ {\theta} (\mathbf {x} _ {t}, t, \mathrm {n u l l}), \tag {1} +$$ + +where $t$ denotes diffusion time step, the null indicates a null text and $w$ represents the guidance scale of classifier-free guidance where $w \geq 1$ . Note that $\epsilon_{\theta}$ is typically parameterized by the UNet [40]. + +RL-based T2I Diffusion Model Fine-tuning: Given a reward signal from generated images, the goal of RL-tuning for T2I diffusion models is to optimize the policy defined as one denoising step of T2I diffusion models. In particular, Black et al. [4] apply policy gradient algorithm, which regards the denoising process of diffusion models as a Markov decision process (MDP) by performing multiple denoising steps iteratively. Subsequently, a black box reward model $r(\cdot ,\cdot)$ predicts a single scalar value from sampled image $\mathbf{x}_0$ . Given text condition $c\sim p(c)$ and image $\mathbf{x}_0$ , objective function $\mathcal{I}$ can be defined to maximize the expected reward as follows: + +$$ +\mathcal {J} _ {\theta} = \mathbb {E} _ {p (c)} \mathbb {E} _ {p _ {\theta} (\mathbf {x} _ {0} | c)} [ r (\mathbf {x} _ {0}, c) ], \tag {2} +$$ + +where the pre-trained diffusion model $p_{\theta}$ produces a sample distribution $p_{\theta}(\mathbf{x}_0|c)$ using text condition $c$ . Modifying this equation, Fan et al. [12] demonstrate that the gradient of objective function $\nabla \mathcal{I}_{\theta}$ can be calculated through gradient ascent algorithm without using the gradient of reward model as follows: + +$$ +\nabla \mathcal {J} _ {\theta} = \mathbb {E} \left[ r \left(\mathbf {x} _ {0}, c\right) \sum_ {t = 1} ^ {T} \nabla_ {\theta} \log p _ {\theta} \left(\mathbf {x} _ {t - 1} \mid c, t, \mathbf {x} _ {t}\right) \right], \tag {3} +$$ + +where $T$ denotes the total time step of the diffusion sampling. With parameters $\theta$ , the expectation value can be taken over the trajectories of diffusion sampling. + +# 4 Method + +# 4.1 Parrot Overview + +Fig. 2 shows the overview of Parrot, which consists of the prompt expansion network (PEN) $p_{\phi}$ and the T2I diffusion model $p_{\theta}$ . The PEN is first initialized from a supervised fine-tuning checkpoint on demonstrations of prompt expansion pairs, and the T2I model is initialized from pretrained diffusion model. Given the original prompt $c$ , the PEN generates an expanded prompt $\hat{c}$ , and the T2I model generates images based on this expanded prompt. During the multi-reward RL fine-tuning, a batch of $N$ images is sampled, and multiple quality rewards are calculated for each image, encompassing aspects like text-image alignment, aesthetics, human preference, and image sentiment. Based on these reward scores, Parrot identifies the batch-wise Pareto-optimal set using a non-dominated sorting algorithm. This optimal set of images is then used for joint optimization of the PEN and T2I model parameters through RL policy gradient update. During inference, Parrot leverages both the original prompt and its expansion, striking a balance between maintaining faithfulness to the original prompt and incorporating additional details for higher quality. + +# 4.2 Batch-wise Pareto-optimal Selection + +Li et al. [30] has demonstrated that using batchwise Pareto-set learning and selecting good samples in a batch can approximate Pareto-optimality across multiple objectives. Backed up by this theory, we propose to select non-dominated + +![](images/108ff289898dda4a663cb1763b479d7e4373fcb644164b567d72c8c0a3ddbd91.jpg) +Fig. 2: Overview of Parrot. During the training, $N$ images are sampled from the T2I model using the expanded prompt from the prompt expansion network. Multiple quality rewards are calculated for each image, and the Pareto-optimal set is identified using the non-dominated sorting algorithm. These optimal images are then used to perform policy gradient update of the parameters of T2I model and prompt expansion network jointly. During the inference, both the original prompt and the expanded prompt are provided to the T2I model, enabling better faithfulness while adding detail. + +points in a batch for policy gradient update in RL to achieve Pareto-optimality. Algorithm 1 outlines the procedure of Parrot. Rather than updating the gradients using all images, Parrot focuses on high-quality samples, considering multiple quality rewards in each mini-batch. In the multi-reward RL, each sample generated by the T2I model presents distinct trade-offs for different rewards. Among these samples, a subset with varied optimal trade-offs across multiple objectives, also known as the Pareto set, exists. For a Pareto-optimal sample, none of its objective values can be further improved without damaging others. In other words, the Pareto-optimal set is not dominated by any data points, also known as the non-dominated set. To achieve a Pareto-optimal solution with text-to-image generation diffusion model, Parrot selectively uses data points from the non-dominated set using non-dominated sorting algorithm. This naturally encourages the T2I model to produce Pareto-optimal samples with respect to the multi-reward objectives. + +Reward-specific Preference: Inspired by the use of preference information in multi-objective optimization [30], Parrot incorporates the preference information through reward-specific identifiers. This enables Parrot to automatically determine the importance for each reward objective. Concretely, we enrich the expanded prompt $\hat{c}$ by preponding reward-specific identifier “$ ” for $k$ -th reward. Based on this reward-specific prompt, $N$ images are generated and are used for maximizing the corresponding $k$ -th reward model during gradient update. At inference time, a concatenation of all the reward identifiers “$ , ..., $ ” is used for image generation. + +Non-dominated Sorting: Parrot constructs Pareto set with non-dominated points based on trade-offs among multiple rewards. These non-dominated points are superior to the remaining solutions and are not dominated by each other. Formally, the dominance relationship is defined as follows: the image $\mathbf{x}_0^a$ domi + +# Algorithm Parrot: Pareto-optimal Multi-Reward RL + +Input: Prompt $c$ , Batch size $N$ , Total iteration $E$ , the number of rewards: $K$ , Prompt expansion network $p_{\phi}$ , T2I diffusion model: $p_{\theta}$ , Total diffusion time step $T$ , Non-dominated set: $\mathcal{P}$ + +for $e = 1$ to $E$ do + +Sample text prompt $c \sim p(c)$ + +for $k = 1$ to $K$ do + +Expand text prompt $\hat{c} \sim p_{\phi}(\hat{c} | c)$ + +Prepend reward-specific tokens $<$ reward $k>$ to $\hat{c}$ + +Sample a set of images $\{\mathbf{x}_0^1,\dots ,\mathbf{x}_0^N\} \sim p_\theta (x_0|\hat{c})$ + +A set of reward vector $\mathcal{R} = \{R_1,\dots,R_N\}$ + +$\mathcal{P}\gets \mathrm{NDSET}(\{\mathbf{x}_0^1,\dots,\mathbf{x}_0^N\})$ + +$\nabla \mathcal{J}_{\phi} + = -r_k(\mathbf{x}_0^j,\hat{c})\times \nabla \log p_{\phi}(\hat{c} |c)$ + +Update the gradient $p_{\theta}$ from Eq. 4 + +Update the gradient $p_{\phi}$ + +Output: Fine-tuned diffusion model $p_{\theta}$ , prompt expansion network $p_{\phi}$ + +function $\mathrm{NDSET}(\{\mathbf{x}_0^1,\dots,\mathbf{x}_0^N\})$ + +$\mathcal{P}\gets \emptyset$ + +for $i = 1$ to $N$ do + +dominance $\leftarrow$ True + +for $j = 1$ to $N$ do + +if $\mathbf{x}_0^j$ dominates $\mathbf{x}_0^i$ then + +dominance $\leftarrow$ False + +if dominance is True then + +Add $i$ to $\mathcal{P}$ + +return $\mathcal{P}$ + +nates the image $\mathbf{x}_0^b$ , denoted as $\mathbf{x}_0^b < \mathbf{x}_0^a$ , if and only if $R_{i}(\mathbf{x}_{0}^{b})\leq R_{i}(\mathbf{x}_{0}^{a})$ for all $i\in 1,2,\ldots ,m$ , and there exists $j\in 1,2,\dots,m$ such that $R_{j}(\mathbf{x}_{0}^{b}) < R_{j}(\mathbf{x}_{0}^{a})$ . For example, given the $i$ -th generated image $\mathbf{x}_0^i$ in a mini-batch, when no point in the mini-batch dominates $\mathbf{x}_0^i$ , it is referred to as a non-dominated point. + +Policy Gradient Update: We assign a reward value of zero to the data points not included in non-dominated sets and only update the gradient of these non-dominated data points as follows: + +$$ +\nabla \mathcal {J} _ {\theta} = \sum_ {k = 1} ^ {K} \frac {1}{n (\mathcal {P})} \sum_ {i = 1, \mathbf {x} _ {0} ^ {i} \in \mathcal {P}} ^ {N} \sum_ {t = 1} ^ {T} r _ {k} \left(\mathbf {x} _ {0} ^ {i}, c _ {k}\right) \times \nabla_ {\theta} \log p _ {\theta} \left(\mathbf {x} _ {t - 1} ^ {i} \mid c _ {k}, t, \mathbf {x} _ {t} ^ {i}\right), \tag {4} +$$ + +where $i$ indicates the index of images in mini-batches, and $\mathcal{P}$ denotes batch-wise a set of non-dominated points. $K$ and $T$ are the total number of reward models and total diffusion time steps, respectively. The same text prompt is used when updating the diffusion model in each batch. + +# 4.3 Original Prompt Centered Guidance + +While prompt expansion enhances details and often improves generation quality, there is a concern that the added context may dilute the main content of the original input. To mitigate this during the inference, we introduce original prompt-centered guidance. When sampling conditioned on the original prompt, the diffusion model $\epsilon_{\theta}$ typically predicts noises by combining the unconditioned score estimate and the prompt-conditioned estimate. Instead of relying solely on the expanded prompt from PEN, we propose using a linear combination of two. guidances for T2I generation: one from the user input and the other from the expanded prompt. The strength of the original prompt is controlled by guidance scales $w_{1}$ and $w_{2}$ . The noise $\bar{\epsilon}_{\theta}$ is estimated, derived from Eq. 1, as follows: + +$$ +\bar {\epsilon} _ {\theta} = w _ {1} \cdot \epsilon_ {\theta} (\mathbf {x} _ {t}, t, c) + (1 - w _ {1} - w _ {2}) \cdot \epsilon_ {\theta} (\mathbf {x} _ {t}, t, \text {n u l l}) + w _ {2} \cdot \epsilon_ {\theta} (\mathbf {x} _ {t}, t, \hat {c}), \tag {5} +$$ + +where null denotes a null text. + +# 5 Experiments + +# 5.1 Experiment Setting + +Dataset: The PEN is first supervised fine-tuned on a large-scale text dataset named the Promptist [15], which has 360K constructed prompt pairs for original prompt and prompt expansion demonstration. The original instruction "Rephrase" is included per pair in Promptist. We modify the instruction into "Input: ". This is a text input for image generation. Expand prompt for improving image quality. Output:". Subsequently, we use the RL tuning prompts (1200K) from Promptist for RL training of the PEN and T2I model. + +T2I Model: Our T2I model is based on the JAX version of Stable Diffusion 1.5 [39] pre-trained with the LAION-5B [43] dataset. We conduct experiments on a machine equipped with 16 NVIDIA RTX A100 GPUs. DDIM [46] with 50 denoising steps is used, and the classifier-free guidance weight is set to 5.0 with the resolution $512 \times 512$ . Instead of updating all layers, we specifically update the cross-attention layer in the Denoising U-Net. For optimization, we employ the Adam [25] optimizer with a learning rate of $1 \times 10^{-5}$ . + +Prompt Expansion Network: For prompt expansion, we use PaLM 2-L-IT [2], one of the PaLM2 variations, which is a multi-layer Transformer [49] decoder with casual language modeling. We optimize LoRA [20] weights for RL-based fine-tuning. The output token length of the PEN is set to 77 to match the maximum number of token length for Stable Diffusion. For original prompt-centered guidance, we set both $w_{1}$ and $w_{2}$ to 5 in Eq. 5. + +Reward Models: We incorporate four quality signals as rewards: Aesthetics, Human preference, Text-Image Alignment, Image Sentiment. For aesthetics, we use the VILA-R [24] pre-trained with the AVA [33] dataset. For human preference, we train a ViT-B/16 [10] using the Pick-a-Pic [26] dataset, which contains 500K examples for human feedback in T2I generation. The ViT-B/16 image encoder consists of 12 transformer layers, and the image resolution is $224 \times 224$ with a patch size of $16 \times 16$ . For text-image alignment, we use CLIP [35] with the image encoder ViT-B/32. For image sentiment, we use the pre-trained model from [44], which outputs three labels: positive, neutral, negative. We use the positive score ranging from 0 to 1 as the sentiment reward. + +# 5.2 Qualitative Analysis + +Comparison with Baselines: Fig. 3 shows the visual comparison of Parrot and multiple baselines. We include results from Stable Diffusion 1.5, DPOK [12] with a weighted sum of rewards, Promptist [15], and Parrot. DPOK exclusively fine-tunes the T2I model, while Promptist focuses on fine-tuning only the prompt expansion network. Parrot shows visually better images, particularly in aspects like color combination, cropping, perspective, and fine details in the image. This improvement can be attributed to Parrot's T2I model being fine-tuned together + +![](images/ae6232071aac53bcf03124b48b9d00f41b179c8709e99c0bc68d1720258677f9.jpg) +SD 1.5 +A kitchen + +![](images/d4748c79abd0214fde03e30f2695b294f6fff77162411f74f2cde462ce33828e.jpg) +DPOK +A kitchen + +![](images/f3f59b1b91f145bfff3becf93703298da87351ead285bf567944c61b0abd7f1c.jpg) +Promptist + +![](images/be9ba214bada9817db1a1920ade4c7d433e15213e9106f291c204b80436d1da7.jpg) +Parrot +A kitchen with a lot of natural light and a beautiful view of the mountains + +![](images/3a0fb3345062f16eafbb92de81c52a302a4e796892c3fce8e74314e884b0d422.jpg) +AF1 + +![](images/08b52d6d94c1cebccf191ab1c03df4c01b1de851b4e31987e8cd89d362b5a3b1.jpg) +AF1 + +![](images/74c59804e71369577864594382c6d779f194095db69ae193d868c6b76f026fec.jpg) +A kitchen, a fantasy digital painting, +trending on Artstation, +highly detailed +F1, concept art oil painting by Jama Jurabaev, extremely detailed, brush hard, artstation + +![](images/ba4568e76894545df4c02602960a8c6246e3f4bf7b50cb1604941114854948cd.jpg) +3d rendering of a f1 car in dark and moody environment with a red and orange color scheme. + +![](images/dff8bd248f9e50e0e809226043dba6e2436c567deb216e1c7e028e945b2657cf.jpg) +A woman is looking out on a winter archipelago. + +![](images/14677736b9eff726b97b475dd96c1c6eda8c0ddfa029e8d7b85473d0bef3a859.jpg) +A woman is looking out on a winter archipelago. + +![](images/4ec405cfaf3fb554d0a36f2eef601d46a48c7d2a86dcda595b57b56131c74d64.jpg) + +![](images/442e140b2561d2d17d592bbb4ec5673671caa514cce76334205728dbd355721a.jpg) +A woman is looking out on winter archipelago. The image is very peaceful and serene. + +![](images/a10b619707dca6b91163d7f9d93c02b5aa64594cf03a6c32941b916d65399e68.jpg) +A Japanese Porcelain Imari vase +Fig. 3: Comparison of Parrot and diffusion-based RL baselines. From left to right, we provide results of Stable diffusion 1.5 [39] (1st column), DPOK [12] (2nd column) with the weighted sum, Promptist [15] (3rd column), and Parrot (4th column). + +![](images/9d758abedc5dc52c4dcbdb8e4fb459a0f71a232438ea3ae857653039757e8b5c.jpg) +A Japanese Porcelain Imari vase + +![](images/d3990557d2afa7becae0faaa66fece9a3117ce00d334735a486402b3ccc58874.jpg) +Painting of a woman looking out on a winter archipelago, ultra realistic, concept art, intricate details, eerie +A Japanese Porcelain Imari vase, intricate, elegant, highly detailed, digital painting, artstation, concept art, smooth, sharp focus, illustration. + +![](images/a7d7fcd3bfa8c03c77d3b056b2bd5e9ef2d2d94c54a50e0008938bbd07015da0.jpg) +A photo of a Japanese Porcelain Imari vase on the ground with a beautiful floral pattern on the front + +![](images/af97f6e4f7cef419bb1ec58ec2b9a03f8d699915b8b980f61b4a77bfbafc55fb.jpg) + +![](images/98afd077dd00f4121c3ecb6c6bbb0f054c14642bf77463f176a2a616a1f3063c.jpg) +Fig.4: Training curve for fine-tuning on weighted sum and Parrot. For weighted sum, WS1 denotes $\{0.7, 0.1, 0.1, 0.1\}$ and WS2 denotes $\{0.25, 0.25, 0.25, 0.25\}$ for aesthetics, human preference, text-image alignment and image sentiment. Using weighted sum leads to decrease in human preference score and image sentiment score despite an improvement in the aesthetic score. In contrast, Parrot exhibits stable increases across all metrics. + +![](images/3f71f1be3076a9147b65313e7a6f37d0f4143330c01fb0134da47b1d0a3d73a2.jpg) + +![](images/8ab51a1ac6692c195c2948ef192e3f0f1a8f5a552fd98abea6a32dc546e2ad23.jpg) + +
ModelQuality Metrics
TIA (↑)Aesth. (↑)HP (↑)Sent. (↑)Average (↑)
SD 1.5 [39]0.23220.57550.19300.30100.3254
DPOK [12] (WS)0.23370.58130.19320.30130.3273 (+0.58%)
Parrot w/o PE0.23550.60340.20090.30180.3354 (+3.07%)
Parrot T2I Model Tuning Only0.25090.70730.33370.30520.3992 (+22.6%)
Promptist [15]0.14490.67830.27590.25180.3377 (+3.77 %)
Parrot with HP Only0.15430.59610.35280.25620.3398 (+4.42 %)
Parrot PEN Tuning Only0.16590.64920.26170.31310.3474 (+6.76 %)
Parrot w/o Joint Optimization0.16610.63080.25660.30840.3404 (+4.60 %)
Parrot w/o ori prompt guidance0.16230.71560.34250.31300.3833 (+17.8 %)
Parrot0.16670.73960.34110.31320.3901 (+19.8 %)
+ +Table 1: Quantitative comparison between Parrot and alternatives on the Parti dataset [53]. Abbreviations: WS - Weighted Sum; PE - Prompt Expansion; TIA - Text-Image Alignment; Aesth. - Aesthetics; HP - Human Preference; Sent. - Image Sentiment. TIA score is measured against the original prompt without expansion. + +with the prompt expansion model that incorporates aesthetic keywords during training. Parrot generates results that are more closely aligned with the input prompt, as well as more visually pleasing. + +Weighted sum vs. Parrot: Fig. 4 shows the training curve comparison of Parrot and using a linear combination of the reward scores. Each subgraph represents a reward. WS1 and WS2 denote two different weights with multiple reward scores. WS1 places greater emphasis on the aesthetic score, while WS2 adopts balanced weights across aesthetics, human preference, text-image alignment, and image sentiment. Employing the weighted sum of multiple rewards leads to a decrease in the image sentiment score, despite notable enhancements in aesthetics and human preference. In contrast, Parrot consistently exhibits improvement across all metrics. + +# 5.3 Quantitative Evaluation + +Comparison with Baselines: Table 1 presents our results of the quality score across four quality rewards: text-image alignment score, aesthetic score, human preference score, and emotion score. The first group shows methods without prompt expansion, and the second group compares methods with expansion. The prompt expansion and T2I generation are performed on the PartiPrompts [53]. Using a set of 1632 prompts, we generate 32 images for each text input and calculate the average for each metric. + +$w / o$ PE indicates the generation of images solely based on original prompt without expansion. Note that $w / o$ PE is not the same as Parrot T2I Model Tuning Only. The former takes a model trained with PEN and removes it during inference, which results in a notable disparity between prompts used in training and testing. The latter trains with the multi-objective RL using only the T2I model, and it shows substantial improvement upon DPOK [12] (weighted sum) with balanced weights of $\{0.25, 0.25, 0.25, 0.25\}$ . This shows that the proposed RL method is indeed effective in multi-objective optimization. + +For Promptist [15], we generate prompt expansion from their model. Parrot with HP Only shows solely using a human preference reward leads to decline in others. Parrot PEN Tuning Only shows suboptimal aesthetics and human preference. Parrot w/o Joint Optimization shows suboptimal results than Parrot which demonstrates the necessity of jointly optimizing PEN and T2I models. + +Our method outperforms both compared methods in aesthetics, human preference and sentiment scores. The text-image alignment score is measured with the original prompt before expansion for fair comparison. As a result, the group without prompt expansion generally shows a higher text-image alignment score. Parrot shows better text-image alignment in each subgroup. + +
AreaQuestion
Aesthetics“Which image shows better aesthetics without blurry texture, unnatural focusing, and poor color combination?”
Human Preference“Which generated image do you prefer?”
Text-Image Alignment“Which image is well aligned with the text?”
Image Sentiment“Which image is closer to amusement, excitement, and contentment?”
+ +Table 2: Questions for user study. For performing user study, we carefully design questions suitable to each quality. + +User Study: We conduct a user study using MTurk [1] with generated images from 100 random prompts in the PartiPrompts [53]. Five models are compared: Stable Diffusion v1.5, DPOK [12] with an equal weighted sum, Promptist [15], Parrot without prompt expansion, and Parrot. Each rater is presented with the original prompt (before expansion) and a set of five generated images, with the image order being randomized. Raters are then tasked with selecting the best image from the group, guided by questions outlined in Table 2. Each question + +![](images/50643f67ee5cbb0900b727734672b7bff184750fe4d7005f80ec79720216fc4e.jpg) +Fig.5: User study results on PartiPrompts [53]. Parrot outperforms baselines across all metrics. + +![](images/60c7bce9ff211310d4c2458526c643ff71c292892d6d409969b66f5973f634a7.jpg) + +![](images/511c367a10f90faea4dd8a326a1f5ec0de1aa1f9241e579495cf800156934a4a.jpg) +Fig. 6: Ablation study. We perform an ablation study by removing one of quality signals. We observe that each quality signal affects their improvement of (a) aesthetics, (b) human preference score, (c) text-image alignment score, (d) image sentiment score. +Fig.7: Comparing Parrot with/without Pareto Optimal. By pretending different reward-specific preferences, Parrot can change the tradeoff between rewards. Specifically, “”, “”, “”, “” indicates aesthetics, human preference, text-image alignment, and image sentiment respectively. With Pareto-optimal selection, each reward increases based on reward-specific preference. + +pertains to a specific quality aspect: aesthetics, human preference, text-image alignment, and image sentiment. For each prompt, 20 rounds of random sampling are conducted and sent to different raters. The user study results, illustrated in Fig. 5, show that Parrot outperforms other baselines across all dimensions. + +Proportion of non-dominated points: Using batch size of 256, we observe that in a batch around $20\%$ to $30\%$ are non-dominated points and the proportion of non-dominated points in single batch slightly increase as training proceeds. + +# 5.4 Ablations + +Effect of Pareto-optimal Multi-reward RL: To show the efficacy of Pareto-optimal Multi-reward RL, we conduct an ablation study by removing one reward model at a time. Fig. 6 shows quantitative results using one hundred random text + +![](images/4945e6cd6998b9566e3a9fb25757a5fe57230c5b9429f8c1cbf73e047211a388.jpg) +Fig. 8: The comparisons of the diffusion fine-tuning between Pareto-optimal multi-reward RL and single reward RL. We show results with same seed from various methods: Stable Diffusion 1.5 [39] (1st column), T2I model fine-tuned with the aesthetic model (2nd column), the human preference model (3rd column), text-image alignment (4th column), image sentiment (5th column), Parrot without Pareto-optimal selection (6th column) and Parrot (7th column). Parrot is effective to generate acceptable images without sacrificing one of quality signals. For example, T2I model fine-tuned with a single quality signal such as aesthetics, human preference and image sentiment results in text-image misalignment, while our method achieves a balanced visual outcome across multiple criteria. + +prompts from the Promptist [15]. We observe that our training scheme improves multiple target objectives. + +To verify whether Parrot achieved better trade-off for different reward scores, we generate 1000 images from common animal dataset [4], where each text prompt consists of the name of a common animal. As shown in Fig. 7, using only text-image alignment with reward-specific preference $\langle reward3\rangle$ generates images with higher text-image alignment score, while using only aesthetic model with reward-specific preference $\langle reward1\rangle$ yields images with higher aesthetic score. In the case of using two reward-specific preferences $\langle reward1\rangle$ , $\langle reward3\rangle$ , we observe that scores are balanced and show that better results across multiple rewards than Parrot without Pareto optimal selection. + +Fig 8 shows the visual comparison between Parrot, Parrot with a single reward, and Parrot without selecting the batch-wise Pareto-optimal solution. Using a single reward model tends to result in degradation of other rewards, especially text-image alignment. In the third column, results of the first row miss the text $a$ tophat in input prompt, even though the Stable Diffusion result includes that attribute. On the other hand, Parrot results capture all prompts, improving other quality signals, such as aesthetics, image sentiment and human preference. + +Effect of Original Prompt Centered Guidance: Fig 9 shows the effect of the proposed original prompt-centered guidance. As evident from the figure, using only the expanded prompt as input often results in the main content be + +ing overwhelmed by the added context. For instance, given the original prompt "A shiba inu", the result from the expanded prompt shows a zoomed-out image and the intended main subject (shiba inu) becomes small. The proposed original prompt-centered guidance effectively addresses this issue, generating an image that faithfully captures the input prompt while incorporating visually more pleasing details. + +![](images/6a19c507fef9901d104d5d9638edfb87354d7390120f50624207b296ddd439f6.jpg) +Fig. 9: Results of original prompt centered guidance. As we expand the prompt, the content in the generated image often fades away. This guidance is helpful for keeping the main content of the original prompt. + +# 6 Conclusion and Limitation + +We propose Parrot, a novel multi-reward optimization algorithm aimed to improve text-to-image generation by effectively optimizing multiple quality rewards using RL. With batch-wise Pareto-optimal selection, Parrot adaptively balance the optimization of multiple quality rewards. By applying Parrot to jointly fine-tune both the T2I model and the prompt expansion model, we achieve the generation of higher-quality images with richer details. Additionally, our original prompt centered guidance technique ensures that the generated image maintains fidelity to the user prompt after prompt expansion during inference. Results from the user study indicate that Parrot significantly improves the quality of generated images across multiple criteria, including text-image alignment, human preference, aesthetics, and image sentiment. While Parrot has shown effectiveness in enhancing generated image quality, its efficacy is limited by the quality metrics it relies on. Therefore, advancements of the generated image quality metrics will directly enhance the capabilities of Parrot. Additionally, Parrot is adaptable to a broader range of rewards that quantify generated image quality. + +Societal Impact: Parrot could potentially raise ethical concerns related to the generation of immoral content. This concern stems from the user's ability to influence T2I generation, allowing for the creation of visual content that may be deemed inappropriate. The risk may be tied to the potential biases in reward models inherited from various datasets. + +# References + +1. Amazon mechanical turk. https://www.mturk.com/ (2005) +2. Anil, R., Dai, A.M., First, O., Johnson, M., Lepikhin, D., Passos, A., Shakeri, S., Taropa, E., Bailey, P., Chen, Z., et al.: Palm 2 technical report. arXiv preprint arXiv:2305.10403 (2023) +3. Bai, Y., Jones, A., Ndousse, K., Askell, A., Chen, A., DasSarma, N., Drain, D., Fort, S., Ganguli, D., Henighan, T., et al.: Training a helpful and harmless assistant with reinforcement learning from human feedback. arXiv preprint arXiv:2204.05862 (2022) +4. Black, K., Janner, M., Du, Y., Kostrikov, I., Levine, S.: Training diffusion models with reinforcement learning. arXiv preprint arXiv:2305.13301 (2023) +5. Chang, H., Zhang, H., Barber, J., Maschinot, A., Lezama, J., Jiang, L., Yang, M.H., Murphy, K., Freeman, W.T., Rubinstein, M., et al.: Muse: Text-to-image generation via masked generative transformers. arXiv preprint arXiv:2301.00704 (2023) +6. Clark, K., Vicol, P., Swersky, K., Fleet, D.J.: Directly fine-tuning diffusion models on differentiable rewards. ICLR (2024) +7. Dai, X., Hou, J., Ma, C.Y., Tsai, S., Wang, J., Wang, R., Zhang, P., Vandenhende, S., Wang, X., Dubey, A., et al.: Emu: Enhancing image generation models using photogenic needles in a haystack. arXiv preprint arXiv:2309.15807 (2023) +8. Deng, F., Wang, Q., Wei, W., Grundmann, M., Hou, T.: Prdp: Proximal reward difference prediction for large-scale reward finetuning of diffusion models. CVPR (2024) +9. Dong, H., Xiong, W., Goyal, D., Pan, R., Dao, S., Zhang, J., Shum, K., Zhang, T.: Raft: Reward ranked finetuning for generative foundation model alignment. arXiv preprint arXiv:2304.06767 (2023) +10. Dosovitskiy, A., Beyer, L., Kolesnikov, A., Weissenborn, D., Zhai, X., Unterthiner, T., Dehghani, M., Minderer, M., Heigold, G., Gelly, S., et al.: An image is worth 16x16 words: Transformers for image recognition at scale. arXiv preprint arXiv:2010.11929 (2020) +11. Fan, Y., Lee, K.: Optimizing ddpm sampling with shortcut fine-tuning. ICML (2023) +12. Fan, Y., Watkins, O., Du, Y., Liu, H., Ryu, M., Boutilier, C., Abbeel, P., Ghavamzadeh, M., Lee, K., Lee, K.: Dpok: Reinforcement learning for fine-tuning text-to-image diffusion models. NeurIPS (2023) +13. Fang, Y., Zhu, H., Zeng, Y., Ma, K., Wang, Z.: Perceptual quality assessment of smartphone photography. In: CVPR (2020) +14. Han, L., Li, Y., Zhang, H., Milanfar, P., Metaxas, D., Yang, F.: Svdiff: Compact parameter space for diffusion fine-tuning. CVPR (2023) +15. Hao, Y., Chi, Z., Dong, L., Wei, F.: Optimizing prompts for text-to-image generation. arXiv preprint arXiv:2212.09611 (2022) +16. He, H., Wang, T., Yang, H., Fu, J., Yuan, N.J., Yin, J., Chao, H., Zhang, Q.: Learning profitable nft image diffusions via multiple visual-policy guided reinforcement learning. arXiv preprint arXiv:2306.11731 (2023) +17. Ho, J., Jain, A., Abbeel, P.: Denoising diffusion probabilistic models. NeurIPS (2020) +18. Ho, J., Salimans, T.: Classifier-free diffusion guidance. arXiv preprint arXiv:2207.12598 (2022) + +19. Hosu, V., Lin, H., Sziranyi, T., Saupe, D.: Koniq-10k: An ecologically valid database for deep learning of blind image quality assessment. TIP 29, 4041-4056 (2020) +20. Hu, E.J., Shen, Y., Wallis, P., Allen-Zhu, Z., Li, Y., Wang, S., Wang, L., Chen, W.: Lora: Low-rank adaptation of large language models. arXiv preprint arXiv:2106.09685 (2021) +21. Jeong, Y., Ryoo, W., Lee, S., Seo, D., Byeon, W., Kim, S., Kim, J.: The power of sound (tpos): Audio reactive video generation with stable diffusion. In: ICCV (2023) +22. Kawar, B., Zada, S., Lang, O., Tov, O., Chang, H., Dekel, T., Mosseri, I., Irani, M.: Imagic: Text-based real image editing with diffusion models. In: CVPR (2023) +23. Ke, J., Wang, Q., Wang, Y., Milanfar, P., Yang, F.: Musiq: Multi-scale image quality transformer. In: ICCV (2021) +24. Ke, J., Ye, K., Yu, J., Wu, Y., Milanfar, P., Yang, F.: Vila: Learning image aesthetics from user comments with vision-language pretraining. In: CVPR (2023) +25. Kingma, D.P., Ba, J.: Adam: A method for stochastic optimization. ICLR (2015) +26. Kirstain, Y., Polyak, A., Singer, U., Matiana, S., Penna, J., Levy, O.: Pick-a-pic: An open dataset of user preferences for text-to-image generation. arXiv preprint arXiv:2305.01569 (2023) +27. Lee, K., Liu, H., Ryu, M., Watkins, O., Du, Y., Boutilier, C., Abbeel, P., Ghavamzadeh, M., Gu, S.S.: Aligning text-to-image models using human feedback. arXiv preprint arXiv:2302.12192 (2023) +28. Lee, S.H., Kim, S., Yoo, I., Yang, F., Cho, D., Kim, Y., Chang, H., Kim, J., Kim, S.: Soundini: Sound-guided diffusion for natural video editing. arXiv preprint arXiv:2304.06818 (2023) +29. Li, Y., Liu, H., Wu, Q., Mu, F., Yang, J., Gao, J., Li, C., Lee, Y.J.: Gligen: Open-set grounded text-to-image generation. In: CVPR (2023) +30. Lin, X., Yang, Z., Zhang, X., Zhang, Q.: Pareto set learning for expensive multi-objective optimization. NeurIPS (2022) +31. Mannor, S., Shimkin, N.: The steering approach for multi-criteria reinforcement learning. NeurIPS (2001) +32. Miettinen, K.: Nonlinear multiobjective optimization, vol. 12. Springer Science & Business Media (1999) +33. Murray, N., Marchesotti, L., Perronnin, F.: Ava: A large-scale database for aesthetic visual analysis. In: CVPR (2012) +34. Ouyang, L., Wu, J., Jiang, X., Almeida, D., Wainwright, C., Mishkin, P., Zhang, C., Agarwal, S., Slama, K., Ray, A., et al.: Training language models to follow instructions with human feedback. NeurIPS (2022) +35. Radford, A., Kim, J.W., Hallacy, C., Ramesh, A., Goh, G., Agarwal, S., Sastry, G., Askell, A., Mishkin, P., Clark, J., et al.: Learning transferable visual models from natural language supervision. In: ICML (2021) +36. Rame, A., Couairon, G., Shukor, M., Dancette, C., Gaya, J.B., Soulier, L., Cord, M.: Rewarded soups: towards pareto-optimal alignment by interpolating weights fine-tuned on diverse rewards. NeurIPS (2023) +37. Ramesh, A., Dhariwal, P., Nichol, A., Chu, C., Chen, M.: Hierarchical text-conditional image generation with clip latents. arXiv preprint arXiv:2204.06125 (2022) +38. Richardson, E., Goldberg, K., Alaluf, Y., Cohen-Or, D.: Conceptlab: Creative generation using diffusion prior constraints. arXiv preprint arXiv:2308.02669 (2023) +39. Rombach, R., Blattmann, A., Lorenz, D., Esser, P., Ommer, B.: High-resolution image synthesis with latent diffusion models. In: CVPR (2022) + +40. Ronneberger, O., Fischer, P., Brox, T.: U-net: Convolutional networks for biomedical image segmentation. In: MICCAI (2015) +41. Sahara, C., Chan, W., Chang, H., Lee, C., Ho, J., Salimans, T., Fleet, D., Norouzi, M.: Palette: Image-to-image diffusion models. In: SIGGRAPH (2022) +42. Sahara, C., Chan, W., Saxena, S., Li, L., Whang, J., Denton, E.L., Ghasemipour, K., Gontijo Lopes, R., Karagol Ayan, B., Salimans, T., et al.: Photorealistic text-to-image diffusion models with deep language understanding. NeurIPS (2022) +43. Schuhmann, C., Beaumont, R., Vencu, R., Gordon, C., Wightman, R., Cherti, M., Coombes, T., Katta, A., Mullis, C., Wortsman, M., et al.: Laion-5b: An open large-scale dataset for training next generation image-text models. NeurIPS (2022) +44. Serra, A., Carrara, F., Tesconi, M., Falchi, F.: The emotions of the crowd: Learning image sentiment from tweets via cross-modal distillation. ECAI (2023) +45. Sohl-Dickstein, J., Weiss, E., Maheswaranathan, N., Ganguli, S.: Deep unsupervised learning using nonequilibrium thermodynamics. In: ICML (2015) +46. Song, J., Meng, C., Ermon, S.: Denoising diffusion implicit models. arXiv preprint arXiv:2010.02502 (2020) +47. Tesauro, G., Das, R., Chan, H., Kephart, J., Levine, D., Rawson, F., Lefury, C.: Managing power consumption and performance of computing systems using reinforcement learning. NeurIPS (2007) +48. Tu, Z., Talebi, H., Zhang, H., Yang, F., Milanfar, P., Bovik, A., Li, Y.: Maxvit: Multi-axis vision transformer. In: ECCV (2022) +49. Vaswani, A., Shazeer, N., Parmar, N., Uszkoreit, J., Jones, L., Gomez, A.N., Kaiser, L., Polosukhin, I.: Attention is all you need. NeurIPS (2017) +50. Wu, X., Sun, K., Zhu, F., Zhao, R., Li, H.: Human preference score: Better aligning text-to-image models with human preference. In: ICCV (2023) +51. Xu, J., Liu, X., Wu, Y., Tong, Y., Li, Q., Ding, M., Tang, J., Dong, Y.: Imagereward: Learning and evaluating human preferences for text-to-image generation. arXiv preprint arXiv:2304.05977 (2023) +52. Ying, Z., Niu, H., Gupta, P., Mahajan, D., Ghadiyaram, D., Bovik, A.: From patches to pictures (paq-2-piq): Mapping the perceptual space of picture quality. In: CVPR (2020) +53. Yu, J., Xu, Y., Koh, J.Y., Luong, T., Baid, G., Wang, Z., Vasudevan, V., Ku, A., Yang, Y., Ayan, B.K., et al.: Scaling autoregressive models for content-rich text-to-image generation. arXiv preprint arXiv:2206.10789 (2022) +54. Yu, L., Cheng, Y., Sohn, K., Lezama, J., Zhang, H., Chang, H., Hauptmann, A.G., Yang, M.H., Hao, Y., Essa, I., et al.: Magvit: Masked generative video transformer. In: CVPR (2023) +55. Zhou, Y., Liu, B., Zhu, Y., Yang, X., Chen, C., Xu, J.: Shifted diffusion for text-to-image generation. In: CVPR (2023) \ No newline at end of file diff --git a/2024/Parrot_ Pareto-optimal Multi-Reward Reinforcement Learning Framework for Text-to-Image Generation/images.zip b/2024/Parrot_ Pareto-optimal Multi-Reward Reinforcement Learning Framework for Text-to-Image Generation/images.zip new file mode 100644 index 0000000000000000000000000000000000000000..3f34a614395f3d99da33560e9b069d8d9ad7b8be --- /dev/null +++ b/2024/Parrot_ Pareto-optimal Multi-Reward Reinforcement Learning Framework for Text-to-Image Generation/images.zip @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:22777725ace414fbf10476484f32e169a23b69ff703714b02a0763c3965345ae +size 717029 diff --git a/2024/Parrot_ Pareto-optimal Multi-Reward Reinforcement Learning Framework for Text-to-Image Generation/layout.json b/2024/Parrot_ Pareto-optimal Multi-Reward Reinforcement Learning Framework for Text-to-Image Generation/layout.json new file mode 100644 index 0000000000000000000000000000000000000000..f021687bcd76ce3bbfcce48746a92a2ef2350e58 --- /dev/null +++ b/2024/Parrot_ Pareto-optimal Multi-Reward Reinforcement Learning Framework for Text-to-Image Generation/layout.json @@ -0,0 +1,11065 @@ +{ + "pdf_info": [ + { + "para_blocks": [ + { + "bbox": [ + 170, + 111, + 445, + 166 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 170, + 111, + 445, + 166 + ], + "spans": [ + { + "bbox": [ + 170, + 111, + 445, + 166 + ], + "type": "text", + "content": "Parrot: Pareto-optimal Multi-Reward Reinforcement Learning Framework for Text-to-Image Generation" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 140, + 185, + 473, + 224 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 140, + 185, + 473, + 224 + ], + "spans": [ + { + "bbox": [ + 140, + 185, + 473, + 224 + ], + "type": "text", + "content": "Seung Hyun Lee" + }, + { + "bbox": [ + 140, + 185, + 473, + 224 + ], + "type": "inline_equation", + "content": "^{1,6*}" + }, + { + "bbox": [ + 140, + 185, + 473, + 224 + ], + "type": "text", + "content": ", Yinxiao Li" + }, + { + "bbox": [ + 140, + 185, + 473, + 224 + ], + "type": "inline_equation", + "content": "^{1}" + }, + { + "bbox": [ + 140, + 185, + 473, + 224 + ], + "type": "text", + "content": ", Junjie Ke" + }, + { + "bbox": [ + 140, + 185, + 473, + 224 + ], + "type": "inline_equation", + "content": "^{1}" + }, + { + "bbox": [ + 140, + 185, + 473, + 224 + ], + "type": "text", + "content": ", Innfarn Yoo" + }, + { + "bbox": [ + 140, + 185, + 473, + 224 + ], + "type": "inline_equation", + "content": "^{2}" + }, + { + "bbox": [ + 140, + 185, + 473, + 224 + ], + "type": "text", + "content": ", Han Zhang" + }, + { + "bbox": [ + 140, + 185, + 473, + 224 + ], + "type": "inline_equation", + "content": "^{3}" + }, + { + "bbox": [ + 140, + 185, + 473, + 224 + ], + "type": "text", + "content": ", Jiahui Yu" + }, + { + "bbox": [ + 140, + 185, + 473, + 224 + ], + "type": "inline_equation", + "content": "^{4\\dagger}" + }, + { + "bbox": [ + 140, + 185, + 473, + 224 + ], + "type": "text", + "content": ", Qifei Wang" + }, + { + "bbox": [ + 140, + 185, + 473, + 224 + ], + "type": "inline_equation", + "content": "^{2}" + }, + { + "bbox": [ + 140, + 185, + 473, + 224 + ], + "type": "text", + "content": ", Fei Deng" + }, + { + "bbox": [ + 140, + 185, + 473, + 224 + ], + "type": "inline_equation", + "content": "^{2,5*}" + }, + { + "bbox": [ + 140, + 185, + 473, + 224 + ], + "type": "text", + "content": ", Glenn Entis" + }, + { + "bbox": [ + 140, + 185, + 473, + 224 + ], + "type": "inline_equation", + "content": "^{1}" + }, + { + "bbox": [ + 140, + 185, + 473, + 224 + ], + "type": "text", + "content": ", Junfeng He" + }, + { + "bbox": [ + 140, + 185, + 473, + 224 + ], + "type": "inline_equation", + "content": "^{1}" + }, + { + "bbox": [ + 140, + 185, + 473, + 224 + ], + "type": "text", + "content": ", Gang Li" + }, + { + "bbox": [ + 140, + 185, + 473, + 224 + ], + "type": "inline_equation", + "content": "^{1}" + }, + { + "bbox": [ + 140, + 185, + 473, + 224 + ], + "type": "text", + "content": ", Sangpil Kim" + }, + { + "bbox": [ + 140, + 185, + 473, + 224 + ], + "type": "inline_equation", + "content": "^{6}" + }, + { + "bbox": [ + 140, + 185, + 473, + 224 + ], + "type": "text", + "content": ", Irfan Essa" + }, + { + "bbox": [ + 140, + 185, + 473, + 224 + ], + "type": "inline_equation", + "content": "^{1}" + }, + { + "bbox": [ + 140, + 185, + 473, + 224 + ], + "type": "text", + "content": ", Feng Yang" + }, + { + "bbox": [ + 140, + 185, + 473, + 224 + ], + "type": "inline_equation", + "content": "^{1}" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 187, + 231, + 426, + 255 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 187, + 231, + 426, + 255 + ], + "spans": [ + { + "bbox": [ + 187, + 231, + 426, + 255 + ], + "type": "text", + "content": "Google Research1, Google2, Google DeepMind3, OpenAI4, Rutgers University5, Korea University6" + } + ] + } + ], + "index": 2 + }, + { + "type": "image", + "bbox": [ + 133, + 257, + 482, + 453 + ], + "blocks": [ + { + "bbox": [ + 133, + 257, + 482, + 453 + ], + "lines": [ + { + "bbox": [ + 133, + 257, + 482, + 453 + ], + "spans": [ + { + "bbox": [ + 133, + 257, + 482, + 453 + ], + "type": "image", + "image_path": "4c4a4ecb41a55053004f374235ae1011858d93e51976d24ef4603162331103c4.jpg" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 132, + 455, + 482, + 489 + ], + "lines": [ + { + "bbox": [ + 132, + 455, + 482, + 489 + ], + "spans": [ + { + "bbox": [ + 132, + 455, + 482, + 489 + ], + "type": "text", + "content": "Fig. 1: Parrot visual examples. Parrot consistently improves the quality of generated images across multiple criteria: aesthetics, human preference, text-image alignment, and image sentiment. Each column shows generated images using the same seed." + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "image_caption" + } + ], + "index": 3 + }, + { + "bbox": [ + 159, + 507, + 455, + 639 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 159, + 507, + 455, + 639 + ], + "spans": [ + { + "bbox": [ + 159, + 507, + 455, + 639 + ], + "type": "text", + "content": "Abstract. Recent works have demonstrated that using reinforcement learning (RL) with multiple quality rewards can improve the quality of generated images in text-to-image (T2I) generation. However, manually adjusting reward weights poses challenges and may cause over-optimization in certain metrics. To solve this, we propose Parrot, which addresses the issue through multi-objective optimization and introduces an effective multi-reward optimization strategy to approximate Pareto optimal. Utilizing batch-wise Pareto optimal selection, Parrot automatically identifies the optimal trade-off among different rewards. We use the novel multi-reward optimization algorithm to jointly optimize the T2I model and a prompt expansion network, resulting in significant improvement of image quality and also allow to control the trade-off of" + } + ] + } + ], + "index": 5 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 141, + 647, + 333, + 656 + ], + "type": "page_footnote", + "angle": 0, + "lines": [ + { + "bbox": [ + 141, + 647, + 333, + 656 + ], + "spans": [ + { + "bbox": [ + 141, + 647, + 333, + 656 + ], + "type": "text", + "content": "* This work was done during an internship at Google." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 141, + 656, + 314, + 665 + ], + "type": "page_footnote", + "angle": 0, + "lines": [ + { + "bbox": [ + 141, + 656, + 314, + 665 + ], + "spans": [ + { + "bbox": [ + 141, + 656, + 314, + 665 + ], + "type": "text", + "content": "† This work was done during working at Google." + } + ] + } + ], + "index": 7 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 0 + }, + { + "para_blocks": [ + { + "bbox": [ + 159, + 117, + 455, + 184 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 159, + 117, + 455, + 184 + ], + "spans": [ + { + "bbox": [ + 159, + 117, + 455, + 184 + ], + "type": "text", + "content": "different rewards using a reward related prompt during inference. Furthermore, we introduce original prompt-centered guidance at inference time, ensuring fidelity to user input after prompt expansion. Extensive experiments and a user study validate the superiority of Parrot over several baselines across various quality criteria, including aesthetics, human preference, text-image alignment, and image sentiment." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 132, + 205, + 230, + 219 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 205, + 230, + 219 + ], + "spans": [ + { + "bbox": [ + 132, + 205, + 230, + 219 + ], + "type": "text", + "content": "1 Introduction" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 130, + 233, + 482, + 341 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 233, + 482, + 341 + ], + "spans": [ + { + "bbox": [ + 130, + 233, + 482, + 341 + ], + "type": "text", + "content": "Despite significant advancements in text-to-image (T2I) generation [17, 29, 38, 45, 55], recent work like Imagen [42] and Stable Diffusion [39], still struggle to produce high quality images. Images in the first row in Fig. 1 illustrates such quality issues in Stable Diffusion [39], including poor composition (e.g. bad cropping), misalignment with input prompts (e.g. missing objects), or overall lack of aesthetic appeal. Assessing the quality of generated images can involve various metrics such as aesthetics [23, 24, 33], human preference [26], text-image alignment [35], and emotional appeal [44]. Enhancing T2I generation across multiple quality metrics remains a challenging task." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 130, + 342, + 483, + 461 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 342, + 483, + 461 + ], + "spans": [ + { + "bbox": [ + 130, + 342, + 483, + 461 + ], + "type": "text", + "content": "Recent works [4,12,15,27] have demonstrated that incorporating quality signals as reward functions in fine-tuning T2I with reinforcement learning (RL) can improve image quality. For instance, Promptist [15] fine-tunes prompt expansion model using RL with the sum of aesthetics and text-image alignment scores. However, the simple weighted sum approach may not effectively handle trade-offs among multiple quality metrics. As the number of rewards increases, manually adjusting reward weights becomes impractical. Moreover, optimizing one quality metric may inadvertently compromise others, as the model might prioritize aesthetics over relevance to the input prompt. Additionally, the trade-off for different reward is not controllable after training." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 130, + 462, + 483, + 666 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 462, + 483, + 666 + ], + "spans": [ + { + "bbox": [ + 130, + 462, + 483, + 666 + ], + "type": "text", + "content": "To address these challenges, we propose Parrot, a Pareto-optimal multi-reward reinforcement learning algorithm to improve text-to-image generation. Unlike previous approaches that treat T2I reward optimization as a single objective optimization problem, Parrot tackles this challenge through multi-objective optimization and introduces an effective multi-reward optimization strategy to achieve Pareto optimal approximation. Intuitively, each generated sample in a batch embodies a distinctive trade-off among various quality rewards, with some samples exhibiting superior trade-offs compared to others. Instead of updating gradients using all batch samples, Parrot uses non-dominated points [32] which have better trade-offs. Consequently, Parrot automatically learns from the optimal trade-off among different rewards. Moreover, Parrot learns reward-specific preference prompts, which can be utilized individually or in combination to control the trade-off among different rewards during inference time. Unlike prior work, which either solely fine-tunes the T2I model [12] or only tunes the prompt expansion network while freezing the T2I model [15], we employ the Parrot multi-reward optimization algorithm to jointly optimize both the T2I model and the prompt expansion network. This collaborative optimization unlocks the" + } + ] + } + ], + "index": 6 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "spans": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "type": "text", + "content": "2" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 166, + 91, + 260, + 102 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 166, + 91, + 260, + 102 + ], + "spans": [ + { + "bbox": [ + 166, + 91, + 260, + 102 + ], + "type": "text", + "content": "Lee, Seung Hyun et al." + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 1 + }, + { + "para_blocks": [ + { + "bbox": [ + 130, + 116, + 482, + 176 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 116, + 482, + 176 + ], + "spans": [ + { + "bbox": [ + 130, + 116, + 482, + 176 + ], + "type": "text", + "content": "full potential of Parrot by encouraging both more details from added context from the prompt expansion model, and the overall quality improvement on the T2I generation. During inference, we further introduce original prompt-centered guidance to ensure the output image is relevant to input prompts after prompt expansion." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 146, + 177, + 389, + 189 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 146, + 177, + 389, + 189 + ], + "spans": [ + { + "bbox": [ + 146, + 177, + 389, + 189 + ], + "type": "text", + "content": "In summary, our contributions can be listed as follows:" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 137, + 190, + 479, + 344 + ], + "type": "list", + "angle": 0, + "index": 8, + "blocks": [ + { + "bbox": [ + 138, + 190, + 479, + 260 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 190, + 479, + 260 + ], + "spans": [ + { + "bbox": [ + 138, + 190, + 479, + 260 + ], + "type": "text", + "content": "- We propose Parrot, a novel multi-reward optimization algorithm for T2I RL fine-tuning. Leveraging batch-wise Pareto-optimal selection, it effectively optimizes multiple T2I rewards, enabling collaborative improvement in aesthetics, human preference, image sentiment, and text-image alignment and also allowing to control the trade-off of different rewards using a reward related prompt during inference." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 137, + 262, + 479, + 285 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 137, + 262, + 479, + 285 + ], + "spans": [ + { + "bbox": [ + 137, + 262, + 479, + 285 + ], + "type": "text", + "content": "- We show the advantage of jointly optimizing both the prompt expansion network and the T2I model, which has never been explored before." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 137, + 286, + 479, + 320 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 137, + 286, + 479, + 320 + ], + "spans": [ + { + "bbox": [ + 137, + 286, + 479, + 320 + ], + "type": "text", + "content": "- We introduce original prompt-centered guidance during inference time after prompt expansion, ensuring better alignment with the original prompt while enriching image details." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 137, + 322, + 479, + 344 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 137, + 322, + 479, + 344 + ], + "spans": [ + { + "bbox": [ + 137, + 322, + 479, + 344 + ], + "type": "text", + "content": "- Extensive results and a user study validate that Parrot outperforms several baseline methods across various quality criteria." + } + ] + } + ], + "index": 7 + } + ], + "sub_type": "text" + }, + { + "bbox": [ + 132, + 370, + 237, + 383 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 370, + 237, + 383 + ], + "spans": [ + { + "bbox": [ + 132, + 370, + 237, + 383 + ], + "type": "text", + "content": "2 Related Work" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 130, + 407, + 482, + 491 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 407, + 482, + 491 + ], + "spans": [ + { + "bbox": [ + 130, + 407, + 482, + 491 + ], + "type": "text", + "content": "T2I Generation: The goal of T2I generation is to create an image given an input text prompt. Several T2I generative models have been proposed and have demonstrated promising results [5, 7, 14, 21, 22, 28, 37, 41, 42, 54]. Stable Diffusion [39] shows impressive generation performance in T2I generation, leveraging latent text representations from LLMs. Despite substantial progress, the images generated by those models still exhibit quality issues, such as bad cropping or misalignment with the input texts." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 130, + 498, + 482, + 665 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 498, + 482, + 665 + ], + "spans": [ + { + "bbox": [ + 130, + 498, + 482, + 665 + ], + "type": "text", + "content": "RL for T2I Fine-tuning: Starting by Fan et al. [11] to explore RL fine-tuning for T2I models, following works [4, 8, 9, 12, 16] have explored RL fine-tuning technique for T2I diffusion model, showcasing superior performance for human preference learning. DPOK [12] improves quality through RL using ImageReward [51] score as a reward with a few prompts. In addition to fine-tuning the T2I model directly using RL, Promptist [15] fine-tunes the prompt expansion model by using a simple sum of aesthetic and text-image alignment scores as reward. DRaFT [6] proposed not only differentiable rewards for efficient fine-tuning but also effectiveness of using linear summation of multi-rewards. These methods treat T2I RL as a single-objective optimization problem, while Parrot employs multi-objective optimization. Additionally, prior approaches either fine-tune the T2I model or the prompt expansion model while freezing the other. In contrast, Parrot proposes joint optimization of the prompt expansion model and the T2I model using multi-reward RL to foster better collaboration." + } + ] + } + ], + "index": 11 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 182, + 91, + 447, + 102 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 182, + 91, + 447, + 102 + ], + "spans": [ + { + "bbox": [ + 182, + 91, + 447, + 102 + ], + "type": "text", + "content": "Pareto-optimal Multi-Reward RL Framework for T2I Generation" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 474, + 92, + 480, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 474, + 92, + 480, + 100 + ], + "spans": [ + { + "bbox": [ + 474, + 92, + 480, + 100 + ], + "type": "text", + "content": "3" + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 2 + }, + { + "para_blocks": [ + { + "bbox": [ + 130, + 116, + 482, + 271 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 116, + 482, + 271 + ], + "spans": [ + { + "bbox": [ + 130, + 116, + 482, + 271 + ], + "type": "text", + "content": "Multi-objective Optimization: Multi-objective optimization problem involves optimizing multiple objective functions simultaneously. The scalarization technique [31, 47] formulates multi-objective problem into single-objective problems with the weighted sum of each score, which requires pre-defined weights for each objective. Rame et al. [36] proposed weighted averaging method to find Pareto frontier, leveraging multiple fine-tuned models. Lin et al. [30] proposes to learn a set model to map trade-off preference vectors to their corresponding Pareto solutions. Inspired by this, Parrot introduces a language-based preference vector constructed from task identifiers for each reward, then encoded by the text-encoder. In the context of multi-reward RL for T2I diffusion models, Promptist [15] uses a simple weighted sum of two reward scores. This approach requires manual tuning of the weights, which makes it time-consuming and hard to scale when the number of rewards increases." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 130, + 277, + 482, + 458 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 277, + 482, + 458 + ], + "spans": [ + { + "bbox": [ + 130, + 277, + 482, + 458 + ], + "type": "text", + "content": "Generated Image Quality: The quality assessment of images generated by T2I models involves multiple dimensions, and various metrics have been proposed. In this paper, we consider using four types of quality metrics as rewards: aesthetics, human preference, text-image alignment, and image sentiment. Aesthetics captures the overall visual appealingness of the image, and it is learned using human ratings for aesthetics in real images [13, 19, 23, 24, 33, 48, 52]. Human preferences, rooted the concept of learning from human feedback [3, 34, 50], involves gathering preferences at scale by having raters to compare generated images [26, 51]. Text-image alignment measures the extent to which the generated image aligns with the input prompt, CLIP [35] score is often employed, measuring the cosine distance of between contrastive image embedding and text embedding. Image sentiment is important for ensuring the generated image evokes positive emotions in the viewer. Serra et al. [44] predict average polarity of sentiments an image elicits and learn estimates for positive, neutral, and negative scores. In Parrot, we use its positive score as a reward for positive emotions." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 132, + 478, + 226, + 493 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 478, + 226, + 493 + ], + "spans": [ + { + "bbox": [ + 132, + 478, + 226, + 493 + ], + "type": "text", + "content": "3 Preliminary" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 130, + 511, + 482, + 596 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 511, + 482, + 596 + ], + "spans": [ + { + "bbox": [ + 130, + 511, + 482, + 596 + ], + "type": "text", + "content": "Diffusion Probabilistic Models: Diffusion probabilistic models [17] generate the image by gradually denoising a noisy image. Specifically, given a real image " + }, + { + "bbox": [ + 130, + 511, + 482, + 596 + ], + "type": "inline_equation", + "content": "\\mathbf{x}_0" + }, + { + "bbox": [ + 130, + 511, + 482, + 596 + ], + "type": "text", + "content": " from the data distribution " + }, + { + "bbox": [ + 130, + 511, + 482, + 596 + ], + "type": "inline_equation", + "content": "\\mathbf{x}_0 \\sim q(\\mathbf{x}_0)" + }, + { + "bbox": [ + 130, + 511, + 482, + 596 + ], + "type": "text", + "content": ", the forward process " + }, + { + "bbox": [ + 130, + 511, + 482, + 596 + ], + "type": "inline_equation", + "content": "q(\\mathbf{x}_t | \\mathbf{x}_0, c)" + }, + { + "bbox": [ + 130, + 511, + 482, + 596 + ], + "type": "text", + "content": " of diffusion probabilistic models produce a noisy image " + }, + { + "bbox": [ + 130, + 511, + 482, + 596 + ], + "type": "inline_equation", + "content": "\\mathbf{x}_t" + }, + { + "bbox": [ + 130, + 511, + 482, + 596 + ], + "type": "text", + "content": ", which induces a distribution " + }, + { + "bbox": [ + 130, + 511, + 482, + 596 + ], + "type": "inline_equation", + "content": "p(\\mathbf{x}_0, c)" + }, + { + "bbox": [ + 130, + 511, + 482, + 596 + ], + "type": "text", + "content": " conditioned on text prompt " + }, + { + "bbox": [ + 130, + 511, + 482, + 596 + ], + "type": "inline_equation", + "content": "c" + }, + { + "bbox": [ + 130, + 511, + 482, + 596 + ], + "type": "text", + "content": ". In classifier-free guidance [18], denoising model predicts noise " + }, + { + "bbox": [ + 130, + 511, + 482, + 596 + ], + "type": "inline_equation", + "content": "\\bar{\\epsilon}_{\\theta}" + }, + { + "bbox": [ + 130, + 511, + 482, + 596 + ], + "type": "text", + "content": " with a linear combination of the unconditional score estimates " + }, + { + "bbox": [ + 130, + 511, + 482, + 596 + ], + "type": "inline_equation", + "content": "\\epsilon_{\\theta}(\\mathbf{x}_t, t)" + }, + { + "bbox": [ + 130, + 511, + 482, + 596 + ], + "type": "text", + "content": " and the conditional score estimates " + }, + { + "bbox": [ + 130, + 511, + 482, + 596 + ], + "type": "inline_equation", + "content": "\\epsilon_{\\theta}(\\mathbf{x}_t, t, c)" + }, + { + "bbox": [ + 130, + 511, + 482, + 596 + ], + "type": "text", + "content": " as follows:" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 212, + 607, + 481, + 620 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 212, + 607, + 481, + 620 + ], + "spans": [ + { + "bbox": [ + 212, + 607, + 481, + 620 + ], + "type": "interline_equation", + "content": "\\bar {\\epsilon} _ {\\theta} = w \\cdot \\epsilon_ {\\theta} (\\mathbf {x} _ {t}, t, c) + (1 - w) \\cdot \\epsilon_ {\\theta} (\\mathbf {x} _ {t}, t, \\mathrm {n u l l}), \\tag {1}", + "image_path": "00acee124627b4ccf7b6693c7821c38c2788144ae8a181680677c35050d3633f.jpg" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 130, + 629, + 482, + 667 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 629, + 482, + 667 + ], + "spans": [ + { + "bbox": [ + 130, + 629, + 482, + 667 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 130, + 629, + 482, + 667 + ], + "type": "inline_equation", + "content": "t" + }, + { + "bbox": [ + 130, + 629, + 482, + 667 + ], + "type": "text", + "content": " denotes diffusion time step, the null indicates a null text and " + }, + { + "bbox": [ + 130, + 629, + 482, + 667 + ], + "type": "inline_equation", + "content": "w" + }, + { + "bbox": [ + 130, + 629, + 482, + 667 + ], + "type": "text", + "content": " represents the guidance scale of classifier-free guidance where " + }, + { + "bbox": [ + 130, + 629, + 482, + 667 + ], + "type": "inline_equation", + "content": "w \\geq 1" + }, + { + "bbox": [ + 130, + 629, + 482, + 667 + ], + "type": "text", + "content": ". Note that " + }, + { + "bbox": [ + 130, + 629, + 482, + 667 + ], + "type": "inline_equation", + "content": "\\epsilon_{\\theta}" + }, + { + "bbox": [ + 130, + 629, + 482, + 667 + ], + "type": "text", + "content": " is typically parameterized by the UNet [40]." + } + ] + } + ], + "index": 7 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "spans": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "type": "text", + "content": "4" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 166, + 91, + 260, + 102 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 166, + 91, + 260, + 102 + ], + "spans": [ + { + "bbox": [ + 166, + 91, + 260, + 102 + ], + "type": "text", + "content": "Lee, Seung Hyun et al." + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 3 + }, + { + "para_blocks": [ + { + "bbox": [ + 130, + 116, + 479, + 224 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 116, + 479, + 224 + ], + "spans": [ + { + "bbox": [ + 130, + 116, + 479, + 224 + ], + "type": "text", + "content": "RL-based T2I Diffusion Model Fine-tuning: Given a reward signal from generated images, the goal of RL-tuning for T2I diffusion models is to optimize the policy defined as one denoising step of T2I diffusion models. In particular, Black et al. [4] apply policy gradient algorithm, which regards the denoising process of diffusion models as a Markov decision process (MDP) by performing multiple denoising steps iteratively. Subsequently, a black box reward model " + }, + { + "bbox": [ + 130, + 116, + 479, + 224 + ], + "type": "inline_equation", + "content": "r(\\cdot ,\\cdot)" + }, + { + "bbox": [ + 130, + 116, + 479, + 224 + ], + "type": "text", + "content": " predicts a single scalar value from sampled image " + }, + { + "bbox": [ + 130, + 116, + 479, + 224 + ], + "type": "inline_equation", + "content": "\\mathbf{x}_0" + }, + { + "bbox": [ + 130, + 116, + 479, + 224 + ], + "type": "text", + "content": ". Given text condition " + }, + { + "bbox": [ + 130, + 116, + 479, + 224 + ], + "type": "inline_equation", + "content": "c\\sim p(c)" + }, + { + "bbox": [ + 130, + 116, + 479, + 224 + ], + "type": "text", + "content": " and image " + }, + { + "bbox": [ + 130, + 116, + 479, + 224 + ], + "type": "inline_equation", + "content": "\\mathbf{x}_0" + }, + { + "bbox": [ + 130, + 116, + 479, + 224 + ], + "type": "text", + "content": ", objective function " + }, + { + "bbox": [ + 130, + 116, + 479, + 224 + ], + "type": "inline_equation", + "content": "\\mathcal{I}" + }, + { + "bbox": [ + 130, + 116, + 479, + 224 + ], + "type": "text", + "content": " can be defined to maximize the expected reward as follows:" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 246, + 231, + 480, + 244 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 246, + 231, + 480, + 244 + ], + "spans": [ + { + "bbox": [ + 246, + 231, + 480, + 244 + ], + "type": "interline_equation", + "content": "\\mathcal {J} _ {\\theta} = \\mathbb {E} _ {p (c)} \\mathbb {E} _ {p _ {\\theta} (\\mathbf {x} _ {0} | c)} [ r (\\mathbf {x} _ {0}, c) ], \\tag {2}", + "image_path": "4fa0b072fea518bd57e86e5e4b221f601a7bd17ea4ebb9b49f007100dcb6685e.jpg" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 130, + 248, + 479, + 297 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 248, + 479, + 297 + ], + "spans": [ + { + "bbox": [ + 130, + 248, + 479, + 297 + ], + "type": "text", + "content": "where the pre-trained diffusion model " + }, + { + "bbox": [ + 130, + 248, + 479, + 297 + ], + "type": "inline_equation", + "content": "p_{\\theta}" + }, + { + "bbox": [ + 130, + 248, + 479, + 297 + ], + "type": "text", + "content": " produces a sample distribution " + }, + { + "bbox": [ + 130, + 248, + 479, + 297 + ], + "type": "inline_equation", + "content": "p_{\\theta}(\\mathbf{x}_0|c)" + }, + { + "bbox": [ + 130, + 248, + 479, + 297 + ], + "type": "text", + "content": " using text condition " + }, + { + "bbox": [ + 130, + 248, + 479, + 297 + ], + "type": "inline_equation", + "content": "c" + }, + { + "bbox": [ + 130, + 248, + 479, + 297 + ], + "type": "text", + "content": ". Modifying this equation, Fan et al. [12] demonstrate that the gradient of objective function " + }, + { + "bbox": [ + 130, + 248, + 479, + 297 + ], + "type": "inline_equation", + "content": "\\nabla \\mathcal{I}_{\\theta}" + }, + { + "bbox": [ + 130, + 248, + 479, + 297 + ], + "type": "text", + "content": " can be calculated through gradient ascent algorithm without using the gradient of reward model as follows:" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 211, + 302, + 480, + 334 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 211, + 302, + 480, + 334 + ], + "spans": [ + { + "bbox": [ + 211, + 302, + 480, + 334 + ], + "type": "interline_equation", + "content": "\\nabla \\mathcal {J} _ {\\theta} = \\mathbb {E} \\left[ r \\left(\\mathbf {x} _ {0}, c\\right) \\sum_ {t = 1} ^ {T} \\nabla_ {\\theta} \\log p _ {\\theta} \\left(\\mathbf {x} _ {t - 1} \\mid c, t, \\mathbf {x} _ {t}\\right) \\right], \\tag {3}", + "image_path": "f38818f7de0250de4373ef2dafe2db833b6c7f1a027b877382460a73a5bbf433.jpg" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 131, + 338, + 479, + 363 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 131, + 338, + 479, + 363 + ], + "spans": [ + { + "bbox": [ + 131, + 338, + 479, + 363 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 131, + 338, + 479, + 363 + ], + "type": "inline_equation", + "content": "T" + }, + { + "bbox": [ + 131, + 338, + 479, + 363 + ], + "type": "text", + "content": " denotes the total time step of the diffusion sampling. With parameters " + }, + { + "bbox": [ + 131, + 338, + 479, + 363 + ], + "type": "inline_equation", + "content": "\\theta" + }, + { + "bbox": [ + 131, + 338, + 479, + 363 + ], + "type": "text", + "content": ", the expectation value can be taken over the trajectories of diffusion sampling." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 131, + 378, + 202, + 391 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 131, + 378, + 202, + 391 + ], + "spans": [ + { + "bbox": [ + 131, + 378, + 202, + 391 + ], + "type": "text", + "content": "4 Method" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 131, + 401, + 245, + 412 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 131, + 401, + 245, + 412 + ], + "spans": [ + { + "bbox": [ + 131, + 401, + 245, + 412 + ], + "type": "text", + "content": "4.1 Parrot Overview" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 130, + 418, + 482, + 598 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 418, + 482, + 598 + ], + "spans": [ + { + "bbox": [ + 130, + 418, + 482, + 598 + ], + "type": "text", + "content": "Fig. 2 shows the overview of Parrot, which consists of the prompt expansion network (PEN) " + }, + { + "bbox": [ + 130, + 418, + 482, + 598 + ], + "type": "inline_equation", + "content": "p_{\\phi}" + }, + { + "bbox": [ + 130, + 418, + 482, + 598 + ], + "type": "text", + "content": " and the T2I diffusion model " + }, + { + "bbox": [ + 130, + 418, + 482, + 598 + ], + "type": "inline_equation", + "content": "p_{\\theta}" + }, + { + "bbox": [ + 130, + 418, + 482, + 598 + ], + "type": "text", + "content": ". The PEN is first initialized from a supervised fine-tuning checkpoint on demonstrations of prompt expansion pairs, and the T2I model is initialized from pretrained diffusion model. Given the original prompt " + }, + { + "bbox": [ + 130, + 418, + 482, + 598 + ], + "type": "inline_equation", + "content": "c" + }, + { + "bbox": [ + 130, + 418, + 482, + 598 + ], + "type": "text", + "content": ", the PEN generates an expanded prompt " + }, + { + "bbox": [ + 130, + 418, + 482, + 598 + ], + "type": "inline_equation", + "content": "\\hat{c}" + }, + { + "bbox": [ + 130, + 418, + 482, + 598 + ], + "type": "text", + "content": ", and the T2I model generates images based on this expanded prompt. During the multi-reward RL fine-tuning, a batch of " + }, + { + "bbox": [ + 130, + 418, + 482, + 598 + ], + "type": "inline_equation", + "content": "N" + }, + { + "bbox": [ + 130, + 418, + 482, + 598 + ], + "type": "text", + "content": " images is sampled, and multiple quality rewards are calculated for each image, encompassing aspects like text-image alignment, aesthetics, human preference, and image sentiment. Based on these reward scores, Parrot identifies the batch-wise Pareto-optimal set using a non-dominated sorting algorithm. This optimal set of images is then used for joint optimization of the PEN and T2I model parameters through RL policy gradient update. During inference, Parrot leverages both the original prompt and its expansion, striking a balance between maintaining faithfulness to the original prompt and incorporating additional details for higher quality." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 131, + 613, + 344, + 624 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 131, + 613, + 344, + 624 + ], + "spans": [ + { + "bbox": [ + 131, + 613, + 344, + 624 + ], + "type": "text", + "content": "4.2 Batch-wise Pareto-optimal Selection" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 130, + 629, + 481, + 666 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 629, + 481, + 666 + ], + "spans": [ + { + "bbox": [ + 130, + 629, + 481, + 666 + ], + "type": "text", + "content": "Li et al. [30] has demonstrated that using batchwise Pareto-set learning and selecting good samples in a batch can approximate Pareto-optimality across multiple objectives. Backed up by this theory, we propose to select non-dominated" + } + ] + } + ], + "index": 11 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 182, + 91, + 448, + 102 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 182, + 91, + 448, + 102 + ], + "spans": [ + { + "bbox": [ + 182, + 91, + 448, + 102 + ], + "type": "text", + "content": "Pareto-optimal Multi-Reward RL Framework for T2I Generation" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 474, + 92, + 480, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 474, + 92, + 480, + 100 + ], + "spans": [ + { + "bbox": [ + 474, + 92, + 480, + 100 + ], + "type": "text", + "content": "5" + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 4 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 134, + 114, + 482, + 225 + ], + "blocks": [ + { + "bbox": [ + 134, + 114, + 482, + 225 + ], + "lines": [ + { + "bbox": [ + 134, + 114, + 482, + 225 + ], + "spans": [ + { + "bbox": [ + 134, + 114, + 482, + 225 + ], + "type": "image", + "image_path": "108ff289898dda4a663cb1763b479d7e4373fcb644164b567d72c8c0a3ddbd91.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 130, + 228, + 482, + 306 + ], + "lines": [ + { + "bbox": [ + 130, + 228, + 482, + 306 + ], + "spans": [ + { + "bbox": [ + 130, + 228, + 482, + 306 + ], + "type": "text", + "content": "Fig. 2: Overview of Parrot. During the training, " + }, + { + "bbox": [ + 130, + 228, + 482, + 306 + ], + "type": "inline_equation", + "content": "N" + }, + { + "bbox": [ + 130, + 228, + 482, + 306 + ], + "type": "text", + "content": " images are sampled from the T2I model using the expanded prompt from the prompt expansion network. Multiple quality rewards are calculated for each image, and the Pareto-optimal set is identified using the non-dominated sorting algorithm. These optimal images are then used to perform policy gradient update of the parameters of T2I model and prompt expansion network jointly. During the inference, both the original prompt and the expanded prompt are provided to the T2I model, enabling better faithfulness while adding detail." + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_caption" + } + ], + "index": 2 + }, + { + "bbox": [ + 130, + 329, + 482, + 498 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 329, + 482, + 498 + ], + "spans": [ + { + "bbox": [ + 130, + 329, + 482, + 498 + ], + "type": "text", + "content": "points in a batch for policy gradient update in RL to achieve Pareto-optimality. Algorithm 1 outlines the procedure of Parrot. Rather than updating the gradients using all images, Parrot focuses on high-quality samples, considering multiple quality rewards in each mini-batch. In the multi-reward RL, each sample generated by the T2I model presents distinct trade-offs for different rewards. Among these samples, a subset with varied optimal trade-offs across multiple objectives, also known as the Pareto set, exists. For a Pareto-optimal sample, none of its objective values can be further improved without damaging others. In other words, the Pareto-optimal set is not dominated by any data points, also known as the non-dominated set. To achieve a Pareto-optimal solution with text-to-image generation diffusion model, Parrot selectively uses data points from the non-dominated set using non-dominated sorting algorithm. This naturally encourages the T2I model to produce Pareto-optimal samples with respect to the multi-reward objectives." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 130, + 503, + 483, + 612 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 503, + 483, + 612 + ], + "spans": [ + { + "bbox": [ + 130, + 503, + 483, + 612 + ], + "type": "text", + "content": "Reward-specific Preference: Inspired by the use of preference information in multi-objective optimization [30], Parrot incorporates the preference information through reward-specific identifiers. This enables Parrot to automatically determine the importance for each reward objective. Concretely, we enrich the expanded prompt " + }, + { + "bbox": [ + 130, + 503, + 483, + 612 + ], + "type": "inline_equation", + "content": "\\hat{c}" + }, + { + "bbox": [ + 130, + 503, + 483, + 612 + ], + "type": "text", + "content": " by preponding reward-specific identifier “" + }, + { + "bbox": [ + 130, + 503, + 483, + 612 + ], + "type": "text", + "content": "” for " + }, + { + "bbox": [ + 130, + 503, + 483, + 612 + ], + "type": "inline_equation", + "content": "k" + }, + { + "bbox": [ + 130, + 503, + 483, + 612 + ], + "type": "text", + "content": "-th reward. Based on this reward-specific prompt, " + }, + { + "bbox": [ + 130, + 503, + 483, + 612 + ], + "type": "inline_equation", + "content": "N" + }, + { + "bbox": [ + 130, + 503, + 483, + 612 + ], + "type": "text", + "content": " images are generated and are used for maximizing the corresponding " + }, + { + "bbox": [ + 130, + 503, + 483, + 612 + ], + "type": "inline_equation", + "content": "k" + }, + { + "bbox": [ + 130, + 503, + 483, + 612 + ], + "type": "text", + "content": "-th reward model during gradient update. At inference time, a concatenation of all the reward identifiers “" + }, + { + "bbox": [ + 130, + 503, + 483, + 612 + ], + "type": "text", + "content": ", ..., " + }, + { + "bbox": [ + 130, + 503, + 483, + 612 + ], + "type": "text", + "content": "” is used for image generation." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 130, + 617, + 482, + 667 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 617, + 482, + 667 + ], + "spans": [ + { + "bbox": [ + 130, + 617, + 482, + 667 + ], + "type": "text", + "content": "Non-dominated Sorting: Parrot constructs Pareto set with non-dominated points based on trade-offs among multiple rewards. These non-dominated points are superior to the remaining solutions and are not dominated by each other. Formally, the dominance relationship is defined as follows: the image " + }, + { + "bbox": [ + 130, + 617, + 482, + 667 + ], + "type": "inline_equation", + "content": "\\mathbf{x}_0^a" + }, + { + "bbox": [ + 130, + 617, + 482, + 667 + ], + "type": "text", + "content": " domi" + } + ] + } + ], + "index": 6 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "spans": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "type": "text", + "content": "6" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 166, + 91, + 260, + 102 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 166, + 91, + 260, + 102 + ], + "spans": [ + { + "bbox": [ + 166, + 91, + 260, + 102 + ], + "type": "text", + "content": "Lee, Seung Hyun et al." + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 5 + }, + { + "para_blocks": [ + { + "bbox": [ + 132, + 115, + 321, + 125 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 115, + 321, + 125 + ], + "spans": [ + { + "bbox": [ + 132, + 115, + 321, + 125 + ], + "type": "text", + "content": "Algorithm Parrot: Pareto-optimal Multi-Reward RL" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 132, + 126, + 479, + 143 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 126, + 479, + 143 + ], + "spans": [ + { + "bbox": [ + 132, + 126, + 479, + 143 + ], + "type": "text", + "content": "Input: Prompt " + }, + { + "bbox": [ + 132, + 126, + 479, + 143 + ], + "type": "inline_equation", + "content": "c" + }, + { + "bbox": [ + 132, + 126, + 479, + 143 + ], + "type": "text", + "content": ", Batch size " + }, + { + "bbox": [ + 132, + 126, + 479, + 143 + ], + "type": "inline_equation", + "content": "N" + }, + { + "bbox": [ + 132, + 126, + 479, + 143 + ], + "type": "text", + "content": ", Total iteration " + }, + { + "bbox": [ + 132, + 126, + 479, + 143 + ], + "type": "inline_equation", + "content": "E" + }, + { + "bbox": [ + 132, + 126, + 479, + 143 + ], + "type": "text", + "content": ", the number of rewards: " + }, + { + "bbox": [ + 132, + 126, + 479, + 143 + ], + "type": "inline_equation", + "content": "K" + }, + { + "bbox": [ + 132, + 126, + 479, + 143 + ], + "type": "text", + "content": ", Prompt expansion network " + }, + { + "bbox": [ + 132, + 126, + 479, + 143 + ], + "type": "inline_equation", + "content": "p_{\\phi}" + }, + { + "bbox": [ + 132, + 126, + 479, + 143 + ], + "type": "text", + "content": ", T2I diffusion model: " + }, + { + "bbox": [ + 132, + 126, + 479, + 143 + ], + "type": "inline_equation", + "content": "p_{\\theta}" + }, + { + "bbox": [ + 132, + 126, + 479, + 143 + ], + "type": "text", + "content": ", Total diffusion time step " + }, + { + "bbox": [ + 132, + 126, + 479, + 143 + ], + "type": "inline_equation", + "content": "T" + }, + { + "bbox": [ + 132, + 126, + 479, + 143 + ], + "type": "text", + "content": ", Non-dominated set: " + }, + { + "bbox": [ + 132, + 126, + 479, + 143 + ], + "type": "inline_equation", + "content": "\\mathcal{P}" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 140, + 143, + 206, + 151 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 140, + 143, + 206, + 151 + ], + "spans": [ + { + "bbox": [ + 140, + 143, + 206, + 151 + ], + "type": "text", + "content": "for " + }, + { + "bbox": [ + 140, + 143, + 206, + 151 + ], + "type": "inline_equation", + "content": "e = 1" + }, + { + "bbox": [ + 140, + 143, + 206, + 151 + ], + "type": "text", + "content": " to " + }, + { + "bbox": [ + 140, + 143, + 206, + 151 + ], + "type": "inline_equation", + "content": "E" + }, + { + "bbox": [ + 140, + 143, + 206, + 151 + ], + "type": "text", + "content": " do" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 153, + 151, + 255, + 159 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 153, + 151, + 255, + 159 + ], + "spans": [ + { + "bbox": [ + 153, + 151, + 255, + 159 + ], + "type": "text", + "content": "Sample text prompt " + }, + { + "bbox": [ + 153, + 151, + 255, + 159 + ], + "type": "inline_equation", + "content": "c \\sim p(c)" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 154, + 159, + 219, + 166 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 154, + 159, + 219, + 166 + ], + "spans": [ + { + "bbox": [ + 154, + 159, + 219, + 166 + ], + "type": "text", + "content": "for " + }, + { + "bbox": [ + 154, + 159, + 219, + 166 + ], + "type": "inline_equation", + "content": "k = 1" + }, + { + "bbox": [ + 154, + 159, + 219, + 166 + ], + "type": "text", + "content": " to " + }, + { + "bbox": [ + 154, + 159, + 219, + 166 + ], + "type": "inline_equation", + "content": "K" + }, + { + "bbox": [ + 154, + 159, + 219, + 166 + ], + "type": "text", + "content": " do" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 165, + 167, + 279, + 175 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 165, + 167, + 279, + 175 + ], + "spans": [ + { + "bbox": [ + 165, + 167, + 279, + 175 + ], + "type": "text", + "content": "Expand text prompt " + }, + { + "bbox": [ + 165, + 167, + 279, + 175 + ], + "type": "inline_equation", + "content": "\\hat{c} \\sim p_{\\phi}(\\hat{c} | c)" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 165, + 175, + 342, + 182 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 165, + 175, + 342, + 182 + ], + "spans": [ + { + "bbox": [ + 165, + 175, + 342, + 182 + ], + "type": "text", + "content": "Prepend reward-specific tokens " + }, + { + "bbox": [ + 165, + 175, + 342, + 182 + ], + "type": "inline_equation", + "content": "<" + }, + { + "bbox": [ + 165, + 175, + 342, + 182 + ], + "type": "text", + "content": " reward " + }, + { + "bbox": [ + 165, + 175, + 342, + 182 + ], + "type": "inline_equation", + "content": "k>" + }, + { + "bbox": [ + 165, + 175, + 342, + 182 + ], + "type": "text", + "content": " to " + }, + { + "bbox": [ + 165, + 175, + 342, + 182 + ], + "type": "inline_equation", + "content": "\\hat{c}" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 165, + 182, + 331, + 191 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 165, + 182, + 331, + 191 + ], + "spans": [ + { + "bbox": [ + 165, + 182, + 331, + 191 + ], + "type": "text", + "content": "Sample a set of images " + }, + { + "bbox": [ + 165, + 182, + 331, + 191 + ], + "type": "inline_equation", + "content": "\\{\\mathbf{x}_0^1,\\dots ,\\mathbf{x}_0^N\\} \\sim p_\\theta (x_0|\\hat{c})" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 165, + 191, + 309, + 199 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 165, + 191, + 309, + 199 + ], + "spans": [ + { + "bbox": [ + 165, + 191, + 309, + 199 + ], + "type": "text", + "content": "A set of reward vector " + }, + { + "bbox": [ + 165, + 191, + 309, + 199 + ], + "type": "inline_equation", + "content": "\\mathcal{R} = \\{R_1,\\dots,R_N\\}" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 165, + 199, + 260, + 209 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 165, + 199, + 260, + 209 + ], + "spans": [ + { + "bbox": [ + 165, + 199, + 260, + 209 + ], + "type": "inline_equation", + "content": "\\mathcal{P}\\gets \\mathrm{NDSET}(\\{\\mathbf{x}_0^1,\\dots,\\mathbf{x}_0^N\\})" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 165, + 209, + 290, + 219 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 165, + 209, + 290, + 219 + ], + "spans": [ + { + "bbox": [ + 165, + 209, + 290, + 219 + ], + "type": "inline_equation", + "content": "\\nabla \\mathcal{J}_{\\phi} + = -r_k(\\mathbf{x}_0^j,\\hat{c})\\times \\nabla \\log p_{\\phi}(\\hat{c} |c)" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 153, + 219, + 275, + 228 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 153, + 219, + 275, + 228 + ], + "spans": [ + { + "bbox": [ + 153, + 219, + 275, + 228 + ], + "type": "text", + "content": "Update the gradient " + }, + { + "bbox": [ + 153, + 219, + 275, + 228 + ], + "type": "inline_equation", + "content": "p_{\\theta}" + }, + { + "bbox": [ + 153, + 219, + 275, + 228 + ], + "type": "text", + "content": " from Eq. 4" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 153, + 228, + 236, + 236 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 153, + 228, + 236, + 236 + ], + "spans": [ + { + "bbox": [ + 153, + 228, + 236, + 236 + ], + "type": "text", + "content": "Update the gradient " + }, + { + "bbox": [ + 153, + 228, + 236, + 236 + ], + "type": "inline_equation", + "content": "p_{\\phi}" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 133, + 236, + 382, + 244 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 236, + 382, + 244 + ], + "spans": [ + { + "bbox": [ + 133, + 236, + 382, + 244 + ], + "type": "text", + "content": "Output: Fine-tuned diffusion model " + }, + { + "bbox": [ + 133, + 236, + 382, + 244 + ], + "type": "inline_equation", + "content": "p_{\\theta}" + }, + { + "bbox": [ + 133, + 236, + 382, + 244 + ], + "type": "text", + "content": ", prompt expansion network " + }, + { + "bbox": [ + 133, + 236, + 382, + 244 + ], + "type": "inline_equation", + "content": "p_{\\phi}" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 349, + 148, + 460, + 158 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 349, + 148, + 460, + 158 + ], + "spans": [ + { + "bbox": [ + 349, + 148, + 460, + 158 + ], + "type": "text", + "content": "function " + }, + { + "bbox": [ + 349, + 148, + 460, + 158 + ], + "type": "inline_equation", + "content": "\\mathrm{NDSET}(\\{\\mathbf{x}_0^1,\\dots,\\mathbf{x}_0^N\\})" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 361, + 158, + 385, + 166 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 361, + 158, + 385, + 166 + ], + "spans": [ + { + "bbox": [ + 361, + 158, + 385, + 166 + ], + "type": "inline_equation", + "content": "\\mathcal{P}\\gets \\emptyset" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 361, + 167, + 425, + 175 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 361, + 167, + 425, + 175 + ], + "spans": [ + { + "bbox": [ + 361, + 167, + 425, + 175 + ], + "type": "text", + "content": "for " + }, + { + "bbox": [ + 361, + 167, + 425, + 175 + ], + "type": "inline_equation", + "content": "i = 1" + }, + { + "bbox": [ + 361, + 167, + 425, + 175 + ], + "type": "text", + "content": " to " + }, + { + "bbox": [ + 361, + 167, + 425, + 175 + ], + "type": "inline_equation", + "content": "N" + }, + { + "bbox": [ + 361, + 167, + 425, + 175 + ], + "type": "text", + "content": " do" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 373, + 175, + 440, + 182 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 373, + 175, + 440, + 182 + ], + "spans": [ + { + "bbox": [ + 373, + 175, + 440, + 182 + ], + "type": "text", + "content": "dominance " + }, + { + "bbox": [ + 373, + 175, + 440, + 182 + ], + "type": "inline_equation", + "content": "\\leftarrow" + }, + { + "bbox": [ + 373, + 175, + 440, + 182 + ], + "type": "text", + "content": " True" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 372, + 182, + 438, + 190 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 372, + 182, + 438, + 190 + ], + "spans": [ + { + "bbox": [ + 372, + 182, + 438, + 190 + ], + "type": "text", + "content": "for " + }, + { + "bbox": [ + 372, + 182, + 438, + 190 + ], + "type": "inline_equation", + "content": "j = 1" + }, + { + "bbox": [ + 372, + 182, + 438, + 190 + ], + "type": "text", + "content": " to " + }, + { + "bbox": [ + 372, + 182, + 438, + 190 + ], + "type": "inline_equation", + "content": "N" + }, + { + "bbox": [ + 372, + 182, + 438, + 190 + ], + "type": "text", + "content": " do" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 384, + 190, + 473, + 199 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 384, + 190, + 473, + 199 + ], + "spans": [ + { + "bbox": [ + 384, + 190, + 473, + 199 + ], + "type": "text", + "content": "if " + }, + { + "bbox": [ + 384, + 190, + 473, + 199 + ], + "type": "inline_equation", + "content": "\\mathbf{x}_0^j" + }, + { + "bbox": [ + 384, + 190, + 473, + 199 + ], + "type": "text", + "content": " dominates " + }, + { + "bbox": [ + 384, + 190, + 473, + 199 + ], + "type": "inline_equation", + "content": "\\mathbf{x}_0^i" + }, + { + "bbox": [ + 384, + 190, + 473, + 199 + ], + "type": "text", + "content": " then" + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 384, + 199, + 465, + 206 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 384, + 199, + 465, + 206 + ], + "spans": [ + { + "bbox": [ + 384, + 199, + 465, + 206 + ], + "type": "text", + "content": "dominance " + }, + { + "bbox": [ + 384, + 199, + 465, + 206 + ], + "type": "inline_equation", + "content": "\\leftarrow" + }, + { + "bbox": [ + 384, + 199, + 465, + 206 + ], + "type": "text", + "content": " False" + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 372, + 207, + 466, + 215 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 372, + 207, + 466, + 215 + ], + "spans": [ + { + "bbox": [ + 372, + 207, + 466, + 215 + ], + "type": "text", + "content": "if dominance is True then" + } + ] + } + ], + "index": 23 + }, + { + "bbox": [ + 384, + 215, + 425, + 222 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 384, + 215, + 425, + 222 + ], + "spans": [ + { + "bbox": [ + 384, + 215, + 425, + 222 + ], + "type": "text", + "content": "Add " + }, + { + "bbox": [ + 384, + 215, + 425, + 222 + ], + "type": "inline_equation", + "content": "i" + }, + { + "bbox": [ + 384, + 215, + 425, + 222 + ], + "type": "text", + "content": " to " + }, + { + "bbox": [ + 384, + 215, + 425, + 222 + ], + "type": "inline_equation", + "content": "\\mathcal{P}" + } + ] + } + ], + "index": 24 + }, + { + "bbox": [ + 361, + 223, + 396, + 230 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 361, + 223, + 396, + 230 + ], + "spans": [ + { + "bbox": [ + 361, + 223, + 396, + 230 + ], + "type": "text", + "content": "return " + }, + { + "bbox": [ + 361, + 223, + 396, + 230 + ], + "type": "inline_equation", + "content": "\\mathcal{P}" + } + ] + } + ], + "index": 25 + }, + { + "bbox": [ + 132, + 268, + 479, + 317 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 268, + 479, + 317 + ], + "spans": [ + { + "bbox": [ + 132, + 268, + 479, + 317 + ], + "type": "text", + "content": "nates the image " + }, + { + "bbox": [ + 132, + 268, + 479, + 317 + ], + "type": "inline_equation", + "content": "\\mathbf{x}_0^b" + }, + { + "bbox": [ + 132, + 268, + 479, + 317 + ], + "type": "text", + "content": ", denoted as " + }, + { + "bbox": [ + 132, + 268, + 479, + 317 + ], + "type": "inline_equation", + "content": "\\mathbf{x}_0^b < \\mathbf{x}_0^a" + }, + { + "bbox": [ + 132, + 268, + 479, + 317 + ], + "type": "text", + "content": ", if and only if " + }, + { + "bbox": [ + 132, + 268, + 479, + 317 + ], + "type": "inline_equation", + "content": "R_{i}(\\mathbf{x}_{0}^{b})\\leq R_{i}(\\mathbf{x}_{0}^{a})" + }, + { + "bbox": [ + 132, + 268, + 479, + 317 + ], + "type": "text", + "content": " for all " + }, + { + "bbox": [ + 132, + 268, + 479, + 317 + ], + "type": "inline_equation", + "content": "i\\in 1,2,\\ldots ,m" + }, + { + "bbox": [ + 132, + 268, + 479, + 317 + ], + "type": "text", + "content": ", and there exists " + }, + { + "bbox": [ + 132, + 268, + 479, + 317 + ], + "type": "inline_equation", + "content": "j\\in 1,2,\\dots,m" + }, + { + "bbox": [ + 132, + 268, + 479, + 317 + ], + "type": "text", + "content": " such that " + }, + { + "bbox": [ + 132, + 268, + 479, + 317 + ], + "type": "inline_equation", + "content": "R_{j}(\\mathbf{x}_{0}^{b}) < R_{j}(\\mathbf{x}_{0}^{a})" + }, + { + "bbox": [ + 132, + 268, + 479, + 317 + ], + "type": "text", + "content": ". For example, given the " + }, + { + "bbox": [ + 132, + 268, + 479, + 317 + ], + "type": "inline_equation", + "content": "i" + }, + { + "bbox": [ + 132, + 268, + 479, + 317 + ], + "type": "text", + "content": "-th generated image " + }, + { + "bbox": [ + 132, + 268, + 479, + 317 + ], + "type": "inline_equation", + "content": "\\mathbf{x}_0^i" + }, + { + "bbox": [ + 132, + 268, + 479, + 317 + ], + "type": "text", + "content": " in a mini-batch, when no point in the mini-batch dominates " + }, + { + "bbox": [ + 132, + 268, + 479, + 317 + ], + "type": "inline_equation", + "content": "\\mathbf{x}_0^i" + }, + { + "bbox": [ + 132, + 268, + 479, + 317 + ], + "type": "text", + "content": ", it is referred to as a non-dominated point." + } + ] + } + ], + "index": 26 + }, + { + "bbox": [ + 131, + 323, + 479, + 358 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 131, + 323, + 479, + 358 + ], + "spans": [ + { + "bbox": [ + 131, + 323, + 479, + 358 + ], + "type": "text", + "content": "Policy Gradient Update: We assign a reward value of zero to the data points not included in non-dominated sets and only update the gradient of these non-dominated data points as follows:" + } + ] + } + ], + "index": 27 + }, + { + "bbox": [ + 162, + 365, + 480, + 401 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 162, + 365, + 480, + 401 + ], + "spans": [ + { + "bbox": [ + 162, + 365, + 480, + 401 + ], + "type": "interline_equation", + "content": "\\nabla \\mathcal {J} _ {\\theta} = \\sum_ {k = 1} ^ {K} \\frac {1}{n (\\mathcal {P})} \\sum_ {i = 1, \\mathbf {x} _ {0} ^ {i} \\in \\mathcal {P}} ^ {N} \\sum_ {t = 1} ^ {T} r _ {k} \\left(\\mathbf {x} _ {0} ^ {i}, c _ {k}\\right) \\times \\nabla_ {\\theta} \\log p _ {\\theta} \\left(\\mathbf {x} _ {t - 1} ^ {i} \\mid c _ {k}, t, \\mathbf {x} _ {t} ^ {i}\\right), \\tag {4}", + "image_path": "6c78b61ced8f5e7dc7fef88813546680cc551315ba85f4d25a728e1a9eec9656.jpg" + } + ] + } + ], + "index": 28 + }, + { + "bbox": [ + 131, + 408, + 479, + 456 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 131, + 408, + 479, + 456 + ], + "spans": [ + { + "bbox": [ + 131, + 408, + 479, + 456 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 131, + 408, + 479, + 456 + ], + "type": "inline_equation", + "content": "i" + }, + { + "bbox": [ + 131, + 408, + 479, + 456 + ], + "type": "text", + "content": " indicates the index of images in mini-batches, and " + }, + { + "bbox": [ + 131, + 408, + 479, + 456 + ], + "type": "inline_equation", + "content": "\\mathcal{P}" + }, + { + "bbox": [ + 131, + 408, + 479, + 456 + ], + "type": "text", + "content": " denotes batch-wise a set of non-dominated points. " + }, + { + "bbox": [ + 131, + 408, + 479, + 456 + ], + "type": "inline_equation", + "content": "K" + }, + { + "bbox": [ + 131, + 408, + 479, + 456 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 131, + 408, + 479, + 456 + ], + "type": "inline_equation", + "content": "T" + }, + { + "bbox": [ + 131, + 408, + 479, + 456 + ], + "type": "text", + "content": " are the total number of reward models and total diffusion time steps, respectively. The same text prompt is used when updating the diffusion model in each batch." + } + ] + } + ], + "index": 29 + }, + { + "bbox": [ + 132, + 473, + 344, + 486 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 473, + 344, + 486 + ], + "spans": [ + { + "bbox": [ + 132, + 473, + 344, + 486 + ], + "type": "text", + "content": "4.3 Original Prompt Centered Guidance" + } + ] + } + ], + "index": 30 + }, + { + "bbox": [ + 130, + 492, + 480, + 612 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 492, + 480, + 612 + ], + "spans": [ + { + "bbox": [ + 130, + 492, + 480, + 612 + ], + "type": "text", + "content": "While prompt expansion enhances details and often improves generation quality, there is a concern that the added context may dilute the main content of the original input. To mitigate this during the inference, we introduce original prompt-centered guidance. When sampling conditioned on the original prompt, the diffusion model " + }, + { + "bbox": [ + 130, + 492, + 480, + 612 + ], + "type": "inline_equation", + "content": "\\epsilon_{\\theta}" + }, + { + "bbox": [ + 130, + 492, + 480, + 612 + ], + "type": "text", + "content": " typically predicts noises by combining the unconditioned score estimate and the prompt-conditioned estimate. Instead of relying solely on the expanded prompt from PEN, we propose using a linear combination of two. guidances for T2I generation: one from the user input and the other from the expanded prompt. The strength of the original prompt is controlled by guidance scales " + }, + { + "bbox": [ + 130, + 492, + 480, + 612 + ], + "type": "inline_equation", + "content": "w_{1}" + }, + { + "bbox": [ + 130, + 492, + 480, + 612 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 130, + 492, + 480, + 612 + ], + "type": "inline_equation", + "content": "w_{2}" + }, + { + "bbox": [ + 130, + 492, + 480, + 612 + ], + "type": "text", + "content": ". The noise " + }, + { + "bbox": [ + 130, + 492, + 480, + 612 + ], + "type": "inline_equation", + "content": "\\bar{\\epsilon}_{\\theta}" + }, + { + "bbox": [ + 130, + 492, + 480, + 612 + ], + "type": "text", + "content": " is estimated, derived from Eq. 1, as follows:" + } + ] + } + ], + "index": 31 + }, + { + "bbox": [ + 157, + 632, + 480, + 645 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 157, + 632, + 480, + 645 + ], + "spans": [ + { + "bbox": [ + 157, + 632, + 480, + 645 + ], + "type": "interline_equation", + "content": "\\bar {\\epsilon} _ {\\theta} = w _ {1} \\cdot \\epsilon_ {\\theta} (\\mathbf {x} _ {t}, t, c) + (1 - w _ {1} - w _ {2}) \\cdot \\epsilon_ {\\theta} (\\mathbf {x} _ {t}, t, \\text {n u l l}) + w _ {2} \\cdot \\epsilon_ {\\theta} (\\mathbf {x} _ {t}, t, \\hat {c}), \\tag {5}", + "image_path": "2e6093e52b390f974650552761691b42cbd0415f04d75cd42480fc6d4a9c7b54.jpg" + } + ] + } + ], + "index": 32 + }, + { + "bbox": [ + 132, + 653, + 266, + 664 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 653, + 266, + 664 + ], + "spans": [ + { + "bbox": [ + 132, + 653, + 266, + 664 + ], + "type": "text", + "content": "where null denotes a null text." + } + ] + } + ], + "index": 33 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 182, + 91, + 447, + 101 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 182, + 91, + 447, + 101 + ], + "spans": [ + { + "bbox": [ + 182, + 91, + 447, + 101 + ], + "type": "text", + "content": "Pareto-optimal Multi-Reward RL Framework for T2I Generation" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 474, + 92, + 480, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 474, + 92, + 480, + 100 + ], + "spans": [ + { + "bbox": [ + 474, + 92, + 480, + 100 + ], + "type": "text", + "content": "7" + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 6 + }, + { + "para_blocks": [ + { + "bbox": [ + 132, + 114, + 230, + 129 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 114, + 230, + 129 + ], + "spans": [ + { + "bbox": [ + 132, + 114, + 230, + 129 + ], + "type": "text", + "content": "5 Experiments" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 132, + 138, + 261, + 152 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 138, + 261, + 152 + ], + "spans": [ + { + "bbox": [ + 132, + 138, + 261, + 152 + ], + "type": "text", + "content": "5.1 Experiment Setting" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 130, + 163, + 485, + 247 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 163, + 485, + 247 + ], + "spans": [ + { + "bbox": [ + 130, + 163, + 485, + 247 + ], + "type": "text", + "content": "Dataset: The PEN is first supervised fine-tuned on a large-scale text dataset named the Promptist [15], which has 360K constructed prompt pairs for original prompt and prompt expansion demonstration. The original instruction \"Rephrase\" is included per pair in Promptist. We modify the instruction into \"Input: \". This is a text input for image generation. Expand prompt for improving image quality. Output:\". Subsequently, we use the RL tuning prompts (1200K) from Promptist for RL training of the PEN and T2I model." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 130, + 253, + 483, + 338 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 253, + 483, + 338 + ], + "spans": [ + { + "bbox": [ + 130, + 253, + 483, + 338 + ], + "type": "text", + "content": "T2I Model: Our T2I model is based on the JAX version of Stable Diffusion 1.5 [39] pre-trained with the LAION-5B [43] dataset. We conduct experiments on a machine equipped with 16 NVIDIA RTX A100 GPUs. DDIM [46] with 50 denoising steps is used, and the classifier-free guidance weight is set to 5.0 with the resolution " + }, + { + "bbox": [ + 130, + 253, + 483, + 338 + ], + "type": "inline_equation", + "content": "512 \\times 512" + }, + { + "bbox": [ + 130, + 253, + 483, + 338 + ], + "type": "text", + "content": ". Instead of updating all layers, we specifically update the cross-attention layer in the Denoising U-Net. For optimization, we employ the Adam [25] optimizer with a learning rate of " + }, + { + "bbox": [ + 130, + 253, + 483, + 338 + ], + "type": "inline_equation", + "content": "1 \\times 10^{-5}" + }, + { + "bbox": [ + 130, + 253, + 483, + 338 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 130, + 342, + 483, + 416 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 342, + 483, + 416 + ], + "spans": [ + { + "bbox": [ + 130, + 342, + 483, + 416 + ], + "type": "text", + "content": "Prompt Expansion Network: For prompt expansion, we use PaLM 2-L-IT [2], one of the PaLM2 variations, which is a multi-layer Transformer [49] decoder with casual language modeling. We optimize LoRA [20] weights for RL-based fine-tuning. The output token length of the PEN is set to 77 to match the maximum number of token length for Stable Diffusion. For original prompt-centered guidance, we set both " + }, + { + "bbox": [ + 130, + 342, + 483, + 416 + ], + "type": "inline_equation", + "content": "w_{1}" + }, + { + "bbox": [ + 130, + 342, + 483, + 416 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 130, + 342, + 483, + 416 + ], + "type": "inline_equation", + "content": "w_{2}" + }, + { + "bbox": [ + 130, + 342, + 483, + 416 + ], + "type": "text", + "content": " to 5 in Eq. 5." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 130, + 420, + 483, + 541 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 420, + 483, + 541 + ], + "spans": [ + { + "bbox": [ + 130, + 420, + 483, + 541 + ], + "type": "text", + "content": "Reward Models: We incorporate four quality signals as rewards: Aesthetics, Human preference, Text-Image Alignment, Image Sentiment. For aesthetics, we use the VILA-R [24] pre-trained with the AVA [33] dataset. For human preference, we train a ViT-B/16 [10] using the Pick-a-Pic [26] dataset, which contains 500K examples for human feedback in T2I generation. The ViT-B/16 image encoder consists of 12 transformer layers, and the image resolution is " + }, + { + "bbox": [ + 130, + 420, + 483, + 541 + ], + "type": "inline_equation", + "content": "224 \\times 224" + }, + { + "bbox": [ + 130, + 420, + 483, + 541 + ], + "type": "text", + "content": " with a patch size of " + }, + { + "bbox": [ + 130, + 420, + 483, + 541 + ], + "type": "inline_equation", + "content": "16 \\times 16" + }, + { + "bbox": [ + 130, + 420, + 483, + 541 + ], + "type": "text", + "content": ". For text-image alignment, we use CLIP [35] with the image encoder ViT-B/32. For image sentiment, we use the pre-trained model from [44], which outputs three labels: positive, neutral, negative. We use the positive score ranging from 0 to 1 as the sentiment reward." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 132, + 557, + 263, + 570 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 557, + 263, + 570 + ], + "spans": [ + { + "bbox": [ + 132, + 557, + 263, + 570 + ], + "type": "text", + "content": "5.2 Qualitative Analysis" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 130, + 581, + 483, + 667 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 581, + 483, + 667 + ], + "spans": [ + { + "bbox": [ + 130, + 581, + 483, + 667 + ], + "type": "text", + "content": "Comparison with Baselines: Fig. 3 shows the visual comparison of Parrot and multiple baselines. We include results from Stable Diffusion 1.5, DPOK [12] with a weighted sum of rewards, Promptist [15], and Parrot. DPOK exclusively fine-tunes the T2I model, while Promptist focuses on fine-tuning only the prompt expansion network. Parrot shows visually better images, particularly in aspects like color combination, cropping, perspective, and fine details in the image. This improvement can be attributed to Parrot's T2I model being fine-tuned together" + } + ] + } + ], + "index": 9 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "spans": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "type": "text", + "content": "8" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 166, + 91, + 260, + 102 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 166, + 91, + 260, + 102 + ], + "spans": [ + { + "bbox": [ + 166, + 91, + 260, + 102 + ], + "type": "text", + "content": "Lee, Seung Hyun et al." + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 7 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 137, + 129, + 221, + 213 + ], + "blocks": [ + { + "bbox": [ + 164, + 117, + 192, + 126 + ], + "lines": [ + { + "bbox": [ + 164, + 117, + 192, + 126 + ], + "spans": [ + { + "bbox": [ + 164, + 117, + 192, + 126 + ], + "type": "text", + "content": "SD 1.5" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 137, + 129, + 221, + 213 + ], + "lines": [ + { + "bbox": [ + 137, + 129, + 221, + 213 + ], + "spans": [ + { + "bbox": [ + 137, + 129, + 221, + 213 + ], + "type": "image", + "image_path": "ae6232071aac53bcf03124b48b9d00f41b179c8709e99c0bc68d1720258677f9.jpg" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 140, + 215, + 170, + 222 + ], + "lines": [ + { + "bbox": [ + 140, + 215, + 170, + 222 + ], + "spans": [ + { + "bbox": [ + 140, + 215, + 170, + 222 + ], + "type": "text", + "content": "A kitchen" + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "image_caption" + } + ], + "index": 3 + }, + { + "type": "image", + "bbox": [ + 221, + 129, + 306, + 213 + ], + "blocks": [ + { + "bbox": [ + 250, + 117, + 277, + 126 + ], + "lines": [ + { + "bbox": [ + 250, + 117, + 277, + 126 + ], + "spans": [ + { + "bbox": [ + 250, + 117, + 277, + 126 + ], + "type": "text", + "content": "DPOK" + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 221, + 129, + 306, + 213 + ], + "lines": [ + { + "bbox": [ + 221, + 129, + 306, + 213 + ], + "spans": [ + { + "bbox": [ + 221, + 129, + 306, + 213 + ], + "type": "image", + "image_path": "d4748c79abd0214fde03e30f2695b294f6fff77162411f74f2cde462ce33828e.jpg" + } + ] + } + ], + "index": 6, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 222, + 215, + 252, + 222 + ], + "lines": [ + { + "bbox": [ + 222, + 215, + 252, + 222 + ], + "spans": [ + { + "bbox": [ + 222, + 215, + 252, + 222 + ], + "type": "text", + "content": "A kitchen" + } + ] + } + ], + "index": 7, + "angle": 0, + "type": "image_caption" + } + ], + "index": 6 + }, + { + "type": "image", + "bbox": [ + 306, + 129, + 389, + 213 + ], + "blocks": [ + { + "bbox": [ + 329, + 117, + 367, + 126 + ], + "lines": [ + { + "bbox": [ + 329, + 117, + 367, + 126 + ], + "spans": [ + { + "bbox": [ + 329, + 117, + 367, + 126 + ], + "type": "text", + "content": "Promptist" + } + ] + } + ], + "index": 8, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 306, + 129, + 389, + 213 + ], + "lines": [ + { + "bbox": [ + 306, + 129, + 389, + 213 + ], + "spans": [ + { + "bbox": [ + 306, + 129, + 389, + 213 + ], + "type": "image", + "image_path": "f3f59b1b91f145bfff3becf93703298da87351ead285bf567944c61b0abd7f1c.jpg" + } + ] + } + ], + "index": 9, + "angle": 0, + "type": "image_body" + } + ], + "index": 9 + }, + { + "type": "image", + "bbox": [ + 390, + 129, + 475, + 213 + ], + "blocks": [ + { + "bbox": [ + 420, + 117, + 444, + 125 + ], + "lines": [ + { + "bbox": [ + 420, + 117, + 444, + 125 + ], + "spans": [ + { + "bbox": [ + 420, + 117, + 444, + 125 + ], + "type": "text", + "content": "Parrot" + } + ] + } + ], + "index": 11, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 390, + 129, + 475, + 213 + ], + "lines": [ + { + "bbox": [ + 390, + 129, + 475, + 213 + ], + "spans": [ + { + "bbox": [ + 390, + 129, + 475, + 213 + ], + "type": "image", + "image_path": "be9ba214bada9817db1a1920ade4c7d433e15213e9106f291c204b80436d1da7.jpg" + } + ] + } + ], + "index": 12, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 394, + 215, + 464, + 243 + ], + "lines": [ + { + "bbox": [ + 394, + 215, + 464, + 243 + ], + "spans": [ + { + "bbox": [ + 394, + 215, + 464, + 243 + ], + "type": "text", + "content": "A kitchen with a lot of natural light and a beautiful view of the mountains" + } + ] + } + ], + "index": 13, + "angle": 0, + "type": "image_caption" + } + ], + "index": 12 + }, + { + "type": "image", + "bbox": [ + 137, + 247, + 220, + 331 + ], + "blocks": [ + { + "bbox": [ + 137, + 247, + 220, + 331 + ], + "lines": [ + { + "bbox": [ + 137, + 247, + 220, + 331 + ], + "spans": [ + { + "bbox": [ + 137, + 247, + 220, + 331 + ], + "type": "image", + "image_path": "3a0fb3345062f16eafbb92de81c52a302a4e796892c3fce8e74314e884b0d422.jpg" + } + ] + } + ], + "index": 14, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 140, + 335, + 154, + 341 + ], + "lines": [ + { + "bbox": [ + 140, + 335, + 154, + 341 + ], + "spans": [ + { + "bbox": [ + 140, + 335, + 154, + 341 + ], + "type": "text", + "content": "AF1" + } + ] + } + ], + "index": 15, + "angle": 0, + "type": "image_caption" + } + ], + "index": 14 + }, + { + "type": "image", + "bbox": [ + 221, + 247, + 306, + 331 + ], + "blocks": [ + { + "bbox": [ + 221, + 247, + 306, + 331 + ], + "lines": [ + { + "bbox": [ + 221, + 247, + 306, + 331 + ], + "spans": [ + { + "bbox": [ + 221, + 247, + 306, + 331 + ], + "type": "image", + "image_path": "08b52d6d94c1cebccf191ab1c03df4c01b1de851b4e31987e8cd89d362b5a3b1.jpg" + } + ] + } + ], + "index": 16, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 224, + 335, + 239, + 341 + ], + "lines": [ + { + "bbox": [ + 224, + 335, + 239, + 341 + ], + "spans": [ + { + "bbox": [ + 224, + 335, + 239, + 341 + ], + "type": "text", + "content": "AF1" + } + ] + } + ], + "index": 17, + "angle": 0, + "type": "image_caption" + } + ], + "index": 16 + }, + { + "type": "image", + "bbox": [ + 306, + 247, + 390, + 331 + ], + "blocks": [ + { + "bbox": [ + 310, + 215, + 383, + 246 + ], + "lines": [ + { + "bbox": [ + 310, + 215, + 383, + 246 + ], + "spans": [ + { + "bbox": [ + 310, + 215, + 383, + 246 + ], + "type": "text", + "content": "A kitchen, a fantasy digital painting, \ntrending on Artstation, \nhighly detailed" + } + ] + } + ], + "index": 10, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 306, + 247, + 390, + 331 + ], + "lines": [ + { + "bbox": [ + 306, + 247, + 390, + 331 + ], + "spans": [ + { + "bbox": [ + 306, + 247, + 390, + 331 + ], + "type": "image", + "image_path": "74c59804e71369577864594382c6d779f194095db69ae193d868c6b76f026fec.jpg" + } + ] + } + ], + "index": 18, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 309, + 334, + 374, + 369 + ], + "lines": [ + { + "bbox": [ + 309, + 334, + 374, + 369 + ], + "spans": [ + { + "bbox": [ + 309, + 334, + 374, + 369 + ], + "type": "text", + "content": "F1, concept art oil painting by Jama Jurabaev, extremely detailed, brush hard, artstation" + } + ] + } + ], + "index": 19, + "angle": 0, + "type": "image_caption" + } + ], + "index": 18 + }, + { + "type": "image", + "bbox": [ + 391, + 247, + 475, + 333 + ], + "blocks": [ + { + "bbox": [ + 391, + 247, + 475, + 333 + ], + "lines": [ + { + "bbox": [ + 391, + 247, + 475, + 333 + ], + "spans": [ + { + "bbox": [ + 391, + 247, + 475, + 333 + ], + "type": "image", + "image_path": "ba4568e76894545df4c02602960a8c6246e3f4bf7b50cb1604941114854948cd.jpg" + } + ] + } + ], + "index": 20, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 394, + 334, + 462, + 369 + ], + "lines": [ + { + "bbox": [ + 394, + 334, + 462, + 369 + ], + "spans": [ + { + "bbox": [ + 394, + 334, + 462, + 369 + ], + "type": "text", + "content": "3d rendering of a f1 car in dark and moody environment with a red and orange color scheme." + } + ] + } + ], + "index": 21, + "angle": 0, + "type": "image_caption" + } + ], + "index": 20 + }, + { + "type": "image", + "bbox": [ + 137, + 372, + 220, + 457 + ], + "blocks": [ + { + "bbox": [ + 137, + 372, + 220, + 457 + ], + "lines": [ + { + "bbox": [ + 137, + 372, + 220, + 457 + ], + "spans": [ + { + "bbox": [ + 137, + 372, + 220, + 457 + ], + "type": "image", + "image_path": "dff8bd248f9e50e0e809226043dba6e2436c567deb216e1c7e028e945b2657cf.jpg" + } + ] + } + ], + "index": 22, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 141, + 460, + 204, + 481 + ], + "lines": [ + { + "bbox": [ + 141, + 460, + 204, + 481 + ], + "spans": [ + { + "bbox": [ + 141, + 460, + 204, + 481 + ], + "type": "text", + "content": "A woman is looking out on a winter archipelago." + } + ] + } + ], + "index": 23, + "angle": 0, + "type": "image_caption" + } + ], + "index": 22 + }, + { + "type": "image", + "bbox": [ + 221, + 372, + 306, + 457 + ], + "blocks": [ + { + "bbox": [ + 221, + 372, + 306, + 457 + ], + "lines": [ + { + "bbox": [ + 221, + 372, + 306, + 457 + ], + "spans": [ + { + "bbox": [ + 221, + 372, + 306, + 457 + ], + "type": "image", + "image_path": "14677736b9eff726b97b475dd96c1c6eda8c0ddfa029e8d7b85473d0bef3a859.jpg" + } + ] + } + ], + "index": 24, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 223, + 460, + 285, + 481 + ], + "lines": [ + { + "bbox": [ + 223, + 460, + 285, + 481 + ], + "spans": [ + { + "bbox": [ + 223, + 460, + 285, + 481 + ], + "type": "text", + "content": "A woman is looking out on a winter archipelago." + } + ] + } + ], + "index": 25, + "angle": 0, + "type": "image_caption" + } + ], + "index": 24 + }, + { + "type": "image", + "bbox": [ + 306, + 372, + 389, + 457 + ], + "blocks": [ + { + "bbox": [ + 306, + 372, + 389, + 457 + ], + "lines": [ + { + "bbox": [ + 306, + 372, + 389, + 457 + ], + "spans": [ + { + "bbox": [ + 306, + 372, + 389, + 457 + ], + "type": "image", + "image_path": "4ec405cfaf3fb554d0a36f2eef601d46a48c7d2a86dcda595b57b56131c74d64.jpg" + } + ] + } + ], + "index": 26, + "angle": 0, + "type": "image_body" + } + ], + "index": 26 + }, + { + "type": "image", + "bbox": [ + 390, + 372, + 475, + 457 + ], + "blocks": [ + { + "bbox": [ + 390, + 372, + 475, + 457 + ], + "lines": [ + { + "bbox": [ + 390, + 372, + 475, + 457 + ], + "spans": [ + { + "bbox": [ + 390, + 372, + 475, + 457 + ], + "type": "image", + "image_path": "442e140b2561d2d17d592bbb4ec5673671caa514cce76334205728dbd355721a.jpg" + } + ] + } + ], + "index": 28, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 392, + 460, + 462, + 487 + ], + "lines": [ + { + "bbox": [ + 392, + 460, + 462, + 487 + ], + "spans": [ + { + "bbox": [ + 392, + 460, + 462, + 487 + ], + "type": "text", + "content": "A woman is looking out on winter archipelago. The image is very peaceful and serene." + } + ] + } + ], + "index": 29, + "angle": 0, + "type": "image_caption" + } + ], + "index": 28 + }, + { + "type": "image", + "bbox": [ + 137, + 497, + 219, + 582 + ], + "blocks": [ + { + "bbox": [ + 137, + 497, + 219, + 582 + ], + "lines": [ + { + "bbox": [ + 137, + 497, + 219, + 582 + ], + "spans": [ + { + "bbox": [ + 137, + 497, + 219, + 582 + ], + "type": "image", + "image_path": "a10b619707dca6b91163d7f9d93c02b5aa64594cf03a6c32941b916d65399e68.jpg" + } + ] + } + ], + "index": 30, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 140, + 586, + 202, + 601 + ], + "lines": [ + { + "bbox": [ + 140, + 586, + 202, + 601 + ], + "spans": [ + { + "bbox": [ + 140, + 586, + 202, + 601 + ], + "type": "text", + "content": "A Japanese Porcelain Imari vase" + } + ] + } + ], + "index": 31, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 132, + 638, + 481, + 672 + ], + "lines": [ + { + "bbox": [ + 132, + 638, + 481, + 672 + ], + "spans": [ + { + "bbox": [ + 132, + 638, + 481, + 672 + ], + "type": "text", + "content": "Fig. 3: Comparison of Parrot and diffusion-based RL baselines. From left to right, we provide results of Stable diffusion 1.5 [39] (1st column), DPOK [12] (2nd column) with the weighted sum, Promptist [15] (3rd column), and Parrot (4th column)." + } + ] + } + ], + "index": 38, + "angle": 0, + "type": "image_caption" + } + ], + "index": 30 + }, + { + "type": "image", + "bbox": [ + 221, + 497, + 306, + 582 + ], + "blocks": [ + { + "bbox": [ + 221, + 497, + 306, + 582 + ], + "lines": [ + { + "bbox": [ + 221, + 497, + 306, + 582 + ], + "spans": [ + { + "bbox": [ + 221, + 497, + 306, + 582 + ], + "type": "image", + "image_path": "9d758abedc5dc52c4dcbdb8e4fb459a0f71a232438ea3ae857653039757e8b5c.jpg" + } + ] + } + ], + "index": 32, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 224, + 586, + 287, + 601 + ], + "lines": [ + { + "bbox": [ + 224, + 586, + 287, + 601 + ], + "spans": [ + { + "bbox": [ + 224, + 586, + 287, + 601 + ], + "type": "text", + "content": "A Japanese Porcelain Imari vase" + } + ] + } + ], + "index": 33, + "angle": 0, + "type": "image_caption" + } + ], + "index": 32 + }, + { + "type": "image", + "bbox": [ + 306, + 497, + 390, + 582 + ], + "blocks": [ + { + "bbox": [ + 310, + 460, + 380, + 495 + ], + "lines": [ + { + "bbox": [ + 310, + 460, + 380, + 495 + ], + "spans": [ + { + "bbox": [ + 310, + 460, + 380, + 495 + ], + "type": "text", + "content": "Painting of a woman looking out on a winter archipelago, ultra realistic, concept art, intricate details, eerie" + } + ] + } + ], + "index": 27, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 306, + 497, + 390, + 582 + ], + "lines": [ + { + "bbox": [ + 306, + 497, + 390, + 582 + ], + "spans": [ + { + "bbox": [ + 306, + 497, + 390, + 582 + ], + "type": "image", + "image_path": "d3990557d2afa7becae0faaa66fece9a3117ce00d334735a486402b3ccc58874.jpg" + } + ] + } + ], + "index": 34, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 307, + 584, + 390, + 627 + ], + "lines": [ + { + "bbox": [ + 307, + 584, + 390, + 627 + ], + "spans": [ + { + "bbox": [ + 307, + 584, + 390, + 627 + ], + "type": "text", + "content": "A Japanese Porcelain Imari vase, intricate, elegant, highly detailed, digital painting, artstation, concept art, smooth, sharp focus, illustration." + } + ] + } + ], + "index": 35, + "angle": 0, + "type": "image_caption" + } + ], + "index": 34 + }, + { + "type": "image", + "bbox": [ + 391, + 497, + 476, + 582 + ], + "blocks": [ + { + "bbox": [ + 391, + 497, + 476, + 582 + ], + "lines": [ + { + "bbox": [ + 391, + 497, + 476, + 582 + ], + "spans": [ + { + "bbox": [ + 391, + 497, + 476, + 582 + ], + "type": "image", + "image_path": "a7d7fcd3bfa8c03c77d3b056b2bd5e9ef2d2d94c54a50e0008938bbd07015da0.jpg" + } + ] + } + ], + "index": 36, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 392, + 584, + 466, + 620 + ], + "lines": [ + { + "bbox": [ + 392, + 584, + 466, + 620 + ], + "spans": [ + { + "bbox": [ + 392, + 584, + 466, + 620 + ], + "type": "text", + "content": "A photo of a Japanese Porcelain Imari vase on the ground with a beautiful floral pattern on the front" + } + ] + } + ], + "index": 37, + "angle": 0, + "type": "image_caption" + } + ], + "index": 36 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 182, + 91, + 447, + 102 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 182, + 91, + 447, + 102 + ], + "spans": [ + { + "bbox": [ + 182, + 91, + 447, + 102 + ], + "type": "text", + "content": "Pareto-optimal Multi-Reward RL Framework for T2I Generation" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 474, + 92, + 480, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 474, + 92, + 480, + 100 + ], + "spans": [ + { + "bbox": [ + 474, + 92, + 480, + 100 + ], + "type": "text", + "content": "9" + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 8 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 134, + 113, + 223, + 199 + ], + "blocks": [ + { + "bbox": [ + 134, + 113, + 223, + 199 + ], + "lines": [ + { + "bbox": [ + 134, + 113, + 223, + 199 + ], + "spans": [ + { + "bbox": [ + 134, + 113, + 223, + 199 + ], + "type": "image", + "image_path": "af97f6e4f7cef419bb1ec58ec2b9a03f8d699915b8b980f61b4a77bfbafc55fb.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_body" + } + ], + "index": 2 + }, + { + "type": "image", + "bbox": [ + 228, + 113, + 307, + 201 + ], + "blocks": [ + { + "bbox": [ + 228, + 113, + 307, + 201 + ], + "lines": [ + { + "bbox": [ + 228, + 113, + 307, + 201 + ], + "spans": [ + { + "bbox": [ + 228, + 113, + 307, + 201 + ], + "type": "image", + "image_path": "98afd077dd00f4121c3ecb6c6bbb0f054c14642bf77463f176a2a616a1f3063c.jpg" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 130, + 202, + 482, + 269 + ], + "lines": [ + { + "bbox": [ + 130, + 202, + 482, + 269 + ], + "spans": [ + { + "bbox": [ + 130, + 202, + 482, + 269 + ], + "type": "text", + "content": "Fig.4: Training curve for fine-tuning on weighted sum and Parrot. For weighted sum, WS1 denotes " + }, + { + "bbox": [ + 130, + 202, + 482, + 269 + ], + "type": "inline_equation", + "content": "\\{0.7, 0.1, 0.1, 0.1\\}" + }, + { + "bbox": [ + 130, + 202, + 482, + 269 + ], + "type": "text", + "content": " and WS2 denotes " + }, + { + "bbox": [ + 130, + 202, + 482, + 269 + ], + "type": "inline_equation", + "content": "\\{0.25, 0.25, 0.25, 0.25\\}" + }, + { + "bbox": [ + 130, + 202, + 482, + 269 + ], + "type": "text", + "content": " for aesthetics, human preference, text-image alignment and image sentiment. Using weighted sum leads to decrease in human preference score and image sentiment score despite an improvement in the aesthetic score. In contrast, Parrot exhibits stable increases across all metrics." + } + ] + } + ], + "index": 6, + "angle": 0, + "type": "image_caption" + } + ], + "index": 3 + }, + { + "type": "image", + "bbox": [ + 308, + 113, + 394, + 201 + ], + "blocks": [ + { + "bbox": [ + 308, + 113, + 394, + 201 + ], + "lines": [ + { + "bbox": [ + 308, + 113, + 394, + 201 + ], + "spans": [ + { + "bbox": [ + 308, + 113, + 394, + 201 + ], + "type": "image", + "image_path": "3f71f1be3076a9147b65313e7a6f37d0f4143330c01fb0134da47b1d0a3d73a2.jpg" + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "image_body" + } + ], + "index": 4 + }, + { + "type": "image", + "bbox": [ + 394, + 113, + 479, + 200 + ], + "blocks": [ + { + "bbox": [ + 394, + 113, + 479, + 200 + ], + "lines": [ + { + "bbox": [ + 394, + 113, + 479, + 200 + ], + "spans": [ + { + "bbox": [ + 394, + 113, + 479, + 200 + ], + "type": "image", + "image_path": "8ab51a1ac6692c195c2948ef192e3f0f1a8f5a552fd98abea6a32dc546e2ad23.jpg" + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "image_body" + } + ], + "index": 5 + }, + { + "type": "table", + "bbox": [ + 132, + 302, + 482, + 419 + ], + "blocks": [ + { + "bbox": [ + 132, + 302, + 482, + 419 + ], + "lines": [ + { + "bbox": [ + 132, + 302, + 482, + 419 + ], + "spans": [ + { + "bbox": [ + 132, + 302, + 482, + 419 + ], + "type": "table", + "html": "
ModelQuality Metrics
TIA (↑)Aesth. (↑)HP (↑)Sent. (↑)Average (↑)
SD 1.5 [39]0.23220.57550.19300.30100.3254
DPOK [12] (WS)0.23370.58130.19320.30130.3273 (+0.58%)
Parrot w/o PE0.23550.60340.20090.30180.3354 (+3.07%)
Parrot T2I Model Tuning Only0.25090.70730.33370.30520.3992 (+22.6%)
Promptist [15]0.14490.67830.27590.25180.3377 (+3.77 %)
Parrot with HP Only0.15430.59610.35280.25620.3398 (+4.42 %)
Parrot PEN Tuning Only0.16590.64920.26170.31310.3474 (+6.76 %)
Parrot w/o Joint Optimization0.16610.63080.25660.30840.3404 (+4.60 %)
Parrot w/o ori prompt guidance0.16230.71560.34250.31300.3833 (+17.8 %)
Parrot0.16670.73960.34110.31320.3901 (+19.8 %)
", + "image_path": "d400cd7619e318489381f63fe1bf3444b61a9b562550151e7a47e3cc0b71fbce.jpg" + } + ] + } + ], + "index": 7, + "angle": 0, + "type": "table_body" + } + ], + "index": 7 + }, + { + "bbox": [ + 130, + 420, + 482, + 464 + ], + "lines": [ + { + "bbox": [ + 130, + 420, + 482, + 464 + ], + "spans": [ + { + "bbox": [ + 130, + 420, + 482, + 464 + ], + "type": "text", + "content": "Table 1: Quantitative comparison between Parrot and alternatives on the Parti dataset [53]. Abbreviations: WS - Weighted Sum; PE - Prompt Expansion; TIA - Text-Image Alignment; Aesth. - Aesthetics; HP - Human Preference; Sent. - Image Sentiment. TIA score is measured against the original prompt without expansion." + } + ] + } + ], + "index": 8, + "angle": 0, + "type": "text" + }, + { + "bbox": [ + 130, + 504, + 482, + 541 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 504, + 482, + 541 + ], + "spans": [ + { + "bbox": [ + 130, + 504, + 482, + 541 + ], + "type": "text", + "content": "with the prompt expansion model that incorporates aesthetic keywords during training. Parrot generates results that are more closely aligned with the input prompt, as well as more visually pleasing." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 130, + 557, + 482, + 666 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 557, + 482, + 666 + ], + "spans": [ + { + "bbox": [ + 130, + 557, + 482, + 666 + ], + "type": "text", + "content": "Weighted sum vs. Parrot: Fig. 4 shows the training curve comparison of Parrot and using a linear combination of the reward scores. Each subgraph represents a reward. WS1 and WS2 denote two different weights with multiple reward scores. WS1 places greater emphasis on the aesthetic score, while WS2 adopts balanced weights across aesthetics, human preference, text-image alignment, and image sentiment. Employing the weighted sum of multiple rewards leads to a decrease in the image sentiment score, despite notable enhancements in aesthetics and human preference. In contrast, Parrot consistently exhibits improvement across all metrics." + } + ] + } + ], + "index": 10 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "spans": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "type": "text", + "content": "10" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 166, + 91, + 260, + 102 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 166, + 91, + 260, + 102 + ], + "spans": [ + { + "bbox": [ + 166, + 91, + 260, + 102 + ], + "type": "text", + "content": "Lee, Seung Hyun et al." + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 9 + }, + { + "para_blocks": [ + { + "bbox": [ + 132, + 116, + 281, + 128 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 116, + 281, + 128 + ], + "spans": [ + { + "bbox": [ + 132, + 116, + 281, + 128 + ], + "type": "text", + "content": "5.3 Quantitative Evaluation" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 130, + 141, + 482, + 224 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 141, + 482, + 224 + ], + "spans": [ + { + "bbox": [ + 130, + 141, + 482, + 224 + ], + "type": "text", + "content": "Comparison with Baselines: Table 1 presents our results of the quality score across four quality rewards: text-image alignment score, aesthetic score, human preference score, and emotion score. The first group shows methods without prompt expansion, and the second group compares methods with expansion. The prompt expansion and T2I generation are performed on the PartiPrompts [53]. Using a set of 1632 prompts, we generate 32 images for each text input and calculate the average for each metric." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 130, + 225, + 482, + 320 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 225, + 482, + 320 + ], + "spans": [ + { + "bbox": [ + 130, + 225, + 482, + 320 + ], + "type": "inline_equation", + "content": "w / o" + }, + { + "bbox": [ + 130, + 225, + 482, + 320 + ], + "type": "text", + "content": " PE indicates the generation of images solely based on original prompt without expansion. Note that " + }, + { + "bbox": [ + 130, + 225, + 482, + 320 + ], + "type": "inline_equation", + "content": "w / o" + }, + { + "bbox": [ + 130, + 225, + 482, + 320 + ], + "type": "text", + "content": " PE is not the same as Parrot T2I Model Tuning Only. The former takes a model trained with PEN and removes it during inference, which results in a notable disparity between prompts used in training and testing. The latter trains with the multi-objective RL using only the T2I model, and it shows substantial improvement upon DPOK [12] (weighted sum) with balanced weights of " + }, + { + "bbox": [ + 130, + 225, + 482, + 320 + ], + "type": "inline_equation", + "content": "\\{0.25, 0.25, 0.25, 0.25\\}" + }, + { + "bbox": [ + 130, + 225, + 482, + 320 + ], + "type": "text", + "content": ". This shows that the proposed RL method is indeed effective in multi-objective optimization." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 130, + 320, + 482, + 380 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 320, + 482, + 380 + ], + "spans": [ + { + "bbox": [ + 130, + 320, + 482, + 380 + ], + "type": "text", + "content": "For Promptist [15], we generate prompt expansion from their model. Parrot with HP Only shows solely using a human preference reward leads to decline in others. Parrot PEN Tuning Only shows suboptimal aesthetics and human preference. Parrot w/o Joint Optimization shows suboptimal results than Parrot which demonstrates the necessity of jointly optimizing PEN and T2I models." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 130, + 380, + 482, + 441 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 380, + 482, + 441 + ], + "spans": [ + { + "bbox": [ + 130, + 380, + 482, + 441 + ], + "type": "text", + "content": "Our method outperforms both compared methods in aesthetics, human preference and sentiment scores. The text-image alignment score is measured with the original prompt before expansion for fair comparison. As a result, the group without prompt expansion generally shows a higher text-image alignment score. Parrot shows better text-image alignment in each subgroup." + } + ] + } + ], + "index": 6 + }, + { + "type": "table", + "bbox": [ + 132, + 460, + 482, + 520 + ], + "blocks": [ + { + "bbox": [ + 132, + 460, + 482, + 520 + ], + "lines": [ + { + "bbox": [ + 132, + 460, + 482, + 520 + ], + "spans": [ + { + "bbox": [ + 132, + 460, + 482, + 520 + ], + "type": "table", + "html": "
AreaQuestion
Aesthetics“Which image shows better aesthetics without blurry texture, unnatural focusing, and poor color combination?”
Human Preference“Which generated image do you prefer?”
Text-Image Alignment“Which image is well aligned with the text?”
Image Sentiment“Which image is closer to amusement, excitement, and contentment?”
", + "image_path": "ab58a733a019a0ddbb54e1d279b97fd87bca774fdccae52907d98e4ef7f72144.jpg" + } + ] + } + ], + "index": 7, + "angle": 0, + "type": "table_body" + } + ], + "index": 7 + }, + { + "bbox": [ + 132, + 521, + 482, + 542 + ], + "lines": [ + { + "bbox": [ + 132, + 521, + 482, + 542 + ], + "spans": [ + { + "bbox": [ + 132, + 521, + 482, + 542 + ], + "type": "text", + "content": "Table 2: Questions for user study. For performing user study, we carefully design questions suitable to each quality." + } + ] + } + ], + "index": 8, + "angle": 0, + "type": "text" + }, + { + "bbox": [ + 130, + 581, + 482, + 666 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 581, + 482, + 666 + ], + "spans": [ + { + "bbox": [ + 130, + 581, + 482, + 666 + ], + "type": "text", + "content": "User Study: We conduct a user study using MTurk [1] with generated images from 100 random prompts in the PartiPrompts [53]. Five models are compared: Stable Diffusion v1.5, DPOK [12] with an equal weighted sum, Promptist [15], Parrot without prompt expansion, and Parrot. Each rater is presented with the original prompt (before expansion) and a set of five generated images, with the image order being randomized. Raters are then tasked with selecting the best image from the group, guided by questions outlined in Table 2. Each question" + } + ] + } + ], + "index": 9 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 182, + 90, + 447, + 102 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 182, + 90, + 447, + 102 + ], + "spans": [ + { + "bbox": [ + 182, + 90, + 447, + 102 + ], + "type": "text", + "content": "Pareto-optimal Multi-Reward RL Framework for T2I Generation" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 470, + 91, + 479, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 470, + 91, + 479, + 100 + ], + "spans": [ + { + "bbox": [ + 470, + 91, + 479, + 100 + ], + "type": "text", + "content": "11" + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 10 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 139, + 118, + 292, + 232 + ], + "blocks": [ + { + "bbox": [ + 139, + 118, + 292, + 232 + ], + "lines": [ + { + "bbox": [ + 139, + 118, + 292, + 232 + ], + "spans": [ + { + "bbox": [ + 139, + 118, + 292, + 232 + ], + "type": "image", + "image_path": "50643f67ee5cbb0900b727734672b7bff184750fe4d7005f80ec79720216fc4e.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 137, + 239, + 296, + 271 + ], + "lines": [ + { + "bbox": [ + 137, + 239, + 296, + 271 + ], + "spans": [ + { + "bbox": [ + 137, + 239, + 296, + 271 + ], + "type": "text", + "content": "Fig.5: User study results on PartiPrompts [53]. Parrot outperforms baselines across all metrics." + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_caption" + } + ], + "index": 2 + }, + { + "type": "image", + "bbox": [ + 301, + 114, + 475, + 225 + ], + "blocks": [ + { + "bbox": [ + 301, + 114, + 475, + 225 + ], + "lines": [ + { + "bbox": [ + 301, + 114, + 475, + 225 + ], + "spans": [ + { + "bbox": [ + 301, + 114, + 475, + 225 + ], + "type": "image", + "image_path": "60c7bce9ff211310d4c2458526c643ff71c292892d6d409969b66f5973f634a7.jpg" + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "image_body" + } + ], + "index": 4 + }, + { + "type": "image", + "bbox": [ + 133, + 293, + 480, + 416 + ], + "blocks": [ + { + "bbox": [ + 299, + 227, + 476, + 293 + ], + "lines": [ + { + "bbox": [ + 299, + 227, + 476, + 293 + ], + "spans": [ + { + "bbox": [ + 299, + 227, + 476, + 293 + ], + "type": "text", + "content": "Fig. 6: Ablation study. We perform an ablation study by removing one of quality signals. We observe that each quality signal affects their improvement of (a) aesthetics, (b) human preference score, (c) text-image alignment score, (d) image sentiment score." + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 133, + 293, + 480, + 416 + ], + "lines": [ + { + "bbox": [ + 133, + 293, + 480, + 416 + ], + "spans": [ + { + "bbox": [ + 133, + 293, + 480, + 416 + ], + "type": "image", + "image_path": "511c367a10f90faea4dd8a326a1f5ec0de1aa1f9241e579495cf800156934a4a.jpg" + } + ] + } + ], + "index": 6, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 130, + 418, + 482, + 473 + ], + "lines": [ + { + "bbox": [ + 130, + 418, + 482, + 473 + ], + "spans": [ + { + "bbox": [ + 130, + 418, + 482, + 473 + ], + "type": "text", + "content": "Fig.7: Comparing Parrot with/without Pareto Optimal. By pretending different reward-specific preferences, Parrot can change the tradeoff between rewards. Specifically, “”, “”, “”, “” indicates aesthetics, human preference, text-image alignment, and image sentiment respectively. With Pareto-optimal selection, each reward increases based on reward-specific preference." + } + ] + } + ], + "index": 7, + "angle": 0, + "type": "image_caption" + } + ], + "index": 6 + }, + { + "bbox": [ + 130, + 496, + 482, + 544 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 496, + 482, + 544 + ], + "spans": [ + { + "bbox": [ + 130, + 496, + 482, + 544 + ], + "type": "text", + "content": "pertains to a specific quality aspect: aesthetics, human preference, text-image alignment, and image sentiment. For each prompt, 20 rounds of random sampling are conducted and sent to different raters. The user study results, illustrated in Fig. 5, show that Parrot outperforms other baselines across all dimensions." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 130, + 550, + 482, + 586 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 550, + 482, + 586 + ], + "spans": [ + { + "bbox": [ + 130, + 550, + 482, + 586 + ], + "type": "text", + "content": "Proportion of non-dominated points: Using batch size of 256, we observe that in a batch around " + }, + { + "bbox": [ + 130, + 550, + 482, + 586 + ], + "type": "inline_equation", + "content": "20\\%" + }, + { + "bbox": [ + 130, + 550, + 482, + 586 + ], + "type": "text", + "content": " to " + }, + { + "bbox": [ + 130, + 550, + 482, + 586 + ], + "type": "inline_equation", + "content": "30\\%" + }, + { + "bbox": [ + 130, + 550, + 482, + 586 + ], + "type": "text", + "content": " are non-dominated points and the proportion of non-dominated points in single batch slightly increase as training proceeds." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 131, + 604, + 209, + 615 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 131, + 604, + 209, + 615 + ], + "spans": [ + { + "bbox": [ + 131, + 604, + 209, + 615 + ], + "type": "text", + "content": "5.4 Ablations" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 130, + 629, + 481, + 665 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 629, + 481, + 665 + ], + "spans": [ + { + "bbox": [ + 130, + 629, + 481, + 665 + ], + "type": "text", + "content": "Effect of Pareto-optimal Multi-reward RL: To show the efficacy of Pareto-optimal Multi-reward RL, we conduct an ablation study by removing one reward model at a time. Fig. 6 shows quantitative results using one hundred random text" + } + ] + } + ], + "index": 11 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "spans": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "type": "text", + "content": "12" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 166, + 91, + 260, + 102 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 166, + 91, + 260, + 102 + ], + "spans": [ + { + "bbox": [ + 166, + 91, + 260, + 102 + ], + "type": "text", + "content": "Lee, Seung Hyun et al." + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 11 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 134, + 117, + 480, + 266 + ], + "blocks": [ + { + "bbox": [ + 134, + 117, + 480, + 266 + ], + "lines": [ + { + "bbox": [ + 134, + 117, + 480, + 266 + ], + "spans": [ + { + "bbox": [ + 134, + 117, + 480, + 266 + ], + "type": "image", + "image_path": "4945e6cd6998b9566e3a9fb25757a5fe57230c5b9429f8c1cbf73e047211a388.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 130, + 270, + 482, + 380 + ], + "lines": [ + { + "bbox": [ + 130, + 270, + 482, + 380 + ], + "spans": [ + { + "bbox": [ + 130, + 270, + 482, + 380 + ], + "type": "text", + "content": "Fig. 8: The comparisons of the diffusion fine-tuning between Pareto-optimal multi-reward RL and single reward RL. We show results with same seed from various methods: Stable Diffusion 1.5 [39] (1st column), T2I model fine-tuned with the aesthetic model (2nd column), the human preference model (3rd column), text-image alignment (4th column), image sentiment (5th column), Parrot without Pareto-optimal selection (6th column) and Parrot (7th column). Parrot is effective to generate acceptable images without sacrificing one of quality signals. For example, T2I model fine-tuned with a single quality signal such as aesthetics, human preference and image sentiment results in text-image misalignment, while our method achieves a balanced visual outcome across multiple criteria." + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_caption" + } + ], + "index": 2 + }, + { + "bbox": [ + 130, + 405, + 480, + 429 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 405, + 480, + 429 + ], + "spans": [ + { + "bbox": [ + 130, + 405, + 480, + 429 + ], + "type": "text", + "content": "prompts from the Promptist [15]. We observe that our training scheme improves multiple target objectives." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 130, + 430, + 482, + 537 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 430, + 482, + 537 + ], + "spans": [ + { + "bbox": [ + 130, + 430, + 482, + 537 + ], + "type": "text", + "content": "To verify whether Parrot achieved better trade-off for different reward scores, we generate 1000 images from common animal dataset [4], where each text prompt consists of the name of a common animal. As shown in Fig. 7, using only text-image alignment with reward-specific preference " + }, + { + "bbox": [ + 130, + 430, + 482, + 537 + ], + "type": "inline_equation", + "content": "\\langle reward3\\rangle" + }, + { + "bbox": [ + 130, + 430, + 482, + 537 + ], + "type": "text", + "content": " generates images with higher text-image alignment score, while using only aesthetic model with reward-specific preference " + }, + { + "bbox": [ + 130, + 430, + 482, + 537 + ], + "type": "inline_equation", + "content": "\\langle reward1\\rangle" + }, + { + "bbox": [ + 130, + 430, + 482, + 537 + ], + "type": "text", + "content": " yields images with higher aesthetic score. In the case of using two reward-specific preferences " + }, + { + "bbox": [ + 130, + 430, + 482, + 537 + ], + "type": "inline_equation", + "content": "\\langle reward1\\rangle" + }, + { + "bbox": [ + 130, + 430, + 482, + 537 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 130, + 430, + 482, + 537 + ], + "type": "inline_equation", + "content": "\\langle reward3\\rangle" + }, + { + "bbox": [ + 130, + 430, + 482, + 537 + ], + "type": "text", + "content": ", we observe that scores are balanced and show that better results across multiple rewards than Parrot without Pareto optimal selection." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 130, + 539, + 482, + 624 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 539, + 482, + 624 + ], + "spans": [ + { + "bbox": [ + 130, + 539, + 482, + 624 + ], + "type": "text", + "content": "Fig 8 shows the visual comparison between Parrot, Parrot with a single reward, and Parrot without selecting the batch-wise Pareto-optimal solution. Using a single reward model tends to result in degradation of other rewards, especially text-image alignment. In the third column, results of the first row miss the text " + }, + { + "bbox": [ + 130, + 539, + 482, + 624 + ], + "type": "inline_equation", + "content": "a" + }, + { + "bbox": [ + 130, + 539, + 482, + 624 + ], + "type": "text", + "content": " tophat in input prompt, even though the Stable Diffusion result includes that attribute. On the other hand, Parrot results capture all prompts, improving other quality signals, such as aesthetics, image sentiment and human preference." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 130, + 629, + 482, + 666 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 629, + 482, + 666 + ], + "spans": [ + { + "bbox": [ + 130, + 629, + 482, + 666 + ], + "type": "text", + "content": "Effect of Original Prompt Centered Guidance: Fig 9 shows the effect of the proposed original prompt-centered guidance. As evident from the figure, using only the expanded prompt as input often results in the main content be" + } + ] + } + ], + "index": 7 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 182, + 91, + 447, + 101 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 182, + 91, + 447, + 101 + ], + "spans": [ + { + "bbox": [ + 182, + 91, + 447, + 101 + ], + "type": "text", + "content": "Pareto-optimal Multi-Reward RL Framework for T2I Generation" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 470, + 92, + 480, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 470, + 92, + 480, + 100 + ], + "spans": [ + { + "bbox": [ + 470, + 92, + 480, + 100 + ], + "type": "text", + "content": "13" + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 12 + }, + { + "para_blocks": [ + { + "bbox": [ + 130, + 116, + 482, + 189 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 116, + 482, + 189 + ], + "spans": [ + { + "bbox": [ + 130, + 116, + 482, + 189 + ], + "type": "text", + "content": "ing overwhelmed by the added context. For instance, given the original prompt \"A shiba inu\", the result from the expanded prompt shows a zoomed-out image and the intended main subject (shiba inu) becomes small. The proposed original prompt-centered guidance effectively addresses this issue, generating an image that faithfully captures the input prompt while incorporating visually more pleasing details." + } + ] + } + ], + "index": 2 + }, + { + "type": "image", + "bbox": [ + 133, + 211, + 482, + 297 + ], + "blocks": [ + { + "bbox": [ + 133, + 211, + 482, + 297 + ], + "lines": [ + { + "bbox": [ + 133, + 211, + 482, + 297 + ], + "spans": [ + { + "bbox": [ + 133, + 211, + 482, + 297 + ], + "type": "image", + "image_path": "6a19c507fef9901d104d5d9638edfb87354d7390120f50624207b296ddd439f6.jpg" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 130, + 298, + 482, + 333 + ], + "lines": [ + { + "bbox": [ + 130, + 298, + 482, + 333 + ], + "spans": [ + { + "bbox": [ + 130, + 298, + 482, + 333 + ], + "type": "text", + "content": "Fig. 9: Results of original prompt centered guidance. As we expand the prompt, the content in the generated image often fades away. This guidance is helpful for keeping the main content of the original prompt." + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "image_caption" + } + ], + "index": 3 + }, + { + "bbox": [ + 132, + 384, + 312, + 397 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 384, + 312, + 397 + ], + "spans": [ + { + "bbox": [ + 132, + 384, + 312, + 397 + ], + "type": "text", + "content": "6 Conclusion and Limitation" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 130, + 418, + 482, + 598 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 418, + 482, + 598 + ], + "spans": [ + { + "bbox": [ + 130, + 418, + 482, + 598 + ], + "type": "text", + "content": "We propose Parrot, a novel multi-reward optimization algorithm aimed to improve text-to-image generation by effectively optimizing multiple quality rewards using RL. With batch-wise Pareto-optimal selection, Parrot adaptively balance the optimization of multiple quality rewards. By applying Parrot to jointly fine-tune both the T2I model and the prompt expansion model, we achieve the generation of higher-quality images with richer details. Additionally, our original prompt centered guidance technique ensures that the generated image maintains fidelity to the user prompt after prompt expansion during inference. Results from the user study indicate that Parrot significantly improves the quality of generated images across multiple criteria, including text-image alignment, human preference, aesthetics, and image sentiment. While Parrot has shown effectiveness in enhancing generated image quality, its efficacy is limited by the quality metrics it relies on. Therefore, advancements of the generated image quality metrics will directly enhance the capabilities of Parrot. Additionally, Parrot is adaptable to a broader range of rewards that quantify generated image quality." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 130, + 605, + 482, + 665 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 605, + 482, + 665 + ], + "spans": [ + { + "bbox": [ + 130, + 605, + 482, + 665 + ], + "type": "text", + "content": "Societal Impact: Parrot could potentially raise ethical concerns related to the generation of immoral content. This concern stems from the user's ability to influence T2I generation, allowing for the creation of visual content that may be deemed inappropriate. The risk may be tied to the potential biases in reward models inherited from various datasets." + } + ] + } + ], + "index": 7 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "spans": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "type": "text", + "content": "14" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 166, + 91, + 260, + 102 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 166, + 91, + 260, + 102 + ], + "spans": [ + { + "bbox": [ + 166, + 91, + 260, + 102 + ], + "type": "text", + "content": "Lee, Seung Hyun et al." + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 13 + }, + { + "para_blocks": [ + { + "bbox": [ + 132, + 114, + 198, + 127 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 114, + 198, + 127 + ], + "spans": [ + { + "bbox": [ + 132, + 114, + 198, + 127 + ], + "type": "text", + "content": "References" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 134, + 141, + 481, + 665 + ], + "type": "list", + "angle": 0, + "index": 21, + "blocks": [ + { + "bbox": [ + 138, + 141, + 391, + 152 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 141, + 391, + 152 + ], + "spans": [ + { + "bbox": [ + 138, + 141, + 391, + 152 + ], + "type": "text", + "content": "1. Amazon mechanical turk. https://www.mturk.com/ (2005)" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 138, + 153, + 481, + 186 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 153, + 481, + 186 + ], + "spans": [ + { + "bbox": [ + 138, + 153, + 481, + 186 + ], + "type": "text", + "content": "2. Anil, R., Dai, A.M., First, O., Johnson, M., Lepikhin, D., Passos, A., Shakeri, S., Taropa, E., Bailey, P., Chen, Z., et al.: Palm 2 technical report. arXiv preprint arXiv:2305.10403 (2023)" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 138, + 186, + 481, + 231 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 186, + 481, + 231 + ], + "spans": [ + { + "bbox": [ + 138, + 186, + 481, + 231 + ], + "type": "text", + "content": "3. Bai, Y., Jones, A., Ndousse, K., Askell, A., Chen, A., DasSarma, N., Drain, D., Fort, S., Ganguli, D., Henighan, T., et al.: Training a helpful and harmless assistant with reinforcement learning from human feedback. arXiv preprint arXiv:2204.05862 (2022)" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 138, + 232, + 481, + 253 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 232, + 481, + 253 + ], + "spans": [ + { + "bbox": [ + 138, + 232, + 481, + 253 + ], + "type": "text", + "content": "4. Black, K., Janner, M., Du, Y., Kostrikov, I., Levine, S.: Training diffusion models with reinforcement learning. arXiv preprint arXiv:2305.13301 (2023)" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 138, + 254, + 481, + 297 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 254, + 481, + 297 + ], + "spans": [ + { + "bbox": [ + 138, + 254, + 481, + 297 + ], + "type": "text", + "content": "5. Chang, H., Zhang, H., Barber, J., Maschinot, A., Lezama, J., Jiang, L., Yang, M.H., Murphy, K., Freeman, W.T., Rubinstein, M., et al.: Muse: Text-to-image generation via masked generative transformers. arXiv preprint arXiv:2301.00704 (2023)" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 138, + 298, + 481, + 319 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 298, + 481, + 319 + ], + "spans": [ + { + "bbox": [ + 138, + 298, + 481, + 319 + ], + "type": "text", + "content": "6. Clark, K., Vicol, P., Swersky, K., Fleet, D.J.: Directly fine-tuning diffusion models on differentiable rewards. ICLR (2024)" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 138, + 320, + 481, + 354 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 320, + 481, + 354 + ], + "spans": [ + { + "bbox": [ + 138, + 320, + 481, + 354 + ], + "type": "text", + "content": "7. Dai, X., Hou, J., Ma, C.Y., Tsai, S., Wang, J., Wang, R., Zhang, P., Vandenhende, S., Wang, X., Dubey, A., et al.: Emu: Enhancing image generation models using photogenic needles in a haystack. arXiv preprint arXiv:2309.15807 (2023)" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 138, + 354, + 481, + 387 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 354, + 481, + 387 + ], + "spans": [ + { + "bbox": [ + 138, + 354, + 481, + 387 + ], + "type": "text", + "content": "8. Deng, F., Wang, Q., Wei, W., Grundmann, M., Hou, T.: Prdp: Proximal reward difference prediction for large-scale reward finetuning of diffusion models. CVPR (2024)" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 138, + 387, + 481, + 420 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 387, + 481, + 420 + ], + "spans": [ + { + "bbox": [ + 138, + 387, + 481, + 420 + ], + "type": "text", + "content": "9. Dong, H., Xiong, W., Goyal, D., Pan, R., Dao, S., Zhang, J., Shum, K., Zhang, T.: Raft: Reward ranked finetuning for generative foundation model alignment. arXiv preprint arXiv:2304.06767 (2023)" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 134, + 420, + 481, + 464 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 134, + 420, + 481, + 464 + ], + "spans": [ + { + "bbox": [ + 134, + 420, + 481, + 464 + ], + "type": "text", + "content": "10. Dosovitskiy, A., Beyer, L., Kolesnikov, A., Weissenborn, D., Zhai, X., Unterthiner, T., Dehghani, M., Minderer, M., Heigold, G., Gelly, S., et al.: An image is worth 16x16 words: Transformers for image recognition at scale. arXiv preprint arXiv:2010.11929 (2020)" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 134, + 464, + 481, + 486 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 134, + 464, + 481, + 486 + ], + "spans": [ + { + "bbox": [ + 134, + 464, + 481, + 486 + ], + "type": "text", + "content": "11. Fan, Y., Lee, K.: Optimizing ddpm sampling with shortcut fine-tuning. ICML (2023)" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 134, + 487, + 481, + 520 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 134, + 487, + 481, + 520 + ], + "spans": [ + { + "bbox": [ + 134, + 487, + 481, + 520 + ], + "type": "text", + "content": "12. Fan, Y., Watkins, O., Du, Y., Liu, H., Ryu, M., Boutilier, C., Abbeel, P., Ghavamzadeh, M., Lee, K., Lee, K.: Dpok: Reinforcement learning for fine-tuning text-to-image diffusion models. NeurIPS (2023)" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 134, + 521, + 481, + 542 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 134, + 521, + 481, + 542 + ], + "spans": [ + { + "bbox": [ + 134, + 521, + 481, + 542 + ], + "type": "text", + "content": "13. Fang, Y., Zhu, H., Zeng, Y., Ma, K., Wang, Z.: Perceptual quality assessment of smartphone photography. In: CVPR (2020)" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 134, + 543, + 481, + 564 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 134, + 543, + 481, + 564 + ], + "spans": [ + { + "bbox": [ + 134, + 543, + 481, + 564 + ], + "type": "text", + "content": "14. Han, L., Li, Y., Zhang, H., Milanfar, P., Metaxas, D., Yang, F.: Svdiff: Compact parameter space for diffusion fine-tuning. CVPR (2023)" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 134, + 565, + 481, + 587 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 134, + 565, + 481, + 587 + ], + "spans": [ + { + "bbox": [ + 134, + 565, + 481, + 587 + ], + "type": "text", + "content": "15. Hao, Y., Chi, Z., Dong, L., Wei, F.: Optimizing prompts for text-to-image generation. arXiv preprint arXiv:2212.09611 (2022)" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 134, + 587, + 481, + 620 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 134, + 587, + 481, + 620 + ], + "spans": [ + { + "bbox": [ + 134, + 587, + 481, + 620 + ], + "type": "text", + "content": "16. He, H., Wang, T., Yang, H., Fu, J., Yuan, N.J., Yin, J., Chao, H., Zhang, Q.: Learning profitable nft image diffusions via multiple visual-policy guided reinforcement learning. arXiv preprint arXiv:2306.11731 (2023)" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 134, + 621, + 481, + 643 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 134, + 621, + 481, + 643 + ], + "spans": [ + { + "bbox": [ + 134, + 621, + 481, + 643 + ], + "type": "text", + "content": "17. Ho, J., Jain, A., Abbeel, P.: Denoising diffusion probabilistic models. NeurIPS (2020)" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 134, + 643, + 481, + 665 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 134, + 643, + 481, + 665 + ], + "spans": [ + { + "bbox": [ + 134, + 643, + 481, + 665 + ], + "type": "text", + "content": "18. Ho, J., Salimans, T.: Classifier-free diffusion guidance. arXiv preprint arXiv:2207.12598 (2022)" + } + ] + } + ], + "index": 20 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 182, + 91, + 447, + 102 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 182, + 91, + 447, + 102 + ], + "spans": [ + { + "bbox": [ + 182, + 91, + 447, + 102 + ], + "type": "text", + "content": "Pareto-optimal Multi-Reward RL Framework for T2I Generation" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 470, + 92, + 481, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 470, + 92, + 481, + 100 + ], + "spans": [ + { + "bbox": [ + 470, + 92, + 481, + 100 + ], + "type": "text", + "content": "15" + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 14 + }, + { + "para_blocks": [ + { + "bbox": [ + 132, + 116, + 481, + 665 + ], + "type": "list", + "angle": 0, + "index": 23, + "blocks": [ + { + "bbox": [ + 133, + 116, + 481, + 148 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 116, + 481, + 148 + ], + "spans": [ + { + "bbox": [ + 133, + 116, + 481, + 148 + ], + "type": "text", + "content": "19. Hosu, V., Lin, H., Sziranyi, T., Saupe, D.: Koniq-10k: An ecologically valid database for deep learning of blind image quality assessment. TIP 29, 4041-4056 (2020)" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 132, + 149, + 481, + 182 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 149, + 481, + 182 + ], + "spans": [ + { + "bbox": [ + 132, + 149, + 481, + 182 + ], + "type": "text", + "content": "20. Hu, E.J., Shen, Y., Wallis, P., Allen-Zhu, Z., Li, Y., Wang, S., Wang, L., Chen, W.: Lora: Low-rank adaptation of large language models. arXiv preprint arXiv:2106.09685 (2021)" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 132, + 182, + 481, + 214 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 182, + 481, + 214 + ], + "spans": [ + { + "bbox": [ + 132, + 182, + 481, + 214 + ], + "type": "text", + "content": "21. Jeong, Y., Ryoo, W., Lee, S., Seo, D., Byeon, W., Kim, S., Kim, J.: The power of sound (tpos): Audio reactive video generation with stable diffusion. In: ICCV (2023)" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 132, + 214, + 481, + 236 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 214, + 481, + 236 + ], + "spans": [ + { + "bbox": [ + 132, + 214, + 481, + 236 + ], + "type": "text", + "content": "22. Kawar, B., Zada, S., Lang, O., Tov, O., Chang, H., Dekel, T., Mosseri, I., Irani, M.: Imagic: Text-based real image editing with diffusion models. In: CVPR (2023)" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 132, + 236, + 481, + 257 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 236, + 481, + 257 + ], + "spans": [ + { + "bbox": [ + 132, + 236, + 481, + 257 + ], + "type": "text", + "content": "23. Ke, J., Wang, Q., Wang, Y., Milanfar, P., Yang, F.: Musiq: Multi-scale image quality transformer. In: ICCV (2021)" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 132, + 257, + 481, + 278 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 257, + 481, + 278 + ], + "spans": [ + { + "bbox": [ + 132, + 257, + 481, + 278 + ], + "type": "text", + "content": "24. Ke, J., Ye, K., Yu, J., Wu, Y., Milanfar, P., Yang, F.: Vila: Learning image aesthetics from user comments with vision-language pretraining. In: CVPR (2023)" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 132, + 278, + 481, + 289 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 278, + 481, + 289 + ], + "spans": [ + { + "bbox": [ + 132, + 278, + 481, + 289 + ], + "type": "text", + "content": "25. Kingma, D.P., Ba, J.: Adam: A method for stochastic optimization. ICLR (2015)" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 132, + 289, + 481, + 320 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 289, + 481, + 320 + ], + "spans": [ + { + "bbox": [ + 132, + 289, + 481, + 320 + ], + "type": "text", + "content": "26. Kirstain, Y., Polyak, A., Singer, U., Matiana, S., Penna, J., Levy, O.: Pick-a-pic: An open dataset of user preferences for text-to-image generation. arXiv preprint arXiv:2305.01569 (2023)" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 132, + 320, + 481, + 354 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 320, + 481, + 354 + ], + "spans": [ + { + "bbox": [ + 132, + 320, + 481, + 354 + ], + "type": "text", + "content": "27. Lee, K., Liu, H., Ryu, M., Watkins, O., Du, Y., Boutilier, C., Abbeel, P., Ghavamzadeh, M., Gu, S.S.: Aligning text-to-image models using human feedback. arXiv preprint arXiv:2302.12192 (2023)" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 132, + 354, + 481, + 386 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 354, + 481, + 386 + ], + "spans": [ + { + "bbox": [ + 132, + 354, + 481, + 386 + ], + "type": "text", + "content": "28. Lee, S.H., Kim, S., Yoo, I., Yang, F., Cho, D., Kim, Y., Chang, H., Kim, J., Kim, S.: Soundini: Sound-guided diffusion for natural video editing. arXiv preprint arXiv:2304.06818 (2023)" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 132, + 386, + 481, + 407 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 386, + 481, + 407 + ], + "spans": [ + { + "bbox": [ + 132, + 386, + 481, + 407 + ], + "type": "text", + "content": "29. Li, Y., Liu, H., Wu, Q., Mu, F., Yang, J., Gao, J., Li, C., Lee, Y.J.: Gligen: Open-set grounded text-to-image generation. In: CVPR (2023)" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 132, + 407, + 481, + 429 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 407, + 481, + 429 + ], + "spans": [ + { + "bbox": [ + 132, + 407, + 481, + 429 + ], + "type": "text", + "content": "30. Lin, X., Yang, Z., Zhang, X., Zhang, Q.: Pareto set learning for expensive multi-objective optimization. NeurIPS (2022)" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 132, + 429, + 481, + 450 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 429, + 481, + 450 + ], + "spans": [ + { + "bbox": [ + 132, + 429, + 481, + 450 + ], + "type": "text", + "content": "31. Mannor, S., Shimkin, N.: The steering approach for multi-criteria reinforcement learning. NeurIPS (2001)" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 132, + 450, + 481, + 471 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 450, + 481, + 471 + ], + "spans": [ + { + "bbox": [ + 132, + 450, + 481, + 471 + ], + "type": "text", + "content": "32. Miettinen, K.: Nonlinear multiobjective optimization, vol. 12. Springer Science & Business Media (1999)" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 132, + 471, + 481, + 492 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 471, + 481, + 492 + ], + "spans": [ + { + "bbox": [ + 132, + 471, + 481, + 492 + ], + "type": "text", + "content": "33. Murray, N., Marchesotti, L., Perronnin, F.: Ava: A large-scale database for aesthetic visual analysis. In: CVPR (2012)" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 132, + 492, + 481, + 525 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 492, + 481, + 525 + ], + "spans": [ + { + "bbox": [ + 132, + 492, + 481, + 525 + ], + "type": "text", + "content": "34. Ouyang, L., Wu, J., Jiang, X., Almeida, D., Wainwright, C., Mishkin, P., Zhang, C., Agarwal, S., Slama, K., Ray, A., et al.: Training language models to follow instructions with human feedback. NeurIPS (2022)" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 132, + 525, + 481, + 557 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 525, + 481, + 557 + ], + "spans": [ + { + "bbox": [ + 132, + 525, + 481, + 557 + ], + "type": "text", + "content": "35. Radford, A., Kim, J.W., Hallacy, C., Ramesh, A., Goh, G., Agarwal, S., Sastry, G., Askell, A., Mishkin, P., Clark, J., et al.: Learning transferable visual models from natural language supervision. In: ICML (2021)" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 132, + 557, + 481, + 590 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 557, + 481, + 590 + ], + "spans": [ + { + "bbox": [ + 132, + 557, + 481, + 590 + ], + "type": "text", + "content": "36. Rame, A., Couairon, G., Shukor, M., Dancette, C., Gaya, J.B., Soulier, L., Cord, M.: Rewarded soups: towards pareto-optimal alignment by interpolating weights fine-tuned on diverse rewards. NeurIPS (2023)" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 132, + 590, + 481, + 622 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 590, + 481, + 622 + ], + "spans": [ + { + "bbox": [ + 132, + 590, + 481, + 622 + ], + "type": "text", + "content": "37. Ramesh, A., Dhariwal, P., Nichol, A., Chu, C., Chen, M.: Hierarchical text-conditional image generation with clip latents. arXiv preprint arXiv:2204.06125 (2022)" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 132, + 622, + 481, + 643 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 622, + 481, + 643 + ], + "spans": [ + { + "bbox": [ + 132, + 622, + 481, + 643 + ], + "type": "text", + "content": "38. Richardson, E., Goldberg, K., Alaluf, Y., Cohen-Or, D.: Conceptlab: Creative generation using diffusion prior constraints. arXiv preprint arXiv:2308.02669 (2023)" + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 132, + 643, + 481, + 665 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 643, + 481, + 665 + ], + "spans": [ + { + "bbox": [ + 132, + 643, + 481, + 665 + ], + "type": "text", + "content": "39. Rombach, R., Blattmann, A., Lorenz, D., Esser, P., Ommer, B.: High-resolution image synthesis with latent diffusion models. In: CVPR (2022)" + } + ] + } + ], + "index": 22 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "spans": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "type": "text", + "content": "16" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 166, + 91, + 260, + 101 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 166, + 91, + 260, + 101 + ], + "spans": [ + { + "bbox": [ + 166, + 91, + 260, + 101 + ], + "type": "text", + "content": "Lee, Seung Hyun et al." + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 15 + }, + { + "para_blocks": [ + { + "bbox": [ + 132, + 116, + 480, + 544 + ], + "type": "list", + "angle": 0, + "index": 18, + "blocks": [ + { + "bbox": [ + 132, + 116, + 480, + 138 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 116, + 480, + 138 + ], + "spans": [ + { + "bbox": [ + 132, + 116, + 480, + 138 + ], + "type": "text", + "content": "40. Ronneberger, O., Fischer, P., Brox, T.: U-net: Convolutional networks for biomedical image segmentation. In: MICCAI (2015)" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 132, + 139, + 480, + 160 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 139, + 480, + 160 + ], + "spans": [ + { + "bbox": [ + 132, + 139, + 480, + 160 + ], + "type": "text", + "content": "41. Sahara, C., Chan, W., Chang, H., Lee, C., Ho, J., Salimans, T., Fleet, D., Norouzi, M.: Palette: Image-to-image diffusion models. In: SIGGRAPH (2022)" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 132, + 161, + 480, + 194 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 161, + 480, + 194 + ], + "spans": [ + { + "bbox": [ + 132, + 161, + 480, + 194 + ], + "type": "text", + "content": "42. Sahara, C., Chan, W., Saxena, S., Li, L., Whang, J., Denton, E.L., Ghasemipour, K., Gontijo Lopes, R., Karagol Ayan, B., Salimans, T., et al.: Photorealistic text-to-image diffusion models with deep language understanding. NeurIPS (2022)" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 132, + 194, + 480, + 226 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 194, + 480, + 226 + ], + "spans": [ + { + "bbox": [ + 132, + 194, + 480, + 226 + ], + "type": "text", + "content": "43. Schuhmann, C., Beaumont, R., Vencu, R., Gordon, C., Wightman, R., Cherti, M., Coombes, T., Katta, A., Mullis, C., Wortsman, M., et al.: Laion-5b: An open large-scale dataset for training next generation image-text models. NeurIPS (2022)" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 132, + 227, + 480, + 248 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 227, + 480, + 248 + ], + "spans": [ + { + "bbox": [ + 132, + 227, + 480, + 248 + ], + "type": "text", + "content": "44. Serra, A., Carrara, F., Tesconi, M., Falchi, F.: The emotions of the crowd: Learning image sentiment from tweets via cross-modal distillation. ECAI (2023)" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 132, + 249, + 480, + 270 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 249, + 480, + 270 + ], + "spans": [ + { + "bbox": [ + 132, + 249, + 480, + 270 + ], + "type": "text", + "content": "45. Sohl-Dickstein, J., Weiss, E., Maheswaranathan, N., Ganguli, S.: Deep unsupervised learning using nonequilibrium thermodynamics. In: ICML (2015)" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 132, + 271, + 480, + 292 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 271, + 480, + 292 + ], + "spans": [ + { + "bbox": [ + 132, + 271, + 480, + 292 + ], + "type": "text", + "content": "46. Song, J., Meng, C., Ermon, S.: Denoising diffusion implicit models. arXiv preprint arXiv:2010.02502 (2020)" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 132, + 293, + 480, + 325 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 293, + 480, + 325 + ], + "spans": [ + { + "bbox": [ + 132, + 293, + 480, + 325 + ], + "type": "text", + "content": "47. Tesauro, G., Das, R., Chan, H., Kephart, J., Levine, D., Rawson, F., Lefury, C.: Managing power consumption and performance of computing systems using reinforcement learning. NeurIPS (2007)" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 132, + 326, + 480, + 346 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 326, + 480, + 346 + ], + "spans": [ + { + "bbox": [ + 132, + 326, + 480, + 346 + ], + "type": "text", + "content": "48. Tu, Z., Talebi, H., Zhang, H., Yang, F., Milanfar, P., Bovik, A., Li, Y.: Maxvit: Multi-axis vision transformer. In: ECCV (2022)" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 132, + 347, + 480, + 369 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 347, + 480, + 369 + ], + "spans": [ + { + "bbox": [ + 132, + 347, + 480, + 369 + ], + "type": "text", + "content": "49. Vaswani, A., Shazeer, N., Parmar, N., Uszkoreit, J., Jones, L., Gomez, A.N., Kaiser, L., Polosukhin, I.: Attention is all you need. NeurIPS (2017)" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 132, + 369, + 480, + 391 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 369, + 480, + 391 + ], + "spans": [ + { + "bbox": [ + 132, + 369, + 480, + 391 + ], + "type": "text", + "content": "50. Wu, X., Sun, K., Zhu, F., Zhao, R., Li, H.: Human preference score: Better aligning text-to-image models with human preference. In: ICCV (2023)" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 132, + 392, + 480, + 423 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 392, + 480, + 423 + ], + "spans": [ + { + "bbox": [ + 132, + 392, + 480, + 423 + ], + "type": "text", + "content": "51. Xu, J., Liu, X., Wu, Y., Tong, Y., Li, Q., Ding, M., Tang, J., Dong, Y.: Imagereward: Learning and evaluating human preferences for text-to-image generation. arXiv preprint arXiv:2304.05977 (2023)" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 132, + 424, + 480, + 456 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 424, + 480, + 456 + ], + "spans": [ + { + "bbox": [ + 132, + 424, + 480, + 456 + ], + "type": "text", + "content": "52. Ying, Z., Niu, H., Gupta, P., Mahajan, D., Ghadiyaram, D., Bovik, A.: From patches to pictures (paq-2-piq): Mapping the perceptual space of picture quality. In: CVPR (2020)" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 132, + 457, + 480, + 490 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 457, + 480, + 490 + ], + "spans": [ + { + "bbox": [ + 132, + 457, + 480, + 490 + ], + "type": "text", + "content": "53. Yu, J., Xu, Y., Koh, J.Y., Luong, T., Baid, G., Wang, Z., Vasudevan, V., Ku, A., Yang, Y., Ayan, B.K., et al.: Scaling autoregressive models for content-rich text-to-image generation. arXiv preprint arXiv:2206.10789 (2022)" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 132, + 491, + 480, + 522 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 491, + 480, + 522 + ], + "spans": [ + { + "bbox": [ + 132, + 491, + 480, + 522 + ], + "type": "text", + "content": "54. Yu, L., Cheng, Y., Sohn, K., Lezama, J., Zhang, H., Chang, H., Hauptmann, A.G., Yang, M.H., Hao, Y., Essa, I., et al.: Magvit: Masked generative video transformer. In: CVPR (2023)" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 132, + 523, + 480, + 544 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 523, + 480, + 544 + ], + "spans": [ + { + "bbox": [ + 132, + 523, + 480, + 544 + ], + "type": "text", + "content": "55. Zhou, Y., Liu, B., Zhu, Y., Yang, X., Chen, C., Xu, J.: Shifted diffusion for text-to-image generation. In: CVPR (2023)" + } + ] + } + ], + "index": 17 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 182, + 91, + 447, + 102 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 182, + 91, + 447, + 102 + ], + "spans": [ + { + "bbox": [ + 182, + 91, + 447, + 102 + ], + "type": "text", + "content": "Pareto-optimal Multi-Reward RL Framework for T2I Generation" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 470, + 91, + 481, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 470, + 91, + 481, + 100 + ], + "spans": [ + { + "bbox": [ + 470, + 91, + 481, + 100 + ], + "type": "text", + "content": "17" + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 16 + } + ], + "_backend": "vlm", + "_version_name": "2.6.4" +} \ No newline at end of file diff --git a/2024/Part2Object_ Hierarchical Unsupervised 3D Instance Segmentation/918bb93b-c2e7-4e11-b966-4fcc81659f0e_content_list.json b/2024/Part2Object_ Hierarchical Unsupervised 3D Instance Segmentation/918bb93b-c2e7-4e11-b966-4fcc81659f0e_content_list.json new file mode 100644 index 0000000000000000000000000000000000000000..cde8d68cd326e7c0ac52cc145af05e1acbea8bcc --- /dev/null +++ b/2024/Part2Object_ Hierarchical Unsupervised 3D Instance Segmentation/918bb93b-c2e7-4e11-b966-4fcc81659f0e_content_list.json @@ -0,0 +1,1489 @@ +[ + { + "type": "text", + "text": "Part2Object: Hierarchical Unsupervised 3D Instance Segmentation", + "text_level": 1, + "bbox": [ + 254, + 140, + 750, + 185 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Cheng Shi*, Yulin Zhang*, Bin Yang, Jiajin Tang, Yuexin Ma, and Sibei Yang", + "bbox": [ + 320, + 212, + 679, + 243 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "School of Information Science and Technology ShanghaiTech University {shicheng2022,yangsb}@shanghaiitech.edu.cn", + "bbox": [ + 341, + 255, + 656, + 297 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Abstract. Unsupervised 3D instance segmentation aims to segment objects from a 3D point cloud without any annotations. Existing methods face the challenge of either too loose or too tight clustering, leading to under-segmentation or over-segmentation. To address this issue, we propose Part2Object, hierarchical clustering with object guidance. Part2Object employs multi-layer clustering from points to object parts and objects, allowing objects to manifest at any layer. Additionally, it extracts and utilizes 3D objectness priors from temporally consecutive 2D RGB frames to guide the clustering process. Moreover, we propose Hi-Mask3D to support hierarchical 3D object part and instance segmentation. By training Hi-Mask3D on the objects and object parts extracted from Part2Object, we achieve consistent and superior performance compared to state-of-the-art models in various settings, including unsupervised instance segmentation, data-efficient fine-tuning, and cross-dataset generalization. Code is release at https://github.com/ChengShiest/Part2object.", + "bbox": [ + 261, + 337, + 738, + 559 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Keywords: 3D Instance Segmentation $\\cdot$ Unsupervised Learning", + "bbox": [ + 261, + 573, + 692, + 587 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "1 Introduction", + "text_level": 1, + "bbox": [ + 217, + 614, + 374, + 630 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "3D instance segmentation, parsing semantic compositions from complex point cloud scenes into distinct objects, is fundamental and crucial in real-world applications, such as mixed reality [51], autonomous navigation [3, 52], planning [23], and manipulation [17]. In recent years, 3D instance segmentation has seen substantial advancements [19, 37, 38, 42, 55, 62], improving both the accuracy and efficiency of object localization and recognition. Nevertheless, it often demands large-scale human annotations for fully-supervised training, resulting in costly and manpower-intensive efforts. While a few studies [33, 47] have initiated to address 3D instance segmentation without human annotations, their reliance on 3D scene flows from sequential point clouds restricts their applicability to single point cloud scenarios, such as indoor scenes in ScanNet [12].", + "bbox": [ + 217, + 648, + 785, + 814 + ], + "page_idx": 0 + }, + { + "type": "page_footnote", + "text": "* Equal contribution. $\\boxtimes$ Corresponding author.", + "bbox": [ + 232, + 825, + 550, + 839 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/9e2c399d4d0af5c900b794d764d3a5c384bacf2840baf4a16c5ae60fa0c5a65a.jpg", + "image_caption": [ + "Fig. 1: Motivation of our hierarchical clustering. Single-level clustering results in a trade-off between under-segmentation for certain objects and over-segmentation for others. In contrast, our hierarchical clustering allows for gathering and identifying objects at varying levels of clustering granularity." + ], + "image_footnote": [], + "bbox": [ + 225, + 202, + 777, + 412 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "In this paper, we study unsupervised 3D instance segmentation, aiming to segment class-agnostic instances from indoor 3D scenes without relying on any human labels. To address this problem, two straightforward strategies have been explored recently: 1) to apply traditional clustering [16, 32] or graph-cut methods [66] to group points into objects based on their RGB-D data, such as coordinates, colors, normal vectors, and self-supervised pretraining features [41]. 2) To project 2D instance segmentation results from 2D RGB frames onto the point cloud using the corresponding camera pose. Typically, 2D RGB frames are captured concurrently with indoor point cloud data [12, 39], while 2D instance masks are extracted using 2D unsupervised instance segmentation advancements [59-61]. However, both strategies can only work in straightforward scenes with few salient objects, as they cannot tackle the fundamental challenges inherent in unsupervised 3D instance segmentation, as outlined below.", + "bbox": [ + 212, + 444, + 787, + 641 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "In complex scenes, achieving consistent segmentation granularity for different objects within a single clustering and graph-cut result is challenging, given the significant variations in geometric shape, color, and size among 3D objects, as well as the diverse compositions of these objects within scenes. This entails a trade-off between under-segmentation for certain objects and over-segmentation for others, making it almost impossible to attain satisfactory segmentation granularity for all of them. As shown in Fig 1, tighter grouping can accurately segment objects with simple geometric shapes and textures, such as the pillow on the bed (see layer one). However, it can also result in the over-segmentation of larger and more complex objects like the toilet in the same layer (layer one), and vice versa. In addition, there is a lack of an effective strategy to identify whether a point cluster or group represents a 3D instance. For example, the toilet back (see layer two) forms a well-grouped clus", + "bbox": [ + 212, + 643, + 787, + 840 + ], + "page_idx": 1 + }, + { + "type": "page_number", + "text": "2", + "bbox": [ + 217, + 114, + 228, + 126 + ], + "page_idx": 1 + }, + { + "type": "header", + "text": "Cheng et al.", + "bbox": [ + 271, + 114, + 354, + 128 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "ter, but it does not constitute a meaningful object. The 3D geometric features are fundamentally semantically insufficient to identify objects. Recently, it has been discovered that self-supervised vision transformer (ViT) features contain information about object segmentations in 2D images [6, 60]. Building on this discovery, 3D unsupervised instance segmentation approaches [41] project the 2D self-supervised ViT features (or their inferred object segmentations) from 2D RGB frames onto 3D point clouds, thereby enhancing the perception of objectness for 3D point features. However, a 3D point typically corresponds to multiple pixels in different 2D RGB frames, and this many-to-one mapping is fragile, easily disrupted by noise features or imprecise masks on any frame, resulting in an inability to distinguish one object from others or the segmentation of objects into fragmented segments, as illustrated in Fig 3(b) and (c). In addition, despite enhanced point features, discerning nearby objects or those sharing similar semantics remains challenging due to the inherently greater complexity of 3D scenes compared to 2D images.", + "bbox": [ + 212, + 146, + 787, + 375 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "To tackle these challenges, we introduce a simple yet effective principle, \"Gather & Aim\", for the automatic discovery and segmentation of 3D objects. For the gathering, inspired by the observation that tighter or looser groupings can respectively include certain objects within a scene, we propose allowing the gathering of potential objects at varying levels of grouping granularity, rather than being restricted to a single level. Specifically, we perform hierarchical clustering in the 3D scene, gradually grouping points into larger clusters, as illustrated in Fig 1. Therefore, instead of striving to improve single-layer clustering results, ours is expected that as long as each potential instance appears in one of the layers, we will be able to lock onto the segmentation mask of that instance. This mitigates excessive reliance on any single-level clustering results, thereby enhancing the clustering's adaptability to significant variations among objects and scenes. For the aiming, we propose to target the 3D objectness priors, a set of candidate objects, from temporally consecutive 2D RGB frames. We leverage both the temporal consistency across frames and the prior knowledge of unsupervised object discovery within frames to extract the 3D objectness priors. Note that we jointly estimate 2D objects from multiple frames first and then approximate object-level 3D bounding boxes, departing from the point-level projection used in previous methods [41], to address the fragility issue of the many-to-one mapping from pixels to points. It avoids adhesive or scattered 3D fragments caused by inconsistencies and occlusions over 2D RGB frames. Finally, the 3D bounding boxes serve as indicators to identify the 3D instance masks within the multi-level clusters. Our \"Gather & Aim\" strategy enables leverage of 2D semantic priors for 3D object identification while harnessing 3D geometric priors for precise instance segmentation.", + "bbox": [ + 212, + 380, + 787, + 758 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Furthermore, the taxonomy of how object parts are composed into an object is a strong yet disregarded prior information for unsupervised 3D instance segmentation. For example, the taxonomy based on the parts of a chair, including the back, arm, seat, and leg, can assist in addressing highly challenging scenarios, such as distinguishing two closely positioned and semantically similar", + "bbox": [ + 212, + 763, + 787, + 840 + ], + "page_idx": 2 + }, + { + "type": "header", + "text": "Part2Object", + "bbox": [ + 645, + 114, + 730, + 128 + ], + "page_idx": 2 + }, + { + "type": "page_number", + "text": "3", + "bbox": [ + 774, + 116, + 785, + 126 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "chairs as two separate instances. Thanks to our hierarchical clustering, we cannot only identify objects but also trace their constituent parts (see the toilet and its parts in layers three and two, respectively). Therefore, we extract both objects and object parts from our hierarchical clustering and name our clustering algorithm Part2Object. Expanding on this, we enhance the 3D instance segmentation framework, Mask3D [42], to support hierarchical 3D part and instance segmentation. Our improved model, named Hi-Mask3D, takes the object and object parts from Part2Object as pseudo-labels for learning 3D instance segmentation by incorporating explicit interactions between object parts and objects.", + "bbox": [ + 212, + 146, + 787, + 297 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "To evaluate the effectiveness of our Part2Object clustering and Hi-Mask3D model, we conduct experiments on the challenging and cluttered indoor environments [4,12,39] in three settings: 1) direct evaluation on unsupervised instance segmentation, 2) application to data-efficient fine-tuning, and 3) cross-dataset generalization, all showing significant performance improvement. In summary, our contributions are multi-fold:", + "bbox": [ + 212, + 297, + 787, + 388 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "- We propose two key insights for unsupervised 3D instance segmentation: 1) Employing hierarchical clustering enables the gathering of objects at varying levels of clustering granularity. 2) Leveraging 3D objectness priors from temporally consecutive 2D frames as guidance, while harnessing 3D geometric priors to clustering on the point cloud for precise instance segmentation.", + "bbox": [ + 225, + 402, + 782, + 477 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "- Based on our insights, we propose an innovative hierarchical clustering approach, Part2Object. It progressively groups points into object parts and objects while extracting and leveraging 3D objectness priors to guide the clustering process. Our Part2Object significantly outperforms the state-of-the-art training-free unsupervised 3D instance segmentation methods by $16.8\\%$ mAP@50 on the ScanNet dataset.", + "bbox": [ + 225, + 479, + 784, + 566 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "- We propose Hi-Mask3D, an extension of 3D instance segmentation to support hierarchical unsupervised 3D part and instance segmentation. Experiments demonstrate that Hi-Mask3D consistently and significantly outperforms state-of-the-art models in all the settings.", + "bbox": [ + 225, + 568, + 784, + 628 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "2 Related Work", + "text_level": 1, + "bbox": [ + 215, + 654, + 387, + 669 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Unsupervised 3D instance segmentation. 3D instance segmentation is an essential task in 3D scene understanding, which aims to locate and recognize different objects in a 3D point cloud. However, the cost and labor intensity of 3D scene-level instance annotation underscore the significance of exploring challenging unsupervised instance segmentation methods. Early works utilize raw geometric information, including coordinates, colors, and normal vectors, to perform traditional clustering [15,16,32] for segmentation. Recently, inspired by the new paradigm in 2D unsupervised instance segmentation, some works [41,66] introduce pseudo labels and self-training to unsupervised 3D instance segmentation. Specifically, these methods use 2D or 3D self-supervised models to extract", + "bbox": [ + 212, + 688, + 787, + 840 + ], + "page_idx": 3 + }, + { + "type": "page_number", + "text": "4", + "bbox": [ + 217, + 114, + 228, + 126 + ], + "page_idx": 3 + }, + { + "type": "header", + "text": "Cheng et al.", + "bbox": [ + 271, + 114, + 354, + 128 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "point cloud features, applying graph-cut algorithms to generate pseudo-labels for training and prediction. As illustrated earlier, the clustering method encounters a trade-off between under-segmentation and over-segmentation. Relying on a single graph-cut algorithm, as recent works do, falls short of achieving consistent granularity in complex indoor environments.", + "bbox": [ + 212, + 146, + 787, + 222 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Transfer 2D foundation models into 3D. The rapid advancements in 2D vision have given rise to powerful foundation models [2,6,13,18,26,28,43-45,53], proving beneficial across a variety of visual tasks. In contrast to the 2D domain, 3D vision encounters challenges with data scarcity and training limitations, hindering the development of foundation models. Consequently, several works in the 3D turn to leveraging 2D foundation models to address 3D problems. In 3D instance segmentation, some works [7,64] utilize 2D vision foundation models to generate pseudo-labels for 2D images, aiding in the training and prediction of 3D models through 2D to 3D projection algorithms. Other works [31,36] employ features rather than outputs obtained from 2D foundation models, enhancing the feature extraction capabilities of 3D models through pixel-point alignment. However, relying solely on 2D features or 2D pseudo-labels poses challenges in addressing cluttered indoor scenes with stacked objects.", + "bbox": [ + 212, + 222, + 787, + 419 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Supervised point cloud segmentation. Supervised point cloud segmentation can be divided into instance-level and part-level segmentation. The former has been significant breakthroughs in recent years [8, 19, 20, 22, 25, 27, 37, 38, 42, 50, 55, 56, 58, 62, 63]. However, these approaches demand a substantial amount of annotated data and focus on learning object-level attributes, neglecting the understanding of parts and their contributions to the composition of objects. In the realm of 3D part segmentation, various setups, including supervised [30,37,65], weakly supervised [9,10,57] instance segmentation and semantic segmentation, as well as open-world semantic segmentation [24, 36], have been extensively explored. While these approaches have demonstrated promising result, their primary emphasis lies in determining how to divide an object into parts. To integrate instance-level and part-level perspectives, we perform hierarchical clustering with the prior knowledge of how parts compose objects, unifying instance-level object identification and part-level constituent tracing.", + "bbox": [ + 212, + 419, + 787, + 633 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "3 Methodology", + "text_level": 1, + "bbox": [ + 215, + 655, + 380, + 672 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Problem definition. Unsupervised 3D instance segmentation [41, 67] is the task of generating class-agnostic masks for all foreground objects in a 3D point cloud scene solely based on raw RGB-D data, without relying on any object detection or instance segmentation annotations. The raw RGB-D data [12, 39] consists of the point cloud $P \\in \\mathbb{R}^{N \\times 3}$ and corresponding 2D RGB frames $\\{I_i\\}_{i=1}^M$ , where each image frame $I_m \\in \\mathbb{R}^{H \\times W \\times 3}$ . With raw RGB-D data, 3D instance segmentation task aims to output all object masks $P_{\\mathrm{object}}$ in the scene.", + "bbox": [ + 212, + 688, + 787, + 794 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Method overview. As shown in Fig 2, our unsupervised framework has two major components: 1) the hierarchical clustering algorithm (Sec 3.1 and Sec 3.2), Part2Object, which progressively group points into larger clusters. Duringclus", + "bbox": [ + 212, + 795, + 787, + 840 + ], + "page_idx": 4 + }, + { + "type": "header", + "text": "Part2Object", + "bbox": [ + 645, + 114, + 730, + 128 + ], + "page_idx": 4 + }, + { + "type": "page_number", + "text": "5", + "bbox": [ + 774, + 116, + 784, + 126 + ], + "page_idx": 4 + }, + { + "type": "image", + "img_path": "images/2c25a136c403368b85f58a5414f7feb2309379dbd617bc02a166b98383d86538.jpg", + "image_caption": [ + "Fig. 2: Overview of our Part2Object hierarchical clustering and Hi-Mask3D instance segmentation framework. Part2Object extracts 3D objectness priors from consecutive 2D RGB frames and uses them to guide hierarchical clustering from points to object parts and objects. Hi-Mask3D utilizes objects and parts identified by Part2Object as pseudo-labels, learning for improving instance segmentation through the utilization of object parts." + ], + "image_footnote": [], + "bbox": [ + 222, + 243, + 784, + 453 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "tering, aided by objectness priors from 2D frames, we can effectively obtain both object-level and part-level segmentation masks, which serve as pseudo-labels for learning unsupervised instance segmentation. 2) The end-to-end 3D instance segmentation model Hi-Mask3D (Sec 3.3) extends the original Mask3D model [42] by explicitly predicting segmentation results of object parts and leveraging them to aid instance segmentation. By self-training on the pseudo-labels extracted from the first-stage Part2Object, Hi-Mask3D can effectively predict 3D instance segments.", + "bbox": [ + 212, + 497, + 787, + 619 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "3.1 Hierarchical Clustering on 3D Point Cloud", + "text_level": 1, + "bbox": [ + 214, + 643, + 614, + 660 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "In this section, we introduce the hierarchical clustering process of our Part2Object which starts by grouping initial points into initial clusters and then progressively merging them into larger clusters. We will sequentially cover: 1) the feature representation of points, 2) the initialization of clusters from points, and 3) the hierarchical clustering process for merging clusters. Point Feature Representation. Given a 3D surface point cloud $P$ , we denote its norm and color as $P^{\\mathrm{Norm}}$ and $P^{\\mathrm{RGB}}$ , respectively. The 3D feature $f_{i}^{3\\mathrm{D}}$ of point $p_i \\in P$ comprises $p_i$ , $p_i^{\\mathrm{Norm}}$ , and $p_i^{\\mathrm{RGB}}$ , representing the location, color, and geometry information, respectively. Additionally, we define point cloud 2D semantics feature $f_{i}$ by projecting 2D features [6] of RGB frames into 3D space using the corresponding camera pose. Here, 2D features are extracted using the 2D self-supervised", + "bbox": [ + 212, + 672, + 789, + 840 + ], + "page_idx": 5 + }, + { + "type": "page_number", + "text": "6", + "bbox": [ + 217, + 114, + 228, + 126 + ], + "page_idx": 5 + }, + { + "type": "header", + "text": "Cheng et al.", + "bbox": [ + 271, + 114, + 354, + 128 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "method DINO [6], without rely on any 2D instance annotations. Cluster Initialization and Feature Representation. Given the vast number of points in the point cloud data [1], we initially apply the VCCS algorithm [34] to generate super-points on the point cloud, forming the first-layer clusters $\\{c_i^0\\}_{i = 1}^{N_0}$ as follows:", + "bbox": [ + 212, + 146, + 787, + 220 + ], + "page_idx": 6 + }, + { + "type": "equation", + "text": "\n$$\n\\{c _ {i} ^ {0} \\} _ {i = 1} ^ {N _ {0}} = \\mathrm {V C C S} (P, P ^ {\\mathrm {N o r m}}, P ^ {\\mathrm {R G B}}), \\tag {1}\n$$\n", + "text_format": "latex", + "bbox": [ + 379, + 223, + 785, + 242 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "where VCCS algorithm takes points' coordinates $P$ , norms $P^{\\mathrm{Norm}}$ , and colors $P^{\\mathrm{RGB}}$ as input and group original $N$ points into $N_0$ super-points as $\\{c_i^0\\}_{i=1}^{N_0}$ . We consider each super-point $c_i^0$ as a cluster, consisting of multiple points that are closely positioned and share similar color and norm characteristics. Then, we define the feature representation for the cluster based on the features of the points within the cluster. Instead of directly averaging point features within the cluster, we compute a weight for each point and perform a weighted sum of features. These weights are determined by the similarity between each point's feature and the average feature of points within the cluster. This approach helps mitigate the influence of noisy points in the cluster, leading to more robust representations. Specifically, for the cluster $c_i^0$ , its cluster feature $f_i^0$ is computed as follows,", + "bbox": [ + 212, + 251, + 787, + 434 + ], + "page_idx": 6 + }, + { + "type": "equation", + "text": "\n$$\n\\boldsymbol {f} _ {i} ^ {0} = \\sum_ {p _ {j} \\in c _ {i} ^ {0}} \\frac {\\operatorname {s i m} \\left(\\boldsymbol {f} _ {j} , \\bar {\\boldsymbol {f}} _ {i} ^ {0}\\right)}{\\sum p _ {j} \\in c _ {j} ^ {0} \\operatorname {s i m} \\left(\\boldsymbol {f} _ {j} , \\bar {\\boldsymbol {f}} _ {i} ^ {0}\\right)} \\boldsymbol {f} _ {j}, \\text {w h e r e} \\bar {\\boldsymbol {f}} _ {i} ^ {0} = \\sum_ {p _ {j} \\in c _ {i} ^ {0}} \\frac {1}{\\left| c _ {i} ^ {0} \\right|} \\boldsymbol {f} _ {j}. \\tag {2}\n$$\n", + "text_format": "latex", + "bbox": [ + 281, + 446, + 785, + 491 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "The $p_j \\in c_i^0$ denotes each point $p_j$ within the cluster $c_i^0$ , and $\\pmb{f}_j$ is the feature of point $p_j$ . The $\\bar{\\pmb{f}}_i^0$ is the average feature of points within $c_i^0$ , and $\\sin(\\cdot, \\cdot)$ denotes the computation of cosine similarity. As depicted in the example in Fig 3(a), our cluster feature $\\pmb{f}_i^0$ demonstrates greater robustness compared to $\\bar{\\pmb{f}}_i^0$ . For simplicity, we abbreviate the operations in Equ 2 as function $\\mathsf{FU}(\\cdot)$ , which computes the cluster feature for the inputted cluster based on point features, e.g., $\\pmb{f}_i^0 = \\mathsf{FU}(c_i^0)$ .", + "bbox": [ + 212, + 505, + 787, + 618 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Hierarchical Clustering with One Stop Criteria. Next, we group and merge the first-layer clusters $\\{c_i^0\\}_{i = 1}^{N_0}$ to the next-layer clusters $\\{c_i^1\\}_{i = 1}^{N_1}$ based on their features $\\{\\pmb {f}_i^0\\}_{i = 0}^{N_0}$ and spatial coordinates, progressively and iteratively forming higher-hierarchical clusters $\\{c_{ij}^{t}\\}_{i = 1}^{N_t}$ , where $t$ represents the $t$ -th layer in hierarchical clustering. In each single clustering layer, our clustering principle is that two clusters can only be merged when they are semantically similar and spatially adjacent: 1) the feature similarity between them ranks among the top $K$ similarities between any pair of clusters. 2) The closest points between two clusters are adjacent. Specifically, when considering two clusters in the $t$ -th layer, $c_{i}^{t}$ and $c_{j}^{t}$ , along with their features $\\pmb {f}_i^t$ and $\\pmb {f}_j^t$ , the merging process to form the next-layer cluster $c_k^{t + 1}$ and feature $\\pmb {f}_k^{t + 1}$ is as follows:", + "bbox": [ + 212, + 619, + 787, + 792 + ], + "page_idx": 6 + }, + { + "type": "equation", + "text": "\n$$\n\\begin{array}{l} c _ {k} ^ {t + 1} = c _ {i} ^ {t} \\cup c _ {j} ^ {t} \\quad \\mathrm {i f} \\operatorname {r a n k} (\\operatorname {s i m} (\\pmb {f} _ {i} ^ {t}, \\pmb {f} _ {j} ^ {t})) \\leq K \\mathrm {a n d} \\operatorname {d i s t} (c _ {i} ^ {t}, c _ {j} ^ {t}) \\leq T, \\\\ \\boldsymbol {f} _ {k} ^ {t + 1} = \\mathrm {F U} (c _ {i} ^ {t} \\cup c _ {j} ^ {t}), \\\\ \\end{array}\n$$\n", + "text_format": "latex", + "bbox": [ + 276, + 803, + 728, + 844 + ], + "page_idx": 6 + }, + { + "type": "header", + "text": "Part2Object", + "bbox": [ + 645, + 114, + 730, + 128 + ], + "page_idx": 6 + }, + { + "type": "page_number", + "text": "7", + "bbox": [ + 774, + 114, + 785, + 126 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "where $\\mathrm{rank}(\\cdot)$ denotes the ranking of feature similarity among pairwise cluster similarities in this clustering layer, where a higher rank signifies a higher similarity, and $\\mathrm{dist}(\\cdot,\\cdot)$ computes the Euclidean distance between closest points of the two clusters. The $K$ and $T$ represent the threshold values for ranking and distance, respectively. However, we observe that solely using the merging metric in Equ 3 can result in the incorrect merging of parts from different objects in the early stages of clustering (i.e., shallow layers of clustering). This occurs because clusters at the shallow layers are typically scattered fragments with highly local features, which hinder them from perceiving the entire object. To track this issue, we first extract 3D objectness priors, represented by a set of 3D bounding boxes of potential objects $B^{3\\mathrm{D}}$ , detailed in Sec 3.2. Then, we design a stopping criteria based on the 3D objectness priors to prevent clusters belonging to different objects from being merged. Criteria: Stop merging clusters belonging to different objects! Specifically, for any pair of clusters $c_{i}^{t}$ and $c_{j}^{t}$ that meet the merging metric, we extra assess their spatial relationship relative to each 3D object $b_{k}^{3\\mathrm{D}} \\in B^{3\\mathrm{D}}$ . If clusters $c_{i}^{t}$ and $c_{j}^{t}$ are respectively inside and outside the object $b_{k}^{3\\mathrm{D}}$ , then reject their merging because they do not belong to the same object.", + "bbox": [ + 212, + 146, + 787, + 419 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Collect Objects and Parts from Clustering. By employing hierarchical clustering with the stopping criteria, we identify clusters that stop merging with others as 3D objects, denoted as $\\hat{P}_{\\mathrm{object}}$ . Additionally, in hierarchical clustering, objects with complex geometric are typically merged from scattered fragments into meaningful parts, and then into complete objects. As shown in Fig 1, scattered \"toilet\" fragments are combined into object parts such as the toilet back and seat (see layer two), which are then formed into the object \"toilet\" (layer three). Therefore, we can trace back from objects $\\hat{P}_{\\mathrm{object}}$ to identify the clusters from the previous level that form them, considering these clusters as potential object parts, denoted as $\\hat{P}_{\\mathrm{part}}$ . Both objects $\\hat{P}_{\\mathrm{object}}$ and object parts $\\hat{P}_{\\mathrm{part}}$ identified in hierarchical clustering serve as pseudo-labels for training our instance segmentation model, Hi-Mask3D (Sec 3.3).", + "bbox": [ + 212, + 420, + 787, + 602 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "3.2 3D Objectness Priors From 2D Frames", + "text_level": 1, + "bbox": [ + 215, + 621, + 583, + 636 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "In this section, we introduce our grouping-first-then-projection pipeline to extract 3D objectness priors $B^{3\\mathrm{D}}$ from 2D RGB frames [12], where $B^{3\\mathrm{D}}$ is the set of 3D bounding boxes for potential objects in the scene.", + "bbox": [ + 212, + 643, + 785, + 688 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Grouping-first-then-projection Pipeline. To extract 3D objectness priors, a simple strategy involves projection-first-then-grouping: projecting 2D unsupervised instance segmentation masks [46, 60] from 2D frames onto the 3D point cloud to acquire 3D masks, then grouping and merging these 3D masks based on their spatial relationships to generate 3D objects, as shown in Fig 3(c). However, due to variations in object sizes and corresponding camera poses of 2D frames across different scenes, it is nearly impossible to establish a unified grouping criterion for different objects. If the grouping criterion is too strict, allowing only masks with high overlap to be grouped, masks from multiple parts of a large object cannot be merged (see Fig 3c-\"strict\"). Conversely, if the criterion", + "bbox": [ + 212, + 689, + 787, + 840 + ], + "page_idx": 7 + }, + { + "type": "page_number", + "text": "8", + "bbox": [ + 217, + 114, + 228, + 126 + ], + "page_idx": 7 + }, + { + "type": "header", + "text": "Cheng et al.", + "bbox": [ + 271, + 114, + 354, + 128 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/2cfae4e55033010216d29745460a386c2922614fd575d5d2edc95b255a06f0d8.jpg", + "image_caption": [ + "Fig. 3: Robustness of our cluster features and 3D objectness priors. (a) Visualization of the first 3 PCA components and our computed weights of points in Equ 2. (b) Visualization of 2D DINO features' PCA components, 3D points' normal vectors, and our 3D object priors. The comparison between (c) the projection-first-then-grouping pipeline and (d) our grouping-first-then-projection pipeline." + ], + "image_footnote": [], + "bbox": [ + 218, + 224, + 784, + 383 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "is relaxed, masks that should belong to multiple objects may be merged into a single object, e.g., the \"chair\" and \"sofa\" in Fig 3c-\"relaxed\". We attribute this limitation to the inability to rely solely on 3D spatial relationships and 3D point features (see Fig 3b) to determine if multiple 3D masks belong to the same object. To tackle this issue, we propose a grouping-first-then-projection pipeline: 1) it co-segments objects across multiple 2D frames by exploiting the temporal consistency of the frames, naturally identifying and grouping multiple 2D masks for each single object. 2) It projects the 2D masks corresponding to each object onto the 3D point cloud and then directly merges them.", + "bbox": [ + 212, + 431, + 787, + 568 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "Object Co-segmentation from Consecutive 2D RGB Frames. Given that 2D RGB frames are captured consecutively, we leverage their temporal consistency to co-segment multi-frame 2D masks corresponding to the same objects. First, for each RGB frame $I_{m}$ , we apply the 2D unsupervised instance segmentation method MaskCut [60] to obtain the 2D masks of objects in it, denoted as $O_{m}$ . As MaskCut uses the image encoder of 2D self-supervised DINO [6] to extract image feature map, we employ the same encoder to extract features for objects $O_{m}$ by masked average pooling each object mask $o_{m}^{i} \\in O_{m}$ over the frame's feature map. Next, for any adjacent pair of frames $I_{m}$ and $I_{m+1}$ , we compute the pairwise similarity between their object masks $O_{m}$ and $O_{m+1}$ based on mask features. Then, for each object mask $o_{m}^{i} \\in O_{m}$ in frame $I_{m}$ , we find the object mask $o_{m+1}^{j^{*}}$ in frame $I_{m+1}$ with the highest similarity as its corresponding candidate mask as follows,", + "bbox": [ + 212, + 571, + 787, + 770 + ], + "page_idx": 8 + }, + { + "type": "equation", + "text": "\n$$\n\\begin{array}{l} o _ {m + 1} ^ {j ^ {*}} = \\operatorname * {a r g m a x} _ {i} \\operatorname * {s i m} (o _ {m + 1} ^ {j}, o _ {m} ^ {i}), \\\\ o _ {m + 1} ^ {\\jmath} \\in O _ {m + 1} \\tag {4} \\\\ g _ {m \\rightarrow m + 1} ^ {i} = o _ {m + 1} ^ {j ^ {*}} \\mathrm {i f} \\mathrm {s i m} (o _ {m + 1} ^ {j ^ {*}}, o _ {m} ^ {i}) > \\tau \\mathrm {\\bf e l s e N u l l}, \\\\ \\end{array}\n$$\n", + "text_format": "latex", + "bbox": [ + 310, + 787, + 785, + 844 + ], + "page_idx": 8 + }, + { + "type": "header", + "text": "Part2Object", + "bbox": [ + 645, + 114, + 730, + 128 + ], + "page_idx": 8 + }, + { + "type": "page_number", + "text": "9", + "bbox": [ + 774, + 114, + 785, + 126 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "where $\\mathrm{sim}(o_{m + 1}^j,o_m^i)$ is to compute the cosine similarity between mask features of object masks $o_{m + 1}^{j}$ and $o_{m}^{i}$ . The $o_{m + 1}^{j^{*}}$ in frame $I_{m + 1}$ is the candidate object mask of the object mask $o_{m}^{i}$ in frame $I_{m}$ . Notably, only when the similarity between the object mask $o_{m}^{i}$ and its candidate object mask $o_{m + 1}^{j^{*}}$ exceeds the threshold $\\pi$ , they are considered to belong to the same object. After determining the same objects between any adjacent frames, $g_{m\\rightarrow m + 1}^{i}$ , we propagate this \"sameness\" to all adjacent frames to identify the same objects across all frames. Consequently, we can gather objects and identify all 2D masks in different frames of each object.", + "bbox": [ + 212, + 145, + 787, + 272 + ], + "page_idx": 9 + }, + { + "type": "text", + "text": "3D Objectness Priors. Given that we have grouped 2D masks into different objects, for each object, we can directly project its 2D masks onto the 3D point cloud and seamlessly merge their 3D masks into a unified whole [64], forming the object's 3D mask. Furthermore, instead of directly utilizing the 3D masks of objects, we utilize the 3D bounding boxes $B^{3\\mathrm{D}}$ of their masks to guide hierarchical clustering. This decision is based on treating the collected 3D objects from 2D frames as objectness priors and leaving precise segmentation to the clustering process itself.", + "bbox": [ + 212, + 273, + 787, + 393 + ], + "page_idx": 9 + }, + { + "type": "text", + "text": "3.3 Hi-Mask3D and Learning Objective", + "text_level": 1, + "bbox": [ + 215, + 419, + 555, + 435 + ], + "page_idx": 9 + }, + { + "type": "text", + "text": "In this section, we introduce our 3D instance segmentation framework, HiMask3D. It extends and adapts the 3D instance segmentation model, Mask3D [42], to predict and utilize object parts to improve 3D instance segmentation.", + "bbox": [ + 212, + 448, + 792, + 493 + ], + "page_idx": 9 + }, + { + "type": "text", + "text": "Introduction to Mask3D. The 3D instance segmentation model Mask3D [42] mainly consists of three components: 1) a feature backbone, a sparse convolutional U-net [11] that takes a colored point cloud as input and yields multi-resolution feature matrices $\\{\\pmb{F}_r\\}$ , where each $\\pmb{F}_r \\in \\mathbb{R}^{M_r \\times C}$ of size $M_r$ . 2) A mask module predicts instance masks $B$ corresponding to instance queries $Q \\in \\mathbb{R}^{K \\times C}$ by computing the dot product between queries $Q$ and scene features $\\{\\pmb{F}_r\\}$ . 3) A query refinement module, a transformer decoder [54] that cross-attends $K$ instance queries $Q$ with multi-resolution feature matrices $\\{\\pmb{F}_r\\}$ . In the $r$ -th decoder layer of query refinement module, the cross-attention between instance queries $Q_r$ and feature matrix $\\pmb{F}_r$ is formulated as follows,", + "bbox": [ + 212, + 494, + 787, + 645 + ], + "page_idx": 9 + }, + { + "type": "equation", + "text": "\n$$\nQ _ {r} = \\operatorname {s o f t m a x} \\left(Q _ {r - 1} \\boldsymbol {F} _ {r} ^ {T} / \\sqrt {C}\\right) \\boldsymbol {F} _ {r}. \\tag {5}\n$$\n", + "text_format": "latex", + "bbox": [ + 375, + 657, + 785, + 676 + ], + "page_idx": 9 + }, + { + "type": "text", + "text": "Note that for simplicity of demonstration, we omit the QKV projection [54] and intermediate instance mask $B$ [42] in the cross-attention formulation.", + "bbox": [ + 212, + 686, + 785, + 718 + ], + "page_idx": 9 + }, + { + "type": "text", + "text": "Hi-Mask3D with Hierarchical-Aware Detector. Our Hi-Mask3D modifies the query refinement module of Mask3D to adapt for both object and object part segmentation. Specifically, we start with two sets of queries $Q^o$ and $Q^p$ representing object queries and object part queries, respectively. Next, in each decoder layer, we no longer directly perform cross-attention between object queries and feature matrix to update object queries. Instead, we perform hierarchical attention: 1) conducting cross-attention between object part queries and feature matrix, and 2) updating object queries using part queries. This enables explicit", + "bbox": [ + 212, + 719, + 787, + 840 + ], + "page_idx": 9 + }, + { + "type": "page_number", + "text": "10", + "bbox": [ + 217, + 114, + 235, + 126 + ], + "page_idx": 9 + }, + { + "type": "header", + "text": "Cheng et al.", + "bbox": [ + 271, + 114, + 354, + 128 + ], + "page_idx": 9 + }, + { + "type": "text", + "text": "interaction between object part queries and object queries, which helps to leverage the information from object parts to improve object segmentation. In the $r$ -th decoder layer, the hierarchical attention is formulated as follows:", + "bbox": [ + 212, + 146, + 782, + 191 + ], + "page_idx": 10 + }, + { + "type": "equation", + "text": "\n$$\nQ _ {r} ^ {p} = \\operatorname {s o f t m a x} \\left(Q _ {r - 1} ^ {p} \\boldsymbol {F} _ {r} ^ {T} / \\sqrt {C}\\right) \\boldsymbol {F} _ {r}, \\tag {6}\n$$\n", + "text_format": "latex", + "bbox": [ + 377, + 204, + 784, + 229 + ], + "page_idx": 10 + }, + { + "type": "equation", + "text": "\n$$\nQ _ {r} ^ {o} = \\operatorname {s o f t m a x} \\left(Q _ {r - 1} ^ {o} \\left(Q _ {r} ^ {p}\\right) ^ {T} / \\sqrt {C}\\right) Q _ {r} ^ {p}.\n$$\n", + "text_format": "latex", + "bbox": [ + 375, + 224, + 625, + 243 + ], + "page_idx": 10 + }, + { + "type": "text", + "text": "The last-layer part queries and object queries undergo the mask module to obtain object part and part segmentation results, $P_{\\mathrm{part}}$ and $P_{\\mathrm{object}}$ , respectively.", + "bbox": [ + 212, + 253, + 782, + 282 + ], + "page_idx": 10 + }, + { + "type": "text", + "text": "Training Objective. Following Mask3D [42], we adopt bipartite graph algorithm [5, 48] to match prediction $P_{\\mathrm{object}}$ and $P_{\\mathrm{part}}$ with pseudo-labels $\\hat{P}_{\\mathrm{object}}$ and $\\hat{P}_{\\mathrm{part}}$ , respectively. To supervise segmentation masks, we combine DICE [14] loss and Binary Cross Entropy (BCE) loss [29]. Similar to [41], we also employ iterative rounds of self-training to enhance the segmentation capabilities of HiMask3D. During self-training, predictions from the previous round are utilized as pseudo-labels for the subsequent round.", + "bbox": [ + 212, + 284, + 784, + 391 + ], + "page_idx": 10 + }, + { + "type": "text", + "text": "4 Experiment", + "text_level": 1, + "bbox": [ + 215, + 412, + 366, + 430 + ], + "page_idx": 10 + }, + { + "type": "text", + "text": "Datasets and Implementations. We evaluate our method on four public indoor point cloud datasets, including ScanNet [12], Scannet200 [40], S3DIS [4] and Replica [49]. Following Mask3D [42], we employ a Sparse CNN, Res16UNet34C [11], implemented in MinkowskiEngine [11], as the point cloud encoder. Our HiMask3D is trained for 600 epochs with a learning rate 1e-4 and a batch size 4. Experiments are conducted using PyTorch [35] and executed on NVIDIA Tesla A40 GPUs. We evaluate the performance of class-agnostic instance segmentation using standard average precision scores. The scores consist of mAP@25, mAP@50, and mAP, representing performance metrics at IoU thresholds of 0.25, 0.5, and mean average precision, respectively. Hyper-parameter $T = 0.05$ in Equ 3 and $\\tau = 0.3$ in Equ 4.", + "bbox": [ + 212, + 445, + 802, + 612 + ], + "page_idx": 10 + }, + { + "type": "text", + "text": "4.1 Comparison with State-of-the-Art Unsupervised Methods", + "text_level": 1, + "bbox": [ + 212, + 633, + 736, + 648 + ], + "page_idx": 10 + }, + { + "type": "text", + "text": "Table 1 shows the results on ScanNet dataset [12] under two different settings: Traning-free Setting. In the training-free setting, all methods rely solely on clustering [6] or graph-cut [46, 60] algorithms for prediction, as no annotations are available for use. Table 1a compares our method, Part2Object, with the previous state-of-the-art method, Unscene3D [41], which also leverages DINO features but applies graph-cut algorithms on 3D point clouds. Compared to Unscene3D, our approach demonstrates significant enhancements across all metrics, with increases of $35.2\\%$ , $16.8\\%$ , and $6.7\\%$ in mAP@25, mAP@50, and mAP, respectively. These improvements are attributed to our utilization of 3D objectness priors, which guide the clustering process.", + "bbox": [ + 212, + 657, + 782, + 809 + ], + "page_idx": 10 + }, + { + "type": "text", + "text": "Data-efficient Fine-tuning Setting. In the data-efficient fine-tuning setting, all methods initially undergo either pretraining or self-training. Subsequently,", + "bbox": [ + 212, + 809, + 782, + 839 + ], + "page_idx": 10 + }, + { + "type": "header", + "text": "Part2Object", + "bbox": [ + 645, + 114, + 730, + 128 + ], + "page_idx": 10 + }, + { + "type": "page_number", + "text": "11", + "bbox": [ + 767, + 114, + 782, + 126 + ], + "page_idx": 10 + }, + { + "type": "table", + "img_path": "images/7a9ba7435b61df920ca62bd392a4ce3d518c3037d64eb9191d3db1aab9c19b13.jpg", + "table_caption": [ + "Table 1: Comparison of instance segmentation results on ScanNet val.", + "(a) Training-free Setting. All methods do not require gradient backpropagation to obtain prediction results.", + "(b) Data-efficient Setting with $\\mathbf{x}\\%$ available data annotations. A $0\\%$ value indicates that the method can provide pseudo-labels, and training is conducted on them." + ], + "table_footnote": [], + "table_body": "
w/o training MethodsScanNet Val [12]w/ training MethodsmAP@50 ScanNet Val [12]
mAP@25mAP@50mAP0%1%5%10%20%
HDBSCAN [32]32.15.51.6Scratch [42]/14.133.339.243.4
Nunes [33]30.57.32.3CSC [21]/22.139.943.848.9
Felzenswalb [16]38.912.75.0HDBSCAN [32]10.515.136.340.042.7
CutLER [60]7.00.20.3Felzenswalb [16]15.225.337.245.750.0
Unscene3D [41]19.910.05.9Unscene3D [41]23.228.446.855.760.7
Ours55.126.812.6Ours32.644.164.268.072.1
", + "bbox": [ + 215, + 210, + 792, + 354 + ], + "page_idx": 11 + }, + { + "type": "text", + "text": "they fine-tune on downstream data with varying percentages (0% represents pseudo-labels solely). Under this setting, the extraction of pseudo-labels is the primary requirement, followed by the necessity for models to exhibit strong learning capabilities. As demonstrated in Table 1b, Hi-Mask3D exhibits remarkable performance even with limited data, showcasing the robust 3D representation capabilities learned on the pseudo-labels from Part2object. Specifically, our method surpasses the previous SOTA by $15.7\\%$ and $11.4\\%$ when considering the utilization of only $1\\%$ and $20\\%$ of the available data, respectively.", + "bbox": [ + 212, + 366, + 787, + 488 + ], + "page_idx": 11 + }, + { + "type": "text", + "text": "4.2 Comparison on Cross-dataset Generalization", + "text_level": 1, + "bbox": [ + 215, + 510, + 630, + 526 + ], + "page_idx": 11 + }, + { + "type": "text", + "text": "Table 2 and Table 3 show our comparisons with fully-supervised Mask3D [42] on cross-dataset generalization, including both zero-shot and dataset-efficient settings. In both settings, our Hi-Mask3D is trained on pseudo-labels from our unsupervised Part2Object on the ScanNet dataset [12], while Mask3D is trained fully supervised on ScanNet.", + "bbox": [ + 212, + 536, + 787, + 612 + ], + "page_idx": 11 + }, + { + "type": "text", + "text": "Cross-dataset Zero-shot Generalization Setting. In Table 2, we compare unsupervised Hi-Mask3D with fully-supervised Mask3D on three downstream datasets, including ScanNet200 [39], S3DIS [4] and Replica [49]. Compared with the fully supervised Mask3D trained on ScanNet, our Hi-Mask3D trained on pseudo-labels from Part2Object achieves consistent performance improvement without using any manually annotated data. Hi-Mask3D surpasses fully-supervised Mask3D by $10.7\\%$ , $4.8\\%$ , and $6.4\\%$ in terms of mAP@50 on ScanNet200, S3DIS, and Replica datasets, respectively. Thanks to the learning from pseudo-labels in Part2Object, Hi-Mask3D has acquired more generalizable representations of 3D objects. Unlike fully-supervised methods, which are constrained to annotated classes in the training dataset, Hi-Mask3D demonstrates robust performance in cross-dataset generalization.", + "bbox": [ + 212, + 613, + 787, + 792 + ], + "page_idx": 11 + }, + { + "type": "text", + "text": "Cross-dataset Data-efficient Generalization Setting. In Table 3, we conduct data-efficient training on downstream datasets, such as ScanNet200 [39] and S3DIS [4], after pre-training our Hi-Mask3D on ScanNet's pseudo-labels ex", + "bbox": [ + 212, + 794, + 787, + 840 + ], + "page_idx": 11 + }, + { + "type": "page_number", + "text": "12", + "bbox": [ + 217, + 114, + 235, + 126 + ], + "page_idx": 11 + }, + { + "type": "header", + "text": "Cheng et al.", + "bbox": [ + 271, + 114, + 354, + 128 + ], + "page_idx": 11 + }, + { + "type": "table", + "img_path": "images/e32b36ef9531d7bcea8e4301ed1e2ae8fb45543f616de5d099b8f269c36b0fdb.jpg", + "table_caption": [ + "Table 2: Comparison on cross-dataset zero-shot generalization setting. We compare the zero-shot generalization ability of unsupervised Hi-Mask3D with fully-supervised Mask3D on three downstream datasets: ScanNet200, S3DIS and Replica." + ], + "table_footnote": [], + "table_body": "
Zero-shot MethodsScanNet200 Val [39]S3DIS 6-fold [4]Replica [49]
m@25m@50mAPm@25m@50mAPm@25m@50mAP
Mask3D [42]30.824.215.417.611.77.720.815.69.7
Ours63.234.916.324.516.58.536.522.011.2
", + "bbox": [ + 251, + 198, + 751, + 282 + ], + "page_idx": 12 + }, + { + "type": "table", + "img_path": "images/f3b751c476b6c6d460421b0d3d9531214f8139cc3a606a7b4f764fb7553ce98a.jpg", + "table_caption": [ + "Table 3: Comparison on cross-dataset data-efficient generalization setting. We utilize pseudo-labels from Part2Object to pre-train our Hi-Mask3D and conduct data-efficient finetuning on downstream datasets, comparing them with Mask3D." + ], + "table_footnote": [], + "table_body": "
MethodDatasetmAP@25mAP@50
1%5%10%20%1%5%10%20%
ScratchScanNet200 [39]2.122.530.640.21.38.218.422.9
OursScanNet200 [39]59.470.472.377.835.452.856.262.8
ScratchS3DIS [4]1.710.120.845.41.02.99.125.9
OursS3DIS [4]49.155.965.170.225.735.046.449.4
", + "bbox": [ + 251, + 348, + 736, + 467 + ], + "page_idx": 12 + }, + { + "type": "text", + "text": "tracted from Part2Object. For ScanNet200 and S3DIS, Hi-Mask3D outperforms Mask3D $47.4\\%$ and $24.7\\%$ mAP@50 with $20\\%$ data, respectively. When limited data is available, the improvement brought by pretraining on pseudo-labels from Part2Object is even greater, with the enhancement increasing from $37.6\\%$ to $57.3\\%$ as the available data decreases from $20\\%$ to $1\\%$ .", + "bbox": [ + 212, + 477, + 787, + 551 + ], + "page_idx": 12 + }, + { + "type": "text", + "text": "4.3 Ablation Study", + "text_level": 1, + "bbox": [ + 214, + 573, + 390, + 589 + ], + "page_idx": 12 + }, + { + "type": "text", + "text": "To evaluate the effectiveness of our hierarchical clustering Part2Object and 3D instance segmentation architecture Hi-Mask3D, as well as the effect of hyperparameters, we conduct comprehensive and detailed ablation experiments, as outlined in Table 4a and Table 4b.", + "bbox": [ + 212, + 598, + 785, + 657 + ], + "page_idx": 12 + }, + { + "type": "text", + "text": "Ablation on Part2Object Clustering. Table 4a illustrates the variants of Part2Object clustering, starting with a baseline utilizing the single-layer clustering algorithm. The performance of this baseline yields only $6.1\\%$ in mAP and $13.8\\%$ in mAP@50. This limitation arises due to the single-layer clustering being either too loose or too tight, leading to over-clustered or under-clustered objects. Subsequently, we remove the guidance of objectness priors from our proposed clustering algorithm, denoted as \"w/o OG\". Without object priors, while objects may appear at every layer, it becomes challenging to determine which layer they belong to. Furthermore, we replace our cluster feature computation function, \"FU(\\cdot)\", as simple averaged over features of points within the cluster, \"w/o FU\". The $2.0\\%$ decline in mAP highlights the importance of removing noise features for more robust cluster feature representations.", + "bbox": [ + 212, + 657, + 787, + 839 + ], + "page_idx": 12 + }, + { + "type": "header", + "text": "Part2Object", + "bbox": [ + 645, + 114, + 730, + 128 + ], + "page_idx": 12 + }, + { + "type": "page_number", + "text": "13", + "bbox": [ + 767, + 114, + 785, + 126 + ], + "page_idx": 12 + }, + { + "type": "text", + "text": "Table 4: Ablation study. We conduct ablation studies on ScanNet val. We validate the effectiveness of our Part2Object clustering, Hi-Mask3D, and self-training procedure.", + "bbox": [ + 215, + 143, + 785, + 172 + ], + "page_idx": 13 + }, + { + "type": "table", + "img_path": "images/f05d3d2b49e436f2e2e9e286de18319344b0eda29ba5ed82438c5d400427bf9e.jpg", + "table_caption": [ + "(a) Ablation study on Part2Object clustering and Hi-Mask3D architecture. OG, FU denotes objectness guidance and cluster feature function, respectively." + ], + "table_footnote": [], + "table_body": "
Ablation on ClusteringScanNet Val [12]
m@25m@50mAP
Baseline37.813.86.1
w/o OG43.220.710.4
w/o FU44.221.510.6
Ours55.126.812.6
Architecture
Mask3D59.031.014.2
Hi-Mask3D64.936.016.9
", + "bbox": [ + 233, + 227, + 464, + 388 + ], + "page_idx": 13 + }, + { + "type": "table", + "img_path": "images/e719ce9b95c8651cf03b8f4959e3d50976344adc3b8ede21a6ebdbd2f3f203fd.jpg", + "table_caption": [ + "(b) Ablation study on Self-training and Hyper-parameter." + ], + "table_footnote": [], + "table_body": "
Ablation on Self-TrainingScanNet Val [12]
m@25m@50mAP
Pseudo labels55.126.812.6
Round 160.732.615.3
Round 264.335.216.5
Round 364.936.016.9
Hyper-parameter
K = 0.455.722.610.3
K = 0.557.025.312.0
K = 0.655.126.812.6
K = 0.744.521.110.6
", + "bbox": [ + 504, + 208, + 764, + 396 + ], + "page_idx": 13 + }, + { + "type": "text", + "text": "Ablation on Hi-Mask3D. We compare our Hi-Mask3D with Mask3D using the same pseudo-labels extracted from Part2Object. Hi-Mask3D improves Mask3D by $2.7\\%$ and $5.0\\%$ in terms of mAP and mAP@50, respectively. This demonstrates that additional part information aids in object understanding. Furthermore, through reporting the performance of self-training, we observe that after several rounds, the mAP increases from $15.3\\%$ to $16.9\\%$ .", + "bbox": [ + 212, + 416, + 784, + 506 + ], + "page_idx": 13 + }, + { + "type": "text", + "text": "Ablation on Hyper-parameter. We conduct ablation on the import hyperparameter $K$ in Equ 3. Since the number of clusters at each layer varies, we use percentages (K=0.6 indicates the top 60%) to measure how many of the top clusters from each layer can be aggregated. Table 4b shows the performance of Part2Object at different $K$ . We use grey to indicate the default setting.", + "bbox": [ + 212, + 508, + 785, + 583 + ], + "page_idx": 13 + }, + { + "type": "text", + "text": "5 Conclusion and Limitations", + "text_level": 1, + "bbox": [ + 215, + 609, + 517, + 625 + ], + "page_idx": 13 + }, + { + "type": "text", + "text": "We introduce Part2Object, an efficient hierarchical clustering algorithm, that progressively groups point clouds into object parts and objects, while leveraging 3D object-ness prior to precisely target objects. Furthermore, we propose the Hierarchical-Aware Mask3D, facilitating self-training with pseudo-object and part labels from Part2Object. Experimental results demonstrate that we consistently surpass all existing methods in both unsupervised settings and data-efficient settings. Ethics Statement: Given that our 2D knowledge is derived from the self-supervised models DINO, we acknowledge that biases and controversies inherent in the training data for these models may be introduced into our model. Acknowledgment: This work was supported by the National Natural Science Foundation of China (No.62206174) and MoE Key Laboratory of Intelligent Perception and Human-Machine Collaboration (ShanghaiTech University).", + "bbox": [ + 212, + 643, + 787, + 840 + ], + "page_idx": 13 + }, + { + "type": "page_number", + "text": "14", + "bbox": [ + 217, + 114, + 235, + 126 + ], + "page_idx": 13 + }, + { + "type": "header", + "text": "Cheng et al.", + "bbox": [ + 271, + 114, + 354, + 128 + ], + "page_idx": 13 + }, + { + "type": "text", + "text": "References", + "text_level": 1, + "bbox": [ + 215, + 143, + 323, + 160 + ], + "page_idx": 14 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "1. Adams, R., Bischof, L.: Seeded region growing. IEEE Transactions on pattern analysis and machine intelligence 16(6), 641-647 (1994)", + "2. Amir, S., Gandelsman, Y., Bagon, S., Dekel, T.: Deep vit features as dense visual descriptors. arXiv preprint arXiv:2112.05814 (2021)", + "3. An, D., Wang, H., Wang, W., Wang, Z., Huang, Y., He, K., Wang, L.: Etpnav: Evolving topological planning for vision-language navigation in continuous environments. arXiv preprint arXiv:2304.03047 (2023)", + "4. Armeni, I., Sener, O., Zamir, A.R., Jiang, H., Brilakis, I., Fischer, M., Savarese, S.: 3d semantic parsing of large-scale indoor spaces. In: Proceedings of the IEEE conference on computer vision and pattern recognition. pp. 1534-1543 (2016)", + "5. Carion, N., Massa, F., Synnaeve, G., Usunier, N., Kirillov, A., Zagoruyko, S.: End-to-end object detection with transformers. In: European conference on computer vision. pp. 213-229. Springer (2020)", + "6. Caron, M., Touvron, H., Misra, I., Jégou, H., Mairal, J., Bojanowski, P., Joulin, A.: Emerging properties in self-supervised vision transformers. In: Proceedings of the International Conference on Computer Vision (ICCV) (2021)", + "7. Chen, R., Liu, Y., Kong, L., Chen, N., Xinge, Z., Ma, Y., Liu, T., Wang, W.: Towards label-free scene understanding by vision foundation models. In: Thirty-seventh Conference on Neural Information Processing Systems (2023)", + "8. Chen, S., Fang, J., Zhang, Q., Liu, W., Wang, X.: Hierarchical aggregation for 3d instance segmentation. In: Proceedings of the IEEE/CVF International Conference on Computer Vision. pp. 15467-15476 (2021)", + "9. Chen, Z., Yin, K., Fisher, M., Chaudhuri, S., Zhang, H.: Bae-net: Branched autoencoder for shape co-segmentation. In: Proceedings of the IEEE/CVF International Conference on Computer Vision. pp. 8490-8499 (2019)", + "0. Chibane, J., Engelmann, F., Anh Tran, T., Pons-Moll, G.: Box2mask: Weakly supervised 3d semantic instance segmentation using bounding boxes. In: European Conference on Computer Vision. pp. 681-699. Springer (2022)", + "1. Choy, C., Gwak, J., Savarese, S.: 4d spatio-temporal convnets: Minkowski convolutional neural networks. In: Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition. pp. 3075-3084 (2019)", + "2. Dai, A., Chang, A.X., Savva, M., Halber, M., Funkhouser, T., Nießner, M.: Scannet: Richly-annotated 3d reconstructions of indoor scenes. In: Proc. Computer Vision and Pattern Recognition (CVPR), IEEE (2017)", + "3. Dai, Q., Yang, S.: Curriculum point prompting for weakly-supervised referring segmentation (2024)", + "4. Deng, R., Shen, C., Liu, S., Wang, H., Liu, X.: Learning to predict crisp boundaries. In: Proceedings of the European conference on computer vision (ECCV). pp. 562-578 (2018)", + "5. Ester, M., Kriegel, H.P., Sander, J., Xu, X.: A density-based algorithm for discovering clusters in large spatial databases with noise. In: Proceedings of the Second International Conference on Knowledge Discovery and Data Mining. p. 226-231. KDD'96, AAAI Press (1996)", + "6. Felzenszwalb, P.F., Huttenlocher, D.P.: Efficient graph-based image segmentation. International journal of computer vision 59, 167-181 (2004)", + "7. Geng, H., Xu, H., Zhao, C., Xu, C., Yi, L., Huang, S., Wang, H.: Gapartnet: Cross-category domain-generalizable object perception and manipulation via generalizable and actionable parts. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 7081-7091 (2023)" + ], + "bbox": [ + 225, + 172, + 784, + 839 + ], + "page_idx": 14 + }, + { + "type": "header", + "text": "Part2Object", + "bbox": [ + 645, + 114, + 730, + 128 + ], + "page_idx": 14 + }, + { + "type": "page_number", + "text": "15", + "bbox": [ + 767, + 114, + 785, + 126 + ], + "page_idx": 14 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "18. Ghiasi, G., Gu, X., Cui, Y., Lin, T.Y.: Open-vocabulary image segmentation. arXiv preprint arXiv:2112.12143 (2021)", + "19. Han, L., Zheng, T., Xu, L., Fang, L.: Occuseg: Occupancy-aware 3d instance segmentation. In: 2020 IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR). pp. 2937-2946 (2020). https://doi.org/10.1109/CVPR42600.2020.00301", + "20. Hou, J., Dai, A., Nießner, M.: 3d-sis: 3d semantic instance segmentation of rgb-d scans. In: Proceedings of the IEEE/CVF conference on computer vision and pattern recognition. pp. 4421-4430 (2019)", + "21. Hou, J., Graham, B., Nießner, M., Xie, S.: Exploring data-efficient 3d scene understanding with contrastive scene contexts. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 15587-15597 (2021)", + "22. Hu, Q., Yang, B., Xie, L., Rosa, S., Guo, Y., Wang, Z., Trigoni, N., Markham, A.: Randla-net: Efficient semantic segmentation of large-scale point clouds. In: Proceedings of the IEEE/CVF conference on computer vision and pattern recognition. pp. 11108-11117 (2020)", + "23. Huang, S., Wang, Z., Li, P., Jia, B., Liu, T., Zhu, Y., Liang, W., Zhu, S.C.: Diffusion-based generation, optimization, and planning in 3d scenes. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 16750-16761 (2023)", + "24. Huang, Z., Wu, X., Chen, X., Zhao, H., Zhu, L., Lasenby, J.: Openins3d: Snap and lookup for 3d open-vocabulary instance segmentation. arXiv preprint arXiv:2309.00616 (2023)", + "25. Hui, L., Tang, L., Shen, Y., Xie, J., Yang, J.: Learning superpoint graph cut for 3d instance segmentation. In: NeurIPS (2022)", + "26. Kirillov, A., Mintun, E., Ravi, N., Mao, H., Rolland, C., Gustafson, L., Xiao, T., Whitehead, S., Berg, A.C., Lo, W.Y., et al.: Segment anything. arXiv preprint arXiv:2304.02643 (2023)", + "27. Kolodiazhnyi, M., Rukhovich, D., Vorontsova, A., Konushin, A.: Top-down beats bottom-up in 3d instance segmentation (2023). https://doi.org/10.48550/ ARXIV.2302.02871, https://arxiv.org/abs/2302.02871", + "28. Li, B., Weinberger, K.Q., Belongie, S., Koltun, V., Ranftl, R.: Language-driven semantic segmentation. In: International Conference on Learning Representations (2022), https://openreview.net/forum?id=RriDjddCLN", + "29. Lin, T.Y., Goyal, P., Girshick, R., He, K., Dollar, P.: Focal loss for dense object detection. In: Proceedings of the IEEE international conference on computer vision. pp. 2980-2988 (2017)", + "30. Liu, J., Yu, M., Ni, B., Chen, Y.: Self-prediction for joint instance and semantic segmentation of point clouds. In: Computer Vision-ECCV 2020: 16th European Conference, Glasgow, UK, August 23-28, 2020, Proceedings, Part XXII 16. pp. 187-204. Springer (2020)", + "31. Liu, Y., Kong, L., Cen, J., Chen, R., Zhang, W., Pan, L., Chen, K., Liu, Z.: Segment any point cloud sequences by distilling vision foundation models. arXiv preprint arXiv:2306.09347 (2023)", + "32. McInnes, L., Healy, J.: Accelerated hierarchical density based clustering. In: 2017 IEEE International Conference on Data Mining Workshops (ICDMW). pp. 33-42. IEEE (2017)", + "33. Nunes, L., Chen, X., Marcuzzi, R., Osep, A., Leal-Taixe, L., Stachniss, C., Behley, J.: Unsupervised class-agnostic instance segmentation of 3d lidar data for autonomous vehicles. IEEE Robotics and Automation Letters 7(4), 8713-8720 (2022)" + ], + "bbox": [ + 215, + 146, + 785, + 839 + ], + "page_idx": 15 + }, + { + "type": "page_number", + "text": "16", + "bbox": [ + 217, + 114, + 235, + 126 + ], + "page_idx": 15 + }, + { + "type": "header", + "text": "Cheng et al.", + "bbox": [ + 271, + 114, + 354, + 128 + ], + "page_idx": 15 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "34. Papon, J., Abramov, A., Schoeler, M., Worgotter, F.: Voxel cloud connectivity segmentation-supervoxels for point clouds. In: Proceedings of the IEEE conference on computer vision and pattern recognition. pp. 2027-2034 (2013)", + "35. Paszke, A., Gross, S., Massa, F., Lerer, A., Bradbury, J., Chanan, G., Killeen, T., Lin, Z., Gimelshein, N., Antiga, L., et al.: Pytorch: An imperative style, high-performance deep learning library. Advances in neural information processing systems 32 (2019)", + "36. Peng, S., Genova, K., Jiang, C., Tagliasacchi, A., Pollefeys, M., Funkhouser, T., et al.: Openscene: 3d scene understanding with open vocabularies. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 815-824 (2023)", + "37. Qi, C.R., Yi, L., Su, H., Guibas, L.J.: Pointnet++: Deep hierarchical feature learning on point sets in a metric space. Advances in neural information processing systems 30 (2017)", + "38. Rethage, D., Wald, J., Sturm, J., Navab, N., Tombari, F.: Fully-convolutional point networks for large-scale point clouds. In: Proceedings of the European Conference on Computer Vision (ECCV). pp. 596-611 (2018)", + "39. Rozenberszki, D., Litany, O., Dai, A.: Language-grounded indoor 3d semantic segmentation in the wild. In: European Conference on Computer Vision. pp. 125-141. Springer (2022)", + "40. Rozenberszki, D., Litany, O., Dai, A.: Language-grounded indoor 3d semantic segmentation in the wild. In: Proceedings of the European Conference on Computer Vision (ECCV) (2022)", + "41. Rozenberszki, D., Litany, O., Dai, A.: Unscene3d: Unsupervised 3d instance segmentation for indoor scenes. arXiv preprint arXiv:2303.14541 (2023)", + "42. Schult, J., Engelmann, F., Hermans, A., Litany, O., Tang, S., Leibe, B.: Mask3D for 3D Semantic Instance Segmentation. In: International Conference on Robotics and Automation (ICRA) (2023)", + "43. Shi, C., Yang, S.: Edadet: Open-vocabulary object detection using early dense alignment. In: Proceedings of the IEEE/CVF International Conference on Computer Vision. pp. 15724-15734 (2023)", + "44. Shi, C., Yang, S.: Logoprompt: Synthetic text images can be good visual prompts for vision-language models. arXiv preprint arXiv:2309.01155 (2023)", + "45. Shi, C., Yang, S.: The devil is in the object boundary: Towards annotation-free instance segmentation using foundation models. In: The Twelfth International Conference on Learning Representations (2024), https://openreview.net/forum?id=4JbrdrHxYy", + "46. Shi, J., Malik, J.: Normalized cuts and image segmentation. IEEE Transactions on pattern analysis and machine intelligence 22(8), 888-905 (2000)", + "47. Song, Z., Yang, B.: Ogc: Unsupervised 3d object segmentation from rigid dynamics of point clouds. Advances in Neural Information Processing Systems 35, 30798-30812 (2022)", + "48. Stewart, R., Andriluka, M., Ng, A.Y.: End-to-end people detection in crowded scenes. In: Proceedings of the IEEE conference on computer vision and pattern recognition. pp. 2325-2333 (2016)", + "49. Straub, J., Whelan, T., Ma, L., Chen, Y., Wijmans, E., Green, S., Engel, J.J., Mur-Artal, R., Ren, C., Verma, S., et al.: The replica dataset: A digital replica of indoor spaces. arXiv preprint arXiv:1906.05797 (2019)", + "50. Sun, J., Qing, C., Tan, J., Xu, X.: Superpoint transformer for 3d scene instance segmentation (2022)" + ], + "bbox": [ + 212, + 146, + 787, + 840 + ], + "page_idx": 16 + }, + { + "type": "header", + "text": "Part2Object", + "bbox": [ + 645, + 114, + 730, + 128 + ], + "page_idx": 16 + }, + { + "type": "page_number", + "text": "17", + "bbox": [ + 767, + 114, + 785, + 126 + ], + "page_idx": 16 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "51. Suo, S., Wong, K., Xu, J., Tu, J., Cui, A., Casas, S., Urtasun, R.: Mixsim: A hierarchical framework for mixed reality traffic simulation. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 9622-9631 (2023)", + "52. Suomela, L., Kalliola, J., Dag, A., Edelman, H., Kämäräinen, J.K.: Benchmarking visual localization for autonomous navigation. In: Proceedings of the IEEE/CVF Winter Conference on Applications of Computer Vision. pp. 2945-2955 (2023)", + "53. Tang, J., Zheng, G., Shi, C., Yang, S.: Contrastive grouping with transformer for referring image segmentation. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 23570-23580 (2023)", + "54. Vaswani, A., Shazeer, N., Parmar, N., Uszkoreit, J., Jones, L., Gomez, A.N., Kaiser, L., Polosukhin, I.: Attention is all you need. Advances in neural information processing systems 30 (2017)", + "55. Vu, T., Kim, K., Luu, T.M., Nguyen, T., Kim, J., Yoo, C.D.: Softgroup++: Scalable 3d instance segmentation with octree pyramid grouping. arXiv preprint arXiv:2209.08263 (2022)", + "56. Vu, T., Kim, K., Luu, T.M., Nguyen, X.T., Yoo, C.D.: Softgroup for 3d instance segmentation on 3d point clouds. In: CVPR (2022)", + "57. Wang, R., Zhang, Y., Mao, J., Zhang, R., Cheng, C.Y., Wu, J.: Ikea-manual: Seeing shape assembly step by step. Advances in Neural Information Processing Systems 35, 28428-28440 (2022)", + "58. Wang, W., Yu, R., Huang, Q., Neumann, U.: Sgpn: Similarity group proposal network for 3d point cloud instance segmentation. In: Proceedings of the IEEE conference on computer vision and pattern recognition. pp. 2569-2578 (2018)", + "59. Wang, X., Yu, Z., De Mello, S., Kautz, J., Anandkumar, A., Shen, C., Alvarez, J.M.: FreeSOLO: Learning to segment objects without annotations. arXiv preprint arXiv:2202.12181 (2022)", + "60. Wang, X., Girdhar, R., Yu, S.X., Misra, I.: Cut and learn for unsupervised object detection and instance segmentation. arXiv preprint arXiv:2301.11320 (2023)", + "61. Wang, Y., Shen, X., Hu, S.X., Yuan, Y., Crowley, J.L., Vaufreydaz, D.: Self-supervised transformers for unsupervised object discovery using normalized cut. In: Conference on Computer Vision and Pattern Recognition. New Orleans, LA, USA (June 2022)", + "62. Wang, Y., Sun, Y., Liu, Z., Sarma, S.E., Bronstein, M.M., Solomon, J.M.: Dynamic graph cnn for learning on point clouds. Acm Transactions On Graphics (tog) 38(5), 1-12 (2019)", + "63. Wu, W., Qi, Z., Fuxin, L.: Pointconv: Deep convolutional networks on 3d point clouds. In: Proceedings of the IEEE/CVF Conference on computer vision and pattern recognition. pp. 9621-9630 (2019)", + "64. Yang, Y., Wu, X., He, T., Zhao, H., Liu, X.: Sam3d: Segment anything in 3d scenes. arXiv preprint arXiv:2306.03908 (2023)", + "65. Zhang, B., Wonka, P.: Point cloud instance segmentation using probabilistic embeddings. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 8883-8892 (2021)", + "66. Zhang, Z., Ding, J., Jiang, L., Dai, D., Xia, G.S.: Freepoint: Unsupervised point cloud instance segmentation. arXiv preprint arXiv:2305.06973 (2023)", + "67. Zhang, Z., Yang, B., Wang, B., Li, B.: Growsp: Unsupervised semantic segmentation of 3d point clouds. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 17619-17629 (2023)" + ], + "bbox": [ + 215, + 146, + 787, + 825 + ], + "page_idx": 17 + }, + { + "type": "page_number", + "text": "18", + "bbox": [ + 217, + 114, + 235, + 126 + ], + "page_idx": 17 + }, + { + "type": "header", + "text": "Cheng et al.", + "bbox": [ + 271, + 114, + 354, + 128 + ], + "page_idx": 17 + } +] \ No newline at end of file diff --git a/2024/Part2Object_ Hierarchical Unsupervised 3D Instance Segmentation/918bb93b-c2e7-4e11-b966-4fcc81659f0e_model.json b/2024/Part2Object_ Hierarchical Unsupervised 3D Instance Segmentation/918bb93b-c2e7-4e11-b966-4fcc81659f0e_model.json new file mode 100644 index 0000000000000000000000000000000000000000..737ce82d45179f241856d292f2ede9a0c45a45f2 --- /dev/null +++ b/2024/Part2Object_ Hierarchical Unsupervised 3D Instance Segmentation/918bb93b-c2e7-4e11-b966-4fcc81659f0e_model.json @@ -0,0 +1,2238 @@ +[ + [ + { + "type": "title", + "bbox": [ + 0.255, + 0.141, + 0.75, + 0.186 + ], + "angle": 0, + "content": "Part2Object: Hierarchical Unsupervised 3D Instance Segmentation" + }, + { + "type": "text", + "bbox": [ + 0.321, + 0.213, + 0.681, + 0.244 + ], + "angle": 0, + "content": "Cheng Shi*, Yulin Zhang*, Bin Yang, Jiajin Tang, Yuexin Ma, and Sibei Yang" + }, + { + "type": "text", + "bbox": [ + 0.342, + 0.256, + 0.658, + 0.298 + ], + "angle": 0, + "content": "School of Information Science and Technology ShanghaiTech University {shicheng2022,yangsb}@shanghaiitech.edu.cn" + }, + { + "type": "text", + "bbox": [ + 0.263, + 0.338, + 0.74, + 0.56 + ], + "angle": 0, + "content": "Abstract. Unsupervised 3D instance segmentation aims to segment objects from a 3D point cloud without any annotations. Existing methods face the challenge of either too loose or too tight clustering, leading to under-segmentation or over-segmentation. To address this issue, we propose Part2Object, hierarchical clustering with object guidance. Part2Object employs multi-layer clustering from points to object parts and objects, allowing objects to manifest at any layer. Additionally, it extracts and utilizes 3D objectness priors from temporally consecutive 2D RGB frames to guide the clustering process. Moreover, we propose Hi-Mask3D to support hierarchical 3D object part and instance segmentation. By training Hi-Mask3D on the objects and object parts extracted from Part2Object, we achieve consistent and superior performance compared to state-of-the-art models in various settings, including unsupervised instance segmentation, data-efficient fine-tuning, and cross-dataset generalization. Code is release at https://github.com/ChengShiest/Part2object." + }, + { + "type": "text", + "bbox": [ + 0.263, + 0.574, + 0.694, + 0.588 + ], + "angle": 0, + "content": "Keywords: 3D Instance Segmentation \\(\\cdot\\) Unsupervised Learning" + }, + { + "type": "title", + "bbox": [ + 0.218, + 0.615, + 0.375, + 0.631 + ], + "angle": 0, + "content": "1 Introduction" + }, + { + "type": "text", + "bbox": [ + 0.218, + 0.649, + 0.787, + 0.815 + ], + "angle": 0, + "content": "3D instance segmentation, parsing semantic compositions from complex point cloud scenes into distinct objects, is fundamental and crucial in real-world applications, such as mixed reality [51], autonomous navigation [3, 52], planning [23], and manipulation [17]. In recent years, 3D instance segmentation has seen substantial advancements [19, 37, 38, 42, 55, 62], improving both the accuracy and efficiency of object localization and recognition. Nevertheless, it often demands large-scale human annotations for fully-supervised training, resulting in costly and manpower-intensive efforts. While a few studies [33, 47] have initiated to address 3D instance segmentation without human annotations, their reliance on 3D scene flows from sequential point clouds restricts their applicability to single point cloud scenarios, such as indoor scenes in ScanNet [12]." + }, + { + "type": "page_footnote", + "bbox": [ + 0.233, + 0.826, + 0.551, + 0.84 + ], + "angle": 0, + "content": "* Equal contribution. \\( \\boxtimes \\) Corresponding author." + } + ], + [ + { + "type": "page_number", + "bbox": [ + 0.218, + 0.116, + 0.23, + 0.127 + ], + "angle": 0, + "content": "2" + }, + { + "type": "header", + "bbox": [ + 0.272, + 0.115, + 0.355, + 0.129 + ], + "angle": 0, + "content": "Cheng et al." + }, + { + "type": "image_caption", + "bbox": [ + 0.214, + 0.145, + 0.788, + 0.201 + ], + "angle": 0, + "content": "Fig. 1: Motivation of our hierarchical clustering. Single-level clustering results in a trade-off between under-segmentation for certain objects and over-segmentation for others. In contrast, our hierarchical clustering allows for gathering and identifying objects at varying levels of clustering granularity." + }, + { + "type": "image", + "bbox": [ + 0.226, + 0.203, + 0.779, + 0.414 + ], + "angle": 0, + "content": null + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.445, + 0.788, + 0.642 + ], + "angle": 0, + "content": "In this paper, we study unsupervised 3D instance segmentation, aiming to segment class-agnostic instances from indoor 3D scenes without relying on any human labels. To address this problem, two straightforward strategies have been explored recently: 1) to apply traditional clustering [16, 32] or graph-cut methods [66] to group points into objects based on their RGB-D data, such as coordinates, colors, normal vectors, and self-supervised pretraining features [41]. 2) To project 2D instance segmentation results from 2D RGB frames onto the point cloud using the corresponding camera pose. Typically, 2D RGB frames are captured concurrently with indoor point cloud data [12, 39], while 2D instance masks are extracted using 2D unsupervised instance segmentation advancements [59-61]. However, both strategies can only work in straightforward scenes with few salient objects, as they cannot tackle the fundamental challenges inherent in unsupervised 3D instance segmentation, as outlined below." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.645, + 0.789, + 0.841 + ], + "angle": 0, + "content": "In complex scenes, achieving consistent segmentation granularity for different objects within a single clustering and graph-cut result is challenging, given the significant variations in geometric shape, color, and size among 3D objects, as well as the diverse compositions of these objects within scenes. This entails a trade-off between under-segmentation for certain objects and over-segmentation for others, making it almost impossible to attain satisfactory segmentation granularity for all of them. As shown in Fig 1, tighter grouping can accurately segment objects with simple geometric shapes and textures, such as the pillow on the bed (see layer one). However, it can also result in the over-segmentation of larger and more complex objects like the toilet in the same layer (layer one), and vice versa. In addition, there is a lack of an effective strategy to identify whether a point cluster or group represents a 3D instance. For example, the toilet back (see layer two) forms a well-grouped clus" + } + ], + [ + { + "type": "header", + "bbox": [ + 0.647, + 0.115, + 0.732, + 0.129 + ], + "angle": 0, + "content": "Part2Object" + }, + { + "type": "page_number", + "bbox": [ + 0.775, + 0.117, + 0.787, + 0.127 + ], + "angle": 0, + "content": "3" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.147, + 0.788, + 0.375 + ], + "angle": 0, + "content": "ter, but it does not constitute a meaningful object. The 3D geometric features are fundamentally semantically insufficient to identify objects. Recently, it has been discovered that self-supervised vision transformer (ViT) features contain information about object segmentations in 2D images [6, 60]. Building on this discovery, 3D unsupervised instance segmentation approaches [41] project the 2D self-supervised ViT features (or their inferred object segmentations) from 2D RGB frames onto 3D point clouds, thereby enhancing the perception of objectness for 3D point features. However, a 3D point typically corresponds to multiple pixels in different 2D RGB frames, and this many-to-one mapping is fragile, easily disrupted by noise features or imprecise masks on any frame, resulting in an inability to distinguish one object from others or the segmentation of objects into fragmented segments, as illustrated in Fig 3(b) and (c). In addition, despite enhanced point features, discerning nearby objects or those sharing similar semantics remains challenging due to the inherently greater complexity of 3D scenes compared to 2D images." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.381, + 0.788, + 0.76 + ], + "angle": 0, + "content": "To tackle these challenges, we introduce a simple yet effective principle, \"Gather & Aim\", for the automatic discovery and segmentation of 3D objects. For the gathering, inspired by the observation that tighter or looser groupings can respectively include certain objects within a scene, we propose allowing the gathering of potential objects at varying levels of grouping granularity, rather than being restricted to a single level. Specifically, we perform hierarchical clustering in the 3D scene, gradually grouping points into larger clusters, as illustrated in Fig 1. Therefore, instead of striving to improve single-layer clustering results, ours is expected that as long as each potential instance appears in one of the layers, we will be able to lock onto the segmentation mask of that instance. This mitigates excessive reliance on any single-level clustering results, thereby enhancing the clustering's adaptability to significant variations among objects and scenes. For the aiming, we propose to target the 3D objectness priors, a set of candidate objects, from temporally consecutive 2D RGB frames. We leverage both the temporal consistency across frames and the prior knowledge of unsupervised object discovery within frames to extract the 3D objectness priors. Note that we jointly estimate 2D objects from multiple frames first and then approximate object-level 3D bounding boxes, departing from the point-level projection used in previous methods [41], to address the fragility issue of the many-to-one mapping from pixels to points. It avoids adhesive or scattered 3D fragments caused by inconsistencies and occlusions over 2D RGB frames. Finally, the 3D bounding boxes serve as indicators to identify the 3D instance masks within the multi-level clusters. Our \"Gather & Aim\" strategy enables leverage of 2D semantic priors for 3D object identification while harnessing 3D geometric priors for precise instance segmentation." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.765, + 0.788, + 0.842 + ], + "angle": 0, + "content": "Furthermore, the taxonomy of how object parts are composed into an object is a strong yet disregarded prior information for unsupervised 3D instance segmentation. For example, the taxonomy based on the parts of a chair, including the back, arm, seat, and leg, can assist in addressing highly challenging scenarios, such as distinguishing two closely positioned and semantically similar" + } + ], + [ + { + "type": "page_number", + "bbox": [ + 0.218, + 0.116, + 0.23, + 0.127 + ], + "angle": 0, + "content": "4" + }, + { + "type": "header", + "bbox": [ + 0.272, + 0.115, + 0.355, + 0.129 + ], + "angle": 0, + "content": "Cheng et al." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.147, + 0.788, + 0.298 + ], + "angle": 0, + "content": "chairs as two separate instances. Thanks to our hierarchical clustering, we cannot only identify objects but also trace their constituent parts (see the toilet and its parts in layers three and two, respectively). Therefore, we extract both objects and object parts from our hierarchical clustering and name our clustering algorithm Part2Object. Expanding on this, we enhance the 3D instance segmentation framework, Mask3D [42], to support hierarchical 3D part and instance segmentation. Our improved model, named Hi-Mask3D, takes the object and object parts from Part2Object as pseudo-labels for learning 3D instance segmentation by incorporating explicit interactions between object parts and objects." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.298, + 0.788, + 0.389 + ], + "angle": 0, + "content": "To evaluate the effectiveness of our Part2Object clustering and Hi-Mask3D model, we conduct experiments on the challenging and cluttered indoor environments [4,12,39] in three settings: 1) direct evaluation on unsupervised instance segmentation, 2) application to data-efficient fine-tuning, and 3) cross-dataset generalization, all showing significant performance improvement. In summary, our contributions are multi-fold:" + }, + { + "type": "text", + "bbox": [ + 0.226, + 0.404, + 0.784, + 0.478 + ], + "angle": 0, + "content": "- We propose two key insights for unsupervised 3D instance segmentation: 1) Employing hierarchical clustering enables the gathering of objects at varying levels of clustering granularity. 2) Leveraging 3D objectness priors from temporally consecutive 2D frames as guidance, while harnessing 3D geometric priors to clustering on the point cloud for precise instance segmentation." + }, + { + "type": "text", + "bbox": [ + 0.226, + 0.48, + 0.785, + 0.567 + ], + "angle": 0, + "content": "- Based on our insights, we propose an innovative hierarchical clustering approach, Part2Object. It progressively groups points into object parts and objects while extracting and leveraging 3D objectness priors to guide the clustering process. Our Part2Object significantly outperforms the state-of-the-art training-free unsupervised 3D instance segmentation methods by \\(16.8\\%\\) mAP@50 on the ScanNet dataset." + }, + { + "type": "text", + "bbox": [ + 0.226, + 0.569, + 0.785, + 0.629 + ], + "angle": 0, + "content": "- We propose Hi-Mask3D, an extension of 3D instance segmentation to support hierarchical unsupervised 3D part and instance segmentation. Experiments demonstrate that Hi-Mask3D consistently and significantly outperforms state-of-the-art models in all the settings." + }, + { + "type": "title", + "bbox": [ + 0.216, + 0.655, + 0.388, + 0.67 + ], + "angle": 0, + "content": "2 Related Work" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.689, + 0.788, + 0.841 + ], + "angle": 0, + "content": "Unsupervised 3D instance segmentation. 3D instance segmentation is an essential task in 3D scene understanding, which aims to locate and recognize different objects in a 3D point cloud. However, the cost and labor intensity of 3D scene-level instance annotation underscore the significance of exploring challenging unsupervised instance segmentation methods. Early works utilize raw geometric information, including coordinates, colors, and normal vectors, to perform traditional clustering [15,16,32] for segmentation. Recently, inspired by the new paradigm in 2D unsupervised instance segmentation, some works [41,66] introduce pseudo labels and self-training to unsupervised 3D instance segmentation. Specifically, these methods use 2D or 3D self-supervised models to extract" + } + ], + [ + { + "type": "header", + "bbox": [ + 0.647, + 0.115, + 0.732, + 0.129 + ], + "angle": 0, + "content": "Part2Object" + }, + { + "type": "page_number", + "bbox": [ + 0.775, + 0.117, + 0.785, + 0.127 + ], + "angle": 0, + "content": "5" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.147, + 0.788, + 0.223 + ], + "angle": 0, + "content": "point cloud features, applying graph-cut algorithms to generate pseudo-labels for training and prediction. As illustrated earlier, the clustering method encounters a trade-off between under-segmentation and over-segmentation. Relying on a single graph-cut algorithm, as recent works do, falls short of achieving consistent granularity in complex indoor environments." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.223, + 0.788, + 0.42 + ], + "angle": 0, + "content": "Transfer 2D foundation models into 3D. The rapid advancements in 2D vision have given rise to powerful foundation models [2,6,13,18,26,28,43-45,53], proving beneficial across a variety of visual tasks. In contrast to the 2D domain, 3D vision encounters challenges with data scarcity and training limitations, hindering the development of foundation models. Consequently, several works in the 3D turn to leveraging 2D foundation models to address 3D problems. In 3D instance segmentation, some works [7,64] utilize 2D vision foundation models to generate pseudo-labels for 2D images, aiding in the training and prediction of 3D models through 2D to 3D projection algorithms. Other works [31,36] employ features rather than outputs obtained from 2D foundation models, enhancing the feature extraction capabilities of 3D models through pixel-point alignment. However, relying solely on 2D features or 2D pseudo-labels poses challenges in addressing cluttered indoor scenes with stacked objects." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.42, + 0.788, + 0.634 + ], + "angle": 0, + "content": "Supervised point cloud segmentation. Supervised point cloud segmentation can be divided into instance-level and part-level segmentation. The former has been significant breakthroughs in recent years [8, 19, 20, 22, 25, 27, 37, 38, 42, 50, 55, 56, 58, 62, 63]. However, these approaches demand a substantial amount of annotated data and focus on learning object-level attributes, neglecting the understanding of parts and their contributions to the composition of objects. In the realm of 3D part segmentation, various setups, including supervised [30,37,65], weakly supervised [9,10,57] instance segmentation and semantic segmentation, as well as open-world semantic segmentation [24, 36], have been extensively explored. While these approaches have demonstrated promising result, their primary emphasis lies in determining how to divide an object into parts. To integrate instance-level and part-level perspectives, we perform hierarchical clustering with the prior knowledge of how parts compose objects, unifying instance-level object identification and part-level constituent tracing." + }, + { + "type": "title", + "bbox": [ + 0.216, + 0.656, + 0.381, + 0.673 + ], + "angle": 0, + "content": "3 Methodology" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.689, + 0.788, + 0.795 + ], + "angle": 0, + "content": "Problem definition. Unsupervised 3D instance segmentation [41, 67] is the task of generating class-agnostic masks for all foreground objects in a 3D point cloud scene solely based on raw RGB-D data, without relying on any object detection or instance segmentation annotations. The raw RGB-D data [12, 39] consists of the point cloud \\( P \\in \\mathbb{R}^{N \\times 3} \\) and corresponding 2D RGB frames \\( \\{I_i\\}_{i=1}^M \\), where each image frame \\( I_m \\in \\mathbb{R}^{H \\times W \\times 3} \\). With raw RGB-D data, 3D instance segmentation task aims to output all object masks \\( P_{\\mathrm{object}} \\) in the scene." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.796, + 0.788, + 0.842 + ], + "angle": 0, + "content": "Method overview. As shown in Fig 2, our unsupervised framework has two major components: 1) the hierarchical clustering algorithm (Sec 3.1 and Sec 3.2), Part2Object, which progressively group points into larger clusters. Duringclus" + } + ], + [ + { + "type": "page_number", + "bbox": [ + 0.218, + 0.116, + 0.23, + 0.127 + ], + "angle": 0, + "content": "6" + }, + { + "type": "header", + "bbox": [ + 0.272, + 0.115, + 0.355, + 0.129 + ], + "angle": 0, + "content": "Cheng et al." + }, + { + "type": "image_caption", + "bbox": [ + 0.214, + 0.145, + 0.788, + 0.228 + ], + "angle": 0, + "content": "Fig. 2: Overview of our Part2Object hierarchical clustering and Hi-Mask3D instance segmentation framework. Part2Object extracts 3D objectness priors from consecutive 2D RGB frames and uses them to guide hierarchical clustering from points to object parts and objects. Hi-Mask3D utilizes objects and parts identified by Part2Object as pseudo-labels, learning for improving instance segmentation through the utilization of object parts." + }, + { + "type": "image", + "bbox": [ + 0.223, + 0.244, + 0.785, + 0.454 + ], + "angle": 0, + "content": null + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.498, + 0.788, + 0.62 + ], + "angle": 0, + "content": "tering, aided by objectness priors from 2D frames, we can effectively obtain both object-level and part-level segmentation masks, which serve as pseudo-labels for learning unsupervised instance segmentation. 2) The end-to-end 3D instance segmentation model Hi-Mask3D (Sec 3.3) extends the original Mask3D model [42] by explicitly predicting segmentation results of object parts and leveraging them to aid instance segmentation. By self-training on the pseudo-labels extracted from the first-stage Part2Object, Hi-Mask3D can effectively predict 3D instance segments." + }, + { + "type": "title", + "bbox": [ + 0.215, + 0.645, + 0.616, + 0.661 + ], + "angle": 0, + "content": "3.1 Hierarchical Clustering on 3D Point Cloud" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.674, + 0.79, + 0.841 + ], + "angle": 0, + "content": "In this section, we introduce the hierarchical clustering process of our Part2Object which starts by grouping initial points into initial clusters and then progressively merging them into larger clusters. We will sequentially cover: 1) the feature representation of points, 2) the initialization of clusters from points, and 3) the hierarchical clustering process for merging clusters. Point Feature Representation. Given a 3D surface point cloud \\( P \\), we denote its norm and color as \\( P^{\\mathrm{Norm}} \\) and \\( P^{\\mathrm{RGB}} \\), respectively. The 3D feature \\( f_{i}^{3\\mathrm{D}} \\) of point \\( p_i \\in P \\) comprises \\( p_i \\), \\( p_i^{\\mathrm{Norm}} \\), and \\( p_i^{\\mathrm{RGB}} \\), representing the location, color, and geometry information, respectively. Additionally, we define point cloud 2D semantics feature \\( f_{i} \\) by projecting 2D features [6] of RGB frames into 3D space using the corresponding camera pose. Here, 2D features are extracted using the 2D self-supervised" + } + ], + [ + { + "type": "header", + "bbox": [ + 0.647, + 0.115, + 0.732, + 0.129 + ], + "angle": 0, + "content": "Part2Object" + }, + { + "type": "page_number", + "bbox": [ + 0.775, + 0.116, + 0.787, + 0.127 + ], + "angle": 0, + "content": "7" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.147, + 0.788, + 0.221 + ], + "angle": 0, + "content": "method DINO [6], without rely on any 2D instance annotations. Cluster Initialization and Feature Representation. Given the vast number of points in the point cloud data [1], we initially apply the VCCS algorithm [34] to generate super-points on the point cloud, forming the first-layer clusters \\(\\{c_i^0\\}_{i = 1}^{N_0}\\) as follows:" + }, + { + "type": "equation", + "bbox": [ + 0.38, + 0.224, + 0.786, + 0.243 + ], + "angle": 0, + "content": "\\[\n\\{c _ {i} ^ {0} \\} _ {i = 1} ^ {N _ {0}} = \\mathrm {V C C S} (P, P ^ {\\mathrm {N o r m}}, P ^ {\\mathrm {R G B}}), \\tag {1}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.252, + 0.789, + 0.435 + ], + "angle": 0, + "content": "where VCCS algorithm takes points' coordinates \\( P \\), norms \\( P^{\\mathrm{Norm}} \\), and colors \\( P^{\\mathrm{RGB}} \\) as input and group original \\( N \\) points into \\( N_0 \\) super-points as \\( \\{c_i^0\\}_{i=1}^{N_0} \\). We consider each super-point \\( c_i^0 \\) as a cluster, consisting of multiple points that are closely positioned and share similar color and norm characteristics. Then, we define the feature representation for the cluster based on the features of the points within the cluster. Instead of directly averaging point features within the cluster, we compute a weight for each point and perform a weighted sum of features. These weights are determined by the similarity between each point's feature and the average feature of points within the cluster. This approach helps mitigate the influence of noisy points in the cluster, leading to more robust representations. Specifically, for the cluster \\( c_i^0 \\), its cluster feature \\( f_i^0 \\) is computed as follows," + }, + { + "type": "equation", + "bbox": [ + 0.282, + 0.447, + 0.786, + 0.492 + ], + "angle": 0, + "content": "\\[\n\\boldsymbol {f} _ {i} ^ {0} = \\sum_ {p _ {j} \\in c _ {i} ^ {0}} \\frac {\\operatorname {s i m} \\left(\\boldsymbol {f} _ {j} , \\bar {\\boldsymbol {f}} _ {i} ^ {0}\\right)}{\\sum p _ {j} \\in c _ {j} ^ {0} \\operatorname {s i m} \\left(\\boldsymbol {f} _ {j} , \\bar {\\boldsymbol {f}} _ {i} ^ {0}\\right)} \\boldsymbol {f} _ {j}, \\text {w h e r e} \\bar {\\boldsymbol {f}} _ {i} ^ {0} = \\sum_ {p _ {j} \\in c _ {i} ^ {0}} \\frac {1}{\\left| c _ {i} ^ {0} \\right|} \\boldsymbol {f} _ {j}. \\tag {2}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.506, + 0.788, + 0.619 + ], + "angle": 0, + "content": "The \\( p_j \\in c_i^0 \\) denotes each point \\( p_j \\) within the cluster \\( c_i^0 \\), and \\( \\pmb{f}_j \\) is the feature of point \\( p_j \\). The \\( \\bar{\\pmb{f}}_i^0 \\) is the average feature of points within \\( c_i^0 \\), and \\( \\sin(\\cdot, \\cdot) \\) denotes the computation of cosine similarity. As depicted in the example in Fig 3(a), our cluster feature \\( \\pmb{f}_i^0 \\) demonstrates greater robustness compared to \\( \\bar{\\pmb{f}}_i^0 \\). For simplicity, we abbreviate the operations in Equ 2 as function \\( \\mathsf{FU}(\\cdot) \\), which computes the cluster feature for the inputted cluster based on point features, e.g., \\( \\pmb{f}_i^0 = \\mathsf{FU}(c_i^0) \\)." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.62, + 0.789, + 0.793 + ], + "angle": 0, + "content": "Hierarchical Clustering with One Stop Criteria. Next, we group and merge the first-layer clusters \\(\\{c_i^0\\}_{i = 1}^{N_0}\\) to the next-layer clusters \\(\\{c_i^1\\}_{i = 1}^{N_1}\\) based on their features \\(\\{\\pmb {f}_i^0\\}_{i = 0}^{N_0}\\) and spatial coordinates, progressively and iteratively forming higher-hierarchical clusters \\(\\{c_{ij}^{t}\\}_{i = 1}^{N_t}\\), where \\(t\\) represents the \\(t\\)-th layer in hierarchical clustering. In each single clustering layer, our clustering principle is that two clusters can only be merged when they are semantically similar and spatially adjacent: 1) the feature similarity between them ranks among the top \\(K\\) similarities between any pair of clusters. 2) The closest points between two clusters are adjacent. Specifically, when considering two clusters in the \\(t\\)-th layer, \\(c_{i}^{t}\\) and \\(c_{j}^{t}\\), along with their features \\(\\pmb {f}_i^t\\) and \\(\\pmb {f}_j^t\\), the merging process to form the next-layer cluster \\(c_k^{t + 1}\\) and feature \\(\\pmb {f}_k^{t + 1}\\) is as follows:" + }, + { + "type": "equation", + "bbox": [ + 0.277, + 0.804, + 0.73, + 0.845 + ], + "angle": 0, + "content": "\\[\n\\begin{array}{l} c _ {k} ^ {t + 1} = c _ {i} ^ {t} \\cup c _ {j} ^ {t} \\quad \\mathrm {i f} \\operatorname {r a n k} (\\operatorname {s i m} (\\pmb {f} _ {i} ^ {t}, \\pmb {f} _ {j} ^ {t})) \\leq K \\mathrm {a n d} \\operatorname {d i s t} (c _ {i} ^ {t}, c _ {j} ^ {t}) \\leq T, \\\\ \\boldsymbol {f} _ {k} ^ {t + 1} = \\mathrm {F U} (c _ {i} ^ {t} \\cup c _ {j} ^ {t}), \\\\ \\end{array}\n\\]" + } + ], + [ + { + "type": "page_number", + "bbox": [ + 0.218, + 0.116, + 0.23, + 0.127 + ], + "angle": 0, + "content": "8" + }, + { + "type": "header", + "bbox": [ + 0.272, + 0.115, + 0.355, + 0.129 + ], + "angle": 0, + "content": "Cheng et al." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.147, + 0.788, + 0.42 + ], + "angle": 0, + "content": "where \\(\\mathrm{rank}(\\cdot)\\) denotes the ranking of feature similarity among pairwise cluster similarities in this clustering layer, where a higher rank signifies a higher similarity, and \\(\\mathrm{dist}(\\cdot,\\cdot)\\) computes the Euclidean distance between closest points of the two clusters. The \\(K\\) and \\(T\\) represent the threshold values for ranking and distance, respectively. However, we observe that solely using the merging metric in Equ 3 can result in the incorrect merging of parts from different objects in the early stages of clustering (i.e., shallow layers of clustering). This occurs because clusters at the shallow layers are typically scattered fragments with highly local features, which hinder them from perceiving the entire object. To track this issue, we first extract 3D objectness priors, represented by a set of 3D bounding boxes of potential objects \\(B^{3\\mathrm{D}}\\), detailed in Sec 3.2. Then, we design a stopping criteria based on the 3D objectness priors to prevent clusters belonging to different objects from being merged. Criteria: Stop merging clusters belonging to different objects! Specifically, for any pair of clusters \\(c_{i}^{t}\\) and \\(c_{j}^{t}\\) that meet the merging metric, we extra assess their spatial relationship relative to each 3D object \\(b_{k}^{3\\mathrm{D}} \\in B^{3\\mathrm{D}}\\). If clusters \\(c_{i}^{t}\\) and \\(c_{j}^{t}\\) are respectively inside and outside the object \\(b_{k}^{3\\mathrm{D}}\\), then reject their merging because they do not belong to the same object." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.421, + 0.788, + 0.603 + ], + "angle": 0, + "content": "Collect Objects and Parts from Clustering. By employing hierarchical clustering with the stopping criteria, we identify clusters that stop merging with others as 3D objects, denoted as \\(\\hat{P}_{\\mathrm{object}}\\). Additionally, in hierarchical clustering, objects with complex geometric are typically merged from scattered fragments into meaningful parts, and then into complete objects. As shown in Fig 1, scattered \"toilet\" fragments are combined into object parts such as the toilet back and seat (see layer two), which are then formed into the object \"toilet\" (layer three). Therefore, we can trace back from objects \\(\\hat{P}_{\\mathrm{object}}\\) to identify the clusters from the previous level that form them, considering these clusters as potential object parts, denoted as \\(\\hat{P}_{\\mathrm{part}}\\). Both objects \\(\\hat{P}_{\\mathrm{object}}\\) and object parts \\(\\hat{P}_{\\mathrm{part}}\\) identified in hierarchical clustering serve as pseudo-labels for training our instance segmentation model, Hi-Mask3D (Sec 3.3)." + }, + { + "type": "title", + "bbox": [ + 0.216, + 0.622, + 0.584, + 0.637 + ], + "angle": 0, + "content": "3.2 3D Objectness Priors From 2D Frames" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.644, + 0.787, + 0.689 + ], + "angle": 0, + "content": "In this section, we introduce our grouping-first-then-projection pipeline to extract 3D objectness priors \\( B^{3\\mathrm{D}} \\) from 2D RGB frames [12], where \\( B^{3\\mathrm{D}} \\) is the set of 3D bounding boxes for potential objects in the scene." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.69, + 0.788, + 0.841 + ], + "angle": 0, + "content": "Grouping-first-then-projection Pipeline. To extract 3D objectness priors, a simple strategy involves projection-first-then-grouping: projecting 2D unsupervised instance segmentation masks [46, 60] from 2D frames onto the 3D point cloud to acquire 3D masks, then grouping and merging these 3D masks based on their spatial relationships to generate 3D objects, as shown in Fig 3(c). However, due to variations in object sizes and corresponding camera poses of 2D frames across different scenes, it is nearly impossible to establish a unified grouping criterion for different objects. If the grouping criterion is too strict, allowing only masks with high overlap to be grouped, masks from multiple parts of a large object cannot be merged (see Fig 3c-\"strict\"). Conversely, if the criterion" + } + ], + [ + { + "type": "header", + "bbox": [ + 0.647, + 0.115, + 0.732, + 0.129 + ], + "angle": 0, + "content": "Part2Object" + }, + { + "type": "page_number", + "bbox": [ + 0.775, + 0.116, + 0.787, + 0.127 + ], + "angle": 0, + "content": "9" + }, + { + "type": "image_caption", + "bbox": [ + 0.215, + 0.145, + 0.788, + 0.219 + ], + "angle": 0, + "content": "Fig. 3: Robustness of our cluster features and 3D objectness priors. (a) Visualization of the first 3 PCA components and our computed weights of points in Equ 2. (b) Visualization of 2D DINO features' PCA components, 3D points' normal vectors, and our 3D object priors. The comparison between (c) the projection-first-then-grouping pipeline and (d) our grouping-first-then-projection pipeline." + }, + { + "type": "image", + "bbox": [ + 0.22, + 0.226, + 0.785, + 0.384 + ], + "angle": 0, + "content": null + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.433, + 0.788, + 0.569 + ], + "angle": 0, + "content": "is relaxed, masks that should belong to multiple objects may be merged into a single object, e.g., the \"chair\" and \"sofa\" in Fig 3c-\"relaxed\". We attribute this limitation to the inability to rely solely on 3D spatial relationships and 3D point features (see Fig 3b) to determine if multiple 3D masks belong to the same object. To tackle this issue, we propose a grouping-first-then-projection pipeline: 1) it co-segments objects across multiple 2D frames by exploiting the temporal consistency of the frames, naturally identifying and grouping multiple 2D masks for each single object. 2) It projects the 2D masks corresponding to each object onto the 3D point cloud and then directly merges them." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.572, + 0.788, + 0.771 + ], + "angle": 0, + "content": "Object Co-segmentation from Consecutive 2D RGB Frames. Given that 2D RGB frames are captured consecutively, we leverage their temporal consistency to co-segment multi-frame 2D masks corresponding to the same objects. First, for each RGB frame \\( I_{m} \\), we apply the 2D unsupervised instance segmentation method MaskCut [60] to obtain the 2D masks of objects in it, denoted as \\( O_{m} \\). As MaskCut uses the image encoder of 2D self-supervised DINO [6] to extract image feature map, we employ the same encoder to extract features for objects \\( O_{m} \\) by masked average pooling each object mask \\( o_{m}^{i} \\in O_{m} \\) over the frame's feature map. Next, for any adjacent pair of frames \\( I_{m} \\) and \\( I_{m+1} \\), we compute the pairwise similarity between their object masks \\( O_{m} \\) and \\( O_{m+1} \\) based on mask features. Then, for each object mask \\( o_{m}^{i} \\in O_{m} \\) in frame \\( I_{m} \\), we find the object mask \\( o_{m+1}^{j^{*}} \\) in frame \\( I_{m+1} \\) with the highest similarity as its corresponding candidate mask as follows," + }, + { + "type": "equation", + "bbox": [ + 0.312, + 0.788, + 0.786, + 0.845 + ], + "angle": 0, + "content": "\\[\n\\begin{array}{l} o _ {m + 1} ^ {j ^ {*}} = \\operatorname * {a r g m a x} _ {i} \\operatorname * {s i m} (o _ {m + 1} ^ {j}, o _ {m} ^ {i}), \\\\ o _ {m + 1} ^ {\\jmath} \\in O _ {m + 1} \\tag {4} \\\\ g _ {m \\rightarrow m + 1} ^ {i} = o _ {m + 1} ^ {j ^ {*}} \\mathrm {i f} \\mathrm {s i m} (o _ {m + 1} ^ {j ^ {*}}, o _ {m} ^ {i}) > \\tau \\mathrm {\\bf e l s e N u l l}, \\\\ \\end{array}\n\\]" + } + ], + [ + { + "type": "page_number", + "bbox": [ + 0.218, + 0.116, + 0.236, + 0.127 + ], + "angle": 0, + "content": "10" + }, + { + "type": "header", + "bbox": [ + 0.272, + 0.115, + 0.355, + 0.129 + ], + "angle": 0, + "content": "Cheng et al." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.146, + 0.788, + 0.273 + ], + "angle": 0, + "content": "where \\(\\mathrm{sim}(o_{m + 1}^j,o_m^i)\\) is to compute the cosine similarity between mask features of object masks \\(o_{m + 1}^{j}\\) and \\(o_{m}^{i}\\). The \\(o_{m + 1}^{j^{*}}\\) in frame \\(I_{m + 1}\\) is the candidate object mask of the object mask \\(o_{m}^{i}\\) in frame \\(I_{m}\\). Notably, only when the similarity between the object mask \\(o_{m}^{i}\\) and its candidate object mask \\(o_{m + 1}^{j^{*}}\\) exceeds the threshold \\(\\pi\\), they are considered to belong to the same object. After determining the same objects between any adjacent frames, \\(g_{m\\rightarrow m + 1}^{i}\\), we propagate this \"sameness\" to all adjacent frames to identify the same objects across all frames. Consequently, we can gather objects and identify all 2D masks in different frames of each object." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.274, + 0.788, + 0.395 + ], + "angle": 0, + "content": "3D Objectness Priors. Given that we have grouped 2D masks into different objects, for each object, we can directly project its 2D masks onto the 3D point cloud and seamlessly merge their 3D masks into a unified whole [64], forming the object's 3D mask. Furthermore, instead of directly utilizing the 3D masks of objects, we utilize the 3D bounding boxes \\( B^{3\\mathrm{D}} \\) of their masks to guide hierarchical clustering. This decision is based on treating the collected 3D objects from 2D frames as objectness priors and leaving precise segmentation to the clustering process itself." + }, + { + "type": "title", + "bbox": [ + 0.216, + 0.42, + 0.557, + 0.436 + ], + "angle": 0, + "content": "3.3 Hi-Mask3D and Learning Objective" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.449, + 0.793, + 0.494 + ], + "angle": 0, + "content": "In this section, we introduce our 3D instance segmentation framework, HiMask3D. It extends and adapts the 3D instance segmentation model, Mask3D [42], to predict and utilize object parts to improve 3D instance segmentation." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.495, + 0.788, + 0.646 + ], + "angle": 0, + "content": "Introduction to Mask3D. The 3D instance segmentation model Mask3D [42] mainly consists of three components: 1) a feature backbone, a sparse convolutional U-net [11] that takes a colored point cloud as input and yields multi-resolution feature matrices \\(\\{\\pmb{F}_r\\}\\), where each \\(\\pmb{F}_r \\in \\mathbb{R}^{M_r \\times C}\\) of size \\(M_r\\). 2) A mask module predicts instance masks \\(B\\) corresponding to instance queries \\(Q \\in \\mathbb{R}^{K \\times C}\\) by computing the dot product between queries \\(Q\\) and scene features \\(\\{\\pmb{F}_r\\}\\). 3) A query refinement module, a transformer decoder [54] that cross-attends \\(K\\) instance queries \\(Q\\) with multi-resolution feature matrices \\(\\{\\pmb{F}_r\\}\\). In the \\(r\\)-th decoder layer of query refinement module, the cross-attention between instance queries \\(Q_r\\) and feature matrix \\(\\pmb{F}_r\\) is formulated as follows," + }, + { + "type": "equation", + "bbox": [ + 0.377, + 0.658, + 0.786, + 0.677 + ], + "angle": 0, + "content": "\\[\nQ _ {r} = \\operatorname {s o f t m a x} \\left(Q _ {r - 1} \\boldsymbol {F} _ {r} ^ {T} / \\sqrt {C}\\right) \\boldsymbol {F} _ {r}. \\tag {5}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.688, + 0.787, + 0.719 + ], + "angle": 0, + "content": "Note that for simplicity of demonstration, we omit the QKV projection [54] and intermediate instance mask \\( B \\) [42] in the cross-attention formulation." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.72, + 0.788, + 0.841 + ], + "angle": 0, + "content": "Hi-Mask3D with Hierarchical-Aware Detector. Our Hi-Mask3D modifies the query refinement module of Mask3D to adapt for both object and object part segmentation. Specifically, we start with two sets of queries \\( Q^o \\) and \\( Q^p \\) representing object queries and object part queries, respectively. Next, in each decoder layer, we no longer directly perform cross-attention between object queries and feature matrix to update object queries. Instead, we perform hierarchical attention: 1) conducting cross-attention between object part queries and feature matrix, and 2) updating object queries using part queries. This enables explicit" + } + ], + [ + { + "type": "header", + "bbox": [ + 0.647, + 0.115, + 0.732, + 0.129 + ], + "angle": 0, + "content": "Part2Object" + }, + { + "type": "page_number", + "bbox": [ + 0.769, + 0.116, + 0.784, + 0.127 + ], + "angle": 0, + "content": "11" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.147, + 0.784, + 0.193 + ], + "angle": 0, + "content": "interaction between object part queries and object queries, which helps to leverage the information from object parts to improve object segmentation. In the \\( r \\)-th decoder layer, the hierarchical attention is formulated as follows:" + }, + { + "type": "equation", + "bbox": [ + 0.379, + 0.205, + 0.785, + 0.231 + ], + "angle": 0, + "content": "\\[\nQ _ {r} ^ {p} = \\operatorname {s o f t m a x} \\left(Q _ {r - 1} ^ {p} \\boldsymbol {F} _ {r} ^ {T} / \\sqrt {C}\\right) \\boldsymbol {F} _ {r}, \\tag {6}\n\\]" + }, + { + "type": "equation", + "bbox": [ + 0.377, + 0.226, + 0.627, + 0.244 + ], + "angle": 0, + "content": "\\[\nQ _ {r} ^ {o} = \\operatorname {s o f t m a x} \\left(Q _ {r - 1} ^ {o} \\left(Q _ {r} ^ {p}\\right) ^ {T} / \\sqrt {C}\\right) Q _ {r} ^ {p}.\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.254, + 0.784, + 0.284 + ], + "angle": 0, + "content": "The last-layer part queries and object queries undergo the mask module to obtain object part and part segmentation results, \\( P_{\\mathrm{part}} \\) and \\( P_{\\mathrm{object}} \\), respectively." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.285, + 0.785, + 0.392 + ], + "angle": 0, + "content": "Training Objective. Following Mask3D [42], we adopt bipartite graph algorithm [5, 48] to match prediction \\( P_{\\mathrm{object}} \\) and \\( P_{\\mathrm{part}} \\) with pseudo-labels \\( \\hat{P}_{\\mathrm{object}} \\) and \\( \\hat{P}_{\\mathrm{part}} \\), respectively. To supervise segmentation masks, we combine DICE [14] loss and Binary Cross Entropy (BCE) loss [29]. Similar to [41], we also employ iterative rounds of self-training to enhance the segmentation capabilities of HiMask3D. During self-training, predictions from the previous round are utilized as pseudo-labels for the subsequent round." + }, + { + "type": "title", + "bbox": [ + 0.216, + 0.414, + 0.367, + 0.431 + ], + "angle": 0, + "content": "4 Experiment" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.446, + 0.803, + 0.613 + ], + "angle": 0, + "content": "Datasets and Implementations. We evaluate our method on four public indoor point cloud datasets, including ScanNet [12], Scannet200 [40], S3DIS [4] and Replica [49]. Following Mask3D [42], we employ a Sparse CNN, Res16UNet34C [11], implemented in MinkowskiEngine [11], as the point cloud encoder. Our HiMask3D is trained for 600 epochs with a learning rate 1e-4 and a batch size 4. Experiments are conducted using PyTorch [35] and executed on NVIDIA Tesla A40 GPUs. We evaluate the performance of class-agnostic instance segmentation using standard average precision scores. The scores consist of mAP@25, mAP@50, and mAP, representing performance metrics at IoU thresholds of 0.25, 0.5, and mean average precision, respectively. Hyper-parameter \\( T = 0.05 \\) in Equ 3 and \\( \\tau = 0.3 \\) in Equ 4." + }, + { + "type": "title", + "bbox": [ + 0.214, + 0.634, + 0.738, + 0.65 + ], + "angle": 0, + "content": "4.1 Comparison with State-of-the-Art Unsupervised Methods" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.659, + 0.784, + 0.81 + ], + "angle": 0, + "content": "Table 1 shows the results on ScanNet dataset [12] under two different settings: Traning-free Setting. In the training-free setting, all methods rely solely on clustering [6] or graph-cut [46, 60] algorithms for prediction, as no annotations are available for use. Table 1a compares our method, Part2Object, with the previous state-of-the-art method, Unscene3D [41], which also leverages DINO features but applies graph-cut algorithms on 3D point clouds. Compared to Unscene3D, our approach demonstrates significant enhancements across all metrics, with increases of \\(35.2\\%\\), \\(16.8\\%\\), and \\(6.7\\%\\) in mAP@25, mAP@50, and mAP, respectively. These improvements are attributed to our utilization of 3D objectness priors, which guide the clustering process." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.81, + 0.784, + 0.84 + ], + "angle": 0, + "content": "Data-efficient Fine-tuning Setting. In the data-efficient fine-tuning setting, all methods initially undergo either pretraining or self-training. Subsequently," + } + ], + [ + { + "type": "page_number", + "bbox": [ + 0.218, + 0.116, + 0.236, + 0.127 + ], + "angle": 0, + "content": "12" + }, + { + "type": "header", + "bbox": [ + 0.272, + 0.115, + 0.355, + 0.129 + ], + "angle": 0, + "content": "Cheng et al." + }, + { + "type": "table_caption", + "bbox": [ + 0.232, + 0.145, + 0.768, + 0.159 + ], + "angle": 0, + "content": "Table 1: Comparison of instance segmentation results on ScanNet val." + }, + { + "type": "table_caption", + "bbox": [ + 0.217, + 0.165, + 0.495, + 0.197 + ], + "angle": 0, + "content": "(a) Training-free Setting. All methods do not require gradient backpropagation to obtain prediction results." + }, + { + "type": "table_caption", + "bbox": [ + 0.505, + 0.165, + 0.782, + 0.206 + ], + "angle": 0, + "content": "(b) Data-efficient Setting with \\( \\mathbf{x}\\% \\) available data annotations. A \\( 0\\% \\) value indicates that the method can provide pseudo-labels, and training is conducted on them." + }, + { + "type": "table", + "bbox": [ + 0.217, + 0.212, + 0.794, + 0.355 + ], + "angle": 0, + "content": "
w/o training MethodsScanNet Val [12]w/ training MethodsmAP@50 ScanNet Val [12]
mAP@25mAP@50mAP0%1%5%10%20%
HDBSCAN [32]32.15.51.6Scratch [42]/14.133.339.243.4
Nunes [33]30.57.32.3CSC [21]/22.139.943.848.9
Felzenswalb [16]38.912.75.0HDBSCAN [32]10.515.136.340.042.7
CutLER [60]7.00.20.3Felzenswalb [16]15.225.337.245.750.0
Unscene3D [41]19.910.05.9Unscene3D [41]23.228.446.855.760.7
Ours55.126.812.6Ours32.644.164.268.072.1
" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.367, + 0.788, + 0.489 + ], + "angle": 0, + "content": "they fine-tune on downstream data with varying percentages (0% represents pseudo-labels solely). Under this setting, the extraction of pseudo-labels is the primary requirement, followed by the necessity for models to exhibit strong learning capabilities. As demonstrated in Table 1b, Hi-Mask3D exhibits remarkable performance even with limited data, showcasing the robust 3D representation capabilities learned on the pseudo-labels from Part2object. Specifically, our method surpasses the previous SOTA by \\(15.7\\%\\) and \\(11.4\\%\\) when considering the utilization of only \\(1\\%\\) and \\(20\\%\\) of the available data, respectively." + }, + { + "type": "title", + "bbox": [ + 0.216, + 0.511, + 0.631, + 0.527 + ], + "angle": 0, + "content": "4.2 Comparison on Cross-dataset Generalization" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.537, + 0.788, + 0.613 + ], + "angle": 0, + "content": "Table 2 and Table 3 show our comparisons with fully-supervised Mask3D [42] on cross-dataset generalization, including both zero-shot and dataset-efficient settings. In both settings, our Hi-Mask3D is trained on pseudo-labels from our unsupervised Part2Object on the ScanNet dataset [12], while Mask3D is trained fully supervised on ScanNet." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.614, + 0.788, + 0.794 + ], + "angle": 0, + "content": "Cross-dataset Zero-shot Generalization Setting. In Table 2, we compare unsupervised Hi-Mask3D with fully-supervised Mask3D on three downstream datasets, including ScanNet200 [39], S3DIS [4] and Replica [49]. Compared with the fully supervised Mask3D trained on ScanNet, our Hi-Mask3D trained on pseudo-labels from Part2Object achieves consistent performance improvement without using any manually annotated data. Hi-Mask3D surpasses fully-supervised Mask3D by \\(10.7\\%\\), \\(4.8\\%\\), and \\(6.4\\%\\) in terms of mAP@50 on ScanNet200, S3DIS, and Replica datasets, respectively. Thanks to the learning from pseudo-labels in Part2Object, Hi-Mask3D has acquired more generalizable representations of 3D objects. Unlike fully-supervised methods, which are constrained to annotated classes in the training dataset, Hi-Mask3D demonstrates robust performance in cross-dataset generalization." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.795, + 0.788, + 0.842 + ], + "angle": 0, + "content": "Cross-dataset Data-efficient Generalization Setting. In Table 3, we conduct data-efficient training on downstream datasets, such as ScanNet200 [39] and S3DIS [4], after pre-training our Hi-Mask3D on ScanNet's pseudo-labels ex" + } + ], + [ + { + "type": "header", + "bbox": [ + 0.647, + 0.115, + 0.732, + 0.129 + ], + "angle": 0, + "content": "Part2Object" + }, + { + "type": "page_number", + "bbox": [ + 0.769, + 0.116, + 0.786, + 0.127 + ], + "angle": 0, + "content": "13" + }, + { + "type": "table_caption", + "bbox": [ + 0.214, + 0.145, + 0.788, + 0.187 + ], + "angle": 0, + "content": "Table 2: Comparison on cross-dataset zero-shot generalization setting. We compare the zero-shot generalization ability of unsupervised Hi-Mask3D with fully-supervised Mask3D on three downstream datasets: ScanNet200, S3DIS and Replica." + }, + { + "type": "table", + "bbox": [ + 0.252, + 0.199, + 0.753, + 0.284 + ], + "angle": 0, + "content": "
Zero-shot MethodsScanNet200 Val [39]S3DIS 6-fold [4]Replica [49]
m@25m@50mAPm@25m@50mAPm@25m@50mAP
Mask3D [42]30.824.215.417.611.77.720.815.69.7
Ours63.234.916.324.516.58.536.522.011.2
" + }, + { + "type": "table_caption", + "bbox": [ + 0.214, + 0.294, + 0.786, + 0.337 + ], + "angle": 0, + "content": "Table 3: Comparison on cross-dataset data-efficient generalization setting. We utilize pseudo-labels from Part2Object to pre-train our Hi-Mask3D and conduct data-efficient finetuning on downstream datasets, comparing them with Mask3D." + }, + { + "type": "table", + "bbox": [ + 0.252, + 0.349, + 0.738, + 0.468 + ], + "angle": 0, + "content": "
MethodDatasetmAP@25mAP@50
1%5%10%20%1%5%10%20%
ScratchScanNet200 [39]2.122.530.640.21.38.218.422.9
OursScanNet200 [39]59.470.472.377.835.452.856.262.8
ScratchS3DIS [4]1.710.120.845.41.02.99.125.9
OursS3DIS [4]49.155.965.170.225.735.046.449.4
" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.478, + 0.788, + 0.553 + ], + "angle": 0, + "content": "tracted from Part2Object. For ScanNet200 and S3DIS, Hi-Mask3D outperforms Mask3D \\(47.4\\%\\) and \\(24.7\\%\\) mAP@50 with \\(20\\%\\) data, respectively. When limited data is available, the improvement brought by pretraining on pseudo-labels from Part2Object is even greater, with the enhancement increasing from \\(37.6\\%\\) to \\(57.3\\%\\) as the available data decreases from \\(20\\%\\) to \\(1\\%\\)." + }, + { + "type": "title", + "bbox": [ + 0.215, + 0.574, + 0.391, + 0.59 + ], + "angle": 0, + "content": "4.3 Ablation Study" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.599, + 0.787, + 0.658 + ], + "angle": 0, + "content": "To evaluate the effectiveness of our hierarchical clustering Part2Object and 3D instance segmentation architecture Hi-Mask3D, as well as the effect of hyperparameters, we conduct comprehensive and detailed ablation experiments, as outlined in Table 4a and Table 4b." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.659, + 0.788, + 0.84 + ], + "angle": 0, + "content": "Ablation on Part2Object Clustering. Table 4a illustrates the variants of Part2Object clustering, starting with a baseline utilizing the single-layer clustering algorithm. The performance of this baseline yields only \\(6.1\\%\\) in mAP and \\(13.8\\%\\) in mAP@50. This limitation arises due to the single-layer clustering being either too loose or too tight, leading to over-clustered or under-clustered objects. Subsequently, we remove the guidance of objectness priors from our proposed clustering algorithm, denoted as \"w/o OG\". Without object priors, while objects may appear at every layer, it becomes challenging to determine which layer they belong to. Furthermore, we replace our cluster feature computation function, \"FU(\\cdot)\", as simple averaged over features of points within the cluster, \"w/o FU\". The \\(2.0\\%\\) decline in mAP highlights the importance of removing noise features for more robust cluster feature representations." + } + ], + [ + { + "type": "page_number", + "bbox": [ + 0.218, + 0.116, + 0.236, + 0.127 + ], + "angle": 0, + "content": "14" + }, + { + "type": "header", + "bbox": [ + 0.272, + 0.115, + 0.355, + 0.129 + ], + "angle": 0, + "content": "Cheng et al." + }, + { + "type": "table_caption", + "bbox": [ + 0.216, + 0.145, + 0.787, + 0.174 + ], + "angle": 0, + "content": "Table 4: Ablation study. We conduct ablation studies on ScanNet val. We validate the effectiveness of our Part2Object clustering, Hi-Mask3D, and self-training procedure." + }, + { + "type": "table_caption", + "bbox": [ + 0.216, + 0.182, + 0.493, + 0.224 + ], + "angle": 0, + "content": "(a) Ablation study on Part2Object clustering and Hi-Mask3D architecture. OG, FU denotes objectness guidance and cluster feature function, respectively." + }, + { + "type": "table", + "bbox": [ + 0.235, + 0.228, + 0.465, + 0.39 + ], + "angle": 0, + "content": "
Ablation on ClusteringScanNet Val [12]
m@25m@50mAP
Baseline37.813.86.1
w/o OG43.220.710.4
w/o FU44.221.510.6
Ours55.126.812.6
Architecture
Mask3D59.031.014.2
Hi-Mask3D64.936.016.9
" + }, + { + "type": "table_caption", + "bbox": [ + 0.507, + 0.182, + 0.782, + 0.204 + ], + "angle": 0, + "content": "(b) Ablation study on Self-training and Hyper-parameter." + }, + { + "type": "table", + "bbox": [ + 0.506, + 0.209, + 0.766, + 0.397 + ], + "angle": 0, + "content": "
Ablation on Self-TrainingScanNet Val [12]
m@25m@50mAP
Pseudo labels55.126.812.6
Round 160.732.615.3
Round 264.335.216.5
Round 364.936.016.9
Hyper-parameter
K = 0.455.722.610.3
K = 0.557.025.312.0
K = 0.655.126.812.6
K = 0.744.521.110.6
" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.417, + 0.785, + 0.507 + ], + "angle": 0, + "content": "Ablation on Hi-Mask3D. We compare our Hi-Mask3D with Mask3D using the same pseudo-labels extracted from Part2Object. Hi-Mask3D improves Mask3D by \\(2.7\\%\\) and \\(5.0\\%\\) in terms of mAP and mAP@50, respectively. This demonstrates that additional part information aids in object understanding. Furthermore, through reporting the performance of self-training, we observe that after several rounds, the mAP increases from \\(15.3\\%\\) to \\(16.9\\%\\)." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.509, + 0.787, + 0.584 + ], + "angle": 0, + "content": "Ablation on Hyper-parameter. We conduct ablation on the import hyperparameter \\( K \\) in Equ 3. Since the number of clusters at each layer varies, we use percentages (K=0.6 indicates the top 60%) to measure how many of the top clusters from each layer can be aggregated. Table 4b shows the performance of Part2Object at different \\( K \\). We use grey to indicate the default setting." + }, + { + "type": "title", + "bbox": [ + 0.216, + 0.61, + 0.519, + 0.626 + ], + "angle": 0, + "content": "5 Conclusion and Limitations" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.644, + 0.788, + 0.841 + ], + "angle": 0, + "content": "We introduce Part2Object, an efficient hierarchical clustering algorithm, that progressively groups point clouds into object parts and objects, while leveraging 3D object-ness prior to precisely target objects. Furthermore, we propose the Hierarchical-Aware Mask3D, facilitating self-training with pseudo-object and part labels from Part2Object. Experimental results demonstrate that we consistently surpass all existing methods in both unsupervised settings and data-efficient settings. Ethics Statement: Given that our 2D knowledge is derived from the self-supervised models DINO, we acknowledge that biases and controversies inherent in the training data for these models may be introduced into our model. Acknowledgment: This work was supported by the National Natural Science Foundation of China (No.62206174) and MoE Key Laboratory of Intelligent Perception and Human-Machine Collaboration (ShanghaiTech University)." + } + ], + [ + { + "type": "header", + "bbox": [ + 0.647, + 0.115, + 0.732, + 0.129 + ], + "angle": 0, + "content": "Part2Object" + }, + { + "type": "page_number", + "bbox": [ + 0.769, + 0.116, + 0.786, + 0.127 + ], + "angle": 0, + "content": "15" + }, + { + "type": "title", + "bbox": [ + 0.217, + 0.145, + 0.324, + 0.161 + ], + "angle": 0, + "content": "References" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.173, + 0.785, + 0.2 + ], + "angle": 0, + "content": "1. Adams, R., Bischof, L.: Seeded region growing. IEEE Transactions on pattern analysis and machine intelligence 16(6), 641-647 (1994)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.201, + 0.785, + 0.227 + ], + "angle": 0, + "content": "2. Amir, S., Gandelsman, Y., Bagon, S., Dekel, T.: Deep vit features as dense visual descriptors. arXiv preprint arXiv:2112.05814 (2021)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.228, + 0.785, + 0.268 + ], + "angle": 0, + "content": "3. An, D., Wang, H., Wang, W., Wang, Z., Huang, Y., He, K., Wang, L.: Etpnav: Evolving topological planning for vision-language navigation in continuous environments. arXiv preprint arXiv:2304.03047 (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.269, + 0.785, + 0.31 + ], + "angle": 0, + "content": "4. Armeni, I., Sener, O., Zamir, A.R., Jiang, H., Brilakis, I., Fischer, M., Savarese, S.: 3d semantic parsing of large-scale indoor spaces. In: Proceedings of the IEEE conference on computer vision and pattern recognition. pp. 1534-1543 (2016)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.31, + 0.785, + 0.35 + ], + "angle": 0, + "content": "5. Carion, N., Massa, F., Synnaeve, G., Usunier, N., Kirillov, A., Zagoruyko, S.: End-to-end object detection with transformers. In: European conference on computer vision. pp. 213-229. Springer (2020)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.35, + 0.785, + 0.391 + ], + "angle": 0, + "content": "6. Caron, M., Touvron, H., Misra, I., Jégou, H., Mairal, J., Bojanowski, P., Joulin, A.: Emerging properties in self-supervised vision transformers. In: Proceedings of the International Conference on Computer Vision (ICCV) (2021)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.391, + 0.785, + 0.432 + ], + "angle": 0, + "content": "7. Chen, R., Liu, Y., Kong, L., Chen, N., Xinge, Z., Ma, Y., Liu, T., Wang, W.: Towards label-free scene understanding by vision foundation models. In: Thirty-seventh Conference on Neural Information Processing Systems (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.433, + 0.785, + 0.472 + ], + "angle": 0, + "content": "8. Chen, S., Fang, J., Zhang, Q., Liu, W., Wang, X.: Hierarchical aggregation for 3d instance segmentation. In: Proceedings of the IEEE/CVF International Conference on Computer Vision. pp. 15467-15476 (2021)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.473, + 0.785, + 0.513 + ], + "angle": 0, + "content": "9. Chen, Z., Yin, K., Fisher, M., Chaudhuri, S., Zhang, H.: Bae-net: Branched autoencoder for shape co-segmentation. In: Proceedings of the IEEE/CVF International Conference on Computer Vision. pp. 8490-8499 (2019)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.513, + 0.785, + 0.554 + ], + "angle": 0, + "content": "0. Chibane, J., Engelmann, F., Anh Tran, T., Pons-Moll, G.: Box2mask: Weakly supervised 3d semantic instance segmentation using bounding boxes. In: European Conference on Computer Vision. pp. 681-699. Springer (2022)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.554, + 0.785, + 0.595 + ], + "angle": 0, + "content": "1. Choy, C., Gwak, J., Savarese, S.: 4d spatio-temporal convnets: Minkowski convolutional neural networks. In: Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition. pp. 3075-3084 (2019)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.595, + 0.785, + 0.636 + ], + "angle": 0, + "content": "2. Dai, A., Chang, A.X., Savva, M., Halber, M., Funkhouser, T., Nießner, M.: Scannet: Richly-annotated 3d reconstructions of indoor scenes. In: Proc. Computer Vision and Pattern Recognition (CVPR), IEEE (2017)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.636, + 0.785, + 0.663 + ], + "angle": 0, + "content": "3. Dai, Q., Yang, S.: Curriculum point prompting for weakly-supervised referring segmentation (2024)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.663, + 0.785, + 0.703 + ], + "angle": 0, + "content": "4. Deng, R., Shen, C., Liu, S., Wang, H., Liu, X.: Learning to predict crisp boundaries. In: Proceedings of the European conference on computer vision (ECCV). pp. 562-578 (2018)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.703, + 0.785, + 0.758 + ], + "angle": 0, + "content": "5. Ester, M., Kriegel, H.P., Sander, J., Xu, X.: A density-based algorithm for discovering clusters in large spatial databases with noise. In: Proceedings of the Second International Conference on Knowledge Discovery and Data Mining. p. 226-231. KDD'96, AAAI Press (1996)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.759, + 0.785, + 0.785 + ], + "angle": 0, + "content": "6. Felzenszwalb, P.F., Huttenlocher, D.P.: Efficient graph-based image segmentation. International journal of computer vision 59, 167-181 (2004)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.785, + 0.785, + 0.84 + ], + "angle": 0, + "content": "7. Geng, H., Xu, H., Zhao, C., Xu, C., Yi, L., Huang, S., Wang, H.: Gapartnet: Cross-category domain-generalizable object perception and manipulation via generalizable and actionable parts. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 7081-7091 (2023)" + }, + { + "type": "list", + "bbox": [ + 0.226, + 0.173, + 0.785, + 0.84 + ], + "angle": 0, + "content": null + } + ], + [ + { + "type": "page_number", + "bbox": [ + 0.218, + 0.116, + 0.236, + 0.127 + ], + "angle": 0, + "content": "16" + }, + { + "type": "header", + "bbox": [ + 0.272, + 0.115, + 0.355, + 0.129 + ], + "angle": 0, + "content": "Cheng et al." + }, + { + "type": "ref_text", + "bbox": [ + 0.218, + 0.147, + 0.785, + 0.175 + ], + "angle": 0, + "content": "18. Ghiasi, G., Gu, X., Cui, Y., Lin, T.Y.: Open-vocabulary image segmentation. arXiv preprint arXiv:2112.12143 (2021)" + }, + { + "type": "ref_text", + "bbox": [ + 0.218, + 0.176, + 0.787, + 0.23 + ], + "angle": 0, + "content": "19. Han, L., Zheng, T., Xu, L., Fang, L.: Occuseg: Occupancy-aware 3d instance segmentation. In: 2020 IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR). pp. 2937-2946 (2020). https://doi.org/10.1109/CVPR42600.2020.00301" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.231, + 0.785, + 0.272 + ], + "angle": 0, + "content": "20. Hou, J., Dai, A., Nießner, M.: 3d-sis: 3d semantic instance segmentation of rgb-d scans. In: Proceedings of the IEEE/CVF conference on computer vision and pattern recognition. pp. 4421-4430 (2019)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.273, + 0.785, + 0.314 + ], + "angle": 0, + "content": "21. Hou, J., Graham, B., Nießner, M., Xie, S.: Exploring data-efficient 3d scene understanding with contrastive scene contexts. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 15587-15597 (2021)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.315, + 0.785, + 0.369 + ], + "angle": 0, + "content": "22. Hu, Q., Yang, B., Xie, L., Rosa, S., Guo, Y., Wang, Z., Trigoni, N., Markham, A.: Randla-net: Efficient semantic segmentation of large-scale point clouds. In: Proceedings of the IEEE/CVF conference on computer vision and pattern recognition. pp. 11108-11117 (2020)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.37, + 0.785, + 0.424 + ], + "angle": 0, + "content": "23. Huang, S., Wang, Z., Li, P., Jia, B., Liu, T., Zhu, Y., Liang, W., Zhu, S.C.: Diffusion-based generation, optimization, and planning in 3d scenes. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 16750-16761 (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.425, + 0.785, + 0.466 + ], + "angle": 0, + "content": "24. Huang, Z., Wu, X., Chen, X., Zhao, H., Zhu, L., Lasenby, J.: Openins3d: Snap and lookup for 3d open-vocabulary instance segmentation. arXiv preprint arXiv:2309.00616 (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.467, + 0.785, + 0.494 + ], + "angle": 0, + "content": "25. Hui, L., Tang, L., Shen, Y., Xie, J., Yang, J.: Learning superpoint graph cut for 3d instance segmentation. In: NeurIPS (2022)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.495, + 0.785, + 0.535 + ], + "angle": 0, + "content": "26. Kirillov, A., Mintun, E., Ravi, N., Mao, H., Rolland, C., Gustafson, L., Xiao, T., Whitehead, S., Berg, A.C., Lo, W.Y., et al.: Segment anything. arXiv preprint arXiv:2304.02643 (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.536, + 0.785, + 0.577 + ], + "angle": 0, + "content": "27. Kolodiazhnyi, M., Rukhovich, D., Vorontsova, A., Konushin, A.: Top-down beats bottom-up in 3d instance segmentation (2023). https://doi.org/10.48550/ ARXIV.2302.02871, https://arxiv.org/abs/2302.02871" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.578, + 0.785, + 0.618 + ], + "angle": 0, + "content": "28. Li, B., Weinberger, K.Q., Belongie, S., Koltun, V., Ranftl, R.: Language-driven semantic segmentation. In: International Conference on Learning Representations (2022), https://openreview.net/forum?id=RriDjddCLN" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.619, + 0.785, + 0.66 + ], + "angle": 0, + "content": "29. Lin, T.Y., Goyal, P., Girshick, R., He, K., Dollar, P.: Focal loss for dense object detection. In: Proceedings of the IEEE international conference on computer vision. pp. 2980-2988 (2017)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.661, + 0.785, + 0.715 + ], + "angle": 0, + "content": "30. Liu, J., Yu, M., Ni, B., Chen, Y.: Self-prediction for joint instance and semantic segmentation of point clouds. In: Computer Vision-ECCV 2020: 16th European Conference, Glasgow, UK, August 23-28, 2020, Proceedings, Part XXII 16. pp. 187-204. Springer (2020)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.716, + 0.785, + 0.757 + ], + "angle": 0, + "content": "31. Liu, Y., Kong, L., Cen, J., Chen, R., Zhang, W., Pan, L., Chen, K., Liu, Z.: Segment any point cloud sequences by distilling vision foundation models. arXiv preprint arXiv:2306.09347 (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.758, + 0.785, + 0.799 + ], + "angle": 0, + "content": "32. McInnes, L., Healy, J.: Accelerated hierarchical density based clustering. In: 2017 IEEE International Conference on Data Mining Workshops (ICDMW). pp. 33-42. IEEE (2017)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.799, + 0.785, + 0.84 + ], + "angle": 0, + "content": "33. Nunes, L., Chen, X., Marcuzzi, R., Osep, A., Leal-Taixe, L., Stachniss, C., Behley, J.: Unsupervised class-agnostic instance segmentation of 3d lidar data for autonomous vehicles. IEEE Robotics and Automation Letters 7(4), 8713-8720 (2022)" + }, + { + "type": "list", + "bbox": [ + 0.217, + 0.147, + 0.787, + 0.84 + ], + "angle": 0, + "content": null + } + ], + [ + { + "type": "header", + "bbox": [ + 0.647, + 0.115, + 0.732, + 0.129 + ], + "angle": 0, + "content": "Part2Object" + }, + { + "type": "page_number", + "bbox": [ + 0.769, + 0.116, + 0.786, + 0.127 + ], + "angle": 0, + "content": "17" + }, + { + "type": "ref_text", + "bbox": [ + 0.214, + 0.147, + 0.788, + 0.189 + ], + "angle": 0, + "content": "34. Papon, J., Abramov, A., Schoeler, M., Worgotter, F.: Voxel cloud connectivity segmentation-supervoxels for point clouds. In: Proceedings of the IEEE conference on computer vision and pattern recognition. pp. 2027-2034 (2013)" + }, + { + "type": "ref_text", + "bbox": [ + 0.214, + 0.19, + 0.788, + 0.245 + ], + "angle": 0, + "content": "35. Paszke, A., Gross, S., Massa, F., Lerer, A., Bradbury, J., Chanan, G., Killeen, T., Lin, Z., Gimelshein, N., Antiga, L., et al.: Pytorch: An imperative style, high-performance deep learning library. Advances in neural information processing systems 32 (2019)" + }, + { + "type": "ref_text", + "bbox": [ + 0.214, + 0.246, + 0.788, + 0.3 + ], + "angle": 0, + "content": "36. Peng, S., Genova, K., Jiang, C., Tagliasacchi, A., Pollefeys, M., Funkhouser, T., et al.: Openscene: 3d scene understanding with open vocabularies. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 815-824 (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.214, + 0.3, + 0.788, + 0.342 + ], + "angle": 0, + "content": "37. Qi, C.R., Yi, L., Su, H., Guibas, L.J.: Pointnet++: Deep hierarchical feature learning on point sets in a metric space. Advances in neural information processing systems 30 (2017)" + }, + { + "type": "ref_text", + "bbox": [ + 0.214, + 0.342, + 0.788, + 0.384 + ], + "angle": 0, + "content": "38. Rethage, D., Wald, J., Sturm, J., Navab, N., Tombari, F.: Fully-convolutional point networks for large-scale point clouds. In: Proceedings of the European Conference on Computer Vision (ECCV). pp. 596-611 (2018)" + }, + { + "type": "ref_text", + "bbox": [ + 0.214, + 0.384, + 0.788, + 0.425 + ], + "angle": 0, + "content": "39. Rozenberszki, D., Litany, O., Dai, A.: Language-grounded indoor 3d semantic segmentation in the wild. In: European Conference on Computer Vision. pp. 125-141. Springer (2022)" + }, + { + "type": "ref_text", + "bbox": [ + 0.214, + 0.425, + 0.788, + 0.467 + ], + "angle": 0, + "content": "40. Rozenberszki, D., Litany, O., Dai, A.: Language-grounded indoor 3d semantic segmentation in the wild. In: Proceedings of the European Conference on Computer Vision (ECCV) (2022)" + }, + { + "type": "ref_text", + "bbox": [ + 0.214, + 0.467, + 0.788, + 0.495 + ], + "angle": 0, + "content": "41. Rozenberszki, D., Litany, O., Dai, A.: Unscene3d: Unsupervised 3d instance segmentation for indoor scenes. arXiv preprint arXiv:2303.14541 (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.214, + 0.495, + 0.788, + 0.536 + ], + "angle": 0, + "content": "42. Schult, J., Engelmann, F., Hermans, A., Litany, O., Tang, S., Leibe, B.: Mask3D for 3D Semantic Instance Segmentation. In: International Conference on Robotics and Automation (ICRA) (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.214, + 0.536, + 0.788, + 0.578 + ], + "angle": 0, + "content": "43. Shi, C., Yang, S.: Edadet: Open-vocabulary object detection using early dense alignment. In: Proceedings of the IEEE/CVF International Conference on Computer Vision. pp. 15724-15734 (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.214, + 0.578, + 0.788, + 0.606 + ], + "angle": 0, + "content": "44. Shi, C., Yang, S.: Logoprompt: Synthetic text images can be good visual prompts for vision-language models. arXiv preprint arXiv:2309.01155 (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.214, + 0.606, + 0.788, + 0.66 + ], + "angle": 0, + "content": "45. Shi, C., Yang, S.: The devil is in the object boundary: Towards annotation-free instance segmentation using foundation models. In: The Twelfth International Conference on Learning Representations (2024), https://openreview.net/forum?id=4JbrdrHxYy" + }, + { + "type": "ref_text", + "bbox": [ + 0.214, + 0.66, + 0.788, + 0.689 + ], + "angle": 0, + "content": "46. Shi, J., Malik, J.: Normalized cuts and image segmentation. IEEE Transactions on pattern analysis and machine intelligence 22(8), 888-905 (2000)" + }, + { + "type": "ref_text", + "bbox": [ + 0.214, + 0.689, + 0.788, + 0.73 + ], + "angle": 0, + "content": "47. Song, Z., Yang, B.: Ogc: Unsupervised 3d object segmentation from rigid dynamics of point clouds. Advances in Neural Information Processing Systems 35, 30798-30812 (2022)" + }, + { + "type": "ref_text", + "bbox": [ + 0.214, + 0.73, + 0.788, + 0.772 + ], + "angle": 0, + "content": "48. Stewart, R., Andriluka, M., Ng, A.Y.: End-to-end people detection in crowded scenes. In: Proceedings of the IEEE conference on computer vision and pattern recognition. pp. 2325-2333 (2016)" + }, + { + "type": "ref_text", + "bbox": [ + 0.214, + 0.772, + 0.788, + 0.813 + ], + "angle": 0, + "content": "49. Straub, J., Whelan, T., Ma, L., Chen, Y., Wijmans, E., Green, S., Engel, J.J., Mur-Artal, R., Ren, C., Verma, S., et al.: The replica dataset: A digital replica of indoor spaces. arXiv preprint arXiv:1906.05797 (2019)" + }, + { + "type": "ref_text", + "bbox": [ + 0.214, + 0.813, + 0.788, + 0.841 + ], + "angle": 0, + "content": "50. Sun, J., Qing, C., Tan, J., Xu, X.: Superpoint transformer for 3d scene instance segmentation (2022)" + }, + { + "type": "list", + "bbox": [ + 0.214, + 0.147, + 0.788, + 0.841 + ], + "angle": 0, + "content": null + } + ], + [ + { + "type": "page_number", + "bbox": [ + 0.218, + 0.116, + 0.236, + 0.127 + ], + "angle": 0, + "content": "18" + }, + { + "type": "header", + "bbox": [ + 0.272, + 0.115, + 0.355, + 0.129 + ], + "angle": 0, + "content": "Cheng et al." + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.147, + 0.788, + 0.203 + ], + "angle": 0, + "content": "51. Suo, S., Wong, K., Xu, J., Tu, J., Cui, A., Casas, S., Urtasun, R.: Mixsim: A hierarchical framework for mixed reality traffic simulation. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 9622-9631 (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.204, + 0.787, + 0.245 + ], + "angle": 0, + "content": "52. Suomela, L., Kalliola, J., Dag, A., Edelman, H., Kämäräinen, J.K.: Benchmarking visual localization for autonomous navigation. In: Proceedings of the IEEE/CVF Winter Conference on Applications of Computer Vision. pp. 2945-2955 (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.246, + 0.787, + 0.286 + ], + "angle": 0, + "content": "53. Tang, J., Zheng, G., Shi, C., Yang, S.: Contrastive grouping with transformer for referring image segmentation. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 23570-23580 (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.287, + 0.787, + 0.327 + ], + "angle": 0, + "content": "54. Vaswani, A., Shazeer, N., Parmar, N., Uszkoreit, J., Jones, L., Gomez, A.N., Kaiser, L., Polosukhin, I.: Attention is all you need. Advances in neural information processing systems 30 (2017)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.328, + 0.787, + 0.369 + ], + "angle": 0, + "content": "55. Vu, T., Kim, K., Luu, T.M., Nguyen, T., Kim, J., Yoo, C.D.: Softgroup++: Scalable 3d instance segmentation with octree pyramid grouping. arXiv preprint arXiv:2209.08263 (2022)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.37, + 0.787, + 0.397 + ], + "angle": 0, + "content": "56. Vu, T., Kim, K., Luu, T.M., Nguyen, X.T., Yoo, C.D.: Softgroup for 3d instance segmentation on 3d point clouds. In: CVPR (2022)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.398, + 0.787, + 0.438 + ], + "angle": 0, + "content": "57. Wang, R., Zhang, Y., Mao, J., Zhang, R., Cheng, C.Y., Wu, J.: Ikea-manual: Seeing shape assembly step by step. Advances in Neural Information Processing Systems 35, 28428-28440 (2022)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.439, + 0.787, + 0.481 + ], + "angle": 0, + "content": "58. Wang, W., Yu, R., Huang, Q., Neumann, U.: Sgpn: Similarity group proposal network for 3d point cloud instance segmentation. In: Proceedings of the IEEE conference on computer vision and pattern recognition. pp. 2569-2578 (2018)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.481, + 0.787, + 0.522 + ], + "angle": 0, + "content": "59. Wang, X., Yu, Z., De Mello, S., Kautz, J., Anandkumar, A., Shen, C., Alvarez, J.M.: FreeSOLO: Learning to segment objects without annotations. arXiv preprint arXiv:2202.12181 (2022)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.522, + 0.787, + 0.55 + ], + "angle": 0, + "content": "60. Wang, X., Girdhar, R., Yu, S.X., Misra, I.: Cut and learn for unsupervised object detection and instance segmentation. arXiv preprint arXiv:2301.11320 (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.55, + 0.787, + 0.605 + ], + "angle": 0, + "content": "61. Wang, Y., Shen, X., Hu, S.X., Yuan, Y., Crowley, J.L., Vaufreydaz, D.: Self-supervised transformers for unsupervised object discovery using normalized cut. In: Conference on Computer Vision and Pattern Recognition. New Orleans, LA, USA (June 2022)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.606, + 0.787, + 0.646 + ], + "angle": 0, + "content": "62. Wang, Y., Sun, Y., Liu, Z., Sarma, S.E., Bronstein, M.M., Solomon, J.M.: Dynamic graph cnn for learning on point clouds. Acm Transactions On Graphics (tog) 38(5), 1-12 (2019)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.646, + 0.787, + 0.688 + ], + "angle": 0, + "content": "63. Wu, W., Qi, Z., Fuxin, L.: Pointconv: Deep convolutional networks on 3d point clouds. In: Proceedings of the IEEE/CVF Conference on computer vision and pattern recognition. pp. 9621-9630 (2019)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.689, + 0.787, + 0.716 + ], + "angle": 0, + "content": "64. Yang, Y., Wu, X., He, T., Zhao, H., Liu, X.: Sam3d: Segment anything in 3d scenes. arXiv preprint arXiv:2306.03908 (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.717, + 0.787, + 0.757 + ], + "angle": 0, + "content": "65. Zhang, B., Wonka, P.: Point cloud instance segmentation using probabilistic embeddings. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 8883-8892 (2021)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.758, + 0.787, + 0.785 + ], + "angle": 0, + "content": "66. Zhang, Z., Ding, J., Jiang, L., Dai, D., Xia, G.S.: Freepoint: Unsupervised point cloud instance segmentation. arXiv preprint arXiv:2305.06973 (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.786, + 0.787, + 0.827 + ], + "angle": 0, + "content": "67. Zhang, Z., Yang, B., Wang, B., Li, B.: Growsp: Unsupervised semantic segmentation of 3d point clouds. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 17619-17629 (2023)" + }, + { + "type": "list", + "bbox": [ + 0.217, + 0.147, + 0.788, + 0.827 + ], + "angle": 0, + "content": null + } + ] +] \ No newline at end of file diff --git a/2024/Part2Object_ Hierarchical Unsupervised 3D Instance Segmentation/918bb93b-c2e7-4e11-b966-4fcc81659f0e_origin.pdf b/2024/Part2Object_ Hierarchical Unsupervised 3D Instance Segmentation/918bb93b-c2e7-4e11-b966-4fcc81659f0e_origin.pdf new file mode 100644 index 0000000000000000000000000000000000000000..948f69fec2455bcaa159e4b56ecc7a78c6d64421 --- /dev/null +++ b/2024/Part2Object_ Hierarchical Unsupervised 3D Instance Segmentation/918bb93b-c2e7-4e11-b966-4fcc81659f0e_origin.pdf @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:413df9b7b34150687da3f95d87b92ca217dd555017762428019c96c58dfa2726 +size 1645811 diff --git a/2024/Part2Object_ Hierarchical Unsupervised 3D Instance Segmentation/full.md b/2024/Part2Object_ Hierarchical Unsupervised 3D Instance Segmentation/full.md new file mode 100644 index 0000000000000000000000000000000000000000..4658b647405afce8833a22c8384e0450e13fc0cf --- /dev/null +++ b/2024/Part2Object_ Hierarchical Unsupervised 3D Instance Segmentation/full.md @@ -0,0 +1,267 @@ +# Part2Object: Hierarchical Unsupervised 3D Instance Segmentation + +Cheng Shi*, Yulin Zhang*, Bin Yang, Jiajin Tang, Yuexin Ma, and Sibei Yang + +School of Information Science and Technology ShanghaiTech University {shicheng2022,yangsb}@shanghaiitech.edu.cn + +Abstract. Unsupervised 3D instance segmentation aims to segment objects from a 3D point cloud without any annotations. Existing methods face the challenge of either too loose or too tight clustering, leading to under-segmentation or over-segmentation. To address this issue, we propose Part2Object, hierarchical clustering with object guidance. Part2Object employs multi-layer clustering from points to object parts and objects, allowing objects to manifest at any layer. Additionally, it extracts and utilizes 3D objectness priors from temporally consecutive 2D RGB frames to guide the clustering process. Moreover, we propose Hi-Mask3D to support hierarchical 3D object part and instance segmentation. By training Hi-Mask3D on the objects and object parts extracted from Part2Object, we achieve consistent and superior performance compared to state-of-the-art models in various settings, including unsupervised instance segmentation, data-efficient fine-tuning, and cross-dataset generalization. Code is release at https://github.com/ChengShiest/Part2object. + +Keywords: 3D Instance Segmentation $\cdot$ Unsupervised Learning + +# 1 Introduction + +3D instance segmentation, parsing semantic compositions from complex point cloud scenes into distinct objects, is fundamental and crucial in real-world applications, such as mixed reality [51], autonomous navigation [3, 52], planning [23], and manipulation [17]. In recent years, 3D instance segmentation has seen substantial advancements [19, 37, 38, 42, 55, 62], improving both the accuracy and efficiency of object localization and recognition. Nevertheless, it often demands large-scale human annotations for fully-supervised training, resulting in costly and manpower-intensive efforts. While a few studies [33, 47] have initiated to address 3D instance segmentation without human annotations, their reliance on 3D scene flows from sequential point clouds restricts their applicability to single point cloud scenarios, such as indoor scenes in ScanNet [12]. + +![](images/9e2c399d4d0af5c900b794d764d3a5c384bacf2840baf4a16c5ae60fa0c5a65a.jpg) +Fig. 1: Motivation of our hierarchical clustering. Single-level clustering results in a trade-off between under-segmentation for certain objects and over-segmentation for others. In contrast, our hierarchical clustering allows for gathering and identifying objects at varying levels of clustering granularity. + +In this paper, we study unsupervised 3D instance segmentation, aiming to segment class-agnostic instances from indoor 3D scenes without relying on any human labels. To address this problem, two straightforward strategies have been explored recently: 1) to apply traditional clustering [16, 32] or graph-cut methods [66] to group points into objects based on their RGB-D data, such as coordinates, colors, normal vectors, and self-supervised pretraining features [41]. 2) To project 2D instance segmentation results from 2D RGB frames onto the point cloud using the corresponding camera pose. Typically, 2D RGB frames are captured concurrently with indoor point cloud data [12, 39], while 2D instance masks are extracted using 2D unsupervised instance segmentation advancements [59-61]. However, both strategies can only work in straightforward scenes with few salient objects, as they cannot tackle the fundamental challenges inherent in unsupervised 3D instance segmentation, as outlined below. + +In complex scenes, achieving consistent segmentation granularity for different objects within a single clustering and graph-cut result is challenging, given the significant variations in geometric shape, color, and size among 3D objects, as well as the diverse compositions of these objects within scenes. This entails a trade-off between under-segmentation for certain objects and over-segmentation for others, making it almost impossible to attain satisfactory segmentation granularity for all of them. As shown in Fig 1, tighter grouping can accurately segment objects with simple geometric shapes and textures, such as the pillow on the bed (see layer one). However, it can also result in the over-segmentation of larger and more complex objects like the toilet in the same layer (layer one), and vice versa. In addition, there is a lack of an effective strategy to identify whether a point cluster or group represents a 3D instance. For example, the toilet back (see layer two) forms a well-grouped clus + +ter, but it does not constitute a meaningful object. The 3D geometric features are fundamentally semantically insufficient to identify objects. Recently, it has been discovered that self-supervised vision transformer (ViT) features contain information about object segmentations in 2D images [6, 60]. Building on this discovery, 3D unsupervised instance segmentation approaches [41] project the 2D self-supervised ViT features (or their inferred object segmentations) from 2D RGB frames onto 3D point clouds, thereby enhancing the perception of objectness for 3D point features. However, a 3D point typically corresponds to multiple pixels in different 2D RGB frames, and this many-to-one mapping is fragile, easily disrupted by noise features or imprecise masks on any frame, resulting in an inability to distinguish one object from others or the segmentation of objects into fragmented segments, as illustrated in Fig 3(b) and (c). In addition, despite enhanced point features, discerning nearby objects or those sharing similar semantics remains challenging due to the inherently greater complexity of 3D scenes compared to 2D images. + +To tackle these challenges, we introduce a simple yet effective principle, "Gather & Aim", for the automatic discovery and segmentation of 3D objects. For the gathering, inspired by the observation that tighter or looser groupings can respectively include certain objects within a scene, we propose allowing the gathering of potential objects at varying levels of grouping granularity, rather than being restricted to a single level. Specifically, we perform hierarchical clustering in the 3D scene, gradually grouping points into larger clusters, as illustrated in Fig 1. Therefore, instead of striving to improve single-layer clustering results, ours is expected that as long as each potential instance appears in one of the layers, we will be able to lock onto the segmentation mask of that instance. This mitigates excessive reliance on any single-level clustering results, thereby enhancing the clustering's adaptability to significant variations among objects and scenes. For the aiming, we propose to target the 3D objectness priors, a set of candidate objects, from temporally consecutive 2D RGB frames. We leverage both the temporal consistency across frames and the prior knowledge of unsupervised object discovery within frames to extract the 3D objectness priors. Note that we jointly estimate 2D objects from multiple frames first and then approximate object-level 3D bounding boxes, departing from the point-level projection used in previous methods [41], to address the fragility issue of the many-to-one mapping from pixels to points. It avoids adhesive or scattered 3D fragments caused by inconsistencies and occlusions over 2D RGB frames. Finally, the 3D bounding boxes serve as indicators to identify the 3D instance masks within the multi-level clusters. Our "Gather & Aim" strategy enables leverage of 2D semantic priors for 3D object identification while harnessing 3D geometric priors for precise instance segmentation. + +Furthermore, the taxonomy of how object parts are composed into an object is a strong yet disregarded prior information for unsupervised 3D instance segmentation. For example, the taxonomy based on the parts of a chair, including the back, arm, seat, and leg, can assist in addressing highly challenging scenarios, such as distinguishing two closely positioned and semantically similar + +chairs as two separate instances. Thanks to our hierarchical clustering, we cannot only identify objects but also trace their constituent parts (see the toilet and its parts in layers three and two, respectively). Therefore, we extract both objects and object parts from our hierarchical clustering and name our clustering algorithm Part2Object. Expanding on this, we enhance the 3D instance segmentation framework, Mask3D [42], to support hierarchical 3D part and instance segmentation. Our improved model, named Hi-Mask3D, takes the object and object parts from Part2Object as pseudo-labels for learning 3D instance segmentation by incorporating explicit interactions between object parts and objects. + +To evaluate the effectiveness of our Part2Object clustering and Hi-Mask3D model, we conduct experiments on the challenging and cluttered indoor environments [4,12,39] in three settings: 1) direct evaluation on unsupervised instance segmentation, 2) application to data-efficient fine-tuning, and 3) cross-dataset generalization, all showing significant performance improvement. In summary, our contributions are multi-fold: + +- We propose two key insights for unsupervised 3D instance segmentation: 1) Employing hierarchical clustering enables the gathering of objects at varying levels of clustering granularity. 2) Leveraging 3D objectness priors from temporally consecutive 2D frames as guidance, while harnessing 3D geometric priors to clustering on the point cloud for precise instance segmentation. + +- Based on our insights, we propose an innovative hierarchical clustering approach, Part2Object. It progressively groups points into object parts and objects while extracting and leveraging 3D objectness priors to guide the clustering process. Our Part2Object significantly outperforms the state-of-the-art training-free unsupervised 3D instance segmentation methods by $16.8\%$ mAP@50 on the ScanNet dataset. + +- We propose Hi-Mask3D, an extension of 3D instance segmentation to support hierarchical unsupervised 3D part and instance segmentation. Experiments demonstrate that Hi-Mask3D consistently and significantly outperforms state-of-the-art models in all the settings. + +# 2 Related Work + +Unsupervised 3D instance segmentation. 3D instance segmentation is an essential task in 3D scene understanding, which aims to locate and recognize different objects in a 3D point cloud. However, the cost and labor intensity of 3D scene-level instance annotation underscore the significance of exploring challenging unsupervised instance segmentation methods. Early works utilize raw geometric information, including coordinates, colors, and normal vectors, to perform traditional clustering [15,16,32] for segmentation. Recently, inspired by the new paradigm in 2D unsupervised instance segmentation, some works [41,66] introduce pseudo labels and self-training to unsupervised 3D instance segmentation. Specifically, these methods use 2D or 3D self-supervised models to extract + +point cloud features, applying graph-cut algorithms to generate pseudo-labels for training and prediction. As illustrated earlier, the clustering method encounters a trade-off between under-segmentation and over-segmentation. Relying on a single graph-cut algorithm, as recent works do, falls short of achieving consistent granularity in complex indoor environments. + +Transfer 2D foundation models into 3D. The rapid advancements in 2D vision have given rise to powerful foundation models [2,6,13,18,26,28,43-45,53], proving beneficial across a variety of visual tasks. In contrast to the 2D domain, 3D vision encounters challenges with data scarcity and training limitations, hindering the development of foundation models. Consequently, several works in the 3D turn to leveraging 2D foundation models to address 3D problems. In 3D instance segmentation, some works [7,64] utilize 2D vision foundation models to generate pseudo-labels for 2D images, aiding in the training and prediction of 3D models through 2D to 3D projection algorithms. Other works [31,36] employ features rather than outputs obtained from 2D foundation models, enhancing the feature extraction capabilities of 3D models through pixel-point alignment. However, relying solely on 2D features or 2D pseudo-labels poses challenges in addressing cluttered indoor scenes with stacked objects. + +Supervised point cloud segmentation. Supervised point cloud segmentation can be divided into instance-level and part-level segmentation. The former has been significant breakthroughs in recent years [8, 19, 20, 22, 25, 27, 37, 38, 42, 50, 55, 56, 58, 62, 63]. However, these approaches demand a substantial amount of annotated data and focus on learning object-level attributes, neglecting the understanding of parts and their contributions to the composition of objects. In the realm of 3D part segmentation, various setups, including supervised [30,37,65], weakly supervised [9,10,57] instance segmentation and semantic segmentation, as well as open-world semantic segmentation [24, 36], have been extensively explored. While these approaches have demonstrated promising result, their primary emphasis lies in determining how to divide an object into parts. To integrate instance-level and part-level perspectives, we perform hierarchical clustering with the prior knowledge of how parts compose objects, unifying instance-level object identification and part-level constituent tracing. + +# 3 Methodology + +Problem definition. Unsupervised 3D instance segmentation [41, 67] is the task of generating class-agnostic masks for all foreground objects in a 3D point cloud scene solely based on raw RGB-D data, without relying on any object detection or instance segmentation annotations. The raw RGB-D data [12, 39] consists of the point cloud $P \in \mathbb{R}^{N \times 3}$ and corresponding 2D RGB frames $\{I_i\}_{i=1}^M$ , where each image frame $I_m \in \mathbb{R}^{H \times W \times 3}$ . With raw RGB-D data, 3D instance segmentation task aims to output all object masks $P_{\mathrm{object}}$ in the scene. + +Method overview. As shown in Fig 2, our unsupervised framework has two major components: 1) the hierarchical clustering algorithm (Sec 3.1 and Sec 3.2), Part2Object, which progressively group points into larger clusters. Duringclus + +![](images/2c25a136c403368b85f58a5414f7feb2309379dbd617bc02a166b98383d86538.jpg) +Fig. 2: Overview of our Part2Object hierarchical clustering and Hi-Mask3D instance segmentation framework. Part2Object extracts 3D objectness priors from consecutive 2D RGB frames and uses them to guide hierarchical clustering from points to object parts and objects. Hi-Mask3D utilizes objects and parts identified by Part2Object as pseudo-labels, learning for improving instance segmentation through the utilization of object parts. + +tering, aided by objectness priors from 2D frames, we can effectively obtain both object-level and part-level segmentation masks, which serve as pseudo-labels for learning unsupervised instance segmentation. 2) The end-to-end 3D instance segmentation model Hi-Mask3D (Sec 3.3) extends the original Mask3D model [42] by explicitly predicting segmentation results of object parts and leveraging them to aid instance segmentation. By self-training on the pseudo-labels extracted from the first-stage Part2Object, Hi-Mask3D can effectively predict 3D instance segments. + +# 3.1 Hierarchical Clustering on 3D Point Cloud + +In this section, we introduce the hierarchical clustering process of our Part2Object which starts by grouping initial points into initial clusters and then progressively merging them into larger clusters. We will sequentially cover: 1) the feature representation of points, 2) the initialization of clusters from points, and 3) the hierarchical clustering process for merging clusters. Point Feature Representation. Given a 3D surface point cloud $P$ , we denote its norm and color as $P^{\mathrm{Norm}}$ and $P^{\mathrm{RGB}}$ , respectively. The 3D feature $f_{i}^{3\mathrm{D}}$ of point $p_i \in P$ comprises $p_i$ , $p_i^{\mathrm{Norm}}$ , and $p_i^{\mathrm{RGB}}$ , representing the location, color, and geometry information, respectively. Additionally, we define point cloud 2D semantics feature $f_{i}$ by projecting 2D features [6] of RGB frames into 3D space using the corresponding camera pose. Here, 2D features are extracted using the 2D self-supervised + +method DINO [6], without rely on any 2D instance annotations. Cluster Initialization and Feature Representation. Given the vast number of points in the point cloud data [1], we initially apply the VCCS algorithm [34] to generate super-points on the point cloud, forming the first-layer clusters $\{c_i^0\}_{i = 1}^{N_0}$ as follows: + +$$ +\{c _ {i} ^ {0} \} _ {i = 1} ^ {N _ {0}} = \mathrm {V C C S} (P, P ^ {\mathrm {N o r m}}, P ^ {\mathrm {R G B}}), \tag {1} +$$ + +where VCCS algorithm takes points' coordinates $P$ , norms $P^{\mathrm{Norm}}$ , and colors $P^{\mathrm{RGB}}$ as input and group original $N$ points into $N_0$ super-points as $\{c_i^0\}_{i=1}^{N_0}$ . We consider each super-point $c_i^0$ as a cluster, consisting of multiple points that are closely positioned and share similar color and norm characteristics. Then, we define the feature representation for the cluster based on the features of the points within the cluster. Instead of directly averaging point features within the cluster, we compute a weight for each point and perform a weighted sum of features. These weights are determined by the similarity between each point's feature and the average feature of points within the cluster. This approach helps mitigate the influence of noisy points in the cluster, leading to more robust representations. Specifically, for the cluster $c_i^0$ , its cluster feature $f_i^0$ is computed as follows, + +$$ +\boldsymbol {f} _ {i} ^ {0} = \sum_ {p _ {j} \in c _ {i} ^ {0}} \frac {\operatorname {s i m} \left(\boldsymbol {f} _ {j} , \bar {\boldsymbol {f}} _ {i} ^ {0}\right)}{\sum p _ {j} \in c _ {j} ^ {0} \operatorname {s i m} \left(\boldsymbol {f} _ {j} , \bar {\boldsymbol {f}} _ {i} ^ {0}\right)} \boldsymbol {f} _ {j}, \text {w h e r e} \bar {\boldsymbol {f}} _ {i} ^ {0} = \sum_ {p _ {j} \in c _ {i} ^ {0}} \frac {1}{\left| c _ {i} ^ {0} \right|} \boldsymbol {f} _ {j}. \tag {2} +$$ + +The $p_j \in c_i^0$ denotes each point $p_j$ within the cluster $c_i^0$ , and $\pmb{f}_j$ is the feature of point $p_j$ . The $\bar{\pmb{f}}_i^0$ is the average feature of points within $c_i^0$ , and $\sin(\cdot, \cdot)$ denotes the computation of cosine similarity. As depicted in the example in Fig 3(a), our cluster feature $\pmb{f}_i^0$ demonstrates greater robustness compared to $\bar{\pmb{f}}_i^0$ . For simplicity, we abbreviate the operations in Equ 2 as function $\mathsf{FU}(\cdot)$ , which computes the cluster feature for the inputted cluster based on point features, e.g., $\pmb{f}_i^0 = \mathsf{FU}(c_i^0)$ . + +Hierarchical Clustering with One Stop Criteria. Next, we group and merge the first-layer clusters $\{c_i^0\}_{i = 1}^{N_0}$ to the next-layer clusters $\{c_i^1\}_{i = 1}^{N_1}$ based on their features $\{\pmb {f}_i^0\}_{i = 0}^{N_0}$ and spatial coordinates, progressively and iteratively forming higher-hierarchical clusters $\{c_{ij}^{t}\}_{i = 1}^{N_t}$ , where $t$ represents the $t$ -th layer in hierarchical clustering. In each single clustering layer, our clustering principle is that two clusters can only be merged when they are semantically similar and spatially adjacent: 1) the feature similarity between them ranks among the top $K$ similarities between any pair of clusters. 2) The closest points between two clusters are adjacent. Specifically, when considering two clusters in the $t$ -th layer, $c_{i}^{t}$ and $c_{j}^{t}$ , along with their features $\pmb {f}_i^t$ and $\pmb {f}_j^t$ , the merging process to form the next-layer cluster $c_k^{t + 1}$ and feature $\pmb {f}_k^{t + 1}$ is as follows: + +$$ +\begin{array}{l} c _ {k} ^ {t + 1} = c _ {i} ^ {t} \cup c _ {j} ^ {t} \quad \mathrm {i f} \operatorname {r a n k} (\operatorname {s i m} (\pmb {f} _ {i} ^ {t}, \pmb {f} _ {j} ^ {t})) \leq K \mathrm {a n d} \operatorname {d i s t} (c _ {i} ^ {t}, c _ {j} ^ {t}) \leq T, \\ \boldsymbol {f} _ {k} ^ {t + 1} = \mathrm {F U} (c _ {i} ^ {t} \cup c _ {j} ^ {t}), \\ \end{array} +$$ + +where $\mathrm{rank}(\cdot)$ denotes the ranking of feature similarity among pairwise cluster similarities in this clustering layer, where a higher rank signifies a higher similarity, and $\mathrm{dist}(\cdot,\cdot)$ computes the Euclidean distance between closest points of the two clusters. The $K$ and $T$ represent the threshold values for ranking and distance, respectively. However, we observe that solely using the merging metric in Equ 3 can result in the incorrect merging of parts from different objects in the early stages of clustering (i.e., shallow layers of clustering). This occurs because clusters at the shallow layers are typically scattered fragments with highly local features, which hinder them from perceiving the entire object. To track this issue, we first extract 3D objectness priors, represented by a set of 3D bounding boxes of potential objects $B^{3\mathrm{D}}$ , detailed in Sec 3.2. Then, we design a stopping criteria based on the 3D objectness priors to prevent clusters belonging to different objects from being merged. Criteria: Stop merging clusters belonging to different objects! Specifically, for any pair of clusters $c_{i}^{t}$ and $c_{j}^{t}$ that meet the merging metric, we extra assess their spatial relationship relative to each 3D object $b_{k}^{3\mathrm{D}} \in B^{3\mathrm{D}}$ . If clusters $c_{i}^{t}$ and $c_{j}^{t}$ are respectively inside and outside the object $b_{k}^{3\mathrm{D}}$ , then reject their merging because they do not belong to the same object. + +Collect Objects and Parts from Clustering. By employing hierarchical clustering with the stopping criteria, we identify clusters that stop merging with others as 3D objects, denoted as $\hat{P}_{\mathrm{object}}$ . Additionally, in hierarchical clustering, objects with complex geometric are typically merged from scattered fragments into meaningful parts, and then into complete objects. As shown in Fig 1, scattered "toilet" fragments are combined into object parts such as the toilet back and seat (see layer two), which are then formed into the object "toilet" (layer three). Therefore, we can trace back from objects $\hat{P}_{\mathrm{object}}$ to identify the clusters from the previous level that form them, considering these clusters as potential object parts, denoted as $\hat{P}_{\mathrm{part}}$ . Both objects $\hat{P}_{\mathrm{object}}$ and object parts $\hat{P}_{\mathrm{part}}$ identified in hierarchical clustering serve as pseudo-labels for training our instance segmentation model, Hi-Mask3D (Sec 3.3). + +# 3.2 3D Objectness Priors From 2D Frames + +In this section, we introduce our grouping-first-then-projection pipeline to extract 3D objectness priors $B^{3\mathrm{D}}$ from 2D RGB frames [12], where $B^{3\mathrm{D}}$ is the set of 3D bounding boxes for potential objects in the scene. + +Grouping-first-then-projection Pipeline. To extract 3D objectness priors, a simple strategy involves projection-first-then-grouping: projecting 2D unsupervised instance segmentation masks [46, 60] from 2D frames onto the 3D point cloud to acquire 3D masks, then grouping and merging these 3D masks based on their spatial relationships to generate 3D objects, as shown in Fig 3(c). However, due to variations in object sizes and corresponding camera poses of 2D frames across different scenes, it is nearly impossible to establish a unified grouping criterion for different objects. If the grouping criterion is too strict, allowing only masks with high overlap to be grouped, masks from multiple parts of a large object cannot be merged (see Fig 3c-"strict"). Conversely, if the criterion + +![](images/2cfae4e55033010216d29745460a386c2922614fd575d5d2edc95b255a06f0d8.jpg) +Fig. 3: Robustness of our cluster features and 3D objectness priors. (a) Visualization of the first 3 PCA components and our computed weights of points in Equ 2. (b) Visualization of 2D DINO features' PCA components, 3D points' normal vectors, and our 3D object priors. The comparison between (c) the projection-first-then-grouping pipeline and (d) our grouping-first-then-projection pipeline. + +is relaxed, masks that should belong to multiple objects may be merged into a single object, e.g., the "chair" and "sofa" in Fig 3c-"relaxed". We attribute this limitation to the inability to rely solely on 3D spatial relationships and 3D point features (see Fig 3b) to determine if multiple 3D masks belong to the same object. To tackle this issue, we propose a grouping-first-then-projection pipeline: 1) it co-segments objects across multiple 2D frames by exploiting the temporal consistency of the frames, naturally identifying and grouping multiple 2D masks for each single object. 2) It projects the 2D masks corresponding to each object onto the 3D point cloud and then directly merges them. + +Object Co-segmentation from Consecutive 2D RGB Frames. Given that 2D RGB frames are captured consecutively, we leverage their temporal consistency to co-segment multi-frame 2D masks corresponding to the same objects. First, for each RGB frame $I_{m}$ , we apply the 2D unsupervised instance segmentation method MaskCut [60] to obtain the 2D masks of objects in it, denoted as $O_{m}$ . As MaskCut uses the image encoder of 2D self-supervised DINO [6] to extract image feature map, we employ the same encoder to extract features for objects $O_{m}$ by masked average pooling each object mask $o_{m}^{i} \in O_{m}$ over the frame's feature map. Next, for any adjacent pair of frames $I_{m}$ and $I_{m+1}$ , we compute the pairwise similarity between their object masks $O_{m}$ and $O_{m+1}$ based on mask features. Then, for each object mask $o_{m}^{i} \in O_{m}$ in frame $I_{m}$ , we find the object mask $o_{m+1}^{j^{*}}$ in frame $I_{m+1}$ with the highest similarity as its corresponding candidate mask as follows, + +$$ +\begin{array}{l} o _ {m + 1} ^ {j ^ {*}} = \operatorname * {a r g m a x} _ {i} \operatorname * {s i m} (o _ {m + 1} ^ {j}, o _ {m} ^ {i}), \\ o _ {m + 1} ^ {\jmath} \in O _ {m + 1} \tag {4} \\ g _ {m \rightarrow m + 1} ^ {i} = o _ {m + 1} ^ {j ^ {*}} \mathrm {i f} \mathrm {s i m} (o _ {m + 1} ^ {j ^ {*}}, o _ {m} ^ {i}) > \tau \mathrm {\bf e l s e N u l l}, \\ \end{array} +$$ + +where $\mathrm{sim}(o_{m + 1}^j,o_m^i)$ is to compute the cosine similarity between mask features of object masks $o_{m + 1}^{j}$ and $o_{m}^{i}$ . The $o_{m + 1}^{j^{*}}$ in frame $I_{m + 1}$ is the candidate object mask of the object mask $o_{m}^{i}$ in frame $I_{m}$ . Notably, only when the similarity between the object mask $o_{m}^{i}$ and its candidate object mask $o_{m + 1}^{j^{*}}$ exceeds the threshold $\pi$ , they are considered to belong to the same object. After determining the same objects between any adjacent frames, $g_{m\rightarrow m + 1}^{i}$ , we propagate this "sameness" to all adjacent frames to identify the same objects across all frames. Consequently, we can gather objects and identify all 2D masks in different frames of each object. + +3D Objectness Priors. Given that we have grouped 2D masks into different objects, for each object, we can directly project its 2D masks onto the 3D point cloud and seamlessly merge their 3D masks into a unified whole [64], forming the object's 3D mask. Furthermore, instead of directly utilizing the 3D masks of objects, we utilize the 3D bounding boxes $B^{3\mathrm{D}}$ of their masks to guide hierarchical clustering. This decision is based on treating the collected 3D objects from 2D frames as objectness priors and leaving precise segmentation to the clustering process itself. + +# 3.3 Hi-Mask3D and Learning Objective + +In this section, we introduce our 3D instance segmentation framework, HiMask3D. It extends and adapts the 3D instance segmentation model, Mask3D [42], to predict and utilize object parts to improve 3D instance segmentation. + +Introduction to Mask3D. The 3D instance segmentation model Mask3D [42] mainly consists of three components: 1) a feature backbone, a sparse convolutional U-net [11] that takes a colored point cloud as input and yields multi-resolution feature matrices $\{\pmb{F}_r\}$ , where each $\pmb{F}_r \in \mathbb{R}^{M_r \times C}$ of size $M_r$ . 2) A mask module predicts instance masks $B$ corresponding to instance queries $Q \in \mathbb{R}^{K \times C}$ by computing the dot product between queries $Q$ and scene features $\{\pmb{F}_r\}$ . 3) A query refinement module, a transformer decoder [54] that cross-attends $K$ instance queries $Q$ with multi-resolution feature matrices $\{\pmb{F}_r\}$ . In the $r$ -th decoder layer of query refinement module, the cross-attention between instance queries $Q_r$ and feature matrix $\pmb{F}_r$ is formulated as follows, + +$$ +Q _ {r} = \operatorname {s o f t m a x} \left(Q _ {r - 1} \boldsymbol {F} _ {r} ^ {T} / \sqrt {C}\right) \boldsymbol {F} _ {r}. \tag {5} +$$ + +Note that for simplicity of demonstration, we omit the QKV projection [54] and intermediate instance mask $B$ [42] in the cross-attention formulation. + +Hi-Mask3D with Hierarchical-Aware Detector. Our Hi-Mask3D modifies the query refinement module of Mask3D to adapt for both object and object part segmentation. Specifically, we start with two sets of queries $Q^o$ and $Q^p$ representing object queries and object part queries, respectively. Next, in each decoder layer, we no longer directly perform cross-attention between object queries and feature matrix to update object queries. Instead, we perform hierarchical attention: 1) conducting cross-attention between object part queries and feature matrix, and 2) updating object queries using part queries. This enables explicit + +interaction between object part queries and object queries, which helps to leverage the information from object parts to improve object segmentation. In the $r$ -th decoder layer, the hierarchical attention is formulated as follows: + +$$ +Q _ {r} ^ {p} = \operatorname {s o f t m a x} \left(Q _ {r - 1} ^ {p} \boldsymbol {F} _ {r} ^ {T} / \sqrt {C}\right) \boldsymbol {F} _ {r}, \tag {6} +$$ + +$$ +Q _ {r} ^ {o} = \operatorname {s o f t m a x} \left(Q _ {r - 1} ^ {o} \left(Q _ {r} ^ {p}\right) ^ {T} / \sqrt {C}\right) Q _ {r} ^ {p}. +$$ + +The last-layer part queries and object queries undergo the mask module to obtain object part and part segmentation results, $P_{\mathrm{part}}$ and $P_{\mathrm{object}}$ , respectively. + +Training Objective. Following Mask3D [42], we adopt bipartite graph algorithm [5, 48] to match prediction $P_{\mathrm{object}}$ and $P_{\mathrm{part}}$ with pseudo-labels $\hat{P}_{\mathrm{object}}$ and $\hat{P}_{\mathrm{part}}$ , respectively. To supervise segmentation masks, we combine DICE [14] loss and Binary Cross Entropy (BCE) loss [29]. Similar to [41], we also employ iterative rounds of self-training to enhance the segmentation capabilities of HiMask3D. During self-training, predictions from the previous round are utilized as pseudo-labels for the subsequent round. + +# 4 Experiment + +Datasets and Implementations. We evaluate our method on four public indoor point cloud datasets, including ScanNet [12], Scannet200 [40], S3DIS [4] and Replica [49]. Following Mask3D [42], we employ a Sparse CNN, Res16UNet34C [11], implemented in MinkowskiEngine [11], as the point cloud encoder. Our HiMask3D is trained for 600 epochs with a learning rate 1e-4 and a batch size 4. Experiments are conducted using PyTorch [35] and executed on NVIDIA Tesla A40 GPUs. We evaluate the performance of class-agnostic instance segmentation using standard average precision scores. The scores consist of mAP@25, mAP@50, and mAP, representing performance metrics at IoU thresholds of 0.25, 0.5, and mean average precision, respectively. Hyper-parameter $T = 0.05$ in Equ 3 and $\tau = 0.3$ in Equ 4. + +# 4.1 Comparison with State-of-the-Art Unsupervised Methods + +Table 1 shows the results on ScanNet dataset [12] under two different settings: Traning-free Setting. In the training-free setting, all methods rely solely on clustering [6] or graph-cut [46, 60] algorithms for prediction, as no annotations are available for use. Table 1a compares our method, Part2Object, with the previous state-of-the-art method, Unscene3D [41], which also leverages DINO features but applies graph-cut algorithms on 3D point clouds. Compared to Unscene3D, our approach demonstrates significant enhancements across all metrics, with increases of $35.2\%$ , $16.8\%$ , and $6.7\%$ in mAP@25, mAP@50, and mAP, respectively. These improvements are attributed to our utilization of 3D objectness priors, which guide the clustering process. + +Data-efficient Fine-tuning Setting. In the data-efficient fine-tuning setting, all methods initially undergo either pretraining or self-training. Subsequently, + +Table 1: Comparison of instance segmentation results on ScanNet val. +(a) Training-free Setting. All methods do not require gradient backpropagation to obtain prediction results. +(b) Data-efficient Setting with $\mathbf{x}\%$ available data annotations. A $0\%$ value indicates that the method can provide pseudo-labels, and training is conducted on them. + +
w/o training MethodsScanNet Val [12]w/ training MethodsmAP@50 ScanNet Val [12]
mAP@25mAP@50mAP0%1%5%10%20%
HDBSCAN [32]32.15.51.6Scratch [42]/14.133.339.243.4
Nunes [33]30.57.32.3CSC [21]/22.139.943.848.9
Felzenswalb [16]38.912.75.0HDBSCAN [32]10.515.136.340.042.7
CutLER [60]7.00.20.3Felzenswalb [16]15.225.337.245.750.0
Unscene3D [41]19.910.05.9Unscene3D [41]23.228.446.855.760.7
Ours55.126.812.6Ours32.644.164.268.072.1
+ +they fine-tune on downstream data with varying percentages (0% represents pseudo-labels solely). Under this setting, the extraction of pseudo-labels is the primary requirement, followed by the necessity for models to exhibit strong learning capabilities. As demonstrated in Table 1b, Hi-Mask3D exhibits remarkable performance even with limited data, showcasing the robust 3D representation capabilities learned on the pseudo-labels from Part2object. Specifically, our method surpasses the previous SOTA by $15.7\%$ and $11.4\%$ when considering the utilization of only $1\%$ and $20\%$ of the available data, respectively. + +# 4.2 Comparison on Cross-dataset Generalization + +Table 2 and Table 3 show our comparisons with fully-supervised Mask3D [42] on cross-dataset generalization, including both zero-shot and dataset-efficient settings. In both settings, our Hi-Mask3D is trained on pseudo-labels from our unsupervised Part2Object on the ScanNet dataset [12], while Mask3D is trained fully supervised on ScanNet. + +Cross-dataset Zero-shot Generalization Setting. In Table 2, we compare unsupervised Hi-Mask3D with fully-supervised Mask3D on three downstream datasets, including ScanNet200 [39], S3DIS [4] and Replica [49]. Compared with the fully supervised Mask3D trained on ScanNet, our Hi-Mask3D trained on pseudo-labels from Part2Object achieves consistent performance improvement without using any manually annotated data. Hi-Mask3D surpasses fully-supervised Mask3D by $10.7\%$ , $4.8\%$ , and $6.4\%$ in terms of mAP@50 on ScanNet200, S3DIS, and Replica datasets, respectively. Thanks to the learning from pseudo-labels in Part2Object, Hi-Mask3D has acquired more generalizable representations of 3D objects. Unlike fully-supervised methods, which are constrained to annotated classes in the training dataset, Hi-Mask3D demonstrates robust performance in cross-dataset generalization. + +Cross-dataset Data-efficient Generalization Setting. In Table 3, we conduct data-efficient training on downstream datasets, such as ScanNet200 [39] and S3DIS [4], after pre-training our Hi-Mask3D on ScanNet's pseudo-labels ex + +Table 2: Comparison on cross-dataset zero-shot generalization setting. We compare the zero-shot generalization ability of unsupervised Hi-Mask3D with fully-supervised Mask3D on three downstream datasets: ScanNet200, S3DIS and Replica. + +
Zero-shot MethodsScanNet200 Val [39]S3DIS 6-fold [4]Replica [49]
m@25m@50mAPm@25m@50mAPm@25m@50mAP
Mask3D [42]30.824.215.417.611.77.720.815.69.7
Ours63.234.916.324.516.58.536.522.011.2
+ +Table 3: Comparison on cross-dataset data-efficient generalization setting. We utilize pseudo-labels from Part2Object to pre-train our Hi-Mask3D and conduct data-efficient finetuning on downstream datasets, comparing them with Mask3D. + +
MethodDatasetmAP@25mAP@50
1%5%10%20%1%5%10%20%
ScratchScanNet200 [39]2.122.530.640.21.38.218.422.9
OursScanNet200 [39]59.470.472.377.835.452.856.262.8
ScratchS3DIS [4]1.710.120.845.41.02.99.125.9
OursS3DIS [4]49.155.965.170.225.735.046.449.4
+ +tracted from Part2Object. For ScanNet200 and S3DIS, Hi-Mask3D outperforms Mask3D $47.4\%$ and $24.7\%$ mAP@50 with $20\%$ data, respectively. When limited data is available, the improvement brought by pretraining on pseudo-labels from Part2Object is even greater, with the enhancement increasing from $37.6\%$ to $57.3\%$ as the available data decreases from $20\%$ to $1\%$ . + +# 4.3 Ablation Study + +To evaluate the effectiveness of our hierarchical clustering Part2Object and 3D instance segmentation architecture Hi-Mask3D, as well as the effect of hyperparameters, we conduct comprehensive and detailed ablation experiments, as outlined in Table 4a and Table 4b. + +Ablation on Part2Object Clustering. Table 4a illustrates the variants of Part2Object clustering, starting with a baseline utilizing the single-layer clustering algorithm. The performance of this baseline yields only $6.1\%$ in mAP and $13.8\%$ in mAP@50. This limitation arises due to the single-layer clustering being either too loose or too tight, leading to over-clustered or under-clustered objects. Subsequently, we remove the guidance of objectness priors from our proposed clustering algorithm, denoted as "w/o OG". Without object priors, while objects may appear at every layer, it becomes challenging to determine which layer they belong to. Furthermore, we replace our cluster feature computation function, "FU(\cdot)", as simple averaged over features of points within the cluster, "w/o FU". The $2.0\%$ decline in mAP highlights the importance of removing noise features for more robust cluster feature representations. + +Table 4: Ablation study. We conduct ablation studies on ScanNet val. We validate the effectiveness of our Part2Object clustering, Hi-Mask3D, and self-training procedure. + +(a) Ablation study on Part2Object clustering and Hi-Mask3D architecture. OG, FU denotes objectness guidance and cluster feature function, respectively. + +
Ablation on ClusteringScanNet Val [12]
m@25m@50mAP
Baseline37.813.86.1
w/o OG43.220.710.4
w/o FU44.221.510.6
Ours55.126.812.6
Architecture
Mask3D59.031.014.2
Hi-Mask3D64.936.016.9
+ +(b) Ablation study on Self-training and Hyper-parameter. + +
Ablation on Self-TrainingScanNet Val [12]
m@25m@50mAP
Pseudo labels55.126.812.6
Round 160.732.615.3
Round 264.335.216.5
Round 364.936.016.9
Hyper-parameter
K = 0.455.722.610.3
K = 0.557.025.312.0
K = 0.655.126.812.6
K = 0.744.521.110.6
+ +Ablation on Hi-Mask3D. We compare our Hi-Mask3D with Mask3D using the same pseudo-labels extracted from Part2Object. Hi-Mask3D improves Mask3D by $2.7\%$ and $5.0\%$ in terms of mAP and mAP@50, respectively. This demonstrates that additional part information aids in object understanding. Furthermore, through reporting the performance of self-training, we observe that after several rounds, the mAP increases from $15.3\%$ to $16.9\%$ . + +Ablation on Hyper-parameter. We conduct ablation on the import hyperparameter $K$ in Equ 3. Since the number of clusters at each layer varies, we use percentages (K=0.6 indicates the top 60%) to measure how many of the top clusters from each layer can be aggregated. Table 4b shows the performance of Part2Object at different $K$ . We use grey to indicate the default setting. + +# 5 Conclusion and Limitations + +We introduce Part2Object, an efficient hierarchical clustering algorithm, that progressively groups point clouds into object parts and objects, while leveraging 3D object-ness prior to precisely target objects. Furthermore, we propose the Hierarchical-Aware Mask3D, facilitating self-training with pseudo-object and part labels from Part2Object. Experimental results demonstrate that we consistently surpass all existing methods in both unsupervised settings and data-efficient settings. Ethics Statement: Given that our 2D knowledge is derived from the self-supervised models DINO, we acknowledge that biases and controversies inherent in the training data for these models may be introduced into our model. Acknowledgment: This work was supported by the National Natural Science Foundation of China (No.62206174) and MoE Key Laboratory of Intelligent Perception and Human-Machine Collaboration (ShanghaiTech University). + +# References + +1. Adams, R., Bischof, L.: Seeded region growing. IEEE Transactions on pattern analysis and machine intelligence 16(6), 641-647 (1994) +2. Amir, S., Gandelsman, Y., Bagon, S., Dekel, T.: Deep vit features as dense visual descriptors. arXiv preprint arXiv:2112.05814 (2021) +3. An, D., Wang, H., Wang, W., Wang, Z., Huang, Y., He, K., Wang, L.: Etpnav: Evolving topological planning for vision-language navigation in continuous environments. arXiv preprint arXiv:2304.03047 (2023) +4. Armeni, I., Sener, O., Zamir, A.R., Jiang, H., Brilakis, I., Fischer, M., Savarese, S.: 3d semantic parsing of large-scale indoor spaces. In: Proceedings of the IEEE conference on computer vision and pattern recognition. pp. 1534-1543 (2016) +5. Carion, N., Massa, F., Synnaeve, G., Usunier, N., Kirillov, A., Zagoruyko, S.: End-to-end object detection with transformers. In: European conference on computer vision. pp. 213-229. Springer (2020) +6. Caron, M., Touvron, H., Misra, I., Jégou, H., Mairal, J., Bojanowski, P., Joulin, A.: Emerging properties in self-supervised vision transformers. In: Proceedings of the International Conference on Computer Vision (ICCV) (2021) +7. Chen, R., Liu, Y., Kong, L., Chen, N., Xinge, Z., Ma, Y., Liu, T., Wang, W.: Towards label-free scene understanding by vision foundation models. In: Thirty-seventh Conference on Neural Information Processing Systems (2023) +8. Chen, S., Fang, J., Zhang, Q., Liu, W., Wang, X.: Hierarchical aggregation for 3d instance segmentation. In: Proceedings of the IEEE/CVF International Conference on Computer Vision. pp. 15467-15476 (2021) +9. Chen, Z., Yin, K., Fisher, M., Chaudhuri, S., Zhang, H.: Bae-net: Branched autoencoder for shape co-segmentation. In: Proceedings of the IEEE/CVF International Conference on Computer Vision. pp. 8490-8499 (2019) +0. Chibane, J., Engelmann, F., Anh Tran, T., Pons-Moll, G.: Box2mask: Weakly supervised 3d semantic instance segmentation using bounding boxes. In: European Conference on Computer Vision. pp. 681-699. Springer (2022) +1. Choy, C., Gwak, J., Savarese, S.: 4d spatio-temporal convnets: Minkowski convolutional neural networks. In: Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition. pp. 3075-3084 (2019) +2. Dai, A., Chang, A.X., Savva, M., Halber, M., Funkhouser, T., Nießner, M.: Scannet: Richly-annotated 3d reconstructions of indoor scenes. In: Proc. Computer Vision and Pattern Recognition (CVPR), IEEE (2017) +3. Dai, Q., Yang, S.: Curriculum point prompting for weakly-supervised referring segmentation (2024) +4. Deng, R., Shen, C., Liu, S., Wang, H., Liu, X.: Learning to predict crisp boundaries. In: Proceedings of the European conference on computer vision (ECCV). pp. 562-578 (2018) +5. Ester, M., Kriegel, H.P., Sander, J., Xu, X.: A density-based algorithm for discovering clusters in large spatial databases with noise. In: Proceedings of the Second International Conference on Knowledge Discovery and Data Mining. p. 226-231. KDD'96, AAAI Press (1996) +6. Felzenszwalb, P.F., Huttenlocher, D.P.: Efficient graph-based image segmentation. International journal of computer vision 59, 167-181 (2004) +7. Geng, H., Xu, H., Zhao, C., Xu, C., Yi, L., Huang, S., Wang, H.: Gapartnet: Cross-category domain-generalizable object perception and manipulation via generalizable and actionable parts. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 7081-7091 (2023) + +18. Ghiasi, G., Gu, X., Cui, Y., Lin, T.Y.: Open-vocabulary image segmentation. arXiv preprint arXiv:2112.12143 (2021) +19. Han, L., Zheng, T., Xu, L., Fang, L.: Occuseg: Occupancy-aware 3d instance segmentation. In: 2020 IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR). pp. 2937-2946 (2020). https://doi.org/10.1109/CVPR42600.2020.00301 +20. Hou, J., Dai, A., Nießner, M.: 3d-sis: 3d semantic instance segmentation of rgb-d scans. In: Proceedings of the IEEE/CVF conference on computer vision and pattern recognition. pp. 4421-4430 (2019) +21. Hou, J., Graham, B., Nießner, M., Xie, S.: Exploring data-efficient 3d scene understanding with contrastive scene contexts. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 15587-15597 (2021) +22. Hu, Q., Yang, B., Xie, L., Rosa, S., Guo, Y., Wang, Z., Trigoni, N., Markham, A.: Randla-net: Efficient semantic segmentation of large-scale point clouds. In: Proceedings of the IEEE/CVF conference on computer vision and pattern recognition. pp. 11108-11117 (2020) +23. Huang, S., Wang, Z., Li, P., Jia, B., Liu, T., Zhu, Y., Liang, W., Zhu, S.C.: Diffusion-based generation, optimization, and planning in 3d scenes. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 16750-16761 (2023) +24. Huang, Z., Wu, X., Chen, X., Zhao, H., Zhu, L., Lasenby, J.: Openins3d: Snap and lookup for 3d open-vocabulary instance segmentation. arXiv preprint arXiv:2309.00616 (2023) +25. Hui, L., Tang, L., Shen, Y., Xie, J., Yang, J.: Learning superpoint graph cut for 3d instance segmentation. In: NeurIPS (2022) +26. Kirillov, A., Mintun, E., Ravi, N., Mao, H., Rolland, C., Gustafson, L., Xiao, T., Whitehead, S., Berg, A.C., Lo, W.Y., et al.: Segment anything. arXiv preprint arXiv:2304.02643 (2023) +27. Kolodiazhnyi, M., Rukhovich, D., Vorontsova, A., Konushin, A.: Top-down beats bottom-up in 3d instance segmentation (2023). https://doi.org/10.48550/ ARXIV.2302.02871, https://arxiv.org/abs/2302.02871 +28. Li, B., Weinberger, K.Q., Belongie, S., Koltun, V., Ranftl, R.: Language-driven semantic segmentation. In: International Conference on Learning Representations (2022), https://openreview.net/forum?id=RriDjddCLN +29. Lin, T.Y., Goyal, P., Girshick, R., He, K., Dollar, P.: Focal loss for dense object detection. In: Proceedings of the IEEE international conference on computer vision. pp. 2980-2988 (2017) +30. Liu, J., Yu, M., Ni, B., Chen, Y.: Self-prediction for joint instance and semantic segmentation of point clouds. In: Computer Vision-ECCV 2020: 16th European Conference, Glasgow, UK, August 23-28, 2020, Proceedings, Part XXII 16. pp. 187-204. Springer (2020) +31. Liu, Y., Kong, L., Cen, J., Chen, R., Zhang, W., Pan, L., Chen, K., Liu, Z.: Segment any point cloud sequences by distilling vision foundation models. arXiv preprint arXiv:2306.09347 (2023) +32. McInnes, L., Healy, J.: Accelerated hierarchical density based clustering. In: 2017 IEEE International Conference on Data Mining Workshops (ICDMW). pp. 33-42. IEEE (2017) +33. Nunes, L., Chen, X., Marcuzzi, R., Osep, A., Leal-Taixe, L., Stachniss, C., Behley, J.: Unsupervised class-agnostic instance segmentation of 3d lidar data for autonomous vehicles. IEEE Robotics and Automation Letters 7(4), 8713-8720 (2022) + +34. Papon, J., Abramov, A., Schoeler, M., Worgotter, F.: Voxel cloud connectivity segmentation-supervoxels for point clouds. In: Proceedings of the IEEE conference on computer vision and pattern recognition. pp. 2027-2034 (2013) +35. Paszke, A., Gross, S., Massa, F., Lerer, A., Bradbury, J., Chanan, G., Killeen, T., Lin, Z., Gimelshein, N., Antiga, L., et al.: Pytorch: An imperative style, high-performance deep learning library. Advances in neural information processing systems 32 (2019) +36. Peng, S., Genova, K., Jiang, C., Tagliasacchi, A., Pollefeys, M., Funkhouser, T., et al.: Openscene: 3d scene understanding with open vocabularies. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 815-824 (2023) +37. Qi, C.R., Yi, L., Su, H., Guibas, L.J.: Pointnet++: Deep hierarchical feature learning on point sets in a metric space. Advances in neural information processing systems 30 (2017) +38. Rethage, D., Wald, J., Sturm, J., Navab, N., Tombari, F.: Fully-convolutional point networks for large-scale point clouds. In: Proceedings of the European Conference on Computer Vision (ECCV). pp. 596-611 (2018) +39. Rozenberszki, D., Litany, O., Dai, A.: Language-grounded indoor 3d semantic segmentation in the wild. In: European Conference on Computer Vision. pp. 125-141. Springer (2022) +40. Rozenberszki, D., Litany, O., Dai, A.: Language-grounded indoor 3d semantic segmentation in the wild. In: Proceedings of the European Conference on Computer Vision (ECCV) (2022) +41. Rozenberszki, D., Litany, O., Dai, A.: Unscene3d: Unsupervised 3d instance segmentation for indoor scenes. arXiv preprint arXiv:2303.14541 (2023) +42. Schult, J., Engelmann, F., Hermans, A., Litany, O., Tang, S., Leibe, B.: Mask3D for 3D Semantic Instance Segmentation. In: International Conference on Robotics and Automation (ICRA) (2023) +43. Shi, C., Yang, S.: Edadet: Open-vocabulary object detection using early dense alignment. In: Proceedings of the IEEE/CVF International Conference on Computer Vision. pp. 15724-15734 (2023) +44. Shi, C., Yang, S.: Logoprompt: Synthetic text images can be good visual prompts for vision-language models. arXiv preprint arXiv:2309.01155 (2023) +45. Shi, C., Yang, S.: The devil is in the object boundary: Towards annotation-free instance segmentation using foundation models. In: The Twelfth International Conference on Learning Representations (2024), https://openreview.net/forum?id=4JbrdrHxYy +46. Shi, J., Malik, J.: Normalized cuts and image segmentation. IEEE Transactions on pattern analysis and machine intelligence 22(8), 888-905 (2000) +47. Song, Z., Yang, B.: Ogc: Unsupervised 3d object segmentation from rigid dynamics of point clouds. Advances in Neural Information Processing Systems 35, 30798-30812 (2022) +48. Stewart, R., Andriluka, M., Ng, A.Y.: End-to-end people detection in crowded scenes. In: Proceedings of the IEEE conference on computer vision and pattern recognition. pp. 2325-2333 (2016) +49. Straub, J., Whelan, T., Ma, L., Chen, Y., Wijmans, E., Green, S., Engel, J.J., Mur-Artal, R., Ren, C., Verma, S., et al.: The replica dataset: A digital replica of indoor spaces. arXiv preprint arXiv:1906.05797 (2019) +50. Sun, J., Qing, C., Tan, J., Xu, X.: Superpoint transformer for 3d scene instance segmentation (2022) + +51. Suo, S., Wong, K., Xu, J., Tu, J., Cui, A., Casas, S., Urtasun, R.: Mixsim: A hierarchical framework for mixed reality traffic simulation. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 9622-9631 (2023) +52. Suomela, L., Kalliola, J., Dag, A., Edelman, H., Kämäräinen, J.K.: Benchmarking visual localization for autonomous navigation. In: Proceedings of the IEEE/CVF Winter Conference on Applications of Computer Vision. pp. 2945-2955 (2023) +53. Tang, J., Zheng, G., Shi, C., Yang, S.: Contrastive grouping with transformer for referring image segmentation. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 23570-23580 (2023) +54. Vaswani, A., Shazeer, N., Parmar, N., Uszkoreit, J., Jones, L., Gomez, A.N., Kaiser, L., Polosukhin, I.: Attention is all you need. Advances in neural information processing systems 30 (2017) +55. Vu, T., Kim, K., Luu, T.M., Nguyen, T., Kim, J., Yoo, C.D.: Softgroup++: Scalable 3d instance segmentation with octree pyramid grouping. arXiv preprint arXiv:2209.08263 (2022) +56. Vu, T., Kim, K., Luu, T.M., Nguyen, X.T., Yoo, C.D.: Softgroup for 3d instance segmentation on 3d point clouds. In: CVPR (2022) +57. Wang, R., Zhang, Y., Mao, J., Zhang, R., Cheng, C.Y., Wu, J.: Ikea-manual: Seeing shape assembly step by step. Advances in Neural Information Processing Systems 35, 28428-28440 (2022) +58. Wang, W., Yu, R., Huang, Q., Neumann, U.: Sgpn: Similarity group proposal network for 3d point cloud instance segmentation. In: Proceedings of the IEEE conference on computer vision and pattern recognition. pp. 2569-2578 (2018) +59. Wang, X., Yu, Z., De Mello, S., Kautz, J., Anandkumar, A., Shen, C., Alvarez, J.M.: FreeSOLO: Learning to segment objects without annotations. arXiv preprint arXiv:2202.12181 (2022) +60. Wang, X., Girdhar, R., Yu, S.X., Misra, I.: Cut and learn for unsupervised object detection and instance segmentation. arXiv preprint arXiv:2301.11320 (2023) +61. Wang, Y., Shen, X., Hu, S.X., Yuan, Y., Crowley, J.L., Vaufreydaz, D.: Self-supervised transformers for unsupervised object discovery using normalized cut. In: Conference on Computer Vision and Pattern Recognition. New Orleans, LA, USA (June 2022) +62. Wang, Y., Sun, Y., Liu, Z., Sarma, S.E., Bronstein, M.M., Solomon, J.M.: Dynamic graph cnn for learning on point clouds. Acm Transactions On Graphics (tog) 38(5), 1-12 (2019) +63. Wu, W., Qi, Z., Fuxin, L.: Pointconv: Deep convolutional networks on 3d point clouds. In: Proceedings of the IEEE/CVF Conference on computer vision and pattern recognition. pp. 9621-9630 (2019) +64. Yang, Y., Wu, X., He, T., Zhao, H., Liu, X.: Sam3d: Segment anything in 3d scenes. arXiv preprint arXiv:2306.03908 (2023) +65. Zhang, B., Wonka, P.: Point cloud instance segmentation using probabilistic embeddings. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 8883-8892 (2021) +66. Zhang, Z., Ding, J., Jiang, L., Dai, D., Xia, G.S.: Freepoint: Unsupervised point cloud instance segmentation. arXiv preprint arXiv:2305.06973 (2023) +67. Zhang, Z., Yang, B., Wang, B., Li, B.: Growsp: Unsupervised semantic segmentation of 3d point clouds. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 17619-17629 (2023) \ No newline at end of file diff --git a/2024/Part2Object_ Hierarchical Unsupervised 3D Instance Segmentation/images.zip b/2024/Part2Object_ Hierarchical Unsupervised 3D Instance Segmentation/images.zip new file mode 100644 index 0000000000000000000000000000000000000000..c184a222e9f257689f437e61876c5b2e48eeceff --- /dev/null +++ b/2024/Part2Object_ Hierarchical Unsupervised 3D Instance Segmentation/images.zip @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:aa981db2aa825bcf8804f8c1a0a320ef87901124647ec13c96420cf3186d47a4 +size 410924 diff --git a/2024/Part2Object_ Hierarchical Unsupervised 3D Instance Segmentation/layout.json b/2024/Part2Object_ Hierarchical Unsupervised 3D Instance Segmentation/layout.json new file mode 100644 index 0000000000000000000000000000000000000000..a57307a823686b5e63b273f6a927640f51ec49a6 --- /dev/null +++ b/2024/Part2Object_ Hierarchical Unsupervised 3D Instance Segmentation/layout.json @@ -0,0 +1,9776 @@ +{ + "pdf_info": [ + { + "para_blocks": [ + { + "bbox": [ + 156, + 111, + 459, + 147 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 156, + 111, + 459, + 147 + ], + "spans": [ + { + "bbox": [ + 156, + 111, + 459, + 147 + ], + "type": "text", + "content": "Part2Object: Hierarchical Unsupervised 3D Instance Segmentation" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 196, + 168, + 416, + 193 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 196, + 168, + 416, + 193 + ], + "spans": [ + { + "bbox": [ + 196, + 168, + 416, + 193 + ], + "type": "text", + "content": "Cheng Shi*, Yulin Zhang*, Bin Yang, Jiajin Tang, Yuexin Ma, and Sibei Yang" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 209, + 202, + 402, + 236 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 209, + 202, + 402, + 236 + ], + "spans": [ + { + "bbox": [ + 209, + 202, + 402, + 236 + ], + "type": "text", + "content": "School of Information Science and Technology ShanghaiTech University {shicheng2022,yangsb}@shanghaiitech.edu.cn" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 160, + 267, + 452, + 443 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 160, + 267, + 452, + 443 + ], + "spans": [ + { + "bbox": [ + 160, + 267, + 452, + 443 + ], + "type": "text", + "content": "Abstract. Unsupervised 3D instance segmentation aims to segment objects from a 3D point cloud without any annotations. Existing methods face the challenge of either too loose or too tight clustering, leading to under-segmentation or over-segmentation. To address this issue, we propose Part2Object, hierarchical clustering with object guidance. Part2Object employs multi-layer clustering from points to object parts and objects, allowing objects to manifest at any layer. Additionally, it extracts and utilizes 3D objectness priors from temporally consecutive 2D RGB frames to guide the clustering process. Moreover, we propose Hi-Mask3D to support hierarchical 3D object part and instance segmentation. By training Hi-Mask3D on the objects and object parts extracted from Part2Object, we achieve consistent and superior performance compared to state-of-the-art models in various settings, including unsupervised instance segmentation, data-efficient fine-tuning, and cross-dataset generalization. Code is release at https://github.com/ChengShiest/Part2object." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 160, + 454, + 424, + 465 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 160, + 454, + 424, + 465 + ], + "spans": [ + { + "bbox": [ + 160, + 454, + 424, + 465 + ], + "type": "text", + "content": "Keywords: 3D Instance Segmentation " + }, + { + "bbox": [ + 160, + 454, + 424, + 465 + ], + "type": "inline_equation", + "content": "\\cdot" + }, + { + "bbox": [ + 160, + 454, + 424, + 465 + ], + "type": "text", + "content": " Unsupervised Learning" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 133, + 487, + 229, + 499 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 487, + 229, + 499 + ], + "spans": [ + { + "bbox": [ + 133, + 487, + 229, + 499 + ], + "type": "text", + "content": "1 Introduction" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 133, + 514, + 481, + 645 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 514, + 481, + 645 + ], + "spans": [ + { + "bbox": [ + 133, + 514, + 481, + 645 + ], + "type": "text", + "content": "3D instance segmentation, parsing semantic compositions from complex point cloud scenes into distinct objects, is fundamental and crucial in real-world applications, such as mixed reality [51], autonomous navigation [3, 52], planning [23], and manipulation [17]. In recent years, 3D instance segmentation has seen substantial advancements [19, 37, 38, 42, 55, 62], improving both the accuracy and efficiency of object localization and recognition. Nevertheless, it often demands large-scale human annotations for fully-supervised training, resulting in costly and manpower-intensive efforts. While a few studies [33, 47] have initiated to address 3D instance segmentation without human annotations, their reliance on 3D scene flows from sequential point clouds restricts their applicability to single point cloud scenarios, such as indoor scenes in ScanNet [12]." + } + ] + } + ], + "index": 6 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 142, + 654, + 337, + 665 + ], + "type": "page_footnote", + "angle": 0, + "lines": [ + { + "bbox": [ + 142, + 654, + 337, + 665 + ], + "spans": [ + { + "bbox": [ + 142, + 654, + 337, + 665 + ], + "type": "text", + "content": "* Equal contribution. " + }, + { + "bbox": [ + 142, + 654, + 337, + 665 + ], + "type": "inline_equation", + "content": "\\boxtimes" + }, + { + "bbox": [ + 142, + 654, + 337, + 665 + ], + "type": "text", + "content": " Corresponding author." + } + ] + } + ], + "index": 7 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 0 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 138, + 160, + 476, + 327 + ], + "blocks": [ + { + "bbox": [ + 130, + 114, + 482, + 159 + ], + "lines": [ + { + "bbox": [ + 130, + 114, + 482, + 159 + ], + "spans": [ + { + "bbox": [ + 130, + 114, + 482, + 159 + ], + "type": "text", + "content": "Fig. 1: Motivation of our hierarchical clustering. Single-level clustering results in a trade-off between under-segmentation for certain objects and over-segmentation for others. In contrast, our hierarchical clustering allows for gathering and identifying objects at varying levels of clustering granularity." + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 138, + 160, + 476, + 327 + ], + "lines": [ + { + "bbox": [ + 138, + 160, + 476, + 327 + ], + "spans": [ + { + "bbox": [ + 138, + 160, + 476, + 327 + ], + "type": "image", + "image_path": "9e2c399d4d0af5c900b794d764d3a5c384bacf2840baf4a16c5ae60fa0c5a65a.jpg" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_body" + } + ], + "index": 3 + }, + { + "bbox": [ + 130, + 352, + 482, + 508 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 352, + 482, + 508 + ], + "spans": [ + { + "bbox": [ + 130, + 352, + 482, + 508 + ], + "type": "text", + "content": "In this paper, we study unsupervised 3D instance segmentation, aiming to segment class-agnostic instances from indoor 3D scenes without relying on any human labels. To address this problem, two straightforward strategies have been explored recently: 1) to apply traditional clustering [16, 32] or graph-cut methods [66] to group points into objects based on their RGB-D data, such as coordinates, colors, normal vectors, and self-supervised pretraining features [41]. 2) To project 2D instance segmentation results from 2D RGB frames onto the point cloud using the corresponding camera pose. Typically, 2D RGB frames are captured concurrently with indoor point cloud data [12, 39], while 2D instance masks are extracted using 2D unsupervised instance segmentation advancements [59-61]. However, both strategies can only work in straightforward scenes with few salient objects, as they cannot tackle the fundamental challenges inherent in unsupervised 3D instance segmentation, as outlined below." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 130, + 510, + 482, + 666 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 510, + 482, + 666 + ], + "spans": [ + { + "bbox": [ + 130, + 510, + 482, + 666 + ], + "type": "text", + "content": "In complex scenes, achieving consistent segmentation granularity for different objects within a single clustering and graph-cut result is challenging, given the significant variations in geometric shape, color, and size among 3D objects, as well as the diverse compositions of these objects within scenes. This entails a trade-off between under-segmentation for certain objects and over-segmentation for others, making it almost impossible to attain satisfactory segmentation granularity for all of them. As shown in Fig 1, tighter grouping can accurately segment objects with simple geometric shapes and textures, such as the pillow on the bed (see layer one). However, it can also result in the over-segmentation of larger and more complex objects like the toilet in the same layer (layer one), and vice versa. In addition, there is a lack of an effective strategy to identify whether a point cluster or group represents a 3D instance. For example, the toilet back (see layer two) forms a well-grouped clus" + } + ] + } + ], + "index": 5 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "spans": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "type": "text", + "content": "2" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 166, + 91, + 217, + 102 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 166, + 91, + 217, + 102 + ], + "spans": [ + { + "bbox": [ + 166, + 91, + 217, + 102 + ], + "type": "text", + "content": "Cheng et al." + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 1 + }, + { + "para_blocks": [ + { + "bbox": [ + 130, + 116, + 482, + 297 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 116, + 482, + 297 + ], + "spans": [ + { + "bbox": [ + 130, + 116, + 482, + 297 + ], + "type": "text", + "content": "ter, but it does not constitute a meaningful object. The 3D geometric features are fundamentally semantically insufficient to identify objects. Recently, it has been discovered that self-supervised vision transformer (ViT) features contain information about object segmentations in 2D images [6, 60]. Building on this discovery, 3D unsupervised instance segmentation approaches [41] project the 2D self-supervised ViT features (or their inferred object segmentations) from 2D RGB frames onto 3D point clouds, thereby enhancing the perception of objectness for 3D point features. However, a 3D point typically corresponds to multiple pixels in different 2D RGB frames, and this many-to-one mapping is fragile, easily disrupted by noise features or imprecise masks on any frame, resulting in an inability to distinguish one object from others or the segmentation of objects into fragmented segments, as illustrated in Fig 3(b) and (c). In addition, despite enhanced point features, discerning nearby objects or those sharing similar semantics remains challenging due to the inherently greater complexity of 3D scenes compared to 2D images." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 130, + 301, + 482, + 601 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 301, + 482, + 601 + ], + "spans": [ + { + "bbox": [ + 130, + 301, + 482, + 601 + ], + "type": "text", + "content": "To tackle these challenges, we introduce a simple yet effective principle, \"Gather & Aim\", for the automatic discovery and segmentation of 3D objects. For the gathering, inspired by the observation that tighter or looser groupings can respectively include certain objects within a scene, we propose allowing the gathering of potential objects at varying levels of grouping granularity, rather than being restricted to a single level. Specifically, we perform hierarchical clustering in the 3D scene, gradually grouping points into larger clusters, as illustrated in Fig 1. Therefore, instead of striving to improve single-layer clustering results, ours is expected that as long as each potential instance appears in one of the layers, we will be able to lock onto the segmentation mask of that instance. This mitigates excessive reliance on any single-level clustering results, thereby enhancing the clustering's adaptability to significant variations among objects and scenes. For the aiming, we propose to target the 3D objectness priors, a set of candidate objects, from temporally consecutive 2D RGB frames. We leverage both the temporal consistency across frames and the prior knowledge of unsupervised object discovery within frames to extract the 3D objectness priors. Note that we jointly estimate 2D objects from multiple frames first and then approximate object-level 3D bounding boxes, departing from the point-level projection used in previous methods [41], to address the fragility issue of the many-to-one mapping from pixels to points. It avoids adhesive or scattered 3D fragments caused by inconsistencies and occlusions over 2D RGB frames. Finally, the 3D bounding boxes serve as indicators to identify the 3D instance masks within the multi-level clusters. Our \"Gather & Aim\" strategy enables leverage of 2D semantic priors for 3D object identification while harnessing 3D geometric priors for precise instance segmentation." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 130, + 605, + 482, + 666 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 605, + 482, + 666 + ], + "spans": [ + { + "bbox": [ + 130, + 605, + 482, + 666 + ], + "type": "text", + "content": "Furthermore, the taxonomy of how object parts are composed into an object is a strong yet disregarded prior information for unsupervised 3D instance segmentation. For example, the taxonomy based on the parts of a chair, including the back, arm, seat, and leg, can assist in addressing highly challenging scenarios, such as distinguishing two closely positioned and semantically similar" + } + ] + } + ], + "index": 4 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 395, + 91, + 447, + 102 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 395, + 91, + 447, + 102 + ], + "spans": [ + { + "bbox": [ + 395, + 91, + 447, + 102 + ], + "type": "text", + "content": "Part2Object" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 474, + 92, + 481, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 474, + 92, + 481, + 100 + ], + "spans": [ + { + "bbox": [ + 474, + 92, + 481, + 100 + ], + "type": "text", + "content": "3" + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 2 + }, + { + "para_blocks": [ + { + "bbox": [ + 130, + 116, + 482, + 236 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 116, + 482, + 236 + ], + "spans": [ + { + "bbox": [ + 130, + 116, + 482, + 236 + ], + "type": "text", + "content": "chairs as two separate instances. Thanks to our hierarchical clustering, we cannot only identify objects but also trace their constituent parts (see the toilet and its parts in layers three and two, respectively). Therefore, we extract both objects and object parts from our hierarchical clustering and name our clustering algorithm Part2Object. Expanding on this, we enhance the 3D instance segmentation framework, Mask3D [42], to support hierarchical 3D part and instance segmentation. Our improved model, named Hi-Mask3D, takes the object and object parts from Part2Object as pseudo-labels for learning 3D instance segmentation by incorporating explicit interactions between object parts and objects." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 130, + 236, + 482, + 308 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 236, + 482, + 308 + ], + "spans": [ + { + "bbox": [ + 130, + 236, + 482, + 308 + ], + "type": "text", + "content": "To evaluate the effectiveness of our Part2Object clustering and Hi-Mask3D model, we conduct experiments on the challenging and cluttered indoor environments [4,12,39] in three settings: 1) direct evaluation on unsupervised instance segmentation, 2) application to data-efficient fine-tuning, and 3) cross-dataset generalization, all showing significant performance improvement. In summary, our contributions are multi-fold:" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 138, + 319, + 479, + 378 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 319, + 479, + 378 + ], + "spans": [ + { + "bbox": [ + 138, + 319, + 479, + 378 + ], + "type": "text", + "content": "- We propose two key insights for unsupervised 3D instance segmentation: 1) Employing hierarchical clustering enables the gathering of objects at varying levels of clustering granularity. 2) Leveraging 3D objectness priors from temporally consecutive 2D frames as guidance, while harnessing 3D geometric priors to clustering on the point cloud for precise instance segmentation." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 138, + 380, + 480, + 449 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 380, + 480, + 449 + ], + "spans": [ + { + "bbox": [ + 138, + 380, + 480, + 449 + ], + "type": "text", + "content": "- Based on our insights, we propose an innovative hierarchical clustering approach, Part2Object. It progressively groups points into object parts and objects while extracting and leveraging 3D objectness priors to guide the clustering process. Our Part2Object significantly outperforms the state-of-the-art training-free unsupervised 3D instance segmentation methods by " + }, + { + "bbox": [ + 138, + 380, + 480, + 449 + ], + "type": "inline_equation", + "content": "16.8\\%" + }, + { + "bbox": [ + 138, + 380, + 480, + 449 + ], + "type": "text", + "content": " mAP@50 on the ScanNet dataset." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 138, + 450, + 480, + 498 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 450, + 480, + 498 + ], + "spans": [ + { + "bbox": [ + 138, + 450, + 480, + 498 + ], + "type": "text", + "content": "- We propose Hi-Mask3D, an extension of 3D instance segmentation to support hierarchical unsupervised 3D part and instance segmentation. Experiments demonstrate that Hi-Mask3D consistently and significantly outperforms state-of-the-art models in all the settings." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 132, + 518, + 237, + 530 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 518, + 237, + 530 + ], + "spans": [ + { + "bbox": [ + 132, + 518, + 237, + 530 + ], + "type": "text", + "content": "2 Related Work" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 130, + 545, + 482, + 666 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 545, + 482, + 666 + ], + "spans": [ + { + "bbox": [ + 130, + 545, + 482, + 666 + ], + "type": "text", + "content": "Unsupervised 3D instance segmentation. 3D instance segmentation is an essential task in 3D scene understanding, which aims to locate and recognize different objects in a 3D point cloud. However, the cost and labor intensity of 3D scene-level instance annotation underscore the significance of exploring challenging unsupervised instance segmentation methods. Early works utilize raw geometric information, including coordinates, colors, and normal vectors, to perform traditional clustering [15,16,32] for segmentation. Recently, inspired by the new paradigm in 2D unsupervised instance segmentation, some works [41,66] introduce pseudo labels and self-training to unsupervised 3D instance segmentation. Specifically, these methods use 2D or 3D self-supervised models to extract" + } + ] + } + ], + "index": 8 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "spans": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "type": "text", + "content": "4" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 166, + 91, + 217, + 102 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 166, + 91, + 217, + 102 + ], + "spans": [ + { + "bbox": [ + 166, + 91, + 217, + 102 + ], + "type": "text", + "content": "Cheng et al." + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 3 + }, + { + "para_blocks": [ + { + "bbox": [ + 130, + 116, + 482, + 176 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 116, + 482, + 176 + ], + "spans": [ + { + "bbox": [ + 130, + 116, + 482, + 176 + ], + "type": "text", + "content": "point cloud features, applying graph-cut algorithms to generate pseudo-labels for training and prediction. As illustrated earlier, the clustering method encounters a trade-off between under-segmentation and over-segmentation. Relying on a single graph-cut algorithm, as recent works do, falls short of achieving consistent granularity in complex indoor environments." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 130, + 176, + 482, + 332 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 176, + 482, + 332 + ], + "spans": [ + { + "bbox": [ + 130, + 176, + 482, + 332 + ], + "type": "text", + "content": "Transfer 2D foundation models into 3D. The rapid advancements in 2D vision have given rise to powerful foundation models [2,6,13,18,26,28,43-45,53], proving beneficial across a variety of visual tasks. In contrast to the 2D domain, 3D vision encounters challenges with data scarcity and training limitations, hindering the development of foundation models. Consequently, several works in the 3D turn to leveraging 2D foundation models to address 3D problems. In 3D instance segmentation, some works [7,64] utilize 2D vision foundation models to generate pseudo-labels for 2D images, aiding in the training and prediction of 3D models through 2D to 3D projection algorithms. Other works [31,36] employ features rather than outputs obtained from 2D foundation models, enhancing the feature extraction capabilities of 3D models through pixel-point alignment. However, relying solely on 2D features or 2D pseudo-labels poses challenges in addressing cluttered indoor scenes with stacked objects." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 130, + 332, + 482, + 502 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 332, + 482, + 502 + ], + "spans": [ + { + "bbox": [ + 130, + 332, + 482, + 502 + ], + "type": "text", + "content": "Supervised point cloud segmentation. Supervised point cloud segmentation can be divided into instance-level and part-level segmentation. The former has been significant breakthroughs in recent years [8, 19, 20, 22, 25, 27, 37, 38, 42, 50, 55, 56, 58, 62, 63]. However, these approaches demand a substantial amount of annotated data and focus on learning object-level attributes, neglecting the understanding of parts and their contributions to the composition of objects. In the realm of 3D part segmentation, various setups, including supervised [30,37,65], weakly supervised [9,10,57] instance segmentation and semantic segmentation, as well as open-world semantic segmentation [24, 36], have been extensively explored. While these approaches have demonstrated promising result, their primary emphasis lies in determining how to divide an object into parts. To integrate instance-level and part-level perspectives, we perform hierarchical clustering with the prior knowledge of how parts compose objects, unifying instance-level object identification and part-level constituent tracing." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 132, + 519, + 233, + 533 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 519, + 233, + 533 + ], + "spans": [ + { + "bbox": [ + 132, + 519, + 233, + 533 + ], + "type": "text", + "content": "3 Methodology" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 130, + 545, + 482, + 629 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 545, + 482, + 629 + ], + "spans": [ + { + "bbox": [ + 130, + 545, + 482, + 629 + ], + "type": "text", + "content": "Problem definition. Unsupervised 3D instance segmentation [41, 67] is the task of generating class-agnostic masks for all foreground objects in a 3D point cloud scene solely based on raw RGB-D data, without relying on any object detection or instance segmentation annotations. The raw RGB-D data [12, 39] consists of the point cloud " + }, + { + "bbox": [ + 130, + 545, + 482, + 629 + ], + "type": "inline_equation", + "content": "P \\in \\mathbb{R}^{N \\times 3}" + }, + { + "bbox": [ + 130, + 545, + 482, + 629 + ], + "type": "text", + "content": " and corresponding 2D RGB frames " + }, + { + "bbox": [ + 130, + 545, + 482, + 629 + ], + "type": "inline_equation", + "content": "\\{I_i\\}_{i=1}^M" + }, + { + "bbox": [ + 130, + 545, + 482, + 629 + ], + "type": "text", + "content": ", where each image frame " + }, + { + "bbox": [ + 130, + 545, + 482, + 629 + ], + "type": "inline_equation", + "content": "I_m \\in \\mathbb{R}^{H \\times W \\times 3}" + }, + { + "bbox": [ + 130, + 545, + 482, + 629 + ], + "type": "text", + "content": ". With raw RGB-D data, 3D instance segmentation task aims to output all object masks " + }, + { + "bbox": [ + 130, + 545, + 482, + 629 + ], + "type": "inline_equation", + "content": "P_{\\mathrm{object}}" + }, + { + "bbox": [ + 130, + 545, + 482, + 629 + ], + "type": "text", + "content": " in the scene." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 130, + 630, + 482, + 666 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 630, + 482, + 666 + ], + "spans": [ + { + "bbox": [ + 130, + 630, + 482, + 666 + ], + "type": "text", + "content": "Method overview. As shown in Fig 2, our unsupervised framework has two major components: 1) the hierarchical clustering algorithm (Sec 3.1 and Sec 3.2), Part2Object, which progressively group points into larger clusters. Duringclus" + } + ] + } + ], + "index": 7 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 395, + 91, + 447, + 102 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 395, + 91, + 447, + 102 + ], + "spans": [ + { + "bbox": [ + 395, + 91, + 447, + 102 + ], + "type": "text", + "content": "Part2Object" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 474, + 92, + 480, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 474, + 92, + 480, + 100 + ], + "spans": [ + { + "bbox": [ + 474, + 92, + 480, + 100 + ], + "type": "text", + "content": "5" + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 4 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 136, + 193, + 480, + 359 + ], + "blocks": [ + { + "bbox": [ + 130, + 114, + 482, + 180 + ], + "lines": [ + { + "bbox": [ + 130, + 114, + 482, + 180 + ], + "spans": [ + { + "bbox": [ + 130, + 114, + 482, + 180 + ], + "type": "text", + "content": "Fig. 2: Overview of our Part2Object hierarchical clustering and Hi-Mask3D instance segmentation framework. Part2Object extracts 3D objectness priors from consecutive 2D RGB frames and uses them to guide hierarchical clustering from points to object parts and objects. Hi-Mask3D utilizes objects and parts identified by Part2Object as pseudo-labels, learning for improving instance segmentation through the utilization of object parts." + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 136, + 193, + 480, + 359 + ], + "lines": [ + { + "bbox": [ + 136, + 193, + 480, + 359 + ], + "spans": [ + { + "bbox": [ + 136, + 193, + 480, + 359 + ], + "type": "image", + "image_path": "2c25a136c403368b85f58a5414f7feb2309379dbd617bc02a166b98383d86538.jpg" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_body" + } + ], + "index": 3 + }, + { + "bbox": [ + 130, + 394, + 482, + 491 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 394, + 482, + 491 + ], + "spans": [ + { + "bbox": [ + 130, + 394, + 482, + 491 + ], + "type": "text", + "content": "tering, aided by objectness priors from 2D frames, we can effectively obtain both object-level and part-level segmentation masks, which serve as pseudo-labels for learning unsupervised instance segmentation. 2) The end-to-end 3D instance segmentation model Hi-Mask3D (Sec 3.3) extends the original Mask3D model [42] by explicitly predicting segmentation results of object parts and leveraging them to aid instance segmentation. By self-training on the pseudo-labels extracted from the first-stage Part2Object, Hi-Mask3D can effectively predict 3D instance segments." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 131, + 510, + 376, + 523 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 131, + 510, + 376, + 523 + ], + "spans": [ + { + "bbox": [ + 131, + 510, + 376, + 523 + ], + "type": "text", + "content": "3.1 Hierarchical Clustering on 3D Point Cloud" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 130, + 533, + 483, + 666 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 533, + 483, + 666 + ], + "spans": [ + { + "bbox": [ + 130, + 533, + 483, + 666 + ], + "type": "text", + "content": "In this section, we introduce the hierarchical clustering process of our Part2Object which starts by grouping initial points into initial clusters and then progressively merging them into larger clusters. We will sequentially cover: 1) the feature representation of points, 2) the initialization of clusters from points, and 3) the hierarchical clustering process for merging clusters. Point Feature Representation. Given a 3D surface point cloud " + }, + { + "bbox": [ + 130, + 533, + 483, + 666 + ], + "type": "inline_equation", + "content": "P" + }, + { + "bbox": [ + 130, + 533, + 483, + 666 + ], + "type": "text", + "content": ", we denote its norm and color as " + }, + { + "bbox": [ + 130, + 533, + 483, + 666 + ], + "type": "inline_equation", + "content": "P^{\\mathrm{Norm}}" + }, + { + "bbox": [ + 130, + 533, + 483, + 666 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 130, + 533, + 483, + 666 + ], + "type": "inline_equation", + "content": "P^{\\mathrm{RGB}}" + }, + { + "bbox": [ + 130, + 533, + 483, + 666 + ], + "type": "text", + "content": ", respectively. The 3D feature " + }, + { + "bbox": [ + 130, + 533, + 483, + 666 + ], + "type": "inline_equation", + "content": "f_{i}^{3\\mathrm{D}}" + }, + { + "bbox": [ + 130, + 533, + 483, + 666 + ], + "type": "text", + "content": " of point " + }, + { + "bbox": [ + 130, + 533, + 483, + 666 + ], + "type": "inline_equation", + "content": "p_i \\in P" + }, + { + "bbox": [ + 130, + 533, + 483, + 666 + ], + "type": "text", + "content": " comprises " + }, + { + "bbox": [ + 130, + 533, + 483, + 666 + ], + "type": "inline_equation", + "content": "p_i" + }, + { + "bbox": [ + 130, + 533, + 483, + 666 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 130, + 533, + 483, + 666 + ], + "type": "inline_equation", + "content": "p_i^{\\mathrm{Norm}}" + }, + { + "bbox": [ + 130, + 533, + 483, + 666 + ], + "type": "text", + "content": ", and " + }, + { + "bbox": [ + 130, + 533, + 483, + 666 + ], + "type": "inline_equation", + "content": "p_i^{\\mathrm{RGB}}" + }, + { + "bbox": [ + 130, + 533, + 483, + 666 + ], + "type": "text", + "content": ", representing the location, color, and geometry information, respectively. Additionally, we define point cloud 2D semantics feature " + }, + { + "bbox": [ + 130, + 533, + 483, + 666 + ], + "type": "inline_equation", + "content": "f_{i}" + }, + { + "bbox": [ + 130, + 533, + 483, + 666 + ], + "type": "text", + "content": " by projecting 2D features [6] of RGB frames into 3D space using the corresponding camera pose. Here, 2D features are extracted using the 2D self-supervised" + } + ] + } + ], + "index": 6 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "spans": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "type": "text", + "content": "6" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 166, + 91, + 217, + 102 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 166, + 91, + 217, + 102 + ], + "spans": [ + { + "bbox": [ + 166, + 91, + 217, + 102 + ], + "type": "text", + "content": "Cheng et al." + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 5 + }, + { + "para_blocks": [ + { + "bbox": [ + 130, + 116, + 482, + 175 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 116, + 482, + 175 + ], + "spans": [ + { + "bbox": [ + 130, + 116, + 482, + 175 + ], + "type": "text", + "content": "method DINO [6], without rely on any 2D instance annotations. Cluster Initialization and Feature Representation. Given the vast number of points in the point cloud data [1], we initially apply the VCCS algorithm [34] to generate super-points on the point cloud, forming the first-layer clusters " + }, + { + "bbox": [ + 130, + 116, + 482, + 175 + ], + "type": "inline_equation", + "content": "\\{c_i^0\\}_{i = 1}^{N_0}" + }, + { + "bbox": [ + 130, + 116, + 482, + 175 + ], + "type": "text", + "content": " as follows:" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 232, + 177, + 481, + 192 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 232, + 177, + 481, + 192 + ], + "spans": [ + { + "bbox": [ + 232, + 177, + 481, + 192 + ], + "type": "interline_equation", + "content": "\\{c _ {i} ^ {0} \\} _ {i = 1} ^ {N _ {0}} = \\mathrm {V C C S} (P, P ^ {\\mathrm {N o r m}}, P ^ {\\mathrm {R G B}}), \\tag {1}", + "image_path": "51f8b9149dd5b65a74ff3e9c8b4e7d0093df763ea46df9f650744802c41b1f84.jpg" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 130, + 199, + 482, + 344 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 199, + 482, + 344 + ], + "spans": [ + { + "bbox": [ + 130, + 199, + 482, + 344 + ], + "type": "text", + "content": "where VCCS algorithm takes points' coordinates " + }, + { + "bbox": [ + 130, + 199, + 482, + 344 + ], + "type": "inline_equation", + "content": "P" + }, + { + "bbox": [ + 130, + 199, + 482, + 344 + ], + "type": "text", + "content": ", norms " + }, + { + "bbox": [ + 130, + 199, + 482, + 344 + ], + "type": "inline_equation", + "content": "P^{\\mathrm{Norm}}" + }, + { + "bbox": [ + 130, + 199, + 482, + 344 + ], + "type": "text", + "content": ", and colors " + }, + { + "bbox": [ + 130, + 199, + 482, + 344 + ], + "type": "inline_equation", + "content": "P^{\\mathrm{RGB}}" + }, + { + "bbox": [ + 130, + 199, + 482, + 344 + ], + "type": "text", + "content": " as input and group original " + }, + { + "bbox": [ + 130, + 199, + 482, + 344 + ], + "type": "inline_equation", + "content": "N" + }, + { + "bbox": [ + 130, + 199, + 482, + 344 + ], + "type": "text", + "content": " points into " + }, + { + "bbox": [ + 130, + 199, + 482, + 344 + ], + "type": "inline_equation", + "content": "N_0" + }, + { + "bbox": [ + 130, + 199, + 482, + 344 + ], + "type": "text", + "content": " super-points as " + }, + { + "bbox": [ + 130, + 199, + 482, + 344 + ], + "type": "inline_equation", + "content": "\\{c_i^0\\}_{i=1}^{N_0}" + }, + { + "bbox": [ + 130, + 199, + 482, + 344 + ], + "type": "text", + "content": ". We consider each super-point " + }, + { + "bbox": [ + 130, + 199, + 482, + 344 + ], + "type": "inline_equation", + "content": "c_i^0" + }, + { + "bbox": [ + 130, + 199, + 482, + 344 + ], + "type": "text", + "content": " as a cluster, consisting of multiple points that are closely positioned and share similar color and norm characteristics. Then, we define the feature representation for the cluster based on the features of the points within the cluster. Instead of directly averaging point features within the cluster, we compute a weight for each point and perform a weighted sum of features. These weights are determined by the similarity between each point's feature and the average feature of points within the cluster. This approach helps mitigate the influence of noisy points in the cluster, leading to more robust representations. Specifically, for the cluster " + }, + { + "bbox": [ + 130, + 199, + 482, + 344 + ], + "type": "inline_equation", + "content": "c_i^0" + }, + { + "bbox": [ + 130, + 199, + 482, + 344 + ], + "type": "text", + "content": ", its cluster feature " + }, + { + "bbox": [ + 130, + 199, + 482, + 344 + ], + "type": "inline_equation", + "content": "f_i^0" + }, + { + "bbox": [ + 130, + 199, + 482, + 344 + ], + "type": "text", + "content": " is computed as follows," + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 172, + 354, + 481, + 389 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 172, + 354, + 481, + 389 + ], + "spans": [ + { + "bbox": [ + 172, + 354, + 481, + 389 + ], + "type": "interline_equation", + "content": "\\boldsymbol {f} _ {i} ^ {0} = \\sum_ {p _ {j} \\in c _ {i} ^ {0}} \\frac {\\operatorname {s i m} \\left(\\boldsymbol {f} _ {j} , \\bar {\\boldsymbol {f}} _ {i} ^ {0}\\right)}{\\sum p _ {j} \\in c _ {j} ^ {0} \\operatorname {s i m} \\left(\\boldsymbol {f} _ {j} , \\bar {\\boldsymbol {f}} _ {i} ^ {0}\\right)} \\boldsymbol {f} _ {j}, \\text {w h e r e} \\bar {\\boldsymbol {f}} _ {i} ^ {0} = \\sum_ {p _ {j} \\in c _ {i} ^ {0}} \\frac {1}{\\left| c _ {i} ^ {0} \\right|} \\boldsymbol {f} _ {j}. \\tag {2}", + "image_path": "2ac60208928842b2d450f0f61c7acdd9e3aa13335b86d82beeb6ea840cb1654a.jpg" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 130, + 400, + 482, + 490 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 400, + 482, + 490 + ], + "spans": [ + { + "bbox": [ + 130, + 400, + 482, + 490 + ], + "type": "text", + "content": "The " + }, + { + "bbox": [ + 130, + 400, + 482, + 490 + ], + "type": "inline_equation", + "content": "p_j \\in c_i^0" + }, + { + "bbox": [ + 130, + 400, + 482, + 490 + ], + "type": "text", + "content": " denotes each point " + }, + { + "bbox": [ + 130, + 400, + 482, + 490 + ], + "type": "inline_equation", + "content": "p_j" + }, + { + "bbox": [ + 130, + 400, + 482, + 490 + ], + "type": "text", + "content": " within the cluster " + }, + { + "bbox": [ + 130, + 400, + 482, + 490 + ], + "type": "inline_equation", + "content": "c_i^0" + }, + { + "bbox": [ + 130, + 400, + 482, + 490 + ], + "type": "text", + "content": ", and " + }, + { + "bbox": [ + 130, + 400, + 482, + 490 + ], + "type": "inline_equation", + "content": "\\pmb{f}_j" + }, + { + "bbox": [ + 130, + 400, + 482, + 490 + ], + "type": "text", + "content": " is the feature of point " + }, + { + "bbox": [ + 130, + 400, + 482, + 490 + ], + "type": "inline_equation", + "content": "p_j" + }, + { + "bbox": [ + 130, + 400, + 482, + 490 + ], + "type": "text", + "content": ". The " + }, + { + "bbox": [ + 130, + 400, + 482, + 490 + ], + "type": "inline_equation", + "content": "\\bar{\\pmb{f}}_i^0" + }, + { + "bbox": [ + 130, + 400, + 482, + 490 + ], + "type": "text", + "content": " is the average feature of points within " + }, + { + "bbox": [ + 130, + 400, + 482, + 490 + ], + "type": "inline_equation", + "content": "c_i^0" + }, + { + "bbox": [ + 130, + 400, + 482, + 490 + ], + "type": "text", + "content": ", and " + }, + { + "bbox": [ + 130, + 400, + 482, + 490 + ], + "type": "inline_equation", + "content": "\\sin(\\cdot, \\cdot)" + }, + { + "bbox": [ + 130, + 400, + 482, + 490 + ], + "type": "text", + "content": " denotes the computation of cosine similarity. As depicted in the example in Fig 3(a), our cluster feature " + }, + { + "bbox": [ + 130, + 400, + 482, + 490 + ], + "type": "inline_equation", + "content": "\\pmb{f}_i^0" + }, + { + "bbox": [ + 130, + 400, + 482, + 490 + ], + "type": "text", + "content": " demonstrates greater robustness compared to " + }, + { + "bbox": [ + 130, + 400, + 482, + 490 + ], + "type": "inline_equation", + "content": "\\bar{\\pmb{f}}_i^0" + }, + { + "bbox": [ + 130, + 400, + 482, + 490 + ], + "type": "text", + "content": ". For simplicity, we abbreviate the operations in Equ 2 as function " + }, + { + "bbox": [ + 130, + 400, + 482, + 490 + ], + "type": "inline_equation", + "content": "\\mathsf{FU}(\\cdot)" + }, + { + "bbox": [ + 130, + 400, + 482, + 490 + ], + "type": "text", + "content": ", which computes the cluster feature for the inputted cluster based on point features, e.g., " + }, + { + "bbox": [ + 130, + 400, + 482, + 490 + ], + "type": "inline_equation", + "content": "\\pmb{f}_i^0 = \\mathsf{FU}(c_i^0)" + }, + { + "bbox": [ + 130, + 400, + 482, + 490 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 130, + 491, + 482, + 628 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 491, + 482, + 628 + ], + "spans": [ + { + "bbox": [ + 130, + 491, + 482, + 628 + ], + "type": "text", + "content": "Hierarchical Clustering with One Stop Criteria. Next, we group and merge the first-layer clusters " + }, + { + "bbox": [ + 130, + 491, + 482, + 628 + ], + "type": "inline_equation", + "content": "\\{c_i^0\\}_{i = 1}^{N_0}" + }, + { + "bbox": [ + 130, + 491, + 482, + 628 + ], + "type": "text", + "content": " to the next-layer clusters " + }, + { + "bbox": [ + 130, + 491, + 482, + 628 + ], + "type": "inline_equation", + "content": "\\{c_i^1\\}_{i = 1}^{N_1}" + }, + { + "bbox": [ + 130, + 491, + 482, + 628 + ], + "type": "text", + "content": " based on their features " + }, + { + "bbox": [ + 130, + 491, + 482, + 628 + ], + "type": "inline_equation", + "content": "\\{\\pmb {f}_i^0\\}_{i = 0}^{N_0}" + }, + { + "bbox": [ + 130, + 491, + 482, + 628 + ], + "type": "text", + "content": " and spatial coordinates, progressively and iteratively forming higher-hierarchical clusters " + }, + { + "bbox": [ + 130, + 491, + 482, + 628 + ], + "type": "inline_equation", + "content": "\\{c_{ij}^{t}\\}_{i = 1}^{N_t}" + }, + { + "bbox": [ + 130, + 491, + 482, + 628 + ], + "type": "text", + "content": ", where " + }, + { + "bbox": [ + 130, + 491, + 482, + 628 + ], + "type": "inline_equation", + "content": "t" + }, + { + "bbox": [ + 130, + 491, + 482, + 628 + ], + "type": "text", + "content": " represents the " + }, + { + "bbox": [ + 130, + 491, + 482, + 628 + ], + "type": "inline_equation", + "content": "t" + }, + { + "bbox": [ + 130, + 491, + 482, + 628 + ], + "type": "text", + "content": "-th layer in hierarchical clustering. In each single clustering layer, our clustering principle is that two clusters can only be merged when they are semantically similar and spatially adjacent: 1) the feature similarity between them ranks among the top " + }, + { + "bbox": [ + 130, + 491, + 482, + 628 + ], + "type": "inline_equation", + "content": "K" + }, + { + "bbox": [ + 130, + 491, + 482, + 628 + ], + "type": "text", + "content": " similarities between any pair of clusters. 2) The closest points between two clusters are adjacent. Specifically, when considering two clusters in the " + }, + { + "bbox": [ + 130, + 491, + 482, + 628 + ], + "type": "inline_equation", + "content": "t" + }, + { + "bbox": [ + 130, + 491, + 482, + 628 + ], + "type": "text", + "content": "-th layer, " + }, + { + "bbox": [ + 130, + 491, + 482, + 628 + ], + "type": "inline_equation", + "content": "c_{i}^{t}" + }, + { + "bbox": [ + 130, + 491, + 482, + 628 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 130, + 491, + 482, + 628 + ], + "type": "inline_equation", + "content": "c_{j}^{t}" + }, + { + "bbox": [ + 130, + 491, + 482, + 628 + ], + "type": "text", + "content": ", along with their features " + }, + { + "bbox": [ + 130, + 491, + 482, + 628 + ], + "type": "inline_equation", + "content": "\\pmb {f}_i^t" + }, + { + "bbox": [ + 130, + 491, + 482, + 628 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 130, + 491, + 482, + 628 + ], + "type": "inline_equation", + "content": "\\pmb {f}_j^t" + }, + { + "bbox": [ + 130, + 491, + 482, + 628 + ], + "type": "text", + "content": ", the merging process to form the next-layer cluster " + }, + { + "bbox": [ + 130, + 491, + 482, + 628 + ], + "type": "inline_equation", + "content": "c_k^{t + 1}" + }, + { + "bbox": [ + 130, + 491, + 482, + 628 + ], + "type": "text", + "content": " and feature " + }, + { + "bbox": [ + 130, + 491, + 482, + 628 + ], + "type": "inline_equation", + "content": "\\pmb {f}_k^{t + 1}" + }, + { + "bbox": [ + 130, + 491, + 482, + 628 + ], + "type": "text", + "content": " is as follows:" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 169, + 636, + 446, + 669 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 169, + 636, + 446, + 669 + ], + "spans": [ + { + "bbox": [ + 169, + 636, + 446, + 669 + ], + "type": "interline_equation", + "content": "\\begin{array}{l} c _ {k} ^ {t + 1} = c _ {i} ^ {t} \\cup c _ {j} ^ {t} \\quad \\mathrm {i f} \\operatorname {r a n k} (\\operatorname {s i m} (\\pmb {f} _ {i} ^ {t}, \\pmb {f} _ {j} ^ {t})) \\leq K \\mathrm {a n d} \\operatorname {d i s t} (c _ {i} ^ {t}, c _ {j} ^ {t}) \\leq T, \\\\ \\boldsymbol {f} _ {k} ^ {t + 1} = \\mathrm {F U} (c _ {i} ^ {t} \\cup c _ {j} ^ {t}), \\\\ \\end{array}", + "image_path": "594bebaaaba9f6eef45c4387b3ad9c9e2ea194d3e0447e165e3fdc19f0c28f84.jpg" + } + ] + } + ], + "index": 8 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 395, + 91, + 447, + 102 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 395, + 91, + 447, + 102 + ], + "spans": [ + { + "bbox": [ + 395, + 91, + 447, + 102 + ], + "type": "text", + "content": "Part2Object" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 474, + 91, + 481, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 474, + 91, + 481, + 100 + ], + "spans": [ + { + "bbox": [ + 474, + 91, + 481, + 100 + ], + "type": "text", + "content": "7" + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 6 + }, + { + "para_blocks": [ + { + "bbox": [ + 130, + 116, + 482, + 332 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 116, + 482, + 332 + ], + "spans": [ + { + "bbox": [ + 130, + 116, + 482, + 332 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 130, + 116, + 482, + 332 + ], + "type": "inline_equation", + "content": "\\mathrm{rank}(\\cdot)" + }, + { + "bbox": [ + 130, + 116, + 482, + 332 + ], + "type": "text", + "content": " denotes the ranking of feature similarity among pairwise cluster similarities in this clustering layer, where a higher rank signifies a higher similarity, and " + }, + { + "bbox": [ + 130, + 116, + 482, + 332 + ], + "type": "inline_equation", + "content": "\\mathrm{dist}(\\cdot,\\cdot)" + }, + { + "bbox": [ + 130, + 116, + 482, + 332 + ], + "type": "text", + "content": " computes the Euclidean distance between closest points of the two clusters. The " + }, + { + "bbox": [ + 130, + 116, + 482, + 332 + ], + "type": "inline_equation", + "content": "K" + }, + { + "bbox": [ + 130, + 116, + 482, + 332 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 130, + 116, + 482, + 332 + ], + "type": "inline_equation", + "content": "T" + }, + { + "bbox": [ + 130, + 116, + 482, + 332 + ], + "type": "text", + "content": " represent the threshold values for ranking and distance, respectively. However, we observe that solely using the merging metric in Equ 3 can result in the incorrect merging of parts from different objects in the early stages of clustering (i.e., shallow layers of clustering). This occurs because clusters at the shallow layers are typically scattered fragments with highly local features, which hinder them from perceiving the entire object. To track this issue, we first extract 3D objectness priors, represented by a set of 3D bounding boxes of potential objects " + }, + { + "bbox": [ + 130, + 116, + 482, + 332 + ], + "type": "inline_equation", + "content": "B^{3\\mathrm{D}}" + }, + { + "bbox": [ + 130, + 116, + 482, + 332 + ], + "type": "text", + "content": ", detailed in Sec 3.2. Then, we design a stopping criteria based on the 3D objectness priors to prevent clusters belonging to different objects from being merged. Criteria: Stop merging clusters belonging to different objects! Specifically, for any pair of clusters " + }, + { + "bbox": [ + 130, + 116, + 482, + 332 + ], + "type": "inline_equation", + "content": "c_{i}^{t}" + }, + { + "bbox": [ + 130, + 116, + 482, + 332 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 130, + 116, + 482, + 332 + ], + "type": "inline_equation", + "content": "c_{j}^{t}" + }, + { + "bbox": [ + 130, + 116, + 482, + 332 + ], + "type": "text", + "content": " that meet the merging metric, we extra assess their spatial relationship relative to each 3D object " + }, + { + "bbox": [ + 130, + 116, + 482, + 332 + ], + "type": "inline_equation", + "content": "b_{k}^{3\\mathrm{D}} \\in B^{3\\mathrm{D}}" + }, + { + "bbox": [ + 130, + 116, + 482, + 332 + ], + "type": "text", + "content": ". If clusters " + }, + { + "bbox": [ + 130, + 116, + 482, + 332 + ], + "type": "inline_equation", + "content": "c_{i}^{t}" + }, + { + "bbox": [ + 130, + 116, + 482, + 332 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 130, + 116, + 482, + 332 + ], + "type": "inline_equation", + "content": "c_{j}^{t}" + }, + { + "bbox": [ + 130, + 116, + 482, + 332 + ], + "type": "text", + "content": " are respectively inside and outside the object " + }, + { + "bbox": [ + 130, + 116, + 482, + 332 + ], + "type": "inline_equation", + "content": "b_{k}^{3\\mathrm{D}}" + }, + { + "bbox": [ + 130, + 116, + 482, + 332 + ], + "type": "text", + "content": ", then reject their merging because they do not belong to the same object." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 130, + 333, + 482, + 477 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 333, + 482, + 477 + ], + "spans": [ + { + "bbox": [ + 130, + 333, + 482, + 477 + ], + "type": "text", + "content": "Collect Objects and Parts from Clustering. By employing hierarchical clustering with the stopping criteria, we identify clusters that stop merging with others as 3D objects, denoted as " + }, + { + "bbox": [ + 130, + 333, + 482, + 477 + ], + "type": "inline_equation", + "content": "\\hat{P}_{\\mathrm{object}}" + }, + { + "bbox": [ + 130, + 333, + 482, + 477 + ], + "type": "text", + "content": ". Additionally, in hierarchical clustering, objects with complex geometric are typically merged from scattered fragments into meaningful parts, and then into complete objects. As shown in Fig 1, scattered \"toilet\" fragments are combined into object parts such as the toilet back and seat (see layer two), which are then formed into the object \"toilet\" (layer three). Therefore, we can trace back from objects " + }, + { + "bbox": [ + 130, + 333, + 482, + 477 + ], + "type": "inline_equation", + "content": "\\hat{P}_{\\mathrm{object}}" + }, + { + "bbox": [ + 130, + 333, + 482, + 477 + ], + "type": "text", + "content": " to identify the clusters from the previous level that form them, considering these clusters as potential object parts, denoted as " + }, + { + "bbox": [ + 130, + 333, + 482, + 477 + ], + "type": "inline_equation", + "content": "\\hat{P}_{\\mathrm{part}}" + }, + { + "bbox": [ + 130, + 333, + 482, + 477 + ], + "type": "text", + "content": ". Both objects " + }, + { + "bbox": [ + 130, + 333, + 482, + 477 + ], + "type": "inline_equation", + "content": "\\hat{P}_{\\mathrm{object}}" + }, + { + "bbox": [ + 130, + 333, + 482, + 477 + ], + "type": "text", + "content": " and object parts " + }, + { + "bbox": [ + 130, + 333, + 482, + 477 + ], + "type": "inline_equation", + "content": "\\hat{P}_{\\mathrm{part}}" + }, + { + "bbox": [ + 130, + 333, + 482, + 477 + ], + "type": "text", + "content": " identified in hierarchical clustering serve as pseudo-labels for training our instance segmentation model, Hi-Mask3D (Sec 3.3)." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 132, + 492, + 357, + 504 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 492, + 357, + 504 + ], + "spans": [ + { + "bbox": [ + 132, + 492, + 357, + 504 + ], + "type": "text", + "content": "3.2 3D Objectness Priors From 2D Frames" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 130, + 510, + 481, + 545 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 510, + 481, + 545 + ], + "spans": [ + { + "bbox": [ + 130, + 510, + 481, + 545 + ], + "type": "text", + "content": "In this section, we introduce our grouping-first-then-projection pipeline to extract 3D objectness priors " + }, + { + "bbox": [ + 130, + 510, + 481, + 545 + ], + "type": "inline_equation", + "content": "B^{3\\mathrm{D}}" + }, + { + "bbox": [ + 130, + 510, + 481, + 545 + ], + "type": "text", + "content": " from 2D RGB frames [12], where " + }, + { + "bbox": [ + 130, + 510, + 481, + 545 + ], + "type": "inline_equation", + "content": "B^{3\\mathrm{D}}" + }, + { + "bbox": [ + 130, + 510, + 481, + 545 + ], + "type": "text", + "content": " is the set of 3D bounding boxes for potential objects in the scene." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 130, + 546, + 482, + 666 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 546, + 482, + 666 + ], + "spans": [ + { + "bbox": [ + 130, + 546, + 482, + 666 + ], + "type": "text", + "content": "Grouping-first-then-projection Pipeline. To extract 3D objectness priors, a simple strategy involves projection-first-then-grouping: projecting 2D unsupervised instance segmentation masks [46, 60] from 2D frames onto the 3D point cloud to acquire 3D masks, then grouping and merging these 3D masks based on their spatial relationships to generate 3D objects, as shown in Fig 3(c). However, due to variations in object sizes and corresponding camera poses of 2D frames across different scenes, it is nearly impossible to establish a unified grouping criterion for different objects. If the grouping criterion is too strict, allowing only masks with high overlap to be grouped, masks from multiple parts of a large object cannot be merged (see Fig 3c-\"strict\"). Conversely, if the criterion" + } + ] + } + ], + "index": 6 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "spans": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "type": "text", + "content": "8" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 166, + 91, + 217, + 102 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 166, + 91, + 217, + 102 + ], + "spans": [ + { + "bbox": [ + 166, + 91, + 217, + 102 + ], + "type": "text", + "content": "Cheng et al." + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 7 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 134, + 178, + 480, + 304 + ], + "blocks": [ + { + "bbox": [ + 131, + 114, + 482, + 173 + ], + "lines": [ + { + "bbox": [ + 131, + 114, + 482, + 173 + ], + "spans": [ + { + "bbox": [ + 131, + 114, + 482, + 173 + ], + "type": "text", + "content": "Fig. 3: Robustness of our cluster features and 3D objectness priors. (a) Visualization of the first 3 PCA components and our computed weights of points in Equ 2. (b) Visualization of 2D DINO features' PCA components, 3D points' normal vectors, and our 3D object priors. The comparison between (c) the projection-first-then-grouping pipeline and (d) our grouping-first-then-projection pipeline." + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 134, + 178, + 480, + 304 + ], + "lines": [ + { + "bbox": [ + 134, + 178, + 480, + 304 + ], + "spans": [ + { + "bbox": [ + 134, + 178, + 480, + 304 + ], + "type": "image", + "image_path": "2cfae4e55033010216d29745460a386c2922614fd575d5d2edc95b255a06f0d8.jpg" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_body" + } + ], + "index": 3 + }, + { + "bbox": [ + 130, + 342, + 482, + 450 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 342, + 482, + 450 + ], + "spans": [ + { + "bbox": [ + 130, + 342, + 482, + 450 + ], + "type": "text", + "content": "is relaxed, masks that should belong to multiple objects may be merged into a single object, e.g., the \"chair\" and \"sofa\" in Fig 3c-\"relaxed\". We attribute this limitation to the inability to rely solely on 3D spatial relationships and 3D point features (see Fig 3b) to determine if multiple 3D masks belong to the same object. To tackle this issue, we propose a grouping-first-then-projection pipeline: 1) it co-segments objects across multiple 2D frames by exploiting the temporal consistency of the frames, naturally identifying and grouping multiple 2D masks for each single object. 2) It projects the 2D masks corresponding to each object onto the 3D point cloud and then directly merges them." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 130, + 453, + 482, + 610 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 453, + 482, + 610 + ], + "spans": [ + { + "bbox": [ + 130, + 453, + 482, + 610 + ], + "type": "text", + "content": "Object Co-segmentation from Consecutive 2D RGB Frames. Given that 2D RGB frames are captured consecutively, we leverage their temporal consistency to co-segment multi-frame 2D masks corresponding to the same objects. First, for each RGB frame " + }, + { + "bbox": [ + 130, + 453, + 482, + 610 + ], + "type": "inline_equation", + "content": "I_{m}" + }, + { + "bbox": [ + 130, + 453, + 482, + 610 + ], + "type": "text", + "content": ", we apply the 2D unsupervised instance segmentation method MaskCut [60] to obtain the 2D masks of objects in it, denoted as " + }, + { + "bbox": [ + 130, + 453, + 482, + 610 + ], + "type": "inline_equation", + "content": "O_{m}" + }, + { + "bbox": [ + 130, + 453, + 482, + 610 + ], + "type": "text", + "content": ". As MaskCut uses the image encoder of 2D self-supervised DINO [6] to extract image feature map, we employ the same encoder to extract features for objects " + }, + { + "bbox": [ + 130, + 453, + 482, + 610 + ], + "type": "inline_equation", + "content": "O_{m}" + }, + { + "bbox": [ + 130, + 453, + 482, + 610 + ], + "type": "text", + "content": " by masked average pooling each object mask " + }, + { + "bbox": [ + 130, + 453, + 482, + 610 + ], + "type": "inline_equation", + "content": "o_{m}^{i} \\in O_{m}" + }, + { + "bbox": [ + 130, + 453, + 482, + 610 + ], + "type": "text", + "content": " over the frame's feature map. Next, for any adjacent pair of frames " + }, + { + "bbox": [ + 130, + 453, + 482, + 610 + ], + "type": "inline_equation", + "content": "I_{m}" + }, + { + "bbox": [ + 130, + 453, + 482, + 610 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 130, + 453, + 482, + 610 + ], + "type": "inline_equation", + "content": "I_{m+1}" + }, + { + "bbox": [ + 130, + 453, + 482, + 610 + ], + "type": "text", + "content": ", we compute the pairwise similarity between their object masks " + }, + { + "bbox": [ + 130, + 453, + 482, + 610 + ], + "type": "inline_equation", + "content": "O_{m}" + }, + { + "bbox": [ + 130, + 453, + 482, + 610 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 130, + 453, + 482, + 610 + ], + "type": "inline_equation", + "content": "O_{m+1}" + }, + { + "bbox": [ + 130, + 453, + 482, + 610 + ], + "type": "text", + "content": " based on mask features. Then, for each object mask " + }, + { + "bbox": [ + 130, + 453, + 482, + 610 + ], + "type": "inline_equation", + "content": "o_{m}^{i} \\in O_{m}" + }, + { + "bbox": [ + 130, + 453, + 482, + 610 + ], + "type": "text", + "content": " in frame " + }, + { + "bbox": [ + 130, + 453, + 482, + 610 + ], + "type": "inline_equation", + "content": "I_{m}" + }, + { + "bbox": [ + 130, + 453, + 482, + 610 + ], + "type": "text", + "content": ", we find the object mask " + }, + { + "bbox": [ + 130, + 453, + 482, + 610 + ], + "type": "inline_equation", + "content": "o_{m+1}^{j^{*}}" + }, + { + "bbox": [ + 130, + 453, + 482, + 610 + ], + "type": "text", + "content": " in frame " + }, + { + "bbox": [ + 130, + 453, + 482, + 610 + ], + "type": "inline_equation", + "content": "I_{m+1}" + }, + { + "bbox": [ + 130, + 453, + 482, + 610 + ], + "type": "text", + "content": " with the highest similarity as its corresponding candidate mask as follows," + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 190, + 624, + 481, + 669 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 190, + 624, + 481, + 669 + ], + "spans": [ + { + "bbox": [ + 190, + 624, + 481, + 669 + ], + "type": "interline_equation", + "content": "\\begin{array}{l} o _ {m + 1} ^ {j ^ {*}} = \\operatorname * {a r g m a x} _ {i} \\operatorname * {s i m} (o _ {m + 1} ^ {j}, o _ {m} ^ {i}), \\\\ o _ {m + 1} ^ {\\jmath} \\in O _ {m + 1} \\tag {4} \\\\ g _ {m \\rightarrow m + 1} ^ {i} = o _ {m + 1} ^ {j ^ {*}} \\mathrm {i f} \\mathrm {s i m} (o _ {m + 1} ^ {j ^ {*}}, o _ {m} ^ {i}) > \\tau \\mathrm {\\bf e l s e N u l l}, \\\\ \\end{array}", + "image_path": "76d29ae47ea8ea604bc82158f653e9ab19762fab17b848023fedf108bbd00cf0.jpg" + } + ] + } + ], + "index": 6 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 395, + 91, + 447, + 102 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 395, + 91, + 447, + 102 + ], + "spans": [ + { + "bbox": [ + 395, + 91, + 447, + 102 + ], + "type": "text", + "content": "Part2Object" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 474, + 91, + 481, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 474, + 91, + 481, + 100 + ], + "spans": [ + { + "bbox": [ + 474, + 91, + 481, + 100 + ], + "type": "text", + "content": "9" + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 8 + }, + { + "para_blocks": [ + { + "bbox": [ + 130, + 115, + 482, + 216 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 115, + 482, + 216 + ], + "spans": [ + { + "bbox": [ + 130, + 115, + 482, + 216 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 130, + 115, + 482, + 216 + ], + "type": "inline_equation", + "content": "\\mathrm{sim}(o_{m + 1}^j,o_m^i)" + }, + { + "bbox": [ + 130, + 115, + 482, + 216 + ], + "type": "text", + "content": " is to compute the cosine similarity between mask features of object masks " + }, + { + "bbox": [ + 130, + 115, + 482, + 216 + ], + "type": "inline_equation", + "content": "o_{m + 1}^{j}" + }, + { + "bbox": [ + 130, + 115, + 482, + 216 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 130, + 115, + 482, + 216 + ], + "type": "inline_equation", + "content": "o_{m}^{i}" + }, + { + "bbox": [ + 130, + 115, + 482, + 216 + ], + "type": "text", + "content": ". The " + }, + { + "bbox": [ + 130, + 115, + 482, + 216 + ], + "type": "inline_equation", + "content": "o_{m + 1}^{j^{*}}" + }, + { + "bbox": [ + 130, + 115, + 482, + 216 + ], + "type": "text", + "content": " in frame " + }, + { + "bbox": [ + 130, + 115, + 482, + 216 + ], + "type": "inline_equation", + "content": "I_{m + 1}" + }, + { + "bbox": [ + 130, + 115, + 482, + 216 + ], + "type": "text", + "content": " is the candidate object mask of the object mask " + }, + { + "bbox": [ + 130, + 115, + 482, + 216 + ], + "type": "inline_equation", + "content": "o_{m}^{i}" + }, + { + "bbox": [ + 130, + 115, + 482, + 216 + ], + "type": "text", + "content": " in frame " + }, + { + "bbox": [ + 130, + 115, + 482, + 216 + ], + "type": "inline_equation", + "content": "I_{m}" + }, + { + "bbox": [ + 130, + 115, + 482, + 216 + ], + "type": "text", + "content": ". Notably, only when the similarity between the object mask " + }, + { + "bbox": [ + 130, + 115, + 482, + 216 + ], + "type": "inline_equation", + "content": "o_{m}^{i}" + }, + { + "bbox": [ + 130, + 115, + 482, + 216 + ], + "type": "text", + "content": " and its candidate object mask " + }, + { + "bbox": [ + 130, + 115, + 482, + 216 + ], + "type": "inline_equation", + "content": "o_{m + 1}^{j^{*}}" + }, + { + "bbox": [ + 130, + 115, + 482, + 216 + ], + "type": "text", + "content": " exceeds the threshold " + }, + { + "bbox": [ + 130, + 115, + 482, + 216 + ], + "type": "inline_equation", + "content": "\\pi" + }, + { + "bbox": [ + 130, + 115, + 482, + 216 + ], + "type": "text", + "content": ", they are considered to belong to the same object. After determining the same objects between any adjacent frames, " + }, + { + "bbox": [ + 130, + 115, + 482, + 216 + ], + "type": "inline_equation", + "content": "g_{m\\rightarrow m + 1}^{i}" + }, + { + "bbox": [ + 130, + 115, + 482, + 216 + ], + "type": "text", + "content": ", we propagate this \"sameness\" to all adjacent frames to identify the same objects across all frames. Consequently, we can gather objects and identify all 2D masks in different frames of each object." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 130, + 217, + 482, + 312 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 217, + 482, + 312 + ], + "spans": [ + { + "bbox": [ + 130, + 217, + 482, + 312 + ], + "type": "text", + "content": "3D Objectness Priors. Given that we have grouped 2D masks into different objects, for each object, we can directly project its 2D masks onto the 3D point cloud and seamlessly merge their 3D masks into a unified whole [64], forming the object's 3D mask. Furthermore, instead of directly utilizing the 3D masks of objects, we utilize the 3D bounding boxes " + }, + { + "bbox": [ + 130, + 217, + 482, + 312 + ], + "type": "inline_equation", + "content": "B^{3\\mathrm{D}}" + }, + { + "bbox": [ + 130, + 217, + 482, + 312 + ], + "type": "text", + "content": " of their masks to guide hierarchical clustering. This decision is based on treating the collected 3D objects from 2D frames as objectness priors and leaving precise segmentation to the clustering process itself." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 132, + 332, + 340, + 345 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 332, + 340, + 345 + ], + "spans": [ + { + "bbox": [ + 132, + 332, + 340, + 345 + ], + "type": "text", + "content": "3.3 Hi-Mask3D and Learning Objective" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 130, + 355, + 485, + 391 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 355, + 485, + 391 + ], + "spans": [ + { + "bbox": [ + 130, + 355, + 485, + 391 + ], + "type": "text", + "content": "In this section, we introduce our 3D instance segmentation framework, HiMask3D. It extends and adapts the 3D instance segmentation model, Mask3D [42], to predict and utilize object parts to improve 3D instance segmentation." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 130, + 392, + 482, + 511 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 392, + 482, + 511 + ], + "spans": [ + { + "bbox": [ + 130, + 392, + 482, + 511 + ], + "type": "text", + "content": "Introduction to Mask3D. The 3D instance segmentation model Mask3D [42] mainly consists of three components: 1) a feature backbone, a sparse convolutional U-net [11] that takes a colored point cloud as input and yields multi-resolution feature matrices " + }, + { + "bbox": [ + 130, + 392, + 482, + 511 + ], + "type": "inline_equation", + "content": "\\{\\pmb{F}_r\\}" + }, + { + "bbox": [ + 130, + 392, + 482, + 511 + ], + "type": "text", + "content": ", where each " + }, + { + "bbox": [ + 130, + 392, + 482, + 511 + ], + "type": "inline_equation", + "content": "\\pmb{F}_r \\in \\mathbb{R}^{M_r \\times C}" + }, + { + "bbox": [ + 130, + 392, + 482, + 511 + ], + "type": "text", + "content": " of size " + }, + { + "bbox": [ + 130, + 392, + 482, + 511 + ], + "type": "inline_equation", + "content": "M_r" + }, + { + "bbox": [ + 130, + 392, + 482, + 511 + ], + "type": "text", + "content": ". 2) A mask module predicts instance masks " + }, + { + "bbox": [ + 130, + 392, + 482, + 511 + ], + "type": "inline_equation", + "content": "B" + }, + { + "bbox": [ + 130, + 392, + 482, + 511 + ], + "type": "text", + "content": " corresponding to instance queries " + }, + { + "bbox": [ + 130, + 392, + 482, + 511 + ], + "type": "inline_equation", + "content": "Q \\in \\mathbb{R}^{K \\times C}" + }, + { + "bbox": [ + 130, + 392, + 482, + 511 + ], + "type": "text", + "content": " by computing the dot product between queries " + }, + { + "bbox": [ + 130, + 392, + 482, + 511 + ], + "type": "inline_equation", + "content": "Q" + }, + { + "bbox": [ + 130, + 392, + 482, + 511 + ], + "type": "text", + "content": " and scene features " + }, + { + "bbox": [ + 130, + 392, + 482, + 511 + ], + "type": "inline_equation", + "content": "\\{\\pmb{F}_r\\}" + }, + { + "bbox": [ + 130, + 392, + 482, + 511 + ], + "type": "text", + "content": ". 3) A query refinement module, a transformer decoder [54] that cross-attends " + }, + { + "bbox": [ + 130, + 392, + 482, + 511 + ], + "type": "inline_equation", + "content": "K" + }, + { + "bbox": [ + 130, + 392, + 482, + 511 + ], + "type": "text", + "content": " instance queries " + }, + { + "bbox": [ + 130, + 392, + 482, + 511 + ], + "type": "inline_equation", + "content": "Q" + }, + { + "bbox": [ + 130, + 392, + 482, + 511 + ], + "type": "text", + "content": " with multi-resolution feature matrices " + }, + { + "bbox": [ + 130, + 392, + 482, + 511 + ], + "type": "inline_equation", + "content": "\\{\\pmb{F}_r\\}" + }, + { + "bbox": [ + 130, + 392, + 482, + 511 + ], + "type": "text", + "content": ". In the " + }, + { + "bbox": [ + 130, + 392, + 482, + 511 + ], + "type": "inline_equation", + "content": "r" + }, + { + "bbox": [ + 130, + 392, + 482, + 511 + ], + "type": "text", + "content": "-th decoder layer of query refinement module, the cross-attention between instance queries " + }, + { + "bbox": [ + 130, + 392, + 482, + 511 + ], + "type": "inline_equation", + "content": "Q_r" + }, + { + "bbox": [ + 130, + 392, + 482, + 511 + ], + "type": "text", + "content": " and feature matrix " + }, + { + "bbox": [ + 130, + 392, + 482, + 511 + ], + "type": "inline_equation", + "content": "\\pmb{F}_r" + }, + { + "bbox": [ + 130, + 392, + 482, + 511 + ], + "type": "text", + "content": " is formulated as follows," + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 230, + 521, + 481, + 536 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 230, + 521, + 481, + 536 + ], + "spans": [ + { + "bbox": [ + 230, + 521, + 481, + 536 + ], + "type": "interline_equation", + "content": "Q _ {r} = \\operatorname {s o f t m a x} \\left(Q _ {r - 1} \\boldsymbol {F} _ {r} ^ {T} / \\sqrt {C}\\right) \\boldsymbol {F} _ {r}. \\tag {5}", + "image_path": "6e6c42111288505b7c3a7c51083609a074b6843e8a0ea2db51a8f7901a03a44f.jpg" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 130, + 544, + 481, + 569 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 544, + 481, + 569 + ], + "spans": [ + { + "bbox": [ + 130, + 544, + 481, + 569 + ], + "type": "text", + "content": "Note that for simplicity of demonstration, we omit the QKV projection [54] and intermediate instance mask " + }, + { + "bbox": [ + 130, + 544, + 481, + 569 + ], + "type": "inline_equation", + "content": "B" + }, + { + "bbox": [ + 130, + 544, + 481, + 569 + ], + "type": "text", + "content": " [42] in the cross-attention formulation." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 130, + 570, + 482, + 666 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 570, + 482, + 666 + ], + "spans": [ + { + "bbox": [ + 130, + 570, + 482, + 666 + ], + "type": "text", + "content": "Hi-Mask3D with Hierarchical-Aware Detector. Our Hi-Mask3D modifies the query refinement module of Mask3D to adapt for both object and object part segmentation. Specifically, we start with two sets of queries " + }, + { + "bbox": [ + 130, + 570, + 482, + 666 + ], + "type": "inline_equation", + "content": "Q^o" + }, + { + "bbox": [ + 130, + 570, + 482, + 666 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 130, + 570, + 482, + 666 + ], + "type": "inline_equation", + "content": "Q^p" + }, + { + "bbox": [ + 130, + 570, + 482, + 666 + ], + "type": "text", + "content": " representing object queries and object part queries, respectively. Next, in each decoder layer, we no longer directly perform cross-attention between object queries and feature matrix to update object queries. Instead, we perform hierarchical attention: 1) conducting cross-attention between object part queries and feature matrix, and 2) updating object queries using part queries. This enables explicit" + } + ] + } + ], + "index": 9 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "spans": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "type": "text", + "content": "10" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 166, + 91, + 217, + 102 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 166, + 91, + 217, + 102 + ], + "spans": [ + { + "bbox": [ + 166, + 91, + 217, + 102 + ], + "type": "text", + "content": "Cheng et al." + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 9 + }, + { + "para_blocks": [ + { + "bbox": [ + 130, + 116, + 479, + 152 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 116, + 479, + 152 + ], + "spans": [ + { + "bbox": [ + 130, + 116, + 479, + 152 + ], + "type": "text", + "content": "interaction between object part queries and object queries, which helps to leverage the information from object parts to improve object segmentation. In the " + }, + { + "bbox": [ + 130, + 116, + 479, + 152 + ], + "type": "inline_equation", + "content": "r" + }, + { + "bbox": [ + 130, + 116, + 479, + 152 + ], + "type": "text", + "content": "-th decoder layer, the hierarchical attention is formulated as follows:" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 231, + 162, + 480, + 182 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 231, + 162, + 480, + 182 + ], + "spans": [ + { + "bbox": [ + 231, + 162, + 480, + 182 + ], + "type": "interline_equation", + "content": "Q _ {r} ^ {p} = \\operatorname {s o f t m a x} \\left(Q _ {r - 1} ^ {p} \\boldsymbol {F} _ {r} ^ {T} / \\sqrt {C}\\right) \\boldsymbol {F} _ {r}, \\tag {6}", + "image_path": "ed19eebee5e3e89a357ef0f13c0ad3cb5ea1add6da30c74531d185da47de5ffa.jpg" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 230, + 178, + 383, + 193 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 230, + 178, + 383, + 193 + ], + "spans": [ + { + "bbox": [ + 230, + 178, + 383, + 193 + ], + "type": "interline_equation", + "content": "Q _ {r} ^ {o} = \\operatorname {s o f t m a x} \\left(Q _ {r - 1} ^ {o} \\left(Q _ {r} ^ {p}\\right) ^ {T} / \\sqrt {C}\\right) Q _ {r} ^ {p}.", + "image_path": "834a1c5bdde3777ff3e91b3ae245085f5d68f5e77ec8b6b80d3d0ab780f7dab4.jpg" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 130, + 201, + 479, + 224 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 201, + 479, + 224 + ], + "spans": [ + { + "bbox": [ + 130, + 201, + 479, + 224 + ], + "type": "text", + "content": "The last-layer part queries and object queries undergo the mask module to obtain object part and part segmentation results, " + }, + { + "bbox": [ + 130, + 201, + 479, + 224 + ], + "type": "inline_equation", + "content": "P_{\\mathrm{part}}" + }, + { + "bbox": [ + 130, + 201, + 479, + 224 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 130, + 201, + 479, + 224 + ], + "type": "inline_equation", + "content": "P_{\\mathrm{object}}" + }, + { + "bbox": [ + 130, + 201, + 479, + 224 + ], + "type": "text", + "content": ", respectively." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 130, + 225, + 480, + 310 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 225, + 480, + 310 + ], + "spans": [ + { + "bbox": [ + 130, + 225, + 480, + 310 + ], + "type": "text", + "content": "Training Objective. Following Mask3D [42], we adopt bipartite graph algorithm [5, 48] to match prediction " + }, + { + "bbox": [ + 130, + 225, + 480, + 310 + ], + "type": "inline_equation", + "content": "P_{\\mathrm{object}}" + }, + { + "bbox": [ + 130, + 225, + 480, + 310 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 130, + 225, + 480, + 310 + ], + "type": "inline_equation", + "content": "P_{\\mathrm{part}}" + }, + { + "bbox": [ + 130, + 225, + 480, + 310 + ], + "type": "text", + "content": " with pseudo-labels " + }, + { + "bbox": [ + 130, + 225, + 480, + 310 + ], + "type": "inline_equation", + "content": "\\hat{P}_{\\mathrm{object}}" + }, + { + "bbox": [ + 130, + 225, + 480, + 310 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 130, + 225, + 480, + 310 + ], + "type": "inline_equation", + "content": "\\hat{P}_{\\mathrm{part}}" + }, + { + "bbox": [ + 130, + 225, + 480, + 310 + ], + "type": "text", + "content": ", respectively. To supervise segmentation masks, we combine DICE [14] loss and Binary Cross Entropy (BCE) loss [29]. Similar to [41], we also employ iterative rounds of self-training to enhance the segmentation capabilities of HiMask3D. During self-training, predictions from the previous round are utilized as pseudo-labels for the subsequent round." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 132, + 327, + 224, + 341 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 327, + 224, + 341 + ], + "spans": [ + { + "bbox": [ + 132, + 327, + 224, + 341 + ], + "type": "text", + "content": "4 Experiment" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 130, + 353, + 491, + 485 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 353, + 491, + 485 + ], + "spans": [ + { + "bbox": [ + 130, + 353, + 491, + 485 + ], + "type": "text", + "content": "Datasets and Implementations. We evaluate our method on four public indoor point cloud datasets, including ScanNet [12], Scannet200 [40], S3DIS [4] and Replica [49]. Following Mask3D [42], we employ a Sparse CNN, Res16UNet34C [11], implemented in MinkowskiEngine [11], as the point cloud encoder. Our HiMask3D is trained for 600 epochs with a learning rate 1e-4 and a batch size 4. Experiments are conducted using PyTorch [35] and executed on NVIDIA Tesla A40 GPUs. We evaluate the performance of class-agnostic instance segmentation using standard average precision scores. The scores consist of mAP@25, mAP@50, and mAP, representing performance metrics at IoU thresholds of 0.25, 0.5, and mean average precision, respectively. Hyper-parameter " + }, + { + "bbox": [ + 130, + 353, + 491, + 485 + ], + "type": "inline_equation", + "content": "T = 0.05" + }, + { + "bbox": [ + 130, + 353, + 491, + 485 + ], + "type": "text", + "content": " in Equ 3 and " + }, + { + "bbox": [ + 130, + 353, + 491, + 485 + ], + "type": "inline_equation", + "content": "\\tau = 0.3" + }, + { + "bbox": [ + 130, + 353, + 491, + 485 + ], + "type": "text", + "content": " in Equ 4." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 130, + 502, + 451, + 514 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 502, + 451, + 514 + ], + "spans": [ + { + "bbox": [ + 130, + 502, + 451, + 514 + ], + "type": "text", + "content": "4.1 Comparison with State-of-the-Art Unsupervised Methods" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 130, + 521, + 479, + 641 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 521, + 479, + 641 + ], + "spans": [ + { + "bbox": [ + 130, + 521, + 479, + 641 + ], + "type": "text", + "content": "Table 1 shows the results on ScanNet dataset [12] under two different settings: Traning-free Setting. In the training-free setting, all methods rely solely on clustering [6] or graph-cut [46, 60] algorithms for prediction, as no annotations are available for use. Table 1a compares our method, Part2Object, with the previous state-of-the-art method, Unscene3D [41], which also leverages DINO features but applies graph-cut algorithms on 3D point clouds. Compared to Unscene3D, our approach demonstrates significant enhancements across all metrics, with increases of " + }, + { + "bbox": [ + 130, + 521, + 479, + 641 + ], + "type": "inline_equation", + "content": "35.2\\%" + }, + { + "bbox": [ + 130, + 521, + 479, + 641 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 130, + 521, + 479, + 641 + ], + "type": "inline_equation", + "content": "16.8\\%" + }, + { + "bbox": [ + 130, + 521, + 479, + 641 + ], + "type": "text", + "content": ", and " + }, + { + "bbox": [ + 130, + 521, + 479, + 641 + ], + "type": "inline_equation", + "content": "6.7\\%" + }, + { + "bbox": [ + 130, + 521, + 479, + 641 + ], + "type": "text", + "content": " in mAP@25, mAP@50, and mAP, respectively. These improvements are attributed to our utilization of 3D objectness priors, which guide the clustering process." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 130, + 641, + 479, + 665 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 641, + 479, + 665 + ], + "spans": [ + { + "bbox": [ + 130, + 641, + 479, + 665 + ], + "type": "text", + "content": "Data-efficient Fine-tuning Setting. In the data-efficient fine-tuning setting, all methods initially undergo either pretraining or self-training. Subsequently," + } + ] + } + ], + "index": 11 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 395, + 91, + 447, + 102 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 395, + 91, + 447, + 102 + ], + "spans": [ + { + "bbox": [ + 395, + 91, + 447, + 102 + ], + "type": "text", + "content": "Part2Object" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 470, + 91, + 479, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 470, + 91, + 479, + 100 + ], + "spans": [ + { + "bbox": [ + 470, + 91, + 479, + 100 + ], + "type": "text", + "content": "11" + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 10 + }, + { + "para_blocks": [ + { + "type": "table", + "bbox": [ + 132, + 167, + 485, + 281 + ], + "blocks": [ + { + "bbox": [ + 141, + 114, + 470, + 125 + ], + "lines": [ + { + "bbox": [ + 141, + 114, + 470, + 125 + ], + "spans": [ + { + "bbox": [ + 141, + 114, + 470, + 125 + ], + "type": "text", + "content": "Table 1: Comparison of instance segmentation results on ScanNet val." + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 132, + 130, + 302, + 156 + ], + "lines": [ + { + "bbox": [ + 132, + 130, + 302, + 156 + ], + "spans": [ + { + "bbox": [ + 132, + 130, + 302, + 156 + ], + "type": "text", + "content": "(a) Training-free Setting. All methods do not require gradient backpropagation to obtain prediction results." + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 309, + 130, + 478, + 163 + ], + "lines": [ + { + "bbox": [ + 309, + 130, + 478, + 163 + ], + "spans": [ + { + "bbox": [ + 309, + 130, + 478, + 163 + ], + "type": "text", + "content": "(b) Data-efficient Setting with " + }, + { + "bbox": [ + 309, + 130, + 478, + 163 + ], + "type": "inline_equation", + "content": "\\mathbf{x}\\%" + }, + { + "bbox": [ + 309, + 130, + 478, + 163 + ], + "type": "text", + "content": " available data annotations. A " + }, + { + "bbox": [ + 309, + 130, + 478, + 163 + ], + "type": "inline_equation", + "content": "0\\%" + }, + { + "bbox": [ + 309, + 130, + 478, + 163 + ], + "type": "text", + "content": " value indicates that the method can provide pseudo-labels, and training is conducted on them." + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 132, + 167, + 485, + 281 + ], + "lines": [ + { + "bbox": [ + 132, + 167, + 485, + 281 + ], + "spans": [ + { + "bbox": [ + 132, + 167, + 485, + 281 + ], + "type": "table", + "html": "
w/o training MethodsScanNet Val [12]w/ training MethodsmAP@50 ScanNet Val [12]
mAP@25mAP@50mAP0%1%5%10%20%
HDBSCAN [32]32.15.51.6Scratch [42]/14.133.339.243.4
Nunes [33]30.57.32.3CSC [21]/22.139.943.848.9
Felzenswalb [16]38.912.75.0HDBSCAN [32]10.515.136.340.042.7
CutLER [60]7.00.20.3Felzenswalb [16]15.225.337.245.750.0
Unscene3D [41]19.910.05.9Unscene3D [41]23.228.446.855.760.7
Ours55.126.812.6Ours32.644.164.268.072.1
", + "image_path": "7a9ba7435b61df920ca62bd392a4ce3d518c3037d64eb9191d3db1aab9c19b13.jpg" + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "table_body" + } + ], + "index": 5 + }, + { + "bbox": [ + 130, + 290, + 482, + 387 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 290, + 482, + 387 + ], + "spans": [ + { + "bbox": [ + 130, + 290, + 482, + 387 + ], + "type": "text", + "content": "they fine-tune on downstream data with varying percentages (0% represents pseudo-labels solely). Under this setting, the extraction of pseudo-labels is the primary requirement, followed by the necessity for models to exhibit strong learning capabilities. As demonstrated in Table 1b, Hi-Mask3D exhibits remarkable performance even with limited data, showcasing the robust 3D representation capabilities learned on the pseudo-labels from Part2object. Specifically, our method surpasses the previous SOTA by " + }, + { + "bbox": [ + 130, + 290, + 482, + 387 + ], + "type": "inline_equation", + "content": "15.7\\%" + }, + { + "bbox": [ + 130, + 290, + 482, + 387 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 130, + 290, + 482, + 387 + ], + "type": "inline_equation", + "content": "11.4\\%" + }, + { + "bbox": [ + 130, + 290, + 482, + 387 + ], + "type": "text", + "content": " when considering the utilization of only " + }, + { + "bbox": [ + 130, + 290, + 482, + 387 + ], + "type": "inline_equation", + "content": "1\\%" + }, + { + "bbox": [ + 130, + 290, + 482, + 387 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 130, + 290, + 482, + 387 + ], + "type": "inline_equation", + "content": "20\\%" + }, + { + "bbox": [ + 130, + 290, + 482, + 387 + ], + "type": "text", + "content": " of the available data, respectively." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 132, + 404, + 386, + 417 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 404, + 386, + 417 + ], + "spans": [ + { + "bbox": [ + 132, + 404, + 386, + 417 + ], + "type": "text", + "content": "4.2 Comparison on Cross-dataset Generalization" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 130, + 425, + 482, + 485 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 425, + 482, + 485 + ], + "spans": [ + { + "bbox": [ + 130, + 425, + 482, + 485 + ], + "type": "text", + "content": "Table 2 and Table 3 show our comparisons with fully-supervised Mask3D [42] on cross-dataset generalization, including both zero-shot and dataset-efficient settings. In both settings, our Hi-Mask3D is trained on pseudo-labels from our unsupervised Part2Object on the ScanNet dataset [12], while Mask3D is trained fully supervised on ScanNet." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 130, + 486, + 482, + 628 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 486, + 482, + 628 + ], + "spans": [ + { + "bbox": [ + 130, + 486, + 482, + 628 + ], + "type": "text", + "content": "Cross-dataset Zero-shot Generalization Setting. In Table 2, we compare unsupervised Hi-Mask3D with fully-supervised Mask3D on three downstream datasets, including ScanNet200 [39], S3DIS [4] and Replica [49]. Compared with the fully supervised Mask3D trained on ScanNet, our Hi-Mask3D trained on pseudo-labels from Part2Object achieves consistent performance improvement without using any manually annotated data. Hi-Mask3D surpasses fully-supervised Mask3D by " + }, + { + "bbox": [ + 130, + 486, + 482, + 628 + ], + "type": "inline_equation", + "content": "10.7\\%" + }, + { + "bbox": [ + 130, + 486, + 482, + 628 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 130, + 486, + 482, + 628 + ], + "type": "inline_equation", + "content": "4.8\\%" + }, + { + "bbox": [ + 130, + 486, + 482, + 628 + ], + "type": "text", + "content": ", and " + }, + { + "bbox": [ + 130, + 486, + 482, + 628 + ], + "type": "inline_equation", + "content": "6.4\\%" + }, + { + "bbox": [ + 130, + 486, + 482, + 628 + ], + "type": "text", + "content": " in terms of mAP@50 on ScanNet200, S3DIS, and Replica datasets, respectively. Thanks to the learning from pseudo-labels in Part2Object, Hi-Mask3D has acquired more generalizable representations of 3D objects. Unlike fully-supervised methods, which are constrained to annotated classes in the training dataset, Hi-Mask3D demonstrates robust performance in cross-dataset generalization." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 130, + 629, + 482, + 666 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 629, + 482, + 666 + ], + "spans": [ + { + "bbox": [ + 130, + 629, + 482, + 666 + ], + "type": "text", + "content": "Cross-dataset Data-efficient Generalization Setting. In Table 3, we conduct data-efficient training on downstream datasets, such as ScanNet200 [39] and S3DIS [4], after pre-training our Hi-Mask3D on ScanNet's pseudo-labels ex" + } + ] + } + ], + "index": 10 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "spans": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "type": "text", + "content": "12" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 166, + 91, + 217, + 102 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 166, + 91, + 217, + 102 + ], + "spans": [ + { + "bbox": [ + 166, + 91, + 217, + 102 + ], + "type": "text", + "content": "Cheng et al." + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 11 + }, + { + "para_blocks": [ + { + "type": "table", + "bbox": [ + 154, + 157, + 460, + 224 + ], + "blocks": [ + { + "bbox": [ + 130, + 114, + 482, + 148 + ], + "lines": [ + { + "bbox": [ + 130, + 114, + 482, + 148 + ], + "spans": [ + { + "bbox": [ + 130, + 114, + 482, + 148 + ], + "type": "text", + "content": "Table 2: Comparison on cross-dataset zero-shot generalization setting. We compare the zero-shot generalization ability of unsupervised Hi-Mask3D with fully-supervised Mask3D on three downstream datasets: ScanNet200, S3DIS and Replica." + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 154, + 157, + 460, + 224 + ], + "lines": [ + { + "bbox": [ + 154, + 157, + 460, + 224 + ], + "spans": [ + { + "bbox": [ + 154, + 157, + 460, + 224 + ], + "type": "table", + "html": "
Zero-shot MethodsScanNet200 Val [39]S3DIS 6-fold [4]Replica [49]
m@25m@50mAPm@25m@50mAPm@25m@50mAP
Mask3D [42]30.824.215.417.611.77.720.815.69.7
Ours63.234.916.324.516.58.536.522.011.2
", + "image_path": "e32b36ef9531d7bcea8e4301ed1e2ae8fb45543f616de5d099b8f269c36b0fdb.jpg" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "table_body" + } + ], + "index": 3 + }, + { + "type": "table", + "bbox": [ + 154, + 276, + 451, + 370 + ], + "blocks": [ + { + "bbox": [ + 130, + 232, + 481, + 266 + ], + "lines": [ + { + "bbox": [ + 130, + 232, + 481, + 266 + ], + "spans": [ + { + "bbox": [ + 130, + 232, + 481, + 266 + ], + "type": "text", + "content": "Table 3: Comparison on cross-dataset data-efficient generalization setting. We utilize pseudo-labels from Part2Object to pre-train our Hi-Mask3D and conduct data-efficient finetuning on downstream datasets, comparing them with Mask3D." + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 154, + 276, + 451, + 370 + ], + "lines": [ + { + "bbox": [ + 154, + 276, + 451, + 370 + ], + "spans": [ + { + "bbox": [ + 154, + 276, + 451, + 370 + ], + "type": "table", + "html": "
MethodDatasetmAP@25mAP@50
1%5%10%20%1%5%10%20%
ScratchScanNet200 [39]2.122.530.640.21.38.218.422.9
OursScanNet200 [39]59.470.472.377.835.452.856.262.8
ScratchS3DIS [4]1.710.120.845.41.02.99.125.9
OursS3DIS [4]49.155.965.170.225.735.046.449.4
", + "image_path": "f3b751c476b6c6d460421b0d3d9531214f8139cc3a606a7b4f764fb7553ce98a.jpg" + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "table_body" + } + ], + "index": 5 + }, + { + "bbox": [ + 130, + 378, + 482, + 437 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 378, + 482, + 437 + ], + "spans": [ + { + "bbox": [ + 130, + 378, + 482, + 437 + ], + "type": "text", + "content": "tracted from Part2Object. For ScanNet200 and S3DIS, Hi-Mask3D outperforms Mask3D " + }, + { + "bbox": [ + 130, + 378, + 482, + 437 + ], + "type": "inline_equation", + "content": "47.4\\%" + }, + { + "bbox": [ + 130, + 378, + 482, + 437 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 130, + 378, + 482, + 437 + ], + "type": "inline_equation", + "content": "24.7\\%" + }, + { + "bbox": [ + 130, + 378, + 482, + 437 + ], + "type": "text", + "content": " mAP@50 with " + }, + { + "bbox": [ + 130, + 378, + 482, + 437 + ], + "type": "inline_equation", + "content": "20\\%" + }, + { + "bbox": [ + 130, + 378, + 482, + 437 + ], + "type": "text", + "content": " data, respectively. When limited data is available, the improvement brought by pretraining on pseudo-labels from Part2Object is even greater, with the enhancement increasing from " + }, + { + "bbox": [ + 130, + 378, + 482, + 437 + ], + "type": "inline_equation", + "content": "37.6\\%" + }, + { + "bbox": [ + 130, + 378, + 482, + 437 + ], + "type": "text", + "content": " to " + }, + { + "bbox": [ + 130, + 378, + 482, + 437 + ], + "type": "inline_equation", + "content": "57.3\\%" + }, + { + "bbox": [ + 130, + 378, + 482, + 437 + ], + "type": "text", + "content": " as the available data decreases from " + }, + { + "bbox": [ + 130, + 378, + 482, + 437 + ], + "type": "inline_equation", + "content": "20\\%" + }, + { + "bbox": [ + 130, + 378, + 482, + 437 + ], + "type": "text", + "content": " to " + }, + { + "bbox": [ + 130, + 378, + 482, + 437 + ], + "type": "inline_equation", + "content": "1\\%" + }, + { + "bbox": [ + 130, + 378, + 482, + 437 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 131, + 454, + 239, + 467 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 131, + 454, + 239, + 467 + ], + "spans": [ + { + "bbox": [ + 131, + 454, + 239, + 467 + ], + "type": "text", + "content": "4.3 Ablation Study" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 130, + 474, + 481, + 521 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 474, + 481, + 521 + ], + "spans": [ + { + "bbox": [ + 130, + 474, + 481, + 521 + ], + "type": "text", + "content": "To evaluate the effectiveness of our hierarchical clustering Part2Object and 3D instance segmentation architecture Hi-Mask3D, as well as the effect of hyperparameters, we conduct comprehensive and detailed ablation experiments, as outlined in Table 4a and Table 4b." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 130, + 521, + 482, + 665 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 521, + 482, + 665 + ], + "spans": [ + { + "bbox": [ + 130, + 521, + 482, + 665 + ], + "type": "text", + "content": "Ablation on Part2Object Clustering. Table 4a illustrates the variants of Part2Object clustering, starting with a baseline utilizing the single-layer clustering algorithm. The performance of this baseline yields only " + }, + { + "bbox": [ + 130, + 521, + 482, + 665 + ], + "type": "inline_equation", + "content": "6.1\\%" + }, + { + "bbox": [ + 130, + 521, + 482, + 665 + ], + "type": "text", + "content": " in mAP and " + }, + { + "bbox": [ + 130, + 521, + 482, + 665 + ], + "type": "inline_equation", + "content": "13.8\\%" + }, + { + "bbox": [ + 130, + 521, + 482, + 665 + ], + "type": "text", + "content": " in mAP@50. This limitation arises due to the single-layer clustering being either too loose or too tight, leading to over-clustered or under-clustered objects. Subsequently, we remove the guidance of objectness priors from our proposed clustering algorithm, denoted as \"w/o OG\". Without object priors, while objects may appear at every layer, it becomes challenging to determine which layer they belong to. Furthermore, we replace our cluster feature computation function, \"FU(\\cdot)\", as simple averaged over features of points within the cluster, \"w/o FU\". The " + }, + { + "bbox": [ + 130, + 521, + 482, + 665 + ], + "type": "inline_equation", + "content": "2.0\\%" + }, + { + "bbox": [ + 130, + 521, + 482, + 665 + ], + "type": "text", + "content": " decline in mAP highlights the importance of removing noise features for more robust cluster feature representations." + } + ] + } + ], + "index": 9 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 395, + 91, + 447, + 102 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 395, + 91, + 447, + 102 + ], + "spans": [ + { + "bbox": [ + 395, + 91, + 447, + 102 + ], + "type": "text", + "content": "Part2Object" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 470, + 91, + 481, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 470, + 91, + 481, + 100 + ], + "spans": [ + { + "bbox": [ + 470, + 91, + 481, + 100 + ], + "type": "text", + "content": "13" + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 12 + }, + { + "para_blocks": [ + { + "bbox": [ + 132, + 114, + 481, + 137 + ], + "lines": [ + { + "bbox": [ + 132, + 114, + 481, + 137 + ], + "spans": [ + { + "bbox": [ + 132, + 114, + 481, + 137 + ], + "type": "text", + "content": "Table 4: Ablation study. We conduct ablation studies on ScanNet val. We validate the effectiveness of our Part2Object clustering, Hi-Mask3D, and self-training procedure." + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "text" + }, + { + "type": "table", + "bbox": [ + 143, + 180, + 284, + 308 + ], + "blocks": [ + { + "bbox": [ + 132, + 144, + 301, + 177 + ], + "lines": [ + { + "bbox": [ + 132, + 144, + 301, + 177 + ], + "spans": [ + { + "bbox": [ + 132, + 144, + 301, + 177 + ], + "type": "text", + "content": "(a) Ablation study on Part2Object clustering and Hi-Mask3D architecture. OG, FU denotes objectness guidance and cluster feature function, respectively." + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 143, + 180, + 284, + 308 + ], + "lines": [ + { + "bbox": [ + 143, + 180, + 284, + 308 + ], + "spans": [ + { + "bbox": [ + 143, + 180, + 284, + 308 + ], + "type": "table", + "html": "
Ablation on ClusteringScanNet Val [12]
m@25m@50mAP
Baseline37.813.86.1
w/o OG43.220.710.4
w/o FU44.221.510.6
Ours55.126.812.6
Architecture
Mask3D59.031.014.2
Hi-Mask3D64.936.016.9
", + "image_path": "f05d3d2b49e436f2e2e9e286de18319344b0eda29ba5ed82438c5d400427bf9e.jpg" + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "table_body" + } + ], + "index": 4 + }, + { + "type": "table", + "bbox": [ + 309, + 165, + 468, + 314 + ], + "blocks": [ + { + "bbox": [ + 310, + 144, + 478, + 161 + ], + "lines": [ + { + "bbox": [ + 310, + 144, + 478, + 161 + ], + "spans": [ + { + "bbox": [ + 310, + 144, + 478, + 161 + ], + "type": "text", + "content": "(b) Ablation study on Self-training and Hyper-parameter." + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 309, + 165, + 468, + 314 + ], + "lines": [ + { + "bbox": [ + 309, + 165, + 468, + 314 + ], + "spans": [ + { + "bbox": [ + 309, + 165, + 468, + 314 + ], + "type": "table", + "html": "
Ablation on Self-TrainingScanNet Val [12]
m@25m@50mAP
Pseudo labels55.126.812.6
Round 160.732.615.3
Round 264.335.216.5
Round 364.936.016.9
Hyper-parameter
K = 0.455.722.610.3
K = 0.557.025.312.0
K = 0.655.126.812.6
K = 0.744.521.110.6
", + "image_path": "e719ce9b95c8651cf03b8f4959e3d50976344adc3b8ede21a6ebdbd2f3f203fd.jpg" + } + ] + } + ], + "index": 6, + "angle": 0, + "type": "table_body" + } + ], + "index": 6 + }, + { + "bbox": [ + 130, + 330, + 480, + 401 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 330, + 480, + 401 + ], + "spans": [ + { + "bbox": [ + 130, + 330, + 480, + 401 + ], + "type": "text", + "content": "Ablation on Hi-Mask3D. We compare our Hi-Mask3D with Mask3D using the same pseudo-labels extracted from Part2Object. Hi-Mask3D improves Mask3D by " + }, + { + "bbox": [ + 130, + 330, + 480, + 401 + ], + "type": "inline_equation", + "content": "2.7\\%" + }, + { + "bbox": [ + 130, + 330, + 480, + 401 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 130, + 330, + 480, + 401 + ], + "type": "inline_equation", + "content": "5.0\\%" + }, + { + "bbox": [ + 130, + 330, + 480, + 401 + ], + "type": "text", + "content": " in terms of mAP and mAP@50, respectively. This demonstrates that additional part information aids in object understanding. Furthermore, through reporting the performance of self-training, we observe that after several rounds, the mAP increases from " + }, + { + "bbox": [ + 130, + 330, + 480, + 401 + ], + "type": "inline_equation", + "content": "15.3\\%" + }, + { + "bbox": [ + 130, + 330, + 480, + 401 + ], + "type": "text", + "content": " to " + }, + { + "bbox": [ + 130, + 330, + 480, + 401 + ], + "type": "inline_equation", + "content": "16.9\\%" + }, + { + "bbox": [ + 130, + 330, + 480, + 401 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 130, + 403, + 481, + 462 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 403, + 481, + 462 + ], + "spans": [ + { + "bbox": [ + 130, + 403, + 481, + 462 + ], + "type": "text", + "content": "Ablation on Hyper-parameter. We conduct ablation on the import hyperparameter " + }, + { + "bbox": [ + 130, + 403, + 481, + 462 + ], + "type": "inline_equation", + "content": "K" + }, + { + "bbox": [ + 130, + 403, + 481, + 462 + ], + "type": "text", + "content": " in Equ 3. Since the number of clusters at each layer varies, we use percentages (K=0.6 indicates the top 60%) to measure how many of the top clusters from each layer can be aggregated. Table 4b shows the performance of Part2Object at different " + }, + { + "bbox": [ + 130, + 403, + 481, + 462 + ], + "type": "inline_equation", + "content": "K" + }, + { + "bbox": [ + 130, + 403, + 481, + 462 + ], + "type": "text", + "content": ". We use grey to indicate the default setting." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 132, + 483, + 317, + 495 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 483, + 317, + 495 + ], + "spans": [ + { + "bbox": [ + 132, + 483, + 317, + 495 + ], + "type": "text", + "content": "5 Conclusion and Limitations" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 130, + 510, + 482, + 666 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 510, + 482, + 666 + ], + "spans": [ + { + "bbox": [ + 130, + 510, + 482, + 666 + ], + "type": "text", + "content": "We introduce Part2Object, an efficient hierarchical clustering algorithm, that progressively groups point clouds into object parts and objects, while leveraging 3D object-ness prior to precisely target objects. Furthermore, we propose the Hierarchical-Aware Mask3D, facilitating self-training with pseudo-object and part labels from Part2Object. Experimental results demonstrate that we consistently surpass all existing methods in both unsupervised settings and data-efficient settings. Ethics Statement: Given that our 2D knowledge is derived from the self-supervised models DINO, we acknowledge that biases and controversies inherent in the training data for these models may be introduced into our model. Acknowledgment: This work was supported by the National Natural Science Foundation of China (No.62206174) and MoE Key Laboratory of Intelligent Perception and Human-Machine Collaboration (ShanghaiTech University)." + } + ] + } + ], + "index": 10 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "spans": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "type": "text", + "content": "14" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 166, + 91, + 217, + 102 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 166, + 91, + 217, + 102 + ], + "spans": [ + { + "bbox": [ + 166, + 91, + 217, + 102 + ], + "type": "text", + "content": "Cheng et al." + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 13 + }, + { + "para_blocks": [ + { + "bbox": [ + 132, + 114, + 198, + 127 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 114, + 198, + 127 + ], + "spans": [ + { + "bbox": [ + 132, + 114, + 198, + 127 + ], + "type": "text", + "content": "References" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 138, + 137, + 480, + 665 + ], + "type": "list", + "angle": 0, + "index": 20, + "blocks": [ + { + "bbox": [ + 138, + 137, + 480, + 158 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 137, + 480, + 158 + ], + "spans": [ + { + "bbox": [ + 138, + 137, + 480, + 158 + ], + "type": "text", + "content": "1. Adams, R., Bischof, L.: Seeded region growing. IEEE Transactions on pattern analysis and machine intelligence 16(6), 641-647 (1994)" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 138, + 159, + 480, + 179 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 159, + 480, + 179 + ], + "spans": [ + { + "bbox": [ + 138, + 159, + 480, + 179 + ], + "type": "text", + "content": "2. Amir, S., Gandelsman, Y., Bagon, S., Dekel, T.: Deep vit features as dense visual descriptors. arXiv preprint arXiv:2112.05814 (2021)" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 138, + 180, + 480, + 212 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 180, + 480, + 212 + ], + "spans": [ + { + "bbox": [ + 138, + 180, + 480, + 212 + ], + "type": "text", + "content": "3. An, D., Wang, H., Wang, W., Wang, Z., Huang, Y., He, K., Wang, L.: Etpnav: Evolving topological planning for vision-language navigation in continuous environments. arXiv preprint arXiv:2304.03047 (2023)" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 138, + 213, + 480, + 245 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 213, + 480, + 245 + ], + "spans": [ + { + "bbox": [ + 138, + 213, + 480, + 245 + ], + "type": "text", + "content": "4. Armeni, I., Sener, O., Zamir, A.R., Jiang, H., Brilakis, I., Fischer, M., Savarese, S.: 3d semantic parsing of large-scale indoor spaces. In: Proceedings of the IEEE conference on computer vision and pattern recognition. pp. 1534-1543 (2016)" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 138, + 245, + 480, + 277 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 245, + 480, + 277 + ], + "spans": [ + { + "bbox": [ + 138, + 245, + 480, + 277 + ], + "type": "text", + "content": "5. Carion, N., Massa, F., Synnaeve, G., Usunier, N., Kirillov, A., Zagoruyko, S.: End-to-end object detection with transformers. In: European conference on computer vision. pp. 213-229. Springer (2020)" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 138, + 277, + 480, + 309 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 277, + 480, + 309 + ], + "spans": [ + { + "bbox": [ + 138, + 277, + 480, + 309 + ], + "type": "text", + "content": "6. Caron, M., Touvron, H., Misra, I., Jégou, H., Mairal, J., Bojanowski, P., Joulin, A.: Emerging properties in self-supervised vision transformers. In: Proceedings of the International Conference on Computer Vision (ICCV) (2021)" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 138, + 309, + 480, + 342 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 309, + 480, + 342 + ], + "spans": [ + { + "bbox": [ + 138, + 309, + 480, + 342 + ], + "type": "text", + "content": "7. Chen, R., Liu, Y., Kong, L., Chen, N., Xinge, Z., Ma, Y., Liu, T., Wang, W.: Towards label-free scene understanding by vision foundation models. In: Thirty-seventh Conference on Neural Information Processing Systems (2023)" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 138, + 342, + 480, + 373 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 342, + 480, + 373 + ], + "spans": [ + { + "bbox": [ + 138, + 342, + 480, + 373 + ], + "type": "text", + "content": "8. Chen, S., Fang, J., Zhang, Q., Liu, W., Wang, X.: Hierarchical aggregation for 3d instance segmentation. In: Proceedings of the IEEE/CVF International Conference on Computer Vision. pp. 15467-15476 (2021)" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 138, + 374, + 480, + 406 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 374, + 480, + 406 + ], + "spans": [ + { + "bbox": [ + 138, + 374, + 480, + 406 + ], + "type": "text", + "content": "9. Chen, Z., Yin, K., Fisher, M., Chaudhuri, S., Zhang, H.: Bae-net: Branched autoencoder for shape co-segmentation. In: Proceedings of the IEEE/CVF International Conference on Computer Vision. pp. 8490-8499 (2019)" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 138, + 406, + 480, + 438 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 406, + 480, + 438 + ], + "spans": [ + { + "bbox": [ + 138, + 406, + 480, + 438 + ], + "type": "text", + "content": "0. Chibane, J., Engelmann, F., Anh Tran, T., Pons-Moll, G.: Box2mask: Weakly supervised 3d semantic instance segmentation using bounding boxes. In: European Conference on Computer Vision. pp. 681-699. Springer (2022)" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 138, + 438, + 480, + 471 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 438, + 480, + 471 + ], + "spans": [ + { + "bbox": [ + 138, + 438, + 480, + 471 + ], + "type": "text", + "content": "1. Choy, C., Gwak, J., Savarese, S.: 4d spatio-temporal convnets: Minkowski convolutional neural networks. In: Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition. pp. 3075-3084 (2019)" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 138, + 471, + 480, + 503 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 471, + 480, + 503 + ], + "spans": [ + { + "bbox": [ + 138, + 471, + 480, + 503 + ], + "type": "text", + "content": "2. Dai, A., Chang, A.X., Savva, M., Halber, M., Funkhouser, T., Nießner, M.: Scannet: Richly-annotated 3d reconstructions of indoor scenes. In: Proc. Computer Vision and Pattern Recognition (CVPR), IEEE (2017)" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 138, + 503, + 480, + 525 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 503, + 480, + 525 + ], + "spans": [ + { + "bbox": [ + 138, + 503, + 480, + 525 + ], + "type": "text", + "content": "3. Dai, Q., Yang, S.: Curriculum point prompting for weakly-supervised referring segmentation (2024)" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 138, + 525, + 480, + 556 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 525, + 480, + 556 + ], + "spans": [ + { + "bbox": [ + 138, + 525, + 480, + 556 + ], + "type": "text", + "content": "4. Deng, R., Shen, C., Liu, S., Wang, H., Liu, X.: Learning to predict crisp boundaries. In: Proceedings of the European conference on computer vision (ECCV). pp. 562-578 (2018)" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 138, + 556, + 480, + 600 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 556, + 480, + 600 + ], + "spans": [ + { + "bbox": [ + 138, + 556, + 480, + 600 + ], + "type": "text", + "content": "5. Ester, M., Kriegel, H.P., Sander, J., Xu, X.: A density-based algorithm for discovering clusters in large spatial databases with noise. In: Proceedings of the Second International Conference on Knowledge Discovery and Data Mining. p. 226-231. KDD'96, AAAI Press (1996)" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 138, + 601, + 480, + 621 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 601, + 480, + 621 + ], + "spans": [ + { + "bbox": [ + 138, + 601, + 480, + 621 + ], + "type": "text", + "content": "6. Felzenszwalb, P.F., Huttenlocher, D.P.: Efficient graph-based image segmentation. International journal of computer vision 59, 167-181 (2004)" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 138, + 621, + 480, + 665 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 621, + 480, + 665 + ], + "spans": [ + { + "bbox": [ + 138, + 621, + 480, + 665 + ], + "type": "text", + "content": "7. Geng, H., Xu, H., Zhao, C., Xu, C., Yi, L., Huang, S., Wang, H.: Gapartnet: Cross-category domain-generalizable object perception and manipulation via generalizable and actionable parts. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 7081-7091 (2023)" + } + ] + } + ], + "index": 19 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 395, + 91, + 447, + 102 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 395, + 91, + 447, + 102 + ], + "spans": [ + { + "bbox": [ + 395, + 91, + 447, + 102 + ], + "type": "text", + "content": "Part2Object" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 470, + 91, + 481, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 470, + 91, + 481, + 100 + ], + "spans": [ + { + "bbox": [ + 470, + 91, + 481, + 100 + ], + "type": "text", + "content": "15" + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 14 + }, + { + "para_blocks": [ + { + "bbox": [ + 132, + 116, + 481, + 665 + ], + "type": "list", + "angle": 0, + "index": 18, + "blocks": [ + { + "bbox": [ + 133, + 116, + 480, + 138 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 116, + 480, + 138 + ], + "spans": [ + { + "bbox": [ + 133, + 116, + 480, + 138 + ], + "type": "text", + "content": "18. Ghiasi, G., Gu, X., Cui, Y., Lin, T.Y.: Open-vocabulary image segmentation. arXiv preprint arXiv:2112.12143 (2021)" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 133, + 139, + 481, + 182 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 139, + 481, + 182 + ], + "spans": [ + { + "bbox": [ + 133, + 139, + 481, + 182 + ], + "type": "text", + "content": "19. Han, L., Zheng, T., Xu, L., Fang, L.: Occuseg: Occupancy-aware 3d instance segmentation. In: 2020 IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR). pp. 2937-2946 (2020). https://doi.org/10.1109/CVPR42600.2020.00301" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 132, + 182, + 480, + 215 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 182, + 480, + 215 + ], + "spans": [ + { + "bbox": [ + 132, + 182, + 480, + 215 + ], + "type": "text", + "content": "20. Hou, J., Dai, A., Nießner, M.: 3d-sis: 3d semantic instance segmentation of rgb-d scans. In: Proceedings of the IEEE/CVF conference on computer vision and pattern recognition. pp. 4421-4430 (2019)" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 132, + 216, + 480, + 248 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 216, + 480, + 248 + ], + "spans": [ + { + "bbox": [ + 132, + 216, + 480, + 248 + ], + "type": "text", + "content": "21. Hou, J., Graham, B., Nießner, M., Xie, S.: Exploring data-efficient 3d scene understanding with contrastive scene contexts. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 15587-15597 (2021)" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 132, + 249, + 480, + 292 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 249, + 480, + 292 + ], + "spans": [ + { + "bbox": [ + 132, + 249, + 480, + 292 + ], + "type": "text", + "content": "22. Hu, Q., Yang, B., Xie, L., Rosa, S., Guo, Y., Wang, Z., Trigoni, N., Markham, A.: Randla-net: Efficient semantic segmentation of large-scale point clouds. In: Proceedings of the IEEE/CVF conference on computer vision and pattern recognition. pp. 11108-11117 (2020)" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 132, + 293, + 480, + 335 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 293, + 480, + 335 + ], + "spans": [ + { + "bbox": [ + 132, + 293, + 480, + 335 + ], + "type": "text", + "content": "23. Huang, S., Wang, Z., Li, P., Jia, B., Liu, T., Zhu, Y., Liang, W., Zhu, S.C.: Diffusion-based generation, optimization, and planning in 3d scenes. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 16750-16761 (2023)" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 132, + 336, + 480, + 369 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 336, + 480, + 369 + ], + "spans": [ + { + "bbox": [ + 132, + 336, + 480, + 369 + ], + "type": "text", + "content": "24. Huang, Z., Wu, X., Chen, X., Zhao, H., Zhu, L., Lasenby, J.: Openins3d: Snap and lookup for 3d open-vocabulary instance segmentation. arXiv preprint arXiv:2309.00616 (2023)" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 132, + 369, + 480, + 391 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 369, + 480, + 391 + ], + "spans": [ + { + "bbox": [ + 132, + 369, + 480, + 391 + ], + "type": "text", + "content": "25. Hui, L., Tang, L., Shen, Y., Xie, J., Yang, J.: Learning superpoint graph cut for 3d instance segmentation. In: NeurIPS (2022)" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 132, + 392, + 480, + 423 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 392, + 480, + 423 + ], + "spans": [ + { + "bbox": [ + 132, + 392, + 480, + 423 + ], + "type": "text", + "content": "26. Kirillov, A., Mintun, E., Ravi, N., Mao, H., Rolland, C., Gustafson, L., Xiao, T., Whitehead, S., Berg, A.C., Lo, W.Y., et al.: Segment anything. arXiv preprint arXiv:2304.02643 (2023)" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 132, + 424, + 480, + 456 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 424, + 480, + 456 + ], + "spans": [ + { + "bbox": [ + 132, + 424, + 480, + 456 + ], + "type": "text", + "content": "27. Kolodiazhnyi, M., Rukhovich, D., Vorontsova, A., Konushin, A.: Top-down beats bottom-up in 3d instance segmentation (2023). https://doi.org/10.48550/ ARXIV.2302.02871, https://arxiv.org/abs/2302.02871" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 132, + 457, + 480, + 489 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 457, + 480, + 489 + ], + "spans": [ + { + "bbox": [ + 132, + 457, + 480, + 489 + ], + "type": "text", + "content": "28. Li, B., Weinberger, K.Q., Belongie, S., Koltun, V., Ranftl, R.: Language-driven semantic segmentation. In: International Conference on Learning Representations (2022), https://openreview.net/forum?id=RriDjddCLN" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 132, + 490, + 480, + 522 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 490, + 480, + 522 + ], + "spans": [ + { + "bbox": [ + 132, + 490, + 480, + 522 + ], + "type": "text", + "content": "29. Lin, T.Y., Goyal, P., Girshick, R., He, K., Dollar, P.: Focal loss for dense object detection. In: Proceedings of the IEEE international conference on computer vision. pp. 2980-2988 (2017)" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 132, + 523, + 480, + 566 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 523, + 480, + 566 + ], + "spans": [ + { + "bbox": [ + 132, + 523, + 480, + 566 + ], + "type": "text", + "content": "30. Liu, J., Yu, M., Ni, B., Chen, Y.: Self-prediction for joint instance and semantic segmentation of point clouds. In: Computer Vision-ECCV 2020: 16th European Conference, Glasgow, UK, August 23-28, 2020, Proceedings, Part XXII 16. pp. 187-204. Springer (2020)" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 132, + 567, + 480, + 599 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 567, + 480, + 599 + ], + "spans": [ + { + "bbox": [ + 132, + 567, + 480, + 599 + ], + "type": "text", + "content": "31. Liu, Y., Kong, L., Cen, J., Chen, R., Zhang, W., Pan, L., Chen, K., Liu, Z.: Segment any point cloud sequences by distilling vision foundation models. arXiv preprint arXiv:2306.09347 (2023)" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 132, + 600, + 480, + 632 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 600, + 480, + 632 + ], + "spans": [ + { + "bbox": [ + 132, + 600, + 480, + 632 + ], + "type": "text", + "content": "32. McInnes, L., Healy, J.: Accelerated hierarchical density based clustering. In: 2017 IEEE International Conference on Data Mining Workshops (ICDMW). pp. 33-42. IEEE (2017)" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 132, + 632, + 480, + 665 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 632, + 480, + 665 + ], + "spans": [ + { + "bbox": [ + 132, + 632, + 480, + 665 + ], + "type": "text", + "content": "33. Nunes, L., Chen, X., Marcuzzi, R., Osep, A., Leal-Taixe, L., Stachniss, C., Behley, J.: Unsupervised class-agnostic instance segmentation of 3d lidar data for autonomous vehicles. IEEE Robotics and Automation Letters 7(4), 8713-8720 (2022)" + } + ] + } + ], + "index": 17 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "spans": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "type": "text", + "content": "16" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 166, + 91, + 217, + 102 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 166, + 91, + 217, + 102 + ], + "spans": [ + { + "bbox": [ + 166, + 91, + 217, + 102 + ], + "type": "text", + "content": "Cheng et al." + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 15 + }, + { + "para_blocks": [ + { + "bbox": [ + 130, + 116, + 482, + 666 + ], + "type": "list", + "angle": 0, + "index": 19, + "blocks": [ + { + "bbox": [ + 130, + 116, + 482, + 149 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 116, + 482, + 149 + ], + "spans": [ + { + "bbox": [ + 130, + 116, + 482, + 149 + ], + "type": "text", + "content": "34. Papon, J., Abramov, A., Schoeler, M., Worgotter, F.: Voxel cloud connectivity segmentation-supervoxels for point clouds. In: Proceedings of the IEEE conference on computer vision and pattern recognition. pp. 2027-2034 (2013)" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 130, + 150, + 482, + 194 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 150, + 482, + 194 + ], + "spans": [ + { + "bbox": [ + 130, + 150, + 482, + 194 + ], + "type": "text", + "content": "35. Paszke, A., Gross, S., Massa, F., Lerer, A., Bradbury, J., Chanan, G., Killeen, T., Lin, Z., Gimelshein, N., Antiga, L., et al.: Pytorch: An imperative style, high-performance deep learning library. Advances in neural information processing systems 32 (2019)" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 130, + 194, + 482, + 237 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 194, + 482, + 237 + ], + "spans": [ + { + "bbox": [ + 130, + 194, + 482, + 237 + ], + "type": "text", + "content": "36. Peng, S., Genova, K., Jiang, C., Tagliasacchi, A., Pollefeys, M., Funkhouser, T., et al.: Openscene: 3d scene understanding with open vocabularies. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 815-824 (2023)" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 130, + 237, + 482, + 270 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 237, + 482, + 270 + ], + "spans": [ + { + "bbox": [ + 130, + 237, + 482, + 270 + ], + "type": "text", + "content": "37. Qi, C.R., Yi, L., Su, H., Guibas, L.J.: Pointnet++: Deep hierarchical feature learning on point sets in a metric space. Advances in neural information processing systems 30 (2017)" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 130, + 270, + 482, + 304 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 270, + 482, + 304 + ], + "spans": [ + { + "bbox": [ + 130, + 270, + 482, + 304 + ], + "type": "text", + "content": "38. Rethage, D., Wald, J., Sturm, J., Navab, N., Tombari, F.: Fully-convolutional point networks for large-scale point clouds. In: Proceedings of the European Conference on Computer Vision (ECCV). pp. 596-611 (2018)" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 130, + 304, + 482, + 336 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 304, + 482, + 336 + ], + "spans": [ + { + "bbox": [ + 130, + 304, + 482, + 336 + ], + "type": "text", + "content": "39. Rozenberszki, D., Litany, O., Dai, A.: Language-grounded indoor 3d semantic segmentation in the wild. In: European Conference on Computer Vision. pp. 125-141. Springer (2022)" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 130, + 336, + 482, + 369 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 336, + 482, + 369 + ], + "spans": [ + { + "bbox": [ + 130, + 336, + 482, + 369 + ], + "type": "text", + "content": "40. Rozenberszki, D., Litany, O., Dai, A.: Language-grounded indoor 3d semantic segmentation in the wild. In: Proceedings of the European Conference on Computer Vision (ECCV) (2022)" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 130, + 369, + 482, + 392 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 369, + 482, + 392 + ], + "spans": [ + { + "bbox": [ + 130, + 369, + 482, + 392 + ], + "type": "text", + "content": "41. Rozenberszki, D., Litany, O., Dai, A.: Unscene3d: Unsupervised 3d instance segmentation for indoor scenes. arXiv preprint arXiv:2303.14541 (2023)" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 130, + 392, + 482, + 424 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 392, + 482, + 424 + ], + "spans": [ + { + "bbox": [ + 130, + 392, + 482, + 424 + ], + "type": "text", + "content": "42. Schult, J., Engelmann, F., Hermans, A., Litany, O., Tang, S., Leibe, B.: Mask3D for 3D Semantic Instance Segmentation. In: International Conference on Robotics and Automation (ICRA) (2023)" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 130, + 424, + 482, + 457 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 424, + 482, + 457 + ], + "spans": [ + { + "bbox": [ + 130, + 424, + 482, + 457 + ], + "type": "text", + "content": "43. Shi, C., Yang, S.: Edadet: Open-vocabulary object detection using early dense alignment. In: Proceedings of the IEEE/CVF International Conference on Computer Vision. pp. 15724-15734 (2023)" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 130, + 457, + 482, + 479 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 457, + 482, + 479 + ], + "spans": [ + { + "bbox": [ + 130, + 457, + 482, + 479 + ], + "type": "text", + "content": "44. Shi, C., Yang, S.: Logoprompt: Synthetic text images can be good visual prompts for vision-language models. arXiv preprint arXiv:2309.01155 (2023)" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 130, + 479, + 482, + 522 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 479, + 482, + 522 + ], + "spans": [ + { + "bbox": [ + 130, + 479, + 482, + 522 + ], + "type": "text", + "content": "45. Shi, C., Yang, S.: The devil is in the object boundary: Towards annotation-free instance segmentation using foundation models. In: The Twelfth International Conference on Learning Representations (2024), https://openreview.net/forum?id=4JbrdrHxYy" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 130, + 522, + 482, + 545 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 522, + 482, + 545 + ], + "spans": [ + { + "bbox": [ + 130, + 522, + 482, + 545 + ], + "type": "text", + "content": "46. Shi, J., Malik, J.: Normalized cuts and image segmentation. IEEE Transactions on pattern analysis and machine intelligence 22(8), 888-905 (2000)" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 130, + 545, + 482, + 578 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 545, + 482, + 578 + ], + "spans": [ + { + "bbox": [ + 130, + 545, + 482, + 578 + ], + "type": "text", + "content": "47. Song, Z., Yang, B.: Ogc: Unsupervised 3d object segmentation from rigid dynamics of point clouds. Advances in Neural Information Processing Systems 35, 30798-30812 (2022)" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 130, + 578, + 482, + 611 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 578, + 482, + 611 + ], + "spans": [ + { + "bbox": [ + 130, + 578, + 482, + 611 + ], + "type": "text", + "content": "48. Stewart, R., Andriluka, M., Ng, A.Y.: End-to-end people detection in crowded scenes. In: Proceedings of the IEEE conference on computer vision and pattern recognition. pp. 2325-2333 (2016)" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 130, + 611, + 482, + 643 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 611, + 482, + 643 + ], + "spans": [ + { + "bbox": [ + 130, + 611, + 482, + 643 + ], + "type": "text", + "content": "49. Straub, J., Whelan, T., Ma, L., Chen, Y., Wijmans, E., Green, S., Engel, J.J., Mur-Artal, R., Ren, C., Verma, S., et al.: The replica dataset: A digital replica of indoor spaces. arXiv preprint arXiv:1906.05797 (2019)" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 130, + 643, + 482, + 666 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 643, + 482, + 666 + ], + "spans": [ + { + "bbox": [ + 130, + 643, + 482, + 666 + ], + "type": "text", + "content": "50. Sun, J., Qing, C., Tan, J., Xu, X.: Superpoint transformer for 3d scene instance segmentation (2022)" + } + ] + } + ], + "index": 18 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 395, + 91, + 447, + 102 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 395, + 91, + 447, + 102 + ], + "spans": [ + { + "bbox": [ + 395, + 91, + 447, + 102 + ], + "type": "text", + "content": "Part2Object" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 470, + 91, + 481, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 470, + 91, + 481, + 100 + ], + "spans": [ + { + "bbox": [ + 470, + 91, + 481, + 100 + ], + "type": "text", + "content": "17" + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 16 + }, + { + "para_blocks": [ + { + "bbox": [ + 132, + 116, + 482, + 654 + ], + "type": "list", + "angle": 0, + "index": 19, + "blocks": [ + { + "bbox": [ + 132, + 116, + 482, + 160 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 116, + 482, + 160 + ], + "spans": [ + { + "bbox": [ + 132, + 116, + 482, + 160 + ], + "type": "text", + "content": "51. Suo, S., Wong, K., Xu, J., Tu, J., Cui, A., Casas, S., Urtasun, R.: Mixsim: A hierarchical framework for mixed reality traffic simulation. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 9622-9631 (2023)" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 132, + 161, + 481, + 194 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 161, + 481, + 194 + ], + "spans": [ + { + "bbox": [ + 132, + 161, + 481, + 194 + ], + "type": "text", + "content": "52. Suomela, L., Kalliola, J., Dag, A., Edelman, H., Kämäräinen, J.K.: Benchmarking visual localization for autonomous navigation. In: Proceedings of the IEEE/CVF Winter Conference on Applications of Computer Vision. pp. 2945-2955 (2023)" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 132, + 194, + 481, + 226 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 194, + 481, + 226 + ], + "spans": [ + { + "bbox": [ + 132, + 194, + 481, + 226 + ], + "type": "text", + "content": "53. Tang, J., Zheng, G., Shi, C., Yang, S.: Contrastive grouping with transformer for referring image segmentation. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 23570-23580 (2023)" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 132, + 227, + 481, + 258 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 227, + 481, + 258 + ], + "spans": [ + { + "bbox": [ + 132, + 227, + 481, + 258 + ], + "type": "text", + "content": "54. Vaswani, A., Shazeer, N., Parmar, N., Uszkoreit, J., Jones, L., Gomez, A.N., Kaiser, L., Polosukhin, I.: Attention is all you need. Advances in neural information processing systems 30 (2017)" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 132, + 259, + 481, + 292 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 259, + 481, + 292 + ], + "spans": [ + { + "bbox": [ + 132, + 259, + 481, + 292 + ], + "type": "text", + "content": "55. Vu, T., Kim, K., Luu, T.M., Nguyen, T., Kim, J., Yoo, C.D.: Softgroup++: Scalable 3d instance segmentation with octree pyramid grouping. arXiv preprint arXiv:2209.08263 (2022)" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 132, + 293, + 481, + 314 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 293, + 481, + 314 + ], + "spans": [ + { + "bbox": [ + 132, + 293, + 481, + 314 + ], + "type": "text", + "content": "56. Vu, T., Kim, K., Luu, T.M., Nguyen, X.T., Yoo, C.D.: Softgroup for 3d instance segmentation on 3d point clouds. In: CVPR (2022)" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 132, + 315, + 481, + 346 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 315, + 481, + 346 + ], + "spans": [ + { + "bbox": [ + 132, + 315, + 481, + 346 + ], + "type": "text", + "content": "57. Wang, R., Zhang, Y., Mao, J., Zhang, R., Cheng, C.Y., Wu, J.: Ikea-manual: Seeing shape assembly step by step. Advances in Neural Information Processing Systems 35, 28428-28440 (2022)" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 132, + 347, + 481, + 380 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 347, + 481, + 380 + ], + "spans": [ + { + "bbox": [ + 132, + 347, + 481, + 380 + ], + "type": "text", + "content": "58. Wang, W., Yu, R., Huang, Q., Neumann, U.: Sgpn: Similarity group proposal network for 3d point cloud instance segmentation. In: Proceedings of the IEEE conference on computer vision and pattern recognition. pp. 2569-2578 (2018)" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 132, + 380, + 481, + 413 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 380, + 481, + 413 + ], + "spans": [ + { + "bbox": [ + 132, + 380, + 481, + 413 + ], + "type": "text", + "content": "59. Wang, X., Yu, Z., De Mello, S., Kautz, J., Anandkumar, A., Shen, C., Alvarez, J.M.: FreeSOLO: Learning to segment objects without annotations. arXiv preprint arXiv:2202.12181 (2022)" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 132, + 413, + 481, + 435 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 413, + 481, + 435 + ], + "spans": [ + { + "bbox": [ + 132, + 413, + 481, + 435 + ], + "type": "text", + "content": "60. Wang, X., Girdhar, R., Yu, S.X., Misra, I.: Cut and learn for unsupervised object detection and instance segmentation. arXiv preprint arXiv:2301.11320 (2023)" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 132, + 435, + 481, + 479 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 435, + 481, + 479 + ], + "spans": [ + { + "bbox": [ + 132, + 435, + 481, + 479 + ], + "type": "text", + "content": "61. Wang, Y., Shen, X., Hu, S.X., Yuan, Y., Crowley, J.L., Vaufreydaz, D.: Self-supervised transformers for unsupervised object discovery using normalized cut. In: Conference on Computer Vision and Pattern Recognition. New Orleans, LA, USA (June 2022)" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 132, + 479, + 481, + 511 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 479, + 481, + 511 + ], + "spans": [ + { + "bbox": [ + 132, + 479, + 481, + 511 + ], + "type": "text", + "content": "62. Wang, Y., Sun, Y., Liu, Z., Sarma, S.E., Bronstein, M.M., Solomon, J.M.: Dynamic graph cnn for learning on point clouds. Acm Transactions On Graphics (tog) 38(5), 1-12 (2019)" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 132, + 511, + 481, + 544 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 511, + 481, + 544 + ], + "spans": [ + { + "bbox": [ + 132, + 511, + 481, + 544 + ], + "type": "text", + "content": "63. Wu, W., Qi, Z., Fuxin, L.: Pointconv: Deep convolutional networks on 3d point clouds. In: Proceedings of the IEEE/CVF Conference on computer vision and pattern recognition. pp. 9621-9630 (2019)" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 132, + 545, + 481, + 567 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 545, + 481, + 567 + ], + "spans": [ + { + "bbox": [ + 132, + 545, + 481, + 567 + ], + "type": "text", + "content": "64. Yang, Y., Wu, X., He, T., Zhao, H., Liu, X.: Sam3d: Segment anything in 3d scenes. arXiv preprint arXiv:2306.03908 (2023)" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 132, + 567, + 481, + 599 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 567, + 481, + 599 + ], + "spans": [ + { + "bbox": [ + 132, + 567, + 481, + 599 + ], + "type": "text", + "content": "65. Zhang, B., Wonka, P.: Point cloud instance segmentation using probabilistic embeddings. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 8883-8892 (2021)" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 132, + 600, + 481, + 621 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 600, + 481, + 621 + ], + "spans": [ + { + "bbox": [ + 132, + 600, + 481, + 621 + ], + "type": "text", + "content": "66. Zhang, Z., Ding, J., Jiang, L., Dai, D., Xia, G.S.: Freepoint: Unsupervised point cloud instance segmentation. arXiv preprint arXiv:2305.06973 (2023)" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 132, + 622, + 481, + 654 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 622, + 481, + 654 + ], + "spans": [ + { + "bbox": [ + 132, + 622, + 481, + 654 + ], + "type": "text", + "content": "67. Zhang, Z., Yang, B., Wang, B., Li, B.: Growsp: Unsupervised semantic segmentation of 3d point clouds. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 17619-17629 (2023)" + } + ] + } + ], + "index": 18 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "spans": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "type": "text", + "content": "18" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 166, + 91, + 217, + 102 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 166, + 91, + 217, + 102 + ], + "spans": [ + { + "bbox": [ + 166, + 91, + 217, + 102 + ], + "type": "text", + "content": "Cheng et al." + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 17 + } + ], + "_backend": "vlm", + "_version_name": "2.6.4" +} \ No newline at end of file diff --git a/2024/PartCraft_ Crafting Creative Objects by Parts/4747571e-3f49-4aac-b94c-924e39815a4f_content_list.json b/2024/PartCraft_ Crafting Creative Objects by Parts/4747571e-3f49-4aac-b94c-924e39815a4f_content_list.json new file mode 100644 index 0000000000000000000000000000000000000000..047d775f2d014c9aa5f313da3d930c1450a1a59f --- /dev/null +++ b/2024/PartCraft_ Crafting Creative Objects by Parts/4747571e-3f49-4aac-b94c-924e39815a4f_content_list.json @@ -0,0 +1,1720 @@ +[ + { + "type": "text", + "text": "PartCraft: Crafting Creative Objects by Parts", + "text_level": 1, + "bbox": [ + 236, + 140, + 767, + 162 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Kam Woh $\\mathrm{Ng}^{1}$ , Xiatian Zhu $^{1,2}$ , Yi-Zhe Song $^{1}$ , and Tao Xiang $^{1}$", + "bbox": [ + 256, + 188, + 743, + 205 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "$^{1}$ CVSSP, University of Surrey, United Kingdom", + "bbox": [ + 334, + 215, + 696, + 231 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Surrey Institute for People-Centred AI", + "bbox": [ + 362, + 231, + 638, + 244 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "{kamwoh.ng, xiatian.zhu, y song, t.xiang}@surrey.ac.uk", + "bbox": [ + 303, + 244, + 699, + 258 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Abstract. This paper propels creative control in generative visual AI by allowing users to \"select\". Departing from traditional text or sketch-based methods, we for the first time allow users to choose visual concepts by parts for their creative endeavors. The outcome is fine-grained generation that precisely captures selected visual concepts, ensuring a holistically faithful and plausible result. To achieve this, we first parse objects into parts through unsupervised feature clustering. Then, we encode parts into text tokens and introduce an entropy-based normalized attention loss that operates on them. This loss design enables our model to learn generic prior topology knowledge about object's part composition, and further generalize to novel part compositions to ensure the generation looks holistically faithful. Lastly, we employ a bottleneck encoder to project the part tokens. This not only enhances fidelity but also accelerates learning, by leveraging shared knowledge and facilitating information exchange among instances. Visual results in the paper and supplementary material showcase the compelling power of PartCraft in crafting highly customized, innovative creations, exemplified by the \"charming\" and creative birds in Fig. 1. Code is released at https://github.com/kamwoh/partcraft.", + "bbox": [ + 261, + 295, + 743, + 559 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Keywords: Part Composition $\\cdot$ Controllable Text-to-image Generation", + "bbox": [ + 261, + 571, + 740, + 585 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "1 Introduction", + "text_level": 1, + "bbox": [ + 215, + 612, + 375, + 628 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Humans are creators; AI, on the other hand, hallucinates. Creativity, arguably, is the very force driving humanity forward. Recently, generative AI has garnered considerable attention for its perceived \"creativity\" [6, 19, 20, 43, 47, 49, 52, 55]. Despite its promise, the challenge of control has swiftly surfaced - how can humans infuse their creativity into the generation process and regulate the extent to which AI hallucinates?", + "bbox": [ + 212, + 643, + 787, + 733 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Creativity starts with an idea. The immediate challenge is how to articulate that idea and integrate it with generative AI. Text is the most commonly employed medium. For instance, imagine being a bird enthusiast wanting to craft the most unique bird akin to Fig. 1. The go-to approach would be furnishing Stable Diffusion [52] (or an equivalent model) with the following textual prompt (the idea): \"generate a bird with the head of $X$ , wings of $Y$ , and body of $Z$ \". While you might be presented with remarkably looking birds like Fig. 1(a), they", + "bbox": [ + 212, + 734, + 787, + 840 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/e5c3dba8567cb6b6bdad61f2d12b18caababc60418696b678e4cb1bb0f2171d4.jpg", + "image_caption": [ + "Fig. 1: People often form creative concepts based on existing ones [9, 42, 54, 65]. For instance, a bird enthusiast may want to craft a unique bird with different parts (e.g., heads, bodies, and wings) from common bird types (e.g., blue jay, cardinal and song sparrow). (a) Using text prompts in T2I models often results in a lack of control and deviation from the intended details, especially those visual details that are difficult to describe. (b) While sketching is a direct way, not everyone possesses the ability to sketch, particularly in intricate detail. (c) Even with reference images, existing methods (e.g., DreamBooth [53]) did not consider learning object parts, thus cannot generate with desired parts. (d) Using additional control is even cumbersome, requiring too many inputs! (e) We aim to create an object by simply selecting desired parts. PartCraft learns from visual examples to generate the object with a faithful holistic structure, seamlessly integrating the chosen parts into a natural and coherent entity." + ], + "image_footnote": [], + "bbox": [ + 218, + 143, + 784, + 321 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "may bear little resemblance to the envisioned concept. Recent literature suggests that a swift sketch could serve as a viable alternative [40, 63, 71], providing fine-grained shape control. However, the caveat is that not everyone possesses the ability to sketch, particularly in intricate detail.", + "bbox": [ + 212, + 517, + 787, + 577 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "In this paper, our primary focus revolves around addressing the issue of \"control\" in generative AI. We advance by introducing fine-grained control into the generative process, inviting you to \"select\". While this selection mechanism might seem modest at first, it closely mirrors the human creative process, where new concepts often emerge from existing ones [9,42,54,65]. Recall those moments when you desired an \"ideal\" pair of shoes with selected features from different pairs, or when you aimed to get creative with a cat (for that matter!).", + "bbox": [ + 212, + 580, + 787, + 685 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "It follows that rather than relying on writing (text) or drawing (sketch), all that is required is to choose the distinct visual concepts you specifically desire in your creative endeavor. Our model then ensures that all selected concepts are seamlessly and precisely composed into a faithful novel object in the final generation, without resorting to additional control such as bounding boxes [7,38, 40,66,71]. To illustrate with the \"unique bird\" example once more, our approach literally involves selecting the head of X, wings of Y, and body of Z! (see Fig. 1(e) and 2).", + "bbox": [ + 212, + 686, + 787, + 806 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Our solution is intuitive and centers around the well-studied computer vision concept of objects and their parts [8,14,22,27,35]. The challenge then boils down", + "bbox": [ + 212, + 809, + 787, + 840 + ], + "page_idx": 1 + }, + { + "type": "page_number", + "text": "2", + "bbox": [ + 217, + 114, + 227, + 126 + ], + "page_idx": 1 + }, + { + "type": "header", + "text": "KW. Ng et al.", + "bbox": [ + 271, + 114, + 367, + 128 + ], + "page_idx": 1 + }, + { + "type": "image", + "img_path": "images/4cf0ceb7b2a760d2da60d869c667e63255c96c0b5a3360d3a537ac5cc3e6e273.jpg", + "image_caption": [ + "Fig. 2: Two sets of images were generated from their original parts (sources A and B). We can integrate a specific part (e.g., body, head, or even background (BG) of source B to target A seamlessly without effort." + ], + "image_footnote": [], + "bbox": [ + 217, + 142, + 787, + 335 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "to two aspects: (i) how to parse known objects into their parts (i.e., recognizing that birds have eyes and tails), and (ii) how to assemble parts from different visual concepts to form a faithful creative concept (i.e., ensuring the output is recognizably a bird).", + "bbox": [ + 212, + 415, + 782, + 474 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "The former is easier. We make clever use of DINOv2 [45] feature maps and perform unsupervised clustering to identify common parts. The idea is that each cluster will then correspond to a semantic part of a common object (e.g., head and wings of birds). We specifically chose DINOv2 for its superior fine-grained perception [2] compared to others [50, 57, 58]. Further, this way has a higher flexibility to enhance fine-grained parsing by using a higher cluster count.", + "bbox": [ + 212, + 479, + 784, + 571 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Our major contribution lies in addressing the latter challenge. The solution is intuitive - it essentially revolves around the fine-grained selection and placement of the chosen parts. Inspired by recent efforts in personalization [3,25,36,53,64], primarily designed for learning entire objects, we introduce a tailored attention loss that specifically operates on parts. With this loss, our model learns the object part composition, ensuring the final generation appears holistically faithful with mixed parts (i.e., head and wings of a bird actually appear at the right places).", + "bbox": [ + 212, + 577, + 784, + 683 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "More specifically, we introduce an entropy-based attention loss that maximizes the attention of a specific part at a particular location while minimizing the attention where no parts appear. This is achieved by first selecting the attention maps that correspond to the parts. A normalization over all parts is then performed to ensure that each image region is occupied by no more than one part. Finally, we minimize the entropy loss between each normalized individual part and the semantic maps obtained during object parsing, containing the correct part location. This not only facilitates stronger part disentanglement but is also the key to generating a faithful holistic structure of an object as it learns a generic prior topology knowledge about object parts.", + "bbox": [ + 212, + 689, + 785, + 840 + ], + "page_idx": 2 + }, + { + "type": "header", + "text": "PartCraft: Crafting Creative Objects by Parts", + "bbox": [ + 421, + 114, + 732, + 128 + ], + "page_idx": 2 + }, + { + "type": "page_number", + "text": "3", + "bbox": [ + 774, + 114, + 785, + 126 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "To enhance generation fidelity, we further employ a bottleneck encoder to project the text tokens. This approach accelerates learning by leveraging shared knowledge (common parts) and facilitating information exchange among instances in each part. Each instance adjusts slightly to adapt to the fine-grained part details during optimization.", + "bbox": [ + 212, + 146, + 787, + 222 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Our contributions are as follows: (i) We pioneer a unique approach for fine-grained part-level control in Text-to-Image (T2I) models, empowering users to craft objects by selecting desired parts. This method streamlines the creative process, marking a significant advancement in our capacity to manipulate and reimagine visual content. (ii) We introduce PartCraft, a technique that autonomously parses object parts and orchestrates them from different visual concepts, resulting in the faithful creation of a novel object. (iii) For enhanced part disentanglement and generation fidelity, we propose an entropy-based normalized attention loss and leverage a bottleneck encoder. (iv) We present two problem-specific quantitative metrics. Comprehensive experiments on CUB-200-2011 (birds) and the Stanford Dogs dataset demonstrate the superior performance of our method in generating novel objects, surpassing alternative approaches in both qualitative and quantitative evaluations.", + "bbox": [ + 212, + 222, + 789, + 420 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "2 Related Work", + "text_level": 1, + "bbox": [ + 215, + 431, + 387, + 448 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Creative editing and generation. Creativity involves generating innovative ideas or artifacts across various domains [12]. Extensive research has explored the integration of creativity into Generative Adversarial Networks (GANs) [21,44,56] and Variational Autoencoders (VAEs) [16, 18]. For example, DoodlerGAN [26] learns and combines fine-level part components to create sketches of new species. A recent study by [60] demonstrated decomposing personalized concepts into distinct visual aspects, creatively recombined through diffusion models. Instruct-Pix2Pix [10] allows creative image editing through instructions, while ConceptLab [51] aims to identify novel concepts within a specified category, deviating from existing concepts. Different from these works where editing/generation usually focuses on the whole object, we instead focus on training a text-to-image generative model that can understand parts, thus able to creatively generate new objects by seamlessly composing different parts simply through selection.", + "bbox": [ + 212, + 462, + 787, + 657 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Text-to-image generation. Recent advancements in large text-to-image (T2I) diffusion models [6, 19, 20, 43, 49, 52, 55] have made significant improvements over conventional methods [37, 46, 59, 67, 70, 72] in producing high-fidelity images from text prompts. Their application scope has expanded to include both global [10, 32, 41] and localized [4, 17, 28, 68] image editing tasks, demonstrating versatility. Methods such as [5, 7, 13, 15, 23, 24, 34, 38, 66, 69, 71] have introduced more granular spatial control, such as incorporating semantic segmentation masks or bounding boxes, into large pretrained diffusion models to guide image generation. Contrary to these approaches, we focus on enhancing the control by enabling a straightforward discrete selection of desired parts. We minimize the complexity and manual intervention required by spatial controls, yet the model can compose selected parts as a coherent object seamlessly. This not", + "bbox": [ + 212, + 659, + 787, + 840 + ], + "page_idx": 3 + }, + { + "type": "page_number", + "text": "4", + "bbox": [ + 217, + 114, + 228, + 126 + ], + "page_idx": 3 + }, + { + "type": "header", + "text": "KW. Ng et al.", + "bbox": [ + 271, + 114, + 367, + 128 + ], + "page_idx": 3 + }, + { + "type": "image", + "img_path": "images/07aab029c3b582c0c91383aecfd2e6532c96a881fbc5b46150f13f502b722ee8.jpg", + "image_caption": [ + "Fig. 3: Overview of our PartCraft. All parts are organized into a dictionary, and their semantic embeddings are learned through a textual inversion approach. For instance, a text description like \"a photo of a [Head,42] [Wing,87]...\" guides the optimization of the corresponding textual embedding by reconstructing the associated image. To improve generation fidelity, we incorporate a bottleneck encoder $f$ (MLP) to compute the embedding $y$ (Eq. (4)) as input to the text encoder. To promote disentanglement among learned parts, we minimize a specially designed attention loss, denoted as $\\mathcal{L}_{\\text{att}}$ ." + ], + "image_footnote": [], + "bbox": [ + 220, + 157, + 789, + 299 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "only simplifies the user's role in the generative process but also ensures that the compositional logic and coherence are inherently managed by the model's ability. As such, our research is distinguished by focusing on the model's inherent ability to understand and apply part relationships instead of providing additional controls.", + "bbox": [ + 212, + 426, + 787, + 500 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Abstracting visual knowledge as a text token. The effectiveness of T2I models is constrained by the user's ability to articulate their desired image through text. These models face challenges in faithfully replicating visual characteristics from a reference set and generating innovative interpretations in diverse contexts, even with detailed textual descriptions. To address this challenge, various personalization techniques have been developed. These techniques obtain a new word embedding from multiple images depicting the same concept [1, 25, 36, 53, 61] or multiple new word embeddings for various concepts within a single image [3] through inversion. The learned visual concepts can be creatively reused in many image editing tasks. Nonetheless, most of these approaches struggle to learn object parts, often not able to follow the part selections due to part entanglement as they were designed to learn object as a whole. In this work, we introduce a customized attention loss that serves a dual purpose: ensure accurate positioning of each part and enforce each image region occupied by no more than one part. This greatly improves the part disentanglement, creating novel concepts with correct appearances (see Fig. 2).", + "bbox": [ + 212, + 503, + 787, + 744 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "3 Methodology", + "text_level": 1, + "bbox": [ + 215, + 772, + 380, + 791 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Given a set of unlabeled images depicting the same object (e.g., bird) with different part details, we aim to train a T2I generative model that decomposes", + "bbox": [ + 212, + 809, + 785, + 840 + ], + "page_idx": 4 + }, + { + "type": "header", + "text": "PartCraft: Crafting Creative Objects by Parts", + "bbox": [ + 421, + 114, + 732, + 130 + ], + "page_idx": 4 + }, + { + "type": "page_number", + "text": "5", + "bbox": [ + 774, + 114, + 784, + 126 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "parts of objects into text tokens and can recompose them in a novel way. To that end, we propose PartCraft, as depicted in Fig. 3.", + "bbox": [ + 212, + 146, + 782, + 176 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "We start by discovering the parts in a three-tier hierarchy, as detailed in Sec. 3.1. Paired with the training images $\\{x_{i}\\}_{i = 1}^{N}$ , this semantic hierarchy subsequently serves as the supervision to fine-tune a pre-trained text-to-image model, say a latent diffusion model [52], denoted as $\\{\\epsilon_{\\theta},\\tau_{\\theta},\\mathcal{E},\\mathcal{D}\\}$ , where $\\epsilon_{\\theta}$ represents the diffusion denoiser, $\\tau_{\\theta}$ the text encoder, and $\\mathcal{E} / \\mathcal{D}$ the autoencoder respectively. We adopt the textual inversion technique [25]. Concretely, we learn a set of pseudo-words $p^*$ for each part in the word embedding space with:", + "bbox": [ + 212, + 176, + 784, + 282 + ], + "page_idx": 5 + }, + { + "type": "equation", + "text": "\n$$\n\\mathcal {L} _ {l d m} = \\mathbb {E} _ {z, t, p, \\epsilon} \\left[ | | \\epsilon - \\epsilon_ {\\theta} \\left(z _ {t}, t, \\tau_ {\\theta} (y _ {p})\\right) | | _ {2} ^ {2} \\right], \\tag {1}\n$$\n", + "text_format": "latex", + "bbox": [ + 359, + 292, + 785, + 311 + ], + "page_idx": 5 + }, + { + "type": "equation", + "text": "\n$$\np ^ {*} = \\underset {p} {\\operatorname {a r g m i n}} \\mathcal {L} _ {l d m}, \\tag {2}\n$$\n", + "text_format": "latex", + "bbox": [ + 377, + 313, + 785, + 338 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "where $\\epsilon \\sim \\mathcal{N}(0,1)$ denotes the unscaled noise, $t$ is the time step, $z = \\mathcal{E}(x)$ is the latent representation of the image, $z_{t}$ is the latent noise at time $t$ , and $y_{p}$ is the text condition that includes $p$ as part of the text tokens. $\\mathcal{L}_{ldm}$ is a standard diffusion loss [30] to reconstruct the parts. As each object is composed of a set of parts, its reconstruction is achieved by the reconstruction of the associated set of parts. In other words, when all parts are reconstructed properly, it will become a valid object.", + "bbox": [ + 212, + 349, + 784, + 455 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "3.1 Unsupervised Part Discovery", + "text_level": 1, + "bbox": [ + 212, + 479, + 503, + 494 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/7cdf1b8f24fb6024e9ccc6323e592f8e95ce299405de4bc25888338d91dd2b32.jpg", + "image_caption": [ + "Fig. 4: Part discovery within a semantic hierarchy involves partitioning each image into distinct parts and forming semantic clusters across unlabeled training data." + ], + "image_footnote": [], + "bbox": [ + 215, + 515, + 475, + 627 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "To minimize the labeling cost, we develop a scalable process to reveal the underlying semantic hierarchy with parts in an unsupervised fashion. We leverage the off-the-shelf vision model for image decomposition and clustering. Specifically, given an image $x_{i}$ , we employ DINOv2 [45] to extract the feature map $F = \\{F_{i} = \\mathrm{DINOv2}(x_{i})\\}_{i}^{N}$ . We then conduct three-level hierarchical clustering (see Fig. 4): (i) At the top level, $k$ -means is applied on all patches in $F$ with $k = 2$ to separate foregrounds and backgrounds $B$ . (ii) At the middle level, $k$ -means is further ap", + "bbox": [ + 483, + 505, + 787, + 715 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "plied on all foreground patches to acquire $M$ clusters representing common parts, such as the heads of birds. (iii) At the bottom level, we further group each of the $M$ clusters as well as the background cluster $B$ into $K$ splits. Each split refers to finer meanings, such as the head of a specific bird species, or a specific background style. Lastly, each region of an image will be tagged with the corresponding cluster index. We represent these cluster tags as follows:", + "bbox": [ + 212, + 717, + 787, + 808 + ], + "page_idx": 5 + }, + { + "type": "equation", + "text": "\n$$\np = (0, k _ {0}), (1, k _ {1}), \\dots , (M, k _ {M}), \\tag {3}\n$$\n", + "text_format": "latex", + "bbox": [ + 387, + 819, + 785, + 835 + ], + "page_idx": 5 + }, + { + "type": "page_number", + "text": "6", + "bbox": [ + 215, + 114, + 228, + 126 + ], + "page_idx": 5 + }, + { + "type": "header", + "text": "KW. Ng et al.", + "bbox": [ + 271, + 114, + 367, + 128 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "where the first pair refers to the background style, and the following $M$ pairs denote the combinations of $M$ parts (e.g., head, body, wings) each associated with a specific object (e.g., sparrow), and $k \\in \\{1, \\dots, K\\}$ . This description will be used as the textual prompt in model training, such as \"a photo of a $[p]$ \". Please refer to the supplementary material for more examples of the discovered semantic hierarchy. This process also yields the segmentation mask of each $m$ -th part, which we define as $S_{m}$ .", + "bbox": [ + 212, + 146, + 782, + 252 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Motivation. While we can leverage off-the-shelf segmentation models such as VLPart [58], the robustness relies on the generalizability of the model and the part segmentation result is usually pre-defined and may be unstable for unseen domains. As a result, we rely on our feature clustering method to obtain the segmentation map, which also has a higher flexibility in choosing the number of clusters (parts).", + "bbox": [ + 212, + 252, + 784, + 342 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "3.2 Part Token Bottleneck", + "text_level": 1, + "bbox": [ + 214, + 366, + 449, + 380 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "In contrast to prior text inversion studies [25], our task requires learning a greater quantity at the same time - specifically, $(M + 1)K$ -of word tokens derived from a collection of discovered parts marked by inherent imperfections (such as partial overlap and over splitting). This makes the learning task more demanding. To enhance the learning process, we propose a neural network $f(\\cdot)$ comprising a two-layer MLP with ReLU activation:", + "bbox": [ + 212, + 392, + 782, + 482 + ], + "page_idx": 6 + }, + { + "type": "equation", + "text": "\n$$\ny _ {p} = f (e (p)), \\tag {4}\n$$\n", + "text_format": "latex", + "bbox": [ + 455, + 497, + 784, + 512 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "where $y_{p}$ will be subsequently used as the input3 to the text encoder $\\tau_{\\theta}$ and $e \\in \\mathbb{R}^{MK \\times D}$ is a learnable word embedding dictionary that maps $p$ to their respective embeddings.", + "bbox": [ + 212, + 523, + 782, + 568 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Our design demonstrates quicker convergence than directly learning the final word embeddings $e(\\cdot)$ [25] (see supplementary material). This could be attributed to the entanglement of word embeddings in the conventional design, where there is no information exchange among them during optimization. For instance, each token doesn't know they are learning for a specific part of a specific species. This lack of communication leads to lower data efficiency and slower learning. With the bottleneck $f$ , it will first project the token into a common part embedding space (e.g., head), then slightly adjust itself to adapt the fine-grained part details. It's worth noting that the conventional design is a specific instance of our approach when $f$ is an identity function.", + "bbox": [ + 212, + 569, + 784, + 720 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "3.3 Learning to Craft by Parts", + "text_level": 1, + "bbox": [ + 214, + 743, + 483, + 758 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Fine-tuning the T2I model, rather than sorely learning pseudo-words, has been shown to achieve better reconstruction of target concepts as demonstrated in [36, 53]. However, this comes with a significant training cost. Thus, we apply LoRA", + "bbox": [ + 212, + 768, + 782, + 816 + ], + "page_idx": 6 + }, + { + "type": "header", + "text": "PartCraft: Crafting Creative Objects by Parts", + "bbox": [ + 421, + 114, + 732, + 130 + ], + "page_idx": 6 + }, + { + "type": "page_number", + "text": "7", + "bbox": [ + 774, + 114, + 785, + 126 + ], + "page_idx": 6 + }, + { + "type": "page_footnote", + "text": "3 Word templates such as \"a photo of a [*]\" will be used.", + "bbox": [ + 217, + 824, + 598, + 840 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "(low-rank adaptation) [31] to the cross-attention block for efficient training. We then minimize the diffusion loss $\\mathcal{L}_{ldm}$ (Eq. (1)) to learn both pseudo-words and LoRA adapters.", + "bbox": [ + 212, + 146, + 782, + 190 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "While training with only $\\mathcal{L}_{ldm}$ , entanglement happens between parts, as evident from the attention maps in the cross-attention block of the denoiser $\\epsilon_{\\theta}$ (see Fig. 9). This entanglement arises due to the correlation between parts (e.g., a bird head code is consistently paired with a bird body code to represent the same species). To address this issue, we introduce an entropy-based attention loss as regularization:", + "bbox": [ + 212, + 191, + 784, + 282 + ], + "page_idx": 7 + }, + { + "type": "equation", + "text": "\n$$\n\\mathcal {L} _ {\\text {a t t n}} = \\mathbb {E} _ {z, t, m} \\left[ - \\left(S _ {m} \\log \\hat {A} _ {m} + (1 - S _ {m}) \\log (1 - \\hat {A} _ {m})\\right) \\right], \\tag {5}\n$$\n", + "text_format": "latex", + "bbox": [ + 303, + 294, + 784, + 311 + ], + "page_idx": 7 + }, + { + "type": "equation", + "text": "\n$$\n\\hat {A} _ {m, i, j} = \\frac {\\bar {A} _ {m , i , j}}{\\sum_ {k} \\bar {A} _ {k , i , j}}, \\quad \\bar {A} _ {m} = \\frac {1}{L} \\sum_ {l} ^ {L} A _ {l, m}, \\tag {6}\n$$\n", + "text_format": "latex", + "bbox": [ + 299, + 316, + 784, + 354 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "where $A \\in [0,1]^{M \\times HW}$ represents the cross-attention map between the $m$ -th part and the noisy latent $z_{t}$ , $L$ represents the number of specific layers to select attention maps, $\\hat{A} \\in [0,1]^{M \\times HW}$ represents the averaged and normalized cross-attention map over all parts and $S_{m} \\in \\{0,1\\}^{M \\times HW}$ serves as the mask that indicates the location of $m$ -th part. In cases where the part is not present in the image (e.g., occluded), we set both $S_{m}$ and $\\hat{A}_{m}$ as 0 to exclude them. Thus, the overall learning objective is defined as:", + "bbox": [ + 212, + 368, + 787, + 474 + ], + "page_idx": 7 + }, + { + "type": "equation", + "text": "\n$$\n\\mathcal {L} _ {\\text {t o t a l}} = \\mathcal {L} _ {l d m} + \\lambda_ {\\text {a t t n}} \\mathcal {L} _ {\\text {a t t n}}, \\tag {7}\n$$\n", + "text_format": "latex", + "bbox": [ + 405, + 488, + 784, + 503 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "where $\\lambda_{attn} = 0.01$ . We focus on attention maps at the resolution of $16 \\times 16$ where rich semantic information is captured [28]. Normalization is performed at each location to ensure that the sum of a patch location equals 1. This aims to maximize the attention of a specific part at a particular location which implicitly minimizing the attention of other parts similar to a softmax classification task. Compared to the mean-square based attention loss [3], this intuitively ensures that a part only appears once at a particular location, facilitating stronger disentanglement from other parts during the denoising operation. When generating a part for a particular location, the diffusion model $\\epsilon_{\\theta}$ should only attend to the part instead of other non-related parts.", + "bbox": [ + 212, + 513, + 787, + 667 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "4 Experiments", + "text_level": 1, + "bbox": [ + 214, + 683, + 375, + 700 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Datasets. We demonstrate our selection task on two fine-grained object datasets: CUB-200-2011 (birds) [62] which contains 5,994 training images, and Stanford Dogs [33] which contains 12,000 training images.", + "bbox": [ + 212, + 715, + 787, + 761 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Implementation. For part composition, we assess the model's ability to combine up to 4 different parts from 4 distinct species/objects. We set $M = 5$ for bird generation (head, front body/breast area, wings, legs, tail) and $M = 7$ for dog generation (forehead, eyes, mouth/nose, ears, neck, body/tail, legs). For both datasets, $K$ is set as 256, ensuring sufficient coverage of all fine-grained", + "bbox": [ + 212, + 763, + 787, + 840 + ], + "page_idx": 7 + }, + { + "type": "page_number", + "text": "8", + "bbox": [ + 217, + 114, + 228, + 126 + ], + "page_idx": 7 + }, + { + "type": "header", + "text": "KW. Ng et al.", + "bbox": [ + 271, + 114, + 367, + 128 + ], + "page_idx": 7 + }, + { + "type": "table", + "img_path": "images/0987285f288d20cfba543303676758392d729a124ed36ad26d042f7567494416.jpg", + "table_caption": [], + "table_footnote": [], + "table_body": "
MethodLearnable TokenFine TuneDisentanglementBottleneck
Textual Inversion [25]✓✗✗✗
DreamBooth [53]✓LoRA*✗✗
CustomDiffusion [36]✓K/V✗✗
Break-a-scene [3]✓LoRA*MSE✗
PartCraft (Ours)✓LoRAEq. (5)✓
", + "bbox": [ + 246, + 143, + 750, + 224 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "Table 1: Comparing our and alternative methods in design properties. *: We finetuned the added LoRA [31] adapter rather than the entire diffusion model $\\epsilon_{\\theta}$ due to resource limit. MSE is a mean-square based attention loss used in [3].", + "bbox": [ + 212, + 227, + 782, + 268 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "classes (i.e., 200 for birds and 120 for dogs). We randomly generate 500 images by sampling 500 sets of parts. For each image, we randomly replace an original part with any part from another 500 non-overlapping sets of parts. The resulting set of parts may take the form of $\\left( {0,{k}_{A}}\\right) \\left( {1,{k}_{B}}\\right) \\left( {2,{k}_{C}}\\right) \\ldots \\left( {M,{k}_{D}}\\right)$ ,representing a composition from species A,B,C,and D. Stable Diffusion v1.5 [52] is used. Please see the supplementary material for further training details.", + "bbox": [ + 212, + 332, + 784, + 422 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "We compare our method with the recent personalization methods: Textual Inversion (TI) [25], DreamBooth (DB) [53], Custom Diffusion (CD) [36], Break-a-scene (BaS) [3]. These personalization methods were designed to take single or multiple images with associated labeled objects as input. The text prompt for each image is as simple as \"a photo of $[p]$ \" where $p$ is expressed in Eq. (3), since we do not rely on complex prompts. We employ the official implementations released by the original authors for training. We summarize the main design properties of all compared methods in Tab. 1.", + "bbox": [ + 212, + 441, + 787, + 561 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "Evaluation metrics. To assess a model's ability to disentangle and composite parts, we introduce two metrics: (a) exact matching rate (EMR) and (b) cosine similarity (CoSim) between the $k$ -means embeddings of the parts of real and generated images. Utilizing the pre-trained $k$ -means from Sec. 3.1, we predict the parts of generated images. EMR quantifies how accurately the cluster index of parts of generated images matches the parts of the corresponding real images whereas CoSim measures the cosine similarity between the $k$ -means centroid vector that the part belongs to between generated and real images. These metrics assess the model's ability to follow the input parts and accurately reconstruct them, with perfect disentanglement indicated by EMR of 1 and CoSim of 1. A detailed algorithm is provided in the supplementary material. We also measure image generation quality using FID [29] to assess model performance in terms of image distribution. Additionally, we compute the average pairwise cosine similarity between CLIP [48]/DINO [11] embeddings of generated and real class-specific images following [53]. Each generated image is conditioned on the parts of the corresponding real image. This results in 5,994 generated images for birds and 12,000 generated images for dogs.", + "bbox": [ + 212, + 582, + 787, + 840 + ], + "page_idx": 8 + }, + { + "type": "header", + "text": "PartCraft: Crafting Creative Objects by Parts", + "bbox": [ + 421, + 114, + 732, + 128 + ], + "page_idx": 8 + }, + { + "type": "page_number", + "text": "9", + "bbox": [ + 774, + 114, + 784, + 126 + ], + "page_idx": 8 + }, + { + "type": "table", + "img_path": "images/4a9820926c0f178845398e9b0317c31c5961cf7e7b9b43e194301fade74a3471.jpg", + "table_caption": [], + "table_footnote": [], + "table_body": "
MethodBirds: CUB-200-2011Dogs: Stanford Dogs
FIDCLIPDINOEMRCoSimFIDCLIPDINOEMRCoSim
Textual Inversion [25]10.100.7840.6070.3050.84223.360.6520.5320.2180.754
DreamBooth [53]12.940.7750.5940.3550.85622.650.6600.5630.2750.777
Custom Diffusion [36]37.610.6940.5040.3380.83342.410.5930.4910.2530.755
Break-a-Scene [3]20.050.7420.5490.3900.85424.200.6330.5320.3000.775
PartCraft (Ours)12.860.7830.6180.4600.88216.920.6690.5730.3580.796
", + "bbox": [ + 220, + 142, + 782, + 253 + ], + "page_idx": 9 + }, + { + "type": "text", + "text": "Table 2: Quantitative comparison for part reconstruction.", + "bbox": [ + 303, + 257, + 696, + 271 + ], + "page_idx": 9 + }, + { + "type": "image", + "img_path": "images/2461d17a447063d53bc670ae77ba2c86222d1a10e0b4f5ded508227952c9bdca.jpg", + "image_caption": [ + "(a) TI" + ], + "image_footnote": [], + "bbox": [ + 217, + 279, + 330, + 450 + ], + "page_idx": 9 + }, + { + "type": "image", + "img_path": "images/525a983d3e6f635dbc54686bdefc5eba78e2170f720894cf71cbe0bfcc2c2c58.jpg", + "image_caption": [ + "(b) DB" + ], + "image_footnote": [], + "bbox": [ + 331, + 280, + 442, + 450 + ], + "page_idx": 9 + }, + { + "type": "image", + "img_path": "images/3797cf22eda737f60fc7102a73197117e5728876b1aedb06ed9568e3e2bd3579.jpg", + "image_caption": [ + "(c) CD" + ], + "image_footnote": [], + "bbox": [ + 444, + 279, + 555, + 450 + ], + "page_idx": 9 + }, + { + "type": "image", + "img_path": "images/c83aeab8a8fb6ae0ca334696b241f93777adbfb9ffd4947b627896e96b631219.jpg", + "image_caption": [ + "(d) BaS", + "Fig. 5: Visual comparison under the part reconstruction setting. All images are generated by using the original parts of respective objects." + ], + "image_footnote": [], + "bbox": [ + 558, + 280, + 669, + 450 + ], + "page_idx": 9 + }, + { + "type": "image", + "img_path": "images/76a89cfde8e5a63f6a92ff9c9eaaa58b2e37f1cb6a11a5c61dd7216815adbfb7.jpg", + "image_caption": [ + "(e) PartCraft" + ], + "image_footnote": [], + "bbox": [ + 671, + 280, + 784, + 450 + ], + "page_idx": 9 + }, + { + "type": "text", + "text": "4.1 Part Reconstruction and Image Quality Evaluation", + "text_level": 1, + "bbox": [ + 212, + 527, + 681, + 542 + ], + "page_idx": 9 + }, + { + "type": "text", + "text": "We first assess the ability of different methods to learn parts as text tokens by evaluating how well they can accurately reconstruct the parts (this also means image generation with original parts).", + "bbox": [ + 212, + 553, + 784, + 598 + ], + "page_idx": 9 + }, + { + "type": "text", + "text": "In Tab. 2, we summarize the performance of respective methods on the bird and dog generation, respectively. We highlight four observations: (i) Textual Inversion performs quite well compared to DreamBooth, CustomDiffusion, and Break-a-scene in terms of FID, CLIP, and DINO scores although did not finetune the diffusion model $\\epsilon_{\\theta}$ . This may be due to the potential risk of overfitting when fine-tuning $\\epsilon_{\\theta}$ especially when learning a vast array of new concepts with many update iterations. It is also not uncommon to carefully tune the learning rate and the training iterations in these models when fine-tuning new concepts (e.g., only 800-1000 steps of updates to learn a new concept in [3]). (ii) Nonetheless, fine-tuning the diffusion model $\\epsilon_{\\theta}$ can help improve the ability to follow prompts as shown by increased EMR and CoSim scores (e.g., EMR of at least $5\\%$ in DreamBooth). (iii) Break-a-scene has a better ability to reconstruct the parts as shown by EMR and CoSim, this is due to the attention loss explicitly forcing the parts to focus on the respective semantic region. (iv) PartCraft achieves the best performance in DINO, EMR, and CoSim scores (e.g., $7\\%$ better in EMR compared to Break-a-scene). This indicates that not only does our", + "bbox": [ + 212, + 598, + 785, + 840 + ], + "page_idx": 9 + }, + { + "type": "page_number", + "text": "10", + "bbox": [ + 217, + 114, + 235, + 126 + ], + "page_idx": 9 + }, + { + "type": "header", + "text": "KW. Ng et al.", + "bbox": [ + 271, + 114, + 367, + 128 + ], + "page_idx": 9 + }, + { + "type": "image", + "img_path": "images/cb410159b384692c2139554d69c2674923f06e73f46b6f4d561006ff1e811c7f.jpg", + "image_caption": [ + "Fig. 6: Quantitative comparisons of part composition in terms of EMR and CoSim." + ], + "image_footnote": [], + "bbox": [ + 217, + 143, + 785, + 256 + ], + "page_idx": 10 + }, + { + "type": "text", + "text": "image-generation ability perform comparably well with textual inversion, but PartCraft is also able to disentangle the parts learning so that it can follow the prompt instructions more accurately to generate the parts in a cohort. In Fig. 5, we present generated images from different methods, with CustomDiffusion exhibiting high-contrast images, possibly due to unconstrained fine-tuning on the cross-attention components $K / V$ and resulting in worse FID scores.", + "bbox": [ + 212, + 294, + 787, + 386 + ], + "page_idx": 10 + }, + { + "type": "text", + "text": "4.2 Part Composition Evaluation", + "text_level": 1, + "bbox": [ + 215, + 406, + 503, + 422 + ], + "page_idx": 10 + }, + { + "type": "text", + "text": "In this section, we assess the part composition ability of different methods. In this experiment, we generate the image by mixing different parts from different species. Our findings, as shown in Fig. 6, can be summarized as follows: (i) As the number of composited parts increases, EMR and CoSim decrease, reflecting the challenge of composing multiple diverse parts. (ii) Break-a-scene and PartCraft achieve notably higher EMR and CoSim scores, thanks to disentanglement through attention loss minimization. (iii) PartCraft outperforms Break-a-scene significantly by token bottleneck and tailored attention loss.", + "bbox": [ + 212, + 430, + 787, + 551 + ], + "page_idx": 10 + }, + { + "type": "text", + "text": "In Fig. 7, we visualize the results of composing 4 different parts. While all images appear realistic, most methods struggle to assemble all 4 parts. For instance, Break-a-scene missed out on the flurry body of kerry blue terrier (right-most column). In contrast, our methods successfully combine 4 different parts from 4 different species, demonstrating the superior ability of our approach to part composition. We also visualize additional examples of our method in the supplementary material.", + "bbox": [ + 212, + 551, + 789, + 656 + ], + "page_idx": 10 + }, + { + "type": "text", + "text": "Furthermore, we explore the versatility of the adapted model by generating images with simple styles such as pencil drawing. While most methods successfully incorporate specific styles into the generated image, Custom Diffusion often fails to do so, possibly due to the unconstrained fine-tuning of the cross-attention components $K / V$ .", + "bbox": [ + 212, + 657, + 787, + 734 + ], + "page_idx": 10 + }, + { + "type": "text", + "text": "4.3 Ablation Studies", + "text_level": 1, + "bbox": [ + 215, + 753, + 401, + 768 + ], + "page_idx": 10 + }, + { + "type": "text", + "text": "Component analysis. In Fig. 8, we evaluate the effect of our proposed components (token bottleneck and attention loss) on creating novel bird species. (i) Removing the bottleneck outlined in Eq. (4) degrades the generation quality as evidenced by a higher FID score $(12.86 \\rightarrow 16.36)$ even though both EMR", + "bbox": [ + 212, + 779, + 787, + 840 + ], + "page_idx": 10 + }, + { + "type": "header", + "text": "PartCraft: Crafting Creative Objects by Parts", + "bbox": [ + 421, + 114, + 732, + 130 + ], + "page_idx": 10 + }, + { + "type": "page_number", + "text": "11", + "bbox": [ + 767, + 114, + 782, + 126 + ], + "page_idx": 10 + }, + { + "type": "image", + "img_path": "images/258f2c1e1980ccfb10ad38c91986c23f8aabb54eaea0a085f6905dca02ef4683.jpg", + "image_caption": [ + "Fig. 7: Visual comparison on 4-species (specified on the top row) mixed generation. The last column indicates generated images with different styles (i.e., $DSLR$ , Van Gogh, Oil Painting, Pencil Drawing)." + ], + "image_footnote": [], + "bbox": [ + 217, + 142, + 785, + 631 + ], + "page_idx": 11 + }, + { + "type": "text", + "text": "and CoSim remain. (ii) By replacing our $\\mathcal{L}_{attn}$ with the MSE loss as proposed in [3], we observe significant deterioration in both EMR and CoSim. (iii) Finally, incorporating both the projector and our attention loss performs the best. This improvement highlights the necessity of incorporating interactions between multiple parts to achieve more effective part disentanglement and optimization.", + "bbox": [ + 212, + 726, + 787, + 803 + ], + "page_idx": 11 + }, + { + "type": "text", + "text": "Cross-attention visualization. Our attention loss plays a crucial role in token disentanglement. We demonstrate the impact of this loss in Fig. 9, where we", + "bbox": [ + 212, + 809, + 787, + 840 + ], + "page_idx": 11 + }, + { + "type": "page_number", + "text": "12", + "bbox": [ + 217, + 114, + 235, + 126 + ], + "page_idx": 11 + }, + { + "type": "header", + "text": "KW. Ng et al.", + "bbox": [ + 271, + 113, + 367, + 128 + ], + "page_idx": 11 + }, + { + "type": "image", + "img_path": "images/d489f73860ad7af2326e6c5eff7b9f787e900d76288dc443dfe88408a3be9928.jpg", + "image_caption": [ + "Fig.9: Cross-attention map of each part (top) without and (bottom) with our attention loss." + ], + "image_footnote": [], + "bbox": [ + 217, + 143, + 496, + 233 + ], + "page_idx": 12 + }, + { + "type": "image", + "img_path": "images/c6d1fbb7d14a0a6f64c8532f51b664666b768c548fc6106232cf540a2a5d45c8.jpg", + "image_caption": [ + "Fig.8: Ablation on our part token bottleneck and attention loss under the part composition on CUB-200-2011 birds." + ], + "image_footnote": [], + "bbox": [ + 504, + 161, + 787, + 219 + ], + "page_idx": 12 + }, + { + "type": "image", + "img_path": "images/0880ca23d405fb532d96c74a65064acb8cd1bf1482e8482d2a97615297a63633.jpg", + "image_caption": [ + "(a) Textual Inversion" + ], + "image_footnote": [], + "bbox": [ + 225, + 323, + 395, + 428 + ], + "page_idx": 12 + }, + { + "type": "image", + "img_path": "images/e1c58d70a21aa6c21e133b0b69ca32b989a2437b162aab0c753b51860fc799c3.jpg", + "image_caption": [ + "(b) Break-a-scene" + ], + "image_footnote": [], + "bbox": [ + 418, + 323, + 588, + 429 + ], + "page_idx": 12 + }, + { + "type": "image", + "img_path": "images/313f6382013f8b85457e6d24af5d5f9acb802c9bc491b7cbddfa00cfce25161e.jpg", + "image_caption": [ + "(c) PartCraft (Ours)", + "Fig. 10: 2D tSNE [39] projection of word embeddings. Different colors represent different common parts. (correspond to the segmentation mask in Fig. 4)." + ], + "image_footnote": [], + "bbox": [ + 609, + 323, + 781, + 430 + ], + "page_idx": 12 + }, + { + "type": "text", + "text": "observe significantly enhanced disentanglement after explicitly guiding attention to focus on distinct semantic regions.", + "bbox": [ + 212, + 535, + 784, + 566 + ], + "page_idx": 12 + }, + { + "type": "text", + "text": "Part word embedding space. We visualize the word embeddings of learned tokens of Textual Inversion [25], Break-a-scene [3] and our PartCraft for birds generation (CUB-200-2011 [62]) through tSNE [39] in Fig. 10. In our PartCraft, the word embeddings are the projected embeddings through Eq. (4). We can see that our projected version has a better semantic meaning such that the part embeddings are clustered together by their semantic meaning (e.g., head). We believe this is one of the reasons that our PartCraft outperforms previous methods in which we can compose all parts seamlessly yet with higher quality.", + "bbox": [ + 212, + 574, + 787, + 696 + ], + "page_idx": 12 + }, + { + "type": "text", + "text": "Transferability for Creativity. (i) In Fig. 11, we demonstrate that not only it can compose parts within the training domain (e.g., birds), but it can also transfer the learned parts to and combine with other domains (e.g., cat). This enables the creation of unique combinations, such as a cat with a dog's ear. (ii) Leveraging the prior knowledge embedded in Stable Diffusion, PartCraft can also repurpose learned parts for creative image generation. An example of this is the generation of a bird-shaped robot adorned with various parts. These examples showcase PartCraft's immense potential for diverse and limitless creative applications. Please see the supplementary material for more examples.", + "bbox": [ + 212, + 703, + 787, + 840 + ], + "page_idx": 12 + }, + { + "type": "header", + "text": "PartCraft: Crafting Creative Objects by Parts", + "bbox": [ + 421, + 114, + 732, + 128 + ], + "page_idx": 12 + }, + { + "type": "page_number", + "text": "13", + "bbox": [ + 767, + 114, + 785, + 126 + ], + "page_idx": 12 + }, + { + "type": "image", + "img_path": "images/1fc2c2041a4f1456f00ab034420aaa44a14a9f89290790e8b24d13acb8381d53.jpg", + "image_caption": [ + "Fig. 11: (Top): Using learned parts to modify the property of other domains such as cat, and lion with prompt such as \"A cat with [beagle's ear)\". (Bottom) We can also repurpose learned parts for creative image generation using prior knowledge in Stable Diffusion with prompt such as \"A robot designed inspired by [red header woodpecker's head] and [blue jay's body)\"." + ], + "image_footnote": [], + "bbox": [ + 235, + 143, + 767, + 305 + ], + "page_idx": 13 + }, + { + "type": "text", + "text": "5 Conclusion", + "text_level": 1, + "bbox": [ + 215, + 397, + 357, + 414 + ], + "page_idx": 13 + }, + { + "type": "text", + "text": "We propose a new way of control in generative AI. Instead of text or sketch, we \"select\" desired parts to create an object. We addressed the challenge of learning parts in T2I models by introducing a customized attention loss. This loss serves a dual purpose: to ensure parts are at the right location and to ensure each location is occupied by not more than one part. This greatly improves the part disentanglement. We further employ a non-linear bottleneck encoder to improve generation fidelity. Our model, PartCraft, can seamlessly compose different parts from different objects, creating objects that do not exist yet holistically correct and plausible objects by mixing them. Extensive experiments demonstrated PartCraft's superior performance in both qualitative and quantitative evaluation. Moreover, the learned parts demonstrate strong transferability. We hope that our PartCraft will empower artists, designers, and enthusiasts to bring the creations of their dreams to reality.", + "bbox": [ + 212, + 425, + 787, + 621 + ], + "page_idx": 13 + }, + { + "type": "text", + "text": "6 Limitations and Future Works", + "text_level": 1, + "bbox": [ + 215, + 643, + 545, + 659 + ], + "page_idx": 13 + }, + { + "type": "text", + "text": "It is worth noting that the accuracy of obtained parts may be affected by using a self-supervised pre-trained feature extractor, incorporating encoders [64] may help. We also observed challenges in composing relatively small parts, like tails and legs. Additionally, we are also exploring cross-domain generation, i.e., combining learned parts from different datasets to create objects with even more diverse parts. For instance, we can merge non-rigid parts (e.g., parts from quadrupled animals) with rigid parts (e.g., parts from cars/airplanes) and form a creative structure (e.g., a car that has horse legs instead of wheels). This not only further improves the applicability of PartCraft but also serves as a stepping stone to creative generation, as generative AI progresses, continually expanding the limits of achievable creativity and artistic expression.", + "bbox": [ + 212, + 672, + 787, + 840 + ], + "page_idx": 13 + }, + { + "type": "page_number", + "text": "14", + "bbox": [ + 217, + 114, + 235, + 126 + ], + "page_idx": 13 + }, + { + "type": "header", + "text": "KW. Ng et al.", + "bbox": [ + 271, + 114, + 367, + 128 + ], + "page_idx": 13 + }, + { + "type": "text", + "text": "Acknowledgements", + "text_level": 1, + "bbox": [ + 215, + 143, + 401, + 162 + ], + "page_idx": 14 + }, + { + "type": "text", + "text": "We extend our special thanks to Jia Wei Sii for her help in creating figures and discussing the main concept. We are also grateful to Ruoyi Du, Zhiyu Qu, Chee Seng Chan, and the reviewers for their fruitful comments and corrections on our draft, methodology, and experiments.", + "bbox": [ + 212, + 178, + 787, + 239 + ], + "page_idx": 14 + }, + { + "type": "text", + "text": "References", + "text_level": 1, + "bbox": [ + 215, + 263, + 321, + 277 + ], + "page_idx": 14 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "1. Alaluf, Y., Richardson, E., Metzer, G., Cohen-Or, D.: A neural space-time representation for text-to-image personalization. arXiv preprint arXiv:2305.15391 (2023)", + "2. Amir, S., Gandelsman, Y., Bagon, S., Dekel, T.: On the effectiveness of vit features as local semantic descriptors. In: ECCV (2022)", + "3. Avrahami, O., Aberman, K., Fried, O., Cohen-Or, D., Lischinski, D.: Break-a-scene: Extracting multiple concepts from a single image. In: SIGGRAPH Asia (2023)", + "4. Avrahami, O., Fried, O., Lischinski, D.: Blended latent diffusion. ACM TOG (2023)", + "5. Avrahami, O., Hayes, T., Gafni, O., Gupta, S., Taigman, Y., Parikh, D., Lischinski, D., Fried, O., Yin, X.: Spatext: Spatio-textual representation for controllable image generation. In: CVPR (2023)", + "6. Balaji, Y., Nah, S., Huang, X., Vahdat, A., Song, J., Kreis, K., Aittala, M., Aila, T., Laine, S., Catanzaro, B., et al.: edifi: Text-to-image diffusion models with an ensemble of expert denoisers. arXiv preprint arXiv:2211.01324 (2022)", + "7. Bar-Tal, O., Yariv, L., Lipman, Y., Dekel, T.: Multidiffusion: Fusing diffusion paths for controlled image generation. In: ICML (2023)", + "8. Biederman, I.: Recognition-by-components: a theory of human image understanding. Psychological review (1987)", + "9. Bonnardel, N., Marmèche, E.: Towards supporting evocation processes in creative design: A cognitive approach. International Journal of Human-Computer Studies (2005)", + "10. Brooks, T., Holynski, A., Efros, A.A.: Instructpix2pix: Learning to follow image editing instructions. In: CVPR (2023)", + "11. Caron, M., Touvron, H., Misra, I., Jégou, H., Mairal, J., Bojanowski, P., Joulin, A.: Emerging properties in self-supervised vision transformers. In: ICCV (2021)", + "12. Cetinic, E., She, J.: Understanding and creating art with ai: Review and outlook. ACM Transactions on Multimedia Computing, Communications, and Applications (2022)", + "13. Chefer, H., Alaluf, Y., Vinker, Y., Wolf, L., Cohen-Or, D.: Attend-and-excite: Attention-based semantic guidance for text-to-image diffusion models. ACM TOG (2023)", + "14. Chen, C., Li, O., Tao, D., Barnett, A., Rudin, C., Su, J.K.: This looks like that: deep learning for interpretable image recognition. NeurIPS (2019)", + "15. Chen, M., Laina, I., Vedaldi, A.: Training-free layout control with cross-attention guidance (2024)", + "16. Cintas, C., Das, P., Quanz, B., Tadesse, G.A., Speakman, S., Chen, P.Y.: Towards creativity characterization of generative models via group-based subset scanning. In: IJCAI (2022)", + "17. Couairon, G., Verbeek, J., Schwenk, H., Cord, M.: Diffedit: Diffusion-based semantic image editing with mask guidance. In: ICLR (2023)" + ], + "bbox": [ + 217, + 295, + 784, + 839 + ], + "page_idx": 14 + }, + { + "type": "header", + "text": "PartCraft: Crafting Creative Objects by Parts", + "bbox": [ + 421, + 114, + 730, + 128 + ], + "page_idx": 14 + }, + { + "type": "page_number", + "text": "15", + "bbox": [ + 767, + 114, + 782, + 126 + ], + "page_idx": 14 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "18. Das, P., Quanz, B., Chen, P.Y., Ahn, J.w., Shah, D.: Toward a neuro-inspired creative decoder. In: IJCAI (2020)", + "19. Ding, M., Yang, Z., Hong, W., Zheng, W., Zhou, C., Yin, D., Lin, J., Zou, X., Shao, Z., Yang, H., et al.: Cogview: Mastering text-to-image generation via transformers. In: NeurIPS (2021)", + "20. Ding, M., Zheng, W., Hong, W., Tang, J.: Cogview2: Faster and better text-to-image generation via hierarchical transformers. In: NeurIPS. vol. 35 (2022)", + "21. Elgammal, A., Liu, B., Elhoseiny, M., Mazzone, M.: Can: Creative adversarial networks generating \"art\" by learning about styles and deviating from style norms. In: ICCC (2017)", + "22. Felzenszwalb, P.F., Huttenlocher, D.P.: Pictorial structures for object recognition. IJCV (2005)", + "23. Feng, W., He, X., Fu, T.J., Jampani, V., Akula, A.R., Narayana, P., Basu, S., Wang, X.E., Wang, W.Y.: Training-free structured diffusion guidance for compositional text-to-image synthesis. In: ICLR (2022)", + "24. Gafni, O., Polyak, A., Ashual, O., Sheynin, S., Parikh, D., Taigman, Y.: Make-ascene: Scene-based text-to-image generation with human priors. In: ECCV (2022)", + "25. Gal, R., Alaluf, Y., Atzmon, Y., Patashnik, O., Bermano, A.H., Chechik, G., Cohen-Or, D.: An image is worth one word: Personalizing text-to-image generation using textual inversion. In: ICLR (2023)", + "26. Ge, S., Goswami, V., Zitnick, C.L., Parikh, D.: Creative sketch generation. In: ICLR (2021)", + "27. He, J., Chen, J., Lin, M.X., Yu, Q., Yuille, A.L.: Compositor: Bottom-up clustering and compositing for robust part and object segmentation. In: CVPR (2023)", + "28. Hertz, A., Mokady, R., Tenenbaum, J., Aberman, K., Pritch, Y., Cohen-Or, D.: Prompt-to-prompt image editing with cross attention control. In: ICLR (2023)", + "29. Heusel, M., Ramsauer, H., Unterthiner, T., Nessler, B., Hochreiter, S.: Gans trained by a two time-scale update rule converge to a local nash equilibrium. NeurIPS 30 (2017)", + "30. Ho, J., Jain, A., Abbeel, P.: Denoising diffusion probabilistic models. In: NeurIPS (2020)", + "31. Hu, E.J., Shen, Y., Wallis, P., Allen-Zhu, Z., Li, Y., Wang, S., Wang, L., Chen, W.: LoRA: Low-rank adaptation of large language models. In: ICLR (2022)", + "32. Kawar, B., Zada, S., Lang, O., Tov, O., Chang, H., Dekel, T., Mosseri, I., Irani, M.: Imagic: Text-based real image editing with diffusion models. In: CVPR (2023)", + "33. Khosla, A., Jayadevaprakash, N., Yao, B., Li, F.F.: Novel dataset for fine-grained image categorization: Stanford dogs. In: CVPRW (2011)", + "34. Kim, Y., Lee, J., Kim, J.H., Ha, J.W., Zhu, J.Y.: Dense text-to-image generation with attention modulation. In: ICCV (2023)", + "35. Krause, J., Jin, H., Yang, J., Fei-Fei, L.: Fine-grained recognition without part annotations. In: CVPR (2015)", + "36. Kumari, N., Zhang, B., Zhang, R., Shechtman, E., Zhu, J.Y.: Multi-concept customization of text-to-image diffusion. In: CVPR (2023)", + "37. Li, B., Qi, X., Lukasiewicz, T., Torr, P.: Controllable text-to-image generation. In: NeurIPS (2019)", + "38. Li, Y., Liu, H., Wu, Q., Mu, F., Yang, J., Gao, J., Li, C., Lee, Y.J.: Gligen: Open-set grounded text-to-image generation. In: CVPR (2023)", + "39. Van der Maaten, L., Hinton, G.: Visualizing data using t-sne. Journal of Machine Learning Research (2008)" + ], + "bbox": [ + 215, + 146, + 784, + 839 + ], + "page_idx": 15 + }, + { + "type": "page_number", + "text": "16", + "bbox": [ + 217, + 114, + 235, + 126 + ], + "page_idx": 15 + }, + { + "type": "header", + "text": "KW. Ng et al.", + "bbox": [ + 271, + 114, + 367, + 128 + ], + "page_idx": 15 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "40. Mo, S., Mu, F., Lin, K.H., Liu, Y., Guan, B., Li, Y., Zhou, B.: Freecontrol: Training-free spatial control of any text-to-image diffusion model with any condition. arXiv preprint arXiv:2312.07536 (2023)", + "41. Mokady, R., Hertz, A., Aberman, K., Pritch, Y., Cohen-Or, D.: Null-text inversion for editing real images using guided diffusion models. In: CVPR (2023)", + "42. Nagai, Y., Taura, T., Mukai, F.: Concept blending and dissimilarity: factors for creative concept generation process. Design studies (2009)", + "43. Nichol, A., Dhariwal, P., Ramesh, A., Shyam, P., Mishkin, P., McGrew, B., Sutskever, I., Chen, M.: Glide: Towards photorealistic image generation and editing with text-guided diffusion models. arXiv preprint arXiv:2112.10741 (2021)", + "44. Nobari, A.H., Rashad, M.F., Ahmed, F.: Creativegan: Editing generative adversarial networks for creative design synthesis. arXiv preprint arXiv:2103.06242 (2021)", + "45. Oquab, M., Darcet, T., Moutakanni, T., Vo, H., Szafraniec, M., Khalidov, V., Fernandez, P., Haziza, D., Massa, F., El-Nouby, A., Assran, M., Ballas, N., Galuba, W., Howes, R., Huang, P.Y., Li, S.W., Misra, I., Rabbat, M., Sharma, V., Synnaeve, G., Xu, H., Jegou, H., Mairal, J., Labatut, P., Joulin, A., Bojanowski, P.: Dinov2: Learning robust visual features without supervision (2023)", + "46. Qiao, T., Zhang, J., Xu, D., Tao, D.: Mirrorgan: Learning text-to-image generation by redescription. In: CVPR (2019)", + "47. Qu, Z., Xiang, T., Song, Y.Z.: Sketchdreamer: Interactive text-augmented creative sketch ideation. In: BMVC (2023)", + "48. Radford, A., Kim, J.W., Hallacy, C., Ramesh, A., Goh, G., Agarwal, S., Sastry, G., Askell, A., Mishkin, P., Clark, J., et al.: Learning transferable visual models from natural language supervision. In: ICML (2021)", + "49. Ramesh, A., Dhariwal, P., Nichol, A., Chu, C., Chen, M.: Hierarchical text-conditional image generation with clip latents. arXiv preprint arXiv:2204.06125 (2022)", + "50. Ren, T., Liu, S., Zeng, A., Lin, J., Li, K., Cao, H., Chen, J., Huang, X., Chen, Y., Yan, F., Zeng, Z., Zhang, H., Li, F., Yang, J., Li, H., Jiang, Q., Zhang, L.: Grounded sam: Assembling open-world models for diverse visual tasks. arXiv preprint arXiv:2401.14159 (2024)", + "51. Richardson, E., Goldberg, K., Alaluf, Y., Cohen-Or, D.: Conceptlab: Creative generation using diffusion prior constraints. arXiv preprint arXiv:2308.02669 (2023)", + "52. Rombach, R., Blattmann, A., Lorenz, D., Esser, P., Ommer, B.: High-resolution image synthesis with latent diffusion models. In: CVPR (2022)", + "53. Ruiz, N., Li, Y., Jampani, V., Pritch, Y., Rubinstein, M., Aberman, K.: Dreambooth: Fine tuning text-to-image diffusion models for subject-driven generation. In: CVPR (2023)", + "54. Runco, M.A., Jaeger, G.J.: The standard definition of creativity. Creativity research journal (2012)", + "55. Saharia, C., Chan, W., Saxena, S., Li, L., Whang, J., Denton, E.L., Ghasemipour, K., Gontijo Lopes, R., Karagol Ayan, B., Salimans, T., et al.: Photorealistic text-to-image diffusion models with deep language understanding. In: NeurIPS. vol. 35 (2022)", + "56. Sbai, O., Elhoseiny, M., Bordes, A., LeCun, Y., Couprie, C.: Design: Design inspiration from generative networks. In: ECCVW (2019)", + "57. Sun, P., Chen, S., Luo, P.: Grounded segment anything: From objects to parts. https://github.com/Cheems-Seminar/grounded-segment-any-parts (2023)", + "58. Sun, P., Chen, S., Zhu, C., Xiao, F., Luo, P., Xie, S., Yan, Z.: Going denser with open-vocabulary part segmentation. arXiv preprint arXiv:2305.11173 (2023)" + ], + "bbox": [ + 212, + 146, + 787, + 840 + ], + "page_idx": 16 + }, + { + "type": "header", + "text": "PartCraft: Crafting Creative Objects by Parts", + "bbox": [ + 421, + 114, + 730, + 128 + ], + "page_idx": 16 + }, + { + "type": "page_number", + "text": "17", + "bbox": [ + 767, + 114, + 784, + 126 + ], + "page_idx": 16 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "59. Tao, M., Tang, H., Wu, F., Jing, X.Y., Bao, B.K., Xu, C.: Df-gan: A simple and effective baseline for text-to-image synthesis. In: CVPR (2022)", + "60. Vinker, Y., Voynov, A., Cohen-Or, D., Shamir, A.: Concept decomposition for visual exploration and inspiration. In: SIGGRAPH Asia (2023)", + "61. Voynov, A., Chu, Q., Cohen-Or, D., Aberman, K.: P+: Extended textual conditioning in text-to-image generation. arXiv preprint arXiv:2303.09522 (2023)", + "62. Wah, C., Branson, S., Welinder, P., Perona, P., Belongie, S.: The Caltech-UCSD Birds-200-2011 Dataset (2011)", + "63. Wang, X., Darrell, T., Rambhatla, S.S., Girdhar, R., Misra, I.: Instancediffusion: Instance-level control for image generation. arXiv preprint arXiv:2402.03290 (2024)", + "64. Wei, Y., Zhang, Y., Ji, Z., Bai, J., Zhang, L., Zuo, W.: Elite: Encoding visual concepts into textual embeddings for customized text-to-image generation. In: ICCV (2023)", + "65. Wilkenfeld, M.J., Ward, T.B.: Similarity and emergence in conceptual combination. Journal of Memory and Language (2001)", + "66. Xie, J., Li, Y., Huang, Y., Liu, H., Zhang, W., Zheng, Y., Shou, M.Z.: Boxdiff: Text-to-image synthesis with training-free box-constrained diffusion. In: ICCV (2023)", + "67. Xu, T., Zhang, P., Huang, Q., Zhang, H., Gan, Z., Huang, X., He, X.: Attngan: Fine-grained text to image generation with attentional generative adversarial networks. In: CVPR (2018)", + "68. Yang, B., Gu, S., Zhang, B., Zhang, T., Chen, X., Sun, X., Chen, D., Wen, F.: Paint by example: Exemplar-based image editing with diffusion models. In: CVPR (2023)", + "69. Yang, Z., Wang, J., Gan, Z., Li, L., Lin, K., Wu, C., Duan, N., Liu, Z., Liu, C., Zeng, M., et al.: Reco: Region-controlled text-to-image generation. In: CVPR (2023)", + "70. Yin, G., Liu, B., Sheng, L., Yu, N., Wang, X., Shao, J.: Semantics disentangling for text-to-image generation. In: CVPR (2019)", + "71. Zhang, L., Rao, A., Agrawala, M.: Adding conditional control to text-to-image diffusion models. In: ICCV (2023)", + "72. Zhu, M., Pan, P., Chen, W., Yang, Y.: Dm-gan: Dynamic memory generative adversarial networks for text-to-image synthesis. In: CVPR (2019)" + ], + "bbox": [ + 215, + 146, + 784, + 575 + ], + "page_idx": 17 + }, + { + "type": "page_number", + "text": "18", + "bbox": [ + 217, + 114, + 235, + 126 + ], + "page_idx": 17 + }, + { + "type": "header", + "text": "KW. Ng et al.", + "bbox": [ + 271, + 114, + 367, + 128 + ], + "page_idx": 17 + } +] \ No newline at end of file diff --git a/2024/PartCraft_ Crafting Creative Objects by Parts/4747571e-3f49-4aac-b94c-924e39815a4f_model.json b/2024/PartCraft_ Crafting Creative Objects by Parts/4747571e-3f49-4aac-b94c-924e39815a4f_model.json new file mode 100644 index 0000000000000000000000000000000000000000..a0cd1805ac6c5800723e0508bf04a88730b4923d --- /dev/null +++ b/2024/PartCraft_ Crafting Creative Objects by Parts/4747571e-3f49-4aac-b94c-924e39815a4f_model.json @@ -0,0 +1,2579 @@ +[ + [ + { + "type": "title", + "bbox": [ + 0.237, + 0.141, + 0.768, + 0.164 + ], + "angle": 0, + "content": "PartCraft: Crafting Creative Objects by Parts" + }, + { + "type": "text", + "bbox": [ + 0.258, + 0.189, + 0.744, + 0.207 + ], + "angle": 0, + "content": "Kam Woh \\(\\mathrm{Ng}^{1}\\), Xiatian Zhu\\(^{1,2}\\), Yi-Zhe Song\\(^{1}\\), and Tao Xiang\\(^{1}\\)" + }, + { + "type": "text", + "bbox": [ + 0.336, + 0.217, + 0.697, + 0.232 + ], + "angle": 0, + "content": "\\(^{1}\\) CVSSP, University of Surrey, United Kingdom" + }, + { + "type": "text", + "bbox": [ + 0.364, + 0.232, + 0.64, + 0.246 + ], + "angle": 0, + "content": "Surrey Institute for People-Centred AI" + }, + { + "type": "text", + "bbox": [ + 0.305, + 0.246, + 0.7, + 0.26 + ], + "angle": 0, + "content": "{kamwoh.ng, xiatian.zhu, y song, t.xiang}@surrey.ac.uk" + }, + { + "type": "text", + "bbox": [ + 0.262, + 0.296, + 0.744, + 0.56 + ], + "angle": 0, + "content": "Abstract. This paper propels creative control in generative visual AI by allowing users to \"select\". Departing from traditional text or sketch-based methods, we for the first time allow users to choose visual concepts by parts for their creative endeavors. The outcome is fine-grained generation that precisely captures selected visual concepts, ensuring a holistically faithful and plausible result. To achieve this, we first parse objects into parts through unsupervised feature clustering. Then, we encode parts into text tokens and introduce an entropy-based normalized attention loss that operates on them. This loss design enables our model to learn generic prior topology knowledge about object's part composition, and further generalize to novel part compositions to ensure the generation looks holistically faithful. Lastly, we employ a bottleneck encoder to project the part tokens. This not only enhances fidelity but also accelerates learning, by leveraging shared knowledge and facilitating information exchange among instances. Visual results in the paper and supplementary material showcase the compelling power of PartCraft in crafting highly customized, innovative creations, exemplified by the \"charming\" and creative birds in Fig. 1. Code is released at https://github.com/kamwoh/partcraft." + }, + { + "type": "text", + "bbox": [ + 0.262, + 0.572, + 0.741, + 0.587 + ], + "angle": 0, + "content": "Keywords: Part Composition \\(\\cdot\\) Controllable Text-to-image Generation" + }, + { + "type": "title", + "bbox": [ + 0.216, + 0.613, + 0.377, + 0.629 + ], + "angle": 0, + "content": "1 Introduction" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.644, + 0.788, + 0.734 + ], + "angle": 0, + "content": "Humans are creators; AI, on the other hand, hallucinates. Creativity, arguably, is the very force driving humanity forward. Recently, generative AI has garnered considerable attention for its perceived \"creativity\" [6, 19, 20, 43, 47, 49, 52, 55]. Despite its promise, the challenge of control has swiftly surfaced - how can humans infuse their creativity into the generation process and regulate the extent to which AI hallucinates?" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.735, + 0.788, + 0.842 + ], + "angle": 0, + "content": "Creativity starts with an idea. The immediate challenge is how to articulate that idea and integrate it with generative AI. Text is the most commonly employed medium. For instance, imagine being a bird enthusiast wanting to craft the most unique bird akin to Fig. 1. The go-to approach would be furnishing Stable Diffusion [52] (or an equivalent model) with the following textual prompt (the idea): \"generate a bird with the head of \\( X \\), wings of \\( Y \\), and body of \\( Z \\)\". While you might be presented with remarkably looking birds like Fig. 1(a), they" + } + ], + [ + { + "type": "page_number", + "bbox": [ + 0.218, + 0.116, + 0.228, + 0.127 + ], + "angle": 0, + "content": "2" + }, + { + "type": "header", + "bbox": [ + 0.272, + 0.115, + 0.369, + 0.129 + ], + "angle": 0, + "content": "KW. Ng et al." + }, + { + "type": "image", + "bbox": [ + 0.219, + 0.144, + 0.785, + 0.322 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.214, + 0.332, + 0.788, + 0.499 + ], + "angle": 0, + "content": "Fig. 1: People often form creative concepts based on existing ones [9, 42, 54, 65]. For instance, a bird enthusiast may want to craft a unique bird with different parts (e.g., heads, bodies, and wings) from common bird types (e.g., blue jay, cardinal and song sparrow). (a) Using text prompts in T2I models often results in a lack of control and deviation from the intended details, especially those visual details that are difficult to describe. (b) While sketching is a direct way, not everyone possesses the ability to sketch, particularly in intricate detail. (c) Even with reference images, existing methods (e.g., DreamBooth [53]) did not consider learning object parts, thus cannot generate with desired parts. (d) Using additional control is even cumbersome, requiring too many inputs! (e) We aim to create an object by simply selecting desired parts. PartCraft learns from visual examples to generate the object with a faithful holistic structure, seamlessly integrating the chosen parts into a natural and coherent entity." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.518, + 0.788, + 0.578 + ], + "angle": 0, + "content": "may bear little resemblance to the envisioned concept. Recent literature suggests that a swift sketch could serve as a viable alternative [40, 63, 71], providing fine-grained shape control. However, the caveat is that not everyone possesses the ability to sketch, particularly in intricate detail." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.581, + 0.788, + 0.686 + ], + "angle": 0, + "content": "In this paper, our primary focus revolves around addressing the issue of \"control\" in generative AI. We advance by introducing fine-grained control into the generative process, inviting you to \"select\". While this selection mechanism might seem modest at first, it closely mirrors the human creative process, where new concepts often emerge from existing ones [9,42,54,65]. Recall those moments when you desired an \"ideal\" pair of shoes with selected features from different pairs, or when you aimed to get creative with a cat (for that matter!)." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.688, + 0.788, + 0.808 + ], + "angle": 0, + "content": "It follows that rather than relying on writing (text) or drawing (sketch), all that is required is to choose the distinct visual concepts you specifically desire in your creative endeavor. Our model then ensures that all selected concepts are seamlessly and precisely composed into a faithful novel object in the final generation, without resorting to additional control such as bounding boxes [7,38, 40,66,71]. To illustrate with the \"unique bird\" example once more, our approach literally involves selecting the head of X, wings of Y, and body of Z! (see Fig. 1(e) and 2)." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.81, + 0.788, + 0.841 + ], + "angle": 0, + "content": "Our solution is intuitive and centers around the well-studied computer vision concept of objects and their parts [8,14,22,27,35]. The challenge then boils down" + } + ], + [ + { + "type": "header", + "bbox": [ + 0.423, + 0.115, + 0.733, + 0.13 + ], + "angle": 0, + "content": "PartCraft: Crafting Creative Objects by Parts" + }, + { + "type": "page_number", + "bbox": [ + 0.775, + 0.116, + 0.787, + 0.127 + ], + "angle": 0, + "content": "3" + }, + { + "type": "image", + "bbox": [ + 0.218, + 0.143, + 0.789, + 0.337 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.214, + 0.346, + 0.788, + 0.389 + ], + "angle": 0, + "content": "Fig. 2: Two sets of images were generated from their original parts (sources A and B). We can integrate a specific part (e.g., body, head, or even background (BG) of source B to target A seamlessly without effort." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.416, + 0.784, + 0.476 + ], + "angle": 0, + "content": "to two aspects: (i) how to parse known objects into their parts (i.e., recognizing that birds have eyes and tails), and (ii) how to assemble parts from different visual concepts to form a faithful creative concept (i.e., ensuring the output is recognizably a bird)." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.481, + 0.785, + 0.572 + ], + "angle": 0, + "content": "The former is easier. We make clever use of DINOv2 [45] feature maps and perform unsupervised clustering to identify common parts. The idea is that each cluster will then correspond to a semantic part of a common object (e.g., head and wings of birds). We specifically chose DINOv2 for its superior fine-grained perception [2] compared to others [50, 57, 58]. Further, this way has a higher flexibility to enhance fine-grained parsing by using a higher cluster count." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.578, + 0.785, + 0.684 + ], + "angle": 0, + "content": "Our major contribution lies in addressing the latter challenge. The solution is intuitive - it essentially revolves around the fine-grained selection and placement of the chosen parts. Inspired by recent efforts in personalization [3,25,36,53,64], primarily designed for learning entire objects, we introduce a tailored attention loss that specifically operates on parts. With this loss, our model learns the object part composition, ensuring the final generation appears holistically faithful with mixed parts (i.e., head and wings of a bird actually appear at the right places)." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.69, + 0.787, + 0.841 + ], + "angle": 0, + "content": "More specifically, we introduce an entropy-based attention loss that maximizes the attention of a specific part at a particular location while minimizing the attention where no parts appear. This is achieved by first selecting the attention maps that correspond to the parts. A normalization over all parts is then performed to ensure that each image region is occupied by no more than one part. Finally, we minimize the entropy loss between each normalized individual part and the semantic maps obtained during object parsing, containing the correct part location. This not only facilitates stronger part disentanglement but is also the key to generating a faithful holistic structure of an object as it learns a generic prior topology knowledge about object parts." + } + ], + [ + { + "type": "page_number", + "bbox": [ + 0.218, + 0.116, + 0.23, + 0.127 + ], + "angle": 0, + "content": "4" + }, + { + "type": "header", + "bbox": [ + 0.272, + 0.115, + 0.368, + 0.129 + ], + "angle": 0, + "content": "KW. Ng et al." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.147, + 0.788, + 0.223 + ], + "angle": 0, + "content": "To enhance generation fidelity, we further employ a bottleneck encoder to project the text tokens. This approach accelerates learning by leveraging shared knowledge (common parts) and facilitating information exchange among instances in each part. Each instance adjusts slightly to adapt to the fine-grained part details during optimization." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.223, + 0.79, + 0.421 + ], + "angle": 0, + "content": "Our contributions are as follows: (i) We pioneer a unique approach for fine-grained part-level control in Text-to-Image (T2I) models, empowering users to craft objects by selecting desired parts. This method streamlines the creative process, marking a significant advancement in our capacity to manipulate and reimagine visual content. (ii) We introduce PartCraft, a technique that autonomously parses object parts and orchestrates them from different visual concepts, resulting in the faithful creation of a novel object. (iii) For enhanced part disentanglement and generation fidelity, we propose an entropy-based normalized attention loss and leverage a bottleneck encoder. (iv) We present two problem-specific quantitative metrics. Comprehensive experiments on CUB-200-2011 (birds) and the Stanford Dogs dataset demonstrate the superior performance of our method in generating novel objects, surpassing alternative approaches in both qualitative and quantitative evaluations." + }, + { + "type": "title", + "bbox": [ + 0.216, + 0.433, + 0.388, + 0.449 + ], + "angle": 0, + "content": "2 Related Work" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.463, + 0.789, + 0.658 + ], + "angle": 0, + "content": "Creative editing and generation. Creativity involves generating innovative ideas or artifacts across various domains [12]. Extensive research has explored the integration of creativity into Generative Adversarial Networks (GANs) [21,44,56] and Variational Autoencoders (VAEs) [16, 18]. For example, DoodlerGAN [26] learns and combines fine-level part components to create sketches of new species. A recent study by [60] demonstrated decomposing personalized concepts into distinct visual aspects, creatively recombined through diffusion models. Instruct-Pix2Pix [10] allows creative image editing through instructions, while ConceptLab [51] aims to identify novel concepts within a specified category, deviating from existing concepts. Different from these works where editing/generation usually focuses on the whole object, we instead focus on training a text-to-image generative model that can understand parts, thus able to creatively generate new objects by seamlessly composing different parts simply through selection." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.66, + 0.789, + 0.841 + ], + "angle": 0, + "content": "Text-to-image generation. Recent advancements in large text-to-image (T2I) diffusion models [6, 19, 20, 43, 49, 52, 55] have made significant improvements over conventional methods [37, 46, 59, 67, 70, 72] in producing high-fidelity images from text prompts. Their application scope has expanded to include both global [10, 32, 41] and localized [4, 17, 28, 68] image editing tasks, demonstrating versatility. Methods such as [5, 7, 13, 15, 23, 24, 34, 38, 66, 69, 71] have introduced more granular spatial control, such as incorporating semantic segmentation masks or bounding boxes, into large pretrained diffusion models to guide image generation. Contrary to these approaches, we focus on enhancing the control by enabling a straightforward discrete selection of desired parts. We minimize the complexity and manual intervention required by spatial controls, yet the model can compose selected parts as a coherent object seamlessly. This not" + } + ], + [ + { + "type": "header", + "bbox": [ + 0.423, + 0.115, + 0.733, + 0.131 + ], + "angle": 0, + "content": "PartCraft: Crafting Creative Objects by Parts" + }, + { + "type": "page_number", + "bbox": [ + 0.775, + 0.116, + 0.785, + 0.127 + ], + "angle": 0, + "content": "5" + }, + { + "type": "image", + "bbox": [ + 0.222, + 0.159, + 0.79, + 0.3 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.214, + 0.312, + 0.788, + 0.412 + ], + "angle": 0, + "content": "Fig. 3: Overview of our PartCraft. All parts are organized into a dictionary, and their semantic embeddings are learned through a textual inversion approach. For instance, a text description like \"a photo of a [Head,42] [Wing,87]...\" guides the optimization of the corresponding textual embedding by reconstructing the associated image. To improve generation fidelity, we incorporate a bottleneck encoder \\( f \\) (MLP) to compute the embedding \\( y \\) (Eq. (4)) as input to the text encoder. To promote disentanglement among learned parts, we minimize a specially designed attention loss, denoted as \\( \\mathcal{L}_{\\text{att}} \\)." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.427, + 0.788, + 0.501 + ], + "angle": 0, + "content": "only simplifies the user's role in the generative process but also ensures that the compositional logic and coherence are inherently managed by the model's ability. As such, our research is distinguished by focusing on the model's inherent ability to understand and apply part relationships instead of providing additional controls." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.504, + 0.788, + 0.746 + ], + "angle": 0, + "content": "Abstracting visual knowledge as a text token. The effectiveness of T2I models is constrained by the user's ability to articulate their desired image through text. These models face challenges in faithfully replicating visual characteristics from a reference set and generating innovative interpretations in diverse contexts, even with detailed textual descriptions. To address this challenge, various personalization techniques have been developed. These techniques obtain a new word embedding from multiple images depicting the same concept [1, 25, 36, 53, 61] or multiple new word embeddings for various concepts within a single image [3] through inversion. The learned visual concepts can be creatively reused in many image editing tasks. Nonetheless, most of these approaches struggle to learn object parts, often not able to follow the part selections due to part entanglement as they were designed to learn object as a whole. In this work, we introduce a customized attention loss that serves a dual purpose: ensure accurate positioning of each part and enforce each image region occupied by no more than one part. This greatly improves the part disentanglement, creating novel concepts with correct appearances (see Fig. 2)." + }, + { + "type": "title", + "bbox": [ + 0.216, + 0.773, + 0.382, + 0.792 + ], + "angle": 0, + "content": "3 Methodology" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.81, + 0.787, + 0.842 + ], + "angle": 0, + "content": "Given a set of unlabeled images depicting the same object (e.g., bird) with different part details, we aim to train a T2I generative model that decomposes" + } + ], + [ + { + "type": "page_number", + "bbox": [ + 0.217, + 0.116, + 0.23, + 0.127 + ], + "angle": 0, + "content": "6" + }, + { + "type": "header", + "bbox": [ + 0.272, + 0.115, + 0.369, + 0.129 + ], + "angle": 0, + "content": "KW. Ng et al." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.147, + 0.784, + 0.177 + ], + "angle": 0, + "content": "parts of objects into text tokens and can recompose them in a novel way. To that end, we propose PartCraft, as depicted in Fig. 3." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.178, + 0.785, + 0.283 + ], + "angle": 0, + "content": "We start by discovering the parts in a three-tier hierarchy, as detailed in Sec. 3.1. Paired with the training images \\(\\{x_{i}\\}_{i = 1}^{N}\\), this semantic hierarchy subsequently serves as the supervision to fine-tune a pre-trained text-to-image model, say a latent diffusion model [52], denoted as \\(\\{\\epsilon_{\\theta},\\tau_{\\theta},\\mathcal{E},\\mathcal{D}\\}\\), where \\(\\epsilon_{\\theta}\\) represents the diffusion denoiser, \\(\\tau_{\\theta}\\) the text encoder, and \\(\\mathcal{E} / \\mathcal{D}\\) the autoencoder respectively. We adopt the textual inversion technique [25]. Concretely, we learn a set of pseudo-words \\(p^*\\) for each part in the word embedding space with:" + }, + { + "type": "equation", + "bbox": [ + 0.36, + 0.294, + 0.786, + 0.313 + ], + "angle": 0, + "content": "\\[\n\\mathcal {L} _ {l d m} = \\mathbb {E} _ {z, t, p, \\epsilon} \\left[ | | \\epsilon - \\epsilon_ {\\theta} \\left(z _ {t}, t, \\tau_ {\\theta} (y _ {p})\\right) | | _ {2} ^ {2} \\right], \\tag {1}\n\\]" + }, + { + "type": "equation", + "bbox": [ + 0.379, + 0.314, + 0.786, + 0.339 + ], + "angle": 0, + "content": "\\[\np ^ {*} = \\underset {p} {\\operatorname {a r g m i n}} \\mathcal {L} _ {l d m}, \\tag {2}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.351, + 0.785, + 0.457 + ], + "angle": 0, + "content": "where \\(\\epsilon \\sim \\mathcal{N}(0,1)\\) denotes the unscaled noise, \\(t\\) is the time step, \\(z = \\mathcal{E}(x)\\) is the latent representation of the image, \\(z_{t}\\) is the latent noise at time \\(t\\), and \\(y_{p}\\) is the text condition that includes \\(p\\) as part of the text tokens. \\(\\mathcal{L}_{ldm}\\) is a standard diffusion loss [30] to reconstruct the parts. As each object is composed of a set of parts, its reconstruction is achieved by the reconstruction of the associated set of parts. In other words, when all parts are reconstructed properly, it will become a valid object." + }, + { + "type": "title", + "bbox": [ + 0.214, + 0.48, + 0.504, + 0.495 + ], + "angle": 0, + "content": "3.1 Unsupervised Part Discovery" + }, + { + "type": "image", + "bbox": [ + 0.216, + 0.516, + 0.477, + 0.628 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.214, + 0.638, + 0.476, + 0.707 + ], + "angle": 0, + "content": "Fig. 4: Part discovery within a semantic hierarchy involves partitioning each image into distinct parts and forming semantic clusters across unlabeled training data." + }, + { + "type": "text", + "bbox": [ + 0.485, + 0.506, + 0.788, + 0.717 + ], + "angle": 0, + "content": "To minimize the labeling cost, we develop a scalable process to reveal the underlying semantic hierarchy with parts in an unsupervised fashion. We leverage the off-the-shelf vision model for image decomposition and clustering. Specifically, given an image \\( x_{i} \\), we employ DINOv2 [45] to extract the feature map \\( F = \\{F_{i} = \\mathrm{DINOv2}(x_{i})\\}_{i}^{N} \\). We then conduct three-level hierarchical clustering (see Fig. 4): (i) At the top level, \\( k \\)-means is applied on all patches in \\( F \\) with \\( k = 2 \\) to separate foregrounds and backgrounds \\( B \\). (ii) At the middle level, \\( k \\)-means is further ap" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.718, + 0.788, + 0.809 + ], + "angle": 0, + "content": "plied on all foreground patches to acquire \\( M \\) clusters representing common parts, such as the heads of birds. (iii) At the bottom level, we further group each of the \\( M \\) clusters as well as the background cluster \\( B \\) into \\( K \\) splits. Each split refers to finer meanings, such as the head of a specific bird species, or a specific background style. Lastly, each region of an image will be tagged with the corresponding cluster index. We represent these cluster tags as follows:" + }, + { + "type": "equation", + "bbox": [ + 0.388, + 0.82, + 0.786, + 0.837 + ], + "angle": 0, + "content": "\\[\np = (0, k _ {0}), (1, k _ {1}), \\dots , (M, k _ {M}), \\tag {3}\n\\]" + } + ], + [ + { + "type": "header", + "bbox": [ + 0.423, + 0.115, + 0.733, + 0.131 + ], + "angle": 0, + "content": "PartCraft: Crafting Creative Objects by Parts" + }, + { + "type": "page_number", + "bbox": [ + 0.775, + 0.116, + 0.787, + 0.127 + ], + "angle": 0, + "content": "7" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.147, + 0.784, + 0.253 + ], + "angle": 0, + "content": "where the first pair refers to the background style, and the following \\(M\\) pairs denote the combinations of \\(M\\) parts (e.g., head, body, wings) each associated with a specific object (e.g., sparrow), and \\(k \\in \\{1, \\dots, K\\}\\). This description will be used as the textual prompt in model training, such as \"a photo of a \\([p]\\)\". Please refer to the supplementary material for more examples of the discovered semantic hierarchy. This process also yields the segmentation mask of each \\(m\\)-th part, which we define as \\(S_{m}\\)." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.253, + 0.785, + 0.343 + ], + "angle": 0, + "content": "Motivation. While we can leverage off-the-shelf segmentation models such as VLPart [58], the robustness relies on the generalizability of the model and the part segmentation result is usually pre-defined and may be unstable for unseen domains. As a result, we rely on our feature clustering method to obtain the segmentation map, which also has a higher flexibility in choosing the number of clusters (parts)." + }, + { + "type": "title", + "bbox": [ + 0.215, + 0.367, + 0.45, + 0.381 + ], + "angle": 0, + "content": "3.2 Part Token Bottleneck" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.393, + 0.784, + 0.483 + ], + "angle": 0, + "content": "In contrast to prior text inversion studies [25], our task requires learning a greater quantity at the same time - specifically, \\((M + 1)K\\)-of word tokens derived from a collection of discovered parts marked by inherent imperfections (such as partial overlap and over splitting). This makes the learning task more demanding. To enhance the learning process, we propose a neural network \\(f(\\cdot)\\) comprising a two-layer MLP with ReLU activation:" + }, + { + "type": "equation", + "bbox": [ + 0.456, + 0.498, + 0.785, + 0.513 + ], + "angle": 0, + "content": "\\[\ny _ {p} = f (e (p)), \\tag {4}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.524, + 0.784, + 0.569 + ], + "angle": 0, + "content": "where \\( y_{p} \\) will be subsequently used as the input3 to the text encoder \\( \\tau_{\\theta} \\) and \\( e \\in \\mathbb{R}^{MK \\times D} \\) is a learnable word embedding dictionary that maps \\( p \\) to their respective embeddings." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.57, + 0.785, + 0.721 + ], + "angle": 0, + "content": "Our design demonstrates quicker convergence than directly learning the final word embeddings \\( e(\\cdot) \\) [25] (see supplementary material). This could be attributed to the entanglement of word embeddings in the conventional design, where there is no information exchange among them during optimization. For instance, each token doesn't know they are learning for a specific part of a specific species. This lack of communication leads to lower data efficiency and slower learning. With the bottleneck \\( f \\), it will first project the token into a common part embedding space (e.g., head), then slightly adjust itself to adapt the fine-grained part details. It's worth noting that the conventional design is a specific instance of our approach when \\( f \\) is an identity function." + }, + { + "type": "title", + "bbox": [ + 0.215, + 0.744, + 0.485, + 0.76 + ], + "angle": 0, + "content": "3.3 Learning to Craft by Parts" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.77, + 0.784, + 0.817 + ], + "angle": 0, + "content": "Fine-tuning the T2I model, rather than sorely learning pseudo-words, has been shown to achieve better reconstruction of target concepts as demonstrated in [36, 53]. However, this comes with a significant training cost. Thus, we apply LoRA" + }, + { + "type": "page_footnote", + "bbox": [ + 0.218, + 0.825, + 0.599, + 0.841 + ], + "angle": 0, + "content": "3 Word templates such as \"a photo of a [*]\" will be used." + } + ], + [ + { + "type": "page_number", + "bbox": [ + 0.218, + 0.116, + 0.23, + 0.127 + ], + "angle": 0, + "content": "8" + }, + { + "type": "header", + "bbox": [ + 0.272, + 0.115, + 0.369, + 0.13 + ], + "angle": 0, + "content": "KW. Ng et al." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.147, + 0.784, + 0.191 + ], + "angle": 0, + "content": "(low-rank adaptation) [31] to the cross-attention block for efficient training. We then minimize the diffusion loss \\(\\mathcal{L}_{ldm}\\) (Eq. (1)) to learn both pseudo-words and LoRA adapters." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.192, + 0.785, + 0.283 + ], + "angle": 0, + "content": "While training with only \\(\\mathcal{L}_{ldm}\\), entanglement happens between parts, as evident from the attention maps in the cross-attention block of the denoiser \\(\\epsilon_{\\theta}\\) (see Fig. 9). This entanglement arises due to the correlation between parts (e.g., a bird head code is consistently paired with a bird body code to represent the same species). To address this issue, we introduce an entropy-based attention loss as regularization:" + }, + { + "type": "equation", + "bbox": [ + 0.305, + 0.295, + 0.785, + 0.313 + ], + "angle": 0, + "content": "\\[\n\\mathcal {L} _ {\\text {a t t n}} = \\mathbb {E} _ {z, t, m} \\left[ - \\left(S _ {m} \\log \\hat {A} _ {m} + (1 - S _ {m}) \\log (1 - \\hat {A} _ {m})\\right) \\right], \\tag {5}\n\\]" + }, + { + "type": "equation", + "bbox": [ + 0.3, + 0.317, + 0.785, + 0.356 + ], + "angle": 0, + "content": "\\[\n\\hat {A} _ {m, i, j} = \\frac {\\bar {A} _ {m , i , j}}{\\sum_ {k} \\bar {A} _ {k , i , j}}, \\quad \\bar {A} _ {m} = \\frac {1}{L} \\sum_ {l} ^ {L} A _ {l, m}, \\tag {6}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.369, + 0.788, + 0.476 + ], + "angle": 0, + "content": "where \\(A \\in [0,1]^{M \\times HW}\\) represents the cross-attention map between the \\(m\\)-th part and the noisy latent \\(z_{t}\\), \\(L\\) represents the number of specific layers to select attention maps, \\(\\hat{A} \\in [0,1]^{M \\times HW}\\) represents the averaged and normalized cross-attention map over all parts and \\(S_{m} \\in \\{0,1\\}^{M \\times HW}\\) serves as the mask that indicates the location of \\(m\\)-th part. In cases where the part is not present in the image (e.g., occluded), we set both \\(S_{m}\\) and \\(\\hat{A}_{m}\\) as 0 to exclude them. Thus, the overall learning objective is defined as:" + }, + { + "type": "equation", + "bbox": [ + 0.406, + 0.489, + 0.785, + 0.504 + ], + "angle": 0, + "content": "\\[\n\\mathcal {L} _ {\\text {t o t a l}} = \\mathcal {L} _ {l d m} + \\lambda_ {\\text {a t t n}} \\mathcal {L} _ {\\text {a t t n}}, \\tag {7}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.515, + 0.788, + 0.668 + ], + "angle": 0, + "content": "where \\(\\lambda_{attn} = 0.01\\). We focus on attention maps at the resolution of \\(16 \\times 16\\) where rich semantic information is captured [28]. Normalization is performed at each location to ensure that the sum of a patch location equals 1. This aims to maximize the attention of a specific part at a particular location which implicitly minimizing the attention of other parts similar to a softmax classification task. Compared to the mean-square based attention loss [3], this intuitively ensures that a part only appears once at a particular location, facilitating stronger disentanglement from other parts during the denoising operation. When generating a part for a particular location, the diffusion model \\(\\epsilon_{\\theta}\\) should only attend to the part instead of other non-related parts." + }, + { + "type": "title", + "bbox": [ + 0.215, + 0.684, + 0.377, + 0.701 + ], + "angle": 0, + "content": "4 Experiments" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.716, + 0.788, + 0.762 + ], + "angle": 0, + "content": "Datasets. We demonstrate our selection task on two fine-grained object datasets: CUB-200-2011 (birds) [62] which contains 5,994 training images, and Stanford Dogs [33] which contains 12,000 training images." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.765, + 0.788, + 0.841 + ], + "angle": 0, + "content": "Implementation. For part composition, we assess the model's ability to combine up to 4 different parts from 4 distinct species/objects. We set \\( M = 5 \\) for bird generation (head, front body/breast area, wings, legs, tail) and \\( M = 7 \\) for dog generation (forehead, eyes, mouth/nose, ears, neck, body/tail, legs). For both datasets, \\( K \\) is set as 256, ensuring sufficient coverage of all fine-grained" + } + ], + [ + { + "type": "header", + "bbox": [ + 0.423, + 0.115, + 0.733, + 0.13 + ], + "angle": 0, + "content": "PartCraft: Crafting Creative Objects by Parts" + }, + { + "type": "page_number", + "bbox": [ + 0.775, + 0.116, + 0.785, + 0.127 + ], + "angle": 0, + "content": "9" + }, + { + "type": "table", + "bbox": [ + 0.248, + 0.144, + 0.75, + 0.226 + ], + "angle": 0, + "content": "
MethodLearnable TokenFine TuneDisentanglementBottleneck
Textual Inversion [25]✓✗✗✗
DreamBooth [53]✓LoRA*✗✗
CustomDiffusion [36]✓K/V✗✗
Break-a-scene [3]✓LoRA*MSE✗
PartCraft (Ours)✓LoRAEq. (5)✓
" + }, + { + "type": "table_caption", + "bbox": [ + 0.214, + 0.228, + 0.784, + 0.27 + ], + "angle": 0, + "content": "Table 1: Comparing our and alternative methods in design properties. *: We finetuned the added LoRA [31] adapter rather than the entire diffusion model \\(\\epsilon_{\\theta}\\) due to resource limit. MSE is a mean-square based attention loss used in [3]." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.333, + 0.785, + 0.424 + ], + "angle": 0, + "content": "classes (i.e., 200 for birds and 120 for dogs). We randomly generate 500 images by sampling 500 sets of parts. For each image, we randomly replace an original part with any part from another 500 non-overlapping sets of parts. The resulting set of parts may take the form of \\( \\left( {0,{k}_{A}}\\right) \\left( {1,{k}_{B}}\\right) \\left( {2,{k}_{C}}\\right) \\ldots \\left( {M,{k}_{D}}\\right) \\) ,representing a composition from species A,B,C,and D. Stable Diffusion v1.5 [52] is used. Please see the supplementary material for further training details." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.443, + 0.788, + 0.563 + ], + "angle": 0, + "content": "We compare our method with the recent personalization methods: Textual Inversion (TI) [25], DreamBooth (DB) [53], Custom Diffusion (CD) [36], Break-a-scene (BaS) [3]. These personalization methods were designed to take single or multiple images with associated labeled objects as input. The text prompt for each image is as simple as \"a photo of \\([p]\\)\" where \\( p \\) is expressed in Eq. (3), since we do not rely on complex prompts. We employ the official implementations released by the original authors for training. We summarize the main design properties of all compared methods in Tab. 1." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.583, + 0.788, + 0.841 + ], + "angle": 0, + "content": "Evaluation metrics. To assess a model's ability to disentangle and composite parts, we introduce two metrics: (a) exact matching rate (EMR) and (b) cosine similarity (CoSim) between the \\( k \\)-means embeddings of the parts of real and generated images. Utilizing the pre-trained \\( k \\)-means from Sec. 3.1, we predict the parts of generated images. EMR quantifies how accurately the cluster index of parts of generated images matches the parts of the corresponding real images whereas CoSim measures the cosine similarity between the \\( k \\)-means centroid vector that the part belongs to between generated and real images. These metrics assess the model's ability to follow the input parts and accurately reconstruct them, with perfect disentanglement indicated by EMR of 1 and CoSim of 1. A detailed algorithm is provided in the supplementary material. We also measure image generation quality using FID [29] to assess model performance in terms of image distribution. Additionally, we compute the average pairwise cosine similarity between CLIP [48]/DINO [11] embeddings of generated and real class-specific images following [53]. Each generated image is conditioned on the parts of the corresponding real image. This results in 5,994 generated images for birds and 12,000 generated images for dogs." + } + ], + [ + { + "type": "page_number", + "bbox": [ + 0.218, + 0.116, + 0.236, + 0.127 + ], + "angle": 0, + "content": "10" + }, + { + "type": "header", + "bbox": [ + 0.272, + 0.115, + 0.368, + 0.129 + ], + "angle": 0, + "content": "KW. Ng et al." + }, + { + "type": "table", + "bbox": [ + 0.221, + 0.143, + 0.784, + 0.254 + ], + "angle": 0, + "content": "
MethodBirds: CUB-200-2011Dogs: Stanford Dogs
FIDCLIPDINOEMRCoSimFIDCLIPDINOEMRCoSim
Textual Inversion [25]10.100.7840.6070.3050.84223.360.6520.5320.2180.754
DreamBooth [53]12.940.7750.5940.3550.85622.650.6600.5630.2750.777
Custom Diffusion [36]37.610.6940.5040.3380.83342.410.5930.4910.2530.755
Break-a-Scene [3]20.050.7420.5490.3900.85424.200.6330.5320.3000.775
PartCraft (Ours)12.860.7830.6180.4600.88216.920.6690.5730.3580.796
" + }, + { + "type": "table_caption", + "bbox": [ + 0.305, + 0.258, + 0.697, + 0.272 + ], + "angle": 0, + "content": "Table 2: Quantitative comparison for part reconstruction." + }, + { + "type": "image", + "bbox": [ + 0.218, + 0.28, + 0.331, + 0.452 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.255, + 0.453, + 0.296, + 0.466 + ], + "angle": 0, + "content": "(a) TI" + }, + { + "type": "image", + "bbox": [ + 0.332, + 0.281, + 0.444, + 0.451 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.366, + 0.453, + 0.411, + 0.466 + ], + "angle": 0, + "content": "(b) DB" + }, + { + "type": "image", + "bbox": [ + 0.446, + 0.28, + 0.557, + 0.452 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.48, + 0.453, + 0.525, + 0.465 + ], + "angle": 0, + "content": "(c) CD" + }, + { + "type": "image", + "bbox": [ + 0.559, + 0.281, + 0.671, + 0.452 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.591, + 0.453, + 0.641, + 0.465 + ], + "angle": 0, + "content": "(d) BaS" + }, + { + "type": "image", + "bbox": [ + 0.673, + 0.281, + 0.785, + 0.452 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.689, + 0.453, + 0.768, + 0.466 + ], + "angle": 0, + "content": "(e) PartCraft" + }, + { + "type": "image_caption", + "bbox": [ + 0.214, + 0.476, + 0.785, + 0.505 + ], + "angle": 0, + "content": "Fig. 5: Visual comparison under the part reconstruction setting. All images are generated by using the original parts of respective objects." + }, + { + "type": "title", + "bbox": [ + 0.214, + 0.528, + 0.682, + 0.543 + ], + "angle": 0, + "content": "4.1 Part Reconstruction and Image Quality Evaluation" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.554, + 0.785, + 0.599 + ], + "angle": 0, + "content": "We first assess the ability of different methods to learn parts as text tokens by evaluating how well they can accurately reconstruct the parts (this also means image generation with original parts)." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.599, + 0.787, + 0.841 + ], + "angle": 0, + "content": "In Tab. 2, we summarize the performance of respective methods on the bird and dog generation, respectively. We highlight four observations: (i) Textual Inversion performs quite well compared to DreamBooth, CustomDiffusion, and Break-a-scene in terms of FID, CLIP, and DINO scores although did not finetune the diffusion model \\(\\epsilon_{\\theta}\\). This may be due to the potential risk of overfitting when fine-tuning \\(\\epsilon_{\\theta}\\) especially when learning a vast array of new concepts with many update iterations. It is also not uncommon to carefully tune the learning rate and the training iterations in these models when fine-tuning new concepts (e.g., only 800-1000 steps of updates to learn a new concept in [3]). (ii) Nonetheless, fine-tuning the diffusion model \\(\\epsilon_{\\theta}\\) can help improve the ability to follow prompts as shown by increased EMR and CoSim scores (e.g., EMR of at least \\(5\\%\\) in DreamBooth). (iii) Break-a-scene has a better ability to reconstruct the parts as shown by EMR and CoSim, this is due to the attention loss explicitly forcing the parts to focus on the respective semantic region. (iv) PartCraft achieves the best performance in DINO, EMR, and CoSim scores (e.g., \\(7\\%\\) better in EMR compared to Break-a-scene). This indicates that not only does our" + } + ], + [ + { + "type": "header", + "bbox": [ + 0.423, + 0.115, + 0.733, + 0.131 + ], + "angle": 0, + "content": "PartCraft: Crafting Creative Objects by Parts" + }, + { + "type": "page_number", + "bbox": [ + 0.768, + 0.116, + 0.784, + 0.127 + ], + "angle": 0, + "content": "11" + }, + { + "type": "image", + "bbox": [ + 0.218, + 0.145, + 0.787, + 0.257 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.221, + 0.265, + 0.78, + 0.28 + ], + "angle": 0, + "content": "Fig. 6: Quantitative comparisons of part composition in terms of EMR and CoSim." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.295, + 0.789, + 0.387 + ], + "angle": 0, + "content": "image-generation ability perform comparably well with textual inversion, but PartCraft is also able to disentangle the parts learning so that it can follow the prompt instructions more accurately to generate the parts in a cohort. In Fig. 5, we present generated images from different methods, with CustomDiffusion exhibiting high-contrast images, possibly due to unconstrained fine-tuning on the cross-attention components \\( K / V \\) and resulting in worse FID scores." + }, + { + "type": "title", + "bbox": [ + 0.216, + 0.407, + 0.504, + 0.423 + ], + "angle": 0, + "content": "4.2 Part Composition Evaluation" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.431, + 0.789, + 0.552 + ], + "angle": 0, + "content": "In this section, we assess the part composition ability of different methods. In this experiment, we generate the image by mixing different parts from different species. Our findings, as shown in Fig. 6, can be summarized as follows: (i) As the number of composited parts increases, EMR and CoSim decrease, reflecting the challenge of composing multiple diverse parts. (ii) Break-a-scene and PartCraft achieve notably higher EMR and CoSim scores, thanks to disentanglement through attention loss minimization. (iii) PartCraft outperforms Break-a-scene significantly by token bottleneck and tailored attention loss." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.553, + 0.79, + 0.657 + ], + "angle": 0, + "content": "In Fig. 7, we visualize the results of composing 4 different parts. While all images appear realistic, most methods struggle to assemble all 4 parts. For instance, Break-a-scene missed out on the flurry body of kerry blue terrier (right-most column). In contrast, our methods successfully combine 4 different parts from 4 different species, demonstrating the superior ability of our approach to part composition. We also visualize additional examples of our method in the supplementary material." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.658, + 0.789, + 0.735 + ], + "angle": 0, + "content": "Furthermore, we explore the versatility of the adapted model by generating images with simple styles such as pencil drawing. While most methods successfully incorporate specific styles into the generated image, Custom Diffusion often fails to do so, possibly due to the unconstrained fine-tuning of the cross-attention components \\( K / V \\)." + }, + { + "type": "title", + "bbox": [ + 0.216, + 0.755, + 0.403, + 0.77 + ], + "angle": 0, + "content": "4.3 Ablation Studies" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.78, + 0.789, + 0.842 + ], + "angle": 0, + "content": "Component analysis. In Fig. 8, we evaluate the effect of our proposed components (token bottleneck and attention loss) on creating novel bird species. (i) Removing the bottleneck outlined in Eq. (4) degrades the generation quality as evidenced by a higher FID score \\((12.86 \\rightarrow 16.36)\\) even though both EMR" + } + ], + [ + { + "type": "page_number", + "bbox": [ + 0.218, + 0.116, + 0.236, + 0.127 + ], + "angle": 0, + "content": "12" + }, + { + "type": "header", + "bbox": [ + 0.272, + 0.114, + 0.369, + 0.129 + ], + "angle": 0, + "content": "KW. Ng et al." + }, + { + "type": "image", + "bbox": [ + 0.218, + 0.143, + 0.787, + 0.632 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.214, + 0.641, + 0.788, + 0.684 + ], + "angle": 0, + "content": "Fig. 7: Visual comparison on 4-species (specified on the top row) mixed generation. The last column indicates generated images with different styles (i.e., \\(DSLR\\), Van Gogh, Oil Painting, Pencil Drawing)." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.727, + 0.788, + 0.804 + ], + "angle": 0, + "content": "and CoSim remain. (ii) By replacing our \\(\\mathcal{L}_{attn}\\) with the MSE loss as proposed in [3], we observe significant deterioration in both EMR and CoSim. (iii) Finally, incorporating both the projector and our attention loss performs the best. This improvement highlights the necessity of incorporating interactions between multiple parts to achieve more effective part disentanglement and optimization." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.81, + 0.788, + 0.842 + ], + "angle": 0, + "content": "Cross-attention visualization. Our attention loss plays a crucial role in token disentanglement. We demonstrate the impact of this loss in Fig. 9, where we" + } + ], + [ + { + "type": "header", + "bbox": [ + 0.423, + 0.115, + 0.733, + 0.13 + ], + "angle": 0, + "content": "PartCraft: Crafting Creative Objects by Parts" + }, + { + "type": "page_number", + "bbox": [ + 0.768, + 0.116, + 0.786, + 0.127 + ], + "angle": 0, + "content": "13" + }, + { + "type": "image", + "bbox": [ + 0.218, + 0.145, + 0.498, + 0.234 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.215, + 0.248, + 0.498, + 0.29 + ], + "angle": 0, + "content": "Fig.8: Ablation on our part token bottleneck and attention loss under the part composition on CUB-200-2011 birds." + }, + { + "type": "image", + "bbox": [ + 0.505, + 0.162, + 0.788, + 0.22 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.503, + 0.228, + 0.785, + 0.269 + ], + "angle": 0, + "content": "Fig.9: Cross-attention map of each part (top) without and (bottom) with our attention loss." + }, + { + "type": "image", + "bbox": [ + 0.226, + 0.324, + 0.396, + 0.429 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.247, + 0.437, + 0.372, + 0.449 + ], + "angle": 0, + "content": "(a) Textual Inversion" + }, + { + "type": "image", + "bbox": [ + 0.419, + 0.324, + 0.589, + 0.43 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.449, + 0.437, + 0.555, + 0.449 + ], + "angle": 0, + "content": "(b) Break-a-scene" + }, + { + "type": "image", + "bbox": [ + 0.61, + 0.324, + 0.782, + 0.431 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.633, + 0.437, + 0.754, + 0.449 + ], + "angle": 0, + "content": "(c) PartCraft (Ours)" + }, + { + "type": "image_caption", + "bbox": [ + 0.214, + 0.46, + 0.785, + 0.489 + ], + "angle": 0, + "content": "Fig. 10: 2D tSNE [39] projection of word embeddings. Different colors represent different common parts. (correspond to the segmentation mask in Fig. 4)." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.536, + 0.785, + 0.567 + ], + "angle": 0, + "content": "observe significantly enhanced disentanglement after explicitly guiding attention to focus on distinct semantic regions." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.575, + 0.788, + 0.697 + ], + "angle": 0, + "content": "Part word embedding space. We visualize the word embeddings of learned tokens of Textual Inversion [25], Break-a-scene [3] and our PartCraft for birds generation (CUB-200-2011 [62]) through tSNE [39] in Fig. 10. In our PartCraft, the word embeddings are the projected embeddings through Eq. (4). We can see that our projected version has a better semantic meaning such that the part embeddings are clustered together by their semantic meaning (e.g., head). We believe this is one of the reasons that our PartCraft outperforms previous methods in which we can compose all parts seamlessly yet with higher quality." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.704, + 0.788, + 0.841 + ], + "angle": 0, + "content": "Transferability for Creativity. (i) In Fig. 11, we demonstrate that not only it can compose parts within the training domain (e.g., birds), but it can also transfer the learned parts to and combine with other domains (e.g., cat). This enables the creation of unique combinations, such as a cat with a dog's ear. (ii) Leveraging the prior knowledge embedded in Stable Diffusion, PartCraft can also repurpose learned parts for creative image generation. An example of this is the generation of a bird-shaped robot adorned with various parts. These examples showcase PartCraft's immense potential for diverse and limitless creative applications. Please see the supplementary material for more examples." + } + ], + [ + { + "type": "page_number", + "bbox": [ + 0.218, + 0.116, + 0.236, + 0.127 + ], + "angle": 0, + "content": "14" + }, + { + "type": "header", + "bbox": [ + 0.272, + 0.115, + 0.368, + 0.129 + ], + "angle": 0, + "content": "KW. Ng et al." + }, + { + "type": "image", + "bbox": [ + 0.236, + 0.144, + 0.768, + 0.306 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.214, + 0.316, + 0.788, + 0.388 + ], + "angle": 0, + "content": "Fig. 11: (Top): Using learned parts to modify the property of other domains such as cat, and lion with prompt such as \"A cat with [beagle's ear)\". (Bottom) We can also repurpose learned parts for creative image generation using prior knowledge in Stable Diffusion with prompt such as \"A robot designed inspired by [red header woodpecker's head] and [blue jay's body)\"." + }, + { + "type": "title", + "bbox": [ + 0.216, + 0.398, + 0.359, + 0.415 + ], + "angle": 0, + "content": "5 Conclusion" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.426, + 0.789, + 0.622 + ], + "angle": 0, + "content": "We propose a new way of control in generative AI. Instead of text or sketch, we \"select\" desired parts to create an object. We addressed the challenge of learning parts in T2I models by introducing a customized attention loss. This loss serves a dual purpose: to ensure parts are at the right location and to ensure each location is occupied by not more than one part. This greatly improves the part disentanglement. We further employ a non-linear bottleneck encoder to improve generation fidelity. Our model, PartCraft, can seamlessly compose different parts from different objects, creating objects that do not exist yet holistically correct and plausible objects by mixing them. Extensive experiments demonstrated PartCraft's superior performance in both qualitative and quantitative evaluation. Moreover, the learned parts demonstrate strong transferability. We hope that our PartCraft will empower artists, designers, and enthusiasts to bring the creations of their dreams to reality." + }, + { + "type": "title", + "bbox": [ + 0.216, + 0.644, + 0.547, + 0.66 + ], + "angle": 0, + "content": "6 Limitations and Future Works" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.674, + 0.789, + 0.841 + ], + "angle": 0, + "content": "It is worth noting that the accuracy of obtained parts may be affected by using a self-supervised pre-trained feature extractor, incorporating encoders [64] may help. We also observed challenges in composing relatively small parts, like tails and legs. Additionally, we are also exploring cross-domain generation, i.e., combining learned parts from different datasets to create objects with even more diverse parts. For instance, we can merge non-rigid parts (e.g., parts from quadrupled animals) with rigid parts (e.g., parts from cars/airplanes) and form a creative structure (e.g., a car that has horse legs instead of wheels). This not only further improves the applicability of PartCraft but also serves as a stepping stone to creative generation, as generative AI progresses, continually expanding the limits of achievable creativity and artistic expression." + } + ], + [ + { + "type": "header", + "bbox": [ + 0.423, + 0.115, + 0.732, + 0.129 + ], + "angle": 0, + "content": "PartCraft: Crafting Creative Objects by Parts" + }, + { + "type": "page_number", + "bbox": [ + 0.768, + 0.116, + 0.784, + 0.127 + ], + "angle": 0, + "content": "15" + }, + { + "type": "title", + "bbox": [ + 0.217, + 0.145, + 0.402, + 0.163 + ], + "angle": 0, + "content": "Acknowledgements" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.179, + 0.788, + 0.24 + ], + "angle": 0, + "content": "We extend our special thanks to Jia Wei Sii for her help in creating figures and discussing the main concept. We are also grateful to Ruoyi Du, Zhiyu Qu, Chee Seng Chan, and the reviewers for their fruitful comments and corrections on our draft, methodology, and experiments." + }, + { + "type": "title", + "bbox": [ + 0.216, + 0.264, + 0.323, + 0.279 + ], + "angle": 0, + "content": "References" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.296, + 0.785, + 0.324 + ], + "angle": 0, + "content": "1. Alaluf, Y., Richardson, E., Metzer, G., Cohen-Or, D.: A neural space-time representation for text-to-image personalization. arXiv preprint arXiv:2305.15391 (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.325, + 0.785, + 0.352 + ], + "angle": 0, + "content": "2. Amir, S., Gandelsman, Y., Bagon, S., Dekel, T.: On the effectiveness of vit features as local semantic descriptors. In: ECCV (2022)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.353, + 0.785, + 0.38 + ], + "angle": 0, + "content": "3. Avrahami, O., Aberman, K., Fried, O., Cohen-Or, D., Lischinski, D.: Break-a-scene: Extracting multiple concepts from a single image. In: SIGGRAPH Asia (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.381, + 0.785, + 0.394 + ], + "angle": 0, + "content": "4. Avrahami, O., Fried, O., Lischinski, D.: Blended latent diffusion. ACM TOG (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.395, + 0.785, + 0.435 + ], + "angle": 0, + "content": "5. Avrahami, O., Hayes, T., Gafni, O., Gupta, S., Taigman, Y., Parikh, D., Lischinski, D., Fried, O., Yin, X.: Spatext: Spatio-textual representation for controllable image generation. In: CVPR (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.437, + 0.785, + 0.478 + ], + "angle": 0, + "content": "6. Balaji, Y., Nah, S., Huang, X., Vahdat, A., Song, J., Kreis, K., Aittala, M., Aila, T., Laine, S., Catanzaro, B., et al.: edifi: Text-to-image diffusion models with an ensemble of expert denoisers. arXiv preprint arXiv:2211.01324 (2022)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.479, + 0.785, + 0.505 + ], + "angle": 0, + "content": "7. Bar-Tal, O., Yariv, L., Lipman, Y., Dekel, T.: Multidiffusion: Fusing diffusion paths for controlled image generation. In: ICML (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.507, + 0.785, + 0.534 + ], + "angle": 0, + "content": "8. Biederman, I.: Recognition-by-components: a theory of human image understanding. Psychological review (1987)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.535, + 0.785, + 0.575 + ], + "angle": 0, + "content": "9. Bonnardel, N., Marmèche, E.: Towards supporting evocation processes in creative design: A cognitive approach. International Journal of Human-Computer Studies (2005)" + }, + { + "type": "ref_text", + "bbox": [ + 0.218, + 0.577, + 0.785, + 0.603 + ], + "angle": 0, + "content": "10. Brooks, T., Holynski, A., Efros, A.A.: Instructpix2pix: Learning to follow image editing instructions. In: CVPR (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.218, + 0.604, + 0.785, + 0.632 + ], + "angle": 0, + "content": "11. Caron, M., Touvron, H., Misra, I., Jégou, H., Mairal, J., Bojanowski, P., Joulin, A.: Emerging properties in self-supervised vision transformers. In: ICCV (2021)" + }, + { + "type": "ref_text", + "bbox": [ + 0.218, + 0.633, + 0.785, + 0.672 + ], + "angle": 0, + "content": "12. Cetinic, E., She, J.: Understanding and creating art with ai: Review and outlook. ACM Transactions on Multimedia Computing, Communications, and Applications (2022)" + }, + { + "type": "ref_text", + "bbox": [ + 0.218, + 0.673, + 0.785, + 0.714 + ], + "angle": 0, + "content": "13. Chefer, H., Alaluf, Y., Vinker, Y., Wolf, L., Cohen-Or, D.: Attend-and-excite: Attention-based semantic guidance for text-to-image diffusion models. ACM TOG (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.218, + 0.715, + 0.785, + 0.743 + ], + "angle": 0, + "content": "14. Chen, C., Li, O., Tao, D., Barnett, A., Rudin, C., Su, J.K.: This looks like that: deep learning for interpretable image recognition. NeurIPS (2019)" + }, + { + "type": "ref_text", + "bbox": [ + 0.218, + 0.744, + 0.785, + 0.771 + ], + "angle": 0, + "content": "15. Chen, M., Laina, I., Vedaldi, A.: Training-free layout control with cross-attention guidance (2024)" + }, + { + "type": "ref_text", + "bbox": [ + 0.218, + 0.772, + 0.785, + 0.812 + ], + "angle": 0, + "content": "16. Cintas, C., Das, P., Quanz, B., Tadesse, G.A., Speakman, S., Chen, P.Y.: Towards creativity characterization of generative models via group-based subset scanning. In: IJCAI (2022)" + }, + { + "type": "ref_text", + "bbox": [ + 0.218, + 0.813, + 0.785, + 0.84 + ], + "angle": 0, + "content": "17. Couairon, G., Verbeek, J., Schwenk, H., Cord, M.: Diffedit: Diffusion-based semantic image editing with mask guidance. In: ICLR (2023)" + }, + { + "type": "list", + "bbox": [ + 0.218, + 0.296, + 0.785, + 0.84 + ], + "angle": 0, + "content": null + } + ], + [ + { + "type": "page_number", + "bbox": [ + 0.218, + 0.116, + 0.236, + 0.127 + ], + "angle": 0, + "content": "16" + }, + { + "type": "header", + "bbox": [ + 0.272, + 0.115, + 0.368, + 0.129 + ], + "angle": 0, + "content": "KW. Ng et al." + }, + { + "type": "ref_text", + "bbox": [ + 0.218, + 0.147, + 0.785, + 0.175 + ], + "angle": 0, + "content": "18. Das, P., Quanz, B., Chen, P.Y., Ahn, J.w., Shah, D.: Toward a neuro-inspired creative decoder. In: IJCAI (2020)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.177, + 0.785, + 0.218 + ], + "angle": 0, + "content": "19. Ding, M., Yang, Z., Hong, W., Zheng, W., Zhou, C., Yin, D., Lin, J., Zou, X., Shao, Z., Yang, H., et al.: Cogview: Mastering text-to-image generation via transformers. In: NeurIPS (2021)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.219, + 0.785, + 0.246 + ], + "angle": 0, + "content": "20. Ding, M., Zheng, W., Hong, W., Tang, J.: Cogview2: Faster and better text-to-image generation via hierarchical transformers. In: NeurIPS. vol. 35 (2022)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.247, + 0.785, + 0.288 + ], + "angle": 0, + "content": "21. Elgammal, A., Liu, B., Elhoseiny, M., Mazzone, M.: Can: Creative adversarial networks generating \"art\" by learning about styles and deviating from style norms. In: ICCC (2017)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.289, + 0.785, + 0.316 + ], + "angle": 0, + "content": "22. Felzenszwalb, P.F., Huttenlocher, D.P.: Pictorial structures for object recognition. IJCV (2005)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.317, + 0.785, + 0.359 + ], + "angle": 0, + "content": "23. Feng, W., He, X., Fu, T.J., Jampani, V., Akula, A.R., Narayana, P., Basu, S., Wang, X.E., Wang, W.Y.: Training-free structured diffusion guidance for compositional text-to-image synthesis. In: ICLR (2022)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.36, + 0.785, + 0.387 + ], + "angle": 0, + "content": "24. Gafni, O., Polyak, A., Ashual, O., Sheynin, S., Parikh, D., Taigman, Y.: Make-ascene: Scene-based text-to-image generation with human priors. In: ECCV (2022)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.388, + 0.785, + 0.43 + ], + "angle": 0, + "content": "25. Gal, R., Alaluf, Y., Atzmon, Y., Patashnik, O., Bermano, A.H., Chechik, G., Cohen-Or, D.: An image is worth one word: Personalizing text-to-image generation using textual inversion. In: ICLR (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.43, + 0.785, + 0.457 + ], + "angle": 0, + "content": "26. Ge, S., Goswami, V., Zitnick, C.L., Parikh, D.: Creative sketch generation. In: ICLR (2021)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.458, + 0.785, + 0.486 + ], + "angle": 0, + "content": "27. He, J., Chen, J., Lin, M.X., Yu, Q., Yuille, A.L.: Compositor: Bottom-up clustering and compositing for robust part and object segmentation. In: CVPR (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.487, + 0.785, + 0.514 + ], + "angle": 0, + "content": "28. Hertz, A., Mokady, R., Tenenbaum, J., Aberman, K., Pritch, Y., Cohen-Or, D.: Prompt-to-prompt image editing with cross attention control. In: ICLR (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.515, + 0.785, + 0.556 + ], + "angle": 0, + "content": "29. Heusel, M., Ramsauer, H., Unterthiner, T., Nessler, B., Hochreiter, S.: Gans trained by a two time-scale update rule converge to a local nash equilibrium. NeurIPS 30 (2017)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.557, + 0.785, + 0.584 + ], + "angle": 0, + "content": "30. Ho, J., Jain, A., Abbeel, P.: Denoising diffusion probabilistic models. In: NeurIPS (2020)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.585, + 0.785, + 0.613 + ], + "angle": 0, + "content": "31. Hu, E.J., Shen, Y., Wallis, P., Allen-Zhu, Z., Li, Y., Wang, S., Wang, L., Chen, W.: LoRA: Low-rank adaptation of large language models. In: ICLR (2022)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.614, + 0.785, + 0.642 + ], + "angle": 0, + "content": "32. Kawar, B., Zada, S., Lang, O., Tov, O., Chang, H., Dekel, T., Mosseri, I., Irani, M.: Imagic: Text-based real image editing with diffusion models. In: CVPR (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.642, + 0.785, + 0.67 + ], + "angle": 0, + "content": "33. Khosla, A., Jayadevaprakash, N., Yao, B., Li, F.F.: Novel dataset for fine-grained image categorization: Stanford dogs. In: CVPRW (2011)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.671, + 0.785, + 0.698 + ], + "angle": 0, + "content": "34. Kim, Y., Lee, J., Kim, J.H., Ha, J.W., Zhu, J.Y.: Dense text-to-image generation with attention modulation. In: ICCV (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.699, + 0.785, + 0.727 + ], + "angle": 0, + "content": "35. Krause, J., Jin, H., Yang, J., Fei-Fei, L.: Fine-grained recognition without part annotations. In: CVPR (2015)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.728, + 0.785, + 0.755 + ], + "angle": 0, + "content": "36. Kumari, N., Zhang, B., Zhang, R., Shechtman, E., Zhu, J.Y.: Multi-concept customization of text-to-image diffusion. In: CVPR (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.756, + 0.785, + 0.783 + ], + "angle": 0, + "content": "37. Li, B., Qi, X., Lukasiewicz, T., Torr, P.: Controllable text-to-image generation. In: NeurIPS (2019)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.784, + 0.785, + 0.812 + ], + "angle": 0, + "content": "38. Li, Y., Liu, H., Wu, Q., Mu, F., Yang, J., Gao, J., Li, C., Lee, Y.J.: Gligen: Open-set grounded text-to-image generation. In: CVPR (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.813, + 0.785, + 0.84 + ], + "angle": 0, + "content": "39. Van der Maaten, L., Hinton, G.: Visualizing data using t-sne. Journal of Machine Learning Research (2008)" + }, + { + "type": "list", + "bbox": [ + 0.217, + 0.147, + 0.785, + 0.84 + ], + "angle": 0, + "content": null + } + ], + [ + { + "type": "header", + "bbox": [ + 0.423, + 0.115, + 0.732, + 0.129 + ], + "angle": 0, + "content": "PartCraft: Crafting Creative Objects by Parts" + }, + { + "type": "page_number", + "bbox": [ + 0.768, + 0.116, + 0.785, + 0.127 + ], + "angle": 0, + "content": "17" + }, + { + "type": "ref_text", + "bbox": [ + 0.214, + 0.147, + 0.788, + 0.189 + ], + "angle": 0, + "content": "40. Mo, S., Mu, F., Lin, K.H., Liu, Y., Guan, B., Li, Y., Zhou, B.: Freecontrol: Training-free spatial control of any text-to-image diffusion model with any condition. arXiv preprint arXiv:2312.07536 (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.214, + 0.19, + 0.787, + 0.218 + ], + "angle": 0, + "content": "41. Mokady, R., Hertz, A., Aberman, K., Pritch, Y., Cohen-Or, D.: Null-text inversion for editing real images using guided diffusion models. In: CVPR (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.214, + 0.218, + 0.787, + 0.245 + ], + "angle": 0, + "content": "42. Nagai, Y., Taura, T., Mukai, F.: Concept blending and dissimilarity: factors for creative concept generation process. Design studies (2009)" + }, + { + "type": "ref_text", + "bbox": [ + 0.214, + 0.245, + 0.787, + 0.287 + ], + "angle": 0, + "content": "43. Nichol, A., Dhariwal, P., Ramesh, A., Shyam, P., Mishkin, P., McGrew, B., Sutskever, I., Chen, M.: Glide: Towards photorealistic image generation and editing with text-guided diffusion models. arXiv preprint arXiv:2112.10741 (2021)" + }, + { + "type": "ref_text", + "bbox": [ + 0.214, + 0.287, + 0.787, + 0.314 + ], + "angle": 0, + "content": "44. Nobari, A.H., Rashad, M.F., Ahmed, F.: Creativegan: Editing generative adversarial networks for creative design synthesis. arXiv preprint arXiv:2103.06242 (2021)" + }, + { + "type": "ref_text", + "bbox": [ + 0.214, + 0.315, + 0.787, + 0.384 + ], + "angle": 0, + "content": "45. Oquab, M., Darcet, T., Moutakanni, T., Vo, H., Szafraniec, M., Khalidov, V., Fernandez, P., Haziza, D., Massa, F., El-Nouby, A., Assran, M., Ballas, N., Galuba, W., Howes, R., Huang, P.Y., Li, S.W., Misra, I., Rabbat, M., Sharma, V., Synnaeve, G., Xu, H., Jegou, H., Mairal, J., Labatut, P., Joulin, A., Bojanowski, P.: Dinov2: Learning robust visual features without supervision (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.214, + 0.384, + 0.787, + 0.412 + ], + "angle": 0, + "content": "46. Qiao, T., Zhang, J., Xu, D., Tao, D.: Mirrorgan: Learning text-to-image generation by redescription. In: CVPR (2019)" + }, + { + "type": "ref_text", + "bbox": [ + 0.214, + 0.412, + 0.787, + 0.438 + ], + "angle": 0, + "content": "47. Qu, Z., Xiang, T., Song, Y.Z.: Sketchdreamer: Interactive text-augmented creative sketch ideation. In: BMVC (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.214, + 0.439, + 0.787, + 0.481 + ], + "angle": 0, + "content": "48. Radford, A., Kim, J.W., Hallacy, C., Ramesh, A., Goh, G., Agarwal, S., Sastry, G., Askell, A., Mishkin, P., Clark, J., et al.: Learning transferable visual models from natural language supervision. In: ICML (2021)" + }, + { + "type": "ref_text", + "bbox": [ + 0.214, + 0.481, + 0.787, + 0.522 + ], + "angle": 0, + "content": "49. Ramesh, A., Dhariwal, P., Nichol, A., Chu, C., Chen, M.: Hierarchical text-conditional image generation with clip latents. arXiv preprint arXiv:2204.06125 (2022)" + }, + { + "type": "ref_text", + "bbox": [ + 0.214, + 0.522, + 0.787, + 0.578 + ], + "angle": 0, + "content": "50. Ren, T., Liu, S., Zeng, A., Lin, J., Li, K., Cao, H., Chen, J., Huang, X., Chen, Y., Yan, F., Zeng, Z., Zhang, H., Li, F., Yang, J., Li, H., Jiang, Q., Zhang, L.: Grounded sam: Assembling open-world models for diverse visual tasks. arXiv preprint arXiv:2401.14159 (2024)" + }, + { + "type": "ref_text", + "bbox": [ + 0.214, + 0.578, + 0.787, + 0.606 + ], + "angle": 0, + "content": "51. Richardson, E., Goldberg, K., Alaluf, Y., Cohen-Or, D.: Conceptlab: Creative generation using diffusion prior constraints. arXiv preprint arXiv:2308.02669 (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.214, + 0.606, + 0.787, + 0.633 + ], + "angle": 0, + "content": "52. Rombach, R., Blattmann, A., Lorenz, D., Esser, P., Ommer, B.: High-resolution image synthesis with latent diffusion models. In: CVPR (2022)" + }, + { + "type": "ref_text", + "bbox": [ + 0.214, + 0.633, + 0.787, + 0.674 + ], + "angle": 0, + "content": "53. Ruiz, N., Li, Y., Jampani, V., Pritch, Y., Rubinstein, M., Aberman, K.: Dreambooth: Fine tuning text-to-image diffusion models for subject-driven generation. In: CVPR (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.214, + 0.674, + 0.787, + 0.702 + ], + "angle": 0, + "content": "54. Runco, M.A., Jaeger, G.J.: The standard definition of creativity. Creativity research journal (2012)" + }, + { + "type": "ref_text", + "bbox": [ + 0.214, + 0.702, + 0.787, + 0.758 + ], + "angle": 0, + "content": "55. Saharia, C., Chan, W., Saxena, S., Li, L., Whang, J., Denton, E.L., Ghasemipour, K., Gontijo Lopes, R., Karagol Ayan, B., Salimans, T., et al.: Photorealistic text-to-image diffusion models with deep language understanding. In: NeurIPS. vol. 35 (2022)" + }, + { + "type": "ref_text", + "bbox": [ + 0.214, + 0.758, + 0.787, + 0.785 + ], + "angle": 0, + "content": "56. Sbai, O., Elhoseiny, M., Bordes, A., LeCun, Y., Couprie, C.: Design: Design inspiration from generative networks. In: ECCVW (2019)" + }, + { + "type": "ref_text", + "bbox": [ + 0.214, + 0.785, + 0.787, + 0.814 + ], + "angle": 0, + "content": "57. Sun, P., Chen, S., Luo, P.: Grounded segment anything: From objects to parts. https://github.com/Cheems-Seminar/grounded-segment-any-parts (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.214, + 0.814, + 0.787, + 0.842 + ], + "angle": 0, + "content": "58. Sun, P., Chen, S., Zhu, C., Xiao, F., Luo, P., Xie, S., Yan, Z.: Going denser with open-vocabulary part segmentation. arXiv preprint arXiv:2305.11173 (2023)" + }, + { + "type": "list", + "bbox": [ + 0.214, + 0.147, + 0.788, + 0.842 + ], + "angle": 0, + "content": null + } + ], + [ + { + "type": "page_number", + "bbox": [ + 0.218, + 0.116, + 0.236, + 0.127 + ], + "angle": 0, + "content": "18" + }, + { + "type": "header", + "bbox": [ + 0.272, + 0.115, + 0.368, + 0.129 + ], + "angle": 0, + "content": "KW. Ng et al." + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.147, + 0.785, + 0.175 + ], + "angle": 0, + "content": "59. Tao, M., Tang, H., Wu, F., Jing, X.Y., Bao, B.K., Xu, C.: Df-gan: A simple and effective baseline for text-to-image synthesis. In: CVPR (2022)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.176, + 0.785, + 0.203 + ], + "angle": 0, + "content": "60. Vinker, Y., Voynov, A., Cohen-Or, D., Shamir, A.: Concept decomposition for visual exploration and inspiration. In: SIGGRAPH Asia (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.204, + 0.785, + 0.231 + ], + "angle": 0, + "content": "61. Voynov, A., Chu, Q., Cohen-Or, D., Aberman, K.: P+: Extended textual conditioning in text-to-image generation. arXiv preprint arXiv:2303.09522 (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.232, + 0.785, + 0.258 + ], + "angle": 0, + "content": "62. Wah, C., Branson, S., Welinder, P., Perona, P., Belongie, S.: The Caltech-UCSD Birds-200-2011 Dataset (2011)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.259, + 0.785, + 0.286 + ], + "angle": 0, + "content": "63. Wang, X., Darrell, T., Rambhatla, S.S., Girdhar, R., Misra, I.: Instancediffusion: Instance-level control for image generation. arXiv preprint arXiv:2402.03290 (2024)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.287, + 0.785, + 0.327 + ], + "angle": 0, + "content": "64. Wei, Y., Zhang, Y., Ji, Z., Bai, J., Zhang, L., Zuo, W.: Elite: Encoding visual concepts into textual embeddings for customized text-to-image generation. In: ICCV (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.328, + 0.785, + 0.355 + ], + "angle": 0, + "content": "65. Wilkenfeld, M.J., Ward, T.B.: Similarity and emergence in conceptual combination. Journal of Memory and Language (2001)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.356, + 0.785, + 0.383 + ], + "angle": 0, + "content": "66. Xie, J., Li, Y., Huang, Y., Liu, H., Zhang, W., Zheng, Y., Shou, M.Z.: Boxdiff: Text-to-image synthesis with training-free box-constrained diffusion. In: ICCV (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.384, + 0.785, + 0.424 + ], + "angle": 0, + "content": "67. Xu, T., Zhang, P., Huang, Q., Zhang, H., Gan, Z., Huang, X., He, X.: Attngan: Fine-grained text to image generation with attentional generative adversarial networks. In: CVPR (2018)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.425, + 0.785, + 0.466 + ], + "angle": 0, + "content": "68. Yang, B., Gu, S., Zhang, B., Zhang, T., Chen, X., Sun, X., Chen, D., Wen, F.: Paint by example: Exemplar-based image editing with diffusion models. In: CVPR (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.467, + 0.785, + 0.494 + ], + "angle": 0, + "content": "69. Yang, Z., Wang, J., Gan, Z., Li, L., Lin, K., Wu, C., Duan, N., Liu, Z., Liu, C., Zeng, M., et al.: Reco: Region-controlled text-to-image generation. In: CVPR (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.495, + 0.785, + 0.522 + ], + "angle": 0, + "content": "70. Yin, G., Liu, B., Sheng, L., Yu, N., Wang, X., Shao, J.: Semantics disentangling for text-to-image generation. In: CVPR (2019)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.522, + 0.785, + 0.549 + ], + "angle": 0, + "content": "71. Zhang, L., Rao, A., Agrawala, M.: Adding conditional control to text-to-image diffusion models. In: ICCV (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.55, + 0.785, + 0.577 + ], + "angle": 0, + "content": "72. Zhu, M., Pan, P., Chen, W., Yang, Y.: Dm-gan: Dynamic memory generative adversarial networks for text-to-image synthesis. In: CVPR (2019)" + }, + { + "type": "list", + "bbox": [ + 0.217, + 0.147, + 0.785, + 0.577 + ], + "angle": 0, + "content": null + } + ] +] \ No newline at end of file diff --git a/2024/PartCraft_ Crafting Creative Objects by Parts/4747571e-3f49-4aac-b94c-924e39815a4f_origin.pdf b/2024/PartCraft_ Crafting Creative Objects by Parts/4747571e-3f49-4aac-b94c-924e39815a4f_origin.pdf new file mode 100644 index 0000000000000000000000000000000000000000..29461b5a6d0e3c66f943c5de9a3ae7ff4aa8118b --- /dev/null +++ b/2024/PartCraft_ Crafting Creative Objects by Parts/4747571e-3f49-4aac-b94c-924e39815a4f_origin.pdf @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:4074199f0fe10bd6aa91a8e90885d631efc0ccdbc741ffd9f83328288a26b061 +size 6693439 diff --git a/2024/PartCraft_ Crafting Creative Objects by Parts/full.md b/2024/PartCraft_ Crafting Creative Objects by Parts/full.md new file mode 100644 index 0000000000000000000000000000000000000000..d5213a754d92648feffdf72b4adf7ca50e16c2c7 --- /dev/null +++ b/2024/PartCraft_ Crafting Creative Objects by Parts/full.md @@ -0,0 +1,310 @@ +# PartCraft: Crafting Creative Objects by Parts + +Kam Woh $\mathrm{Ng}^{1}$ , Xiatian Zhu $^{1,2}$ , Yi-Zhe Song $^{1}$ , and Tao Xiang $^{1}$ + +$^{1}$ CVSSP, University of Surrey, United Kingdom + +Surrey Institute for People-Centred AI + +{kamwoh.ng, xiatian.zhu, y song, t.xiang}@surrey.ac.uk + +Abstract. This paper propels creative control in generative visual AI by allowing users to "select". Departing from traditional text or sketch-based methods, we for the first time allow users to choose visual concepts by parts for their creative endeavors. The outcome is fine-grained generation that precisely captures selected visual concepts, ensuring a holistically faithful and plausible result. To achieve this, we first parse objects into parts through unsupervised feature clustering. Then, we encode parts into text tokens and introduce an entropy-based normalized attention loss that operates on them. This loss design enables our model to learn generic prior topology knowledge about object's part composition, and further generalize to novel part compositions to ensure the generation looks holistically faithful. Lastly, we employ a bottleneck encoder to project the part tokens. This not only enhances fidelity but also accelerates learning, by leveraging shared knowledge and facilitating information exchange among instances. Visual results in the paper and supplementary material showcase the compelling power of PartCraft in crafting highly customized, innovative creations, exemplified by the "charming" and creative birds in Fig. 1. Code is released at https://github.com/kamwoh/partcraft. + +Keywords: Part Composition $\cdot$ Controllable Text-to-image Generation + +# 1 Introduction + +Humans are creators; AI, on the other hand, hallucinates. Creativity, arguably, is the very force driving humanity forward. Recently, generative AI has garnered considerable attention for its perceived "creativity" [6, 19, 20, 43, 47, 49, 52, 55]. Despite its promise, the challenge of control has swiftly surfaced - how can humans infuse their creativity into the generation process and regulate the extent to which AI hallucinates? + +Creativity starts with an idea. The immediate challenge is how to articulate that idea and integrate it with generative AI. Text is the most commonly employed medium. For instance, imagine being a bird enthusiast wanting to craft the most unique bird akin to Fig. 1. The go-to approach would be furnishing Stable Diffusion [52] (or an equivalent model) with the following textual prompt (the idea): "generate a bird with the head of $X$ , wings of $Y$ , and body of $Z$ ". While you might be presented with remarkably looking birds like Fig. 1(a), they + +![](images/e5c3dba8567cb6b6bdad61f2d12b18caababc60418696b678e4cb1bb0f2171d4.jpg) +Fig. 1: People often form creative concepts based on existing ones [9, 42, 54, 65]. For instance, a bird enthusiast may want to craft a unique bird with different parts (e.g., heads, bodies, and wings) from common bird types (e.g., blue jay, cardinal and song sparrow). (a) Using text prompts in T2I models often results in a lack of control and deviation from the intended details, especially those visual details that are difficult to describe. (b) While sketching is a direct way, not everyone possesses the ability to sketch, particularly in intricate detail. (c) Even with reference images, existing methods (e.g., DreamBooth [53]) did not consider learning object parts, thus cannot generate with desired parts. (d) Using additional control is even cumbersome, requiring too many inputs! (e) We aim to create an object by simply selecting desired parts. PartCraft learns from visual examples to generate the object with a faithful holistic structure, seamlessly integrating the chosen parts into a natural and coherent entity. + +may bear little resemblance to the envisioned concept. Recent literature suggests that a swift sketch could serve as a viable alternative [40, 63, 71], providing fine-grained shape control. However, the caveat is that not everyone possesses the ability to sketch, particularly in intricate detail. + +In this paper, our primary focus revolves around addressing the issue of "control" in generative AI. We advance by introducing fine-grained control into the generative process, inviting you to "select". While this selection mechanism might seem modest at first, it closely mirrors the human creative process, where new concepts often emerge from existing ones [9,42,54,65]. Recall those moments when you desired an "ideal" pair of shoes with selected features from different pairs, or when you aimed to get creative with a cat (for that matter!). + +It follows that rather than relying on writing (text) or drawing (sketch), all that is required is to choose the distinct visual concepts you specifically desire in your creative endeavor. Our model then ensures that all selected concepts are seamlessly and precisely composed into a faithful novel object in the final generation, without resorting to additional control such as bounding boxes [7,38, 40,66,71]. To illustrate with the "unique bird" example once more, our approach literally involves selecting the head of X, wings of Y, and body of Z! (see Fig. 1(e) and 2). + +Our solution is intuitive and centers around the well-studied computer vision concept of objects and their parts [8,14,22,27,35]. The challenge then boils down + +![](images/4cf0ceb7b2a760d2da60d869c667e63255c96c0b5a3360d3a537ac5cc3e6e273.jpg) +Fig. 2: Two sets of images were generated from their original parts (sources A and B). We can integrate a specific part (e.g., body, head, or even background (BG) of source B to target A seamlessly without effort. + +to two aspects: (i) how to parse known objects into their parts (i.e., recognizing that birds have eyes and tails), and (ii) how to assemble parts from different visual concepts to form a faithful creative concept (i.e., ensuring the output is recognizably a bird). + +The former is easier. We make clever use of DINOv2 [45] feature maps and perform unsupervised clustering to identify common parts. The idea is that each cluster will then correspond to a semantic part of a common object (e.g., head and wings of birds). We specifically chose DINOv2 for its superior fine-grained perception [2] compared to others [50, 57, 58]. Further, this way has a higher flexibility to enhance fine-grained parsing by using a higher cluster count. + +Our major contribution lies in addressing the latter challenge. The solution is intuitive - it essentially revolves around the fine-grained selection and placement of the chosen parts. Inspired by recent efforts in personalization [3,25,36,53,64], primarily designed for learning entire objects, we introduce a tailored attention loss that specifically operates on parts. With this loss, our model learns the object part composition, ensuring the final generation appears holistically faithful with mixed parts (i.e., head and wings of a bird actually appear at the right places). + +More specifically, we introduce an entropy-based attention loss that maximizes the attention of a specific part at a particular location while minimizing the attention where no parts appear. This is achieved by first selecting the attention maps that correspond to the parts. A normalization over all parts is then performed to ensure that each image region is occupied by no more than one part. Finally, we minimize the entropy loss between each normalized individual part and the semantic maps obtained during object parsing, containing the correct part location. This not only facilitates stronger part disentanglement but is also the key to generating a faithful holistic structure of an object as it learns a generic prior topology knowledge about object parts. + +To enhance generation fidelity, we further employ a bottleneck encoder to project the text tokens. This approach accelerates learning by leveraging shared knowledge (common parts) and facilitating information exchange among instances in each part. Each instance adjusts slightly to adapt to the fine-grained part details during optimization. + +Our contributions are as follows: (i) We pioneer a unique approach for fine-grained part-level control in Text-to-Image (T2I) models, empowering users to craft objects by selecting desired parts. This method streamlines the creative process, marking a significant advancement in our capacity to manipulate and reimagine visual content. (ii) We introduce PartCraft, a technique that autonomously parses object parts and orchestrates them from different visual concepts, resulting in the faithful creation of a novel object. (iii) For enhanced part disentanglement and generation fidelity, we propose an entropy-based normalized attention loss and leverage a bottleneck encoder. (iv) We present two problem-specific quantitative metrics. Comprehensive experiments on CUB-200-2011 (birds) and the Stanford Dogs dataset demonstrate the superior performance of our method in generating novel objects, surpassing alternative approaches in both qualitative and quantitative evaluations. + +# 2 Related Work + +Creative editing and generation. Creativity involves generating innovative ideas or artifacts across various domains [12]. Extensive research has explored the integration of creativity into Generative Adversarial Networks (GANs) [21,44,56] and Variational Autoencoders (VAEs) [16, 18]. For example, DoodlerGAN [26] learns and combines fine-level part components to create sketches of new species. A recent study by [60] demonstrated decomposing personalized concepts into distinct visual aspects, creatively recombined through diffusion models. Instruct-Pix2Pix [10] allows creative image editing through instructions, while ConceptLab [51] aims to identify novel concepts within a specified category, deviating from existing concepts. Different from these works where editing/generation usually focuses on the whole object, we instead focus on training a text-to-image generative model that can understand parts, thus able to creatively generate new objects by seamlessly composing different parts simply through selection. + +Text-to-image generation. Recent advancements in large text-to-image (T2I) diffusion models [6, 19, 20, 43, 49, 52, 55] have made significant improvements over conventional methods [37, 46, 59, 67, 70, 72] in producing high-fidelity images from text prompts. Their application scope has expanded to include both global [10, 32, 41] and localized [4, 17, 28, 68] image editing tasks, demonstrating versatility. Methods such as [5, 7, 13, 15, 23, 24, 34, 38, 66, 69, 71] have introduced more granular spatial control, such as incorporating semantic segmentation masks or bounding boxes, into large pretrained diffusion models to guide image generation. Contrary to these approaches, we focus on enhancing the control by enabling a straightforward discrete selection of desired parts. We minimize the complexity and manual intervention required by spatial controls, yet the model can compose selected parts as a coherent object seamlessly. This not + +![](images/07aab029c3b582c0c91383aecfd2e6532c96a881fbc5b46150f13f502b722ee8.jpg) +Fig. 3: Overview of our PartCraft. All parts are organized into a dictionary, and their semantic embeddings are learned through a textual inversion approach. For instance, a text description like "a photo of a [Head,42] [Wing,87]..." guides the optimization of the corresponding textual embedding by reconstructing the associated image. To improve generation fidelity, we incorporate a bottleneck encoder $f$ (MLP) to compute the embedding $y$ (Eq. (4)) as input to the text encoder. To promote disentanglement among learned parts, we minimize a specially designed attention loss, denoted as $\mathcal{L}_{\text{att}}$ . + +only simplifies the user's role in the generative process but also ensures that the compositional logic and coherence are inherently managed by the model's ability. As such, our research is distinguished by focusing on the model's inherent ability to understand and apply part relationships instead of providing additional controls. + +Abstracting visual knowledge as a text token. The effectiveness of T2I models is constrained by the user's ability to articulate their desired image through text. These models face challenges in faithfully replicating visual characteristics from a reference set and generating innovative interpretations in diverse contexts, even with detailed textual descriptions. To address this challenge, various personalization techniques have been developed. These techniques obtain a new word embedding from multiple images depicting the same concept [1, 25, 36, 53, 61] or multiple new word embeddings for various concepts within a single image [3] through inversion. The learned visual concepts can be creatively reused in many image editing tasks. Nonetheless, most of these approaches struggle to learn object parts, often not able to follow the part selections due to part entanglement as they were designed to learn object as a whole. In this work, we introduce a customized attention loss that serves a dual purpose: ensure accurate positioning of each part and enforce each image region occupied by no more than one part. This greatly improves the part disentanglement, creating novel concepts with correct appearances (see Fig. 2). + +# 3 Methodology + +Given a set of unlabeled images depicting the same object (e.g., bird) with different part details, we aim to train a T2I generative model that decomposes + +parts of objects into text tokens and can recompose them in a novel way. To that end, we propose PartCraft, as depicted in Fig. 3. + +We start by discovering the parts in a three-tier hierarchy, as detailed in Sec. 3.1. Paired with the training images $\{x_{i}\}_{i = 1}^{N}$ , this semantic hierarchy subsequently serves as the supervision to fine-tune a pre-trained text-to-image model, say a latent diffusion model [52], denoted as $\{\epsilon_{\theta},\tau_{\theta},\mathcal{E},\mathcal{D}\}$ , where $\epsilon_{\theta}$ represents the diffusion denoiser, $\tau_{\theta}$ the text encoder, and $\mathcal{E} / \mathcal{D}$ the autoencoder respectively. We adopt the textual inversion technique [25]. Concretely, we learn a set of pseudo-words $p^*$ for each part in the word embedding space with: + +$$ +\mathcal {L} _ {l d m} = \mathbb {E} _ {z, t, p, \epsilon} \left[ | | \epsilon - \epsilon_ {\theta} \left(z _ {t}, t, \tau_ {\theta} (y _ {p})\right) | | _ {2} ^ {2} \right], \tag {1} +$$ + +$$ +p ^ {*} = \underset {p} {\operatorname {a r g m i n}} \mathcal {L} _ {l d m}, \tag {2} +$$ + +where $\epsilon \sim \mathcal{N}(0,1)$ denotes the unscaled noise, $t$ is the time step, $z = \mathcal{E}(x)$ is the latent representation of the image, $z_{t}$ is the latent noise at time $t$ , and $y_{p}$ is the text condition that includes $p$ as part of the text tokens. $\mathcal{L}_{ldm}$ is a standard diffusion loss [30] to reconstruct the parts. As each object is composed of a set of parts, its reconstruction is achieved by the reconstruction of the associated set of parts. In other words, when all parts are reconstructed properly, it will become a valid object. + +# 3.1 Unsupervised Part Discovery + +![](images/7cdf1b8f24fb6024e9ccc6323e592f8e95ce299405de4bc25888338d91dd2b32.jpg) +Fig. 4: Part discovery within a semantic hierarchy involves partitioning each image into distinct parts and forming semantic clusters across unlabeled training data. + +To minimize the labeling cost, we develop a scalable process to reveal the underlying semantic hierarchy with parts in an unsupervised fashion. We leverage the off-the-shelf vision model for image decomposition and clustering. Specifically, given an image $x_{i}$ , we employ DINOv2 [45] to extract the feature map $F = \{F_{i} = \mathrm{DINOv2}(x_{i})\}_{i}^{N}$ . We then conduct three-level hierarchical clustering (see Fig. 4): (i) At the top level, $k$ -means is applied on all patches in $F$ with $k = 2$ to separate foregrounds and backgrounds $B$ . (ii) At the middle level, $k$ -means is further ap + +plied on all foreground patches to acquire $M$ clusters representing common parts, such as the heads of birds. (iii) At the bottom level, we further group each of the $M$ clusters as well as the background cluster $B$ into $K$ splits. Each split refers to finer meanings, such as the head of a specific bird species, or a specific background style. Lastly, each region of an image will be tagged with the corresponding cluster index. We represent these cluster tags as follows: + +$$ +p = (0, k _ {0}), (1, k _ {1}), \dots , (M, k _ {M}), \tag {3} +$$ + +where the first pair refers to the background style, and the following $M$ pairs denote the combinations of $M$ parts (e.g., head, body, wings) each associated with a specific object (e.g., sparrow), and $k \in \{1, \dots, K\}$ . This description will be used as the textual prompt in model training, such as "a photo of a $[p]$ ". Please refer to the supplementary material for more examples of the discovered semantic hierarchy. This process also yields the segmentation mask of each $m$ -th part, which we define as $S_{m}$ . + +Motivation. While we can leverage off-the-shelf segmentation models such as VLPart [58], the robustness relies on the generalizability of the model and the part segmentation result is usually pre-defined and may be unstable for unseen domains. As a result, we rely on our feature clustering method to obtain the segmentation map, which also has a higher flexibility in choosing the number of clusters (parts). + +# 3.2 Part Token Bottleneck + +In contrast to prior text inversion studies [25], our task requires learning a greater quantity at the same time - specifically, $(M + 1)K$ -of word tokens derived from a collection of discovered parts marked by inherent imperfections (such as partial overlap and over splitting). This makes the learning task more demanding. To enhance the learning process, we propose a neural network $f(\cdot)$ comprising a two-layer MLP with ReLU activation: + +$$ +y _ {p} = f (e (p)), \tag {4} +$$ + +where $y_{p}$ will be subsequently used as the input3 to the text encoder $\tau_{\theta}$ and $e \in \mathbb{R}^{MK \times D}$ is a learnable word embedding dictionary that maps $p$ to their respective embeddings. + +Our design demonstrates quicker convergence than directly learning the final word embeddings $e(\cdot)$ [25] (see supplementary material). This could be attributed to the entanglement of word embeddings in the conventional design, where there is no information exchange among them during optimization. For instance, each token doesn't know they are learning for a specific part of a specific species. This lack of communication leads to lower data efficiency and slower learning. With the bottleneck $f$ , it will first project the token into a common part embedding space (e.g., head), then slightly adjust itself to adapt the fine-grained part details. It's worth noting that the conventional design is a specific instance of our approach when $f$ is an identity function. + +# 3.3 Learning to Craft by Parts + +Fine-tuning the T2I model, rather than sorely learning pseudo-words, has been shown to achieve better reconstruction of target concepts as demonstrated in [36, 53]. However, this comes with a significant training cost. Thus, we apply LoRA + +(low-rank adaptation) [31] to the cross-attention block for efficient training. We then minimize the diffusion loss $\mathcal{L}_{ldm}$ (Eq. (1)) to learn both pseudo-words and LoRA adapters. + +While training with only $\mathcal{L}_{ldm}$ , entanglement happens between parts, as evident from the attention maps in the cross-attention block of the denoiser $\epsilon_{\theta}$ (see Fig. 9). This entanglement arises due to the correlation between parts (e.g., a bird head code is consistently paired with a bird body code to represent the same species). To address this issue, we introduce an entropy-based attention loss as regularization: + +$$ +\mathcal {L} _ {\text {a t t n}} = \mathbb {E} _ {z, t, m} \left[ - \left(S _ {m} \log \hat {A} _ {m} + (1 - S _ {m}) \log (1 - \hat {A} _ {m})\right) \right], \tag {5} +$$ + +$$ +\hat {A} _ {m, i, j} = \frac {\bar {A} _ {m , i , j}}{\sum_ {k} \bar {A} _ {k , i , j}}, \quad \bar {A} _ {m} = \frac {1}{L} \sum_ {l} ^ {L} A _ {l, m}, \tag {6} +$$ + +where $A \in [0,1]^{M \times HW}$ represents the cross-attention map between the $m$ -th part and the noisy latent $z_{t}$ , $L$ represents the number of specific layers to select attention maps, $\hat{A} \in [0,1]^{M \times HW}$ represents the averaged and normalized cross-attention map over all parts and $S_{m} \in \{0,1\}^{M \times HW}$ serves as the mask that indicates the location of $m$ -th part. In cases where the part is not present in the image (e.g., occluded), we set both $S_{m}$ and $\hat{A}_{m}$ as 0 to exclude them. Thus, the overall learning objective is defined as: + +$$ +\mathcal {L} _ {\text {t o t a l}} = \mathcal {L} _ {l d m} + \lambda_ {\text {a t t n}} \mathcal {L} _ {\text {a t t n}}, \tag {7} +$$ + +where $\lambda_{attn} = 0.01$ . We focus on attention maps at the resolution of $16 \times 16$ where rich semantic information is captured [28]. Normalization is performed at each location to ensure that the sum of a patch location equals 1. This aims to maximize the attention of a specific part at a particular location which implicitly minimizing the attention of other parts similar to a softmax classification task. Compared to the mean-square based attention loss [3], this intuitively ensures that a part only appears once at a particular location, facilitating stronger disentanglement from other parts during the denoising operation. When generating a part for a particular location, the diffusion model $\epsilon_{\theta}$ should only attend to the part instead of other non-related parts. + +# 4 Experiments + +Datasets. We demonstrate our selection task on two fine-grained object datasets: CUB-200-2011 (birds) [62] which contains 5,994 training images, and Stanford Dogs [33] which contains 12,000 training images. + +Implementation. For part composition, we assess the model's ability to combine up to 4 different parts from 4 distinct species/objects. We set $M = 5$ for bird generation (head, front body/breast area, wings, legs, tail) and $M = 7$ for dog generation (forehead, eyes, mouth/nose, ears, neck, body/tail, legs). For both datasets, $K$ is set as 256, ensuring sufficient coverage of all fine-grained + +
MethodLearnable TokenFine TuneDisentanglementBottleneck
Textual Inversion [25]✓✗✗✗
DreamBooth [53]✓LoRA*✗✗
CustomDiffusion [36]✓K/V✗✗
Break-a-scene [3]✓LoRA*MSE✗
PartCraft (Ours)✓LoRAEq. (5)✓
+ +Table 1: Comparing our and alternative methods in design properties. *: We finetuned the added LoRA [31] adapter rather than the entire diffusion model $\epsilon_{\theta}$ due to resource limit. MSE is a mean-square based attention loss used in [3]. + +classes (i.e., 200 for birds and 120 for dogs). We randomly generate 500 images by sampling 500 sets of parts. For each image, we randomly replace an original part with any part from another 500 non-overlapping sets of parts. The resulting set of parts may take the form of $\left( {0,{k}_{A}}\right) \left( {1,{k}_{B}}\right) \left( {2,{k}_{C}}\right) \ldots \left( {M,{k}_{D}}\right)$ ,representing a composition from species A,B,C,and D. Stable Diffusion v1.5 [52] is used. Please see the supplementary material for further training details. + +We compare our method with the recent personalization methods: Textual Inversion (TI) [25], DreamBooth (DB) [53], Custom Diffusion (CD) [36], Break-a-scene (BaS) [3]. These personalization methods were designed to take single or multiple images with associated labeled objects as input. The text prompt for each image is as simple as "a photo of $[p]$ " where $p$ is expressed in Eq. (3), since we do not rely on complex prompts. We employ the official implementations released by the original authors for training. We summarize the main design properties of all compared methods in Tab. 1. + +Evaluation metrics. To assess a model's ability to disentangle and composite parts, we introduce two metrics: (a) exact matching rate (EMR) and (b) cosine similarity (CoSim) between the $k$ -means embeddings of the parts of real and generated images. Utilizing the pre-trained $k$ -means from Sec. 3.1, we predict the parts of generated images. EMR quantifies how accurately the cluster index of parts of generated images matches the parts of the corresponding real images whereas CoSim measures the cosine similarity between the $k$ -means centroid vector that the part belongs to between generated and real images. These metrics assess the model's ability to follow the input parts and accurately reconstruct them, with perfect disentanglement indicated by EMR of 1 and CoSim of 1. A detailed algorithm is provided in the supplementary material. We also measure image generation quality using FID [29] to assess model performance in terms of image distribution. Additionally, we compute the average pairwise cosine similarity between CLIP [48]/DINO [11] embeddings of generated and real class-specific images following [53]. Each generated image is conditioned on the parts of the corresponding real image. This results in 5,994 generated images for birds and 12,000 generated images for dogs. + +
MethodBirds: CUB-200-2011Dogs: Stanford Dogs
FIDCLIPDINOEMRCoSimFIDCLIPDINOEMRCoSim
Textual Inversion [25]10.100.7840.6070.3050.84223.360.6520.5320.2180.754
DreamBooth [53]12.940.7750.5940.3550.85622.650.6600.5630.2750.777
Custom Diffusion [36]37.610.6940.5040.3380.83342.410.5930.4910.2530.755
Break-a-Scene [3]20.050.7420.5490.3900.85424.200.6330.5320.3000.775
PartCraft (Ours)12.860.7830.6180.4600.88216.920.6690.5730.3580.796
+ +Table 2: Quantitative comparison for part reconstruction. + +![](images/2461d17a447063d53bc670ae77ba2c86222d1a10e0b4f5ded508227952c9bdca.jpg) +(a) TI + +![](images/525a983d3e6f635dbc54686bdefc5eba78e2170f720894cf71cbe0bfcc2c2c58.jpg) +(b) DB + +![](images/3797cf22eda737f60fc7102a73197117e5728876b1aedb06ed9568e3e2bd3579.jpg) +(c) CD + +![](images/c83aeab8a8fb6ae0ca334696b241f93777adbfb9ffd4947b627896e96b631219.jpg) +(d) BaS +Fig. 5: Visual comparison under the part reconstruction setting. All images are generated by using the original parts of respective objects. + +![](images/76a89cfde8e5a63f6a92ff9c9eaaa58b2e37f1cb6a11a5c61dd7216815adbfb7.jpg) +(e) PartCraft + +# 4.1 Part Reconstruction and Image Quality Evaluation + +We first assess the ability of different methods to learn parts as text tokens by evaluating how well they can accurately reconstruct the parts (this also means image generation with original parts). + +In Tab. 2, we summarize the performance of respective methods on the bird and dog generation, respectively. We highlight four observations: (i) Textual Inversion performs quite well compared to DreamBooth, CustomDiffusion, and Break-a-scene in terms of FID, CLIP, and DINO scores although did not finetune the diffusion model $\epsilon_{\theta}$ . This may be due to the potential risk of overfitting when fine-tuning $\epsilon_{\theta}$ especially when learning a vast array of new concepts with many update iterations. It is also not uncommon to carefully tune the learning rate and the training iterations in these models when fine-tuning new concepts (e.g., only 800-1000 steps of updates to learn a new concept in [3]). (ii) Nonetheless, fine-tuning the diffusion model $\epsilon_{\theta}$ can help improve the ability to follow prompts as shown by increased EMR and CoSim scores (e.g., EMR of at least $5\%$ in DreamBooth). (iii) Break-a-scene has a better ability to reconstruct the parts as shown by EMR and CoSim, this is due to the attention loss explicitly forcing the parts to focus on the respective semantic region. (iv) PartCraft achieves the best performance in DINO, EMR, and CoSim scores (e.g., $7\%$ better in EMR compared to Break-a-scene). This indicates that not only does our + +![](images/cb410159b384692c2139554d69c2674923f06e73f46b6f4d561006ff1e811c7f.jpg) +Fig. 6: Quantitative comparisons of part composition in terms of EMR and CoSim. + +image-generation ability perform comparably well with textual inversion, but PartCraft is also able to disentangle the parts learning so that it can follow the prompt instructions more accurately to generate the parts in a cohort. In Fig. 5, we present generated images from different methods, with CustomDiffusion exhibiting high-contrast images, possibly due to unconstrained fine-tuning on the cross-attention components $K / V$ and resulting in worse FID scores. + +# 4.2 Part Composition Evaluation + +In this section, we assess the part composition ability of different methods. In this experiment, we generate the image by mixing different parts from different species. Our findings, as shown in Fig. 6, can be summarized as follows: (i) As the number of composited parts increases, EMR and CoSim decrease, reflecting the challenge of composing multiple diverse parts. (ii) Break-a-scene and PartCraft achieve notably higher EMR and CoSim scores, thanks to disentanglement through attention loss minimization. (iii) PartCraft outperforms Break-a-scene significantly by token bottleneck and tailored attention loss. + +In Fig. 7, we visualize the results of composing 4 different parts. While all images appear realistic, most methods struggle to assemble all 4 parts. For instance, Break-a-scene missed out on the flurry body of kerry blue terrier (right-most column). In contrast, our methods successfully combine 4 different parts from 4 different species, demonstrating the superior ability of our approach to part composition. We also visualize additional examples of our method in the supplementary material. + +Furthermore, we explore the versatility of the adapted model by generating images with simple styles such as pencil drawing. While most methods successfully incorporate specific styles into the generated image, Custom Diffusion often fails to do so, possibly due to the unconstrained fine-tuning of the cross-attention components $K / V$ . + +# 4.3 Ablation Studies + +Component analysis. In Fig. 8, we evaluate the effect of our proposed components (token bottleneck and attention loss) on creating novel bird species. (i) Removing the bottleneck outlined in Eq. (4) degrades the generation quality as evidenced by a higher FID score $(12.86 \rightarrow 16.36)$ even though both EMR + +![](images/258f2c1e1980ccfb10ad38c91986c23f8aabb54eaea0a085f6905dca02ef4683.jpg) +Fig. 7: Visual comparison on 4-species (specified on the top row) mixed generation. The last column indicates generated images with different styles (i.e., $DSLR$ , Van Gogh, Oil Painting, Pencil Drawing). + +and CoSim remain. (ii) By replacing our $\mathcal{L}_{attn}$ with the MSE loss as proposed in [3], we observe significant deterioration in both EMR and CoSim. (iii) Finally, incorporating both the projector and our attention loss performs the best. This improvement highlights the necessity of incorporating interactions between multiple parts to achieve more effective part disentanglement and optimization. + +Cross-attention visualization. Our attention loss plays a crucial role in token disentanglement. We demonstrate the impact of this loss in Fig. 9, where we + +![](images/d489f73860ad7af2326e6c5eff7b9f787e900d76288dc443dfe88408a3be9928.jpg) +Fig.9: Cross-attention map of each part (top) without and (bottom) with our attention loss. + +![](images/c6d1fbb7d14a0a6f64c8532f51b664666b768c548fc6106232cf540a2a5d45c8.jpg) +Fig.8: Ablation on our part token bottleneck and attention loss under the part composition on CUB-200-2011 birds. + +![](images/0880ca23d405fb532d96c74a65064acb8cd1bf1482e8482d2a97615297a63633.jpg) +(a) Textual Inversion + +![](images/e1c58d70a21aa6c21e133b0b69ca32b989a2437b162aab0c753b51860fc799c3.jpg) +(b) Break-a-scene + +![](images/313f6382013f8b85457e6d24af5d5f9acb802c9bc491b7cbddfa00cfce25161e.jpg) +(c) PartCraft (Ours) +Fig. 10: 2D tSNE [39] projection of word embeddings. Different colors represent different common parts. (correspond to the segmentation mask in Fig. 4). + +observe significantly enhanced disentanglement after explicitly guiding attention to focus on distinct semantic regions. + +Part word embedding space. We visualize the word embeddings of learned tokens of Textual Inversion [25], Break-a-scene [3] and our PartCraft for birds generation (CUB-200-2011 [62]) through tSNE [39] in Fig. 10. In our PartCraft, the word embeddings are the projected embeddings through Eq. (4). We can see that our projected version has a better semantic meaning such that the part embeddings are clustered together by their semantic meaning (e.g., head). We believe this is one of the reasons that our PartCraft outperforms previous methods in which we can compose all parts seamlessly yet with higher quality. + +Transferability for Creativity. (i) In Fig. 11, we demonstrate that not only it can compose parts within the training domain (e.g., birds), but it can also transfer the learned parts to and combine with other domains (e.g., cat). This enables the creation of unique combinations, such as a cat with a dog's ear. (ii) Leveraging the prior knowledge embedded in Stable Diffusion, PartCraft can also repurpose learned parts for creative image generation. An example of this is the generation of a bird-shaped robot adorned with various parts. These examples showcase PartCraft's immense potential for diverse and limitless creative applications. Please see the supplementary material for more examples. + +![](images/1fc2c2041a4f1456f00ab034420aaa44a14a9f89290790e8b24d13acb8381d53.jpg) +Fig. 11: (Top): Using learned parts to modify the property of other domains such as cat, and lion with prompt such as "A cat with [beagle's ear)". (Bottom) We can also repurpose learned parts for creative image generation using prior knowledge in Stable Diffusion with prompt such as "A robot designed inspired by [red header woodpecker's head] and [blue jay's body)". + +# 5 Conclusion + +We propose a new way of control in generative AI. Instead of text or sketch, we "select" desired parts to create an object. We addressed the challenge of learning parts in T2I models by introducing a customized attention loss. This loss serves a dual purpose: to ensure parts are at the right location and to ensure each location is occupied by not more than one part. This greatly improves the part disentanglement. We further employ a non-linear bottleneck encoder to improve generation fidelity. Our model, PartCraft, can seamlessly compose different parts from different objects, creating objects that do not exist yet holistically correct and plausible objects by mixing them. Extensive experiments demonstrated PartCraft's superior performance in both qualitative and quantitative evaluation. Moreover, the learned parts demonstrate strong transferability. We hope that our PartCraft will empower artists, designers, and enthusiasts to bring the creations of their dreams to reality. + +# 6 Limitations and Future Works + +It is worth noting that the accuracy of obtained parts may be affected by using a self-supervised pre-trained feature extractor, incorporating encoders [64] may help. We also observed challenges in composing relatively small parts, like tails and legs. Additionally, we are also exploring cross-domain generation, i.e., combining learned parts from different datasets to create objects with even more diverse parts. For instance, we can merge non-rigid parts (e.g., parts from quadrupled animals) with rigid parts (e.g., parts from cars/airplanes) and form a creative structure (e.g., a car that has horse legs instead of wheels). This not only further improves the applicability of PartCraft but also serves as a stepping stone to creative generation, as generative AI progresses, continually expanding the limits of achievable creativity and artistic expression. + +# Acknowledgements + +We extend our special thanks to Jia Wei Sii for her help in creating figures and discussing the main concept. We are also grateful to Ruoyi Du, Zhiyu Qu, Chee Seng Chan, and the reviewers for their fruitful comments and corrections on our draft, methodology, and experiments. + +# References + +1. Alaluf, Y., Richardson, E., Metzer, G., Cohen-Or, D.: A neural space-time representation for text-to-image personalization. arXiv preprint arXiv:2305.15391 (2023) +2. Amir, S., Gandelsman, Y., Bagon, S., Dekel, T.: On the effectiveness of vit features as local semantic descriptors. In: ECCV (2022) +3. Avrahami, O., Aberman, K., Fried, O., Cohen-Or, D., Lischinski, D.: Break-a-scene: Extracting multiple concepts from a single image. In: SIGGRAPH Asia (2023) +4. Avrahami, O., Fried, O., Lischinski, D.: Blended latent diffusion. ACM TOG (2023) +5. Avrahami, O., Hayes, T., Gafni, O., Gupta, S., Taigman, Y., Parikh, D., Lischinski, D., Fried, O., Yin, X.: Spatext: Spatio-textual representation for controllable image generation. In: CVPR (2023) +6. Balaji, Y., Nah, S., Huang, X., Vahdat, A., Song, J., Kreis, K., Aittala, M., Aila, T., Laine, S., Catanzaro, B., et al.: edifi: Text-to-image diffusion models with an ensemble of expert denoisers. arXiv preprint arXiv:2211.01324 (2022) +7. Bar-Tal, O., Yariv, L., Lipman, Y., Dekel, T.: Multidiffusion: Fusing diffusion paths for controlled image generation. In: ICML (2023) +8. Biederman, I.: Recognition-by-components: a theory of human image understanding. Psychological review (1987) +9. Bonnardel, N., Marmèche, E.: Towards supporting evocation processes in creative design: A cognitive approach. International Journal of Human-Computer Studies (2005) +10. Brooks, T., Holynski, A., Efros, A.A.: Instructpix2pix: Learning to follow image editing instructions. In: CVPR (2023) +11. Caron, M., Touvron, H., Misra, I., Jégou, H., Mairal, J., Bojanowski, P., Joulin, A.: Emerging properties in self-supervised vision transformers. In: ICCV (2021) +12. Cetinic, E., She, J.: Understanding and creating art with ai: Review and outlook. ACM Transactions on Multimedia Computing, Communications, and Applications (2022) +13. Chefer, H., Alaluf, Y., Vinker, Y., Wolf, L., Cohen-Or, D.: Attend-and-excite: Attention-based semantic guidance for text-to-image diffusion models. ACM TOG (2023) +14. Chen, C., Li, O., Tao, D., Barnett, A., Rudin, C., Su, J.K.: This looks like that: deep learning for interpretable image recognition. NeurIPS (2019) +15. Chen, M., Laina, I., Vedaldi, A.: Training-free layout control with cross-attention guidance (2024) +16. Cintas, C., Das, P., Quanz, B., Tadesse, G.A., Speakman, S., Chen, P.Y.: Towards creativity characterization of generative models via group-based subset scanning. In: IJCAI (2022) +17. Couairon, G., Verbeek, J., Schwenk, H., Cord, M.: Diffedit: Diffusion-based semantic image editing with mask guidance. In: ICLR (2023) + +18. Das, P., Quanz, B., Chen, P.Y., Ahn, J.w., Shah, D.: Toward a neuro-inspired creative decoder. In: IJCAI (2020) +19. Ding, M., Yang, Z., Hong, W., Zheng, W., Zhou, C., Yin, D., Lin, J., Zou, X., Shao, Z., Yang, H., et al.: Cogview: Mastering text-to-image generation via transformers. In: NeurIPS (2021) +20. Ding, M., Zheng, W., Hong, W., Tang, J.: Cogview2: Faster and better text-to-image generation via hierarchical transformers. In: NeurIPS. vol. 35 (2022) +21. Elgammal, A., Liu, B., Elhoseiny, M., Mazzone, M.: Can: Creative adversarial networks generating "art" by learning about styles and deviating from style norms. In: ICCC (2017) +22. Felzenszwalb, P.F., Huttenlocher, D.P.: Pictorial structures for object recognition. IJCV (2005) +23. Feng, W., He, X., Fu, T.J., Jampani, V., Akula, A.R., Narayana, P., Basu, S., Wang, X.E., Wang, W.Y.: Training-free structured diffusion guidance for compositional text-to-image synthesis. In: ICLR (2022) +24. Gafni, O., Polyak, A., Ashual, O., Sheynin, S., Parikh, D., Taigman, Y.: Make-ascene: Scene-based text-to-image generation with human priors. In: ECCV (2022) +25. Gal, R., Alaluf, Y., Atzmon, Y., Patashnik, O., Bermano, A.H., Chechik, G., Cohen-Or, D.: An image is worth one word: Personalizing text-to-image generation using textual inversion. In: ICLR (2023) +26. Ge, S., Goswami, V., Zitnick, C.L., Parikh, D.: Creative sketch generation. In: ICLR (2021) +27. He, J., Chen, J., Lin, M.X., Yu, Q., Yuille, A.L.: Compositor: Bottom-up clustering and compositing for robust part and object segmentation. In: CVPR (2023) +28. Hertz, A., Mokady, R., Tenenbaum, J., Aberman, K., Pritch, Y., Cohen-Or, D.: Prompt-to-prompt image editing with cross attention control. In: ICLR (2023) +29. Heusel, M., Ramsauer, H., Unterthiner, T., Nessler, B., Hochreiter, S.: Gans trained by a two time-scale update rule converge to a local nash equilibrium. NeurIPS 30 (2017) +30. Ho, J., Jain, A., Abbeel, P.: Denoising diffusion probabilistic models. In: NeurIPS (2020) +31. Hu, E.J., Shen, Y., Wallis, P., Allen-Zhu, Z., Li, Y., Wang, S., Wang, L., Chen, W.: LoRA: Low-rank adaptation of large language models. In: ICLR (2022) +32. Kawar, B., Zada, S., Lang, O., Tov, O., Chang, H., Dekel, T., Mosseri, I., Irani, M.: Imagic: Text-based real image editing with diffusion models. In: CVPR (2023) +33. Khosla, A., Jayadevaprakash, N., Yao, B., Li, F.F.: Novel dataset for fine-grained image categorization: Stanford dogs. In: CVPRW (2011) +34. Kim, Y., Lee, J., Kim, J.H., Ha, J.W., Zhu, J.Y.: Dense text-to-image generation with attention modulation. In: ICCV (2023) +35. Krause, J., Jin, H., Yang, J., Fei-Fei, L.: Fine-grained recognition without part annotations. In: CVPR (2015) +36. Kumari, N., Zhang, B., Zhang, R., Shechtman, E., Zhu, J.Y.: Multi-concept customization of text-to-image diffusion. In: CVPR (2023) +37. Li, B., Qi, X., Lukasiewicz, T., Torr, P.: Controllable text-to-image generation. In: NeurIPS (2019) +38. Li, Y., Liu, H., Wu, Q., Mu, F., Yang, J., Gao, J., Li, C., Lee, Y.J.: Gligen: Open-set grounded text-to-image generation. In: CVPR (2023) +39. Van der Maaten, L., Hinton, G.: Visualizing data using t-sne. Journal of Machine Learning Research (2008) + +40. Mo, S., Mu, F., Lin, K.H., Liu, Y., Guan, B., Li, Y., Zhou, B.: Freecontrol: Training-free spatial control of any text-to-image diffusion model with any condition. arXiv preprint arXiv:2312.07536 (2023) +41. Mokady, R., Hertz, A., Aberman, K., Pritch, Y., Cohen-Or, D.: Null-text inversion for editing real images using guided diffusion models. In: CVPR (2023) +42. Nagai, Y., Taura, T., Mukai, F.: Concept blending and dissimilarity: factors for creative concept generation process. Design studies (2009) +43. Nichol, A., Dhariwal, P., Ramesh, A., Shyam, P., Mishkin, P., McGrew, B., Sutskever, I., Chen, M.: Glide: Towards photorealistic image generation and editing with text-guided diffusion models. arXiv preprint arXiv:2112.10741 (2021) +44. Nobari, A.H., Rashad, M.F., Ahmed, F.: Creativegan: Editing generative adversarial networks for creative design synthesis. arXiv preprint arXiv:2103.06242 (2021) +45. Oquab, M., Darcet, T., Moutakanni, T., Vo, H., Szafraniec, M., Khalidov, V., Fernandez, P., Haziza, D., Massa, F., El-Nouby, A., Assran, M., Ballas, N., Galuba, W., Howes, R., Huang, P.Y., Li, S.W., Misra, I., Rabbat, M., Sharma, V., Synnaeve, G., Xu, H., Jegou, H., Mairal, J., Labatut, P., Joulin, A., Bojanowski, P.: Dinov2: Learning robust visual features without supervision (2023) +46. Qiao, T., Zhang, J., Xu, D., Tao, D.: Mirrorgan: Learning text-to-image generation by redescription. In: CVPR (2019) +47. Qu, Z., Xiang, T., Song, Y.Z.: Sketchdreamer: Interactive text-augmented creative sketch ideation. In: BMVC (2023) +48. Radford, A., Kim, J.W., Hallacy, C., Ramesh, A., Goh, G., Agarwal, S., Sastry, G., Askell, A., Mishkin, P., Clark, J., et al.: Learning transferable visual models from natural language supervision. In: ICML (2021) +49. Ramesh, A., Dhariwal, P., Nichol, A., Chu, C., Chen, M.: Hierarchical text-conditional image generation with clip latents. arXiv preprint arXiv:2204.06125 (2022) +50. Ren, T., Liu, S., Zeng, A., Lin, J., Li, K., Cao, H., Chen, J., Huang, X., Chen, Y., Yan, F., Zeng, Z., Zhang, H., Li, F., Yang, J., Li, H., Jiang, Q., Zhang, L.: Grounded sam: Assembling open-world models for diverse visual tasks. arXiv preprint arXiv:2401.14159 (2024) +51. Richardson, E., Goldberg, K., Alaluf, Y., Cohen-Or, D.: Conceptlab: Creative generation using diffusion prior constraints. arXiv preprint arXiv:2308.02669 (2023) +52. Rombach, R., Blattmann, A., Lorenz, D., Esser, P., Ommer, B.: High-resolution image synthesis with latent diffusion models. In: CVPR (2022) +53. Ruiz, N., Li, Y., Jampani, V., Pritch, Y., Rubinstein, M., Aberman, K.: Dreambooth: Fine tuning text-to-image diffusion models for subject-driven generation. In: CVPR (2023) +54. Runco, M.A., Jaeger, G.J.: The standard definition of creativity. Creativity research journal (2012) +55. Saharia, C., Chan, W., Saxena, S., Li, L., Whang, J., Denton, E.L., Ghasemipour, K., Gontijo Lopes, R., Karagol Ayan, B., Salimans, T., et al.: Photorealistic text-to-image diffusion models with deep language understanding. In: NeurIPS. vol. 35 (2022) +56. Sbai, O., Elhoseiny, M., Bordes, A., LeCun, Y., Couprie, C.: Design: Design inspiration from generative networks. In: ECCVW (2019) +57. Sun, P., Chen, S., Luo, P.: Grounded segment anything: From objects to parts. https://github.com/Cheems-Seminar/grounded-segment-any-parts (2023) +58. Sun, P., Chen, S., Zhu, C., Xiao, F., Luo, P., Xie, S., Yan, Z.: Going denser with open-vocabulary part segmentation. arXiv preprint arXiv:2305.11173 (2023) + +59. Tao, M., Tang, H., Wu, F., Jing, X.Y., Bao, B.K., Xu, C.: Df-gan: A simple and effective baseline for text-to-image synthesis. In: CVPR (2022) +60. Vinker, Y., Voynov, A., Cohen-Or, D., Shamir, A.: Concept decomposition for visual exploration and inspiration. In: SIGGRAPH Asia (2023) +61. Voynov, A., Chu, Q., Cohen-Or, D., Aberman, K.: P+: Extended textual conditioning in text-to-image generation. arXiv preprint arXiv:2303.09522 (2023) +62. Wah, C., Branson, S., Welinder, P., Perona, P., Belongie, S.: The Caltech-UCSD Birds-200-2011 Dataset (2011) +63. Wang, X., Darrell, T., Rambhatla, S.S., Girdhar, R., Misra, I.: Instancediffusion: Instance-level control for image generation. arXiv preprint arXiv:2402.03290 (2024) +64. Wei, Y., Zhang, Y., Ji, Z., Bai, J., Zhang, L., Zuo, W.: Elite: Encoding visual concepts into textual embeddings for customized text-to-image generation. In: ICCV (2023) +65. Wilkenfeld, M.J., Ward, T.B.: Similarity and emergence in conceptual combination. Journal of Memory and Language (2001) +66. Xie, J., Li, Y., Huang, Y., Liu, H., Zhang, W., Zheng, Y., Shou, M.Z.: Boxdiff: Text-to-image synthesis with training-free box-constrained diffusion. In: ICCV (2023) +67. Xu, T., Zhang, P., Huang, Q., Zhang, H., Gan, Z., Huang, X., He, X.: Attngan: Fine-grained text to image generation with attentional generative adversarial networks. In: CVPR (2018) +68. Yang, B., Gu, S., Zhang, B., Zhang, T., Chen, X., Sun, X., Chen, D., Wen, F.: Paint by example: Exemplar-based image editing with diffusion models. In: CVPR (2023) +69. Yang, Z., Wang, J., Gan, Z., Li, L., Lin, K., Wu, C., Duan, N., Liu, Z., Liu, C., Zeng, M., et al.: Reco: Region-controlled text-to-image generation. In: CVPR (2023) +70. Yin, G., Liu, B., Sheng, L., Yu, N., Wang, X., Shao, J.: Semantics disentangling for text-to-image generation. In: CVPR (2019) +71. Zhang, L., Rao, A., Agrawala, M.: Adding conditional control to text-to-image diffusion models. In: ICCV (2023) +72. Zhu, M., Pan, P., Chen, W., Yang, Y.: Dm-gan: Dynamic memory generative adversarial networks for text-to-image synthesis. In: CVPR (2019) \ No newline at end of file diff --git a/2024/PartCraft_ Crafting Creative Objects by Parts/images.zip b/2024/PartCraft_ Crafting Creative Objects by Parts/images.zip new file mode 100644 index 0000000000000000000000000000000000000000..6c8125bc535637e3e1e30fecfd44acc6934fa7cf --- /dev/null +++ b/2024/PartCraft_ Crafting Creative Objects by Parts/images.zip @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:aa8ac48a6c3faddf7e5597192ffd38e3b7b791d374a8e7becb111c8f794b8005 +size 900791 diff --git a/2024/PartCraft_ Crafting Creative Objects by Parts/layout.json b/2024/PartCraft_ Crafting Creative Objects by Parts/layout.json new file mode 100644 index 0000000000000000000000000000000000000000..23e386ac6638273b7b3f7d6d87927f21f2475da0 --- /dev/null +++ b/2024/PartCraft_ Crafting Creative Objects by Parts/layout.json @@ -0,0 +1,9707 @@ +{ + "pdf_info": [ + { + "para_blocks": [ + { + "bbox": [ + 145, + 111, + 470, + 129 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 145, + 111, + 470, + 129 + ], + "spans": [ + { + "bbox": [ + 145, + 111, + 470, + 129 + ], + "type": "text", + "content": "PartCraft: Crafting Creative Objects by Parts" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 157, + 149, + 455, + 163 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 157, + 149, + 455, + 163 + ], + "spans": [ + { + "bbox": [ + 157, + 149, + 455, + 163 + ], + "type": "text", + "content": "Kam Woh " + }, + { + "bbox": [ + 157, + 149, + 455, + 163 + ], + "type": "inline_equation", + "content": "\\mathrm{Ng}^{1}" + }, + { + "bbox": [ + 157, + 149, + 455, + 163 + ], + "type": "text", + "content": ", Xiatian Zhu" + }, + { + "bbox": [ + 157, + 149, + 455, + 163 + ], + "type": "inline_equation", + "content": "^{1,2}" + }, + { + "bbox": [ + 157, + 149, + 455, + 163 + ], + "type": "text", + "content": ", Yi-Zhe Song" + }, + { + "bbox": [ + 157, + 149, + 455, + 163 + ], + "type": "inline_equation", + "content": "^{1}" + }, + { + "bbox": [ + 157, + 149, + 455, + 163 + ], + "type": "text", + "content": ", and Tao Xiang" + }, + { + "bbox": [ + 157, + 149, + 455, + 163 + ], + "type": "inline_equation", + "content": "^{1}" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 205, + 171, + 426, + 183 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 205, + 171, + 426, + 183 + ], + "spans": [ + { + "bbox": [ + 205, + 171, + 426, + 183 + ], + "type": "inline_equation", + "content": "^{1}" + }, + { + "bbox": [ + 205, + 171, + 426, + 183 + ], + "type": "text", + "content": " CVSSP, University of Surrey, United Kingdom" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 222, + 183, + 391, + 194 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 222, + 183, + 391, + 194 + ], + "spans": [ + { + "bbox": [ + 222, + 183, + 391, + 194 + ], + "type": "text", + "content": "Surrey Institute for People-Centred AI" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 186, + 194, + 428, + 205 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 186, + 194, + 428, + 205 + ], + "spans": [ + { + "bbox": [ + 186, + 194, + 428, + 205 + ], + "type": "text", + "content": "{kamwoh.ng, xiatian.zhu, y song, t.xiang}@surrey.ac.uk" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 160, + 234, + 455, + 443 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 160, + 234, + 455, + 443 + ], + "spans": [ + { + "bbox": [ + 160, + 234, + 455, + 443 + ], + "type": "text", + "content": "Abstract. This paper propels creative control in generative visual AI by allowing users to \"select\". Departing from traditional text or sketch-based methods, we for the first time allow users to choose visual concepts by parts for their creative endeavors. The outcome is fine-grained generation that precisely captures selected visual concepts, ensuring a holistically faithful and plausible result. To achieve this, we first parse objects into parts through unsupervised feature clustering. Then, we encode parts into text tokens and introduce an entropy-based normalized attention loss that operates on them. This loss design enables our model to learn generic prior topology knowledge about object's part composition, and further generalize to novel part compositions to ensure the generation looks holistically faithful. Lastly, we employ a bottleneck encoder to project the part tokens. This not only enhances fidelity but also accelerates learning, by leveraging shared knowledge and facilitating information exchange among instances. Visual results in the paper and supplementary material showcase the compelling power of PartCraft in crafting highly customized, innovative creations, exemplified by the \"charming\" and creative birds in Fig. 1. Code is released at https://github.com/kamwoh/partcraft." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 160, + 453, + 453, + 464 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 160, + 453, + 453, + 464 + ], + "spans": [ + { + "bbox": [ + 160, + 453, + 453, + 464 + ], + "type": "text", + "content": "Keywords: Part Composition " + }, + { + "bbox": [ + 160, + 453, + 453, + 464 + ], + "type": "inline_equation", + "content": "\\cdot" + }, + { + "bbox": [ + 160, + 453, + 453, + 464 + ], + "type": "text", + "content": " Controllable Text-to-image Generation" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 132, + 485, + 230, + 498 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 485, + 230, + 498 + ], + "spans": [ + { + "bbox": [ + 132, + 485, + 230, + 498 + ], + "type": "text", + "content": "1 Introduction" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 130, + 510, + 482, + 581 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 510, + 482, + 581 + ], + "spans": [ + { + "bbox": [ + 130, + 510, + 482, + 581 + ], + "type": "text", + "content": "Humans are creators; AI, on the other hand, hallucinates. Creativity, arguably, is the very force driving humanity forward. Recently, generative AI has garnered considerable attention for its perceived \"creativity\" [6, 19, 20, 43, 47, 49, 52, 55]. Despite its promise, the challenge of control has swiftly surfaced - how can humans infuse their creativity into the generation process and regulate the extent to which AI hallucinates?" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 130, + 582, + 482, + 666 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 582, + 482, + 666 + ], + "spans": [ + { + "bbox": [ + 130, + 582, + 482, + 666 + ], + "type": "text", + "content": "Creativity starts with an idea. The immediate challenge is how to articulate that idea and integrate it with generative AI. Text is the most commonly employed medium. For instance, imagine being a bird enthusiast wanting to craft the most unique bird akin to Fig. 1. The go-to approach would be furnishing Stable Diffusion [52] (or an equivalent model) with the following textual prompt (the idea): \"generate a bird with the head of " + }, + { + "bbox": [ + 130, + 582, + 482, + 666 + ], + "type": "inline_equation", + "content": "X" + }, + { + "bbox": [ + 130, + 582, + 482, + 666 + ], + "type": "text", + "content": ", wings of " + }, + { + "bbox": [ + 130, + 582, + 482, + 666 + ], + "type": "inline_equation", + "content": "Y" + }, + { + "bbox": [ + 130, + 582, + 482, + 666 + ], + "type": "text", + "content": ", and body of " + }, + { + "bbox": [ + 130, + 582, + 482, + 666 + ], + "type": "inline_equation", + "content": "Z" + }, + { + "bbox": [ + 130, + 582, + 482, + 666 + ], + "type": "text", + "content": "\". While you might be presented with remarkably looking birds like Fig. 1(a), they" + } + ] + } + ], + "index": 9 + } + ], + "discarded_blocks": [], + "page_size": [ + 612, + 792 + ], + "page_idx": 0 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 134, + 114, + 480, + 255 + ], + "blocks": [ + { + "bbox": [ + 134, + 114, + 480, + 255 + ], + "lines": [ + { + "bbox": [ + 134, + 114, + 480, + 255 + ], + "spans": [ + { + "bbox": [ + 134, + 114, + 480, + 255 + ], + "type": "image", + "image_path": "e5c3dba8567cb6b6bdad61f2d12b18caababc60418696b678e4cb1bb0f2171d4.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 130, + 262, + 482, + 395 + ], + "lines": [ + { + "bbox": [ + 130, + 262, + 482, + 395 + ], + "spans": [ + { + "bbox": [ + 130, + 262, + 482, + 395 + ], + "type": "text", + "content": "Fig. 1: People often form creative concepts based on existing ones [9, 42, 54, 65]. For instance, a bird enthusiast may want to craft a unique bird with different parts (e.g., heads, bodies, and wings) from common bird types (e.g., blue jay, cardinal and song sparrow). (a) Using text prompts in T2I models often results in a lack of control and deviation from the intended details, especially those visual details that are difficult to describe. (b) While sketching is a direct way, not everyone possesses the ability to sketch, particularly in intricate detail. (c) Even with reference images, existing methods (e.g., DreamBooth [53]) did not consider learning object parts, thus cannot generate with desired parts. (d) Using additional control is even cumbersome, requiring too many inputs! (e) We aim to create an object by simply selecting desired parts. PartCraft learns from visual examples to generate the object with a faithful holistic structure, seamlessly integrating the chosen parts into a natural and coherent entity." + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_caption" + } + ], + "index": 2 + }, + { + "bbox": [ + 130, + 410, + 482, + 457 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 410, + 482, + 457 + ], + "spans": [ + { + "bbox": [ + 130, + 410, + 482, + 457 + ], + "type": "text", + "content": "may bear little resemblance to the envisioned concept. Recent literature suggests that a swift sketch could serve as a viable alternative [40, 63, 71], providing fine-grained shape control. However, the caveat is that not everyone possesses the ability to sketch, particularly in intricate detail." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 130, + 460, + 482, + 543 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 460, + 482, + 543 + ], + "spans": [ + { + "bbox": [ + 130, + 460, + 482, + 543 + ], + "type": "text", + "content": "In this paper, our primary focus revolves around addressing the issue of \"control\" in generative AI. We advance by introducing fine-grained control into the generative process, inviting you to \"select\". While this selection mechanism might seem modest at first, it closely mirrors the human creative process, where new concepts often emerge from existing ones [9,42,54,65]. Recall those moments when you desired an \"ideal\" pair of shoes with selected features from different pairs, or when you aimed to get creative with a cat (for that matter!)." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 130, + 544, + 482, + 639 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 544, + 482, + 639 + ], + "spans": [ + { + "bbox": [ + 130, + 544, + 482, + 639 + ], + "type": "text", + "content": "It follows that rather than relying on writing (text) or drawing (sketch), all that is required is to choose the distinct visual concepts you specifically desire in your creative endeavor. Our model then ensures that all selected concepts are seamlessly and precisely composed into a faithful novel object in the final generation, without resorting to additional control such as bounding boxes [7,38, 40,66,71]. To illustrate with the \"unique bird\" example once more, our approach literally involves selecting the head of X, wings of Y, and body of Z! (see Fig. 1(e) and 2)." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 130, + 641, + 482, + 666 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 641, + 482, + 666 + ], + "spans": [ + { + "bbox": [ + 130, + 641, + 482, + 666 + ], + "type": "text", + "content": "Our solution is intuitive and centers around the well-studied computer vision concept of objects and their parts [8,14,22,27,35]. The challenge then boils down" + } + ] + } + ], + "index": 7 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 133, + 91, + 139, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 91, + 139, + 100 + ], + "spans": [ + { + "bbox": [ + 133, + 91, + 139, + 100 + ], + "type": "text", + "content": "2" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 166, + 91, + 225, + 102 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 166, + 91, + 225, + 102 + ], + "spans": [ + { + "bbox": [ + 166, + 91, + 225, + 102 + ], + "type": "text", + "content": "KW. Ng et al." + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 1 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 133, + 113, + 482, + 266 + ], + "blocks": [ + { + "bbox": [ + 133, + 113, + 482, + 266 + ], + "lines": [ + { + "bbox": [ + 133, + 113, + 482, + 266 + ], + "spans": [ + { + "bbox": [ + 133, + 113, + 482, + 266 + ], + "type": "image", + "image_path": "4cf0ceb7b2a760d2da60d869c667e63255c96c0b5a3360d3a537ac5cc3e6e273.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 130, + 274, + 482, + 308 + ], + "lines": [ + { + "bbox": [ + 130, + 274, + 482, + 308 + ], + "spans": [ + { + "bbox": [ + 130, + 274, + 482, + 308 + ], + "type": "text", + "content": "Fig. 2: Two sets of images were generated from their original parts (sources A and B). We can integrate a specific part (e.g., body, head, or even background (BG) of source B to target A seamlessly without effort." + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_caption" + } + ], + "index": 2 + }, + { + "bbox": [ + 130, + 329, + 479, + 376 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 329, + 479, + 376 + ], + "spans": [ + { + "bbox": [ + 130, + 329, + 479, + 376 + ], + "type": "text", + "content": "to two aspects: (i) how to parse known objects into their parts (i.e., recognizing that birds have eyes and tails), and (ii) how to assemble parts from different visual concepts to form a faithful creative concept (i.e., ensuring the output is recognizably a bird)." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 130, + 380, + 480, + 453 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 380, + 480, + 453 + ], + "spans": [ + { + "bbox": [ + 130, + 380, + 480, + 453 + ], + "type": "text", + "content": "The former is easier. We make clever use of DINOv2 [45] feature maps and perform unsupervised clustering to identify common parts. The idea is that each cluster will then correspond to a semantic part of a common object (e.g., head and wings of birds). We specifically chose DINOv2 for its superior fine-grained perception [2] compared to others [50, 57, 58]. Further, this way has a higher flexibility to enhance fine-grained parsing by using a higher cluster count." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 130, + 457, + 480, + 541 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 457, + 480, + 541 + ], + "spans": [ + { + "bbox": [ + 130, + 457, + 480, + 541 + ], + "type": "text", + "content": "Our major contribution lies in addressing the latter challenge. The solution is intuitive - it essentially revolves around the fine-grained selection and placement of the chosen parts. Inspired by recent efforts in personalization [3,25,36,53,64], primarily designed for learning entire objects, we introduce a tailored attention loss that specifically operates on parts. With this loss, our model learns the object part composition, ensuring the final generation appears holistically faithful with mixed parts (i.e., head and wings of a bird actually appear at the right places)." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 130, + 546, + 481, + 666 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 546, + 481, + 666 + ], + "spans": [ + { + "bbox": [ + 130, + 546, + 481, + 666 + ], + "type": "text", + "content": "More specifically, we introduce an entropy-based attention loss that maximizes the attention of a specific part at a particular location while minimizing the attention where no parts appear. This is achieved by first selecting the attention maps that correspond to the parts. A normalization over all parts is then performed to ensure that each image region is occupied by no more than one part. Finally, we minimize the entropy loss between each normalized individual part and the semantic maps obtained during object parsing, containing the correct part location. This not only facilitates stronger part disentanglement but is also the key to generating a faithful holistic structure of an object as it learns a generic prior topology knowledge about object parts." + } + ] + } + ], + "index": 7 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 258, + 91, + 448, + 102 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 258, + 91, + 448, + 102 + ], + "spans": [ + { + "bbox": [ + 258, + 91, + 448, + 102 + ], + "type": "text", + "content": "PartCraft: Crafting Creative Objects by Parts" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 474, + 91, + 481, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 474, + 91, + 481, + 100 + ], + "spans": [ + { + "bbox": [ + 474, + 91, + 481, + 100 + ], + "type": "text", + "content": "3" + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 2 + }, + { + "para_blocks": [ + { + "bbox": [ + 130, + 116, + 482, + 176 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 116, + 482, + 176 + ], + "spans": [ + { + "bbox": [ + 130, + 116, + 482, + 176 + ], + "type": "text", + "content": "To enhance generation fidelity, we further employ a bottleneck encoder to project the text tokens. This approach accelerates learning by leveraging shared knowledge (common parts) and facilitating information exchange among instances in each part. Each instance adjusts slightly to adapt to the fine-grained part details during optimization." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 130, + 176, + 483, + 333 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 176, + 483, + 333 + ], + "spans": [ + { + "bbox": [ + 130, + 176, + 483, + 333 + ], + "type": "text", + "content": "Our contributions are as follows: (i) We pioneer a unique approach for fine-grained part-level control in Text-to-Image (T2I) models, empowering users to craft objects by selecting desired parts. This method streamlines the creative process, marking a significant advancement in our capacity to manipulate and reimagine visual content. (ii) We introduce PartCraft, a technique that autonomously parses object parts and orchestrates them from different visual concepts, resulting in the faithful creation of a novel object. (iii) For enhanced part disentanglement and generation fidelity, we propose an entropy-based normalized attention loss and leverage a bottleneck encoder. (iv) We present two problem-specific quantitative metrics. Comprehensive experiments on CUB-200-2011 (birds) and the Stanford Dogs dataset demonstrate the superior performance of our method in generating novel objects, surpassing alternative approaches in both qualitative and quantitative evaluations." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 132, + 342, + 237, + 355 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 342, + 237, + 355 + ], + "spans": [ + { + "bbox": [ + 132, + 342, + 237, + 355 + ], + "type": "text", + "content": "2 Related Work" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 130, + 366, + 482, + 521 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 366, + 482, + 521 + ], + "spans": [ + { + "bbox": [ + 130, + 366, + 482, + 521 + ], + "type": "text", + "content": "Creative editing and generation. Creativity involves generating innovative ideas or artifacts across various domains [12]. Extensive research has explored the integration of creativity into Generative Adversarial Networks (GANs) [21,44,56] and Variational Autoencoders (VAEs) [16, 18]. For example, DoodlerGAN [26] learns and combines fine-level part components to create sketches of new species. A recent study by [60] demonstrated decomposing personalized concepts into distinct visual aspects, creatively recombined through diffusion models. Instruct-Pix2Pix [10] allows creative image editing through instructions, while ConceptLab [51] aims to identify novel concepts within a specified category, deviating from existing concepts. Different from these works where editing/generation usually focuses on the whole object, we instead focus on training a text-to-image generative model that can understand parts, thus able to creatively generate new objects by seamlessly composing different parts simply through selection." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 130, + 522, + 482, + 666 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 522, + 482, + 666 + ], + "spans": [ + { + "bbox": [ + 130, + 522, + 482, + 666 + ], + "type": "text", + "content": "Text-to-image generation. Recent advancements in large text-to-image (T2I) diffusion models [6, 19, 20, 43, 49, 52, 55] have made significant improvements over conventional methods [37, 46, 59, 67, 70, 72] in producing high-fidelity images from text prompts. Their application scope has expanded to include both global [10, 32, 41] and localized [4, 17, 28, 68] image editing tasks, demonstrating versatility. Methods such as [5, 7, 13, 15, 23, 24, 34, 38, 66, 69, 71] have introduced more granular spatial control, such as incorporating semantic segmentation masks or bounding boxes, into large pretrained diffusion models to guide image generation. Contrary to these approaches, we focus on enhancing the control by enabling a straightforward discrete selection of desired parts. We minimize the complexity and manual intervention required by spatial controls, yet the model can compose selected parts as a coherent object seamlessly. This not" + } + ] + } + ], + "index": 6 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "spans": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "type": "text", + "content": "4" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 166, + 91, + 225, + 102 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 166, + 91, + 225, + 102 + ], + "spans": [ + { + "bbox": [ + 166, + 91, + 225, + 102 + ], + "type": "text", + "content": "KW. Ng et al." + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 3 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 135, + 125, + 483, + 237 + ], + "blocks": [ + { + "bbox": [ + 135, + 125, + 483, + 237 + ], + "lines": [ + { + "bbox": [ + 135, + 125, + 483, + 237 + ], + "spans": [ + { + "bbox": [ + 135, + 125, + 483, + 237 + ], + "type": "image", + "image_path": "07aab029c3b582c0c91383aecfd2e6532c96a881fbc5b46150f13f502b722ee8.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 130, + 247, + 482, + 326 + ], + "lines": [ + { + "bbox": [ + 130, + 247, + 482, + 326 + ], + "spans": [ + { + "bbox": [ + 130, + 247, + 482, + 326 + ], + "type": "text", + "content": "Fig. 3: Overview of our PartCraft. All parts are organized into a dictionary, and their semantic embeddings are learned through a textual inversion approach. For instance, a text description like \"a photo of a [Head,42] [Wing,87]...\" guides the optimization of the corresponding textual embedding by reconstructing the associated image. To improve generation fidelity, we incorporate a bottleneck encoder " + }, + { + "bbox": [ + 130, + 247, + 482, + 326 + ], + "type": "inline_equation", + "content": "f" + }, + { + "bbox": [ + 130, + 247, + 482, + 326 + ], + "type": "text", + "content": " (MLP) to compute the embedding " + }, + { + "bbox": [ + 130, + 247, + 482, + 326 + ], + "type": "inline_equation", + "content": "y" + }, + { + "bbox": [ + 130, + 247, + 482, + 326 + ], + "type": "text", + "content": " (Eq. (4)) as input to the text encoder. To promote disentanglement among learned parts, we minimize a specially designed attention loss, denoted as " + }, + { + "bbox": [ + 130, + 247, + 482, + 326 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_{\\text{att}}" + }, + { + "bbox": [ + 130, + 247, + 482, + 326 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_caption" + } + ], + "index": 2 + }, + { + "bbox": [ + 130, + 338, + 482, + 396 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 338, + 482, + 396 + ], + "spans": [ + { + "bbox": [ + 130, + 338, + 482, + 396 + ], + "type": "text", + "content": "only simplifies the user's role in the generative process but also ensures that the compositional logic and coherence are inherently managed by the model's ability. As such, our research is distinguished by focusing on the model's inherent ability to understand and apply part relationships instead of providing additional controls." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 130, + 399, + 482, + 590 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 399, + 482, + 590 + ], + "spans": [ + { + "bbox": [ + 130, + 399, + 482, + 590 + ], + "type": "text", + "content": "Abstracting visual knowledge as a text token. The effectiveness of T2I models is constrained by the user's ability to articulate their desired image through text. These models face challenges in faithfully replicating visual characteristics from a reference set and generating innovative interpretations in diverse contexts, even with detailed textual descriptions. To address this challenge, various personalization techniques have been developed. These techniques obtain a new word embedding from multiple images depicting the same concept [1, 25, 36, 53, 61] or multiple new word embeddings for various concepts within a single image [3] through inversion. The learned visual concepts can be creatively reused in many image editing tasks. Nonetheless, most of these approaches struggle to learn object parts, often not able to follow the part selections due to part entanglement as they were designed to learn object as a whole. In this work, we introduce a customized attention loss that serves a dual purpose: ensure accurate positioning of each part and enforce each image region occupied by no more than one part. This greatly improves the part disentanglement, creating novel concepts with correct appearances (see Fig. 2)." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 132, + 612, + 233, + 627 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 612, + 233, + 627 + ], + "spans": [ + { + "bbox": [ + 132, + 612, + 233, + 627 + ], + "type": "text", + "content": "3 Methodology" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 130, + 641, + 481, + 666 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 641, + 481, + 666 + ], + "spans": [ + { + "bbox": [ + 130, + 641, + 481, + 666 + ], + "type": "text", + "content": "Given a set of unlabeled images depicting the same object (e.g., bird) with different part details, we aim to train a T2I generative model that decomposes" + } + ] + } + ], + "index": 7 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 258, + 91, + 448, + 103 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 258, + 91, + 448, + 103 + ], + "spans": [ + { + "bbox": [ + 258, + 91, + 448, + 103 + ], + "type": "text", + "content": "PartCraft: Crafting Creative Objects by Parts" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 474, + 91, + 480, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 474, + 91, + 480, + 100 + ], + "spans": [ + { + "bbox": [ + 474, + 91, + 480, + 100 + ], + "type": "text", + "content": "5" + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 4 + }, + { + "para_blocks": [ + { + "bbox": [ + 130, + 116, + 479, + 140 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 116, + 479, + 140 + ], + "spans": [ + { + "bbox": [ + 130, + 116, + 479, + 140 + ], + "type": "text", + "content": "parts of objects into text tokens and can recompose them in a novel way. To that end, we propose PartCraft, as depicted in Fig. 3." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 130, + 140, + 480, + 224 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 140, + 480, + 224 + ], + "spans": [ + { + "bbox": [ + 130, + 140, + 480, + 224 + ], + "type": "text", + "content": "We start by discovering the parts in a three-tier hierarchy, as detailed in Sec. 3.1. Paired with the training images " + }, + { + "bbox": [ + 130, + 140, + 480, + 224 + ], + "type": "inline_equation", + "content": "\\{x_{i}\\}_{i = 1}^{N}" + }, + { + "bbox": [ + 130, + 140, + 480, + 224 + ], + "type": "text", + "content": ", this semantic hierarchy subsequently serves as the supervision to fine-tune a pre-trained text-to-image model, say a latent diffusion model [52], denoted as " + }, + { + "bbox": [ + 130, + 140, + 480, + 224 + ], + "type": "inline_equation", + "content": "\\{\\epsilon_{\\theta},\\tau_{\\theta},\\mathcal{E},\\mathcal{D}\\}" + }, + { + "bbox": [ + 130, + 140, + 480, + 224 + ], + "type": "text", + "content": ", where " + }, + { + "bbox": [ + 130, + 140, + 480, + 224 + ], + "type": "inline_equation", + "content": "\\epsilon_{\\theta}" + }, + { + "bbox": [ + 130, + 140, + 480, + 224 + ], + "type": "text", + "content": " represents the diffusion denoiser, " + }, + { + "bbox": [ + 130, + 140, + 480, + 224 + ], + "type": "inline_equation", + "content": "\\tau_{\\theta}" + }, + { + "bbox": [ + 130, + 140, + 480, + 224 + ], + "type": "text", + "content": " the text encoder, and " + }, + { + "bbox": [ + 130, + 140, + 480, + 224 + ], + "type": "inline_equation", + "content": "\\mathcal{E} / \\mathcal{D}" + }, + { + "bbox": [ + 130, + 140, + 480, + 224 + ], + "type": "text", + "content": " the autoencoder respectively. We adopt the textual inversion technique [25]. Concretely, we learn a set of pseudo-words " + }, + { + "bbox": [ + 130, + 140, + 480, + 224 + ], + "type": "inline_equation", + "content": "p^*" + }, + { + "bbox": [ + 130, + 140, + 480, + 224 + ], + "type": "text", + "content": " for each part in the word embedding space with:" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 220, + 232, + 481, + 247 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 220, + 232, + 481, + 247 + ], + "spans": [ + { + "bbox": [ + 220, + 232, + 481, + 247 + ], + "type": "interline_equation", + "content": "\\mathcal {L} _ {l d m} = \\mathbb {E} _ {z, t, p, \\epsilon} \\left[ | | \\epsilon - \\epsilon_ {\\theta} \\left(z _ {t}, t, \\tau_ {\\theta} (y _ {p})\\right) | | _ {2} ^ {2} \\right], \\tag {1}", + "image_path": "d7180ac0e2525a51a074ffc9a6f0f1010a40d65f017d02e1a4e9badf314be412.jpg" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 231, + 248, + 481, + 268 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 231, + 248, + 481, + 268 + ], + "spans": [ + { + "bbox": [ + 231, + 248, + 481, + 268 + ], + "type": "interline_equation", + "content": "p ^ {*} = \\underset {p} {\\operatorname {a r g m i n}} \\mathcal {L} _ {l d m}, \\tag {2}", + "image_path": "b09065e62513a74f9c244ddab9dcfb591a26d711b57b09e8d466e32a48180601.jpg" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 130, + 277, + 480, + 361 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 277, + 480, + 361 + ], + "spans": [ + { + "bbox": [ + 130, + 277, + 480, + 361 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 130, + 277, + 480, + 361 + ], + "type": "inline_equation", + "content": "\\epsilon \\sim \\mathcal{N}(0,1)" + }, + { + "bbox": [ + 130, + 277, + 480, + 361 + ], + "type": "text", + "content": " denotes the unscaled noise, " + }, + { + "bbox": [ + 130, + 277, + 480, + 361 + ], + "type": "inline_equation", + "content": "t" + }, + { + "bbox": [ + 130, + 277, + 480, + 361 + ], + "type": "text", + "content": " is the time step, " + }, + { + "bbox": [ + 130, + 277, + 480, + 361 + ], + "type": "inline_equation", + "content": "z = \\mathcal{E}(x)" + }, + { + "bbox": [ + 130, + 277, + 480, + 361 + ], + "type": "text", + "content": " is the latent representation of the image, " + }, + { + "bbox": [ + 130, + 277, + 480, + 361 + ], + "type": "inline_equation", + "content": "z_{t}" + }, + { + "bbox": [ + 130, + 277, + 480, + 361 + ], + "type": "text", + "content": " is the latent noise at time " + }, + { + "bbox": [ + 130, + 277, + 480, + 361 + ], + "type": "inline_equation", + "content": "t" + }, + { + "bbox": [ + 130, + 277, + 480, + 361 + ], + "type": "text", + "content": ", and " + }, + { + "bbox": [ + 130, + 277, + 480, + 361 + ], + "type": "inline_equation", + "content": "y_{p}" + }, + { + "bbox": [ + 130, + 277, + 480, + 361 + ], + "type": "text", + "content": " is the text condition that includes " + }, + { + "bbox": [ + 130, + 277, + 480, + 361 + ], + "type": "inline_equation", + "content": "p" + }, + { + "bbox": [ + 130, + 277, + 480, + 361 + ], + "type": "text", + "content": " as part of the text tokens. " + }, + { + "bbox": [ + 130, + 277, + 480, + 361 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_{ldm}" + }, + { + "bbox": [ + 130, + 277, + 480, + 361 + ], + "type": "text", + "content": " is a standard diffusion loss [30] to reconstruct the parts. As each object is composed of a set of parts, its reconstruction is achieved by the reconstruction of the associated set of parts. In other words, when all parts are reconstructed properly, it will become a valid object." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 130, + 380, + 308, + 392 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 380, + 308, + 392 + ], + "spans": [ + { + "bbox": [ + 130, + 380, + 308, + 392 + ], + "type": "text", + "content": "3.1 Unsupervised Part Discovery" + } + ] + } + ], + "index": 7 + }, + { + "type": "image", + "bbox": [ + 132, + 408, + 291, + 497 + ], + "blocks": [ + { + "bbox": [ + 132, + 408, + 291, + 497 + ], + "lines": [ + { + "bbox": [ + 132, + 408, + 291, + 497 + ], + "spans": [ + { + "bbox": [ + 132, + 408, + 291, + 497 + ], + "type": "image", + "image_path": "7cdf1b8f24fb6024e9ccc6323e592f8e95ce299405de4bc25888338d91dd2b32.jpg" + } + ] + } + ], + "index": 8, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 130, + 505, + 291, + 559 + ], + "lines": [ + { + "bbox": [ + 130, + 505, + 291, + 559 + ], + "spans": [ + { + "bbox": [ + 130, + 505, + 291, + 559 + ], + "type": "text", + "content": "Fig. 4: Part discovery within a semantic hierarchy involves partitioning each image into distinct parts and forming semantic clusters across unlabeled training data." + } + ] + } + ], + "index": 9, + "angle": 0, + "type": "image_caption" + } + ], + "index": 8 + }, + { + "bbox": [ + 296, + 400, + 482, + 567 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 296, + 400, + 482, + 567 + ], + "spans": [ + { + "bbox": [ + 296, + 400, + 482, + 567 + ], + "type": "text", + "content": "To minimize the labeling cost, we develop a scalable process to reveal the underlying semantic hierarchy with parts in an unsupervised fashion. We leverage the off-the-shelf vision model for image decomposition and clustering. Specifically, given an image " + }, + { + "bbox": [ + 296, + 400, + 482, + 567 + ], + "type": "inline_equation", + "content": "x_{i}" + }, + { + "bbox": [ + 296, + 400, + 482, + 567 + ], + "type": "text", + "content": ", we employ DINOv2 [45] to extract the feature map " + }, + { + "bbox": [ + 296, + 400, + 482, + 567 + ], + "type": "inline_equation", + "content": "F = \\{F_{i} = \\mathrm{DINOv2}(x_{i})\\}_{i}^{N}" + }, + { + "bbox": [ + 296, + 400, + 482, + 567 + ], + "type": "text", + "content": ". We then conduct three-level hierarchical clustering (see Fig. 4): (i) At the top level, " + }, + { + "bbox": [ + 296, + 400, + 482, + 567 + ], + "type": "inline_equation", + "content": "k" + }, + { + "bbox": [ + 296, + 400, + 482, + 567 + ], + "type": "text", + "content": "-means is applied on all patches in " + }, + { + "bbox": [ + 296, + 400, + 482, + 567 + ], + "type": "inline_equation", + "content": "F" + }, + { + "bbox": [ + 296, + 400, + 482, + 567 + ], + "type": "text", + "content": " with " + }, + { + "bbox": [ + 296, + 400, + 482, + 567 + ], + "type": "inline_equation", + "content": "k = 2" + }, + { + "bbox": [ + 296, + 400, + 482, + 567 + ], + "type": "text", + "content": " to separate foregrounds and backgrounds " + }, + { + "bbox": [ + 296, + 400, + 482, + 567 + ], + "type": "inline_equation", + "content": "B" + }, + { + "bbox": [ + 296, + 400, + 482, + 567 + ], + "type": "text", + "content": ". (ii) At the middle level, " + }, + { + "bbox": [ + 296, + 400, + 482, + 567 + ], + "type": "inline_equation", + "content": "k" + }, + { + "bbox": [ + 296, + 400, + 482, + 567 + ], + "type": "text", + "content": "-means is further ap" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 130, + 568, + 482, + 640 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 568, + 482, + 640 + ], + "spans": [ + { + "bbox": [ + 130, + 568, + 482, + 640 + ], + "type": "text", + "content": "plied on all foreground patches to acquire " + }, + { + "bbox": [ + 130, + 568, + 482, + 640 + ], + "type": "inline_equation", + "content": "M" + }, + { + "bbox": [ + 130, + 568, + 482, + 640 + ], + "type": "text", + "content": " clusters representing common parts, such as the heads of birds. (iii) At the bottom level, we further group each of the " + }, + { + "bbox": [ + 130, + 568, + 482, + 640 + ], + "type": "inline_equation", + "content": "M" + }, + { + "bbox": [ + 130, + 568, + 482, + 640 + ], + "type": "text", + "content": " clusters as well as the background cluster " + }, + { + "bbox": [ + 130, + 568, + 482, + 640 + ], + "type": "inline_equation", + "content": "B" + }, + { + "bbox": [ + 130, + 568, + 482, + 640 + ], + "type": "text", + "content": " into " + }, + { + "bbox": [ + 130, + 568, + 482, + 640 + ], + "type": "inline_equation", + "content": "K" + }, + { + "bbox": [ + 130, + 568, + 482, + 640 + ], + "type": "text", + "content": " splits. Each split refers to finer meanings, such as the head of a specific bird species, or a specific background style. Lastly, each region of an image will be tagged with the corresponding cluster index. We represent these cluster tags as follows:" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 237, + 649, + 481, + 662 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 237, + 649, + 481, + 662 + ], + "spans": [ + { + "bbox": [ + 237, + 649, + 481, + 662 + ], + "type": "interline_equation", + "content": "p = (0, k _ {0}), (1, k _ {1}), \\dots , (M, k _ {M}), \\tag {3}", + "image_path": "fdbc3e0d4d71b110d533fe892fd005afa0954dd52f2061237faaae8ca738a30d.jpg" + } + ] + } + ], + "index": 12 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 132, + 91, + 140, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 91, + 140, + 100 + ], + "spans": [ + { + "bbox": [ + 132, + 91, + 140, + 100 + ], + "type": "text", + "content": "6" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 166, + 91, + 225, + 102 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 166, + 91, + 225, + 102 + ], + "spans": [ + { + "bbox": [ + 166, + 91, + 225, + 102 + ], + "type": "text", + "content": "KW. Ng et al." + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 5 + }, + { + "para_blocks": [ + { + "bbox": [ + 130, + 116, + 479, + 200 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 116, + 479, + 200 + ], + "spans": [ + { + "bbox": [ + 130, + 116, + 479, + 200 + ], + "type": "text", + "content": "where the first pair refers to the background style, and the following " + }, + { + "bbox": [ + 130, + 116, + 479, + 200 + ], + "type": "inline_equation", + "content": "M" + }, + { + "bbox": [ + 130, + 116, + 479, + 200 + ], + "type": "text", + "content": " pairs denote the combinations of " + }, + { + "bbox": [ + 130, + 116, + 479, + 200 + ], + "type": "inline_equation", + "content": "M" + }, + { + "bbox": [ + 130, + 116, + 479, + 200 + ], + "type": "text", + "content": " parts (e.g., head, body, wings) each associated with a specific object (e.g., sparrow), and " + }, + { + "bbox": [ + 130, + 116, + 479, + 200 + ], + "type": "inline_equation", + "content": "k \\in \\{1, \\dots, K\\}" + }, + { + "bbox": [ + 130, + 116, + 479, + 200 + ], + "type": "text", + "content": ". This description will be used as the textual prompt in model training, such as \"a photo of a " + }, + { + "bbox": [ + 130, + 116, + 479, + 200 + ], + "type": "inline_equation", + "content": "[p]" + }, + { + "bbox": [ + 130, + 116, + 479, + 200 + ], + "type": "text", + "content": "\". Please refer to the supplementary material for more examples of the discovered semantic hierarchy. This process also yields the segmentation mask of each " + }, + { + "bbox": [ + 130, + 116, + 479, + 200 + ], + "type": "inline_equation", + "content": "m" + }, + { + "bbox": [ + 130, + 116, + 479, + 200 + ], + "type": "text", + "content": "-th part, which we define as " + }, + { + "bbox": [ + 130, + 116, + 479, + 200 + ], + "type": "inline_equation", + "content": "S_{m}" + }, + { + "bbox": [ + 130, + 116, + 479, + 200 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 130, + 200, + 480, + 271 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 200, + 480, + 271 + ], + "spans": [ + { + "bbox": [ + 130, + 200, + 480, + 271 + ], + "type": "text", + "content": "Motivation. While we can leverage off-the-shelf segmentation models such as VLPart [58], the robustness relies on the generalizability of the model and the part segmentation result is usually pre-defined and may be unstable for unseen domains. As a result, we rely on our feature clustering method to obtain the segmentation map, which also has a higher flexibility in choosing the number of clusters (parts)." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 131, + 290, + 275, + 301 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 131, + 290, + 275, + 301 + ], + "spans": [ + { + "bbox": [ + 131, + 290, + 275, + 301 + ], + "type": "text", + "content": "3.2 Part Token Bottleneck" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 130, + 311, + 479, + 382 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 311, + 479, + 382 + ], + "spans": [ + { + "bbox": [ + 130, + 311, + 479, + 382 + ], + "type": "text", + "content": "In contrast to prior text inversion studies [25], our task requires learning a greater quantity at the same time - specifically, " + }, + { + "bbox": [ + 130, + 311, + 479, + 382 + ], + "type": "inline_equation", + "content": "(M + 1)K" + }, + { + "bbox": [ + 130, + 311, + 479, + 382 + ], + "type": "text", + "content": "-of word tokens derived from a collection of discovered parts marked by inherent imperfections (such as partial overlap and over splitting). This makes the learning task more demanding. To enhance the learning process, we propose a neural network " + }, + { + "bbox": [ + 130, + 311, + 479, + 382 + ], + "type": "inline_equation", + "content": "f(\\cdot)" + }, + { + "bbox": [ + 130, + 311, + 479, + 382 + ], + "type": "text", + "content": " comprising a two-layer MLP with ReLU activation:" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 279, + 394, + 480, + 406 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 279, + 394, + 480, + 406 + ], + "spans": [ + { + "bbox": [ + 279, + 394, + 480, + 406 + ], + "type": "interline_equation", + "content": "y _ {p} = f (e (p)), \\tag {4}", + "image_path": "44ec2be5b5ab09ef78f657d6ca654a16e733a56007a8ab5a2487edbabe0e5618.jpg" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 130, + 415, + 479, + 450 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 415, + 479, + 450 + ], + "spans": [ + { + "bbox": [ + 130, + 415, + 479, + 450 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 130, + 415, + 479, + 450 + ], + "type": "inline_equation", + "content": "y_{p}" + }, + { + "bbox": [ + 130, + 415, + 479, + 450 + ], + "type": "text", + "content": " will be subsequently used as the input3 to the text encoder " + }, + { + "bbox": [ + 130, + 415, + 479, + 450 + ], + "type": "inline_equation", + "content": "\\tau_{\\theta}" + }, + { + "bbox": [ + 130, + 415, + 479, + 450 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 130, + 415, + 479, + 450 + ], + "type": "inline_equation", + "content": "e \\in \\mathbb{R}^{MK \\times D}" + }, + { + "bbox": [ + 130, + 415, + 479, + 450 + ], + "type": "text", + "content": " is a learnable word embedding dictionary that maps " + }, + { + "bbox": [ + 130, + 415, + 479, + 450 + ], + "type": "inline_equation", + "content": "p" + }, + { + "bbox": [ + 130, + 415, + 479, + 450 + ], + "type": "text", + "content": " to their respective embeddings." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 130, + 451, + 480, + 571 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 451, + 480, + 571 + ], + "spans": [ + { + "bbox": [ + 130, + 451, + 480, + 571 + ], + "type": "text", + "content": "Our design demonstrates quicker convergence than directly learning the final word embeddings " + }, + { + "bbox": [ + 130, + 451, + 480, + 571 + ], + "type": "inline_equation", + "content": "e(\\cdot)" + }, + { + "bbox": [ + 130, + 451, + 480, + 571 + ], + "type": "text", + "content": " [25] (see supplementary material). This could be attributed to the entanglement of word embeddings in the conventional design, where there is no information exchange among them during optimization. For instance, each token doesn't know they are learning for a specific part of a specific species. This lack of communication leads to lower data efficiency and slower learning. With the bottleneck " + }, + { + "bbox": [ + 130, + 451, + 480, + 571 + ], + "type": "inline_equation", + "content": "f" + }, + { + "bbox": [ + 130, + 451, + 480, + 571 + ], + "type": "text", + "content": ", it will first project the token into a common part embedding space (e.g., head), then slightly adjust itself to adapt the fine-grained part details. It's worth noting that the conventional design is a specific instance of our approach when " + }, + { + "bbox": [ + 130, + 451, + 480, + 571 + ], + "type": "inline_equation", + "content": "f" + }, + { + "bbox": [ + 130, + 451, + 480, + 571 + ], + "type": "text", + "content": " is an identity function." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 131, + 589, + 296, + 601 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 131, + 589, + 296, + 601 + ], + "spans": [ + { + "bbox": [ + 131, + 589, + 296, + 601 + ], + "type": "text", + "content": "3.3 Learning to Craft by Parts" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 130, + 609, + 479, + 647 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 609, + 479, + 647 + ], + "spans": [ + { + "bbox": [ + 130, + 609, + 479, + 647 + ], + "type": "text", + "content": "Fine-tuning the T2I model, rather than sorely learning pseudo-words, has been shown to achieve better reconstruction of target concepts as demonstrated in [36, 53]. However, this comes with a significant training cost. Thus, we apply LoRA" + } + ] + } + ], + "index": 10 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 258, + 91, + 448, + 103 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 258, + 91, + 448, + 103 + ], + "spans": [ + { + "bbox": [ + 258, + 91, + 448, + 103 + ], + "type": "text", + "content": "PartCraft: Crafting Creative Objects by Parts" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 474, + 91, + 481, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 474, + 91, + 481, + 100 + ], + "spans": [ + { + "bbox": [ + 474, + 91, + 481, + 100 + ], + "type": "text", + "content": "7" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 133, + 653, + 366, + 666 + ], + "type": "page_footnote", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 653, + 366, + 666 + ], + "spans": [ + { + "bbox": [ + 133, + 653, + 366, + 666 + ], + "type": "text", + "content": "3 Word templates such as \"a photo of a [*]\" will be used." + } + ] + } + ], + "index": 11 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 6 + }, + { + "para_blocks": [ + { + "bbox": [ + 130, + 116, + 479, + 151 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 116, + 479, + 151 + ], + "spans": [ + { + "bbox": [ + 130, + 116, + 479, + 151 + ], + "type": "text", + "content": "(low-rank adaptation) [31] to the cross-attention block for efficient training. We then minimize the diffusion loss " + }, + { + "bbox": [ + 130, + 116, + 479, + 151 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_{ldm}" + }, + { + "bbox": [ + 130, + 116, + 479, + 151 + ], + "type": "text", + "content": " (Eq. (1)) to learn both pseudo-words and LoRA adapters." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 130, + 152, + 480, + 224 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 152, + 480, + 224 + ], + "spans": [ + { + "bbox": [ + 130, + 152, + 480, + 224 + ], + "type": "text", + "content": "While training with only " + }, + { + "bbox": [ + 130, + 152, + 480, + 224 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_{ldm}" + }, + { + "bbox": [ + 130, + 152, + 480, + 224 + ], + "type": "text", + "content": ", entanglement happens between parts, as evident from the attention maps in the cross-attention block of the denoiser " + }, + { + "bbox": [ + 130, + 152, + 480, + 224 + ], + "type": "inline_equation", + "content": "\\epsilon_{\\theta}" + }, + { + "bbox": [ + 130, + 152, + 480, + 224 + ], + "type": "text", + "content": " (see Fig. 9). This entanglement arises due to the correlation between parts (e.g., a bird head code is consistently paired with a bird body code to represent the same species). To address this issue, we introduce an entropy-based attention loss as regularization:" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 186, + 233, + 480, + 247 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 186, + 233, + 480, + 247 + ], + "spans": [ + { + "bbox": [ + 186, + 233, + 480, + 247 + ], + "type": "interline_equation", + "content": "\\mathcal {L} _ {\\text {a t t n}} = \\mathbb {E} _ {z, t, m} \\left[ - \\left(S _ {m} \\log \\hat {A} _ {m} + (1 - S _ {m}) \\log (1 - \\hat {A} _ {m})\\right) \\right], \\tag {5}", + "image_path": "2afd6c756a86e52283d8eab66d1a534891588e39b77da1a0241979b68ba93f41.jpg" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 183, + 251, + 480, + 281 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 183, + 251, + 480, + 281 + ], + "spans": [ + { + "bbox": [ + 183, + 251, + 480, + 281 + ], + "type": "interline_equation", + "content": "\\hat {A} _ {m, i, j} = \\frac {\\bar {A} _ {m , i , j}}{\\sum_ {k} \\bar {A} _ {k , i , j}}, \\quad \\bar {A} _ {m} = \\frac {1}{L} \\sum_ {l} ^ {L} A _ {l, m}, \\tag {6}", + "image_path": "f3ed1647ec821574ec32e924b3b75fb3fde7689c87e44c65c7e93e2738508e40.jpg" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 130, + 292, + 482, + 376 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 292, + 482, + 376 + ], + "spans": [ + { + "bbox": [ + 130, + 292, + 482, + 376 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 130, + 292, + 482, + 376 + ], + "type": "inline_equation", + "content": "A \\in [0,1]^{M \\times HW}" + }, + { + "bbox": [ + 130, + 292, + 482, + 376 + ], + "type": "text", + "content": " represents the cross-attention map between the " + }, + { + "bbox": [ + 130, + 292, + 482, + 376 + ], + "type": "inline_equation", + "content": "m" + }, + { + "bbox": [ + 130, + 292, + 482, + 376 + ], + "type": "text", + "content": "-th part and the noisy latent " + }, + { + "bbox": [ + 130, + 292, + 482, + 376 + ], + "type": "inline_equation", + "content": "z_{t}" + }, + { + "bbox": [ + 130, + 292, + 482, + 376 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 130, + 292, + 482, + 376 + ], + "type": "inline_equation", + "content": "L" + }, + { + "bbox": [ + 130, + 292, + 482, + 376 + ], + "type": "text", + "content": " represents the number of specific layers to select attention maps, " + }, + { + "bbox": [ + 130, + 292, + 482, + 376 + ], + "type": "inline_equation", + "content": "\\hat{A} \\in [0,1]^{M \\times HW}" + }, + { + "bbox": [ + 130, + 292, + 482, + 376 + ], + "type": "text", + "content": " represents the averaged and normalized cross-attention map over all parts and " + }, + { + "bbox": [ + 130, + 292, + 482, + 376 + ], + "type": "inline_equation", + "content": "S_{m} \\in \\{0,1\\}^{M \\times HW}" + }, + { + "bbox": [ + 130, + 292, + 482, + 376 + ], + "type": "text", + "content": " serves as the mask that indicates the location of " + }, + { + "bbox": [ + 130, + 292, + 482, + 376 + ], + "type": "inline_equation", + "content": "m" + }, + { + "bbox": [ + 130, + 292, + 482, + 376 + ], + "type": "text", + "content": "-th part. In cases where the part is not present in the image (e.g., occluded), we set both " + }, + { + "bbox": [ + 130, + 292, + 482, + 376 + ], + "type": "inline_equation", + "content": "S_{m}" + }, + { + "bbox": [ + 130, + 292, + 482, + 376 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 130, + 292, + 482, + 376 + ], + "type": "inline_equation", + "content": "\\hat{A}_{m}" + }, + { + "bbox": [ + 130, + 292, + 482, + 376 + ], + "type": "text", + "content": " as 0 to exclude them. Thus, the overall learning objective is defined as:" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 248, + 387, + 480, + 399 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 248, + 387, + 480, + 399 + ], + "spans": [ + { + "bbox": [ + 248, + 387, + 480, + 399 + ], + "type": "interline_equation", + "content": "\\mathcal {L} _ {\\text {t o t a l}} = \\mathcal {L} _ {l d m} + \\lambda_ {\\text {a t t n}} \\mathcal {L} _ {\\text {a t t n}}, \\tag {7}", + "image_path": "4a77dfbafbc0a2d73872acb0476b33cb5a4ac3444479b7529663d6b788d97286.jpg" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 130, + 407, + 482, + 529 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 407, + 482, + 529 + ], + "spans": [ + { + "bbox": [ + 130, + 407, + 482, + 529 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 130, + 407, + 482, + 529 + ], + "type": "inline_equation", + "content": "\\lambda_{attn} = 0.01" + }, + { + "bbox": [ + 130, + 407, + 482, + 529 + ], + "type": "text", + "content": ". We focus on attention maps at the resolution of " + }, + { + "bbox": [ + 130, + 407, + 482, + 529 + ], + "type": "inline_equation", + "content": "16 \\times 16" + }, + { + "bbox": [ + 130, + 407, + 482, + 529 + ], + "type": "text", + "content": " where rich semantic information is captured [28]. Normalization is performed at each location to ensure that the sum of a patch location equals 1. This aims to maximize the attention of a specific part at a particular location which implicitly minimizing the attention of other parts similar to a softmax classification task. Compared to the mean-square based attention loss [3], this intuitively ensures that a part only appears once at a particular location, facilitating stronger disentanglement from other parts during the denoising operation. When generating a part for a particular location, the diffusion model " + }, + { + "bbox": [ + 130, + 407, + 482, + 529 + ], + "type": "inline_equation", + "content": "\\epsilon_{\\theta}" + }, + { + "bbox": [ + 130, + 407, + 482, + 529 + ], + "type": "text", + "content": " should only attend to the part instead of other non-related parts." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 131, + 541, + 230, + 555 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 131, + 541, + 230, + 555 + ], + "spans": [ + { + "bbox": [ + 131, + 541, + 230, + 555 + ], + "type": "text", + "content": "4 Experiments" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 130, + 567, + 482, + 603 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 567, + 482, + 603 + ], + "spans": [ + { + "bbox": [ + 130, + 567, + 482, + 603 + ], + "type": "text", + "content": "Datasets. We demonstrate our selection task on two fine-grained object datasets: CUB-200-2011 (birds) [62] which contains 5,994 training images, and Stanford Dogs [33] which contains 12,000 training images." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 130, + 605, + 482, + 666 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 605, + 482, + 666 + ], + "spans": [ + { + "bbox": [ + 130, + 605, + 482, + 666 + ], + "type": "text", + "content": "Implementation. For part composition, we assess the model's ability to combine up to 4 different parts from 4 distinct species/objects. We set " + }, + { + "bbox": [ + 130, + 605, + 482, + 666 + ], + "type": "inline_equation", + "content": "M = 5" + }, + { + "bbox": [ + 130, + 605, + 482, + 666 + ], + "type": "text", + "content": " for bird generation (head, front body/breast area, wings, legs, tail) and " + }, + { + "bbox": [ + 130, + 605, + 482, + 666 + ], + "type": "inline_equation", + "content": "M = 7" + }, + { + "bbox": [ + 130, + 605, + 482, + 666 + ], + "type": "text", + "content": " for dog generation (forehead, eyes, mouth/nose, ears, neck, body/tail, legs). For both datasets, " + }, + { + "bbox": [ + 130, + 605, + 482, + 666 + ], + "type": "inline_equation", + "content": "K" + }, + { + "bbox": [ + 130, + 605, + 482, + 666 + ], + "type": "text", + "content": " is set as 256, ensuring sufficient coverage of all fine-grained" + } + ] + } + ], + "index": 11 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "spans": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "type": "text", + "content": "8" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 166, + 91, + 225, + 102 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 166, + 91, + 225, + 102 + ], + "spans": [ + { + "bbox": [ + 166, + 91, + 225, + 102 + ], + "type": "text", + "content": "KW. Ng et al." + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 7 + }, + { + "para_blocks": [ + { + "type": "table", + "bbox": [ + 151, + 114, + 459, + 178 + ], + "blocks": [ + { + "bbox": [ + 151, + 114, + 459, + 178 + ], + "lines": [ + { + "bbox": [ + 151, + 114, + 459, + 178 + ], + "spans": [ + { + "bbox": [ + 151, + 114, + 459, + 178 + ], + "type": "table", + "html": "
MethodLearnable TokenFine TuneDisentanglementBottleneck
Textual Inversion [25]✓✗✗✗
DreamBooth [53]✓LoRA*✗✗
CustomDiffusion [36]✓K/V✗✗
Break-a-scene [3]✓LoRA*MSE✗
PartCraft (Ours)✓LoRAEq. (5)✓
", + "image_path": "0987285f288d20cfba543303676758392d729a124ed36ad26d042f7567494416.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "table_body" + } + ], + "index": 2 + }, + { + "bbox": [ + 130, + 180, + 479, + 213 + ], + "lines": [ + { + "bbox": [ + 130, + 180, + 479, + 213 + ], + "spans": [ + { + "bbox": [ + 130, + 180, + 479, + 213 + ], + "type": "text", + "content": "Table 1: Comparing our and alternative methods in design properties. *: We finetuned the added LoRA [31] adapter rather than the entire diffusion model " + }, + { + "bbox": [ + 130, + 180, + 479, + 213 + ], + "type": "inline_equation", + "content": "\\epsilon_{\\theta}" + }, + { + "bbox": [ + 130, + 180, + 479, + 213 + ], + "type": "text", + "content": " due to resource limit. MSE is a mean-square based attention loss used in [3]." + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "text" + }, + { + "bbox": [ + 130, + 263, + 480, + 335 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 263, + 480, + 335 + ], + "spans": [ + { + "bbox": [ + 130, + 263, + 480, + 335 + ], + "type": "text", + "content": "classes (i.e., 200 for birds and 120 for dogs). We randomly generate 500 images by sampling 500 sets of parts. For each image, we randomly replace an original part with any part from another 500 non-overlapping sets of parts. The resulting set of parts may take the form of " + }, + { + "bbox": [ + 130, + 263, + 480, + 335 + ], + "type": "inline_equation", + "content": "\\left( {0,{k}_{A}}\\right) \\left( {1,{k}_{B}}\\right) \\left( {2,{k}_{C}}\\right) \\ldots \\left( {M,{k}_{D}}\\right)" + }, + { + "bbox": [ + 130, + 263, + 480, + 335 + ], + "type": "text", + "content": " ,representing a composition from species A,B,C,and D. Stable Diffusion v1.5 [52] is used. Please see the supplementary material for further training details." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 130, + 350, + 482, + 445 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 350, + 482, + 445 + ], + "spans": [ + { + "bbox": [ + 130, + 350, + 482, + 445 + ], + "type": "text", + "content": "We compare our method with the recent personalization methods: Textual Inversion (TI) [25], DreamBooth (DB) [53], Custom Diffusion (CD) [36], Break-a-scene (BaS) [3]. These personalization methods were designed to take single or multiple images with associated labeled objects as input. The text prompt for each image is as simple as \"a photo of " + }, + { + "bbox": [ + 130, + 350, + 482, + 445 + ], + "type": "inline_equation", + "content": "[p]" + }, + { + "bbox": [ + 130, + 350, + 482, + 445 + ], + "type": "text", + "content": "\" where " + }, + { + "bbox": [ + 130, + 350, + 482, + 445 + ], + "type": "inline_equation", + "content": "p" + }, + { + "bbox": [ + 130, + 350, + 482, + 445 + ], + "type": "text", + "content": " is expressed in Eq. (3), since we do not rely on complex prompts. We employ the official implementations released by the original authors for training. We summarize the main design properties of all compared methods in Tab. 1." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 130, + 461, + 482, + 666 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 461, + 482, + 666 + ], + "spans": [ + { + "bbox": [ + 130, + 461, + 482, + 666 + ], + "type": "text", + "content": "Evaluation metrics. To assess a model's ability to disentangle and composite parts, we introduce two metrics: (a) exact matching rate (EMR) and (b) cosine similarity (CoSim) between the " + }, + { + "bbox": [ + 130, + 461, + 482, + 666 + ], + "type": "inline_equation", + "content": "k" + }, + { + "bbox": [ + 130, + 461, + 482, + 666 + ], + "type": "text", + "content": "-means embeddings of the parts of real and generated images. Utilizing the pre-trained " + }, + { + "bbox": [ + 130, + 461, + 482, + 666 + ], + "type": "inline_equation", + "content": "k" + }, + { + "bbox": [ + 130, + 461, + 482, + 666 + ], + "type": "text", + "content": "-means from Sec. 3.1, we predict the parts of generated images. EMR quantifies how accurately the cluster index of parts of generated images matches the parts of the corresponding real images whereas CoSim measures the cosine similarity between the " + }, + { + "bbox": [ + 130, + 461, + 482, + 666 + ], + "type": "inline_equation", + "content": "k" + }, + { + "bbox": [ + 130, + 461, + 482, + 666 + ], + "type": "text", + "content": "-means centroid vector that the part belongs to between generated and real images. These metrics assess the model's ability to follow the input parts and accurately reconstruct them, with perfect disentanglement indicated by EMR of 1 and CoSim of 1. A detailed algorithm is provided in the supplementary material. We also measure image generation quality using FID [29] to assess model performance in terms of image distribution. Additionally, we compute the average pairwise cosine similarity between CLIP [48]/DINO [11] embeddings of generated and real class-specific images following [53]. Each generated image is conditioned on the parts of the corresponding real image. This results in 5,994 generated images for birds and 12,000 generated images for dogs." + } + ] + } + ], + "index": 6 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 258, + 91, + 448, + 102 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 258, + 91, + 448, + 102 + ], + "spans": [ + { + "bbox": [ + 258, + 91, + 448, + 102 + ], + "type": "text", + "content": "PartCraft: Crafting Creative Objects by Parts" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 474, + 91, + 480, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 474, + 91, + 480, + 100 + ], + "spans": [ + { + "bbox": [ + 474, + 91, + 480, + 100 + ], + "type": "text", + "content": "9" + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 8 + }, + { + "para_blocks": [ + { + "type": "table", + "bbox": [ + 135, + 113, + 479, + 201 + ], + "blocks": [ + { + "bbox": [ + 135, + 113, + 479, + 201 + ], + "lines": [ + { + "bbox": [ + 135, + 113, + 479, + 201 + ], + "spans": [ + { + "bbox": [ + 135, + 113, + 479, + 201 + ], + "type": "table", + "html": "
MethodBirds: CUB-200-2011Dogs: Stanford Dogs
FIDCLIPDINOEMRCoSimFIDCLIPDINOEMRCoSim
Textual Inversion [25]10.100.7840.6070.3050.84223.360.6520.5320.2180.754
DreamBooth [53]12.940.7750.5940.3550.85622.650.6600.5630.2750.777
Custom Diffusion [36]37.610.6940.5040.3380.83342.410.5930.4910.2530.755
Break-a-Scene [3]20.050.7420.5490.3900.85424.200.6330.5320.3000.775
PartCraft (Ours)12.860.7830.6180.4600.88216.920.6690.5730.3580.796
", + "image_path": "4a9820926c0f178845398e9b0317c31c5961cf7e7b9b43e194301fade74a3471.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "table_body" + } + ], + "index": 2 + }, + { + "bbox": [ + 186, + 204, + 426, + 215 + ], + "lines": [ + { + "bbox": [ + 186, + 204, + 426, + 215 + ], + "spans": [ + { + "bbox": [ + 186, + 204, + 426, + 215 + ], + "type": "text", + "content": "Table 2: Quantitative comparison for part reconstruction." + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "text" + }, + { + "type": "image", + "bbox": [ + 133, + 221, + 202, + 357 + ], + "blocks": [ + { + "bbox": [ + 133, + 221, + 202, + 357 + ], + "lines": [ + { + "bbox": [ + 133, + 221, + 202, + 357 + ], + "spans": [ + { + "bbox": [ + 133, + 221, + 202, + 357 + ], + "type": "image", + "image_path": "2461d17a447063d53bc670ae77ba2c86222d1a10e0b4f5ded508227952c9bdca.jpg" + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 156, + 358, + 181, + 369 + ], + "lines": [ + { + "bbox": [ + 156, + 358, + 181, + 369 + ], + "spans": [ + { + "bbox": [ + 156, + 358, + 181, + 369 + ], + "type": "text", + "content": "(a) TI" + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "image_caption" + } + ], + "index": 4 + }, + { + "type": "image", + "bbox": [ + 203, + 222, + 271, + 357 + ], + "blocks": [ + { + "bbox": [ + 203, + 222, + 271, + 357 + ], + "lines": [ + { + "bbox": [ + 203, + 222, + 271, + 357 + ], + "spans": [ + { + "bbox": [ + 203, + 222, + 271, + 357 + ], + "type": "image", + "image_path": "525a983d3e6f635dbc54686bdefc5eba78e2170f720894cf71cbe0bfcc2c2c58.jpg" + } + ] + } + ], + "index": 6, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 223, + 358, + 251, + 369 + ], + "lines": [ + { + "bbox": [ + 223, + 358, + 251, + 369 + ], + "spans": [ + { + "bbox": [ + 223, + 358, + 251, + 369 + ], + "type": "text", + "content": "(b) DB" + } + ] + } + ], + "index": 7, + "angle": 0, + "type": "image_caption" + } + ], + "index": 6 + }, + { + "type": "image", + "bbox": [ + 272, + 221, + 340, + 357 + ], + "blocks": [ + { + "bbox": [ + 272, + 221, + 340, + 357 + ], + "lines": [ + { + "bbox": [ + 272, + 221, + 340, + 357 + ], + "spans": [ + { + "bbox": [ + 272, + 221, + 340, + 357 + ], + "type": "image", + "image_path": "3797cf22eda737f60fc7102a73197117e5728876b1aedb06ed9568e3e2bd3579.jpg" + } + ] + } + ], + "index": 8, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 293, + 358, + 321, + 368 + ], + "lines": [ + { + "bbox": [ + 293, + 358, + 321, + 368 + ], + "spans": [ + { + "bbox": [ + 293, + 358, + 321, + 368 + ], + "type": "text", + "content": "(c) CD" + } + ] + } + ], + "index": 9, + "angle": 0, + "type": "image_caption" + } + ], + "index": 8 + }, + { + "type": "image", + "bbox": [ + 342, + 222, + 410, + 357 + ], + "blocks": [ + { + "bbox": [ + 342, + 222, + 410, + 357 + ], + "lines": [ + { + "bbox": [ + 342, + 222, + 410, + 357 + ], + "spans": [ + { + "bbox": [ + 342, + 222, + 410, + 357 + ], + "type": "image", + "image_path": "c83aeab8a8fb6ae0ca334696b241f93777adbfb9ffd4947b627896e96b631219.jpg" + } + ] + } + ], + "index": 10, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 361, + 358, + 392, + 368 + ], + "lines": [ + { + "bbox": [ + 361, + 358, + 392, + 368 + ], + "spans": [ + { + "bbox": [ + 361, + 358, + 392, + 368 + ], + "type": "text", + "content": "(d) BaS" + } + ] + } + ], + "index": 11, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 130, + 376, + 480, + 399 + ], + "lines": [ + { + "bbox": [ + 130, + 376, + 480, + 399 + ], + "spans": [ + { + "bbox": [ + 130, + 376, + 480, + 399 + ], + "type": "text", + "content": "Fig. 5: Visual comparison under the part reconstruction setting. All images are generated by using the original parts of respective objects." + } + ] + } + ], + "index": 14, + "angle": 0, + "type": "image_caption" + } + ], + "index": 10 + }, + { + "type": "image", + "bbox": [ + 411, + 222, + 480, + 357 + ], + "blocks": [ + { + "bbox": [ + 411, + 222, + 480, + 357 + ], + "lines": [ + { + "bbox": [ + 411, + 222, + 480, + 357 + ], + "spans": [ + { + "bbox": [ + 411, + 222, + 480, + 357 + ], + "type": "image", + "image_path": "76a89cfde8e5a63f6a92ff9c9eaaa58b2e37f1cb6a11a5c61dd7216815adbfb7.jpg" + } + ] + } + ], + "index": 12, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 421, + 358, + 470, + 369 + ], + "lines": [ + { + "bbox": [ + 421, + 358, + 470, + 369 + ], + "spans": [ + { + "bbox": [ + 421, + 358, + 470, + 369 + ], + "type": "text", + "content": "(e) PartCraft" + } + ] + } + ], + "index": 13, + "angle": 0, + "type": "image_caption" + } + ], + "index": 12 + }, + { + "bbox": [ + 130, + 418, + 417, + 430 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 418, + 417, + 430 + ], + "spans": [ + { + "bbox": [ + 130, + 418, + 417, + 430 + ], + "type": "text", + "content": "4.1 Part Reconstruction and Image Quality Evaluation" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 130, + 438, + 480, + 474 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 438, + 480, + 474 + ], + "spans": [ + { + "bbox": [ + 130, + 438, + 480, + 474 + ], + "type": "text", + "content": "We first assess the ability of different methods to learn parts as text tokens by evaluating how well they can accurately reconstruct the parts (this also means image generation with original parts)." + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 130, + 474, + 481, + 666 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 474, + 481, + 666 + ], + "spans": [ + { + "bbox": [ + 130, + 474, + 481, + 666 + ], + "type": "text", + "content": "In Tab. 2, we summarize the performance of respective methods on the bird and dog generation, respectively. We highlight four observations: (i) Textual Inversion performs quite well compared to DreamBooth, CustomDiffusion, and Break-a-scene in terms of FID, CLIP, and DINO scores although did not finetune the diffusion model " + }, + { + "bbox": [ + 130, + 474, + 481, + 666 + ], + "type": "inline_equation", + "content": "\\epsilon_{\\theta}" + }, + { + "bbox": [ + 130, + 474, + 481, + 666 + ], + "type": "text", + "content": ". This may be due to the potential risk of overfitting when fine-tuning " + }, + { + "bbox": [ + 130, + 474, + 481, + 666 + ], + "type": "inline_equation", + "content": "\\epsilon_{\\theta}" + }, + { + "bbox": [ + 130, + 474, + 481, + 666 + ], + "type": "text", + "content": " especially when learning a vast array of new concepts with many update iterations. It is also not uncommon to carefully tune the learning rate and the training iterations in these models when fine-tuning new concepts (e.g., only 800-1000 steps of updates to learn a new concept in [3]). (ii) Nonetheless, fine-tuning the diffusion model " + }, + { + "bbox": [ + 130, + 474, + 481, + 666 + ], + "type": "inline_equation", + "content": "\\epsilon_{\\theta}" + }, + { + "bbox": [ + 130, + 474, + 481, + 666 + ], + "type": "text", + "content": " can help improve the ability to follow prompts as shown by increased EMR and CoSim scores (e.g., EMR of at least " + }, + { + "bbox": [ + 130, + 474, + 481, + 666 + ], + "type": "inline_equation", + "content": "5\\%" + }, + { + "bbox": [ + 130, + 474, + 481, + 666 + ], + "type": "text", + "content": " in DreamBooth). (iii) Break-a-scene has a better ability to reconstruct the parts as shown by EMR and CoSim, this is due to the attention loss explicitly forcing the parts to focus on the respective semantic region. (iv) PartCraft achieves the best performance in DINO, EMR, and CoSim scores (e.g., " + }, + { + "bbox": [ + 130, + 474, + 481, + 666 + ], + "type": "inline_equation", + "content": "7\\%" + }, + { + "bbox": [ + 130, + 474, + 481, + 666 + ], + "type": "text", + "content": " better in EMR compared to Break-a-scene). This indicates that not only does our" + } + ] + } + ], + "index": 17 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "spans": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "type": "text", + "content": "10" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 166, + 91, + 225, + 102 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 166, + 91, + 225, + 102 + ], + "spans": [ + { + "bbox": [ + 166, + 91, + 225, + 102 + ], + "type": "text", + "content": "KW. Ng et al." + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 9 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 133, + 114, + 481, + 203 + ], + "blocks": [ + { + "bbox": [ + 133, + 114, + 481, + 203 + ], + "lines": [ + { + "bbox": [ + 133, + 114, + 481, + 203 + ], + "spans": [ + { + "bbox": [ + 133, + 114, + 481, + 203 + ], + "type": "image", + "image_path": "cb410159b384692c2139554d69c2674923f06e73f46b6f4d561006ff1e811c7f.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 135, + 209, + 477, + 221 + ], + "lines": [ + { + "bbox": [ + 135, + 209, + 477, + 221 + ], + "spans": [ + { + "bbox": [ + 135, + 209, + 477, + 221 + ], + "type": "text", + "content": "Fig. 6: Quantitative comparisons of part composition in terms of EMR and CoSim." + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_caption" + } + ], + "index": 2 + }, + { + "bbox": [ + 130, + 233, + 482, + 306 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 233, + 482, + 306 + ], + "spans": [ + { + "bbox": [ + 130, + 233, + 482, + 306 + ], + "type": "text", + "content": "image-generation ability perform comparably well with textual inversion, but PartCraft is also able to disentangle the parts learning so that it can follow the prompt instructions more accurately to generate the parts in a cohort. In Fig. 5, we present generated images from different methods, with CustomDiffusion exhibiting high-contrast images, possibly due to unconstrained fine-tuning on the cross-attention components " + }, + { + "bbox": [ + 130, + 233, + 482, + 306 + ], + "type": "inline_equation", + "content": "K / V" + }, + { + "bbox": [ + 130, + 233, + 482, + 306 + ], + "type": "text", + "content": " and resulting in worse FID scores." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 132, + 322, + 308, + 335 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 322, + 308, + 335 + ], + "spans": [ + { + "bbox": [ + 132, + 322, + 308, + 335 + ], + "type": "text", + "content": "4.2 Part Composition Evaluation" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 130, + 341, + 482, + 437 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 341, + 482, + 437 + ], + "spans": [ + { + "bbox": [ + 130, + 341, + 482, + 437 + ], + "type": "text", + "content": "In this section, we assess the part composition ability of different methods. In this experiment, we generate the image by mixing different parts from different species. Our findings, as shown in Fig. 6, can be summarized as follows: (i) As the number of composited parts increases, EMR and CoSim decrease, reflecting the challenge of composing multiple diverse parts. (ii) Break-a-scene and PartCraft achieve notably higher EMR and CoSim scores, thanks to disentanglement through attention loss minimization. (iii) PartCraft outperforms Break-a-scene significantly by token bottleneck and tailored attention loss." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 130, + 437, + 483, + 520 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 437, + 483, + 520 + ], + "spans": [ + { + "bbox": [ + 130, + 437, + 483, + 520 + ], + "type": "text", + "content": "In Fig. 7, we visualize the results of composing 4 different parts. While all images appear realistic, most methods struggle to assemble all 4 parts. For instance, Break-a-scene missed out on the flurry body of kerry blue terrier (right-most column). In contrast, our methods successfully combine 4 different parts from 4 different species, demonstrating the superior ability of our approach to part composition. We also visualize additional examples of our method in the supplementary material." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 130, + 521, + 482, + 582 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 521, + 482, + 582 + ], + "spans": [ + { + "bbox": [ + 130, + 521, + 482, + 582 + ], + "type": "text", + "content": "Furthermore, we explore the versatility of the adapted model by generating images with simple styles such as pencil drawing. While most methods successfully incorporate specific styles into the generated image, Custom Diffusion often fails to do so, possibly due to the unconstrained fine-tuning of the cross-attention components " + }, + { + "bbox": [ + 130, + 521, + 482, + 582 + ], + "type": "inline_equation", + "content": "K / V" + }, + { + "bbox": [ + 130, + 521, + 482, + 582 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 132, + 597, + 246, + 609 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 597, + 246, + 609 + ], + "spans": [ + { + "bbox": [ + 132, + 597, + 246, + 609 + ], + "type": "text", + "content": "4.3 Ablation Studies" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 130, + 617, + 482, + 666 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 617, + 482, + 666 + ], + "spans": [ + { + "bbox": [ + 130, + 617, + 482, + 666 + ], + "type": "text", + "content": "Component analysis. In Fig. 8, we evaluate the effect of our proposed components (token bottleneck and attention loss) on creating novel bird species. (i) Removing the bottleneck outlined in Eq. (4) degrades the generation quality as evidenced by a higher FID score " + }, + { + "bbox": [ + 130, + 617, + 482, + 666 + ], + "type": "inline_equation", + "content": "(12.86 \\rightarrow 16.36)" + }, + { + "bbox": [ + 130, + 617, + 482, + 666 + ], + "type": "text", + "content": " even though both EMR" + } + ] + } + ], + "index": 10 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 258, + 91, + 448, + 103 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 258, + 91, + 448, + 103 + ], + "spans": [ + { + "bbox": [ + 258, + 91, + 448, + 103 + ], + "type": "text", + "content": "PartCraft: Crafting Creative Objects by Parts" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 470, + 91, + 479, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 470, + 91, + 479, + 100 + ], + "spans": [ + { + "bbox": [ + 470, + 91, + 479, + 100 + ], + "type": "text", + "content": "11" + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 10 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 133, + 113, + 481, + 500 + ], + "blocks": [ + { + "bbox": [ + 133, + 113, + 481, + 500 + ], + "lines": [ + { + "bbox": [ + 133, + 113, + 481, + 500 + ], + "spans": [ + { + "bbox": [ + 133, + 113, + 481, + 500 + ], + "type": "image", + "image_path": "258f2c1e1980ccfb10ad38c91986c23f8aabb54eaea0a085f6905dca02ef4683.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 130, + 507, + 482, + 541 + ], + "lines": [ + { + "bbox": [ + 130, + 507, + 482, + 541 + ], + "spans": [ + { + "bbox": [ + 130, + 507, + 482, + 541 + ], + "type": "text", + "content": "Fig. 7: Visual comparison on 4-species (specified on the top row) mixed generation. The last column indicates generated images with different styles (i.e., " + }, + { + "bbox": [ + 130, + 507, + 482, + 541 + ], + "type": "inline_equation", + "content": "DSLR" + }, + { + "bbox": [ + 130, + 507, + 482, + 541 + ], + "type": "text", + "content": ", Van Gogh, Oil Painting, Pencil Drawing)." + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_caption" + } + ], + "index": 2 + }, + { + "bbox": [ + 130, + 575, + 482, + 636 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 575, + 482, + 636 + ], + "spans": [ + { + "bbox": [ + 130, + 575, + 482, + 636 + ], + "type": "text", + "content": "and CoSim remain. (ii) By replacing our " + }, + { + "bbox": [ + 130, + 575, + 482, + 636 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_{attn}" + }, + { + "bbox": [ + 130, + 575, + 482, + 636 + ], + "type": "text", + "content": " with the MSE loss as proposed in [3], we observe significant deterioration in both EMR and CoSim. (iii) Finally, incorporating both the projector and our attention loss performs the best. This improvement highlights the necessity of incorporating interactions between multiple parts to achieve more effective part disentanglement and optimization." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 130, + 641, + 482, + 666 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 641, + 482, + 666 + ], + "spans": [ + { + "bbox": [ + 130, + 641, + 482, + 666 + ], + "type": "text", + "content": "Cross-attention visualization. Our attention loss plays a crucial role in token disentanglement. We demonstrate the impact of this loss in Fig. 9, where we" + } + ] + } + ], + "index": 5 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "spans": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "type": "text", + "content": "12" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 166, + 90, + 225, + 102 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 166, + 90, + 225, + 102 + ], + "spans": [ + { + "bbox": [ + 166, + 90, + 225, + 102 + ], + "type": "text", + "content": "KW. Ng et al." + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 11 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 133, + 114, + 304, + 185 + ], + "blocks": [ + { + "bbox": [ + 133, + 114, + 304, + 185 + ], + "lines": [ + { + "bbox": [ + 133, + 114, + 304, + 185 + ], + "spans": [ + { + "bbox": [ + 133, + 114, + 304, + 185 + ], + "type": "image", + "image_path": "d489f73860ad7af2326e6c5eff7b9f787e900d76288dc443dfe88408a3be9928.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 307, + 180, + 480, + 213 + ], + "lines": [ + { + "bbox": [ + 307, + 180, + 480, + 213 + ], + "spans": [ + { + "bbox": [ + 307, + 180, + 480, + 213 + ], + "type": "text", + "content": "Fig.9: Cross-attention map of each part (top) without and (bottom) with our attention loss." + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "image_caption" + } + ], + "index": 2 + }, + { + "type": "image", + "bbox": [ + 309, + 128, + 482, + 174 + ], + "blocks": [ + { + "bbox": [ + 131, + 196, + 304, + 229 + ], + "lines": [ + { + "bbox": [ + 131, + 196, + 304, + 229 + ], + "spans": [ + { + "bbox": [ + 131, + 196, + 304, + 229 + ], + "type": "text", + "content": "Fig.8: Ablation on our part token bottleneck and attention loss under the part composition on CUB-200-2011 birds." + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 309, + 128, + 482, + 174 + ], + "lines": [ + { + "bbox": [ + 309, + 128, + 482, + 174 + ], + "spans": [ + { + "bbox": [ + 309, + 128, + 482, + 174 + ], + "type": "image", + "image_path": "c6d1fbb7d14a0a6f64c8532f51b664666b768c548fc6106232cf540a2a5d45c8.jpg" + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "image_body" + } + ], + "index": 4 + }, + { + "type": "image", + "bbox": [ + 138, + 256, + 242, + 339 + ], + "blocks": [ + { + "bbox": [ + 138, + 256, + 242, + 339 + ], + "lines": [ + { + "bbox": [ + 138, + 256, + 242, + 339 + ], + "spans": [ + { + "bbox": [ + 138, + 256, + 242, + 339 + ], + "type": "image", + "image_path": "0880ca23d405fb532d96c74a65064acb8cd1bf1482e8482d2a97615297a63633.jpg" + } + ] + } + ], + "index": 6, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 151, + 346, + 227, + 355 + ], + "lines": [ + { + "bbox": [ + 151, + 346, + 227, + 355 + ], + "spans": [ + { + "bbox": [ + 151, + 346, + 227, + 355 + ], + "type": "text", + "content": "(a) Textual Inversion" + } + ] + } + ], + "index": 7, + "angle": 0, + "type": "image_caption" + } + ], + "index": 6 + }, + { + "type": "image", + "bbox": [ + 256, + 256, + 360, + 340 + ], + "blocks": [ + { + "bbox": [ + 256, + 256, + 360, + 340 + ], + "lines": [ + { + "bbox": [ + 256, + 256, + 360, + 340 + ], + "spans": [ + { + "bbox": [ + 256, + 256, + 360, + 340 + ], + "type": "image", + "image_path": "e1c58d70a21aa6c21e133b0b69ca32b989a2437b162aab0c753b51860fc799c3.jpg" + } + ] + } + ], + "index": 8, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 274, + 346, + 339, + 355 + ], + "lines": [ + { + "bbox": [ + 274, + 346, + 339, + 355 + ], + "spans": [ + { + "bbox": [ + 274, + 346, + 339, + 355 + ], + "type": "text", + "content": "(b) Break-a-scene" + } + ] + } + ], + "index": 9, + "angle": 0, + "type": "image_caption" + } + ], + "index": 8 + }, + { + "type": "image", + "bbox": [ + 373, + 256, + 478, + 341 + ], + "blocks": [ + { + "bbox": [ + 373, + 256, + 478, + 341 + ], + "lines": [ + { + "bbox": [ + 373, + 256, + 478, + 341 + ], + "spans": [ + { + "bbox": [ + 373, + 256, + 478, + 341 + ], + "type": "image", + "image_path": "313f6382013f8b85457e6d24af5d5f9acb802c9bc491b7cbddfa00cfce25161e.jpg" + } + ] + } + ], + "index": 10, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 387, + 346, + 461, + 355 + ], + "lines": [ + { + "bbox": [ + 387, + 346, + 461, + 355 + ], + "spans": [ + { + "bbox": [ + 387, + 346, + 461, + 355 + ], + "type": "text", + "content": "(c) PartCraft (Ours)" + } + ] + } + ], + "index": 11, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 130, + 364, + 480, + 387 + ], + "lines": [ + { + "bbox": [ + 130, + 364, + 480, + 387 + ], + "spans": [ + { + "bbox": [ + 130, + 364, + 480, + 387 + ], + "type": "text", + "content": "Fig. 10: 2D tSNE [39] projection of word embeddings. Different colors represent different common parts. (correspond to the segmentation mask in Fig. 4)." + } + ] + } + ], + "index": 12, + "angle": 0, + "type": "image_caption" + } + ], + "index": 10 + }, + { + "bbox": [ + 130, + 424, + 480, + 449 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 424, + 480, + 449 + ], + "spans": [ + { + "bbox": [ + 130, + 424, + 480, + 449 + ], + "type": "text", + "content": "observe significantly enhanced disentanglement after explicitly guiding attention to focus on distinct semantic regions." + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 130, + 455, + 482, + 552 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 455, + 482, + 552 + ], + "spans": [ + { + "bbox": [ + 130, + 455, + 482, + 552 + ], + "type": "text", + "content": "Part word embedding space. We visualize the word embeddings of learned tokens of Textual Inversion [25], Break-a-scene [3] and our PartCraft for birds generation (CUB-200-2011 [62]) through tSNE [39] in Fig. 10. In our PartCraft, the word embeddings are the projected embeddings through Eq. (4). We can see that our projected version has a better semantic meaning such that the part embeddings are clustered together by their semantic meaning (e.g., head). We believe this is one of the reasons that our PartCraft outperforms previous methods in which we can compose all parts seamlessly yet with higher quality." + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 130, + 557, + 482, + 666 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 557, + 482, + 666 + ], + "spans": [ + { + "bbox": [ + 130, + 557, + 482, + 666 + ], + "type": "text", + "content": "Transferability for Creativity. (i) In Fig. 11, we demonstrate that not only it can compose parts within the training domain (e.g., birds), but it can also transfer the learned parts to and combine with other domains (e.g., cat). This enables the creation of unique combinations, such as a cat with a dog's ear. (ii) Leveraging the prior knowledge embedded in Stable Diffusion, PartCraft can also repurpose learned parts for creative image generation. An example of this is the generation of a bird-shaped robot adorned with various parts. These examples showcase PartCraft's immense potential for diverse and limitless creative applications. Please see the supplementary material for more examples." + } + ] + } + ], + "index": 15 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 258, + 91, + 448, + 102 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 258, + 91, + 448, + 102 + ], + "spans": [ + { + "bbox": [ + 258, + 91, + 448, + 102 + ], + "type": "text", + "content": "PartCraft: Crafting Creative Objects by Parts" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 470, + 91, + 481, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 470, + 91, + 481, + 100 + ], + "spans": [ + { + "bbox": [ + 470, + 91, + 481, + 100 + ], + "type": "text", + "content": "13" + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 12 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 144, + 114, + 470, + 242 + ], + "blocks": [ + { + "bbox": [ + 144, + 114, + 470, + 242 + ], + "lines": [ + { + "bbox": [ + 144, + 114, + 470, + 242 + ], + "spans": [ + { + "bbox": [ + 144, + 114, + 470, + 242 + ], + "type": "image", + "image_path": "1fc2c2041a4f1456f00ab034420aaa44a14a9f89290790e8b24d13acb8381d53.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 130, + 250, + 482, + 307 + ], + "lines": [ + { + "bbox": [ + 130, + 250, + 482, + 307 + ], + "spans": [ + { + "bbox": [ + 130, + 250, + 482, + 307 + ], + "type": "text", + "content": "Fig. 11: (Top): Using learned parts to modify the property of other domains such as cat, and lion with prompt such as \"A cat with [beagle's ear)\". (Bottom) We can also repurpose learned parts for creative image generation using prior knowledge in Stable Diffusion with prompt such as \"A robot designed inspired by [red header woodpecker's head] and [blue jay's body)\"." + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_caption" + } + ], + "index": 2 + }, + { + "bbox": [ + 132, + 315, + 219, + 328 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 315, + 219, + 328 + ], + "spans": [ + { + "bbox": [ + 132, + 315, + 219, + 328 + ], + "type": "text", + "content": "5 Conclusion" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 130, + 337, + 482, + 492 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 337, + 482, + 492 + ], + "spans": [ + { + "bbox": [ + 130, + 337, + 482, + 492 + ], + "type": "text", + "content": "We propose a new way of control in generative AI. Instead of text or sketch, we \"select\" desired parts to create an object. We addressed the challenge of learning parts in T2I models by introducing a customized attention loss. This loss serves a dual purpose: to ensure parts are at the right location and to ensure each location is occupied by not more than one part. This greatly improves the part disentanglement. We further employ a non-linear bottleneck encoder to improve generation fidelity. Our model, PartCraft, can seamlessly compose different parts from different objects, creating objects that do not exist yet holistically correct and plausible objects by mixing them. Extensive experiments demonstrated PartCraft's superior performance in both qualitative and quantitative evaluation. Moreover, the learned parts demonstrate strong transferability. We hope that our PartCraft will empower artists, designers, and enthusiasts to bring the creations of their dreams to reality." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 132, + 510, + 334, + 522 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 510, + 334, + 522 + ], + "spans": [ + { + "bbox": [ + 132, + 510, + 334, + 522 + ], + "type": "text", + "content": "6 Limitations and Future Works" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 130, + 533, + 482, + 666 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 533, + 482, + 666 + ], + "spans": [ + { + "bbox": [ + 130, + 533, + 482, + 666 + ], + "type": "text", + "content": "It is worth noting that the accuracy of obtained parts may be affected by using a self-supervised pre-trained feature extractor, incorporating encoders [64] may help. We also observed challenges in composing relatively small parts, like tails and legs. Additionally, we are also exploring cross-domain generation, i.e., combining learned parts from different datasets to create objects with even more diverse parts. For instance, we can merge non-rigid parts (e.g., parts from quadrupled animals) with rigid parts (e.g., parts from cars/airplanes) and form a creative structure (e.g., a car that has horse legs instead of wheels). This not only further improves the applicability of PartCraft but also serves as a stepping stone to creative generation, as generative AI progresses, continually expanding the limits of achievable creativity and artistic expression." + } + ] + } + ], + "index": 7 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "spans": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "type": "text", + "content": "14" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 166, + 91, + 225, + 102 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 166, + 91, + 225, + 102 + ], + "spans": [ + { + "bbox": [ + 166, + 91, + 225, + 102 + ], + "type": "text", + "content": "KW. Ng et al." + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 13 + }, + { + "para_blocks": [ + { + "bbox": [ + 132, + 114, + 246, + 129 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 114, + 246, + 129 + ], + "spans": [ + { + "bbox": [ + 132, + 114, + 246, + 129 + ], + "type": "text", + "content": "Acknowledgements" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 130, + 141, + 482, + 190 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 141, + 482, + 190 + ], + "spans": [ + { + "bbox": [ + 130, + 141, + 482, + 190 + ], + "type": "text", + "content": "We extend our special thanks to Jia Wei Sii for her help in creating figures and discussing the main concept. We are also grateful to Ruoyi Du, Zhiyu Qu, Chee Seng Chan, and the reviewers for their fruitful comments and corrections on our draft, methodology, and experiments." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 132, + 209, + 197, + 220 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 209, + 197, + 220 + ], + "spans": [ + { + "bbox": [ + 132, + 209, + 197, + 220 + ], + "type": "text", + "content": "References" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 133, + 234, + 480, + 665 + ], + "type": "list", + "angle": 0, + "index": 22, + "blocks": [ + { + "bbox": [ + 138, + 234, + 480, + 256 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 234, + 480, + 256 + ], + "spans": [ + { + "bbox": [ + 138, + 234, + 480, + 256 + ], + "type": "text", + "content": "1. Alaluf, Y., Richardson, E., Metzer, G., Cohen-Or, D.: A neural space-time representation for text-to-image personalization. arXiv preprint arXiv:2305.15391 (2023)" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 138, + 257, + 480, + 278 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 257, + 480, + 278 + ], + "spans": [ + { + "bbox": [ + 138, + 257, + 480, + 278 + ], + "type": "text", + "content": "2. Amir, S., Gandelsman, Y., Bagon, S., Dekel, T.: On the effectiveness of vit features as local semantic descriptors. In: ECCV (2022)" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 138, + 279, + 480, + 300 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 279, + 480, + 300 + ], + "spans": [ + { + "bbox": [ + 138, + 279, + 480, + 300 + ], + "type": "text", + "content": "3. Avrahami, O., Aberman, K., Fried, O., Cohen-Or, D., Lischinski, D.: Break-a-scene: Extracting multiple concepts from a single image. In: SIGGRAPH Asia (2023)" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 138, + 301, + 480, + 312 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 301, + 480, + 312 + ], + "spans": [ + { + "bbox": [ + 138, + 301, + 480, + 312 + ], + "type": "text", + "content": "4. Avrahami, O., Fried, O., Lischinski, D.: Blended latent diffusion. ACM TOG (2023)" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 138, + 312, + 480, + 344 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 312, + 480, + 344 + ], + "spans": [ + { + "bbox": [ + 138, + 312, + 480, + 344 + ], + "type": "text", + "content": "5. Avrahami, O., Hayes, T., Gafni, O., Gupta, S., Taigman, Y., Parikh, D., Lischinski, D., Fried, O., Yin, X.: Spatext: Spatio-textual representation for controllable image generation. In: CVPR (2023)" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 138, + 346, + 480, + 378 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 346, + 480, + 378 + ], + "spans": [ + { + "bbox": [ + 138, + 346, + 480, + 378 + ], + "type": "text", + "content": "6. Balaji, Y., Nah, S., Huang, X., Vahdat, A., Song, J., Kreis, K., Aittala, M., Aila, T., Laine, S., Catanzaro, B., et al.: edifi: Text-to-image diffusion models with an ensemble of expert denoisers. arXiv preprint arXiv:2211.01324 (2022)" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 138, + 379, + 480, + 399 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 379, + 480, + 399 + ], + "spans": [ + { + "bbox": [ + 138, + 379, + 480, + 399 + ], + "type": "text", + "content": "7. Bar-Tal, O., Yariv, L., Lipman, Y., Dekel, T.: Multidiffusion: Fusing diffusion paths for controlled image generation. In: ICML (2023)" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 138, + 401, + 480, + 422 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 401, + 480, + 422 + ], + "spans": [ + { + "bbox": [ + 138, + 401, + 480, + 422 + ], + "type": "text", + "content": "8. Biederman, I.: Recognition-by-components: a theory of human image understanding. Psychological review (1987)" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 138, + 423, + 480, + 455 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 423, + 480, + 455 + ], + "spans": [ + { + "bbox": [ + 138, + 423, + 480, + 455 + ], + "type": "text", + "content": "9. Bonnardel, N., Marmèche, E.: Towards supporting evocation processes in creative design: A cognitive approach. International Journal of Human-Computer Studies (2005)" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 133, + 456, + 480, + 477 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 456, + 480, + 477 + ], + "spans": [ + { + "bbox": [ + 133, + 456, + 480, + 477 + ], + "type": "text", + "content": "10. Brooks, T., Holynski, A., Efros, A.A.: Instructpix2pix: Learning to follow image editing instructions. In: CVPR (2023)" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 133, + 478, + 480, + 500 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 478, + 480, + 500 + ], + "spans": [ + { + "bbox": [ + 133, + 478, + 480, + 500 + ], + "type": "text", + "content": "11. Caron, M., Touvron, H., Misra, I., Jégou, H., Mairal, J., Bojanowski, P., Joulin, A.: Emerging properties in self-supervised vision transformers. In: ICCV (2021)" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 133, + 501, + 480, + 532 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 501, + 480, + 532 + ], + "spans": [ + { + "bbox": [ + 133, + 501, + 480, + 532 + ], + "type": "text", + "content": "12. Cetinic, E., She, J.: Understanding and creating art with ai: Review and outlook. ACM Transactions on Multimedia Computing, Communications, and Applications (2022)" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 133, + 533, + 480, + 565 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 533, + 480, + 565 + ], + "spans": [ + { + "bbox": [ + 133, + 533, + 480, + 565 + ], + "type": "text", + "content": "13. Chefer, H., Alaluf, Y., Vinker, Y., Wolf, L., Cohen-Or, D.: Attend-and-excite: Attention-based semantic guidance for text-to-image diffusion models. ACM TOG (2023)" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 133, + 566, + 480, + 588 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 566, + 480, + 588 + ], + "spans": [ + { + "bbox": [ + 133, + 566, + 480, + 588 + ], + "type": "text", + "content": "14. Chen, C., Li, O., Tao, D., Barnett, A., Rudin, C., Su, J.K.: This looks like that: deep learning for interpretable image recognition. NeurIPS (2019)" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 133, + 589, + 480, + 610 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 589, + 480, + 610 + ], + "spans": [ + { + "bbox": [ + 133, + 589, + 480, + 610 + ], + "type": "text", + "content": "15. Chen, M., Laina, I., Vedaldi, A.: Training-free layout control with cross-attention guidance (2024)" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 133, + 611, + 480, + 643 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 611, + 480, + 643 + ], + "spans": [ + { + "bbox": [ + 133, + 611, + 480, + 643 + ], + "type": "text", + "content": "16. Cintas, C., Das, P., Quanz, B., Tadesse, G.A., Speakman, S., Chen, P.Y.: Towards creativity characterization of generative models via group-based subset scanning. In: IJCAI (2022)" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 133, + 643, + 480, + 665 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 643, + 480, + 665 + ], + "spans": [ + { + "bbox": [ + 133, + 643, + 480, + 665 + ], + "type": "text", + "content": "17. Couairon, G., Verbeek, J., Schwenk, H., Cord, M.: Diffedit: Diffusion-based semantic image editing with mask guidance. In: ICLR (2023)" + } + ] + } + ], + "index": 21 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 258, + 91, + 447, + 102 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 258, + 91, + 447, + 102 + ], + "spans": [ + { + "bbox": [ + 258, + 91, + 447, + 102 + ], + "type": "text", + "content": "PartCraft: Crafting Creative Objects by Parts" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 470, + 91, + 479, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 470, + 91, + 479, + 100 + ], + "spans": [ + { + "bbox": [ + 470, + 91, + 479, + 100 + ], + "type": "text", + "content": "15" + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 14 + }, + { + "para_blocks": [ + { + "bbox": [ + 132, + 116, + 480, + 665 + ], + "type": "list", + "angle": 0, + "index": 24, + "blocks": [ + { + "bbox": [ + 133, + 116, + 480, + 138 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 116, + 480, + 138 + ], + "spans": [ + { + "bbox": [ + 133, + 116, + 480, + 138 + ], + "type": "text", + "content": "18. Das, P., Quanz, B., Chen, P.Y., Ahn, J.w., Shah, D.: Toward a neuro-inspired creative decoder. In: IJCAI (2020)" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 132, + 140, + 480, + 172 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 140, + 480, + 172 + ], + "spans": [ + { + "bbox": [ + 132, + 140, + 480, + 172 + ], + "type": "text", + "content": "19. Ding, M., Yang, Z., Hong, W., Zheng, W., Zhou, C., Yin, D., Lin, J., Zou, X., Shao, Z., Yang, H., et al.: Cogview: Mastering text-to-image generation via transformers. In: NeurIPS (2021)" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 132, + 173, + 480, + 194 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 173, + 480, + 194 + ], + "spans": [ + { + "bbox": [ + 132, + 173, + 480, + 194 + ], + "type": "text", + "content": "20. Ding, M., Zheng, W., Hong, W., Tang, J.: Cogview2: Faster and better text-to-image generation via hierarchical transformers. In: NeurIPS. vol. 35 (2022)" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 132, + 195, + 480, + 228 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 195, + 480, + 228 + ], + "spans": [ + { + "bbox": [ + 132, + 195, + 480, + 228 + ], + "type": "text", + "content": "21. Elgammal, A., Liu, B., Elhoseiny, M., Mazzone, M.: Can: Creative adversarial networks generating \"art\" by learning about styles and deviating from style norms. In: ICCC (2017)" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 132, + 228, + 480, + 250 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 228, + 480, + 250 + ], + "spans": [ + { + "bbox": [ + 132, + 228, + 480, + 250 + ], + "type": "text", + "content": "22. Felzenszwalb, P.F., Huttenlocher, D.P.: Pictorial structures for object recognition. IJCV (2005)" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 132, + 251, + 480, + 284 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 251, + 480, + 284 + ], + "spans": [ + { + "bbox": [ + 132, + 251, + 480, + 284 + ], + "type": "text", + "content": "23. Feng, W., He, X., Fu, T.J., Jampani, V., Akula, A.R., Narayana, P., Basu, S., Wang, X.E., Wang, W.Y.: Training-free structured diffusion guidance for compositional text-to-image synthesis. In: ICLR (2022)" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 132, + 285, + 480, + 306 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 285, + 480, + 306 + ], + "spans": [ + { + "bbox": [ + 132, + 285, + 480, + 306 + ], + "type": "text", + "content": "24. Gafni, O., Polyak, A., Ashual, O., Sheynin, S., Parikh, D., Taigman, Y.: Make-ascene: Scene-based text-to-image generation with human priors. In: ECCV (2022)" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 132, + 307, + 480, + 340 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 307, + 480, + 340 + ], + "spans": [ + { + "bbox": [ + 132, + 307, + 480, + 340 + ], + "type": "text", + "content": "25. Gal, R., Alaluf, Y., Atzmon, Y., Patashnik, O., Bermano, A.H., Chechik, G., Cohen-Or, D.: An image is worth one word: Personalizing text-to-image generation using textual inversion. In: ICLR (2023)" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 132, + 340, + 480, + 361 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 340, + 480, + 361 + ], + "spans": [ + { + "bbox": [ + 132, + 340, + 480, + 361 + ], + "type": "text", + "content": "26. Ge, S., Goswami, V., Zitnick, C.L., Parikh, D.: Creative sketch generation. In: ICLR (2021)" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 132, + 362, + 480, + 384 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 362, + 480, + 384 + ], + "spans": [ + { + "bbox": [ + 132, + 362, + 480, + 384 + ], + "type": "text", + "content": "27. He, J., Chen, J., Lin, M.X., Yu, Q., Yuille, A.L.: Compositor: Bottom-up clustering and compositing for robust part and object segmentation. In: CVPR (2023)" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 132, + 385, + 480, + 407 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 385, + 480, + 407 + ], + "spans": [ + { + "bbox": [ + 132, + 385, + 480, + 407 + ], + "type": "text", + "content": "28. Hertz, A., Mokady, R., Tenenbaum, J., Aberman, K., Pritch, Y., Cohen-Or, D.: Prompt-to-prompt image editing with cross attention control. In: ICLR (2023)" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 132, + 407, + 480, + 440 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 407, + 480, + 440 + ], + "spans": [ + { + "bbox": [ + 132, + 407, + 480, + 440 + ], + "type": "text", + "content": "29. Heusel, M., Ramsauer, H., Unterthiner, T., Nessler, B., Hochreiter, S.: Gans trained by a two time-scale update rule converge to a local nash equilibrium. NeurIPS 30 (2017)" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 132, + 441, + 480, + 462 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 441, + 480, + 462 + ], + "spans": [ + { + "bbox": [ + 132, + 441, + 480, + 462 + ], + "type": "text", + "content": "30. Ho, J., Jain, A., Abbeel, P.: Denoising diffusion probabilistic models. In: NeurIPS (2020)" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 132, + 463, + 480, + 485 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 463, + 480, + 485 + ], + "spans": [ + { + "bbox": [ + 132, + 463, + 480, + 485 + ], + "type": "text", + "content": "31. Hu, E.J., Shen, Y., Wallis, P., Allen-Zhu, Z., Li, Y., Wang, S., Wang, L., Chen, W.: LoRA: Low-rank adaptation of large language models. In: ICLR (2022)" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 132, + 486, + 480, + 508 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 486, + 480, + 508 + ], + "spans": [ + { + "bbox": [ + 132, + 486, + 480, + 508 + ], + "type": "text", + "content": "32. Kawar, B., Zada, S., Lang, O., Tov, O., Chang, H., Dekel, T., Mosseri, I., Irani, M.: Imagic: Text-based real image editing with diffusion models. In: CVPR (2023)" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 132, + 508, + 480, + 530 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 508, + 480, + 530 + ], + "spans": [ + { + "bbox": [ + 132, + 508, + 480, + 530 + ], + "type": "text", + "content": "33. Khosla, A., Jayadevaprakash, N., Yao, B., Li, F.F.: Novel dataset for fine-grained image categorization: Stanford dogs. In: CVPRW (2011)" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 132, + 531, + 480, + 552 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 531, + 480, + 552 + ], + "spans": [ + { + "bbox": [ + 132, + 531, + 480, + 552 + ], + "type": "text", + "content": "34. Kim, Y., Lee, J., Kim, J.H., Ha, J.W., Zhu, J.Y.: Dense text-to-image generation with attention modulation. In: ICCV (2023)" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 132, + 553, + 480, + 575 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 553, + 480, + 575 + ], + "spans": [ + { + "bbox": [ + 132, + 553, + 480, + 575 + ], + "type": "text", + "content": "35. Krause, J., Jin, H., Yang, J., Fei-Fei, L.: Fine-grained recognition without part annotations. In: CVPR (2015)" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 132, + 576, + 480, + 597 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 576, + 480, + 597 + ], + "spans": [ + { + "bbox": [ + 132, + 576, + 480, + 597 + ], + "type": "text", + "content": "36. Kumari, N., Zhang, B., Zhang, R., Shechtman, E., Zhu, J.Y.: Multi-concept customization of text-to-image diffusion. In: CVPR (2023)" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 132, + 598, + 480, + 620 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 598, + 480, + 620 + ], + "spans": [ + { + "bbox": [ + 132, + 598, + 480, + 620 + ], + "type": "text", + "content": "37. Li, B., Qi, X., Lukasiewicz, T., Torr, P.: Controllable text-to-image generation. In: NeurIPS (2019)" + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 132, + 620, + 480, + 643 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 620, + 480, + 643 + ], + "spans": [ + { + "bbox": [ + 132, + 620, + 480, + 643 + ], + "type": "text", + "content": "38. Li, Y., Liu, H., Wu, Q., Mu, F., Yang, J., Gao, J., Li, C., Lee, Y.J.: Gligen: Open-set grounded text-to-image generation. In: CVPR (2023)" + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 132, + 643, + 480, + 665 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 643, + 480, + 665 + ], + "spans": [ + { + "bbox": [ + 132, + 643, + 480, + 665 + ], + "type": "text", + "content": "39. Van der Maaten, L., Hinton, G.: Visualizing data using t-sne. Journal of Machine Learning Research (2008)" + } + ] + } + ], + "index": 23 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "spans": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "type": "text", + "content": "16" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 166, + 91, + 225, + 102 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 166, + 91, + 225, + 102 + ], + "spans": [ + { + "bbox": [ + 166, + 91, + 225, + 102 + ], + "type": "text", + "content": "KW. Ng et al." + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 15 + }, + { + "para_blocks": [ + { + "bbox": [ + 130, + 116, + 482, + 666 + ], + "type": "list", + "angle": 0, + "index": 21, + "blocks": [ + { + "bbox": [ + 130, + 116, + 482, + 149 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 116, + 482, + 149 + ], + "spans": [ + { + "bbox": [ + 130, + 116, + 482, + 149 + ], + "type": "text", + "content": "40. Mo, S., Mu, F., Lin, K.H., Liu, Y., Guan, B., Li, Y., Zhou, B.: Freecontrol: Training-free spatial control of any text-to-image diffusion model with any condition. arXiv preprint arXiv:2312.07536 (2023)" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 130, + 150, + 481, + 172 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 150, + 481, + 172 + ], + "spans": [ + { + "bbox": [ + 130, + 150, + 481, + 172 + ], + "type": "text", + "content": "41. Mokady, R., Hertz, A., Aberman, K., Pritch, Y., Cohen-Or, D.: Null-text inversion for editing real images using guided diffusion models. In: CVPR (2023)" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 130, + 172, + 481, + 194 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 172, + 481, + 194 + ], + "spans": [ + { + "bbox": [ + 130, + 172, + 481, + 194 + ], + "type": "text", + "content": "42. Nagai, Y., Taura, T., Mukai, F.: Concept blending and dissimilarity: factors for creative concept generation process. Design studies (2009)" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 130, + 194, + 481, + 227 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 194, + 481, + 227 + ], + "spans": [ + { + "bbox": [ + 130, + 194, + 481, + 227 + ], + "type": "text", + "content": "43. Nichol, A., Dhariwal, P., Ramesh, A., Shyam, P., Mishkin, P., McGrew, B., Sutskever, I., Chen, M.: Glide: Towards photorealistic image generation and editing with text-guided diffusion models. arXiv preprint arXiv:2112.10741 (2021)" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 130, + 227, + 481, + 248 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 227, + 481, + 248 + ], + "spans": [ + { + "bbox": [ + 130, + 227, + 481, + 248 + ], + "type": "text", + "content": "44. Nobari, A.H., Rashad, M.F., Ahmed, F.: Creativegan: Editing generative adversarial networks for creative design synthesis. arXiv preprint arXiv:2103.06242 (2021)" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 130, + 249, + 481, + 304 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 249, + 481, + 304 + ], + "spans": [ + { + "bbox": [ + 130, + 249, + 481, + 304 + ], + "type": "text", + "content": "45. Oquab, M., Darcet, T., Moutakanni, T., Vo, H., Szafraniec, M., Khalidov, V., Fernandez, P., Haziza, D., Massa, F., El-Nouby, A., Assran, M., Ballas, N., Galuba, W., Howes, R., Huang, P.Y., Li, S.W., Misra, I., Rabbat, M., Sharma, V., Synnaeve, G., Xu, H., Jegou, H., Mairal, J., Labatut, P., Joulin, A., Bojanowski, P.: Dinov2: Learning robust visual features without supervision (2023)" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 130, + 304, + 481, + 326 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 304, + 481, + 326 + ], + "spans": [ + { + "bbox": [ + 130, + 304, + 481, + 326 + ], + "type": "text", + "content": "46. Qiao, T., Zhang, J., Xu, D., Tao, D.: Mirrorgan: Learning text-to-image generation by redescription. In: CVPR (2019)" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 130, + 326, + 481, + 346 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 326, + 481, + 346 + ], + "spans": [ + { + "bbox": [ + 130, + 326, + 481, + 346 + ], + "type": "text", + "content": "47. Qu, Z., Xiang, T., Song, Y.Z.: Sketchdreamer: Interactive text-augmented creative sketch ideation. In: BMVC (2023)" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 130, + 347, + 481, + 380 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 347, + 481, + 380 + ], + "spans": [ + { + "bbox": [ + 130, + 347, + 481, + 380 + ], + "type": "text", + "content": "48. Radford, A., Kim, J.W., Hallacy, C., Ramesh, A., Goh, G., Agarwal, S., Sastry, G., Askell, A., Mishkin, P., Clark, J., et al.: Learning transferable visual models from natural language supervision. In: ICML (2021)" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 130, + 380, + 481, + 413 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 380, + 481, + 413 + ], + "spans": [ + { + "bbox": [ + 130, + 380, + 481, + 413 + ], + "type": "text", + "content": "49. Ramesh, A., Dhariwal, P., Nichol, A., Chu, C., Chen, M.: Hierarchical text-conditional image generation with clip latents. arXiv preprint arXiv:2204.06125 (2022)" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 130, + 413, + 481, + 457 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 413, + 481, + 457 + ], + "spans": [ + { + "bbox": [ + 130, + 413, + 481, + 457 + ], + "type": "text", + "content": "50. Ren, T., Liu, S., Zeng, A., Lin, J., Li, K., Cao, H., Chen, J., Huang, X., Chen, Y., Yan, F., Zeng, Z., Zhang, H., Li, F., Yang, J., Li, H., Jiang, Q., Zhang, L.: Grounded sam: Assembling open-world models for diverse visual tasks. arXiv preprint arXiv:2401.14159 (2024)" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 130, + 457, + 481, + 479 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 457, + 481, + 479 + ], + "spans": [ + { + "bbox": [ + 130, + 457, + 481, + 479 + ], + "type": "text", + "content": "51. Richardson, E., Goldberg, K., Alaluf, Y., Cohen-Or, D.: Conceptlab: Creative generation using diffusion prior constraints. arXiv preprint arXiv:2308.02669 (2023)" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 130, + 479, + 481, + 501 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 479, + 481, + 501 + ], + "spans": [ + { + "bbox": [ + 130, + 479, + 481, + 501 + ], + "type": "text", + "content": "52. Rombach, R., Blattmann, A., Lorenz, D., Esser, P., Ommer, B.: High-resolution image synthesis with latent diffusion models. In: CVPR (2022)" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 130, + 501, + 481, + 533 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 501, + 481, + 533 + ], + "spans": [ + { + "bbox": [ + 130, + 501, + 481, + 533 + ], + "type": "text", + "content": "53. Ruiz, N., Li, Y., Jampani, V., Pritch, Y., Rubinstein, M., Aberman, K.: Dreambooth: Fine tuning text-to-image diffusion models for subject-driven generation. In: CVPR (2023)" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 130, + 533, + 481, + 555 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 533, + 481, + 555 + ], + "spans": [ + { + "bbox": [ + 130, + 533, + 481, + 555 + ], + "type": "text", + "content": "54. Runco, M.A., Jaeger, G.J.: The standard definition of creativity. Creativity research journal (2012)" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 130, + 555, + 481, + 600 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 555, + 481, + 600 + ], + "spans": [ + { + "bbox": [ + 130, + 555, + 481, + 600 + ], + "type": "text", + "content": "55. Saharia, C., Chan, W., Saxena, S., Li, L., Whang, J., Denton, E.L., Ghasemipour, K., Gontijo Lopes, R., Karagol Ayan, B., Salimans, T., et al.: Photorealistic text-to-image diffusion models with deep language understanding. In: NeurIPS. vol. 35 (2022)" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 130, + 600, + 481, + 621 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 600, + 481, + 621 + ], + "spans": [ + { + "bbox": [ + 130, + 600, + 481, + 621 + ], + "type": "text", + "content": "56. Sbai, O., Elhoseiny, M., Bordes, A., LeCun, Y., Couprie, C.: Design: Design inspiration from generative networks. In: ECCVW (2019)" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 130, + 621, + 481, + 644 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 621, + 481, + 644 + ], + "spans": [ + { + "bbox": [ + 130, + 621, + 481, + 644 + ], + "type": "text", + "content": "57. Sun, P., Chen, S., Luo, P.: Grounded segment anything: From objects to parts. https://github.com/Cheems-Seminar/grounded-segment-any-parts (2023)" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 130, + 644, + 481, + 666 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 644, + 481, + 666 + ], + "spans": [ + { + "bbox": [ + 130, + 644, + 481, + 666 + ], + "type": "text", + "content": "58. Sun, P., Chen, S., Zhu, C., Xiao, F., Luo, P., Xie, S., Yan, Z.: Going denser with open-vocabulary part segmentation. arXiv preprint arXiv:2305.11173 (2023)" + } + ] + } + ], + "index": 20 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 258, + 91, + 447, + 102 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 258, + 91, + 447, + 102 + ], + "spans": [ + { + "bbox": [ + 258, + 91, + 447, + 102 + ], + "type": "text", + "content": "PartCraft: Crafting Creative Objects by Parts" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 470, + 91, + 480, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 470, + 91, + 480, + 100 + ], + "spans": [ + { + "bbox": [ + 470, + 91, + 480, + 100 + ], + "type": "text", + "content": "17" + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 16 + }, + { + "para_blocks": [ + { + "bbox": [ + 132, + 116, + 480, + 456 + ], + "type": "list", + "angle": 0, + "index": 16, + "blocks": [ + { + "bbox": [ + 132, + 116, + 480, + 138 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 116, + 480, + 138 + ], + "spans": [ + { + "bbox": [ + 132, + 116, + 480, + 138 + ], + "type": "text", + "content": "59. Tao, M., Tang, H., Wu, F., Jing, X.Y., Bao, B.K., Xu, C.: Df-gan: A simple and effective baseline for text-to-image synthesis. In: CVPR (2022)" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 132, + 139, + 480, + 160 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 139, + 480, + 160 + ], + "spans": [ + { + "bbox": [ + 132, + 139, + 480, + 160 + ], + "type": "text", + "content": "60. Vinker, Y., Voynov, A., Cohen-Or, D., Shamir, A.: Concept decomposition for visual exploration and inspiration. In: SIGGRAPH Asia (2023)" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 132, + 161, + 480, + 182 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 161, + 480, + 182 + ], + "spans": [ + { + "bbox": [ + 132, + 161, + 480, + 182 + ], + "type": "text", + "content": "61. Voynov, A., Chu, Q., Cohen-Or, D., Aberman, K.: P+: Extended textual conditioning in text-to-image generation. arXiv preprint arXiv:2303.09522 (2023)" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 132, + 183, + 480, + 204 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 183, + 480, + 204 + ], + "spans": [ + { + "bbox": [ + 132, + 183, + 480, + 204 + ], + "type": "text", + "content": "62. Wah, C., Branson, S., Welinder, P., Perona, P., Belongie, S.: The Caltech-UCSD Birds-200-2011 Dataset (2011)" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 132, + 205, + 480, + 226 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 205, + 480, + 226 + ], + "spans": [ + { + "bbox": [ + 132, + 205, + 480, + 226 + ], + "type": "text", + "content": "63. Wang, X., Darrell, T., Rambhatla, S.S., Girdhar, R., Misra, I.: Instancediffusion: Instance-level control for image generation. arXiv preprint arXiv:2402.03290 (2024)" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 132, + 227, + 480, + 258 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 227, + 480, + 258 + ], + "spans": [ + { + "bbox": [ + 132, + 227, + 480, + 258 + ], + "type": "text", + "content": "64. Wei, Y., Zhang, Y., Ji, Z., Bai, J., Zhang, L., Zuo, W.: Elite: Encoding visual concepts into textual embeddings for customized text-to-image generation. In: ICCV (2023)" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 132, + 259, + 480, + 281 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 259, + 480, + 281 + ], + "spans": [ + { + "bbox": [ + 132, + 259, + 480, + 281 + ], + "type": "text", + "content": "65. Wilkenfeld, M.J., Ward, T.B.: Similarity and emergence in conceptual combination. Journal of Memory and Language (2001)" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 132, + 281, + 480, + 303 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 281, + 480, + 303 + ], + "spans": [ + { + "bbox": [ + 132, + 281, + 480, + 303 + ], + "type": "text", + "content": "66. Xie, J., Li, Y., Huang, Y., Liu, H., Zhang, W., Zheng, Y., Shou, M.Z.: Boxdiff: Text-to-image synthesis with training-free box-constrained diffusion. In: ICCV (2023)" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 132, + 304, + 480, + 335 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 304, + 480, + 335 + ], + "spans": [ + { + "bbox": [ + 132, + 304, + 480, + 335 + ], + "type": "text", + "content": "67. Xu, T., Zhang, P., Huang, Q., Zhang, H., Gan, Z., Huang, X., He, X.: Attngan: Fine-grained text to image generation with attentional generative adversarial networks. In: CVPR (2018)" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 132, + 336, + 480, + 369 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 336, + 480, + 369 + ], + "spans": [ + { + "bbox": [ + 132, + 336, + 480, + 369 + ], + "type": "text", + "content": "68. Yang, B., Gu, S., Zhang, B., Zhang, T., Chen, X., Sun, X., Chen, D., Wen, F.: Paint by example: Exemplar-based image editing with diffusion models. In: CVPR (2023)" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 132, + 369, + 480, + 391 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 369, + 480, + 391 + ], + "spans": [ + { + "bbox": [ + 132, + 369, + 480, + 391 + ], + "type": "text", + "content": "69. Yang, Z., Wang, J., Gan, Z., Li, L., Lin, K., Wu, C., Duan, N., Liu, Z., Liu, C., Zeng, M., et al.: Reco: Region-controlled text-to-image generation. In: CVPR (2023)" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 132, + 392, + 480, + 413 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 392, + 480, + 413 + ], + "spans": [ + { + "bbox": [ + 132, + 392, + 480, + 413 + ], + "type": "text", + "content": "70. Yin, G., Liu, B., Sheng, L., Yu, N., Wang, X., Shao, J.: Semantics disentangling for text-to-image generation. In: CVPR (2019)" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 132, + 413, + 480, + 434 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 413, + 480, + 434 + ], + "spans": [ + { + "bbox": [ + 132, + 413, + 480, + 434 + ], + "type": "text", + "content": "71. Zhang, L., Rao, A., Agrawala, M.: Adding conditional control to text-to-image diffusion models. In: ICCV (2023)" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 132, + 435, + 480, + 456 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 435, + 480, + 456 + ], + "spans": [ + { + "bbox": [ + 132, + 435, + 480, + 456 + ], + "type": "text", + "content": "72. Zhu, M., Pan, P., Chen, W., Yang, Y.: Dm-gan: Dynamic memory generative adversarial networks for text-to-image synthesis. In: CVPR (2019)" + } + ] + } + ], + "index": 15 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "spans": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "type": "text", + "content": "18" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 166, + 91, + 225, + 102 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 166, + 91, + 225, + 102 + ], + "spans": [ + { + "bbox": [ + 166, + 91, + 225, + 102 + ], + "type": "text", + "content": "KW. Ng et al." + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 17 + } + ], + "_backend": "vlm", + "_version_name": "2.6.4" +} \ No newline at end of file diff --git a/2024/PartGLEE_ A Foundation Model for Recognizing and Parsing Any Objects/a546ffc1-52e4-4141-be35-2478ed6c5802_content_list.json b/2024/PartGLEE_ A Foundation Model for Recognizing and Parsing Any Objects/a546ffc1-52e4-4141-be35-2478ed6c5802_content_list.json new file mode 100644 index 0000000000000000000000000000000000000000..5e581141095983a045d9482be2f85b1334d80085 --- /dev/null +++ b/2024/PartGLEE_ A Foundation Model for Recognizing and Parsing Any Objects/a546ffc1-52e4-4141-be35-2478ed6c5802_content_list.json @@ -0,0 +1,1572 @@ +[ + { + "type": "text", + "text": "PartGLEE: A Foundation Model for Recognizing and Parsing Any Objects", + "text_level": 1, + "bbox": [ + 222, + 140, + 779, + 186 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Junyi Li\\*, Junfeng Wu\\*, Weizhi Zhao\\*, Song Bai\\*, and Xiang Bai", + "bbox": [ + 250, + 210, + 750, + 228 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "1 Huazhong University of Science and Technology \n2 ByteDance Inc.", + "bbox": [ + 333, + 239, + 666, + 267 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Abstract. We present PartGLEE, a part-level foundation model for locating and identifying both objects and parts in images. Through a unified framework, PartGLEE accomplishes detection, segmentation, and grounding of instances at any granularity in the open world scenario. Specifically, we propose a Q-Former to construct the hierarchical relationship between objects and parts, parsing every object into corresponding semantic parts. By incorporating a large amount of object-level data, the hierarchical relationships can be extended, enabling PartGLEE to recognize a rich variety of parts. We conduct comprehensive studies to validate the effectiveness of our method, PartGLEE achieves the state-of-the-art performance across various part-level tasks and obtain competitive results on object-level tasks. The proposed PartGLEE significantly enhances hierarchical modeling capabilities and part-level perception over our previous GLEE model. Further analysis indicates that the hierarchical cognitive ability of PartGLEE is able to facilitate a detailed comprehension in images for mLLMs. The model and code will be released at https://provencstar.github.io/PartGLEE-Vision/.", + "bbox": [ + 261, + 309, + 738, + 544 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Keywords: Foundation Model $\\cdot$ Hierarchical Recognition $\\cdot$ Part Segmentation", + "bbox": [ + 261, + 559, + 738, + 585 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "1 Introduction", + "text_level": 1, + "bbox": [ + 215, + 614, + 375, + 630 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "In recent years, foundation models have dominated the majority of tasks in the fields of Natural Language Processing [3, 9, 54] and Computer Vision [19, 23, 53, 60, 71, 72]. CLIP family [12, 13, 21, 53, 81] have made significant advancements in transfer learning and have demonstrated impressive zero-shot capabilities on vision-language tasks. SAM [23] has revolutionized the development of segmentation tasks and is able to provide multi-level class-agnostic masks. GLEE [71] utilized diverse object-level data to develop general object representations, enabling detection, segmentation, tracking, grounding, and identification of objects in open-world scenarios. Their remarkable achievement can be attributed to the integration of extensive and diverse range of datasets.", + "bbox": [ + 212, + 648, + 787, + 800 + ], + "page_idx": 0 + }, + { + "type": "page_footnote", + "text": "* Equal Technical Contribution. Work done during Junfeng's internship at ByteDance. †Correspondence to Xiang Bai .", + "bbox": [ + 217, + 810, + 785, + 839 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/47f7c330d14fe208579b04eeb8af74ec4ab188c2b13614d4336fbf924c3b0dda.jpg", + "image_caption": [ + "Fig. 1: An illustrative example demonstrating image annotations at diverse granularities across multiple datasets. The annotations at hierarchical levels with corresponding relationships are depicted on the right side. Below is a visualization of our segmentation results at multiple granularities." + ], + "image_footnote": [], + "bbox": [ + 228, + 147, + 779, + 518 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Different from the vast quantity of object-level data, the scale of part-level data is relatively small, which turns out to be a major bottleneck hindering vision models from recognizing part-level instances. Thus, most vision models lack the hierarchical comprehension between objects and parts. However, it is evident that the ability to recognize parts from objects is essential for various practical applications such as image editing [22,33,38], behavior analysis [50,75], pose estimation [10,76], robotics manipulation [2,49], etc. Moreover, we observe that part-level information is able to help multi-modal Large Language Models (mLLMs) in achieving a more detailed understanding of image content. Since part-level comprehension is a critical ability for foundation models to tackle a broader range of problems, it leads to a natural question: How could we break through data limitations to build a part-level vision foundation model?", + "bbox": [ + 212, + 625, + 787, + 805 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "To enable object foundation model with part-level cognitive ability, we emphasize that the model should achieve two key objectives: (1) Hierarchical", + "bbox": [ + 212, + 809, + 785, + 840 + ], + "page_idx": 1 + }, + { + "type": "page_number", + "text": "2", + "bbox": [ + 217, + 114, + 228, + 126 + ], + "page_idx": 1 + }, + { + "type": "header", + "text": "J. Li, J. Wu, et al.", + "bbox": [ + 271, + 114, + 395, + 127 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Comprehension, the model is supposed to understand the intrinsic relationship between objects and parts, and extend this hierarchical connection to any novel object, (2) Semantic Granularity, the model should be capable of learning a universal feature representation, enabling it to recognize semantic instances at any granularity. Consequently, we present a method to jointly detect and segment both objects and parts in a top-down manner. A lightweight Querying Transformer (Q-Former) is proposed to construct the hierarchical relationship between objects and parts. Specifically, it employs a set of universal parsing queries to interact with object queries, consequently generating multiple part-level queries that are capable of predicting corresponding semantic parts for each object. The Q-Former acts as a decomposer, which first recognizes individual objects in the images and subsequently parsing them into parts. Such model design is built upon the observation that various common objects often exhibit shared characteristics of parts. For example, cats, dogs, and dinosaurs all have parts such as torso, legs, and tails. In this way, two sets of query embeddings at different levels are generated, which are then used to predict object-level and part-level instances respectively. Through this approach, the relationship between objects and parts is established via the Q-Former design. Meanwhile, the hierarchical levels of objects and parts are distinguished, which is different from previous research [6,55,62,68,70] that consider parts as fine-grained objects. This paradigm enables vision models to better understand the features on different levels during training, thereby achieving improved performance.", + "bbox": [ + 217, + 146, + 785, + 478 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Our complete solution, PartGLEE, for jointly detecting and segmenting instances at both object and part levels, makes it possible for vision models to achieve favorable outcomes on both object and part levels. Some previous research have devised specialized training paradigms to utilize abundant image-text pair data [79, 89] as well as grounding data [30, 39, 71], thereby enhancing the cognitive and generalization capabilities of the models. On the contrary, the quantity of part-level data is much smaller compared to object-level data. So far, the largest dataset incorporating the concepts of both objects and parts is the recently proposed PACO [55] dataset. The scarcity of data has limited research on part-level recognition and restricted the generalization improvement of vision models. Although VLPart [62] has attempted to utilize pseudo-labeling schemes to generate part-level annotations for both object-level and image-level datasets, the quality of the pseudo-labels is relatively poor. Our innovative algorithm that parsing objects into their corresponding parts facilitates the transfer of generalization capability from objects to parts. Consequently, parts are generated from objects, which enables vision models to maintain generalization performance when predicting parts for novel objects without labeling extensive part-level data. To facilitate the training process of Q-Former, we standardize the annotation granularity across various part-level datasets and introduce a vast amount of object-level datasets, an intuitive display of the overall training data is shown in Fig. 1. Unlike VLPart, which exhibits unsatisfactory performance at object-level datasets after joint-training, our method demonstrates favorable outcomes at both object and part levels after joint-training. Moreover, it turns out that", + "bbox": [ + 217, + 493, + 785, + 838 + ], + "page_idx": 2 + }, + { + "type": "header", + "text": "PartGLEE", + "bbox": [ + 656, + 114, + 730, + 126 + ], + "page_idx": 2 + }, + { + "type": "page_number", + "text": "3", + "bbox": [ + 774, + 116, + 784, + 126 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "using object-level datasets is able to improve the performance of the model on part-level tasks, indicating a beneficial interaction between objects and parts.", + "bbox": [ + 212, + 146, + 782, + 176 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Extensive experiments demonstrate that our method significantly improve the open-vocabulary part segmentation performance, concurrently ensuring a decent performance on object detection and segmentation. We verify its effectiveness on various popular datasets. To validate the generalization performance of our model in identifying various parts of novel objects, we conduct experiments on PartImageNet [18] and Pascal Part [6] datasets in cross-dataset and cross-category manners respectively. Our method exhibits strong transferability and generalization ability when adding extra object-level datasets during training. To evaluate the decomposition capability of our model, we conduct experiments on both ADE20K-Part and Pascal Part datasets follow OV-PARTS [70]. As a result, our approach significantly outperforms one-stage baselines of OV-PARTS, with an increase of $8.16\\%$ and $2.07\\%$ on harmonic mean IoU (hIoU) in ADE20K-Part-234 and Pascal-Part-116 respectively. Additionally, by incorporating a large amount of object-level data for joint-training, our method establishes generic hierarchical relationships and breaks through the limitations of scarce part-level data, achieving state-of-the-art performance across various part-level tasks.", + "bbox": [ + 212, + 176, + 787, + 417 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "In conclusion, our main contributions can be summarized as follows:", + "bbox": [ + 238, + 417, + 732, + 431 + ], + "page_idx": 3 + }, + { + "type": "list", + "sub_type": "text", + "list_items": [ + "1. We construct the hierarchical relationship between objects and parts via the Q-Former, facilitating part segmentation to acquire advantages from various object-level datasets.", + "2. We propose a unified pipeline for hierarchical detection and segmentation, where we first recognize objects and then parsing them into corresponding semantic parts. This algorithm enables us to jointly detect and segment both object-level and part-level instances.", + "3. We standardize the annotation granularity across various part-level datasets by incorporating corresponding object-level annotations, complementing the hierarchical correspondences for current part-level datasets, promoting the development of vision foundation models." + ], + "bbox": [ + 220, + 443, + 782, + 604 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "2 Related Work", + "text_level": 1, + "bbox": [ + 215, + 632, + 387, + 648 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "2.1 Visual Foundation Models and General Models", + "text_level": 1, + "bbox": [ + 215, + 662, + 651, + 679 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Vision foundation models and generalist models are considered as a milestone in the development of the intelligent vision system. For instance, multi-modal visual foundation models [1, 21, 53, 67, 81] have significantly advanced efficient transfer learning and exhibit impressive zero-shot capabilities on vision-language tasks by using contrastive learning with large-scale image-text pairs. Generative foundation models [11, 56, 57, 60] are trained on vast collections of images and captions, empowering them to generate image content conditioned on textual prompts. Self-supervised foundation models [4, 12, 13, 19] have learned general visual representations from large-scale image datasets, enhancing their ability to transfer to downstream tasks. However, the image-level features learned by these", + "bbox": [ + 212, + 688, + 787, + 840 + ], + "page_idx": 3 + }, + { + "type": "page_number", + "text": "4", + "bbox": [ + 217, + 114, + 228, + 126 + ], + "page_idx": 3 + }, + { + "type": "header", + "text": "J. Li, J. Wu, et al.", + "bbox": [ + 271, + 114, + 395, + 127 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "foundation models are not well-suited for direct application to dense prediction tasks that involve precise object and part localization.", + "bbox": [ + 212, + 146, + 782, + 176 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Transformer-based generalist methods [5,41,66,77,87] adopt a sequence generation pipeline to unify the output of text and spatial coordinates. However, they mainly focus on image-level comprehension, which results in relatively weak localization capabilities. Works such as UNINEXT, etc. [28, 73, 74], built upon strong detectors [27, 86], demonstrating a strong localization capability across multiple datasets. But they fail to exhibit zero-shot transfer ability and generalization capability due to their closed-set training paradigm. Some works about open-vocabulary detection (OVD) [30, 34, 35, 42, 43, 78, 82, 83] have explored zero-shot generalization capabilities on novel categories. X-Decoder [89] and SEEM [90] have developed a versatile decoding architecture that are able to generate accurate pixel-level segmentation predictions. GLEE [71] addresses various object-level tasks through a unified architecture and training paradigm. However, current generalist models and foundation models are trained mainly on image-level and object-level datasets, thus their ability to extract more fine-grained information is limited, making it difficult for them to recognize corresponding semantic parts of any object. Our work focuses on empowering hierarchical cognitive capability for vision foundation models, thereby further advancing the development of comprehensive visual systems.", + "bbox": [ + 212, + 176, + 787, + 449 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "2.2 Part Segmentation", + "text_level": 1, + "bbox": [ + 214, + 468, + 418, + 484 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "The growing interest in achieving a more fine-grained understanding of objects has sparked a surge in research focused on part level recognition. Some pioneering studies have introduced datasets with part-level annotations, concentrating on objects of some specific categories such as human body parts [15, 29, 75], animal body parts [64] and vehicle components [58]. More general part annotations for common objects such as Pascal-Part [6], PartImageNet [18], ADE20K [84], CityscapesPanoptic-Parts [45] and more recent PACO [55] were then proposed to promote more in-depth research in the field of parts. Most of the previous works [14, 31, 46, 85] were conducted based on a closed-set configuration, thus only capable of detecting and segmenting closed-set objects and parts. Recently, VLPart [62] present a pipeline for detecting and segmenting both open-vocabulary objects and their corresponding part regions, while OVPARTS [70] utilize adapters to transfer the generalization abilities of CLIP into open-vocabulary part segmentation task. However, due to the limited quantity of data, the generalization capability of previous models [15, 29, 51, 62, 63, 68, 70, 75] still relies heavily on the training datasets. Furthermore, in prior works, both objects and parts are treated equally, they consider part as a special type of object. On the contrary, we distinguish them by considering parts as integral components subordinate to objects and generate parts from corresponding objects in a top-down manner. Our work is aimed at building hierarchical relationships while unifying the training paradigm for object and part-level data. By incorporating a large amount of object-level data, the hierarchical relationships can be extended to any object, enabling our method to recognize a rich variety of parts.", + "bbox": [ + 212, + 492, + 787, + 843 + ], + "page_idx": 4 + }, + { + "type": "header", + "text": "PartGLEE", + "bbox": [ + 656, + 114, + 730, + 126 + ], + "page_idx": 4 + }, + { + "type": "page_number", + "text": "5", + "bbox": [ + 774, + 116, + 785, + 126 + ], + "page_idx": 4 + }, + { + "type": "image", + "img_path": "images/d2e25db7a14cb03ec2bc08abfeda2e0926d6c4b89e67d93ac90204d9c787ef5a.jpg", + "image_caption": [ + "Fig. 2: Framework of PartGLEE. The Q-Former takes each object query as input and output the corresponding part queries. These queries are then fed into the object decoder and the part decoder respectively to generate hierarchical predictions." + ], + "image_footnote": [], + "bbox": [ + 250, + 143, + 759, + 340 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "2.3 Hierarchical Learning of Objects and Parts", + "text_level": 1, + "bbox": [ + 214, + 424, + 617, + 439 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Learning objects through parts has been a long-standing research topic as part annotations provide more detailed semantic information of objects. Morabia et al. [48] first introduced a pipeline employing an attention mechanism for simultaneous detection of both objects and parts. Deepflux [69] designed an image context flux representation which enables better object parts interaction for skeleton detection. Leopart [88] demonstrated that learning object parts can provide spatially diverse representation which facilitates self-supervised semantic segmentation. Wang et al. [65] proposed a method to predict both parts and objects simultaneously on Pascal-Part dataset [6]. Recent studies such as SAM [23] and Semantic-SAM [26] have studied on class-agnostic multi-granularity interactive segmentation task. However, they have not explored the relationship between objects and their corresponding semantic parts. Recently, Compositor [17] designed a bottom-up pipeline to predict parts and then cluster them into objects, while AIMS [52] utilized an independent relation decoder to construct the hierarchical association between objects and parts. Different from these works, our approach introduces a Querying Transformer to model the hierarchical relationship, allowing our model to parse any object into its corresponding parts.", + "bbox": [ + 212, + 446, + 787, + 705 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "3 Method", + "text_level": 1, + "bbox": [ + 214, + 726, + 330, + 739 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "3.1 Overall Framework", + "text_level": 1, + "bbox": [ + 214, + 755, + 419, + 768 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Following [68,71], we propose PartGLEE, which comprises of an image encoder, a Q-Former, two independent decoders and a text encoder, as shown in Fig. 2.", + "bbox": [ + 212, + 779, + 782, + 808 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Given an input image $I \\in \\mathcal{R}^{H \\times W \\times 3}$ , the backbone and the pixel decoder first extract multi-scale image features $F_s \\in \\mathcal{R}_{\\frac{H}{2^S} \\times \\frac{W}{2^S} \\times C}$ and $s = \\{2, 3, 4, 5\\}$ with", + "bbox": [ + 212, + 809, + 784, + 839 + ], + "page_idx": 5 + }, + { + "type": "page_number", + "text": "6", + "bbox": [ + 217, + 114, + 228, + 126 + ], + "page_idx": 5 + }, + { + "type": "header", + "text": "J. Li, J. Wu, et al.", + "bbox": [ + 271, + 114, + 395, + 127 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "backbones such as ResNet [20] or Swin Transformer [40]. Then we feed them into the object decoder, where the object-level query embeddings $q_{obj} \\in \\mathcal{R}^{N \\times C}$ are generated in a two-stage process. These object queries are utilized to perform object-level classification, detection as well as segmentation tasks through three independent prediction heads. Besides, the object queries $q_{obj}$ are fed into the Q-Former simultaneously, where $L$ learnable universal parsing queries are initialized to interact with object queries. It takes object queries as input and generate part-level queries $q_{part} \\in \\mathcal{R}^{N \\cdot L \\times C}$ which are then passed into the part decoder to yield part-level predictions (detailed in Sec. 3.2). To enhance the semantic-awareness, an early fusion module is adopted before Transformer encoder following [74], which takes image feature from backbone and text embedding as input and perform bi-directional cross-attention between them. In line with previous segmentation models [7,27,32], a pixel embedding map $M_p \\in \\mathcal{R}^{\\frac{H}{4} \\times \\frac{W}{4} \\times C}$ at $1/4$ resolution is constructed by upsampling and integrating multi-scale feature maps from the backbone and the pixel decoder. Eventually, we dot product each object query or part query with the pixel embedding map to derive an output mask $m \\in \\mathcal{R}^{\\frac{H}{4} \\times \\frac{W}{4}}$ :", + "bbox": [ + 212, + 145, + 787, + 402 + ], + "page_idx": 6 + }, + { + "type": "equation", + "text": "\n$$\nm = F F N (q _ {l}) \\otimes M _ {p}, \\quad l \\in \\{o b j, p a r t \\} \\tag {1}\n$$\n", + "text_format": "latex", + "bbox": [ + 366, + 412, + 784, + 426 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "where FFN is comprised of 3 layers feed forward network with ReLU activation functions and linear layers.", + "bbox": [ + 212, + 434, + 785, + 465 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "3.2 Parsing Objects into Parts", + "text_level": 1, + "bbox": [ + 214, + 484, + 482, + 500 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "We propose a Q-Former to establish the hierarchical relationship between objects and parts. As Various common objects tend to manifest shared attributes in their constituent parts, for example, both lizards and birds exhibit similar components, such as heads and torsos. Thus, we initialize a set of query embeddings in the Q-Former to parse any object into semantic parts. We denote these universal parsing query embeddings as $q_{parse} \\in \\mathcal{R}^{L \\times C}$ , where $L$ represents the number of the parsing queries. As shown in Fig. 2, the Q-Former is comprised of $M$ cascaded attention modules, each module includes a self-attention layer, a cross-attention layer, and a feed forward network. The universal parsing queries are first fed into the self-attention layer and then perform cross-attention with the object queries. Note that every object query is interacted with all universal parsing queries. Hence, assume $N$ object queries $(q_{obj} \\in \\mathcal{R}^{N \\times C})$ are generated from the object decoder, and $L$ universal parsing queries $q_{parse} \\in \\mathcal{R}^{L \\times C}$ are initialized in the Q-Former, we obtain $N \\cdot L$ part-level queries which can be denoted as $q_{part} \\in \\mathcal{R}^{N \\cdot L \\times C}$ . We refer to this process as:", + "bbox": [ + 212, + 506, + 787, + 734 + ], + "page_idx": 6 + }, + { + "type": "equation", + "text": "\n$$\nq _ {p a r t} = Q - F o r m e r \\left(q _ {p a r s e}; q _ {o b j}\\right) \\tag {2}\n$$\n", + "text_format": "latex", + "bbox": [ + 392, + 742, + 784, + 757 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Our proposed Q-Former functions as a decomposer, extracting and representing parts from object queries. Hence, by training jointly on extensive object-level datasets and limited hierarchical datasets which contain object-part correspondences, our Q-Former obtains strong generalization ability to parse any novel object into its corresponding parts.", + "bbox": [ + 212, + 763, + 787, + 840 + ], + "page_idx": 6 + }, + { + "type": "header", + "text": "PartGLEE", + "bbox": [ + 656, + 114, + 730, + 126 + ], + "page_idx": 6 + }, + { + "type": "page_number", + "text": "7", + "bbox": [ + 774, + 116, + 785, + 126 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/9a2fdac931e61e57734068d20d91b4215414125b93fa7a763beb1991e846600d.jpg", + "image_caption": [ + "Fig. 3: Matching mechanisms of PartGLEE. Two separate forward passes are performed on the same image to obtain hierarchical segmentation results." + ], + "image_footnote": [], + "bbox": [ + 222, + 147, + 777, + 311 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "3.3 Unified Training Paradigm for Objects and Parts", + "text_level": 1, + "bbox": [ + 214, + 377, + 668, + 393 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Since the Q-Former requires hierarchical data to learn how to parse objects into parts, we enrich part-level data with corresponding object-level annotations, details of which are provided in the supplementary materials.", + "bbox": [ + 212, + 398, + 782, + 444 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Since the annotation granularity across part-level datasets is standardized, our model can first learn the characteristics of objects and then acquire the ability to parse any object into its semantic parts. To facilitate open-vocabulary detection and segmentation, we substitute the similarity scores between the instance embeddings and the text embeddings for the original class head. Given $K$ object-level and part-level input categories as separate sentences, we feed them into the text encoder and utilize the average of each individual sentence tokens as the output text embedding $T_{l}$ for each category. Then the similarity scores $S_{l} \\in \\mathcal{R}^{N \\times K}$ are calculated through a dot product operation between the hierarchical instance embeddings $q_{l} \\in \\mathcal{R}^{N \\times C}$ from detector and the text embeddings $T_{l} \\in \\mathcal{R}^{K \\times D}$ from text encoder, which can be denoted as:", + "bbox": [ + 212, + 444, + 787, + 609 + ], + "page_idx": 7 + }, + { + "type": "equation", + "text": "\n$$\nS _ {l} = q _ {l} \\cdot W _ {p r o j} \\otimes T _ {l}, \\quad l \\in \\{o b j, p a r t \\} \\tag {3}\n$$\n", + "text_format": "latex", + "bbox": [ + 366, + 617, + 784, + 633 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "where $W_{proj} \\in \\mathcal{R}^{C \\times D}$ is a trainable projection weight for fine-tuning text embedding space especially for part-level descriptions. Following [26, 68], we perform Hungarian matching of objects and parts individually, where object-level predictions are only matched with object-level targets, and the same applies to the part-level output, as shown in Fig. 3.", + "bbox": [ + 212, + 640, + 787, + 715 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "We then introduce a constraint loss to ensure the part-level predictions to be the component of the objects. We denote this novel loss function as restriction loss $L_{res}$ . Due to memory limitations, we only calculate our restriction loss on the predicted bounding boxes between different levels, while leaving the predicted masks unconstrained. Our restriction loss can be calculated as follow:", + "bbox": [ + 212, + 717, + 787, + 792 + ], + "page_idx": 7 + }, + { + "type": "equation", + "text": "\n$$\nL _ {r e s} = \\sum_ {i} ^ {L} \\left(1 - \\frac {\\left| S _ {o b j} \\cap S _ {p a r t} ^ {i} \\right|}{S _ {p a r t} ^ {i}}\\right) \\tag {4}\n$$\n", + "text_format": "latex", + "bbox": [ + 390, + 795, + 785, + 837 + ], + "page_idx": 7 + }, + { + "type": "page_number", + "text": "8", + "bbox": [ + 217, + 114, + 228, + 126 + ], + "page_idx": 7 + }, + { + "type": "header", + "text": "J. Li, J. Wu, et al.", + "bbox": [ + 271, + 114, + 395, + 127 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "where $S_{obj}$ represents the area of the object-level bounding box prediction, and $S_{part}^{i}$ stands for the area of the $i - th$ part-level bounding box prediction. Note that each object query can generate $L$ part queries through Q-Former. This loss function is only applied to the matched predictions in part-level datasets, thereby strengthening the mutual correspondence between different hierarchies.", + "bbox": [ + 212, + 146, + 782, + 220 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "PartGLEE is trained with a linear combination of losses for object-level tasks and part-level tasks, which can be formulated as:", + "bbox": [ + 212, + 222, + 782, + 251 + ], + "page_idx": 8 + }, + { + "type": "equation", + "text": "\n$$\nL = \\lambda_ {1} (L _ {c l s} ^ {o b j} + L _ {c l s} ^ {p a r t}) + \\lambda_ {2} (L _ {b o x} ^ {o b j} + L _ {b o x} ^ {p a r t}) + \\lambda_ {3} (L _ {m a s k} ^ {o b j} + L _ {m a s k} ^ {p a r t}) + \\lambda_ {4} L _ {r e s} \\quad (5)\n$$\n", + "text_format": "latex", + "bbox": [ + 235, + 262, + 785, + 282 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "where $L_{cls}^{l}, L_{box}^{l}, L_{mask}^{l}$ are the classification, box, and mask loss at different levels ( $l \\in \\{obj, part\\}$ ), while $L_{res}$ is the restriction loss, and $\\lambda$ are their corresponding weights. We apply Focal Loss [36] as the classification loss on the similarity scores $S_{l}$ to align the text concepts with instance features. A combination of L1 loss and generalized IoU loss [59] is utilized for box predictions, while we employ both Dice Loss [47] and Focal Loss to calculate mask loss. We follow MaskDINO to set our hyperparameters to $\\lambda_{1} = 4, \\lambda_{2} = 2, \\lambda_{3} = 5, \\lambda_{4} = 5$ . Based on the above designs, PartGLEE is able to leverage both object-level data and part-level data thus obtaining a strong generalization capability.", + "bbox": [ + 212, + 292, + 784, + 429 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "4 Experiments", + "text_level": 1, + "bbox": [ + 215, + 450, + 375, + 468 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "4.1 Experimental Setup", + "text_level": 1, + "bbox": [ + 215, + 483, + 426, + 498 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "We conduct comprehensive experiments to exhibit the effectiveness of PartGLEE across a wide range of object-level and part-level tasks.", + "bbox": [ + 212, + 507, + 782, + 536 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "Data Unification. We utilize object-level datasets such as COCO [37], LVIS [16], Object365 [61], OpenImages [25], Visual Genome [24] and RefCOCO series [44,80], etc, while using part-level datasets PACO [55], PartImageNet [18], Pascal Part [6], ADE20K-Part [70] and SA-1B [23] with varying annotation granularity for joint-training. For Visual Genome and SA-1B, we categorize their corresponding part-level annotations based on semantic and mask overlap relationships to construct hierarchical data versions. For part-level data, we integrate the original part-level annotations with corresponding object-level annotations according to their associated object-level dataset. The details of these dataset preprocessing steps are left in the supplementary materials.", + "bbox": [ + 212, + 537, + 784, + 686 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "Implementation Details. In our experiments, we utilize ResNet-50 [20] and Swin-Large [40] as the vision encoder. Following MaskDINO [27], we adopt deformable transformer in the decoder, and use 300 object queries while setting the number of parsing queries $L$ to be 10. The $M$ of Q-Former is set to 6. We select the top 50 object queries based on the similarity scores and input them into the Q-Former, ultimately yielding 500 part queries. We use both query denoising and hybrid matching strategies to facilitate convergence and enhance performance. We conduct experiments on part-level datasets following the methodologies of VLPart [62] and OV-PARTS [70] in order to evaluate the generalization performance and the ability to parse novel objects of our model. For joint-training, we", + "bbox": [ + 212, + 688, + 785, + 839 + ], + "page_idx": 8 + }, + { + "type": "header", + "text": "PartGLEE", + "bbox": [ + 656, + 114, + 730, + 126 + ], + "page_idx": 8 + }, + { + "type": "page_number", + "text": "9", + "bbox": [ + 774, + 116, + 785, + 126 + ], + "page_idx": 8 + }, + { + "type": "table", + "img_path": "images/7e186bea69deed2a6ad4786fc2254ecb0a7086b8e91bca07792f25fc6652fcd3.jpg", + "table_caption": [ + "Table 1: Cross-dataset generalization performance compared with VLPart. The evaluation metric is $mAP_{mask}$ on the validation set of PartImageNet. All models utilize ResNet-50 as backbone and use the text embeddings of the category names as the classifier. PartImageNet denotes the fully-supervised method reported for comparison." + ], + "table_footnote": [], + "table_body": "
MethodDatasetsAll (40)quadruped
headbodyfoottail
VLPart [62]Pascal Part4.517.40.10.02.9
+ IN-S11 label5.423.63.40.81.2
+ Parsing IN-S117.835.015.23.58.9
vs. baseline+3.3+17.6+15.1+3.5+6.0
PartImageNet29.757.325.822.922.9
PartGLEEPascal Part9.923.64.51.34.6
+ Parsing IN-S1114.955.327.27.023.6
vs. baseline+5.0+31.7+22.7+5.7+19.0
PartImageNet40.267.037.636.540.7
", + "bbox": [ + 305, + 224, + 702, + 375 + ], + "page_idx": 9 + }, + { + "type": "text", + "text": "train PartGLEE based on the weights of GLEE [71], continuing training on 32 A100 GPUs. The settings for the part-level zero-shot experiments are described separately in each section.", + "bbox": [ + 215, + 417, + 784, + 463 + ], + "page_idx": 9 + }, + { + "type": "text", + "text": "4.2 Zero-Shot Part Segmentation Results", + "text_level": 1, + "bbox": [ + 217, + 489, + 568, + 506 + ], + "page_idx": 9 + }, + { + "type": "text", + "text": "1) Cross-dataset Part Segmentation on PartImageNet. We follow VLPart [62] to conduct experiments on cross-dataset generalization performance by directly evaluating on PartImageNet [18] validation set. We report the metrics of all (40) part categories and the detailed metrics of quadruped are also provided. The baseline approach only utilize Pascal Part as the training set and directly perform evaluation on PartImageNet in a zero-shot manner. Note that IN-S11 label represents adding image-level classification data for training in order to improve performance. Meanwhile, Parsing IN-S11 stands for training with the pseudo-labels generated from the parsing pipeline proposed by VLPart. However, both of these methods expose the model to categories and images from the PartImageNet dataset. We first perform our training process exclusively on the Pascal Part dataset to verify our zero-shot capabilities, and then we incorporate pseudo-labels to assess the ability of our model to utilize low-quality annotations.", + "bbox": [ + 215, + 521, + 785, + 717 + ], + "page_idx": 9 + }, + { + "type": "text", + "text": "Given that Pascal Part does not provide semantic labels for categories like quadruped in PartImageNet, the model needs to generalize from annotated parts of dog, cat, etc. in Pascal Part to parts of quadruped in PartImageNet. As shown in Tab. 1, our model significantly outperforms VLPart when only training on Pascal Part, even surpassing the model trained with Parsing IN-S11. After incorporating pseudo-labeled data into training, our model shows higher performance gains, indicating better utilization of low-quality data. This result illustrates the importance of hierarchical modeling, which enables our model to recognize and", + "bbox": [ + 215, + 719, + 785, + 839 + ], + "page_idx": 9 + }, + { + "type": "page_number", + "text": "10", + "bbox": [ + 217, + 114, + 235, + 126 + ], + "page_idx": 9 + }, + { + "type": "header", + "text": "J. Li, J. Wu, et al.", + "bbox": [ + 271, + 114, + 393, + 127 + ], + "page_idx": 9 + }, + { + "type": "table", + "img_path": "images/d8f86c5fc781e12d68b6fb00a84da010f2213b793afaf9dde6858a4f8fac1a89.jpg", + "table_caption": [ + "Table 2: Cross-category generalization performance compared with VLPart. The evaluation metric is $mAP_{mask}$ on the validation set of PascalPart but report AP50 specifically for dog parts. All models utilize ResNet-50 as backbone and use the text embeddings of the category names as the classifier. Base part represents the base split from Pascal Part. VOC object is added to the training process to improve the cognitive ability of the model thus reach a better performance. Pascal Part denotes the fully-supervised method reported for comparison." + ], + "table_footnote": [], + "table_body": "
MethodDatasetsAll AP (93)BaseAP (77)NovelAP (16)dogNovelAP Increment
headtorsolegpawtail
VLPart [62]Base Part15.017.81.56.17.92.913.83.2-
+ VOC object16.819.92.129.922.63.212.42.1+0.6
+ IN-S20 label17.420.81.112.817.82.05.90.9-0.4
+ Parsing IN-S2018.421.34.228.734.817.25.714.3+2.7
Pascal Part19.418.822.488.049.638.348.925.8-
PartGLEEBase Part25.630.52.112.615.68.25.26.2-
+ VOC object26.931.25.846.535.027.014.715.1+3.7
+ Parsing IN-S2026.628.915.580.357.336.717.037.4+9.7
Pascal Part35.534.639.995.988.575.076.772.9-
", + "bbox": [ + 225, + 266, + 782, + 404 + ], + "page_idx": 10 + }, + { + "type": "text", + "text": "parse novel objects into their corresponding parts based on the generalization capability brought by CLIP.", + "bbox": [ + 212, + 446, + 782, + 474 + ], + "page_idx": 10 + }, + { + "type": "list", + "sub_type": "text", + "list_items": [ + "2) Cross-category Part Segmentation on Pascal Part. We follow the evaluation setting proposed by VLPart to assess the cross-category generalization performance of our model on the Pascal Part dataset. A total of 93 part categories are divided into 77 base part categories and 16 novel part categories. Tab. 2 presents the evaluation results for all (93), base (77), and novel (16) parts. The model is trained only on the base categories, and is directly evaluated on the entire datasets. Note that IN-S20 label represents adding image-level classification data and Parsing IN-S20 is on behalf of the pseudo-labels generated by VLPart [62] on ImageNet [8]. We further introduce a metric called NovelAP Increment on top of VLPart to assess the improvement of our model when adding extra object datasets into the training process. It is calculated by subtracting the baseline Novel AP from the Novel AP achieved after incorporating extra datasets. The results shown in Tab. 2 demonstrate that our method surpasses the performance of VLPart by a large margin. By comparing the NovelAP Increment, we observe that our method achieves a greater increment after incorporating extra object dataset. Since the VOC dataset includes object categories corresponding to novel parts, the hierarchical relationships of the Q-Former can be extended to novel part categories, resulting in a higher NovelAP Increment.", + "3) Generalized Zero-Shot Part Segmentation. We adopt the Oracle-Obj setting proposed by OV-PARTS [70] to conduct experiments on ADE-Part234 and Pascal-Part-116 datasets. This setting assumes that the ground-truth masks and categories of object-level instances are known during the inference process, aiming to evaluate the capability of the model to parse any novel object. All categories in the datasets are divided into a base set and a novel set," + ], + "bbox": [ + 212, + 477, + 787, + 840 + ], + "page_idx": 10 + }, + { + "type": "header", + "text": "PartGLEE", + "bbox": [ + 656, + 114, + 730, + 126 + ], + "page_idx": 10 + }, + { + "type": "page_number", + "text": "11", + "bbox": [ + 767, + 114, + 782, + 126 + ], + "page_idx": 10 + }, + { + "type": "table", + "img_path": "images/ef6ba8175a945b50696fe6cad2e37606c3f9ebc9f0f803144782ac6208b0d3e5.jpg", + "table_caption": [ + "Table 3: Generalized zero-Shot part segmentation performance on ADE-Part-234 and Pascal-Part-116 compared with baselines proposed by OV-PARTS." + ], + "table_footnote": [], + "table_body": "
MethodModelBackboneFinetuningOracle-Obj
ADE-Part-234Oracle-obj
SeenUnseenHarmonicSeenUnseenHarmonic
Fully SupervisedMask2FormerResNet-50X46.2547.86-55.2852.14-
Two-StageZSeg+ResNet-50CPTCoOp43.1927.8433.8555.3319.1728.48
ResNet-50CPTCoCoOp39.6725.1530.7854.4319.0428.21
ResNet-101cCPTCoOp43.4125.7032.2857.8821.9331.81
One-StageCATSegResNet-101&ViT-B/16X11.498.569.8114.8910.2912.17
ResNet-101&ViT-B/16B+D31.4025.7728.3143.9726.1132.76
CLIPSegViT-B/16X15.2718.0116.5322.3319.7320.95
ViT-B/16VA+L+F+D38.9629.6533.6748.6827.3735.04
PartGLEEResNet-50X51.2935.3341.8357.4327.4137.11
", + "bbox": [ + 225, + 196, + 779, + 337 + ], + "page_idx": 11 + }, + { + "type": "text", + "text": "and the training process is performed only on the base set, while we evaluate the performance of the model on all categories. As shown in Tab. 3, our model achieves a superior performance on both datasets, which indicates the importance of hierarchical modeling. The establishment of hierarchical relationships between objects and parts enables our model to extend to novel objects, thereby effectively parsing them into corresponding semantic parts. Consequently, our model exhibits outstanding performance across both datasets.", + "bbox": [ + 212, + 383, + 784, + 491 + ], + "page_idx": 11 + }, + { + "type": "text", + "text": "4.3 Joint-training results on Detection and Segmentation", + "text_level": 1, + "bbox": [ + 215, + 527, + 699, + 542 + ], + "page_idx": 11 + }, + { + "type": "text", + "text": "To endow our model with robust generalization capability, we perform joint training on various datasets and evaluate its performance on both object-level and part-level tasks. We compare our model with specialist and generalist models to evaluate its performance on object-level data. Additionally, we contrast it with VLPart to assess its performance on part-level datasets as well as the effectiveness of joint-training process on both types of datasets. As shown in Tab. 4, PartGLEE significantly outperforms VLPart on both object-level and part-level tasks after joint-training, while achieving comparable performance on object-level tasks compared with previous SOTA. Through joint-training, our model has acquired strong generalization performance, allowing it to simultaneously address tasks for different hierarchies. We also observe that VLPart fails to achieve satisfactory performance on both object-level and part-level tasks. For example, VLPart obtains better performance on Pascal Part than its dataset-specific oracle, while decreasing its performance on COCO and LVIS. We attribute the performance drop of VLPart to the absence of hierarchical relationships, which causes confusion in modeling parts and objects and impairs object-level performance. PartGLEE effectively addresses this problem and extends the generalization capabilities from object-level to part-level tasks.", + "bbox": [ + 212, + 568, + 785, + 840 + ], + "page_idx": 11 + }, + { + "type": "page_number", + "text": "12", + "bbox": [ + 217, + 114, + 235, + 126 + ], + "page_idx": 11 + }, + { + "type": "header", + "text": "J. Li, J. Wu, et al.", + "bbox": [ + 271, + 114, + 393, + 127 + ], + "page_idx": 11 + }, + { + "type": "table", + "img_path": "images/9d0326b2d8877f17c4f15a4d1aed5717cf1b7ec8e84eecc7c4ad8a08481013cd.jpg", + "table_caption": [ + "Table 4: Joint-Training Performance of PartGLEE. Note that Oracle represents the dataset-specific training paradigm. We directly evaluate the generalist models on PACO to assess their recognition capability at the part level, as indicated by the results annotated in the grey font." + ], + "table_footnote": [], + "table_body": "
TypeMethodPart-level TasksObject-level Tasks
PartImageNetPascal PartPACOCOCO-valLVIS-minimalLVIS-val
APboxAPmaskAPboxAPmaskAPmaskAPobjmaskAPpartmaskAPboxAPmaskAPboxAPmaskAPbox
SpecialistMask2Former(R50) [7]------46.243.7----
Mask2Former(L) [7]------50.1-----
MaskDINO(R50) [27]------50.546.0----
MaskDINO(L) [27]------58.352.1----
ViTDet-L [32]------57.649.8--51.246.0
ViTDet-H [32]------57.649.8--53.448.1
EVA-02-L [12]------64.255.0--65.257.3
PACO(R50) [55]----32.612.5------
PACO(L) [55]----43.417.7------
GeneralistPix2Seq v2 [5]------46.538.2----
X-Decoder(L) [89]---2.6911.90.94-46.7----
SEEM(L) [90]---1.998.420.69-47.7----
HIPIE(R50) [68]------53.945.9----
Florence-2(B) [72]------41.4-----
Florence-2(L) [72]------43.4-----
UNINEXT(R50) [74]------51.344.9--36.4-
UNINEXT(L) [74]------58.149.6----
GLEE(R50) [71]---3.4415.31.2955.048.450.545.944.240.2
HierarchicalVLPart(R50) [62]30.731.623.924.013.836.99.628.5--26.2-
VLPart(R50)-Oracle [62]29.229.718.919.413.328.010.638.0--28.1-
VLPart(B) [62]43.941.233.531.722.155.015.940.3--39.6-
VLPart(B)-Oracle [62]44.341.729.227.419.137.715.252.5--43.1-
PartGLEE (R50)40.940.235.035.521.850.515.454.447.648.743.542.7
PartGLEE (L)52.750.939.639.127.855.721.359.552.056.550.645.0
", + "bbox": [ + 222, + 224, + 784, + 441 + ], + "page_idx": 12 + }, + { + "type": "text", + "text": "4.4 Ablation Study", + "text_level": 1, + "bbox": [ + 215, + 484, + 388, + 500 + ], + "page_idx": 12 + }, + { + "type": "text", + "text": "To demonstrate that our model design achieves satisfactory results on both object-level and part-level tasks, we conduct an ablation study (depicted in Fig. 4) on the model architecture and present results in Tab. 5. We ablate with a backbone of ResNet-50 and perform joint-training on COCO [37], LVIS [16], PartImageNet [18], Pascal Part [6] and PACO [55] with 90K iterations. From this study, we draw several important conclusions: (1) The utilization of parallel pixel decoders only results in slight improvements in mask predictions on few datasets, indicating that the influence of feature maps at different granularities is negligible. (2) Adopting independent decoders to obtain predictions at different levels demonstrates superior performance across the majority of datasets, manifesting the effectiveness of independent decoders. As adopting parallel pixel decoders (b) results in significant GPU memory costs without considerable gains, and all metrics for (a) are lower than (c), we select (c) as our final model design. Additional ablation studies, extensive qualitative analysis, and experiments on mLLM can be found in the supplementary materials.", + "bbox": [ + 217, + 513, + 787, + 739 + ], + "page_idx": 12 + }, + { + "type": "text", + "text": "4.5 Limitations", + "text_level": 1, + "bbox": [ + 215, + 766, + 356, + 780 + ], + "page_idx": 12 + }, + { + "type": "text", + "text": "In this work, we still adopt CLIP as the text encoder, which is trained on text-image pairs and thus lacks the ability to perceive fine-grained descriptions of object or part instances. This limitation may restrict the improvement of model", + "bbox": [ + 215, + 794, + 785, + 839 + ], + "page_idx": 12 + }, + { + "type": "header", + "text": "PartGLEE", + "bbox": [ + 656, + 114, + 730, + 126 + ], + "page_idx": 12 + }, + { + "type": "page_number", + "text": "13", + "bbox": [ + 767, + 114, + 785, + 126 + ], + "page_idx": 12 + }, + { + "type": "table", + "img_path": "images/380a1a8ccedf3ac0c73b931153dde8949c159fc9a688bdbba265db0f4c15f336.jpg", + "table_caption": [ + "Table 5: An ablation study on different model designs, as depicted in Fig 4. Note that Parallel Pixel Decoders refers to the utilization of two pixel decoders to generate feature maps at different hierarchies respectively. Independent Decoders denote the usage of two decoders, which facilitate the interaction between feature maps and queries at different hierarchies. Our final choice is scheme (c), which is highlighted in gray." + ], + "table_footnote": [], + "table_body": "
SchemeModel DesignPart-level TasksObject-level Tasks
Parrallel Pixel DecodersIndependent DecodersPartImageNetPascal PartPACOCOCO-valLVIS-minimal
APmaskAPmaskAPmaskAPmaskAPobjmaskAPopartmaskAPboxAPmaskAPboxAPmask
(a)XX39.034.120.147.413.547.843.534.833.4
(b)✓X38.334.520.848.813.848.544.334.934.2
(c)X✓39.034.720.947.914.249.344.235.633.8
", + "bbox": [ + 222, + 238, + 784, + 311 + ], + "page_idx": 13 + }, + { + "type": "image", + "img_path": "images/4dcd8f982810a9b1fe787c65d99b5e0e3b46a23f84bf7f62b07359704c85f0f0.jpg", + "image_caption": [ + "(a)" + ], + "image_footnote": [], + "bbox": [ + 223, + 354, + 403, + 445 + ], + "page_idx": 13 + }, + { + "type": "image", + "img_path": "images/1974bef7631da668f780e78acdb603bc1363bc5433c6727de90ac6ef375762d6.jpg", + "image_caption": [ + "(b)", + "(c)", + "Fig. 4: Various designs for generating predictions at different hierarchies. In scheme (a), we only utilize a single decoder to generate predictions for both objects and parts. In scheme (b), two parallel pixel decoders are employed to generate feature maps at different levels, aiming to explore the effectiveness of feature maps at different granularity. In scheme (c), we use two independent decoders to generate predictions for objects and parts respectively." + ], + "image_footnote": [], + "bbox": [ + 411, + 349, + 589, + 450 + ], + "page_idx": 13 + }, + { + "type": "image", + "img_path": "images/5f24408bc94c656ce390a6f85c1be5a280f2124a8feaa04d680d885d7a1149f2.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 596, + 354, + 774, + 443 + ], + "page_idx": 13 + }, + { + "type": "text", + "text": "performance and prompts us to consider how to enhance the perception capabilities of region-level models, which will be our future work.", + "bbox": [ + 212, + 590, + 785, + 619 + ], + "page_idx": 13 + }, + { + "type": "text", + "text": "5 Conclusion", + "text_level": 1, + "bbox": [ + 215, + 650, + 359, + 666 + ], + "page_idx": 13 + }, + { + "type": "text", + "text": "In this paper, we introduce PartGLEE, a groundbreaking foundation model designed towards a complete comprehension of both objects and parts in images. Through the generic hierarchical relationships established by the Q-Former, we are able to break through the limitation of scarce part-level data by introducing a large amount of object-level data, thereby transferring the powerful generalization capabilities from objects to parts. Through extensive training on diverse datasets, PartGLEE achieves SOTA performance across various part-level tasks while maintaining competitive results on object-level tasks, enabling it to parse any objects into parts and serve as a foundation model for general fine-grained region-level perception tasks.", + "bbox": [ + 212, + 688, + 787, + 840 + ], + "page_idx": 13 + }, + { + "type": "page_number", + "text": "14", + "bbox": [ + 217, + 114, + 235, + 126 + ], + "page_idx": 13 + }, + { + "type": "header", + "text": "J. Li, J. Wu, et al.", + "bbox": [ + 271, + 114, + 395, + 127 + ], + "page_idx": 13 + }, + { + "type": "text", + "text": "Acknowledgements", + "text_level": 1, + "bbox": [ + 217, + 143, + 401, + 162 + ], + "page_idx": 14 + }, + { + "type": "text", + "text": "This work was supported by the National Science Fund for Distinguished Young Scholars of China (Grant No.62225603).", + "bbox": [ + 215, + 172, + 784, + 204 + ], + "page_idx": 14 + }, + { + "type": "text", + "text": "References", + "text_level": 1, + "bbox": [ + 217, + 224, + 321, + 239 + ], + "page_idx": 14 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "1. Alayrac, J.B., Donahue, J., Luc, P., Miech, A., Barr, I., Hasson, Y., Lenc, K., Mensch, A., Millican, K., Reynolds, M., et al.: Flamingo: a visual language model for few-shot learning. In: Advances in neural information processing systems. vol. 35, pp. 23716-23736 (2022)", + "2. Brohan, A., Brown, N., Carbajal, J., Chebotar, Y., Dabis, J., Finn, C., Gopalakrishnan, K., Hausman, K., Herzog, A., Hsu, J., et al.: Rt-1: Robotics transformer for real-world control at scale. arXiv preprint arXiv:2212.06817 (2022)", + "3. Brown, T., Mann, B., Ryder, N., Subbiah, M., Kaplan, J.D., Dhariwal, P., Neelakantan, A., Shyam, P., Sastry, G., Askell, A., et al.: Language models are few-shot learners. In: Advances in neural information processing systems. vol. 33, pp. 1877-1901 (2020)", + "4. Caron, M., Touvron, H., Misra, I., Jégou, H., Mairal, J., Bojanowski, P., Joulin, A.: Emerging properties in self-supervised vision transformers. In: Proceedings of the IEEE/CVF international conference on computer vision. pp. 9650-9660 (2021)", + "5. Chen, T., Saxena, S., Li, L., Lin, T.Y., Fleet, D.J., Hinton, G.E.: A unified sequence interface for vision tasks. Advances in Neural Information Processing Systems 35, 31333-31346 (2022)", + "6. Chen, X., Mottaghi, R., Liu, X., Fidler, S., Urtasun, R., Yuille, A.: Detect what you can: Detecting and representing objects using holistic models and body parts. In: Proceedings of the IEEE conference on computer vision and pattern recognition. pp. 1971-1978 (2014)", + "7. Cheng, B., Misra, I., Schwing, A.G., Kirillov, A., Girdhar, R.: Masked-attention mask transformer for universal image segmentation. In: Proceedings of the IEEE/CVF conference on computer vision and pattern recognition. pp. 1290-1299 (2022)", + "8. Deng, J., Dong, W., Socher, R., Li, L.J., Li, K., Fei-Fei, L.: Imagenet: A large-scale hierarchical image database. In: Proceedings of the IEEE/CVF International Conference on Computer Vision. pp. 248-255 (2009)", + "9. Devlin, J., Chang, M.W., Lee, K., Toutanova, K.: Bert: Pre-training of deep bidirectional transformers for language understanding. arXiv preprint arXiv:1810.04805 (2018)", + "0. Dong, J., Chen, Q., Shen, X., Yang, J., Yan, S.: Towards unified human parsing and pose estimation. In: Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition. pp. 843-850 (2014)", + "1. Esser, P., Rombach, R., Ommer, B.: Taming transformers for high-resolution image synthesis. In: Proceedings of the IEEE/CVF conference on computer vision and pattern recognition. pp. 12873-12883 (2021)", + "2. Fang, Y., Sun, Q., Wang, X., Huang, T., Wang, X., Cao, Y.: Eva-02: A visual representation for neon genesis. arXiv preprint arXiv:2303.11331 (2023)", + "3. Fang, Y., Wang, W., Xie, B., Sun, Q., Wu, L., Wang, X., Huang, T., Wang, X., Cao, Y.: Eva: Exploring the limits of masked visual representation learning at scale. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 19358-19369 (2023)" + ], + "bbox": [ + 225, + 252, + 785, + 839 + ], + "page_idx": 14 + }, + { + "type": "header", + "text": "PartGLEE", + "bbox": [ + 656, + 114, + 730, + 126 + ], + "page_idx": 14 + }, + { + "type": "page_number", + "text": "15", + "bbox": [ + 767, + 116, + 785, + 126 + ], + "page_idx": 14 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "14. de Geus, D., Meletis, P., Lu, C., Wen, X., Dubbelman, G.: Part-aware panoptic segmentation. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 5485-5494 (2021)", + "15. Gong, K., Liang, X., Zhang, D., Shen, X., Lin, L.: Look into person: Self-supervised structure-sensitive learning and a new benchmark for human parsing. In: Proceedings of the IEEE conference on computer vision and pattern recognition. pp. 932-940 (2017)", + "16. Gupta, A., Dollar, P., Girshick, R.: Lvis: A dataset for large vocabulary instance segmentation. In: Proceedings of the IEEE/CVF conference on computer vision and pattern recognition. pp. 5356-5364 (2019)", + "17. He, J., Chen, J., Lin, M.X., Yu, Q., Yuille, A.L.: Compositor: Bottom-up clustering and compositing for robust part and object segmentation. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 11259-11268 (2023)", + "18. He, J., Yang, S., Yang, S., Kortylewski, A., Yuan, X., Chen, J.N., Liu, S., Yang, C., Yu, Q., Yuille, A.: Partimagenet: A large, high-quality dataset of parts. In: European Conference on Computer Vision. pp. 128-145. Springer (2022)", + "19. He, K., Chen, X., Xie, S., Li, Y., Dólár, P., Girshick, R.: Masked autoencoders are scalable vision learners. In: Proceedings of the IEEE/CVF conference on computer vision and pattern recognition. pp. 16000-16009 (2022)", + "20. He, K., Zhang, X., Ren, S., Sun, J.: Deep residual learning for image recognition. In: Proceedings of the IEEE conference on computer vision and pattern recognition. pp. 770-778 (2016)", + "21. Jia, C., Yang, Y., Xia, Y., Chen, Y.T., Parekh, Z., Pham, H., Le, Q., Sung, Y.H., Li, Z., Duerig, T.: Scaling up visual and vision-language representation learning with noisy text supervision. In: International conference on machine learning. pp. 4904-4916. PMLR (2021)", + "22. Kawar, B., Zada, S., Lang, O., Tov, O., Chang, H., Dekel, T., Mosseri, I., Irani, M.: Imagic: Text-based real image editing with diffusion models. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 6007-6017 (2023)", + "23. Kirillov, A., Mintun, E., Ravi, N., Mao, H., Rolland, C., Gustafson, L., Xiao, T., Whitehead, S., Berg, A.C., Lo, W.Y., et al.: Segment anything. In: Proceedings of the IEEE/CVF International Conference on Computer Vision. pp. 4015-4026 (2023)", + "24. Krishna, R., Zhu, Y., Groth, O., Johnson, J., Hata, K., Kravitz, J., Chen, S., Kalantidis, Y., Li, L.J., Shamma, D.A., et al.: Visual genome: Connecting language and vision using crowdsourced dense image annotations. International journal of computer vision 123, 32-73 (2017)", + "25. Kuznetsova, A., Rom, H., Alldrin, N., Uijlings, J., Krasin, I., Pont-Tuset, J., Kamali, S., Popov, S., Malloci, M., Kolesnikov, A., et al.: The open images dataset v4: Unified image classification, object detection, and visual relationship detection at scale. International journal of computer vision 128(7), 1956-1981 (2020)", + "26. Li, F., Zhang, H., Sun, P., Zou, X., Liu, S., Yang, J., Li, C., Zhang, L., Gao, J.: Semantic-sam: Segment and recognize anything at any granularity. arXiv preprint arXiv:2307.04767 (2023)", + "27. Li, F., Zhang, H., Xu, H., Liu, S., Zhang, L., Ni, L.M., Shum, H.Y.: Mask dino: Towards a unified transformer-based framework for object detection and segmentation. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 3041-3050 (2023)" + ], + "bbox": [ + 215, + 146, + 785, + 839 + ], + "page_idx": 15 + }, + { + "type": "page_number", + "text": "16", + "bbox": [ + 217, + 114, + 235, + 126 + ], + "page_idx": 15 + }, + { + "type": "header", + "text": "J. Li, J. Wu, et al.", + "bbox": [ + 271, + 114, + 393, + 127 + ], + "page_idx": 15 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "28. Li, H., Zhu, J., Jiang, X., Zhu, X., Li, H., Yuan, C., Wang, X., Qiao, Y., Wang, X., Wang, W., et al.: Uni-perceiver v2: A generalist model for large-scale vision and vision-language tasks. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 2691-2700 (2023)", + "29. Li, J., Zhao, J., Wei, Y., Lang, C., Li, Y., Sim, T., Yan, S., Feng, J.: Multiple-human parsing in the wild. arXiv preprint arXiv:1705.07206 (2017)", + "30. Li, L.H., Zhang, P., Zhang, H., Yang, J., Li, C., Zhong, Y., Wang, L., Yuan, L., Zhang, L., Hwang, J.N., et al.: Grounded language-image pre-training. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 10965-10975 (2022)", + "31. Li, X., Xu, S., Yang, Y., Cheng, G., Tong, Y., Tao, D.: Panoptic-partformer: Learning a unified model for panoptic part segmentation. In: European Conference on Computer Vision. pp. 729-747. Springer (2022)", + "32. Li, Y., Mao, H., Girshick, R., He, K.: Exploring plain vision transformer backbones for object detection. In: European Conference on Computer Vision. pp. 280-296. Springer (2022)", + "33. Li, Y., Singh, K.K., Xue, Y., Lee, Y.J.: Partgan: Weakly-supervised part decomposition for image generation and segmentation. In: British Machine Vision Conference (BMVC) (2021)", + "34. Lin, C., Jiang, Y., Qu, L., Yuan, Z., Cai, J.: Generative region-language pretraining for open-ended object detection. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 13958-13968 (2024)", + "35. Lin, C., Sun, P., Jiang, Y., Luo, P., Qu, L., Haffari, G., Yuan, Z., Cai, J.: Learning object-language alignments for open-vocabulary object detection. In: The Eleventh International Conference on Learning Representations (2023)", + "36. Lin, T.Y., Goyal, P., Girshick, R., He, K., Dollar, P.: Focal loss for dense object detection. In: Proceedings of the IEEE international conference on computer vision. pp. 2980-2988 (2017)", + "37. Lin, T.Y., Maire, M., Belongie, S., Hays, J., Perona, P., Ramanan, D., Dollar, P., Zitnick, C.L.: Microsoft coco: Common objects in context. In: European Conference on Computer Vision. pp. 740-755. Springer (2014)", + "38. Ling, H., Kreis, K., Li, D., Kim, S.W., Torralba, A., Fidler, S.: Editgan: High-precision semantic image editing. In: Advances in Neural Information Processing Systems. vol. 34, pp. 16331-16345 (2021)", + "39. Liu, S., Zeng, Z., Ren, T., Li, F., Zhang, H., Yang, J., Li, C., Yang, J., Su, H., Zhu, J., et al.: Grounding dino: Marrying dino with grounded pre-training for open-set object detection. arXiv preprint arXiv:2303.05499 (2023)", + "40. Liu, Z., Lin, Y., Cao, Y., Hu, H., Wei, Y., Zhang, Z., Lin, S., Guo, B.: Swin transformer: Hierarchical vision transformer using shifted windows. In: Proceedings of the IEEE/CVF international conference on computer vision. pp. 10012-10022 (2021)", + "41. Lu, J., Clark, C., Zellers, R., Mottaghi, R., Kembhavi, A.: Unified-io: A unified model for vision, language, and multi-modal tasks. In: The Eleventh International Conference on Learning Representations (2022)", + "42. Ma, C., Jiang, Y., Wen, X., Yuan, Z., Qi, X.: Codet: Co-occurrence guided region-word alignment for open-vocabulary object detection. In: Advances in neural information processing systems. vol. 36 (2023)", + "43. Ma, C., Jiang, Y., Wu, J., Yuan, Z., Qi, X.: Groma: Localized visual tokenization for grounding multimodal large language models. arXiv preprint arXiv:2404.13013 (2024)" + ], + "bbox": [ + 212, + 146, + 787, + 840 + ], + "page_idx": 16 + }, + { + "type": "header", + "text": "PartGLEE", + "bbox": [ + 656, + 114, + 730, + 126 + ], + "page_idx": 16 + }, + { + "type": "page_number", + "text": "17", + "bbox": [ + 767, + 114, + 785, + 126 + ], + "page_idx": 16 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "44. Mao, J., Huang, J., Toshev, A., Camburu, O., Yuille, A.L., Murphy, K.: Generation and comprehension of unambiguous object descriptions. In: Proceedings of the IEEE conference on computer vision and pattern recognition. pp. 11-20 (2016)", + "45. Meletis, P., Wen, X., Lu, C., de Geus, D., Dubbelman, G.: Cityscapes-panoptic-parts and pascal-panoptic-parts datasets for scene understanding. arXiv preprint arXiv:2004.07944 (2020)", + "46. Michieli, U., Borsato, E., Rossi, L., Zanuttigh, P.: Gmnet: Graph matching network for large scale part semantic segmentation in the wild. In: European Conference on Computer Vision. pp. 397-414. Springer (2020)", + "47. Miletari, F., Navab, N., Ahmadi, S.A.: V-net: Fully convolutional neural networks for volumetric medical image segmentation. In: 2016 fourth international conference on 3D vision (3DV) (2016)", + "48. Morabia, K., Arora, J., Vijaykumar, T.: Attention-based joint detection of object and semantic part. arXiv preprint arXiv:2007.02419 (2020)", + "49. Nair, S., Rajeswaran, A., Kumar, V., Finn, C., Gupta, A.: R3m: A universal visual representation for robot manipulation. In: Conference on Robot Learning. pp. 892-909. PMLR (2023)", + "50. Ng, X.L., Ong, K.E., Zheng, Q., Ni, Y., Yeo, S.Y., Liu, J.: Animal kingdom: A large and diverse dataset for animal behavior understanding. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 19023-19034 (2022)", + "51. Pan, T.Y., Liu, Q., Chao, W.L., Price, B.: Towards open-world segmentation of parts. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 15392-15401 (2023)", + "52. Qi, L., Kuen, J., Guo, W., Gu, J., Lin, Z., Du, B., Xu, Y., Yang, M.H.: Aims: All-inclusive multi-level segmentation for anything. In: Advances in Neural Information Processing Systems. vol. 36 (2023)", + "53. Radford, A., Kim, J.W., Hallacy, C., Ramesh, A., Goh, G., Agarwal, S., Sastry, G., Askell, A., Mishkin, P., Clark, J., et al.: Learning transferable visual models from natural language supervision. In: International conference on machine learning. pp. 8748-8763. PMLR (2021)", + "54. Raffel, C., Shazeer, N., Roberts, A., Lee, K., Narang, S., Matena, M., Zhou, Y., Li, W., Liu, P.J.: Exploring the limits of transfer learning with a unified text-to-text transformer. The Journal of Machine Learning Research 21(1), 5485-5551 (2020)", + "55. Ramanathan, V., Kalia, A., Petrovic, V., Wen, Y., Zheng, B., Guo, B., Wang, R., Marquez, A., Kovvuri, R., Kadian, A., et al.: Paco: Parts and attributes of common objects. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 7141-7151 (2023)", + "56. Ramesh, A., Dhariwal, P., Nichol, A., Chu, C., Chen, M.: Hierarchical text-conditional image generation with clip latents. arXiv preprint arXiv:2204.06125 1(2), 3 (2022)", + "57. Ramesh, A., Pavlov, M., Goh, G., Gray, S., Voss, C., Radford, A., Chen, M., Sutskever, I.: Zero-shot text-to-image generation. In: International Conference on Machine Learning. pp. 8821-8831. PMLR (2021)", + "58. Reddy, N.D., Vo, M., Narasimhan, S.G.: Carfusion: Combining point tracking and part detection for dynamic 3d reconstruction of vehicles. In: Proceedings of the IEEE conference on computer vision and pattern recognition. pp. 1906-1915 (2018)", + "59. Rezatofighi, H., Tsoi, N., Gwak, J., Sadeghian, A., Reid, I., Savarese, S.: Generalized intersection over union: A metric and a loss for bounding box regression. In: Proceedings of the IEEE/CVF conference on computer vision and pattern recognition. pp. 658-666 (2019)" + ], + "bbox": [ + 215, + 147, + 784, + 839 + ], + "page_idx": 17 + }, + { + "type": "page_number", + "text": "18", + "bbox": [ + 217, + 114, + 235, + 126 + ], + "page_idx": 17 + }, + { + "type": "header", + "text": "J. Li, J. Wu, et al.", + "bbox": [ + 271, + 114, + 393, + 127 + ], + "page_idx": 17 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "60. Rombach, R., Blattmann, A., Lorenz, D., Esser, P., Ommer, B.: High-resolution image synthesis with latent diffusion models. In: Proceedings of the IEEE/CVF conference on computer vision and pattern recognition. pp. 10684-10695 (2022)", + "61. Shao, S., Li, Z., Zhang, T., Peng, C., Yu, G., Zhang, X., Li, J., Sun, J.: Objects365: A large-scale, high-quality dataset for object detection. In: Proceedings of the IEEE/CVF international conference on computer vision. pp. 8430-8439 (2019)", + "62. Sun, P., Chen, S., Zhu, C., Xiao, F., Luo, P., Xie, S., Yan, Z.: Going denser with open-vocabulary part segmentation. In: Proceedings of the IEEE/CVF International Conference on Computer Vision. pp. 15453-15465 (2023)", + "63. Tang, C., Xie, L., Zhang, X., Hu, X., Tian, Q.: Visual recognition by request. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 15265-15274 (2023)", + "64. Wah, C., Branson, S., Welinder, P., Perona, P., Belongie, S.: Caltech-ucsd birds200-2011 (cub-200-2011). Tech. Rep. CNS-TR-2011-001, California Institute of Technology (2011)", + "65. Wang, P., Shen, X., Lin, Z., Cohen, S., Price, B., Yuille, A.L.: Joint object and part segmentation using deep learned potentials. In: Proceedings of the IEEE International Conference on Computer Vision. pp. 1573-1581 (2015)", + "66. Wang, P., Yang, A., Men, R., Lin, J., Bai, S., Li, Z., Ma, J., Zhou, C., Zhou, J., Yang, H.: Ofa: Unifying architectures, tasks, and modalities through a simple sequence-to-sequence learning framework. In: International Conference on Machine Learning. pp. 23318-23340. PMLR (2022)", + "67. Wang, W., Bao, H., Dong, L., Bjorck, J., Peng, Z., Liu, Q., Aggarwal, K., Mohammed, O.K., Singhal, S., Som, S., et al.: Image as a foreign language: Beit pretraining for all vision and vision-language tasks. arXiv preprint arXiv:2208.10442 (2022)", + "68. Wang, X., Li, S., Kallidromitis, K., Kato, Y., Kozuka, K., Darrell, T.: Hierarchical open-vocabulary universal image segmentation. In: Advances in Neural Information Processing Systems. vol. 36 (2023)", + "69. Wang, Y., Xu, Y., Tsogkas, S., Bai, X., Dickinson, S., Siddiqi, K.: Deepflux for skeletons in the wild. In: Proceedings of the IEEE/CVF conference on computer vision and pattern recognition. pp. 5287-5296 (2019)", + "70. Wei, M., Yue, X., Zhang, W., Kong, S., Liu, X., Pang, J.: Ov-parts: Towards open-vocabulary part segmentation. In: Thirty-seventh Conference on Neural Information Processing Systems Datasets and Benchmarks Track (2023)", + "71. Wu, J., Jiang, Y., Liu, Q., Yuan, Z., Bai, X., Bai, S.: General object foundation model for images and videos at scale. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 3783-3795 (2024)", + "72. Xiao, B., Wu, H., Xu, W., Dai, X., Hu, H., Lu, Y., Zeng, M., Liu, C., Yuan, L.: Florence-2: Advancing a unified representation for a variety of vision tasks. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 4818-4829 (2024)", + "73. Yan, B., Jiang, Y., Sun, P., Wang, D., Yuan, Z., Luo, P., Lu, H.: Towards grand unification of object tracking. In: European Conference on Computer Vision (2022)", + "74. Yan, B., Jiang, Y., Wu, J., Wang, D., Luo, P., Yuan, Z., Lu, H.: Universal instance perception as object discovery and retrieval. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 15325-15336 (2023)", + "75. Yang, L., Song, Q., Wang, Z., Jiang, M.: Parsing r-cnn for instance-level human analysis. In: Proceedings of the IEEE/CVF conference on computer vision and pattern recognition. pp. 364-373 (2019)" + ], + "bbox": [ + 212, + 146, + 787, + 840 + ], + "page_idx": 18 + }, + { + "type": "header", + "text": "PartGLEE", + "bbox": [ + 656, + 114, + 730, + 126 + ], + "page_idx": 18 + }, + { + "type": "page_number", + "text": "19", + "bbox": [ + 767, + 114, + 785, + 126 + ], + "page_idx": 18 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "76. Yang, Y., Ramanan, D.: Articulated pose estimation with flexible mixtures-of-parts. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 1385-1392. IEEE (2011)", + "77. Yang, Z., Gan, Z., Wang, J., Hu, X., Ahmed, F., Liu, Z., Lu, Y., Wang, L.: Unitab: Unifying text and box outputs for grounded vision-language modeling. In: European Conference on Computer Vision. pp. 521-539. Springer (2022)", + "78. Yao, L., Han, J., Liang, X., Xu, D., Zhang, W., Li, Z., Xu, H.: Detclipv2: Scalable open-vocabulary object detection pre-training via word-region alignment. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 23497-23506 (2023)", + "79. Yao, L., Han, J., Wen, Y., Liang, X., Xu, D., Zhang, W., Li, Z., Xu, C., Xu, H.: Detclip: Dictionary-enriched visual-concept paralleled pre-training for open-world detection. In: Advances in Neural Information Processing Systems. vol. 35, pp. 9125-9138 (2022)", + "80. Yu, L., Poirson, P., Yang, S., Berg, A.C., Berg, T.L.: Modeling context in referring expressions. In: European Conference on Computer Vision. pp. 69-85. Springer (2016)", + "81. Yuan, L., Chen, D., Chen, Y.L., Codella, N., Dai, X., Gao, J., Hu, H., Huang, X., Li, B., Li, C., et al.: Florence: A new foundation model for computer vision. arXiv preprint arXiv:2111.11432 (2021)", + "82. Zareian, A., Rosa, K.D., Hu, D.H., Chang, S.F.: Open-vocabulary object detection using captions. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 14393-14402 (2021)", + "83. Zhong, Y., Yang, J., Zhang, P., Li, C., Codella, N., Li, L.H., Zhou, L., Dai, X., Yuan, L., Li, Y., et al.: Regionclip: Region-based language-image pretraining. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 16793-16803 (2022)", + "84. Zhou, B., Zhao, H., Puig, X., Xiao, T., Fidler, S., Barriuso, A., Torralba, A.: Semantic understanding of scenes through the ade20k dataset. International Journal on Computer Vision (2018)", + "85. Zhou, T., Wang, W., Liu, S., Yang, Y., Van Gool, L.: Differentiable multigranularity human representation learning for instance-aware human semantic parsing. In: Proceedings of the IEEE/CVF conference on computer vision and pattern recognition. pp. 1622-1631 (2021)", + "86. Zhu, X., Su, W., Lu, L., Li, B., Wang, X., Dai, J.: Deformable detr: Deformable transformers for end-to-end object detection. In: International Conference on Learning Representations (2021)", + "87. Zhu, X., Zhu, J., Li, H., Wu, X., Li, H., Wang, X., Dai, J.: Uni-perceiver: Pretraining unified architecture for generic perception for zero-shot and few-shot tasks. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 16804-16815 (2022)", + "88. Ziegler, A., Asano, Y.M.: Self-supervised learning of object parts for semantic segmentation. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 14502-14511 (2022)", + "89. Zou, X., Dou, Z.Y., Yang, J., Gan, Z., Li, L., Li, C., Dai, X., Behl, H., Wang, J., Yuan, L., et al.: Generalized decoding for pixel, image, and language. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 15116-15127 (2023)", + "90. Zou, X., Yang, J., Zhang, H., Li, F., Li, L., Wang, J., Wang, L., Gao, J., Lee, Y.J.: Segment everything everywhere all at once. In: Advances in Neural Information Processing Systems. vol. 36 (2023)" + ], + "bbox": [ + 215, + 147, + 784, + 839 + ], + "page_idx": 19 + }, + { + "type": "page_number", + "text": "20", + "bbox": [ + 217, + 114, + 235, + 126 + ], + "page_idx": 19 + }, + { + "type": "header", + "text": "J. Li, J. Wu, et al.", + "bbox": [ + 271, + 114, + 393, + 127 + ], + "page_idx": 19 + } +] \ No newline at end of file diff --git a/2024/PartGLEE_ A Foundation Model for Recognizing and Parsing Any Objects/a546ffc1-52e4-4141-be35-2478ed6c5802_model.json b/2024/PartGLEE_ A Foundation Model for Recognizing and Parsing Any Objects/a546ffc1-52e4-4141-be35-2478ed6c5802_model.json new file mode 100644 index 0000000000000000000000000000000000000000..5e8ec8b8377513dfbd6f7b5f63c6a41ab7c7722c --- /dev/null +++ b/2024/PartGLEE_ A Foundation Model for Recognizing and Parsing Any Objects/a546ffc1-52e4-4141-be35-2478ed6c5802_model.json @@ -0,0 +1,2605 @@ +[ + [ + { + "type": "title", + "bbox": [ + 0.223, + 0.141, + 0.78, + 0.187 + ], + "angle": 0, + "content": "PartGLEE: A Foundation Model for Recognizing and Parsing Any Objects" + }, + { + "type": "text", + "bbox": [ + 0.251, + 0.212, + 0.751, + 0.229 + ], + "angle": 0, + "content": "Junyi Li\\*, Junfeng Wu\\*, Weizhi Zhao\\*, Song Bai\\*, and Xiang Bai" + }, + { + "type": "text", + "bbox": [ + 0.334, + 0.24, + 0.668, + 0.268 + ], + "angle": 0, + "content": "1 Huazhong University of Science and Technology \n2 ByteDance Inc." + }, + { + "type": "text", + "bbox": [ + 0.262, + 0.31, + 0.74, + 0.545 + ], + "angle": 0, + "content": "Abstract. We present PartGLEE, a part-level foundation model for locating and identifying both objects and parts in images. Through a unified framework, PartGLEE accomplishes detection, segmentation, and grounding of instances at any granularity in the open world scenario. Specifically, we propose a Q-Former to construct the hierarchical relationship between objects and parts, parsing every object into corresponding semantic parts. By incorporating a large amount of object-level data, the hierarchical relationships can be extended, enabling PartGLEE to recognize a rich variety of parts. We conduct comprehensive studies to validate the effectiveness of our method, PartGLEE achieves the state-of-the-art performance across various part-level tasks and obtain competitive results on object-level tasks. The proposed PartGLEE significantly enhances hierarchical modeling capabilities and part-level perception over our previous GLEE model. Further analysis indicates that the hierarchical cognitive ability of PartGLEE is able to facilitate a detailed comprehension in images for mLLMs. The model and code will be released at https://provencstar.github.io/PartGLEE-Vision/." + }, + { + "type": "text", + "bbox": [ + 0.262, + 0.56, + 0.74, + 0.587 + ], + "angle": 0, + "content": "Keywords: Foundation Model \\(\\cdot\\) Hierarchical Recognition \\(\\cdot\\) Part Segmentation" + }, + { + "type": "title", + "bbox": [ + 0.217, + 0.615, + 0.377, + 0.631 + ], + "angle": 0, + "content": "1 Introduction" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.649, + 0.788, + 0.801 + ], + "angle": 0, + "content": "In recent years, foundation models have dominated the majority of tasks in the fields of Natural Language Processing [3, 9, 54] and Computer Vision [19, 23, 53, 60, 71, 72]. CLIP family [12, 13, 21, 53, 81] have made significant advancements in transfer learning and have demonstrated impressive zero-shot capabilities on vision-language tasks. SAM [23] has revolutionized the development of segmentation tasks and is able to provide multi-level class-agnostic masks. GLEE [71] utilized diverse object-level data to develop general object representations, enabling detection, segmentation, tracking, grounding, and identification of objects in open-world scenarios. Their remarkable achievement can be attributed to the integration of extensive and diverse range of datasets." + }, + { + "type": "page_footnote", + "bbox": [ + 0.218, + 0.811, + 0.786, + 0.84 + ], + "angle": 0, + "content": "* Equal Technical Contribution. Work done during Junfeng's internship at ByteDance. †Correspondence to Xiang Bai ." + } + ], + [ + { + "type": "page_number", + "bbox": [ + 0.218, + 0.116, + 0.23, + 0.127 + ], + "angle": 0, + "content": "2" + }, + { + "type": "header", + "bbox": [ + 0.272, + 0.115, + 0.396, + 0.128 + ], + "angle": 0, + "content": "J. Li, J. Wu, et al." + }, + { + "type": "image", + "bbox": [ + 0.229, + 0.148, + 0.78, + 0.52 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.216, + 0.532, + 0.788, + 0.588 + ], + "angle": 0, + "content": "Fig. 1: An illustrative example demonstrating image annotations at diverse granularities across multiple datasets. The annotations at hierarchical levels with corresponding relationships are depicted on the right side. Below is a visualization of our segmentation results at multiple granularities." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.625, + 0.788, + 0.806 + ], + "angle": 0, + "content": "Different from the vast quantity of object-level data, the scale of part-level data is relatively small, which turns out to be a major bottleneck hindering vision models from recognizing part-level instances. Thus, most vision models lack the hierarchical comprehension between objects and parts. However, it is evident that the ability to recognize parts from objects is essential for various practical applications such as image editing [22,33,38], behavior analysis [50,75], pose estimation [10,76], robotics manipulation [2,49], etc. Moreover, we observe that part-level information is able to help multi-modal Large Language Models (mLLMs) in achieving a more detailed understanding of image content. Since part-level comprehension is a critical ability for foundation models to tackle a broader range of problems, it leads to a natural question: How could we break through data limitations to build a part-level vision foundation model?" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.81, + 0.787, + 0.841 + ], + "angle": 0, + "content": "To enable object foundation model with part-level cognitive ability, we emphasize that the model should achieve two key objectives: (1) Hierarchical" + } + ], + [ + { + "type": "header", + "bbox": [ + 0.658, + 0.115, + 0.732, + 0.127 + ], + "angle": 0, + "content": "PartGLEE" + }, + { + "type": "page_number", + "bbox": [ + 0.775, + 0.117, + 0.785, + 0.127 + ], + "angle": 0, + "content": "3" + }, + { + "type": "text", + "bbox": [ + 0.218, + 0.147, + 0.786, + 0.479 + ], + "angle": 0, + "content": "Comprehension, the model is supposed to understand the intrinsic relationship between objects and parts, and extend this hierarchical connection to any novel object, (2) Semantic Granularity, the model should be capable of learning a universal feature representation, enabling it to recognize semantic instances at any granularity. Consequently, we present a method to jointly detect and segment both objects and parts in a top-down manner. A lightweight Querying Transformer (Q-Former) is proposed to construct the hierarchical relationship between objects and parts. Specifically, it employs a set of universal parsing queries to interact with object queries, consequently generating multiple part-level queries that are capable of predicting corresponding semantic parts for each object. The Q-Former acts as a decomposer, which first recognizes individual objects in the images and subsequently parsing them into parts. Such model design is built upon the observation that various common objects often exhibit shared characteristics of parts. For example, cats, dogs, and dinosaurs all have parts such as torso, legs, and tails. In this way, two sets of query embeddings at different levels are generated, which are then used to predict object-level and part-level instances respectively. Through this approach, the relationship between objects and parts is established via the Q-Former design. Meanwhile, the hierarchical levels of objects and parts are distinguished, which is different from previous research [6,55,62,68,70] that consider parts as fine-grained objects. This paradigm enables vision models to better understand the features on different levels during training, thereby achieving improved performance." + }, + { + "type": "text", + "bbox": [ + 0.218, + 0.494, + 0.786, + 0.839 + ], + "angle": 0, + "content": "Our complete solution, PartGLEE, for jointly detecting and segmenting instances at both object and part levels, makes it possible for vision models to achieve favorable outcomes on both object and part levels. Some previous research have devised specialized training paradigms to utilize abundant image-text pair data [79, 89] as well as grounding data [30, 39, 71], thereby enhancing the cognitive and generalization capabilities of the models. On the contrary, the quantity of part-level data is much smaller compared to object-level data. So far, the largest dataset incorporating the concepts of both objects and parts is the recently proposed PACO [55] dataset. The scarcity of data has limited research on part-level recognition and restricted the generalization improvement of vision models. Although VLPart [62] has attempted to utilize pseudo-labeling schemes to generate part-level annotations for both object-level and image-level datasets, the quality of the pseudo-labels is relatively poor. Our innovative algorithm that parsing objects into their corresponding parts facilitates the transfer of generalization capability from objects to parts. Consequently, parts are generated from objects, which enables vision models to maintain generalization performance when predicting parts for novel objects without labeling extensive part-level data. To facilitate the training process of Q-Former, we standardize the annotation granularity across various part-level datasets and introduce a vast amount of object-level datasets, an intuitive display of the overall training data is shown in Fig. 1. Unlike VLPart, which exhibits unsatisfactory performance at object-level datasets after joint-training, our method demonstrates favorable outcomes at both object and part levels after joint-training. Moreover, it turns out that" + } + ], + [ + { + "type": "page_number", + "bbox": [ + 0.218, + 0.116, + 0.23, + 0.127 + ], + "angle": 0, + "content": "4" + }, + { + "type": "header", + "bbox": [ + 0.272, + 0.115, + 0.397, + 0.128 + ], + "angle": 0, + "content": "J. Li, J. Wu, et al." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.147, + 0.784, + 0.177 + ], + "angle": 0, + "content": "using object-level datasets is able to improve the performance of the model on part-level tasks, indicating a beneficial interaction between objects and parts." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.178, + 0.788, + 0.419 + ], + "angle": 0, + "content": "Extensive experiments demonstrate that our method significantly improve the open-vocabulary part segmentation performance, concurrently ensuring a decent performance on object detection and segmentation. We verify its effectiveness on various popular datasets. To validate the generalization performance of our model in identifying various parts of novel objects, we conduct experiments on PartImageNet [18] and Pascal Part [6] datasets in cross-dataset and cross-category manners respectively. Our method exhibits strong transferability and generalization ability when adding extra object-level datasets during training. To evaluate the decomposition capability of our model, we conduct experiments on both ADE20K-Part and Pascal Part datasets follow OV-PARTS [70]. As a result, our approach significantly outperforms one-stage baselines of OV-PARTS, with an increase of \\(8.16\\%\\) and \\(2.07\\%\\) on harmonic mean IoU (hIoU) in ADE20K-Part-234 and Pascal-Part-116 respectively. Additionally, by incorporating a large amount of object-level data for joint-training, our method establishes generic hierarchical relationships and breaks through the limitations of scarce part-level data, achieving state-of-the-art performance across various part-level tasks." + }, + { + "type": "text", + "bbox": [ + 0.24, + 0.419, + 0.733, + 0.433 + ], + "angle": 0, + "content": "In conclusion, our main contributions can be summarized as follows:" + }, + { + "type": "text", + "bbox": [ + 0.223, + 0.444, + 0.784, + 0.485 + ], + "angle": 0, + "content": "1. We construct the hierarchical relationship between objects and parts via the Q-Former, facilitating part segmentation to acquire advantages from various object-level datasets." + }, + { + "type": "text", + "bbox": [ + 0.222, + 0.489, + 0.784, + 0.546 + ], + "angle": 0, + "content": "2. We propose a unified pipeline for hierarchical detection and segmentation, where we first recognize objects and then parsing them into corresponding semantic parts. This algorithm enables us to jointly detect and segment both object-level and part-level instances." + }, + { + "type": "text", + "bbox": [ + 0.222, + 0.549, + 0.784, + 0.606 + ], + "angle": 0, + "content": "3. We standardize the annotation granularity across various part-level datasets by incorporating corresponding object-level annotations, complementing the hierarchical correspondences for current part-level datasets, promoting the development of vision foundation models." + }, + { + "type": "list", + "bbox": [ + 0.222, + 0.444, + 0.784, + 0.606 + ], + "angle": 0, + "content": null + }, + { + "type": "title", + "bbox": [ + 0.216, + 0.633, + 0.388, + 0.649 + ], + "angle": 0, + "content": "2 Related Work" + }, + { + "type": "title", + "bbox": [ + 0.216, + 0.664, + 0.652, + 0.68 + ], + "angle": 0, + "content": "2.1 Visual Foundation Models and General Models" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.689, + 0.788, + 0.841 + ], + "angle": 0, + "content": "Vision foundation models and generalist models are considered as a milestone in the development of the intelligent vision system. For instance, multi-modal visual foundation models [1, 21, 53, 67, 81] have significantly advanced efficient transfer learning and exhibit impressive zero-shot capabilities on vision-language tasks by using contrastive learning with large-scale image-text pairs. Generative foundation models [11, 56, 57, 60] are trained on vast collections of images and captions, empowering them to generate image content conditioned on textual prompts. Self-supervised foundation models [4, 12, 13, 19] have learned general visual representations from large-scale image datasets, enhancing their ability to transfer to downstream tasks. However, the image-level features learned by these" + } + ], + [ + { + "type": "header", + "bbox": [ + 0.658, + 0.115, + 0.732, + 0.127 + ], + "angle": 0, + "content": "PartGLEE" + }, + { + "type": "page_number", + "bbox": [ + 0.776, + 0.117, + 0.786, + 0.127 + ], + "angle": 0, + "content": "5" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.147, + 0.784, + 0.177 + ], + "angle": 0, + "content": "foundation models are not well-suited for direct application to dense prediction tasks that involve precise object and part localization." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.178, + 0.788, + 0.45 + ], + "angle": 0, + "content": "Transformer-based generalist methods [5,41,66,77,87] adopt a sequence generation pipeline to unify the output of text and spatial coordinates. However, they mainly focus on image-level comprehension, which results in relatively weak localization capabilities. Works such as UNINEXT, etc. [28, 73, 74], built upon strong detectors [27, 86], demonstrating a strong localization capability across multiple datasets. But they fail to exhibit zero-shot transfer ability and generalization capability due to their closed-set training paradigm. Some works about open-vocabulary detection (OVD) [30, 34, 35, 42, 43, 78, 82, 83] have explored zero-shot generalization capabilities on novel categories. X-Decoder [89] and SEEM [90] have developed a versatile decoding architecture that are able to generate accurate pixel-level segmentation predictions. GLEE [71] addresses various object-level tasks through a unified architecture and training paradigm. However, current generalist models and foundation models are trained mainly on image-level and object-level datasets, thus their ability to extract more fine-grained information is limited, making it difficult for them to recognize corresponding semantic parts of any object. Our work focuses on empowering hierarchical cognitive capability for vision foundation models, thereby further advancing the development of comprehensive visual systems." + }, + { + "type": "title", + "bbox": [ + 0.215, + 0.469, + 0.419, + 0.485 + ], + "angle": 0, + "content": "2.2 Part Segmentation" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.493, + 0.789, + 0.844 + ], + "angle": 0, + "content": "The growing interest in achieving a more fine-grained understanding of objects has sparked a surge in research focused on part level recognition. Some pioneering studies have introduced datasets with part-level annotations, concentrating on objects of some specific categories such as human body parts [15, 29, 75], animal body parts [64] and vehicle components [58]. More general part annotations for common objects such as Pascal-Part [6], PartImageNet [18], ADE20K [84], CityscapesPanoptic-Parts [45] and more recent PACO [55] were then proposed to promote more in-depth research in the field of parts. Most of the previous works [14, 31, 46, 85] were conducted based on a closed-set configuration, thus only capable of detecting and segmenting closed-set objects and parts. Recently, VLPart [62] present a pipeline for detecting and segmenting both open-vocabulary objects and their corresponding part regions, while OVPARTS [70] utilize adapters to transfer the generalization abilities of CLIP into open-vocabulary part segmentation task. However, due to the limited quantity of data, the generalization capability of previous models [15, 29, 51, 62, 63, 68, 70, 75] still relies heavily on the training datasets. Furthermore, in prior works, both objects and parts are treated equally, they consider part as a special type of object. On the contrary, we distinguish them by considering parts as integral components subordinate to objects and generate parts from corresponding objects in a top-down manner. Our work is aimed at building hierarchical relationships while unifying the training paradigm for object and part-level data. By incorporating a large amount of object-level data, the hierarchical relationships can be extended to any object, enabling our method to recognize a rich variety of parts." + } + ], + [ + { + "type": "page_number", + "bbox": [ + 0.218, + 0.116, + 0.23, + 0.127 + ], + "angle": 0, + "content": "6" + }, + { + "type": "header", + "bbox": [ + 0.272, + 0.115, + 0.396, + 0.128 + ], + "angle": 0, + "content": "J. Li, J. Wu, et al." + }, + { + "type": "image", + "bbox": [ + 0.25, + 0.145, + 0.761, + 0.342 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.214, + 0.354, + 0.788, + 0.397 + ], + "angle": 0, + "content": "Fig. 2: Framework of PartGLEE. The Q-Former takes each object query as input and output the corresponding part queries. These queries are then fed into the object decoder and the part decoder respectively to generate hierarchical predictions." + }, + { + "type": "title", + "bbox": [ + 0.215, + 0.425, + 0.618, + 0.44 + ], + "angle": 0, + "content": "2.3 Hierarchical Learning of Objects and Parts" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.448, + 0.789, + 0.706 + ], + "angle": 0, + "content": "Learning objects through parts has been a long-standing research topic as part annotations provide more detailed semantic information of objects. Morabia et al. [48] first introduced a pipeline employing an attention mechanism for simultaneous detection of both objects and parts. Deepflux [69] designed an image context flux representation which enables better object parts interaction for skeleton detection. Leopart [88] demonstrated that learning object parts can provide spatially diverse representation which facilitates self-supervised semantic segmentation. Wang et al. [65] proposed a method to predict both parts and objects simultaneously on Pascal-Part dataset [6]. Recent studies such as SAM [23] and Semantic-SAM [26] have studied on class-agnostic multi-granularity interactive segmentation task. However, they have not explored the relationship between objects and their corresponding semantic parts. Recently, Compositor [17] designed a bottom-up pipeline to predict parts and then cluster them into objects, while AIMS [52] utilized an independent relation decoder to construct the hierarchical association between objects and parts. Different from these works, our approach introduces a Querying Transformer to model the hierarchical relationship, allowing our model to parse any object into its corresponding parts." + }, + { + "type": "title", + "bbox": [ + 0.215, + 0.727, + 0.331, + 0.741 + ], + "angle": 0, + "content": "3 Method" + }, + { + "type": "title", + "bbox": [ + 0.215, + 0.756, + 0.421, + 0.77 + ], + "angle": 0, + "content": "3.1 Overall Framework" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.78, + 0.784, + 0.809 + ], + "angle": 0, + "content": "Following [68,71], we propose PartGLEE, which comprises of an image encoder, a Q-Former, two independent decoders and a text encoder, as shown in Fig. 2." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.81, + 0.785, + 0.84 + ], + "angle": 0, + "content": "Given an input image \\(I \\in \\mathcal{R}^{H \\times W \\times 3}\\), the backbone and the pixel decoder first extract multi-scale image features \\(F_s \\in \\mathcal{R}_{\\frac{H}{2^S} \\times \\frac{W}{2^S} \\times C}\\) and \\(s = \\{2, 3, 4, 5\\}\\) with" + } + ], + [ + { + "type": "header", + "bbox": [ + 0.657, + 0.115, + 0.732, + 0.127 + ], + "angle": 0, + "content": "PartGLEE" + }, + { + "type": "page_number", + "bbox": [ + 0.776, + 0.117, + 0.786, + 0.127 + ], + "angle": 0, + "content": "7" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.146, + 0.788, + 0.404 + ], + "angle": 0, + "content": "backbones such as ResNet [20] or Swin Transformer [40]. Then we feed them into the object decoder, where the object-level query embeddings \\( q_{obj} \\in \\mathcal{R}^{N \\times C} \\) are generated in a two-stage process. These object queries are utilized to perform object-level classification, detection as well as segmentation tasks through three independent prediction heads. Besides, the object queries \\( q_{obj} \\) are fed into the Q-Former simultaneously, where \\( L \\) learnable universal parsing queries are initialized to interact with object queries. It takes object queries as input and generate part-level queries \\( q_{part} \\in \\mathcal{R}^{N \\cdot L \\times C} \\) which are then passed into the part decoder to yield part-level predictions (detailed in Sec. 3.2). To enhance the semantic-awareness, an early fusion module is adopted before Transformer encoder following [74], which takes image feature from backbone and text embedding as input and perform bi-directional cross-attention between them. In line with previous segmentation models [7,27,32], a pixel embedding map \\( M_p \\in \\mathcal{R}^{\\frac{H}{4} \\times \\frac{W}{4} \\times C} \\) at \\( 1/4 \\) resolution is constructed by upsampling and integrating multi-scale feature maps from the backbone and the pixel decoder. Eventually, we dot product each object query or part query with the pixel embedding map to derive an output mask \\( m \\in \\mathcal{R}^{\\frac{H}{4} \\times \\frac{W}{4}} \\):" + }, + { + "type": "equation", + "bbox": [ + 0.367, + 0.413, + 0.785, + 0.428 + ], + "angle": 0, + "content": "\\[\nm = F F N (q _ {l}) \\otimes M _ {p}, \\quad l \\in \\{o b j, p a r t \\} \\tag {1}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.435, + 0.786, + 0.466 + ], + "angle": 0, + "content": "where FFN is comprised of 3 layers feed forward network with ReLU activation functions and linear layers." + }, + { + "type": "title", + "bbox": [ + 0.215, + 0.485, + 0.483, + 0.5 + ], + "angle": 0, + "content": "3.2 Parsing Objects into Parts" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.507, + 0.788, + 0.735 + ], + "angle": 0, + "content": "We propose a Q-Former to establish the hierarchical relationship between objects and parts. As Various common objects tend to manifest shared attributes in their constituent parts, for example, both lizards and birds exhibit similar components, such as heads and torsos. Thus, we initialize a set of query embeddings in the Q-Former to parse any object into semantic parts. We denote these universal parsing query embeddings as \\( q_{parse} \\in \\mathcal{R}^{L \\times C} \\), where \\( L \\) represents the number of the parsing queries. As shown in Fig. 2, the Q-Former is comprised of \\( M \\) cascaded attention modules, each module includes a self-attention layer, a cross-attention layer, and a feed forward network. The universal parsing queries are first fed into the self-attention layer and then perform cross-attention with the object queries. Note that every object query is interacted with all universal parsing queries. Hence, assume \\( N \\) object queries \\( (q_{obj} \\in \\mathcal{R}^{N \\times C}) \\) are generated from the object decoder, and \\( L \\) universal parsing queries \\( q_{parse} \\in \\mathcal{R}^{L \\times C} \\) are initialized in the Q-Former, we obtain \\( N \\cdot L \\) part-level queries which can be denoted as \\( q_{part} \\in \\mathcal{R}^{N \\cdot L \\times C} \\). We refer to this process as:" + }, + { + "type": "equation", + "bbox": [ + 0.393, + 0.743, + 0.785, + 0.758 + ], + "angle": 0, + "content": "\\[\nq _ {p a r t} = Q - F o r m e r \\left(q _ {p a r s e}; q _ {o b j}\\right) \\tag {2}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.765, + 0.788, + 0.841 + ], + "angle": 0, + "content": "Our proposed Q-Former functions as a decomposer, extracting and representing parts from object queries. Hence, by training jointly on extensive object-level datasets and limited hierarchical datasets which contain object-part correspondences, our Q-Former obtains strong generalization ability to parse any novel object into its corresponding parts." + } + ], + [ + { + "type": "page_number", + "bbox": [ + 0.218, + 0.116, + 0.23, + 0.127 + ], + "angle": 0, + "content": "8" + }, + { + "type": "header", + "bbox": [ + 0.272, + 0.115, + 0.396, + 0.128 + ], + "angle": 0, + "content": "J. Li, J. Wu, et al." + }, + { + "type": "image", + "bbox": [ + 0.223, + 0.148, + 0.778, + 0.312 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.215, + 0.324, + 0.788, + 0.355 + ], + "angle": 0, + "content": "Fig. 3: Matching mechanisms of PartGLEE. Two separate forward passes are performed on the same image to obtain hierarchical segmentation results." + }, + { + "type": "title", + "bbox": [ + 0.215, + 0.378, + 0.669, + 0.394 + ], + "angle": 0, + "content": "3.3 Unified Training Paradigm for Objects and Parts" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.399, + 0.784, + 0.445 + ], + "angle": 0, + "content": "Since the Q-Former requires hierarchical data to learn how to parse objects into parts, we enrich part-level data with corresponding object-level annotations, details of which are provided in the supplementary materials." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.445, + 0.788, + 0.61 + ], + "angle": 0, + "content": "Since the annotation granularity across part-level datasets is standardized, our model can first learn the characteristics of objects and then acquire the ability to parse any object into its semantic parts. To facilitate open-vocabulary detection and segmentation, we substitute the similarity scores between the instance embeddings and the text embeddings for the original class head. Given \\(K\\) object-level and part-level input categories as separate sentences, we feed them into the text encoder and utilize the average of each individual sentence tokens as the output text embedding \\(T_{l}\\) for each category. Then the similarity scores \\(S_{l} \\in \\mathcal{R}^{N \\times K}\\) are calculated through a dot product operation between the hierarchical instance embeddings \\(q_{l} \\in \\mathcal{R}^{N \\times C}\\) from detector and the text embeddings \\(T_{l} \\in \\mathcal{R}^{K \\times D}\\) from text encoder, which can be denoted as:" + }, + { + "type": "equation", + "bbox": [ + 0.367, + 0.618, + 0.785, + 0.634 + ], + "angle": 0, + "content": "\\[\nS _ {l} = q _ {l} \\cdot W _ {p r o j} \\otimes T _ {l}, \\quad l \\in \\{o b j, p a r t \\} \\tag {3}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.641, + 0.788, + 0.717 + ], + "angle": 0, + "content": "where \\( W_{proj} \\in \\mathcal{R}^{C \\times D} \\) is a trainable projection weight for fine-tuning text embedding space especially for part-level descriptions. Following [26, 68], we perform Hungarian matching of objects and parts individually, where object-level predictions are only matched with object-level targets, and the same applies to the part-level output, as shown in Fig. 3." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.718, + 0.788, + 0.793 + ], + "angle": 0, + "content": "We then introduce a constraint loss to ensure the part-level predictions to be the component of the objects. We denote this novel loss function as restriction loss \\( L_{res} \\). Due to memory limitations, we only calculate our restriction loss on the predicted bounding boxes between different levels, while leaving the predicted masks unconstrained. Our restriction loss can be calculated as follow:" + }, + { + "type": "equation", + "bbox": [ + 0.391, + 0.796, + 0.787, + 0.838 + ], + "angle": 0, + "content": "\\[\nL _ {r e s} = \\sum_ {i} ^ {L} \\left(1 - \\frac {\\left| S _ {o b j} \\cap S _ {p a r t} ^ {i} \\right|}{S _ {p a r t} ^ {i}}\\right) \\tag {4}\n\\]" + } + ], + [ + { + "type": "header", + "bbox": [ + 0.657, + 0.115, + 0.732, + 0.127 + ], + "angle": 0, + "content": "PartGLEE" + }, + { + "type": "page_number", + "bbox": [ + 0.776, + 0.117, + 0.786, + 0.127 + ], + "angle": 0, + "content": "9" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.147, + 0.784, + 0.221 + ], + "angle": 0, + "content": "where \\( S_{obj} \\) represents the area of the object-level bounding box prediction, and \\( S_{part}^{i} \\) stands for the area of the \\( i - th \\) part-level bounding box prediction. Note that each object query can generate \\( L \\) part queries through Q-Former. This loss function is only applied to the matched predictions in part-level datasets, thereby strengthening the mutual correspondence between different hierarchies." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.223, + 0.784, + 0.252 + ], + "angle": 0, + "content": "PartGLEE is trained with a linear combination of losses for object-level tasks and part-level tasks, which can be formulated as:" + }, + { + "type": "equation", + "bbox": [ + 0.236, + 0.263, + 0.786, + 0.283 + ], + "angle": 0, + "content": "\\[\nL = \\lambda_ {1} (L _ {c l s} ^ {o b j} + L _ {c l s} ^ {p a r t}) + \\lambda_ {2} (L _ {b o x} ^ {o b j} + L _ {b o x} ^ {p a r t}) + \\lambda_ {3} (L _ {m a s k} ^ {o b j} + L _ {m a s k} ^ {p a r t}) + \\lambda_ {4} L _ {r e s} \\quad (5)\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.293, + 0.785, + 0.43 + ], + "angle": 0, + "content": "where \\( L_{cls}^{l}, L_{box}^{l}, L_{mask}^{l} \\) are the classification, box, and mask loss at different levels (\\( l \\in \\{obj, part\\} \\)), while \\( L_{res} \\) is the restriction loss, and \\( \\lambda \\) are their corresponding weights. We apply Focal Loss [36] as the classification loss on the similarity scores \\( S_{l} \\) to align the text concepts with instance features. A combination of L1 loss and generalized IoU loss [59] is utilized for box predictions, while we employ both Dice Loss [47] and Focal Loss to calculate mask loss. We follow MaskDINO to set our hyperparameters to \\( \\lambda_{1} = 4, \\lambda_{2} = 2, \\lambda_{3} = 5, \\lambda_{4} = 5 \\). Based on the above designs, PartGLEE is able to leverage both object-level data and part-level data thus obtaining a strong generalization capability." + }, + { + "type": "title", + "bbox": [ + 0.216, + 0.452, + 0.376, + 0.469 + ], + "angle": 0, + "content": "4 Experiments" + }, + { + "type": "title", + "bbox": [ + 0.216, + 0.484, + 0.427, + 0.499 + ], + "angle": 0, + "content": "4.1 Experimental Setup" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.508, + 0.784, + 0.537 + ], + "angle": 0, + "content": "We conduct comprehensive experiments to exhibit the effectiveness of PartGLEE across a wide range of object-level and part-level tasks." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.538, + 0.785, + 0.688 + ], + "angle": 0, + "content": "Data Unification. We utilize object-level datasets such as COCO [37], LVIS [16], Object365 [61], OpenImages [25], Visual Genome [24] and RefCOCO series [44,80], etc, while using part-level datasets PACO [55], PartImageNet [18], Pascal Part [6], ADE20K-Part [70] and SA-1B [23] with varying annotation granularity for joint-training. For Visual Genome and SA-1B, we categorize their corresponding part-level annotations based on semantic and mask overlap relationships to construct hierarchical data versions. For part-level data, we integrate the original part-level annotations with corresponding object-level annotations according to their associated object-level dataset. The details of these dataset preprocessing steps are left in the supplementary materials." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.689, + 0.786, + 0.84 + ], + "angle": 0, + "content": "Implementation Details. In our experiments, we utilize ResNet-50 [20] and Swin-Large [40] as the vision encoder. Following MaskDINO [27], we adopt deformable transformer in the decoder, and use 300 object queries while setting the number of parsing queries \\( L \\) to be 10. The \\( M \\) of Q-Former is set to 6. We select the top 50 object queries based on the similarity scores and input them into the Q-Former, ultimately yielding 500 part queries. We use both query denoising and hybrid matching strategies to facilitate convergence and enhance performance. We conduct experiments on part-level datasets following the methodologies of VLPart [62] and OV-PARTS [70] in order to evaluate the generalization performance and the ability to parse novel objects of our model. For joint-training, we" + } + ], + [ + { + "type": "page_number", + "bbox": [ + 0.218, + 0.116, + 0.236, + 0.127 + ], + "angle": 0, + "content": "10" + }, + { + "type": "header", + "bbox": [ + 0.272, + 0.115, + 0.395, + 0.128 + ], + "angle": 0, + "content": "J. Li, J. Wu, et al." + }, + { + "type": "table_caption", + "bbox": [ + 0.217, + 0.145, + 0.785, + 0.2 + ], + "angle": 0, + "content": "Table 1: Cross-dataset generalization performance compared with VLPart. The evaluation metric is \\( mAP_{mask} \\) on the validation set of PartImageNet. All models utilize ResNet-50 as backbone and use the text embeddings of the category names as the classifier. PartImageNet denotes the fully-supervised method reported for comparison." + }, + { + "type": "table", + "bbox": [ + 0.306, + 0.226, + 0.703, + 0.376 + ], + "angle": 0, + "content": "
MethodDatasetsAll (40)quadruped
headbodyfoottail
VLPart [62]Pascal Part4.517.40.10.02.9
+ IN-S11 label5.423.63.40.81.2
+ Parsing IN-S117.835.015.23.58.9
vs. baseline+3.3+17.6+15.1+3.5+6.0
PartImageNet29.757.325.822.922.9
PartGLEEPascal Part9.923.64.51.34.6
+ Parsing IN-S1114.955.327.27.023.6
vs. baseline+5.0+31.7+22.7+5.7+19.0
PartImageNet40.267.037.636.540.7
" + }, + { + "type": "text", + "bbox": [ + 0.217, + 0.418, + 0.785, + 0.464 + ], + "angle": 0, + "content": "train PartGLEE based on the weights of GLEE [71], continuing training on 32 A100 GPUs. The settings for the part-level zero-shot experiments are described separately in each section." + }, + { + "type": "title", + "bbox": [ + 0.218, + 0.491, + 0.57, + 0.507 + ], + "angle": 0, + "content": "4.2 Zero-Shot Part Segmentation Results" + }, + { + "type": "text", + "bbox": [ + 0.217, + 0.522, + 0.786, + 0.718 + ], + "angle": 0, + "content": "1) Cross-dataset Part Segmentation on PartImageNet. We follow VLPart [62] to conduct experiments on cross-dataset generalization performance by directly evaluating on PartImageNet [18] validation set. We report the metrics of all (40) part categories and the detailed metrics of quadruped are also provided. The baseline approach only utilize Pascal Part as the training set and directly perform evaluation on PartImageNet in a zero-shot manner. Note that IN-S11 label represents adding image-level classification data for training in order to improve performance. Meanwhile, Parsing IN-S11 stands for training with the pseudo-labels generated from the parsing pipeline proposed by VLPart. However, both of these methods expose the model to categories and images from the PartImageNet dataset. We first perform our training process exclusively on the Pascal Part dataset to verify our zero-shot capabilities, and then we incorporate pseudo-labels to assess the ability of our model to utilize low-quality annotations." + }, + { + "type": "text", + "bbox": [ + 0.217, + 0.72, + 0.786, + 0.84 + ], + "angle": 0, + "content": "Given that Pascal Part does not provide semantic labels for categories like quadruped in PartImageNet, the model needs to generalize from annotated parts of dog, cat, etc. in Pascal Part to parts of quadruped in PartImageNet. As shown in Tab. 1, our model significantly outperforms VLPart when only training on Pascal Part, even surpassing the model trained with Parsing IN-S11. After incorporating pseudo-labeled data into training, our model shows higher performance gains, indicating better utilization of low-quality data. This result illustrates the importance of hierarchical modeling, which enables our model to recognize and" + } + ], + [ + { + "type": "header", + "bbox": [ + 0.658, + 0.115, + 0.732, + 0.127 + ], + "angle": 0, + "content": "PartGLEE" + }, + { + "type": "page_number", + "bbox": [ + 0.769, + 0.116, + 0.784, + 0.127 + ], + "angle": 0, + "content": "11" + }, + { + "type": "table_caption", + "bbox": [ + 0.214, + 0.145, + 0.788, + 0.243 + ], + "angle": 0, + "content": "Table 2: Cross-category generalization performance compared with VLPart. The evaluation metric is \\( mAP_{mask} \\) on the validation set of PascalPart but report AP50 specifically for dog parts. All models utilize ResNet-50 as backbone and use the text embeddings of the category names as the classifier. Base part represents the base split from Pascal Part. VOC object is added to the training process to improve the cognitive ability of the model thus reach a better performance. Pascal Part denotes the fully-supervised method reported for comparison." + }, + { + "type": "table", + "bbox": [ + 0.227, + 0.267, + 0.784, + 0.405 + ], + "angle": 0, + "content": "
MethodDatasetsAll AP (93)BaseAP (77)NovelAP (16)dogNovelAP Increment
headtorsolegpawtail
VLPart [62]Base Part15.017.81.56.17.92.913.83.2-
+ VOC object16.819.92.129.922.63.212.42.1+0.6
+ IN-S20 label17.420.81.112.817.82.05.90.9-0.4
+ Parsing IN-S2018.421.34.228.734.817.25.714.3+2.7
Pascal Part19.418.822.488.049.638.348.925.8-
PartGLEEBase Part25.630.52.112.615.68.25.26.2-
+ VOC object26.931.25.846.535.027.014.715.1+3.7
+ Parsing IN-S2026.628.915.580.357.336.717.037.4+9.7
Pascal Part35.534.639.995.988.575.076.772.9-
" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.447, + 0.784, + 0.476 + ], + "angle": 0, + "content": "parse novel objects into their corresponding parts based on the generalization capability brought by CLIP." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.478, + 0.788, + 0.749 + ], + "angle": 0, + "content": "2) Cross-category Part Segmentation on Pascal Part. We follow the evaluation setting proposed by VLPart to assess the cross-category generalization performance of our model on the Pascal Part dataset. A total of 93 part categories are divided into 77 base part categories and 16 novel part categories. Tab. 2 presents the evaluation results for all (93), base (77), and novel (16) parts. The model is trained only on the base categories, and is directly evaluated on the entire datasets. Note that IN-S20 label represents adding image-level classification data and Parsing IN-S20 is on behalf of the pseudo-labels generated by VLPart [62] on ImageNet [8]. We further introduce a metric called NovelAP Increment on top of VLPart to assess the improvement of our model when adding extra object datasets into the training process. It is calculated by subtracting the baseline Novel AP from the Novel AP achieved after incorporating extra datasets. The results shown in Tab. 2 demonstrate that our method surpasses the performance of VLPart by a large margin. By comparing the NovelAP Increment, we observe that our method achieves a greater increment after incorporating extra object dataset. Since the VOC dataset includes object categories corresponding to novel parts, the hierarchical relationships of the Q-Former can be extended to novel part categories, resulting in a higher NovelAP Increment." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.75, + 0.788, + 0.841 + ], + "angle": 0, + "content": "3) Generalized Zero-Shot Part Segmentation. We adopt the Oracle-Obj setting proposed by OV-PARTS [70] to conduct experiments on ADE-Part234 and Pascal-Part-116 datasets. This setting assumes that the ground-truth masks and categories of object-level instances are known during the inference process, aiming to evaluate the capability of the model to parse any novel object. All categories in the datasets are divided into a base set and a novel set," + }, + { + "type": "list", + "bbox": [ + 0.214, + 0.478, + 0.788, + 0.841 + ], + "angle": 0, + "content": null + } + ], + [ + { + "type": "page_number", + "bbox": [ + 0.218, + 0.116, + 0.236, + 0.127 + ], + "angle": 0, + "content": "12" + }, + { + "type": "header", + "bbox": [ + 0.272, + 0.115, + 0.395, + 0.128 + ], + "angle": 0, + "content": "J. Li, J. Wu, et al." + }, + { + "type": "table_caption", + "bbox": [ + 0.217, + 0.145, + 0.785, + 0.174 + ], + "angle": 0, + "content": "Table 3: Generalized zero-Shot part segmentation performance on ADE-Part-234 and Pascal-Part-116 compared with baselines proposed by OV-PARTS." + }, + { + "type": "table", + "bbox": [ + 0.226, + 0.198, + 0.781, + 0.338 + ], + "angle": 0, + "content": "
MethodModelBackboneFinetuningOracle-Obj
ADE-Part-234Oracle-obj
SeenUnseenHarmonicSeenUnseenHarmonic
Fully SupervisedMask2FormerResNet-50X46.2547.86-55.2852.14-
Two-StageZSeg+ResNet-50CPTCoOp43.1927.8433.8555.3319.1728.48
ResNet-50CPTCoCoOp39.6725.1530.7854.4319.0428.21
ResNet-101cCPTCoOp43.4125.7032.2857.8821.9331.81
One-StageCATSegResNet-101&ViT-B/16X11.498.569.8114.8910.2912.17
ResNet-101&ViT-B/16B+D31.4025.7728.3143.9726.1132.76
CLIPSegViT-B/16X15.2718.0116.5322.3319.7320.95
ViT-B/16VA+L+F+D38.9629.6533.6748.6827.3735.04
PartGLEEResNet-50X51.2935.3341.8357.4327.4137.11
" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.385, + 0.785, + 0.492 + ], + "angle": 0, + "content": "and the training process is performed only on the base set, while we evaluate the performance of the model on all categories. As shown in Tab. 3, our model achieves a superior performance on both datasets, which indicates the importance of hierarchical modeling. The establishment of hierarchical relationships between objects and parts enables our model to extend to novel objects, thereby effectively parsing them into corresponding semantic parts. Consequently, our model exhibits outstanding performance across both datasets." + }, + { + "type": "title", + "bbox": [ + 0.216, + 0.528, + 0.7, + 0.543 + ], + "angle": 0, + "content": "4.3 Joint-training results on Detection and Segmentation" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.569, + 0.787, + 0.841 + ], + "angle": 0, + "content": "To endow our model with robust generalization capability, we perform joint training on various datasets and evaluate its performance on both object-level and part-level tasks. We compare our model with specialist and generalist models to evaluate its performance on object-level data. Additionally, we contrast it with VLPart to assess its performance on part-level datasets as well as the effectiveness of joint-training process on both types of datasets. As shown in Tab. 4, PartGLEE significantly outperforms VLPart on both object-level and part-level tasks after joint-training, while achieving comparable performance on object-level tasks compared with previous SOTA. Through joint-training, our model has acquired strong generalization performance, allowing it to simultaneously address tasks for different hierarchies. We also observe that VLPart fails to achieve satisfactory performance on both object-level and part-level tasks. For example, VLPart obtains better performance on Pascal Part than its dataset-specific oracle, while decreasing its performance on COCO and LVIS. We attribute the performance drop of VLPart to the absence of hierarchical relationships, which causes confusion in modeling parts and objects and impairs object-level performance. PartGLEE effectively addresses this problem and extends the generalization capabilities from object-level to part-level tasks." + } + ], + [ + { + "type": "header", + "bbox": [ + 0.658, + 0.115, + 0.732, + 0.127 + ], + "angle": 0, + "content": "PartGLEE" + }, + { + "type": "page_number", + "bbox": [ + 0.769, + 0.116, + 0.786, + 0.127 + ], + "angle": 0, + "content": "13" + }, + { + "type": "table_caption", + "bbox": [ + 0.217, + 0.145, + 0.788, + 0.201 + ], + "angle": 0, + "content": "Table 4: Joint-Training Performance of PartGLEE. Note that Oracle represents the dataset-specific training paradigm. We directly evaluate the generalist models on PACO to assess their recognition capability at the part level, as indicated by the results annotated in the grey font." + }, + { + "type": "table", + "bbox": [ + 0.223, + 0.226, + 0.785, + 0.443 + ], + "angle": 0, + "content": "
TypeMethodPart-level TasksObject-level Tasks
PartImageNetPascal PartPACOCOCO-valLVIS-minimalLVIS-val
APboxAPmaskAPboxAPmaskAPmaskAPobjmaskAPpartmaskAPboxAPmaskAPboxAPmaskAPbox
SpecialistMask2Former(R50) [7]------46.243.7----
Mask2Former(L) [7]------50.1-----
MaskDINO(R50) [27]------50.546.0----
MaskDINO(L) [27]------58.352.1----
ViTDet-L [32]------57.649.8--51.246.0
ViTDet-H [32]------57.649.8--53.448.1
EVA-02-L [12]------64.255.0--65.257.3
PACO(R50) [55]----32.612.5------
PACO(L) [55]----43.417.7------
GeneralistPix2Seq v2 [5]------46.538.2----
X-Decoder(L) [89]---2.6911.90.94-46.7----
SEEM(L) [90]---1.998.420.69-47.7----
HIPIE(R50) [68]------53.945.9----
Florence-2(B) [72]------41.4-----
Florence-2(L) [72]------43.4-----
UNINEXT(R50) [74]------51.344.9--36.4-
UNINEXT(L) [74]------58.149.6----
GLEE(R50) [71]---3.4415.31.2955.048.450.545.944.240.2
HierarchicalVLPart(R50) [62]30.731.623.924.013.836.99.628.5--26.2-
VLPart(R50)-Oracle [62]29.229.718.919.413.328.010.638.0--28.1-
VLPart(B) [62]43.941.233.531.722.155.015.940.3--39.6-
VLPart(B)-Oracle [62]44.341.729.227.419.137.715.252.5--43.1-
PartGLEE (R50)40.940.235.035.521.850.515.454.447.648.743.542.7
PartGLEE (L)52.750.939.639.127.855.721.359.552.056.550.645.0
" + }, + { + "type": "title", + "bbox": [ + 0.217, + 0.486, + 0.389, + 0.501 + ], + "angle": 0, + "content": "4.4 Ablation Study" + }, + { + "type": "text", + "bbox": [ + 0.218, + 0.515, + 0.788, + 0.741 + ], + "angle": 0, + "content": "To demonstrate that our model design achieves satisfactory results on both object-level and part-level tasks, we conduct an ablation study (depicted in Fig. 4) on the model architecture and present results in Tab. 5. We ablate with a backbone of ResNet-50 and perform joint-training on COCO [37], LVIS [16], PartImageNet [18], Pascal Part [6] and PACO [55] with 90K iterations. From this study, we draw several important conclusions: (1) The utilization of parallel pixel decoders only results in slight improvements in mask predictions on few datasets, indicating that the influence of feature maps at different granularities is negligible. (2) Adopting independent decoders to obtain predictions at different levels demonstrates superior performance across the majority of datasets, manifesting the effectiveness of independent decoders. As adopting parallel pixel decoders (b) results in significant GPU memory costs without considerable gains, and all metrics for (a) are lower than (c), we select (c) as our final model design. Additional ablation studies, extensive qualitative analysis, and experiments on mLLM can be found in the supplementary materials." + }, + { + "type": "title", + "bbox": [ + 0.217, + 0.767, + 0.357, + 0.781 + ], + "angle": 0, + "content": "4.5 Limitations" + }, + { + "type": "text", + "bbox": [ + 0.217, + 0.795, + 0.787, + 0.84 + ], + "angle": 0, + "content": "In this work, we still adopt CLIP as the text encoder, which is trained on text-image pairs and thus lacks the ability to perceive fine-grained descriptions of object or part instances. This limitation may restrict the improvement of model" + } + ], + [ + { + "type": "page_number", + "bbox": [ + 0.218, + 0.116, + 0.236, + 0.127 + ], + "angle": 0, + "content": "14" + }, + { + "type": "header", + "bbox": [ + 0.272, + 0.115, + 0.396, + 0.128 + ], + "angle": 0, + "content": "J. Li, J. Wu, et al." + }, + { + "type": "table_caption", + "bbox": [ + 0.216, + 0.145, + 0.788, + 0.215 + ], + "angle": 0, + "content": "Table 5: An ablation study on different model designs, as depicted in Fig 4. Note that Parallel Pixel Decoders refers to the utilization of two pixel decoders to generate feature maps at different hierarchies respectively. Independent Decoders denote the usage of two decoders, which facilitate the interaction between feature maps and queries at different hierarchies. Our final choice is scheme (c), which is highlighted in gray." + }, + { + "type": "table", + "bbox": [ + 0.223, + 0.239, + 0.785, + 0.313 + ], + "angle": 0, + "content": "
SchemeModel DesignPart-level TasksObject-level Tasks
Parrallel Pixel DecodersIndependent DecodersPartImageNetPascal PartPACOCOCO-valLVIS-minimal
APmaskAPmaskAPmaskAPmaskAPobjmaskAPopartmaskAPboxAPmaskAPboxAPmask
(a)XX39.034.120.147.413.547.843.534.833.4
(b)✓X38.334.520.848.813.848.544.334.934.2
(c)X✓39.034.720.947.914.249.344.235.633.8
" + }, + { + "type": "image_caption", + "bbox": [ + 0.223, + 0.34, + 0.239, + 0.35 + ], + "angle": 0, + "content": "(a)" + }, + { + "type": "image", + "bbox": [ + 0.224, + 0.355, + 0.404, + 0.446 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.412, + 0.339, + 0.428, + 0.349 + ], + "angle": 0, + "content": "(b)" + }, + { + "type": "image", + "bbox": [ + 0.412, + 0.351, + 0.591, + 0.451 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.597, + 0.34, + 0.613, + 0.35 + ], + "angle": 0, + "content": "(c)" + }, + { + "type": "image", + "bbox": [ + 0.597, + 0.355, + 0.776, + 0.444 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.214, + 0.474, + 0.788, + 0.558 + ], + "angle": 0, + "content": "Fig. 4: Various designs for generating predictions at different hierarchies. In scheme (a), we only utilize a single decoder to generate predictions for both objects and parts. In scheme (b), two parallel pixel decoders are employed to generate feature maps at different levels, aiming to explore the effectiveness of feature maps at different granularity. In scheme (c), we use two independent decoders to generate predictions for objects and parts respectively." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.591, + 0.787, + 0.621 + ], + "angle": 0, + "content": "performance and prompts us to consider how to enhance the perception capabilities of region-level models, which will be our future work." + }, + { + "type": "title", + "bbox": [ + 0.216, + 0.651, + 0.36, + 0.667 + ], + "angle": 0, + "content": "5 Conclusion" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.689, + 0.789, + 0.841 + ], + "angle": 0, + "content": "In this paper, we introduce PartGLEE, a groundbreaking foundation model designed towards a complete comprehension of both objects and parts in images. Through the generic hierarchical relationships established by the Q-Former, we are able to break through the limitation of scarce part-level data by introducing a large amount of object-level data, thereby transferring the powerful generalization capabilities from objects to parts. Through extensive training on diverse datasets, PartGLEE achieves SOTA performance across various part-level tasks while maintaining competitive results on object-level tasks, enabling it to parse any objects into parts and serve as a foundation model for general fine-grained region-level perception tasks." + } + ], + [ + { + "type": "header", + "bbox": [ + 0.658, + 0.115, + 0.732, + 0.127 + ], + "angle": 0, + "content": "PartGLEE" + }, + { + "type": "page_number", + "bbox": [ + 0.769, + 0.117, + 0.786, + 0.127 + ], + "angle": 0, + "content": "15" + }, + { + "type": "title", + "bbox": [ + 0.218, + 0.145, + 0.403, + 0.163 + ], + "angle": 0, + "content": "Acknowledgements" + }, + { + "type": "text", + "bbox": [ + 0.217, + 0.174, + 0.785, + 0.205 + ], + "angle": 0, + "content": "This work was supported by the National Science Fund for Distinguished Young Scholars of China (Grant No.62225603)." + }, + { + "type": "title", + "bbox": [ + 0.218, + 0.225, + 0.323, + 0.241 + ], + "angle": 0, + "content": "References" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.253, + 0.786, + 0.308 + ], + "angle": 0, + "content": "1. Alayrac, J.B., Donahue, J., Luc, P., Miech, A., Barr, I., Hasson, Y., Lenc, K., Mensch, A., Millican, K., Reynolds, M., et al.: Flamingo: a visual language model for few-shot learning. In: Advances in neural information processing systems. vol. 35, pp. 23716-23736 (2022)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.309, + 0.786, + 0.349 + ], + "angle": 0, + "content": "2. Brohan, A., Brown, N., Carbajal, J., Chebotar, Y., Dabis, J., Finn, C., Gopalakrishnan, K., Hausman, K., Herzog, A., Hsu, J., et al.: Rt-1: Robotics transformer for real-world control at scale. arXiv preprint arXiv:2212.06817 (2022)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.35, + 0.786, + 0.404 + ], + "angle": 0, + "content": "3. Brown, T., Mann, B., Ryder, N., Subbiah, M., Kaplan, J.D., Dhariwal, P., Neelakantan, A., Shyam, P., Sastry, G., Askell, A., et al.: Language models are few-shot learners. In: Advances in neural information processing systems. vol. 33, pp. 1877-1901 (2020)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.405, + 0.786, + 0.445 + ], + "angle": 0, + "content": "4. Caron, M., Touvron, H., Misra, I., Jégou, H., Mairal, J., Bojanowski, P., Joulin, A.: Emerging properties in self-supervised vision transformers. In: Proceedings of the IEEE/CVF international conference on computer vision. pp. 9650-9660 (2021)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.445, + 0.786, + 0.485 + ], + "angle": 0, + "content": "5. Chen, T., Saxena, S., Li, L., Lin, T.Y., Fleet, D.J., Hinton, G.E.: A unified sequence interface for vision tasks. Advances in Neural Information Processing Systems 35, 31333-31346 (2022)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.486, + 0.786, + 0.54 + ], + "angle": 0, + "content": "6. Chen, X., Mottaghi, R., Liu, X., Fidler, S., Urtasun, R., Yuille, A.: Detect what you can: Detecting and representing objects using holistic models and body parts. In: Proceedings of the IEEE conference on computer vision and pattern recognition. pp. 1971-1978 (2014)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.541, + 0.786, + 0.595 + ], + "angle": 0, + "content": "7. Cheng, B., Misra, I., Schwing, A.G., Kirillov, A., Girdhar, R.: Masked-attention mask transformer for universal image segmentation. In: Proceedings of the IEEE/CVF conference on computer vision and pattern recognition. pp. 1290-1299 (2022)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.596, + 0.786, + 0.636 + ], + "angle": 0, + "content": "8. Deng, J., Dong, W., Socher, R., Li, L.J., Li, K., Fei-Fei, L.: Imagenet: A large-scale hierarchical image database. In: Proceedings of the IEEE/CVF International Conference on Computer Vision. pp. 248-255 (2009)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.636, + 0.786, + 0.676 + ], + "angle": 0, + "content": "9. Devlin, J., Chang, M.W., Lee, K., Toutanova, K.: Bert: Pre-training of deep bidirectional transformers for language understanding. arXiv preprint arXiv:1810.04805 (2018)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.677, + 0.786, + 0.717 + ], + "angle": 0, + "content": "0. Dong, J., Chen, Q., Shen, X., Yang, J., Yan, S.: Towards unified human parsing and pose estimation. In: Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition. pp. 843-850 (2014)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.718, + 0.786, + 0.759 + ], + "angle": 0, + "content": "1. Esser, P., Rombach, R., Ommer, B.: Taming transformers for high-resolution image synthesis. In: Proceedings of the IEEE/CVF conference on computer vision and pattern recognition. pp. 12873-12883 (2021)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.759, + 0.786, + 0.785 + ], + "angle": 0, + "content": "2. Fang, Y., Sun, Q., Wang, X., Huang, T., Wang, X., Cao, Y.: Eva-02: A visual representation for neon genesis. arXiv preprint arXiv:2303.11331 (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.786, + 0.786, + 0.84 + ], + "angle": 0, + "content": "3. Fang, Y., Wang, W., Xie, B., Sun, Q., Wu, L., Wang, X., Huang, T., Wang, X., Cao, Y.: Eva: Exploring the limits of masked visual representation learning at scale. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 19358-19369 (2023)" + }, + { + "type": "list", + "bbox": [ + 0.226, + 0.253, + 0.786, + 0.84 + ], + "angle": 0, + "content": null + } + ], + [ + { + "type": "page_number", + "bbox": [ + 0.218, + 0.116, + 0.236, + 0.127 + ], + "angle": 0, + "content": "16" + }, + { + "type": "header", + "bbox": [ + 0.272, + 0.115, + 0.395, + 0.128 + ], + "angle": 0, + "content": "J. Li, J. Wu, et al." + }, + { + "type": "ref_text", + "bbox": [ + 0.218, + 0.147, + 0.785, + 0.189 + ], + "angle": 0, + "content": "14. de Geus, D., Meletis, P., Lu, C., Wen, X., Dubbelman, G.: Part-aware panoptic segmentation. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 5485-5494 (2021)" + }, + { + "type": "ref_text", + "bbox": [ + 0.218, + 0.19, + 0.787, + 0.245 + ], + "angle": 0, + "content": "15. Gong, K., Liang, X., Zhang, D., Shen, X., Lin, L.: Look into person: Self-supervised structure-sensitive learning and a new benchmark for human parsing. In: Proceedings of the IEEE conference on computer vision and pattern recognition. pp. 932-940 (2017)" + }, + { + "type": "ref_text", + "bbox": [ + 0.218, + 0.246, + 0.785, + 0.286 + ], + "angle": 0, + "content": "16. Gupta, A., Dollar, P., Girshick, R.: Lvis: A dataset for large vocabulary instance segmentation. In: Proceedings of the IEEE/CVF conference on computer vision and pattern recognition. pp. 5356-5364 (2019)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.287, + 0.785, + 0.342 + ], + "angle": 0, + "content": "17. He, J., Chen, J., Lin, M.X., Yu, Q., Yuille, A.L.: Compositor: Bottom-up clustering and compositing for robust part and object segmentation. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 11259-11268 (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.343, + 0.785, + 0.384 + ], + "angle": 0, + "content": "18. He, J., Yang, S., Yang, S., Kortylewski, A., Yuan, X., Chen, J.N., Liu, S., Yang, C., Yu, Q., Yuille, A.: Partimagenet: A large, high-quality dataset of parts. In: European Conference on Computer Vision. pp. 128-145. Springer (2022)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.384, + 0.785, + 0.425 + ], + "angle": 0, + "content": "19. He, K., Chen, X., Xie, S., Li, Y., Dólár, P., Girshick, R.: Masked autoencoders are scalable vision learners. In: Proceedings of the IEEE/CVF conference on computer vision and pattern recognition. pp. 16000-16009 (2022)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.425, + 0.785, + 0.466 + ], + "angle": 0, + "content": "20. He, K., Zhang, X., Ren, S., Sun, J.: Deep residual learning for image recognition. In: Proceedings of the IEEE conference on computer vision and pattern recognition. pp. 770-778 (2016)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.467, + 0.785, + 0.522 + ], + "angle": 0, + "content": "21. Jia, C., Yang, Y., Xia, Y., Chen, Y.T., Parekh, Z., Pham, H., Le, Q., Sung, Y.H., Li, Z., Duerig, T.: Scaling up visual and vision-language representation learning with noisy text supervision. In: International conference on machine learning. pp. 4904-4916. PMLR (2021)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.523, + 0.785, + 0.577 + ], + "angle": 0, + "content": "22. Kawar, B., Zada, S., Lang, O., Tov, O., Chang, H., Dekel, T., Mosseri, I., Irani, M.: Imagic: Text-based real image editing with diffusion models. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 6007-6017 (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.578, + 0.785, + 0.632 + ], + "angle": 0, + "content": "23. Kirillov, A., Mintun, E., Ravi, N., Mao, H., Rolland, C., Gustafson, L., Xiao, T., Whitehead, S., Berg, A.C., Lo, W.Y., et al.: Segment anything. In: Proceedings of the IEEE/CVF International Conference on Computer Vision. pp. 4015-4026 (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.633, + 0.785, + 0.688 + ], + "angle": 0, + "content": "24. Krishna, R., Zhu, Y., Groth, O., Johnson, J., Hata, K., Kravitz, J., Chen, S., Kalantidis, Y., Li, L.J., Shamma, D.A., et al.: Visual genome: Connecting language and vision using crowdsourced dense image annotations. International journal of computer vision 123, 32-73 (2017)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.689, + 0.785, + 0.743 + ], + "angle": 0, + "content": "25. Kuznetsova, A., Rom, H., Alldrin, N., Uijlings, J., Krasin, I., Pont-Tuset, J., Kamali, S., Popov, S., Malloci, M., Kolesnikov, A., et al.: The open images dataset v4: Unified image classification, object detection, and visual relationship detection at scale. International journal of computer vision 128(7), 1956-1981 (2020)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.744, + 0.785, + 0.785 + ], + "angle": 0, + "content": "26. Li, F., Zhang, H., Sun, P., Zou, X., Liu, S., Yang, J., Li, C., Zhang, L., Gao, J.: Semantic-sam: Segment and recognize anything at any granularity. arXiv preprint arXiv:2307.04767 (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.786, + 0.785, + 0.84 + ], + "angle": 0, + "content": "27. Li, F., Zhang, H., Xu, H., Liu, S., Zhang, L., Ni, L.M., Shum, H.Y.: Mask dino: Towards a unified transformer-based framework for object detection and segmentation. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 3041-3050 (2023)" + }, + { + "type": "list", + "bbox": [ + 0.217, + 0.147, + 0.787, + 0.84 + ], + "angle": 0, + "content": null + } + ], + [ + { + "type": "header", + "bbox": [ + 0.658, + 0.115, + 0.732, + 0.127 + ], + "angle": 0, + "content": "PartGLEE" + }, + { + "type": "page_number", + "bbox": [ + 0.769, + 0.116, + 0.786, + 0.127 + ], + "angle": 0, + "content": "17" + }, + { + "type": "ref_text", + "bbox": [ + 0.214, + 0.147, + 0.788, + 0.204 + ], + "angle": 0, + "content": "28. Li, H., Zhu, J., Jiang, X., Zhu, X., Li, H., Yuan, C., Wang, X., Qiao, Y., Wang, X., Wang, W., et al.: Uni-perceiver v2: A generalist model for large-scale vision and vision-language tasks. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 2691-2700 (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.214, + 0.205, + 0.787, + 0.232 + ], + "angle": 0, + "content": "29. Li, J., Zhao, J., Wei, Y., Lang, C., Li, Y., Sim, T., Yan, S., Feng, J.: Multiple-human parsing in the wild. arXiv preprint arXiv:1705.07206 (2017)" + }, + { + "type": "ref_text", + "bbox": [ + 0.216, + 0.233, + 0.786, + 0.286 + ], + "angle": 0, + "content": "30. Li, L.H., Zhang, P., Zhang, H., Yang, J., Li, C., Zhong, Y., Wang, L., Yuan, L., Zhang, L., Hwang, J.N., et al.: Grounded language-image pre-training. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 10965-10975 (2022)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.288, + 0.787, + 0.328 + ], + "angle": 0, + "content": "31. Li, X., Xu, S., Yang, Y., Cheng, G., Tong, Y., Tao, D.: Panoptic-partformer: Learning a unified model for panoptic part segmentation. In: European Conference on Computer Vision. pp. 729-747. Springer (2022)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.328, + 0.786, + 0.37 + ], + "angle": 0, + "content": "32. Li, Y., Mao, H., Girshick, R., He, K.: Exploring plain vision transformer backbones for object detection. In: European Conference on Computer Vision. pp. 280-296. Springer (2022)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.371, + 0.786, + 0.412 + ], + "angle": 0, + "content": "33. Li, Y., Singh, K.K., Xue, Y., Lee, Y.J.: Partgan: Weakly-supervised part decomposition for image generation and segmentation. In: British Machine Vision Conference (BMVC) (2021)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.412, + 0.787, + 0.453 + ], + "angle": 0, + "content": "34. Lin, C., Jiang, Y., Qu, L., Yuan, Z., Cai, J.: Generative region-language pretraining for open-ended object detection. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 13958-13968 (2024)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.453, + 0.787, + 0.495 + ], + "angle": 0, + "content": "35. Lin, C., Sun, P., Jiang, Y., Luo, P., Qu, L., Haffari, G., Yuan, Z., Cai, J.: Learning object-language alignments for open-vocabulary object detection. In: The Eleventh International Conference on Learning Representations (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.495, + 0.787, + 0.536 + ], + "angle": 0, + "content": "36. Lin, T.Y., Goyal, P., Girshick, R., He, K., Dollar, P.: Focal loss for dense object detection. In: Proceedings of the IEEE international conference on computer vision. pp. 2980-2988 (2017)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.536, + 0.787, + 0.577 + ], + "angle": 0, + "content": "37. Lin, T.Y., Maire, M., Belongie, S., Hays, J., Perona, P., Ramanan, D., Dollar, P., Zitnick, C.L.: Microsoft coco: Common objects in context. In: European Conference on Computer Vision. pp. 740-755. Springer (2014)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.578, + 0.787, + 0.619 + ], + "angle": 0, + "content": "38. Ling, H., Kreis, K., Li, D., Kim, S.W., Torralba, A., Fidler, S.: Editgan: High-precision semantic image editing. In: Advances in Neural Information Processing Systems. vol. 34, pp. 16331-16345 (2021)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.62, + 0.787, + 0.66 + ], + "angle": 0, + "content": "39. Liu, S., Zeng, Z., Ren, T., Li, F., Zhang, H., Yang, J., Li, C., Yang, J., Su, H., Zhu, J., et al.: Grounding dino: Marrying dino with grounded pre-training for open-set object detection. arXiv preprint arXiv:2303.05499 (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.66, + 0.787, + 0.715 + ], + "angle": 0, + "content": "40. Liu, Z., Lin, Y., Cao, Y., Hu, H., Wei, Y., Zhang, Z., Lin, S., Guo, B.: Swin transformer: Hierarchical vision transformer using shifted windows. In: Proceedings of the IEEE/CVF international conference on computer vision. pp. 10012-10022 (2021)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.716, + 0.787, + 0.758 + ], + "angle": 0, + "content": "41. Lu, J., Clark, C., Zellers, R., Mottaghi, R., Kembhavi, A.: Unified-io: A unified model for vision, language, and multi-modal tasks. In: The Eleventh International Conference on Learning Representations (2022)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.759, + 0.787, + 0.799 + ], + "angle": 0, + "content": "42. Ma, C., Jiang, Y., Wen, X., Yuan, Z., Qi, X.: Codet: Co-occurrence guided region-word alignment for open-vocabulary object detection. In: Advances in neural information processing systems. vol. 36 (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.799, + 0.787, + 0.841 + ], + "angle": 0, + "content": "43. Ma, C., Jiang, Y., Wu, J., Yuan, Z., Qi, X.: Groma: Localized visual tokenization for grounding multimodal large language models. arXiv preprint arXiv:2404.13013 (2024)" + }, + { + "type": "list", + "bbox": [ + 0.214, + 0.147, + 0.788, + 0.841 + ], + "angle": 0, + "content": null + } + ], + [ + { + "type": "page_number", + "bbox": [ + 0.218, + 0.116, + 0.236, + 0.127 + ], + "angle": 0, + "content": "18" + }, + { + "type": "header", + "bbox": [ + 0.272, + 0.115, + 0.395, + 0.128 + ], + "angle": 0, + "content": "J. Li, J. Wu, et al." + }, + { + "type": "ref_text", + "bbox": [ + 0.218, + 0.148, + 0.785, + 0.189 + ], + "angle": 0, + "content": "44. Mao, J., Huang, J., Toshev, A., Camburu, O., Yuille, A.L., Murphy, K.: Generation and comprehension of unambiguous object descriptions. In: Proceedings of the IEEE conference on computer vision and pattern recognition. pp. 11-20 (2016)" + }, + { + "type": "ref_text", + "bbox": [ + 0.218, + 0.189, + 0.785, + 0.229 + ], + "angle": 0, + "content": "45. Meletis, P., Wen, X., Lu, C., de Geus, D., Dubbelman, G.: Cityscapes-panoptic-parts and pascal-panoptic-parts datasets for scene understanding. arXiv preprint arXiv:2004.07944 (2020)" + }, + { + "type": "ref_text", + "bbox": [ + 0.218, + 0.229, + 0.785, + 0.271 + ], + "angle": 0, + "content": "46. Michieli, U., Borsato, E., Rossi, L., Zanuttigh, P.: Gmnet: Graph matching network for large scale part semantic segmentation in the wild. In: European Conference on Computer Vision. pp. 397-414. Springer (2020)" + }, + { + "type": "ref_text", + "bbox": [ + 0.218, + 0.271, + 0.785, + 0.311 + ], + "angle": 0, + "content": "47. Miletari, F., Navab, N., Ahmadi, S.A.: V-net: Fully convolutional neural networks for volumetric medical image segmentation. In: 2016 fourth international conference on 3D vision (3DV) (2016)" + }, + { + "type": "ref_text", + "bbox": [ + 0.218, + 0.311, + 0.785, + 0.338 + ], + "angle": 0, + "content": "48. Morabia, K., Arora, J., Vijaykumar, T.: Attention-based joint detection of object and semantic part. arXiv preprint arXiv:2007.02419 (2020)" + }, + { + "type": "ref_text", + "bbox": [ + 0.218, + 0.338, + 0.785, + 0.378 + ], + "angle": 0, + "content": "49. Nair, S., Rajeswaran, A., Kumar, V., Finn, C., Gupta, A.: R3m: A universal visual representation for robot manipulation. In: Conference on Robot Learning. pp. 892-909. PMLR (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.378, + 0.785, + 0.433 + ], + "angle": 0, + "content": "50. Ng, X.L., Ong, K.E., Zheng, Q., Ni, Y., Yeo, S.Y., Liu, J.: Animal kingdom: A large and diverse dataset for animal behavior understanding. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 19023-19034 (2022)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.433, + 0.785, + 0.473 + ], + "angle": 0, + "content": "51. Pan, T.Y., Liu, Q., Chao, W.L., Price, B.: Towards open-world segmentation of parts. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 15392-15401 (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.473, + 0.785, + 0.514 + ], + "angle": 0, + "content": "52. Qi, L., Kuen, J., Guo, W., Gu, J., Lin, Z., Du, B., Xu, Y., Yang, M.H.: Aims: All-inclusive multi-level segmentation for anything. In: Advances in Neural Information Processing Systems. vol. 36 (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.514, + 0.785, + 0.568 + ], + "angle": 0, + "content": "53. Radford, A., Kim, J.W., Hallacy, C., Ramesh, A., Goh, G., Agarwal, S., Sastry, G., Askell, A., Mishkin, P., Clark, J., et al.: Learning transferable visual models from natural language supervision. In: International conference on machine learning. pp. 8748-8763. PMLR (2021)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.568, + 0.785, + 0.609 + ], + "angle": 0, + "content": "54. Raffel, C., Shazeer, N., Roberts, A., Lee, K., Narang, S., Matena, M., Zhou, Y., Li, W., Liu, P.J.: Exploring the limits of transfer learning with a unified text-to-text transformer. The Journal of Machine Learning Research 21(1), 5485-5551 (2020)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.609, + 0.785, + 0.663 + ], + "angle": 0, + "content": "55. Ramanathan, V., Kalia, A., Petrovic, V., Wen, Y., Zheng, B., Guo, B., Wang, R., Marquez, A., Kovvuri, R., Kadian, A., et al.: Paco: Parts and attributes of common objects. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 7141-7151 (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.663, + 0.785, + 0.704 + ], + "angle": 0, + "content": "56. Ramesh, A., Dhariwal, P., Nichol, A., Chu, C., Chen, M.: Hierarchical text-conditional image generation with clip latents. arXiv preprint arXiv:2204.06125 1(2), 3 (2022)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.704, + 0.785, + 0.744 + ], + "angle": 0, + "content": "57. Ramesh, A., Pavlov, M., Goh, G., Gray, S., Voss, C., Radford, A., Chen, M., Sutskever, I.: Zero-shot text-to-image generation. In: International Conference on Machine Learning. pp. 8821-8831. PMLR (2021)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.744, + 0.785, + 0.786 + ], + "angle": 0, + "content": "58. Reddy, N.D., Vo, M., Narasimhan, S.G.: Carfusion: Combining point tracking and part detection for dynamic 3d reconstruction of vehicles. In: Proceedings of the IEEE conference on computer vision and pattern recognition. pp. 1906-1915 (2018)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.786, + 0.785, + 0.84 + ], + "angle": 0, + "content": "59. Rezatofighi, H., Tsoi, N., Gwak, J., Sadeghian, A., Reid, I., Savarese, S.: Generalized intersection over union: A metric and a loss for bounding box regression. In: Proceedings of the IEEE/CVF conference on computer vision and pattern recognition. pp. 658-666 (2019)" + }, + { + "type": "list", + "bbox": [ + 0.217, + 0.148, + 0.785, + 0.84 + ], + "angle": 0, + "content": null + } + ], + [ + { + "type": "header", + "bbox": [ + 0.658, + 0.115, + 0.732, + 0.127 + ], + "angle": 0, + "content": "PartGLEE" + }, + { + "type": "page_number", + "bbox": [ + 0.769, + 0.116, + 0.786, + 0.127 + ], + "angle": 0, + "content": "19" + }, + { + "type": "ref_text", + "bbox": [ + 0.214, + 0.147, + 0.788, + 0.189 + ], + "angle": 0, + "content": "60. Rombach, R., Blattmann, A., Lorenz, D., Esser, P., Ommer, B.: High-resolution image synthesis with latent diffusion models. In: Proceedings of the IEEE/CVF conference on computer vision and pattern recognition. pp. 10684-10695 (2022)" + }, + { + "type": "ref_text", + "bbox": [ + 0.214, + 0.189, + 0.788, + 0.231 + ], + "angle": 0, + "content": "61. Shao, S., Li, Z., Zhang, T., Peng, C., Yu, G., Zhang, X., Li, J., Sun, J.: Objects365: A large-scale, high-quality dataset for object detection. In: Proceedings of the IEEE/CVF international conference on computer vision. pp. 8430-8439 (2019)" + }, + { + "type": "ref_text", + "bbox": [ + 0.214, + 0.231, + 0.788, + 0.272 + ], + "angle": 0, + "content": "62. Sun, P., Chen, S., Zhu, C., Xiao, F., Luo, P., Xie, S., Yan, Z.: Going denser with open-vocabulary part segmentation. In: Proceedings of the IEEE/CVF International Conference on Computer Vision. pp. 15453-15465 (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.214, + 0.272, + 0.788, + 0.313 + ], + "angle": 0, + "content": "63. Tang, C., Xie, L., Zhang, X., Hu, X., Tian, Q.: Visual recognition by request. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 15265-15274 (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.214, + 0.314, + 0.788, + 0.355 + ], + "angle": 0, + "content": "64. Wah, C., Branson, S., Welinder, P., Perona, P., Belongie, S.: Caltech-ucsd birds200-2011 (cub-200-2011). Tech. Rep. CNS-TR-2011-001, California Institute of Technology (2011)" + }, + { + "type": "ref_text", + "bbox": [ + 0.214, + 0.356, + 0.788, + 0.397 + ], + "angle": 0, + "content": "65. Wang, P., Shen, X., Lin, Z., Cohen, S., Price, B., Yuille, A.L.: Joint object and part segmentation using deep learned potentials. In: Proceedings of the IEEE International Conference on Computer Vision. pp. 1573-1581 (2015)" + }, + { + "type": "ref_text", + "bbox": [ + 0.214, + 0.397, + 0.788, + 0.453 + ], + "angle": 0, + "content": "66. Wang, P., Yang, A., Men, R., Lin, J., Bai, S., Li, Z., Ma, J., Zhou, C., Zhou, J., Yang, H.: Ofa: Unifying architectures, tasks, and modalities through a simple sequence-to-sequence learning framework. In: International Conference on Machine Learning. pp. 23318-23340. PMLR (2022)" + }, + { + "type": "ref_text", + "bbox": [ + 0.214, + 0.453, + 0.788, + 0.508 + ], + "angle": 0, + "content": "67. Wang, W., Bao, H., Dong, L., Bjorck, J., Peng, Z., Liu, Q., Aggarwal, K., Mohammed, O.K., Singhal, S., Som, S., et al.: Image as a foreign language: Beit pretraining for all vision and vision-language tasks. arXiv preprint arXiv:2208.10442 (2022)" + }, + { + "type": "ref_text", + "bbox": [ + 0.214, + 0.508, + 0.788, + 0.55 + ], + "angle": 0, + "content": "68. Wang, X., Li, S., Kallidromitis, K., Kato, Y., Kozuka, K., Darrell, T.: Hierarchical open-vocabulary universal image segmentation. In: Advances in Neural Information Processing Systems. vol. 36 (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.214, + 0.55, + 0.788, + 0.591 + ], + "angle": 0, + "content": "69. Wang, Y., Xu, Y., Tsogkas, S., Bai, X., Dickinson, S., Siddiqi, K.: Deepflux for skeletons in the wild. In: Proceedings of the IEEE/CVF conference on computer vision and pattern recognition. pp. 5287-5296 (2019)" + }, + { + "type": "ref_text", + "bbox": [ + 0.214, + 0.591, + 0.788, + 0.633 + ], + "angle": 0, + "content": "70. Wei, M., Yue, X., Zhang, W., Kong, S., Liu, X., Pang, J.: Ov-parts: Towards open-vocabulary part segmentation. In: Thirty-seventh Conference on Neural Information Processing Systems Datasets and Benchmarks Track (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.214, + 0.633, + 0.788, + 0.674 + ], + "angle": 0, + "content": "71. Wu, J., Jiang, Y., Liu, Q., Yuan, Z., Bai, X., Bai, S.: General object foundation model for images and videos at scale. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 3783-3795 (2024)" + }, + { + "type": "ref_text", + "bbox": [ + 0.214, + 0.674, + 0.788, + 0.73 + ], + "angle": 0, + "content": "72. Xiao, B., Wu, H., Xu, W., Dai, X., Hu, H., Lu, Y., Zeng, M., Liu, C., Yuan, L.: Florence-2: Advancing a unified representation for a variety of vision tasks. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 4818-4829 (2024)" + }, + { + "type": "ref_text", + "bbox": [ + 0.214, + 0.73, + 0.788, + 0.758 + ], + "angle": 0, + "content": "73. Yan, B., Jiang, Y., Sun, P., Wang, D., Yuan, Z., Luo, P., Lu, H.: Towards grand unification of object tracking. In: European Conference on Computer Vision (2022)" + }, + { + "type": "ref_text", + "bbox": [ + 0.214, + 0.758, + 0.788, + 0.799 + ], + "angle": 0, + "content": "74. Yan, B., Jiang, Y., Wu, J., Wang, D., Luo, P., Yuan, Z., Lu, H.: Universal instance perception as object discovery and retrieval. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 15325-15336 (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.214, + 0.799, + 0.788, + 0.841 + ], + "angle": 0, + "content": "75. Yang, L., Song, Q., Wang, Z., Jiang, M.: Parsing r-cnn for instance-level human analysis. In: Proceedings of the IEEE/CVF conference on computer vision and pattern recognition. pp. 364-373 (2019)" + }, + { + "type": "list", + "bbox": [ + 0.214, + 0.147, + 0.788, + 0.841 + ], + "angle": 0, + "content": null + } + ], + [ + { + "type": "page_number", + "bbox": [ + 0.218, + 0.116, + 0.236, + 0.127 + ], + "angle": 0, + "content": "20" + }, + { + "type": "header", + "bbox": [ + 0.272, + 0.115, + 0.395, + 0.128 + ], + "angle": 0, + "content": "J. Li, J. Wu, et al." + }, + { + "type": "ref_text", + "bbox": [ + 0.218, + 0.148, + 0.785, + 0.189 + ], + "angle": 0, + "content": "76. Yang, Y., Ramanan, D.: Articulated pose estimation with flexible mixtures-of-parts. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 1385-1392. IEEE (2011)" + }, + { + "type": "ref_text", + "bbox": [ + 0.218, + 0.19, + 0.785, + 0.23 + ], + "angle": 0, + "content": "77. Yang, Z., Gan, Z., Wang, J., Hu, X., Ahmed, F., Liu, Z., Lu, Y., Wang, L.: Unitab: Unifying text and box outputs for grounded vision-language modeling. In: European Conference on Computer Vision. pp. 521-539. Springer (2022)" + }, + { + "type": "ref_text", + "bbox": [ + 0.218, + 0.23, + 0.785, + 0.285 + ], + "angle": 0, + "content": "78. Yao, L., Han, J., Liang, X., Xu, D., Zhang, W., Li, Z., Xu, H.: Detclipv2: Scalable open-vocabulary object detection pre-training via word-region alignment. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 23497-23506 (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.285, + 0.785, + 0.338 + ], + "angle": 0, + "content": "79. Yao, L., Han, J., Wen, Y., Liang, X., Xu, D., Zhang, W., Li, Z., Xu, C., Xu, H.: Detclip: Dictionary-enriched visual-concept paralleled pre-training for open-world detection. In: Advances in Neural Information Processing Systems. vol. 35, pp. 9125-9138 (2022)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.339, + 0.785, + 0.379 + ], + "angle": 0, + "content": "80. Yu, L., Poirson, P., Yang, S., Berg, A.C., Berg, T.L.: Modeling context in referring expressions. In: European Conference on Computer Vision. pp. 69-85. Springer (2016)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.379, + 0.785, + 0.42 + ], + "angle": 0, + "content": "81. Yuan, L., Chen, D., Chen, Y.L., Codella, N., Dai, X., Gao, J., Hu, H., Huang, X., Li, B., Li, C., et al.: Florence: A new foundation model for computer vision. arXiv preprint arXiv:2111.11432 (2021)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.42, + 0.785, + 0.461 + ], + "angle": 0, + "content": "82. Zareian, A., Rosa, K.D., Hu, D.H., Chang, S.F.: Open-vocabulary object detection using captions. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 14393-14402 (2021)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.461, + 0.785, + 0.514 + ], + "angle": 0, + "content": "83. Zhong, Y., Yang, J., Zhang, P., Li, C., Codella, N., Li, L.H., Zhou, L., Dai, X., Yuan, L., Li, Y., et al.: Regionclip: Region-based language-image pretraining. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 16793-16803 (2022)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.514, + 0.785, + 0.555 + ], + "angle": 0, + "content": "84. Zhou, B., Zhao, H., Puig, X., Xiao, T., Fidler, S., Barriuso, A., Torralba, A.: Semantic understanding of scenes through the ade20k dataset. International Journal on Computer Vision (2018)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.555, + 0.785, + 0.61 + ], + "angle": 0, + "content": "85. Zhou, T., Wang, W., Liu, S., Yang, Y., Van Gool, L.: Differentiable multigranularity human representation learning for instance-aware human semantic parsing. In: Proceedings of the IEEE/CVF conference on computer vision and pattern recognition. pp. 1622-1631 (2021)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.61, + 0.785, + 0.651 + ], + "angle": 0, + "content": "86. Zhu, X., Su, W., Lu, L., Li, B., Wang, X., Dai, J.: Deformable detr: Deformable transformers for end-to-end object detection. In: International Conference on Learning Representations (2021)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.651, + 0.785, + 0.704 + ], + "angle": 0, + "content": "87. Zhu, X., Zhu, J., Li, H., Wu, X., Li, H., Wang, X., Dai, J.: Uni-perceiver: Pretraining unified architecture for generic perception for zero-shot and few-shot tasks. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 16804-16815 (2022)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.705, + 0.785, + 0.746 + ], + "angle": 0, + "content": "88. Ziegler, A., Asano, Y.M.: Self-supervised learning of object parts for semantic segmentation. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 14502-14511 (2022)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.746, + 0.785, + 0.8 + ], + "angle": 0, + "content": "89. Zou, X., Dou, Z.Y., Yang, J., Gan, Z., Li, L., Li, C., Dai, X., Behl, H., Wang, J., Yuan, L., et al.: Generalized decoding for pixel, image, and language. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 15116-15127 (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.8, + 0.785, + 0.84 + ], + "angle": 0, + "content": "90. Zou, X., Yang, J., Zhang, H., Li, F., Li, L., Wang, J., Wang, L., Gao, J., Lee, Y.J.: Segment everything everywhere all at once. In: Advances in Neural Information Processing Systems. vol. 36 (2023)" + }, + { + "type": "list", + "bbox": [ + 0.217, + 0.148, + 0.785, + 0.84 + ], + "angle": 0, + "content": null + } + ] +] \ No newline at end of file diff --git a/2024/PartGLEE_ A Foundation Model for Recognizing and Parsing Any Objects/a546ffc1-52e4-4141-be35-2478ed6c5802_origin.pdf b/2024/PartGLEE_ A Foundation Model for Recognizing and Parsing Any Objects/a546ffc1-52e4-4141-be35-2478ed6c5802_origin.pdf new file mode 100644 index 0000000000000000000000000000000000000000..c4e66faf4ae58dab34ad985061f618c894a0fe59 --- /dev/null +++ b/2024/PartGLEE_ A Foundation Model for Recognizing and Parsing Any Objects/a546ffc1-52e4-4141-be35-2478ed6c5802_origin.pdf @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:f3228856dd64f0129d59e6778a5979c32dcdc62395b498153212eaddbfe15f31 +size 4740678 diff --git a/2024/PartGLEE_ A Foundation Model for Recognizing and Parsing Any Objects/full.md b/2024/PartGLEE_ A Foundation Model for Recognizing and Parsing Any Objects/full.md new file mode 100644 index 0000000000000000000000000000000000000000..cbf290ef3187127e6406ffe2eae9b868be1680bc --- /dev/null +++ b/2024/PartGLEE_ A Foundation Model for Recognizing and Parsing Any Objects/full.md @@ -0,0 +1,288 @@ +# PartGLEE: A Foundation Model for Recognizing and Parsing Any Objects + +Junyi Li\*, Junfeng Wu\*, Weizhi Zhao\*, Song Bai\*, and Xiang Bai + +1 Huazhong University of Science and Technology +2 ByteDance Inc. + +Abstract. We present PartGLEE, a part-level foundation model for locating and identifying both objects and parts in images. Through a unified framework, PartGLEE accomplishes detection, segmentation, and grounding of instances at any granularity in the open world scenario. Specifically, we propose a Q-Former to construct the hierarchical relationship between objects and parts, parsing every object into corresponding semantic parts. By incorporating a large amount of object-level data, the hierarchical relationships can be extended, enabling PartGLEE to recognize a rich variety of parts. We conduct comprehensive studies to validate the effectiveness of our method, PartGLEE achieves the state-of-the-art performance across various part-level tasks and obtain competitive results on object-level tasks. The proposed PartGLEE significantly enhances hierarchical modeling capabilities and part-level perception over our previous GLEE model. Further analysis indicates that the hierarchical cognitive ability of PartGLEE is able to facilitate a detailed comprehension in images for mLLMs. The model and code will be released at https://provencstar.github.io/PartGLEE-Vision/. + +Keywords: Foundation Model $\cdot$ Hierarchical Recognition $\cdot$ Part Segmentation + +# 1 Introduction + +In recent years, foundation models have dominated the majority of tasks in the fields of Natural Language Processing [3, 9, 54] and Computer Vision [19, 23, 53, 60, 71, 72]. CLIP family [12, 13, 21, 53, 81] have made significant advancements in transfer learning and have demonstrated impressive zero-shot capabilities on vision-language tasks. SAM [23] has revolutionized the development of segmentation tasks and is able to provide multi-level class-agnostic masks. GLEE [71] utilized diverse object-level data to develop general object representations, enabling detection, segmentation, tracking, grounding, and identification of objects in open-world scenarios. Their remarkable achievement can be attributed to the integration of extensive and diverse range of datasets. + +![](images/47f7c330d14fe208579b04eeb8af74ec4ab188c2b13614d4336fbf924c3b0dda.jpg) +Fig. 1: An illustrative example demonstrating image annotations at diverse granularities across multiple datasets. The annotations at hierarchical levels with corresponding relationships are depicted on the right side. Below is a visualization of our segmentation results at multiple granularities. + +Different from the vast quantity of object-level data, the scale of part-level data is relatively small, which turns out to be a major bottleneck hindering vision models from recognizing part-level instances. Thus, most vision models lack the hierarchical comprehension between objects and parts. However, it is evident that the ability to recognize parts from objects is essential for various practical applications such as image editing [22,33,38], behavior analysis [50,75], pose estimation [10,76], robotics manipulation [2,49], etc. Moreover, we observe that part-level information is able to help multi-modal Large Language Models (mLLMs) in achieving a more detailed understanding of image content. Since part-level comprehension is a critical ability for foundation models to tackle a broader range of problems, it leads to a natural question: How could we break through data limitations to build a part-level vision foundation model? + +To enable object foundation model with part-level cognitive ability, we emphasize that the model should achieve two key objectives: (1) Hierarchical + +Comprehension, the model is supposed to understand the intrinsic relationship between objects and parts, and extend this hierarchical connection to any novel object, (2) Semantic Granularity, the model should be capable of learning a universal feature representation, enabling it to recognize semantic instances at any granularity. Consequently, we present a method to jointly detect and segment both objects and parts in a top-down manner. A lightweight Querying Transformer (Q-Former) is proposed to construct the hierarchical relationship between objects and parts. Specifically, it employs a set of universal parsing queries to interact with object queries, consequently generating multiple part-level queries that are capable of predicting corresponding semantic parts for each object. The Q-Former acts as a decomposer, which first recognizes individual objects in the images and subsequently parsing them into parts. Such model design is built upon the observation that various common objects often exhibit shared characteristics of parts. For example, cats, dogs, and dinosaurs all have parts such as torso, legs, and tails. In this way, two sets of query embeddings at different levels are generated, which are then used to predict object-level and part-level instances respectively. Through this approach, the relationship between objects and parts is established via the Q-Former design. Meanwhile, the hierarchical levels of objects and parts are distinguished, which is different from previous research [6,55,62,68,70] that consider parts as fine-grained objects. This paradigm enables vision models to better understand the features on different levels during training, thereby achieving improved performance. + +Our complete solution, PartGLEE, for jointly detecting and segmenting instances at both object and part levels, makes it possible for vision models to achieve favorable outcomes on both object and part levels. Some previous research have devised specialized training paradigms to utilize abundant image-text pair data [79, 89] as well as grounding data [30, 39, 71], thereby enhancing the cognitive and generalization capabilities of the models. On the contrary, the quantity of part-level data is much smaller compared to object-level data. So far, the largest dataset incorporating the concepts of both objects and parts is the recently proposed PACO [55] dataset. The scarcity of data has limited research on part-level recognition and restricted the generalization improvement of vision models. Although VLPart [62] has attempted to utilize pseudo-labeling schemes to generate part-level annotations for both object-level and image-level datasets, the quality of the pseudo-labels is relatively poor. Our innovative algorithm that parsing objects into their corresponding parts facilitates the transfer of generalization capability from objects to parts. Consequently, parts are generated from objects, which enables vision models to maintain generalization performance when predicting parts for novel objects without labeling extensive part-level data. To facilitate the training process of Q-Former, we standardize the annotation granularity across various part-level datasets and introduce a vast amount of object-level datasets, an intuitive display of the overall training data is shown in Fig. 1. Unlike VLPart, which exhibits unsatisfactory performance at object-level datasets after joint-training, our method demonstrates favorable outcomes at both object and part levels after joint-training. Moreover, it turns out that + +using object-level datasets is able to improve the performance of the model on part-level tasks, indicating a beneficial interaction between objects and parts. + +Extensive experiments demonstrate that our method significantly improve the open-vocabulary part segmentation performance, concurrently ensuring a decent performance on object detection and segmentation. We verify its effectiveness on various popular datasets. To validate the generalization performance of our model in identifying various parts of novel objects, we conduct experiments on PartImageNet [18] and Pascal Part [6] datasets in cross-dataset and cross-category manners respectively. Our method exhibits strong transferability and generalization ability when adding extra object-level datasets during training. To evaluate the decomposition capability of our model, we conduct experiments on both ADE20K-Part and Pascal Part datasets follow OV-PARTS [70]. As a result, our approach significantly outperforms one-stage baselines of OV-PARTS, with an increase of $8.16\%$ and $2.07\%$ on harmonic mean IoU (hIoU) in ADE20K-Part-234 and Pascal-Part-116 respectively. Additionally, by incorporating a large amount of object-level data for joint-training, our method establishes generic hierarchical relationships and breaks through the limitations of scarce part-level data, achieving state-of-the-art performance across various part-level tasks. + +In conclusion, our main contributions can be summarized as follows: + +1. We construct the hierarchical relationship between objects and parts via the Q-Former, facilitating part segmentation to acquire advantages from various object-level datasets. +2. We propose a unified pipeline for hierarchical detection and segmentation, where we first recognize objects and then parsing them into corresponding semantic parts. This algorithm enables us to jointly detect and segment both object-level and part-level instances. +3. We standardize the annotation granularity across various part-level datasets by incorporating corresponding object-level annotations, complementing the hierarchical correspondences for current part-level datasets, promoting the development of vision foundation models. + +# 2 Related Work + +# 2.1 Visual Foundation Models and General Models + +Vision foundation models and generalist models are considered as a milestone in the development of the intelligent vision system. For instance, multi-modal visual foundation models [1, 21, 53, 67, 81] have significantly advanced efficient transfer learning and exhibit impressive zero-shot capabilities on vision-language tasks by using contrastive learning with large-scale image-text pairs. Generative foundation models [11, 56, 57, 60] are trained on vast collections of images and captions, empowering them to generate image content conditioned on textual prompts. Self-supervised foundation models [4, 12, 13, 19] have learned general visual representations from large-scale image datasets, enhancing their ability to transfer to downstream tasks. However, the image-level features learned by these + +foundation models are not well-suited for direct application to dense prediction tasks that involve precise object and part localization. + +Transformer-based generalist methods [5,41,66,77,87] adopt a sequence generation pipeline to unify the output of text and spatial coordinates. However, they mainly focus on image-level comprehension, which results in relatively weak localization capabilities. Works such as UNINEXT, etc. [28, 73, 74], built upon strong detectors [27, 86], demonstrating a strong localization capability across multiple datasets. But they fail to exhibit zero-shot transfer ability and generalization capability due to their closed-set training paradigm. Some works about open-vocabulary detection (OVD) [30, 34, 35, 42, 43, 78, 82, 83] have explored zero-shot generalization capabilities on novel categories. X-Decoder [89] and SEEM [90] have developed a versatile decoding architecture that are able to generate accurate pixel-level segmentation predictions. GLEE [71] addresses various object-level tasks through a unified architecture and training paradigm. However, current generalist models and foundation models are trained mainly on image-level and object-level datasets, thus their ability to extract more fine-grained information is limited, making it difficult for them to recognize corresponding semantic parts of any object. Our work focuses on empowering hierarchical cognitive capability for vision foundation models, thereby further advancing the development of comprehensive visual systems. + +# 2.2 Part Segmentation + +The growing interest in achieving a more fine-grained understanding of objects has sparked a surge in research focused on part level recognition. Some pioneering studies have introduced datasets with part-level annotations, concentrating on objects of some specific categories such as human body parts [15, 29, 75], animal body parts [64] and vehicle components [58]. More general part annotations for common objects such as Pascal-Part [6], PartImageNet [18], ADE20K [84], CityscapesPanoptic-Parts [45] and more recent PACO [55] were then proposed to promote more in-depth research in the field of parts. Most of the previous works [14, 31, 46, 85] were conducted based on a closed-set configuration, thus only capable of detecting and segmenting closed-set objects and parts. Recently, VLPart [62] present a pipeline for detecting and segmenting both open-vocabulary objects and their corresponding part regions, while OVPARTS [70] utilize adapters to transfer the generalization abilities of CLIP into open-vocabulary part segmentation task. However, due to the limited quantity of data, the generalization capability of previous models [15, 29, 51, 62, 63, 68, 70, 75] still relies heavily on the training datasets. Furthermore, in prior works, both objects and parts are treated equally, they consider part as a special type of object. On the contrary, we distinguish them by considering parts as integral components subordinate to objects and generate parts from corresponding objects in a top-down manner. Our work is aimed at building hierarchical relationships while unifying the training paradigm for object and part-level data. By incorporating a large amount of object-level data, the hierarchical relationships can be extended to any object, enabling our method to recognize a rich variety of parts. + +![](images/d2e25db7a14cb03ec2bc08abfeda2e0926d6c4b89e67d93ac90204d9c787ef5a.jpg) +Fig. 2: Framework of PartGLEE. The Q-Former takes each object query as input and output the corresponding part queries. These queries are then fed into the object decoder and the part decoder respectively to generate hierarchical predictions. + +# 2.3 Hierarchical Learning of Objects and Parts + +Learning objects through parts has been a long-standing research topic as part annotations provide more detailed semantic information of objects. Morabia et al. [48] first introduced a pipeline employing an attention mechanism for simultaneous detection of both objects and parts. Deepflux [69] designed an image context flux representation which enables better object parts interaction for skeleton detection. Leopart [88] demonstrated that learning object parts can provide spatially diverse representation which facilitates self-supervised semantic segmentation. Wang et al. [65] proposed a method to predict both parts and objects simultaneously on Pascal-Part dataset [6]. Recent studies such as SAM [23] and Semantic-SAM [26] have studied on class-agnostic multi-granularity interactive segmentation task. However, they have not explored the relationship between objects and their corresponding semantic parts. Recently, Compositor [17] designed a bottom-up pipeline to predict parts and then cluster them into objects, while AIMS [52] utilized an independent relation decoder to construct the hierarchical association between objects and parts. Different from these works, our approach introduces a Querying Transformer to model the hierarchical relationship, allowing our model to parse any object into its corresponding parts. + +# 3 Method + +# 3.1 Overall Framework + +Following [68,71], we propose PartGLEE, which comprises of an image encoder, a Q-Former, two independent decoders and a text encoder, as shown in Fig. 2. + +Given an input image $I \in \mathcal{R}^{H \times W \times 3}$ , the backbone and the pixel decoder first extract multi-scale image features $F_s \in \mathcal{R}_{\frac{H}{2^S} \times \frac{W}{2^S} \times C}$ and $s = \{2, 3, 4, 5\}$ with + +backbones such as ResNet [20] or Swin Transformer [40]. Then we feed them into the object decoder, where the object-level query embeddings $q_{obj} \in \mathcal{R}^{N \times C}$ are generated in a two-stage process. These object queries are utilized to perform object-level classification, detection as well as segmentation tasks through three independent prediction heads. Besides, the object queries $q_{obj}$ are fed into the Q-Former simultaneously, where $L$ learnable universal parsing queries are initialized to interact with object queries. It takes object queries as input and generate part-level queries $q_{part} \in \mathcal{R}^{N \cdot L \times C}$ which are then passed into the part decoder to yield part-level predictions (detailed in Sec. 3.2). To enhance the semantic-awareness, an early fusion module is adopted before Transformer encoder following [74], which takes image feature from backbone and text embedding as input and perform bi-directional cross-attention between them. In line with previous segmentation models [7,27,32], a pixel embedding map $M_p \in \mathcal{R}^{\frac{H}{4} \times \frac{W}{4} \times C}$ at $1/4$ resolution is constructed by upsampling and integrating multi-scale feature maps from the backbone and the pixel decoder. Eventually, we dot product each object query or part query with the pixel embedding map to derive an output mask $m \in \mathcal{R}^{\frac{H}{4} \times \frac{W}{4}}$ : + +$$ +m = F F N (q _ {l}) \otimes M _ {p}, \quad l \in \{o b j, p a r t \} \tag {1} +$$ + +where FFN is comprised of 3 layers feed forward network with ReLU activation functions and linear layers. + +# 3.2 Parsing Objects into Parts + +We propose a Q-Former to establish the hierarchical relationship between objects and parts. As Various common objects tend to manifest shared attributes in their constituent parts, for example, both lizards and birds exhibit similar components, such as heads and torsos. Thus, we initialize a set of query embeddings in the Q-Former to parse any object into semantic parts. We denote these universal parsing query embeddings as $q_{parse} \in \mathcal{R}^{L \times C}$ , where $L$ represents the number of the parsing queries. As shown in Fig. 2, the Q-Former is comprised of $M$ cascaded attention modules, each module includes a self-attention layer, a cross-attention layer, and a feed forward network. The universal parsing queries are first fed into the self-attention layer and then perform cross-attention with the object queries. Note that every object query is interacted with all universal parsing queries. Hence, assume $N$ object queries $(q_{obj} \in \mathcal{R}^{N \times C})$ are generated from the object decoder, and $L$ universal parsing queries $q_{parse} \in \mathcal{R}^{L \times C}$ are initialized in the Q-Former, we obtain $N \cdot L$ part-level queries which can be denoted as $q_{part} \in \mathcal{R}^{N \cdot L \times C}$ . We refer to this process as: + +$$ +q _ {p a r t} = Q - F o r m e r \left(q _ {p a r s e}; q _ {o b j}\right) \tag {2} +$$ + +Our proposed Q-Former functions as a decomposer, extracting and representing parts from object queries. Hence, by training jointly on extensive object-level datasets and limited hierarchical datasets which contain object-part correspondences, our Q-Former obtains strong generalization ability to parse any novel object into its corresponding parts. + +![](images/9a2fdac931e61e57734068d20d91b4215414125b93fa7a763beb1991e846600d.jpg) +Fig. 3: Matching mechanisms of PartGLEE. Two separate forward passes are performed on the same image to obtain hierarchical segmentation results. + +# 3.3 Unified Training Paradigm for Objects and Parts + +Since the Q-Former requires hierarchical data to learn how to parse objects into parts, we enrich part-level data with corresponding object-level annotations, details of which are provided in the supplementary materials. + +Since the annotation granularity across part-level datasets is standardized, our model can first learn the characteristics of objects and then acquire the ability to parse any object into its semantic parts. To facilitate open-vocabulary detection and segmentation, we substitute the similarity scores between the instance embeddings and the text embeddings for the original class head. Given $K$ object-level and part-level input categories as separate sentences, we feed them into the text encoder and utilize the average of each individual sentence tokens as the output text embedding $T_{l}$ for each category. Then the similarity scores $S_{l} \in \mathcal{R}^{N \times K}$ are calculated through a dot product operation between the hierarchical instance embeddings $q_{l} \in \mathcal{R}^{N \times C}$ from detector and the text embeddings $T_{l} \in \mathcal{R}^{K \times D}$ from text encoder, which can be denoted as: + +$$ +S _ {l} = q _ {l} \cdot W _ {p r o j} \otimes T _ {l}, \quad l \in \{o b j, p a r t \} \tag {3} +$$ + +where $W_{proj} \in \mathcal{R}^{C \times D}$ is a trainable projection weight for fine-tuning text embedding space especially for part-level descriptions. Following [26, 68], we perform Hungarian matching of objects and parts individually, where object-level predictions are only matched with object-level targets, and the same applies to the part-level output, as shown in Fig. 3. + +We then introduce a constraint loss to ensure the part-level predictions to be the component of the objects. We denote this novel loss function as restriction loss $L_{res}$ . Due to memory limitations, we only calculate our restriction loss on the predicted bounding boxes between different levels, while leaving the predicted masks unconstrained. Our restriction loss can be calculated as follow: + +$$ +L _ {r e s} = \sum_ {i} ^ {L} \left(1 - \frac {\left| S _ {o b j} \cap S _ {p a r t} ^ {i} \right|}{S _ {p a r t} ^ {i}}\right) \tag {4} +$$ + +where $S_{obj}$ represents the area of the object-level bounding box prediction, and $S_{part}^{i}$ stands for the area of the $i - th$ part-level bounding box prediction. Note that each object query can generate $L$ part queries through Q-Former. This loss function is only applied to the matched predictions in part-level datasets, thereby strengthening the mutual correspondence between different hierarchies. + +PartGLEE is trained with a linear combination of losses for object-level tasks and part-level tasks, which can be formulated as: + +$$ +L = \lambda_ {1} (L _ {c l s} ^ {o b j} + L _ {c l s} ^ {p a r t}) + \lambda_ {2} (L _ {b o x} ^ {o b j} + L _ {b o x} ^ {p a r t}) + \lambda_ {3} (L _ {m a s k} ^ {o b j} + L _ {m a s k} ^ {p a r t}) + \lambda_ {4} L _ {r e s} \quad (5) +$$ + +where $L_{cls}^{l}, L_{box}^{l}, L_{mask}^{l}$ are the classification, box, and mask loss at different levels ( $l \in \{obj, part\}$ ), while $L_{res}$ is the restriction loss, and $\lambda$ are their corresponding weights. We apply Focal Loss [36] as the classification loss on the similarity scores $S_{l}$ to align the text concepts with instance features. A combination of L1 loss and generalized IoU loss [59] is utilized for box predictions, while we employ both Dice Loss [47] and Focal Loss to calculate mask loss. We follow MaskDINO to set our hyperparameters to $\lambda_{1} = 4, \lambda_{2} = 2, \lambda_{3} = 5, \lambda_{4} = 5$ . Based on the above designs, PartGLEE is able to leverage both object-level data and part-level data thus obtaining a strong generalization capability. + +# 4 Experiments + +# 4.1 Experimental Setup + +We conduct comprehensive experiments to exhibit the effectiveness of PartGLEE across a wide range of object-level and part-level tasks. + +Data Unification. We utilize object-level datasets such as COCO [37], LVIS [16], Object365 [61], OpenImages [25], Visual Genome [24] and RefCOCO series [44,80], etc, while using part-level datasets PACO [55], PartImageNet [18], Pascal Part [6], ADE20K-Part [70] and SA-1B [23] with varying annotation granularity for joint-training. For Visual Genome and SA-1B, we categorize their corresponding part-level annotations based on semantic and mask overlap relationships to construct hierarchical data versions. For part-level data, we integrate the original part-level annotations with corresponding object-level annotations according to their associated object-level dataset. The details of these dataset preprocessing steps are left in the supplementary materials. + +Implementation Details. In our experiments, we utilize ResNet-50 [20] and Swin-Large [40] as the vision encoder. Following MaskDINO [27], we adopt deformable transformer in the decoder, and use 300 object queries while setting the number of parsing queries $L$ to be 10. The $M$ of Q-Former is set to 6. We select the top 50 object queries based on the similarity scores and input them into the Q-Former, ultimately yielding 500 part queries. We use both query denoising and hybrid matching strategies to facilitate convergence and enhance performance. We conduct experiments on part-level datasets following the methodologies of VLPart [62] and OV-PARTS [70] in order to evaluate the generalization performance and the ability to parse novel objects of our model. For joint-training, we + +Table 1: Cross-dataset generalization performance compared with VLPart. The evaluation metric is $mAP_{mask}$ on the validation set of PartImageNet. All models utilize ResNet-50 as backbone and use the text embeddings of the category names as the classifier. PartImageNet denotes the fully-supervised method reported for comparison. + +
MethodDatasetsAll (40)quadruped
headbodyfoottail
VLPart [62]Pascal Part4.517.40.10.02.9
+ IN-S11 label5.423.63.40.81.2
+ Parsing IN-S117.835.015.23.58.9
vs. baseline+3.3+17.6+15.1+3.5+6.0
PartImageNet29.757.325.822.922.9
PartGLEEPascal Part9.923.64.51.34.6
+ Parsing IN-S1114.955.327.27.023.6
vs. baseline+5.0+31.7+22.7+5.7+19.0
PartImageNet40.267.037.636.540.7
+ +train PartGLEE based on the weights of GLEE [71], continuing training on 32 A100 GPUs. The settings for the part-level zero-shot experiments are described separately in each section. + +# 4.2 Zero-Shot Part Segmentation Results + +1) Cross-dataset Part Segmentation on PartImageNet. We follow VLPart [62] to conduct experiments on cross-dataset generalization performance by directly evaluating on PartImageNet [18] validation set. We report the metrics of all (40) part categories and the detailed metrics of quadruped are also provided. The baseline approach only utilize Pascal Part as the training set and directly perform evaluation on PartImageNet in a zero-shot manner. Note that IN-S11 label represents adding image-level classification data for training in order to improve performance. Meanwhile, Parsing IN-S11 stands for training with the pseudo-labels generated from the parsing pipeline proposed by VLPart. However, both of these methods expose the model to categories and images from the PartImageNet dataset. We first perform our training process exclusively on the Pascal Part dataset to verify our zero-shot capabilities, and then we incorporate pseudo-labels to assess the ability of our model to utilize low-quality annotations. + +Given that Pascal Part does not provide semantic labels for categories like quadruped in PartImageNet, the model needs to generalize from annotated parts of dog, cat, etc. in Pascal Part to parts of quadruped in PartImageNet. As shown in Tab. 1, our model significantly outperforms VLPart when only training on Pascal Part, even surpassing the model trained with Parsing IN-S11. After incorporating pseudo-labeled data into training, our model shows higher performance gains, indicating better utilization of low-quality data. This result illustrates the importance of hierarchical modeling, which enables our model to recognize and + +Table 2: Cross-category generalization performance compared with VLPart. The evaluation metric is $mAP_{mask}$ on the validation set of PascalPart but report AP50 specifically for dog parts. All models utilize ResNet-50 as backbone and use the text embeddings of the category names as the classifier. Base part represents the base split from Pascal Part. VOC object is added to the training process to improve the cognitive ability of the model thus reach a better performance. Pascal Part denotes the fully-supervised method reported for comparison. + +
MethodDatasetsAll AP (93)BaseAP (77)NovelAP (16)dogNovelAP Increment
headtorsolegpawtail
VLPart [62]Base Part15.017.81.56.17.92.913.83.2-
+ VOC object16.819.92.129.922.63.212.42.1+0.6
+ IN-S20 label17.420.81.112.817.82.05.90.9-0.4
+ Parsing IN-S2018.421.34.228.734.817.25.714.3+2.7
Pascal Part19.418.822.488.049.638.348.925.8-
PartGLEEBase Part25.630.52.112.615.68.25.26.2-
+ VOC object26.931.25.846.535.027.014.715.1+3.7
+ Parsing IN-S2026.628.915.580.357.336.717.037.4+9.7
Pascal Part35.534.639.995.988.575.076.772.9-
+ +parse novel objects into their corresponding parts based on the generalization capability brought by CLIP. + +2) Cross-category Part Segmentation on Pascal Part. We follow the evaluation setting proposed by VLPart to assess the cross-category generalization performance of our model on the Pascal Part dataset. A total of 93 part categories are divided into 77 base part categories and 16 novel part categories. Tab. 2 presents the evaluation results for all (93), base (77), and novel (16) parts. The model is trained only on the base categories, and is directly evaluated on the entire datasets. Note that IN-S20 label represents adding image-level classification data and Parsing IN-S20 is on behalf of the pseudo-labels generated by VLPart [62] on ImageNet [8]. We further introduce a metric called NovelAP Increment on top of VLPart to assess the improvement of our model when adding extra object datasets into the training process. It is calculated by subtracting the baseline Novel AP from the Novel AP achieved after incorporating extra datasets. The results shown in Tab. 2 demonstrate that our method surpasses the performance of VLPart by a large margin. By comparing the NovelAP Increment, we observe that our method achieves a greater increment after incorporating extra object dataset. Since the VOC dataset includes object categories corresponding to novel parts, the hierarchical relationships of the Q-Former can be extended to novel part categories, resulting in a higher NovelAP Increment. +3) Generalized Zero-Shot Part Segmentation. We adopt the Oracle-Obj setting proposed by OV-PARTS [70] to conduct experiments on ADE-Part234 and Pascal-Part-116 datasets. This setting assumes that the ground-truth masks and categories of object-level instances are known during the inference process, aiming to evaluate the capability of the model to parse any novel object. All categories in the datasets are divided into a base set and a novel set, + +Table 3: Generalized zero-Shot part segmentation performance on ADE-Part-234 and Pascal-Part-116 compared with baselines proposed by OV-PARTS. + +
MethodModelBackboneFinetuningOracle-Obj
ADE-Part-234Oracle-obj
SeenUnseenHarmonicSeenUnseenHarmonic
Fully SupervisedMask2FormerResNet-50X46.2547.86-55.2852.14-
Two-StageZSeg+ResNet-50CPTCoOp43.1927.8433.8555.3319.1728.48
ResNet-50CPTCoCoOp39.6725.1530.7854.4319.0428.21
ResNet-101cCPTCoOp43.4125.7032.2857.8821.9331.81
One-StageCATSegResNet-101&ViT-B/16X11.498.569.8114.8910.2912.17
ResNet-101&ViT-B/16B+D31.4025.7728.3143.9726.1132.76
CLIPSegViT-B/16X15.2718.0116.5322.3319.7320.95
ViT-B/16VA+L+F+D38.9629.6533.6748.6827.3735.04
PartGLEEResNet-50X51.2935.3341.8357.4327.4137.11
+ +and the training process is performed only on the base set, while we evaluate the performance of the model on all categories. As shown in Tab. 3, our model achieves a superior performance on both datasets, which indicates the importance of hierarchical modeling. The establishment of hierarchical relationships between objects and parts enables our model to extend to novel objects, thereby effectively parsing them into corresponding semantic parts. Consequently, our model exhibits outstanding performance across both datasets. + +# 4.3 Joint-training results on Detection and Segmentation + +To endow our model with robust generalization capability, we perform joint training on various datasets and evaluate its performance on both object-level and part-level tasks. We compare our model with specialist and generalist models to evaluate its performance on object-level data. Additionally, we contrast it with VLPart to assess its performance on part-level datasets as well as the effectiveness of joint-training process on both types of datasets. As shown in Tab. 4, PartGLEE significantly outperforms VLPart on both object-level and part-level tasks after joint-training, while achieving comparable performance on object-level tasks compared with previous SOTA. Through joint-training, our model has acquired strong generalization performance, allowing it to simultaneously address tasks for different hierarchies. We also observe that VLPart fails to achieve satisfactory performance on both object-level and part-level tasks. For example, VLPart obtains better performance on Pascal Part than its dataset-specific oracle, while decreasing its performance on COCO and LVIS. We attribute the performance drop of VLPart to the absence of hierarchical relationships, which causes confusion in modeling parts and objects and impairs object-level performance. PartGLEE effectively addresses this problem and extends the generalization capabilities from object-level to part-level tasks. + +Table 4: Joint-Training Performance of PartGLEE. Note that Oracle represents the dataset-specific training paradigm. We directly evaluate the generalist models on PACO to assess their recognition capability at the part level, as indicated by the results annotated in the grey font. + +
TypeMethodPart-level TasksObject-level Tasks
PartImageNetPascal PartPACOCOCO-valLVIS-minimalLVIS-val
APboxAPmaskAPboxAPmaskAPmaskAPobjmaskAPpartmaskAPboxAPmaskAPboxAPmaskAPbox
SpecialistMask2Former(R50) [7]------46.243.7----
Mask2Former(L) [7]------50.1-----
MaskDINO(R50) [27]------50.546.0----
MaskDINO(L) [27]------58.352.1----
ViTDet-L [32]------57.649.8--51.246.0
ViTDet-H [32]------57.649.8--53.448.1
EVA-02-L [12]------64.255.0--65.257.3
PACO(R50) [55]----32.612.5------
PACO(L) [55]----43.417.7------
GeneralistPix2Seq v2 [5]------46.538.2----
X-Decoder(L) [89]---2.6911.90.94-46.7----
SEEM(L) [90]---1.998.420.69-47.7----
HIPIE(R50) [68]------53.945.9----
Florence-2(B) [72]------41.4-----
Florence-2(L) [72]------43.4-----
UNINEXT(R50) [74]------51.344.9--36.4-
UNINEXT(L) [74]------58.149.6----
GLEE(R50) [71]---3.4415.31.2955.048.450.545.944.240.2
HierarchicalVLPart(R50) [62]30.731.623.924.013.836.99.628.5--26.2-
VLPart(R50)-Oracle [62]29.229.718.919.413.328.010.638.0--28.1-
VLPart(B) [62]43.941.233.531.722.155.015.940.3--39.6-
VLPart(B)-Oracle [62]44.341.729.227.419.137.715.252.5--43.1-
PartGLEE (R50)40.940.235.035.521.850.515.454.447.648.743.542.7
PartGLEE (L)52.750.939.639.127.855.721.359.552.056.550.645.0
+ +# 4.4 Ablation Study + +To demonstrate that our model design achieves satisfactory results on both object-level and part-level tasks, we conduct an ablation study (depicted in Fig. 4) on the model architecture and present results in Tab. 5. We ablate with a backbone of ResNet-50 and perform joint-training on COCO [37], LVIS [16], PartImageNet [18], Pascal Part [6] and PACO [55] with 90K iterations. From this study, we draw several important conclusions: (1) The utilization of parallel pixel decoders only results in slight improvements in mask predictions on few datasets, indicating that the influence of feature maps at different granularities is negligible. (2) Adopting independent decoders to obtain predictions at different levels demonstrates superior performance across the majority of datasets, manifesting the effectiveness of independent decoders. As adopting parallel pixel decoders (b) results in significant GPU memory costs without considerable gains, and all metrics for (a) are lower than (c), we select (c) as our final model design. Additional ablation studies, extensive qualitative analysis, and experiments on mLLM can be found in the supplementary materials. + +# 4.5 Limitations + +In this work, we still adopt CLIP as the text encoder, which is trained on text-image pairs and thus lacks the ability to perceive fine-grained descriptions of object or part instances. This limitation may restrict the improvement of model + +Table 5: An ablation study on different model designs, as depicted in Fig 4. Note that Parallel Pixel Decoders refers to the utilization of two pixel decoders to generate feature maps at different hierarchies respectively. Independent Decoders denote the usage of two decoders, which facilitate the interaction between feature maps and queries at different hierarchies. Our final choice is scheme (c), which is highlighted in gray. + +
SchemeModel DesignPart-level TasksObject-level Tasks
Parrallel Pixel DecodersIndependent DecodersPartImageNetPascal PartPACOCOCO-valLVIS-minimal
APmaskAPmaskAPmaskAPmaskAPobjmaskAPopartmaskAPboxAPmaskAPboxAPmask
(a)XX39.034.120.147.413.547.843.534.833.4
(b)✓X38.334.520.848.813.848.544.334.934.2
(c)X✓39.034.720.947.914.249.344.235.633.8
+ +![](images/4dcd8f982810a9b1fe787c65d99b5e0e3b46a23f84bf7f62b07359704c85f0f0.jpg) +(a) + +![](images/1974bef7631da668f780e78acdb603bc1363bc5433c6727de90ac6ef375762d6.jpg) +(b) +(c) +Fig. 4: Various designs for generating predictions at different hierarchies. In scheme (a), we only utilize a single decoder to generate predictions for both objects and parts. In scheme (b), two parallel pixel decoders are employed to generate feature maps at different levels, aiming to explore the effectiveness of feature maps at different granularity. In scheme (c), we use two independent decoders to generate predictions for objects and parts respectively. + +![](images/5f24408bc94c656ce390a6f85c1be5a280f2124a8feaa04d680d885d7a1149f2.jpg) + +performance and prompts us to consider how to enhance the perception capabilities of region-level models, which will be our future work. + +# 5 Conclusion + +In this paper, we introduce PartGLEE, a groundbreaking foundation model designed towards a complete comprehension of both objects and parts in images. Through the generic hierarchical relationships established by the Q-Former, we are able to break through the limitation of scarce part-level data by introducing a large amount of object-level data, thereby transferring the powerful generalization capabilities from objects to parts. Through extensive training on diverse datasets, PartGLEE achieves SOTA performance across various part-level tasks while maintaining competitive results on object-level tasks, enabling it to parse any objects into parts and serve as a foundation model for general fine-grained region-level perception tasks. + +# Acknowledgements + +This work was supported by the National Science Fund for Distinguished Young Scholars of China (Grant No.62225603). + +# References + +1. Alayrac, J.B., Donahue, J., Luc, P., Miech, A., Barr, I., Hasson, Y., Lenc, K., Mensch, A., Millican, K., Reynolds, M., et al.: Flamingo: a visual language model for few-shot learning. In: Advances in neural information processing systems. vol. 35, pp. 23716-23736 (2022) +2. Brohan, A., Brown, N., Carbajal, J., Chebotar, Y., Dabis, J., Finn, C., Gopalakrishnan, K., Hausman, K., Herzog, A., Hsu, J., et al.: Rt-1: Robotics transformer for real-world control at scale. arXiv preprint arXiv:2212.06817 (2022) +3. Brown, T., Mann, B., Ryder, N., Subbiah, M., Kaplan, J.D., Dhariwal, P., Neelakantan, A., Shyam, P., Sastry, G., Askell, A., et al.: Language models are few-shot learners. In: Advances in neural information processing systems. vol. 33, pp. 1877-1901 (2020) +4. Caron, M., Touvron, H., Misra, I., Jégou, H., Mairal, J., Bojanowski, P., Joulin, A.: Emerging properties in self-supervised vision transformers. In: Proceedings of the IEEE/CVF international conference on computer vision. pp. 9650-9660 (2021) +5. Chen, T., Saxena, S., Li, L., Lin, T.Y., Fleet, D.J., Hinton, G.E.: A unified sequence interface for vision tasks. Advances in Neural Information Processing Systems 35, 31333-31346 (2022) +6. Chen, X., Mottaghi, R., Liu, X., Fidler, S., Urtasun, R., Yuille, A.: Detect what you can: Detecting and representing objects using holistic models and body parts. In: Proceedings of the IEEE conference on computer vision and pattern recognition. pp. 1971-1978 (2014) +7. Cheng, B., Misra, I., Schwing, A.G., Kirillov, A., Girdhar, R.: Masked-attention mask transformer for universal image segmentation. In: Proceedings of the IEEE/CVF conference on computer vision and pattern recognition. pp. 1290-1299 (2022) +8. Deng, J., Dong, W., Socher, R., Li, L.J., Li, K., Fei-Fei, L.: Imagenet: A large-scale hierarchical image database. In: Proceedings of the IEEE/CVF International Conference on Computer Vision. pp. 248-255 (2009) +9. Devlin, J., Chang, M.W., Lee, K., Toutanova, K.: Bert: Pre-training of deep bidirectional transformers for language understanding. arXiv preprint arXiv:1810.04805 (2018) +0. Dong, J., Chen, Q., Shen, X., Yang, J., Yan, S.: Towards unified human parsing and pose estimation. In: Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition. pp. 843-850 (2014) +1. Esser, P., Rombach, R., Ommer, B.: Taming transformers for high-resolution image synthesis. In: Proceedings of the IEEE/CVF conference on computer vision and pattern recognition. pp. 12873-12883 (2021) +2. Fang, Y., Sun, Q., Wang, X., Huang, T., Wang, X., Cao, Y.: Eva-02: A visual representation for neon genesis. arXiv preprint arXiv:2303.11331 (2023) +3. Fang, Y., Wang, W., Xie, B., Sun, Q., Wu, L., Wang, X., Huang, T., Wang, X., Cao, Y.: Eva: Exploring the limits of masked visual representation learning at scale. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 19358-19369 (2023) + +14. de Geus, D., Meletis, P., Lu, C., Wen, X., Dubbelman, G.: Part-aware panoptic segmentation. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 5485-5494 (2021) +15. Gong, K., Liang, X., Zhang, D., Shen, X., Lin, L.: Look into person: Self-supervised structure-sensitive learning and a new benchmark for human parsing. In: Proceedings of the IEEE conference on computer vision and pattern recognition. pp. 932-940 (2017) +16. Gupta, A., Dollar, P., Girshick, R.: Lvis: A dataset for large vocabulary instance segmentation. In: Proceedings of the IEEE/CVF conference on computer vision and pattern recognition. pp. 5356-5364 (2019) +17. He, J., Chen, J., Lin, M.X., Yu, Q., Yuille, A.L.: Compositor: Bottom-up clustering and compositing for robust part and object segmentation. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 11259-11268 (2023) +18. He, J., Yang, S., Yang, S., Kortylewski, A., Yuan, X., Chen, J.N., Liu, S., Yang, C., Yu, Q., Yuille, A.: Partimagenet: A large, high-quality dataset of parts. In: European Conference on Computer Vision. pp. 128-145. Springer (2022) +19. He, K., Chen, X., Xie, S., Li, Y., Dólár, P., Girshick, R.: Masked autoencoders are scalable vision learners. In: Proceedings of the IEEE/CVF conference on computer vision and pattern recognition. pp. 16000-16009 (2022) +20. He, K., Zhang, X., Ren, S., Sun, J.: Deep residual learning for image recognition. In: Proceedings of the IEEE conference on computer vision and pattern recognition. pp. 770-778 (2016) +21. Jia, C., Yang, Y., Xia, Y., Chen, Y.T., Parekh, Z., Pham, H., Le, Q., Sung, Y.H., Li, Z., Duerig, T.: Scaling up visual and vision-language representation learning with noisy text supervision. In: International conference on machine learning. pp. 4904-4916. PMLR (2021) +22. Kawar, B., Zada, S., Lang, O., Tov, O., Chang, H., Dekel, T., Mosseri, I., Irani, M.: Imagic: Text-based real image editing with diffusion models. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 6007-6017 (2023) +23. Kirillov, A., Mintun, E., Ravi, N., Mao, H., Rolland, C., Gustafson, L., Xiao, T., Whitehead, S., Berg, A.C., Lo, W.Y., et al.: Segment anything. In: Proceedings of the IEEE/CVF International Conference on Computer Vision. pp. 4015-4026 (2023) +24. Krishna, R., Zhu, Y., Groth, O., Johnson, J., Hata, K., Kravitz, J., Chen, S., Kalantidis, Y., Li, L.J., Shamma, D.A., et al.: Visual genome: Connecting language and vision using crowdsourced dense image annotations. International journal of computer vision 123, 32-73 (2017) +25. Kuznetsova, A., Rom, H., Alldrin, N., Uijlings, J., Krasin, I., Pont-Tuset, J., Kamali, S., Popov, S., Malloci, M., Kolesnikov, A., et al.: The open images dataset v4: Unified image classification, object detection, and visual relationship detection at scale. International journal of computer vision 128(7), 1956-1981 (2020) +26. Li, F., Zhang, H., Sun, P., Zou, X., Liu, S., Yang, J., Li, C., Zhang, L., Gao, J.: Semantic-sam: Segment and recognize anything at any granularity. arXiv preprint arXiv:2307.04767 (2023) +27. Li, F., Zhang, H., Xu, H., Liu, S., Zhang, L., Ni, L.M., Shum, H.Y.: Mask dino: Towards a unified transformer-based framework for object detection and segmentation. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 3041-3050 (2023) + +28. Li, H., Zhu, J., Jiang, X., Zhu, X., Li, H., Yuan, C., Wang, X., Qiao, Y., Wang, X., Wang, W., et al.: Uni-perceiver v2: A generalist model for large-scale vision and vision-language tasks. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 2691-2700 (2023) +29. Li, J., Zhao, J., Wei, Y., Lang, C., Li, Y., Sim, T., Yan, S., Feng, J.: Multiple-human parsing in the wild. arXiv preprint arXiv:1705.07206 (2017) +30. Li, L.H., Zhang, P., Zhang, H., Yang, J., Li, C., Zhong, Y., Wang, L., Yuan, L., Zhang, L., Hwang, J.N., et al.: Grounded language-image pre-training. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 10965-10975 (2022) +31. Li, X., Xu, S., Yang, Y., Cheng, G., Tong, Y., Tao, D.: Panoptic-partformer: Learning a unified model for panoptic part segmentation. In: European Conference on Computer Vision. pp. 729-747. Springer (2022) +32. Li, Y., Mao, H., Girshick, R., He, K.: Exploring plain vision transformer backbones for object detection. In: European Conference on Computer Vision. pp. 280-296. Springer (2022) +33. Li, Y., Singh, K.K., Xue, Y., Lee, Y.J.: Partgan: Weakly-supervised part decomposition for image generation and segmentation. In: British Machine Vision Conference (BMVC) (2021) +34. Lin, C., Jiang, Y., Qu, L., Yuan, Z., Cai, J.: Generative region-language pretraining for open-ended object detection. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 13958-13968 (2024) +35. Lin, C., Sun, P., Jiang, Y., Luo, P., Qu, L., Haffari, G., Yuan, Z., Cai, J.: Learning object-language alignments for open-vocabulary object detection. In: The Eleventh International Conference on Learning Representations (2023) +36. Lin, T.Y., Goyal, P., Girshick, R., He, K., Dollar, P.: Focal loss for dense object detection. In: Proceedings of the IEEE international conference on computer vision. pp. 2980-2988 (2017) +37. Lin, T.Y., Maire, M., Belongie, S., Hays, J., Perona, P., Ramanan, D., Dollar, P., Zitnick, C.L.: Microsoft coco: Common objects in context. In: European Conference on Computer Vision. pp. 740-755. Springer (2014) +38. Ling, H., Kreis, K., Li, D., Kim, S.W., Torralba, A., Fidler, S.: Editgan: High-precision semantic image editing. In: Advances in Neural Information Processing Systems. vol. 34, pp. 16331-16345 (2021) +39. Liu, S., Zeng, Z., Ren, T., Li, F., Zhang, H., Yang, J., Li, C., Yang, J., Su, H., Zhu, J., et al.: Grounding dino: Marrying dino with grounded pre-training for open-set object detection. arXiv preprint arXiv:2303.05499 (2023) +40. Liu, Z., Lin, Y., Cao, Y., Hu, H., Wei, Y., Zhang, Z., Lin, S., Guo, B.: Swin transformer: Hierarchical vision transformer using shifted windows. In: Proceedings of the IEEE/CVF international conference on computer vision. pp. 10012-10022 (2021) +41. Lu, J., Clark, C., Zellers, R., Mottaghi, R., Kembhavi, A.: Unified-io: A unified model for vision, language, and multi-modal tasks. In: The Eleventh International Conference on Learning Representations (2022) +42. Ma, C., Jiang, Y., Wen, X., Yuan, Z., Qi, X.: Codet: Co-occurrence guided region-word alignment for open-vocabulary object detection. In: Advances in neural information processing systems. vol. 36 (2023) +43. Ma, C., Jiang, Y., Wu, J., Yuan, Z., Qi, X.: Groma: Localized visual tokenization for grounding multimodal large language models. arXiv preprint arXiv:2404.13013 (2024) + +44. Mao, J., Huang, J., Toshev, A., Camburu, O., Yuille, A.L., Murphy, K.: Generation and comprehension of unambiguous object descriptions. In: Proceedings of the IEEE conference on computer vision and pattern recognition. pp. 11-20 (2016) +45. Meletis, P., Wen, X., Lu, C., de Geus, D., Dubbelman, G.: Cityscapes-panoptic-parts and pascal-panoptic-parts datasets for scene understanding. arXiv preprint arXiv:2004.07944 (2020) +46. Michieli, U., Borsato, E., Rossi, L., Zanuttigh, P.: Gmnet: Graph matching network for large scale part semantic segmentation in the wild. In: European Conference on Computer Vision. pp. 397-414. Springer (2020) +47. Miletari, F., Navab, N., Ahmadi, S.A.: V-net: Fully convolutional neural networks for volumetric medical image segmentation. In: 2016 fourth international conference on 3D vision (3DV) (2016) +48. Morabia, K., Arora, J., Vijaykumar, T.: Attention-based joint detection of object and semantic part. arXiv preprint arXiv:2007.02419 (2020) +49. Nair, S., Rajeswaran, A., Kumar, V., Finn, C., Gupta, A.: R3m: A universal visual representation for robot manipulation. In: Conference on Robot Learning. pp. 892-909. PMLR (2023) +50. Ng, X.L., Ong, K.E., Zheng, Q., Ni, Y., Yeo, S.Y., Liu, J.: Animal kingdom: A large and diverse dataset for animal behavior understanding. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 19023-19034 (2022) +51. Pan, T.Y., Liu, Q., Chao, W.L., Price, B.: Towards open-world segmentation of parts. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 15392-15401 (2023) +52. Qi, L., Kuen, J., Guo, W., Gu, J., Lin, Z., Du, B., Xu, Y., Yang, M.H.: Aims: All-inclusive multi-level segmentation for anything. In: Advances in Neural Information Processing Systems. vol. 36 (2023) +53. Radford, A., Kim, J.W., Hallacy, C., Ramesh, A., Goh, G., Agarwal, S., Sastry, G., Askell, A., Mishkin, P., Clark, J., et al.: Learning transferable visual models from natural language supervision. In: International conference on machine learning. pp. 8748-8763. PMLR (2021) +54. Raffel, C., Shazeer, N., Roberts, A., Lee, K., Narang, S., Matena, M., Zhou, Y., Li, W., Liu, P.J.: Exploring the limits of transfer learning with a unified text-to-text transformer. The Journal of Machine Learning Research 21(1), 5485-5551 (2020) +55. Ramanathan, V., Kalia, A., Petrovic, V., Wen, Y., Zheng, B., Guo, B., Wang, R., Marquez, A., Kovvuri, R., Kadian, A., et al.: Paco: Parts and attributes of common objects. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 7141-7151 (2023) +56. Ramesh, A., Dhariwal, P., Nichol, A., Chu, C., Chen, M.: Hierarchical text-conditional image generation with clip latents. arXiv preprint arXiv:2204.06125 1(2), 3 (2022) +57. Ramesh, A., Pavlov, M., Goh, G., Gray, S., Voss, C., Radford, A., Chen, M., Sutskever, I.: Zero-shot text-to-image generation. In: International Conference on Machine Learning. pp. 8821-8831. PMLR (2021) +58. Reddy, N.D., Vo, M., Narasimhan, S.G.: Carfusion: Combining point tracking and part detection for dynamic 3d reconstruction of vehicles. In: Proceedings of the IEEE conference on computer vision and pattern recognition. pp. 1906-1915 (2018) +59. Rezatofighi, H., Tsoi, N., Gwak, J., Sadeghian, A., Reid, I., Savarese, S.: Generalized intersection over union: A metric and a loss for bounding box regression. In: Proceedings of the IEEE/CVF conference on computer vision and pattern recognition. pp. 658-666 (2019) + +60. Rombach, R., Blattmann, A., Lorenz, D., Esser, P., Ommer, B.: High-resolution image synthesis with latent diffusion models. In: Proceedings of the IEEE/CVF conference on computer vision and pattern recognition. pp. 10684-10695 (2022) +61. Shao, S., Li, Z., Zhang, T., Peng, C., Yu, G., Zhang, X., Li, J., Sun, J.: Objects365: A large-scale, high-quality dataset for object detection. In: Proceedings of the IEEE/CVF international conference on computer vision. pp. 8430-8439 (2019) +62. Sun, P., Chen, S., Zhu, C., Xiao, F., Luo, P., Xie, S., Yan, Z.: Going denser with open-vocabulary part segmentation. In: Proceedings of the IEEE/CVF International Conference on Computer Vision. pp. 15453-15465 (2023) +63. Tang, C., Xie, L., Zhang, X., Hu, X., Tian, Q.: Visual recognition by request. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 15265-15274 (2023) +64. Wah, C., Branson, S., Welinder, P., Perona, P., Belongie, S.: Caltech-ucsd birds200-2011 (cub-200-2011). Tech. Rep. CNS-TR-2011-001, California Institute of Technology (2011) +65. Wang, P., Shen, X., Lin, Z., Cohen, S., Price, B., Yuille, A.L.: Joint object and part segmentation using deep learned potentials. In: Proceedings of the IEEE International Conference on Computer Vision. pp. 1573-1581 (2015) +66. Wang, P., Yang, A., Men, R., Lin, J., Bai, S., Li, Z., Ma, J., Zhou, C., Zhou, J., Yang, H.: Ofa: Unifying architectures, tasks, and modalities through a simple sequence-to-sequence learning framework. In: International Conference on Machine Learning. pp. 23318-23340. PMLR (2022) +67. Wang, W., Bao, H., Dong, L., Bjorck, J., Peng, Z., Liu, Q., Aggarwal, K., Mohammed, O.K., Singhal, S., Som, S., et al.: Image as a foreign language: Beit pretraining for all vision and vision-language tasks. arXiv preprint arXiv:2208.10442 (2022) +68. Wang, X., Li, S., Kallidromitis, K., Kato, Y., Kozuka, K., Darrell, T.: Hierarchical open-vocabulary universal image segmentation. In: Advances in Neural Information Processing Systems. vol. 36 (2023) +69. Wang, Y., Xu, Y., Tsogkas, S., Bai, X., Dickinson, S., Siddiqi, K.: Deepflux for skeletons in the wild. In: Proceedings of the IEEE/CVF conference on computer vision and pattern recognition. pp. 5287-5296 (2019) +70. Wei, M., Yue, X., Zhang, W., Kong, S., Liu, X., Pang, J.: Ov-parts: Towards open-vocabulary part segmentation. In: Thirty-seventh Conference on Neural Information Processing Systems Datasets and Benchmarks Track (2023) +71. Wu, J., Jiang, Y., Liu, Q., Yuan, Z., Bai, X., Bai, S.: General object foundation model for images and videos at scale. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 3783-3795 (2024) +72. Xiao, B., Wu, H., Xu, W., Dai, X., Hu, H., Lu, Y., Zeng, M., Liu, C., Yuan, L.: Florence-2: Advancing a unified representation for a variety of vision tasks. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 4818-4829 (2024) +73. Yan, B., Jiang, Y., Sun, P., Wang, D., Yuan, Z., Luo, P., Lu, H.: Towards grand unification of object tracking. In: European Conference on Computer Vision (2022) +74. Yan, B., Jiang, Y., Wu, J., Wang, D., Luo, P., Yuan, Z., Lu, H.: Universal instance perception as object discovery and retrieval. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 15325-15336 (2023) +75. Yang, L., Song, Q., Wang, Z., Jiang, M.: Parsing r-cnn for instance-level human analysis. In: Proceedings of the IEEE/CVF conference on computer vision and pattern recognition. pp. 364-373 (2019) + +76. Yang, Y., Ramanan, D.: Articulated pose estimation with flexible mixtures-of-parts. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 1385-1392. IEEE (2011) +77. Yang, Z., Gan, Z., Wang, J., Hu, X., Ahmed, F., Liu, Z., Lu, Y., Wang, L.: Unitab: Unifying text and box outputs for grounded vision-language modeling. In: European Conference on Computer Vision. pp. 521-539. Springer (2022) +78. Yao, L., Han, J., Liang, X., Xu, D., Zhang, W., Li, Z., Xu, H.: Detclipv2: Scalable open-vocabulary object detection pre-training via word-region alignment. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 23497-23506 (2023) +79. Yao, L., Han, J., Wen, Y., Liang, X., Xu, D., Zhang, W., Li, Z., Xu, C., Xu, H.: Detclip: Dictionary-enriched visual-concept paralleled pre-training for open-world detection. In: Advances in Neural Information Processing Systems. vol. 35, pp. 9125-9138 (2022) +80. Yu, L., Poirson, P., Yang, S., Berg, A.C., Berg, T.L.: Modeling context in referring expressions. In: European Conference on Computer Vision. pp. 69-85. Springer (2016) +81. Yuan, L., Chen, D., Chen, Y.L., Codella, N., Dai, X., Gao, J., Hu, H., Huang, X., Li, B., Li, C., et al.: Florence: A new foundation model for computer vision. arXiv preprint arXiv:2111.11432 (2021) +82. Zareian, A., Rosa, K.D., Hu, D.H., Chang, S.F.: Open-vocabulary object detection using captions. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 14393-14402 (2021) +83. Zhong, Y., Yang, J., Zhang, P., Li, C., Codella, N., Li, L.H., Zhou, L., Dai, X., Yuan, L., Li, Y., et al.: Regionclip: Region-based language-image pretraining. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 16793-16803 (2022) +84. Zhou, B., Zhao, H., Puig, X., Xiao, T., Fidler, S., Barriuso, A., Torralba, A.: Semantic understanding of scenes through the ade20k dataset. International Journal on Computer Vision (2018) +85. Zhou, T., Wang, W., Liu, S., Yang, Y., Van Gool, L.: Differentiable multigranularity human representation learning for instance-aware human semantic parsing. In: Proceedings of the IEEE/CVF conference on computer vision and pattern recognition. pp. 1622-1631 (2021) +86. Zhu, X., Su, W., Lu, L., Li, B., Wang, X., Dai, J.: Deformable detr: Deformable transformers for end-to-end object detection. In: International Conference on Learning Representations (2021) +87. Zhu, X., Zhu, J., Li, H., Wu, X., Li, H., Wang, X., Dai, J.: Uni-perceiver: Pretraining unified architecture for generic perception for zero-shot and few-shot tasks. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 16804-16815 (2022) +88. Ziegler, A., Asano, Y.M.: Self-supervised learning of object parts for semantic segmentation. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 14502-14511 (2022) +89. Zou, X., Dou, Z.Y., Yang, J., Gan, Z., Li, L., Li, C., Dai, X., Behl, H., Wang, J., Yuan, L., et al.: Generalized decoding for pixel, image, and language. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 15116-15127 (2023) +90. Zou, X., Yang, J., Zhang, H., Li, F., Li, L., Wang, J., Wang, L., Gao, J., Lee, Y.J.: Segment everything everywhere all at once. In: Advances in Neural Information Processing Systems. vol. 36 (2023) \ No newline at end of file diff --git a/2024/PartGLEE_ A Foundation Model for Recognizing and Parsing Any Objects/images.zip b/2024/PartGLEE_ A Foundation Model for Recognizing and Parsing Any Objects/images.zip new file mode 100644 index 0000000000000000000000000000000000000000..b7ced70ef1aeec8513ec213ae3b559555e19e52d --- /dev/null +++ b/2024/PartGLEE_ A Foundation Model for Recognizing and Parsing Any Objects/images.zip @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:47c14c0c042c78c44d29d207861c1a67f318e9ae78467b597e61b31e5f07b763 +size 517978 diff --git a/2024/PartGLEE_ A Foundation Model for Recognizing and Parsing Any Objects/layout.json b/2024/PartGLEE_ A Foundation Model for Recognizing and Parsing Any Objects/layout.json new file mode 100644 index 0000000000000000000000000000000000000000..75d7dbfd3d5cd2dde101894a48ce42d03299caa2 --- /dev/null +++ b/2024/PartGLEE_ A Foundation Model for Recognizing and Parsing Any Objects/layout.json @@ -0,0 +1,8785 @@ +{ + "pdf_info": [ + { + "para_blocks": [ + { + "bbox": [ + 136, + 111, + 477, + 148 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 136, + 111, + 477, + 148 + ], + "spans": [ + { + "bbox": [ + 136, + 111, + 477, + 148 + ], + "type": "text", + "content": "PartGLEE: A Foundation Model for Recognizing and Parsing Any Objects" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 153, + 167, + 459, + 181 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 153, + 167, + 459, + 181 + ], + "spans": [ + { + "bbox": [ + 153, + 167, + 459, + 181 + ], + "type": "text", + "content": "Junyi Li\\*, Junfeng Wu\\*, Weizhi Zhao\\*, Song Bai\\*, and Xiang Bai" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 204, + 190, + 408, + 212 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 204, + 190, + 408, + 212 + ], + "spans": [ + { + "bbox": [ + 204, + 190, + 408, + 212 + ], + "type": "text", + "content": "1 Huazhong University of Science and Technology \n2 ByteDance Inc." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 160, + 245, + 452, + 431 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 160, + 245, + 452, + 431 + ], + "spans": [ + { + "bbox": [ + 160, + 245, + 452, + 431 + ], + "type": "text", + "content": "Abstract. We present PartGLEE, a part-level foundation model for locating and identifying both objects and parts in images. Through a unified framework, PartGLEE accomplishes detection, segmentation, and grounding of instances at any granularity in the open world scenario. Specifically, we propose a Q-Former to construct the hierarchical relationship between objects and parts, parsing every object into corresponding semantic parts. By incorporating a large amount of object-level data, the hierarchical relationships can be extended, enabling PartGLEE to recognize a rich variety of parts. We conduct comprehensive studies to validate the effectiveness of our method, PartGLEE achieves the state-of-the-art performance across various part-level tasks and obtain competitive results on object-level tasks. The proposed PartGLEE significantly enhances hierarchical modeling capabilities and part-level perception over our previous GLEE model. Further analysis indicates that the hierarchical cognitive ability of PartGLEE is able to facilitate a detailed comprehension in images for mLLMs. The model and code will be released at https://provencstar.github.io/PartGLEE-Vision/." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 160, + 443, + 452, + 464 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 160, + 443, + 452, + 464 + ], + "spans": [ + { + "bbox": [ + 160, + 443, + 452, + 464 + ], + "type": "text", + "content": "Keywords: Foundation Model " + }, + { + "bbox": [ + 160, + 443, + 452, + 464 + ], + "type": "inline_equation", + "content": "\\cdot" + }, + { + "bbox": [ + 160, + 443, + 452, + 464 + ], + "type": "text", + "content": " Hierarchical Recognition " + }, + { + "bbox": [ + 160, + 443, + 452, + 464 + ], + "type": "inline_equation", + "content": "\\cdot" + }, + { + "bbox": [ + 160, + 443, + 452, + 464 + ], + "type": "text", + "content": " Part Segmentation" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 132, + 487, + 230, + 499 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 487, + 230, + 499 + ], + "spans": [ + { + "bbox": [ + 132, + 487, + 230, + 499 + ], + "type": "text", + "content": "1 Introduction" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 130, + 514, + 482, + 634 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 514, + 482, + 634 + ], + "spans": [ + { + "bbox": [ + 130, + 514, + 482, + 634 + ], + "type": "text", + "content": "In recent years, foundation models have dominated the majority of tasks in the fields of Natural Language Processing [3, 9, 54] and Computer Vision [19, 23, 53, 60, 71, 72]. CLIP family [12, 13, 21, 53, 81] have made significant advancements in transfer learning and have demonstrated impressive zero-shot capabilities on vision-language tasks. SAM [23] has revolutionized the development of segmentation tasks and is able to provide multi-level class-agnostic masks. GLEE [71] utilized diverse object-level data to develop general object representations, enabling detection, segmentation, tracking, grounding, and identification of objects in open-world scenarios. Their remarkable achievement can be attributed to the integration of extensive and diverse range of datasets." + } + ] + } + ], + "index": 6 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 133, + 642, + 481, + 665 + ], + "type": "page_footnote", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 642, + 481, + 665 + ], + "spans": [ + { + "bbox": [ + 133, + 642, + 481, + 665 + ], + "type": "text", + "content": "* Equal Technical Contribution. Work done during Junfeng's internship at ByteDance. †Correspondence to Xiang Bai ." + } + ] + } + ], + "index": 7 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 0 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 140, + 117, + 477, + 411 + ], + "blocks": [ + { + "bbox": [ + 140, + 117, + 477, + 411 + ], + "lines": [ + { + "bbox": [ + 140, + 117, + 477, + 411 + ], + "spans": [ + { + "bbox": [ + 140, + 117, + 477, + 411 + ], + "type": "image", + "image_path": "47f7c330d14fe208579b04eeb8af74ec4ab188c2b13614d4336fbf924c3b0dda.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 132, + 421, + 482, + 465 + ], + "lines": [ + { + "bbox": [ + 132, + 421, + 482, + 465 + ], + "spans": [ + { + "bbox": [ + 132, + 421, + 482, + 465 + ], + "type": "text", + "content": "Fig. 1: An illustrative example demonstrating image annotations at diverse granularities across multiple datasets. The annotations at hierarchical levels with corresponding relationships are depicted on the right side. Below is a visualization of our segmentation results at multiple granularities." + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_caption" + } + ], + "index": 2 + }, + { + "bbox": [ + 130, + 495, + 482, + 638 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 495, + 482, + 638 + ], + "spans": [ + { + "bbox": [ + 130, + 495, + 482, + 638 + ], + "type": "text", + "content": "Different from the vast quantity of object-level data, the scale of part-level data is relatively small, which turns out to be a major bottleneck hindering vision models from recognizing part-level instances. Thus, most vision models lack the hierarchical comprehension between objects and parts. However, it is evident that the ability to recognize parts from objects is essential for various practical applications such as image editing [22,33,38], behavior analysis [50,75], pose estimation [10,76], robotics manipulation [2,49], etc. Moreover, we observe that part-level information is able to help multi-modal Large Language Models (mLLMs) in achieving a more detailed understanding of image content. Since part-level comprehension is a critical ability for foundation models to tackle a broader range of problems, it leads to a natural question: How could we break through data limitations to build a part-level vision foundation model?" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 130, + 641, + 481, + 666 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 641, + 481, + 666 + ], + "spans": [ + { + "bbox": [ + 130, + 641, + 481, + 666 + ], + "type": "text", + "content": "To enable object foundation model with part-level cognitive ability, we emphasize that the model should achieve two key objectives: (1) Hierarchical" + } + ] + } + ], + "index": 5 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "spans": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "type": "text", + "content": "2" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 166, + 91, + 242, + 101 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 166, + 91, + 242, + 101 + ], + "spans": [ + { + "bbox": [ + 166, + 91, + 242, + 101 + ], + "type": "text", + "content": "J. Li, J. Wu, et al." + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 1 + }, + { + "para_blocks": [ + { + "bbox": [ + 133, + 116, + 481, + 379 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 116, + 481, + 379 + ], + "spans": [ + { + "bbox": [ + 133, + 116, + 481, + 379 + ], + "type": "text", + "content": "Comprehension, the model is supposed to understand the intrinsic relationship between objects and parts, and extend this hierarchical connection to any novel object, (2) Semantic Granularity, the model should be capable of learning a universal feature representation, enabling it to recognize semantic instances at any granularity. Consequently, we present a method to jointly detect and segment both objects and parts in a top-down manner. A lightweight Querying Transformer (Q-Former) is proposed to construct the hierarchical relationship between objects and parts. Specifically, it employs a set of universal parsing queries to interact with object queries, consequently generating multiple part-level queries that are capable of predicting corresponding semantic parts for each object. The Q-Former acts as a decomposer, which first recognizes individual objects in the images and subsequently parsing them into parts. Such model design is built upon the observation that various common objects often exhibit shared characteristics of parts. For example, cats, dogs, and dinosaurs all have parts such as torso, legs, and tails. In this way, two sets of query embeddings at different levels are generated, which are then used to predict object-level and part-level instances respectively. Through this approach, the relationship between objects and parts is established via the Q-Former design. Meanwhile, the hierarchical levels of objects and parts are distinguished, which is different from previous research [6,55,62,68,70] that consider parts as fine-grained objects. This paradigm enables vision models to better understand the features on different levels during training, thereby achieving improved performance." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 133, + 391, + 481, + 664 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 391, + 481, + 664 + ], + "spans": [ + { + "bbox": [ + 133, + 391, + 481, + 664 + ], + "type": "text", + "content": "Our complete solution, PartGLEE, for jointly detecting and segmenting instances at both object and part levels, makes it possible for vision models to achieve favorable outcomes on both object and part levels. Some previous research have devised specialized training paradigms to utilize abundant image-text pair data [79, 89] as well as grounding data [30, 39, 71], thereby enhancing the cognitive and generalization capabilities of the models. On the contrary, the quantity of part-level data is much smaller compared to object-level data. So far, the largest dataset incorporating the concepts of both objects and parts is the recently proposed PACO [55] dataset. The scarcity of data has limited research on part-level recognition and restricted the generalization improvement of vision models. Although VLPart [62] has attempted to utilize pseudo-labeling schemes to generate part-level annotations for both object-level and image-level datasets, the quality of the pseudo-labels is relatively poor. Our innovative algorithm that parsing objects into their corresponding parts facilitates the transfer of generalization capability from objects to parts. Consequently, parts are generated from objects, which enables vision models to maintain generalization performance when predicting parts for novel objects without labeling extensive part-level data. To facilitate the training process of Q-Former, we standardize the annotation granularity across various part-level datasets and introduce a vast amount of object-level datasets, an intuitive display of the overall training data is shown in Fig. 1. Unlike VLPart, which exhibits unsatisfactory performance at object-level datasets after joint-training, our method demonstrates favorable outcomes at both object and part levels after joint-training. Moreover, it turns out that" + } + ] + } + ], + "index": 3 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 402, + 91, + 447, + 100 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 402, + 91, + 447, + 100 + ], + "spans": [ + { + "bbox": [ + 402, + 91, + 447, + 100 + ], + "type": "text", + "content": "PartGLEE" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 474, + 92, + 480, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 474, + 92, + 480, + 100 + ], + "spans": [ + { + "bbox": [ + 474, + 92, + 480, + 100 + ], + "type": "text", + "content": "3" + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 2 + }, + { + "para_blocks": [ + { + "bbox": [ + 130, + 116, + 479, + 140 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 116, + 479, + 140 + ], + "spans": [ + { + "bbox": [ + 130, + 116, + 479, + 140 + ], + "type": "text", + "content": "using object-level datasets is able to improve the performance of the model on part-level tasks, indicating a beneficial interaction between objects and parts." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 130, + 140, + 482, + 331 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 140, + 482, + 331 + ], + "spans": [ + { + "bbox": [ + 130, + 140, + 482, + 331 + ], + "type": "text", + "content": "Extensive experiments demonstrate that our method significantly improve the open-vocabulary part segmentation performance, concurrently ensuring a decent performance on object detection and segmentation. We verify its effectiveness on various popular datasets. To validate the generalization performance of our model in identifying various parts of novel objects, we conduct experiments on PartImageNet [18] and Pascal Part [6] datasets in cross-dataset and cross-category manners respectively. Our method exhibits strong transferability and generalization ability when adding extra object-level datasets during training. To evaluate the decomposition capability of our model, we conduct experiments on both ADE20K-Part and Pascal Part datasets follow OV-PARTS [70]. As a result, our approach significantly outperforms one-stage baselines of OV-PARTS, with an increase of " + }, + { + "bbox": [ + 130, + 140, + 482, + 331 + ], + "type": "inline_equation", + "content": "8.16\\%" + }, + { + "bbox": [ + 130, + 140, + 482, + 331 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 130, + 140, + 482, + 331 + ], + "type": "inline_equation", + "content": "2.07\\%" + }, + { + "bbox": [ + 130, + 140, + 482, + 331 + ], + "type": "text", + "content": " on harmonic mean IoU (hIoU) in ADE20K-Part-234 and Pascal-Part-116 respectively. Additionally, by incorporating a large amount of object-level data for joint-training, our method establishes generic hierarchical relationships and breaks through the limitations of scarce part-level data, achieving state-of-the-art performance across various part-level tasks." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 146, + 331, + 448, + 342 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 146, + 331, + 448, + 342 + ], + "spans": [ + { + "bbox": [ + 146, + 331, + 448, + 342 + ], + "type": "text", + "content": "In conclusion, our main contributions can be summarized as follows:" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 135, + 351, + 479, + 479 + ], + "type": "list", + "angle": 0, + "index": 8, + "blocks": [ + { + "bbox": [ + 136, + 351, + 479, + 384 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 136, + 351, + 479, + 384 + ], + "spans": [ + { + "bbox": [ + 136, + 351, + 479, + 384 + ], + "type": "text", + "content": "1. We construct the hierarchical relationship between objects and parts via the Q-Former, facilitating part segmentation to acquire advantages from various object-level datasets." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 135, + 387, + 479, + 432 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 135, + 387, + 479, + 432 + ], + "spans": [ + { + "bbox": [ + 135, + 387, + 479, + 432 + ], + "type": "text", + "content": "2. We propose a unified pipeline for hierarchical detection and segmentation, where we first recognize objects and then parsing them into corresponding semantic parts. This algorithm enables us to jointly detect and segment both object-level and part-level instances." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 135, + 434, + 479, + 479 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 135, + 434, + 479, + 479 + ], + "spans": [ + { + "bbox": [ + 135, + 434, + 479, + 479 + ], + "type": "text", + "content": "3. We standardize the annotation granularity across various part-level datasets by incorporating corresponding object-level annotations, complementing the hierarchical correspondences for current part-level datasets, promoting the development of vision foundation models." + } + ] + } + ], + "index": 7 + } + ], + "sub_type": "text" + }, + { + "bbox": [ + 132, + 501, + 237, + 514 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 501, + 237, + 514 + ], + "spans": [ + { + "bbox": [ + 132, + 501, + 237, + 514 + ], + "type": "text", + "content": "2 Related Work" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 132, + 525, + 399, + 538 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 525, + 399, + 538 + ], + "spans": [ + { + "bbox": [ + 132, + 525, + 399, + 538 + ], + "type": "text", + "content": "2.1 Visual Foundation Models and General Models" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 130, + 545, + 482, + 666 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 545, + 482, + 666 + ], + "spans": [ + { + "bbox": [ + 130, + 545, + 482, + 666 + ], + "type": "text", + "content": "Vision foundation models and generalist models are considered as a milestone in the development of the intelligent vision system. For instance, multi-modal visual foundation models [1, 21, 53, 67, 81] have significantly advanced efficient transfer learning and exhibit impressive zero-shot capabilities on vision-language tasks by using contrastive learning with large-scale image-text pairs. Generative foundation models [11, 56, 57, 60] are trained on vast collections of images and captions, empowering them to generate image content conditioned on textual prompts. Self-supervised foundation models [4, 12, 13, 19] have learned general visual representations from large-scale image datasets, enhancing their ability to transfer to downstream tasks. However, the image-level features learned by these" + } + ] + } + ], + "index": 11 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "spans": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "type": "text", + "content": "4" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 166, + 91, + 242, + 101 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 166, + 91, + 242, + 101 + ], + "spans": [ + { + "bbox": [ + 166, + 91, + 242, + 101 + ], + "type": "text", + "content": "J. Li, J. Wu, et al." + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 3 + }, + { + "para_blocks": [ + { + "bbox": [ + 130, + 116, + 479, + 140 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 116, + 479, + 140 + ], + "spans": [ + { + "bbox": [ + 130, + 116, + 479, + 140 + ], + "type": "text", + "content": "foundation models are not well-suited for direct application to dense prediction tasks that involve precise object and part localization." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 130, + 140, + 482, + 356 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 140, + 482, + 356 + ], + "spans": [ + { + "bbox": [ + 130, + 140, + 482, + 356 + ], + "type": "text", + "content": "Transformer-based generalist methods [5,41,66,77,87] adopt a sequence generation pipeline to unify the output of text and spatial coordinates. However, they mainly focus on image-level comprehension, which results in relatively weak localization capabilities. Works such as UNINEXT, etc. [28, 73, 74], built upon strong detectors [27, 86], demonstrating a strong localization capability across multiple datasets. But they fail to exhibit zero-shot transfer ability and generalization capability due to their closed-set training paradigm. Some works about open-vocabulary detection (OVD) [30, 34, 35, 42, 43, 78, 82, 83] have explored zero-shot generalization capabilities on novel categories. X-Decoder [89] and SEEM [90] have developed a versatile decoding architecture that are able to generate accurate pixel-level segmentation predictions. GLEE [71] addresses various object-level tasks through a unified architecture and training paradigm. However, current generalist models and foundation models are trained mainly on image-level and object-level datasets, thus their ability to extract more fine-grained information is limited, making it difficult for them to recognize corresponding semantic parts of any object. Our work focuses on empowering hierarchical cognitive capability for vision foundation models, thereby further advancing the development of comprehensive visual systems." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 131, + 371, + 256, + 384 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 131, + 371, + 256, + 384 + ], + "spans": [ + { + "bbox": [ + 131, + 371, + 256, + 384 + ], + "type": "text", + "content": "2.2 Part Segmentation" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 130, + 390, + 482, + 668 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 390, + 482, + 668 + ], + "spans": [ + { + "bbox": [ + 130, + 390, + 482, + 668 + ], + "type": "text", + "content": "The growing interest in achieving a more fine-grained understanding of objects has sparked a surge in research focused on part level recognition. Some pioneering studies have introduced datasets with part-level annotations, concentrating on objects of some specific categories such as human body parts [15, 29, 75], animal body parts [64] and vehicle components [58]. More general part annotations for common objects such as Pascal-Part [6], PartImageNet [18], ADE20K [84], CityscapesPanoptic-Parts [45] and more recent PACO [55] were then proposed to promote more in-depth research in the field of parts. Most of the previous works [14, 31, 46, 85] were conducted based on a closed-set configuration, thus only capable of detecting and segmenting closed-set objects and parts. Recently, VLPart [62] present a pipeline for detecting and segmenting both open-vocabulary objects and their corresponding part regions, while OVPARTS [70] utilize adapters to transfer the generalization abilities of CLIP into open-vocabulary part segmentation task. However, due to the limited quantity of data, the generalization capability of previous models [15, 29, 51, 62, 63, 68, 70, 75] still relies heavily on the training datasets. Furthermore, in prior works, both objects and parts are treated equally, they consider part as a special type of object. On the contrary, we distinguish them by considering parts as integral components subordinate to objects and generate parts from corresponding objects in a top-down manner. Our work is aimed at building hierarchical relationships while unifying the training paradigm for object and part-level data. By incorporating a large amount of object-level data, the hierarchical relationships can be extended to any object, enabling our method to recognize a rich variety of parts." + } + ] + } + ], + "index": 5 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 402, + 91, + 447, + 100 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 402, + 91, + 447, + 100 + ], + "spans": [ + { + "bbox": [ + 402, + 91, + 447, + 100 + ], + "type": "text", + "content": "PartGLEE" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 474, + 92, + 481, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 474, + 92, + 481, + 100 + ], + "spans": [ + { + "bbox": [ + 474, + 92, + 481, + 100 + ], + "type": "text", + "content": "5" + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 4 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 153, + 114, + 465, + 270 + ], + "blocks": [ + { + "bbox": [ + 153, + 114, + 465, + 270 + ], + "lines": [ + { + "bbox": [ + 153, + 114, + 465, + 270 + ], + "spans": [ + { + "bbox": [ + 153, + 114, + 465, + 270 + ], + "type": "image", + "image_path": "d2e25db7a14cb03ec2bc08abfeda2e0926d6c4b89e67d93ac90204d9c787ef5a.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 130, + 280, + 482, + 314 + ], + "lines": [ + { + "bbox": [ + 130, + 280, + 482, + 314 + ], + "spans": [ + { + "bbox": [ + 130, + 280, + 482, + 314 + ], + "type": "text", + "content": "Fig. 2: Framework of PartGLEE. The Q-Former takes each object query as input and output the corresponding part queries. These queries are then fed into the object decoder and the part decoder respectively to generate hierarchical predictions." + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_caption" + } + ], + "index": 2 + }, + { + "bbox": [ + 131, + 336, + 378, + 348 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 131, + 336, + 378, + 348 + ], + "spans": [ + { + "bbox": [ + 131, + 336, + 378, + 348 + ], + "type": "text", + "content": "2.3 Hierarchical Learning of Objects and Parts" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 130, + 354, + 482, + 559 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 354, + 482, + 559 + ], + "spans": [ + { + "bbox": [ + 130, + 354, + 482, + 559 + ], + "type": "text", + "content": "Learning objects through parts has been a long-standing research topic as part annotations provide more detailed semantic information of objects. Morabia et al. [48] first introduced a pipeline employing an attention mechanism for simultaneous detection of both objects and parts. Deepflux [69] designed an image context flux representation which enables better object parts interaction for skeleton detection. Leopart [88] demonstrated that learning object parts can provide spatially diverse representation which facilitates self-supervised semantic segmentation. Wang et al. [65] proposed a method to predict both parts and objects simultaneously on Pascal-Part dataset [6]. Recent studies such as SAM [23] and Semantic-SAM [26] have studied on class-agnostic multi-granularity interactive segmentation task. However, they have not explored the relationship between objects and their corresponding semantic parts. Recently, Compositor [17] designed a bottom-up pipeline to predict parts and then cluster them into objects, while AIMS [52] utilized an independent relation decoder to construct the hierarchical association between objects and parts. Different from these works, our approach introduces a Querying Transformer to model the hierarchical relationship, allowing our model to parse any object into its corresponding parts." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 131, + 575, + 202, + 586 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 131, + 575, + 202, + 586 + ], + "spans": [ + { + "bbox": [ + 131, + 575, + 202, + 586 + ], + "type": "text", + "content": "3 Method" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 131, + 598, + 257, + 609 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 131, + 598, + 257, + 609 + ], + "spans": [ + { + "bbox": [ + 131, + 598, + 257, + 609 + ], + "type": "text", + "content": "3.1 Overall Framework" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 130, + 617, + 479, + 640 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 617, + 479, + 640 + ], + "spans": [ + { + "bbox": [ + 130, + 617, + 479, + 640 + ], + "type": "text", + "content": "Following [68,71], we propose PartGLEE, which comprises of an image encoder, a Q-Former, two independent decoders and a text encoder, as shown in Fig. 2." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 130, + 641, + 480, + 665 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 641, + 480, + 665 + ], + "spans": [ + { + "bbox": [ + 130, + 641, + 480, + 665 + ], + "type": "text", + "content": "Given an input image " + }, + { + "bbox": [ + 130, + 641, + 480, + 665 + ], + "type": "inline_equation", + "content": "I \\in \\mathcal{R}^{H \\times W \\times 3}" + }, + { + "bbox": [ + 130, + 641, + 480, + 665 + ], + "type": "text", + "content": ", the backbone and the pixel decoder first extract multi-scale image features " + }, + { + "bbox": [ + 130, + 641, + 480, + 665 + ], + "type": "inline_equation", + "content": "F_s \\in \\mathcal{R}_{\\frac{H}{2^S} \\times \\frac{W}{2^S} \\times C}" + }, + { + "bbox": [ + 130, + 641, + 480, + 665 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 130, + 641, + 480, + 665 + ], + "type": "inline_equation", + "content": "s = \\{2, 3, 4, 5\\}" + }, + { + "bbox": [ + 130, + 641, + 480, + 665 + ], + "type": "text", + "content": " with" + } + ] + } + ], + "index": 9 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "spans": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "type": "text", + "content": "6" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 166, + 91, + 242, + 101 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 166, + 91, + 242, + 101 + ], + "spans": [ + { + "bbox": [ + 166, + 91, + 242, + 101 + ], + "type": "text", + "content": "J. Li, J. Wu, et al." + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 5 + }, + { + "para_blocks": [ + { + "bbox": [ + 130, + 115, + 482, + 319 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 115, + 482, + 319 + ], + "spans": [ + { + "bbox": [ + 130, + 115, + 482, + 319 + ], + "type": "text", + "content": "backbones such as ResNet [20] or Swin Transformer [40]. Then we feed them into the object decoder, where the object-level query embeddings " + }, + { + "bbox": [ + 130, + 115, + 482, + 319 + ], + "type": "inline_equation", + "content": "q_{obj} \\in \\mathcal{R}^{N \\times C}" + }, + { + "bbox": [ + 130, + 115, + 482, + 319 + ], + "type": "text", + "content": " are generated in a two-stage process. These object queries are utilized to perform object-level classification, detection as well as segmentation tasks through three independent prediction heads. Besides, the object queries " + }, + { + "bbox": [ + 130, + 115, + 482, + 319 + ], + "type": "inline_equation", + "content": "q_{obj}" + }, + { + "bbox": [ + 130, + 115, + 482, + 319 + ], + "type": "text", + "content": " are fed into the Q-Former simultaneously, where " + }, + { + "bbox": [ + 130, + 115, + 482, + 319 + ], + "type": "inline_equation", + "content": "L" + }, + { + "bbox": [ + 130, + 115, + 482, + 319 + ], + "type": "text", + "content": " learnable universal parsing queries are initialized to interact with object queries. It takes object queries as input and generate part-level queries " + }, + { + "bbox": [ + 130, + 115, + 482, + 319 + ], + "type": "inline_equation", + "content": "q_{part} \\in \\mathcal{R}^{N \\cdot L \\times C}" + }, + { + "bbox": [ + 130, + 115, + 482, + 319 + ], + "type": "text", + "content": " which are then passed into the part decoder to yield part-level predictions (detailed in Sec. 3.2). To enhance the semantic-awareness, an early fusion module is adopted before Transformer encoder following [74], which takes image feature from backbone and text embedding as input and perform bi-directional cross-attention between them. In line with previous segmentation models [7,27,32], a pixel embedding map " + }, + { + "bbox": [ + 130, + 115, + 482, + 319 + ], + "type": "inline_equation", + "content": "M_p \\in \\mathcal{R}^{\\frac{H}{4} \\times \\frac{W}{4} \\times C}" + }, + { + "bbox": [ + 130, + 115, + 482, + 319 + ], + "type": "text", + "content": " at " + }, + { + "bbox": [ + 130, + 115, + 482, + 319 + ], + "type": "inline_equation", + "content": "1/4" + }, + { + "bbox": [ + 130, + 115, + 482, + 319 + ], + "type": "text", + "content": " resolution is constructed by upsampling and integrating multi-scale feature maps from the backbone and the pixel decoder. Eventually, we dot product each object query or part query with the pixel embedding map to derive an output mask " + }, + { + "bbox": [ + 130, + 115, + 482, + 319 + ], + "type": "inline_equation", + "content": "m \\in \\mathcal{R}^{\\frac{H}{4} \\times \\frac{W}{4}}" + }, + { + "bbox": [ + 130, + 115, + 482, + 319 + ], + "type": "text", + "content": ":" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 224, + 327, + 480, + 338 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 224, + 327, + 480, + 338 + ], + "spans": [ + { + "bbox": [ + 224, + 327, + 480, + 338 + ], + "type": "interline_equation", + "content": "m = F F N (q _ {l}) \\otimes M _ {p}, \\quad l \\in \\{o b j, p a r t \\} \\tag {1}", + "image_path": "3c1a900cfef3e61bc232493b27f690dc3c24cb40203d79e938d50c24d6d42153.jpg" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 130, + 344, + 481, + 369 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 344, + 481, + 369 + ], + "spans": [ + { + "bbox": [ + 130, + 344, + 481, + 369 + ], + "type": "text", + "content": "where FFN is comprised of 3 layers feed forward network with ReLU activation functions and linear layers." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 131, + 384, + 295, + 396 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 131, + 384, + 295, + 396 + ], + "spans": [ + { + "bbox": [ + 131, + 384, + 295, + 396 + ], + "type": "text", + "content": "3.2 Parsing Objects into Parts" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 130, + 401, + 482, + 582 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 401, + 482, + 582 + ], + "spans": [ + { + "bbox": [ + 130, + 401, + 482, + 582 + ], + "type": "text", + "content": "We propose a Q-Former to establish the hierarchical relationship between objects and parts. As Various common objects tend to manifest shared attributes in their constituent parts, for example, both lizards and birds exhibit similar components, such as heads and torsos. Thus, we initialize a set of query embeddings in the Q-Former to parse any object into semantic parts. We denote these universal parsing query embeddings as " + }, + { + "bbox": [ + 130, + 401, + 482, + 582 + ], + "type": "inline_equation", + "content": "q_{parse} \\in \\mathcal{R}^{L \\times C}" + }, + { + "bbox": [ + 130, + 401, + 482, + 582 + ], + "type": "text", + "content": ", where " + }, + { + "bbox": [ + 130, + 401, + 482, + 582 + ], + "type": "inline_equation", + "content": "L" + }, + { + "bbox": [ + 130, + 401, + 482, + 582 + ], + "type": "text", + "content": " represents the number of the parsing queries. As shown in Fig. 2, the Q-Former is comprised of " + }, + { + "bbox": [ + 130, + 401, + 482, + 582 + ], + "type": "inline_equation", + "content": "M" + }, + { + "bbox": [ + 130, + 401, + 482, + 582 + ], + "type": "text", + "content": " cascaded attention modules, each module includes a self-attention layer, a cross-attention layer, and a feed forward network. The universal parsing queries are first fed into the self-attention layer and then perform cross-attention with the object queries. Note that every object query is interacted with all universal parsing queries. Hence, assume " + }, + { + "bbox": [ + 130, + 401, + 482, + 582 + ], + "type": "inline_equation", + "content": "N" + }, + { + "bbox": [ + 130, + 401, + 482, + 582 + ], + "type": "text", + "content": " object queries " + }, + { + "bbox": [ + 130, + 401, + 482, + 582 + ], + "type": "inline_equation", + "content": "(q_{obj} \\in \\mathcal{R}^{N \\times C})" + }, + { + "bbox": [ + 130, + 401, + 482, + 582 + ], + "type": "text", + "content": " are generated from the object decoder, and " + }, + { + "bbox": [ + 130, + 401, + 482, + 582 + ], + "type": "inline_equation", + "content": "L" + }, + { + "bbox": [ + 130, + 401, + 482, + 582 + ], + "type": "text", + "content": " universal parsing queries " + }, + { + "bbox": [ + 130, + 401, + 482, + 582 + ], + "type": "inline_equation", + "content": "q_{parse} \\in \\mathcal{R}^{L \\times C}" + }, + { + "bbox": [ + 130, + 401, + 482, + 582 + ], + "type": "text", + "content": " are initialized in the Q-Former, we obtain " + }, + { + "bbox": [ + 130, + 401, + 482, + 582 + ], + "type": "inline_equation", + "content": "N \\cdot L" + }, + { + "bbox": [ + 130, + 401, + 482, + 582 + ], + "type": "text", + "content": " part-level queries which can be denoted as " + }, + { + "bbox": [ + 130, + 401, + 482, + 582 + ], + "type": "inline_equation", + "content": "q_{part} \\in \\mathcal{R}^{N \\cdot L \\times C}" + }, + { + "bbox": [ + 130, + 401, + 482, + 582 + ], + "type": "text", + "content": ". We refer to this process as:" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 240, + 588, + 480, + 600 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 240, + 588, + 480, + 600 + ], + "spans": [ + { + "bbox": [ + 240, + 588, + 480, + 600 + ], + "type": "interline_equation", + "content": "q _ {p a r t} = Q - F o r m e r \\left(q _ {p a r s e}; q _ {o b j}\\right) \\tag {2}", + "image_path": "17f3d33aeb30196c3c9df1e211a4e11fa9be33a2795576b6453a9b3782f4301a.jpg" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 130, + 605, + 482, + 666 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 605, + 482, + 666 + ], + "spans": [ + { + "bbox": [ + 130, + 605, + 482, + 666 + ], + "type": "text", + "content": "Our proposed Q-Former functions as a decomposer, extracting and representing parts from object queries. Hence, by training jointly on extensive object-level datasets and limited hierarchical datasets which contain object-part correspondences, our Q-Former obtains strong generalization ability to parse any novel object into its corresponding parts." + } + ] + } + ], + "index": 8 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 402, + 91, + 447, + 100 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 402, + 91, + 447, + 100 + ], + "spans": [ + { + "bbox": [ + 402, + 91, + 447, + 100 + ], + "type": "text", + "content": "PartGLEE" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 474, + 92, + 481, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 474, + 92, + 481, + 100 + ], + "spans": [ + { + "bbox": [ + 474, + 92, + 481, + 100 + ], + "type": "text", + "content": "7" + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 6 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 136, + 117, + 476, + 247 + ], + "blocks": [ + { + "bbox": [ + 136, + 117, + 476, + 247 + ], + "lines": [ + { + "bbox": [ + 136, + 117, + 476, + 247 + ], + "spans": [ + { + "bbox": [ + 136, + 117, + 476, + 247 + ], + "type": "image", + "image_path": "9a2fdac931e61e57734068d20d91b4215414125b93fa7a763beb1991e846600d.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 131, + 256, + 482, + 281 + ], + "lines": [ + { + "bbox": [ + 131, + 256, + 482, + 281 + ], + "spans": [ + { + "bbox": [ + 131, + 256, + 482, + 281 + ], + "type": "text", + "content": "Fig. 3: Matching mechanisms of PartGLEE. Two separate forward passes are performed on the same image to obtain hierarchical segmentation results." + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_caption" + } + ], + "index": 2 + }, + { + "bbox": [ + 131, + 299, + 409, + 312 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 131, + 299, + 409, + 312 + ], + "spans": [ + { + "bbox": [ + 131, + 299, + 409, + 312 + ], + "type": "text", + "content": "3.3 Unified Training Paradigm for Objects and Parts" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 130, + 316, + 479, + 352 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 316, + 479, + 352 + ], + "spans": [ + { + "bbox": [ + 130, + 316, + 479, + 352 + ], + "type": "text", + "content": "Since the Q-Former requires hierarchical data to learn how to parse objects into parts, we enrich part-level data with corresponding object-level annotations, details of which are provided in the supplementary materials." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 130, + 352, + 482, + 483 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 352, + 482, + 483 + ], + "spans": [ + { + "bbox": [ + 130, + 352, + 482, + 483 + ], + "type": "text", + "content": "Since the annotation granularity across part-level datasets is standardized, our model can first learn the characteristics of objects and then acquire the ability to parse any object into its semantic parts. To facilitate open-vocabulary detection and segmentation, we substitute the similarity scores between the instance embeddings and the text embeddings for the original class head. Given " + }, + { + "bbox": [ + 130, + 352, + 482, + 483 + ], + "type": "inline_equation", + "content": "K" + }, + { + "bbox": [ + 130, + 352, + 482, + 483 + ], + "type": "text", + "content": " object-level and part-level input categories as separate sentences, we feed them into the text encoder and utilize the average of each individual sentence tokens as the output text embedding " + }, + { + "bbox": [ + 130, + 352, + 482, + 483 + ], + "type": "inline_equation", + "content": "T_{l}" + }, + { + "bbox": [ + 130, + 352, + 482, + 483 + ], + "type": "text", + "content": " for each category. Then the similarity scores " + }, + { + "bbox": [ + 130, + 352, + 482, + 483 + ], + "type": "inline_equation", + "content": "S_{l} \\in \\mathcal{R}^{N \\times K}" + }, + { + "bbox": [ + 130, + 352, + 482, + 483 + ], + "type": "text", + "content": " are calculated through a dot product operation between the hierarchical instance embeddings " + }, + { + "bbox": [ + 130, + 352, + 482, + 483 + ], + "type": "inline_equation", + "content": "q_{l} \\in \\mathcal{R}^{N \\times C}" + }, + { + "bbox": [ + 130, + 352, + 482, + 483 + ], + "type": "text", + "content": " from detector and the text embeddings " + }, + { + "bbox": [ + 130, + 352, + 482, + 483 + ], + "type": "inline_equation", + "content": "T_{l} \\in \\mathcal{R}^{K \\times D}" + }, + { + "bbox": [ + 130, + 352, + 482, + 483 + ], + "type": "text", + "content": " from text encoder, which can be denoted as:" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 224, + 489, + 480, + 502 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 224, + 489, + 480, + 502 + ], + "spans": [ + { + "bbox": [ + 224, + 489, + 480, + 502 + ], + "type": "interline_equation", + "content": "S _ {l} = q _ {l} \\cdot W _ {p r o j} \\otimes T _ {l}, \\quad l \\in \\{o b j, p a r t \\} \\tag {3}", + "image_path": "cf774148c00a7d62b0bfb81199bebb3c2a31299eaf02efc56d7b999691aac177.jpg" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 130, + 507, + 482, + 567 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 507, + 482, + 567 + ], + "spans": [ + { + "bbox": [ + 130, + 507, + 482, + 567 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 130, + 507, + 482, + 567 + ], + "type": "inline_equation", + "content": "W_{proj} \\in \\mathcal{R}^{C \\times D}" + }, + { + "bbox": [ + 130, + 507, + 482, + 567 + ], + "type": "text", + "content": " is a trainable projection weight for fine-tuning text embedding space especially for part-level descriptions. Following [26, 68], we perform Hungarian matching of objects and parts individually, where object-level predictions are only matched with object-level targets, and the same applies to the part-level output, as shown in Fig. 3." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 130, + 568, + 482, + 628 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 568, + 482, + 628 + ], + "spans": [ + { + "bbox": [ + 130, + 568, + 482, + 628 + ], + "type": "text", + "content": "We then introduce a constraint loss to ensure the part-level predictions to be the component of the objects. We denote this novel loss function as restriction loss " + }, + { + "bbox": [ + 130, + 568, + 482, + 628 + ], + "type": "inline_equation", + "content": "L_{res}" + }, + { + "bbox": [ + 130, + 568, + 482, + 628 + ], + "type": "text", + "content": ". Due to memory limitations, we only calculate our restriction loss on the predicted bounding boxes between different levels, while leaving the predicted masks unconstrained. Our restriction loss can be calculated as follow:" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 239, + 630, + 481, + 663 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 239, + 630, + 481, + 663 + ], + "spans": [ + { + "bbox": [ + 239, + 630, + 481, + 663 + ], + "type": "interline_equation", + "content": "L _ {r e s} = \\sum_ {i} ^ {L} \\left(1 - \\frac {\\left| S _ {o b j} \\cap S _ {p a r t} ^ {i} \\right|}{S _ {p a r t} ^ {i}}\\right) \\tag {4}", + "image_path": "b4b53b7f3ee5826ba0a54ca137633e8405b22281b352a1ec05d6aa10ac9adc65.jpg" + } + ] + } + ], + "index": 10 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "spans": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "type": "text", + "content": "8" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 166, + 91, + 242, + 101 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 166, + 91, + 242, + 101 + ], + "spans": [ + { + "bbox": [ + 166, + 91, + 242, + 101 + ], + "type": "text", + "content": "J. Li, J. Wu, et al." + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 7 + }, + { + "para_blocks": [ + { + "bbox": [ + 130, + 116, + 479, + 175 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 116, + 479, + 175 + ], + "spans": [ + { + "bbox": [ + 130, + 116, + 479, + 175 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 130, + 116, + 479, + 175 + ], + "type": "inline_equation", + "content": "S_{obj}" + }, + { + "bbox": [ + 130, + 116, + 479, + 175 + ], + "type": "text", + "content": " represents the area of the object-level bounding box prediction, and " + }, + { + "bbox": [ + 130, + 116, + 479, + 175 + ], + "type": "inline_equation", + "content": "S_{part}^{i}" + }, + { + "bbox": [ + 130, + 116, + 479, + 175 + ], + "type": "text", + "content": " stands for the area of the " + }, + { + "bbox": [ + 130, + 116, + 479, + 175 + ], + "type": "inline_equation", + "content": "i - th" + }, + { + "bbox": [ + 130, + 116, + 479, + 175 + ], + "type": "text", + "content": " part-level bounding box prediction. Note that each object query can generate " + }, + { + "bbox": [ + 130, + 116, + 479, + 175 + ], + "type": "inline_equation", + "content": "L" + }, + { + "bbox": [ + 130, + 116, + 479, + 175 + ], + "type": "text", + "content": " part queries through Q-Former. This loss function is only applied to the matched predictions in part-level datasets, thereby strengthening the mutual correspondence between different hierarchies." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 130, + 176, + 479, + 199 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 176, + 479, + 199 + ], + "spans": [ + { + "bbox": [ + 130, + 176, + 479, + 199 + ], + "type": "text", + "content": "PartGLEE is trained with a linear combination of losses for object-level tasks and part-level tasks, which can be formulated as:" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 144, + 208, + 481, + 224 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 144, + 208, + 481, + 224 + ], + "spans": [ + { + "bbox": [ + 144, + 208, + 481, + 224 + ], + "type": "interline_equation", + "content": "L = \\lambda_ {1} (L _ {c l s} ^ {o b j} + L _ {c l s} ^ {p a r t}) + \\lambda_ {2} (L _ {b o x} ^ {o b j} + L _ {b o x} ^ {p a r t}) + \\lambda_ {3} (L _ {m a s k} ^ {o b j} + L _ {m a s k} ^ {p a r t}) + \\lambda_ {4} L _ {r e s} \\quad (5)", + "image_path": "f0d901387fea6e89de6cb54791cca37b3877d9359100e48453ad183761c47c2f.jpg" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 130, + 232, + 480, + 340 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 232, + 480, + 340 + ], + "spans": [ + { + "bbox": [ + 130, + 232, + 480, + 340 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 130, + 232, + 480, + 340 + ], + "type": "inline_equation", + "content": "L_{cls}^{l}, L_{box}^{l}, L_{mask}^{l}" + }, + { + "bbox": [ + 130, + 232, + 480, + 340 + ], + "type": "text", + "content": " are the classification, box, and mask loss at different levels (" + }, + { + "bbox": [ + 130, + 232, + 480, + 340 + ], + "type": "inline_equation", + "content": "l \\in \\{obj, part\\}" + }, + { + "bbox": [ + 130, + 232, + 480, + 340 + ], + "type": "text", + "content": "), while " + }, + { + "bbox": [ + 130, + 232, + 480, + 340 + ], + "type": "inline_equation", + "content": "L_{res}" + }, + { + "bbox": [ + 130, + 232, + 480, + 340 + ], + "type": "text", + "content": " is the restriction loss, and " + }, + { + "bbox": [ + 130, + 232, + 480, + 340 + ], + "type": "inline_equation", + "content": "\\lambda" + }, + { + "bbox": [ + 130, + 232, + 480, + 340 + ], + "type": "text", + "content": " are their corresponding weights. We apply Focal Loss [36] as the classification loss on the similarity scores " + }, + { + "bbox": [ + 130, + 232, + 480, + 340 + ], + "type": "inline_equation", + "content": "S_{l}" + }, + { + "bbox": [ + 130, + 232, + 480, + 340 + ], + "type": "text", + "content": " to align the text concepts with instance features. A combination of L1 loss and generalized IoU loss [59] is utilized for box predictions, while we employ both Dice Loss [47] and Focal Loss to calculate mask loss. We follow MaskDINO to set our hyperparameters to " + }, + { + "bbox": [ + 130, + 232, + 480, + 340 + ], + "type": "inline_equation", + "content": "\\lambda_{1} = 4, \\lambda_{2} = 2, \\lambda_{3} = 5, \\lambda_{4} = 5" + }, + { + "bbox": [ + 130, + 232, + 480, + 340 + ], + "type": "text", + "content": ". Based on the above designs, PartGLEE is able to leverage both object-level data and part-level data thus obtaining a strong generalization capability." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 132, + 357, + 230, + 371 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 357, + 230, + 371 + ], + "spans": [ + { + "bbox": [ + 132, + 357, + 230, + 371 + ], + "type": "text", + "content": "4 Experiments" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 132, + 383, + 261, + 395 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 383, + 261, + 395 + ], + "spans": [ + { + "bbox": [ + 132, + 383, + 261, + 395 + ], + "type": "text", + "content": "4.1 Experimental Setup" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 130, + 402, + 479, + 425 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 402, + 479, + 425 + ], + "spans": [ + { + "bbox": [ + 130, + 402, + 479, + 425 + ], + "type": "text", + "content": "We conduct comprehensive experiments to exhibit the effectiveness of PartGLEE across a wide range of object-level and part-level tasks." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 130, + 426, + 480, + 544 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 426, + 480, + 544 + ], + "spans": [ + { + "bbox": [ + 130, + 426, + 480, + 544 + ], + "type": "text", + "content": "Data Unification. We utilize object-level datasets such as COCO [37], LVIS [16], Object365 [61], OpenImages [25], Visual Genome [24] and RefCOCO series [44,80], etc, while using part-level datasets PACO [55], PartImageNet [18], Pascal Part [6], ADE20K-Part [70] and SA-1B [23] with varying annotation granularity for joint-training. For Visual Genome and SA-1B, we categorize their corresponding part-level annotations based on semantic and mask overlap relationships to construct hierarchical data versions. For part-level data, we integrate the original part-level annotations with corresponding object-level annotations according to their associated object-level dataset. The details of these dataset preprocessing steps are left in the supplementary materials." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 130, + 545, + 481, + 665 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 545, + 481, + 665 + ], + "spans": [ + { + "bbox": [ + 130, + 545, + 481, + 665 + ], + "type": "text", + "content": "Implementation Details. In our experiments, we utilize ResNet-50 [20] and Swin-Large [40] as the vision encoder. Following MaskDINO [27], we adopt deformable transformer in the decoder, and use 300 object queries while setting the number of parsing queries " + }, + { + "bbox": [ + 130, + 545, + 481, + 665 + ], + "type": "inline_equation", + "content": "L" + }, + { + "bbox": [ + 130, + 545, + 481, + 665 + ], + "type": "text", + "content": " to be 10. The " + }, + { + "bbox": [ + 130, + 545, + 481, + 665 + ], + "type": "inline_equation", + "content": "M" + }, + { + "bbox": [ + 130, + 545, + 481, + 665 + ], + "type": "text", + "content": " of Q-Former is set to 6. We select the top 50 object queries based on the similarity scores and input them into the Q-Former, ultimately yielding 500 part queries. We use both query denoising and hybrid matching strategies to facilitate convergence and enhance performance. We conduct experiments on part-level datasets following the methodologies of VLPart [62] and OV-PARTS [70] in order to evaluate the generalization performance and the ability to parse novel objects of our model. For joint-training, we" + } + ] + } + ], + "index": 10 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 402, + 91, + 447, + 100 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 402, + 91, + 447, + 100 + ], + "spans": [ + { + "bbox": [ + 402, + 91, + 447, + 100 + ], + "type": "text", + "content": "PartGLEE" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 474, + 92, + 481, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 474, + 92, + 481, + 100 + ], + "spans": [ + { + "bbox": [ + 474, + 92, + 481, + 100 + ], + "type": "text", + "content": "9" + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 8 + }, + { + "para_blocks": [ + { + "type": "table", + "bbox": [ + 187, + 178, + 430, + 297 + ], + "blocks": [ + { + "bbox": [ + 132, + 114, + 480, + 158 + ], + "lines": [ + { + "bbox": [ + 132, + 114, + 480, + 158 + ], + "spans": [ + { + "bbox": [ + 132, + 114, + 480, + 158 + ], + "type": "text", + "content": "Table 1: Cross-dataset generalization performance compared with VLPart. The evaluation metric is " + }, + { + "bbox": [ + 132, + 114, + 480, + 158 + ], + "type": "inline_equation", + "content": "mAP_{mask}" + }, + { + "bbox": [ + 132, + 114, + 480, + 158 + ], + "type": "text", + "content": " on the validation set of PartImageNet. All models utilize ResNet-50 as backbone and use the text embeddings of the category names as the classifier. PartImageNet denotes the fully-supervised method reported for comparison." + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 187, + 178, + 430, + 297 + ], + "lines": [ + { + "bbox": [ + 187, + 178, + 430, + 297 + ], + "spans": [ + { + "bbox": [ + 187, + 178, + 430, + 297 + ], + "type": "table", + "html": "
MethodDatasetsAll (40)quadruped
headbodyfoottail
VLPart [62]Pascal Part4.517.40.10.02.9
+ IN-S11 label5.423.63.40.81.2
+ Parsing IN-S117.835.015.23.58.9
vs. baseline+3.3+17.6+15.1+3.5+6.0
PartImageNet29.757.325.822.922.9
PartGLEEPascal Part9.923.64.51.34.6
+ Parsing IN-S1114.955.327.27.023.6
vs. baseline+5.0+31.7+22.7+5.7+19.0
PartImageNet40.267.037.636.540.7
", + "image_path": "7e186bea69deed2a6ad4786fc2254ecb0a7086b8e91bca07792f25fc6652fcd3.jpg" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "table_body" + } + ], + "index": 3 + }, + { + "bbox": [ + 132, + 331, + 480, + 367 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 331, + 480, + 367 + ], + "spans": [ + { + "bbox": [ + 132, + 331, + 480, + 367 + ], + "type": "text", + "content": "train PartGLEE based on the weights of GLEE [71], continuing training on 32 A100 GPUs. The settings for the part-level zero-shot experiments are described separately in each section." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 133, + 388, + 348, + 401 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 388, + 348, + 401 + ], + "spans": [ + { + "bbox": [ + 133, + 388, + 348, + 401 + ], + "type": "text", + "content": "4.2 Zero-Shot Part Segmentation Results" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 132, + 413, + 481, + 568 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 413, + 481, + 568 + ], + "spans": [ + { + "bbox": [ + 132, + 413, + 481, + 568 + ], + "type": "text", + "content": "1) Cross-dataset Part Segmentation on PartImageNet. We follow VLPart [62] to conduct experiments on cross-dataset generalization performance by directly evaluating on PartImageNet [18] validation set. We report the metrics of all (40) part categories and the detailed metrics of quadruped are also provided. The baseline approach only utilize Pascal Part as the training set and directly perform evaluation on PartImageNet in a zero-shot manner. Note that IN-S11 label represents adding image-level classification data for training in order to improve performance. Meanwhile, Parsing IN-S11 stands for training with the pseudo-labels generated from the parsing pipeline proposed by VLPart. However, both of these methods expose the model to categories and images from the PartImageNet dataset. We first perform our training process exclusively on the Pascal Part dataset to verify our zero-shot capabilities, and then we incorporate pseudo-labels to assess the ability of our model to utilize low-quality annotations." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 132, + 570, + 481, + 665 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 570, + 481, + 665 + ], + "spans": [ + { + "bbox": [ + 132, + 570, + 481, + 665 + ], + "type": "text", + "content": "Given that Pascal Part does not provide semantic labels for categories like quadruped in PartImageNet, the model needs to generalize from annotated parts of dog, cat, etc. in Pascal Part to parts of quadruped in PartImageNet. As shown in Tab. 1, our model significantly outperforms VLPart when only training on Pascal Part, even surpassing the model trained with Parsing IN-S11. After incorporating pseudo-labeled data into training, our model shows higher performance gains, indicating better utilization of low-quality data. This result illustrates the importance of hierarchical modeling, which enables our model to recognize and" + } + ] + } + ], + "index": 7 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "spans": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "type": "text", + "content": "10" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 166, + 91, + 241, + 101 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 166, + 91, + 241, + 101 + ], + "spans": [ + { + "bbox": [ + 166, + 91, + 241, + 101 + ], + "type": "text", + "content": "J. Li, J. Wu, et al." + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 9 + }, + { + "para_blocks": [ + { + "type": "table", + "bbox": [ + 138, + 211, + 479, + 320 + ], + "blocks": [ + { + "bbox": [ + 130, + 114, + 482, + 192 + ], + "lines": [ + { + "bbox": [ + 130, + 114, + 482, + 192 + ], + "spans": [ + { + "bbox": [ + 130, + 114, + 482, + 192 + ], + "type": "text", + "content": "Table 2: Cross-category generalization performance compared with VLPart. The evaluation metric is " + }, + { + "bbox": [ + 130, + 114, + 482, + 192 + ], + "type": "inline_equation", + "content": "mAP_{mask}" + }, + { + "bbox": [ + 130, + 114, + 482, + 192 + ], + "type": "text", + "content": " on the validation set of PascalPart but report AP50 specifically for dog parts. All models utilize ResNet-50 as backbone and use the text embeddings of the category names as the classifier. Base part represents the base split from Pascal Part. VOC object is added to the training process to improve the cognitive ability of the model thus reach a better performance. Pascal Part denotes the fully-supervised method reported for comparison." + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 138, + 211, + 479, + 320 + ], + "lines": [ + { + "bbox": [ + 138, + 211, + 479, + 320 + ], + "spans": [ + { + "bbox": [ + 138, + 211, + 479, + 320 + ], + "type": "table", + "html": "
MethodDatasetsAll AP (93)BaseAP (77)NovelAP (16)dogNovelAP Increment
headtorsolegpawtail
VLPart [62]Base Part15.017.81.56.17.92.913.83.2-
+ VOC object16.819.92.129.922.63.212.42.1+0.6
+ IN-S20 label17.420.81.112.817.82.05.90.9-0.4
+ Parsing IN-S2018.421.34.228.734.817.25.714.3+2.7
Pascal Part19.418.822.488.049.638.348.925.8-
PartGLEEBase Part25.630.52.112.615.68.25.26.2-
+ VOC object26.931.25.846.535.027.014.715.1+3.7
+ Parsing IN-S2026.628.915.580.357.336.717.037.4+9.7
Pascal Part35.534.639.995.988.575.076.772.9-
", + "image_path": "d8f86c5fc781e12d68b6fb00a84da010f2213b793afaf9dde6858a4f8fac1a89.jpg" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "table_body" + } + ], + "index": 3 + }, + { + "bbox": [ + 130, + 354, + 479, + 376 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 354, + 479, + 376 + ], + "spans": [ + { + "bbox": [ + 130, + 354, + 479, + 376 + ], + "type": "text", + "content": "parse novel objects into their corresponding parts based on the generalization capability brought by CLIP." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 130, + 378, + 482, + 666 + ], + "type": "list", + "angle": 0, + "index": 7, + "blocks": [ + { + "bbox": [ + 130, + 378, + 482, + 593 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 378, + 482, + 593 + ], + "spans": [ + { + "bbox": [ + 130, + 378, + 482, + 593 + ], + "type": "text", + "content": "2) Cross-category Part Segmentation on Pascal Part. We follow the evaluation setting proposed by VLPart to assess the cross-category generalization performance of our model on the Pascal Part dataset. A total of 93 part categories are divided into 77 base part categories and 16 novel part categories. Tab. 2 presents the evaluation results for all (93), base (77), and novel (16) parts. The model is trained only on the base categories, and is directly evaluated on the entire datasets. Note that IN-S20 label represents adding image-level classification data and Parsing IN-S20 is on behalf of the pseudo-labels generated by VLPart [62] on ImageNet [8]. We further introduce a metric called NovelAP Increment on top of VLPart to assess the improvement of our model when adding extra object datasets into the training process. It is calculated by subtracting the baseline Novel AP from the Novel AP achieved after incorporating extra datasets. The results shown in Tab. 2 demonstrate that our method surpasses the performance of VLPart by a large margin. By comparing the NovelAP Increment, we observe that our method achieves a greater increment after incorporating extra object dataset. Since the VOC dataset includes object categories corresponding to novel parts, the hierarchical relationships of the Q-Former can be extended to novel part categories, resulting in a higher NovelAP Increment." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 130, + 594, + 482, + 666 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 594, + 482, + 666 + ], + "spans": [ + { + "bbox": [ + 130, + 594, + 482, + 666 + ], + "type": "text", + "content": "3) Generalized Zero-Shot Part Segmentation. We adopt the Oracle-Obj setting proposed by OV-PARTS [70] to conduct experiments on ADE-Part234 and Pascal-Part-116 datasets. This setting assumes that the ground-truth masks and categories of object-level instances are known during the inference process, aiming to evaluate the capability of the model to parse any novel object. All categories in the datasets are divided into a base set and a novel set," + } + ] + } + ], + "index": 6 + } + ], + "sub_type": "text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 402, + 91, + 447, + 100 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 402, + 91, + 447, + 100 + ], + "spans": [ + { + "bbox": [ + 402, + 91, + 447, + 100 + ], + "type": "text", + "content": "PartGLEE" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 470, + 91, + 479, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 470, + 91, + 479, + 100 + ], + "spans": [ + { + "bbox": [ + 470, + 91, + 479, + 100 + ], + "type": "text", + "content": "11" + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 10 + }, + { + "para_blocks": [ + { + "type": "table", + "bbox": [ + 138, + 156, + 477, + 267 + ], + "blocks": [ + { + "bbox": [ + 132, + 114, + 480, + 137 + ], + "lines": [ + { + "bbox": [ + 132, + 114, + 480, + 137 + ], + "spans": [ + { + "bbox": [ + 132, + 114, + 480, + 137 + ], + "type": "text", + "content": "Table 3: Generalized zero-Shot part segmentation performance on ADE-Part-234 and Pascal-Part-116 compared with baselines proposed by OV-PARTS." + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 138, + 156, + 477, + 267 + ], + "lines": [ + { + "bbox": [ + 138, + 156, + 477, + 267 + ], + "spans": [ + { + "bbox": [ + 138, + 156, + 477, + 267 + ], + "type": "table", + "html": "
MethodModelBackboneFinetuningOracle-Obj
ADE-Part-234Oracle-obj
SeenUnseenHarmonicSeenUnseenHarmonic
Fully SupervisedMask2FormerResNet-50X46.2547.86-55.2852.14-
Two-StageZSeg+ResNet-50CPTCoOp43.1927.8433.8555.3319.1728.48
ResNet-50CPTCoCoOp39.6725.1530.7854.4319.0428.21
ResNet-101cCPTCoOp43.4125.7032.2857.8821.9331.81
One-StageCATSegResNet-101&ViT-B/16X11.498.569.8114.8910.2912.17
ResNet-101&ViT-B/16B+D31.4025.7728.3143.9726.1132.76
CLIPSegViT-B/16X15.2718.0116.5322.3319.7320.95
ViT-B/16VA+L+F+D38.9629.6533.6748.6827.3735.04
PartGLEEResNet-50X51.2935.3341.8357.4327.4137.11
", + "image_path": "ef6ba8175a945b50696fe6cad2e37606c3f9ebc9f0f803144782ac6208b0d3e5.jpg" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "table_body" + } + ], + "index": 3 + }, + { + "bbox": [ + 130, + 304, + 480, + 389 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 304, + 480, + 389 + ], + "spans": [ + { + "bbox": [ + 130, + 304, + 480, + 389 + ], + "type": "text", + "content": "and the training process is performed only on the base set, while we evaluate the performance of the model on all categories. As shown in Tab. 3, our model achieves a superior performance on both datasets, which indicates the importance of hierarchical modeling. The establishment of hierarchical relationships between objects and parts enables our model to extend to novel objects, thereby effectively parsing them into corresponding semantic parts. Consequently, our model exhibits outstanding performance across both datasets." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 132, + 418, + 428, + 430 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 418, + 428, + 430 + ], + "spans": [ + { + "bbox": [ + 132, + 418, + 428, + 430 + ], + "type": "text", + "content": "4.3 Joint-training results on Detection and Segmentation" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 130, + 450, + 481, + 666 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 450, + 481, + 666 + ], + "spans": [ + { + "bbox": [ + 130, + 450, + 481, + 666 + ], + "type": "text", + "content": "To endow our model with robust generalization capability, we perform joint training on various datasets and evaluate its performance on both object-level and part-level tasks. We compare our model with specialist and generalist models to evaluate its performance on object-level data. Additionally, we contrast it with VLPart to assess its performance on part-level datasets as well as the effectiveness of joint-training process on both types of datasets. As shown in Tab. 4, PartGLEE significantly outperforms VLPart on both object-level and part-level tasks after joint-training, while achieving comparable performance on object-level tasks compared with previous SOTA. Through joint-training, our model has acquired strong generalization performance, allowing it to simultaneously address tasks for different hierarchies. We also observe that VLPart fails to achieve satisfactory performance on both object-level and part-level tasks. For example, VLPart obtains better performance on Pascal Part than its dataset-specific oracle, while decreasing its performance on COCO and LVIS. We attribute the performance drop of VLPart to the absence of hierarchical relationships, which causes confusion in modeling parts and objects and impairs object-level performance. PartGLEE effectively addresses this problem and extends the generalization capabilities from object-level to part-level tasks." + } + ] + } + ], + "index": 6 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "spans": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "type": "text", + "content": "12" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 166, + 91, + 241, + 101 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 166, + 91, + 241, + 101 + ], + "spans": [ + { + "bbox": [ + 166, + 91, + 241, + 101 + ], + "type": "text", + "content": "J. Li, J. Wu, et al." + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 11 + }, + { + "para_blocks": [ + { + "type": "table", + "bbox": [ + 136, + 178, + 480, + 350 + ], + "blocks": [ + { + "bbox": [ + 132, + 114, + 482, + 159 + ], + "lines": [ + { + "bbox": [ + 132, + 114, + 482, + 159 + ], + "spans": [ + { + "bbox": [ + 132, + 114, + 482, + 159 + ], + "type": "text", + "content": "Table 4: Joint-Training Performance of PartGLEE. Note that Oracle represents the dataset-specific training paradigm. We directly evaluate the generalist models on PACO to assess their recognition capability at the part level, as indicated by the results annotated in the grey font." + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 136, + 178, + 480, + 350 + ], + "lines": [ + { + "bbox": [ + 136, + 178, + 480, + 350 + ], + "spans": [ + { + "bbox": [ + 136, + 178, + 480, + 350 + ], + "type": "table", + "html": "
TypeMethodPart-level TasksObject-level Tasks
PartImageNetPascal PartPACOCOCO-valLVIS-minimalLVIS-val
APboxAPmaskAPboxAPmaskAPmaskAPobjmaskAPpartmaskAPboxAPmaskAPboxAPmaskAPbox
SpecialistMask2Former(R50) [7]------46.243.7----
Mask2Former(L) [7]------50.1-----
MaskDINO(R50) [27]------50.546.0----
MaskDINO(L) [27]------58.352.1----
ViTDet-L [32]------57.649.8--51.246.0
ViTDet-H [32]------57.649.8--53.448.1
EVA-02-L [12]------64.255.0--65.257.3
PACO(R50) [55]----32.612.5------
PACO(L) [55]----43.417.7------
GeneralistPix2Seq v2 [5]------46.538.2----
X-Decoder(L) [89]---2.6911.90.94-46.7----
SEEM(L) [90]---1.998.420.69-47.7----
HIPIE(R50) [68]------53.945.9----
Florence-2(B) [72]------41.4-----
Florence-2(L) [72]------43.4-----
UNINEXT(R50) [74]------51.344.9--36.4-
UNINEXT(L) [74]------58.149.6----
GLEE(R50) [71]---3.4415.31.2955.048.450.545.944.240.2
HierarchicalVLPart(R50) [62]30.731.623.924.013.836.99.628.5--26.2-
VLPart(R50)-Oracle [62]29.229.718.919.413.328.010.638.0--28.1-
VLPart(B) [62]43.941.233.531.722.155.015.940.3--39.6-
VLPart(B)-Oracle [62]44.341.729.227.419.137.715.252.5--43.1-
PartGLEE (R50)40.940.235.035.521.850.515.454.447.648.743.542.7
PartGLEE (L)52.750.939.639.127.855.721.359.552.056.550.645.0
", + "image_path": "9d0326b2d8877f17c4f15a4d1aed5717cf1b7ec8e84eecc7c4ad8a08481013cd.jpg" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "table_body" + } + ], + "index": 3 + }, + { + "bbox": [ + 132, + 384, + 238, + 396 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 384, + 238, + 396 + ], + "spans": [ + { + "bbox": [ + 132, + 384, + 238, + 396 + ], + "type": "text", + "content": "4.4 Ablation Study" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 133, + 407, + 482, + 586 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 407, + 482, + 586 + ], + "spans": [ + { + "bbox": [ + 133, + 407, + 482, + 586 + ], + "type": "text", + "content": "To demonstrate that our model design achieves satisfactory results on both object-level and part-level tasks, we conduct an ablation study (depicted in Fig. 4) on the model architecture and present results in Tab. 5. We ablate with a backbone of ResNet-50 and perform joint-training on COCO [37], LVIS [16], PartImageNet [18], Pascal Part [6] and PACO [55] with 90K iterations. From this study, we draw several important conclusions: (1) The utilization of parallel pixel decoders only results in slight improvements in mask predictions on few datasets, indicating that the influence of feature maps at different granularities is negligible. (2) Adopting independent decoders to obtain predictions at different levels demonstrates superior performance across the majority of datasets, manifesting the effectiveness of independent decoders. As adopting parallel pixel decoders (b) results in significant GPU memory costs without considerable gains, and all metrics for (a) are lower than (c), we select (c) as our final model design. Additional ablation studies, extensive qualitative analysis, and experiments on mLLM can be found in the supplementary materials." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 132, + 607, + 218, + 618 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 607, + 218, + 618 + ], + "spans": [ + { + "bbox": [ + 132, + 607, + 218, + 618 + ], + "type": "text", + "content": "4.5 Limitations" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 132, + 629, + 481, + 665 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 629, + 481, + 665 + ], + "spans": [ + { + "bbox": [ + 132, + 629, + 481, + 665 + ], + "type": "text", + "content": "In this work, we still adopt CLIP as the text encoder, which is trained on text-image pairs and thus lacks the ability to perceive fine-grained descriptions of object or part instances. This limitation may restrict the improvement of model" + } + ] + } + ], + "index": 7 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 402, + 91, + 447, + 100 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 402, + 91, + 447, + 100 + ], + "spans": [ + { + "bbox": [ + 402, + 91, + 447, + 100 + ], + "type": "text", + "content": "PartGLEE" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 470, + 91, + 481, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 470, + 91, + 481, + 100 + ], + "spans": [ + { + "bbox": [ + 470, + 91, + 481, + 100 + ], + "type": "text", + "content": "13" + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 12 + }, + { + "para_blocks": [ + { + "type": "table", + "bbox": [ + 136, + 189, + 480, + 247 + ], + "blocks": [ + { + "bbox": [ + 132, + 114, + 482, + 170 + ], + "lines": [ + { + "bbox": [ + 132, + 114, + 482, + 170 + ], + "spans": [ + { + "bbox": [ + 132, + 114, + 482, + 170 + ], + "type": "text", + "content": "Table 5: An ablation study on different model designs, as depicted in Fig 4. Note that Parallel Pixel Decoders refers to the utilization of two pixel decoders to generate feature maps at different hierarchies respectively. Independent Decoders denote the usage of two decoders, which facilitate the interaction between feature maps and queries at different hierarchies. Our final choice is scheme (c), which is highlighted in gray." + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 136, + 189, + 480, + 247 + ], + "lines": [ + { + "bbox": [ + 136, + 189, + 480, + 247 + ], + "spans": [ + { + "bbox": [ + 136, + 189, + 480, + 247 + ], + "type": "table", + "html": "
SchemeModel DesignPart-level TasksObject-level Tasks
Parrallel Pixel DecodersIndependent DecodersPartImageNetPascal PartPACOCOCO-valLVIS-minimal
APmaskAPmaskAPmaskAPmaskAPobjmaskAPopartmaskAPboxAPmaskAPboxAPmask
(a)XX39.034.120.147.413.547.843.534.833.4
(b)✓X38.334.520.848.813.848.544.334.934.2
(c)X✓39.034.720.947.914.249.344.235.633.8
", + "image_path": "380a1a8ccedf3ac0c73b931153dde8949c159fc9a688bdbba265db0f4c15f336.jpg" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "table_body" + } + ], + "index": 3 + }, + { + "type": "image", + "bbox": [ + 137, + 281, + 247, + 353 + ], + "blocks": [ + { + "bbox": [ + 136, + 269, + 146, + 277 + ], + "lines": [ + { + "bbox": [ + 136, + 269, + 146, + 277 + ], + "spans": [ + { + "bbox": [ + 136, + 269, + 146, + 277 + ], + "type": "text", + "content": "(a)" + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 137, + 281, + 247, + 353 + ], + "lines": [ + { + "bbox": [ + 137, + 281, + 247, + 353 + ], + "spans": [ + { + "bbox": [ + 137, + 281, + 247, + 353 + ], + "type": "image", + "image_path": "4dcd8f982810a9b1fe787c65d99b5e0e3b46a23f84bf7f62b07359704c85f0f0.jpg" + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "image_body" + } + ], + "index": 5 + }, + { + "type": "image", + "bbox": [ + 252, + 277, + 361, + 357 + ], + "blocks": [ + { + "bbox": [ + 252, + 268, + 261, + 276 + ], + "lines": [ + { + "bbox": [ + 252, + 268, + 261, + 276 + ], + "spans": [ + { + "bbox": [ + 252, + 268, + 261, + 276 + ], + "type": "text", + "content": "(b)" + } + ] + } + ], + "index": 6, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 252, + 277, + 361, + 357 + ], + "lines": [ + { + "bbox": [ + 252, + 277, + 361, + 357 + ], + "spans": [ + { + "bbox": [ + 252, + 277, + 361, + 357 + ], + "type": "image", + "image_path": "1974bef7631da668f780e78acdb603bc1363bc5433c6727de90ac6ef375762d6.jpg" + } + ] + } + ], + "index": 7, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 365, + 269, + 375, + 277 + ], + "lines": [ + { + "bbox": [ + 365, + 269, + 375, + 277 + ], + "spans": [ + { + "bbox": [ + 365, + 269, + 375, + 277 + ], + "type": "text", + "content": "(c)" + } + ] + } + ], + "index": 8, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 130, + 375, + 482, + 441 + ], + "lines": [ + { + "bbox": [ + 130, + 375, + 482, + 441 + ], + "spans": [ + { + "bbox": [ + 130, + 375, + 482, + 441 + ], + "type": "text", + "content": "Fig. 4: Various designs for generating predictions at different hierarchies. In scheme (a), we only utilize a single decoder to generate predictions for both objects and parts. In scheme (b), two parallel pixel decoders are employed to generate feature maps at different levels, aiming to explore the effectiveness of feature maps at different granularity. In scheme (c), we use two independent decoders to generate predictions for objects and parts respectively." + } + ] + } + ], + "index": 10, + "angle": 0, + "type": "image_caption" + } + ], + "index": 7 + }, + { + "type": "image", + "bbox": [ + 365, + 281, + 474, + 351 + ], + "blocks": [ + { + "bbox": [ + 365, + 281, + 474, + 351 + ], + "lines": [ + { + "bbox": [ + 365, + 281, + 474, + 351 + ], + "spans": [ + { + "bbox": [ + 365, + 281, + 474, + 351 + ], + "type": "image", + "image_path": "5f24408bc94c656ce390a6f85c1be5a280f2124a8feaa04d680d885d7a1149f2.jpg" + } + ] + } + ], + "index": 9, + "angle": 0, + "type": "image_body" + } + ], + "index": 9 + }, + { + "bbox": [ + 130, + 468, + 481, + 491 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 468, + 481, + 491 + ], + "spans": [ + { + "bbox": [ + 130, + 468, + 481, + 491 + ], + "type": "text", + "content": "performance and prompts us to consider how to enhance the perception capabilities of region-level models, which will be our future work." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 132, + 515, + 220, + 528 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 515, + 220, + 528 + ], + "spans": [ + { + "bbox": [ + 132, + 515, + 220, + 528 + ], + "type": "text", + "content": "5 Conclusion" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 130, + 545, + 482, + 666 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 545, + 482, + 666 + ], + "spans": [ + { + "bbox": [ + 130, + 545, + 482, + 666 + ], + "type": "text", + "content": "In this paper, we introduce PartGLEE, a groundbreaking foundation model designed towards a complete comprehension of both objects and parts in images. Through the generic hierarchical relationships established by the Q-Former, we are able to break through the limitation of scarce part-level data by introducing a large amount of object-level data, thereby transferring the powerful generalization capabilities from objects to parts. Through extensive training on diverse datasets, PartGLEE achieves SOTA performance across various part-level tasks while maintaining competitive results on object-level tasks, enabling it to parse any objects into parts and serve as a foundation model for general fine-grained region-level perception tasks." + } + ] + } + ], + "index": 13 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "spans": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "type": "text", + "content": "14" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 166, + 91, + 242, + 101 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 166, + 91, + 242, + 101 + ], + "spans": [ + { + "bbox": [ + 166, + 91, + 242, + 101 + ], + "type": "text", + "content": "J. Li, J. Wu, et al." + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 13 + }, + { + "para_blocks": [ + { + "bbox": [ + 133, + 114, + 246, + 129 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 114, + 246, + 129 + ], + "spans": [ + { + "bbox": [ + 133, + 114, + 246, + 129 + ], + "type": "text", + "content": "Acknowledgements" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 132, + 137, + 480, + 162 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 137, + 480, + 162 + ], + "spans": [ + { + "bbox": [ + 132, + 137, + 480, + 162 + ], + "type": "text", + "content": "This work was supported by the National Science Fund for Distinguished Young Scholars of China (Grant No.62225603)." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 133, + 178, + 197, + 190 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 178, + 197, + 190 + ], + "spans": [ + { + "bbox": [ + 133, + 178, + 197, + 190 + ], + "type": "text", + "content": "References" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 138, + 200, + 481, + 665 + ], + "type": "list", + "angle": 0, + "index": 18, + "blocks": [ + { + "bbox": [ + 138, + 200, + 481, + 243 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 200, + 481, + 243 + ], + "spans": [ + { + "bbox": [ + 138, + 200, + 481, + 243 + ], + "type": "text", + "content": "1. Alayrac, J.B., Donahue, J., Luc, P., Miech, A., Barr, I., Hasson, Y., Lenc, K., Mensch, A., Millican, K., Reynolds, M., et al.: Flamingo: a visual language model for few-shot learning. In: Advances in neural information processing systems. vol. 35, pp. 23716-23736 (2022)" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 138, + 244, + 481, + 276 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 244, + 481, + 276 + ], + "spans": [ + { + "bbox": [ + 138, + 244, + 481, + 276 + ], + "type": "text", + "content": "2. Brohan, A., Brown, N., Carbajal, J., Chebotar, Y., Dabis, J., Finn, C., Gopalakrishnan, K., Hausman, K., Herzog, A., Hsu, J., et al.: Rt-1: Robotics transformer for real-world control at scale. arXiv preprint arXiv:2212.06817 (2022)" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 138, + 277, + 481, + 319 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 277, + 481, + 319 + ], + "spans": [ + { + "bbox": [ + 138, + 277, + 481, + 319 + ], + "type": "text", + "content": "3. Brown, T., Mann, B., Ryder, N., Subbiah, M., Kaplan, J.D., Dhariwal, P., Neelakantan, A., Shyam, P., Sastry, G., Askell, A., et al.: Language models are few-shot learners. In: Advances in neural information processing systems. vol. 33, pp. 1877-1901 (2020)" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 138, + 320, + 481, + 352 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 320, + 481, + 352 + ], + "spans": [ + { + "bbox": [ + 138, + 320, + 481, + 352 + ], + "type": "text", + "content": "4. Caron, M., Touvron, H., Misra, I., Jégou, H., Mairal, J., Bojanowski, P., Joulin, A.: Emerging properties in self-supervised vision transformers. In: Proceedings of the IEEE/CVF international conference on computer vision. pp. 9650-9660 (2021)" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 138, + 352, + 481, + 384 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 352, + 481, + 384 + ], + "spans": [ + { + "bbox": [ + 138, + 352, + 481, + 384 + ], + "type": "text", + "content": "5. Chen, T., Saxena, S., Li, L., Lin, T.Y., Fleet, D.J., Hinton, G.E.: A unified sequence interface for vision tasks. Advances in Neural Information Processing Systems 35, 31333-31346 (2022)" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 138, + 384, + 481, + 427 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 384, + 481, + 427 + ], + "spans": [ + { + "bbox": [ + 138, + 384, + 481, + 427 + ], + "type": "text", + "content": "6. Chen, X., Mottaghi, R., Liu, X., Fidler, S., Urtasun, R., Yuille, A.: Detect what you can: Detecting and representing objects using holistic models and body parts. In: Proceedings of the IEEE conference on computer vision and pattern recognition. pp. 1971-1978 (2014)" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 138, + 428, + 481, + 471 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 428, + 481, + 471 + ], + "spans": [ + { + "bbox": [ + 138, + 428, + 481, + 471 + ], + "type": "text", + "content": "7. Cheng, B., Misra, I., Schwing, A.G., Kirillov, A., Girdhar, R.: Masked-attention mask transformer for universal image segmentation. In: Proceedings of the IEEE/CVF conference on computer vision and pattern recognition. pp. 1290-1299 (2022)" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 138, + 472, + 481, + 503 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 472, + 481, + 503 + ], + "spans": [ + { + "bbox": [ + 138, + 472, + 481, + 503 + ], + "type": "text", + "content": "8. Deng, J., Dong, W., Socher, R., Li, L.J., Li, K., Fei-Fei, L.: Imagenet: A large-scale hierarchical image database. In: Proceedings of the IEEE/CVF International Conference on Computer Vision. pp. 248-255 (2009)" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 138, + 503, + 481, + 535 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 503, + 481, + 535 + ], + "spans": [ + { + "bbox": [ + 138, + 503, + 481, + 535 + ], + "type": "text", + "content": "9. Devlin, J., Chang, M.W., Lee, K., Toutanova, K.: Bert: Pre-training of deep bidirectional transformers for language understanding. arXiv preprint arXiv:1810.04805 (2018)" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 138, + 536, + 481, + 567 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 536, + 481, + 567 + ], + "spans": [ + { + "bbox": [ + 138, + 536, + 481, + 567 + ], + "type": "text", + "content": "0. Dong, J., Chen, Q., Shen, X., Yang, J., Yan, S.: Towards unified human parsing and pose estimation. In: Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition. pp. 843-850 (2014)" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 138, + 568, + 481, + 601 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 568, + 481, + 601 + ], + "spans": [ + { + "bbox": [ + 138, + 568, + 481, + 601 + ], + "type": "text", + "content": "1. Esser, P., Rombach, R., Ommer, B.: Taming transformers for high-resolution image synthesis. In: Proceedings of the IEEE/CVF conference on computer vision and pattern recognition. pp. 12873-12883 (2021)" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 138, + 601, + 481, + 621 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 601, + 481, + 621 + ], + "spans": [ + { + "bbox": [ + 138, + 601, + 481, + 621 + ], + "type": "text", + "content": "2. Fang, Y., Sun, Q., Wang, X., Huang, T., Wang, X., Cao, Y.: Eva-02: A visual representation for neon genesis. arXiv preprint arXiv:2303.11331 (2023)" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 138, + 622, + 481, + 665 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 622, + 481, + 665 + ], + "spans": [ + { + "bbox": [ + 138, + 622, + 481, + 665 + ], + "type": "text", + "content": "3. Fang, Y., Wang, W., Xie, B., Sun, Q., Wu, L., Wang, X., Huang, T., Wang, X., Cao, Y.: Eva: Exploring the limits of masked visual representation learning at scale. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 19358-19369 (2023)" + } + ] + } + ], + "index": 17 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 402, + 91, + 447, + 100 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 402, + 91, + 447, + 100 + ], + "spans": [ + { + "bbox": [ + 402, + 91, + 447, + 100 + ], + "type": "text", + "content": "PartGLEE" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 470, + 92, + 481, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 470, + 92, + 481, + 100 + ], + "spans": [ + { + "bbox": [ + 470, + 92, + 481, + 100 + ], + "type": "text", + "content": "15" + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 14 + }, + { + "para_blocks": [ + { + "bbox": [ + 132, + 116, + 481, + 665 + ], + "type": "list", + "angle": 0, + "index": 16, + "blocks": [ + { + "bbox": [ + 133, + 116, + 480, + 149 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 116, + 480, + 149 + ], + "spans": [ + { + "bbox": [ + 133, + 116, + 480, + 149 + ], + "type": "text", + "content": "14. de Geus, D., Meletis, P., Lu, C., Wen, X., Dubbelman, G.: Part-aware panoptic segmentation. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 5485-5494 (2021)" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 133, + 150, + 481, + 194 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 150, + 481, + 194 + ], + "spans": [ + { + "bbox": [ + 133, + 150, + 481, + 194 + ], + "type": "text", + "content": "15. Gong, K., Liang, X., Zhang, D., Shen, X., Lin, L.: Look into person: Self-supervised structure-sensitive learning and a new benchmark for human parsing. In: Proceedings of the IEEE conference on computer vision and pattern recognition. pp. 932-940 (2017)" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 133, + 194, + 480, + 226 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 194, + 480, + 226 + ], + "spans": [ + { + "bbox": [ + 133, + 194, + 480, + 226 + ], + "type": "text", + "content": "16. Gupta, A., Dollar, P., Girshick, R.: Lvis: A dataset for large vocabulary instance segmentation. In: Proceedings of the IEEE/CVF conference on computer vision and pattern recognition. pp. 5356-5364 (2019)" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 132, + 227, + 480, + 270 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 227, + 480, + 270 + ], + "spans": [ + { + "bbox": [ + 132, + 227, + 480, + 270 + ], + "type": "text", + "content": "17. He, J., Chen, J., Lin, M.X., Yu, Q., Yuille, A.L.: Compositor: Bottom-up clustering and compositing for robust part and object segmentation. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 11259-11268 (2023)" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 132, + 271, + 480, + 304 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 271, + 480, + 304 + ], + "spans": [ + { + "bbox": [ + 132, + 271, + 480, + 304 + ], + "type": "text", + "content": "18. He, J., Yang, S., Yang, S., Kortylewski, A., Yuan, X., Chen, J.N., Liu, S., Yang, C., Yu, Q., Yuille, A.: Partimagenet: A large, high-quality dataset of parts. In: European Conference on Computer Vision. pp. 128-145. Springer (2022)" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 132, + 304, + 480, + 336 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 304, + 480, + 336 + ], + "spans": [ + { + "bbox": [ + 132, + 304, + 480, + 336 + ], + "type": "text", + "content": "19. He, K., Chen, X., Xie, S., Li, Y., Dólár, P., Girshick, R.: Masked autoencoders are scalable vision learners. In: Proceedings of the IEEE/CVF conference on computer vision and pattern recognition. pp. 16000-16009 (2022)" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 132, + 336, + 480, + 369 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 336, + 480, + 369 + ], + "spans": [ + { + "bbox": [ + 132, + 336, + 480, + 369 + ], + "type": "text", + "content": "20. He, K., Zhang, X., Ren, S., Sun, J.: Deep residual learning for image recognition. In: Proceedings of the IEEE conference on computer vision and pattern recognition. pp. 770-778 (2016)" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 132, + 369, + 480, + 413 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 369, + 480, + 413 + ], + "spans": [ + { + "bbox": [ + 132, + 369, + 480, + 413 + ], + "type": "text", + "content": "21. Jia, C., Yang, Y., Xia, Y., Chen, Y.T., Parekh, Z., Pham, H., Le, Q., Sung, Y.H., Li, Z., Duerig, T.: Scaling up visual and vision-language representation learning with noisy text supervision. In: International conference on machine learning. pp. 4904-4916. PMLR (2021)" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 132, + 414, + 480, + 456 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 414, + 480, + 456 + ], + "spans": [ + { + "bbox": [ + 132, + 414, + 480, + 456 + ], + "type": "text", + "content": "22. Kawar, B., Zada, S., Lang, O., Tov, O., Chang, H., Dekel, T., Mosseri, I., Irani, M.: Imagic: Text-based real image editing with diffusion models. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 6007-6017 (2023)" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 132, + 457, + 480, + 500 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 457, + 480, + 500 + ], + "spans": [ + { + "bbox": [ + 132, + 457, + 480, + 500 + ], + "type": "text", + "content": "23. Kirillov, A., Mintun, E., Ravi, N., Mao, H., Rolland, C., Gustafson, L., Xiao, T., Whitehead, S., Berg, A.C., Lo, W.Y., et al.: Segment anything. In: Proceedings of the IEEE/CVF International Conference on Computer Vision. pp. 4015-4026 (2023)" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 132, + 501, + 480, + 544 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 501, + 480, + 544 + ], + "spans": [ + { + "bbox": [ + 132, + 501, + 480, + 544 + ], + "type": "text", + "content": "24. Krishna, R., Zhu, Y., Groth, O., Johnson, J., Hata, K., Kravitz, J., Chen, S., Kalantidis, Y., Li, L.J., Shamma, D.A., et al.: Visual genome: Connecting language and vision using crowdsourced dense image annotations. International journal of computer vision 123, 32-73 (2017)" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 132, + 545, + 480, + 588 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 545, + 480, + 588 + ], + "spans": [ + { + "bbox": [ + 132, + 545, + 480, + 588 + ], + "type": "text", + "content": "25. Kuznetsova, A., Rom, H., Alldrin, N., Uijlings, J., Krasin, I., Pont-Tuset, J., Kamali, S., Popov, S., Malloci, M., Kolesnikov, A., et al.: The open images dataset v4: Unified image classification, object detection, and visual relationship detection at scale. International journal of computer vision 128(7), 1956-1981 (2020)" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 132, + 589, + 480, + 621 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 589, + 480, + 621 + ], + "spans": [ + { + "bbox": [ + 132, + 589, + 480, + 621 + ], + "type": "text", + "content": "26. Li, F., Zhang, H., Sun, P., Zou, X., Liu, S., Yang, J., Li, C., Zhang, L., Gao, J.: Semantic-sam: Segment and recognize anything at any granularity. arXiv preprint arXiv:2307.04767 (2023)" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 132, + 622, + 480, + 665 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 622, + 480, + 665 + ], + "spans": [ + { + "bbox": [ + 132, + 622, + 480, + 665 + ], + "type": "text", + "content": "27. Li, F., Zhang, H., Xu, H., Liu, S., Zhang, L., Ni, L.M., Shum, H.Y.: Mask dino: Towards a unified transformer-based framework for object detection and segmentation. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 3041-3050 (2023)" + } + ] + } + ], + "index": 15 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "spans": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "type": "text", + "content": "16" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 166, + 91, + 241, + 101 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 166, + 91, + 241, + 101 + ], + "spans": [ + { + "bbox": [ + 166, + 91, + 241, + 101 + ], + "type": "text", + "content": "J. Li, J. Wu, et al." + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 15 + }, + { + "para_blocks": [ + { + "bbox": [ + 130, + 116, + 482, + 666 + ], + "type": "list", + "angle": 0, + "index": 18, + "blocks": [ + { + "bbox": [ + 130, + 116, + 482, + 161 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 116, + 482, + 161 + ], + "spans": [ + { + "bbox": [ + 130, + 116, + 482, + 161 + ], + "type": "text", + "content": "28. Li, H., Zhu, J., Jiang, X., Zhu, X., Li, H., Yuan, C., Wang, X., Qiao, Y., Wang, X., Wang, W., et al.: Uni-perceiver v2: A generalist model for large-scale vision and vision-language tasks. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 2691-2700 (2023)" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 130, + 162, + 481, + 183 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 162, + 481, + 183 + ], + "spans": [ + { + "bbox": [ + 130, + 162, + 481, + 183 + ], + "type": "text", + "content": "29. Li, J., Zhao, J., Wei, Y., Lang, C., Li, Y., Sim, T., Yan, S., Feng, J.: Multiple-human parsing in the wild. arXiv preprint arXiv:1705.07206 (2017)" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 132, + 184, + 481, + 226 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 184, + 481, + 226 + ], + "spans": [ + { + "bbox": [ + 132, + 184, + 481, + 226 + ], + "type": "text", + "content": "30. Li, L.H., Zhang, P., Zhang, H., Yang, J., Li, C., Zhong, Y., Wang, L., Yuan, L., Zhang, L., Hwang, J.N., et al.: Grounded language-image pre-training. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 10965-10975 (2022)" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 132, + 228, + 481, + 259 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 228, + 481, + 259 + ], + "spans": [ + { + "bbox": [ + 132, + 228, + 481, + 259 + ], + "type": "text", + "content": "31. Li, X., Xu, S., Yang, Y., Cheng, G., Tong, Y., Tao, D.: Panoptic-partformer: Learning a unified model for panoptic part segmentation. In: European Conference on Computer Vision. pp. 729-747. Springer (2022)" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 132, + 259, + 481, + 293 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 259, + 481, + 293 + ], + "spans": [ + { + "bbox": [ + 132, + 259, + 481, + 293 + ], + "type": "text", + "content": "32. Li, Y., Mao, H., Girshick, R., He, K.: Exploring plain vision transformer backbones for object detection. In: European Conference on Computer Vision. pp. 280-296. Springer (2022)" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 132, + 293, + 481, + 326 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 293, + 481, + 326 + ], + "spans": [ + { + "bbox": [ + 132, + 293, + 481, + 326 + ], + "type": "text", + "content": "33. Li, Y., Singh, K.K., Xue, Y., Lee, Y.J.: Partgan: Weakly-supervised part decomposition for image generation and segmentation. In: British Machine Vision Conference (BMVC) (2021)" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 132, + 326, + 481, + 358 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 326, + 481, + 358 + ], + "spans": [ + { + "bbox": [ + 132, + 326, + 481, + 358 + ], + "type": "text", + "content": "34. Lin, C., Jiang, Y., Qu, L., Yuan, Z., Cai, J.: Generative region-language pretraining for open-ended object detection. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 13958-13968 (2024)" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 132, + 358, + 481, + 392 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 358, + 481, + 392 + ], + "spans": [ + { + "bbox": [ + 132, + 358, + 481, + 392 + ], + "type": "text", + "content": "35. Lin, C., Sun, P., Jiang, Y., Luo, P., Qu, L., Haffari, G., Yuan, Z., Cai, J.: Learning object-language alignments for open-vocabulary object detection. In: The Eleventh International Conference on Learning Representations (2023)" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 132, + 392, + 481, + 424 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 392, + 481, + 424 + ], + "spans": [ + { + "bbox": [ + 132, + 392, + 481, + 424 + ], + "type": "text", + "content": "36. Lin, T.Y., Goyal, P., Girshick, R., He, K., Dollar, P.: Focal loss for dense object detection. In: Proceedings of the IEEE international conference on computer vision. pp. 2980-2988 (2017)" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 132, + 424, + 481, + 456 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 424, + 481, + 456 + ], + "spans": [ + { + "bbox": [ + 132, + 424, + 481, + 456 + ], + "type": "text", + "content": "37. Lin, T.Y., Maire, M., Belongie, S., Hays, J., Perona, P., Ramanan, D., Dollar, P., Zitnick, C.L.: Microsoft coco: Common objects in context. In: European Conference on Computer Vision. pp. 740-755. Springer (2014)" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 132, + 457, + 481, + 490 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 457, + 481, + 490 + ], + "spans": [ + { + "bbox": [ + 132, + 457, + 481, + 490 + ], + "type": "text", + "content": "38. Ling, H., Kreis, K., Li, D., Kim, S.W., Torralba, A., Fidler, S.: Editgan: High-precision semantic image editing. In: Advances in Neural Information Processing Systems. vol. 34, pp. 16331-16345 (2021)" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 132, + 491, + 481, + 522 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 491, + 481, + 522 + ], + "spans": [ + { + "bbox": [ + 132, + 491, + 481, + 522 + ], + "type": "text", + "content": "39. Liu, S., Zeng, Z., Ren, T., Li, F., Zhang, H., Yang, J., Li, C., Yang, J., Su, H., Zhu, J., et al.: Grounding dino: Marrying dino with grounded pre-training for open-set object detection. arXiv preprint arXiv:2303.05499 (2023)" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 132, + 522, + 481, + 566 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 522, + 481, + 566 + ], + "spans": [ + { + "bbox": [ + 132, + 522, + 481, + 566 + ], + "type": "text", + "content": "40. Liu, Z., Lin, Y., Cao, Y., Hu, H., Wei, Y., Zhang, Z., Lin, S., Guo, B.: Swin transformer: Hierarchical vision transformer using shifted windows. In: Proceedings of the IEEE/CVF international conference on computer vision. pp. 10012-10022 (2021)" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 132, + 567, + 481, + 600 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 567, + 481, + 600 + ], + "spans": [ + { + "bbox": [ + 132, + 567, + 481, + 600 + ], + "type": "text", + "content": "41. Lu, J., Clark, C., Zellers, R., Mottaghi, R., Kembhavi, A.: Unified-io: A unified model for vision, language, and multi-modal tasks. In: The Eleventh International Conference on Learning Representations (2022)" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 132, + 601, + 481, + 632 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 601, + 481, + 632 + ], + "spans": [ + { + "bbox": [ + 132, + 601, + 481, + 632 + ], + "type": "text", + "content": "42. Ma, C., Jiang, Y., Wen, X., Yuan, Z., Qi, X.: Codet: Co-occurrence guided region-word alignment for open-vocabulary object detection. In: Advances in neural information processing systems. vol. 36 (2023)" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 132, + 632, + 481, + 666 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 632, + 481, + 666 + ], + "spans": [ + { + "bbox": [ + 132, + 632, + 481, + 666 + ], + "type": "text", + "content": "43. Ma, C., Jiang, Y., Wu, J., Yuan, Z., Qi, X.: Groma: Localized visual tokenization for grounding multimodal large language models. arXiv preprint arXiv:2404.13013 (2024)" + } + ] + } + ], + "index": 17 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 402, + 91, + 447, + 100 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 402, + 91, + 447, + 100 + ], + "spans": [ + { + "bbox": [ + 402, + 91, + 447, + 100 + ], + "type": "text", + "content": "PartGLEE" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 470, + 91, + 481, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 470, + 91, + 481, + 100 + ], + "spans": [ + { + "bbox": [ + 470, + 91, + 481, + 100 + ], + "type": "text", + "content": "17" + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 16 + }, + { + "para_blocks": [ + { + "bbox": [ + 132, + 117, + 480, + 665 + ], + "type": "list", + "angle": 0, + "index": 18, + "blocks": [ + { + "bbox": [ + 133, + 117, + 480, + 149 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 117, + 480, + 149 + ], + "spans": [ + { + "bbox": [ + 133, + 117, + 480, + 149 + ], + "type": "text", + "content": "44. Mao, J., Huang, J., Toshev, A., Camburu, O., Yuille, A.L., Murphy, K.: Generation and comprehension of unambiguous object descriptions. In: Proceedings of the IEEE conference on computer vision and pattern recognition. pp. 11-20 (2016)" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 133, + 149, + 480, + 181 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 149, + 480, + 181 + ], + "spans": [ + { + "bbox": [ + 133, + 149, + 480, + 181 + ], + "type": "text", + "content": "45. Meletis, P., Wen, X., Lu, C., de Geus, D., Dubbelman, G.: Cityscapes-panoptic-parts and pascal-panoptic-parts datasets for scene understanding. arXiv preprint arXiv:2004.07944 (2020)" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 133, + 181, + 480, + 214 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 181, + 480, + 214 + ], + "spans": [ + { + "bbox": [ + 133, + 181, + 480, + 214 + ], + "type": "text", + "content": "46. Michieli, U., Borsato, E., Rossi, L., Zanuttigh, P.: Gmnet: Graph matching network for large scale part semantic segmentation in the wild. In: European Conference on Computer Vision. pp. 397-414. Springer (2020)" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 133, + 214, + 480, + 246 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 214, + 480, + 246 + ], + "spans": [ + { + "bbox": [ + 133, + 214, + 480, + 246 + ], + "type": "text", + "content": "47. Miletari, F., Navab, N., Ahmadi, S.A.: V-net: Fully convolutional neural networks for volumetric medical image segmentation. In: 2016 fourth international conference on 3D vision (3DV) (2016)" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 133, + 246, + 480, + 267 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 246, + 480, + 267 + ], + "spans": [ + { + "bbox": [ + 133, + 246, + 480, + 267 + ], + "type": "text", + "content": "48. Morabia, K., Arora, J., Vijaykumar, T.: Attention-based joint detection of object and semantic part. arXiv preprint arXiv:2007.02419 (2020)" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 133, + 267, + 480, + 299 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 267, + 480, + 299 + ], + "spans": [ + { + "bbox": [ + 133, + 267, + 480, + 299 + ], + "type": "text", + "content": "49. Nair, S., Rajeswaran, A., Kumar, V., Finn, C., Gupta, A.: R3m: A universal visual representation for robot manipulation. In: Conference on Robot Learning. pp. 892-909. PMLR (2023)" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 132, + 299, + 480, + 342 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 299, + 480, + 342 + ], + "spans": [ + { + "bbox": [ + 132, + 299, + 480, + 342 + ], + "type": "text", + "content": "50. Ng, X.L., Ong, K.E., Zheng, Q., Ni, Y., Yeo, S.Y., Liu, J.: Animal kingdom: A large and diverse dataset for animal behavior understanding. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 19023-19034 (2022)" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 132, + 342, + 480, + 374 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 342, + 480, + 374 + ], + "spans": [ + { + "bbox": [ + 132, + 342, + 480, + 374 + ], + "type": "text", + "content": "51. Pan, T.Y., Liu, Q., Chao, W.L., Price, B.: Towards open-world segmentation of parts. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 15392-15401 (2023)" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 132, + 374, + 480, + 407 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 374, + 480, + 407 + ], + "spans": [ + { + "bbox": [ + 132, + 374, + 480, + 407 + ], + "type": "text", + "content": "52. Qi, L., Kuen, J., Guo, W., Gu, J., Lin, Z., Du, B., Xu, Y., Yang, M.H.: Aims: All-inclusive multi-level segmentation for anything. In: Advances in Neural Information Processing Systems. vol. 36 (2023)" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 132, + 407, + 480, + 449 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 407, + 480, + 449 + ], + "spans": [ + { + "bbox": [ + 132, + 407, + 480, + 449 + ], + "type": "text", + "content": "53. Radford, A., Kim, J.W., Hallacy, C., Ramesh, A., Goh, G., Agarwal, S., Sastry, G., Askell, A., Mishkin, P., Clark, J., et al.: Learning transferable visual models from natural language supervision. In: International conference on machine learning. pp. 8748-8763. PMLR (2021)" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 132, + 449, + 480, + 482 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 449, + 480, + 482 + ], + "spans": [ + { + "bbox": [ + 132, + 449, + 480, + 482 + ], + "type": "text", + "content": "54. Raffel, C., Shazeer, N., Roberts, A., Lee, K., Narang, S., Matena, M., Zhou, Y., Li, W., Liu, P.J.: Exploring the limits of transfer learning with a unified text-to-text transformer. The Journal of Machine Learning Research 21(1), 5485-5551 (2020)" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 132, + 482, + 480, + 525 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 482, + 480, + 525 + ], + "spans": [ + { + "bbox": [ + 132, + 482, + 480, + 525 + ], + "type": "text", + "content": "55. Ramanathan, V., Kalia, A., Petrovic, V., Wen, Y., Zheng, B., Guo, B., Wang, R., Marquez, A., Kovvuri, R., Kadian, A., et al.: Paco: Parts and attributes of common objects. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 7141-7151 (2023)" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 132, + 525, + 480, + 557 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 525, + 480, + 557 + ], + "spans": [ + { + "bbox": [ + 132, + 525, + 480, + 557 + ], + "type": "text", + "content": "56. Ramesh, A., Dhariwal, P., Nichol, A., Chu, C., Chen, M.: Hierarchical text-conditional image generation with clip latents. arXiv preprint arXiv:2204.06125 1(2), 3 (2022)" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 132, + 557, + 480, + 589 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 557, + 480, + 589 + ], + "spans": [ + { + "bbox": [ + 132, + 557, + 480, + 589 + ], + "type": "text", + "content": "57. Ramesh, A., Pavlov, M., Goh, G., Gray, S., Voss, C., Radford, A., Chen, M., Sutskever, I.: Zero-shot text-to-image generation. In: International Conference on Machine Learning. pp. 8821-8831. PMLR (2021)" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 132, + 589, + 480, + 622 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 589, + 480, + 622 + ], + "spans": [ + { + "bbox": [ + 132, + 589, + 480, + 622 + ], + "type": "text", + "content": "58. Reddy, N.D., Vo, M., Narasimhan, S.G.: Carfusion: Combining point tracking and part detection for dynamic 3d reconstruction of vehicles. In: Proceedings of the IEEE conference on computer vision and pattern recognition. pp. 1906-1915 (2018)" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 132, + 622, + 480, + 665 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 622, + 480, + 665 + ], + "spans": [ + { + "bbox": [ + 132, + 622, + 480, + 665 + ], + "type": "text", + "content": "59. Rezatofighi, H., Tsoi, N., Gwak, J., Sadeghian, A., Reid, I., Savarese, S.: Generalized intersection over union: A metric and a loss for bounding box regression. In: Proceedings of the IEEE/CVF conference on computer vision and pattern recognition. pp. 658-666 (2019)" + } + ] + } + ], + "index": 17 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "spans": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "type": "text", + "content": "18" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 166, + 91, + 241, + 101 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 166, + 91, + 241, + 101 + ], + "spans": [ + { + "bbox": [ + 166, + 91, + 241, + 101 + ], + "type": "text", + "content": "J. Li, J. Wu, et al." + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 17 + }, + { + "para_blocks": [ + { + "bbox": [ + 130, + 116, + 482, + 666 + ], + "type": "list", + "angle": 0, + "index": 18, + "blocks": [ + { + "bbox": [ + 130, + 116, + 482, + 149 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 116, + 482, + 149 + ], + "spans": [ + { + "bbox": [ + 130, + 116, + 482, + 149 + ], + "type": "text", + "content": "60. Rombach, R., Blattmann, A., Lorenz, D., Esser, P., Ommer, B.: High-resolution image synthesis with latent diffusion models. In: Proceedings of the IEEE/CVF conference on computer vision and pattern recognition. pp. 10684-10695 (2022)" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 130, + 149, + 482, + 182 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 149, + 482, + 182 + ], + "spans": [ + { + "bbox": [ + 130, + 149, + 482, + 182 + ], + "type": "text", + "content": "61. Shao, S., Li, Z., Zhang, T., Peng, C., Yu, G., Zhang, X., Li, J., Sun, J.: Objects365: A large-scale, high-quality dataset for object detection. In: Proceedings of the IEEE/CVF international conference on computer vision. pp. 8430-8439 (2019)" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 130, + 182, + 482, + 215 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 182, + 482, + 215 + ], + "spans": [ + { + "bbox": [ + 130, + 182, + 482, + 215 + ], + "type": "text", + "content": "62. Sun, P., Chen, S., Zhu, C., Xiao, F., Luo, P., Xie, S., Yan, Z.: Going denser with open-vocabulary part segmentation. In: Proceedings of the IEEE/CVF International Conference on Computer Vision. pp. 15453-15465 (2023)" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 130, + 215, + 482, + 247 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 215, + 482, + 247 + ], + "spans": [ + { + "bbox": [ + 130, + 215, + 482, + 247 + ], + "type": "text", + "content": "63. Tang, C., Xie, L., Zhang, X., Hu, X., Tian, Q.: Visual recognition by request. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 15265-15274 (2023)" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 130, + 248, + 482, + 281 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 248, + 482, + 281 + ], + "spans": [ + { + "bbox": [ + 130, + 248, + 482, + 281 + ], + "type": "text", + "content": "64. Wah, C., Branson, S., Welinder, P., Perona, P., Belongie, S.: Caltech-ucsd birds200-2011 (cub-200-2011). Tech. Rep. CNS-TR-2011-001, California Institute of Technology (2011)" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 130, + 281, + 482, + 314 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 281, + 482, + 314 + ], + "spans": [ + { + "bbox": [ + 130, + 281, + 482, + 314 + ], + "type": "text", + "content": "65. Wang, P., Shen, X., Lin, Z., Cohen, S., Price, B., Yuille, A.L.: Joint object and part segmentation using deep learned potentials. In: Proceedings of the IEEE International Conference on Computer Vision. pp. 1573-1581 (2015)" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 130, + 314, + 482, + 358 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 314, + 482, + 358 + ], + "spans": [ + { + "bbox": [ + 130, + 314, + 482, + 358 + ], + "type": "text", + "content": "66. Wang, P., Yang, A., Men, R., Lin, J., Bai, S., Li, Z., Ma, J., Zhou, C., Zhou, J., Yang, H.: Ofa: Unifying architectures, tasks, and modalities through a simple sequence-to-sequence learning framework. In: International Conference on Machine Learning. pp. 23318-23340. PMLR (2022)" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 130, + 358, + 482, + 402 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 358, + 482, + 402 + ], + "spans": [ + { + "bbox": [ + 130, + 358, + 482, + 402 + ], + "type": "text", + "content": "67. Wang, W., Bao, H., Dong, L., Bjorck, J., Peng, Z., Liu, Q., Aggarwal, K., Mohammed, O.K., Singhal, S., Som, S., et al.: Image as a foreign language: Beit pretraining for all vision and vision-language tasks. arXiv preprint arXiv:2208.10442 (2022)" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 130, + 402, + 482, + 435 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 402, + 482, + 435 + ], + "spans": [ + { + "bbox": [ + 130, + 402, + 482, + 435 + ], + "type": "text", + "content": "68. Wang, X., Li, S., Kallidromitis, K., Kato, Y., Kozuka, K., Darrell, T.: Hierarchical open-vocabulary universal image segmentation. In: Advances in Neural Information Processing Systems. vol. 36 (2023)" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 130, + 435, + 482, + 468 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 435, + 482, + 468 + ], + "spans": [ + { + "bbox": [ + 130, + 435, + 482, + 468 + ], + "type": "text", + "content": "69. Wang, Y., Xu, Y., Tsogkas, S., Bai, X., Dickinson, S., Siddiqi, K.: Deepflux for skeletons in the wild. In: Proceedings of the IEEE/CVF conference on computer vision and pattern recognition. pp. 5287-5296 (2019)" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 130, + 468, + 482, + 501 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 468, + 482, + 501 + ], + "spans": [ + { + "bbox": [ + 130, + 468, + 482, + 501 + ], + "type": "text", + "content": "70. Wei, M., Yue, X., Zhang, W., Kong, S., Liu, X., Pang, J.: Ov-parts: Towards open-vocabulary part segmentation. In: Thirty-seventh Conference on Neural Information Processing Systems Datasets and Benchmarks Track (2023)" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 130, + 501, + 482, + 533 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 501, + 482, + 533 + ], + "spans": [ + { + "bbox": [ + 130, + 501, + 482, + 533 + ], + "type": "text", + "content": "71. Wu, J., Jiang, Y., Liu, Q., Yuan, Z., Bai, X., Bai, S.: General object foundation model for images and videos at scale. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 3783-3795 (2024)" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 130, + 533, + 482, + 578 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 533, + 482, + 578 + ], + "spans": [ + { + "bbox": [ + 130, + 533, + 482, + 578 + ], + "type": "text", + "content": "72. Xiao, B., Wu, H., Xu, W., Dai, X., Hu, H., Lu, Y., Zeng, M., Liu, C., Yuan, L.: Florence-2: Advancing a unified representation for a variety of vision tasks. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 4818-4829 (2024)" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 130, + 578, + 482, + 600 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 578, + 482, + 600 + ], + "spans": [ + { + "bbox": [ + 130, + 578, + 482, + 600 + ], + "type": "text", + "content": "73. Yan, B., Jiang, Y., Sun, P., Wang, D., Yuan, Z., Luo, P., Lu, H.: Towards grand unification of object tracking. In: European Conference on Computer Vision (2022)" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 130, + 600, + 482, + 632 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 600, + 482, + 632 + ], + "spans": [ + { + "bbox": [ + 130, + 600, + 482, + 632 + ], + "type": "text", + "content": "74. Yan, B., Jiang, Y., Wu, J., Wang, D., Luo, P., Yuan, Z., Lu, H.: Universal instance perception as object discovery and retrieval. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 15325-15336 (2023)" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 130, + 632, + 482, + 666 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 632, + 482, + 666 + ], + "spans": [ + { + "bbox": [ + 130, + 632, + 482, + 666 + ], + "type": "text", + "content": "75. Yang, L., Song, Q., Wang, Z., Jiang, M.: Parsing r-cnn for instance-level human analysis. In: Proceedings of the IEEE/CVF conference on computer vision and pattern recognition. pp. 364-373 (2019)" + } + ] + } + ], + "index": 17 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 402, + 91, + 447, + 100 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 402, + 91, + 447, + 100 + ], + "spans": [ + { + "bbox": [ + 402, + 91, + 447, + 100 + ], + "type": "text", + "content": "PartGLEE" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 470, + 91, + 481, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 470, + 91, + 481, + 100 + ], + "spans": [ + { + "bbox": [ + 470, + 91, + 481, + 100 + ], + "type": "text", + "content": "19" + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 18 + }, + { + "para_blocks": [ + { + "bbox": [ + 132, + 117, + 480, + 665 + ], + "type": "list", + "angle": 0, + "index": 17, + "blocks": [ + { + "bbox": [ + 133, + 117, + 480, + 149 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 117, + 480, + 149 + ], + "spans": [ + { + "bbox": [ + 133, + 117, + 480, + 149 + ], + "type": "text", + "content": "76. Yang, Y., Ramanan, D.: Articulated pose estimation with flexible mixtures-of-parts. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 1385-1392. IEEE (2011)" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 133, + 150, + 480, + 182 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 150, + 480, + 182 + ], + "spans": [ + { + "bbox": [ + 133, + 150, + 480, + 182 + ], + "type": "text", + "content": "77. Yang, Z., Gan, Z., Wang, J., Hu, X., Ahmed, F., Liu, Z., Lu, Y., Wang, L.: Unitab: Unifying text and box outputs for grounded vision-language modeling. In: European Conference on Computer Vision. pp. 521-539. Springer (2022)" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 133, + 182, + 480, + 225 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 182, + 480, + 225 + ], + "spans": [ + { + "bbox": [ + 133, + 182, + 480, + 225 + ], + "type": "text", + "content": "78. Yao, L., Han, J., Liang, X., Xu, D., Zhang, W., Li, Z., Xu, H.: Detclipv2: Scalable open-vocabulary object detection pre-training via word-region alignment. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 23497-23506 (2023)" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 132, + 225, + 480, + 267 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 225, + 480, + 267 + ], + "spans": [ + { + "bbox": [ + 132, + 225, + 480, + 267 + ], + "type": "text", + "content": "79. Yao, L., Han, J., Wen, Y., Liang, X., Xu, D., Zhang, W., Li, Z., Xu, C., Xu, H.: Detclip: Dictionary-enriched visual-concept paralleled pre-training for open-world detection. In: Advances in Neural Information Processing Systems. vol. 35, pp. 9125-9138 (2022)" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 132, + 268, + 480, + 300 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 268, + 480, + 300 + ], + "spans": [ + { + "bbox": [ + 132, + 268, + 480, + 300 + ], + "type": "text", + "content": "80. Yu, L., Poirson, P., Yang, S., Berg, A.C., Berg, T.L.: Modeling context in referring expressions. In: European Conference on Computer Vision. pp. 69-85. Springer (2016)" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 132, + 300, + 480, + 332 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 300, + 480, + 332 + ], + "spans": [ + { + "bbox": [ + 132, + 300, + 480, + 332 + ], + "type": "text", + "content": "81. Yuan, L., Chen, D., Chen, Y.L., Codella, N., Dai, X., Gao, J., Hu, H., Huang, X., Li, B., Li, C., et al.: Florence: A new foundation model for computer vision. arXiv preprint arXiv:2111.11432 (2021)" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 132, + 332, + 480, + 365 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 332, + 480, + 365 + ], + "spans": [ + { + "bbox": [ + 132, + 332, + 480, + 365 + ], + "type": "text", + "content": "82. Zareian, A., Rosa, K.D., Hu, D.H., Chang, S.F.: Open-vocabulary object detection using captions. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 14393-14402 (2021)" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 132, + 365, + 480, + 407 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 365, + 480, + 407 + ], + "spans": [ + { + "bbox": [ + 132, + 365, + 480, + 407 + ], + "type": "text", + "content": "83. Zhong, Y., Yang, J., Zhang, P., Li, C., Codella, N., Li, L.H., Zhou, L., Dai, X., Yuan, L., Li, Y., et al.: Regionclip: Region-based language-image pretraining. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 16793-16803 (2022)" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 132, + 407, + 480, + 439 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 407, + 480, + 439 + ], + "spans": [ + { + "bbox": [ + 132, + 407, + 480, + 439 + ], + "type": "text", + "content": "84. Zhou, B., Zhao, H., Puig, X., Xiao, T., Fidler, S., Barriuso, A., Torralba, A.: Semantic understanding of scenes through the ade20k dataset. International Journal on Computer Vision (2018)" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 132, + 439, + 480, + 483 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 439, + 480, + 483 + ], + "spans": [ + { + "bbox": [ + 132, + 439, + 480, + 483 + ], + "type": "text", + "content": "85. Zhou, T., Wang, W., Liu, S., Yang, Y., Van Gool, L.: Differentiable multigranularity human representation learning for instance-aware human semantic parsing. In: Proceedings of the IEEE/CVF conference on computer vision and pattern recognition. pp. 1622-1631 (2021)" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 132, + 483, + 480, + 515 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 483, + 480, + 515 + ], + "spans": [ + { + "bbox": [ + 132, + 483, + 480, + 515 + ], + "type": "text", + "content": "86. Zhu, X., Su, W., Lu, L., Li, B., Wang, X., Dai, J.: Deformable detr: Deformable transformers for end-to-end object detection. In: International Conference on Learning Representations (2021)" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 132, + 515, + 480, + 557 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 515, + 480, + 557 + ], + "spans": [ + { + "bbox": [ + 132, + 515, + 480, + 557 + ], + "type": "text", + "content": "87. Zhu, X., Zhu, J., Li, H., Wu, X., Li, H., Wang, X., Dai, J.: Uni-perceiver: Pretraining unified architecture for generic perception for zero-shot and few-shot tasks. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 16804-16815 (2022)" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 132, + 558, + 480, + 590 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 558, + 480, + 590 + ], + "spans": [ + { + "bbox": [ + 132, + 558, + 480, + 590 + ], + "type": "text", + "content": "88. Ziegler, A., Asano, Y.M.: Self-supervised learning of object parts for semantic segmentation. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 14502-14511 (2022)" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 132, + 590, + 480, + 633 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 590, + 480, + 633 + ], + "spans": [ + { + "bbox": [ + 132, + 590, + 480, + 633 + ], + "type": "text", + "content": "89. Zou, X., Dou, Z.Y., Yang, J., Gan, Z., Li, L., Li, C., Dai, X., Behl, H., Wang, J., Yuan, L., et al.: Generalized decoding for pixel, image, and language. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 15116-15127 (2023)" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 132, + 633, + 480, + 665 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 633, + 480, + 665 + ], + "spans": [ + { + "bbox": [ + 132, + 633, + 480, + 665 + ], + "type": "text", + "content": "90. Zou, X., Yang, J., Zhang, H., Li, F., Li, L., Wang, J., Wang, L., Gao, J., Lee, Y.J.: Segment everything everywhere all at once. In: Advances in Neural Information Processing Systems. vol. 36 (2023)" + } + ] + } + ], + "index": 16 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "spans": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "type": "text", + "content": "20" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 166, + 91, + 241, + 101 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 166, + 91, + 241, + 101 + ], + "spans": [ + { + "bbox": [ + 166, + 91, + 241, + 101 + ], + "type": "text", + "content": "J. Li, J. Wu, et al." + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 19 + } + ], + "_backend": "vlm", + "_version_name": "2.6.4" +} \ No newline at end of file diff --git a/2024/PartImageNet++ Dataset_ Scaling up Part-based Models for Robust Recognition/0d30f97b-2cbe-4f46-92c7-0bf8c016e1e3_content_list.json b/2024/PartImageNet++ Dataset_ Scaling up Part-based Models for Robust Recognition/0d30f97b-2cbe-4f46-92c7-0bf8c016e1e3_content_list.json new file mode 100644 index 0000000000000000000000000000000000000000..7b0dad812440add8af40dd3903c4f5f6f8abde52 --- /dev/null +++ b/2024/PartImageNet++ Dataset_ Scaling up Part-based Models for Robust Recognition/0d30f97b-2cbe-4f46-92c7-0bf8c016e1e3_content_list.json @@ -0,0 +1,1583 @@ +[ + { + "type": "text", + "text": "PartImageNet++ Dataset: Scaling up Part-based Models for Robust Recognition", + "text_level": 1, + "bbox": [ + 218, + 140, + 782, + 186 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Xiao Li $^{1}$ , Yining Liu $^{2}$ , Na Dong $^{3}$ , Sitian Qin $^{2}$ , and Xiaolin Hu $^{1,4*}$", + "bbox": [ + 251, + 210, + 750, + 228 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "$^{1}$ Department of Computer Science and Technology, BNRist, Institute for Artificial Intelligence, IDG/McGovern Institute for Brain Research, Tsinghua Laboratory of Brain and Intelligence, Tsinghua University, Beijing, China \n $^{2}$ Harbin Institute of Technology, Weihai, China \n $^{3}$ Beijing Institute of Technology, Beijing, China \n $^{4}$ Chinese Institute for Brain Research (CIBR), Beijing, China \nlixiao20@mails.tsinghua.edu.cn, 22S030184@stu.hit.edu.cn, 3220220911@bit.edu.cn, qinsitian@hitwh.edu.cn, xlu@mail.tsinghua.edu.cn", + "bbox": [ + 222, + 239, + 779, + 351 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Abstract. Deep learning-based object recognition systems can be easily fooled by various adversarial perturbations. One reason for the weak robustness may be that they do not have part-based inductive bias like the human recognition process. Motivated by this, several part-based recognition models have been proposed to improve the adversarial robustness of recognition. However, due to the lack of part annotations, the effectiveness of these methods is only validated on small-scale non-standard datasets. In this work, we propose PIN++, short for PartImageNet++, a dataset providing high-quality part segmentation annotations for all categories of ImageNet-1K (IN-1K). With these annotations, we build part-based methods directly on the standard IN-1K dataset for robust recognition. Different from previous two-stage part-based models, we propose a Multi-scale Part-supervised Model (MPM), to learn a robust representation with part annotations. Experiments show that MPM yielded better adversarial robustness on the large-scale IN-1K over strong baselines across various attack settings. Furthermore, MPM achieved improved robustness on common corruptions and several out-of-distribution datasets. The dataset, together with these results, enables and encourages researchers to explore the potential of part-based models in more real applications. The dataset and the code are available at https://github.com/LixiaoTHU/PartImageNetPP.", + "bbox": [ + 261, + 391, + 740, + 681 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Keywords: Adversarial robustness $\\cdot$ Part dataset $\\cdot$ Robust object recognition $\\cdot$ Part-based model", + "bbox": [ + 259, + 695, + 740, + 722 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "1 Introduction", + "text_level": 1, + "bbox": [ + 215, + 751, + 375, + 767 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Object recognition has achieved remarkable success due to the rise of Deep Neural Networks (DNNs) [23, 30, 44, 45]. However, DNNs often lack robustness and", + "bbox": [ + 212, + 784, + 787, + 815 + ], + "page_idx": 0 + }, + { + "type": "page_footnote", + "text": "* Corresponding author.", + "bbox": [ + 217, + 824, + 385, + 839 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/99d95f6b34d987bf2a3e8ecefab8c0dd433af43d20be3e12595fd1fc0b7cc4ca.jpg", + "image_caption": [ + "Fig. 1: Examples of annotated images in $\\mathrm{PIN}++$ . High-quality part segmentation annotations are provided on all categories in IN-1K. The object names are shown on the top-right of each image. The part names are hidden here." + ], + "image_footnote": [], + "bbox": [ + 316, + 143, + 493, + 297 + ], + "page_idx": 1 + }, + { + "type": "image", + "img_path": "images/5d7e5d39c16a6afbfcbe90e67c26bd757d3049b7ee7d081737dada980f114120.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 495, + 143, + 591, + 297 + ], + "page_idx": 1 + }, + { + "type": "image", + "img_path": "images/07f2dfb8820e93cc7699018c344d79c33212124cac03f85d618a2ffe5e9342b0.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 591, + 143, + 686, + 297 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "can be easily deceived by adversarial examples [19, 59], common image corruptions [24], and various out-of-distribution (OOD) shifts [34, 62], which greatly hinder applications of DNNs in security-critical scenarios [12, 33]. One possible reason for the weakness may be that DNN-based recognition does not have a part-based inductive bias (i.e., recognition based on object parts) like the human recognition process, which is generally regarded as a robust system [70]. According to the well-known cognitive psychological theory recognition-by-components [4], humans prefer to recognize objects by decomposing objects into parts and taking into account the hierarchical representations and the spatial relationships of parts. This cognitive theory has been supported by numerous psychological evidence [20, 26, 53, 61]. However, several studies revealed that DNNs often exhibit a shortcut recognition behavior, e.g., relying on texture-based features [16] and even unintelligible features for humans [27].", + "bbox": [ + 212, + 383, + 787, + 580 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "If DNNs can perform a human-aligned part-based recognition, both the adversarial and non-adversarial robustness of DNNs could be closer to humans. Motivated by this idea, two recent works [35, 56] investigated how to use additional part-level supervision to build up part-based (recognition) models, which make predictions by recognizing the object's parts in a bottom-up manner. Both works investigated two-stage part-based models, where the first stage involves part segmentation, followed by the classification stage based on the segmentation results. These works have shown the potential of part-based models in enhancing robustness against various perturbations.", + "bbox": [ + 212, + 580, + 787, + 717 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Nevertheless, due to the lack of part annotations, the effectiveness of part-based models has only been preliminary validated by the two works [35, 56] on small-scale nonstandard datasets. The largest dataset for evaluation was PartImageNet (PIN) [21], consisting of about $24\\mathrm{K}$ images of 158 categories selected from the large-scale ImageNet-1K (IN-1K) [11]. PIN has several potential issues. Firstly, it is quite small compared to the standard IN-1K (about $1.3\\mathrm{M}$ images) commonly used in the domain of adversarial robustness [3, 54]. This limited scale may restrict the generalizability of the findings. Secondly, categories included in", + "bbox": [ + 212, + 719, + 787, + 840 + ], + "page_idx": 1 + }, + { + "type": "page_number", + "text": "2", + "bbox": [ + 217, + 114, + 228, + 126 + ], + "page_idx": 1 + }, + { + "type": "header", + "text": "Li et al.", + "bbox": [ + 271, + 114, + 326, + 126 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "PIN are specifically chosen to facilitate part-related tasks (e.g., part segmentation [43] and object parsing [68]), with an emphasis on animal categories (118 out of 158). While this focus can be beneficial for studying part-related tasks themselves, it poses challenges in demonstrating the effectiveness of part-based models for general recognition tasks across a diverse range of objects. Thirdly, the annotation scheme of PIN is simple. It groups 158 categories into 11 super-categories and determines part annotations for each super-category, potentially failing to capture the characteristics of object categories within a super-category. Due to the three issues, the part-based models [35,56] trained on PIN cannot be directly compared with state-of-the-art (SOTA) methods trained on the standard IN-1K.", + "bbox": [ + 212, + 146, + 787, + 313 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "In this work, we introduce PartImageNet++ (PIN++, a new dataset providing high-quality part segmentation annotations for all categories of the widely used IN-1K [11], to facilitate the research on part-based models for robust recognition (classification). To create PIN++, we designed a detailed annotation scheme to ensure high-quality part annotations. With this scheme, we annotated 100 randomly selected images per category, resulting in a total of 100K images with part annotations (about 1/13 of IN-1K). To avoid bias, all the part annotations are created manually without using any auxiliary model. To the best of our knowledge, PIN++ provides large-scale high-quality part segmentation annotations for the most diverse range of object categories, including creatures, artifacts, rigid objects, nonrigid objects, etc, among existing part datasets. Fig. 1 showcases some examples of annotated images. We introduce this dataset in detail in Sec. 3 and show that existing techniques fail to obtain high-quality part segmentation results without such annotations.", + "bbox": [ + 212, + 321, + 787, + 534 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "By leveraging the part annotations of $\\mathrm{PIN}++$ , we develop part-based methods for robust recognition directly on the standard IN-1K dataset. To achieve this, we first train a part segmentation network with the annotations of $\\mathrm{PIN}++$ to obtain pseudo part labels for all unannotated images in IN-1K. We then propose a new Multi-scale Part-supervised recognition Model, termed MPM, to better exploit the part annotations. MPM is expected to learn a robust intermediate representation by adding auxiliary bypass layers to the vanilla recognition model (e.g., ResNet-50 [23]). The auxiliary bypass layers are supervised by the part annotations and the generated pseudo part labels. Without introducing any additional parameter or computation during inference, MPM can use high-resolution part annotations more adequately than two-stage part-based models [35, 56], which are constrained to use low-resolution part labels. Fig. 3 shows an overview of our whole method.", + "bbox": [ + 212, + 542, + 787, + 739 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "The experimental results demonstrate that MPM, combined with adversarial training (AT) [46], achieved better adversarial robustness on the large-scale IN-1K, outperforming strong AT baselines in both the seen and unseen adversarial threats. Furthermore, MPM exhibited improved robustness on common image corruptions [24] and several OOD datasets. In addition, MPM achieved higher alignments with human vision. Beyond object recognition, MPM is shown to", + "bbox": [ + 212, + 750, + 787, + 840 + ], + "page_idx": 2 + }, + { + "type": "header", + "text": "PartImageNet++ Dataset: Scaling up Part-based ...", + "bbox": [ + 383, + 114, + 730, + 128 + ], + "page_idx": 2 + }, + { + "type": "page_number", + "text": "3", + "bbox": [ + 774, + 116, + 784, + 126 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "boost the adversarial robustness of downstream tasks such as object detection [51]. The main contributions of this work can be summarized as follows:", + "bbox": [ + 215, + 146, + 784, + 176 + ], + "page_idx": 3 + }, + { + "type": "list", + "sub_type": "text", + "list_items": [ + "- We created PIN++, a new large-scale dataset with part annotations. This dataset can facilitate further research on part-based models for robust recognition and other part-related visual understanding tasks.", + "- We proposed MPM, a new part-supervised recognition model, to exploit the part annotations better, achieving superior robustness across various settings and benchmarks without introducing any extra inference cost." + ], + "bbox": [ + 223, + 184, + 782, + 272 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "2 Related Work", + "text_level": 1, + "bbox": [ + 215, + 296, + 385, + 311 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Adversarial robustness. Szegedy et al. [59] first revealed the vulnerability of DNNs to adversarial examples. Since then, numerous methods [55,66] have been proposed to enhance the adversarial robustness of DNNs. Among them, AT has emerged as the de facto paradigm for training adversially robust DNNs. Early AT works [64,65] primarily focused on small-scale, low-resolution datasets such as CIFAR-10 [29]. Recently, there has been a growing interest in investigating AT on IN-1K [11], as this dataset serves as a standard benchmark for evaluating SOTA computer vision techniques [23,44,45]. Several studies have demonstrated that AT models trained on IN-1K can transfer their adversarial robustness to downstream dense-prediction tasks and achieve zero-shot adversially robust recognition [9, 33, 36].", + "bbox": [ + 215, + 325, + 785, + 492 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Part datasets. The concept of parts holds significant importance in both human visual cognition [4] and computer vision domains. To conduct research on part-related visual tasks, the availability of part datasets is crucial. However, annotating object parts is challenging and expensive, and most part datasets are only limited to specific domains, such as cars [50, 57, 63] and humans [18, 37, 38, 67]. Furthermore, part datasets targeting common objects often have a limited number of object categories. Cityscapes PanopticParts (Cityscapes PP.) [48], PascalPart [6], and ADE20K provide part annotations for 5, 20, and 80 object categories, respectively. Although the recent PACO dataset [49] offers a large number of part annotations, it is still restricted to only 75 object categories. PIN [21], containing 158 categories from IN-1K [11], represents the dataset with the most object categories in terms of part annotations.", + "bbox": [ + 215, + 492, + 785, + 672 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Part-based object recognition. Modeling objects in terms of parts has a long history in computer vision [5,14,17]. But with the rise of DNNs, parts have been rarely used as auxiliaries for end-to-end general object recognition [23,30,44,45]. Recently, two works [35, 56] revisited part-based recognition models from the robustness perspective and showed the potential to enhance adversarial robustness. Li et al. [35] proposed ROCK, a part-based model that first predicts all parts represented by segmentation results, and then utilizes a judgment block to give category predictions based on the predicted parts. Similarly, Sitawarin et al. [56] also investigated several types of two-stage part-based models, where the first stage involves part segmentation, followed by the recognition stage with a tiny classifier. However, as mentioned before, both works were only validated on", + "bbox": [ + 215, + 674, + 785, + 839 + ], + "page_idx": 3 + }, + { + "type": "page_number", + "text": "4", + "bbox": [ + 217, + 114, + 228, + 126 + ], + "page_idx": 3 + }, + { + "type": "header", + "text": "Li et al.", + "bbox": [ + 271, + 114, + 325, + 126 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "small-scale part-friendly datasets for the lack of part annotations, e.g., Sitawarin et al. [56] performed classification tasks only on 11 super-categories of PIN [21]. Furthermore, the proposed two-stage part-based models inevitably introduced a lot of extra parameters during inference, while the utilization of high-resolution part annotations was limited, as discussed in Sec. 4.2. In addition, ROCK [35] contains a non-differentiable process, which could lead to a potential overestimation of adversarial robustness [2,60]. Our work will address all of these issues.", + "bbox": [ + 212, + 146, + 787, + 253 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "3 PIN++ Dataset", + "text_level": 1, + "bbox": [ + 214, + 273, + 410, + 290 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "We first present the details of how we built $\\mathrm{PIN}++$ to ensure high-quality part annotations, followed by the statistics of this dataset. We then give a discussion on $\\mathrm{PIN}++$ .", + "bbox": [ + 212, + 303, + 785, + 349 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "3.1 Annotation Scheme", + "text_level": 1, + "bbox": [ + 214, + 369, + 426, + 386 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Data source. IN-1K [11], containing about 1.3M images with category labels in the training set, is one of the most widely used datasets for general object recognition. Our aim is to provide part annotations for part-based robust recognition on IN-1K, and thus the images of PIN++ directly reuse the training set of IN-1K. All the images conform to licensing for research purposes. However, due to cost constraints, it is impractical to provide part-level annotations for all training images in IN-1K. Instead, we decided to annotate 100 randomly selected images per category. Once part annotations are provided for all categories, modern supervised segmentation techniques [22] can be used to obtain pseudo part labels for the remaining unannotated images, as we will show in Sec. 4. Note that PIN++ does not provide any part annotations for the validation set of IN-1K.", + "bbox": [ + 212, + 393, + 787, + 559 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "While PIN [21] has provided some part annotations for 158 categories of IN-1K, their annotations do not fully meet our purpose and principles. To minimize annotation costs without compromising quality, we retained annotations of 90 categories from PIN as part of $\\mathrm{PIN}++$ . The details are shown in Appendix A. We then show the annotation design for the remaining 910 categories of IN-1K. Annotation quality control. To ensure high-quality part annotations, three points should be taken into account:", + "bbox": [ + 212, + 560, + 787, + 665 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "1) Deciding which parts to annotate per category. One of the key challenges in the part annotation task is the ambiguity of object part selection (e.g., how to annotate the parts of hammer). PIN [21] evades this problem and focuses primarily on annotating parts for animal categories (most quadrupeds can be regarded as consisting of four part categories: head, body, foot, tail). However, the remaining categories in IN-1K are highly diverse, making it challenging to decide which parts to annotate. To ensure a scientifically grounded division of parts for each category, we initially consult the Wikidata knowledge base5 to obtain part vocabularies for every object category (e.g., hammer consists of three parts according to Wikidata: handle, striker, and hammerhead). Additionally, for object", + "bbox": [ + 212, + 665, + 787, + 816 + ], + "page_idx": 4 + }, + { + "type": "header", + "text": "PartImageNet++ Dataset: Scaling up Part-based ...", + "bbox": [ + 383, + 114, + 730, + 130 + ], + "page_idx": 4 + }, + { + "type": "page_number", + "text": "5", + "bbox": [ + 774, + 116, + 784, + 126 + ], + "page_idx": 4 + }, + { + "type": "page_footnote", + "text": "5 https://www.wikidata.org/wiki/Wikidata:Introduction", + "bbox": [ + 217, + 824, + 632, + 840 + ], + "page_idx": 4 + }, + { + "type": "table", + "img_path": "images/081e7fd9d179973ffa17c4d5df5b73e81fb0fca4b45df7ee669ddd7ad5a77bd8.jpg", + "table_caption": [ + "Table 1: The numbers of annotated object categories, part categories, images and part masks of publicly available part datasets." + ], + "table_footnote": [ + "*: ADE20K is severely category-unbalanced and here we show statistics for categories with more than 10 annotated part masks." + ], + "table_body": "
DatasetObject CategoryPart CategoryImagePart Mask
Cityscapes PP. [48]5233.5K100K
Pascal-Part [6]2019319K363.5K
ADE20K* [69]8056612.6K193.2K
PACO [49]7545676.7K641.4K
PIN [21]15860924K112K
PIN++10003308100K406.4K
", + "bbox": [ + 246, + 184, + 754, + 291 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "categories without a clear definition of parts on Wikidata, we decide which parts to annotate by asking recruited volunteers which parts they think to be important to their cognition of the object. For example, according to the cognition of volunteers, for most quadruped categories, annotating head, body, foot, and tail is enough, while for ram (see Fig. 1 for one example), horn should be annotated as an extra part. For categories that are indeed difficult to decompose into parts, e.g., flatworm, we generalize the concept of \"parts\" and treat the foreground of an object as one part category. In this case, the part is the object itself. Other details on deciding the parts to annotate are further discussed in Appendix B. These principles ensure that all categories in IN-1K can have part annotations.", + "bbox": [ + 212, + 349, + 785, + 501 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "2) Designing the part segmentation principles. We design several part segmentation principles to guide the annotators and ensure high-quality part segmentation annotation. Firstly, the annotated part masks are required to be combined to cover the entire object and have no overlap, unless in the case of the second principle. Secondly, for some parts that indeed should overlap in semantic concepts (e.g., horn and head), the annotators should annotate the inclusion relation of parts (e.g., horn is included in head), and then these part masks can be annotated with overlap. Thirdly, to keep the consistency of part annotations among images of the same object category, following [21], the same annotator should annotate all 100 images of this object category. The other principles are shown in Appendix C.", + "bbox": [ + 212, + 502, + 785, + 667 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "3) Annotation quality inspection. Ten randomly selected images together with the annotation visualizations (see Fig. 1) are sent to inspectors to assess if the annotations for a specific object category meet the defined principles. If the annotations for any two images fail to meet these principles, the entire category's annotations are re-annotated until the requirements are satisfied.", + "bbox": [ + 212, + 667, + 785, + 743 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "3.2 Statistics", + "text_level": 1, + "bbox": [ + 215, + 767, + 338, + 781 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "We finally annotated 100 images per category for 910 categories in IN-1K, resulting in a total of 91K images with part annotations. Taking these data and the data of 90 reused categories of PIN together, $\\mathrm{PIN}++$ for 1K categories is", + "bbox": [ + 212, + 794, + 785, + 839 + ], + "page_idx": 5 + }, + { + "type": "page_number", + "text": "6", + "bbox": [ + 217, + 114, + 228, + 126 + ], + "page_idx": 5 + }, + { + "type": "header", + "text": "Li et al.", + "bbox": [ + 271, + 114, + 326, + 126 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "obtained. Following IN-1K, here we ensure PIN++ is class-balanced. PIN++ is split into train/val sets with a 9:1 ratio for each category. But note that all annotated images are part of the training set of IN-1K. Unless stated otherwise, they are all utilized for training in our following experiments and the evaluation is performed on IN-1K val set.", + "bbox": [ + 212, + 146, + 782, + 220 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "In summary, $\\mathrm{PIN}++$ includes a total of 100K images with part annotations. These annotations cover 3,310 part categories for 1,000 object categories. A total of 406,364 part masks are annotated in $\\mathrm{PIN}++$ . Further statistics of $\\mathrm{PIN}++$ are shown in Appendix D.", + "bbox": [ + 212, + 223, + 782, + 282 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Comparison with other part datasets. We compare PIN++ with publicly available part datasets, as shown in Tab. 1. PIN++ outperforms previous datasets in terms of the number of object and part categories, as well as the total number of images with annotated part masks. It also demonstrates the competitive number of part masks compared to the recent PACO dataset [49]. Notably, PIN++ offers part annotations for a diverse range of objects, including creatures, artifacts, rigid objects, and nonrigid objects, unlike PACO or PIN, which are primarily focused on common tools or animals, respectively. As shown in Appendix D, the annotation quality of PIN++ exhibits an advantage over that of PIN. Moreover, PIN++ is category-balanced and guarantees that each image contains only one foreground category. All these features build the unique advantage of PIN++ for studying object recognition tasks.", + "bbox": [ + 212, + 284, + 784, + 465 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "3.3 Discussion on PIN++", + "text_level": 1, + "bbox": [ + 215, + 492, + 442, + 507 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Recent advancements in segmentation include CLIP-based open-vocabulary segmentation methods [32,39,58] and the SAM model [28]. We compared these techniques with the part annotations of PIN++. We used the largest version of SAM, and for open-vocabulary segmentation, we employed the VLPart model [58], which is trained with part annotations from several part datasets including Pascal-Part [6], PIN [21], and PACO [49]. Visual comparisons are presented in Fig. 2. We can see that VLPart only provides roughly reasonable semantics for part categories it has been trained on. It fails to accurately identify the semantics of parts from unseen objects. On the other hand, SAM seems to perform well in segmenting certain parts based on the object's edges, but it struggles when the part edges are less distinguishable, as shown in the segmentation results for camel (the third row in Fig. 2). In addition, note that all masks segmented by SAM are category-agnostic.", + "bbox": [ + 212, + 521, + 784, + 717 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "The above results highlight the challenges faced by existing techniques in achieving accurate part segmentation for diverse objects without training on PIN++. Additionally, current text-to-image generative models [52] also struggle with understanding part semantics [42]. However, supervised training on PIN++ yields good part segmentation results, as shown in Appendix E. Therefore, besides its usage in robust object recognition (as demonstrated next), we believe that PIN++ and its extensive collections of 3308 part categories (vocabularies) have the potential to enhance general part-related visual understanding tasks.", + "bbox": [ + 212, + 719, + 782, + 839 + ], + "page_idx": 6 + }, + { + "type": "header", + "text": "PartImageNet++ Dataset: Scaling up Part-based ...", + "bbox": [ + 383, + 114, + 730, + 128 + ], + "page_idx": 6 + }, + { + "type": "page_number", + "text": "7", + "bbox": [ + 774, + 116, + 784, + 126 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/e529d980f0f4400e03052c8750ac3ec7e18b9d69b90e6fac69a5eb40752d1a76.jpg", + "image_caption": [ + "Fig. 2: Comparison between part segmentation results of different methods and PIN++ annotations. Without training on PIN++, VLPart and SAM fail to segment objects into specific parts with accurate semantics." + ], + "image_footnote": [], + "bbox": [ + 259, + 146, + 743, + 362 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "4 Part-based Recognition Method", + "text_level": 1, + "bbox": [ + 215, + 444, + 560, + 462 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "With the annotations of $\\mathrm{PIN}++$ , we develop part-based methods directly on the standard IN-1K. Our part-based methods can be divided into two steps. We first obtain pseudo part labels for the remaining unlabeled images of IN-1K and then train the MPM with these part annotations and pseudo labels. The pipeline of our part-based methods is illustrated in Fig. 3.", + "bbox": [ + 212, + 474, + 787, + 551 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "4.1 Pseudo Part Label Generation", + "text_level": 1, + "bbox": [ + 215, + 573, + 514, + 588 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Part segmentation model. Arbitrary instance segmentation models [13, 22] can be leveraged to perform part segmentation and generate pseudo part labels when trained with the part annotations from $\\mathrm{PIN}++$ . Following Sun et al. [58], we use the classical Mask R-CNN [22] model with a vision transformer Swin-B [44] as the backbone network. In training the Mask R-CNN, we consider the masks for each part category as masks for individual object categories directly. During inference for obtaining pseudo part labels, we incorporate a simple post-processing operation alongside the regular inference procedure.", + "bbox": [ + 212, + 597, + 785, + 718 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Post-processing. Given an image $\\mathbf{x} \\in \\mathbb{R}^{3 \\times H \\times W}$ with object category $y_{c}$ , where $H \\times W$ represents the resolution, the regular inference procedure of the trained Mask R-CNN produces pseudo part labels in the form of $\\{(\\mathbf{M}_{\\mathbf{p}}, \\mathbf{v}_{\\mathbf{p}})\\}$ , where $\\mathbf{M}_{\\mathbf{p}} \\in \\{0,1\\}^{H \\times W}$ denotes the binary part mask, $\\mathbf{v}_{\\mathbf{p}} \\in [0,1]^{K}$ denotes the output probability of part categories, and $K$ denotes the number of part categories (3308 in our case). But note that the images that need inference are all in the training set of IN-1K. Their object category $y_{c}$ is not utilized in the regular inference procedure. To leverage it, we apply a Category Filter (CF) operation", + "bbox": [ + 212, + 719, + 787, + 840 + ], + "page_idx": 7 + }, + { + "type": "page_number", + "text": "8", + "bbox": [ + 217, + 114, + 228, + 126 + ], + "page_idx": 7 + }, + { + "type": "header", + "text": "Li et al.", + "bbox": [ + 271, + 114, + 326, + 126 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/0c28df4a0c5bf1f10a9da094faf07570f46ba4a83f395256ee4b8d3467d539fa.jpg", + "image_caption": [ + "Fig. 3: An overview of the generation of pseudo-labels and the structure of MPM. (a) A part segmentation model trained on $\\mathrm{PIN}++$ and used to obtain pseudo-part labels for unannotated images. (b) MPM adds several auxiliary bypass layers to the vanilla recognition model for part segmentation supervision. MPM is trained by part annotations together with the pseudo part labels. During inference, the auxiliary layers are dropped, and the vanilla recognition model gives the final object category prediction." + ], + "image_footnote": [], + "bbox": [ + 217, + 147, + 485, + 320 + ], + "page_idx": 8 + }, + { + "type": "image", + "img_path": "images/830f9567b4d8b79f78cf57d39ae85fce31b21491f9c33169dec4938f65f22068.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 506, + 148, + 784, + 319 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "to obtain $\\hat{\\mathbf{v}}_{\\mathbf{p}}$ , which sets the probabilities of unrelated part categories in $\\mathbf{v}_{\\mathbf{p}}$ to zero (e.g., for cat, except for cat:head, cat:body, etc, other part categories such as radio:button and bird:head are all excluded). The pseudo part category for $\\mathbf{M}_{\\mathbf{p}}$ is then computed as $y_{p} = \\arg \\max \\hat{\\mathbf{v}}_{\\mathbf{p}}$ .", + "bbox": [ + 212, + 444, + 787, + 506 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "4.2 Multi-scale Part-supervised Recognition Model", + "text_level": 1, + "bbox": [ + 214, + 526, + 651, + 542 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "Data preparation. After obtaining the pseudo labels for all remaining training images of IN-1K, we simply treat the pseudo labels and real part annotations equally because we find that the quality of the pseudo labels is roughly satisfactory, as shown in Appendix E. The part labels $\\{(\\mathbf{M_p},\\mathbf{y_p})\\}$ for an image $\\mathbf{x}$ are then converted to a single composite segmentation mask $\\mathbf{M} \\in \\{0,1\\}^{H \\times W \\times (K + 1)}$ where the extra one channel indicates the background. Finally, the data samples for training part-based models are $\\{(\\mathbf{x},y_c,\\mathbf{M})\\}$ .", + "bbox": [ + 212, + 551, + 787, + 656 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "Model design. Both previous works [35, 56] build part-based models in the two-stage way: a segmenter $\\mathcal{F}_{\\mathrm{seg}}: \\mathbb{R}^{3 \\times H \\times W} \\to \\mathbb{R}^{(K + 1) \\times h \\times w}$ and an extra module $\\mathcal{F}_{\\mathrm{cls}}: \\mathbb{R}^{(K + 1) \\times h \\times w} \\to \\mathbb{R}^C$ for classification, where $h \\times w$ denotes the output resolution and $C$ denotes the number of object category. $\\mathcal{F}_{\\mathrm{seg}}$ is a backbone network (e.g., a ResNet-50 without classification head). Most popular backbones [23, 44, 45] are hierarchically composed of several blocks with down-sampling layers and thus they can be written as $\\mathcal{F}_{\\mathrm{seg}} := f_l \\circ \\dots \\circ f_1$ , where $l$ denotes the number of down-sampling layers and $f_{i(1 \\leq i \\leq l)}$ denotes one block with the $i$ th down-sampling layer. $\\mathcal{F}_{\\mathrm{cls}}$ can be a tiny classifier [56] or a nondifferentiable judgment block [35] and its input is the part segmentation results $\\hat{\\mathbf{M}} = \\mathcal{F}_{\\mathrm{seg}}(\\mathbf{x})$ . The overall model for recognition is $\\mathcal{F} := \\mathcal{F}_{\\mathrm{cls}} \\circ \\mathcal{F}_{\\mathrm{seg}}$ . However, as $K$ is significantly larger than $C$ , using $\\hat{M}$ as the intermediate results for", + "bbox": [ + 212, + 657, + 787, + 840 + ], + "page_idx": 8 + }, + { + "type": "header", + "text": "PartImageNet++ Dataset: Scaling up Part-based ...", + "bbox": [ + 383, + 114, + 730, + 128 + ], + "page_idx": 8 + }, + { + "type": "page_number", + "text": "9", + "bbox": [ + 774, + 116, + 784, + 126 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "recognition inevitably introduce extra parameters and computation during inference. This situation gets worse with an increased number of part categories, as shown in Sec. 5.5. Moreover, these methods solely employ part supervision on the output of $f_{l}$ . Without expanding $\\mathcal{F}_{\\mathrm{seg}}$ , part annotations have to be downsampled to match the output resolution of $f_{l}$ , potentially limiting the utilization of high-resolution part annotations.", + "bbox": [ + 212, + 146, + 782, + 234 + ], + "page_idx": 9 + }, + { + "type": "text", + "text": "Instead of these two-stage part-based models, MPM is expected to learn a robust intermediate representation by adding multi-scale bypass layers to the vanilla recognition network. In MPM, the model used for object recognition is the vanilla backbone network directly: $\\mathcal{F} \\coloneqq h_c \\circ f_l \\circ \\dots \\circ f_1$ , where $h_c$ denotes a vanilla classification head (usually a linear layer). MPM utilizes the part annotations to supervise the intermediate features of $\\mathcal{F}$ by several bypass layers $g_i$ : $\\tilde{\\mathbf{M}}_i = g_i \\circ f_i \\circ \\dots \\circ f_1(\\mathbf{x})$ , where $i \\leq l$ . In this way, the intermediate features of $\\mathcal{F}$ can be seen as implicit representations of part segmentation results when $g_i$ is simple enough. Note that here we use part supervision not only on the output of $f_l$ , but also $f_{i(i < l)}$ . As the output of lower layers has a larger resolution, it can be supervised by part annotations with higher resolution. But the outputs of $f_{i(i < l)}$ are relatively low-level, which may not be enough to obtain high-level part segmentation results themselves. To balance this, only the outputs of the last three blocks $f_{i(l - 2 \\leq i \\leq l)}$ are supervised by part mask $\\mathbf{M}$ (e.g., for a ResNet-50 with input size $224 \\times 224$ , the intermediate features of $7 \\times 7$ , $14 \\times 14$ , and $28 \\times 28$ are supervised by corresponding down-sampled $\\mathbf{M}$ , respectively). In addition, following the principle of FPN [40], some top-down layers are used to augment the low-level features with high-level features. Note that distinct from the conventional FPN that generally has comparable parameters with the backbone network [40], these bypass layers are extremely lightweight as our goal is to improve the recognition accuracy rather than the quality of part segmentation results. Fig. 3(b) shows the overall structure of MPM. During inference, these auxiliary layers are dropped, and the vanilla recognition model gives the final object category prediction.", + "bbox": [ + 212, + 237, + 787, + 599 + ], + "page_idx": 9 + }, + { + "type": "text", + "text": "Training objective. The overall loss for training MPM is: $L = L_{\\mathrm{cls}} + \\lambda \\cdot L_{\\mathrm{seg}}$ , where $L_{\\mathrm{cls}}$ denotes the vanilla loss for classification, $L_{\\mathrm{seg}}$ denotes the loss for part segmentation, and $\\lambda$ is a hyper-parameter. We compute $L_{\\mathrm{seg}}$ as the average of losses on three part segmentation results of different resolutions, as mentioned before. Note that when performing AT, adversarial examples are generated by $\\mathbf{x}^{\\star} = \\arg \\max_{\\mathbf{x}^{\\star}:||\\mathbf{x}^{\\star} - \\mathbf{x}||_{p}\\leq \\epsilon}L_{\\mathrm{cls}}(\\mathcal{F}(\\mathbf{x}^{\\star}),y)$ , while MPM is always trained with $L$ .", + "bbox": [ + 212, + 599, + 787, + 694 + ], + "page_idx": 9 + }, + { + "type": "text", + "text": "5 Experiments", + "text_level": 1, + "bbox": [ + 215, + 710, + 375, + 729 + ], + "page_idx": 9 + }, + { + "type": "text", + "text": "5.1 Experimental Setup", + "text_level": 1, + "bbox": [ + 215, + 741, + 426, + 758 + ], + "page_idx": 9 + }, + { + "type": "text", + "text": "Training setup. Unless specified otherwise, following previous works [3,10], we performed AT with $l_{\\infty}$ bound $\\epsilon = 4 / 255$ on IN-1K. The inner optimization for obtaining adversarial examples used PGD [46] with iterative steps $t = 2$ and the step size was set to be $s = 2*\\epsilon /t$ . The input resolution used $224\\times 224$ . We used the vanilla version of the AT [46] and did not incorporate other variants", + "bbox": [ + 212, + 763, + 787, + 840 + ], + "page_idx": 9 + }, + { + "type": "page_number", + "text": "10", + "bbox": [ + 217, + 114, + 235, + 126 + ], + "page_idx": 9 + }, + { + "type": "header", + "text": "Li et al.", + "bbox": [ + 271, + 114, + 325, + 126 + ], + "page_idx": 9 + }, + { + "type": "table", + "img_path": "images/646998addc8cd6ce92e68f821450cd69bc4add88141085ae92cbfb1806eaf704.jpg", + "table_caption": [ + "Table 2: Recognition accuracies (\\%) of methods under different attack threats on IN-1K, following the evaluation setup in Sec. 5.1. Average denotes the average accuracies in four attack settings. We highlight the best results in each column." + ], + "table_footnote": [], + "table_body": "
ArchitectureMethodAdversarial Train wrt l∞ (ε = 4/255)
Cleanl∞l∞(×2)l1l2Average
ResNet-50[54], 202063.935.913.22.013.716.2
ResNet-50[47], 202264.734.311.42.316.416.1
ResNet-50[41], 202365.532.69.53.317.915.8
ResNet-50-Gelu[3], 202167.836.610.83.016.816.8
ResNet-50-Gelu[10], 202366.835.612.83.616.517.1
ResNet-50-Geluours (vanilla)67.138.112.65.021.619.3
ResNet-50-Geluours (MPM)67.839.113.66.224.320.8
", + "bbox": [ + 235, + 198, + 756, + 335 + ], + "page_idx": 10 + }, + { + "type": "text", + "text": "such as TRADES [65] or AWP [64]. These variants have shown effectiveness mainly on small-scale datasets, making their generalization to IN-1K non-trivial. Following previous part-based works [35,56], we used the classical ResNet-50 [23] network with 25.6M parameters as the baseline model. We note that recent works [3,10] used a GELU [25] activation to replace the original RELU [1] activation of ResNet-50 for boosting robustness when performing AT, here we also used GELU, denoted as ResNet-50-Gelu. In addition, we improved the training recipe of ResNet-50 to build a strong baseline. Other details on the training recipe are provided in Appendix F. Note that bypass layers (e.g., point-wise convolutions) with about 4.5M parameters were used by MPM during training but completely dropped during inference. The code is submitted along with the paper.", + "bbox": [ + 212, + 358, + 787, + 525 + ], + "page_idx": 10 + }, + { + "type": "text", + "text": "Evaluation setup. For all experiments in the work, unless specified otherwise, we evaluated the adversarial robustness with AutoAttack [8], which is a combination of various attack methods and is generally recognized as a reliable evaluation [7]. The evaluation was performed at a resolution $224 \\times 224$ on randomly selected 10K images of the 50K IN-1K val set, 10 images for each category. This ensures that the $95\\%$ confidence intervals of the reported average adversarial robustness (see Tab. 2) were less than $\\pm 0.5\\%$ . Note that, unlike previous works [3,10], we did not use the 5K images selected by RobustBench [7] as we found that they were category-unbalanced, containing three categories even without any image for validation.", + "bbox": [ + 212, + 525, + 787, + 674 + ], + "page_idx": 10 + }, + { + "type": "text", + "text": "We expect the AT models with $l_{\\infty}$ bound $\\epsilon = 4 / 255$ to be adversarially robust not only on seen threats (i.e., $l_{\\infty}$ attack with $\\epsilon = 4 / 255$ ) but also on unseen attack threats (e.g., $l_{\\infty}$ attack with larger $\\epsilon, l_1$ and $l_2$ attacks). Thus, we evaluated the models on three attack threats: $l_{\\infty}$ , $l_1$ , and $l_2$ , with the bounds $\\epsilon_{\\infty} = 4 / 255$ , $\\epsilon_1 = 75$ , $\\epsilon_2 = 2$ , respectively. For $l_{\\infty}$ attack, we extra evaluated the models with $\\epsilon_{\\infty} = 8 / 255$ to mimic a stronger threat, denoted as $l_{\\infty}(\\times 2)$ .", + "bbox": [ + 212, + 676, + 787, + 768 + ], + "page_idx": 10 + }, + { + "type": "text", + "text": "5.2 Robustness against Adversarial Attacks", + "text_level": 1, + "bbox": [ + 215, + 786, + 588, + 801 + ], + "page_idx": 10 + }, + { + "type": "text", + "text": "With the above setup, we trained the vanilla ResNet-50-Gelu as a baseline. MPM used the same recipe as the baseline model for a fair comparison, except that it", + "bbox": [ + 212, + 809, + 785, + 840 + ], + "page_idx": 10 + }, + { + "type": "header", + "text": "PartImageNet++ Dataset: Scaling up Part-based ...", + "bbox": [ + 383, + 114, + 730, + 130 + ], + "page_idx": 10 + }, + { + "type": "page_number", + "text": "11", + "bbox": [ + 767, + 114, + 782, + 126 + ], + "page_idx": 10 + }, + { + "type": "table", + "img_path": "images/732f12916ba7eefeed1a26d21a9cb4bbfde33bcc787c9f9a7e7442c70a5c3243.jpg", + "table_caption": [ + "Table 3: Recognition accuracies $(\\%)$ of ResNet-50-Gelu on 15 different common image corruptions [24]. We adopt accuracy as the metric to be consistent with other results, while this metric can be easily converted into the corruption error metric [24]. \"Average\" denotes the accuracy averaged over different common corruptions." + ], + "table_footnote": [], + "table_body": "
Part ATNoiseBlurWeatherDigitalAvg.
Gauss. Shot Impul.Defoc.Glass Moti.ZoomSnow FrostFogBrightCont.Elast.Pixel JPEG
31.230.727.339.028.639.537.835.840.753.869.038.037.047.157.040.8
✓31.531.028.441.929.840.336.737.041.355.369.639.636.952.959.942.1
✓29.628.521.023.332.932.834.434.234.010.858.49.552.258.863.134.9
✓30.129.222.425.235.135.036.436.436.211.959.710.453.159.864.336.4
", + "bbox": [ + 228, + 212, + 781, + 289 + ], + "page_idx": 11 + }, + { + "type": "table", + "img_path": "images/923cc41b2d6caaffae796b79383c3eb722238c7b1a39b9b2a65f8df273629fad.jpg", + "table_caption": [ + "Table 4: Recognition accuracies (%) of models on four OOD datasets." + ], + "table_footnote": [], + "table_body": "
PartATIN-A-PlusIN-SketchSINDINAverage
6.925.86.953.523.3
✓7.425.77.453.923.6
✓5.325.112.555.424.6
✓✓5.726.412.655.925.2
", + "bbox": [ + 339, + 329, + 663, + 395 + ], + "page_idx": 11 + }, + { + "type": "text", + "text": "used an extra $L_{\\mathrm{seg}}$ to introduce part-based supervision. $\\lambda$ was simply set to 1. In addition, we compared MPM with all open-sourced ResNet-50 checkpoints adversarially trained on IN-1K in recent years. The recognition accuracies under different attack threats are shown in Tab. 2. We can see that our baseline model is strong enough, outperforming all other benchmarks by $2.2\\%$ on average robustness (17.1% vs. 19.3%). MPM further shows significantly better adversarial robustness than the vanilla baseline across all these attack threats. In addition, compared with our strong vanilla baseline, the accuracy of clean images is also improved with MPM. To our best knowledge, these are SOTA results of adversarial robustness with ResNet-50 on IN-1K. We additionally show in Appendix G that these results are significantly better than previous results on PIN.", + "bbox": [ + 212, + 425, + 787, + 590 + ], + "page_idx": 11 + }, + { + "type": "text", + "text": "5.3 Robustness on Corruptions and OOD Datasets", + "text_level": 1, + "bbox": [ + 215, + 614, + 648, + 631 + ], + "page_idx": 11 + }, + { + "type": "text", + "text": "Except for robustness against adversarial attacks, which consider the worst-case security threat scenarios, we evaluated the robustness of our part-based model on common image corruptions [24] and several OOD datasets [15, 16, 34, 62], which could be more realistic threats. Considering that AT can have negative effects on these non-adversarial scenarios [35, 41], here we also evaluated models without AT. The training details are also shown in Appendix F.", + "bbox": [ + 212, + 643, + 787, + 734 + ], + "page_idx": 11 + }, + { + "type": "text", + "text": "For common image corruptions [24], we generated different types of corrupted images on the whole IN-1K val set. The results of different models on these images are shown in Tab. 3, where \"Part\" indicates MPM. Here the results of each corruption type are the accuracies averaged on five severity levels [24]. It is seen that MPM has better robustness on nearly all types of image corruptions than models without part supervision, regardless of AT. Besides common corruptions, we evaluated the models on four OOD datasets: ImageNet-A-Plus (IN-A-Plus)", + "bbox": [ + 212, + 734, + 787, + 840 + ], + "page_idx": 11 + }, + { + "type": "page_number", + "text": "12", + "bbox": [ + 217, + 114, + 235, + 126 + ], + "page_idx": 11 + }, + { + "type": "header", + "text": "Li et al.", + "bbox": [ + 271, + 114, + 325, + 126 + ], + "page_idx": 11 + }, + { + "type": "table", + "img_path": "images/5392d36c47977097706d8152cf2fc82b4cde7cd0b21ebee02e47fd73c2ad5916.jpg", + "table_caption": [ + "Table 5: Segmentation accuracies (AP) of Mask R-CNN with and without CF, and recognition accuracies (\\%) of MPM on IN-1K trained with various pseudo labels." + ], + "table_footnote": [], + "table_body": "
Pseudo LabelCFAP\\( AP_{50} \\)Clean\\( l_{\\infty} \\)\\( l_{\\infty}(\\times 2) \\)
--65.533.09.4
✓37.258.667.838.813.1
✓✓40.764.467.839.113.6
", + "bbox": [ + 303, + 184, + 689, + 251 + ], + "page_idx": 12 + }, + { + "type": "text", + "text": "[34], ImageNet Sketch (IN-Sketch) [62], Stylized ImageNet (SIN) [16], and image distortion dataset (denoted as DIN) [15]. See Appendix H for the introduction of these datasets. The results shown in Tab. 4 indicate that MPM also exhibits improved robustness on these OOD shifts.", + "bbox": [ + 214, + 277, + 787, + 339 + ], + "page_idx": 12 + }, + { + "type": "text", + "text": "5.4 Other Advantages of MPM", + "text_level": 1, + "bbox": [ + 214, + 359, + 488, + 376 + ], + "page_idx": 12 + }, + { + "type": "text", + "text": "Models versus humans. The part-based models are motivated by the human recognition process. We were interested in whether the robustness of MPM aligned more with human cognition. To investigate this, we employed the evaluation method introduced by Geirhos et al. [15], which involved comparing the models' decisions on different distorted images with the actual judgments made by human observers. The results shown in Appendix I indicate that MPM measurably improved alignment with the human recognition process.", + "bbox": [ + 212, + 385, + 787, + 491 + ], + "page_idx": 12 + }, + { + "type": "text", + "text": "Boosting the downstream task. Li et al. [33] show that adversarially trained models on the large-scale IN-1K can be utilized to initialize the backbone of downstream networks (e.g., Faster R-CNN [51]). This approach together with downstream AT enables the transfer of adversarial robustness to downstream tasks. We conducted similar experiments by using the checkpoints of the vanilla baseline and MPM. The results shown in Appendix J indicate that MPM enhances both the clean accuracy and adversarial robustness in object detection.", + "bbox": [ + 212, + 491, + 787, + 599 + ], + "page_idx": 12 + }, + { + "type": "text", + "text": "5.5 Ablation Study", + "text_level": 1, + "bbox": [ + 214, + 619, + 390, + 635 + ], + "page_idx": 12 + }, + { + "type": "text", + "text": "Pseudo-label quality. We first investigate the influence of pseudo-label quality on the adversarial robustness of MPM. In Sec. 4.1, we illustrate the use of CF during generating pseudo labels. Here we remove the CF operator and regenerate pseudo labels for images without part annotations. The left half of Tab. 5 (the 2nd and 3rd rows) shows the segmentation accuracies of models with and without CF (here the models were supervised with the train set of $\\mathrm{PIN}++$ , and the AP were calculated on val). It is seen that training with $\\mathrm{PIN}++$ achieved good part segmentation results, indicating its potential for enhancing part-related tasks. CF further significantly improves AP.", + "bbox": [ + 212, + 643, + 787, + 779 + ], + "page_idx": 12 + }, + { + "type": "text", + "text": "We then trained MPM using both pseudo labels and real part annotations, as well as solely using real part annotations. The results, displayed in Tab. 5 (the 1st row), indicate that training MPM solely with real part annotations seems to be ineffective, potentially compromising the model's performance. We guess", + "bbox": [ + 214, + 779, + 787, + 840 + ], + "page_idx": 12 + }, + { + "type": "header", + "text": "PartImageNet++ Dataset: Scaling up Part-based ...", + "bbox": [ + 383, + 114, + 730, + 130 + ], + "page_idx": 12 + }, + { + "type": "page_number", + "text": "13", + "bbox": [ + 767, + 114, + 785, + 127 + ], + "page_idx": 12 + }, + { + "type": "table", + "img_path": "images/51d390e3426bf1ce5614f589f0a9f3b8a5acfecbfb67a18209979da0eba27cbb.jpg", + "table_caption": [ + "Table 6: Recognition accuracies (\\%) of different part-based models on IN-1K. The number of parameters during inference is listed." + ], + "table_footnote": [], + "table_body": "
MethodParam.Cleanl∞l∞(×2)
Two-stage60.9M67.338.713.0
SV167.336.98.8
SV2 w/ TD25.6M67.637.111.7
MPM (SV3 w/ TD)67.839.113.6
SV3 w/o TD67.436.911.7
", + "bbox": [ + 313, + 184, + 679, + 277 + ], + "page_idx": 13 + }, + { + "type": "text", + "text": "it could caused by overfitting on the images with part annotations. Besides, the inclusion of better-quality pseudo labels can improve the robustness of our part-based model, underscoring the importance of high-quality pseudo labels.", + "bbox": [ + 212, + 305, + 782, + 349 + ], + "page_idx": 13 + }, + { + "type": "text", + "text": "Part-based model structure. We then conducted ablation experiments on the structure of the part-based models. Specifically, we compared MPM with the two-stage part-based model designed by Sitawarin et al. [56]. In addition, we performed ablations on the number of part supervisions. MPM incorporates part supervision with three resolutions, referred to as SV3. Here we compared it with SV1 and SV2, which indicate part-based models solely supervised by $f_{l}$ and both $f_{l}$ and $f_{l-1}$ , respectively. Besides these, we extra performed an ablation by removing all top-down (TD) connections of MPM. The results are shown in Tab. 6. We can see that although the two-stage model needs about twice the parameters during inference, its robustness is inferior to the MPM. And all the designs of MPM, including utilizing higher resolution part annotations and top-down connections, contribute to the robustness.", + "bbox": [ + 212, + 349, + 784, + 530 + ], + "page_idx": 13 + }, + { + "type": "text", + "text": "Additional ablations. Additional ablation experiments emphasize the significance of fine-grained part annotations rather than object segmentation annotations. Experiments on Tiny-ImageNet [31] highlight the significance of investigating adversarial robustness on high-resolution images. Moreover, the adversarial robustness of part-based models exhibits low sensitivity to $\\lambda$ . Refer to Appendix K for comprehensive information about these ablation studies.", + "bbox": [ + 212, + 531, + 785, + 621 + ], + "page_idx": 13 + }, + { + "type": "text", + "text": "6 Conclusion and Discussion", + "text_level": 1, + "bbox": [ + 215, + 643, + 508, + 659 + ], + "page_idx": 13 + }, + { + "type": "text", + "text": "In this work, we introduce PIN++, a new large-scale high-quality part dataset, to facilitate the research on part-based models for robust recognition. We propose a new MPM to leverage this dataset. With extensive experiments, we demonstrate the robustness of MPM across diverse scenarios. But we note that MPM represents just a stepping stone in the realm of part-based models for robustness. We hope that this work will inspire and motivate researchers to further explore the potential of part-based models for robust recognition.", + "bbox": [ + 212, + 672, + 784, + 779 + ], + "page_idx": 13 + }, + { + "type": "text", + "text": "Limitation. We believe that except for robust recognition, $\\mathrm{PIN}++$ has the potential to boost general part-related visual understanding tasks. We have not conducted many benchmark experiments specifically for such tasks because they are beyond the scope of this work. We leave these as future work.", + "bbox": [ + 212, + 779, + 785, + 839 + ], + "page_idx": 13 + }, + { + "type": "page_number", + "text": "14", + "bbox": [ + 217, + 114, + 235, + 126 + ], + "page_idx": 13 + }, + { + "type": "header", + "text": "Li et al.", + "bbox": [ + 271, + 114, + 326, + 127 + ], + "page_idx": 13 + }, + { + "type": "text", + "text": "Acknowledgement", + "text_level": 1, + "bbox": [ + 217, + 143, + 393, + 162 + ], + "page_idx": 14 + }, + { + "type": "text", + "text": "This work was supported in part by the National Key Research and Development Program of China (No. 2021ZD0200301) and the National Natural Science Foundation of China (No. U2341228).", + "bbox": [ + 212, + 176, + 787, + 220 + ], + "page_idx": 14 + }, + { + "type": "text", + "text": "References", + "text_level": 1, + "bbox": [ + 217, + 243, + 321, + 258 + ], + "page_idx": 14 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "1. Agarap, A.F.: Deep learning using rectified linear units (relu). arXiv preprint arXiv:1803.08375 (2018)", + "2. Athalye, A., Carlini, N., Wagner, D.A.: Obfuscated gradients give a false sense of security: Circumventing defenses to adversarial examples. In: Int. Conf. Mach. Learn. (ICML). vol. 80, pp. 274-283 (2018)", + "3. Bai, Y., Mei, J., Yuille, A.L., Xie, C.: Are transformers more robust than cnns? In: Adv. Neural Inform. Process. Syst. (NeurIPS). pp. 26831-26843 (2021)", + "4. Biederman, I.: Recognition-by-components: a theory of human image understanding. Psychological Review 94(2), 115 (1987)", + "5. Burl, M.C., Weber, M., Perona, P.: A probabilistic approach to object recognition using local photometry and global geometry. In: Eur. Conf. Comput. Vis. (ECCV). vol. 1407, pp. 628-641 (1998)", + "6. Chen, X., Mottaghi, R., Liu, X., Fidler, S., Urtasun, R., Yuille, A.L.: Detect what you can: Detecting and representing objects using holistic models and body parts. In: IEEE Conf. Comput. Vis. Pattern Recog. (CVPR). pp. 1979-1986 (2014)", + "7. Croce, F., Andriushchenko, M., Sehwag, V., Debenedetti, E., Flammarion, N., Chiang, M., Mittal, P., Hein, M.: Robustbench: a standardized adversarial robustness benchmark. In: NeurIPS Datasets and Benchmarks (2021)", + "8. Croce, F., Hein, M.: Reliable evaluation of adversarial robustness with an ensemble of diverse parameter-free attacks. In: Int. Conf. Mach. Learn. (ICML). Proceedings of Machine Learning Research, vol. 119, pp. 2206-2216 (2020)", + "9. Croce, F., Singh, N.D., Hein, M.: Robust semantic segmentation: Strong adversarial attacks and fast training of robust models. Adv. Neural Inform. Process. Syst. (NeurIPS) (2023)", + "10. Debenedetti, E., Sehwag, V., Mittal, P.: A light recipe to train robust vision transformers. In: First IEEE Conference on Secure and Trustworthy Machine Learning (SaTML). pp. 225-253 (2023)", + "11. Deng, J., Dong, W., Socher, R., Li, L., Li, K., Fei-Fei, L.: Imagenet: A large-scale hierarchical image database. In: IEEE Conf. Comput. Vis. Pattern Recog. (CVPR). pp. 248-255 (2009)", + "12. Eykholt, K., Evtimov, I., Fernandes, E., Li, B., Rahmati, A., Xiao, C., Prakash, A., Kohno, T., Song, D.: Robust physical-world attacks on deep learning visual classification. In: IEEE Conf. Comput. Vis. Pattern Recog. (CVPR). pp. 1625-1634 (2018)", + "13. Fang, Y., Yang, S., Wang, X., Li, Y., Fang, C., Shan, Y., Feng, B., Liu, W.: Instances as queries. In: Int. Conf. Comput. Vis. (ICCV). pp. 6890-6899 (2021)", + "14. Fischler, M.A., Elschlager, R.A.: The representation and matching of pictorial structures. IEEE Trans. Comput. 22(1), 67-92 (1973)", + "15. Geirhos, R., Narayanappa, K., Mitzkus, B., Thieringer, T., Bethge, M., Wichmann, F.A., Brendel, W.: Partial success in closing the gap between human and machine vision. In: Adv. Neural Inform. Process. Syst. (NeurIPS). pp. 23885-23899 (2021)" + ], + "bbox": [ + 218, + 273, + 784, + 839 + ], + "page_idx": 14 + }, + { + "type": "header", + "text": "PartImageNet++ Dataset: Scaling up Part-based ...", + "bbox": [ + 383, + 114, + 730, + 128 + ], + "page_idx": 14 + }, + { + "type": "page_number", + "text": "15", + "bbox": [ + 767, + 116, + 785, + 126 + ], + "page_idx": 14 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "16. Geirhos, R., Rubisch, P., Michaelis, C., Bethge, M., Wichmann, F.A., Brendel, W.: Imagenet-trained cnns are biased towards texture; increasing shape bias improves accuracy and robustness. In: Int. Conf. Learn. Rep. (ICLR) (2019)", + "17. Girshick, R.B., Iandola, F.N., Darrell, T., Malik, J.: Deformable part models are convolutional neural networks. In: IEEE Conf. Comput. Vis. Pattern Recog. (CVPR). pp. 437-446 (2015)", + "18. Gong, K., Liang, X., Li, Y., Chen, Y., Yang, M., Lin, L.: Instance-level human parsing via part grouping network. In: Eur. Conf. Comput. Vis. (ECCV). vol. 11208, pp. 805-822 (2018)", + "19. Goodfellow, I.J., Shlens, J., Szegedy, C.: Explaining and harnessing adversarial examples. In: Int. Conf. Learn. Rep. (ICLR) (2015)", + "20. Haaf, R.A., Fulkerson, A.L., Jablonski, B.J., Hupp, J.M., Shull, S.S., Pescara-Kovach, L.: Object recognition and attention to object components by preschool children and 4-month-old infants. Journal of Experimental Child Psychology 86(2), 108-123 (2003)", + "21. He, J., Yang, S., Yang, S., Kortylewski, A., Yuan, X., Chen, J., Liu, S., Yang, C., Yu, Q., Yuille, A.L.: Partimagenet: A large, high-quality dataset of parts. In: Eur. Conf. Comput. Vis. (ECCV). vol. 13668, pp. 128-145 (2022)", + "22. He, K., Gkioxari, G., Dollár, P., Girshick, R.B.: Mask R-CNN. IEEE Trans. Pattern Anal. Mach. Intell. (TPAMI) 42(2), 386-397 (2020)", + "23. He, K., Zhang, X., Ren, S., Sun, J.: Deep residual learning for image recognition. In: IEEE Conf. Comput. Vis. Pattern Recog. (CVPR). pp. 770-778 (2016)", + "24. Hendrycks, D., Dietterich, T.G.: Benchmarking neural network robustness to common corruptions and perturbations. In: Int. Conf. Learn. Rep. (ICLR) (2019)", + "25. Hendrycks, D., Gimpel, K.: Gaussian error linear units (gelus). arXiv preprint arXiv:1606.08415 (2016)", + "26. Hoffman, J.E., Subramaniam, B.: The role of visual attention in saccadic eye movements. Perception & Psychophysics 57(6), 787-795 (1995)", + "27. Ilyas, A., Santurkar, S., Tsipras, D., Engstrom, L., Tran, B., Madry, A.: Adversarial examples are not bugs, they are features. In: Adv. Neural Inform. Process. Syst. (NeurIPS). pp. 125-136 (2019)", + "28. Kirillov, A., Mintun, E., Ravi, N., Mao, H., Rolland, C., Gustafson, L., Xiao, T., Whitehead, S., Berg, A.C., Lo, W.Y., Dollar, P., Girshick, R.: Segment anything. In: Int. Conf. Comput. Vis. (ICCV). pp. 4015-4026 (2023)", + "29. Krizhevsky, A., Hinton, G., et al.: Learning multiple layers of features from tiny images (2009)", + "30. Krizhevsky, A., Sutskever, I., Hinton, G.E.: Imagenet classification with deep convolutional neural networks. In: Adv. Neural Inform. Process. Syst. (NeurIPS). pp. 1106-1114 (2012)", + "31. Le, Y., Yang, X.: Tiny imagenet visual recognition challenge. CS 231N 7(7), 3 (2015)", + "32. Li, L.H., Zhang, P., Zhang, H., Yang, J., Li, C., Zhong, Y., Wang, L., Yuan, L., Zhang, L., Hwang, J., Chang, K., Gao, J.: Grounded language-image pre-training. In: IEEE Conf. Comput. Vis. Pattern Recog. (CVPR). pp. 10955-10965 (2022)", + "33. Li, X., Chen, H., Hu, X.: On the importance of backbone to the adversarial robustness of object detectors. arXiv preprint arXiv:2305.17438 (2023)", + "34. Li, X., Li, J., Dai, T., Shi, J., Zhu, J., Hu, X.: Rethinking natural adversarial examples for classification models. arXiv preprint arXiv:2102.11731 (2021)", + "35. Li, X., Wang, Z., Zhang, B., Sun, F., Hu, X.: Recognizing object by components with human prior knowledge enhances adversarial robustness of deep neural networks. IEEE Trans. Pattern Anal. Mach. Intell. (TPAMI) 45(7), 8861-8873 (2023)" + ], + "bbox": [ + 215, + 146, + 784, + 839 + ], + "page_idx": 15 + }, + { + "type": "page_number", + "text": "16", + "bbox": [ + 217, + 114, + 235, + 126 + ], + "page_idx": 15 + }, + { + "type": "header", + "text": "Li et al.", + "bbox": [ + 271, + 114, + 325, + 126 + ], + "page_idx": 15 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "36. Li, X., Zhang, W., Liu, Y., Hu, Z., Zhang, B., Hu, X.: Language-driven anchors for zero-shot adversarial robustness. In: IEEE Conf. Comput. Vis. Pattern Recog. (CVPR) (2024)", + "37. Liang, X., Gong, K., Shen, X., Lin, L.: Look into person: Joint body parsing & pose estimation network and a new benchmark. IEEE Trans. Pattern Anal. Mach. Intell. (TPAMI) 41(4), 871-885 (2019)", + "38. Liang, X., Liu, S., Shen, X., Yang, J., Liu, L., Dong, J., Lin, L., Yan, S.: Deep human parsing with active template regression. IEEE Trans. Pattern Anal. Mach. Intell. (TPAMI) 37(12), 2402-2414 (2015)", + "39. Lin, C., Sun, P., Jiang, Y., Luo, P., Qu, L., Haffari, G., Yuan, Z., Cai, J.: Learning object-language alignments for open-vocabulary object detection. In: Int. Conf. Learn. Rep. (ICLR) (2023)", + "40. Lin, T., Dollar, P., Girshick, R.B., He, K., Hariharan, B., Belongie, S.J.: Feature pyramid networks for object detection. In: IEEE Conf. Comput. Vis. Pattern Recog. (CVPR). pp. 936-944 (2017)", + "41. Liu, C., Dong, Y., Xiang, W., Yang, X., Su, H., Zhu, J., Chen, Y., He, Y., Xue, H., Zheng, S.: A comprehensive study on robustness of image classification models: Benchmarking and rethinking. arXiv preprint arXiv:2302.14301 (2023)", + "42. Liu, Q., Kortylewski, A., Bai, Y., Bai, S., Yuille, A.L.: Intriguing properties of text-guided diffusion models. arXiv preprint arXiv:2306.00974 (2023)", + "43. Liu, S., Zhang, L., Yang, X., Su, H., Zhu, J.: Unsupervised part segmentation through disentangling appearance and shape. In: IEEE Conf. Comput. Vis. Pattern Recog. (CVPR). pp. 8355-8364 (2021)", + "44. Liu, Z., Lin, Y., Cao, Y., Hu, H., Wei, Y., Zhang, Z., Lin, S., Guo, B.: Swin transformer: Hierarchical vision transformer using shifted windows. In: Int. Conf. Comput. Vis. (ICCV). pp. 9992-10002 (2021)", + "45. Liu, Z., Mao, H., Wu, C., Feichtenhofer, C., Darrell, T., Xie, S.: A convnet for the 2020s. In: IEEE Conf. Comput. Vis. Pattern Recog. (CVPR). pp. 11966-11976 (2022)", + "46. Madry, A., Makelov, A., Schmidt, L., Tsipras, D., Vladu, A.: Towards deep learning models resistant to adversarial attacks. In: Int. Conf. Learn. Rep. (ICLR) (2018)", + "47. Mao, X., Chen, Y., Li, X., Qi, G., Duan, R., Zhang, R., Xue, H.: Easyrobust: A comprehensive and easy-to-use toolkit for robust computer vision. https://github.com/alibaba/easyrobust (2022)", + "48. Meletis, P., Wen, X., Lu, C., de Geus, D., Dubbelman, G.: Cityscapes-panoptic-parts and pascal-panoptic-parts datasets for scene understanding. arXiv preprint arXiv:2004.07944 (2020)", + "49. Ramanathan, V., Kalia, A., Petrovic, V., Wen, Y., Zheng, B., Guo, B., Wang, R., Marquez, A., Kovvuri, R., Kadian, A., Mousavi, A., Song, Y., Dubey, A., Mahajan, D.: PACO: parts and attributes of common objects. In: IEEE Conf. Comput. Vis. Pattern Recog. (CVPR). pp. 7141-7151 (2023)", + "50. Reddy, N.D., Vo, M., Narasimhan, S.G.: Carfusion: Combining point tracking and part detection for dynamic 3d reconstruction of vehicles. In: IEEE Conf. Comput. Vis. Pattern Recog. (CVPR). pp. 1906-1915 (2018)", + "51. Ren, S., He, K., Girshick, R.B., Sun, J.: Faster R-CNN: towards real-time object detection with region proposal networks. In: Adv. Neural Inform. Process. Syst. (NeurIPS). pp. 91-99 (2015)", + "52. Rombach, R., Blattmann, A., Lorenz, D., Esser, P., Ommer, B.: High-resolution image synthesis with latent diffusion models. In: IEEE Conf. Comput. Vis. Pattern Recog. (CVPR). pp. 10674-10685 (2022)" + ], + "bbox": [ + 212, + 146, + 787, + 840 + ], + "page_idx": 16 + }, + { + "type": "header", + "text": "PartImageNet++ Dataset: Scaling up Part-based ...", + "bbox": [ + 383, + 114, + 730, + 128 + ], + "page_idx": 16 + }, + { + "type": "page_number", + "text": "17", + "bbox": [ + 767, + 114, + 785, + 126 + ], + "page_idx": 16 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "53. Rosch, E., Mervis, C.B.: Family resemblances: Studies in the internal structure of categories. Cognitive Psychology 7(4), 573-605 (1975)", + "54. Salman, H., Ilyas, A., Engstrom, L., Kapoor, A., Madry, A.: Do adversarially robust imagenet models transfer better? In: Adv. Neural Inform. Process. Syst. (NeurIPS) (2020)", + "55. Salman, H., Sun, M., Yang, G., Kapoor, A., Kolter, J.Z.: Denoised smoothing: A provable defense for pretrained classifiers. In: Adv. Neural Inform. Process. Syst. (NeurIPS) (2020)", + "56. Sitawarin, C., Pongmala, K., Chen, Y., Carlini, N., Wagner, D.A.: Part-based models improve adversarial robustness. In: Int. Conf. Learn. Rep. (ICLR) (2023)", + "57. Song, X., Wang, P., Zhou, D., Zhu, R., Guan, C., Dai, Y., Su, H., Li, H., Yang, R.: Apollocar3d: A large 3d car instance understanding benchmark for autonomous driving. In: IEEE Conf. Comput. Vis. Pattern Recog. (CVPR). pp. 5452-5462 (2019)", + "58. Sun, P., Chen, S., Zhu, C., Xiao, F., Luo, P., Xie, S., Yan, Z.: Going denser with open-vocabulary part segmentation. arXiv preprint arXiv:2305.11173 (2023)", + "59. Szegedy, C., Zaremba, W., Sutskever, I., Bruna, J., Erhan, D., Goodfellow, I.J., Fergus, R.: Intriguing properties of neural networks. In: Int. Conf. Learn. Rep. (ICLR) (2014)", + "60. Tramèr, F., Carlini, N., Brendel, W., Madry, A.: On adaptive attacks to adversarial example defenses. In: Adv. Neural Inform. Process. Syst. (NeurIPS) (2020)", + "61. Tversky, B., Hemenway, K.: Objects, parts, and categories. Journal of Experimental Psychology: General 113(2), 169 (1984)", + "62. Wang, H., Ge, S., Lipton, Z.C., Xing, E.P.: Learning robust global representations by penalizing local predictive power. In: Adv. Neural Inform. Process. Syst. (NeurIPS). pp. 10506-10518 (2019)", + "63. Wang, J., Zhang, Z., Xie, C., Premachandran, V., Yuille, A.: Unsupervised learning of object semantic parts from internal states of cnns by population encoding. arXiv preprint arXiv:1511.06855 (2015)", + "64. Wu, D., Xia, S., Wang, Y.: Adversarial weight perturbation helps robust generalization. In: Adv. Neural Inform. Process. Syst. (NeurIPS) (2020)", + "65. Zhang, H., Yu, Y., Jiao, J., Xing, E.P., Ghaoui, L.E., Jordan, M.I.: Theoretically principled trade-off between robustness and accuracy. In: Int. Conf. Mach. Learn. (ICML). vol. 97, pp. 7472-7482 (2019)", + "66. Zhang, Y., Yao, Y., Jia, J., Yi, J., Hong, M., Chang, S., Liu, S.: How to robustify black-box ML models? A zeroth-order optimization perspective. In: Int. Conf. Learn. Rep. (ICLR) (2022)", + "67. Zhao, J., Li, J., Cheng, Y., Sim, T., Yan, S., Feng, J.: Understanding humans in crowded scenes: Deep nested adversarial learning and A new benchmark for multi-human parsing. In: ACM Int. Conf. Multimedia (ACM MM). pp. 792-800 (2018)", + "68. Zhao, Y., Li, J., Zhang, Y., Tian, Y.: Multi-class part parsing with joint boundary-semantic awareness. In: Int. Conf. Comput. Vis. (ICCV). pp. 9176-9185 (2019)", + "69. Zhou, B., Zhao, H., Puig, X., Xiao, T., Fidler, S., Barriuso, A., Torralba, A.: Semantic understanding of scenes through the ADE20K dataset. Int. J. Comput. Vis. (IJCV) 127(3), 302-321 (2019)", + "70. Zhou, Z., Firestone, C.: Humans can decipher adversarial images. Nature Communications 10(1), 1-9 (2019)" + ], + "bbox": [ + 215, + 147, + 784, + 811 + ], + "page_idx": 17 + }, + { + "type": "page_number", + "text": "18", + "bbox": [ + 217, + 114, + 235, + 126 + ], + "page_idx": 17 + }, + { + "type": "header", + "text": "Li et al.", + "bbox": [ + 271, + 114, + 325, + 126 + ], + "page_idx": 17 + } +] \ No newline at end of file diff --git a/2024/PartImageNet++ Dataset_ Scaling up Part-based Models for Robust Recognition/0d30f97b-2cbe-4f46-92c7-0bf8c016e1e3_model.json b/2024/PartImageNet++ Dataset_ Scaling up Part-based Models for Robust Recognition/0d30f97b-2cbe-4f46-92c7-0bf8c016e1e3_model.json new file mode 100644 index 0000000000000000000000000000000000000000..52a90e88bbde97247b0d53b56377bf23a0d3863e --- /dev/null +++ b/2024/PartImageNet++ Dataset_ Scaling up Part-based Models for Robust Recognition/0d30f97b-2cbe-4f46-92c7-0bf8c016e1e3_model.json @@ -0,0 +1,2370 @@ +[ + [ + { + "type": "title", + "bbox": [ + 0.219, + 0.141, + 0.784, + 0.187 + ], + "angle": 0, + "content": "PartImageNet++ Dataset: Scaling up Part-based Models for Robust Recognition" + }, + { + "type": "text", + "bbox": [ + 0.252, + 0.212, + 0.751, + 0.229 + ], + "angle": 0, + "content": "Xiao Li\\(^{1}\\), Yining Liu\\(^{2}\\), Na Dong\\(^{3}\\), Sitian Qin\\(^{2}\\), and Xiaolin Hu\\(^{1,4*}\\)" + }, + { + "type": "text", + "bbox": [ + 0.223, + 0.24, + 0.78, + 0.352 + ], + "angle": 0, + "content": "\\(^{1}\\) Department of Computer Science and Technology, BNRist, Institute for Artificial Intelligence, IDG/McGovern Institute for Brain Research, Tsinghua Laboratory of Brain and Intelligence, Tsinghua University, Beijing, China \n\\(^{2}\\) Harbin Institute of Technology, Weihai, China \n\\(^{3}\\) Beijing Institute of Technology, Beijing, China \n\\(^{4}\\) Chinese Institute for Brain Research (CIBR), Beijing, China \nlixiao20@mails.tsinghua.edu.cn, 22S030184@stu.hit.edu.cn, 3220220911@bit.edu.cn, qinsitian@hitwh.edu.cn, xlu@mail.tsinghua.edu.cn" + }, + { + "type": "text", + "bbox": [ + 0.262, + 0.392, + 0.741, + 0.683 + ], + "angle": 0, + "content": "Abstract. Deep learning-based object recognition systems can be easily fooled by various adversarial perturbations. One reason for the weak robustness may be that they do not have part-based inductive bias like the human recognition process. Motivated by this, several part-based recognition models have been proposed to improve the adversarial robustness of recognition. However, due to the lack of part annotations, the effectiveness of these methods is only validated on small-scale non-standard datasets. In this work, we propose PIN++, short for PartImageNet++, a dataset providing high-quality part segmentation annotations for all categories of ImageNet-1K (IN-1K). With these annotations, we build part-based methods directly on the standard IN-1K dataset for robust recognition. Different from previous two-stage part-based models, we propose a Multi-scale Part-supervised Model (MPM), to learn a robust representation with part annotations. Experiments show that MPM yielded better adversarial robustness on the large-scale IN-1K over strong baselines across various attack settings. Furthermore, MPM achieved improved robustness on common corruptions and several out-of-distribution datasets. The dataset, together with these results, enables and encourages researchers to explore the potential of part-based models in more real applications. The dataset and the code are available at https://github.com/LixiaoTHU/PartImageNetPP." + }, + { + "type": "text", + "bbox": [ + 0.261, + 0.696, + 0.741, + 0.723 + ], + "angle": 0, + "content": "Keywords: Adversarial robustness \\(\\cdot\\) Part dataset \\(\\cdot\\) Robust object recognition \\(\\cdot\\) Part-based model" + }, + { + "type": "title", + "bbox": [ + 0.217, + 0.752, + 0.377, + 0.768 + ], + "angle": 0, + "content": "1 Introduction" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.785, + 0.788, + 0.816 + ], + "angle": 0, + "content": "Object recognition has achieved remarkable success due to the rise of Deep Neural Networks (DNNs) [23, 30, 44, 45]. However, DNNs often lack robustness and" + }, + { + "type": "page_footnote", + "bbox": [ + 0.218, + 0.825, + 0.386, + 0.84 + ], + "angle": 0, + "content": "* Corresponding author." + } + ], + [ + { + "type": "page_number", + "bbox": [ + 0.218, + 0.116, + 0.23, + 0.127 + ], + "angle": 0, + "content": "2" + }, + { + "type": "header", + "bbox": [ + 0.272, + 0.115, + 0.328, + 0.127 + ], + "angle": 0, + "content": "Li et al." + }, + { + "type": "image", + "bbox": [ + 0.317, + 0.144, + 0.495, + 0.299 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.496, + 0.144, + 0.593, + 0.298 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.593, + 0.144, + 0.687, + 0.298 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.214, + 0.309, + 0.788, + 0.352 + ], + "angle": 0, + "content": "Fig. 1: Examples of annotated images in \\(\\mathrm{PIN}++\\). High-quality part segmentation annotations are provided on all categories in IN-1K. The object names are shown on the top-right of each image. The part names are hidden here." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.384, + 0.788, + 0.581 + ], + "angle": 0, + "content": "can be easily deceived by adversarial examples [19, 59], common image corruptions [24], and various out-of-distribution (OOD) shifts [34, 62], which greatly hinder applications of DNNs in security-critical scenarios [12, 33]. One possible reason for the weakness may be that DNN-based recognition does not have a part-based inductive bias (i.e., recognition based on object parts) like the human recognition process, which is generally regarded as a robust system [70]. According to the well-known cognitive psychological theory recognition-by-components [4], humans prefer to recognize objects by decomposing objects into parts and taking into account the hierarchical representations and the spatial relationships of parts. This cognitive theory has been supported by numerous psychological evidence [20, 26, 53, 61]. However, several studies revealed that DNNs often exhibit a shortcut recognition behavior, e.g., relying on texture-based features [16] and even unintelligible features for humans [27]." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.582, + 0.788, + 0.718 + ], + "angle": 0, + "content": "If DNNs can perform a human-aligned part-based recognition, both the adversarial and non-adversarial robustness of DNNs could be closer to humans. Motivated by this idea, two recent works [35, 56] investigated how to use additional part-level supervision to build up part-based (recognition) models, which make predictions by recognizing the object's parts in a bottom-up manner. Both works investigated two-stage part-based models, where the first stage involves part segmentation, followed by the classification stage based on the segmentation results. These works have shown the potential of part-based models in enhancing robustness against various perturbations." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.72, + 0.788, + 0.841 + ], + "angle": 0, + "content": "Nevertheless, due to the lack of part annotations, the effectiveness of part-based models has only been preliminary validated by the two works [35, 56] on small-scale nonstandard datasets. The largest dataset for evaluation was PartImageNet (PIN) [21], consisting of about \\(24\\mathrm{K}\\) images of 158 categories selected from the large-scale ImageNet-1K (IN-1K) [11]. PIN has several potential issues. Firstly, it is quite small compared to the standard IN-1K (about \\(1.3\\mathrm{M}\\) images) commonly used in the domain of adversarial robustness [3, 54]. This limited scale may restrict the generalizability of the findings. Secondly, categories included in" + } + ], + [ + { + "type": "header", + "bbox": [ + 0.384, + 0.115, + 0.732, + 0.129 + ], + "angle": 0, + "content": "PartImageNet++ Dataset: Scaling up Part-based ..." + }, + { + "type": "page_number", + "bbox": [ + 0.775, + 0.117, + 0.785, + 0.127 + ], + "angle": 0, + "content": "3" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.147, + 0.788, + 0.314 + ], + "angle": 0, + "content": "PIN are specifically chosen to facilitate part-related tasks (e.g., part segmentation [43] and object parsing [68]), with an emphasis on animal categories (118 out of 158). While this focus can be beneficial for studying part-related tasks themselves, it poses challenges in demonstrating the effectiveness of part-based models for general recognition tasks across a diverse range of objects. Thirdly, the annotation scheme of PIN is simple. It groups 158 categories into 11 super-categories and determines part annotations for each super-category, potentially failing to capture the characteristics of object categories within a super-category. Due to the three issues, the part-based models [35,56] trained on PIN cannot be directly compared with state-of-the-art (SOTA) methods trained on the standard IN-1K." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.323, + 0.788, + 0.535 + ], + "angle": 0, + "content": "In this work, we introduce PartImageNet++ (PIN++, a new dataset providing high-quality part segmentation annotations for all categories of the widely used IN-1K [11], to facilitate the research on part-based models for robust recognition (classification). To create PIN++, we designed a detailed annotation scheme to ensure high-quality part annotations. With this scheme, we annotated 100 randomly selected images per category, resulting in a total of 100K images with part annotations (about 1/13 of IN-1K). To avoid bias, all the part annotations are created manually without using any auxiliary model. To the best of our knowledge, PIN++ provides large-scale high-quality part segmentation annotations for the most diverse range of object categories, including creatures, artifacts, rigid objects, nonrigid objects, etc, among existing part datasets. Fig. 1 showcases some examples of annotated images. We introduce this dataset in detail in Sec. 3 and show that existing techniques fail to obtain high-quality part segmentation results without such annotations." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.544, + 0.788, + 0.741 + ], + "angle": 0, + "content": "By leveraging the part annotations of \\(\\mathrm{PIN}++\\), we develop part-based methods for robust recognition directly on the standard IN-1K dataset. To achieve this, we first train a part segmentation network with the annotations of \\(\\mathrm{PIN}++\\) to obtain pseudo part labels for all unannotated images in IN-1K. We then propose a new Multi-scale Part-supervised recognition Model, termed MPM, to better exploit the part annotations. MPM is expected to learn a robust intermediate representation by adding auxiliary bypass layers to the vanilla recognition model (e.g., ResNet-50 [23]). The auxiliary bypass layers are supervised by the part annotations and the generated pseudo part labels. Without introducing any additional parameter or computation during inference, MPM can use high-resolution part annotations more adequately than two-stage part-based models [35, 56], which are constrained to use low-resolution part labels. Fig. 3 shows an overview of our whole method." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.75, + 0.788, + 0.842 + ], + "angle": 0, + "content": "The experimental results demonstrate that MPM, combined with adversarial training (AT) [46], achieved better adversarial robustness on the large-scale IN-1K, outperforming strong AT baselines in both the seen and unseen adversarial threats. Furthermore, MPM exhibited improved robustness on common image corruptions [24] and several OOD datasets. In addition, MPM achieved higher alignments with human vision. Beyond object recognition, MPM is shown to" + } + ], + [ + { + "type": "page_number", + "bbox": [ + 0.218, + 0.116, + 0.23, + 0.127 + ], + "angle": 0, + "content": "4" + }, + { + "type": "header", + "bbox": [ + 0.272, + 0.115, + 0.326, + 0.127 + ], + "angle": 0, + "content": "Li et al." + }, + { + "type": "text", + "bbox": [ + 0.217, + 0.147, + 0.785, + 0.178 + ], + "angle": 0, + "content": "boost the adversarial robustness of downstream tasks such as object detection [51]. The main contributions of this work can be summarized as follows:" + }, + { + "type": "text", + "bbox": [ + 0.224, + 0.185, + 0.784, + 0.228 + ], + "angle": 0, + "content": "- We created PIN++, a new large-scale dataset with part annotations. This dataset can facilitate further research on part-based models for robust recognition and other part-related visual understanding tasks." + }, + { + "type": "text", + "bbox": [ + 0.226, + 0.23, + 0.784, + 0.273 + ], + "angle": 0, + "content": "- We proposed MPM, a new part-supervised recognition model, to exploit the part annotations better, achieving superior robustness across various settings and benchmarks without introducing any extra inference cost." + }, + { + "type": "list", + "bbox": [ + 0.224, + 0.185, + 0.784, + 0.273 + ], + "angle": 0, + "content": null + }, + { + "type": "title", + "bbox": [ + 0.217, + 0.297, + 0.387, + 0.312 + ], + "angle": 0, + "content": "2 Related Work" + }, + { + "type": "text", + "bbox": [ + 0.216, + 0.327, + 0.787, + 0.493 + ], + "angle": 0, + "content": "Adversarial robustness. Szegedy et al. [59] first revealed the vulnerability of DNNs to adversarial examples. Since then, numerous methods [55,66] have been proposed to enhance the adversarial robustness of DNNs. Among them, AT has emerged as the de facto paradigm for training adversially robust DNNs. Early AT works [64,65] primarily focused on small-scale, low-resolution datasets such as CIFAR-10 [29]. Recently, there has been a growing interest in investigating AT on IN-1K [11], as this dataset serves as a standard benchmark for evaluating SOTA computer vision techniques [23,44,45]. Several studies have demonstrated that AT models trained on IN-1K can transfer their adversarial robustness to downstream dense-prediction tasks and achieve zero-shot adversially robust recognition [9, 33, 36]." + }, + { + "type": "text", + "bbox": [ + 0.216, + 0.493, + 0.787, + 0.674 + ], + "angle": 0, + "content": "Part datasets. The concept of parts holds significant importance in both human visual cognition [4] and computer vision domains. To conduct research on part-related visual tasks, the availability of part datasets is crucial. However, annotating object parts is challenging and expensive, and most part datasets are only limited to specific domains, such as cars [50, 57, 63] and humans [18, 37, 38, 67]. Furthermore, part datasets targeting common objects often have a limited number of object categories. Cityscapes PanopticParts (Cityscapes PP.) [48], PascalPart [6], and ADE20K provide part annotations for 5, 20, and 80 object categories, respectively. Although the recent PACO dataset [49] offers a large number of part annotations, it is still restricted to only 75 object categories. PIN [21], containing 158 categories from IN-1K [11], represents the dataset with the most object categories in terms of part annotations." + }, + { + "type": "text", + "bbox": [ + 0.216, + 0.675, + 0.787, + 0.84 + ], + "angle": 0, + "content": "Part-based object recognition. Modeling objects in terms of parts has a long history in computer vision [5,14,17]. But with the rise of DNNs, parts have been rarely used as auxiliaries for end-to-end general object recognition [23,30,44,45]. Recently, two works [35, 56] revisited part-based recognition models from the robustness perspective and showed the potential to enhance adversarial robustness. Li et al. [35] proposed ROCK, a part-based model that first predicts all parts represented by segmentation results, and then utilizes a judgment block to give category predictions based on the predicted parts. Similarly, Sitawarin et al. [56] also investigated several types of two-stage part-based models, where the first stage involves part segmentation, followed by the recognition stage with a tiny classifier. However, as mentioned before, both works were only validated on" + } + ], + [ + { + "type": "header", + "bbox": [ + 0.384, + 0.115, + 0.732, + 0.131 + ], + "angle": 0, + "content": "PartImageNet++ Dataset: Scaling up Part-based ..." + }, + { + "type": "page_number", + "bbox": [ + 0.775, + 0.117, + 0.785, + 0.127 + ], + "angle": 0, + "content": "5" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.147, + 0.788, + 0.254 + ], + "angle": 0, + "content": "small-scale part-friendly datasets for the lack of part annotations, e.g., Sitawarin et al. [56] performed classification tasks only on 11 super-categories of PIN [21]. Furthermore, the proposed two-stage part-based models inevitably introduced a lot of extra parameters during inference, while the utilization of high-resolution part annotations was limited, as discussed in Sec. 4.2. In addition, ROCK [35] contains a non-differentiable process, which could lead to a potential overestimation of adversarial robustness [2,60]. Our work will address all of these issues." + }, + { + "type": "title", + "bbox": [ + 0.215, + 0.275, + 0.411, + 0.291 + ], + "angle": 0, + "content": "3 PIN++ Dataset" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.304, + 0.787, + 0.351 + ], + "angle": 0, + "content": "We first present the details of how we built \\(\\mathrm{PIN}++\\) to ensure high-quality part annotations, followed by the statistics of this dataset. We then give a discussion on \\(\\mathrm{PIN}++\\)." + }, + { + "type": "title", + "bbox": [ + 0.215, + 0.371, + 0.427, + 0.387 + ], + "angle": 0, + "content": "3.1 Annotation Scheme" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.394, + 0.789, + 0.56 + ], + "angle": 0, + "content": "Data source. IN-1K [11], containing about 1.3M images with category labels in the training set, is one of the most widely used datasets for general object recognition. Our aim is to provide part annotations for part-based robust recognition on IN-1K, and thus the images of PIN++ directly reuse the training set of IN-1K. All the images conform to licensing for research purposes. However, due to cost constraints, it is impractical to provide part-level annotations for all training images in IN-1K. Instead, we decided to annotate 100 randomly selected images per category. Once part annotations are provided for all categories, modern supervised segmentation techniques [22] can be used to obtain pseudo part labels for the remaining unannotated images, as we will show in Sec. 4. Note that PIN++ does not provide any part annotations for the validation set of IN-1K." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.561, + 0.789, + 0.666 + ], + "angle": 0, + "content": "While PIN [21] has provided some part annotations for 158 categories of IN-1K, their annotations do not fully meet our purpose and principles. To minimize annotation costs without compromising quality, we retained annotations of 90 categories from PIN as part of \\(\\mathrm{PIN}++\\). The details are shown in Appendix A. We then show the annotation design for the remaining 910 categories of IN-1K. Annotation quality control. To ensure high-quality part annotations, three points should be taken into account:" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.666, + 0.789, + 0.818 + ], + "angle": 0, + "content": "1) Deciding which parts to annotate per category. One of the key challenges in the part annotation task is the ambiguity of object part selection (e.g., how to annotate the parts of hammer). PIN [21] evades this problem and focuses primarily on annotating parts for animal categories (most quadrupeds can be regarded as consisting of four part categories: head, body, foot, tail). However, the remaining categories in IN-1K are highly diverse, making it challenging to decide which parts to annotate. To ensure a scientifically grounded division of parts for each category, we initially consult the Wikidata knowledge base5 to obtain part vocabularies for every object category (e.g., hammer consists of three parts according to Wikidata: handle, striker, and hammerhead). Additionally, for object" + }, + { + "type": "page_footnote", + "bbox": [ + 0.218, + 0.825, + 0.633, + 0.841 + ], + "angle": 0, + "content": "5 https://www.wikidata.org/wiki/Wikidata:Introduction" + } + ], + [ + { + "type": "page_number", + "bbox": [ + 0.218, + 0.116, + 0.23, + 0.127 + ], + "angle": 0, + "content": "6" + }, + { + "type": "header", + "bbox": [ + 0.272, + 0.115, + 0.327, + 0.127 + ], + "angle": 0, + "content": "Li et al." + }, + { + "type": "table_caption", + "bbox": [ + 0.216, + 0.145, + 0.785, + 0.173 + ], + "angle": 0, + "content": "Table 1: The numbers of annotated object categories, part categories, images and part masks of publicly available part datasets." + }, + { + "type": "table", + "bbox": [ + 0.248, + 0.185, + 0.756, + 0.292 + ], + "angle": 0, + "content": "
DatasetObject CategoryPart CategoryImagePart Mask
Cityscapes PP. [48]5233.5K100K
Pascal-Part [6]2019319K363.5K
ADE20K* [69]8056612.6K193.2K
PACO [49]7545676.7K641.4K
PIN [21]15860924K112K
PIN++10003308100K406.4K
" + }, + { + "type": "table_footnote", + "bbox": [ + 0.238, + 0.294, + 0.784, + 0.322 + ], + "angle": 0, + "content": "*: ADE20K is severely category-unbalanced and here we show statistics for categories with more than 10 annotated part masks." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.351, + 0.787, + 0.502 + ], + "angle": 0, + "content": "categories without a clear definition of parts on Wikidata, we decide which parts to annotate by asking recruited volunteers which parts they think to be important to their cognition of the object. For example, according to the cognition of volunteers, for most quadruped categories, annotating head, body, foot, and tail is enough, while for ram (see Fig. 1 for one example), horn should be annotated as an extra part. For categories that are indeed difficult to decompose into parts, e.g., flatworm, we generalize the concept of \"parts\" and treat the foreground of an object as one part category. In this case, the part is the object itself. Other details on deciding the parts to annotate are further discussed in Appendix B. These principles ensure that all categories in IN-1K can have part annotations." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.503, + 0.787, + 0.669 + ], + "angle": 0, + "content": "2) Designing the part segmentation principles. We design several part segmentation principles to guide the annotators and ensure high-quality part segmentation annotation. Firstly, the annotated part masks are required to be combined to cover the entire object and have no overlap, unless in the case of the second principle. Secondly, for some parts that indeed should overlap in semantic concepts (e.g., horn and head), the annotators should annotate the inclusion relation of parts (e.g., horn is included in head), and then these part masks can be annotated with overlap. Thirdly, to keep the consistency of part annotations among images of the same object category, following [21], the same annotator should annotate all 100 images of this object category. The other principles are shown in Appendix C." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.669, + 0.787, + 0.744 + ], + "angle": 0, + "content": "3) Annotation quality inspection. Ten randomly selected images together with the annotation visualizations (see Fig. 1) are sent to inspectors to assess if the annotations for a specific object category meet the defined principles. If the annotations for any two images fail to meet these principles, the entire category's annotations are re-annotated until the requirements are satisfied." + }, + { + "type": "title", + "bbox": [ + 0.216, + 0.768, + 0.339, + 0.782 + ], + "angle": 0, + "content": "3.2 Statistics" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.795, + 0.787, + 0.84 + ], + "angle": 0, + "content": "We finally annotated 100 images per category for 910 categories in IN-1K, resulting in a total of 91K images with part annotations. Taking these data and the data of 90 reused categories of PIN together, \\(\\mathrm{PIN}++\\) for 1K categories is" + } + ], + [ + { + "type": "header", + "bbox": [ + 0.384, + 0.115, + 0.731, + 0.129 + ], + "angle": 0, + "content": "PartImageNet++ Dataset: Scaling up Part-based ..." + }, + { + "type": "page_number", + "bbox": [ + 0.775, + 0.117, + 0.785, + 0.127 + ], + "angle": 0, + "content": "7" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.147, + 0.784, + 0.221 + ], + "angle": 0, + "content": "obtained. Following IN-1K, here we ensure PIN++ is class-balanced. PIN++ is split into train/val sets with a 9:1 ratio for each category. But note that all annotated images are part of the training set of IN-1K. Unless stated otherwise, they are all utilized for training in our following experiments and the evaluation is performed on IN-1K val set." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.224, + 0.784, + 0.284 + ], + "angle": 0, + "content": "In summary, \\(\\mathrm{PIN}++\\) includes a total of 100K images with part annotations. These annotations cover 3,310 part categories for 1,000 object categories. A total of 406,364 part masks are annotated in \\(\\mathrm{PIN}++\\). Further statistics of \\(\\mathrm{PIN}++\\) are shown in Appendix D." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.285, + 0.785, + 0.466 + ], + "angle": 0, + "content": "Comparison with other part datasets. We compare PIN++ with publicly available part datasets, as shown in Tab. 1. PIN++ outperforms previous datasets in terms of the number of object and part categories, as well as the total number of images with annotated part masks. It also demonstrates the competitive number of part masks compared to the recent PACO dataset [49]. Notably, PIN++ offers part annotations for a diverse range of objects, including creatures, artifacts, rigid objects, and nonrigid objects, unlike PACO or PIN, which are primarily focused on common tools or animals, respectively. As shown in Appendix D, the annotation quality of PIN++ exhibits an advantage over that of PIN. Moreover, PIN++ is category-balanced and guarantees that each image contains only one foreground category. All these features build the unique advantage of PIN++ for studying object recognition tasks." + }, + { + "type": "title", + "bbox": [ + 0.216, + 0.493, + 0.444, + 0.508 + ], + "angle": 0, + "content": "3.3 Discussion on PIN++" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.522, + 0.785, + 0.718 + ], + "angle": 0, + "content": "Recent advancements in segmentation include CLIP-based open-vocabulary segmentation methods [32,39,58] and the SAM model [28]. We compared these techniques with the part annotations of PIN++. We used the largest version of SAM, and for open-vocabulary segmentation, we employed the VLPart model [58], which is trained with part annotations from several part datasets including Pascal-Part [6], PIN [21], and PACO [49]. Visual comparisons are presented in Fig. 2. We can see that VLPart only provides roughly reasonable semantics for part categories it has been trained on. It fails to accurately identify the semantics of parts from unseen objects. On the other hand, SAM seems to perform well in segmenting certain parts based on the object's edges, but it struggles when the part edges are less distinguishable, as shown in the segmentation results for camel (the third row in Fig. 2). In addition, note that all masks segmented by SAM are category-agnostic." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.72, + 0.784, + 0.84 + ], + "angle": 0, + "content": "The above results highlight the challenges faced by existing techniques in achieving accurate part segmentation for diverse objects without training on PIN++. Additionally, current text-to-image generative models [52] also struggle with understanding part semantics [42]. However, supervised training on PIN++ yields good part segmentation results, as shown in Appendix E. Therefore, besides its usage in robust object recognition (as demonstrated next), we believe that PIN++ and its extensive collections of 3308 part categories (vocabularies) have the potential to enhance general part-related visual understanding tasks." + } + ], + [ + { + "type": "page_number", + "bbox": [ + 0.218, + 0.116, + 0.23, + 0.127 + ], + "angle": 0, + "content": "8" + }, + { + "type": "header", + "bbox": [ + 0.272, + 0.115, + 0.327, + 0.127 + ], + "angle": 0, + "content": "Li et al." + }, + { + "type": "image", + "bbox": [ + 0.26, + 0.147, + 0.744, + 0.363 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.214, + 0.375, + 0.788, + 0.417 + ], + "angle": 0, + "content": "Fig. 2: Comparison between part segmentation results of different methods and PIN++ annotations. Without training on PIN++, VLPart and SAM fail to segment objects into specific parts with accurate semantics." + }, + { + "type": "title", + "bbox": [ + 0.216, + 0.445, + 0.562, + 0.463 + ], + "angle": 0, + "content": "4 Part-based Recognition Method" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.476, + 0.788, + 0.553 + ], + "angle": 0, + "content": "With the annotations of \\(\\mathrm{PIN}++\\), we develop part-based methods directly on the standard IN-1K. Our part-based methods can be divided into two steps. We first obtain pseudo part labels for the remaining unlabeled images of IN-1K and then train the MPM with these part annotations and pseudo labels. The pipeline of our part-based methods is illustrated in Fig. 3." + }, + { + "type": "title", + "bbox": [ + 0.216, + 0.574, + 0.515, + 0.589 + ], + "angle": 0, + "content": "4.1 Pseudo Part Label Generation" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.598, + 0.787, + 0.719 + ], + "angle": 0, + "content": "Part segmentation model. Arbitrary instance segmentation models [13, 22] can be leveraged to perform part segmentation and generate pseudo part labels when trained with the part annotations from \\(\\mathrm{PIN}++\\). Following Sun et al. [58], we use the classical Mask R-CNN [22] model with a vision transformer Swin-B [44] as the backbone network. In training the Mask R-CNN, we consider the masks for each part category as masks for individual object categories directly. During inference for obtaining pseudo part labels, we incorporate a simple post-processing operation alongside the regular inference procedure." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.72, + 0.788, + 0.841 + ], + "angle": 0, + "content": "Post-processing. Given an image \\(\\mathbf{x} \\in \\mathbb{R}^{3 \\times H \\times W}\\) with object category \\(y_{c}\\), where \\(H \\times W\\) represents the resolution, the regular inference procedure of the trained Mask R-CNN produces pseudo part labels in the form of \\(\\{(\\mathbf{M}_{\\mathbf{p}}, \\mathbf{v}_{\\mathbf{p}})\\}\\), where \\(\\mathbf{M}_{\\mathbf{p}} \\in \\{0,1\\}^{H \\times W}\\) denotes the binary part mask, \\(\\mathbf{v}_{\\mathbf{p}} \\in [0,1]^{K}\\) denotes the output probability of part categories, and \\(K\\) denotes the number of part categories (3308 in our case). But note that the images that need inference are all in the training set of IN-1K. Their object category \\(y_{c}\\) is not utilized in the regular inference procedure. To leverage it, we apply a Category Filter (CF) operation" + } + ], + [ + { + "type": "header", + "bbox": [ + 0.384, + 0.115, + 0.731, + 0.129 + ], + "angle": 0, + "content": "PartImageNet++ Dataset: Scaling up Part-based ..." + }, + { + "type": "page_number", + "bbox": [ + 0.775, + 0.117, + 0.785, + 0.127 + ], + "angle": 0, + "content": "9" + }, + { + "type": "image", + "bbox": [ + 0.218, + 0.148, + 0.486, + 0.321 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.507, + 0.149, + 0.785, + 0.32 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.214, + 0.331, + 0.788, + 0.415 + ], + "angle": 0, + "content": "Fig. 3: An overview of the generation of pseudo-labels and the structure of MPM. (a) A part segmentation model trained on \\(\\mathrm{PIN}++\\) and used to obtain pseudo-part labels for unannotated images. (b) MPM adds several auxiliary bypass layers to the vanilla recognition model for part segmentation supervision. MPM is trained by part annotations together with the pseudo part labels. During inference, the auxiliary layers are dropped, and the vanilla recognition model gives the final object category prediction." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.445, + 0.788, + 0.507 + ], + "angle": 0, + "content": "to obtain \\(\\hat{\\mathbf{v}}_{\\mathbf{p}}\\), which sets the probabilities of unrelated part categories in \\(\\mathbf{v}_{\\mathbf{p}}\\) to zero (e.g., for cat, except for cat:head, cat:body, etc, other part categories such as radio:button and bird:head are all excluded). The pseudo part category for \\(\\mathbf{M}_{\\mathbf{p}}\\) is then computed as \\(y_{p} = \\arg \\max \\hat{\\mathbf{v}}_{\\mathbf{p}}\\)." + }, + { + "type": "title", + "bbox": [ + 0.215, + 0.527, + 0.652, + 0.543 + ], + "angle": 0, + "content": "4.2 Multi-scale Part-supervised Recognition Model" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.552, + 0.788, + 0.657 + ], + "angle": 0, + "content": "Data preparation. After obtaining the pseudo labels for all remaining training images of IN-1K, we simply treat the pseudo labels and real part annotations equally because we find that the quality of the pseudo labels is roughly satisfactory, as shown in Appendix E. The part labels \\(\\{(\\mathbf{M_p},\\mathbf{y_p})\\}\\) for an image \\(\\mathbf{x}\\) are then converted to a single composite segmentation mask \\(\\mathbf{M} \\in \\{0,1\\}^{H \\times W \\times (K + 1)}\\) where the extra one channel indicates the background. Finally, the data samples for training part-based models are \\(\\{(\\mathbf{x},y_c,\\mathbf{M})\\}\\)." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.658, + 0.789, + 0.841 + ], + "angle": 0, + "content": "Model design. Both previous works [35, 56] build part-based models in the two-stage way: a segmenter \\(\\mathcal{F}_{\\mathrm{seg}}: \\mathbb{R}^{3 \\times H \\times W} \\to \\mathbb{R}^{(K + 1) \\times h \\times w}\\) and an extra module \\(\\mathcal{F}_{\\mathrm{cls}}: \\mathbb{R}^{(K + 1) \\times h \\times w} \\to \\mathbb{R}^C\\) for classification, where \\(h \\times w\\) denotes the output resolution and \\(C\\) denotes the number of object category. \\(\\mathcal{F}_{\\mathrm{seg}}\\) is a backbone network (e.g., a ResNet-50 without classification head). Most popular backbones [23, 44, 45] are hierarchically composed of several blocks with down-sampling layers and thus they can be written as \\(\\mathcal{F}_{\\mathrm{seg}} := f_l \\circ \\dots \\circ f_1\\), where \\(l\\) denotes the number of down-sampling layers and \\(f_{i(1 \\leq i \\leq l)}\\) denotes one block with the \\(i\\)th down-sampling layer. \\(\\mathcal{F}_{\\mathrm{cls}}\\) can be a tiny classifier [56] or a nondifferentiable judgment block [35] and its input is the part segmentation results \\(\\hat{\\mathbf{M}} = \\mathcal{F}_{\\mathrm{seg}}(\\mathbf{x})\\). The overall model for recognition is \\(\\mathcal{F} := \\mathcal{F}_{\\mathrm{cls}} \\circ \\mathcal{F}_{\\mathrm{seg}}\\). However, as \\(K\\) is significantly larger than \\(C\\), using \\(\\hat{M}\\) as the intermediate results for" + } + ], + [ + { + "type": "page_number", + "bbox": [ + 0.218, + 0.116, + 0.236, + 0.127 + ], + "angle": 0, + "content": "10" + }, + { + "type": "header", + "bbox": [ + 0.272, + 0.115, + 0.326, + 0.127 + ], + "angle": 0, + "content": "Li et al." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.147, + 0.784, + 0.236 + ], + "angle": 0, + "content": "recognition inevitably introduce extra parameters and computation during inference. This situation gets worse with an increased number of part categories, as shown in Sec. 5.5. Moreover, these methods solely employ part supervision on the output of \\( f_{l} \\). Without expanding \\( \\mathcal{F}_{\\mathrm{seg}} \\), part annotations have to be downsampled to match the output resolution of \\( f_{l} \\), potentially limiting the utilization of high-resolution part annotations." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.238, + 0.788, + 0.6 + ], + "angle": 0, + "content": "Instead of these two-stage part-based models, MPM is expected to learn a robust intermediate representation by adding multi-scale bypass layers to the vanilla recognition network. In MPM, the model used for object recognition is the vanilla backbone network directly: \\(\\mathcal{F} \\coloneqq h_c \\circ f_l \\circ \\dots \\circ f_1\\), where \\(h_c\\) denotes a vanilla classification head (usually a linear layer). MPM utilizes the part annotations to supervise the intermediate features of \\(\\mathcal{F}\\) by several bypass layers \\(g_i\\): \\(\\tilde{\\mathbf{M}}_i = g_i \\circ f_i \\circ \\dots \\circ f_1(\\mathbf{x})\\), where \\(i \\leq l\\). In this way, the intermediate features of \\(\\mathcal{F}\\) can be seen as implicit representations of part segmentation results when \\(g_i\\) is simple enough. Note that here we use part supervision not only on the output of \\(f_l\\), but also \\(f_{i(i < l)}\\). As the output of lower layers has a larger resolution, it can be supervised by part annotations with higher resolution. But the outputs of \\(f_{i(i < l)}\\) are relatively low-level, which may not be enough to obtain high-level part segmentation results themselves. To balance this, only the outputs of the last three blocks \\(f_{i(l - 2 \\leq i \\leq l)}\\) are supervised by part mask \\(\\mathbf{M}\\) (e.g., for a ResNet-50 with input size \\(224 \\times 224\\), the intermediate features of \\(7 \\times 7\\), \\(14 \\times 14\\), and \\(28 \\times 28\\) are supervised by corresponding down-sampled \\(\\mathbf{M}\\), respectively). In addition, following the principle of FPN [40], some top-down layers are used to augment the low-level features with high-level features. Note that distinct from the conventional FPN that generally has comparable parameters with the backbone network [40], these bypass layers are extremely lightweight as our goal is to improve the recognition accuracy rather than the quality of part segmentation results. Fig. 3(b) shows the overall structure of MPM. During inference, these auxiliary layers are dropped, and the vanilla recognition model gives the final object category prediction." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.6, + 0.788, + 0.695 + ], + "angle": 0, + "content": "Training objective. The overall loss for training MPM is: \\( L = L_{\\mathrm{cls}} + \\lambda \\cdot L_{\\mathrm{seg}} \\), where \\( L_{\\mathrm{cls}} \\) denotes the vanilla loss for classification, \\( L_{\\mathrm{seg}} \\) denotes the loss for part segmentation, and \\( \\lambda \\) is a hyper-parameter. We compute \\( L_{\\mathrm{seg}} \\) as the average of losses on three part segmentation results of different resolutions, as mentioned before. Note that when performing AT, adversarial examples are generated by \\( \\mathbf{x}^{\\star} = \\arg \\max_{\\mathbf{x}^{\\star}:||\\mathbf{x}^{\\star} - \\mathbf{x}||_{p}\\leq \\epsilon}L_{\\mathrm{cls}}(\\mathcal{F}(\\mathbf{x}^{\\star}),y) \\), while MPM is always trained with \\( L \\)." + }, + { + "type": "title", + "bbox": [ + 0.216, + 0.712, + 0.376, + 0.73 + ], + "angle": 0, + "content": "5 Experiments" + }, + { + "type": "title", + "bbox": [ + 0.216, + 0.742, + 0.428, + 0.759 + ], + "angle": 0, + "content": "5.1 Experimental Setup" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.765, + 0.788, + 0.841 + ], + "angle": 0, + "content": "Training setup. Unless specified otherwise, following previous works [3,10], we performed AT with \\( l_{\\infty} \\) bound \\( \\epsilon = 4 / 255 \\) on IN-1K. The inner optimization for obtaining adversarial examples used PGD [46] with iterative steps \\( t = 2 \\) and the step size was set to be \\( s = 2*\\epsilon /t \\). The input resolution used \\( 224\\times 224 \\). We used the vanilla version of the AT [46] and did not incorporate other variants" + } + ], + [ + { + "type": "header", + "bbox": [ + 0.384, + 0.115, + 0.732, + 0.131 + ], + "angle": 0, + "content": "PartImageNet++ Dataset: Scaling up Part-based ..." + }, + { + "type": "page_number", + "bbox": [ + 0.768, + 0.116, + 0.784, + 0.127 + ], + "angle": 0, + "content": "11" + }, + { + "type": "table_caption", + "bbox": [ + 0.216, + 0.145, + 0.788, + 0.189 + ], + "angle": 0, + "content": "Table 2: Recognition accuracies (\\%) of methods under different attack threats on IN-1K, following the evaluation setup in Sec. 5.1. Average denotes the average accuracies in four attack settings. We highlight the best results in each column." + }, + { + "type": "table", + "bbox": [ + 0.236, + 0.199, + 0.758, + 0.336 + ], + "angle": 0, + "content": "
ArchitectureMethodAdversarial Train wrt l∞ (ε = 4/255)
Cleanl∞l∞(×2)l1l2Average
ResNet-50[54], 202063.935.913.22.013.716.2
ResNet-50[47], 202264.734.311.42.316.416.1
ResNet-50[41], 202365.532.69.53.317.915.8
ResNet-50-Gelu[3], 202167.836.610.83.016.816.8
ResNet-50-Gelu[10], 202366.835.612.83.616.517.1
ResNet-50-Geluours (vanilla)67.138.112.65.021.619.3
ResNet-50-Geluours (MPM)67.839.113.66.224.320.8
" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.359, + 0.788, + 0.526 + ], + "angle": 0, + "content": "such as TRADES [65] or AWP [64]. These variants have shown effectiveness mainly on small-scale datasets, making their generalization to IN-1K non-trivial. Following previous part-based works [35,56], we used the classical ResNet-50 [23] network with 25.6M parameters as the baseline model. We note that recent works [3,10] used a GELU [25] activation to replace the original RELU [1] activation of ResNet-50 for boosting robustness when performing AT, here we also used GELU, denoted as ResNet-50-Gelu. In addition, we improved the training recipe of ResNet-50 to build a strong baseline. Other details on the training recipe are provided in Appendix F. Note that bypass layers (e.g., point-wise convolutions) with about 4.5M parameters were used by MPM during training but completely dropped during inference. The code is submitted along with the paper." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.526, + 0.788, + 0.675 + ], + "angle": 0, + "content": "Evaluation setup. For all experiments in the work, unless specified otherwise, we evaluated the adversarial robustness with AutoAttack [8], which is a combination of various attack methods and is generally recognized as a reliable evaluation [7]. The evaluation was performed at a resolution \\(224 \\times 224\\) on randomly selected 10K images of the 50K IN-1K val set, 10 images for each category. This ensures that the \\(95\\%\\) confidence intervals of the reported average adversarial robustness (see Tab. 2) were less than \\(\\pm 0.5\\%\\). Note that, unlike previous works [3,10], we did not use the 5K images selected by RobustBench [7] as we found that they were category-unbalanced, containing three categories even without any image for validation." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.677, + 0.788, + 0.769 + ], + "angle": 0, + "content": "We expect the AT models with \\( l_{\\infty} \\) bound \\( \\epsilon = 4 / 255 \\) to be adversarially robust not only on seen threats (i.e., \\( l_{\\infty} \\) attack with \\( \\epsilon = 4 / 255 \\)) but also on unseen attack threats (e.g., \\( l_{\\infty} \\) attack with larger \\( \\epsilon, l_1 \\) and \\( l_2 \\) attacks). Thus, we evaluated the models on three attack threats: \\( l_{\\infty} \\), \\( l_1 \\), and \\( l_2 \\), with the bounds \\( \\epsilon_{\\infty} = 4 / 255 \\), \\( \\epsilon_1 = 75 \\), \\( \\epsilon_2 = 2 \\), respectively. For \\( l_{\\infty} \\) attack, we extra evaluated the models with \\( \\epsilon_{\\infty} = 8 / 255 \\) to mimic a stronger threat, denoted as \\( l_{\\infty}(\\times 2) \\)." + }, + { + "type": "title", + "bbox": [ + 0.216, + 0.787, + 0.589, + 0.803 + ], + "angle": 0, + "content": "5.2 Robustness against Adversarial Attacks" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.81, + 0.787, + 0.842 + ], + "angle": 0, + "content": "With the above setup, we trained the vanilla ResNet-50-Gelu as a baseline. MPM used the same recipe as the baseline model for a fair comparison, except that it" + } + ], + [ + { + "type": "page_number", + "bbox": [ + 0.218, + 0.116, + 0.236, + 0.127 + ], + "angle": 0, + "content": "12" + }, + { + "type": "header", + "bbox": [ + 0.272, + 0.115, + 0.326, + 0.127 + ], + "angle": 0, + "content": "Li et al." + }, + { + "type": "table_caption", + "bbox": [ + 0.216, + 0.145, + 0.788, + 0.201 + ], + "angle": 0, + "content": "Table 3: Recognition accuracies \\((\\%)\\) of ResNet-50-Gelu on 15 different common image corruptions [24]. We adopt accuracy as the metric to be consistent with other results, while this metric can be easily converted into the corruption error metric [24]. \"Average\" denotes the accuracy averaged over different common corruptions." + }, + { + "type": "table", + "bbox": [ + 0.229, + 0.213, + 0.782, + 0.29 + ], + "angle": 0, + "content": "
Part ATNoiseBlurWeatherDigitalAvg.
Gauss. Shot Impul.Defoc.Glass Moti.ZoomSnow FrostFogBrightCont.Elast.Pixel JPEG
31.230.727.339.028.639.537.835.840.753.869.038.037.047.157.040.8
✓31.531.028.441.929.840.336.737.041.355.369.639.636.952.959.942.1
✓29.628.521.023.332.932.834.434.234.010.858.49.552.258.863.134.9
✓30.129.222.425.235.135.036.436.436.211.959.710.453.159.864.336.4
" + }, + { + "type": "table_caption", + "bbox": [ + 0.265, + 0.304, + 0.737, + 0.318 + ], + "angle": 0, + "content": "Table 4: Recognition accuracies (%) of models on four OOD datasets." + }, + { + "type": "table", + "bbox": [ + 0.34, + 0.33, + 0.664, + 0.396 + ], + "angle": 0, + "content": "
PartATIN-A-PlusIN-SketchSINDINAverage
6.925.86.953.523.3
✓7.425.77.453.923.6
✓5.325.112.555.424.6
✓✓5.726.412.655.925.2
" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.426, + 0.788, + 0.592 + ], + "angle": 0, + "content": "used an extra \\( L_{\\mathrm{seg}} \\) to introduce part-based supervision. \\( \\lambda \\) was simply set to 1. In addition, we compared MPM with all open-sourced ResNet-50 checkpoints adversarially trained on IN-1K in recent years. The recognition accuracies under different attack threats are shown in Tab. 2. We can see that our baseline model is strong enough, outperforming all other benchmarks by \\( 2.2\\% \\) on average robustness (17.1% vs. 19.3%). MPM further shows significantly better adversarial robustness than the vanilla baseline across all these attack threats. In addition, compared with our strong vanilla baseline, the accuracy of clean images is also improved with MPM. To our best knowledge, these are SOTA results of adversarial robustness with ResNet-50 on IN-1K. We additionally show in Appendix G that these results are significantly better than previous results on PIN." + }, + { + "type": "title", + "bbox": [ + 0.216, + 0.616, + 0.649, + 0.632 + ], + "angle": 0, + "content": "5.3 Robustness on Corruptions and OOD Datasets" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.644, + 0.788, + 0.735 + ], + "angle": 0, + "content": "Except for robustness against adversarial attacks, which consider the worst-case security threat scenarios, we evaluated the robustness of our part-based model on common image corruptions [24] and several OOD datasets [15, 16, 34, 62], which could be more realistic threats. Considering that AT can have negative effects on these non-adversarial scenarios [35, 41], here we also evaluated models without AT. The training details are also shown in Appendix F." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.735, + 0.788, + 0.841 + ], + "angle": 0, + "content": "For common image corruptions [24], we generated different types of corrupted images on the whole IN-1K val set. The results of different models on these images are shown in Tab. 3, where \"Part\" indicates MPM. Here the results of each corruption type are the accuracies averaged on five severity levels [24]. It is seen that MPM has better robustness on nearly all types of image corruptions than models without part supervision, regardless of AT. Besides common corruptions, we evaluated the models on four OOD datasets: ImageNet-A-Plus (IN-A-Plus)" + } + ], + [ + { + "type": "header", + "bbox": [ + 0.384, + 0.115, + 0.732, + 0.131 + ], + "angle": 0, + "content": "PartImageNet++ Dataset: Scaling up Part-based ..." + }, + { + "type": "page_number", + "bbox": [ + 0.768, + 0.116, + 0.786, + 0.128 + ], + "angle": 0, + "content": "13" + }, + { + "type": "table_caption", + "bbox": [ + 0.215, + 0.145, + 0.788, + 0.174 + ], + "angle": 0, + "content": "Table 5: Segmentation accuracies (AP) of Mask R-CNN with and without CF, and recognition accuracies (\\%) of MPM on IN-1K trained with various pseudo labels." + }, + { + "type": "table", + "bbox": [ + 0.304, + 0.185, + 0.691, + 0.252 + ], + "angle": 0, + "content": "
Pseudo LabelCFAP\\( AP_{50} \\)Clean\\( l_{\\infty} \\)\\( l_{\\infty}(\\times 2) \\)
--65.533.09.4
✓37.258.667.838.813.1
✓✓40.764.467.839.113.6
" + }, + { + "type": "text", + "bbox": [ + 0.215, + 0.278, + 0.788, + 0.34 + ], + "angle": 0, + "content": "[34], ImageNet Sketch (IN-Sketch) [62], Stylized ImageNet (SIN) [16], and image distortion dataset (denoted as DIN) [15]. See Appendix H for the introduction of these datasets. The results shown in Tab. 4 indicate that MPM also exhibits improved robustness on these OOD shifts." + }, + { + "type": "title", + "bbox": [ + 0.215, + 0.361, + 0.489, + 0.377 + ], + "angle": 0, + "content": "5.4 Other Advantages of MPM" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.386, + 0.788, + 0.492 + ], + "angle": 0, + "content": "Models versus humans. The part-based models are motivated by the human recognition process. We were interested in whether the robustness of MPM aligned more with human cognition. To investigate this, we employed the evaluation method introduced by Geirhos et al. [15], which involved comparing the models' decisions on different distorted images with the actual judgments made by human observers. The results shown in Appendix I indicate that MPM measurably improved alignment with the human recognition process." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.492, + 0.788, + 0.6 + ], + "angle": 0, + "content": "Boosting the downstream task. Li et al. [33] show that adversarially trained models on the large-scale IN-1K can be utilized to initialize the backbone of downstream networks (e.g., Faster R-CNN [51]). This approach together with downstream AT enables the transfer of adversarial robustness to downstream tasks. We conducted similar experiments by using the checkpoints of the vanilla baseline and MPM. The results shown in Appendix J indicate that MPM enhances both the clean accuracy and adversarial robustness in object detection." + }, + { + "type": "title", + "bbox": [ + 0.215, + 0.62, + 0.391, + 0.636 + ], + "angle": 0, + "content": "5.5 Ablation Study" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.644, + 0.788, + 0.78 + ], + "angle": 0, + "content": "Pseudo-label quality. We first investigate the influence of pseudo-label quality on the adversarial robustness of MPM. In Sec. 4.1, we illustrate the use of CF during generating pseudo labels. Here we remove the CF operator and regenerate pseudo labels for images without part annotations. The left half of Tab. 5 (the 2nd and 3rd rows) shows the segmentation accuracies of models with and without CF (here the models were supervised with the train set of \\(\\mathrm{PIN}++\\), and the AP were calculated on val). It is seen that training with \\(\\mathrm{PIN}++\\) achieved good part segmentation results, indicating its potential for enhancing part-related tasks. CF further significantly improves AP." + }, + { + "type": "text", + "bbox": [ + 0.215, + 0.78, + 0.788, + 0.842 + ], + "angle": 0, + "content": "We then trained MPM using both pseudo labels and real part annotations, as well as solely using real part annotations. The results, displayed in Tab. 5 (the 1st row), indicate that training MPM solely with real part annotations seems to be ineffective, potentially compromising the model's performance. We guess" + } + ], + [ + { + "type": "page_number", + "bbox": [ + 0.218, + 0.116, + 0.236, + 0.127 + ], + "angle": 0, + "content": "14" + }, + { + "type": "header", + "bbox": [ + 0.272, + 0.115, + 0.328, + 0.128 + ], + "angle": 0, + "content": "Li et al." + }, + { + "type": "table_caption", + "bbox": [ + 0.216, + 0.145, + 0.788, + 0.174 + ], + "angle": 0, + "content": "Table 6: Recognition accuracies (\\%) of different part-based models on IN-1K. The number of parameters during inference is listed." + }, + { + "type": "table", + "bbox": [ + 0.314, + 0.185, + 0.681, + 0.279 + ], + "angle": 0, + "content": "
MethodParam.Cleanl∞l∞(×2)
Two-stage60.9M67.338.713.0
SV167.336.98.8
SV2 w/ TD25.6M67.637.111.7
MPM (SV3 w/ TD)67.839.113.6
SV3 w/o TD67.436.911.7
" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.306, + 0.784, + 0.351 + ], + "angle": 0, + "content": "it could caused by overfitting on the images with part annotations. Besides, the inclusion of better-quality pseudo labels can improve the robustness of our part-based model, underscoring the importance of high-quality pseudo labels." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.351, + 0.785, + 0.531 + ], + "angle": 0, + "content": "Part-based model structure. We then conducted ablation experiments on the structure of the part-based models. Specifically, we compared MPM with the two-stage part-based model designed by Sitawarin et al. [56]. In addition, we performed ablations on the number of part supervisions. MPM incorporates part supervision with three resolutions, referred to as SV3. Here we compared it with SV1 and SV2, which indicate part-based models solely supervised by \\( f_{l} \\) and both \\( f_{l} \\) and \\( f_{l-1} \\), respectively. Besides these, we extra performed an ablation by removing all top-down (TD) connections of MPM. The results are shown in Tab. 6. We can see that although the two-stage model needs about twice the parameters during inference, its robustness is inferior to the MPM. And all the designs of MPM, including utilizing higher resolution part annotations and top-down connections, contribute to the robustness." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.532, + 0.787, + 0.622 + ], + "angle": 0, + "content": "Additional ablations. Additional ablation experiments emphasize the significance of fine-grained part annotations rather than object segmentation annotations. Experiments on Tiny-ImageNet [31] highlight the significance of investigating adversarial robustness on high-resolution images. Moreover, the adversarial robustness of part-based models exhibits low sensitivity to \\(\\lambda\\). Refer to Appendix K for comprehensive information about these ablation studies." + }, + { + "type": "title", + "bbox": [ + 0.216, + 0.644, + 0.509, + 0.66 + ], + "angle": 0, + "content": "6 Conclusion and Discussion" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.674, + 0.785, + 0.78 + ], + "angle": 0, + "content": "In this work, we introduce PIN++, a new large-scale high-quality part dataset, to facilitate the research on part-based models for robust recognition. We propose a new MPM to leverage this dataset. With extensive experiments, we demonstrate the robustness of MPM across diverse scenarios. But we note that MPM represents just a stepping stone in the realm of part-based models for robustness. We hope that this work will inspire and motivate researchers to further explore the potential of part-based models for robust recognition." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.78, + 0.787, + 0.84 + ], + "angle": 0, + "content": "Limitation. We believe that except for robust recognition, \\(\\mathrm{PIN}++\\) has the potential to boost general part-related visual understanding tasks. We have not conducted many benchmark experiments specifically for such tasks because they are beyond the scope of this work. We leave these as future work." + } + ], + [ + { + "type": "header", + "bbox": [ + 0.384, + 0.115, + 0.732, + 0.129 + ], + "angle": 0, + "content": "PartImageNet++ Dataset: Scaling up Part-based ..." + }, + { + "type": "page_number", + "bbox": [ + 0.768, + 0.117, + 0.786, + 0.127 + ], + "angle": 0, + "content": "15" + }, + { + "type": "title", + "bbox": [ + 0.218, + 0.145, + 0.395, + 0.163 + ], + "angle": 0, + "content": "Acknowledgement" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.177, + 0.788, + 0.222 + ], + "angle": 0, + "content": "This work was supported in part by the National Key Research and Development Program of China (No. 2021ZD0200301) and the National Natural Science Foundation of China (No. U2341228)." + }, + { + "type": "title", + "bbox": [ + 0.218, + 0.244, + 0.323, + 0.26 + ], + "angle": 0, + "content": "References" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.274, + 0.785, + 0.302 + ], + "angle": 0, + "content": "1. Agarap, A.F.: Deep learning using rectified linear units (relu). arXiv preprint arXiv:1803.08375 (2018)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.304, + 0.785, + 0.344 + ], + "angle": 0, + "content": "2. Athalye, A., Carlini, N., Wagner, D.A.: Obfuscated gradients give a false sense of security: Circumventing defenses to adversarial examples. In: Int. Conf. Mach. Learn. (ICML). vol. 80, pp. 274-283 (2018)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.346, + 0.785, + 0.371 + ], + "angle": 0, + "content": "3. Bai, Y., Mei, J., Yuille, A.L., Xie, C.: Are transformers more robust than cnns? In: Adv. Neural Inform. Process. Syst. (NeurIPS). pp. 26831-26843 (2021)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.373, + 0.785, + 0.399 + ], + "angle": 0, + "content": "4. Biederman, I.: Recognition-by-components: a theory of human image understanding. Psychological Review 94(2), 115 (1987)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.401, + 0.785, + 0.44 + ], + "angle": 0, + "content": "5. Burl, M.C., Weber, M., Perona, P.: A probabilistic approach to object recognition using local photometry and global geometry. In: Eur. Conf. Comput. Vis. (ECCV). vol. 1407, pp. 628-641 (1998)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.442, + 0.785, + 0.482 + ], + "angle": 0, + "content": "6. Chen, X., Mottaghi, R., Liu, X., Fidler, S., Urtasun, R., Yuille, A.L.: Detect what you can: Detecting and representing objects using holistic models and body parts. In: IEEE Conf. Comput. Vis. Pattern Recog. (CVPR). pp. 1979-1986 (2014)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.484, + 0.785, + 0.523 + ], + "angle": 0, + "content": "7. Croce, F., Andriushchenko, M., Sehwag, V., Debenedetti, E., Flammarion, N., Chiang, M., Mittal, P., Hein, M.: Robustbench: a standardized adversarial robustness benchmark. In: NeurIPS Datasets and Benchmarks (2021)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.525, + 0.785, + 0.564 + ], + "angle": 0, + "content": "8. Croce, F., Hein, M.: Reliable evaluation of adversarial robustness with an ensemble of diverse parameter-free attacks. In: Int. Conf. Mach. Learn. (ICML). Proceedings of Machine Learning Research, vol. 119, pp. 2206-2216 (2020)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.565, + 0.785, + 0.605 + ], + "angle": 0, + "content": "9. Croce, F., Singh, N.D., Hein, M.: Robust semantic segmentation: Strong adversarial attacks and fast training of robust models. Adv. Neural Inform. Process. Syst. (NeurIPS) (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.22, + 0.607, + 0.785, + 0.646 + ], + "angle": 0, + "content": "10. Debenedetti, E., Sehwag, V., Mittal, P.: A light recipe to train robust vision transformers. In: First IEEE Conference on Secure and Trustworthy Machine Learning (SaTML). pp. 225-253 (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.219, + 0.649, + 0.785, + 0.688 + ], + "angle": 0, + "content": "11. Deng, J., Dong, W., Socher, R., Li, L., Li, K., Fei-Fei, L.: Imagenet: A large-scale hierarchical image database. In: IEEE Conf. Comput. Vis. Pattern Recog. (CVPR). pp. 248-255 (2009)" + }, + { + "type": "ref_text", + "bbox": [ + 0.219, + 0.689, + 0.785, + 0.743 + ], + "angle": 0, + "content": "12. Eykholt, K., Evtimov, I., Fernandes, E., Li, B., Rahmati, A., Xiao, C., Prakash, A., Kohno, T., Song, D.: Robust physical-world attacks on deep learning visual classification. In: IEEE Conf. Comput. Vis. Pattern Recog. (CVPR). pp. 1625-1634 (2018)" + }, + { + "type": "ref_text", + "bbox": [ + 0.219, + 0.745, + 0.785, + 0.771 + ], + "angle": 0, + "content": "13. Fang, Y., Yang, S., Wang, X., Li, Y., Fang, C., Shan, Y., Feng, B., Liu, W.: Instances as queries. In: Int. Conf. Comput. Vis. (ICCV). pp. 6890-6899 (2021)" + }, + { + "type": "ref_text", + "bbox": [ + 0.219, + 0.772, + 0.785, + 0.799 + ], + "angle": 0, + "content": "14. Fischler, M.A., Elschlager, R.A.: The representation and matching of pictorial structures. IEEE Trans. Comput. 22(1), 67-92 (1973)" + }, + { + "type": "ref_text", + "bbox": [ + 0.219, + 0.8, + 0.785, + 0.84 + ], + "angle": 0, + "content": "15. Geirhos, R., Narayanappa, K., Mitzkus, B., Thieringer, T., Bethge, M., Wichmann, F.A., Brendel, W.: Partial success in closing the gap between human and machine vision. In: Adv. Neural Inform. Process. Syst. (NeurIPS). pp. 23885-23899 (2021)" + }, + { + "type": "list", + "bbox": [ + 0.219, + 0.274, + 0.785, + 0.84 + ], + "angle": 0, + "content": null + } + ], + [ + { + "type": "page_number", + "bbox": [ + 0.218, + 0.116, + 0.236, + 0.127 + ], + "angle": 0, + "content": "16" + }, + { + "type": "header", + "bbox": [ + 0.272, + 0.115, + 0.326, + 0.127 + ], + "angle": 0, + "content": "Li et al." + }, + { + "type": "ref_text", + "bbox": [ + 0.218, + 0.147, + 0.785, + 0.189 + ], + "angle": 0, + "content": "16. Geirhos, R., Rubisch, P., Michaelis, C., Bethge, M., Wichmann, F.A., Brendel, W.: Imagenet-trained cnns are biased towards texture; increasing shape bias improves accuracy and robustness. In: Int. Conf. Learn. Rep. (ICLR) (2019)" + }, + { + "type": "ref_text", + "bbox": [ + 0.218, + 0.189, + 0.785, + 0.23 + ], + "angle": 0, + "content": "17. Girshick, R.B., Iandola, F.N., Darrell, T., Malik, J.: Deformable part models are convolutional neural networks. In: IEEE Conf. Comput. Vis. Pattern Recog. (CVPR). pp. 437-446 (2015)" + }, + { + "type": "ref_text", + "bbox": [ + 0.218, + 0.23, + 0.785, + 0.271 + ], + "angle": 0, + "content": "18. Gong, K., Liang, X., Li, Y., Chen, Y., Yang, M., Lin, L.: Instance-level human parsing via part grouping network. In: Eur. Conf. Comput. Vis. (ECCV). vol. 11208, pp. 805-822 (2018)" + }, + { + "type": "ref_text", + "bbox": [ + 0.218, + 0.271, + 0.785, + 0.298 + ], + "angle": 0, + "content": "19. Goodfellow, I.J., Shlens, J., Szegedy, C.: Explaining and harnessing adversarial examples. In: Int. Conf. Learn. Rep. (ICLR) (2015)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.298, + 0.785, + 0.352 + ], + "angle": 0, + "content": "20. Haaf, R.A., Fulkerson, A.L., Jablonski, B.J., Hupp, J.M., Shull, S.S., Pescara-Kovach, L.: Object recognition and attention to object components by preschool children and 4-month-old infants. Journal of Experimental Child Psychology 86(2), 108-123 (2003)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.352, + 0.785, + 0.393 + ], + "angle": 0, + "content": "21. He, J., Yang, S., Yang, S., Kortylewski, A., Yuan, X., Chen, J., Liu, S., Yang, C., Yu, Q., Yuille, A.L.: Partimagenet: A large, high-quality dataset of parts. In: Eur. Conf. Comput. Vis. (ECCV). vol. 13668, pp. 128-145 (2022)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.393, + 0.785, + 0.421 + ], + "angle": 0, + "content": "22. He, K., Gkioxari, G., Dollár, P., Girshick, R.B.: Mask R-CNN. IEEE Trans. Pattern Anal. Mach. Intell. (TPAMI) 42(2), 386-397 (2020)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.421, + 0.785, + 0.448 + ], + "angle": 0, + "content": "23. He, K., Zhang, X., Ren, S., Sun, J.: Deep residual learning for image recognition. In: IEEE Conf. Comput. Vis. Pattern Recog. (CVPR). pp. 770-778 (2016)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.448, + 0.785, + 0.474 + ], + "angle": 0, + "content": "24. Hendrycks, D., Dietterich, T.G.: Benchmarking neural network robustness to common corruptions and perturbations. In: Int. Conf. Learn. Rep. (ICLR) (2019)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.474, + 0.785, + 0.501 + ], + "angle": 0, + "content": "25. Hendrycks, D., Gimpel, K.: Gaussian error linear units (gelus). arXiv preprint arXiv:1606.08415 (2016)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.501, + 0.785, + 0.528 + ], + "angle": 0, + "content": "26. Hoffman, J.E., Subramaniam, B.: The role of visual attention in saccadic eye movements. Perception & Psychophysics 57(6), 787-795 (1995)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.528, + 0.785, + 0.569 + ], + "angle": 0, + "content": "27. Ilyas, A., Santurkar, S., Tsipras, D., Engstrom, L., Tran, B., Madry, A.: Adversarial examples are not bugs, they are features. In: Adv. Neural Inform. Process. Syst. (NeurIPS). pp. 125-136 (2019)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.569, + 0.785, + 0.61 + ], + "angle": 0, + "content": "28. Kirillov, A., Mintun, E., Ravi, N., Mao, H., Rolland, C., Gustafson, L., Xiao, T., Whitehead, S., Berg, A.C., Lo, W.Y., Dollar, P., Girshick, R.: Segment anything. In: Int. Conf. Comput. Vis. (ICCV). pp. 4015-4026 (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.61, + 0.785, + 0.637 + ], + "angle": 0, + "content": "29. Krizhevsky, A., Hinton, G., et al.: Learning multiple layers of features from tiny images (2009)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.637, + 0.785, + 0.677 + ], + "angle": 0, + "content": "30. Krizhevsky, A., Sutskever, I., Hinton, G.E.: Imagenet classification with deep convolutional neural networks. In: Adv. Neural Inform. Process. Syst. (NeurIPS). pp. 1106-1114 (2012)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.677, + 0.785, + 0.704 + ], + "angle": 0, + "content": "31. Le, Y., Yang, X.: Tiny imagenet visual recognition challenge. CS 231N 7(7), 3 (2015)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.704, + 0.785, + 0.746 + ], + "angle": 0, + "content": "32. Li, L.H., Zhang, P., Zhang, H., Yang, J., Li, C., Zhong, Y., Wang, L., Yuan, L., Zhang, L., Hwang, J., Chang, K., Gao, J.: Grounded language-image pre-training. In: IEEE Conf. Comput. Vis. Pattern Recog. (CVPR). pp. 10955-10965 (2022)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.746, + 0.785, + 0.772 + ], + "angle": 0, + "content": "33. Li, X., Chen, H., Hu, X.: On the importance of backbone to the adversarial robustness of object detectors. arXiv preprint arXiv:2305.17438 (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.772, + 0.785, + 0.8 + ], + "angle": 0, + "content": "34. Li, X., Li, J., Dai, T., Shi, J., Zhu, J., Hu, X.: Rethinking natural adversarial examples for classification models. arXiv preprint arXiv:2102.11731 (2021)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.8, + 0.785, + 0.84 + ], + "angle": 0, + "content": "35. Li, X., Wang, Z., Zhang, B., Sun, F., Hu, X.: Recognizing object by components with human prior knowledge enhances adversarial robustness of deep neural networks. IEEE Trans. Pattern Anal. Mach. Intell. (TPAMI) 45(7), 8861-8873 (2023)" + }, + { + "type": "list", + "bbox": [ + 0.217, + 0.147, + 0.785, + 0.84 + ], + "angle": 0, + "content": null + } + ], + [ + { + "type": "header", + "bbox": [ + 0.384, + 0.115, + 0.732, + 0.129 + ], + "angle": 0, + "content": "PartImageNet++ Dataset: Scaling up Part-based ..." + }, + { + "type": "page_number", + "bbox": [ + 0.768, + 0.116, + 0.786, + 0.127 + ], + "angle": 0, + "content": "17" + }, + { + "type": "ref_text", + "bbox": [ + 0.214, + 0.147, + 0.788, + 0.189 + ], + "angle": 0, + "content": "36. Li, X., Zhang, W., Liu, Y., Hu, Z., Zhang, B., Hu, X.: Language-driven anchors for zero-shot adversarial robustness. In: IEEE Conf. Comput. Vis. Pattern Recog. (CVPR) (2024)" + }, + { + "type": "ref_text", + "bbox": [ + 0.214, + 0.19, + 0.787, + 0.232 + ], + "angle": 0, + "content": "37. Liang, X., Gong, K., Shen, X., Lin, L.: Look into person: Joint body parsing & pose estimation network and a new benchmark. IEEE Trans. Pattern Anal. Mach. Intell. (TPAMI) 41(4), 871-885 (2019)" + }, + { + "type": "ref_text", + "bbox": [ + 0.214, + 0.232, + 0.787, + 0.273 + ], + "angle": 0, + "content": "38. Liang, X., Liu, S., Shen, X., Yang, J., Liu, L., Dong, J., Lin, L., Yan, S.: Deep human parsing with active template regression. IEEE Trans. Pattern Anal. Mach. Intell. (TPAMI) 37(12), 2402-2414 (2015)" + }, + { + "type": "ref_text", + "bbox": [ + 0.214, + 0.273, + 0.787, + 0.314 + ], + "angle": 0, + "content": "39. Lin, C., Sun, P., Jiang, Y., Luo, P., Qu, L., Haffari, G., Yuan, Z., Cai, J.: Learning object-language alignments for open-vocabulary object detection. In: Int. Conf. Learn. Rep. (ICLR) (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.214, + 0.315, + 0.787, + 0.357 + ], + "angle": 0, + "content": "40. Lin, T., Dollar, P., Girshick, R.B., He, K., Hariharan, B., Belongie, S.J.: Feature pyramid networks for object detection. In: IEEE Conf. Comput. Vis. Pattern Recog. (CVPR). pp. 936-944 (2017)" + }, + { + "type": "ref_text", + "bbox": [ + 0.214, + 0.357, + 0.787, + 0.398 + ], + "angle": 0, + "content": "41. Liu, C., Dong, Y., Xiang, W., Yang, X., Su, H., Zhu, J., Chen, Y., He, Y., Xue, H., Zheng, S.: A comprehensive study on robustness of image classification models: Benchmarking and rethinking. arXiv preprint arXiv:2302.14301 (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.214, + 0.398, + 0.787, + 0.424 + ], + "angle": 0, + "content": "42. Liu, Q., Kortylewski, A., Bai, Y., Bai, S., Yuille, A.L.: Intriguing properties of text-guided diffusion models. arXiv preprint arXiv:2306.00974 (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.214, + 0.425, + 0.787, + 0.467 + ], + "angle": 0, + "content": "43. Liu, S., Zhang, L., Yang, X., Su, H., Zhu, J.: Unsupervised part segmentation through disentangling appearance and shape. In: IEEE Conf. Comput. Vis. Pattern Recog. (CVPR). pp. 8355-8364 (2021)" + }, + { + "type": "ref_text", + "bbox": [ + 0.214, + 0.467, + 0.787, + 0.509 + ], + "angle": 0, + "content": "44. Liu, Z., Lin, Y., Cao, Y., Hu, H., Wei, Y., Zhang, Z., Lin, S., Guo, B.: Swin transformer: Hierarchical vision transformer using shifted windows. In: Int. Conf. Comput. Vis. (ICCV). pp. 9992-10002 (2021)" + }, + { + "type": "ref_text", + "bbox": [ + 0.214, + 0.509, + 0.787, + 0.55 + ], + "angle": 0, + "content": "45. Liu, Z., Mao, H., Wu, C., Feichtenhofer, C., Darrell, T., Xie, S.: A convnet for the 2020s. In: IEEE Conf. Comput. Vis. Pattern Recog. (CVPR). pp. 11966-11976 (2022)" + }, + { + "type": "ref_text", + "bbox": [ + 0.214, + 0.55, + 0.787, + 0.577 + ], + "angle": 0, + "content": "46. Madry, A., Makelov, A., Schmidt, L., Tsipras, D., Vladu, A.: Towards deep learning models resistant to adversarial attacks. In: Int. Conf. Learn. Rep. (ICLR) (2018)" + }, + { + "type": "ref_text", + "bbox": [ + 0.214, + 0.577, + 0.787, + 0.619 + ], + "angle": 0, + "content": "47. Mao, X., Chen, Y., Li, X., Qi, G., Duan, R., Zhang, R., Xue, H.: Easyrobust: A comprehensive and easy-to-use toolkit for robust computer vision. https://github.com/alibaba/easyrobust (2022)" + }, + { + "type": "ref_text", + "bbox": [ + 0.214, + 0.619, + 0.787, + 0.66 + ], + "angle": 0, + "content": "48. Meletis, P., Wen, X., Lu, C., de Geus, D., Dubbelman, G.: Cityscapes-panoptic-parts and pascal-panoptic-parts datasets for scene understanding. arXiv preprint arXiv:2004.07944 (2020)" + }, + { + "type": "ref_text", + "bbox": [ + 0.214, + 0.66, + 0.787, + 0.716 + ], + "angle": 0, + "content": "49. Ramanathan, V., Kalia, A., Petrovic, V., Wen, Y., Zheng, B., Guo, B., Wang, R., Marquez, A., Kovvuri, R., Kadian, A., Mousavi, A., Song, Y., Dubey, A., Mahajan, D.: PACO: parts and attributes of common objects. In: IEEE Conf. Comput. Vis. Pattern Recog. (CVPR). pp. 7141-7151 (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.214, + 0.716, + 0.787, + 0.758 + ], + "angle": 0, + "content": "50. Reddy, N.D., Vo, M., Narasimhan, S.G.: Carfusion: Combining point tracking and part detection for dynamic 3d reconstruction of vehicles. In: IEEE Conf. Comput. Vis. Pattern Recog. (CVPR). pp. 1906-1915 (2018)" + }, + { + "type": "ref_text", + "bbox": [ + 0.214, + 0.758, + 0.787, + 0.799 + ], + "angle": 0, + "content": "51. Ren, S., He, K., Girshick, R.B., Sun, J.: Faster R-CNN: towards real-time object detection with region proposal networks. In: Adv. Neural Inform. Process. Syst. (NeurIPS). pp. 91-99 (2015)" + }, + { + "type": "ref_text", + "bbox": [ + 0.214, + 0.799, + 0.787, + 0.841 + ], + "angle": 0, + "content": "52. Rombach, R., Blattmann, A., Lorenz, D., Esser, P., Ommer, B.: High-resolution image synthesis with latent diffusion models. In: IEEE Conf. Comput. Vis. Pattern Recog. (CVPR). pp. 10674-10685 (2022)" + }, + { + "type": "list", + "bbox": [ + 0.214, + 0.147, + 0.788, + 0.841 + ], + "angle": 0, + "content": null + } + ], + [ + { + "type": "page_number", + "bbox": [ + 0.218, + 0.116, + 0.236, + 0.127 + ], + "angle": 0, + "content": "18" + }, + { + "type": "header", + "bbox": [ + 0.272, + 0.115, + 0.326, + 0.127 + ], + "angle": 0, + "content": "Li et al." + }, + { + "type": "ref_text", + "bbox": [ + 0.218, + 0.148, + 0.785, + 0.175 + ], + "angle": 0, + "content": "53. Rosch, E., Mervis, C.B.: Family resemblances: Studies in the internal structure of categories. Cognitive Psychology 7(4), 573-605 (1975)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.177, + 0.785, + 0.217 + ], + "angle": 0, + "content": "54. Salman, H., Ilyas, A., Engstrom, L., Kapoor, A., Madry, A.: Do adversarially robust imagenet models transfer better? In: Adv. Neural Inform. Process. Syst. (NeurIPS) (2020)" + }, + { + "type": "ref_text", + "bbox": [ + 0.218, + 0.218, + 0.785, + 0.258 + ], + "angle": 0, + "content": "55. Salman, H., Sun, M., Yang, G., Kapoor, A., Kolter, J.Z.: Denoised smoothing: A provable defense for pretrained classifiers. In: Adv. Neural Inform. Process. Syst. (NeurIPS) (2020)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.259, + 0.785, + 0.286 + ], + "angle": 0, + "content": "56. Sitawarin, C., Pongmala, K., Chen, Y., Carlini, N., Wagner, D.A.: Part-based models improve adversarial robustness. In: Int. Conf. Learn. Rep. (ICLR) (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.287, + 0.785, + 0.34 + ], + "angle": 0, + "content": "57. Song, X., Wang, P., Zhou, D., Zhu, R., Guan, C., Dai, Y., Su, H., Li, H., Yang, R.: Apollocar3d: A large 3d car instance understanding benchmark for autonomous driving. In: IEEE Conf. Comput. Vis. Pattern Recog. (CVPR). pp. 5452-5462 (2019)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.342, + 0.785, + 0.37 + ], + "angle": 0, + "content": "58. Sun, P., Chen, S., Zhu, C., Xiao, F., Luo, P., Xie, S., Yan, Z.: Going denser with open-vocabulary part segmentation. arXiv preprint arXiv:2305.11173 (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.371, + 0.785, + 0.411 + ], + "angle": 0, + "content": "59. Szegedy, C., Zaremba, W., Sutskever, I., Bruna, J., Erhan, D., Goodfellow, I.J., Fergus, R.: Intriguing properties of neural networks. In: Int. Conf. Learn. Rep. (ICLR) (2014)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.412, + 0.785, + 0.438 + ], + "angle": 0, + "content": "60. Tramèr, F., Carlini, N., Brendel, W., Madry, A.: On adaptive attacks to adversarial example defenses. In: Adv. Neural Inform. Process. Syst. (NeurIPS) (2020)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.439, + 0.785, + 0.466 + ], + "angle": 0, + "content": "61. Tversky, B., Hemenway, K.: Objects, parts, and categories. Journal of Experimental Psychology: General 113(2), 169 (1984)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.467, + 0.785, + 0.508 + ], + "angle": 0, + "content": "62. Wang, H., Ge, S., Lipton, Z.C., Xing, E.P.: Learning robust global representations by penalizing local predictive power. In: Adv. Neural Inform. Process. Syst. (NeurIPS). pp. 10506-10518 (2019)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.509, + 0.785, + 0.55 + ], + "angle": 0, + "content": "63. Wang, J., Zhang, Z., Xie, C., Premachandran, V., Yuille, A.: Unsupervised learning of object semantic parts from internal states of cnns by population encoding. arXiv preprint arXiv:1511.06855 (2015)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.55, + 0.785, + 0.577 + ], + "angle": 0, + "content": "64. Wu, D., Xia, S., Wang, Y.: Adversarial weight perturbation helps robust generalization. In: Adv. Neural Inform. Process. Syst. (NeurIPS) (2020)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.578, + 0.785, + 0.618 + ], + "angle": 0, + "content": "65. Zhang, H., Yu, Y., Jiao, J., Xing, E.P., Ghaoui, L.E., Jordan, M.I.: Theoretically principled trade-off between robustness and accuracy. In: Int. Conf. Mach. Learn. (ICML). vol. 97, pp. 7472-7482 (2019)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.619, + 0.785, + 0.66 + ], + "angle": 0, + "content": "66. Zhang, Y., Yao, Y., Jia, J., Yi, J., Hong, M., Chang, S., Liu, S.: How to robustify black-box ML models? A zeroth-order optimization perspective. In: Int. Conf. Learn. Rep. (ICLR) (2022)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.661, + 0.785, + 0.715 + ], + "angle": 0, + "content": "67. Zhao, J., Li, J., Cheng, Y., Sim, T., Yan, S., Feng, J.: Understanding humans in crowded scenes: Deep nested adversarial learning and A new benchmark for multi-human parsing. In: ACM Int. Conf. Multimedia (ACM MM). pp. 792-800 (2018)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.716, + 0.785, + 0.743 + ], + "angle": 0, + "content": "68. Zhao, Y., Li, J., Zhang, Y., Tian, Y.: Multi-class part parsing with joint boundary-semantic awareness. In: Int. Conf. Comput. Vis. (ICCV). pp. 9176-9185 (2019)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.744, + 0.785, + 0.785 + ], + "angle": 0, + "content": "69. Zhou, B., Zhao, H., Puig, X., Xiao, T., Fidler, S., Barriuso, A., Torralba, A.: Semantic understanding of scenes through the ADE20K dataset. Int. J. Comput. Vis. (IJCV) 127(3), 302-321 (2019)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.786, + 0.785, + 0.813 + ], + "angle": 0, + "content": "70. Zhou, Z., Firestone, C.: Humans can decipher adversarial images. Nature Communications 10(1), 1-9 (2019)" + }, + { + "type": "list", + "bbox": [ + 0.217, + 0.148, + 0.785, + 0.813 + ], + "angle": 0, + "content": null + } + ] +] \ No newline at end of file diff --git a/2024/PartImageNet++ Dataset_ Scaling up Part-based Models for Robust Recognition/0d30f97b-2cbe-4f46-92c7-0bf8c016e1e3_origin.pdf b/2024/PartImageNet++ Dataset_ Scaling up Part-based Models for Robust Recognition/0d30f97b-2cbe-4f46-92c7-0bf8c016e1e3_origin.pdf new file mode 100644 index 0000000000000000000000000000000000000000..0be30ac0326bdfb24811c29e2d248c8ff7e20e8e --- /dev/null +++ b/2024/PartImageNet++ Dataset_ Scaling up Part-based Models for Robust Recognition/0d30f97b-2cbe-4f46-92c7-0bf8c016e1e3_origin.pdf @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:76e63332bbd01dd97845322a03a9c9822b67d303f8fa174c30a829c8296278f4 +size 1509931 diff --git a/2024/PartImageNet++ Dataset_ Scaling up Part-based Models for Robust Recognition/full.md b/2024/PartImageNet++ Dataset_ Scaling up Part-based Models for Robust Recognition/full.md new file mode 100644 index 0000000000000000000000000000000000000000..a8b4bd3d27c0dae33f837f4e4d7b575bdb52f1c1 --- /dev/null +++ b/2024/PartImageNet++ Dataset_ Scaling up Part-based Models for Robust Recognition/full.md @@ -0,0 +1,275 @@ +# PartImageNet++ Dataset: Scaling up Part-based Models for Robust Recognition + +Xiao Li $^{1}$ , Yining Liu $^{2}$ , Na Dong $^{3}$ , Sitian Qin $^{2}$ , and Xiaolin Hu $^{1,4*}$ + +$^{1}$ Department of Computer Science and Technology, BNRist, Institute for Artificial Intelligence, IDG/McGovern Institute for Brain Research, Tsinghua Laboratory of Brain and Intelligence, Tsinghua University, Beijing, China + $^{2}$ Harbin Institute of Technology, Weihai, China + $^{3}$ Beijing Institute of Technology, Beijing, China + $^{4}$ Chinese Institute for Brain Research (CIBR), Beijing, China +lixiao20@mails.tsinghua.edu.cn, 22S030184@stu.hit.edu.cn, 3220220911@bit.edu.cn, qinsitian@hitwh.edu.cn, xlu@mail.tsinghua.edu.cn + +Abstract. Deep learning-based object recognition systems can be easily fooled by various adversarial perturbations. One reason for the weak robustness may be that they do not have part-based inductive bias like the human recognition process. Motivated by this, several part-based recognition models have been proposed to improve the adversarial robustness of recognition. However, due to the lack of part annotations, the effectiveness of these methods is only validated on small-scale non-standard datasets. In this work, we propose PIN++, short for PartImageNet++, a dataset providing high-quality part segmentation annotations for all categories of ImageNet-1K (IN-1K). With these annotations, we build part-based methods directly on the standard IN-1K dataset for robust recognition. Different from previous two-stage part-based models, we propose a Multi-scale Part-supervised Model (MPM), to learn a robust representation with part annotations. Experiments show that MPM yielded better adversarial robustness on the large-scale IN-1K over strong baselines across various attack settings. Furthermore, MPM achieved improved robustness on common corruptions and several out-of-distribution datasets. The dataset, together with these results, enables and encourages researchers to explore the potential of part-based models in more real applications. The dataset and the code are available at https://github.com/LixiaoTHU/PartImageNetPP. + +Keywords: Adversarial robustness $\cdot$ Part dataset $\cdot$ Robust object recognition $\cdot$ Part-based model + +# 1 Introduction + +Object recognition has achieved remarkable success due to the rise of Deep Neural Networks (DNNs) [23, 30, 44, 45]. However, DNNs often lack robustness and + +![](images/99d95f6b34d987bf2a3e8ecefab8c0dd433af43d20be3e12595fd1fc0b7cc4ca.jpg) +Fig. 1: Examples of annotated images in $\mathrm{PIN}++$ . High-quality part segmentation annotations are provided on all categories in IN-1K. The object names are shown on the top-right of each image. The part names are hidden here. + +![](images/5d7e5d39c16a6afbfcbe90e67c26bd757d3049b7ee7d081737dada980f114120.jpg) + +![](images/07f2dfb8820e93cc7699018c344d79c33212124cac03f85d618a2ffe5e9342b0.jpg) + +can be easily deceived by adversarial examples [19, 59], common image corruptions [24], and various out-of-distribution (OOD) shifts [34, 62], which greatly hinder applications of DNNs in security-critical scenarios [12, 33]. One possible reason for the weakness may be that DNN-based recognition does not have a part-based inductive bias (i.e., recognition based on object parts) like the human recognition process, which is generally regarded as a robust system [70]. According to the well-known cognitive psychological theory recognition-by-components [4], humans prefer to recognize objects by decomposing objects into parts and taking into account the hierarchical representations and the spatial relationships of parts. This cognitive theory has been supported by numerous psychological evidence [20, 26, 53, 61]. However, several studies revealed that DNNs often exhibit a shortcut recognition behavior, e.g., relying on texture-based features [16] and even unintelligible features for humans [27]. + +If DNNs can perform a human-aligned part-based recognition, both the adversarial and non-adversarial robustness of DNNs could be closer to humans. Motivated by this idea, two recent works [35, 56] investigated how to use additional part-level supervision to build up part-based (recognition) models, which make predictions by recognizing the object's parts in a bottom-up manner. Both works investigated two-stage part-based models, where the first stage involves part segmentation, followed by the classification stage based on the segmentation results. These works have shown the potential of part-based models in enhancing robustness against various perturbations. + +Nevertheless, due to the lack of part annotations, the effectiveness of part-based models has only been preliminary validated by the two works [35, 56] on small-scale nonstandard datasets. The largest dataset for evaluation was PartImageNet (PIN) [21], consisting of about $24\mathrm{K}$ images of 158 categories selected from the large-scale ImageNet-1K (IN-1K) [11]. PIN has several potential issues. Firstly, it is quite small compared to the standard IN-1K (about $1.3\mathrm{M}$ images) commonly used in the domain of adversarial robustness [3, 54]. This limited scale may restrict the generalizability of the findings. Secondly, categories included in + +PIN are specifically chosen to facilitate part-related tasks (e.g., part segmentation [43] and object parsing [68]), with an emphasis on animal categories (118 out of 158). While this focus can be beneficial for studying part-related tasks themselves, it poses challenges in demonstrating the effectiveness of part-based models for general recognition tasks across a diverse range of objects. Thirdly, the annotation scheme of PIN is simple. It groups 158 categories into 11 super-categories and determines part annotations for each super-category, potentially failing to capture the characteristics of object categories within a super-category. Due to the three issues, the part-based models [35,56] trained on PIN cannot be directly compared with state-of-the-art (SOTA) methods trained on the standard IN-1K. + +In this work, we introduce PartImageNet++ (PIN++, a new dataset providing high-quality part segmentation annotations for all categories of the widely used IN-1K [11], to facilitate the research on part-based models for robust recognition (classification). To create PIN++, we designed a detailed annotation scheme to ensure high-quality part annotations. With this scheme, we annotated 100 randomly selected images per category, resulting in a total of 100K images with part annotations (about 1/13 of IN-1K). To avoid bias, all the part annotations are created manually without using any auxiliary model. To the best of our knowledge, PIN++ provides large-scale high-quality part segmentation annotations for the most diverse range of object categories, including creatures, artifacts, rigid objects, nonrigid objects, etc, among existing part datasets. Fig. 1 showcases some examples of annotated images. We introduce this dataset in detail in Sec. 3 and show that existing techniques fail to obtain high-quality part segmentation results without such annotations. + +By leveraging the part annotations of $\mathrm{PIN}++$ , we develop part-based methods for robust recognition directly on the standard IN-1K dataset. To achieve this, we first train a part segmentation network with the annotations of $\mathrm{PIN}++$ to obtain pseudo part labels for all unannotated images in IN-1K. We then propose a new Multi-scale Part-supervised recognition Model, termed MPM, to better exploit the part annotations. MPM is expected to learn a robust intermediate representation by adding auxiliary bypass layers to the vanilla recognition model (e.g., ResNet-50 [23]). The auxiliary bypass layers are supervised by the part annotations and the generated pseudo part labels. Without introducing any additional parameter or computation during inference, MPM can use high-resolution part annotations more adequately than two-stage part-based models [35, 56], which are constrained to use low-resolution part labels. Fig. 3 shows an overview of our whole method. + +The experimental results demonstrate that MPM, combined with adversarial training (AT) [46], achieved better adversarial robustness on the large-scale IN-1K, outperforming strong AT baselines in both the seen and unseen adversarial threats. Furthermore, MPM exhibited improved robustness on common image corruptions [24] and several OOD datasets. In addition, MPM achieved higher alignments with human vision. Beyond object recognition, MPM is shown to + +boost the adversarial robustness of downstream tasks such as object detection [51]. The main contributions of this work can be summarized as follows: + +- We created PIN++, a new large-scale dataset with part annotations. This dataset can facilitate further research on part-based models for robust recognition and other part-related visual understanding tasks. +- We proposed MPM, a new part-supervised recognition model, to exploit the part annotations better, achieving superior robustness across various settings and benchmarks without introducing any extra inference cost. + +# 2 Related Work + +Adversarial robustness. Szegedy et al. [59] first revealed the vulnerability of DNNs to adversarial examples. Since then, numerous methods [55,66] have been proposed to enhance the adversarial robustness of DNNs. Among them, AT has emerged as the de facto paradigm for training adversially robust DNNs. Early AT works [64,65] primarily focused on small-scale, low-resolution datasets such as CIFAR-10 [29]. Recently, there has been a growing interest in investigating AT on IN-1K [11], as this dataset serves as a standard benchmark for evaluating SOTA computer vision techniques [23,44,45]. Several studies have demonstrated that AT models trained on IN-1K can transfer their adversarial robustness to downstream dense-prediction tasks and achieve zero-shot adversially robust recognition [9, 33, 36]. + +Part datasets. The concept of parts holds significant importance in both human visual cognition [4] and computer vision domains. To conduct research on part-related visual tasks, the availability of part datasets is crucial. However, annotating object parts is challenging and expensive, and most part datasets are only limited to specific domains, such as cars [50, 57, 63] and humans [18, 37, 38, 67]. Furthermore, part datasets targeting common objects often have a limited number of object categories. Cityscapes PanopticParts (Cityscapes PP.) [48], PascalPart [6], and ADE20K provide part annotations for 5, 20, and 80 object categories, respectively. Although the recent PACO dataset [49] offers a large number of part annotations, it is still restricted to only 75 object categories. PIN [21], containing 158 categories from IN-1K [11], represents the dataset with the most object categories in terms of part annotations. + +Part-based object recognition. Modeling objects in terms of parts has a long history in computer vision [5,14,17]. But with the rise of DNNs, parts have been rarely used as auxiliaries for end-to-end general object recognition [23,30,44,45]. Recently, two works [35, 56] revisited part-based recognition models from the robustness perspective and showed the potential to enhance adversarial robustness. Li et al. [35] proposed ROCK, a part-based model that first predicts all parts represented by segmentation results, and then utilizes a judgment block to give category predictions based on the predicted parts. Similarly, Sitawarin et al. [56] also investigated several types of two-stage part-based models, where the first stage involves part segmentation, followed by the recognition stage with a tiny classifier. However, as mentioned before, both works were only validated on + +small-scale part-friendly datasets for the lack of part annotations, e.g., Sitawarin et al. [56] performed classification tasks only on 11 super-categories of PIN [21]. Furthermore, the proposed two-stage part-based models inevitably introduced a lot of extra parameters during inference, while the utilization of high-resolution part annotations was limited, as discussed in Sec. 4.2. In addition, ROCK [35] contains a non-differentiable process, which could lead to a potential overestimation of adversarial robustness [2,60]. Our work will address all of these issues. + +# 3 PIN++ Dataset + +We first present the details of how we built $\mathrm{PIN}++$ to ensure high-quality part annotations, followed by the statistics of this dataset. We then give a discussion on $\mathrm{PIN}++$ . + +# 3.1 Annotation Scheme + +Data source. IN-1K [11], containing about 1.3M images with category labels in the training set, is one of the most widely used datasets for general object recognition. Our aim is to provide part annotations for part-based robust recognition on IN-1K, and thus the images of PIN++ directly reuse the training set of IN-1K. All the images conform to licensing for research purposes. However, due to cost constraints, it is impractical to provide part-level annotations for all training images in IN-1K. Instead, we decided to annotate 100 randomly selected images per category. Once part annotations are provided for all categories, modern supervised segmentation techniques [22] can be used to obtain pseudo part labels for the remaining unannotated images, as we will show in Sec. 4. Note that PIN++ does not provide any part annotations for the validation set of IN-1K. + +While PIN [21] has provided some part annotations for 158 categories of IN-1K, their annotations do not fully meet our purpose and principles. To minimize annotation costs without compromising quality, we retained annotations of 90 categories from PIN as part of $\mathrm{PIN}++$ . The details are shown in Appendix A. We then show the annotation design for the remaining 910 categories of IN-1K. Annotation quality control. To ensure high-quality part annotations, three points should be taken into account: + +1) Deciding which parts to annotate per category. One of the key challenges in the part annotation task is the ambiguity of object part selection (e.g., how to annotate the parts of hammer). PIN [21] evades this problem and focuses primarily on annotating parts for animal categories (most quadrupeds can be regarded as consisting of four part categories: head, body, foot, tail). However, the remaining categories in IN-1K are highly diverse, making it challenging to decide which parts to annotate. To ensure a scientifically grounded division of parts for each category, we initially consult the Wikidata knowledge base5 to obtain part vocabularies for every object category (e.g., hammer consists of three parts according to Wikidata: handle, striker, and hammerhead). Additionally, for object + +Table 1: The numbers of annotated object categories, part categories, images and part masks of publicly available part datasets. + +
DatasetObject CategoryPart CategoryImagePart Mask
Cityscapes PP. [48]5233.5K100K
Pascal-Part [6]2019319K363.5K
ADE20K* [69]8056612.6K193.2K
PACO [49]7545676.7K641.4K
PIN [21]15860924K112K
PIN++10003308100K406.4K
+ +*: ADE20K is severely category-unbalanced and here we show statistics for categories with more than 10 annotated part masks. + +categories without a clear definition of parts on Wikidata, we decide which parts to annotate by asking recruited volunteers which parts they think to be important to their cognition of the object. For example, according to the cognition of volunteers, for most quadruped categories, annotating head, body, foot, and tail is enough, while for ram (see Fig. 1 for one example), horn should be annotated as an extra part. For categories that are indeed difficult to decompose into parts, e.g., flatworm, we generalize the concept of "parts" and treat the foreground of an object as one part category. In this case, the part is the object itself. Other details on deciding the parts to annotate are further discussed in Appendix B. These principles ensure that all categories in IN-1K can have part annotations. + +2) Designing the part segmentation principles. We design several part segmentation principles to guide the annotators and ensure high-quality part segmentation annotation. Firstly, the annotated part masks are required to be combined to cover the entire object and have no overlap, unless in the case of the second principle. Secondly, for some parts that indeed should overlap in semantic concepts (e.g., horn and head), the annotators should annotate the inclusion relation of parts (e.g., horn is included in head), and then these part masks can be annotated with overlap. Thirdly, to keep the consistency of part annotations among images of the same object category, following [21], the same annotator should annotate all 100 images of this object category. The other principles are shown in Appendix C. + +3) Annotation quality inspection. Ten randomly selected images together with the annotation visualizations (see Fig. 1) are sent to inspectors to assess if the annotations for a specific object category meet the defined principles. If the annotations for any two images fail to meet these principles, the entire category's annotations are re-annotated until the requirements are satisfied. + +# 3.2 Statistics + +We finally annotated 100 images per category for 910 categories in IN-1K, resulting in a total of 91K images with part annotations. Taking these data and the data of 90 reused categories of PIN together, $\mathrm{PIN}++$ for 1K categories is + +obtained. Following IN-1K, here we ensure PIN++ is class-balanced. PIN++ is split into train/val sets with a 9:1 ratio for each category. But note that all annotated images are part of the training set of IN-1K. Unless stated otherwise, they are all utilized for training in our following experiments and the evaluation is performed on IN-1K val set. + +In summary, $\mathrm{PIN}++$ includes a total of 100K images with part annotations. These annotations cover 3,310 part categories for 1,000 object categories. A total of 406,364 part masks are annotated in $\mathrm{PIN}++$ . Further statistics of $\mathrm{PIN}++$ are shown in Appendix D. + +Comparison with other part datasets. We compare PIN++ with publicly available part datasets, as shown in Tab. 1. PIN++ outperforms previous datasets in terms of the number of object and part categories, as well as the total number of images with annotated part masks. It also demonstrates the competitive number of part masks compared to the recent PACO dataset [49]. Notably, PIN++ offers part annotations for a diverse range of objects, including creatures, artifacts, rigid objects, and nonrigid objects, unlike PACO or PIN, which are primarily focused on common tools or animals, respectively. As shown in Appendix D, the annotation quality of PIN++ exhibits an advantage over that of PIN. Moreover, PIN++ is category-balanced and guarantees that each image contains only one foreground category. All these features build the unique advantage of PIN++ for studying object recognition tasks. + +# 3.3 Discussion on PIN++ + +Recent advancements in segmentation include CLIP-based open-vocabulary segmentation methods [32,39,58] and the SAM model [28]. We compared these techniques with the part annotations of PIN++. We used the largest version of SAM, and for open-vocabulary segmentation, we employed the VLPart model [58], which is trained with part annotations from several part datasets including Pascal-Part [6], PIN [21], and PACO [49]. Visual comparisons are presented in Fig. 2. We can see that VLPart only provides roughly reasonable semantics for part categories it has been trained on. It fails to accurately identify the semantics of parts from unseen objects. On the other hand, SAM seems to perform well in segmenting certain parts based on the object's edges, but it struggles when the part edges are less distinguishable, as shown in the segmentation results for camel (the third row in Fig. 2). In addition, note that all masks segmented by SAM are category-agnostic. + +The above results highlight the challenges faced by existing techniques in achieving accurate part segmentation for diverse objects without training on PIN++. Additionally, current text-to-image generative models [52] also struggle with understanding part semantics [42]. However, supervised training on PIN++ yields good part segmentation results, as shown in Appendix E. Therefore, besides its usage in robust object recognition (as demonstrated next), we believe that PIN++ and its extensive collections of 3308 part categories (vocabularies) have the potential to enhance general part-related visual understanding tasks. + +![](images/e529d980f0f4400e03052c8750ac3ec7e18b9d69b90e6fac69a5eb40752d1a76.jpg) +Fig. 2: Comparison between part segmentation results of different methods and PIN++ annotations. Without training on PIN++, VLPart and SAM fail to segment objects into specific parts with accurate semantics. + +# 4 Part-based Recognition Method + +With the annotations of $\mathrm{PIN}++$ , we develop part-based methods directly on the standard IN-1K. Our part-based methods can be divided into two steps. We first obtain pseudo part labels for the remaining unlabeled images of IN-1K and then train the MPM with these part annotations and pseudo labels. The pipeline of our part-based methods is illustrated in Fig. 3. + +# 4.1 Pseudo Part Label Generation + +Part segmentation model. Arbitrary instance segmentation models [13, 22] can be leveraged to perform part segmentation and generate pseudo part labels when trained with the part annotations from $\mathrm{PIN}++$ . Following Sun et al. [58], we use the classical Mask R-CNN [22] model with a vision transformer Swin-B [44] as the backbone network. In training the Mask R-CNN, we consider the masks for each part category as masks for individual object categories directly. During inference for obtaining pseudo part labels, we incorporate a simple post-processing operation alongside the regular inference procedure. + +Post-processing. Given an image $\mathbf{x} \in \mathbb{R}^{3 \times H \times W}$ with object category $y_{c}$ , where $H \times W$ represents the resolution, the regular inference procedure of the trained Mask R-CNN produces pseudo part labels in the form of $\{(\mathbf{M}_{\mathbf{p}}, \mathbf{v}_{\mathbf{p}})\}$ , where $\mathbf{M}_{\mathbf{p}} \in \{0,1\}^{H \times W}$ denotes the binary part mask, $\mathbf{v}_{\mathbf{p}} \in [0,1]^{K}$ denotes the output probability of part categories, and $K$ denotes the number of part categories (3308 in our case). But note that the images that need inference are all in the training set of IN-1K. Their object category $y_{c}$ is not utilized in the regular inference procedure. To leverage it, we apply a Category Filter (CF) operation + +![](images/0c28df4a0c5bf1f10a9da094faf07570f46ba4a83f395256ee4b8d3467d539fa.jpg) +Fig. 3: An overview of the generation of pseudo-labels and the structure of MPM. (a) A part segmentation model trained on $\mathrm{PIN}++$ and used to obtain pseudo-part labels for unannotated images. (b) MPM adds several auxiliary bypass layers to the vanilla recognition model for part segmentation supervision. MPM is trained by part annotations together with the pseudo part labels. During inference, the auxiliary layers are dropped, and the vanilla recognition model gives the final object category prediction. + +![](images/830f9567b4d8b79f78cf57d39ae85fce31b21491f9c33169dec4938f65f22068.jpg) + +to obtain $\hat{\mathbf{v}}_{\mathbf{p}}$ , which sets the probabilities of unrelated part categories in $\mathbf{v}_{\mathbf{p}}$ to zero (e.g., for cat, except for cat:head, cat:body, etc, other part categories such as radio:button and bird:head are all excluded). The pseudo part category for $\mathbf{M}_{\mathbf{p}}$ is then computed as $y_{p} = \arg \max \hat{\mathbf{v}}_{\mathbf{p}}$ . + +# 4.2 Multi-scale Part-supervised Recognition Model + +Data preparation. After obtaining the pseudo labels for all remaining training images of IN-1K, we simply treat the pseudo labels and real part annotations equally because we find that the quality of the pseudo labels is roughly satisfactory, as shown in Appendix E. The part labels $\{(\mathbf{M_p},\mathbf{y_p})\}$ for an image $\mathbf{x}$ are then converted to a single composite segmentation mask $\mathbf{M} \in \{0,1\}^{H \times W \times (K + 1)}$ where the extra one channel indicates the background. Finally, the data samples for training part-based models are $\{(\mathbf{x},y_c,\mathbf{M})\}$ . + +Model design. Both previous works [35, 56] build part-based models in the two-stage way: a segmenter $\mathcal{F}_{\mathrm{seg}}: \mathbb{R}^{3 \times H \times W} \to \mathbb{R}^{(K + 1) \times h \times w}$ and an extra module $\mathcal{F}_{\mathrm{cls}}: \mathbb{R}^{(K + 1) \times h \times w} \to \mathbb{R}^C$ for classification, where $h \times w$ denotes the output resolution and $C$ denotes the number of object category. $\mathcal{F}_{\mathrm{seg}}$ is a backbone network (e.g., a ResNet-50 without classification head). Most popular backbones [23, 44, 45] are hierarchically composed of several blocks with down-sampling layers and thus they can be written as $\mathcal{F}_{\mathrm{seg}} := f_l \circ \dots \circ f_1$ , where $l$ denotes the number of down-sampling layers and $f_{i(1 \leq i \leq l)}$ denotes one block with the $i$ th down-sampling layer. $\mathcal{F}_{\mathrm{cls}}$ can be a tiny classifier [56] or a nondifferentiable judgment block [35] and its input is the part segmentation results $\hat{\mathbf{M}} = \mathcal{F}_{\mathrm{seg}}(\mathbf{x})$ . The overall model for recognition is $\mathcal{F} := \mathcal{F}_{\mathrm{cls}} \circ \mathcal{F}_{\mathrm{seg}}$ . However, as $K$ is significantly larger than $C$ , using $\hat{M}$ as the intermediate results for + +recognition inevitably introduce extra parameters and computation during inference. This situation gets worse with an increased number of part categories, as shown in Sec. 5.5. Moreover, these methods solely employ part supervision on the output of $f_{l}$ . Without expanding $\mathcal{F}_{\mathrm{seg}}$ , part annotations have to be downsampled to match the output resolution of $f_{l}$ , potentially limiting the utilization of high-resolution part annotations. + +Instead of these two-stage part-based models, MPM is expected to learn a robust intermediate representation by adding multi-scale bypass layers to the vanilla recognition network. In MPM, the model used for object recognition is the vanilla backbone network directly: $\mathcal{F} \coloneqq h_c \circ f_l \circ \dots \circ f_1$ , where $h_c$ denotes a vanilla classification head (usually a linear layer). MPM utilizes the part annotations to supervise the intermediate features of $\mathcal{F}$ by several bypass layers $g_i$ : $\tilde{\mathbf{M}}_i = g_i \circ f_i \circ \dots \circ f_1(\mathbf{x})$ , where $i \leq l$ . In this way, the intermediate features of $\mathcal{F}$ can be seen as implicit representations of part segmentation results when $g_i$ is simple enough. Note that here we use part supervision not only on the output of $f_l$ , but also $f_{i(i < l)}$ . As the output of lower layers has a larger resolution, it can be supervised by part annotations with higher resolution. But the outputs of $f_{i(i < l)}$ are relatively low-level, which may not be enough to obtain high-level part segmentation results themselves. To balance this, only the outputs of the last three blocks $f_{i(l - 2 \leq i \leq l)}$ are supervised by part mask $\mathbf{M}$ (e.g., for a ResNet-50 with input size $224 \times 224$ , the intermediate features of $7 \times 7$ , $14 \times 14$ , and $28 \times 28$ are supervised by corresponding down-sampled $\mathbf{M}$ , respectively). In addition, following the principle of FPN [40], some top-down layers are used to augment the low-level features with high-level features. Note that distinct from the conventional FPN that generally has comparable parameters with the backbone network [40], these bypass layers are extremely lightweight as our goal is to improve the recognition accuracy rather than the quality of part segmentation results. Fig. 3(b) shows the overall structure of MPM. During inference, these auxiliary layers are dropped, and the vanilla recognition model gives the final object category prediction. + +Training objective. The overall loss for training MPM is: $L = L_{\mathrm{cls}} + \lambda \cdot L_{\mathrm{seg}}$ , where $L_{\mathrm{cls}}$ denotes the vanilla loss for classification, $L_{\mathrm{seg}}$ denotes the loss for part segmentation, and $\lambda$ is a hyper-parameter. We compute $L_{\mathrm{seg}}$ as the average of losses on three part segmentation results of different resolutions, as mentioned before. Note that when performing AT, adversarial examples are generated by $\mathbf{x}^{\star} = \arg \max_{\mathbf{x}^{\star}:||\mathbf{x}^{\star} - \mathbf{x}||_{p}\leq \epsilon}L_{\mathrm{cls}}(\mathcal{F}(\mathbf{x}^{\star}),y)$ , while MPM is always trained with $L$ . + +# 5 Experiments + +# 5.1 Experimental Setup + +Training setup. Unless specified otherwise, following previous works [3,10], we performed AT with $l_{\infty}$ bound $\epsilon = 4 / 255$ on IN-1K. The inner optimization for obtaining adversarial examples used PGD [46] with iterative steps $t = 2$ and the step size was set to be $s = 2*\epsilon /t$ . The input resolution used $224\times 224$ . We used the vanilla version of the AT [46] and did not incorporate other variants + +Table 2: Recognition accuracies (\%) of methods under different attack threats on IN-1K, following the evaluation setup in Sec. 5.1. Average denotes the average accuracies in four attack settings. We highlight the best results in each column. + +
ArchitectureMethodAdversarial Train wrt l∞ (ε = 4/255)
Cleanl∞l∞(×2)l1l2Average
ResNet-50[54], 202063.935.913.22.013.716.2
ResNet-50[47], 202264.734.311.42.316.416.1
ResNet-50[41], 202365.532.69.53.317.915.8
ResNet-50-Gelu[3], 202167.836.610.83.016.816.8
ResNet-50-Gelu[10], 202366.835.612.83.616.517.1
ResNet-50-Geluours (vanilla)67.138.112.65.021.619.3
ResNet-50-Geluours (MPM)67.839.113.66.224.320.8
+ +such as TRADES [65] or AWP [64]. These variants have shown effectiveness mainly on small-scale datasets, making their generalization to IN-1K non-trivial. Following previous part-based works [35,56], we used the classical ResNet-50 [23] network with 25.6M parameters as the baseline model. We note that recent works [3,10] used a GELU [25] activation to replace the original RELU [1] activation of ResNet-50 for boosting robustness when performing AT, here we also used GELU, denoted as ResNet-50-Gelu. In addition, we improved the training recipe of ResNet-50 to build a strong baseline. Other details on the training recipe are provided in Appendix F. Note that bypass layers (e.g., point-wise convolutions) with about 4.5M parameters were used by MPM during training but completely dropped during inference. The code is submitted along with the paper. + +Evaluation setup. For all experiments in the work, unless specified otherwise, we evaluated the adversarial robustness with AutoAttack [8], which is a combination of various attack methods and is generally recognized as a reliable evaluation [7]. The evaluation was performed at a resolution $224 \times 224$ on randomly selected 10K images of the 50K IN-1K val set, 10 images for each category. This ensures that the $95\%$ confidence intervals of the reported average adversarial robustness (see Tab. 2) were less than $\pm 0.5\%$ . Note that, unlike previous works [3,10], we did not use the 5K images selected by RobustBench [7] as we found that they were category-unbalanced, containing three categories even without any image for validation. + +We expect the AT models with $l_{\infty}$ bound $\epsilon = 4 / 255$ to be adversarially robust not only on seen threats (i.e., $l_{\infty}$ attack with $\epsilon = 4 / 255$ ) but also on unseen attack threats (e.g., $l_{\infty}$ attack with larger $\epsilon, l_1$ and $l_2$ attacks). Thus, we evaluated the models on three attack threats: $l_{\infty}$ , $l_1$ , and $l_2$ , with the bounds $\epsilon_{\infty} = 4 / 255$ , $\epsilon_1 = 75$ , $\epsilon_2 = 2$ , respectively. For $l_{\infty}$ attack, we extra evaluated the models with $\epsilon_{\infty} = 8 / 255$ to mimic a stronger threat, denoted as $l_{\infty}(\times 2)$ . + +# 5.2 Robustness against Adversarial Attacks + +With the above setup, we trained the vanilla ResNet-50-Gelu as a baseline. MPM used the same recipe as the baseline model for a fair comparison, except that it + +Table 3: Recognition accuracies $(\%)$ of ResNet-50-Gelu on 15 different common image corruptions [24]. We adopt accuracy as the metric to be consistent with other results, while this metric can be easily converted into the corruption error metric [24]. "Average" denotes the accuracy averaged over different common corruptions. + +
Part ATNoiseBlurWeatherDigitalAvg.
Gauss. Shot Impul.Defoc.Glass Moti.ZoomSnow FrostFogBrightCont.Elast.Pixel JPEG
31.230.727.339.028.639.537.835.840.753.869.038.037.047.157.040.8
✓31.531.028.441.929.840.336.737.041.355.369.639.636.952.959.942.1
✓29.628.521.023.332.932.834.434.234.010.858.49.552.258.863.134.9
✓30.129.222.425.235.135.036.436.436.211.959.710.453.159.864.336.4
+ +Table 4: Recognition accuracies (%) of models on four OOD datasets. + +
PartATIN-A-PlusIN-SketchSINDINAverage
6.925.86.953.523.3
✓7.425.77.453.923.6
✓5.325.112.555.424.6
✓✓5.726.412.655.925.2
+ +used an extra $L_{\mathrm{seg}}$ to introduce part-based supervision. $\lambda$ was simply set to 1. In addition, we compared MPM with all open-sourced ResNet-50 checkpoints adversarially trained on IN-1K in recent years. The recognition accuracies under different attack threats are shown in Tab. 2. We can see that our baseline model is strong enough, outperforming all other benchmarks by $2.2\%$ on average robustness (17.1% vs. 19.3%). MPM further shows significantly better adversarial robustness than the vanilla baseline across all these attack threats. In addition, compared with our strong vanilla baseline, the accuracy of clean images is also improved with MPM. To our best knowledge, these are SOTA results of adversarial robustness with ResNet-50 on IN-1K. We additionally show in Appendix G that these results are significantly better than previous results on PIN. + +# 5.3 Robustness on Corruptions and OOD Datasets + +Except for robustness against adversarial attacks, which consider the worst-case security threat scenarios, we evaluated the robustness of our part-based model on common image corruptions [24] and several OOD datasets [15, 16, 34, 62], which could be more realistic threats. Considering that AT can have negative effects on these non-adversarial scenarios [35, 41], here we also evaluated models without AT. The training details are also shown in Appendix F. + +For common image corruptions [24], we generated different types of corrupted images on the whole IN-1K val set. The results of different models on these images are shown in Tab. 3, where "Part" indicates MPM. Here the results of each corruption type are the accuracies averaged on five severity levels [24]. It is seen that MPM has better robustness on nearly all types of image corruptions than models without part supervision, regardless of AT. Besides common corruptions, we evaluated the models on four OOD datasets: ImageNet-A-Plus (IN-A-Plus) + +Table 5: Segmentation accuracies (AP) of Mask R-CNN with and without CF, and recognition accuracies (\%) of MPM on IN-1K trained with various pseudo labels. + +
Pseudo LabelCFAP\( AP_{50} \)Clean\( l_{\infty} \)\( l_{\infty}(\times 2) \)
--65.533.09.4
✓37.258.667.838.813.1
✓✓40.764.467.839.113.6
+ +[34], ImageNet Sketch (IN-Sketch) [62], Stylized ImageNet (SIN) [16], and image distortion dataset (denoted as DIN) [15]. See Appendix H for the introduction of these datasets. The results shown in Tab. 4 indicate that MPM also exhibits improved robustness on these OOD shifts. + +# 5.4 Other Advantages of MPM + +Models versus humans. The part-based models are motivated by the human recognition process. We were interested in whether the robustness of MPM aligned more with human cognition. To investigate this, we employed the evaluation method introduced by Geirhos et al. [15], which involved comparing the models' decisions on different distorted images with the actual judgments made by human observers. The results shown in Appendix I indicate that MPM measurably improved alignment with the human recognition process. + +Boosting the downstream task. Li et al. [33] show that adversarially trained models on the large-scale IN-1K can be utilized to initialize the backbone of downstream networks (e.g., Faster R-CNN [51]). This approach together with downstream AT enables the transfer of adversarial robustness to downstream tasks. We conducted similar experiments by using the checkpoints of the vanilla baseline and MPM. The results shown in Appendix J indicate that MPM enhances both the clean accuracy and adversarial robustness in object detection. + +# 5.5 Ablation Study + +Pseudo-label quality. We first investigate the influence of pseudo-label quality on the adversarial robustness of MPM. In Sec. 4.1, we illustrate the use of CF during generating pseudo labels. Here we remove the CF operator and regenerate pseudo labels for images without part annotations. The left half of Tab. 5 (the 2nd and 3rd rows) shows the segmentation accuracies of models with and without CF (here the models were supervised with the train set of $\mathrm{PIN}++$ , and the AP were calculated on val). It is seen that training with $\mathrm{PIN}++$ achieved good part segmentation results, indicating its potential for enhancing part-related tasks. CF further significantly improves AP. + +We then trained MPM using both pseudo labels and real part annotations, as well as solely using real part annotations. The results, displayed in Tab. 5 (the 1st row), indicate that training MPM solely with real part annotations seems to be ineffective, potentially compromising the model's performance. We guess + +Table 6: Recognition accuracies (\%) of different part-based models on IN-1K. The number of parameters during inference is listed. + +
MethodParam.Cleanl∞l∞(×2)
Two-stage60.9M67.338.713.0
SV167.336.98.8
SV2 w/ TD25.6M67.637.111.7
MPM (SV3 w/ TD)67.839.113.6
SV3 w/o TD67.436.911.7
+ +it could caused by overfitting on the images with part annotations. Besides, the inclusion of better-quality pseudo labels can improve the robustness of our part-based model, underscoring the importance of high-quality pseudo labels. + +Part-based model structure. We then conducted ablation experiments on the structure of the part-based models. Specifically, we compared MPM with the two-stage part-based model designed by Sitawarin et al. [56]. In addition, we performed ablations on the number of part supervisions. MPM incorporates part supervision with three resolutions, referred to as SV3. Here we compared it with SV1 and SV2, which indicate part-based models solely supervised by $f_{l}$ and both $f_{l}$ and $f_{l-1}$ , respectively. Besides these, we extra performed an ablation by removing all top-down (TD) connections of MPM. The results are shown in Tab. 6. We can see that although the two-stage model needs about twice the parameters during inference, its robustness is inferior to the MPM. And all the designs of MPM, including utilizing higher resolution part annotations and top-down connections, contribute to the robustness. + +Additional ablations. Additional ablation experiments emphasize the significance of fine-grained part annotations rather than object segmentation annotations. Experiments on Tiny-ImageNet [31] highlight the significance of investigating adversarial robustness on high-resolution images. Moreover, the adversarial robustness of part-based models exhibits low sensitivity to $\lambda$ . Refer to Appendix K for comprehensive information about these ablation studies. + +# 6 Conclusion and Discussion + +In this work, we introduce PIN++, a new large-scale high-quality part dataset, to facilitate the research on part-based models for robust recognition. We propose a new MPM to leverage this dataset. With extensive experiments, we demonstrate the robustness of MPM across diverse scenarios. But we note that MPM represents just a stepping stone in the realm of part-based models for robustness. We hope that this work will inspire and motivate researchers to further explore the potential of part-based models for robust recognition. + +Limitation. We believe that except for robust recognition, $\mathrm{PIN}++$ has the potential to boost general part-related visual understanding tasks. We have not conducted many benchmark experiments specifically for such tasks because they are beyond the scope of this work. We leave these as future work. + +# Acknowledgement + +This work was supported in part by the National Key Research and Development Program of China (No. 2021ZD0200301) and the National Natural Science Foundation of China (No. U2341228). + +# References + +1. Agarap, A.F.: Deep learning using rectified linear units (relu). arXiv preprint arXiv:1803.08375 (2018) +2. Athalye, A., Carlini, N., Wagner, D.A.: Obfuscated gradients give a false sense of security: Circumventing defenses to adversarial examples. In: Int. Conf. Mach. Learn. (ICML). vol. 80, pp. 274-283 (2018) +3. Bai, Y., Mei, J., Yuille, A.L., Xie, C.: Are transformers more robust than cnns? In: Adv. Neural Inform. Process. Syst. (NeurIPS). pp. 26831-26843 (2021) +4. Biederman, I.: Recognition-by-components: a theory of human image understanding. Psychological Review 94(2), 115 (1987) +5. Burl, M.C., Weber, M., Perona, P.: A probabilistic approach to object recognition using local photometry and global geometry. In: Eur. Conf. Comput. Vis. (ECCV). vol. 1407, pp. 628-641 (1998) +6. Chen, X., Mottaghi, R., Liu, X., Fidler, S., Urtasun, R., Yuille, A.L.: Detect what you can: Detecting and representing objects using holistic models and body parts. In: IEEE Conf. Comput. Vis. Pattern Recog. (CVPR). pp. 1979-1986 (2014) +7. Croce, F., Andriushchenko, M., Sehwag, V., Debenedetti, E., Flammarion, N., Chiang, M., Mittal, P., Hein, M.: Robustbench: a standardized adversarial robustness benchmark. In: NeurIPS Datasets and Benchmarks (2021) +8. Croce, F., Hein, M.: Reliable evaluation of adversarial robustness with an ensemble of diverse parameter-free attacks. In: Int. Conf. Mach. Learn. (ICML). Proceedings of Machine Learning Research, vol. 119, pp. 2206-2216 (2020) +9. Croce, F., Singh, N.D., Hein, M.: Robust semantic segmentation: Strong adversarial attacks and fast training of robust models. Adv. Neural Inform. Process. Syst. (NeurIPS) (2023) +10. Debenedetti, E., Sehwag, V., Mittal, P.: A light recipe to train robust vision transformers. In: First IEEE Conference on Secure and Trustworthy Machine Learning (SaTML). pp. 225-253 (2023) +11. Deng, J., Dong, W., Socher, R., Li, L., Li, K., Fei-Fei, L.: Imagenet: A large-scale hierarchical image database. In: IEEE Conf. Comput. Vis. Pattern Recog. (CVPR). pp. 248-255 (2009) +12. Eykholt, K., Evtimov, I., Fernandes, E., Li, B., Rahmati, A., Xiao, C., Prakash, A., Kohno, T., Song, D.: Robust physical-world attacks on deep learning visual classification. In: IEEE Conf. Comput. Vis. Pattern Recog. (CVPR). pp. 1625-1634 (2018) +13. Fang, Y., Yang, S., Wang, X., Li, Y., Fang, C., Shan, Y., Feng, B., Liu, W.: Instances as queries. In: Int. Conf. Comput. Vis. (ICCV). pp. 6890-6899 (2021) +14. Fischler, M.A., Elschlager, R.A.: The representation and matching of pictorial structures. IEEE Trans. Comput. 22(1), 67-92 (1973) +15. Geirhos, R., Narayanappa, K., Mitzkus, B., Thieringer, T., Bethge, M., Wichmann, F.A., Brendel, W.: Partial success in closing the gap between human and machine vision. In: Adv. Neural Inform. Process. Syst. (NeurIPS). pp. 23885-23899 (2021) + +16. Geirhos, R., Rubisch, P., Michaelis, C., Bethge, M., Wichmann, F.A., Brendel, W.: Imagenet-trained cnns are biased towards texture; increasing shape bias improves accuracy and robustness. In: Int. Conf. Learn. Rep. (ICLR) (2019) +17. Girshick, R.B., Iandola, F.N., Darrell, T., Malik, J.: Deformable part models are convolutional neural networks. In: IEEE Conf. Comput. Vis. Pattern Recog. (CVPR). pp. 437-446 (2015) +18. Gong, K., Liang, X., Li, Y., Chen, Y., Yang, M., Lin, L.: Instance-level human parsing via part grouping network. In: Eur. Conf. Comput. Vis. (ECCV). vol. 11208, pp. 805-822 (2018) +19. Goodfellow, I.J., Shlens, J., Szegedy, C.: Explaining and harnessing adversarial examples. In: Int. Conf. Learn. Rep. (ICLR) (2015) +20. Haaf, R.A., Fulkerson, A.L., Jablonski, B.J., Hupp, J.M., Shull, S.S., Pescara-Kovach, L.: Object recognition and attention to object components by preschool children and 4-month-old infants. Journal of Experimental Child Psychology 86(2), 108-123 (2003) +21. He, J., Yang, S., Yang, S., Kortylewski, A., Yuan, X., Chen, J., Liu, S., Yang, C., Yu, Q., Yuille, A.L.: Partimagenet: A large, high-quality dataset of parts. In: Eur. Conf. Comput. Vis. (ECCV). vol. 13668, pp. 128-145 (2022) +22. He, K., Gkioxari, G., Dollár, P., Girshick, R.B.: Mask R-CNN. IEEE Trans. Pattern Anal. Mach. Intell. (TPAMI) 42(2), 386-397 (2020) +23. He, K., Zhang, X., Ren, S., Sun, J.: Deep residual learning for image recognition. In: IEEE Conf. Comput. Vis. Pattern Recog. (CVPR). pp. 770-778 (2016) +24. Hendrycks, D., Dietterich, T.G.: Benchmarking neural network robustness to common corruptions and perturbations. In: Int. Conf. Learn. Rep. (ICLR) (2019) +25. Hendrycks, D., Gimpel, K.: Gaussian error linear units (gelus). arXiv preprint arXiv:1606.08415 (2016) +26. Hoffman, J.E., Subramaniam, B.: The role of visual attention in saccadic eye movements. Perception & Psychophysics 57(6), 787-795 (1995) +27. Ilyas, A., Santurkar, S., Tsipras, D., Engstrom, L., Tran, B., Madry, A.: Adversarial examples are not bugs, they are features. In: Adv. Neural Inform. Process. Syst. (NeurIPS). pp. 125-136 (2019) +28. Kirillov, A., Mintun, E., Ravi, N., Mao, H., Rolland, C., Gustafson, L., Xiao, T., Whitehead, S., Berg, A.C., Lo, W.Y., Dollar, P., Girshick, R.: Segment anything. In: Int. Conf. Comput. Vis. (ICCV). pp. 4015-4026 (2023) +29. Krizhevsky, A., Hinton, G., et al.: Learning multiple layers of features from tiny images (2009) +30. Krizhevsky, A., Sutskever, I., Hinton, G.E.: Imagenet classification with deep convolutional neural networks. In: Adv. Neural Inform. Process. Syst. (NeurIPS). pp. 1106-1114 (2012) +31. Le, Y., Yang, X.: Tiny imagenet visual recognition challenge. CS 231N 7(7), 3 (2015) +32. Li, L.H., Zhang, P., Zhang, H., Yang, J., Li, C., Zhong, Y., Wang, L., Yuan, L., Zhang, L., Hwang, J., Chang, K., Gao, J.: Grounded language-image pre-training. In: IEEE Conf. Comput. Vis. Pattern Recog. (CVPR). pp. 10955-10965 (2022) +33. Li, X., Chen, H., Hu, X.: On the importance of backbone to the adversarial robustness of object detectors. arXiv preprint arXiv:2305.17438 (2023) +34. Li, X., Li, J., Dai, T., Shi, J., Zhu, J., Hu, X.: Rethinking natural adversarial examples for classification models. arXiv preprint arXiv:2102.11731 (2021) +35. Li, X., Wang, Z., Zhang, B., Sun, F., Hu, X.: Recognizing object by components with human prior knowledge enhances adversarial robustness of deep neural networks. IEEE Trans. Pattern Anal. Mach. Intell. (TPAMI) 45(7), 8861-8873 (2023) + +36. Li, X., Zhang, W., Liu, Y., Hu, Z., Zhang, B., Hu, X.: Language-driven anchors for zero-shot adversarial robustness. In: IEEE Conf. Comput. Vis. Pattern Recog. (CVPR) (2024) +37. Liang, X., Gong, K., Shen, X., Lin, L.: Look into person: Joint body parsing & pose estimation network and a new benchmark. IEEE Trans. Pattern Anal. Mach. Intell. (TPAMI) 41(4), 871-885 (2019) +38. Liang, X., Liu, S., Shen, X., Yang, J., Liu, L., Dong, J., Lin, L., Yan, S.: Deep human parsing with active template regression. IEEE Trans. Pattern Anal. Mach. Intell. (TPAMI) 37(12), 2402-2414 (2015) +39. Lin, C., Sun, P., Jiang, Y., Luo, P., Qu, L., Haffari, G., Yuan, Z., Cai, J.: Learning object-language alignments for open-vocabulary object detection. In: Int. Conf. Learn. Rep. (ICLR) (2023) +40. Lin, T., Dollar, P., Girshick, R.B., He, K., Hariharan, B., Belongie, S.J.: Feature pyramid networks for object detection. In: IEEE Conf. Comput. Vis. Pattern Recog. (CVPR). pp. 936-944 (2017) +41. Liu, C., Dong, Y., Xiang, W., Yang, X., Su, H., Zhu, J., Chen, Y., He, Y., Xue, H., Zheng, S.: A comprehensive study on robustness of image classification models: Benchmarking and rethinking. arXiv preprint arXiv:2302.14301 (2023) +42. Liu, Q., Kortylewski, A., Bai, Y., Bai, S., Yuille, A.L.: Intriguing properties of text-guided diffusion models. arXiv preprint arXiv:2306.00974 (2023) +43. Liu, S., Zhang, L., Yang, X., Su, H., Zhu, J.: Unsupervised part segmentation through disentangling appearance and shape. In: IEEE Conf. Comput. Vis. Pattern Recog. (CVPR). pp. 8355-8364 (2021) +44. Liu, Z., Lin, Y., Cao, Y., Hu, H., Wei, Y., Zhang, Z., Lin, S., Guo, B.: Swin transformer: Hierarchical vision transformer using shifted windows. In: Int. Conf. Comput. Vis. (ICCV). pp. 9992-10002 (2021) +45. Liu, Z., Mao, H., Wu, C., Feichtenhofer, C., Darrell, T., Xie, S.: A convnet for the 2020s. In: IEEE Conf. Comput. Vis. Pattern Recog. (CVPR). pp. 11966-11976 (2022) +46. Madry, A., Makelov, A., Schmidt, L., Tsipras, D., Vladu, A.: Towards deep learning models resistant to adversarial attacks. In: Int. Conf. Learn. Rep. (ICLR) (2018) +47. Mao, X., Chen, Y., Li, X., Qi, G., Duan, R., Zhang, R., Xue, H.: Easyrobust: A comprehensive and easy-to-use toolkit for robust computer vision. https://github.com/alibaba/easyrobust (2022) +48. Meletis, P., Wen, X., Lu, C., de Geus, D., Dubbelman, G.: Cityscapes-panoptic-parts and pascal-panoptic-parts datasets for scene understanding. arXiv preprint arXiv:2004.07944 (2020) +49. Ramanathan, V., Kalia, A., Petrovic, V., Wen, Y., Zheng, B., Guo, B., Wang, R., Marquez, A., Kovvuri, R., Kadian, A., Mousavi, A., Song, Y., Dubey, A., Mahajan, D.: PACO: parts and attributes of common objects. In: IEEE Conf. Comput. Vis. Pattern Recog. (CVPR). pp. 7141-7151 (2023) +50. Reddy, N.D., Vo, M., Narasimhan, S.G.: Carfusion: Combining point tracking and part detection for dynamic 3d reconstruction of vehicles. In: IEEE Conf. Comput. Vis. Pattern Recog. (CVPR). pp. 1906-1915 (2018) +51. Ren, S., He, K., Girshick, R.B., Sun, J.: Faster R-CNN: towards real-time object detection with region proposal networks. In: Adv. Neural Inform. Process. Syst. (NeurIPS). pp. 91-99 (2015) +52. Rombach, R., Blattmann, A., Lorenz, D., Esser, P., Ommer, B.: High-resolution image synthesis with latent diffusion models. In: IEEE Conf. Comput. Vis. Pattern Recog. (CVPR). pp. 10674-10685 (2022) + +53. Rosch, E., Mervis, C.B.: Family resemblances: Studies in the internal structure of categories. Cognitive Psychology 7(4), 573-605 (1975) +54. Salman, H., Ilyas, A., Engstrom, L., Kapoor, A., Madry, A.: Do adversarially robust imagenet models transfer better? In: Adv. Neural Inform. Process. Syst. (NeurIPS) (2020) +55. Salman, H., Sun, M., Yang, G., Kapoor, A., Kolter, J.Z.: Denoised smoothing: A provable defense for pretrained classifiers. In: Adv. Neural Inform. Process. Syst. (NeurIPS) (2020) +56. Sitawarin, C., Pongmala, K., Chen, Y., Carlini, N., Wagner, D.A.: Part-based models improve adversarial robustness. In: Int. Conf. Learn. Rep. (ICLR) (2023) +57. Song, X., Wang, P., Zhou, D., Zhu, R., Guan, C., Dai, Y., Su, H., Li, H., Yang, R.: Apollocar3d: A large 3d car instance understanding benchmark for autonomous driving. In: IEEE Conf. Comput. Vis. Pattern Recog. (CVPR). pp. 5452-5462 (2019) +58. Sun, P., Chen, S., Zhu, C., Xiao, F., Luo, P., Xie, S., Yan, Z.: Going denser with open-vocabulary part segmentation. arXiv preprint arXiv:2305.11173 (2023) +59. Szegedy, C., Zaremba, W., Sutskever, I., Bruna, J., Erhan, D., Goodfellow, I.J., Fergus, R.: Intriguing properties of neural networks. In: Int. Conf. Learn. Rep. (ICLR) (2014) +60. Tramèr, F., Carlini, N., Brendel, W., Madry, A.: On adaptive attacks to adversarial example defenses. In: Adv. Neural Inform. Process. Syst. (NeurIPS) (2020) +61. Tversky, B., Hemenway, K.: Objects, parts, and categories. Journal of Experimental Psychology: General 113(2), 169 (1984) +62. Wang, H., Ge, S., Lipton, Z.C., Xing, E.P.: Learning robust global representations by penalizing local predictive power. In: Adv. Neural Inform. Process. Syst. (NeurIPS). pp. 10506-10518 (2019) +63. Wang, J., Zhang, Z., Xie, C., Premachandran, V., Yuille, A.: Unsupervised learning of object semantic parts from internal states of cnns by population encoding. arXiv preprint arXiv:1511.06855 (2015) +64. Wu, D., Xia, S., Wang, Y.: Adversarial weight perturbation helps robust generalization. In: Adv. Neural Inform. Process. Syst. (NeurIPS) (2020) +65. Zhang, H., Yu, Y., Jiao, J., Xing, E.P., Ghaoui, L.E., Jordan, M.I.: Theoretically principled trade-off between robustness and accuracy. In: Int. Conf. Mach. Learn. (ICML). vol. 97, pp. 7472-7482 (2019) +66. Zhang, Y., Yao, Y., Jia, J., Yi, J., Hong, M., Chang, S., Liu, S.: How to robustify black-box ML models? A zeroth-order optimization perspective. In: Int. Conf. Learn. Rep. (ICLR) (2022) +67. Zhao, J., Li, J., Cheng, Y., Sim, T., Yan, S., Feng, J.: Understanding humans in crowded scenes: Deep nested adversarial learning and A new benchmark for multi-human parsing. In: ACM Int. Conf. Multimedia (ACM MM). pp. 792-800 (2018) +68. Zhao, Y., Li, J., Zhang, Y., Tian, Y.: Multi-class part parsing with joint boundary-semantic awareness. In: Int. Conf. Comput. Vis. (ICCV). pp. 9176-9185 (2019) +69. Zhou, B., Zhao, H., Puig, X., Xiao, T., Fidler, S., Barriuso, A., Torralba, A.: Semantic understanding of scenes through the ADE20K dataset. Int. J. Comput. Vis. (IJCV) 127(3), 302-321 (2019) +70. Zhou, Z., Firestone, C.: Humans can decipher adversarial images. Nature Communications 10(1), 1-9 (2019) \ No newline at end of file diff --git a/2024/PartImageNet++ Dataset_ Scaling up Part-based Models for Robust Recognition/images.zip b/2024/PartImageNet++ Dataset_ Scaling up Part-based Models for Robust Recognition/images.zip new file mode 100644 index 0000000000000000000000000000000000000000..e29bf91d0f90f26506e788898253436ed3ba45f4 --- /dev/null +++ b/2024/PartImageNet++ Dataset_ Scaling up Part-based Models for Robust Recognition/images.zip @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:a08678f824bb82f5fa60e111abc2f0d88ba1b7c27cd935c453a420c9f0ed6752 +size 335992 diff --git a/2024/PartImageNet++ Dataset_ Scaling up Part-based Models for Robust Recognition/layout.json b/2024/PartImageNet++ Dataset_ Scaling up Part-based Models for Robust Recognition/layout.json new file mode 100644 index 0000000000000000000000000000000000000000..2483e5860516c7fc7a0c4bb8fcafe307ae15dd90 --- /dev/null +++ b/2024/PartImageNet++ Dataset_ Scaling up Part-based Models for Robust Recognition/layout.json @@ -0,0 +1,9655 @@ +{ + "pdf_info": [ + { + "para_blocks": [ + { + "bbox": [ + 134, + 111, + 479, + 148 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 134, + 111, + 479, + 148 + ], + "spans": [ + { + "bbox": [ + 134, + 111, + 479, + 148 + ], + "type": "text", + "content": "PartImageNet++ Dataset: Scaling up Part-based Models for Robust Recognition" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 154, + 167, + 459, + 181 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 154, + 167, + 459, + 181 + ], + "spans": [ + { + "bbox": [ + 154, + 167, + 459, + 181 + ], + "type": "text", + "content": "Xiao Li" + }, + { + "bbox": [ + 154, + 167, + 459, + 181 + ], + "type": "inline_equation", + "content": "^{1}" + }, + { + "bbox": [ + 154, + 167, + 459, + 181 + ], + "type": "text", + "content": ", Yining Liu" + }, + { + "bbox": [ + 154, + 167, + 459, + 181 + ], + "type": "inline_equation", + "content": "^{2}" + }, + { + "bbox": [ + 154, + 167, + 459, + 181 + ], + "type": "text", + "content": ", Na Dong" + }, + { + "bbox": [ + 154, + 167, + 459, + 181 + ], + "type": "inline_equation", + "content": "^{3}" + }, + { + "bbox": [ + 154, + 167, + 459, + 181 + ], + "type": "text", + "content": ", Sitian Qin" + }, + { + "bbox": [ + 154, + 167, + 459, + 181 + ], + "type": "inline_equation", + "content": "^{2}" + }, + { + "bbox": [ + 154, + 167, + 459, + 181 + ], + "type": "text", + "content": ", and Xiaolin Hu" + }, + { + "bbox": [ + 154, + 167, + 459, + 181 + ], + "type": "inline_equation", + "content": "^{1,4*}" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 136, + 190, + 477, + 278 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 136, + 190, + 477, + 278 + ], + "spans": [ + { + "bbox": [ + 136, + 190, + 477, + 278 + ], + "type": "inline_equation", + "content": "^{1}" + }, + { + "bbox": [ + 136, + 190, + 477, + 278 + ], + "type": "text", + "content": " Department of Computer Science and Technology, BNRist, Institute for Artificial Intelligence, IDG/McGovern Institute for Brain Research, Tsinghua Laboratory of Brain and Intelligence, Tsinghua University, Beijing, China \n" + }, + { + "bbox": [ + 136, + 190, + 477, + 278 + ], + "type": "inline_equation", + "content": "^{2}" + }, + { + "bbox": [ + 136, + 190, + 477, + 278 + ], + "type": "text", + "content": " Harbin Institute of Technology, Weihai, China \n" + }, + { + "bbox": [ + 136, + 190, + 477, + 278 + ], + "type": "inline_equation", + "content": "^{3}" + }, + { + "bbox": [ + 136, + 190, + 477, + 278 + ], + "type": "text", + "content": " Beijing Institute of Technology, Beijing, China \n" + }, + { + "bbox": [ + 136, + 190, + 477, + 278 + ], + "type": "inline_equation", + "content": "^{4}" + }, + { + "bbox": [ + 136, + 190, + 477, + 278 + ], + "type": "text", + "content": " Chinese Institute for Brain Research (CIBR), Beijing, China \nlixiao20@mails.tsinghua.edu.cn, 22S030184@stu.hit.edu.cn, 3220220911@bit.edu.cn, qinsitian@hitwh.edu.cn, xlu@mail.tsinghua.edu.cn" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 160, + 310, + 453, + 540 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 160, + 310, + 453, + 540 + ], + "spans": [ + { + "bbox": [ + 160, + 310, + 453, + 540 + ], + "type": "text", + "content": "Abstract. Deep learning-based object recognition systems can be easily fooled by various adversarial perturbations. One reason for the weak robustness may be that they do not have part-based inductive bias like the human recognition process. Motivated by this, several part-based recognition models have been proposed to improve the adversarial robustness of recognition. However, due to the lack of part annotations, the effectiveness of these methods is only validated on small-scale non-standard datasets. In this work, we propose PIN++, short for PartImageNet++, a dataset providing high-quality part segmentation annotations for all categories of ImageNet-1K (IN-1K). With these annotations, we build part-based methods directly on the standard IN-1K dataset for robust recognition. Different from previous two-stage part-based models, we propose a Multi-scale Part-supervised Model (MPM), to learn a robust representation with part annotations. Experiments show that MPM yielded better adversarial robustness on the large-scale IN-1K over strong baselines across various attack settings. Furthermore, MPM achieved improved robustness on common corruptions and several out-of-distribution datasets. The dataset, together with these results, enables and encourages researchers to explore the potential of part-based models in more real applications. The dataset and the code are available at https://github.com/LixiaoTHU/PartImageNetPP." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 159, + 551, + 453, + 572 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 159, + 551, + 453, + 572 + ], + "spans": [ + { + "bbox": [ + 159, + 551, + 453, + 572 + ], + "type": "text", + "content": "Keywords: Adversarial robustness " + }, + { + "bbox": [ + 159, + 551, + 453, + 572 + ], + "type": "inline_equation", + "content": "\\cdot" + }, + { + "bbox": [ + 159, + 551, + 453, + 572 + ], + "type": "text", + "content": " Part dataset " + }, + { + "bbox": [ + 159, + 551, + 453, + 572 + ], + "type": "inline_equation", + "content": "\\cdot" + }, + { + "bbox": [ + 159, + 551, + 453, + 572 + ], + "type": "text", + "content": " Robust object recognition " + }, + { + "bbox": [ + 159, + 551, + 453, + 572 + ], + "type": "inline_equation", + "content": "\\cdot" + }, + { + "bbox": [ + 159, + 551, + 453, + 572 + ], + "type": "text", + "content": " Part-based model" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 132, + 595, + 230, + 608 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 595, + 230, + 608 + ], + "spans": [ + { + "bbox": [ + 132, + 595, + 230, + 608 + ], + "type": "text", + "content": "1 Introduction" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 130, + 621, + 482, + 646 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 621, + 482, + 646 + ], + "spans": [ + { + "bbox": [ + 130, + 621, + 482, + 646 + ], + "type": "text", + "content": "Object recognition has achieved remarkable success due to the rise of Deep Neural Networks (DNNs) [23, 30, 44, 45]. However, DNNs often lack robustness and" + } + ] + } + ], + "index": 6 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 133, + 653, + 236, + 665 + ], + "type": "page_footnote", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 653, + 236, + 665 + ], + "spans": [ + { + "bbox": [ + 133, + 653, + 236, + 665 + ], + "type": "text", + "content": "* Corresponding author." + } + ] + } + ], + "index": 7 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 0 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 194, + 114, + 302, + 236 + ], + "blocks": [ + { + "bbox": [ + 194, + 114, + 302, + 236 + ], + "lines": [ + { + "bbox": [ + 194, + 114, + 302, + 236 + ], + "spans": [ + { + "bbox": [ + 194, + 114, + 302, + 236 + ], + "type": "image", + "image_path": "99d95f6b34d987bf2a3e8ecefab8c0dd433af43d20be3e12595fd1fc0b7cc4ca.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 130, + 244, + 482, + 278 + ], + "lines": [ + { + "bbox": [ + 130, + 244, + 482, + 278 + ], + "spans": [ + { + "bbox": [ + 130, + 244, + 482, + 278 + ], + "type": "text", + "content": "Fig. 1: Examples of annotated images in " + }, + { + "bbox": [ + 130, + 244, + 482, + 278 + ], + "type": "inline_equation", + "content": "\\mathrm{PIN}++" + }, + { + "bbox": [ + 130, + 244, + 482, + 278 + ], + "type": "text", + "content": ". High-quality part segmentation annotations are provided on all categories in IN-1K. The object names are shown on the top-right of each image. The part names are hidden here." + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "image_caption" + } + ], + "index": 2 + }, + { + "type": "image", + "bbox": [ + 303, + 114, + 362, + 236 + ], + "blocks": [ + { + "bbox": [ + 303, + 114, + 362, + 236 + ], + "lines": [ + { + "bbox": [ + 303, + 114, + 362, + 236 + ], + "spans": [ + { + "bbox": [ + 303, + 114, + 362, + 236 + ], + "type": "image", + "image_path": "5d7e5d39c16a6afbfcbe90e67c26bd757d3049b7ee7d081737dada980f114120.jpg" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_body" + } + ], + "index": 3 + }, + { + "type": "image", + "bbox": [ + 362, + 114, + 420, + 236 + ], + "blocks": [ + { + "bbox": [ + 362, + 114, + 420, + 236 + ], + "lines": [ + { + "bbox": [ + 362, + 114, + 420, + 236 + ], + "spans": [ + { + "bbox": [ + 362, + 114, + 420, + 236 + ], + "type": "image", + "image_path": "07f2dfb8820e93cc7699018c344d79c33212124cac03f85d618a2ffe5e9342b0.jpg" + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "image_body" + } + ], + "index": 4 + }, + { + "bbox": [ + 130, + 304, + 482, + 460 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 304, + 482, + 460 + ], + "spans": [ + { + "bbox": [ + 130, + 304, + 482, + 460 + ], + "type": "text", + "content": "can be easily deceived by adversarial examples [19, 59], common image corruptions [24], and various out-of-distribution (OOD) shifts [34, 62], which greatly hinder applications of DNNs in security-critical scenarios [12, 33]. One possible reason for the weakness may be that DNN-based recognition does not have a part-based inductive bias (i.e., recognition based on object parts) like the human recognition process, which is generally regarded as a robust system [70]. According to the well-known cognitive psychological theory recognition-by-components [4], humans prefer to recognize objects by decomposing objects into parts and taking into account the hierarchical representations and the spatial relationships of parts. This cognitive theory has been supported by numerous psychological evidence [20, 26, 53, 61]. However, several studies revealed that DNNs often exhibit a shortcut recognition behavior, e.g., relying on texture-based features [16] and even unintelligible features for humans [27]." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 130, + 460, + 482, + 568 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 460, + 482, + 568 + ], + "spans": [ + { + "bbox": [ + 130, + 460, + 482, + 568 + ], + "type": "text", + "content": "If DNNs can perform a human-aligned part-based recognition, both the adversarial and non-adversarial robustness of DNNs could be closer to humans. Motivated by this idea, two recent works [35, 56] investigated how to use additional part-level supervision to build up part-based (recognition) models, which make predictions by recognizing the object's parts in a bottom-up manner. Both works investigated two-stage part-based models, where the first stage involves part segmentation, followed by the classification stage based on the segmentation results. These works have shown the potential of part-based models in enhancing robustness against various perturbations." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 130, + 570, + 482, + 666 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 570, + 482, + 666 + ], + "spans": [ + { + "bbox": [ + 130, + 570, + 482, + 666 + ], + "type": "text", + "content": "Nevertheless, due to the lack of part annotations, the effectiveness of part-based models has only been preliminary validated by the two works [35, 56] on small-scale nonstandard datasets. The largest dataset for evaluation was PartImageNet (PIN) [21], consisting of about " + }, + { + "bbox": [ + 130, + 570, + 482, + 666 + ], + "type": "inline_equation", + "content": "24\\mathrm{K}" + }, + { + "bbox": [ + 130, + 570, + 482, + 666 + ], + "type": "text", + "content": " images of 158 categories selected from the large-scale ImageNet-1K (IN-1K) [11]. PIN has several potential issues. Firstly, it is quite small compared to the standard IN-1K (about " + }, + { + "bbox": [ + 130, + 570, + 482, + 666 + ], + "type": "inline_equation", + "content": "1.3\\mathrm{M}" + }, + { + "bbox": [ + 130, + 570, + 482, + 666 + ], + "type": "text", + "content": " images) commonly used in the domain of adversarial robustness [3, 54]. This limited scale may restrict the generalizability of the findings. Secondly, categories included in" + } + ] + } + ], + "index": 8 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "spans": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "type": "text", + "content": "2" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 166, + 91, + 200, + 100 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 166, + 91, + 200, + 100 + ], + "spans": [ + { + "bbox": [ + 166, + 91, + 200, + 100 + ], + "type": "text", + "content": "Li et al." + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 1 + }, + { + "para_blocks": [ + { + "bbox": [ + 130, + 116, + 482, + 248 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 116, + 482, + 248 + ], + "spans": [ + { + "bbox": [ + 130, + 116, + 482, + 248 + ], + "type": "text", + "content": "PIN are specifically chosen to facilitate part-related tasks (e.g., part segmentation [43] and object parsing [68]), with an emphasis on animal categories (118 out of 158). While this focus can be beneficial for studying part-related tasks themselves, it poses challenges in demonstrating the effectiveness of part-based models for general recognition tasks across a diverse range of objects. Thirdly, the annotation scheme of PIN is simple. It groups 158 categories into 11 super-categories and determines part annotations for each super-category, potentially failing to capture the characteristics of object categories within a super-category. Due to the three issues, the part-based models [35,56] trained on PIN cannot be directly compared with state-of-the-art (SOTA) methods trained on the standard IN-1K." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 130, + 255, + 482, + 423 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 255, + 482, + 423 + ], + "spans": [ + { + "bbox": [ + 130, + 255, + 482, + 423 + ], + "type": "text", + "content": "In this work, we introduce PartImageNet++ (PIN++, a new dataset providing high-quality part segmentation annotations for all categories of the widely used IN-1K [11], to facilitate the research on part-based models for robust recognition (classification). To create PIN++, we designed a detailed annotation scheme to ensure high-quality part annotations. With this scheme, we annotated 100 randomly selected images per category, resulting in a total of 100K images with part annotations (about 1/13 of IN-1K). To avoid bias, all the part annotations are created manually without using any auxiliary model. To the best of our knowledge, PIN++ provides large-scale high-quality part segmentation annotations for the most diverse range of object categories, including creatures, artifacts, rigid objects, nonrigid objects, etc, among existing part datasets. Fig. 1 showcases some examples of annotated images. We introduce this dataset in detail in Sec. 3 and show that existing techniques fail to obtain high-quality part segmentation results without such annotations." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 130, + 430, + 482, + 586 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 430, + 482, + 586 + ], + "spans": [ + { + "bbox": [ + 130, + 430, + 482, + 586 + ], + "type": "text", + "content": "By leveraging the part annotations of " + }, + { + "bbox": [ + 130, + 430, + 482, + 586 + ], + "type": "inline_equation", + "content": "\\mathrm{PIN}++" + }, + { + "bbox": [ + 130, + 430, + 482, + 586 + ], + "type": "text", + "content": ", we develop part-based methods for robust recognition directly on the standard IN-1K dataset. To achieve this, we first train a part segmentation network with the annotations of " + }, + { + "bbox": [ + 130, + 430, + 482, + 586 + ], + "type": "inline_equation", + "content": "\\mathrm{PIN}++" + }, + { + "bbox": [ + 130, + 430, + 482, + 586 + ], + "type": "text", + "content": " to obtain pseudo part labels for all unannotated images in IN-1K. We then propose a new Multi-scale Part-supervised recognition Model, termed MPM, to better exploit the part annotations. MPM is expected to learn a robust intermediate representation by adding auxiliary bypass layers to the vanilla recognition model (e.g., ResNet-50 [23]). The auxiliary bypass layers are supervised by the part annotations and the generated pseudo part labels. Without introducing any additional parameter or computation during inference, MPM can use high-resolution part annotations more adequately than two-stage part-based models [35, 56], which are constrained to use low-resolution part labels. Fig. 3 shows an overview of our whole method." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 130, + 594, + 482, + 666 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 594, + 482, + 666 + ], + "spans": [ + { + "bbox": [ + 130, + 594, + 482, + 666 + ], + "type": "text", + "content": "The experimental results demonstrate that MPM, combined with adversarial training (AT) [46], achieved better adversarial robustness on the large-scale IN-1K, outperforming strong AT baselines in both the seen and unseen adversarial threats. Furthermore, MPM exhibited improved robustness on common image corruptions [24] and several OOD datasets. In addition, MPM achieved higher alignments with human vision. Beyond object recognition, MPM is shown to" + } + ] + } + ], + "index": 5 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 235, + 91, + 447, + 102 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 235, + 91, + 447, + 102 + ], + "spans": [ + { + "bbox": [ + 235, + 91, + 447, + 102 + ], + "type": "text", + "content": "PartImageNet++ Dataset: Scaling up Part-based ..." + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 474, + 92, + 480, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 474, + 92, + 480, + 100 + ], + "spans": [ + { + "bbox": [ + 474, + 92, + 480, + 100 + ], + "type": "text", + "content": "3" + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 2 + }, + { + "para_blocks": [ + { + "bbox": [ + 132, + 116, + 480, + 140 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 116, + 480, + 140 + ], + "spans": [ + { + "bbox": [ + 132, + 116, + 480, + 140 + ], + "type": "text", + "content": "boost the adversarial robustness of downstream tasks such as object detection [51]. The main contributions of this work can be summarized as follows:" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 137, + 146, + 479, + 216 + ], + "type": "list", + "angle": 0, + "index": 5, + "blocks": [ + { + "bbox": [ + 137, + 146, + 479, + 180 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 137, + 146, + 479, + 180 + ], + "spans": [ + { + "bbox": [ + 137, + 146, + 479, + 180 + ], + "type": "text", + "content": "- We created PIN++, a new large-scale dataset with part annotations. This dataset can facilitate further research on part-based models for robust recognition and other part-related visual understanding tasks." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 138, + 182, + 479, + 216 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 182, + 479, + 216 + ], + "spans": [ + { + "bbox": [ + 138, + 182, + 479, + 216 + ], + "type": "text", + "content": "- We proposed MPM, a new part-supervised recognition model, to exploit the part annotations better, achieving superior robustness across various settings and benchmarks without introducing any extra inference cost." + } + ] + } + ], + "index": 4 + } + ], + "sub_type": "text" + }, + { + "bbox": [ + 132, + 235, + 236, + 247 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 235, + 236, + 247 + ], + "spans": [ + { + "bbox": [ + 132, + 235, + 236, + 247 + ], + "type": "text", + "content": "2 Related Work" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 132, + 258, + 481, + 390 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 258, + 481, + 390 + ], + "spans": [ + { + "bbox": [ + 132, + 258, + 481, + 390 + ], + "type": "text", + "content": "Adversarial robustness. Szegedy et al. [59] first revealed the vulnerability of DNNs to adversarial examples. Since then, numerous methods [55,66] have been proposed to enhance the adversarial robustness of DNNs. Among them, AT has emerged as the de facto paradigm for training adversially robust DNNs. Early AT works [64,65] primarily focused on small-scale, low-resolution datasets such as CIFAR-10 [29]. Recently, there has been a growing interest in investigating AT on IN-1K [11], as this dataset serves as a standard benchmark for evaluating SOTA computer vision techniques [23,44,45]. Several studies have demonstrated that AT models trained on IN-1K can transfer their adversarial robustness to downstream dense-prediction tasks and achieve zero-shot adversially robust recognition [9, 33, 36]." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 132, + 390, + 481, + 533 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 390, + 481, + 533 + ], + "spans": [ + { + "bbox": [ + 132, + 390, + 481, + 533 + ], + "type": "text", + "content": "Part datasets. The concept of parts holds significant importance in both human visual cognition [4] and computer vision domains. To conduct research on part-related visual tasks, the availability of part datasets is crucial. However, annotating object parts is challenging and expensive, and most part datasets are only limited to specific domains, such as cars [50, 57, 63] and humans [18, 37, 38, 67]. Furthermore, part datasets targeting common objects often have a limited number of object categories. Cityscapes PanopticParts (Cityscapes PP.) [48], PascalPart [6], and ADE20K provide part annotations for 5, 20, and 80 object categories, respectively. Although the recent PACO dataset [49] offers a large number of part annotations, it is still restricted to only 75 object categories. PIN [21], containing 158 categories from IN-1K [11], represents the dataset with the most object categories in terms of part annotations." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 132, + 534, + 481, + 665 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 534, + 481, + 665 + ], + "spans": [ + { + "bbox": [ + 132, + 534, + 481, + 665 + ], + "type": "text", + "content": "Part-based object recognition. Modeling objects in terms of parts has a long history in computer vision [5,14,17]. But with the rise of DNNs, parts have been rarely used as auxiliaries for end-to-end general object recognition [23,30,44,45]. Recently, two works [35, 56] revisited part-based recognition models from the robustness perspective and showed the potential to enhance adversarial robustness. Li et al. [35] proposed ROCK, a part-based model that first predicts all parts represented by segmentation results, and then utilizes a judgment block to give category predictions based on the predicted parts. Similarly, Sitawarin et al. [56] also investigated several types of two-stage part-based models, where the first stage involves part segmentation, followed by the recognition stage with a tiny classifier. However, as mentioned before, both works were only validated on" + } + ] + } + ], + "index": 9 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "spans": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "type": "text", + "content": "4" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 166, + 91, + 199, + 100 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 166, + 91, + 199, + 100 + ], + "spans": [ + { + "bbox": [ + 166, + 91, + 199, + 100 + ], + "type": "text", + "content": "Li et al." + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 3 + }, + { + "para_blocks": [ + { + "bbox": [ + 130, + 116, + 482, + 201 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 116, + 482, + 201 + ], + "spans": [ + { + "bbox": [ + 130, + 116, + 482, + 201 + ], + "type": "text", + "content": "small-scale part-friendly datasets for the lack of part annotations, e.g., Sitawarin et al. [56] performed classification tasks only on 11 super-categories of PIN [21]. Furthermore, the proposed two-stage part-based models inevitably introduced a lot of extra parameters during inference, while the utilization of high-resolution part annotations was limited, as discussed in Sec. 4.2. In addition, ROCK [35] contains a non-differentiable process, which could lead to a potential overestimation of adversarial robustness [2,60]. Our work will address all of these issues." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 131, + 217, + 251, + 230 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 131, + 217, + 251, + 230 + ], + "spans": [ + { + "bbox": [ + 131, + 217, + 251, + 230 + ], + "type": "text", + "content": "3 PIN++ Dataset" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 130, + 240, + 481, + 277 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 240, + 481, + 277 + ], + "spans": [ + { + "bbox": [ + 130, + 240, + 481, + 277 + ], + "type": "text", + "content": "We first present the details of how we built " + }, + { + "bbox": [ + 130, + 240, + 481, + 277 + ], + "type": "inline_equation", + "content": "\\mathrm{PIN}++" + }, + { + "bbox": [ + 130, + 240, + 481, + 277 + ], + "type": "text", + "content": " to ensure high-quality part annotations, followed by the statistics of this dataset. We then give a discussion on " + }, + { + "bbox": [ + 130, + 240, + 481, + 277 + ], + "type": "inline_equation", + "content": "\\mathrm{PIN}++" + }, + { + "bbox": [ + 130, + 240, + 481, + 277 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 131, + 293, + 261, + 306 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 131, + 293, + 261, + 306 + ], + "spans": [ + { + "bbox": [ + 131, + 293, + 261, + 306 + ], + "type": "text", + "content": "3.1 Annotation Scheme" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 130, + 312, + 482, + 443 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 312, + 482, + 443 + ], + "spans": [ + { + "bbox": [ + 130, + 312, + 482, + 443 + ], + "type": "text", + "content": "Data source. IN-1K [11], containing about 1.3M images with category labels in the training set, is one of the most widely used datasets for general object recognition. Our aim is to provide part annotations for part-based robust recognition on IN-1K, and thus the images of PIN++ directly reuse the training set of IN-1K. All the images conform to licensing for research purposes. However, due to cost constraints, it is impractical to provide part-level annotations for all training images in IN-1K. Instead, we decided to annotate 100 randomly selected images per category. Once part annotations are provided for all categories, modern supervised segmentation techniques [22] can be used to obtain pseudo part labels for the remaining unannotated images, as we will show in Sec. 4. Note that PIN++ does not provide any part annotations for the validation set of IN-1K." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 130, + 444, + 482, + 527 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 444, + 482, + 527 + ], + "spans": [ + { + "bbox": [ + 130, + 444, + 482, + 527 + ], + "type": "text", + "content": "While PIN [21] has provided some part annotations for 158 categories of IN-1K, their annotations do not fully meet our purpose and principles. To minimize annotation costs without compromising quality, we retained annotations of 90 categories from PIN as part of " + }, + { + "bbox": [ + 130, + 444, + 482, + 527 + ], + "type": "inline_equation", + "content": "\\mathrm{PIN}++" + }, + { + "bbox": [ + 130, + 444, + 482, + 527 + ], + "type": "text", + "content": ". The details are shown in Appendix A. We then show the annotation design for the remaining 910 categories of IN-1K. Annotation quality control. To ensure high-quality part annotations, three points should be taken into account:" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 130, + 527, + 482, + 647 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 527, + 482, + 647 + ], + "spans": [ + { + "bbox": [ + 130, + 527, + 482, + 647 + ], + "type": "text", + "content": "1) Deciding which parts to annotate per category. One of the key challenges in the part annotation task is the ambiguity of object part selection (e.g., how to annotate the parts of hammer). PIN [21] evades this problem and focuses primarily on annotating parts for animal categories (most quadrupeds can be regarded as consisting of four part categories: head, body, foot, tail). However, the remaining categories in IN-1K are highly diverse, making it challenging to decide which parts to annotate. To ensure a scientifically grounded division of parts for each category, we initially consult the Wikidata knowledge base5 to obtain part vocabularies for every object category (e.g., hammer consists of three parts according to Wikidata: handle, striker, and hammerhead). Additionally, for object" + } + ] + } + ], + "index": 8 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 235, + 91, + 447, + 103 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 235, + 91, + 447, + 103 + ], + "spans": [ + { + "bbox": [ + 235, + 91, + 447, + 103 + ], + "type": "text", + "content": "PartImageNet++ Dataset: Scaling up Part-based ..." + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 474, + 92, + 480, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 474, + 92, + 480, + 100 + ], + "spans": [ + { + "bbox": [ + 474, + 92, + 480, + 100 + ], + "type": "text", + "content": "5" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 133, + 653, + 387, + 666 + ], + "type": "page_footnote", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 653, + 387, + 666 + ], + "spans": [ + { + "bbox": [ + 133, + 653, + 387, + 666 + ], + "type": "text", + "content": "5 https://www.wikidata.org/wiki/Wikidata:Introduction" + } + ] + } + ], + "index": 9 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 4 + }, + { + "para_blocks": [ + { + "type": "table", + "bbox": [ + 151, + 146, + 462, + 231 + ], + "blocks": [ + { + "bbox": [ + 132, + 114, + 480, + 137 + ], + "lines": [ + { + "bbox": [ + 132, + 114, + 480, + 137 + ], + "spans": [ + { + "bbox": [ + 132, + 114, + 480, + 137 + ], + "type": "text", + "content": "Table 1: The numbers of annotated object categories, part categories, images and part masks of publicly available part datasets." + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 151, + 146, + 462, + 231 + ], + "lines": [ + { + "bbox": [ + 151, + 146, + 462, + 231 + ], + "spans": [ + { + "bbox": [ + 151, + 146, + 462, + 231 + ], + "type": "table", + "html": "
DatasetObject CategoryPart CategoryImagePart Mask
Cityscapes PP. [48]5233.5K100K
Pascal-Part [6]2019319K363.5K
ADE20K* [69]8056612.6K193.2K
PACO [49]7545676.7K641.4K
PIN [21]15860924K112K
PIN++10003308100K406.4K
", + "image_path": "081e7fd9d179973ffa17c4d5df5b73e81fb0fca4b45df7ee669ddd7ad5a77bd8.jpg" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "table_body" + }, + { + "bbox": [ + 145, + 232, + 479, + 255 + ], + "lines": [ + { + "bbox": [ + 145, + 232, + 479, + 255 + ], + "spans": [ + { + "bbox": [ + 145, + 232, + 479, + 255 + ], + "type": "text", + "content": "*: ADE20K is severely category-unbalanced and here we show statistics for categories with more than 10 annotated part masks." + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "table_footnote" + } + ], + "index": 3 + }, + { + "bbox": [ + 130, + 277, + 481, + 397 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 277, + 481, + 397 + ], + "spans": [ + { + "bbox": [ + 130, + 277, + 481, + 397 + ], + "type": "text", + "content": "categories without a clear definition of parts on Wikidata, we decide which parts to annotate by asking recruited volunteers which parts they think to be important to their cognition of the object. For example, according to the cognition of volunteers, for most quadruped categories, annotating head, body, foot, and tail is enough, while for ram (see Fig. 1 for one example), horn should be annotated as an extra part. For categories that are indeed difficult to decompose into parts, e.g., flatworm, we generalize the concept of \"parts\" and treat the foreground of an object as one part category. In this case, the part is the object itself. Other details on deciding the parts to annotate are further discussed in Appendix B. These principles ensure that all categories in IN-1K can have part annotations." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 130, + 398, + 481, + 529 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 398, + 481, + 529 + ], + "spans": [ + { + "bbox": [ + 130, + 398, + 481, + 529 + ], + "type": "text", + "content": "2) Designing the part segmentation principles. We design several part segmentation principles to guide the annotators and ensure high-quality part segmentation annotation. Firstly, the annotated part masks are required to be combined to cover the entire object and have no overlap, unless in the case of the second principle. Secondly, for some parts that indeed should overlap in semantic concepts (e.g., horn and head), the annotators should annotate the inclusion relation of parts (e.g., horn is included in head), and then these part masks can be annotated with overlap. Thirdly, to keep the consistency of part annotations among images of the same object category, following [21], the same annotator should annotate all 100 images of this object category. The other principles are shown in Appendix C." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 130, + 529, + 481, + 589 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 529, + 481, + 589 + ], + "spans": [ + { + "bbox": [ + 130, + 529, + 481, + 589 + ], + "type": "text", + "content": "3) Annotation quality inspection. Ten randomly selected images together with the annotation visualizations (see Fig. 1) are sent to inspectors to assess if the annotations for a specific object category meet the defined principles. If the annotations for any two images fail to meet these principles, the entire category's annotations are re-annotated until the requirements are satisfied." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 132, + 608, + 207, + 619 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 608, + 207, + 619 + ], + "spans": [ + { + "bbox": [ + 132, + 608, + 207, + 619 + ], + "type": "text", + "content": "3.2 Statistics" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 130, + 629, + 481, + 665 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 629, + 481, + 665 + ], + "spans": [ + { + "bbox": [ + 130, + 629, + 481, + 665 + ], + "type": "text", + "content": "We finally annotated 100 images per category for 910 categories in IN-1K, resulting in a total of 91K images with part annotations. Taking these data and the data of 90 reused categories of PIN together, " + }, + { + "bbox": [ + 130, + 629, + 481, + 665 + ], + "type": "inline_equation", + "content": "\\mathrm{PIN}++" + }, + { + "bbox": [ + 130, + 629, + 481, + 665 + ], + "type": "text", + "content": " for 1K categories is" + } + ] + } + ], + "index": 9 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "spans": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "type": "text", + "content": "6" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 166, + 91, + 200, + 100 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 166, + 91, + 200, + 100 + ], + "spans": [ + { + "bbox": [ + 166, + 91, + 200, + 100 + ], + "type": "text", + "content": "Li et al." + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 5 + }, + { + "para_blocks": [ + { + "bbox": [ + 130, + 116, + 479, + 175 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 116, + 479, + 175 + ], + "spans": [ + { + "bbox": [ + 130, + 116, + 479, + 175 + ], + "type": "text", + "content": "obtained. Following IN-1K, here we ensure PIN++ is class-balanced. PIN++ is split into train/val sets with a 9:1 ratio for each category. But note that all annotated images are part of the training set of IN-1K. Unless stated otherwise, they are all utilized for training in our following experiments and the evaluation is performed on IN-1K val set." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 130, + 177, + 479, + 224 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 177, + 479, + 224 + ], + "spans": [ + { + "bbox": [ + 130, + 177, + 479, + 224 + ], + "type": "text", + "content": "In summary, " + }, + { + "bbox": [ + 130, + 177, + 479, + 224 + ], + "type": "inline_equation", + "content": "\\mathrm{PIN}++" + }, + { + "bbox": [ + 130, + 177, + 479, + 224 + ], + "type": "text", + "content": " includes a total of 100K images with part annotations. These annotations cover 3,310 part categories for 1,000 object categories. A total of 406,364 part masks are annotated in " + }, + { + "bbox": [ + 130, + 177, + 479, + 224 + ], + "type": "inline_equation", + "content": "\\mathrm{PIN}++" + }, + { + "bbox": [ + 130, + 177, + 479, + 224 + ], + "type": "text", + "content": ". Further statistics of " + }, + { + "bbox": [ + 130, + 177, + 479, + 224 + ], + "type": "inline_equation", + "content": "\\mathrm{PIN}++" + }, + { + "bbox": [ + 130, + 177, + 479, + 224 + ], + "type": "text", + "content": " are shown in Appendix D." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 130, + 225, + 480, + 369 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 225, + 480, + 369 + ], + "spans": [ + { + "bbox": [ + 130, + 225, + 480, + 369 + ], + "type": "text", + "content": "Comparison with other part datasets. We compare PIN++ with publicly available part datasets, as shown in Tab. 1. PIN++ outperforms previous datasets in terms of the number of object and part categories, as well as the total number of images with annotated part masks. It also demonstrates the competitive number of part masks compared to the recent PACO dataset [49]. Notably, PIN++ offers part annotations for a diverse range of objects, including creatures, artifacts, rigid objects, and nonrigid objects, unlike PACO or PIN, which are primarily focused on common tools or animals, respectively. As shown in Appendix D, the annotation quality of PIN++ exhibits an advantage over that of PIN. Moreover, PIN++ is category-balanced and guarantees that each image contains only one foreground category. All these features build the unique advantage of PIN++ for studying object recognition tasks." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 132, + 390, + 271, + 402 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 390, + 271, + 402 + ], + "spans": [ + { + "bbox": [ + 132, + 390, + 271, + 402 + ], + "type": "text", + "content": "3.3 Discussion on PIN++" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 130, + 413, + 480, + 568 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 413, + 480, + 568 + ], + "spans": [ + { + "bbox": [ + 130, + 413, + 480, + 568 + ], + "type": "text", + "content": "Recent advancements in segmentation include CLIP-based open-vocabulary segmentation methods [32,39,58] and the SAM model [28]. We compared these techniques with the part annotations of PIN++. We used the largest version of SAM, and for open-vocabulary segmentation, we employed the VLPart model [58], which is trained with part annotations from several part datasets including Pascal-Part [6], PIN [21], and PACO [49]. Visual comparisons are presented in Fig. 2. We can see that VLPart only provides roughly reasonable semantics for part categories it has been trained on. It fails to accurately identify the semantics of parts from unseen objects. On the other hand, SAM seems to perform well in segmenting certain parts based on the object's edges, but it struggles when the part edges are less distinguishable, as shown in the segmentation results for camel (the third row in Fig. 2). In addition, note that all masks segmented by SAM are category-agnostic." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 130, + 570, + 479, + 665 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 570, + 479, + 665 + ], + "spans": [ + { + "bbox": [ + 130, + 570, + 479, + 665 + ], + "type": "text", + "content": "The above results highlight the challenges faced by existing techniques in achieving accurate part segmentation for diverse objects without training on PIN++. Additionally, current text-to-image generative models [52] also struggle with understanding part semantics [42]. However, supervised training on PIN++ yields good part segmentation results, as shown in Appendix E. Therefore, besides its usage in robust object recognition (as demonstrated next), we believe that PIN++ and its extensive collections of 3308 part categories (vocabularies) have the potential to enhance general part-related visual understanding tasks." + } + ] + } + ], + "index": 7 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 235, + 91, + 447, + 102 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 235, + 91, + 447, + 102 + ], + "spans": [ + { + "bbox": [ + 235, + 91, + 447, + 102 + ], + "type": "text", + "content": "PartImageNet++ Dataset: Scaling up Part-based ..." + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 474, + 92, + 480, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 474, + 92, + 480, + 100 + ], + "spans": [ + { + "bbox": [ + 474, + 92, + 480, + 100 + ], + "type": "text", + "content": "7" + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 6 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 159, + 116, + 455, + 287 + ], + "blocks": [ + { + "bbox": [ + 159, + 116, + 455, + 287 + ], + "lines": [ + { + "bbox": [ + 159, + 116, + 455, + 287 + ], + "spans": [ + { + "bbox": [ + 159, + 116, + 455, + 287 + ], + "type": "image", + "image_path": "e529d980f0f4400e03052c8750ac3ec7e18b9d69b90e6fac69a5eb40752d1a76.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 130, + 297, + 482, + 330 + ], + "lines": [ + { + "bbox": [ + 130, + 297, + 482, + 330 + ], + "spans": [ + { + "bbox": [ + 130, + 297, + 482, + 330 + ], + "type": "text", + "content": "Fig. 2: Comparison between part segmentation results of different methods and PIN++ annotations. Without training on PIN++, VLPart and SAM fail to segment objects into specific parts with accurate semantics." + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_caption" + } + ], + "index": 2 + }, + { + "bbox": [ + 132, + 352, + 343, + 366 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 352, + 343, + 366 + ], + "spans": [ + { + "bbox": [ + 132, + 352, + 343, + 366 + ], + "type": "text", + "content": "4 Part-based Recognition Method" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 130, + 376, + 482, + 437 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 376, + 482, + 437 + ], + "spans": [ + { + "bbox": [ + 130, + 376, + 482, + 437 + ], + "type": "text", + "content": "With the annotations of " + }, + { + "bbox": [ + 130, + 376, + 482, + 437 + ], + "type": "inline_equation", + "content": "\\mathrm{PIN}++" + }, + { + "bbox": [ + 130, + 376, + 482, + 437 + ], + "type": "text", + "content": ", we develop part-based methods directly on the standard IN-1K. Our part-based methods can be divided into two steps. We first obtain pseudo part labels for the remaining unlabeled images of IN-1K and then train the MPM with these part annotations and pseudo labels. The pipeline of our part-based methods is illustrated in Fig. 3." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 132, + 454, + 315, + 466 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 454, + 315, + 466 + ], + "spans": [ + { + "bbox": [ + 132, + 454, + 315, + 466 + ], + "type": "text", + "content": "4.1 Pseudo Part Label Generation" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 130, + 473, + 481, + 569 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 473, + 481, + 569 + ], + "spans": [ + { + "bbox": [ + 130, + 473, + 481, + 569 + ], + "type": "text", + "content": "Part segmentation model. Arbitrary instance segmentation models [13, 22] can be leveraged to perform part segmentation and generate pseudo part labels when trained with the part annotations from " + }, + { + "bbox": [ + 130, + 473, + 481, + 569 + ], + "type": "inline_equation", + "content": "\\mathrm{PIN}++" + }, + { + "bbox": [ + 130, + 473, + 481, + 569 + ], + "type": "text", + "content": ". Following Sun et al. [58], we use the classical Mask R-CNN [22] model with a vision transformer Swin-B [44] as the backbone network. In training the Mask R-CNN, we consider the masks for each part category as masks for individual object categories directly. During inference for obtaining pseudo part labels, we incorporate a simple post-processing operation alongside the regular inference procedure." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 130, + 570, + 482, + 666 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 570, + 482, + 666 + ], + "spans": [ + { + "bbox": [ + 130, + 570, + 482, + 666 + ], + "type": "text", + "content": "Post-processing. Given an image " + }, + { + "bbox": [ + 130, + 570, + 482, + 666 + ], + "type": "inline_equation", + "content": "\\mathbf{x} \\in \\mathbb{R}^{3 \\times H \\times W}" + }, + { + "bbox": [ + 130, + 570, + 482, + 666 + ], + "type": "text", + "content": " with object category " + }, + { + "bbox": [ + 130, + 570, + 482, + 666 + ], + "type": "inline_equation", + "content": "y_{c}" + }, + { + "bbox": [ + 130, + 570, + 482, + 666 + ], + "type": "text", + "content": ", where " + }, + { + "bbox": [ + 130, + 570, + 482, + 666 + ], + "type": "inline_equation", + "content": "H \\times W" + }, + { + "bbox": [ + 130, + 570, + 482, + 666 + ], + "type": "text", + "content": " represents the resolution, the regular inference procedure of the trained Mask R-CNN produces pseudo part labels in the form of " + }, + { + "bbox": [ + 130, + 570, + 482, + 666 + ], + "type": "inline_equation", + "content": "\\{(\\mathbf{M}_{\\mathbf{p}}, \\mathbf{v}_{\\mathbf{p}})\\}" + }, + { + "bbox": [ + 130, + 570, + 482, + 666 + ], + "type": "text", + "content": ", where " + }, + { + "bbox": [ + 130, + 570, + 482, + 666 + ], + "type": "inline_equation", + "content": "\\mathbf{M}_{\\mathbf{p}} \\in \\{0,1\\}^{H \\times W}" + }, + { + "bbox": [ + 130, + 570, + 482, + 666 + ], + "type": "text", + "content": " denotes the binary part mask, " + }, + { + "bbox": [ + 130, + 570, + 482, + 666 + ], + "type": "inline_equation", + "content": "\\mathbf{v}_{\\mathbf{p}} \\in [0,1]^{K}" + }, + { + "bbox": [ + 130, + 570, + 482, + 666 + ], + "type": "text", + "content": " denotes the output probability of part categories, and " + }, + { + "bbox": [ + 130, + 570, + 482, + 666 + ], + "type": "inline_equation", + "content": "K" + }, + { + "bbox": [ + 130, + 570, + 482, + 666 + ], + "type": "text", + "content": " denotes the number of part categories (3308 in our case). But note that the images that need inference are all in the training set of IN-1K. Their object category " + }, + { + "bbox": [ + 130, + 570, + 482, + 666 + ], + "type": "inline_equation", + "content": "y_{c}" + }, + { + "bbox": [ + 130, + 570, + 482, + 666 + ], + "type": "text", + "content": " is not utilized in the regular inference procedure. To leverage it, we apply a Category Filter (CF) operation" + } + ] + } + ], + "index": 8 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "spans": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "type": "text", + "content": "8" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 166, + 91, + 200, + 100 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 166, + 91, + 200, + 100 + ], + "spans": [ + { + "bbox": [ + 166, + 91, + 200, + 100 + ], + "type": "text", + "content": "Li et al." + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 7 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 133, + 117, + 297, + 254 + ], + "blocks": [ + { + "bbox": [ + 133, + 117, + 297, + 254 + ], + "lines": [ + { + "bbox": [ + 133, + 117, + 297, + 254 + ], + "spans": [ + { + "bbox": [ + 133, + 117, + 297, + 254 + ], + "type": "image", + "image_path": "0c28df4a0c5bf1f10a9da094faf07570f46ba4a83f395256ee4b8d3467d539fa.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 130, + 262, + 482, + 328 + ], + "lines": [ + { + "bbox": [ + 130, + 262, + 482, + 328 + ], + "spans": [ + { + "bbox": [ + 130, + 262, + 482, + 328 + ], + "type": "text", + "content": "Fig. 3: An overview of the generation of pseudo-labels and the structure of MPM. (a) A part segmentation model trained on " + }, + { + "bbox": [ + 130, + 262, + 482, + 328 + ], + "type": "inline_equation", + "content": "\\mathrm{PIN}++" + }, + { + "bbox": [ + 130, + 262, + 482, + 328 + ], + "type": "text", + "content": " and used to obtain pseudo-part labels for unannotated images. (b) MPM adds several auxiliary bypass layers to the vanilla recognition model for part segmentation supervision. MPM is trained by part annotations together with the pseudo part labels. During inference, the auxiliary layers are dropped, and the vanilla recognition model gives the final object category prediction." + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "image_caption" + } + ], + "index": 2 + }, + { + "type": "image", + "bbox": [ + 310, + 118, + 480, + 253 + ], + "blocks": [ + { + "bbox": [ + 310, + 118, + 480, + 253 + ], + "lines": [ + { + "bbox": [ + 310, + 118, + 480, + 253 + ], + "spans": [ + { + "bbox": [ + 310, + 118, + 480, + 253 + ], + "type": "image", + "image_path": "830f9567b4d8b79f78cf57d39ae85fce31b21491f9c33169dec4938f65f22068.jpg" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_body" + } + ], + "index": 3 + }, + { + "bbox": [ + 130, + 352, + 482, + 401 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 352, + 482, + 401 + ], + "spans": [ + { + "bbox": [ + 130, + 352, + 482, + 401 + ], + "type": "text", + "content": "to obtain " + }, + { + "bbox": [ + 130, + 352, + 482, + 401 + ], + "type": "inline_equation", + "content": "\\hat{\\mathbf{v}}_{\\mathbf{p}}" + }, + { + "bbox": [ + 130, + 352, + 482, + 401 + ], + "type": "text", + "content": ", which sets the probabilities of unrelated part categories in " + }, + { + "bbox": [ + 130, + 352, + 482, + 401 + ], + "type": "inline_equation", + "content": "\\mathbf{v}_{\\mathbf{p}}" + }, + { + "bbox": [ + 130, + 352, + 482, + 401 + ], + "type": "text", + "content": " to zero (e.g., for cat, except for cat:head, cat:body, etc, other part categories such as radio:button and bird:head are all excluded). The pseudo part category for " + }, + { + "bbox": [ + 130, + 352, + 482, + 401 + ], + "type": "inline_equation", + "content": "\\mathbf{M}_{\\mathbf{p}}" + }, + { + "bbox": [ + 130, + 352, + 482, + 401 + ], + "type": "text", + "content": " is then computed as " + }, + { + "bbox": [ + 130, + 352, + 482, + 401 + ], + "type": "inline_equation", + "content": "y_{p} = \\arg \\max \\hat{\\mathbf{v}}_{\\mathbf{p}}" + }, + { + "bbox": [ + 130, + 352, + 482, + 401 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 131, + 417, + 399, + 430 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 131, + 417, + 399, + 430 + ], + "spans": [ + { + "bbox": [ + 131, + 417, + 399, + 430 + ], + "type": "text", + "content": "4.2 Multi-scale Part-supervised Recognition Model" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 130, + 437, + 482, + 520 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 437, + 482, + 520 + ], + "spans": [ + { + "bbox": [ + 130, + 437, + 482, + 520 + ], + "type": "text", + "content": "Data preparation. After obtaining the pseudo labels for all remaining training images of IN-1K, we simply treat the pseudo labels and real part annotations equally because we find that the quality of the pseudo labels is roughly satisfactory, as shown in Appendix E. The part labels " + }, + { + "bbox": [ + 130, + 437, + 482, + 520 + ], + "type": "inline_equation", + "content": "\\{(\\mathbf{M_p},\\mathbf{y_p})\\}" + }, + { + "bbox": [ + 130, + 437, + 482, + 520 + ], + "type": "text", + "content": " for an image " + }, + { + "bbox": [ + 130, + 437, + 482, + 520 + ], + "type": "inline_equation", + "content": "\\mathbf{x}" + }, + { + "bbox": [ + 130, + 437, + 482, + 520 + ], + "type": "text", + "content": " are then converted to a single composite segmentation mask " + }, + { + "bbox": [ + 130, + 437, + 482, + 520 + ], + "type": "inline_equation", + "content": "\\mathbf{M} \\in \\{0,1\\}^{H \\times W \\times (K + 1)}" + }, + { + "bbox": [ + 130, + 437, + 482, + 520 + ], + "type": "text", + "content": " where the extra one channel indicates the background. Finally, the data samples for training part-based models are " + }, + { + "bbox": [ + 130, + 437, + 482, + 520 + ], + "type": "inline_equation", + "content": "\\{(\\mathbf{x},y_c,\\mathbf{M})\\}" + }, + { + "bbox": [ + 130, + 437, + 482, + 520 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 130, + 521, + 482, + 666 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 521, + 482, + 666 + ], + "spans": [ + { + "bbox": [ + 130, + 521, + 482, + 666 + ], + "type": "text", + "content": "Model design. Both previous works [35, 56] build part-based models in the two-stage way: a segmenter " + }, + { + "bbox": [ + 130, + 521, + 482, + 666 + ], + "type": "inline_equation", + "content": "\\mathcal{F}_{\\mathrm{seg}}: \\mathbb{R}^{3 \\times H \\times W} \\to \\mathbb{R}^{(K + 1) \\times h \\times w}" + }, + { + "bbox": [ + 130, + 521, + 482, + 666 + ], + "type": "text", + "content": " and an extra module " + }, + { + "bbox": [ + 130, + 521, + 482, + 666 + ], + "type": "inline_equation", + "content": "\\mathcal{F}_{\\mathrm{cls}}: \\mathbb{R}^{(K + 1) \\times h \\times w} \\to \\mathbb{R}^C" + }, + { + "bbox": [ + 130, + 521, + 482, + 666 + ], + "type": "text", + "content": " for classification, where " + }, + { + "bbox": [ + 130, + 521, + 482, + 666 + ], + "type": "inline_equation", + "content": "h \\times w" + }, + { + "bbox": [ + 130, + 521, + 482, + 666 + ], + "type": "text", + "content": " denotes the output resolution and " + }, + { + "bbox": [ + 130, + 521, + 482, + 666 + ], + "type": "inline_equation", + "content": "C" + }, + { + "bbox": [ + 130, + 521, + 482, + 666 + ], + "type": "text", + "content": " denotes the number of object category. " + }, + { + "bbox": [ + 130, + 521, + 482, + 666 + ], + "type": "inline_equation", + "content": "\\mathcal{F}_{\\mathrm{seg}}" + }, + { + "bbox": [ + 130, + 521, + 482, + 666 + ], + "type": "text", + "content": " is a backbone network (e.g., a ResNet-50 without classification head). Most popular backbones [23, 44, 45] are hierarchically composed of several blocks with down-sampling layers and thus they can be written as " + }, + { + "bbox": [ + 130, + 521, + 482, + 666 + ], + "type": "inline_equation", + "content": "\\mathcal{F}_{\\mathrm{seg}} := f_l \\circ \\dots \\circ f_1" + }, + { + "bbox": [ + 130, + 521, + 482, + 666 + ], + "type": "text", + "content": ", where " + }, + { + "bbox": [ + 130, + 521, + 482, + 666 + ], + "type": "inline_equation", + "content": "l" + }, + { + "bbox": [ + 130, + 521, + 482, + 666 + ], + "type": "text", + "content": " denotes the number of down-sampling layers and " + }, + { + "bbox": [ + 130, + 521, + 482, + 666 + ], + "type": "inline_equation", + "content": "f_{i(1 \\leq i \\leq l)}" + }, + { + "bbox": [ + 130, + 521, + 482, + 666 + ], + "type": "text", + "content": " denotes one block with the " + }, + { + "bbox": [ + 130, + 521, + 482, + 666 + ], + "type": "inline_equation", + "content": "i" + }, + { + "bbox": [ + 130, + 521, + 482, + 666 + ], + "type": "text", + "content": "th down-sampling layer. " + }, + { + "bbox": [ + 130, + 521, + 482, + 666 + ], + "type": "inline_equation", + "content": "\\mathcal{F}_{\\mathrm{cls}}" + }, + { + "bbox": [ + 130, + 521, + 482, + 666 + ], + "type": "text", + "content": " can be a tiny classifier [56] or a nondifferentiable judgment block [35] and its input is the part segmentation results " + }, + { + "bbox": [ + 130, + 521, + 482, + 666 + ], + "type": "inline_equation", + "content": "\\hat{\\mathbf{M}} = \\mathcal{F}_{\\mathrm{seg}}(\\mathbf{x})" + }, + { + "bbox": [ + 130, + 521, + 482, + 666 + ], + "type": "text", + "content": ". The overall model for recognition is " + }, + { + "bbox": [ + 130, + 521, + 482, + 666 + ], + "type": "inline_equation", + "content": "\\mathcal{F} := \\mathcal{F}_{\\mathrm{cls}} \\circ \\mathcal{F}_{\\mathrm{seg}}" + }, + { + "bbox": [ + 130, + 521, + 482, + 666 + ], + "type": "text", + "content": ". However, as " + }, + { + "bbox": [ + 130, + 521, + 482, + 666 + ], + "type": "inline_equation", + "content": "K" + }, + { + "bbox": [ + 130, + 521, + 482, + 666 + ], + "type": "text", + "content": " is significantly larger than " + }, + { + "bbox": [ + 130, + 521, + 482, + 666 + ], + "type": "inline_equation", + "content": "C" + }, + { + "bbox": [ + 130, + 521, + 482, + 666 + ], + "type": "text", + "content": ", using " + }, + { + "bbox": [ + 130, + 521, + 482, + 666 + ], + "type": "inline_equation", + "content": "\\hat{M}" + }, + { + "bbox": [ + 130, + 521, + 482, + 666 + ], + "type": "text", + "content": " as the intermediate results for" + } + ] + } + ], + "index": 8 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 235, + 91, + 447, + 102 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 235, + 91, + 447, + 102 + ], + "spans": [ + { + "bbox": [ + 235, + 91, + 447, + 102 + ], + "type": "text", + "content": "PartImageNet++ Dataset: Scaling up Part-based ..." + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 474, + 92, + 480, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 474, + 92, + 480, + 100 + ], + "spans": [ + { + "bbox": [ + 474, + 92, + 480, + 100 + ], + "type": "text", + "content": "9" + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 8 + }, + { + "para_blocks": [ + { + "bbox": [ + 130, + 116, + 479, + 186 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 116, + 479, + 186 + ], + "spans": [ + { + "bbox": [ + 130, + 116, + 479, + 186 + ], + "type": "text", + "content": "recognition inevitably introduce extra parameters and computation during inference. This situation gets worse with an increased number of part categories, as shown in Sec. 5.5. Moreover, these methods solely employ part supervision on the output of " + }, + { + "bbox": [ + 130, + 116, + 479, + 186 + ], + "type": "inline_equation", + "content": "f_{l}" + }, + { + "bbox": [ + 130, + 116, + 479, + 186 + ], + "type": "text", + "content": ". Without expanding " + }, + { + "bbox": [ + 130, + 116, + 479, + 186 + ], + "type": "inline_equation", + "content": "\\mathcal{F}_{\\mathrm{seg}}" + }, + { + "bbox": [ + 130, + 116, + 479, + 186 + ], + "type": "text", + "content": ", part annotations have to be downsampled to match the output resolution of " + }, + { + "bbox": [ + 130, + 116, + 479, + 186 + ], + "type": "inline_equation", + "content": "f_{l}" + }, + { + "bbox": [ + 130, + 116, + 479, + 186 + ], + "type": "text", + "content": ", potentially limiting the utilization of high-resolution part annotations." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 130, + 188, + 482, + 475 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 188, + 482, + 475 + ], + "spans": [ + { + "bbox": [ + 130, + 188, + 482, + 475 + ], + "type": "text", + "content": "Instead of these two-stage part-based models, MPM is expected to learn a robust intermediate representation by adding multi-scale bypass layers to the vanilla recognition network. In MPM, the model used for object recognition is the vanilla backbone network directly: " + }, + { + "bbox": [ + 130, + 188, + 482, + 475 + ], + "type": "inline_equation", + "content": "\\mathcal{F} \\coloneqq h_c \\circ f_l \\circ \\dots \\circ f_1" + }, + { + "bbox": [ + 130, + 188, + 482, + 475 + ], + "type": "text", + "content": ", where " + }, + { + "bbox": [ + 130, + 188, + 482, + 475 + ], + "type": "inline_equation", + "content": "h_c" + }, + { + "bbox": [ + 130, + 188, + 482, + 475 + ], + "type": "text", + "content": " denotes a vanilla classification head (usually a linear layer). MPM utilizes the part annotations to supervise the intermediate features of " + }, + { + "bbox": [ + 130, + 188, + 482, + 475 + ], + "type": "inline_equation", + "content": "\\mathcal{F}" + }, + { + "bbox": [ + 130, + 188, + 482, + 475 + ], + "type": "text", + "content": " by several bypass layers " + }, + { + "bbox": [ + 130, + 188, + 482, + 475 + ], + "type": "inline_equation", + "content": "g_i" + }, + { + "bbox": [ + 130, + 188, + 482, + 475 + ], + "type": "text", + "content": ": " + }, + { + "bbox": [ + 130, + 188, + 482, + 475 + ], + "type": "inline_equation", + "content": "\\tilde{\\mathbf{M}}_i = g_i \\circ f_i \\circ \\dots \\circ f_1(\\mathbf{x})" + }, + { + "bbox": [ + 130, + 188, + 482, + 475 + ], + "type": "text", + "content": ", where " + }, + { + "bbox": [ + 130, + 188, + 482, + 475 + ], + "type": "inline_equation", + "content": "i \\leq l" + }, + { + "bbox": [ + 130, + 188, + 482, + 475 + ], + "type": "text", + "content": ". In this way, the intermediate features of " + }, + { + "bbox": [ + 130, + 188, + 482, + 475 + ], + "type": "inline_equation", + "content": "\\mathcal{F}" + }, + { + "bbox": [ + 130, + 188, + 482, + 475 + ], + "type": "text", + "content": " can be seen as implicit representations of part segmentation results when " + }, + { + "bbox": [ + 130, + 188, + 482, + 475 + ], + "type": "inline_equation", + "content": "g_i" + }, + { + "bbox": [ + 130, + 188, + 482, + 475 + ], + "type": "text", + "content": " is simple enough. Note that here we use part supervision not only on the output of " + }, + { + "bbox": [ + 130, + 188, + 482, + 475 + ], + "type": "inline_equation", + "content": "f_l" + }, + { + "bbox": [ + 130, + 188, + 482, + 475 + ], + "type": "text", + "content": ", but also " + }, + { + "bbox": [ + 130, + 188, + 482, + 475 + ], + "type": "inline_equation", + "content": "f_{i(i < l)}" + }, + { + "bbox": [ + 130, + 188, + 482, + 475 + ], + "type": "text", + "content": ". As the output of lower layers has a larger resolution, it can be supervised by part annotations with higher resolution. But the outputs of " + }, + { + "bbox": [ + 130, + 188, + 482, + 475 + ], + "type": "inline_equation", + "content": "f_{i(i < l)}" + }, + { + "bbox": [ + 130, + 188, + 482, + 475 + ], + "type": "text", + "content": " are relatively low-level, which may not be enough to obtain high-level part segmentation results themselves. To balance this, only the outputs of the last three blocks " + }, + { + "bbox": [ + 130, + 188, + 482, + 475 + ], + "type": "inline_equation", + "content": "f_{i(l - 2 \\leq i \\leq l)}" + }, + { + "bbox": [ + 130, + 188, + 482, + 475 + ], + "type": "text", + "content": " are supervised by part mask " + }, + { + "bbox": [ + 130, + 188, + 482, + 475 + ], + "type": "inline_equation", + "content": "\\mathbf{M}" + }, + { + "bbox": [ + 130, + 188, + 482, + 475 + ], + "type": "text", + "content": " (e.g., for a ResNet-50 with input size " + }, + { + "bbox": [ + 130, + 188, + 482, + 475 + ], + "type": "inline_equation", + "content": "224 \\times 224" + }, + { + "bbox": [ + 130, + 188, + 482, + 475 + ], + "type": "text", + "content": ", the intermediate features of " + }, + { + "bbox": [ + 130, + 188, + 482, + 475 + ], + "type": "inline_equation", + "content": "7 \\times 7" + }, + { + "bbox": [ + 130, + 188, + 482, + 475 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 130, + 188, + 482, + 475 + ], + "type": "inline_equation", + "content": "14 \\times 14" + }, + { + "bbox": [ + 130, + 188, + 482, + 475 + ], + "type": "text", + "content": ", and " + }, + { + "bbox": [ + 130, + 188, + 482, + 475 + ], + "type": "inline_equation", + "content": "28 \\times 28" + }, + { + "bbox": [ + 130, + 188, + 482, + 475 + ], + "type": "text", + "content": " are supervised by corresponding down-sampled " + }, + { + "bbox": [ + 130, + 188, + 482, + 475 + ], + "type": "inline_equation", + "content": "\\mathbf{M}" + }, + { + "bbox": [ + 130, + 188, + 482, + 475 + ], + "type": "text", + "content": ", respectively). In addition, following the principle of FPN [40], some top-down layers are used to augment the low-level features with high-level features. Note that distinct from the conventional FPN that generally has comparable parameters with the backbone network [40], these bypass layers are extremely lightweight as our goal is to improve the recognition accuracy rather than the quality of part segmentation results. Fig. 3(b) shows the overall structure of MPM. During inference, these auxiliary layers are dropped, and the vanilla recognition model gives the final object category prediction." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 130, + 475, + 482, + 550 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 475, + 482, + 550 + ], + "spans": [ + { + "bbox": [ + 130, + 475, + 482, + 550 + ], + "type": "text", + "content": "Training objective. The overall loss for training MPM is: " + }, + { + "bbox": [ + 130, + 475, + 482, + 550 + ], + "type": "inline_equation", + "content": "L = L_{\\mathrm{cls}} + \\lambda \\cdot L_{\\mathrm{seg}}" + }, + { + "bbox": [ + 130, + 475, + 482, + 550 + ], + "type": "text", + "content": ", where " + }, + { + "bbox": [ + 130, + 475, + 482, + 550 + ], + "type": "inline_equation", + "content": "L_{\\mathrm{cls}}" + }, + { + "bbox": [ + 130, + 475, + 482, + 550 + ], + "type": "text", + "content": " denotes the vanilla loss for classification, " + }, + { + "bbox": [ + 130, + 475, + 482, + 550 + ], + "type": "inline_equation", + "content": "L_{\\mathrm{seg}}" + }, + { + "bbox": [ + 130, + 475, + 482, + 550 + ], + "type": "text", + "content": " denotes the loss for part segmentation, and " + }, + { + "bbox": [ + 130, + 475, + 482, + 550 + ], + "type": "inline_equation", + "content": "\\lambda" + }, + { + "bbox": [ + 130, + 475, + 482, + 550 + ], + "type": "text", + "content": " is a hyper-parameter. We compute " + }, + { + "bbox": [ + 130, + 475, + 482, + 550 + ], + "type": "inline_equation", + "content": "L_{\\mathrm{seg}}" + }, + { + "bbox": [ + 130, + 475, + 482, + 550 + ], + "type": "text", + "content": " as the average of losses on three part segmentation results of different resolutions, as mentioned before. Note that when performing AT, adversarial examples are generated by " + }, + { + "bbox": [ + 130, + 475, + 482, + 550 + ], + "type": "inline_equation", + "content": "\\mathbf{x}^{\\star} = \\arg \\max_{\\mathbf{x}^{\\star}:||\\mathbf{x}^{\\star} - \\mathbf{x}||_{p}\\leq \\epsilon}L_{\\mathrm{cls}}(\\mathcal{F}(\\mathbf{x}^{\\star}),y)" + }, + { + "bbox": [ + 130, + 475, + 482, + 550 + ], + "type": "text", + "content": ", while MPM is always trained with " + }, + { + "bbox": [ + 130, + 475, + 482, + 550 + ], + "type": "inline_equation", + "content": "L" + }, + { + "bbox": [ + 130, + 475, + 482, + 550 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 132, + 563, + 230, + 578 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 563, + 230, + 578 + ], + "spans": [ + { + "bbox": [ + 132, + 563, + 230, + 578 + ], + "type": "text", + "content": "5 Experiments" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 132, + 587, + 261, + 601 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 587, + 261, + 601 + ], + "spans": [ + { + "bbox": [ + 132, + 587, + 261, + 601 + ], + "type": "text", + "content": "5.1 Experimental Setup" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 130, + 605, + 482, + 666 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 605, + 482, + 666 + ], + "spans": [ + { + "bbox": [ + 130, + 605, + 482, + 666 + ], + "type": "text", + "content": "Training setup. Unless specified otherwise, following previous works [3,10], we performed AT with " + }, + { + "bbox": [ + 130, + 605, + 482, + 666 + ], + "type": "inline_equation", + "content": "l_{\\infty}" + }, + { + "bbox": [ + 130, + 605, + 482, + 666 + ], + "type": "text", + "content": " bound " + }, + { + "bbox": [ + 130, + 605, + 482, + 666 + ], + "type": "inline_equation", + "content": "\\epsilon = 4 / 255" + }, + { + "bbox": [ + 130, + 605, + 482, + 666 + ], + "type": "text", + "content": " on IN-1K. The inner optimization for obtaining adversarial examples used PGD [46] with iterative steps " + }, + { + "bbox": [ + 130, + 605, + 482, + 666 + ], + "type": "inline_equation", + "content": "t = 2" + }, + { + "bbox": [ + 130, + 605, + 482, + 666 + ], + "type": "text", + "content": " and the step size was set to be " + }, + { + "bbox": [ + 130, + 605, + 482, + 666 + ], + "type": "inline_equation", + "content": "s = 2*\\epsilon /t" + }, + { + "bbox": [ + 130, + 605, + 482, + 666 + ], + "type": "text", + "content": ". The input resolution used " + }, + { + "bbox": [ + 130, + 605, + 482, + 666 + ], + "type": "inline_equation", + "content": "224\\times 224" + }, + { + "bbox": [ + 130, + 605, + 482, + 666 + ], + "type": "text", + "content": ". We used the vanilla version of the AT [46] and did not incorporate other variants" + } + ] + } + ], + "index": 7 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "spans": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "type": "text", + "content": "10" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 166, + 91, + 199, + 100 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 166, + 91, + 199, + 100 + ], + "spans": [ + { + "bbox": [ + 166, + 91, + 199, + 100 + ], + "type": "text", + "content": "Li et al." + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 9 + }, + { + "para_blocks": [ + { + "type": "table", + "bbox": [ + 144, + 157, + 463, + 266 + ], + "blocks": [ + { + "bbox": [ + 132, + 114, + 482, + 149 + ], + "lines": [ + { + "bbox": [ + 132, + 114, + 482, + 149 + ], + "spans": [ + { + "bbox": [ + 132, + 114, + 482, + 149 + ], + "type": "text", + "content": "Table 2: Recognition accuracies (\\%) of methods under different attack threats on IN-1K, following the evaluation setup in Sec. 5.1. Average denotes the average accuracies in four attack settings. We highlight the best results in each column." + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 144, + 157, + 463, + 266 + ], + "lines": [ + { + "bbox": [ + 144, + 157, + 463, + 266 + ], + "spans": [ + { + "bbox": [ + 144, + 157, + 463, + 266 + ], + "type": "table", + "html": "
ArchitectureMethodAdversarial Train wrt l∞ (ε = 4/255)
Cleanl∞l∞(×2)l1l2Average
ResNet-50[54], 202063.935.913.22.013.716.2
ResNet-50[47], 202264.734.311.42.316.416.1
ResNet-50[41], 202365.532.69.53.317.915.8
ResNet-50-Gelu[3], 202167.836.610.83.016.816.8
ResNet-50-Gelu[10], 202366.835.612.83.616.517.1
ResNet-50-Geluours (vanilla)67.138.112.65.021.619.3
ResNet-50-Geluours (MPM)67.839.113.66.224.320.8
", + "image_path": "646998addc8cd6ce92e68f821450cd69bc4add88141085ae92cbfb1806eaf704.jpg" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "table_body" + } + ], + "index": 3 + }, + { + "bbox": [ + 130, + 284, + 482, + 416 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 284, + 482, + 416 + ], + "spans": [ + { + "bbox": [ + 130, + 284, + 482, + 416 + ], + "type": "text", + "content": "such as TRADES [65] or AWP [64]. These variants have shown effectiveness mainly on small-scale datasets, making their generalization to IN-1K non-trivial. Following previous part-based works [35,56], we used the classical ResNet-50 [23] network with 25.6M parameters as the baseline model. We note that recent works [3,10] used a GELU [25] activation to replace the original RELU [1] activation of ResNet-50 for boosting robustness when performing AT, here we also used GELU, denoted as ResNet-50-Gelu. In addition, we improved the training recipe of ResNet-50 to build a strong baseline. Other details on the training recipe are provided in Appendix F. Note that bypass layers (e.g., point-wise convolutions) with about 4.5M parameters were used by MPM during training but completely dropped during inference. The code is submitted along with the paper." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 130, + 416, + 482, + 534 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 416, + 482, + 534 + ], + "spans": [ + { + "bbox": [ + 130, + 416, + 482, + 534 + ], + "type": "text", + "content": "Evaluation setup. For all experiments in the work, unless specified otherwise, we evaluated the adversarial robustness with AutoAttack [8], which is a combination of various attack methods and is generally recognized as a reliable evaluation [7]. The evaluation was performed at a resolution " + }, + { + "bbox": [ + 130, + 416, + 482, + 534 + ], + "type": "inline_equation", + "content": "224 \\times 224" + }, + { + "bbox": [ + 130, + 416, + 482, + 534 + ], + "type": "text", + "content": " on randomly selected 10K images of the 50K IN-1K val set, 10 images for each category. This ensures that the " + }, + { + "bbox": [ + 130, + 416, + 482, + 534 + ], + "type": "inline_equation", + "content": "95\\%" + }, + { + "bbox": [ + 130, + 416, + 482, + 534 + ], + "type": "text", + "content": " confidence intervals of the reported average adversarial robustness (see Tab. 2) were less than " + }, + { + "bbox": [ + 130, + 416, + 482, + 534 + ], + "type": "inline_equation", + "content": "\\pm 0.5\\%" + }, + { + "bbox": [ + 130, + 416, + 482, + 534 + ], + "type": "text", + "content": ". Note that, unlike previous works [3,10], we did not use the 5K images selected by RobustBench [7] as we found that they were category-unbalanced, containing three categories even without any image for validation." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 130, + 536, + 482, + 609 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 536, + 482, + 609 + ], + "spans": [ + { + "bbox": [ + 130, + 536, + 482, + 609 + ], + "type": "text", + "content": "We expect the AT models with " + }, + { + "bbox": [ + 130, + 536, + 482, + 609 + ], + "type": "inline_equation", + "content": "l_{\\infty}" + }, + { + "bbox": [ + 130, + 536, + 482, + 609 + ], + "type": "text", + "content": " bound " + }, + { + "bbox": [ + 130, + 536, + 482, + 609 + ], + "type": "inline_equation", + "content": "\\epsilon = 4 / 255" + }, + { + "bbox": [ + 130, + 536, + 482, + 609 + ], + "type": "text", + "content": " to be adversarially robust not only on seen threats (i.e., " + }, + { + "bbox": [ + 130, + 536, + 482, + 609 + ], + "type": "inline_equation", + "content": "l_{\\infty}" + }, + { + "bbox": [ + 130, + 536, + 482, + 609 + ], + "type": "text", + "content": " attack with " + }, + { + "bbox": [ + 130, + 536, + 482, + 609 + ], + "type": "inline_equation", + "content": "\\epsilon = 4 / 255" + }, + { + "bbox": [ + 130, + 536, + 482, + 609 + ], + "type": "text", + "content": ") but also on unseen attack threats (e.g., " + }, + { + "bbox": [ + 130, + 536, + 482, + 609 + ], + "type": "inline_equation", + "content": "l_{\\infty}" + }, + { + "bbox": [ + 130, + 536, + 482, + 609 + ], + "type": "text", + "content": " attack with larger " + }, + { + "bbox": [ + 130, + 536, + 482, + 609 + ], + "type": "inline_equation", + "content": "\\epsilon, l_1" + }, + { + "bbox": [ + 130, + 536, + 482, + 609 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 130, + 536, + 482, + 609 + ], + "type": "inline_equation", + "content": "l_2" + }, + { + "bbox": [ + 130, + 536, + 482, + 609 + ], + "type": "text", + "content": " attacks). Thus, we evaluated the models on three attack threats: " + }, + { + "bbox": [ + 130, + 536, + 482, + 609 + ], + "type": "inline_equation", + "content": "l_{\\infty}" + }, + { + "bbox": [ + 130, + 536, + 482, + 609 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 130, + 536, + 482, + 609 + ], + "type": "inline_equation", + "content": "l_1" + }, + { + "bbox": [ + 130, + 536, + 482, + 609 + ], + "type": "text", + "content": ", and " + }, + { + "bbox": [ + 130, + 536, + 482, + 609 + ], + "type": "inline_equation", + "content": "l_2" + }, + { + "bbox": [ + 130, + 536, + 482, + 609 + ], + "type": "text", + "content": ", with the bounds " + }, + { + "bbox": [ + 130, + 536, + 482, + 609 + ], + "type": "inline_equation", + "content": "\\epsilon_{\\infty} = 4 / 255" + }, + { + "bbox": [ + 130, + 536, + 482, + 609 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 130, + 536, + 482, + 609 + ], + "type": "inline_equation", + "content": "\\epsilon_1 = 75" + }, + { + "bbox": [ + 130, + 536, + 482, + 609 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 130, + 536, + 482, + 609 + ], + "type": "inline_equation", + "content": "\\epsilon_2 = 2" + }, + { + "bbox": [ + 130, + 536, + 482, + 609 + ], + "type": "text", + "content": ", respectively. For " + }, + { + "bbox": [ + 130, + 536, + 482, + 609 + ], + "type": "inline_equation", + "content": "l_{\\infty}" + }, + { + "bbox": [ + 130, + 536, + 482, + 609 + ], + "type": "text", + "content": " attack, we extra evaluated the models with " + }, + { + "bbox": [ + 130, + 536, + 482, + 609 + ], + "type": "inline_equation", + "content": "\\epsilon_{\\infty} = 8 / 255" + }, + { + "bbox": [ + 130, + 536, + 482, + 609 + ], + "type": "text", + "content": " to mimic a stronger threat, denoted as " + }, + { + "bbox": [ + 130, + 536, + 482, + 609 + ], + "type": "inline_equation", + "content": "l_{\\infty}(\\times 2)" + }, + { + "bbox": [ + 130, + 536, + 482, + 609 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 132, + 623, + 360, + 635 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 623, + 360, + 635 + ], + "spans": [ + { + "bbox": [ + 132, + 623, + 360, + 635 + ], + "type": "text", + "content": "5.2 Robustness against Adversarial Attacks" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 130, + 641, + 481, + 666 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 641, + 481, + 666 + ], + "spans": [ + { + "bbox": [ + 130, + 641, + 481, + 666 + ], + "type": "text", + "content": "With the above setup, we trained the vanilla ResNet-50-Gelu as a baseline. MPM used the same recipe as the baseline model for a fair comparison, except that it" + } + ] + } + ], + "index": 8 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 235, + 91, + 447, + 103 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 235, + 91, + 447, + 103 + ], + "spans": [ + { + "bbox": [ + 235, + 91, + 447, + 103 + ], + "type": "text", + "content": "PartImageNet++ Dataset: Scaling up Part-based ..." + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 470, + 91, + 479, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 470, + 91, + 479, + 100 + ], + "spans": [ + { + "bbox": [ + 470, + 91, + 479, + 100 + ], + "type": "text", + "content": "11" + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 10 + }, + { + "para_blocks": [ + { + "type": "table", + "bbox": [ + 140, + 168, + 478, + 229 + ], + "blocks": [ + { + "bbox": [ + 132, + 114, + 482, + 159 + ], + "lines": [ + { + "bbox": [ + 132, + 114, + 482, + 159 + ], + "spans": [ + { + "bbox": [ + 132, + 114, + 482, + 159 + ], + "type": "text", + "content": "Table 3: Recognition accuracies " + }, + { + "bbox": [ + 132, + 114, + 482, + 159 + ], + "type": "inline_equation", + "content": "(\\%)" + }, + { + "bbox": [ + 132, + 114, + 482, + 159 + ], + "type": "text", + "content": " of ResNet-50-Gelu on 15 different common image corruptions [24]. We adopt accuracy as the metric to be consistent with other results, while this metric can be easily converted into the corruption error metric [24]. \"Average\" denotes the accuracy averaged over different common corruptions." + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 140, + 168, + 478, + 229 + ], + "lines": [ + { + "bbox": [ + 140, + 168, + 478, + 229 + ], + "spans": [ + { + "bbox": [ + 140, + 168, + 478, + 229 + ], + "type": "table", + "html": "
Part ATNoiseBlurWeatherDigitalAvg.
Gauss. Shot Impul.Defoc.Glass Moti.ZoomSnow FrostFogBrightCont.Elast.Pixel JPEG
31.230.727.339.028.639.537.835.840.753.869.038.037.047.157.040.8
✓31.531.028.441.929.840.336.737.041.355.369.639.636.952.959.942.1
✓29.628.521.023.332.932.834.434.234.010.858.49.552.258.863.134.9
✓30.129.222.425.235.135.036.436.436.211.959.710.453.159.864.336.4
", + "image_path": "732f12916ba7eefeed1a26d21a9cb4bbfde33bcc787c9f9a7e7442c70a5c3243.jpg" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "table_body" + } + ], + "index": 3 + }, + { + "type": "table", + "bbox": [ + 208, + 261, + 406, + 313 + ], + "blocks": [ + { + "bbox": [ + 162, + 240, + 451, + 251 + ], + "lines": [ + { + "bbox": [ + 162, + 240, + 451, + 251 + ], + "spans": [ + { + "bbox": [ + 162, + 240, + 451, + 251 + ], + "type": "text", + "content": "Table 4: Recognition accuracies (%) of models on four OOD datasets." + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 208, + 261, + 406, + 313 + ], + "lines": [ + { + "bbox": [ + 208, + 261, + 406, + 313 + ], + "spans": [ + { + "bbox": [ + 208, + 261, + 406, + 313 + ], + "type": "table", + "html": "
PartATIN-A-PlusIN-SketchSINDINAverage
6.925.86.953.523.3
✓7.425.77.453.923.6
✓5.325.112.555.424.6
✓✓5.726.412.655.925.2
", + "image_path": "923cc41b2d6caaffae796b79383c3eb722238c7b1a39b9b2a65f8df273629fad.jpg" + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "table_body" + } + ], + "index": 5 + }, + { + "bbox": [ + 130, + 337, + 482, + 468 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 337, + 482, + 468 + ], + "spans": [ + { + "bbox": [ + 130, + 337, + 482, + 468 + ], + "type": "text", + "content": "used an extra " + }, + { + "bbox": [ + 130, + 337, + 482, + 468 + ], + "type": "inline_equation", + "content": "L_{\\mathrm{seg}}" + }, + { + "bbox": [ + 130, + 337, + 482, + 468 + ], + "type": "text", + "content": " to introduce part-based supervision. " + }, + { + "bbox": [ + 130, + 337, + 482, + 468 + ], + "type": "inline_equation", + "content": "\\lambda" + }, + { + "bbox": [ + 130, + 337, + 482, + 468 + ], + "type": "text", + "content": " was simply set to 1. In addition, we compared MPM with all open-sourced ResNet-50 checkpoints adversarially trained on IN-1K in recent years. The recognition accuracies under different attack threats are shown in Tab. 2. We can see that our baseline model is strong enough, outperforming all other benchmarks by " + }, + { + "bbox": [ + 130, + 337, + 482, + 468 + ], + "type": "inline_equation", + "content": "2.2\\%" + }, + { + "bbox": [ + 130, + 337, + 482, + 468 + ], + "type": "text", + "content": " on average robustness (17.1% vs. 19.3%). MPM further shows significantly better adversarial robustness than the vanilla baseline across all these attack threats. In addition, compared with our strong vanilla baseline, the accuracy of clean images is also improved with MPM. To our best knowledge, these are SOTA results of adversarial robustness with ResNet-50 on IN-1K. We additionally show in Appendix G that these results are significantly better than previous results on PIN." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 132, + 487, + 397, + 500 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 487, + 397, + 500 + ], + "spans": [ + { + "bbox": [ + 132, + 487, + 397, + 500 + ], + "type": "text", + "content": "5.3 Robustness on Corruptions and OOD Datasets" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 130, + 510, + 482, + 582 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 510, + 482, + 582 + ], + "spans": [ + { + "bbox": [ + 130, + 510, + 482, + 582 + ], + "type": "text", + "content": "Except for robustness against adversarial attacks, which consider the worst-case security threat scenarios, we evaluated the robustness of our part-based model on common image corruptions [24] and several OOD datasets [15, 16, 34, 62], which could be more realistic threats. Considering that AT can have negative effects on these non-adversarial scenarios [35, 41], here we also evaluated models without AT. The training details are also shown in Appendix F." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 130, + 582, + 482, + 666 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 582, + 482, + 666 + ], + "spans": [ + { + "bbox": [ + 130, + 582, + 482, + 666 + ], + "type": "text", + "content": "For common image corruptions [24], we generated different types of corrupted images on the whole IN-1K val set. The results of different models on these images are shown in Tab. 3, where \"Part\" indicates MPM. Here the results of each corruption type are the accuracies averaged on five severity levels [24]. It is seen that MPM has better robustness on nearly all types of image corruptions than models without part supervision, regardless of AT. Besides common corruptions, we evaluated the models on four OOD datasets: ImageNet-A-Plus (IN-A-Plus)" + } + ] + } + ], + "index": 9 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "spans": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "type": "text", + "content": "12" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 166, + 91, + 199, + 100 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 166, + 91, + 199, + 100 + ], + "spans": [ + { + "bbox": [ + 166, + 91, + 199, + 100 + ], + "type": "text", + "content": "Li et al." + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 11 + }, + { + "para_blocks": [ + { + "type": "table", + "bbox": [ + 186, + 146, + 422, + 199 + ], + "blocks": [ + { + "bbox": [ + 131, + 114, + 482, + 137 + ], + "lines": [ + { + "bbox": [ + 131, + 114, + 482, + 137 + ], + "spans": [ + { + "bbox": [ + 131, + 114, + 482, + 137 + ], + "type": "text", + "content": "Table 5: Segmentation accuracies (AP) of Mask R-CNN with and without CF, and recognition accuracies (\\%) of MPM on IN-1K trained with various pseudo labels." + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 186, + 146, + 422, + 199 + ], + "lines": [ + { + "bbox": [ + 186, + 146, + 422, + 199 + ], + "spans": [ + { + "bbox": [ + 186, + 146, + 422, + 199 + ], + "type": "table", + "html": "
Pseudo LabelCFAP\\( AP_{50} \\)Clean\\( l_{\\infty} \\)\\( l_{\\infty}(\\times 2) \\)
--65.533.09.4
✓37.258.667.838.813.1
✓✓40.764.467.839.113.6
", + "image_path": "5392d36c47977097706d8152cf2fc82b4cde7cd0b21ebee02e47fd73c2ad5916.jpg" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "table_body" + } + ], + "index": 3 + }, + { + "bbox": [ + 131, + 220, + 482, + 269 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 131, + 220, + 482, + 269 + ], + "spans": [ + { + "bbox": [ + 131, + 220, + 482, + 269 + ], + "type": "text", + "content": "[34], ImageNet Sketch (IN-Sketch) [62], Stylized ImageNet (SIN) [16], and image distortion dataset (denoted as DIN) [15]. See Appendix H for the introduction of these datasets. The results shown in Tab. 4 indicate that MPM also exhibits improved robustness on these OOD shifts." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 131, + 285, + 299, + 298 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 131, + 285, + 299, + 298 + ], + "spans": [ + { + "bbox": [ + 131, + 285, + 299, + 298 + ], + "type": "text", + "content": "5.4 Other Advantages of MPM" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 130, + 305, + 482, + 389 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 305, + 482, + 389 + ], + "spans": [ + { + "bbox": [ + 130, + 305, + 482, + 389 + ], + "type": "text", + "content": "Models versus humans. The part-based models are motivated by the human recognition process. We were interested in whether the robustness of MPM aligned more with human cognition. To investigate this, we employed the evaluation method introduced by Geirhos et al. [15], which involved comparing the models' decisions on different distorted images with the actual judgments made by human observers. The results shown in Appendix I indicate that MPM measurably improved alignment with the human recognition process." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 130, + 389, + 482, + 475 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 389, + 482, + 475 + ], + "spans": [ + { + "bbox": [ + 130, + 389, + 482, + 475 + ], + "type": "text", + "content": "Boosting the downstream task. Li et al. [33] show that adversarially trained models on the large-scale IN-1K can be utilized to initialize the backbone of downstream networks (e.g., Faster R-CNN [51]). This approach together with downstream AT enables the transfer of adversarial robustness to downstream tasks. We conducted similar experiments by using the checkpoints of the vanilla baseline and MPM. The results shown in Appendix J indicate that MPM enhances both the clean accuracy and adversarial robustness in object detection." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 131, + 491, + 239, + 503 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 131, + 491, + 239, + 503 + ], + "spans": [ + { + "bbox": [ + 131, + 491, + 239, + 503 + ], + "type": "text", + "content": "5.5 Ablation Study" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 130, + 510, + 482, + 617 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 510, + 482, + 617 + ], + "spans": [ + { + "bbox": [ + 130, + 510, + 482, + 617 + ], + "type": "text", + "content": "Pseudo-label quality. We first investigate the influence of pseudo-label quality on the adversarial robustness of MPM. In Sec. 4.1, we illustrate the use of CF during generating pseudo labels. Here we remove the CF operator and regenerate pseudo labels for images without part annotations. The left half of Tab. 5 (the 2nd and 3rd rows) shows the segmentation accuracies of models with and without CF (here the models were supervised with the train set of " + }, + { + "bbox": [ + 130, + 510, + 482, + 617 + ], + "type": "inline_equation", + "content": "\\mathrm{PIN}++" + }, + { + "bbox": [ + 130, + 510, + 482, + 617 + ], + "type": "text", + "content": ", and the AP were calculated on val). It is seen that training with " + }, + { + "bbox": [ + 130, + 510, + 482, + 617 + ], + "type": "inline_equation", + "content": "\\mathrm{PIN}++" + }, + { + "bbox": [ + 130, + 510, + 482, + 617 + ], + "type": "text", + "content": " achieved good part segmentation results, indicating its potential for enhancing part-related tasks. CF further significantly improves AP." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 131, + 617, + 482, + 666 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 131, + 617, + 482, + 666 + ], + "spans": [ + { + "bbox": [ + 131, + 617, + 482, + 666 + ], + "type": "text", + "content": "We then trained MPM using both pseudo labels and real part annotations, as well as solely using real part annotations. The results, displayed in Tab. 5 (the 1st row), indicate that training MPM solely with real part annotations seems to be ineffective, potentially compromising the model's performance. We guess" + } + ] + } + ], + "index": 10 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 235, + 91, + 447, + 103 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 235, + 91, + 447, + 103 + ], + "spans": [ + { + "bbox": [ + 235, + 91, + 447, + 103 + ], + "type": "text", + "content": "PartImageNet++ Dataset: Scaling up Part-based ..." + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 470, + 91, + 481, + 101 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 470, + 91, + 481, + 101 + ], + "spans": [ + { + "bbox": [ + 470, + 91, + 481, + 101 + ], + "type": "text", + "content": "13" + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 12 + }, + { + "para_blocks": [ + { + "type": "table", + "bbox": [ + 192, + 146, + 416, + 220 + ], + "blocks": [ + { + "bbox": [ + 132, + 114, + 482, + 137 + ], + "lines": [ + { + "bbox": [ + 132, + 114, + 482, + 137 + ], + "spans": [ + { + "bbox": [ + 132, + 114, + 482, + 137 + ], + "type": "text", + "content": "Table 6: Recognition accuracies (\\%) of different part-based models on IN-1K. The number of parameters during inference is listed." + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 192, + 146, + 416, + 220 + ], + "lines": [ + { + "bbox": [ + 192, + 146, + 416, + 220 + ], + "spans": [ + { + "bbox": [ + 192, + 146, + 416, + 220 + ], + "type": "table", + "html": "
MethodParam.Cleanl∞l∞(×2)
Two-stage60.9M67.338.713.0
SV167.336.98.8
SV2 w/ TD25.6M67.637.111.7
MPM (SV3 w/ TD)67.839.113.6
SV3 w/o TD67.436.911.7
", + "image_path": "51d390e3426bf1ce5614f589f0a9f3b8a5acfecbfb67a18209979da0eba27cbb.jpg" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "table_body" + } + ], + "index": 3 + }, + { + "bbox": [ + 130, + 242, + 479, + 277 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 242, + 479, + 277 + ], + "spans": [ + { + "bbox": [ + 130, + 242, + 479, + 277 + ], + "type": "text", + "content": "it could caused by overfitting on the images with part annotations. Besides, the inclusion of better-quality pseudo labels can improve the robustness of our part-based model, underscoring the importance of high-quality pseudo labels." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 130, + 277, + 480, + 420 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 277, + 480, + 420 + ], + "spans": [ + { + "bbox": [ + 130, + 277, + 480, + 420 + ], + "type": "text", + "content": "Part-based model structure. We then conducted ablation experiments on the structure of the part-based models. Specifically, we compared MPM with the two-stage part-based model designed by Sitawarin et al. [56]. In addition, we performed ablations on the number of part supervisions. MPM incorporates part supervision with three resolutions, referred to as SV3. Here we compared it with SV1 and SV2, which indicate part-based models solely supervised by " + }, + { + "bbox": [ + 130, + 277, + 480, + 420 + ], + "type": "inline_equation", + "content": "f_{l}" + }, + { + "bbox": [ + 130, + 277, + 480, + 420 + ], + "type": "text", + "content": " and both " + }, + { + "bbox": [ + 130, + 277, + 480, + 420 + ], + "type": "inline_equation", + "content": "f_{l}" + }, + { + "bbox": [ + 130, + 277, + 480, + 420 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 130, + 277, + 480, + 420 + ], + "type": "inline_equation", + "content": "f_{l-1}" + }, + { + "bbox": [ + 130, + 277, + 480, + 420 + ], + "type": "text", + "content": ", respectively. Besides these, we extra performed an ablation by removing all top-down (TD) connections of MPM. The results are shown in Tab. 6. We can see that although the two-stage model needs about twice the parameters during inference, its robustness is inferior to the MPM. And all the designs of MPM, including utilizing higher resolution part annotations and top-down connections, contribute to the robustness." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 130, + 421, + 481, + 492 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 421, + 481, + 492 + ], + "spans": [ + { + "bbox": [ + 130, + 421, + 481, + 492 + ], + "type": "text", + "content": "Additional ablations. Additional ablation experiments emphasize the significance of fine-grained part annotations rather than object segmentation annotations. Experiments on Tiny-ImageNet [31] highlight the significance of investigating adversarial robustness on high-resolution images. Moreover, the adversarial robustness of part-based models exhibits low sensitivity to " + }, + { + "bbox": [ + 130, + 421, + 481, + 492 + ], + "type": "inline_equation", + "content": "\\lambda" + }, + { + "bbox": [ + 130, + 421, + 481, + 492 + ], + "type": "text", + "content": ". Refer to Appendix K for comprehensive information about these ablation studies." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 132, + 510, + 311, + 522 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 510, + 311, + 522 + ], + "spans": [ + { + "bbox": [ + 132, + 510, + 311, + 522 + ], + "type": "text", + "content": "6 Conclusion and Discussion" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 130, + 533, + 480, + 617 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 533, + 480, + 617 + ], + "spans": [ + { + "bbox": [ + 130, + 533, + 480, + 617 + ], + "type": "text", + "content": "In this work, we introduce PIN++, a new large-scale high-quality part dataset, to facilitate the research on part-based models for robust recognition. We propose a new MPM to leverage this dataset. With extensive experiments, we demonstrate the robustness of MPM across diverse scenarios. But we note that MPM represents just a stepping stone in the realm of part-based models for robustness. We hope that this work will inspire and motivate researchers to further explore the potential of part-based models for robust recognition." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 130, + 617, + 481, + 665 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 617, + 481, + 665 + ], + "spans": [ + { + "bbox": [ + 130, + 617, + 481, + 665 + ], + "type": "text", + "content": "Limitation. We believe that except for robust recognition, " + }, + { + "bbox": [ + 130, + 617, + 481, + 665 + ], + "type": "inline_equation", + "content": "\\mathrm{PIN}++" + }, + { + "bbox": [ + 130, + 617, + 481, + 665 + ], + "type": "text", + "content": " has the potential to boost general part-related visual understanding tasks. We have not conducted many benchmark experiments specifically for such tasks because they are beyond the scope of this work. We leave these as future work." + } + ] + } + ], + "index": 9 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "spans": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "type": "text", + "content": "14" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 166, + 91, + 200, + 101 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 166, + 91, + 200, + 101 + ], + "spans": [ + { + "bbox": [ + 166, + 91, + 200, + 101 + ], + "type": "text", + "content": "Li et al." + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 13 + }, + { + "para_blocks": [ + { + "bbox": [ + 133, + 114, + 241, + 129 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 114, + 241, + 129 + ], + "spans": [ + { + "bbox": [ + 133, + 114, + 241, + 129 + ], + "type": "text", + "content": "Acknowledgement" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 130, + 140, + 482, + 175 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 140, + 482, + 175 + ], + "spans": [ + { + "bbox": [ + 130, + 140, + 482, + 175 + ], + "type": "text", + "content": "This work was supported in part by the National Key Research and Development Program of China (No. 2021ZD0200301) and the National Natural Science Foundation of China (No. U2341228)." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 133, + 193, + 197, + 205 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 193, + 197, + 205 + ], + "spans": [ + { + "bbox": [ + 133, + 193, + 197, + 205 + ], + "type": "text", + "content": "References" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 134, + 217, + 480, + 665 + ], + "type": "list", + "angle": 0, + "index": 20, + "blocks": [ + { + "bbox": [ + 138, + 217, + 480, + 239 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 217, + 480, + 239 + ], + "spans": [ + { + "bbox": [ + 138, + 217, + 480, + 239 + ], + "type": "text", + "content": "1. Agarap, A.F.: Deep learning using rectified linear units (relu). arXiv preprint arXiv:1803.08375 (2018)" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 138, + 240, + 480, + 272 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 240, + 480, + 272 + ], + "spans": [ + { + "bbox": [ + 138, + 240, + 480, + 272 + ], + "type": "text", + "content": "2. Athalye, A., Carlini, N., Wagner, D.A.: Obfuscated gradients give a false sense of security: Circumventing defenses to adversarial examples. In: Int. Conf. Mach. Learn. (ICML). vol. 80, pp. 274-283 (2018)" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 138, + 274, + 480, + 293 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 274, + 480, + 293 + ], + "spans": [ + { + "bbox": [ + 138, + 274, + 480, + 293 + ], + "type": "text", + "content": "3. Bai, Y., Mei, J., Yuille, A.L., Xie, C.: Are transformers more robust than cnns? In: Adv. Neural Inform. Process. Syst. (NeurIPS). pp. 26831-26843 (2021)" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 138, + 295, + 480, + 316 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 295, + 480, + 316 + ], + "spans": [ + { + "bbox": [ + 138, + 295, + 480, + 316 + ], + "type": "text", + "content": "4. Biederman, I.: Recognition-by-components: a theory of human image understanding. Psychological Review 94(2), 115 (1987)" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 138, + 317, + 480, + 348 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 317, + 480, + 348 + ], + "spans": [ + { + "bbox": [ + 138, + 317, + 480, + 348 + ], + "type": "text", + "content": "5. Burl, M.C., Weber, M., Perona, P.: A probabilistic approach to object recognition using local photometry and global geometry. In: Eur. Conf. Comput. Vis. (ECCV). vol. 1407, pp. 628-641 (1998)" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 138, + 350, + 480, + 381 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 350, + 480, + 381 + ], + "spans": [ + { + "bbox": [ + 138, + 350, + 480, + 381 + ], + "type": "text", + "content": "6. Chen, X., Mottaghi, R., Liu, X., Fidler, S., Urtasun, R., Yuille, A.L.: Detect what you can: Detecting and representing objects using holistic models and body parts. In: IEEE Conf. Comput. Vis. Pattern Recog. (CVPR). pp. 1979-1986 (2014)" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 138, + 383, + 480, + 414 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 383, + 480, + 414 + ], + "spans": [ + { + "bbox": [ + 138, + 383, + 480, + 414 + ], + "type": "text", + "content": "7. Croce, F., Andriushchenko, M., Sehwag, V., Debenedetti, E., Flammarion, N., Chiang, M., Mittal, P., Hein, M.: Robustbench: a standardized adversarial robustness benchmark. In: NeurIPS Datasets and Benchmarks (2021)" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 138, + 415, + 480, + 446 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 415, + 480, + 446 + ], + "spans": [ + { + "bbox": [ + 138, + 415, + 480, + 446 + ], + "type": "text", + "content": "8. Croce, F., Hein, M.: Reliable evaluation of adversarial robustness with an ensemble of diverse parameter-free attacks. In: Int. Conf. Mach. Learn. (ICML). Proceedings of Machine Learning Research, vol. 119, pp. 2206-2216 (2020)" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 138, + 447, + 480, + 479 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 447, + 480, + 479 + ], + "spans": [ + { + "bbox": [ + 138, + 447, + 480, + 479 + ], + "type": "text", + "content": "9. Croce, F., Singh, N.D., Hein, M.: Robust semantic segmentation: Strong adversarial attacks and fast training of robust models. Adv. Neural Inform. Process. Syst. (NeurIPS) (2023)" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 134, + 480, + 480, + 511 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 134, + 480, + 480, + 511 + ], + "spans": [ + { + "bbox": [ + 134, + 480, + 480, + 511 + ], + "type": "text", + "content": "10. Debenedetti, E., Sehwag, V., Mittal, P.: A light recipe to train robust vision transformers. In: First IEEE Conference on Secure and Trustworthy Machine Learning (SaTML). pp. 225-253 (2023)" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 134, + 514, + 480, + 544 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 134, + 514, + 480, + 544 + ], + "spans": [ + { + "bbox": [ + 134, + 514, + 480, + 544 + ], + "type": "text", + "content": "11. Deng, J., Dong, W., Socher, R., Li, L., Li, K., Fei-Fei, L.: Imagenet: A large-scale hierarchical image database. In: IEEE Conf. Comput. Vis. Pattern Recog. (CVPR). pp. 248-255 (2009)" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 134, + 545, + 480, + 588 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 134, + 545, + 480, + 588 + ], + "spans": [ + { + "bbox": [ + 134, + 545, + 480, + 588 + ], + "type": "text", + "content": "12. Eykholt, K., Evtimov, I., Fernandes, E., Li, B., Rahmati, A., Xiao, C., Prakash, A., Kohno, T., Song, D.: Robust physical-world attacks on deep learning visual classification. In: IEEE Conf. Comput. Vis. Pattern Recog. (CVPR). pp. 1625-1634 (2018)" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 134, + 590, + 480, + 610 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 134, + 590, + 480, + 610 + ], + "spans": [ + { + "bbox": [ + 134, + 590, + 480, + 610 + ], + "type": "text", + "content": "13. Fang, Y., Yang, S., Wang, X., Li, Y., Fang, C., Shan, Y., Feng, B., Liu, W.: Instances as queries. In: Int. Conf. Comput. Vis. (ICCV). pp. 6890-6899 (2021)" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 134, + 611, + 480, + 632 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 134, + 611, + 480, + 632 + ], + "spans": [ + { + "bbox": [ + 134, + 611, + 480, + 632 + ], + "type": "text", + "content": "14. Fischler, M.A., Elschlager, R.A.: The representation and matching of pictorial structures. IEEE Trans. Comput. 22(1), 67-92 (1973)" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 134, + 633, + 480, + 665 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 134, + 633, + 480, + 665 + ], + "spans": [ + { + "bbox": [ + 134, + 633, + 480, + 665 + ], + "type": "text", + "content": "15. Geirhos, R., Narayanappa, K., Mitzkus, B., Thieringer, T., Bethge, M., Wichmann, F.A., Brendel, W.: Partial success in closing the gap between human and machine vision. In: Adv. Neural Inform. Process. Syst. (NeurIPS). pp. 23885-23899 (2021)" + } + ] + } + ], + "index": 19 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 235, + 91, + 447, + 102 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 235, + 91, + 447, + 102 + ], + "spans": [ + { + "bbox": [ + 235, + 91, + 447, + 102 + ], + "type": "text", + "content": "PartImageNet++ Dataset: Scaling up Part-based ..." + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 470, + 92, + 481, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 470, + 92, + 481, + 100 + ], + "spans": [ + { + "bbox": [ + 470, + 92, + 481, + 100 + ], + "type": "text", + "content": "15" + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 14 + }, + { + "para_blocks": [ + { + "bbox": [ + 132, + 116, + 480, + 665 + ], + "type": "list", + "angle": 0, + "index": 22, + "blocks": [ + { + "bbox": [ + 133, + 116, + 480, + 149 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 116, + 480, + 149 + ], + "spans": [ + { + "bbox": [ + 133, + 116, + 480, + 149 + ], + "type": "text", + "content": "16. Geirhos, R., Rubisch, P., Michaelis, C., Bethge, M., Wichmann, F.A., Brendel, W.: Imagenet-trained cnns are biased towards texture; increasing shape bias improves accuracy and robustness. In: Int. Conf. Learn. Rep. (ICLR) (2019)" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 133, + 149, + 480, + 182 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 149, + 480, + 182 + ], + "spans": [ + { + "bbox": [ + 133, + 149, + 480, + 182 + ], + "type": "text", + "content": "17. Girshick, R.B., Iandola, F.N., Darrell, T., Malik, J.: Deformable part models are convolutional neural networks. In: IEEE Conf. Comput. Vis. Pattern Recog. (CVPR). pp. 437-446 (2015)" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 133, + 182, + 480, + 214 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 182, + 480, + 214 + ], + "spans": [ + { + "bbox": [ + 133, + 182, + 480, + 214 + ], + "type": "text", + "content": "18. Gong, K., Liang, X., Li, Y., Chen, Y., Yang, M., Lin, L.: Instance-level human parsing via part grouping network. In: Eur. Conf. Comput. Vis. (ECCV). vol. 11208, pp. 805-822 (2018)" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 133, + 214, + 480, + 236 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 214, + 480, + 236 + ], + "spans": [ + { + "bbox": [ + 133, + 214, + 480, + 236 + ], + "type": "text", + "content": "19. Goodfellow, I.J., Shlens, J., Szegedy, C.: Explaining and harnessing adversarial examples. In: Int. Conf. Learn. Rep. (ICLR) (2015)" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 132, + 236, + 480, + 278 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 236, + 480, + 278 + ], + "spans": [ + { + "bbox": [ + 132, + 236, + 480, + 278 + ], + "type": "text", + "content": "20. Haaf, R.A., Fulkerson, A.L., Jablonski, B.J., Hupp, J.M., Shull, S.S., Pescara-Kovach, L.: Object recognition and attention to object components by preschool children and 4-month-old infants. Journal of Experimental Child Psychology 86(2), 108-123 (2003)" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 132, + 278, + 480, + 311 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 278, + 480, + 311 + ], + "spans": [ + { + "bbox": [ + 132, + 278, + 480, + 311 + ], + "type": "text", + "content": "21. He, J., Yang, S., Yang, S., Kortylewski, A., Yuan, X., Chen, J., Liu, S., Yang, C., Yu, Q., Yuille, A.L.: Partimagenet: A large, high-quality dataset of parts. In: Eur. Conf. Comput. Vis. (ECCV). vol. 13668, pp. 128-145 (2022)" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 132, + 311, + 480, + 333 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 311, + 480, + 333 + ], + "spans": [ + { + "bbox": [ + 132, + 311, + 480, + 333 + ], + "type": "text", + "content": "22. He, K., Gkioxari, G., Dollár, P., Girshick, R.B.: Mask R-CNN. IEEE Trans. Pattern Anal. Mach. Intell. (TPAMI) 42(2), 386-397 (2020)" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 132, + 333, + 480, + 354 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 333, + 480, + 354 + ], + "spans": [ + { + "bbox": [ + 132, + 333, + 480, + 354 + ], + "type": "text", + "content": "23. He, K., Zhang, X., Ren, S., Sun, J.: Deep residual learning for image recognition. In: IEEE Conf. Comput. Vis. Pattern Recog. (CVPR). pp. 770-778 (2016)" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 132, + 354, + 480, + 375 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 354, + 480, + 375 + ], + "spans": [ + { + "bbox": [ + 132, + 354, + 480, + 375 + ], + "type": "text", + "content": "24. Hendrycks, D., Dietterich, T.G.: Benchmarking neural network robustness to common corruptions and perturbations. In: Int. Conf. Learn. Rep. (ICLR) (2019)" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 132, + 375, + 480, + 396 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 375, + 480, + 396 + ], + "spans": [ + { + "bbox": [ + 132, + 375, + 480, + 396 + ], + "type": "text", + "content": "25. Hendrycks, D., Gimpel, K.: Gaussian error linear units (gelus). arXiv preprint arXiv:1606.08415 (2016)" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 132, + 396, + 480, + 418 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 396, + 480, + 418 + ], + "spans": [ + { + "bbox": [ + 132, + 396, + 480, + 418 + ], + "type": "text", + "content": "26. Hoffman, J.E., Subramaniam, B.: The role of visual attention in saccadic eye movements. Perception & Psychophysics 57(6), 787-795 (1995)" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 132, + 418, + 480, + 450 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 418, + 480, + 450 + ], + "spans": [ + { + "bbox": [ + 132, + 418, + 480, + 450 + ], + "type": "text", + "content": "27. Ilyas, A., Santurkar, S., Tsipras, D., Engstrom, L., Tran, B., Madry, A.: Adversarial examples are not bugs, they are features. In: Adv. Neural Inform. Process. Syst. (NeurIPS). pp. 125-136 (2019)" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 132, + 450, + 480, + 483 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 450, + 480, + 483 + ], + "spans": [ + { + "bbox": [ + 132, + 450, + 480, + 483 + ], + "type": "text", + "content": "28. Kirillov, A., Mintun, E., Ravi, N., Mao, H., Rolland, C., Gustafson, L., Xiao, T., Whitehead, S., Berg, A.C., Lo, W.Y., Dollar, P., Girshick, R.: Segment anything. In: Int. Conf. Comput. Vis. (ICCV). pp. 4015-4026 (2023)" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 132, + 483, + 480, + 504 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 483, + 480, + 504 + ], + "spans": [ + { + "bbox": [ + 132, + 483, + 480, + 504 + ], + "type": "text", + "content": "29. Krizhevsky, A., Hinton, G., et al.: Learning multiple layers of features from tiny images (2009)" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 132, + 504, + 480, + 536 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 504, + 480, + 536 + ], + "spans": [ + { + "bbox": [ + 132, + 504, + 480, + 536 + ], + "type": "text", + "content": "30. Krizhevsky, A., Sutskever, I., Hinton, G.E.: Imagenet classification with deep convolutional neural networks. In: Adv. Neural Inform. Process. Syst. (NeurIPS). pp. 1106-1114 (2012)" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 132, + 536, + 480, + 557 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 536, + 480, + 557 + ], + "spans": [ + { + "bbox": [ + 132, + 536, + 480, + 557 + ], + "type": "text", + "content": "31. Le, Y., Yang, X.: Tiny imagenet visual recognition challenge. CS 231N 7(7), 3 (2015)" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 132, + 557, + 480, + 590 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 557, + 480, + 590 + ], + "spans": [ + { + "bbox": [ + 132, + 557, + 480, + 590 + ], + "type": "text", + "content": "32. Li, L.H., Zhang, P., Zhang, H., Yang, J., Li, C., Zhong, Y., Wang, L., Yuan, L., Zhang, L., Hwang, J., Chang, K., Gao, J.: Grounded language-image pre-training. In: IEEE Conf. Comput. Vis. Pattern Recog. (CVPR). pp. 10955-10965 (2022)" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 132, + 590, + 480, + 611 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 590, + 480, + 611 + ], + "spans": [ + { + "bbox": [ + 132, + 590, + 480, + 611 + ], + "type": "text", + "content": "33. Li, X., Chen, H., Hu, X.: On the importance of backbone to the adversarial robustness of object detectors. arXiv preprint arXiv:2305.17438 (2023)" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 132, + 611, + 480, + 633 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 611, + 480, + 633 + ], + "spans": [ + { + "bbox": [ + 132, + 611, + 480, + 633 + ], + "type": "text", + "content": "34. Li, X., Li, J., Dai, T., Shi, J., Zhu, J., Hu, X.: Rethinking natural adversarial examples for classification models. arXiv preprint arXiv:2102.11731 (2021)" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 132, + 633, + 480, + 665 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 633, + 480, + 665 + ], + "spans": [ + { + "bbox": [ + 132, + 633, + 480, + 665 + ], + "type": "text", + "content": "35. Li, X., Wang, Z., Zhang, B., Sun, F., Hu, X.: Recognizing object by components with human prior knowledge enhances adversarial robustness of deep neural networks. IEEE Trans. Pattern Anal. Mach. Intell. (TPAMI) 45(7), 8861-8873 (2023)" + } + ] + } + ], + "index": 21 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "spans": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "type": "text", + "content": "16" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 166, + 91, + 199, + 100 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 166, + 91, + 199, + 100 + ], + "spans": [ + { + "bbox": [ + 166, + 91, + 199, + 100 + ], + "type": "text", + "content": "Li et al." + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 15 + }, + { + "para_blocks": [ + { + "bbox": [ + 130, + 116, + 482, + 666 + ], + "type": "list", + "angle": 0, + "index": 19, + "blocks": [ + { + "bbox": [ + 130, + 116, + 482, + 149 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 116, + 482, + 149 + ], + "spans": [ + { + "bbox": [ + 130, + 116, + 482, + 149 + ], + "type": "text", + "content": "36. Li, X., Zhang, W., Liu, Y., Hu, Z., Zhang, B., Hu, X.: Language-driven anchors for zero-shot adversarial robustness. In: IEEE Conf. Comput. Vis. Pattern Recog. (CVPR) (2024)" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 130, + 150, + 481, + 183 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 150, + 481, + 183 + ], + "spans": [ + { + "bbox": [ + 130, + 150, + 481, + 183 + ], + "type": "text", + "content": "37. Liang, X., Gong, K., Shen, X., Lin, L.: Look into person: Joint body parsing & pose estimation network and a new benchmark. IEEE Trans. Pattern Anal. Mach. Intell. (TPAMI) 41(4), 871-885 (2019)" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 130, + 183, + 481, + 216 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 183, + 481, + 216 + ], + "spans": [ + { + "bbox": [ + 130, + 183, + 481, + 216 + ], + "type": "text", + "content": "38. Liang, X., Liu, S., Shen, X., Yang, J., Liu, L., Dong, J., Lin, L., Yan, S.: Deep human parsing with active template regression. IEEE Trans. Pattern Anal. Mach. Intell. (TPAMI) 37(12), 2402-2414 (2015)" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 130, + 216, + 481, + 248 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 216, + 481, + 248 + ], + "spans": [ + { + "bbox": [ + 130, + 216, + 481, + 248 + ], + "type": "text", + "content": "39. Lin, C., Sun, P., Jiang, Y., Luo, P., Qu, L., Haffari, G., Yuan, Z., Cai, J.: Learning object-language alignments for open-vocabulary object detection. In: Int. Conf. Learn. Rep. (ICLR) (2023)" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 130, + 249, + 481, + 282 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 249, + 481, + 282 + ], + "spans": [ + { + "bbox": [ + 130, + 249, + 481, + 282 + ], + "type": "text", + "content": "40. Lin, T., Dollar, P., Girshick, R.B., He, K., Hariharan, B., Belongie, S.J.: Feature pyramid networks for object detection. In: IEEE Conf. Comput. Vis. Pattern Recog. (CVPR). pp. 936-944 (2017)" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 130, + 282, + 481, + 315 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 282, + 481, + 315 + ], + "spans": [ + { + "bbox": [ + 130, + 282, + 481, + 315 + ], + "type": "text", + "content": "41. Liu, C., Dong, Y., Xiang, W., Yang, X., Su, H., Zhu, J., Chen, Y., He, Y., Xue, H., Zheng, S.: A comprehensive study on robustness of image classification models: Benchmarking and rethinking. arXiv preprint arXiv:2302.14301 (2023)" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 130, + 315, + 481, + 335 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 315, + 481, + 335 + ], + "spans": [ + { + "bbox": [ + 130, + 315, + 481, + 335 + ], + "type": "text", + "content": "42. Liu, Q., Kortylewski, A., Bai, Y., Bai, S., Yuille, A.L.: Intriguing properties of text-guided diffusion models. arXiv preprint arXiv:2306.00974 (2023)" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 130, + 336, + 481, + 369 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 336, + 481, + 369 + ], + "spans": [ + { + "bbox": [ + 130, + 336, + 481, + 369 + ], + "type": "text", + "content": "43. Liu, S., Zhang, L., Yang, X., Su, H., Zhu, J.: Unsupervised part segmentation through disentangling appearance and shape. In: IEEE Conf. Comput. Vis. Pattern Recog. (CVPR). pp. 8355-8364 (2021)" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 130, + 369, + 481, + 403 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 369, + 481, + 403 + ], + "spans": [ + { + "bbox": [ + 130, + 369, + 481, + 403 + ], + "type": "text", + "content": "44. Liu, Z., Lin, Y., Cao, Y., Hu, H., Wei, Y., Zhang, Z., Lin, S., Guo, B.: Swin transformer: Hierarchical vision transformer using shifted windows. In: Int. Conf. Comput. Vis. (ICCV). pp. 9992-10002 (2021)" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 130, + 403, + 481, + 435 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 403, + 481, + 435 + ], + "spans": [ + { + "bbox": [ + 130, + 403, + 481, + 435 + ], + "type": "text", + "content": "45. Liu, Z., Mao, H., Wu, C., Feichtenhofer, C., Darrell, T., Xie, S.: A convnet for the 2020s. In: IEEE Conf. Comput. Vis. Pattern Recog. (CVPR). pp. 11966-11976 (2022)" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 130, + 435, + 481, + 456 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 435, + 481, + 456 + ], + "spans": [ + { + "bbox": [ + 130, + 435, + 481, + 456 + ], + "type": "text", + "content": "46. Madry, A., Makelov, A., Schmidt, L., Tsipras, D., Vladu, A.: Towards deep learning models resistant to adversarial attacks. In: Int. Conf. Learn. Rep. (ICLR) (2018)" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 130, + 456, + 481, + 490 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 456, + 481, + 490 + ], + "spans": [ + { + "bbox": [ + 130, + 456, + 481, + 490 + ], + "type": "text", + "content": "47. Mao, X., Chen, Y., Li, X., Qi, G., Duan, R., Zhang, R., Xue, H.: Easyrobust: A comprehensive and easy-to-use toolkit for robust computer vision. https://github.com/alibaba/easyrobust (2022)" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 130, + 490, + 481, + 522 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 490, + 481, + 522 + ], + "spans": [ + { + "bbox": [ + 130, + 490, + 481, + 522 + ], + "type": "text", + "content": "48. Meletis, P., Wen, X., Lu, C., de Geus, D., Dubbelman, G.: Cityscapes-panoptic-parts and pascal-panoptic-parts datasets for scene understanding. arXiv preprint arXiv:2004.07944 (2020)" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 130, + 522, + 481, + 567 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 522, + 481, + 567 + ], + "spans": [ + { + "bbox": [ + 130, + 522, + 481, + 567 + ], + "type": "text", + "content": "49. Ramanathan, V., Kalia, A., Petrovic, V., Wen, Y., Zheng, B., Guo, B., Wang, R., Marquez, A., Kovvuri, R., Kadian, A., Mousavi, A., Song, Y., Dubey, A., Mahajan, D.: PACO: parts and attributes of common objects. In: IEEE Conf. Comput. Vis. Pattern Recog. (CVPR). pp. 7141-7151 (2023)" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 130, + 567, + 481, + 600 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 567, + 481, + 600 + ], + "spans": [ + { + "bbox": [ + 130, + 567, + 481, + 600 + ], + "type": "text", + "content": "50. Reddy, N.D., Vo, M., Narasimhan, S.G.: Carfusion: Combining point tracking and part detection for dynamic 3d reconstruction of vehicles. In: IEEE Conf. Comput. Vis. Pattern Recog. (CVPR). pp. 1906-1915 (2018)" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 130, + 600, + 481, + 632 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 600, + 481, + 632 + ], + "spans": [ + { + "bbox": [ + 130, + 600, + 481, + 632 + ], + "type": "text", + "content": "51. Ren, S., He, K., Girshick, R.B., Sun, J.: Faster R-CNN: towards real-time object detection with region proposal networks. In: Adv. Neural Inform. Process. Syst. (NeurIPS). pp. 91-99 (2015)" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 130, + 632, + 481, + 666 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 632, + 481, + 666 + ], + "spans": [ + { + "bbox": [ + 130, + 632, + 481, + 666 + ], + "type": "text", + "content": "52. Rombach, R., Blattmann, A., Lorenz, D., Esser, P., Ommer, B.: High-resolution image synthesis with latent diffusion models. In: IEEE Conf. Comput. Vis. Pattern Recog. (CVPR). pp. 10674-10685 (2022)" + } + ] + } + ], + "index": 18 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 235, + 91, + 447, + 102 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 235, + 91, + 447, + 102 + ], + "spans": [ + { + "bbox": [ + 235, + 91, + 447, + 102 + ], + "type": "text", + "content": "PartImageNet++ Dataset: Scaling up Part-based ..." + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 470, + 91, + 481, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 470, + 91, + 481, + 100 + ], + "spans": [ + { + "bbox": [ + 470, + 91, + 481, + 100 + ], + "type": "text", + "content": "17" + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 16 + }, + { + "para_blocks": [ + { + "bbox": [ + 132, + 117, + 480, + 643 + ], + "type": "list", + "angle": 0, + "index": 20, + "blocks": [ + { + "bbox": [ + 133, + 117, + 480, + 138 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 117, + 480, + 138 + ], + "spans": [ + { + "bbox": [ + 133, + 117, + 480, + 138 + ], + "type": "text", + "content": "53. Rosch, E., Mervis, C.B.: Family resemblances: Studies in the internal structure of categories. Cognitive Psychology 7(4), 573-605 (1975)" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 132, + 140, + 480, + 171 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 140, + 480, + 171 + ], + "spans": [ + { + "bbox": [ + 132, + 140, + 480, + 171 + ], + "type": "text", + "content": "54. Salman, H., Ilyas, A., Engstrom, L., Kapoor, A., Madry, A.: Do adversarially robust imagenet models transfer better? In: Adv. Neural Inform. Process. Syst. (NeurIPS) (2020)" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 133, + 172, + 480, + 204 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 172, + 480, + 204 + ], + "spans": [ + { + "bbox": [ + 133, + 172, + 480, + 204 + ], + "type": "text", + "content": "55. Salman, H., Sun, M., Yang, G., Kapoor, A., Kolter, J.Z.: Denoised smoothing: A provable defense for pretrained classifiers. In: Adv. Neural Inform. Process. Syst. (NeurIPS) (2020)" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 132, + 205, + 480, + 226 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 205, + 480, + 226 + ], + "spans": [ + { + "bbox": [ + 132, + 205, + 480, + 226 + ], + "type": "text", + "content": "56. Sitawarin, C., Pongmala, K., Chen, Y., Carlini, N., Wagner, D.A.: Part-based models improve adversarial robustness. In: Int. Conf. Learn. Rep. (ICLR) (2023)" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 132, + 227, + 480, + 269 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 227, + 480, + 269 + ], + "spans": [ + { + "bbox": [ + 132, + 227, + 480, + 269 + ], + "type": "text", + "content": "57. Song, X., Wang, P., Zhou, D., Zhu, R., Guan, C., Dai, Y., Su, H., Li, H., Yang, R.: Apollocar3d: A large 3d car instance understanding benchmark for autonomous driving. In: IEEE Conf. Comput. Vis. Pattern Recog. (CVPR). pp. 5452-5462 (2019)" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 132, + 270, + 480, + 293 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 270, + 480, + 293 + ], + "spans": [ + { + "bbox": [ + 132, + 270, + 480, + 293 + ], + "type": "text", + "content": "58. Sun, P., Chen, S., Zhu, C., Xiao, F., Luo, P., Xie, S., Yan, Z.: Going denser with open-vocabulary part segmentation. arXiv preprint arXiv:2305.11173 (2023)" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 132, + 293, + 480, + 325 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 293, + 480, + 325 + ], + "spans": [ + { + "bbox": [ + 132, + 293, + 480, + 325 + ], + "type": "text", + "content": "59. Szegedy, C., Zaremba, W., Sutskever, I., Bruna, J., Erhan, D., Goodfellow, I.J., Fergus, R.: Intriguing properties of neural networks. In: Int. Conf. Learn. Rep. (ICLR) (2014)" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 132, + 326, + 480, + 346 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 326, + 480, + 346 + ], + "spans": [ + { + "bbox": [ + 132, + 326, + 480, + 346 + ], + "type": "text", + "content": "60. Tramèr, F., Carlini, N., Brendel, W., Madry, A.: On adaptive attacks to adversarial example defenses. In: Adv. Neural Inform. Process. Syst. (NeurIPS) (2020)" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 132, + 347, + 480, + 369 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 347, + 480, + 369 + ], + "spans": [ + { + "bbox": [ + 132, + 347, + 480, + 369 + ], + "type": "text", + "content": "61. Tversky, B., Hemenway, K.: Objects, parts, and categories. Journal of Experimental Psychology: General 113(2), 169 (1984)" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 132, + 369, + 480, + 402 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 369, + 480, + 402 + ], + "spans": [ + { + "bbox": [ + 132, + 369, + 480, + 402 + ], + "type": "text", + "content": "62. Wang, H., Ge, S., Lipton, Z.C., Xing, E.P.: Learning robust global representations by penalizing local predictive power. In: Adv. Neural Inform. Process. Syst. (NeurIPS). pp. 10506-10518 (2019)" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 132, + 403, + 480, + 435 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 403, + 480, + 435 + ], + "spans": [ + { + "bbox": [ + 132, + 403, + 480, + 435 + ], + "type": "text", + "content": "63. Wang, J., Zhang, Z., Xie, C., Premachandran, V., Yuille, A.: Unsupervised learning of object semantic parts from internal states of cnns by population encoding. arXiv preprint arXiv:1511.06855 (2015)" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 132, + 435, + 480, + 456 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 435, + 480, + 456 + ], + "spans": [ + { + "bbox": [ + 132, + 435, + 480, + 456 + ], + "type": "text", + "content": "64. Wu, D., Xia, S., Wang, Y.: Adversarial weight perturbation helps robust generalization. In: Adv. Neural Inform. Process. Syst. (NeurIPS) (2020)" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 132, + 457, + 480, + 489 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 457, + 480, + 489 + ], + "spans": [ + { + "bbox": [ + 132, + 457, + 480, + 489 + ], + "type": "text", + "content": "65. Zhang, H., Yu, Y., Jiao, J., Xing, E.P., Ghaoui, L.E., Jordan, M.I.: Theoretically principled trade-off between robustness and accuracy. In: Int. Conf. Mach. Learn. (ICML). vol. 97, pp. 7472-7482 (2019)" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 132, + 490, + 480, + 522 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 490, + 480, + 522 + ], + "spans": [ + { + "bbox": [ + 132, + 490, + 480, + 522 + ], + "type": "text", + "content": "66. Zhang, Y., Yao, Y., Jia, J., Yi, J., Hong, M., Chang, S., Liu, S.: How to robustify black-box ML models? A zeroth-order optimization perspective. In: Int. Conf. Learn. Rep. (ICLR) (2022)" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 132, + 523, + 480, + 566 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 523, + 480, + 566 + ], + "spans": [ + { + "bbox": [ + 132, + 523, + 480, + 566 + ], + "type": "text", + "content": "67. Zhao, J., Li, J., Cheng, Y., Sim, T., Yan, S., Feng, J.: Understanding humans in crowded scenes: Deep nested adversarial learning and A new benchmark for multi-human parsing. In: ACM Int. Conf. Multimedia (ACM MM). pp. 792-800 (2018)" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 132, + 567, + 480, + 588 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 567, + 480, + 588 + ], + "spans": [ + { + "bbox": [ + 132, + 567, + 480, + 588 + ], + "type": "text", + "content": "68. Zhao, Y., Li, J., Zhang, Y., Tian, Y.: Multi-class part parsing with joint boundary-semantic awareness. In: Int. Conf. Comput. Vis. (ICCV). pp. 9176-9185 (2019)" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 132, + 589, + 480, + 621 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 589, + 480, + 621 + ], + "spans": [ + { + "bbox": [ + 132, + 589, + 480, + 621 + ], + "type": "text", + "content": "69. Zhou, B., Zhao, H., Puig, X., Xiao, T., Fidler, S., Barriuso, A., Torralba, A.: Semantic understanding of scenes through the ADE20K dataset. Int. J. Comput. Vis. (IJCV) 127(3), 302-321 (2019)" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 132, + 622, + 480, + 643 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 622, + 480, + 643 + ], + "spans": [ + { + "bbox": [ + 132, + 622, + 480, + 643 + ], + "type": "text", + "content": "70. Zhou, Z., Firestone, C.: Humans can decipher adversarial images. Nature Communications 10(1), 1-9 (2019)" + } + ] + } + ], + "index": 19 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "spans": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "type": "text", + "content": "18" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 166, + 91, + 199, + 100 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 166, + 91, + 199, + 100 + ], + "spans": [ + { + "bbox": [ + 166, + 91, + 199, + 100 + ], + "type": "text", + "content": "Li et al." + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 17 + } + ], + "_backend": "vlm", + "_version_name": "2.6.4" +} \ No newline at end of file diff --git a/2024/PartSTAD_ 2D-to-3D Part Segmentation Task Adaptation/bbdaa935-8663-4268-b38a-384a769ccb82_content_list.json b/2024/PartSTAD_ 2D-to-3D Part Segmentation Task Adaptation/bbdaa935-8663-4268-b38a-384a769ccb82_content_list.json new file mode 100644 index 0000000000000000000000000000000000000000..91a5a12c0f582fc17a7d01aaaf38aa5bd2391397 --- /dev/null +++ b/2024/PartSTAD_ 2D-to-3D Part Segmentation Task Adaptation/bbdaa935-8663-4268-b38a-384a769ccb82_content_list.json @@ -0,0 +1,1724 @@ +[ + { + "type": "text", + "text": "PartSTAD: 2D-to-3D Part Segmentation Task Adaptation", + "text_level": 1, + "bbox": [ + 238, + 140, + 767, + 186 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Hyunjin $\\mathrm{Kim}^{1\\dagger}$ and Minhyuk $\\mathrm{Sung}^{2}$", + "bbox": [ + 359, + 210, + 643, + 228 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "1 KRAFTON Inc., South Korea", + "bbox": [ + 392, + 238, + 609, + 252 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "2 KAIST, South Korea", + "bbox": [ + 421, + 253, + 580, + 266 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "rlaguswls98@krafton.com, mhsung@kaist.ac.kr", + "bbox": [ + 334, + 268, + 666, + 281 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/826464d72d32b4853d3b3fdf9c02fc29e39f268677baaa949e049661e789fe67.jpg", + "image_caption": [ + "Figure 1: We introduce PARTSTAD, a novel few-shot 3D point cloud part segmentation method that leverages 2D-to-3D task adaptation. By obtaining 2D segmentation masks in multi-view images from GLIP [25] and SAM [20] and optimizing the mask weights for 3D segmentation as a learning objective, it can successfully predict fine-grained parts with accurate boundaries, as shown in the figure above." + ], + "image_footnote": [], + "bbox": [ + 220, + 290, + 787, + 426 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Abstract. We introduce PARTSTAD, a method designed for the task adaptation of 2D-to-3D segmentation lifting. Recent studies have highlighted the advantages of utilizing 2D segmentation models to achieve high-quality 3D segmentation through few-shot adaptation. However, previous approaches have focused on adapting 2D segmentation models for domain shift to rendered images and synthetic text descriptions, rather than optimizing the model specifically for 3D segmentation. Our proposed task adaptation method finetunes a 2D bounding box prediction model with an objective function for 3D segmentation. We introduce weights for 2D bounding boxes for adaptive merging and learn the weights using a small additional neural network. Additionally, we incorporate SAM, a foreground segmentation model on a bounding box, to improve the boundaries of 2D segments and consequently those of 3D segmentation. Our experiments on the PartNet-Mobility dataset show significant improvements with our task adaptation approach, achieving a $7.0\\%$ p increase in mIoU and a $5.2\\%$ p improvement in $\\mathrm{mAP}_{50}$ for semantic and instance segmentation compared to the SotA few-shot 3D segmentation model. The code is available at https://github.com/KAIST-Visual-AI-Group/PartSTAD.", + "bbox": [ + 259, + 529, + 743, + 792 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Keywords: part segmentation $\\cdot$ few-shot learning $\\cdot$ 3D deep learning", + "bbox": [ + 261, + 805, + 727, + 820 + ], + "page_idx": 0 + }, + { + "type": "page_footnote", + "text": "This work was conducted when the author was at KAIST.", + "bbox": [ + 220, + 824, + 624, + 838 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "1 Introduction", + "text_level": 1, + "bbox": [ + 217, + 143, + 375, + 161 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "3D segmentation has been a subject of extensive research in computer vision due to its central role in various downstream applications involving the understanding of shape structure, functionality, mobility, and semantics. However, the limited availability of annotated 3D shapes has remained a bottleneck in achieving generalizability in learned segmentation models for diverse 3D data. Annotating 3D segmentation is particularly labor-intensive, time-consuming, and requires expertise in handling 3D models. For this reason, the scale of 3D part-level annotations remains in the tens of thousands [37], while the scale of 2D annotation datasets in the image domain, for example, exceeds a million [20, 30].", + "bbox": [ + 212, + 175, + 785, + 311 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Recent research [2,32] has illuminated the potential of visual-language models bridging textual descriptions with images in accomplishing zero-shot or few-shot 3D segmentation results. The basic idea is to render a 3D model from various viewpoints, conduct 2D detection or segmentation on the rendered images, and then aggregate the 2D segmentation results into a 3D representation using either a voting mechanism [32] or a label propagation scheme [2]. This approach is not only effective in enabling zero-shot and few-shot 3D segmentation but also offers an advantage in that the set of part names does not need to be predefined in training but can be determined at test time.", + "bbox": [ + 212, + 311, + 785, + 446 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "PartSLIP [32] is a notable example that achieves 3D part segmentation results comparable to fully supervised methods while adapting a pretrained 2D detection model with few-shot training for 3D segmentation. It leverages the pretrained GLIP [25] model for 2D detection, associating the output 2D bounding boxes with one of the tokens (part names) provided in the input prompt. Its key component is the GLIP finetuning process, which introduces a small number of learnable parameters to the frozen GLIP model and trains them using few-shot synthetic images and texts employed in the 3D segmentation pipeline. As GLIP was initially trained with real photos and natural language descriptions of objects by humans, this adaptation to rendered images and the unconventional description (a sequence of part names) results in a substantial improvement in the 3D segmentation task.", + "bbox": [ + 212, + 446, + 785, + 628 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Despite promising initiatives, it is important to note that the previous approach was limited to achieving domain adaptation. In contrast, we emphasize that in the 2D-to-3D segmentation lifting task, not only does the data domain change, but the task itself also shifts from 2D segmentation to 3D segmentation. Thus, the process of applying the pretrained model to the new task requires task adaptation, involving modifying the model with an objective function associated with the new task. Particularly for 3D segmentation, it is necessary to integrate 2D segmentation results from multiple viewpoints into a coherent 3D representation. Therefore, it is crucial to control the noise of 2D bounding box predictions in the context of its impact on the final 3D segmentation after integration.", + "bbox": [ + 212, + 628, + 785, + 779 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Specifically, our Part Segmentation Task ADaptation method, PARTSTAD, adapts the pretrained GLIP model with a relaxed 3D mIoU loss while incorporating bounding boxes across all different views. Instead of typical finetuning, we draw inspiration from recent work [16, 32] that introduces small learnable", + "bbox": [ + 212, + 779, + 785, + 840 + ], + "page_idx": 1 + }, + { + "type": "page_number", + "text": "2", + "bbox": [ + 217, + 114, + 228, + 126 + ], + "page_idx": 1 + }, + { + "type": "header", + "text": "H. Kim and M. Sung", + "bbox": [ + 271, + 114, + 413, + 128 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "parameters to the existing model while keeping existing parameters frozen. As a result, we train a small MLP using 2D bounding box features extracted from the pretrained GLIP model. Since the relaxed mIoU loss is non-differentiable with respect to the bounding box positions, we suggest predicting a weight for each bounding box instead of adjusting its position. This results in minimal modification in the 2D-to-3D voting scheme while achieving a significant improvement in 3D segmentation, even with few-shot training, e.g., with eight objects per category, as done in PartSLIP. To further enhance 3D segmentation performance, we leverage SAM [20], a 2D instance segmentation model, to segment the foreground region within each bounding box, obtaining an accurate boundary for each 2D segment.", + "bbox": [ + 212, + 146, + 787, + 311 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "In our experiments on the PartNet-Mobility [56] dataset, we showcase the superior performance of our method in comparison to recent zero-shot/few-shot 3D semantic and instance segmentation methods that leverage 2D segmentation models. In contrast to the SotA few-shot method, our approach achieves a $7.0\\% \\mathbf{p}$ improvement in semantic segmentation mIoU and a $5.2\\% \\mathbf{p}$ improvement in instance segmentation $\\mathrm{mAP}_{50}$ . These improvements are consistent across all object classes.", + "bbox": [ + 212, + 311, + 787, + 417 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "2 Related Work", + "text_level": 1, + "bbox": [ + 215, + 441, + 387, + 458 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "2.1 Supervised 3D Segmentation", + "text_level": 1, + "bbox": [ + 215, + 476, + 501, + 491 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Given the availability of open segment-annotated 3D datasets [3,5,7,8,12,14,37, 56], 3D segmentation has been extensively researched in the last several years, focusing on the development of novel network architectures for supervised learning. Regarding the architecture of semantic segmentation, diverse models have been explored, including PointNet [42] and its variants [36,43,44,57,71], those using CNN [27,34,49,59], GCN [17,28,53], and Transformers [55,60,67,73]. For instance segmentation, various approaches based on proposing 3D bounding boxes [15,61,66], clustering learned features [11,19,31,38,50,51,69,74], and combining both semantic and instance segmentations [29,40,46,52] have been introduced. Despite these great advances, the performance of supervised methods has been limited by the scale of 3D datasets; the largest part-annotated 3D dataset, PartNet [37], includes fewer than $30\\mathrm{k}$ models.", + "bbox": [ + 212, + 502, + 787, + 683 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "There have been some attempts to overcome the limitation by leveraging additional weak supervisions, such as text descriptions of objects [21], 3D bounding box annotations [10], sparse labels [26, 35], and geometric priors [23, 24]. However, the scale of this additional data has also been limited to tens of thousands, while datasets in the image domain are on the scale of millions.", + "bbox": [ + 212, + 684, + 787, + 758 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "2.2 3D Segmentation Using Vision-Language Models", + "text_level": 1, + "bbox": [ + 215, + 782, + 665, + 797 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Recent work introduced ideas about exploiting vision-language models (VLM) for 3D segmentation. The advantage of utilizing the learned visual-language", + "bbox": [ + 212, + 809, + 785, + 839 + ], + "page_idx": 2 + }, + { + "type": "header", + "text": "PartSTAD: 2D-to-3D Part Segmentation Task Adaptation", + "bbox": [ + 343, + 114, + 730, + 128 + ], + "page_idx": 2 + }, + { + "type": "page_number", + "text": "3", + "bbox": [ + 774, + 116, + 784, + 126 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "grounding in 3D segmentation lies in the strong generalizability to arbitrary 3D models with zero-shot or few-shot training, and also in the open vocabulary understanding that enables labeling points without specifying the set of part or object names during training. For a VLM, CLIP [45] has been adapted in multiple previous works to lift the 2D grounding to the 3D, detecting parts in an object [13] and objects in a scene [9,39,47,70], while the results are limited to highlighting some regions without providing clear segment boundaries.", + "bbox": [ + 212, + 146, + 787, + 252 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "To obtain precise segments in 3D, other 2D object detection and segmentation models, such as SAM [20], GLIP [25], and GroundingDINO [33], are also leveraged. SAM3D [63] was an example of directly lifting the 2D masks from SAM to 3D with a bottom-up merging approach, although it was limited to segmenting the parts, not labeling them. SA3D [6] utilized NeRF representation and allowed user interaction for 3D segmentation, but it can be inefficient when the 3D objects are given as a mesh, point cloud, or other common 3D representations. OpenMask3D [47] and OpenIns3D [18] are concurrent works, both of which combine multiple foundation models, SAM [20] and CLIP [45] for OpenMask3D, and GroundingDINO [33] and LISA [22] for OpenIns3D. These works focus on segmenting objects in 3D scenes, while we aim to do finer-grained segmentation, finding parts in 3D objects.", + "bbox": [ + 212, + 255, + 787, + 436 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "PartSLIP [32] and SATR [2] are notable examples that achieve part segmentation of 3D objects based on a 2D object detection model, GLIP [25]. They obtain 2D bounding boxes on the rendered images from multiple views and integrate them over the 3D object with different merging schemes. The finetuning of GLIP introduced by PartSLIP particularly leads to a performance improvement, but it is limited to domain adaptation, not task adaptation. Building on this, we introduce a novel task adaptation technique as well as integration with SAM [20] to achieve a further significant improvement in 3D segmentation accuracy.", + "bbox": [ + 212, + 440, + 787, + 561 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "2.3 Task Adaptation", + "text_level": 1, + "bbox": [ + 215, + 599, + 401, + 616 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Taskonomy [68] is a seminal work that introduced the concept of task transfer learning for the first time, enabling the transformation of an image dense prediction model trained for a base task to perform other tasks through finetuning. Pruksachatkun et al. [41] also introduced a similar idea for language-domain tasks. Later, task adaptation has been extensively studied to leverage features learned from generative models, such as GAN [72] and diffusion models [4, 48, 58, 62], adapting them with a small network for various tasks, including image segmentation [4, 58, 72], depth prediction [48], and keypoint detection [62]. For 3D tasks, Abdelreheem et al. [1] are notable examples of using vision language-grounding models for 3D correspondence but are limited to directly applying 2D priors without finetuning. We introduce a novel task adaptation technique that lifts 2D segments to 3D by adapting the 2D features during the training of a small network with a task-specific objective function.", + "bbox": [ + 212, + 643, + 787, + 840 + ], + "page_idx": 3 + }, + { + "type": "page_number", + "text": "4", + "bbox": [ + 217, + 114, + 228, + 126 + ], + "page_idx": 3 + }, + { + "type": "header", + "text": "H. Kim and M. Sung", + "bbox": [ + 271, + 114, + 413, + 128 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "3 Background and Preliminaries", + "text_level": 1, + "bbox": [ + 215, + 143, + 545, + 162 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Recent research [2,13,32] has demonstrated how vision-language models designed for image object detection and segmentation can be applied to segment parts in 3D objects. The main idea involves rendering a 3D model from various viewpoints, performing 2D detection or segmentation for the rendered image at each view, and then combining the 2D segmentation results over the 3D model.", + "bbox": [ + 212, + 176, + 782, + 252 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Specifically, SATR [2] utilizes GLIP [25] as the 2D detection model. To obtain 2D bounding boxes for each of the specified part names, it feeds a concatenated list of these part names (separated by commas) as the input prompt to the pretrained GLIP model, along with a rendered image. The 2D bounding boxes are then mapped to a single face of the input 3D mesh, located at the center of the respective 2D bounding box, and the part names are propagated throughout the entire mesh based on geodesic distances. Similarly, PartSLIP [32] also leverages GLIP [25] for 2D detection but employs a different voting scheme for the 2D bounding box integration over the 3D object, which utilizes the oversegmentation of the input point cloud. One key distinction between PartSLIP and SATR is that PartSLIP proposes to modify the GLIP model for domain adaptation to the rendered images and text prompts used in 3D segmentation. Specifically, the text prompt is given as a concatenation of part names in the PartSLIP pipeline, which is not a typical and natural description of the object. Therefore, it learns a category-specific constant offset vector for the language embedding feature in the given pairs of prompts and rendered images, while keeping the GLIP parameters frozen.", + "bbox": [ + 212, + 252, + 787, + 508 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Our work builds upon the PartSLIP pipeline. In the following subsections, we describe the details of its major components: the voting scheme, and the finetuning process.", + "bbox": [ + 212, + 510, + 785, + 555 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "3.1 PartSLIP [32]", + "text_level": 1, + "bbox": [ + 214, + 579, + 375, + 595 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "The main technical components of PartSLIP are the 1) voting scheme based on super points and 2) adaptation to text prompts1. We describe the details of each component below.", + "bbox": [ + 212, + 604, + 784, + 651 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Voting for Super Points. PartSLIP first over-segments the input 3D object represented as a point cloud $\\mathcal{P}$ into a set of super points $P_{i} \\subseteq \\mathcal{P}$ , which provide geometric priors for the boundaries of the parts. The 3D object is rendered into $K$ viewpoints, and the 2D bounding boxes for each image are predicted using GLIP [25]. Let $\\mathcal{B}$ denote the entire set of 2D bounding boxes from all views, and $B_{jk} \\subseteq \\mathcal{B}$ indicates a subset that includes the bounding boxes classified as the $j$ -th part label from the $k$ -th view. $\\mathrm{V}_k: \\mathcal{P} \\to \\{0,1\\}$ is a function indicating whether the input point is visible from the $k$ -th viewpoint. Also, for each 2D", + "bbox": [ + 212, + 666, + 787, + 787 + ], + "page_idx": 4 + }, + { + "type": "header", + "text": "PartSTAD: 2D-to-3D Part Segmentation Task Adaptation", + "bbox": [ + 341, + 114, + 730, + 128 + ], + "page_idx": 4 + }, + { + "type": "page_number", + "text": "5", + "bbox": [ + 774, + 116, + 784, + 126 + ], + "page_idx": 4 + }, + { + "type": "page_footnote", + "text": "1 Although multi-view feature aggregation is also introduced, the improvement by this component is marginal, and the author of PartSLIP also did not include this component in their official code. Therefore, we will omit this part in our review.", + "bbox": [ + 217, + 796, + 787, + 839 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "bounding box $b \\in \\mathcal{B}$ , $\\mathrm{I}_b : \\mathcal{P} \\to \\{0,1\\}$ is a function indicating whether the input point is included in the bounding box $b$ . Given these, the voting aggregating the semantic labels of the 2D bounding boxes to the 3D point cloud is performed by calculating the following score $s_{ij}$ for each pair of the $i$ -th super point and the $j$ -th label:", + "bbox": [ + 212, + 146, + 787, + 220 + ], + "page_idx": 5 + }, + { + "type": "equation", + "text": "\n$$\ns _ {i j} = \\frac {\\sum_ {k} \\sum_ {p \\in P _ {i}} \\mathrm {V} _ {k} (p) \\left(\\max _ {b \\in \\left(B _ {j k} \\cup \\{b _ {\\emptyset} \\}\\right)} \\mathrm {I} _ {b} (p)\\right)}{\\sum_ {k} \\sum_ {p \\in P _ {i}} \\mathrm {V} _ {k} (p)}, \\tag {1}\n$$\n", + "text_format": "latex", + "bbox": [ + 334, + 232, + 785, + 268 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "where $b_{\\emptyset}$ is a null bounding box, and we assume that $\\mathrm{I}_{b_{\\emptyset}}$ for the null bounding box returns zero for any input point (simplifying notations for the case when $B_{jk} = \\emptyset$ ). Note that this $s_{ij} \\in [0,1]$ denotes the ratio of visible points from each view included in any of the 2D bounding boxes labeled with the $j$ -th label. Finally, for each $i$ -th super point, the label with the highest score $s_{ij}$ across all the labels is assigned to the super point. Exceptionally, if the highest score is less than a null label threshold $s_{\\emptyset}$ , which is set to 0.5 in our experiments, the null label is assigned to the super point.", + "bbox": [ + 214, + 277, + 787, + 398 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "GLIP Finetuning. While GLIP demonstrates impressive generalizability for unseen images and prompts, it still has limitations when it comes to handling synthetic images and text prompts, such as the rendered images and sequences of part names used in the PartSLIP pipeline. To address this, the authors of PartSLIP propose adding small, learnable parameters to the GLIP model while keeping all the pretrained GLIP parameters frozen. These new parameters are learned with very few-shot examples: 8 annotated 3D objects per category. They represent offset feature vectors for each part name and remain constant as global variables for each category, rather than changing for each input. This adaptation of GLIP significantly improves the accuracy of 3D part segmentation.", + "bbox": [ + 212, + 411, + 787, + 561 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Instance Segmentation. The voting scheme described above is designed for semantic segmentation, but PartSLIP also introduces a straightforward merging-based method to achieve instance segmentation based on semantic segmentation. To obtain instance segments, it merges adjacent super points under two conditions: 1) they share the same semantic label, and 2) they are either both included or both excluded for all bounding boxes.", + "bbox": [ + 212, + 575, + 787, + 667 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "4 PARTSTAD: Task Adaptation for 2D-to-3D", + "text_level": 1, + "bbox": [ + 214, + 688, + 671, + 707 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "PartSLIP has demonstrated that adapting a 2D detection model for 3D segmentation with a small set of trainable parameters can significantly enhance 3D segmentation performance. However, we observe that PartSLIP's adaptation is limited to domain adaptation, where it learns new parameters using synthetic images and data used in the PartSLIP framework without altering the objective function during training.", + "bbox": [ + 212, + 719, + 787, + 809 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "When applying the 2D segmentation model to 3D segmentation, it is important to note that not only does the data domain change, but the task itself", + "bbox": [ + 212, + 809, + 787, + 839 + ], + "page_idx": 5 + }, + { + "type": "page_number", + "text": "6", + "bbox": [ + 217, + 114, + 228, + 126 + ], + "page_idx": 5 + }, + { + "type": "header", + "text": "H. Kim and M. Sung", + "bbox": [ + 271, + 114, + 413, + 128 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/23e10151c3738afe3e36eff6d18c2f98fa62a94a5f38d19207d1137cf4c0bd08.jpg", + "image_caption": [ + "Figure 2: Overall pipeline of PARTSTAD. Our approach begins by rendering the provided 3D point cloud from multiple viewpoints. Subsequently, we extract 2D bounding boxes for its parts using GLIP [25] (Bounding Box Prediction); note that we utilize the finetuned GLIP model from PartSLIP [32]. Following this, we convert the bounding boxes into segmentation masks using SAM [20], extracting the foreground region for each bounding box (SAM Mask Integration). Next, we predict weights for all the masks and adaptively combine them into a 3D representation (2D-to-3D task adaptation). The final step involves obtaining the segmentation label for the input point cloud. The GLIP and SAM models are frozen, while only our novel weight prediction network is trained per category in a few-shot setting (8 objects)." + ], + "image_footnote": [], + "bbox": [ + 215, + 142, + 787, + 256 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "changes. Therefore, through the process of incorporating 2D bounding boxes into the 3D object, the 2D bounding box prediction model must align with our ultimate goal: 3D part segmentation.", + "bbox": [ + 212, + 445, + 784, + 491 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "To address this, we introduce Part Segmentation Task ADaptation method, PA STAD, a task adaptation approach designed to finetune the 2D segmentation model for 3D part segmentation. In Sec. 4.1, we introduce our objective function to finetune the GLIP for the 3D segmentation task. In Sec. 4.2, we describe how we adapt the pretrained GLIP model with the new objective function by introducing additional learnable parameters and modifying the voting scheme (Eq. 1). PartSLIP also faces a performance limitation due to the GLIP that provides 2D bounding boxes, not 2D segments. In Sec. 4.3, we also propose to combine the SAM [20] foreground mask for each bounding box to obtain more precise 2D segments and, consequently, further improve the 3D segmentation.", + "bbox": [ + 212, + 491, + 787, + 643 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "4.1 3D mRIoU Loss", + "text_level": 1, + "bbox": [ + 215, + 666, + 395, + 681 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "We begin by describing the loss function that we use for adapting the GLIP to the 3D segmentation task. Let $\\mathbf{l}_j$ and $\\hat{\\mathbf{l}}_j \\in \\{0,1\\}^{|\\mathcal{P}|}$ be binary vectors indicating whether each point $p$ in $\\mathcal{P}$ has the $j$ -th label in the ground truth and label prediction, respectively. The mean Intersection over Union (mIoU) for the predicted 3D segmentation, the standard evaluation metric for 3D segmentation, is defined as follows:", + "bbox": [ + 212, + 693, + 787, + 784 + ], + "page_idx": 6 + }, + { + "type": "equation", + "text": "\n$$\n\\operatorname {m I o U} \\left(\\left\\{\\mathbf {l} _ {j} \\right\\}, \\left\\{\\hat {\\mathbf {l}} _ {j} \\right\\}\\right) = \\frac {1}{M} \\sum_ {j = 1} ^ {M} \\frac {\\mathbf {l} _ {j} ^ {\\top} \\hat {\\mathbf {l}} _ {j}}{\\left\\| \\mathbf {l} _ {j} \\right\\| _ {1} + \\left\\| \\hat {\\mathbf {l}} _ {j} \\right\\| _ {1} - \\left. \\mathbf {l} _ {j} ^ {\\top} \\hat {\\mathbf {l}} _ {j} \\right.}, \\tag {2}\n$$\n", + "text_format": "latex", + "bbox": [ + 328, + 794, + 785, + 838 + ], + "page_idx": 6 + }, + { + "type": "header", + "text": "PartSTAD: 2D-to-3D Part Segmentation Task Adaptation", + "bbox": [ + 343, + 114, + 730, + 128 + ], + "page_idx": 6 + }, + { + "type": "page_number", + "text": "7", + "bbox": [ + 774, + 116, + 784, + 126 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "where $M$ is the number of labels.", + "bbox": [ + 215, + 146, + 455, + 160 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Our goal is to use the 3D mIoU as the objective function directly when adapting the GLIP model. However, the mIoU is non-differentiable. Thus, we employ a relaxed version known as mean Relaxed IoU (mRIOU) [24,65], which simply allows the predicted indicator vectors $\\hat{\\mathbf{l}}_j$ to be non-binary: $\\hat{\\mathbf{l}}_j\\in [0,1]^{|\\mathcal{P}|}$ . The loss $\\mathcal{L}_{\\mathrm{mRIOU}}$ is specifically defined as follows:", + "bbox": [ + 212, + 162, + 787, + 238 + ], + "page_idx": 7 + }, + { + "type": "equation", + "text": "\n$$\n\\mathcal {L} _ {\\mathrm {m R I o U}} = 1 - \\operatorname {m I o U} \\left(\\left\\{\\mathbf {l} _ {j} \\right\\}, \\left\\{\\hat {\\mathbf {l}} _ {j} \\right\\}\\right). \\tag {3}\n$$\n", + "text_format": "latex", + "bbox": [ + 388, + 244, + 785, + 262 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "In the supplementary material, we demonstrate that the choice of the loss function, as compared to other alternatives such as the cross-entropy loss, is crucial to achieving a substantial improvement in our task adaptation.", + "bbox": [ + 212, + 267, + 784, + 313 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Note that in the PartSLIP framework, where super points are utilized in 3D segmentation, and $s_{ij} \\in [0,1]$ indicates the likelihood of the $i$ -th super point belonging to the $j$ -label, the mRIoU can be calculated by defining $\\hat{\\mathbf{l}}_j$ based on $\\mathbf{s}_j = [s_{1j}, s_{2j}, \\dots]^{\\top}$ as follows:", + "bbox": [ + 212, + 313, + 784, + 377 + ], + "page_idx": 7 + }, + { + "type": "equation", + "text": "\n$$\n\\hat {\\mathbf {l}} _ {j} = \\mathbf {M} \\mathbf {s} _ {j}, \\tag {4}\n$$\n", + "text_format": "latex", + "bbox": [ + 464, + 395, + 785, + 414 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "where $\\mathbf{M} \\in \\{0,1\\}^{|\\mathcal{P}| \\times |\\{P_i\\}|}$ is a binary matrix that describes the memberships from each point to the super points.", + "bbox": [ + 212, + 420, + 784, + 452 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "4.2 Bounding Box Weight Prediction", + "text_level": 1, + "bbox": [ + 215, + 469, + 535, + 484 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "A typical approach for adapting a pretrained model to a new task would be finetuning the model using an objective function tailored to that new task. However, the finetuning method generally fails to yield meaningful improvements, especially when the goal is to achieve few-shot adaptation (e.g., with only 8 annotated 3D objects per class, as done in our case), while the pretrained model has been trained on an extensive dataset. To address this challenge, recent works like LoRA [16] and PartSLIP [32] have introduced the concept of adding small learnable parameters while keeping the existing pretrained model parameters frozen. This approach not only makes training very efficient but also enables the model to generalize effectively to unseen data.", + "bbox": [ + 212, + 489, + 787, + 641 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "While we also follow this approach, our specific challenge arises from the fact that when we use the 3D mRIoU (Eq. 3) as the objective function for 3D segmentation, it becomes non-differentiable with respect to the 2D bounding box location, which is the output of GLIP. This non-differentiability arises from the computation of the score $s_{ij}$ (Eq. 1) for each pair of super point and label.", + "bbox": [ + 212, + 642, + 787, + 717 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "To address this, we propose training a small network that does not refine the location but instead predicts a weight for each 2D bounding box, taking the learned features of the boxes as input. If we denote the output weight of the network for bounding box $b$ as $\\mathrm{W}(b) \\in \\mathbb{R}^{+}$ , the score $s_{ij}$ for a pair of super point and label is modified to $\\tilde{s}_{ij}$ as follows:", + "bbox": [ + 212, + 717, + 787, + 794 + ], + "page_idx": 7 + }, + { + "type": "equation", + "text": "\n$$\n\\tilde {s} _ {i j} = \\frac {\\sum_ {k} \\sum_ {p \\in P _ {i}} \\mathrm {V} _ {k} (p) \\left(\\max _ {b \\in \\left(B _ {j k} \\cup \\{b _ {\\emptyset} \\}\\right)} \\mathrm {I} _ {b} (p) \\mathrm {W} (b)\\right)}{\\sum_ {k} \\sum_ {p \\in P _ {i}} \\mathrm {V} _ {k} (p)}. \\tag {5}\n$$\n", + "text_format": "latex", + "bbox": [ + 316, + 800, + 785, + 838 + ], + "page_idx": 7 + }, + { + "type": "page_number", + "text": "8", + "bbox": [ + 217, + 114, + 228, + 126 + ], + "page_idx": 7 + }, + { + "type": "header", + "text": "H. Kim and M. Sung", + "bbox": [ + 271, + 114, + 413, + 128 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Note that the change is simply to multiply the weight $\\mathrm{W}(b)$ by $\\mathrm{I}_b(p)$ , while only $\\mathrm{I}_b(p)$ , indicating whether the point $p$ is included in bounding box $b$ , has been used in Eq. 1. This modified score no longer falls within the range of [0,1]. The final score $\\bar{s}_{ij}$ is thus defined by normalizing $\\tilde{s}_{ij}$ using the softmax function over the set of labels:", + "bbox": [ + 212, + 146, + 787, + 220 + ], + "page_idx": 8 + }, + { + "type": "equation", + "text": "\n$$\n\\bar {s} _ {i j} = \\frac {\\exp (\\tilde {s} _ {i j})}{\\sum_ {j} \\exp (\\tilde {s} _ {i j})}. \\tag {6}\n$$\n", + "text_format": "latex", + "bbox": [ + 434, + 229, + 785, + 263 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "We additionally propose to make the unnormalized score for the null label $\\tilde{s}_{\\emptyset}$ learnable in our case, initialized with 10. We thus compute the softmax above while including $\\tilde{s}_{\\emptyset}$ . The label of each super point is still chosen as the label giving the highest score $\\bar{s}_{ij}$ , becoming null if the normalized null label score $\\bar{s}_{\\emptyset}$ becomes the highest.", + "bbox": [ + 212, + 271, + 787, + 347 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "Network Architecture. We design the weight predictor network to take the bounding box feature vectors $\\mathbf{f}_b$ from the pretrained GLIP model as input. The feature vectors of all bounding boxes across all views are fed to the network at once and processed with a small shared two-layer MLP. Context normalization [64] is added in the middle of the two layers to incorporate contextual information from the global set of boxes for each box. The output of the MLP for each bounding box is further processed with the following modified ReLU function $\\phi(\\cdot)$ :", + "bbox": [ + 212, + 359, + 787, + 481 + ], + "page_idx": 8 + }, + { + "type": "equation", + "text": "\n$$\n\\phi (\\mathbf {x}) = \\max (\\tau + \\mathbf {x}, 0), \\tag {7}\n$$\n", + "text_format": "latex", + "bbox": [ + 421, + 492, + 784, + 507 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "where $\\tau$ is a user-defined constant offset (10 in our experiments).", + "bbox": [ + 214, + 517, + 683, + 532 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "4.3 SAM [20] Mask Integration", + "text_level": 1, + "bbox": [ + 214, + 551, + 490, + 569 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "Another limitation of PartSLIP is its reliance on GLIP [25] for 2D segmentation. Since GLIP is, in turn, a 2D object detection model, it produces bounding boxes instead of 2D segments, which cannot provide accurate boundaries of segments. We propose to address this issue by incorporating another pretrained 2D segmentation model, SAM [20]. SAM has the functionality of taking a bounding box as input and producing the foreground mask. Using this, we replace the set of 2D bounding boxes $\\mathcal{B}$ with a set of 2D masks, while preserving the remaining steps in the framework. Note that the point-to-bounding-box membership $\\mathrm{I}_b$ is changed to point-to-mask membership, and we still use the same bounding box features extracted from GLIP corresponding to each mask to train the weight prediction network. We demonstrate the effectiveness of applying the SAM mask in our experiments in Sec. 5.", + "bbox": [ + 212, + 575, + 787, + 758 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "5 Experiment Results", + "text_level": 1, + "bbox": [ + 214, + 779, + 444, + 797 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "In the experiments, we compare our method with the SotA 2D-prior-based zero-shot/few-shot 3D segmentation models using the PartNet-Mobility [56] dataset", + "bbox": [ + 212, + 809, + 785, + 840 + ], + "page_idx": 8 + }, + { + "type": "header", + "text": "PartSTAD: 2D-to-3D Part Segmentation Task Adaptation", + "bbox": [ + 341, + 114, + 732, + 130 + ], + "page_idx": 8 + }, + { + "type": "page_number", + "text": "9", + "bbox": [ + 774, + 116, + 784, + 126 + ], + "page_idx": 8 + }, + { + "type": "image", + "img_path": "images/60bec06afde2cda26bafb12f6d9cf23921e60a7fa0a6ec758435e00c0078f545.jpg", + "image_caption": [ + "Figure 3: Qualitative comparison of semantic segmentation results. Our PARTSTAD segments 3D parts more precisely with clearer boundaries, even for small (Camera, Chair) and thin (Clock) parts." + ], + "image_footnote": [], + "bbox": [ + 225, + 143, + 785, + 330 + ], + "page_idx": 9 + }, + { + "type": "table", + "img_path": "images/ccdc6a4fd0d3ef030f7013bc6bfa0a566f4b213d17f73c7e46a48fd2eae5af78.jpg", + "table_caption": [ + "Table 1: Quantitative comparison of the semantic segmentation results (mIoU(%) on the PartNet-Mobility dataset. Ours achieves a $7.0\\%$ p improvement in average mIoU compared to PartSLIP [32], the SotA few-shot 3D segmentation method, while consistently increasing mIoU across all categories. Please refer to the supplementary material for the complete table with results for all 45 categories." + ], + "table_footnote": [], + "table_body": "
MethodmIoUStorage FurnitureTableChairSwitchToiletLaptopUSBRemoteScissors
SATR [2]29.320.623.333.121.417.611.230.217.236.8
SATR [2]+SP34.828.928.037.737.022.112.433.428.043.0
PartSLIP [32]58.052.344.682.852.150.431.252.136.661.4
Ablation Study
w/o Weight Pred.61.956.645.085.051.956.631.557.136.660.8
w/o SAM Integ.62.154.045.783.153.049.433.653.646.867.5
PARTSTAD (Ours)65.059.547.885.357.957.534.659.953.468.5
", + "bbox": [ + 215, + 481, + 787, + 604 + ], + "page_idx": 9 + }, + { + "type": "text", + "text": "and OmniObject3D [54] dataset. In the supplementary material, we present additional results on the ablation study, outcomes with scanned data, and comprehensive results covering all PartNet-Mobility categories.", + "bbox": [ + 212, + 630, + 787, + 678 + ], + "page_idx": 9 + }, + { + "type": "text", + "text": "5.1 Experiment Setup", + "text_level": 1, + "bbox": [ + 214, + 696, + 413, + 712 + ], + "page_idx": 9 + }, + { + "type": "text", + "text": "Dataset. We use the PartNet-Mobility [56] dataset, which is also used in the experiments of PartSLIP [32]. The PartNet-Mobility [56] dataset includes 45 object categories. For each category, the training split includes 8 shapes, while the test split has a range of the number of objects from 6 to 338, totaling 1,906 objects across all 45 categories. Following PartSLIP, 8 shapes from the test set of each category are designated as the validation set. To obtain the 2D bounding boxes for each object, we use 10 images from fixed viewpoints across all our experiments. Note that the PartNet-Mobility [56] dataset is identical to the subset", + "bbox": [ + 212, + 718, + 787, + 840 + ], + "page_idx": 9 + }, + { + "type": "page_number", + "text": "10", + "bbox": [ + 217, + 114, + 235, + 126 + ], + "page_idx": 9 + }, + { + "type": "header", + "text": "H. Kim and M. Sung", + "bbox": [ + 271, + 114, + 413, + 128 + ], + "page_idx": 9 + }, + { + "type": "text", + "text": "of PartNet-Ensembled dataset used to train and evaluate PartSLIP [32]. (The remaining part of PartNet-Ensembled was only used to train other supervised baseline models, not PartSLIP [32].)", + "bbox": [ + 212, + 145, + 787, + 194 + ], + "page_idx": 10 + }, + { + "type": "text", + "text": "Evaluation Metrics. As evaluation metrics, we employ the mean Intersection over Union (mIoU) metric for semantic segmentation and mean Average Precision (mAP) for instance segmentation.", + "bbox": [ + 212, + 207, + 785, + 253 + ], + "page_idx": 10 + }, + { + "type": "text", + "text": "For semantic segmentation, the average mIoU over the entire dataset is computed by first calculating it per semantic part, averaging them across the parts in the same category, resulting in category-level mIoU, and then averaging the category-level mIoUs again across all the categories.", + "bbox": [ + 212, + 253, + 785, + 313 + ], + "page_idx": 10 + }, + { + "type": "text", + "text": "For instance segmentation, we compute two types of mAP: part-aware mAP and part-agnostic mAP. For part-aware mAP, similar to mIoU in semantic segmentation, we first calculate the mAP for each semantic part with instances having the part label. We then average them for each category, and then across categories. For part-agnostic mAP, we do not consider the semantic labels and directly compute the category-level mAP with part instances in the same object category, and then average them. All mAP values correspond to $\\mathrm{mAP}_{50}$ with an IoU threshold of $50\\%$ .", + "bbox": [ + 212, + 313, + 785, + 434 + ], + "page_idx": 10 + }, + { + "type": "text", + "text": "Baselines. We compare our method with three baselines: PartSLIP [32], SATR [2], and SAM3D [63].", + "bbox": [ + 214, + 449, + 805, + 481 + ], + "page_idx": 10 + }, + { + "type": "list", + "sub_type": "text", + "list_items": [ + "- PartSLIP [32] is the method that we adopt as our base. Note that PartSLIP uses a finetuned GLIP model, and we also employ the same model in our framework to extract bounding box features for our task adaptation.", + "- SATR [2] is similar to PartSLIP, conducting 3D segmentation using 2D bounding boxes from GLIP [25]. However, SATR differs from PartSLIP in four key aspects: 1) it does not involve finetuning the GLIP model, 2) it does not utilize super points, 3) it is designed for mesh segmentation, not point cloud segmentation, and 4) it employs a distinct approach for integrating 2D bounding boxes into 3D (label propagation). To ensure a fair comparison between our method and SATR, we implement the following changes to SATR. First, since we employ the GLIP model finetuned by PartSLIP in our framework, we use the same finetuned model within the SATR framework. Second, we convert the point cloud input to a mesh, performing mesh segmentation with SATR, and apply the results to the input point cloud by finding the closest vertex for each point. Lastly, we introduce a version of SATR using super points, resulting in a comparison between our method and two versions of SATR: one without super points (denoted as $SATR$ ) and one with super points (denoted as $SATR + SP$ ). Since SATR does not perform instance segmentation, our comparison with it is focused on semantic segmentation.", + "- SAM3D [63] is another zero-shot 3D segmentation method using a 2D segmentation model, not GLIP [25], but SAM [20]. Since it only performs instance segmentation but does not label each segment, we compare our" + ], + "bbox": [ + 225, + 492, + 785, + 840 + ], + "page_idx": 10 + }, + { + "type": "header", + "text": "PartSTAD: 2D-to-3D Part Segmentation Task Adaptation", + "bbox": [ + 341, + 113, + 730, + 130 + ], + "page_idx": 10 + }, + { + "type": "page_number", + "text": "11", + "bbox": [ + 767, + 114, + 782, + 126 + ], + "page_idx": 10 + }, + { + "type": "text", + "text": "method with it only for instance segmentation, using the part-agnostic mAP metric.", + "bbox": [ + 241, + 146, + 784, + 175 + ], + "page_idx": 11 + }, + { + "type": "text", + "text": "Ablation Study. We also conduct an ablation study, comparing our method with two major components being ablated: the weight prediction (Sec. 4.2) and the SAM mask integration (Sec. 4.3). In the supplementary material, we also compare our method with cases using cross-entropy loss instead of mR IoU loss (Sec. 4.1) and using GLIP bounding box confidence score instead of our learned weights in the weighted voting (Sec. 4.2). Note that if both weight prediction and SAM mask integration are excluded, our PARTSTAD is identical to PartSLIP [32].", + "bbox": [ + 212, + 190, + 787, + 311 + ], + "page_idx": 11 + }, + { + "type": "text", + "text": "5.2 Semantic Segmentation Results", + "text_level": 1, + "bbox": [ + 214, + 332, + 522, + 348 + ], + "page_idx": 11 + }, + { + "type": "text", + "text": "Tab. 1 and Fig. 3 present the quantitative and qualitative comparisons of part semantic segmentation results, respectively. Please refer to the supplementary material for the complete table encompassing all 45 categories. In comparison to PartSLIP [32], a SotA few-shot 3D segmentation method, we achieve a significant $7.0\\%$ p mIoU improvement over the entire set of 45 categories. Furthermore, for each specific category, we consistently demonstrate improvement, surpassing by more than $15\\%$ p in some categories (e.g., Remote). Qualitatively, the results also highlight the effectiveness of our task adaptation and the SAM mask integration in segmenting parts more precisely, even for small (e.g., Camera, Chair) and thin (e.g., Clock) parts of objects.", + "bbox": [ + 212, + 356, + 787, + 507 + ], + "page_idx": 11 + }, + { + "type": "text", + "text": "SATR exhibits inferior performance, even when leveraging the finetuned GLIP model and super points (see SATR+SP results) due to the label-propagation-based 2D bounding box integration, which often fails to provide clear boundaries in 3D segmentation, as illustrated in Fig. 3.", + "bbox": [ + 212, + 507, + 797, + 568 + ], + "page_idx": 11 + }, + { + "type": "text", + "text": "The ablation study results demonstrate the influence of each major component in our framework. The quantitative findings indicate that the weight prediction network plays a crucial role in substantial improvement, leading to a $3.1\\%$ decrease in average mIoU when it is ablated. The SAM mask integration also contributes significantly, resulting in a $2.9\\%$ average mIoU decrease when ablated.", + "bbox": [ + 212, + 568, + 787, + 657 + ], + "page_idx": 11 + }, + { + "type": "text", + "text": "5.3 Instance Segmentation Results", + "text_level": 1, + "bbox": [ + 214, + 680, + 516, + 695 + ], + "page_idx": 11 + }, + { + "type": "text", + "text": "Following the instance segmentation idea of PartSLIP, we also perform instance segmentation using our method and compare the quantitative results with those of other methods in Tab. 2 and Tab. 3, reporting part-aware mAPs and part-agnostic mAPs, respectively. Fig. 4 also displays qualitative comparisons across different methods. Please refer to the supplementary material for the complete table of all 45 categories. Similar to semantic segmentation, ours achieves $4.0\\%$ p and $5.2\\%$ p improvements for both average part-aware mAPs and part-agnostic mAPs, respectively. Both the weight prediction network and SAM mask integration also exhibit meaningful improvements in performance, as their influence is", + "bbox": [ + 212, + 704, + 787, + 839 + ], + "page_idx": 11 + }, + { + "type": "page_number", + "text": "12", + "bbox": [ + 217, + 114, + 235, + 126 + ], + "page_idx": 11 + }, + { + "type": "header", + "text": "H. Kim and M. Sung", + "bbox": [ + 271, + 114, + 413, + 128 + ], + "page_idx": 11 + }, + { + "type": "image", + "img_path": "images/a6d16926218f69eb912161d12a84bfd50612da83e1ef4eb6d887164cad58f27f.jpg", + "image_caption": [ + "Figure 4: Qualitative comparison of instance segmentation results shows that our PARTSTAD successfully segments tiny 3D parts, such as the keys of keyboards and buttons of remote controls, with clear segment boundaries." + ], + "image_footnote": [], + "bbox": [ + 220, + 145, + 787, + 333 + ], + "page_idx": 12 + }, + { + "type": "table", + "img_path": "images/5463fd0e69b7d0dcb78c904cb04baf8a63b8f83bb548b508ab52841d7e0ed036.jpg", + "table_caption": [ + "Table 2: Part-aware mAPs(\\%) of the instance segmentation results on the PartNet-Mobility dataset. Ours achieves $4.0\\%$ p improvement in mean part-aware mAPs compared to PartSLIP [32], with consistent enhancements across all categories. Please refer to the supplementary material for the complete table with results for all 45 categories." + ], + "table_footnote": [], + "table_body": "
MethodmAPStorage FurnitureTableChairSwitchToiletLaptopUSBRemoteScissors
PartSLIP [32]41.629.132.682.221.236.217.820.919.923.6
Ablation Study
w/o Weight Pred.44.733.830.882.522.240.824.417.420.325.5
w/o SAM Integ.44.229.432.382.219.736.023.624.829.623.1
PARTSTAD(Ours)45.635.533.282.522.140.628.926.533.726.5
", + "bbox": [ + 215, + 474, + 785, + 578 + ], + "page_idx": 12 + }, + { + "type": "table", + "img_path": "images/cc479a0ed7b244059a70793cad37e664ade4de084d10782d29ed4774a73587a2.jpg", + "table_caption": [ + "Table 3: Part-agnostic mAPs(%) of the instance segmentation results on the PartNet-Mobility dataset. Ours demonstrates $5.2\\%$ p improvement in mean part-agnostic mAPs compared to PartSLIP [32], surpassing SAM3D with a mean mAP that is more than three times larger. Please refer to the supplementary material for the complete table with results for all 45 categories. (*SAM3D is a zero-shot method.)" + ], + "table_footnote": [], + "table_body": "
MethodmAPStorage FurnitureTableChairSwitchToiletLaptopUSBRemoteScissors
SAM3D* [63]12.11.210.65.55.43.71.522.61.07.2
PartSLIP [32]38.929.728.780.721.235.419.520.519.927.0
Ablation Study
w/o Weight Pred.42.635.929.281.122.240.026.515.220.329.4
w/o SAM Integ.42.634.827.681.219.736.627.427.329.624.9
PARTSTAD (Ours)44.141.728.283.322.441.034.123.533.726.2
", + "bbox": [ + 215, + 678, + 787, + 790 + ], + "page_idx": 12 + }, + { + "type": "text", + "text": "depicted in the ablation study results. Compared to SAM3D [63], whose part-", + "bbox": [ + 215, + 824, + 785, + 840 + ], + "page_idx": 12 + }, + { + "type": "header", + "text": "PartSTAD: 2D-to-3D Part Segmentation Task Adaptation", + "bbox": [ + 343, + 114, + 730, + 128 + ], + "page_idx": 12 + }, + { + "type": "page_number", + "text": "13", + "bbox": [ + 767, + 116, + 784, + 126 + ], + "page_idx": 12 + }, + { + "type": "image", + "img_path": "images/6a310f47bdd29d93f1e3a2cfeea8e30a524f4d32022b4b55cd3941fbda53abfe.jpg", + "image_caption": [ + "Figure 5: Qualitative comparison of semantic segmentation results on OmniObect3D [54] dataset, a high quality real scanned 3D objects dataset. Our PART-STD predicts precise boundaries, even in the case that the object has an appearance significantly different from training data such as box category." + ], + "image_footnote": [], + "bbox": [ + 228, + 145, + 787, + 265 + ], + "page_idx": 13 + }, + { + "type": "text", + "text": "agnostic mAPs are shown in Tab. 3, our method achieves much better performance with more than three times greater mAP. Qualitative results also illustrate the outstanding performance of our method, accurately identifying all the tiny instances of parts, such as keys on keyboards, buttons on remotes, and clear boundaries of these instances.", + "bbox": [ + 212, + 358, + 787, + 434 + ], + "page_idx": 13 + }, + { + "type": "text", + "text": "5.4 Part Segmentation with Real Scanned Dataset", + "text_level": 1, + "bbox": [ + 214, + 452, + 645, + 468 + ], + "page_idx": 13 + }, + { + "type": "text", + "text": "Fig. 5 displays additional results on the OmniObject3D [54] dataset, a high-quality dataset featuring real-scanned 3D objects. We present qualitative results only, as ground truth segmentation for OmniObject3D [54] is unavailable. Note that our PARTSTAD accurately identifies parts in real scans, demonstrating its effectiveness beyond synthetic data. Particularly impressive is its performance in the box category depicted in Fig. 5, where despite significant discrepancies in appearance compared to the training data, PARTSTAD achieves remarkably accurate predictions.", + "bbox": [ + 212, + 474, + 787, + 595 + ], + "page_idx": 13 + }, + { + "type": "text", + "text": "6 Conclusion and Future Work", + "text_level": 1, + "bbox": [ + 214, + 614, + 532, + 632 + ], + "page_idx": 13 + }, + { + "type": "text", + "text": "We presented PARTSTAD, a task adaptation method for lifting 2D segmentation to 3D. Instead of directly finetuning a 2D segmentation network, our method learns a small neural network predicting weights for each 2D bounding box with an objective function for 3D segmentation, and then performs 3D segmentation via weighted merging of the boxes in 3D. We further improve the performance of 3D segmentation by integrating SAM foreground masks for each bounding box. We achieve the SotA results in few-shot 3D part segmentation, demonstrating significant improvements in both semantic and instance segmentations.", + "bbox": [ + 212, + 643, + 787, + 763 + ], + "page_idx": 13 + }, + { + "type": "text", + "text": "As PARTSTAD still depends on 2D representations for 3D predictions, its understanding of 3D geometry is relatively limited, and thus it cannot account for occluded or interior points, which poses a limitation. In future work, we plan to further investigate combining our approach with the direct utilization of 3D representations, such as leveraging 3D features, for more accurate predictions.", + "bbox": [ + 212, + 763, + 787, + 840 + ], + "page_idx": 13 + }, + { + "type": "page_number", + "text": "14", + "bbox": [ + 217, + 114, + 235, + 126 + ], + "page_idx": 13 + }, + { + "type": "header", + "text": "H. Kim and M. Sung", + "bbox": [ + 271, + 114, + 413, + 128 + ], + "page_idx": 13 + }, + { + "type": "text", + "text": "Acknowledgements", + "text_level": 1, + "bbox": [ + 217, + 143, + 401, + 162 + ], + "page_idx": 14 + }, + { + "type": "text", + "text": "We sincerely thank Minghua Liu for providing the code of PartSLIP and answering questions. This work was supported by NRF grant (RS-2023-00209723), IITP grants (2022-0-00594, RS-2023-00227592, RS-2024-00399817), and Alchymist Project Program (RS-2024-00423625) funded by the Korean government (MSIT and MOTIE), and grants from the DRB-KAIST SketchTheFuture Research Center, Naver-intel, Adobe Research, Hyundai NGV, KT, and Samsung Electronics.", + "bbox": [ + 212, + 179, + 812, + 270 + ], + "page_idx": 14 + }, + { + "type": "text", + "text": "References", + "text_level": 1, + "bbox": [ + 215, + 296, + 321, + 313 + ], + "page_idx": 14 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "1. Abdelreheem, A., Eldesokey, A., Ovsjanikov, M., Wonka, P.: Zero-shot 3D shape correspondence. In: SIGGRAPH Asia (2023)", + "2. Abdelreheem, A., Skorokhodov, I., Ovsjanikov, M., Wonka, P.: SATR: Zero-shot semantic segmentation of 3D shapes. In: ICCV (2023)", + "3. Armeni, I., Sax, S., Zamir, A.R., Savarese, S.: Joint 2D-3D-semantic data for indoor scene understanding. arXiv preprint arXiv:1702.01105 (2017)", + "4. Baranchuk, D., Voynov, A., Rubachev, I., Khrulkov, V., Babenko, A.: Label-efficient semantic segmentation with diffusion models. In: ICLR (2022)", + "5. Behley, J., Garbade, M., Milioto, A., Quenzel, J., Behnke, S., Stachniss, C., Gall, J.: SemanticKITTI: A dataset for semantic scene understanding of LiDAR sequences. In: ICCV (2019)", + "6. Cen, J., Zhou, Z., Fang, J., Yang, C., Shen, W., Xie, L., Zhang, X., Tian, Q.: Segment anything in 3D with nerfs. In: NeurIPS (2023)", + "7. Chang, A., Dai, A., Funkhouser, T., Halber, M., Niessner, M., Savva, M., Song, S., Zeng, A., Zhang, Y.: Matterport3D: Learning from rgb-d data in indoor environments. In: 3DV (2017)", + "8. Chang, A.X., Funkhouser, T., Guibas, L., Hanrahan, P., Huang, Q., Li, Z., Savarese, S., Savva, M., Song, S., Su, H., et al.: ShapeNet: An information-rich 3D model repository. arXiv preprint arXiv:1512.03012 (2015)", + "9. Chen, R., Liu, Y., Kong, L., Zhu, X., Ma, Y., Li, Y., Hou, Y., Qiao, Y., Wang, W.: CLIP2Scene: Towards label-efficient 3D scene understanding by clip. In: CVPR (2023)", + "0. Chibane, J., Engelmann, F., Tran, T.A., Pons-Moll, G.: Box2Mask: Weakly supervised 3D semantic instance segmentation using bounding boxes. In: ECCV (2022)", + "1. Chu, R., Chen, Y., Kong, T., Qi, L., Li, L.: ICM-3D: Instantiated category modeling for 3D instance segmentation. IEEE Robotics and Automation Letters (2021)", + "2. Dai, A., Chang, A.X., Savva, M., Halber, M., Funkhouser, T., Nießner, M.: ScanNet: Richly-annotated 3D reconstructions of indoor scenes. In: CVPR (2017)", + "3. Decatur, D., Lang, I., Hanocka, R.: 3D Highlighter: Localizing regions on 3D shapes via text descriptions. In: CVPR (2023)", + "4. Hackel, T., Savinov, N., Ladicky, L., Wegner, J.D., Schindler, K., Pollefeys, M.: SEMANTIC3D.NET: A new large-scale point cloud classification benchmark. In: ISPRS Annals of the Photogrammetry, Remote Sensing and Spatial Information Sciences (2017)", + "5. Hou, J., Dai, A., Nießner, M.: 3D-SIS: 3D semantic instance segmentation of rgb-d scans. In: CVPR (2019)" + ], + "bbox": [ + 225, + 330, + 785, + 839 + ], + "page_idx": 14 + }, + { + "type": "header", + "text": "PartSTAD: 2D-to-3D Part Segmentation Task Adaptation", + "bbox": [ + 341, + 114, + 730, + 128 + ], + "page_idx": 14 + }, + { + "type": "page_number", + "text": "15", + "bbox": [ + 767, + 116, + 785, + 126 + ], + "page_idx": 14 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "16. Hu, E.J., Shen, Y., Wallis, P., Allen-Zhu, Z., Li, Y., Wang, S., Wang, L., Chen, W.: LoRA: Low-rank adaptation of large language models. arXiv preprint arXiv:2106.09685 (2021)", + "17. Huang, R., Huang, C., Liu, Y., Dai, G., Kong, W.: LSGCN: Long short-term traffic prediction with graph convolutional networks. In: IJCAI (2020)", + "18. Huang, Z., Wu, X., Chen, X., Zhao, H., Zhu, L., Lasenby, J.: OpenIns3D: Snap and lookup for 3D open-vocabulary instance segmentation. arXiv preprint (2023)", + "19. Jiang, L., Zhao, H., Shi, S., Liu, S., Fu, C.W., Jia, J.: PointGroup: Dual-set point grouping for 3D instance segmentation. In: CVPR (2020)", + "20. Kirillov, A., Mintun, E., Ravi, N., Mao, H., Rolland, C., Gustafson, L., Xiao, T., Whitehead, S., Berg, A.C., Lo, W.Y., et al.: Segment anything. arXiv preprint arXiv:2304.02643 (2023)", + "21. Koo, J., Huang, I., Achlioptas, P., Guibas, L.J., Sung, M.: PartGlot: Learning shape part segmentation from language reference games. In: CVPR (2022)", + "22. Lai, X., Tian, Z., Chen, Y., Li, Y., Yuan, Y., Liu, S., Jia, J.: LISA: Reasoning segmentation via large language model. arXiv preprint arXiv:2308.00692 (2023)", + "23. Lé, E.T., Sung, M., Ceylan, D., Mech, R., Boubekeur, T., Mitra, N.J.: CPFN: Cascaded primitive fitting networks for high-resolution point clouds. In: ICCV (2021)", + "24. Li, L., Sung, M., Dubrovina, A., Yi, L., Guibas, L.J.: Supervised fitting of geometric primitives to 3D point clouds. In: CVPR (2019)", + "25. Li, L.H., Zhang, P., Zhang, H., Yang, J., Li, C., Zhong, Y., Wang, L., Yuan, L., Zhang, L., Hwang, J.N., et al.: Grounded language-image pre-training. In: CVPR (2022)", + "26. Li, M., Xie, Y., Shen, Y., Ke, B., Qiao, R., Ren, B., Lin, S., Ma, L.: Hybridcr: Weakly-supervised 3d point cloud semantic segmentation via hybrid contrastive regularization. In: CVPR (2022)", + "27. Li, Y., Bu, R., Sun, M., Wu, W., Di, X., Chen, B.: PointCNN: Convolution on x-transformed points. In: NeurIPS (2018)", + "28. Li, Y., Ma, L., Zhong, Z., Cao, D., Li, J.: TGNet: Geometric graph cnn on 3D point cloud segmentation. IEEE Transactions on Geoscience and Remote Sensing (2019)", + "29. Liang, Z., Yang, M., Li, H., Wang, C.: 3D instance embedding learning with a structure-aware loss function for point cloud segmentation. IEEE Robotics and Automation Letters (2020)", + "30. Lin, T.Y., Maire, M., Belongie, S., Hays, J., Perona, P., Ramanan, D., Dólar, P., Zitnick, C.L.: Microsoft COCO: Common objects in context. In: ECCV (2014)", + "31. Liu, J., Yu, M., Ni, B., Chen, Y.: Self-prediction for joint instance and semantic segmentation of point clouds. In: ECCV (2020)", + "32. Liu, M., Zhu, Y., Cai, H., Han, S., Ling, Z., Porikli, F., Su, H.: PartSLIP: Low-shot part segmentation for 3d point clouds via pretrained image-language models. In: CVPR (2023)", + "33. Liu, S., Zeng, Z., Ren, T., Li, F., Zhang, H., Yang, J., Li, C., Yang, J., Su, H., Zhu, J., et al.: Grounding DINO: Marrying dino with grounded pre-training for open-set object detection. arXiv preprint arXiv:2303.05499 (2023)", + "34. Liu, Y., Fan, B., Meng, G., Lu, J., Xiang, S., Pan, C.: DensePoint: Learning densely contextual representation for efficient point cloud processing. In: ICCV (2019)", + "35. Liu, Z., Qi, X., Fu, C.W.: One thing one click: A self-training approach for weakly supervised 3D semantic segmentation. In: CVPR (2021)", + "36. Ma, X., Qin, C., You, H., Ran, H., Fu, Y.: Rethinking network design and local geometry in point cloud: A simple residual mlp framework. In: ICLR (2022)" + ], + "bbox": [ + 215, + 146, + 784, + 840 + ], + "page_idx": 15 + }, + { + "type": "page_number", + "text": "16", + "bbox": [ + 217, + 114, + 235, + 126 + ], + "page_idx": 15 + }, + { + "type": "header", + "text": "H. Kim and M. Sung", + "bbox": [ + 271, + 114, + 413, + 128 + ], + "page_idx": 15 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "37. Mo, K., Zhu, S., Chang, A.X., Yi, L., Tripathi, S., Guibas, L.J., Su, H.: PartNet: A large-scale benchmark for fine-grained and hierarchical part-level 3D object understanding. In: CVPR (2019)", + "38. Narita, G., Seno, T., Ishikawa, T., Kaji, Y.: Panoptic fusion: Online volumetric semantic mapping at the level of stuff and things. In: IROS (2019)", + "39. Peng, S., Genova, K., Jiang, C., Tagliasacchi, A., Pollefeys, M., Funkhouser, T., et al.: OpenScene: 3D scene understanding with open vocabularies. In: CVPR (2023)", + "40. Pham, Q.H., Nguyen, D.T., Hua, B.S., Roig, G., Yeung, S.K.: JSIS3D: Joint semantic-instance segmentation of 3D point clouds with multi-task pointwise networks and multi-value conditional random fields. In: CVPR (2019)", + "41. Pruksachatkun, Y., Phang, J., Liu, H., Htut, P.M., Zhang, X., Pang, R.Y., Vania, C., Kann, K., Bowman, S.R.: Intermediate-Task Transfer Learning with Pretrained Language Models: When and why does it work? In: ACL (2020)", + "42. Qi, C.R., Su, H., Mo, K., Guibas, L.J.: PointNet: Deep learning on point sets for 3d classification and segmentation. In: CVPR (2017)", + "43. Qi, C.R., Yi, L., Su, H., Guibas, L.J.: PointNet++: Deep hierarchical feature learning on point sets in a metric space. In: NeurIPS (2017)", + "44. Qian, G., Li, Y., Peng, H., Mai, J., Hammoud, H., Elhoseiny, M., Ghanem, B.: PointNext: Revisiting pointnet++ with improved training and scaling strategies. In: NeurIPS (2022)", + "45. Radford, A., Kim, J.W., Hallacy, C., Ramesh, A., Goh, G., Agarwal, S., Sastry, G., Askell, A., Mishkin, P., Clark, J., et al.: Learning transferable visual models from natural language supervision. In: ICML (2021)", + "46. Schult, J., Engelmann, F., Hermans, A., Litany, O., Tang, S., Leibe, B.: Mask3D: Mask transformer for 3D semantic instance segmentation. In: ICRA (2023)", + "47. Takmaz, A., Fedele, E., Sumner, R.W., Pollefeys, M., Tombari, F., Engelmann, F.: OpenMask3D: Open-vocabulary 3D instance segmentation. In: NeurIPS (2023)", + "48. Tang, L., Jia, M., Wang, Q., Phoo, C., Hariharan, B.: Emergent correspondence from image diffusion. In: NeurIPS (2023)", + "49. Thomas, H., Qi, C.R., Deschaud, J.E., Marcotegui, B., Goulette, F., Guibas, L.J.: Kpconv: Flexible and deformable convolution for point clouds. In: ICCV (2019)", + "50. Vu, T., Kim, K., Luu, T.M., Nguyen, X.T., Yoo, C.D.: Softgroup for 3d instance segmentation on 3D point clouds. In: CVPR (2022)", + "51. Wang, W., Yu, R., Huang, Q., Neumann, U.: SGPN: Similarity group proposal network for 3D point cloud instance segmentation. In: CVPR (2018)", + "52. Wang, X., Liu, S., Shen, X., Shen, C., Jia, J.: Associatively segmenting instances and semantics in point clouds. In: CVPR (2019)", + "53. Wang, Y., Sun, Y., Liu, Z., Sarma, S.E., Bronstein, M.M., Solomon, J.M.: Dynamic graph cnn for learning on point clouds. ACM TOG (2019)", + "54. Wu, T., Zhang, J., Fu, X., Wang, Y., Ren, J., Pan, L., Wu, W., Yang, L., Wang, J., Qian, C., Lin, D., Liu, Z.: OmniObject3D: Large-vocabulary 3D object dataset for realistic perception, reconstruction and generation. In: CVPR (2023)", + "55. Wu, X., Lao, Y., Jiang, L., Liu, X., Zhao, H.: Point transformer v2: Grouped vector attention and partition-based pooling. In: NeurIPS (2022)", + "56. Xiang, F., Qin, Y., Mo, K., Xia, Y., Zhu, H., Liu, F., Liu, M., Jiang, H., Yuan, Y., Wang, H., Yi, L., Chang, A.X., Guibas, L.J., Su, H.: SAPIEN: A simulated part-based interactive environment. In: CVPR (2020)", + "57. Xiang, T., Zhang, C., Song, Y., Yu, J., Cai, W.: Walk in the Cloud: Learning curves for point clouds shape analysis. In: ICCV (2021)" + ], + "bbox": [ + 215, + 146, + 784, + 840 + ], + "page_idx": 16 + }, + { + "type": "header", + "text": "PartSTAD: 2D-to-3D Part Segmentation Task Adaptation", + "bbox": [ + 343, + 114, + 730, + 128 + ], + "page_idx": 16 + }, + { + "type": "page_number", + "text": "17", + "bbox": [ + 767, + 116, + 784, + 126 + ], + "page_idx": 16 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "58. Xu, J., Liu, S., Vahdat, A., Byeon, W., Wang, X., De Mello, S.: Open-vocabulary panoptic segmentation with text-to-image diffusion models. In: CVPR (2023)", + "59. Xu, M., Ding, R., Zhao, H., Qi, X.: PACov: Position adaptive convolution with dynamic kernel assembling on point clouds. In: CVPR (2021)", + "60. Yan, X., Zheng, C., Li, Z., Wang, S., Cui, S.: PointANSL: Robust point clouds processing using nonlocal neural networks with adaptive sampling. In: CVPR (2020)", + "61. Yang, B., Wang, J., Clark, R., Hu, Q., Wang, S., Markham, A., Trigoni, N.: Learning object bounding boxes for 3D instance segmentation on point clouds. In: NeurIPS (2019)", + "62. Yang, X., Wang, X.: Diffusion model as representation learner. In: ICCV (2023)", + "63. Yang, Y., Wu, X., He, T., Zhao, H., Liu, X.: SAM3D: Segment anything in 3D scenes. arXiv preprint arXiv:2306.03908 (2023)", + "64. Yi, K.M., Trulls, E., Ono, Y., Lepetit, V., Salzmann, M., Fua, P.: Learning to find good correspondences. In: CVPR (2018)", + "65. Yi, L., Huang, H., Liu, D., Kalogerakis, E., Su, H., Guibas, L.: Deep part induction from articulated object pairs. ACM TOG (2018)", + "66. Yi, L., Zhao, W., Wang, H., Sung, M., Guibas, L.J.: GSPN: Generative shape proposal network for 3D instance segmentation in point cloud. In: CVPR (2019)", + "67. Yu, X., Tang, L., Rao, Y., Huang, T., Zhou, J., Lu, J.: Point-BERT: Pre-training 3D point cloud transformers with masked point modeling. In: CVPR (2022)", + "68. Zamir, A.R., Sax, A., Shen, W.B., Guibas, L.J., Malik, J., Savarese, S.: Taskonomy: Disentangling task transfer learning. In: CVPR (2018)", + "69. Zhang, B., Wonka, P.: Point cloud instance segmentation using probabilistic embeddings. In: CVPR (2021)", + "70. Zhang, J., Dong, R., Ma, K.: CLIP-FO3D: Learning free open-world 3D scene representations from 2D dense clip. arXiv preprint arXiv:2303.04748 (2023)", + "71. Zhang, R., Wang, L., Wang, Y., Gao, P., Li, H., Shi, J.: Parameter is not all you need: Starting from non-parametric networks for 3D point cloud analysis. In: CVPR (2023)", + "72. Zhang, Y., Ling, H., Gao, J., Yin, K., Lafleche, J.F., Barriuso, A., Torralba, A., Fidler, S.: DatasetGAN: Efficient labeled data factory with minimal human effort. In: CVPR (2021)", + "73. Zhao, H., Jiang, L., Jia, J., Torr, P.H., Koltun, V.: Point transformer. In: ICCV (2021)", + "74. Zhao, N., Chua, T.S., Lee, G.H.: Few-shot 3D point cloud semantic segmentation. In: CVPR (2021)" + ], + "bbox": [ + 215, + 146, + 784, + 645 + ], + "page_idx": 17 + }, + { + "type": "page_number", + "text": "18", + "bbox": [ + 217, + 114, + 235, + 126 + ], + "page_idx": 17 + }, + { + "type": "header", + "text": "H. Kim and M. Sung", + "bbox": [ + 271, + 114, + 413, + 128 + ], + "page_idx": 17 + } +] \ No newline at end of file diff --git a/2024/PartSTAD_ 2D-to-3D Part Segmentation Task Adaptation/bbdaa935-8663-4268-b38a-384a769ccb82_model.json b/2024/PartSTAD_ 2D-to-3D Part Segmentation Task Adaptation/bbdaa935-8663-4268-b38a-384a769ccb82_model.json new file mode 100644 index 0000000000000000000000000000000000000000..b37785d364073c8d6eaafe9082a7badc9e4c715c --- /dev/null +++ b/2024/PartSTAD_ 2D-to-3D Part Segmentation Task Adaptation/bbdaa935-8663-4268-b38a-384a769ccb82_model.json @@ -0,0 +1,2546 @@ +[ + [ + { + "type": "title", + "bbox": [ + 0.239, + 0.141, + 0.768, + 0.187 + ], + "angle": 0, + "content": "PartSTAD: 2D-to-3D Part Segmentation Task Adaptation" + }, + { + "type": "text", + "bbox": [ + 0.36, + 0.212, + 0.644, + 0.229 + ], + "angle": 0, + "content": "Hyunjin \\(\\mathrm{Kim}^{1\\dagger}\\) and Minhyuk \\(\\mathrm{Sung}^{2}\\)" + }, + { + "type": "text", + "bbox": [ + 0.393, + 0.239, + 0.61, + 0.253 + ], + "angle": 0, + "content": "1 KRAFTON Inc., South Korea" + }, + { + "type": "text", + "bbox": [ + 0.423, + 0.254, + 0.581, + 0.267 + ], + "angle": 0, + "content": "2 KAIST, South Korea" + }, + { + "type": "text", + "bbox": [ + 0.335, + 0.269, + 0.668, + 0.282 + ], + "angle": 0, + "content": "rlaguswls98@krafton.com, mhsung@kaist.ac.kr" + }, + { + "type": "image", + "bbox": [ + 0.221, + 0.291, + 0.789, + 0.427 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.215, + 0.44, + 0.789, + 0.511 + ], + "angle": 0, + "content": "Figure 1: We introduce PARTSTAD, a novel few-shot 3D point cloud part segmentation method that leverages 2D-to-3D task adaptation. By obtaining 2D segmentation masks in multi-view images from GLIP [25] and SAM [20] and optimizing the mask weights for 3D segmentation as a learning objective, it can successfully predict fine-grained parts with accurate boundaries, as shown in the figure above." + }, + { + "type": "text", + "bbox": [ + 0.261, + 0.53, + 0.744, + 0.794 + ], + "angle": 0, + "content": "Abstract. We introduce PARTSTAD, a method designed for the task adaptation of 2D-to-3D segmentation lifting. Recent studies have highlighted the advantages of utilizing 2D segmentation models to achieve high-quality 3D segmentation through few-shot adaptation. However, previous approaches have focused on adapting 2D segmentation models for domain shift to rendered images and synthetic text descriptions, rather than optimizing the model specifically for 3D segmentation. Our proposed task adaptation method finetunes a 2D bounding box prediction model with an objective function for 3D segmentation. We introduce weights for 2D bounding boxes for adaptive merging and learn the weights using a small additional neural network. Additionally, we incorporate SAM, a foreground segmentation model on a bounding box, to improve the boundaries of 2D segments and consequently those of 3D segmentation. Our experiments on the PartNet-Mobility dataset show significant improvements with our task adaptation approach, achieving a \\(7.0\\%\\)p increase in mIoU and a \\(5.2\\%\\)p improvement in \\(\\mathrm{mAP}_{50}\\) for semantic and instance segmentation compared to the SotA few-shot 3D segmentation model. The code is available at https://github.com/KAIST-Visual-AI-Group/PartSTAD." + }, + { + "type": "text", + "bbox": [ + 0.262, + 0.806, + 0.728, + 0.821 + ], + "angle": 0, + "content": "Keywords: part segmentation \\(\\cdot\\) few-shot learning \\(\\cdot\\) 3D deep learning" + }, + { + "type": "page_footnote", + "bbox": [ + 0.221, + 0.825, + 0.625, + 0.839 + ], + "angle": 0, + "content": "This work was conducted when the author was at KAIST." + } + ], + [ + { + "type": "page_number", + "bbox": [ + 0.218, + 0.116, + 0.23, + 0.127 + ], + "angle": 0, + "content": "2" + }, + { + "type": "header", + "bbox": [ + 0.272, + 0.115, + 0.415, + 0.129 + ], + "angle": 0, + "content": "H. Kim and M. Sung" + }, + { + "type": "title", + "bbox": [ + 0.218, + 0.145, + 0.377, + 0.162 + ], + "angle": 0, + "content": "1 Introduction" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.176, + 0.787, + 0.312 + ], + "angle": 0, + "content": "3D segmentation has been a subject of extensive research in computer vision due to its central role in various downstream applications involving the understanding of shape structure, functionality, mobility, and semantics. However, the limited availability of annotated 3D shapes has remained a bottleneck in achieving generalizability in learned segmentation models for diverse 3D data. Annotating 3D segmentation is particularly labor-intensive, time-consuming, and requires expertise in handling 3D models. For this reason, the scale of 3D part-level annotations remains in the tens of thousands [37], while the scale of 2D annotation datasets in the image domain, for example, exceeds a million [20, 30]." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.312, + 0.787, + 0.447 + ], + "angle": 0, + "content": "Recent research [2,32] has illuminated the potential of visual-language models bridging textual descriptions with images in accomplishing zero-shot or few-shot 3D segmentation results. The basic idea is to render a 3D model from various viewpoints, conduct 2D detection or segmentation on the rendered images, and then aggregate the 2D segmentation results into a 3D representation using either a voting mechanism [32] or a label propagation scheme [2]. This approach is not only effective in enabling zero-shot and few-shot 3D segmentation but also offers an advantage in that the set of part names does not need to be predefined in training but can be determined at test time." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.448, + 0.787, + 0.629 + ], + "angle": 0, + "content": "PartSLIP [32] is a notable example that achieves 3D part segmentation results comparable to fully supervised methods while adapting a pretrained 2D detection model with few-shot training for 3D segmentation. It leverages the pretrained GLIP [25] model for 2D detection, associating the output 2D bounding boxes with one of the tokens (part names) provided in the input prompt. Its key component is the GLIP finetuning process, which introduces a small number of learnable parameters to the frozen GLIP model and trains them using few-shot synthetic images and texts employed in the 3D segmentation pipeline. As GLIP was initially trained with real photos and natural language descriptions of objects by humans, this adaptation to rendered images and the unconventional description (a sequence of part names) results in a substantial improvement in the 3D segmentation task." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.63, + 0.787, + 0.78 + ], + "angle": 0, + "content": "Despite promising initiatives, it is important to note that the previous approach was limited to achieving domain adaptation. In contrast, we emphasize that in the 2D-to-3D segmentation lifting task, not only does the data domain change, but the task itself also shifts from 2D segmentation to 3D segmentation. Thus, the process of applying the pretrained model to the new task requires task adaptation, involving modifying the model with an objective function associated with the new task. Particularly for 3D segmentation, it is necessary to integrate 2D segmentation results from multiple viewpoints into a coherent 3D representation. Therefore, it is crucial to control the noise of 2D bounding box predictions in the context of its impact on the final 3D segmentation after integration." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.78, + 0.787, + 0.841 + ], + "angle": 0, + "content": "Specifically, our Part Segmentation Task ADaptation method, PARTSTAD, adapts the pretrained GLIP model with a relaxed 3D mIoU loss while incorporating bounding boxes across all different views. Instead of typical finetuning, we draw inspiration from recent work [16, 32] that introduces small learnable" + } + ], + [ + { + "type": "header", + "bbox": [ + 0.344, + 0.115, + 0.732, + 0.129 + ], + "angle": 0, + "content": "PartSTAD: 2D-to-3D Part Segmentation Task Adaptation" + }, + { + "type": "page_number", + "bbox": [ + 0.775, + 0.117, + 0.785, + 0.127 + ], + "angle": 0, + "content": "3" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.147, + 0.788, + 0.312 + ], + "angle": 0, + "content": "parameters to the existing model while keeping existing parameters frozen. As a result, we train a small MLP using 2D bounding box features extracted from the pretrained GLIP model. Since the relaxed mIoU loss is non-differentiable with respect to the bounding box positions, we suggest predicting a weight for each bounding box instead of adjusting its position. This results in minimal modification in the 2D-to-3D voting scheme while achieving a significant improvement in 3D segmentation, even with few-shot training, e.g., with eight objects per category, as done in PartSLIP. To further enhance 3D segmentation performance, we leverage SAM [20], a 2D instance segmentation model, to segment the foreground region within each bounding box, obtaining an accurate boundary for each 2D segment." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.313, + 0.788, + 0.419 + ], + "angle": 0, + "content": "In our experiments on the PartNet-Mobility [56] dataset, we showcase the superior performance of our method in comparison to recent zero-shot/few-shot 3D semantic and instance segmentation methods that leverage 2D segmentation models. In contrast to the SotA few-shot method, our approach achieves a \\(7.0\\% \\mathbf{p}\\) improvement in semantic segmentation mIoU and a \\(5.2\\% \\mathbf{p}\\) improvement in instance segmentation \\(\\mathrm{mAP}_{50}\\). These improvements are consistent across all object classes." + }, + { + "type": "title", + "bbox": [ + 0.216, + 0.443, + 0.388, + 0.459 + ], + "angle": 0, + "content": "2 Related Work" + }, + { + "type": "title", + "bbox": [ + 0.216, + 0.477, + 0.502, + 0.492 + ], + "angle": 0, + "content": "2.1 Supervised 3D Segmentation" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.503, + 0.788, + 0.684 + ], + "angle": 0, + "content": "Given the availability of open segment-annotated 3D datasets [3,5,7,8,12,14,37, 56], 3D segmentation has been extensively researched in the last several years, focusing on the development of novel network architectures for supervised learning. Regarding the architecture of semantic segmentation, diverse models have been explored, including PointNet [42] and its variants [36,43,44,57,71], those using CNN [27,34,49,59], GCN [17,28,53], and Transformers [55,60,67,73]. For instance segmentation, various approaches based on proposing 3D bounding boxes [15,61,66], clustering learned features [11,19,31,38,50,51,69,74], and combining both semantic and instance segmentations [29,40,46,52] have been introduced. Despite these great advances, the performance of supervised methods has been limited by the scale of 3D datasets; the largest part-annotated 3D dataset, PartNet [37], includes fewer than \\(30\\mathrm{k}\\) models." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.685, + 0.788, + 0.76 + ], + "angle": 0, + "content": "There have been some attempts to overcome the limitation by leveraging additional weak supervisions, such as text descriptions of objects [21], 3D bounding box annotations [10], sparse labels [26, 35], and geometric priors [23, 24]. However, the scale of this additional data has also been limited to tens of thousands, while datasets in the image domain are on the scale of millions." + }, + { + "type": "title", + "bbox": [ + 0.216, + 0.784, + 0.666, + 0.799 + ], + "angle": 0, + "content": "2.2 3D Segmentation Using Vision-Language Models" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.81, + 0.787, + 0.84 + ], + "angle": 0, + "content": "Recent work introduced ideas about exploiting vision-language models (VLM) for 3D segmentation. The advantage of utilizing the learned visual-language" + } + ], + [ + { + "type": "page_number", + "bbox": [ + 0.218, + 0.116, + 0.23, + 0.127 + ], + "angle": 0, + "content": "4" + }, + { + "type": "header", + "bbox": [ + 0.272, + 0.115, + 0.415, + 0.129 + ], + "angle": 0, + "content": "H. Kim and M. Sung" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.147, + 0.788, + 0.253 + ], + "angle": 0, + "content": "grounding in 3D segmentation lies in the strong generalizability to arbitrary 3D models with zero-shot or few-shot training, and also in the open vocabulary understanding that enables labeling points without specifying the set of part or object names during training. For a VLM, CLIP [45] has been adapted in multiple previous works to lift the 2D grounding to the 3D, detecting parts in an object [13] and objects in a scene [9,39,47,70], while the results are limited to highlighting some regions without providing clear segment boundaries." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.256, + 0.788, + 0.438 + ], + "angle": 0, + "content": "To obtain precise segments in 3D, other 2D object detection and segmentation models, such as SAM [20], GLIP [25], and GroundingDINO [33], are also leveraged. SAM3D [63] was an example of directly lifting the 2D masks from SAM to 3D with a bottom-up merging approach, although it was limited to segmenting the parts, not labeling them. SA3D [6] utilized NeRF representation and allowed user interaction for 3D segmentation, but it can be inefficient when the 3D objects are given as a mesh, point cloud, or other common 3D representations. OpenMask3D [47] and OpenIns3D [18] are concurrent works, both of which combine multiple foundation models, SAM [20] and CLIP [45] for OpenMask3D, and GroundingDINO [33] and LISA [22] for OpenIns3D. These works focus on segmenting objects in 3D scenes, while we aim to do finer-grained segmentation, finding parts in 3D objects." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.441, + 0.788, + 0.562 + ], + "angle": 0, + "content": "PartSLIP [32] and SATR [2] are notable examples that achieve part segmentation of 3D objects based on a 2D object detection model, GLIP [25]. They obtain 2D bounding boxes on the rendered images from multiple views and integrate them over the 3D object with different merging schemes. The finetuning of GLIP introduced by PartSLIP particularly leads to a performance improvement, but it is limited to domain adaptation, not task adaptation. Building on this, we introduce a novel task adaptation technique as well as integration with SAM [20] to achieve a further significant improvement in 3D segmentation accuracy." + }, + { + "type": "title", + "bbox": [ + 0.216, + 0.601, + 0.402, + 0.617 + ], + "angle": 0, + "content": "2.3 Task Adaptation" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.644, + 0.788, + 0.841 + ], + "angle": 0, + "content": "Taskonomy [68] is a seminal work that introduced the concept of task transfer learning for the first time, enabling the transformation of an image dense prediction model trained for a base task to perform other tasks through finetuning. Pruksachatkun et al. [41] also introduced a similar idea for language-domain tasks. Later, task adaptation has been extensively studied to leverage features learned from generative models, such as GAN [72] and diffusion models [4, 48, 58, 62], adapting them with a small network for various tasks, including image segmentation [4, 58, 72], depth prediction [48], and keypoint detection [62]. For 3D tasks, Abdelreheem et al. [1] are notable examples of using vision language-grounding models for 3D correspondence but are limited to directly applying 2D priors without finetuning. We introduce a novel task adaptation technique that lifts 2D segments to 3D by adapting the 2D features during the training of a small network with a task-specific objective function." + } + ], + [ + { + "type": "header", + "bbox": [ + 0.343, + 0.115, + 0.732, + 0.129 + ], + "angle": 0, + "content": "PartSTAD: 2D-to-3D Part Segmentation Task Adaptation" + }, + { + "type": "page_number", + "bbox": [ + 0.775, + 0.117, + 0.785, + 0.127 + ], + "angle": 0, + "content": "5" + }, + { + "type": "title", + "bbox": [ + 0.216, + 0.145, + 0.547, + 0.163 + ], + "angle": 0, + "content": "3 Background and Preliminaries" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.178, + 0.784, + 0.253 + ], + "angle": 0, + "content": "Recent research [2,13,32] has demonstrated how vision-language models designed for image object detection and segmentation can be applied to segment parts in 3D objects. The main idea involves rendering a 3D model from various viewpoints, performing 2D detection or segmentation for the rendered image at each view, and then combining the 2D segmentation results over the 3D model." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.253, + 0.788, + 0.51 + ], + "angle": 0, + "content": "Specifically, SATR [2] utilizes GLIP [25] as the 2D detection model. To obtain 2D bounding boxes for each of the specified part names, it feeds a concatenated list of these part names (separated by commas) as the input prompt to the pretrained GLIP model, along with a rendered image. The 2D bounding boxes are then mapped to a single face of the input 3D mesh, located at the center of the respective 2D bounding box, and the part names are propagated throughout the entire mesh based on geodesic distances. Similarly, PartSLIP [32] also leverages GLIP [25] for 2D detection but employs a different voting scheme for the 2D bounding box integration over the 3D object, which utilizes the oversegmentation of the input point cloud. One key distinction between PartSLIP and SATR is that PartSLIP proposes to modify the GLIP model for domain adaptation to the rendered images and text prompts used in 3D segmentation. Specifically, the text prompt is given as a concatenation of part names in the PartSLIP pipeline, which is not a typical and natural description of the object. Therefore, it learns a category-specific constant offset vector for the language embedding feature in the given pairs of prompts and rendered images, while keeping the GLIP parameters frozen." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.511, + 0.786, + 0.556 + ], + "angle": 0, + "content": "Our work builds upon the PartSLIP pipeline. In the following subsections, we describe the details of its major components: the voting scheme, and the finetuning process." + }, + { + "type": "title", + "bbox": [ + 0.215, + 0.58, + 0.377, + 0.596 + ], + "angle": 0, + "content": "3.1 PartSLIP [32]" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.606, + 0.785, + 0.652 + ], + "angle": 0, + "content": "The main technical components of PartSLIP are the 1) voting scheme based on super points and 2) adaptation to text prompts1. We describe the details of each component below." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.667, + 0.788, + 0.789 + ], + "angle": 0, + "content": "Voting for Super Points. PartSLIP first over-segments the input 3D object represented as a point cloud \\(\\mathcal{P}\\) into a set of super points \\(P_{i} \\subseteq \\mathcal{P}\\), which provide geometric priors for the boundaries of the parts. The 3D object is rendered into \\(K\\) viewpoints, and the 2D bounding boxes for each image are predicted using GLIP [25]. Let \\(\\mathcal{B}\\) denote the entire set of 2D bounding boxes from all views, and \\(B_{jk} \\subseteq \\mathcal{B}\\) indicates a subset that includes the bounding boxes classified as the \\(j\\)-th part label from the \\(k\\)-th view. \\(\\mathrm{V}_k: \\mathcal{P} \\to \\{0,1\\}\\) is a function indicating whether the input point is visible from the \\(k\\)-th viewpoint. Also, for each 2D" + }, + { + "type": "page_footnote", + "bbox": [ + 0.218, + 0.797, + 0.788, + 0.84 + ], + "angle": 0, + "content": "1 Although multi-view feature aggregation is also introduced, the improvement by this component is marginal, and the author of PartSLIP also did not include this component in their official code. Therefore, we will omit this part in our review." + } + ], + [ + { + "type": "page_number", + "bbox": [ + 0.218, + 0.116, + 0.23, + 0.127 + ], + "angle": 0, + "content": "6" + }, + { + "type": "header", + "bbox": [ + 0.272, + 0.115, + 0.415, + 0.129 + ], + "angle": 0, + "content": "H. Kim and M. Sung" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.147, + 0.788, + 0.222 + ], + "angle": 0, + "content": "bounding box \\( b \\in \\mathcal{B} \\), \\( \\mathrm{I}_b : \\mathcal{P} \\to \\{0,1\\} \\) is a function indicating whether the input point is included in the bounding box \\( b \\). Given these, the voting aggregating the semantic labels of the 2D bounding boxes to the 3D point cloud is performed by calculating the following score \\( s_{ij} \\) for each pair of the \\( i \\)-th super point and the \\( j \\)-th label:" + }, + { + "type": "equation", + "bbox": [ + 0.335, + 0.233, + 0.786, + 0.27 + ], + "angle": 0, + "content": "\\[\ns _ {i j} = \\frac {\\sum_ {k} \\sum_ {p \\in P _ {i}} \\mathrm {V} _ {k} (p) \\left(\\max _ {b \\in \\left(B _ {j k} \\cup \\{b _ {\\emptyset} \\}\\right)} \\mathrm {I} _ {b} (p)\\right)}{\\sum_ {k} \\sum_ {p \\in P _ {i}} \\mathrm {V} _ {k} (p)}, \\tag {1}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.215, + 0.278, + 0.789, + 0.4 + ], + "angle": 0, + "content": "where \\( b_{\\emptyset} \\) is a null bounding box, and we assume that \\( \\mathrm{I}_{b_{\\emptyset}} \\) for the null bounding box returns zero for any input point (simplifying notations for the case when \\( B_{jk} = \\emptyset \\)). Note that this \\( s_{ij} \\in [0,1] \\) denotes the ratio of visible points from each view included in any of the 2D bounding boxes labeled with the \\( j \\)-th label. Finally, for each \\( i \\)-th super point, the label with the highest score \\( s_{ij} \\) across all the labels is assigned to the super point. Exceptionally, if the highest score is less than a null label threshold \\( s_{\\emptyset} \\), which is set to 0.5 in our experiments, the null label is assigned to the super point." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.412, + 0.789, + 0.563 + ], + "angle": 0, + "content": "GLIP Finetuning. While GLIP demonstrates impressive generalizability for unseen images and prompts, it still has limitations when it comes to handling synthetic images and text prompts, such as the rendered images and sequences of part names used in the PartSLIP pipeline. To address this, the authors of PartSLIP propose adding small, learnable parameters to the GLIP model while keeping all the pretrained GLIP parameters frozen. These new parameters are learned with very few-shot examples: 8 annotated 3D objects per category. They represent offset feature vectors for each part name and remain constant as global variables for each category, rather than changing for each input. This adaptation of GLIP significantly improves the accuracy of 3D part segmentation." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.577, + 0.788, + 0.668 + ], + "angle": 0, + "content": "Instance Segmentation. The voting scheme described above is designed for semantic segmentation, but PartSLIP also introduces a straightforward merging-based method to achieve instance segmentation based on semantic segmentation. To obtain instance segments, it merges adjacent super points under two conditions: 1) they share the same semantic label, and 2) they are either both included or both excluded for all bounding boxes." + }, + { + "type": "title", + "bbox": [ + 0.215, + 0.689, + 0.673, + 0.708 + ], + "angle": 0, + "content": "4 PARTSTAD: Task Adaptation for 2D-to-3D" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.72, + 0.788, + 0.81 + ], + "angle": 0, + "content": "PartSLIP has demonstrated that adapting a 2D detection model for 3D segmentation with a small set of trainable parameters can significantly enhance 3D segmentation performance. However, we observe that PartSLIP's adaptation is limited to domain adaptation, where it learns new parameters using synthetic images and data used in the PartSLIP framework without altering the objective function during training." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.81, + 0.789, + 0.84 + ], + "angle": 0, + "content": "When applying the 2D segmentation model to 3D segmentation, it is important to note that not only does the data domain change, but the task itself" + } + ], + [ + { + "type": "header", + "bbox": [ + 0.344, + 0.115, + 0.731, + 0.129 + ], + "angle": 0, + "content": "PartSTAD: 2D-to-3D Part Segmentation Task Adaptation" + }, + { + "type": "page_number", + "bbox": [ + 0.775, + 0.117, + 0.785, + 0.127 + ], + "angle": 0, + "content": "7" + }, + { + "type": "image", + "bbox": [ + 0.217, + 0.143, + 0.788, + 0.257 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.214, + 0.276, + 0.788, + 0.415 + ], + "angle": 0, + "content": "Figure 2: Overall pipeline of PARTSTAD. Our approach begins by rendering the provided 3D point cloud from multiple viewpoints. Subsequently, we extract 2D bounding boxes for its parts using GLIP [25] (Bounding Box Prediction); note that we utilize the finetuned GLIP model from PartSLIP [32]. Following this, we convert the bounding boxes into segmentation masks using SAM [20], extracting the foreground region for each bounding box (SAM Mask Integration). Next, we predict weights for all the masks and adaptively combine them into a 3D representation (2D-to-3D task adaptation). The final step involves obtaining the segmentation label for the input point cloud. The GLIP and SAM models are frozen, while only our novel weight prediction network is trained per category in a few-shot setting (8 objects)." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.446, + 0.785, + 0.492 + ], + "angle": 0, + "content": "changes. Therefore, through the process of incorporating 2D bounding boxes into the 3D object, the 2D bounding box prediction model must align with our ultimate goal: 3D part segmentation." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.492, + 0.788, + 0.645 + ], + "angle": 0, + "content": "To address this, we introduce Part Segmentation Task ADaptation method, PA STAD, a task adaptation approach designed to finetune the 2D segmentation model for 3D part segmentation. In Sec. 4.1, we introduce our objective function to finetune the GLIP for the 3D segmentation task. In Sec. 4.2, we describe how we adapt the pretrained GLIP model with the new objective function by introducing additional learnable parameters and modifying the voting scheme (Eq. 1). PartSLIP also faces a performance limitation due to the GLIP that provides 2D bounding boxes, not 2D segments. In Sec. 4.3, we also propose to combine the SAM [20] foreground mask for each bounding box to obtain more precise 2D segments and, consequently, further improve the 3D segmentation." + }, + { + "type": "title", + "bbox": [ + 0.216, + 0.667, + 0.396, + 0.682 + ], + "angle": 0, + "content": "4.1 3D mRIoU Loss" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.694, + 0.788, + 0.785 + ], + "angle": 0, + "content": "We begin by describing the loss function that we use for adapting the GLIP to the 3D segmentation task. Let \\(\\mathbf{l}_j\\) and \\(\\hat{\\mathbf{l}}_j \\in \\{0,1\\}^{|\\mathcal{P}|}\\) be binary vectors indicating whether each point \\(p\\) in \\(\\mathcal{P}\\) has the \\(j\\)-th label in the ground truth and label prediction, respectively. The mean Intersection over Union (mIoU) for the predicted 3D segmentation, the standard evaluation metric for 3D segmentation, is defined as follows:" + }, + { + "type": "equation", + "bbox": [ + 0.329, + 0.795, + 0.786, + 0.839 + ], + "angle": 0, + "content": "\\[\n\\operatorname {m I o U} \\left(\\left\\{\\mathbf {l} _ {j} \\right\\}, \\left\\{\\hat {\\mathbf {l}} _ {j} \\right\\}\\right) = \\frac {1}{M} \\sum_ {j = 1} ^ {M} \\frac {\\mathbf {l} _ {j} ^ {\\top} \\hat {\\mathbf {l}} _ {j}}{\\left\\| \\mathbf {l} _ {j} \\right\\| _ {1} + \\left\\| \\hat {\\mathbf {l}} _ {j} \\right\\| _ {1} - \\left. \\mathbf {l} _ {j} ^ {\\top} \\hat {\\mathbf {l}} _ {j} \\right.}, \\tag {2}\n\\]" + } + ], + [ + { + "type": "page_number", + "bbox": [ + 0.218, + 0.116, + 0.23, + 0.127 + ], + "angle": 0, + "content": "8" + }, + { + "type": "header", + "bbox": [ + 0.272, + 0.115, + 0.415, + 0.129 + ], + "angle": 0, + "content": "H. Kim and M. Sung" + }, + { + "type": "text", + "bbox": [ + 0.216, + 0.147, + 0.457, + 0.161 + ], + "angle": 0, + "content": "where \\(M\\) is the number of labels." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.163, + 0.788, + 0.239 + ], + "angle": 0, + "content": "Our goal is to use the 3D mIoU as the objective function directly when adapting the GLIP model. However, the mIoU is non-differentiable. Thus, we employ a relaxed version known as mean Relaxed IoU (mRIOU) [24,65], which simply allows the predicted indicator vectors \\(\\hat{\\mathbf{l}}_j\\) to be non-binary: \\(\\hat{\\mathbf{l}}_j\\in [0,1]^{|\\mathcal{P}|}\\). The loss \\(\\mathcal{L}_{\\mathrm{mRIOU}}\\) is specifically defined as follows:" + }, + { + "type": "equation", + "bbox": [ + 0.389, + 0.245, + 0.786, + 0.263 + ], + "angle": 0, + "content": "\\[\n\\mathcal {L} _ {\\mathrm {m R I o U}} = 1 - \\operatorname {m I o U} \\left(\\left\\{\\mathbf {l} _ {j} \\right\\}, \\left\\{\\hat {\\mathbf {l}} _ {j} \\right\\}\\right). \\tag {3}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.268, + 0.785, + 0.314 + ], + "angle": 0, + "content": "In the supplementary material, we demonstrate that the choice of the loss function, as compared to other alternatives such as the cross-entropy loss, is crucial to achieving a substantial improvement in our task adaptation." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.314, + 0.785, + 0.378 + ], + "angle": 0, + "content": "Note that in the PartSLIP framework, where super points are utilized in 3D segmentation, and \\( s_{ij} \\in [0,1] \\) indicates the likelihood of the \\( i \\)-th super point belonging to the \\( j \\)-label, the mRIoU can be calculated by defining \\( \\hat{\\mathbf{l}}_j \\) based on \\( \\mathbf{s}_j = [s_{1j}, s_{2j}, \\dots]^{\\top} \\) as follows:" + }, + { + "type": "equation", + "bbox": [ + 0.465, + 0.396, + 0.786, + 0.415 + ], + "angle": 0, + "content": "\\[\n\\hat {\\mathbf {l}} _ {j} = \\mathbf {M} \\mathbf {s} _ {j}, \\tag {4}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.421, + 0.785, + 0.453 + ], + "angle": 0, + "content": "where \\(\\mathbf{M} \\in \\{0,1\\}^{|\\mathcal{P}| \\times |\\{P_i\\}|}\\) is a binary matrix that describes the memberships from each point to the super points." + }, + { + "type": "title", + "bbox": [ + 0.216, + 0.47, + 0.536, + 0.485 + ], + "angle": 0, + "content": "4.2 Bounding Box Weight Prediction" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.491, + 0.788, + 0.642 + ], + "angle": 0, + "content": "A typical approach for adapting a pretrained model to a new task would be finetuning the model using an objective function tailored to that new task. However, the finetuning method generally fails to yield meaningful improvements, especially when the goal is to achieve few-shot adaptation (e.g., with only 8 annotated 3D objects per class, as done in our case), while the pretrained model has been trained on an extensive dataset. To address this challenge, recent works like LoRA [16] and PartSLIP [32] have introduced the concept of adding small learnable parameters while keeping the existing pretrained model parameters frozen. This approach not only makes training very efficient but also enables the model to generalize effectively to unseen data." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.643, + 0.788, + 0.718 + ], + "angle": 0, + "content": "While we also follow this approach, our specific challenge arises from the fact that when we use the 3D mRIoU (Eq. 3) as the objective function for 3D segmentation, it becomes non-differentiable with respect to the 2D bounding box location, which is the output of GLIP. This non-differentiability arises from the computation of the score \\( s_{ij} \\) (Eq. 1) for each pair of super point and label." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.718, + 0.788, + 0.795 + ], + "angle": 0, + "content": "To address this, we propose training a small network that does not refine the location but instead predicts a weight for each 2D bounding box, taking the learned features of the boxes as input. If we denote the output weight of the network for bounding box \\( b \\) as \\( \\mathrm{W}(b) \\in \\mathbb{R}^{+} \\), the score \\( s_{ij} \\) for a pair of super point and label is modified to \\( \\tilde{s}_{ij} \\) as follows:" + }, + { + "type": "equation", + "bbox": [ + 0.317, + 0.801, + 0.786, + 0.839 + ], + "angle": 0, + "content": "\\[\n\\tilde {s} _ {i j} = \\frac {\\sum_ {k} \\sum_ {p \\in P _ {i}} \\mathrm {V} _ {k} (p) \\left(\\max _ {b \\in \\left(B _ {j k} \\cup \\{b _ {\\emptyset} \\}\\right)} \\mathrm {I} _ {b} (p) \\mathrm {W} (b)\\right)}{\\sum_ {k} \\sum_ {p \\in P _ {i}} \\mathrm {V} _ {k} (p)}. \\tag {5}\n\\]" + } + ], + [ + { + "type": "header", + "bbox": [ + 0.343, + 0.115, + 0.733, + 0.131 + ], + "angle": 0, + "content": "PartSTAD: 2D-to-3D Part Segmentation Task Adaptation" + }, + { + "type": "page_number", + "bbox": [ + 0.775, + 0.117, + 0.785, + 0.127 + ], + "angle": 0, + "content": "9" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.147, + 0.788, + 0.222 + ], + "angle": 0, + "content": "Note that the change is simply to multiply the weight \\( \\mathrm{W}(b) \\) by \\( \\mathrm{I}_b(p) \\), while only \\( \\mathrm{I}_b(p) \\), indicating whether the point \\( p \\) is included in bounding box \\( b \\), has been used in Eq. 1. This modified score no longer falls within the range of [0,1]. The final score \\( \\bar{s}_{ij} \\) is thus defined by normalizing \\( \\tilde{s}_{ij} \\) using the softmax function over the set of labels:" + }, + { + "type": "equation", + "bbox": [ + 0.436, + 0.23, + 0.786, + 0.264 + ], + "angle": 0, + "content": "\\[\n\\bar {s} _ {i j} = \\frac {\\exp (\\tilde {s} _ {i j})}{\\sum_ {j} \\exp (\\tilde {s} _ {i j})}. \\tag {6}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.272, + 0.788, + 0.348 + ], + "angle": 0, + "content": "We additionally propose to make the unnormalized score for the null label \\(\\tilde{s}_{\\emptyset}\\) learnable in our case, initialized with 10. We thus compute the softmax above while including \\(\\tilde{s}_{\\emptyset}\\). The label of each super point is still chosen as the label giving the highest score \\(\\bar{s}_{ij}\\), becoming null if the normalized null label score \\(\\bar{s}_{\\emptyset}\\) becomes the highest." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.361, + 0.788, + 0.482 + ], + "angle": 0, + "content": "Network Architecture. We design the weight predictor network to take the bounding box feature vectors \\(\\mathbf{f}_b\\) from the pretrained GLIP model as input. The feature vectors of all bounding boxes across all views are fed to the network at once and processed with a small shared two-layer MLP. Context normalization [64] is added in the middle of the two layers to incorporate contextual information from the global set of boxes for each box. The output of the MLP for each bounding box is further processed with the following modified ReLU function \\(\\phi(\\cdot)\\):" + }, + { + "type": "equation", + "bbox": [ + 0.423, + 0.493, + 0.785, + 0.508 + ], + "angle": 0, + "content": "\\[\n\\phi (\\mathbf {x}) = \\max (\\tau + \\mathbf {x}, 0), \\tag {7}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.215, + 0.518, + 0.684, + 0.534 + ], + "angle": 0, + "content": "where \\(\\tau\\) is a user-defined constant offset (10 in our experiments)." + }, + { + "type": "title", + "bbox": [ + 0.215, + 0.553, + 0.491, + 0.57 + ], + "angle": 0, + "content": "4.3 SAM [20] Mask Integration" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.577, + 0.788, + 0.76 + ], + "angle": 0, + "content": "Another limitation of PartSLIP is its reliance on GLIP [25] for 2D segmentation. Since GLIP is, in turn, a 2D object detection model, it produces bounding boxes instead of 2D segments, which cannot provide accurate boundaries of segments. We propose to address this issue by incorporating another pretrained 2D segmentation model, SAM [20]. SAM has the functionality of taking a bounding box as input and producing the foreground mask. Using this, we replace the set of 2D bounding boxes \\(\\mathcal{B}\\) with a set of 2D masks, while preserving the remaining steps in the framework. Note that the point-to-bounding-box membership \\(\\mathrm{I}_b\\) is changed to point-to-mask membership, and we still use the same bounding box features extracted from GLIP corresponding to each mask to train the weight prediction network. We demonstrate the effectiveness of applying the SAM mask in our experiments in Sec. 5." + }, + { + "type": "title", + "bbox": [ + 0.215, + 0.78, + 0.446, + 0.798 + ], + "angle": 0, + "content": "5 Experiment Results" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.81, + 0.787, + 0.842 + ], + "angle": 0, + "content": "In the experiments, we compare our method with the SotA 2D-prior-based zero-shot/few-shot 3D segmentation models using the PartNet-Mobility [56] dataset" + } + ], + [ + { + "type": "page_number", + "bbox": [ + 0.218, + 0.116, + 0.236, + 0.127 + ], + "angle": 0, + "content": "10" + }, + { + "type": "header", + "bbox": [ + 0.272, + 0.115, + 0.415, + 0.129 + ], + "angle": 0, + "content": "H. Kim and M. Sung" + }, + { + "type": "image", + "bbox": [ + 0.226, + 0.145, + 0.787, + 0.332 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.214, + 0.344, + 0.788, + 0.388 + ], + "angle": 0, + "content": "Figure 3: Qualitative comparison of semantic segmentation results. Our PARTSTAD segments 3D parts more precisely with clearer boundaries, even for small (Camera, Chair) and thin (Clock) parts." + }, + { + "type": "table_caption", + "bbox": [ + 0.214, + 0.399, + 0.789, + 0.47 + ], + "angle": 0, + "content": "Table 1: Quantitative comparison of the semantic segmentation results (mIoU(%) on the PartNet-Mobility dataset. Ours achieves a \\(7.0\\%\\)p improvement in average mIoU compared to PartSLIP [32], the SotA few-shot 3D segmentation method, while consistently increasing mIoU across all categories. Please refer to the supplementary material for the complete table with results for all 45 categories." + }, + { + "type": "table", + "bbox": [ + 0.216, + 0.482, + 0.789, + 0.605 + ], + "angle": 0, + "content": "
MethodmIoUStorage FurnitureTableChairSwitchToiletLaptopUSBRemoteScissors
SATR [2]29.320.623.333.121.417.611.230.217.236.8
SATR [2]+SP34.828.928.037.737.022.112.433.428.043.0
PartSLIP [32]58.052.344.682.852.150.431.252.136.661.4
Ablation Study
w/o Weight Pred.61.956.645.085.051.956.631.557.136.660.8
w/o SAM Integ.62.154.045.783.153.049.433.653.646.867.5
PARTSTAD (Ours)65.059.547.885.357.957.534.659.953.468.5
" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.631, + 0.788, + 0.679 + ], + "angle": 0, + "content": "and OmniObject3D [54] dataset. In the supplementary material, we present additional results on the ablation study, outcomes with scanned data, and comprehensive results covering all PartNet-Mobility categories." + }, + { + "type": "title", + "bbox": [ + 0.215, + 0.697, + 0.414, + 0.713 + ], + "angle": 0, + "content": "5.1 Experiment Setup" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.719, + 0.789, + 0.842 + ], + "angle": 0, + "content": "Dataset. We use the PartNet-Mobility [56] dataset, which is also used in the experiments of PartSLIP [32]. The PartNet-Mobility [56] dataset includes 45 object categories. For each category, the training split includes 8 shapes, while the test split has a range of the number of objects from 6 to 338, totaling 1,906 objects across all 45 categories. Following PartSLIP, 8 shapes from the test set of each category are designated as the validation set. To obtain the 2D bounding boxes for each object, we use 10 images from fixed viewpoints across all our experiments. Note that the PartNet-Mobility [56] dataset is identical to the subset" + } + ], + [ + { + "type": "header", + "bbox": [ + 0.343, + 0.114, + 0.732, + 0.131 + ], + "angle": 0, + "content": "PartSTAD: 2D-to-3D Part Segmentation Task Adaptation" + }, + { + "type": "page_number", + "bbox": [ + 0.768, + 0.116, + 0.784, + 0.127 + ], + "angle": 0, + "content": "11" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.146, + 0.788, + 0.195 + ], + "angle": 0, + "content": "of PartNet-Ensembled dataset used to train and evaluate PartSLIP [32]. (The remaining part of PartNet-Ensembled was only used to train other supervised baseline models, not PartSLIP [32].)" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.208, + 0.787, + 0.254 + ], + "angle": 0, + "content": "Evaluation Metrics. As evaluation metrics, we employ the mean Intersection over Union (mIoU) metric for semantic segmentation and mean Average Precision (mAP) for instance segmentation." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.254, + 0.787, + 0.314 + ], + "angle": 0, + "content": "For semantic segmentation, the average mIoU over the entire dataset is computed by first calculating it per semantic part, averaging them across the parts in the same category, resulting in category-level mIoU, and then averaging the category-level mIoUs again across all the categories." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.314, + 0.787, + 0.435 + ], + "angle": 0, + "content": "For instance segmentation, we compute two types of mAP: part-aware mAP and part-agnostic mAP. For part-aware mAP, similar to mIoU in semantic segmentation, we first calculate the mAP for each semantic part with instances having the part label. We then average them for each category, and then across categories. For part-agnostic mAP, we do not consider the semantic labels and directly compute the category-level mAP with part instances in the same object category, and then average them. All mAP values correspond to \\(\\mathrm{mAP}_{50}\\) with an IoU threshold of \\(50\\%\\)." + }, + { + "type": "text", + "bbox": [ + 0.215, + 0.45, + 0.807, + 0.482 + ], + "angle": 0, + "content": "Baselines. We compare our method with three baselines: PartSLIP [32], SATR [2], and SAM3D [63]." + }, + { + "type": "text", + "bbox": [ + 0.226, + 0.493, + 0.787, + 0.538 + ], + "angle": 0, + "content": "- PartSLIP [32] is the method that we adopt as our base. Note that PartSLIP uses a finetuned GLIP model, and we also employ the same model in our framework to extract bounding box features for our task adaptation." + }, + { + "type": "text", + "bbox": [ + 0.226, + 0.538, + 0.787, + 0.794 + ], + "angle": 0, + "content": "- SATR [2] is similar to PartSLIP, conducting 3D segmentation using 2D bounding boxes from GLIP [25]. However, SATR differs from PartSLIP in four key aspects: 1) it does not involve finetuning the GLIP model, 2) it does not utilize super points, 3) it is designed for mesh segmentation, not point cloud segmentation, and 4) it employs a distinct approach for integrating 2D bounding boxes into 3D (label propagation). To ensure a fair comparison between our method and SATR, we implement the following changes to SATR. First, since we employ the GLIP model finetuned by PartSLIP in our framework, we use the same finetuned model within the SATR framework. Second, we convert the point cloud input to a mesh, performing mesh segmentation with SATR, and apply the results to the input point cloud by finding the closest vertex for each point. Lastly, we introduce a version of SATR using super points, resulting in a comparison between our method and two versions of SATR: one without super points (denoted as \\(SATR\\)) and one with super points (denoted as \\(SATR + SP\\)). Since SATR does not perform instance segmentation, our comparison with it is focused on semantic segmentation." + }, + { + "type": "text", + "bbox": [ + 0.226, + 0.795, + 0.787, + 0.842 + ], + "angle": 0, + "content": "- SAM3D [63] is another zero-shot 3D segmentation method using a 2D segmentation model, not GLIP [25], but SAM [20]. Since it only performs instance segmentation but does not label each segment, we compare our" + }, + { + "type": "list", + "bbox": [ + 0.226, + 0.493, + 0.787, + 0.842 + ], + "angle": 0, + "content": null + } + ], + [ + { + "type": "page_number", + "bbox": [ + 0.218, + 0.116, + 0.236, + 0.127 + ], + "angle": 0, + "content": "12" + }, + { + "type": "header", + "bbox": [ + 0.272, + 0.115, + 0.415, + 0.129 + ], + "angle": 0, + "content": "H. Kim and M. Sung" + }, + { + "type": "text", + "bbox": [ + 0.243, + 0.147, + 0.785, + 0.176 + ], + "angle": 0, + "content": "method with it only for instance segmentation, using the part-agnostic mAP metric." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.191, + 0.788, + 0.313 + ], + "angle": 0, + "content": "Ablation Study. We also conduct an ablation study, comparing our method with two major components being ablated: the weight prediction (Sec. 4.2) and the SAM mask integration (Sec. 4.3). In the supplementary material, we also compare our method with cases using cross-entropy loss instead of mR IoU loss (Sec. 4.1) and using GLIP bounding box confidence score instead of our learned weights in the weighted voting (Sec. 4.2). Note that if both weight prediction and SAM mask integration are excluded, our PARTSTAD is identical to PartSLIP [32]." + }, + { + "type": "title", + "bbox": [ + 0.215, + 0.333, + 0.523, + 0.349 + ], + "angle": 0, + "content": "5.2 Semantic Segmentation Results" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.357, + 0.788, + 0.508 + ], + "angle": 0, + "content": "Tab. 1 and Fig. 3 present the quantitative and qualitative comparisons of part semantic segmentation results, respectively. Please refer to the supplementary material for the complete table encompassing all 45 categories. In comparison to PartSLIP [32], a SotA few-shot 3D segmentation method, we achieve a significant \\(7.0\\%\\) p mIoU improvement over the entire set of 45 categories. Furthermore, for each specific category, we consistently demonstrate improvement, surpassing by more than \\(15\\%\\) p in some categories (e.g., Remote). Qualitatively, the results also highlight the effectiveness of our task adaptation and the SAM mask integration in segmenting parts more precisely, even for small (e.g., Camera, Chair) and thin (e.g., Clock) parts of objects." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.508, + 0.799, + 0.569 + ], + "angle": 0, + "content": "SATR exhibits inferior performance, even when leveraging the finetuned GLIP model and super points (see SATR+SP results) due to the label-propagation-based 2D bounding box integration, which often fails to provide clear boundaries in 3D segmentation, as illustrated in Fig. 3." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.569, + 0.788, + 0.658 + ], + "angle": 0, + "content": "The ablation study results demonstrate the influence of each major component in our framework. The quantitative findings indicate that the weight prediction network plays a crucial role in substantial improvement, leading to a \\(3.1\\%\\) decrease in average mIoU when it is ablated. The SAM mask integration also contributes significantly, resulting in a \\(2.9\\%\\) average mIoU decrease when ablated." + }, + { + "type": "title", + "bbox": [ + 0.215, + 0.681, + 0.517, + 0.696 + ], + "angle": 0, + "content": "5.3 Instance Segmentation Results" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.705, + 0.788, + 0.84 + ], + "angle": 0, + "content": "Following the instance segmentation idea of PartSLIP, we also perform instance segmentation using our method and compare the quantitative results with those of other methods in Tab. 2 and Tab. 3, reporting part-aware mAPs and part-agnostic mAPs, respectively. Fig. 4 also displays qualitative comparisons across different methods. Please refer to the supplementary material for the complete table of all 45 categories. Similar to semantic segmentation, ours achieves \\(4.0\\%\\)p and \\(5.2\\%\\)p improvements for both average part-aware mAPs and part-agnostic mAPs, respectively. Both the weight prediction network and SAM mask integration also exhibit meaningful improvements in performance, as their influence is" + } + ], + [ + { + "type": "header", + "bbox": [ + 0.344, + 0.115, + 0.732, + 0.129 + ], + "angle": 0, + "content": "PartSTAD: 2D-to-3D Part Segmentation Task Adaptation" + }, + { + "type": "page_number", + "bbox": [ + 0.769, + 0.117, + 0.785, + 0.127 + ], + "angle": 0, + "content": "13" + }, + { + "type": "image", + "bbox": [ + 0.221, + 0.146, + 0.788, + 0.334 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.216, + 0.347, + 0.788, + 0.389 + ], + "angle": 0, + "content": "Figure 4: Qualitative comparison of instance segmentation results shows that our PARTSTAD successfully segments tiny 3D parts, such as the keys of keyboards and buttons of remote controls, with clear segment boundaries." + }, + { + "type": "table_caption", + "bbox": [ + 0.216, + 0.408, + 0.788, + 0.464 + ], + "angle": 0, + "content": "Table 2: Part-aware mAPs(\\%) of the instance segmentation results on the PartNet-Mobility dataset. Ours achieves \\(4.0\\%\\)p improvement in mean part-aware mAPs compared to PartSLIP [32], with consistent enhancements across all categories. Please refer to the supplementary material for the complete table with results for all 45 categories." + }, + { + "type": "table", + "bbox": [ + 0.217, + 0.476, + 0.787, + 0.579 + ], + "angle": 0, + "content": "
MethodmAPStorage FurnitureTableChairSwitchToiletLaptopUSBRemoteScissors
PartSLIP [32]41.629.132.682.221.236.217.820.919.923.6
Ablation Study
w/o Weight Pred.44.733.830.882.522.240.824.417.420.325.5
w/o SAM Integ.44.229.432.382.219.736.023.624.829.623.1
PARTSTAD(Ours)45.635.533.282.522.140.628.926.533.726.5
" + }, + { + "type": "table_caption", + "bbox": [ + 0.216, + 0.597, + 0.788, + 0.667 + ], + "angle": 0, + "content": "Table 3: Part-agnostic mAPs(%) of the instance segmentation results on the PartNet-Mobility dataset. Ours demonstrates \\(5.2\\%\\)p improvement in mean part-agnostic mAPs compared to PartSLIP [32], surpassing SAM3D with a mean mAP that is more than three times larger. Please refer to the supplementary material for the complete table with results for all 45 categories. (*SAM3D is a zero-shot method.)" + }, + { + "type": "table", + "bbox": [ + 0.217, + 0.679, + 0.788, + 0.791 + ], + "angle": 0, + "content": "
MethodmAPStorage FurnitureTableChairSwitchToiletLaptopUSBRemoteScissors
SAM3D* [63]12.11.210.65.55.43.71.522.61.07.2
PartSLIP [32]38.929.728.780.721.235.419.520.519.927.0
Ablation Study
w/o Weight Pred.42.635.929.281.122.240.026.515.220.329.4
w/o SAM Integ.42.634.827.681.219.736.627.427.329.624.9
PARTSTAD (Ours)44.141.728.283.322.441.034.123.533.726.2
" + }, + { + "type": "text", + "bbox": [ + 0.216, + 0.825, + 0.787, + 0.841 + ], + "angle": 0, + "content": "depicted in the ablation study results. Compared to SAM3D [63], whose part-" + } + ], + [ + { + "type": "page_number", + "bbox": [ + 0.218, + 0.116, + 0.236, + 0.127 + ], + "angle": 0, + "content": "14" + }, + { + "type": "header", + "bbox": [ + 0.272, + 0.115, + 0.415, + 0.129 + ], + "angle": 0, + "content": "H. Kim and M. Sung" + }, + { + "type": "image", + "bbox": [ + 0.229, + 0.146, + 0.788, + 0.266 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.214, + 0.277, + 0.788, + 0.334 + ], + "angle": 0, + "content": "Figure 5: Qualitative comparison of semantic segmentation results on OmniObect3D [54] dataset, a high quality real scanned 3D objects dataset. Our PART-STD predicts precise boundaries, even in the case that the object has an appearance significantly different from training data such as box category." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.359, + 0.788, + 0.435 + ], + "angle": 0, + "content": "agnostic mAPs are shown in Tab. 3, our method achieves much better performance with more than three times greater mAP. Qualitative results also illustrate the outstanding performance of our method, accurately identifying all the tiny instances of parts, such as keys on keyboards, buttons on remotes, and clear boundaries of these instances." + }, + { + "type": "title", + "bbox": [ + 0.215, + 0.453, + 0.646, + 0.469 + ], + "angle": 0, + "content": "5.4 Part Segmentation with Real Scanned Dataset" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.475, + 0.789, + 0.596 + ], + "angle": 0, + "content": "Fig. 5 displays additional results on the OmniObject3D [54] dataset, a high-quality dataset featuring real-scanned 3D objects. We present qualitative results only, as ground truth segmentation for OmniObject3D [54] is unavailable. Note that our PARTSTAD accurately identifies parts in real scans, demonstrating its effectiveness beyond synthetic data. Particularly impressive is its performance in the box category depicted in Fig. 5, where despite significant discrepancies in appearance compared to the training data, PARTSTAD achieves remarkably accurate predictions." + }, + { + "type": "title", + "bbox": [ + 0.215, + 0.616, + 0.534, + 0.633 + ], + "angle": 0, + "content": "6 Conclusion and Future Work" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.644, + 0.788, + 0.764 + ], + "angle": 0, + "content": "We presented PARTSTAD, a task adaptation method for lifting 2D segmentation to 3D. Instead of directly finetuning a 2D segmentation network, our method learns a small neural network predicting weights for each 2D bounding box with an objective function for 3D segmentation, and then performs 3D segmentation via weighted merging of the boxes in 3D. We further improve the performance of 3D segmentation by integrating SAM foreground masks for each bounding box. We achieve the SotA results in few-shot 3D part segmentation, demonstrating significant improvements in both semantic and instance segmentations." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.765, + 0.788, + 0.841 + ], + "angle": 0, + "content": "As PARTSTAD still depends on 2D representations for 3D predictions, its understanding of 3D geometry is relatively limited, and thus it cannot account for occluded or interior points, which poses a limitation. In future work, we plan to further investigate combining our approach with the direct utilization of 3D representations, such as leveraging 3D features, for more accurate predictions." + } + ], + [ + { + "type": "header", + "bbox": [ + 0.343, + 0.115, + 0.732, + 0.13 + ], + "angle": 0, + "content": "PartSTAD: 2D-to-3D Part Segmentation Task Adaptation" + }, + { + "type": "page_number", + "bbox": [ + 0.768, + 0.117, + 0.786, + 0.127 + ], + "angle": 0, + "content": "15" + }, + { + "type": "title", + "bbox": [ + 0.218, + 0.145, + 0.403, + 0.163 + ], + "angle": 0, + "content": "Acknowledgements" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.18, + 0.813, + 0.271 + ], + "angle": 0, + "content": "We sincerely thank Minghua Liu for providing the code of PartSLIP and answering questions. This work was supported by NRF grant (RS-2023-00209723), IITP grants (2022-0-00594, RS-2023-00227592, RS-2024-00399817), and Alchymist Project Program (RS-2024-00423625) funded by the Korean government (MSIT and MOTIE), and grants from the DRB-KAIST SketchTheFuture Research Center, Naver-intel, Adobe Research, Hyundai NGV, KT, and Samsung Electronics." + }, + { + "type": "title", + "bbox": [ + 0.216, + 0.297, + 0.323, + 0.314 + ], + "angle": 0, + "content": "References" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.332, + 0.787, + 0.36 + ], + "angle": 0, + "content": "1. Abdelreheem, A., Eldesokey, A., Ovsjanikov, M., Wonka, P.: Zero-shot 3D shape correspondence. In: SIGGRAPH Asia (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.361, + 0.786, + 0.389 + ], + "angle": 0, + "content": "2. Abdelreheem, A., Skorokhodov, I., Ovsjanikov, M., Wonka, P.: SATR: Zero-shot semantic segmentation of 3D shapes. In: ICCV (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.39, + 0.786, + 0.417 + ], + "angle": 0, + "content": "3. Armeni, I., Sax, S., Zamir, A.R., Savarese, S.: Joint 2D-3D-semantic data for indoor scene understanding. arXiv preprint arXiv:1702.01105 (2017)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.418, + 0.786, + 0.445 + ], + "angle": 0, + "content": "4. Baranchuk, D., Voynov, A., Rubachev, I., Khrulkov, V., Babenko, A.: Label-efficient semantic segmentation with diffusion models. In: ICLR (2022)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.446, + 0.786, + 0.488 + ], + "angle": 0, + "content": "5. Behley, J., Garbade, M., Milioto, A., Quenzel, J., Behnke, S., Stachniss, C., Gall, J.: SemanticKITTI: A dataset for semantic scene understanding of LiDAR sequences. In: ICCV (2019)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.489, + 0.786, + 0.516 + ], + "angle": 0, + "content": "6. Cen, J., Zhou, Z., Fang, J., Yang, C., Shen, W., Xie, L., Zhang, X., Tian, Q.: Segment anything in 3D with nerfs. In: NeurIPS (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.517, + 0.786, + 0.558 + ], + "angle": 0, + "content": "7. Chang, A., Dai, A., Funkhouser, T., Halber, M., Niessner, M., Savva, M., Song, S., Zeng, A., Zhang, Y.: Matterport3D: Learning from rgb-d data in indoor environments. In: 3DV (2017)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.559, + 0.786, + 0.601 + ], + "angle": 0, + "content": "8. Chang, A.X., Funkhouser, T., Guibas, L., Hanrahan, P., Huang, Q., Li, Z., Savarese, S., Savva, M., Song, S., Su, H., et al.: ShapeNet: An information-rich 3D model repository. arXiv preprint arXiv:1512.03012 (2015)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.602, + 0.786, + 0.643 + ], + "angle": 0, + "content": "9. Chen, R., Liu, Y., Kong, L., Zhu, X., Ma, Y., Li, Y., Hou, Y., Qiao, Y., Wang, W.: CLIP2Scene: Towards label-efficient 3D scene understanding by clip. In: CVPR (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.644, + 0.786, + 0.671 + ], + "angle": 0, + "content": "0. Chibane, J., Engelmann, F., Tran, T.A., Pons-Moll, G.: Box2Mask: Weakly supervised 3D semantic instance segmentation using bounding boxes. In: ECCV (2022)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.672, + 0.786, + 0.699 + ], + "angle": 0, + "content": "1. Chu, R., Chen, Y., Kong, T., Qi, L., Li, L.: ICM-3D: Instantiated category modeling for 3D instance segmentation. IEEE Robotics and Automation Letters (2021)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.7, + 0.786, + 0.728 + ], + "angle": 0, + "content": "2. Dai, A., Chang, A.X., Savva, M., Halber, M., Funkhouser, T., Nießner, M.: ScanNet: Richly-annotated 3D reconstructions of indoor scenes. In: CVPR (2017)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.729, + 0.786, + 0.756 + ], + "angle": 0, + "content": "3. Decatur, D., Lang, I., Hanocka, R.: 3D Highlighter: Localizing regions on 3D shapes via text descriptions. In: CVPR (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.757, + 0.786, + 0.811 + ], + "angle": 0, + "content": "4. Hackel, T., Savinov, N., Ladicky, L., Wegner, J.D., Schindler, K., Pollefeys, M.: SEMANTIC3D.NET: A new large-scale point cloud classification benchmark. In: ISPRS Annals of the Photogrammetry, Remote Sensing and Spatial Information Sciences (2017)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.812, + 0.786, + 0.84 + ], + "angle": 0, + "content": "5. Hou, J., Dai, A., Nießner, M.: 3D-SIS: 3D semantic instance segmentation of rgb-d scans. In: CVPR (2019)" + }, + { + "type": "list", + "bbox": [ + 0.226, + 0.332, + 0.787, + 0.84 + ], + "angle": 0, + "content": null + } + ], + [ + { + "type": "page_number", + "bbox": [ + 0.218, + 0.116, + 0.236, + 0.127 + ], + "angle": 0, + "content": "16" + }, + { + "type": "header", + "bbox": [ + 0.272, + 0.115, + 0.414, + 0.129 + ], + "angle": 0, + "content": "H. Kim and M. Sung" + }, + { + "type": "ref_text", + "bbox": [ + 0.218, + 0.147, + 0.785, + 0.189 + ], + "angle": 0, + "content": "16. Hu, E.J., Shen, Y., Wallis, P., Allen-Zhu, Z., Li, Y., Wang, S., Wang, L., Chen, W.: LoRA: Low-rank adaptation of large language models. arXiv preprint arXiv:2106.09685 (2021)" + }, + { + "type": "ref_text", + "bbox": [ + 0.218, + 0.189, + 0.785, + 0.217 + ], + "angle": 0, + "content": "17. Huang, R., Huang, C., Liu, Y., Dai, G., Kong, W.: LSGCN: Long short-term traffic prediction with graph convolutional networks. In: IJCAI (2020)" + }, + { + "type": "ref_text", + "bbox": [ + 0.218, + 0.217, + 0.785, + 0.243 + ], + "angle": 0, + "content": "18. Huang, Z., Wu, X., Chen, X., Zhao, H., Zhu, L., Lasenby, J.: OpenIns3D: Snap and lookup for 3D open-vocabulary instance segmentation. arXiv preprint (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.218, + 0.243, + 0.785, + 0.27 + ], + "angle": 0, + "content": "19. Jiang, L., Zhao, H., Shi, S., Liu, S., Fu, C.W., Jia, J.: PointGroup: Dual-set point grouping for 3D instance segmentation. In: CVPR (2020)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.27, + 0.785, + 0.311 + ], + "angle": 0, + "content": "20. Kirillov, A., Mintun, E., Ravi, N., Mao, H., Rolland, C., Gustafson, L., Xiao, T., Whitehead, S., Berg, A.C., Lo, W.Y., et al.: Segment anything. arXiv preprint arXiv:2304.02643 (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.311, + 0.785, + 0.338 + ], + "angle": 0, + "content": "21. Koo, J., Huang, I., Achlioptas, P., Guibas, L.J., Sung, M.: PartGlot: Learning shape part segmentation from language reference games. In: CVPR (2022)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.339, + 0.785, + 0.366 + ], + "angle": 0, + "content": "22. Lai, X., Tian, Z., Chen, Y., Li, Y., Yuan, Y., Liu, S., Jia, J.: LISA: Reasoning segmentation via large language model. arXiv preprint arXiv:2308.00692 (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.366, + 0.785, + 0.405 + ], + "angle": 0, + "content": "23. Lé, E.T., Sung, M., Ceylan, D., Mech, R., Boubekeur, T., Mitra, N.J.: CPFN: Cascaded primitive fitting networks for high-resolution point clouds. In: ICCV (2021)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.405, + 0.785, + 0.433 + ], + "angle": 0, + "content": "24. Li, L., Sung, M., Dubrovina, A., Yi, L., Guibas, L.J.: Supervised fitting of geometric primitives to 3D point clouds. In: CVPR (2019)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.433, + 0.785, + 0.473 + ], + "angle": 0, + "content": "25. Li, L.H., Zhang, P., Zhang, H., Yang, J., Li, C., Zhong, Y., Wang, L., Yuan, L., Zhang, L., Hwang, J.N., et al.: Grounded language-image pre-training. In: CVPR (2022)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.473, + 0.785, + 0.514 + ], + "angle": 0, + "content": "26. Li, M., Xie, Y., Shen, Y., Ke, B., Qiao, R., Ren, B., Lin, S., Ma, L.: Hybridcr: Weakly-supervised 3d point cloud semantic segmentation via hybrid contrastive regularization. In: CVPR (2022)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.514, + 0.785, + 0.542 + ], + "angle": 0, + "content": "27. Li, Y., Bu, R., Sun, M., Wu, W., Di, X., Chen, B.: PointCNN: Convolution on x-transformed points. In: NeurIPS (2018)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.542, + 0.785, + 0.582 + ], + "angle": 0, + "content": "28. Li, Y., Ma, L., Zhong, Z., Cao, D., Li, J.: TGNet: Geometric graph cnn on 3D point cloud segmentation. IEEE Transactions on Geoscience and Remote Sensing (2019)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.582, + 0.785, + 0.623 + ], + "angle": 0, + "content": "29. Liang, Z., Yang, M., Li, H., Wang, C.: 3D instance embedding learning with a structure-aware loss function for point cloud segmentation. IEEE Robotics and Automation Letters (2020)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.623, + 0.785, + 0.651 + ], + "angle": 0, + "content": "30. Lin, T.Y., Maire, M., Belongie, S., Hays, J., Perona, P., Ramanan, D., Dólar, P., Zitnick, C.L.: Microsoft COCO: Common objects in context. In: ECCV (2014)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.651, + 0.785, + 0.677 + ], + "angle": 0, + "content": "31. Liu, J., Yu, M., Ni, B., Chen, Y.: Self-prediction for joint instance and semantic segmentation of point clouds. In: ECCV (2020)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.677, + 0.785, + 0.717 + ], + "angle": 0, + "content": "32. Liu, M., Zhu, Y., Cai, H., Han, S., Ling, Z., Porikli, F., Su, H.: PartSLIP: Low-shot part segmentation for 3d point clouds via pretrained image-language models. In: CVPR (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.718, + 0.785, + 0.759 + ], + "angle": 0, + "content": "33. Liu, S., Zeng, Z., Ren, T., Li, F., Zhang, H., Yang, J., Li, C., Yang, J., Su, H., Zhu, J., et al.: Grounding DINO: Marrying dino with grounded pre-training for open-set object detection. arXiv preprint arXiv:2303.05499 (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.759, + 0.785, + 0.786 + ], + "angle": 0, + "content": "34. Liu, Y., Fan, B., Meng, G., Lu, J., Xiang, S., Pan, C.: DensePoint: Learning densely contextual representation for efficient point cloud processing. In: ICCV (2019)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.786, + 0.785, + 0.813 + ], + "angle": 0, + "content": "35. Liu, Z., Qi, X., Fu, C.W.: One thing one click: A self-training approach for weakly supervised 3D semantic segmentation. In: CVPR (2021)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.813, + 0.785, + 0.841 + ], + "angle": 0, + "content": "36. Ma, X., Qin, C., You, H., Ran, H., Fu, Y.: Rethinking network design and local geometry in point cloud: A simple residual mlp framework. In: ICLR (2022)" + }, + { + "type": "list", + "bbox": [ + 0.217, + 0.147, + 0.785, + 0.841 + ], + "angle": 0, + "content": null + } + ], + [ + { + "type": "header", + "bbox": [ + 0.344, + 0.115, + 0.732, + 0.129 + ], + "angle": 0, + "content": "PartSTAD: 2D-to-3D Part Segmentation Task Adaptation" + }, + { + "type": "page_number", + "bbox": [ + 0.769, + 0.117, + 0.785, + 0.127 + ], + "angle": 0, + "content": "17" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.147, + 0.785, + 0.189 + ], + "angle": 0, + "content": "37. Mo, K., Zhu, S., Chang, A.X., Yi, L., Tripathi, S., Guibas, L.J., Su, H.: PartNet: A large-scale benchmark for fine-grained and hierarchical part-level 3D object understanding. In: CVPR (2019)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.19, + 0.785, + 0.218 + ], + "angle": 0, + "content": "38. Narita, G., Seno, T., Ishikawa, T., Kaji, Y.: Panoptic fusion: Online volumetric semantic mapping at the level of stuff and things. In: IROS (2019)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.218, + 0.785, + 0.259 + ], + "angle": 0, + "content": "39. Peng, S., Genova, K., Jiang, C., Tagliasacchi, A., Pollefeys, M., Funkhouser, T., et al.: OpenScene: 3D scene understanding with open vocabularies. In: CVPR (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.259, + 0.785, + 0.3 + ], + "angle": 0, + "content": "40. Pham, Q.H., Nguyen, D.T., Hua, B.S., Roig, G., Yeung, S.K.: JSIS3D: Joint semantic-instance segmentation of 3D point clouds with multi-task pointwise networks and multi-value conditional random fields. In: CVPR (2019)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.3, + 0.785, + 0.342 + ], + "angle": 0, + "content": "41. Pruksachatkun, Y., Phang, J., Liu, H., Htut, P.M., Zhang, X., Pang, R.Y., Vania, C., Kann, K., Bowman, S.R.: Intermediate-Task Transfer Learning with Pretrained Language Models: When and why does it work? In: ACL (2020)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.342, + 0.785, + 0.37 + ], + "angle": 0, + "content": "42. Qi, C.R., Su, H., Mo, K., Guibas, L.J.: PointNet: Deep learning on point sets for 3d classification and segmentation. In: CVPR (2017)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.37, + 0.785, + 0.398 + ], + "angle": 0, + "content": "43. Qi, C.R., Yi, L., Su, H., Guibas, L.J.: PointNet++: Deep hierarchical feature learning on point sets in a metric space. In: NeurIPS (2017)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.398, + 0.785, + 0.439 + ], + "angle": 0, + "content": "44. Qian, G., Li, Y., Peng, H., Mai, J., Hammoud, H., Elhoseiny, M., Ghanem, B.: PointNext: Revisiting pointnet++ with improved training and scaling strategies. In: NeurIPS (2022)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.439, + 0.785, + 0.481 + ], + "angle": 0, + "content": "45. Radford, A., Kim, J.W., Hallacy, C., Ramesh, A., Goh, G., Agarwal, S., Sastry, G., Askell, A., Mishkin, P., Clark, J., et al.: Learning transferable visual models from natural language supervision. In: ICML (2021)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.481, + 0.785, + 0.509 + ], + "angle": 0, + "content": "46. Schult, J., Engelmann, F., Hermans, A., Litany, O., Tang, S., Leibe, B.: Mask3D: Mask transformer for 3D semantic instance segmentation. In: ICRA (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.509, + 0.785, + 0.536 + ], + "angle": 0, + "content": "47. Takmaz, A., Fedele, E., Sumner, R.W., Pollefeys, M., Tombari, F., Engelmann, F.: OpenMask3D: Open-vocabulary 3D instance segmentation. In: NeurIPS (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.536, + 0.785, + 0.564 + ], + "angle": 0, + "content": "48. Tang, L., Jia, M., Wang, Q., Phoo, C., Hariharan, B.: Emergent correspondence from image diffusion. In: NeurIPS (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.564, + 0.785, + 0.592 + ], + "angle": 0, + "content": "49. Thomas, H., Qi, C.R., Deschaud, J.E., Marcotegui, B., Goulette, F., Guibas, L.J.: Kpconv: Flexible and deformable convolution for point clouds. In: ICCV (2019)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.592, + 0.785, + 0.619 + ], + "angle": 0, + "content": "50. Vu, T., Kim, K., Luu, T.M., Nguyen, X.T., Yoo, C.D.: Softgroup for 3d instance segmentation on 3D point clouds. In: CVPR (2022)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.619, + 0.785, + 0.647 + ], + "angle": 0, + "content": "51. Wang, W., Yu, R., Huang, Q., Neumann, U.: SGPN: Similarity group proposal network for 3D point cloud instance segmentation. In: CVPR (2018)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.647, + 0.785, + 0.674 + ], + "angle": 0, + "content": "52. Wang, X., Liu, S., Shen, X., Shen, C., Jia, J.: Associatively segmenting instances and semantics in point clouds. In: CVPR (2019)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.674, + 0.785, + 0.702 + ], + "angle": 0, + "content": "53. Wang, Y., Sun, Y., Liu, Z., Sarma, S.E., Bronstein, M.M., Solomon, J.M.: Dynamic graph cnn for learning on point clouds. ACM TOG (2019)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.702, + 0.785, + 0.743 + ], + "angle": 0, + "content": "54. Wu, T., Zhang, J., Fu, X., Wang, Y., Ren, J., Pan, L., Wu, W., Yang, L., Wang, J., Qian, C., Lin, D., Liu, Z.: OmniObject3D: Large-vocabulary 3D object dataset for realistic perception, reconstruction and generation. In: CVPR (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.743, + 0.785, + 0.772 + ], + "angle": 0, + "content": "55. Wu, X., Lao, Y., Jiang, L., Liu, X., Zhao, H.: Point transformer v2: Grouped vector attention and partition-based pooling. In: NeurIPS (2022)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.772, + 0.785, + 0.813 + ], + "angle": 0, + "content": "56. Xiang, F., Qin, Y., Mo, K., Xia, Y., Zhu, H., Liu, F., Liu, M., Jiang, H., Yuan, Y., Wang, H., Yi, L., Chang, A.X., Guibas, L.J., Su, H.: SAPIEN: A simulated part-based interactive environment. In: CVPR (2020)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.813, + 0.785, + 0.841 + ], + "angle": 0, + "content": "57. Xiang, T., Zhang, C., Song, Y., Yu, J., Cai, W.: Walk in the Cloud: Learning curves for point clouds shape analysis. In: ICCV (2021)" + }, + { + "type": "list", + "bbox": [ + 0.217, + 0.147, + 0.785, + 0.841 + ], + "angle": 0, + "content": null + } + ], + [ + { + "type": "page_number", + "bbox": [ + 0.218, + 0.116, + 0.236, + 0.127 + ], + "angle": 0, + "content": "18" + }, + { + "type": "header", + "bbox": [ + 0.272, + 0.115, + 0.415, + 0.129 + ], + "angle": 0, + "content": "H. Kim and M. Sung" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.147, + 0.785, + 0.175 + ], + "angle": 0, + "content": "58. Xu, J., Liu, S., Vahdat, A., Byeon, W., Wang, X., De Mello, S.: Open-vocabulary panoptic segmentation with text-to-image diffusion models. In: CVPR (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.177, + 0.785, + 0.203 + ], + "angle": 0, + "content": "59. Xu, M., Ding, R., Zhao, H., Qi, X.: PACov: Position adaptive convolution with dynamic kernel assembling on point clouds. In: CVPR (2021)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.204, + 0.785, + 0.231 + ], + "angle": 0, + "content": "60. Yan, X., Zheng, C., Li, Z., Wang, S., Cui, S.: PointANSL: Robust point clouds processing using nonlocal neural networks with adaptive sampling. In: CVPR (2020)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.232, + 0.785, + 0.272 + ], + "angle": 0, + "content": "61. Yang, B., Wang, J., Clark, R., Hu, Q., Wang, S., Markham, A., Trigoni, N.: Learning object bounding boxes for 3D instance segmentation on point clouds. In: NeurIPS (2019)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.273, + 0.774, + 0.286 + ], + "angle": 0, + "content": "62. Yang, X., Wang, X.: Diffusion model as representation learner. In: ICCV (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.287, + 0.785, + 0.314 + ], + "angle": 0, + "content": "63. Yang, Y., Wu, X., He, T., Zhao, H., Liu, X.: SAM3D: Segment anything in 3D scenes. arXiv preprint arXiv:2306.03908 (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.315, + 0.785, + 0.342 + ], + "angle": 0, + "content": "64. Yi, K.M., Trulls, E., Ono, Y., Lepetit, V., Salzmann, M., Fua, P.: Learning to find good correspondences. In: CVPR (2018)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.343, + 0.785, + 0.37 + ], + "angle": 0, + "content": "65. Yi, L., Huang, H., Liu, D., Kalogerakis, E., Su, H., Guibas, L.: Deep part induction from articulated object pairs. ACM TOG (2018)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.371, + 0.785, + 0.397 + ], + "angle": 0, + "content": "66. Yi, L., Zhao, W., Wang, H., Sung, M., Guibas, L.J.: GSPN: Generative shape proposal network for 3D instance segmentation in point cloud. In: CVPR (2019)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.398, + 0.785, + 0.425 + ], + "angle": 0, + "content": "67. Yu, X., Tang, L., Rao, Y., Huang, T., Zhou, J., Lu, J.: Point-BERT: Pre-training 3D point cloud transformers with masked point modeling. In: CVPR (2022)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.426, + 0.785, + 0.452 + ], + "angle": 0, + "content": "68. Zamir, A.R., Sax, A., Shen, W.B., Guibas, L.J., Malik, J., Savarese, S.: Taskonomy: Disentangling task transfer learning. In: CVPR (2018)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.453, + 0.785, + 0.479 + ], + "angle": 0, + "content": "69. Zhang, B., Wonka, P.: Point cloud instance segmentation using probabilistic embeddings. In: CVPR (2021)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.48, + 0.785, + 0.508 + ], + "angle": 0, + "content": "70. Zhang, J., Dong, R., Ma, K.: CLIP-FO3D: Learning free open-world 3D scene representations from 2D dense clip. arXiv preprint arXiv:2303.04748 (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.509, + 0.785, + 0.549 + ], + "angle": 0, + "content": "71. Zhang, R., Wang, L., Wang, Y., Gao, P., Li, H., Shi, J.: Parameter is not all you need: Starting from non-parametric networks for 3D point cloud analysis. In: CVPR (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.55, + 0.785, + 0.591 + ], + "angle": 0, + "content": "72. Zhang, Y., Ling, H., Gao, J., Yin, K., Lafleche, J.F., Barriuso, A., Torralba, A., Fidler, S.: DatasetGAN: Efficient labeled data factory with minimal human effort. In: CVPR (2021)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.592, + 0.785, + 0.619 + ], + "angle": 0, + "content": "73. Zhao, H., Jiang, L., Jia, J., Torr, P.H., Koltun, V.: Point transformer. In: ICCV (2021)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.62, + 0.785, + 0.646 + ], + "angle": 0, + "content": "74. Zhao, N., Chua, T.S., Lee, G.H.: Few-shot 3D point cloud semantic segmentation. In: CVPR (2021)" + }, + { + "type": "list", + "bbox": [ + 0.217, + 0.147, + 0.785, + 0.646 + ], + "angle": 0, + "content": null + } + ] +] \ No newline at end of file diff --git a/2024/PartSTAD_ 2D-to-3D Part Segmentation Task Adaptation/bbdaa935-8663-4268-b38a-384a769ccb82_origin.pdf b/2024/PartSTAD_ 2D-to-3D Part Segmentation Task Adaptation/bbdaa935-8663-4268-b38a-384a769ccb82_origin.pdf new file mode 100644 index 0000000000000000000000000000000000000000..44be0f8b198cca1cffcf8ac685f479d5fcf83c16 --- /dev/null +++ b/2024/PartSTAD_ 2D-to-3D Part Segmentation Task Adaptation/bbdaa935-8663-4268-b38a-384a769ccb82_origin.pdf @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:2e1fefcd8573b028030b7842a2ad097991d2ee57e772a2dea72224ff4dc908aa +size 6022761 diff --git a/2024/PartSTAD_ 2D-to-3D Part Segmentation Task Adaptation/full.md b/2024/PartSTAD_ 2D-to-3D Part Segmentation Task Adaptation/full.md new file mode 100644 index 0000000000000000000000000000000000000000..bc141b64c3cc5c7452f8c64c992e2b926791002f --- /dev/null +++ b/2024/PartSTAD_ 2D-to-3D Part Segmentation Task Adaptation/full.md @@ -0,0 +1,310 @@ +# PartSTAD: 2D-to-3D Part Segmentation Task Adaptation + +Hyunjin $\mathrm{Kim}^{1\dagger}$ and Minhyuk $\mathrm{Sung}^{2}$ + +1 KRAFTON Inc., South Korea + +2 KAIST, South Korea + +rlaguswls98@krafton.com, mhsung@kaist.ac.kr + +![](images/826464d72d32b4853d3b3fdf9c02fc29e39f268677baaa949e049661e789fe67.jpg) +Figure 1: We introduce PARTSTAD, a novel few-shot 3D point cloud part segmentation method that leverages 2D-to-3D task adaptation. By obtaining 2D segmentation masks in multi-view images from GLIP [25] and SAM [20] and optimizing the mask weights for 3D segmentation as a learning objective, it can successfully predict fine-grained parts with accurate boundaries, as shown in the figure above. + +Abstract. We introduce PARTSTAD, a method designed for the task adaptation of 2D-to-3D segmentation lifting. Recent studies have highlighted the advantages of utilizing 2D segmentation models to achieve high-quality 3D segmentation through few-shot adaptation. However, previous approaches have focused on adapting 2D segmentation models for domain shift to rendered images and synthetic text descriptions, rather than optimizing the model specifically for 3D segmentation. Our proposed task adaptation method finetunes a 2D bounding box prediction model with an objective function for 3D segmentation. We introduce weights for 2D bounding boxes for adaptive merging and learn the weights using a small additional neural network. Additionally, we incorporate SAM, a foreground segmentation model on a bounding box, to improve the boundaries of 2D segments and consequently those of 3D segmentation. Our experiments on the PartNet-Mobility dataset show significant improvements with our task adaptation approach, achieving a $7.0\%$ p increase in mIoU and a $5.2\%$ p improvement in $\mathrm{mAP}_{50}$ for semantic and instance segmentation compared to the SotA few-shot 3D segmentation model. The code is available at https://github.com/KAIST-Visual-AI-Group/PartSTAD. + +Keywords: part segmentation $\cdot$ few-shot learning $\cdot$ 3D deep learning + +# 1 Introduction + +3D segmentation has been a subject of extensive research in computer vision due to its central role in various downstream applications involving the understanding of shape structure, functionality, mobility, and semantics. However, the limited availability of annotated 3D shapes has remained a bottleneck in achieving generalizability in learned segmentation models for diverse 3D data. Annotating 3D segmentation is particularly labor-intensive, time-consuming, and requires expertise in handling 3D models. For this reason, the scale of 3D part-level annotations remains in the tens of thousands [37], while the scale of 2D annotation datasets in the image domain, for example, exceeds a million [20, 30]. + +Recent research [2,32] has illuminated the potential of visual-language models bridging textual descriptions with images in accomplishing zero-shot or few-shot 3D segmentation results. The basic idea is to render a 3D model from various viewpoints, conduct 2D detection or segmentation on the rendered images, and then aggregate the 2D segmentation results into a 3D representation using either a voting mechanism [32] or a label propagation scheme [2]. This approach is not only effective in enabling zero-shot and few-shot 3D segmentation but also offers an advantage in that the set of part names does not need to be predefined in training but can be determined at test time. + +PartSLIP [32] is a notable example that achieves 3D part segmentation results comparable to fully supervised methods while adapting a pretrained 2D detection model with few-shot training for 3D segmentation. It leverages the pretrained GLIP [25] model for 2D detection, associating the output 2D bounding boxes with one of the tokens (part names) provided in the input prompt. Its key component is the GLIP finetuning process, which introduces a small number of learnable parameters to the frozen GLIP model and trains them using few-shot synthetic images and texts employed in the 3D segmentation pipeline. As GLIP was initially trained with real photos and natural language descriptions of objects by humans, this adaptation to rendered images and the unconventional description (a sequence of part names) results in a substantial improvement in the 3D segmentation task. + +Despite promising initiatives, it is important to note that the previous approach was limited to achieving domain adaptation. In contrast, we emphasize that in the 2D-to-3D segmentation lifting task, not only does the data domain change, but the task itself also shifts from 2D segmentation to 3D segmentation. Thus, the process of applying the pretrained model to the new task requires task adaptation, involving modifying the model with an objective function associated with the new task. Particularly for 3D segmentation, it is necessary to integrate 2D segmentation results from multiple viewpoints into a coherent 3D representation. Therefore, it is crucial to control the noise of 2D bounding box predictions in the context of its impact on the final 3D segmentation after integration. + +Specifically, our Part Segmentation Task ADaptation method, PARTSTAD, adapts the pretrained GLIP model with a relaxed 3D mIoU loss while incorporating bounding boxes across all different views. Instead of typical finetuning, we draw inspiration from recent work [16, 32] that introduces small learnable + +parameters to the existing model while keeping existing parameters frozen. As a result, we train a small MLP using 2D bounding box features extracted from the pretrained GLIP model. Since the relaxed mIoU loss is non-differentiable with respect to the bounding box positions, we suggest predicting a weight for each bounding box instead of adjusting its position. This results in minimal modification in the 2D-to-3D voting scheme while achieving a significant improvement in 3D segmentation, even with few-shot training, e.g., with eight objects per category, as done in PartSLIP. To further enhance 3D segmentation performance, we leverage SAM [20], a 2D instance segmentation model, to segment the foreground region within each bounding box, obtaining an accurate boundary for each 2D segment. + +In our experiments on the PartNet-Mobility [56] dataset, we showcase the superior performance of our method in comparison to recent zero-shot/few-shot 3D semantic and instance segmentation methods that leverage 2D segmentation models. In contrast to the SotA few-shot method, our approach achieves a $7.0\% \mathbf{p}$ improvement in semantic segmentation mIoU and a $5.2\% \mathbf{p}$ improvement in instance segmentation $\mathrm{mAP}_{50}$ . These improvements are consistent across all object classes. + +# 2 Related Work + +# 2.1 Supervised 3D Segmentation + +Given the availability of open segment-annotated 3D datasets [3,5,7,8,12,14,37, 56], 3D segmentation has been extensively researched in the last several years, focusing on the development of novel network architectures for supervised learning. Regarding the architecture of semantic segmentation, diverse models have been explored, including PointNet [42] and its variants [36,43,44,57,71], those using CNN [27,34,49,59], GCN [17,28,53], and Transformers [55,60,67,73]. For instance segmentation, various approaches based on proposing 3D bounding boxes [15,61,66], clustering learned features [11,19,31,38,50,51,69,74], and combining both semantic and instance segmentations [29,40,46,52] have been introduced. Despite these great advances, the performance of supervised methods has been limited by the scale of 3D datasets; the largest part-annotated 3D dataset, PartNet [37], includes fewer than $30\mathrm{k}$ models. + +There have been some attempts to overcome the limitation by leveraging additional weak supervisions, such as text descriptions of objects [21], 3D bounding box annotations [10], sparse labels [26, 35], and geometric priors [23, 24]. However, the scale of this additional data has also been limited to tens of thousands, while datasets in the image domain are on the scale of millions. + +# 2.2 3D Segmentation Using Vision-Language Models + +Recent work introduced ideas about exploiting vision-language models (VLM) for 3D segmentation. The advantage of utilizing the learned visual-language + +grounding in 3D segmentation lies in the strong generalizability to arbitrary 3D models with zero-shot or few-shot training, and also in the open vocabulary understanding that enables labeling points without specifying the set of part or object names during training. For a VLM, CLIP [45] has been adapted in multiple previous works to lift the 2D grounding to the 3D, detecting parts in an object [13] and objects in a scene [9,39,47,70], while the results are limited to highlighting some regions without providing clear segment boundaries. + +To obtain precise segments in 3D, other 2D object detection and segmentation models, such as SAM [20], GLIP [25], and GroundingDINO [33], are also leveraged. SAM3D [63] was an example of directly lifting the 2D masks from SAM to 3D with a bottom-up merging approach, although it was limited to segmenting the parts, not labeling them. SA3D [6] utilized NeRF representation and allowed user interaction for 3D segmentation, but it can be inefficient when the 3D objects are given as a mesh, point cloud, or other common 3D representations. OpenMask3D [47] and OpenIns3D [18] are concurrent works, both of which combine multiple foundation models, SAM [20] and CLIP [45] for OpenMask3D, and GroundingDINO [33] and LISA [22] for OpenIns3D. These works focus on segmenting objects in 3D scenes, while we aim to do finer-grained segmentation, finding parts in 3D objects. + +PartSLIP [32] and SATR [2] are notable examples that achieve part segmentation of 3D objects based on a 2D object detection model, GLIP [25]. They obtain 2D bounding boxes on the rendered images from multiple views and integrate them over the 3D object with different merging schemes. The finetuning of GLIP introduced by PartSLIP particularly leads to a performance improvement, but it is limited to domain adaptation, not task adaptation. Building on this, we introduce a novel task adaptation technique as well as integration with SAM [20] to achieve a further significant improvement in 3D segmentation accuracy. + +# 2.3 Task Adaptation + +Taskonomy [68] is a seminal work that introduced the concept of task transfer learning for the first time, enabling the transformation of an image dense prediction model trained for a base task to perform other tasks through finetuning. Pruksachatkun et al. [41] also introduced a similar idea for language-domain tasks. Later, task adaptation has been extensively studied to leverage features learned from generative models, such as GAN [72] and diffusion models [4, 48, 58, 62], adapting them with a small network for various tasks, including image segmentation [4, 58, 72], depth prediction [48], and keypoint detection [62]. For 3D tasks, Abdelreheem et al. [1] are notable examples of using vision language-grounding models for 3D correspondence but are limited to directly applying 2D priors without finetuning. We introduce a novel task adaptation technique that lifts 2D segments to 3D by adapting the 2D features during the training of a small network with a task-specific objective function. + +# 3 Background and Preliminaries + +Recent research [2,13,32] has demonstrated how vision-language models designed for image object detection and segmentation can be applied to segment parts in 3D objects. The main idea involves rendering a 3D model from various viewpoints, performing 2D detection or segmentation for the rendered image at each view, and then combining the 2D segmentation results over the 3D model. + +Specifically, SATR [2] utilizes GLIP [25] as the 2D detection model. To obtain 2D bounding boxes for each of the specified part names, it feeds a concatenated list of these part names (separated by commas) as the input prompt to the pretrained GLIP model, along with a rendered image. The 2D bounding boxes are then mapped to a single face of the input 3D mesh, located at the center of the respective 2D bounding box, and the part names are propagated throughout the entire mesh based on geodesic distances. Similarly, PartSLIP [32] also leverages GLIP [25] for 2D detection but employs a different voting scheme for the 2D bounding box integration over the 3D object, which utilizes the oversegmentation of the input point cloud. One key distinction between PartSLIP and SATR is that PartSLIP proposes to modify the GLIP model for domain adaptation to the rendered images and text prompts used in 3D segmentation. Specifically, the text prompt is given as a concatenation of part names in the PartSLIP pipeline, which is not a typical and natural description of the object. Therefore, it learns a category-specific constant offset vector for the language embedding feature in the given pairs of prompts and rendered images, while keeping the GLIP parameters frozen. + +Our work builds upon the PartSLIP pipeline. In the following subsections, we describe the details of its major components: the voting scheme, and the finetuning process. + +# 3.1 PartSLIP [32] + +The main technical components of PartSLIP are the 1) voting scheme based on super points and 2) adaptation to text prompts1. We describe the details of each component below. + +Voting for Super Points. PartSLIP first over-segments the input 3D object represented as a point cloud $\mathcal{P}$ into a set of super points $P_{i} \subseteq \mathcal{P}$ , which provide geometric priors for the boundaries of the parts. The 3D object is rendered into $K$ viewpoints, and the 2D bounding boxes for each image are predicted using GLIP [25]. Let $\mathcal{B}$ denote the entire set of 2D bounding boxes from all views, and $B_{jk} \subseteq \mathcal{B}$ indicates a subset that includes the bounding boxes classified as the $j$ -th part label from the $k$ -th view. $\mathrm{V}_k: \mathcal{P} \to \{0,1\}$ is a function indicating whether the input point is visible from the $k$ -th viewpoint. Also, for each 2D + +bounding box $b \in \mathcal{B}$ , $\mathrm{I}_b : \mathcal{P} \to \{0,1\}$ is a function indicating whether the input point is included in the bounding box $b$ . Given these, the voting aggregating the semantic labels of the 2D bounding boxes to the 3D point cloud is performed by calculating the following score $s_{ij}$ for each pair of the $i$ -th super point and the $j$ -th label: + +$$ +s _ {i j} = \frac {\sum_ {k} \sum_ {p \in P _ {i}} \mathrm {V} _ {k} (p) \left(\max _ {b \in \left(B _ {j k} \cup \{b _ {\emptyset} \}\right)} \mathrm {I} _ {b} (p)\right)}{\sum_ {k} \sum_ {p \in P _ {i}} \mathrm {V} _ {k} (p)}, \tag {1} +$$ + +where $b_{\emptyset}$ is a null bounding box, and we assume that $\mathrm{I}_{b_{\emptyset}}$ for the null bounding box returns zero for any input point (simplifying notations for the case when $B_{jk} = \emptyset$ ). Note that this $s_{ij} \in [0,1]$ denotes the ratio of visible points from each view included in any of the 2D bounding boxes labeled with the $j$ -th label. Finally, for each $i$ -th super point, the label with the highest score $s_{ij}$ across all the labels is assigned to the super point. Exceptionally, if the highest score is less than a null label threshold $s_{\emptyset}$ , which is set to 0.5 in our experiments, the null label is assigned to the super point. + +GLIP Finetuning. While GLIP demonstrates impressive generalizability for unseen images and prompts, it still has limitations when it comes to handling synthetic images and text prompts, such as the rendered images and sequences of part names used in the PartSLIP pipeline. To address this, the authors of PartSLIP propose adding small, learnable parameters to the GLIP model while keeping all the pretrained GLIP parameters frozen. These new parameters are learned with very few-shot examples: 8 annotated 3D objects per category. They represent offset feature vectors for each part name and remain constant as global variables for each category, rather than changing for each input. This adaptation of GLIP significantly improves the accuracy of 3D part segmentation. + +Instance Segmentation. The voting scheme described above is designed for semantic segmentation, but PartSLIP also introduces a straightforward merging-based method to achieve instance segmentation based on semantic segmentation. To obtain instance segments, it merges adjacent super points under two conditions: 1) they share the same semantic label, and 2) they are either both included or both excluded for all bounding boxes. + +# 4 PARTSTAD: Task Adaptation for 2D-to-3D + +PartSLIP has demonstrated that adapting a 2D detection model for 3D segmentation with a small set of trainable parameters can significantly enhance 3D segmentation performance. However, we observe that PartSLIP's adaptation is limited to domain adaptation, where it learns new parameters using synthetic images and data used in the PartSLIP framework without altering the objective function during training. + +When applying the 2D segmentation model to 3D segmentation, it is important to note that not only does the data domain change, but the task itself + +![](images/23e10151c3738afe3e36eff6d18c2f98fa62a94a5f38d19207d1137cf4c0bd08.jpg) +Figure 2: Overall pipeline of PARTSTAD. Our approach begins by rendering the provided 3D point cloud from multiple viewpoints. Subsequently, we extract 2D bounding boxes for its parts using GLIP [25] (Bounding Box Prediction); note that we utilize the finetuned GLIP model from PartSLIP [32]. Following this, we convert the bounding boxes into segmentation masks using SAM [20], extracting the foreground region for each bounding box (SAM Mask Integration). Next, we predict weights for all the masks and adaptively combine them into a 3D representation (2D-to-3D task adaptation). The final step involves obtaining the segmentation label for the input point cloud. The GLIP and SAM models are frozen, while only our novel weight prediction network is trained per category in a few-shot setting (8 objects). + +changes. Therefore, through the process of incorporating 2D bounding boxes into the 3D object, the 2D bounding box prediction model must align with our ultimate goal: 3D part segmentation. + +To address this, we introduce Part Segmentation Task ADaptation method, PA STAD, a task adaptation approach designed to finetune the 2D segmentation model for 3D part segmentation. In Sec. 4.1, we introduce our objective function to finetune the GLIP for the 3D segmentation task. In Sec. 4.2, we describe how we adapt the pretrained GLIP model with the new objective function by introducing additional learnable parameters and modifying the voting scheme (Eq. 1). PartSLIP also faces a performance limitation due to the GLIP that provides 2D bounding boxes, not 2D segments. In Sec. 4.3, we also propose to combine the SAM [20] foreground mask for each bounding box to obtain more precise 2D segments and, consequently, further improve the 3D segmentation. + +# 4.1 3D mRIoU Loss + +We begin by describing the loss function that we use for adapting the GLIP to the 3D segmentation task. Let $\mathbf{l}_j$ and $\hat{\mathbf{l}}_j \in \{0,1\}^{|\mathcal{P}|}$ be binary vectors indicating whether each point $p$ in $\mathcal{P}$ has the $j$ -th label in the ground truth and label prediction, respectively. The mean Intersection over Union (mIoU) for the predicted 3D segmentation, the standard evaluation metric for 3D segmentation, is defined as follows: + +$$ +\operatorname {m I o U} \left(\left\{\mathbf {l} _ {j} \right\}, \left\{\hat {\mathbf {l}} _ {j} \right\}\right) = \frac {1}{M} \sum_ {j = 1} ^ {M} \frac {\mathbf {l} _ {j} ^ {\top} \hat {\mathbf {l}} _ {j}}{\left\| \mathbf {l} _ {j} \right\| _ {1} + \left\| \hat {\mathbf {l}} _ {j} \right\| _ {1} - \left. \mathbf {l} _ {j} ^ {\top} \hat {\mathbf {l}} _ {j} \right.}, \tag {2} +$$ + +where $M$ is the number of labels. + +Our goal is to use the 3D mIoU as the objective function directly when adapting the GLIP model. However, the mIoU is non-differentiable. Thus, we employ a relaxed version known as mean Relaxed IoU (mRIOU) [24,65], which simply allows the predicted indicator vectors $\hat{\mathbf{l}}_j$ to be non-binary: $\hat{\mathbf{l}}_j\in [0,1]^{|\mathcal{P}|}$ . The loss $\mathcal{L}_{\mathrm{mRIOU}}$ is specifically defined as follows: + +$$ +\mathcal {L} _ {\mathrm {m R I o U}} = 1 - \operatorname {m I o U} \left(\left\{\mathbf {l} _ {j} \right\}, \left\{\hat {\mathbf {l}} _ {j} \right\}\right). \tag {3} +$$ + +In the supplementary material, we demonstrate that the choice of the loss function, as compared to other alternatives such as the cross-entropy loss, is crucial to achieving a substantial improvement in our task adaptation. + +Note that in the PartSLIP framework, where super points are utilized in 3D segmentation, and $s_{ij} \in [0,1]$ indicates the likelihood of the $i$ -th super point belonging to the $j$ -label, the mRIoU can be calculated by defining $\hat{\mathbf{l}}_j$ based on $\mathbf{s}_j = [s_{1j}, s_{2j}, \dots]^{\top}$ as follows: + +$$ +\hat {\mathbf {l}} _ {j} = \mathbf {M} \mathbf {s} _ {j}, \tag {4} +$$ + +where $\mathbf{M} \in \{0,1\}^{|\mathcal{P}| \times |\{P_i\}|}$ is a binary matrix that describes the memberships from each point to the super points. + +# 4.2 Bounding Box Weight Prediction + +A typical approach for adapting a pretrained model to a new task would be finetuning the model using an objective function tailored to that new task. However, the finetuning method generally fails to yield meaningful improvements, especially when the goal is to achieve few-shot adaptation (e.g., with only 8 annotated 3D objects per class, as done in our case), while the pretrained model has been trained on an extensive dataset. To address this challenge, recent works like LoRA [16] and PartSLIP [32] have introduced the concept of adding small learnable parameters while keeping the existing pretrained model parameters frozen. This approach not only makes training very efficient but also enables the model to generalize effectively to unseen data. + +While we also follow this approach, our specific challenge arises from the fact that when we use the 3D mRIoU (Eq. 3) as the objective function for 3D segmentation, it becomes non-differentiable with respect to the 2D bounding box location, which is the output of GLIP. This non-differentiability arises from the computation of the score $s_{ij}$ (Eq. 1) for each pair of super point and label. + +To address this, we propose training a small network that does not refine the location but instead predicts a weight for each 2D bounding box, taking the learned features of the boxes as input. If we denote the output weight of the network for bounding box $b$ as $\mathrm{W}(b) \in \mathbb{R}^{+}$ , the score $s_{ij}$ for a pair of super point and label is modified to $\tilde{s}_{ij}$ as follows: + +$$ +\tilde {s} _ {i j} = \frac {\sum_ {k} \sum_ {p \in P _ {i}} \mathrm {V} _ {k} (p) \left(\max _ {b \in \left(B _ {j k} \cup \{b _ {\emptyset} \}\right)} \mathrm {I} _ {b} (p) \mathrm {W} (b)\right)}{\sum_ {k} \sum_ {p \in P _ {i}} \mathrm {V} _ {k} (p)}. \tag {5} +$$ + +Note that the change is simply to multiply the weight $\mathrm{W}(b)$ by $\mathrm{I}_b(p)$ , while only $\mathrm{I}_b(p)$ , indicating whether the point $p$ is included in bounding box $b$ , has been used in Eq. 1. This modified score no longer falls within the range of [0,1]. The final score $\bar{s}_{ij}$ is thus defined by normalizing $\tilde{s}_{ij}$ using the softmax function over the set of labels: + +$$ +\bar {s} _ {i j} = \frac {\exp (\tilde {s} _ {i j})}{\sum_ {j} \exp (\tilde {s} _ {i j})}. \tag {6} +$$ + +We additionally propose to make the unnormalized score for the null label $\tilde{s}_{\emptyset}$ learnable in our case, initialized with 10. We thus compute the softmax above while including $\tilde{s}_{\emptyset}$ . The label of each super point is still chosen as the label giving the highest score $\bar{s}_{ij}$ , becoming null if the normalized null label score $\bar{s}_{\emptyset}$ becomes the highest. + +Network Architecture. We design the weight predictor network to take the bounding box feature vectors $\mathbf{f}_b$ from the pretrained GLIP model as input. The feature vectors of all bounding boxes across all views are fed to the network at once and processed with a small shared two-layer MLP. Context normalization [64] is added in the middle of the two layers to incorporate contextual information from the global set of boxes for each box. The output of the MLP for each bounding box is further processed with the following modified ReLU function $\phi(\cdot)$ : + +$$ +\phi (\mathbf {x}) = \max (\tau + \mathbf {x}, 0), \tag {7} +$$ + +where $\tau$ is a user-defined constant offset (10 in our experiments). + +# 4.3 SAM [20] Mask Integration + +Another limitation of PartSLIP is its reliance on GLIP [25] for 2D segmentation. Since GLIP is, in turn, a 2D object detection model, it produces bounding boxes instead of 2D segments, which cannot provide accurate boundaries of segments. We propose to address this issue by incorporating another pretrained 2D segmentation model, SAM [20]. SAM has the functionality of taking a bounding box as input and producing the foreground mask. Using this, we replace the set of 2D bounding boxes $\mathcal{B}$ with a set of 2D masks, while preserving the remaining steps in the framework. Note that the point-to-bounding-box membership $\mathrm{I}_b$ is changed to point-to-mask membership, and we still use the same bounding box features extracted from GLIP corresponding to each mask to train the weight prediction network. We demonstrate the effectiveness of applying the SAM mask in our experiments in Sec. 5. + +# 5 Experiment Results + +In the experiments, we compare our method with the SotA 2D-prior-based zero-shot/few-shot 3D segmentation models using the PartNet-Mobility [56] dataset + +![](images/60bec06afde2cda26bafb12f6d9cf23921e60a7fa0a6ec758435e00c0078f545.jpg) +Figure 3: Qualitative comparison of semantic segmentation results. Our PARTSTAD segments 3D parts more precisely with clearer boundaries, even for small (Camera, Chair) and thin (Clock) parts. + +Table 1: Quantitative comparison of the semantic segmentation results (mIoU(%) on the PartNet-Mobility dataset. Ours achieves a $7.0\%$ p improvement in average mIoU compared to PartSLIP [32], the SotA few-shot 3D segmentation method, while consistently increasing mIoU across all categories. Please refer to the supplementary material for the complete table with results for all 45 categories. + +
MethodmIoUStorage FurnitureTableChairSwitchToiletLaptopUSBRemoteScissors
SATR [2]29.320.623.333.121.417.611.230.217.236.8
SATR [2]+SP34.828.928.037.737.022.112.433.428.043.0
PartSLIP [32]58.052.344.682.852.150.431.252.136.661.4
Ablation Study
w/o Weight Pred.61.956.645.085.051.956.631.557.136.660.8
w/o SAM Integ.62.154.045.783.153.049.433.653.646.867.5
PARTSTAD (Ours)65.059.547.885.357.957.534.659.953.468.5
+ +and OmniObject3D [54] dataset. In the supplementary material, we present additional results on the ablation study, outcomes with scanned data, and comprehensive results covering all PartNet-Mobility categories. + +# 5.1 Experiment Setup + +Dataset. We use the PartNet-Mobility [56] dataset, which is also used in the experiments of PartSLIP [32]. The PartNet-Mobility [56] dataset includes 45 object categories. For each category, the training split includes 8 shapes, while the test split has a range of the number of objects from 6 to 338, totaling 1,906 objects across all 45 categories. Following PartSLIP, 8 shapes from the test set of each category are designated as the validation set. To obtain the 2D bounding boxes for each object, we use 10 images from fixed viewpoints across all our experiments. Note that the PartNet-Mobility [56] dataset is identical to the subset + +of PartNet-Ensembled dataset used to train and evaluate PartSLIP [32]. (The remaining part of PartNet-Ensembled was only used to train other supervised baseline models, not PartSLIP [32].) + +Evaluation Metrics. As evaluation metrics, we employ the mean Intersection over Union (mIoU) metric for semantic segmentation and mean Average Precision (mAP) for instance segmentation. + +For semantic segmentation, the average mIoU over the entire dataset is computed by first calculating it per semantic part, averaging them across the parts in the same category, resulting in category-level mIoU, and then averaging the category-level mIoUs again across all the categories. + +For instance segmentation, we compute two types of mAP: part-aware mAP and part-agnostic mAP. For part-aware mAP, similar to mIoU in semantic segmentation, we first calculate the mAP for each semantic part with instances having the part label. We then average them for each category, and then across categories. For part-agnostic mAP, we do not consider the semantic labels and directly compute the category-level mAP with part instances in the same object category, and then average them. All mAP values correspond to $\mathrm{mAP}_{50}$ with an IoU threshold of $50\%$ . + +Baselines. We compare our method with three baselines: PartSLIP [32], SATR [2], and SAM3D [63]. + +- PartSLIP [32] is the method that we adopt as our base. Note that PartSLIP uses a finetuned GLIP model, and we also employ the same model in our framework to extract bounding box features for our task adaptation. +- SATR [2] is similar to PartSLIP, conducting 3D segmentation using 2D bounding boxes from GLIP [25]. However, SATR differs from PartSLIP in four key aspects: 1) it does not involve finetuning the GLIP model, 2) it does not utilize super points, 3) it is designed for mesh segmentation, not point cloud segmentation, and 4) it employs a distinct approach for integrating 2D bounding boxes into 3D (label propagation). To ensure a fair comparison between our method and SATR, we implement the following changes to SATR. First, since we employ the GLIP model finetuned by PartSLIP in our framework, we use the same finetuned model within the SATR framework. Second, we convert the point cloud input to a mesh, performing mesh segmentation with SATR, and apply the results to the input point cloud by finding the closest vertex for each point. Lastly, we introduce a version of SATR using super points, resulting in a comparison between our method and two versions of SATR: one without super points (denoted as $SATR$ ) and one with super points (denoted as $SATR + SP$ ). Since SATR does not perform instance segmentation, our comparison with it is focused on semantic segmentation. +- SAM3D [63] is another zero-shot 3D segmentation method using a 2D segmentation model, not GLIP [25], but SAM [20]. Since it only performs instance segmentation but does not label each segment, we compare our + +method with it only for instance segmentation, using the part-agnostic mAP metric. + +Ablation Study. We also conduct an ablation study, comparing our method with two major components being ablated: the weight prediction (Sec. 4.2) and the SAM mask integration (Sec. 4.3). In the supplementary material, we also compare our method with cases using cross-entropy loss instead of mR IoU loss (Sec. 4.1) and using GLIP bounding box confidence score instead of our learned weights in the weighted voting (Sec. 4.2). Note that if both weight prediction and SAM mask integration are excluded, our PARTSTAD is identical to PartSLIP [32]. + +# 5.2 Semantic Segmentation Results + +Tab. 1 and Fig. 3 present the quantitative and qualitative comparisons of part semantic segmentation results, respectively. Please refer to the supplementary material for the complete table encompassing all 45 categories. In comparison to PartSLIP [32], a SotA few-shot 3D segmentation method, we achieve a significant $7.0\%$ p mIoU improvement over the entire set of 45 categories. Furthermore, for each specific category, we consistently demonstrate improvement, surpassing by more than $15\%$ p in some categories (e.g., Remote). Qualitatively, the results also highlight the effectiveness of our task adaptation and the SAM mask integration in segmenting parts more precisely, even for small (e.g., Camera, Chair) and thin (e.g., Clock) parts of objects. + +SATR exhibits inferior performance, even when leveraging the finetuned GLIP model and super points (see SATR+SP results) due to the label-propagation-based 2D bounding box integration, which often fails to provide clear boundaries in 3D segmentation, as illustrated in Fig. 3. + +The ablation study results demonstrate the influence of each major component in our framework. The quantitative findings indicate that the weight prediction network plays a crucial role in substantial improvement, leading to a $3.1\%$ decrease in average mIoU when it is ablated. The SAM mask integration also contributes significantly, resulting in a $2.9\%$ average mIoU decrease when ablated. + +# 5.3 Instance Segmentation Results + +Following the instance segmentation idea of PartSLIP, we also perform instance segmentation using our method and compare the quantitative results with those of other methods in Tab. 2 and Tab. 3, reporting part-aware mAPs and part-agnostic mAPs, respectively. Fig. 4 also displays qualitative comparisons across different methods. Please refer to the supplementary material for the complete table of all 45 categories. Similar to semantic segmentation, ours achieves $4.0\%$ p and $5.2\%$ p improvements for both average part-aware mAPs and part-agnostic mAPs, respectively. Both the weight prediction network and SAM mask integration also exhibit meaningful improvements in performance, as their influence is + +![](images/a6d16926218f69eb912161d12a84bfd50612da83e1ef4eb6d887164cad58f27f.jpg) +Figure 4: Qualitative comparison of instance segmentation results shows that our PARTSTAD successfully segments tiny 3D parts, such as the keys of keyboards and buttons of remote controls, with clear segment boundaries. + +Table 2: Part-aware mAPs(\%) of the instance segmentation results on the PartNet-Mobility dataset. Ours achieves $4.0\%$ p improvement in mean part-aware mAPs compared to PartSLIP [32], with consistent enhancements across all categories. Please refer to the supplementary material for the complete table with results for all 45 categories. + +
MethodmAPStorage FurnitureTableChairSwitchToiletLaptopUSBRemoteScissors
PartSLIP [32]41.629.132.682.221.236.217.820.919.923.6
Ablation Study
w/o Weight Pred.44.733.830.882.522.240.824.417.420.325.5
w/o SAM Integ.44.229.432.382.219.736.023.624.829.623.1
PARTSTAD(Ours)45.635.533.282.522.140.628.926.533.726.5
+ +Table 3: Part-agnostic mAPs(%) of the instance segmentation results on the PartNet-Mobility dataset. Ours demonstrates $5.2\%$ p improvement in mean part-agnostic mAPs compared to PartSLIP [32], surpassing SAM3D with a mean mAP that is more than three times larger. Please refer to the supplementary material for the complete table with results for all 45 categories. (*SAM3D is a zero-shot method.) + +
MethodmAPStorage FurnitureTableChairSwitchToiletLaptopUSBRemoteScissors
SAM3D* [63]12.11.210.65.55.43.71.522.61.07.2
PartSLIP [32]38.929.728.780.721.235.419.520.519.927.0
Ablation Study
w/o Weight Pred.42.635.929.281.122.240.026.515.220.329.4
w/o SAM Integ.42.634.827.681.219.736.627.427.329.624.9
PARTSTAD (Ours)44.141.728.283.322.441.034.123.533.726.2
+ +depicted in the ablation study results. Compared to SAM3D [63], whose part- + +![](images/6a310f47bdd29d93f1e3a2cfeea8e30a524f4d32022b4b55cd3941fbda53abfe.jpg) +Figure 5: Qualitative comparison of semantic segmentation results on OmniObect3D [54] dataset, a high quality real scanned 3D objects dataset. Our PART-STD predicts precise boundaries, even in the case that the object has an appearance significantly different from training data such as box category. + +agnostic mAPs are shown in Tab. 3, our method achieves much better performance with more than three times greater mAP. Qualitative results also illustrate the outstanding performance of our method, accurately identifying all the tiny instances of parts, such as keys on keyboards, buttons on remotes, and clear boundaries of these instances. + +# 5.4 Part Segmentation with Real Scanned Dataset + +Fig. 5 displays additional results on the OmniObject3D [54] dataset, a high-quality dataset featuring real-scanned 3D objects. We present qualitative results only, as ground truth segmentation for OmniObject3D [54] is unavailable. Note that our PARTSTAD accurately identifies parts in real scans, demonstrating its effectiveness beyond synthetic data. Particularly impressive is its performance in the box category depicted in Fig. 5, where despite significant discrepancies in appearance compared to the training data, PARTSTAD achieves remarkably accurate predictions. + +# 6 Conclusion and Future Work + +We presented PARTSTAD, a task adaptation method for lifting 2D segmentation to 3D. Instead of directly finetuning a 2D segmentation network, our method learns a small neural network predicting weights for each 2D bounding box with an objective function for 3D segmentation, and then performs 3D segmentation via weighted merging of the boxes in 3D. We further improve the performance of 3D segmentation by integrating SAM foreground masks for each bounding box. We achieve the SotA results in few-shot 3D part segmentation, demonstrating significant improvements in both semantic and instance segmentations. + +As PARTSTAD still depends on 2D representations for 3D predictions, its understanding of 3D geometry is relatively limited, and thus it cannot account for occluded or interior points, which poses a limitation. In future work, we plan to further investigate combining our approach with the direct utilization of 3D representations, such as leveraging 3D features, for more accurate predictions. + +# Acknowledgements + +We sincerely thank Minghua Liu for providing the code of PartSLIP and answering questions. This work was supported by NRF grant (RS-2023-00209723), IITP grants (2022-0-00594, RS-2023-00227592, RS-2024-00399817), and Alchymist Project Program (RS-2024-00423625) funded by the Korean government (MSIT and MOTIE), and grants from the DRB-KAIST SketchTheFuture Research Center, Naver-intel, Adobe Research, Hyundai NGV, KT, and Samsung Electronics. + +# References + +1. Abdelreheem, A., Eldesokey, A., Ovsjanikov, M., Wonka, P.: Zero-shot 3D shape correspondence. In: SIGGRAPH Asia (2023) +2. Abdelreheem, A., Skorokhodov, I., Ovsjanikov, M., Wonka, P.: SATR: Zero-shot semantic segmentation of 3D shapes. In: ICCV (2023) +3. Armeni, I., Sax, S., Zamir, A.R., Savarese, S.: Joint 2D-3D-semantic data for indoor scene understanding. arXiv preprint arXiv:1702.01105 (2017) +4. Baranchuk, D., Voynov, A., Rubachev, I., Khrulkov, V., Babenko, A.: Label-efficient semantic segmentation with diffusion models. In: ICLR (2022) +5. Behley, J., Garbade, M., Milioto, A., Quenzel, J., Behnke, S., Stachniss, C., Gall, J.: SemanticKITTI: A dataset for semantic scene understanding of LiDAR sequences. In: ICCV (2019) +6. Cen, J., Zhou, Z., Fang, J., Yang, C., Shen, W., Xie, L., Zhang, X., Tian, Q.: Segment anything in 3D with nerfs. In: NeurIPS (2023) +7. Chang, A., Dai, A., Funkhouser, T., Halber, M., Niessner, M., Savva, M., Song, S., Zeng, A., Zhang, Y.: Matterport3D: Learning from rgb-d data in indoor environments. In: 3DV (2017) +8. Chang, A.X., Funkhouser, T., Guibas, L., Hanrahan, P., Huang, Q., Li, Z., Savarese, S., Savva, M., Song, S., Su, H., et al.: ShapeNet: An information-rich 3D model repository. arXiv preprint arXiv:1512.03012 (2015) +9. Chen, R., Liu, Y., Kong, L., Zhu, X., Ma, Y., Li, Y., Hou, Y., Qiao, Y., Wang, W.: CLIP2Scene: Towards label-efficient 3D scene understanding by clip. In: CVPR (2023) +0. Chibane, J., Engelmann, F., Tran, T.A., Pons-Moll, G.: Box2Mask: Weakly supervised 3D semantic instance segmentation using bounding boxes. In: ECCV (2022) +1. Chu, R., Chen, Y., Kong, T., Qi, L., Li, L.: ICM-3D: Instantiated category modeling for 3D instance segmentation. IEEE Robotics and Automation Letters (2021) +2. Dai, A., Chang, A.X., Savva, M., Halber, M., Funkhouser, T., Nießner, M.: ScanNet: Richly-annotated 3D reconstructions of indoor scenes. In: CVPR (2017) +3. Decatur, D., Lang, I., Hanocka, R.: 3D Highlighter: Localizing regions on 3D shapes via text descriptions. In: CVPR (2023) +4. Hackel, T., Savinov, N., Ladicky, L., Wegner, J.D., Schindler, K., Pollefeys, M.: SEMANTIC3D.NET: A new large-scale point cloud classification benchmark. In: ISPRS Annals of the Photogrammetry, Remote Sensing and Spatial Information Sciences (2017) +5. Hou, J., Dai, A., Nießner, M.: 3D-SIS: 3D semantic instance segmentation of rgb-d scans. In: CVPR (2019) + +16. Hu, E.J., Shen, Y., Wallis, P., Allen-Zhu, Z., Li, Y., Wang, S., Wang, L., Chen, W.: LoRA: Low-rank adaptation of large language models. arXiv preprint arXiv:2106.09685 (2021) +17. Huang, R., Huang, C., Liu, Y., Dai, G., Kong, W.: LSGCN: Long short-term traffic prediction with graph convolutional networks. In: IJCAI (2020) +18. Huang, Z., Wu, X., Chen, X., Zhao, H., Zhu, L., Lasenby, J.: OpenIns3D: Snap and lookup for 3D open-vocabulary instance segmentation. arXiv preprint (2023) +19. Jiang, L., Zhao, H., Shi, S., Liu, S., Fu, C.W., Jia, J.: PointGroup: Dual-set point grouping for 3D instance segmentation. In: CVPR (2020) +20. Kirillov, A., Mintun, E., Ravi, N., Mao, H., Rolland, C., Gustafson, L., Xiao, T., Whitehead, S., Berg, A.C., Lo, W.Y., et al.: Segment anything. arXiv preprint arXiv:2304.02643 (2023) +21. Koo, J., Huang, I., Achlioptas, P., Guibas, L.J., Sung, M.: PartGlot: Learning shape part segmentation from language reference games. In: CVPR (2022) +22. Lai, X., Tian, Z., Chen, Y., Li, Y., Yuan, Y., Liu, S., Jia, J.: LISA: Reasoning segmentation via large language model. arXiv preprint arXiv:2308.00692 (2023) +23. Lé, E.T., Sung, M., Ceylan, D., Mech, R., Boubekeur, T., Mitra, N.J.: CPFN: Cascaded primitive fitting networks for high-resolution point clouds. In: ICCV (2021) +24. Li, L., Sung, M., Dubrovina, A., Yi, L., Guibas, L.J.: Supervised fitting of geometric primitives to 3D point clouds. In: CVPR (2019) +25. Li, L.H., Zhang, P., Zhang, H., Yang, J., Li, C., Zhong, Y., Wang, L., Yuan, L., Zhang, L., Hwang, J.N., et al.: Grounded language-image pre-training. In: CVPR (2022) +26. Li, M., Xie, Y., Shen, Y., Ke, B., Qiao, R., Ren, B., Lin, S., Ma, L.: Hybridcr: Weakly-supervised 3d point cloud semantic segmentation via hybrid contrastive regularization. In: CVPR (2022) +27. Li, Y., Bu, R., Sun, M., Wu, W., Di, X., Chen, B.: PointCNN: Convolution on x-transformed points. In: NeurIPS (2018) +28. Li, Y., Ma, L., Zhong, Z., Cao, D., Li, J.: TGNet: Geometric graph cnn on 3D point cloud segmentation. IEEE Transactions on Geoscience and Remote Sensing (2019) +29. Liang, Z., Yang, M., Li, H., Wang, C.: 3D instance embedding learning with a structure-aware loss function for point cloud segmentation. IEEE Robotics and Automation Letters (2020) +30. Lin, T.Y., Maire, M., Belongie, S., Hays, J., Perona, P., Ramanan, D., Dólar, P., Zitnick, C.L.: Microsoft COCO: Common objects in context. In: ECCV (2014) +31. Liu, J., Yu, M., Ni, B., Chen, Y.: Self-prediction for joint instance and semantic segmentation of point clouds. In: ECCV (2020) +32. Liu, M., Zhu, Y., Cai, H., Han, S., Ling, Z., Porikli, F., Su, H.: PartSLIP: Low-shot part segmentation for 3d point clouds via pretrained image-language models. In: CVPR (2023) +33. Liu, S., Zeng, Z., Ren, T., Li, F., Zhang, H., Yang, J., Li, C., Yang, J., Su, H., Zhu, J., et al.: Grounding DINO: Marrying dino with grounded pre-training for open-set object detection. arXiv preprint arXiv:2303.05499 (2023) +34. Liu, Y., Fan, B., Meng, G., Lu, J., Xiang, S., Pan, C.: DensePoint: Learning densely contextual representation for efficient point cloud processing. In: ICCV (2019) +35. Liu, Z., Qi, X., Fu, C.W.: One thing one click: A self-training approach for weakly supervised 3D semantic segmentation. In: CVPR (2021) +36. Ma, X., Qin, C., You, H., Ran, H., Fu, Y.: Rethinking network design and local geometry in point cloud: A simple residual mlp framework. In: ICLR (2022) + +37. Mo, K., Zhu, S., Chang, A.X., Yi, L., Tripathi, S., Guibas, L.J., Su, H.: PartNet: A large-scale benchmark for fine-grained and hierarchical part-level 3D object understanding. In: CVPR (2019) +38. Narita, G., Seno, T., Ishikawa, T., Kaji, Y.: Panoptic fusion: Online volumetric semantic mapping at the level of stuff and things. In: IROS (2019) +39. Peng, S., Genova, K., Jiang, C., Tagliasacchi, A., Pollefeys, M., Funkhouser, T., et al.: OpenScene: 3D scene understanding with open vocabularies. In: CVPR (2023) +40. Pham, Q.H., Nguyen, D.T., Hua, B.S., Roig, G., Yeung, S.K.: JSIS3D: Joint semantic-instance segmentation of 3D point clouds with multi-task pointwise networks and multi-value conditional random fields. In: CVPR (2019) +41. Pruksachatkun, Y., Phang, J., Liu, H., Htut, P.M., Zhang, X., Pang, R.Y., Vania, C., Kann, K., Bowman, S.R.: Intermediate-Task Transfer Learning with Pretrained Language Models: When and why does it work? In: ACL (2020) +42. Qi, C.R., Su, H., Mo, K., Guibas, L.J.: PointNet: Deep learning on point sets for 3d classification and segmentation. In: CVPR (2017) +43. Qi, C.R., Yi, L., Su, H., Guibas, L.J.: PointNet++: Deep hierarchical feature learning on point sets in a metric space. In: NeurIPS (2017) +44. Qian, G., Li, Y., Peng, H., Mai, J., Hammoud, H., Elhoseiny, M., Ghanem, B.: PointNext: Revisiting pointnet++ with improved training and scaling strategies. In: NeurIPS (2022) +45. Radford, A., Kim, J.W., Hallacy, C., Ramesh, A., Goh, G., Agarwal, S., Sastry, G., Askell, A., Mishkin, P., Clark, J., et al.: Learning transferable visual models from natural language supervision. In: ICML (2021) +46. Schult, J., Engelmann, F., Hermans, A., Litany, O., Tang, S., Leibe, B.: Mask3D: Mask transformer for 3D semantic instance segmentation. In: ICRA (2023) +47. Takmaz, A., Fedele, E., Sumner, R.W., Pollefeys, M., Tombari, F., Engelmann, F.: OpenMask3D: Open-vocabulary 3D instance segmentation. In: NeurIPS (2023) +48. Tang, L., Jia, M., Wang, Q., Phoo, C., Hariharan, B.: Emergent correspondence from image diffusion. In: NeurIPS (2023) +49. Thomas, H., Qi, C.R., Deschaud, J.E., Marcotegui, B., Goulette, F., Guibas, L.J.: Kpconv: Flexible and deformable convolution for point clouds. In: ICCV (2019) +50. Vu, T., Kim, K., Luu, T.M., Nguyen, X.T., Yoo, C.D.: Softgroup for 3d instance segmentation on 3D point clouds. In: CVPR (2022) +51. Wang, W., Yu, R., Huang, Q., Neumann, U.: SGPN: Similarity group proposal network for 3D point cloud instance segmentation. In: CVPR (2018) +52. Wang, X., Liu, S., Shen, X., Shen, C., Jia, J.: Associatively segmenting instances and semantics in point clouds. In: CVPR (2019) +53. Wang, Y., Sun, Y., Liu, Z., Sarma, S.E., Bronstein, M.M., Solomon, J.M.: Dynamic graph cnn for learning on point clouds. ACM TOG (2019) +54. Wu, T., Zhang, J., Fu, X., Wang, Y., Ren, J., Pan, L., Wu, W., Yang, L., Wang, J., Qian, C., Lin, D., Liu, Z.: OmniObject3D: Large-vocabulary 3D object dataset for realistic perception, reconstruction and generation. In: CVPR (2023) +55. Wu, X., Lao, Y., Jiang, L., Liu, X., Zhao, H.: Point transformer v2: Grouped vector attention and partition-based pooling. In: NeurIPS (2022) +56. Xiang, F., Qin, Y., Mo, K., Xia, Y., Zhu, H., Liu, F., Liu, M., Jiang, H., Yuan, Y., Wang, H., Yi, L., Chang, A.X., Guibas, L.J., Su, H.: SAPIEN: A simulated part-based interactive environment. In: CVPR (2020) +57. Xiang, T., Zhang, C., Song, Y., Yu, J., Cai, W.: Walk in the Cloud: Learning curves for point clouds shape analysis. In: ICCV (2021) + +58. Xu, J., Liu, S., Vahdat, A., Byeon, W., Wang, X., De Mello, S.: Open-vocabulary panoptic segmentation with text-to-image diffusion models. In: CVPR (2023) +59. Xu, M., Ding, R., Zhao, H., Qi, X.: PACov: Position adaptive convolution with dynamic kernel assembling on point clouds. In: CVPR (2021) +60. Yan, X., Zheng, C., Li, Z., Wang, S., Cui, S.: PointANSL: Robust point clouds processing using nonlocal neural networks with adaptive sampling. In: CVPR (2020) +61. Yang, B., Wang, J., Clark, R., Hu, Q., Wang, S., Markham, A., Trigoni, N.: Learning object bounding boxes for 3D instance segmentation on point clouds. In: NeurIPS (2019) +62. Yang, X., Wang, X.: Diffusion model as representation learner. In: ICCV (2023) +63. Yang, Y., Wu, X., He, T., Zhao, H., Liu, X.: SAM3D: Segment anything in 3D scenes. arXiv preprint arXiv:2306.03908 (2023) +64. Yi, K.M., Trulls, E., Ono, Y., Lepetit, V., Salzmann, M., Fua, P.: Learning to find good correspondences. In: CVPR (2018) +65. Yi, L., Huang, H., Liu, D., Kalogerakis, E., Su, H., Guibas, L.: Deep part induction from articulated object pairs. ACM TOG (2018) +66. Yi, L., Zhao, W., Wang, H., Sung, M., Guibas, L.J.: GSPN: Generative shape proposal network for 3D instance segmentation in point cloud. In: CVPR (2019) +67. Yu, X., Tang, L., Rao, Y., Huang, T., Zhou, J., Lu, J.: Point-BERT: Pre-training 3D point cloud transformers with masked point modeling. In: CVPR (2022) +68. Zamir, A.R., Sax, A., Shen, W.B., Guibas, L.J., Malik, J., Savarese, S.: Taskonomy: Disentangling task transfer learning. In: CVPR (2018) +69. Zhang, B., Wonka, P.: Point cloud instance segmentation using probabilistic embeddings. In: CVPR (2021) +70. Zhang, J., Dong, R., Ma, K.: CLIP-FO3D: Learning free open-world 3D scene representations from 2D dense clip. arXiv preprint arXiv:2303.04748 (2023) +71. Zhang, R., Wang, L., Wang, Y., Gao, P., Li, H., Shi, J.: Parameter is not all you need: Starting from non-parametric networks for 3D point cloud analysis. In: CVPR (2023) +72. Zhang, Y., Ling, H., Gao, J., Yin, K., Lafleche, J.F., Barriuso, A., Torralba, A., Fidler, S.: DatasetGAN: Efficient labeled data factory with minimal human effort. In: CVPR (2021) +73. Zhao, H., Jiang, L., Jia, J., Torr, P.H., Koltun, V.: Point transformer. In: ICCV (2021) +74. Zhao, N., Chua, T.S., Lee, G.H.: Few-shot 3D point cloud semantic segmentation. In: CVPR (2021) \ No newline at end of file diff --git a/2024/PartSTAD_ 2D-to-3D Part Segmentation Task Adaptation/images.zip b/2024/PartSTAD_ 2D-to-3D Part Segmentation Task Adaptation/images.zip new file mode 100644 index 0000000000000000000000000000000000000000..83756ac04ee8fb4f7fe00b08006025e88b52bf9c --- /dev/null +++ b/2024/PartSTAD_ 2D-to-3D Part Segmentation Task Adaptation/images.zip @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:5844858b15a45b5668b85e6af6fa3f8ceb64007b97175956b53b031ac90e9659 +size 411158 diff --git a/2024/PartSTAD_ 2D-to-3D Part Segmentation Task Adaptation/layout.json b/2024/PartSTAD_ 2D-to-3D Part Segmentation Task Adaptation/layout.json new file mode 100644 index 0000000000000000000000000000000000000000..03e08a408f166d26d9ef52cbdf98f60d5ab8d7c5 --- /dev/null +++ b/2024/PartSTAD_ 2D-to-3D Part Segmentation Task Adaptation/layout.json @@ -0,0 +1,9389 @@ +{ + "pdf_info": [ + { + "para_blocks": [ + { + "bbox": [ + 146, + 111, + 470, + 148 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 146, + 111, + 470, + 148 + ], + "spans": [ + { + "bbox": [ + 146, + 111, + 470, + 148 + ], + "type": "text", + "content": "PartSTAD: 2D-to-3D Part Segmentation Task Adaptation" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 220, + 167, + 394, + 181 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 220, + 167, + 394, + 181 + ], + "spans": [ + { + "bbox": [ + 220, + 167, + 394, + 181 + ], + "type": "text", + "content": "Hyunjin " + }, + { + "bbox": [ + 220, + 167, + 394, + 181 + ], + "type": "inline_equation", + "content": "\\mathrm{Kim}^{1\\dagger}" + }, + { + "bbox": [ + 220, + 167, + 394, + 181 + ], + "type": "text", + "content": " and Minhyuk " + }, + { + "bbox": [ + 220, + 167, + 394, + 181 + ], + "type": "inline_equation", + "content": "\\mathrm{Sung}^{2}" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 240, + 189, + 373, + 200 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 240, + 189, + 373, + 200 + ], + "spans": [ + { + "bbox": [ + 240, + 189, + 373, + 200 + ], + "type": "text", + "content": "1 KRAFTON Inc., South Korea" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 258, + 201, + 355, + 211 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 258, + 201, + 355, + 211 + ], + "spans": [ + { + "bbox": [ + 258, + 201, + 355, + 211 + ], + "type": "text", + "content": "2 KAIST, South Korea" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 205, + 213, + 408, + 223 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 205, + 213, + 408, + 223 + ], + "spans": [ + { + "bbox": [ + 205, + 213, + 408, + 223 + ], + "type": "text", + "content": "rlaguswls98@krafton.com, mhsung@kaist.ac.kr" + } + ] + } + ], + "index": 4 + }, + { + "type": "image", + "bbox": [ + 135, + 230, + 482, + 338 + ], + "blocks": [ + { + "bbox": [ + 135, + 230, + 482, + 338 + ], + "lines": [ + { + "bbox": [ + 135, + 230, + 482, + 338 + ], + "spans": [ + { + "bbox": [ + 135, + 230, + 482, + 338 + ], + "type": "image", + "image_path": "826464d72d32b4853d3b3fdf9c02fc29e39f268677baaa949e049661e789fe67.jpg" + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 131, + 348, + 482, + 404 + ], + "lines": [ + { + "bbox": [ + 131, + 348, + 482, + 404 + ], + "spans": [ + { + "bbox": [ + 131, + 348, + 482, + 404 + ], + "type": "text", + "content": "Figure 1: We introduce PARTSTAD, a novel few-shot 3D point cloud part segmentation method that leverages 2D-to-3D task adaptation. By obtaining 2D segmentation masks in multi-view images from GLIP [25] and SAM [20] and optimizing the mask weights for 3D segmentation as a learning objective, it can successfully predict fine-grained parts with accurate boundaries, as shown in the figure above." + } + ] + } + ], + "index": 6, + "angle": 0, + "type": "image_caption" + } + ], + "index": 5 + }, + { + "bbox": [ + 159, + 419, + 455, + 628 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 159, + 419, + 455, + 628 + ], + "spans": [ + { + "bbox": [ + 159, + 419, + 455, + 628 + ], + "type": "text", + "content": "Abstract. We introduce PARTSTAD, a method designed for the task adaptation of 2D-to-3D segmentation lifting. Recent studies have highlighted the advantages of utilizing 2D segmentation models to achieve high-quality 3D segmentation through few-shot adaptation. However, previous approaches have focused on adapting 2D segmentation models for domain shift to rendered images and synthetic text descriptions, rather than optimizing the model specifically for 3D segmentation. Our proposed task adaptation method finetunes a 2D bounding box prediction model with an objective function for 3D segmentation. We introduce weights for 2D bounding boxes for adaptive merging and learn the weights using a small additional neural network. Additionally, we incorporate SAM, a foreground segmentation model on a bounding box, to improve the boundaries of 2D segments and consequently those of 3D segmentation. Our experiments on the PartNet-Mobility dataset show significant improvements with our task adaptation approach, achieving a " + }, + { + "bbox": [ + 159, + 419, + 455, + 628 + ], + "type": "inline_equation", + "content": "7.0\\%" + }, + { + "bbox": [ + 159, + 419, + 455, + 628 + ], + "type": "text", + "content": "p increase in mIoU and a " + }, + { + "bbox": [ + 159, + 419, + 455, + 628 + ], + "type": "inline_equation", + "content": "5.2\\%" + }, + { + "bbox": [ + 159, + 419, + 455, + 628 + ], + "type": "text", + "content": "p improvement in " + }, + { + "bbox": [ + 159, + 419, + 455, + 628 + ], + "type": "inline_equation", + "content": "\\mathrm{mAP}_{50}" + }, + { + "bbox": [ + 159, + 419, + 455, + 628 + ], + "type": "text", + "content": " for semantic and instance segmentation compared to the SotA few-shot 3D segmentation model. The code is available at https://github.com/KAIST-Visual-AI-Group/PartSTAD." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 160, + 638, + 445, + 650 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 160, + 638, + 445, + 650 + ], + "spans": [ + { + "bbox": [ + 160, + 638, + 445, + 650 + ], + "type": "text", + "content": "Keywords: part segmentation " + }, + { + "bbox": [ + 160, + 638, + 445, + 650 + ], + "type": "inline_equation", + "content": "\\cdot" + }, + { + "bbox": [ + 160, + 638, + 445, + 650 + ], + "type": "text", + "content": " few-shot learning " + }, + { + "bbox": [ + 160, + 638, + 445, + 650 + ], + "type": "inline_equation", + "content": "\\cdot" + }, + { + "bbox": [ + 160, + 638, + 445, + 650 + ], + "type": "text", + "content": " 3D deep learning" + } + ] + } + ], + "index": 8 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 135, + 653, + 382, + 664 + ], + "type": "page_footnote", + "angle": 0, + "lines": [ + { + "bbox": [ + 135, + 653, + 382, + 664 + ], + "spans": [ + { + "bbox": [ + 135, + 653, + 382, + 664 + ], + "type": "text", + "content": "This work was conducted when the author was at KAIST." + } + ] + } + ], + "index": 9 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 0 + }, + { + "para_blocks": [ + { + "bbox": [ + 133, + 114, + 230, + 128 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 114, + 230, + 128 + ], + "spans": [ + { + "bbox": [ + 133, + 114, + 230, + 128 + ], + "type": "text", + "content": "1 Introduction" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 130, + 139, + 481, + 247 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 139, + 481, + 247 + ], + "spans": [ + { + "bbox": [ + 130, + 139, + 481, + 247 + ], + "type": "text", + "content": "3D segmentation has been a subject of extensive research in computer vision due to its central role in various downstream applications involving the understanding of shape structure, functionality, mobility, and semantics. However, the limited availability of annotated 3D shapes has remained a bottleneck in achieving generalizability in learned segmentation models for diverse 3D data. Annotating 3D segmentation is particularly labor-intensive, time-consuming, and requires expertise in handling 3D models. For this reason, the scale of 3D part-level annotations remains in the tens of thousands [37], while the scale of 2D annotation datasets in the image domain, for example, exceeds a million [20, 30]." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 130, + 247, + 481, + 354 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 247, + 481, + 354 + ], + "spans": [ + { + "bbox": [ + 130, + 247, + 481, + 354 + ], + "type": "text", + "content": "Recent research [2,32] has illuminated the potential of visual-language models bridging textual descriptions with images in accomplishing zero-shot or few-shot 3D segmentation results. The basic idea is to render a 3D model from various viewpoints, conduct 2D detection or segmentation on the rendered images, and then aggregate the 2D segmentation results into a 3D representation using either a voting mechanism [32] or a label propagation scheme [2]. This approach is not only effective in enabling zero-shot and few-shot 3D segmentation but also offers an advantage in that the set of part names does not need to be predefined in training but can be determined at test time." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 130, + 354, + 481, + 498 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 354, + 481, + 498 + ], + "spans": [ + { + "bbox": [ + 130, + 354, + 481, + 498 + ], + "type": "text", + "content": "PartSLIP [32] is a notable example that achieves 3D part segmentation results comparable to fully supervised methods while adapting a pretrained 2D detection model with few-shot training for 3D segmentation. It leverages the pretrained GLIP [25] model for 2D detection, associating the output 2D bounding boxes with one of the tokens (part names) provided in the input prompt. Its key component is the GLIP finetuning process, which introduces a small number of learnable parameters to the frozen GLIP model and trains them using few-shot synthetic images and texts employed in the 3D segmentation pipeline. As GLIP was initially trained with real photos and natural language descriptions of objects by humans, this adaptation to rendered images and the unconventional description (a sequence of part names) results in a substantial improvement in the 3D segmentation task." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 130, + 498, + 481, + 617 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 498, + 481, + 617 + ], + "spans": [ + { + "bbox": [ + 130, + 498, + 481, + 617 + ], + "type": "text", + "content": "Despite promising initiatives, it is important to note that the previous approach was limited to achieving domain adaptation. In contrast, we emphasize that in the 2D-to-3D segmentation lifting task, not only does the data domain change, but the task itself also shifts from 2D segmentation to 3D segmentation. Thus, the process of applying the pretrained model to the new task requires task adaptation, involving modifying the model with an objective function associated with the new task. Particularly for 3D segmentation, it is necessary to integrate 2D segmentation results from multiple viewpoints into a coherent 3D representation. Therefore, it is crucial to control the noise of 2D bounding box predictions in the context of its impact on the final 3D segmentation after integration." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 130, + 617, + 481, + 666 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 617, + 481, + 666 + ], + "spans": [ + { + "bbox": [ + 130, + 617, + 481, + 666 + ], + "type": "text", + "content": "Specifically, our Part Segmentation Task ADaptation method, PARTSTAD, adapts the pretrained GLIP model with a relaxed 3D mIoU loss while incorporating bounding boxes across all different views. Instead of typical finetuning, we draw inspiration from recent work [16, 32] that introduces small learnable" + } + ] + } + ], + "index": 7 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "spans": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "type": "text", + "content": "2" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 166, + 91, + 253, + 102 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 166, + 91, + 253, + 102 + ], + "spans": [ + { + "bbox": [ + 166, + 91, + 253, + 102 + ], + "type": "text", + "content": "H. Kim and M. Sung" + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 1 + }, + { + "para_blocks": [ + { + "bbox": [ + 130, + 116, + 482, + 247 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 116, + 482, + 247 + ], + "spans": [ + { + "bbox": [ + 130, + 116, + 482, + 247 + ], + "type": "text", + "content": "parameters to the existing model while keeping existing parameters frozen. As a result, we train a small MLP using 2D bounding box features extracted from the pretrained GLIP model. Since the relaxed mIoU loss is non-differentiable with respect to the bounding box positions, we suggest predicting a weight for each bounding box instead of adjusting its position. This results in minimal modification in the 2D-to-3D voting scheme while achieving a significant improvement in 3D segmentation, even with few-shot training, e.g., with eight objects per category, as done in PartSLIP. To further enhance 3D segmentation performance, we leverage SAM [20], a 2D instance segmentation model, to segment the foreground region within each bounding box, obtaining an accurate boundary for each 2D segment." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 130, + 247, + 482, + 331 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 247, + 482, + 331 + ], + "spans": [ + { + "bbox": [ + 130, + 247, + 482, + 331 + ], + "type": "text", + "content": "In our experiments on the PartNet-Mobility [56] dataset, we showcase the superior performance of our method in comparison to recent zero-shot/few-shot 3D semantic and instance segmentation methods that leverage 2D segmentation models. In contrast to the SotA few-shot method, our approach achieves a " + }, + { + "bbox": [ + 130, + 247, + 482, + 331 + ], + "type": "inline_equation", + "content": "7.0\\% \\mathbf{p}" + }, + { + "bbox": [ + 130, + 247, + 482, + 331 + ], + "type": "text", + "content": " improvement in semantic segmentation mIoU and a " + }, + { + "bbox": [ + 130, + 247, + 482, + 331 + ], + "type": "inline_equation", + "content": "5.2\\% \\mathbf{p}" + }, + { + "bbox": [ + 130, + 247, + 482, + 331 + ], + "type": "text", + "content": " improvement in instance segmentation " + }, + { + "bbox": [ + 130, + 247, + 482, + 331 + ], + "type": "inline_equation", + "content": "\\mathrm{mAP}_{50}" + }, + { + "bbox": [ + 130, + 247, + 482, + 331 + ], + "type": "text", + "content": ". These improvements are consistent across all object classes." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 132, + 350, + 237, + 363 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 350, + 237, + 363 + ], + "spans": [ + { + "bbox": [ + 132, + 350, + 237, + 363 + ], + "type": "text", + "content": "2 Related Work" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 132, + 377, + 307, + 389 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 377, + 307, + 389 + ], + "spans": [ + { + "bbox": [ + 132, + 377, + 307, + 389 + ], + "type": "text", + "content": "2.1 Supervised 3D Segmentation" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 130, + 398, + 482, + 541 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 398, + 482, + 541 + ], + "spans": [ + { + "bbox": [ + 130, + 398, + 482, + 541 + ], + "type": "text", + "content": "Given the availability of open segment-annotated 3D datasets [3,5,7,8,12,14,37, 56], 3D segmentation has been extensively researched in the last several years, focusing on the development of novel network architectures for supervised learning. Regarding the architecture of semantic segmentation, diverse models have been explored, including PointNet [42] and its variants [36,43,44,57,71], those using CNN [27,34,49,59], GCN [17,28,53], and Transformers [55,60,67,73]. For instance segmentation, various approaches based on proposing 3D bounding boxes [15,61,66], clustering learned features [11,19,31,38,50,51,69,74], and combining both semantic and instance segmentations [29,40,46,52] have been introduced. Despite these great advances, the performance of supervised methods has been limited by the scale of 3D datasets; the largest part-annotated 3D dataset, PartNet [37], includes fewer than " + }, + { + "bbox": [ + 130, + 398, + 482, + 541 + ], + "type": "inline_equation", + "content": "30\\mathrm{k}" + }, + { + "bbox": [ + 130, + 398, + 482, + 541 + ], + "type": "text", + "content": " models." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 130, + 542, + 482, + 601 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 542, + 482, + 601 + ], + "spans": [ + { + "bbox": [ + 130, + 542, + 482, + 601 + ], + "type": "text", + "content": "There have been some attempts to overcome the limitation by leveraging additional weak supervisions, such as text descriptions of objects [21], 3D bounding box annotations [10], sparse labels [26, 35], and geometric priors [23, 24]. However, the scale of this additional data has also been limited to tens of thousands, while datasets in the image domain are on the scale of millions." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 132, + 620, + 407, + 632 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 620, + 407, + 632 + ], + "spans": [ + { + "bbox": [ + 132, + 620, + 407, + 632 + ], + "type": "text", + "content": "2.2 3D Segmentation Using Vision-Language Models" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 130, + 641, + 481, + 665 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 641, + 481, + 665 + ], + "spans": [ + { + "bbox": [ + 130, + 641, + 481, + 665 + ], + "type": "text", + "content": "Recent work introduced ideas about exploiting vision-language models (VLM) for 3D segmentation. The advantage of utilizing the learned visual-language" + } + ] + } + ], + "index": 9 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 210, + 91, + 447, + 102 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 210, + 91, + 447, + 102 + ], + "spans": [ + { + "bbox": [ + 210, + 91, + 447, + 102 + ], + "type": "text", + "content": "PartSTAD: 2D-to-3D Part Segmentation Task Adaptation" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 474, + 92, + 480, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 474, + 92, + 480, + 100 + ], + "spans": [ + { + "bbox": [ + 474, + 92, + 480, + 100 + ], + "type": "text", + "content": "3" + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 2 + }, + { + "para_blocks": [ + { + "bbox": [ + 130, + 116, + 482, + 200 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 116, + 482, + 200 + ], + "spans": [ + { + "bbox": [ + 130, + 116, + 482, + 200 + ], + "type": "text", + "content": "grounding in 3D segmentation lies in the strong generalizability to arbitrary 3D models with zero-shot or few-shot training, and also in the open vocabulary understanding that enables labeling points without specifying the set of part or object names during training. For a VLM, CLIP [45] has been adapted in multiple previous works to lift the 2D grounding to the 3D, detecting parts in an object [13] and objects in a scene [9,39,47,70], while the results are limited to highlighting some regions without providing clear segment boundaries." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 130, + 202, + 482, + 346 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 202, + 482, + 346 + ], + "spans": [ + { + "bbox": [ + 130, + 202, + 482, + 346 + ], + "type": "text", + "content": "To obtain precise segments in 3D, other 2D object detection and segmentation models, such as SAM [20], GLIP [25], and GroundingDINO [33], are also leveraged. SAM3D [63] was an example of directly lifting the 2D masks from SAM to 3D with a bottom-up merging approach, although it was limited to segmenting the parts, not labeling them. SA3D [6] utilized NeRF representation and allowed user interaction for 3D segmentation, but it can be inefficient when the 3D objects are given as a mesh, point cloud, or other common 3D representations. OpenMask3D [47] and OpenIns3D [18] are concurrent works, both of which combine multiple foundation models, SAM [20] and CLIP [45] for OpenMask3D, and GroundingDINO [33] and LISA [22] for OpenIns3D. These works focus on segmenting objects in 3D scenes, while we aim to do finer-grained segmentation, finding parts in 3D objects." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 130, + 349, + 482, + 445 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 349, + 482, + 445 + ], + "spans": [ + { + "bbox": [ + 130, + 349, + 482, + 445 + ], + "type": "text", + "content": "PartSLIP [32] and SATR [2] are notable examples that achieve part segmentation of 3D objects based on a 2D object detection model, GLIP [25]. They obtain 2D bounding boxes on the rendered images from multiple views and integrate them over the 3D object with different merging schemes. The finetuning of GLIP introduced by PartSLIP particularly leads to a performance improvement, but it is limited to domain adaptation, not task adaptation. Building on this, we introduce a novel task adaptation technique as well as integration with SAM [20] to achieve a further significant improvement in 3D segmentation accuracy." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 132, + 475, + 246, + 488 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 475, + 246, + 488 + ], + "spans": [ + { + "bbox": [ + 132, + 475, + 246, + 488 + ], + "type": "text", + "content": "2.3 Task Adaptation" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 130, + 510, + 482, + 666 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 510, + 482, + 666 + ], + "spans": [ + { + "bbox": [ + 130, + 510, + 482, + 666 + ], + "type": "text", + "content": "Taskonomy [68] is a seminal work that introduced the concept of task transfer learning for the first time, enabling the transformation of an image dense prediction model trained for a base task to perform other tasks through finetuning. Pruksachatkun et al. [41] also introduced a similar idea for language-domain tasks. Later, task adaptation has been extensively studied to leverage features learned from generative models, such as GAN [72] and diffusion models [4, 48, 58, 62], adapting them with a small network for various tasks, including image segmentation [4, 58, 72], depth prediction [48], and keypoint detection [62]. For 3D tasks, Abdelreheem et al. [1] are notable examples of using vision language-grounding models for 3D correspondence but are limited to directly applying 2D priors without finetuning. We introduce a novel task adaptation technique that lifts 2D segments to 3D by adapting the 2D features during the training of a small network with a task-specific objective function." + } + ] + } + ], + "index": 6 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "spans": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "type": "text", + "content": "4" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 166, + 91, + 253, + 102 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 166, + 91, + 253, + 102 + ], + "spans": [ + { + "bbox": [ + 166, + 91, + 253, + 102 + ], + "type": "text", + "content": "H. Kim and M. Sung" + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 3 + }, + { + "para_blocks": [ + { + "bbox": [ + 132, + 114, + 334, + 129 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 114, + 334, + 129 + ], + "spans": [ + { + "bbox": [ + 132, + 114, + 334, + 129 + ], + "type": "text", + "content": "3 Background and Preliminaries" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 130, + 140, + 479, + 200 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 140, + 479, + 200 + ], + "spans": [ + { + "bbox": [ + 130, + 140, + 479, + 200 + ], + "type": "text", + "content": "Recent research [2,13,32] has demonstrated how vision-language models designed for image object detection and segmentation can be applied to segment parts in 3D objects. The main idea involves rendering a 3D model from various viewpoints, performing 2D detection or segmentation for the rendered image at each view, and then combining the 2D segmentation results over the 3D model." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 130, + 200, + 482, + 403 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 200, + 482, + 403 + ], + "spans": [ + { + "bbox": [ + 130, + 200, + 482, + 403 + ], + "type": "text", + "content": "Specifically, SATR [2] utilizes GLIP [25] as the 2D detection model. To obtain 2D bounding boxes for each of the specified part names, it feeds a concatenated list of these part names (separated by commas) as the input prompt to the pretrained GLIP model, along with a rendered image. The 2D bounding boxes are then mapped to a single face of the input 3D mesh, located at the center of the respective 2D bounding box, and the part names are propagated throughout the entire mesh based on geodesic distances. Similarly, PartSLIP [32] also leverages GLIP [25] for 2D detection but employs a different voting scheme for the 2D bounding box integration over the 3D object, which utilizes the oversegmentation of the input point cloud. One key distinction between PartSLIP and SATR is that PartSLIP proposes to modify the GLIP model for domain adaptation to the rendered images and text prompts used in 3D segmentation. Specifically, the text prompt is given as a concatenation of part names in the PartSLIP pipeline, which is not a typical and natural description of the object. Therefore, it learns a category-specific constant offset vector for the language embedding feature in the given pairs of prompts and rendered images, while keeping the GLIP parameters frozen." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 130, + 404, + 481, + 440 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 404, + 481, + 440 + ], + "spans": [ + { + "bbox": [ + 130, + 404, + 481, + 440 + ], + "type": "text", + "content": "Our work builds upon the PartSLIP pipeline. In the following subsections, we describe the details of its major components: the voting scheme, and the finetuning process." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 131, + 459, + 230, + 472 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 131, + 459, + 230, + 472 + ], + "spans": [ + { + "bbox": [ + 131, + 459, + 230, + 472 + ], + "type": "text", + "content": "3.1 PartSLIP [32]" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 130, + 479, + 480, + 516 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 479, + 480, + 516 + ], + "spans": [ + { + "bbox": [ + 130, + 479, + 480, + 516 + ], + "type": "text", + "content": "The main technical components of PartSLIP are the 1) voting scheme based on super points and 2) adaptation to text prompts1. We describe the details of each component below." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 130, + 528, + 482, + 624 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 528, + 482, + 624 + ], + "spans": [ + { + "bbox": [ + 130, + 528, + 482, + 624 + ], + "type": "text", + "content": "Voting for Super Points. PartSLIP first over-segments the input 3D object represented as a point cloud " + }, + { + "bbox": [ + 130, + 528, + 482, + 624 + ], + "type": "inline_equation", + "content": "\\mathcal{P}" + }, + { + "bbox": [ + 130, + 528, + 482, + 624 + ], + "type": "text", + "content": " into a set of super points " + }, + { + "bbox": [ + 130, + 528, + 482, + 624 + ], + "type": "inline_equation", + "content": "P_{i} \\subseteq \\mathcal{P}" + }, + { + "bbox": [ + 130, + 528, + 482, + 624 + ], + "type": "text", + "content": ", which provide geometric priors for the boundaries of the parts. The 3D object is rendered into " + }, + { + "bbox": [ + 130, + 528, + 482, + 624 + ], + "type": "inline_equation", + "content": "K" + }, + { + "bbox": [ + 130, + 528, + 482, + 624 + ], + "type": "text", + "content": " viewpoints, and the 2D bounding boxes for each image are predicted using GLIP [25]. Let " + }, + { + "bbox": [ + 130, + 528, + 482, + 624 + ], + "type": "inline_equation", + "content": "\\mathcal{B}" + }, + { + "bbox": [ + 130, + 528, + 482, + 624 + ], + "type": "text", + "content": " denote the entire set of 2D bounding boxes from all views, and " + }, + { + "bbox": [ + 130, + 528, + 482, + 624 + ], + "type": "inline_equation", + "content": "B_{jk} \\subseteq \\mathcal{B}" + }, + { + "bbox": [ + 130, + 528, + 482, + 624 + ], + "type": "text", + "content": " indicates a subset that includes the bounding boxes classified as the " + }, + { + "bbox": [ + 130, + 528, + 482, + 624 + ], + "type": "inline_equation", + "content": "j" + }, + { + "bbox": [ + 130, + 528, + 482, + 624 + ], + "type": "text", + "content": "-th part label from the " + }, + { + "bbox": [ + 130, + 528, + 482, + 624 + ], + "type": "inline_equation", + "content": "k" + }, + { + "bbox": [ + 130, + 528, + 482, + 624 + ], + "type": "text", + "content": "-th view. " + }, + { + "bbox": [ + 130, + 528, + 482, + 624 + ], + "type": "inline_equation", + "content": "\\mathrm{V}_k: \\mathcal{P} \\to \\{0,1\\}" + }, + { + "bbox": [ + 130, + 528, + 482, + 624 + ], + "type": "text", + "content": " is a function indicating whether the input point is visible from the " + }, + { + "bbox": [ + 130, + 528, + 482, + 624 + ], + "type": "inline_equation", + "content": "k" + }, + { + "bbox": [ + 130, + 528, + 482, + 624 + ], + "type": "text", + "content": "-th viewpoint. Also, for each 2D" + } + ] + } + ], + "index": 8 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 209, + 91, + 447, + 102 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 209, + 91, + 447, + 102 + ], + "spans": [ + { + "bbox": [ + 209, + 91, + 447, + 102 + ], + "type": "text", + "content": "PartSTAD: 2D-to-3D Part Segmentation Task Adaptation" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 474, + 92, + 480, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 474, + 92, + 480, + 100 + ], + "spans": [ + { + "bbox": [ + 474, + 92, + 480, + 100 + ], + "type": "text", + "content": "5" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 133, + 631, + 482, + 665 + ], + "type": "page_footnote", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 631, + 482, + 665 + ], + "spans": [ + { + "bbox": [ + 133, + 631, + 482, + 665 + ], + "type": "text", + "content": "1 Although multi-view feature aggregation is also introduced, the improvement by this component is marginal, and the author of PartSLIP also did not include this component in their official code. Therefore, we will omit this part in our review." + } + ] + } + ], + "index": 9 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 4 + }, + { + "para_blocks": [ + { + "bbox": [ + 130, + 116, + 482, + 175 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 116, + 482, + 175 + ], + "spans": [ + { + "bbox": [ + 130, + 116, + 482, + 175 + ], + "type": "text", + "content": "bounding box " + }, + { + "bbox": [ + 130, + 116, + 482, + 175 + ], + "type": "inline_equation", + "content": "b \\in \\mathcal{B}" + }, + { + "bbox": [ + 130, + 116, + 482, + 175 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 130, + 116, + 482, + 175 + ], + "type": "inline_equation", + "content": "\\mathrm{I}_b : \\mathcal{P} \\to \\{0,1\\}" + }, + { + "bbox": [ + 130, + 116, + 482, + 175 + ], + "type": "text", + "content": " is a function indicating whether the input point is included in the bounding box " + }, + { + "bbox": [ + 130, + 116, + 482, + 175 + ], + "type": "inline_equation", + "content": "b" + }, + { + "bbox": [ + 130, + 116, + 482, + 175 + ], + "type": "text", + "content": ". Given these, the voting aggregating the semantic labels of the 2D bounding boxes to the 3D point cloud is performed by calculating the following score " + }, + { + "bbox": [ + 130, + 116, + 482, + 175 + ], + "type": "inline_equation", + "content": "s_{ij}" + }, + { + "bbox": [ + 130, + 116, + 482, + 175 + ], + "type": "text", + "content": " for each pair of the " + }, + { + "bbox": [ + 130, + 116, + 482, + 175 + ], + "type": "inline_equation", + "content": "i" + }, + { + "bbox": [ + 130, + 116, + 482, + 175 + ], + "type": "text", + "content": "-th super point and the " + }, + { + "bbox": [ + 130, + 116, + 482, + 175 + ], + "type": "inline_equation", + "content": "j" + }, + { + "bbox": [ + 130, + 116, + 482, + 175 + ], + "type": "text", + "content": "-th label:" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 205, + 184, + 481, + 213 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 205, + 184, + 481, + 213 + ], + "spans": [ + { + "bbox": [ + 205, + 184, + 481, + 213 + ], + "type": "interline_equation", + "content": "s _ {i j} = \\frac {\\sum_ {k} \\sum_ {p \\in P _ {i}} \\mathrm {V} _ {k} (p) \\left(\\max _ {b \\in \\left(B _ {j k} \\cup \\{b _ {\\emptyset} \\}\\right)} \\mathrm {I} _ {b} (p)\\right)}{\\sum_ {k} \\sum_ {p \\in P _ {i}} \\mathrm {V} _ {k} (p)}, \\tag {1}", + "image_path": "9085fc4afa2b030f4a45032acc6b2b8f8f99ec7c1636d499baa58ab07b75cbb7.jpg" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 131, + 220, + 482, + 316 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 131, + 220, + 482, + 316 + ], + "spans": [ + { + "bbox": [ + 131, + 220, + 482, + 316 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 131, + 220, + 482, + 316 + ], + "type": "inline_equation", + "content": "b_{\\emptyset}" + }, + { + "bbox": [ + 131, + 220, + 482, + 316 + ], + "type": "text", + "content": " is a null bounding box, and we assume that " + }, + { + "bbox": [ + 131, + 220, + 482, + 316 + ], + "type": "inline_equation", + "content": "\\mathrm{I}_{b_{\\emptyset}}" + }, + { + "bbox": [ + 131, + 220, + 482, + 316 + ], + "type": "text", + "content": " for the null bounding box returns zero for any input point (simplifying notations for the case when " + }, + { + "bbox": [ + 131, + 220, + 482, + 316 + ], + "type": "inline_equation", + "content": "B_{jk} = \\emptyset" + }, + { + "bbox": [ + 131, + 220, + 482, + 316 + ], + "type": "text", + "content": "). Note that this " + }, + { + "bbox": [ + 131, + 220, + 482, + 316 + ], + "type": "inline_equation", + "content": "s_{ij} \\in [0,1]" + }, + { + "bbox": [ + 131, + 220, + 482, + 316 + ], + "type": "text", + "content": " denotes the ratio of visible points from each view included in any of the 2D bounding boxes labeled with the " + }, + { + "bbox": [ + 131, + 220, + 482, + 316 + ], + "type": "inline_equation", + "content": "j" + }, + { + "bbox": [ + 131, + 220, + 482, + 316 + ], + "type": "text", + "content": "-th label. Finally, for each " + }, + { + "bbox": [ + 131, + 220, + 482, + 316 + ], + "type": "inline_equation", + "content": "i" + }, + { + "bbox": [ + 131, + 220, + 482, + 316 + ], + "type": "text", + "content": "-th super point, the label with the highest score " + }, + { + "bbox": [ + 131, + 220, + 482, + 316 + ], + "type": "inline_equation", + "content": "s_{ij}" + }, + { + "bbox": [ + 131, + 220, + 482, + 316 + ], + "type": "text", + "content": " across all the labels is assigned to the super point. Exceptionally, if the highest score is less than a null label threshold " + }, + { + "bbox": [ + 131, + 220, + 482, + 316 + ], + "type": "inline_equation", + "content": "s_{\\emptyset}" + }, + { + "bbox": [ + 131, + 220, + 482, + 316 + ], + "type": "text", + "content": ", which is set to 0.5 in our experiments, the null label is assigned to the super point." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 130, + 326, + 482, + 445 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 326, + 482, + 445 + ], + "spans": [ + { + "bbox": [ + 130, + 326, + 482, + 445 + ], + "type": "text", + "content": "GLIP Finetuning. While GLIP demonstrates impressive generalizability for unseen images and prompts, it still has limitations when it comes to handling synthetic images and text prompts, such as the rendered images and sequences of part names used in the PartSLIP pipeline. To address this, the authors of PartSLIP propose adding small, learnable parameters to the GLIP model while keeping all the pretrained GLIP parameters frozen. These new parameters are learned with very few-shot examples: 8 annotated 3D objects per category. They represent offset feature vectors for each part name and remain constant as global variables for each category, rather than changing for each input. This adaptation of GLIP significantly improves the accuracy of 3D part segmentation." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 130, + 456, + 482, + 529 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 456, + 482, + 529 + ], + "spans": [ + { + "bbox": [ + 130, + 456, + 482, + 529 + ], + "type": "text", + "content": "Instance Segmentation. The voting scheme described above is designed for semantic segmentation, but PartSLIP also introduces a straightforward merging-based method to achieve instance segmentation based on semantic segmentation. To obtain instance segments, it merges adjacent super points under two conditions: 1) they share the same semantic label, and 2) they are either both included or both excluded for all bounding boxes." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 131, + 545, + 411, + 560 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 131, + 545, + 411, + 560 + ], + "spans": [ + { + "bbox": [ + 131, + 545, + 411, + 560 + ], + "type": "text", + "content": "4 PARTSTAD: Task Adaptation for 2D-to-3D" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 130, + 570, + 482, + 641 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 570, + 482, + 641 + ], + "spans": [ + { + "bbox": [ + 130, + 570, + 482, + 641 + ], + "type": "text", + "content": "PartSLIP has demonstrated that adapting a 2D detection model for 3D segmentation with a small set of trainable parameters can significantly enhance 3D segmentation performance. However, we observe that PartSLIP's adaptation is limited to domain adaptation, where it learns new parameters using synthetic images and data used in the PartSLIP framework without altering the objective function during training." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 130, + 641, + 482, + 665 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 641, + 482, + 665 + ], + "spans": [ + { + "bbox": [ + 130, + 641, + 482, + 665 + ], + "type": "text", + "content": "When applying the 2D segmentation model to 3D segmentation, it is important to note that not only does the data domain change, but the task itself" + } + ] + } + ], + "index": 9 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "spans": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "type": "text", + "content": "6" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 166, + 91, + 253, + 102 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 166, + 91, + 253, + 102 + ], + "spans": [ + { + "bbox": [ + 166, + 91, + 253, + 102 + ], + "type": "text", + "content": "H. Kim and M. Sung" + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 5 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 132, + 113, + 482, + 203 + ], + "blocks": [ + { + "bbox": [ + 132, + 113, + 482, + 203 + ], + "lines": [ + { + "bbox": [ + 132, + 113, + 482, + 203 + ], + "spans": [ + { + "bbox": [ + 132, + 113, + 482, + 203 + ], + "type": "image", + "image_path": "23e10151c3738afe3e36eff6d18c2f98fa62a94a5f38d19207d1137cf4c0bd08.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 130, + 218, + 482, + 328 + ], + "lines": [ + { + "bbox": [ + 130, + 218, + 482, + 328 + ], + "spans": [ + { + "bbox": [ + 130, + 218, + 482, + 328 + ], + "type": "text", + "content": "Figure 2: Overall pipeline of PARTSTAD. Our approach begins by rendering the provided 3D point cloud from multiple viewpoints. Subsequently, we extract 2D bounding boxes for its parts using GLIP [25] (Bounding Box Prediction); note that we utilize the finetuned GLIP model from PartSLIP [32]. Following this, we convert the bounding boxes into segmentation masks using SAM [20], extracting the foreground region for each bounding box (SAM Mask Integration). Next, we predict weights for all the masks and adaptively combine them into a 3D representation (2D-to-3D task adaptation). The final step involves obtaining the segmentation label for the input point cloud. The GLIP and SAM models are frozen, while only our novel weight prediction network is trained per category in a few-shot setting (8 objects)." + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_caption" + } + ], + "index": 2 + }, + { + "bbox": [ + 130, + 353, + 480, + 389 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 353, + 480, + 389 + ], + "spans": [ + { + "bbox": [ + 130, + 353, + 480, + 389 + ], + "type": "text", + "content": "changes. Therefore, through the process of incorporating 2D bounding boxes into the 3D object, the 2D bounding box prediction model must align with our ultimate goal: 3D part segmentation." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 130, + 389, + 482, + 510 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 389, + 482, + 510 + ], + "spans": [ + { + "bbox": [ + 130, + 389, + 482, + 510 + ], + "type": "text", + "content": "To address this, we introduce Part Segmentation Task ADaptation method, PA STAD, a task adaptation approach designed to finetune the 2D segmentation model for 3D part segmentation. In Sec. 4.1, we introduce our objective function to finetune the GLIP for the 3D segmentation task. In Sec. 4.2, we describe how we adapt the pretrained GLIP model with the new objective function by introducing additional learnable parameters and modifying the voting scheme (Eq. 1). PartSLIP also faces a performance limitation due to the GLIP that provides 2D bounding boxes, not 2D segments. In Sec. 4.3, we also propose to combine the SAM [20] foreground mask for each bounding box to obtain more precise 2D segments and, consequently, further improve the 3D segmentation." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 132, + 528, + 242, + 540 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 528, + 242, + 540 + ], + "spans": [ + { + "bbox": [ + 132, + 528, + 242, + 540 + ], + "type": "text", + "content": "4.1 3D mRIoU Loss" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 130, + 549, + 482, + 621 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 549, + 482, + 621 + ], + "spans": [ + { + "bbox": [ + 130, + 549, + 482, + 621 + ], + "type": "text", + "content": "We begin by describing the loss function that we use for adapting the GLIP to the 3D segmentation task. Let " + }, + { + "bbox": [ + 130, + 549, + 482, + 621 + ], + "type": "inline_equation", + "content": "\\mathbf{l}_j" + }, + { + "bbox": [ + 130, + 549, + 482, + 621 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 130, + 549, + 482, + 621 + ], + "type": "inline_equation", + "content": "\\hat{\\mathbf{l}}_j \\in \\{0,1\\}^{|\\mathcal{P}|}" + }, + { + "bbox": [ + 130, + 549, + 482, + 621 + ], + "type": "text", + "content": " be binary vectors indicating whether each point " + }, + { + "bbox": [ + 130, + 549, + 482, + 621 + ], + "type": "inline_equation", + "content": "p" + }, + { + "bbox": [ + 130, + 549, + 482, + 621 + ], + "type": "text", + "content": " in " + }, + { + "bbox": [ + 130, + 549, + 482, + 621 + ], + "type": "inline_equation", + "content": "\\mathcal{P}" + }, + { + "bbox": [ + 130, + 549, + 482, + 621 + ], + "type": "text", + "content": " has the " + }, + { + "bbox": [ + 130, + 549, + 482, + 621 + ], + "type": "inline_equation", + "content": "j" + }, + { + "bbox": [ + 130, + 549, + 482, + 621 + ], + "type": "text", + "content": "-th label in the ground truth and label prediction, respectively. The mean Intersection over Union (mIoU) for the predicted 3D segmentation, the standard evaluation metric for 3D segmentation, is defined as follows:" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 201, + 629, + 481, + 664 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 201, + 629, + 481, + 664 + ], + "spans": [ + { + "bbox": [ + 201, + 629, + 481, + 664 + ], + "type": "interline_equation", + "content": "\\operatorname {m I o U} \\left(\\left\\{\\mathbf {l} _ {j} \\right\\}, \\left\\{\\hat {\\mathbf {l}} _ {j} \\right\\}\\right) = \\frac {1}{M} \\sum_ {j = 1} ^ {M} \\frac {\\mathbf {l} _ {j} ^ {\\top} \\hat {\\mathbf {l}} _ {j}}{\\left\\| \\mathbf {l} _ {j} \\right\\| _ {1} + \\left\\| \\hat {\\mathbf {l}} _ {j} \\right\\| _ {1} - \\left. \\mathbf {l} _ {j} ^ {\\top} \\hat {\\mathbf {l}} _ {j} \\right.}, \\tag {2}", + "image_path": "2d0e666776b98c7a7b6c2cfbb21bff3d3a05be4b71ffcd43c6c84d7e49e34f6d.jpg" + } + ] + } + ], + "index": 8 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 210, + 91, + 447, + 102 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 210, + 91, + 447, + 102 + ], + "spans": [ + { + "bbox": [ + 210, + 91, + 447, + 102 + ], + "type": "text", + "content": "PartSTAD: 2D-to-3D Part Segmentation Task Adaptation" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 474, + 92, + 480, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 474, + 92, + 480, + 100 + ], + "spans": [ + { + "bbox": [ + 474, + 92, + 480, + 100 + ], + "type": "text", + "content": "7" + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 6 + }, + { + "para_blocks": [ + { + "bbox": [ + 132, + 116, + 279, + 127 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 116, + 279, + 127 + ], + "spans": [ + { + "bbox": [ + 132, + 116, + 279, + 127 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 132, + 116, + 279, + 127 + ], + "type": "inline_equation", + "content": "M" + }, + { + "bbox": [ + 132, + 116, + 279, + 127 + ], + "type": "text", + "content": " is the number of labels." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 130, + 129, + 482, + 189 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 129, + 482, + 189 + ], + "spans": [ + { + "bbox": [ + 130, + 129, + 482, + 189 + ], + "type": "text", + "content": "Our goal is to use the 3D mIoU as the objective function directly when adapting the GLIP model. However, the mIoU is non-differentiable. Thus, we employ a relaxed version known as mean Relaxed IoU (mRIOU) [24,65], which simply allows the predicted indicator vectors " + }, + { + "bbox": [ + 130, + 129, + 482, + 189 + ], + "type": "inline_equation", + "content": "\\hat{\\mathbf{l}}_j" + }, + { + "bbox": [ + 130, + 129, + 482, + 189 + ], + "type": "text", + "content": " to be non-binary: " + }, + { + "bbox": [ + 130, + 129, + 482, + 189 + ], + "type": "inline_equation", + "content": "\\hat{\\mathbf{l}}_j\\in [0,1]^{|\\mathcal{P}|}" + }, + { + "bbox": [ + 130, + 129, + 482, + 189 + ], + "type": "text", + "content": ". The loss " + }, + { + "bbox": [ + 130, + 129, + 482, + 189 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_{\\mathrm{mRIOU}}" + }, + { + "bbox": [ + 130, + 129, + 482, + 189 + ], + "type": "text", + "content": " is specifically defined as follows:" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 238, + 194, + 481, + 208 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 238, + 194, + 481, + 208 + ], + "spans": [ + { + "bbox": [ + 238, + 194, + 481, + 208 + ], + "type": "interline_equation", + "content": "\\mathcal {L} _ {\\mathrm {m R I o U}} = 1 - \\operatorname {m I o U} \\left(\\left\\{\\mathbf {l} _ {j} \\right\\}, \\left\\{\\hat {\\mathbf {l}} _ {j} \\right\\}\\right). \\tag {3}", + "image_path": "ca3dc9014799a3b928ca35cbca76b68d9c8a289ecc4dd5d6ca41a101e0406e9b.jpg" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 130, + 212, + 480, + 248 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 212, + 480, + 248 + ], + "spans": [ + { + "bbox": [ + 130, + 212, + 480, + 248 + ], + "type": "text", + "content": "In the supplementary material, we demonstrate that the choice of the loss function, as compared to other alternatives such as the cross-entropy loss, is crucial to achieving a substantial improvement in our task adaptation." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 130, + 248, + 480, + 299 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 248, + 480, + 299 + ], + "spans": [ + { + "bbox": [ + 130, + 248, + 480, + 299 + ], + "type": "text", + "content": "Note that in the PartSLIP framework, where super points are utilized in 3D segmentation, and " + }, + { + "bbox": [ + 130, + 248, + 480, + 299 + ], + "type": "inline_equation", + "content": "s_{ij} \\in [0,1]" + }, + { + "bbox": [ + 130, + 248, + 480, + 299 + ], + "type": "text", + "content": " indicates the likelihood of the " + }, + { + "bbox": [ + 130, + 248, + 480, + 299 + ], + "type": "inline_equation", + "content": "i" + }, + { + "bbox": [ + 130, + 248, + 480, + 299 + ], + "type": "text", + "content": "-th super point belonging to the " + }, + { + "bbox": [ + 130, + 248, + 480, + 299 + ], + "type": "inline_equation", + "content": "j" + }, + { + "bbox": [ + 130, + 248, + 480, + 299 + ], + "type": "text", + "content": "-label, the mRIoU can be calculated by defining " + }, + { + "bbox": [ + 130, + 248, + 480, + 299 + ], + "type": "inline_equation", + "content": "\\hat{\\mathbf{l}}_j" + }, + { + "bbox": [ + 130, + 248, + 480, + 299 + ], + "type": "text", + "content": " based on " + }, + { + "bbox": [ + 130, + 248, + 480, + 299 + ], + "type": "inline_equation", + "content": "\\mathbf{s}_j = [s_{1j}, s_{2j}, \\dots]^{\\top}" + }, + { + "bbox": [ + 130, + 248, + 480, + 299 + ], + "type": "text", + "content": " as follows:" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 284, + 313, + 481, + 328 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 284, + 313, + 481, + 328 + ], + "spans": [ + { + "bbox": [ + 284, + 313, + 481, + 328 + ], + "type": "interline_equation", + "content": "\\hat {\\mathbf {l}} _ {j} = \\mathbf {M} \\mathbf {s} _ {j}, \\tag {4}", + "image_path": "338f1f0aad3e896a0ced9d5deb612963e036ff2d0151b8c4cb1efec427e85ab9.jpg" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 130, + 333, + 480, + 358 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 333, + 480, + 358 + ], + "spans": [ + { + "bbox": [ + 130, + 333, + 480, + 358 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 130, + 333, + 480, + 358 + ], + "type": "inline_equation", + "content": "\\mathbf{M} \\in \\{0,1\\}^{|\\mathcal{P}| \\times |\\{P_i\\}|}" + }, + { + "bbox": [ + 130, + 333, + 480, + 358 + ], + "type": "text", + "content": " is a binary matrix that describes the memberships from each point to the super points." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 132, + 372, + 328, + 384 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 372, + 328, + 384 + ], + "spans": [ + { + "bbox": [ + 132, + 372, + 328, + 384 + ], + "type": "text", + "content": "4.2 Bounding Box Weight Prediction" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 130, + 388, + 482, + 508 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 388, + 482, + 508 + ], + "spans": [ + { + "bbox": [ + 130, + 388, + 482, + 508 + ], + "type": "text", + "content": "A typical approach for adapting a pretrained model to a new task would be finetuning the model using an objective function tailored to that new task. However, the finetuning method generally fails to yield meaningful improvements, especially when the goal is to achieve few-shot adaptation (e.g., with only 8 annotated 3D objects per class, as done in our case), while the pretrained model has been trained on an extensive dataset. To address this challenge, recent works like LoRA [16] and PartSLIP [32] have introduced the concept of adding small learnable parameters while keeping the existing pretrained model parameters frozen. This approach not only makes training very efficient but also enables the model to generalize effectively to unseen data." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 130, + 509, + 482, + 568 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 509, + 482, + 568 + ], + "spans": [ + { + "bbox": [ + 130, + 509, + 482, + 568 + ], + "type": "text", + "content": "While we also follow this approach, our specific challenge arises from the fact that when we use the 3D mRIoU (Eq. 3) as the objective function for 3D segmentation, it becomes non-differentiable with respect to the 2D bounding box location, which is the output of GLIP. This non-differentiability arises from the computation of the score " + }, + { + "bbox": [ + 130, + 509, + 482, + 568 + ], + "type": "inline_equation", + "content": "s_{ij}" + }, + { + "bbox": [ + 130, + 509, + 482, + 568 + ], + "type": "text", + "content": " (Eq. 1) for each pair of super point and label." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 130, + 568, + 482, + 629 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 568, + 482, + 629 + ], + "spans": [ + { + "bbox": [ + 130, + 568, + 482, + 629 + ], + "type": "text", + "content": "To address this, we propose training a small network that does not refine the location but instead predicts a weight for each 2D bounding box, taking the learned features of the boxes as input. If we denote the output weight of the network for bounding box " + }, + { + "bbox": [ + 130, + 568, + 482, + 629 + ], + "type": "inline_equation", + "content": "b" + }, + { + "bbox": [ + 130, + 568, + 482, + 629 + ], + "type": "text", + "content": " as " + }, + { + "bbox": [ + 130, + 568, + 482, + 629 + ], + "type": "inline_equation", + "content": "\\mathrm{W}(b) \\in \\mathbb{R}^{+}" + }, + { + "bbox": [ + 130, + 568, + 482, + 629 + ], + "type": "text", + "content": ", the score " + }, + { + "bbox": [ + 130, + 568, + 482, + 629 + ], + "type": "inline_equation", + "content": "s_{ij}" + }, + { + "bbox": [ + 130, + 568, + 482, + 629 + ], + "type": "text", + "content": " for a pair of super point and label is modified to " + }, + { + "bbox": [ + 130, + 568, + 482, + 629 + ], + "type": "inline_equation", + "content": "\\tilde{s}_{ij}" + }, + { + "bbox": [ + 130, + 568, + 482, + 629 + ], + "type": "text", + "content": " as follows:" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 194, + 634, + 481, + 664 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 194, + 634, + 481, + 664 + ], + "spans": [ + { + "bbox": [ + 194, + 634, + 481, + 664 + ], + "type": "interline_equation", + "content": "\\tilde {s} _ {i j} = \\frac {\\sum_ {k} \\sum_ {p \\in P _ {i}} \\mathrm {V} _ {k} (p) \\left(\\max _ {b \\in \\left(B _ {j k} \\cup \\{b _ {\\emptyset} \\}\\right)} \\mathrm {I} _ {b} (p) \\mathrm {W} (b)\\right)}{\\sum_ {k} \\sum_ {p \\in P _ {i}} \\mathrm {V} _ {k} (p)}. \\tag {5}", + "image_path": "bf02ff4e709973fb2e69445c5c167122caa52ccafe4d21d3a58b6a63fa9f3b33.jpg" + } + ] + } + ], + "index": 13 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "spans": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "type": "text", + "content": "8" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 166, + 91, + 253, + 102 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 166, + 91, + 253, + 102 + ], + "spans": [ + { + "bbox": [ + 166, + 91, + 253, + 102 + ], + "type": "text", + "content": "H. Kim and M. Sung" + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 7 + }, + { + "para_blocks": [ + { + "bbox": [ + 130, + 116, + 482, + 175 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 116, + 482, + 175 + ], + "spans": [ + { + "bbox": [ + 130, + 116, + 482, + 175 + ], + "type": "text", + "content": "Note that the change is simply to multiply the weight " + }, + { + "bbox": [ + 130, + 116, + 482, + 175 + ], + "type": "inline_equation", + "content": "\\mathrm{W}(b)" + }, + { + "bbox": [ + 130, + 116, + 482, + 175 + ], + "type": "text", + "content": " by " + }, + { + "bbox": [ + 130, + 116, + 482, + 175 + ], + "type": "inline_equation", + "content": "\\mathrm{I}_b(p)" + }, + { + "bbox": [ + 130, + 116, + 482, + 175 + ], + "type": "text", + "content": ", while only " + }, + { + "bbox": [ + 130, + 116, + 482, + 175 + ], + "type": "inline_equation", + "content": "\\mathrm{I}_b(p)" + }, + { + "bbox": [ + 130, + 116, + 482, + 175 + ], + "type": "text", + "content": ", indicating whether the point " + }, + { + "bbox": [ + 130, + 116, + 482, + 175 + ], + "type": "inline_equation", + "content": "p" + }, + { + "bbox": [ + 130, + 116, + 482, + 175 + ], + "type": "text", + "content": " is included in bounding box " + }, + { + "bbox": [ + 130, + 116, + 482, + 175 + ], + "type": "inline_equation", + "content": "b" + }, + { + "bbox": [ + 130, + 116, + 482, + 175 + ], + "type": "text", + "content": ", has been used in Eq. 1. This modified score no longer falls within the range of [0,1]. The final score " + }, + { + "bbox": [ + 130, + 116, + 482, + 175 + ], + "type": "inline_equation", + "content": "\\bar{s}_{ij}" + }, + { + "bbox": [ + 130, + 116, + 482, + 175 + ], + "type": "text", + "content": " is thus defined by normalizing " + }, + { + "bbox": [ + 130, + 116, + 482, + 175 + ], + "type": "inline_equation", + "content": "\\tilde{s}_{ij}" + }, + { + "bbox": [ + 130, + 116, + 482, + 175 + ], + "type": "text", + "content": " using the softmax function over the set of labels:" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 266, + 182, + 481, + 209 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 266, + 182, + 481, + 209 + ], + "spans": [ + { + "bbox": [ + 266, + 182, + 481, + 209 + ], + "type": "interline_equation", + "content": "\\bar {s} _ {i j} = \\frac {\\exp (\\tilde {s} _ {i j})}{\\sum_ {j} \\exp (\\tilde {s} _ {i j})}. \\tag {6}", + "image_path": "4f8773f2a801fbda30739de85d4145891ced9f2bc0833aed2f701b8a35801842.jpg" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 130, + 215, + 482, + 275 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 215, + 482, + 275 + ], + "spans": [ + { + "bbox": [ + 130, + 215, + 482, + 275 + ], + "type": "text", + "content": "We additionally propose to make the unnormalized score for the null label " + }, + { + "bbox": [ + 130, + 215, + 482, + 275 + ], + "type": "inline_equation", + "content": "\\tilde{s}_{\\emptyset}" + }, + { + "bbox": [ + 130, + 215, + 482, + 275 + ], + "type": "text", + "content": " learnable in our case, initialized with 10. We thus compute the softmax above while including " + }, + { + "bbox": [ + 130, + 215, + 482, + 275 + ], + "type": "inline_equation", + "content": "\\tilde{s}_{\\emptyset}" + }, + { + "bbox": [ + 130, + 215, + 482, + 275 + ], + "type": "text", + "content": ". The label of each super point is still chosen as the label giving the highest score " + }, + { + "bbox": [ + 130, + 215, + 482, + 275 + ], + "type": "inline_equation", + "content": "\\bar{s}_{ij}" + }, + { + "bbox": [ + 130, + 215, + 482, + 275 + ], + "type": "text", + "content": ", becoming null if the normalized null label score " + }, + { + "bbox": [ + 130, + 215, + 482, + 275 + ], + "type": "inline_equation", + "content": "\\bar{s}_{\\emptyset}" + }, + { + "bbox": [ + 130, + 215, + 482, + 275 + ], + "type": "text", + "content": " becomes the highest." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 130, + 285, + 482, + 381 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 285, + 482, + 381 + ], + "spans": [ + { + "bbox": [ + 130, + 285, + 482, + 381 + ], + "type": "text", + "content": "Network Architecture. We design the weight predictor network to take the bounding box feature vectors " + }, + { + "bbox": [ + 130, + 285, + 482, + 381 + ], + "type": "inline_equation", + "content": "\\mathbf{f}_b" + }, + { + "bbox": [ + 130, + 285, + 482, + 381 + ], + "type": "text", + "content": " from the pretrained GLIP model as input. The feature vectors of all bounding boxes across all views are fed to the network at once and processed with a small shared two-layer MLP. Context normalization [64] is added in the middle of the two layers to incorporate contextual information from the global set of boxes for each box. The output of the MLP for each bounding box is further processed with the following modified ReLU function " + }, + { + "bbox": [ + 130, + 285, + 482, + 381 + ], + "type": "inline_equation", + "content": "\\phi(\\cdot)" + }, + { + "bbox": [ + 130, + 285, + 482, + 381 + ], + "type": "text", + "content": ":" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 258, + 390, + 480, + 402 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 258, + 390, + 480, + 402 + ], + "spans": [ + { + "bbox": [ + 258, + 390, + 480, + 402 + ], + "type": "interline_equation", + "content": "\\phi (\\mathbf {x}) = \\max (\\tau + \\mathbf {x}, 0), \\tag {7}", + "image_path": "19fed8aa628180261316566de0d1ef10c81f990d0cbd02e56f0333d732035dbc.jpg" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 131, + 410, + 418, + 422 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 131, + 410, + 418, + 422 + ], + "spans": [ + { + "bbox": [ + 131, + 410, + 418, + 422 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 131, + 410, + 418, + 422 + ], + "type": "inline_equation", + "content": "\\tau" + }, + { + "bbox": [ + 131, + 410, + 418, + 422 + ], + "type": "text", + "content": " is a user-defined constant offset (10 in our experiments)." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 131, + 437, + 300, + 451 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 131, + 437, + 300, + 451 + ], + "spans": [ + { + "bbox": [ + 131, + 437, + 300, + 451 + ], + "type": "text", + "content": "4.3 SAM [20] Mask Integration" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 130, + 456, + 482, + 601 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 456, + 482, + 601 + ], + "spans": [ + { + "bbox": [ + 130, + 456, + 482, + 601 + ], + "type": "text", + "content": "Another limitation of PartSLIP is its reliance on GLIP [25] for 2D segmentation. Since GLIP is, in turn, a 2D object detection model, it produces bounding boxes instead of 2D segments, which cannot provide accurate boundaries of segments. We propose to address this issue by incorporating another pretrained 2D segmentation model, SAM [20]. SAM has the functionality of taking a bounding box as input and producing the foreground mask. Using this, we replace the set of 2D bounding boxes " + }, + { + "bbox": [ + 130, + 456, + 482, + 601 + ], + "type": "inline_equation", + "content": "\\mathcal{B}" + }, + { + "bbox": [ + 130, + 456, + 482, + 601 + ], + "type": "text", + "content": " with a set of 2D masks, while preserving the remaining steps in the framework. Note that the point-to-bounding-box membership " + }, + { + "bbox": [ + 130, + 456, + 482, + 601 + ], + "type": "inline_equation", + "content": "\\mathrm{I}_b" + }, + { + "bbox": [ + 130, + 456, + 482, + 601 + ], + "type": "text", + "content": " is changed to point-to-mask membership, and we still use the same bounding box features extracted from GLIP corresponding to each mask to train the weight prediction network. We demonstrate the effectiveness of applying the SAM mask in our experiments in Sec. 5." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 131, + 617, + 272, + 632 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 131, + 617, + 272, + 632 + ], + "spans": [ + { + "bbox": [ + 131, + 617, + 272, + 632 + ], + "type": "text", + "content": "5 Experiment Results" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 130, + 641, + 481, + 666 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 641, + 481, + 666 + ], + "spans": [ + { + "bbox": [ + 130, + 641, + 481, + 666 + ], + "type": "text", + "content": "In the experiments, we compare our method with the SotA 2D-prior-based zero-shot/few-shot 3D segmentation models using the PartNet-Mobility [56] dataset" + } + ] + } + ], + "index": 11 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 209, + 91, + 448, + 103 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 209, + 91, + 448, + 103 + ], + "spans": [ + { + "bbox": [ + 209, + 91, + 448, + 103 + ], + "type": "text", + "content": "PartSTAD: 2D-to-3D Part Segmentation Task Adaptation" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 474, + 92, + 480, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 474, + 92, + 480, + 100 + ], + "spans": [ + { + "bbox": [ + 474, + 92, + 480, + 100 + ], + "type": "text", + "content": "9" + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 8 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 138, + 114, + 481, + 262 + ], + "blocks": [ + { + "bbox": [ + 138, + 114, + 481, + 262 + ], + "lines": [ + { + "bbox": [ + 138, + 114, + 481, + 262 + ], + "spans": [ + { + "bbox": [ + 138, + 114, + 481, + 262 + ], + "type": "image", + "image_path": "60bec06afde2cda26bafb12f6d9cf23921e60a7fa0a6ec758435e00c0078f545.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 130, + 272, + 482, + 307 + ], + "lines": [ + { + "bbox": [ + 130, + 272, + 482, + 307 + ], + "spans": [ + { + "bbox": [ + 130, + 272, + 482, + 307 + ], + "type": "text", + "content": "Figure 3: Qualitative comparison of semantic segmentation results. Our PARTSTAD segments 3D parts more precisely with clearer boundaries, even for small (Camera, Chair) and thin (Clock) parts." + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_caption" + } + ], + "index": 2 + }, + { + "type": "table", + "bbox": [ + 132, + 381, + 482, + 479 + ], + "blocks": [ + { + "bbox": [ + 130, + 316, + 482, + 372 + ], + "lines": [ + { + "bbox": [ + 130, + 316, + 482, + 372 + ], + "spans": [ + { + "bbox": [ + 130, + 316, + 482, + 372 + ], + "type": "text", + "content": "Table 1: Quantitative comparison of the semantic segmentation results (mIoU(%) on the PartNet-Mobility dataset. Ours achieves a " + }, + { + "bbox": [ + 130, + 316, + 482, + 372 + ], + "type": "inline_equation", + "content": "7.0\\%" + }, + { + "bbox": [ + 130, + 316, + 482, + 372 + ], + "type": "text", + "content": "p improvement in average mIoU compared to PartSLIP [32], the SotA few-shot 3D segmentation method, while consistently increasing mIoU across all categories. Please refer to the supplementary material for the complete table with results for all 45 categories." + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 132, + 381, + 482, + 479 + ], + "lines": [ + { + "bbox": [ + 132, + 381, + 482, + 479 + ], + "spans": [ + { + "bbox": [ + 132, + 381, + 482, + 479 + ], + "type": "table", + "html": "
MethodmIoUStorage FurnitureTableChairSwitchToiletLaptopUSBRemoteScissors
SATR [2]29.320.623.333.121.417.611.230.217.236.8
SATR [2]+SP34.828.928.037.737.022.112.433.428.043.0
PartSLIP [32]58.052.344.682.852.150.431.252.136.661.4
Ablation Study
w/o Weight Pred.61.956.645.085.051.956.631.557.136.660.8
w/o SAM Integ.62.154.045.783.153.049.433.653.646.867.5
PARTSTAD (Ours)65.059.547.885.357.957.534.659.953.468.5
", + "image_path": "ccdc6a4fd0d3ef030f7013bc6bfa0a566f4b213d17f73c7e46a48fd2eae5af78.jpg" + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "table_body" + } + ], + "index": 5 + }, + { + "bbox": [ + 130, + 499, + 482, + 537 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 499, + 482, + 537 + ], + "spans": [ + { + "bbox": [ + 130, + 499, + 482, + 537 + ], + "type": "text", + "content": "and OmniObject3D [54] dataset. In the supplementary material, we present additional results on the ablation study, outcomes with scanned data, and comprehensive results covering all PartNet-Mobility categories." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 131, + 552, + 253, + 564 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 131, + 552, + 253, + 564 + ], + "spans": [ + { + "bbox": [ + 131, + 552, + 253, + 564 + ], + "type": "text", + "content": "5.1 Experiment Setup" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 130, + 569, + 482, + 666 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 569, + 482, + 666 + ], + "spans": [ + { + "bbox": [ + 130, + 569, + 482, + 666 + ], + "type": "text", + "content": "Dataset. We use the PartNet-Mobility [56] dataset, which is also used in the experiments of PartSLIP [32]. The PartNet-Mobility [56] dataset includes 45 object categories. For each category, the training split includes 8 shapes, while the test split has a range of the number of objects from 6 to 338, totaling 1,906 objects across all 45 categories. Following PartSLIP, 8 shapes from the test set of each category are designated as the validation set. To obtain the 2D bounding boxes for each object, we use 10 images from fixed viewpoints across all our experiments. Note that the PartNet-Mobility [56] dataset is identical to the subset" + } + ] + } + ], + "index": 8 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "spans": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "type": "text", + "content": "10" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 166, + 91, + 253, + 102 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 166, + 91, + 253, + 102 + ], + "spans": [ + { + "bbox": [ + 166, + 91, + 253, + 102 + ], + "type": "text", + "content": "H. Kim and M. Sung" + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 9 + }, + { + "para_blocks": [ + { + "bbox": [ + 130, + 115, + 482, + 154 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 115, + 482, + 154 + ], + "spans": [ + { + "bbox": [ + 130, + 115, + 482, + 154 + ], + "type": "text", + "content": "of PartNet-Ensembled dataset used to train and evaluate PartSLIP [32]. (The remaining part of PartNet-Ensembled was only used to train other supervised baseline models, not PartSLIP [32].)" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 130, + 164, + 481, + 201 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 164, + 481, + 201 + ], + "spans": [ + { + "bbox": [ + 130, + 164, + 481, + 201 + ], + "type": "text", + "content": "Evaluation Metrics. As evaluation metrics, we employ the mean Intersection over Union (mIoU) metric for semantic segmentation and mean Average Precision (mAP) for instance segmentation." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 130, + 201, + 481, + 248 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 201, + 481, + 248 + ], + "spans": [ + { + "bbox": [ + 130, + 201, + 481, + 248 + ], + "type": "text", + "content": "For semantic segmentation, the average mIoU over the entire dataset is computed by first calculating it per semantic part, averaging them across the parts in the same category, resulting in category-level mIoU, and then averaging the category-level mIoUs again across all the categories." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 130, + 248, + 481, + 344 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 248, + 481, + 344 + ], + "spans": [ + { + "bbox": [ + 130, + 248, + 481, + 344 + ], + "type": "text", + "content": "For instance segmentation, we compute two types of mAP: part-aware mAP and part-agnostic mAP. For part-aware mAP, similar to mIoU in semantic segmentation, we first calculate the mAP for each semantic part with instances having the part label. We then average them for each category, and then across categories. For part-agnostic mAP, we do not consider the semantic labels and directly compute the category-level mAP with part instances in the same object category, and then average them. All mAP values correspond to " + }, + { + "bbox": [ + 130, + 248, + 481, + 344 + ], + "type": "inline_equation", + "content": "\\mathrm{mAP}_{50}" + }, + { + "bbox": [ + 130, + 248, + 481, + 344 + ], + "type": "text", + "content": " with an IoU threshold of " + }, + { + "bbox": [ + 130, + 248, + 481, + 344 + ], + "type": "inline_equation", + "content": "50\\%" + }, + { + "bbox": [ + 130, + 248, + 481, + 344 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 131, + 356, + 493, + 381 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 131, + 356, + 493, + 381 + ], + "spans": [ + { + "bbox": [ + 131, + 356, + 493, + 381 + ], + "type": "text", + "content": "Baselines. We compare our method with three baselines: PartSLIP [32], SATR [2], and SAM3D [63]." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 138, + 390, + 481, + 666 + ], + "type": "list", + "angle": 0, + "index": 10, + "blocks": [ + { + "bbox": [ + 138, + 390, + 481, + 426 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 390, + 481, + 426 + ], + "spans": [ + { + "bbox": [ + 138, + 390, + 481, + 426 + ], + "type": "text", + "content": "- PartSLIP [32] is the method that we adopt as our base. Note that PartSLIP uses a finetuned GLIP model, and we also employ the same model in our framework to extract bounding box features for our task adaptation." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 138, + 426, + 481, + 628 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 426, + 481, + 628 + ], + "spans": [ + { + "bbox": [ + 138, + 426, + 481, + 628 + ], + "type": "text", + "content": "- SATR [2] is similar to PartSLIP, conducting 3D segmentation using 2D bounding boxes from GLIP [25]. However, SATR differs from PartSLIP in four key aspects: 1) it does not involve finetuning the GLIP model, 2) it does not utilize super points, 3) it is designed for mesh segmentation, not point cloud segmentation, and 4) it employs a distinct approach for integrating 2D bounding boxes into 3D (label propagation). To ensure a fair comparison between our method and SATR, we implement the following changes to SATR. First, since we employ the GLIP model finetuned by PartSLIP in our framework, we use the same finetuned model within the SATR framework. Second, we convert the point cloud input to a mesh, performing mesh segmentation with SATR, and apply the results to the input point cloud by finding the closest vertex for each point. Lastly, we introduce a version of SATR using super points, resulting in a comparison between our method and two versions of SATR: one without super points (denoted as " + }, + { + "bbox": [ + 138, + 426, + 481, + 628 + ], + "type": "inline_equation", + "content": "SATR" + }, + { + "bbox": [ + 138, + 426, + 481, + 628 + ], + "type": "text", + "content": ") and one with super points (denoted as " + }, + { + "bbox": [ + 138, + 426, + 481, + 628 + ], + "type": "inline_equation", + "content": "SATR + SP" + }, + { + "bbox": [ + 138, + 426, + 481, + 628 + ], + "type": "text", + "content": "). Since SATR does not perform instance segmentation, our comparison with it is focused on semantic segmentation." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 138, + 629, + 481, + 666 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 629, + 481, + 666 + ], + "spans": [ + { + "bbox": [ + 138, + 629, + 481, + 666 + ], + "type": "text", + "content": "- SAM3D [63] is another zero-shot 3D segmentation method using a 2D segmentation model, not GLIP [25], but SAM [20]. Since it only performs instance segmentation but does not label each segment, we compare our" + } + ] + } + ], + "index": 9 + } + ], + "sub_type": "text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 209, + 90, + 447, + 103 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 209, + 90, + 447, + 103 + ], + "spans": [ + { + "bbox": [ + 209, + 90, + 447, + 103 + ], + "type": "text", + "content": "PartSTAD: 2D-to-3D Part Segmentation Task Adaptation" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 470, + 91, + 479, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 470, + 91, + 479, + 100 + ], + "spans": [ + { + "bbox": [ + 470, + 91, + 479, + 100 + ], + "type": "text", + "content": "11" + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 10 + }, + { + "para_blocks": [ + { + "bbox": [ + 148, + 116, + 480, + 139 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 148, + 116, + 480, + 139 + ], + "spans": [ + { + "bbox": [ + 148, + 116, + 480, + 139 + ], + "type": "text", + "content": "method with it only for instance segmentation, using the part-agnostic mAP metric." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 130, + 151, + 482, + 247 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 151, + 482, + 247 + ], + "spans": [ + { + "bbox": [ + 130, + 151, + 482, + 247 + ], + "type": "text", + "content": "Ablation Study. We also conduct an ablation study, comparing our method with two major components being ablated: the weight prediction (Sec. 4.2) and the SAM mask integration (Sec. 4.3). In the supplementary material, we also compare our method with cases using cross-entropy loss instead of mR IoU loss (Sec. 4.1) and using GLIP bounding box confidence score instead of our learned weights in the weighted voting (Sec. 4.2). Note that if both weight prediction and SAM mask integration are excluded, our PARTSTAD is identical to PartSLIP [32]." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 131, + 263, + 320, + 276 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 131, + 263, + 320, + 276 + ], + "spans": [ + { + "bbox": [ + 131, + 263, + 320, + 276 + ], + "type": "text", + "content": "5.2 Semantic Segmentation Results" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 130, + 282, + 482, + 402 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 282, + 482, + 402 + ], + "spans": [ + { + "bbox": [ + 130, + 282, + 482, + 402 + ], + "type": "text", + "content": "Tab. 1 and Fig. 3 present the quantitative and qualitative comparisons of part semantic segmentation results, respectively. Please refer to the supplementary material for the complete table encompassing all 45 categories. In comparison to PartSLIP [32], a SotA few-shot 3D segmentation method, we achieve a significant " + }, + { + "bbox": [ + 130, + 282, + 482, + 402 + ], + "type": "inline_equation", + "content": "7.0\\%" + }, + { + "bbox": [ + 130, + 282, + 482, + 402 + ], + "type": "text", + "content": " p mIoU improvement over the entire set of 45 categories. Furthermore, for each specific category, we consistently demonstrate improvement, surpassing by more than " + }, + { + "bbox": [ + 130, + 282, + 482, + 402 + ], + "type": "inline_equation", + "content": "15\\%" + }, + { + "bbox": [ + 130, + 282, + 482, + 402 + ], + "type": "text", + "content": " p in some categories (e.g., Remote). Qualitatively, the results also highlight the effectiveness of our task adaptation and the SAM mask integration in segmenting parts more precisely, even for small (e.g., Camera, Chair) and thin (e.g., Clock) parts of objects." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 130, + 402, + 488, + 450 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 402, + 488, + 450 + ], + "spans": [ + { + "bbox": [ + 130, + 402, + 488, + 450 + ], + "type": "text", + "content": "SATR exhibits inferior performance, even when leveraging the finetuned GLIP model and super points (see SATR+SP results) due to the label-propagation-based 2D bounding box integration, which often fails to provide clear boundaries in 3D segmentation, as illustrated in Fig. 3." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 130, + 450, + 482, + 521 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 450, + 482, + 521 + ], + "spans": [ + { + "bbox": [ + 130, + 450, + 482, + 521 + ], + "type": "text", + "content": "The ablation study results demonstrate the influence of each major component in our framework. The quantitative findings indicate that the weight prediction network plays a crucial role in substantial improvement, leading to a " + }, + { + "bbox": [ + 130, + 450, + 482, + 521 + ], + "type": "inline_equation", + "content": "3.1\\%" + }, + { + "bbox": [ + 130, + 450, + 482, + 521 + ], + "type": "text", + "content": " decrease in average mIoU when it is ablated. The SAM mask integration also contributes significantly, resulting in a " + }, + { + "bbox": [ + 130, + 450, + 482, + 521 + ], + "type": "inline_equation", + "content": "2.9\\%" + }, + { + "bbox": [ + 130, + 450, + 482, + 521 + ], + "type": "text", + "content": " average mIoU decrease when ablated." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 131, + 539, + 316, + 551 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 131, + 539, + 316, + 551 + ], + "spans": [ + { + "bbox": [ + 131, + 539, + 316, + 551 + ], + "type": "text", + "content": "5.3 Instance Segmentation Results" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 130, + 558, + 482, + 665 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 558, + 482, + 665 + ], + "spans": [ + { + "bbox": [ + 130, + 558, + 482, + 665 + ], + "type": "text", + "content": "Following the instance segmentation idea of PartSLIP, we also perform instance segmentation using our method and compare the quantitative results with those of other methods in Tab. 2 and Tab. 3, reporting part-aware mAPs and part-agnostic mAPs, respectively. Fig. 4 also displays qualitative comparisons across different methods. Please refer to the supplementary material for the complete table of all 45 categories. Similar to semantic segmentation, ours achieves " + }, + { + "bbox": [ + 130, + 558, + 482, + 665 + ], + "type": "inline_equation", + "content": "4.0\\%" + }, + { + "bbox": [ + 130, + 558, + 482, + 665 + ], + "type": "text", + "content": "p and " + }, + { + "bbox": [ + 130, + 558, + 482, + 665 + ], + "type": "inline_equation", + "content": "5.2\\%" + }, + { + "bbox": [ + 130, + 558, + 482, + 665 + ], + "type": "text", + "content": "p improvements for both average part-aware mAPs and part-agnostic mAPs, respectively. Both the weight prediction network and SAM mask integration also exhibit meaningful improvements in performance, as their influence is" + } + ] + } + ], + "index": 9 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "spans": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "type": "text", + "content": "12" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 166, + 91, + 253, + 102 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 166, + 91, + 253, + 102 + ], + "spans": [ + { + "bbox": [ + 166, + 91, + 253, + 102 + ], + "type": "text", + "content": "H. Kim and M. Sung" + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 11 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 135, + 115, + 482, + 264 + ], + "blocks": [ + { + "bbox": [ + 135, + 115, + 482, + 264 + ], + "lines": [ + { + "bbox": [ + 135, + 115, + 482, + 264 + ], + "spans": [ + { + "bbox": [ + 135, + 115, + 482, + 264 + ], + "type": "image", + "image_path": "a6d16926218f69eb912161d12a84bfd50612da83e1ef4eb6d887164cad58f27f.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 132, + 274, + 482, + 308 + ], + "lines": [ + { + "bbox": [ + 132, + 274, + 482, + 308 + ], + "spans": [ + { + "bbox": [ + 132, + 274, + 482, + 308 + ], + "type": "text", + "content": "Figure 4: Qualitative comparison of instance segmentation results shows that our PARTSTAD successfully segments tiny 3D parts, such as the keys of keyboards and buttons of remote controls, with clear segment boundaries." + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_caption" + } + ], + "index": 2 + }, + { + "type": "table", + "bbox": [ + 132, + 376, + 481, + 458 + ], + "blocks": [ + { + "bbox": [ + 132, + 323, + 482, + 367 + ], + "lines": [ + { + "bbox": [ + 132, + 323, + 482, + 367 + ], + "spans": [ + { + "bbox": [ + 132, + 323, + 482, + 367 + ], + "type": "text", + "content": "Table 2: Part-aware mAPs(\\%) of the instance segmentation results on the PartNet-Mobility dataset. Ours achieves " + }, + { + "bbox": [ + 132, + 323, + 482, + 367 + ], + "type": "inline_equation", + "content": "4.0\\%" + }, + { + "bbox": [ + 132, + 323, + 482, + 367 + ], + "type": "text", + "content": "p improvement in mean part-aware mAPs compared to PartSLIP [32], with consistent enhancements across all categories. Please refer to the supplementary material for the complete table with results for all 45 categories." + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 132, + 376, + 481, + 458 + ], + "lines": [ + { + "bbox": [ + 132, + 376, + 481, + 458 + ], + "spans": [ + { + "bbox": [ + 132, + 376, + 481, + 458 + ], + "type": "table", + "html": "
MethodmAPStorage FurnitureTableChairSwitchToiletLaptopUSBRemoteScissors
PartSLIP [32]41.629.132.682.221.236.217.820.919.923.6
Ablation Study
w/o Weight Pred.44.733.830.882.522.240.824.417.420.325.5
w/o SAM Integ.44.229.432.382.219.736.023.624.829.623.1
PARTSTAD(Ours)45.635.533.282.522.140.628.926.533.726.5
", + "image_path": "5463fd0e69b7d0dcb78c904cb04baf8a63b8f83bb548b508ab52841d7e0ed036.jpg" + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "table_body" + } + ], + "index": 5 + }, + { + "type": "table", + "bbox": [ + 132, + 537, + 482, + 626 + ], + "blocks": [ + { + "bbox": [ + 132, + 472, + 482, + 528 + ], + "lines": [ + { + "bbox": [ + 132, + 472, + 482, + 528 + ], + "spans": [ + { + "bbox": [ + 132, + 472, + 482, + 528 + ], + "type": "text", + "content": "Table 3: Part-agnostic mAPs(%) of the instance segmentation results on the PartNet-Mobility dataset. Ours demonstrates " + }, + { + "bbox": [ + 132, + 472, + 482, + 528 + ], + "type": "inline_equation", + "content": "5.2\\%" + }, + { + "bbox": [ + 132, + 472, + 482, + 528 + ], + "type": "text", + "content": "p improvement in mean part-agnostic mAPs compared to PartSLIP [32], surpassing SAM3D with a mean mAP that is more than three times larger. Please refer to the supplementary material for the complete table with results for all 45 categories. (*SAM3D is a zero-shot method.)" + } + ] + } + ], + "index": 6, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 132, + 537, + 482, + 626 + ], + "lines": [ + { + "bbox": [ + 132, + 537, + 482, + 626 + ], + "spans": [ + { + "bbox": [ + 132, + 537, + 482, + 626 + ], + "type": "table", + "html": "
MethodmAPStorage FurnitureTableChairSwitchToiletLaptopUSBRemoteScissors
SAM3D* [63]12.11.210.65.55.43.71.522.61.07.2
PartSLIP [32]38.929.728.780.721.235.419.520.519.927.0
Ablation Study
w/o Weight Pred.42.635.929.281.122.240.026.515.220.329.4
w/o SAM Integ.42.634.827.681.219.736.627.427.329.624.9
PARTSTAD (Ours)44.141.728.283.322.441.034.123.533.726.2
", + "image_path": "cc479a0ed7b244059a70793cad37e664ade4de084d10782d29ed4774a73587a2.jpg" + } + ] + } + ], + "index": 7, + "angle": 0, + "type": "table_body" + } + ], + "index": 7 + }, + { + "bbox": [ + 132, + 653, + 481, + 666 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 653, + 481, + 666 + ], + "spans": [ + { + "bbox": [ + 132, + 653, + 481, + 666 + ], + "type": "text", + "content": "depicted in the ablation study results. Compared to SAM3D [63], whose part-" + } + ] + } + ], + "index": 8 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 210, + 91, + 447, + 102 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 210, + 91, + 447, + 102 + ], + "spans": [ + { + "bbox": [ + 210, + 91, + 447, + 102 + ], + "type": "text", + "content": "PartSTAD: 2D-to-3D Part Segmentation Task Adaptation" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 470, + 92, + 480, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 470, + 92, + 480, + 100 + ], + "spans": [ + { + "bbox": [ + 470, + 92, + 480, + 100 + ], + "type": "text", + "content": "13" + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 12 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 140, + 115, + 482, + 210 + ], + "blocks": [ + { + "bbox": [ + 140, + 115, + 482, + 210 + ], + "lines": [ + { + "bbox": [ + 140, + 115, + 482, + 210 + ], + "spans": [ + { + "bbox": [ + 140, + 115, + 482, + 210 + ], + "type": "image", + "image_path": "6a310f47bdd29d93f1e3a2cfeea8e30a524f4d32022b4b55cd3941fbda53abfe.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 130, + 219, + 482, + 264 + ], + "lines": [ + { + "bbox": [ + 130, + 219, + 482, + 264 + ], + "spans": [ + { + "bbox": [ + 130, + 219, + 482, + 264 + ], + "type": "text", + "content": "Figure 5: Qualitative comparison of semantic segmentation results on OmniObect3D [54] dataset, a high quality real scanned 3D objects dataset. Our PART-STD predicts precise boundaries, even in the case that the object has an appearance significantly different from training data such as box category." + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_caption" + } + ], + "index": 2 + }, + { + "bbox": [ + 130, + 284, + 482, + 344 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 284, + 482, + 344 + ], + "spans": [ + { + "bbox": [ + 130, + 284, + 482, + 344 + ], + "type": "text", + "content": "agnostic mAPs are shown in Tab. 3, our method achieves much better performance with more than three times greater mAP. Qualitative results also illustrate the outstanding performance of our method, accurately identifying all the tiny instances of parts, such as keys on keyboards, buttons on remotes, and clear boundaries of these instances." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 131, + 358, + 395, + 371 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 131, + 358, + 395, + 371 + ], + "spans": [ + { + "bbox": [ + 131, + 358, + 395, + 371 + ], + "type": "text", + "content": "5.4 Part Segmentation with Real Scanned Dataset" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 130, + 376, + 482, + 472 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 376, + 482, + 472 + ], + "spans": [ + { + "bbox": [ + 130, + 376, + 482, + 472 + ], + "type": "text", + "content": "Fig. 5 displays additional results on the OmniObject3D [54] dataset, a high-quality dataset featuring real-scanned 3D objects. We present qualitative results only, as ground truth segmentation for OmniObject3D [54] is unavailable. Note that our PARTSTAD accurately identifies parts in real scans, demonstrating its effectiveness beyond synthetic data. Particularly impressive is its performance in the box category depicted in Fig. 5, where despite significant discrepancies in appearance compared to the training data, PARTSTAD achieves remarkably accurate predictions." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 131, + 487, + 326, + 501 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 131, + 487, + 326, + 501 + ], + "spans": [ + { + "bbox": [ + 131, + 487, + 326, + 501 + ], + "type": "text", + "content": "6 Conclusion and Future Work" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 130, + 510, + 482, + 605 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 510, + 482, + 605 + ], + "spans": [ + { + "bbox": [ + 130, + 510, + 482, + 605 + ], + "type": "text", + "content": "We presented PARTSTAD, a task adaptation method for lifting 2D segmentation to 3D. Instead of directly finetuning a 2D segmentation network, our method learns a small neural network predicting weights for each 2D bounding box with an objective function for 3D segmentation, and then performs 3D segmentation via weighted merging of the boxes in 3D. We further improve the performance of 3D segmentation by integrating SAM foreground masks for each bounding box. We achieve the SotA results in few-shot 3D part segmentation, demonstrating significant improvements in both semantic and instance segmentations." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 130, + 605, + 482, + 666 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 605, + 482, + 666 + ], + "spans": [ + { + "bbox": [ + 130, + 605, + 482, + 666 + ], + "type": "text", + "content": "As PARTSTAD still depends on 2D representations for 3D predictions, its understanding of 3D geometry is relatively limited, and thus it cannot account for occluded or interior points, which poses a limitation. In future work, we plan to further investigate combining our approach with the direct utilization of 3D representations, such as leveraging 3D features, for more accurate predictions." + } + ] + } + ], + "index": 9 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "spans": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "type": "text", + "content": "14" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 166, + 91, + 253, + 102 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 166, + 91, + 253, + 102 + ], + "spans": [ + { + "bbox": [ + 166, + 91, + 253, + 102 + ], + "type": "text", + "content": "H. Kim and M. Sung" + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 13 + }, + { + "para_blocks": [ + { + "bbox": [ + 133, + 114, + 246, + 129 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 114, + 246, + 129 + ], + "spans": [ + { + "bbox": [ + 133, + 114, + 246, + 129 + ], + "type": "text", + "content": "Acknowledgements" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 130, + 142, + 497, + 214 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 142, + 497, + 214 + ], + "spans": [ + { + "bbox": [ + 130, + 142, + 497, + 214 + ], + "type": "text", + "content": "We sincerely thank Minghua Liu for providing the code of PartSLIP and answering questions. This work was supported by NRF grant (RS-2023-00209723), IITP grants (2022-0-00594, RS-2023-00227592, RS-2024-00399817), and Alchymist Project Program (RS-2024-00423625) funded by the Korean government (MSIT and MOTIE), and grants from the DRB-KAIST SketchTheFuture Research Center, Naver-intel, Adobe Research, Hyundai NGV, KT, and Samsung Electronics." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 132, + 235, + 197, + 248 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 235, + 197, + 248 + ], + "spans": [ + { + "bbox": [ + 132, + 235, + 197, + 248 + ], + "type": "text", + "content": "References" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 138, + 262, + 481, + 665 + ], + "type": "list", + "angle": 0, + "index": 20, + "blocks": [ + { + "bbox": [ + 138, + 262, + 481, + 285 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 262, + 481, + 285 + ], + "spans": [ + { + "bbox": [ + 138, + 262, + 481, + 285 + ], + "type": "text", + "content": "1. Abdelreheem, A., Eldesokey, A., Ovsjanikov, M., Wonka, P.: Zero-shot 3D shape correspondence. In: SIGGRAPH Asia (2023)" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 138, + 285, + 481, + 308 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 285, + 481, + 308 + ], + "spans": [ + { + "bbox": [ + 138, + 285, + 481, + 308 + ], + "type": "text", + "content": "2. Abdelreheem, A., Skorokhodov, I., Ovsjanikov, M., Wonka, P.: SATR: Zero-shot semantic segmentation of 3D shapes. In: ICCV (2023)" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 138, + 308, + 481, + 330 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 308, + 481, + 330 + ], + "spans": [ + { + "bbox": [ + 138, + 308, + 481, + 330 + ], + "type": "text", + "content": "3. Armeni, I., Sax, S., Zamir, A.R., Savarese, S.: Joint 2D-3D-semantic data for indoor scene understanding. arXiv preprint arXiv:1702.01105 (2017)" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 138, + 331, + 481, + 352 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 331, + 481, + 352 + ], + "spans": [ + { + "bbox": [ + 138, + 331, + 481, + 352 + ], + "type": "text", + "content": "4. Baranchuk, D., Voynov, A., Rubachev, I., Khrulkov, V., Babenko, A.: Label-efficient semantic segmentation with diffusion models. In: ICLR (2022)" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 138, + 353, + 481, + 386 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 353, + 481, + 386 + ], + "spans": [ + { + "bbox": [ + 138, + 353, + 481, + 386 + ], + "type": "text", + "content": "5. Behley, J., Garbade, M., Milioto, A., Quenzel, J., Behnke, S., Stachniss, C., Gall, J.: SemanticKITTI: A dataset for semantic scene understanding of LiDAR sequences. In: ICCV (2019)" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 138, + 387, + 481, + 408 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 387, + 481, + 408 + ], + "spans": [ + { + "bbox": [ + 138, + 387, + 481, + 408 + ], + "type": "text", + "content": "6. Cen, J., Zhou, Z., Fang, J., Yang, C., Shen, W., Xie, L., Zhang, X., Tian, Q.: Segment anything in 3D with nerfs. In: NeurIPS (2023)" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 138, + 409, + 481, + 441 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 409, + 481, + 441 + ], + "spans": [ + { + "bbox": [ + 138, + 409, + 481, + 441 + ], + "type": "text", + "content": "7. Chang, A., Dai, A., Funkhouser, T., Halber, M., Niessner, M., Savva, M., Song, S., Zeng, A., Zhang, Y.: Matterport3D: Learning from rgb-d data in indoor environments. In: 3DV (2017)" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 138, + 442, + 481, + 475 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 442, + 481, + 475 + ], + "spans": [ + { + "bbox": [ + 138, + 442, + 481, + 475 + ], + "type": "text", + "content": "8. Chang, A.X., Funkhouser, T., Guibas, L., Hanrahan, P., Huang, Q., Li, Z., Savarese, S., Savva, M., Song, S., Su, H., et al.: ShapeNet: An information-rich 3D model repository. arXiv preprint arXiv:1512.03012 (2015)" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 138, + 476, + 481, + 509 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 476, + 481, + 509 + ], + "spans": [ + { + "bbox": [ + 138, + 476, + 481, + 509 + ], + "type": "text", + "content": "9. Chen, R., Liu, Y., Kong, L., Zhu, X., Ma, Y., Li, Y., Hou, Y., Qiao, Y., Wang, W.: CLIP2Scene: Towards label-efficient 3D scene understanding by clip. In: CVPR (2023)" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 138, + 510, + 481, + 531 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 510, + 481, + 531 + ], + "spans": [ + { + "bbox": [ + 138, + 510, + 481, + 531 + ], + "type": "text", + "content": "0. Chibane, J., Engelmann, F., Tran, T.A., Pons-Moll, G.: Box2Mask: Weakly supervised 3D semantic instance segmentation using bounding boxes. In: ECCV (2022)" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 138, + 532, + 481, + 553 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 532, + 481, + 553 + ], + "spans": [ + { + "bbox": [ + 138, + 532, + 481, + 553 + ], + "type": "text", + "content": "1. Chu, R., Chen, Y., Kong, T., Qi, L., Li, L.: ICM-3D: Instantiated category modeling for 3D instance segmentation. IEEE Robotics and Automation Letters (2021)" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 138, + 554, + 481, + 576 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 554, + 481, + 576 + ], + "spans": [ + { + "bbox": [ + 138, + 554, + 481, + 576 + ], + "type": "text", + "content": "2. Dai, A., Chang, A.X., Savva, M., Halber, M., Funkhouser, T., Nießner, M.: ScanNet: Richly-annotated 3D reconstructions of indoor scenes. In: CVPR (2017)" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 138, + 577, + 481, + 598 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 577, + 481, + 598 + ], + "spans": [ + { + "bbox": [ + 138, + 577, + 481, + 598 + ], + "type": "text", + "content": "3. Decatur, D., Lang, I., Hanocka, R.: 3D Highlighter: Localizing regions on 3D shapes via text descriptions. In: CVPR (2023)" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 138, + 599, + 481, + 642 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 599, + 481, + 642 + ], + "spans": [ + { + "bbox": [ + 138, + 599, + 481, + 642 + ], + "type": "text", + "content": "4. Hackel, T., Savinov, N., Ladicky, L., Wegner, J.D., Schindler, K., Pollefeys, M.: SEMANTIC3D.NET: A new large-scale point cloud classification benchmark. In: ISPRS Annals of the Photogrammetry, Remote Sensing and Spatial Information Sciences (2017)" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 138, + 643, + 481, + 665 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 643, + 481, + 665 + ], + "spans": [ + { + "bbox": [ + 138, + 643, + 481, + 665 + ], + "type": "text", + "content": "5. Hou, J., Dai, A., Nießner, M.: 3D-SIS: 3D semantic instance segmentation of rgb-d scans. In: CVPR (2019)" + } + ] + } + ], + "index": 19 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 209, + 91, + 447, + 102 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 209, + 91, + 447, + 102 + ], + "spans": [ + { + "bbox": [ + 209, + 91, + 447, + 102 + ], + "type": "text", + "content": "PartSTAD: 2D-to-3D Part Segmentation Task Adaptation" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 470, + 92, + 481, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 470, + 92, + 481, + 100 + ], + "spans": [ + { + "bbox": [ + 470, + 92, + 481, + 100 + ], + "type": "text", + "content": "15" + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 14 + }, + { + "para_blocks": [ + { + "bbox": [ + 132, + 116, + 480, + 666 + ], + "type": "list", + "angle": 0, + "index": 23, + "blocks": [ + { + "bbox": [ + 133, + 116, + 480, + 149 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 116, + 480, + 149 + ], + "spans": [ + { + "bbox": [ + 133, + 116, + 480, + 149 + ], + "type": "text", + "content": "16. Hu, E.J., Shen, Y., Wallis, P., Allen-Zhu, Z., Li, Y., Wang, S., Wang, L., Chen, W.: LoRA: Low-rank adaptation of large language models. arXiv preprint arXiv:2106.09685 (2021)" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 133, + 149, + 480, + 171 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 149, + 480, + 171 + ], + "spans": [ + { + "bbox": [ + 133, + 149, + 480, + 171 + ], + "type": "text", + "content": "17. Huang, R., Huang, C., Liu, Y., Dai, G., Kong, W.: LSGCN: Long short-term traffic prediction with graph convolutional networks. In: IJCAI (2020)" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 133, + 171, + 480, + 192 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 171, + 480, + 192 + ], + "spans": [ + { + "bbox": [ + 133, + 171, + 480, + 192 + ], + "type": "text", + "content": "18. Huang, Z., Wu, X., Chen, X., Zhao, H., Zhu, L., Lasenby, J.: OpenIns3D: Snap and lookup for 3D open-vocabulary instance segmentation. arXiv preprint (2023)" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 133, + 192, + 480, + 213 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 192, + 480, + 213 + ], + "spans": [ + { + "bbox": [ + 133, + 192, + 480, + 213 + ], + "type": "text", + "content": "19. Jiang, L., Zhao, H., Shi, S., Liu, S., Fu, C.W., Jia, J.: PointGroup: Dual-set point grouping for 3D instance segmentation. In: CVPR (2020)" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 132, + 213, + 480, + 246 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 213, + 480, + 246 + ], + "spans": [ + { + "bbox": [ + 132, + 213, + 480, + 246 + ], + "type": "text", + "content": "20. Kirillov, A., Mintun, E., Ravi, N., Mao, H., Rolland, C., Gustafson, L., Xiao, T., Whitehead, S., Berg, A.C., Lo, W.Y., et al.: Segment anything. arXiv preprint arXiv:2304.02643 (2023)" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 132, + 246, + 480, + 267 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 246, + 480, + 267 + ], + "spans": [ + { + "bbox": [ + 132, + 246, + 480, + 267 + ], + "type": "text", + "content": "21. Koo, J., Huang, I., Achlioptas, P., Guibas, L.J., Sung, M.: PartGlot: Learning shape part segmentation from language reference games. In: CVPR (2022)" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 132, + 268, + 480, + 289 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 268, + 480, + 289 + ], + "spans": [ + { + "bbox": [ + 132, + 268, + 480, + 289 + ], + "type": "text", + "content": "22. Lai, X., Tian, Z., Chen, Y., Li, Y., Yuan, Y., Liu, S., Jia, J.: LISA: Reasoning segmentation via large language model. arXiv preprint arXiv:2308.00692 (2023)" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 132, + 289, + 480, + 320 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 289, + 480, + 320 + ], + "spans": [ + { + "bbox": [ + 132, + 289, + 480, + 320 + ], + "type": "text", + "content": "23. Lé, E.T., Sung, M., Ceylan, D., Mech, R., Boubekeur, T., Mitra, N.J.: CPFN: Cascaded primitive fitting networks for high-resolution point clouds. In: ICCV (2021)" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 132, + 320, + 480, + 342 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 320, + 480, + 342 + ], + "spans": [ + { + "bbox": [ + 132, + 320, + 480, + 342 + ], + "type": "text", + "content": "24. Li, L., Sung, M., Dubrovina, A., Yi, L., Guibas, L.J.: Supervised fitting of geometric primitives to 3D point clouds. In: CVPR (2019)" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 132, + 342, + 480, + 374 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 342, + 480, + 374 + ], + "spans": [ + { + "bbox": [ + 132, + 342, + 480, + 374 + ], + "type": "text", + "content": "25. Li, L.H., Zhang, P., Zhang, H., Yang, J., Li, C., Zhong, Y., Wang, L., Yuan, L., Zhang, L., Hwang, J.N., et al.: Grounded language-image pre-training. In: CVPR (2022)" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 132, + 374, + 480, + 407 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 374, + 480, + 407 + ], + "spans": [ + { + "bbox": [ + 132, + 374, + 480, + 407 + ], + "type": "text", + "content": "26. Li, M., Xie, Y., Shen, Y., Ke, B., Qiao, R., Ren, B., Lin, S., Ma, L.: Hybridcr: Weakly-supervised 3d point cloud semantic segmentation via hybrid contrastive regularization. In: CVPR (2022)" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 132, + 407, + 480, + 429 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 407, + 480, + 429 + ], + "spans": [ + { + "bbox": [ + 132, + 407, + 480, + 429 + ], + "type": "text", + "content": "27. Li, Y., Bu, R., Sun, M., Wu, W., Di, X., Chen, B.: PointCNN: Convolution on x-transformed points. In: NeurIPS (2018)" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 132, + 429, + 480, + 460 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 429, + 480, + 460 + ], + "spans": [ + { + "bbox": [ + 132, + 429, + 480, + 460 + ], + "type": "text", + "content": "28. Li, Y., Ma, L., Zhong, Z., Cao, D., Li, J.: TGNet: Geometric graph cnn on 3D point cloud segmentation. IEEE Transactions on Geoscience and Remote Sensing (2019)" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 132, + 460, + 480, + 493 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 460, + 480, + 493 + ], + "spans": [ + { + "bbox": [ + 132, + 460, + 480, + 493 + ], + "type": "text", + "content": "29. Liang, Z., Yang, M., Li, H., Wang, C.: 3D instance embedding learning with a structure-aware loss function for point cloud segmentation. IEEE Robotics and Automation Letters (2020)" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 132, + 493, + 480, + 515 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 493, + 480, + 515 + ], + "spans": [ + { + "bbox": [ + 132, + 493, + 480, + 515 + ], + "type": "text", + "content": "30. Lin, T.Y., Maire, M., Belongie, S., Hays, J., Perona, P., Ramanan, D., Dólar, P., Zitnick, C.L.: Microsoft COCO: Common objects in context. In: ECCV (2014)" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 132, + 515, + 480, + 536 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 515, + 480, + 536 + ], + "spans": [ + { + "bbox": [ + 132, + 515, + 480, + 536 + ], + "type": "text", + "content": "31. Liu, J., Yu, M., Ni, B., Chen, Y.: Self-prediction for joint instance and semantic segmentation of point clouds. In: ECCV (2020)" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 132, + 536, + 480, + 567 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 536, + 480, + 567 + ], + "spans": [ + { + "bbox": [ + 132, + 536, + 480, + 567 + ], + "type": "text", + "content": "32. Liu, M., Zhu, Y., Cai, H., Han, S., Ling, Z., Porikli, F., Su, H.: PartSLIP: Low-shot part segmentation for 3d point clouds via pretrained image-language models. In: CVPR (2023)" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 132, + 568, + 480, + 601 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 568, + 480, + 601 + ], + "spans": [ + { + "bbox": [ + 132, + 568, + 480, + 601 + ], + "type": "text", + "content": "33. Liu, S., Zeng, Z., Ren, T., Li, F., Zhang, H., Yang, J., Li, C., Yang, J., Su, H., Zhu, J., et al.: Grounding DINO: Marrying dino with grounded pre-training for open-set object detection. arXiv preprint arXiv:2303.05499 (2023)" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 132, + 601, + 480, + 622 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 601, + 480, + 622 + ], + "spans": [ + { + "bbox": [ + 132, + 601, + 480, + 622 + ], + "type": "text", + "content": "34. Liu, Y., Fan, B., Meng, G., Lu, J., Xiang, S., Pan, C.: DensePoint: Learning densely contextual representation for efficient point cloud processing. In: ICCV (2019)" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 132, + 622, + 480, + 643 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 622, + 480, + 643 + ], + "spans": [ + { + "bbox": [ + 132, + 622, + 480, + 643 + ], + "type": "text", + "content": "35. Liu, Z., Qi, X., Fu, C.W.: One thing one click: A self-training approach for weakly supervised 3D semantic segmentation. In: CVPR (2021)" + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 132, + 643, + 480, + 666 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 643, + 480, + 666 + ], + "spans": [ + { + "bbox": [ + 132, + 643, + 480, + 666 + ], + "type": "text", + "content": "36. Ma, X., Qin, C., You, H., Ran, H., Fu, Y.: Rethinking network design and local geometry in point cloud: A simple residual mlp framework. In: ICLR (2022)" + } + ] + } + ], + "index": 22 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "spans": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "type": "text", + "content": "16" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 166, + 91, + 253, + 102 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 166, + 91, + 253, + 102 + ], + "spans": [ + { + "bbox": [ + 166, + 91, + 253, + 102 + ], + "type": "text", + "content": "H. Kim and M. Sung" + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 15 + }, + { + "para_blocks": [ + { + "bbox": [ + 132, + 116, + 480, + 666 + ], + "type": "list", + "angle": 0, + "index": 23, + "blocks": [ + { + "bbox": [ + 132, + 116, + 480, + 149 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 116, + 480, + 149 + ], + "spans": [ + { + "bbox": [ + 132, + 116, + 480, + 149 + ], + "type": "text", + "content": "37. Mo, K., Zhu, S., Chang, A.X., Yi, L., Tripathi, S., Guibas, L.J., Su, H.: PartNet: A large-scale benchmark for fine-grained and hierarchical part-level 3D object understanding. In: CVPR (2019)" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 132, + 150, + 480, + 172 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 150, + 480, + 172 + ], + "spans": [ + { + "bbox": [ + 132, + 150, + 480, + 172 + ], + "type": "text", + "content": "38. Narita, G., Seno, T., Ishikawa, T., Kaji, Y.: Panoptic fusion: Online volumetric semantic mapping at the level of stuff and things. In: IROS (2019)" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 132, + 172, + 480, + 205 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 172, + 480, + 205 + ], + "spans": [ + { + "bbox": [ + 132, + 172, + 480, + 205 + ], + "type": "text", + "content": "39. Peng, S., Genova, K., Jiang, C., Tagliasacchi, A., Pollefeys, M., Funkhouser, T., et al.: OpenScene: 3D scene understanding with open vocabularies. In: CVPR (2023)" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 132, + 205, + 480, + 237 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 205, + 480, + 237 + ], + "spans": [ + { + "bbox": [ + 132, + 205, + 480, + 237 + ], + "type": "text", + "content": "40. Pham, Q.H., Nguyen, D.T., Hua, B.S., Roig, G., Yeung, S.K.: JSIS3D: Joint semantic-instance segmentation of 3D point clouds with multi-task pointwise networks and multi-value conditional random fields. In: CVPR (2019)" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 132, + 237, + 480, + 270 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 237, + 480, + 270 + ], + "spans": [ + { + "bbox": [ + 132, + 237, + 480, + 270 + ], + "type": "text", + "content": "41. Pruksachatkun, Y., Phang, J., Liu, H., Htut, P.M., Zhang, X., Pang, R.Y., Vania, C., Kann, K., Bowman, S.R.: Intermediate-Task Transfer Learning with Pretrained Language Models: When and why does it work? In: ACL (2020)" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 132, + 270, + 480, + 293 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 270, + 480, + 293 + ], + "spans": [ + { + "bbox": [ + 132, + 270, + 480, + 293 + ], + "type": "text", + "content": "42. Qi, C.R., Su, H., Mo, K., Guibas, L.J.: PointNet: Deep learning on point sets for 3d classification and segmentation. In: CVPR (2017)" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 132, + 293, + 480, + 315 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 293, + 480, + 315 + ], + "spans": [ + { + "bbox": [ + 132, + 293, + 480, + 315 + ], + "type": "text", + "content": "43. Qi, C.R., Yi, L., Su, H., Guibas, L.J.: PointNet++: Deep hierarchical feature learning on point sets in a metric space. In: NeurIPS (2017)" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 132, + 315, + 480, + 347 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 315, + 480, + 347 + ], + "spans": [ + { + "bbox": [ + 132, + 315, + 480, + 347 + ], + "type": "text", + "content": "44. Qian, G., Li, Y., Peng, H., Mai, J., Hammoud, H., Elhoseiny, M., Ghanem, B.: PointNext: Revisiting pointnet++ with improved training and scaling strategies. In: NeurIPS (2022)" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 132, + 347, + 480, + 380 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 347, + 480, + 380 + ], + "spans": [ + { + "bbox": [ + 132, + 347, + 480, + 380 + ], + "type": "text", + "content": "45. Radford, A., Kim, J.W., Hallacy, C., Ramesh, A., Goh, G., Agarwal, S., Sastry, G., Askell, A., Mishkin, P., Clark, J., et al.: Learning transferable visual models from natural language supervision. In: ICML (2021)" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 132, + 380, + 480, + 403 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 380, + 480, + 403 + ], + "spans": [ + { + "bbox": [ + 132, + 380, + 480, + 403 + ], + "type": "text", + "content": "46. Schult, J., Engelmann, F., Hermans, A., Litany, O., Tang, S., Leibe, B.: Mask3D: Mask transformer for 3D semantic instance segmentation. In: ICRA (2023)" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 132, + 403, + 480, + 424 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 403, + 480, + 424 + ], + "spans": [ + { + "bbox": [ + 132, + 403, + 480, + 424 + ], + "type": "text", + "content": "47. Takmaz, A., Fedele, E., Sumner, R.W., Pollefeys, M., Tombari, F., Engelmann, F.: OpenMask3D: Open-vocabulary 3D instance segmentation. In: NeurIPS (2023)" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 132, + 424, + 480, + 446 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 424, + 480, + 446 + ], + "spans": [ + { + "bbox": [ + 132, + 424, + 480, + 446 + ], + "type": "text", + "content": "48. Tang, L., Jia, M., Wang, Q., Phoo, C., Hariharan, B.: Emergent correspondence from image diffusion. In: NeurIPS (2023)" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 132, + 446, + 480, + 468 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 446, + 480, + 468 + ], + "spans": [ + { + "bbox": [ + 132, + 446, + 480, + 468 + ], + "type": "text", + "content": "49. Thomas, H., Qi, C.R., Deschaud, J.E., Marcotegui, B., Goulette, F., Guibas, L.J.: Kpconv: Flexible and deformable convolution for point clouds. In: ICCV (2019)" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 132, + 468, + 480, + 490 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 468, + 480, + 490 + ], + "spans": [ + { + "bbox": [ + 132, + 468, + 480, + 490 + ], + "type": "text", + "content": "50. Vu, T., Kim, K., Luu, T.M., Nguyen, X.T., Yoo, C.D.: Softgroup for 3d instance segmentation on 3D point clouds. In: CVPR (2022)" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 132, + 490, + 480, + 512 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 490, + 480, + 512 + ], + "spans": [ + { + "bbox": [ + 132, + 490, + 480, + 512 + ], + "type": "text", + "content": "51. Wang, W., Yu, R., Huang, Q., Neumann, U.: SGPN: Similarity group proposal network for 3D point cloud instance segmentation. In: CVPR (2018)" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 132, + 512, + 480, + 533 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 512, + 480, + 533 + ], + "spans": [ + { + "bbox": [ + 132, + 512, + 480, + 533 + ], + "type": "text", + "content": "52. Wang, X., Liu, S., Shen, X., Shen, C., Jia, J.: Associatively segmenting instances and semantics in point clouds. In: CVPR (2019)" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 132, + 533, + 480, + 555 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 533, + 480, + 555 + ], + "spans": [ + { + "bbox": [ + 132, + 533, + 480, + 555 + ], + "type": "text", + "content": "53. Wang, Y., Sun, Y., Liu, Z., Sarma, S.E., Bronstein, M.M., Solomon, J.M.: Dynamic graph cnn for learning on point clouds. ACM TOG (2019)" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 132, + 555, + 480, + 588 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 555, + 480, + 588 + ], + "spans": [ + { + "bbox": [ + 132, + 555, + 480, + 588 + ], + "type": "text", + "content": "54. Wu, T., Zhang, J., Fu, X., Wang, Y., Ren, J., Pan, L., Wu, W., Yang, L., Wang, J., Qian, C., Lin, D., Liu, Z.: OmniObject3D: Large-vocabulary 3D object dataset for realistic perception, reconstruction and generation. In: CVPR (2023)" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 132, + 588, + 480, + 611 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 588, + 480, + 611 + ], + "spans": [ + { + "bbox": [ + 132, + 588, + 480, + 611 + ], + "type": "text", + "content": "55. Wu, X., Lao, Y., Jiang, L., Liu, X., Zhao, H.: Point transformer v2: Grouped vector attention and partition-based pooling. In: NeurIPS (2022)" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 132, + 611, + 480, + 643 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 611, + 480, + 643 + ], + "spans": [ + { + "bbox": [ + 132, + 611, + 480, + 643 + ], + "type": "text", + "content": "56. Xiang, F., Qin, Y., Mo, K., Xia, Y., Zhu, H., Liu, F., Liu, M., Jiang, H., Yuan, Y., Wang, H., Yi, L., Chang, A.X., Guibas, L.J., Su, H.: SAPIEN: A simulated part-based interactive environment. In: CVPR (2020)" + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 132, + 643, + 480, + 666 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 643, + 480, + 666 + ], + "spans": [ + { + "bbox": [ + 132, + 643, + 480, + 666 + ], + "type": "text", + "content": "57. Xiang, T., Zhang, C., Song, Y., Yu, J., Cai, W.: Walk in the Cloud: Learning curves for point clouds shape analysis. In: ICCV (2021)" + } + ] + } + ], + "index": 22 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 210, + 91, + 447, + 102 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 210, + 91, + 447, + 102 + ], + "spans": [ + { + "bbox": [ + 210, + 91, + 447, + 102 + ], + "type": "text", + "content": "PartSTAD: 2D-to-3D Part Segmentation Task Adaptation" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 470, + 92, + 480, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 470, + 92, + 480, + 100 + ], + "spans": [ + { + "bbox": [ + 470, + 92, + 480, + 100 + ], + "type": "text", + "content": "17" + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 16 + }, + { + "para_blocks": [ + { + "bbox": [ + 132, + 116, + 480, + 511 + ], + "type": "list", + "angle": 0, + "index": 19, + "blocks": [ + { + "bbox": [ + 132, + 116, + 480, + 138 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 116, + 480, + 138 + ], + "spans": [ + { + "bbox": [ + 132, + 116, + 480, + 138 + ], + "type": "text", + "content": "58. Xu, J., Liu, S., Vahdat, A., Byeon, W., Wang, X., De Mello, S.: Open-vocabulary panoptic segmentation with text-to-image diffusion models. In: CVPR (2023)" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 132, + 140, + 480, + 160 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 140, + 480, + 160 + ], + "spans": [ + { + "bbox": [ + 132, + 140, + 480, + 160 + ], + "type": "text", + "content": "59. Xu, M., Ding, R., Zhao, H., Qi, X.: PACov: Position adaptive convolution with dynamic kernel assembling on point clouds. In: CVPR (2021)" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 132, + 161, + 480, + 182 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 161, + 480, + 182 + ], + "spans": [ + { + "bbox": [ + 132, + 161, + 480, + 182 + ], + "type": "text", + "content": "60. Yan, X., Zheng, C., Li, Z., Wang, S., Cui, S.: PointANSL: Robust point clouds processing using nonlocal neural networks with adaptive sampling. In: CVPR (2020)" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 132, + 183, + 480, + 215 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 183, + 480, + 215 + ], + "spans": [ + { + "bbox": [ + 132, + 183, + 480, + 215 + ], + "type": "text", + "content": "61. Yang, B., Wang, J., Clark, R., Hu, Q., Wang, S., Markham, A., Trigoni, N.: Learning object bounding boxes for 3D instance segmentation on point clouds. In: NeurIPS (2019)" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 132, + 216, + 473, + 226 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 216, + 473, + 226 + ], + "spans": [ + { + "bbox": [ + 132, + 216, + 473, + 226 + ], + "type": "text", + "content": "62. Yang, X., Wang, X.: Diffusion model as representation learner. In: ICCV (2023)" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 132, + 227, + 480, + 248 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 227, + 480, + 248 + ], + "spans": [ + { + "bbox": [ + 132, + 227, + 480, + 248 + ], + "type": "text", + "content": "63. Yang, Y., Wu, X., He, T., Zhao, H., Liu, X.: SAM3D: Segment anything in 3D scenes. arXiv preprint arXiv:2306.03908 (2023)" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 132, + 249, + 480, + 270 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 249, + 480, + 270 + ], + "spans": [ + { + "bbox": [ + 132, + 249, + 480, + 270 + ], + "type": "text", + "content": "64. Yi, K.M., Trulls, E., Ono, Y., Lepetit, V., Salzmann, M., Fua, P.: Learning to find good correspondences. In: CVPR (2018)" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 132, + 271, + 480, + 293 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 271, + 480, + 293 + ], + "spans": [ + { + "bbox": [ + 132, + 271, + 480, + 293 + ], + "type": "text", + "content": "65. Yi, L., Huang, H., Liu, D., Kalogerakis, E., Su, H., Guibas, L.: Deep part induction from articulated object pairs. ACM TOG (2018)" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 132, + 293, + 480, + 314 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 293, + 480, + 314 + ], + "spans": [ + { + "bbox": [ + 132, + 293, + 480, + 314 + ], + "type": "text", + "content": "66. Yi, L., Zhao, W., Wang, H., Sung, M., Guibas, L.J.: GSPN: Generative shape proposal network for 3D instance segmentation in point cloud. In: CVPR (2019)" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 132, + 315, + 480, + 336 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 315, + 480, + 336 + ], + "spans": [ + { + "bbox": [ + 132, + 315, + 480, + 336 + ], + "type": "text", + "content": "67. Yu, X., Tang, L., Rao, Y., Huang, T., Zhou, J., Lu, J.: Point-BERT: Pre-training 3D point cloud transformers with masked point modeling. In: CVPR (2022)" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 132, + 337, + 480, + 357 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 337, + 480, + 357 + ], + "spans": [ + { + "bbox": [ + 132, + 337, + 480, + 357 + ], + "type": "text", + "content": "68. Zamir, A.R., Sax, A., Shen, W.B., Guibas, L.J., Malik, J., Savarese, S.: Taskonomy: Disentangling task transfer learning. In: CVPR (2018)" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 132, + 358, + 480, + 379 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 358, + 480, + 379 + ], + "spans": [ + { + "bbox": [ + 132, + 358, + 480, + 379 + ], + "type": "text", + "content": "69. Zhang, B., Wonka, P.: Point cloud instance segmentation using probabilistic embeddings. In: CVPR (2021)" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 132, + 380, + 480, + 402 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 380, + 480, + 402 + ], + "spans": [ + { + "bbox": [ + 132, + 380, + 480, + 402 + ], + "type": "text", + "content": "70. Zhang, J., Dong, R., Ma, K.: CLIP-FO3D: Learning free open-world 3D scene representations from 2D dense clip. arXiv preprint arXiv:2303.04748 (2023)" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 132, + 403, + 480, + 434 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 403, + 480, + 434 + ], + "spans": [ + { + "bbox": [ + 132, + 403, + 480, + 434 + ], + "type": "text", + "content": "71. Zhang, R., Wang, L., Wang, Y., Gao, P., Li, H., Shi, J.: Parameter is not all you need: Starting from non-parametric networks for 3D point cloud analysis. In: CVPR (2023)" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 132, + 435, + 480, + 468 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 435, + 480, + 468 + ], + "spans": [ + { + "bbox": [ + 132, + 435, + 480, + 468 + ], + "type": "text", + "content": "72. Zhang, Y., Ling, H., Gao, J., Yin, K., Lafleche, J.F., Barriuso, A., Torralba, A., Fidler, S.: DatasetGAN: Efficient labeled data factory with minimal human effort. In: CVPR (2021)" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 132, + 468, + 480, + 490 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 468, + 480, + 490 + ], + "spans": [ + { + "bbox": [ + 132, + 468, + 480, + 490 + ], + "type": "text", + "content": "73. Zhao, H., Jiang, L., Jia, J., Torr, P.H., Koltun, V.: Point transformer. In: ICCV (2021)" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 132, + 491, + 480, + 511 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 491, + 480, + 511 + ], + "spans": [ + { + "bbox": [ + 132, + 491, + 480, + 511 + ], + "type": "text", + "content": "74. Zhao, N., Chua, T.S., Lee, G.H.: Few-shot 3D point cloud semantic segmentation. In: CVPR (2021)" + } + ] + } + ], + "index": 18 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "spans": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "type": "text", + "content": "18" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 166, + 91, + 253, + 102 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 166, + 91, + 253, + 102 + ], + "spans": [ + { + "bbox": [ + 166, + 91, + 253, + 102 + ], + "type": "text", + "content": "H. Kim and M. Sung" + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 17 + } + ], + "_backend": "vlm", + "_version_name": "2.6.4" +} \ No newline at end of file diff --git a/2024/PatchRefiner_ Leveraging Synthetic Data for Real-Domain High-Resolution Monocular Metric Depth Estimation/3be463cc-22b0-43d6-8e8c-f91c9fa34d5b_content_list.json b/2024/PatchRefiner_ Leveraging Synthetic Data for Real-Domain High-Resolution Monocular Metric Depth Estimation/3be463cc-22b0-43d6-8e8c-f91c9fa34d5b_content_list.json new file mode 100644 index 0000000000000000000000000000000000000000..38b256b304a055e1a2ebc42efd855ebc7570b632 --- /dev/null +++ b/2024/PatchRefiner_ Leveraging Synthetic Data for Real-Domain High-Resolution Monocular Metric Depth Estimation/3be463cc-22b0-43d6-8e8c-f91c9fa34d5b_content_list.json @@ -0,0 +1,2207 @@ +[ + { + "type": "text", + "text": "PatchRefiner: Leveraging Synthetic Data for Real-Domain High-Resolution Monocular Metric Depth Estimation", + "text_level": 1, + "bbox": [ + 222, + 140, + 779, + 209 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Zhenyu Li, Shariq Farooq Bhat, Peter Wonka", + "bbox": [ + 334, + 234, + 666, + 251 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "King Abdullah University of Science and Technology (KAUST) zhenyu.li.1@kaust.edu.sa", + "bbox": [ + 289, + 261, + 712, + 291 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Abstract. This paper introduces PatchRefiner, an advanced framework for metric single image depth estimation aimed at high-resolution real-domain inputs. While depth estimation is crucial for applications such as autonomous driving, 3D generative modeling, and 3D reconstruction, achieving accurate high-resolution depth in real-world scenarios is challenging due to the constraints of existing architectures and the scarcity of detailed real-world depth data. PatchRefiner adopts a tile-based methodology, reconceptualizing high-resolution depth estimation as a refinement process, which results in notable performance enhancements. Utilizing a pseudo-labeling strategy that leverages synthetic data, PatchRefiner incorporates a Detail and Scale Disentangling (DSD) loss to enhance detail capture while maintaining scale accuracy, thus facilitating the effective transfer of knowledge from synthetic to real-world data. Our extensive evaluations demonstrate PatchRefiner's superior performance, significantly outperforming existing benchmarks on the Unreal4KStereo dataset by $18.1\\%$ in terms of the root mean squared error (RMSE) and showing marked improvements in detail accuracy and consistent scale estimation on diverse real-world datasets like CityScape, ScanNet++, and ETH3D.", + "bbox": [ + 261, + 325, + 743, + 589 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Keywords: High-Resolution Metric Depth Estimation $\\cdot$ Synthetic Data", + "bbox": [ + 261, + 603, + 740, + 617 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "1 Introduction", + "text_level": 1, + "bbox": [ + 215, + 642, + 375, + 657 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "This paper delves into the field of metric single-image depth estimation, focusing on high-resolution inputs from the real domain. High-resolution depth estimation plays a key role in autonomous driving, augmented reality, content creation, and 3D reconstruction [2,11,30,59]. Despite significant progress, the high-resolution depth estimation in real-world scenarios remains daunting. This challenge is primarily due to the resolution limitations inherent in most state-of-the-art depth estimation architectures [2, 56, 61] and the scarcity of high-quality real-world depth datasets.", + "bbox": [ + 212, + 672, + 787, + 794 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "The current state-of-the-art in high-resolution depth estimation, PatchFusion [27, 34, 43], employs a tile-based strategy to navigate the resolution constraints, posing the task as a fusion process of coarse and fine depth estimations.", + "bbox": [ + 212, + 795, + 787, + 840 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/131911c50077e8a56732c678d77d0484938b30bd000a405316e52ed57133234a.jpg", + "image_caption": [ + "(a) LR Depth" + ], + "image_footnote": [], + "bbox": [ + 223, + 145, + 310, + 232 + ], + "page_idx": 1 + }, + { + "type": "image", + "img_path": "images/8cfde7296b386600ec18cc38cf4e70a1db57087fe160e66543be58f1c6f3df81.jpg", + "image_caption": [ + "(b) Fusion-Based Framework" + ], + "image_footnote": [], + "bbox": [ + 341, + 146, + 555, + 233 + ], + "page_idx": 1 + }, + { + "type": "image", + "img_path": "images/b48e13f85a3b73ccff89df2117da75a4ef32fdab0fc767f7398d3c4e06a4a624.jpg", + "image_caption": [ + "(c) Refiner-Based Framework", + "Fig. 1: Framework Comparison. (a) Low resolution depth estimation framework with single forward pass. (b) Fusion-based high-resolution framework combining the best of coarse and fine depth predictions [27, 34]. (c) Our refiner-based framework predicts a residual to refine the coarse prediction." + ], + "image_footnote": [], + "bbox": [ + 591, + 146, + 763, + 232 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Because of the scarcity of real high-resolution depth datasets, PatchFusion resorts to training on a synthetic 4K dataset [27, 51]. There are two limitations of PatchFusion that we would like to improve upon: 1) It utilizes a three-step training process, which is not only time-consuming and expensive but also risks the framework achieving the stage-wise local optima and limits the potential performance gains from end-to-end learning. 2) PatchFusion demonstrates poor generalization to the real domain.", + "bbox": [ + 212, + 349, + 787, + 455 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "The poor synthetic-to-real generalization is a long-standing problem in general [12,53]. In the context of depth estimation, the difference in the metric scale and depth distributions of synthetic datasets and the real domain further exacerbates the domain shift. This results in particularly subpar scale accuracy of depth models on real data when trained on synthetic datasets. On the other hand, the real datasets are not only low resolution but also often have missing ground truth values due to sensor constraints, occlusion, etc (see Fig. 3 and Sec. 3.4). This adds to the inability of depth models to capture sharp details when trained on real datasets. Thus, one is confronted with a dilemma - Training on real datasets leads to good scale accuracy but poor high-frequency details, while training on synthetic datasets leads to sharper results but poor scale performance on real images.", + "bbox": [ + 212, + 458, + 787, + 638 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "We introduce PatchRefiner, a novel framework that reformulates high-resolution depth estimation as a process of refining coarse depth. We propose improvements on two levels:", + "bbox": [ + 212, + 641, + 787, + 685 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "First, unlike direct depth regression approaches [27,39,43], PatchRefiner utilizes a frozen coarse depth model and enhances the quality by predicting residual depth for refinement. This approach not only streamlines model training but also markedly enhances performance.", + "bbox": [ + 212, + 686, + 787, + 747 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Second, we propose a method to exploit the best of both worlds to solve the above synthetic-real dilemma. We employ a teacher-student framework, leveraging the sharpness of synthetic data while learning the scale from real data. The teacher model, pre-trained on synthetic data, generates pseudo labels for real-domain training samples. Recognizing that these pseudo labels offer detailed features albeit with scale inaccuracies, we introduce the Detail and Scale Dis-", + "bbox": [ + 212, + 750, + 787, + 839 + ], + "page_idx": 1 + }, + { + "type": "page_number", + "text": "2", + "bbox": [ + 217, + 114, + 228, + 126 + ], + "page_idx": 1 + }, + { + "type": "header", + "text": "Zhenyu Li, Shariq Farooq Bhat, Peter Wonka", + "bbox": [ + 269, + 114, + 578, + 128 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "entangling (DSD) loss. This loss integrates ranking supervision and scale-shift invariance, drawing inspiration from recent advances in relative depth estimation [5, 42, 54, 56]. It leads to a framework capable of delivering high-resolution depth estimates with both precise scale and sharp details in real-world settings.", + "bbox": [ + 212, + 146, + 782, + 205 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Our evaluation of PatchRefiner on the Unreal4KStereo synthetic dataset [51] demonstrates a substantial improvement over the current state-of-the-art, reducing RMSE by $18.1\\%$ and REL by $15.7\\%$ . Further, we assess the framework's efficacy in leveraging synthetic data across diverse real-world datasets, including CityScape [9] (outdoor, stereo), ScanNet++ [58] (indoor, LiDAR and reconstruction), and ETH3D [46] (mixed, LiDAR). Our findings reveal notable enhancements in depth boundary delineation (e.g., a $19.2\\%$ increase in boundary recall on CityScape) while maintaining accurate scale estimation, showcasing the framework's adaptability and effectiveness across varying settings and sensor technologies.", + "bbox": [ + 212, + 207, + 784, + 357 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "2 Related Work", + "text_level": 1, + "bbox": [ + 215, + 383, + 387, + 398 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "High-Resolution Monocular Metric Depth Estimation. Monocular depth estimation has achieved tremendous progress [1, 11, 13, 15, 24, 29, 30]. Current SOTA approaches often employ complex network architectures yet grapple with the limitations imposed by their input resolution [2, 56]. This stands in stark contrast to the advancements in modern imaging devices that capture images at increasing resolutions, and the growing demand among users for high-resolution depth estimation. Initial efforts to address this gap included the use of Guided Depth Super-Resolution (GDSR) [19,33,63,65] and Implicit Functions [6,35]. Recently, Tile-Based Methods have emerged as a potent strategy for high-resolution depth estimation [27,34,43], segmenting images into patches for individual processing before reassembling them into a comprehensive depth map. This paper extends the tile-based approach, seeking to elevate the quality of depth estimation further.", + "bbox": [ + 212, + 416, + 784, + 611 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Synthetic Data for Depth Estimation. The challenge of acquiring high-quality, real-domain data for high-resolution depth training has prompted the use of synthetic datasets [41]. Traditionally, synthetic data has been employed within unsupervised domain adaptation frameworks, utilizing labeled synthetic and unlabeled real-domain data to enhance depth estimation accuracy on real-world images [7,22,23,32,60,64]. Techniques vary from pixel-level style transfer and image translation to feature-level adversarial learning [7,23,60,64], with some methods integrating additional information such as stereo pairs or segmentation maps for enhanced adaptation [22,32,60]. Contrasting with these approaches, our work explores a practical scenario where labeled data from both synthetic and real domains are leveraged to improve real-world, high-resolution depth estimation, delving into a relatively underexplored application of synthetic data.", + "bbox": [ + 212, + 613, + 785, + 792 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Pseudo-Labeling for Depth Estimation. Pseudo-labeling, a cornerstone of semi-supervised learning, has been widely applied across various domains, including classification [3,17,40,44,50] and scene understanding [8,26,28,36,37,47,62,", + "bbox": [ + 212, + 795, + 785, + 839 + ], + "page_idx": 2 + }, + { + "type": "header", + "text": "PatchRefiner", + "bbox": [ + 642, + 114, + 730, + 127 + ], + "page_idx": 2 + }, + { + "type": "page_number", + "text": "3", + "bbox": [ + 774, + 116, + 784, + 126 + ], + "page_idx": 2 + }, + { + "type": "image", + "img_path": "images/fc5f3c6def0d0be422667eafcc0ca334f82bc0e18f9c873309a33f77b5f3c887.jpg", + "image_caption": [ + "Fig. 2: Architecture Illustration. PatchRefiner contains a pre-trained frozen coarse depth estimation model $\\mathcal{N}_c$ and a refiner model $\\mathcal{N}_r$ predicts residual depth map $\\mathcal{D}_r$ to refine the coarse depth $\\mathcal{D}_c$ . The refiner contains one base depth model $\\mathcal{N}_d$ that has the same architecture as $\\mathcal{N}_c$ , and a light-weight decoder to aggregate information and make the final prediction." + ], + "image_footnote": [], + "bbox": [ + 248, + 146, + 754, + 294 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "66], to extrapolate knowledge from labeled data to unlabeled datasets. In depth estimation, pseudo-labeling often serves to provide supplementary supervision in unsupervised domain adaptation settings [32, 55, 57]. While recent state-of-the-art methods like Depth-Anything utilize pseudo-labeling to enhance model generalization [56], these techniques predominantly aim to refine or enhance the pseudo labels themselves. Our approach diverges by utilizing real-domain data with accurate depth labels, focusing on a novel Detail and Scale Disentangling loss. This loss mechanism uniquely leverages the detailed insights from pseudo labels to enrich real-domain depth estimation without compromising the scale accuracy derived from real ground-truth data.", + "bbox": [ + 212, + 405, + 789, + 556 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "3 Method", + "text_level": 1, + "bbox": [ + 215, + 579, + 330, + 595 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "In this section, we first present the overall PatchRefiner framework in Sec. 3.1. Then, we introduce the limitation of adopting real-domain data for high-resolution depth estimation 3.2 and the proposed teacher-student framework in Sec. 3.3 with the Detail and Scale Disentangling loss (DSD) in Sec. 3.4.", + "bbox": [ + 212, + 611, + 792, + 674 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "3.1 PatchRefiner Framework", + "text_level": 1, + "bbox": [ + 214, + 694, + 468, + 709 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "PatchRefiner follows the tile-based strategy to address the prohibitive memory and computational demands for high resolutions such as 4K [27, 34]. However, recognizing the limitations of existing models, we propose a simplified two-step approach for high-resolution depth estimation: (i) Coarse Scale-Aware Estimation, and (ii) Fine-Grained Depth Refinement, shown in Fig. 2.", + "bbox": [ + 212, + 719, + 787, + 794 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "(i) Coarse Scale-Aware Estimation: The foundation of PatchRefiner is the Coarse Depth Estimation network, $\\mathcal{N}_c$ , which processes downsampled versions of the input images to produce a global depth prediction, $\\mathbf{D}_c$ . Similar to", + "bbox": [ + 212, + 795, + 787, + 840 + ], + "page_idx": 3 + }, + { + "type": "page_number", + "text": "4", + "bbox": [ + 217, + 114, + 228, + 126 + ], + "page_idx": 3 + }, + { + "type": "header", + "text": "Zhenyu Li, Shariq Farooq Bhat, Peter Wonka", + "bbox": [ + 269, + 114, + 578, + 128 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "previous work [27, 34], this step is crucial for establishing a baseline depth map that captures the scene's overall structure and depth consistency, albeit without high-resolution details. Moreover, $\\mathcal{N}_c$ can be an arbitrary depth estimation model and it is frozen after the first step of training.", + "bbox": [ + 212, + 146, + 782, + 205 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "(ii) Depth Refinement Process: Different from the conventional approach of using a separate fine depth network and a fusion mechanism [27, 39], our framework introduces a unified refinement network, $\\mathcal{N}_r$ . This network is designed to refine the coarse depth map by focusing on the recovery of lost details and enhancing depth precision at a patch-wise level.", + "bbox": [ + 212, + 207, + 784, + 281 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "The input to $\\mathcal{N}_r$ is the cropped original image $I$ , which is processed by the base depth model $\\mathcal{N}_d$ with the same architecture as $\\mathcal{N}_c$ in the refiner module. Then, we collect $L$ -level multi-scale immediate features from both $\\mathcal{N}_d$ and $\\mathcal{N}_c$ , denoting them as $\\mathcal{F}_d = \\{f_d^i\\}_{i=1}^L$ and $\\tilde{\\mathcal{F}}_c = \\{\\tilde{f}_c^i\\}_{i=1}^L$ , respectively. Following [27], we apply the roi [16] operation to fetch features of the corresponding cropped area as $\\tilde{f}_c^i = \\text{roi}(f_c^i)$ .", + "bbox": [ + 212, + 282, + 784, + 373 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Next, we adopt a lightweight decoder to obtain high-resolution predictions. Given $\\mathcal{F}_d$ and $\\bar{\\mathcal{F}}_c$ , we aggregate them with concatenation operators (Cat) following by simple convolutional blocks (CB):", + "bbox": [ + 212, + 375, + 782, + 417 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\nf _ {r} ^ {i} = \\operatorname {C B} \\left(\\operatorname {C a t} \\left(f _ {d} ^ {i}, \\tilde {f} _ {c} ^ {i}\\right)\\right), \\tag {1}\n$$\n", + "text_format": "latex", + "bbox": [ + 426, + 433, + 784, + 450 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "The output feature set $\\mathcal{F}_r = \\{f_r^i\\}_{i=1}^L$ is then fed to a successive series of up-sampling layers [25], in order to construct the residual depth map $\\mathbf{D}_r$ at the input resolution. Associated with skip-connections, they form our refiner decoder. Further details about the architecture with immediate feature shapes are described in the supplementary material. Finally, we calculate the final patchwise depth estimation as $\\mathbf{D} = \\mathsf{roi}(\\mathbf{D}_c) + \\mathbf{D}_r$ .", + "bbox": [ + 212, + 457, + 784, + 547 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "3.2 Limitation of Real-Domain Depth Estimation Datasets", + "text_level": 1, + "bbox": [ + 214, + 571, + 712, + 585 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Our second goal in this paper is to train a real-domain high-resolution depth estimation model with both synthetic dataset $\\mathcal{S}$ and real-domain dataset $\\mathcal{R}$ . Our main insight is to distinguish between scale errors and boundary errors.", + "bbox": [ + 212, + 598, + 782, + 642 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "In the field of high-resolution depth estimation, the prevailing state-of-the-art methodology [27] trains models using synthetic datasets, which offers paired high-resolution images and corresponding dense high-resolution depth ground-truth maps [18,51]. This synthetic training regime, while beneficial in a controlled setting, introduces significant challenges when models are applied to real-world data due to the intrinsic domain gap between synthetic and real-world environments. This gap often manifests as substantial scale errors during inference on real-domain datasets as shown in Tab. 2.", + "bbox": [ + 212, + 643, + 784, + 762 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Addressing this issue directly by training on real-domain datasets presents its own set of challenges. Real-world high-resolution depth datasets are scarce and typically constrained by limited resolution and unavailable missing ground-truth pixels. These limitations stem from the methods employed in generating real-world depth annotations, such as Kinect [20, 48, 49], LiDAR [14, 46, 58], or", + "bbox": [ + 212, + 765, + 784, + 839 + ], + "page_idx": 4 + }, + { + "type": "header", + "text": "PatchRefiner", + "bbox": [ + 642, + 114, + 730, + 126 + ], + "page_idx": 4 + }, + { + "type": "page_number", + "text": "5", + "bbox": [ + 774, + 116, + 785, + 126 + ], + "page_idx": 4 + }, + { + "type": "image", + "img_path": "images/fe6917163c2e50005745bdc52f8546710cd5931788b69dc1a8829b4ae0d8c133.jpg", + "image_caption": [ + "Fig.3: Visualization of Real-Domain Data Pairs. Points lacking ground-truth data are depicted in gray. Due to sparse annotations near edges, models trained on real-domain data exhibit blurred boundary estimations." + ], + "image_footnote": [], + "bbox": [ + 218, + 143, + 504, + 344 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/87ecd62ae9318cd8e358b71389c009ef71e1a69a508d615974efba15c63abe2b.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 506, + 143, + 784, + 345 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "stereo vision techniques [9, 45], each with their inherent drawbacks as shown in Fig. 3.", + "bbox": [ + 212, + 429, + 782, + 458 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "For instance, depth maps generated using Kinect technology are confined to a resolution of $640 \\times 480$ [48, 49], insufficient for high-resolution depth estimation [27]. LiDAR-based depth maps, while useful, tend to be sparse [14, 46] or limited in resolution (e.g., 256x192 in Scannet++ [58]), and the process of reconstructing high-resolution dense depth maps from LiDAR point clouds is fragile with cascade errors and omissions [58]. Stereo vision techniques, on the other hand, can also introduce missing values around object boundaries due to rectification transformations [9, 45]. It is important to realize that existing high-resolution real datasets with missing values around edges do not help in reducing boundary errors.", + "bbox": [ + 212, + 460, + 787, + 611 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "These limitations highlight the inherent difficulties in employing real-world datasets for training high-resolution depth estimators [27, 39]. The lack of high-quality, high-resolution ground-truth depth maps in the real domain makes it challenging to train models that can accurately predict sharp depth around fine object boundaries [41]. Our main idea is to devise a training strategy that can improve the scale errors when fine-tuning on real data, while maintaining high-resolution information around edges to minimize boundary errors.", + "bbox": [ + 212, + 613, + 787, + 719 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "3.3 Overall Pipeline Illustration", + "text_level": 1, + "bbox": [ + 214, + 747, + 493, + 763 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Building on the success of semi-supervised learning frameworks in depth estimation tasks [32,56,57], our proposed framework employs a teacher-student architecture, as shown in Fig. 4, to effectively integrate synthetic and real-domain data for high-resolution depth estimation. It employs a two-step process:", + "bbox": [ + 212, + 779, + 785, + 840 + ], + "page_idx": 5 + }, + { + "type": "page_number", + "text": "6", + "bbox": [ + 217, + 114, + 228, + 126 + ], + "page_idx": 5 + }, + { + "type": "header", + "text": "Zhenyu Li, Shariq Farooq Bhat, Peter Wonka", + "bbox": [ + 271, + 114, + 578, + 128 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/238c474efd7684cb61068d3f08b54809b865fb85cfca8dae1fc1dc8e02c765ca.jpg", + "image_caption": [ + "Fig. 4: Enhancing Real-Domain Learning with Synthetic Data. A teacher model trained on synthetic data produces pseudo labels for real-domain training. The student model benefits from a DSD dual-supervision approach: loss on pseudo labels for detail enhancement and loss on ground truth for scale accuracy. This method ensures detailed depth perception without compromising scale accuracy." + ], + "image_footnote": [], + "bbox": [ + 235, + 145, + 769, + 262 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Teacher Model Training: The teacher model is initially pretrained on a synthetic dataset $S$ , which, due to its high-quality and detailed annotations, enables the model to predict high-resolution depth maps with precise boundaries.", + "bbox": [ + 212, + 375, + 782, + 421 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Student Model Training with both Pseudo Labels and Ground-Truth Depth: In the subsequent phase, the teacher model is frozen, and the student model is trained on the real-domain dataset $\\mathcal{R}$ , utilizing both the ground truth depth labels $\\tilde{\\mathbf{D}}$ and pseudo labels $\\hat{\\mathbf{D}}$ generated by the teacher model. As identified in the limitations of real-domain datasets, while the ground truth labels offer accurate depth information, they miss crucial information near boundaries essential for high-resolution depth learning. To address this, the student model is guided by the teacher's pseudo labels, which excel near boundaries. However, these pseudo labels, while sharp, exhibit scale discrepancies due to the domain gap between synthetic and real-world data as shown in Tab. 2.", + "bbox": [ + 212, + 422, + 784, + 571 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Therefore, we introduce the Detail and Scale Disentangling (DSD) loss. It ensures that the final high-resolution depth predictions from the student model maintain an accurate scale while benefiting from the enhanced boundary details provided by the teacher model trained on the synthetic data.", + "bbox": [ + 212, + 573, + 784, + 633 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "3.4 Detail and Scale Disentangling Loss", + "text_level": 1, + "bbox": [ + 214, + 659, + 555, + 674 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "The student model is subject to two sources of supervision. The primary one is the standard scale-invariant loss [2,11,27,30], $\\mathcal{L}_{\\text{silog}}$ , calculated with the ground truth (GT) data, which ensures reliable and scale-consistent depth estimation on the real-domain data.", + "bbox": [ + 212, + 688, + 782, + 747 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "The secondary source of supervision comes from pseudo labels generated by the teacher model. These labels, while detailed in capturing boundaries and fine details, exhibit low scale accuracy. Directly applying conventional depth losses [11,24,31] could lead to an imbalance between enhancing detail and maintaining scale accuracy. To address this, motivated by current impressive progress in relative depth estimation [5,42,54,56], we adopt the ranking loss $\\mathcal{L}_{\\text{rank}}$ [5,54]", + "bbox": [ + 212, + 750, + 784, + 840 + ], + "page_idx": 6 + }, + { + "type": "header", + "text": "PatchRefiner", + "bbox": [ + 642, + 114, + 730, + 126 + ], + "page_idx": 6 + }, + { + "type": "page_number", + "text": "7", + "bbox": [ + 774, + 114, + 784, + 125 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "and the Scale and Shift Invariant losses $\\mathcal{L}_{s s i}$ [42,56] to inject the detail information. Notably, The $\\mathcal{L}_{\\text {rank }}$ only provides supervision on the prediction relationship whereas $\\mathcal{L}_{s s i}$ ignores scale and shift. Both losses tackle the same challenge with a slightly different approach.", + "bbox": [ + 212, + 146, + 782, + 205 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Ranking loss [5, 54] is designed for sparse sets with ordinal data. This adaptation allows us to leverage dense detailed ranking information from the pseudo labels, enhancing high-resolution estimation without compromising scale accuracy. Specifically, for a certain pair of points $[p_{i,1}, p_{i,2}]$ with predicted depth values $[d_{i,1}, d_{i,2}]$ and pseudo depth values $[\\hat{d}_{i,1}, \\hat{d}_{i,2}]$ in a set of $N$ sampled point pairs $\\mathcal{P} = \\{[p_{i,1}, p_{i,2}], i = 1, 2, \\dots, N\\}$ , the ranking loss is defined as", + "bbox": [ + 212, + 207, + 784, + 299 + ], + "page_idx": 7 + }, + { + "type": "equation", + "text": "\n$$\n\\mathcal {L} _ {\\text {r a n k}} = \\frac {1}{N} \\sum_ {i} \\phi \\left(p _ {i, 1}, p _ {i, 2}\\right), \\tag {2}\n$$\n", + "text_format": "latex", + "bbox": [ + 406, + 311, + 784, + 344 + ], + "page_idx": 7 + }, + { + "type": "equation", + "text": "\n$$\n\\phi \\left(p _ {i, 1}, p _ {i, 2}\\right) = \\left\\{ \\begin{array}{l l} \\log \\left(1 + \\exp \\left(- \\ell \\times \\left(d _ {i, 1} - d _ {i, 2}\\right)\\right)\\right), & \\ell \\neq 0, \\\\ \\left(d _ {i, 1} - d _ {i, 2}\\right) ^ {2}, & \\ell = 0, \\end{array} \\right. \\tag {3}\n$$\n", + "text_format": "latex", + "bbox": [ + 303, + 353, + 782, + 393 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "where $\\ell$ is the pseudo ordinal label, which can be induced by:", + "bbox": [ + 214, + 398, + 656, + 412 + ], + "page_idx": 7 + }, + { + "type": "equation", + "text": "\n$$\n\\ell = \\left\\{ \\begin{array}{l l} + 1, & \\hat {d} _ {i, 1} / \\hat {d} _ {i, 2} > = 1 + \\tau , \\\\ - 1, & \\hat {d} _ {i, 1} / \\hat {d} _ {i, 2} < = \\frac {1}{1 + \\tau}, \\\\ 0, & \\text {o t h e r w i s e .} \\end{array} \\right. \\tag {4}\n$$\n", + "text_format": "latex", + "bbox": [ + 392, + 425, + 782, + 479 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Here $\\tau$ is a tolerance threshold, which is set to 0.03 in our experiments following [54]. This loss not only encourages the predicted depth values for closely related points to align but also emphasizes the differentiation of points when the pseudo depth value of the two points is different.", + "bbox": [ + 212, + 489, + 782, + 549 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Scale and Shift Invariant loss [42, 56] is proposed to learn relative depth estimation with ignoring the unknown scale and shift of each sample:", + "bbox": [ + 212, + 551, + 782, + 580 + ], + "page_idx": 7 + }, + { + "type": "equation", + "text": "\n$$\n\\mathcal {L} _ {s s i} = \\frac {1}{M} \\sum_ {i = 1} ^ {M} \\rho \\left(d _ {i} ^ {*} - \\hat {d} _ {i} ^ {*}\\right), \\tag {5}\n$$\n", + "text_format": "latex", + "bbox": [ + 411, + 593, + 782, + 632 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "where $d_i^*$ and $\\hat{d}_i^*$ are scaled and shifted versions of the predicted depth $d_i$ and pseudo label $\\hat{d}_i$ , and $\\rho$ is the mean absolute error loss. $M$ is the number of pixels. We use the least-squares criterion to align the prediction to the ground truth:", + "bbox": [ + 214, + 641, + 782, + 688 + ], + "page_idx": 7 + }, + { + "type": "equation", + "text": "\n$$\n(s, t) = \\operatorname {a r g m i n} _ {s, t} \\sum_ {i = 1} ^ {M} \\left(s d _ {i} + t - \\hat {d} _ {i}\\right) ^ {2}, \\tag {6}\n$$\n", + "text_format": "latex", + "bbox": [ + 375, + 700, + 782, + 739 + ], + "page_idx": 7 + }, + { + "type": "equation", + "text": "\n$$\nd ^ {*} = s d + t, \\hat {d} ^ {*} = \\hat {d} \\tag {7}\n$$\n", + "text_format": "latex", + "bbox": [ + 431, + 756, + 782, + 772 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "where the scale $s$ and shift $t$ factors are effectively determined with the closed form [42].", + "bbox": [ + 212, + 779, + 782, + 808 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "The final Detail and Scale Disentangling loss for training the student model combines the $\\mathcal{L}_{\\text{silog}}$ , $\\mathcal{L}_{\\text{rank}}$ , and $\\mathcal{L}_{\\text{ssi}}$ as follows:", + "bbox": [ + 212, + 809, + 782, + 840 + ], + "page_idx": 7 + }, + { + "type": "page_number", + "text": "8", + "bbox": [ + 217, + 114, + 228, + 126 + ], + "page_idx": 7 + }, + { + "type": "header", + "text": "Zhenyu Li, Shariq Farooq Bhat, Peter Wonka", + "bbox": [ + 271, + 114, + 575, + 128 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/8dab28b88d7563df30b6187fecfc8857fe6795683a82a2f92436616104e182c4.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 215, + 143, + 357, + 205 + ], + "page_idx": 8 + }, + { + "type": "image", + "img_path": "images/6c68fbedfa8a1c82b2a1a77f63b210f2f0a2b87089f50957cb3ab25ae05d0b1a.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 217, + 205, + 357, + 267 + ], + "page_idx": 8 + }, + { + "type": "image", + "img_path": "images/2f08459c366f5842c705c642757da8ba579c4758c4f284199217bf174f13f01e.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 217, + 268, + 357, + 330 + ], + "page_idx": 8 + }, + { + "type": "image", + "img_path": "images/c58e792f875c97aa3ada614d43537838c8498edb30d5d798e995884d3ae81ad4.jpg", + "image_caption": [ + "ZoeDepth [2]" + ], + "image_footnote": [], + "bbox": [ + 217, + 330, + 357, + 393 + ], + "page_idx": 8 + }, + { + "type": "image", + "img_path": "images/c4beefae1114f0d2774628194d7a91ebd4f07cb291c7cad5096deee59acb2f47.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 359, + 143, + 498, + 204 + ], + "page_idx": 8 + }, + { + "type": "image", + "img_path": "images/5c7b289c8e4930dd5ac194640954128a479160058341fa71256fb1d0f1a7adbe.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 359, + 205, + 500, + 268 + ], + "page_idx": 8 + }, + { + "type": "image", + "img_path": "images/bc860e2d7bd90318b0e40008d643c053271d41c6bacfe815438e4b85be9ea37d.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 359, + 268, + 500, + 330 + ], + "page_idx": 8 + }, + { + "type": "image", + "img_path": "images/88f4b67b583fa246a8775c1d39a11c928c5c40654a3fb401f9273a2b99b41149.jpg", + "image_caption": [ + "PatchFusion [27]" + ], + "image_footnote": [], + "bbox": [ + 359, + 330, + 500, + 393 + ], + "page_idx": 8 + }, + { + "type": "image", + "img_path": "images/e01f47a4253b0550142c5aee49050ef1092532c0b4d91a39f896709714fb5fa0.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 501, + 143, + 640, + 204 + ], + "page_idx": 8 + }, + { + "type": "image", + "img_path": "images/df734763aadd7f059378f04649a6248048b487a13c8e2865277e42ef36437d4a.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 501, + 205, + 640, + 268 + ], + "page_idx": 8 + }, + { + "type": "image", + "img_path": "images/24cc49880004440dee69805bbaff368ce5bf55ebf7d39c818f81658e7ee300d8.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 501, + 268, + 640, + 330 + ], + "page_idx": 8 + }, + { + "type": "image", + "img_path": "images/9255df48973122c1d1596726e4505f4de5bbffa868d5d0a5d1c6c056042c6ebb.jpg", + "image_caption": [ + "Ours", + "Fig. 5: Qualitative Comparison on UnrealStereo4K. We show the depth prediction and corresponding error map, respectively. The qualitative comparisons showcased here indicate our framework outperforms counterparts [2, 27] with sharper edges and lower error around boundaries. We show individual patches in all images to emphasize details near depth boundaries." + ], + "image_footnote": [], + "bbox": [ + 501, + 330, + 640, + 393 + ], + "page_idx": 8 + }, + { + "type": "image", + "img_path": "images/eb9bf0c1d9113e5f8e80a0e473b30eb83a45a3c45a23b9a9e5be5fe5faec0d19.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 642, + 143, + 784, + 204 + ], + "page_idx": 8 + }, + { + "type": "image", + "img_path": "images/fc54c1ccfc9b2dbb82684191c04d34522f7074e00f0d5f2e88e82aa030e3b234.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 642, + 205, + 784, + 268 + ], + "page_idx": 8 + }, + { + "type": "image", + "img_path": "images/aad4f5d5ccbd16761a3442768302a5b38dad7bf541a49195a72e78a232a2538c.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 642, + 268, + 784, + 330 + ], + "page_idx": 8 + }, + { + "type": "image", + "img_path": "images/53ca2cb852d8fe3b66c8d74328dba9e51b2b253b2d1a2d517b80900fc05222ff.jpg", + "image_caption": [ + "GT, Input" + ], + "image_footnote": [], + "bbox": [ + 642, + 330, + 784, + 393 + ], + "page_idx": 8 + }, + { + "type": "equation", + "text": "\n$$\n\\mathcal {L} _ {D S D} = \\overbrace {\\mathcal {L} _ {s i l o g}} ^ {\\mathcal {L} _ {g t}} + \\overbrace {\\lambda_ {1} \\mathcal {L} _ {r a n k} + \\lambda_ {2} \\mathcal {L} _ {s s i}} ^ {\\mathcal {L} _ {p l}}, \\tag {8}\n$$\n", + "text_format": "latex", + "bbox": [ + 372, + 535, + 784, + 571 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "where $\\lambda_{1}$ and $\\lambda_{2}$ are two balancing factors, both empirically set to 0.1 in our experiments, respectively. The terms $\\mathcal{L}_{gt}$ and $\\mathcal{L}_{pl}$ represent the supervision signals derived from the ground truth and pseudo labels, respectively.", + "bbox": [ + 214, + 577, + 787, + 623 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "4 Experiments", + "text_level": 1, + "bbox": [ + 215, + 646, + 375, + 662 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "4.1 Datasets and Metrics", + "text_level": 1, + "bbox": [ + 215, + 678, + 439, + 691 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "UnrealStereo4K (Synthetic): The UnrealStereo4K dataset [51] offers synthetic stereo images at a 4K resolution $(2160\\times 3840)$ , each paired with accurate, boundary-complete pixel-wise ground truth. Following the procedure in [27], we exclude mislabeled images using the Structural Similarity Index (SSIM) [52] and compute ground truth depth maps from provided disparity maps using camera parameters. Adhering to the dataset splits in [27, 51], we employ a default patch size of $540\\times 960$ for compatibility with [27].", + "bbox": [ + 212, + 703, + 787, + 809 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "CityScapes (Real, Stereo): Cityscapes [9] offers a comprehensive suite of urban scene images, segmentation masks, and disparity maps at a resolution", + "bbox": [ + 214, + 809, + 787, + 839 + ], + "page_idx": 8 + }, + { + "type": "header", + "text": "PatchRefiner", + "bbox": [ + 642, + 114, + 730, + 126 + ], + "page_idx": 8 + }, + { + "type": "page_number", + "text": "9", + "bbox": [ + 774, + 114, + 784, + 126 + ], + "page_idx": 8 + }, + { + "type": "image", + "img_path": "images/9a82ae2f5eeeb936bab130c8da7434df19693f3fefe365ba4597cee22ef16bd6.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 220, + 143, + 331, + 196 + ], + "page_idx": 9 + }, + { + "type": "image", + "img_path": "images/f121f3099f538b47ec7e9c78c35d53f79c13b3ea5969cc3debf791f97014636e.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 220, + 198, + 331, + 253 + ], + "page_idx": 9 + }, + { + "type": "image", + "img_path": "images/5989e62ce3a0c89bd616266826c0151a011da40d38c82361081ba6d95847c282.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 220, + 253, + 331, + 309 + ], + "page_idx": 9 + }, + { + "type": "image", + "img_path": "images/43494bd5ec155036c39632cc68277204c4e25ca6f9ebf30efd0ebeace4e4177c.jpg", + "image_caption": [ + "Image" + ], + "image_footnote": [], + "bbox": [ + 220, + 310, + 331, + 364 + ], + "page_idx": 9 + }, + { + "type": "image", + "img_path": "images/291c1f44d547b4fb8bba2ec6a94c0a880d9ba5b44709b5dee0bffa46cba658ec.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 333, + 143, + 442, + 198 + ], + "page_idx": 9 + }, + { + "type": "image", + "img_path": "images/7a4342bbaa38dd6ecf58b592662bf78ba9c000ca50d02d1bdc8b29398a8ea9f4.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 333, + 199, + 442, + 253 + ], + "page_idx": 9 + }, + { + "type": "image", + "img_path": "images/4642ed713d37ed4b63ce59ed15c6f41918acdd089a15b5f26f7e71a42ef6f9cc.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 333, + 253, + 442, + 309 + ], + "page_idx": 9 + }, + { + "type": "image", + "img_path": "images/b19b9e13cc89a4bc5588dcc2e2b775a88ed1d41b6ad93f8dba25041b436eca27.jpg", + "image_caption": [ + "ZoeDepth [2]" + ], + "image_footnote": [], + "bbox": [ + 333, + 310, + 442, + 364 + ], + "page_idx": 9 + }, + { + "type": "image", + "img_path": "images/e9a3ca30e4056ab2eb9607b245365d894e49378b63ba9ba373a4e011beecfbaf.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 446, + 143, + 555, + 198 + ], + "page_idx": 9 + }, + { + "type": "image", + "img_path": "images/c0502f40097d064392fc97ab1269e4c981c13d4fa6e2c4cdf3f31355b611e0d6.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 446, + 199, + 555, + 253 + ], + "page_idx": 9 + }, + { + "type": "image", + "img_path": "images/863fabd6a3c138b8e0471cd71e00baf78300bc1742a3ae7c3c105e2861df0824.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 446, + 253, + 555, + 309 + ], + "page_idx": 9 + }, + { + "type": "image", + "img_path": "images/00e2026fb09e55e8d5039a008a63e4884e54eb2a6da4ff3fcfbf0c5a5719c3fb.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 446, + 310, + 555, + 364 + ], + "page_idx": 9 + }, + { + "type": "image", + "img_path": "images/a8f651b632aa099bbac8a3b95114804ba25dca6b6c202cd2cffaa60a8b4bf746.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 558, + 143, + 668, + 198 + ], + "page_idx": 9 + }, + { + "type": "image", + "img_path": "images/84b8e82a90297997638098c428dfddb86c5b55666d54b8d43c3c18f776f7e6a3.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 558, + 199, + 668, + 253 + ], + "page_idx": 9 + }, + { + "type": "image", + "img_path": "images/efca61ec132355ce07605ec69b5119fbdfe6a3702778d838c9fece8f2131430b.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 558, + 253, + 668, + 309 + ], + "page_idx": 9 + }, + { + "type": "image", + "img_path": "images/f5ce4825eedcf7cbff277dddffaf217a205310d89de00c536c1c4bca6d1412dc.jpg", + "image_caption": [ + "PR R", + "Fig. 6: Qualitative Comparison on ETH3D and ScanNet++. The first two rows depict results for ETH3D and the last two for ScanNet++. The baseline ZoeDepth and PatchRefiner with conventional fine tuning (PR $\\mathcal{R}$ ) both struggle to create high-resolution depth details near boundaries. Our proposed strategy yields crisp boundaries on both datasets." + ], + "image_footnote": [], + "bbox": [ + 558, + 310, + 668, + 364 + ], + "page_idx": 9 + }, + { + "type": "image", + "img_path": "images/b459ecba07ad557a1075c3a3e623799e31e01f0b7d84d36429e807a6394b85b9.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 669, + 143, + 779, + 198 + ], + "page_idx": 9 + }, + { + "type": "image", + "img_path": "images/b373343d15848305d659867a958f194cae8795048ebee5b883b89e78237d05f4.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 669, + 199, + 779, + 253 + ], + "page_idx": 9 + }, + { + "type": "image", + "img_path": "images/ba004701f1ea2d7cb45f30200b87a83a2e7c9c1bccf0976d5571cd0a2399bd98.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 669, + 253, + 779, + 309 + ], + "page_idx": 9 + }, + { + "type": "image", + "img_path": "images/27c246a3de63a7759baf0f2a00b0e2aaa257713ad856c9a24a5e3fd655bbeff4.jpg", + "image_caption": [ + "Ours 1" + ], + "image_footnote": [], + "bbox": [ + 669, + 310, + 779, + 364 + ], + "page_idx": 9 + }, + { + "type": "text", + "text": "of $1024 \\times 2048$ , surpassing many real-domain datasets in density, quantity, and resolution [45, 46, 48, 49]. We use a standard patch size of $256 \\times 512$ and conduct most experiments on this dataset.", + "bbox": [ + 212, + 506, + 784, + 551 + ], + "page_idx": 9 + }, + { + "type": "text", + "text": "ScanNet++ (Real, LiDAR, Reconstruction): ScanNet++ [58] is an extensive indoor dataset providing high-resolution images $(1440\\times 1920)$ , depth maps from iPhone LiDAR $(192\\times 256)$ , and high-resolution depth maps sampled from reconstructions of laser scans $(1440\\times 1920)$ . Our chosen patch size is $720\\times 960$ . We use the low-resolution ground-truth for training since the high-resolution version contains much noise as shown in Fig. 3.", + "bbox": [ + 212, + 556, + 787, + 648 + ], + "page_idx": 9 + }, + { + "type": "text", + "text": "ETH3D (Real, LiDAR): The ETH3D benchmark [46] provides high-resolution indoor and outdoor images $(6048\\times 4032)$ with ground-truth depth from laser sensors. We downsample the image-depth pairs to $2160\\times 3840$ and select a patch size of $540\\times 960$ .", + "bbox": [ + 212, + 652, + 787, + 712 + ], + "page_idx": 9 + }, + { + "type": "text", + "text": "Metrics: We adopt standard depth evaluation metrics from [2, 11, 38] and the Soft Edge Error (SEE) from [4, 27, 51] for scale evaluation. Additional metric details are in the supplementary material. Given real-world depth map boundary incompleteness, we design metrics to focus on evaluating boundary accuracy, leveraging high-resolution fine-grained segmentation masks as depth boundary quality proxies. Specifically, we use the Sobel operator [21] on both predicted depth and segmentation maps to generate edge masks, and then compute Precision, Recall, and F1 Score.", + "bbox": [ + 212, + 719, + 787, + 839 + ], + "page_idx": 9 + }, + { + "type": "page_number", + "text": "10", + "bbox": [ + 217, + 114, + 235, + 126 + ], + "page_idx": 9 + }, + { + "type": "header", + "text": "Zhenyu Li, Shariq Farooq Bhat, Peter Wonka", + "bbox": [ + 271, + 114, + 575, + 128 + ], + "page_idx": 9 + }, + { + "type": "image", + "img_path": "images/3eebc3454996388cb42a2c5423d6d7d2eae964d238e942613c36dfeb31f78dd4.jpg", + "image_caption": [ + "Fig. 7: Qualitative Comparison on CityScapes. This figure illustrates depth estimation comparisons between the base ZoeDepth model, PatchRefiner (PR) trained on CityScapes, and our method. We display outcomes under varying levels of $\\mathcal{L}_{pl}$ supervision ( $\\lambda_1 = \\lambda_2 = 1e^{-1}$ or 1), featuring zoomed-in sections of each image to highlight detail fidelity near depth discontinuities." + ], + "image_footnote": [], + "bbox": [ + 220, + 143, + 331, + 422 + ], + "page_idx": 10 + }, + { + "type": "image", + "img_path": "images/cbf42d6b563403d4ce2a678da840a98d7d2a104b7e2b532d11f0db1cddb31694.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 334, + 143, + 442, + 422 + ], + "page_idx": 10 + }, + { + "type": "image", + "img_path": "images/8ee3869a0da1c121b5c6d2d06617e3fda7827f30812088e868cf52a6d6699fac.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 446, + 143, + 555, + 422 + ], + "page_idx": 10 + }, + { + "type": "image", + "img_path": "images/8f34a358dd8d4a536260f08c20f741e16c67555430f728f0ceb03e24d447b9a3.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 558, + 143, + 666, + 422 + ], + "page_idx": 10 + }, + { + "type": "image", + "img_path": "images/910705c5cb23bf68a742fe54f336dfc9e91ef3446490140b06f65a5650d98a5e.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 669, + 143, + 779, + 422 + ], + "page_idx": 10 + }, + { + "type": "text", + "text": "4.2 Implementation Details", + "text_level": 1, + "bbox": [ + 215, + 540, + 455, + 555 + ], + "page_idx": 10 + }, + { + "type": "text", + "text": "PatchRefiner on Synthetic Dataset: For training on synthetic datasets, we employ the scale-invariant log loss $\\mathcal{L}_{\\text{silog}}$ , as outlined in [2, 11, 27, 30]. Initialization of the coarse network $\\mathcal{N}_c$ leverages pretrained weights from the NYU-v2 dataset [48], adhering to the approach in [27]. We dedicate 24 epochs to training $\\mathcal{N}_c$ , which is subsequently frozen to ensure stability in subsequent training phases. For the refinement network $\\mathcal{N}_r$ , initialization employs $\\mathcal{N}_c$ 's parameters, and training extends for an additional 36 epochs. Standard augmentation strategies from the baseline depth model are incorporated to enhance training effectiveness. During inference, we implement Consistency-Aware Inference, as described in [27], to optimize performance.", + "bbox": [ + 212, + 580, + 787, + 731 + ], + "page_idx": 10 + }, + { + "type": "text", + "text": "Learning on Real-Domain Dataset: Since the coarse model within the framework offers scale-consistent predictions, we first pretrain a coarse model on real-domain data to establish a reliable depth scale foundation. This model is subsequently frozen to preserve scale consistency throughout the training process. The refiner model in PatchRefiner is then initialized with parameters pretrained on synthetic data, enabling it to maintain high-frequency detail knowledge acquired from the synthetic domain. Initially, the student model is trained solely", + "bbox": [ + 212, + 734, + 787, + 840 + ], + "page_idx": 10 + }, + { + "type": "header", + "text": "PatchRefiner", + "bbox": [ + 642, + 114, + 730, + 126 + ], + "page_idx": 10 + }, + { + "type": "page_number", + "text": "11", + "bbox": [ + 767, + 114, + 784, + 126 + ], + "page_idx": 10 + }, + { + "type": "table", + "img_path": "images/122b8bbfa8120f6c0e61d22bf2087987dacf1a98bc14e65aaeb1be74878a393a.jpg", + "table_caption": [ + "Table 1: Quantitative comparison on UnrealStereo4K. We color code the corresponding best competitor and our method within each block. PF and PR are short for PatchFusion [27] and PatchRefiner, respectively. The reported numbers are from [27]." + ], + "table_footnote": [], + "table_body": "
Methodδ1(%)↑REL↓RMS↓SiLog↓SEE↓Reference
iDisc [38]96.9400.0531.4048.5021.070ICCV 2023
SMD-Net [51]97.7740.0441.2827.3890.883CVPR 2021
Graph-GDSR [10]97.9320.0441.2647.4690.872CVPR 2022
BoostingDepth [34]98.1040.0441.1236.6620.939CVPR 2021
ZoeDepth [2]97.7170.0461.2897.4480.914-
ZoeDepth+PFp=16 [27]98.4190.0401.0886.2120.838CVPR 2024
ZoeDepth+PFp=49 [27]98.4500.0391.0756.1310.846
ZoeDepth+PFR=128 [27]98.4690.0391.0666.0850.849
ZoeDepth+PRp=1698.8210.0330.8925.4170.750Ours
ZoeDepth+PRp=4998.8590.0330.8705.3190.751
ZoeDepth+PRR=12898.8640.0330.8725.3770.738
Depth-Anything [56]97.7730.0411.2357.1920.911CVPR 2024
Depth-Anything+PFp=16 [27]98.5580.0361.0155.8830.811CVPR 2024
Depth-Anything+PFp=49 [27]98.6070.0350.9875.7460.812
Depth-Anything+PFR=128 [27]98.6160.0350.9845.7750.813
Depth-Anything+PRp=1698.8260.0330.8895.2890.768Ours
Depth-Anything+PRp=4998.8780.0330.8605.1490.767
Depth-Anything+PRR=12898.8780.0330.8605.2060.759
", + "bbox": [ + 225, + 198, + 776, + 454 + ], + "page_idx": 11 + }, + { + "type": "table", + "img_path": "images/8637ab8ce4d95789d649ea22289681017732fd0ea26c9b7b169882df9c45d605.jpg", + "table_caption": [ + "Table 2: Quantitative comparison on CityScapes. FT, and MIX are short for fine-tuning and mixed-data strategies, which are our main competitors. Baseline is highlighted in gray and each of the three competitors in a different color." + ], + "table_footnote": [], + "table_body": "
MethodDataScaleBoundary
SRδ1(%)↑REL↓RMS↓Precision↑Recall↑
ZoeDepth [2]✓94.5020.0704.40613.3237.59
ZoeDepth [2] + FT✓✓94.4980.0714.41812.9337.89
PatchRefiner (zero-shot)✓5.7050.39912.20328.6851.28
PatchRefiner✓95.2840.0664.04716.6739.53
PatchRefiner + FT✓✓95.4180.0653.99217.0940.92
PatchRefiner + mix [39]✓✓89.1080.1124.73223.0841.17
PatchRefiner + DSD✓✓95.3590.0663.98218.9248.78
", + "bbox": [ + 230, + 521, + 771, + 642 + ], + "page_idx": 11 + }, + { + "type": "text", + "text": "with $\\mathcal{L}_{\\text{silog}}$ for 24 epochs. Subsequent fine-tuning with our Detail and Scale Disentangling loss $\\mathcal{L}_{DSD}$ over an additional 6 epochs to refine depth estimations.", + "bbox": [ + 215, + 670, + 782, + 700 + ], + "page_idx": 11 + }, + { + "type": "text", + "text": "4.3 Main Results", + "text_level": 1, + "bbox": [ + 215, + 723, + 372, + 736 + ], + "page_idx": 11 + }, + { + "type": "text", + "text": "Synthetic Dataset: On the UnrealStereo4K dataset, PatchRefiner not only outperforms the base depth model but also shows substantial improvements over PatchFusion, reducing RMSE by $18.1\\%$ and REL by $15.7\\%$ . This advancement is further underscored by achieving the lowest SEE, highlighting our model's proficiency in capturing edge details. Qualitative comparisons, illustrated in Fig. 5, indicate the superior boundary delineation achieved by PatchRefiner.", + "bbox": [ + 215, + 750, + 782, + 839 + ], + "page_idx": 11 + }, + { + "type": "page_number", + "text": "12", + "bbox": [ + 217, + 114, + 235, + 126 + ], + "page_idx": 11 + }, + { + "type": "header", + "text": "Zhenyu Li, Shariq Farooq Bhat, Peter Wonka", + "bbox": [ + 271, + 114, + 573, + 128 + ], + "page_idx": 11 + }, + { + "type": "table", + "img_path": "images/e12c7381bb8a16edcb3e164843734ea41ce8f7e3986a7794c73443d1d9b0de46.jpg", + "table_caption": [ + "Table 3: Ablation study of architecture variations and formulation of final depth prediction. Ours is highlighted with color. We highlight the best in bold." + ], + "table_footnote": [], + "table_body": "
MethodType of outputFeature Levelsδ1(%)↑REL↓RMS↓SiLog↓SEE↓
PatchFusion [27]Direct6 features98.4190.0401.0886.2120.838
PatchRefinerDcresidual1 features98.7340.0340.9265.5500.782
Dcresidual2 features98.8140.0330.9055.5110.750
Dcresidual3 features98.8150.0340.9005.5830.753
Dcresidual4 features98.8150.0330.8995.4940.752
Dcresidual5 features98.8140.0330.8945.4680.752
Dcresidual6 features98.8210.0330.8925.4170.750
Ddresidual6 features98.8040.0330.8995.4480.753
Direct6 features98.7490.0340.9255.5910.765
", + "bbox": [ + 222, + 184, + 779, + 316 + ], + "page_idx": 12 + }, + { + "type": "table", + "img_path": "images/986358a00d4dad570fe0089a8c082b78b81f03dc4f6a5810221d2dafde222d66.jpg", + "table_caption": [ + "Table 4: Variations of $\\mathcal{L}_{pl}$ . We analyse various options for $\\mathcal{L}_{pl}$ and compare them against PR. $\\mathcal{S}$ and PR. $\\mathcal{R}$ that serve as baselines for training on Synthetic and Real data individually. We set $\\lambda = \\lambda_1 = \\lambda_2$ to analyze the influence of DSD weight. The highlighted result is achieved with $\\lambda = 1e^{-1}$ ." + ], + "table_footnote": [], + "table_body": "
Variations (Lpl)ScaleBoundary
δ1(%)↑REL↓RMS↓Precision↑Recall↑F1↑
Baseline PR. S5.7050.39912.2028.6851.2836.34
Baseline PR. R95.4180.0653.99217.0940.9223.68
Lsilog82.5500.1555.91426.3458.3535.81
Lsilog + mask81.4250.1486.51330.0446.8936.19
Lrank95.4130.0663.97318.3847.8126.12
Lssi95.4650.0653.97419.2644.0126.39
Lrank + Lssi (Ours)95.3590.0663.98218.9248.7826.84
λ = 195.0770.0694.22221.3258.4830.90
λ = 3e-195.2960.0684.08620.1153.8328.91
λ = 3e-295.4620.0653.95318.0143.9025.11
", + "bbox": [ + 222, + 397, + 774, + 561 + ], + "page_idx": 12 + }, + { + "type": "text", + "text": "Real-Domain Dataset: Tab. 2 and Fig. 6, 7 delineate the performance disparity when leveraging synthetic data for real-domain learning. Although the synthetic-trained model excels in boundary details, it fails in scale accuracy due to the domain gap. Sole training on real-domain data enhances baseline's scale prediction yet falls short in detail accuracy due to the missing depth ground truth around boundaries. Neither fine-tuning nor mixed-training substantially elevates performance across scale and detail metrics, reflecting the inherent challenges in our task. Contrastingly, our strategy propels the model to notable gains in boundary accuracy (19.2% increase in boundary recall) while sustaining scale precision comparable to the baseline model.", + "bbox": [ + 212, + 584, + 784, + 733 + ], + "page_idx": 12 + }, + { + "type": "text", + "text": "4.4 Ablation Studies and Discussion", + "text_level": 1, + "bbox": [ + 215, + 755, + 527, + 768 + ], + "page_idx": 12 + }, + { + "type": "text", + "text": "We ablate and discuss the contributions of individual components. As default, we utilize the UnrealStereo4K dataset for synthetic-dataset training and the PatchRefiner variant with $P = 16$ patches for clarity and ease of comparison. We ablate our teacher-student framework using the CityScapes dataset.", + "bbox": [ + 212, + 779, + 782, + 839 + ], + "page_idx": 12 + }, + { + "type": "header", + "text": "PatchRefiner", + "bbox": [ + 642, + 114, + 730, + 127 + ], + "page_idx": 12 + }, + { + "type": "page_number", + "text": "13", + "bbox": [ + 767, + 114, + 785, + 126 + ], + "page_idx": 12 + }, + { + "type": "text", + "text": "Architecture and Formulation Variations: We first ablate the effectiveness of our architecture design. Utilizing a ZoeDepth model, we extract six levels of intermediate features during the forward pass, represented as $\\mathcal{F} = f_1, f_2, \\dots, f_6$ , with progressively increasing resolution. We then sequentially omit lower resolution feature maps from the refiner's decoder input. As depicted in Tab. 3, performance diminishes with fewer feature maps, yet even with solely the highest resolution feature map, our model outperforms PatchFusion.", + "bbox": [ + 212, + 146, + 787, + 252 + ], + "page_idx": 13 + }, + { + "type": "text", + "text": "Further, we explore our refinement formulation's efficacy by comparing it against two alternatives: (1) Residual depth adjustment based on the base model $\\mathcal{N}_d$ within the refiner. (2) Direct depth prediction from the refiner's decoder, similar to PatchFusion ( $\\mathbf{D} = \\mathbf{D}_r$ in this context). Our residual-based approach demonstrates superior performance, underscoring its advantages in optimization and training efficiency.", + "bbox": [ + 212, + 252, + 787, + 343 + ], + "page_idx": 13 + }, + { + "type": "text", + "text": "Effectiveness of Pseudo Label Supervision: We evaluate the efficacy of our Detail and Scale Disentangling (DSD) loss against the conventional scale-invariant loss, $\\mathcal{L}_{silog}$ , for pseudo label supervision. Tab. 4 illustrates that while $\\mathcal{L}_{silog}$ aids in detail transfer from the teacher to the student model, it compromises scale accuracy due to significant discrepancies in the pseudo labels. A masking approach, focusing only on areas lacking depth, does not mitigate this issue, indicating pervasive negative effects. In contrast, the combination of $\\mathcal{L}_{rank}$ and $\\mathcal{L}_{ssi}$ within $\\mathcal{L}_{pl}$ not only improves detail fidelity but also maintains scale accuracy, demonstrating that ranking constraints and scale-shift invariance are effective, orthogonal strategies for enhancing high-resolution detail without sacrificing scale accuracy.", + "bbox": [ + 212, + 362, + 787, + 529 + ], + "page_idx": 13 + }, + { + "type": "text", + "text": "Exploration of Variant DSD Loss Weight: This study examines the impact of different DSD loss weights $(\\lambda_{1}$ and $\\lambda_{2})$ to elucidate the loss's efficiency. As shown in Tab. 4 and Fig. 7, increasing the loss weight marginally affects scale accuracy while significantly improving boundaries, validating the DSD loss's role in balancing detail enhancement and scale preservation. Notably, when both weights are set to 1, the model's boundary recall surpasses the teacher's performance. Furthermore, when DSD loss achieves a comparable boundary metric to $\\mathcal{L}_{silog}$ , it exhibits a smaller performance decline, underscoring its effectiveness.", + "bbox": [ + 212, + 547, + 787, + 672 + ], + "page_idx": 13 + }, + { + "type": "text", + "text": "5 Conclusion", + "text_level": 1, + "bbox": [ + 214, + 689, + 359, + 707 + ], + "page_idx": 13 + }, + { + "type": "text", + "text": "We presented PatchRefiner, a tile-based framework tailored to real-world high-resolution monocular metric depth estimation. It reconceptualizes high-resolution depth estimation as a refinement process. With a pseudo-labeling strategy that leverages synthetic data, we propose a Detail and Scale Disentangling (DSD) loss to enhance detail capture while maintaining scale accuracy. Our proposed framework decisively surpasses the current SOTA method for UnrealStereo4K (17.3% in RMSE), while demonstrating marked improvements in detail accuracy on real-world datasets such as CityScape, ScanNet++, and ETH3D.", + "bbox": [ + 212, + 719, + 787, + 840 + ], + "page_idx": 13 + }, + { + "type": "page_number", + "text": "14", + "bbox": [ + 217, + 114, + 235, + 126 + ], + "page_idx": 13 + }, + { + "type": "header", + "text": "Zhenyu Li, Shariq Farooq Bhat, Peter Wonka", + "bbox": [ + 271, + 114, + 573, + 128 + ], + "page_idx": 13 + }, + { + "type": "text", + "text": "Acknowledgements", + "text_level": 1, + "bbox": [ + 217, + 143, + 401, + 162 + ], + "page_idx": 14 + }, + { + "type": "text", + "text": "This publication is funded in part by KAUST under award #ORA-2023-5241 (for the NTGI-AI project).", + "bbox": [ + 215, + 174, + 782, + 204 + ], + "page_idx": 14 + }, + { + "type": "text", + "text": "References", + "text_level": 1, + "bbox": [ + 217, + 224, + 321, + 239 + ], + "page_idx": 14 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "1. Bhat, S.F., Alhashim, I., Wonka, P.: Adabins: Depth estimation using adaptive bins. In: CVPR. pp. 4009-4018 (2021) 3", + "2. Bhat, S.F., Birkl, R., Wofk, D., Wonka, P., Müller, M.: Zoedepth: Zero-shot transfer by combining relative and metric depth. arXiv preprint arXiv:2302.12288 (2023) 1, 3, 7, 9, 10, 11, 12", + "3. Chen, C., Xie, W., Huang, W., Rong, Y., Ding, X., Huang, Y., Xu, T., Huang, J.: Progressive feature alignment for unsupervised domain adaptation. In: Proceedings of the IEEE/CVF conference on computer vision and pattern recognition. pp. 627-636 (2019) 3", + "4. Chen, C., Chen, X., Cheng, H.: On the over-smoothing problem of cnn based disparity estimation. In: ICCV. pp. 8997-9005 (2019) 10", + "5. Chen, W., Fu, Z., Yang, D., Deng, J.: Single-image depth perception in the wild. NeurIPS 29 (2016) 3, 7, 8", + "6. Chen, Y., Liu, S., Wang, X.: Learning continuous image representation with local implicit image function. In: CVPR. pp. 8628-8638 (2021) 3", + "7. Chen, Y.C., Lin, Y.Y., Yang, M.H., Huang, J.B.: Crdoco: Pixel-level domain transfer with cross-domain consistency. In: Proceedings of the IEEE/CVF conference on computer vision and pattern recognition. pp. 1791-1800 (2019) 3", + "8. Chen, Z., Zhang, R., Zhang, G., Ma, Z., Lei, T.: Digging into pseudo label: a low-budget approach for semi-supervised semantic segmentation. IEEE Access 8, 41830-41837 (2020) 3", + "9. Cordts, M., Omran, M., Ramos, S., Rehfeld, T., Enzweiler, M., Benenson, R., Franke, U., Roth, S., Schiele, B.: The cityscapes dataset for semantic urban scene understanding. In: CVPR. pp. 3213-3223 (2016) 3, 6, 9", + "10. De Lutio, R., Becker, A., D'Aronco, S., Russo, S., Wegner, J.D., Schindler, K.: Learning graph regularisation for guided super-resolution. In: CVPR. pp. 1979-1988 (2022) 12", + "11. Eigen, D., Puhrsch, C., Fergus, R.: Depth map prediction from a single image using a multi-scale deep network. NeurIPS 27 (2014) 1, 3, 7, 10, 11", + "12. Farahani, A., Voghoei, S., Rasheed, K., Arabnia, H.R.: A brief review of domain adaptation. Advances in data science and information engineering: proceedings from ICDATA 2020 and IKE 2020 pp. 877-894 (2021) 2", + "13. Fu, H., Gong, M., Wang, C., Batmanghelich, K., Tao, D.: Deep ordinal regression network for monocular depth estimation. In: CVPR. pp. 2002-2011 (2018) 3", + "14. Geiger, A., Lenz, P., Urtasun, R.: Are we ready for autonomous driving? the kitti vision benchmark suite. In: CVPR. pp. 3354-3361. IEEE (2012) 5, 6", + "15. Godard, C., Mac Aodha, O., Firman, M., Brostow, G.J.: Digging into self-supervised monocular depth estimation. In: ICCV. pp. 3828-3838 (2019) 3", + "16. He, K., Gkioxari, G., Dollár, P., Girshick, R.: Mask r-cnn. In: Proceedings of the IEEE international conference on computer vision. pp. 2961-2969 (2017) 5", + "17. Hu, Z., Yang, Z., Hu, X., Nevatia, R.: Simple: Similar pseudo label exploitation for semi-supervised classification. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 15099-15108 (2021) 3" + ], + "bbox": [ + 218, + 253, + 784, + 839 + ], + "page_idx": 14 + }, + { + "type": "header", + "text": "PatchRefiner", + "bbox": [ + 642, + 114, + 730, + 127 + ], + "page_idx": 14 + }, + { + "type": "page_number", + "text": "15", + "bbox": [ + 767, + 116, + 785, + 126 + ], + "page_idx": 14 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "18. Huang, P.H., Matzen, K., Kopf, J., Ahuja, N., Huang, J.B.: Deepmvs: Learning multi-view stereopsis. In: CVPR (2018) 5", + "19. Hui, T.W., Loy, C.C., Tang, X.: Depth map super-resolution by deep multi-scale guidance. In: ECCV. pp. 353-369. Springer (2016) 3", + "20. Janoch, A., Karayev, S., Jia, Y., Barron, J.T., Fritz, M., Saenko, K., Darrell, T.: A category-level 3d object dataset: Putting the kinect to work. Consumer Depth Cameras for Computer Vision: Research Topics and Applications pp. 141-165 (2013) 5", + "21. Kanopoulos, N., Vasanthavada, N., Baker, R.L.: Design of an image edge detection filter using the sobel operator. IEEE Journal of solid-state circuits 23(2), 358-367 (1988) 10", + "22. Koutilya, P., Zhou, H., Jacobs, D.: Sharingan: Combining synthetic and real data for unsupervised geometry estimation. In: CVPR. vol. 2, p. 5 (2020) 3", + "23. Kundu, J.N., Uppala, P.K., Pahuja, A., Babu, R.V.: Adadepth: Unsupervised content congruent adaptation for depth estimation. In: Proceedings of the IEEE conference on computer vision and pattern recognition. pp. 2656-2665 (2018) 3", + "24. Lee, J.H., Kim, C.S.: Multi-loss rebalancing algorithm for monocular depth estimation. In: ECCV. pp. 785-801. Springer (2020) 3, 7", + "25. Lehtinen, J., Munkberg, J., Hasselgren, J., Laine, S., Karras, T., Aittala, M., Aila, T.: Noise2noise: Learning image restoration without clean data. arXiv preprint arXiv:1803.04189 (2018) 5", + "26. Li, Y., Yuan, L., Vasconcelos, N.: Bidirectional learning for domain adaptation of semantic segmentation. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 6936-6945 (2019) 3", + "27. Li, Z., Bhat, S.F., Wonka, P.: Patchfusion: An end-to-end tile-based framework for high-resolution monocular metric depth estimation. arXiv preprint arXiv:2312.02284 (2023) 1, 2, 3, 4, 5, 6, 7, 9, 10, 11, 12, 13", + "28. Li, Z., Chen, Z., Li, A., Fang, L., Jiang, Q., Liu, X., Jiang, J.: Unsupervised domain adaptation for monocular 3d object detection via self-training. In: European Conference on Computer Vision. pp. 245-262. Springer (2022) 3", + "29. Li, Z., Chen, Z., Liu, X., Jiang, J.: Depthformer: Exploiting long-range correlation and local information for accurate monocular depth estimation. Machine Intelligence Research pp. 1-18 (2023) 3", + "30. Li, Z., Wang, X., Liu, X., Jiang, J.: Binsformer: Revisiting adaptive bins for monocular depth estimation. arXiv preprint arXiv:2204.00987 (2022) 1, 3, 7, 11", + "31. Liu, C., Kumar, S., Gu, S., Timofte, R., Van Gool, L.: Single image depth prediction made better: A multivariate gaussian take. In: CVPR. pp. 17346-17356 (June 2023) 7", + "32. Lopez-Rodriguez, A., Mikolajczyk, K.: Desc: Domain adaptation for depth estimation via semantic consistency. International Journal of Computer Vision 131(3), 752-771 (2023) 3, 4, 6", + "33. Metzger, N., Daudt, R.C., Schindler, K.: Guided depth super-resolution by deep anisotropic diffusion. In: CVPR. pp. 18237-18246 (2023) 3", + "34. Miangoleh, S.M.H., Dille, S., Mai, L., Paris, S., Aksoy, Y.: Boosting monocular depth estimation models to high-resolution via content-adaptive multi-resolution merging. In: CVPR. pp. 9685-9694 (2021) 1, 2, 3, 4, 5, 12", + "35. Mildenhall, B., Srinivasan, P.P., Tancik, M., Barron, J.T., Ramamoorthi, R., Ng, R.: Nerf: Representing scenes as neural radiance fields for view synthesis. Communications of the ACM 65(1), 99-106 (2021) 3" + ], + "bbox": [ + 215, + 146, + 784, + 839 + ], + "page_idx": 15 + }, + { + "type": "page_number", + "text": "16", + "bbox": [ + 217, + 114, + 235, + 126 + ], + "page_idx": 15 + }, + { + "type": "header", + "text": "Zhenyu Li, Shariq Farooq Bhat, Peter Wonka", + "bbox": [ + 271, + 114, + 575, + 128 + ], + "page_idx": 15 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "36. Pastore, G., Cermelli, F., Xian, Y., Mancini, M., Akata, Z., Caputo, B.: A closer look at self-training for zero-label semantic segmentation. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 2693-2702 (2021) 3", + "37. Paul, S., Tsai, Y.H., Schulter, S., Roy-Chowdhury, A.K., Chandraker, M.: Domain adaptive semantic segmentation using weak labels. In: Computer Vision-ECCV 2020: 16th European Conference, Glasgow, UK, August 23-28, 2020, Proceedings, Part IX 16. pp. 571-587. Springer (2020) 3", + "38. Piccinelli, L., Sakaridis, C., Yu, F.: idisc: Internal discretization for monocular depth estimation. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 21477-21487 (2023) 10, 12", + "39. Poucin, F., Kraus, A., Simon, M.: Boosting instance segmentation with synthetic data: A study to overcome the limits of real world data sets. In: Int. Conf. Comput. Vis. Workshop. pp. 945-953 (2021) 2, 5, 6, 12", + "40. Pseudo-Label, D.H.L.: The simple and efficient semi-supervised learning method for deep neural networks. In: ICML 2013 Workshop: Challenges in Representation Learning. pp. 1-6 (2013) 3", + "41. Rajpal, A., Cheema, N., Illgner-Fehns, K., Slusallek, P., Jaiswal, S.: High-resolution synthetic rgb-d datasets for monocular depth estimation. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 1188-1198 (2023) 3, 6", + "42. Ranftl, R., Lasinger, K., Hafner, D., Schindler, K., Koltun, V.: Towards robust monocular depth estimation: Mixing datasets for zero-shot cross-dataset transfer. IEEE TPAMI 44(3) (2022) 3, 7, 8", + "43. Rey-Area, M., Yuan, M., Richardt, C.: 360monodepth: High-resolution 360deg monocular depth estimation. In: CVPR. pp. 3762-3772 (2022) 1, 2, 3", + "44. Saito, K., Ushiku, Y., Harada, T.: Asymmetric tri-training for unsupervised domain adaptation. In: International Conference on Machine Learning. pp. 2988-2997. PMLR (2017) 3", + "45. Scharstein, D., Hirschmüller, H., Kitajima, Y., Krathwohl, G., Nesic, N., Wang, X., Westling, P.: High-resolution stereo datasets with subpixel-accurate ground truth. In: Pattern Recognition: 36th German Conference, GCPR 2014, Münster, Germany, September 2-5, 2014, Proceedings 36. pp. 31-42. Springer (2014) 6, 10", + "46. Schops, T., Schonberger, J.L., Galliani, S., Sattler, T., Schindler, K., Pollefeys, M., Geiger, A.: A multi-view stereo benchmark with high-resolution images and multi-camera videos. In: CVPR. pp. 3260-3269 (2017) 3, 5, 6, 10", + "47. Shin, I., Tsai, Y.H., Zhuang, B., Schulter, S., Liu, B., Garg, S., Kweon, I.S., Yoon, K.J.: Mm-tta: multi-modal test-time adaptation for 3d semantic segmentation. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 16928-16937 (2022) 3", + "48. Silberman, N., Hoiem, D., Kohli, P., Fergus, R.: Indoor segmentation and support inference from rgbd images. In: ECCV. pp. 746-760. Springer (2012) 5, 6, 10, 11", + "49. Song, S., Lichtenberg, S.P., Xiao, J.: Sun rgb-d: A rgb-d scene understanding benchmark suite. In: CVPR. pp. 567-576 (2015) 5, 6, 10", + "50. Taherkhani, F., Dabouei, A., Soleymani, S., Dawson, J., Nasrabadi, N.M.: Self-supervised Wasserstein pseudo-labeling for semi-supervised image classification. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 12267-12277 (2021) 3", + "51. Tosi, F., Liao, Y., Schmitt, C., Geiger, A.: Smd-nets: Stereo mixture density networks. In: CVPR. pp. 8942-8952 (2021) 2, 3, 5, 9, 10, 12" + ], + "bbox": [ + 212, + 146, + 787, + 839 + ], + "page_idx": 16 + }, + { + "type": "header", + "text": "PatchRefiner", + "bbox": [ + 642, + 114, + 730, + 126 + ], + "page_idx": 16 + }, + { + "type": "page_number", + "text": "17", + "bbox": [ + 767, + 114, + 785, + 126 + ], + "page_idx": 16 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "52. Wang, Z., Bovik, A.C., Sheikh, H.R., Simoncelli, E.P.: Image quality assessment: from error visibility to structural similarity. IEEE TIP 13(4), 600-612 (2004) 9", + "53. Weiss, K., Khoshgoftaar, T.M., Wang, D.: A survey of transfer learning. Journal of Big data 3(1), 1-40 (2016) 2", + "54. Xian, K., Zhang, J., Wang, O., Mai, L., Lin, Z., Cao, Z.: Structure-guided ranking loss for single image depth prediction. In: CVPR. pp. 611-620 (2020) 3, 7, 8", + "55. Yang, J., Alvarez, J.M., Liu, M.: Self-supervised learning of depth inference for multi-view stereo. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 7526-7534 (2021) 4", + "56. Yang, L., Kang, B., Huang, Z., Xu, X., Feng, J., Zhao, H.: Depth anything: Unleashing the power of large-scale unlabeled data. arXiv preprint arXiv:2401.10891 (2024) 1, 3, 4, 6, 7, 8, 12", + "57. Yen, Y.T., Lu, C.N., Chiu, W.C., Tsai, Y.H.: 3d-pl: Domain adaptive depth estimation with 3d-aware pseudo-labeling. In: European Conference on Computer Vision. pp. 710-728. Springer (2022) 4, 6", + "58. Yeshwanth, C., Liu, Y.C., Nießner, M., Dai, A.: Scannet++: A high-fidelity dataset of 3d indoor scenes. In: ICCV. pp. 12-22 (2023) 3, 5, 6, 10", + "59. Zhang, L., Rao, A., Agrawala, M.: Adding conditional control to text-to-image diffusion models. In: ICCV. pp. 3836-3847 (2023) 1", + "60. Zhao, S., Fu, H., Gong, M., Tao, D.: Geometry-aware symmetric domain adaptation for monocular depth estimation. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 9788-9798 (2019) 3", + "61. Zhao, W., Rao, Y., Liu, Z., Liu, B., Zhou, J., Lu, J.: Unleashing text-to-image diffusion models for visual perception. ICCV (2023) 1", + "62. Zhao, X., Schulter, S., Sharma, G., Tsai, Y.H., Chandraker, M., Wu, Y.: Object detection with a unified label space from multiple datasets. In: Computer Vision-ECCV 2020: 16th European Conference, Glasgow, UK, August 23-28, 2020, Proceedings, Part XIV 16. pp. 178-193. Springer (2020) 3", + "63. Zhao, Z., Zhang, J., Xu, S., Lin, Z., Pfister, H.: Discrete cosine transform network for guided depth map super-resolution. In: CVPR. pp. 5697-5707 (2022) 3", + "64. Zheng, C., Cham, T.J., Cai, J.: T2net: Synthetic-to-realistic translation for solving single-image depth estimation tasks. In: Proceedings of the European conference on computer vision (ECCV). pp. 767-783 (2018) 3", + "65. Zhong, Z., Liu, X., Jiang, J., Zhao, D., Ji, X.: Guided depth map super-resolution: A survey. ACM Computing Surveys (2023) 3", + "66. Zou, Y., Yu, Z., Kumar, B., Wang, J.: Unsupervised domain adaptation for semantic segmentation via class-balanced self-training. In: Proceedings of the European conference on computer vision (ECCV). pp. 289-305 (2018) 3" + ], + "bbox": [ + 215, + 146, + 784, + 672 + ], + "page_idx": 17 + }, + { + "type": "page_number", + "text": "18", + "bbox": [ + 217, + 114, + 235, + 126 + ], + "page_idx": 17 + }, + { + "type": "header", + "text": "Zhenyu Li, Shariq Farooq Bhat, Peter Wonka", + "bbox": [ + 271, + 114, + 575, + 128 + ], + "page_idx": 17 + } +] \ No newline at end of file diff --git a/2024/PatchRefiner_ Leveraging Synthetic Data for Real-Domain High-Resolution Monocular Metric Depth Estimation/3be463cc-22b0-43d6-8e8c-f91c9fa34d5b_model.json b/2024/PatchRefiner_ Leveraging Synthetic Data for Real-Domain High-Resolution Monocular Metric Depth Estimation/3be463cc-22b0-43d6-8e8c-f91c9fa34d5b_model.json new file mode 100644 index 0000000000000000000000000000000000000000..f3c6158d057bf37ed2813bbb32a928c3aff27ac2 --- /dev/null +++ b/2024/PatchRefiner_ Leveraging Synthetic Data for Real-Domain High-Resolution Monocular Metric Depth Estimation/3be463cc-22b0-43d6-8e8c-f91c9fa34d5b_model.json @@ -0,0 +1,2964 @@ +[ + [ + { + "type": "title", + "bbox": [ + 0.223, + 0.141, + 0.78, + 0.21 + ], + "angle": 0, + "content": "PatchRefiner: Leveraging Synthetic Data for Real-Domain High-Resolution Monocular Metric Depth Estimation" + }, + { + "type": "text", + "bbox": [ + 0.336, + 0.235, + 0.668, + 0.252 + ], + "angle": 0, + "content": "Zhenyu Li, Shariq Farooq Bhat, Peter Wonka" + }, + { + "type": "text", + "bbox": [ + 0.29, + 0.262, + 0.713, + 0.292 + ], + "angle": 0, + "content": "King Abdullah University of Science and Technology (KAUST) zhenyu.li.1@kaust.edu.sa" + }, + { + "type": "text", + "bbox": [ + 0.262, + 0.327, + 0.744, + 0.59 + ], + "angle": 0, + "content": "Abstract. This paper introduces PatchRefiner, an advanced framework for metric single image depth estimation aimed at high-resolution real-domain inputs. While depth estimation is crucial for applications such as autonomous driving, 3D generative modeling, and 3D reconstruction, achieving accurate high-resolution depth in real-world scenarios is challenging due to the constraints of existing architectures and the scarcity of detailed real-world depth data. PatchRefiner adopts a tile-based methodology, reconceptualizing high-resolution depth estimation as a refinement process, which results in notable performance enhancements. Utilizing a pseudo-labeling strategy that leverages synthetic data, PatchRefiner incorporates a Detail and Scale Disentangling (DSD) loss to enhance detail capture while maintaining scale accuracy, thus facilitating the effective transfer of knowledge from synthetic to real-world data. Our extensive evaluations demonstrate PatchRefiner's superior performance, significantly outperforming existing benchmarks on the Unreal4KStereo dataset by \\(18.1\\%\\) in terms of the root mean squared error (RMSE) and showing marked improvements in detail accuracy and consistent scale estimation on diverse real-world datasets like CityScape, ScanNet++, and ETH3D." + }, + { + "type": "text", + "bbox": [ + 0.262, + 0.604, + 0.741, + 0.618 + ], + "angle": 0, + "content": "Keywords: High-Resolution Metric Depth Estimation \\(\\cdot\\) Synthetic Data" + }, + { + "type": "title", + "bbox": [ + 0.216, + 0.643, + 0.377, + 0.659 + ], + "angle": 0, + "content": "1 Introduction" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.674, + 0.788, + 0.795 + ], + "angle": 0, + "content": "This paper delves into the field of metric single-image depth estimation, focusing on high-resolution inputs from the real domain. High-resolution depth estimation plays a key role in autonomous driving, augmented reality, content creation, and 3D reconstruction [2,11,30,59]. Despite significant progress, the high-resolution depth estimation in real-world scenarios remains daunting. This challenge is primarily due to the resolution limitations inherent in most state-of-the-art depth estimation architectures [2, 56, 61] and the scarcity of high-quality real-world depth datasets." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.796, + 0.788, + 0.842 + ], + "angle": 0, + "content": "The current state-of-the-art in high-resolution depth estimation, PatchFusion [27, 34, 43], employs a tile-based strategy to navigate the resolution constraints, posing the task as a fusion process of coarse and fine depth estimations." + } + ], + [ + { + "type": "page_number", + "bbox": [ + 0.218, + 0.116, + 0.23, + 0.127 + ], + "angle": 0, + "content": "2" + }, + { + "type": "header", + "bbox": [ + 0.271, + 0.115, + 0.579, + 0.13 + ], + "angle": 0, + "content": "Zhenyu Li, Shariq Farooq Bhat, Peter Wonka" + }, + { + "type": "image", + "bbox": [ + 0.225, + 0.146, + 0.312, + 0.233 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.232, + 0.239, + 0.315, + 0.252 + ], + "angle": 0, + "content": "(a) LR Depth" + }, + { + "type": "image", + "bbox": [ + 0.342, + 0.147, + 0.557, + 0.234 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.365, + 0.239, + 0.536, + 0.251 + ], + "angle": 0, + "content": "(b) Fusion-Based Framework" + }, + { + "type": "image", + "bbox": [ + 0.593, + 0.147, + 0.764, + 0.233 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.589, + 0.239, + 0.761, + 0.251 + ], + "angle": 0, + "content": "(c) Refiner-Based Framework" + }, + { + "type": "image_caption", + "bbox": [ + 0.214, + 0.263, + 0.789, + 0.319 + ], + "angle": 0, + "content": "Fig. 1: Framework Comparison. (a) Low resolution depth estimation framework with single forward pass. (b) Fusion-based high-resolution framework combining the best of coarse and fine depth predictions [27, 34]. (c) Our refiner-based framework predicts a residual to refine the coarse prediction." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.351, + 0.789, + 0.457 + ], + "angle": 0, + "content": "Because of the scarcity of real high-resolution depth datasets, PatchFusion resorts to training on a synthetic 4K dataset [27, 51]. There are two limitations of PatchFusion that we would like to improve upon: 1) It utilizes a three-step training process, which is not only time-consuming and expensive but also risks the framework achieving the stage-wise local optima and limits the potential performance gains from end-to-end learning. 2) PatchFusion demonstrates poor generalization to the real domain." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.459, + 0.789, + 0.639 + ], + "angle": 0, + "content": "The poor synthetic-to-real generalization is a long-standing problem in general [12,53]. In the context of depth estimation, the difference in the metric scale and depth distributions of synthetic datasets and the real domain further exacerbates the domain shift. This results in particularly subpar scale accuracy of depth models on real data when trained on synthetic datasets. On the other hand, the real datasets are not only low resolution but also often have missing ground truth values due to sensor constraints, occlusion, etc (see Fig. 3 and Sec. 3.4). This adds to the inability of depth models to capture sharp details when trained on real datasets. Thus, one is confronted with a dilemma - Training on real datasets leads to good scale accuracy but poor high-frequency details, while training on synthetic datasets leads to sharper results but poor scale performance on real images." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.642, + 0.789, + 0.686 + ], + "angle": 0, + "content": "We introduce PatchRefiner, a novel framework that reformulates high-resolution depth estimation as a process of refining coarse depth. We propose improvements on two levels:" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.688, + 0.789, + 0.748 + ], + "angle": 0, + "content": "First, unlike direct depth regression approaches [27,39,43], PatchRefiner utilizes a frozen coarse depth model and enhances the quality by predicting residual depth for refinement. This approach not only streamlines model training but also markedly enhances performance." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.75, + 0.789, + 0.84 + ], + "angle": 0, + "content": "Second, we propose a method to exploit the best of both worlds to solve the above synthetic-real dilemma. We employ a teacher-student framework, leveraging the sharpness of synthetic data while learning the scale from real data. The teacher model, pre-trained on synthetic data, generates pseudo labels for real-domain training samples. Recognizing that these pseudo labels offer detailed features albeit with scale inaccuracies, we introduce the Detail and Scale Dis-" + } + ], + [ + { + "type": "header", + "bbox": [ + 0.643, + 0.115, + 0.732, + 0.128 + ], + "angle": 0, + "content": "PatchRefiner" + }, + { + "type": "page_number", + "bbox": [ + 0.775, + 0.117, + 0.785, + 0.127 + ], + "angle": 0, + "content": "3" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.147, + 0.784, + 0.207 + ], + "angle": 0, + "content": "entangling (DSD) loss. This loss integrates ranking supervision and scale-shift invariance, drawing inspiration from recent advances in relative depth estimation [5, 42, 54, 56]. It leads to a framework capable of delivering high-resolution depth estimates with both precise scale and sharp details in real-world settings." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.208, + 0.785, + 0.358 + ], + "angle": 0, + "content": "Our evaluation of PatchRefiner on the Unreal4KStereo synthetic dataset [51] demonstrates a substantial improvement over the current state-of-the-art, reducing RMSE by \\(18.1\\%\\) and REL by \\(15.7\\%\\). Further, we assess the framework's efficacy in leveraging synthetic data across diverse real-world datasets, including CityScape [9] (outdoor, stereo), ScanNet++ [58] (indoor, LiDAR and reconstruction), and ETH3D [46] (mixed, LiDAR). Our findings reveal notable enhancements in depth boundary delineation (e.g., a \\(19.2\\%\\) increase in boundary recall on CityScape) while maintaining accurate scale estimation, showcasing the framework's adaptability and effectiveness across varying settings and sensor technologies." + }, + { + "type": "title", + "bbox": [ + 0.216, + 0.384, + 0.388, + 0.399 + ], + "angle": 0, + "content": "2 Related Work" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.417, + 0.785, + 0.612 + ], + "angle": 0, + "content": "High-Resolution Monocular Metric Depth Estimation. Monocular depth estimation has achieved tremendous progress [1, 11, 13, 15, 24, 29, 30]. Current SOTA approaches often employ complex network architectures yet grapple with the limitations imposed by their input resolution [2, 56]. This stands in stark contrast to the advancements in modern imaging devices that capture images at increasing resolutions, and the growing demand among users for high-resolution depth estimation. Initial efforts to address this gap included the use of Guided Depth Super-Resolution (GDSR) [19,33,63,65] and Implicit Functions [6,35]. Recently, Tile-Based Methods have emerged as a potent strategy for high-resolution depth estimation [27,34,43], segmenting images into patches for individual processing before reassembling them into a comprehensive depth map. This paper extends the tile-based approach, seeking to elevate the quality of depth estimation further." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.614, + 0.786, + 0.794 + ], + "angle": 0, + "content": "Synthetic Data for Depth Estimation. The challenge of acquiring high-quality, real-domain data for high-resolution depth training has prompted the use of synthetic datasets [41]. Traditionally, synthetic data has been employed within unsupervised domain adaptation frameworks, utilizing labeled synthetic and unlabeled real-domain data to enhance depth estimation accuracy on real-world images [7,22,23,32,60,64]. Techniques vary from pixel-level style transfer and image translation to feature-level adversarial learning [7,23,60,64], with some methods integrating additional information such as stereo pairs or segmentation maps for enhanced adaptation [22,32,60]. Contrasting with these approaches, our work explores a practical scenario where labeled data from both synthetic and real domains are leveraged to improve real-world, high-resolution depth estimation, delving into a relatively underexplored application of synthetic data." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.796, + 0.787, + 0.84 + ], + "angle": 0, + "content": "Pseudo-Labeling for Depth Estimation. Pseudo-labeling, a cornerstone of semi-supervised learning, has been widely applied across various domains, including classification [3,17,40,44,50] and scene understanding [8,26,28,36,37,47,62," + } + ], + [ + { + "type": "page_number", + "bbox": [ + 0.218, + 0.116, + 0.23, + 0.127 + ], + "angle": 0, + "content": "4" + }, + { + "type": "header", + "bbox": [ + 0.271, + 0.115, + 0.579, + 0.129 + ], + "angle": 0, + "content": "Zhenyu Li, Shariq Farooq Bhat, Peter Wonka" + }, + { + "type": "image", + "bbox": [ + 0.249, + 0.147, + 0.756, + 0.295 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.214, + 0.306, + 0.789, + 0.378 + ], + "angle": 0, + "content": "Fig. 2: Architecture Illustration. PatchRefiner contains a pre-trained frozen coarse depth estimation model \\(\\mathcal{N}_c\\) and a refiner model \\(\\mathcal{N}_r\\) predicts residual depth map \\(\\mathcal{D}_r\\) to refine the coarse depth \\(\\mathcal{D}_c\\). The refiner contains one base depth model \\(\\mathcal{N}_d\\) that has the same architecture as \\(\\mathcal{N}_c\\), and a light-weight decoder to aggregate information and make the final prediction." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.406, + 0.79, + 0.557 + ], + "angle": 0, + "content": "66], to extrapolate knowledge from labeled data to unlabeled datasets. In depth estimation, pseudo-labeling often serves to provide supplementary supervision in unsupervised domain adaptation settings [32, 55, 57]. While recent state-of-the-art methods like Depth-Anything utilize pseudo-labeling to enhance model generalization [56], these techniques predominantly aim to refine or enhance the pseudo labels themselves. Our approach diverges by utilizing real-domain data with accurate depth labels, focusing on a novel Detail and Scale Disentangling loss. This loss mechanism uniquely leverages the detailed insights from pseudo labels to enrich real-domain depth estimation without compromising the scale accuracy derived from real ground-truth data." + }, + { + "type": "title", + "bbox": [ + 0.216, + 0.58, + 0.331, + 0.597 + ], + "angle": 0, + "content": "3 Method" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.612, + 0.794, + 0.675 + ], + "angle": 0, + "content": "In this section, we first present the overall PatchRefiner framework in Sec. 3.1. Then, we introduce the limitation of adopting real-domain data for high-resolution depth estimation 3.2 and the proposed teacher-student framework in Sec. 3.3 with the Detail and Scale Disentangling loss (DSD) in Sec. 3.4." + }, + { + "type": "title", + "bbox": [ + 0.215, + 0.695, + 0.469, + 0.71 + ], + "angle": 0, + "content": "3.1 PatchRefiner Framework" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.72, + 0.788, + 0.795 + ], + "angle": 0, + "content": "PatchRefiner follows the tile-based strategy to address the prohibitive memory and computational demands for high resolutions such as 4K [27, 34]. However, recognizing the limitations of existing models, we propose a simplified two-step approach for high-resolution depth estimation: (i) Coarse Scale-Aware Estimation, and (ii) Fine-Grained Depth Refinement, shown in Fig. 2." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.796, + 0.788, + 0.842 + ], + "angle": 0, + "content": "(i) Coarse Scale-Aware Estimation: The foundation of PatchRefiner is the Coarse Depth Estimation network, \\(\\mathcal{N}_c\\), which processes downsampled versions of the input images to produce a global depth prediction, \\(\\mathbf{D}_c\\). Similar to" + } + ], + [ + { + "type": "header", + "bbox": [ + 0.643, + 0.115, + 0.732, + 0.127 + ], + "angle": 0, + "content": "PatchRefiner" + }, + { + "type": "page_number", + "bbox": [ + 0.776, + 0.117, + 0.786, + 0.127 + ], + "angle": 0, + "content": "5" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.147, + 0.784, + 0.207 + ], + "angle": 0, + "content": "previous work [27, 34], this step is crucial for establishing a baseline depth map that captures the scene's overall structure and depth consistency, albeit without high-resolution details. Moreover, \\(\\mathcal{N}_c\\) can be an arbitrary depth estimation model and it is frozen after the first step of training." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.208, + 0.785, + 0.282 + ], + "angle": 0, + "content": "(ii) Depth Refinement Process: Different from the conventional approach of using a separate fine depth network and a fusion mechanism [27, 39], our framework introduces a unified refinement network, \\(\\mathcal{N}_r\\). This network is designed to refine the coarse depth map by focusing on the recovery of lost details and enhancing depth precision at a patch-wise level." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.283, + 0.785, + 0.374 + ], + "angle": 0, + "content": "The input to \\(\\mathcal{N}_r\\) is the cropped original image \\(I\\), which is processed by the base depth model \\(\\mathcal{N}_d\\) with the same architecture as \\(\\mathcal{N}_c\\) in the refiner module. Then, we collect \\(L\\)-level multi-scale immediate features from both \\(\\mathcal{N}_d\\) and \\(\\mathcal{N}_c\\), denoting them as \\(\\mathcal{F}_d = \\{f_d^i\\}_{i=1}^L\\) and \\(\\tilde{\\mathcal{F}}_c = \\{\\tilde{f}_c^i\\}_{i=1}^L\\), respectively. Following [27], we apply the roi [16] operation to fetch features of the corresponding cropped area as \\(\\tilde{f}_c^i = \\text{roi}(f_c^i)\\)." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.375, + 0.784, + 0.419 + ], + "angle": 0, + "content": "Next, we adopt a lightweight decoder to obtain high-resolution predictions. Given \\(\\mathcal{F}_d\\) and \\(\\bar{\\mathcal{F}}_c\\), we aggregate them with concatenation operators (Cat) following by simple convolutional blocks (CB):" + }, + { + "type": "equation", + "bbox": [ + 0.428, + 0.434, + 0.785, + 0.452 + ], + "angle": 0, + "content": "\\[\nf _ {r} ^ {i} = \\operatorname {C B} \\left(\\operatorname {C a t} \\left(f _ {d} ^ {i}, \\tilde {f} _ {c} ^ {i}\\right)\\right), \\tag {1}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.458, + 0.785, + 0.549 + ], + "angle": 0, + "content": "The output feature set \\(\\mathcal{F}_r = \\{f_r^i\\}_{i=1}^L\\) is then fed to a successive series of up-sampling layers [25], in order to construct the residual depth map \\(\\mathbf{D}_r\\) at the input resolution. Associated with skip-connections, they form our refiner decoder. Further details about the architecture with immediate feature shapes are described in the supplementary material. Finally, we calculate the final patchwise depth estimation as \\(\\mathbf{D} = \\mathsf{roi}(\\mathbf{D}_c) + \\mathbf{D}_r\\)." + }, + { + "type": "title", + "bbox": [ + 0.215, + 0.572, + 0.713, + 0.587 + ], + "angle": 0, + "content": "3.2 Limitation of Real-Domain Depth Estimation Datasets" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.599, + 0.784, + 0.643 + ], + "angle": 0, + "content": "Our second goal in this paper is to train a real-domain high-resolution depth estimation model with both synthetic dataset \\(\\mathcal{S}\\) and real-domain dataset \\(\\mathcal{R}\\). Our main insight is to distinguish between scale errors and boundary errors." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.645, + 0.785, + 0.763 + ], + "angle": 0, + "content": "In the field of high-resolution depth estimation, the prevailing state-of-the-art methodology [27] trains models using synthetic datasets, which offers paired high-resolution images and corresponding dense high-resolution depth ground-truth maps [18,51]. This synthetic training regime, while beneficial in a controlled setting, introduces significant challenges when models are applied to real-world data due to the intrinsic domain gap between synthetic and real-world environments. This gap often manifests as substantial scale errors during inference on real-domain datasets as shown in Tab. 2." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.766, + 0.785, + 0.84 + ], + "angle": 0, + "content": "Addressing this issue directly by training on real-domain datasets presents its own set of challenges. Real-world high-resolution depth datasets are scarce and typically constrained by limited resolution and unavailable missing ground-truth pixels. These limitations stem from the methods employed in generating real-world depth annotations, such as Kinect [20, 48, 49], LiDAR [14, 46, 58], or" + } + ], + [ + { + "type": "page_number", + "bbox": [ + 0.218, + 0.116, + 0.23, + 0.127 + ], + "angle": 0, + "content": "6" + }, + { + "type": "header", + "bbox": [ + 0.272, + 0.115, + 0.579, + 0.129 + ], + "angle": 0, + "content": "Zhenyu Li, Shariq Farooq Bhat, Peter Wonka" + }, + { + "type": "image", + "bbox": [ + 0.219, + 0.144, + 0.506, + 0.345 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.508, + 0.144, + 0.785, + 0.346 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.214, + 0.355, + 0.788, + 0.397 + ], + "angle": 0, + "content": "Fig.3: Visualization of Real-Domain Data Pairs. Points lacking ground-truth data are depicted in gray. Due to sparse annotations near edges, models trained on real-domain data exhibit blurred boundary estimations." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.43, + 0.784, + 0.459 + ], + "angle": 0, + "content": "stereo vision techniques [9, 45], each with their inherent drawbacks as shown in Fig. 3." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.462, + 0.788, + 0.612 + ], + "angle": 0, + "content": "For instance, depth maps generated using Kinect technology are confined to a resolution of \\(640 \\times 480\\) [48, 49], insufficient for high-resolution depth estimation [27]. LiDAR-based depth maps, while useful, tend to be sparse [14, 46] or limited in resolution (e.g., 256x192 in Scannet++ [58]), and the process of reconstructing high-resolution dense depth maps from LiDAR point clouds is fragile with cascade errors and omissions [58]. Stereo vision techniques, on the other hand, can also introduce missing values around object boundaries due to rectification transformations [9, 45]. It is important to realize that existing high-resolution real datasets with missing values around edges do not help in reducing boundary errors." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.614, + 0.788, + 0.72 + ], + "angle": 0, + "content": "These limitations highlight the inherent difficulties in employing real-world datasets for training high-resolution depth estimators [27, 39]. The lack of high-quality, high-resolution ground-truth depth maps in the real domain makes it challenging to train models that can accurately predict sharp depth around fine object boundaries [41]. Our main idea is to devise a training strategy that can improve the scale errors when fine-tuning on real data, while maintaining high-resolution information around edges to minimize boundary errors." + }, + { + "type": "title", + "bbox": [ + 0.215, + 0.748, + 0.494, + 0.765 + ], + "angle": 0, + "content": "3.3 Overall Pipeline Illustration" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.78, + 0.787, + 0.842 + ], + "angle": 0, + "content": "Building on the success of semi-supervised learning frameworks in depth estimation tasks [32,56,57], our proposed framework employs a teacher-student architecture, as shown in Fig. 4, to effectively integrate synthetic and real-domain data for high-resolution depth estimation. It employs a two-step process:" + } + ], + [ + { + "type": "header", + "bbox": [ + 0.643, + 0.115, + 0.732, + 0.127 + ], + "angle": 0, + "content": "PatchRefiner" + }, + { + "type": "page_number", + "bbox": [ + 0.775, + 0.116, + 0.785, + 0.126 + ], + "angle": 0, + "content": "7" + }, + { + "type": "image", + "bbox": [ + 0.236, + 0.146, + 0.77, + 0.263 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.214, + 0.275, + 0.788, + 0.346 + ], + "angle": 0, + "content": "Fig. 4: Enhancing Real-Domain Learning with Synthetic Data. A teacher model trained on synthetic data produces pseudo labels for real-domain training. The student model benefits from a DSD dual-supervision approach: loss on pseudo labels for detail enhancement and loss on ground truth for scale accuracy. This method ensures detailed depth perception without compromising scale accuracy." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.376, + 0.784, + 0.422 + ], + "angle": 0, + "content": "Teacher Model Training: The teacher model is initially pretrained on a synthetic dataset \\( S \\), which, due to its high-quality and detailed annotations, enables the model to predict high-resolution depth maps with precise boundaries." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.423, + 0.785, + 0.572 + ], + "angle": 0, + "content": "Student Model Training with both Pseudo Labels and Ground-Truth Depth: In the subsequent phase, the teacher model is frozen, and the student model is trained on the real-domain dataset \\(\\mathcal{R}\\), utilizing both the ground truth depth labels \\(\\tilde{\\mathbf{D}}\\) and pseudo labels \\(\\hat{\\mathbf{D}}\\) generated by the teacher model. As identified in the limitations of real-domain datasets, while the ground truth labels offer accurate depth information, they miss crucial information near boundaries essential for high-resolution depth learning. To address this, the student model is guided by the teacher's pseudo labels, which excel near boundaries. However, these pseudo labels, while sharp, exhibit scale discrepancies due to the domain gap between synthetic and real-world data as shown in Tab. 2." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.574, + 0.785, + 0.635 + ], + "angle": 0, + "content": "Therefore, we introduce the Detail and Scale Disentangling (DSD) loss. It ensures that the final high-resolution depth predictions from the student model maintain an accurate scale while benefiting from the enhanced boundary details provided by the teacher model trained on the synthetic data." + }, + { + "type": "title", + "bbox": [ + 0.215, + 0.66, + 0.557, + 0.675 + ], + "angle": 0, + "content": "3.4 Detail and Scale Disentangling Loss" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.689, + 0.784, + 0.748 + ], + "angle": 0, + "content": "The student model is subject to two sources of supervision. The primary one is the standard scale-invariant loss [2,11,27,30], \\(\\mathcal{L}_{\\text{silog}}\\), calculated with the ground truth (GT) data, which ensures reliable and scale-consistent depth estimation on the real-domain data." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.75, + 0.785, + 0.841 + ], + "angle": 0, + "content": "The secondary source of supervision comes from pseudo labels generated by the teacher model. These labels, while detailed in capturing boundaries and fine details, exhibit low scale accuracy. Directly applying conventional depth losses [11,24,31] could lead to an imbalance between enhancing detail and maintaining scale accuracy. To address this, motivated by current impressive progress in relative depth estimation [5,42,54,56], we adopt the ranking loss \\(\\mathcal{L}_{\\text{rank}}\\) [5,54]" + } + ], + [ + { + "type": "page_number", + "bbox": [ + 0.218, + 0.116, + 0.23, + 0.127 + ], + "angle": 0, + "content": "8" + }, + { + "type": "header", + "bbox": [ + 0.272, + 0.115, + 0.576, + 0.129 + ], + "angle": 0, + "content": "Zhenyu Li, Shariq Farooq Bhat, Peter Wonka" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.147, + 0.784, + 0.207 + ], + "angle": 0, + "content": "and the Scale and Shift Invariant losses \\(\\mathcal{L}_{s s i}\\) [42,56] to inject the detail information. Notably, The \\(\\mathcal{L}_{\\text {rank }}\\) only provides supervision on the prediction relationship whereas \\(\\mathcal{L}_{s s i}\\) ignores scale and shift. Both losses tackle the same challenge with a slightly different approach." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.208, + 0.785, + 0.3 + ], + "angle": 0, + "content": "Ranking loss [5, 54] is designed for sparse sets with ordinal data. This adaptation allows us to leverage dense detailed ranking information from the pseudo labels, enhancing high-resolution estimation without compromising scale accuracy. Specifically, for a certain pair of points \\([p_{i,1}, p_{i,2}]\\) with predicted depth values \\([d_{i,1}, d_{i,2}]\\) and pseudo depth values \\([\\hat{d}_{i,1}, \\hat{d}_{i,2}]\\) in a set of \\(N\\) sampled point pairs \\(\\mathcal{P} = \\{[p_{i,1}, p_{i,2}], i = 1, 2, \\dots, N\\}\\), the ranking loss is defined as" + }, + { + "type": "equation", + "bbox": [ + 0.407, + 0.312, + 0.785, + 0.345 + ], + "angle": 0, + "content": "\\[\n\\mathcal {L} _ {\\text {r a n k}} = \\frac {1}{N} \\sum_ {i} \\phi \\left(p _ {i, 1}, p _ {i, 2}\\right), \\tag {2}\n\\]" + }, + { + "type": "equation", + "bbox": [ + 0.304, + 0.354, + 0.784, + 0.394 + ], + "angle": 0, + "content": "\\[\n\\phi \\left(p _ {i, 1}, p _ {i, 2}\\right) = \\left\\{ \\begin{array}{l l} \\log \\left(1 + \\exp \\left(- \\ell \\times \\left(d _ {i, 1} - d _ {i, 2}\\right)\\right)\\right), & \\ell \\neq 0, \\\\ \\left(d _ {i, 1} - d _ {i, 2}\\right) ^ {2}, & \\ell = 0, \\end{array} \\right. \\tag {3}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.215, + 0.399, + 0.658, + 0.414 + ], + "angle": 0, + "content": "where \\(\\ell\\) is the pseudo ordinal label, which can be induced by:" + }, + { + "type": "equation", + "bbox": [ + 0.393, + 0.426, + 0.784, + 0.48 + ], + "angle": 0, + "content": "\\[\n\\ell = \\left\\{ \\begin{array}{l l} + 1, & \\hat {d} _ {i, 1} / \\hat {d} _ {i, 2} > = 1 + \\tau , \\\\ - 1, & \\hat {d} _ {i, 1} / \\hat {d} _ {i, 2} < = \\frac {1}{1 + \\tau}, \\\\ 0, & \\text {o t h e r w i s e .} \\end{array} \\right. \\tag {4}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.491, + 0.784, + 0.55 + ], + "angle": 0, + "content": "Here \\(\\tau\\) is a tolerance threshold, which is set to 0.03 in our experiments following [54]. This loss not only encourages the predicted depth values for closely related points to align but also emphasizes the differentiation of points when the pseudo depth value of the two points is different." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.552, + 0.784, + 0.581 + ], + "angle": 0, + "content": "Scale and Shift Invariant loss [42, 56] is proposed to learn relative depth estimation with ignoring the unknown scale and shift of each sample:" + }, + { + "type": "equation", + "bbox": [ + 0.413, + 0.594, + 0.784, + 0.633 + ], + "angle": 0, + "content": "\\[\n\\mathcal {L} _ {s s i} = \\frac {1}{M} \\sum_ {i = 1} ^ {M} \\rho \\left(d _ {i} ^ {*} - \\hat {d} _ {i} ^ {*}\\right), \\tag {5}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.215, + 0.642, + 0.784, + 0.689 + ], + "angle": 0, + "content": "where \\( d_i^* \\) and \\( \\hat{d}_i^* \\) are scaled and shifted versions of the predicted depth \\( d_i \\) and pseudo label \\( \\hat{d}_i \\), and \\( \\rho \\) is the mean absolute error loss. \\( M \\) is the number of pixels. We use the least-squares criterion to align the prediction to the ground truth:" + }, + { + "type": "equation", + "bbox": [ + 0.377, + 0.701, + 0.784, + 0.74 + ], + "angle": 0, + "content": "\\[\n(s, t) = \\operatorname {a r g m i n} _ {s, t} \\sum_ {i = 1} ^ {M} \\left(s d _ {i} + t - \\hat {d} _ {i}\\right) ^ {2}, \\tag {6}\n\\]" + }, + { + "type": "equation", + "bbox": [ + 0.433, + 0.757, + 0.784, + 0.773 + ], + "angle": 0, + "content": "\\[\nd ^ {*} = s d + t, \\hat {d} ^ {*} = \\hat {d} \\tag {7}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.78, + 0.784, + 0.809 + ], + "angle": 0, + "content": "where the scale \\( s \\) and shift \\( t \\) factors are effectively determined with the closed form [42]." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.81, + 0.784, + 0.841 + ], + "angle": 0, + "content": "The final Detail and Scale Disentangling loss for training the student model combines the \\(\\mathcal{L}_{\\text{silog}}\\), \\(\\mathcal{L}_{\\text{rank}}\\), and \\(\\mathcal{L}_{\\text{ssi}}\\) as follows:" + } + ], + [ + { + "type": "header", + "bbox": [ + 0.643, + 0.115, + 0.732, + 0.127 + ], + "angle": 0, + "content": "PatchRefiner" + }, + { + "type": "page_number", + "bbox": [ + 0.775, + 0.116, + 0.785, + 0.127 + ], + "angle": 0, + "content": "9" + }, + { + "type": "image", + "bbox": [ + 0.217, + 0.144, + 0.359, + 0.206 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.218, + 0.207, + 0.358, + 0.268 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.218, + 0.269, + 0.358, + 0.331 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.218, + 0.332, + 0.358, + 0.394 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.244, + 0.397, + 0.333, + 0.412 + ], + "angle": 0, + "content": "ZoeDepth [2]" + }, + { + "type": "image", + "bbox": [ + 0.36, + 0.144, + 0.499, + 0.205 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.36, + 0.207, + 0.5, + 0.269 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.36, + 0.27, + 0.5, + 0.331 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.36, + 0.332, + 0.5, + 0.394 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.375, + 0.397, + 0.487, + 0.413 + ], + "angle": 0, + "content": "PatchFusion [27]" + }, + { + "type": "image", + "bbox": [ + 0.503, + 0.144, + 0.642, + 0.205 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.503, + 0.207, + 0.642, + 0.269 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.503, + 0.27, + 0.642, + 0.331 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.503, + 0.332, + 0.642, + 0.394 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.556, + 0.398, + 0.59, + 0.411 + ], + "angle": 0, + "content": "Ours" + }, + { + "type": "image", + "bbox": [ + 0.643, + 0.144, + 0.785, + 0.205 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.643, + 0.207, + 0.785, + 0.269 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.643, + 0.27, + 0.785, + 0.331 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.643, + 0.332, + 0.785, + 0.394 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.68, + 0.398, + 0.75, + 0.412 + ], + "angle": 0, + "content": "GT, Input" + }, + { + "type": "image_caption", + "bbox": [ + 0.215, + 0.424, + 0.788, + 0.493 + ], + "angle": 0, + "content": "Fig. 5: Qualitative Comparison on UnrealStereo4K. We show the depth prediction and corresponding error map, respectively. The qualitative comparisons showcased here indicate our framework outperforms counterparts [2, 27] with sharper edges and lower error around boundaries. We show individual patches in all images to emphasize details near depth boundaries." + }, + { + "type": "equation", + "bbox": [ + 0.373, + 0.536, + 0.785, + 0.572 + ], + "angle": 0, + "content": "\\[\n\\mathcal {L} _ {D S D} = \\overbrace {\\mathcal {L} _ {s i l o g}} ^ {\\mathcal {L} _ {g t}} + \\overbrace {\\lambda_ {1} \\mathcal {L} _ {r a n k} + \\lambda_ {2} \\mathcal {L} _ {s s i}} ^ {\\mathcal {L} _ {p l}}, \\tag {8}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.215, + 0.578, + 0.788, + 0.624 + ], + "angle": 0, + "content": "where \\(\\lambda_{1}\\) and \\(\\lambda_{2}\\) are two balancing factors, both empirically set to 0.1 in our experiments, respectively. The terms \\(\\mathcal{L}_{gt}\\) and \\(\\mathcal{L}_{pl}\\) represent the supervision signals derived from the ground truth and pseudo labels, respectively." + }, + { + "type": "title", + "bbox": [ + 0.216, + 0.647, + 0.377, + 0.664 + ], + "angle": 0, + "content": "4 Experiments" + }, + { + "type": "title", + "bbox": [ + 0.216, + 0.679, + 0.44, + 0.693 + ], + "angle": 0, + "content": "4.1 Datasets and Metrics" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.704, + 0.788, + 0.81 + ], + "angle": 0, + "content": "UnrealStereo4K (Synthetic): The UnrealStereo4K dataset [51] offers synthetic stereo images at a 4K resolution \\((2160\\times 3840)\\), each paired with accurate, boundary-complete pixel-wise ground truth. Following the procedure in [27], we exclude mislabeled images using the Structural Similarity Index (SSIM) [52] and compute ground truth depth maps from provided disparity maps using camera parameters. Adhering to the dataset splits in [27, 51], we employ a default patch size of \\(540\\times 960\\) for compatibility with [27]." + }, + { + "type": "text", + "bbox": [ + 0.215, + 0.81, + 0.788, + 0.84 + ], + "angle": 0, + "content": "CityScapes (Real, Stereo): Cityscapes [9] offers a comprehensive suite of urban scene images, segmentation masks, and disparity maps at a resolution" + } + ], + [ + { + "type": "page_number", + "bbox": [ + 0.218, + 0.116, + 0.236, + 0.127 + ], + "angle": 0, + "content": "10" + }, + { + "type": "header", + "bbox": [ + 0.272, + 0.115, + 0.576, + 0.129 + ], + "angle": 0, + "content": "Zhenyu Li, Shariq Farooq Bhat, Peter Wonka" + }, + { + "type": "image", + "bbox": [ + 0.221, + 0.144, + 0.332, + 0.198 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.221, + 0.199, + 0.332, + 0.254 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.221, + 0.255, + 0.332, + 0.31 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.221, + 0.311, + 0.332, + 0.365 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.334, + 0.145, + 0.444, + 0.199 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.334, + 0.2, + 0.444, + 0.254 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.334, + 0.255, + 0.444, + 0.31 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.334, + 0.311, + 0.444, + 0.365 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.313, + 0.369, + 0.356, + 0.383 + ], + "angle": 0, + "content": "Image" + }, + { + "type": "image", + "bbox": [ + 0.447, + 0.144, + 0.557, + 0.199 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.447, + 0.2, + 0.557, + 0.254 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.447, + 0.255, + 0.557, + 0.31 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.447, + 0.311, + 0.557, + 0.365 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.457, + 0.369, + 0.545, + 0.383 + ], + "angle": 0, + "content": "ZoeDepth [2]" + }, + { + "type": "image", + "bbox": [ + 0.559, + 0.144, + 0.669, + 0.199 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.559, + 0.2, + 0.669, + 0.254 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.559, + 0.255, + 0.669, + 0.31 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.559, + 0.311, + 0.669, + 0.365 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.593, + 0.369, + 0.634, + 0.381 + ], + "angle": 0, + "content": "PR R" + }, + { + "type": "image", + "bbox": [ + 0.671, + 0.145, + 0.781, + 0.199 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.671, + 0.2, + 0.781, + 0.254 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.671, + 0.255, + 0.781, + 0.31 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.671, + 0.311, + 0.781, + 0.365 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.703, + 0.369, + 0.748, + 0.381 + ], + "angle": 0, + "content": "Ours 1" + }, + { + "type": "image_caption", + "bbox": [ + 0.214, + 0.395, + 0.788, + 0.464 + ], + "angle": 0, + "content": "Fig. 6: Qualitative Comparison on ETH3D and ScanNet++. The first two rows depict results for ETH3D and the last two for ScanNet++. The baseline ZoeDepth and PatchRefiner with conventional fine tuning (PR \\(\\mathcal{R}\\)) both struggle to create high-resolution depth details near boundaries. Our proposed strategy yields crisp boundaries on both datasets." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.507, + 0.785, + 0.552 + ], + "angle": 0, + "content": "of \\(1024 \\times 2048\\), surpassing many real-domain datasets in density, quantity, and resolution [45, 46, 48, 49]. We use a standard patch size of \\(256 \\times 512\\) and conduct most experiments on this dataset." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.557, + 0.788, + 0.649 + ], + "angle": 0, + "content": "ScanNet++ (Real, LiDAR, Reconstruction): ScanNet++ [58] is an extensive indoor dataset providing high-resolution images \\((1440\\times 1920)\\), depth maps from iPhone LiDAR \\((192\\times 256)\\), and high-resolution depth maps sampled from reconstructions of laser scans \\((1440\\times 1920)\\). Our chosen patch size is \\(720\\times 960\\). We use the low-resolution ground-truth for training since the high-resolution version contains much noise as shown in Fig. 3." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.653, + 0.788, + 0.713 + ], + "angle": 0, + "content": "ETH3D (Real, LiDAR): The ETH3D benchmark [46] provides high-resolution indoor and outdoor images \\((6048\\times 4032)\\) with ground-truth depth from laser sensors. We downsample the image-depth pairs to \\(2160\\times 3840\\) and select a patch size of \\(540\\times 960\\)." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.72, + 0.788, + 0.84 + ], + "angle": 0, + "content": "Metrics: We adopt standard depth evaluation metrics from [2, 11, 38] and the Soft Edge Error (SEE) from [4, 27, 51] for scale evaluation. Additional metric details are in the supplementary material. Given real-world depth map boundary incompleteness, we design metrics to focus on evaluating boundary accuracy, leveraging high-resolution fine-grained segmentation masks as depth boundary quality proxies. Specifically, we use the Sobel operator [21] on both predicted depth and segmentation maps to generate edge masks, and then compute Precision, Recall, and F1 Score." + } + ], + [ + { + "type": "header", + "bbox": [ + 0.643, + 0.115, + 0.732, + 0.127 + ], + "angle": 0, + "content": "PatchRefiner" + }, + { + "type": "page_number", + "bbox": [ + 0.769, + 0.116, + 0.785, + 0.127 + ], + "angle": 0, + "content": "11" + }, + { + "type": "image", + "bbox": [ + 0.221, + 0.144, + 0.332, + 0.424 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.335, + 0.144, + 0.444, + 0.424 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.447, + 0.144, + 0.557, + 0.424 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.559, + 0.144, + 0.668, + 0.424 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.67, + 0.144, + 0.781, + 0.424 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.214, + 0.435, + 0.788, + 0.507 + ], + "angle": 0, + "content": "Fig. 7: Qualitative Comparison on CityScapes. This figure illustrates depth estimation comparisons between the base ZoeDepth model, PatchRefiner (PR) trained on CityScapes, and our method. We display outcomes under varying levels of \\(\\mathcal{L}_{pl}\\) supervision (\\(\\lambda_1 = \\lambda_2 = 1e^{-1}\\) or 1), featuring zoomed-in sections of each image to highlight detail fidelity near depth discontinuities." + }, + { + "type": "title", + "bbox": [ + 0.216, + 0.541, + 0.456, + 0.556 + ], + "angle": 0, + "content": "4.2 Implementation Details" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.581, + 0.788, + 0.732 + ], + "angle": 0, + "content": "PatchRefiner on Synthetic Dataset: For training on synthetic datasets, we employ the scale-invariant log loss \\(\\mathcal{L}_{\\text{silog}}\\), as outlined in [2, 11, 27, 30]. Initialization of the coarse network \\(\\mathcal{N}_c\\) leverages pretrained weights from the NYU-v2 dataset [48], adhering to the approach in [27]. We dedicate 24 epochs to training \\(\\mathcal{N}_c\\), which is subsequently frozen to ensure stability in subsequent training phases. For the refinement network \\(\\mathcal{N}_r\\), initialization employs \\(\\mathcal{N}_c\\)'s parameters, and training extends for an additional 36 epochs. Standard augmentation strategies from the baseline depth model are incorporated to enhance training effectiveness. During inference, we implement Consistency-Aware Inference, as described in [27], to optimize performance." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.735, + 0.788, + 0.842 + ], + "angle": 0, + "content": "Learning on Real-Domain Dataset: Since the coarse model within the framework offers scale-consistent predictions, we first pretrain a coarse model on real-domain data to establish a reliable depth scale foundation. This model is subsequently frozen to preserve scale consistency throughout the training process. The refiner model in PatchRefiner is then initialized with parameters pretrained on synthetic data, enabling it to maintain high-frequency detail knowledge acquired from the synthetic domain. Initially, the student model is trained solely" + } + ], + [ + { + "type": "page_number", + "bbox": [ + 0.218, + 0.116, + 0.236, + 0.127 + ], + "angle": 0, + "content": "12" + }, + { + "type": "header", + "bbox": [ + 0.272, + 0.115, + 0.575, + 0.129 + ], + "angle": 0, + "content": "Zhenyu Li, Shariq Farooq Bhat, Peter Wonka" + }, + { + "type": "table_caption", + "bbox": [ + 0.216, + 0.145, + 0.784, + 0.187 + ], + "angle": 0, + "content": "Table 1: Quantitative comparison on UnrealStereo4K. We color code the corresponding best competitor and our method within each block. PF and PR are short for PatchFusion [27] and PatchRefiner, respectively. The reported numbers are from [27]." + }, + { + "type": "table", + "bbox": [ + 0.227, + 0.199, + 0.777, + 0.455 + ], + "angle": 0, + "content": "
Methodδ1(%)↑REL↓RMS↓SiLog↓SEE↓Reference
iDisc [38]96.9400.0531.4048.5021.070ICCV 2023
SMD-Net [51]97.7740.0441.2827.3890.883CVPR 2021
Graph-GDSR [10]97.9320.0441.2647.4690.872CVPR 2022
BoostingDepth [34]98.1040.0441.1236.6620.939CVPR 2021
ZoeDepth [2]97.7170.0461.2897.4480.914-
ZoeDepth+PFp=16 [27]98.4190.0401.0886.2120.838CVPR 2024
ZoeDepth+PFp=49 [27]98.4500.0391.0756.1310.846
ZoeDepth+PFR=128 [27]98.4690.0391.0666.0850.849
ZoeDepth+PRp=1698.8210.0330.8925.4170.750Ours
ZoeDepth+PRp=4998.8590.0330.8705.3190.751
ZoeDepth+PRR=12898.8640.0330.8725.3770.738
Depth-Anything [56]97.7730.0411.2357.1920.911CVPR 2024
Depth-Anything+PFp=16 [27]98.5580.0361.0155.8830.811CVPR 2024
Depth-Anything+PFp=49 [27]98.6070.0350.9875.7460.812
Depth-Anything+PFR=128 [27]98.6160.0350.9845.7750.813
Depth-Anything+PRp=1698.8260.0330.8895.2890.768Ours
Depth-Anything+PRp=4998.8780.0330.8605.1490.767
Depth-Anything+PRR=12898.8780.0330.8605.2060.759
" + }, + { + "type": "table_caption", + "bbox": [ + 0.216, + 0.468, + 0.784, + 0.509 + ], + "angle": 0, + "content": "Table 2: Quantitative comparison on CityScapes. FT, and MIX are short for fine-tuning and mixed-data strategies, which are our main competitors. Baseline is highlighted in gray and each of the three competitors in a different color." + }, + { + "type": "table", + "bbox": [ + 0.232, + 0.522, + 0.772, + 0.643 + ], + "angle": 0, + "content": "
MethodDataScaleBoundary
SRδ1(%)↑REL↓RMS↓Precision↑Recall↑
ZoeDepth [2]✓94.5020.0704.40613.3237.59
ZoeDepth [2] + FT✓✓94.4980.0714.41812.9337.89
PatchRefiner (zero-shot)✓5.7050.39912.20328.6851.28
PatchRefiner✓95.2840.0664.04716.6739.53
PatchRefiner + FT✓✓95.4180.0653.99217.0940.92
PatchRefiner + mix [39]✓✓89.1080.1124.73223.0841.17
PatchRefiner + DSD✓✓95.3590.0663.98218.9248.78
" + }, + { + "type": "text", + "bbox": [ + 0.216, + 0.671, + 0.784, + 0.701 + ], + "angle": 0, + "content": "with \\(\\mathcal{L}_{\\text{silog}}\\) for 24 epochs. Subsequent fine-tuning with our Detail and Scale Disentangling loss \\(\\mathcal{L}_{DSD}\\) over an additional 6 epochs to refine depth estimations." + }, + { + "type": "title", + "bbox": [ + 0.216, + 0.724, + 0.373, + 0.737 + ], + "angle": 0, + "content": "4.3 Main Results" + }, + { + "type": "text", + "bbox": [ + 0.216, + 0.75, + 0.784, + 0.84 + ], + "angle": 0, + "content": "Synthetic Dataset: On the UnrealStereo4K dataset, PatchRefiner not only outperforms the base depth model but also shows substantial improvements over PatchFusion, reducing RMSE by \\(18.1\\%\\) and REL by \\(15.7\\%\\). This advancement is further underscored by achieving the lowest SEE, highlighting our model's proficiency in capturing edge details. Qualitative comparisons, illustrated in Fig. 5, indicate the superior boundary delineation achieved by PatchRefiner." + } + ], + [ + { + "type": "header", + "bbox": [ + 0.643, + 0.115, + 0.732, + 0.128 + ], + "angle": 0, + "content": "PatchRefiner" + }, + { + "type": "page_number", + "bbox": [ + 0.769, + 0.116, + 0.786, + 0.127 + ], + "angle": 0, + "content": "13" + }, + { + "type": "table_caption", + "bbox": [ + 0.216, + 0.145, + 0.785, + 0.174 + ], + "angle": 0, + "content": "Table 3: Ablation study of architecture variations and formulation of final depth prediction. Ours is highlighted with color. We highlight the best in bold." + }, + { + "type": "table", + "bbox": [ + 0.223, + 0.185, + 0.781, + 0.317 + ], + "angle": 0, + "content": "
MethodType of outputFeature Levelsδ1(%)↑REL↓RMS↓SiLog↓SEE↓
PatchFusion [27]Direct6 features98.4190.0401.0886.2120.838
PatchRefinerDcresidual1 features98.7340.0340.9265.5500.782
Dcresidual2 features98.8140.0330.9055.5110.750
Dcresidual3 features98.8150.0340.9005.5830.753
Dcresidual4 features98.8150.0330.8995.4940.752
Dcresidual5 features98.8140.0330.8945.4680.752
Dcresidual6 features98.8210.0330.8925.4170.750
Ddresidual6 features98.8040.0330.8995.4480.753
Direct6 features98.7490.0340.9255.5910.765
" + }, + { + "type": "table_caption", + "bbox": [ + 0.214, + 0.329, + 0.784, + 0.384 + ], + "angle": 0, + "content": "Table 4: Variations of \\(\\mathcal{L}_{pl}\\). We analyse various options for \\(\\mathcal{L}_{pl}\\) and compare them against PR. \\(\\mathcal{S}\\) and PR. \\(\\mathcal{R}\\) that serve as baselines for training on Synthetic and Real data individually. We set \\(\\lambda = \\lambda_1 = \\lambda_2\\) to analyze the influence of DSD weight. The highlighted result is achieved with \\(\\lambda = 1e^{-1}\\)." + }, + { + "type": "table", + "bbox": [ + 0.223, + 0.398, + 0.776, + 0.563 + ], + "angle": 0, + "content": "
Variations (Lpl)ScaleBoundary
δ1(%)↑REL↓RMS↓Precision↑Recall↑F1↑
Baseline PR. S5.7050.39912.2028.6851.2836.34
Baseline PR. R95.4180.0653.99217.0940.9223.68
Lsilog82.5500.1555.91426.3458.3535.81
Lsilog + mask81.4250.1486.51330.0446.8936.19
Lrank95.4130.0663.97318.3847.8126.12
Lssi95.4650.0653.97419.2644.0126.39
Lrank + Lssi (Ours)95.3590.0663.98218.9248.7826.84
λ = 195.0770.0694.22221.3258.4830.90
λ = 3e-195.2960.0684.08620.1153.8328.91
λ = 3e-295.4620.0653.95318.0143.9025.11
" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.585, + 0.785, + 0.734 + ], + "angle": 0, + "content": "Real-Domain Dataset: Tab. 2 and Fig. 6, 7 delineate the performance disparity when leveraging synthetic data for real-domain learning. Although the synthetic-trained model excels in boundary details, it fails in scale accuracy due to the domain gap. Sole training on real-domain data enhances baseline's scale prediction yet falls short in detail accuracy due to the missing depth ground truth around boundaries. Neither fine-tuning nor mixed-training substantially elevates performance across scale and detail metrics, reflecting the inherent challenges in our task. Contrastingly, our strategy propels the model to notable gains in boundary accuracy (19.2% increase in boundary recall) while sustaining scale precision comparable to the baseline model." + }, + { + "type": "title", + "bbox": [ + 0.216, + 0.756, + 0.528, + 0.77 + ], + "angle": 0, + "content": "4.4 Ablation Studies and Discussion" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.78, + 0.784, + 0.84 + ], + "angle": 0, + "content": "We ablate and discuss the contributions of individual components. As default, we utilize the UnrealStereo4K dataset for synthetic-dataset training and the PatchRefiner variant with \\( P = 16 \\) patches for clarity and ease of comparison. We ablate our teacher-student framework using the CityScapes dataset." + } + ], + [ + { + "type": "page_number", + "bbox": [ + 0.218, + 0.116, + 0.236, + 0.127 + ], + "angle": 0, + "content": "14" + }, + { + "type": "header", + "bbox": [ + 0.272, + 0.115, + 0.575, + 0.129 + ], + "angle": 0, + "content": "Zhenyu Li, Shariq Farooq Bhat, Peter Wonka" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.147, + 0.788, + 0.253 + ], + "angle": 0, + "content": "Architecture and Formulation Variations: We first ablate the effectiveness of our architecture design. Utilizing a ZoeDepth model, we extract six levels of intermediate features during the forward pass, represented as \\(\\mathcal{F} = f_1, f_2, \\dots, f_6\\), with progressively increasing resolution. We then sequentially omit lower resolution feature maps from the refiner's decoder input. As depicted in Tab. 3, performance diminishes with fewer feature maps, yet even with solely the highest resolution feature map, our model outperforms PatchFusion." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.253, + 0.788, + 0.344 + ], + "angle": 0, + "content": "Further, we explore our refinement formulation's efficacy by comparing it against two alternatives: (1) Residual depth adjustment based on the base model \\(\\mathcal{N}_d\\) within the refiner. (2) Direct depth prediction from the refiner's decoder, similar to PatchFusion (\\(\\mathbf{D} = \\mathbf{D}_r\\) in this context). Our residual-based approach demonstrates superior performance, underscoring its advantages in optimization and training efficiency." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.363, + 0.788, + 0.53 + ], + "angle": 0, + "content": "Effectiveness of Pseudo Label Supervision: We evaluate the efficacy of our Detail and Scale Disentangling (DSD) loss against the conventional scale-invariant loss, \\(\\mathcal{L}_{silog}\\), for pseudo label supervision. Tab. 4 illustrates that while \\(\\mathcal{L}_{silog}\\) aids in detail transfer from the teacher to the student model, it compromises scale accuracy due to significant discrepancies in the pseudo labels. A masking approach, focusing only on areas lacking depth, does not mitigate this issue, indicating pervasive negative effects. In contrast, the combination of \\(\\mathcal{L}_{rank}\\) and \\(\\mathcal{L}_{ssi}\\) within \\(\\mathcal{L}_{pl}\\) not only improves detail fidelity but also maintains scale accuracy, demonstrating that ranking constraints and scale-shift invariance are effective, orthogonal strategies for enhancing high-resolution detail without sacrificing scale accuracy." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.549, + 0.788, + 0.673 + ], + "angle": 0, + "content": "Exploration of Variant DSD Loss Weight: This study examines the impact of different DSD loss weights \\((\\lambda_{1}\\) and \\(\\lambda_{2})\\) to elucidate the loss's efficiency. As shown in Tab. 4 and Fig. 7, increasing the loss weight marginally affects scale accuracy while significantly improving boundaries, validating the DSD loss's role in balancing detail enhancement and scale preservation. Notably, when both weights are set to 1, the model's boundary recall surpasses the teacher's performance. Furthermore, when DSD loss achieves a comparable boundary metric to \\(\\mathcal{L}_{silog}\\), it exhibits a smaller performance decline, underscoring its effectiveness." + }, + { + "type": "title", + "bbox": [ + 0.215, + 0.69, + 0.36, + 0.708 + ], + "angle": 0, + "content": "5 Conclusion" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.72, + 0.788, + 0.841 + ], + "angle": 0, + "content": "We presented PatchRefiner, a tile-based framework tailored to real-world high-resolution monocular metric depth estimation. It reconceptualizes high-resolution depth estimation as a refinement process. With a pseudo-labeling strategy that leverages synthetic data, we propose a Detail and Scale Disentangling (DSD) loss to enhance detail capture while maintaining scale accuracy. Our proposed framework decisively surpasses the current SOTA method for UnrealStereo4K (17.3% in RMSE), while demonstrating marked improvements in detail accuracy on real-world datasets such as CityScape, ScanNet++, and ETH3D." + } + ], + [ + { + "type": "header", + "bbox": [ + 0.643, + 0.115, + 0.732, + 0.128 + ], + "angle": 0, + "content": "PatchRefiner" + }, + { + "type": "page_number", + "bbox": [ + 0.769, + 0.117, + 0.786, + 0.127 + ], + "angle": 0, + "content": "15" + }, + { + "type": "title", + "bbox": [ + 0.218, + 0.145, + 0.403, + 0.163 + ], + "angle": 0, + "content": "Acknowledgements" + }, + { + "type": "text", + "bbox": [ + 0.217, + 0.175, + 0.784, + 0.205 + ], + "angle": 0, + "content": "This publication is funded in part by KAUST under award #ORA-2023-5241 (for the NTGI-AI project)." + }, + { + "type": "title", + "bbox": [ + 0.218, + 0.226, + 0.323, + 0.241 + ], + "angle": 0, + "content": "References" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.254, + 0.785, + 0.281 + ], + "angle": 0, + "content": "1. Bhat, S.F., Alhashim, I., Wonka, P.: Adabins: Depth estimation using adaptive bins. In: CVPR. pp. 4009-4018 (2021) 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.283, + 0.785, + 0.322 + ], + "angle": 0, + "content": "2. Bhat, S.F., Birkl, R., Wofk, D., Wonka, P., Müller, M.: Zoedepth: Zero-shot transfer by combining relative and metric depth. arXiv preprint arXiv:2302.12288 (2023) 1, 3, 7, 9, 10, 11, 12" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.324, + 0.785, + 0.377 + ], + "angle": 0, + "content": "3. Chen, C., Xie, W., Huang, W., Rong, Y., Ding, X., Huang, Y., Xu, T., Huang, J.: Progressive feature alignment for unsupervised domain adaptation. In: Proceedings of the IEEE/CVF conference on computer vision and pattern recognition. pp. 627-636 (2019) 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.379, + 0.785, + 0.404 + ], + "angle": 0, + "content": "4. Chen, C., Chen, X., Cheng, H.: On the over-smoothing problem of cnn based disparity estimation. In: ICCV. pp. 8997-9005 (2019) 10" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.406, + 0.785, + 0.431 + ], + "angle": 0, + "content": "5. Chen, W., Fu, Z., Yang, D., Deng, J.: Single-image depth perception in the wild. NeurIPS 29 (2016) 3, 7, 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.433, + 0.785, + 0.459 + ], + "angle": 0, + "content": "6. Chen, Y., Liu, S., Wang, X.: Learning continuous image representation with local implicit image function. In: CVPR. pp. 8628-8638 (2021) 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.46, + 0.785, + 0.499 + ], + "angle": 0, + "content": "7. Chen, Y.C., Lin, Y.Y., Yang, M.H., Huang, J.B.: Crdoco: Pixel-level domain transfer with cross-domain consistency. In: Proceedings of the IEEE/CVF conference on computer vision and pattern recognition. pp. 1791-1800 (2019) 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.501, + 0.785, + 0.54 + ], + "angle": 0, + "content": "8. Chen, Z., Zhang, R., Zhang, G., Ma, Z., Lei, T.: Digging into pseudo label: a low-budget approach for semi-supervised semantic segmentation. IEEE Access 8, 41830-41837 (2020) 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.542, + 0.785, + 0.581 + ], + "angle": 0, + "content": "9. Cordts, M., Omran, M., Ramos, S., Rehfeld, T., Enzweiler, M., Benenson, R., Franke, U., Roth, S., Schiele, B.: The cityscapes dataset for semantic urban scene understanding. In: CVPR. pp. 3213-3223 (2016) 3, 6, 9" + }, + { + "type": "ref_text", + "bbox": [ + 0.22, + 0.583, + 0.785, + 0.622 + ], + "angle": 0, + "content": "10. De Lutio, R., Becker, A., D'Aronco, S., Russo, S., Wegner, J.D., Schindler, K.: Learning graph regularisation for guided super-resolution. In: CVPR. pp. 1979-1988 (2022) 12" + }, + { + "type": "ref_text", + "bbox": [ + 0.22, + 0.624, + 0.785, + 0.649 + ], + "angle": 0, + "content": "11. Eigen, D., Puhrsch, C., Fergus, R.: Depth map prediction from a single image using a multi-scale deep network. NeurIPS 27 (2014) 1, 3, 7, 10, 11" + }, + { + "type": "ref_text", + "bbox": [ + 0.22, + 0.651, + 0.785, + 0.69 + ], + "angle": 0, + "content": "12. Farahani, A., Voghoei, S., Rasheed, K., Arabnia, H.R.: A brief review of domain adaptation. Advances in data science and information engineering: proceedings from ICDATA 2020 and IKE 2020 pp. 877-894 (2021) 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.22, + 0.692, + 0.785, + 0.718 + ], + "angle": 0, + "content": "13. Fu, H., Gong, M., Wang, C., Batmanghelich, K., Tao, D.: Deep ordinal regression network for monocular depth estimation. In: CVPR. pp. 2002-2011 (2018) 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.22, + 0.719, + 0.785, + 0.744 + ], + "angle": 0, + "content": "14. Geiger, A., Lenz, P., Urtasun, R.: Are we ready for autonomous driving? the kitti vision benchmark suite. In: CVPR. pp. 3354-3361. IEEE (2012) 5, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.22, + 0.746, + 0.785, + 0.771 + ], + "angle": 0, + "content": "15. Godard, C., Mac Aodha, O., Firman, M., Brostow, G.J.: Digging into self-supervised monocular depth estimation. In: ICCV. pp. 3828-3838 (2019) 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.22, + 0.773, + 0.785, + 0.799 + ], + "angle": 0, + "content": "16. He, K., Gkioxari, G., Dollár, P., Girshick, R.: Mask r-cnn. In: Proceedings of the IEEE international conference on computer vision. pp. 2961-2969 (2017) 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.22, + 0.8, + 0.785, + 0.84 + ], + "angle": 0, + "content": "17. Hu, Z., Yang, Z., Hu, X., Nevatia, R.: Simple: Similar pseudo label exploitation for semi-supervised classification. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 15099-15108 (2021) 3" + }, + { + "type": "list", + "bbox": [ + 0.22, + 0.254, + 0.785, + 0.84 + ], + "angle": 0, + "content": null + } + ], + [ + { + "type": "page_number", + "bbox": [ + 0.218, + 0.116, + 0.236, + 0.127 + ], + "angle": 0, + "content": "16" + }, + { + "type": "header", + "bbox": [ + 0.272, + 0.115, + 0.576, + 0.129 + ], + "angle": 0, + "content": "Zhenyu Li, Shariq Farooq Bhat, Peter Wonka" + }, + { + "type": "ref_text", + "bbox": [ + 0.218, + 0.147, + 0.785, + 0.175 + ], + "angle": 0, + "content": "18. Huang, P.H., Matzen, K., Kopf, J., Ahuja, N., Huang, J.B.: Deepmvs: Learning multi-view stereopsis. In: CVPR (2018) 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.177, + 0.785, + 0.204 + ], + "angle": 0, + "content": "19. Hui, T.W., Loy, C.C., Tang, X.: Depth map super-resolution by deep multi-scale guidance. In: ECCV. pp. 353-369. Springer (2016) 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.206, + 0.785, + 0.26 + ], + "angle": 0, + "content": "20. Janoch, A., Karayev, S., Jia, Y., Barron, J.T., Fritz, M., Saenko, K., Darrell, T.: A category-level 3d object dataset: Putting the kinect to work. Consumer Depth Cameras for Computer Vision: Research Topics and Applications pp. 141-165 (2013) 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.261, + 0.785, + 0.302 + ], + "angle": 0, + "content": "21. Kanopoulos, N., Vasanthavada, N., Baker, R.L.: Design of an image edge detection filter using the sobel operator. IEEE Journal of solid-state circuits 23(2), 358-367 (1988) 10" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.304, + 0.785, + 0.332 + ], + "angle": 0, + "content": "22. Koutilya, P., Zhou, H., Jacobs, D.: Sharingan: Combining synthetic and real data for unsupervised geometry estimation. In: CVPR. vol. 2, p. 5 (2020) 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.333, + 0.785, + 0.374 + ], + "angle": 0, + "content": "23. Kundu, J.N., Uppala, P.K., Pahuja, A., Babu, R.V.: Adadepth: Unsupervised content congruent adaptation for depth estimation. In: Proceedings of the IEEE conference on computer vision and pattern recognition. pp. 2656-2665 (2018) 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.375, + 0.785, + 0.403 + ], + "angle": 0, + "content": "24. Lee, J.H., Kim, C.S.: Multi-loss rebalancing algorithm for monocular depth estimation. In: ECCV. pp. 785-801. Springer (2020) 3, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.404, + 0.785, + 0.445 + ], + "angle": 0, + "content": "25. Lehtinen, J., Munkberg, J., Hasselgren, J., Laine, S., Karras, T., Aittala, M., Aila, T.: Noise2noise: Learning image restoration without clean data. arXiv preprint arXiv:1803.04189 (2018) 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.446, + 0.785, + 0.488 + ], + "angle": 0, + "content": "26. Li, Y., Yuan, L., Vasconcelos, N.: Bidirectional learning for domain adaptation of semantic segmentation. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 6936-6945 (2019) 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.488, + 0.785, + 0.529 + ], + "angle": 0, + "content": "27. Li, Z., Bhat, S.F., Wonka, P.: Patchfusion: An end-to-end tile-based framework for high-resolution monocular metric depth estimation. arXiv preprint arXiv:2312.02284 (2023) 1, 2, 3, 4, 5, 6, 7, 9, 10, 11, 12, 13" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.53, + 0.785, + 0.572 + ], + "angle": 0, + "content": "28. Li, Z., Chen, Z., Li, A., Fang, L., Jiang, Q., Liu, X., Jiang, J.: Unsupervised domain adaptation for monocular 3d object detection via self-training. In: European Conference on Computer Vision. pp. 245-262. Springer (2022) 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.573, + 0.785, + 0.614 + ], + "angle": 0, + "content": "29. Li, Z., Chen, Z., Liu, X., Jiang, J.: Depthformer: Exploiting long-range correlation and local information for accurate monocular depth estimation. Machine Intelligence Research pp. 1-18 (2023) 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.615, + 0.785, + 0.643 + ], + "angle": 0, + "content": "30. Li, Z., Wang, X., Liu, X., Jiang, J.: Binsformer: Revisiting adaptive bins for monocular depth estimation. arXiv preprint arXiv:2204.00987 (2022) 1, 3, 7, 11" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.643, + 0.785, + 0.683 + ], + "angle": 0, + "content": "31. Liu, C., Kumar, S., Gu, S., Timofte, R., Van Gool, L.: Single image depth prediction made better: A multivariate gaussian take. In: CVPR. pp. 17346-17356 (June 2023) 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.686, + 0.785, + 0.727 + ], + "angle": 0, + "content": "32. Lopez-Rodriguez, A., Mikolajczyk, K.: Desc: Domain adaptation for depth estimation via semantic consistency. International Journal of Computer Vision 131(3), 752-771 (2023) 3, 4, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.728, + 0.785, + 0.756 + ], + "angle": 0, + "content": "33. Metzger, N., Daudt, R.C., Schindler, K.: Guided depth super-resolution by deep anisotropic diffusion. In: CVPR. pp. 18237-18246 (2023) 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.757, + 0.785, + 0.798 + ], + "angle": 0, + "content": "34. Miangoleh, S.M.H., Dille, S., Mai, L., Paris, S., Aksoy, Y.: Boosting monocular depth estimation models to high-resolution via content-adaptive multi-resolution merging. In: CVPR. pp. 9685-9694 (2021) 1, 2, 3, 4, 5, 12" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.799, + 0.785, + 0.84 + ], + "angle": 0, + "content": "35. Mildenhall, B., Srinivasan, P.P., Tancik, M., Barron, J.T., Ramamoorthi, R., Ng, R.: Nerf: Representing scenes as neural radiance fields for view synthesis. Communications of the ACM 65(1), 99-106 (2021) 3" + }, + { + "type": "list", + "bbox": [ + 0.217, + 0.147, + 0.785, + 0.84 + ], + "angle": 0, + "content": null + } + ], + [ + { + "type": "header", + "bbox": [ + 0.643, + 0.115, + 0.732, + 0.127 + ], + "angle": 0, + "content": "PatchRefiner" + }, + { + "type": "page_number", + "bbox": [ + 0.769, + 0.116, + 0.786, + 0.127 + ], + "angle": 0, + "content": "17" + }, + { + "type": "ref_text", + "bbox": [ + 0.214, + 0.147, + 0.788, + 0.203 + ], + "angle": 0, + "content": "36. Pastore, G., Cermelli, F., Xian, Y., Mancini, M., Akata, Z., Caputo, B.: A closer look at self-training for zero-label semantic segmentation. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 2693-2702 (2021) 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.214, + 0.204, + 0.787, + 0.259 + ], + "angle": 0, + "content": "37. Paul, S., Tsai, Y.H., Schulter, S., Roy-Chowdhury, A.K., Chandraker, M.: Domain adaptive semantic segmentation using weak labels. In: Computer Vision-ECCV 2020: 16th European Conference, Glasgow, UK, August 23-28, 2020, Proceedings, Part IX 16. pp. 571-587. Springer (2020) 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.216, + 0.26, + 0.787, + 0.3 + ], + "angle": 0, + "content": "38. Piccinelli, L., Sakaridis, C., Yu, F.: idisc: Internal discretization for monocular depth estimation. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 21477-21487 (2023) 10, 12" + }, + { + "type": "ref_text", + "bbox": [ + 0.216, + 0.301, + 0.787, + 0.342 + ], + "angle": 0, + "content": "39. Poucin, F., Kraus, A., Simon, M.: Boosting instance segmentation with synthetic data: A study to overcome the limits of real world data sets. In: Int. Conf. Comput. Vis. Workshop. pp. 945-953 (2021) 2, 5, 6, 12" + }, + { + "type": "ref_text", + "bbox": [ + 0.216, + 0.343, + 0.787, + 0.383 + ], + "angle": 0, + "content": "40. Pseudo-Label, D.H.L.: The simple and efficient semi-supervised learning method for deep neural networks. In: ICML 2013 Workshop: Challenges in Representation Learning. pp. 1-6 (2013) 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.216, + 0.384, + 0.787, + 0.438 + ], + "angle": 0, + "content": "41. Rajpal, A., Cheema, N., Illgner-Fehns, K., Slusallek, P., Jaiswal, S.: High-resolution synthetic rgb-d datasets for monocular depth estimation. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 1188-1198 (2023) 3, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.216, + 0.439, + 0.787, + 0.481 + ], + "angle": 0, + "content": "42. Ranftl, R., Lasinger, K., Hafner, D., Schindler, K., Koltun, V.: Towards robust monocular depth estimation: Mixing datasets for zero-shot cross-dataset transfer. IEEE TPAMI 44(3) (2022) 3, 7, 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.216, + 0.481, + 0.787, + 0.508 + ], + "angle": 0, + "content": "43. Rey-Area, M., Yuan, M., Richardt, C.: 360monodepth: High-resolution 360deg monocular depth estimation. In: CVPR. pp. 3762-3772 (2022) 1, 2, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.216, + 0.509, + 0.787, + 0.549 + ], + "angle": 0, + "content": "44. Saito, K., Ushiku, Y., Harada, T.: Asymmetric tri-training for unsupervised domain adaptation. In: International Conference on Machine Learning. pp. 2988-2997. PMLR (2017) 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.216, + 0.55, + 0.787, + 0.605 + ], + "angle": 0, + "content": "45. Scharstein, D., Hirschmüller, H., Kitajima, Y., Krathwohl, G., Nesic, N., Wang, X., Westling, P.: High-resolution stereo datasets with subpixel-accurate ground truth. In: Pattern Recognition: 36th German Conference, GCPR 2014, Münster, Germany, September 2-5, 2014, Proceedings 36. pp. 31-42. Springer (2014) 6, 10" + }, + { + "type": "ref_text", + "bbox": [ + 0.216, + 0.606, + 0.787, + 0.646 + ], + "angle": 0, + "content": "46. Schops, T., Schonberger, J.L., Galliani, S., Sattler, T., Schindler, K., Pollefeys, M., Geiger, A.: A multi-view stereo benchmark with high-resolution images and multi-camera videos. In: CVPR. pp. 3260-3269 (2017) 3, 5, 6, 10" + }, + { + "type": "ref_text", + "bbox": [ + 0.216, + 0.647, + 0.787, + 0.702 + ], + "angle": 0, + "content": "47. Shin, I., Tsai, Y.H., Zhuang, B., Schulter, S., Liu, B., Garg, S., Kweon, I.S., Yoon, K.J.: Mm-tta: multi-modal test-time adaptation for 3d semantic segmentation. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 16928-16937 (2022) 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.216, + 0.702, + 0.787, + 0.73 + ], + "angle": 0, + "content": "48. Silberman, N., Hoiem, D., Kohli, P., Fergus, R.: Indoor segmentation and support inference from rgbd images. In: ECCV. pp. 746-760. Springer (2012) 5, 6, 10, 11" + }, + { + "type": "ref_text", + "bbox": [ + 0.216, + 0.73, + 0.787, + 0.757 + ], + "angle": 0, + "content": "49. Song, S., Lichtenberg, S.P., Xiao, J.: Sun rgb-d: A rgb-d scene understanding benchmark suite. In: CVPR. pp. 567-576 (2015) 5, 6, 10" + }, + { + "type": "ref_text", + "bbox": [ + 0.216, + 0.758, + 0.787, + 0.813 + ], + "angle": 0, + "content": "50. Taherkhani, F., Dabouei, A., Soleymani, S., Dawson, J., Nasrabadi, N.M.: Self-supervised Wasserstein pseudo-labeling for semi-supervised image classification. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 12267-12277 (2021) 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.216, + 0.814, + 0.787, + 0.84 + ], + "angle": 0, + "content": "51. Tosi, F., Liao, Y., Schmitt, C., Geiger, A.: Smd-nets: Stereo mixture density networks. In: CVPR. pp. 8942-8952 (2021) 2, 3, 5, 9, 10, 12" + }, + { + "type": "list", + "bbox": [ + 0.214, + 0.147, + 0.788, + 0.84 + ], + "angle": 0, + "content": null + } + ], + [ + { + "type": "page_number", + "bbox": [ + 0.218, + 0.116, + 0.236, + 0.127 + ], + "angle": 0, + "content": "18" + }, + { + "type": "header", + "bbox": [ + 0.272, + 0.115, + 0.576, + 0.129 + ], + "angle": 0, + "content": "Zhenyu Li, Shariq Farooq Bhat, Peter Wonka" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.147, + 0.785, + 0.175 + ], + "angle": 0, + "content": "52. Wang, Z., Bovik, A.C., Sheikh, H.R., Simoncelli, E.P.: Image quality assessment: from error visibility to structural similarity. IEEE TIP 13(4), 600-612 (2004) 9" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.177, + 0.785, + 0.203 + ], + "angle": 0, + "content": "53. Weiss, K., Khoshgoftaar, T.M., Wang, D.: A survey of transfer learning. Journal of Big data 3(1), 1-40 (2016) 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.218, + 0.204, + 0.785, + 0.231 + ], + "angle": 0, + "content": "54. Xian, K., Zhang, J., Wang, O., Mai, L., Lin, Z., Cao, Z.: Structure-guided ranking loss for single image depth prediction. In: CVPR. pp. 611-620 (2020) 3, 7, 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.218, + 0.232, + 0.785, + 0.272 + ], + "angle": 0, + "content": "55. Yang, J., Alvarez, J.M., Liu, M.: Self-supervised learning of depth inference for multi-view stereo. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 7526-7534 (2021) 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.273, + 0.785, + 0.313 + ], + "angle": 0, + "content": "56. Yang, L., Kang, B., Huang, Z., Xu, X., Feng, J., Zhao, H.: Depth anything: Unleashing the power of large-scale unlabeled data. arXiv preprint arXiv:2401.10891 (2024) 1, 3, 4, 6, 7, 8, 12" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.315, + 0.785, + 0.355 + ], + "angle": 0, + "content": "57. Yen, Y.T., Lu, C.N., Chiu, W.C., Tsai, Y.H.: 3d-pl: Domain adaptive depth estimation with 3d-aware pseudo-labeling. In: European Conference on Computer Vision. pp. 710-728. Springer (2022) 4, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.357, + 0.785, + 0.383 + ], + "angle": 0, + "content": "58. Yeshwanth, C., Liu, Y.C., Nießner, M., Dai, A.: Scannet++: A high-fidelity dataset of 3d indoor scenes. In: ICCV. pp. 12-22 (2023) 3, 5, 6, 10" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.384, + 0.785, + 0.411 + ], + "angle": 0, + "content": "59. Zhang, L., Rao, A., Agrawala, M.: Adding conditional control to text-to-image diffusion models. In: ICCV. pp. 3836-3847 (2023) 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.412, + 0.785, + 0.452 + ], + "angle": 0, + "content": "60. Zhao, S., Fu, H., Gong, M., Tao, D.: Geometry-aware symmetric domain adaptation for monocular depth estimation. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 9788-9798 (2019) 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.453, + 0.785, + 0.479 + ], + "angle": 0, + "content": "61. Zhao, W., Rao, Y., Liu, Z., Liu, B., Zhou, J., Lu, J.: Unleashing text-to-image diffusion models for visual perception. ICCV (2023) 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.218, + 0.481, + 0.785, + 0.535 + ], + "angle": 0, + "content": "62. Zhao, X., Schulter, S., Sharma, G., Tsai, Y.H., Chandraker, M., Wu, Y.: Object detection with a unified label space from multiple datasets. In: Computer Vision-ECCV 2020: 16th European Conference, Glasgow, UK, August 23-28, 2020, Proceedings, Part XIV 16. pp. 178-193. Springer (2020) 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.536, + 0.785, + 0.563 + ], + "angle": 0, + "content": "63. Zhao, Z., Zhang, J., Xu, S., Lin, Z., Pfister, H.: Discrete cosine transform network for guided depth map super-resolution. In: CVPR. pp. 5697-5707 (2022) 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.565, + 0.785, + 0.605 + ], + "angle": 0, + "content": "64. Zheng, C., Cham, T.J., Cai, J.: T2net: Synthetic-to-realistic translation for solving single-image depth estimation tasks. In: Proceedings of the European conference on computer vision (ECCV). pp. 767-783 (2018) 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.606, + 0.785, + 0.632 + ], + "angle": 0, + "content": "65. Zhong, Z., Liu, X., Jiang, J., Zhao, D., Ji, X.: Guided depth map super-resolution: A survey. ACM Computing Surveys (2023) 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.633, + 0.785, + 0.674 + ], + "angle": 0, + "content": "66. Zou, Y., Yu, Z., Kumar, B., Wang, J.: Unsupervised domain adaptation for semantic segmentation via class-balanced self-training. In: Proceedings of the European conference on computer vision (ECCV). pp. 289-305 (2018) 3" + }, + { + "type": "list", + "bbox": [ + 0.217, + 0.147, + 0.785, + 0.674 + ], + "angle": 0, + "content": null + } + ] +] \ No newline at end of file diff --git a/2024/PatchRefiner_ Leveraging Synthetic Data for Real-Domain High-Resolution Monocular Metric Depth Estimation/3be463cc-22b0-43d6-8e8c-f91c9fa34d5b_origin.pdf b/2024/PatchRefiner_ Leveraging Synthetic Data for Real-Domain High-Resolution Monocular Metric Depth Estimation/3be463cc-22b0-43d6-8e8c-f91c9fa34d5b_origin.pdf new file mode 100644 index 0000000000000000000000000000000000000000..4258c75a555cce2687b7ac91a946f1f375f69093 --- /dev/null +++ b/2024/PatchRefiner_ Leveraging Synthetic Data for Real-Domain High-Resolution Monocular Metric Depth Estimation/3be463cc-22b0-43d6-8e8c-f91c9fa34d5b_origin.pdf @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:8fc0492b09265f390c88a3e7f6078b116cf8f2b0ba8ccebb9ac74eb4672ee082 +size 4611551 diff --git a/2024/PatchRefiner_ Leveraging Synthetic Data for Real-Domain High-Resolution Monocular Metric Depth Estimation/full.md b/2024/PatchRefiner_ Leveraging Synthetic Data for Real-Domain High-Resolution Monocular Metric Depth Estimation/full.md new file mode 100644 index 0000000000000000000000000000000000000000..9b8174e5e5dd35ed94c3ee64719951f763964590 --- /dev/null +++ b/2024/PatchRefiner_ Leveraging Synthetic Data for Real-Domain High-Resolution Monocular Metric Depth Estimation/full.md @@ -0,0 +1,391 @@ +# PatchRefiner: Leveraging Synthetic Data for Real-Domain High-Resolution Monocular Metric Depth Estimation + +Zhenyu Li, Shariq Farooq Bhat, Peter Wonka + +King Abdullah University of Science and Technology (KAUST) zhenyu.li.1@kaust.edu.sa + +Abstract. This paper introduces PatchRefiner, an advanced framework for metric single image depth estimation aimed at high-resolution real-domain inputs. While depth estimation is crucial for applications such as autonomous driving, 3D generative modeling, and 3D reconstruction, achieving accurate high-resolution depth in real-world scenarios is challenging due to the constraints of existing architectures and the scarcity of detailed real-world depth data. PatchRefiner adopts a tile-based methodology, reconceptualizing high-resolution depth estimation as a refinement process, which results in notable performance enhancements. Utilizing a pseudo-labeling strategy that leverages synthetic data, PatchRefiner incorporates a Detail and Scale Disentangling (DSD) loss to enhance detail capture while maintaining scale accuracy, thus facilitating the effective transfer of knowledge from synthetic to real-world data. Our extensive evaluations demonstrate PatchRefiner's superior performance, significantly outperforming existing benchmarks on the Unreal4KStereo dataset by $18.1\%$ in terms of the root mean squared error (RMSE) and showing marked improvements in detail accuracy and consistent scale estimation on diverse real-world datasets like CityScape, ScanNet++, and ETH3D. + +Keywords: High-Resolution Metric Depth Estimation $\cdot$ Synthetic Data + +# 1 Introduction + +This paper delves into the field of metric single-image depth estimation, focusing on high-resolution inputs from the real domain. High-resolution depth estimation plays a key role in autonomous driving, augmented reality, content creation, and 3D reconstruction [2,11,30,59]. Despite significant progress, the high-resolution depth estimation in real-world scenarios remains daunting. This challenge is primarily due to the resolution limitations inherent in most state-of-the-art depth estimation architectures [2, 56, 61] and the scarcity of high-quality real-world depth datasets. + +The current state-of-the-art in high-resolution depth estimation, PatchFusion [27, 34, 43], employs a tile-based strategy to navigate the resolution constraints, posing the task as a fusion process of coarse and fine depth estimations. + +![](images/131911c50077e8a56732c678d77d0484938b30bd000a405316e52ed57133234a.jpg) +(a) LR Depth + +![](images/8cfde7296b386600ec18cc38cf4e70a1db57087fe160e66543be58f1c6f3df81.jpg) +(b) Fusion-Based Framework + +![](images/b48e13f85a3b73ccff89df2117da75a4ef32fdab0fc767f7398d3c4e06a4a624.jpg) +(c) Refiner-Based Framework +Fig. 1: Framework Comparison. (a) Low resolution depth estimation framework with single forward pass. (b) Fusion-based high-resolution framework combining the best of coarse and fine depth predictions [27, 34]. (c) Our refiner-based framework predicts a residual to refine the coarse prediction. + +Because of the scarcity of real high-resolution depth datasets, PatchFusion resorts to training on a synthetic 4K dataset [27, 51]. There are two limitations of PatchFusion that we would like to improve upon: 1) It utilizes a three-step training process, which is not only time-consuming and expensive but also risks the framework achieving the stage-wise local optima and limits the potential performance gains from end-to-end learning. 2) PatchFusion demonstrates poor generalization to the real domain. + +The poor synthetic-to-real generalization is a long-standing problem in general [12,53]. In the context of depth estimation, the difference in the metric scale and depth distributions of synthetic datasets and the real domain further exacerbates the domain shift. This results in particularly subpar scale accuracy of depth models on real data when trained on synthetic datasets. On the other hand, the real datasets are not only low resolution but also often have missing ground truth values due to sensor constraints, occlusion, etc (see Fig. 3 and Sec. 3.4). This adds to the inability of depth models to capture sharp details when trained on real datasets. Thus, one is confronted with a dilemma - Training on real datasets leads to good scale accuracy but poor high-frequency details, while training on synthetic datasets leads to sharper results but poor scale performance on real images. + +We introduce PatchRefiner, a novel framework that reformulates high-resolution depth estimation as a process of refining coarse depth. We propose improvements on two levels: + +First, unlike direct depth regression approaches [27,39,43], PatchRefiner utilizes a frozen coarse depth model and enhances the quality by predicting residual depth for refinement. This approach not only streamlines model training but also markedly enhances performance. + +Second, we propose a method to exploit the best of both worlds to solve the above synthetic-real dilemma. We employ a teacher-student framework, leveraging the sharpness of synthetic data while learning the scale from real data. The teacher model, pre-trained on synthetic data, generates pseudo labels for real-domain training samples. Recognizing that these pseudo labels offer detailed features albeit with scale inaccuracies, we introduce the Detail and Scale Dis- + +entangling (DSD) loss. This loss integrates ranking supervision and scale-shift invariance, drawing inspiration from recent advances in relative depth estimation [5, 42, 54, 56]. It leads to a framework capable of delivering high-resolution depth estimates with both precise scale and sharp details in real-world settings. + +Our evaluation of PatchRefiner on the Unreal4KStereo synthetic dataset [51] demonstrates a substantial improvement over the current state-of-the-art, reducing RMSE by $18.1\%$ and REL by $15.7\%$ . Further, we assess the framework's efficacy in leveraging synthetic data across diverse real-world datasets, including CityScape [9] (outdoor, stereo), ScanNet++ [58] (indoor, LiDAR and reconstruction), and ETH3D [46] (mixed, LiDAR). Our findings reveal notable enhancements in depth boundary delineation (e.g., a $19.2\%$ increase in boundary recall on CityScape) while maintaining accurate scale estimation, showcasing the framework's adaptability and effectiveness across varying settings and sensor technologies. + +# 2 Related Work + +High-Resolution Monocular Metric Depth Estimation. Monocular depth estimation has achieved tremendous progress [1, 11, 13, 15, 24, 29, 30]. Current SOTA approaches often employ complex network architectures yet grapple with the limitations imposed by their input resolution [2, 56]. This stands in stark contrast to the advancements in modern imaging devices that capture images at increasing resolutions, and the growing demand among users for high-resolution depth estimation. Initial efforts to address this gap included the use of Guided Depth Super-Resolution (GDSR) [19,33,63,65] and Implicit Functions [6,35]. Recently, Tile-Based Methods have emerged as a potent strategy for high-resolution depth estimation [27,34,43], segmenting images into patches for individual processing before reassembling them into a comprehensive depth map. This paper extends the tile-based approach, seeking to elevate the quality of depth estimation further. + +Synthetic Data for Depth Estimation. The challenge of acquiring high-quality, real-domain data for high-resolution depth training has prompted the use of synthetic datasets [41]. Traditionally, synthetic data has been employed within unsupervised domain adaptation frameworks, utilizing labeled synthetic and unlabeled real-domain data to enhance depth estimation accuracy on real-world images [7,22,23,32,60,64]. Techniques vary from pixel-level style transfer and image translation to feature-level adversarial learning [7,23,60,64], with some methods integrating additional information such as stereo pairs or segmentation maps for enhanced adaptation [22,32,60]. Contrasting with these approaches, our work explores a practical scenario where labeled data from both synthetic and real domains are leveraged to improve real-world, high-resolution depth estimation, delving into a relatively underexplored application of synthetic data. + +Pseudo-Labeling for Depth Estimation. Pseudo-labeling, a cornerstone of semi-supervised learning, has been widely applied across various domains, including classification [3,17,40,44,50] and scene understanding [8,26,28,36,37,47,62, + +![](images/fc5f3c6def0d0be422667eafcc0ca334f82bc0e18f9c873309a33f77b5f3c887.jpg) +Fig. 2: Architecture Illustration. PatchRefiner contains a pre-trained frozen coarse depth estimation model $\mathcal{N}_c$ and a refiner model $\mathcal{N}_r$ predicts residual depth map $\mathcal{D}_r$ to refine the coarse depth $\mathcal{D}_c$ . The refiner contains one base depth model $\mathcal{N}_d$ that has the same architecture as $\mathcal{N}_c$ , and a light-weight decoder to aggregate information and make the final prediction. + +66], to extrapolate knowledge from labeled data to unlabeled datasets. In depth estimation, pseudo-labeling often serves to provide supplementary supervision in unsupervised domain adaptation settings [32, 55, 57]. While recent state-of-the-art methods like Depth-Anything utilize pseudo-labeling to enhance model generalization [56], these techniques predominantly aim to refine or enhance the pseudo labels themselves. Our approach diverges by utilizing real-domain data with accurate depth labels, focusing on a novel Detail and Scale Disentangling loss. This loss mechanism uniquely leverages the detailed insights from pseudo labels to enrich real-domain depth estimation without compromising the scale accuracy derived from real ground-truth data. + +# 3 Method + +In this section, we first present the overall PatchRefiner framework in Sec. 3.1. Then, we introduce the limitation of adopting real-domain data for high-resolution depth estimation 3.2 and the proposed teacher-student framework in Sec. 3.3 with the Detail and Scale Disentangling loss (DSD) in Sec. 3.4. + +# 3.1 PatchRefiner Framework + +PatchRefiner follows the tile-based strategy to address the prohibitive memory and computational demands for high resolutions such as 4K [27, 34]. However, recognizing the limitations of existing models, we propose a simplified two-step approach for high-resolution depth estimation: (i) Coarse Scale-Aware Estimation, and (ii) Fine-Grained Depth Refinement, shown in Fig. 2. + +(i) Coarse Scale-Aware Estimation: The foundation of PatchRefiner is the Coarse Depth Estimation network, $\mathcal{N}_c$ , which processes downsampled versions of the input images to produce a global depth prediction, $\mathbf{D}_c$ . Similar to + +previous work [27, 34], this step is crucial for establishing a baseline depth map that captures the scene's overall structure and depth consistency, albeit without high-resolution details. Moreover, $\mathcal{N}_c$ can be an arbitrary depth estimation model and it is frozen after the first step of training. + +(ii) Depth Refinement Process: Different from the conventional approach of using a separate fine depth network and a fusion mechanism [27, 39], our framework introduces a unified refinement network, $\mathcal{N}_r$ . This network is designed to refine the coarse depth map by focusing on the recovery of lost details and enhancing depth precision at a patch-wise level. + +The input to $\mathcal{N}_r$ is the cropped original image $I$ , which is processed by the base depth model $\mathcal{N}_d$ with the same architecture as $\mathcal{N}_c$ in the refiner module. Then, we collect $L$ -level multi-scale immediate features from both $\mathcal{N}_d$ and $\mathcal{N}_c$ , denoting them as $\mathcal{F}_d = \{f_d^i\}_{i=1}^L$ and $\tilde{\mathcal{F}}_c = \{\tilde{f}_c^i\}_{i=1}^L$ , respectively. Following [27], we apply the roi [16] operation to fetch features of the corresponding cropped area as $\tilde{f}_c^i = \text{roi}(f_c^i)$ . + +Next, we adopt a lightweight decoder to obtain high-resolution predictions. Given $\mathcal{F}_d$ and $\bar{\mathcal{F}}_c$ , we aggregate them with concatenation operators (Cat) following by simple convolutional blocks (CB): + +$$ +f _ {r} ^ {i} = \operatorname {C B} \left(\operatorname {C a t} \left(f _ {d} ^ {i}, \tilde {f} _ {c} ^ {i}\right)\right), \tag {1} +$$ + +The output feature set $\mathcal{F}_r = \{f_r^i\}_{i=1}^L$ is then fed to a successive series of up-sampling layers [25], in order to construct the residual depth map $\mathbf{D}_r$ at the input resolution. Associated with skip-connections, they form our refiner decoder. Further details about the architecture with immediate feature shapes are described in the supplementary material. Finally, we calculate the final patchwise depth estimation as $\mathbf{D} = \mathsf{roi}(\mathbf{D}_c) + \mathbf{D}_r$ . + +# 3.2 Limitation of Real-Domain Depth Estimation Datasets + +Our second goal in this paper is to train a real-domain high-resolution depth estimation model with both synthetic dataset $\mathcal{S}$ and real-domain dataset $\mathcal{R}$ . Our main insight is to distinguish between scale errors and boundary errors. + +In the field of high-resolution depth estimation, the prevailing state-of-the-art methodology [27] trains models using synthetic datasets, which offers paired high-resolution images and corresponding dense high-resolution depth ground-truth maps [18,51]. This synthetic training regime, while beneficial in a controlled setting, introduces significant challenges when models are applied to real-world data due to the intrinsic domain gap between synthetic and real-world environments. This gap often manifests as substantial scale errors during inference on real-domain datasets as shown in Tab. 2. + +Addressing this issue directly by training on real-domain datasets presents its own set of challenges. Real-world high-resolution depth datasets are scarce and typically constrained by limited resolution and unavailable missing ground-truth pixels. These limitations stem from the methods employed in generating real-world depth annotations, such as Kinect [20, 48, 49], LiDAR [14, 46, 58], or + +![](images/fe6917163c2e50005745bdc52f8546710cd5931788b69dc1a8829b4ae0d8c133.jpg) +Fig.3: Visualization of Real-Domain Data Pairs. Points lacking ground-truth data are depicted in gray. Due to sparse annotations near edges, models trained on real-domain data exhibit blurred boundary estimations. + +![](images/87ecd62ae9318cd8e358b71389c009ef71e1a69a508d615974efba15c63abe2b.jpg) + +stereo vision techniques [9, 45], each with their inherent drawbacks as shown in Fig. 3. + +For instance, depth maps generated using Kinect technology are confined to a resolution of $640 \times 480$ [48, 49], insufficient for high-resolution depth estimation [27]. LiDAR-based depth maps, while useful, tend to be sparse [14, 46] or limited in resolution (e.g., 256x192 in Scannet++ [58]), and the process of reconstructing high-resolution dense depth maps from LiDAR point clouds is fragile with cascade errors and omissions [58]. Stereo vision techniques, on the other hand, can also introduce missing values around object boundaries due to rectification transformations [9, 45]. It is important to realize that existing high-resolution real datasets with missing values around edges do not help in reducing boundary errors. + +These limitations highlight the inherent difficulties in employing real-world datasets for training high-resolution depth estimators [27, 39]. The lack of high-quality, high-resolution ground-truth depth maps in the real domain makes it challenging to train models that can accurately predict sharp depth around fine object boundaries [41]. Our main idea is to devise a training strategy that can improve the scale errors when fine-tuning on real data, while maintaining high-resolution information around edges to minimize boundary errors. + +# 3.3 Overall Pipeline Illustration + +Building on the success of semi-supervised learning frameworks in depth estimation tasks [32,56,57], our proposed framework employs a teacher-student architecture, as shown in Fig. 4, to effectively integrate synthetic and real-domain data for high-resolution depth estimation. It employs a two-step process: + +![](images/238c474efd7684cb61068d3f08b54809b865fb85cfca8dae1fc1dc8e02c765ca.jpg) +Fig. 4: Enhancing Real-Domain Learning with Synthetic Data. A teacher model trained on synthetic data produces pseudo labels for real-domain training. The student model benefits from a DSD dual-supervision approach: loss on pseudo labels for detail enhancement and loss on ground truth for scale accuracy. This method ensures detailed depth perception without compromising scale accuracy. + +Teacher Model Training: The teacher model is initially pretrained on a synthetic dataset $S$ , which, due to its high-quality and detailed annotations, enables the model to predict high-resolution depth maps with precise boundaries. + +Student Model Training with both Pseudo Labels and Ground-Truth Depth: In the subsequent phase, the teacher model is frozen, and the student model is trained on the real-domain dataset $\mathcal{R}$ , utilizing both the ground truth depth labels $\tilde{\mathbf{D}}$ and pseudo labels $\hat{\mathbf{D}}$ generated by the teacher model. As identified in the limitations of real-domain datasets, while the ground truth labels offer accurate depth information, they miss crucial information near boundaries essential for high-resolution depth learning. To address this, the student model is guided by the teacher's pseudo labels, which excel near boundaries. However, these pseudo labels, while sharp, exhibit scale discrepancies due to the domain gap between synthetic and real-world data as shown in Tab. 2. + +Therefore, we introduce the Detail and Scale Disentangling (DSD) loss. It ensures that the final high-resolution depth predictions from the student model maintain an accurate scale while benefiting from the enhanced boundary details provided by the teacher model trained on the synthetic data. + +# 3.4 Detail and Scale Disentangling Loss + +The student model is subject to two sources of supervision. The primary one is the standard scale-invariant loss [2,11,27,30], $\mathcal{L}_{\text{silog}}$ , calculated with the ground truth (GT) data, which ensures reliable and scale-consistent depth estimation on the real-domain data. + +The secondary source of supervision comes from pseudo labels generated by the teacher model. These labels, while detailed in capturing boundaries and fine details, exhibit low scale accuracy. Directly applying conventional depth losses [11,24,31] could lead to an imbalance between enhancing detail and maintaining scale accuracy. To address this, motivated by current impressive progress in relative depth estimation [5,42,54,56], we adopt the ranking loss $\mathcal{L}_{\text{rank}}$ [5,54] + +and the Scale and Shift Invariant losses $\mathcal{L}_{s s i}$ [42,56] to inject the detail information. Notably, The $\mathcal{L}_{\text {rank }}$ only provides supervision on the prediction relationship whereas $\mathcal{L}_{s s i}$ ignores scale and shift. Both losses tackle the same challenge with a slightly different approach. + +Ranking loss [5, 54] is designed for sparse sets with ordinal data. This adaptation allows us to leverage dense detailed ranking information from the pseudo labels, enhancing high-resolution estimation without compromising scale accuracy. Specifically, for a certain pair of points $[p_{i,1}, p_{i,2}]$ with predicted depth values $[d_{i,1}, d_{i,2}]$ and pseudo depth values $[\hat{d}_{i,1}, \hat{d}_{i,2}]$ in a set of $N$ sampled point pairs $\mathcal{P} = \{[p_{i,1}, p_{i,2}], i = 1, 2, \dots, N\}$ , the ranking loss is defined as + +$$ +\mathcal {L} _ {\text {r a n k}} = \frac {1}{N} \sum_ {i} \phi \left(p _ {i, 1}, p _ {i, 2}\right), \tag {2} +$$ + +$$ +\phi \left(p _ {i, 1}, p _ {i, 2}\right) = \left\{ \begin{array}{l l} \log \left(1 + \exp \left(- \ell \times \left(d _ {i, 1} - d _ {i, 2}\right)\right)\right), & \ell \neq 0, \\ \left(d _ {i, 1} - d _ {i, 2}\right) ^ {2}, & \ell = 0, \end{array} \right. \tag {3} +$$ + +where $\ell$ is the pseudo ordinal label, which can be induced by: + +$$ +\ell = \left\{ \begin{array}{l l} + 1, & \hat {d} _ {i, 1} / \hat {d} _ {i, 2} > = 1 + \tau , \\ - 1, & \hat {d} _ {i, 1} / \hat {d} _ {i, 2} < = \frac {1}{1 + \tau}, \\ 0, & \text {o t h e r w i s e .} \end{array} \right. \tag {4} +$$ + +Here $\tau$ is a tolerance threshold, which is set to 0.03 in our experiments following [54]. This loss not only encourages the predicted depth values for closely related points to align but also emphasizes the differentiation of points when the pseudo depth value of the two points is different. + +Scale and Shift Invariant loss [42, 56] is proposed to learn relative depth estimation with ignoring the unknown scale and shift of each sample: + +$$ +\mathcal {L} _ {s s i} = \frac {1}{M} \sum_ {i = 1} ^ {M} \rho \left(d _ {i} ^ {*} - \hat {d} _ {i} ^ {*}\right), \tag {5} +$$ + +where $d_i^*$ and $\hat{d}_i^*$ are scaled and shifted versions of the predicted depth $d_i$ and pseudo label $\hat{d}_i$ , and $\rho$ is the mean absolute error loss. $M$ is the number of pixels. We use the least-squares criterion to align the prediction to the ground truth: + +$$ +(s, t) = \operatorname {a r g m i n} _ {s, t} \sum_ {i = 1} ^ {M} \left(s d _ {i} + t - \hat {d} _ {i}\right) ^ {2}, \tag {6} +$$ + +$$ +d ^ {*} = s d + t, \hat {d} ^ {*} = \hat {d} \tag {7} +$$ + +where the scale $s$ and shift $t$ factors are effectively determined with the closed form [42]. + +The final Detail and Scale Disentangling loss for training the student model combines the $\mathcal{L}_{\text{silog}}$ , $\mathcal{L}_{\text{rank}}$ , and $\mathcal{L}_{\text{ssi}}$ as follows: + +![](images/8dab28b88d7563df30b6187fecfc8857fe6795683a82a2f92436616104e182c4.jpg) + +![](images/6c68fbedfa8a1c82b2a1a77f63b210f2f0a2b87089f50957cb3ab25ae05d0b1a.jpg) + +![](images/2f08459c366f5842c705c642757da8ba579c4758c4f284199217bf174f13f01e.jpg) + +![](images/c58e792f875c97aa3ada614d43537838c8498edb30d5d798e995884d3ae81ad4.jpg) +ZoeDepth [2] + +![](images/c4beefae1114f0d2774628194d7a91ebd4f07cb291c7cad5096deee59acb2f47.jpg) + +![](images/5c7b289c8e4930dd5ac194640954128a479160058341fa71256fb1d0f1a7adbe.jpg) + +![](images/bc860e2d7bd90318b0e40008d643c053271d41c6bacfe815438e4b85be9ea37d.jpg) + +![](images/88f4b67b583fa246a8775c1d39a11c928c5c40654a3fb401f9273a2b99b41149.jpg) +PatchFusion [27] + +![](images/e01f47a4253b0550142c5aee49050ef1092532c0b4d91a39f896709714fb5fa0.jpg) + +![](images/df734763aadd7f059378f04649a6248048b487a13c8e2865277e42ef36437d4a.jpg) + +![](images/24cc49880004440dee69805bbaff368ce5bf55ebf7d39c818f81658e7ee300d8.jpg) + +![](images/9255df48973122c1d1596726e4505f4de5bbffa868d5d0a5d1c6c056042c6ebb.jpg) +Ours +Fig. 5: Qualitative Comparison on UnrealStereo4K. We show the depth prediction and corresponding error map, respectively. The qualitative comparisons showcased here indicate our framework outperforms counterparts [2, 27] with sharper edges and lower error around boundaries. We show individual patches in all images to emphasize details near depth boundaries. + +![](images/eb9bf0c1d9113e5f8e80a0e473b30eb83a45a3c45a23b9a9e5be5fe5faec0d19.jpg) + +![](images/fc54c1ccfc9b2dbb82684191c04d34522f7074e00f0d5f2e88e82aa030e3b234.jpg) + +![](images/aad4f5d5ccbd16761a3442768302a5b38dad7bf541a49195a72e78a232a2538c.jpg) + +![](images/53ca2cb852d8fe3b66c8d74328dba9e51b2b253b2d1a2d517b80900fc05222ff.jpg) +GT, Input + +$$ +\mathcal {L} _ {D S D} = \overbrace {\mathcal {L} _ {s i l o g}} ^ {\mathcal {L} _ {g t}} + \overbrace {\lambda_ {1} \mathcal {L} _ {r a n k} + \lambda_ {2} \mathcal {L} _ {s s i}} ^ {\mathcal {L} _ {p l}}, \tag {8} +$$ + +where $\lambda_{1}$ and $\lambda_{2}$ are two balancing factors, both empirically set to 0.1 in our experiments, respectively. The terms $\mathcal{L}_{gt}$ and $\mathcal{L}_{pl}$ represent the supervision signals derived from the ground truth and pseudo labels, respectively. + +# 4 Experiments + +# 4.1 Datasets and Metrics + +UnrealStereo4K (Synthetic): The UnrealStereo4K dataset [51] offers synthetic stereo images at a 4K resolution $(2160\times 3840)$ , each paired with accurate, boundary-complete pixel-wise ground truth. Following the procedure in [27], we exclude mislabeled images using the Structural Similarity Index (SSIM) [52] and compute ground truth depth maps from provided disparity maps using camera parameters. Adhering to the dataset splits in [27, 51], we employ a default patch size of $540\times 960$ for compatibility with [27]. + +CityScapes (Real, Stereo): Cityscapes [9] offers a comprehensive suite of urban scene images, segmentation masks, and disparity maps at a resolution + +![](images/9a82ae2f5eeeb936bab130c8da7434df19693f3fefe365ba4597cee22ef16bd6.jpg) + +![](images/f121f3099f538b47ec7e9c78c35d53f79c13b3ea5969cc3debf791f97014636e.jpg) + +![](images/5989e62ce3a0c89bd616266826c0151a011da40d38c82361081ba6d95847c282.jpg) + +![](images/43494bd5ec155036c39632cc68277204c4e25ca6f9ebf30efd0ebeace4e4177c.jpg) +Image + +![](images/291c1f44d547b4fb8bba2ec6a94c0a880d9ba5b44709b5dee0bffa46cba658ec.jpg) + +![](images/7a4342bbaa38dd6ecf58b592662bf78ba9c000ca50d02d1bdc8b29398a8ea9f4.jpg) + +![](images/4642ed713d37ed4b63ce59ed15c6f41918acdd089a15b5f26f7e71a42ef6f9cc.jpg) + +![](images/b19b9e13cc89a4bc5588dcc2e2b775a88ed1d41b6ad93f8dba25041b436eca27.jpg) +ZoeDepth [2] + +![](images/e9a3ca30e4056ab2eb9607b245365d894e49378b63ba9ba373a4e011beecfbaf.jpg) + +![](images/c0502f40097d064392fc97ab1269e4c981c13d4fa6e2c4cdf3f31355b611e0d6.jpg) + +![](images/863fabd6a3c138b8e0471cd71e00baf78300bc1742a3ae7c3c105e2861df0824.jpg) + +![](images/00e2026fb09e55e8d5039a008a63e4884e54eb2a6da4ff3fcfbf0c5a5719c3fb.jpg) + +![](images/a8f651b632aa099bbac8a3b95114804ba25dca6b6c202cd2cffaa60a8b4bf746.jpg) + +![](images/84b8e82a90297997638098c428dfddb86c5b55666d54b8d43c3c18f776f7e6a3.jpg) + +![](images/efca61ec132355ce07605ec69b5119fbdfe6a3702778d838c9fece8f2131430b.jpg) + +![](images/f5ce4825eedcf7cbff277dddffaf217a205310d89de00c536c1c4bca6d1412dc.jpg) +PR R +Fig. 6: Qualitative Comparison on ETH3D and ScanNet++. The first two rows depict results for ETH3D and the last two for ScanNet++. The baseline ZoeDepth and PatchRefiner with conventional fine tuning (PR $\mathcal{R}$ ) both struggle to create high-resolution depth details near boundaries. Our proposed strategy yields crisp boundaries on both datasets. + +![](images/b459ecba07ad557a1075c3a3e623799e31e01f0b7d84d36429e807a6394b85b9.jpg) + +![](images/b373343d15848305d659867a958f194cae8795048ebee5b883b89e78237d05f4.jpg) + +![](images/ba004701f1ea2d7cb45f30200b87a83a2e7c9c1bccf0976d5571cd0a2399bd98.jpg) + +![](images/27c246a3de63a7759baf0f2a00b0e2aaa257713ad856c9a24a5e3fd655bbeff4.jpg) +Ours 1 + +of $1024 \times 2048$ , surpassing many real-domain datasets in density, quantity, and resolution [45, 46, 48, 49]. We use a standard patch size of $256 \times 512$ and conduct most experiments on this dataset. + +ScanNet++ (Real, LiDAR, Reconstruction): ScanNet++ [58] is an extensive indoor dataset providing high-resolution images $(1440\times 1920)$ , depth maps from iPhone LiDAR $(192\times 256)$ , and high-resolution depth maps sampled from reconstructions of laser scans $(1440\times 1920)$ . Our chosen patch size is $720\times 960$ . We use the low-resolution ground-truth for training since the high-resolution version contains much noise as shown in Fig. 3. + +ETH3D (Real, LiDAR): The ETH3D benchmark [46] provides high-resolution indoor and outdoor images $(6048\times 4032)$ with ground-truth depth from laser sensors. We downsample the image-depth pairs to $2160\times 3840$ and select a patch size of $540\times 960$ . + +Metrics: We adopt standard depth evaluation metrics from [2, 11, 38] and the Soft Edge Error (SEE) from [4, 27, 51] for scale evaluation. Additional metric details are in the supplementary material. Given real-world depth map boundary incompleteness, we design metrics to focus on evaluating boundary accuracy, leveraging high-resolution fine-grained segmentation masks as depth boundary quality proxies. Specifically, we use the Sobel operator [21] on both predicted depth and segmentation maps to generate edge masks, and then compute Precision, Recall, and F1 Score. + +![](images/3eebc3454996388cb42a2c5423d6d7d2eae964d238e942613c36dfeb31f78dd4.jpg) +Fig. 7: Qualitative Comparison on CityScapes. This figure illustrates depth estimation comparisons between the base ZoeDepth model, PatchRefiner (PR) trained on CityScapes, and our method. We display outcomes under varying levels of $\mathcal{L}_{pl}$ supervision ( $\lambda_1 = \lambda_2 = 1e^{-1}$ or 1), featuring zoomed-in sections of each image to highlight detail fidelity near depth discontinuities. + +![](images/cbf42d6b563403d4ce2a678da840a98d7d2a104b7e2b532d11f0db1cddb31694.jpg) + +![](images/8ee3869a0da1c121b5c6d2d06617e3fda7827f30812088e868cf52a6d6699fac.jpg) + +![](images/8f34a358dd8d4a536260f08c20f741e16c67555430f728f0ceb03e24d447b9a3.jpg) + +![](images/910705c5cb23bf68a742fe54f336dfc9e91ef3446490140b06f65a5650d98a5e.jpg) + +# 4.2 Implementation Details + +PatchRefiner on Synthetic Dataset: For training on synthetic datasets, we employ the scale-invariant log loss $\mathcal{L}_{\text{silog}}$ , as outlined in [2, 11, 27, 30]. Initialization of the coarse network $\mathcal{N}_c$ leverages pretrained weights from the NYU-v2 dataset [48], adhering to the approach in [27]. We dedicate 24 epochs to training $\mathcal{N}_c$ , which is subsequently frozen to ensure stability in subsequent training phases. For the refinement network $\mathcal{N}_r$ , initialization employs $\mathcal{N}_c$ 's parameters, and training extends for an additional 36 epochs. Standard augmentation strategies from the baseline depth model are incorporated to enhance training effectiveness. During inference, we implement Consistency-Aware Inference, as described in [27], to optimize performance. + +Learning on Real-Domain Dataset: Since the coarse model within the framework offers scale-consistent predictions, we first pretrain a coarse model on real-domain data to establish a reliable depth scale foundation. This model is subsequently frozen to preserve scale consistency throughout the training process. The refiner model in PatchRefiner is then initialized with parameters pretrained on synthetic data, enabling it to maintain high-frequency detail knowledge acquired from the synthetic domain. Initially, the student model is trained solely + +Table 1: Quantitative comparison on UnrealStereo4K. We color code the corresponding best competitor and our method within each block. PF and PR are short for PatchFusion [27] and PatchRefiner, respectively. The reported numbers are from [27]. + +
Methodδ1(%)↑REL↓RMS↓SiLog↓SEE↓Reference
iDisc [38]96.9400.0531.4048.5021.070ICCV 2023
SMD-Net [51]97.7740.0441.2827.3890.883CVPR 2021
Graph-GDSR [10]97.9320.0441.2647.4690.872CVPR 2022
BoostingDepth [34]98.1040.0441.1236.6620.939CVPR 2021
ZoeDepth [2]97.7170.0461.2897.4480.914-
ZoeDepth+PFp=16 [27]98.4190.0401.0886.2120.838CVPR 2024
ZoeDepth+PFp=49 [27]98.4500.0391.0756.1310.846
ZoeDepth+PFR=128 [27]98.4690.0391.0666.0850.849
ZoeDepth+PRp=1698.8210.0330.8925.4170.750Ours
ZoeDepth+PRp=4998.8590.0330.8705.3190.751
ZoeDepth+PRR=12898.8640.0330.8725.3770.738
Depth-Anything [56]97.7730.0411.2357.1920.911CVPR 2024
Depth-Anything+PFp=16 [27]98.5580.0361.0155.8830.811CVPR 2024
Depth-Anything+PFp=49 [27]98.6070.0350.9875.7460.812
Depth-Anything+PFR=128 [27]98.6160.0350.9845.7750.813
Depth-Anything+PRp=1698.8260.0330.8895.2890.768Ours
Depth-Anything+PRp=4998.8780.0330.8605.1490.767
Depth-Anything+PRR=12898.8780.0330.8605.2060.759
+ +Table 2: Quantitative comparison on CityScapes. FT, and MIX are short for fine-tuning and mixed-data strategies, which are our main competitors. Baseline is highlighted in gray and each of the three competitors in a different color. + +
MethodDataScaleBoundary
SRδ1(%)↑REL↓RMS↓Precision↑Recall↑
ZoeDepth [2]✓94.5020.0704.40613.3237.59
ZoeDepth [2] + FT✓✓94.4980.0714.41812.9337.89
PatchRefiner (zero-shot)✓5.7050.39912.20328.6851.28
PatchRefiner✓95.2840.0664.04716.6739.53
PatchRefiner + FT✓✓95.4180.0653.99217.0940.92
PatchRefiner + mix [39]✓✓89.1080.1124.73223.0841.17
PatchRefiner + DSD✓✓95.3590.0663.98218.9248.78
+ +with $\mathcal{L}_{\text{silog}}$ for 24 epochs. Subsequent fine-tuning with our Detail and Scale Disentangling loss $\mathcal{L}_{DSD}$ over an additional 6 epochs to refine depth estimations. + +# 4.3 Main Results + +Synthetic Dataset: On the UnrealStereo4K dataset, PatchRefiner not only outperforms the base depth model but also shows substantial improvements over PatchFusion, reducing RMSE by $18.1\%$ and REL by $15.7\%$ . This advancement is further underscored by achieving the lowest SEE, highlighting our model's proficiency in capturing edge details. Qualitative comparisons, illustrated in Fig. 5, indicate the superior boundary delineation achieved by PatchRefiner. + +Table 3: Ablation study of architecture variations and formulation of final depth prediction. Ours is highlighted with color. We highlight the best in bold. + +
MethodType of outputFeature Levelsδ1(%)↑REL↓RMS↓SiLog↓SEE↓
PatchFusion [27]Direct6 features98.4190.0401.0886.2120.838
PatchRefinerDcresidual1 features98.7340.0340.9265.5500.782
Dcresidual2 features98.8140.0330.9055.5110.750
Dcresidual3 features98.8150.0340.9005.5830.753
Dcresidual4 features98.8150.0330.8995.4940.752
Dcresidual5 features98.8140.0330.8945.4680.752
Dcresidual6 features98.8210.0330.8925.4170.750
Ddresidual6 features98.8040.0330.8995.4480.753
Direct6 features98.7490.0340.9255.5910.765
+ +Table 4: Variations of $\mathcal{L}_{pl}$ . We analyse various options for $\mathcal{L}_{pl}$ and compare them against PR. $\mathcal{S}$ and PR. $\mathcal{R}$ that serve as baselines for training on Synthetic and Real data individually. We set $\lambda = \lambda_1 = \lambda_2$ to analyze the influence of DSD weight. The highlighted result is achieved with $\lambda = 1e^{-1}$ . + +
Variations (Lpl)ScaleBoundary
δ1(%)↑REL↓RMS↓Precision↑Recall↑F1↑
Baseline PR. S5.7050.39912.2028.6851.2836.34
Baseline PR. R95.4180.0653.99217.0940.9223.68
Lsilog82.5500.1555.91426.3458.3535.81
Lsilog + mask81.4250.1486.51330.0446.8936.19
Lrank95.4130.0663.97318.3847.8126.12
Lssi95.4650.0653.97419.2644.0126.39
Lrank + Lssi (Ours)95.3590.0663.98218.9248.7826.84
λ = 195.0770.0694.22221.3258.4830.90
λ = 3e-195.2960.0684.08620.1153.8328.91
λ = 3e-295.4620.0653.95318.0143.9025.11
+ +Real-Domain Dataset: Tab. 2 and Fig. 6, 7 delineate the performance disparity when leveraging synthetic data for real-domain learning. Although the synthetic-trained model excels in boundary details, it fails in scale accuracy due to the domain gap. Sole training on real-domain data enhances baseline's scale prediction yet falls short in detail accuracy due to the missing depth ground truth around boundaries. Neither fine-tuning nor mixed-training substantially elevates performance across scale and detail metrics, reflecting the inherent challenges in our task. Contrastingly, our strategy propels the model to notable gains in boundary accuracy (19.2% increase in boundary recall) while sustaining scale precision comparable to the baseline model. + +# 4.4 Ablation Studies and Discussion + +We ablate and discuss the contributions of individual components. As default, we utilize the UnrealStereo4K dataset for synthetic-dataset training and the PatchRefiner variant with $P = 16$ patches for clarity and ease of comparison. We ablate our teacher-student framework using the CityScapes dataset. + +Architecture and Formulation Variations: We first ablate the effectiveness of our architecture design. Utilizing a ZoeDepth model, we extract six levels of intermediate features during the forward pass, represented as $\mathcal{F} = f_1, f_2, \dots, f_6$ , with progressively increasing resolution. We then sequentially omit lower resolution feature maps from the refiner's decoder input. As depicted in Tab. 3, performance diminishes with fewer feature maps, yet even with solely the highest resolution feature map, our model outperforms PatchFusion. + +Further, we explore our refinement formulation's efficacy by comparing it against two alternatives: (1) Residual depth adjustment based on the base model $\mathcal{N}_d$ within the refiner. (2) Direct depth prediction from the refiner's decoder, similar to PatchFusion ( $\mathbf{D} = \mathbf{D}_r$ in this context). Our residual-based approach demonstrates superior performance, underscoring its advantages in optimization and training efficiency. + +Effectiveness of Pseudo Label Supervision: We evaluate the efficacy of our Detail and Scale Disentangling (DSD) loss against the conventional scale-invariant loss, $\mathcal{L}_{silog}$ , for pseudo label supervision. Tab. 4 illustrates that while $\mathcal{L}_{silog}$ aids in detail transfer from the teacher to the student model, it compromises scale accuracy due to significant discrepancies in the pseudo labels. A masking approach, focusing only on areas lacking depth, does not mitigate this issue, indicating pervasive negative effects. In contrast, the combination of $\mathcal{L}_{rank}$ and $\mathcal{L}_{ssi}$ within $\mathcal{L}_{pl}$ not only improves detail fidelity but also maintains scale accuracy, demonstrating that ranking constraints and scale-shift invariance are effective, orthogonal strategies for enhancing high-resolution detail without sacrificing scale accuracy. + +Exploration of Variant DSD Loss Weight: This study examines the impact of different DSD loss weights $(\lambda_{1}$ and $\lambda_{2})$ to elucidate the loss's efficiency. As shown in Tab. 4 and Fig. 7, increasing the loss weight marginally affects scale accuracy while significantly improving boundaries, validating the DSD loss's role in balancing detail enhancement and scale preservation. Notably, when both weights are set to 1, the model's boundary recall surpasses the teacher's performance. Furthermore, when DSD loss achieves a comparable boundary metric to $\mathcal{L}_{silog}$ , it exhibits a smaller performance decline, underscoring its effectiveness. + +# 5 Conclusion + +We presented PatchRefiner, a tile-based framework tailored to real-world high-resolution monocular metric depth estimation. It reconceptualizes high-resolution depth estimation as a refinement process. With a pseudo-labeling strategy that leverages synthetic data, we propose a Detail and Scale Disentangling (DSD) loss to enhance detail capture while maintaining scale accuracy. Our proposed framework decisively surpasses the current SOTA method for UnrealStereo4K (17.3% in RMSE), while demonstrating marked improvements in detail accuracy on real-world datasets such as CityScape, ScanNet++, and ETH3D. + +# Acknowledgements + +This publication is funded in part by KAUST under award #ORA-2023-5241 (for the NTGI-AI project). + +# References + +1. Bhat, S.F., Alhashim, I., Wonka, P.: Adabins: Depth estimation using adaptive bins. In: CVPR. pp. 4009-4018 (2021) 3 +2. Bhat, S.F., Birkl, R., Wofk, D., Wonka, P., Müller, M.: Zoedepth: Zero-shot transfer by combining relative and metric depth. arXiv preprint arXiv:2302.12288 (2023) 1, 3, 7, 9, 10, 11, 12 +3. Chen, C., Xie, W., Huang, W., Rong, Y., Ding, X., Huang, Y., Xu, T., Huang, J.: Progressive feature alignment for unsupervised domain adaptation. In: Proceedings of the IEEE/CVF conference on computer vision and pattern recognition. pp. 627-636 (2019) 3 +4. Chen, C., Chen, X., Cheng, H.: On the over-smoothing problem of cnn based disparity estimation. In: ICCV. pp. 8997-9005 (2019) 10 +5. Chen, W., Fu, Z., Yang, D., Deng, J.: Single-image depth perception in the wild. NeurIPS 29 (2016) 3, 7, 8 +6. Chen, Y., Liu, S., Wang, X.: Learning continuous image representation with local implicit image function. In: CVPR. pp. 8628-8638 (2021) 3 +7. Chen, Y.C., Lin, Y.Y., Yang, M.H., Huang, J.B.: Crdoco: Pixel-level domain transfer with cross-domain consistency. In: Proceedings of the IEEE/CVF conference on computer vision and pattern recognition. pp. 1791-1800 (2019) 3 +8. Chen, Z., Zhang, R., Zhang, G., Ma, Z., Lei, T.: Digging into pseudo label: a low-budget approach for semi-supervised semantic segmentation. IEEE Access 8, 41830-41837 (2020) 3 +9. Cordts, M., Omran, M., Ramos, S., Rehfeld, T., Enzweiler, M., Benenson, R., Franke, U., Roth, S., Schiele, B.: The cityscapes dataset for semantic urban scene understanding. In: CVPR. pp. 3213-3223 (2016) 3, 6, 9 +10. De Lutio, R., Becker, A., D'Aronco, S., Russo, S., Wegner, J.D., Schindler, K.: Learning graph regularisation for guided super-resolution. In: CVPR. pp. 1979-1988 (2022) 12 +11. Eigen, D., Puhrsch, C., Fergus, R.: Depth map prediction from a single image using a multi-scale deep network. NeurIPS 27 (2014) 1, 3, 7, 10, 11 +12. Farahani, A., Voghoei, S., Rasheed, K., Arabnia, H.R.: A brief review of domain adaptation. Advances in data science and information engineering: proceedings from ICDATA 2020 and IKE 2020 pp. 877-894 (2021) 2 +13. Fu, H., Gong, M., Wang, C., Batmanghelich, K., Tao, D.: Deep ordinal regression network for monocular depth estimation. In: CVPR. pp. 2002-2011 (2018) 3 +14. Geiger, A., Lenz, P., Urtasun, R.: Are we ready for autonomous driving? the kitti vision benchmark suite. In: CVPR. pp. 3354-3361. IEEE (2012) 5, 6 +15. Godard, C., Mac Aodha, O., Firman, M., Brostow, G.J.: Digging into self-supervised monocular depth estimation. In: ICCV. pp. 3828-3838 (2019) 3 +16. He, K., Gkioxari, G., Dollár, P., Girshick, R.: Mask r-cnn. In: Proceedings of the IEEE international conference on computer vision. pp. 2961-2969 (2017) 5 +17. Hu, Z., Yang, Z., Hu, X., Nevatia, R.: Simple: Similar pseudo label exploitation for semi-supervised classification. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 15099-15108 (2021) 3 + +18. Huang, P.H., Matzen, K., Kopf, J., Ahuja, N., Huang, J.B.: Deepmvs: Learning multi-view stereopsis. In: CVPR (2018) 5 +19. Hui, T.W., Loy, C.C., Tang, X.: Depth map super-resolution by deep multi-scale guidance. In: ECCV. pp. 353-369. Springer (2016) 3 +20. Janoch, A., Karayev, S., Jia, Y., Barron, J.T., Fritz, M., Saenko, K., Darrell, T.: A category-level 3d object dataset: Putting the kinect to work. Consumer Depth Cameras for Computer Vision: Research Topics and Applications pp. 141-165 (2013) 5 +21. Kanopoulos, N., Vasanthavada, N., Baker, R.L.: Design of an image edge detection filter using the sobel operator. IEEE Journal of solid-state circuits 23(2), 358-367 (1988) 10 +22. Koutilya, P., Zhou, H., Jacobs, D.: Sharingan: Combining synthetic and real data for unsupervised geometry estimation. In: CVPR. vol. 2, p. 5 (2020) 3 +23. Kundu, J.N., Uppala, P.K., Pahuja, A., Babu, R.V.: Adadepth: Unsupervised content congruent adaptation for depth estimation. In: Proceedings of the IEEE conference on computer vision and pattern recognition. pp. 2656-2665 (2018) 3 +24. Lee, J.H., Kim, C.S.: Multi-loss rebalancing algorithm for monocular depth estimation. In: ECCV. pp. 785-801. Springer (2020) 3, 7 +25. Lehtinen, J., Munkberg, J., Hasselgren, J., Laine, S., Karras, T., Aittala, M., Aila, T.: Noise2noise: Learning image restoration without clean data. arXiv preprint arXiv:1803.04189 (2018) 5 +26. Li, Y., Yuan, L., Vasconcelos, N.: Bidirectional learning for domain adaptation of semantic segmentation. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 6936-6945 (2019) 3 +27. Li, Z., Bhat, S.F., Wonka, P.: Patchfusion: An end-to-end tile-based framework for high-resolution monocular metric depth estimation. arXiv preprint arXiv:2312.02284 (2023) 1, 2, 3, 4, 5, 6, 7, 9, 10, 11, 12, 13 +28. Li, Z., Chen, Z., Li, A., Fang, L., Jiang, Q., Liu, X., Jiang, J.: Unsupervised domain adaptation for monocular 3d object detection via self-training. In: European Conference on Computer Vision. pp. 245-262. Springer (2022) 3 +29. Li, Z., Chen, Z., Liu, X., Jiang, J.: Depthformer: Exploiting long-range correlation and local information for accurate monocular depth estimation. Machine Intelligence Research pp. 1-18 (2023) 3 +30. Li, Z., Wang, X., Liu, X., Jiang, J.: Binsformer: Revisiting adaptive bins for monocular depth estimation. arXiv preprint arXiv:2204.00987 (2022) 1, 3, 7, 11 +31. Liu, C., Kumar, S., Gu, S., Timofte, R., Van Gool, L.: Single image depth prediction made better: A multivariate gaussian take. In: CVPR. pp. 17346-17356 (June 2023) 7 +32. Lopez-Rodriguez, A., Mikolajczyk, K.: Desc: Domain adaptation for depth estimation via semantic consistency. International Journal of Computer Vision 131(3), 752-771 (2023) 3, 4, 6 +33. Metzger, N., Daudt, R.C., Schindler, K.: Guided depth super-resolution by deep anisotropic diffusion. In: CVPR. pp. 18237-18246 (2023) 3 +34. Miangoleh, S.M.H., Dille, S., Mai, L., Paris, S., Aksoy, Y.: Boosting monocular depth estimation models to high-resolution via content-adaptive multi-resolution merging. In: CVPR. pp. 9685-9694 (2021) 1, 2, 3, 4, 5, 12 +35. Mildenhall, B., Srinivasan, P.P., Tancik, M., Barron, J.T., Ramamoorthi, R., Ng, R.: Nerf: Representing scenes as neural radiance fields for view synthesis. Communications of the ACM 65(1), 99-106 (2021) 3 + +36. Pastore, G., Cermelli, F., Xian, Y., Mancini, M., Akata, Z., Caputo, B.: A closer look at self-training for zero-label semantic segmentation. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 2693-2702 (2021) 3 +37. Paul, S., Tsai, Y.H., Schulter, S., Roy-Chowdhury, A.K., Chandraker, M.: Domain adaptive semantic segmentation using weak labels. In: Computer Vision-ECCV 2020: 16th European Conference, Glasgow, UK, August 23-28, 2020, Proceedings, Part IX 16. pp. 571-587. Springer (2020) 3 +38. Piccinelli, L., Sakaridis, C., Yu, F.: idisc: Internal discretization for monocular depth estimation. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 21477-21487 (2023) 10, 12 +39. Poucin, F., Kraus, A., Simon, M.: Boosting instance segmentation with synthetic data: A study to overcome the limits of real world data sets. In: Int. Conf. Comput. Vis. Workshop. pp. 945-953 (2021) 2, 5, 6, 12 +40. Pseudo-Label, D.H.L.: The simple and efficient semi-supervised learning method for deep neural networks. In: ICML 2013 Workshop: Challenges in Representation Learning. pp. 1-6 (2013) 3 +41. Rajpal, A., Cheema, N., Illgner-Fehns, K., Slusallek, P., Jaiswal, S.: High-resolution synthetic rgb-d datasets for monocular depth estimation. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 1188-1198 (2023) 3, 6 +42. Ranftl, R., Lasinger, K., Hafner, D., Schindler, K., Koltun, V.: Towards robust monocular depth estimation: Mixing datasets for zero-shot cross-dataset transfer. IEEE TPAMI 44(3) (2022) 3, 7, 8 +43. Rey-Area, M., Yuan, M., Richardt, C.: 360monodepth: High-resolution 360deg monocular depth estimation. In: CVPR. pp. 3762-3772 (2022) 1, 2, 3 +44. Saito, K., Ushiku, Y., Harada, T.: Asymmetric tri-training for unsupervised domain adaptation. In: International Conference on Machine Learning. pp. 2988-2997. PMLR (2017) 3 +45. Scharstein, D., Hirschmüller, H., Kitajima, Y., Krathwohl, G., Nesic, N., Wang, X., Westling, P.: High-resolution stereo datasets with subpixel-accurate ground truth. In: Pattern Recognition: 36th German Conference, GCPR 2014, Münster, Germany, September 2-5, 2014, Proceedings 36. pp. 31-42. Springer (2014) 6, 10 +46. Schops, T., Schonberger, J.L., Galliani, S., Sattler, T., Schindler, K., Pollefeys, M., Geiger, A.: A multi-view stereo benchmark with high-resolution images and multi-camera videos. In: CVPR. pp. 3260-3269 (2017) 3, 5, 6, 10 +47. Shin, I., Tsai, Y.H., Zhuang, B., Schulter, S., Liu, B., Garg, S., Kweon, I.S., Yoon, K.J.: Mm-tta: multi-modal test-time adaptation for 3d semantic segmentation. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 16928-16937 (2022) 3 +48. Silberman, N., Hoiem, D., Kohli, P., Fergus, R.: Indoor segmentation and support inference from rgbd images. In: ECCV. pp. 746-760. Springer (2012) 5, 6, 10, 11 +49. Song, S., Lichtenberg, S.P., Xiao, J.: Sun rgb-d: A rgb-d scene understanding benchmark suite. In: CVPR. pp. 567-576 (2015) 5, 6, 10 +50. Taherkhani, F., Dabouei, A., Soleymani, S., Dawson, J., Nasrabadi, N.M.: Self-supervised Wasserstein pseudo-labeling for semi-supervised image classification. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 12267-12277 (2021) 3 +51. Tosi, F., Liao, Y., Schmitt, C., Geiger, A.: Smd-nets: Stereo mixture density networks. In: CVPR. pp. 8942-8952 (2021) 2, 3, 5, 9, 10, 12 + +52. Wang, Z., Bovik, A.C., Sheikh, H.R., Simoncelli, E.P.: Image quality assessment: from error visibility to structural similarity. IEEE TIP 13(4), 600-612 (2004) 9 +53. Weiss, K., Khoshgoftaar, T.M., Wang, D.: A survey of transfer learning. Journal of Big data 3(1), 1-40 (2016) 2 +54. Xian, K., Zhang, J., Wang, O., Mai, L., Lin, Z., Cao, Z.: Structure-guided ranking loss for single image depth prediction. In: CVPR. pp. 611-620 (2020) 3, 7, 8 +55. Yang, J., Alvarez, J.M., Liu, M.: Self-supervised learning of depth inference for multi-view stereo. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 7526-7534 (2021) 4 +56. Yang, L., Kang, B., Huang, Z., Xu, X., Feng, J., Zhao, H.: Depth anything: Unleashing the power of large-scale unlabeled data. arXiv preprint arXiv:2401.10891 (2024) 1, 3, 4, 6, 7, 8, 12 +57. Yen, Y.T., Lu, C.N., Chiu, W.C., Tsai, Y.H.: 3d-pl: Domain adaptive depth estimation with 3d-aware pseudo-labeling. In: European Conference on Computer Vision. pp. 710-728. Springer (2022) 4, 6 +58. Yeshwanth, C., Liu, Y.C., Nießner, M., Dai, A.: Scannet++: A high-fidelity dataset of 3d indoor scenes. In: ICCV. pp. 12-22 (2023) 3, 5, 6, 10 +59. Zhang, L., Rao, A., Agrawala, M.: Adding conditional control to text-to-image diffusion models. In: ICCV. pp. 3836-3847 (2023) 1 +60. Zhao, S., Fu, H., Gong, M., Tao, D.: Geometry-aware symmetric domain adaptation for monocular depth estimation. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 9788-9798 (2019) 3 +61. Zhao, W., Rao, Y., Liu, Z., Liu, B., Zhou, J., Lu, J.: Unleashing text-to-image diffusion models for visual perception. ICCV (2023) 1 +62. Zhao, X., Schulter, S., Sharma, G., Tsai, Y.H., Chandraker, M., Wu, Y.: Object detection with a unified label space from multiple datasets. In: Computer Vision-ECCV 2020: 16th European Conference, Glasgow, UK, August 23-28, 2020, Proceedings, Part XIV 16. pp. 178-193. Springer (2020) 3 +63. Zhao, Z., Zhang, J., Xu, S., Lin, Z., Pfister, H.: Discrete cosine transform network for guided depth map super-resolution. In: CVPR. pp. 5697-5707 (2022) 3 +64. Zheng, C., Cham, T.J., Cai, J.: T2net: Synthetic-to-realistic translation for solving single-image depth estimation tasks. In: Proceedings of the European conference on computer vision (ECCV). pp. 767-783 (2018) 3 +65. Zhong, Z., Liu, X., Jiang, J., Zhao, D., Ji, X.: Guided depth map super-resolution: A survey. ACM Computing Surveys (2023) 3 +66. Zou, Y., Yu, Z., Kumar, B., Wang, J.: Unsupervised domain adaptation for semantic segmentation via class-balanced self-training. In: Proceedings of the European conference on computer vision (ECCV). pp. 289-305 (2018) 3 \ No newline at end of file diff --git a/2024/PatchRefiner_ Leveraging Synthetic Data for Real-Domain High-Resolution Monocular Metric Depth Estimation/images.zip b/2024/PatchRefiner_ Leveraging Synthetic Data for Real-Domain High-Resolution Monocular Metric Depth Estimation/images.zip new file mode 100644 index 0000000000000000000000000000000000000000..aee5277d66801f0f71ace1d7c2a9ac310b8fc713 --- /dev/null +++ b/2024/PatchRefiner_ Leveraging Synthetic Data for Real-Domain High-Resolution Monocular Metric Depth Estimation/images.zip @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:285fd5f2989c4e08df8ec0a45a076fe0e585a9d4c5e2b127b2e162dde36eee25 +size 710961 diff --git a/2024/PatchRefiner_ Leveraging Synthetic Data for Real-Domain High-Resolution Monocular Metric Depth Estimation/layout.json b/2024/PatchRefiner_ Leveraging Synthetic Data for Real-Domain High-Resolution Monocular Metric Depth Estimation/layout.json new file mode 100644 index 0000000000000000000000000000000000000000..bf8fcc5ef5ebe7f276689483903d0680b07f2106 --- /dev/null +++ b/2024/PatchRefiner_ Leveraging Synthetic Data for Real-Domain High-Resolution Monocular Metric Depth Estimation/layout.json @@ -0,0 +1,11641 @@ +{ + "pdf_info": [ + { + "para_blocks": [ + { + "bbox": [ + 136, + 111, + 477, + 166 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 136, + 111, + 477, + 166 + ], + "spans": [ + { + "bbox": [ + 136, + 111, + 477, + 166 + ], + "type": "text", + "content": "PatchRefiner: Leveraging Synthetic Data for Real-Domain High-Resolution Monocular Metric Depth Estimation" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 205, + 186, + 408, + 199 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 205, + 186, + 408, + 199 + ], + "spans": [ + { + "bbox": [ + 205, + 186, + 408, + 199 + ], + "type": "text", + "content": "Zhenyu Li, Shariq Farooq Bhat, Peter Wonka" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 177, + 207, + 436, + 231 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 177, + 207, + 436, + 231 + ], + "spans": [ + { + "bbox": [ + 177, + 207, + 436, + 231 + ], + "type": "text", + "content": "King Abdullah University of Science and Technology (KAUST) zhenyu.li.1@kaust.edu.sa" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 160, + 258, + 455, + 467 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 160, + 258, + 455, + 467 + ], + "spans": [ + { + "bbox": [ + 160, + 258, + 455, + 467 + ], + "type": "text", + "content": "Abstract. This paper introduces PatchRefiner, an advanced framework for metric single image depth estimation aimed at high-resolution real-domain inputs. While depth estimation is crucial for applications such as autonomous driving, 3D generative modeling, and 3D reconstruction, achieving accurate high-resolution depth in real-world scenarios is challenging due to the constraints of existing architectures and the scarcity of detailed real-world depth data. PatchRefiner adopts a tile-based methodology, reconceptualizing high-resolution depth estimation as a refinement process, which results in notable performance enhancements. Utilizing a pseudo-labeling strategy that leverages synthetic data, PatchRefiner incorporates a Detail and Scale Disentangling (DSD) loss to enhance detail capture while maintaining scale accuracy, thus facilitating the effective transfer of knowledge from synthetic to real-world data. Our extensive evaluations demonstrate PatchRefiner's superior performance, significantly outperforming existing benchmarks on the Unreal4KStereo dataset by " + }, + { + "bbox": [ + 160, + 258, + 455, + 467 + ], + "type": "inline_equation", + "content": "18.1\\%" + }, + { + "bbox": [ + 160, + 258, + 455, + 467 + ], + "type": "text", + "content": " in terms of the root mean squared error (RMSE) and showing marked improvements in detail accuracy and consistent scale estimation on diverse real-world datasets like CityScape, ScanNet++, and ETH3D." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 160, + 478, + 453, + 489 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 160, + 478, + 453, + 489 + ], + "spans": [ + { + "bbox": [ + 160, + 478, + 453, + 489 + ], + "type": "text", + "content": "Keywords: High-Resolution Metric Depth Estimation " + }, + { + "bbox": [ + 160, + 478, + 453, + 489 + ], + "type": "inline_equation", + "content": "\\cdot" + }, + { + "bbox": [ + 160, + 478, + 453, + 489 + ], + "type": "text", + "content": " Synthetic Data" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 132, + 509, + 230, + 521 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 509, + 230, + 521 + ], + "spans": [ + { + "bbox": [ + 132, + 509, + 230, + 521 + ], + "type": "text", + "content": "1 Introduction" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 130, + 533, + 482, + 629 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 533, + 482, + 629 + ], + "spans": [ + { + "bbox": [ + 130, + 533, + 482, + 629 + ], + "type": "text", + "content": "This paper delves into the field of metric single-image depth estimation, focusing on high-resolution inputs from the real domain. High-resolution depth estimation plays a key role in autonomous driving, augmented reality, content creation, and 3D reconstruction [2,11,30,59]. Despite significant progress, the high-resolution depth estimation in real-world scenarios remains daunting. This challenge is primarily due to the resolution limitations inherent in most state-of-the-art depth estimation architectures [2, 56, 61] and the scarcity of high-quality real-world depth datasets." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 130, + 630, + 482, + 666 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 630, + 482, + 666 + ], + "spans": [ + { + "bbox": [ + 130, + 630, + 482, + 666 + ], + "type": "text", + "content": "The current state-of-the-art in high-resolution depth estimation, PatchFusion [27, 34, 43], employs a tile-based strategy to navigate the resolution constraints, posing the task as a fusion process of coarse and fine depth estimations." + } + ] + } + ], + "index": 7 + } + ], + "discarded_blocks": [], + "page_size": [ + 612, + 792 + ], + "page_idx": 0 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 137, + 115, + 190, + 184 + ], + "blocks": [ + { + "bbox": [ + 137, + 115, + 190, + 184 + ], + "lines": [ + { + "bbox": [ + 137, + 115, + 190, + 184 + ], + "spans": [ + { + "bbox": [ + 137, + 115, + 190, + 184 + ], + "type": "image", + "image_path": "131911c50077e8a56732c678d77d0484938b30bd000a405316e52ed57133234a.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 141, + 189, + 192, + 199 + ], + "lines": [ + { + "bbox": [ + 141, + 189, + 192, + 199 + ], + "spans": [ + { + "bbox": [ + 141, + 189, + 192, + 199 + ], + "type": "text", + "content": "(a) LR Depth" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_caption" + } + ], + "index": 2 + }, + { + "type": "image", + "bbox": [ + 209, + 116, + 340, + 185 + ], + "blocks": [ + { + "bbox": [ + 209, + 116, + 340, + 185 + ], + "lines": [ + { + "bbox": [ + 209, + 116, + 340, + 185 + ], + "spans": [ + { + "bbox": [ + 209, + 116, + 340, + 185 + ], + "type": "image", + "image_path": "8cfde7296b386600ec18cc38cf4e70a1db57087fe160e66543be58f1c6f3df81.jpg" + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 223, + 189, + 328, + 198 + ], + "lines": [ + { + "bbox": [ + 223, + 189, + 328, + 198 + ], + "spans": [ + { + "bbox": [ + 223, + 189, + 328, + 198 + ], + "type": "text", + "content": "(b) Fusion-Based Framework" + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "image_caption" + } + ], + "index": 4 + }, + { + "type": "image", + "bbox": [ + 362, + 116, + 467, + 184 + ], + "blocks": [ + { + "bbox": [ + 362, + 116, + 467, + 184 + ], + "lines": [ + { + "bbox": [ + 362, + 116, + 467, + 184 + ], + "spans": [ + { + "bbox": [ + 362, + 116, + 467, + 184 + ], + "type": "image", + "image_path": "b48e13f85a3b73ccff89df2117da75a4ef32fdab0fc767f7398d3c4e06a4a624.jpg" + } + ] + } + ], + "index": 6, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 360, + 189, + 465, + 198 + ], + "lines": [ + { + "bbox": [ + 360, + 189, + 465, + 198 + ], + "spans": [ + { + "bbox": [ + 360, + 189, + 465, + 198 + ], + "type": "text", + "content": "(c) Refiner-Based Framework" + } + ] + } + ], + "index": 7, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 130, + 208, + 482, + 252 + ], + "lines": [ + { + "bbox": [ + 130, + 208, + 482, + 252 + ], + "spans": [ + { + "bbox": [ + 130, + 208, + 482, + 252 + ], + "type": "text", + "content": "Fig. 1: Framework Comparison. (a) Low resolution depth estimation framework with single forward pass. (b) Fusion-based high-resolution framework combining the best of coarse and fine depth predictions [27, 34]. (c) Our refiner-based framework predicts a residual to refine the coarse prediction." + } + ] + } + ], + "index": 8, + "angle": 0, + "type": "image_caption" + } + ], + "index": 6 + }, + { + "bbox": [ + 130, + 277, + 482, + 361 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 277, + 482, + 361 + ], + "spans": [ + { + "bbox": [ + 130, + 277, + 482, + 361 + ], + "type": "text", + "content": "Because of the scarcity of real high-resolution depth datasets, PatchFusion resorts to training on a synthetic 4K dataset [27, 51]. There are two limitations of PatchFusion that we would like to improve upon: 1) It utilizes a three-step training process, which is not only time-consuming and expensive but also risks the framework achieving the stage-wise local optima and limits the potential performance gains from end-to-end learning. 2) PatchFusion demonstrates poor generalization to the real domain." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 130, + 363, + 482, + 506 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 363, + 482, + 506 + ], + "spans": [ + { + "bbox": [ + 130, + 363, + 482, + 506 + ], + "type": "text", + "content": "The poor synthetic-to-real generalization is a long-standing problem in general [12,53]. In the context of depth estimation, the difference in the metric scale and depth distributions of synthetic datasets and the real domain further exacerbates the domain shift. This results in particularly subpar scale accuracy of depth models on real data when trained on synthetic datasets. On the other hand, the real datasets are not only low resolution but also often have missing ground truth values due to sensor constraints, occlusion, etc (see Fig. 3 and Sec. 3.4). This adds to the inability of depth models to capture sharp details when trained on real datasets. Thus, one is confronted with a dilemma - Training on real datasets leads to good scale accuracy but poor high-frequency details, while training on synthetic datasets leads to sharper results but poor scale performance on real images." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 130, + 508, + 482, + 543 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 508, + 482, + 543 + ], + "spans": [ + { + "bbox": [ + 130, + 508, + 482, + 543 + ], + "type": "text", + "content": "We introduce PatchRefiner, a novel framework that reformulates high-resolution depth estimation as a process of refining coarse depth. We propose improvements on two levels:" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 130, + 544, + 482, + 592 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 544, + 482, + 592 + ], + "spans": [ + { + "bbox": [ + 130, + 544, + 482, + 592 + ], + "type": "text", + "content": "First, unlike direct depth regression approaches [27,39,43], PatchRefiner utilizes a frozen coarse depth model and enhances the quality by predicting residual depth for refinement. This approach not only streamlines model training but also markedly enhances performance." + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 130, + 594, + 482, + 665 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 594, + 482, + 665 + ], + "spans": [ + { + "bbox": [ + 130, + 594, + 482, + 665 + ], + "type": "text", + "content": "Second, we propose a method to exploit the best of both worlds to solve the above synthetic-real dilemma. We employ a teacher-student framework, leveraging the sharpness of synthetic data while learning the scale from real data. The teacher model, pre-trained on synthetic data, generates pseudo labels for real-domain training samples. Recognizing that these pseudo labels offer detailed features albeit with scale inaccuracies, we introduce the Detail and Scale Dis-" + } + ] + } + ], + "index": 13 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "spans": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "type": "text", + "content": "2" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 165, + 91, + 354, + 102 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 165, + 91, + 354, + 102 + ], + "spans": [ + { + "bbox": [ + 165, + 91, + 354, + 102 + ], + "type": "text", + "content": "Zhenyu Li, Shariq Farooq Bhat, Peter Wonka" + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 1 + }, + { + "para_blocks": [ + { + "bbox": [ + 130, + 116, + 479, + 163 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 116, + 479, + 163 + ], + "spans": [ + { + "bbox": [ + 130, + 116, + 479, + 163 + ], + "type": "text", + "content": "entangling (DSD) loss. This loss integrates ranking supervision and scale-shift invariance, drawing inspiration from recent advances in relative depth estimation [5, 42, 54, 56]. It leads to a framework capable of delivering high-resolution depth estimates with both precise scale and sharp details in real-world settings." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 130, + 164, + 480, + 283 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 164, + 480, + 283 + ], + "spans": [ + { + "bbox": [ + 130, + 164, + 480, + 283 + ], + "type": "text", + "content": "Our evaluation of PatchRefiner on the Unreal4KStereo synthetic dataset [51] demonstrates a substantial improvement over the current state-of-the-art, reducing RMSE by " + }, + { + "bbox": [ + 130, + 164, + 480, + 283 + ], + "type": "inline_equation", + "content": "18.1\\%" + }, + { + "bbox": [ + 130, + 164, + 480, + 283 + ], + "type": "text", + "content": " and REL by " + }, + { + "bbox": [ + 130, + 164, + 480, + 283 + ], + "type": "inline_equation", + "content": "15.7\\%" + }, + { + "bbox": [ + 130, + 164, + 480, + 283 + ], + "type": "text", + "content": ". Further, we assess the framework's efficacy in leveraging synthetic data across diverse real-world datasets, including CityScape [9] (outdoor, stereo), ScanNet++ [58] (indoor, LiDAR and reconstruction), and ETH3D [46] (mixed, LiDAR). Our findings reveal notable enhancements in depth boundary delineation (e.g., a " + }, + { + "bbox": [ + 130, + 164, + 480, + 283 + ], + "type": "inline_equation", + "content": "19.2\\%" + }, + { + "bbox": [ + 130, + 164, + 480, + 283 + ], + "type": "text", + "content": " increase in boundary recall on CityScape) while maintaining accurate scale estimation, showcasing the framework's adaptability and effectiveness across varying settings and sensor technologies." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 132, + 304, + 237, + 316 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 304, + 237, + 316 + ], + "spans": [ + { + "bbox": [ + 132, + 304, + 237, + 316 + ], + "type": "text", + "content": "2 Related Work" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 130, + 330, + 480, + 484 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 330, + 480, + 484 + ], + "spans": [ + { + "bbox": [ + 130, + 330, + 480, + 484 + ], + "type": "text", + "content": "High-Resolution Monocular Metric Depth Estimation. Monocular depth estimation has achieved tremendous progress [1, 11, 13, 15, 24, 29, 30]. Current SOTA approaches often employ complex network architectures yet grapple with the limitations imposed by their input resolution [2, 56]. This stands in stark contrast to the advancements in modern imaging devices that capture images at increasing resolutions, and the growing demand among users for high-resolution depth estimation. Initial efforts to address this gap included the use of Guided Depth Super-Resolution (GDSR) [19,33,63,65] and Implicit Functions [6,35]. Recently, Tile-Based Methods have emerged as a potent strategy for high-resolution depth estimation [27,34,43], segmenting images into patches for individual processing before reassembling them into a comprehensive depth map. This paper extends the tile-based approach, seeking to elevate the quality of depth estimation further." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 130, + 486, + 481, + 628 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 486, + 481, + 628 + ], + "spans": [ + { + "bbox": [ + 130, + 486, + 481, + 628 + ], + "type": "text", + "content": "Synthetic Data for Depth Estimation. The challenge of acquiring high-quality, real-domain data for high-resolution depth training has prompted the use of synthetic datasets [41]. Traditionally, synthetic data has been employed within unsupervised domain adaptation frameworks, utilizing labeled synthetic and unlabeled real-domain data to enhance depth estimation accuracy on real-world images [7,22,23,32,60,64]. Techniques vary from pixel-level style transfer and image translation to feature-level adversarial learning [7,23,60,64], with some methods integrating additional information such as stereo pairs or segmentation maps for enhanced adaptation [22,32,60]. Contrasting with these approaches, our work explores a practical scenario where labeled data from both synthetic and real domains are leveraged to improve real-world, high-resolution depth estimation, delving into a relatively underexplored application of synthetic data." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 130, + 630, + 481, + 665 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 630, + 481, + 665 + ], + "spans": [ + { + "bbox": [ + 130, + 630, + 481, + 665 + ], + "type": "text", + "content": "Pseudo-Labeling for Depth Estimation. Pseudo-labeling, a cornerstone of semi-supervised learning, has been widely applied across various domains, including classification [3,17,40,44,50] and scene understanding [8,26,28,36,37,47,62," + } + ] + } + ], + "index": 7 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 393, + 91, + 447, + 101 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 393, + 91, + 447, + 101 + ], + "spans": [ + { + "bbox": [ + 393, + 91, + 447, + 101 + ], + "type": "text", + "content": "PatchRefiner" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 474, + 92, + 480, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 474, + 92, + 480, + 100 + ], + "spans": [ + { + "bbox": [ + 474, + 92, + 480, + 100 + ], + "type": "text", + "content": "3" + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 2 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 152, + 116, + 462, + 233 + ], + "blocks": [ + { + "bbox": [ + 152, + 116, + 462, + 233 + ], + "lines": [ + { + "bbox": [ + 152, + 116, + 462, + 233 + ], + "spans": [ + { + "bbox": [ + 152, + 116, + 462, + 233 + ], + "type": "image", + "image_path": "fc5f3c6def0d0be422667eafcc0ca334f82bc0e18f9c873309a33f77b5f3c887.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 130, + 242, + 482, + 299 + ], + "lines": [ + { + "bbox": [ + 130, + 242, + 482, + 299 + ], + "spans": [ + { + "bbox": [ + 130, + 242, + 482, + 299 + ], + "type": "text", + "content": "Fig. 2: Architecture Illustration. PatchRefiner contains a pre-trained frozen coarse depth estimation model " + }, + { + "bbox": [ + 130, + 242, + 482, + 299 + ], + "type": "inline_equation", + "content": "\\mathcal{N}_c" + }, + { + "bbox": [ + 130, + 242, + 482, + 299 + ], + "type": "text", + "content": " and a refiner model " + }, + { + "bbox": [ + 130, + 242, + 482, + 299 + ], + "type": "inline_equation", + "content": "\\mathcal{N}_r" + }, + { + "bbox": [ + 130, + 242, + 482, + 299 + ], + "type": "text", + "content": " predicts residual depth map " + }, + { + "bbox": [ + 130, + 242, + 482, + 299 + ], + "type": "inline_equation", + "content": "\\mathcal{D}_r" + }, + { + "bbox": [ + 130, + 242, + 482, + 299 + ], + "type": "text", + "content": " to refine the coarse depth " + }, + { + "bbox": [ + 130, + 242, + 482, + 299 + ], + "type": "inline_equation", + "content": "\\mathcal{D}_c" + }, + { + "bbox": [ + 130, + 242, + 482, + 299 + ], + "type": "text", + "content": ". The refiner contains one base depth model " + }, + { + "bbox": [ + 130, + 242, + 482, + 299 + ], + "type": "inline_equation", + "content": "\\mathcal{N}_d" + }, + { + "bbox": [ + 130, + 242, + 482, + 299 + ], + "type": "text", + "content": " that has the same architecture as " + }, + { + "bbox": [ + 130, + 242, + 482, + 299 + ], + "type": "inline_equation", + "content": "\\mathcal{N}_c" + }, + { + "bbox": [ + 130, + 242, + 482, + 299 + ], + "type": "text", + "content": ", and a light-weight decoder to aggregate information and make the final prediction." + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_caption" + } + ], + "index": 2 + }, + { + "bbox": [ + 130, + 321, + 483, + 441 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 321, + 483, + 441 + ], + "spans": [ + { + "bbox": [ + 130, + 321, + 483, + 441 + ], + "type": "text", + "content": "66], to extrapolate knowledge from labeled data to unlabeled datasets. In depth estimation, pseudo-labeling often serves to provide supplementary supervision in unsupervised domain adaptation settings [32, 55, 57]. While recent state-of-the-art methods like Depth-Anything utilize pseudo-labeling to enhance model generalization [56], these techniques predominantly aim to refine or enhance the pseudo labels themselves. Our approach diverges by utilizing real-domain data with accurate depth labels, focusing on a novel Detail and Scale Disentangling loss. This loss mechanism uniquely leverages the detailed insights from pseudo labels to enrich real-domain depth estimation without compromising the scale accuracy derived from real ground-truth data." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 132, + 459, + 202, + 472 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 459, + 202, + 472 + ], + "spans": [ + { + "bbox": [ + 132, + 459, + 202, + 472 + ], + "type": "text", + "content": "3 Method" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 130, + 484, + 485, + 534 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 484, + 485, + 534 + ], + "spans": [ + { + "bbox": [ + 130, + 484, + 485, + 534 + ], + "type": "text", + "content": "In this section, we first present the overall PatchRefiner framework in Sec. 3.1. Then, we introduce the limitation of adopting real-domain data for high-resolution depth estimation 3.2 and the proposed teacher-student framework in Sec. 3.3 with the Detail and Scale Disentangling loss (DSD) in Sec. 3.4." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 131, + 550, + 287, + 562 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 131, + 550, + 287, + 562 + ], + "spans": [ + { + "bbox": [ + 131, + 550, + 287, + 562 + ], + "type": "text", + "content": "3.1 PatchRefiner Framework" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 130, + 570, + 482, + 629 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 570, + 482, + 629 + ], + "spans": [ + { + "bbox": [ + 130, + 570, + 482, + 629 + ], + "type": "text", + "content": "PatchRefiner follows the tile-based strategy to address the prohibitive memory and computational demands for high resolutions such as 4K [27, 34]. However, recognizing the limitations of existing models, we propose a simplified two-step approach for high-resolution depth estimation: (i) Coarse Scale-Aware Estimation, and (ii) Fine-Grained Depth Refinement, shown in Fig. 2." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 130, + 630, + 482, + 666 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 630, + 482, + 666 + ], + "spans": [ + { + "bbox": [ + 130, + 630, + 482, + 666 + ], + "type": "text", + "content": "(i) Coarse Scale-Aware Estimation: The foundation of PatchRefiner is the Coarse Depth Estimation network, " + }, + { + "bbox": [ + 130, + 630, + 482, + 666 + ], + "type": "inline_equation", + "content": "\\mathcal{N}_c" + }, + { + "bbox": [ + 130, + 630, + 482, + 666 + ], + "type": "text", + "content": ", which processes downsampled versions of the input images to produce a global depth prediction, " + }, + { + "bbox": [ + 130, + 630, + 482, + 666 + ], + "type": "inline_equation", + "content": "\\mathbf{D}_c" + }, + { + "bbox": [ + 130, + 630, + 482, + 666 + ], + "type": "text", + "content": ". Similar to" + } + ] + } + ], + "index": 9 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "spans": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "type": "text", + "content": "4" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 165, + 91, + 354, + 102 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 165, + 91, + 354, + 102 + ], + "spans": [ + { + "bbox": [ + 165, + 91, + 354, + 102 + ], + "type": "text", + "content": "Zhenyu Li, Shariq Farooq Bhat, Peter Wonka" + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 3 + }, + { + "para_blocks": [ + { + "bbox": [ + 130, + 116, + 479, + 163 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 116, + 479, + 163 + ], + "spans": [ + { + "bbox": [ + 130, + 116, + 479, + 163 + ], + "type": "text", + "content": "previous work [27, 34], this step is crucial for establishing a baseline depth map that captures the scene's overall structure and depth consistency, albeit without high-resolution details. Moreover, " + }, + { + "bbox": [ + 130, + 116, + 479, + 163 + ], + "type": "inline_equation", + "content": "\\mathcal{N}_c" + }, + { + "bbox": [ + 130, + 116, + 479, + 163 + ], + "type": "text", + "content": " can be an arbitrary depth estimation model and it is frozen after the first step of training." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 130, + 164, + 480, + 223 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 164, + 480, + 223 + ], + "spans": [ + { + "bbox": [ + 130, + 164, + 480, + 223 + ], + "type": "text", + "content": "(ii) Depth Refinement Process: Different from the conventional approach of using a separate fine depth network and a fusion mechanism [27, 39], our framework introduces a unified refinement network, " + }, + { + "bbox": [ + 130, + 164, + 480, + 223 + ], + "type": "inline_equation", + "content": "\\mathcal{N}_r" + }, + { + "bbox": [ + 130, + 164, + 480, + 223 + ], + "type": "text", + "content": ". This network is designed to refine the coarse depth map by focusing on the recovery of lost details and enhancing depth precision at a patch-wise level." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 130, + 224, + 480, + 296 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 224, + 480, + 296 + ], + "spans": [ + { + "bbox": [ + 130, + 224, + 480, + 296 + ], + "type": "text", + "content": "The input to " + }, + { + "bbox": [ + 130, + 224, + 480, + 296 + ], + "type": "inline_equation", + "content": "\\mathcal{N}_r" + }, + { + "bbox": [ + 130, + 224, + 480, + 296 + ], + "type": "text", + "content": " is the cropped original image " + }, + { + "bbox": [ + 130, + 224, + 480, + 296 + ], + "type": "inline_equation", + "content": "I" + }, + { + "bbox": [ + 130, + 224, + 480, + 296 + ], + "type": "text", + "content": ", which is processed by the base depth model " + }, + { + "bbox": [ + 130, + 224, + 480, + 296 + ], + "type": "inline_equation", + "content": "\\mathcal{N}_d" + }, + { + "bbox": [ + 130, + 224, + 480, + 296 + ], + "type": "text", + "content": " with the same architecture as " + }, + { + "bbox": [ + 130, + 224, + 480, + 296 + ], + "type": "inline_equation", + "content": "\\mathcal{N}_c" + }, + { + "bbox": [ + 130, + 224, + 480, + 296 + ], + "type": "text", + "content": " in the refiner module. Then, we collect " + }, + { + "bbox": [ + 130, + 224, + 480, + 296 + ], + "type": "inline_equation", + "content": "L" + }, + { + "bbox": [ + 130, + 224, + 480, + 296 + ], + "type": "text", + "content": "-level multi-scale immediate features from both " + }, + { + "bbox": [ + 130, + 224, + 480, + 296 + ], + "type": "inline_equation", + "content": "\\mathcal{N}_d" + }, + { + "bbox": [ + 130, + 224, + 480, + 296 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 130, + 224, + 480, + 296 + ], + "type": "inline_equation", + "content": "\\mathcal{N}_c" + }, + { + "bbox": [ + 130, + 224, + 480, + 296 + ], + "type": "text", + "content": ", denoting them as " + }, + { + "bbox": [ + 130, + 224, + 480, + 296 + ], + "type": "inline_equation", + "content": "\\mathcal{F}_d = \\{f_d^i\\}_{i=1}^L" + }, + { + "bbox": [ + 130, + 224, + 480, + 296 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 130, + 224, + 480, + 296 + ], + "type": "inline_equation", + "content": "\\tilde{\\mathcal{F}}_c = \\{\\tilde{f}_c^i\\}_{i=1}^L" + }, + { + "bbox": [ + 130, + 224, + 480, + 296 + ], + "type": "text", + "content": ", respectively. Following [27], we apply the roi [16] operation to fetch features of the corresponding cropped area as " + }, + { + "bbox": [ + 130, + 224, + 480, + 296 + ], + "type": "inline_equation", + "content": "\\tilde{f}_c^i = \\text{roi}(f_c^i)" + }, + { + "bbox": [ + 130, + 224, + 480, + 296 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 130, + 297, + 479, + 331 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 297, + 479, + 331 + ], + "spans": [ + { + "bbox": [ + 130, + 297, + 479, + 331 + ], + "type": "text", + "content": "Next, we adopt a lightweight decoder to obtain high-resolution predictions. Given " + }, + { + "bbox": [ + 130, + 297, + 479, + 331 + ], + "type": "inline_equation", + "content": "\\mathcal{F}_d" + }, + { + "bbox": [ + 130, + 297, + 479, + 331 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 130, + 297, + 479, + 331 + ], + "type": "inline_equation", + "content": "\\bar{\\mathcal{F}}_c" + }, + { + "bbox": [ + 130, + 297, + 479, + 331 + ], + "type": "text", + "content": ", we aggregate them with concatenation operators (Cat) following by simple convolutional blocks (CB):" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 261, + 343, + 480, + 357 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 261, + 343, + 480, + 357 + ], + "spans": [ + { + "bbox": [ + 261, + 343, + 480, + 357 + ], + "type": "interline_equation", + "content": "f _ {r} ^ {i} = \\operatorname {C B} \\left(\\operatorname {C a t} \\left(f _ {d} ^ {i}, \\tilde {f} _ {c} ^ {i}\\right)\\right), \\tag {1}", + "image_path": "c0c8ceb8c353d644bee976c69ff959bc1fef0c2ec88707d1dfa40ea7fc0b74a3.jpg" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 130, + 362, + 480, + 434 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 362, + 480, + 434 + ], + "spans": [ + { + "bbox": [ + 130, + 362, + 480, + 434 + ], + "type": "text", + "content": "The output feature set " + }, + { + "bbox": [ + 130, + 362, + 480, + 434 + ], + "type": "inline_equation", + "content": "\\mathcal{F}_r = \\{f_r^i\\}_{i=1}^L" + }, + { + "bbox": [ + 130, + 362, + 480, + 434 + ], + "type": "text", + "content": " is then fed to a successive series of up-sampling layers [25], in order to construct the residual depth map " + }, + { + "bbox": [ + 130, + 362, + 480, + 434 + ], + "type": "inline_equation", + "content": "\\mathbf{D}_r" + }, + { + "bbox": [ + 130, + 362, + 480, + 434 + ], + "type": "text", + "content": " at the input resolution. Associated with skip-connections, they form our refiner decoder. Further details about the architecture with immediate feature shapes are described in the supplementary material. Finally, we calculate the final patchwise depth estimation as " + }, + { + "bbox": [ + 130, + 362, + 480, + 434 + ], + "type": "inline_equation", + "content": "\\mathbf{D} = \\mathsf{roi}(\\mathbf{D}_c) + \\mathbf{D}_r" + }, + { + "bbox": [ + 130, + 362, + 480, + 434 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 131, + 453, + 436, + 464 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 131, + 453, + 436, + 464 + ], + "spans": [ + { + "bbox": [ + 131, + 453, + 436, + 464 + ], + "type": "text", + "content": "3.2 Limitation of Real-Domain Depth Estimation Datasets" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 130, + 474, + 479, + 509 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 474, + 479, + 509 + ], + "spans": [ + { + "bbox": [ + 130, + 474, + 479, + 509 + ], + "type": "text", + "content": "Our second goal in this paper is to train a real-domain high-resolution depth estimation model with both synthetic dataset " + }, + { + "bbox": [ + 130, + 474, + 479, + 509 + ], + "type": "inline_equation", + "content": "\\mathcal{S}" + }, + { + "bbox": [ + 130, + 474, + 479, + 509 + ], + "type": "text", + "content": " and real-domain dataset " + }, + { + "bbox": [ + 130, + 474, + 479, + 509 + ], + "type": "inline_equation", + "content": "\\mathcal{R}" + }, + { + "bbox": [ + 130, + 474, + 479, + 509 + ], + "type": "text", + "content": ". Our main insight is to distinguish between scale errors and boundary errors." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 130, + 510, + 480, + 604 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 510, + 480, + 604 + ], + "spans": [ + { + "bbox": [ + 130, + 510, + 480, + 604 + ], + "type": "text", + "content": "In the field of high-resolution depth estimation, the prevailing state-of-the-art methodology [27] trains models using synthetic datasets, which offers paired high-resolution images and corresponding dense high-resolution depth ground-truth maps [18,51]. This synthetic training regime, while beneficial in a controlled setting, introduces significant challenges when models are applied to real-world data due to the intrinsic domain gap between synthetic and real-world environments. This gap often manifests as substantial scale errors during inference on real-domain datasets as shown in Tab. 2." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 130, + 606, + 480, + 665 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 606, + 480, + 665 + ], + "spans": [ + { + "bbox": [ + 130, + 606, + 480, + 665 + ], + "type": "text", + "content": "Addressing this issue directly by training on real-domain datasets presents its own set of challenges. Real-world high-resolution depth datasets are scarce and typically constrained by limited resolution and unavailable missing ground-truth pixels. These limitations stem from the methods employed in generating real-world depth annotations, such as Kinect [20, 48, 49], LiDAR [14, 46, 58], or" + } + ] + } + ], + "index": 11 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 393, + 91, + 447, + 100 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 393, + 91, + 447, + 100 + ], + "spans": [ + { + "bbox": [ + 393, + 91, + 447, + 100 + ], + "type": "text", + "content": "PatchRefiner" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 474, + 92, + 481, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 474, + 92, + 481, + 100 + ], + "spans": [ + { + "bbox": [ + 474, + 92, + 481, + 100 + ], + "type": "text", + "content": "5" + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 4 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 134, + 114, + 309, + 273 + ], + "blocks": [ + { + "bbox": [ + 134, + 114, + 309, + 273 + ], + "lines": [ + { + "bbox": [ + 134, + 114, + 309, + 273 + ], + "spans": [ + { + "bbox": [ + 134, + 114, + 309, + 273 + ], + "type": "image", + "image_path": "fe6917163c2e50005745bdc52f8546710cd5931788b69dc1a8829b4ae0d8c133.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 130, + 281, + 482, + 314 + ], + "lines": [ + { + "bbox": [ + 130, + 281, + 482, + 314 + ], + "spans": [ + { + "bbox": [ + 130, + 281, + 482, + 314 + ], + "type": "text", + "content": "Fig.3: Visualization of Real-Domain Data Pairs. Points lacking ground-truth data are depicted in gray. Due to sparse annotations near edges, models trained on real-domain data exhibit blurred boundary estimations." + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "image_caption" + } + ], + "index": 2 + }, + { + "type": "image", + "bbox": [ + 310, + 114, + 480, + 274 + ], + "blocks": [ + { + "bbox": [ + 310, + 114, + 480, + 274 + ], + "lines": [ + { + "bbox": [ + 310, + 114, + 480, + 274 + ], + "spans": [ + { + "bbox": [ + 310, + 114, + 480, + 274 + ], + "type": "image", + "image_path": "87ecd62ae9318cd8e358b71389c009ef71e1a69a508d615974efba15c63abe2b.jpg" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_body" + } + ], + "index": 3 + }, + { + "bbox": [ + 130, + 340, + 479, + 363 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 340, + 479, + 363 + ], + "spans": [ + { + "bbox": [ + 130, + 340, + 479, + 363 + ], + "type": "text", + "content": "stereo vision techniques [9, 45], each with their inherent drawbacks as shown in Fig. 3." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 130, + 365, + 482, + 484 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 365, + 482, + 484 + ], + "spans": [ + { + "bbox": [ + 130, + 365, + 482, + 484 + ], + "type": "text", + "content": "For instance, depth maps generated using Kinect technology are confined to a resolution of " + }, + { + "bbox": [ + 130, + 365, + 482, + 484 + ], + "type": "inline_equation", + "content": "640 \\times 480" + }, + { + "bbox": [ + 130, + 365, + 482, + 484 + ], + "type": "text", + "content": " [48, 49], insufficient for high-resolution depth estimation [27]. LiDAR-based depth maps, while useful, tend to be sparse [14, 46] or limited in resolution (e.g., 256x192 in Scannet++ [58]), and the process of reconstructing high-resolution dense depth maps from LiDAR point clouds is fragile with cascade errors and omissions [58]. Stereo vision techniques, on the other hand, can also introduce missing values around object boundaries due to rectification transformations [9, 45]. It is important to realize that existing high-resolution real datasets with missing values around edges do not help in reducing boundary errors." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 130, + 486, + 482, + 570 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 486, + 482, + 570 + ], + "spans": [ + { + "bbox": [ + 130, + 486, + 482, + 570 + ], + "type": "text", + "content": "These limitations highlight the inherent difficulties in employing real-world datasets for training high-resolution depth estimators [27, 39]. The lack of high-quality, high-resolution ground-truth depth maps in the real domain makes it challenging to train models that can accurately predict sharp depth around fine object boundaries [41]. Our main idea is to devise a training strategy that can improve the scale errors when fine-tuning on real data, while maintaining high-resolution information around edges to minimize boundary errors." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 131, + 592, + 302, + 605 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 131, + 592, + 302, + 605 + ], + "spans": [ + { + "bbox": [ + 131, + 592, + 302, + 605 + ], + "type": "text", + "content": "3.3 Overall Pipeline Illustration" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 130, + 617, + 481, + 666 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 617, + 481, + 666 + ], + "spans": [ + { + "bbox": [ + 130, + 617, + 481, + 666 + ], + "type": "text", + "content": "Building on the success of semi-supervised learning frameworks in depth estimation tasks [32,56,57], our proposed framework employs a teacher-student architecture, as shown in Fig. 4, to effectively integrate synthetic and real-domain data for high-resolution depth estimation. It employs a two-step process:" + } + ] + } + ], + "index": 9 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "spans": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "type": "text", + "content": "6" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 166, + 91, + 354, + 102 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 166, + 91, + 354, + 102 + ], + "spans": [ + { + "bbox": [ + 166, + 91, + 354, + 102 + ], + "type": "text", + "content": "Zhenyu Li, Shariq Farooq Bhat, Peter Wonka" + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 5 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 144, + 115, + 471, + 208 + ], + "blocks": [ + { + "bbox": [ + 144, + 115, + 471, + 208 + ], + "lines": [ + { + "bbox": [ + 144, + 115, + 471, + 208 + ], + "spans": [ + { + "bbox": [ + 144, + 115, + 471, + 208 + ], + "type": "image", + "image_path": "238c474efd7684cb61068d3f08b54809b865fb85cfca8dae1fc1dc8e02c765ca.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 130, + 217, + 482, + 274 + ], + "lines": [ + { + "bbox": [ + 130, + 217, + 482, + 274 + ], + "spans": [ + { + "bbox": [ + 130, + 217, + 482, + 274 + ], + "type": "text", + "content": "Fig. 4: Enhancing Real-Domain Learning with Synthetic Data. A teacher model trained on synthetic data produces pseudo labels for real-domain training. The student model benefits from a DSD dual-supervision approach: loss on pseudo labels for detail enhancement and loss on ground truth for scale accuracy. This method ensures detailed depth perception without compromising scale accuracy." + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_caption" + } + ], + "index": 2 + }, + { + "bbox": [ + 130, + 297, + 479, + 334 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 297, + 479, + 334 + ], + "spans": [ + { + "bbox": [ + 130, + 297, + 479, + 334 + ], + "type": "text", + "content": "Teacher Model Training: The teacher model is initially pretrained on a synthetic dataset " + }, + { + "bbox": [ + 130, + 297, + 479, + 334 + ], + "type": "inline_equation", + "content": "S" + }, + { + "bbox": [ + 130, + 297, + 479, + 334 + ], + "type": "text", + "content": ", which, due to its high-quality and detailed annotations, enables the model to predict high-resolution depth maps with precise boundaries." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 130, + 335, + 480, + 453 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 335, + 480, + 453 + ], + "spans": [ + { + "bbox": [ + 130, + 335, + 480, + 453 + ], + "type": "text", + "content": "Student Model Training with both Pseudo Labels and Ground-Truth Depth: In the subsequent phase, the teacher model is frozen, and the student model is trained on the real-domain dataset " + }, + { + "bbox": [ + 130, + 335, + 480, + 453 + ], + "type": "inline_equation", + "content": "\\mathcal{R}" + }, + { + "bbox": [ + 130, + 335, + 480, + 453 + ], + "type": "text", + "content": ", utilizing both the ground truth depth labels " + }, + { + "bbox": [ + 130, + 335, + 480, + 453 + ], + "type": "inline_equation", + "content": "\\tilde{\\mathbf{D}}" + }, + { + "bbox": [ + 130, + 335, + 480, + 453 + ], + "type": "text", + "content": " and pseudo labels " + }, + { + "bbox": [ + 130, + 335, + 480, + 453 + ], + "type": "inline_equation", + "content": "\\hat{\\mathbf{D}}" + }, + { + "bbox": [ + 130, + 335, + 480, + 453 + ], + "type": "text", + "content": " generated by the teacher model. As identified in the limitations of real-domain datasets, while the ground truth labels offer accurate depth information, they miss crucial information near boundaries essential for high-resolution depth learning. To address this, the student model is guided by the teacher's pseudo labels, which excel near boundaries. However, these pseudo labels, while sharp, exhibit scale discrepancies due to the domain gap between synthetic and real-world data as shown in Tab. 2." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 130, + 454, + 480, + 502 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 454, + 480, + 502 + ], + "spans": [ + { + "bbox": [ + 130, + 454, + 480, + 502 + ], + "type": "text", + "content": "Therefore, we introduce the Detail and Scale Disentangling (DSD) loss. It ensures that the final high-resolution depth predictions from the student model maintain an accurate scale while benefiting from the enhanced boundary details provided by the teacher model trained on the synthetic data." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 131, + 522, + 340, + 534 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 131, + 522, + 340, + 534 + ], + "spans": [ + { + "bbox": [ + 131, + 522, + 340, + 534 + ], + "type": "text", + "content": "3.4 Detail and Scale Disentangling Loss" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 130, + 545, + 479, + 592 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 545, + 479, + 592 + ], + "spans": [ + { + "bbox": [ + 130, + 545, + 479, + 592 + ], + "type": "text", + "content": "The student model is subject to two sources of supervision. The primary one is the standard scale-invariant loss [2,11,27,30], " + }, + { + "bbox": [ + 130, + 545, + 479, + 592 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_{\\text{silog}}" + }, + { + "bbox": [ + 130, + 545, + 479, + 592 + ], + "type": "text", + "content": ", calculated with the ground truth (GT) data, which ensures reliable and scale-consistent depth estimation on the real-domain data." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 130, + 594, + 480, + 666 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 594, + 480, + 666 + ], + "spans": [ + { + "bbox": [ + 130, + 594, + 480, + 666 + ], + "type": "text", + "content": "The secondary source of supervision comes from pseudo labels generated by the teacher model. These labels, while detailed in capturing boundaries and fine details, exhibit low scale accuracy. Directly applying conventional depth losses [11,24,31] could lead to an imbalance between enhancing detail and maintaining scale accuracy. To address this, motivated by current impressive progress in relative depth estimation [5,42,54,56], we adopt the ranking loss " + }, + { + "bbox": [ + 130, + 594, + 480, + 666 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_{\\text{rank}}" + }, + { + "bbox": [ + 130, + 594, + 480, + 666 + ], + "type": "text", + "content": " [5,54]" + } + ] + } + ], + "index": 9 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 393, + 91, + 447, + 100 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 393, + 91, + 447, + 100 + ], + "spans": [ + { + "bbox": [ + 393, + 91, + 447, + 100 + ], + "type": "text", + "content": "PatchRefiner" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 474, + 91, + 480, + 99 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 474, + 91, + 480, + 99 + ], + "spans": [ + { + "bbox": [ + 474, + 91, + 480, + 99 + ], + "type": "text", + "content": "7" + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 6 + }, + { + "para_blocks": [ + { + "bbox": [ + 130, + 116, + 479, + 163 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 116, + 479, + 163 + ], + "spans": [ + { + "bbox": [ + 130, + 116, + 479, + 163 + ], + "type": "text", + "content": "and the Scale and Shift Invariant losses " + }, + { + "bbox": [ + 130, + 116, + 479, + 163 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_{s s i}" + }, + { + "bbox": [ + 130, + 116, + 479, + 163 + ], + "type": "text", + "content": " [42,56] to inject the detail information. Notably, The " + }, + { + "bbox": [ + 130, + 116, + 479, + 163 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_{\\text {rank }}" + }, + { + "bbox": [ + 130, + 116, + 479, + 163 + ], + "type": "text", + "content": " only provides supervision on the prediction relationship whereas " + }, + { + "bbox": [ + 130, + 116, + 479, + 163 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_{s s i}" + }, + { + "bbox": [ + 130, + 116, + 479, + 163 + ], + "type": "text", + "content": " ignores scale and shift. Both losses tackle the same challenge with a slightly different approach." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 130, + 164, + 480, + 237 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 164, + 480, + 237 + ], + "spans": [ + { + "bbox": [ + 130, + 164, + 480, + 237 + ], + "type": "text", + "content": "Ranking loss [5, 54] is designed for sparse sets with ordinal data. This adaptation allows us to leverage dense detailed ranking information from the pseudo labels, enhancing high-resolution estimation without compromising scale accuracy. Specifically, for a certain pair of points " + }, + { + "bbox": [ + 130, + 164, + 480, + 237 + ], + "type": "inline_equation", + "content": "[p_{i,1}, p_{i,2}]" + }, + { + "bbox": [ + 130, + 164, + 480, + 237 + ], + "type": "text", + "content": " with predicted depth values " + }, + { + "bbox": [ + 130, + 164, + 480, + 237 + ], + "type": "inline_equation", + "content": "[d_{i,1}, d_{i,2}]" + }, + { + "bbox": [ + 130, + 164, + 480, + 237 + ], + "type": "text", + "content": " and pseudo depth values " + }, + { + "bbox": [ + 130, + 164, + 480, + 237 + ], + "type": "inline_equation", + "content": "[\\hat{d}_{i,1}, \\hat{d}_{i,2}]" + }, + { + "bbox": [ + 130, + 164, + 480, + 237 + ], + "type": "text", + "content": " in a set of " + }, + { + "bbox": [ + 130, + 164, + 480, + 237 + ], + "type": "inline_equation", + "content": "N" + }, + { + "bbox": [ + 130, + 164, + 480, + 237 + ], + "type": "text", + "content": " sampled point pairs " + }, + { + "bbox": [ + 130, + 164, + 480, + 237 + ], + "type": "inline_equation", + "content": "\\mathcal{P} = \\{[p_{i,1}, p_{i,2}], i = 1, 2, \\dots, N\\}" + }, + { + "bbox": [ + 130, + 164, + 480, + 237 + ], + "type": "text", + "content": ", the ranking loss is defined as" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 249, + 247, + 480, + 273 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 249, + 247, + 480, + 273 + ], + "spans": [ + { + "bbox": [ + 249, + 247, + 480, + 273 + ], + "type": "interline_equation", + "content": "\\mathcal {L} _ {\\text {r a n k}} = \\frac {1}{N} \\sum_ {i} \\phi \\left(p _ {i, 1}, p _ {i, 2}\\right), \\tag {2}", + "image_path": "616e7555670dd81d2a20aa44bca22a23604c0abda5c0e72216111bc29affccb1.jpg" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 186, + 280, + 479, + 312 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 186, + 280, + 479, + 312 + ], + "spans": [ + { + "bbox": [ + 186, + 280, + 479, + 312 + ], + "type": "interline_equation", + "content": "\\phi \\left(p _ {i, 1}, p _ {i, 2}\\right) = \\left\\{ \\begin{array}{l l} \\log \\left(1 + \\exp \\left(- \\ell \\times \\left(d _ {i, 1} - d _ {i, 2}\\right)\\right)\\right), & \\ell \\neq 0, \\\\ \\left(d _ {i, 1} - d _ {i, 2}\\right) ^ {2}, & \\ell = 0, \\end{array} \\right. \\tag {3}", + "image_path": "079346f60690d11e004e3209cf31f3c821e612397a79578b4ea13d413a5e8832.jpg" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 131, + 316, + 402, + 327 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 131, + 316, + 402, + 327 + ], + "spans": [ + { + "bbox": [ + 131, + 316, + 402, + 327 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 131, + 316, + 402, + 327 + ], + "type": "inline_equation", + "content": "\\ell" + }, + { + "bbox": [ + 131, + 316, + 402, + 327 + ], + "type": "text", + "content": " is the pseudo ordinal label, which can be induced by:" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 240, + 337, + 479, + 380 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 240, + 337, + 479, + 380 + ], + "spans": [ + { + "bbox": [ + 240, + 337, + 479, + 380 + ], + "type": "interline_equation", + "content": "\\ell = \\left\\{ \\begin{array}{l l} + 1, & \\hat {d} _ {i, 1} / \\hat {d} _ {i, 2} > = 1 + \\tau , \\\\ - 1, & \\hat {d} _ {i, 1} / \\hat {d} _ {i, 2} < = \\frac {1}{1 + \\tau}, \\\\ 0, & \\text {o t h e r w i s e .} \\end{array} \\right. \\tag {4}", + "image_path": "e4c54b1f94edab189bb382748ee37b1bcc4432eb75c7edb43be5bea1fcc1dbef.jpg" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 130, + 388, + 479, + 435 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 388, + 479, + 435 + ], + "spans": [ + { + "bbox": [ + 130, + 388, + 479, + 435 + ], + "type": "text", + "content": "Here " + }, + { + "bbox": [ + 130, + 388, + 479, + 435 + ], + "type": "inline_equation", + "content": "\\tau" + }, + { + "bbox": [ + 130, + 388, + 479, + 435 + ], + "type": "text", + "content": " is a tolerance threshold, which is set to 0.03 in our experiments following [54]. This loss not only encourages the predicted depth values for closely related points to align but also emphasizes the differentiation of points when the pseudo depth value of the two points is different." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 130, + 437, + 479, + 460 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 437, + 479, + 460 + ], + "spans": [ + { + "bbox": [ + 130, + 437, + 479, + 460 + ], + "type": "text", + "content": "Scale and Shift Invariant loss [42, 56] is proposed to learn relative depth estimation with ignoring the unknown scale and shift of each sample:" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 252, + 470, + 479, + 501 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 252, + 470, + 479, + 501 + ], + "spans": [ + { + "bbox": [ + 252, + 470, + 479, + 501 + ], + "type": "interline_equation", + "content": "\\mathcal {L} _ {s s i} = \\frac {1}{M} \\sum_ {i = 1} ^ {M} \\rho \\left(d _ {i} ^ {*} - \\hat {d} _ {i} ^ {*}\\right), \\tag {5}", + "image_path": "96ccfa0ba3d17ae184065ac4b28337f0be44a107cd4d1038591c4619252b510e.jpg" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 131, + 508, + 479, + 545 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 131, + 508, + 479, + 545 + ], + "spans": [ + { + "bbox": [ + 131, + 508, + 479, + 545 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 131, + 508, + 479, + 545 + ], + "type": "inline_equation", + "content": "d_i^*" + }, + { + "bbox": [ + 131, + 508, + 479, + 545 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 131, + 508, + 479, + 545 + ], + "type": "inline_equation", + "content": "\\hat{d}_i^*" + }, + { + "bbox": [ + 131, + 508, + 479, + 545 + ], + "type": "text", + "content": " are scaled and shifted versions of the predicted depth " + }, + { + "bbox": [ + 131, + 508, + 479, + 545 + ], + "type": "inline_equation", + "content": "d_i" + }, + { + "bbox": [ + 131, + 508, + 479, + 545 + ], + "type": "text", + "content": " and pseudo label " + }, + { + "bbox": [ + 131, + 508, + 479, + 545 + ], + "type": "inline_equation", + "content": "\\hat{d}_i" + }, + { + "bbox": [ + 131, + 508, + 479, + 545 + ], + "type": "text", + "content": ", and " + }, + { + "bbox": [ + 131, + 508, + 479, + 545 + ], + "type": "inline_equation", + "content": "\\rho" + }, + { + "bbox": [ + 131, + 508, + 479, + 545 + ], + "type": "text", + "content": " is the mean absolute error loss. " + }, + { + "bbox": [ + 131, + 508, + 479, + 545 + ], + "type": "inline_equation", + "content": "M" + }, + { + "bbox": [ + 131, + 508, + 479, + 545 + ], + "type": "text", + "content": " is the number of pixels. We use the least-squares criterion to align the prediction to the ground truth:" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 230, + 555, + 479, + 586 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 230, + 555, + 479, + 586 + ], + "spans": [ + { + "bbox": [ + 230, + 555, + 479, + 586 + ], + "type": "interline_equation", + "content": "(s, t) = \\operatorname {a r g m i n} _ {s, t} \\sum_ {i = 1} ^ {M} \\left(s d _ {i} + t - \\hat {d} _ {i}\\right) ^ {2}, \\tag {6}", + "image_path": "39de73e3cbf978299375f48b6ceaff99d0db856d19d57770f1eeaca152622bf1.jpg" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 264, + 599, + 479, + 612 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 264, + 599, + 479, + 612 + ], + "spans": [ + { + "bbox": [ + 264, + 599, + 479, + 612 + ], + "type": "interline_equation", + "content": "d ^ {*} = s d + t, \\hat {d} ^ {*} = \\hat {d} \\tag {7}", + "image_path": "4b0ea4359d31f53266e8aa46b1b5c973d24c8c6d9686b0e6bf3d7af7cffac55e.jpg" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 130, + 617, + 479, + 640 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 617, + 479, + 640 + ], + "spans": [ + { + "bbox": [ + 130, + 617, + 479, + 640 + ], + "type": "text", + "content": "where the scale " + }, + { + "bbox": [ + 130, + 617, + 479, + 640 + ], + "type": "inline_equation", + "content": "s" + }, + { + "bbox": [ + 130, + 617, + 479, + 640 + ], + "type": "text", + "content": " and shift " + }, + { + "bbox": [ + 130, + 617, + 479, + 640 + ], + "type": "inline_equation", + "content": "t" + }, + { + "bbox": [ + 130, + 617, + 479, + 640 + ], + "type": "text", + "content": " factors are effectively determined with the closed form [42]." + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 130, + 641, + 479, + 666 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 641, + 479, + 666 + ], + "spans": [ + { + "bbox": [ + 130, + 641, + 479, + 666 + ], + "type": "text", + "content": "The final Detail and Scale Disentangling loss for training the student model combines the " + }, + { + "bbox": [ + 130, + 641, + 479, + 666 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_{\\text{silog}}" + }, + { + "bbox": [ + 130, + 641, + 479, + 666 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 130, + 641, + 479, + 666 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_{\\text{rank}}" + }, + { + "bbox": [ + 130, + 641, + 479, + 666 + ], + "type": "text", + "content": ", and " + }, + { + "bbox": [ + 130, + 641, + 479, + 666 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_{\\text{ssi}}" + }, + { + "bbox": [ + 130, + 641, + 479, + 666 + ], + "type": "text", + "content": " as follows:" + } + ] + } + ], + "index": 15 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "spans": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "type": "text", + "content": "8" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 166, + 91, + 352, + 102 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 166, + 91, + 352, + 102 + ], + "spans": [ + { + "bbox": [ + 166, + 91, + 352, + 102 + ], + "type": "text", + "content": "Zhenyu Li, Shariq Farooq Bhat, Peter Wonka" + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 7 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 132, + 114, + 219, + 163 + ], + "blocks": [ + { + "bbox": [ + 132, + 114, + 219, + 163 + ], + "lines": [ + { + "bbox": [ + 132, + 114, + 219, + 163 + ], + "spans": [ + { + "bbox": [ + 132, + 114, + 219, + 163 + ], + "type": "image", + "image_path": "8dab28b88d7563df30b6187fecfc8857fe6795683a82a2f92436616104e182c4.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_body" + } + ], + "index": 2 + }, + { + "type": "image", + "bbox": [ + 133, + 163, + 219, + 212 + ], + "blocks": [ + { + "bbox": [ + 133, + 163, + 219, + 212 + ], + "lines": [ + { + "bbox": [ + 133, + 163, + 219, + 212 + ], + "spans": [ + { + "bbox": [ + 133, + 163, + 219, + 212 + ], + "type": "image", + "image_path": "6c68fbedfa8a1c82b2a1a77f63b210f2f0a2b87089f50957cb3ab25ae05d0b1a.jpg" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_body" + } + ], + "index": 3 + }, + { + "type": "image", + "bbox": [ + 133, + 213, + 219, + 262 + ], + "blocks": [ + { + "bbox": [ + 133, + 213, + 219, + 262 + ], + "lines": [ + { + "bbox": [ + 133, + 213, + 219, + 262 + ], + "spans": [ + { + "bbox": [ + 133, + 213, + 219, + 262 + ], + "type": "image", + "image_path": "2f08459c366f5842c705c642757da8ba579c4758c4f284199217bf174f13f01e.jpg" + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "image_body" + } + ], + "index": 4 + }, + { + "type": "image", + "bbox": [ + 133, + 262, + 219, + 312 + ], + "blocks": [ + { + "bbox": [ + 133, + 262, + 219, + 312 + ], + "lines": [ + { + "bbox": [ + 133, + 262, + 219, + 312 + ], + "spans": [ + { + "bbox": [ + 133, + 262, + 219, + 312 + ], + "type": "image", + "image_path": "c58e792f875c97aa3ada614d43537838c8498edb30d5d798e995884d3ae81ad4.jpg" + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 149, + 314, + 203, + 326 + ], + "lines": [ + { + "bbox": [ + 149, + 314, + 203, + 326 + ], + "spans": [ + { + "bbox": [ + 149, + 314, + 203, + 326 + ], + "type": "text", + "content": "ZoeDepth [2]" + } + ] + } + ], + "index": 6, + "angle": 0, + "type": "image_caption" + } + ], + "index": 5 + }, + { + "type": "image", + "bbox": [ + 220, + 114, + 305, + 162 + ], + "blocks": [ + { + "bbox": [ + 220, + 114, + 305, + 162 + ], + "lines": [ + { + "bbox": [ + 220, + 114, + 305, + 162 + ], + "spans": [ + { + "bbox": [ + 220, + 114, + 305, + 162 + ], + "type": "image", + "image_path": "c4beefae1114f0d2774628194d7a91ebd4f07cb291c7cad5096deee59acb2f47.jpg" + } + ] + } + ], + "index": 7, + "angle": 0, + "type": "image_body" + } + ], + "index": 7 + }, + { + "type": "image", + "bbox": [ + 220, + 163, + 306, + 213 + ], + "blocks": [ + { + "bbox": [ + 220, + 163, + 306, + 213 + ], + "lines": [ + { + "bbox": [ + 220, + 163, + 306, + 213 + ], + "spans": [ + { + "bbox": [ + 220, + 163, + 306, + 213 + ], + "type": "image", + "image_path": "5c7b289c8e4930dd5ac194640954128a479160058341fa71256fb1d0f1a7adbe.jpg" + } + ] + } + ], + "index": 8, + "angle": 0, + "type": "image_body" + } + ], + "index": 8 + }, + { + "type": "image", + "bbox": [ + 220, + 213, + 306, + 262 + ], + "blocks": [ + { + "bbox": [ + 220, + 213, + 306, + 262 + ], + "lines": [ + { + "bbox": [ + 220, + 213, + 306, + 262 + ], + "spans": [ + { + "bbox": [ + 220, + 213, + 306, + 262 + ], + "type": "image", + "image_path": "bc860e2d7bd90318b0e40008d643c053271d41c6bacfe815438e4b85be9ea37d.jpg" + } + ] + } + ], + "index": 9, + "angle": 0, + "type": "image_body" + } + ], + "index": 9 + }, + { + "type": "image", + "bbox": [ + 220, + 262, + 306, + 312 + ], + "blocks": [ + { + "bbox": [ + 220, + 262, + 306, + 312 + ], + "lines": [ + { + "bbox": [ + 220, + 262, + 306, + 312 + ], + "spans": [ + { + "bbox": [ + 220, + 262, + 306, + 312 + ], + "type": "image", + "image_path": "88f4b67b583fa246a8775c1d39a11c928c5c40654a3fb401f9273a2b99b41149.jpg" + } + ] + } + ], + "index": 10, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 229, + 314, + 298, + 327 + ], + "lines": [ + { + "bbox": [ + 229, + 314, + 298, + 327 + ], + "spans": [ + { + "bbox": [ + 229, + 314, + 298, + 327 + ], + "type": "text", + "content": "PatchFusion [27]" + } + ] + } + ], + "index": 11, + "angle": 0, + "type": "image_caption" + } + ], + "index": 10 + }, + { + "type": "image", + "bbox": [ + 307, + 114, + 392, + 162 + ], + "blocks": [ + { + "bbox": [ + 307, + 114, + 392, + 162 + ], + "lines": [ + { + "bbox": [ + 307, + 114, + 392, + 162 + ], + "spans": [ + { + "bbox": [ + 307, + 114, + 392, + 162 + ], + "type": "image", + "image_path": "e01f47a4253b0550142c5aee49050ef1092532c0b4d91a39f896709714fb5fa0.jpg" + } + ] + } + ], + "index": 12, + "angle": 0, + "type": "image_body" + } + ], + "index": 12 + }, + { + "type": "image", + "bbox": [ + 307, + 163, + 392, + 213 + ], + "blocks": [ + { + "bbox": [ + 307, + 163, + 392, + 213 + ], + "lines": [ + { + "bbox": [ + 307, + 163, + 392, + 213 + ], + "spans": [ + { + "bbox": [ + 307, + 163, + 392, + 213 + ], + "type": "image", + "image_path": "df734763aadd7f059378f04649a6248048b487a13c8e2865277e42ef36437d4a.jpg" + } + ] + } + ], + "index": 13, + "angle": 0, + "type": "image_body" + } + ], + "index": 13 + }, + { + "type": "image", + "bbox": [ + 307, + 213, + 392, + 262 + ], + "blocks": [ + { + "bbox": [ + 307, + 213, + 392, + 262 + ], + "lines": [ + { + "bbox": [ + 307, + 213, + 392, + 262 + ], + "spans": [ + { + "bbox": [ + 307, + 213, + 392, + 262 + ], + "type": "image", + "image_path": "24cc49880004440dee69805bbaff368ce5bf55ebf7d39c818f81658e7ee300d8.jpg" + } + ] + } + ], + "index": 14, + "angle": 0, + "type": "image_body" + } + ], + "index": 14 + }, + { + "type": "image", + "bbox": [ + 307, + 262, + 392, + 312 + ], + "blocks": [ + { + "bbox": [ + 307, + 262, + 392, + 312 + ], + "lines": [ + { + "bbox": [ + 307, + 262, + 392, + 312 + ], + "spans": [ + { + "bbox": [ + 307, + 262, + 392, + 312 + ], + "type": "image", + "image_path": "9255df48973122c1d1596726e4505f4de5bbffa868d5d0a5d1c6c056042c6ebb.jpg" + } + ] + } + ], + "index": 15, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 340, + 315, + 361, + 325 + ], + "lines": [ + { + "bbox": [ + 340, + 315, + 361, + 325 + ], + "spans": [ + { + "bbox": [ + 340, + 315, + 361, + 325 + ], + "type": "text", + "content": "Ours" + } + ] + } + ], + "index": 16, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 131, + 335, + 482, + 390 + ], + "lines": [ + { + "bbox": [ + 131, + 335, + 482, + 390 + ], + "spans": [ + { + "bbox": [ + 131, + 335, + 482, + 390 + ], + "type": "text", + "content": "Fig. 5: Qualitative Comparison on UnrealStereo4K. We show the depth prediction and corresponding error map, respectively. The qualitative comparisons showcased here indicate our framework outperforms counterparts [2, 27] with sharper edges and lower error around boundaries. We show individual patches in all images to emphasize details near depth boundaries." + } + ] + } + ], + "index": 22, + "angle": 0, + "type": "image_caption" + } + ], + "index": 15 + }, + { + "type": "image", + "bbox": [ + 393, + 114, + 480, + 162 + ], + "blocks": [ + { + "bbox": [ + 393, + 114, + 480, + 162 + ], + "lines": [ + { + "bbox": [ + 393, + 114, + 480, + 162 + ], + "spans": [ + { + "bbox": [ + 393, + 114, + 480, + 162 + ], + "type": "image", + "image_path": "eb9bf0c1d9113e5f8e80a0e473b30eb83a45a3c45a23b9a9e5be5fe5faec0d19.jpg" + } + ] + } + ], + "index": 17, + "angle": 0, + "type": "image_body" + } + ], + "index": 17 + }, + { + "type": "image", + "bbox": [ + 393, + 163, + 480, + 213 + ], + "blocks": [ + { + "bbox": [ + 393, + 163, + 480, + 213 + ], + "lines": [ + { + "bbox": [ + 393, + 163, + 480, + 213 + ], + "spans": [ + { + "bbox": [ + 393, + 163, + 480, + 213 + ], + "type": "image", + "image_path": "fc54c1ccfc9b2dbb82684191c04d34522f7074e00f0d5f2e88e82aa030e3b234.jpg" + } + ] + } + ], + "index": 18, + "angle": 0, + "type": "image_body" + } + ], + "index": 18 + }, + { + "type": "image", + "bbox": [ + 393, + 213, + 480, + 262 + ], + "blocks": [ + { + "bbox": [ + 393, + 213, + 480, + 262 + ], + "lines": [ + { + "bbox": [ + 393, + 213, + 480, + 262 + ], + "spans": [ + { + "bbox": [ + 393, + 213, + 480, + 262 + ], + "type": "image", + "image_path": "aad4f5d5ccbd16761a3442768302a5b38dad7bf541a49195a72e78a232a2538c.jpg" + } + ] + } + ], + "index": 19, + "angle": 0, + "type": "image_body" + } + ], + "index": 19 + }, + { + "type": "image", + "bbox": [ + 393, + 262, + 480, + 312 + ], + "blocks": [ + { + "bbox": [ + 393, + 262, + 480, + 312 + ], + "lines": [ + { + "bbox": [ + 393, + 262, + 480, + 312 + ], + "spans": [ + { + "bbox": [ + 393, + 262, + 480, + 312 + ], + "type": "image", + "image_path": "53ca2cb852d8fe3b66c8d74328dba9e51b2b253b2d1a2d517b80900fc05222ff.jpg" + } + ] + } + ], + "index": 20, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 416, + 315, + 459, + 326 + ], + "lines": [ + { + "bbox": [ + 416, + 315, + 459, + 326 + ], + "spans": [ + { + "bbox": [ + 416, + 315, + 459, + 326 + ], + "type": "text", + "content": "GT, Input" + } + ] + } + ], + "index": 21, + "angle": 0, + "type": "image_caption" + } + ], + "index": 20 + }, + { + "bbox": [ + 228, + 424, + 480, + 453 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 228, + 424, + 480, + 453 + ], + "spans": [ + { + "bbox": [ + 228, + 424, + 480, + 453 + ], + "type": "interline_equation", + "content": "\\mathcal {L} _ {D S D} = \\overbrace {\\mathcal {L} _ {s i l o g}} ^ {\\mathcal {L} _ {g t}} + \\overbrace {\\lambda_ {1} \\mathcal {L} _ {r a n k} + \\lambda_ {2} \\mathcal {L} _ {s s i}} ^ {\\mathcal {L} _ {p l}}, \\tag {8}", + "image_path": "67ff9ba6f96e467dd7064a65f19207bbec436065fa7df4dcd53ab1de4ad5cf47.jpg" + } + ] + } + ], + "index": 23 + }, + { + "bbox": [ + 131, + 457, + 482, + 494 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 131, + 457, + 482, + 494 + ], + "spans": [ + { + "bbox": [ + 131, + 457, + 482, + 494 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 131, + 457, + 482, + 494 + ], + "type": "inline_equation", + "content": "\\lambda_{1}" + }, + { + "bbox": [ + 131, + 457, + 482, + 494 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 131, + 457, + 482, + 494 + ], + "type": "inline_equation", + "content": "\\lambda_{2}" + }, + { + "bbox": [ + 131, + 457, + 482, + 494 + ], + "type": "text", + "content": " are two balancing factors, both empirically set to 0.1 in our experiments, respectively. The terms " + }, + { + "bbox": [ + 131, + 457, + 482, + 494 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_{gt}" + }, + { + "bbox": [ + 131, + 457, + 482, + 494 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 131, + 457, + 482, + 494 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_{pl}" + }, + { + "bbox": [ + 131, + 457, + 482, + 494 + ], + "type": "text", + "content": " represent the supervision signals derived from the ground truth and pseudo labels, respectively." + } + ] + } + ], + "index": 24 + }, + { + "bbox": [ + 132, + 512, + 230, + 525 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 512, + 230, + 525 + ], + "spans": [ + { + "bbox": [ + 132, + 512, + 230, + 525 + ], + "type": "text", + "content": "4 Experiments" + } + ] + } + ], + "index": 25 + }, + { + "bbox": [ + 132, + 537, + 269, + 548 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 537, + 269, + 548 + ], + "spans": [ + { + "bbox": [ + 132, + 537, + 269, + 548 + ], + "type": "text", + "content": "4.1 Datasets and Metrics" + } + ] + } + ], + "index": 26 + }, + { + "bbox": [ + 130, + 557, + 482, + 641 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 557, + 482, + 641 + ], + "spans": [ + { + "bbox": [ + 130, + 557, + 482, + 641 + ], + "type": "text", + "content": "UnrealStereo4K (Synthetic): The UnrealStereo4K dataset [51] offers synthetic stereo images at a 4K resolution " + }, + { + "bbox": [ + 130, + 557, + 482, + 641 + ], + "type": "inline_equation", + "content": "(2160\\times 3840)" + }, + { + "bbox": [ + 130, + 557, + 482, + 641 + ], + "type": "text", + "content": ", each paired with accurate, boundary-complete pixel-wise ground truth. Following the procedure in [27], we exclude mislabeled images using the Structural Similarity Index (SSIM) [52] and compute ground truth depth maps from provided disparity maps using camera parameters. Adhering to the dataset splits in [27, 51], we employ a default patch size of " + }, + { + "bbox": [ + 130, + 557, + 482, + 641 + ], + "type": "inline_equation", + "content": "540\\times 960" + }, + { + "bbox": [ + 130, + 557, + 482, + 641 + ], + "type": "text", + "content": " for compatibility with [27]." + } + ] + } + ], + "index": 27 + }, + { + "bbox": [ + 131, + 641, + 482, + 665 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 131, + 641, + 482, + 665 + ], + "spans": [ + { + "bbox": [ + 131, + 641, + 482, + 665 + ], + "type": "text", + "content": "CityScapes (Real, Stereo): Cityscapes [9] offers a comprehensive suite of urban scene images, segmentation masks, and disparity maps at a resolution" + } + ] + } + ], + "index": 28 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 393, + 91, + 447, + 100 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 393, + 91, + 447, + 100 + ], + "spans": [ + { + "bbox": [ + 393, + 91, + 447, + 100 + ], + "type": "text", + "content": "PatchRefiner" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 474, + 91, + 480, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 474, + 91, + 480, + 100 + ], + "spans": [ + { + "bbox": [ + 474, + 91, + 480, + 100 + ], + "type": "text", + "content": "9" + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 8 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 135, + 114, + 203, + 156 + ], + "blocks": [ + { + "bbox": [ + 135, + 114, + 203, + 156 + ], + "lines": [ + { + "bbox": [ + 135, + 114, + 203, + 156 + ], + "spans": [ + { + "bbox": [ + 135, + 114, + 203, + 156 + ], + "type": "image", + "image_path": "9a82ae2f5eeeb936bab130c8da7434df19693f3fefe365ba4597cee22ef16bd6.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_body" + } + ], + "index": 2 + }, + { + "type": "image", + "bbox": [ + 135, + 157, + 203, + 201 + ], + "blocks": [ + { + "bbox": [ + 135, + 157, + 203, + 201 + ], + "lines": [ + { + "bbox": [ + 135, + 157, + 203, + 201 + ], + "spans": [ + { + "bbox": [ + 135, + 157, + 203, + 201 + ], + "type": "image", + "image_path": "f121f3099f538b47ec7e9c78c35d53f79c13b3ea5969cc3debf791f97014636e.jpg" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_body" + } + ], + "index": 3 + }, + { + "type": "image", + "bbox": [ + 135, + 201, + 203, + 245 + ], + "blocks": [ + { + "bbox": [ + 135, + 201, + 203, + 245 + ], + "lines": [ + { + "bbox": [ + 135, + 201, + 203, + 245 + ], + "spans": [ + { + "bbox": [ + 135, + 201, + 203, + 245 + ], + "type": "image", + "image_path": "5989e62ce3a0c89bd616266826c0151a011da40d38c82361081ba6d95847c282.jpg" + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "image_body" + } + ], + "index": 4 + }, + { + "type": "image", + "bbox": [ + 135, + 246, + 203, + 289 + ], + "blocks": [ + { + "bbox": [ + 135, + 246, + 203, + 289 + ], + "lines": [ + { + "bbox": [ + 135, + 246, + 203, + 289 + ], + "spans": [ + { + "bbox": [ + 135, + 246, + 203, + 289 + ], + "type": "image", + "image_path": "43494bd5ec155036c39632cc68277204c4e25ca6f9ebf30efd0ebeace4e4177c.jpg" + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 191, + 292, + 217, + 303 + ], + "lines": [ + { + "bbox": [ + 191, + 292, + 217, + 303 + ], + "spans": [ + { + "bbox": [ + 191, + 292, + 217, + 303 + ], + "type": "text", + "content": "Image" + } + ] + } + ], + "index": 10, + "angle": 0, + "type": "image_caption" + } + ], + "index": 5 + }, + { + "type": "image", + "bbox": [ + 204, + 114, + 271, + 157 + ], + "blocks": [ + { + "bbox": [ + 204, + 114, + 271, + 157 + ], + "lines": [ + { + "bbox": [ + 204, + 114, + 271, + 157 + ], + "spans": [ + { + "bbox": [ + 204, + 114, + 271, + 157 + ], + "type": "image", + "image_path": "291c1f44d547b4fb8bba2ec6a94c0a880d9ba5b44709b5dee0bffa46cba658ec.jpg" + } + ] + } + ], + "index": 6, + "angle": 0, + "type": "image_body" + } + ], + "index": 6 + }, + { + "type": "image", + "bbox": [ + 204, + 158, + 271, + 201 + ], + "blocks": [ + { + "bbox": [ + 204, + 158, + 271, + 201 + ], + "lines": [ + { + "bbox": [ + 204, + 158, + 271, + 201 + ], + "spans": [ + { + "bbox": [ + 204, + 158, + 271, + 201 + ], + "type": "image", + "image_path": "7a4342bbaa38dd6ecf58b592662bf78ba9c000ca50d02d1bdc8b29398a8ea9f4.jpg" + } + ] + } + ], + "index": 7, + "angle": 0, + "type": "image_body" + } + ], + "index": 7 + }, + { + "type": "image", + "bbox": [ + 204, + 201, + 271, + 245 + ], + "blocks": [ + { + "bbox": [ + 204, + 201, + 271, + 245 + ], + "lines": [ + { + "bbox": [ + 204, + 201, + 271, + 245 + ], + "spans": [ + { + "bbox": [ + 204, + 201, + 271, + 245 + ], + "type": "image", + "image_path": "4642ed713d37ed4b63ce59ed15c6f41918acdd089a15b5f26f7e71a42ef6f9cc.jpg" + } + ] + } + ], + "index": 8, + "angle": 0, + "type": "image_body" + } + ], + "index": 8 + }, + { + "type": "image", + "bbox": [ + 204, + 246, + 271, + 289 + ], + "blocks": [ + { + "bbox": [ + 204, + 246, + 271, + 289 + ], + "lines": [ + { + "bbox": [ + 204, + 246, + 271, + 289 + ], + "spans": [ + { + "bbox": [ + 204, + 246, + 271, + 289 + ], + "type": "image", + "image_path": "b19b9e13cc89a4bc5588dcc2e2b775a88ed1d41b6ad93f8dba25041b436eca27.jpg" + } + ] + } + ], + "index": 9, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 279, + 292, + 333, + 303 + ], + "lines": [ + { + "bbox": [ + 279, + 292, + 333, + 303 + ], + "spans": [ + { + "bbox": [ + 279, + 292, + 333, + 303 + ], + "type": "text", + "content": "ZoeDepth [2]" + } + ] + } + ], + "index": 15, + "angle": 0, + "type": "image_caption" + } + ], + "index": 9 + }, + { + "type": "image", + "bbox": [ + 273, + 114, + 340, + 157 + ], + "blocks": [ + { + "bbox": [ + 273, + 114, + 340, + 157 + ], + "lines": [ + { + "bbox": [ + 273, + 114, + 340, + 157 + ], + "spans": [ + { + "bbox": [ + 273, + 114, + 340, + 157 + ], + "type": "image", + "image_path": "e9a3ca30e4056ab2eb9607b245365d894e49378b63ba9ba373a4e011beecfbaf.jpg" + } + ] + } + ], + "index": 11, + "angle": 0, + "type": "image_body" + } + ], + "index": 11 + }, + { + "type": "image", + "bbox": [ + 273, + 158, + 340, + 201 + ], + "blocks": [ + { + "bbox": [ + 273, + 158, + 340, + 201 + ], + "lines": [ + { + "bbox": [ + 273, + 158, + 340, + 201 + ], + "spans": [ + { + "bbox": [ + 273, + 158, + 340, + 201 + ], + "type": "image", + "image_path": "c0502f40097d064392fc97ab1269e4c981c13d4fa6e2c4cdf3f31355b611e0d6.jpg" + } + ] + } + ], + "index": 12, + "angle": 0, + "type": "image_body" + } + ], + "index": 12 + }, + { + "type": "image", + "bbox": [ + 273, + 201, + 340, + 245 + ], + "blocks": [ + { + "bbox": [ + 273, + 201, + 340, + 245 + ], + "lines": [ + { + "bbox": [ + 273, + 201, + 340, + 245 + ], + "spans": [ + { + "bbox": [ + 273, + 201, + 340, + 245 + ], + "type": "image", + "image_path": "863fabd6a3c138b8e0471cd71e00baf78300bc1742a3ae7c3c105e2861df0824.jpg" + } + ] + } + ], + "index": 13, + "angle": 0, + "type": "image_body" + } + ], + "index": 13 + }, + { + "type": "image", + "bbox": [ + 273, + 246, + 340, + 289 + ], + "blocks": [ + { + "bbox": [ + 273, + 246, + 340, + 289 + ], + "lines": [ + { + "bbox": [ + 273, + 246, + 340, + 289 + ], + "spans": [ + { + "bbox": [ + 273, + 246, + 340, + 289 + ], + "type": "image", + "image_path": "00e2026fb09e55e8d5039a008a63e4884e54eb2a6da4ff3fcfbf0c5a5719c3fb.jpg" + } + ] + } + ], + "index": 14, + "angle": 0, + "type": "image_body" + } + ], + "index": 14 + }, + { + "type": "image", + "bbox": [ + 342, + 114, + 409, + 157 + ], + "blocks": [ + { + "bbox": [ + 342, + 114, + 409, + 157 + ], + "lines": [ + { + "bbox": [ + 342, + 114, + 409, + 157 + ], + "spans": [ + { + "bbox": [ + 342, + 114, + 409, + 157 + ], + "type": "image", + "image_path": "a8f651b632aa099bbac8a3b95114804ba25dca6b6c202cd2cffaa60a8b4bf746.jpg" + } + ] + } + ], + "index": 16, + "angle": 0, + "type": "image_body" + } + ], + "index": 16 + }, + { + "type": "image", + "bbox": [ + 342, + 158, + 409, + 201 + ], + "blocks": [ + { + "bbox": [ + 342, + 158, + 409, + 201 + ], + "lines": [ + { + "bbox": [ + 342, + 158, + 409, + 201 + ], + "spans": [ + { + "bbox": [ + 342, + 158, + 409, + 201 + ], + "type": "image", + "image_path": "84b8e82a90297997638098c428dfddb86c5b55666d54b8d43c3c18f776f7e6a3.jpg" + } + ] + } + ], + "index": 17, + "angle": 0, + "type": "image_body" + } + ], + "index": 17 + }, + { + "type": "image", + "bbox": [ + 342, + 201, + 409, + 245 + ], + "blocks": [ + { + "bbox": [ + 342, + 201, + 409, + 245 + ], + "lines": [ + { + "bbox": [ + 342, + 201, + 409, + 245 + ], + "spans": [ + { + "bbox": [ + 342, + 201, + 409, + 245 + ], + "type": "image", + "image_path": "efca61ec132355ce07605ec69b5119fbdfe6a3702778d838c9fece8f2131430b.jpg" + } + ] + } + ], + "index": 18, + "angle": 0, + "type": "image_body" + } + ], + "index": 18 + }, + { + "type": "image", + "bbox": [ + 342, + 246, + 409, + 289 + ], + "blocks": [ + { + "bbox": [ + 342, + 246, + 409, + 289 + ], + "lines": [ + { + "bbox": [ + 342, + 246, + 409, + 289 + ], + "spans": [ + { + "bbox": [ + 342, + 246, + 409, + 289 + ], + "type": "image", + "image_path": "f5ce4825eedcf7cbff277dddffaf217a205310d89de00c536c1c4bca6d1412dc.jpg" + } + ] + } + ], + "index": 19, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 362, + 292, + 388, + 301 + ], + "lines": [ + { + "bbox": [ + 362, + 292, + 388, + 301 + ], + "spans": [ + { + "bbox": [ + 362, + 292, + 388, + 301 + ], + "type": "text", + "content": "PR R" + } + ] + } + ], + "index": 20, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 130, + 312, + 482, + 367 + ], + "lines": [ + { + "bbox": [ + 130, + 312, + 482, + 367 + ], + "spans": [ + { + "bbox": [ + 130, + 312, + 482, + 367 + ], + "type": "text", + "content": "Fig. 6: Qualitative Comparison on ETH3D and ScanNet++. The first two rows depict results for ETH3D and the last two for ScanNet++. The baseline ZoeDepth and PatchRefiner with conventional fine tuning (PR " + }, + { + "bbox": [ + 130, + 312, + 482, + 367 + ], + "type": "inline_equation", + "content": "\\mathcal{R}" + }, + { + "bbox": [ + 130, + 312, + 482, + 367 + ], + "type": "text", + "content": ") both struggle to create high-resolution depth details near boundaries. Our proposed strategy yields crisp boundaries on both datasets." + } + ] + } + ], + "index": 26, + "angle": 0, + "type": "image_caption" + } + ], + "index": 19 + }, + { + "type": "image", + "bbox": [ + 410, + 114, + 477, + 157 + ], + "blocks": [ + { + "bbox": [ + 410, + 114, + 477, + 157 + ], + "lines": [ + { + "bbox": [ + 410, + 114, + 477, + 157 + ], + "spans": [ + { + "bbox": [ + 410, + 114, + 477, + 157 + ], + "type": "image", + "image_path": "b459ecba07ad557a1075c3a3e623799e31e01f0b7d84d36429e807a6394b85b9.jpg" + } + ] + } + ], + "index": 21, + "angle": 0, + "type": "image_body" + } + ], + "index": 21 + }, + { + "type": "image", + "bbox": [ + 410, + 158, + 477, + 201 + ], + "blocks": [ + { + "bbox": [ + 410, + 158, + 477, + 201 + ], + "lines": [ + { + "bbox": [ + 410, + 158, + 477, + 201 + ], + "spans": [ + { + "bbox": [ + 410, + 158, + 477, + 201 + ], + "type": "image", + "image_path": "b373343d15848305d659867a958f194cae8795048ebee5b883b89e78237d05f4.jpg" + } + ] + } + ], + "index": 22, + "angle": 0, + "type": "image_body" + } + ], + "index": 22 + }, + { + "type": "image", + "bbox": [ + 410, + 201, + 477, + 245 + ], + "blocks": [ + { + "bbox": [ + 410, + 201, + 477, + 245 + ], + "lines": [ + { + "bbox": [ + 410, + 201, + 477, + 245 + ], + "spans": [ + { + "bbox": [ + 410, + 201, + 477, + 245 + ], + "type": "image", + "image_path": "ba004701f1ea2d7cb45f30200b87a83a2e7c9c1bccf0976d5571cd0a2399bd98.jpg" + } + ] + } + ], + "index": 23, + "angle": 0, + "type": "image_body" + } + ], + "index": 23 + }, + { + "type": "image", + "bbox": [ + 410, + 246, + 477, + 289 + ], + "blocks": [ + { + "bbox": [ + 410, + 246, + 477, + 289 + ], + "lines": [ + { + "bbox": [ + 410, + 246, + 477, + 289 + ], + "spans": [ + { + "bbox": [ + 410, + 246, + 477, + 289 + ], + "type": "image", + "image_path": "27c246a3de63a7759baf0f2a00b0e2aaa257713ad856c9a24a5e3fd655bbeff4.jpg" + } + ] + } + ], + "index": 24, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 430, + 292, + 457, + 301 + ], + "lines": [ + { + "bbox": [ + 430, + 292, + 457, + 301 + ], + "spans": [ + { + "bbox": [ + 430, + 292, + 457, + 301 + ], + "type": "text", + "content": "Ours 1" + } + ] + } + ], + "index": 25, + "angle": 0, + "type": "image_caption" + } + ], + "index": 24 + }, + { + "bbox": [ + 130, + 401, + 480, + 437 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 401, + 480, + 437 + ], + "spans": [ + { + "bbox": [ + 130, + 401, + 480, + 437 + ], + "type": "text", + "content": "of " + }, + { + "bbox": [ + 130, + 401, + 480, + 437 + ], + "type": "inline_equation", + "content": "1024 \\times 2048" + }, + { + "bbox": [ + 130, + 401, + 480, + 437 + ], + "type": "text", + "content": ", surpassing many real-domain datasets in density, quantity, and resolution [45, 46, 48, 49]. We use a standard patch size of " + }, + { + "bbox": [ + 130, + 401, + 480, + 437 + ], + "type": "inline_equation", + "content": "256 \\times 512" + }, + { + "bbox": [ + 130, + 401, + 480, + 437 + ], + "type": "text", + "content": " and conduct most experiments on this dataset." + } + ] + } + ], + "index": 27 + }, + { + "bbox": [ + 130, + 441, + 482, + 514 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 441, + 482, + 514 + ], + "spans": [ + { + "bbox": [ + 130, + 441, + 482, + 514 + ], + "type": "text", + "content": "ScanNet++ (Real, LiDAR, Reconstruction): ScanNet++ [58] is an extensive indoor dataset providing high-resolution images " + }, + { + "bbox": [ + 130, + 441, + 482, + 514 + ], + "type": "inline_equation", + "content": "(1440\\times 1920)" + }, + { + "bbox": [ + 130, + 441, + 482, + 514 + ], + "type": "text", + "content": ", depth maps from iPhone LiDAR " + }, + { + "bbox": [ + 130, + 441, + 482, + 514 + ], + "type": "inline_equation", + "content": "(192\\times 256)" + }, + { + "bbox": [ + 130, + 441, + 482, + 514 + ], + "type": "text", + "content": ", and high-resolution depth maps sampled from reconstructions of laser scans " + }, + { + "bbox": [ + 130, + 441, + 482, + 514 + ], + "type": "inline_equation", + "content": "(1440\\times 1920)" + }, + { + "bbox": [ + 130, + 441, + 482, + 514 + ], + "type": "text", + "content": ". Our chosen patch size is " + }, + { + "bbox": [ + 130, + 441, + 482, + 514 + ], + "type": "inline_equation", + "content": "720\\times 960" + }, + { + "bbox": [ + 130, + 441, + 482, + 514 + ], + "type": "text", + "content": ". We use the low-resolution ground-truth for training since the high-resolution version contains much noise as shown in Fig. 3." + } + ] + } + ], + "index": 28 + }, + { + "bbox": [ + 130, + 517, + 482, + 564 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 517, + 482, + 564 + ], + "spans": [ + { + "bbox": [ + 130, + 517, + 482, + 564 + ], + "type": "text", + "content": "ETH3D (Real, LiDAR): The ETH3D benchmark [46] provides high-resolution indoor and outdoor images " + }, + { + "bbox": [ + 130, + 517, + 482, + 564 + ], + "type": "inline_equation", + "content": "(6048\\times 4032)" + }, + { + "bbox": [ + 130, + 517, + 482, + 564 + ], + "type": "text", + "content": " with ground-truth depth from laser sensors. We downsample the image-depth pairs to " + }, + { + "bbox": [ + 130, + 517, + 482, + 564 + ], + "type": "inline_equation", + "content": "2160\\times 3840" + }, + { + "bbox": [ + 130, + 517, + 482, + 564 + ], + "type": "text", + "content": " and select a patch size of " + }, + { + "bbox": [ + 130, + 517, + 482, + 564 + ], + "type": "inline_equation", + "content": "540\\times 960" + }, + { + "bbox": [ + 130, + 517, + 482, + 564 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 29 + }, + { + "bbox": [ + 130, + 570, + 482, + 665 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 570, + 482, + 665 + ], + "spans": [ + { + "bbox": [ + 130, + 570, + 482, + 665 + ], + "type": "text", + "content": "Metrics: We adopt standard depth evaluation metrics from [2, 11, 38] and the Soft Edge Error (SEE) from [4, 27, 51] for scale evaluation. Additional metric details are in the supplementary material. Given real-world depth map boundary incompleteness, we design metrics to focus on evaluating boundary accuracy, leveraging high-resolution fine-grained segmentation masks as depth boundary quality proxies. Specifically, we use the Sobel operator [21] on both predicted depth and segmentation maps to generate edge masks, and then compute Precision, Recall, and F1 Score." + } + ] + } + ], + "index": 30 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "spans": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "type": "text", + "content": "10" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 166, + 91, + 352, + 102 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 166, + 91, + 352, + 102 + ], + "spans": [ + { + "bbox": [ + 166, + 91, + 352, + 102 + ], + "type": "text", + "content": "Zhenyu Li, Shariq Farooq Bhat, Peter Wonka" + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 9 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 135, + 114, + 203, + 335 + ], + "blocks": [ + { + "bbox": [ + 135, + 114, + 203, + 335 + ], + "lines": [ + { + "bbox": [ + 135, + 114, + 203, + 335 + ], + "spans": [ + { + "bbox": [ + 135, + 114, + 203, + 335 + ], + "type": "image", + "image_path": "3eebc3454996388cb42a2c5423d6d7d2eae964d238e942613c36dfeb31f78dd4.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 130, + 344, + 482, + 401 + ], + "lines": [ + { + "bbox": [ + 130, + 344, + 482, + 401 + ], + "spans": [ + { + "bbox": [ + 130, + 344, + 482, + 401 + ], + "type": "text", + "content": "Fig. 7: Qualitative Comparison on CityScapes. This figure illustrates depth estimation comparisons between the base ZoeDepth model, PatchRefiner (PR) trained on CityScapes, and our method. We display outcomes under varying levels of " + }, + { + "bbox": [ + 130, + 344, + 482, + 401 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_{pl}" + }, + { + "bbox": [ + 130, + 344, + 482, + 401 + ], + "type": "text", + "content": " supervision (" + }, + { + "bbox": [ + 130, + 344, + 482, + 401 + ], + "type": "inline_equation", + "content": "\\lambda_1 = \\lambda_2 = 1e^{-1}" + }, + { + "bbox": [ + 130, + 344, + 482, + 401 + ], + "type": "text", + "content": " or 1), featuring zoomed-in sections of each image to highlight detail fidelity near depth discontinuities." + } + ] + } + ], + "index": 7, + "angle": 0, + "type": "image_caption" + } + ], + "index": 2 + }, + { + "type": "image", + "bbox": [ + 205, + 114, + 271, + 335 + ], + "blocks": [ + { + "bbox": [ + 205, + 114, + 271, + 335 + ], + "lines": [ + { + "bbox": [ + 205, + 114, + 271, + 335 + ], + "spans": [ + { + "bbox": [ + 205, + 114, + 271, + 335 + ], + "type": "image", + "image_path": "cbf42d6b563403d4ce2a678da840a98d7d2a104b7e2b532d11f0db1cddb31694.jpg" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_body" + } + ], + "index": 3 + }, + { + "type": "image", + "bbox": [ + 273, + 114, + 340, + 335 + ], + "blocks": [ + { + "bbox": [ + 273, + 114, + 340, + 335 + ], + "lines": [ + { + "bbox": [ + 273, + 114, + 340, + 335 + ], + "spans": [ + { + "bbox": [ + 273, + 114, + 340, + 335 + ], + "type": "image", + "image_path": "8ee3869a0da1c121b5c6d2d06617e3fda7827f30812088e868cf52a6d6699fac.jpg" + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "image_body" + } + ], + "index": 4 + }, + { + "type": "image", + "bbox": [ + 342, + 114, + 408, + 335 + ], + "blocks": [ + { + "bbox": [ + 342, + 114, + 408, + 335 + ], + "lines": [ + { + "bbox": [ + 342, + 114, + 408, + 335 + ], + "spans": [ + { + "bbox": [ + 342, + 114, + 408, + 335 + ], + "type": "image", + "image_path": "8f34a358dd8d4a536260f08c20f741e16c67555430f728f0ceb03e24d447b9a3.jpg" + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "image_body" + } + ], + "index": 5 + }, + { + "type": "image", + "bbox": [ + 410, + 114, + 477, + 335 + ], + "blocks": [ + { + "bbox": [ + 410, + 114, + 477, + 335 + ], + "lines": [ + { + "bbox": [ + 410, + 114, + 477, + 335 + ], + "spans": [ + { + "bbox": [ + 410, + 114, + 477, + 335 + ], + "type": "image", + "image_path": "910705c5cb23bf68a742fe54f336dfc9e91ef3446490140b06f65a5650d98a5e.jpg" + } + ] + } + ], + "index": 6, + "angle": 0, + "type": "image_body" + } + ], + "index": 6 + }, + { + "bbox": [ + 132, + 428, + 279, + 440 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 428, + 279, + 440 + ], + "spans": [ + { + "bbox": [ + 132, + 428, + 279, + 440 + ], + "type": "text", + "content": "4.2 Implementation Details" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 130, + 460, + 482, + 579 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 460, + 482, + 579 + ], + "spans": [ + { + "bbox": [ + 130, + 460, + 482, + 579 + ], + "type": "text", + "content": "PatchRefiner on Synthetic Dataset: For training on synthetic datasets, we employ the scale-invariant log loss " + }, + { + "bbox": [ + 130, + 460, + 482, + 579 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_{\\text{silog}}" + }, + { + "bbox": [ + 130, + 460, + 482, + 579 + ], + "type": "text", + "content": ", as outlined in [2, 11, 27, 30]. Initialization of the coarse network " + }, + { + "bbox": [ + 130, + 460, + 482, + 579 + ], + "type": "inline_equation", + "content": "\\mathcal{N}_c" + }, + { + "bbox": [ + 130, + 460, + 482, + 579 + ], + "type": "text", + "content": " leverages pretrained weights from the NYU-v2 dataset [48], adhering to the approach in [27]. We dedicate 24 epochs to training " + }, + { + "bbox": [ + 130, + 460, + 482, + 579 + ], + "type": "inline_equation", + "content": "\\mathcal{N}_c" + }, + { + "bbox": [ + 130, + 460, + 482, + 579 + ], + "type": "text", + "content": ", which is subsequently frozen to ensure stability in subsequent training phases. For the refinement network " + }, + { + "bbox": [ + 130, + 460, + 482, + 579 + ], + "type": "inline_equation", + "content": "\\mathcal{N}_r" + }, + { + "bbox": [ + 130, + 460, + 482, + 579 + ], + "type": "text", + "content": ", initialization employs " + }, + { + "bbox": [ + 130, + 460, + 482, + 579 + ], + "type": "inline_equation", + "content": "\\mathcal{N}_c" + }, + { + "bbox": [ + 130, + 460, + 482, + 579 + ], + "type": "text", + "content": "'s parameters, and training extends for an additional 36 epochs. Standard augmentation strategies from the baseline depth model are incorporated to enhance training effectiveness. During inference, we implement Consistency-Aware Inference, as described in [27], to optimize performance." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 130, + 582, + 482, + 666 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 582, + 482, + 666 + ], + "spans": [ + { + "bbox": [ + 130, + 582, + 482, + 666 + ], + "type": "text", + "content": "Learning on Real-Domain Dataset: Since the coarse model within the framework offers scale-consistent predictions, we first pretrain a coarse model on real-domain data to establish a reliable depth scale foundation. This model is subsequently frozen to preserve scale consistency throughout the training process. The refiner model in PatchRefiner is then initialized with parameters pretrained on synthetic data, enabling it to maintain high-frequency detail knowledge acquired from the synthetic domain. Initially, the student model is trained solely" + } + ] + } + ], + "index": 10 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 393, + 91, + 447, + 100 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 393, + 91, + 447, + 100 + ], + "spans": [ + { + "bbox": [ + 393, + 91, + 447, + 100 + ], + "type": "text", + "content": "PatchRefiner" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 470, + 91, + 480, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 470, + 91, + 480, + 100 + ], + "spans": [ + { + "bbox": [ + 470, + 91, + 480, + 100 + ], + "type": "text", + "content": "11" + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 10 + }, + { + "para_blocks": [ + { + "type": "table", + "bbox": [ + 138, + 157, + 475, + 360 + ], + "blocks": [ + { + "bbox": [ + 132, + 114, + 479, + 148 + ], + "lines": [ + { + "bbox": [ + 132, + 114, + 479, + 148 + ], + "spans": [ + { + "bbox": [ + 132, + 114, + 479, + 148 + ], + "type": "text", + "content": "Table 1: Quantitative comparison on UnrealStereo4K. We color code the corresponding best competitor and our method within each block. PF and PR are short for PatchFusion [27] and PatchRefiner, respectively. The reported numbers are from [27]." + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 138, + 157, + 475, + 360 + ], + "lines": [ + { + "bbox": [ + 138, + 157, + 475, + 360 + ], + "spans": [ + { + "bbox": [ + 138, + 157, + 475, + 360 + ], + "type": "table", + "html": "
Methodδ1(%)↑REL↓RMS↓SiLog↓SEE↓Reference
iDisc [38]96.9400.0531.4048.5021.070ICCV 2023
SMD-Net [51]97.7740.0441.2827.3890.883CVPR 2021
Graph-GDSR [10]97.9320.0441.2647.4690.872CVPR 2022
BoostingDepth [34]98.1040.0441.1236.6620.939CVPR 2021
ZoeDepth [2]97.7170.0461.2897.4480.914-
ZoeDepth+PFp=16 [27]98.4190.0401.0886.2120.838CVPR 2024
ZoeDepth+PFp=49 [27]98.4500.0391.0756.1310.846
ZoeDepth+PFR=128 [27]98.4690.0391.0666.0850.849
ZoeDepth+PRp=1698.8210.0330.8925.4170.750Ours
ZoeDepth+PRp=4998.8590.0330.8705.3190.751
ZoeDepth+PRR=12898.8640.0330.8725.3770.738
Depth-Anything [56]97.7730.0411.2357.1920.911CVPR 2024
Depth-Anything+PFp=16 [27]98.5580.0361.0155.8830.811CVPR 2024
Depth-Anything+PFp=49 [27]98.6070.0350.9875.7460.812
Depth-Anything+PFR=128 [27]98.6160.0350.9845.7750.813
Depth-Anything+PRp=1698.8260.0330.8895.2890.768Ours
Depth-Anything+PRp=4998.8780.0330.8605.1490.767
Depth-Anything+PRR=12898.8780.0330.8605.2060.759
", + "image_path": "122b8bbfa8120f6c0e61d22bf2087987dacf1a98bc14e65aaeb1be74878a393a.jpg" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "table_body" + } + ], + "index": 3 + }, + { + "type": "table", + "bbox": [ + 141, + 413, + 472, + 509 + ], + "blocks": [ + { + "bbox": [ + 132, + 370, + 479, + 403 + ], + "lines": [ + { + "bbox": [ + 132, + 370, + 479, + 403 + ], + "spans": [ + { + "bbox": [ + 132, + 370, + 479, + 403 + ], + "type": "text", + "content": "Table 2: Quantitative comparison on CityScapes. FT, and MIX are short for fine-tuning and mixed-data strategies, which are our main competitors. Baseline is highlighted in gray and each of the three competitors in a different color." + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 141, + 413, + 472, + 509 + ], + "lines": [ + { + "bbox": [ + 141, + 413, + 472, + 509 + ], + "spans": [ + { + "bbox": [ + 141, + 413, + 472, + 509 + ], + "type": "table", + "html": "
MethodDataScaleBoundary
SRδ1(%)↑REL↓RMS↓Precision↑Recall↑
ZoeDepth [2]✓94.5020.0704.40613.3237.59
ZoeDepth [2] + FT✓✓94.4980.0714.41812.9337.89
PatchRefiner (zero-shot)✓5.7050.39912.20328.6851.28
PatchRefiner✓95.2840.0664.04716.6739.53
PatchRefiner + FT✓✓95.4180.0653.99217.0940.92
PatchRefiner + mix [39]✓✓89.1080.1124.73223.0841.17
PatchRefiner + DSD✓✓95.3590.0663.98218.9248.78
", + "image_path": "8637ab8ce4d95789d649ea22289681017732fd0ea26c9b7b169882df9c45d605.jpg" + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "table_body" + } + ], + "index": 5 + }, + { + "bbox": [ + 132, + 531, + 479, + 555 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 531, + 479, + 555 + ], + "spans": [ + { + "bbox": [ + 132, + 531, + 479, + 555 + ], + "type": "text", + "content": "with " + }, + { + "bbox": [ + 132, + 531, + 479, + 555 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_{\\text{silog}}" + }, + { + "bbox": [ + 132, + 531, + 479, + 555 + ], + "type": "text", + "content": " for 24 epochs. Subsequent fine-tuning with our Detail and Scale Disentangling loss " + }, + { + "bbox": [ + 132, + 531, + 479, + 555 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_{DSD}" + }, + { + "bbox": [ + 132, + 531, + 479, + 555 + ], + "type": "text", + "content": " over an additional 6 epochs to refine depth estimations." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 132, + 573, + 228, + 583 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 573, + 228, + 583 + ], + "spans": [ + { + "bbox": [ + 132, + 573, + 228, + 583 + ], + "type": "text", + "content": "4.3 Main Results" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 132, + 594, + 479, + 665 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 594, + 479, + 665 + ], + "spans": [ + { + "bbox": [ + 132, + 594, + 479, + 665 + ], + "type": "text", + "content": "Synthetic Dataset: On the UnrealStereo4K dataset, PatchRefiner not only outperforms the base depth model but also shows substantial improvements over PatchFusion, reducing RMSE by " + }, + { + "bbox": [ + 132, + 594, + 479, + 665 + ], + "type": "inline_equation", + "content": "18.1\\%" + }, + { + "bbox": [ + 132, + 594, + 479, + 665 + ], + "type": "text", + "content": " and REL by " + }, + { + "bbox": [ + 132, + 594, + 479, + 665 + ], + "type": "inline_equation", + "content": "15.7\\%" + }, + { + "bbox": [ + 132, + 594, + 479, + 665 + ], + "type": "text", + "content": ". This advancement is further underscored by achieving the lowest SEE, highlighting our model's proficiency in capturing edge details. Qualitative comparisons, illustrated in Fig. 5, indicate the superior boundary delineation achieved by PatchRefiner." + } + ] + } + ], + "index": 8 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "spans": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "type": "text", + "content": "12" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 166, + 91, + 351, + 102 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 166, + 91, + 351, + 102 + ], + "spans": [ + { + "bbox": [ + 166, + 91, + 351, + 102 + ], + "type": "text", + "content": "Zhenyu Li, Shariq Farooq Bhat, Peter Wonka" + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 11 + }, + { + "para_blocks": [ + { + "type": "table", + "bbox": [ + 136, + 146, + 477, + 251 + ], + "blocks": [ + { + "bbox": [ + 132, + 114, + 480, + 137 + ], + "lines": [ + { + "bbox": [ + 132, + 114, + 480, + 137 + ], + "spans": [ + { + "bbox": [ + 132, + 114, + 480, + 137 + ], + "type": "text", + "content": "Table 3: Ablation study of architecture variations and formulation of final depth prediction. Ours is highlighted with color. We highlight the best in bold." + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 136, + 146, + 477, + 251 + ], + "lines": [ + { + "bbox": [ + 136, + 146, + 477, + 251 + ], + "spans": [ + { + "bbox": [ + 136, + 146, + 477, + 251 + ], + "type": "table", + "html": "
MethodType of outputFeature Levelsδ1(%)↑REL↓RMS↓SiLog↓SEE↓
PatchFusion [27]Direct6 features98.4190.0401.0886.2120.838
PatchRefinerDcresidual1 features98.7340.0340.9265.5500.782
Dcresidual2 features98.8140.0330.9055.5110.750
Dcresidual3 features98.8150.0340.9005.5830.753
Dcresidual4 features98.8150.0330.8995.4940.752
Dcresidual5 features98.8140.0330.8945.4680.752
Dcresidual6 features98.8210.0330.8925.4170.750
Ddresidual6 features98.8040.0330.8995.4480.753
Direct6 features98.7490.0340.9255.5910.765
", + "image_path": "e12c7381bb8a16edcb3e164843734ea41ce8f7e3986a7794c73443d1d9b0de46.jpg" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "table_body" + } + ], + "index": 3 + }, + { + "type": "table", + "bbox": [ + 136, + 315, + 474, + 445 + ], + "blocks": [ + { + "bbox": [ + 130, + 260, + 479, + 304 + ], + "lines": [ + { + "bbox": [ + 130, + 260, + 479, + 304 + ], + "spans": [ + { + "bbox": [ + 130, + 260, + 479, + 304 + ], + "type": "text", + "content": "Table 4: Variations of " + }, + { + "bbox": [ + 130, + 260, + 479, + 304 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_{pl}" + }, + { + "bbox": [ + 130, + 260, + 479, + 304 + ], + "type": "text", + "content": ". We analyse various options for " + }, + { + "bbox": [ + 130, + 260, + 479, + 304 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_{pl}" + }, + { + "bbox": [ + 130, + 260, + 479, + 304 + ], + "type": "text", + "content": " and compare them against PR. " + }, + { + "bbox": [ + 130, + 260, + 479, + 304 + ], + "type": "inline_equation", + "content": "\\mathcal{S}" + }, + { + "bbox": [ + 130, + 260, + 479, + 304 + ], + "type": "text", + "content": " and PR. " + }, + { + "bbox": [ + 130, + 260, + 479, + 304 + ], + "type": "inline_equation", + "content": "\\mathcal{R}" + }, + { + "bbox": [ + 130, + 260, + 479, + 304 + ], + "type": "text", + "content": " that serve as baselines for training on Synthetic and Real data individually. We set " + }, + { + "bbox": [ + 130, + 260, + 479, + 304 + ], + "type": "inline_equation", + "content": "\\lambda = \\lambda_1 = \\lambda_2" + }, + { + "bbox": [ + 130, + 260, + 479, + 304 + ], + "type": "text", + "content": " to analyze the influence of DSD weight. The highlighted result is achieved with " + }, + { + "bbox": [ + 130, + 260, + 479, + 304 + ], + "type": "inline_equation", + "content": "\\lambda = 1e^{-1}" + }, + { + "bbox": [ + 130, + 260, + 479, + 304 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 136, + 315, + 474, + 445 + ], + "lines": [ + { + "bbox": [ + 136, + 315, + 474, + 445 + ], + "spans": [ + { + "bbox": [ + 136, + 315, + 474, + 445 + ], + "type": "table", + "html": "
Variations (Lpl)ScaleBoundary
δ1(%)↑REL↓RMS↓Precision↑Recall↑F1↑
Baseline PR. S5.7050.39912.2028.6851.2836.34
Baseline PR. R95.4180.0653.99217.0940.9223.68
Lsilog82.5500.1555.91426.3458.3535.81
Lsilog + mask81.4250.1486.51330.0446.8936.19
Lrank95.4130.0663.97318.3847.8126.12
Lssi95.4650.0653.97419.2644.0126.39
Lrank + Lssi (Ours)95.3590.0663.98218.9248.7826.84
λ = 195.0770.0694.22221.3258.4830.90
λ = 3e-195.2960.0684.08620.1153.8328.91
λ = 3e-295.4620.0653.95318.0143.9025.11
", + "image_path": "986358a00d4dad570fe0089a8c082b78b81f03dc4f6a5810221d2dafde222d66.jpg" + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "table_body" + } + ], + "index": 5 + }, + { + "bbox": [ + 130, + 463, + 480, + 581 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 463, + 480, + 581 + ], + "spans": [ + { + "bbox": [ + 130, + 463, + 480, + 581 + ], + "type": "text", + "content": "Real-Domain Dataset: Tab. 2 and Fig. 6, 7 delineate the performance disparity when leveraging synthetic data for real-domain learning. Although the synthetic-trained model excels in boundary details, it fails in scale accuracy due to the domain gap. Sole training on real-domain data enhances baseline's scale prediction yet falls short in detail accuracy due to the missing depth ground truth around boundaries. Neither fine-tuning nor mixed-training substantially elevates performance across scale and detail metrics, reflecting the inherent challenges in our task. Contrastingly, our strategy propels the model to notable gains in boundary accuracy (19.2% increase in boundary recall) while sustaining scale precision comparable to the baseline model." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 132, + 598, + 323, + 609 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 598, + 323, + 609 + ], + "spans": [ + { + "bbox": [ + 132, + 598, + 323, + 609 + ], + "type": "text", + "content": "4.4 Ablation Studies and Discussion" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 130, + 617, + 479, + 665 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 617, + 479, + 665 + ], + "spans": [ + { + "bbox": [ + 130, + 617, + 479, + 665 + ], + "type": "text", + "content": "We ablate and discuss the contributions of individual components. As default, we utilize the UnrealStereo4K dataset for synthetic-dataset training and the PatchRefiner variant with " + }, + { + "bbox": [ + 130, + 617, + 479, + 665 + ], + "type": "inline_equation", + "content": "P = 16" + }, + { + "bbox": [ + 130, + 617, + 479, + 665 + ], + "type": "text", + "content": " patches for clarity and ease of comparison. We ablate our teacher-student framework using the CityScapes dataset." + } + ] + } + ], + "index": 8 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 393, + 91, + 447, + 101 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 393, + 91, + 447, + 101 + ], + "spans": [ + { + "bbox": [ + 393, + 91, + 447, + 101 + ], + "type": "text", + "content": "PatchRefiner" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 470, + 91, + 481, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 470, + 91, + 481, + 100 + ], + "spans": [ + { + "bbox": [ + 470, + 91, + 481, + 100 + ], + "type": "text", + "content": "13" + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 12 + }, + { + "para_blocks": [ + { + "bbox": [ + 130, + 116, + 482, + 200 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 116, + 482, + 200 + ], + "spans": [ + { + "bbox": [ + 130, + 116, + 482, + 200 + ], + "type": "text", + "content": "Architecture and Formulation Variations: We first ablate the effectiveness of our architecture design. Utilizing a ZoeDepth model, we extract six levels of intermediate features during the forward pass, represented as " + }, + { + "bbox": [ + 130, + 116, + 482, + 200 + ], + "type": "inline_equation", + "content": "\\mathcal{F} = f_1, f_2, \\dots, f_6" + }, + { + "bbox": [ + 130, + 116, + 482, + 200 + ], + "type": "text", + "content": ", with progressively increasing resolution. We then sequentially omit lower resolution feature maps from the refiner's decoder input. As depicted in Tab. 3, performance diminishes with fewer feature maps, yet even with solely the highest resolution feature map, our model outperforms PatchFusion." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 130, + 200, + 482, + 272 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 200, + 482, + 272 + ], + "spans": [ + { + "bbox": [ + 130, + 200, + 482, + 272 + ], + "type": "text", + "content": "Further, we explore our refinement formulation's efficacy by comparing it against two alternatives: (1) Residual depth adjustment based on the base model " + }, + { + "bbox": [ + 130, + 200, + 482, + 272 + ], + "type": "inline_equation", + "content": "\\mathcal{N}_d" + }, + { + "bbox": [ + 130, + 200, + 482, + 272 + ], + "type": "text", + "content": " within the refiner. (2) Direct depth prediction from the refiner's decoder, similar to PatchFusion (" + }, + { + "bbox": [ + 130, + 200, + 482, + 272 + ], + "type": "inline_equation", + "content": "\\mathbf{D} = \\mathbf{D}_r" + }, + { + "bbox": [ + 130, + 200, + 482, + 272 + ], + "type": "text", + "content": " in this context). Our residual-based approach demonstrates superior performance, underscoring its advantages in optimization and training efficiency." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 130, + 287, + 482, + 419 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 287, + 482, + 419 + ], + "spans": [ + { + "bbox": [ + 130, + 287, + 482, + 419 + ], + "type": "text", + "content": "Effectiveness of Pseudo Label Supervision: We evaluate the efficacy of our Detail and Scale Disentangling (DSD) loss against the conventional scale-invariant loss, " + }, + { + "bbox": [ + 130, + 287, + 482, + 419 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_{silog}" + }, + { + "bbox": [ + 130, + 287, + 482, + 419 + ], + "type": "text", + "content": ", for pseudo label supervision. Tab. 4 illustrates that while " + }, + { + "bbox": [ + 130, + 287, + 482, + 419 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_{silog}" + }, + { + "bbox": [ + 130, + 287, + 482, + 419 + ], + "type": "text", + "content": " aids in detail transfer from the teacher to the student model, it compromises scale accuracy due to significant discrepancies in the pseudo labels. A masking approach, focusing only on areas lacking depth, does not mitigate this issue, indicating pervasive negative effects. In contrast, the combination of " + }, + { + "bbox": [ + 130, + 287, + 482, + 419 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_{rank}" + }, + { + "bbox": [ + 130, + 287, + 482, + 419 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 130, + 287, + 482, + 419 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_{ssi}" + }, + { + "bbox": [ + 130, + 287, + 482, + 419 + ], + "type": "text", + "content": " within " + }, + { + "bbox": [ + 130, + 287, + 482, + 419 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_{pl}" + }, + { + "bbox": [ + 130, + 287, + 482, + 419 + ], + "type": "text", + "content": " not only improves detail fidelity but also maintains scale accuracy, demonstrating that ranking constraints and scale-shift invariance are effective, orthogonal strategies for enhancing high-resolution detail without sacrificing scale accuracy." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 130, + 434, + 482, + 533 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 434, + 482, + 533 + ], + "spans": [ + { + "bbox": [ + 130, + 434, + 482, + 533 + ], + "type": "text", + "content": "Exploration of Variant DSD Loss Weight: This study examines the impact of different DSD loss weights " + }, + { + "bbox": [ + 130, + 434, + 482, + 533 + ], + "type": "inline_equation", + "content": "(\\lambda_{1}" + }, + { + "bbox": [ + 130, + 434, + 482, + 533 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 130, + 434, + 482, + 533 + ], + "type": "inline_equation", + "content": "\\lambda_{2})" + }, + { + "bbox": [ + 130, + 434, + 482, + 533 + ], + "type": "text", + "content": " to elucidate the loss's efficiency. As shown in Tab. 4 and Fig. 7, increasing the loss weight marginally affects scale accuracy while significantly improving boundaries, validating the DSD loss's role in balancing detail enhancement and scale preservation. Notably, when both weights are set to 1, the model's boundary recall surpasses the teacher's performance. Furthermore, when DSD loss achieves a comparable boundary metric to " + }, + { + "bbox": [ + 130, + 434, + 482, + 533 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_{silog}" + }, + { + "bbox": [ + 130, + 434, + 482, + 533 + ], + "type": "text", + "content": ", it exhibits a smaller performance decline, underscoring its effectiveness." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 131, + 546, + 220, + 560 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 131, + 546, + 220, + 560 + ], + "spans": [ + { + "bbox": [ + 131, + 546, + 220, + 560 + ], + "type": "text", + "content": "5 Conclusion" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 130, + 570, + 482, + 666 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 570, + 482, + 666 + ], + "spans": [ + { + "bbox": [ + 130, + 570, + 482, + 666 + ], + "type": "text", + "content": "We presented PatchRefiner, a tile-based framework tailored to real-world high-resolution monocular metric depth estimation. It reconceptualizes high-resolution depth estimation as a refinement process. With a pseudo-labeling strategy that leverages synthetic data, we propose a Detail and Scale Disentangling (DSD) loss to enhance detail capture while maintaining scale accuracy. Our proposed framework decisively surpasses the current SOTA method for UnrealStereo4K (17.3% in RMSE), while demonstrating marked improvements in detail accuracy on real-world datasets such as CityScape, ScanNet++, and ETH3D." + } + ] + } + ], + "index": 7 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "spans": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "type": "text", + "content": "14" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 166, + 91, + 351, + 102 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 166, + 91, + 351, + 102 + ], + "spans": [ + { + "bbox": [ + 166, + 91, + 351, + 102 + ], + "type": "text", + "content": "Zhenyu Li, Shariq Farooq Bhat, Peter Wonka" + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 13 + }, + { + "para_blocks": [ + { + "bbox": [ + 133, + 114, + 246, + 129 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 114, + 246, + 129 + ], + "spans": [ + { + "bbox": [ + 133, + 114, + 246, + 129 + ], + "type": "text", + "content": "Acknowledgements" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 132, + 138, + 479, + 162 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 138, + 479, + 162 + ], + "spans": [ + { + "bbox": [ + 132, + 138, + 479, + 162 + ], + "type": "text", + "content": "This publication is funded in part by KAUST under award #ORA-2023-5241 (for the NTGI-AI project)." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 133, + 178, + 197, + 190 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 178, + 197, + 190 + ], + "spans": [ + { + "bbox": [ + 133, + 178, + 197, + 190 + ], + "type": "text", + "content": "References" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 134, + 201, + 480, + 665 + ], + "type": "list", + "angle": 0, + "index": 22, + "blocks": [ + { + "bbox": [ + 138, + 201, + 480, + 222 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 201, + 480, + 222 + ], + "spans": [ + { + "bbox": [ + 138, + 201, + 480, + 222 + ], + "type": "text", + "content": "1. Bhat, S.F., Alhashim, I., Wonka, P.: Adabins: Depth estimation using adaptive bins. In: CVPR. pp. 4009-4018 (2021) 3" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 138, + 224, + 480, + 255 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 224, + 480, + 255 + ], + "spans": [ + { + "bbox": [ + 138, + 224, + 480, + 255 + ], + "type": "text", + "content": "2. Bhat, S.F., Birkl, R., Wofk, D., Wonka, P., Müller, M.: Zoedepth: Zero-shot transfer by combining relative and metric depth. arXiv preprint arXiv:2302.12288 (2023) 1, 3, 7, 9, 10, 11, 12" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 138, + 256, + 480, + 298 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 256, + 480, + 298 + ], + "spans": [ + { + "bbox": [ + 138, + 256, + 480, + 298 + ], + "type": "text", + "content": "3. Chen, C., Xie, W., Huang, W., Rong, Y., Ding, X., Huang, Y., Xu, T., Huang, J.: Progressive feature alignment for unsupervised domain adaptation. In: Proceedings of the IEEE/CVF conference on computer vision and pattern recognition. pp. 627-636 (2019) 3" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 138, + 300, + 480, + 319 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 300, + 480, + 319 + ], + "spans": [ + { + "bbox": [ + 138, + 300, + 480, + 319 + ], + "type": "text", + "content": "4. Chen, C., Chen, X., Cheng, H.: On the over-smoothing problem of cnn based disparity estimation. In: ICCV. pp. 8997-9005 (2019) 10" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 138, + 321, + 480, + 341 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 321, + 480, + 341 + ], + "spans": [ + { + "bbox": [ + 138, + 321, + 480, + 341 + ], + "type": "text", + "content": "5. Chen, W., Fu, Z., Yang, D., Deng, J.: Single-image depth perception in the wild. NeurIPS 29 (2016) 3, 7, 8" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 138, + 342, + 480, + 363 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 342, + 480, + 363 + ], + "spans": [ + { + "bbox": [ + 138, + 342, + 480, + 363 + ], + "type": "text", + "content": "6. Chen, Y., Liu, S., Wang, X.: Learning continuous image representation with local implicit image function. In: CVPR. pp. 8628-8638 (2021) 3" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 138, + 364, + 480, + 395 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 364, + 480, + 395 + ], + "spans": [ + { + "bbox": [ + 138, + 364, + 480, + 395 + ], + "type": "text", + "content": "7. Chen, Y.C., Lin, Y.Y., Yang, M.H., Huang, J.B.: Crdoco: Pixel-level domain transfer with cross-domain consistency. In: Proceedings of the IEEE/CVF conference on computer vision and pattern recognition. pp. 1791-1800 (2019) 3" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 138, + 396, + 480, + 427 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 396, + 480, + 427 + ], + "spans": [ + { + "bbox": [ + 138, + 396, + 480, + 427 + ], + "type": "text", + "content": "8. Chen, Z., Zhang, R., Zhang, G., Ma, Z., Lei, T.: Digging into pseudo label: a low-budget approach for semi-supervised semantic segmentation. IEEE Access 8, 41830-41837 (2020) 3" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 138, + 429, + 480, + 460 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 429, + 480, + 460 + ], + "spans": [ + { + "bbox": [ + 138, + 429, + 480, + 460 + ], + "type": "text", + "content": "9. Cordts, M., Omran, M., Ramos, S., Rehfeld, T., Enzweiler, M., Benenson, R., Franke, U., Roth, S., Schiele, B.: The cityscapes dataset for semantic urban scene understanding. In: CVPR. pp. 3213-3223 (2016) 3, 6, 9" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 134, + 461, + 480, + 492 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 134, + 461, + 480, + 492 + ], + "spans": [ + { + "bbox": [ + 134, + 461, + 480, + 492 + ], + "type": "text", + "content": "10. De Lutio, R., Becker, A., D'Aronco, S., Russo, S., Wegner, J.D., Schindler, K.: Learning graph regularisation for guided super-resolution. In: CVPR. pp. 1979-1988 (2022) 12" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 134, + 494, + 480, + 514 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 134, + 494, + 480, + 514 + ], + "spans": [ + { + "bbox": [ + 134, + 494, + 480, + 514 + ], + "type": "text", + "content": "11. Eigen, D., Puhrsch, C., Fergus, R.: Depth map prediction from a single image using a multi-scale deep network. NeurIPS 27 (2014) 1, 3, 7, 10, 11" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 134, + 515, + 480, + 546 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 134, + 515, + 480, + 546 + ], + "spans": [ + { + "bbox": [ + 134, + 515, + 480, + 546 + ], + "type": "text", + "content": "12. Farahani, A., Voghoei, S., Rasheed, K., Arabnia, H.R.: A brief review of domain adaptation. Advances in data science and information engineering: proceedings from ICDATA 2020 and IKE 2020 pp. 877-894 (2021) 2" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 134, + 548, + 480, + 568 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 134, + 548, + 480, + 568 + ], + "spans": [ + { + "bbox": [ + 134, + 548, + 480, + 568 + ], + "type": "text", + "content": "13. Fu, H., Gong, M., Wang, C., Batmanghelich, K., Tao, D.: Deep ordinal regression network for monocular depth estimation. In: CVPR. pp. 2002-2011 (2018) 3" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 134, + 569, + 480, + 589 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 134, + 569, + 480, + 589 + ], + "spans": [ + { + "bbox": [ + 134, + 569, + 480, + 589 + ], + "type": "text", + "content": "14. Geiger, A., Lenz, P., Urtasun, R.: Are we ready for autonomous driving? the kitti vision benchmark suite. In: CVPR. pp. 3354-3361. IEEE (2012) 5, 6" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 134, + 590, + 480, + 610 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 134, + 590, + 480, + 610 + ], + "spans": [ + { + "bbox": [ + 134, + 590, + 480, + 610 + ], + "type": "text", + "content": "15. Godard, C., Mac Aodha, O., Firman, M., Brostow, G.J.: Digging into self-supervised monocular depth estimation. In: ICCV. pp. 3828-3838 (2019) 3" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 134, + 612, + 480, + 632 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 134, + 612, + 480, + 632 + ], + "spans": [ + { + "bbox": [ + 134, + 612, + 480, + 632 + ], + "type": "text", + "content": "16. He, K., Gkioxari, G., Dollár, P., Girshick, R.: Mask r-cnn. In: Proceedings of the IEEE international conference on computer vision. pp. 2961-2969 (2017) 5" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 134, + 633, + 480, + 665 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 134, + 633, + 480, + 665 + ], + "spans": [ + { + "bbox": [ + 134, + 633, + 480, + 665 + ], + "type": "text", + "content": "17. Hu, Z., Yang, Z., Hu, X., Nevatia, R.: Simple: Similar pseudo label exploitation for semi-supervised classification. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 15099-15108 (2021) 3" + } + ] + } + ], + "index": 21 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 393, + 91, + 447, + 101 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 393, + 91, + 447, + 101 + ], + "spans": [ + { + "bbox": [ + 393, + 91, + 447, + 101 + ], + "type": "text", + "content": "PatchRefiner" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 470, + 92, + 481, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 470, + 92, + 481, + 100 + ], + "spans": [ + { + "bbox": [ + 470, + 92, + 481, + 100 + ], + "type": "text", + "content": "15" + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 14 + }, + { + "para_blocks": [ + { + "bbox": [ + 132, + 116, + 480, + 665 + ], + "type": "list", + "angle": 0, + "index": 20, + "blocks": [ + { + "bbox": [ + 133, + 116, + 480, + 138 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 116, + 480, + 138 + ], + "spans": [ + { + "bbox": [ + 133, + 116, + 480, + 138 + ], + "type": "text", + "content": "18. Huang, P.H., Matzen, K., Kopf, J., Ahuja, N., Huang, J.B.: Deepmvs: Learning multi-view stereopsis. In: CVPR (2018) 5" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 132, + 140, + 480, + 161 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 140, + 480, + 161 + ], + "spans": [ + { + "bbox": [ + 132, + 140, + 480, + 161 + ], + "type": "text", + "content": "19. Hui, T.W., Loy, C.C., Tang, X.: Depth map super-resolution by deep multi-scale guidance. In: ECCV. pp. 353-369. Springer (2016) 3" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 132, + 163, + 480, + 205 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 163, + 480, + 205 + ], + "spans": [ + { + "bbox": [ + 132, + 163, + 480, + 205 + ], + "type": "text", + "content": "20. Janoch, A., Karayev, S., Jia, Y., Barron, J.T., Fritz, M., Saenko, K., Darrell, T.: A category-level 3d object dataset: Putting the kinect to work. Consumer Depth Cameras for Computer Vision: Research Topics and Applications pp. 141-165 (2013) 5" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 132, + 206, + 480, + 239 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 206, + 480, + 239 + ], + "spans": [ + { + "bbox": [ + 132, + 206, + 480, + 239 + ], + "type": "text", + "content": "21. Kanopoulos, N., Vasanthavada, N., Baker, R.L.: Design of an image edge detection filter using the sobel operator. IEEE Journal of solid-state circuits 23(2), 358-367 (1988) 10" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 132, + 240, + 480, + 262 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 240, + 480, + 262 + ], + "spans": [ + { + "bbox": [ + 132, + 240, + 480, + 262 + ], + "type": "text", + "content": "22. Koutilya, P., Zhou, H., Jacobs, D.: Sharingan: Combining synthetic and real data for unsupervised geometry estimation. In: CVPR. vol. 2, p. 5 (2020) 3" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 132, + 263, + 480, + 296 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 263, + 480, + 296 + ], + "spans": [ + { + "bbox": [ + 132, + 263, + 480, + 296 + ], + "type": "text", + "content": "23. Kundu, J.N., Uppala, P.K., Pahuja, A., Babu, R.V.: Adadepth: Unsupervised content congruent adaptation for depth estimation. In: Proceedings of the IEEE conference on computer vision and pattern recognition. pp. 2656-2665 (2018) 3" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 132, + 297, + 480, + 319 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 297, + 480, + 319 + ], + "spans": [ + { + "bbox": [ + 132, + 297, + 480, + 319 + ], + "type": "text", + "content": "24. Lee, J.H., Kim, C.S.: Multi-loss rebalancing algorithm for monocular depth estimation. In: ECCV. pp. 785-801. Springer (2020) 3, 7" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 132, + 319, + 480, + 352 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 319, + 480, + 352 + ], + "spans": [ + { + "bbox": [ + 132, + 319, + 480, + 352 + ], + "type": "text", + "content": "25. Lehtinen, J., Munkberg, J., Hasselgren, J., Laine, S., Karras, T., Aittala, M., Aila, T.: Noise2noise: Learning image restoration without clean data. arXiv preprint arXiv:1803.04189 (2018) 5" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 132, + 353, + 480, + 386 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 353, + 480, + 386 + ], + "spans": [ + { + "bbox": [ + 132, + 353, + 480, + 386 + ], + "type": "text", + "content": "26. Li, Y., Yuan, L., Vasconcelos, N.: Bidirectional learning for domain adaptation of semantic segmentation. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 6936-6945 (2019) 3" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 132, + 386, + 480, + 418 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 386, + 480, + 418 + ], + "spans": [ + { + "bbox": [ + 132, + 386, + 480, + 418 + ], + "type": "text", + "content": "27. Li, Z., Bhat, S.F., Wonka, P.: Patchfusion: An end-to-end tile-based framework for high-resolution monocular metric depth estimation. arXiv preprint arXiv:2312.02284 (2023) 1, 2, 3, 4, 5, 6, 7, 9, 10, 11, 12, 13" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 132, + 419, + 480, + 453 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 419, + 480, + 453 + ], + "spans": [ + { + "bbox": [ + 132, + 419, + 480, + 453 + ], + "type": "text", + "content": "28. Li, Z., Chen, Z., Li, A., Fang, L., Jiang, Q., Liu, X., Jiang, J.: Unsupervised domain adaptation for monocular 3d object detection via self-training. In: European Conference on Computer Vision. pp. 245-262. Springer (2022) 3" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 132, + 453, + 480, + 486 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 453, + 480, + 486 + ], + "spans": [ + { + "bbox": [ + 132, + 453, + 480, + 486 + ], + "type": "text", + "content": "29. Li, Z., Chen, Z., Liu, X., Jiang, J.: Depthformer: Exploiting long-range correlation and local information for accurate monocular depth estimation. Machine Intelligence Research pp. 1-18 (2023) 3" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 132, + 487, + 480, + 509 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 487, + 480, + 509 + ], + "spans": [ + { + "bbox": [ + 132, + 487, + 480, + 509 + ], + "type": "text", + "content": "30. Li, Z., Wang, X., Liu, X., Jiang, J.: Binsformer: Revisiting adaptive bins for monocular depth estimation. arXiv preprint arXiv:2204.00987 (2022) 1, 3, 7, 11" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 132, + 509, + 480, + 540 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 509, + 480, + 540 + ], + "spans": [ + { + "bbox": [ + 132, + 509, + 480, + 540 + ], + "type": "text", + "content": "31. Liu, C., Kumar, S., Gu, S., Timofte, R., Van Gool, L.: Single image depth prediction made better: A multivariate gaussian take. In: CVPR. pp. 17346-17356 (June 2023) 7" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 132, + 543, + 480, + 575 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 543, + 480, + 575 + ], + "spans": [ + { + "bbox": [ + 132, + 543, + 480, + 575 + ], + "type": "text", + "content": "32. Lopez-Rodriguez, A., Mikolajczyk, K.: Desc: Domain adaptation for depth estimation via semantic consistency. International Journal of Computer Vision 131(3), 752-771 (2023) 3, 4, 6" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 132, + 576, + 480, + 598 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 576, + 480, + 598 + ], + "spans": [ + { + "bbox": [ + 132, + 576, + 480, + 598 + ], + "type": "text", + "content": "33. Metzger, N., Daudt, R.C., Schindler, K.: Guided depth super-resolution by deep anisotropic diffusion. In: CVPR. pp. 18237-18246 (2023) 3" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 132, + 599, + 480, + 632 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 599, + 480, + 632 + ], + "spans": [ + { + "bbox": [ + 132, + 599, + 480, + 632 + ], + "type": "text", + "content": "34. Miangoleh, S.M.H., Dille, S., Mai, L., Paris, S., Aksoy, Y.: Boosting monocular depth estimation models to high-resolution via content-adaptive multi-resolution merging. In: CVPR. pp. 9685-9694 (2021) 1, 2, 3, 4, 5, 12" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 132, + 632, + 480, + 665 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 632, + 480, + 665 + ], + "spans": [ + { + "bbox": [ + 132, + 632, + 480, + 665 + ], + "type": "text", + "content": "35. Mildenhall, B., Srinivasan, P.P., Tancik, M., Barron, J.T., Ramamoorthi, R., Ng, R.: Nerf: Representing scenes as neural radiance fields for view synthesis. Communications of the ACM 65(1), 99-106 (2021) 3" + } + ] + } + ], + "index": 19 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "spans": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "type": "text", + "content": "16" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 166, + 91, + 352, + 102 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 166, + 91, + 352, + 102 + ], + "spans": [ + { + "bbox": [ + 166, + 91, + 352, + 102 + ], + "type": "text", + "content": "Zhenyu Li, Shariq Farooq Bhat, Peter Wonka" + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 15 + }, + { + "para_blocks": [ + { + "bbox": [ + 130, + 116, + 482, + 665 + ], + "type": "list", + "angle": 0, + "index": 18, + "blocks": [ + { + "bbox": [ + 130, + 116, + 482, + 160 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 116, + 482, + 160 + ], + "spans": [ + { + "bbox": [ + 130, + 116, + 482, + 160 + ], + "type": "text", + "content": "36. Pastore, G., Cermelli, F., Xian, Y., Mancini, M., Akata, Z., Caputo, B.: A closer look at self-training for zero-label semantic segmentation. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 2693-2702 (2021) 3" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 130, + 161, + 481, + 205 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 161, + 481, + 205 + ], + "spans": [ + { + "bbox": [ + 130, + 161, + 481, + 205 + ], + "type": "text", + "content": "37. Paul, S., Tsai, Y.H., Schulter, S., Roy-Chowdhury, A.K., Chandraker, M.: Domain adaptive semantic segmentation using weak labels. In: Computer Vision-ECCV 2020: 16th European Conference, Glasgow, UK, August 23-28, 2020, Proceedings, Part IX 16. pp. 571-587. Springer (2020) 3" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 132, + 205, + 481, + 237 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 205, + 481, + 237 + ], + "spans": [ + { + "bbox": [ + 132, + 205, + 481, + 237 + ], + "type": "text", + "content": "38. Piccinelli, L., Sakaridis, C., Yu, F.: idisc: Internal discretization for monocular depth estimation. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 21477-21487 (2023) 10, 12" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 132, + 238, + 481, + 270 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 238, + 481, + 270 + ], + "spans": [ + { + "bbox": [ + 132, + 238, + 481, + 270 + ], + "type": "text", + "content": "39. Poucin, F., Kraus, A., Simon, M.: Boosting instance segmentation with synthetic data: A study to overcome the limits of real world data sets. In: Int. Conf. Comput. Vis. Workshop. pp. 945-953 (2021) 2, 5, 6, 12" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 132, + 271, + 481, + 303 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 271, + 481, + 303 + ], + "spans": [ + { + "bbox": [ + 132, + 271, + 481, + 303 + ], + "type": "text", + "content": "40. Pseudo-Label, D.H.L.: The simple and efficient semi-supervised learning method for deep neural networks. In: ICML 2013 Workshop: Challenges in Representation Learning. pp. 1-6 (2013) 3" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 132, + 304, + 481, + 346 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 304, + 481, + 346 + ], + "spans": [ + { + "bbox": [ + 132, + 304, + 481, + 346 + ], + "type": "text", + "content": "41. Rajpal, A., Cheema, N., Illgner-Fehns, K., Slusallek, P., Jaiswal, S.: High-resolution synthetic rgb-d datasets for monocular depth estimation. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 1188-1198 (2023) 3, 6" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 132, + 347, + 481, + 380 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 347, + 481, + 380 + ], + "spans": [ + { + "bbox": [ + 132, + 347, + 481, + 380 + ], + "type": "text", + "content": "42. Ranftl, R., Lasinger, K., Hafner, D., Schindler, K., Koltun, V.: Towards robust monocular depth estimation: Mixing datasets for zero-shot cross-dataset transfer. IEEE TPAMI 44(3) (2022) 3, 7, 8" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 132, + 380, + 481, + 402 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 380, + 481, + 402 + ], + "spans": [ + { + "bbox": [ + 132, + 380, + 481, + 402 + ], + "type": "text", + "content": "43. Rey-Area, M., Yuan, M., Richardt, C.: 360monodepth: High-resolution 360deg monocular depth estimation. In: CVPR. pp. 3762-3772 (2022) 1, 2, 3" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 132, + 403, + 481, + 434 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 403, + 481, + 434 + ], + "spans": [ + { + "bbox": [ + 132, + 403, + 481, + 434 + ], + "type": "text", + "content": "44. Saito, K., Ushiku, Y., Harada, T.: Asymmetric tri-training for unsupervised domain adaptation. In: International Conference on Machine Learning. pp. 2988-2997. PMLR (2017) 3" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 132, + 435, + 481, + 479 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 435, + 481, + 479 + ], + "spans": [ + { + "bbox": [ + 132, + 435, + 481, + 479 + ], + "type": "text", + "content": "45. Scharstein, D., Hirschmüller, H., Kitajima, Y., Krathwohl, G., Nesic, N., Wang, X., Westling, P.: High-resolution stereo datasets with subpixel-accurate ground truth. In: Pattern Recognition: 36th German Conference, GCPR 2014, Münster, Germany, September 2-5, 2014, Proceedings 36. pp. 31-42. Springer (2014) 6, 10" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 132, + 479, + 481, + 511 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 479, + 481, + 511 + ], + "spans": [ + { + "bbox": [ + 132, + 479, + 481, + 511 + ], + "type": "text", + "content": "46. Schops, T., Schonberger, J.L., Galliani, S., Sattler, T., Schindler, K., Pollefeys, M., Geiger, A.: A multi-view stereo benchmark with high-resolution images and multi-camera videos. In: CVPR. pp. 3260-3269 (2017) 3, 5, 6, 10" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 132, + 512, + 481, + 555 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 512, + 481, + 555 + ], + "spans": [ + { + "bbox": [ + 132, + 512, + 481, + 555 + ], + "type": "text", + "content": "47. Shin, I., Tsai, Y.H., Zhuang, B., Schulter, S., Liu, B., Garg, S., Kweon, I.S., Yoon, K.J.: Mm-tta: multi-modal test-time adaptation for 3d semantic segmentation. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 16928-16937 (2022) 3" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 132, + 555, + 481, + 578 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 555, + 481, + 578 + ], + "spans": [ + { + "bbox": [ + 132, + 555, + 481, + 578 + ], + "type": "text", + "content": "48. Silberman, N., Hoiem, D., Kohli, P., Fergus, R.: Indoor segmentation and support inference from rgbd images. In: ECCV. pp. 746-760. Springer (2012) 5, 6, 10, 11" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 132, + 578, + 481, + 599 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 578, + 481, + 599 + ], + "spans": [ + { + "bbox": [ + 132, + 578, + 481, + 599 + ], + "type": "text", + "content": "49. Song, S., Lichtenberg, S.P., Xiao, J.: Sun rgb-d: A rgb-d scene understanding benchmark suite. In: CVPR. pp. 567-576 (2015) 5, 6, 10" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 132, + 600, + 481, + 643 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 600, + 481, + 643 + ], + "spans": [ + { + "bbox": [ + 132, + 600, + 481, + 643 + ], + "type": "text", + "content": "50. Taherkhani, F., Dabouei, A., Soleymani, S., Dawson, J., Nasrabadi, N.M.: Self-supervised Wasserstein pseudo-labeling for semi-supervised image classification. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 12267-12277 (2021) 3" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 132, + 644, + 481, + 665 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 644, + 481, + 665 + ], + "spans": [ + { + "bbox": [ + 132, + 644, + 481, + 665 + ], + "type": "text", + "content": "51. Tosi, F., Liao, Y., Schmitt, C., Geiger, A.: Smd-nets: Stereo mixture density networks. In: CVPR. pp. 8942-8952 (2021) 2, 3, 5, 9, 10, 12" + } + ] + } + ], + "index": 17 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 393, + 91, + 447, + 100 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 393, + 91, + 447, + 100 + ], + "spans": [ + { + "bbox": [ + 393, + 91, + 447, + 100 + ], + "type": "text", + "content": "PatchRefiner" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 470, + 91, + 481, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 470, + 91, + 481, + 100 + ], + "spans": [ + { + "bbox": [ + 470, + 91, + 481, + 100 + ], + "type": "text", + "content": "17" + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 16 + }, + { + "para_blocks": [ + { + "bbox": [ + 132, + 116, + 480, + 533 + ], + "type": "list", + "angle": 0, + "index": 17, + "blocks": [ + { + "bbox": [ + 132, + 116, + 480, + 138 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 116, + 480, + 138 + ], + "spans": [ + { + "bbox": [ + 132, + 116, + 480, + 138 + ], + "type": "text", + "content": "52. Wang, Z., Bovik, A.C., Sheikh, H.R., Simoncelli, E.P.: Image quality assessment: from error visibility to structural similarity. IEEE TIP 13(4), 600-612 (2004) 9" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 132, + 140, + 480, + 160 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 140, + 480, + 160 + ], + "spans": [ + { + "bbox": [ + 132, + 140, + 480, + 160 + ], + "type": "text", + "content": "53. Weiss, K., Khoshgoftaar, T.M., Wang, D.: A survey of transfer learning. Journal of Big data 3(1), 1-40 (2016) 2" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 133, + 161, + 480, + 182 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 161, + 480, + 182 + ], + "spans": [ + { + "bbox": [ + 133, + 161, + 480, + 182 + ], + "type": "text", + "content": "54. Xian, K., Zhang, J., Wang, O., Mai, L., Lin, Z., Cao, Z.: Structure-guided ranking loss for single image depth prediction. In: CVPR. pp. 611-620 (2020) 3, 7, 8" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 133, + 183, + 480, + 215 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 183, + 480, + 215 + ], + "spans": [ + { + "bbox": [ + 133, + 183, + 480, + 215 + ], + "type": "text", + "content": "55. Yang, J., Alvarez, J.M., Liu, M.: Self-supervised learning of depth inference for multi-view stereo. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 7526-7534 (2021) 4" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 132, + 216, + 480, + 247 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 216, + 480, + 247 + ], + "spans": [ + { + "bbox": [ + 132, + 216, + 480, + 247 + ], + "type": "text", + "content": "56. Yang, L., Kang, B., Huang, Z., Xu, X., Feng, J., Zhao, H.: Depth anything: Unleashing the power of large-scale unlabeled data. arXiv preprint arXiv:2401.10891 (2024) 1, 3, 4, 6, 7, 8, 12" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 132, + 249, + 480, + 281 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 249, + 480, + 281 + ], + "spans": [ + { + "bbox": [ + 132, + 249, + 480, + 281 + ], + "type": "text", + "content": "57. Yen, Y.T., Lu, C.N., Chiu, W.C., Tsai, Y.H.: 3d-pl: Domain adaptive depth estimation with 3d-aware pseudo-labeling. In: European Conference on Computer Vision. pp. 710-728. Springer (2022) 4, 6" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 132, + 282, + 480, + 303 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 282, + 480, + 303 + ], + "spans": [ + { + "bbox": [ + 132, + 282, + 480, + 303 + ], + "type": "text", + "content": "58. Yeshwanth, C., Liu, Y.C., Nießner, M., Dai, A.: Scannet++: A high-fidelity dataset of 3d indoor scenes. In: ICCV. pp. 12-22 (2023) 3, 5, 6, 10" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 132, + 304, + 480, + 325 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 304, + 480, + 325 + ], + "spans": [ + { + "bbox": [ + 132, + 304, + 480, + 325 + ], + "type": "text", + "content": "59. Zhang, L., Rao, A., Agrawala, M.: Adding conditional control to text-to-image diffusion models. In: ICCV. pp. 3836-3847 (2023) 1" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 132, + 326, + 480, + 357 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 326, + 480, + 357 + ], + "spans": [ + { + "bbox": [ + 132, + 326, + 480, + 357 + ], + "type": "text", + "content": "60. Zhao, S., Fu, H., Gong, M., Tao, D.: Geometry-aware symmetric domain adaptation for monocular depth estimation. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 9788-9798 (2019) 3" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 132, + 358, + 480, + 379 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 358, + 480, + 379 + ], + "spans": [ + { + "bbox": [ + 132, + 358, + 480, + 379 + ], + "type": "text", + "content": "61. Zhao, W., Rao, Y., Liu, Z., Liu, B., Zhou, J., Lu, J.: Unleashing text-to-image diffusion models for visual perception. ICCV (2023) 1" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 133, + 380, + 480, + 423 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 380, + 480, + 423 + ], + "spans": [ + { + "bbox": [ + 133, + 380, + 480, + 423 + ], + "type": "text", + "content": "62. Zhao, X., Schulter, S., Sharma, G., Tsai, Y.H., Chandraker, M., Wu, Y.: Object detection with a unified label space from multiple datasets. In: Computer Vision-ECCV 2020: 16th European Conference, Glasgow, UK, August 23-28, 2020, Proceedings, Part XIV 16. pp. 178-193. Springer (2020) 3" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 132, + 424, + 480, + 445 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 424, + 480, + 445 + ], + "spans": [ + { + "bbox": [ + 132, + 424, + 480, + 445 + ], + "type": "text", + "content": "63. Zhao, Z., Zhang, J., Xu, S., Lin, Z., Pfister, H.: Discrete cosine transform network for guided depth map super-resolution. In: CVPR. pp. 5697-5707 (2022) 3" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 132, + 447, + 480, + 479 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 447, + 480, + 479 + ], + "spans": [ + { + "bbox": [ + 132, + 447, + 480, + 479 + ], + "type": "text", + "content": "64. Zheng, C., Cham, T.J., Cai, J.: T2net: Synthetic-to-realistic translation for solving single-image depth estimation tasks. In: Proceedings of the European conference on computer vision (ECCV). pp. 767-783 (2018) 3" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 132, + 479, + 480, + 500 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 479, + 480, + 500 + ], + "spans": [ + { + "bbox": [ + 132, + 479, + 480, + 500 + ], + "type": "text", + "content": "65. Zhong, Z., Liu, X., Jiang, J., Zhao, D., Ji, X.: Guided depth map super-resolution: A survey. ACM Computing Surveys (2023) 3" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 132, + 501, + 480, + 533 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 501, + 480, + 533 + ], + "spans": [ + { + "bbox": [ + 132, + 501, + 480, + 533 + ], + "type": "text", + "content": "66. Zou, Y., Yu, Z., Kumar, B., Wang, J.: Unsupervised domain adaptation for semantic segmentation via class-balanced self-training. In: Proceedings of the European conference on computer vision (ECCV). pp. 289-305 (2018) 3" + } + ] + } + ], + "index": 16 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "spans": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "type": "text", + "content": "18" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 166, + 91, + 352, + 102 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 166, + 91, + 352, + 102 + ], + "spans": [ + { + "bbox": [ + 166, + 91, + 352, + 102 + ], + "type": "text", + "content": "Zhenyu Li, Shariq Farooq Bhat, Peter Wonka" + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 17 + } + ], + "_backend": "vlm", + "_version_name": "2.6.4" +} \ No newline at end of file diff --git a/2024/PathMMU_ A Massive Multimodal Expert-Level Benchmark for Understanding and Reasoning in Pathology/0e28fdb1-202a-41b0-8b8d-6b1d1b08e3fe_content_list.json b/2024/PathMMU_ A Massive Multimodal Expert-Level Benchmark for Understanding and Reasoning in Pathology/0e28fdb1-202a-41b0-8b8d-6b1d1b08e3fe_content_list.json new file mode 100644 index 0000000000000000000000000000000000000000..107ab2334de16384953a83cc9fe14d026499bb86 --- /dev/null +++ b/2024/PathMMU_ A Massive Multimodal Expert-Level Benchmark for Understanding and Reasoning in Pathology/0e28fdb1-202a-41b0-8b8d-6b1d1b08e3fe_content_list.json @@ -0,0 +1,1561 @@ +[ + { + "type": "text", + "text": "PathMMU: A Massive Multimodal Expert-Level Benchmark for Understanding and Reasoning in Pathology", + "text_level": 1, + "bbox": [ + 223, + 140, + 779, + 209 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Yuxuan Sun $^{1,2}$ , Hao Wu $^{3}$ , Chenglu Zhu $^{2}$ , Sunyi Zheng $^{2}$ , Qizi Chen $^{4}$ , Kai Zhang $^{5}$ , Yunlong Zhang $^{1,2}$ , Dan Wan $^{6}$ , Xiaoxiao Lan $^{1}$ , Mengyue Zheng $^{2}$ , Jingxiong Li $^{1,2}$ , Xinheng Lyu $^{2}$ , Tao Lin $^{2,\\dagger}$ , and Lin Yang $^{2,\\dagger}$", + "bbox": [ + 217, + 233, + 785, + 282 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "1 Zhejiang University, Hangzhou 310058, China", + "bbox": [ + 341, + 291, + 661, + 306 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "$^{2}$ Westlake University, Hangzhou 310030, China", + "bbox": [ + 339, + 306, + 661, + 320 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "{sunyuxuan,zhuchenglu,lintao,yanglin}@westlake.edu.cn", + "bbox": [ + 297, + 321, + 707, + 334 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "3 Macau University of Science and Technology, Macau 999078, China", + "bbox": [ + 269, + 334, + 730, + 347 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "2220024311@student.must.edu.mo", + "bbox": [ + 383, + 349, + 617, + 361 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "$^{4}$ Jiangnan University, Wuxi 214122, China", + "bbox": [ + 354, + 361, + 647, + 375 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "5 The Ohio State University, Columbus 43210, USA", + "bbox": [ + 326, + 375, + 674, + 388 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "$^{6}$ Fujian University of Traditional Chinese Medicine, Fuzhou 350122, China", + "bbox": [ + 250, + 388, + 751, + 402 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "https://pathmmu-benchmark.github.io/", + "bbox": [ + 359, + 404, + 640, + 417 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Abstract. The emergence of Large Multimodal Models (LMMs) has unlocked remarkable potential in AI, particularly in pathology. However, the lack of specialized, high-quality benchmark impeded their development and precise evaluation. To address this, we introduce PathMMU, the largest and highest-quality expert validated pathology benchmark for LMMs. It comprises 33,428 multimodal multi-choice questions and 24,067 images from various sources, each accompanied by an explanation for the correct answer. The construction of PathMMU leverages GPT-4V's advanced capabilities, utilizing over 30,000 image-caption pairs to enrich the descriptive quality of captions and generate corresponding Q&As in a cascading process. To maximize PathMMU's authority, we invite seven pathologists to scrutinize each question under strict standards in PathMMU's validation and test sets, while simultaneously setting an expert-level performance benchmark for PathMMU. We conduct extensive evaluations, including zero-shot assessments of 14 open-sourced and 4 closed-sourced LMMs and their robustness to image corruption. We also fine-tune representative LMMs to assess their adaptability to PathMMU. The empirical findings indicate that advanced LMMs struggle with the challenging PathMMU benchmark, with the top-performing LMM, GPT-4V, achieving only a $49.8\\%$ zero-shot performance, significantly lower than the $71.8\\%$ demonstrated by human pathologists. After fine-tuning, substantially smaller open-sourced LMMs can outperform GPT-4V but still fall short of the expertise shown by pathologists. We hope that the PathMMU will offer valuable insights and foster the development of more specialized, next-generation LMMs for pathology.", + "bbox": [ + 259, + 446, + 740, + 792 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Keywords: Large multimodal model $\\cdot$ Pathology $\\cdot$ Benchmark", + "bbox": [ + 261, + 805, + 686, + 820 + ], + "page_idx": 0 + }, + { + "type": "page_footnote", + "text": "† Corresponding author.", + "bbox": [ + 220, + 824, + 385, + 839 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "1 Introduction", + "text_level": 1, + "bbox": [ + 217, + 143, + 374, + 160 + ], + "page_idx": 1 + }, + { + "type": "image", + "img_path": "images/50434f96daebcdae5f9726c2f375bba90c25bf354d535fffe668de2d0405ab41.jpg", + "image_caption": [ + "Fig. 1: An overview of the PathMMU benchmark: PathMMU is built from diverse, rich data sources. It comprises expert-level, multimodal, multiple-choice questions in pathology, collaboratively crafted by AI and pathology experts. The benchmark reveals even the most advanced LMMs significantly underperform compared to human experts." + ], + "image_footnote": [], + "bbox": [ + 218, + 200, + 372, + 345 + ], + "page_idx": 1 + }, + { + "type": "image", + "img_path": "images/fab67f6ef220219bc7e6430a568fd532310a1b2d6726a7c43dbbf7fdec5810f3.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 383, + 200, + 584, + 347 + ], + "page_idx": 1 + }, + { + "type": "image", + "img_path": "images/688b82a269cf86c14034923178aa4720a73a259f9bc7399aa7c9cd3a6df5c92c.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 589, + 200, + 784, + 347 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Pathology is integral to modern medicine, serving as the foundation for diagnosing and understanding diseases. For instance, liquid-based cytology is crucial for early cervical cancer detection. Assessing biomarkers like HER2 in breast cancer guides the selection of targeted therapies and immunotherapies [23].", + "bbox": [ + 212, + 446, + 784, + 508 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "In recent years, the field of pathology has undergone a significant transformation, driven by advances in digital pathology and the integration of AI. This shift marks a departure from conventional microscope-based slide reading to AI-powered image analysis, greatly easing the workload of pathologists. Traditional pathology models are tailored for specific tasks, such as cervical cytology and liver lesion classification, leading to an abundance of task-specific models [57,59]. In contrast, recent advancements in Large Multimodal Models (LMMs) focus on offering general task-solving capabilities, thereby making universal pathological region recognition possible. This breakthrough not only represents a significant stride in the field, but also paves the way for more versatile and efficient diagnostic approaches in pathology [13,27,42,44].", + "bbox": [ + 212, + 508, + 785, + 672 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Given the demands of precise interpretation in pathology, conducting comprehensive evaluations of LMMs' abilities in interpreting pathology images is essential. However, the field faces a notable scarcity of high-quality benchmark datasets. Currently, the primary large-scale dataset available is PathVQA [18], which offers over 30,000 visual Q&A samples with 4,998 images. Nevertheless, these samples are derived from limited sources of textbooks, with image captions being converted into questions through heuristic approaches, constraining the generation of diverse and logical Q&As. Additionally, image captions in the textbook might not accurately reflect the corresponding images. For instance, some captions may introduce extraneous information not evident in the images or only partially cover the image content, thereby leading to unsolvable Q&As", + "bbox": [ + 212, + 674, + 787, + 840 + ], + "page_idx": 1 + }, + { + "type": "page_number", + "text": "2", + "bbox": [ + 217, + 114, + 228, + 126 + ], + "page_idx": 1 + }, + { + "type": "header", + "text": "Y. Sun et al.", + "bbox": [ + 271, + 114, + 357, + 126 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "or questions that inadequately capture the essence of the images. Moreover, the absence of expert review and validation during the data curation process may introduce substantial noise into the dataset. These challenges pose significant obstacles to effectively validating LMMs' capabilities in pathology.", + "bbox": [ + 212, + 146, + 787, + 207 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "To address these challenges, we introduce PathMMU, a multimodal expert-level benchmark designed to evaluate LMMs in pathology image understanding and reasoning, comprising 33,428 Q&As along with 24,067 pathology images. As depicted in Fig. 1, PathMMU draws from a diverse range of high-quality sources, including PubMed scientific papers, pathology atlas from guidelines, Twitter posts by pathology experts, commonly used pathology classification datasets and educational content from YouTube videos, covering multiple organ systems (e.g., gastrointestinal, pulmonary, etc.) and multiple subjects (e.g., dermatopathology, hematopathology, etc.). In developing PathMMU, we adopt a meticulously designed cascading approach. Initially, we prompt GPT-4V with collected image_caption pairs to enhance the original captions, crafting descriptions that delve deeper into details and highlight crucial morphological features. Subsequently, we use these enhanced descriptions along with the images to prompt GPT-4V to generate professional multi-choice, multimodal pathology Q&As, each accompanied by detailed explanations for their answers. To ensure the questions are specifically designed for multimodal pathology understanding, we employ a collaboration of multiple Large Language Models (LLMs) to eliminate questions that can be solved or guessed using text alone. Most importantly, we invite seven pathology experts to manually review approximately 12,000 Q&As from the test and validation sets of PathMMU. The above carefully executed procedures, ensure the generation of professional, logical, and high-quality Q&As.", + "bbox": [ + 212, + 213, + 787, + 532 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "We evaluate the zero-shot performance of 18 advanced LMMs and their robustness on the PathMMU test sets. We also fine-tune two representative LMMs on the training set to assess their transfer learning capabilities. The key insights from this extensive analysis are: (1) Cutting-edge LMMs struggle with the PathMMU, with 15 out of the 18 achieving no more than $40\\%$ accuracy. Even the most advanced model currently available, GPT-4V, only attains an accuracy of $49.8\\%$ . This reveals a notable discrepancy of nearly $20\\%$ compared to professional pathologists, indicating significant deficiencies of current LMMs in the specialized field of pathology and emphasizing the challenging nature of the PathMMU benchmark. (2) While these LMMs show considerable robustness in robustness tests, the full extent of their robustness is still in question. On the one hand, given their limited performance, the scope for further performance degradation is somewhat constrained. On the other hand, these models tend to capture only superficial image features, and sometimes may not even utilize image information for reasoning. As a result, even when image corruption obscures minor details, the impact on these models' performance might not be significantly pronounced. (3) The text-only performance of GPT-4 Turbo surpasses the top-performing open-source LMMs on the PathMMU benchmark. We observe that GPT-4 Turbo sometimes takes shortcuts to guess answers. This involves choosing the correct answer based", + "bbox": [ + 212, + 537, + 789, + 843 + ], + "page_idx": 2 + }, + { + "type": "header", + "text": "PathMMU", + "bbox": [ + 656, + 114, + 730, + 126 + ], + "page_idx": 2 + }, + { + "type": "page_number", + "text": "3", + "bbox": [ + 774, + 116, + 785, + 126 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "either on the options most commonly encountered in real-world pathological scenarios, or by identifying the most distinctive option, rather than conducting an in-depth analysis of the pathology image. (4) Fine-tuning general-purpose LMMs with substantial amounts of data significantly boosts their ability to comprehend and reason with pathological images, enabling them to easily surpass the zero-shot performance of GPT-4V. However, there still remains a discernible gap between their performance and that of human experts.", + "bbox": [ + 212, + 146, + 787, + 253 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "2 Related Works", + "text_level": 1, + "bbox": [ + 215, + 277, + 398, + 292 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Multimodal Models. The emergence of powerful large language models such as BERT [12], GPT-3 [8], T5 [38], LLaMa [46], ChatGPT [33], GPT-4 [34], and Vicuna [10] has significantly advanced the field of Natural Language Processing (NLP). These models demonstrate exceptional general capabilities, motivating researchers to explore the integration of LLMs with vision models to develop versatile multimodal models. This integration is primarily achieved through the pretraining approach, leading to the creation of groundbreaking LMMs. Representative models of this approach include CLIP [37], Flamingo [1], BLIP-2 [29], and Fuyu [5], all of which demonstrate remarkable multimodal understanding abilities. Furthermore, by adapting instruction-tuning techniques from the NLP domain, a variety of multimodal instruction-tuning datasets have been developed. This initiative aims to guide LMMs in generating outputs that are more controllable, practical, and adaptable to various tasks. Simultaneously, this approach facilitates the integration of LLMs into LMMs using lightweight fine-tuning, significantly reducing both costs and time. Key models that embody this advancement include GPT-4V [35], Gemini Pro Vision [45], Qwen-VL [4], LLaVA [30], MiniGPT-4 [60], BLIP-2 [29] and InstructBLIP [11].", + "bbox": [ + 212, + 310, + 787, + 568 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "LMM Benchmarks. In the general domain, the rise of various large models has led to the creation of extensive benchmark datasets, designed to evaluate the universal capabilities of these models across a range of tasks and domains. Notable examples include LAMM [52], LVLM-eHub [51], SEED [26], MMBench [31], MMVet [53] and BenchLMM [9], which have been used to assess the basic perception abilities of large models. More recently, MMMU [54], a massive and challenging dataset covering 30 university-level subjects, has been developed. It is specifically designed to evaluate the general multimodal understanding and reasoning capabilities of LMMs. Furthermore, HaELM [48] and HallusionBench [16] are proposed to evaluate the degree of hallucination in LMMs.", + "bbox": [ + 212, + 568, + 789, + 718 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "In the medical field, datasets such as VQA-RAD [24] and VQA-Med [6] provide a collection of Q&A pairs based on radiology images, facilitating the evaluation of the capabilities of AI in radiological diagnostics. Additionally, PMC-VQA [55] is a large-scale dataset within this domain. It generates a vast number of Q&A pairs by prompting ChatGPT with text-only image captions from PubMed. In the pathology domain, limited dataset construction efforts have been performed to date. The representative large-scale dataset available, PathVQA [18], is constructed in a manner similar to PMC-VQA. It utilizes", + "bbox": [ + 212, + 719, + 787, + 840 + ], + "page_idx": 3 + }, + { + "type": "page_number", + "text": "4", + "bbox": [ + 217, + 114, + 230, + 126 + ], + "page_idx": 3 + }, + { + "type": "header", + "text": "Y. Sun et al.", + "bbox": [ + 271, + 114, + 359, + 128 + ], + "page_idx": 3 + }, + { + "type": "image", + "img_path": "images/1c94cd14cf7507c2335dd8ab8c118fad25088a4fc7f6416c15ae7aad867fb3c4.jpg", + "image_caption": [ + "Fig. 2: The comparison between PathMMU and existing benchmarks. The Q&A pairs in PathMMU are sourced extensively and comprehensively, undergoing rigorous multi-tiered filtering. This includes the initial filtering by multiple LLMs and the strict reviews by professional pathologists. Additionally, each question is accompanied by a detailed explanation. These attributes establish PathMMU as the most professionally curated, comprehensive, and highest-quality large-scale pathology dataset available." + ], + "image_footnote": [], + "bbox": [ + 222, + 143, + 395, + 250 + ], + "page_idx": 4 + }, + { + "type": "table", + "img_path": "images/e63abd2c46baa4a1e1e83c211a1b439f3148e16d827b3ce243a771e3f2fd48f2.jpg", + "table_caption": [], + "table_footnote": [], + "table_body": "
DatasetDomainSources# Q&As / ImagesAnswer Explanation?LLM Filtered?Expert Annotated?
VQA-MedMedicalMedPix5000 / 5000XX✓
VQA-RADMedicalMedPix3515 / 315XX✓
PMC-VQAMedicalPMC227k / 149kXXX
PathVQAPathologyBook32799 / 4998XXX
Quilt-VQAPathologyEdu-Content1283 / 985XX✓
Quilt-Red
Social Media
PathMMUPathologyEdu-Content, PubMed33428 / 24067✓✓✓
Atlas, CLS-Data
", + "bbox": [ + 401, + 143, + 782, + 250 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "a heuristic approach to generate questions based on the text-only captions of pathology textbook images, yielding over 30,000 Q&As and 4,998 images. More recently, Quilt-VQA [39] is proposed, which is constructed by automatically extracting question and answer pairs from the narrations of professional teachers in YouTube videos. This approach guarantees the professionalism of the Q&A pairs, although it comprises a relatively smaller dataset with 1,283 Q&As and 985 images. Nonetheless, datasets reliant solely on image captions for question generation, face several notable limitations: (1) The creation of questions typically requires the consideration of both the image and its accompanying caption. Relying only on captions can introduce inaccuracies, making some questions unanswerable or solvable using only the text, bypassing the need for image analysis. (2) Since captions tend to be simplistic and may not capture the detailed information evident in images, generating an excessive number of Q&As from the same image's caption can lead to either too simplistic or insufficient questions. (3) Automatically generated questions, whether through predefined rules or by prompting ChatGPT, are susceptible to containing numerous errors.", + "bbox": [ + 212, + 372, + 787, + 613 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Given the current lack of high-quality benchmarks in pathology, we develop PathMMU, a large-scale, high-quality, and expert-verified pathology benchmark, to fill this gap. Our approach involves generating Q&As from a diverse range of image-caption pairs, followed by rigorous expert review and filtering applied to both the images and the questions. Moreover, we provide detailed explanations for each answer as a reference to enhance the interpretability of answers, which is absent in previous works. As shown in Fig. 2, PathMMU distinguishes itself in the pathology domain through its professionalism and superior quality.", + "bbox": [ + 212, + 614, + 789, + 737 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "3 The Proposed PathMMU Benchmark", + "text_level": 1, + "bbox": [ + 215, + 756, + 617, + 773 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "3.1 Design Principle", + "text_level": 1, + "bbox": [ + 215, + 786, + 401, + 801 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "PathMMU is designed to provide the community with a specialized dataset for evaluating pathology LMMs. We adhere to five principal guidelines in its devel-", + "bbox": [ + 212, + 809, + 785, + 840 + ], + "page_idx": 4 + }, + { + "type": "header", + "text": "PathMMU", + "bbox": [ + 656, + 114, + 730, + 126 + ], + "page_idx": 4 + }, + { + "type": "page_number", + "text": "5", + "bbox": [ + 774, + 116, + 784, + 126 + ], + "page_idx": 4 + }, + { + "type": "image", + "img_path": "images/61f327ce61ac000781b5bac66790e28ac7c27c92b9a49089e792b123e8757cbd.jpg", + "image_caption": [ + "Fig. 3: An illustrative overview of the three main processes in PathMMU Q&A generation: data collection and preprocessing, detailed pathology image description generation, and question generation with LLMs filtering and expert validation." + ], + "image_footnote": [], + "bbox": [ + 218, + 143, + 403, + 325 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/cd66e1d8d36649ca58e541ac02307491fba018b121bd091fdd88fa46466fa080.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 411, + 143, + 578, + 325 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/54244a918b274a764ccc3b2174f41c816e068bec2f476862e438d7de41d3fd3e.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 584, + 143, + 782, + 325 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "opment: (1) Comprehensive and Specialized Data: Our benchmark is gathered from a diverse range of authoritative materials, including respected scientific publications from PubMed, pathology atlas from textbooks and guidelines, educational videos on YouTube, pathologist-shared images with explanations on Twitter, and widely recognized pathology classification datasets. Additionally, we involve seven human experts to manually review the dataset to ensure it meets professional standards. (2) Effective and Valuable Questions: The questions in PathMMU are carefully formulated to necessitate answers based on detailed observation of pathology images, rather than being solvable through mere textual interpretation or guesswork. We ensure that these questions are answerable, effective, and aligned with the standards of professional pathology examinations. (3) Large-Scale: We present the largest pathology dataset currently available. This scale enables researchers to thoroughly explore the full potential of LMMs in the field of pathology. (4) High-Quality Images: We prioritize clarity in our images to ensure that every detail is discernible, as evidenced by our dataset's average resolution of approximately $900 \\times 700$ pixels. (5) Explainable Answer Choices: In line with the ongoing interest in understanding the decision-making logic of large models, our dataset includes reference explanations for each answer choice. This feature enhances the interpretability of model responses, contributing to the broader research on model explainability.", + "bbox": [ + 212, + 405, + 787, + 709 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "3.2 The Construction of PathMMU", + "text_level": 1, + "bbox": [ + 215, + 727, + 524, + 742 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "To guarantee the quality of our benchmark, we meticulously develop a three-step data processing and generation protocol. The overall data collection and generation process is illustrated in Fig. 3.", + "bbox": [ + 212, + 750, + 787, + 795 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Step 1: Data Collection and Preprocessing. Our data collection draws from a wide variety of sources, which we integrate as subsets into the PathMMU framework. We name these subsets and detail their collection workflows as follows.", + "bbox": [ + 212, + 794, + 787, + 839 + ], + "page_idx": 5 + }, + { + "type": "page_number", + "text": "6", + "bbox": [ + 217, + 114, + 228, + 126 + ], + "page_idx": 5 + }, + { + "type": "header", + "text": "Y. Sun et al.", + "bbox": [ + 271, + 114, + 357, + 127 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "PubMed: This subset consists of scientific pathology image-text pairs sourced from scientific documents. We gather these pairs from the open-access section of PubMed Central. Given the significant presence of non-pathological data within this resource, we meticulously annotate 20,000 samples as either pathological or non-pathological, and subsequently train a ConvNeXt [32] model to identify pathology data within the remaining dataset. EduContent: This subset, derived from educational YouTube videos, is sourced from Quilt1M [20]. Since YouTube videos often include a variety of visuals unrelated to pathology, such as computer desktops and unrelated imagery, the initial phase involves the annotation of pathological regions in 3,000 YouTube video images. This critical step is followed by the training of a YOLOv6-based detector [28] to automate the identification of pathological regions within remaining YouTube content. Atlas: This subset is a compilation of authoritative pathology textbooks and guidelines, we transform them from PDFs to HTML to facilitate the extraction of image-caption pairs. SocialPath: This subset aggregates image-text pairs from Twitter posts by pathology experts, using Twitter URLs provided by OpenPath [19]. PathCLS: Encompassing widely recognized pathology classification datasets, including PatchCamelyon [47], CRC-100K [21], SICAPv2 [40], BACH [2], Osteo [3], SkinCancer [22], MHIST [50], WSSSLUAD [17], LC-Lung and LC-Colon [7]. We reformulate these datasets into a Q&A format and randomly sample from these datasets. We directly make the gathered samples from PathCLS a part of the PathMMU dataset without undergoing the following two steps Q&A generation process. Finally, we manually review and filter irrelevant or unclear images to ensure the quality of collected data. As a result, approximately 30,000 high-quality image-text pairs are obtained, forming the foundation of the PathMMU.", + "bbox": [ + 217, + 145, + 787, + 523 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Step 2: Detailed Description Generation and Question-answer Pairs Generation. When dealing with data sources from platforms like Twitter and YouTube, it's common to find a weak correlation between an image and its accompanying caption. For instance, in the SocialPath dataset, captions might only describe the aesthetic appeal of the image, which is irrelevant to pathological findings. Similarly, in other data sources, captions tend to describe only a fraction of the features present in the image. To tackle this issue, we prompt GPT-4V to focus on describing the morphology of cells and tissues. It is crucial to note that merely prompting GPT-4V to output image descriptions might lead to incoherent content. Therefore, we provide GPT-4V with original captions for reference, significantly reducing the incidence of such occurrences and thus ensuring the generation of more precise and relevant descriptions.", + "bbox": [ + 217, + 527, + 787, + 710 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Step 3: Question Generation and Expert Validation. Following the generation of image descriptions, GPT-4V is tasked with creating three questions per image, each accompanied by multiple-choice options, the correct answer, and detailed explanations for the answer. The creation of these explanations serves a dual purpose: it not only offers valuable insights into the model reasoning process, but also facilitates the subsequent manual review phase.", + "bbox": [ + 217, + 714, + 787, + 806 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "It's crucial to note that despite the careful design of question generation prompts for GPT-4V, the text-only GPT-4 is capable of correctly answering", + "bbox": [ + 215, + 809, + 785, + 839 + ], + "page_idx": 6 + }, + { + "type": "header", + "text": "PathMMU", + "bbox": [ + 656, + 114, + 730, + 126 + ], + "page_idx": 6 + }, + { + "type": "page_number", + "text": "7", + "bbox": [ + 774, + 114, + 785, + 126 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "over $60\\%$ of these questions through educated guesses, bypassing the need for visual cues. GPT-4 explains that its deductions are drawn from patterns such as one option being more common in typical pathological scenarios or exhibiting noticeable differences from other options. We delve deeper into this phenomenon in Sec. 4.3 to offer a more thorough experimental analysis. To ensure that LMMs do not rely on educated guesses, thereby obscuring their true multimodal capabilities, we employ GPT-3.5 Turbo, GPT-4 Turbo, Gemini Pro, and ERNIE-Bot-4 to perform educated guesses. Questions correctly guessed by at least three of these models are subsequently excluded.", + "bbox": [ + 212, + 146, + 787, + 282 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "We partitioned the remaining dataset into training, validation, and test sets. The training set comprises 16,344 images and 23,041 Q&As, while the validation and test sets contain roughly 12,000 Q&As and over 8,000 images, respectively. For the validation and test sets, we invite seven professional pathologists to conduct a thorough manual review. These pathologists initially attempt to answer the questions independently and then assess them based on the following criteria: (1) Whether the question can be answered without an accompanying image. (2) Whether the answer can be inferred from the provided question and image. (3) Whether the supplied answer is incorrect, if there is no correct answer, or if multiple correct answers are possible. (4) Whether the generated question appears unusual or atypical compared to standard questions in pathology examinations. Questions failing to meet these standards are deemed invalid and consequently removed from the PathMMU dataset.", + "bbox": [ + 212, + 282, + 787, + 477 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Ultimately, after thorough expert review, we obtain 710 Q&As accompanied by 510 images for the validation set, and 9,677 Q&As with 7,213 images for the test set. Additionally, to establish a standard for expert performance, we extract a smaller subset from the test set, named 'test-tiny', which includes 1,156 Q&As. We invite two groups of professional pathologists to participate in this subset as an examination, and we average their performance to set a benchmark for expert performance in the PathMMU, serving as a comparison reference for LMMs.", + "bbox": [ + 212, + 479, + 787, + 585 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "4 Experiments", + "text_level": 1, + "bbox": [ + 215, + 604, + 375, + 622 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "In this section, we conduct extensive experiments on PathMMU. Initially, we assess the zero-shot performance of cutting-edge LMMs. To verify whether LMMs effectively utilize the visual information from pathology images, we also evaluate the performance of text-only LLMs on PathMMU as a reference. Additionally, we apply common corruptions to the images in the test-tiny set to test the robustness of the LMMs against image corruptions. Furthermore, we select representative LMMs and evaluate their performance to explore their transfer-learning capabilities. Finally, we conduct experimental analysis on the issue of LLMs being able to guess answers identified during the data construction process.", + "bbox": [ + 212, + 632, + 787, + 770 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "4.1 Zero-shot Evaluation of LMMs and LLMs", + "text_level": 1, + "bbox": [ + 215, + 787, + 606, + 801 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "In this study, we evaluate the zero-shot capabilities of the latest and most advanced LMMs on the PathMMU. Specifically, we use the PathMMU validation", + "bbox": [ + 212, + 809, + 785, + 839 + ], + "page_idx": 7 + }, + { + "type": "page_number", + "text": "8", + "bbox": [ + 217, + 114, + 228, + 126 + ], + "page_idx": 7 + }, + { + "type": "header", + "text": "Y. Sun et al.", + "bbox": [ + 271, + 114, + 357, + 126 + ], + "page_idx": 7 + }, + { + "type": "table", + "img_path": "images/f507d52c43380b2c009f0fd5ee3afbb06a3add12125d26456fd408bee9b9d79e.jpg", + "table_caption": [ + "Table 1: Overall results of models on the PathMMU validation and test set. Besides reporting the performance of LMMs, we add text-only LLM baselines that purely accept text as inputs. The best-performing LMM in each subset is in- bold, and the top-performing LLM is underlined." + ], + "table_footnote": [], + "table_body": "
Validation OverallTestOverallPubMedSocialPathEduContentAtlasPathCLS
- (710)Tiny (1156)ALL (9677)Tiny (281)ALL (3068)Tiny (235)All (1855)Tiny (255)All (1938)Tiny (208)ALL (1007)Tiny (177)ALL (1809)
Random Choice24.622.123.722.125.125.526.525.526.019.723.015.316.3
Frequent Choice27.527.725.528.826.127.726.729.826.528.427.522.021.0
Expert performance-71.8-72.9-71.5-69.0-68.3-78.9-
Large Multimodal Models (LMMs): Text + Image as Input
OpenFlamingo2-9B [1]26.224.124.224.925.525.525.319.624.428.825.522.020.4
Kosmos2 [36]26.926.124.927.825.827.224.325.125.923.124.227.122.8
LLaMA-Adapter2-7B [14]26.226.626.426.028.327.726.527.128.030.827.420.920.5
MiniGPT4-Vicuna-13B [60]27.225.527.728.830.124.327.225.529.327.928.619.221.8
Otter [25]26.129.328.134.930.226.028.430.630.029.328.223.221.3
Fuyu-8B [5]27.530.129.235.930.528.930.730.629.827.930.624.323.9
CogVLM [49]29.930.629.732.032.131.130.430.629.935.133.722.622.4
Qwen-VL-7B [4]29.432.131.534.932.933.635.834.133.635.633.818.621.7
BLIP-2 FLAN-T5-XL [29]32.832.631.835.934.634.933.633.732.834.637.420.320.8
BLIP-2 FLAN-T5-XXL [29]33.433.333.537.037.435.734.630.234.539.440.719.820.6
InstructBLIP FLAN-T5-XL [11]33.134.931.837.033.235.333.936.533.337.036.726.623.1
InstructBLIP FLAN-T5-XXL [11]32.734.333.939.137.233.634.334.536.038.539.322.622.7
LLaVA-1.5-7B [30]36.634.935.441.639.937.938.132.536.535.639.223.221.9
LLaVA-1.5-13B [30]37.938.837.644.541.040.440.434.139.447.144.324.923.5
Qwen-VL-PLUS [4]38.039.334.343.537.741.336.039.636.044.737.123.223.3
Qwen-VL-MAX [4]43.649.245.953.050.953.649.352.247.951.449.830.529.6
Gemini Pro Vision [45]41.942.842.743.844.942.442.043.543.749.549.432.834.7
GPT-4V-1106 [35]49.353.949.859.453.558.753.960.453.648.152.836.233.8
Large Language Models (LLMs): Only Text as Input
ERNIE-Bot 4.0 [41]31.834.330.937.031.233.632.940.034.536.137.420.920.6
Gemini Pro [45]31.031.631.031.333.531.931.433.332.738.037.721.520.9
Vicuna-v1.5-13B [10]32.031.231.635.233.830.634.331.033.534.635.722.020.7
GPT-3.5 Turbo [33]30.429.428.432.731.231.530.131.030.126.926.822.020.8
GPT-4 Turbo [34]36.541.838.148.842.443.442.347.140.641.343.222.321.1
", + "bbox": [ + 218, + 212, + 784, + 545 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "set to conduct prompt engineering for these models, and then test their performance on the test set. The tested models include 14 open-source LMMs: OpenFlamingo [1], Kosmos2 [36], LLaMA-Adapter2 [14], MiniGPT-4 [60], Otter [25], Fuyu [5], CogVLM [49], Qwen-VL [4], BLIP-2 [29], InstructBLIP [11], and LLaVA-1.5 [30], as well as 4 closed-source LMMs: Qwen-VL-PLUS, Qwen-VL-MAX [4], Gemini Pro Vision [45], and GPT-4V [35]. For BLIP-2, Instruct-BLIP, and LLaVA-1.5, we deploy various sizes of these models to examine the applicability of scaling rules in LMMs. Moreover, we extend our evaluation to purely text-based LLMs, including ERNIE-Bot 4.0 [41], Gemini Pro [45], Vicuna-v1.5-13B [10], GPT-3.5 Turbo [33], and GPT-4 Turbo [34]. This assessment aims to quantify their ability to infer correct answers through educated guesses without visual inputs, reflecting the extent to which LMMs integrate and leverage visual information compared to their text-only counterparts.", + "bbox": [ + 212, + 580, + 787, + 777 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "Results and Discussion. The advanced LMMs struggle with the PathMMU dataset. As shown in Tab. 1, among the 18 models tested, 15 show an accuracy below $40\\%$ . Notably, the highest-performing open-source model, LLaVa-1.5-13B, and the top-performing closed-source model, GPT-4V, achieve", + "bbox": [ + 212, + 779, + 787, + 840 + ], + "page_idx": 8 + }, + { + "type": "header", + "text": "PathMMU", + "bbox": [ + 656, + 114, + 730, + 126 + ], + "page_idx": 8 + }, + { + "type": "page_number", + "text": "9", + "bbox": [ + 774, + 114, + 785, + 126 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "only $37.6\\%$ and $49.8\\%$ accuracy, respectively. Those are significantly lower than human expert performance, which stands at $71.8\\%$ , underscoring a substantial gap between current LMMs and professional pathologist standards. This gap indicates that the practical application of LMMs in real-world pathological scenarios remains significantly constrained.", + "bbox": [ + 212, + 146, + 782, + 220 + ], + "page_idx": 9 + }, + { + "type": "text", + "text": "Closed-source LLMs achieve performance that is on par with or even surpasses that of the most advanced open-source LMMs. When considering text-only models, we prompt them to make educated guesses without providing the image. We observe that GPT-4 Turbo, Vicuna-v1.5-13B, GeminiPro, and ERNIE-Bot 4.0 outperform nearly half of the open-source LMMs. Notably, GPT-4 Turbo leads with a significant margin, achieving an accuracy rate of $38.1\\%$ , exceeding the second-best LLM by $6.5\\%$ , and even outperforming the best-performing open-source LMM, LLaVA-1.5-13B. The $38.1\\%$ accuracy of GPT-4 Turbo, on the one hand, demonstrates that under our strict filtering criteria, even the best LLM, when provided with text alone, makes incorrect answers for the majority of questions. This underscores the effectiveness of our dataset in testing the capabilities of LMMs in pathological image analysis. On the other hand, it reaffirms the analysis in Sec. 3.2 that GPT-4 possesses strong logical reasoning abilities, enabling it to correctly guess a portion of the questions, significantly outperforming random selection. However, this raises critical concerns in pathology diagnosis: the reliability of model outputs in clinical scenarios remains questionable, especially when they rely on educated guesses, particularly probabilistic ones, rather than comprehensive analysis. This emphasizes the urgent necessity to develop interpretable AI models and advocate for their cautious application in pathology diagnostics.", + "bbox": [ + 212, + 222, + 784, + 523 + ], + "page_idx": 9 + }, + { + "type": "text", + "text": "The larger LMMs exhibit better performance. Specifically, LLaVa-1.5-13B exceeds LLaVa-1.5-7B model by $2.2\\%$ in overall performance, while Instruct-BLIP FLAN-T5-XXL and BLIP-2 FLAN-T5 XXL surpass their smaller versions by $2.1\\%$ and $1.7\\%$ , respectively. This suggests that larger models typically possess stronger multimodal capabilities in the field of pathology.", + "bbox": [ + 212, + 523, + 784, + 599 + ], + "page_idx": 9 + }, + { + "type": "text", + "text": "4.2 Robustness of LMMs to Image Corruption", + "text_level": 1, + "bbox": [ + 214, + 619, + 612, + 636 + ], + "page_idx": 9 + }, + { + "type": "text", + "text": "In practical pathology, the interpretation of models significantly influences the subsequent medical decisions and treatment strategies. Therefore, models with strong robustness are crucial for clinical applications. The quality of pathological slides can be affected by various factors during staining, scanning, and storage, including JPEG compression, pixelation, blur (e.g., bubble blur, defocus blur, and motion blur), and color variations (e.g., brightness, saturation, and hue).", + "bbox": [ + 212, + 643, + 782, + 733 + ], + "page_idx": 9 + }, + { + "type": "text", + "text": "Drawing inspiration from the studies on the robustness of encoder-based models to pathology image corruptions [43, 56, 58], we incorporate these aforementioned types of corruptions into our analysis. To be specific, we simulate five levels of each corruption type on the PathMMU to explore the robustness of LMMs against these corruptions, as depicted in the left half of Fig. 4.", + "bbox": [ + 212, + 734, + 782, + 808 + ], + "page_idx": 9 + }, + { + "type": "text", + "text": "Results and Discussion. LMMs demonstrate notable robustness to various types and levels of image corruptions, yet their true robustness", + "bbox": [ + 214, + 809, + 782, + 839 + ], + "page_idx": 9 + }, + { + "type": "page_number", + "text": "10", + "bbox": [ + 217, + 114, + 235, + 126 + ], + "page_idx": 9 + }, + { + "type": "header", + "text": "Y. Sun et al.", + "bbox": [ + 271, + 114, + 357, + 126 + ], + "page_idx": 9 + }, + { + "type": "image", + "img_path": "images/59332a6b7df98c041fbee5fd8a42a4b8d1bd2ee1b6d02a92a0ee7d0f6c399e7a.jpg", + "image_caption": [ + "Fig. 4: Left: Illustration of corrupted pathology images. Right: LMM's performance across various levels of color-related (brightness, hue, saturation) and image quality-related (pixelation, JPEG compression, bubble blur, motion blur, defocus blur) corruptions on the PathMMU test-tiny set, with level 0 representing the uncorrupted images." + ], + "image_footnote": [], + "bbox": [ + 220, + 143, + 454, + 281 + ], + "page_idx": 10 + }, + { + "type": "image", + "img_path": "images/f87f47b50d448db9d0665be2e553f9cba32e39ef9892b286011374a5501a8260.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 457, + 146, + 617, + 280 + ], + "page_idx": 10 + }, + { + "type": "image", + "img_path": "images/912bfc972cbd9be60ab82e937ed09f2153ffddd3ca5b3d6db9766466beaf853a.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 619, + 146, + 779, + 277 + ], + "page_idx": 10 + }, + { + "type": "table", + "img_path": "images/0a6411cc0fdc7ba3b488e32446441d7c9d648388db24832c29726e98527ef8e9.jpg", + "table_caption": [ + "Table 2: Evaluation of the model's robustness on the PathMMU test-tiny set, where green indicates performance improvement and red signifies performance degradation compared to the model's performance with uncorrupted images." + ], + "table_footnote": [], + "table_body": "
BrightnessBubbleDefocusHueJPEGMotionPixelateSaturateOverall
Qwen-VL-7B32.7↑0.632.3↑0.232.1↓0.032.4↓0.332.3↑0.233.8↑1.733.1↑1.033.0↑0.932.7↑0.6
BLIP-2 FLAN-T5-XL32.9↑0.332.4↓0.232.8↑0.232.9↑0.332.4↓0.232.8↑0.232.6↑0.032.7↑0.132.7↓0.1
BLIP-2 FLAN-T5-XXL33.1↓0.233.2↓0.133.5↑0.233.7↑0.433.5↑0.233.1↓0.233.1↓0.233.6↑0.333.3↓0.0
InstructBLIP FLAN-T5-XL34.0↓0.934.1↓0.833.9↓1.034.1↓0.833.7↓1.234.0↓0.934.4↓0.533.8↓1.134.0↓0.9
InstructBLIP FLAN-T5-XXL34.0↓0.333.8↓0.534.2↓0.134.1↓0.234.3↓0.234.3↑0.034.1↓0.233.8↓0.534.1↓0.2
LLaVA-1.5-7B34.8↓0.135.3↑0.434.1↓0.834.3↓0.634.0↓0.934.3↓0.634.6↓0.334.7↓0.234.5↓0.4
LLaVA-1.5-13B38.9↑0.138.3↓0.538.3↓0.538.4↓0.438.0↓0.838.6↓0.238.7↓0.139.1↑0.338.5↓0.3
", + "bbox": [ + 218, + 419, + 785, + 506 + ], + "page_idx": 10 + }, + { + "type": "table", + "img_path": "images/a46eefe9d1117c46165e711f23bc49f327b756303fa2fc4c090b08e50e88f03e.jpg", + "table_caption": [ + "Table 3: Results of LMMs on the PathMMU test set with the original images substituted by random Gaussian noise images." + ], + "table_footnote": [], + "table_body": "
OverallPubMedSocialPathEduContentAtlasPathCLS
TinyALLTinyALLTinyAllTinyAllTinyALLTinyALL
Qwen-VL-7B28.8↓3.328.3↓3.239.5↑4.630.4↓2.524.7↓8.930.5↓5.327.8↓6.328.9↓4.727.4↓8.230.1↓3.720.3↑1.720.8↓0.9
BLIP-FLAN-T5-XL30.9↓1.730.8↓1.030.6↓5.333.1↓1.534.5↑0.433.4↓0.232.2↑1.531.9↓0.934.1↓0.535.1↓2.320.9↑0.620.8↓0.0
BLIP-FLAN-T5-XXL32.4↓0.931.6↓1.934.5↓2.534.6↓2.834.0↑1.732.5↓2.130.6↑0.432.8↑1.738.5↓0.938.1↓2.622.2↑2.420.8↑0.2
InstructBLIP FLAN-T5-XL31.1↓3.829.7↓2.132.4↓4.630.8↓2.432.8↓2.532.0↓1.933.7↓2.832.4↓0.932.7↓4.330.4↓6.321.5↓5.122.4↓0.7
InstructBLIP FLAN-T5-XXL30.4↓3.930.8↓3.132.0↓7.133.5↓3.728.9↑4.731.9↓2.432.2↑2.331.7↓4.335.1↑3.435.2↓4.121.5↓1.121.7↓1.0
LLAVA-1.5-7B30.2↓4.730.9↓4.537.0↓4.633.9↓6.031.5↓6.432.5↓5.628.2↓4.332.7↓3.829.8↓5.833.9↓5.320.9↓2.320.8↓1.1
LLAVA-1.5-13B32.7↓6.133.2↓4.438.4↓6.136.0↓5.035.3↓5.136.3↓4.129.0↓5.134.1↓5.335.6↓11.538.8↓5.522.0↓2.921.0↓2.5
", + "bbox": [ + 217, + 561, + 785, + 642 + ], + "page_idx": 10 + }, + { + "type": "text", + "text": "is questionable, as shown in Tab. 2 and the right half of Fig. 4. Notably, the Qwen-VL-7B even shows a $0.6\\%$ overall performance increase under corruption compared to its baseline. It is more plausible to hypothesize that these corruptions mainly alter minute details of the pathology image that are difficult for LMMs to discern (such as chromatin morphology of the nucleus or vacuolization of the cytoplasm). This limitation stems from their pre-training in general domains, where they learn to recognize larger and more prominent features (e.g., humans, houses, cars, etc.) rather than the nuanced details that are vital in pathology. Furthermore, they may even resort to exploiting spurious correlations as shortcuts to answer questions (e.g., deducing answers from textual patterns instead of focusing on essential image details.)", + "bbox": [ + 212, + 672, + 787, + 840 + ], + "page_idx": 10 + }, + { + "type": "header", + "text": "PathMMU", + "bbox": [ + 656, + 114, + 730, + 126 + ], + "page_idx": 10 + }, + { + "type": "page_number", + "text": "11", + "bbox": [ + 767, + 114, + 782, + 126 + ], + "page_idx": 10 + }, + { + "type": "text", + "text": "To substantiate our argument, we employ an extreme test of corruption by replacing the images with Gaussian noise-generated random images. As shown in Tab. 3, we discover that even when the images contain no relevant information, LMMs can still achieve results significantly better than random choice. The drop in performance when using these random images ranged from only $1.0\\%$ to $4.5\\%$ . This finding suggests that LMMs might rely on shortcuts to accomplish the multimodal task, such as depending exclusively on textual information to make predictions. Additionally, it is intriguing that the performance drop across different sizes of the same model is remarkably similar. In other words, the benefit that images provide to models of various model sizes appears to be consistent, suggesting that the improvement in performance for different-sized models may primarily stem from their LLM component, rather than the vision aspect.", + "bbox": [ + 212, + 146, + 787, + 328 + ], + "page_idx": 11 + }, + { + "type": "text", + "text": "4.3 Analysis of LLMs' Ability to Guess Answers", + "text_level": 1, + "bbox": [ + 214, + 348, + 627, + 364 + ], + "page_idx": 11 + }, + { + "type": "text", + "text": "In this section, we delve deeper into the phenomenon of LLM's ability to make educated guesses with more comprehensive experiments and analysis.", + "bbox": [ + 212, + 371, + 785, + 401 + ], + "page_idx": 11 + }, + { + "type": "text", + "text": "To more clearly demonstrate the guessing abilities of LLMs, we randomly select 100 samples that are filtered during the Q&A generation process and can be correctly guessed by multiple LLMs. We invite pathology experts to answer these questions with reference images. As shown in Fig. 5, despite having access to images, their performance is significantly lower than that of the closed-source LLMs, which achieve about $90\\%$ accuracy, and only slightly better than the smaller open-source Vicuna-v1.5-13B. This indicates that LLMs may outperform humans in guessing correct answers by identifying shortcuts within the questions.", + "bbox": [ + 212, + 402, + 785, + 521 + ], + "page_idx": 11 + }, + { + "type": "text", + "text": "We hypothesize several key reasons for the guessing behavior exhibited by LLMs: (1) Frequency-based guessing: LLMs may guess based on the prevalence of certain options in real-world pathological instances. (2) Based on the options that present a pattern of one supporting and three contradicting, or vice versa: For example, when a question identifies a tumour diagnosis, one option may match the pathological traits of a lesion, while the others suggest non-lesional characteristics. (3) When only one option aligns with the question's subject: For example, if a question asks which feature in the image supports the diagnosis of a Low-grade Squamous Intraepithelial Lesion (LSIL) cell, with only one option describes LSIL's pathological characteristics.", + "bbox": [ + 212, + 523, + 787, + 672 + ], + "page_idx": 11 + }, + { + "type": "text", + "text": "To support our hypothesis, we expand the sample size to 1,600, with 400 from each source (except PathCLS), to empirically analyze how LLMs guess answers. Specifically, we swap the questions among these samples while keeping the options unchanged, creating samples where the questions and options are completely mismatched. As shown in the right half of Fig. 5, we observe that LLMs still manage to guess approximately $50\\%$ cases correctly, significantly higher than random choice accuracy. This finding suggests that the models tend to select the most common or prominent option as the answer, supporting our hypotheses (1) and (2). To further explore hypothesis (3), we design experiments using BERT-large [12] and BiomedBERT-large [15], the latter being specifically pre-trained on biomedical data. Given that the BERT series incorporates Next-Sentence", + "bbox": [ + 212, + 674, + 787, + 839 + ], + "page_idx": 11 + }, + { + "type": "page_number", + "text": "12", + "bbox": [ + 217, + 114, + 235, + 126 + ], + "page_idx": 11 + }, + { + "type": "header", + "text": "Y. Sun et al.", + "bbox": [ + 271, + 114, + 357, + 126 + ], + "page_idx": 11 + }, + { + "type": "image", + "img_path": "images/75c52ac8bc697d5ee19715be4da0962f7a25300525ce2189db23a99c036a280c.jpg", + "image_caption": [ + "Fig. 5: Left: The performance comparison between different LLMs and human experts on 100 filtered samples where the answer can be guessed through text-only. Right: Expand the sample size to 1600 to validate the source of LLMs' ability to guess answers, which includes: (1) randomly replacing the original questions with others from the dataset while keeping the options unchanged, and (2) utilizing the BERT series for answer selection, specifically through its Next Sentence Prediction (NSP), to assess whether an option is the sequential sentence following a question." + ], + "image_footnote": [], + "bbox": [ + 223, + 142, + 475, + 263 + ], + "page_idx": 12 + }, + { + "type": "table", + "img_path": "images/b0b589b8581cd5b594bbdcbf9fb0c7137863c945a09435698fec3daf2fe15e31.jpg", + "table_caption": [], + "table_footnote": [], + "table_body": "
PubMed SocialPath EduContentAtlas
Random Choice25.824.325.525.825.3
GPT-4 Turbo93.091.886.883.888.8
+ Random Swap Question44.345.550.051.547.8
GPT-3.5 Turbo84.386.579.585.383.9
+ Random Swap Question44.044.553.851.348.4
Gemini Pro85.590.176.571.584.1
+ Random Swap Question43.844.850.351.347.5
ERNIE-Bot 4.094.592.389.382.089.5
+ Random Swap Question47.046.556.551.050.3
BERT-large w/ NSP24.536.037.534.333.1
BiomedBERT-large w/ NSP44.344.842.337.542.3
", + "bbox": [ + 483, + 143, + 787, + 260 + ], + "page_idx": 12 + }, + { + "type": "text", + "text": "Prediction (NSP) in their pretraining, which inherently relies on the similarity between two sentences, we apply this to predict the relationship between questions and options. We select the option with the highest probability of being the correct next sentence to the question. Our findings reveal that both BERT and BiomedBERT substantially surpass random guessing, indicating that direct matching of questions and options is a viable method for models to guess the correct answers, thereby supporting hypothesis (3). Moreover, BiomedBERT-large demonstrates a notable performance enhancement compared to BERT-large. This suggests that pretraining on biomedical data equips the model with a broader understanding of pathological knowledge.", + "bbox": [ + 212, + 398, + 787, + 551 + ], + "page_idx": 12 + }, + { + "type": "text", + "text": "4.4 Fine-tuning Results", + "text_level": 1, + "bbox": [ + 215, + 573, + 426, + 589 + ], + "page_idx": 12 + }, + { + "type": "text", + "text": "To explore the adaptability of LMMs to the pathology domain, we select representative LMMs, InstructBLIP FLAN-T5-XL and InstructBLIP FLAN-T5-XXL for fine-tuning experiments. Our experiments consist of two parts: (1) training the LMMs to directly generate answers, and (2) fine-tuning the LMMs to generate a reasoning process before delivering the final answer.", + "bbox": [ + 212, + 598, + 785, + 672 + ], + "page_idx": 12 + }, + { + "type": "text", + "text": "Results and Discussion. All models exhibit significant improvements on the PathMMU test set after fine-tuning on its training set, as detailed in Tab. 4. Notably, during the fine-tuning for direct answer generation, InstructBLIP FLAN-T5-XL and InstructBLIP FLAN-T5-XXL achieve significant improvements of $21.5\\%$ and $21.3\\%$ , respectively. This substantial improvement enables them to outperform the current leading model, GPT-4V. While still lagging behind expert performance, these results demonstrate a trend toward approaching expert-level proficiency. This also reflects the effectiveness of PathMMU in enhancing the LMMs' abilities for pathology image analysis.", + "bbox": [ + 212, + 672, + 787, + 809 + ], + "page_idx": 12 + }, + { + "type": "text", + "text": "Unexpectedly, generating explanations before answers during fine-tuning does not yield improvements. Instead, we observe slight performance decreases of $0.3\\%$", + "bbox": [ + 212, + 809, + 785, + 839 + ], + "page_idx": 12 + }, + { + "type": "header", + "text": "PathMMU", + "bbox": [ + 656, + 114, + 730, + 126 + ], + "page_idx": 12 + }, + { + "type": "page_number", + "text": "13", + "bbox": [ + 767, + 114, + 785, + 126 + ], + "page_idx": 12 + }, + { + "type": "table", + "img_path": "images/8acc7e401da6292596b496a2de53603855cd1ff24b760d47ff73c9e3f84437c2.jpg", + "table_caption": [ + "Table 4: Results of LMMs on the PathMMU test set after the fine-tuning. The 'w/ A' and 'w/ A&E' denote the model is fine-tuned to output the answer directly or to output the answer with an explanation for its answer." + ], + "table_footnote": [], + "table_body": "
OverallPubMedSocialPathEduContentAtlasPathCLS
TinyALLTinyALLTinyALLTinyAllTinyALLTinyALL
InstructBLIP FLAN-T5-XL34.931.837.033.235.333.936.533.337.036.726.623.1
+ fine-tune w/A55.7 ± 20.853.3 ± 21.553.7 ± 16.750.4 ± 17.252.8 ± 17.552.3 ± 18.456.5 ± 20.052.0 ± 18.756.2 ± 19.253.1 ± 16.461.0 ± 34.460.8 ± 37.7
+ fine-tune w/A&E54.9 ± 20.053.0 ± 21.255.5 ± 18.551.4 ± 18.251.9 ± 16.649.7 ± 15.855.3 ± 18.851.5 ± 18.253.4 ± 16.451.6 ± 14.959.3 ± 32.761.4 ± 38.3
InstructBLIP FLAN-T5-XXL34.333.939.137.233.634.334.536.038.539.322.622.7
+ fine-tune w/A56.8 ± 22.555.2 ± 21.355.2 ± 16.151.5 ± 14.359.6 ± 26.055.2 ± 20.958.4 ± 23.954.1 ± 18.150.5 ± 12.053.7 ± 14.461.0 ± 38.463.7 ± 41.0
+ fine-tune w/A&E51.0 ± 16.752.9 ± 19.048.8 ± 9.750.8 ± 13.655.3 ± 21.751.2 ± 16.952.2 ± 17.751.9 ± 15.944.2 ± 5.750.1 ± 10.854.8 ± 32.260.5 ± 37.8
", + "bbox": [ + 217, + 198, + 785, + 267 + ], + "page_idx": 13 + }, + { + "type": "text", + "text": "and $2.3\\%$ compared to direct answer generation, respectively. We speculate that generating explanations is relatively more challenging, and incorporating it with answer generation might impede the models' capacity to generate correct answers. This finding raises an important question: how can we effectively leverage the interpretability information within PathMMU to enhance models' training?", + "bbox": [ + 215, + 291, + 784, + 367 + ], + "page_idx": 13 + }, + { + "type": "text", + "text": "5 Conclusion", + "text_level": 1, + "bbox": [ + 217, + 388, + 356, + 402 + ], + "page_idx": 13 + }, + { + "type": "text", + "text": "In this study, we introduce PathMMU, the largest and highest-quality pathology benchmark to date, specifically crafted to evaluate the capabilities of LMMs in interpreting and reasoning with pathology images. The construction of PathMMU involves a meticulous data collection and curation process, supplemented by a strict manual review by seven professional pathologists to ensure its quality and professionalism. Moreover, we establish a human expert performance benchmark to quantify the gap between cutting-edge LMMs and human experts. Our experimental results reveal that advanced LMMs significantly lag behind on PathMMU, with these models demonstrating poor performance in observing details in pathology images and sometimes even neglecting visual information, highlighting a substantial gap in practical pathology application. However, LMMs demonstrate notable performance improvements after fine-tuning on PathMMU, even surpassing GPT-4V. While they do not achieve human expert-level performance, these LMMs show promising potential for analyzing pathology images.", + "bbox": [ + 215, + 416, + 785, + 628 + ], + "page_idx": 13 + }, + { + "type": "text", + "text": "Future Directions for Pathology LMMs: Our experience with PathMMU highlights key areas for development in pathology LMMs: (1) Current LMMs, which are primarily based on LLMs and fine-tuned in a lightweight manner, tend to over-rely on textual information while neglecting visual data. There is a significant need to explore training methodologies or model structures that better integrate visual and textual modalities. (2) There's a tendency for LMMs to take shortcuts, solving problems in a \"lazy\" manner. This necessitates the development of trustworthy models for real-world clinical applications. (3) Given that most current LMMs do not support multi-image inputs, PathMMU does not include a benchmark for processing multiple images. However, in practical scenarios, pathologists typically analyze samples at various magnifications and perspectives, underscoring the importance of developing LMMs capable of handling multi-image inputs. Finally, we believe PathMMU will catalyze significant advancements in the development of next-generation LMMs in pathology.", + "bbox": [ + 215, + 628, + 785, + 839 + ], + "page_idx": 13 + }, + { + "type": "page_number", + "text": "14", + "bbox": [ + 217, + 114, + 235, + 126 + ], + "page_idx": 13 + }, + { + "type": "header", + "text": "Y. Sun et al.", + "bbox": [ + 271, + 114, + 357, + 126 + ], + "page_idx": 13 + }, + { + "type": "text", + "text": "Acknowledgements", + "text_level": 1, + "bbox": [ + 217, + 143, + 401, + 162 + ], + "page_idx": 14 + }, + { + "type": "text", + "text": "This study was partially supported by the National Natural Science Foundation of China (Grant No.92270108), Zhejiang Provincial Natural Science Foundation of China (Grant No.XHD23F0201), the Research Center for Industries of the Future (RCIF) at Westlake University, and the Westlake Education Foundation.", + "bbox": [ + 212, + 172, + 787, + 233 + ], + "page_idx": 14 + }, + { + "type": "text", + "text": "References", + "text_level": 1, + "bbox": [ + 217, + 253, + 321, + 268 + ], + "page_idx": 14 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "1. Alayrac, J.B., Donahue, J., Luc, P., Miech, A., Barr, I., Hasson, Y., Lenc, K., Mensch, A., Millican, K., Reynolds, M., et al.: Flamingo: a visual language model for few-shot learning. In: NeurIPS. pp. 23716-23736 (2022)", + "2. Aresta, G., Araújo, T., Kwok, S., Chennamsetty, S.S., Safwan, M., Alex, V., Marami, B., Prastawa, M., Chan, M., Donovan, M., et al.: Bach: Grand challenge on breast cancer histology images. Medical image analysis 56, 122-139 (2019)", + "3. Arunachalam, H.B., Mishra, R., Daescu, O., Cederberg, K., Rakheja, D., Sengupta, A., Leonard, D., Hallac, R., Leavey, P.: Viable and necrotic tumor assessment from whole slide images of osteosarcoma using machine-learning and deep-learning models. PloS one 14(4), e0210706 (2019)", + "4. Bai, J., Bai, S., Yang, S., Wang, S., Tan, S., Wang, P., Lin, J., Zhou, C., Zhou, J.: Qwen-vl: A frontier large vision-language model with versatile abilities. arXiv preprint arXiv:2308.12966 (2023)", + "5. Bavishi, R., Elsen, E., Hawthorne, C., Nye, M., Odena, A., Somani, A., Taşırlar, S.: Introducing our multimodal models (2023), https://www.adept.ai/blog/fuyu-8b", + "6. Ben Abacha, A., Sarrouti, M., Demner-Fushman, D., Hasan, S.A., Müller, H.: Overview of the vqa-med task at imagecef 2021: Visual question answering and generation in the medical domain. In: Proceedings of the CLEF 2021 Conference and Labs of the Evaluation Forum-working notes (2021)", + "7. Borkowski, A.A., Bui, M.M., Thomas, L.B., Wilson, C.P., DeLand, L.A., Mastorides, S.M.: Lung and colon cancer histopathological image dataset (lc25000). arXiv preprint arXiv:1912.12142 (2019)", + "8. Brown, T., Mann, B., Ryder, N., Subbiah, M., Kaplan, J.D., Dhariwal, P., Neelakantan, A., Shyam, P., Sastry, G., Askell, A., et al.: Language models are few-shot learners. In: NeurIPS. pp. 1877-1901 (2020)", + "9. Cai, R., Song, Z., Guan, D., Chen, Z., Luo, X., Yi, C., Kot, A.: Benchlmm: Benchmarking cross-style visual capability of large multimodal models. arXiv preprint arXiv:2312.02896 (2023)", + "10. Chiang, W.L., Li, Z., Lin, Z., Sheng, Y., Wu, Z., Zhang, H., Zheng, L., Zhuang, S., Zhuang, Y., Gonzalez, J.E., Stoica, I., Xing, E.P.: Vicuna: An open-source chatbot impressing gpt-4 with $90\\%$ chatgpt quality (March 2023), https://lmsys.org/blog/2023-03-30-vicuna/", + "11. Dai, W., Li, J., Li, D., Tiong, A.M.H., Zhao, J., Wang, W., Li, B., Fung, P., Hoi, S.: Instructclip: Towards general-purpose vision-language models with instruction tuning. arXiv preprint arXiv:2305.06500 (2023)", + "12. Devlin, J., Chang, M.W., Lee, K., Toutanova, K.: Bert: Pre-training of deep bidirectional transformers for language understanding. In: NAACL. pp. 4171-4186 (2019)", + "13. Driess, D., Xia, F., Sajjadi, M.S., Lynch, C., Chowdhery, A., Ichter, B., Wahid, A., Tompson, J., Vuong, Q., Yu, T., et al.: Palm-e: An embodied multimodal language model. In: ICML. pp. 8469-8488 (2023)" + ], + "bbox": [ + 218, + 281, + 784, + 839 + ], + "page_idx": 14 + }, + { + "type": "header", + "text": "PathMMU", + "bbox": [ + 656, + 114, + 730, + 126 + ], + "page_idx": 14 + }, + { + "type": "page_number", + "text": "15", + "bbox": [ + 767, + 116, + 785, + 126 + ], + "page_idx": 14 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "14. Gao, P., Han, J., Zhang, R., Lin, Z., Geng, S., Zhou, A., Zhang, W., Lu, P., He, C., Yue, X., et al.: Llama-adapter v2: Parameter-efficient visual instruction model. arXiv preprint arXiv:2304.15010 (2023)", + "15. Gu, Y., Tinn, R., Cheng, H., Lucas, M., Usuyama, N., Liu, X., Naumann, T., Gao, J., Poon, H.: Domain-specific language model pretraining for biomedical natural language processing. ACM Transactions on Computing for Healthcare (HEALTH) 3(1), 1-23 (2021)", + "16. Guan, T., Liu, F., Wu, X., Xian, R., Li, Z., Liu, X., Wang, X., Chen, L., Huang, F., Yacoob, Y., et al.: Hallusionbench: an advanced diagnostic suite for entangled language hallucination and visual illusion in large vision-language models. In: CVPR. pp. 14375-14385 (2024)", + "17. Han, C., Pan, X., Yan, L., Lin, H., Li, B., Yao, S., Lv, S., Shi, Z., Mai, J., Lin, J., et al.: Wss54luad: Grand challenge on weakly-supervised tissue semantic segmentation for lung adenocarcinoma. arXiv preprint arXiv:2204.06455 (2022)", + "18. He, X., Zhang, Y., Mou, L., Xing, E., Xie, P.: Pathvqa: 30000+ questions for medical visual question answering. arXiv preprint arXiv:2003.10286 (2020)", + "19. Huang, Z., Bianchi, F., Yuksekgonul, M., Montine, T.J., Zou, J.: A visual-language foundation model for pathology image analysis using medical twitter. Nature medicine 29(9), 2307-2316 (2023)", + "20. Ikezogwo, W., Seyfioglu, S., Ghezloo, F., Geva, D., Sheikh Mohammed, F., Anand, P.K., Krishna, R., Shapiro, L.: Quilt-1m: One million image-text pairs for histopathology. In: NeurIPS. pp. 37995–38017 (2023)", + "21. Kather, J.N., Halama, N., Marx, A.: 100,000 histological images of human colorectal cancer and healthy tissue. Zenodo10 5281 (2018)", + "22. Kriegsmann, K., Lovers, F., Zgorzelski, C., Kriegsmann, J., Janssen, C., Meliss, R.R., Muley, T., Sack, U., Steinbuss, G., Kriegsmann, M.: Deep learning for the detection of anatomical tissue structures and neoplasms of the skin on scanned histopathological tissue sections. Frontiers in Oncology 12, 1022967 (2022)", + "23. Kumar, V., Abbas, A.K., Fausto, N., Aster, J.C.: Robbins and Cotran pathologic basis of disease, professional edition e-book. Elsevier health sciences (2014)", + "24. Lau, J.J., Gayen, S., Ben Abacha, A., Demner-Fushman, D.: A dataset of clinically generated visual questions and answers about radiology images. Scientific data 5(1), 1-10 (2018)", + "25. Li, B., Zhang, Y., Chen, L., Wang, J., Yang, J., Liu, Z.: Otter: A multi-modal model with in-context instruction tuning. arXiv preprint arXiv:2305.03726 (2023)", + "26. Li, B., Wang, R., Wang, G., Ge, Y., Ge, Y., Shan, Y.: benseed-bench: Benchmarking multimodal llms with generative comprehension. arXiv preprint arXiv:2307.16125 (2023)", + "27. Li, C., Wong, C., Zhang, S., Usuyama, N., Liu, H., Yang, J., Naumann, T., Poon, H., Gao, J.: Llava-med: Training a large language-and-vision assistant for biomedicine in one day. In: NeurIPS. pp. 28541-28564 (2023)", + "28. Li, C., Li, L., Jiang, H., Weng, K., Geng, Y., Li, L., Ke, Z., Li, Q., Cheng, M., Nie, W., et al.: Yolov6: A single-stage object detection framework for industrial applications. arXiv preprint arXiv:2209.02976 (2022)", + "29. Li, J., Li, D., Savarese, S., Hoi, S.: Blip-2: Bootstrapping language-image pretraining with frozen image encoders and large language models. In: ICML. pp. 19730-19742 (2023)", + "30. Liu, H., Li, C., Wu, Q., Lee, Y.J.: Visual instruction tuning. In: NeurIPS. pp. 34892-34916 (2023)" + ], + "bbox": [ + 215, + 146, + 785, + 839 + ], + "page_idx": 15 + }, + { + "type": "page_number", + "text": "16", + "bbox": [ + 217, + 114, + 235, + 126 + ], + "page_idx": 15 + }, + { + "type": "header", + "text": "Y. Sun et al.", + "bbox": [ + 271, + 114, + 357, + 126 + ], + "page_idx": 15 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "31. Liu, Y., Duan, H., Zhang, Y., Li, B., Zhang, S., Zhao, W., Yuan, Y., Wang, J., He, C., Liu, Z., et al.: Mmbench: Is your multi-modal model an all-around player? arXiv preprint arXiv:2307.06281 (2023)", + "32. Liu, Z., Mao, H., Wu, C.Y., Feichtenhofer, C., Darrell, T., Xie, S.: A convnet for the 2020s. In: CVPR. pp. 11976-11986 (2022)", + "33. OpenAI: Introducing chatgpt. https://openai.com/blog/chatgpt (2022)", + "34. OpenAI: Gpt-4 technical report (2023)", + "35. OpenAI: Gpt-4v(ision) system card. https://cdn.openai.com/papers/GPTV_System/Card.pdf (2023)", + "36. Peng, Z., Wang, W., Dong, L., Hao, Y., Huang, S., Ma, S., Wei, F.: Kosmos-2: Grounding multimodal large language models to the world. arXiv preprint arXiv:2306.14824 (2023)", + "37. Radford, A., Kim, J.W., Hallacy, C., Ramesh, A., Goh, G., Agarwal, S., Sastry, G., Askell, A., Mishkin, P., Clark, J., et al.: Learning transferable visual models from natural language supervision. In: ICML. pp. 8748-8763 (2021)", + "38. Raffel, C., Shazeer, N., Roberts, A., Lee, K., Narang, S., Matena, M., Zhou, Y., Li, W., Liu, P.J.: Exploring the limits of transfer learning with a unified text-to-text transformer. The Journal of Machine Learning Research 21(1), 5485-5551 (2020)", + "39. Seyfioglu, M.S., Ikezogwo, W.O., Ghezloo, F., Krishna, R., Shapiro, L.: Quilt-lava: Visual instruction tuning by extracting localized narratives from open-source histopathology videos. In: CVPR. pp. 13183-13192 (2024)", + "40. Silva-Rodriguez, J., Colomer, A., Sales, M.A., Molina, R., Naranjo, V.: Going deeper through the gleason scoring scale: An automatic end-to-end system for histology prostate grading and cribriform pattern detection. Computer methods and programs in biomedicine 195, 105637 (2020)", + "41. Sun, Y., Wang, S., Feng, S., Ding, S., Pang, C., Shang, J., Liu, J., Chen, X., Zhao, Y., Lu, Y., et al.: Ernie 3.0: Large-scale knowledge enhanced pre-training for language understanding and generation. arXiv preprint arXiv:2107.02137 (2021)", + "42. Sun, Y., Zhang, Y., Si, Y., Zhu, C., Shui, Z., Zhang, K., Li, J., Lyu, X., Lin, T., Yang, L.: Pathgen-1.6m: 1.6 million pathology image-text pairs generation through multi-agent collaboration (2024), https://arxiv.org/abs/2407.00203", + "43. Sun, Y., Zhu, C., Zhang, Y., Li, H., Chen, P., Yang, L.: Assessing the robustness of deep learning-assisted pathological image analysis under practical variables of imaging system. In: ICASSP. pp. 1-5 (2023). https://doi.org/10.1109/ICASSP49357.2023.10095887", + "44. Sun, Y., Zhu, C., Zheng, S., Zhang, K., Sun, L., Shui, Z., Zhang, Y., Li, H., Yang, L.: Pathasst: A generative foundation ai assistant towards artificial general intelligence of pathology. In: AAAI. pp. 5034-5042 (2024)", + "45. Team, G., Anil, R., Borgeaud, S., Wu, Y., Alayrac, J.B., Yu, J., Soricut, R., Schalkwyk, J., Dai, A.M., Hauth, A., et al.: Gemini: a family of highly capable multimodal models. arXiv preprint arXiv:2312.11805 (2023)", + "46. Touvron, H., Lavril, T., Izacard, G., Martinet, X., Lachaux, M.A., Lacroix, T., Rozière, B., Goyal, N., Hambro, E., Azhar, F., et al.: Llama: Open and efficient foundation language models. arXiv preprint arXiv:2302.13971 (2023)", + "47. Veeling, B.S., Linmans, J., Winkens, J., Cohen, T., Welling, M.: Rotation equivariant cnns for digital pathology. In: MICCAI. pp. 210-218. Springer (2018)", + "48. Wang, J., Zhou, Y., Xu, G., Shi, P., Zhao, C., Xu, H., Ye, Q., Yan, M., Zhang, J., Zhu, J., et al.: Evaluation and analysis of hallucination in large vision-language models. arXiv preprint arXiv:2308.15126 (2023)" + ], + "bbox": [ + 212, + 146, + 787, + 839 + ], + "page_idx": 16 + }, + { + "type": "header", + "text": "PathMMU", + "bbox": [ + 656, + 114, + 730, + 126 + ], + "page_idx": 16 + }, + { + "type": "page_number", + "text": "17", + "bbox": [ + 767, + 114, + 785, + 126 + ], + "page_idx": 16 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "49. Wang, W., Lv, Q., Yu, W., Hong, W., Qi, J., Wang, Y., Ji, J., Yang, Z., Zhao, L., Song, X., et al.: Cogvlm: Visual expert for pretrained language models. arXiv preprint arXiv:2311.03079 (2023)", + "50. Wei, J., Suriawinata, A., Ren, B., Liu, X., Lisovsky, M., Vaickus, L., Brown, C., Baker, M., Tomita, N., Torresani, L., et al.: A petri dish for histopathology image analysis. In: Artificial Intelligence in Medicine: 19th International Conference on Artificial Intelligence in Medicine, AIME 2021, Virtual Event, June 15-18, 2021, Proceedings. pp. 11-24. Springer (2021)", + "51. Xu, P., Shao, W., Zhang, K., Gao, P., Liu, S., Lei, M., Meng, F., Huang, S., Qiao, Y., Luo, P.: Lvlm-ehub: A comprehensive evaluation benchmark for large vision-language models. arXiv preprint arXiv:2306.09265 (2023)", + "52. Yin, Z., Wang, J., Cao, J., Shi, Z., Liu, D., Li, M., Huang, X., Wang, Z., Sheng, L., BAI, L., Shao, J., Ouyang, W.: Lamm: Language-assisted multi-modal instruction-tuning dataset, framework, and benchmark. In: NeurIPS. pp. 26650-26685 (2023)", + "53. Yu, W., Yang, Z., Li, L., Wang, J., Lin, K., Liu, Z., Wang, X., Wang, L.: Mmvet: Evaluating large multimodal models for integrated capabilities. arXiv preprint arXiv:2308.02490 (2023)", + "54. Yue, X., Ni, Y., Zhang, K., Zheng, T., Liu, R., Zhang, G., Stevens, S., Jiang, D., Ren, W., Sun, Y., Wei, C., Yu, B., Yuan, R., Sun, R., Yin, M., Zheng, B., Yang, Z., Liu, Y., Huang, W., Sun, H., Su, Y., Chen, W.: Mmmu: A massive multi-discipline multimodal understanding and reasoning benchmark for expert agi. In: CVPR. pp. 9556-9567 (2024)", + "55. Zhang, X., Wu, C., Zhao, Z., Lin, W., Zhang, Y., Wang, Y., Xie, W.: Pmc-vqa: Visual instruction tuning for medical visual question answering. arXiv preprint arXiv:2305.10415 (2023)", + "56. Zhang, Y., Sun, Y., Li, H., Zheng, S., Zhu, C., Yang, L.: Benchmarking the robustness of deep neural networks to common corruptions in digital pathology. In: MICCAI. pp. 242-252 (2022)", + "57. Zhang, Z., Chen, P., McGough, M., Xing, F., Wang, C., Bui, M., Xie, Y., Sapkota, M., Cui, L., Dhillon, J., et al.: Pathologist-level interpretable whole-slide cancer diagnosis with deep learning. Nature Machine Intelligence 1(5), 236-245 (2019)", + "58. Zheng, S., Cui, X., Sun, Y., Li, J., Li, H., Zhang, Y., Chen, P., Jing, X., Ye, Z., Yang, L.: Benchmarking pathclip for pathology image analysis. Journal of Imaging Informatics in Medicine pp. 1-17 (2024)", + "59. Zhu, C., Sun, Y., Li, H., Cui, C., Zhang, S., Cai, J., Ling, Y.: Weakly supervised classification using multi-level instance-aware optimization on cervical cytologic image. In: 2022 IEEE 19th International Symposium on Biomedical Imaging (ISBI). pp. 1-5. IEEE (2022)", + "60. Zhu, D., Chen, J., Shen, X., Li, X., Elhoseiny, M.: Minigpt-4: Enhancing vision-language understanding with advanced large language models. arXiv preprint arXiv:2304.10592 (2023)" + ], + "bbox": [ + 215, + 146, + 785, + 714 + ], + "page_idx": 17 + }, + { + "type": "page_number", + "text": "18", + "bbox": [ + 217, + 114, + 235, + 126 + ], + "page_idx": 17 + }, + { + "type": "header", + "text": "Y. Sun et al.", + "bbox": [ + 271, + 114, + 357, + 126 + ], + "page_idx": 17 + } +] \ No newline at end of file diff --git a/2024/PathMMU_ A Massive Multimodal Expert-Level Benchmark for Understanding and Reasoning in Pathology/0e28fdb1-202a-41b0-8b8d-6b1d1b08e3fe_model.json b/2024/PathMMU_ A Massive Multimodal Expert-Level Benchmark for Understanding and Reasoning in Pathology/0e28fdb1-202a-41b0-8b8d-6b1d1b08e3fe_model.json new file mode 100644 index 0000000000000000000000000000000000000000..4686fd87f65bc493740ce80c5dd2dbf916e01fe4 --- /dev/null +++ b/2024/PathMMU_ A Massive Multimodal Expert-Level Benchmark for Understanding and Reasoning in Pathology/0e28fdb1-202a-41b0-8b8d-6b1d1b08e3fe_model.json @@ -0,0 +1,2216 @@ +[ + [ + { + "type": "title", + "bbox": [ + 0.224, + 0.141, + 0.78, + 0.21 + ], + "angle": 0, + "content": "PathMMU: A Massive Multimodal Expert-Level Benchmark for Understanding and Reasoning in Pathology" + }, + { + "type": "text", + "bbox": [ + 0.218, + 0.234, + 0.787, + 0.283 + ], + "angle": 0, + "content": "Yuxuan Sun\\(^{1,2}\\), Hao Wu\\(^{3}\\), Chenglu Zhu\\(^{2}\\), Sunyi Zheng\\(^{2}\\), Qizi Chen\\(^{4}\\), Kai Zhang\\(^{5}\\), Yunlong Zhang\\(^{1,2}\\), Dan Wan\\(^{6}\\), Xiaoxiao Lan\\(^{1}\\), Mengyue Zheng\\(^{2}\\), Jingxiong Li\\(^{1,2}\\), Xinheng Lyu\\(^{2}\\), Tao Lin\\(^{2,\\dagger}\\), and Lin Yang\\(^{2,\\dagger}\\)" + }, + { + "type": "text", + "bbox": [ + 0.342, + 0.292, + 0.662, + 0.308 + ], + "angle": 0, + "content": "1 Zhejiang University, Hangzhou 310058, China" + }, + { + "type": "text", + "bbox": [ + 0.341, + 0.308, + 0.663, + 0.321 + ], + "angle": 0, + "content": "\\(^{2}\\) Westlake University, Hangzhou 310030, China" + }, + { + "type": "text", + "bbox": [ + 0.298, + 0.322, + 0.708, + 0.335 + ], + "angle": 0, + "content": "{sunyuxuan,zhuchenglu,lintao,yanglin}@westlake.edu.cn" + }, + { + "type": "text", + "bbox": [ + 0.27, + 0.335, + 0.732, + 0.348 + ], + "angle": 0, + "content": "3 Macau University of Science and Technology, Macau 999078, China" + }, + { + "type": "text", + "bbox": [ + 0.385, + 0.35, + 0.619, + 0.362 + ], + "angle": 0, + "content": "2220024311@student.must.edu.mo" + }, + { + "type": "text", + "bbox": [ + 0.356, + 0.362, + 0.648, + 0.376 + ], + "angle": 0, + "content": "\\(^{4}\\) Jiangnan University, Wuxi 214122, China" + }, + { + "type": "text", + "bbox": [ + 0.328, + 0.376, + 0.676, + 0.39 + ], + "angle": 0, + "content": "5 The Ohio State University, Columbus 43210, USA" + }, + { + "type": "text", + "bbox": [ + 0.25, + 0.39, + 0.753, + 0.404 + ], + "angle": 0, + "content": "\\(^{6}\\) Fujian University of Traditional Chinese Medicine, Fuzhou 350122, China" + }, + { + "type": "text", + "bbox": [ + 0.361, + 0.405, + 0.642, + 0.418 + ], + "angle": 0, + "content": "https://pathmmu-benchmark.github.io/" + }, + { + "type": "text", + "bbox": [ + 0.261, + 0.447, + 0.741, + 0.794 + ], + "angle": 0, + "content": "Abstract. The emergence of Large Multimodal Models (LMMs) has unlocked remarkable potential in AI, particularly in pathology. However, the lack of specialized, high-quality benchmark impeded their development and precise evaluation. To address this, we introduce PathMMU, the largest and highest-quality expert validated pathology benchmark for LMMs. It comprises 33,428 multimodal multi-choice questions and 24,067 images from various sources, each accompanied by an explanation for the correct answer. The construction of PathMMU leverages GPT-4V's advanced capabilities, utilizing over 30,000 image-caption pairs to enrich the descriptive quality of captions and generate corresponding Q&As in a cascading process. To maximize PathMMU's authority, we invite seven pathologists to scrutinize each question under strict standards in PathMMU's validation and test sets, while simultaneously setting an expert-level performance benchmark for PathMMU. We conduct extensive evaluations, including zero-shot assessments of 14 open-sourced and 4 closed-sourced LMMs and their robustness to image corruption. We also fine-tune representative LMMs to assess their adaptability to PathMMU. The empirical findings indicate that advanced LMMs struggle with the challenging PathMMU benchmark, with the top-performing LMM, GPT-4V, achieving only a \\(49.8\\%\\) zero-shot performance, significantly lower than the \\(71.8\\%\\) demonstrated by human pathologists. After fine-tuning, substantially smaller open-sourced LMMs can outperform GPT-4V but still fall short of the expertise shown by pathologists. We hope that the PathMMU will offer valuable insights and foster the development of more specialized, next-generation LMMs for pathology." + }, + { + "type": "text", + "bbox": [ + 0.262, + 0.806, + 0.687, + 0.821 + ], + "angle": 0, + "content": "Keywords: Large multimodal model \\(\\cdot\\) Pathology \\(\\cdot\\) Benchmark" + }, + { + "type": "page_footnote", + "bbox": [ + 0.221, + 0.825, + 0.386, + 0.84 + ], + "angle": 0, + "content": "† Corresponding author." + } + ], + [ + { + "type": "page_number", + "bbox": [ + 0.218, + 0.116, + 0.23, + 0.127 + ], + "angle": 0, + "content": "2" + }, + { + "type": "header", + "bbox": [ + 0.272, + 0.115, + 0.359, + 0.127 + ], + "angle": 0, + "content": "Y. Sun et al." + }, + { + "type": "title", + "bbox": [ + 0.218, + 0.145, + 0.375, + 0.161 + ], + "angle": 0, + "content": "1 Introduction" + }, + { + "type": "image", + "bbox": [ + 0.22, + 0.201, + 0.373, + 0.347 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.384, + 0.202, + 0.586, + 0.348 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.59, + 0.201, + 0.785, + 0.348 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.214, + 0.359, + 0.788, + 0.416 + ], + "angle": 0, + "content": "Fig. 1: An overview of the PathMMU benchmark: PathMMU is built from diverse, rich data sources. It comprises expert-level, multimodal, multiple-choice questions in pathology, collaboratively crafted by AI and pathology experts. The benchmark reveals even the most advanced LMMs significantly underperform compared to human experts." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.448, + 0.785, + 0.509 + ], + "angle": 0, + "content": "Pathology is integral to modern medicine, serving as the foundation for diagnosing and understanding diseases. For instance, liquid-based cytology is crucial for early cervical cancer detection. Assessing biomarkers like HER2 in breast cancer guides the selection of targeted therapies and immunotherapies [23]." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.509, + 0.787, + 0.674 + ], + "angle": 0, + "content": "In recent years, the field of pathology has undergone a significant transformation, driven by advances in digital pathology and the integration of AI. This shift marks a departure from conventional microscope-based slide reading to AI-powered image analysis, greatly easing the workload of pathologists. Traditional pathology models are tailored for specific tasks, such as cervical cytology and liver lesion classification, leading to an abundance of task-specific models [57,59]. In contrast, recent advancements in Large Multimodal Models (LMMs) focus on offering general task-solving capabilities, thereby making universal pathological region recognition possible. This breakthrough not only represents a significant stride in the field, but also paves the way for more versatile and efficient diagnostic approaches in pathology [13,27,42,44]." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.675, + 0.788, + 0.841 + ], + "angle": 0, + "content": "Given the demands of precise interpretation in pathology, conducting comprehensive evaluations of LMMs' abilities in interpreting pathology images is essential. However, the field faces a notable scarcity of high-quality benchmark datasets. Currently, the primary large-scale dataset available is PathVQA [18], which offers over 30,000 visual Q&A samples with 4,998 images. Nevertheless, these samples are derived from limited sources of textbooks, with image captions being converted into questions through heuristic approaches, constraining the generation of diverse and logical Q&As. Additionally, image captions in the textbook might not accurately reflect the corresponding images. For instance, some captions may introduce extraneous information not evident in the images or only partially cover the image content, thereby leading to unsolvable Q&As" + } + ], + [ + { + "type": "header", + "bbox": [ + 0.658, + 0.115, + 0.732, + 0.127 + ], + "angle": 0, + "content": "PathMMU" + }, + { + "type": "page_number", + "bbox": [ + 0.775, + 0.117, + 0.787, + 0.127 + ], + "angle": 0, + "content": "3" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.147, + 0.788, + 0.208 + ], + "angle": 0, + "content": "or questions that inadequately capture the essence of the images. Moreover, the absence of expert review and validation during the data curation process may introduce substantial noise into the dataset. These challenges pose significant obstacles to effectively validating LMMs' capabilities in pathology." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.214, + 0.789, + 0.533 + ], + "angle": 0, + "content": "To address these challenges, we introduce PathMMU, a multimodal expert-level benchmark designed to evaluate LMMs in pathology image understanding and reasoning, comprising 33,428 Q&As along with 24,067 pathology images. As depicted in Fig. 1, PathMMU draws from a diverse range of high-quality sources, including PubMed scientific papers, pathology atlas from guidelines, Twitter posts by pathology experts, commonly used pathology classification datasets and educational content from YouTube videos, covering multiple organ systems (e.g., gastrointestinal, pulmonary, etc.) and multiple subjects (e.g., dermatopathology, hematopathology, etc.). In developing PathMMU, we adopt a meticulously designed cascading approach. Initially, we prompt GPT-4V with collected image_caption pairs to enhance the original captions, crafting descriptions that delve deeper into details and highlight crucial morphological features. Subsequently, we use these enhanced descriptions along with the images to prompt GPT-4V to generate professional multi-choice, multimodal pathology Q&As, each accompanied by detailed explanations for their answers. To ensure the questions are specifically designed for multimodal pathology understanding, we employ a collaboration of multiple Large Language Models (LLMs) to eliminate questions that can be solved or guessed using text alone. Most importantly, we invite seven pathology experts to manually review approximately 12,000 Q&As from the test and validation sets of PathMMU. The above carefully executed procedures, ensure the generation of professional, logical, and high-quality Q&As." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.538, + 0.79, + 0.844 + ], + "angle": 0, + "content": "We evaluate the zero-shot performance of 18 advanced LMMs and their robustness on the PathMMU test sets. We also fine-tune two representative LMMs on the training set to assess their transfer learning capabilities. The key insights from this extensive analysis are: (1) Cutting-edge LMMs struggle with the PathMMU, with 15 out of the 18 achieving no more than \\(40\\%\\) accuracy. Even the most advanced model currently available, GPT-4V, only attains an accuracy of \\(49.8\\%\\). This reveals a notable discrepancy of nearly \\(20\\%\\) compared to professional pathologists, indicating significant deficiencies of current LMMs in the specialized field of pathology and emphasizing the challenging nature of the PathMMU benchmark. (2) While these LMMs show considerable robustness in robustness tests, the full extent of their robustness is still in question. On the one hand, given their limited performance, the scope for further performance degradation is somewhat constrained. On the other hand, these models tend to capture only superficial image features, and sometimes may not even utilize image information for reasoning. As a result, even when image corruption obscures minor details, the impact on these models' performance might not be significantly pronounced. (3) The text-only performance of GPT-4 Turbo surpasses the top-performing open-source LMMs on the PathMMU benchmark. We observe that GPT-4 Turbo sometimes takes shortcuts to guess answers. This involves choosing the correct answer based" + } + ], + [ + { + "type": "page_number", + "bbox": [ + 0.218, + 0.116, + 0.232, + 0.127 + ], + "angle": 0, + "content": "4" + }, + { + "type": "header", + "bbox": [ + 0.272, + 0.115, + 0.361, + 0.129 + ], + "angle": 0, + "content": "Y. Sun et al." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.147, + 0.788, + 0.254 + ], + "angle": 0, + "content": "either on the options most commonly encountered in real-world pathological scenarios, or by identifying the most distinctive option, rather than conducting an in-depth analysis of the pathology image. (4) Fine-tuning general-purpose LMMs with substantial amounts of data significantly boosts their ability to comprehend and reason with pathological images, enabling them to easily surpass the zero-shot performance of GPT-4V. However, there still remains a discernible gap between their performance and that of human experts." + }, + { + "type": "title", + "bbox": [ + 0.216, + 0.278, + 0.4, + 0.294 + ], + "angle": 0, + "content": "2 Related Works" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.311, + 0.789, + 0.569 + ], + "angle": 0, + "content": "Multimodal Models. The emergence of powerful large language models such as BERT [12], GPT-3 [8], T5 [38], LLaMa [46], ChatGPT [33], GPT-4 [34], and Vicuna [10] has significantly advanced the field of Natural Language Processing (NLP). These models demonstrate exceptional general capabilities, motivating researchers to explore the integration of LLMs with vision models to develop versatile multimodal models. This integration is primarily achieved through the pretraining approach, leading to the creation of groundbreaking LMMs. Representative models of this approach include CLIP [37], Flamingo [1], BLIP-2 [29], and Fuyu [5], all of which demonstrate remarkable multimodal understanding abilities. Furthermore, by adapting instruction-tuning techniques from the NLP domain, a variety of multimodal instruction-tuning datasets have been developed. This initiative aims to guide LMMs in generating outputs that are more controllable, practical, and adaptable to various tasks. Simultaneously, this approach facilitates the integration of LLMs into LMMs using lightweight fine-tuning, significantly reducing both costs and time. Key models that embody this advancement include GPT-4V [35], Gemini Pro Vision [45], Qwen-VL [4], LLaVA [30], MiniGPT-4 [60], BLIP-2 [29] and InstructBLIP [11]." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.569, + 0.79, + 0.719 + ], + "angle": 0, + "content": "LMM Benchmarks. In the general domain, the rise of various large models has led to the creation of extensive benchmark datasets, designed to evaluate the universal capabilities of these models across a range of tasks and domains. Notable examples include LAMM [52], LVLM-eHub [51], SEED [26], MMBench [31], MMVet [53] and BenchLMM [9], which have been used to assess the basic perception abilities of large models. More recently, MMMU [54], a massive and challenging dataset covering 30 university-level subjects, has been developed. It is specifically designed to evaluate the general multimodal understanding and reasoning capabilities of LMMs. Furthermore, HaELM [48] and HallusionBench [16] are proposed to evaluate the degree of hallucination in LMMs." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.72, + 0.789, + 0.841 + ], + "angle": 0, + "content": "In the medical field, datasets such as VQA-RAD [24] and VQA-Med [6] provide a collection of Q&A pairs based on radiology images, facilitating the evaluation of the capabilities of AI in radiological diagnostics. Additionally, PMC-VQA [55] is a large-scale dataset within this domain. It generates a vast number of Q&A pairs by prompting ChatGPT with text-only image captions from PubMed. In the pathology domain, limited dataset construction efforts have been performed to date. The representative large-scale dataset available, PathVQA [18], is constructed in a manner similar to PMC-VQA. It utilizes" + } + ], + [ + { + "type": "header", + "bbox": [ + 0.658, + 0.115, + 0.732, + 0.127 + ], + "angle": 0, + "content": "PathMMU" + }, + { + "type": "page_number", + "bbox": [ + 0.775, + 0.117, + 0.785, + 0.127 + ], + "angle": 0, + "content": "5" + }, + { + "type": "image", + "bbox": [ + 0.223, + 0.144, + 0.396, + 0.25 + ], + "angle": 0, + "content": null + }, + { + "type": "table", + "bbox": [ + 0.403, + 0.144, + 0.784, + 0.25 + ], + "angle": 0, + "content": "
DatasetDomainSources# Q&As / ImagesAnswer Explanation?LLM Filtered?Expert Annotated?
VQA-MedMedicalMedPix5000 / 5000XX✓
VQA-RADMedicalMedPix3515 / 315XX✓
PMC-VQAMedicalPMC227k / 149kXXX
PathVQAPathologyBook32799 / 4998XXX
Quilt-VQAPathologyEdu-Content1283 / 985XX✓
Quilt-Red
Social Media
PathMMUPathologyEdu-Content, PubMed33428 / 24067✓✓✓
Atlas, CLS-Data
" + }, + { + "type": "image_caption", + "bbox": [ + 0.214, + 0.261, + 0.788, + 0.348 + ], + "angle": 0, + "content": "Fig. 2: The comparison between PathMMU and existing benchmarks. The Q&A pairs in PathMMU are sourced extensively and comprehensively, undergoing rigorous multi-tiered filtering. This includes the initial filtering by multiple LLMs and the strict reviews by professional pathologists. Additionally, each question is accompanied by a detailed explanation. These attributes establish PathMMU as the most professionally curated, comprehensive, and highest-quality large-scale pathology dataset available." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.373, + 0.789, + 0.614 + ], + "angle": 0, + "content": "a heuristic approach to generate questions based on the text-only captions of pathology textbook images, yielding over 30,000 Q&As and 4,998 images. More recently, Quilt-VQA [39] is proposed, which is constructed by automatically extracting question and answer pairs from the narrations of professional teachers in YouTube videos. This approach guarantees the professionalism of the Q&A pairs, although it comprises a relatively smaller dataset with 1,283 Q&As and 985 images. Nonetheless, datasets reliant solely on image captions for question generation, face several notable limitations: (1) The creation of questions typically requires the consideration of both the image and its accompanying caption. Relying only on captions can introduce inaccuracies, making some questions unanswerable or solvable using only the text, bypassing the need for image analysis. (2) Since captions tend to be simplistic and may not capture the detailed information evident in images, generating an excessive number of Q&As from the same image's caption can lead to either too simplistic or insufficient questions. (3) Automatically generated questions, whether through predefined rules or by prompting ChatGPT, are susceptible to containing numerous errors." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.615, + 0.79, + 0.738 + ], + "angle": 0, + "content": "Given the current lack of high-quality benchmarks in pathology, we develop PathMMU, a large-scale, high-quality, and expert-verified pathology benchmark, to fill this gap. Our approach involves generating Q&As from a diverse range of image-caption pairs, followed by rigorous expert review and filtering applied to both the images and the questions. Moreover, we provide detailed explanations for each answer as a reference to enhance the interpretability of answers, which is absent in previous works. As shown in Fig. 2, PathMMU distinguishes itself in the pathology domain through its professionalism and superior quality." + }, + { + "type": "title", + "bbox": [ + 0.216, + 0.757, + 0.618, + 0.775 + ], + "angle": 0, + "content": "3 The Proposed PathMMU Benchmark" + }, + { + "type": "title", + "bbox": [ + 0.216, + 0.787, + 0.402, + 0.803 + ], + "angle": 0, + "content": "3.1 Design Principle" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.81, + 0.787, + 0.842 + ], + "angle": 0, + "content": "PathMMU is designed to provide the community with a specialized dataset for evaluating pathology LMMs. We adhere to five principal guidelines in its devel-" + } + ], + [ + { + "type": "page_number", + "bbox": [ + 0.218, + 0.116, + 0.23, + 0.127 + ], + "angle": 0, + "content": "6" + }, + { + "type": "header", + "bbox": [ + 0.272, + 0.115, + 0.359, + 0.128 + ], + "angle": 0, + "content": "Y. Sun et al." + }, + { + "type": "image", + "bbox": [ + 0.219, + 0.145, + 0.404, + 0.327 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.412, + 0.145, + 0.58, + 0.327 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.585, + 0.145, + 0.784, + 0.327 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.215, + 0.337, + 0.788, + 0.381 + ], + "angle": 0, + "content": "Fig. 3: An illustrative overview of the three main processes in PathMMU Q&A generation: data collection and preprocessing, detailed pathology image description generation, and question generation with LLMs filtering and expert validation." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.406, + 0.789, + 0.71 + ], + "angle": 0, + "content": "opment: (1) Comprehensive and Specialized Data: Our benchmark is gathered from a diverse range of authoritative materials, including respected scientific publications from PubMed, pathology atlas from textbooks and guidelines, educational videos on YouTube, pathologist-shared images with explanations on Twitter, and widely recognized pathology classification datasets. Additionally, we involve seven human experts to manually review the dataset to ensure it meets professional standards. (2) Effective and Valuable Questions: The questions in PathMMU are carefully formulated to necessitate answers based on detailed observation of pathology images, rather than being solvable through mere textual interpretation or guesswork. We ensure that these questions are answerable, effective, and aligned with the standards of professional pathology examinations. (3) Large-Scale: We present the largest pathology dataset currently available. This scale enables researchers to thoroughly explore the full potential of LMMs in the field of pathology. (4) High-Quality Images: We prioritize clarity in our images to ensure that every detail is discernible, as evidenced by our dataset's average resolution of approximately \\(900 \\times 700\\) pixels. (5) Explainable Answer Choices: In line with the ongoing interest in understanding the decision-making logic of large models, our dataset includes reference explanations for each answer choice. This feature enhances the interpretability of model responses, contributing to the broader research on model explainability." + }, + { + "type": "title", + "bbox": [ + 0.216, + 0.728, + 0.525, + 0.743 + ], + "angle": 0, + "content": "3.2 The Construction of PathMMU" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.75, + 0.788, + 0.796 + ], + "angle": 0, + "content": "To guarantee the quality of our benchmark, we meticulously develop a three-step data processing and generation protocol. The overall data collection and generation process is illustrated in Fig. 3." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.795, + 0.788, + 0.84 + ], + "angle": 0, + "content": "Step 1: Data Collection and Preprocessing. Our data collection draws from a wide variety of sources, which we integrate as subsets into the PathMMU framework. We name these subsets and detail their collection workflows as follows." + } + ], + [ + { + "type": "header", + "bbox": [ + 0.658, + 0.115, + 0.732, + 0.127 + ], + "angle": 0, + "content": "PathMMU" + }, + { + "type": "page_number", + "bbox": [ + 0.775, + 0.116, + 0.787, + 0.127 + ], + "angle": 0, + "content": "7" + }, + { + "type": "text", + "bbox": [ + 0.218, + 0.146, + 0.788, + 0.525 + ], + "angle": 0, + "content": "PubMed: This subset consists of scientific pathology image-text pairs sourced from scientific documents. We gather these pairs from the open-access section of PubMed Central. Given the significant presence of non-pathological data within this resource, we meticulously annotate 20,000 samples as either pathological or non-pathological, and subsequently train a ConvNeXt [32] model to identify pathology data within the remaining dataset. EduContent: This subset, derived from educational YouTube videos, is sourced from Quilt1M [20]. Since YouTube videos often include a variety of visuals unrelated to pathology, such as computer desktops and unrelated imagery, the initial phase involves the annotation of pathological regions in 3,000 YouTube video images. This critical step is followed by the training of a YOLOv6-based detector [28] to automate the identification of pathological regions within remaining YouTube content. Atlas: This subset is a compilation of authoritative pathology textbooks and guidelines, we transform them from PDFs to HTML to facilitate the extraction of image-caption pairs. SocialPath: This subset aggregates image-text pairs from Twitter posts by pathology experts, using Twitter URLs provided by OpenPath [19]. PathCLS: Encompassing widely recognized pathology classification datasets, including PatchCamelyon [47], CRC-100K [21], SICAPv2 [40], BACH [2], Osteo [3], SkinCancer [22], MHIST [50], WSSSLUAD [17], LC-Lung and LC-Colon [7]. We reformulate these datasets into a Q&A format and randomly sample from these datasets. We directly make the gathered samples from PathCLS a part of the PathMMU dataset without undergoing the following two steps Q&A generation process. Finally, we manually review and filter irrelevant or unclear images to ensure the quality of collected data. As a result, approximately 30,000 high-quality image-text pairs are obtained, forming the foundation of the PathMMU." + }, + { + "type": "text", + "bbox": [ + 0.218, + 0.529, + 0.788, + 0.711 + ], + "angle": 0, + "content": "Step 2: Detailed Description Generation and Question-answer Pairs Generation. When dealing with data sources from platforms like Twitter and YouTube, it's common to find a weak correlation between an image and its accompanying caption. For instance, in the SocialPath dataset, captions might only describe the aesthetic appeal of the image, which is irrelevant to pathological findings. Similarly, in other data sources, captions tend to describe only a fraction of the features present in the image. To tackle this issue, we prompt GPT-4V to focus on describing the morphology of cells and tissues. It is crucial to note that merely prompting GPT-4V to output image descriptions might lead to incoherent content. Therefore, we provide GPT-4V with original captions for reference, significantly reducing the incidence of such occurrences and thus ensuring the generation of more precise and relevant descriptions." + }, + { + "type": "text", + "bbox": [ + 0.218, + 0.715, + 0.788, + 0.807 + ], + "angle": 0, + "content": "Step 3: Question Generation and Expert Validation. Following the generation of image descriptions, GPT-4V is tasked with creating three questions per image, each accompanied by multiple-choice options, the correct answer, and detailed explanations for the answer. The creation of these explanations serves a dual purpose: it not only offers valuable insights into the model reasoning process, but also facilitates the subsequent manual review phase." + }, + { + "type": "text", + "bbox": [ + 0.217, + 0.81, + 0.787, + 0.84 + ], + "angle": 0, + "content": "It's crucial to note that despite the careful design of question generation prompts for GPT-4V, the text-only GPT-4 is capable of correctly answering" + } + ], + [ + { + "type": "page_number", + "bbox": [ + 0.218, + 0.116, + 0.23, + 0.127 + ], + "angle": 0, + "content": "8" + }, + { + "type": "header", + "bbox": [ + 0.272, + 0.115, + 0.358, + 0.127 + ], + "angle": 0, + "content": "Y. Sun et al." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.147, + 0.788, + 0.283 + ], + "angle": 0, + "content": "over \\(60\\%\\) of these questions through educated guesses, bypassing the need for visual cues. GPT-4 explains that its deductions are drawn from patterns such as one option being more common in typical pathological scenarios or exhibiting noticeable differences from other options. We delve deeper into this phenomenon in Sec. 4.3 to offer a more thorough experimental analysis. To ensure that LMMs do not rely on educated guesses, thereby obscuring their true multimodal capabilities, we employ GPT-3.5 Turbo, GPT-4 Turbo, Gemini Pro, and ERNIE-Bot-4 to perform educated guesses. Questions correctly guessed by at least three of these models are subsequently excluded." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.283, + 0.788, + 0.478 + ], + "angle": 0, + "content": "We partitioned the remaining dataset into training, validation, and test sets. The training set comprises 16,344 images and 23,041 Q&As, while the validation and test sets contain roughly 12,000 Q&As and over 8,000 images, respectively. For the validation and test sets, we invite seven professional pathologists to conduct a thorough manual review. These pathologists initially attempt to answer the questions independently and then assess them based on the following criteria: (1) Whether the question can be answered without an accompanying image. (2) Whether the answer can be inferred from the provided question and image. (3) Whether the supplied answer is incorrect, if there is no correct answer, or if multiple correct answers are possible. (4) Whether the generated question appears unusual or atypical compared to standard questions in pathology examinations. Questions failing to meet these standards are deemed invalid and consequently removed from the PathMMU dataset." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.48, + 0.788, + 0.587 + ], + "angle": 0, + "content": "Ultimately, after thorough expert review, we obtain 710 Q&As accompanied by 510 images for the validation set, and 9,677 Q&As with 7,213 images for the test set. Additionally, to establish a standard for expert performance, we extract a smaller subset from the test set, named 'test-tiny', which includes 1,156 Q&As. We invite two groups of professional pathologists to participate in this subset as an examination, and we average their performance to set a benchmark for expert performance in the PathMMU, serving as a comparison reference for LMMs." + }, + { + "type": "title", + "bbox": [ + 0.216, + 0.605, + 0.377, + 0.623 + ], + "angle": 0, + "content": "4 Experiments" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.633, + 0.788, + 0.771 + ], + "angle": 0, + "content": "In this section, we conduct extensive experiments on PathMMU. Initially, we assess the zero-shot performance of cutting-edge LMMs. To verify whether LMMs effectively utilize the visual information from pathology images, we also evaluate the performance of text-only LLMs on PathMMU as a reference. Additionally, we apply common corruptions to the images in the test-tiny set to test the robustness of the LMMs against image corruptions. Furthermore, we select representative LMMs and evaluate their performance to explore their transfer-learning capabilities. Finally, we conduct experimental analysis on the issue of LLMs being able to guess answers identified during the data construction process." + }, + { + "type": "title", + "bbox": [ + 0.216, + 0.788, + 0.607, + 0.803 + ], + "angle": 0, + "content": "4.1 Zero-shot Evaluation of LMMs and LLMs" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.81, + 0.787, + 0.84 + ], + "angle": 0, + "content": "In this study, we evaluate the zero-shot capabilities of the latest and most advanced LMMs on the PathMMU. Specifically, we use the PathMMU validation" + } + ], + [ + { + "type": "header", + "bbox": [ + 0.658, + 0.115, + 0.732, + 0.127 + ], + "angle": 0, + "content": "PathMMU" + }, + { + "type": "page_number", + "bbox": [ + 0.775, + 0.116, + 0.787, + 0.127 + ], + "angle": 0, + "content": "9" + }, + { + "type": "table_caption", + "bbox": [ + 0.215, + 0.145, + 0.788, + 0.201 + ], + "angle": 0, + "content": "Table 1: Overall results of models on the PathMMU validation and test set. Besides reporting the performance of LMMs, we add text-only LLM baselines that purely accept text as inputs. The best-performing LMM in each subset is in- bold, and the top-performing LLM is underlined." + }, + { + "type": "table", + "bbox": [ + 0.219, + 0.213, + 0.785, + 0.546 + ], + "angle": 0, + "content": "
Validation OverallTestOverallPubMedSocialPathEduContentAtlasPathCLS
- (710)Tiny (1156)ALL (9677)Tiny (281)ALL (3068)Tiny (235)All (1855)Tiny (255)All (1938)Tiny (208)ALL (1007)Tiny (177)ALL (1809)
Random Choice24.622.123.722.125.125.526.525.526.019.723.015.316.3
Frequent Choice27.527.725.528.826.127.726.729.826.528.427.522.021.0
Expert performance-71.8-72.9-71.5-69.0-68.3-78.9-
Large Multimodal Models (LMMs): Text + Image as Input
OpenFlamingo2-9B [1]26.224.124.224.925.525.525.319.624.428.825.522.020.4
Kosmos2 [36]26.926.124.927.825.827.224.325.125.923.124.227.122.8
LLaMA-Adapter2-7B [14]26.226.626.426.028.327.726.527.128.030.827.420.920.5
MiniGPT4-Vicuna-13B [60]27.225.527.728.830.124.327.225.529.327.928.619.221.8
Otter [25]26.129.328.134.930.226.028.430.630.029.328.223.221.3
Fuyu-8B [5]27.530.129.235.930.528.930.730.629.827.930.624.323.9
CogVLM [49]29.930.629.732.032.131.130.430.629.935.133.722.622.4
Qwen-VL-7B [4]29.432.131.534.932.933.635.834.133.635.633.818.621.7
BLIP-2 FLAN-T5-XL [29]32.832.631.835.934.634.933.633.732.834.637.420.320.8
BLIP-2 FLAN-T5-XXL [29]33.433.333.537.037.435.734.630.234.539.440.719.820.6
InstructBLIP FLAN-T5-XL [11]33.134.931.837.033.235.333.936.533.337.036.726.623.1
InstructBLIP FLAN-T5-XXL [11]32.734.333.939.137.233.634.334.536.038.539.322.622.7
LLaVA-1.5-7B [30]36.634.935.441.639.937.938.132.536.535.639.223.221.9
LLaVA-1.5-13B [30]37.938.837.644.541.040.440.434.139.447.144.324.923.5
Qwen-VL-PLUS [4]38.039.334.343.537.741.336.039.636.044.737.123.223.3
Qwen-VL-MAX [4]43.649.245.953.050.953.649.352.247.951.449.830.529.6
Gemini Pro Vision [45]41.942.842.743.844.942.442.043.543.749.549.432.834.7
GPT-4V-1106 [35]49.353.949.859.453.558.753.960.453.648.152.836.233.8
Large Language Models (LLMs): Only Text as Input
ERNIE-Bot 4.0 [41]31.834.330.937.031.233.632.940.034.536.137.420.920.6
Gemini Pro [45]31.031.631.031.333.531.931.433.332.738.037.721.520.9
Vicuna-v1.5-13B [10]32.031.231.635.233.830.634.331.033.534.635.722.020.7
GPT-3.5 Turbo [33]30.429.428.432.731.231.530.131.030.126.926.822.020.8
GPT-4 Turbo [34]36.541.838.148.842.443.442.347.140.641.343.222.321.1
" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.581, + 0.788, + 0.778 + ], + "angle": 0, + "content": "set to conduct prompt engineering for these models, and then test their performance on the test set. The tested models include 14 open-source LMMs: OpenFlamingo [1], Kosmos2 [36], LLaMA-Adapter2 [14], MiniGPT-4 [60], Otter [25], Fuyu [5], CogVLM [49], Qwen-VL [4], BLIP-2 [29], InstructBLIP [11], and LLaVA-1.5 [30], as well as 4 closed-source LMMs: Qwen-VL-PLUS, Qwen-VL-MAX [4], Gemini Pro Vision [45], and GPT-4V [35]. For BLIP-2, Instruct-BLIP, and LLaVA-1.5, we deploy various sizes of these models to examine the applicability of scaling rules in LMMs. Moreover, we extend our evaluation to purely text-based LLMs, including ERNIE-Bot 4.0 [41], Gemini Pro [45], Vicuna-v1.5-13B [10], GPT-3.5 Turbo [33], and GPT-4 Turbo [34]. This assessment aims to quantify their ability to infer correct answers through educated guesses without visual inputs, reflecting the extent to which LMMs integrate and leverage visual information compared to their text-only counterparts." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.78, + 0.788, + 0.841 + ], + "angle": 0, + "content": "Results and Discussion. The advanced LMMs struggle with the PathMMU dataset. As shown in Tab. 1, among the 18 models tested, 15 show an accuracy below \\(40\\%\\). Notably, the highest-performing open-source model, LLaVa-1.5-13B, and the top-performing closed-source model, GPT-4V, achieve" + } + ], + [ + { + "type": "page_number", + "bbox": [ + 0.218, + 0.116, + 0.236, + 0.127 + ], + "angle": 0, + "content": "10" + }, + { + "type": "header", + "bbox": [ + 0.272, + 0.115, + 0.358, + 0.127 + ], + "angle": 0, + "content": "Y. Sun et al." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.147, + 0.783, + 0.221 + ], + "angle": 0, + "content": "only \\(37.6\\%\\) and \\(49.8\\%\\) accuracy, respectively. Those are significantly lower than human expert performance, which stands at \\(71.8\\%\\), underscoring a substantial gap between current LMMs and professional pathologist standards. This gap indicates that the practical application of LMMs in real-world pathological scenarios remains significantly constrained." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.223, + 0.785, + 0.524 + ], + "angle": 0, + "content": "Closed-source LLMs achieve performance that is on par with or even surpasses that of the most advanced open-source LMMs. When considering text-only models, we prompt them to make educated guesses without providing the image. We observe that GPT-4 Turbo, Vicuna-v1.5-13B, GeminiPro, and ERNIE-Bot 4.0 outperform nearly half of the open-source LMMs. Notably, GPT-4 Turbo leads with a significant margin, achieving an accuracy rate of \\(38.1\\%\\), exceeding the second-best LLM by \\(6.5\\%\\), and even outperforming the best-performing open-source LMM, LLaVA-1.5-13B. The \\(38.1\\%\\) accuracy of GPT-4 Turbo, on the one hand, demonstrates that under our strict filtering criteria, even the best LLM, when provided with text alone, makes incorrect answers for the majority of questions. This underscores the effectiveness of our dataset in testing the capabilities of LMMs in pathological image analysis. On the other hand, it reaffirms the analysis in Sec. 3.2 that GPT-4 possesses strong logical reasoning abilities, enabling it to correctly guess a portion of the questions, significantly outperforming random selection. However, this raises critical concerns in pathology diagnosis: the reliability of model outputs in clinical scenarios remains questionable, especially when they rely on educated guesses, particularly probabilistic ones, rather than comprehensive analysis. This emphasizes the urgent necessity to develop interpretable AI models and advocate for their cautious application in pathology diagnostics." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.525, + 0.785, + 0.601 + ], + "angle": 0, + "content": "The larger LMMs exhibit better performance. Specifically, LLaVa-1.5-13B exceeds LLaVa-1.5-7B model by \\(2.2\\%\\) in overall performance, while Instruct-BLIP FLAN-T5-XXL and BLIP-2 FLAN-T5 XXL surpass their smaller versions by \\(2.1\\%\\) and \\(1.7\\%\\), respectively. This suggests that larger models typically possess stronger multimodal capabilities in the field of pathology." + }, + { + "type": "title", + "bbox": [ + 0.215, + 0.621, + 0.614, + 0.637 + ], + "angle": 0, + "content": "4.2 Robustness of LMMs to Image Corruption" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.644, + 0.783, + 0.734 + ], + "angle": 0, + "content": "In practical pathology, the interpretation of models significantly influences the subsequent medical decisions and treatment strategies. Therefore, models with strong robustness are crucial for clinical applications. The quality of pathological slides can be affected by various factors during staining, scanning, and storage, including JPEG compression, pixelation, blur (e.g., bubble blur, defocus blur, and motion blur), and color variations (e.g., brightness, saturation, and hue)." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.735, + 0.783, + 0.809 + ], + "angle": 0, + "content": "Drawing inspiration from the studies on the robustness of encoder-based models to pathology image corruptions [43, 56, 58], we incorporate these aforementioned types of corruptions into our analysis. To be specific, we simulate five levels of each corruption type on the PathMMU to explore the robustness of LMMs against these corruptions, as depicted in the left half of Fig. 4." + }, + { + "type": "text", + "bbox": [ + 0.215, + 0.81, + 0.783, + 0.84 + ], + "angle": 0, + "content": "Results and Discussion. LMMs demonstrate notable robustness to various types and levels of image corruptions, yet their true robustness" + } + ], + [ + { + "type": "header", + "bbox": [ + 0.658, + 0.115, + 0.732, + 0.127 + ], + "angle": 0, + "content": "PathMMU" + }, + { + "type": "page_number", + "bbox": [ + 0.769, + 0.116, + 0.784, + 0.127 + ], + "angle": 0, + "content": "11" + }, + { + "type": "image", + "bbox": [ + 0.221, + 0.144, + 0.455, + 0.282 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.458, + 0.147, + 0.619, + 0.281 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.62, + 0.147, + 0.78, + 0.278 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.214, + 0.292, + 0.784, + 0.349 + ], + "angle": 0, + "content": "Fig. 4: Left: Illustration of corrupted pathology images. Right: LMM's performance across various levels of color-related (brightness, hue, saturation) and image quality-related (pixelation, JPEG compression, bubble blur, motion blur, defocus blur) corruptions on the PathMMU test-tiny set, with level 0 representing the uncorrupted images." + }, + { + "type": "table_caption", + "bbox": [ + 0.214, + 0.365, + 0.785, + 0.408 + ], + "angle": 0, + "content": "Table 2: Evaluation of the model's robustness on the PathMMU test-tiny set, where green indicates performance improvement and red signifies performance degradation compared to the model's performance with uncorrupted images." + }, + { + "type": "table", + "bbox": [ + 0.22, + 0.42, + 0.787, + 0.507 + ], + "angle": 0, + "content": "
BrightnessBubbleDefocusHueJPEGMotionPixelateSaturateOverall
Qwen-VL-7B32.7↑0.632.3↑0.232.1↓0.032.4↓0.332.3↑0.233.8↑1.733.1↑1.033.0↑0.932.7↑0.6
BLIP-2 FLAN-T5-XL32.9↑0.332.4↓0.232.8↑0.232.9↑0.332.4↓0.232.8↑0.232.6↑0.032.7↑0.132.7↓0.1
BLIP-2 FLAN-T5-XXL33.1↓0.233.2↓0.133.5↑0.233.7↑0.433.5↑0.233.1↓0.233.1↓0.233.6↑0.333.3↓0.0
InstructBLIP FLAN-T5-XL34.0↓0.934.1↓0.833.9↓1.034.1↓0.833.7↓1.234.0↓0.934.4↓0.533.8↓1.134.0↓0.9
InstructBLIP FLAN-T5-XXL34.0↓0.333.8↓0.534.2↓0.134.1↓0.234.3↓0.234.3↑0.034.1↓0.233.8↓0.534.1↓0.2
LLaVA-1.5-7B34.8↓0.135.3↑0.434.1↓0.834.3↓0.634.0↓0.934.3↓0.634.6↓0.334.7↓0.234.5↓0.4
LLaVA-1.5-13B38.9↑0.138.3↓0.538.3↓0.538.4↓0.438.0↓0.838.6↓0.238.7↓0.139.1↑0.338.5↓0.3
" + }, + { + "type": "table_caption", + "bbox": [ + 0.214, + 0.523, + 0.784, + 0.551 + ], + "angle": 0, + "content": "Table 3: Results of LMMs on the PathMMU test set with the original images substituted by random Gaussian noise images." + }, + { + "type": "table", + "bbox": [ + 0.218, + 0.563, + 0.787, + 0.643 + ], + "angle": 0, + "content": "
OverallPubMedSocialPathEduContentAtlasPathCLS
TinyALLTinyALLTinyAllTinyAllTinyALLTinyALL
Qwen-VL-7B28.8↓3.328.3↓3.239.5↑4.630.4↓2.524.7↓8.930.5↓5.327.8↓6.328.9↓4.727.4↓8.230.1↓3.720.3↑1.720.8↓0.9
BLIP-FLAN-T5-XL30.9↓1.730.8↓1.030.6↓5.333.1↓1.534.5↑0.433.4↓0.232.2↑1.531.9↓0.934.1↓0.535.1↓2.320.9↑0.620.8↓0.0
BLIP-FLAN-T5-XXL32.4↓0.931.6↓1.934.5↓2.534.6↓2.834.0↑1.732.5↓2.130.6↑0.432.8↑1.738.5↓0.938.1↓2.622.2↑2.420.8↑0.2
InstructBLIP FLAN-T5-XL31.1↓3.829.7↓2.132.4↓4.630.8↓2.432.8↓2.532.0↓1.933.7↓2.832.4↓0.932.7↓4.330.4↓6.321.5↓5.122.4↓0.7
InstructBLIP FLAN-T5-XXL30.4↓3.930.8↓3.132.0↓7.133.5↓3.728.9↑4.731.9↓2.432.2↑2.331.7↓4.335.1↑3.435.2↓4.121.5↓1.121.7↓1.0
LLAVA-1.5-7B30.2↓4.730.9↓4.537.0↓4.633.9↓6.031.5↓6.432.5↓5.628.2↓4.332.7↓3.829.8↓5.833.9↓5.320.9↓2.320.8↓1.1
LLAVA-1.5-13B32.7↓6.133.2↓4.438.4↓6.136.0↓5.035.3↓5.136.3↓4.129.0↓5.134.1↓5.335.6↓11.538.8↓5.522.0↓2.921.0↓2.5
" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.674, + 0.788, + 0.841 + ], + "angle": 0, + "content": "is questionable, as shown in Tab. 2 and the right half of Fig. 4. Notably, the Qwen-VL-7B even shows a \\(0.6\\%\\) overall performance increase under corruption compared to its baseline. It is more plausible to hypothesize that these corruptions mainly alter minute details of the pathology image that are difficult for LMMs to discern (such as chromatin morphology of the nucleus or vacuolization of the cytoplasm). This limitation stems from their pre-training in general domains, where they learn to recognize larger and more prominent features (e.g., humans, houses, cars, etc.) rather than the nuanced details that are vital in pathology. Furthermore, they may even resort to exploiting spurious correlations as shortcuts to answer questions (e.g., deducing answers from textual patterns instead of focusing on essential image details.)" + } + ], + [ + { + "type": "page_number", + "bbox": [ + 0.218, + 0.116, + 0.236, + 0.127 + ], + "angle": 0, + "content": "12" + }, + { + "type": "header", + "bbox": [ + 0.272, + 0.115, + 0.358, + 0.127 + ], + "angle": 0, + "content": "Y. Sun et al." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.147, + 0.788, + 0.329 + ], + "angle": 0, + "content": "To substantiate our argument, we employ an extreme test of corruption by replacing the images with Gaussian noise-generated random images. As shown in Tab. 3, we discover that even when the images contain no relevant information, LMMs can still achieve results significantly better than random choice. The drop in performance when using these random images ranged from only \\(1.0\\%\\) to \\(4.5\\%\\). This finding suggests that LMMs might rely on shortcuts to accomplish the multimodal task, such as depending exclusively on textual information to make predictions. Additionally, it is intriguing that the performance drop across different sizes of the same model is remarkably similar. In other words, the benefit that images provide to models of various model sizes appears to be consistent, suggesting that the improvement in performance for different-sized models may primarily stem from their LLM component, rather than the vision aspect." + }, + { + "type": "title", + "bbox": [ + 0.215, + 0.349, + 0.629, + 0.365 + ], + "angle": 0, + "content": "4.3 Analysis of LLMs' Ability to Guess Answers" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.372, + 0.786, + 0.402 + ], + "angle": 0, + "content": "In this section, we delve deeper into the phenomenon of LLM's ability to make educated guesses with more comprehensive experiments and analysis." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.403, + 0.787, + 0.522 + ], + "angle": 0, + "content": "To more clearly demonstrate the guessing abilities of LLMs, we randomly select 100 samples that are filtered during the Q&A generation process and can be correctly guessed by multiple LLMs. We invite pathology experts to answer these questions with reference images. As shown in Fig. 5, despite having access to images, their performance is significantly lower than that of the closed-source LLMs, which achieve about \\(90\\%\\) accuracy, and only slightly better than the smaller open-source Vicuna-v1.5-13B. This indicates that LLMs may outperform humans in guessing correct answers by identifying shortcuts within the questions." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.524, + 0.788, + 0.674 + ], + "angle": 0, + "content": "We hypothesize several key reasons for the guessing behavior exhibited by LLMs: (1) Frequency-based guessing: LLMs may guess based on the prevalence of certain options in real-world pathological instances. (2) Based on the options that present a pattern of one supporting and three contradicting, or vice versa: For example, when a question identifies a tumour diagnosis, one option may match the pathological traits of a lesion, while the others suggest non-lesional characteristics. (3) When only one option aligns with the question's subject: For example, if a question asks which feature in the image supports the diagnosis of a Low-grade Squamous Intraepithelial Lesion (LSIL) cell, with only one option describes LSIL's pathological characteristics." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.675, + 0.788, + 0.84 + ], + "angle": 0, + "content": "To support our hypothesis, we expand the sample size to 1,600, with 400 from each source (except PathCLS), to empirically analyze how LLMs guess answers. Specifically, we swap the questions among these samples while keeping the options unchanged, creating samples where the questions and options are completely mismatched. As shown in the right half of Fig. 5, we observe that LLMs still manage to guess approximately \\(50\\%\\) cases correctly, significantly higher than random choice accuracy. This finding suggests that the models tend to select the most common or prominent option as the answer, supporting our hypotheses (1) and (2). To further explore hypothesis (3), we design experiments using BERT-large [12] and BiomedBERT-large [15], the latter being specifically pre-trained on biomedical data. Given that the BERT series incorporates Next-Sentence" + } + ], + [ + { + "type": "header", + "bbox": [ + 0.658, + 0.115, + 0.732, + 0.127 + ], + "angle": 0, + "content": "PathMMU" + }, + { + "type": "page_number", + "bbox": [ + 0.769, + 0.116, + 0.786, + 0.127 + ], + "angle": 0, + "content": "13" + }, + { + "type": "image", + "bbox": [ + 0.225, + 0.143, + 0.476, + 0.264 + ], + "angle": 0, + "content": null + }, + { + "type": "table", + "bbox": [ + 0.484, + 0.145, + 0.788, + 0.261 + ], + "angle": 0, + "content": "
PubMed SocialPath EduContentAtlas
Random Choice25.824.325.525.825.3
GPT-4 Turbo93.091.886.883.888.8
+ Random Swap Question44.345.550.051.547.8
GPT-3.5 Turbo84.386.579.585.383.9
+ Random Swap Question44.044.553.851.348.4
Gemini Pro85.590.176.571.584.1
+ Random Swap Question43.844.850.351.347.5
ERNIE-Bot 4.094.592.389.382.089.5
+ Random Swap Question47.046.556.551.050.3
BERT-large w/ NSP24.536.037.534.333.1
BiomedBERT-large w/ NSP44.344.842.337.542.3
" + }, + { + "type": "image_caption", + "bbox": [ + 0.214, + 0.273, + 0.788, + 0.373 + ], + "angle": 0, + "content": "Fig. 5: Left: The performance comparison between different LLMs and human experts on 100 filtered samples where the answer can be guessed through text-only. Right: Expand the sample size to 1600 to validate the source of LLMs' ability to guess answers, which includes: (1) randomly replacing the original questions with others from the dataset while keeping the options unchanged, and (2) utilizing the BERT series for answer selection, specifically through its Next Sentence Prediction (NSP), to assess whether an option is the sequential sentence following a question." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.4, + 0.789, + 0.553 + ], + "angle": 0, + "content": "Prediction (NSP) in their pretraining, which inherently relies on the similarity between two sentences, we apply this to predict the relationship between questions and options. We select the option with the highest probability of being the correct next sentence to the question. Our findings reveal that both BERT and BiomedBERT substantially surpass random guessing, indicating that direct matching of questions and options is a viable method for models to guess the correct answers, thereby supporting hypothesis (3). Moreover, BiomedBERT-large demonstrates a notable performance enhancement compared to BERT-large. This suggests that pretraining on biomedical data equips the model with a broader understanding of pathological knowledge." + }, + { + "type": "title", + "bbox": [ + 0.216, + 0.574, + 0.427, + 0.59 + ], + "angle": 0, + "content": "4.4 Fine-tuning Results" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.599, + 0.787, + 0.674 + ], + "angle": 0, + "content": "To explore the adaptability of LMMs to the pathology domain, we select representative LMMs, InstructBLIP FLAN-T5-XL and InstructBLIP FLAN-T5-XXL for fine-tuning experiments. Our experiments consist of two parts: (1) training the LMMs to directly generate answers, and (2) fine-tuning the LMMs to generate a reasoning process before delivering the final answer." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.674, + 0.788, + 0.81 + ], + "angle": 0, + "content": "Results and Discussion. All models exhibit significant improvements on the PathMMU test set after fine-tuning on its training set, as detailed in Tab. 4. Notably, during the fine-tuning for direct answer generation, InstructBLIP FLAN-T5-XL and InstructBLIP FLAN-T5-XXL achieve significant improvements of \\(21.5\\%\\) and \\(21.3\\%\\), respectively. This substantial improvement enables them to outperform the current leading model, GPT-4V. While still lagging behind expert performance, these results demonstrate a trend toward approaching expert-level proficiency. This also reflects the effectiveness of PathMMU in enhancing the LMMs' abilities for pathology image analysis." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.81, + 0.787, + 0.84 + ], + "angle": 0, + "content": "Unexpectedly, generating explanations before answers during fine-tuning does not yield improvements. Instead, we observe slight performance decreases of \\(0.3\\%\\)" + } + ], + [ + { + "type": "page_number", + "bbox": [ + 0.218, + 0.116, + 0.236, + 0.127 + ], + "angle": 0, + "content": "14" + }, + { + "type": "header", + "bbox": [ + 0.272, + 0.115, + 0.358, + 0.127 + ], + "angle": 0, + "content": "Y. Sun et al." + }, + { + "type": "table_caption", + "bbox": [ + 0.217, + 0.145, + 0.785, + 0.187 + ], + "angle": 0, + "content": "Table 4: Results of LMMs on the PathMMU test set after the fine-tuning. The 'w/ A' and 'w/ A&E' denote the model is fine-tuned to output the answer directly or to output the answer with an explanation for its answer." + }, + { + "type": "table", + "bbox": [ + 0.218, + 0.199, + 0.787, + 0.268 + ], + "angle": 0, + "content": "
OverallPubMedSocialPathEduContentAtlasPathCLS
TinyALLTinyALLTinyALLTinyAllTinyALLTinyALL
InstructBLIP FLAN-T5-XL34.931.837.033.235.333.936.533.337.036.726.623.1
+ fine-tune w/A55.7 ± 20.853.3 ± 21.553.7 ± 16.750.4 ± 17.252.8 ± 17.552.3 ± 18.456.5 ± 20.052.0 ± 18.756.2 ± 19.253.1 ± 16.461.0 ± 34.460.8 ± 37.7
+ fine-tune w/A&E54.9 ± 20.053.0 ± 21.255.5 ± 18.551.4 ± 18.251.9 ± 16.649.7 ± 15.855.3 ± 18.851.5 ± 18.253.4 ± 16.451.6 ± 14.959.3 ± 32.761.4 ± 38.3
InstructBLIP FLAN-T5-XXL34.333.939.137.233.634.334.536.038.539.322.622.7
+ fine-tune w/A56.8 ± 22.555.2 ± 21.355.2 ± 16.151.5 ± 14.359.6 ± 26.055.2 ± 20.958.4 ± 23.954.1 ± 18.150.5 ± 12.053.7 ± 14.461.0 ± 38.463.7 ± 41.0
+ fine-tune w/A&E51.0 ± 16.752.9 ± 19.048.8 ± 9.750.8 ± 13.655.3 ± 21.751.2 ± 16.952.2 ± 17.751.9 ± 15.944.2 ± 5.750.1 ± 10.854.8 ± 32.260.5 ± 37.8
" + }, + { + "type": "text", + "bbox": [ + 0.217, + 0.292, + 0.785, + 0.368 + ], + "angle": 0, + "content": "and \\(2.3\\%\\) compared to direct answer generation, respectively. We speculate that generating explanations is relatively more challenging, and incorporating it with answer generation might impede the models' capacity to generate correct answers. This finding raises an important question: how can we effectively leverage the interpretability information within PathMMU to enhance models' training?" + }, + { + "type": "title", + "bbox": [ + 0.218, + 0.389, + 0.357, + 0.404 + ], + "angle": 0, + "content": "5 Conclusion" + }, + { + "type": "text", + "bbox": [ + 0.217, + 0.417, + 0.786, + 0.629 + ], + "angle": 0, + "content": "In this study, we introduce PathMMU, the largest and highest-quality pathology benchmark to date, specifically crafted to evaluate the capabilities of LMMs in interpreting and reasoning with pathology images. The construction of PathMMU involves a meticulous data collection and curation process, supplemented by a strict manual review by seven professional pathologists to ensure its quality and professionalism. Moreover, we establish a human expert performance benchmark to quantify the gap between cutting-edge LMMs and human experts. Our experimental results reveal that advanced LMMs significantly lag behind on PathMMU, with these models demonstrating poor performance in observing details in pathology images and sometimes even neglecting visual information, highlighting a substantial gap in practical pathology application. However, LMMs demonstrate notable performance improvements after fine-tuning on PathMMU, even surpassing GPT-4V. While they do not achieve human expert-level performance, these LMMs show promising potential for analyzing pathology images." + }, + { + "type": "text", + "bbox": [ + 0.217, + 0.63, + 0.786, + 0.84 + ], + "angle": 0, + "content": "Future Directions for Pathology LMMs: Our experience with PathMMU highlights key areas for development in pathology LMMs: (1) Current LMMs, which are primarily based on LLMs and fine-tuned in a lightweight manner, tend to over-rely on textual information while neglecting visual data. There is a significant need to explore training methodologies or model structures that better integrate visual and textual modalities. (2) There's a tendency for LMMs to take shortcuts, solving problems in a \"lazy\" manner. This necessitates the development of trustworthy models for real-world clinical applications. (3) Given that most current LMMs do not support multi-image inputs, PathMMU does not include a benchmark for processing multiple images. However, in practical scenarios, pathologists typically analyze samples at various magnifications and perspectives, underscoring the importance of developing LMMs capable of handling multi-image inputs. Finally, we believe PathMMU will catalyze significant advancements in the development of next-generation LMMs in pathology." + } + ], + [ + { + "type": "header", + "bbox": [ + 0.658, + 0.115, + 0.732, + 0.127 + ], + "angle": 0, + "content": "PathMMU" + }, + { + "type": "page_number", + "bbox": [ + 0.769, + 0.117, + 0.786, + 0.127 + ], + "angle": 0, + "content": "15" + }, + { + "type": "title", + "bbox": [ + 0.218, + 0.145, + 0.403, + 0.163 + ], + "angle": 0, + "content": "Acknowledgements" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.174, + 0.788, + 0.234 + ], + "angle": 0, + "content": "This study was partially supported by the National Natural Science Foundation of China (Grant No.92270108), Zhejiang Provincial Natural Science Foundation of China (Grant No.XHD23F0201), the Research Center for Industries of the Future (RCIF) at Westlake University, and the Westlake Education Foundation." + }, + { + "type": "title", + "bbox": [ + 0.218, + 0.255, + 0.323, + 0.27 + ], + "angle": 0, + "content": "References" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.282, + 0.785, + 0.323 + ], + "angle": 0, + "content": "1. Alayrac, J.B., Donahue, J., Luc, P., Miech, A., Barr, I., Hasson, Y., Lenc, K., Mensch, A., Millican, K., Reynolds, M., et al.: Flamingo: a visual language model for few-shot learning. In: NeurIPS. pp. 23716-23736 (2022)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.324, + 0.785, + 0.363 + ], + "angle": 0, + "content": "2. Aresta, G., Araújo, T., Kwok, S., Chennamsetty, S.S., Safwan, M., Alex, V., Marami, B., Prastawa, M., Chan, M., Donovan, M., et al.: Bach: Grand challenge on breast cancer histology images. Medical image analysis 56, 122-139 (2019)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.364, + 0.785, + 0.418 + ], + "angle": 0, + "content": "3. Arunachalam, H.B., Mishra, R., Daescu, O., Cederberg, K., Rakheja, D., Sengupta, A., Leonard, D., Hallac, R., Leavey, P.: Viable and necrotic tumor assessment from whole slide images of osteosarcoma using machine-learning and deep-learning models. PloS one 14(4), e0210706 (2019)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.419, + 0.785, + 0.46 + ], + "angle": 0, + "content": "4. Bai, J., Bai, S., Yang, S., Wang, S., Tan, S., Wang, P., Lin, J., Zhou, C., Zhou, J.: Qwen-vl: A frontier large vision-language model with versatile abilities. arXiv preprint arXiv:2308.12966 (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.46, + 0.785, + 0.487 + ], + "angle": 0, + "content": "5. Bavishi, R., Elsen, E., Hawthorne, C., Nye, M., Odena, A., Somani, A., Taşırlar, S.: Introducing our multimodal models (2023), https://www.adept.ai/blog/fuyu-8b" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.487, + 0.785, + 0.54 + ], + "angle": 0, + "content": "6. Ben Abacha, A., Sarrouti, M., Demner-Fushman, D., Hasan, S.A., Müller, H.: Overview of the vqa-med task at imagecef 2021: Visual question answering and generation in the medical domain. In: Proceedings of the CLEF 2021 Conference and Labs of the Evaluation Forum-working notes (2021)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.541, + 0.785, + 0.581 + ], + "angle": 0, + "content": "7. Borkowski, A.A., Bui, M.M., Thomas, L.B., Wilson, C.P., DeLand, L.A., Mastorides, S.M.: Lung and colon cancer histopathological image dataset (lc25000). arXiv preprint arXiv:1912.12142 (2019)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.582, + 0.785, + 0.622 + ], + "angle": 0, + "content": "8. Brown, T., Mann, B., Ryder, N., Subbiah, M., Kaplan, J.D., Dhariwal, P., Neelakantan, A., Shyam, P., Sastry, G., Askell, A., et al.: Language models are few-shot learners. In: NeurIPS. pp. 1877-1901 (2020)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.622, + 0.785, + 0.662 + ], + "angle": 0, + "content": "9. Cai, R., Song, Z., Guan, D., Chen, Z., Luo, X., Yi, C., Kot, A.: Benchlmm: Benchmarking cross-style visual capability of large multimodal models. arXiv preprint arXiv:2312.02896 (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.22, + 0.663, + 0.785, + 0.717 + ], + "angle": 0, + "content": "10. Chiang, W.L., Li, Z., Lin, Z., Sheng, Y., Wu, Z., Zhang, H., Zheng, L., Zhuang, S., Zhuang, Y., Gonzalez, J.E., Stoica, I., Xing, E.P.: Vicuna: An open-source chatbot impressing gpt-4 with \\(90\\%\\) chatgpt quality (March 2023), https://lmsys.org/blog/2023-03-30-vicuna/" + }, + { + "type": "ref_text", + "bbox": [ + 0.22, + 0.718, + 0.785, + 0.758 + ], + "angle": 0, + "content": "11. Dai, W., Li, J., Li, D., Tiong, A.M.H., Zhao, J., Wang, W., Li, B., Fung, P., Hoi, S.: Instructclip: Towards general-purpose vision-language models with instruction tuning. arXiv preprint arXiv:2305.06500 (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.22, + 0.759, + 0.785, + 0.799 + ], + "angle": 0, + "content": "12. Devlin, J., Chang, M.W., Lee, K., Toutanova, K.: Bert: Pre-training of deep bidirectional transformers for language understanding. In: NAACL. pp. 4171-4186 (2019)" + }, + { + "type": "ref_text", + "bbox": [ + 0.22, + 0.799, + 0.785, + 0.84 + ], + "angle": 0, + "content": "13. Driess, D., Xia, F., Sajjadi, M.S., Lynch, C., Chowdhery, A., Ichter, B., Wahid, A., Tompson, J., Vuong, Q., Yu, T., et al.: Palm-e: An embodied multimodal language model. In: ICML. pp. 8469-8488 (2023)" + }, + { + "type": "list", + "bbox": [ + 0.22, + 0.282, + 0.785, + 0.84 + ], + "angle": 0, + "content": null + } + ], + [ + { + "type": "page_number", + "bbox": [ + 0.218, + 0.116, + 0.236, + 0.127 + ], + "angle": 0, + "content": "16" + }, + { + "type": "header", + "bbox": [ + 0.272, + 0.115, + 0.358, + 0.127 + ], + "angle": 0, + "content": "Y. Sun et al." + }, + { + "type": "ref_text", + "bbox": [ + 0.218, + 0.147, + 0.785, + 0.189 + ], + "angle": 0, + "content": "14. Gao, P., Han, J., Zhang, R., Lin, Z., Geng, S., Zhou, A., Zhang, W., Lu, P., He, C., Yue, X., et al.: Llama-adapter v2: Parameter-efficient visual instruction model. arXiv preprint arXiv:2304.15010 (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.218, + 0.19, + 0.787, + 0.245 + ], + "angle": 0, + "content": "15. Gu, Y., Tinn, R., Cheng, H., Lucas, M., Usuyama, N., Liu, X., Naumann, T., Gao, J., Poon, H.: Domain-specific language model pretraining for biomedical natural language processing. ACM Transactions on Computing for Healthcare (HEALTH) 3(1), 1-23 (2021)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.247, + 0.785, + 0.302 + ], + "angle": 0, + "content": "16. Guan, T., Liu, F., Wu, X., Xian, R., Li, Z., Liu, X., Wang, X., Chen, L., Huang, F., Yacoob, Y., et al.: Hallusionbench: an advanced diagnostic suite for entangled language hallucination and visual illusion in large vision-language models. In: CVPR. pp. 14375-14385 (2024)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.304, + 0.785, + 0.345 + ], + "angle": 0, + "content": "17. Han, C., Pan, X., Yan, L., Lin, H., Li, B., Yao, S., Lv, S., Shi, Z., Mai, J., Lin, J., et al.: Wss54luad: Grand challenge on weakly-supervised tissue semantic segmentation for lung adenocarcinoma. arXiv preprint arXiv:2204.06455 (2022)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.347, + 0.785, + 0.373 + ], + "angle": 0, + "content": "18. He, X., Zhang, Y., Mou, L., Xing, E., Xie, P.: Pathvqa: 30000+ questions for medical visual question answering. arXiv preprint arXiv:2003.10286 (2020)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.375, + 0.785, + 0.415 + ], + "angle": 0, + "content": "19. Huang, Z., Bianchi, F., Yuksekgonul, M., Montine, T.J., Zou, J.: A visual-language foundation model for pathology image analysis using medical twitter. Nature medicine 29(9), 2307-2316 (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.417, + 0.785, + 0.458 + ], + "angle": 0, + "content": "20. Ikezogwo, W., Seyfioglu, S., Ghezloo, F., Geva, D., Sheikh Mohammed, F., Anand, P.K., Krishna, R., Shapiro, L.: Quilt-1m: One million image-text pairs for histopathology. In: NeurIPS. pp. 37995–38017 (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.46, + 0.785, + 0.486 + ], + "angle": 0, + "content": "21. Kather, J.N., Halama, N., Marx, A.: 100,000 histological images of human colorectal cancer and healthy tissue. Zenodo10 5281 (2018)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.488, + 0.785, + 0.543 + ], + "angle": 0, + "content": "22. Kriegsmann, K., Lovers, F., Zgorzelski, C., Kriegsmann, J., Janssen, C., Meliss, R.R., Muley, T., Sack, U., Steinbuss, G., Kriegsmann, M.: Deep learning for the detection of anatomical tissue structures and neoplasms of the skin on scanned histopathological tissue sections. Frontiers in Oncology 12, 1022967 (2022)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.545, + 0.785, + 0.571 + ], + "angle": 0, + "content": "23. Kumar, V., Abbas, A.K., Fausto, N., Aster, J.C.: Robbins and Cotran pathologic basis of disease, professional edition e-book. Elsevier health sciences (2014)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.573, + 0.785, + 0.613 + ], + "angle": 0, + "content": "24. Lau, J.J., Gayen, S., Ben Abacha, A., Demner-Fushman, D.: A dataset of clinically generated visual questions and answers about radiology images. Scientific data 5(1), 1-10 (2018)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.615, + 0.785, + 0.642 + ], + "angle": 0, + "content": "25. Li, B., Zhang, Y., Chen, L., Wang, J., Yang, J., Liu, Z.: Otter: A multi-modal model with in-context instruction tuning. arXiv preprint arXiv:2305.03726 (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.644, + 0.785, + 0.685 + ], + "angle": 0, + "content": "26. Li, B., Wang, R., Wang, G., Ge, Y., Ge, Y., Shan, Y.: benseed-bench: Benchmarking multimodal llms with generative comprehension. arXiv preprint arXiv:2307.16125 (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.687, + 0.785, + 0.727 + ], + "angle": 0, + "content": "27. Li, C., Wong, C., Zhang, S., Usuyama, N., Liu, H., Yang, J., Naumann, T., Poon, H., Gao, J.: Llava-med: Training a large language-and-vision assistant for biomedicine in one day. In: NeurIPS. pp. 28541-28564 (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.729, + 0.785, + 0.769 + ], + "angle": 0, + "content": "28. Li, C., Li, L., Jiang, H., Weng, K., Geng, Y., Li, L., Ke, Z., Li, Q., Cheng, M., Nie, W., et al.: Yolov6: A single-stage object detection framework for industrial applications. arXiv preprint arXiv:2209.02976 (2022)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.771, + 0.785, + 0.811 + ], + "angle": 0, + "content": "29. Li, J., Li, D., Savarese, S., Hoi, S.: Blip-2: Bootstrapping language-image pretraining with frozen image encoders and large language models. In: ICML. pp. 19730-19742 (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.813, + 0.785, + 0.84 + ], + "angle": 0, + "content": "30. Liu, H., Li, C., Wu, Q., Lee, Y.J.: Visual instruction tuning. In: NeurIPS. pp. 34892-34916 (2023)" + }, + { + "type": "list", + "bbox": [ + 0.217, + 0.147, + 0.787, + 0.84 + ], + "angle": 0, + "content": null + } + ], + [ + { + "type": "header", + "bbox": [ + 0.658, + 0.115, + 0.732, + 0.127 + ], + "angle": 0, + "content": "PathMMU" + }, + { + "type": "page_number", + "bbox": [ + 0.769, + 0.116, + 0.786, + 0.127 + ], + "angle": 0, + "content": "17" + }, + { + "type": "ref_text", + "bbox": [ + 0.214, + 0.147, + 0.788, + 0.189 + ], + "angle": 0, + "content": "31. Liu, Y., Duan, H., Zhang, Y., Li, B., Zhang, S., Zhao, W., Yuan, Y., Wang, J., He, C., Liu, Z., et al.: Mmbench: Is your multi-modal model an all-around player? arXiv preprint arXiv:2307.06281 (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.214, + 0.19, + 0.787, + 0.218 + ], + "angle": 0, + "content": "32. Liu, Z., Mao, H., Wu, C.Y., Feichtenhofer, C., Darrell, T., Xie, S.: A convnet for the 2020s. In: CVPR. pp. 11976-11986 (2022)" + }, + { + "type": "ref_text", + "bbox": [ + 0.216, + 0.219, + 0.715, + 0.233 + ], + "angle": 0, + "content": "33. OpenAI: Introducing chatgpt. https://openai.com/blog/chatgpt (2022)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.234, + 0.503, + 0.248 + ], + "angle": 0, + "content": "34. OpenAI: Gpt-4 technical report (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.249, + 0.786, + 0.276 + ], + "angle": 0, + "content": "35. OpenAI: Gpt-4v(ision) system card. https://cdn.openai.com/papers/GPTV_System/Card.pdf (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.277, + 0.787, + 0.318 + ], + "angle": 0, + "content": "36. Peng, Z., Wang, W., Dong, L., Hao, Y., Huang, S., Ma, S., Wei, F.: Kosmos-2: Grounding multimodal large language models to the world. arXiv preprint arXiv:2306.14824 (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.319, + 0.787, + 0.36 + ], + "angle": 0, + "content": "37. Radford, A., Kim, J.W., Hallacy, C., Ramesh, A., Goh, G., Agarwal, S., Sastry, G., Askell, A., Mishkin, P., Clark, J., et al.: Learning transferable visual models from natural language supervision. In: ICML. pp. 8748-8763 (2021)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.362, + 0.787, + 0.403 + ], + "angle": 0, + "content": "38. Raffel, C., Shazeer, N., Roberts, A., Lee, K., Narang, S., Matena, M., Zhou, Y., Li, W., Liu, P.J.: Exploring the limits of transfer learning with a unified text-to-text transformer. The Journal of Machine Learning Research 21(1), 5485-5551 (2020)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.404, + 0.787, + 0.445 + ], + "angle": 0, + "content": "39. Seyfioglu, M.S., Ikezogwo, W.O., Ghezloo, F., Krishna, R., Shapiro, L.: Quilt-lava: Visual instruction tuning by extracting localized narratives from open-source histopathology videos. In: CVPR. pp. 13183-13192 (2024)" + }, + { + "type": "ref_text", + "bbox": [ + 0.218, + 0.446, + 0.787, + 0.5 + ], + "angle": 0, + "content": "40. Silva-Rodriguez, J., Colomer, A., Sales, M.A., Molina, R., Naranjo, V.: Going deeper through the gleason scoring scale: An automatic end-to-end system for histology prostate grading and cribriform pattern detection. Computer methods and programs in biomedicine 195, 105637 (2020)" + }, + { + "type": "ref_text", + "bbox": [ + 0.218, + 0.502, + 0.787, + 0.543 + ], + "angle": 0, + "content": "41. Sun, Y., Wang, S., Feng, S., Ding, S., Pang, C., Shang, J., Liu, J., Chen, X., Zhao, Y., Lu, Y., et al.: Ernie 3.0: Large-scale knowledge enhanced pre-training for language understanding and generation. arXiv preprint arXiv:2107.02137 (2021)" + }, + { + "type": "ref_text", + "bbox": [ + 0.218, + 0.545, + 0.787, + 0.586 + ], + "angle": 0, + "content": "42. Sun, Y., Zhang, Y., Si, Y., Zhu, C., Shui, Z., Zhang, K., Li, J., Lyu, X., Lin, T., Yang, L.: Pathgen-1.6m: 1.6 million pathology image-text pairs generation through multi-agent collaboration (2024), https://arxiv.org/abs/2407.00203" + }, + { + "type": "ref_text", + "bbox": [ + 0.218, + 0.588, + 0.787, + 0.641 + ], + "angle": 0, + "content": "43. Sun, Y., Zhu, C., Zhang, Y., Li, H., Chen, P., Yang, L.: Assessing the robustness of deep learning-assisted pathological image analysis under practical variables of imaging system. In: ICASSP. pp. 1-5 (2023). https://doi.org/10.1109/ICASSP49357.2023.10095887" + }, + { + "type": "ref_text", + "bbox": [ + 0.218, + 0.643, + 0.787, + 0.685 + ], + "angle": 0, + "content": "44. Sun, Y., Zhu, C., Zheng, S., Zhang, K., Sun, L., Shui, Z., Zhang, Y., Li, H., Yang, L.: Pathasst: A generative foundation ai assistant towards artificial general intelligence of pathology. In: AAAI. pp. 5034-5042 (2024)" + }, + { + "type": "ref_text", + "bbox": [ + 0.218, + 0.686, + 0.787, + 0.727 + ], + "angle": 0, + "content": "45. Team, G., Anil, R., Borgeaud, S., Wu, Y., Alayrac, J.B., Yu, J., Soricut, R., Schalkwyk, J., Dai, A.M., Hauth, A., et al.: Gemini: a family of highly capable multimodal models. arXiv preprint arXiv:2312.11805 (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.218, + 0.728, + 0.787, + 0.769 + ], + "angle": 0, + "content": "46. Touvron, H., Lavril, T., Izacard, G., Martinet, X., Lachaux, M.A., Lacroix, T., Rozière, B., Goyal, N., Hambro, E., Azhar, F., et al.: Llama: Open and efficient foundation language models. arXiv preprint arXiv:2302.13971 (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.218, + 0.77, + 0.787, + 0.798 + ], + "angle": 0, + "content": "47. Veeling, B.S., Linmans, J., Winkens, J., Cohen, T., Welling, M.: Rotation equivariant cnns for digital pathology. In: MICCAI. pp. 210-218. Springer (2018)" + }, + { + "type": "ref_text", + "bbox": [ + 0.218, + 0.799, + 0.787, + 0.84 + ], + "angle": 0, + "content": "48. Wang, J., Zhou, Y., Xu, G., Shi, P., Zhao, C., Xu, H., Ye, Q., Yan, M., Zhang, J., Zhu, J., et al.: Evaluation and analysis of hallucination in large vision-language models. arXiv preprint arXiv:2308.15126 (2023)" + }, + { + "type": "list", + "bbox": [ + 0.214, + 0.147, + 0.788, + 0.84 + ], + "angle": 0, + "content": null + } + ], + [ + { + "type": "page_number", + "bbox": [ + 0.218, + 0.116, + 0.236, + 0.127 + ], + "angle": 0, + "content": "18" + }, + { + "type": "header", + "bbox": [ + 0.272, + 0.115, + 0.358, + 0.127 + ], + "angle": 0, + "content": "Y. Sun et al." + }, + { + "type": "ref_text", + "bbox": [ + 0.218, + 0.147, + 0.785, + 0.189 + ], + "angle": 0, + "content": "49. Wang, W., Lv, Q., Yu, W., Hong, W., Qi, J., Wang, Y., Ji, J., Yang, Z., Zhao, L., Song, X., et al.: Cogvlm: Visual expert for pretrained language models. arXiv preprint arXiv:2311.03079 (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.19, + 0.787, + 0.259 + ], + "angle": 0, + "content": "50. Wei, J., Suriawinata, A., Ren, B., Liu, X., Lisovsky, M., Vaickus, L., Brown, C., Baker, M., Tomita, N., Torresani, L., et al.: A petri dish for histopathology image analysis. In: Artificial Intelligence in Medicine: 19th International Conference on Artificial Intelligence in Medicine, AIME 2021, Virtual Event, June 15-18, 2021, Proceedings. pp. 11-24. Springer (2021)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.26, + 0.785, + 0.3 + ], + "angle": 0, + "content": "51. Xu, P., Shao, W., Zhang, K., Gao, P., Liu, S., Lei, M., Meng, F., Huang, S., Qiao, Y., Luo, P.: Lvlm-ehub: A comprehensive evaluation benchmark for large vision-language models. arXiv preprint arXiv:2306.09265 (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.301, + 0.785, + 0.342 + ], + "angle": 0, + "content": "52. Yin, Z., Wang, J., Cao, J., Shi, Z., Liu, D., Li, M., Huang, X., Wang, Z., Sheng, L., BAI, L., Shao, J., Ouyang, W.: Lamm: Language-assisted multi-modal instruction-tuning dataset, framework, and benchmark. In: NeurIPS. pp. 26650-26685 (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.343, + 0.785, + 0.383 + ], + "angle": 0, + "content": "53. Yu, W., Yang, Z., Li, L., Wang, J., Lin, K., Liu, Z., Wang, X., Wang, L.: Mmvet: Evaluating large multimodal models for integrated capabilities. arXiv preprint arXiv:2308.02490 (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.384, + 0.785, + 0.452 + ], + "angle": 0, + "content": "54. Yue, X., Ni, Y., Zhang, K., Zheng, T., Liu, R., Zhang, G., Stevens, S., Jiang, D., Ren, W., Sun, Y., Wei, C., Yu, B., Yuan, R., Sun, R., Yin, M., Zheng, B., Yang, Z., Liu, Y., Huang, W., Sun, H., Su, Y., Chen, W.: Mmmu: A massive multi-discipline multimodal understanding and reasoning benchmark for expert agi. In: CVPR. pp. 9556-9567 (2024)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.453, + 0.785, + 0.493 + ], + "angle": 0, + "content": "55. Zhang, X., Wu, C., Zhao, Z., Lin, W., Zhang, Y., Wang, Y., Xie, W.: Pmc-vqa: Visual instruction tuning for medical visual question answering. arXiv preprint arXiv:2305.10415 (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.494, + 0.785, + 0.535 + ], + "angle": 0, + "content": "56. Zhang, Y., Sun, Y., Li, H., Zheng, S., Zhu, C., Yang, L.: Benchmarking the robustness of deep neural networks to common corruptions in digital pathology. In: MICCAI. pp. 242-252 (2022)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.536, + 0.785, + 0.577 + ], + "angle": 0, + "content": "57. Zhang, Z., Chen, P., McGough, M., Xing, F., Wang, C., Bui, M., Xie, Y., Sapkota, M., Cui, L., Dhillon, J., et al.: Pathologist-level interpretable whole-slide cancer diagnosis with deep learning. Nature Machine Intelligence 1(5), 236-245 (2019)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.578, + 0.785, + 0.618 + ], + "angle": 0, + "content": "58. Zheng, S., Cui, X., Sun, Y., Li, J., Li, H., Zhang, Y., Chen, P., Jing, X., Ye, Z., Yang, L.: Benchmarking pathclip for pathology image analysis. Journal of Imaging Informatics in Medicine pp. 1-17 (2024)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.619, + 0.785, + 0.673 + ], + "angle": 0, + "content": "59. Zhu, C., Sun, Y., Li, H., Cui, C., Zhang, S., Cai, J., Ling, Y.: Weakly supervised classification using multi-level instance-aware optimization on cervical cytologic image. In: 2022 IEEE 19th International Symposium on Biomedical Imaging (ISBI). pp. 1-5. IEEE (2022)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.674, + 0.785, + 0.715 + ], + "angle": 0, + "content": "60. Zhu, D., Chen, J., Shen, X., Li, X., Elhoseiny, M.: Minigpt-4: Enhancing vision-language understanding with advanced large language models. arXiv preprint arXiv:2304.10592 (2023)" + }, + { + "type": "list", + "bbox": [ + 0.217, + 0.147, + 0.787, + 0.715 + ], + "angle": 0, + "content": null + } + ] +] \ No newline at end of file diff --git a/2024/PathMMU_ A Massive Multimodal Expert-Level Benchmark for Understanding and Reasoning in Pathology/0e28fdb1-202a-41b0-8b8d-6b1d1b08e3fe_origin.pdf b/2024/PathMMU_ A Massive Multimodal Expert-Level Benchmark for Understanding and Reasoning in Pathology/0e28fdb1-202a-41b0-8b8d-6b1d1b08e3fe_origin.pdf new file mode 100644 index 0000000000000000000000000000000000000000..6fc37d2eeba111b0e2f5f766d9f205af4539f9ec --- /dev/null +++ b/2024/PathMMU_ A Massive Multimodal Expert-Level Benchmark for Understanding and Reasoning in Pathology/0e28fdb1-202a-41b0-8b8d-6b1d1b08e3fe_origin.pdf @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:01a0a0bae5bed79c4873e3d914c0c757d61500e8b73efbba7310c9e25a520635 +size 2996299 diff --git a/2024/PathMMU_ A Massive Multimodal Expert-Level Benchmark for Understanding and Reasoning in Pathology/full.md b/2024/PathMMU_ A Massive Multimodal Expert-Level Benchmark for Understanding and Reasoning in Pathology/full.md new file mode 100644 index 0000000000000000000000000000000000000000..5ac7060fad795de58c465a0713f0da6561a056f1 --- /dev/null +++ b/2024/PathMMU_ A Massive Multimodal Expert-Level Benchmark for Understanding and Reasoning in Pathology/full.md @@ -0,0 +1,256 @@ +# PathMMU: A Massive Multimodal Expert-Level Benchmark for Understanding and Reasoning in Pathology + +Yuxuan Sun $^{1,2}$ , Hao Wu $^{3}$ , Chenglu Zhu $^{2}$ , Sunyi Zheng $^{2}$ , Qizi Chen $^{4}$ , Kai Zhang $^{5}$ , Yunlong Zhang $^{1,2}$ , Dan Wan $^{6}$ , Xiaoxiao Lan $^{1}$ , Mengyue Zheng $^{2}$ , Jingxiong Li $^{1,2}$ , Xinheng Lyu $^{2}$ , Tao Lin $^{2,\dagger}$ , and Lin Yang $^{2,\dagger}$ + +1 Zhejiang University, Hangzhou 310058, China + +$^{2}$ Westlake University, Hangzhou 310030, China + +{sunyuxuan,zhuchenglu,lintao,yanglin}@westlake.edu.cn + +3 Macau University of Science and Technology, Macau 999078, China + +2220024311@student.must.edu.mo + +$^{4}$ Jiangnan University, Wuxi 214122, China + +5 The Ohio State University, Columbus 43210, USA + +$^{6}$ Fujian University of Traditional Chinese Medicine, Fuzhou 350122, China + +https://pathmmu-benchmark.github.io/ + +Abstract. The emergence of Large Multimodal Models (LMMs) has unlocked remarkable potential in AI, particularly in pathology. However, the lack of specialized, high-quality benchmark impeded their development and precise evaluation. To address this, we introduce PathMMU, the largest and highest-quality expert validated pathology benchmark for LMMs. It comprises 33,428 multimodal multi-choice questions and 24,067 images from various sources, each accompanied by an explanation for the correct answer. The construction of PathMMU leverages GPT-4V's advanced capabilities, utilizing over 30,000 image-caption pairs to enrich the descriptive quality of captions and generate corresponding Q&As in a cascading process. To maximize PathMMU's authority, we invite seven pathologists to scrutinize each question under strict standards in PathMMU's validation and test sets, while simultaneously setting an expert-level performance benchmark for PathMMU. We conduct extensive evaluations, including zero-shot assessments of 14 open-sourced and 4 closed-sourced LMMs and their robustness to image corruption. We also fine-tune representative LMMs to assess their adaptability to PathMMU. The empirical findings indicate that advanced LMMs struggle with the challenging PathMMU benchmark, with the top-performing LMM, GPT-4V, achieving only a $49.8\%$ zero-shot performance, significantly lower than the $71.8\%$ demonstrated by human pathologists. After fine-tuning, substantially smaller open-sourced LMMs can outperform GPT-4V but still fall short of the expertise shown by pathologists. We hope that the PathMMU will offer valuable insights and foster the development of more specialized, next-generation LMMs for pathology. + +Keywords: Large multimodal model $\cdot$ Pathology $\cdot$ Benchmark + +# 1 Introduction + +![](images/50434f96daebcdae5f9726c2f375bba90c25bf354d535fffe668de2d0405ab41.jpg) +Fig. 1: An overview of the PathMMU benchmark: PathMMU is built from diverse, rich data sources. It comprises expert-level, multimodal, multiple-choice questions in pathology, collaboratively crafted by AI and pathology experts. The benchmark reveals even the most advanced LMMs significantly underperform compared to human experts. + +![](images/fab67f6ef220219bc7e6430a568fd532310a1b2d6726a7c43dbbf7fdec5810f3.jpg) + +![](images/688b82a269cf86c14034923178aa4720a73a259f9bc7399aa7c9cd3a6df5c92c.jpg) + +Pathology is integral to modern medicine, serving as the foundation for diagnosing and understanding diseases. For instance, liquid-based cytology is crucial for early cervical cancer detection. Assessing biomarkers like HER2 in breast cancer guides the selection of targeted therapies and immunotherapies [23]. + +In recent years, the field of pathology has undergone a significant transformation, driven by advances in digital pathology and the integration of AI. This shift marks a departure from conventional microscope-based slide reading to AI-powered image analysis, greatly easing the workload of pathologists. Traditional pathology models are tailored for specific tasks, such as cervical cytology and liver lesion classification, leading to an abundance of task-specific models [57,59]. In contrast, recent advancements in Large Multimodal Models (LMMs) focus on offering general task-solving capabilities, thereby making universal pathological region recognition possible. This breakthrough not only represents a significant stride in the field, but also paves the way for more versatile and efficient diagnostic approaches in pathology [13,27,42,44]. + +Given the demands of precise interpretation in pathology, conducting comprehensive evaluations of LMMs' abilities in interpreting pathology images is essential. However, the field faces a notable scarcity of high-quality benchmark datasets. Currently, the primary large-scale dataset available is PathVQA [18], which offers over 30,000 visual Q&A samples with 4,998 images. Nevertheless, these samples are derived from limited sources of textbooks, with image captions being converted into questions through heuristic approaches, constraining the generation of diverse and logical Q&As. Additionally, image captions in the textbook might not accurately reflect the corresponding images. For instance, some captions may introduce extraneous information not evident in the images or only partially cover the image content, thereby leading to unsolvable Q&As + +or questions that inadequately capture the essence of the images. Moreover, the absence of expert review and validation during the data curation process may introduce substantial noise into the dataset. These challenges pose significant obstacles to effectively validating LMMs' capabilities in pathology. + +To address these challenges, we introduce PathMMU, a multimodal expert-level benchmark designed to evaluate LMMs in pathology image understanding and reasoning, comprising 33,428 Q&As along with 24,067 pathology images. As depicted in Fig. 1, PathMMU draws from a diverse range of high-quality sources, including PubMed scientific papers, pathology atlas from guidelines, Twitter posts by pathology experts, commonly used pathology classification datasets and educational content from YouTube videos, covering multiple organ systems (e.g., gastrointestinal, pulmonary, etc.) and multiple subjects (e.g., dermatopathology, hematopathology, etc.). In developing PathMMU, we adopt a meticulously designed cascading approach. Initially, we prompt GPT-4V with collected image_caption pairs to enhance the original captions, crafting descriptions that delve deeper into details and highlight crucial morphological features. Subsequently, we use these enhanced descriptions along with the images to prompt GPT-4V to generate professional multi-choice, multimodal pathology Q&As, each accompanied by detailed explanations for their answers. To ensure the questions are specifically designed for multimodal pathology understanding, we employ a collaboration of multiple Large Language Models (LLMs) to eliminate questions that can be solved or guessed using text alone. Most importantly, we invite seven pathology experts to manually review approximately 12,000 Q&As from the test and validation sets of PathMMU. The above carefully executed procedures, ensure the generation of professional, logical, and high-quality Q&As. + +We evaluate the zero-shot performance of 18 advanced LMMs and their robustness on the PathMMU test sets. We also fine-tune two representative LMMs on the training set to assess their transfer learning capabilities. The key insights from this extensive analysis are: (1) Cutting-edge LMMs struggle with the PathMMU, with 15 out of the 18 achieving no more than $40\%$ accuracy. Even the most advanced model currently available, GPT-4V, only attains an accuracy of $49.8\%$ . This reveals a notable discrepancy of nearly $20\%$ compared to professional pathologists, indicating significant deficiencies of current LMMs in the specialized field of pathology and emphasizing the challenging nature of the PathMMU benchmark. (2) While these LMMs show considerable robustness in robustness tests, the full extent of their robustness is still in question. On the one hand, given their limited performance, the scope for further performance degradation is somewhat constrained. On the other hand, these models tend to capture only superficial image features, and sometimes may not even utilize image information for reasoning. As a result, even when image corruption obscures minor details, the impact on these models' performance might not be significantly pronounced. (3) The text-only performance of GPT-4 Turbo surpasses the top-performing open-source LMMs on the PathMMU benchmark. We observe that GPT-4 Turbo sometimes takes shortcuts to guess answers. This involves choosing the correct answer based + +either on the options most commonly encountered in real-world pathological scenarios, or by identifying the most distinctive option, rather than conducting an in-depth analysis of the pathology image. (4) Fine-tuning general-purpose LMMs with substantial amounts of data significantly boosts their ability to comprehend and reason with pathological images, enabling them to easily surpass the zero-shot performance of GPT-4V. However, there still remains a discernible gap between their performance and that of human experts. + +# 2 Related Works + +Multimodal Models. The emergence of powerful large language models such as BERT [12], GPT-3 [8], T5 [38], LLaMa [46], ChatGPT [33], GPT-4 [34], and Vicuna [10] has significantly advanced the field of Natural Language Processing (NLP). These models demonstrate exceptional general capabilities, motivating researchers to explore the integration of LLMs with vision models to develop versatile multimodal models. This integration is primarily achieved through the pretraining approach, leading to the creation of groundbreaking LMMs. Representative models of this approach include CLIP [37], Flamingo [1], BLIP-2 [29], and Fuyu [5], all of which demonstrate remarkable multimodal understanding abilities. Furthermore, by adapting instruction-tuning techniques from the NLP domain, a variety of multimodal instruction-tuning datasets have been developed. This initiative aims to guide LMMs in generating outputs that are more controllable, practical, and adaptable to various tasks. Simultaneously, this approach facilitates the integration of LLMs into LMMs using lightweight fine-tuning, significantly reducing both costs and time. Key models that embody this advancement include GPT-4V [35], Gemini Pro Vision [45], Qwen-VL [4], LLaVA [30], MiniGPT-4 [60], BLIP-2 [29] and InstructBLIP [11]. + +LMM Benchmarks. In the general domain, the rise of various large models has led to the creation of extensive benchmark datasets, designed to evaluate the universal capabilities of these models across a range of tasks and domains. Notable examples include LAMM [52], LVLM-eHub [51], SEED [26], MMBench [31], MMVet [53] and BenchLMM [9], which have been used to assess the basic perception abilities of large models. More recently, MMMU [54], a massive and challenging dataset covering 30 university-level subjects, has been developed. It is specifically designed to evaluate the general multimodal understanding and reasoning capabilities of LMMs. Furthermore, HaELM [48] and HallusionBench [16] are proposed to evaluate the degree of hallucination in LMMs. + +In the medical field, datasets such as VQA-RAD [24] and VQA-Med [6] provide a collection of Q&A pairs based on radiology images, facilitating the evaluation of the capabilities of AI in radiological diagnostics. Additionally, PMC-VQA [55] is a large-scale dataset within this domain. It generates a vast number of Q&A pairs by prompting ChatGPT with text-only image captions from PubMed. In the pathology domain, limited dataset construction efforts have been performed to date. The representative large-scale dataset available, PathVQA [18], is constructed in a manner similar to PMC-VQA. It utilizes + +![](images/1c94cd14cf7507c2335dd8ab8c118fad25088a4fc7f6416c15ae7aad867fb3c4.jpg) +Fig. 2: The comparison between PathMMU and existing benchmarks. The Q&A pairs in PathMMU are sourced extensively and comprehensively, undergoing rigorous multi-tiered filtering. This includes the initial filtering by multiple LLMs and the strict reviews by professional pathologists. Additionally, each question is accompanied by a detailed explanation. These attributes establish PathMMU as the most professionally curated, comprehensive, and highest-quality large-scale pathology dataset available. + +
DatasetDomainSources# Q&As / ImagesAnswer Explanation?LLM Filtered?Expert Annotated?
VQA-MedMedicalMedPix5000 / 5000XX✓
VQA-RADMedicalMedPix3515 / 315XX✓
PMC-VQAMedicalPMC227k / 149kXXX
PathVQAPathologyBook32799 / 4998XXX
Quilt-VQAPathologyEdu-Content1283 / 985XX✓
Quilt-Red
Social Media
PathMMUPathologyEdu-Content, PubMed33428 / 24067✓✓✓
Atlas, CLS-Data
+ +a heuristic approach to generate questions based on the text-only captions of pathology textbook images, yielding over 30,000 Q&As and 4,998 images. More recently, Quilt-VQA [39] is proposed, which is constructed by automatically extracting question and answer pairs from the narrations of professional teachers in YouTube videos. This approach guarantees the professionalism of the Q&A pairs, although it comprises a relatively smaller dataset with 1,283 Q&As and 985 images. Nonetheless, datasets reliant solely on image captions for question generation, face several notable limitations: (1) The creation of questions typically requires the consideration of both the image and its accompanying caption. Relying only on captions can introduce inaccuracies, making some questions unanswerable or solvable using only the text, bypassing the need for image analysis. (2) Since captions tend to be simplistic and may not capture the detailed information evident in images, generating an excessive number of Q&As from the same image's caption can lead to either too simplistic or insufficient questions. (3) Automatically generated questions, whether through predefined rules or by prompting ChatGPT, are susceptible to containing numerous errors. + +Given the current lack of high-quality benchmarks in pathology, we develop PathMMU, a large-scale, high-quality, and expert-verified pathology benchmark, to fill this gap. Our approach involves generating Q&As from a diverse range of image-caption pairs, followed by rigorous expert review and filtering applied to both the images and the questions. Moreover, we provide detailed explanations for each answer as a reference to enhance the interpretability of answers, which is absent in previous works. As shown in Fig. 2, PathMMU distinguishes itself in the pathology domain through its professionalism and superior quality. + +# 3 The Proposed PathMMU Benchmark + +# 3.1 Design Principle + +PathMMU is designed to provide the community with a specialized dataset for evaluating pathology LMMs. We adhere to five principal guidelines in its devel- + +![](images/61f327ce61ac000781b5bac66790e28ac7c27c92b9a49089e792b123e8757cbd.jpg) +Fig. 3: An illustrative overview of the three main processes in PathMMU Q&A generation: data collection and preprocessing, detailed pathology image description generation, and question generation with LLMs filtering and expert validation. + +![](images/cd66e1d8d36649ca58e541ac02307491fba018b121bd091fdd88fa46466fa080.jpg) + +![](images/54244a918b274a764ccc3b2174f41c816e068bec2f476862e438d7de41d3fd3e.jpg) + +opment: (1) Comprehensive and Specialized Data: Our benchmark is gathered from a diverse range of authoritative materials, including respected scientific publications from PubMed, pathology atlas from textbooks and guidelines, educational videos on YouTube, pathologist-shared images with explanations on Twitter, and widely recognized pathology classification datasets. Additionally, we involve seven human experts to manually review the dataset to ensure it meets professional standards. (2) Effective and Valuable Questions: The questions in PathMMU are carefully formulated to necessitate answers based on detailed observation of pathology images, rather than being solvable through mere textual interpretation or guesswork. We ensure that these questions are answerable, effective, and aligned with the standards of professional pathology examinations. (3) Large-Scale: We present the largest pathology dataset currently available. This scale enables researchers to thoroughly explore the full potential of LMMs in the field of pathology. (4) High-Quality Images: We prioritize clarity in our images to ensure that every detail is discernible, as evidenced by our dataset's average resolution of approximately $900 \times 700$ pixels. (5) Explainable Answer Choices: In line with the ongoing interest in understanding the decision-making logic of large models, our dataset includes reference explanations for each answer choice. This feature enhances the interpretability of model responses, contributing to the broader research on model explainability. + +# 3.2 The Construction of PathMMU + +To guarantee the quality of our benchmark, we meticulously develop a three-step data processing and generation protocol. The overall data collection and generation process is illustrated in Fig. 3. + +Step 1: Data Collection and Preprocessing. Our data collection draws from a wide variety of sources, which we integrate as subsets into the PathMMU framework. We name these subsets and detail their collection workflows as follows. + +PubMed: This subset consists of scientific pathology image-text pairs sourced from scientific documents. We gather these pairs from the open-access section of PubMed Central. Given the significant presence of non-pathological data within this resource, we meticulously annotate 20,000 samples as either pathological or non-pathological, and subsequently train a ConvNeXt [32] model to identify pathology data within the remaining dataset. EduContent: This subset, derived from educational YouTube videos, is sourced from Quilt1M [20]. Since YouTube videos often include a variety of visuals unrelated to pathology, such as computer desktops and unrelated imagery, the initial phase involves the annotation of pathological regions in 3,000 YouTube video images. This critical step is followed by the training of a YOLOv6-based detector [28] to automate the identification of pathological regions within remaining YouTube content. Atlas: This subset is a compilation of authoritative pathology textbooks and guidelines, we transform them from PDFs to HTML to facilitate the extraction of image-caption pairs. SocialPath: This subset aggregates image-text pairs from Twitter posts by pathology experts, using Twitter URLs provided by OpenPath [19]. PathCLS: Encompassing widely recognized pathology classification datasets, including PatchCamelyon [47], CRC-100K [21], SICAPv2 [40], BACH [2], Osteo [3], SkinCancer [22], MHIST [50], WSSSLUAD [17], LC-Lung and LC-Colon [7]. We reformulate these datasets into a Q&A format and randomly sample from these datasets. We directly make the gathered samples from PathCLS a part of the PathMMU dataset without undergoing the following two steps Q&A generation process. Finally, we manually review and filter irrelevant or unclear images to ensure the quality of collected data. As a result, approximately 30,000 high-quality image-text pairs are obtained, forming the foundation of the PathMMU. + +Step 2: Detailed Description Generation and Question-answer Pairs Generation. When dealing with data sources from platforms like Twitter and YouTube, it's common to find a weak correlation between an image and its accompanying caption. For instance, in the SocialPath dataset, captions might only describe the aesthetic appeal of the image, which is irrelevant to pathological findings. Similarly, in other data sources, captions tend to describe only a fraction of the features present in the image. To tackle this issue, we prompt GPT-4V to focus on describing the morphology of cells and tissues. It is crucial to note that merely prompting GPT-4V to output image descriptions might lead to incoherent content. Therefore, we provide GPT-4V with original captions for reference, significantly reducing the incidence of such occurrences and thus ensuring the generation of more precise and relevant descriptions. + +Step 3: Question Generation and Expert Validation. Following the generation of image descriptions, GPT-4V is tasked with creating three questions per image, each accompanied by multiple-choice options, the correct answer, and detailed explanations for the answer. The creation of these explanations serves a dual purpose: it not only offers valuable insights into the model reasoning process, but also facilitates the subsequent manual review phase. + +It's crucial to note that despite the careful design of question generation prompts for GPT-4V, the text-only GPT-4 is capable of correctly answering + +over $60\%$ of these questions through educated guesses, bypassing the need for visual cues. GPT-4 explains that its deductions are drawn from patterns such as one option being more common in typical pathological scenarios or exhibiting noticeable differences from other options. We delve deeper into this phenomenon in Sec. 4.3 to offer a more thorough experimental analysis. To ensure that LMMs do not rely on educated guesses, thereby obscuring their true multimodal capabilities, we employ GPT-3.5 Turbo, GPT-4 Turbo, Gemini Pro, and ERNIE-Bot-4 to perform educated guesses. Questions correctly guessed by at least three of these models are subsequently excluded. + +We partitioned the remaining dataset into training, validation, and test sets. The training set comprises 16,344 images and 23,041 Q&As, while the validation and test sets contain roughly 12,000 Q&As and over 8,000 images, respectively. For the validation and test sets, we invite seven professional pathologists to conduct a thorough manual review. These pathologists initially attempt to answer the questions independently and then assess them based on the following criteria: (1) Whether the question can be answered without an accompanying image. (2) Whether the answer can be inferred from the provided question and image. (3) Whether the supplied answer is incorrect, if there is no correct answer, or if multiple correct answers are possible. (4) Whether the generated question appears unusual or atypical compared to standard questions in pathology examinations. Questions failing to meet these standards are deemed invalid and consequently removed from the PathMMU dataset. + +Ultimately, after thorough expert review, we obtain 710 Q&As accompanied by 510 images for the validation set, and 9,677 Q&As with 7,213 images for the test set. Additionally, to establish a standard for expert performance, we extract a smaller subset from the test set, named 'test-tiny', which includes 1,156 Q&As. We invite two groups of professional pathologists to participate in this subset as an examination, and we average their performance to set a benchmark for expert performance in the PathMMU, serving as a comparison reference for LMMs. + +# 4 Experiments + +In this section, we conduct extensive experiments on PathMMU. Initially, we assess the zero-shot performance of cutting-edge LMMs. To verify whether LMMs effectively utilize the visual information from pathology images, we also evaluate the performance of text-only LLMs on PathMMU as a reference. Additionally, we apply common corruptions to the images in the test-tiny set to test the robustness of the LMMs against image corruptions. Furthermore, we select representative LMMs and evaluate their performance to explore their transfer-learning capabilities. Finally, we conduct experimental analysis on the issue of LLMs being able to guess answers identified during the data construction process. + +# 4.1 Zero-shot Evaluation of LMMs and LLMs + +In this study, we evaluate the zero-shot capabilities of the latest and most advanced LMMs on the PathMMU. Specifically, we use the PathMMU validation + +Table 1: Overall results of models on the PathMMU validation and test set. Besides reporting the performance of LMMs, we add text-only LLM baselines that purely accept text as inputs. The best-performing LMM in each subset is in- bold, and the top-performing LLM is underlined. + +
Validation OverallTestOverallPubMedSocialPathEduContentAtlasPathCLS
- (710)Tiny (1156)ALL (9677)Tiny (281)ALL (3068)Tiny (235)All (1855)Tiny (255)All (1938)Tiny (208)ALL (1007)Tiny (177)ALL (1809)
Random Choice24.622.123.722.125.125.526.525.526.019.723.015.316.3
Frequent Choice27.527.725.528.826.127.726.729.826.528.427.522.021.0
Expert performance-71.8-72.9-71.5-69.0-68.3-78.9-
Large Multimodal Models (LMMs): Text + Image as Input
OpenFlamingo2-9B [1]26.224.124.224.925.525.525.319.624.428.825.522.020.4
Kosmos2 [36]26.926.124.927.825.827.224.325.125.923.124.227.122.8
LLaMA-Adapter2-7B [14]26.226.626.426.028.327.726.527.128.030.827.420.920.5
MiniGPT4-Vicuna-13B [60]27.225.527.728.830.124.327.225.529.327.928.619.221.8
Otter [25]26.129.328.134.930.226.028.430.630.029.328.223.221.3
Fuyu-8B [5]27.530.129.235.930.528.930.730.629.827.930.624.323.9
CogVLM [49]29.930.629.732.032.131.130.430.629.935.133.722.622.4
Qwen-VL-7B [4]29.432.131.534.932.933.635.834.133.635.633.818.621.7
BLIP-2 FLAN-T5-XL [29]32.832.631.835.934.634.933.633.732.834.637.420.320.8
BLIP-2 FLAN-T5-XXL [29]33.433.333.537.037.435.734.630.234.539.440.719.820.6
InstructBLIP FLAN-T5-XL [11]33.134.931.837.033.235.333.936.533.337.036.726.623.1
InstructBLIP FLAN-T5-XXL [11]32.734.333.939.137.233.634.334.536.038.539.322.622.7
LLaVA-1.5-7B [30]36.634.935.441.639.937.938.132.536.535.639.223.221.9
LLaVA-1.5-13B [30]37.938.837.644.541.040.440.434.139.447.144.324.923.5
Qwen-VL-PLUS [4]38.039.334.343.537.741.336.039.636.044.737.123.223.3
Qwen-VL-MAX [4]43.649.245.953.050.953.649.352.247.951.449.830.529.6
Gemini Pro Vision [45]41.942.842.743.844.942.442.043.543.749.549.432.834.7
GPT-4V-1106 [35]49.353.949.859.453.558.753.960.453.648.152.836.233.8
Large Language Models (LLMs): Only Text as Input
ERNIE-Bot 4.0 [41]31.834.330.937.031.233.632.940.034.536.137.420.920.6
Gemini Pro [45]31.031.631.031.333.531.931.433.332.738.037.721.520.9
Vicuna-v1.5-13B [10]32.031.231.635.233.830.634.331.033.534.635.722.020.7
GPT-3.5 Turbo [33]30.429.428.432.731.231.530.131.030.126.926.822.020.8
GPT-4 Turbo [34]36.541.838.148.842.443.442.347.140.641.343.222.321.1
+ +set to conduct prompt engineering for these models, and then test their performance on the test set. The tested models include 14 open-source LMMs: OpenFlamingo [1], Kosmos2 [36], LLaMA-Adapter2 [14], MiniGPT-4 [60], Otter [25], Fuyu [5], CogVLM [49], Qwen-VL [4], BLIP-2 [29], InstructBLIP [11], and LLaVA-1.5 [30], as well as 4 closed-source LMMs: Qwen-VL-PLUS, Qwen-VL-MAX [4], Gemini Pro Vision [45], and GPT-4V [35]. For BLIP-2, Instruct-BLIP, and LLaVA-1.5, we deploy various sizes of these models to examine the applicability of scaling rules in LMMs. Moreover, we extend our evaluation to purely text-based LLMs, including ERNIE-Bot 4.0 [41], Gemini Pro [45], Vicuna-v1.5-13B [10], GPT-3.5 Turbo [33], and GPT-4 Turbo [34]. This assessment aims to quantify their ability to infer correct answers through educated guesses without visual inputs, reflecting the extent to which LMMs integrate and leverage visual information compared to their text-only counterparts. + +Results and Discussion. The advanced LMMs struggle with the PathMMU dataset. As shown in Tab. 1, among the 18 models tested, 15 show an accuracy below $40\%$ . Notably, the highest-performing open-source model, LLaVa-1.5-13B, and the top-performing closed-source model, GPT-4V, achieve + +only $37.6\%$ and $49.8\%$ accuracy, respectively. Those are significantly lower than human expert performance, which stands at $71.8\%$ , underscoring a substantial gap between current LMMs and professional pathologist standards. This gap indicates that the practical application of LMMs in real-world pathological scenarios remains significantly constrained. + +Closed-source LLMs achieve performance that is on par with or even surpasses that of the most advanced open-source LMMs. When considering text-only models, we prompt them to make educated guesses without providing the image. We observe that GPT-4 Turbo, Vicuna-v1.5-13B, GeminiPro, and ERNIE-Bot 4.0 outperform nearly half of the open-source LMMs. Notably, GPT-4 Turbo leads with a significant margin, achieving an accuracy rate of $38.1\%$ , exceeding the second-best LLM by $6.5\%$ , and even outperforming the best-performing open-source LMM, LLaVA-1.5-13B. The $38.1\%$ accuracy of GPT-4 Turbo, on the one hand, demonstrates that under our strict filtering criteria, even the best LLM, when provided with text alone, makes incorrect answers for the majority of questions. This underscores the effectiveness of our dataset in testing the capabilities of LMMs in pathological image analysis. On the other hand, it reaffirms the analysis in Sec. 3.2 that GPT-4 possesses strong logical reasoning abilities, enabling it to correctly guess a portion of the questions, significantly outperforming random selection. However, this raises critical concerns in pathology diagnosis: the reliability of model outputs in clinical scenarios remains questionable, especially when they rely on educated guesses, particularly probabilistic ones, rather than comprehensive analysis. This emphasizes the urgent necessity to develop interpretable AI models and advocate for their cautious application in pathology diagnostics. + +The larger LMMs exhibit better performance. Specifically, LLaVa-1.5-13B exceeds LLaVa-1.5-7B model by $2.2\%$ in overall performance, while Instruct-BLIP FLAN-T5-XXL and BLIP-2 FLAN-T5 XXL surpass their smaller versions by $2.1\%$ and $1.7\%$ , respectively. This suggests that larger models typically possess stronger multimodal capabilities in the field of pathology. + +# 4.2 Robustness of LMMs to Image Corruption + +In practical pathology, the interpretation of models significantly influences the subsequent medical decisions and treatment strategies. Therefore, models with strong robustness are crucial for clinical applications. The quality of pathological slides can be affected by various factors during staining, scanning, and storage, including JPEG compression, pixelation, blur (e.g., bubble blur, defocus blur, and motion blur), and color variations (e.g., brightness, saturation, and hue). + +Drawing inspiration from the studies on the robustness of encoder-based models to pathology image corruptions [43, 56, 58], we incorporate these aforementioned types of corruptions into our analysis. To be specific, we simulate five levels of each corruption type on the PathMMU to explore the robustness of LMMs against these corruptions, as depicted in the left half of Fig. 4. + +Results and Discussion. LMMs demonstrate notable robustness to various types and levels of image corruptions, yet their true robustness + +![](images/59332a6b7df98c041fbee5fd8a42a4b8d1bd2ee1b6d02a92a0ee7d0f6c399e7a.jpg) +Fig. 4: Left: Illustration of corrupted pathology images. Right: LMM's performance across various levels of color-related (brightness, hue, saturation) and image quality-related (pixelation, JPEG compression, bubble blur, motion blur, defocus blur) corruptions on the PathMMU test-tiny set, with level 0 representing the uncorrupted images. + +![](images/f87f47b50d448db9d0665be2e553f9cba32e39ef9892b286011374a5501a8260.jpg) + +![](images/912bfc972cbd9be60ab82e937ed09f2153ffddd3ca5b3d6db9766466beaf853a.jpg) + +Table 2: Evaluation of the model's robustness on the PathMMU test-tiny set, where green indicates performance improvement and red signifies performance degradation compared to the model's performance with uncorrupted images. + +
BrightnessBubbleDefocusHueJPEGMotionPixelateSaturateOverall
Qwen-VL-7B32.7↑0.632.3↑0.232.1↓0.032.4↓0.332.3↑0.233.8↑1.733.1↑1.033.0↑0.932.7↑0.6
BLIP-2 FLAN-T5-XL32.9↑0.332.4↓0.232.8↑0.232.9↑0.332.4↓0.232.8↑0.232.6↑0.032.7↑0.132.7↓0.1
BLIP-2 FLAN-T5-XXL33.1↓0.233.2↓0.133.5↑0.233.7↑0.433.5↑0.233.1↓0.233.1↓0.233.6↑0.333.3↓0.0
InstructBLIP FLAN-T5-XL34.0↓0.934.1↓0.833.9↓1.034.1↓0.833.7↓1.234.0↓0.934.4↓0.533.8↓1.134.0↓0.9
InstructBLIP FLAN-T5-XXL34.0↓0.333.8↓0.534.2↓0.134.1↓0.234.3↓0.234.3↑0.034.1↓0.233.8↓0.534.1↓0.2
LLaVA-1.5-7B34.8↓0.135.3↑0.434.1↓0.834.3↓0.634.0↓0.934.3↓0.634.6↓0.334.7↓0.234.5↓0.4
LLaVA-1.5-13B38.9↑0.138.3↓0.538.3↓0.538.4↓0.438.0↓0.838.6↓0.238.7↓0.139.1↑0.338.5↓0.3
+ +Table 3: Results of LMMs on the PathMMU test set with the original images substituted by random Gaussian noise images. + +
OverallPubMedSocialPathEduContentAtlasPathCLS
TinyALLTinyALLTinyAllTinyAllTinyALLTinyALL
Qwen-VL-7B28.8↓3.328.3↓3.239.5↑4.630.4↓2.524.7↓8.930.5↓5.327.8↓6.328.9↓4.727.4↓8.230.1↓3.720.3↑1.720.8↓0.9
BLIP-FLAN-T5-XL30.9↓1.730.8↓1.030.6↓5.333.1↓1.534.5↑0.433.4↓0.232.2↑1.531.9↓0.934.1↓0.535.1↓2.320.9↑0.620.8↓0.0
BLIP-FLAN-T5-XXL32.4↓0.931.6↓1.934.5↓2.534.6↓2.834.0↑1.732.5↓2.130.6↑0.432.8↑1.738.5↓0.938.1↓2.622.2↑2.420.8↑0.2
InstructBLIP FLAN-T5-XL31.1↓3.829.7↓2.132.4↓4.630.8↓2.432.8↓2.532.0↓1.933.7↓2.832.4↓0.932.7↓4.330.4↓6.321.5↓5.122.4↓0.7
InstructBLIP FLAN-T5-XXL30.4↓3.930.8↓3.132.0↓7.133.5↓3.728.9↑4.731.9↓2.432.2↑2.331.7↓4.335.1↑3.435.2↓4.121.5↓1.121.7↓1.0
LLAVA-1.5-7B30.2↓4.730.9↓4.537.0↓4.633.9↓6.031.5↓6.432.5↓5.628.2↓4.332.7↓3.829.8↓5.833.9↓5.320.9↓2.320.8↓1.1
LLAVA-1.5-13B32.7↓6.133.2↓4.438.4↓6.136.0↓5.035.3↓5.136.3↓4.129.0↓5.134.1↓5.335.6↓11.538.8↓5.522.0↓2.921.0↓2.5
+ +is questionable, as shown in Tab. 2 and the right half of Fig. 4. Notably, the Qwen-VL-7B even shows a $0.6\%$ overall performance increase under corruption compared to its baseline. It is more plausible to hypothesize that these corruptions mainly alter minute details of the pathology image that are difficult for LMMs to discern (such as chromatin morphology of the nucleus or vacuolization of the cytoplasm). This limitation stems from their pre-training in general domains, where they learn to recognize larger and more prominent features (e.g., humans, houses, cars, etc.) rather than the nuanced details that are vital in pathology. Furthermore, they may even resort to exploiting spurious correlations as shortcuts to answer questions (e.g., deducing answers from textual patterns instead of focusing on essential image details.) + +To substantiate our argument, we employ an extreme test of corruption by replacing the images with Gaussian noise-generated random images. As shown in Tab. 3, we discover that even when the images contain no relevant information, LMMs can still achieve results significantly better than random choice. The drop in performance when using these random images ranged from only $1.0\%$ to $4.5\%$ . This finding suggests that LMMs might rely on shortcuts to accomplish the multimodal task, such as depending exclusively on textual information to make predictions. Additionally, it is intriguing that the performance drop across different sizes of the same model is remarkably similar. In other words, the benefit that images provide to models of various model sizes appears to be consistent, suggesting that the improvement in performance for different-sized models may primarily stem from their LLM component, rather than the vision aspect. + +# 4.3 Analysis of LLMs' Ability to Guess Answers + +In this section, we delve deeper into the phenomenon of LLM's ability to make educated guesses with more comprehensive experiments and analysis. + +To more clearly demonstrate the guessing abilities of LLMs, we randomly select 100 samples that are filtered during the Q&A generation process and can be correctly guessed by multiple LLMs. We invite pathology experts to answer these questions with reference images. As shown in Fig. 5, despite having access to images, their performance is significantly lower than that of the closed-source LLMs, which achieve about $90\%$ accuracy, and only slightly better than the smaller open-source Vicuna-v1.5-13B. This indicates that LLMs may outperform humans in guessing correct answers by identifying shortcuts within the questions. + +We hypothesize several key reasons for the guessing behavior exhibited by LLMs: (1) Frequency-based guessing: LLMs may guess based on the prevalence of certain options in real-world pathological instances. (2) Based on the options that present a pattern of one supporting and three contradicting, or vice versa: For example, when a question identifies a tumour diagnosis, one option may match the pathological traits of a lesion, while the others suggest non-lesional characteristics. (3) When only one option aligns with the question's subject: For example, if a question asks which feature in the image supports the diagnosis of a Low-grade Squamous Intraepithelial Lesion (LSIL) cell, with only one option describes LSIL's pathological characteristics. + +To support our hypothesis, we expand the sample size to 1,600, with 400 from each source (except PathCLS), to empirically analyze how LLMs guess answers. Specifically, we swap the questions among these samples while keeping the options unchanged, creating samples where the questions and options are completely mismatched. As shown in the right half of Fig. 5, we observe that LLMs still manage to guess approximately $50\%$ cases correctly, significantly higher than random choice accuracy. This finding suggests that the models tend to select the most common or prominent option as the answer, supporting our hypotheses (1) and (2). To further explore hypothesis (3), we design experiments using BERT-large [12] and BiomedBERT-large [15], the latter being specifically pre-trained on biomedical data. Given that the BERT series incorporates Next-Sentence + +![](images/75c52ac8bc697d5ee19715be4da0962f7a25300525ce2189db23a99c036a280c.jpg) +Fig. 5: Left: The performance comparison between different LLMs and human experts on 100 filtered samples where the answer can be guessed through text-only. Right: Expand the sample size to 1600 to validate the source of LLMs' ability to guess answers, which includes: (1) randomly replacing the original questions with others from the dataset while keeping the options unchanged, and (2) utilizing the BERT series for answer selection, specifically through its Next Sentence Prediction (NSP), to assess whether an option is the sequential sentence following a question. + +
PubMed SocialPath EduContentAtlas
Random Choice25.824.325.525.825.3
GPT-4 Turbo93.091.886.883.888.8
+ Random Swap Question44.345.550.051.547.8
GPT-3.5 Turbo84.386.579.585.383.9
+ Random Swap Question44.044.553.851.348.4
Gemini Pro85.590.176.571.584.1
+ Random Swap Question43.844.850.351.347.5
ERNIE-Bot 4.094.592.389.382.089.5
+ Random Swap Question47.046.556.551.050.3
BERT-large w/ NSP24.536.037.534.333.1
BiomedBERT-large w/ NSP44.344.842.337.542.3
+ +Prediction (NSP) in their pretraining, which inherently relies on the similarity between two sentences, we apply this to predict the relationship between questions and options. We select the option with the highest probability of being the correct next sentence to the question. Our findings reveal that both BERT and BiomedBERT substantially surpass random guessing, indicating that direct matching of questions and options is a viable method for models to guess the correct answers, thereby supporting hypothesis (3). Moreover, BiomedBERT-large demonstrates a notable performance enhancement compared to BERT-large. This suggests that pretraining on biomedical data equips the model with a broader understanding of pathological knowledge. + +# 4.4 Fine-tuning Results + +To explore the adaptability of LMMs to the pathology domain, we select representative LMMs, InstructBLIP FLAN-T5-XL and InstructBLIP FLAN-T5-XXL for fine-tuning experiments. Our experiments consist of two parts: (1) training the LMMs to directly generate answers, and (2) fine-tuning the LMMs to generate a reasoning process before delivering the final answer. + +Results and Discussion. All models exhibit significant improvements on the PathMMU test set after fine-tuning on its training set, as detailed in Tab. 4. Notably, during the fine-tuning for direct answer generation, InstructBLIP FLAN-T5-XL and InstructBLIP FLAN-T5-XXL achieve significant improvements of $21.5\%$ and $21.3\%$ , respectively. This substantial improvement enables them to outperform the current leading model, GPT-4V. While still lagging behind expert performance, these results demonstrate a trend toward approaching expert-level proficiency. This also reflects the effectiveness of PathMMU in enhancing the LMMs' abilities for pathology image analysis. + +Unexpectedly, generating explanations before answers during fine-tuning does not yield improvements. Instead, we observe slight performance decreases of $0.3\%$ + +Table 4: Results of LMMs on the PathMMU test set after the fine-tuning. The 'w/ A' and 'w/ A&E' denote the model is fine-tuned to output the answer directly or to output the answer with an explanation for its answer. + +
OverallPubMedSocialPathEduContentAtlasPathCLS
TinyALLTinyALLTinyALLTinyAllTinyALLTinyALL
InstructBLIP FLAN-T5-XL34.931.837.033.235.333.936.533.337.036.726.623.1
+ fine-tune w/A55.7 ± 20.853.3 ± 21.553.7 ± 16.750.4 ± 17.252.8 ± 17.552.3 ± 18.456.5 ± 20.052.0 ± 18.756.2 ± 19.253.1 ± 16.461.0 ± 34.460.8 ± 37.7
+ fine-tune w/A&E54.9 ± 20.053.0 ± 21.255.5 ± 18.551.4 ± 18.251.9 ± 16.649.7 ± 15.855.3 ± 18.851.5 ± 18.253.4 ± 16.451.6 ± 14.959.3 ± 32.761.4 ± 38.3
InstructBLIP FLAN-T5-XXL34.333.939.137.233.634.334.536.038.539.322.622.7
+ fine-tune w/A56.8 ± 22.555.2 ± 21.355.2 ± 16.151.5 ± 14.359.6 ± 26.055.2 ± 20.958.4 ± 23.954.1 ± 18.150.5 ± 12.053.7 ± 14.461.0 ± 38.463.7 ± 41.0
+ fine-tune w/A&E51.0 ± 16.752.9 ± 19.048.8 ± 9.750.8 ± 13.655.3 ± 21.751.2 ± 16.952.2 ± 17.751.9 ± 15.944.2 ± 5.750.1 ± 10.854.8 ± 32.260.5 ± 37.8
+ +and $2.3\%$ compared to direct answer generation, respectively. We speculate that generating explanations is relatively more challenging, and incorporating it with answer generation might impede the models' capacity to generate correct answers. This finding raises an important question: how can we effectively leverage the interpretability information within PathMMU to enhance models' training? + +# 5 Conclusion + +In this study, we introduce PathMMU, the largest and highest-quality pathology benchmark to date, specifically crafted to evaluate the capabilities of LMMs in interpreting and reasoning with pathology images. The construction of PathMMU involves a meticulous data collection and curation process, supplemented by a strict manual review by seven professional pathologists to ensure its quality and professionalism. Moreover, we establish a human expert performance benchmark to quantify the gap between cutting-edge LMMs and human experts. Our experimental results reveal that advanced LMMs significantly lag behind on PathMMU, with these models demonstrating poor performance in observing details in pathology images and sometimes even neglecting visual information, highlighting a substantial gap in practical pathology application. However, LMMs demonstrate notable performance improvements after fine-tuning on PathMMU, even surpassing GPT-4V. While they do not achieve human expert-level performance, these LMMs show promising potential for analyzing pathology images. + +Future Directions for Pathology LMMs: Our experience with PathMMU highlights key areas for development in pathology LMMs: (1) Current LMMs, which are primarily based on LLMs and fine-tuned in a lightweight manner, tend to over-rely on textual information while neglecting visual data. There is a significant need to explore training methodologies or model structures that better integrate visual and textual modalities. (2) There's a tendency for LMMs to take shortcuts, solving problems in a "lazy" manner. This necessitates the development of trustworthy models for real-world clinical applications. (3) Given that most current LMMs do not support multi-image inputs, PathMMU does not include a benchmark for processing multiple images. However, in practical scenarios, pathologists typically analyze samples at various magnifications and perspectives, underscoring the importance of developing LMMs capable of handling multi-image inputs. Finally, we believe PathMMU will catalyze significant advancements in the development of next-generation LMMs in pathology. + +# Acknowledgements + +This study was partially supported by the National Natural Science Foundation of China (Grant No.92270108), Zhejiang Provincial Natural Science Foundation of China (Grant No.XHD23F0201), the Research Center for Industries of the Future (RCIF) at Westlake University, and the Westlake Education Foundation. + +# References + +1. Alayrac, J.B., Donahue, J., Luc, P., Miech, A., Barr, I., Hasson, Y., Lenc, K., Mensch, A., Millican, K., Reynolds, M., et al.: Flamingo: a visual language model for few-shot learning. In: NeurIPS. pp. 23716-23736 (2022) +2. Aresta, G., Araújo, T., Kwok, S., Chennamsetty, S.S., Safwan, M., Alex, V., Marami, B., Prastawa, M., Chan, M., Donovan, M., et al.: Bach: Grand challenge on breast cancer histology images. Medical image analysis 56, 122-139 (2019) +3. Arunachalam, H.B., Mishra, R., Daescu, O., Cederberg, K., Rakheja, D., Sengupta, A., Leonard, D., Hallac, R., Leavey, P.: Viable and necrotic tumor assessment from whole slide images of osteosarcoma using machine-learning and deep-learning models. PloS one 14(4), e0210706 (2019) +4. Bai, J., Bai, S., Yang, S., Wang, S., Tan, S., Wang, P., Lin, J., Zhou, C., Zhou, J.: Qwen-vl: A frontier large vision-language model with versatile abilities. arXiv preprint arXiv:2308.12966 (2023) +5. Bavishi, R., Elsen, E., Hawthorne, C., Nye, M., Odena, A., Somani, A., Taşırlar, S.: Introducing our multimodal models (2023), https://www.adept.ai/blog/fuyu-8b +6. Ben Abacha, A., Sarrouti, M., Demner-Fushman, D., Hasan, S.A., Müller, H.: Overview of the vqa-med task at imagecef 2021: Visual question answering and generation in the medical domain. In: Proceedings of the CLEF 2021 Conference and Labs of the Evaluation Forum-working notes (2021) +7. Borkowski, A.A., Bui, M.M., Thomas, L.B., Wilson, C.P., DeLand, L.A., Mastorides, S.M.: Lung and colon cancer histopathological image dataset (lc25000). arXiv preprint arXiv:1912.12142 (2019) +8. Brown, T., Mann, B., Ryder, N., Subbiah, M., Kaplan, J.D., Dhariwal, P., Neelakantan, A., Shyam, P., Sastry, G., Askell, A., et al.: Language models are few-shot learners. In: NeurIPS. pp. 1877-1901 (2020) +9. Cai, R., Song, Z., Guan, D., Chen, Z., Luo, X., Yi, C., Kot, A.: Benchlmm: Benchmarking cross-style visual capability of large multimodal models. arXiv preprint arXiv:2312.02896 (2023) +10. Chiang, W.L., Li, Z., Lin, Z., Sheng, Y., Wu, Z., Zhang, H., Zheng, L., Zhuang, S., Zhuang, Y., Gonzalez, J.E., Stoica, I., Xing, E.P.: Vicuna: An open-source chatbot impressing gpt-4 with $90\%$ chatgpt quality (March 2023), https://lmsys.org/blog/2023-03-30-vicuna/ +11. Dai, W., Li, J., Li, D., Tiong, A.M.H., Zhao, J., Wang, W., Li, B., Fung, P., Hoi, S.: Instructclip: Towards general-purpose vision-language models with instruction tuning. arXiv preprint arXiv:2305.06500 (2023) +12. Devlin, J., Chang, M.W., Lee, K., Toutanova, K.: Bert: Pre-training of deep bidirectional transformers for language understanding. In: NAACL. pp. 4171-4186 (2019) +13. Driess, D., Xia, F., Sajjadi, M.S., Lynch, C., Chowdhery, A., Ichter, B., Wahid, A., Tompson, J., Vuong, Q., Yu, T., et al.: Palm-e: An embodied multimodal language model. In: ICML. pp. 8469-8488 (2023) + +14. Gao, P., Han, J., Zhang, R., Lin, Z., Geng, S., Zhou, A., Zhang, W., Lu, P., He, C., Yue, X., et al.: Llama-adapter v2: Parameter-efficient visual instruction model. arXiv preprint arXiv:2304.15010 (2023) +15. Gu, Y., Tinn, R., Cheng, H., Lucas, M., Usuyama, N., Liu, X., Naumann, T., Gao, J., Poon, H.: Domain-specific language model pretraining for biomedical natural language processing. ACM Transactions on Computing for Healthcare (HEALTH) 3(1), 1-23 (2021) +16. Guan, T., Liu, F., Wu, X., Xian, R., Li, Z., Liu, X., Wang, X., Chen, L., Huang, F., Yacoob, Y., et al.: Hallusionbench: an advanced diagnostic suite for entangled language hallucination and visual illusion in large vision-language models. In: CVPR. pp. 14375-14385 (2024) +17. Han, C., Pan, X., Yan, L., Lin, H., Li, B., Yao, S., Lv, S., Shi, Z., Mai, J., Lin, J., et al.: Wss54luad: Grand challenge on weakly-supervised tissue semantic segmentation for lung adenocarcinoma. arXiv preprint arXiv:2204.06455 (2022) +18. He, X., Zhang, Y., Mou, L., Xing, E., Xie, P.: Pathvqa: 30000+ questions for medical visual question answering. arXiv preprint arXiv:2003.10286 (2020) +19. Huang, Z., Bianchi, F., Yuksekgonul, M., Montine, T.J., Zou, J.: A visual-language foundation model for pathology image analysis using medical twitter. Nature medicine 29(9), 2307-2316 (2023) +20. Ikezogwo, W., Seyfioglu, S., Ghezloo, F., Geva, D., Sheikh Mohammed, F., Anand, P.K., Krishna, R., Shapiro, L.: Quilt-1m: One million image-text pairs for histopathology. In: NeurIPS. pp. 37995–38017 (2023) +21. Kather, J.N., Halama, N., Marx, A.: 100,000 histological images of human colorectal cancer and healthy tissue. Zenodo10 5281 (2018) +22. Kriegsmann, K., Lovers, F., Zgorzelski, C., Kriegsmann, J., Janssen, C., Meliss, R.R., Muley, T., Sack, U., Steinbuss, G., Kriegsmann, M.: Deep learning for the detection of anatomical tissue structures and neoplasms of the skin on scanned histopathological tissue sections. Frontiers in Oncology 12, 1022967 (2022) +23. Kumar, V., Abbas, A.K., Fausto, N., Aster, J.C.: Robbins and Cotran pathologic basis of disease, professional edition e-book. Elsevier health sciences (2014) +24. Lau, J.J., Gayen, S., Ben Abacha, A., Demner-Fushman, D.: A dataset of clinically generated visual questions and answers about radiology images. Scientific data 5(1), 1-10 (2018) +25. Li, B., Zhang, Y., Chen, L., Wang, J., Yang, J., Liu, Z.: Otter: A multi-modal model with in-context instruction tuning. arXiv preprint arXiv:2305.03726 (2023) +26. Li, B., Wang, R., Wang, G., Ge, Y., Ge, Y., Shan, Y.: benseed-bench: Benchmarking multimodal llms with generative comprehension. arXiv preprint arXiv:2307.16125 (2023) +27. Li, C., Wong, C., Zhang, S., Usuyama, N., Liu, H., Yang, J., Naumann, T., Poon, H., Gao, J.: Llava-med: Training a large language-and-vision assistant for biomedicine in one day. In: NeurIPS. pp. 28541-28564 (2023) +28. Li, C., Li, L., Jiang, H., Weng, K., Geng, Y., Li, L., Ke, Z., Li, Q., Cheng, M., Nie, W., et al.: Yolov6: A single-stage object detection framework for industrial applications. arXiv preprint arXiv:2209.02976 (2022) +29. Li, J., Li, D., Savarese, S., Hoi, S.: Blip-2: Bootstrapping language-image pretraining with frozen image encoders and large language models. In: ICML. pp. 19730-19742 (2023) +30. Liu, H., Li, C., Wu, Q., Lee, Y.J.: Visual instruction tuning. In: NeurIPS. pp. 34892-34916 (2023) + +31. Liu, Y., Duan, H., Zhang, Y., Li, B., Zhang, S., Zhao, W., Yuan, Y., Wang, J., He, C., Liu, Z., et al.: Mmbench: Is your multi-modal model an all-around player? arXiv preprint arXiv:2307.06281 (2023) +32. Liu, Z., Mao, H., Wu, C.Y., Feichtenhofer, C., Darrell, T., Xie, S.: A convnet for the 2020s. In: CVPR. pp. 11976-11986 (2022) +33. OpenAI: Introducing chatgpt. https://openai.com/blog/chatgpt (2022) +34. OpenAI: Gpt-4 technical report (2023) +35. OpenAI: Gpt-4v(ision) system card. https://cdn.openai.com/papers/GPTV_System/Card.pdf (2023) +36. Peng, Z., Wang, W., Dong, L., Hao, Y., Huang, S., Ma, S., Wei, F.: Kosmos-2: Grounding multimodal large language models to the world. arXiv preprint arXiv:2306.14824 (2023) +37. Radford, A., Kim, J.W., Hallacy, C., Ramesh, A., Goh, G., Agarwal, S., Sastry, G., Askell, A., Mishkin, P., Clark, J., et al.: Learning transferable visual models from natural language supervision. In: ICML. pp. 8748-8763 (2021) +38. Raffel, C., Shazeer, N., Roberts, A., Lee, K., Narang, S., Matena, M., Zhou, Y., Li, W., Liu, P.J.: Exploring the limits of transfer learning with a unified text-to-text transformer. The Journal of Machine Learning Research 21(1), 5485-5551 (2020) +39. Seyfioglu, M.S., Ikezogwo, W.O., Ghezloo, F., Krishna, R., Shapiro, L.: Quilt-lava: Visual instruction tuning by extracting localized narratives from open-source histopathology videos. In: CVPR. pp. 13183-13192 (2024) +40. Silva-Rodriguez, J., Colomer, A., Sales, M.A., Molina, R., Naranjo, V.: Going deeper through the gleason scoring scale: An automatic end-to-end system for histology prostate grading and cribriform pattern detection. Computer methods and programs in biomedicine 195, 105637 (2020) +41. Sun, Y., Wang, S., Feng, S., Ding, S., Pang, C., Shang, J., Liu, J., Chen, X., Zhao, Y., Lu, Y., et al.: Ernie 3.0: Large-scale knowledge enhanced pre-training for language understanding and generation. arXiv preprint arXiv:2107.02137 (2021) +42. Sun, Y., Zhang, Y., Si, Y., Zhu, C., Shui, Z., Zhang, K., Li, J., Lyu, X., Lin, T., Yang, L.: Pathgen-1.6m: 1.6 million pathology image-text pairs generation through multi-agent collaboration (2024), https://arxiv.org/abs/2407.00203 +43. Sun, Y., Zhu, C., Zhang, Y., Li, H., Chen, P., Yang, L.: Assessing the robustness of deep learning-assisted pathological image analysis under practical variables of imaging system. In: ICASSP. pp. 1-5 (2023). https://doi.org/10.1109/ICASSP49357.2023.10095887 +44. Sun, Y., Zhu, C., Zheng, S., Zhang, K., Sun, L., Shui, Z., Zhang, Y., Li, H., Yang, L.: Pathasst: A generative foundation ai assistant towards artificial general intelligence of pathology. In: AAAI. pp. 5034-5042 (2024) +45. Team, G., Anil, R., Borgeaud, S., Wu, Y., Alayrac, J.B., Yu, J., Soricut, R., Schalkwyk, J., Dai, A.M., Hauth, A., et al.: Gemini: a family of highly capable multimodal models. arXiv preprint arXiv:2312.11805 (2023) +46. Touvron, H., Lavril, T., Izacard, G., Martinet, X., Lachaux, M.A., Lacroix, T., Rozière, B., Goyal, N., Hambro, E., Azhar, F., et al.: Llama: Open and efficient foundation language models. arXiv preprint arXiv:2302.13971 (2023) +47. Veeling, B.S., Linmans, J., Winkens, J., Cohen, T., Welling, M.: Rotation equivariant cnns for digital pathology. In: MICCAI. pp. 210-218. Springer (2018) +48. Wang, J., Zhou, Y., Xu, G., Shi, P., Zhao, C., Xu, H., Ye, Q., Yan, M., Zhang, J., Zhu, J., et al.: Evaluation and analysis of hallucination in large vision-language models. arXiv preprint arXiv:2308.15126 (2023) + +49. Wang, W., Lv, Q., Yu, W., Hong, W., Qi, J., Wang, Y., Ji, J., Yang, Z., Zhao, L., Song, X., et al.: Cogvlm: Visual expert for pretrained language models. arXiv preprint arXiv:2311.03079 (2023) +50. Wei, J., Suriawinata, A., Ren, B., Liu, X., Lisovsky, M., Vaickus, L., Brown, C., Baker, M., Tomita, N., Torresani, L., et al.: A petri dish for histopathology image analysis. In: Artificial Intelligence in Medicine: 19th International Conference on Artificial Intelligence in Medicine, AIME 2021, Virtual Event, June 15-18, 2021, Proceedings. pp. 11-24. Springer (2021) +51. Xu, P., Shao, W., Zhang, K., Gao, P., Liu, S., Lei, M., Meng, F., Huang, S., Qiao, Y., Luo, P.: Lvlm-ehub: A comprehensive evaluation benchmark for large vision-language models. arXiv preprint arXiv:2306.09265 (2023) +52. Yin, Z., Wang, J., Cao, J., Shi, Z., Liu, D., Li, M., Huang, X., Wang, Z., Sheng, L., BAI, L., Shao, J., Ouyang, W.: Lamm: Language-assisted multi-modal instruction-tuning dataset, framework, and benchmark. In: NeurIPS. pp. 26650-26685 (2023) +53. Yu, W., Yang, Z., Li, L., Wang, J., Lin, K., Liu, Z., Wang, X., Wang, L.: Mmvet: Evaluating large multimodal models for integrated capabilities. arXiv preprint arXiv:2308.02490 (2023) +54. Yue, X., Ni, Y., Zhang, K., Zheng, T., Liu, R., Zhang, G., Stevens, S., Jiang, D., Ren, W., Sun, Y., Wei, C., Yu, B., Yuan, R., Sun, R., Yin, M., Zheng, B., Yang, Z., Liu, Y., Huang, W., Sun, H., Su, Y., Chen, W.: Mmmu: A massive multi-discipline multimodal understanding and reasoning benchmark for expert agi. In: CVPR. pp. 9556-9567 (2024) +55. Zhang, X., Wu, C., Zhao, Z., Lin, W., Zhang, Y., Wang, Y., Xie, W.: Pmc-vqa: Visual instruction tuning for medical visual question answering. arXiv preprint arXiv:2305.10415 (2023) +56. Zhang, Y., Sun, Y., Li, H., Zheng, S., Zhu, C., Yang, L.: Benchmarking the robustness of deep neural networks to common corruptions in digital pathology. In: MICCAI. pp. 242-252 (2022) +57. Zhang, Z., Chen, P., McGough, M., Xing, F., Wang, C., Bui, M., Xie, Y., Sapkota, M., Cui, L., Dhillon, J., et al.: Pathologist-level interpretable whole-slide cancer diagnosis with deep learning. Nature Machine Intelligence 1(5), 236-245 (2019) +58. Zheng, S., Cui, X., Sun, Y., Li, J., Li, H., Zhang, Y., Chen, P., Jing, X., Ye, Z., Yang, L.: Benchmarking pathclip for pathology image analysis. Journal of Imaging Informatics in Medicine pp. 1-17 (2024) +59. Zhu, C., Sun, Y., Li, H., Cui, C., Zhang, S., Cai, J., Ling, Y.: Weakly supervised classification using multi-level instance-aware optimization on cervical cytologic image. In: 2022 IEEE 19th International Symposium on Biomedical Imaging (ISBI). pp. 1-5. IEEE (2022) +60. Zhu, D., Chen, J., Shen, X., Li, X., Elhoseiny, M.: Minigpt-4: Enhancing vision-language understanding with advanced large language models. arXiv preprint arXiv:2304.10592 (2023) \ No newline at end of file diff --git a/2024/PathMMU_ A Massive Multimodal Expert-Level Benchmark for Understanding and Reasoning in Pathology/images.zip b/2024/PathMMU_ A Massive Multimodal Expert-Level Benchmark for Understanding and Reasoning in Pathology/images.zip new file mode 100644 index 0000000000000000000000000000000000000000..073fcf6eae358fde3476aa97bb87753ccca41bdd --- /dev/null +++ b/2024/PathMMU_ A Massive Multimodal Expert-Level Benchmark for Understanding and Reasoning in Pathology/images.zip @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:0c7c894fd0dee20c3ef779cfc935b75d9bd342b45718a0e3b5838e0f695c5426 +size 586985 diff --git a/2024/PathMMU_ A Massive Multimodal Expert-Level Benchmark for Understanding and Reasoning in Pathology/layout.json b/2024/PathMMU_ A Massive Multimodal Expert-Level Benchmark for Understanding and Reasoning in Pathology/layout.json new file mode 100644 index 0000000000000000000000000000000000000000..e43cf1c6c4d0629200b31ad216fe87a2f9b2e653 --- /dev/null +++ b/2024/PathMMU_ A Massive Multimodal Expert-Level Benchmark for Understanding and Reasoning in Pathology/layout.json @@ -0,0 +1,7742 @@ +{ + "pdf_info": [ + { + "para_blocks": [ + { + "bbox": [ + 137, + 111, + 477, + 166 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 137, + 111, + 477, + 166 + ], + "spans": [ + { + "bbox": [ + 137, + 111, + 477, + 166 + ], + "type": "text", + "content": "PathMMU: A Massive Multimodal Expert-Level Benchmark for Understanding and Reasoning in Pathology" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 133, + 185, + 481, + 224 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 185, + 481, + 224 + ], + "spans": [ + { + "bbox": [ + 133, + 185, + 481, + 224 + ], + "type": "text", + "content": "Yuxuan Sun" + }, + { + "bbox": [ + 133, + 185, + 481, + 224 + ], + "type": "inline_equation", + "content": "^{1,2}" + }, + { + "bbox": [ + 133, + 185, + 481, + 224 + ], + "type": "text", + "content": ", Hao Wu" + }, + { + "bbox": [ + 133, + 185, + 481, + 224 + ], + "type": "inline_equation", + "content": "^{3}" + }, + { + "bbox": [ + 133, + 185, + 481, + 224 + ], + "type": "text", + "content": ", Chenglu Zhu" + }, + { + "bbox": [ + 133, + 185, + 481, + 224 + ], + "type": "inline_equation", + "content": "^{2}" + }, + { + "bbox": [ + 133, + 185, + 481, + 224 + ], + "type": "text", + "content": ", Sunyi Zheng" + }, + { + "bbox": [ + 133, + 185, + 481, + 224 + ], + "type": "inline_equation", + "content": "^{2}" + }, + { + "bbox": [ + 133, + 185, + 481, + 224 + ], + "type": "text", + "content": ", Qizi Chen" + }, + { + "bbox": [ + 133, + 185, + 481, + 224 + ], + "type": "inline_equation", + "content": "^{4}" + }, + { + "bbox": [ + 133, + 185, + 481, + 224 + ], + "type": "text", + "content": ", Kai Zhang" + }, + { + "bbox": [ + 133, + 185, + 481, + 224 + ], + "type": "inline_equation", + "content": "^{5}" + }, + { + "bbox": [ + 133, + 185, + 481, + 224 + ], + "type": "text", + "content": ", Yunlong Zhang" + }, + { + "bbox": [ + 133, + 185, + 481, + 224 + ], + "type": "inline_equation", + "content": "^{1,2}" + }, + { + "bbox": [ + 133, + 185, + 481, + 224 + ], + "type": "text", + "content": ", Dan Wan" + }, + { + "bbox": [ + 133, + 185, + 481, + 224 + ], + "type": "inline_equation", + "content": "^{6}" + }, + { + "bbox": [ + 133, + 185, + 481, + 224 + ], + "type": "text", + "content": ", Xiaoxiao Lan" + }, + { + "bbox": [ + 133, + 185, + 481, + 224 + ], + "type": "inline_equation", + "content": "^{1}" + }, + { + "bbox": [ + 133, + 185, + 481, + 224 + ], + "type": "text", + "content": ", Mengyue Zheng" + }, + { + "bbox": [ + 133, + 185, + 481, + 224 + ], + "type": "inline_equation", + "content": "^{2}" + }, + { + "bbox": [ + 133, + 185, + 481, + 224 + ], + "type": "text", + "content": ", Jingxiong Li" + }, + { + "bbox": [ + 133, + 185, + 481, + 224 + ], + "type": "inline_equation", + "content": "^{1,2}" + }, + { + "bbox": [ + 133, + 185, + 481, + 224 + ], + "type": "text", + "content": ", Xinheng Lyu" + }, + { + "bbox": [ + 133, + 185, + 481, + 224 + ], + "type": "inline_equation", + "content": "^{2}" + }, + { + "bbox": [ + 133, + 185, + 481, + 224 + ], + "type": "text", + "content": ", Tao Lin" + }, + { + "bbox": [ + 133, + 185, + 481, + 224 + ], + "type": "inline_equation", + "content": "^{2,\\dagger}" + }, + { + "bbox": [ + 133, + 185, + 481, + 224 + ], + "type": "text", + "content": ", and Lin Yang" + }, + { + "bbox": [ + 133, + 185, + 481, + 224 + ], + "type": "inline_equation", + "content": "^{2,\\dagger}" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 209, + 231, + 405, + 243 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 209, + 231, + 405, + 243 + ], + "spans": [ + { + "bbox": [ + 209, + 231, + 405, + 243 + ], + "type": "text", + "content": "1 Zhejiang University, Hangzhou 310058, China" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 208, + 243, + 405, + 254 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 208, + 243, + 405, + 254 + ], + "spans": [ + { + "bbox": [ + 208, + 243, + 405, + 254 + ], + "type": "inline_equation", + "content": "^{2}" + }, + { + "bbox": [ + 208, + 243, + 405, + 254 + ], + "type": "text", + "content": " Westlake University, Hangzhou 310030, China" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 182, + 255, + 433, + 265 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 182, + 255, + 433, + 265 + ], + "spans": [ + { + "bbox": [ + 182, + 255, + 433, + 265 + ], + "type": "text", + "content": "{sunyuxuan,zhuchenglu,lintao,yanglin}@westlake.edu.cn" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 165, + 265, + 447, + 275 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 165, + 265, + 447, + 275 + ], + "spans": [ + { + "bbox": [ + 165, + 265, + 447, + 275 + ], + "type": "text", + "content": "3 Macau University of Science and Technology, Macau 999078, China" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 235, + 277, + 378, + 286 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 235, + 277, + 378, + 286 + ], + "spans": [ + { + "bbox": [ + 235, + 277, + 378, + 286 + ], + "type": "text", + "content": "2220024311@student.must.edu.mo" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 217, + 286, + 396, + 297 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 217, + 286, + 396, + 297 + ], + "spans": [ + { + "bbox": [ + 217, + 286, + 396, + 297 + ], + "type": "inline_equation", + "content": "^{4}" + }, + { + "bbox": [ + 217, + 286, + 396, + 297 + ], + "type": "text", + "content": " Jiangnan University, Wuxi 214122, China" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 200, + 297, + 413, + 308 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 200, + 297, + 413, + 308 + ], + "spans": [ + { + "bbox": [ + 200, + 297, + 413, + 308 + ], + "type": "text", + "content": "5 The Ohio State University, Columbus 43210, USA" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 153, + 308, + 460, + 319 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 153, + 308, + 460, + 319 + ], + "spans": [ + { + "bbox": [ + 153, + 308, + 460, + 319 + ], + "type": "inline_equation", + "content": "^{6}" + }, + { + "bbox": [ + 153, + 308, + 460, + 319 + ], + "type": "text", + "content": " Fujian University of Traditional Chinese Medicine, Fuzhou 350122, China" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 220, + 320, + 392, + 331 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 220, + 320, + 392, + 331 + ], + "spans": [ + { + "bbox": [ + 220, + 320, + 392, + 331 + ], + "type": "text", + "content": "https://pathmmu-benchmark.github.io/" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 159, + 354, + 453, + 628 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 159, + 354, + 453, + 628 + ], + "spans": [ + { + "bbox": [ + 159, + 354, + 453, + 628 + ], + "type": "text", + "content": "Abstract. The emergence of Large Multimodal Models (LMMs) has unlocked remarkable potential in AI, particularly in pathology. However, the lack of specialized, high-quality benchmark impeded their development and precise evaluation. To address this, we introduce PathMMU, the largest and highest-quality expert validated pathology benchmark for LMMs. It comprises 33,428 multimodal multi-choice questions and 24,067 images from various sources, each accompanied by an explanation for the correct answer. The construction of PathMMU leverages GPT-4V's advanced capabilities, utilizing over 30,000 image-caption pairs to enrich the descriptive quality of captions and generate corresponding Q&As in a cascading process. To maximize PathMMU's authority, we invite seven pathologists to scrutinize each question under strict standards in PathMMU's validation and test sets, while simultaneously setting an expert-level performance benchmark for PathMMU. We conduct extensive evaluations, including zero-shot assessments of 14 open-sourced and 4 closed-sourced LMMs and their robustness to image corruption. We also fine-tune representative LMMs to assess their adaptability to PathMMU. The empirical findings indicate that advanced LMMs struggle with the challenging PathMMU benchmark, with the top-performing LMM, GPT-4V, achieving only a " + }, + { + "bbox": [ + 159, + 354, + 453, + 628 + ], + "type": "inline_equation", + "content": "49.8\\%" + }, + { + "bbox": [ + 159, + 354, + 453, + 628 + ], + "type": "text", + "content": " zero-shot performance, significantly lower than the " + }, + { + "bbox": [ + 159, + 354, + 453, + 628 + ], + "type": "inline_equation", + "content": "71.8\\%" + }, + { + "bbox": [ + 159, + 354, + 453, + 628 + ], + "type": "text", + "content": " demonstrated by human pathologists. After fine-tuning, substantially smaller open-sourced LMMs can outperform GPT-4V but still fall short of the expertise shown by pathologists. We hope that the PathMMU will offer valuable insights and foster the development of more specialized, next-generation LMMs for pathology." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 160, + 638, + 420, + 650 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 160, + 638, + 420, + 650 + ], + "spans": [ + { + "bbox": [ + 160, + 638, + 420, + 650 + ], + "type": "text", + "content": "Keywords: Large multimodal model " + }, + { + "bbox": [ + 160, + 638, + 420, + 650 + ], + "type": "inline_equation", + "content": "\\cdot" + }, + { + "bbox": [ + 160, + 638, + 420, + 650 + ], + "type": "text", + "content": " Pathology " + }, + { + "bbox": [ + 160, + 638, + 420, + 650 + ], + "type": "inline_equation", + "content": "\\cdot" + }, + { + "bbox": [ + 160, + 638, + 420, + 650 + ], + "type": "text", + "content": " Benchmark" + } + ] + } + ], + "index": 12 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 135, + 653, + 236, + 665 + ], + "type": "page_footnote", + "angle": 0, + "lines": [ + { + "bbox": [ + 135, + 653, + 236, + 665 + ], + "spans": [ + { + "bbox": [ + 135, + 653, + 236, + 665 + ], + "type": "text", + "content": "† Corresponding author." + } + ] + } + ], + "index": 13 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 0 + }, + { + "para_blocks": [ + { + "bbox": [ + 133, + 114, + 229, + 127 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 114, + 229, + 127 + ], + "spans": [ + { + "bbox": [ + 133, + 114, + 229, + 127 + ], + "type": "text", + "content": "1 Introduction" + } + ] + } + ], + "index": 2 + }, + { + "type": "image", + "bbox": [ + 134, + 159, + 228, + 274 + ], + "blocks": [ + { + "bbox": [ + 134, + 159, + 228, + 274 + ], + "lines": [ + { + "bbox": [ + 134, + 159, + 228, + 274 + ], + "spans": [ + { + "bbox": [ + 134, + 159, + 228, + 274 + ], + "type": "image", + "image_path": "50434f96daebcdae5f9726c2f375bba90c25bf354d535fffe668de2d0405ab41.jpg" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 130, + 284, + 482, + 329 + ], + "lines": [ + { + "bbox": [ + 130, + 284, + 482, + 329 + ], + "spans": [ + { + "bbox": [ + 130, + 284, + 482, + 329 + ], + "type": "text", + "content": "Fig. 1: An overview of the PathMMU benchmark: PathMMU is built from diverse, rich data sources. It comprises expert-level, multimodal, multiple-choice questions in pathology, collaboratively crafted by AI and pathology experts. The benchmark reveals even the most advanced LMMs significantly underperform compared to human experts." + } + ] + } + ], + "index": 6, + "angle": 0, + "type": "image_caption" + } + ], + "index": 3 + }, + { + "type": "image", + "bbox": [ + 235, + 159, + 358, + 275 + ], + "blocks": [ + { + "bbox": [ + 235, + 159, + 358, + 275 + ], + "lines": [ + { + "bbox": [ + 235, + 159, + 358, + 275 + ], + "spans": [ + { + "bbox": [ + 235, + 159, + 358, + 275 + ], + "type": "image", + "image_path": "fab67f6ef220219bc7e6430a568fd532310a1b2d6726a7c43dbbf7fdec5810f3.jpg" + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "image_body" + } + ], + "index": 4 + }, + { + "type": "image", + "bbox": [ + 361, + 159, + 480, + 275 + ], + "blocks": [ + { + "bbox": [ + 361, + 159, + 480, + 275 + ], + "lines": [ + { + "bbox": [ + 361, + 159, + 480, + 275 + ], + "spans": [ + { + "bbox": [ + 361, + 159, + 480, + 275 + ], + "type": "image", + "image_path": "688b82a269cf86c14034923178aa4720a73a259f9bc7399aa7c9cd3a6df5c92c.jpg" + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "image_body" + } + ], + "index": 5 + }, + { + "bbox": [ + 130, + 354, + 480, + 403 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 354, + 480, + 403 + ], + "spans": [ + { + "bbox": [ + 130, + 354, + 480, + 403 + ], + "type": "text", + "content": "Pathology is integral to modern medicine, serving as the foundation for diagnosing and understanding diseases. For instance, liquid-based cytology is crucial for early cervical cancer detection. Assessing biomarkers like HER2 in breast cancer guides the selection of targeted therapies and immunotherapies [23]." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 130, + 403, + 481, + 533 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 403, + 481, + 533 + ], + "spans": [ + { + "bbox": [ + 130, + 403, + 481, + 533 + ], + "type": "text", + "content": "In recent years, the field of pathology has undergone a significant transformation, driven by advances in digital pathology and the integration of AI. This shift marks a departure from conventional microscope-based slide reading to AI-powered image analysis, greatly easing the workload of pathologists. Traditional pathology models are tailored for specific tasks, such as cervical cytology and liver lesion classification, leading to an abundance of task-specific models [57,59]. In contrast, recent advancements in Large Multimodal Models (LMMs) focus on offering general task-solving capabilities, thereby making universal pathological region recognition possible. This breakthrough not only represents a significant stride in the field, but also paves the way for more versatile and efficient diagnostic approaches in pathology [13,27,42,44]." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 130, + 534, + 482, + 666 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 534, + 482, + 666 + ], + "spans": [ + { + "bbox": [ + 130, + 534, + 482, + 666 + ], + "type": "text", + "content": "Given the demands of precise interpretation in pathology, conducting comprehensive evaluations of LMMs' abilities in interpreting pathology images is essential. However, the field faces a notable scarcity of high-quality benchmark datasets. Currently, the primary large-scale dataset available is PathVQA [18], which offers over 30,000 visual Q&A samples with 4,998 images. Nevertheless, these samples are derived from limited sources of textbooks, with image captions being converted into questions through heuristic approaches, constraining the generation of diverse and logical Q&As. Additionally, image captions in the textbook might not accurately reflect the corresponding images. For instance, some captions may introduce extraneous information not evident in the images or only partially cover the image content, thereby leading to unsolvable Q&As" + } + ] + } + ], + "index": 9 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "spans": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "type": "text", + "content": "2" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 166, + 91, + 219, + 100 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 166, + 91, + 219, + 100 + ], + "spans": [ + { + "bbox": [ + 166, + 91, + 219, + 100 + ], + "type": "text", + "content": "Y. Sun et al." + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 1 + }, + { + "para_blocks": [ + { + "bbox": [ + 130, + 116, + 482, + 164 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 116, + 482, + 164 + ], + "spans": [ + { + "bbox": [ + 130, + 116, + 482, + 164 + ], + "type": "text", + "content": "or questions that inadequately capture the essence of the images. Moreover, the absence of expert review and validation during the data curation process may introduce substantial noise into the dataset. These challenges pose significant obstacles to effectively validating LMMs' capabilities in pathology." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 130, + 169, + 482, + 422 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 169, + 482, + 422 + ], + "spans": [ + { + "bbox": [ + 130, + 169, + 482, + 422 + ], + "type": "text", + "content": "To address these challenges, we introduce PathMMU, a multimodal expert-level benchmark designed to evaluate LMMs in pathology image understanding and reasoning, comprising 33,428 Q&As along with 24,067 pathology images. As depicted in Fig. 1, PathMMU draws from a diverse range of high-quality sources, including PubMed scientific papers, pathology atlas from guidelines, Twitter posts by pathology experts, commonly used pathology classification datasets and educational content from YouTube videos, covering multiple organ systems (e.g., gastrointestinal, pulmonary, etc.) and multiple subjects (e.g., dermatopathology, hematopathology, etc.). In developing PathMMU, we adopt a meticulously designed cascading approach. Initially, we prompt GPT-4V with collected image_caption pairs to enhance the original captions, crafting descriptions that delve deeper into details and highlight crucial morphological features. Subsequently, we use these enhanced descriptions along with the images to prompt GPT-4V to generate professional multi-choice, multimodal pathology Q&As, each accompanied by detailed explanations for their answers. To ensure the questions are specifically designed for multimodal pathology understanding, we employ a collaboration of multiple Large Language Models (LLMs) to eliminate questions that can be solved or guessed using text alone. Most importantly, we invite seven pathology experts to manually review approximately 12,000 Q&As from the test and validation sets of PathMMU. The above carefully executed procedures, ensure the generation of professional, logical, and high-quality Q&As." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 130, + 426, + 483, + 668 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 426, + 483, + 668 + ], + "spans": [ + { + "bbox": [ + 130, + 426, + 483, + 668 + ], + "type": "text", + "content": "We evaluate the zero-shot performance of 18 advanced LMMs and their robustness on the PathMMU test sets. We also fine-tune two representative LMMs on the training set to assess their transfer learning capabilities. The key insights from this extensive analysis are: (1) Cutting-edge LMMs struggle with the PathMMU, with 15 out of the 18 achieving no more than " + }, + { + "bbox": [ + 130, + 426, + 483, + 668 + ], + "type": "inline_equation", + "content": "40\\%" + }, + { + "bbox": [ + 130, + 426, + 483, + 668 + ], + "type": "text", + "content": " accuracy. Even the most advanced model currently available, GPT-4V, only attains an accuracy of " + }, + { + "bbox": [ + 130, + 426, + 483, + 668 + ], + "type": "inline_equation", + "content": "49.8\\%" + }, + { + "bbox": [ + 130, + 426, + 483, + 668 + ], + "type": "text", + "content": ". This reveals a notable discrepancy of nearly " + }, + { + "bbox": [ + 130, + 426, + 483, + 668 + ], + "type": "inline_equation", + "content": "20\\%" + }, + { + "bbox": [ + 130, + 426, + 483, + 668 + ], + "type": "text", + "content": " compared to professional pathologists, indicating significant deficiencies of current LMMs in the specialized field of pathology and emphasizing the challenging nature of the PathMMU benchmark. (2) While these LMMs show considerable robustness in robustness tests, the full extent of their robustness is still in question. On the one hand, given their limited performance, the scope for further performance degradation is somewhat constrained. On the other hand, these models tend to capture only superficial image features, and sometimes may not even utilize image information for reasoning. As a result, even when image corruption obscures minor details, the impact on these models' performance might not be significantly pronounced. (3) The text-only performance of GPT-4 Turbo surpasses the top-performing open-source LMMs on the PathMMU benchmark. We observe that GPT-4 Turbo sometimes takes shortcuts to guess answers. This involves choosing the correct answer based" + } + ] + } + ], + "index": 4 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 402, + 91, + 447, + 100 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 402, + 91, + 447, + 100 + ], + "spans": [ + { + "bbox": [ + 402, + 91, + 447, + 100 + ], + "type": "text", + "content": "PathMMU" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 474, + 92, + 481, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 474, + 92, + 481, + 100 + ], + "spans": [ + { + "bbox": [ + 474, + 92, + 481, + 100 + ], + "type": "text", + "content": "3" + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 2 + }, + { + "para_blocks": [ + { + "bbox": [ + 130, + 116, + 482, + 201 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 116, + 482, + 201 + ], + "spans": [ + { + "bbox": [ + 130, + 116, + 482, + 201 + ], + "type": "text", + "content": "either on the options most commonly encountered in real-world pathological scenarios, or by identifying the most distinctive option, rather than conducting an in-depth analysis of the pathology image. (4) Fine-tuning general-purpose LMMs with substantial amounts of data significantly boosts their ability to comprehend and reason with pathological images, enabling them to easily surpass the zero-shot performance of GPT-4V. However, there still remains a discernible gap between their performance and that of human experts." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 132, + 220, + 244, + 232 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 220, + 244, + 232 + ], + "spans": [ + { + "bbox": [ + 132, + 220, + 244, + 232 + ], + "type": "text", + "content": "2 Related Works" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 130, + 246, + 482, + 450 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 246, + 482, + 450 + ], + "spans": [ + { + "bbox": [ + 130, + 246, + 482, + 450 + ], + "type": "text", + "content": "Multimodal Models. The emergence of powerful large language models such as BERT [12], GPT-3 [8], T5 [38], LLaMa [46], ChatGPT [33], GPT-4 [34], and Vicuna [10] has significantly advanced the field of Natural Language Processing (NLP). These models demonstrate exceptional general capabilities, motivating researchers to explore the integration of LLMs with vision models to develop versatile multimodal models. This integration is primarily achieved through the pretraining approach, leading to the creation of groundbreaking LMMs. Representative models of this approach include CLIP [37], Flamingo [1], BLIP-2 [29], and Fuyu [5], all of which demonstrate remarkable multimodal understanding abilities. Furthermore, by adapting instruction-tuning techniques from the NLP domain, a variety of multimodal instruction-tuning datasets have been developed. This initiative aims to guide LMMs in generating outputs that are more controllable, practical, and adaptable to various tasks. Simultaneously, this approach facilitates the integration of LLMs into LMMs using lightweight fine-tuning, significantly reducing both costs and time. Key models that embody this advancement include GPT-4V [35], Gemini Pro Vision [45], Qwen-VL [4], LLaVA [30], MiniGPT-4 [60], BLIP-2 [29] and InstructBLIP [11]." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 130, + 450, + 483, + 569 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 450, + 483, + 569 + ], + "spans": [ + { + "bbox": [ + 130, + 450, + 483, + 569 + ], + "type": "text", + "content": "LMM Benchmarks. In the general domain, the rise of various large models has led to the creation of extensive benchmark datasets, designed to evaluate the universal capabilities of these models across a range of tasks and domains. Notable examples include LAMM [52], LVLM-eHub [51], SEED [26], MMBench [31], MMVet [53] and BenchLMM [9], which have been used to assess the basic perception abilities of large models. More recently, MMMU [54], a massive and challenging dataset covering 30 university-level subjects, has been developed. It is specifically designed to evaluate the general multimodal understanding and reasoning capabilities of LMMs. Furthermore, HaELM [48] and HallusionBench [16] are proposed to evaluate the degree of hallucination in LMMs." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 130, + 570, + 482, + 666 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 570, + 482, + 666 + ], + "spans": [ + { + "bbox": [ + 130, + 570, + 482, + 666 + ], + "type": "text", + "content": "In the medical field, datasets such as VQA-RAD [24] and VQA-Med [6] provide a collection of Q&A pairs based on radiology images, facilitating the evaluation of the capabilities of AI in radiological diagnostics. Additionally, PMC-VQA [55] is a large-scale dataset within this domain. It generates a vast number of Q&A pairs by prompting ChatGPT with text-only image captions from PubMed. In the pathology domain, limited dataset construction efforts have been performed to date. The representative large-scale dataset available, PathVQA [18], is constructed in a manner similar to PMC-VQA. It utilizes" + } + ] + } + ], + "index": 6 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 133, + 91, + 141, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 91, + 141, + 100 + ], + "spans": [ + { + "bbox": [ + 133, + 91, + 141, + 100 + ], + "type": "text", + "content": "4" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 166, + 91, + 220, + 102 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 166, + 91, + 220, + 102 + ], + "spans": [ + { + "bbox": [ + 166, + 91, + 220, + 102 + ], + "type": "text", + "content": "Y. Sun et al." + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 3 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 136, + 114, + 242, + 198 + ], + "blocks": [ + { + "bbox": [ + 136, + 114, + 242, + 198 + ], + "lines": [ + { + "bbox": [ + 136, + 114, + 242, + 198 + ], + "spans": [ + { + "bbox": [ + 136, + 114, + 242, + 198 + ], + "type": "image", + "image_path": "1c94cd14cf7507c2335dd8ab8c118fad25088a4fc7f6416c15ae7aad867fb3c4.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 130, + 206, + 482, + 275 + ], + "lines": [ + { + "bbox": [ + 130, + 206, + 482, + 275 + ], + "spans": [ + { + "bbox": [ + 130, + 206, + 482, + 275 + ], + "type": "text", + "content": "Fig. 2: The comparison between PathMMU and existing benchmarks. The Q&A pairs in PathMMU are sourced extensively and comprehensively, undergoing rigorous multi-tiered filtering. This includes the initial filtering by multiple LLMs and the strict reviews by professional pathologists. Additionally, each question is accompanied by a detailed explanation. These attributes establish PathMMU as the most professionally curated, comprehensive, and highest-quality large-scale pathology dataset available." + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "image_caption" + } + ], + "index": 2 + }, + { + "type": "table", + "bbox": [ + 246, + 114, + 479, + 198 + ], + "blocks": [ + { + "bbox": [ + 246, + 114, + 479, + 198 + ], + "lines": [ + { + "bbox": [ + 246, + 114, + 479, + 198 + ], + "spans": [ + { + "bbox": [ + 246, + 114, + 479, + 198 + ], + "type": "table", + "html": "
DatasetDomainSources# Q&As / ImagesAnswer Explanation?LLM Filtered?Expert Annotated?
VQA-MedMedicalMedPix5000 / 5000XX✓
VQA-RADMedicalMedPix3515 / 315XX✓
PMC-VQAMedicalPMC227k / 149kXXX
PathVQAPathologyBook32799 / 4998XXX
Quilt-VQAPathologyEdu-Content1283 / 985XX✓
Quilt-Red
Social Media
PathMMUPathologyEdu-Content, PubMed33428 / 24067✓✓✓
Atlas, CLS-Data
", + "image_path": "e63abd2c46baa4a1e1e83c211a1b439f3148e16d827b3ce243a771e3f2fd48f2.jpg" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "table_body" + } + ], + "index": 3 + }, + { + "bbox": [ + 130, + 295, + 482, + 486 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 295, + 482, + 486 + ], + "spans": [ + { + "bbox": [ + 130, + 295, + 482, + 486 + ], + "type": "text", + "content": "a heuristic approach to generate questions based on the text-only captions of pathology textbook images, yielding over 30,000 Q&As and 4,998 images. More recently, Quilt-VQA [39] is proposed, which is constructed by automatically extracting question and answer pairs from the narrations of professional teachers in YouTube videos. This approach guarantees the professionalism of the Q&A pairs, although it comprises a relatively smaller dataset with 1,283 Q&As and 985 images. Nonetheless, datasets reliant solely on image captions for question generation, face several notable limitations: (1) The creation of questions typically requires the consideration of both the image and its accompanying caption. Relying only on captions can introduce inaccuracies, making some questions unanswerable or solvable using only the text, bypassing the need for image analysis. (2) Since captions tend to be simplistic and may not capture the detailed information evident in images, generating an excessive number of Q&As from the same image's caption can lead to either too simplistic or insufficient questions. (3) Automatically generated questions, whether through predefined rules or by prompting ChatGPT, are susceptible to containing numerous errors." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 130, + 487, + 483, + 584 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 487, + 483, + 584 + ], + "spans": [ + { + "bbox": [ + 130, + 487, + 483, + 584 + ], + "type": "text", + "content": "Given the current lack of high-quality benchmarks in pathology, we develop PathMMU, a large-scale, high-quality, and expert-verified pathology benchmark, to fill this gap. Our approach involves generating Q&As from a diverse range of image-caption pairs, followed by rigorous expert review and filtering applied to both the images and the questions. Moreover, we provide detailed explanations for each answer as a reference to enhance the interpretability of answers, which is absent in previous works. As shown in Fig. 2, PathMMU distinguishes itself in the pathology domain through its professionalism and superior quality." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 132, + 599, + 378, + 613 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 599, + 378, + 613 + ], + "spans": [ + { + "bbox": [ + 132, + 599, + 378, + 613 + ], + "type": "text", + "content": "3 The Proposed PathMMU Benchmark" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 132, + 623, + 246, + 635 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 623, + 246, + 635 + ], + "spans": [ + { + "bbox": [ + 132, + 623, + 246, + 635 + ], + "type": "text", + "content": "3.1 Design Principle" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 130, + 641, + 481, + 666 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 641, + 481, + 666 + ], + "spans": [ + { + "bbox": [ + 130, + 641, + 481, + 666 + ], + "type": "text", + "content": "PathMMU is designed to provide the community with a specialized dataset for evaluating pathology LMMs. We adhere to five principal guidelines in its devel-" + } + ] + } + ], + "index": 9 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 402, + 91, + 447, + 100 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 402, + 91, + 447, + 100 + ], + "spans": [ + { + "bbox": [ + 402, + 91, + 447, + 100 + ], + "type": "text", + "content": "PathMMU" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 474, + 92, + 480, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 474, + 92, + 480, + 100 + ], + "spans": [ + { + "bbox": [ + 474, + 92, + 480, + 100 + ], + "type": "text", + "content": "5" + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 4 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 134, + 114, + 247, + 258 + ], + "blocks": [ + { + "bbox": [ + 134, + 114, + 247, + 258 + ], + "lines": [ + { + "bbox": [ + 134, + 114, + 247, + 258 + ], + "spans": [ + { + "bbox": [ + 134, + 114, + 247, + 258 + ], + "type": "image", + "image_path": "61f327ce61ac000781b5bac66790e28ac7c27c92b9a49089e792b123e8757cbd.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 131, + 266, + 482, + 301 + ], + "lines": [ + { + "bbox": [ + 131, + 266, + 482, + 301 + ], + "spans": [ + { + "bbox": [ + 131, + 266, + 482, + 301 + ], + "type": "text", + "content": "Fig. 3: An illustrative overview of the three main processes in PathMMU Q&A generation: data collection and preprocessing, detailed pathology image description generation, and question generation with LLMs filtering and expert validation." + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "image_caption" + } + ], + "index": 2 + }, + { + "type": "image", + "bbox": [ + 252, + 114, + 354, + 258 + ], + "blocks": [ + { + "bbox": [ + 252, + 114, + 354, + 258 + ], + "lines": [ + { + "bbox": [ + 252, + 114, + 354, + 258 + ], + "spans": [ + { + "bbox": [ + 252, + 114, + 354, + 258 + ], + "type": "image", + "image_path": "cd66e1d8d36649ca58e541ac02307491fba018b121bd091fdd88fa46466fa080.jpg" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_body" + } + ], + "index": 3 + }, + { + "type": "image", + "bbox": [ + 358, + 114, + 479, + 258 + ], + "blocks": [ + { + "bbox": [ + 358, + 114, + 479, + 258 + ], + "lines": [ + { + "bbox": [ + 358, + 114, + 479, + 258 + ], + "spans": [ + { + "bbox": [ + 358, + 114, + 479, + 258 + ], + "type": "image", + "image_path": "54244a918b274a764ccc3b2174f41c816e068bec2f476862e438d7de41d3fd3e.jpg" + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "image_body" + } + ], + "index": 4 + }, + { + "bbox": [ + 130, + 321, + 482, + 562 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 321, + 482, + 562 + ], + "spans": [ + { + "bbox": [ + 130, + 321, + 482, + 562 + ], + "type": "text", + "content": "opment: (1) Comprehensive and Specialized Data: Our benchmark is gathered from a diverse range of authoritative materials, including respected scientific publications from PubMed, pathology atlas from textbooks and guidelines, educational videos on YouTube, pathologist-shared images with explanations on Twitter, and widely recognized pathology classification datasets. Additionally, we involve seven human experts to manually review the dataset to ensure it meets professional standards. (2) Effective and Valuable Questions: The questions in PathMMU are carefully formulated to necessitate answers based on detailed observation of pathology images, rather than being solvable through mere textual interpretation or guesswork. We ensure that these questions are answerable, effective, and aligned with the standards of professional pathology examinations. (3) Large-Scale: We present the largest pathology dataset currently available. This scale enables researchers to thoroughly explore the full potential of LMMs in the field of pathology. (4) High-Quality Images: We prioritize clarity in our images to ensure that every detail is discernible, as evidenced by our dataset's average resolution of approximately " + }, + { + "bbox": [ + 130, + 321, + 482, + 562 + ], + "type": "inline_equation", + "content": "900 \\times 700" + }, + { + "bbox": [ + 130, + 321, + 482, + 562 + ], + "type": "text", + "content": " pixels. (5) Explainable Answer Choices: In line with the ongoing interest in understanding the decision-making logic of large models, our dataset includes reference explanations for each answer choice. This feature enhances the interpretability of model responses, contributing to the broader research on model explainability." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 132, + 576, + 321, + 588 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 576, + 321, + 588 + ], + "spans": [ + { + "bbox": [ + 132, + 576, + 321, + 588 + ], + "type": "text", + "content": "3.2 The Construction of PathMMU" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 130, + 594, + 482, + 630 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 594, + 482, + 630 + ], + "spans": [ + { + "bbox": [ + 130, + 594, + 482, + 630 + ], + "type": "text", + "content": "To guarantee the quality of our benchmark, we meticulously develop a three-step data processing and generation protocol. The overall data collection and generation process is illustrated in Fig. 3." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 130, + 629, + 482, + 665 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 629, + 482, + 665 + ], + "spans": [ + { + "bbox": [ + 130, + 629, + 482, + 665 + ], + "type": "text", + "content": "Step 1: Data Collection and Preprocessing. Our data collection draws from a wide variety of sources, which we integrate as subsets into the PathMMU framework. We name these subsets and detail their collection workflows as follows." + } + ] + } + ], + "index": 9 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "spans": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "type": "text", + "content": "6" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 166, + 91, + 219, + 101 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 166, + 91, + 219, + 101 + ], + "spans": [ + { + "bbox": [ + 166, + 91, + 219, + 101 + ], + "type": "text", + "content": "Y. Sun et al." + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 5 + }, + { + "para_blocks": [ + { + "bbox": [ + 133, + 115, + 482, + 415 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 115, + 482, + 415 + ], + "spans": [ + { + "bbox": [ + 133, + 115, + 482, + 415 + ], + "type": "text", + "content": "PubMed: This subset consists of scientific pathology image-text pairs sourced from scientific documents. We gather these pairs from the open-access section of PubMed Central. Given the significant presence of non-pathological data within this resource, we meticulously annotate 20,000 samples as either pathological or non-pathological, and subsequently train a ConvNeXt [32] model to identify pathology data within the remaining dataset. EduContent: This subset, derived from educational YouTube videos, is sourced from Quilt1M [20]. Since YouTube videos often include a variety of visuals unrelated to pathology, such as computer desktops and unrelated imagery, the initial phase involves the annotation of pathological regions in 3,000 YouTube video images. This critical step is followed by the training of a YOLOv6-based detector [28] to automate the identification of pathological regions within remaining YouTube content. Atlas: This subset is a compilation of authoritative pathology textbooks and guidelines, we transform them from PDFs to HTML to facilitate the extraction of image-caption pairs. SocialPath: This subset aggregates image-text pairs from Twitter posts by pathology experts, using Twitter URLs provided by OpenPath [19]. PathCLS: Encompassing widely recognized pathology classification datasets, including PatchCamelyon [47], CRC-100K [21], SICAPv2 [40], BACH [2], Osteo [3], SkinCancer [22], MHIST [50], WSSSLUAD [17], LC-Lung and LC-Colon [7]. We reformulate these datasets into a Q&A format and randomly sample from these datasets. We directly make the gathered samples from PathCLS a part of the PathMMU dataset without undergoing the following two steps Q&A generation process. Finally, we manually review and filter irrelevant or unclear images to ensure the quality of collected data. As a result, approximately 30,000 high-quality image-text pairs are obtained, forming the foundation of the PathMMU." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 133, + 418, + 482, + 563 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 418, + 482, + 563 + ], + "spans": [ + { + "bbox": [ + 133, + 418, + 482, + 563 + ], + "type": "text", + "content": "Step 2: Detailed Description Generation and Question-answer Pairs Generation. When dealing with data sources from platforms like Twitter and YouTube, it's common to find a weak correlation between an image and its accompanying caption. For instance, in the SocialPath dataset, captions might only describe the aesthetic appeal of the image, which is irrelevant to pathological findings. Similarly, in other data sources, captions tend to describe only a fraction of the features present in the image. To tackle this issue, we prompt GPT-4V to focus on describing the morphology of cells and tissues. It is crucial to note that merely prompting GPT-4V to output image descriptions might lead to incoherent content. Therefore, we provide GPT-4V with original captions for reference, significantly reducing the incidence of such occurrences and thus ensuring the generation of more precise and relevant descriptions." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 133, + 566, + 482, + 639 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 566, + 482, + 639 + ], + "spans": [ + { + "bbox": [ + 133, + 566, + 482, + 639 + ], + "type": "text", + "content": "Step 3: Question Generation and Expert Validation. Following the generation of image descriptions, GPT-4V is tasked with creating three questions per image, each accompanied by multiple-choice options, the correct answer, and detailed explanations for the answer. The creation of these explanations serves a dual purpose: it not only offers valuable insights into the model reasoning process, but also facilitates the subsequent manual review phase." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 132, + 641, + 481, + 665 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 641, + 481, + 665 + ], + "spans": [ + { + "bbox": [ + 132, + 641, + 481, + 665 + ], + "type": "text", + "content": "It's crucial to note that despite the careful design of question generation prompts for GPT-4V, the text-only GPT-4 is capable of correctly answering" + } + ] + } + ], + "index": 5 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 402, + 91, + 447, + 100 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 402, + 91, + 447, + 100 + ], + "spans": [ + { + "bbox": [ + 402, + 91, + 447, + 100 + ], + "type": "text", + "content": "PathMMU" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 474, + 91, + 481, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 474, + 91, + 481, + 100 + ], + "spans": [ + { + "bbox": [ + 474, + 91, + 481, + 100 + ], + "type": "text", + "content": "7" + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 6 + }, + { + "para_blocks": [ + { + "bbox": [ + 130, + 116, + 482, + 224 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 116, + 482, + 224 + ], + "spans": [ + { + "bbox": [ + 130, + 116, + 482, + 224 + ], + "type": "text", + "content": "over " + }, + { + "bbox": [ + 130, + 116, + 482, + 224 + ], + "type": "inline_equation", + "content": "60\\%" + }, + { + "bbox": [ + 130, + 116, + 482, + 224 + ], + "type": "text", + "content": " of these questions through educated guesses, bypassing the need for visual cues. GPT-4 explains that its deductions are drawn from patterns such as one option being more common in typical pathological scenarios or exhibiting noticeable differences from other options. We delve deeper into this phenomenon in Sec. 4.3 to offer a more thorough experimental analysis. To ensure that LMMs do not rely on educated guesses, thereby obscuring their true multimodal capabilities, we employ GPT-3.5 Turbo, GPT-4 Turbo, Gemini Pro, and ERNIE-Bot-4 to perform educated guesses. Questions correctly guessed by at least three of these models are subsequently excluded." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 130, + 224, + 482, + 378 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 224, + 482, + 378 + ], + "spans": [ + { + "bbox": [ + 130, + 224, + 482, + 378 + ], + "type": "text", + "content": "We partitioned the remaining dataset into training, validation, and test sets. The training set comprises 16,344 images and 23,041 Q&As, while the validation and test sets contain roughly 12,000 Q&As and over 8,000 images, respectively. For the validation and test sets, we invite seven professional pathologists to conduct a thorough manual review. These pathologists initially attempt to answer the questions independently and then assess them based on the following criteria: (1) Whether the question can be answered without an accompanying image. (2) Whether the answer can be inferred from the provided question and image. (3) Whether the supplied answer is incorrect, if there is no correct answer, or if multiple correct answers are possible. (4) Whether the generated question appears unusual or atypical compared to standard questions in pathology examinations. Questions failing to meet these standards are deemed invalid and consequently removed from the PathMMU dataset." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 130, + 380, + 482, + 464 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 380, + 482, + 464 + ], + "spans": [ + { + "bbox": [ + 130, + 380, + 482, + 464 + ], + "type": "text", + "content": "Ultimately, after thorough expert review, we obtain 710 Q&As accompanied by 510 images for the validation set, and 9,677 Q&As with 7,213 images for the test set. Additionally, to establish a standard for expert performance, we extract a smaller subset from the test set, named 'test-tiny', which includes 1,156 Q&As. We invite two groups of professional pathologists to participate in this subset as an examination, and we average their performance to set a benchmark for expert performance in the PathMMU, serving as a comparison reference for LMMs." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 132, + 479, + 230, + 493 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 479, + 230, + 493 + ], + "spans": [ + { + "bbox": [ + 132, + 479, + 230, + 493 + ], + "type": "text", + "content": "4 Experiments" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 130, + 501, + 482, + 610 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 501, + 482, + 610 + ], + "spans": [ + { + "bbox": [ + 130, + 501, + 482, + 610 + ], + "type": "text", + "content": "In this section, we conduct extensive experiments on PathMMU. Initially, we assess the zero-shot performance of cutting-edge LMMs. To verify whether LMMs effectively utilize the visual information from pathology images, we also evaluate the performance of text-only LLMs on PathMMU as a reference. Additionally, we apply common corruptions to the images in the test-tiny set to test the robustness of the LMMs against image corruptions. Furthermore, we select representative LMMs and evaluate their performance to explore their transfer-learning capabilities. Finally, we conduct experimental analysis on the issue of LLMs being able to guess answers identified during the data construction process." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 132, + 624, + 371, + 635 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 624, + 371, + 635 + ], + "spans": [ + { + "bbox": [ + 132, + 624, + 371, + 635 + ], + "type": "text", + "content": "4.1 Zero-shot Evaluation of LMMs and LLMs" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 130, + 641, + 481, + 665 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 641, + 481, + 665 + ], + "spans": [ + { + "bbox": [ + 130, + 641, + 481, + 665 + ], + "type": "text", + "content": "In this study, we evaluate the zero-shot capabilities of the latest and most advanced LMMs on the PathMMU. Specifically, we use the PathMMU validation" + } + ] + } + ], + "index": 8 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "spans": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "type": "text", + "content": "8" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 166, + 91, + 219, + 100 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 166, + 91, + 219, + 100 + ], + "spans": [ + { + "bbox": [ + 166, + 91, + 219, + 100 + ], + "type": "text", + "content": "Y. Sun et al." + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 7 + }, + { + "para_blocks": [ + { + "type": "table", + "bbox": [ + 134, + 168, + 480, + 432 + ], + "blocks": [ + { + "bbox": [ + 131, + 114, + 482, + 159 + ], + "lines": [ + { + "bbox": [ + 131, + 114, + 482, + 159 + ], + "spans": [ + { + "bbox": [ + 131, + 114, + 482, + 159 + ], + "type": "text", + "content": "Table 1: Overall results of models on the PathMMU validation and test set. Besides reporting the performance of LMMs, we add text-only LLM baselines that purely accept text as inputs. The best-performing LMM in each subset is in- bold, and the top-performing LLM is underlined." + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 134, + 168, + 480, + 432 + ], + "lines": [ + { + "bbox": [ + 134, + 168, + 480, + 432 + ], + "spans": [ + { + "bbox": [ + 134, + 168, + 480, + 432 + ], + "type": "table", + "html": "
Validation OverallTestOverallPubMedSocialPathEduContentAtlasPathCLS
- (710)Tiny (1156)ALL (9677)Tiny (281)ALL (3068)Tiny (235)All (1855)Tiny (255)All (1938)Tiny (208)ALL (1007)Tiny (177)ALL (1809)
Random Choice24.622.123.722.125.125.526.525.526.019.723.015.316.3
Frequent Choice27.527.725.528.826.127.726.729.826.528.427.522.021.0
Expert performance-71.8-72.9-71.5-69.0-68.3-78.9-
Large Multimodal Models (LMMs): Text + Image as Input
OpenFlamingo2-9B [1]26.224.124.224.925.525.525.319.624.428.825.522.020.4
Kosmos2 [36]26.926.124.927.825.827.224.325.125.923.124.227.122.8
LLaMA-Adapter2-7B [14]26.226.626.426.028.327.726.527.128.030.827.420.920.5
MiniGPT4-Vicuna-13B [60]27.225.527.728.830.124.327.225.529.327.928.619.221.8
Otter [25]26.129.328.134.930.226.028.430.630.029.328.223.221.3
Fuyu-8B [5]27.530.129.235.930.528.930.730.629.827.930.624.323.9
CogVLM [49]29.930.629.732.032.131.130.430.629.935.133.722.622.4
Qwen-VL-7B [4]29.432.131.534.932.933.635.834.133.635.633.818.621.7
BLIP-2 FLAN-T5-XL [29]32.832.631.835.934.634.933.633.732.834.637.420.320.8
BLIP-2 FLAN-T5-XXL [29]33.433.333.537.037.435.734.630.234.539.440.719.820.6
InstructBLIP FLAN-T5-XL [11]33.134.931.837.033.235.333.936.533.337.036.726.623.1
InstructBLIP FLAN-T5-XXL [11]32.734.333.939.137.233.634.334.536.038.539.322.622.7
LLaVA-1.5-7B [30]36.634.935.441.639.937.938.132.536.535.639.223.221.9
LLaVA-1.5-13B [30]37.938.837.644.541.040.440.434.139.447.144.324.923.5
Qwen-VL-PLUS [4]38.039.334.343.537.741.336.039.636.044.737.123.223.3
Qwen-VL-MAX [4]43.649.245.953.050.953.649.352.247.951.449.830.529.6
Gemini Pro Vision [45]41.942.842.743.844.942.442.043.543.749.549.432.834.7
GPT-4V-1106 [35]49.353.949.859.453.558.753.960.453.648.152.836.233.8
Large Language Models (LLMs): Only Text as Input
ERNIE-Bot 4.0 [41]31.834.330.937.031.233.632.940.034.536.137.420.920.6
Gemini Pro [45]31.031.631.031.333.531.931.433.332.738.037.721.520.9
Vicuna-v1.5-13B [10]32.031.231.635.233.830.634.331.033.534.635.722.020.7
GPT-3.5 Turbo [33]30.429.428.432.731.231.530.131.030.126.926.822.020.8
GPT-4 Turbo [34]36.541.838.148.842.443.442.347.140.641.343.222.321.1
", + "image_path": "f507d52c43380b2c009f0fd5ee3afbb06a3add12125d26456fd408bee9b9d79e.jpg" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "table_body" + } + ], + "index": 3 + }, + { + "bbox": [ + 130, + 460, + 482, + 616 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 460, + 482, + 616 + ], + "spans": [ + { + "bbox": [ + 130, + 460, + 482, + 616 + ], + "type": "text", + "content": "set to conduct prompt engineering for these models, and then test their performance on the test set. The tested models include 14 open-source LMMs: OpenFlamingo [1], Kosmos2 [36], LLaMA-Adapter2 [14], MiniGPT-4 [60], Otter [25], Fuyu [5], CogVLM [49], Qwen-VL [4], BLIP-2 [29], InstructBLIP [11], and LLaVA-1.5 [30], as well as 4 closed-source LMMs: Qwen-VL-PLUS, Qwen-VL-MAX [4], Gemini Pro Vision [45], and GPT-4V [35]. For BLIP-2, Instruct-BLIP, and LLaVA-1.5, we deploy various sizes of these models to examine the applicability of scaling rules in LMMs. Moreover, we extend our evaluation to purely text-based LLMs, including ERNIE-Bot 4.0 [41], Gemini Pro [45], Vicuna-v1.5-13B [10], GPT-3.5 Turbo [33], and GPT-4 Turbo [34]. This assessment aims to quantify their ability to infer correct answers through educated guesses without visual inputs, reflecting the extent to which LMMs integrate and leverage visual information compared to their text-only counterparts." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 130, + 617, + 482, + 666 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 617, + 482, + 666 + ], + "spans": [ + { + "bbox": [ + 130, + 617, + 482, + 666 + ], + "type": "text", + "content": "Results and Discussion. The advanced LMMs struggle with the PathMMU dataset. As shown in Tab. 1, among the 18 models tested, 15 show an accuracy below " + }, + { + "bbox": [ + 130, + 617, + 482, + 666 + ], + "type": "inline_equation", + "content": "40\\%" + }, + { + "bbox": [ + 130, + 617, + 482, + 666 + ], + "type": "text", + "content": ". Notably, the highest-performing open-source model, LLaVa-1.5-13B, and the top-performing closed-source model, GPT-4V, achieve" + } + ] + } + ], + "index": 5 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 402, + 91, + 447, + 100 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 402, + 91, + 447, + 100 + ], + "spans": [ + { + "bbox": [ + 402, + 91, + 447, + 100 + ], + "type": "text", + "content": "PathMMU" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 474, + 91, + 481, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 474, + 91, + 481, + 100 + ], + "spans": [ + { + "bbox": [ + 474, + 91, + 481, + 100 + ], + "type": "text", + "content": "9" + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 8 + }, + { + "para_blocks": [ + { + "bbox": [ + 130, + 116, + 479, + 175 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 116, + 479, + 175 + ], + "spans": [ + { + "bbox": [ + 130, + 116, + 479, + 175 + ], + "type": "text", + "content": "only " + }, + { + "bbox": [ + 130, + 116, + 479, + 175 + ], + "type": "inline_equation", + "content": "37.6\\%" + }, + { + "bbox": [ + 130, + 116, + 479, + 175 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 130, + 116, + 479, + 175 + ], + "type": "inline_equation", + "content": "49.8\\%" + }, + { + "bbox": [ + 130, + 116, + 479, + 175 + ], + "type": "text", + "content": " accuracy, respectively. Those are significantly lower than human expert performance, which stands at " + }, + { + "bbox": [ + 130, + 116, + 479, + 175 + ], + "type": "inline_equation", + "content": "71.8\\%" + }, + { + "bbox": [ + 130, + 116, + 479, + 175 + ], + "type": "text", + "content": ", underscoring a substantial gap between current LMMs and professional pathologist standards. This gap indicates that the practical application of LMMs in real-world pathological scenarios remains significantly constrained." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 130, + 176, + 480, + 415 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 176, + 480, + 415 + ], + "spans": [ + { + "bbox": [ + 130, + 176, + 480, + 415 + ], + "type": "text", + "content": "Closed-source LLMs achieve performance that is on par with or even surpasses that of the most advanced open-source LMMs. When considering text-only models, we prompt them to make educated guesses without providing the image. We observe that GPT-4 Turbo, Vicuna-v1.5-13B, GeminiPro, and ERNIE-Bot 4.0 outperform nearly half of the open-source LMMs. Notably, GPT-4 Turbo leads with a significant margin, achieving an accuracy rate of " + }, + { + "bbox": [ + 130, + 176, + 480, + 415 + ], + "type": "inline_equation", + "content": "38.1\\%" + }, + { + "bbox": [ + 130, + 176, + 480, + 415 + ], + "type": "text", + "content": ", exceeding the second-best LLM by " + }, + { + "bbox": [ + 130, + 176, + 480, + 415 + ], + "type": "inline_equation", + "content": "6.5\\%" + }, + { + "bbox": [ + 130, + 176, + 480, + 415 + ], + "type": "text", + "content": ", and even outperforming the best-performing open-source LMM, LLaVA-1.5-13B. The " + }, + { + "bbox": [ + 130, + 176, + 480, + 415 + ], + "type": "inline_equation", + "content": "38.1\\%" + }, + { + "bbox": [ + 130, + 176, + 480, + 415 + ], + "type": "text", + "content": " accuracy of GPT-4 Turbo, on the one hand, demonstrates that under our strict filtering criteria, even the best LLM, when provided with text alone, makes incorrect answers for the majority of questions. This underscores the effectiveness of our dataset in testing the capabilities of LMMs in pathological image analysis. On the other hand, it reaffirms the analysis in Sec. 3.2 that GPT-4 possesses strong logical reasoning abilities, enabling it to correctly guess a portion of the questions, significantly outperforming random selection. However, this raises critical concerns in pathology diagnosis: the reliability of model outputs in clinical scenarios remains questionable, especially when they rely on educated guesses, particularly probabilistic ones, rather than comprehensive analysis. This emphasizes the urgent necessity to develop interpretable AI models and advocate for their cautious application in pathology diagnostics." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 130, + 415, + 480, + 475 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 415, + 480, + 475 + ], + "spans": [ + { + "bbox": [ + 130, + 415, + 480, + 475 + ], + "type": "text", + "content": "The larger LMMs exhibit better performance. Specifically, LLaVa-1.5-13B exceeds LLaVa-1.5-7B model by " + }, + { + "bbox": [ + 130, + 415, + 480, + 475 + ], + "type": "inline_equation", + "content": "2.2\\%" + }, + { + "bbox": [ + 130, + 415, + 480, + 475 + ], + "type": "text", + "content": " in overall performance, while Instruct-BLIP FLAN-T5-XXL and BLIP-2 FLAN-T5 XXL surpass their smaller versions by " + }, + { + "bbox": [ + 130, + 415, + 480, + 475 + ], + "type": "inline_equation", + "content": "2.1\\%" + }, + { + "bbox": [ + 130, + 415, + 480, + 475 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 130, + 415, + 480, + 475 + ], + "type": "inline_equation", + "content": "1.7\\%" + }, + { + "bbox": [ + 130, + 415, + 480, + 475 + ], + "type": "text", + "content": ", respectively. This suggests that larger models typically possess stronger multimodal capabilities in the field of pathology." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 131, + 491, + 375, + 504 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 131, + 491, + 375, + 504 + ], + "spans": [ + { + "bbox": [ + 131, + 491, + 375, + 504 + ], + "type": "text", + "content": "4.2 Robustness of LMMs to Image Corruption" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 130, + 510, + 479, + 581 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 510, + 479, + 581 + ], + "spans": [ + { + "bbox": [ + 130, + 510, + 479, + 581 + ], + "type": "text", + "content": "In practical pathology, the interpretation of models significantly influences the subsequent medical decisions and treatment strategies. Therefore, models with strong robustness are crucial for clinical applications. The quality of pathological slides can be affected by various factors during staining, scanning, and storage, including JPEG compression, pixelation, blur (e.g., bubble blur, defocus blur, and motion blur), and color variations (e.g., brightness, saturation, and hue)." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 130, + 582, + 479, + 640 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 582, + 479, + 640 + ], + "spans": [ + { + "bbox": [ + 130, + 582, + 479, + 640 + ], + "type": "text", + "content": "Drawing inspiration from the studies on the robustness of encoder-based models to pathology image corruptions [43, 56, 58], we incorporate these aforementioned types of corruptions into our analysis. To be specific, we simulate five levels of each corruption type on the PathMMU to explore the robustness of LMMs against these corruptions, as depicted in the left half of Fig. 4." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 131, + 641, + 479, + 665 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 131, + 641, + 479, + 665 + ], + "spans": [ + { + "bbox": [ + 131, + 641, + 479, + 665 + ], + "type": "text", + "content": "Results and Discussion. LMMs demonstrate notable robustness to various types and levels of image corruptions, yet their true robustness" + } + ] + } + ], + "index": 8 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "spans": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "type": "text", + "content": "10" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 166, + 91, + 219, + 100 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 166, + 91, + 219, + 100 + ], + "spans": [ + { + "bbox": [ + 166, + 91, + 219, + 100 + ], + "type": "text", + "content": "Y. Sun et al." + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 9 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 135, + 114, + 278, + 223 + ], + "blocks": [ + { + "bbox": [ + 135, + 114, + 278, + 223 + ], + "lines": [ + { + "bbox": [ + 135, + 114, + 278, + 223 + ], + "spans": [ + { + "bbox": [ + 135, + 114, + 278, + 223 + ], + "type": "image", + "image_path": "59332a6b7df98c041fbee5fd8a42a4b8d1bd2ee1b6d02a92a0ee7d0f6c399e7a.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 130, + 231, + 479, + 276 + ], + "lines": [ + { + "bbox": [ + 130, + 231, + 479, + 276 + ], + "spans": [ + { + "bbox": [ + 130, + 231, + 479, + 276 + ], + "type": "text", + "content": "Fig. 4: Left: Illustration of corrupted pathology images. Right: LMM's performance across various levels of color-related (brightness, hue, saturation) and image quality-related (pixelation, JPEG compression, bubble blur, motion blur, defocus blur) corruptions on the PathMMU test-tiny set, with level 0 representing the uncorrupted images." + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "image_caption" + } + ], + "index": 2 + }, + { + "type": "image", + "bbox": [ + 280, + 116, + 378, + 222 + ], + "blocks": [ + { + "bbox": [ + 280, + 116, + 378, + 222 + ], + "lines": [ + { + "bbox": [ + 280, + 116, + 378, + 222 + ], + "spans": [ + { + "bbox": [ + 280, + 116, + 378, + 222 + ], + "type": "image", + "image_path": "f87f47b50d448db9d0665be2e553f9cba32e39ef9892b286011374a5501a8260.jpg" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_body" + } + ], + "index": 3 + }, + { + "type": "image", + "bbox": [ + 379, + 116, + 477, + 220 + ], + "blocks": [ + { + "bbox": [ + 379, + 116, + 477, + 220 + ], + "lines": [ + { + "bbox": [ + 379, + 116, + 477, + 220 + ], + "spans": [ + { + "bbox": [ + 379, + 116, + 477, + 220 + ], + "type": "image", + "image_path": "912bfc972cbd9be60ab82e937ed09f2153ffddd3ca5b3d6db9766466beaf853a.jpg" + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "image_body" + } + ], + "index": 4 + }, + { + "type": "table", + "bbox": [ + 134, + 332, + 481, + 401 + ], + "blocks": [ + { + "bbox": [ + 130, + 289, + 480, + 323 + ], + "lines": [ + { + "bbox": [ + 130, + 289, + 480, + 323 + ], + "spans": [ + { + "bbox": [ + 130, + 289, + 480, + 323 + ], + "type": "text", + "content": "Table 2: Evaluation of the model's robustness on the PathMMU test-tiny set, where green indicates performance improvement and red signifies performance degradation compared to the model's performance with uncorrupted images." + } + ] + } + ], + "index": 6, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 134, + 332, + 481, + 401 + ], + "lines": [ + { + "bbox": [ + 134, + 332, + 481, + 401 + ], + "spans": [ + { + "bbox": [ + 134, + 332, + 481, + 401 + ], + "type": "table", + "html": "
BrightnessBubbleDefocusHueJPEGMotionPixelateSaturateOverall
Qwen-VL-7B32.7↑0.632.3↑0.232.1↓0.032.4↓0.332.3↑0.233.8↑1.733.1↑1.033.0↑0.932.7↑0.6
BLIP-2 FLAN-T5-XL32.9↑0.332.4↓0.232.8↑0.232.9↑0.332.4↓0.232.8↑0.232.6↑0.032.7↑0.132.7↓0.1
BLIP-2 FLAN-T5-XXL33.1↓0.233.2↓0.133.5↑0.233.7↑0.433.5↑0.233.1↓0.233.1↓0.233.6↑0.333.3↓0.0
InstructBLIP FLAN-T5-XL34.0↓0.934.1↓0.833.9↓1.034.1↓0.833.7↓1.234.0↓0.934.4↓0.533.8↓1.134.0↓0.9
InstructBLIP FLAN-T5-XXL34.0↓0.333.8↓0.534.2↓0.134.1↓0.234.3↓0.234.3↑0.034.1↓0.233.8↓0.534.1↓0.2
LLaVA-1.5-7B34.8↓0.135.3↑0.434.1↓0.834.3↓0.634.0↓0.934.3↓0.634.6↓0.334.7↓0.234.5↓0.4
LLaVA-1.5-13B38.9↑0.138.3↓0.538.3↓0.538.4↓0.438.0↓0.838.6↓0.238.7↓0.139.1↑0.338.5↓0.3
", + "image_path": "0a6411cc0fdc7ba3b488e32446441d7c9d648388db24832c29726e98527ef8e9.jpg" + } + ] + } + ], + "index": 7, + "angle": 0, + "type": "table_body" + } + ], + "index": 7 + }, + { + "type": "table", + "bbox": [ + 133, + 445, + 481, + 509 + ], + "blocks": [ + { + "bbox": [ + 130, + 414, + 479, + 436 + ], + "lines": [ + { + "bbox": [ + 130, + 414, + 479, + 436 + ], + "spans": [ + { + "bbox": [ + 130, + 414, + 479, + 436 + ], + "type": "text", + "content": "Table 3: Results of LMMs on the PathMMU test set with the original images substituted by random Gaussian noise images." + } + ] + } + ], + "index": 8, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 133, + 445, + 481, + 509 + ], + "lines": [ + { + "bbox": [ + 133, + 445, + 481, + 509 + ], + "spans": [ + { + "bbox": [ + 133, + 445, + 481, + 509 + ], + "type": "table", + "html": "
OverallPubMedSocialPathEduContentAtlasPathCLS
TinyALLTinyALLTinyAllTinyAllTinyALLTinyALL
Qwen-VL-7B28.8↓3.328.3↓3.239.5↑4.630.4↓2.524.7↓8.930.5↓5.327.8↓6.328.9↓4.727.4↓8.230.1↓3.720.3↑1.720.8↓0.9
BLIP-FLAN-T5-XL30.9↓1.730.8↓1.030.6↓5.333.1↓1.534.5↑0.433.4↓0.232.2↑1.531.9↓0.934.1↓0.535.1↓2.320.9↑0.620.8↓0.0
BLIP-FLAN-T5-XXL32.4↓0.931.6↓1.934.5↓2.534.6↓2.834.0↑1.732.5↓2.130.6↑0.432.8↑1.738.5↓0.938.1↓2.622.2↑2.420.8↑0.2
InstructBLIP FLAN-T5-XL31.1↓3.829.7↓2.132.4↓4.630.8↓2.432.8↓2.532.0↓1.933.7↓2.832.4↓0.932.7↓4.330.4↓6.321.5↓5.122.4↓0.7
InstructBLIP FLAN-T5-XXL30.4↓3.930.8↓3.132.0↓7.133.5↓3.728.9↑4.731.9↓2.432.2↑2.331.7↓4.335.1↑3.435.2↓4.121.5↓1.121.7↓1.0
LLAVA-1.5-7B30.2↓4.730.9↓4.537.0↓4.633.9↓6.031.5↓6.432.5↓5.628.2↓4.332.7↓3.829.8↓5.833.9↓5.320.9↓2.320.8↓1.1
LLAVA-1.5-13B32.7↓6.133.2↓4.438.4↓6.136.0↓5.035.3↓5.136.3↓4.129.0↓5.134.1↓5.335.6↓11.538.8↓5.522.0↓2.921.0↓2.5
", + "image_path": "a46eefe9d1117c46165e711f23bc49f327b756303fa2fc4c090b08e50e88f03e.jpg" + } + ] + } + ], + "index": 9, + "angle": 0, + "type": "table_body" + } + ], + "index": 9 + }, + { + "bbox": [ + 130, + 533, + 482, + 666 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 533, + 482, + 666 + ], + "spans": [ + { + "bbox": [ + 130, + 533, + 482, + 666 + ], + "type": "text", + "content": "is questionable, as shown in Tab. 2 and the right half of Fig. 4. Notably, the Qwen-VL-7B even shows a " + }, + { + "bbox": [ + 130, + 533, + 482, + 666 + ], + "type": "inline_equation", + "content": "0.6\\%" + }, + { + "bbox": [ + 130, + 533, + 482, + 666 + ], + "type": "text", + "content": " overall performance increase under corruption compared to its baseline. It is more plausible to hypothesize that these corruptions mainly alter minute details of the pathology image that are difficult for LMMs to discern (such as chromatin morphology of the nucleus or vacuolization of the cytoplasm). This limitation stems from their pre-training in general domains, where they learn to recognize larger and more prominent features (e.g., humans, houses, cars, etc.) rather than the nuanced details that are vital in pathology. Furthermore, they may even resort to exploiting spurious correlations as shortcuts to answer questions (e.g., deducing answers from textual patterns instead of focusing on essential image details.)" + } + ] + } + ], + "index": 10 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 402, + 91, + 447, + 100 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 402, + 91, + 447, + 100 + ], + "spans": [ + { + "bbox": [ + 402, + 91, + 447, + 100 + ], + "type": "text", + "content": "PathMMU" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 470, + 91, + 479, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 470, + 91, + 479, + 100 + ], + "spans": [ + { + "bbox": [ + 470, + 91, + 479, + 100 + ], + "type": "text", + "content": "11" + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 10 + }, + { + "para_blocks": [ + { + "bbox": [ + 130, + 116, + 482, + 260 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 116, + 482, + 260 + ], + "spans": [ + { + "bbox": [ + 130, + 116, + 482, + 260 + ], + "type": "text", + "content": "To substantiate our argument, we employ an extreme test of corruption by replacing the images with Gaussian noise-generated random images. As shown in Tab. 3, we discover that even when the images contain no relevant information, LMMs can still achieve results significantly better than random choice. The drop in performance when using these random images ranged from only " + }, + { + "bbox": [ + 130, + 116, + 482, + 260 + ], + "type": "inline_equation", + "content": "1.0\\%" + }, + { + "bbox": [ + 130, + 116, + 482, + 260 + ], + "type": "text", + "content": " to " + }, + { + "bbox": [ + 130, + 116, + 482, + 260 + ], + "type": "inline_equation", + "content": "4.5\\%" + }, + { + "bbox": [ + 130, + 116, + 482, + 260 + ], + "type": "text", + "content": ". This finding suggests that LMMs might rely on shortcuts to accomplish the multimodal task, such as depending exclusively on textual information to make predictions. Additionally, it is intriguing that the performance drop across different sizes of the same model is remarkably similar. In other words, the benefit that images provide to models of various model sizes appears to be consistent, suggesting that the improvement in performance for different-sized models may primarily stem from their LLM component, rather than the vision aspect." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 131, + 276, + 384, + 289 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 131, + 276, + 384, + 289 + ], + "spans": [ + { + "bbox": [ + 131, + 276, + 384, + 289 + ], + "type": "text", + "content": "4.3 Analysis of LLMs' Ability to Guess Answers" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 130, + 294, + 481, + 318 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 294, + 481, + 318 + ], + "spans": [ + { + "bbox": [ + 130, + 294, + 481, + 318 + ], + "type": "text", + "content": "In this section, we delve deeper into the phenomenon of LLM's ability to make educated guesses with more comprehensive experiments and analysis." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 130, + 319, + 481, + 413 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 319, + 481, + 413 + ], + "spans": [ + { + "bbox": [ + 130, + 319, + 481, + 413 + ], + "type": "text", + "content": "To more clearly demonstrate the guessing abilities of LLMs, we randomly select 100 samples that are filtered during the Q&A generation process and can be correctly guessed by multiple LLMs. We invite pathology experts to answer these questions with reference images. As shown in Fig. 5, despite having access to images, their performance is significantly lower than that of the closed-source LLMs, which achieve about " + }, + { + "bbox": [ + 130, + 319, + 481, + 413 + ], + "type": "inline_equation", + "content": "90\\%" + }, + { + "bbox": [ + 130, + 319, + 481, + 413 + ], + "type": "text", + "content": " accuracy, and only slightly better than the smaller open-source Vicuna-v1.5-13B. This indicates that LLMs may outperform humans in guessing correct answers by identifying shortcuts within the questions." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 130, + 415, + 482, + 533 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 415, + 482, + 533 + ], + "spans": [ + { + "bbox": [ + 130, + 415, + 482, + 533 + ], + "type": "text", + "content": "We hypothesize several key reasons for the guessing behavior exhibited by LLMs: (1) Frequency-based guessing: LLMs may guess based on the prevalence of certain options in real-world pathological instances. (2) Based on the options that present a pattern of one supporting and three contradicting, or vice versa: For example, when a question identifies a tumour diagnosis, one option may match the pathological traits of a lesion, while the others suggest non-lesional characteristics. (3) When only one option aligns with the question's subject: For example, if a question asks which feature in the image supports the diagnosis of a Low-grade Squamous Intraepithelial Lesion (LSIL) cell, with only one option describes LSIL's pathological characteristics." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 130, + 534, + 482, + 665 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 534, + 482, + 665 + ], + "spans": [ + { + "bbox": [ + 130, + 534, + 482, + 665 + ], + "type": "text", + "content": "To support our hypothesis, we expand the sample size to 1,600, with 400 from each source (except PathCLS), to empirically analyze how LLMs guess answers. Specifically, we swap the questions among these samples while keeping the options unchanged, creating samples where the questions and options are completely mismatched. As shown in the right half of Fig. 5, we observe that LLMs still manage to guess approximately " + }, + { + "bbox": [ + 130, + 534, + 482, + 665 + ], + "type": "inline_equation", + "content": "50\\%" + }, + { + "bbox": [ + 130, + 534, + 482, + 665 + ], + "type": "text", + "content": " cases correctly, significantly higher than random choice accuracy. This finding suggests that the models tend to select the most common or prominent option as the answer, supporting our hypotheses (1) and (2). To further explore hypothesis (3), we design experiments using BERT-large [12] and BiomedBERT-large [15], the latter being specifically pre-trained on biomedical data. Given that the BERT series incorporates Next-Sentence" + } + ] + } + ], + "index": 7 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "spans": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "type": "text", + "content": "12" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 166, + 91, + 219, + 100 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 166, + 91, + 219, + 100 + ], + "spans": [ + { + "bbox": [ + 166, + 91, + 219, + 100 + ], + "type": "text", + "content": "Y. Sun et al." + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 11 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 137, + 113, + 291, + 209 + ], + "blocks": [ + { + "bbox": [ + 137, + 113, + 291, + 209 + ], + "lines": [ + { + "bbox": [ + 137, + 113, + 291, + 209 + ], + "spans": [ + { + "bbox": [ + 137, + 113, + 291, + 209 + ], + "type": "image", + "image_path": "75c52ac8bc697d5ee19715be4da0962f7a25300525ce2189db23a99c036a280c.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 130, + 216, + 482, + 295 + ], + "lines": [ + { + "bbox": [ + 130, + 216, + 482, + 295 + ], + "spans": [ + { + "bbox": [ + 130, + 216, + 482, + 295 + ], + "type": "text", + "content": "Fig. 5: Left: The performance comparison between different LLMs and human experts on 100 filtered samples where the answer can be guessed through text-only. Right: Expand the sample size to 1600 to validate the source of LLMs' ability to guess answers, which includes: (1) randomly replacing the original questions with others from the dataset while keeping the options unchanged, and (2) utilizing the BERT series for answer selection, specifically through its Next Sentence Prediction (NSP), to assess whether an option is the sequential sentence following a question." + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "image_caption" + } + ], + "index": 2 + }, + { + "type": "table", + "bbox": [ + 296, + 114, + 482, + 206 + ], + "blocks": [ + { + "bbox": [ + 296, + 114, + 482, + 206 + ], + "lines": [ + { + "bbox": [ + 296, + 114, + 482, + 206 + ], + "spans": [ + { + "bbox": [ + 296, + 114, + 482, + 206 + ], + "type": "table", + "html": "
PubMed SocialPath EduContentAtlas
Random Choice25.824.325.525.825.3
GPT-4 Turbo93.091.886.883.888.8
+ Random Swap Question44.345.550.051.547.8
GPT-3.5 Turbo84.386.579.585.383.9
+ Random Swap Question44.044.553.851.348.4
Gemini Pro85.590.176.571.584.1
+ Random Swap Question43.844.850.351.347.5
ERNIE-Bot 4.094.592.389.382.089.5
+ Random Swap Question47.046.556.551.050.3
BERT-large w/ NSP24.536.037.534.333.1
BiomedBERT-large w/ NSP44.344.842.337.542.3
", + "image_path": "b0b589b8581cd5b594bbdcbf9fb0c7137863c945a09435698fec3daf2fe15e31.jpg" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "table_body" + } + ], + "index": 3 + }, + { + "bbox": [ + 130, + 316, + 482, + 437 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 316, + 482, + 437 + ], + "spans": [ + { + "bbox": [ + 130, + 316, + 482, + 437 + ], + "type": "text", + "content": "Prediction (NSP) in their pretraining, which inherently relies on the similarity between two sentences, we apply this to predict the relationship between questions and options. We select the option with the highest probability of being the correct next sentence to the question. Our findings reveal that both BERT and BiomedBERT substantially surpass random guessing, indicating that direct matching of questions and options is a viable method for models to guess the correct answers, thereby supporting hypothesis (3). Moreover, BiomedBERT-large demonstrates a notable performance enhancement compared to BERT-large. This suggests that pretraining on biomedical data equips the model with a broader understanding of pathological knowledge." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 132, + 454, + 261, + 467 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 454, + 261, + 467 + ], + "spans": [ + { + "bbox": [ + 132, + 454, + 261, + 467 + ], + "type": "text", + "content": "4.4 Fine-tuning Results" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 130, + 474, + 481, + 533 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 474, + 481, + 533 + ], + "spans": [ + { + "bbox": [ + 130, + 474, + 481, + 533 + ], + "type": "text", + "content": "To explore the adaptability of LMMs to the pathology domain, we select representative LMMs, InstructBLIP FLAN-T5-XL and InstructBLIP FLAN-T5-XXL for fine-tuning experiments. Our experiments consist of two parts: (1) training the LMMs to directly generate answers, and (2) fine-tuning the LMMs to generate a reasoning process before delivering the final answer." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 130, + 533, + 482, + 641 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 533, + 482, + 641 + ], + "spans": [ + { + "bbox": [ + 130, + 533, + 482, + 641 + ], + "type": "text", + "content": "Results and Discussion. All models exhibit significant improvements on the PathMMU test set after fine-tuning on its training set, as detailed in Tab. 4. Notably, during the fine-tuning for direct answer generation, InstructBLIP FLAN-T5-XL and InstructBLIP FLAN-T5-XXL achieve significant improvements of " + }, + { + "bbox": [ + 130, + 533, + 482, + 641 + ], + "type": "inline_equation", + "content": "21.5\\%" + }, + { + "bbox": [ + 130, + 533, + 482, + 641 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 130, + 533, + 482, + 641 + ], + "type": "inline_equation", + "content": "21.3\\%" + }, + { + "bbox": [ + 130, + 533, + 482, + 641 + ], + "type": "text", + "content": ", respectively. This substantial improvement enables them to outperform the current leading model, GPT-4V. While still lagging behind expert performance, these results demonstrate a trend toward approaching expert-level proficiency. This also reflects the effectiveness of PathMMU in enhancing the LMMs' abilities for pathology image analysis." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 130, + 641, + 481, + 665 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 641, + 481, + 665 + ], + "spans": [ + { + "bbox": [ + 130, + 641, + 481, + 665 + ], + "type": "text", + "content": "Unexpectedly, generating explanations before answers during fine-tuning does not yield improvements. Instead, we observe slight performance decreases of " + }, + { + "bbox": [ + 130, + 641, + 481, + 665 + ], + "type": "inline_equation", + "content": "0.3\\%" + } + ] + } + ], + "index": 9 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 402, + 91, + 447, + 100 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 402, + 91, + 447, + 100 + ], + "spans": [ + { + "bbox": [ + 402, + 91, + 447, + 100 + ], + "type": "text", + "content": "PathMMU" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 470, + 91, + 481, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 470, + 91, + 481, + 100 + ], + "spans": [ + { + "bbox": [ + 470, + 91, + 481, + 100 + ], + "type": "text", + "content": "13" + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 12 + }, + { + "para_blocks": [ + { + "type": "table", + "bbox": [ + 133, + 157, + 481, + 212 + ], + "blocks": [ + { + "bbox": [ + 132, + 114, + 480, + 148 + ], + "lines": [ + { + "bbox": [ + 132, + 114, + 480, + 148 + ], + "spans": [ + { + "bbox": [ + 132, + 114, + 480, + 148 + ], + "type": "text", + "content": "Table 4: Results of LMMs on the PathMMU test set after the fine-tuning. The 'w/ A' and 'w/ A&E' denote the model is fine-tuned to output the answer directly or to output the answer with an explanation for its answer." + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 133, + 157, + 481, + 212 + ], + "lines": [ + { + "bbox": [ + 133, + 157, + 481, + 212 + ], + "spans": [ + { + "bbox": [ + 133, + 157, + 481, + 212 + ], + "type": "table", + "html": "
OverallPubMedSocialPathEduContentAtlasPathCLS
TinyALLTinyALLTinyALLTinyAllTinyALLTinyALL
InstructBLIP FLAN-T5-XL34.931.837.033.235.333.936.533.337.036.726.623.1
+ fine-tune w/A55.7 ± 20.853.3 ± 21.553.7 ± 16.750.4 ± 17.252.8 ± 17.552.3 ± 18.456.5 ± 20.052.0 ± 18.756.2 ± 19.253.1 ± 16.461.0 ± 34.460.8 ± 37.7
+ fine-tune w/A&E54.9 ± 20.053.0 ± 21.255.5 ± 18.551.4 ± 18.251.9 ± 16.649.7 ± 15.855.3 ± 18.851.5 ± 18.253.4 ± 16.451.6 ± 14.959.3 ± 32.761.4 ± 38.3
InstructBLIP FLAN-T5-XXL34.333.939.137.233.634.334.536.038.539.322.622.7
+ fine-tune w/A56.8 ± 22.555.2 ± 21.355.2 ± 16.151.5 ± 14.359.6 ± 26.055.2 ± 20.958.4 ± 23.954.1 ± 18.150.5 ± 12.053.7 ± 14.461.0 ± 38.463.7 ± 41.0
+ fine-tune w/A&E51.0 ± 16.752.9 ± 19.048.8 ± 9.750.8 ± 13.655.3 ± 21.751.2 ± 16.952.2 ± 17.751.9 ± 15.944.2 ± 5.750.1 ± 10.854.8 ± 32.260.5 ± 37.8
", + "image_path": "8acc7e401da6292596b496a2de53603855cd1ff24b760d47ff73c9e3f84437c2.jpg" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "table_body" + } + ], + "index": 3 + }, + { + "bbox": [ + 132, + 231, + 480, + 291 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 231, + 480, + 291 + ], + "spans": [ + { + "bbox": [ + 132, + 231, + 480, + 291 + ], + "type": "text", + "content": "and " + }, + { + "bbox": [ + 132, + 231, + 480, + 291 + ], + "type": "inline_equation", + "content": "2.3\\%" + }, + { + "bbox": [ + 132, + 231, + 480, + 291 + ], + "type": "text", + "content": " compared to direct answer generation, respectively. We speculate that generating explanations is relatively more challenging, and incorporating it with answer generation might impede the models' capacity to generate correct answers. This finding raises an important question: how can we effectively leverage the interpretability information within PathMMU to enhance models' training?" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 133, + 308, + 218, + 319 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 308, + 218, + 319 + ], + "spans": [ + { + "bbox": [ + 133, + 308, + 218, + 319 + ], + "type": "text", + "content": "5 Conclusion" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 132, + 330, + 481, + 498 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 330, + 481, + 498 + ], + "spans": [ + { + "bbox": [ + 132, + 330, + 481, + 498 + ], + "type": "text", + "content": "In this study, we introduce PathMMU, the largest and highest-quality pathology benchmark to date, specifically crafted to evaluate the capabilities of LMMs in interpreting and reasoning with pathology images. The construction of PathMMU involves a meticulous data collection and curation process, supplemented by a strict manual review by seven professional pathologists to ensure its quality and professionalism. Moreover, we establish a human expert performance benchmark to quantify the gap between cutting-edge LMMs and human experts. Our experimental results reveal that advanced LMMs significantly lag behind on PathMMU, with these models demonstrating poor performance in observing details in pathology images and sometimes even neglecting visual information, highlighting a substantial gap in practical pathology application. However, LMMs demonstrate notable performance improvements after fine-tuning on PathMMU, even surpassing GPT-4V. While they do not achieve human expert-level performance, these LMMs show promising potential for analyzing pathology images." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 132, + 498, + 481, + 665 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 498, + 481, + 665 + ], + "spans": [ + { + "bbox": [ + 132, + 498, + 481, + 665 + ], + "type": "text", + "content": "Future Directions for Pathology LMMs: Our experience with PathMMU highlights key areas for development in pathology LMMs: (1) Current LMMs, which are primarily based on LLMs and fine-tuned in a lightweight manner, tend to over-rely on textual information while neglecting visual data. There is a significant need to explore training methodologies or model structures that better integrate visual and textual modalities. (2) There's a tendency for LMMs to take shortcuts, solving problems in a \"lazy\" manner. This necessitates the development of trustworthy models for real-world clinical applications. (3) Given that most current LMMs do not support multi-image inputs, PathMMU does not include a benchmark for processing multiple images. However, in practical scenarios, pathologists typically analyze samples at various magnifications and perspectives, underscoring the importance of developing LMMs capable of handling multi-image inputs. Finally, we believe PathMMU will catalyze significant advancements in the development of next-generation LMMs in pathology." + } + ] + } + ], + "index": 7 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "spans": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "type": "text", + "content": "14" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 166, + 91, + 219, + 100 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 166, + 91, + 219, + 100 + ], + "spans": [ + { + "bbox": [ + 166, + 91, + 219, + 100 + ], + "type": "text", + "content": "Y. Sun et al." + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 13 + }, + { + "para_blocks": [ + { + "bbox": [ + 133, + 114, + 246, + 129 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 114, + 246, + 129 + ], + "spans": [ + { + "bbox": [ + 133, + 114, + 246, + 129 + ], + "type": "text", + "content": "Acknowledgements" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 130, + 137, + 482, + 185 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 137, + 482, + 185 + ], + "spans": [ + { + "bbox": [ + 130, + 137, + 482, + 185 + ], + "type": "text", + "content": "This study was partially supported by the National Natural Science Foundation of China (Grant No.92270108), Zhejiang Provincial Natural Science Foundation of China (Grant No.XHD23F0201), the Research Center for Industries of the Future (RCIF) at Westlake University, and the Westlake Education Foundation." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 133, + 201, + 197, + 213 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 201, + 197, + 213 + ], + "spans": [ + { + "bbox": [ + 133, + 201, + 197, + 213 + ], + "type": "text", + "content": "References" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 134, + 223, + 480, + 665 + ], + "type": "list", + "angle": 0, + "index": 18, + "blocks": [ + { + "bbox": [ + 138, + 223, + 480, + 255 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 223, + 480, + 255 + ], + "spans": [ + { + "bbox": [ + 138, + 223, + 480, + 255 + ], + "type": "text", + "content": "1. Alayrac, J.B., Donahue, J., Luc, P., Miech, A., Barr, I., Hasson, Y., Lenc, K., Mensch, A., Millican, K., Reynolds, M., et al.: Flamingo: a visual language model for few-shot learning. In: NeurIPS. pp. 23716-23736 (2022)" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 138, + 256, + 480, + 287 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 256, + 480, + 287 + ], + "spans": [ + { + "bbox": [ + 138, + 256, + 480, + 287 + ], + "type": "text", + "content": "2. Aresta, G., Araújo, T., Kwok, S., Chennamsetty, S.S., Safwan, M., Alex, V., Marami, B., Prastawa, M., Chan, M., Donovan, M., et al.: Bach: Grand challenge on breast cancer histology images. Medical image analysis 56, 122-139 (2019)" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 138, + 288, + 480, + 331 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 288, + 480, + 331 + ], + "spans": [ + { + "bbox": [ + 138, + 288, + 480, + 331 + ], + "type": "text", + "content": "3. Arunachalam, H.B., Mishra, R., Daescu, O., Cederberg, K., Rakheja, D., Sengupta, A., Leonard, D., Hallac, R., Leavey, P.: Viable and necrotic tumor assessment from whole slide images of osteosarcoma using machine-learning and deep-learning models. PloS one 14(4), e0210706 (2019)" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 138, + 331, + 480, + 364 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 331, + 480, + 364 + ], + "spans": [ + { + "bbox": [ + 138, + 331, + 480, + 364 + ], + "type": "text", + "content": "4. Bai, J., Bai, S., Yang, S., Wang, S., Tan, S., Wang, P., Lin, J., Zhou, C., Zhou, J.: Qwen-vl: A frontier large vision-language model with versatile abilities. arXiv preprint arXiv:2308.12966 (2023)" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 138, + 364, + 480, + 385 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 364, + 480, + 385 + ], + "spans": [ + { + "bbox": [ + 138, + 364, + 480, + 385 + ], + "type": "text", + "content": "5. Bavishi, R., Elsen, E., Hawthorne, C., Nye, M., Odena, A., Somani, A., Taşırlar, S.: Introducing our multimodal models (2023), https://www.adept.ai/blog/fuyu-8b" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 138, + 385, + 480, + 427 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 385, + 480, + 427 + ], + "spans": [ + { + "bbox": [ + 138, + 385, + 480, + 427 + ], + "type": "text", + "content": "6. Ben Abacha, A., Sarrouti, M., Demner-Fushman, D., Hasan, S.A., Müller, H.: Overview of the vqa-med task at imagecef 2021: Visual question answering and generation in the medical domain. In: Proceedings of the CLEF 2021 Conference and Labs of the Evaluation Forum-working notes (2021)" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 138, + 428, + 480, + 460 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 428, + 480, + 460 + ], + "spans": [ + { + "bbox": [ + 138, + 428, + 480, + 460 + ], + "type": "text", + "content": "7. Borkowski, A.A., Bui, M.M., Thomas, L.B., Wilson, C.P., DeLand, L.A., Mastorides, S.M.: Lung and colon cancer histopathological image dataset (lc25000). arXiv preprint arXiv:1912.12142 (2019)" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 138, + 460, + 480, + 492 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 460, + 480, + 492 + ], + "spans": [ + { + "bbox": [ + 138, + 460, + 480, + 492 + ], + "type": "text", + "content": "8. Brown, T., Mann, B., Ryder, N., Subbiah, M., Kaplan, J.D., Dhariwal, P., Neelakantan, A., Shyam, P., Sastry, G., Askell, A., et al.: Language models are few-shot learners. In: NeurIPS. pp. 1877-1901 (2020)" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 138, + 492, + 480, + 524 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 492, + 480, + 524 + ], + "spans": [ + { + "bbox": [ + 138, + 492, + 480, + 524 + ], + "type": "text", + "content": "9. Cai, R., Song, Z., Guan, D., Chen, Z., Luo, X., Yi, C., Kot, A.: Benchlmm: Benchmarking cross-style visual capability of large multimodal models. arXiv preprint arXiv:2312.02896 (2023)" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 134, + 525, + 480, + 567 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 134, + 525, + 480, + 567 + ], + "spans": [ + { + "bbox": [ + 134, + 525, + 480, + 567 + ], + "type": "text", + "content": "10. Chiang, W.L., Li, Z., Lin, Z., Sheng, Y., Wu, Z., Zhang, H., Zheng, L., Zhuang, S., Zhuang, Y., Gonzalez, J.E., Stoica, I., Xing, E.P.: Vicuna: An open-source chatbot impressing gpt-4 with " + }, + { + "bbox": [ + 134, + 525, + 480, + 567 + ], + "type": "inline_equation", + "content": "90\\%" + }, + { + "bbox": [ + 134, + 525, + 480, + 567 + ], + "type": "text", + "content": " chatgpt quality (March 2023), https://lmsys.org/blog/2023-03-30-vicuna/" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 134, + 568, + 480, + 600 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 134, + 568, + 480, + 600 + ], + "spans": [ + { + "bbox": [ + 134, + 568, + 480, + 600 + ], + "type": "text", + "content": "11. Dai, W., Li, J., Li, D., Tiong, A.M.H., Zhao, J., Wang, W., Li, B., Fung, P., Hoi, S.: Instructclip: Towards general-purpose vision-language models with instruction tuning. arXiv preprint arXiv:2305.06500 (2023)" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 134, + 601, + 480, + 632 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 134, + 601, + 480, + 632 + ], + "spans": [ + { + "bbox": [ + 134, + 601, + 480, + 632 + ], + "type": "text", + "content": "12. Devlin, J., Chang, M.W., Lee, K., Toutanova, K.: Bert: Pre-training of deep bidirectional transformers for language understanding. In: NAACL. pp. 4171-4186 (2019)" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 134, + 632, + 480, + 665 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 134, + 632, + 480, + 665 + ], + "spans": [ + { + "bbox": [ + 134, + 632, + 480, + 665 + ], + "type": "text", + "content": "13. Driess, D., Xia, F., Sajjadi, M.S., Lynch, C., Chowdhery, A., Ichter, B., Wahid, A., Tompson, J., Vuong, Q., Yu, T., et al.: Palm-e: An embodied multimodal language model. In: ICML. pp. 8469-8488 (2023)" + } + ] + } + ], + "index": 17 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 402, + 91, + 447, + 100 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 402, + 91, + 447, + 100 + ], + "spans": [ + { + "bbox": [ + 402, + 91, + 447, + 100 + ], + "type": "text", + "content": "PathMMU" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 470, + 92, + 481, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 470, + 92, + 481, + 100 + ], + "spans": [ + { + "bbox": [ + 470, + 92, + 481, + 100 + ], + "type": "text", + "content": "15" + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 14 + }, + { + "para_blocks": [ + { + "bbox": [ + 132, + 116, + 481, + 665 + ], + "type": "list", + "angle": 0, + "index": 19, + "blocks": [ + { + "bbox": [ + 133, + 116, + 480, + 149 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 116, + 480, + 149 + ], + "spans": [ + { + "bbox": [ + 133, + 116, + 480, + 149 + ], + "type": "text", + "content": "14. Gao, P., Han, J., Zhang, R., Lin, Z., Geng, S., Zhou, A., Zhang, W., Lu, P., He, C., Yue, X., et al.: Llama-adapter v2: Parameter-efficient visual instruction model. arXiv preprint arXiv:2304.15010 (2023)" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 133, + 150, + 481, + 194 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 150, + 481, + 194 + ], + "spans": [ + { + "bbox": [ + 133, + 150, + 481, + 194 + ], + "type": "text", + "content": "15. Gu, Y., Tinn, R., Cheng, H., Lucas, M., Usuyama, N., Liu, X., Naumann, T., Gao, J., Poon, H.: Domain-specific language model pretraining for biomedical natural language processing. ACM Transactions on Computing for Healthcare (HEALTH) 3(1), 1-23 (2021)" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 132, + 195, + 480, + 239 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 195, + 480, + 239 + ], + "spans": [ + { + "bbox": [ + 132, + 195, + 480, + 239 + ], + "type": "text", + "content": "16. Guan, T., Liu, F., Wu, X., Xian, R., Li, Z., Liu, X., Wang, X., Chen, L., Huang, F., Yacoob, Y., et al.: Hallusionbench: an advanced diagnostic suite for entangled language hallucination and visual illusion in large vision-language models. In: CVPR. pp. 14375-14385 (2024)" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 132, + 240, + 480, + 273 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 240, + 480, + 273 + ], + "spans": [ + { + "bbox": [ + 132, + 240, + 480, + 273 + ], + "type": "text", + "content": "17. Han, C., Pan, X., Yan, L., Lin, H., Li, B., Yao, S., Lv, S., Shi, Z., Mai, J., Lin, J., et al.: Wss54luad: Grand challenge on weakly-supervised tissue semantic segmentation for lung adenocarcinoma. arXiv preprint arXiv:2204.06455 (2022)" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 132, + 274, + 480, + 295 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 274, + 480, + 295 + ], + "spans": [ + { + "bbox": [ + 132, + 274, + 480, + 295 + ], + "type": "text", + "content": "18. He, X., Zhang, Y., Mou, L., Xing, E., Xie, P.: Pathvqa: 30000+ questions for medical visual question answering. arXiv preprint arXiv:2003.10286 (2020)" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 132, + 297, + 480, + 328 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 297, + 480, + 328 + ], + "spans": [ + { + "bbox": [ + 132, + 297, + 480, + 328 + ], + "type": "text", + "content": "19. Huang, Z., Bianchi, F., Yuksekgonul, M., Montine, T.J., Zou, J.: A visual-language foundation model for pathology image analysis using medical twitter. Nature medicine 29(9), 2307-2316 (2023)" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 132, + 330, + 480, + 362 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 330, + 480, + 362 + ], + "spans": [ + { + "bbox": [ + 132, + 330, + 480, + 362 + ], + "type": "text", + "content": "20. Ikezogwo, W., Seyfioglu, S., Ghezloo, F., Geva, D., Sheikh Mohammed, F., Anand, P.K., Krishna, R., Shapiro, L.: Quilt-1m: One million image-text pairs for histopathology. In: NeurIPS. pp. 37995–38017 (2023)" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 132, + 364, + 480, + 384 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 364, + 480, + 384 + ], + "spans": [ + { + "bbox": [ + 132, + 364, + 480, + 384 + ], + "type": "text", + "content": "21. Kather, J.N., Halama, N., Marx, A.: 100,000 histological images of human colorectal cancer and healthy tissue. Zenodo10 5281 (2018)" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 132, + 386, + 480, + 430 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 386, + 480, + 430 + ], + "spans": [ + { + "bbox": [ + 132, + 386, + 480, + 430 + ], + "type": "text", + "content": "22. Kriegsmann, K., Lovers, F., Zgorzelski, C., Kriegsmann, J., Janssen, C., Meliss, R.R., Muley, T., Sack, U., Steinbuss, G., Kriegsmann, M.: Deep learning for the detection of anatomical tissue structures and neoplasms of the skin on scanned histopathological tissue sections. Frontiers in Oncology 12, 1022967 (2022)" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 132, + 431, + 480, + 452 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 431, + 480, + 452 + ], + "spans": [ + { + "bbox": [ + 132, + 431, + 480, + 452 + ], + "type": "text", + "content": "23. Kumar, V., Abbas, A.K., Fausto, N., Aster, J.C.: Robbins and Cotran pathologic basis of disease, professional edition e-book. Elsevier health sciences (2014)" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 132, + 453, + 480, + 485 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 453, + 480, + 485 + ], + "spans": [ + { + "bbox": [ + 132, + 453, + 480, + 485 + ], + "type": "text", + "content": "24. Lau, J.J., Gayen, S., Ben Abacha, A., Demner-Fushman, D.: A dataset of clinically generated visual questions and answers about radiology images. Scientific data 5(1), 1-10 (2018)" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 132, + 487, + 480, + 508 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 487, + 480, + 508 + ], + "spans": [ + { + "bbox": [ + 132, + 487, + 480, + 508 + ], + "type": "text", + "content": "25. Li, B., Zhang, Y., Chen, L., Wang, J., Yang, J., Liu, Z.: Otter: A multi-modal model with in-context instruction tuning. arXiv preprint arXiv:2305.03726 (2023)" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 132, + 510, + 480, + 542 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 510, + 480, + 542 + ], + "spans": [ + { + "bbox": [ + 132, + 510, + 480, + 542 + ], + "type": "text", + "content": "26. Li, B., Wang, R., Wang, G., Ge, Y., Ge, Y., Shan, Y.: benseed-bench: Benchmarking multimodal llms with generative comprehension. arXiv preprint arXiv:2307.16125 (2023)" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 132, + 544, + 480, + 575 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 544, + 480, + 575 + ], + "spans": [ + { + "bbox": [ + 132, + 544, + 480, + 575 + ], + "type": "text", + "content": "27. Li, C., Wong, C., Zhang, S., Usuyama, N., Liu, H., Yang, J., Naumann, T., Poon, H., Gao, J.: Llava-med: Training a large language-and-vision assistant for biomedicine in one day. In: NeurIPS. pp. 28541-28564 (2023)" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 132, + 577, + 480, + 609 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 577, + 480, + 609 + ], + "spans": [ + { + "bbox": [ + 132, + 577, + 480, + 609 + ], + "type": "text", + "content": "28. Li, C., Li, L., Jiang, H., Weng, K., Geng, Y., Li, L., Ke, Z., Li, Q., Cheng, M., Nie, W., et al.: Yolov6: A single-stage object detection framework for industrial applications. arXiv preprint arXiv:2209.02976 (2022)" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 132, + 610, + 480, + 642 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 610, + 480, + 642 + ], + "spans": [ + { + "bbox": [ + 132, + 610, + 480, + 642 + ], + "type": "text", + "content": "29. Li, J., Li, D., Savarese, S., Hoi, S.: Blip-2: Bootstrapping language-image pretraining with frozen image encoders and large language models. In: ICML. pp. 19730-19742 (2023)" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 132, + 643, + 480, + 665 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 643, + 480, + 665 + ], + "spans": [ + { + "bbox": [ + 132, + 643, + 480, + 665 + ], + "type": "text", + "content": "30. Liu, H., Li, C., Wu, Q., Lee, Y.J.: Visual instruction tuning. In: NeurIPS. pp. 34892-34916 (2023)" + } + ] + } + ], + "index": 18 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "spans": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "type": "text", + "content": "16" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 166, + 91, + 219, + 100 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 166, + 91, + 219, + 100 + ], + "spans": [ + { + "bbox": [ + 166, + 91, + 219, + 100 + ], + "type": "text", + "content": "Y. Sun et al." + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 15 + }, + { + "para_blocks": [ + { + "bbox": [ + 130, + 116, + 482, + 665 + ], + "type": "list", + "angle": 0, + "index": 20, + "blocks": [ + { + "bbox": [ + 130, + 116, + 482, + 149 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 116, + 482, + 149 + ], + "spans": [ + { + "bbox": [ + 130, + 116, + 482, + 149 + ], + "type": "text", + "content": "31. Liu, Y., Duan, H., Zhang, Y., Li, B., Zhang, S., Zhao, W., Yuan, Y., Wang, J., He, C., Liu, Z., et al.: Mmbench: Is your multi-modal model an all-around player? arXiv preprint arXiv:2307.06281 (2023)" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 130, + 150, + 481, + 172 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 150, + 481, + 172 + ], + "spans": [ + { + "bbox": [ + 130, + 150, + 481, + 172 + ], + "type": "text", + "content": "32. Liu, Z., Mao, H., Wu, C.Y., Feichtenhofer, C., Darrell, T., Xie, S.: A convnet for the 2020s. In: CVPR. pp. 11976-11986 (2022)" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 132, + 173, + 437, + 184 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 173, + 437, + 184 + ], + "spans": [ + { + "bbox": [ + 132, + 173, + 437, + 184 + ], + "type": "text", + "content": "33. OpenAI: Introducing chatgpt. https://openai.com/blog/chatgpt (2022)" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 132, + 185, + 307, + 196 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 185, + 307, + 196 + ], + "spans": [ + { + "bbox": [ + 132, + 185, + 307, + 196 + ], + "type": "text", + "content": "34. OpenAI: Gpt-4 technical report (2023)" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 132, + 197, + 481, + 218 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 197, + 481, + 218 + ], + "spans": [ + { + "bbox": [ + 132, + 197, + 481, + 218 + ], + "type": "text", + "content": "35. OpenAI: Gpt-4v(ision) system card. https://cdn.openai.com/papers/GPTV_System/Card.pdf (2023)" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 132, + 219, + 481, + 251 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 219, + 481, + 251 + ], + "spans": [ + { + "bbox": [ + 132, + 219, + 481, + 251 + ], + "type": "text", + "content": "36. Peng, Z., Wang, W., Dong, L., Hao, Y., Huang, S., Ma, S., Wei, F.: Kosmos-2: Grounding multimodal large language models to the world. arXiv preprint arXiv:2306.14824 (2023)" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 132, + 252, + 481, + 285 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 252, + 481, + 285 + ], + "spans": [ + { + "bbox": [ + 132, + 252, + 481, + 285 + ], + "type": "text", + "content": "37. Radford, A., Kim, J.W., Hallacy, C., Ramesh, A., Goh, G., Agarwal, S., Sastry, G., Askell, A., Mishkin, P., Clark, J., et al.: Learning transferable visual models from natural language supervision. In: ICML. pp. 8748-8763 (2021)" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 132, + 286, + 481, + 319 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 286, + 481, + 319 + ], + "spans": [ + { + "bbox": [ + 132, + 286, + 481, + 319 + ], + "type": "text", + "content": "38. Raffel, C., Shazeer, N., Roberts, A., Lee, K., Narang, S., Matena, M., Zhou, Y., Li, W., Liu, P.J.: Exploring the limits of transfer learning with a unified text-to-text transformer. The Journal of Machine Learning Research 21(1), 5485-5551 (2020)" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 132, + 319, + 481, + 352 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 319, + 481, + 352 + ], + "spans": [ + { + "bbox": [ + 132, + 319, + 481, + 352 + ], + "type": "text", + "content": "39. Seyfioglu, M.S., Ikezogwo, W.O., Ghezloo, F., Krishna, R., Shapiro, L.: Quilt-lava: Visual instruction tuning by extracting localized narratives from open-source histopathology videos. In: CVPR. pp. 13183-13192 (2024)" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 133, + 353, + 481, + 396 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 353, + 481, + 396 + ], + "spans": [ + { + "bbox": [ + 133, + 353, + 481, + 396 + ], + "type": "text", + "content": "40. Silva-Rodriguez, J., Colomer, A., Sales, M.A., Molina, R., Naranjo, V.: Going deeper through the gleason scoring scale: An automatic end-to-end system for histology prostate grading and cribriform pattern detection. Computer methods and programs in biomedicine 195, 105637 (2020)" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 133, + 397, + 481, + 430 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 397, + 481, + 430 + ], + "spans": [ + { + "bbox": [ + 133, + 397, + 481, + 430 + ], + "type": "text", + "content": "41. Sun, Y., Wang, S., Feng, S., Ding, S., Pang, C., Shang, J., Liu, J., Chen, X., Zhao, Y., Lu, Y., et al.: Ernie 3.0: Large-scale knowledge enhanced pre-training for language understanding and generation. arXiv preprint arXiv:2107.02137 (2021)" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 133, + 431, + 481, + 464 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 431, + 481, + 464 + ], + "spans": [ + { + "bbox": [ + 133, + 431, + 481, + 464 + ], + "type": "text", + "content": "42. Sun, Y., Zhang, Y., Si, Y., Zhu, C., Shui, Z., Zhang, K., Li, J., Lyu, X., Lin, T., Yang, L.: Pathgen-1.6m: 1.6 million pathology image-text pairs generation through multi-agent collaboration (2024), https://arxiv.org/abs/2407.00203" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 133, + 465, + 481, + 507 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 465, + 481, + 507 + ], + "spans": [ + { + "bbox": [ + 133, + 465, + 481, + 507 + ], + "type": "text", + "content": "43. Sun, Y., Zhu, C., Zhang, Y., Li, H., Chen, P., Yang, L.: Assessing the robustness of deep learning-assisted pathological image analysis under practical variables of imaging system. In: ICASSP. pp. 1-5 (2023). https://doi.org/10.1109/ICASSP49357.2023.10095887" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 133, + 509, + 481, + 542 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 509, + 481, + 542 + ], + "spans": [ + { + "bbox": [ + 133, + 509, + 481, + 542 + ], + "type": "text", + "content": "44. Sun, Y., Zhu, C., Zheng, S., Zhang, K., Sun, L., Shui, Z., Zhang, Y., Li, H., Yang, L.: Pathasst: A generative foundation ai assistant towards artificial general intelligence of pathology. In: AAAI. pp. 5034-5042 (2024)" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 133, + 543, + 481, + 575 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 543, + 481, + 575 + ], + "spans": [ + { + "bbox": [ + 133, + 543, + 481, + 575 + ], + "type": "text", + "content": "45. Team, G., Anil, R., Borgeaud, S., Wu, Y., Alayrac, J.B., Yu, J., Soricut, R., Schalkwyk, J., Dai, A.M., Hauth, A., et al.: Gemini: a family of highly capable multimodal models. arXiv preprint arXiv:2312.11805 (2023)" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 133, + 576, + 481, + 609 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 576, + 481, + 609 + ], + "spans": [ + { + "bbox": [ + 133, + 576, + 481, + 609 + ], + "type": "text", + "content": "46. Touvron, H., Lavril, T., Izacard, G., Martinet, X., Lachaux, M.A., Lacroix, T., Rozière, B., Goyal, N., Hambro, E., Azhar, F., et al.: Llama: Open and efficient foundation language models. arXiv preprint arXiv:2302.13971 (2023)" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 133, + 609, + 481, + 632 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 609, + 481, + 632 + ], + "spans": [ + { + "bbox": [ + 133, + 609, + 481, + 632 + ], + "type": "text", + "content": "47. Veeling, B.S., Linmans, J., Winkens, J., Cohen, T., Welling, M.: Rotation equivariant cnns for digital pathology. In: MICCAI. pp. 210-218. Springer (2018)" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 133, + 632, + 481, + 665 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 632, + 481, + 665 + ], + "spans": [ + { + "bbox": [ + 133, + 632, + 481, + 665 + ], + "type": "text", + "content": "48. Wang, J., Zhou, Y., Xu, G., Shi, P., Zhao, C., Xu, H., Ye, Q., Yan, M., Zhang, J., Zhu, J., et al.: Evaluation and analysis of hallucination in large vision-language models. arXiv preprint arXiv:2308.15126 (2023)" + } + ] + } + ], + "index": 19 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 402, + 91, + 447, + 100 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 402, + 91, + 447, + 100 + ], + "spans": [ + { + "bbox": [ + 402, + 91, + 447, + 100 + ], + "type": "text", + "content": "PathMMU" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 470, + 91, + 481, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 470, + 91, + 481, + 100 + ], + "spans": [ + { + "bbox": [ + 470, + 91, + 481, + 100 + ], + "type": "text", + "content": "17" + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 16 + }, + { + "para_blocks": [ + { + "bbox": [ + 132, + 116, + 481, + 566 + ], + "type": "list", + "angle": 0, + "index": 14, + "blocks": [ + { + "bbox": [ + 133, + 116, + 480, + 149 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 116, + 480, + 149 + ], + "spans": [ + { + "bbox": [ + 133, + 116, + 480, + 149 + ], + "type": "text", + "content": "49. Wang, W., Lv, Q., Yu, W., Hong, W., Qi, J., Wang, Y., Ji, J., Yang, Z., Zhao, L., Song, X., et al.: Cogvlm: Visual expert for pretrained language models. arXiv preprint arXiv:2311.03079 (2023)" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 132, + 150, + 481, + 205 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 150, + 481, + 205 + ], + "spans": [ + { + "bbox": [ + 132, + 150, + 481, + 205 + ], + "type": "text", + "content": "50. Wei, J., Suriawinata, A., Ren, B., Liu, X., Lisovsky, M., Vaickus, L., Brown, C., Baker, M., Tomita, N., Torresani, L., et al.: A petri dish for histopathology image analysis. In: Artificial Intelligence in Medicine: 19th International Conference on Artificial Intelligence in Medicine, AIME 2021, Virtual Event, June 15-18, 2021, Proceedings. pp. 11-24. Springer (2021)" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 132, + 205, + 480, + 237 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 205, + 480, + 237 + ], + "spans": [ + { + "bbox": [ + 132, + 205, + 480, + 237 + ], + "type": "text", + "content": "51. Xu, P., Shao, W., Zhang, K., Gao, P., Liu, S., Lei, M., Meng, F., Huang, S., Qiao, Y., Luo, P.: Lvlm-ehub: A comprehensive evaluation benchmark for large vision-language models. arXiv preprint arXiv:2306.09265 (2023)" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 132, + 238, + 480, + 270 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 238, + 480, + 270 + ], + "spans": [ + { + "bbox": [ + 132, + 238, + 480, + 270 + ], + "type": "text", + "content": "52. Yin, Z., Wang, J., Cao, J., Shi, Z., Liu, D., Li, M., Huang, X., Wang, Z., Sheng, L., BAI, L., Shao, J., Ouyang, W.: Lamm: Language-assisted multi-modal instruction-tuning dataset, framework, and benchmark. In: NeurIPS. pp. 26650-26685 (2023)" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 132, + 271, + 480, + 303 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 271, + 480, + 303 + ], + "spans": [ + { + "bbox": [ + 132, + 271, + 480, + 303 + ], + "type": "text", + "content": "53. Yu, W., Yang, Z., Li, L., Wang, J., Lin, K., Liu, Z., Wang, X., Wang, L.: Mmvet: Evaluating large multimodal models for integrated capabilities. arXiv preprint arXiv:2308.02490 (2023)" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 132, + 304, + 480, + 357 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 304, + 480, + 357 + ], + "spans": [ + { + "bbox": [ + 132, + 304, + 480, + 357 + ], + "type": "text", + "content": "54. Yue, X., Ni, Y., Zhang, K., Zheng, T., Liu, R., Zhang, G., Stevens, S., Jiang, D., Ren, W., Sun, Y., Wei, C., Yu, B., Yuan, R., Sun, R., Yin, M., Zheng, B., Yang, Z., Liu, Y., Huang, W., Sun, H., Su, Y., Chen, W.: Mmmu: A massive multi-discipline multimodal understanding and reasoning benchmark for expert agi. In: CVPR. pp. 9556-9567 (2024)" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 132, + 358, + 480, + 390 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 358, + 480, + 390 + ], + "spans": [ + { + "bbox": [ + 132, + 358, + 480, + 390 + ], + "type": "text", + "content": "55. Zhang, X., Wu, C., Zhao, Z., Lin, W., Zhang, Y., Wang, Y., Xie, W.: Pmc-vqa: Visual instruction tuning for medical visual question answering. arXiv preprint arXiv:2305.10415 (2023)" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 132, + 391, + 480, + 423 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 391, + 480, + 423 + ], + "spans": [ + { + "bbox": [ + 132, + 391, + 480, + 423 + ], + "type": "text", + "content": "56. Zhang, Y., Sun, Y., Li, H., Zheng, S., Zhu, C., Yang, L.: Benchmarking the robustness of deep neural networks to common corruptions in digital pathology. In: MICCAI. pp. 242-252 (2022)" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 132, + 424, + 480, + 456 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 424, + 480, + 456 + ], + "spans": [ + { + "bbox": [ + 132, + 424, + 480, + 456 + ], + "type": "text", + "content": "57. Zhang, Z., Chen, P., McGough, M., Xing, F., Wang, C., Bui, M., Xie, Y., Sapkota, M., Cui, L., Dhillon, J., et al.: Pathologist-level interpretable whole-slide cancer diagnosis with deep learning. Nature Machine Intelligence 1(5), 236-245 (2019)" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 132, + 457, + 480, + 489 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 457, + 480, + 489 + ], + "spans": [ + { + "bbox": [ + 132, + 457, + 480, + 489 + ], + "type": "text", + "content": "58. Zheng, S., Cui, X., Sun, Y., Li, J., Li, H., Zhang, Y., Chen, P., Jing, X., Ye, Z., Yang, L.: Benchmarking pathclip for pathology image analysis. Journal of Imaging Informatics in Medicine pp. 1-17 (2024)" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 132, + 490, + 480, + 533 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 490, + 480, + 533 + ], + "spans": [ + { + "bbox": [ + 132, + 490, + 480, + 533 + ], + "type": "text", + "content": "59. Zhu, C., Sun, Y., Li, H., Cui, C., Zhang, S., Cai, J., Ling, Y.: Weakly supervised classification using multi-level instance-aware optimization on cervical cytologic image. In: 2022 IEEE 19th International Symposium on Biomedical Imaging (ISBI). pp. 1-5. IEEE (2022)" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 132, + 533, + 480, + 566 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 533, + 480, + 566 + ], + "spans": [ + { + "bbox": [ + 132, + 533, + 480, + 566 + ], + "type": "text", + "content": "60. Zhu, D., Chen, J., Shen, X., Li, X., Elhoseiny, M.: Minigpt-4: Enhancing vision-language understanding with advanced large language models. arXiv preprint arXiv:2304.10592 (2023)" + } + ] + } + ], + "index": 13 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "spans": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "type": "text", + "content": "18" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 166, + 91, + 219, + 100 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 166, + 91, + 219, + 100 + ], + "spans": [ + { + "bbox": [ + 166, + 91, + 219, + 100 + ], + "type": "text", + "content": "Y. Sun et al." + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 17 + } + ], + "_backend": "vlm", + "_version_name": "2.6.4" +} \ No newline at end of file diff --git "a/2024/Pathformer3D_ A 3D Scanpath Transformer for 360\302\260 Images/8acb38cc-358d-422c-8280-ca63ea0a1521_content_list.json" "b/2024/Pathformer3D_ A 3D Scanpath Transformer for 360\302\260 Images/8acb38cc-358d-422c-8280-ca63ea0a1521_content_list.json" new file mode 100644 index 0000000000000000000000000000000000000000..584281ecf67d3d0c8b1979001f2597e9134bd9eb --- /dev/null +++ "b/2024/Pathformer3D_ A 3D Scanpath Transformer for 360\302\260 Images/8acb38cc-358d-422c-8280-ca63ea0a1521_content_list.json" @@ -0,0 +1,1808 @@ +[ + { + "type": "text", + "text": "Pathformer3D: A 3D Scanpath Transformer for $360^{\\circ}$ Images", + "text_level": 1, + "bbox": [ + 228, + 140, + 774, + 186 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Rong Quan $^{1}$ , Yantao Lai $^{1,2}$ , Mengyu Qiu $^{2}$ , and Dong Liang $^{2,3}$ (x̄)", + "bbox": [ + 261, + 210, + 740, + 229 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "1 College of Artificial Intelligence, Nanjing University of Aeronautics and Astronautics, the Key Laboratory of Brain-Machine Intelligence Technology, Ministry of Education, Nanjing, 211106, China rongquan0806@gmail.com, yantaolai@nuaa.edu.cn", + "bbox": [ + 215, + 238, + 785, + 294 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "$^{2}$ MIIT Key Laboratory of Pattern Analysis and Machine Intelligence, College of Computer Science and Technology, Nanjing University of Aeronautics and Astronautics, Nanjing, China {qmengyu, liangdong}@nuaa.edu.cn", + "bbox": [ + 228, + 295, + 772, + 349 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "3 Shenzhen Research Institute, Nanjing University of Aeronautics and Astronautics, Shenzhen, China", + "bbox": [ + 220, + 349, + 781, + 378 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Abstract. Scanpath prediction in $360^{\\circ}$ images can help realize rapid rendering and better user interaction in Virtual/Augmented Reality applications. However, existing scanpath prediction models for $360^{\\circ}$ images execute scanpath prediction on 2D equirectangular projection plane, which always result in big computation error owing to the 2D plane's distortion and coordinate discontinuity. In this work, we perform scanpath prediction for $360^{\\circ}$ images in 3D spherical coordinate system and proposed a novel 3D scanpath Transformer named Pathformer3D. Specifically, a 3D Transformer encoder is first used to extract 3D contextual feature representation for the $360^{\\circ}$ image. Then, the contextual feature representation and historical fixation information are input into a Transformer decoder to output current time step's fixation embedding, where the self-attention module is used to imitate the visual working memory mechanism of human visual system and directly model the time dependencies among the fixations. Finally, a 3D Gaussian distribution is learned from each fixation embedding, from which the fixation position can be sampled. Evaluation on four panoramic eye-tracking datasets demonstrates that Pathformer3D outperforms the current state-of-the-art methods. Code is available at https://github.com/lsztzp/Pathformer3D.", + "bbox": [ + 259, + 415, + 743, + 693 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Keywords: Scanpath $\\cdot$ $360^{\\circ}$ image $\\cdot$ 3D Transformer", + "bbox": [ + 261, + 707, + 620, + 722 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "1 Introduction", + "text_level": 1, + "bbox": [ + 215, + 747, + 375, + 763 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Virtual reality (VR) and augmented reality (AR) [50] technologies have made significant advancements in recent years, providing the users immersive experiences. Meanwhile, understanding and imitating the way human beings explore the $360^{\\circ}$ images of virtual environments is becoming more and more important,", + "bbox": [ + 212, + 779, + 787, + 840 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "since it can help realize more practical and fast rendering and thus improve user interaction in the immersive environment [24, 51]. Scanpath prediction on $360^{\\circ}$ images refers to predict human's gaze shift path when exploring $360^{\\circ}$ images.", + "bbox": [ + 212, + 146, + 784, + 191 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Existing research on scanpath prediction has primarily focused on 2D images [4, 22, 42, 46, 53], while $360^{\\circ}$ images possess distinct characteristics in the VR/AR context. $360^{\\circ}$ images provide immersive interactive environments where users can change their perspectives by physically moving their heads, resulting in a wider distribution of fixations. $360^{\\circ}$ images encompass richer visual information that requires more time for users to process and absorb. Furthermore, the data structure of $360^{\\circ}$ images differs from that of 2D images, introducing new requirements for data processing and analysis. Therefore, existing 2D image scanpath prediction methods cannot be directly applied to predict scanpaths in $360^{\\circ}$ images.", + "bbox": [ + 212, + 196, + 787, + 348 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Early methods [6, 43, 57] for predicting scanpaths in 360 images involved sampling fixations based on saliency information to obtain the entire scanpath. Subsequently, with the advancement of generative adversarial networks [2], some researchers [5,35] have utilized generative networks to directly generate the entire path from the $360^{\\circ}$ image. These methods have shown preliminary results but overlook the modeling of time dependency between fixations, which has long been demonstrated to be a very important characteristic of human visual attention mechanism [22], thereby they often lead to unstable prediction results. Recently, some research works [18,33,38,45] have recognized the need for a comprehensive treatment of the temporal dependencies in viewing behavior, and modeled them through recurrent neural networks (RNNs) or Markov chains. However, these methods integrate all historical fixation information into a single hidden unit, and generate the current fixation only from this hidden unit. In this case, the time dependent relationships among fixations can only be modeled indirectly and the influences from historical fixations will be greatly weakened. Besides, all above methods predict fixations on the 2D equirectangular projection of $360^{\\circ}$ image, which has the issue of coordinate discontinuity caused by longitude (where $-180^{\\circ}$ and $180^{\\circ}$ represent the same fixation despite their significant numerical difference), and different degrees of distortion in different locations. Consequently, the fixation predicted on this 2D equirectangular projection will have a large margin of error.", + "bbox": [ + 212, + 351, + 787, + 669 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "To solve the above problems, we propose a novel scanpath prediction model for $360^{\\circ}$ images named Pathformer3D, which predicts fixations in 3D spherical coordinates of the $360^{\\circ}$ image and directly model the time dependent relationships among fixations, to mimic human viewing behavior in immersive environment more realistically. Specifically, given a 2D equirectangular projection of $360^{\\circ}$ image, we first transform it into 3D spherical coordinate system, and then exploit a spherical convolution to extract its visual features. After extracting the 3D visual features, we exploit a 3D Transformer encoder to learn each $360^{\\circ}$ image region a contextual feature representation, by considering their long range spatial dependencies between each other. Next, we exploit a Transformer Decoder to learn each time step's fixation embedding directly from the whole image's", + "bbox": [ + 212, + 672, + 787, + 840 + ], + "page_idx": 1 + }, + { + "type": "page_number", + "text": "2", + "bbox": [ + 217, + 114, + 228, + 126 + ], + "page_idx": 1 + }, + { + "type": "header", + "text": "R. Quan et al.", + "bbox": [ + 271, + 114, + 367, + 128 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "visual feature and the historical fixations, where the self-attention module is used to mimic human visual working memory mechanism, which can model the time dependencies between the current fixation and all historical fixations. After obtaining each time step's fixation embedding, we employ a 3D mixture density network to learn the 3D Gaussian distribution for each fixation, from which the fixation location is sampled. Using 3D Gaussian distribution to model the positional likelihood of fixations in 3D space takes into account the differences in scanpaths of different individuals, and can thus result in more robust and real scanpaths.", + "bbox": [ + 212, + 146, + 787, + 281 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "In conclusion, this work has the following contributions:", + "bbox": [ + 238, + 282, + 643, + 297 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "- We perform visual scanpath prediction for $360^{\\circ}$ images in a 3D spherical coordinate system for the first time, which can realize a more real imitation of human visual exploration process in immersive environment and simultaneously avoid the error caused by the distortion and positional discontinuity in 2D equirectangular projection of $360^{\\circ}$ image.", + "bbox": [ + 225, + 305, + 782, + 378 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "- We propose a novel scanpath prediction method named Pathformer3D, to learn contextual feature representation of the $360^{\\circ}$ images in a 3D spherical coordinate system and directly consider the influences from historical fixations on current fixation, which is more in line with the visual working memory mechanism of human vision system.", + "bbox": [ + 225, + 380, + 784, + 455 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "- The proposed method is comprehensively evaluated on four eye-tracking datasets of $360^{\\circ}$ images, consistently achieving state-of-the-art performance, demonstrating the effectiveness of our approach.", + "bbox": [ + 225, + 455, + 784, + 500 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "2 Related Works", + "text_level": 1, + "bbox": [ + 215, + 522, + 395, + 537 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Scanpath Prediction for 2D Images In recent years, numerous models have been proposed for predicting scanpaths in 2D images. According to the different sources of inspiration, they can be categorized into different classes.", + "bbox": [ + 212, + 553, + 784, + 598 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Some of these models draw inspiration from neuroscience and vision science, often leveraging human gaze behavior to guide scanpath prediction. For instance, some researchers [17, 22, 46] have utilized Inhibit-of-Return(IOR) strategies to guide the generation of scanpaths. Additionally, Wang et al. [52] compute response maps of sparse coding filters applied to foveated images at each step and select new fixations from the residual perceptual information map(RPI) using the information maximization principle. Adeli et al. [1] first apply a retina transformation to input images, then compute corresponding priority maps and select the next fixation. Furthermore, some researchers [48, 56] attempt to generate new fixations by leveraging low-level semantic information.", + "bbox": [ + 212, + 598, + 785, + 748 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Another subset of work is inspired by the statistical distribution characteristics of human scanpaths. Many researchers [7,8,30,32,34,42,53] in this group attempt to use image saliency information to predict new scanpaths. For example, VSPT [42] uses SalGAN [39] to extract saliency information and employs an autoregressive approach [21] to generate fixations. Others in this category adopt alternative methods, such as Sun et al. [47], who conduct super-Gaussian quality", + "bbox": [ + 212, + 750, + 787, + 840 + ], + "page_idx": 2 + }, + { + "type": "header", + "text": "Pathformer3D: A 3D Scanpath Transformer for $360^{\\circ}$ Images", + "bbox": [ + 330, + 114, + 732, + 130 + ], + "page_idx": 2 + }, + { + "type": "page_number", + "text": "3", + "bbox": [ + 774, + 116, + 785, + 126 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "analysis on image patches and select the $k$ th fixation to be at the patch with the highest response to the $k$ th super Gaussian component. Coutrot et al. [13] model scanpaths using image and task dependent hidden markov models(HMM). Clarke et al. [11] implement a Gaussian jump distribution, which is polynomially dependent on the previous fixed distribution.", + "bbox": [ + 212, + 146, + 782, + 220 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "In addition to the aforementioned models, there are also some deep learning models [5, 28, 55] that have achieved relatively good performance, despite their limited interpretability.", + "bbox": [ + 212, + 222, + 784, + 267 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Scanpath Prediction for $360^{\\circ}$ Images Unlike 2D images, there are currently limited models specifically designed for scanpath prediction in $360^{\\circ}$ images. However, with the gradual development of VR/AR technologies, this field has attracted increasing attention. In this domain, SaltiNet [6] is the first proposed method, which uses saliency volume to process $360^{\\circ}$ images. PathGAN [5] adopts the method of Generative Adversarial [20] to generate scanpath of $360^{\\circ}$ image. Zhu et al. [58] proposed a new method to estimate the saliency of $360^{\\circ}$ images, and introduced mechanisms such as visual uncertainty and visual balance to predict the scanpath. The subsequent ScanGAN [35] is also a method that uses Generative Adversarial Networks [20] to generate scapath. Different from PathGAN [5], ScanGAN [35] adds a soft-dtw loss function [15] to the generator to generate scanpath that are more similar to real scanpath. ScanDMM [45] is a method that uses Deep Markov model [27] to generate scanpath, which takes into account the past historical information.", + "bbox": [ + 212, + 287, + 785, + 498 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "All above methods predict fixations as well as scanpaths in 2D equirectangular projection of $360^{\\circ}$ image, which has the weaknesses of distortion and coordinate discontinuity and always introduce big prediction error. Different from them, we perform scanpath prediction for $360^{\\circ}$ images in 3D spherical coordinate system, which is more consistent with the real-world scenarios of human exploring immersive environments.", + "bbox": [ + 212, + 500, + 784, + 589 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "3 Approach", + "text_level": 1, + "bbox": [ + 215, + 609, + 336, + 628 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "3.1 Problem Definition", + "text_level": 1, + "bbox": [ + 215, + 638, + 419, + 652 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "A scanpath can be represented as a sequence of fixations $X = (p_{1}, p_{2}, \\ldots, p_{T}) \\in \\mathbb{R}^{T \\times 3}$ , where $T$ represents the length of the scanpath. For a given $360^{\\circ}$ image, the task of a scanpath prediction model is to generate a scanpath $\\widehat{X}$ that resembles a real human scanpath. We use a 3D Cartesian coordinate system to represent coordinates, where each fixation is given in the form of $p = (x, y, z)$ , with $x$ , $y$ , and $z$ being real numbers ranging from -1 to 1. This representation form helps address the issue of coordinate discontinuity.", + "bbox": [ + 212, + 661, + 784, + 767 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "3.2 Overview", + "text_level": 1, + "bbox": [ + 215, + 787, + 339, + 801 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "The architecture of our model is shown in Fig. 1. For each input $360^{\\circ}$ image, we first transform it into 3D spherical coordinate system, and then extract the", + "bbox": [ + 212, + 809, + 784, + 839 + ], + "page_idx": 3 + }, + { + "type": "page_number", + "text": "4", + "bbox": [ + 217, + 114, + 228, + 126 + ], + "page_idx": 3 + }, + { + "type": "header", + "text": "R. Quan et al.", + "bbox": [ + 271, + 114, + 367, + 128 + ], + "page_idx": 3 + }, + { + "type": "image", + "img_path": "images/874335bb1b481e953e4d5ee6aec4ebfffad20fb561e97e054a4f218b81b583c5.jpg", + "image_caption": [ + "Fig. 1: Overall architecture of our methods. The Transformer Encoder is utilized to contextualize the feature extracted by the SphereNet [12]. The contextual feature representation and the historical fixations are fed into the Transformer Decoder to output fixation embeddings. Each fixation embedding is input into a 3D Mixture Density Network to output its 3D Gaussian distribution, from which the fixation is sampled." + ], + "image_footnote": [], + "bbox": [ + 220, + 143, + 782, + 383 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "image features using SphereNet [12]. Subsequently, we feed the extracted feature representation into a 3D Transformer Encoder to contextualize them. The Transformer Decoder takes historical fixation information and contextual feature representations as input and generates embedding information for new fixations in an autoregressive manner [21]. Then, a 3D Mixture Density Network (MDN) takes the embedding information of the fixations as input and outputs their 3D Gaussian position distributions. Finally, the fixations are sampled from the 3D Gaussian distributions. Next, we will introduce a detailed module-wise description of our model.", + "bbox": [ + 212, + 496, + 787, + 632 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "3.3 3D Contextual Visual Feature Representation", + "text_level": 1, + "bbox": [ + 214, + 659, + 638, + 674 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "3D feature extraction. We perform scanpath prediction in 3D spherical coordinate system. Therefore, after projecting the $360^{\\circ}$ image into 3D spherical coordinate system, we utilize SphereNet [12] to extract its 3D feature representation. Specifically, for a $360^{\\circ}$ image $I$ , after inputting it into SphereNet, we extract 3 different scales of feature maps and concatenate them along the channel axis, resulting in a feature representation $F$ with dimensions of $C = 448$ , $H = 128$ , $W = 256$ .", + "bbox": [ + 212, + 688, + 787, + 777 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Before inputting into the 3D Transformer Encoder, we first execute a simple size transformation to $F$ . Firstly, we execute average pooling to $F$ in a convolutional kernel of size $8 \\times 8$ , to reduce the size of the feature maps to $C = 448$ , $H = 16$ , $W = 32$ . Next, we flatten the pooled feature maps, resulting", + "bbox": [ + 212, + 779, + 787, + 840 + ], + "page_idx": 4 + }, + { + "type": "header", + "text": "Pathformer3D: A 3D Scanpath Transformer for $360^{\\circ}$ Images", + "bbox": [ + 331, + 114, + 732, + 128 + ], + "page_idx": 4 + }, + { + "type": "page_number", + "text": "5", + "bbox": [ + 774, + 116, + 784, + 126 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "in a feature $F_{L}$ with size of $L = 16 \\times 32 = 512$ , $C = 448$ . This series of operations can be formulated as:", + "bbox": [ + 212, + 146, + 784, + 175 + ], + "page_idx": 5 + }, + { + "type": "equation", + "text": "\n$$\nF = \\operatorname {S p h e r e N e t} (I, \\theta) \\tag {1}\n$$\n", + "text_format": "latex", + "bbox": [ + 413, + 186, + 785, + 202 + ], + "page_idx": 5 + }, + { + "type": "equation", + "text": "\n$$\nF _ {L} = \\operatorname {F l a t t e n} (\\operatorname {A v g P o o l} (F)) \\tag {2}\n$$\n", + "text_format": "latex", + "bbox": [ + 405, + 205, + 785, + 220 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "here, $\\theta$ represents the parameters of SphereNet.", + "bbox": [ + 214, + 229, + 560, + 244 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "3D feature encoder. A linear embedding layer is first applied to map $F_{L}$ into feature embeddings. Subsequently, the obtained feature embeddings are added to a 3D positional encoding and fed into $M = 4$ standard Transformer Encoder Layers to extract contextual feature representations. The above process can be expressed by the following formulas.", + "bbox": [ + 212, + 266, + 784, + 340 + ], + "page_idx": 5 + }, + { + "type": "equation", + "text": "\n$$\nF _ {L} ^ {\\prime} = \\operatorname {E m b e d d i n g} \\left(F _ {L}\\right) \\tag {3}\n$$\n", + "text_format": "latex", + "bbox": [ + 303, + 351, + 784, + 367 + ], + "page_idx": 5 + }, + { + "type": "equation", + "text": "\n$$\nP _ {0} = F _ {L} ^ {\\prime} + E _ {\\text {p o s} 3 d} \\quad E _ {\\text {p o s} 3 d} \\in R ^ {L \\times D} \\tag {4}\n$$\n", + "text_format": "latex", + "bbox": [ + 307, + 371, + 784, + 388 + ], + "page_idx": 5 + }, + { + "type": "equation", + "text": "\n$$\nP _ {i} = \\operatorname {E n c o d e r L a y e r} \\left(P _ {i - 1}\\right) \\quad i = 1 \\dots M \\tag {5}\n$$\n", + "text_format": "latex", + "bbox": [ + 308, + 391, + 784, + 407 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "where $D$ represents the feature embedding dimension of the 3D Transformer Encoder, $M$ represents the number of EncoderLayer in the architecture, and $P_{i}$ represents the output of the i-th EncoderLayer.", + "bbox": [ + 212, + 415, + 784, + 460 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "3.4 Fixation Sequence Generation", + "text_level": 1, + "bbox": [ + 214, + 481, + 511, + 496 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Fixation decoder. The contextual feature representation and historical fixation information are input into a Transformer decoder to output the fixation embedding. The fixation embedding is then fed into a 3D mixture density network to output the fixation's 3D Gaussian distribution, from which the fixation is sampled. We use autoregressive method [21] to predict one fixation each time and iteratively generate the whole fixation sequence.", + "bbox": [ + 212, + 503, + 784, + 595 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "On time step $t$ , we initialize the fixation query vector $q_{t}$ using fixation $p_{t-1}$ of the previous time step. Specially, for the first fixation query $q_{1}$ , we initialize it using the center of 3D space, i.e., $p_{0} = (0,0,0)$ . Subsequently, the initialized query vector $q_{t}$ is concatenated with the historical query vectors to generate the joint query vector $Q_{0}'$ . After adding one-dimensional positional encoding information, $Q_{0}'$ becomes $Q_{0}$ . Finally, after passing through $N = 4$ standard Transformer Decoder Layers, we obtain the hidden states $Q_{N} = (Z_{1}, Z_{2}, \\ldots, Z_{t})$ of the fixations. This process is defined as follows:", + "bbox": [ + 212, + 595, + 785, + 715 + ], + "page_idx": 5 + }, + { + "type": "equation", + "text": "\n$$\nq _ {t} = \\operatorname {E m b e d d i n g} \\left(p _ {t - 1}\\right) \\quad t \\in [ 1, 2, \\dots , T ] \\tag {6}\n$$\n", + "text_format": "latex", + "bbox": [ + 300, + 727, + 784, + 742 + ], + "page_idx": 5 + }, + { + "type": "equation", + "text": "\n$$\nQ _ {0} ^ {\\prime} = \\left[ q _ {1}; q _ {2}; \\dots ; q _ {t} \\right] \\tag {7}\n$$\n", + "text_format": "latex", + "bbox": [ + 295, + 744, + 784, + 761 + ], + "page_idx": 5 + }, + { + "type": "equation", + "text": "\n$$\nQ _ {0} = Q _ {0} ^ {\\prime} + E _ {p o s \\_ 1 d} \\quad E _ {p o s \\_ 1 d} \\in R ^ {t \\times D} \\tag {8}\n$$\n", + "text_format": "latex", + "bbox": [ + 294, + 763, + 784, + 782 + ], + "page_idx": 5 + }, + { + "type": "equation", + "text": "\n$$\nQ _ {i} = \\operatorname {D e c o d e r L a y e r} \\left(Q _ {i - 1}, P _ {M}\\right) \\quad i = 1 \\dots N \\tag {9}\n$$\n", + "text_format": "latex", + "bbox": [ + 295, + 785, + 784, + 800 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "where $P_{M}$ represents the output of the Transformer Encoder, $T$ represents the length of the predicted fixation sequence, $p_{t-1}$ represents the fixation at the time", + "bbox": [ + 212, + 809, + 785, + 840 + ], + "page_idx": 5 + }, + { + "type": "page_number", + "text": "6", + "bbox": [ + 217, + 114, + 228, + 126 + ], + "page_idx": 5 + }, + { + "type": "header", + "text": "R. Quan et al.", + "bbox": [ + 271, + 114, + 369, + 128 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "step $t - 1$ , $N$ represents the number of decoder layers. $Q_{i}$ represents the output of the i-th DecoderLayer.", + "bbox": [ + 212, + 146, + 782, + 176 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "The fixation decoder process in the training process has some difference from that in the generation process. During the training phase, we input the fixations $(p_0,p_1,\\ldots ,p_{T - 1})\\in R^{T\\times 3}$ into the Transformer Decoder, and it can predict the hidden states $Q_{N} = (Z_{1},Z_{2},\\dots,Z_{T})$ of the true scanpath $(p_1,p_2,\\dots,p_T)\\in$ $R^{T\\times 3}$ in a parallel manner, which is a parallel process. During the generation phase, we iteratively generate the hidden states of new coordinates $Z_{1},Z_{2},\\ldots Z_{T}$ in sequence starting from the initial coordinate $p_0$ , which is a sequential process.", + "bbox": [ + 212, + 178, + 787, + 285 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Fixation generation. After obtaining $Z_{t}$ , we then employ a 3D mixture density network to model the probability distribution of the fixations. The 3D mixture density network decodes the hidden state $Z_{t}$ of the fixation at time step $t$ into $K$ sets of Gaussian kernel parameters. We use $\\mu_t^i = (\\mu_{x,t}^i, \\mu_{y,t}^i, \\mu_{z,t}^i)$ , $\\Sigma_t^i$ , and $\\pi_t^i$ to represent the mean vector, covariance matrix, and weight of the $i$ -th Gaussian kernel at time step $t$ , respectively. These three parameters can be obtained through three separate MLP layers, which can be defined as:", + "bbox": [ + 212, + 311, + 787, + 417 + ], + "page_idx": 6 + }, + { + "type": "equation", + "text": "\n$$\n\\left\\{\\mu_ {t} ^ {i}, \\Sigma_ {t} ^ {i}, \\pi_ {t} ^ {i} \\right\\} = f _ {m d n} \\left(Z _ {t}; \\theta_ {m d n}\\right) \\quad i = 1 \\dots K \\tag {10}\n$$\n", + "text_format": "latex", + "bbox": [ + 313, + 431, + 785, + 450 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "where $\\theta_{mdn}$ refers to the parameters of the mixture density network.", + "bbox": [ + 212, + 460, + 709, + 477 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Next, the probability density function for a 3D Gaussian distribution $P(p_{t})$ can be defined as follows:", + "bbox": [ + 212, + 477, + 785, + 506 + ], + "page_idx": 6 + }, + { + "type": "equation", + "text": "\n$$\n\\mathcal {N} \\left(p _ {t} \\mid \\mu_ {t} ^ {i}, \\Sigma_ {t} ^ {i}\\right) = \\frac {\\exp \\left(- \\frac {1}{2} \\left(p _ {t} - \\mu_ {t} ^ {i}\\right) ^ {\\top} \\left(\\Sigma_ {t} ^ {i}\\right) ^ {- 1} \\left(p _ {t} - \\mu_ {t} ^ {i}\\right)\\right)}{\\sqrt {(2 \\pi) ^ {3} \\left| \\Sigma_ {t} ^ {i} \\right|}} \\tag {11}\n$$\n", + "text_format": "latex", + "bbox": [ + 310, + 518, + 785, + 555 + ], + "page_idx": 6 + }, + { + "type": "equation", + "text": "\n$$\nP \\left(p _ {t}\\right) = \\sum_ {i = 1} ^ {K} \\pi_ {t} ^ {i} \\mathcal {N} \\left(p _ {t} \\mid \\mu_ {t} ^ {i}, \\Sigma_ {t} ^ {i}\\right) \\tag {12}\n$$\n", + "text_format": "latex", + "bbox": [ + 370, + 559, + 785, + 599 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "where the $|\\Sigma_t^i|$ means the determinant of the covariance matrix $\\Sigma_t^i$ and the $\\mathcal{N}(p_t\\mid \\mu_t^i,\\Sigma_t^i)$ represents the Gaussian distribution function associated with the $i$ -th 3D Gaussian function.", + "bbox": [ + 212, + 612, + 785, + 657 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "The prediction fixation $\\hat{p}_t$ is sampled probabilistically from a 3D surface, which can be expressed as:", + "bbox": [ + 212, + 659, + 785, + 690 + ], + "page_idx": 6 + }, + { + "type": "equation", + "text": "\n$$\n\\hat {p} _ {t} \\sim P \\left(p _ {t}\\right) \\tag {13}\n$$\n", + "text_format": "latex", + "bbox": [ + 464, + 705, + 785, + 720 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "here, it is worth noting that we do not simply sample the fixation with the highest probability, since such sampling strategy always makes the model rapidly converge to the same fixation positions for all images. In our study, we generated a total of 32,768 fixation points along the latitude-longitude grid, with a resolution of $128 \\times 256$ . We independently computed the probability for each fixation point. Finally, we randomly selected a fixation point based on its probability, with higher probabilities being more likely to be chosen.", + "bbox": [ + 212, + 734, + 787, + 840 + ], + "page_idx": 6 + }, + { + "type": "header", + "text": "Pathformer3D: A 3D Scanpath Transformer for $360^{\\circ}$ Images", + "bbox": [ + 330, + 114, + 732, + 130 + ], + "page_idx": 6 + }, + { + "type": "page_number", + "text": "7", + "bbox": [ + 774, + 114, + 785, + 126 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "3.5 Loss Function", + "text_level": 1, + "bbox": [ + 215, + 146, + 377, + 160 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "During training, the objective is to minimize the discrepancy between the predicted distribution and the real human scanpath. To achieve this, we employ the negative log-likelihood loss function, which encourages higher probabilities for the ground truth fixations. The loss function can be defined as follows:", + "bbox": [ + 212, + 170, + 787, + 231 + ], + "page_idx": 7 + }, + { + "type": "equation", + "text": "\n$$\n\\mathcal {L} = - \\frac {1}{T} \\sum_ {t = 1} ^ {T} \\log \\left(\\sum_ {i = 1} ^ {K} \\pi_ {t} ^ {i} \\mathcal {N} \\left(p _ {t} ^ {*} \\mid \\mu_ {t} ^ {i}, \\Sigma_ {t} ^ {i}\\right)\\right) \\tag {14}\n$$\n", + "text_format": "latex", + "bbox": [ + 362, + 242, + 785, + 282 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "where $p_t^*$ represents the real human fixation position of time step $t$ .", + "bbox": [ + 215, + 292, + 702, + 308 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "4 Experiments", + "text_level": 1, + "bbox": [ + 215, + 330, + 375, + 348 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "4.1 Datasets", + "text_level": 1, + "bbox": [ + 215, + 361, + 334, + 375 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "We conducted experiments on four $360^{\\circ}$ image datasets: Sitzmann [44], Salient360! [43], AOI [54], and JUFE [19]. The Sitzmann dataset [44] comprises 22 images with a total of 1920 scanpaths. Only the training set of the Salient360! dataset [43] is available, which contains 85 images and 3036 scanpaths. The AOI dataset [54] consists of 600 high-resolution images and approximately 18,000 scanpaths. The JUFE dataset [19] includes 1032 images and 30,960 scanpaths. For training, 19 images from Sitzmann [44] and 60 images from Salient360! [43] were utilized, and all remaining data were used for validation. To increase the number of training sets, we rotated the image along with the corresponding scanpaths six times along the direction of longitude. We followed ScanDMM [45] to preprocess the gaze data of Sitzmann, Salient360!, and JUFE datasets, and obtain visual scanpaths containing 30, 25, and 15 fixations for Sitzmann, Salient360!, and JUFE. The AOI dataset has already been pre-processed, with an average scanpath length of approximately 8.", + "bbox": [ + 212, + 385, + 823, + 598 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "4.2 Evaluation Metrics", + "text_level": 1, + "bbox": [ + 215, + 619, + 419, + 633 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "We employed six metrics to assess the performances of our model and other models, including Levenshtein Distance (LEV) [41], Dynamic Time Warping (DTW) [25], Time Delay Embedding (TDE), ScanMatch [14], Recursive Metric (REC) [3] and Sequence Score(SS) [55]. LEV [41] is a metric that quantifies the dissimilarity between two strings by measuring the minimum number of single-character edit operations needed to transform one string into another. ScanMatch [14] involves converting eye-tracking scanpaths into string sequences and then employing an algorithm similar to the Levenshtein distance to compare the similarity of the two sequences. DTW [25] and TDE [52] are metrics based on time series analysis, while REC [3] is a metric for recursive quantification analysis. SS [55] is calculated by converting scanpaths into strings of fixation cluster IDs, and then measuring the similarity between two strings using a string matching algorithm [37].", + "bbox": [ + 212, + 643, + 787, + 842 + ], + "page_idx": 7 + }, + { + "type": "page_number", + "text": "8", + "bbox": [ + 217, + 114, + 228, + 126 + ], + "page_idx": 7 + }, + { + "type": "header", + "text": "R. Quan et al.", + "bbox": [ + 271, + 114, + 369, + 128 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "For each image, every model generates $m = 10$ predicted scanpaths $\\widehat{X}$ , which are evaluated against all real human scanpaths $X$ , and then averaged to determine the final prediction score. Note that our model predicts scanpath with a length of 30. When comparing across different datasets, we truncate the predicted scanpath to the length of the actual scanpath before conducting the comparison. Taking DTW [25] as an example, the score can be calculated as follows:", + "bbox": [ + 212, + 145, + 787, + 238 + ], + "page_idx": 8 + }, + { + "type": "equation", + "text": "\n$$\nD T W = \\frac {1}{n} \\frac {1}{m} \\sum_ {i = 1} ^ {n} \\sum_ {j = 1} ^ {m} D T W \\left(\\widehat {X} _ {j}, X _ {i}\\right) \\tag {15}\n$$\n", + "text_format": "latex", + "bbox": [ + 375, + 250, + 785, + 289 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "where $\\widehat{X}_j$ is the $j$ -th predicted scanpath, $X_i$ is the $i$ -th real human scanpath, $n$ is the number of the real human scanpaths, $m$ is the number of predicted scanpaths.", + "bbox": [ + 212, + 297, + 785, + 347 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "4.3 Implementation Details", + "text_level": 1, + "bbox": [ + 214, + 369, + 455, + 383 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "The internal vector dimension of 3D Transformer Encoder is set to $D = 128$ . We utilize a multi-head attention mechanism with 8 attention heads. The hidden layer size of the feed-forward network (FFN) is set to 64. In the 3D mixture density network model, we configure $K = 5$ Gaussian kernels, and the internal hidden layer dimension is set to 16. During the training process, our batch size is set to 18. As for the optimizer, we use AdamW [26] with a learning rate of 1e-5. We employ a learning rate warm-up and stage-wise adjustment strategy. The warm-up period lasts for 10 epochs, and the learning rate is halved every 10 epochs. The total training duration is 50 epochs.", + "bbox": [ + 212, + 395, + 787, + 530 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "Our experiments were conducted on a single NVIDIA GeForce RTX 3090, and the development was conducted using the PyTorch framework.", + "bbox": [ + 212, + 532, + 785, + 563 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "4.4 Performance Comparison", + "text_level": 1, + "bbox": [ + 214, + 585, + 477, + 602 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "We compared our model with SaliTiNet [6], ScanGAN [35], and ScanDMM [45], utilizing the models provided officially by their respective sources to obtain prediction results. These three models are all designed for scanpath prediction on $360^{\\circ}$ images. PathGAN [5] and Zhu et al. [58] are two additional models specifically designed for generating scanpath on $360^{\\circ}$ images. However, it has been demonstrated that these models are unable to effectively generate realistic scanpaths [35]. Therefore, we have not included them in our comparison.", + "bbox": [ + 212, + 612, + 787, + 718 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "We additionally established two baseline reference values to assess the relative quality of the generated scanpath. The first one is the random baseline (Random walk). This baseline randomly generate ten scanpaths for each image, which provides the reference for lowest performance. The second one is the human baseline (Human), which compares each real human scanpath in the dataset with all the other real human scanpaths and taking the average. This baseline represents the average similarity among real human scanpaths, serving as a reference for highest performance.", + "bbox": [ + 212, + 719, + 787, + 840 + ], + "page_idx": 8 + }, + { + "type": "header", + "text": "Pathformer3D: A 3D Scanpath Transformer for $360^{\\circ}$ Images", + "bbox": [ + 330, + 114, + 732, + 130 + ], + "page_idx": 8 + }, + { + "type": "page_number", + "text": "9", + "bbox": [ + 774, + 116, + 785, + 126 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "Quantitative evaluation. Tab. 1 shows the quantitative comparison results between our method and other competitors on the metrics of LEV [41], DTW [25], TDE [52], ScanMatch [14], REC [3] and SS [55]. As we can see, our method outperforms all the competitors on all four datasets, achieving performance close to 'Human'.", + "bbox": [ + 212, + 146, + 787, + 220 + ], + "page_idx": 9 + }, + { + "type": "table", + "img_path": "images/0b1be78a80a0d6ca4d9b10abe5303baaa87ed717968657719a121151258b7beb.jpg", + "table_caption": [ + "Table 1: Performance of scanpath prediction models on different datasets. The best result for each metric is highlighted in bold, while the second-best result is underlined." + ], + "table_footnote": [], + "table_body": "
MethodSitzmannSalient360!
LEV ↓DTW ↓TDE ↓ScanMatch ↑REC ↑SS ↑LEV ↓DTW ↓TDE ↓ScanMatch ↑REC ↑SS ↑
Random walk51.5222372.38728.0380.3951.5140.19843.6702062.08029.7530.3751.3810.176
SaltiNet [6]51.1512246.05726.3030.3741.7340.19841.0421860.39725.0010.4292.2500.220
ScanGAN [35]45.8521950.48819.0600.4772.9600.26538.9771752.37119.5270.4603.0410.244
ScanDMM [45]45.7481939.98019.0980.4853.1560.27038.5451706.66120.2380.4673.1250.257
Ours44.6911939.66219.8870.4864.1540.28036.2771515.88319.2610.4844.2500.302
Human42.0061841.53416.1920.5265.1630.36635.0901376.43916.6500.5135.0660.377
MethodAOIJUFE
LEV ↓DTW ↓TDE ↓ScanMatch ↑REC ↑SS ↑LEV ↓DTW ↓TDE ↓ScanMatch ↑REC ↑SS ↑
Random walk15.866666.65238.1640.2935.6500.24126.4391270.89338.1640.3521.8470.191
SaltiNet [6]14.978622.07935.7240.3376.3060.24526.1151264.49232.2610.3641.9720.192
ScanGAN [35]14.473569.54932.6020.3617.3790.25824.2591110.82625.9370.4313.2080.217
ScanDMM [45]14.185569.17032.5580.3717.3280.24923.3721081.46525.7870.4523.5270.221
Ours13.904561.89530.9870.3768.2860.22522.5951054.22322.9550.4624.4490.218
Human14.317501.95230.2480.36713.0080.37718.4451030.97020.3050.5647.8640.330
", + "bbox": [ + 218, + 286, + 787, + 483 + ], + "page_idx": 9 + }, + { + "type": "text", + "text": "Qualitative evaluation. We visualized the prediction results of all methods on one image of each dataset, as shown in Fig. 2. Similarly, we visualized the ground truth of human scanpath accordingly. The transition from purple to red in the scanpath indicates the beginning and end of the scanpath, respectively. It is noticeable that most ground truth scanpaths are located within the salient regions near the equator. SaltiNet [6], although designed for $360^{\\circ}$ images, suffers from large displacements. In comparison, the methods proposed by ScanGAN [35] and ScanDMM [45] have the capability to generate relatively realistic scanpaths to some extent. However, our model, in contrast to these approaches, excels at capturing salient information within $360^{\\circ}$ images and is capable of generating scanpath that are closer to the real human scanpaths. In addition, a notable flaw in the tower image is that our model tends to focus more on the prominent plants at both ends rather than the central tower, which is a common issue shared by other models as well.", + "bbox": [ + 212, + 535, + 787, + 746 + ], + "page_idx": 9 + }, + { + "type": "text", + "text": "4.5 Saliency Comparison", + "text_level": 1, + "bbox": [ + 214, + 768, + 436, + 786 + ], + "page_idx": 9 + }, + { + "type": "text", + "text": "A good scanpath prediction model often assigns more attention to salient regions during the prediction process. In our study, we generated 1000 scanpath and performed post-processing on all fixations to generate corresponding saliency", + "bbox": [ + 212, + 794, + 785, + 842 + ], + "page_idx": 9 + }, + { + "type": "page_number", + "text": "10", + "bbox": [ + 217, + 114, + 235, + 126 + ], + "page_idx": 9 + }, + { + "type": "header", + "text": "R. Quan et al.", + "bbox": [ + 271, + 114, + 367, + 128 + ], + "page_idx": 9 + }, + { + "type": "image", + "img_path": "images/f4aad0f1d160d12dc5b8a0092434cb1cafde543cb2325ee239228189b10f8503.jpg", + "image_caption": [ + "Fig. 2: Qualitative representation results of different models on four datasets. From top to bottom, the four images are sourced from Sitzmann [44], Salient360! [43], AOI [54], and JUFE [19], respectively. From left to right, we display the a scanpath sampled from ground truth, our model, ScanDMM [45], ScanGAN [35], and SaltiNet [6]." + ], + "image_footnote": [], + "bbox": [ + 218, + 143, + 782, + 325 + ], + "page_idx": 10 + }, + { + "type": "text", + "text": "maps [49]. The fixation maps was obtained from all unprocessed fixations. We used sizes $\\mathrm{H} = 128$ and $\\mathrm{W} = 256$ for saliency maps (salmap) and fixation maps (fixmap). We compared our results with ScanGAN [35] and ScanDMM [45] on the Salient360! [43] and JUFE [19] datasets, under the saliency metrics AUC_Judd [9], NSS [40], CC [31], SIM [9], and KLD [29].", + "bbox": [ + 212, + 426, + 787, + 505 + ], + "page_idx": 10 + }, + { + "type": "table", + "img_path": "images/02b0aa9f9062aef40b24c7970fe205b2714663a1be8ea6158b5d45b3578b920b.jpg", + "table_caption": [ + "Table 2: Quantitative comparison results of significance comparison on Salient360! [43] and JUFE [19] datasets. The best result for each metric is highlighted in bold, while the second-best result is underlined." + ], + "table_footnote": [], + "table_body": "
DatasetMethodAUC_Judd↑NSS ↑CC ↑SIM ↑KLD ↓
Salient360! [43]ScanGAN [35]0.7970.9860.2600.17321.210
ScanDMM [45]0.8301.2100.3220.18320.882
Ours0.8381.1950.3520.25918.985
JUFE [19]ScanGAN [35]0.7950.8370.1980.23619.529
ScanDMM [45]0.8661.2030.2750.28618.248
Ours0.8721.0350.2450.32016.339
", + "bbox": [ + 277, + 587, + 730, + 688 + ], + "page_idx": 10 + }, + { + "type": "text", + "text": "Based on Tab. 2, our model has demonstrated favorable performance across the majority of metrics. Our model did not achieve optimal results in a few metrics, one potential explanation is that our model exhibits a stronger emphasis on larger salient objects, which may lead to comparatively less attention towards smaller targets compared to models with more evenly distributed saliency maps. As evident in Fig. 3, ScanDMM [45] and ScanGAN [35] exhibit a characteristic of having more dispersed saliency maps. Furthermore, it can be observed that our model has the ability to focus more on salient regions compared to the other", + "bbox": [ + 212, + 719, + 787, + 840 + ], + "page_idx": 10 + }, + { + "type": "header", + "text": "Pathformer3D: A 3D Scanpath Transformer for $360^{\\circ}$ Images", + "bbox": [ + 331, + 114, + 732, + 128 + ], + "page_idx": 10 + }, + { + "type": "page_number", + "text": "11", + "bbox": [ + 767, + 116, + 784, + 126 + ], + "page_idx": 10 + }, + { + "type": "image", + "img_path": "images/441392e9cb3366db93419d574e62c506a078fcae62b96e2f7d159564535b5718.jpg", + "image_caption": [ + "Fig. 3: Qualitative comparison results of significance comparison. The first two images is from Salient360! [43] and the next three images are from JUFE [19]. From top to bottom are the real image, the significance maps of the real images, Ours, ScanDMM [45], ScanGAN [35]." + ], + "image_footnote": [], + "bbox": [ + 222, + 146, + 785, + 363 + ], + "page_idx": 11 + }, + { + "type": "table", + "img_path": "images/9a2e81fe431336051a04a08e779d44158b7f7180f3ea876194c7132eadfeca37.jpg", + "table_caption": [ + "Table 3: Efficiency comparison in terms of model size and running time" + ], + "table_footnote": [], + "table_body": "
SaltiNet [6]ScanGAN [35]ScanDMM [45]Ours
Paramters(MB)98.7315.13417.77919.280
Running Time(ms)127206.482118.250266.420
", + "bbox": [ + 277, + 468, + 730, + 522 + ], + "page_idx": 11 + }, + { + "type": "text", + "text": "models. As a result, the post-processed saliency maps obtained from our model exhibit a closer resemblance to human-generated saliency maps.", + "bbox": [ + 212, + 545, + 787, + 575 + ], + "page_idx": 11 + }, + { + "type": "text", + "text": "4.6 Efficiency Comparison", + "text_level": 1, + "bbox": [ + 214, + 595, + 447, + 609 + ], + "page_idx": 11 + }, + { + "type": "text", + "text": "We also compared our model's size and running time for generating a scanpath with other methods. In Tab. 3, we present the model size and the runtime for generating a single scanpath (measured in milliseconds). As can be seen, we have a similar model size with other models. In terms of runtime, our model is capable of generating approximately four scanpaths of length 30 per second, which is a little slower compared to ScanGAN and ScanDMM. This is primarily because our model generates fixations in an autoregressive manner [21]. Additionally, the use of a mixture density network is another factor leading to slower running speed. However, compared to a single Gaussian distribution, multiple Gaussian distributions can more accurately represent the distribution of fixations.", + "bbox": [ + 212, + 617, + 787, + 768 + ], + "page_idx": 11 + }, + { + "type": "text", + "text": "4.7 Ablation Studies", + "text_level": 1, + "bbox": [ + 214, + 787, + 401, + 801 + ], + "page_idx": 11 + }, + { + "type": "text", + "text": "We conducted ablation experiments on the Salient360! [43] and JUFE [19] datasets. The experimental results are presented in Tab. 4, showing the optimal results", + "bbox": [ + 212, + 809, + 795, + 840 + ], + "page_idx": 11 + }, + { + "type": "page_number", + "text": "12", + "bbox": [ + 217, + 114, + 235, + 126 + ], + "page_idx": 11 + }, + { + "type": "header", + "text": "R. Quan et al.", + "bbox": [ + 271, + 114, + 367, + 128 + ], + "page_idx": 11 + }, + { + "type": "text", + "text": "obtained for each configuration. Next, we will provide a brief overview of each model architecture. For more detailed architecture information, please refer to the supplementary file.", + "bbox": [ + 212, + 146, + 787, + 193 + ], + "page_idx": 12 + }, + { + "type": "table", + "img_path": "images/b6086fcad6ce57e1922b4d1baf67a6f05ec0d7bb3e07ae7d195f597dfca98d1d.jpg", + "table_caption": [ + "Table 4: Ablation studies on Salient360! [43] and JUFE [19] datasets. The best result for each metric is highlighted in bold, while the second-best result is underlined." + ], + "table_footnote": [], + "table_body": "
MethodSalient360!JUFE
LEV ↓DTW ↓TDE ↓ScanMatch ↑REC ↑SS ↑LEV ↓DTW ↓TDE ↓ScanMatch ↑REC ↑SS ↑
Pure ViT37.6431739.71419.2910.4703.7100.26722.8911072.75323.7510.4573.8470.221
Pure 2D CNN36.8481727.24119.6410.4774.1190.27222.6131060.20523.0960.4634.1810.220
Saliency38.1961864.91420.0590.4753.1560.25323.8331094.29024.6360.4423.6690.208
w/o EncoderLayer37.3521798.34020.7270.4634.1710.26222.9261072.26323.1960.4554.2600.218
w/o MDN + MSE Loss39.4631793.94825.9140.4263.7810.09523.9151132.12725.6750.4163.8330.102
K=138.6431804.07420.0740.4623.3260.24823.2391084.73423.5770.4513.8050.218
K=336.2502073.90522.3260.4904.0390.24622.7581077.36523.3020.4604.3600.214
K=838.1171700.87720.4550.4673.3890.25822.9171065.24023.4700.4553.9140.210
Parallel36.7221577.99323.7260.4893.9390.28822.5671054.64428.7750.4513.6880.221
Ours36.2771515.88319.2610.4844.2530.30222.5951054.22322.9550.4624.4490.218
", + "bbox": [ + 218, + 260, + 787, + 398 + ], + "page_idx": 12 + }, + { + "type": "text", + "text": "The feature extractor. To demonstrate the effectiveness of our 3D feature extractor, SphereNet, we try to employ three distinct architectures for feature extraction. Firstly, we extracted features using the patch-based approach of the Vision Transformer (ViT) method [16] (denoted as 'Pure ViT'). Secondly, we replaced SphereNet with a standard 2D convolutional network (denoted as 'Pure 2D CNN'). Lastly, We followed VSPT [42] to extract image features based on the saliency information. Specifically, we employed EPSNet [59] to extract saliency information (denoted as 'Saliency'). As shown in Tab. 4, 'Ours' obtains better performance than 'Pure ViT', 'Pure 2D CNN', and 'Saliency', which demonstrates the superiority of our feature extractor.", + "bbox": [ + 212, + 454, + 787, + 607 + ], + "page_idx": 12 + }, + { + "type": "text", + "text": "The 3D Transformer encoder. To validate the effectiveness of our 3D Transformer Encoder, we attempted to remove the EncoderLayer in the 3D Transformer Encoder and directly utilize the feature embedding of the Transformer Encoder as input of the Transformer Decoder (denoted as 'w/o EncoderLayer'). As we can see from Tab. 4, without our 3D Transformer Encoder, the scanpath prediction performance dropped significantly, which demonstrates the effectiveness of our 3D Transformer Encoder.", + "bbox": [ + 212, + 632, + 787, + 739 + ], + "page_idx": 12 + }, + { + "type": "text", + "text": "3D MDN. To validate the effectiveness of the 3D MDN, we directly exploited linear regression to predict the fixations after the Transformer Decoder. The Mean Squared Error (MSE) was utilized as the loss function (denoted as 'w/o MDN + MSE Loss'). As can be seen in Tab. 4, 'Ours' outperforms 'w/o MDN + MSE Loss' a lot, proving the advantage of our 3D MDN.", + "bbox": [ + 212, + 763, + 787, + 840 + ], + "page_idx": 12 + }, + { + "type": "header", + "text": "Pathformer3D: A 3D Scanpath Transformer for $360^{\\circ}$ Images", + "bbox": [ + 330, + 113, + 730, + 130 + ], + "page_idx": 12 + }, + { + "type": "page_number", + "text": "13", + "bbox": [ + 767, + 114, + 785, + 126 + ], + "page_idx": 12 + }, + { + "type": "text", + "text": "Different number of Gaussian kernels. We compared with the model results obtained by using different numbers of Gaussian kernels in the mixture density network. These comparisons are respectively presented in the rows labeled \"K=1,3,8,\" while our model has 5 Gaussian kernels. As we can see from Tab. 4, using 5 Gaussian kernels can provide a better representation of the fixations' distribution.", + "bbox": [ + 212, + 146, + 787, + 238 + ], + "page_idx": 13 + }, + { + "type": "text", + "text": "Parallel in the Transformer decoder. In our Transformer decoder, we use autoregressive method to predict one fixation each time and iteratively generate the whole fixation sequence. Through our observation, more and more models try to adopt a parallel structure in the Transformer decoder [10, 23, 36] for higher efficiency. Therefore, we also tried to build a parallel Transformer decoder in our method (denoted as 'Parallel'). In the parallel model, we use a random embedding with length $L$ as the input to the Transformer decoder, and we can get the hidden state of a scanpath of length $L$ at one time, all the other parts of the model are unchanged.", + "bbox": [ + 212, + 256, + 789, + 391 + ], + "page_idx": 13 + }, + { + "type": "text", + "text": "We can observe from Tab. 4 that the parallel structure of the decoder performs poorly than the autoregressive Transformer decoder, especially in the TDE [52] metric. Such comparison results demonstrate that an autoregressive Transformer decoder is more suitable for our scanpath prediction task.", + "bbox": [ + 212, + 393, + 789, + 455 + ], + "page_idx": 13 + }, + { + "type": "text", + "text": "5 Limitaions and Future Works", + "text_level": 1, + "bbox": [ + 214, + 474, + 539, + 491 + ], + "page_idx": 13 + }, + { + "type": "text", + "text": "In our work, we employed SphereNet [12] for feature extraction and 3D Mixture Density Network (MDN) for fixation generation. However, we acknowledge that these components are replaceable, and we are contemplating whether there might be better alternatives for both parts.", + "bbox": [ + 212, + 503, + 787, + 563 + ], + "page_idx": 13 + }, + { + "type": "text", + "text": "In addition, we are considering incorporating relevant semantic information to improve the performance further, as objects with important semantic information tend to attract human attention more easily.", + "bbox": [ + 212, + 564, + 787, + 609 + ], + "page_idx": 13 + }, + { + "type": "text", + "text": "6 Conclusion", + "text_level": 1, + "bbox": [ + 214, + 628, + 362, + 646 + ], + "page_idx": 13 + }, + { + "type": "text", + "text": "Considering that the 2D equirectangular projection of $360^{\\circ}$ images has the issues of distortion and coordinate discontinuity, we proposed to execute scanpath prediction for $360^{\\circ}$ images in 3D spherical coordinate system. Correspondingly, we proposed a novel 3D Scanpath Transformer named Pathformer3D, where a 3D Transformer Encoder was first used to obtain 3D contextual feature representation of the $360^{\\circ}$ image. Then, Pathformer3D input the 3D contextual feature representation and historical fixation information into the Transformer Decoder to predict the hidden state for each fixation. Each fixation's hidden state was then fed into a 3D mixture density network to output the 3D Gaussian distribution of the fixation, from which the fixation location can be sampled. Finally, Pathformer3D model was compared with three state-of-the-art $360^{\\circ}$ image scan-path prediction models, to demonstrate the superiority of its performance.", + "bbox": [ + 212, + 657, + 789, + 840 + ], + "page_idx": 13 + }, + { + "type": "page_number", + "text": "14", + "bbox": [ + 217, + 114, + 235, + 126 + ], + "page_idx": 13 + }, + { + "type": "header", + "text": "R. Quan et al.", + "bbox": [ + 271, + 114, + 367, + 128 + ], + "page_idx": 13 + }, + { + "type": "text", + "text": "Acknowledgements This work is supported in part by the National Natural Science Foundation of China under grant 62206127 and 62272229, the Natural Science Foundation of Jiangsu Province under grant BK20222012, and Shenzhen Science and Technology Program JCYJ20230807142001004. Rong Quan and Yantao Lai contributed equally to this work. Dong Liang is the corresponding author.", + "bbox": [ + 212, + 146, + 787, + 238 + ], + "page_idx": 14 + }, + { + "type": "text", + "text": "References", + "text_level": 1, + "bbox": [ + 217, + 262, + 323, + 277 + ], + "page_idx": 14 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "1. Adeli, H., Vitu, F., Zelinsky, G.J.: A model of the superior colliculus predicts fixation locations during scene viewing and visual search. Journal of Neuroscience 37(6), 1453-1467 (2017)", + "2. Aggarwal, A., Mittal, M., Battineni, G.: Generative adversarial network: An overview of theory and applications. International Journal of Information Management Data Insights 1(1), 100004 (2021)", + "3. Anderson, N.C., Bischof, W.F., Laidlaw, K.E., Risko, E.F., Kingstone, A.: Recurrence quantification analysis of eye movements. Behavior research methods 45, 842-856 (2013)", + "4. Antol, S., Agrawal, A., Lu, J., Mitchell, M., Batra, D., Zitnick, C.L., Parikh, D.: Vqa: Visual question answering. In: ICCV. pp. 2425-2433 (2015)", + "5. Assens, M., Giro-i Nieto, X., McGuinness, K., O'Connor, N.E.: Pathgan: Visual scanpath prediction with generative adversarial networks. In: Proceedings of the European Conference on Computer Vision (ECCV) Workshops. pp. 0-0 (2018)", + "6. Assens Reina, M., Giro-i Nieto, X., McGuinness, K., O'Connor, N.E.: Saltinet: Scan-path prediction on 360 degree images using saliency volumes. In: Proceedings of the IEEE International Conference on Computer Vision Workshops. pp. 2331-2338 (2017)", + "7. Boccignone, G., Ferraro, M.: Modelling gaze shift as a constrained random walk. Physica A: Statistical Mechanics and its Applications 331(1-2), 207-218 (2004)", + "8. Brockmann, D., Geisel, T.: The ecology of gaze shifts. Neurocomputing 32, 643-650 (2000)", + "9. Bylinskii, Z., Judd, T., Borji, A., Itti, L., Durand, F., Oliva, A., Torralba, A.: Mit saliency benchmark (2015)", + "0. Carion, N., Massa, F., Synnaeve, G., Usunier, N., Kirillov, A., Zagoruyko, S.: End-to-end object detection with transformers. In: European conference on computer vision. pp. 213-229. Springer (2020)", + "1. Clarke, A.D., Stainer, M.J., Tatler, B.W., Hunt, A.R.: The saccadic flow baseline: Accounting for image-independent biases in fixation behavior. Journal of vision 17(11), 12-12 (2017)", + "2. Coors, B., Condurache, A.P., Geiger, A.: Spherenet: Learning spherical representations for detection and classification in omnidirectional images. In: Proceedings of the European conference on computer vision (ECCV). pp. 518-533 (2018)", + "3. Coutrot, A., Hsiao, J.H., Chan, A.B.: Scanpath modeling and classification with hidden markov models. Behavior research methods 50(1), 362-379 (2018)", + "4. Cristino, F., Mathôt, S., Theeuwes, J., Gilchrist, I.D.: Scanmatch: A novel method for comparing fixation sequences. Behavior research methods 42, 692-700 (2010)", + "5. Cuturi, M., Blondel, M.: Soft-dtw: a differentiable loss function for time-series. In: International conference on machine learning. pp. 894-903. PMLR (2017)" + ], + "bbox": [ + 225, + 294, + 785, + 839 + ], + "page_idx": 14 + }, + { + "type": "header", + "text": "Pathformer3D: A 3D Scanpath Transformer for $360^{\\circ}$ Images", + "bbox": [ + 330, + 114, + 732, + 128 + ], + "page_idx": 14 + }, + { + "type": "page_number", + "text": "15", + "bbox": [ + 767, + 116, + 785, + 126 + ], + "page_idx": 14 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "16. Dosovitskiy, A., Beyer, L., Kolesnikov, A., Weissenborn, D., Zhai, X., Unterthiner, T., Dehghani, M., Minderer, M., Heigold, G., Gelly, S., et al.: An image is worth 16x16 words: Transformers for image recognition at scale. arXiv preprint arXiv:2010.11929 (2020)", + "17. Engbert, R., Trukenbrod, H.A., Barthelme, S., Wichmann, F.A.: Spatial statistics and attentional dynamics in scene viewing. Journal of vision 15(1), 14-14 (2015)", + "18. Fan, C.L., Lee, J., Lo, W.C., Huang, C.Y., Chen, K.T., Hsu, C.H.: Fixation prediction for 360 video streaming in head-mounted virtual reality. In: Proceedings of the 27th Workshop on Network and Operating Systems Support for Digital Audio and Video. pp. 67-72 (2017)", + "19. Fang, Y., Huang, L., Yan, J., Liu, X., Liu, Y.: Perceptual quality assessment of omnidirectional images. In: Proceedings of the AAAI Conference on Artificial Intelligence. vol. 36, pp. 580-588 (2022)", + "20. Goodfellow, I., Pouget-Abadie, J., Mirza, M., Xu, B., Warde-Farley, D., Ozair, S., Courville, A., Bengio, Y.: Generative adversarial nets. Advances in neural information processing systems 27 (2014)", + "21. Graves, A.: Generating sequences with recurrent neural networks. arXiv preprint arXiv:1308.0850 (2013)", + "22. Itti, L., Koch, C., Niebur, E.: A model of saliency-based visual attention for rapid scene analysis. IEEE Transactions on pattern analysis and machine intelligence 20(11), 1254-1259 (1998)", + "23. Kamath, A., Singh, M., LeCun, Y., Synnaeve, G., Misra, I., Carion, N.: Mdetr-modulated detection for end-to-end multi-modal understanding. In: Proceedings of the IEEE/CVF International Conference on Computer Vision. pp. 1780-1790 (2021)", + "24. Kay, A.: User interface: A personal view. The art of human-computer interface design pp. 191-207 (1990)", + "25. Keogh, E., Ratanamahatana, C.A.: Exact indexing of dynamic time warping. Knowledge and information systems 7, 358-386 (2005)", + "26. Kinga, D., Adam, J.B., et al.: A method for stochastic optimization. In: International conference on learning representations (ICLR). vol. 5, p. 6. San Diego, California; (2015)", + "27. Krishnan, R., Shalit, U., Sontag, D.: Structured inference networks for nonlinear state space models. In: Proceedings of the AAAI Conference on Artificial Intelligence. vol. 31 (2017)", + "28. Kümmerer, M., Bethge, M., Wallis, T.S.: Deepgaze iii: Modeling free-viewing human scanpaths with deep learning. Journal of Vision 22(5), 7-7 (2022)", + "29. Kümmerer, M., Wallis, T.S., Bethge, M.: Information-theoretic model comparison unifies saliency metrics. Proceedings of the National Academy of Sciences 112(52), 16054-16059 (2015)", + "30. Le Meur, O., Coutrot, A.: Introducing context-dependent and spatially-variant viewing biases in saccadic models. Vision research 121, 72-84 (2016)", + "31. Le Meur, O., Le Callet, P., Barba, D.: Predicting visual fixations on video based on low-level visual features. Vision research 47(19), 2483-2498 (2007)", + "32. Le Meur, O., Liu, Z.: Saccadic model of eye movements for free-viewing condition. Vision research 116, 152-164 (2015)", + "33. Li, C., Zhang, W., Liu, Y., Wang, Y.: Very long term field of view prediction for 360-degree video streaming. In: 2019 IEEE conference on multimedia information processing and retrieval (MIPR). pp. 297-302. IEEE (2019)" + ], + "bbox": [ + 215, + 146, + 785, + 840 + ], + "page_idx": 15 + }, + { + "type": "page_number", + "text": "16", + "bbox": [ + 217, + 114, + 235, + 126 + ], + "page_idx": 15 + }, + { + "type": "header", + "text": "R. Quan et al.", + "bbox": [ + 271, + 114, + 367, + 128 + ], + "page_idx": 15 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "34. Liu, H., Xu, D., Huang, Q., Li, W., Xu, M., Lin, S.: Semantically-based human scanpath estimation with hmms. In: Proceedings of the IEEE International Conference on Computer Vision. pp. 3232-3239 (2013)", + "35. Martin, D., Serrano, A., Bergman, A.W., Wetzstein, G., Masia, B.: Scangan360: A generative model of realistic scanpaths for 360 images. IEEE Transactions on Visualization and Computer Graphics 28(5), 2003-2013 (2022)", + "36. Mondal, S., Yang, Z., Ahn, S., Samaras, D., Zelinsky, G., Hoai, M.: Gazeformer: Scalable, effective and fast prediction of goal-directed human attention. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 1441-1450 (2023)", + "37. Needleman, S.B., Wunsch, C.D.: A general method applicable to the search for similarities in the amino acid sequence of two proteins. Journal of molecular biology 48(3), 443-453 (1970)", + "38. Nguyen, A., Yan, Z., Nahrstedt, K.: Your attention is unique: Detecting 360-degree video saliency in head-mounted display for head movement prediction. In: Proceedings of the 26th ACM international conference on Multimedia. pp. 1190-1198 (2018)", + "39. Pan, J., Ferrer, C.C., McGuinness, K., O'Connor, N.E., Torres, J., Sayrol, E., Giro-i Nieto, X.: Salgan: Visual saliency prediction with generative adversarial networks. arXiv preprint arXiv:1701.01081 (2017)", + "40. Peters, R.J., Iyer, A., Itti, L., Koch, C.: Components of bottom-up gaze allocation in natural images. Vision research 45(18), 2397-2416 (2005)", + "41. Privitera, C.M., Stark, L.W.: Algorithms for defining visual regions-of-interest: Comparison with eye fixations. IEEE Transactions on pattern analysis and machine intelligence 22(9), 970-982 (2000)", + "42. Qiu, M., Rong, Q., Liang, D., Tu, H.: Visual scanpath transformer: Guiding computers to see the world. In: 2023 IEEE International Symposium on Mixed and Augmented Reality (ISMAR). pp. 223-232. IEEE (2023)", + "43. Rai, Y., Gutierrez, J., Le Callet, P.: A dataset of head and eye movements for 360 degree images. In: Proceedings of the 8th ACM on Multimedia Systems Conference. pp. 205-210 (2017)", + "44. Sitzmann, V., Serrano, A., Pavel, A., Agrawala, M., Gutierrez, D., Masia, B., Wetzstein, G.: Saliency in vr: How do people explore virtual environments? IEEE transactions on visualization and computer graphics 24(4), 1633-1642 (2018)", + "45. Sui, X., Fang, Y., Zhu, H., Wang, S., Wang, Z.: Scandmm: A deep markov model of scanpath prediction for $360^{\\circ}$ images. In: IEEE Conference on Computer Vision and Pattern Recognition (2023)", + "46. Sun, W., Chen, Z., Wu, F.: Visual scanpath prediction using ior-roi recurrent mixture density network. IEEE transactions on pattern analysis and machine intelligence 43(6), 2101-2118 (2019)", + "47. Sun, X., Yao, H., Ji, R., Liu, X.M.: Toward statistical modeling of saccadic eye movement and visual saliency. IEEE Transactions on Image Processing 23(11), 4649-4662 (2014)", + "48. Tatler, B.W., Brockmole, J.R., Carpenter, R.H.: Latest: A model of saccadic decisions in space and time. Psychological review 124(3), 267 (2017)", + "49. Upenik, E., Ebrahimi, T.: A simple method to obtain visual attention data in head mounted virtual reality. In: 2017 IEEE International Conference on Multimedia & Expo Workshops (ICMEW). pp. 73-78. IEEE (2017)", + "50. Verma, J.K., Paul, S.: Advances in Augmented Reality and Virtual Reality. Springer (2022)" + ], + "bbox": [ + 212, + 146, + 787, + 840 + ], + "page_idx": 16 + }, + { + "type": "header", + "text": "Pathformer3D: A 3D Scanpath Transformer for $360^{\\circ}$ Images", + "bbox": [ + 330, + 114, + 732, + 128 + ], + "page_idx": 16 + }, + { + "type": "page_number", + "text": "17", + "bbox": [ + 767, + 116, + 785, + 126 + ], + "page_idx": 16 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "51. Viswanath, B., Mislove, A., Cha, M., Gummadi, K.P.: On the evolution of user interaction in facebook. In: Proceedings of the 2nd ACM workshop on Online social networks. pp. 37-42 (2009)", + "52. Wang, W., Chen, C., Wang, Y., Jiang, T., Fang, F., Yao, Y.: Simulating human saccadic scanpaths on natural images. In: CVPR 2011. pp. 441-448. IEEE (2011)", + "53. Xia, C., Han, J., Qi, F., Shi, G.: Predicting human saccadic scanpaths based on iterative representation learning. IEEE Transactions on Image Processing 28(7), 3502-3515 (2019)", + "54. Xu, M., Yang, L., Tao, X., Duan, Y., Wang, Z.: Saliency prediction on omnidirectional image with generative adversarial imitation learning. IEEE Transactions on Image Processing 30, 2087-2102 (2021)", + "55. Yang, Z., Huang, L., Chen, Y., Wei, Z., Ahn, S., Zelinsky, G., Samaras, D., Hoai, M.: Predicting goal-directed human attention using inverse reinforcement learning. In: Proceedings of the IEEE/CVF conference on computer vision and pattern recognition. pp. 193-202 (2020)", + "56. Zanca, D., Melacci, S., Gori, M.: Gravitational laws of focus of attention. IEEE transactions on pattern analysis and machine intelligence 42(12), 2983-2995 (2019)", + "57. Zhu, Y., Zhai, G., Min, X.: The prediction of head and eye movement for 360 degree images. Signal Processing: Image Communication 69, 15-25 (2018)", + "58. Zhu, Y., Zhai, G., Min, X., Zhou, J.: The prediction of saliency map for head and eye movements in 360 degree images. IEEE Transactions on Multimedia 22(9), 2331-2344 (2019)", + "59. Zou, Z., Ye, M., Li, S., Li, X., Dufaux, F.: $360^{\\circ}$ image saliency prediction by embedding self-supervised proxy task. IEEE Transactions on Broadcasting (2023)" + ], + "bbox": [ + 215, + 146, + 787, + 479 + ], + "page_idx": 17 + }, + { + "type": "page_number", + "text": "18", + "bbox": [ + 217, + 114, + 235, + 126 + ], + "page_idx": 17 + }, + { + "type": "header", + "text": "R. Quan et al.", + "bbox": [ + 271, + 114, + 367, + 128 + ], + "page_idx": 17 + } +] \ No newline at end of file diff --git "a/2024/Pathformer3D_ A 3D Scanpath Transformer for 360\302\260 Images/8acb38cc-358d-422c-8280-ca63ea0a1521_model.json" "b/2024/Pathformer3D_ A 3D Scanpath Transformer for 360\302\260 Images/8acb38cc-358d-422c-8280-ca63ea0a1521_model.json" new file mode 100644 index 0000000000000000000000000000000000000000..6cfdbbf45c4bcef93061f164392868d180872b87 --- /dev/null +++ "b/2024/Pathformer3D_ A 3D Scanpath Transformer for 360\302\260 Images/8acb38cc-358d-422c-8280-ca63ea0a1521_model.json" @@ -0,0 +1,2436 @@ +[ + [ + { + "type": "title", + "bbox": [ + 0.23, + 0.141, + 0.776, + 0.187 + ], + "angle": 0, + "content": "Pathformer3D: A 3D Scanpath Transformer for \\(360^{\\circ}\\) Images" + }, + { + "type": "text", + "bbox": [ + 0.262, + 0.212, + 0.741, + 0.23 + ], + "angle": 0, + "content": "Rong Quan\\(^{1}\\), Yantao Lai\\(^{1,2}\\), Mengyu Qiu\\(^{2}\\), and Dong Liang\\(^{2,3}\\) (x̄)" + }, + { + "type": "text", + "bbox": [ + 0.217, + 0.239, + 0.787, + 0.295 + ], + "angle": 0, + "content": "1 College of Artificial Intelligence, Nanjing University of Aeronautics and Astronautics, the Key Laboratory of Brain-Machine Intelligence Technology, Ministry of Education, Nanjing, 211106, China rongquan0806@gmail.com, yantaolai@nuaa.edu.cn" + }, + { + "type": "text", + "bbox": [ + 0.23, + 0.296, + 0.774, + 0.351 + ], + "angle": 0, + "content": "\\(^{2}\\) MIIT Key Laboratory of Pattern Analysis and Machine Intelligence, College of Computer Science and Technology, Nanjing University of Aeronautics and Astronautics, Nanjing, China {qmengyu, liangdong}@nuaa.edu.cn" + }, + { + "type": "text", + "bbox": [ + 0.221, + 0.351, + 0.782, + 0.379 + ], + "angle": 0, + "content": "3 Shenzhen Research Institute, Nanjing University of Aeronautics and Astronautics, Shenzhen, China" + }, + { + "type": "text", + "bbox": [ + 0.261, + 0.416, + 0.744, + 0.694 + ], + "angle": 0, + "content": "Abstract. Scanpath prediction in \\(360^{\\circ}\\) images can help realize rapid rendering and better user interaction in Virtual/Augmented Reality applications. However, existing scanpath prediction models for \\(360^{\\circ}\\) images execute scanpath prediction on 2D equirectangular projection plane, which always result in big computation error owing to the 2D plane's distortion and coordinate discontinuity. In this work, we perform scanpath prediction for \\(360^{\\circ}\\) images in 3D spherical coordinate system and proposed a novel 3D scanpath Transformer named Pathformer3D. Specifically, a 3D Transformer encoder is first used to extract 3D contextual feature representation for the \\(360^{\\circ}\\) image. Then, the contextual feature representation and historical fixation information are input into a Transformer decoder to output current time step's fixation embedding, where the self-attention module is used to imitate the visual working memory mechanism of human visual system and directly model the time dependencies among the fixations. Finally, a 3D Gaussian distribution is learned from each fixation embedding, from which the fixation position can be sampled. Evaluation on four panoramic eye-tracking datasets demonstrates that Pathformer3D outperforms the current state-of-the-art methods. Code is available at https://github.com/lsztzp/Pathformer3D." + }, + { + "type": "text", + "bbox": [ + 0.262, + 0.708, + 0.621, + 0.723 + ], + "angle": 0, + "content": "Keywords: Scanpath \\(\\cdot\\) \\(360^{\\circ}\\) image \\(\\cdot\\) 3D Transformer" + }, + { + "type": "title", + "bbox": [ + 0.216, + 0.748, + 0.377, + 0.765 + ], + "angle": 0, + "content": "1 Introduction" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.78, + 0.788, + 0.842 + ], + "angle": 0, + "content": "Virtual reality (VR) and augmented reality (AR) [50] technologies have made significant advancements in recent years, providing the users immersive experiences. Meanwhile, understanding and imitating the way human beings explore the \\(360^{\\circ}\\) images of virtual environments is becoming more and more important," + } + ], + [ + { + "type": "page_number", + "bbox": [ + 0.218, + 0.116, + 0.23, + 0.127 + ], + "angle": 0, + "content": "2" + }, + { + "type": "header", + "bbox": [ + 0.272, + 0.115, + 0.369, + 0.129 + ], + "angle": 0, + "content": "R. Quan et al." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.147, + 0.785, + 0.193 + ], + "angle": 0, + "content": "since it can help realize more practical and fast rendering and thus improve user interaction in the immersive environment [24, 51]. Scanpath prediction on \\(360^{\\circ}\\) images refers to predict human's gaze shift path when exploring \\(360^{\\circ}\\) images." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.197, + 0.788, + 0.349 + ], + "angle": 0, + "content": "Existing research on scanpath prediction has primarily focused on 2D images [4, 22, 42, 46, 53], while \\(360^{\\circ}\\) images possess distinct characteristics in the VR/AR context. \\(360^{\\circ}\\) images provide immersive interactive environments where users can change their perspectives by physically moving their heads, resulting in a wider distribution of fixations. \\(360^{\\circ}\\) images encompass richer visual information that requires more time for users to process and absorb. Furthermore, the data structure of \\(360^{\\circ}\\) images differs from that of 2D images, introducing new requirements for data processing and analysis. Therefore, existing 2D image scanpath prediction methods cannot be directly applied to predict scanpaths in \\(360^{\\circ}\\) images." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.352, + 0.788, + 0.67 + ], + "angle": 0, + "content": "Early methods [6, 43, 57] for predicting scanpaths in 360 images involved sampling fixations based on saliency information to obtain the entire scanpath. Subsequently, with the advancement of generative adversarial networks [2], some researchers [5,35] have utilized generative networks to directly generate the entire path from the \\(360^{\\circ}\\) image. These methods have shown preliminary results but overlook the modeling of time dependency between fixations, which has long been demonstrated to be a very important characteristic of human visual attention mechanism [22], thereby they often lead to unstable prediction results. Recently, some research works [18,33,38,45] have recognized the need for a comprehensive treatment of the temporal dependencies in viewing behavior, and modeled them through recurrent neural networks (RNNs) or Markov chains. However, these methods integrate all historical fixation information into a single hidden unit, and generate the current fixation only from this hidden unit. In this case, the time dependent relationships among fixations can only be modeled indirectly and the influences from historical fixations will be greatly weakened. Besides, all above methods predict fixations on the 2D equirectangular projection of \\(360^{\\circ}\\) image, which has the issue of coordinate discontinuity caused by longitude (where \\(-180^{\\circ}\\) and \\(180^{\\circ}\\) represent the same fixation despite their significant numerical difference), and different degrees of distortion in different locations. Consequently, the fixation predicted on this 2D equirectangular projection will have a large margin of error." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.674, + 0.788, + 0.841 + ], + "angle": 0, + "content": "To solve the above problems, we propose a novel scanpath prediction model for \\(360^{\\circ}\\) images named Pathformer3D, which predicts fixations in 3D spherical coordinates of the \\(360^{\\circ}\\) image and directly model the time dependent relationships among fixations, to mimic human viewing behavior in immersive environment more realistically. Specifically, given a 2D equirectangular projection of \\(360^{\\circ}\\) image, we first transform it into 3D spherical coordinate system, and then exploit a spherical convolution to extract its visual features. After extracting the 3D visual features, we exploit a 3D Transformer encoder to learn each \\(360^{\\circ}\\) image region a contextual feature representation, by considering their long range spatial dependencies between each other. Next, we exploit a Transformer Decoder to learn each time step's fixation embedding directly from the whole image's" + } + ], + [ + { + "type": "header", + "bbox": [ + 0.331, + 0.115, + 0.733, + 0.131 + ], + "angle": 0, + "content": "Pathformer3D: A 3D Scanpath Transformer for \\(360^{\\circ}\\) Images" + }, + { + "type": "page_number", + "bbox": [ + 0.775, + 0.117, + 0.787, + 0.127 + ], + "angle": 0, + "content": "3" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.147, + 0.788, + 0.282 + ], + "angle": 0, + "content": "visual feature and the historical fixations, where the self-attention module is used to mimic human visual working memory mechanism, which can model the time dependencies between the current fixation and all historical fixations. After obtaining each time step's fixation embedding, we employ a 3D mixture density network to learn the 3D Gaussian distribution for each fixation, from which the fixation location is sampled. Using 3D Gaussian distribution to model the positional likelihood of fixations in 3D space takes into account the differences in scanpaths of different individuals, and can thus result in more robust and real scanpaths." + }, + { + "type": "text", + "bbox": [ + 0.239, + 0.283, + 0.644, + 0.298 + ], + "angle": 0, + "content": "In conclusion, this work has the following contributions:" + }, + { + "type": "text", + "bbox": [ + 0.226, + 0.306, + 0.784, + 0.38 + ], + "angle": 0, + "content": "- We perform visual scanpath prediction for \\(360^{\\circ}\\) images in a 3D spherical coordinate system for the first time, which can realize a more real imitation of human visual exploration process in immersive environment and simultaneously avoid the error caused by the distortion and positional discontinuity in 2D equirectangular projection of \\(360^{\\circ}\\) image." + }, + { + "type": "text", + "bbox": [ + 0.226, + 0.381, + 0.785, + 0.456 + ], + "angle": 0, + "content": "- We propose a novel scanpath prediction method named Pathformer3D, to learn contextual feature representation of the \\(360^{\\circ}\\) images in a 3D spherical coordinate system and directly consider the influences from historical fixations on current fixation, which is more in line with the visual working memory mechanism of human vision system." + }, + { + "type": "text", + "bbox": [ + 0.226, + 0.457, + 0.785, + 0.501 + ], + "angle": 0, + "content": "- The proposed method is comprehensively evaluated on four eye-tracking datasets of \\(360^{\\circ}\\) images, consistently achieving state-of-the-art performance, demonstrating the effectiveness of our approach." + }, + { + "type": "title", + "bbox": [ + 0.216, + 0.523, + 0.396, + 0.539 + ], + "angle": 0, + "content": "2 Related Works" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.554, + 0.785, + 0.599 + ], + "angle": 0, + "content": "Scanpath Prediction for 2D Images In recent years, numerous models have been proposed for predicting scanpaths in 2D images. According to the different sources of inspiration, they can be categorized into different classes." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.599, + 0.787, + 0.749 + ], + "angle": 0, + "content": "Some of these models draw inspiration from neuroscience and vision science, often leveraging human gaze behavior to guide scanpath prediction. For instance, some researchers [17, 22, 46] have utilized Inhibit-of-Return(IOR) strategies to guide the generation of scanpaths. Additionally, Wang et al. [52] compute response maps of sparse coding filters applied to foveated images at each step and select new fixations from the residual perceptual information map(RPI) using the information maximization principle. Adeli et al. [1] first apply a retina transformation to input images, then compute corresponding priority maps and select the next fixation. Furthermore, some researchers [48, 56] attempt to generate new fixations by leveraging low-level semantic information." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.75, + 0.788, + 0.841 + ], + "angle": 0, + "content": "Another subset of work is inspired by the statistical distribution characteristics of human scanpaths. Many researchers [7,8,30,32,34,42,53] in this group attempt to use image saliency information to predict new scanpaths. For example, VSPT [42] uses SalGAN [39] to extract saliency information and employs an autoregressive approach [21] to generate fixations. Others in this category adopt alternative methods, such as Sun et al. [47], who conduct super-Gaussian quality" + } + ], + [ + { + "type": "page_number", + "bbox": [ + 0.218, + 0.116, + 0.23, + 0.127 + ], + "angle": 0, + "content": "4" + }, + { + "type": "header", + "bbox": [ + 0.272, + 0.115, + 0.369, + 0.129 + ], + "angle": 0, + "content": "R. Quan et al." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.147, + 0.784, + 0.221 + ], + "angle": 0, + "content": "analysis on image patches and select the \\(k\\)th fixation to be at the patch with the highest response to the \\(k\\)th super Gaussian component. Coutrot et al. [13] model scanpaths using image and task dependent hidden markov models(HMM). Clarke et al. [11] implement a Gaussian jump distribution, which is polynomially dependent on the previous fixed distribution." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.223, + 0.785, + 0.268 + ], + "angle": 0, + "content": "In addition to the aforementioned models, there are also some deep learning models [5, 28, 55] that have achieved relatively good performance, despite their limited interpretability." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.288, + 0.787, + 0.499 + ], + "angle": 0, + "content": "Scanpath Prediction for \\(360^{\\circ}\\) Images Unlike 2D images, there are currently limited models specifically designed for scanpath prediction in \\(360^{\\circ}\\) images. However, with the gradual development of VR/AR technologies, this field has attracted increasing attention. In this domain, SaltiNet [6] is the first proposed method, which uses saliency volume to process \\(360^{\\circ}\\) images. PathGAN [5] adopts the method of Generative Adversarial [20] to generate scanpath of \\(360^{\\circ}\\) image. Zhu et al. [58] proposed a new method to estimate the saliency of \\(360^{\\circ}\\) images, and introduced mechanisms such as visual uncertainty and visual balance to predict the scanpath. The subsequent ScanGAN [35] is also a method that uses Generative Adversarial Networks [20] to generate scapath. Different from PathGAN [5], ScanGAN [35] adds a soft-dtw loss function [15] to the generator to generate scanpath that are more similar to real scanpath. ScanDMM [45] is a method that uses Deep Markov model [27] to generate scanpath, which takes into account the past historical information." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.5, + 0.785, + 0.59 + ], + "angle": 0, + "content": "All above methods predict fixations as well as scanpaths in 2D equirectangular projection of \\(360^{\\circ}\\) image, which has the weaknesses of distortion and coordinate discontinuity and always introduce big prediction error. Different from them, we perform scanpath prediction for \\(360^{\\circ}\\) images in 3D spherical coordinate system, which is more consistent with the real-world scenarios of human exploring immersive environments." + }, + { + "type": "title", + "bbox": [ + 0.216, + 0.611, + 0.338, + 0.629 + ], + "angle": 0, + "content": "3 Approach" + }, + { + "type": "title", + "bbox": [ + 0.216, + 0.64, + 0.421, + 0.654 + ], + "angle": 0, + "content": "3.1 Problem Definition" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.662, + 0.785, + 0.768 + ], + "angle": 0, + "content": "A scanpath can be represented as a sequence of fixations \\( X = (p_{1}, p_{2}, \\ldots, p_{T}) \\in \\mathbb{R}^{T \\times 3} \\), where \\( T \\) represents the length of the scanpath. For a given \\( 360^{\\circ} \\) image, the task of a scanpath prediction model is to generate a scanpath \\( \\widehat{X} \\) that resembles a real human scanpath. We use a 3D Cartesian coordinate system to represent coordinates, where each fixation is given in the form of \\( p = (x, y, z) \\), with \\( x \\), \\( y \\), and \\( z \\) being real numbers ranging from -1 to 1. This representation form helps address the issue of coordinate discontinuity." + }, + { + "type": "title", + "bbox": [ + 0.216, + 0.788, + 0.341, + 0.802 + ], + "angle": 0, + "content": "3.2 Overview" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.81, + 0.785, + 0.84 + ], + "angle": 0, + "content": "The architecture of our model is shown in Fig. 1. For each input \\(360^{\\circ}\\) image, we first transform it into 3D spherical coordinate system, and then extract the" + } + ], + [ + { + "type": "header", + "bbox": [ + 0.332, + 0.115, + 0.733, + 0.13 + ], + "angle": 0, + "content": "Pathformer3D: A 3D Scanpath Transformer for \\(360^{\\circ}\\) Images" + }, + { + "type": "page_number", + "bbox": [ + 0.775, + 0.117, + 0.785, + 0.127 + ], + "angle": 0, + "content": "5" + }, + { + "type": "image", + "bbox": [ + 0.221, + 0.144, + 0.784, + 0.384 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.214, + 0.395, + 0.788, + 0.466 + ], + "angle": 0, + "content": "Fig. 1: Overall architecture of our methods. The Transformer Encoder is utilized to contextualize the feature extracted by the SphereNet [12]. The contextual feature representation and the historical fixations are fed into the Transformer Decoder to output fixation embeddings. Each fixation embedding is input into a 3D Mixture Density Network to output its 3D Gaussian distribution, from which the fixation is sampled." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.497, + 0.788, + 0.633 + ], + "angle": 0, + "content": "image features using SphereNet [12]. Subsequently, we feed the extracted feature representation into a 3D Transformer Encoder to contextualize them. The Transformer Decoder takes historical fixation information and contextual feature representations as input and generates embedding information for new fixations in an autoregressive manner [21]. Then, a 3D Mixture Density Network (MDN) takes the embedding information of the fixations as input and outputs their 3D Gaussian position distributions. Finally, the fixations are sampled from the 3D Gaussian distributions. Next, we will introduce a detailed module-wise description of our model." + }, + { + "type": "title", + "bbox": [ + 0.215, + 0.66, + 0.64, + 0.675 + ], + "angle": 0, + "content": "3.3 3D Contextual Visual Feature Representation" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.689, + 0.788, + 0.779 + ], + "angle": 0, + "content": "3D feature extraction. We perform scanpath prediction in 3D spherical coordinate system. Therefore, after projecting the \\(360^{\\circ}\\) image into 3D spherical coordinate system, we utilize SphereNet [12] to extract its 3D feature representation. Specifically, for a \\(360^{\\circ}\\) image \\(I\\), after inputting it into SphereNet, we extract 3 different scales of feature maps and concatenate them along the channel axis, resulting in a feature representation \\(F\\) with dimensions of \\(C = 448\\), \\(H = 128\\), \\(W = 256\\)." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.78, + 0.788, + 0.841 + ], + "angle": 0, + "content": "Before inputting into the 3D Transformer Encoder, we first execute a simple size transformation to \\( F \\). Firstly, we execute average pooling to \\( F \\) in a convolutional kernel of size \\( 8 \\times 8 \\), to reduce the size of the feature maps to \\( C = 448 \\), \\( H = 16 \\), \\( W = 32 \\). Next, we flatten the pooled feature maps, resulting" + } + ], + [ + { + "type": "page_number", + "bbox": [ + 0.218, + 0.116, + 0.23, + 0.127 + ], + "angle": 0, + "content": "6" + }, + { + "type": "header", + "bbox": [ + 0.272, + 0.115, + 0.37, + 0.129 + ], + "angle": 0, + "content": "R. Quan et al." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.147, + 0.785, + 0.176 + ], + "angle": 0, + "content": "in a feature \\( F_{L} \\) with size of \\( L = 16 \\times 32 = 512 \\), \\( C = 448 \\). This series of operations can be formulated as:" + }, + { + "type": "equation", + "bbox": [ + 0.414, + 0.188, + 0.786, + 0.203 + ], + "angle": 0, + "content": "\\[\nF = \\operatorname {S p h e r e N e t} (I, \\theta) \\tag {1}\n\\]" + }, + { + "type": "equation", + "bbox": [ + 0.406, + 0.206, + 0.786, + 0.222 + ], + "angle": 0, + "content": "\\[\nF _ {L} = \\operatorname {F l a t t e n} (\\operatorname {A v g P o o l} (F)) \\tag {2}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.215, + 0.231, + 0.561, + 0.246 + ], + "angle": 0, + "content": "here, \\(\\theta\\) represents the parameters of SphereNet." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.267, + 0.785, + 0.342 + ], + "angle": 0, + "content": "3D feature encoder. A linear embedding layer is first applied to map \\( F_{L} \\) into feature embeddings. Subsequently, the obtained feature embeddings are added to a 3D positional encoding and fed into \\( M = 4 \\) standard Transformer Encoder Layers to extract contextual feature representations. The above process can be expressed by the following formulas." + }, + { + "type": "equation", + "bbox": [ + 0.305, + 0.352, + 0.785, + 0.368 + ], + "angle": 0, + "content": "\\[\nF _ {L} ^ {\\prime} = \\operatorname {E m b e d d i n g} \\left(F _ {L}\\right) \\tag {3}\n\\]" + }, + { + "type": "equation", + "bbox": [ + 0.308, + 0.372, + 0.785, + 0.39 + ], + "angle": 0, + "content": "\\[\nP _ {0} = F _ {L} ^ {\\prime} + E _ {\\text {p o s} 3 d} \\quad E _ {\\text {p o s} 3 d} \\in R ^ {L \\times D} \\tag {4}\n\\]" + }, + { + "type": "equation", + "bbox": [ + 0.31, + 0.392, + 0.785, + 0.408 + ], + "angle": 0, + "content": "\\[\nP _ {i} = \\operatorname {E n c o d e r L a y e r} \\left(P _ {i - 1}\\right) \\quad i = 1 \\dots M \\tag {5}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.416, + 0.785, + 0.462 + ], + "angle": 0, + "content": "where \\( D \\) represents the feature embedding dimension of the 3D Transformer Encoder, \\( M \\) represents the number of EncoderLayer in the architecture, and \\( P_{i} \\) represents the output of the i-th EncoderLayer." + }, + { + "type": "title", + "bbox": [ + 0.215, + 0.482, + 0.512, + 0.497 + ], + "angle": 0, + "content": "3.4 Fixation Sequence Generation" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.505, + 0.785, + 0.596 + ], + "angle": 0, + "content": "Fixation decoder. The contextual feature representation and historical fixation information are input into a Transformer decoder to output the fixation embedding. The fixation embedding is then fed into a 3D mixture density network to output the fixation's 3D Gaussian distribution, from which the fixation is sampled. We use autoregressive method [21] to predict one fixation each time and iteratively generate the whole fixation sequence." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.597, + 0.786, + 0.716 + ], + "angle": 0, + "content": "On time step \\( t \\), we initialize the fixation query vector \\( q_{t} \\) using fixation \\( p_{t-1} \\) of the previous time step. Specially, for the first fixation query \\( q_{1} \\), we initialize it using the center of 3D space, i.e., \\( p_{0} = (0,0,0) \\). Subsequently, the initialized query vector \\( q_{t} \\) is concatenated with the historical query vectors to generate the joint query vector \\( Q_{0}' \\). After adding one-dimensional positional encoding information, \\( Q_{0}' \\) becomes \\( Q_{0} \\). Finally, after passing through \\( N = 4 \\) standard Transformer Decoder Layers, we obtain the hidden states \\( Q_{N} = (Z_{1}, Z_{2}, \\ldots, Z_{t}) \\) of the fixations. This process is defined as follows:" + }, + { + "type": "equation", + "bbox": [ + 0.302, + 0.728, + 0.785, + 0.743 + ], + "angle": 0, + "content": "\\[\nq _ {t} = \\operatorname {E m b e d d i n g} \\left(p _ {t - 1}\\right) \\quad t \\in [ 1, 2, \\dots , T ] \\tag {6}\n\\]" + }, + { + "type": "equation", + "bbox": [ + 0.296, + 0.746, + 0.785, + 0.762 + ], + "angle": 0, + "content": "\\[\nQ _ {0} ^ {\\prime} = \\left[ q _ {1}; q _ {2}; \\dots ; q _ {t} \\right] \\tag {7}\n\\]" + }, + { + "type": "equation", + "bbox": [ + 0.295, + 0.765, + 0.785, + 0.784 + ], + "angle": 0, + "content": "\\[\nQ _ {0} = Q _ {0} ^ {\\prime} + E _ {p o s \\_ 1 d} \\quad E _ {p o s \\_ 1 d} \\in R ^ {t \\times D} \\tag {8}\n\\]" + }, + { + "type": "equation", + "bbox": [ + 0.296, + 0.786, + 0.785, + 0.801 + ], + "angle": 0, + "content": "\\[\nQ _ {i} = \\operatorname {D e c o d e r L a y e r} \\left(Q _ {i - 1}, P _ {M}\\right) \\quad i = 1 \\dots N \\tag {9}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.81, + 0.786, + 0.841 + ], + "angle": 0, + "content": "where \\( P_{M} \\) represents the output of the Transformer Encoder, \\( T \\) represents the length of the predicted fixation sequence, \\( p_{t-1} \\) represents the fixation at the time" + } + ], + [ + { + "type": "header", + "bbox": [ + 0.331, + 0.115, + 0.733, + 0.131 + ], + "angle": 0, + "content": "Pathformer3D: A 3D Scanpath Transformer for \\(360^{\\circ}\\) Images" + }, + { + "type": "page_number", + "bbox": [ + 0.775, + 0.116, + 0.787, + 0.127 + ], + "angle": 0, + "content": "7" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.147, + 0.784, + 0.177 + ], + "angle": 0, + "content": "step \\( t - 1 \\), \\( N \\) represents the number of decoder layers. \\( Q_{i} \\) represents the output of the i-th DecoderLayer." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.179, + 0.788, + 0.286 + ], + "angle": 0, + "content": "The fixation decoder process in the training process has some difference from that in the generation process. During the training phase, we input the fixations \\((p_0,p_1,\\ldots ,p_{T - 1})\\in R^{T\\times 3}\\) into the Transformer Decoder, and it can predict the hidden states \\(Q_{N} = (Z_{1},Z_{2},\\dots,Z_{T})\\) of the true scanpath \\((p_1,p_2,\\dots,p_T)\\in\\) \\(R^{T\\times 3}\\) in a parallel manner, which is a parallel process. During the generation phase, we iteratively generate the hidden states of new coordinates \\(Z_{1},Z_{2},\\ldots Z_{T}\\) in sequence starting from the initial coordinate \\(p_0\\), which is a sequential process." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.312, + 0.788, + 0.419 + ], + "angle": 0, + "content": "Fixation generation. After obtaining \\( Z_{t} \\), we then employ a 3D mixture density network to model the probability distribution of the fixations. The 3D mixture density network decodes the hidden state \\( Z_{t} \\) of the fixation at time step \\( t \\) into \\( K \\) sets of Gaussian kernel parameters. We use \\( \\mu_t^i = (\\mu_{x,t}^i, \\mu_{y,t}^i, \\mu_{z,t}^i) \\), \\( \\Sigma_t^i \\), and \\( \\pi_t^i \\) to represent the mean vector, covariance matrix, and weight of the \\( i \\)-th Gaussian kernel at time step \\( t \\), respectively. These three parameters can be obtained through three separate MLP layers, which can be defined as:" + }, + { + "type": "equation", + "bbox": [ + 0.314, + 0.432, + 0.786, + 0.451 + ], + "angle": 0, + "content": "\\[\n\\left\\{\\mu_ {t} ^ {i}, \\Sigma_ {t} ^ {i}, \\pi_ {t} ^ {i} \\right\\} = f _ {m d n} \\left(Z _ {t}; \\theta_ {m d n}\\right) \\quad i = 1 \\dots K \\tag {10}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.462, + 0.71, + 0.478 + ], + "angle": 0, + "content": "where \\(\\theta_{mdn}\\) refers to the parameters of the mixture density network." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.478, + 0.787, + 0.507 + ], + "angle": 0, + "content": "Next, the probability density function for a 3D Gaussian distribution \\( P(p_{t}) \\) can be defined as follows:" + }, + { + "type": "equation", + "bbox": [ + 0.312, + 0.519, + 0.786, + 0.556 + ], + "angle": 0, + "content": "\\[\n\\mathcal {N} \\left(p _ {t} \\mid \\mu_ {t} ^ {i}, \\Sigma_ {t} ^ {i}\\right) = \\frac {\\exp \\left(- \\frac {1}{2} \\left(p _ {t} - \\mu_ {t} ^ {i}\\right) ^ {\\top} \\left(\\Sigma_ {t} ^ {i}\\right) ^ {- 1} \\left(p _ {t} - \\mu_ {t} ^ {i}\\right)\\right)}{\\sqrt {(2 \\pi) ^ {3} \\left| \\Sigma_ {t} ^ {i} \\right|}} \\tag {11}\n\\]" + }, + { + "type": "equation", + "bbox": [ + 0.372, + 0.56, + 0.786, + 0.6 + ], + "angle": 0, + "content": "\\[\nP \\left(p _ {t}\\right) = \\sum_ {i = 1} ^ {K} \\pi_ {t} ^ {i} \\mathcal {N} \\left(p _ {t} \\mid \\mu_ {t} ^ {i}, \\Sigma_ {t} ^ {i}\\right) \\tag {12}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.613, + 0.787, + 0.658 + ], + "angle": 0, + "content": "where the \\(|\\Sigma_t^i|\\) means the determinant of the covariance matrix \\(\\Sigma_t^i\\) and the \\(\\mathcal{N}(p_t\\mid \\mu_t^i,\\Sigma_t^i)\\) represents the Gaussian distribution function associated with the \\(i\\)-th 3D Gaussian function." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.66, + 0.787, + 0.691 + ], + "angle": 0, + "content": "The prediction fixation \\(\\hat{p}_t\\) is sampled probabilistically from a 3D surface, which can be expressed as:" + }, + { + "type": "equation", + "bbox": [ + 0.465, + 0.706, + 0.786, + 0.721 + ], + "angle": 0, + "content": "\\[\n\\hat {p} _ {t} \\sim P \\left(p _ {t}\\right) \\tag {13}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.735, + 0.788, + 0.842 + ], + "angle": 0, + "content": "here, it is worth noting that we do not simply sample the fixation with the highest probability, since such sampling strategy always makes the model rapidly converge to the same fixation positions for all images. In our study, we generated a total of 32,768 fixation points along the latitude-longitude grid, with a resolution of \\(128 \\times 256\\). We independently computed the probability for each fixation point. Finally, we randomly selected a fixation point based on its probability, with higher probabilities being more likely to be chosen." + } + ], + [ + { + "type": "page_number", + "bbox": [ + 0.218, + 0.116, + 0.23, + 0.127 + ], + "angle": 0, + "content": "8" + }, + { + "type": "header", + "bbox": [ + 0.272, + 0.115, + 0.37, + 0.129 + ], + "angle": 0, + "content": "R. Quan et al." + }, + { + "type": "title", + "bbox": [ + 0.216, + 0.147, + 0.378, + 0.161 + ], + "angle": 0, + "content": "3.5 Loss Function" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.171, + 0.788, + 0.232 + ], + "angle": 0, + "content": "During training, the objective is to minimize the discrepancy between the predicted distribution and the real human scanpath. To achieve this, we employ the negative log-likelihood loss function, which encourages higher probabilities for the ground truth fixations. The loss function can be defined as follows:" + }, + { + "type": "equation", + "bbox": [ + 0.364, + 0.243, + 0.786, + 0.284 + ], + "angle": 0, + "content": "\\[\n\\mathcal {L} = - \\frac {1}{T} \\sum_ {t = 1} ^ {T} \\log \\left(\\sum_ {i = 1} ^ {K} \\pi_ {t} ^ {i} \\mathcal {N} \\left(p _ {t} ^ {*} \\mid \\mu_ {t} ^ {i}, \\Sigma_ {t} ^ {i}\\right)\\right) \\tag {14}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.216, + 0.293, + 0.704, + 0.309 + ], + "angle": 0, + "content": "where \\( p_t^* \\) represents the real human fixation position of time step \\( t \\)." + }, + { + "type": "title", + "bbox": [ + 0.216, + 0.331, + 0.377, + 0.349 + ], + "angle": 0, + "content": "4 Experiments" + }, + { + "type": "title", + "bbox": [ + 0.216, + 0.362, + 0.336, + 0.376 + ], + "angle": 0, + "content": "4.1 Datasets" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.386, + 0.825, + 0.599 + ], + "angle": 0, + "content": "We conducted experiments on four \\(360^{\\circ}\\) image datasets: Sitzmann [44], Salient360! [43], AOI [54], and JUFE [19]. The Sitzmann dataset [44] comprises 22 images with a total of 1920 scanpaths. Only the training set of the Salient360! dataset [43] is available, which contains 85 images and 3036 scanpaths. The AOI dataset [54] consists of 600 high-resolution images and approximately 18,000 scanpaths. The JUFE dataset [19] includes 1032 images and 30,960 scanpaths. For training, 19 images from Sitzmann [44] and 60 images from Salient360! [43] were utilized, and all remaining data were used for validation. To increase the number of training sets, we rotated the image along with the corresponding scanpaths six times along the direction of longitude. We followed ScanDMM [45] to preprocess the gaze data of Sitzmann, Salient360!, and JUFE datasets, and obtain visual scanpaths containing 30, 25, and 15 fixations for Sitzmann, Salient360!, and JUFE. The AOI dataset has already been pre-processed, with an average scanpath length of approximately 8." + }, + { + "type": "title", + "bbox": [ + 0.216, + 0.62, + 0.421, + 0.634 + ], + "angle": 0, + "content": "4.2 Evaluation Metrics" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.644, + 0.789, + 0.843 + ], + "angle": 0, + "content": "We employed six metrics to assess the performances of our model and other models, including Levenshtein Distance (LEV) [41], Dynamic Time Warping (DTW) [25], Time Delay Embedding (TDE), ScanMatch [14], Recursive Metric (REC) [3] and Sequence Score(SS) [55]. LEV [41] is a metric that quantifies the dissimilarity between two strings by measuring the minimum number of single-character edit operations needed to transform one string into another. ScanMatch [14] involves converting eye-tracking scanpaths into string sequences and then employing an algorithm similar to the Levenshtein distance to compare the similarity of the two sequences. DTW [25] and TDE [52] are metrics based on time series analysis, while REC [3] is a metric for recursive quantification analysis. SS [55] is calculated by converting scanpaths into strings of fixation cluster IDs, and then measuring the similarity between two strings using a string matching algorithm [37]." + } + ], + [ + { + "type": "header", + "bbox": [ + 0.331, + 0.115, + 0.733, + 0.131 + ], + "angle": 0, + "content": "Pathformer3D: A 3D Scanpath Transformer for \\(360^{\\circ}\\) Images" + }, + { + "type": "page_number", + "bbox": [ + 0.775, + 0.117, + 0.787, + 0.127 + ], + "angle": 0, + "content": "9" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.146, + 0.788, + 0.239 + ], + "angle": 0, + "content": "For each image, every model generates \\( m = 10 \\) predicted scanpaths \\( \\widehat{X} \\), which are evaluated against all real human scanpaths \\( X \\), and then averaged to determine the final prediction score. Note that our model predicts scanpath with a length of 30. When comparing across different datasets, we truncate the predicted scanpath to the length of the actual scanpath before conducting the comparison. Taking DTW [25] as an example, the score can be calculated as follows:" + }, + { + "type": "equation", + "bbox": [ + 0.376, + 0.251, + 0.786, + 0.29 + ], + "angle": 0, + "content": "\\[\nD T W = \\frac {1}{n} \\frac {1}{m} \\sum_ {i = 1} ^ {n} \\sum_ {j = 1} ^ {m} D T W \\left(\\widehat {X} _ {j}, X _ {i}\\right) \\tag {15}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.299, + 0.787, + 0.348 + ], + "angle": 0, + "content": "where \\(\\widehat{X}_j\\) is the \\(j\\)-th predicted scanpath, \\(X_i\\) is the \\(i\\)-th real human scanpath, \\(n\\) is the number of the real human scanpaths, \\(m\\) is the number of predicted scanpaths." + }, + { + "type": "title", + "bbox": [ + 0.215, + 0.37, + 0.457, + 0.385 + ], + "angle": 0, + "content": "4.3 Implementation Details" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.396, + 0.789, + 0.531 + ], + "angle": 0, + "content": "The internal vector dimension of 3D Transformer Encoder is set to \\( D = 128 \\). We utilize a multi-head attention mechanism with 8 attention heads. The hidden layer size of the feed-forward network (FFN) is set to 64. In the 3D mixture density network model, we configure \\( K = 5 \\) Gaussian kernels, and the internal hidden layer dimension is set to 16. During the training process, our batch size is set to 18. As for the optimizer, we use AdamW [26] with a learning rate of 1e-5. We employ a learning rate warm-up and stage-wise adjustment strategy. The warm-up period lasts for 10 epochs, and the learning rate is halved every 10 epochs. The total training duration is 50 epochs." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.533, + 0.787, + 0.564 + ], + "angle": 0, + "content": "Our experiments were conducted on a single NVIDIA GeForce RTX 3090, and the development was conducted using the PyTorch framework." + }, + { + "type": "title", + "bbox": [ + 0.215, + 0.587, + 0.478, + 0.603 + ], + "angle": 0, + "content": "4.4 Performance Comparison" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.613, + 0.788, + 0.719 + ], + "angle": 0, + "content": "We compared our model with SaliTiNet [6], ScanGAN [35], and ScanDMM [45], utilizing the models provided officially by their respective sources to obtain prediction results. These three models are all designed for scanpath prediction on \\(360^{\\circ}\\) images. PathGAN [5] and Zhu et al. [58] are two additional models specifically designed for generating scanpath on \\(360^{\\circ}\\) images. However, it has been demonstrated that these models are unable to effectively generate realistic scanpaths [35]. Therefore, we have not included them in our comparison." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.72, + 0.789, + 0.841 + ], + "angle": 0, + "content": "We additionally established two baseline reference values to assess the relative quality of the generated scanpath. The first one is the random baseline (Random walk). This baseline randomly generate ten scanpaths for each image, which provides the reference for lowest performance. The second one is the human baseline (Human), which compares each real human scanpath in the dataset with all the other real human scanpaths and taking the average. This baseline represents the average similarity among real human scanpaths, serving as a reference for highest performance." + } + ], + [ + { + "type": "page_number", + "bbox": [ + 0.218, + 0.116, + 0.236, + 0.127 + ], + "angle": 0, + "content": "10" + }, + { + "type": "header", + "bbox": [ + 0.272, + 0.115, + 0.369, + 0.129 + ], + "angle": 0, + "content": "R. Quan et al." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.147, + 0.788, + 0.222 + ], + "angle": 0, + "content": "Quantitative evaluation. Tab. 1 shows the quantitative comparison results between our method and other competitors on the metrics of LEV [41], DTW [25], TDE [52], ScanMatch [14], REC [3] and SS [55]. As we can see, our method outperforms all the competitors on all four datasets, achieving performance close to 'Human'." + }, + { + "type": "table_caption", + "bbox": [ + 0.214, + 0.246, + 0.788, + 0.276 + ], + "angle": 0, + "content": "Table 1: Performance of scanpath prediction models on different datasets. The best result for each metric is highlighted in bold, while the second-best result is underlined." + }, + { + "type": "table", + "bbox": [ + 0.22, + 0.287, + 0.788, + 0.484 + ], + "angle": 0, + "content": "
MethodSitzmannSalient360!
LEV ↓DTW ↓TDE ↓ScanMatch ↑REC ↑SS ↑LEV ↓DTW ↓TDE ↓ScanMatch ↑REC ↑SS ↑
Random walk51.5222372.38728.0380.3951.5140.19843.6702062.08029.7530.3751.3810.176
SaltiNet [6]51.1512246.05726.3030.3741.7340.19841.0421860.39725.0010.4292.2500.220
ScanGAN [35]45.8521950.48819.0600.4772.9600.26538.9771752.37119.5270.4603.0410.244
ScanDMM [45]45.7481939.98019.0980.4853.1560.27038.5451706.66120.2380.4673.1250.257
Ours44.6911939.66219.8870.4864.1540.28036.2771515.88319.2610.4844.2500.302
Human42.0061841.53416.1920.5265.1630.36635.0901376.43916.6500.5135.0660.377
MethodAOIJUFE
LEV ↓DTW ↓TDE ↓ScanMatch ↑REC ↑SS ↑LEV ↓DTW ↓TDE ↓ScanMatch ↑REC ↑SS ↑
Random walk15.866666.65238.1640.2935.6500.24126.4391270.89338.1640.3521.8470.191
SaltiNet [6]14.978622.07935.7240.3376.3060.24526.1151264.49232.2610.3641.9720.192
ScanGAN [35]14.473569.54932.6020.3617.3790.25824.2591110.82625.9370.4313.2080.217
ScanDMM [45]14.185569.17032.5580.3717.3280.24923.3721081.46525.7870.4523.5270.221
Ours13.904561.89530.9870.3768.2860.22522.5951054.22322.9550.4624.4490.218
Human14.317501.95230.2480.36713.0080.37718.4451030.97020.3050.5647.8640.330
" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.536, + 0.788, + 0.747 + ], + "angle": 0, + "content": "Qualitative evaluation. We visualized the prediction results of all methods on one image of each dataset, as shown in Fig. 2. Similarly, we visualized the ground truth of human scanpath accordingly. The transition from purple to red in the scanpath indicates the beginning and end of the scanpath, respectively. It is noticeable that most ground truth scanpaths are located within the salient regions near the equator. SaltiNet [6], although designed for \\(360^{\\circ}\\) images, suffers from large displacements. In comparison, the methods proposed by ScanGAN [35] and ScanDMM [45] have the capability to generate relatively realistic scanpaths to some extent. However, our model, in contrast to these approaches, excels at capturing salient information within \\(360^{\\circ}\\) images and is capable of generating scanpath that are closer to the real human scanpaths. In addition, a notable flaw in the tower image is that our model tends to focus more on the prominent plants at both ends rather than the central tower, which is a common issue shared by other models as well." + }, + { + "type": "title", + "bbox": [ + 0.215, + 0.77, + 0.437, + 0.787 + ], + "angle": 0, + "content": "4.5 Saliency Comparison" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.795, + 0.787, + 0.843 + ], + "angle": 0, + "content": "A good scanpath prediction model often assigns more attention to salient regions during the prediction process. In our study, we generated 1000 scanpath and performed post-processing on all fixations to generate corresponding saliency" + } + ], + [ + { + "type": "header", + "bbox": [ + 0.332, + 0.115, + 0.733, + 0.129 + ], + "angle": 0, + "content": "Pathformer3D: A 3D Scanpath Transformer for \\(360^{\\circ}\\) Images" + }, + { + "type": "page_number", + "bbox": [ + 0.769, + 0.117, + 0.785, + 0.127 + ], + "angle": 0, + "content": "11" + }, + { + "type": "image", + "bbox": [ + 0.219, + 0.144, + 0.784, + 0.327 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.214, + 0.339, + 0.788, + 0.396 + ], + "angle": 0, + "content": "Fig. 2: Qualitative representation results of different models on four datasets. From top to bottom, the four images are sourced from Sitzmann [44], Salient360! [43], AOI [54], and JUFE [19], respectively. From left to right, we display the a scanpath sampled from ground truth, our model, ScanDMM [45], ScanGAN [35], and SaltiNet [6]." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.428, + 0.788, + 0.506 + ], + "angle": 0, + "content": "maps [49]. The fixation maps was obtained from all unprocessed fixations. We used sizes \\( \\mathrm{H} = 128 \\) and \\( \\mathrm{W} = 256 \\) for saliency maps (salmap) and fixation maps (fixmap). We compared our results with ScanGAN [35] and ScanDMM [45] on the Salient360! [43] and JUFE [19] datasets, under the saliency metrics AUC_Judd [9], NSS [40], CC [31], SIM [9], and KLD [29]." + }, + { + "type": "table_caption", + "bbox": [ + 0.214, + 0.534, + 0.788, + 0.577 + ], + "angle": 0, + "content": "Table 2: Quantitative comparison results of significance comparison on Salient360! [43] and JUFE [19] datasets. The best result for each metric is highlighted in bold, while the second-best result is underlined." + }, + { + "type": "table", + "bbox": [ + 0.278, + 0.588, + 0.731, + 0.689 + ], + "angle": 0, + "content": "
DatasetMethodAUC_Judd↑NSS ↑CC ↑SIM ↑KLD ↓
Salient360! [43]ScanGAN [35]0.7970.9860.2600.17321.210
ScanDMM [45]0.8301.2100.3220.18320.882
Ours0.8381.1950.3520.25918.985
JUFE [19]ScanGAN [35]0.7950.8370.1980.23619.529
ScanDMM [45]0.8661.2030.2750.28618.248
Ours0.8721.0350.2450.32016.339
" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.72, + 0.788, + 0.841 + ], + "angle": 0, + "content": "Based on Tab. 2, our model has demonstrated favorable performance across the majority of metrics. Our model did not achieve optimal results in a few metrics, one potential explanation is that our model exhibits a stronger emphasis on larger salient objects, which may lead to comparatively less attention towards smaller targets compared to models with more evenly distributed saliency maps. As evident in Fig. 3, ScanDMM [45] and ScanGAN [35] exhibit a characteristic of having more dispersed saliency maps. Furthermore, it can be observed that our model has the ability to focus more on salient regions compared to the other" + } + ], + [ + { + "type": "page_number", + "bbox": [ + 0.218, + 0.116, + 0.236, + 0.127 + ], + "angle": 0, + "content": "12" + }, + { + "type": "header", + "bbox": [ + 0.272, + 0.115, + 0.369, + 0.129 + ], + "angle": 0, + "content": "R. Quan et al." + }, + { + "type": "image", + "bbox": [ + 0.223, + 0.147, + 0.787, + 0.364 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.214, + 0.374, + 0.788, + 0.432 + ], + "angle": 0, + "content": "Fig. 3: Qualitative comparison results of significance comparison. The first two images is from Salient360! [43] and the next three images are from JUFE [19]. From top to bottom are the real image, the significance maps of the real images, Ours, ScanDMM [45], ScanGAN [35]." + }, + { + "type": "table_caption", + "bbox": [ + 0.259, + 0.443, + 0.744, + 0.459 + ], + "angle": 0, + "content": "Table 3: Efficiency comparison in terms of model size and running time" + }, + { + "type": "table", + "bbox": [ + 0.278, + 0.469, + 0.731, + 0.523 + ], + "angle": 0, + "content": "
SaltiNet [6]ScanGAN [35]ScanDMM [45]Ours
Paramters(MB)98.7315.13417.77919.280
Running Time(ms)127206.482118.250266.420
" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.546, + 0.788, + 0.577 + ], + "angle": 0, + "content": "models. As a result, the post-processed saliency maps obtained from our model exhibit a closer resemblance to human-generated saliency maps." + }, + { + "type": "title", + "bbox": [ + 0.215, + 0.596, + 0.448, + 0.611 + ], + "angle": 0, + "content": "4.6 Efficiency Comparison" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.618, + 0.789, + 0.77 + ], + "angle": 0, + "content": "We also compared our model's size and running time for generating a scanpath with other methods. In Tab. 3, we present the model size and the runtime for generating a single scanpath (measured in milliseconds). As can be seen, we have a similar model size with other models. In terms of runtime, our model is capable of generating approximately four scanpaths of length 30 per second, which is a little slower compared to ScanGAN and ScanDMM. This is primarily because our model generates fixations in an autoregressive manner [21]. Additionally, the use of a mixture density network is another factor leading to slower running speed. However, compared to a single Gaussian distribution, multiple Gaussian distributions can more accurately represent the distribution of fixations." + }, + { + "type": "title", + "bbox": [ + 0.215, + 0.788, + 0.403, + 0.803 + ], + "angle": 0, + "content": "4.7 Ablation Studies" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.81, + 0.797, + 0.842 + ], + "angle": 0, + "content": "We conducted ablation experiments on the Salient360! [43] and JUFE [19] datasets. The experimental results are presented in Tab. 4, showing the optimal results" + } + ], + [ + { + "type": "header", + "bbox": [ + 0.331, + 0.114, + 0.732, + 0.131 + ], + "angle": 0, + "content": "Pathformer3D: A 3D Scanpath Transformer for \\(360^{\\circ}\\) Images" + }, + { + "type": "page_number", + "bbox": [ + 0.769, + 0.116, + 0.786, + 0.127 + ], + "angle": 0, + "content": "13" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.147, + 0.788, + 0.194 + ], + "angle": 0, + "content": "obtained for each configuration. Next, we will provide a brief overview of each model architecture. For more detailed architecture information, please refer to the supplementary file." + }, + { + "type": "table_caption", + "bbox": [ + 0.216, + 0.22, + 0.787, + 0.25 + ], + "angle": 0, + "content": "Table 4: Ablation studies on Salient360! [43] and JUFE [19] datasets. The best result for each metric is highlighted in bold, while the second-best result is underlined." + }, + { + "type": "table", + "bbox": [ + 0.22, + 0.261, + 0.788, + 0.4 + ], + "angle": 0, + "content": "
MethodSalient360!JUFE
LEV ↓DTW ↓TDE ↓ScanMatch ↑REC ↑SS ↑LEV ↓DTW ↓TDE ↓ScanMatch ↑REC ↑SS ↑
Pure ViT37.6431739.71419.2910.4703.7100.26722.8911072.75323.7510.4573.8470.221
Pure 2D CNN36.8481727.24119.6410.4774.1190.27222.6131060.20523.0960.4634.1810.220
Saliency38.1961864.91420.0590.4753.1560.25323.8331094.29024.6360.4423.6690.208
w/o EncoderLayer37.3521798.34020.7270.4634.1710.26222.9261072.26323.1960.4554.2600.218
w/o MDN + MSE Loss39.4631793.94825.9140.4263.7810.09523.9151132.12725.6750.4163.8330.102
K=138.6431804.07420.0740.4623.3260.24823.2391084.73423.5770.4513.8050.218
K=336.2502073.90522.3260.4904.0390.24622.7581077.36523.3020.4604.3600.214
K=838.1171700.87720.4550.4673.3890.25822.9171065.24023.4700.4553.9140.210
Parallel36.7221577.99323.7260.4893.9390.28822.5671054.64428.7750.4513.6880.221
Ours36.2771515.88319.2610.4844.2530.30222.5951054.22322.9550.4624.4490.218
" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.455, + 0.788, + 0.608 + ], + "angle": 0, + "content": "The feature extractor. To demonstrate the effectiveness of our 3D feature extractor, SphereNet, we try to employ three distinct architectures for feature extraction. Firstly, we extracted features using the patch-based approach of the Vision Transformer (ViT) method [16] (denoted as 'Pure ViT'). Secondly, we replaced SphereNet with a standard 2D convolutional network (denoted as 'Pure 2D CNN'). Lastly, We followed VSPT [42] to extract image features based on the saliency information. Specifically, we employed EPSNet [59] to extract saliency information (denoted as 'Saliency'). As shown in Tab. 4, 'Ours' obtains better performance than 'Pure ViT', 'Pure 2D CNN', and 'Saliency', which demonstrates the superiority of our feature extractor." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.633, + 0.788, + 0.74 + ], + "angle": 0, + "content": "The 3D Transformer encoder. To validate the effectiveness of our 3D Transformer Encoder, we attempted to remove the EncoderLayer in the 3D Transformer Encoder and directly utilize the feature embedding of the Transformer Encoder as input of the Transformer Decoder (denoted as 'w/o EncoderLayer'). As we can see from Tab. 4, without our 3D Transformer Encoder, the scanpath prediction performance dropped significantly, which demonstrates the effectiveness of our 3D Transformer Encoder." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.765, + 0.788, + 0.842 + ], + "angle": 0, + "content": "3D MDN. To validate the effectiveness of the 3D MDN, we directly exploited linear regression to predict the fixations after the Transformer Decoder. The Mean Squared Error (MSE) was utilized as the loss function (denoted as 'w/o MDN + MSE Loss'). As can be seen in Tab. 4, 'Ours' outperforms 'w/o MDN + MSE Loss' a lot, proving the advantage of our 3D MDN." + } + ], + [ + { + "type": "page_number", + "bbox": [ + 0.218, + 0.116, + 0.236, + 0.127 + ], + "angle": 0, + "content": "14" + }, + { + "type": "header", + "bbox": [ + 0.272, + 0.115, + 0.369, + 0.129 + ], + "angle": 0, + "content": "R. Quan et al." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.147, + 0.788, + 0.239 + ], + "angle": 0, + "content": "Different number of Gaussian kernels. We compared with the model results obtained by using different numbers of Gaussian kernels in the mixture density network. These comparisons are respectively presented in the rows labeled \"K=1,3,8,\" while our model has 5 Gaussian kernels. As we can see from Tab. 4, using 5 Gaussian kernels can provide a better representation of the fixations' distribution." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.257, + 0.79, + 0.392 + ], + "angle": 0, + "content": "Parallel in the Transformer decoder. In our Transformer decoder, we use autoregressive method to predict one fixation each time and iteratively generate the whole fixation sequence. Through our observation, more and more models try to adopt a parallel structure in the Transformer decoder [10, 23, 36] for higher efficiency. Therefore, we also tried to build a parallel Transformer decoder in our method (denoted as 'Parallel'). In the parallel model, we use a random embedding with length \\( L \\) as the input to the Transformer decoder, and we can get the hidden state of a scanpath of length \\( L \\) at one time, all the other parts of the model are unchanged." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.394, + 0.79, + 0.456 + ], + "angle": 0, + "content": "We can observe from Tab. 4 that the parallel structure of the decoder performs poorly than the autoregressive Transformer decoder, especially in the TDE [52] metric. Such comparison results demonstrate that an autoregressive Transformer decoder is more suitable for our scanpath prediction task." + }, + { + "type": "title", + "bbox": [ + 0.215, + 0.475, + 0.54, + 0.492 + ], + "angle": 0, + "content": "5 Limitaions and Future Works" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.504, + 0.789, + 0.564 + ], + "angle": 0, + "content": "In our work, we employed SphereNet [12] for feature extraction and 3D Mixture Density Network (MDN) for fixation generation. However, we acknowledge that these components are replaceable, and we are contemplating whether there might be better alternatives for both parts." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.565, + 0.789, + 0.61 + ], + "angle": 0, + "content": "In addition, we are considering incorporating relevant semantic information to improve the performance further, as objects with important semantic information tend to attract human attention more easily." + }, + { + "type": "title", + "bbox": [ + 0.215, + 0.63, + 0.363, + 0.647 + ], + "angle": 0, + "content": "6 Conclusion" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.659, + 0.79, + 0.842 + ], + "angle": 0, + "content": "Considering that the 2D equirectangular projection of \\(360^{\\circ}\\) images has the issues of distortion and coordinate discontinuity, we proposed to execute scanpath prediction for \\(360^{\\circ}\\) images in 3D spherical coordinate system. Correspondingly, we proposed a novel 3D Scanpath Transformer named Pathformer3D, where a 3D Transformer Encoder was first used to obtain 3D contextual feature representation of the \\(360^{\\circ}\\) image. Then, Pathformer3D input the 3D contextual feature representation and historical fixation information into the Transformer Decoder to predict the hidden state for each fixation. Each fixation's hidden state was then fed into a 3D mixture density network to output the 3D Gaussian distribution of the fixation, from which the fixation location can be sampled. Finally, Pathformer3D model was compared with three state-of-the-art \\(360^{\\circ}\\) image scan-path prediction models, to demonstrate the superiority of its performance." + } + ], + [ + { + "type": "header", + "bbox": [ + 0.331, + 0.115, + 0.733, + 0.13 + ], + "angle": 0, + "content": "Pathformer3D: A 3D Scanpath Transformer for \\(360^{\\circ}\\) Images" + }, + { + "type": "page_number", + "bbox": [ + 0.769, + 0.117, + 0.786, + 0.127 + ], + "angle": 0, + "content": "15" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.147, + 0.788, + 0.239 + ], + "angle": 0, + "content": "Acknowledgements This work is supported in part by the National Natural Science Foundation of China under grant 62206127 and 62272229, the Natural Science Foundation of Jiangsu Province under grant BK20222012, and Shenzhen Science and Technology Program JCYJ20230807142001004. Rong Quan and Yantao Lai contributed equally to this work. Dong Liang is the corresponding author." + }, + { + "type": "title", + "bbox": [ + 0.218, + 0.263, + 0.324, + 0.279 + ], + "angle": 0, + "content": "References" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.295, + 0.787, + 0.338 + ], + "angle": 0, + "content": "1. Adeli, H., Vitu, F., Zelinsky, G.J.: A model of the superior colliculus predicts fixation locations during scene viewing and visual search. Journal of Neuroscience 37(6), 1453-1467 (2017)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.339, + 0.787, + 0.38 + ], + "angle": 0, + "content": "2. Aggarwal, A., Mittal, M., Battineni, G.: Generative adversarial network: An overview of theory and applications. International Journal of Information Management Data Insights 1(1), 100004 (2021)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.381, + 0.787, + 0.421 + ], + "angle": 0, + "content": "3. Anderson, N.C., Bischof, W.F., Laidlaw, K.E., Risko, E.F., Kingstone, A.: Recurrence quantification analysis of eye movements. Behavior research methods 45, 842-856 (2013)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.422, + 0.787, + 0.449 + ], + "angle": 0, + "content": "4. Antol, S., Agrawal, A., Lu, J., Mitchell, M., Batra, D., Zitnick, C.L., Parikh, D.: Vqa: Visual question answering. In: ICCV. pp. 2425-2433 (2015)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.45, + 0.787, + 0.491 + ], + "angle": 0, + "content": "5. Assens, M., Giro-i Nieto, X., McGuinness, K., O'Connor, N.E.: Pathgan: Visual scanpath prediction with generative adversarial networks. In: Proceedings of the European Conference on Computer Vision (ECCV) Workshops. pp. 0-0 (2018)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.492, + 0.787, + 0.547 + ], + "angle": 0, + "content": "6. Assens Reina, M., Giro-i Nieto, X., McGuinness, K., O'Connor, N.E.: Saltinet: Scan-path prediction on 360 degree images using saliency volumes. In: Proceedings of the IEEE International Conference on Computer Vision Workshops. pp. 2331-2338 (2017)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.548, + 0.787, + 0.575 + ], + "angle": 0, + "content": "7. Boccignone, G., Ferraro, M.: Modelling gaze shift as a constrained random walk. Physica A: Statistical Mechanics and its Applications 331(1-2), 207-218 (2004)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.576, + 0.787, + 0.602 + ], + "angle": 0, + "content": "8. Brockmann, D., Geisel, T.: The ecology of gaze shifts. Neurocomputing 32, 643-650 (2000)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.603, + 0.787, + 0.631 + ], + "angle": 0, + "content": "9. Bylinskii, Z., Judd, T., Borji, A., Itti, L., Durand, F., Oliva, A., Torralba, A.: Mit saliency benchmark (2015)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.632, + 0.787, + 0.673 + ], + "angle": 0, + "content": "0. Carion, N., Massa, F., Synnaeve, G., Usunier, N., Kirillov, A., Zagoruyko, S.: End-to-end object detection with transformers. In: European conference on computer vision. pp. 213-229. Springer (2020)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.674, + 0.787, + 0.714 + ], + "angle": 0, + "content": "1. Clarke, A.D., Stainer, M.J., Tatler, B.W., Hunt, A.R.: The saccadic flow baseline: Accounting for image-independent biases in fixation behavior. Journal of vision 17(11), 12-12 (2017)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.715, + 0.787, + 0.757 + ], + "angle": 0, + "content": "2. Coors, B., Condurache, A.P., Geiger, A.: Spherenet: Learning spherical representations for detection and classification in omnidirectional images. In: Proceedings of the European conference on computer vision (ECCV). pp. 518-533 (2018)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.758, + 0.787, + 0.784 + ], + "angle": 0, + "content": "3. Coutrot, A., Hsiao, J.H., Chan, A.B.: Scanpath modeling and classification with hidden markov models. Behavior research methods 50(1), 362-379 (2018)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.785, + 0.787, + 0.812 + ], + "angle": 0, + "content": "4. Cristino, F., Mathôt, S., Theeuwes, J., Gilchrist, I.D.: Scanmatch: A novel method for comparing fixation sequences. Behavior research methods 42, 692-700 (2010)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.813, + 0.787, + 0.84 + ], + "angle": 0, + "content": "5. Cuturi, M., Blondel, M.: Soft-dtw: a differentiable loss function for time-series. In: International conference on machine learning. pp. 894-903. PMLR (2017)" + }, + { + "type": "list", + "bbox": [ + 0.226, + 0.295, + 0.787, + 0.84 + ], + "angle": 0, + "content": null + } + ], + [ + { + "type": "page_number", + "bbox": [ + 0.218, + 0.116, + 0.236, + 0.127 + ], + "angle": 0, + "content": "16" + }, + { + "type": "header", + "bbox": [ + 0.272, + 0.115, + 0.369, + 0.129 + ], + "angle": 0, + "content": "R. Quan et al." + }, + { + "type": "ref_text", + "bbox": [ + 0.218, + 0.147, + 0.787, + 0.204 + ], + "angle": 0, + "content": "16. Dosovitskiy, A., Beyer, L., Kolesnikov, A., Weissenborn, D., Zhai, X., Unterthiner, T., Dehghani, M., Minderer, M., Heigold, G., Gelly, S., et al.: An image is worth 16x16 words: Transformers for image recognition at scale. arXiv preprint arXiv:2010.11929 (2020)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.205, + 0.786, + 0.232 + ], + "angle": 0, + "content": "17. Engbert, R., Trukenbrod, H.A., Barthelme, S., Wichmann, F.A.: Spatial statistics and attentional dynamics in scene viewing. Journal of vision 15(1), 14-14 (2015)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.233, + 0.787, + 0.288 + ], + "angle": 0, + "content": "18. Fan, C.L., Lee, J., Lo, W.C., Huang, C.Y., Chen, K.T., Hsu, C.H.: Fixation prediction for 360 video streaming in head-mounted virtual reality. In: Proceedings of the 27th Workshop on Network and Operating Systems Support for Digital Audio and Video. pp. 67-72 (2017)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.289, + 0.787, + 0.33 + ], + "angle": 0, + "content": "19. Fang, Y., Huang, L., Yan, J., Liu, X., Liu, Y.: Perceptual quality assessment of omnidirectional images. In: Proceedings of the AAAI Conference on Artificial Intelligence. vol. 36, pp. 580-588 (2022)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.331, + 0.786, + 0.373 + ], + "angle": 0, + "content": "20. Goodfellow, I., Pouget-Abadie, J., Mirza, M., Xu, B., Warde-Farley, D., Ozair, S., Courville, A., Bengio, Y.: Generative adversarial nets. Advances in neural information processing systems 27 (2014)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.374, + 0.786, + 0.401 + ], + "angle": 0, + "content": "21. Graves, A.: Generating sequences with recurrent neural networks. arXiv preprint arXiv:1308.0850 (2013)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.402, + 0.786, + 0.443 + ], + "angle": 0, + "content": "22. Itti, L., Koch, C., Niebur, E.: A model of saliency-based visual attention for rapid scene analysis. IEEE Transactions on pattern analysis and machine intelligence 20(11), 1254-1259 (1998)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.444, + 0.786, + 0.5 + ], + "angle": 0, + "content": "23. Kamath, A., Singh, M., LeCun, Y., Synnaeve, G., Misra, I., Carion, N.: Mdetr-modulated detection for end-to-end multi-modal understanding. In: Proceedings of the IEEE/CVF International Conference on Computer Vision. pp. 1780-1790 (2021)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.501, + 0.786, + 0.528 + ], + "angle": 0, + "content": "24. Kay, A.: User interface: A personal view. The art of human-computer interface design pp. 191-207 (1990)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.53, + 0.786, + 0.556 + ], + "angle": 0, + "content": "25. Keogh, E., Ratanamahatana, C.A.: Exact indexing of dynamic time warping. Knowledge and information systems 7, 358-386 (2005)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.557, + 0.786, + 0.599 + ], + "angle": 0, + "content": "26. Kinga, D., Adam, J.B., et al.: A method for stochastic optimization. In: International conference on learning representations (ICLR). vol. 5, p. 6. San Diego, California; (2015)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.6, + 0.786, + 0.641 + ], + "angle": 0, + "content": "27. Krishnan, R., Shalit, U., Sontag, D.: Structured inference networks for nonlinear state space models. In: Proceedings of the AAAI Conference on Artificial Intelligence. vol. 31 (2017)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.642, + 0.786, + 0.67 + ], + "angle": 0, + "content": "28. Kümmerer, M., Bethge, M., Wallis, T.S.: Deepgaze iii: Modeling free-viewing human scanpaths with deep learning. Journal of Vision 22(5), 7-7 (2022)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.671, + 0.786, + 0.712 + ], + "angle": 0, + "content": "29. Kümmerer, M., Wallis, T.S., Bethge, M.: Information-theoretic model comparison unifies saliency metrics. Proceedings of the National Academy of Sciences 112(52), 16054-16059 (2015)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.713, + 0.786, + 0.741 + ], + "angle": 0, + "content": "30. Le Meur, O., Coutrot, A.: Introducing context-dependent and spatially-variant viewing biases in saccadic models. Vision research 121, 72-84 (2016)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.741, + 0.786, + 0.77 + ], + "angle": 0, + "content": "31. Le Meur, O., Le Callet, P., Barba, D.: Predicting visual fixations on video based on low-level visual features. Vision research 47(19), 2483-2498 (2007)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.771, + 0.786, + 0.798 + ], + "angle": 0, + "content": "32. Le Meur, O., Liu, Z.: Saccadic model of eye movements for free-viewing condition. Vision research 116, 152-164 (2015)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.799, + 0.786, + 0.841 + ], + "angle": 0, + "content": "33. Li, C., Zhang, W., Liu, Y., Wang, Y.: Very long term field of view prediction for 360-degree video streaming. In: 2019 IEEE conference on multimedia information processing and retrieval (MIPR). pp. 297-302. IEEE (2019)" + }, + { + "type": "list", + "bbox": [ + 0.217, + 0.147, + 0.787, + 0.841 + ], + "angle": 0, + "content": null + } + ], + [ + { + "type": "header", + "bbox": [ + 0.331, + 0.115, + 0.733, + 0.129 + ], + "angle": 0, + "content": "Pathformer3D: A 3D Scanpath Transformer for \\(360^{\\circ}\\) Images" + }, + { + "type": "page_number", + "bbox": [ + 0.769, + 0.117, + 0.786, + 0.127 + ], + "angle": 0, + "content": "17" + }, + { + "type": "ref_text", + "bbox": [ + 0.214, + 0.147, + 0.788, + 0.189 + ], + "angle": 0, + "content": "34. Liu, H., Xu, D., Huang, Q., Li, W., Xu, M., Lin, S.: Semantically-based human scanpath estimation with hmms. In: Proceedings of the IEEE International Conference on Computer Vision. pp. 3232-3239 (2013)" + }, + { + "type": "ref_text", + "bbox": [ + 0.214, + 0.19, + 0.788, + 0.232 + ], + "angle": 0, + "content": "35. Martin, D., Serrano, A., Bergman, A.W., Wetzstein, G., Masia, B.: Scangan360: A generative model of realistic scanpaths for 360 images. IEEE Transactions on Visualization and Computer Graphics 28(5), 2003-2013 (2022)" + }, + { + "type": "ref_text", + "bbox": [ + 0.214, + 0.232, + 0.788, + 0.286 + ], + "angle": 0, + "content": "36. Mondal, S., Yang, Z., Ahn, S., Samaras, D., Zelinsky, G., Hoai, M.: Gazeformer: Scalable, effective and fast prediction of goal-directed human attention. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 1441-1450 (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.214, + 0.286, + 0.788, + 0.328 + ], + "angle": 0, + "content": "37. Needleman, S.B., Wunsch, C.D.: A general method applicable to the search for similarities in the amino acid sequence of two proteins. Journal of molecular biology 48(3), 443-453 (1970)" + }, + { + "type": "ref_text", + "bbox": [ + 0.214, + 0.328, + 0.788, + 0.383 + ], + "angle": 0, + "content": "38. Nguyen, A., Yan, Z., Nahrstedt, K.: Your attention is unique: Detecting 360-degree video saliency in head-mounted display for head movement prediction. In: Proceedings of the 26th ACM international conference on Multimedia. pp. 1190-1198 (2018)" + }, + { + "type": "ref_text", + "bbox": [ + 0.214, + 0.383, + 0.788, + 0.425 + ], + "angle": 0, + "content": "39. Pan, J., Ferrer, C.C., McGuinness, K., O'Connor, N.E., Torres, J., Sayrol, E., Giro-i Nieto, X.: Salgan: Visual saliency prediction with generative adversarial networks. arXiv preprint arXiv:1701.01081 (2017)" + }, + { + "type": "ref_text", + "bbox": [ + 0.214, + 0.425, + 0.788, + 0.452 + ], + "angle": 0, + "content": "40. Peters, R.J., Iyer, A., Itti, L., Koch, C.: Components of bottom-up gaze allocation in natural images. Vision research 45(18), 2397-2416 (2005)" + }, + { + "type": "ref_text", + "bbox": [ + 0.214, + 0.452, + 0.788, + 0.493 + ], + "angle": 0, + "content": "41. Privitera, C.M., Stark, L.W.: Algorithms for defining visual regions-of-interest: Comparison with eye fixations. IEEE Transactions on pattern analysis and machine intelligence 22(9), 970-982 (2000)" + }, + { + "type": "ref_text", + "bbox": [ + 0.214, + 0.493, + 0.788, + 0.536 + ], + "angle": 0, + "content": "42. Qiu, M., Rong, Q., Liang, D., Tu, H.: Visual scanpath transformer: Guiding computers to see the world. In: 2023 IEEE International Symposium on Mixed and Augmented Reality (ISMAR). pp. 223-232. IEEE (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.214, + 0.536, + 0.788, + 0.577 + ], + "angle": 0, + "content": "43. Rai, Y., Gutierrez, J., Le Callet, P.: A dataset of head and eye movements for 360 degree images. In: Proceedings of the 8th ACM on Multimedia Systems Conference. pp. 205-210 (2017)" + }, + { + "type": "ref_text", + "bbox": [ + 0.214, + 0.577, + 0.788, + 0.619 + ], + "angle": 0, + "content": "44. Sitzmann, V., Serrano, A., Pavel, A., Agrawala, M., Gutierrez, D., Masia, B., Wetzstein, G.: Saliency in vr: How do people explore virtual environments? IEEE transactions on visualization and computer graphics 24(4), 1633-1642 (2018)" + }, + { + "type": "ref_text", + "bbox": [ + 0.214, + 0.619, + 0.788, + 0.66 + ], + "angle": 0, + "content": "45. Sui, X., Fang, Y., Zhu, H., Wang, S., Wang, Z.: Scandmm: A deep markov model of scanpath prediction for \\(360^{\\circ}\\) images. In: IEEE Conference on Computer Vision and Pattern Recognition (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.214, + 0.66, + 0.788, + 0.702 + ], + "angle": 0, + "content": "46. Sun, W., Chen, Z., Wu, F.: Visual scanpath prediction using ior-roi recurrent mixture density network. IEEE transactions on pattern analysis and machine intelligence 43(6), 2101-2118 (2019)" + }, + { + "type": "ref_text", + "bbox": [ + 0.214, + 0.702, + 0.788, + 0.743 + ], + "angle": 0, + "content": "47. Sun, X., Yao, H., Ji, R., Liu, X.M.: Toward statistical modeling of saccadic eye movement and visual saliency. IEEE Transactions on Image Processing 23(11), 4649-4662 (2014)" + }, + { + "type": "ref_text", + "bbox": [ + 0.214, + 0.743, + 0.788, + 0.771 + ], + "angle": 0, + "content": "48. Tatler, B.W., Brockmole, J.R., Carpenter, R.H.: Latest: A model of saccadic decisions in space and time. Psychological review 124(3), 267 (2017)" + }, + { + "type": "ref_text", + "bbox": [ + 0.214, + 0.771, + 0.788, + 0.813 + ], + "angle": 0, + "content": "49. Upenik, E., Ebrahimi, T.: A simple method to obtain visual attention data in head mounted virtual reality. In: 2017 IEEE International Conference on Multimedia & Expo Workshops (ICMEW). pp. 73-78. IEEE (2017)" + }, + { + "type": "ref_text", + "bbox": [ + 0.214, + 0.813, + 0.788, + 0.841 + ], + "angle": 0, + "content": "50. Verma, J.K., Paul, S.: Advances in Augmented Reality and Virtual Reality. Springer (2022)" + }, + { + "type": "list", + "bbox": [ + 0.214, + 0.147, + 0.788, + 0.841 + ], + "angle": 0, + "content": null + } + ], + [ + { + "type": "page_number", + "bbox": [ + 0.218, + 0.116, + 0.236, + 0.127 + ], + "angle": 0, + "content": "18" + }, + { + "type": "header", + "bbox": [ + 0.272, + 0.115, + 0.369, + 0.129 + ], + "angle": 0, + "content": "R. Quan et al." + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.147, + 0.788, + 0.189 + ], + "angle": 0, + "content": "51. Viswanath, B., Mislove, A., Cha, M., Gummadi, K.P.: On the evolution of user interaction in facebook. In: Proceedings of the 2nd ACM workshop on Online social networks. pp. 37-42 (2009)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.19, + 0.787, + 0.218 + ], + "angle": 0, + "content": "52. Wang, W., Chen, C., Wang, Y., Jiang, T., Fang, F., Yao, Y.: Simulating human saccadic scanpaths on natural images. In: CVPR 2011. pp. 441-448. IEEE (2011)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.218, + 0.787, + 0.259 + ], + "angle": 0, + "content": "53. Xia, C., Han, J., Qi, F., Shi, G.: Predicting human saccadic scanpaths based on iterative representation learning. IEEE Transactions on Image Processing 28(7), 3502-3515 (2019)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.259, + 0.787, + 0.3 + ], + "angle": 0, + "content": "54. Xu, M., Yang, L., Tao, X., Duan, Y., Wang, Z.: Saliency prediction on omnidirectional image with generative adversarial imitation learning. IEEE Transactions on Image Processing 30, 2087-2102 (2021)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.3, + 0.787, + 0.357 + ], + "angle": 0, + "content": "55. Yang, Z., Huang, L., Chen, Y., Wei, Z., Ahn, S., Zelinsky, G., Samaras, D., Hoai, M.: Predicting goal-directed human attention using inverse reinforcement learning. In: Proceedings of the IEEE/CVF conference on computer vision and pattern recognition. pp. 193-202 (2020)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.357, + 0.787, + 0.384 + ], + "angle": 0, + "content": "56. Zanca, D., Melacci, S., Gori, M.: Gravitational laws of focus of attention. IEEE transactions on pattern analysis and machine intelligence 42(12), 2983-2995 (2019)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.384, + 0.787, + 0.411 + ], + "angle": 0, + "content": "57. Zhu, Y., Zhai, G., Min, X.: The prediction of head and eye movement for 360 degree images. Signal Processing: Image Communication 69, 15-25 (2018)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.411, + 0.787, + 0.453 + ], + "angle": 0, + "content": "58. Zhu, Y., Zhai, G., Min, X., Zhou, J.: The prediction of saliency map for head and eye movements in 360 degree images. IEEE Transactions on Multimedia 22(9), 2331-2344 (2019)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.453, + 0.787, + 0.481 + ], + "angle": 0, + "content": "59. Zou, Z., Ye, M., Li, S., Li, X., Dufaux, F.: \\(360^{\\circ}\\) image saliency prediction by embedding self-supervised proxy task. IEEE Transactions on Broadcasting (2023)" + }, + { + "type": "list", + "bbox": [ + 0.217, + 0.147, + 0.788, + 0.481 + ], + "angle": 0, + "content": null + } + ] +] \ No newline at end of file diff --git "a/2024/Pathformer3D_ A 3D Scanpath Transformer for 360\302\260 Images/8acb38cc-358d-422c-8280-ca63ea0a1521_origin.pdf" "b/2024/Pathformer3D_ A 3D Scanpath Transformer for 360\302\260 Images/8acb38cc-358d-422c-8280-ca63ea0a1521_origin.pdf" new file mode 100644 index 0000000000000000000000000000000000000000..9b07949327f5ebbf30861ac4b84cf8ec0eca1e1b --- /dev/null +++ "b/2024/Pathformer3D_ A 3D Scanpath Transformer for 360\302\260 Images/8acb38cc-358d-422c-8280-ca63ea0a1521_origin.pdf" @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:253121c8fe125f000f49c5ea08cc0920589d318dd5acbbe9a3346df8cd75eefa +size 1061987 diff --git "a/2024/Pathformer3D_ A 3D Scanpath Transformer for 360\302\260 Images/full.md" "b/2024/Pathformer3D_ A 3D Scanpath Transformer for 360\302\260 Images/full.md" new file mode 100644 index 0000000000000000000000000000000000000000..7af786639dc482d644c075f21e61112fa4882126 --- /dev/null +++ "b/2024/Pathformer3D_ A 3D Scanpath Transformer for 360\302\260 Images/full.md" @@ -0,0 +1,331 @@ +# Pathformer3D: A 3D Scanpath Transformer for $360^{\circ}$ Images + +Rong Quan $^{1}$ , Yantao Lai $^{1,2}$ , Mengyu Qiu $^{2}$ , and Dong Liang $^{2,3}$ (x̄) + +1 College of Artificial Intelligence, Nanjing University of Aeronautics and Astronautics, the Key Laboratory of Brain-Machine Intelligence Technology, Ministry of Education, Nanjing, 211106, China rongquan0806@gmail.com, yantaolai@nuaa.edu.cn + +$^{2}$ MIIT Key Laboratory of Pattern Analysis and Machine Intelligence, College of Computer Science and Technology, Nanjing University of Aeronautics and Astronautics, Nanjing, China {qmengyu, liangdong}@nuaa.edu.cn + +3 Shenzhen Research Institute, Nanjing University of Aeronautics and Astronautics, Shenzhen, China + +Abstract. Scanpath prediction in $360^{\circ}$ images can help realize rapid rendering and better user interaction in Virtual/Augmented Reality applications. However, existing scanpath prediction models for $360^{\circ}$ images execute scanpath prediction on 2D equirectangular projection plane, which always result in big computation error owing to the 2D plane's distortion and coordinate discontinuity. In this work, we perform scanpath prediction for $360^{\circ}$ images in 3D spherical coordinate system and proposed a novel 3D scanpath Transformer named Pathformer3D. Specifically, a 3D Transformer encoder is first used to extract 3D contextual feature representation for the $360^{\circ}$ image. Then, the contextual feature representation and historical fixation information are input into a Transformer decoder to output current time step's fixation embedding, where the self-attention module is used to imitate the visual working memory mechanism of human visual system and directly model the time dependencies among the fixations. Finally, a 3D Gaussian distribution is learned from each fixation embedding, from which the fixation position can be sampled. Evaluation on four panoramic eye-tracking datasets demonstrates that Pathformer3D outperforms the current state-of-the-art methods. Code is available at https://github.com/lsztzp/Pathformer3D. + +Keywords: Scanpath $\cdot$ $360^{\circ}$ image $\cdot$ 3D Transformer + +# 1 Introduction + +Virtual reality (VR) and augmented reality (AR) [50] technologies have made significant advancements in recent years, providing the users immersive experiences. Meanwhile, understanding and imitating the way human beings explore the $360^{\circ}$ images of virtual environments is becoming more and more important, + +since it can help realize more practical and fast rendering and thus improve user interaction in the immersive environment [24, 51]. Scanpath prediction on $360^{\circ}$ images refers to predict human's gaze shift path when exploring $360^{\circ}$ images. + +Existing research on scanpath prediction has primarily focused on 2D images [4, 22, 42, 46, 53], while $360^{\circ}$ images possess distinct characteristics in the VR/AR context. $360^{\circ}$ images provide immersive interactive environments where users can change their perspectives by physically moving their heads, resulting in a wider distribution of fixations. $360^{\circ}$ images encompass richer visual information that requires more time for users to process and absorb. Furthermore, the data structure of $360^{\circ}$ images differs from that of 2D images, introducing new requirements for data processing and analysis. Therefore, existing 2D image scanpath prediction methods cannot be directly applied to predict scanpaths in $360^{\circ}$ images. + +Early methods [6, 43, 57] for predicting scanpaths in 360 images involved sampling fixations based on saliency information to obtain the entire scanpath. Subsequently, with the advancement of generative adversarial networks [2], some researchers [5,35] have utilized generative networks to directly generate the entire path from the $360^{\circ}$ image. These methods have shown preliminary results but overlook the modeling of time dependency between fixations, which has long been demonstrated to be a very important characteristic of human visual attention mechanism [22], thereby they often lead to unstable prediction results. Recently, some research works [18,33,38,45] have recognized the need for a comprehensive treatment of the temporal dependencies in viewing behavior, and modeled them through recurrent neural networks (RNNs) or Markov chains. However, these methods integrate all historical fixation information into a single hidden unit, and generate the current fixation only from this hidden unit. In this case, the time dependent relationships among fixations can only be modeled indirectly and the influences from historical fixations will be greatly weakened. Besides, all above methods predict fixations on the 2D equirectangular projection of $360^{\circ}$ image, which has the issue of coordinate discontinuity caused by longitude (where $-180^{\circ}$ and $180^{\circ}$ represent the same fixation despite their significant numerical difference), and different degrees of distortion in different locations. Consequently, the fixation predicted on this 2D equirectangular projection will have a large margin of error. + +To solve the above problems, we propose a novel scanpath prediction model for $360^{\circ}$ images named Pathformer3D, which predicts fixations in 3D spherical coordinates of the $360^{\circ}$ image and directly model the time dependent relationships among fixations, to mimic human viewing behavior in immersive environment more realistically. Specifically, given a 2D equirectangular projection of $360^{\circ}$ image, we first transform it into 3D spherical coordinate system, and then exploit a spherical convolution to extract its visual features. After extracting the 3D visual features, we exploit a 3D Transformer encoder to learn each $360^{\circ}$ image region a contextual feature representation, by considering their long range spatial dependencies between each other. Next, we exploit a Transformer Decoder to learn each time step's fixation embedding directly from the whole image's + +visual feature and the historical fixations, where the self-attention module is used to mimic human visual working memory mechanism, which can model the time dependencies between the current fixation and all historical fixations. After obtaining each time step's fixation embedding, we employ a 3D mixture density network to learn the 3D Gaussian distribution for each fixation, from which the fixation location is sampled. Using 3D Gaussian distribution to model the positional likelihood of fixations in 3D space takes into account the differences in scanpaths of different individuals, and can thus result in more robust and real scanpaths. + +In conclusion, this work has the following contributions: + +- We perform visual scanpath prediction for $360^{\circ}$ images in a 3D spherical coordinate system for the first time, which can realize a more real imitation of human visual exploration process in immersive environment and simultaneously avoid the error caused by the distortion and positional discontinuity in 2D equirectangular projection of $360^{\circ}$ image. + +- We propose a novel scanpath prediction method named Pathformer3D, to learn contextual feature representation of the $360^{\circ}$ images in a 3D spherical coordinate system and directly consider the influences from historical fixations on current fixation, which is more in line with the visual working memory mechanism of human vision system. + +- The proposed method is comprehensively evaluated on four eye-tracking datasets of $360^{\circ}$ images, consistently achieving state-of-the-art performance, demonstrating the effectiveness of our approach. + +# 2 Related Works + +Scanpath Prediction for 2D Images In recent years, numerous models have been proposed for predicting scanpaths in 2D images. According to the different sources of inspiration, they can be categorized into different classes. + +Some of these models draw inspiration from neuroscience and vision science, often leveraging human gaze behavior to guide scanpath prediction. For instance, some researchers [17, 22, 46] have utilized Inhibit-of-Return(IOR) strategies to guide the generation of scanpaths. Additionally, Wang et al. [52] compute response maps of sparse coding filters applied to foveated images at each step and select new fixations from the residual perceptual information map(RPI) using the information maximization principle. Adeli et al. [1] first apply a retina transformation to input images, then compute corresponding priority maps and select the next fixation. Furthermore, some researchers [48, 56] attempt to generate new fixations by leveraging low-level semantic information. + +Another subset of work is inspired by the statistical distribution characteristics of human scanpaths. Many researchers [7,8,30,32,34,42,53] in this group attempt to use image saliency information to predict new scanpaths. For example, VSPT [42] uses SalGAN [39] to extract saliency information and employs an autoregressive approach [21] to generate fixations. Others in this category adopt alternative methods, such as Sun et al. [47], who conduct super-Gaussian quality + +analysis on image patches and select the $k$ th fixation to be at the patch with the highest response to the $k$ th super Gaussian component. Coutrot et al. [13] model scanpaths using image and task dependent hidden markov models(HMM). Clarke et al. [11] implement a Gaussian jump distribution, which is polynomially dependent on the previous fixed distribution. + +In addition to the aforementioned models, there are also some deep learning models [5, 28, 55] that have achieved relatively good performance, despite their limited interpretability. + +Scanpath Prediction for $360^{\circ}$ Images Unlike 2D images, there are currently limited models specifically designed for scanpath prediction in $360^{\circ}$ images. However, with the gradual development of VR/AR technologies, this field has attracted increasing attention. In this domain, SaltiNet [6] is the first proposed method, which uses saliency volume to process $360^{\circ}$ images. PathGAN [5] adopts the method of Generative Adversarial [20] to generate scanpath of $360^{\circ}$ image. Zhu et al. [58] proposed a new method to estimate the saliency of $360^{\circ}$ images, and introduced mechanisms such as visual uncertainty and visual balance to predict the scanpath. The subsequent ScanGAN [35] is also a method that uses Generative Adversarial Networks [20] to generate scapath. Different from PathGAN [5], ScanGAN [35] adds a soft-dtw loss function [15] to the generator to generate scanpath that are more similar to real scanpath. ScanDMM [45] is a method that uses Deep Markov model [27] to generate scanpath, which takes into account the past historical information. + +All above methods predict fixations as well as scanpaths in 2D equirectangular projection of $360^{\circ}$ image, which has the weaknesses of distortion and coordinate discontinuity and always introduce big prediction error. Different from them, we perform scanpath prediction for $360^{\circ}$ images in 3D spherical coordinate system, which is more consistent with the real-world scenarios of human exploring immersive environments. + +# 3 Approach + +# 3.1 Problem Definition + +A scanpath can be represented as a sequence of fixations $X = (p_{1}, p_{2}, \ldots, p_{T}) \in \mathbb{R}^{T \times 3}$ , where $T$ represents the length of the scanpath. For a given $360^{\circ}$ image, the task of a scanpath prediction model is to generate a scanpath $\widehat{X}$ that resembles a real human scanpath. We use a 3D Cartesian coordinate system to represent coordinates, where each fixation is given in the form of $p = (x, y, z)$ , with $x$ , $y$ , and $z$ being real numbers ranging from -1 to 1. This representation form helps address the issue of coordinate discontinuity. + +# 3.2 Overview + +The architecture of our model is shown in Fig. 1. For each input $360^{\circ}$ image, we first transform it into 3D spherical coordinate system, and then extract the + +![](images/874335bb1b481e953e4d5ee6aec4ebfffad20fb561e97e054a4f218b81b583c5.jpg) +Fig. 1: Overall architecture of our methods. The Transformer Encoder is utilized to contextualize the feature extracted by the SphereNet [12]. The contextual feature representation and the historical fixations are fed into the Transformer Decoder to output fixation embeddings. Each fixation embedding is input into a 3D Mixture Density Network to output its 3D Gaussian distribution, from which the fixation is sampled. + +image features using SphereNet [12]. Subsequently, we feed the extracted feature representation into a 3D Transformer Encoder to contextualize them. The Transformer Decoder takes historical fixation information and contextual feature representations as input and generates embedding information for new fixations in an autoregressive manner [21]. Then, a 3D Mixture Density Network (MDN) takes the embedding information of the fixations as input and outputs their 3D Gaussian position distributions. Finally, the fixations are sampled from the 3D Gaussian distributions. Next, we will introduce a detailed module-wise description of our model. + +# 3.3 3D Contextual Visual Feature Representation + +3D feature extraction. We perform scanpath prediction in 3D spherical coordinate system. Therefore, after projecting the $360^{\circ}$ image into 3D spherical coordinate system, we utilize SphereNet [12] to extract its 3D feature representation. Specifically, for a $360^{\circ}$ image $I$ , after inputting it into SphereNet, we extract 3 different scales of feature maps and concatenate them along the channel axis, resulting in a feature representation $F$ with dimensions of $C = 448$ , $H = 128$ , $W = 256$ . + +Before inputting into the 3D Transformer Encoder, we first execute a simple size transformation to $F$ . Firstly, we execute average pooling to $F$ in a convolutional kernel of size $8 \times 8$ , to reduce the size of the feature maps to $C = 448$ , $H = 16$ , $W = 32$ . Next, we flatten the pooled feature maps, resulting + +in a feature $F_{L}$ with size of $L = 16 \times 32 = 512$ , $C = 448$ . This series of operations can be formulated as: + +$$ +F = \operatorname {S p h e r e N e t} (I, \theta) \tag {1} +$$ + +$$ +F _ {L} = \operatorname {F l a t t e n} (\operatorname {A v g P o o l} (F)) \tag {2} +$$ + +here, $\theta$ represents the parameters of SphereNet. + +3D feature encoder. A linear embedding layer is first applied to map $F_{L}$ into feature embeddings. Subsequently, the obtained feature embeddings are added to a 3D positional encoding and fed into $M = 4$ standard Transformer Encoder Layers to extract contextual feature representations. The above process can be expressed by the following formulas. + +$$ +F _ {L} ^ {\prime} = \operatorname {E m b e d d i n g} \left(F _ {L}\right) \tag {3} +$$ + +$$ +P _ {0} = F _ {L} ^ {\prime} + E _ {\text {p o s} 3 d} \quad E _ {\text {p o s} 3 d} \in R ^ {L \times D} \tag {4} +$$ + +$$ +P _ {i} = \operatorname {E n c o d e r L a y e r} \left(P _ {i - 1}\right) \quad i = 1 \dots M \tag {5} +$$ + +where $D$ represents the feature embedding dimension of the 3D Transformer Encoder, $M$ represents the number of EncoderLayer in the architecture, and $P_{i}$ represents the output of the i-th EncoderLayer. + +# 3.4 Fixation Sequence Generation + +Fixation decoder. The contextual feature representation and historical fixation information are input into a Transformer decoder to output the fixation embedding. The fixation embedding is then fed into a 3D mixture density network to output the fixation's 3D Gaussian distribution, from which the fixation is sampled. We use autoregressive method [21] to predict one fixation each time and iteratively generate the whole fixation sequence. + +On time step $t$ , we initialize the fixation query vector $q_{t}$ using fixation $p_{t-1}$ of the previous time step. Specially, for the first fixation query $q_{1}$ , we initialize it using the center of 3D space, i.e., $p_{0} = (0,0,0)$ . Subsequently, the initialized query vector $q_{t}$ is concatenated with the historical query vectors to generate the joint query vector $Q_{0}'$ . After adding one-dimensional positional encoding information, $Q_{0}'$ becomes $Q_{0}$ . Finally, after passing through $N = 4$ standard Transformer Decoder Layers, we obtain the hidden states $Q_{N} = (Z_{1}, Z_{2}, \ldots, Z_{t})$ of the fixations. This process is defined as follows: + +$$ +q _ {t} = \operatorname {E m b e d d i n g} \left(p _ {t - 1}\right) \quad t \in [ 1, 2, \dots , T ] \tag {6} +$$ + +$$ +Q _ {0} ^ {\prime} = \left[ q _ {1}; q _ {2}; \dots ; q _ {t} \right] \tag {7} +$$ + +$$ +Q _ {0} = Q _ {0} ^ {\prime} + E _ {p o s \_ 1 d} \quad E _ {p o s \_ 1 d} \in R ^ {t \times D} \tag {8} +$$ + +$$ +Q _ {i} = \operatorname {D e c o d e r L a y e r} \left(Q _ {i - 1}, P _ {M}\right) \quad i = 1 \dots N \tag {9} +$$ + +where $P_{M}$ represents the output of the Transformer Encoder, $T$ represents the length of the predicted fixation sequence, $p_{t-1}$ represents the fixation at the time + +step $t - 1$ , $N$ represents the number of decoder layers. $Q_{i}$ represents the output of the i-th DecoderLayer. + +The fixation decoder process in the training process has some difference from that in the generation process. During the training phase, we input the fixations $(p_0,p_1,\ldots ,p_{T - 1})\in R^{T\times 3}$ into the Transformer Decoder, and it can predict the hidden states $Q_{N} = (Z_{1},Z_{2},\dots,Z_{T})$ of the true scanpath $(p_1,p_2,\dots,p_T)\in$ $R^{T\times 3}$ in a parallel manner, which is a parallel process. During the generation phase, we iteratively generate the hidden states of new coordinates $Z_{1},Z_{2},\ldots Z_{T}$ in sequence starting from the initial coordinate $p_0$ , which is a sequential process. + +Fixation generation. After obtaining $Z_{t}$ , we then employ a 3D mixture density network to model the probability distribution of the fixations. The 3D mixture density network decodes the hidden state $Z_{t}$ of the fixation at time step $t$ into $K$ sets of Gaussian kernel parameters. We use $\mu_t^i = (\mu_{x,t}^i, \mu_{y,t}^i, \mu_{z,t}^i)$ , $\Sigma_t^i$ , and $\pi_t^i$ to represent the mean vector, covariance matrix, and weight of the $i$ -th Gaussian kernel at time step $t$ , respectively. These three parameters can be obtained through three separate MLP layers, which can be defined as: + +$$ +\left\{\mu_ {t} ^ {i}, \Sigma_ {t} ^ {i}, \pi_ {t} ^ {i} \right\} = f _ {m d n} \left(Z _ {t}; \theta_ {m d n}\right) \quad i = 1 \dots K \tag {10} +$$ + +where $\theta_{mdn}$ refers to the parameters of the mixture density network. + +Next, the probability density function for a 3D Gaussian distribution $P(p_{t})$ can be defined as follows: + +$$ +\mathcal {N} \left(p _ {t} \mid \mu_ {t} ^ {i}, \Sigma_ {t} ^ {i}\right) = \frac {\exp \left(- \frac {1}{2} \left(p _ {t} - \mu_ {t} ^ {i}\right) ^ {\top} \left(\Sigma_ {t} ^ {i}\right) ^ {- 1} \left(p _ {t} - \mu_ {t} ^ {i}\right)\right)}{\sqrt {(2 \pi) ^ {3} \left| \Sigma_ {t} ^ {i} \right|}} \tag {11} +$$ + +$$ +P \left(p _ {t}\right) = \sum_ {i = 1} ^ {K} \pi_ {t} ^ {i} \mathcal {N} \left(p _ {t} \mid \mu_ {t} ^ {i}, \Sigma_ {t} ^ {i}\right) \tag {12} +$$ + +where the $|\Sigma_t^i|$ means the determinant of the covariance matrix $\Sigma_t^i$ and the $\mathcal{N}(p_t\mid \mu_t^i,\Sigma_t^i)$ represents the Gaussian distribution function associated with the $i$ -th 3D Gaussian function. + +The prediction fixation $\hat{p}_t$ is sampled probabilistically from a 3D surface, which can be expressed as: + +$$ +\hat {p} _ {t} \sim P \left(p _ {t}\right) \tag {13} +$$ + +here, it is worth noting that we do not simply sample the fixation with the highest probability, since such sampling strategy always makes the model rapidly converge to the same fixation positions for all images. In our study, we generated a total of 32,768 fixation points along the latitude-longitude grid, with a resolution of $128 \times 256$ . We independently computed the probability for each fixation point. Finally, we randomly selected a fixation point based on its probability, with higher probabilities being more likely to be chosen. + +# 3.5 Loss Function + +During training, the objective is to minimize the discrepancy between the predicted distribution and the real human scanpath. To achieve this, we employ the negative log-likelihood loss function, which encourages higher probabilities for the ground truth fixations. The loss function can be defined as follows: + +$$ +\mathcal {L} = - \frac {1}{T} \sum_ {t = 1} ^ {T} \log \left(\sum_ {i = 1} ^ {K} \pi_ {t} ^ {i} \mathcal {N} \left(p _ {t} ^ {*} \mid \mu_ {t} ^ {i}, \Sigma_ {t} ^ {i}\right)\right) \tag {14} +$$ + +where $p_t^*$ represents the real human fixation position of time step $t$ . + +# 4 Experiments + +# 4.1 Datasets + +We conducted experiments on four $360^{\circ}$ image datasets: Sitzmann [44], Salient360! [43], AOI [54], and JUFE [19]. The Sitzmann dataset [44] comprises 22 images with a total of 1920 scanpaths. Only the training set of the Salient360! dataset [43] is available, which contains 85 images and 3036 scanpaths. The AOI dataset [54] consists of 600 high-resolution images and approximately 18,000 scanpaths. The JUFE dataset [19] includes 1032 images and 30,960 scanpaths. For training, 19 images from Sitzmann [44] and 60 images from Salient360! [43] were utilized, and all remaining data were used for validation. To increase the number of training sets, we rotated the image along with the corresponding scanpaths six times along the direction of longitude. We followed ScanDMM [45] to preprocess the gaze data of Sitzmann, Salient360!, and JUFE datasets, and obtain visual scanpaths containing 30, 25, and 15 fixations for Sitzmann, Salient360!, and JUFE. The AOI dataset has already been pre-processed, with an average scanpath length of approximately 8. + +# 4.2 Evaluation Metrics + +We employed six metrics to assess the performances of our model and other models, including Levenshtein Distance (LEV) [41], Dynamic Time Warping (DTW) [25], Time Delay Embedding (TDE), ScanMatch [14], Recursive Metric (REC) [3] and Sequence Score(SS) [55]. LEV [41] is a metric that quantifies the dissimilarity between two strings by measuring the minimum number of single-character edit operations needed to transform one string into another. ScanMatch [14] involves converting eye-tracking scanpaths into string sequences and then employing an algorithm similar to the Levenshtein distance to compare the similarity of the two sequences. DTW [25] and TDE [52] are metrics based on time series analysis, while REC [3] is a metric for recursive quantification analysis. SS [55] is calculated by converting scanpaths into strings of fixation cluster IDs, and then measuring the similarity between two strings using a string matching algorithm [37]. + +For each image, every model generates $m = 10$ predicted scanpaths $\widehat{X}$ , which are evaluated against all real human scanpaths $X$ , and then averaged to determine the final prediction score. Note that our model predicts scanpath with a length of 30. When comparing across different datasets, we truncate the predicted scanpath to the length of the actual scanpath before conducting the comparison. Taking DTW [25] as an example, the score can be calculated as follows: + +$$ +D T W = \frac {1}{n} \frac {1}{m} \sum_ {i = 1} ^ {n} \sum_ {j = 1} ^ {m} D T W \left(\widehat {X} _ {j}, X _ {i}\right) \tag {15} +$$ + +where $\widehat{X}_j$ is the $j$ -th predicted scanpath, $X_i$ is the $i$ -th real human scanpath, $n$ is the number of the real human scanpaths, $m$ is the number of predicted scanpaths. + +# 4.3 Implementation Details + +The internal vector dimension of 3D Transformer Encoder is set to $D = 128$ . We utilize a multi-head attention mechanism with 8 attention heads. The hidden layer size of the feed-forward network (FFN) is set to 64. In the 3D mixture density network model, we configure $K = 5$ Gaussian kernels, and the internal hidden layer dimension is set to 16. During the training process, our batch size is set to 18. As for the optimizer, we use AdamW [26] with a learning rate of 1e-5. We employ a learning rate warm-up and stage-wise adjustment strategy. The warm-up period lasts for 10 epochs, and the learning rate is halved every 10 epochs. The total training duration is 50 epochs. + +Our experiments were conducted on a single NVIDIA GeForce RTX 3090, and the development was conducted using the PyTorch framework. + +# 4.4 Performance Comparison + +We compared our model with SaliTiNet [6], ScanGAN [35], and ScanDMM [45], utilizing the models provided officially by their respective sources to obtain prediction results. These three models are all designed for scanpath prediction on $360^{\circ}$ images. PathGAN [5] and Zhu et al. [58] are two additional models specifically designed for generating scanpath on $360^{\circ}$ images. However, it has been demonstrated that these models are unable to effectively generate realistic scanpaths [35]. Therefore, we have not included them in our comparison. + +We additionally established two baseline reference values to assess the relative quality of the generated scanpath. The first one is the random baseline (Random walk). This baseline randomly generate ten scanpaths for each image, which provides the reference for lowest performance. The second one is the human baseline (Human), which compares each real human scanpath in the dataset with all the other real human scanpaths and taking the average. This baseline represents the average similarity among real human scanpaths, serving as a reference for highest performance. + +Quantitative evaluation. Tab. 1 shows the quantitative comparison results between our method and other competitors on the metrics of LEV [41], DTW [25], TDE [52], ScanMatch [14], REC [3] and SS [55]. As we can see, our method outperforms all the competitors on all four datasets, achieving performance close to 'Human'. + +Table 1: Performance of scanpath prediction models on different datasets. The best result for each metric is highlighted in bold, while the second-best result is underlined. + +
MethodSitzmannSalient360!
LEV ↓DTW ↓TDE ↓ScanMatch ↑REC ↑SS ↑LEV ↓DTW ↓TDE ↓ScanMatch ↑REC ↑SS ↑
Random walk51.5222372.38728.0380.3951.5140.19843.6702062.08029.7530.3751.3810.176
SaltiNet [6]51.1512246.05726.3030.3741.7340.19841.0421860.39725.0010.4292.2500.220
ScanGAN [35]45.8521950.48819.0600.4772.9600.26538.9771752.37119.5270.4603.0410.244
ScanDMM [45]45.7481939.98019.0980.4853.1560.27038.5451706.66120.2380.4673.1250.257
Ours44.6911939.66219.8870.4864.1540.28036.2771515.88319.2610.4844.2500.302
Human42.0061841.53416.1920.5265.1630.36635.0901376.43916.6500.5135.0660.377
MethodAOIJUFE
LEV ↓DTW ↓TDE ↓ScanMatch ↑REC ↑SS ↑LEV ↓DTW ↓TDE ↓ScanMatch ↑REC ↑SS ↑
Random walk15.866666.65238.1640.2935.6500.24126.4391270.89338.1640.3521.8470.191
SaltiNet [6]14.978622.07935.7240.3376.3060.24526.1151264.49232.2610.3641.9720.192
ScanGAN [35]14.473569.54932.6020.3617.3790.25824.2591110.82625.9370.4313.2080.217
ScanDMM [45]14.185569.17032.5580.3717.3280.24923.3721081.46525.7870.4523.5270.221
Ours13.904561.89530.9870.3768.2860.22522.5951054.22322.9550.4624.4490.218
Human14.317501.95230.2480.36713.0080.37718.4451030.97020.3050.5647.8640.330
+ +Qualitative evaluation. We visualized the prediction results of all methods on one image of each dataset, as shown in Fig. 2. Similarly, we visualized the ground truth of human scanpath accordingly. The transition from purple to red in the scanpath indicates the beginning and end of the scanpath, respectively. It is noticeable that most ground truth scanpaths are located within the salient regions near the equator. SaltiNet [6], although designed for $360^{\circ}$ images, suffers from large displacements. In comparison, the methods proposed by ScanGAN [35] and ScanDMM [45] have the capability to generate relatively realistic scanpaths to some extent. However, our model, in contrast to these approaches, excels at capturing salient information within $360^{\circ}$ images and is capable of generating scanpath that are closer to the real human scanpaths. In addition, a notable flaw in the tower image is that our model tends to focus more on the prominent plants at both ends rather than the central tower, which is a common issue shared by other models as well. + +# 4.5 Saliency Comparison + +A good scanpath prediction model often assigns more attention to salient regions during the prediction process. In our study, we generated 1000 scanpath and performed post-processing on all fixations to generate corresponding saliency + +![](images/f4aad0f1d160d12dc5b8a0092434cb1cafde543cb2325ee239228189b10f8503.jpg) +Fig. 2: Qualitative representation results of different models on four datasets. From top to bottom, the four images are sourced from Sitzmann [44], Salient360! [43], AOI [54], and JUFE [19], respectively. From left to right, we display the a scanpath sampled from ground truth, our model, ScanDMM [45], ScanGAN [35], and SaltiNet [6]. + +maps [49]. The fixation maps was obtained from all unprocessed fixations. We used sizes $\mathrm{H} = 128$ and $\mathrm{W} = 256$ for saliency maps (salmap) and fixation maps (fixmap). We compared our results with ScanGAN [35] and ScanDMM [45] on the Salient360! [43] and JUFE [19] datasets, under the saliency metrics AUC_Judd [9], NSS [40], CC [31], SIM [9], and KLD [29]. + +Table 2: Quantitative comparison results of significance comparison on Salient360! [43] and JUFE [19] datasets. The best result for each metric is highlighted in bold, while the second-best result is underlined. + +
DatasetMethodAUC_Judd↑NSS ↑CC ↑SIM ↑KLD ↓
Salient360! [43]ScanGAN [35]0.7970.9860.2600.17321.210
ScanDMM [45]0.8301.2100.3220.18320.882
Ours0.8381.1950.3520.25918.985
JUFE [19]ScanGAN [35]0.7950.8370.1980.23619.529
ScanDMM [45]0.8661.2030.2750.28618.248
Ours0.8721.0350.2450.32016.339
+ +Based on Tab. 2, our model has demonstrated favorable performance across the majority of metrics. Our model did not achieve optimal results in a few metrics, one potential explanation is that our model exhibits a stronger emphasis on larger salient objects, which may lead to comparatively less attention towards smaller targets compared to models with more evenly distributed saliency maps. As evident in Fig. 3, ScanDMM [45] and ScanGAN [35] exhibit a characteristic of having more dispersed saliency maps. Furthermore, it can be observed that our model has the ability to focus more on salient regions compared to the other + +![](images/441392e9cb3366db93419d574e62c506a078fcae62b96e2f7d159564535b5718.jpg) +Fig. 3: Qualitative comparison results of significance comparison. The first two images is from Salient360! [43] and the next three images are from JUFE [19]. From top to bottom are the real image, the significance maps of the real images, Ours, ScanDMM [45], ScanGAN [35]. + +Table 3: Efficiency comparison in terms of model size and running time + +
SaltiNet [6]ScanGAN [35]ScanDMM [45]Ours
Paramters(MB)98.7315.13417.77919.280
Running Time(ms)127206.482118.250266.420
+ +models. As a result, the post-processed saliency maps obtained from our model exhibit a closer resemblance to human-generated saliency maps. + +# 4.6 Efficiency Comparison + +We also compared our model's size and running time for generating a scanpath with other methods. In Tab. 3, we present the model size and the runtime for generating a single scanpath (measured in milliseconds). As can be seen, we have a similar model size with other models. In terms of runtime, our model is capable of generating approximately four scanpaths of length 30 per second, which is a little slower compared to ScanGAN and ScanDMM. This is primarily because our model generates fixations in an autoregressive manner [21]. Additionally, the use of a mixture density network is another factor leading to slower running speed. However, compared to a single Gaussian distribution, multiple Gaussian distributions can more accurately represent the distribution of fixations. + +# 4.7 Ablation Studies + +We conducted ablation experiments on the Salient360! [43] and JUFE [19] datasets. The experimental results are presented in Tab. 4, showing the optimal results + +obtained for each configuration. Next, we will provide a brief overview of each model architecture. For more detailed architecture information, please refer to the supplementary file. + +Table 4: Ablation studies on Salient360! [43] and JUFE [19] datasets. The best result for each metric is highlighted in bold, while the second-best result is underlined. + +
MethodSalient360!JUFE
LEV ↓DTW ↓TDE ↓ScanMatch ↑REC ↑SS ↑LEV ↓DTW ↓TDE ↓ScanMatch ↑REC ↑SS ↑
Pure ViT37.6431739.71419.2910.4703.7100.26722.8911072.75323.7510.4573.8470.221
Pure 2D CNN36.8481727.24119.6410.4774.1190.27222.6131060.20523.0960.4634.1810.220
Saliency38.1961864.91420.0590.4753.1560.25323.8331094.29024.6360.4423.6690.208
w/o EncoderLayer37.3521798.34020.7270.4634.1710.26222.9261072.26323.1960.4554.2600.218
w/o MDN + MSE Loss39.4631793.94825.9140.4263.7810.09523.9151132.12725.6750.4163.8330.102
K=138.6431804.07420.0740.4623.3260.24823.2391084.73423.5770.4513.8050.218
K=336.2502073.90522.3260.4904.0390.24622.7581077.36523.3020.4604.3600.214
K=838.1171700.87720.4550.4673.3890.25822.9171065.24023.4700.4553.9140.210
Parallel36.7221577.99323.7260.4893.9390.28822.5671054.64428.7750.4513.6880.221
Ours36.2771515.88319.2610.4844.2530.30222.5951054.22322.9550.4624.4490.218
+ +The feature extractor. To demonstrate the effectiveness of our 3D feature extractor, SphereNet, we try to employ three distinct architectures for feature extraction. Firstly, we extracted features using the patch-based approach of the Vision Transformer (ViT) method [16] (denoted as 'Pure ViT'). Secondly, we replaced SphereNet with a standard 2D convolutional network (denoted as 'Pure 2D CNN'). Lastly, We followed VSPT [42] to extract image features based on the saliency information. Specifically, we employed EPSNet [59] to extract saliency information (denoted as 'Saliency'). As shown in Tab. 4, 'Ours' obtains better performance than 'Pure ViT', 'Pure 2D CNN', and 'Saliency', which demonstrates the superiority of our feature extractor. + +The 3D Transformer encoder. To validate the effectiveness of our 3D Transformer Encoder, we attempted to remove the EncoderLayer in the 3D Transformer Encoder and directly utilize the feature embedding of the Transformer Encoder as input of the Transformer Decoder (denoted as 'w/o EncoderLayer'). As we can see from Tab. 4, without our 3D Transformer Encoder, the scanpath prediction performance dropped significantly, which demonstrates the effectiveness of our 3D Transformer Encoder. + +3D MDN. To validate the effectiveness of the 3D MDN, we directly exploited linear regression to predict the fixations after the Transformer Decoder. The Mean Squared Error (MSE) was utilized as the loss function (denoted as 'w/o MDN + MSE Loss'). As can be seen in Tab. 4, 'Ours' outperforms 'w/o MDN + MSE Loss' a lot, proving the advantage of our 3D MDN. + +Different number of Gaussian kernels. We compared with the model results obtained by using different numbers of Gaussian kernels in the mixture density network. These comparisons are respectively presented in the rows labeled "K=1,3,8," while our model has 5 Gaussian kernels. As we can see from Tab. 4, using 5 Gaussian kernels can provide a better representation of the fixations' distribution. + +Parallel in the Transformer decoder. In our Transformer decoder, we use autoregressive method to predict one fixation each time and iteratively generate the whole fixation sequence. Through our observation, more and more models try to adopt a parallel structure in the Transformer decoder [10, 23, 36] for higher efficiency. Therefore, we also tried to build a parallel Transformer decoder in our method (denoted as 'Parallel'). In the parallel model, we use a random embedding with length $L$ as the input to the Transformer decoder, and we can get the hidden state of a scanpath of length $L$ at one time, all the other parts of the model are unchanged. + +We can observe from Tab. 4 that the parallel structure of the decoder performs poorly than the autoregressive Transformer decoder, especially in the TDE [52] metric. Such comparison results demonstrate that an autoregressive Transformer decoder is more suitable for our scanpath prediction task. + +# 5 Limitaions and Future Works + +In our work, we employed SphereNet [12] for feature extraction and 3D Mixture Density Network (MDN) for fixation generation. However, we acknowledge that these components are replaceable, and we are contemplating whether there might be better alternatives for both parts. + +In addition, we are considering incorporating relevant semantic information to improve the performance further, as objects with important semantic information tend to attract human attention more easily. + +# 6 Conclusion + +Considering that the 2D equirectangular projection of $360^{\circ}$ images has the issues of distortion and coordinate discontinuity, we proposed to execute scanpath prediction for $360^{\circ}$ images in 3D spherical coordinate system. Correspondingly, we proposed a novel 3D Scanpath Transformer named Pathformer3D, where a 3D Transformer Encoder was first used to obtain 3D contextual feature representation of the $360^{\circ}$ image. Then, Pathformer3D input the 3D contextual feature representation and historical fixation information into the Transformer Decoder to predict the hidden state for each fixation. Each fixation's hidden state was then fed into a 3D mixture density network to output the 3D Gaussian distribution of the fixation, from which the fixation location can be sampled. Finally, Pathformer3D model was compared with three state-of-the-art $360^{\circ}$ image scan-path prediction models, to demonstrate the superiority of its performance. + +Acknowledgements This work is supported in part by the National Natural Science Foundation of China under grant 62206127 and 62272229, the Natural Science Foundation of Jiangsu Province under grant BK20222012, and Shenzhen Science and Technology Program JCYJ20230807142001004. Rong Quan and Yantao Lai contributed equally to this work. Dong Liang is the corresponding author. + +# References + +1. Adeli, H., Vitu, F., Zelinsky, G.J.: A model of the superior colliculus predicts fixation locations during scene viewing and visual search. Journal of Neuroscience 37(6), 1453-1467 (2017) +2. Aggarwal, A., Mittal, M., Battineni, G.: Generative adversarial network: An overview of theory and applications. International Journal of Information Management Data Insights 1(1), 100004 (2021) +3. Anderson, N.C., Bischof, W.F., Laidlaw, K.E., Risko, E.F., Kingstone, A.: Recurrence quantification analysis of eye movements. Behavior research methods 45, 842-856 (2013) +4. Antol, S., Agrawal, A., Lu, J., Mitchell, M., Batra, D., Zitnick, C.L., Parikh, D.: Vqa: Visual question answering. In: ICCV. pp. 2425-2433 (2015) +5. Assens, M., Giro-i Nieto, X., McGuinness, K., O'Connor, N.E.: Pathgan: Visual scanpath prediction with generative adversarial networks. In: Proceedings of the European Conference on Computer Vision (ECCV) Workshops. pp. 0-0 (2018) +6. Assens Reina, M., Giro-i Nieto, X., McGuinness, K., O'Connor, N.E.: Saltinet: Scan-path prediction on 360 degree images using saliency volumes. In: Proceedings of the IEEE International Conference on Computer Vision Workshops. pp. 2331-2338 (2017) +7. Boccignone, G., Ferraro, M.: Modelling gaze shift as a constrained random walk. Physica A: Statistical Mechanics and its Applications 331(1-2), 207-218 (2004) +8. Brockmann, D., Geisel, T.: The ecology of gaze shifts. Neurocomputing 32, 643-650 (2000) +9. Bylinskii, Z., Judd, T., Borji, A., Itti, L., Durand, F., Oliva, A., Torralba, A.: Mit saliency benchmark (2015) +0. Carion, N., Massa, F., Synnaeve, G., Usunier, N., Kirillov, A., Zagoruyko, S.: End-to-end object detection with transformers. In: European conference on computer vision. pp. 213-229. Springer (2020) +1. Clarke, A.D., Stainer, M.J., Tatler, B.W., Hunt, A.R.: The saccadic flow baseline: Accounting for image-independent biases in fixation behavior. Journal of vision 17(11), 12-12 (2017) +2. Coors, B., Condurache, A.P., Geiger, A.: Spherenet: Learning spherical representations for detection and classification in omnidirectional images. In: Proceedings of the European conference on computer vision (ECCV). pp. 518-533 (2018) +3. Coutrot, A., Hsiao, J.H., Chan, A.B.: Scanpath modeling and classification with hidden markov models. Behavior research methods 50(1), 362-379 (2018) +4. Cristino, F., Mathôt, S., Theeuwes, J., Gilchrist, I.D.: Scanmatch: A novel method for comparing fixation sequences. Behavior research methods 42, 692-700 (2010) +5. Cuturi, M., Blondel, M.: Soft-dtw: a differentiable loss function for time-series. In: International conference on machine learning. pp. 894-903. PMLR (2017) + +16. Dosovitskiy, A., Beyer, L., Kolesnikov, A., Weissenborn, D., Zhai, X., Unterthiner, T., Dehghani, M., Minderer, M., Heigold, G., Gelly, S., et al.: An image is worth 16x16 words: Transformers for image recognition at scale. arXiv preprint arXiv:2010.11929 (2020) +17. Engbert, R., Trukenbrod, H.A., Barthelme, S., Wichmann, F.A.: Spatial statistics and attentional dynamics in scene viewing. Journal of vision 15(1), 14-14 (2015) +18. Fan, C.L., Lee, J., Lo, W.C., Huang, C.Y., Chen, K.T., Hsu, C.H.: Fixation prediction for 360 video streaming in head-mounted virtual reality. In: Proceedings of the 27th Workshop on Network and Operating Systems Support for Digital Audio and Video. pp. 67-72 (2017) +19. Fang, Y., Huang, L., Yan, J., Liu, X., Liu, Y.: Perceptual quality assessment of omnidirectional images. In: Proceedings of the AAAI Conference on Artificial Intelligence. vol. 36, pp. 580-588 (2022) +20. Goodfellow, I., Pouget-Abadie, J., Mirza, M., Xu, B., Warde-Farley, D., Ozair, S., Courville, A., Bengio, Y.: Generative adversarial nets. Advances in neural information processing systems 27 (2014) +21. Graves, A.: Generating sequences with recurrent neural networks. arXiv preprint arXiv:1308.0850 (2013) +22. Itti, L., Koch, C., Niebur, E.: A model of saliency-based visual attention for rapid scene analysis. IEEE Transactions on pattern analysis and machine intelligence 20(11), 1254-1259 (1998) +23. Kamath, A., Singh, M., LeCun, Y., Synnaeve, G., Misra, I., Carion, N.: Mdetr-modulated detection for end-to-end multi-modal understanding. In: Proceedings of the IEEE/CVF International Conference on Computer Vision. pp. 1780-1790 (2021) +24. Kay, A.: User interface: A personal view. The art of human-computer interface design pp. 191-207 (1990) +25. Keogh, E., Ratanamahatana, C.A.: Exact indexing of dynamic time warping. Knowledge and information systems 7, 358-386 (2005) +26. Kinga, D., Adam, J.B., et al.: A method for stochastic optimization. In: International conference on learning representations (ICLR). vol. 5, p. 6. San Diego, California; (2015) +27. Krishnan, R., Shalit, U., Sontag, D.: Structured inference networks for nonlinear state space models. In: Proceedings of the AAAI Conference on Artificial Intelligence. vol. 31 (2017) +28. Kümmerer, M., Bethge, M., Wallis, T.S.: Deepgaze iii: Modeling free-viewing human scanpaths with deep learning. Journal of Vision 22(5), 7-7 (2022) +29. Kümmerer, M., Wallis, T.S., Bethge, M.: Information-theoretic model comparison unifies saliency metrics. Proceedings of the National Academy of Sciences 112(52), 16054-16059 (2015) +30. Le Meur, O., Coutrot, A.: Introducing context-dependent and spatially-variant viewing biases in saccadic models. Vision research 121, 72-84 (2016) +31. Le Meur, O., Le Callet, P., Barba, D.: Predicting visual fixations on video based on low-level visual features. Vision research 47(19), 2483-2498 (2007) +32. Le Meur, O., Liu, Z.: Saccadic model of eye movements for free-viewing condition. Vision research 116, 152-164 (2015) +33. Li, C., Zhang, W., Liu, Y., Wang, Y.: Very long term field of view prediction for 360-degree video streaming. In: 2019 IEEE conference on multimedia information processing and retrieval (MIPR). pp. 297-302. IEEE (2019) + +34. Liu, H., Xu, D., Huang, Q., Li, W., Xu, M., Lin, S.: Semantically-based human scanpath estimation with hmms. In: Proceedings of the IEEE International Conference on Computer Vision. pp. 3232-3239 (2013) +35. Martin, D., Serrano, A., Bergman, A.W., Wetzstein, G., Masia, B.: Scangan360: A generative model of realistic scanpaths for 360 images. IEEE Transactions on Visualization and Computer Graphics 28(5), 2003-2013 (2022) +36. Mondal, S., Yang, Z., Ahn, S., Samaras, D., Zelinsky, G., Hoai, M.: Gazeformer: Scalable, effective and fast prediction of goal-directed human attention. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 1441-1450 (2023) +37. Needleman, S.B., Wunsch, C.D.: A general method applicable to the search for similarities in the amino acid sequence of two proteins. Journal of molecular biology 48(3), 443-453 (1970) +38. Nguyen, A., Yan, Z., Nahrstedt, K.: Your attention is unique: Detecting 360-degree video saliency in head-mounted display for head movement prediction. In: Proceedings of the 26th ACM international conference on Multimedia. pp. 1190-1198 (2018) +39. Pan, J., Ferrer, C.C., McGuinness, K., O'Connor, N.E., Torres, J., Sayrol, E., Giro-i Nieto, X.: Salgan: Visual saliency prediction with generative adversarial networks. arXiv preprint arXiv:1701.01081 (2017) +40. Peters, R.J., Iyer, A., Itti, L., Koch, C.: Components of bottom-up gaze allocation in natural images. Vision research 45(18), 2397-2416 (2005) +41. Privitera, C.M., Stark, L.W.: Algorithms for defining visual regions-of-interest: Comparison with eye fixations. IEEE Transactions on pattern analysis and machine intelligence 22(9), 970-982 (2000) +42. Qiu, M., Rong, Q., Liang, D., Tu, H.: Visual scanpath transformer: Guiding computers to see the world. In: 2023 IEEE International Symposium on Mixed and Augmented Reality (ISMAR). pp. 223-232. IEEE (2023) +43. Rai, Y., Gutierrez, J., Le Callet, P.: A dataset of head and eye movements for 360 degree images. In: Proceedings of the 8th ACM on Multimedia Systems Conference. pp. 205-210 (2017) +44. Sitzmann, V., Serrano, A., Pavel, A., Agrawala, M., Gutierrez, D., Masia, B., Wetzstein, G.: Saliency in vr: How do people explore virtual environments? IEEE transactions on visualization and computer graphics 24(4), 1633-1642 (2018) +45. Sui, X., Fang, Y., Zhu, H., Wang, S., Wang, Z.: Scandmm: A deep markov model of scanpath prediction for $360^{\circ}$ images. In: IEEE Conference on Computer Vision and Pattern Recognition (2023) +46. Sun, W., Chen, Z., Wu, F.: Visual scanpath prediction using ior-roi recurrent mixture density network. IEEE transactions on pattern analysis and machine intelligence 43(6), 2101-2118 (2019) +47. Sun, X., Yao, H., Ji, R., Liu, X.M.: Toward statistical modeling of saccadic eye movement and visual saliency. IEEE Transactions on Image Processing 23(11), 4649-4662 (2014) +48. Tatler, B.W., Brockmole, J.R., Carpenter, R.H.: Latest: A model of saccadic decisions in space and time. Psychological review 124(3), 267 (2017) +49. Upenik, E., Ebrahimi, T.: A simple method to obtain visual attention data in head mounted virtual reality. In: 2017 IEEE International Conference on Multimedia & Expo Workshops (ICMEW). pp. 73-78. IEEE (2017) +50. Verma, J.K., Paul, S.: Advances in Augmented Reality and Virtual Reality. Springer (2022) + +51. Viswanath, B., Mislove, A., Cha, M., Gummadi, K.P.: On the evolution of user interaction in facebook. In: Proceedings of the 2nd ACM workshop on Online social networks. pp. 37-42 (2009) +52. Wang, W., Chen, C., Wang, Y., Jiang, T., Fang, F., Yao, Y.: Simulating human saccadic scanpaths on natural images. In: CVPR 2011. pp. 441-448. IEEE (2011) +53. Xia, C., Han, J., Qi, F., Shi, G.: Predicting human saccadic scanpaths based on iterative representation learning. IEEE Transactions on Image Processing 28(7), 3502-3515 (2019) +54. Xu, M., Yang, L., Tao, X., Duan, Y., Wang, Z.: Saliency prediction on omnidirectional image with generative adversarial imitation learning. IEEE Transactions on Image Processing 30, 2087-2102 (2021) +55. Yang, Z., Huang, L., Chen, Y., Wei, Z., Ahn, S., Zelinsky, G., Samaras, D., Hoai, M.: Predicting goal-directed human attention using inverse reinforcement learning. In: Proceedings of the IEEE/CVF conference on computer vision and pattern recognition. pp. 193-202 (2020) +56. Zanca, D., Melacci, S., Gori, M.: Gravitational laws of focus of attention. IEEE transactions on pattern analysis and machine intelligence 42(12), 2983-2995 (2019) +57. Zhu, Y., Zhai, G., Min, X.: The prediction of head and eye movement for 360 degree images. Signal Processing: Image Communication 69, 15-25 (2018) +58. Zhu, Y., Zhai, G., Min, X., Zhou, J.: The prediction of saliency map for head and eye movements in 360 degree images. IEEE Transactions on Multimedia 22(9), 2331-2344 (2019) +59. Zou, Z., Ye, M., Li, S., Li, X., Dufaux, F.: $360^{\circ}$ image saliency prediction by embedding self-supervised proxy task. IEEE Transactions on Broadcasting (2023) \ No newline at end of file diff --git "a/2024/Pathformer3D_ A 3D Scanpath Transformer for 360\302\260 Images/images.zip" "b/2024/Pathformer3D_ A 3D Scanpath Transformer for 360\302\260 Images/images.zip" new file mode 100644 index 0000000000000000000000000000000000000000..1882aeea9a69bcf5c3210cf2f637501c90325169 --- /dev/null +++ "b/2024/Pathformer3D_ A 3D Scanpath Transformer for 360\302\260 Images/images.zip" @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:8f3323f7de5b83797409caf56cbe2dd7138a0243acb21a69fa282741a08c4c5f +size 516918 diff --git "a/2024/Pathformer3D_ A 3D Scanpath Transformer for 360\302\260 Images/layout.json" "b/2024/Pathformer3D_ A 3D Scanpath Transformer for 360\302\260 Images/layout.json" new file mode 100644 index 0000000000000000000000000000000000000000..426e0c06319d22ef0aa3e061130bb95cce68c9fc --- /dev/null +++ "b/2024/Pathformer3D_ A 3D Scanpath Transformer for 360\302\260 Images/layout.json" @@ -0,0 +1,10374 @@ +{ + "pdf_info": [ + { + "para_blocks": [ + { + "bbox": [ + 140, + 111, + 474, + 148 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 140, + 111, + 474, + 148 + ], + "spans": [ + { + "bbox": [ + 140, + 111, + 474, + 148 + ], + "type": "text", + "content": "Pathformer3D: A 3D Scanpath Transformer for " + }, + { + "bbox": [ + 140, + 111, + 474, + 148 + ], + "type": "inline_equation", + "content": "360^{\\circ}" + }, + { + "bbox": [ + 140, + 111, + 474, + 148 + ], + "type": "text", + "content": " Images" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 160, + 167, + 453, + 182 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 160, + 167, + 453, + 182 + ], + "spans": [ + { + "bbox": [ + 160, + 167, + 453, + 182 + ], + "type": "text", + "content": "Rong Quan" + }, + { + "bbox": [ + 160, + 167, + 453, + 182 + ], + "type": "inline_equation", + "content": "^{1}" + }, + { + "bbox": [ + 160, + 167, + 453, + 182 + ], + "type": "text", + "content": ", Yantao Lai" + }, + { + "bbox": [ + 160, + 167, + 453, + 182 + ], + "type": "inline_equation", + "content": "^{1,2}" + }, + { + "bbox": [ + 160, + 167, + 453, + 182 + ], + "type": "text", + "content": ", Mengyu Qiu" + }, + { + "bbox": [ + 160, + 167, + 453, + 182 + ], + "type": "inline_equation", + "content": "^{2}" + }, + { + "bbox": [ + 160, + 167, + 453, + 182 + ], + "type": "text", + "content": ", and Dong Liang" + }, + { + "bbox": [ + 160, + 167, + 453, + 182 + ], + "type": "inline_equation", + "content": "^{2,3}" + }, + { + "bbox": [ + 160, + 167, + 453, + 182 + ], + "type": "text", + "content": " (x̄)" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 132, + 189, + 481, + 233 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 189, + 481, + 233 + ], + "spans": [ + { + "bbox": [ + 132, + 189, + 481, + 233 + ], + "type": "text", + "content": "1 College of Artificial Intelligence, Nanjing University of Aeronautics and Astronautics, the Key Laboratory of Brain-Machine Intelligence Technology, Ministry of Education, Nanjing, 211106, China rongquan0806@gmail.com, yantaolai@nuaa.edu.cn" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 140, + 234, + 473, + 277 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 140, + 234, + 473, + 277 + ], + "spans": [ + { + "bbox": [ + 140, + 234, + 473, + 277 + ], + "type": "inline_equation", + "content": "^{2}" + }, + { + "bbox": [ + 140, + 234, + 473, + 277 + ], + "type": "text", + "content": " MIIT Key Laboratory of Pattern Analysis and Machine Intelligence, College of Computer Science and Technology, Nanjing University of Aeronautics and Astronautics, Nanjing, China {qmengyu, liangdong}@nuaa.edu.cn" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 135, + 277, + 478, + 300 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 135, + 277, + 478, + 300 + ], + "spans": [ + { + "bbox": [ + 135, + 277, + 478, + 300 + ], + "type": "text", + "content": "3 Shenzhen Research Institute, Nanjing University of Aeronautics and Astronautics, Shenzhen, China" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 159, + 329, + 455, + 549 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 159, + 329, + 455, + 549 + ], + "spans": [ + { + "bbox": [ + 159, + 329, + 455, + 549 + ], + "type": "text", + "content": "Abstract. Scanpath prediction in " + }, + { + "bbox": [ + 159, + 329, + 455, + 549 + ], + "type": "inline_equation", + "content": "360^{\\circ}" + }, + { + "bbox": [ + 159, + 329, + 455, + 549 + ], + "type": "text", + "content": " images can help realize rapid rendering and better user interaction in Virtual/Augmented Reality applications. However, existing scanpath prediction models for " + }, + { + "bbox": [ + 159, + 329, + 455, + 549 + ], + "type": "inline_equation", + "content": "360^{\\circ}" + }, + { + "bbox": [ + 159, + 329, + 455, + 549 + ], + "type": "text", + "content": " images execute scanpath prediction on 2D equirectangular projection plane, which always result in big computation error owing to the 2D plane's distortion and coordinate discontinuity. In this work, we perform scanpath prediction for " + }, + { + "bbox": [ + 159, + 329, + 455, + 549 + ], + "type": "inline_equation", + "content": "360^{\\circ}" + }, + { + "bbox": [ + 159, + 329, + 455, + 549 + ], + "type": "text", + "content": " images in 3D spherical coordinate system and proposed a novel 3D scanpath Transformer named Pathformer3D. Specifically, a 3D Transformer encoder is first used to extract 3D contextual feature representation for the " + }, + { + "bbox": [ + 159, + 329, + 455, + 549 + ], + "type": "inline_equation", + "content": "360^{\\circ}" + }, + { + "bbox": [ + 159, + 329, + 455, + 549 + ], + "type": "text", + "content": " image. Then, the contextual feature representation and historical fixation information are input into a Transformer decoder to output current time step's fixation embedding, where the self-attention module is used to imitate the visual working memory mechanism of human visual system and directly model the time dependencies among the fixations. Finally, a 3D Gaussian distribution is learned from each fixation embedding, from which the fixation position can be sampled. Evaluation on four panoramic eye-tracking datasets demonstrates that Pathformer3D outperforms the current state-of-the-art methods. Code is available at https://github.com/lsztzp/Pathformer3D." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 160, + 560, + 380, + 572 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 160, + 560, + 380, + 572 + ], + "spans": [ + { + "bbox": [ + 160, + 560, + 380, + 572 + ], + "type": "text", + "content": "Keywords: Scanpath " + }, + { + "bbox": [ + 160, + 560, + 380, + 572 + ], + "type": "inline_equation", + "content": "\\cdot" + }, + { + "bbox": [ + 160, + 560, + 380, + 572 + ], + "type": "inline_equation", + "content": "360^{\\circ}" + }, + { + "bbox": [ + 160, + 560, + 380, + 572 + ], + "type": "text", + "content": " image " + }, + { + "bbox": [ + 160, + 560, + 380, + 572 + ], + "type": "inline_equation", + "content": "\\cdot" + }, + { + "bbox": [ + 160, + 560, + 380, + 572 + ], + "type": "text", + "content": " 3D Transformer" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 132, + 592, + 230, + 605 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 592, + 230, + 605 + ], + "spans": [ + { + "bbox": [ + 132, + 592, + 230, + 605 + ], + "type": "text", + "content": "1 Introduction" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 130, + 617, + 482, + 666 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 617, + 482, + 666 + ], + "spans": [ + { + "bbox": [ + 130, + 617, + 482, + 666 + ], + "type": "text", + "content": "Virtual reality (VR) and augmented reality (AR) [50] technologies have made significant advancements in recent years, providing the users immersive experiences. Meanwhile, understanding and imitating the way human beings explore the " + }, + { + "bbox": [ + 130, + 617, + 482, + 666 + ], + "type": "inline_equation", + "content": "360^{\\circ}" + }, + { + "bbox": [ + 130, + 617, + 482, + 666 + ], + "type": "text", + "content": " images of virtual environments is becoming more and more important," + } + ] + } + ], + "index": 8 + } + ], + "discarded_blocks": [], + "page_size": [ + 612, + 792 + ], + "page_idx": 0 + }, + { + "para_blocks": [ + { + "bbox": [ + 130, + 116, + 480, + 152 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 116, + 480, + 152 + ], + "spans": [ + { + "bbox": [ + 130, + 116, + 480, + 152 + ], + "type": "text", + "content": "since it can help realize more practical and fast rendering and thus improve user interaction in the immersive environment [24, 51]. Scanpath prediction on " + }, + { + "bbox": [ + 130, + 116, + 480, + 152 + ], + "type": "inline_equation", + "content": "360^{\\circ}" + }, + { + "bbox": [ + 130, + 116, + 480, + 152 + ], + "type": "text", + "content": " images refers to predict human's gaze shift path when exploring " + }, + { + "bbox": [ + 130, + 116, + 480, + 152 + ], + "type": "inline_equation", + "content": "360^{\\circ}" + }, + { + "bbox": [ + 130, + 116, + 480, + 152 + ], + "type": "text", + "content": " images." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 130, + 156, + 482, + 276 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 156, + 482, + 276 + ], + "spans": [ + { + "bbox": [ + 130, + 156, + 482, + 276 + ], + "type": "text", + "content": "Existing research on scanpath prediction has primarily focused on 2D images [4, 22, 42, 46, 53], while " + }, + { + "bbox": [ + 130, + 156, + 482, + 276 + ], + "type": "inline_equation", + "content": "360^{\\circ}" + }, + { + "bbox": [ + 130, + 156, + 482, + 276 + ], + "type": "text", + "content": " images possess distinct characteristics in the VR/AR context. " + }, + { + "bbox": [ + 130, + 156, + 482, + 276 + ], + "type": "inline_equation", + "content": "360^{\\circ}" + }, + { + "bbox": [ + 130, + 156, + 482, + 276 + ], + "type": "text", + "content": " images provide immersive interactive environments where users can change their perspectives by physically moving their heads, resulting in a wider distribution of fixations. " + }, + { + "bbox": [ + 130, + 156, + 482, + 276 + ], + "type": "inline_equation", + "content": "360^{\\circ}" + }, + { + "bbox": [ + 130, + 156, + 482, + 276 + ], + "type": "text", + "content": " images encompass richer visual information that requires more time for users to process and absorb. Furthermore, the data structure of " + }, + { + "bbox": [ + 130, + 156, + 482, + 276 + ], + "type": "inline_equation", + "content": "360^{\\circ}" + }, + { + "bbox": [ + 130, + 156, + 482, + 276 + ], + "type": "text", + "content": " images differs from that of 2D images, introducing new requirements for data processing and analysis. Therefore, existing 2D image scanpath prediction methods cannot be directly applied to predict scanpaths in " + }, + { + "bbox": [ + 130, + 156, + 482, + 276 + ], + "type": "inline_equation", + "content": "360^{\\circ}" + }, + { + "bbox": [ + 130, + 156, + 482, + 276 + ], + "type": "text", + "content": " images." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 130, + 278, + 482, + 530 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 278, + 482, + 530 + ], + "spans": [ + { + "bbox": [ + 130, + 278, + 482, + 530 + ], + "type": "text", + "content": "Early methods [6, 43, 57] for predicting scanpaths in 360 images involved sampling fixations based on saliency information to obtain the entire scanpath. Subsequently, with the advancement of generative adversarial networks [2], some researchers [5,35] have utilized generative networks to directly generate the entire path from the " + }, + { + "bbox": [ + 130, + 278, + 482, + 530 + ], + "type": "inline_equation", + "content": "360^{\\circ}" + }, + { + "bbox": [ + 130, + 278, + 482, + 530 + ], + "type": "text", + "content": " image. These methods have shown preliminary results but overlook the modeling of time dependency between fixations, which has long been demonstrated to be a very important characteristic of human visual attention mechanism [22], thereby they often lead to unstable prediction results. Recently, some research works [18,33,38,45] have recognized the need for a comprehensive treatment of the temporal dependencies in viewing behavior, and modeled them through recurrent neural networks (RNNs) or Markov chains. However, these methods integrate all historical fixation information into a single hidden unit, and generate the current fixation only from this hidden unit. In this case, the time dependent relationships among fixations can only be modeled indirectly and the influences from historical fixations will be greatly weakened. Besides, all above methods predict fixations on the 2D equirectangular projection of " + }, + { + "bbox": [ + 130, + 278, + 482, + 530 + ], + "type": "inline_equation", + "content": "360^{\\circ}" + }, + { + "bbox": [ + 130, + 278, + 482, + 530 + ], + "type": "text", + "content": " image, which has the issue of coordinate discontinuity caused by longitude (where " + }, + { + "bbox": [ + 130, + 278, + 482, + 530 + ], + "type": "inline_equation", + "content": "-180^{\\circ}" + }, + { + "bbox": [ + 130, + 278, + 482, + 530 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 130, + 278, + 482, + 530 + ], + "type": "inline_equation", + "content": "180^{\\circ}" + }, + { + "bbox": [ + 130, + 278, + 482, + 530 + ], + "type": "text", + "content": " represent the same fixation despite their significant numerical difference), and different degrees of distortion in different locations. Consequently, the fixation predicted on this 2D equirectangular projection will have a large margin of error." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 130, + 533, + 482, + 666 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 533, + 482, + 666 + ], + "spans": [ + { + "bbox": [ + 130, + 533, + 482, + 666 + ], + "type": "text", + "content": "To solve the above problems, we propose a novel scanpath prediction model for " + }, + { + "bbox": [ + 130, + 533, + 482, + 666 + ], + "type": "inline_equation", + "content": "360^{\\circ}" + }, + { + "bbox": [ + 130, + 533, + 482, + 666 + ], + "type": "text", + "content": " images named Pathformer3D, which predicts fixations in 3D spherical coordinates of the " + }, + { + "bbox": [ + 130, + 533, + 482, + 666 + ], + "type": "inline_equation", + "content": "360^{\\circ}" + }, + { + "bbox": [ + 130, + 533, + 482, + 666 + ], + "type": "text", + "content": " image and directly model the time dependent relationships among fixations, to mimic human viewing behavior in immersive environment more realistically. Specifically, given a 2D equirectangular projection of " + }, + { + "bbox": [ + 130, + 533, + 482, + 666 + ], + "type": "inline_equation", + "content": "360^{\\circ}" + }, + { + "bbox": [ + 130, + 533, + 482, + 666 + ], + "type": "text", + "content": " image, we first transform it into 3D spherical coordinate system, and then exploit a spherical convolution to extract its visual features. After extracting the 3D visual features, we exploit a 3D Transformer encoder to learn each " + }, + { + "bbox": [ + 130, + 533, + 482, + 666 + ], + "type": "inline_equation", + "content": "360^{\\circ}" + }, + { + "bbox": [ + 130, + 533, + 482, + 666 + ], + "type": "text", + "content": " image region a contextual feature representation, by considering their long range spatial dependencies between each other. Next, we exploit a Transformer Decoder to learn each time step's fixation embedding directly from the whole image's" + } + ] + } + ], + "index": 5 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "spans": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "type": "text", + "content": "2" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 166, + 91, + 225, + 102 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 166, + 91, + 225, + 102 + ], + "spans": [ + { + "bbox": [ + 166, + 91, + 225, + 102 + ], + "type": "text", + "content": "R. Quan et al." + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 1 + }, + { + "para_blocks": [ + { + "bbox": [ + 130, + 116, + 482, + 223 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 116, + 482, + 223 + ], + "spans": [ + { + "bbox": [ + 130, + 116, + 482, + 223 + ], + "type": "text", + "content": "visual feature and the historical fixations, where the self-attention module is used to mimic human visual working memory mechanism, which can model the time dependencies between the current fixation and all historical fixations. After obtaining each time step's fixation embedding, we employ a 3D mixture density network to learn the 3D Gaussian distribution for each fixation, from which the fixation location is sampled. Using 3D Gaussian distribution to model the positional likelihood of fixations in 3D space takes into account the differences in scanpaths of different individuals, and can thus result in more robust and real scanpaths." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 146, + 224, + 394, + 236 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 146, + 224, + 394, + 236 + ], + "spans": [ + { + "bbox": [ + 146, + 224, + 394, + 236 + ], + "type": "text", + "content": "In conclusion, this work has the following contributions:" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 138, + 242, + 479, + 300 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 242, + 479, + 300 + ], + "spans": [ + { + "bbox": [ + 138, + 242, + 479, + 300 + ], + "type": "text", + "content": "- We perform visual scanpath prediction for " + }, + { + "bbox": [ + 138, + 242, + 479, + 300 + ], + "type": "inline_equation", + "content": "360^{\\circ}" + }, + { + "bbox": [ + 138, + 242, + 479, + 300 + ], + "type": "text", + "content": " images in a 3D spherical coordinate system for the first time, which can realize a more real imitation of human visual exploration process in immersive environment and simultaneously avoid the error caused by the distortion and positional discontinuity in 2D equirectangular projection of " + }, + { + "bbox": [ + 138, + 242, + 479, + 300 + ], + "type": "inline_equation", + "content": "360^{\\circ}" + }, + { + "bbox": [ + 138, + 242, + 479, + 300 + ], + "type": "text", + "content": " image." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 138, + 301, + 480, + 361 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 301, + 480, + 361 + ], + "spans": [ + { + "bbox": [ + 138, + 301, + 480, + 361 + ], + "type": "text", + "content": "- We propose a novel scanpath prediction method named Pathformer3D, to learn contextual feature representation of the " + }, + { + "bbox": [ + 138, + 301, + 480, + 361 + ], + "type": "inline_equation", + "content": "360^{\\circ}" + }, + { + "bbox": [ + 138, + 301, + 480, + 361 + ], + "type": "text", + "content": " images in a 3D spherical coordinate system and directly consider the influences from historical fixations on current fixation, which is more in line with the visual working memory mechanism of human vision system." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 138, + 361, + 480, + 396 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 361, + 480, + 396 + ], + "spans": [ + { + "bbox": [ + 138, + 361, + 480, + 396 + ], + "type": "text", + "content": "- The proposed method is comprehensively evaluated on four eye-tracking datasets of " + }, + { + "bbox": [ + 138, + 361, + 480, + 396 + ], + "type": "inline_equation", + "content": "360^{\\circ}" + }, + { + "bbox": [ + 138, + 361, + 480, + 396 + ], + "type": "text", + "content": " images, consistently achieving state-of-the-art performance, demonstrating the effectiveness of our approach." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 132, + 414, + 242, + 426 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 414, + 242, + 426 + ], + "spans": [ + { + "bbox": [ + 132, + 414, + 242, + 426 + ], + "type": "text", + "content": "2 Related Works" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 130, + 438, + 480, + 474 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 438, + 480, + 474 + ], + "spans": [ + { + "bbox": [ + 130, + 438, + 480, + 474 + ], + "type": "text", + "content": "Scanpath Prediction for 2D Images In recent years, numerous models have been proposed for predicting scanpaths in 2D images. According to the different sources of inspiration, they can be categorized into different classes." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 130, + 474, + 481, + 593 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 474, + 481, + 593 + ], + "spans": [ + { + "bbox": [ + 130, + 474, + 481, + 593 + ], + "type": "text", + "content": "Some of these models draw inspiration from neuroscience and vision science, often leveraging human gaze behavior to guide scanpath prediction. For instance, some researchers [17, 22, 46] have utilized Inhibit-of-Return(IOR) strategies to guide the generation of scanpaths. Additionally, Wang et al. [52] compute response maps of sparse coding filters applied to foveated images at each step and select new fixations from the residual perceptual information map(RPI) using the information maximization principle. Adeli et al. [1] first apply a retina transformation to input images, then compute corresponding priority maps and select the next fixation. Furthermore, some researchers [48, 56] attempt to generate new fixations by leveraging low-level semantic information." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 130, + 594, + 482, + 666 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 594, + 482, + 666 + ], + "spans": [ + { + "bbox": [ + 130, + 594, + 482, + 666 + ], + "type": "text", + "content": "Another subset of work is inspired by the statistical distribution characteristics of human scanpaths. Many researchers [7,8,30,32,34,42,53] in this group attempt to use image saliency information to predict new scanpaths. For example, VSPT [42] uses SalGAN [39] to extract saliency information and employs an autoregressive approach [21] to generate fixations. Others in this category adopt alternative methods, such as Sun et al. [47], who conduct super-Gaussian quality" + } + ] + } + ], + "index": 10 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 202, + 91, + 448, + 103 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 202, + 91, + 448, + 103 + ], + "spans": [ + { + "bbox": [ + 202, + 91, + 448, + 103 + ], + "type": "text", + "content": "Pathformer3D: A 3D Scanpath Transformer for " + }, + { + "bbox": [ + 202, + 91, + 448, + 103 + ], + "type": "inline_equation", + "content": "360^{\\circ}" + }, + { + "bbox": [ + 202, + 91, + 448, + 103 + ], + "type": "text", + "content": " Images" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 474, + 92, + 481, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 474, + 92, + 481, + 100 + ], + "spans": [ + { + "bbox": [ + 474, + 92, + 481, + 100 + ], + "type": "text", + "content": "3" + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 2 + }, + { + "para_blocks": [ + { + "bbox": [ + 130, + 116, + 479, + 175 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 116, + 479, + 175 + ], + "spans": [ + { + "bbox": [ + 130, + 116, + 479, + 175 + ], + "type": "text", + "content": "analysis on image patches and select the " + }, + { + "bbox": [ + 130, + 116, + 479, + 175 + ], + "type": "inline_equation", + "content": "k" + }, + { + "bbox": [ + 130, + 116, + 479, + 175 + ], + "type": "text", + "content": "th fixation to be at the patch with the highest response to the " + }, + { + "bbox": [ + 130, + 116, + 479, + 175 + ], + "type": "inline_equation", + "content": "k" + }, + { + "bbox": [ + 130, + 116, + 479, + 175 + ], + "type": "text", + "content": "th super Gaussian component. Coutrot et al. [13] model scanpaths using image and task dependent hidden markov models(HMM). Clarke et al. [11] implement a Gaussian jump distribution, which is polynomially dependent on the previous fixed distribution." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 130, + 176, + 480, + 212 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 176, + 480, + 212 + ], + "spans": [ + { + "bbox": [ + 130, + 176, + 480, + 212 + ], + "type": "text", + "content": "In addition to the aforementioned models, there are also some deep learning models [5, 28, 55] that have achieved relatively good performance, despite their limited interpretability." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 130, + 228, + 481, + 395 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 228, + 481, + 395 + ], + "spans": [ + { + "bbox": [ + 130, + 228, + 481, + 395 + ], + "type": "text", + "content": "Scanpath Prediction for " + }, + { + "bbox": [ + 130, + 228, + 481, + 395 + ], + "type": "inline_equation", + "content": "360^{\\circ}" + }, + { + "bbox": [ + 130, + 228, + 481, + 395 + ], + "type": "text", + "content": " Images Unlike 2D images, there are currently limited models specifically designed for scanpath prediction in " + }, + { + "bbox": [ + 130, + 228, + 481, + 395 + ], + "type": "inline_equation", + "content": "360^{\\circ}" + }, + { + "bbox": [ + 130, + 228, + 481, + 395 + ], + "type": "text", + "content": " images. However, with the gradual development of VR/AR technologies, this field has attracted increasing attention. In this domain, SaltiNet [6] is the first proposed method, which uses saliency volume to process " + }, + { + "bbox": [ + 130, + 228, + 481, + 395 + ], + "type": "inline_equation", + "content": "360^{\\circ}" + }, + { + "bbox": [ + 130, + 228, + 481, + 395 + ], + "type": "text", + "content": " images. PathGAN [5] adopts the method of Generative Adversarial [20] to generate scanpath of " + }, + { + "bbox": [ + 130, + 228, + 481, + 395 + ], + "type": "inline_equation", + "content": "360^{\\circ}" + }, + { + "bbox": [ + 130, + 228, + 481, + 395 + ], + "type": "text", + "content": " image. Zhu et al. [58] proposed a new method to estimate the saliency of " + }, + { + "bbox": [ + 130, + 228, + 481, + 395 + ], + "type": "inline_equation", + "content": "360^{\\circ}" + }, + { + "bbox": [ + 130, + 228, + 481, + 395 + ], + "type": "text", + "content": " images, and introduced mechanisms such as visual uncertainty and visual balance to predict the scanpath. The subsequent ScanGAN [35] is also a method that uses Generative Adversarial Networks [20] to generate scapath. Different from PathGAN [5], ScanGAN [35] adds a soft-dtw loss function [15] to the generator to generate scanpath that are more similar to real scanpath. ScanDMM [45] is a method that uses Deep Markov model [27] to generate scanpath, which takes into account the past historical information." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 130, + 396, + 480, + 467 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 396, + 480, + 467 + ], + "spans": [ + { + "bbox": [ + 130, + 396, + 480, + 467 + ], + "type": "text", + "content": "All above methods predict fixations as well as scanpaths in 2D equirectangular projection of " + }, + { + "bbox": [ + 130, + 396, + 480, + 467 + ], + "type": "inline_equation", + "content": "360^{\\circ}" + }, + { + "bbox": [ + 130, + 396, + 480, + 467 + ], + "type": "text", + "content": " image, which has the weaknesses of distortion and coordinate discontinuity and always introduce big prediction error. Different from them, we perform scanpath prediction for " + }, + { + "bbox": [ + 130, + 396, + 480, + 467 + ], + "type": "inline_equation", + "content": "360^{\\circ}" + }, + { + "bbox": [ + 130, + 396, + 480, + 467 + ], + "type": "text", + "content": " images in 3D spherical coordinate system, which is more consistent with the real-world scenarios of human exploring immersive environments." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 132, + 483, + 206, + 498 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 483, + 206, + 498 + ], + "spans": [ + { + "bbox": [ + 132, + 483, + 206, + 498 + ], + "type": "text", + "content": "3 Approach" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 132, + 506, + 257, + 517 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 506, + 257, + 517 + ], + "spans": [ + { + "bbox": [ + 132, + 506, + 257, + 517 + ], + "type": "text", + "content": "3.1 Problem Definition" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 130, + 524, + 480, + 608 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 524, + 480, + 608 + ], + "spans": [ + { + "bbox": [ + 130, + 524, + 480, + 608 + ], + "type": "text", + "content": "A scanpath can be represented as a sequence of fixations " + }, + { + "bbox": [ + 130, + 524, + 480, + 608 + ], + "type": "inline_equation", + "content": "X = (p_{1}, p_{2}, \\ldots, p_{T}) \\in \\mathbb{R}^{T \\times 3}" + }, + { + "bbox": [ + 130, + 524, + 480, + 608 + ], + "type": "text", + "content": ", where " + }, + { + "bbox": [ + 130, + 524, + 480, + 608 + ], + "type": "inline_equation", + "content": "T" + }, + { + "bbox": [ + 130, + 524, + 480, + 608 + ], + "type": "text", + "content": " represents the length of the scanpath. For a given " + }, + { + "bbox": [ + 130, + 524, + 480, + 608 + ], + "type": "inline_equation", + "content": "360^{\\circ}" + }, + { + "bbox": [ + 130, + 524, + 480, + 608 + ], + "type": "text", + "content": " image, the task of a scanpath prediction model is to generate a scanpath " + }, + { + "bbox": [ + 130, + 524, + 480, + 608 + ], + "type": "inline_equation", + "content": "\\widehat{X}" + }, + { + "bbox": [ + 130, + 524, + 480, + 608 + ], + "type": "text", + "content": " that resembles a real human scanpath. We use a 3D Cartesian coordinate system to represent coordinates, where each fixation is given in the form of " + }, + { + "bbox": [ + 130, + 524, + 480, + 608 + ], + "type": "inline_equation", + "content": "p = (x, y, z)" + }, + { + "bbox": [ + 130, + 524, + 480, + 608 + ], + "type": "text", + "content": ", with " + }, + { + "bbox": [ + 130, + 524, + 480, + 608 + ], + "type": "inline_equation", + "content": "x" + }, + { + "bbox": [ + 130, + 524, + 480, + 608 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 130, + 524, + 480, + 608 + ], + "type": "inline_equation", + "content": "y" + }, + { + "bbox": [ + 130, + 524, + 480, + 608 + ], + "type": "text", + "content": ", and " + }, + { + "bbox": [ + 130, + 524, + 480, + 608 + ], + "type": "inline_equation", + "content": "z" + }, + { + "bbox": [ + 130, + 524, + 480, + 608 + ], + "type": "text", + "content": " being real numbers ranging from -1 to 1. This representation form helps address the issue of coordinate discontinuity." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 132, + 624, + 208, + 635 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 624, + 208, + 635 + ], + "spans": [ + { + "bbox": [ + 132, + 624, + 208, + 635 + ], + "type": "text", + "content": "3.2 Overview" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 130, + 641, + 480, + 665 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 641, + 480, + 665 + ], + "spans": [ + { + "bbox": [ + 130, + 641, + 480, + 665 + ], + "type": "text", + "content": "The architecture of our model is shown in Fig. 1. For each input " + }, + { + "bbox": [ + 130, + 641, + 480, + 665 + ], + "type": "inline_equation", + "content": "360^{\\circ}" + }, + { + "bbox": [ + 130, + 641, + 480, + 665 + ], + "type": "text", + "content": " image, we first transform it into 3D spherical coordinate system, and then extract the" + } + ] + } + ], + "index": 10 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "spans": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "type": "text", + "content": "4" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 166, + 91, + 225, + 102 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 166, + 91, + 225, + 102 + ], + "spans": [ + { + "bbox": [ + 166, + 91, + 225, + 102 + ], + "type": "text", + "content": "R. Quan et al." + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 3 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 135, + 114, + 479, + 304 + ], + "blocks": [ + { + "bbox": [ + 135, + 114, + 479, + 304 + ], + "lines": [ + { + "bbox": [ + 135, + 114, + 479, + 304 + ], + "spans": [ + { + "bbox": [ + 135, + 114, + 479, + 304 + ], + "type": "image", + "image_path": "874335bb1b481e953e4d5ee6aec4ebfffad20fb561e97e054a4f218b81b583c5.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 130, + 312, + 482, + 369 + ], + "lines": [ + { + "bbox": [ + 130, + 312, + 482, + 369 + ], + "spans": [ + { + "bbox": [ + 130, + 312, + 482, + 369 + ], + "type": "text", + "content": "Fig. 1: Overall architecture of our methods. The Transformer Encoder is utilized to contextualize the feature extracted by the SphereNet [12]. The contextual feature representation and the historical fixations are fed into the Transformer Decoder to output fixation embeddings. Each fixation embedding is input into a 3D Mixture Density Network to output its 3D Gaussian distribution, from which the fixation is sampled." + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_caption" + } + ], + "index": 2 + }, + { + "bbox": [ + 130, + 393, + 482, + 501 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 393, + 482, + 501 + ], + "spans": [ + { + "bbox": [ + 130, + 393, + 482, + 501 + ], + "type": "text", + "content": "image features using SphereNet [12]. Subsequently, we feed the extracted feature representation into a 3D Transformer Encoder to contextualize them. The Transformer Decoder takes historical fixation information and contextual feature representations as input and generates embedding information for new fixations in an autoregressive manner [21]. Then, a 3D Mixture Density Network (MDN) takes the embedding information of the fixations as input and outputs their 3D Gaussian position distributions. Finally, the fixations are sampled from the 3D Gaussian distributions. Next, we will introduce a detailed module-wise description of our model." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 131, + 522, + 391, + 534 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 131, + 522, + 391, + 534 + ], + "spans": [ + { + "bbox": [ + 131, + 522, + 391, + 534 + ], + "type": "text", + "content": "3.3 3D Contextual Visual Feature Representation" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 130, + 545, + 482, + 616 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 545, + 482, + 616 + ], + "spans": [ + { + "bbox": [ + 130, + 545, + 482, + 616 + ], + "type": "text", + "content": "3D feature extraction. We perform scanpath prediction in 3D spherical coordinate system. Therefore, after projecting the " + }, + { + "bbox": [ + 130, + 545, + 482, + 616 + ], + "type": "inline_equation", + "content": "360^{\\circ}" + }, + { + "bbox": [ + 130, + 545, + 482, + 616 + ], + "type": "text", + "content": " image into 3D spherical coordinate system, we utilize SphereNet [12] to extract its 3D feature representation. Specifically, for a " + }, + { + "bbox": [ + 130, + 545, + 482, + 616 + ], + "type": "inline_equation", + "content": "360^{\\circ}" + }, + { + "bbox": [ + 130, + 545, + 482, + 616 + ], + "type": "text", + "content": " image " + }, + { + "bbox": [ + 130, + 545, + 482, + 616 + ], + "type": "inline_equation", + "content": "I" + }, + { + "bbox": [ + 130, + 545, + 482, + 616 + ], + "type": "text", + "content": ", after inputting it into SphereNet, we extract 3 different scales of feature maps and concatenate them along the channel axis, resulting in a feature representation " + }, + { + "bbox": [ + 130, + 545, + 482, + 616 + ], + "type": "inline_equation", + "content": "F" + }, + { + "bbox": [ + 130, + 545, + 482, + 616 + ], + "type": "text", + "content": " with dimensions of " + }, + { + "bbox": [ + 130, + 545, + 482, + 616 + ], + "type": "inline_equation", + "content": "C = 448" + }, + { + "bbox": [ + 130, + 545, + 482, + 616 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 130, + 545, + 482, + 616 + ], + "type": "inline_equation", + "content": "H = 128" + }, + { + "bbox": [ + 130, + 545, + 482, + 616 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 130, + 545, + 482, + 616 + ], + "type": "inline_equation", + "content": "W = 256" + }, + { + "bbox": [ + 130, + 545, + 482, + 616 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 130, + 617, + 482, + 666 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 617, + 482, + 666 + ], + "spans": [ + { + "bbox": [ + 130, + 617, + 482, + 666 + ], + "type": "text", + "content": "Before inputting into the 3D Transformer Encoder, we first execute a simple size transformation to " + }, + { + "bbox": [ + 130, + 617, + 482, + 666 + ], + "type": "inline_equation", + "content": "F" + }, + { + "bbox": [ + 130, + 617, + 482, + 666 + ], + "type": "text", + "content": ". Firstly, we execute average pooling to " + }, + { + "bbox": [ + 130, + 617, + 482, + 666 + ], + "type": "inline_equation", + "content": "F" + }, + { + "bbox": [ + 130, + 617, + 482, + 666 + ], + "type": "text", + "content": " in a convolutional kernel of size " + }, + { + "bbox": [ + 130, + 617, + 482, + 666 + ], + "type": "inline_equation", + "content": "8 \\times 8" + }, + { + "bbox": [ + 130, + 617, + 482, + 666 + ], + "type": "text", + "content": ", to reduce the size of the feature maps to " + }, + { + "bbox": [ + 130, + 617, + 482, + 666 + ], + "type": "inline_equation", + "content": "C = 448" + }, + { + "bbox": [ + 130, + 617, + 482, + 666 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 130, + 617, + 482, + 666 + ], + "type": "inline_equation", + "content": "H = 16" + }, + { + "bbox": [ + 130, + 617, + 482, + 666 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 130, + 617, + 482, + 666 + ], + "type": "inline_equation", + "content": "W = 32" + }, + { + "bbox": [ + 130, + 617, + 482, + 666 + ], + "type": "text", + "content": ". Next, we flatten the pooled feature maps, resulting" + } + ] + } + ], + "index": 7 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 203, + 91, + 448, + 102 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 203, + 91, + 448, + 102 + ], + "spans": [ + { + "bbox": [ + 203, + 91, + 448, + 102 + ], + "type": "text", + "content": "Pathformer3D: A 3D Scanpath Transformer for " + }, + { + "bbox": [ + 203, + 91, + 448, + 102 + ], + "type": "inline_equation", + "content": "360^{\\circ}" + }, + { + "bbox": [ + 203, + 91, + 448, + 102 + ], + "type": "text", + "content": " Images" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 474, + 92, + 480, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 474, + 92, + 480, + 100 + ], + "spans": [ + { + "bbox": [ + 474, + 92, + 480, + 100 + ], + "type": "text", + "content": "5" + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 4 + }, + { + "para_blocks": [ + { + "bbox": [ + 130, + 116, + 480, + 139 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 116, + 480, + 139 + ], + "spans": [ + { + "bbox": [ + 130, + 116, + 480, + 139 + ], + "type": "text", + "content": "in a feature " + }, + { + "bbox": [ + 130, + 116, + 480, + 139 + ], + "type": "inline_equation", + "content": "F_{L}" + }, + { + "bbox": [ + 130, + 116, + 480, + 139 + ], + "type": "text", + "content": " with size of " + }, + { + "bbox": [ + 130, + 116, + 480, + 139 + ], + "type": "inline_equation", + "content": "L = 16 \\times 32 = 512" + }, + { + "bbox": [ + 130, + 116, + 480, + 139 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 130, + 116, + 480, + 139 + ], + "type": "inline_equation", + "content": "C = 448" + }, + { + "bbox": [ + 130, + 116, + 480, + 139 + ], + "type": "text", + "content": ". This series of operations can be formulated as:" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 253, + 148, + 481, + 160 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 253, + 148, + 481, + 160 + ], + "spans": [ + { + "bbox": [ + 253, + 148, + 481, + 160 + ], + "type": "interline_equation", + "content": "F = \\operatorname {S p h e r e N e t} (I, \\theta) \\tag {1}", + "image_path": "11e5d587d8e687637b39d1ec865c68a30e0752a047ff56a57aa73644d2f63aa6.jpg" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 248, + 163, + 481, + 175 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 248, + 163, + 481, + 175 + ], + "spans": [ + { + "bbox": [ + 248, + 163, + 481, + 175 + ], + "type": "interline_equation", + "content": "F _ {L} = \\operatorname {F l a t t e n} (\\operatorname {A v g P o o l} (F)) \\tag {2}", + "image_path": "2cc20c56e0c6eb576d5ad702a1616d8bcac0e2246d063e637d455b1ba473e01f.jpg" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 131, + 182, + 343, + 194 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 131, + 182, + 343, + 194 + ], + "spans": [ + { + "bbox": [ + 131, + 182, + 343, + 194 + ], + "type": "text", + "content": "here, " + }, + { + "bbox": [ + 131, + 182, + 343, + 194 + ], + "type": "inline_equation", + "content": "\\theta" + }, + { + "bbox": [ + 131, + 182, + 343, + 194 + ], + "type": "text", + "content": " represents the parameters of SphereNet." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 130, + 211, + 480, + 270 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 211, + 480, + 270 + ], + "spans": [ + { + "bbox": [ + 130, + 211, + 480, + 270 + ], + "type": "text", + "content": "3D feature encoder. A linear embedding layer is first applied to map " + }, + { + "bbox": [ + 130, + 211, + 480, + 270 + ], + "type": "inline_equation", + "content": "F_{L}" + }, + { + "bbox": [ + 130, + 211, + 480, + 270 + ], + "type": "text", + "content": " into feature embeddings. Subsequently, the obtained feature embeddings are added to a 3D positional encoding and fed into " + }, + { + "bbox": [ + 130, + 211, + 480, + 270 + ], + "type": "inline_equation", + "content": "M = 4" + }, + { + "bbox": [ + 130, + 211, + 480, + 270 + ], + "type": "text", + "content": " standard Transformer Encoder Layers to extract contextual feature representations. The above process can be expressed by the following formulas." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 186, + 278, + 480, + 291 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 186, + 278, + 480, + 291 + ], + "spans": [ + { + "bbox": [ + 186, + 278, + 480, + 291 + ], + "type": "interline_equation", + "content": "F _ {L} ^ {\\prime} = \\operatorname {E m b e d d i n g} \\left(F _ {L}\\right) \\tag {3}", + "image_path": "7ceaaad4a7cd5802b9ca519dedca24b7f39275c609ec8528c1655db6075455f4.jpg" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 188, + 294, + 480, + 308 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 188, + 294, + 480, + 308 + ], + "spans": [ + { + "bbox": [ + 188, + 294, + 480, + 308 + ], + "type": "interline_equation", + "content": "P _ {0} = F _ {L} ^ {\\prime} + E _ {\\text {p o s} 3 d} \\quad E _ {\\text {p o s} 3 d} \\in R ^ {L \\times D} \\tag {4}", + "image_path": "8022da5de4f9caa03178edd029def25a817afcb8cc158520d318ed964d06b249.jpg" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 189, + 310, + 480, + 323 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 189, + 310, + 480, + 323 + ], + "spans": [ + { + "bbox": [ + 189, + 310, + 480, + 323 + ], + "type": "interline_equation", + "content": "P _ {i} = \\operatorname {E n c o d e r L a y e r} \\left(P _ {i - 1}\\right) \\quad i = 1 \\dots M \\tag {5}", + "image_path": "338fb2567c060bf0f7b98bfb959200c86c1035267f5687fcf6490ece74c8ced0.jpg" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 130, + 329, + 480, + 365 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 329, + 480, + 365 + ], + "spans": [ + { + "bbox": [ + 130, + 329, + 480, + 365 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 130, + 329, + 480, + 365 + ], + "type": "inline_equation", + "content": "D" + }, + { + "bbox": [ + 130, + 329, + 480, + 365 + ], + "type": "text", + "content": " represents the feature embedding dimension of the 3D Transformer Encoder, " + }, + { + "bbox": [ + 130, + 329, + 480, + 365 + ], + "type": "inline_equation", + "content": "M" + }, + { + "bbox": [ + 130, + 329, + 480, + 365 + ], + "type": "text", + "content": " represents the number of EncoderLayer in the architecture, and " + }, + { + "bbox": [ + 130, + 329, + 480, + 365 + ], + "type": "inline_equation", + "content": "P_{i}" + }, + { + "bbox": [ + 130, + 329, + 480, + 365 + ], + "type": "text", + "content": " represents the output of the i-th EncoderLayer." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 131, + 381, + 313, + 393 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 131, + 381, + 313, + 393 + ], + "spans": [ + { + "bbox": [ + 131, + 381, + 313, + 393 + ], + "type": "text", + "content": "3.4 Fixation Sequence Generation" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 130, + 399, + 480, + 472 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 399, + 480, + 472 + ], + "spans": [ + { + "bbox": [ + 130, + 399, + 480, + 472 + ], + "type": "text", + "content": "Fixation decoder. The contextual feature representation and historical fixation information are input into a Transformer decoder to output the fixation embedding. The fixation embedding is then fed into a 3D mixture density network to output the fixation's 3D Gaussian distribution, from which the fixation is sampled. We use autoregressive method [21] to predict one fixation each time and iteratively generate the whole fixation sequence." + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 130, + 472, + 481, + 567 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 472, + 481, + 567 + ], + "spans": [ + { + "bbox": [ + 130, + 472, + 481, + 567 + ], + "type": "text", + "content": "On time step " + }, + { + "bbox": [ + 130, + 472, + 481, + 567 + ], + "type": "inline_equation", + "content": "t" + }, + { + "bbox": [ + 130, + 472, + 481, + 567 + ], + "type": "text", + "content": ", we initialize the fixation query vector " + }, + { + "bbox": [ + 130, + 472, + 481, + 567 + ], + "type": "inline_equation", + "content": "q_{t}" + }, + { + "bbox": [ + 130, + 472, + 481, + 567 + ], + "type": "text", + "content": " using fixation " + }, + { + "bbox": [ + 130, + 472, + 481, + 567 + ], + "type": "inline_equation", + "content": "p_{t-1}" + }, + { + "bbox": [ + 130, + 472, + 481, + 567 + ], + "type": "text", + "content": " of the previous time step. Specially, for the first fixation query " + }, + { + "bbox": [ + 130, + 472, + 481, + 567 + ], + "type": "inline_equation", + "content": "q_{1}" + }, + { + "bbox": [ + 130, + 472, + 481, + 567 + ], + "type": "text", + "content": ", we initialize it using the center of 3D space, i.e., " + }, + { + "bbox": [ + 130, + 472, + 481, + 567 + ], + "type": "inline_equation", + "content": "p_{0} = (0,0,0)" + }, + { + "bbox": [ + 130, + 472, + 481, + 567 + ], + "type": "text", + "content": ". Subsequently, the initialized query vector " + }, + { + "bbox": [ + 130, + 472, + 481, + 567 + ], + "type": "inline_equation", + "content": "q_{t}" + }, + { + "bbox": [ + 130, + 472, + 481, + 567 + ], + "type": "text", + "content": " is concatenated with the historical query vectors to generate the joint query vector " + }, + { + "bbox": [ + 130, + 472, + 481, + 567 + ], + "type": "inline_equation", + "content": "Q_{0}'" + }, + { + "bbox": [ + 130, + 472, + 481, + 567 + ], + "type": "text", + "content": ". After adding one-dimensional positional encoding information, " + }, + { + "bbox": [ + 130, + 472, + 481, + 567 + ], + "type": "inline_equation", + "content": "Q_{0}'" + }, + { + "bbox": [ + 130, + 472, + 481, + 567 + ], + "type": "text", + "content": " becomes " + }, + { + "bbox": [ + 130, + 472, + 481, + 567 + ], + "type": "inline_equation", + "content": "Q_{0}" + }, + { + "bbox": [ + 130, + 472, + 481, + 567 + ], + "type": "text", + "content": ". Finally, after passing through " + }, + { + "bbox": [ + 130, + 472, + 481, + 567 + ], + "type": "inline_equation", + "content": "N = 4" + }, + { + "bbox": [ + 130, + 472, + 481, + 567 + ], + "type": "text", + "content": " standard Transformer Decoder Layers, we obtain the hidden states " + }, + { + "bbox": [ + 130, + 472, + 481, + 567 + ], + "type": "inline_equation", + "content": "Q_{N} = (Z_{1}, Z_{2}, \\ldots, Z_{t})" + }, + { + "bbox": [ + 130, + 472, + 481, + 567 + ], + "type": "text", + "content": " of the fixations. This process is defined as follows:" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 184, + 576, + 480, + 588 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 184, + 576, + 480, + 588 + ], + "spans": [ + { + "bbox": [ + 184, + 576, + 480, + 588 + ], + "type": "interline_equation", + "content": "q _ {t} = \\operatorname {E m b e d d i n g} \\left(p _ {t - 1}\\right) \\quad t \\in [ 1, 2, \\dots , T ] \\tag {6}", + "image_path": "1c0cd11b650f94ee7c421aa3737c328599879e370bd97824bba8dc09819d5b42.jpg" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 181, + 590, + 480, + 603 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 181, + 590, + 480, + 603 + ], + "spans": [ + { + "bbox": [ + 181, + 590, + 480, + 603 + ], + "type": "interline_equation", + "content": "Q _ {0} ^ {\\prime} = \\left[ q _ {1}; q _ {2}; \\dots ; q _ {t} \\right] \\tag {7}", + "image_path": "b6dce35a08137060693d8351bb3866e47ae3fb5f18b76d8246f9b723302860f4.jpg" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 180, + 605, + 480, + 620 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 180, + 605, + 480, + 620 + ], + "spans": [ + { + "bbox": [ + 180, + 605, + 480, + 620 + ], + "type": "interline_equation", + "content": "Q _ {0} = Q _ {0} ^ {\\prime} + E _ {p o s \\_ 1 d} \\quad E _ {p o s \\_ 1 d} \\in R ^ {t \\times D} \\tag {8}", + "image_path": "1f573bc46e789d41ffa0d97d35639473e490e56e038d5a73e5bab5dcfa391319.jpg" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 181, + 622, + 480, + 634 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 181, + 622, + 480, + 634 + ], + "spans": [ + { + "bbox": [ + 181, + 622, + 480, + 634 + ], + "type": "interline_equation", + "content": "Q _ {i} = \\operatorname {D e c o d e r L a y e r} \\left(Q _ {i - 1}, P _ {M}\\right) \\quad i = 1 \\dots N \\tag {9}", + "image_path": "b26f0f8974fd0db4c147b66e6979cefb57e6d260fe4361bede9f3021717fdde0.jpg" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 130, + 641, + 481, + 666 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 641, + 481, + 666 + ], + "spans": [ + { + "bbox": [ + 130, + 641, + 481, + 666 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 130, + 641, + 481, + 666 + ], + "type": "inline_equation", + "content": "P_{M}" + }, + { + "bbox": [ + 130, + 641, + 481, + 666 + ], + "type": "text", + "content": " represents the output of the Transformer Encoder, " + }, + { + "bbox": [ + 130, + 641, + 481, + 666 + ], + "type": "inline_equation", + "content": "T" + }, + { + "bbox": [ + 130, + 641, + 481, + 666 + ], + "type": "text", + "content": " represents the length of the predicted fixation sequence, " + }, + { + "bbox": [ + 130, + 641, + 481, + 666 + ], + "type": "inline_equation", + "content": "p_{t-1}" + }, + { + "bbox": [ + 130, + 641, + 481, + 666 + ], + "type": "text", + "content": " represents the fixation at the time" + } + ] + } + ], + "index": 18 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "spans": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "type": "text", + "content": "6" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 166, + 91, + 226, + 102 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 166, + 91, + 226, + 102 + ], + "spans": [ + { + "bbox": [ + 166, + 91, + 226, + 102 + ], + "type": "text", + "content": "R. Quan et al." + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 5 + }, + { + "para_blocks": [ + { + "bbox": [ + 130, + 116, + 479, + 140 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 116, + 479, + 140 + ], + "spans": [ + { + "bbox": [ + 130, + 116, + 479, + 140 + ], + "type": "text", + "content": "step " + }, + { + "bbox": [ + 130, + 116, + 479, + 140 + ], + "type": "inline_equation", + "content": "t - 1" + }, + { + "bbox": [ + 130, + 116, + 479, + 140 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 130, + 116, + 479, + 140 + ], + "type": "inline_equation", + "content": "N" + }, + { + "bbox": [ + 130, + 116, + 479, + 140 + ], + "type": "text", + "content": " represents the number of decoder layers. " + }, + { + "bbox": [ + 130, + 116, + 479, + 140 + ], + "type": "inline_equation", + "content": "Q_{i}" + }, + { + "bbox": [ + 130, + 116, + 479, + 140 + ], + "type": "text", + "content": " represents the output of the i-th DecoderLayer." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 130, + 141, + 482, + 226 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 141, + 482, + 226 + ], + "spans": [ + { + "bbox": [ + 130, + 141, + 482, + 226 + ], + "type": "text", + "content": "The fixation decoder process in the training process has some difference from that in the generation process. During the training phase, we input the fixations " + }, + { + "bbox": [ + 130, + 141, + 482, + 226 + ], + "type": "inline_equation", + "content": "(p_0,p_1,\\ldots ,p_{T - 1})\\in R^{T\\times 3}" + }, + { + "bbox": [ + 130, + 141, + 482, + 226 + ], + "type": "text", + "content": " into the Transformer Decoder, and it can predict the hidden states " + }, + { + "bbox": [ + 130, + 141, + 482, + 226 + ], + "type": "inline_equation", + "content": "Q_{N} = (Z_{1},Z_{2},\\dots,Z_{T})" + }, + { + "bbox": [ + 130, + 141, + 482, + 226 + ], + "type": "text", + "content": " of the true scanpath " + }, + { + "bbox": [ + 130, + 141, + 482, + 226 + ], + "type": "inline_equation", + "content": "(p_1,p_2,\\dots,p_T)\\in" + }, + { + "bbox": [ + 130, + 141, + 482, + 226 + ], + "type": "inline_equation", + "content": "R^{T\\times 3}" + }, + { + "bbox": [ + 130, + 141, + 482, + 226 + ], + "type": "text", + "content": " in a parallel manner, which is a parallel process. During the generation phase, we iteratively generate the hidden states of new coordinates " + }, + { + "bbox": [ + 130, + 141, + 482, + 226 + ], + "type": "inline_equation", + "content": "Z_{1},Z_{2},\\ldots Z_{T}" + }, + { + "bbox": [ + 130, + 141, + 482, + 226 + ], + "type": "text", + "content": " in sequence starting from the initial coordinate " + }, + { + "bbox": [ + 130, + 141, + 482, + 226 + ], + "type": "inline_equation", + "content": "p_0" + }, + { + "bbox": [ + 130, + 141, + 482, + 226 + ], + "type": "text", + "content": ", which is a sequential process." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 130, + 247, + 482, + 331 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 247, + 482, + 331 + ], + "spans": [ + { + "bbox": [ + 130, + 247, + 482, + 331 + ], + "type": "text", + "content": "Fixation generation. After obtaining " + }, + { + "bbox": [ + 130, + 247, + 482, + 331 + ], + "type": "inline_equation", + "content": "Z_{t}" + }, + { + "bbox": [ + 130, + 247, + 482, + 331 + ], + "type": "text", + "content": ", we then employ a 3D mixture density network to model the probability distribution of the fixations. The 3D mixture density network decodes the hidden state " + }, + { + "bbox": [ + 130, + 247, + 482, + 331 + ], + "type": "inline_equation", + "content": "Z_{t}" + }, + { + "bbox": [ + 130, + 247, + 482, + 331 + ], + "type": "text", + "content": " of the fixation at time step " + }, + { + "bbox": [ + 130, + 247, + 482, + 331 + ], + "type": "inline_equation", + "content": "t" + }, + { + "bbox": [ + 130, + 247, + 482, + 331 + ], + "type": "text", + "content": " into " + }, + { + "bbox": [ + 130, + 247, + 482, + 331 + ], + "type": "inline_equation", + "content": "K" + }, + { + "bbox": [ + 130, + 247, + 482, + 331 + ], + "type": "text", + "content": " sets of Gaussian kernel parameters. We use " + }, + { + "bbox": [ + 130, + 247, + 482, + 331 + ], + "type": "inline_equation", + "content": "\\mu_t^i = (\\mu_{x,t}^i, \\mu_{y,t}^i, \\mu_{z,t}^i)" + }, + { + "bbox": [ + 130, + 247, + 482, + 331 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 130, + 247, + 482, + 331 + ], + "type": "inline_equation", + "content": "\\Sigma_t^i" + }, + { + "bbox": [ + 130, + 247, + 482, + 331 + ], + "type": "text", + "content": ", and " + }, + { + "bbox": [ + 130, + 247, + 482, + 331 + ], + "type": "inline_equation", + "content": "\\pi_t^i" + }, + { + "bbox": [ + 130, + 247, + 482, + 331 + ], + "type": "text", + "content": " to represent the mean vector, covariance matrix, and weight of the " + }, + { + "bbox": [ + 130, + 247, + 482, + 331 + ], + "type": "inline_equation", + "content": "i" + }, + { + "bbox": [ + 130, + 247, + 482, + 331 + ], + "type": "text", + "content": "-th Gaussian kernel at time step " + }, + { + "bbox": [ + 130, + 247, + 482, + 331 + ], + "type": "inline_equation", + "content": "t" + }, + { + "bbox": [ + 130, + 247, + 482, + 331 + ], + "type": "text", + "content": ", respectively. These three parameters can be obtained through three separate MLP layers, which can be defined as:" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 192, + 342, + 481, + 357 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 192, + 342, + 481, + 357 + ], + "spans": [ + { + "bbox": [ + 192, + 342, + 481, + 357 + ], + "type": "interline_equation", + "content": "\\left\\{\\mu_ {t} ^ {i}, \\Sigma_ {t} ^ {i}, \\pi_ {t} ^ {i} \\right\\} = f _ {m d n} \\left(Z _ {t}; \\theta_ {m d n}\\right) \\quad i = 1 \\dots K \\tag {10}", + "image_path": "516fdcb95ce7f507275960902933bf8678e0fdd68c141412882297b8616996a8.jpg" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 130, + 365, + 434, + 378 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 365, + 434, + 378 + ], + "spans": [ + { + "bbox": [ + 130, + 365, + 434, + 378 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 130, + 365, + 434, + 378 + ], + "type": "inline_equation", + "content": "\\theta_{mdn}" + }, + { + "bbox": [ + 130, + 365, + 434, + 378 + ], + "type": "text", + "content": " refers to the parameters of the mixture density network." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 130, + 378, + 481, + 401 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 378, + 481, + 401 + ], + "spans": [ + { + "bbox": [ + 130, + 378, + 481, + 401 + ], + "type": "text", + "content": "Next, the probability density function for a 3D Gaussian distribution " + }, + { + "bbox": [ + 130, + 378, + 481, + 401 + ], + "type": "inline_equation", + "content": "P(p_{t})" + }, + { + "bbox": [ + 130, + 378, + 481, + 401 + ], + "type": "text", + "content": " can be defined as follows:" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 190, + 411, + 481, + 440 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 190, + 411, + 481, + 440 + ], + "spans": [ + { + "bbox": [ + 190, + 411, + 481, + 440 + ], + "type": "interline_equation", + "content": "\\mathcal {N} \\left(p _ {t} \\mid \\mu_ {t} ^ {i}, \\Sigma_ {t} ^ {i}\\right) = \\frac {\\exp \\left(- \\frac {1}{2} \\left(p _ {t} - \\mu_ {t} ^ {i}\\right) ^ {\\top} \\left(\\Sigma_ {t} ^ {i}\\right) ^ {- 1} \\left(p _ {t} - \\mu_ {t} ^ {i}\\right)\\right)}{\\sqrt {(2 \\pi) ^ {3} \\left| \\Sigma_ {t} ^ {i} \\right|}} \\tag {11}", + "image_path": "e3dc864740e89b66c64a3df3934ddc3d661f74b5f27c22822fcc40c814e73ded.jpg" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 227, + 443, + 481, + 475 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 227, + 443, + 481, + 475 + ], + "spans": [ + { + "bbox": [ + 227, + 443, + 481, + 475 + ], + "type": "interline_equation", + "content": "P \\left(p _ {t}\\right) = \\sum_ {i = 1} ^ {K} \\pi_ {t} ^ {i} \\mathcal {N} \\left(p _ {t} \\mid \\mu_ {t} ^ {i}, \\Sigma_ {t} ^ {i}\\right) \\tag {12}", + "image_path": "bbd852bb7dc07aee3abf7242ad2097c6fa99467cdadf14e8fba62c34ad190316.jpg" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 130, + 485, + 481, + 521 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 485, + 481, + 521 + ], + "spans": [ + { + "bbox": [ + 130, + 485, + 481, + 521 + ], + "type": "text", + "content": "where the " + }, + { + "bbox": [ + 130, + 485, + 481, + 521 + ], + "type": "inline_equation", + "content": "|\\Sigma_t^i|" + }, + { + "bbox": [ + 130, + 485, + 481, + 521 + ], + "type": "text", + "content": " means the determinant of the covariance matrix " + }, + { + "bbox": [ + 130, + 485, + 481, + 521 + ], + "type": "inline_equation", + "content": "\\Sigma_t^i" + }, + { + "bbox": [ + 130, + 485, + 481, + 521 + ], + "type": "text", + "content": " and the " + }, + { + "bbox": [ + 130, + 485, + 481, + 521 + ], + "type": "inline_equation", + "content": "\\mathcal{N}(p_t\\mid \\mu_t^i,\\Sigma_t^i)" + }, + { + "bbox": [ + 130, + 485, + 481, + 521 + ], + "type": "text", + "content": " represents the Gaussian distribution function associated with the " + }, + { + "bbox": [ + 130, + 485, + 481, + 521 + ], + "type": "inline_equation", + "content": "i" + }, + { + "bbox": [ + 130, + 485, + 481, + 521 + ], + "type": "text", + "content": "-th 3D Gaussian function." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 130, + 522, + 481, + 547 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 522, + 481, + 547 + ], + "spans": [ + { + "bbox": [ + 130, + 522, + 481, + 547 + ], + "type": "text", + "content": "The prediction fixation " + }, + { + "bbox": [ + 130, + 522, + 481, + 547 + ], + "type": "inline_equation", + "content": "\\hat{p}_t" + }, + { + "bbox": [ + 130, + 522, + 481, + 547 + ], + "type": "text", + "content": " is sampled probabilistically from a 3D surface, which can be expressed as:" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 284, + 559, + 481, + 571 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 284, + 559, + 481, + 571 + ], + "spans": [ + { + "bbox": [ + 284, + 559, + 481, + 571 + ], + "type": "interline_equation", + "content": "\\hat {p} _ {t} \\sim P \\left(p _ {t}\\right) \\tag {13}", + "image_path": "b650aaec583fb1af8999d05c18a1bc2c1c7f9b2e223ec6e325fa94203de52dcf.jpg" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 130, + 582, + 482, + 666 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 582, + 482, + 666 + ], + "spans": [ + { + "bbox": [ + 130, + 582, + 482, + 666 + ], + "type": "text", + "content": "here, it is worth noting that we do not simply sample the fixation with the highest probability, since such sampling strategy always makes the model rapidly converge to the same fixation positions for all images. In our study, we generated a total of 32,768 fixation points along the latitude-longitude grid, with a resolution of " + }, + { + "bbox": [ + 130, + 582, + 482, + 666 + ], + "type": "inline_equation", + "content": "128 \\times 256" + }, + { + "bbox": [ + 130, + 582, + 482, + 666 + ], + "type": "text", + "content": ". We independently computed the probability for each fixation point. Finally, we randomly selected a fixation point based on its probability, with higher probabilities being more likely to be chosen." + } + ] + } + ], + "index": 13 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 202, + 91, + 448, + 103 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 202, + 91, + 448, + 103 + ], + "spans": [ + { + "bbox": [ + 202, + 91, + 448, + 103 + ], + "type": "text", + "content": "Pathformer3D: A 3D Scanpath Transformer for " + }, + { + "bbox": [ + 202, + 91, + 448, + 103 + ], + "type": "inline_equation", + "content": "360^{\\circ}" + }, + { + "bbox": [ + 202, + 91, + 448, + 103 + ], + "type": "text", + "content": " Images" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 474, + 91, + 481, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 474, + 91, + 481, + 100 + ], + "spans": [ + { + "bbox": [ + 474, + 91, + 481, + 100 + ], + "type": "text", + "content": "7" + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 6 + }, + { + "para_blocks": [ + { + "bbox": [ + 132, + 116, + 231, + 127 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 116, + 231, + 127 + ], + "spans": [ + { + "bbox": [ + 132, + 116, + 231, + 127 + ], + "type": "text", + "content": "3.5 Loss Function" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 130, + 135, + 482, + 183 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 135, + 482, + 183 + ], + "spans": [ + { + "bbox": [ + 130, + 135, + 482, + 183 + ], + "type": "text", + "content": "During training, the objective is to minimize the discrepancy between the predicted distribution and the real human scanpath. To achieve this, we employ the negative log-likelihood loss function, which encourages higher probabilities for the ground truth fixations. The loss function can be defined as follows:" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 222, + 192, + 481, + 224 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 222, + 192, + 481, + 224 + ], + "spans": [ + { + "bbox": [ + 222, + 192, + 481, + 224 + ], + "type": "interline_equation", + "content": "\\mathcal {L} = - \\frac {1}{T} \\sum_ {t = 1} ^ {T} \\log \\left(\\sum_ {i = 1} ^ {K} \\pi_ {t} ^ {i} \\mathcal {N} \\left(p _ {t} ^ {*} \\mid \\mu_ {t} ^ {i}, \\Sigma_ {t} ^ {i}\\right)\\right) \\tag {14}", + "image_path": "6ad4bc9bb146073cb6a6033c68e6bb6ac61f87daba9c39a385b009afd6d30552.jpg" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 132, + 232, + 430, + 244 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 232, + 430, + 244 + ], + "spans": [ + { + "bbox": [ + 132, + 232, + 430, + 244 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 132, + 232, + 430, + 244 + ], + "type": "inline_equation", + "content": "p_t^*" + }, + { + "bbox": [ + 132, + 232, + 430, + 244 + ], + "type": "text", + "content": " represents the real human fixation position of time step " + }, + { + "bbox": [ + 132, + 232, + 430, + 244 + ], + "type": "inline_equation", + "content": "t" + }, + { + "bbox": [ + 132, + 232, + 430, + 244 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 132, + 262, + 230, + 276 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 262, + 230, + 276 + ], + "spans": [ + { + "bbox": [ + 132, + 262, + 230, + 276 + ], + "type": "text", + "content": "4 Experiments" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 132, + 286, + 205, + 297 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 286, + 205, + 297 + ], + "spans": [ + { + "bbox": [ + 132, + 286, + 205, + 297 + ], + "type": "text", + "content": "4.1 Datasets" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 130, + 305, + 504, + 474 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 305, + 504, + 474 + ], + "spans": [ + { + "bbox": [ + 130, + 305, + 504, + 474 + ], + "type": "text", + "content": "We conducted experiments on four " + }, + { + "bbox": [ + 130, + 305, + 504, + 474 + ], + "type": "inline_equation", + "content": "360^{\\circ}" + }, + { + "bbox": [ + 130, + 305, + 504, + 474 + ], + "type": "text", + "content": " image datasets: Sitzmann [44], Salient360! [43], AOI [54], and JUFE [19]. The Sitzmann dataset [44] comprises 22 images with a total of 1920 scanpaths. Only the training set of the Salient360! dataset [43] is available, which contains 85 images and 3036 scanpaths. The AOI dataset [54] consists of 600 high-resolution images and approximately 18,000 scanpaths. The JUFE dataset [19] includes 1032 images and 30,960 scanpaths. For training, 19 images from Sitzmann [44] and 60 images from Salient360! [43] were utilized, and all remaining data were used for validation. To increase the number of training sets, we rotated the image along with the corresponding scanpaths six times along the direction of longitude. We followed ScanDMM [45] to preprocess the gaze data of Sitzmann, Salient360!, and JUFE datasets, and obtain visual scanpaths containing 30, 25, and 15 fixations for Sitzmann, Salient360!, and JUFE. The AOI dataset has already been pre-processed, with an average scanpath length of approximately 8." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 132, + 491, + 257, + 502 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 491, + 257, + 502 + ], + "spans": [ + { + "bbox": [ + 132, + 491, + 257, + 502 + ], + "type": "text", + "content": "4.2 Evaluation Metrics" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 130, + 510, + 482, + 667 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 510, + 482, + 667 + ], + "spans": [ + { + "bbox": [ + 130, + 510, + 482, + 667 + ], + "type": "text", + "content": "We employed six metrics to assess the performances of our model and other models, including Levenshtein Distance (LEV) [41], Dynamic Time Warping (DTW) [25], Time Delay Embedding (TDE), ScanMatch [14], Recursive Metric (REC) [3] and Sequence Score(SS) [55]. LEV [41] is a metric that quantifies the dissimilarity between two strings by measuring the minimum number of single-character edit operations needed to transform one string into another. ScanMatch [14] involves converting eye-tracking scanpaths into string sequences and then employing an algorithm similar to the Levenshtein distance to compare the similarity of the two sequences. DTW [25] and TDE [52] are metrics based on time series analysis, while REC [3] is a metric for recursive quantification analysis. SS [55] is calculated by converting scanpaths into strings of fixation cluster IDs, and then measuring the similarity between two strings using a string matching algorithm [37]." + } + ] + } + ], + "index": 10 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "spans": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "type": "text", + "content": "8" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 166, + 91, + 226, + 102 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 166, + 91, + 226, + 102 + ], + "spans": [ + { + "bbox": [ + 166, + 91, + 226, + 102 + ], + "type": "text", + "content": "R. Quan et al." + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 7 + }, + { + "para_blocks": [ + { + "bbox": [ + 130, + 115, + 482, + 189 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 115, + 482, + 189 + ], + "spans": [ + { + "bbox": [ + 130, + 115, + 482, + 189 + ], + "type": "text", + "content": "For each image, every model generates " + }, + { + "bbox": [ + 130, + 115, + 482, + 189 + ], + "type": "inline_equation", + "content": "m = 10" + }, + { + "bbox": [ + 130, + 115, + 482, + 189 + ], + "type": "text", + "content": " predicted scanpaths " + }, + { + "bbox": [ + 130, + 115, + 482, + 189 + ], + "type": "inline_equation", + "content": "\\widehat{X}" + }, + { + "bbox": [ + 130, + 115, + 482, + 189 + ], + "type": "text", + "content": ", which are evaluated against all real human scanpaths " + }, + { + "bbox": [ + 130, + 115, + 482, + 189 + ], + "type": "inline_equation", + "content": "X" + }, + { + "bbox": [ + 130, + 115, + 482, + 189 + ], + "type": "text", + "content": ", and then averaged to determine the final prediction score. Note that our model predicts scanpath with a length of 30. When comparing across different datasets, we truncate the predicted scanpath to the length of the actual scanpath before conducting the comparison. Taking DTW [25] as an example, the score can be calculated as follows:" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 230, + 198, + 481, + 229 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 230, + 198, + 481, + 229 + ], + "spans": [ + { + "bbox": [ + 230, + 198, + 481, + 229 + ], + "type": "interline_equation", + "content": "D T W = \\frac {1}{n} \\frac {1}{m} \\sum_ {i = 1} ^ {n} \\sum_ {j = 1} ^ {m} D T W \\left(\\widehat {X} _ {j}, X _ {i}\\right) \\tag {15}", + "image_path": "6f0ff50ce6a13611f8a7113305647d189767165bab2fbfa3d54bc9718b2dd93b.jpg" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 130, + 236, + 481, + 275 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 236, + 481, + 275 + ], + "spans": [ + { + "bbox": [ + 130, + 236, + 481, + 275 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 130, + 236, + 481, + 275 + ], + "type": "inline_equation", + "content": "\\widehat{X}_j" + }, + { + "bbox": [ + 130, + 236, + 481, + 275 + ], + "type": "text", + "content": " is the " + }, + { + "bbox": [ + 130, + 236, + 481, + 275 + ], + "type": "inline_equation", + "content": "j" + }, + { + "bbox": [ + 130, + 236, + 481, + 275 + ], + "type": "text", + "content": "-th predicted scanpath, " + }, + { + "bbox": [ + 130, + 236, + 481, + 275 + ], + "type": "inline_equation", + "content": "X_i" + }, + { + "bbox": [ + 130, + 236, + 481, + 275 + ], + "type": "text", + "content": " is the " + }, + { + "bbox": [ + 130, + 236, + 481, + 275 + ], + "type": "inline_equation", + "content": "i" + }, + { + "bbox": [ + 130, + 236, + 481, + 275 + ], + "type": "text", + "content": "-th real human scanpath, " + }, + { + "bbox": [ + 130, + 236, + 481, + 275 + ], + "type": "inline_equation", + "content": "n" + }, + { + "bbox": [ + 130, + 236, + 481, + 275 + ], + "type": "text", + "content": " is the number of the real human scanpaths, " + }, + { + "bbox": [ + 130, + 236, + 481, + 275 + ], + "type": "inline_equation", + "content": "m" + }, + { + "bbox": [ + 130, + 236, + 481, + 275 + ], + "type": "text", + "content": " is the number of predicted scanpaths." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 131, + 293, + 279, + 304 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 131, + 293, + 279, + 304 + ], + "spans": [ + { + "bbox": [ + 131, + 293, + 279, + 304 + ], + "type": "text", + "content": "4.3 Implementation Details" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 130, + 313, + 482, + 420 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 313, + 482, + 420 + ], + "spans": [ + { + "bbox": [ + 130, + 313, + 482, + 420 + ], + "type": "text", + "content": "The internal vector dimension of 3D Transformer Encoder is set to " + }, + { + "bbox": [ + 130, + 313, + 482, + 420 + ], + "type": "inline_equation", + "content": "D = 128" + }, + { + "bbox": [ + 130, + 313, + 482, + 420 + ], + "type": "text", + "content": ". We utilize a multi-head attention mechanism with 8 attention heads. The hidden layer size of the feed-forward network (FFN) is set to 64. In the 3D mixture density network model, we configure " + }, + { + "bbox": [ + 130, + 313, + 482, + 420 + ], + "type": "inline_equation", + "content": "K = 5" + }, + { + "bbox": [ + 130, + 313, + 482, + 420 + ], + "type": "text", + "content": " Gaussian kernels, and the internal hidden layer dimension is set to 16. During the training process, our batch size is set to 18. As for the optimizer, we use AdamW [26] with a learning rate of 1e-5. We employ a learning rate warm-up and stage-wise adjustment strategy. The warm-up period lasts for 10 epochs, and the learning rate is halved every 10 epochs. The total training duration is 50 epochs." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 130, + 422, + 481, + 446 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 422, + 481, + 446 + ], + "spans": [ + { + "bbox": [ + 130, + 422, + 481, + 446 + ], + "type": "text", + "content": "Our experiments were conducted on a single NVIDIA GeForce RTX 3090, and the development was conducted using the PyTorch framework." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 131, + 464, + 292, + 477 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 131, + 464, + 292, + 477 + ], + "spans": [ + { + "bbox": [ + 131, + 464, + 292, + 477 + ], + "type": "text", + "content": "4.4 Performance Comparison" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 130, + 485, + 482, + 569 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 485, + 482, + 569 + ], + "spans": [ + { + "bbox": [ + 130, + 485, + 482, + 569 + ], + "type": "text", + "content": "We compared our model with SaliTiNet [6], ScanGAN [35], and ScanDMM [45], utilizing the models provided officially by their respective sources to obtain prediction results. These three models are all designed for scanpath prediction on " + }, + { + "bbox": [ + 130, + 485, + 482, + 569 + ], + "type": "inline_equation", + "content": "360^{\\circ}" + }, + { + "bbox": [ + 130, + 485, + 482, + 569 + ], + "type": "text", + "content": " images. PathGAN [5] and Zhu et al. [58] are two additional models specifically designed for generating scanpath on " + }, + { + "bbox": [ + 130, + 485, + 482, + 569 + ], + "type": "inline_equation", + "content": "360^{\\circ}" + }, + { + "bbox": [ + 130, + 485, + 482, + 569 + ], + "type": "text", + "content": " images. However, it has been demonstrated that these models are unable to effectively generate realistic scanpaths [35]. Therefore, we have not included them in our comparison." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 130, + 570, + 482, + 666 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 570, + 482, + 666 + ], + "spans": [ + { + "bbox": [ + 130, + 570, + 482, + 666 + ], + "type": "text", + "content": "We additionally established two baseline reference values to assess the relative quality of the generated scanpath. The first one is the random baseline (Random walk). This baseline randomly generate ten scanpaths for each image, which provides the reference for lowest performance. The second one is the human baseline (Human), which compares each real human scanpath in the dataset with all the other real human scanpaths and taking the average. This baseline represents the average similarity among real human scanpaths, serving as a reference for highest performance." + } + ] + } + ], + "index": 10 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 202, + 91, + 448, + 103 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 202, + 91, + 448, + 103 + ], + "spans": [ + { + "bbox": [ + 202, + 91, + 448, + 103 + ], + "type": "text", + "content": "Pathformer3D: A 3D Scanpath Transformer for " + }, + { + "bbox": [ + 202, + 91, + 448, + 103 + ], + "type": "inline_equation", + "content": "360^{\\circ}" + }, + { + "bbox": [ + 202, + 91, + 448, + 103 + ], + "type": "text", + "content": " Images" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 474, + 92, + 481, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 474, + 92, + 481, + 100 + ], + "spans": [ + { + "bbox": [ + 474, + 92, + 481, + 100 + ], + "type": "text", + "content": "9" + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 8 + }, + { + "para_blocks": [ + { + "bbox": [ + 130, + 116, + 482, + 175 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 116, + 482, + 175 + ], + "spans": [ + { + "bbox": [ + 130, + 116, + 482, + 175 + ], + "type": "text", + "content": "Quantitative evaluation. Tab. 1 shows the quantitative comparison results between our method and other competitors on the metrics of LEV [41], DTW [25], TDE [52], ScanMatch [14], REC [3] and SS [55]. As we can see, our method outperforms all the competitors on all four datasets, achieving performance close to 'Human'." + } + ] + } + ], + "index": 2 + }, + { + "type": "table", + "bbox": [ + 134, + 227, + 482, + 383 + ], + "blocks": [ + { + "bbox": [ + 130, + 194, + 482, + 218 + ], + "lines": [ + { + "bbox": [ + 130, + 194, + 482, + 218 + ], + "spans": [ + { + "bbox": [ + 130, + 194, + 482, + 218 + ], + "type": "text", + "content": "Table 1: Performance of scanpath prediction models on different datasets. The best result for each metric is highlighted in bold, while the second-best result is underlined." + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 134, + 227, + 482, + 383 + ], + "lines": [ + { + "bbox": [ + 134, + 227, + 482, + 383 + ], + "spans": [ + { + "bbox": [ + 134, + 227, + 482, + 383 + ], + "type": "table", + "html": "
MethodSitzmannSalient360!
LEV ↓DTW ↓TDE ↓ScanMatch ↑REC ↑SS ↑LEV ↓DTW ↓TDE ↓ScanMatch ↑REC ↑SS ↑
Random walk51.5222372.38728.0380.3951.5140.19843.6702062.08029.7530.3751.3810.176
SaltiNet [6]51.1512246.05726.3030.3741.7340.19841.0421860.39725.0010.4292.2500.220
ScanGAN [35]45.8521950.48819.0600.4772.9600.26538.9771752.37119.5270.4603.0410.244
ScanDMM [45]45.7481939.98019.0980.4853.1560.27038.5451706.66120.2380.4673.1250.257
Ours44.6911939.66219.8870.4864.1540.28036.2771515.88319.2610.4844.2500.302
Human42.0061841.53416.1920.5265.1630.36635.0901376.43916.6500.5135.0660.377
MethodAOIJUFE
LEV ↓DTW ↓TDE ↓ScanMatch ↑REC ↑SS ↑LEV ↓DTW ↓TDE ↓ScanMatch ↑REC ↑SS ↑
Random walk15.866666.65238.1640.2935.6500.24126.4391270.89338.1640.3521.8470.191
SaltiNet [6]14.978622.07935.7240.3376.3060.24526.1151264.49232.2610.3641.9720.192
ScanGAN [35]14.473569.54932.6020.3617.3790.25824.2591110.82625.9370.4313.2080.217
ScanDMM [45]14.185569.17032.5580.3717.3280.24923.3721081.46525.7870.4523.5270.221
Ours13.904561.89530.9870.3768.2860.22522.5951054.22322.9550.4624.4490.218
Human14.317501.95230.2480.36713.0080.37718.4451030.97020.3050.5647.8640.330
", + "image_path": "0b1be78a80a0d6ca4d9b10abe5303baaa87ed717968657719a121151258b7beb.jpg" + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "table_body" + } + ], + "index": 4 + }, + { + "bbox": [ + 130, + 424, + 482, + 591 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 424, + 482, + 591 + ], + "spans": [ + { + "bbox": [ + 130, + 424, + 482, + 591 + ], + "type": "text", + "content": "Qualitative evaluation. We visualized the prediction results of all methods on one image of each dataset, as shown in Fig. 2. Similarly, we visualized the ground truth of human scanpath accordingly. The transition from purple to red in the scanpath indicates the beginning and end of the scanpath, respectively. It is noticeable that most ground truth scanpaths are located within the salient regions near the equator. SaltiNet [6], although designed for " + }, + { + "bbox": [ + 130, + 424, + 482, + 591 + ], + "type": "inline_equation", + "content": "360^{\\circ}" + }, + { + "bbox": [ + 130, + 424, + 482, + 591 + ], + "type": "text", + "content": " images, suffers from large displacements. In comparison, the methods proposed by ScanGAN [35] and ScanDMM [45] have the capability to generate relatively realistic scanpaths to some extent. However, our model, in contrast to these approaches, excels at capturing salient information within " + }, + { + "bbox": [ + 130, + 424, + 482, + 591 + ], + "type": "inline_equation", + "content": "360^{\\circ}" + }, + { + "bbox": [ + 130, + 424, + 482, + 591 + ], + "type": "text", + "content": " images and is capable of generating scanpath that are closer to the real human scanpaths. In addition, a notable flaw in the tower image is that our model tends to focus more on the prominent plants at both ends rather than the central tower, which is a common issue shared by other models as well." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 131, + 609, + 267, + 623 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 131, + 609, + 267, + 623 + ], + "spans": [ + { + "bbox": [ + 131, + 609, + 267, + 623 + ], + "type": "text", + "content": "4.5 Saliency Comparison" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 130, + 629, + 481, + 667 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 629, + 481, + 667 + ], + "spans": [ + { + "bbox": [ + 130, + 629, + 481, + 667 + ], + "type": "text", + "content": "A good scanpath prediction model often assigns more attention to salient regions during the prediction process. In our study, we generated 1000 scanpath and performed post-processing on all fixations to generate corresponding saliency" + } + ] + } + ], + "index": 7 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "spans": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "type": "text", + "content": "10" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 166, + 91, + 225, + 102 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 166, + 91, + 225, + 102 + ], + "spans": [ + { + "bbox": [ + 166, + 91, + 225, + 102 + ], + "type": "text", + "content": "R. Quan et al." + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 9 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 134, + 114, + 479, + 258 + ], + "blocks": [ + { + "bbox": [ + 134, + 114, + 479, + 258 + ], + "lines": [ + { + "bbox": [ + 134, + 114, + 479, + 258 + ], + "spans": [ + { + "bbox": [ + 134, + 114, + 479, + 258 + ], + "type": "image", + "image_path": "f4aad0f1d160d12dc5b8a0092434cb1cafde543cb2325ee239228189b10f8503.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 130, + 268, + 482, + 313 + ], + "lines": [ + { + "bbox": [ + 130, + 268, + 482, + 313 + ], + "spans": [ + { + "bbox": [ + 130, + 268, + 482, + 313 + ], + "type": "text", + "content": "Fig. 2: Qualitative representation results of different models on four datasets. From top to bottom, the four images are sourced from Sitzmann [44], Salient360! [43], AOI [54], and JUFE [19], respectively. From left to right, we display the a scanpath sampled from ground truth, our model, ScanDMM [45], ScanGAN [35], and SaltiNet [6]." + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_caption" + } + ], + "index": 2 + }, + { + "bbox": [ + 130, + 338, + 482, + 400 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 338, + 482, + 400 + ], + "spans": [ + { + "bbox": [ + 130, + 338, + 482, + 400 + ], + "type": "text", + "content": "maps [49]. The fixation maps was obtained from all unprocessed fixations. We used sizes " + }, + { + "bbox": [ + 130, + 338, + 482, + 400 + ], + "type": "inline_equation", + "content": "\\mathrm{H} = 128" + }, + { + "bbox": [ + 130, + 338, + 482, + 400 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 130, + 338, + 482, + 400 + ], + "type": "inline_equation", + "content": "\\mathrm{W} = 256" + }, + { + "bbox": [ + 130, + 338, + 482, + 400 + ], + "type": "text", + "content": " for saliency maps (salmap) and fixation maps (fixmap). We compared our results with ScanGAN [35] and ScanDMM [45] on the Salient360! [43] and JUFE [19] datasets, under the saliency metrics AUC_Judd [9], NSS [40], CC [31], SIM [9], and KLD [29]." + } + ] + } + ], + "index": 4 + }, + { + "type": "table", + "bbox": [ + 170, + 465, + 447, + 545 + ], + "blocks": [ + { + "bbox": [ + 130, + 422, + 482, + 456 + ], + "lines": [ + { + "bbox": [ + 130, + 422, + 482, + 456 + ], + "spans": [ + { + "bbox": [ + 130, + 422, + 482, + 456 + ], + "type": "text", + "content": "Table 2: Quantitative comparison results of significance comparison on Salient360! [43] and JUFE [19] datasets. The best result for each metric is highlighted in bold, while the second-best result is underlined." + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 170, + 465, + 447, + 545 + ], + "lines": [ + { + "bbox": [ + 170, + 465, + 447, + 545 + ], + "spans": [ + { + "bbox": [ + 170, + 465, + 447, + 545 + ], + "type": "table", + "html": "
DatasetMethodAUC_Judd↑NSS ↑CC ↑SIM ↑KLD ↓
Salient360! [43]ScanGAN [35]0.7970.9860.2600.17321.210
ScanDMM [45]0.8301.2100.3220.18320.882
Ours0.8381.1950.3520.25918.985
JUFE [19]ScanGAN [35]0.7950.8370.1980.23619.529
ScanDMM [45]0.8661.2030.2750.28618.248
Ours0.8721.0350.2450.32016.339
", + "image_path": "02b0aa9f9062aef40b24c7970fe205b2714663a1be8ea6158b5d45b3578b920b.jpg" + } + ] + } + ], + "index": 6, + "angle": 0, + "type": "table_body" + } + ], + "index": 6 + }, + { + "bbox": [ + 130, + 570, + 482, + 666 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 570, + 482, + 666 + ], + "spans": [ + { + "bbox": [ + 130, + 570, + 482, + 666 + ], + "type": "text", + "content": "Based on Tab. 2, our model has demonstrated favorable performance across the majority of metrics. Our model did not achieve optimal results in a few metrics, one potential explanation is that our model exhibits a stronger emphasis on larger salient objects, which may lead to comparatively less attention towards smaller targets compared to models with more evenly distributed saliency maps. As evident in Fig. 3, ScanDMM [45] and ScanGAN [35] exhibit a characteristic of having more dispersed saliency maps. Furthermore, it can be observed that our model has the ability to focus more on salient regions compared to the other" + } + ] + } + ], + "index": 7 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 203, + 91, + 448, + 102 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 203, + 91, + 448, + 102 + ], + "spans": [ + { + "bbox": [ + 203, + 91, + 448, + 102 + ], + "type": "text", + "content": "Pathformer3D: A 3D Scanpath Transformer for " + }, + { + "bbox": [ + 203, + 91, + 448, + 102 + ], + "type": "inline_equation", + "content": "360^{\\circ}" + }, + { + "bbox": [ + 203, + 91, + 448, + 102 + ], + "type": "text", + "content": " Images" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 470, + 92, + 480, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 470, + 92, + 480, + 100 + ], + "spans": [ + { + "bbox": [ + 470, + 92, + 480, + 100 + ], + "type": "text", + "content": "11" + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 10 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 136, + 116, + 481, + 288 + ], + "blocks": [ + { + "bbox": [ + 136, + 116, + 481, + 288 + ], + "lines": [ + { + "bbox": [ + 136, + 116, + 481, + 288 + ], + "spans": [ + { + "bbox": [ + 136, + 116, + 481, + 288 + ], + "type": "image", + "image_path": "441392e9cb3366db93419d574e62c506a078fcae62b96e2f7d159564535b5718.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 130, + 296, + 482, + 342 + ], + "lines": [ + { + "bbox": [ + 130, + 296, + 482, + 342 + ], + "spans": [ + { + "bbox": [ + 130, + 296, + 482, + 342 + ], + "type": "text", + "content": "Fig. 3: Qualitative comparison results of significance comparison. The first two images is from Salient360! [43] and the next three images are from JUFE [19]. From top to bottom are the real image, the significance maps of the real images, Ours, ScanDMM [45], ScanGAN [35]." + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_caption" + } + ], + "index": 2 + }, + { + "type": "table", + "bbox": [ + 170, + 371, + 447, + 414 + ], + "blocks": [ + { + "bbox": [ + 158, + 350, + 455, + 363 + ], + "lines": [ + { + "bbox": [ + 158, + 350, + 455, + 363 + ], + "spans": [ + { + "bbox": [ + 158, + 350, + 455, + 363 + ], + "type": "text", + "content": "Table 3: Efficiency comparison in terms of model size and running time" + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 170, + 371, + 447, + 414 + ], + "lines": [ + { + "bbox": [ + 170, + 371, + 447, + 414 + ], + "spans": [ + { + "bbox": [ + 170, + 371, + 447, + 414 + ], + "type": "table", + "html": "
SaltiNet [6]ScanGAN [35]ScanDMM [45]Ours
Paramters(MB)98.7315.13417.77919.280
Running Time(ms)127206.482118.250266.420
", + "image_path": "9a2e81fe431336051a04a08e779d44158b7f7180f3ea876194c7132eadfeca37.jpg" + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "table_body" + } + ], + "index": 5 + }, + { + "bbox": [ + 130, + 432, + 482, + 456 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 432, + 482, + 456 + ], + "spans": [ + { + "bbox": [ + 130, + 432, + 482, + 456 + ], + "type": "text", + "content": "models. As a result, the post-processed saliency maps obtained from our model exhibit a closer resemblance to human-generated saliency maps." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 131, + 472, + 274, + 483 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 131, + 472, + 274, + 483 + ], + "spans": [ + { + "bbox": [ + 131, + 472, + 274, + 483 + ], + "type": "text", + "content": "4.6 Efficiency Comparison" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 130, + 489, + 482, + 609 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 489, + 482, + 609 + ], + "spans": [ + { + "bbox": [ + 130, + 489, + 482, + 609 + ], + "type": "text", + "content": "We also compared our model's size and running time for generating a scanpath with other methods. In Tab. 3, we present the model size and the runtime for generating a single scanpath (measured in milliseconds). As can be seen, we have a similar model size with other models. In terms of runtime, our model is capable of generating approximately four scanpaths of length 30 per second, which is a little slower compared to ScanGAN and ScanDMM. This is primarily because our model generates fixations in an autoregressive manner [21]. Additionally, the use of a mixture density network is another factor leading to slower running speed. However, compared to a single Gaussian distribution, multiple Gaussian distributions can more accurately represent the distribution of fixations." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 131, + 624, + 246, + 635 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 131, + 624, + 246, + 635 + ], + "spans": [ + { + "bbox": [ + 131, + 624, + 246, + 635 + ], + "type": "text", + "content": "4.7 Ablation Studies" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 130, + 641, + 487, + 666 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 641, + 487, + 666 + ], + "spans": [ + { + "bbox": [ + 130, + 641, + 487, + 666 + ], + "type": "text", + "content": "We conducted ablation experiments on the Salient360! [43] and JUFE [19] datasets. The experimental results are presented in Tab. 4, showing the optimal results" + } + ] + } + ], + "index": 10 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "spans": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "type": "text", + "content": "12" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 166, + 91, + 225, + 102 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 166, + 91, + 225, + 102 + ], + "spans": [ + { + "bbox": [ + 166, + 91, + 225, + 102 + ], + "type": "text", + "content": "R. Quan et al." + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 11 + }, + { + "para_blocks": [ + { + "bbox": [ + 130, + 116, + 482, + 153 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 116, + 482, + 153 + ], + "spans": [ + { + "bbox": [ + 130, + 116, + 482, + 153 + ], + "type": "text", + "content": "obtained for each configuration. Next, we will provide a brief overview of each model architecture. For more detailed architecture information, please refer to the supplementary file." + } + ] + } + ], + "index": 2 + }, + { + "type": "table", + "bbox": [ + 134, + 206, + 482, + 316 + ], + "blocks": [ + { + "bbox": [ + 132, + 174, + 481, + 198 + ], + "lines": [ + { + "bbox": [ + 132, + 174, + 481, + 198 + ], + "spans": [ + { + "bbox": [ + 132, + 174, + 481, + 198 + ], + "type": "text", + "content": "Table 4: Ablation studies on Salient360! [43] and JUFE [19] datasets. The best result for each metric is highlighted in bold, while the second-best result is underlined." + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 134, + 206, + 482, + 316 + ], + "lines": [ + { + "bbox": [ + 134, + 206, + 482, + 316 + ], + "spans": [ + { + "bbox": [ + 134, + 206, + 482, + 316 + ], + "type": "table", + "html": "
MethodSalient360!JUFE
LEV ↓DTW ↓TDE ↓ScanMatch ↑REC ↑SS ↑LEV ↓DTW ↓TDE ↓ScanMatch ↑REC ↑SS ↑
Pure ViT37.6431739.71419.2910.4703.7100.26722.8911072.75323.7510.4573.8470.221
Pure 2D CNN36.8481727.24119.6410.4774.1190.27222.6131060.20523.0960.4634.1810.220
Saliency38.1961864.91420.0590.4753.1560.25323.8331094.29024.6360.4423.6690.208
w/o EncoderLayer37.3521798.34020.7270.4634.1710.26222.9261072.26323.1960.4554.2600.218
w/o MDN + MSE Loss39.4631793.94825.9140.4263.7810.09523.9151132.12725.6750.4163.8330.102
K=138.6431804.07420.0740.4623.3260.24823.2391084.73423.5770.4513.8050.218
K=336.2502073.90522.3260.4904.0390.24622.7581077.36523.3020.4604.3600.214
K=838.1171700.87720.4550.4673.3890.25822.9171065.24023.4700.4553.9140.210
Parallel36.7221577.99323.7260.4893.9390.28822.5671054.64428.7750.4513.6880.221
Ours36.2771515.88319.2610.4844.2530.30222.5951054.22322.9550.4624.4490.218
", + "image_path": "b6086fcad6ce57e1922b4d1baf67a6f05ec0d7bb3e07ae7d195f597dfca98d1d.jpg" + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "table_body" + } + ], + "index": 4 + }, + { + "bbox": [ + 130, + 360, + 482, + 481 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 360, + 482, + 481 + ], + "spans": [ + { + "bbox": [ + 130, + 360, + 482, + 481 + ], + "type": "text", + "content": "The feature extractor. To demonstrate the effectiveness of our 3D feature extractor, SphereNet, we try to employ three distinct architectures for feature extraction. Firstly, we extracted features using the patch-based approach of the Vision Transformer (ViT) method [16] (denoted as 'Pure ViT'). Secondly, we replaced SphereNet with a standard 2D convolutional network (denoted as 'Pure 2D CNN'). Lastly, We followed VSPT [42] to extract image features based on the saliency information. Specifically, we employed EPSNet [59] to extract saliency information (denoted as 'Saliency'). As shown in Tab. 4, 'Ours' obtains better performance than 'Pure ViT', 'Pure 2D CNN', and 'Saliency', which demonstrates the superiority of our feature extractor." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 130, + 501, + 482, + 586 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 501, + 482, + 586 + ], + "spans": [ + { + "bbox": [ + 130, + 501, + 482, + 586 + ], + "type": "text", + "content": "The 3D Transformer encoder. To validate the effectiveness of our 3D Transformer Encoder, we attempted to remove the EncoderLayer in the 3D Transformer Encoder and directly utilize the feature embedding of the Transformer Encoder as input of the Transformer Decoder (denoted as 'w/o EncoderLayer'). As we can see from Tab. 4, without our 3D Transformer Encoder, the scanpath prediction performance dropped significantly, which demonstrates the effectiveness of our 3D Transformer Encoder." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 130, + 605, + 482, + 666 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 605, + 482, + 666 + ], + "spans": [ + { + "bbox": [ + 130, + 605, + 482, + 666 + ], + "type": "text", + "content": "3D MDN. To validate the effectiveness of the 3D MDN, we directly exploited linear regression to predict the fixations after the Transformer Decoder. The Mean Squared Error (MSE) was utilized as the loss function (denoted as 'w/o MDN + MSE Loss'). As can be seen in Tab. 4, 'Ours' outperforms 'w/o MDN + MSE Loss' a lot, proving the advantage of our 3D MDN." + } + ] + } + ], + "index": 7 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 202, + 90, + 447, + 103 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 202, + 90, + 447, + 103 + ], + "spans": [ + { + "bbox": [ + 202, + 90, + 447, + 103 + ], + "type": "text", + "content": "Pathformer3D: A 3D Scanpath Transformer for " + }, + { + "bbox": [ + 202, + 90, + 447, + 103 + ], + "type": "inline_equation", + "content": "360^{\\circ}" + }, + { + "bbox": [ + 202, + 90, + 447, + 103 + ], + "type": "text", + "content": " Images" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 470, + 91, + 481, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 470, + 91, + 481, + 100 + ], + "spans": [ + { + "bbox": [ + 470, + 91, + 481, + 100 + ], + "type": "text", + "content": "13" + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 12 + }, + { + "para_blocks": [ + { + "bbox": [ + 130, + 116, + 482, + 189 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 116, + 482, + 189 + ], + "spans": [ + { + "bbox": [ + 130, + 116, + 482, + 189 + ], + "type": "text", + "content": "Different number of Gaussian kernels. We compared with the model results obtained by using different numbers of Gaussian kernels in the mixture density network. These comparisons are respectively presented in the rows labeled \"K=1,3,8,\" while our model has 5 Gaussian kernels. As we can see from Tab. 4, using 5 Gaussian kernels can provide a better representation of the fixations' distribution." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 130, + 203, + 483, + 310 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 203, + 483, + 310 + ], + "spans": [ + { + "bbox": [ + 130, + 203, + 483, + 310 + ], + "type": "text", + "content": "Parallel in the Transformer decoder. In our Transformer decoder, we use autoregressive method to predict one fixation each time and iteratively generate the whole fixation sequence. Through our observation, more and more models try to adopt a parallel structure in the Transformer decoder [10, 23, 36] for higher efficiency. Therefore, we also tried to build a parallel Transformer decoder in our method (denoted as 'Parallel'). In the parallel model, we use a random embedding with length " + }, + { + "bbox": [ + 130, + 203, + 483, + 310 + ], + "type": "inline_equation", + "content": "L" + }, + { + "bbox": [ + 130, + 203, + 483, + 310 + ], + "type": "text", + "content": " as the input to the Transformer decoder, and we can get the hidden state of a scanpath of length " + }, + { + "bbox": [ + 130, + 203, + 483, + 310 + ], + "type": "inline_equation", + "content": "L" + }, + { + "bbox": [ + 130, + 203, + 483, + 310 + ], + "type": "text", + "content": " at one time, all the other parts of the model are unchanged." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 130, + 312, + 483, + 361 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 312, + 483, + 361 + ], + "spans": [ + { + "bbox": [ + 130, + 312, + 483, + 361 + ], + "type": "text", + "content": "We can observe from Tab. 4 that the parallel structure of the decoder performs poorly than the autoregressive Transformer decoder, especially in the TDE [52] metric. Such comparison results demonstrate that an autoregressive Transformer decoder is more suitable for our scanpath prediction task." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 131, + 376, + 330, + 389 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 131, + 376, + 330, + 389 + ], + "spans": [ + { + "bbox": [ + 131, + 376, + 330, + 389 + ], + "type": "text", + "content": "5 Limitaions and Future Works" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 130, + 399, + 482, + 446 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 399, + 482, + 446 + ], + "spans": [ + { + "bbox": [ + 130, + 399, + 482, + 446 + ], + "type": "text", + "content": "In our work, we employed SphereNet [12] for feature extraction and 3D Mixture Density Network (MDN) for fixation generation. However, we acknowledge that these components are replaceable, and we are contemplating whether there might be better alternatives for both parts." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 130, + 447, + 482, + 483 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 447, + 482, + 483 + ], + "spans": [ + { + "bbox": [ + 130, + 447, + 482, + 483 + ], + "type": "text", + "content": "In addition, we are considering incorporating relevant semantic information to improve the performance further, as objects with important semantic information tend to attract human attention more easily." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 131, + 498, + 222, + 512 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 131, + 498, + 222, + 512 + ], + "spans": [ + { + "bbox": [ + 131, + 498, + 222, + 512 + ], + "type": "text", + "content": "6 Conclusion" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 130, + 521, + 483, + 666 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 521, + 483, + 666 + ], + "spans": [ + { + "bbox": [ + 130, + 521, + 483, + 666 + ], + "type": "text", + "content": "Considering that the 2D equirectangular projection of " + }, + { + "bbox": [ + 130, + 521, + 483, + 666 + ], + "type": "inline_equation", + "content": "360^{\\circ}" + }, + { + "bbox": [ + 130, + 521, + 483, + 666 + ], + "type": "text", + "content": " images has the issues of distortion and coordinate discontinuity, we proposed to execute scanpath prediction for " + }, + { + "bbox": [ + 130, + 521, + 483, + 666 + ], + "type": "inline_equation", + "content": "360^{\\circ}" + }, + { + "bbox": [ + 130, + 521, + 483, + 666 + ], + "type": "text", + "content": " images in 3D spherical coordinate system. Correspondingly, we proposed a novel 3D Scanpath Transformer named Pathformer3D, where a 3D Transformer Encoder was first used to obtain 3D contextual feature representation of the " + }, + { + "bbox": [ + 130, + 521, + 483, + 666 + ], + "type": "inline_equation", + "content": "360^{\\circ}" + }, + { + "bbox": [ + 130, + 521, + 483, + 666 + ], + "type": "text", + "content": " image. Then, Pathformer3D input the 3D contextual feature representation and historical fixation information into the Transformer Decoder to predict the hidden state for each fixation. Each fixation's hidden state was then fed into a 3D mixture density network to output the 3D Gaussian distribution of the fixation, from which the fixation location can be sampled. Finally, Pathformer3D model was compared with three state-of-the-art " + }, + { + "bbox": [ + 130, + 521, + 483, + 666 + ], + "type": "inline_equation", + "content": "360^{\\circ}" + }, + { + "bbox": [ + 130, + 521, + 483, + 666 + ], + "type": "text", + "content": " image scan-path prediction models, to demonstrate the superiority of its performance." + } + ] + } + ], + "index": 9 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "spans": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "type": "text", + "content": "14" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 166, + 91, + 225, + 102 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 166, + 91, + 225, + 102 + ], + "spans": [ + { + "bbox": [ + 166, + 91, + 225, + 102 + ], + "type": "text", + "content": "R. Quan et al." + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 13 + }, + { + "para_blocks": [ + { + "bbox": [ + 130, + 116, + 482, + 189 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 116, + 482, + 189 + ], + "spans": [ + { + "bbox": [ + 130, + 116, + 482, + 189 + ], + "type": "text", + "content": "Acknowledgements This work is supported in part by the National Natural Science Foundation of China under grant 62206127 and 62272229, the Natural Science Foundation of Jiangsu Province under grant BK20222012, and Shenzhen Science and Technology Program JCYJ20230807142001004. Rong Quan and Yantao Lai contributed equally to this work. Dong Liang is the corresponding author." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 133, + 208, + 198, + 220 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 208, + 198, + 220 + ], + "spans": [ + { + "bbox": [ + 133, + 208, + 198, + 220 + ], + "type": "text", + "content": "References" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 138, + 233, + 481, + 665 + ], + "type": "list", + "angle": 0, + "index": 19, + "blocks": [ + { + "bbox": [ + 138, + 233, + 481, + 267 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 233, + 481, + 267 + ], + "spans": [ + { + "bbox": [ + 138, + 233, + 481, + 267 + ], + "type": "text", + "content": "1. Adeli, H., Vitu, F., Zelinsky, G.J.: A model of the superior colliculus predicts fixation locations during scene viewing and visual search. Journal of Neuroscience 37(6), 1453-1467 (2017)" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 138, + 268, + 481, + 300 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 268, + 481, + 300 + ], + "spans": [ + { + "bbox": [ + 138, + 268, + 481, + 300 + ], + "type": "text", + "content": "2. Aggarwal, A., Mittal, M., Battineni, G.: Generative adversarial network: An overview of theory and applications. International Journal of Information Management Data Insights 1(1), 100004 (2021)" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 138, + 301, + 481, + 333 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 301, + 481, + 333 + ], + "spans": [ + { + "bbox": [ + 138, + 301, + 481, + 333 + ], + "type": "text", + "content": "3. Anderson, N.C., Bischof, W.F., Laidlaw, K.E., Risko, E.F., Kingstone, A.: Recurrence quantification analysis of eye movements. Behavior research methods 45, 842-856 (2013)" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 138, + 334, + 481, + 355 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 334, + 481, + 355 + ], + "spans": [ + { + "bbox": [ + 138, + 334, + 481, + 355 + ], + "type": "text", + "content": "4. Antol, S., Agrawal, A., Lu, J., Mitchell, M., Batra, D., Zitnick, C.L., Parikh, D.: Vqa: Visual question answering. In: ICCV. pp. 2425-2433 (2015)" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 138, + 356, + 481, + 388 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 356, + 481, + 388 + ], + "spans": [ + { + "bbox": [ + 138, + 356, + 481, + 388 + ], + "type": "text", + "content": "5. Assens, M., Giro-i Nieto, X., McGuinness, K., O'Connor, N.E.: Pathgan: Visual scanpath prediction with generative adversarial networks. In: Proceedings of the European Conference on Computer Vision (ECCV) Workshops. pp. 0-0 (2018)" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 138, + 389, + 481, + 433 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 389, + 481, + 433 + ], + "spans": [ + { + "bbox": [ + 138, + 389, + 481, + 433 + ], + "type": "text", + "content": "6. Assens Reina, M., Giro-i Nieto, X., McGuinness, K., O'Connor, N.E.: Saltinet: Scan-path prediction on 360 degree images using saliency volumes. In: Proceedings of the IEEE International Conference on Computer Vision Workshops. pp. 2331-2338 (2017)" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 138, + 434, + 481, + 455 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 434, + 481, + 455 + ], + "spans": [ + { + "bbox": [ + 138, + 434, + 481, + 455 + ], + "type": "text", + "content": "7. Boccignone, G., Ferraro, M.: Modelling gaze shift as a constrained random walk. Physica A: Statistical Mechanics and its Applications 331(1-2), 207-218 (2004)" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 138, + 456, + 481, + 476 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 456, + 481, + 476 + ], + "spans": [ + { + "bbox": [ + 138, + 456, + 481, + 476 + ], + "type": "text", + "content": "8. Brockmann, D., Geisel, T.: The ecology of gaze shifts. Neurocomputing 32, 643-650 (2000)" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 138, + 477, + 481, + 499 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 477, + 481, + 499 + ], + "spans": [ + { + "bbox": [ + 138, + 477, + 481, + 499 + ], + "type": "text", + "content": "9. Bylinskii, Z., Judd, T., Borji, A., Itti, L., Durand, F., Oliva, A., Torralba, A.: Mit saliency benchmark (2015)" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 138, + 500, + 481, + 533 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 500, + 481, + 533 + ], + "spans": [ + { + "bbox": [ + 138, + 500, + 481, + 533 + ], + "type": "text", + "content": "0. Carion, N., Massa, F., Synnaeve, G., Usunier, N., Kirillov, A., Zagoruyko, S.: End-to-end object detection with transformers. In: European conference on computer vision. pp. 213-229. Springer (2020)" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 138, + 533, + 481, + 565 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 533, + 481, + 565 + ], + "spans": [ + { + "bbox": [ + 138, + 533, + 481, + 565 + ], + "type": "text", + "content": "1. Clarke, A.D., Stainer, M.J., Tatler, B.W., Hunt, A.R.: The saccadic flow baseline: Accounting for image-independent biases in fixation behavior. Journal of vision 17(11), 12-12 (2017)" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 138, + 566, + 481, + 599 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 566, + 481, + 599 + ], + "spans": [ + { + "bbox": [ + 138, + 566, + 481, + 599 + ], + "type": "text", + "content": "2. Coors, B., Condurache, A.P., Geiger, A.: Spherenet: Learning spherical representations for detection and classification in omnidirectional images. In: Proceedings of the European conference on computer vision (ECCV). pp. 518-533 (2018)" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 138, + 600, + 481, + 620 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 600, + 481, + 620 + ], + "spans": [ + { + "bbox": [ + 138, + 600, + 481, + 620 + ], + "type": "text", + "content": "3. Coutrot, A., Hsiao, J.H., Chan, A.B.: Scanpath modeling and classification with hidden markov models. Behavior research methods 50(1), 362-379 (2018)" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 138, + 621, + 481, + 643 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 621, + 481, + 643 + ], + "spans": [ + { + "bbox": [ + 138, + 621, + 481, + 643 + ], + "type": "text", + "content": "4. Cristino, F., Mathôt, S., Theeuwes, J., Gilchrist, I.D.: Scanmatch: A novel method for comparing fixation sequences. Behavior research methods 42, 692-700 (2010)" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 138, + 643, + 481, + 665 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 643, + 481, + 665 + ], + "spans": [ + { + "bbox": [ + 138, + 643, + 481, + 665 + ], + "type": "text", + "content": "5. Cuturi, M., Blondel, M.: Soft-dtw: a differentiable loss function for time-series. In: International conference on machine learning. pp. 894-903. PMLR (2017)" + } + ] + } + ], + "index": 18 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 202, + 91, + 448, + 102 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 202, + 91, + 448, + 102 + ], + "spans": [ + { + "bbox": [ + 202, + 91, + 448, + 102 + ], + "type": "text", + "content": "Pathformer3D: A 3D Scanpath Transformer for " + }, + { + "bbox": [ + 202, + 91, + 448, + 102 + ], + "type": "inline_equation", + "content": "360^{\\circ}" + }, + { + "bbox": [ + 202, + 91, + 448, + 102 + ], + "type": "text", + "content": " Images" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 470, + 92, + 481, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 470, + 92, + 481, + 100 + ], + "spans": [ + { + "bbox": [ + 470, + 92, + 481, + 100 + ], + "type": "text", + "content": "15" + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 14 + }, + { + "para_blocks": [ + { + "bbox": [ + 132, + 116, + 481, + 666 + ], + "type": "list", + "angle": 0, + "index": 20, + "blocks": [ + { + "bbox": [ + 133, + 116, + 481, + 161 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 116, + 481, + 161 + ], + "spans": [ + { + "bbox": [ + 133, + 116, + 481, + 161 + ], + "type": "text", + "content": "16. Dosovitskiy, A., Beyer, L., Kolesnikov, A., Weissenborn, D., Zhai, X., Unterthiner, T., Dehghani, M., Minderer, M., Heigold, G., Gelly, S., et al.: An image is worth 16x16 words: Transformers for image recognition at scale. arXiv preprint arXiv:2010.11929 (2020)" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 132, + 162, + 481, + 183 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 162, + 481, + 183 + ], + "spans": [ + { + "bbox": [ + 132, + 162, + 481, + 183 + ], + "type": "text", + "content": "17. Engbert, R., Trukenbrod, H.A., Barthelme, S., Wichmann, F.A.: Spatial statistics and attentional dynamics in scene viewing. Journal of vision 15(1), 14-14 (2015)" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 132, + 184, + 481, + 228 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 184, + 481, + 228 + ], + "spans": [ + { + "bbox": [ + 132, + 184, + 481, + 228 + ], + "type": "text", + "content": "18. Fan, C.L., Lee, J., Lo, W.C., Huang, C.Y., Chen, K.T., Hsu, C.H.: Fixation prediction for 360 video streaming in head-mounted virtual reality. In: Proceedings of the 27th Workshop on Network and Operating Systems Support for Digital Audio and Video. pp. 67-72 (2017)" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 132, + 228, + 481, + 261 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 228, + 481, + 261 + ], + "spans": [ + { + "bbox": [ + 132, + 228, + 481, + 261 + ], + "type": "text", + "content": "19. Fang, Y., Huang, L., Yan, J., Liu, X., Liu, Y.: Perceptual quality assessment of omnidirectional images. In: Proceedings of the AAAI Conference on Artificial Intelligence. vol. 36, pp. 580-588 (2022)" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 132, + 262, + 481, + 295 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 262, + 481, + 295 + ], + "spans": [ + { + "bbox": [ + 132, + 262, + 481, + 295 + ], + "type": "text", + "content": "20. Goodfellow, I., Pouget-Abadie, J., Mirza, M., Xu, B., Warde-Farley, D., Ozair, S., Courville, A., Bengio, Y.: Generative adversarial nets. Advances in neural information processing systems 27 (2014)" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 132, + 296, + 481, + 317 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 296, + 481, + 317 + ], + "spans": [ + { + "bbox": [ + 132, + 296, + 481, + 317 + ], + "type": "text", + "content": "21. Graves, A.: Generating sequences with recurrent neural networks. arXiv preprint arXiv:1308.0850 (2013)" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 132, + 318, + 481, + 350 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 318, + 481, + 350 + ], + "spans": [ + { + "bbox": [ + 132, + 318, + 481, + 350 + ], + "type": "text", + "content": "22. Itti, L., Koch, C., Niebur, E.: A model of saliency-based visual attention for rapid scene analysis. IEEE Transactions on pattern analysis and machine intelligence 20(11), 1254-1259 (1998)" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 132, + 351, + 481, + 396 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 351, + 481, + 396 + ], + "spans": [ + { + "bbox": [ + 132, + 351, + 481, + 396 + ], + "type": "text", + "content": "23. Kamath, A., Singh, M., LeCun, Y., Synnaeve, G., Misra, I., Carion, N.: Mdetr-modulated detection for end-to-end multi-modal understanding. In: Proceedings of the IEEE/CVF International Conference on Computer Vision. pp. 1780-1790 (2021)" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 132, + 396, + 481, + 418 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 396, + 481, + 418 + ], + "spans": [ + { + "bbox": [ + 132, + 396, + 481, + 418 + ], + "type": "text", + "content": "24. Kay, A.: User interface: A personal view. The art of human-computer interface design pp. 191-207 (1990)" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 132, + 419, + 481, + 440 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 419, + 481, + 440 + ], + "spans": [ + { + "bbox": [ + 132, + 419, + 481, + 440 + ], + "type": "text", + "content": "25. Keogh, E., Ratanamahatana, C.A.: Exact indexing of dynamic time warping. Knowledge and information systems 7, 358-386 (2005)" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 132, + 441, + 481, + 474 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 441, + 481, + 474 + ], + "spans": [ + { + "bbox": [ + 132, + 441, + 481, + 474 + ], + "type": "text", + "content": "26. Kinga, D., Adam, J.B., et al.: A method for stochastic optimization. In: International conference on learning representations (ICLR). vol. 5, p. 6. San Diego, California; (2015)" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 132, + 475, + 481, + 507 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 475, + 481, + 507 + ], + "spans": [ + { + "bbox": [ + 132, + 475, + 481, + 507 + ], + "type": "text", + "content": "27. Krishnan, R., Shalit, U., Sontag, D.: Structured inference networks for nonlinear state space models. In: Proceedings of the AAAI Conference on Artificial Intelligence. vol. 31 (2017)" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 132, + 508, + 481, + 530 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 508, + 481, + 530 + ], + "spans": [ + { + "bbox": [ + 132, + 508, + 481, + 530 + ], + "type": "text", + "content": "28. Kümmerer, M., Bethge, M., Wallis, T.S.: Deepgaze iii: Modeling free-viewing human scanpaths with deep learning. Journal of Vision 22(5), 7-7 (2022)" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 132, + 531, + 481, + 563 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 531, + 481, + 563 + ], + "spans": [ + { + "bbox": [ + 132, + 531, + 481, + 563 + ], + "type": "text", + "content": "29. Kümmerer, M., Wallis, T.S., Bethge, M.: Information-theoretic model comparison unifies saliency metrics. Proceedings of the National Academy of Sciences 112(52), 16054-16059 (2015)" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 132, + 564, + 481, + 586 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 564, + 481, + 586 + ], + "spans": [ + { + "bbox": [ + 132, + 564, + 481, + 586 + ], + "type": "text", + "content": "30. Le Meur, O., Coutrot, A.: Introducing context-dependent and spatially-variant viewing biases in saccadic models. Vision research 121, 72-84 (2016)" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 132, + 586, + 481, + 609 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 586, + 481, + 609 + ], + "spans": [ + { + "bbox": [ + 132, + 586, + 481, + 609 + ], + "type": "text", + "content": "31. Le Meur, O., Le Callet, P., Barba, D.: Predicting visual fixations on video based on low-level visual features. Vision research 47(19), 2483-2498 (2007)" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 132, + 610, + 481, + 632 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 610, + 481, + 632 + ], + "spans": [ + { + "bbox": [ + 132, + 610, + 481, + 632 + ], + "type": "text", + "content": "32. Le Meur, O., Liu, Z.: Saccadic model of eye movements for free-viewing condition. Vision research 116, 152-164 (2015)" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 132, + 632, + 481, + 666 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 632, + 481, + 666 + ], + "spans": [ + { + "bbox": [ + 132, + 632, + 481, + 666 + ], + "type": "text", + "content": "33. Li, C., Zhang, W., Liu, Y., Wang, Y.: Very long term field of view prediction for 360-degree video streaming. In: 2019 IEEE conference on multimedia information processing and retrieval (MIPR). pp. 297-302. IEEE (2019)" + } + ] + } + ], + "index": 19 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "spans": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "type": "text", + "content": "16" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 166, + 91, + 225, + 102 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 166, + 91, + 225, + 102 + ], + "spans": [ + { + "bbox": [ + 166, + 91, + 225, + 102 + ], + "type": "text", + "content": "R. Quan et al." + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 15 + }, + { + "para_blocks": [ + { + "bbox": [ + 130, + 116, + 482, + 666 + ], + "type": "list", + "angle": 0, + "index": 19, + "blocks": [ + { + "bbox": [ + 130, + 116, + 482, + 149 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 116, + 482, + 149 + ], + "spans": [ + { + "bbox": [ + 130, + 116, + 482, + 149 + ], + "type": "text", + "content": "34. Liu, H., Xu, D., Huang, Q., Li, W., Xu, M., Lin, S.: Semantically-based human scanpath estimation with hmms. In: Proceedings of the IEEE International Conference on Computer Vision. pp. 3232-3239 (2013)" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 130, + 150, + 482, + 183 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 150, + 482, + 183 + ], + "spans": [ + { + "bbox": [ + 130, + 150, + 482, + 183 + ], + "type": "text", + "content": "35. Martin, D., Serrano, A., Bergman, A.W., Wetzstein, G., Masia, B.: Scangan360: A generative model of realistic scanpaths for 360 images. IEEE Transactions on Visualization and Computer Graphics 28(5), 2003-2013 (2022)" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 130, + 183, + 482, + 226 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 183, + 482, + 226 + ], + "spans": [ + { + "bbox": [ + 130, + 183, + 482, + 226 + ], + "type": "text", + "content": "36. Mondal, S., Yang, Z., Ahn, S., Samaras, D., Zelinsky, G., Hoai, M.: Gazeformer: Scalable, effective and fast prediction of goal-directed human attention. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 1441-1450 (2023)" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 130, + 226, + 482, + 259 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 226, + 482, + 259 + ], + "spans": [ + { + "bbox": [ + 130, + 226, + 482, + 259 + ], + "type": "text", + "content": "37. Needleman, S.B., Wunsch, C.D.: A general method applicable to the search for similarities in the amino acid sequence of two proteins. Journal of molecular biology 48(3), 443-453 (1970)" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 130, + 259, + 482, + 303 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 259, + 482, + 303 + ], + "spans": [ + { + "bbox": [ + 130, + 259, + 482, + 303 + ], + "type": "text", + "content": "38. Nguyen, A., Yan, Z., Nahrstedt, K.: Your attention is unique: Detecting 360-degree video saliency in head-mounted display for head movement prediction. In: Proceedings of the 26th ACM international conference on Multimedia. pp. 1190-1198 (2018)" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 130, + 303, + 482, + 336 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 303, + 482, + 336 + ], + "spans": [ + { + "bbox": [ + 130, + 303, + 482, + 336 + ], + "type": "text", + "content": "39. Pan, J., Ferrer, C.C., McGuinness, K., O'Connor, N.E., Torres, J., Sayrol, E., Giro-i Nieto, X.: Salgan: Visual saliency prediction with generative adversarial networks. arXiv preprint arXiv:1701.01081 (2017)" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 130, + 336, + 482, + 357 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 336, + 482, + 357 + ], + "spans": [ + { + "bbox": [ + 130, + 336, + 482, + 357 + ], + "type": "text", + "content": "40. Peters, R.J., Iyer, A., Itti, L., Koch, C.: Components of bottom-up gaze allocation in natural images. Vision research 45(18), 2397-2416 (2005)" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 130, + 357, + 482, + 390 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 357, + 482, + 390 + ], + "spans": [ + { + "bbox": [ + 130, + 357, + 482, + 390 + ], + "type": "text", + "content": "41. Privitera, C.M., Stark, L.W.: Algorithms for defining visual regions-of-interest: Comparison with eye fixations. IEEE Transactions on pattern analysis and machine intelligence 22(9), 970-982 (2000)" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 130, + 390, + 482, + 424 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 390, + 482, + 424 + ], + "spans": [ + { + "bbox": [ + 130, + 390, + 482, + 424 + ], + "type": "text", + "content": "42. Qiu, M., Rong, Q., Liang, D., Tu, H.: Visual scanpath transformer: Guiding computers to see the world. In: 2023 IEEE International Symposium on Mixed and Augmented Reality (ISMAR). pp. 223-232. IEEE (2023)" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 130, + 424, + 482, + 456 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 424, + 482, + 456 + ], + "spans": [ + { + "bbox": [ + 130, + 424, + 482, + 456 + ], + "type": "text", + "content": "43. Rai, Y., Gutierrez, J., Le Callet, P.: A dataset of head and eye movements for 360 degree images. In: Proceedings of the 8th ACM on Multimedia Systems Conference. pp. 205-210 (2017)" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 130, + 456, + 482, + 490 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 456, + 482, + 490 + ], + "spans": [ + { + "bbox": [ + 130, + 456, + 482, + 490 + ], + "type": "text", + "content": "44. Sitzmann, V., Serrano, A., Pavel, A., Agrawala, M., Gutierrez, D., Masia, B., Wetzstein, G.: Saliency in vr: How do people explore virtual environments? IEEE transactions on visualization and computer graphics 24(4), 1633-1642 (2018)" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 130, + 490, + 482, + 522 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 490, + 482, + 522 + ], + "spans": [ + { + "bbox": [ + 130, + 490, + 482, + 522 + ], + "type": "text", + "content": "45. Sui, X., Fang, Y., Zhu, H., Wang, S., Wang, Z.: Scandmm: A deep markov model of scanpath prediction for " + }, + { + "bbox": [ + 130, + 490, + 482, + 522 + ], + "type": "inline_equation", + "content": "360^{\\circ}" + }, + { + "bbox": [ + 130, + 490, + 482, + 522 + ], + "type": "text", + "content": " images. In: IEEE Conference on Computer Vision and Pattern Recognition (2023)" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 130, + 522, + 482, + 555 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 522, + 482, + 555 + ], + "spans": [ + { + "bbox": [ + 130, + 522, + 482, + 555 + ], + "type": "text", + "content": "46. Sun, W., Chen, Z., Wu, F.: Visual scanpath prediction using ior-roi recurrent mixture density network. IEEE transactions on pattern analysis and machine intelligence 43(6), 2101-2118 (2019)" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 130, + 555, + 482, + 588 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 555, + 482, + 588 + ], + "spans": [ + { + "bbox": [ + 130, + 555, + 482, + 588 + ], + "type": "text", + "content": "47. Sun, X., Yao, H., Ji, R., Liu, X.M.: Toward statistical modeling of saccadic eye movement and visual saliency. IEEE Transactions on Image Processing 23(11), 4649-4662 (2014)" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 130, + 588, + 482, + 610 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 588, + 482, + 610 + ], + "spans": [ + { + "bbox": [ + 130, + 588, + 482, + 610 + ], + "type": "text", + "content": "48. Tatler, B.W., Brockmole, J.R., Carpenter, R.H.: Latest: A model of saccadic decisions in space and time. Psychological review 124(3), 267 (2017)" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 130, + 610, + 482, + 643 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 610, + 482, + 643 + ], + "spans": [ + { + "bbox": [ + 130, + 610, + 482, + 643 + ], + "type": "text", + "content": "49. Upenik, E., Ebrahimi, T.: A simple method to obtain visual attention data in head mounted virtual reality. In: 2017 IEEE International Conference on Multimedia & Expo Workshops (ICMEW). pp. 73-78. IEEE (2017)" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 130, + 643, + 482, + 666 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 643, + 482, + 666 + ], + "spans": [ + { + "bbox": [ + 130, + 643, + 482, + 666 + ], + "type": "text", + "content": "50. Verma, J.K., Paul, S.: Advances in Augmented Reality and Virtual Reality. Springer (2022)" + } + ] + } + ], + "index": 18 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 202, + 91, + 448, + 102 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 202, + 91, + 448, + 102 + ], + "spans": [ + { + "bbox": [ + 202, + 91, + 448, + 102 + ], + "type": "text", + "content": "Pathformer3D: A 3D Scanpath Transformer for " + }, + { + "bbox": [ + 202, + 91, + 448, + 102 + ], + "type": "inline_equation", + "content": "360^{\\circ}" + }, + { + "bbox": [ + 202, + 91, + 448, + 102 + ], + "type": "text", + "content": " Images" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 470, + 92, + 481, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 470, + 92, + 481, + 100 + ], + "spans": [ + { + "bbox": [ + 470, + 92, + 481, + 100 + ], + "type": "text", + "content": "17" + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 16 + }, + { + "para_blocks": [ + { + "bbox": [ + 132, + 116, + 482, + 380 + ], + "type": "list", + "angle": 0, + "index": 11, + "blocks": [ + { + "bbox": [ + 132, + 116, + 482, + 149 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 116, + 482, + 149 + ], + "spans": [ + { + "bbox": [ + 132, + 116, + 482, + 149 + ], + "type": "text", + "content": "51. Viswanath, B., Mislove, A., Cha, M., Gummadi, K.P.: On the evolution of user interaction in facebook. In: Proceedings of the 2nd ACM workshop on Online social networks. pp. 37-42 (2009)" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 132, + 150, + 481, + 172 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 150, + 481, + 172 + ], + "spans": [ + { + "bbox": [ + 132, + 150, + 481, + 172 + ], + "type": "text", + "content": "52. Wang, W., Chen, C., Wang, Y., Jiang, T., Fang, F., Yao, Y.: Simulating human saccadic scanpaths on natural images. In: CVPR 2011. pp. 441-448. IEEE (2011)" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 132, + 172, + 481, + 205 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 172, + 481, + 205 + ], + "spans": [ + { + "bbox": [ + 132, + 172, + 481, + 205 + ], + "type": "text", + "content": "53. Xia, C., Han, J., Qi, F., Shi, G.: Predicting human saccadic scanpaths based on iterative representation learning. IEEE Transactions on Image Processing 28(7), 3502-3515 (2019)" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 132, + 205, + 481, + 237 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 205, + 481, + 237 + ], + "spans": [ + { + "bbox": [ + 132, + 205, + 481, + 237 + ], + "type": "text", + "content": "54. Xu, M., Yang, L., Tao, X., Duan, Y., Wang, Z.: Saliency prediction on omnidirectional image with generative adversarial imitation learning. IEEE Transactions on Image Processing 30, 2087-2102 (2021)" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 132, + 237, + 481, + 282 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 237, + 481, + 282 + ], + "spans": [ + { + "bbox": [ + 132, + 237, + 481, + 282 + ], + "type": "text", + "content": "55. Yang, Z., Huang, L., Chen, Y., Wei, Z., Ahn, S., Zelinsky, G., Samaras, D., Hoai, M.: Predicting goal-directed human attention using inverse reinforcement learning. In: Proceedings of the IEEE/CVF conference on computer vision and pattern recognition. pp. 193-202 (2020)" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 132, + 282, + 481, + 304 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 282, + 481, + 304 + ], + "spans": [ + { + "bbox": [ + 132, + 282, + 481, + 304 + ], + "type": "text", + "content": "56. Zanca, D., Melacci, S., Gori, M.: Gravitational laws of focus of attention. IEEE transactions on pattern analysis and machine intelligence 42(12), 2983-2995 (2019)" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 132, + 304, + 481, + 325 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 304, + 481, + 325 + ], + "spans": [ + { + "bbox": [ + 132, + 304, + 481, + 325 + ], + "type": "text", + "content": "57. Zhu, Y., Zhai, G., Min, X.: The prediction of head and eye movement for 360 degree images. Signal Processing: Image Communication 69, 15-25 (2018)" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 132, + 325, + 481, + 358 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 325, + 481, + 358 + ], + "spans": [ + { + "bbox": [ + 132, + 325, + 481, + 358 + ], + "type": "text", + "content": "58. Zhu, Y., Zhai, G., Min, X., Zhou, J.: The prediction of saliency map for head and eye movements in 360 degree images. IEEE Transactions on Multimedia 22(9), 2331-2344 (2019)" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 132, + 358, + 481, + 380 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 358, + 481, + 380 + ], + "spans": [ + { + "bbox": [ + 132, + 358, + 481, + 380 + ], + "type": "text", + "content": "59. Zou, Z., Ye, M., Li, S., Li, X., Dufaux, F.: " + }, + { + "bbox": [ + 132, + 358, + 481, + 380 + ], + "type": "inline_equation", + "content": "360^{\\circ}" + }, + { + "bbox": [ + 132, + 358, + 481, + 380 + ], + "type": "text", + "content": " image saliency prediction by embedding self-supervised proxy task. IEEE Transactions on Broadcasting (2023)" + } + ] + } + ], + "index": 10 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "spans": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "type": "text", + "content": "18" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 166, + 91, + 225, + 102 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 166, + 91, + 225, + 102 + ], + "spans": [ + { + "bbox": [ + 166, + 91, + 225, + 102 + ], + "type": "text", + "content": "R. Quan et al." + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 17 + } + ], + "_backend": "vlm", + "_version_name": "2.6.4" +} \ No newline at end of file diff --git a/2024/Pathology-knowledge Enhanced Multi-instance Prompt Learning for Few-shot Whole Slide Image Classification/3e473ac8-930f-46c0-b97b-e8e52a7c0bfd_content_list.json b/2024/Pathology-knowledge Enhanced Multi-instance Prompt Learning for Few-shot Whole Slide Image Classification/3e473ac8-930f-46c0-b97b-e8e52a7c0bfd_content_list.json new file mode 100644 index 0000000000000000000000000000000000000000..2a8a6efc70f1f0053afee9f046af66fe26c0c7db --- /dev/null +++ b/2024/Pathology-knowledge Enhanced Multi-instance Prompt Learning for Few-shot Whole Slide Image Classification/3e473ac8-930f-46c0-b97b-e8e52a7c0bfd_content_list.json @@ -0,0 +1,1587 @@ +[ + { + "type": "text", + "text": "Pathology-knowledge Enhanced Multi-instance Prompt Learning for Few-shot Whole Slide Image Classification", + "text_level": 1, + "bbox": [ + 259, + 140, + 743, + 209 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Linhao Qu $^{1}$ , Dingkang Yang $^{2}$ , Dan Huang $^{3}$ , Qinhao Guo $^{4}$ , Rongkui Luo $^{5}$ , Shaoting Zhang $^{1}$ , and Xiaosong Wang $^{1\\boxtimes}$", + "bbox": [ + 217, + 233, + 782, + 273 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "1 Shanghai Artificial Intelligence Laboratory, Shanghai, China", + "bbox": [ + 292, + 282, + 709, + 297 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "$^{2}$ Academy for Engineering and Technology, Fudan University, Shanghai, China", + "bbox": [ + 235, + 297, + 766, + 311 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "$^{3}$ Department of Pathology, Fudan University Shanghai Cancer Center, China", + "bbox": [ + 238, + 311, + 761, + 325 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "$^{4}$ Department of Gynecologic Oncology, Fudan University Shanghai Cancer Center", + "bbox": [ + 223, + 325, + 777, + 339 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "$^{5}$ Department of Pathology, Zhongshan Hospital, Fudan University, Shanghai, China wangxiaosong@pjlab.org.cn;", + "bbox": [ + 220, + 339, + 782, + 368 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Abstract. Current multi-instance learning algorithms for pathology image analysis often require a substantial number of Whole Slide Images for effective training but exhibit suboptimal performance in scenarios with limited learning data. In clinical settings, restricted access to pathology slides is inevitable due to patient privacy concerns and the prevalence of rare or emerging diseases. The emergence of the Few-shot Weakly Supervised WSI Classification accommodates the significant challenge of the limited slide data and sparse slide-level labels for diagnosis. Prompt learning based on the pre-trained models (e.g., CLIP) appears to be a promising scheme for this setting; however, current research in this area is limited, and existing algorithms often focus solely on patch-level prompts or confine themselves to language prompts. This paper proposes a multi-instance prompt learning framework enhanced with pathology knowledge, i.e., integrating visual and textual prior knowledge into prompts at both patch and slide levels. The training process employs a combination of static and learnable prompts, effectively guiding the activation of pre-trained models and further facilitating the diagnosis of key pathology patterns. Lightweight Messenger (self-attention) and Summary (attention-pooling) layers are introduced to model relationships between patches and slides within the same patient data. Additionally, alignment-wise contrastive losses ensure the feature-level alignment between visual and textual learnable prompts for both patches and slides. Our method demonstrates superior performance in three challenging clinical tasks, significantly outperforming comparative few-shot methods.", + "bbox": [ + 259, + 398, + 743, + 731 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Keywords: Pathology image analysis $\\cdot$ Prompt learning", + "bbox": [ + 261, + 743, + 643, + 757 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "1 Introduction", + "text_level": 1, + "bbox": [ + 215, + 780, + 375, + 796 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Deep learning for pathology Whole Slide Image (WSI) classification offers a rapid, objective, and precise diagnostic approach, enhancing prognosis predic", + "bbox": [ + 212, + 809, + 787, + 840 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/404a3d468aa48cec21122f2dfb34cb762b56fb207372dc71391b81cf728087ec.jpg", + "image_caption": [ + "Fig. 1: Existing methods such as CoOp [42] do not fully consider task-related pathology visual features and its association with specific terms in FSWC. We utilize aligned task-specific image examples and language descriptions to enhance visual and textual prompt learning at both patch and slide levels." + ], + "image_footnote": [], + "bbox": [ + 333, + 146, + 671, + 306 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "tion and treatment response assessment [3, 4, 8, 13, 20, 26]. Given the gigapixel resolution of WSIs, segmentation into multiple patches is essential for processing. However, obtaining detailed annotations for each patch is challenging. Instead, it often results in slide-level ground truth labels. Multiple Instance Learning (MIL), a prominent weakly supervised paradigm, has emerged as a key technique for WSI classification [5, 9, 25, 30]. In MIL, each WSI is treated as a bag, with patches as instances. During training, only slide-level labels are available, necessitating a comprehensive model to capture relationships between patches and slide labels for accurate classification [28].", + "bbox": [ + 212, + 411, + 782, + 547 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "MIL algorithms achieve slide-level classification by aggregating patch-level features or predictions. Though effective, existing algorithms typically demand a considerable number of slides for practical training, performing inadequately in scenarios with limited samples. In clinical practice, challenges such as preserving patient privacy, the rarity of certain diseases, and emerging conditions, particularly for prognosis and rare tumor subtypes, hinder the acquisition of a substantial quantity of WSIs [9,26,33]. Recently, the challenging problem of Few-shot weakly Supervised WSI Classification (FSWC) has emerged [27]. FSWC is characterized by the limited availability of training WSIs (e.g., typically 2, 4, 8, 16, or 32 per class) and sparse slide-level annotations, with unknown fine-grained annotations for numerous patches in each WSI.", + "bbox": [ + 212, + 550, + 784, + 715 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "In image classification, recent advancements highlight the effectiveness of Vision-Language Models (VLMs) like CLIP [29], BLIP [21], and Flamingo [1]. These models undergo training on large-scale language-image pairs using alignment contrastive learning, demonstrating notable zero-shot and few-shot transfer learning capabilities. In the pathology image domain, despite valuable contributions from initiatives like MI-Zero [24] and PLIP [14] in constructing VLMs for computational pathology, challenges including ethical concerns, annotation scarcity, and storage limitations make it impractical to train foundational mod-", + "bbox": [ + 212, + 719, + 785, + 839 + ], + "page_idx": 1 + }, + { + "type": "page_number", + "text": "2", + "bbox": [ + 217, + 114, + 228, + 126 + ], + "page_idx": 1 + }, + { + "type": "header", + "text": "L. Qu et al.", + "bbox": [ + 271, + 114, + 351, + 128 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "els for each specific downstream task [27]. Furthermore, existing VLMs exhibit restricted direct zero-shot transfer capabilities for specific downstream tasks [7,36,38]. Considering these challenges, the application of existing VLMs for developing prompt learning algorithms tailored to few-shot downstream tasks appears promising. Prompt learning involves adapting a pre-trained and then fixed VLM to new tasks by integrating task-specific prompts as additional model inputs. In the context of few-shot learning, these prompts can be either fixed or trainable, with trainable prompts enabling contrastive learning between visual and language features [11,42].", + "bbox": [ + 212, + 146, + 787, + 282 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Presently, research on prompt learning for FSWC is limited [27]. Previous efforts, such as MI-Zero [24], PLIP [14], and CITE [41], have predominantly focused on the utilization of \"patch-level\" prompts—textual prompts associated with pathology regions of interest. Nonetheless, visual or textual prompts in domain-specific downstream tasks may not be understandable (often unseen) for the pre-trained model and thus may not properly activate the relevant features of the pre-trained model. The TOP framework [27] represents the initial endeavor to implement prompt learning by introducing language-based prior descriptions in the prompts. However, pathology images encompass highly specialized visual features and corresponding pathology-specific terms, posing a challenge for networks to comprehensively understand and learn pathology-related features and specialized descriptions solely through language explanations for classification.", + "bbox": [ + 212, + 282, + 787, + 464 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "To address this issue, we propose Pathology-knowledge Enhanced Multi-instance Prompt Learning for FSWC named PEMP, with a concise architecture depicted in Fig. 1. PEMP is a multi-instance prompt learning framework specifically designed for the FSWC problem, incorporating both visual and textual samples (as a form of prior knowledge) at both patch and slide levels. Drawing inspiration from pathologists learning pathology images and language knowledge from textbooks, we introduce task-specific image examples and corresponding language descriptions, injected in both the visual and textual prompts. During prompt learning, the model is fed with a more understandable and effective guide (in image examples and plain words) and then activates more meaningful features for the diagnosis of key patch-level and slide-level pathology patterns in scenarios with limited training samples.", + "bbox": [ + 212, + 464, + 787, + 645 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "The contribution of this work is threefold:", + "bbox": [ + 238, + 646, + 544, + 660 + ], + "page_idx": 2 + }, + { + "type": "list", + "sub_type": "text", + "list_items": [ + "- We propose a multi-instance prompt learning framework, PEMP, introducing vision and text prior knowledge in the designed prompts at both patch and slide levels. A combination of static and learnable prompts is employed in the training, effectively adapting the pre-trained model to focus on key pathology patterns in patches and slides.", + "- We explore the alignment of visual and textual prompts using contrastive losses at both patch and slide levels. We design lightweight Messenger (self-attention) and Summary (attention-pooling) layers to model relationships of patch-to-patch and patch-to-slide within the same patient data.", + "- We evaluate PEMP's performance in addressing three challenging pathology tasks in clinical practice, including predicting patient survival and lymph" + ], + "bbox": [ + 225, + 672, + 785, + 840 + ], + "page_idx": 2 + }, + { + "type": "header", + "text": "Pathology-knowledge Enhanced Multi-instance Prompt Learning", + "bbox": [ + 300, + 114, + 730, + 130 + ], + "page_idx": 2 + }, + { + "type": "page_number", + "text": "3", + "bbox": [ + 774, + 116, + 784, + 126 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "node metastasis in early cervical cancer H&E slides and subtype classification for a rare tumor (round cell tumor). PEMP achieves superior performance in all these tasks under the few-shot setting, notably outperforming all comparative methods by a large margin (4% on average).", + "bbox": [ + 241, + 146, + 787, + 208 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "2 Related Work", + "text_level": 1, + "bbox": [ + 215, + 231, + 388, + 247 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "2.1 MIL for WSI Classification", + "text_level": 1, + "bbox": [ + 215, + 263, + 485, + 279 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Current MIL algorithms for WSI classification can be broadly categorized as instance-based [2, 10, 22, 28, 37] and bag-based [6, 12, 16, 18, 19, 23, 25, 31, 32, 34, 35, 39, 40, 43] methods. Instance-based methods involve generating pseudo-labels for each patch and training patch classifiers to score all patches within a slide, whereas bag-based methods, the current mainstream, extract patch features and employ techniques like attention weighting [12, 16, 18, 25, 32, 39, 40, 43] to aggregate these features into slide-level features. While existing algorithms perform well with ample training data, they lack specificity for FSWC tasks, struggling with limited samples. Our proposed PEMP, falling under the bag-based MIL paradigm, distinguishes itself by innovatively addressing FSWC through VLMs and prompt learning, deviating from traditional bag-based MIL approaches.", + "bbox": [ + 212, + 291, + 787, + 458 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "2.2 VLM-Based Prompt Learning", + "text_level": 1, + "bbox": [ + 215, + 479, + 511, + 496 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "VLMs, comprising image and text encoders, are trained on large-scale image-text pairs using alignment contrastive learning to establish a shared feature space. Pre-trained VLMs like CLIP [29], BLIP [21], and Flamingo [1] exhibit robust zero-shot and few-shot transferability in image recognition tasks. However, limited accuracy in zero-shot transfer has fueled interest in few-shot learning based on VLMs, where the prompt learning paradigm, demonstrating strong performance [7,36,38], is emerging. Prompt learning leverages a pre-trained VLM in the adaptation to new downstream tasks by incorporating task-specific prompts through enriched model inputs. These prompts could be static with manual designs or learnable vectors [11,42].", + "bbox": [ + 212, + 507, + 787, + 657 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "In pathology, recent research emphasizes constructing specialized VLMs, such as MI-Zero [24], PLIP [14], and QUILT [15]. However, privacy concerns, annotation scarcity, and storage demands make training separate foundational models for each downstream task impractical. Existing VLMs exhibit limited zero-shot transferability, especially for rare diseases. Unlike fine-tuning VLMs at the data level or constructing large datasets like these works, we introduce a novel prompt learning paradigm based on VLMs (e.g., CLIP with frozen parameters) to address the FSWC problem. Importantly, our scheme is versatile across various foundation models. Moreover, MI-Zero, PLIP and QUILT mainly contribute by collecting large datasets to construct pathology VLMs but lack effective methods for FSWC. Their simple adapting strategies for FSWC (e.g., LinearProbe and Attention Pooling) limit their performance in challenging clinical tasks.", + "bbox": [ + 212, + 659, + 787, + 840 + ], + "page_idx": 3 + }, + { + "type": "page_number", + "text": "4", + "bbox": [ + 217, + 114, + 228, + 126 + ], + "page_idx": 3 + }, + { + "type": "header", + "text": "L. Qu et al.", + "bbox": [ + 271, + 114, + 351, + 128 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "In pathology, prompt learning research for FSWC is limited. Prior works like MI-Zero [24], PLIP [14], and CITE [41] only explore \"patch-level\" prompts, limiting slide-level classification performance. TOP [27] introduces prompt learning at both patch and slide levels but lacks task-specific visual knowledge on the vision side. Relying solely on language explanations is challenging for the networks to understand pathology features with limited training samples. In contrast, PEMP's learning process operates at patch and slide levels, covers visual and textual prompt design, and introduces task-specific image examples and descriptions separately as prior knowledge.", + "bbox": [ + 212, + 146, + 787, + 282 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "3 FSWC Baseline Framework", + "text_level": 1, + "bbox": [ + 214, + 304, + 519, + 320 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "3.1 Problem Formulation", + "text_level": 1, + "bbox": [ + 214, + 334, + 437, + 348 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "We provide a brief introduction to the problem definition of FSWC and the pre-trained VLM used in our study. Given a dataset $W = \\{W_{1}, W_{2}, \\ldots, W_{N}\\}$ containing $N$ WSIs, each WSI $W_{i}$ is divided into non-overlapping small image patches $\\{p_{i,j}, j = 1, 2, \\ldots, n_{i}\\}$ , where $n_{i}$ represents the number of patches obtained from $W_{i}$ . It is important to note that the number of patches obtained from each slide may vary. All patches in $W_{i}$ collectively form a bag, and each patch serves as an instance of this bag. Each bag is assigned a label $Y_{i} \\in \\{0, 1, \\ldots, U\\}$ , while the label of each instance is unknown, and $U$ is the number of classes. FSWC is highly challenging as it allows training with only a limited number of slides (with labels). In the FSWC task, \"shot\" refers to the number of WSIs, e.g., $K$ pairs of WSIs and labels for training in a $K$ -shot setting, while the trained model is evaluated on a completely independent test set. Typically, $K$ can take values such as 2, 4, 8, 16, or 32.", + "bbox": [ + 212, + 358, + 787, + 555 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "3.2 Attention Aggregation for WSI Classification", + "text_level": 1, + "bbox": [ + 214, + 575, + 635, + 590 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "The attention-based pooling strategy [16] is a classical MIL aggregation method used for WSI classification. Initially, an encoder $\\phi_{img}$ is used to extract features $f_{i,j}$ for all patches $\\{p_{i,j}, j = 1,2,\\dots n_i\\}$ in slide $W_{i}$ , and then aggregate these patch features using an attention-based trainable aggregation function $A(\\cdot)$ to obtain the bag feature $F_{i}$ :", + "bbox": [ + 212, + 599, + 787, + 676 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\nF _ {i} = A \\left(\\phi_ {i m g} \\left(p _ {i, 1}\\right), \\phi_ {i m g} \\left(p _ {i, 2}\\right), \\dots\\right) = \\sum_ {j = 1} ^ {n _ {i}} a _ {i, j} f _ {i, j}, \\tag {1}\n$$\n", + "text_format": "latex", + "bbox": [ + 326, + 686, + 785, + 727 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\na _ {i, j} = \\frac {\\exp \\left\\{w ^ {\\top} \\tanh \\left(V f _ {i , j} ^ {\\top}\\right) \\right\\}}{\\sum_ {j = 1} ^ {n _ {i}} \\exp \\left\\{w ^ {\\top} \\tanh \\left(V f _ {i , j} ^ {\\top}\\right) \\right\\}}, \\tag {2}\n$$\n", + "text_format": "latex", + "bbox": [ + 372, + 737, + 785, + 776 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "where $a_{i,j}$ is the attention score predicted by the self-attention network, which is parameterized by $w$ and $V$ . The weights $a_{i,j}$ reflect the contribution of each patch in making the slide-level prediction. The patch-level encoder $\\phi_{img}$ is from the pre-trained CLIP [29] and remains fixed during the training process.", + "bbox": [ + 212, + 779, + 787, + 840 + ], + "page_idx": 4 + }, + { + "type": "header", + "text": "Pathology-knowledge Enhanced Multi-instance Prompt Learning", + "bbox": [ + 300, + 114, + 732, + 130 + ], + "page_idx": 4 + }, + { + "type": "page_number", + "text": "5", + "bbox": [ + 774, + 116, + 784, + 126 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "3.3 VLM backbone and Few-shot Prompt Learning", + "text_level": 1, + "bbox": [ + 215, + 146, + 653, + 162 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "We utilize CLIP [29] as the backbone of the proposed framework and also the pretrained VLM for feature extraction, which is trained on a massive dataset of 400 million image-text pairs, including medical data, facilitating efficient knowledge transfer of diverse visual concepts to downstream tasks.", + "bbox": [ + 212, + 169, + 782, + 228 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Prompt learning enables the efficient adaptation of a fixed pre-trained CLIP in few-shot downstream classification tasks by introducing task-specific prompts into the model input. These prompts, whether fixed or learnable, can be applied to the text encoder, visual encoder, or both simultaneously. In the few-shot setting, the training could also involve aligning visual and text features through alignment contrastive learning (AC-Loss) in the specific domain. Formally, we denote the visual encoder as $\\phi_{img}$ and the text encoder as $\\phi_{text}$ . For a downstream classification task with $U$ classes, a commonly used fixed hand-crafted prompt, denoted as [Class Prompt], guides classification on the text side. Specifically, assuming the name of the $i$ -th class is “[class-name],\" the corresponding hand-crafted prompt is typically designed as “a photo of a [class-name].” These [Class Prompts] are pre-encoded into a vector $C = \\{c_i\\}_{i=1}^U$ .", + "bbox": [ + 212, + 229, + 784, + 412 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "To further enhance generalization to downstream tasks, it is also popular to explore learning a set of continuous context vectors as learnable prompts, denoted as $\\mathbb{V} = \\{v_1,v_2,\\ldots ,v_M\\}$ . These learnable context vectors $\\mathbb{V}$ , concatenated with the vectors from [Class Prompt] $C$ , form the complete text prompt $t_i = \\{v_1,v_2,\\dots ,v_M,c_i\\}$ . This prompt is then fed into the text encoder $\\phi_{text}$ to obtain the final textual features for each class, $T_{i} = \\phi_{text}(t_{i})$ . On the visual side, given an image $W_{i}$ and its label $y$ , its slide features $F_{i}$ is computed via Eq. (1). Similar to the text side, learnable prompts could also be appended, which will be inputted into the visual encoder $\\phi_{img}$ .", + "bbox": [ + 212, + 412, + 784, + 547 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "With a limited number of image-text pairs, the optimization involves minimizing the negative log-likelihood between the visual slide feature $F_{i}$ and its class text feature $T_{y}$ , updating all learnable parameters. During testing, the extracted visual features $F_{i}$ are used to retrieve the class text features $T_{y}$ to obtain the probabilities for each class.", + "bbox": [ + 212, + 549, + 785, + 625 + ], + "page_idx": 5 + }, + { + "type": "equation", + "text": "\n$$\np (y \\mid F _ {i}) = \\frac {\\exp (\\operatorname {s i m} \\left(F _ {i} , T _ {y}\\right) / \\tau)}{\\sum_ {i = 1} ^ {U} \\exp (\\operatorname {s i m} \\left(F _ {i} , T _ {i}\\right) / \\tau)}, \\tag {3}\n$$\n", + "text_format": "latex", + "bbox": [ + 366, + 633, + 785, + 670 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "where, $y$ represents the ground truth class, $U$ is the total number of classes, $\\mathrm{sim}(\\cdot)$ denotes cosine similarity, and $\\tau$ is a learnable temperature parameter. It is crucial to note that during the training process, the parameters of CLIP's image encoder and text encoder remain fixed and unchanged.", + "bbox": [ + 212, + 675, + 787, + 737 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "4 Knowledge-enhanced Prompt Learning", + "text_level": 1, + "bbox": [ + 215, + 756, + 630, + 776 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "4.1 Overview", + "text_level": 1, + "bbox": [ + 215, + 786, + 339, + 801 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Fig. 2 illustrates the primary workflow of our proposed PEMP, which comprises three learning processes: Visual prompt learning, Textual prompt learning, and", + "bbox": [ + 212, + 809, + 785, + 840 + ], + "page_idx": 5 + }, + { + "type": "page_number", + "text": "6", + "bbox": [ + 217, + 114, + 228, + 126 + ], + "page_idx": 5 + }, + { + "type": "header", + "text": "L. Qu et al.", + "bbox": [ + 271, + 114, + 351, + 128 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/f73bce66b1b82e2f233c4d1f4045d8cd1037ffd1aca6f64d147ac22b37c9c289.jpg", + "image_caption": [ + "Fig. 2: Overview of PEMP, where flames represent optimized parameters, and snowflakes indicate frozen parameters during training." + ], + "image_footnote": [], + "bbox": [ + 223, + 143, + 782, + 444 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Two-level prompt alignment via contrastive learning. In visual prompt learning, we address three key tasks. Firstly, we integrate prior visual pathology knowledge into the prompts at both patch-level and slide-level. This guides the model, particularly with limited training samples, to understand the prompt better and produce more meaningful activations for task-specific key pathology patterns. Additionally, we also include the learnable prompts on the slide level, triggering the model's precise adaptation to specific tasks. Thirdly, we introduce a lightweight Messenger (self-attention) Layer and Summary (attention-pooling) Layer. The former efficiently communicates and models features of different patches within the same slide, while the latter aggregates patch features into slide features using an attention mechanism.", + "bbox": [ + 212, + 526, + 787, + 691 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "In a similar initiative, textual prompt learning concentrates on modeling language descriptions related to pathology tasks, also incorporating patch-level and slide-level priors. Pathologists are employed to provide straightforward language descriptions in the overall task, task-related slide-level pathology, and task-related patch-level pathology, avoiding obscure professional terms.", + "bbox": [ + 212, + 698, + 787, + 776 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Recognizing the challenges faced by the VLM in understanding key visual features and complex pathology terms in visual examples, we establish a close alignment between textual prompt learning (with integrated pathology language descriptions) and visual prompt learning (with injected visual examples) at both", + "bbox": [ + 212, + 779, + 787, + 840 + ], + "page_idx": 6 + }, + { + "type": "header", + "text": "Pathology-knowledge Enhanced Multi-instance Prompt Learning", + "bbox": [ + 300, + 114, + 732, + 128 + ], + "page_idx": 6 + }, + { + "type": "page_number", + "text": "7", + "bbox": [ + 774, + 116, + 784, + 126 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/6db8b8b77b6bd76bff780a10a8034b7e1e89c17c9ffa2f89887e3cd128f3d194.jpg", + "image_caption": [ + "Fig. 3: (A) and (B) examples of visual and textual token constructions related to the prognosis of early cervical cancer patients." + ], + "image_footnote": [ + "(a) Examples for predicting the prognosis of early cervical cancer patients from H&E pathology slides of the primary lesion, derived from authoritative textbooks and pathology experts. Top: visual patch examples. Bottom: visual slide examples." + ], + "bbox": [ + 217, + 143, + 485, + 224 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/442e7c81a69bbfc81aba21262bee527b54204d7880347cf5c89af58349d7a4bf.jpg", + "image_caption": [], + "image_footnote": [ + "(b) Top: Construction of Patch-level descriptive [Token]. Bottom: Construction of Slide-level descriptive [Token]. Examples for the prognosis of early cervical cancer patients from H&E pathology slides of the primary lesion, serving as task-related patch-level and slide-level pathology priors on the text side." + ], + "bbox": [ + 529, + 143, + 790, + 224 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "slide and patch levels. This alignment enhances the organization and cohesion of knowledge-enhanced prompts on both visual and textual sides.", + "bbox": [ + 212, + 369, + 785, + 402 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "4.2 Visual Prompt Learning", + "text_level": 1, + "bbox": [ + 215, + 430, + 462, + 446 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Construction of visual prompts. In the FSWC task, limited training samples present challenges in acquiring effective knowledge. Moreover, tasks like survival analysis require finding key factors both locally (e.g., patch-level: necrosis, vascular invasion) and globally (e.g., slide-level: tumor boundary, tumor stroma ratio). To overcome these, we out-source prior visual pathology knowledge at both patch and slide levels (shown in the top part of Fig. 3a), guiding the model with task-specific pathology patterns for diagnosis. For challenging slide-level diagnostic tasks, relevant patterns manifest at both levels. Leveraging expertise from pathology professionals, we construct visual patch and slide examples as fixed prompts, as shown in Fig. 3a. For each classification task, highly relevant visual patches and slides indicating high and low risks are retrieved from authoritative sources or pathology experts. For instance, in the prognosis of early cervical cancer, visual patch examples for poor prognosis include \"high-grade atypia\", \"vascular invasion,\" and \"necrosis,\" while slide examples encompass \"overall infiltration with fuzzy tumor boundaries and low tumor stroma ratio.\" Conversely, low-risk examples include \"low-grade dysplasia,\" \"no vascular invasion,\" and \"overall pushing-type with clear tumor boundaries and high tumor stroma ratio.\"", + "bbox": [ + 212, + 460, + 787, + 733 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "In total, 6 slides and 6 patch images are used as visual prompts, including 3 typical slides per prognosis category (good and poor), each with 3 local patches in the prognosis of early cervical cancer. For the Lymph Node Metastasis Prediction Task, we used 4 slides and 4 patches, with two representative slides per category (metastasis and non-metastasis) and two local images per slide. For Round Cell Subtype Diagnosis, 10 slides and 10 patches were used, with two slides per category, each with two local images. Selection of these visual exam", + "bbox": [ + 212, + 734, + 787, + 840 + ], + "page_idx": 7 + }, + { + "type": "page_number", + "text": "8", + "bbox": [ + 217, + 114, + 228, + 126 + ], + "page_idx": 7 + }, + { + "type": "header", + "text": "L. Qu et al.", + "bbox": [ + 271, + 114, + 351, + 128 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "samples was guided by clinical experts, with each risk factor represented by one slide of the patch and its local patch.", + "bbox": [ + 212, + 146, + 782, + 176 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "Learning pipeline. As illustrated in Fig. 2, we assume $W_{i}$ represents the training slide containing $n_i$ patches named $p_{i,j}$ and utilize the fixed CLIP image encoder $\\phi_{img}$ to extract patch features $f_{i,j} = \\phi_{img}(p_{i,j}) \\in \\mathbb{R}^{1 \\times v}$ , where $v$ represents the dimension of the features. Subsequently, the constructed visual patch examples $e_l$ are also transformed into patch features $z_l = \\phi_{img}(e_l)$ . $z_l$ remains unchanged throughout the training process and is combined with the current slide's patch features as a fixed visual prompt. Specifically, we match the cosine similarity between $f_{i,j}$ and $z_l$ , concatenate the most matched example patch feature with the patch feature as $f_{i,j}^e$ , and ultimately concatenate all new patch features for the current slide $F_i^e = \\{f_{i,j}^e\\}_{j=1}^{n_i} \\in \\mathbb{R}^{n_i \\times v \\times 2}$ . To model all patch features within a slide effectively, we introduce a lightweight self-attention layer named Messenger Layer (ML). Taking $F_i^e$ as input, it outputs $F_i^{ML}$ :", + "bbox": [ + 212, + 176, + 787, + 364 + ], + "page_idx": 8 + }, + { + "type": "equation", + "text": "\n$$\nF _ {i} ^ {M L} = \\operatorname {s o f t m a x} \\left(\\frac {Q K ^ {\\top}}{\\sqrt {d _ {w}}}\\right) V, \\tag {4}\n$$\n", + "text_format": "latex", + "bbox": [ + 401, + 375, + 785, + 407 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "where $Q$ , $K$ , and $V$ are Query, Key, and Value vectors obtained through linear mapping of $F_{i}^{e}$ , and $d_{w}$ is the feature dimension. Next, we construct a lightweight Attention Aggregation Layer to aggregate all patch features $F_{i}^{ML}$ within a slide into a slide feature $F_{i}^{S}$ . The attention weights are learnable, and the detailed structure is shown in Section 3.2. Following this, the constructed visual slide examples $E_{m}$ undergo the aforementioned process like $W_{i}$ (i.e., patch extraction, patch example matching, Messenger Layer, and Summary Layer) to obtain their slide features $Z_{m}$ . These slide example features are also integrated as prompts into slide-level training. Similar to the patch-level matching, we match the cosine similarity between $F_{i}^{S}$ and $Z_{m}$ and concatenate the most matched slide example feature with the slide feature. Additionally, we introduce learnable vector $F_{i}^{P}$ as a slide-level learnable prompt, concatenate it together, and input them into a one-layer fully-connected layer named \"projector\" to adjust the feature dimension, yielding the final slide visual features $F_{i} \\in \\mathbb{R}^{1 \\times v}$ .", + "bbox": [ + 212, + 416, + 787, + 628 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "4.3 Textual Prompt Learning", + "text_level": 1, + "bbox": [ + 215, + 648, + 472, + 666 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "Construction of textual prompts. Textual prompt learning primarily focuses on modeling language descriptions related to pathology tasks, patch-level, and slide-level pathology priors on the text side. This knowledge is integrated into the training process through fixed descriptive prompts and learnable prompts (shown in the middle of Fig. 2). The construction of textual prompts could be divided into mainly three parts. The following detailed explanation uses the task \"predicting the prognosis of early cervical cancer patients from H&E pathology slides of the primary lesion\" as an example, also illustrated in Fig. 3b.", + "bbox": [ + 212, + 672, + 787, + 794 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "First, the \"Slide task [Token]\" primarily serves to prompt the main category of the current task, such as \"A Whole Slide Image of cervical cancer from primary tumors with a [poor/good] prognosis.\" Furthermore, we concatenate", + "bbox": [ + 212, + 795, + 785, + 840 + ], + "page_idx": 8 + }, + { + "type": "header", + "text": "Pathology-knowledge Enhanced Multi-instance Prompt Learning", + "bbox": [ + 300, + 114, + 730, + 130 + ], + "page_idx": 8 + }, + { + "type": "page_number", + "text": "9", + "bbox": [ + 774, + 116, + 784, + 126 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "learnable prompts $(([\\alpha]_1[\\alpha]_2\\ldots [\\alpha]_M))$ with the \"Slide task [Token]\" to adaptively adjust and complement during few-shot training.", + "bbox": [ + 212, + 145, + 782, + 176 + ], + "page_idx": 9 + }, + { + "type": "text", + "text": "Then, the \"Slide-level descriptive [Token]\" is a straightforward language description corresponding to the aforementioned visual slide examples. These descriptions are provided by experienced pathology experts, avoiding obscure professional terminology. For instance, the description corresponding to the \"Poor prognosis\" category is: \"Fuzzy Tumor Boundaries. The edges or borders of the tumor are fuzzy or hard to distinguish on the slide level. Low tumor stroma ratio. Few stroma tissues surrounding the tumors.\" In contrast, the description corresponding to the \"Good prognosis\" category is: \"Clear Tumor Boundaries. The edges or borders of the tumor are easy to see and well-defined on the slide level. High tumor stroma ratio. Lots of stroma tissue surrounding the tumors.\" Similarly, we also append learnable prompts $(([\\beta]_{1}[\\beta]_{2}\\ldots [\\beta]_{M}))$ .", + "bbox": [ + 212, + 176, + 785, + 343 + ], + "page_idx": 9 + }, + { + "type": "text", + "text": "Finally, the \"Patch-level descriptive [Token]\" provides the language description corresponding to the aforementioned visual patch examples. For example, the description corresponding to the \"Poor prognosis\" category is: \"Vascular and lymphatic invasion: The presence of cancer cells within the blood and lymphatic vessels. High-level atypia: Pronounced and abnormal variations in the size, shape, and staining pattern of the cells. Necrosis: Empty spaces or regions with a lack of normal cell structure.\" Similarly, immediately following this, we also include learnable prompts $(([\\gamma]_{1}[\\gamma]_{2}\\ldots [\\gamma]_{M}))$ .", + "bbox": [ + 212, + 345, + 787, + 465 + ], + "page_idx": 9 + }, + { + "type": "text", + "text": "It is crucial to note that the second and third parts correspond closely to the same number of visual slide and patch examples in visual prompt learning. We input each part's visual descriptions and learnable vectors together into the CLIP Text Encoder, obtaining the final slide textual features $T_{i} \\in \\mathbb{R}^{1 \\times v}$ .", + "bbox": [ + 212, + 465, + 787, + 526 + ], + "page_idx": 9 + }, + { + "type": "text", + "text": "4.4 Alignment of Visual and Textual Prompts", + "text_level": 1, + "bbox": [ + 215, + 551, + 609, + 566 + ], + "page_idx": 9 + }, + { + "type": "text", + "text": "Alignment contrastive learning (AC-Loss) is originally used to align corresponding visual representation and textual representation. Nonetheless, the overall loss function in this paper consists of three parts of AC-Loss, expressed as $\\mathcal{L}_{\\mathrm{total}} = \\mathcal{L}_t + \\lambda_1\\mathcal{L}_s + \\lambda_2\\mathcal{L}_p$ . Conceptually, $\\mathcal{L}_t$ aligns the overall slide visual features $F_{i}$ with the slide textual features $T_{i}$ corresponding to the task category, thereby completing slide-level classification. $\\mathcal{L}_s$ aligns the features of slide-level visual examples in visual prompt learning with the slide-level pathological language descriptions in textual prompt learning. $\\mathcal{L}_p$ aligns the features of patch-level visual examples with patch-level pathology language descriptions (shown in the bottom of Fig. 2).", + "bbox": [ + 212, + 579, + 787, + 731 + ], + "page_idx": 9 + }, + { + "type": "text", + "text": "Mathematically, the forms of these three parts of AC-Loss are identical. Here, we take $\\mathcal{L}_t$ as an example. AC-Loss optimizes the learnable parameters by minimizing the negative log-likelihood between the image features $F_{i}$ and the ground truth class text features $T_{y}$ . Its basic form is as follows:", + "bbox": [ + 212, + 731, + 787, + 792 + ], + "page_idx": 9 + }, + { + "type": "equation", + "text": "\n$$\n\\mathcal {L} = - \\sum_ {F _ {i}} \\log \\frac {\\exp \\left(\\operatorname {s i m} \\left(F _ {i} , T _ {y}\\right) / \\tau\\right)}{\\sum_ {i = 1} ^ {U} \\exp \\left(\\operatorname {s i m} \\left(F _ {i} , T _ {i}\\right) / \\tau\\right)}, \\tag {5}\n$$\n", + "text_format": "latex", + "bbox": [ + 357, + 805, + 785, + 844 + ], + "page_idx": 9 + }, + { + "type": "page_number", + "text": "10", + "bbox": [ + 217, + 114, + 235, + 126 + ], + "page_idx": 9 + }, + { + "type": "header", + "text": "L. Qu et al.", + "bbox": [ + 271, + 114, + 351, + 128 + ], + "page_idx": 9 + }, + { + "type": "text", + "text": "where $y$ is the ground truth category, $U$ is the total number of categories, $\\mathrm{sim}(\\cdot)$ denotes the cosine similarity, and $\\tau$ is a learnable temperature parameter.", + "bbox": [ + 212, + 146, + 782, + 176 + ], + "page_idx": 10 + }, + { + "type": "text", + "text": "During testing, we input the unseen WSI and obtain its slide visual feature in the same fashion as training. Subsequently, we match it separately with the textual features for \"poor prognosis\" and \"good prognosis\" and use softmax to obtain the predicted risks for the two categories.", + "bbox": [ + 212, + 176, + 787, + 238 + ], + "page_idx": 10 + }, + { + "type": "text", + "text": "5 Experiment and Result", + "text_level": 1, + "bbox": [ + 214, + 262, + 478, + 279 + ], + "page_idx": 10 + }, + { + "type": "text", + "text": "Datasets and Evaluation Metrics. This paper addresses clinical tasks in H&E pathology diagnosis, specifically focusing on data label scarcity, the complexity of direct diagnosis by pathologists, and limited sample sizes for rare diseases. This study encompasses three clinical tasks for five tumor types across five datasets: (1) Predicting patient survival prognosis and lymph node metastasis from H&E slides of early-stage cervical cancer primary lesions using an in-house dataset (For testing, 80 cases and 74 cases, individually) and a public TCGA-CESC dataset (For testing, 80 cases and 70 cases, individually). Note that the prompts utilized in the public TCGA datasets are the same as those used in the corresponding in-house datasets for each task. We ensure no overlap between our training data and visual prompts (obtained from external authoritative sources or pathology experts) and testing data, across both in-house and TCGA datasets. Furthermore, the lymph node metastasis detection tasks from primary lesions assessed in this study present a significantly greater challenge compared to those from sentinel lymph nodes, as utilized in prior research, with samples and labels being notably rarer. (2) Classifying round cell tumors, including neuroendocrine tumors, malignant melanoma, lymphohematopoietic tumors, and soft tissue tumors, into subtypes using an in-house dataset (For testing, 112 cases). More details are in the Supplementary Materials. For the prognosis prediction task in cervical cancer, the concordance index (C-index) is used, while for other tasks, the Area Under the Curve (AUC) is employed. Higher values for both metrics indicate improved performance.", + "bbox": [ + 212, + 295, + 787, + 628 + ], + "page_idx": 10 + }, + { + "type": "text", + "text": "Compared Methods. Currently, few methods directly address FSWC tasks through prompt learning, with the SOTA being represented by the recently proposed TOP method [27] in the pathology domain. In the natural domain, prompt learning work based on VLMs is limited to single image patches and does not readily apply to slide-level classification. Following the TOP approach, we adopt advanced prompt learning methods from the natural domain for the MIL problem. These methods include: (1) Linear-Probe [29], (2) VPT [17], (3) CoOp [42], and (4) KgCoOp [38]. The Messenger Layer and Summary Layer, as described in this paper, aggregate patch features into a slide feature. All compared methods use the same Messenger & Summary layers for fair comparisons. For Linear-Probe, a linear layer performs slide-level classification on the aggregated feature. VPT incorporates a learnable prompt on the vision side for slide-level classification. CoOp utilizes CLIP's prompt learning by adding the Slide Task [Token] and learnable textual prompts trained through AC-Loss. Kg-", + "bbox": [ + 212, + 628, + 787, + 840 + ], + "page_idx": 10 + }, + { + "type": "header", + "text": "Pathology-knowledge Enhanced Multi-instance Prompt Learning", + "bbox": [ + 300, + 114, + 730, + 130 + ], + "page_idx": 10 + }, + { + "type": "page_number", + "text": "11", + "bbox": [ + 767, + 114, + 782, + 126 + ], + "page_idx": 10 + }, + { + "type": "table", + "img_path": "images/bb1c12fcd4f83d8de30a3854a5d076be72a8eff59c35f726d16bfc153f19c7db.jpg", + "table_caption": [ + "Table 1: Comparison for Survival Prognosis Prediction Task." + ], + "table_footnote": [], + "table_body": "
DatasetIn-house DatasetPublic TCGA-CESC Dataset
Method32-shot16-shot8-shot4-shot2-shot32-shot16-shot8-shot4-shot2-shot
LinearProbe0.6200.5620.5430.5010.4580.5840.5770.5430.5110.485
VPT(ECCV'22)0.6260.5690.5450.5010.4640.5890.5730.5500.5180.482
CoOp(IJCV'22)0.6410.5940.5610.5170.4900.6040.5850.5580.5230.487
KgCoOp(CVPR'23)0.6350.5870.5520.5130.4870.5970.5840.5500.5190.486
TOP(NeurIPS'23)0.6520.6080.5740.5390.5080.6110.5970.5660.5360.518
PEMP (ours)0.6670.6370.6140.5870.5620.6370.6240.6020.5770.551
", + "bbox": [ + 272, + 171, + 727, + 239 + ], + "page_idx": 11 + }, + { + "type": "table", + "img_path": "images/1f2dd4f90b6e8026354a48d95b0282ec9b7c5a81fa7ea8260f55b232885ea20c.jpg", + "table_caption": [ + "Table 2: Comparison for Lymph Node Metastasis Prediction Task." + ], + "table_footnote": [], + "table_body": "
DatasetIn-house DatasetPublic TCGA-CESC Dataset
Method32-shot16-shot8-shot4-shot2-shot32-shot16-shot8-shot4-shot2-shot
LinearProbe0.7810.7350.7280.7160.6870.7630.7190.6960.6710.655
VPT(ECCV'22)0.7850.7420.7280.7190.6890.7690.7240.7120.6780.655
CoOp(IJCV'22)0.7920.7570.7350.7270.6930.7860.7410.7230.6910.659
KgCoOp(CVPR'23)0.7890.7530.7310.7240.6930.7750.7330.7100.6870.652
TOP(NeurIPS'23)0.8250.8190.8010.7870.7620.7990.7610.7440.7080.679
PEMP(ours)0.8490.8380.8240.8010.7830.8180.7950.7600.7260.704
", + "bbox": [ + 272, + 279, + 727, + 348 + ], + "page_idx": 11 + }, + { + "type": "table", + "img_path": "images/5bc3176b43b6e30e152e1cdd4c6b0aebf71af604ecc3417f31e9296a458c09f7.jpg", + "table_caption": [ + "Table 3: Comparison for Round Cell Subtype Diagnosis Task." + ], + "table_footnote": [], + "table_body": "
Method32-shot16-shot8-shot4-shot2-shot
LinearProbe0.6540.6060.5490.5170.483
VPT(ECCV'22)0.6590.6130.5560.5150.511
CoOp(IJCV'22)0.6670.6270.5680.5530.549
KgCoOp(CVPR'23)0.6610.6190.5470.5350.534
TOP(NeurIPS'23)0.6820.6520.6330.5840.560
PEMP (ours)0.7510.7180.6850.6430.625
", + "bbox": [ + 330, + 388, + 669, + 463 + ], + "page_idx": 11 + }, + { + "type": "text", + "text": "CoOp imposes constraints on learnable textual prompts. Few-shot experiments are conducted for each category with 2, 4, 8, 16, and 32 slides (with labels). Results are reported as an average over five experimental runs. Standard deviation of our main results are provided in the supplementary materials.", + "bbox": [ + 212, + 489, + 787, + 551 + ], + "page_idx": 11 + }, + { + "type": "text", + "text": "5.1 Results", + "text_level": 1, + "bbox": [ + 215, + 573, + 323, + 587 + ], + "page_idx": 11 + }, + { + "type": "text", + "text": "Survival prognosis prediction. Predicting survival prognosis, especially directly from H&E slides with a limited number of samples, poses a significant challenge. Table 1 showcases the performance of all methods in the in-house dataset and the public TCGA-CESC dataset, with PEMP consistently demonstrating superior performance across all settings with few samples, outperforming all comparative methods significantly. In scenarios with limited samples, LinearProbe and VPT, relying solely on features extracted by the CLIP Image Encoder on the visual side without incorporating language-based prior knowledge, exhibit suboptimal performance. CoOp and KgCoOp, which combine visual and language aspects, struggle to comprehend complex medical terms related to prognosis effectively, limiting their performance. TOP enhances language descriptions at both slide and patch levels but provides explanations solely on the language side, making it challenging to establish a direct relationship between crucial visual features and language descriptions. In contrast, PEMP introduces typical patch and slide examples related to prognosis on the visual side, along with straightforward language descriptions on the language side. Through AC-", + "bbox": [ + 212, + 597, + 787, + 840 + ], + "page_idx": 11 + }, + { + "type": "page_number", + "text": "12", + "bbox": [ + 217, + 114, + 235, + 126 + ], + "page_idx": 11 + }, + { + "type": "header", + "text": "L. Qu et al.", + "bbox": [ + 271, + 114, + 351, + 128 + ], + "page_idx": 11 + }, + { + "type": "table", + "img_path": "images/1d89706e47051ede058dc8bb2c110bc49acce3e95b6a9a56585bfde6f1fe70e0.jpg", + "table_caption": [ + "Table 4: Ablation studies of the modules." + ], + "table_footnote": [], + "table_body": "
Method32-shot16-shot8-shot4-shot2-shot
PEMP w/o v&t em.0.6410.5940.5610.5170.490
PEMP vision only0.6490.6070.5680.5430.527
PEMP w/o vision em.0.6550.6130.5770.5400.511
PEMP w/o text em.0.6580.6190.5750.5570.533
PEMP w/o Summary Layer0.6320.5870.5580.5110.487
PEMP w/o Messenger Layer0.6640.6290.5940.5750.554
PEMP w/o Slide-level Prompts0.6560.6200.5810.5520.525
PEMP w/o AC-Loss0.6600.6260.5890.5680.549
PEMP (ours)0.6670.6370.6140.5870.562
", + "bbox": [ + 330, + 170, + 671, + 262 + ], + "page_idx": 12 + }, + { + "type": "text", + "text": "Loss, the model efficiently completes the task under the guidance of pathology prior knowledge. Additionally, PEMP excels when the sample size is very small (e.g., 2-shot, 4-shot), showcasing its effective utilization of visual and language prior knowledge related to prognosis in H&E pathology slides for diagnosis.", + "bbox": [ + 212, + 291, + 782, + 353 + ], + "page_idx": 12 + }, + { + "type": "text", + "text": "Lymph node metastasis prediction. Table 2 presents the performance of all methods in the in-house and TCGA-CESC datasets. Across all settings with limited samples, PEMP consistently achieves the best performance and significantly outperforms all comparative methods. This indicates that PEMP is also capable of efficiently understanding and leveraging pathological prior knowledge related to metastasis in primary lesion H&E pathology slides for diagnosis. The results of both tasks highlight the immense potential of PEMP in clinical tasks characterized by challenges in label acquisition and diagnostic complexity.", + "bbox": [ + 212, + 354, + 785, + 474 + ], + "page_idx": 12 + }, + { + "type": "text", + "text": "Round cell subtype diagnosis. The classification diagnosis for rare diseases in the round cell subtype task presents an increasingly challenging yet vital clinical issue. Table 3 illustrates the performance of all methods. Notably, with a limited number of shots, most methods face challenges in achieving satisfactory diagnostic results. However, leveraging visual and pathological expertise from pathology experts, PEMP achieves significantly better results for all the settings (6.2% higher AUC on average). This underscores the substantial potential of PEMP in the diagnosis of rare diseases.", + "bbox": [ + 212, + 476, + 787, + 595 + ], + "page_idx": 12 + }, + { + "type": "text", + "text": "Ablation Study. As shown in Table 4, we designed eight variants and evaluated the roles of each component in PEMP in the prognosis task. (1) PEMP w/o v&t em. represents the removal of all vision and text examples, reducing PEMP to CoOp; (2) PEMP vision only denotes the exclusion of the text end, training solely with vision examples as guidance for prior knowledge; (3) PEMP w/o vision em. indicates the removal of all vision examples; (4) PEMP w/o text em. signifies the exclusion of all text examples; (5) PEMP w/o Summary Layer denotes the replacement from Summary Layer to simple average pooling. Please note that the Summary Layer is essentially an Attention Pooling layer, recognized as a crucial component in MIL problems; (6) PEMP w/o Messenger Layer denotes the removal of the Messenger Layer; (7) PEMP w/o Slide-level Prompt denotes the removal of all slide-level prompts; (8) PEMP w/o AC-Loss denotes the removal of AC-Loss for vision and text examples while retaining the AC-Loss for the visual and text features of the slide task token.", + "bbox": [ + 212, + 597, + 787, + 808 + ], + "page_idx": 12 + }, + { + "type": "text", + "text": "Visualization and Interpretability. To illustrate the robust performance of PEMP, we briefly visualize typical cases where visual patch examples and", + "bbox": [ + 212, + 809, + 785, + 840 + ], + "page_idx": 12 + }, + { + "type": "header", + "text": "Pathology-knowledge Enhanced Multi-instance Prompt Learning", + "bbox": [ + 300, + 114, + 730, + 130 + ], + "page_idx": 12 + }, + { + "type": "page_number", + "text": "13", + "bbox": [ + 767, + 114, + 785, + 127 + ], + "page_idx": 12 + }, + { + "type": "image", + "img_path": "images/cb2d44ba90b186e9cf6faf3a1be08efa2a48624c996022d2b57abe93e26c43d5.jpg", + "image_caption": [ + "Fig. 4: Visualization of key pathology patterns indicating both good and poor prognosis retrieved by PEMP from the test set." + ], + "image_footnote": [], + "bbox": [ + 302, + 142, + 702, + 342 + ], + "page_idx": 13 + }, + { + "type": "text", + "text": "slide examples exhibit the highest similarity on the prognosis task's test dataset (Fig. 4). PEMP demonstrates accurate learning of pathology prior knowledge, effectively addressing the challenge of predicting survival outcomes with limited training WSIs. More importantly, PEMP also shows high interpretability. PEMP efficiently integrates the high-risk and low-risk pathology image patterns and textual descriptions that pathologists have summarized over an extended period. PEMP achieves excellent performance under the guidance and facilitation of this prior knowledge.", + "bbox": [ + 212, + 409, + 787, + 532 + ], + "page_idx": 13 + }, + { + "type": "text", + "text": "6 Conclusion", + "text_level": 1, + "bbox": [ + 215, + 551, + 359, + 568 + ], + "page_idx": 13 + }, + { + "type": "text", + "text": "This paper presents PEMP, a pathology-knowledge enhanced multi-instance prompt learning framework, excelling in challenging tasks such as prognosis prediction with difficult label acquisition and rare tumor classification in Few-shot Weakly Supervised WSI Classification. PEMP integrates vision and text prior knowledge into prompts at both patch and slide levels, using a combination of static and learnable prompts during training to guide the effective diagnosis of key pathology patterns. Alignment-wise contrastive losses ensure feature-level alignment between knowledge-enhanced visual and textual prompts at both patch and slide levels. In specific clinical settings, collecting a large number of WSIs may be challenging due to patient privacy issues, the prevalence of rare diseases, new treatment protocols, or small patient groups. PEMP has shown substantial benefits across five datasets, including prognosis and rare diseases, demonstrating significant diagnostic benefits in fields with limited medical resources. PEMP also demonstrates high interpretability in medical diagnoses, indicating significant potential for future clinical applications. Moreover, such prompt-description-based model adaptation could be widely adopted as a general learning paradigm for specialized domain applications.", + "bbox": [ + 212, + 583, + 787, + 843 + ], + "page_idx": 13 + }, + { + "type": "page_number", + "text": "14", + "bbox": [ + 217, + 114, + 235, + 126 + ], + "page_idx": 13 + }, + { + "type": "header", + "text": "L. Qu et al.", + "bbox": [ + 271, + 114, + 351, + 128 + ], + "page_idx": 13 + }, + { + "type": "text", + "text": "Acknowledgements", + "text_level": 1, + "bbox": [ + 217, + 143, + 401, + 162 + ], + "page_idx": 14 + }, + { + "type": "text", + "text": "This work is funded by the National Key R&D Program of China (2022ZD0160700) and Shanghai AI Laboratory.", + "bbox": [ + 215, + 178, + 795, + 209 + ], + "page_idx": 14 + }, + { + "type": "text", + "text": "References", + "text_level": 1, + "bbox": [ + 217, + 234, + 321, + 250 + ], + "page_idx": 14 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "1. Alayrac, J.B., Donahue, J., Luc, P., Miech, A., Barr, I., Hasson, Y., Lenc, K., Mensch, A., Millican, K., Reynolds, M., et al.: Flamingo: a visual language model for few-shot learning. Advances in Neural Information Processing Systems (NeurIPS) 35, 23716-23736 (2022)", + "2. Campanella, G., Hanna, M.G., Geneslaw, L., Miraflor, A., Werneck Krauss Silva, V., Busam, K.J., Brogi, E., Reuter, V.E., Klimstra, D.S., Fuchs, T.J.: Clinical-grade computational pathology using weakly supervised deep learning on whole slide images. Nature Medicine 25(8), 1301-1309 (2019)", + "3. Chan, T.H., Cendra, F.J., Ma, L., Yin, G., Yu, L.: Histopathology whole slide image analysis with heterogeneous graph representation learning. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR). pp. 15661-15670 (2023)", + "4. Chen, R.J., Chen, C., Li, Y., Chen, T.Y., Trister, A.D., Krishnan, R.G., Mahmood, F.: Scaling vision transformers to gigapixel images via hierarchical self-supervised learning. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR). pp. 16144-16155 (2022)", + "5. Chen, R.J., Lu, M.Y., Wang, J., Williamson, D.F., Rodig, S.J., Lindeman, N.I., Mahmood, F.: Pathomic fusion: an integrated framework for fusing histopathology and genomic features for cancer diagnosis and prognosis. IEEE Transactions on Medical Imaging 41(4), 757-770 (2020)", + "6. Chen, R.J., Lu, M.Y., Weng, W.H., Chen, T.Y., Williamson, D.F., Manz, T., Shady, M., Mahmood, F.: Multimodal co-attention transformer for survival prediction in gigapixel whole slide images. In: Proceedings of the IEEE/CVF International Conference on Computer Vision (ICCV). pp. 4015-4025 (2021)", + "7. Chen, W., Si, C., Zhang, Z., Wang, L., Wang, Z., Tan, T.: Semantic prompt for few-shot image recognition. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR). pp. 23581-23591 (2023)", + "8. Chen, Y.C., Lu, C.S.: Rankmix: Data augmentation for weakly supervised learning of classifying whole slide images with diverse sizes and imbalanced categories. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR). pp. 23936-23945 (2023)", + "9. Cheplygina, V., de Bruijne, M., Pluim, J.P.: Not-so-supervised: a survey of semi-supervised, multi-instance, and transfer learning in medical image analysis. Medical Image Analysis 54, 280-296 (2019)", + "10. Chikontwe, P., Kim, M., Nam, S.J., Go, H., Park, S.H.: Multiple instance learning with center embeddings for histopathology classification. In: Medical Image Computing and Computer Assisted Intervention (MICCAI). pp. 519-528. Springer (2020)", + "11. Gu, J., Han, Z., Chen, S., Beirami, A., He, B., Zhang, G., Liao, R., Qin, Y., Tresp, V., Torr, P.: A systematic survey of prompt engineering on vision-language foundation models. arXiv preprint arXiv:2307.12980 (2023)" + ], + "bbox": [ + 218, + 267, + 785, + 839 + ], + "page_idx": 14 + }, + { + "type": "header", + "text": "Pathology-knowledge Enhanced Multi-instance Prompt Learning", + "bbox": [ + 300, + 114, + 730, + 128 + ], + "page_idx": 14 + }, + { + "type": "page_number", + "text": "15", + "bbox": [ + 767, + 116, + 784, + 126 + ], + "page_idx": 14 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "12. Hashimoto, N., Fukushima, D., Koga, R., Takagi, Y., Ko, K., Kohno, K., Nakaguro, M., Nakamura, S., Hontani, H., Takeuchi, I.: Multi-scale domain-adversarial multiple-instance cnn for cancer subtype classification with unannotated histopathological images. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR). pp. 3852-3861 (2020)", + "13. Huang, Y., Zhao, W., Wang, S., Fu, Y., Jiang, Y., Yu, L.: Conslide: Asynchronous hierarchical interaction transformer with breakup-reorganize rehearsal for continual whole slide image analysis. In: Proceedings of the IEEE/CVF International Conference on Computer Vision (ICCV). pp. 21349-21360 (2023)", + "14. Huang, Z., Bianchi, F., Yuksekgonul, M., Montine, T.J., Zou, J.: A visual-language foundation model for pathology image analysis using medical twitter. Nature Medicine pp. 1-10 (2023)", + "15. Ikezogwo, W., Seyfioglu, S., Ghezloo, F., Geva, D., Sheikh Mohammed, F., Anand, P.K., Krishna, R., Shapiro, L.: Quilt-1m: One million image-text pairs for histopathology. Advances in Neural Information Processing Systems (NeurIPS) 36 (2024)", + "16. Ilse, M., Tomczak, J., Welling, M.: Attention-based deep multiple instance learning. In: International Conference on Machine Learning (ICML). pp. 2127-2136. PMLR (2018)", + "17. Jia, M., Tang, L., Chen, B.C., Cardie, C., Belongie, S., Hariharan, B., Lim, S.N.: Visual prompt tuning. In: European Conference on Computer Vision (ECCV). pp. 709-727. Springer (2022)", + "18. Li, B., Li, Y., Eliceiri, K.W.: Dual-stream multiple instance learning network for whole slide image classification with self-supervised contrastive learning. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR). pp. 14318-14328 (2021)", + "19. Li, H., Yang, F., Zhao, Y., Xing, X., Zhang, J., Gao, M., Huang, J., Wang, L., Yao, J.: Dt-mil: deformable transformer for multi-instance learning on histopathological image. In: Medical Image Computing and Computer Assisted Intervention (MICCAI). pp. 206-216. Springer (2021)", + "20. Li, H., Zhu, C., Zhang, Y., Sun, Y., Shui, Z., Kuang, W., Zheng, S., Yang, L.: Task-specific fine-tuning via variational information bottleneck for weakly-supervised pathology whole slide image classification. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR). pp. 7454-7463 (2023)", + "21. Li, J., Li, D., Xiong, C., Hoi, S.: Blip: Bootstrapping language-image pre-training for unified vision-language understanding and generation. In: International Conference on Machine Learning (ICML). pp. 12888-12900. PMLR (2022)", + "22. Lin, T., Xu, H., Yang, C., Xu, Y.: Interventional multi-instance learning with deconfounded instance-level prediction. In: Proceedings of the AAAI Conference on Artificial Intelligence (AAAI). vol. 36, pp. 1601-1609 (2022)", + "23. Lin, T., Yu, Z., Hu, H., Xu, Y., Chen, C.W.: Interventional bag multi-instance learning on whole-slide pathological images. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR). pp. 19830-19839 (2023)", + "24. Lu, M.Y., Chen, B., Zhang, A., Williamson, D.F., Chen, R.J., Ding, T., Le, L.P., Chuang, Y.S., Mahmood, F.: Visual language pretrained multiple instance zero-shot transfer for histopathology images. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR). pp. 19764-19775 (2023)" + ], + "bbox": [ + 215, + 146, + 785, + 840 + ], + "page_idx": 15 + }, + { + "type": "page_number", + "text": "16", + "bbox": [ + 217, + 114, + 235, + 126 + ], + "page_idx": 15 + }, + { + "type": "header", + "text": "L. Qu et al.", + "bbox": [ + 271, + 114, + 351, + 128 + ], + "page_idx": 15 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "25. Lu, M.Y., Williamson, D.F., Chen, T.Y., Chen, R.J., Barbieri, M., Mahmood, F.: Data-efficient and weakly supervised computational pathology on whole-slide images. Nature Biomedical Engineering 5(6), 555-570 (2021)", + "26. Qu, L., Liu, S., Liu, X., Wang, M., Song, Z.: Towards label-efficient automatic diagnosis and analysis: a comprehensive survey of advanced deep learning-based weakly-supervised, semi-supervised and self-supervised techniques in histopathological image analysis. Physics in Medicine & Biology (2022)", + "27. Qu, L., Luo, X., Fu, K., Wang, M., Song, Z.: The rise of ai language pathologists: Exploring two-level prompt learning for few-shot weakly-supervised whole slide image classification. arXiv preprint arXiv:2305.17891 (2023)", + "28. Qu, L., Luo, X., Liu, S., Wang, M., Song, Z.: Dgmil: Distribution guided multiple instance learning for whole slide image classification. In: Medical Image Computing and Computer Assisted Intervention (MICCAI). pp. 24-34. Springer (2022)", + "29. Radford, A., Kim, J.W., Hallacy, C., Ramesh, A., Goh, G., Agarwal, S., Sastry, G., Askell, A., Mishkin, P., Clark, J., et al.: Learning transferable visual models from natural language supervision. In: International Conference on Machine Learning (ICML). pp. 8748-8763. PMLR (2021)", + "30. Rony, J., Belharbi, S., Dolz, J., Ayed, I.B., McCaffrey, L., Granger, E.: Deep weakly-supervised learning methods for classification and localization in histology images: a survey. arXiv preprint arXiv:1909.03354 (2019)", + "31. Shao, Z., Bian, H., Chen, Y., Wang, Y., Zhang, J., Ji, X., et al.: Transmil: Transformer based correlated multiple instance learning for whole slide image classification. Advances in Neural Information Processing Systems (NeurIPS) 34, 2136-2147 (2021)", + "32. Shi, X., Xing, F., Xie, Y., Zhang, Z., Cui, L., Yang, L.: Loss-based attention for deep multiple instance learning. In: Proceedings of the AAAI Conference on Artificial Intelligence (AAAI). vol. 34, pp. 5742-5749 (2020)", + "33. Song, A.H., Jaume, G., Williamson, D.F., Lu, M.Y., Vaidya, A., Miller, T.R., Mahmood, F.: Artificial intelligence for digital and computational pathology. Nature Reviews Bioengineering pp. 1-20 (2023)", + "34. Tu, C., Zhang, Y., Ning, Z.: Dual-curriculum contrastive multi-instance learning for cancer prognosis analysis with whole slide images. Advances in Neural Information Processing Systems (NeurIPS) 35, 29484-29497 (2022)", + "35. Wang, X., Xiang, J., Zhang, J., Yang, S., Yang, Z., Wang, M.H., Zhang, J., Yang, W., Huang, J., Han, X.: Scl-wc: Cross-slide contrastive learning for weakly-supervised whole-slide image classification. Advances in Neural Information Processing Systems (NeurIPS) 35, 18009-18021 (2022)", + "36. Wasim, S.T., Naseer, M., Khan, S., Khan, F.S., Shah, M.: Vita-clip: Video and text adaptive clip via multimodal prompting. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR). pp. 23034-23044 (2023)", + "37. Xu, G., Song, Z., Sun, Z., Ku, C., Yang, Z., Liu, C., Wang, S., Ma, J., Xu, W.: Camel: A weakly supervised learning framework for histopathology image segmentation. In: Proceedings of the IEEE/CVF International Conference on Computer Vision (ICCV). pp. 10682-10691 (2019)", + "38. Yao, H., Zhang, R., Xu, C.: Visual-language prompt tuning with knowledge-guided context optimization. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR). pp. 6757-6767 (2023)", + "39. Yao, J., Zhu, X., Jonnagaddala, J., Hawkins, N., Huang, J.: Whole slide images based cancer survival prediction using attention guided deep multiple instance learning networks. Medical Image Analysis 65, 101789 (2020)" + ], + "bbox": [ + 212, + 146, + 787, + 839 + ], + "page_idx": 16 + }, + { + "type": "header", + "text": "Pathology-knowledge Enhanced Multi-instance Prompt Learning", + "bbox": [ + 300, + 114, + 730, + 128 + ], + "page_idx": 16 + }, + { + "type": "page_number", + "text": "17", + "bbox": [ + 767, + 116, + 784, + 126 + ], + "page_idx": 16 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "40. Zhang, H., Meng, Y., Zhao, Y., Qiao, Y., Yang, X., Coupland, S.E., Zheng, Y.: Dtdf-mil: Double-tier feature distillation multiple instance learning for histopathology whole slide image classification. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR). pp. 18802-18812 (2022)", + "41. Zhang, Y., Gao, J., Zhou, M., Wang, X., Qiao, Y., Zhang, S., Wang, D.: Text-guided foundation model adaptation for pathological image classification. In: International Conference on Medical Image Computing and Computer-Assisted Intervention (MICCAI). pp. 272-282. Springer (2023)", + "42. Zhou, K., Yang, J., Loy, C.C., Liu, Z.: Learning to prompt for vision-language models. International Journal of Computer Vision 130(9), 2337-2348 (2022)", + "43. Zhu, X., Yao, J., Zhu, F., Huang, J.: Wsisa: Making survival prediction from whole slide histopathological images. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR). pp. 7234-7242 (2017)" + ], + "bbox": [ + 215, + 146, + 785, + 327 + ], + "page_idx": 17 + }, + { + "type": "page_number", + "text": "18", + "bbox": [ + 217, + 114, + 235, + 126 + ], + "page_idx": 17 + }, + { + "type": "header", + "text": "L. Qu et al.", + "bbox": [ + 271, + 114, + 351, + 128 + ], + "page_idx": 17 + } +] \ No newline at end of file diff --git a/2024/Pathology-knowledge Enhanced Multi-instance Prompt Learning for Few-shot Whole Slide Image Classification/3e473ac8-930f-46c0-b97b-e8e52a7c0bfd_model.json b/2024/Pathology-knowledge Enhanced Multi-instance Prompt Learning for Few-shot Whole Slide Image Classification/3e473ac8-930f-46c0-b97b-e8e52a7c0bfd_model.json new file mode 100644 index 0000000000000000000000000000000000000000..987d283181e42d8db40544fca381a1077b7b91c3 --- /dev/null +++ b/2024/Pathology-knowledge Enhanced Multi-instance Prompt Learning for Few-shot Whole Slide Image Classification/3e473ac8-930f-46c0-b97b-e8e52a7c0bfd_model.json @@ -0,0 +1,2117 @@ +[ + [ + { + "type": "title", + "bbox": [ + 0.26, + 0.141, + 0.744, + 0.21 + ], + "angle": 0, + "content": "Pathology-knowledge Enhanced Multi-instance Prompt Learning for Few-shot Whole Slide Image Classification" + }, + { + "type": "text", + "bbox": [ + 0.218, + 0.234, + 0.784, + 0.274 + ], + "angle": 0, + "content": "Linhao Qu\\(^{1}\\), Dingkang Yang\\(^{2}\\), Dan Huang\\(^{3}\\), Qinhao Guo\\(^{4}\\), Rongkui Luo\\(^{5}\\), Shaoting Zhang\\(^{1}\\), and Xiaosong Wang\\(^{1\\boxtimes}\\)" + }, + { + "type": "text", + "bbox": [ + 0.293, + 0.284, + 0.71, + 0.299 + ], + "angle": 0, + "content": "1 Shanghai Artificial Intelligence Laboratory, Shanghai, China" + }, + { + "type": "text", + "bbox": [ + 0.236, + 0.299, + 0.767, + 0.312 + ], + "angle": 0, + "content": "\\(^{2}\\) Academy for Engineering and Technology, Fudan University, Shanghai, China" + }, + { + "type": "text", + "bbox": [ + 0.239, + 0.313, + 0.762, + 0.327 + ], + "angle": 0, + "content": "\\(^{3}\\) Department of Pathology, Fudan University Shanghai Cancer Center, China" + }, + { + "type": "text", + "bbox": [ + 0.225, + 0.327, + 0.778, + 0.34 + ], + "angle": 0, + "content": "\\(^{4}\\) Department of Gynecologic Oncology, Fudan University Shanghai Cancer Center" + }, + { + "type": "text", + "bbox": [ + 0.221, + 0.34, + 0.783, + 0.369 + ], + "angle": 0, + "content": "\\(^{5}\\) Department of Pathology, Zhongshan Hospital, Fudan University, Shanghai, China wangxiaosong@pjlab.org.cn;" + }, + { + "type": "text", + "bbox": [ + 0.261, + 0.399, + 0.744, + 0.732 + ], + "angle": 0, + "content": "Abstract. Current multi-instance learning algorithms for pathology image analysis often require a substantial number of Whole Slide Images for effective training but exhibit suboptimal performance in scenarios with limited learning data. In clinical settings, restricted access to pathology slides is inevitable due to patient privacy concerns and the prevalence of rare or emerging diseases. The emergence of the Few-shot Weakly Supervised WSI Classification accommodates the significant challenge of the limited slide data and sparse slide-level labels for diagnosis. Prompt learning based on the pre-trained models (e.g., CLIP) appears to be a promising scheme for this setting; however, current research in this area is limited, and existing algorithms often focus solely on patch-level prompts or confine themselves to language prompts. This paper proposes a multi-instance prompt learning framework enhanced with pathology knowledge, i.e., integrating visual and textual prior knowledge into prompts at both patch and slide levels. The training process employs a combination of static and learnable prompts, effectively guiding the activation of pre-trained models and further facilitating the diagnosis of key pathology patterns. Lightweight Messenger (self-attention) and Summary (attention-pooling) layers are introduced to model relationships between patches and slides within the same patient data. Additionally, alignment-wise contrastive losses ensure the feature-level alignment between visual and textual learnable prompts for both patches and slides. Our method demonstrates superior performance in three challenging clinical tasks, significantly outperforming comparative few-shot methods." + }, + { + "type": "text", + "bbox": [ + 0.262, + 0.744, + 0.644, + 0.758 + ], + "angle": 0, + "content": "Keywords: Pathology image analysis \\(\\cdot\\) Prompt learning" + }, + { + "type": "title", + "bbox": [ + 0.216, + 0.781, + 0.377, + 0.797 + ], + "angle": 0, + "content": "1 Introduction" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.81, + 0.788, + 0.842 + ], + "angle": 0, + "content": "Deep learning for pathology Whole Slide Image (WSI) classification offers a rapid, objective, and precise diagnostic approach, enhancing prognosis predic" + } + ], + [ + { + "type": "page_number", + "bbox": [ + 0.218, + 0.116, + 0.23, + 0.127 + ], + "angle": 0, + "content": "2" + }, + { + "type": "header", + "bbox": [ + 0.272, + 0.115, + 0.352, + 0.129 + ], + "angle": 0, + "content": "L. Qu et al." + }, + { + "type": "image", + "bbox": [ + 0.334, + 0.147, + 0.672, + 0.307 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.216, + 0.321, + 0.788, + 0.378 + ], + "angle": 0, + "content": "Fig. 1: Existing methods such as CoOp [42] do not fully consider task-related pathology visual features and its association with specific terms in FSWC. We utilize aligned task-specific image examples and language descriptions to enhance visual and textual prompt learning at both patch and slide levels." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.412, + 0.784, + 0.548 + ], + "angle": 0, + "content": "tion and treatment response assessment [3, 4, 8, 13, 20, 26]. Given the gigapixel resolution of WSIs, segmentation into multiple patches is essential for processing. However, obtaining detailed annotations for each patch is challenging. Instead, it often results in slide-level ground truth labels. Multiple Instance Learning (MIL), a prominent weakly supervised paradigm, has emerged as a key technique for WSI classification [5, 9, 25, 30]. In MIL, each WSI is treated as a bag, with patches as instances. During training, only slide-level labels are available, necessitating a comprehensive model to capture relationships between patches and slide labels for accurate classification [28]." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.551, + 0.785, + 0.717 + ], + "angle": 0, + "content": "MIL algorithms achieve slide-level classification by aggregating patch-level features or predictions. Though effective, existing algorithms typically demand a considerable number of slides for practical training, performing inadequately in scenarios with limited samples. In clinical practice, challenges such as preserving patient privacy, the rarity of certain diseases, and emerging conditions, particularly for prognosis and rare tumor subtypes, hinder the acquisition of a substantial quantity of WSIs [9,26,33]. Recently, the challenging problem of Few-shot weakly Supervised WSI Classification (FSWC) has emerged [27]. FSWC is characterized by the limited availability of training WSIs (e.g., typically 2, 4, 8, 16, or 32 per class) and sparse slide-level annotations, with unknown fine-grained annotations for numerous patches in each WSI." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.72, + 0.787, + 0.84 + ], + "angle": 0, + "content": "In image classification, recent advancements highlight the effectiveness of Vision-Language Models (VLMs) like CLIP [29], BLIP [21], and Flamingo [1]. These models undergo training on large-scale language-image pairs using alignment contrastive learning, demonstrating notable zero-shot and few-shot transfer learning capabilities. In the pathology image domain, despite valuable contributions from initiatives like MI-Zero [24] and PLIP [14] in constructing VLMs for computational pathology, challenges including ethical concerns, annotation scarcity, and storage limitations make it impractical to train foundational mod-" + } + ], + [ + { + "type": "header", + "bbox": [ + 0.301, + 0.115, + 0.732, + 0.131 + ], + "angle": 0, + "content": "Pathology-knowledge Enhanced Multi-instance Prompt Learning" + }, + { + "type": "page_number", + "bbox": [ + 0.775, + 0.117, + 0.785, + 0.127 + ], + "angle": 0, + "content": "3" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.147, + 0.788, + 0.283 + ], + "angle": 0, + "content": "els for each specific downstream task [27]. Furthermore, existing VLMs exhibit restricted direct zero-shot transfer capabilities for specific downstream tasks [7,36,38]. Considering these challenges, the application of existing VLMs for developing prompt learning algorithms tailored to few-shot downstream tasks appears promising. Prompt learning involves adapting a pre-trained and then fixed VLM to new tasks by integrating task-specific prompts as additional model inputs. In the context of few-shot learning, these prompts can be either fixed or trainable, with trainable prompts enabling contrastive learning between visual and language features [11,42]." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.283, + 0.789, + 0.465 + ], + "angle": 0, + "content": "Presently, research on prompt learning for FSWC is limited [27]. Previous efforts, such as MI-Zero [24], PLIP [14], and CITE [41], have predominantly focused on the utilization of \"patch-level\" prompts—textual prompts associated with pathology regions of interest. Nonetheless, visual or textual prompts in domain-specific downstream tasks may not be understandable (often unseen) for the pre-trained model and thus may not properly activate the relevant features of the pre-trained model. The TOP framework [27] represents the initial endeavor to implement prompt learning by introducing language-based prior descriptions in the prompts. However, pathology images encompass highly specialized visual features and corresponding pathology-specific terms, posing a challenge for networks to comprehensively understand and learn pathology-related features and specialized descriptions solely through language explanations for classification." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.465, + 0.789, + 0.646 + ], + "angle": 0, + "content": "To address this issue, we propose Pathology-knowledge Enhanced Multi-instance Prompt Learning for FSWC named PEMP, with a concise architecture depicted in Fig. 1. PEMP is a multi-instance prompt learning framework specifically designed for the FSWC problem, incorporating both visual and textual samples (as a form of prior knowledge) at both patch and slide levels. Drawing inspiration from pathologists learning pathology images and language knowledge from textbooks, we introduce task-specific image examples and corresponding language descriptions, injected in both the visual and textual prompts. During prompt learning, the model is fed with a more understandable and effective guide (in image examples and plain words) and then activates more meaningful features for the diagnosis of key patch-level and slide-level pathology patterns in scenarios with limited training samples." + }, + { + "type": "text", + "bbox": [ + 0.24, + 0.647, + 0.545, + 0.661 + ], + "angle": 0, + "content": "The contribution of this work is threefold:" + }, + { + "type": "text", + "bbox": [ + 0.226, + 0.673, + 0.787, + 0.749 + ], + "angle": 0, + "content": "- We propose a multi-instance prompt learning framework, PEMP, introducing vision and text prior knowledge in the designed prompts at both patch and slide levels. A combination of static and learnable prompts is employed in the training, effectively adapting the pre-trained model to focus on key pathology patterns in patches and slides." + }, + { + "type": "text", + "bbox": [ + 0.226, + 0.749, + 0.787, + 0.81 + ], + "angle": 0, + "content": "- We explore the alignment of visual and textual prompts using contrastive losses at both patch and slide levels. We design lightweight Messenger (self-attention) and Summary (attention-pooling) layers to model relationships of patch-to-patch and patch-to-slide within the same patient data." + }, + { + "type": "text", + "bbox": [ + 0.226, + 0.81, + 0.787, + 0.842 + ], + "angle": 0, + "content": "- We evaluate PEMP's performance in addressing three challenging pathology tasks in clinical practice, including predicting patient survival and lymph" + }, + { + "type": "list", + "bbox": [ + 0.226, + 0.673, + 0.787, + 0.842 + ], + "angle": 0, + "content": null + } + ], + [ + { + "type": "page_number", + "bbox": [ + 0.218, + 0.116, + 0.23, + 0.127 + ], + "angle": 0, + "content": "4" + }, + { + "type": "header", + "bbox": [ + 0.272, + 0.115, + 0.352, + 0.129 + ], + "angle": 0, + "content": "L. Qu et al." + }, + { + "type": "text", + "bbox": [ + 0.243, + 0.147, + 0.788, + 0.209 + ], + "angle": 0, + "content": "node metastasis in early cervical cancer H&E slides and subtype classification for a rare tumor (round cell tumor). PEMP achieves superior performance in all these tasks under the few-shot setting, notably outperforming all comparative methods by a large margin (4% on average)." + }, + { + "type": "title", + "bbox": [ + 0.216, + 0.232, + 0.39, + 0.248 + ], + "angle": 0, + "content": "2 Related Work" + }, + { + "type": "title", + "bbox": [ + 0.216, + 0.265, + 0.486, + 0.28 + ], + "angle": 0, + "content": "2.1 MIL for WSI Classification" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.292, + 0.789, + 0.459 + ], + "angle": 0, + "content": "Current MIL algorithms for WSI classification can be broadly categorized as instance-based [2, 10, 22, 28, 37] and bag-based [6, 12, 16, 18, 19, 23, 25, 31, 32, 34, 35, 39, 40, 43] methods. Instance-based methods involve generating pseudo-labels for each patch and training patch classifiers to score all patches within a slide, whereas bag-based methods, the current mainstream, extract patch features and employ techniques like attention weighting [12, 16, 18, 25, 32, 39, 40, 43] to aggregate these features into slide-level features. While existing algorithms perform well with ample training data, they lack specificity for FSWC tasks, struggling with limited samples. Our proposed PEMP, falling under the bag-based MIL paradigm, distinguishes itself by innovatively addressing FSWC through VLMs and prompt learning, deviating from traditional bag-based MIL approaches." + }, + { + "type": "title", + "bbox": [ + 0.216, + 0.481, + 0.512, + 0.497 + ], + "angle": 0, + "content": "2.2 VLM-Based Prompt Learning" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.508, + 0.789, + 0.659 + ], + "angle": 0, + "content": "VLMs, comprising image and text encoders, are trained on large-scale image-text pairs using alignment contrastive learning to establish a shared feature space. Pre-trained VLMs like CLIP [29], BLIP [21], and Flamingo [1] exhibit robust zero-shot and few-shot transferability in image recognition tasks. However, limited accuracy in zero-shot transfer has fueled interest in few-shot learning based on VLMs, where the prompt learning paradigm, demonstrating strong performance [7,36,38], is emerging. Prompt learning leverages a pre-trained VLM in the adaptation to new downstream tasks by incorporating task-specific prompts through enriched model inputs. These prompts could be static with manual designs or learnable vectors [11,42]." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.66, + 0.789, + 0.841 + ], + "angle": 0, + "content": "In pathology, recent research emphasizes constructing specialized VLMs, such as MI-Zero [24], PLIP [14], and QUILT [15]. However, privacy concerns, annotation scarcity, and storage demands make training separate foundational models for each downstream task impractical. Existing VLMs exhibit limited zero-shot transferability, especially for rare diseases. Unlike fine-tuning VLMs at the data level or constructing large datasets like these works, we introduce a novel prompt learning paradigm based on VLMs (e.g., CLIP with frozen parameters) to address the FSWC problem. Importantly, our scheme is versatile across various foundation models. Moreover, MI-Zero, PLIP and QUILT mainly contribute by collecting large datasets to construct pathology VLMs but lack effective methods for FSWC. Their simple adapting strategies for FSWC (e.g., LinearProbe and Attention Pooling) limit their performance in challenging clinical tasks." + } + ], + [ + { + "type": "header", + "bbox": [ + 0.301, + 0.115, + 0.733, + 0.131 + ], + "angle": 0, + "content": "Pathology-knowledge Enhanced Multi-instance Prompt Learning" + }, + { + "type": "page_number", + "bbox": [ + 0.775, + 0.117, + 0.785, + 0.127 + ], + "angle": 0, + "content": "5" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.147, + 0.788, + 0.284 + ], + "angle": 0, + "content": "In pathology, prompt learning research for FSWC is limited. Prior works like MI-Zero [24], PLIP [14], and CITE [41] only explore \"patch-level\" prompts, limiting slide-level classification performance. TOP [27] introduces prompt learning at both patch and slide levels but lacks task-specific visual knowledge on the vision side. Relying solely on language explanations is challenging for the networks to understand pathology features with limited training samples. In contrast, PEMP's learning process operates at patch and slide levels, covers visual and textual prompt design, and introduces task-specific image examples and descriptions separately as prior knowledge." + }, + { + "type": "title", + "bbox": [ + 0.215, + 0.305, + 0.52, + 0.321 + ], + "angle": 0, + "content": "3 FSWC Baseline Framework" + }, + { + "type": "title", + "bbox": [ + 0.215, + 0.335, + 0.438, + 0.349 + ], + "angle": 0, + "content": "3.1 Problem Formulation" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.359, + 0.789, + 0.556 + ], + "angle": 0, + "content": "We provide a brief introduction to the problem definition of FSWC and the pre-trained VLM used in our study. Given a dataset \\( W = \\{W_{1}, W_{2}, \\ldots, W_{N}\\} \\) containing \\( N \\) WSIs, each WSI \\( W_{i} \\) is divided into non-overlapping small image patches \\( \\{p_{i,j}, j = 1, 2, \\ldots, n_{i}\\} \\), where \\( n_{i} \\) represents the number of patches obtained from \\( W_{i} \\). It is important to note that the number of patches obtained from each slide may vary. All patches in \\( W_{i} \\) collectively form a bag, and each patch serves as an instance of this bag. Each bag is assigned a label \\( Y_{i} \\in \\{0, 1, \\ldots, U\\} \\), while the label of each instance is unknown, and \\( U \\) is the number of classes. FSWC is highly challenging as it allows training with only a limited number of slides (with labels). In the FSWC task, \"shot\" refers to the number of WSIs, e.g., \\( K \\) pairs of WSIs and labels for training in a \\( K \\)-shot setting, while the trained model is evaluated on a completely independent test set. Typically, \\( K \\) can take values such as 2, 4, 8, 16, or 32." + }, + { + "type": "title", + "bbox": [ + 0.215, + 0.577, + 0.636, + 0.592 + ], + "angle": 0, + "content": "3.2 Attention Aggregation for WSI Classification" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.601, + 0.788, + 0.678 + ], + "angle": 0, + "content": "The attention-based pooling strategy [16] is a classical MIL aggregation method used for WSI classification. Initially, an encoder \\(\\phi_{img}\\) is used to extract features \\(f_{i,j}\\) for all patches \\(\\{p_{i,j}, j = 1,2,\\dots n_i\\}\\) in slide \\(W_{i}\\), and then aggregate these patch features using an attention-based trainable aggregation function \\(A(\\cdot)\\) to obtain the bag feature \\(F_{i}\\):" + }, + { + "type": "equation", + "bbox": [ + 0.327, + 0.687, + 0.787, + 0.728 + ], + "angle": 0, + "content": "\\[\nF _ {i} = A \\left(\\phi_ {i m g} \\left(p _ {i, 1}\\right), \\phi_ {i m g} \\left(p _ {i, 2}\\right), \\dots\\right) = \\sum_ {j = 1} ^ {n _ {i}} a _ {i, j} f _ {i, j}, \\tag {1}\n\\]" + }, + { + "type": "equation", + "bbox": [ + 0.373, + 0.738, + 0.786, + 0.777 + ], + "angle": 0, + "content": "\\[\na _ {i, j} = \\frac {\\exp \\left\\{w ^ {\\top} \\tanh \\left(V f _ {i , j} ^ {\\top}\\right) \\right\\}}{\\sum_ {j = 1} ^ {n _ {i}} \\exp \\left\\{w ^ {\\top} \\tanh \\left(V f _ {i , j} ^ {\\top}\\right) \\right\\}}, \\tag {2}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.78, + 0.788, + 0.842 + ], + "angle": 0, + "content": "where \\( a_{i,j} \\) is the attention score predicted by the self-attention network, which is parameterized by \\( w \\) and \\( V \\). The weights \\( a_{i,j} \\) reflect the contribution of each patch in making the slide-level prediction. The patch-level encoder \\( \\phi_{img} \\) is from the pre-trained CLIP [29] and remains fixed during the training process." + } + ], + [ + { + "type": "page_number", + "bbox": [ + 0.218, + 0.116, + 0.23, + 0.127 + ], + "angle": 0, + "content": "6" + }, + { + "type": "header", + "bbox": [ + 0.272, + 0.115, + 0.352, + 0.129 + ], + "angle": 0, + "content": "L. Qu et al." + }, + { + "type": "title", + "bbox": [ + 0.216, + 0.147, + 0.654, + 0.163 + ], + "angle": 0, + "content": "3.3 VLM backbone and Few-shot Prompt Learning" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.17, + 0.784, + 0.229 + ], + "angle": 0, + "content": "We utilize CLIP [29] as the backbone of the proposed framework and also the pretrained VLM for feature extraction, which is trained on a massive dataset of 400 million image-text pairs, including medical data, facilitating efficient knowledge transfer of diverse visual concepts to downstream tasks." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.231, + 0.785, + 0.414 + ], + "angle": 0, + "content": "Prompt learning enables the efficient adaptation of a fixed pre-trained CLIP in few-shot downstream classification tasks by introducing task-specific prompts into the model input. These prompts, whether fixed or learnable, can be applied to the text encoder, visual encoder, or both simultaneously. In the few-shot setting, the training could also involve aligning visual and text features through alignment contrastive learning (AC-Loss) in the specific domain. Formally, we denote the visual encoder as \\(\\phi_{img}\\) and the text encoder as \\(\\phi_{text}\\). For a downstream classification task with \\(U\\) classes, a commonly used fixed hand-crafted prompt, denoted as [Class Prompt], guides classification on the text side. Specifically, assuming the name of the \\(i\\)-th class is “[class-name],\" the corresponding hand-crafted prompt is typically designed as “a photo of a [class-name].” These [Class Prompts] are pre-encoded into a vector \\(C = \\{c_i\\}_{i=1}^U\\)." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.414, + 0.785, + 0.549 + ], + "angle": 0, + "content": "To further enhance generalization to downstream tasks, it is also popular to explore learning a set of continuous context vectors as learnable prompts, denoted as \\(\\mathbb{V} = \\{v_1,v_2,\\ldots ,v_M\\}\\). These learnable context vectors \\(\\mathbb{V}\\), concatenated with the vectors from [Class Prompt] \\(C\\), form the complete text prompt \\(t_i = \\{v_1,v_2,\\dots ,v_M,c_i\\}\\). This prompt is then fed into the text encoder \\(\\phi_{text}\\) to obtain the final textual features for each class, \\(T_{i} = \\phi_{text}(t_{i})\\). On the visual side, given an image \\(W_{i}\\) and its label \\(y\\), its slide features \\(F_{i}\\) is computed via Eq. (1). Similar to the text side, learnable prompts could also be appended, which will be inputted into the visual encoder \\(\\phi_{img}\\)." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.55, + 0.787, + 0.625 + ], + "angle": 0, + "content": "With a limited number of image-text pairs, the optimization involves minimizing the negative log-likelihood between the visual slide feature \\( F_{i} \\) and its class text feature \\( T_{y} \\), updating all learnable parameters. During testing, the extracted visual features \\( F_{i} \\) are used to retrieve the class text features \\( T_{y} \\) to obtain the probabilities for each class." + }, + { + "type": "equation", + "bbox": [ + 0.367, + 0.634, + 0.786, + 0.671 + ], + "angle": 0, + "content": "\\[\np (y \\mid F _ {i}) = \\frac {\\exp (\\operatorname {s i m} \\left(F _ {i} , T _ {y}\\right) / \\tau)}{\\sum_ {i = 1} ^ {U} \\exp (\\operatorname {s i m} \\left(F _ {i} , T _ {i}\\right) / \\tau)}, \\tag {3}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.676, + 0.788, + 0.738 + ], + "angle": 0, + "content": "where, \\( y \\) represents the ground truth class, \\( U \\) is the total number of classes, \\( \\mathrm{sim}(\\cdot) \\) denotes cosine similarity, and \\( \\tau \\) is a learnable temperature parameter. It is crucial to note that during the training process, the parameters of CLIP's image encoder and text encoder remain fixed and unchanged." + }, + { + "type": "title", + "bbox": [ + 0.216, + 0.757, + 0.631, + 0.777 + ], + "angle": 0, + "content": "4 Knowledge-enhanced Prompt Learning" + }, + { + "type": "title", + "bbox": [ + 0.216, + 0.787, + 0.341, + 0.802 + ], + "angle": 0, + "content": "4.1 Overview" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.81, + 0.787, + 0.842 + ], + "angle": 0, + "content": "Fig. 2 illustrates the primary workflow of our proposed PEMP, which comprises three learning processes: Visual prompt learning, Textual prompt learning, and" + } + ], + [ + { + "type": "header", + "bbox": [ + 0.301, + 0.115, + 0.733, + 0.13 + ], + "angle": 0, + "content": "Pathology-knowledge Enhanced Multi-instance Prompt Learning" + }, + { + "type": "page_number", + "bbox": [ + 0.775, + 0.117, + 0.785, + 0.127 + ], + "angle": 0, + "content": "7" + }, + { + "type": "image", + "bbox": [ + 0.225, + 0.144, + 0.784, + 0.445 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.216, + 0.457, + 0.788, + 0.486 + ], + "angle": 0, + "content": "Fig. 2: Overview of PEMP, where flames represent optimized parameters, and snowflakes indicate frozen parameters during training." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.527, + 0.788, + 0.693 + ], + "angle": 0, + "content": "Two-level prompt alignment via contrastive learning. In visual prompt learning, we address three key tasks. Firstly, we integrate prior visual pathology knowledge into the prompts at both patch-level and slide-level. This guides the model, particularly with limited training samples, to understand the prompt better and produce more meaningful activations for task-specific key pathology patterns. Additionally, we also include the learnable prompts on the slide level, triggering the model's precise adaptation to specific tasks. Thirdly, we introduce a lightweight Messenger (self-attention) Layer and Summary (attention-pooling) Layer. The former efficiently communicates and models features of different patches within the same slide, while the latter aggregates patch features into slide features using an attention mechanism." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.699, + 0.788, + 0.777 + ], + "angle": 0, + "content": "In a similar initiative, textual prompt learning concentrates on modeling language descriptions related to pathology tasks, also incorporating patch-level and slide-level priors. Pathologists are employed to provide straightforward language descriptions in the overall task, task-related slide-level pathology, and task-related patch-level pathology, avoiding obscure professional terms." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.78, + 0.788, + 0.842 + ], + "angle": 0, + "content": "Recognizing the challenges faced by the VLM in understanding key visual features and complex pathology terms in visual examples, we establish a close alignment between textual prompt learning (with integrated pathology language descriptions) and visual prompt learning (with injected visual examples) at both" + } + ], + [ + { + "type": "page_number", + "bbox": [ + 0.218, + 0.116, + 0.23, + 0.127 + ], + "angle": 0, + "content": "8" + }, + { + "type": "header", + "bbox": [ + 0.272, + 0.115, + 0.352, + 0.129 + ], + "angle": 0, + "content": "L. Qu et al." + }, + { + "type": "image", + "bbox": [ + 0.218, + 0.144, + 0.486, + 0.226 + ], + "angle": 0, + "content": null + }, + { + "type": "image_footnote", + "bbox": [ + 0.217, + 0.227, + 0.477, + 0.289 + ], + "angle": 0, + "content": "(a) Examples for predicting the prognosis of early cervical cancer patients from H&E pathology slides of the primary lesion, derived from authoritative textbooks and pathology experts. Top: visual patch examples. Bottom: visual slide examples." + }, + { + "type": "image", + "bbox": [ + 0.53, + 0.144, + 0.792, + 0.226 + ], + "angle": 0, + "content": null + }, + { + "type": "image_footnote", + "bbox": [ + 0.527, + 0.227, + 0.788, + 0.3 + ], + "angle": 0, + "content": "(b) Top: Construction of Patch-level descriptive [Token]. Bottom: Construction of Slide-level descriptive [Token]. Examples for the prognosis of early cervical cancer patients from H&E pathology slides of the primary lesion, serving as task-related patch-level and slide-level pathology priors on the text side." + }, + { + "type": "image_caption", + "bbox": [ + 0.214, + 0.31, + 0.788, + 0.34 + ], + "angle": 0, + "content": "Fig. 3: (A) and (B) examples of visual and textual token constructions related to the prognosis of early cervical cancer patients." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.371, + 0.786, + 0.403 + ], + "angle": 0, + "content": "slide and patch levels. This alignment enhances the organization and cohesion of knowledge-enhanced prompts on both visual and textual sides." + }, + { + "type": "title", + "bbox": [ + 0.216, + 0.431, + 0.464, + 0.447 + ], + "angle": 0, + "content": "4.2 Visual Prompt Learning" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.462, + 0.788, + 0.734 + ], + "angle": 0, + "content": "Construction of visual prompts. In the FSWC task, limited training samples present challenges in acquiring effective knowledge. Moreover, tasks like survival analysis require finding key factors both locally (e.g., patch-level: necrosis, vascular invasion) and globally (e.g., slide-level: tumor boundary, tumor stroma ratio). To overcome these, we out-source prior visual pathology knowledge at both patch and slide levels (shown in the top part of Fig. 3a), guiding the model with task-specific pathology patterns for diagnosis. For challenging slide-level diagnostic tasks, relevant patterns manifest at both levels. Leveraging expertise from pathology professionals, we construct visual patch and slide examples as fixed prompts, as shown in Fig. 3a. For each classification task, highly relevant visual patches and slides indicating high and low risks are retrieved from authoritative sources or pathology experts. For instance, in the prognosis of early cervical cancer, visual patch examples for poor prognosis include \"high-grade atypia\", \"vascular invasion,\" and \"necrosis,\" while slide examples encompass \"overall infiltration with fuzzy tumor boundaries and low tumor stroma ratio.\" Conversely, low-risk examples include \"low-grade dysplasia,\" \"no vascular invasion,\" and \"overall pushing-type with clear tumor boundaries and high tumor stroma ratio.\"" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.735, + 0.788, + 0.841 + ], + "angle": 0, + "content": "In total, 6 slides and 6 patch images are used as visual prompts, including 3 typical slides per prognosis category (good and poor), each with 3 local patches in the prognosis of early cervical cancer. For the Lymph Node Metastasis Prediction Task, we used 4 slides and 4 patches, with two representative slides per category (metastasis and non-metastasis) and two local images per slide. For Round Cell Subtype Diagnosis, 10 slides and 10 patches were used, with two slides per category, each with two local images. Selection of these visual exam" + } + ], + [ + { + "type": "header", + "bbox": [ + 0.301, + 0.115, + 0.731, + 0.131 + ], + "angle": 0, + "content": "Pathology-knowledge Enhanced Multi-instance Prompt Learning" + }, + { + "type": "page_number", + "bbox": [ + 0.775, + 0.117, + 0.785, + 0.127 + ], + "angle": 0, + "content": "9" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.147, + 0.784, + 0.177 + ], + "angle": 0, + "content": "samples was guided by clinical experts, with each risk factor represented by one slide of the patch and its local patch." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.178, + 0.788, + 0.365 + ], + "angle": 0, + "content": "Learning pipeline. As illustrated in Fig. 2, we assume \\( W_{i} \\) represents the training slide containing \\( n_i \\) patches named \\( p_{i,j} \\) and utilize the fixed CLIP image encoder \\( \\phi_{img} \\) to extract patch features \\( f_{i,j} = \\phi_{img}(p_{i,j}) \\in \\mathbb{R}^{1 \\times v} \\), where \\( v \\) represents the dimension of the features. Subsequently, the constructed visual patch examples \\( e_l \\) are also transformed into patch features \\( z_l = \\phi_{img}(e_l) \\). \\( z_l \\) remains unchanged throughout the training process and is combined with the current slide's patch features as a fixed visual prompt. Specifically, we match the cosine similarity between \\( f_{i,j} \\) and \\( z_l \\), concatenate the most matched example patch feature with the patch feature as \\( f_{i,j}^e \\), and ultimately concatenate all new patch features for the current slide \\( F_i^e = \\{f_{i,j}^e\\}_{j=1}^{n_i} \\in \\mathbb{R}^{n_i \\times v \\times 2} \\). To model all patch features within a slide effectively, we introduce a lightweight self-attention layer named Messenger Layer (ML). Taking \\( F_i^e \\) as input, it outputs \\( F_i^{ML} \\):" + }, + { + "type": "equation", + "bbox": [ + 0.403, + 0.375, + 0.786, + 0.408 + ], + "angle": 0, + "content": "\\[\nF _ {i} ^ {M L} = \\operatorname {s o f t m a x} \\left(\\frac {Q K ^ {\\top}}{\\sqrt {d _ {w}}}\\right) V, \\tag {4}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.417, + 0.788, + 0.629 + ], + "angle": 0, + "content": "where \\( Q \\), \\( K \\), and \\( V \\) are Query, Key, and Value vectors obtained through linear mapping of \\( F_{i}^{e} \\), and \\( d_{w} \\) is the feature dimension. Next, we construct a lightweight Attention Aggregation Layer to aggregate all patch features \\( F_{i}^{ML} \\) within a slide into a slide feature \\( F_{i}^{S} \\). The attention weights are learnable, and the detailed structure is shown in Section 3.2. Following this, the constructed visual slide examples \\( E_{m} \\) undergo the aforementioned process like \\( W_{i} \\) (i.e., patch extraction, patch example matching, Messenger Layer, and Summary Layer) to obtain their slide features \\( Z_{m} \\). These slide example features are also integrated as prompts into slide-level training. Similar to the patch-level matching, we match the cosine similarity between \\( F_{i}^{S} \\) and \\( Z_{m} \\) and concatenate the most matched slide example feature with the slide feature. Additionally, we introduce learnable vector \\( F_{i}^{P} \\) as a slide-level learnable prompt, concatenate it together, and input them into a one-layer fully-connected layer named \"projector\" to adjust the feature dimension, yielding the final slide visual features \\( F_{i} \\in \\mathbb{R}^{1 \\times v} \\)." + }, + { + "type": "title", + "bbox": [ + 0.216, + 0.65, + 0.473, + 0.667 + ], + "angle": 0, + "content": "4.3 Textual Prompt Learning" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.674, + 0.788, + 0.795 + ], + "angle": 0, + "content": "Construction of textual prompts. Textual prompt learning primarily focuses on modeling language descriptions related to pathology tasks, patch-level, and slide-level pathology priors on the text side. This knowledge is integrated into the training process through fixed descriptive prompts and learnable prompts (shown in the middle of Fig. 2). The construction of textual prompts could be divided into mainly three parts. The following detailed explanation uses the task \"predicting the prognosis of early cervical cancer patients from H&E pathology slides of the primary lesion\" as an example, also illustrated in Fig. 3b." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.796, + 0.787, + 0.842 + ], + "angle": 0, + "content": "First, the \"Slide task [Token]\" primarily serves to prompt the main category of the current task, such as \"A Whole Slide Image of cervical cancer from primary tumors with a [poor/good] prognosis.\" Furthermore, we concatenate" + } + ], + [ + { + "type": "page_number", + "bbox": [ + 0.218, + 0.116, + 0.236, + 0.127 + ], + "angle": 0, + "content": "10" + }, + { + "type": "header", + "bbox": [ + 0.272, + 0.115, + 0.352, + 0.129 + ], + "angle": 0, + "content": "L. Qu et al." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.146, + 0.784, + 0.177 + ], + "angle": 0, + "content": "learnable prompts \\((([\\alpha]_1[\\alpha]_2\\ldots [\\alpha]_M))\\) with the \"Slide task [Token]\" to adaptively adjust and complement during few-shot training." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.178, + 0.787, + 0.344 + ], + "angle": 0, + "content": "Then, the \"Slide-level descriptive [Token]\" is a straightforward language description corresponding to the aforementioned visual slide examples. These descriptions are provided by experienced pathology experts, avoiding obscure professional terminology. For instance, the description corresponding to the \"Poor prognosis\" category is: \"Fuzzy Tumor Boundaries. The edges or borders of the tumor are fuzzy or hard to distinguish on the slide level. Low tumor stroma ratio. Few stroma tissues surrounding the tumors.\" In contrast, the description corresponding to the \"Good prognosis\" category is: \"Clear Tumor Boundaries. The edges or borders of the tumor are easy to see and well-defined on the slide level. High tumor stroma ratio. Lots of stroma tissue surrounding the tumors.\" Similarly, we also append learnable prompts \\((([\\beta]_{1}[\\beta]_{2}\\ldots [\\beta]_{M}))\\)." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.346, + 0.788, + 0.466 + ], + "angle": 0, + "content": "Finally, the \"Patch-level descriptive [Token]\" provides the language description corresponding to the aforementioned visual patch examples. For example, the description corresponding to the \"Poor prognosis\" category is: \"Vascular and lymphatic invasion: The presence of cancer cells within the blood and lymphatic vessels. High-level atypia: Pronounced and abnormal variations in the size, shape, and staining pattern of the cells. Necrosis: Empty spaces or regions with a lack of normal cell structure.\" Similarly, immediately following this, we also include learnable prompts \\((([\\gamma]_{1}[\\gamma]_{2}\\ldots [\\gamma]_{M}))\\)." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.467, + 0.788, + 0.527 + ], + "angle": 0, + "content": "It is crucial to note that the second and third parts correspond closely to the same number of visual slide and patch examples in visual prompt learning. We input each part's visual descriptions and learnable vectors together into the CLIP Text Encoder, obtaining the final slide textual features \\( T_{i} \\in \\mathbb{R}^{1 \\times v} \\)." + }, + { + "type": "title", + "bbox": [ + 0.216, + 0.552, + 0.61, + 0.568 + ], + "angle": 0, + "content": "4.4 Alignment of Visual and Textual Prompts" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.58, + 0.788, + 0.732 + ], + "angle": 0, + "content": "Alignment contrastive learning (AC-Loss) is originally used to align corresponding visual representation and textual representation. Nonetheless, the overall loss function in this paper consists of three parts of AC-Loss, expressed as \\(\\mathcal{L}_{\\mathrm{total}} = \\mathcal{L}_t + \\lambda_1\\mathcal{L}_s + \\lambda_2\\mathcal{L}_p\\). Conceptually, \\(\\mathcal{L}_t\\) aligns the overall slide visual features \\(F_{i}\\) with the slide textual features \\(T_{i}\\) corresponding to the task category, thereby completing slide-level classification. \\(\\mathcal{L}_s\\) aligns the features of slide-level visual examples in visual prompt learning with the slide-level pathological language descriptions in textual prompt learning. \\(\\mathcal{L}_p\\) aligns the features of patch-level visual examples with patch-level pathology language descriptions (shown in the bottom of Fig. 2)." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.732, + 0.788, + 0.793 + ], + "angle": 0, + "content": "Mathematically, the forms of these three parts of AC-Loss are identical. Here, we take \\(\\mathcal{L}_t\\) as an example. AC-Loss optimizes the learnable parameters by minimizing the negative log-likelihood between the image features \\(F_{i}\\) and the ground truth class text features \\(T_{y}\\). Its basic form is as follows:" + }, + { + "type": "equation", + "bbox": [ + 0.358, + 0.806, + 0.787, + 0.845 + ], + "angle": 0, + "content": "\\[\n\\mathcal {L} = - \\sum_ {F _ {i}} \\log \\frac {\\exp \\left(\\operatorname {s i m} \\left(F _ {i} , T _ {y}\\right) / \\tau\\right)}{\\sum_ {i = 1} ^ {U} \\exp \\left(\\operatorname {s i m} \\left(F _ {i} , T _ {i}\\right) / \\tau\\right)}, \\tag {5}\n\\]" + } + ], + [ + { + "type": "header", + "bbox": [ + 0.301, + 0.115, + 0.732, + 0.131 + ], + "angle": 0, + "content": "Pathology-knowledge Enhanced Multi-instance Prompt Learning" + }, + { + "type": "page_number", + "bbox": [ + 0.768, + 0.116, + 0.784, + 0.127 + ], + "angle": 0, + "content": "11" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.147, + 0.784, + 0.177 + ], + "angle": 0, + "content": "where \\(y\\) is the ground truth category, \\(U\\) is the total number of categories, \\(\\mathrm{sim}(\\cdot)\\) denotes the cosine similarity, and \\(\\tau\\) is a learnable temperature parameter." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.178, + 0.788, + 0.239 + ], + "angle": 0, + "content": "During testing, we input the unseen WSI and obtain its slide visual feature in the same fashion as training. Subsequently, we match it separately with the textual features for \"poor prognosis\" and \"good prognosis\" and use softmax to obtain the predicted risks for the two categories." + }, + { + "type": "title", + "bbox": [ + 0.215, + 0.263, + 0.479, + 0.28 + ], + "angle": 0, + "content": "5 Experiment and Result" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.296, + 0.789, + 0.629 + ], + "angle": 0, + "content": "Datasets and Evaluation Metrics. This paper addresses clinical tasks in H&E pathology diagnosis, specifically focusing on data label scarcity, the complexity of direct diagnosis by pathologists, and limited sample sizes for rare diseases. This study encompasses three clinical tasks for five tumor types across five datasets: (1) Predicting patient survival prognosis and lymph node metastasis from H&E slides of early-stage cervical cancer primary lesions using an in-house dataset (For testing, 80 cases and 74 cases, individually) and a public TCGA-CESC dataset (For testing, 80 cases and 70 cases, individually). Note that the prompts utilized in the public TCGA datasets are the same as those used in the corresponding in-house datasets for each task. We ensure no overlap between our training data and visual prompts (obtained from external authoritative sources or pathology experts) and testing data, across both in-house and TCGA datasets. Furthermore, the lymph node metastasis detection tasks from primary lesions assessed in this study present a significantly greater challenge compared to those from sentinel lymph nodes, as utilized in prior research, with samples and labels being notably rarer. (2) Classifying round cell tumors, including neuroendocrine tumors, malignant melanoma, lymphohematopoietic tumors, and soft tissue tumors, into subtypes using an in-house dataset (For testing, 112 cases). More details are in the Supplementary Materials. For the prognosis prediction task in cervical cancer, the concordance index (C-index) is used, while for other tasks, the Area Under the Curve (AUC) is employed. Higher values for both metrics indicate improved performance." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.63, + 0.789, + 0.841 + ], + "angle": 0, + "content": "Compared Methods. Currently, few methods directly address FSWC tasks through prompt learning, with the SOTA being represented by the recently proposed TOP method [27] in the pathology domain. In the natural domain, prompt learning work based on VLMs is limited to single image patches and does not readily apply to slide-level classification. Following the TOP approach, we adopt advanced prompt learning methods from the natural domain for the MIL problem. These methods include: (1) Linear-Probe [29], (2) VPT [17], (3) CoOp [42], and (4) KgCoOp [38]. The Messenger Layer and Summary Layer, as described in this paper, aggregate patch features into a slide feature. All compared methods use the same Messenger & Summary layers for fair comparisons. For Linear-Probe, a linear layer performs slide-level classification on the aggregated feature. VPT incorporates a learnable prompt on the vision side for slide-level classification. CoOp utilizes CLIP's prompt learning by adding the Slide Task [Token] and learnable textual prompts trained through AC-Loss. Kg-" + } + ], + [ + { + "type": "page_number", + "bbox": [ + 0.218, + 0.116, + 0.236, + 0.127 + ], + "angle": 0, + "content": "12" + }, + { + "type": "header", + "bbox": [ + 0.272, + 0.115, + 0.352, + 0.129 + ], + "angle": 0, + "content": "L. Qu et al." + }, + { + "type": "table_caption", + "bbox": [ + 0.295, + 0.145, + 0.707, + 0.16 + ], + "angle": 0, + "content": "Table 1: Comparison for Survival Prognosis Prediction Task." + }, + { + "type": "table", + "bbox": [ + 0.274, + 0.172, + 0.728, + 0.24 + ], + "angle": 0, + "content": "
DatasetIn-house DatasetPublic TCGA-CESC Dataset
Method32-shot16-shot8-shot4-shot2-shot32-shot16-shot8-shot4-shot2-shot
LinearProbe0.6200.5620.5430.5010.4580.5840.5770.5430.5110.485
VPT(ECCV'22)0.6260.5690.5450.5010.4640.5890.5730.5500.5180.482
CoOp(IJCV'22)0.6410.5940.5610.5170.4900.6040.5850.5580.5230.487
KgCoOp(CVPR'23)0.6350.5870.5520.5130.4870.5970.5840.5500.5190.486
TOP(NeurIPS'23)0.6520.6080.5740.5390.5080.6110.5970.5660.5360.518
PEMP (ours)0.6670.6370.6140.5870.5620.6370.6240.6020.5770.551
" + }, + { + "type": "table_caption", + "bbox": [ + 0.275, + 0.254, + 0.725, + 0.268 + ], + "angle": 0, + "content": "Table 2: Comparison for Lymph Node Metastasis Prediction Task." + }, + { + "type": "table", + "bbox": [ + 0.274, + 0.28, + 0.728, + 0.349 + ], + "angle": 0, + "content": "
DatasetIn-house DatasetPublic TCGA-CESC Dataset
Method32-shot16-shot8-shot4-shot2-shot32-shot16-shot8-shot4-shot2-shot
LinearProbe0.7810.7350.7280.7160.6870.7630.7190.6960.6710.655
VPT(ECCV'22)0.7850.7420.7280.7190.6890.7690.7240.7120.6780.655
CoOp(IJCV'22)0.7920.7570.7350.7270.6930.7860.7410.7230.6910.659
KgCoOp(CVPR'23)0.7890.7530.7310.7240.6930.7750.7330.7100.6870.652
TOP(NeurIPS'23)0.8250.8190.8010.7870.7620.7990.7610.7440.7080.679
PEMP(ours)0.8490.8380.8240.8010.7830.8180.7950.7600.7260.704
" + }, + { + "type": "table_caption", + "bbox": [ + 0.291, + 0.362, + 0.71, + 0.377 + ], + "angle": 0, + "content": "Table 3: Comparison for Round Cell Subtype Diagnosis Task." + }, + { + "type": "table", + "bbox": [ + 0.331, + 0.389, + 0.671, + 0.464 + ], + "angle": 0, + "content": "
Method32-shot16-shot8-shot4-shot2-shot
LinearProbe0.6540.6060.5490.5170.483
VPT(ECCV'22)0.6590.6130.5560.5150.511
CoOp(IJCV'22)0.6670.6270.5680.5530.549
KgCoOp(CVPR'23)0.6610.6190.5470.5350.534
TOP(NeurIPS'23)0.6820.6520.6330.5840.560
PEMP (ours)0.7510.7180.6850.6430.625
" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.491, + 0.788, + 0.552 + ], + "angle": 0, + "content": "CoOp imposes constraints on learnable textual prompts. Few-shot experiments are conducted for each category with 2, 4, 8, 16, and 32 slides (with labels). Results are reported as an average over five experimental runs. Standard deviation of our main results are provided in the supplementary materials." + }, + { + "type": "title", + "bbox": [ + 0.216, + 0.574, + 0.325, + 0.588 + ], + "angle": 0, + "content": "5.1 Results" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.598, + 0.789, + 0.842 + ], + "angle": 0, + "content": "Survival prognosis prediction. Predicting survival prognosis, especially directly from H&E slides with a limited number of samples, poses a significant challenge. Table 1 showcases the performance of all methods in the in-house dataset and the public TCGA-CESC dataset, with PEMP consistently demonstrating superior performance across all settings with few samples, outperforming all comparative methods significantly. In scenarios with limited samples, LinearProbe and VPT, relying solely on features extracted by the CLIP Image Encoder on the visual side without incorporating language-based prior knowledge, exhibit suboptimal performance. CoOp and KgCoOp, which combine visual and language aspects, struggle to comprehend complex medical terms related to prognosis effectively, limiting their performance. TOP enhances language descriptions at both slide and patch levels but provides explanations solely on the language side, making it challenging to establish a direct relationship between crucial visual features and language descriptions. In contrast, PEMP introduces typical patch and slide examples related to prognosis on the visual side, along with straightforward language descriptions on the language side. Through AC-" + } + ], + [ + { + "type": "header", + "bbox": [ + 0.301, + 0.115, + 0.732, + 0.131 + ], + "angle": 0, + "content": "Pathology-knowledge Enhanced Multi-instance Prompt Learning" + }, + { + "type": "page_number", + "bbox": [ + 0.768, + 0.116, + 0.786, + 0.128 + ], + "angle": 0, + "content": "13" + }, + { + "type": "table_caption", + "bbox": [ + 0.359, + 0.145, + 0.645, + 0.159 + ], + "angle": 0, + "content": "Table 4: Ablation studies of the modules." + }, + { + "type": "table", + "bbox": [ + 0.331, + 0.171, + 0.672, + 0.263 + ], + "angle": 0, + "content": "
Method32-shot16-shot8-shot4-shot2-shot
PEMP w/o v&t em.0.6410.5940.5610.5170.490
PEMP vision only0.6490.6070.5680.5430.527
PEMP w/o vision em.0.6550.6130.5770.5400.511
PEMP w/o text em.0.6580.6190.5750.5570.533
PEMP w/o Summary Layer0.6320.5870.5580.5110.487
PEMP w/o Messenger Layer0.6640.6290.5940.5750.554
PEMP w/o Slide-level Prompts0.6560.6200.5810.5520.525
PEMP w/o AC-Loss0.6600.6260.5890.5680.549
PEMP (ours)0.6670.6370.6140.5870.562
" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.292, + 0.784, + 0.354 + ], + "angle": 0, + "content": "Loss, the model efficiently completes the task under the guidance of pathology prior knowledge. Additionally, PEMP excels when the sample size is very small (e.g., 2-shot, 4-shot), showcasing its effective utilization of visual and language prior knowledge related to prognosis in H&E pathology slides for diagnosis." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.355, + 0.787, + 0.476 + ], + "angle": 0, + "content": "Lymph node metastasis prediction. Table 2 presents the performance of all methods in the in-house and TCGA-CESC datasets. Across all settings with limited samples, PEMP consistently achieves the best performance and significantly outperforms all comparative methods. This indicates that PEMP is also capable of efficiently understanding and leveraging pathological prior knowledge related to metastasis in primary lesion H&E pathology slides for diagnosis. The results of both tasks highlight the immense potential of PEMP in clinical tasks characterized by challenges in label acquisition and diagnostic complexity." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.477, + 0.788, + 0.596 + ], + "angle": 0, + "content": "Round cell subtype diagnosis. The classification diagnosis for rare diseases in the round cell subtype task presents an increasingly challenging yet vital clinical issue. Table 3 illustrates the performance of all methods. Notably, with a limited number of shots, most methods face challenges in achieving satisfactory diagnostic results. However, leveraging visual and pathological expertise from pathology experts, PEMP achieves significantly better results for all the settings (6.2% higher AUC on average). This underscores the substantial potential of PEMP in the diagnosis of rare diseases." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.598, + 0.788, + 0.809 + ], + "angle": 0, + "content": "Ablation Study. As shown in Table 4, we designed eight variants and evaluated the roles of each component in PEMP in the prognosis task. (1) PEMP w/o v&t em. represents the removal of all vision and text examples, reducing PEMP to CoOp; (2) PEMP vision only denotes the exclusion of the text end, training solely with vision examples as guidance for prior knowledge; (3) PEMP w/o vision em. indicates the removal of all vision examples; (4) PEMP w/o text em. signifies the exclusion of all text examples; (5) PEMP w/o Summary Layer denotes the replacement from Summary Layer to simple average pooling. Please note that the Summary Layer is essentially an Attention Pooling layer, recognized as a crucial component in MIL problems; (6) PEMP w/o Messenger Layer denotes the removal of the Messenger Layer; (7) PEMP w/o Slide-level Prompt denotes the removal of all slide-level prompts; (8) PEMP w/o AC-Loss denotes the removal of AC-Loss for vision and text examples while retaining the AC-Loss for the visual and text features of the slide task token." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.81, + 0.787, + 0.842 + ], + "angle": 0, + "content": "Visualization and Interpretability. To illustrate the robust performance of PEMP, we briefly visualize typical cases where visual patch examples and" + } + ], + [ + { + "type": "page_number", + "bbox": [ + 0.218, + 0.116, + 0.236, + 0.127 + ], + "angle": 0, + "content": "14" + }, + { + "type": "header", + "bbox": [ + 0.272, + 0.115, + 0.352, + 0.129 + ], + "angle": 0, + "content": "L. Qu et al." + }, + { + "type": "image", + "bbox": [ + 0.303, + 0.143, + 0.703, + 0.343 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.216, + 0.353, + 0.785, + 0.381 + ], + "angle": 0, + "content": "Fig. 4: Visualization of key pathology patterns indicating both good and poor prognosis retrieved by PEMP from the test set." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.41, + 0.788, + 0.533 + ], + "angle": 0, + "content": "slide examples exhibit the highest similarity on the prognosis task's test dataset (Fig. 4). PEMP demonstrates accurate learning of pathology prior knowledge, effectively addressing the challenge of predicting survival outcomes with limited training WSIs. More importantly, PEMP also shows high interpretability. PEMP efficiently integrates the high-risk and low-risk pathology image patterns and textual descriptions that pathologists have summarized over an extended period. PEMP achieves excellent performance under the guidance and facilitation of this prior knowledge." + }, + { + "type": "title", + "bbox": [ + 0.216, + 0.553, + 0.36, + 0.569 + ], + "angle": 0, + "content": "6 Conclusion" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.584, + 0.789, + 0.844 + ], + "angle": 0, + "content": "This paper presents PEMP, a pathology-knowledge enhanced multi-instance prompt learning framework, excelling in challenging tasks such as prognosis prediction with difficult label acquisition and rare tumor classification in Few-shot Weakly Supervised WSI Classification. PEMP integrates vision and text prior knowledge into prompts at both patch and slide levels, using a combination of static and learnable prompts during training to guide the effective diagnosis of key pathology patterns. Alignment-wise contrastive losses ensure feature-level alignment between knowledge-enhanced visual and textual prompts at both patch and slide levels. In specific clinical settings, collecting a large number of WSIs may be challenging due to patient privacy issues, the prevalence of rare diseases, new treatment protocols, or small patient groups. PEMP has shown substantial benefits across five datasets, including prognosis and rare diseases, demonstrating significant diagnostic benefits in fields with limited medical resources. PEMP also demonstrates high interpretability in medical diagnoses, indicating significant potential for future clinical applications. Moreover, such prompt-description-based model adaptation could be widely adopted as a general learning paradigm for specialized domain applications." + } + ], + [ + { + "type": "header", + "bbox": [ + 0.301, + 0.115, + 0.732, + 0.13 + ], + "angle": 0, + "content": "Pathology-knowledge Enhanced Multi-instance Prompt Learning" + }, + { + "type": "page_number", + "bbox": [ + 0.768, + 0.117, + 0.785, + 0.127 + ], + "angle": 0, + "content": "15" + }, + { + "type": "title", + "bbox": [ + 0.218, + 0.145, + 0.403, + 0.163 + ], + "angle": 0, + "content": "Acknowledgements" + }, + { + "type": "text", + "bbox": [ + 0.217, + 0.179, + 0.797, + 0.21 + ], + "angle": 0, + "content": "This work is funded by the National Key R&D Program of China (2022ZD0160700) and Shanghai AI Laboratory." + }, + { + "type": "title", + "bbox": [ + 0.218, + 0.235, + 0.323, + 0.251 + ], + "angle": 0, + "content": "References" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.268, + 0.786, + 0.323 + ], + "angle": 0, + "content": "1. Alayrac, J.B., Donahue, J., Luc, P., Miech, A., Barr, I., Hasson, Y., Lenc, K., Mensch, A., Millican, K., Reynolds, M., et al.: Flamingo: a visual language model for few-shot learning. Advances in Neural Information Processing Systems (NeurIPS) 35, 23716-23736 (2022)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.324, + 0.786, + 0.379 + ], + "angle": 0, + "content": "2. Campanella, G., Hanna, M.G., Geneslaw, L., Miraflor, A., Werneck Krauss Silva, V., Busam, K.J., Brogi, E., Reuter, V.E., Klimstra, D.S., Fuchs, T.J.: Clinical-grade computational pathology using weakly supervised deep learning on whole slide images. Nature Medicine 25(8), 1301-1309 (2019)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.381, + 0.786, + 0.435 + ], + "angle": 0, + "content": "3. Chan, T.H., Cendra, F.J., Ma, L., Yin, G., Yu, L.: Histopathology whole slide image analysis with heterogeneous graph representation learning. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR). pp. 15661-15670 (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.436, + 0.786, + 0.49 + ], + "angle": 0, + "content": "4. Chen, R.J., Chen, C., Li, Y., Chen, T.Y., Trister, A.D., Krishnan, R.G., Mahmood, F.: Scaling vision transformers to gigapixel images via hierarchical self-supervised learning. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR). pp. 16144-16155 (2022)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.491, + 0.786, + 0.547 + ], + "angle": 0, + "content": "5. Chen, R.J., Lu, M.Y., Wang, J., Williamson, D.F., Rodig, S.J., Lindeman, N.I., Mahmood, F.: Pathomic fusion: an integrated framework for fusing histopathology and genomic features for cancer diagnosis and prognosis. IEEE Transactions on Medical Imaging 41(4), 757-770 (2020)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.548, + 0.786, + 0.603 + ], + "angle": 0, + "content": "6. Chen, R.J., Lu, M.Y., Weng, W.H., Chen, T.Y., Williamson, D.F., Manz, T., Shady, M., Mahmood, F.: Multimodal co-attention transformer for survival prediction in gigapixel whole slide images. In: Proceedings of the IEEE/CVF International Conference on Computer Vision (ICCV). pp. 4015-4025 (2021)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.604, + 0.786, + 0.645 + ], + "angle": 0, + "content": "7. Chen, W., Si, C., Zhang, Z., Wang, L., Wang, Z., Tan, T.: Semantic prompt for few-shot image recognition. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR). pp. 23581-23591 (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.645, + 0.786, + 0.7 + ], + "angle": 0, + "content": "8. Chen, Y.C., Lu, C.S.: Rankmix: Data augmentation for weakly supervised learning of classifying whole slide images with diverse sizes and imbalanced categories. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR). pp. 23936-23945 (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.701, + 0.786, + 0.742 + ], + "angle": 0, + "content": "9. Cheplygina, V., de Bruijne, M., Pluim, J.P.: Not-so-supervised: a survey of semi-supervised, multi-instance, and transfer learning in medical image analysis. Medical Image Analysis 54, 280-296 (2019)" + }, + { + "type": "ref_text", + "bbox": [ + 0.22, + 0.743, + 0.786, + 0.798 + ], + "angle": 0, + "content": "10. Chikontwe, P., Kim, M., Nam, S.J., Go, H., Park, S.H.: Multiple instance learning with center embeddings for histopathology classification. In: Medical Image Computing and Computer Assisted Intervention (MICCAI). pp. 519-528. Springer (2020)" + }, + { + "type": "ref_text", + "bbox": [ + 0.22, + 0.799, + 0.786, + 0.84 + ], + "angle": 0, + "content": "11. Gu, J., Han, Z., Chen, S., Beirami, A., He, B., Zhang, G., Liao, R., Qin, Y., Tresp, V., Torr, P.: A systematic survey of prompt engineering on vision-language foundation models. arXiv preprint arXiv:2307.12980 (2023)" + }, + { + "type": "list", + "bbox": [ + 0.22, + 0.268, + 0.786, + 0.84 + ], + "angle": 0, + "content": null + } + ], + [ + { + "type": "page_number", + "bbox": [ + 0.218, + 0.116, + 0.236, + 0.127 + ], + "angle": 0, + "content": "16" + }, + { + "type": "header", + "bbox": [ + 0.272, + 0.115, + 0.352, + 0.129 + ], + "angle": 0, + "content": "L. Qu et al." + }, + { + "type": "ref_text", + "bbox": [ + 0.218, + 0.147, + 0.787, + 0.218 + ], + "angle": 0, + "content": "12. Hashimoto, N., Fukushima, D., Koga, R., Takagi, Y., Ko, K., Kohno, K., Nakaguro, M., Nakamura, S., Hontani, H., Takeuchi, I.: Multi-scale domain-adversarial multiple-instance cnn for cancer subtype classification with unannotated histopathological images. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR). pp. 3852-3861 (2020)" + }, + { + "type": "ref_text", + "bbox": [ + 0.218, + 0.218, + 0.787, + 0.272 + ], + "angle": 0, + "content": "13. Huang, Y., Zhao, W., Wang, S., Fu, Y., Jiang, Y., Yu, L.: Conslide: Asynchronous hierarchical interaction transformer with breakup-reorganize rehearsal for continual whole slide image analysis. In: Proceedings of the IEEE/CVF International Conference on Computer Vision (ICCV). pp. 21349-21360 (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.218, + 0.273, + 0.787, + 0.314 + ], + "angle": 0, + "content": "14. Huang, Z., Bianchi, F., Yuksekgonul, M., Montine, T.J., Zou, J.: A visual-language foundation model for pathology image analysis using medical twitter. Nature Medicine pp. 1-10 (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.218, + 0.315, + 0.787, + 0.369 + ], + "angle": 0, + "content": "15. Ikezogwo, W., Seyfioglu, S., Ghezloo, F., Geva, D., Sheikh Mohammed, F., Anand, P.K., Krishna, R., Shapiro, L.: Quilt-1m: One million image-text pairs for histopathology. Advances in Neural Information Processing Systems (NeurIPS) 36 (2024)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.37, + 0.787, + 0.41 + ], + "angle": 0, + "content": "16. Ilse, M., Tomczak, J., Welling, M.: Attention-based deep multiple instance learning. In: International Conference on Machine Learning (ICML). pp. 2127-2136. PMLR (2018)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.411, + 0.787, + 0.453 + ], + "angle": 0, + "content": "17. Jia, M., Tang, L., Chen, B.C., Cardie, C., Belongie, S., Hariharan, B., Lim, S.N.: Visual prompt tuning. In: European Conference on Computer Vision (ECCV). pp. 709-727. Springer (2022)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.453, + 0.787, + 0.508 + ], + "angle": 0, + "content": "18. Li, B., Li, Y., Eliceiri, K.W.: Dual-stream multiple instance learning network for whole slide image classification with self-supervised contrastive learning. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR). pp. 14318-14328 (2021)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.508, + 0.787, + 0.564 + ], + "angle": 0, + "content": "19. Li, H., Yang, F., Zhao, Y., Xing, X., Zhang, J., Gao, M., Huang, J., Wang, L., Yao, J.: Dt-mil: deformable transformer for multi-instance learning on histopathological image. In: Medical Image Computing and Computer Assisted Intervention (MICCAI). pp. 206-216. Springer (2021)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.565, + 0.787, + 0.632 + ], + "angle": 0, + "content": "20. Li, H., Zhu, C., Zhang, Y., Sun, Y., Shui, Z., Kuang, W., Zheng, S., Yang, L.: Task-specific fine-tuning via variational information bottleneck for weakly-supervised pathology whole slide image classification. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR). pp. 7454-7463 (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.633, + 0.787, + 0.674 + ], + "angle": 0, + "content": "21. Li, J., Li, D., Xiong, C., Hoi, S.: Blip: Bootstrapping language-image pre-training for unified vision-language understanding and generation. In: International Conference on Machine Learning (ICML). pp. 12888-12900. PMLR (2022)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.675, + 0.787, + 0.715 + ], + "angle": 0, + "content": "22. Lin, T., Xu, H., Yang, C., Xu, Y.: Interventional multi-instance learning with deconfounded instance-level prediction. In: Proceedings of the AAAI Conference on Artificial Intelligence (AAAI). vol. 36, pp. 1601-1609 (2022)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.716, + 0.787, + 0.771 + ], + "angle": 0, + "content": "23. Lin, T., Yu, Z., Hu, H., Xu, Y., Chen, C.W.: Interventional bag multi-instance learning on whole-slide pathological images. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR). pp. 19830-19839 (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.771, + 0.787, + 0.841 + ], + "angle": 0, + "content": "24. Lu, M.Y., Chen, B., Zhang, A., Williamson, D.F., Chen, R.J., Ding, T., Le, L.P., Chuang, Y.S., Mahmood, F.: Visual language pretrained multiple instance zero-shot transfer for histopathology images. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR). pp. 19764-19775 (2023)" + }, + { + "type": "list", + "bbox": [ + 0.217, + 0.147, + 0.787, + 0.841 + ], + "angle": 0, + "content": null + } + ], + [ + { + "type": "header", + "bbox": [ + 0.301, + 0.115, + 0.732, + 0.129 + ], + "angle": 0, + "content": "Pathology-knowledge Enhanced Multi-instance Prompt Learning" + }, + { + "type": "page_number", + "bbox": [ + 0.768, + 0.117, + 0.785, + 0.127 + ], + "angle": 0, + "content": "17" + }, + { + "type": "ref_text", + "bbox": [ + 0.214, + 0.147, + 0.788, + 0.189 + ], + "angle": 0, + "content": "25. Lu, M.Y., Williamson, D.F., Chen, T.Y., Chen, R.J., Barbieri, M., Mahmood, F.: Data-efficient and weakly supervised computational pathology on whole-slide images. Nature Biomedical Engineering 5(6), 555-570 (2021)" + }, + { + "type": "ref_text", + "bbox": [ + 0.214, + 0.189, + 0.788, + 0.243 + ], + "angle": 0, + "content": "26. Qu, L., Liu, S., Liu, X., Wang, M., Song, Z.: Towards label-efficient automatic diagnosis and analysis: a comprehensive survey of advanced deep learning-based weakly-supervised, semi-supervised and self-supervised techniques in histopathological image analysis. Physics in Medicine & Biology (2022)" + }, + { + "type": "ref_text", + "bbox": [ + 0.214, + 0.243, + 0.788, + 0.284 + ], + "angle": 0, + "content": "27. Qu, L., Luo, X., Fu, K., Wang, M., Song, Z.: The rise of ai language pathologists: Exploring two-level prompt learning for few-shot weakly-supervised whole slide image classification. arXiv preprint arXiv:2305.17891 (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.214, + 0.284, + 0.788, + 0.325 + ], + "angle": 0, + "content": "28. Qu, L., Luo, X., Liu, S., Wang, M., Song, Z.: Dgmil: Distribution guided multiple instance learning for whole slide image classification. In: Medical Image Computing and Computer Assisted Intervention (MICCAI). pp. 24-34. Springer (2022)" + }, + { + "type": "ref_text", + "bbox": [ + 0.214, + 0.325, + 0.788, + 0.379 + ], + "angle": 0, + "content": "29. Radford, A., Kim, J.W., Hallacy, C., Ramesh, A., Goh, G., Agarwal, S., Sastry, G., Askell, A., Mishkin, P., Clark, J., et al.: Learning transferable visual models from natural language supervision. In: International Conference on Machine Learning (ICML). pp. 8748-8763. PMLR (2021)" + }, + { + "type": "ref_text", + "bbox": [ + 0.214, + 0.379, + 0.788, + 0.42 + ], + "angle": 0, + "content": "30. Rony, J., Belharbi, S., Dolz, J., Ayed, I.B., McCaffrey, L., Granger, E.: Deep weakly-supervised learning methods for classification and localization in histology images: a survey. arXiv preprint arXiv:1909.03354 (2019)" + }, + { + "type": "ref_text", + "bbox": [ + 0.214, + 0.42, + 0.788, + 0.473 + ], + "angle": 0, + "content": "31. Shao, Z., Bian, H., Chen, Y., Wang, Y., Zhang, J., Ji, X., et al.: Transmil: Transformer based correlated multiple instance learning for whole slide image classification. Advances in Neural Information Processing Systems (NeurIPS) 34, 2136-2147 (2021)" + }, + { + "type": "ref_text", + "bbox": [ + 0.214, + 0.473, + 0.788, + 0.514 + ], + "angle": 0, + "content": "32. Shi, X., Xing, F., Xie, Y., Zhang, Z., Cui, L., Yang, L.: Loss-based attention for deep multiple instance learning. In: Proceedings of the AAAI Conference on Artificial Intelligence (AAAI). vol. 34, pp. 5742-5749 (2020)" + }, + { + "type": "ref_text", + "bbox": [ + 0.214, + 0.514, + 0.788, + 0.555 + ], + "angle": 0, + "content": "33. Song, A.H., Jaume, G., Williamson, D.F., Lu, M.Y., Vaidya, A., Miller, T.R., Mahmood, F.: Artificial intelligence for digital and computational pathology. Nature Reviews Bioengineering pp. 1-20 (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.214, + 0.555, + 0.788, + 0.596 + ], + "angle": 0, + "content": "34. Tu, C., Zhang, Y., Ning, Z.: Dual-curriculum contrastive multi-instance learning for cancer prognosis analysis with whole slide images. Advances in Neural Information Processing Systems (NeurIPS) 35, 29484-29497 (2022)" + }, + { + "type": "ref_text", + "bbox": [ + 0.214, + 0.596, + 0.788, + 0.65 + ], + "angle": 0, + "content": "35. Wang, X., Xiang, J., Zhang, J., Yang, S., Yang, Z., Wang, M.H., Zhang, J., Yang, W., Huang, J., Han, X.: Scl-wc: Cross-slide contrastive learning for weakly-supervised whole-slide image classification. Advances in Neural Information Processing Systems (NeurIPS) 35, 18009-18021 (2022)" + }, + { + "type": "ref_text", + "bbox": [ + 0.214, + 0.65, + 0.788, + 0.704 + ], + "angle": 0, + "content": "36. Wasim, S.T., Naseer, M., Khan, S., Khan, F.S., Shah, M.: Vita-clip: Video and text adaptive clip via multimodal prompting. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR). pp. 23034-23044 (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.214, + 0.704, + 0.788, + 0.759 + ], + "angle": 0, + "content": "37. Xu, G., Song, Z., Sun, Z., Ku, C., Yang, Z., Liu, C., Wang, S., Ma, J., Xu, W.: Camel: A weakly supervised learning framework for histopathology image segmentation. In: Proceedings of the IEEE/CVF International Conference on Computer Vision (ICCV). pp. 10682-10691 (2019)" + }, + { + "type": "ref_text", + "bbox": [ + 0.214, + 0.759, + 0.788, + 0.8 + ], + "angle": 0, + "content": "38. Yao, H., Zhang, R., Xu, C.: Visual-language prompt tuning with knowledge-guided context optimization. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR). pp. 6757-6767 (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.214, + 0.8, + 0.788, + 0.84 + ], + "angle": 0, + "content": "39. Yao, J., Zhu, X., Jonnagaddala, J., Hawkins, N., Huang, J.: Whole slide images based cancer survival prediction using attention guided deep multiple instance learning networks. Medical Image Analysis 65, 101789 (2020)" + }, + { + "type": "list", + "bbox": [ + 0.214, + 0.147, + 0.788, + 0.84 + ], + "angle": 0, + "content": null + } + ], + [ + { + "type": "page_number", + "bbox": [ + 0.218, + 0.116, + 0.236, + 0.127 + ], + "angle": 0, + "content": "18" + }, + { + "type": "header", + "bbox": [ + 0.272, + 0.115, + 0.352, + 0.129 + ], + "angle": 0, + "content": "L. Qu et al." + }, + { + "type": "ref_text", + "bbox": [ + 0.218, + 0.147, + 0.787, + 0.203 + ], + "angle": 0, + "content": "40. Zhang, H., Meng, Y., Zhao, Y., Qiao, Y., Yang, X., Coupland, S.E., Zheng, Y.: Dtdf-mil: Double-tier feature distillation multiple instance learning for histopathology whole slide image classification. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR). pp. 18802-18812 (2022)" + }, + { + "type": "ref_text", + "bbox": [ + 0.218, + 0.204, + 0.787, + 0.259 + ], + "angle": 0, + "content": "41. Zhang, Y., Gao, J., Zhou, M., Wang, X., Qiao, Y., Zhang, S., Wang, D.: Text-guided foundation model adaptation for pathological image classification. In: International Conference on Medical Image Computing and Computer-Assisted Intervention (MICCAI). pp. 272-282. Springer (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.26, + 0.787, + 0.286 + ], + "angle": 0, + "content": "42. Zhou, K., Yang, J., Loy, C.C., Liu, Z.: Learning to prompt for vision-language models. International Journal of Computer Vision 130(9), 2337-2348 (2022)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.287, + 0.787, + 0.328 + ], + "angle": 0, + "content": "43. Zhu, X., Yao, J., Zhu, F., Huang, J.: Wsisa: Making survival prediction from whole slide histopathological images. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR). pp. 7234-7242 (2017)" + }, + { + "type": "list", + "bbox": [ + 0.217, + 0.147, + 0.787, + 0.328 + ], + "angle": 0, + "content": null + } + ] +] \ No newline at end of file diff --git a/2024/Pathology-knowledge Enhanced Multi-instance Prompt Learning for Few-shot Whole Slide Image Classification/3e473ac8-930f-46c0-b97b-e8e52a7c0bfd_origin.pdf b/2024/Pathology-knowledge Enhanced Multi-instance Prompt Learning for Few-shot Whole Slide Image Classification/3e473ac8-930f-46c0-b97b-e8e52a7c0bfd_origin.pdf new file mode 100644 index 0000000000000000000000000000000000000000..ea49b4033b52f6cad207b1c290a58c751325bef4 --- /dev/null +++ b/2024/Pathology-knowledge Enhanced Multi-instance Prompt Learning for Few-shot Whole Slide Image Classification/3e473ac8-930f-46c0-b97b-e8e52a7c0bfd_origin.pdf @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:26637409545ae130d3e0a51241b72240d5228ec93c43c458b7261062d8567065 +size 9699077 diff --git a/2024/Pathology-knowledge Enhanced Multi-instance Prompt Learning for Few-shot Whole Slide Image Classification/full.md b/2024/Pathology-knowledge Enhanced Multi-instance Prompt Learning for Few-shot Whole Slide Image Classification/full.md new file mode 100644 index 0000000000000000000000000000000000000000..b25668f9776a65923d3afda10caeddd86a097522 --- /dev/null +++ b/2024/Pathology-knowledge Enhanced Multi-instance Prompt Learning for Few-shot Whole Slide Image Classification/full.md @@ -0,0 +1,262 @@ +# Pathology-knowledge Enhanced Multi-instance Prompt Learning for Few-shot Whole Slide Image Classification + +Linhao Qu $^{1}$ , Dingkang Yang $^{2}$ , Dan Huang $^{3}$ , Qinhao Guo $^{4}$ , Rongkui Luo $^{5}$ , Shaoting Zhang $^{1}$ , and Xiaosong Wang $^{1\boxtimes}$ + +1 Shanghai Artificial Intelligence Laboratory, Shanghai, China + +$^{2}$ Academy for Engineering and Technology, Fudan University, Shanghai, China + +$^{3}$ Department of Pathology, Fudan University Shanghai Cancer Center, China + +$^{4}$ Department of Gynecologic Oncology, Fudan University Shanghai Cancer Center + +$^{5}$ Department of Pathology, Zhongshan Hospital, Fudan University, Shanghai, China wangxiaosong@pjlab.org.cn; + +Abstract. Current multi-instance learning algorithms for pathology image analysis often require a substantial number of Whole Slide Images for effective training but exhibit suboptimal performance in scenarios with limited learning data. In clinical settings, restricted access to pathology slides is inevitable due to patient privacy concerns and the prevalence of rare or emerging diseases. The emergence of the Few-shot Weakly Supervised WSI Classification accommodates the significant challenge of the limited slide data and sparse slide-level labels for diagnosis. Prompt learning based on the pre-trained models (e.g., CLIP) appears to be a promising scheme for this setting; however, current research in this area is limited, and existing algorithms often focus solely on patch-level prompts or confine themselves to language prompts. This paper proposes a multi-instance prompt learning framework enhanced with pathology knowledge, i.e., integrating visual and textual prior knowledge into prompts at both patch and slide levels. The training process employs a combination of static and learnable prompts, effectively guiding the activation of pre-trained models and further facilitating the diagnosis of key pathology patterns. Lightweight Messenger (self-attention) and Summary (attention-pooling) layers are introduced to model relationships between patches and slides within the same patient data. Additionally, alignment-wise contrastive losses ensure the feature-level alignment between visual and textual learnable prompts for both patches and slides. Our method demonstrates superior performance in three challenging clinical tasks, significantly outperforming comparative few-shot methods. + +Keywords: Pathology image analysis $\cdot$ Prompt learning + +# 1 Introduction + +Deep learning for pathology Whole Slide Image (WSI) classification offers a rapid, objective, and precise diagnostic approach, enhancing prognosis predic + +![](images/404a3d468aa48cec21122f2dfb34cb762b56fb207372dc71391b81cf728087ec.jpg) +Fig. 1: Existing methods such as CoOp [42] do not fully consider task-related pathology visual features and its association with specific terms in FSWC. We utilize aligned task-specific image examples and language descriptions to enhance visual and textual prompt learning at both patch and slide levels. + +tion and treatment response assessment [3, 4, 8, 13, 20, 26]. Given the gigapixel resolution of WSIs, segmentation into multiple patches is essential for processing. However, obtaining detailed annotations for each patch is challenging. Instead, it often results in slide-level ground truth labels. Multiple Instance Learning (MIL), a prominent weakly supervised paradigm, has emerged as a key technique for WSI classification [5, 9, 25, 30]. In MIL, each WSI is treated as a bag, with patches as instances. During training, only slide-level labels are available, necessitating a comprehensive model to capture relationships between patches and slide labels for accurate classification [28]. + +MIL algorithms achieve slide-level classification by aggregating patch-level features or predictions. Though effective, existing algorithms typically demand a considerable number of slides for practical training, performing inadequately in scenarios with limited samples. In clinical practice, challenges such as preserving patient privacy, the rarity of certain diseases, and emerging conditions, particularly for prognosis and rare tumor subtypes, hinder the acquisition of a substantial quantity of WSIs [9,26,33]. Recently, the challenging problem of Few-shot weakly Supervised WSI Classification (FSWC) has emerged [27]. FSWC is characterized by the limited availability of training WSIs (e.g., typically 2, 4, 8, 16, or 32 per class) and sparse slide-level annotations, with unknown fine-grained annotations for numerous patches in each WSI. + +In image classification, recent advancements highlight the effectiveness of Vision-Language Models (VLMs) like CLIP [29], BLIP [21], and Flamingo [1]. These models undergo training on large-scale language-image pairs using alignment contrastive learning, demonstrating notable zero-shot and few-shot transfer learning capabilities. In the pathology image domain, despite valuable contributions from initiatives like MI-Zero [24] and PLIP [14] in constructing VLMs for computational pathology, challenges including ethical concerns, annotation scarcity, and storage limitations make it impractical to train foundational mod- + +els for each specific downstream task [27]. Furthermore, existing VLMs exhibit restricted direct zero-shot transfer capabilities for specific downstream tasks [7,36,38]. Considering these challenges, the application of existing VLMs for developing prompt learning algorithms tailored to few-shot downstream tasks appears promising. Prompt learning involves adapting a pre-trained and then fixed VLM to new tasks by integrating task-specific prompts as additional model inputs. In the context of few-shot learning, these prompts can be either fixed or trainable, with trainable prompts enabling contrastive learning between visual and language features [11,42]. + +Presently, research on prompt learning for FSWC is limited [27]. Previous efforts, such as MI-Zero [24], PLIP [14], and CITE [41], have predominantly focused on the utilization of "patch-level" prompts—textual prompts associated with pathology regions of interest. Nonetheless, visual or textual prompts in domain-specific downstream tasks may not be understandable (often unseen) for the pre-trained model and thus may not properly activate the relevant features of the pre-trained model. The TOP framework [27] represents the initial endeavor to implement prompt learning by introducing language-based prior descriptions in the prompts. However, pathology images encompass highly specialized visual features and corresponding pathology-specific terms, posing a challenge for networks to comprehensively understand and learn pathology-related features and specialized descriptions solely through language explanations for classification. + +To address this issue, we propose Pathology-knowledge Enhanced Multi-instance Prompt Learning for FSWC named PEMP, with a concise architecture depicted in Fig. 1. PEMP is a multi-instance prompt learning framework specifically designed for the FSWC problem, incorporating both visual and textual samples (as a form of prior knowledge) at both patch and slide levels. Drawing inspiration from pathologists learning pathology images and language knowledge from textbooks, we introduce task-specific image examples and corresponding language descriptions, injected in both the visual and textual prompts. During prompt learning, the model is fed with a more understandable and effective guide (in image examples and plain words) and then activates more meaningful features for the diagnosis of key patch-level and slide-level pathology patterns in scenarios with limited training samples. + +The contribution of this work is threefold: + +- We propose a multi-instance prompt learning framework, PEMP, introducing vision and text prior knowledge in the designed prompts at both patch and slide levels. A combination of static and learnable prompts is employed in the training, effectively adapting the pre-trained model to focus on key pathology patterns in patches and slides. +- We explore the alignment of visual and textual prompts using contrastive losses at both patch and slide levels. We design lightweight Messenger (self-attention) and Summary (attention-pooling) layers to model relationships of patch-to-patch and patch-to-slide within the same patient data. +- We evaluate PEMP's performance in addressing three challenging pathology tasks in clinical practice, including predicting patient survival and lymph + +node metastasis in early cervical cancer H&E slides and subtype classification for a rare tumor (round cell tumor). PEMP achieves superior performance in all these tasks under the few-shot setting, notably outperforming all comparative methods by a large margin (4% on average). + +# 2 Related Work + +# 2.1 MIL for WSI Classification + +Current MIL algorithms for WSI classification can be broadly categorized as instance-based [2, 10, 22, 28, 37] and bag-based [6, 12, 16, 18, 19, 23, 25, 31, 32, 34, 35, 39, 40, 43] methods. Instance-based methods involve generating pseudo-labels for each patch and training patch classifiers to score all patches within a slide, whereas bag-based methods, the current mainstream, extract patch features and employ techniques like attention weighting [12, 16, 18, 25, 32, 39, 40, 43] to aggregate these features into slide-level features. While existing algorithms perform well with ample training data, they lack specificity for FSWC tasks, struggling with limited samples. Our proposed PEMP, falling under the bag-based MIL paradigm, distinguishes itself by innovatively addressing FSWC through VLMs and prompt learning, deviating from traditional bag-based MIL approaches. + +# 2.2 VLM-Based Prompt Learning + +VLMs, comprising image and text encoders, are trained on large-scale image-text pairs using alignment contrastive learning to establish a shared feature space. Pre-trained VLMs like CLIP [29], BLIP [21], and Flamingo [1] exhibit robust zero-shot and few-shot transferability in image recognition tasks. However, limited accuracy in zero-shot transfer has fueled interest in few-shot learning based on VLMs, where the prompt learning paradigm, demonstrating strong performance [7,36,38], is emerging. Prompt learning leverages a pre-trained VLM in the adaptation to new downstream tasks by incorporating task-specific prompts through enriched model inputs. These prompts could be static with manual designs or learnable vectors [11,42]. + +In pathology, recent research emphasizes constructing specialized VLMs, such as MI-Zero [24], PLIP [14], and QUILT [15]. However, privacy concerns, annotation scarcity, and storage demands make training separate foundational models for each downstream task impractical. Existing VLMs exhibit limited zero-shot transferability, especially for rare diseases. Unlike fine-tuning VLMs at the data level or constructing large datasets like these works, we introduce a novel prompt learning paradigm based on VLMs (e.g., CLIP with frozen parameters) to address the FSWC problem. Importantly, our scheme is versatile across various foundation models. Moreover, MI-Zero, PLIP and QUILT mainly contribute by collecting large datasets to construct pathology VLMs but lack effective methods for FSWC. Their simple adapting strategies for FSWC (e.g., LinearProbe and Attention Pooling) limit their performance in challenging clinical tasks. + +In pathology, prompt learning research for FSWC is limited. Prior works like MI-Zero [24], PLIP [14], and CITE [41] only explore "patch-level" prompts, limiting slide-level classification performance. TOP [27] introduces prompt learning at both patch and slide levels but lacks task-specific visual knowledge on the vision side. Relying solely on language explanations is challenging for the networks to understand pathology features with limited training samples. In contrast, PEMP's learning process operates at patch and slide levels, covers visual and textual prompt design, and introduces task-specific image examples and descriptions separately as prior knowledge. + +# 3 FSWC Baseline Framework + +# 3.1 Problem Formulation + +We provide a brief introduction to the problem definition of FSWC and the pre-trained VLM used in our study. Given a dataset $W = \{W_{1}, W_{2}, \ldots, W_{N}\}$ containing $N$ WSIs, each WSI $W_{i}$ is divided into non-overlapping small image patches $\{p_{i,j}, j = 1, 2, \ldots, n_{i}\}$ , where $n_{i}$ represents the number of patches obtained from $W_{i}$ . It is important to note that the number of patches obtained from each slide may vary. All patches in $W_{i}$ collectively form a bag, and each patch serves as an instance of this bag. Each bag is assigned a label $Y_{i} \in \{0, 1, \ldots, U\}$ , while the label of each instance is unknown, and $U$ is the number of classes. FSWC is highly challenging as it allows training with only a limited number of slides (with labels). In the FSWC task, "shot" refers to the number of WSIs, e.g., $K$ pairs of WSIs and labels for training in a $K$ -shot setting, while the trained model is evaluated on a completely independent test set. Typically, $K$ can take values such as 2, 4, 8, 16, or 32. + +# 3.2 Attention Aggregation for WSI Classification + +The attention-based pooling strategy [16] is a classical MIL aggregation method used for WSI classification. Initially, an encoder $\phi_{img}$ is used to extract features $f_{i,j}$ for all patches $\{p_{i,j}, j = 1,2,\dots n_i\}$ in slide $W_{i}$ , and then aggregate these patch features using an attention-based trainable aggregation function $A(\cdot)$ to obtain the bag feature $F_{i}$ : + +$$ +F _ {i} = A \left(\phi_ {i m g} \left(p _ {i, 1}\right), \phi_ {i m g} \left(p _ {i, 2}\right), \dots\right) = \sum_ {j = 1} ^ {n _ {i}} a _ {i, j} f _ {i, j}, \tag {1} +$$ + +$$ +a _ {i, j} = \frac {\exp \left\{w ^ {\top} \tanh \left(V f _ {i , j} ^ {\top}\right) \right\}}{\sum_ {j = 1} ^ {n _ {i}} \exp \left\{w ^ {\top} \tanh \left(V f _ {i , j} ^ {\top}\right) \right\}}, \tag {2} +$$ + +where $a_{i,j}$ is the attention score predicted by the self-attention network, which is parameterized by $w$ and $V$ . The weights $a_{i,j}$ reflect the contribution of each patch in making the slide-level prediction. The patch-level encoder $\phi_{img}$ is from the pre-trained CLIP [29] and remains fixed during the training process. + +# 3.3 VLM backbone and Few-shot Prompt Learning + +We utilize CLIP [29] as the backbone of the proposed framework and also the pretrained VLM for feature extraction, which is trained on a massive dataset of 400 million image-text pairs, including medical data, facilitating efficient knowledge transfer of diverse visual concepts to downstream tasks. + +Prompt learning enables the efficient adaptation of a fixed pre-trained CLIP in few-shot downstream classification tasks by introducing task-specific prompts into the model input. These prompts, whether fixed or learnable, can be applied to the text encoder, visual encoder, or both simultaneously. In the few-shot setting, the training could also involve aligning visual and text features through alignment contrastive learning (AC-Loss) in the specific domain. Formally, we denote the visual encoder as $\phi_{img}$ and the text encoder as $\phi_{text}$ . For a downstream classification task with $U$ classes, a commonly used fixed hand-crafted prompt, denoted as [Class Prompt], guides classification on the text side. Specifically, assuming the name of the $i$ -th class is “[class-name]," the corresponding hand-crafted prompt is typically designed as “a photo of a [class-name].” These [Class Prompts] are pre-encoded into a vector $C = \{c_i\}_{i=1}^U$ . + +To further enhance generalization to downstream tasks, it is also popular to explore learning a set of continuous context vectors as learnable prompts, denoted as $\mathbb{V} = \{v_1,v_2,\ldots ,v_M\}$ . These learnable context vectors $\mathbb{V}$ , concatenated with the vectors from [Class Prompt] $C$ , form the complete text prompt $t_i = \{v_1,v_2,\dots ,v_M,c_i\}$ . This prompt is then fed into the text encoder $\phi_{text}$ to obtain the final textual features for each class, $T_{i} = \phi_{text}(t_{i})$ . On the visual side, given an image $W_{i}$ and its label $y$ , its slide features $F_{i}$ is computed via Eq. (1). Similar to the text side, learnable prompts could also be appended, which will be inputted into the visual encoder $\phi_{img}$ . + +With a limited number of image-text pairs, the optimization involves minimizing the negative log-likelihood between the visual slide feature $F_{i}$ and its class text feature $T_{y}$ , updating all learnable parameters. During testing, the extracted visual features $F_{i}$ are used to retrieve the class text features $T_{y}$ to obtain the probabilities for each class. + +$$ +p (y \mid F _ {i}) = \frac {\exp (\operatorname {s i m} \left(F _ {i} , T _ {y}\right) / \tau)}{\sum_ {i = 1} ^ {U} \exp (\operatorname {s i m} \left(F _ {i} , T _ {i}\right) / \tau)}, \tag {3} +$$ + +where, $y$ represents the ground truth class, $U$ is the total number of classes, $\mathrm{sim}(\cdot)$ denotes cosine similarity, and $\tau$ is a learnable temperature parameter. It is crucial to note that during the training process, the parameters of CLIP's image encoder and text encoder remain fixed and unchanged. + +# 4 Knowledge-enhanced Prompt Learning + +# 4.1 Overview + +Fig. 2 illustrates the primary workflow of our proposed PEMP, which comprises three learning processes: Visual prompt learning, Textual prompt learning, and + +![](images/f73bce66b1b82e2f233c4d1f4045d8cd1037ffd1aca6f64d147ac22b37c9c289.jpg) +Fig. 2: Overview of PEMP, where flames represent optimized parameters, and snowflakes indicate frozen parameters during training. + +Two-level prompt alignment via contrastive learning. In visual prompt learning, we address three key tasks. Firstly, we integrate prior visual pathology knowledge into the prompts at both patch-level and slide-level. This guides the model, particularly with limited training samples, to understand the prompt better and produce more meaningful activations for task-specific key pathology patterns. Additionally, we also include the learnable prompts on the slide level, triggering the model's precise adaptation to specific tasks. Thirdly, we introduce a lightweight Messenger (self-attention) Layer and Summary (attention-pooling) Layer. The former efficiently communicates and models features of different patches within the same slide, while the latter aggregates patch features into slide features using an attention mechanism. + +In a similar initiative, textual prompt learning concentrates on modeling language descriptions related to pathology tasks, also incorporating patch-level and slide-level priors. Pathologists are employed to provide straightforward language descriptions in the overall task, task-related slide-level pathology, and task-related patch-level pathology, avoiding obscure professional terms. + +Recognizing the challenges faced by the VLM in understanding key visual features and complex pathology terms in visual examples, we establish a close alignment between textual prompt learning (with integrated pathology language descriptions) and visual prompt learning (with injected visual examples) at both + +Fig. 3: (A) and (B) examples of visual and textual token constructions related to the prognosis of early cervical cancer patients. +![](images/6db8b8b77b6bd76bff780a10a8034b7e1e89c17c9ffa2f89887e3cd128f3d194.jpg) +(a) Examples for predicting the prognosis of early cervical cancer patients from H&E pathology slides of the primary lesion, derived from authoritative textbooks and pathology experts. Top: visual patch examples. Bottom: visual slide examples. + +![](images/442e7c81a69bbfc81aba21262bee527b54204d7880347cf5c89af58349d7a4bf.jpg) +(b) Top: Construction of Patch-level descriptive [Token]. Bottom: Construction of Slide-level descriptive [Token]. Examples for the prognosis of early cervical cancer patients from H&E pathology slides of the primary lesion, serving as task-related patch-level and slide-level pathology priors on the text side. + +slide and patch levels. This alignment enhances the organization and cohesion of knowledge-enhanced prompts on both visual and textual sides. + +# 4.2 Visual Prompt Learning + +Construction of visual prompts. In the FSWC task, limited training samples present challenges in acquiring effective knowledge. Moreover, tasks like survival analysis require finding key factors both locally (e.g., patch-level: necrosis, vascular invasion) and globally (e.g., slide-level: tumor boundary, tumor stroma ratio). To overcome these, we out-source prior visual pathology knowledge at both patch and slide levels (shown in the top part of Fig. 3a), guiding the model with task-specific pathology patterns for diagnosis. For challenging slide-level diagnostic tasks, relevant patterns manifest at both levels. Leveraging expertise from pathology professionals, we construct visual patch and slide examples as fixed prompts, as shown in Fig. 3a. For each classification task, highly relevant visual patches and slides indicating high and low risks are retrieved from authoritative sources or pathology experts. For instance, in the prognosis of early cervical cancer, visual patch examples for poor prognosis include "high-grade atypia", "vascular invasion," and "necrosis," while slide examples encompass "overall infiltration with fuzzy tumor boundaries and low tumor stroma ratio." Conversely, low-risk examples include "low-grade dysplasia," "no vascular invasion," and "overall pushing-type with clear tumor boundaries and high tumor stroma ratio." + +In total, 6 slides and 6 patch images are used as visual prompts, including 3 typical slides per prognosis category (good and poor), each with 3 local patches in the prognosis of early cervical cancer. For the Lymph Node Metastasis Prediction Task, we used 4 slides and 4 patches, with two representative slides per category (metastasis and non-metastasis) and two local images per slide. For Round Cell Subtype Diagnosis, 10 slides and 10 patches were used, with two slides per category, each with two local images. Selection of these visual exam + +samples was guided by clinical experts, with each risk factor represented by one slide of the patch and its local patch. + +Learning pipeline. As illustrated in Fig. 2, we assume $W_{i}$ represents the training slide containing $n_i$ patches named $p_{i,j}$ and utilize the fixed CLIP image encoder $\phi_{img}$ to extract patch features $f_{i,j} = \phi_{img}(p_{i,j}) \in \mathbb{R}^{1 \times v}$ , where $v$ represents the dimension of the features. Subsequently, the constructed visual patch examples $e_l$ are also transformed into patch features $z_l = \phi_{img}(e_l)$ . $z_l$ remains unchanged throughout the training process and is combined with the current slide's patch features as a fixed visual prompt. Specifically, we match the cosine similarity between $f_{i,j}$ and $z_l$ , concatenate the most matched example patch feature with the patch feature as $f_{i,j}^e$ , and ultimately concatenate all new patch features for the current slide $F_i^e = \{f_{i,j}^e\}_{j=1}^{n_i} \in \mathbb{R}^{n_i \times v \times 2}$ . To model all patch features within a slide effectively, we introduce a lightweight self-attention layer named Messenger Layer (ML). Taking $F_i^e$ as input, it outputs $F_i^{ML}$ : + +$$ +F _ {i} ^ {M L} = \operatorname {s o f t m a x} \left(\frac {Q K ^ {\top}}{\sqrt {d _ {w}}}\right) V, \tag {4} +$$ + +where $Q$ , $K$ , and $V$ are Query, Key, and Value vectors obtained through linear mapping of $F_{i}^{e}$ , and $d_{w}$ is the feature dimension. Next, we construct a lightweight Attention Aggregation Layer to aggregate all patch features $F_{i}^{ML}$ within a slide into a slide feature $F_{i}^{S}$ . The attention weights are learnable, and the detailed structure is shown in Section 3.2. Following this, the constructed visual slide examples $E_{m}$ undergo the aforementioned process like $W_{i}$ (i.e., patch extraction, patch example matching, Messenger Layer, and Summary Layer) to obtain their slide features $Z_{m}$ . These slide example features are also integrated as prompts into slide-level training. Similar to the patch-level matching, we match the cosine similarity between $F_{i}^{S}$ and $Z_{m}$ and concatenate the most matched slide example feature with the slide feature. Additionally, we introduce learnable vector $F_{i}^{P}$ as a slide-level learnable prompt, concatenate it together, and input them into a one-layer fully-connected layer named "projector" to adjust the feature dimension, yielding the final slide visual features $F_{i} \in \mathbb{R}^{1 \times v}$ . + +# 4.3 Textual Prompt Learning + +Construction of textual prompts. Textual prompt learning primarily focuses on modeling language descriptions related to pathology tasks, patch-level, and slide-level pathology priors on the text side. This knowledge is integrated into the training process through fixed descriptive prompts and learnable prompts (shown in the middle of Fig. 2). The construction of textual prompts could be divided into mainly three parts. The following detailed explanation uses the task "predicting the prognosis of early cervical cancer patients from H&E pathology slides of the primary lesion" as an example, also illustrated in Fig. 3b. + +First, the "Slide task [Token]" primarily serves to prompt the main category of the current task, such as "A Whole Slide Image of cervical cancer from primary tumors with a [poor/good] prognosis." Furthermore, we concatenate + +learnable prompts $(([\alpha]_1[\alpha]_2\ldots [\alpha]_M))$ with the "Slide task [Token]" to adaptively adjust and complement during few-shot training. + +Then, the "Slide-level descriptive [Token]" is a straightforward language description corresponding to the aforementioned visual slide examples. These descriptions are provided by experienced pathology experts, avoiding obscure professional terminology. For instance, the description corresponding to the "Poor prognosis" category is: "Fuzzy Tumor Boundaries. The edges or borders of the tumor are fuzzy or hard to distinguish on the slide level. Low tumor stroma ratio. Few stroma tissues surrounding the tumors." In contrast, the description corresponding to the "Good prognosis" category is: "Clear Tumor Boundaries. The edges or borders of the tumor are easy to see and well-defined on the slide level. High tumor stroma ratio. Lots of stroma tissue surrounding the tumors." Similarly, we also append learnable prompts $(([\beta]_{1}[\beta]_{2}\ldots [\beta]_{M}))$ . + +Finally, the "Patch-level descriptive [Token]" provides the language description corresponding to the aforementioned visual patch examples. For example, the description corresponding to the "Poor prognosis" category is: "Vascular and lymphatic invasion: The presence of cancer cells within the blood and lymphatic vessels. High-level atypia: Pronounced and abnormal variations in the size, shape, and staining pattern of the cells. Necrosis: Empty spaces or regions with a lack of normal cell structure." Similarly, immediately following this, we also include learnable prompts $(([\gamma]_{1}[\gamma]_{2}\ldots [\gamma]_{M}))$ . + +It is crucial to note that the second and third parts correspond closely to the same number of visual slide and patch examples in visual prompt learning. We input each part's visual descriptions and learnable vectors together into the CLIP Text Encoder, obtaining the final slide textual features $T_{i} \in \mathbb{R}^{1 \times v}$ . + +# 4.4 Alignment of Visual and Textual Prompts + +Alignment contrastive learning (AC-Loss) is originally used to align corresponding visual representation and textual representation. Nonetheless, the overall loss function in this paper consists of three parts of AC-Loss, expressed as $\mathcal{L}_{\mathrm{total}} = \mathcal{L}_t + \lambda_1\mathcal{L}_s + \lambda_2\mathcal{L}_p$ . Conceptually, $\mathcal{L}_t$ aligns the overall slide visual features $F_{i}$ with the slide textual features $T_{i}$ corresponding to the task category, thereby completing slide-level classification. $\mathcal{L}_s$ aligns the features of slide-level visual examples in visual prompt learning with the slide-level pathological language descriptions in textual prompt learning. $\mathcal{L}_p$ aligns the features of patch-level visual examples with patch-level pathology language descriptions (shown in the bottom of Fig. 2). + +Mathematically, the forms of these three parts of AC-Loss are identical. Here, we take $\mathcal{L}_t$ as an example. AC-Loss optimizes the learnable parameters by minimizing the negative log-likelihood between the image features $F_{i}$ and the ground truth class text features $T_{y}$ . Its basic form is as follows: + +$$ +\mathcal {L} = - \sum_ {F _ {i}} \log \frac {\exp \left(\operatorname {s i m} \left(F _ {i} , T _ {y}\right) / \tau\right)}{\sum_ {i = 1} ^ {U} \exp \left(\operatorname {s i m} \left(F _ {i} , T _ {i}\right) / \tau\right)}, \tag {5} +$$ + +where $y$ is the ground truth category, $U$ is the total number of categories, $\mathrm{sim}(\cdot)$ denotes the cosine similarity, and $\tau$ is a learnable temperature parameter. + +During testing, we input the unseen WSI and obtain its slide visual feature in the same fashion as training. Subsequently, we match it separately with the textual features for "poor prognosis" and "good prognosis" and use softmax to obtain the predicted risks for the two categories. + +# 5 Experiment and Result + +Datasets and Evaluation Metrics. This paper addresses clinical tasks in H&E pathology diagnosis, specifically focusing on data label scarcity, the complexity of direct diagnosis by pathologists, and limited sample sizes for rare diseases. This study encompasses three clinical tasks for five tumor types across five datasets: (1) Predicting patient survival prognosis and lymph node metastasis from H&E slides of early-stage cervical cancer primary lesions using an in-house dataset (For testing, 80 cases and 74 cases, individually) and a public TCGA-CESC dataset (For testing, 80 cases and 70 cases, individually). Note that the prompts utilized in the public TCGA datasets are the same as those used in the corresponding in-house datasets for each task. We ensure no overlap between our training data and visual prompts (obtained from external authoritative sources or pathology experts) and testing data, across both in-house and TCGA datasets. Furthermore, the lymph node metastasis detection tasks from primary lesions assessed in this study present a significantly greater challenge compared to those from sentinel lymph nodes, as utilized in prior research, with samples and labels being notably rarer. (2) Classifying round cell tumors, including neuroendocrine tumors, malignant melanoma, lymphohematopoietic tumors, and soft tissue tumors, into subtypes using an in-house dataset (For testing, 112 cases). More details are in the Supplementary Materials. For the prognosis prediction task in cervical cancer, the concordance index (C-index) is used, while for other tasks, the Area Under the Curve (AUC) is employed. Higher values for both metrics indicate improved performance. + +Compared Methods. Currently, few methods directly address FSWC tasks through prompt learning, with the SOTA being represented by the recently proposed TOP method [27] in the pathology domain. In the natural domain, prompt learning work based on VLMs is limited to single image patches and does not readily apply to slide-level classification. Following the TOP approach, we adopt advanced prompt learning methods from the natural domain for the MIL problem. These methods include: (1) Linear-Probe [29], (2) VPT [17], (3) CoOp [42], and (4) KgCoOp [38]. The Messenger Layer and Summary Layer, as described in this paper, aggregate patch features into a slide feature. All compared methods use the same Messenger & Summary layers for fair comparisons. For Linear-Probe, a linear layer performs slide-level classification on the aggregated feature. VPT incorporates a learnable prompt on the vision side for slide-level classification. CoOp utilizes CLIP's prompt learning by adding the Slide Task [Token] and learnable textual prompts trained through AC-Loss. Kg- + +Table 1: Comparison for Survival Prognosis Prediction Task. + +
DatasetIn-house DatasetPublic TCGA-CESC Dataset
Method32-shot16-shot8-shot4-shot2-shot32-shot16-shot8-shot4-shot2-shot
LinearProbe0.6200.5620.5430.5010.4580.5840.5770.5430.5110.485
VPT(ECCV'22)0.6260.5690.5450.5010.4640.5890.5730.5500.5180.482
CoOp(IJCV'22)0.6410.5940.5610.5170.4900.6040.5850.5580.5230.487
KgCoOp(CVPR'23)0.6350.5870.5520.5130.4870.5970.5840.5500.5190.486
TOP(NeurIPS'23)0.6520.6080.5740.5390.5080.6110.5970.5660.5360.518
PEMP (ours)0.6670.6370.6140.5870.5620.6370.6240.6020.5770.551
+ +Table 2: Comparison for Lymph Node Metastasis Prediction Task. + +
DatasetIn-house DatasetPublic TCGA-CESC Dataset
Method32-shot16-shot8-shot4-shot2-shot32-shot16-shot8-shot4-shot2-shot
LinearProbe0.7810.7350.7280.7160.6870.7630.7190.6960.6710.655
VPT(ECCV'22)0.7850.7420.7280.7190.6890.7690.7240.7120.6780.655
CoOp(IJCV'22)0.7920.7570.7350.7270.6930.7860.7410.7230.6910.659
KgCoOp(CVPR'23)0.7890.7530.7310.7240.6930.7750.7330.7100.6870.652
TOP(NeurIPS'23)0.8250.8190.8010.7870.7620.7990.7610.7440.7080.679
PEMP(ours)0.8490.8380.8240.8010.7830.8180.7950.7600.7260.704
+ +Table 3: Comparison for Round Cell Subtype Diagnosis Task. + +
Method32-shot16-shot8-shot4-shot2-shot
LinearProbe0.6540.6060.5490.5170.483
VPT(ECCV'22)0.6590.6130.5560.5150.511
CoOp(IJCV'22)0.6670.6270.5680.5530.549
KgCoOp(CVPR'23)0.6610.6190.5470.5350.534
TOP(NeurIPS'23)0.6820.6520.6330.5840.560
PEMP (ours)0.7510.7180.6850.6430.625
+ +CoOp imposes constraints on learnable textual prompts. Few-shot experiments are conducted for each category with 2, 4, 8, 16, and 32 slides (with labels). Results are reported as an average over five experimental runs. Standard deviation of our main results are provided in the supplementary materials. + +# 5.1 Results + +Survival prognosis prediction. Predicting survival prognosis, especially directly from H&E slides with a limited number of samples, poses a significant challenge. Table 1 showcases the performance of all methods in the in-house dataset and the public TCGA-CESC dataset, with PEMP consistently demonstrating superior performance across all settings with few samples, outperforming all comparative methods significantly. In scenarios with limited samples, LinearProbe and VPT, relying solely on features extracted by the CLIP Image Encoder on the visual side without incorporating language-based prior knowledge, exhibit suboptimal performance. CoOp and KgCoOp, which combine visual and language aspects, struggle to comprehend complex medical terms related to prognosis effectively, limiting their performance. TOP enhances language descriptions at both slide and patch levels but provides explanations solely on the language side, making it challenging to establish a direct relationship between crucial visual features and language descriptions. In contrast, PEMP introduces typical patch and slide examples related to prognosis on the visual side, along with straightforward language descriptions on the language side. Through AC- + +Table 4: Ablation studies of the modules. + +
Method32-shot16-shot8-shot4-shot2-shot
PEMP w/o v&t em.0.6410.5940.5610.5170.490
PEMP vision only0.6490.6070.5680.5430.527
PEMP w/o vision em.0.6550.6130.5770.5400.511
PEMP w/o text em.0.6580.6190.5750.5570.533
PEMP w/o Summary Layer0.6320.5870.5580.5110.487
PEMP w/o Messenger Layer0.6640.6290.5940.5750.554
PEMP w/o Slide-level Prompts0.6560.6200.5810.5520.525
PEMP w/o AC-Loss0.6600.6260.5890.5680.549
PEMP (ours)0.6670.6370.6140.5870.562
+ +Loss, the model efficiently completes the task under the guidance of pathology prior knowledge. Additionally, PEMP excels when the sample size is very small (e.g., 2-shot, 4-shot), showcasing its effective utilization of visual and language prior knowledge related to prognosis in H&E pathology slides for diagnosis. + +Lymph node metastasis prediction. Table 2 presents the performance of all methods in the in-house and TCGA-CESC datasets. Across all settings with limited samples, PEMP consistently achieves the best performance and significantly outperforms all comparative methods. This indicates that PEMP is also capable of efficiently understanding and leveraging pathological prior knowledge related to metastasis in primary lesion H&E pathology slides for diagnosis. The results of both tasks highlight the immense potential of PEMP in clinical tasks characterized by challenges in label acquisition and diagnostic complexity. + +Round cell subtype diagnosis. The classification diagnosis for rare diseases in the round cell subtype task presents an increasingly challenging yet vital clinical issue. Table 3 illustrates the performance of all methods. Notably, with a limited number of shots, most methods face challenges in achieving satisfactory diagnostic results. However, leveraging visual and pathological expertise from pathology experts, PEMP achieves significantly better results for all the settings (6.2% higher AUC on average). This underscores the substantial potential of PEMP in the diagnosis of rare diseases. + +Ablation Study. As shown in Table 4, we designed eight variants and evaluated the roles of each component in PEMP in the prognosis task. (1) PEMP w/o v&t em. represents the removal of all vision and text examples, reducing PEMP to CoOp; (2) PEMP vision only denotes the exclusion of the text end, training solely with vision examples as guidance for prior knowledge; (3) PEMP w/o vision em. indicates the removal of all vision examples; (4) PEMP w/o text em. signifies the exclusion of all text examples; (5) PEMP w/o Summary Layer denotes the replacement from Summary Layer to simple average pooling. Please note that the Summary Layer is essentially an Attention Pooling layer, recognized as a crucial component in MIL problems; (6) PEMP w/o Messenger Layer denotes the removal of the Messenger Layer; (7) PEMP w/o Slide-level Prompt denotes the removal of all slide-level prompts; (8) PEMP w/o AC-Loss denotes the removal of AC-Loss for vision and text examples while retaining the AC-Loss for the visual and text features of the slide task token. + +Visualization and Interpretability. To illustrate the robust performance of PEMP, we briefly visualize typical cases where visual patch examples and + +![](images/cb2d44ba90b186e9cf6faf3a1be08efa2a48624c996022d2b57abe93e26c43d5.jpg) +Fig. 4: Visualization of key pathology patterns indicating both good and poor prognosis retrieved by PEMP from the test set. + +slide examples exhibit the highest similarity on the prognosis task's test dataset (Fig. 4). PEMP demonstrates accurate learning of pathology prior knowledge, effectively addressing the challenge of predicting survival outcomes with limited training WSIs. More importantly, PEMP also shows high interpretability. PEMP efficiently integrates the high-risk and low-risk pathology image patterns and textual descriptions that pathologists have summarized over an extended period. PEMP achieves excellent performance under the guidance and facilitation of this prior knowledge. + +# 6 Conclusion + +This paper presents PEMP, a pathology-knowledge enhanced multi-instance prompt learning framework, excelling in challenging tasks such as prognosis prediction with difficult label acquisition and rare tumor classification in Few-shot Weakly Supervised WSI Classification. PEMP integrates vision and text prior knowledge into prompts at both patch and slide levels, using a combination of static and learnable prompts during training to guide the effective diagnosis of key pathology patterns. Alignment-wise contrastive losses ensure feature-level alignment between knowledge-enhanced visual and textual prompts at both patch and slide levels. In specific clinical settings, collecting a large number of WSIs may be challenging due to patient privacy issues, the prevalence of rare diseases, new treatment protocols, or small patient groups. PEMP has shown substantial benefits across five datasets, including prognosis and rare diseases, demonstrating significant diagnostic benefits in fields with limited medical resources. PEMP also demonstrates high interpretability in medical diagnoses, indicating significant potential for future clinical applications. Moreover, such prompt-description-based model adaptation could be widely adopted as a general learning paradigm for specialized domain applications. + +# Acknowledgements + +This work is funded by the National Key R&D Program of China (2022ZD0160700) and Shanghai AI Laboratory. + +# References + +1. Alayrac, J.B., Donahue, J., Luc, P., Miech, A., Barr, I., Hasson, Y., Lenc, K., Mensch, A., Millican, K., Reynolds, M., et al.: Flamingo: a visual language model for few-shot learning. Advances in Neural Information Processing Systems (NeurIPS) 35, 23716-23736 (2022) +2. Campanella, G., Hanna, M.G., Geneslaw, L., Miraflor, A., Werneck Krauss Silva, V., Busam, K.J., Brogi, E., Reuter, V.E., Klimstra, D.S., Fuchs, T.J.: Clinical-grade computational pathology using weakly supervised deep learning on whole slide images. Nature Medicine 25(8), 1301-1309 (2019) +3. Chan, T.H., Cendra, F.J., Ma, L., Yin, G., Yu, L.: Histopathology whole slide image analysis with heterogeneous graph representation learning. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR). pp. 15661-15670 (2023) +4. Chen, R.J., Chen, C., Li, Y., Chen, T.Y., Trister, A.D., Krishnan, R.G., Mahmood, F.: Scaling vision transformers to gigapixel images via hierarchical self-supervised learning. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR). pp. 16144-16155 (2022) +5. Chen, R.J., Lu, M.Y., Wang, J., Williamson, D.F., Rodig, S.J., Lindeman, N.I., Mahmood, F.: Pathomic fusion: an integrated framework for fusing histopathology and genomic features for cancer diagnosis and prognosis. IEEE Transactions on Medical Imaging 41(4), 757-770 (2020) +6. Chen, R.J., Lu, M.Y., Weng, W.H., Chen, T.Y., Williamson, D.F., Manz, T., Shady, M., Mahmood, F.: Multimodal co-attention transformer for survival prediction in gigapixel whole slide images. In: Proceedings of the IEEE/CVF International Conference on Computer Vision (ICCV). pp. 4015-4025 (2021) +7. Chen, W., Si, C., Zhang, Z., Wang, L., Wang, Z., Tan, T.: Semantic prompt for few-shot image recognition. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR). pp. 23581-23591 (2023) +8. Chen, Y.C., Lu, C.S.: Rankmix: Data augmentation for weakly supervised learning of classifying whole slide images with diverse sizes and imbalanced categories. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR). pp. 23936-23945 (2023) +9. Cheplygina, V., de Bruijne, M., Pluim, J.P.: Not-so-supervised: a survey of semi-supervised, multi-instance, and transfer learning in medical image analysis. Medical Image Analysis 54, 280-296 (2019) +10. Chikontwe, P., Kim, M., Nam, S.J., Go, H., Park, S.H.: Multiple instance learning with center embeddings for histopathology classification. In: Medical Image Computing and Computer Assisted Intervention (MICCAI). pp. 519-528. Springer (2020) +11. Gu, J., Han, Z., Chen, S., Beirami, A., He, B., Zhang, G., Liao, R., Qin, Y., Tresp, V., Torr, P.: A systematic survey of prompt engineering on vision-language foundation models. arXiv preprint arXiv:2307.12980 (2023) + +12. Hashimoto, N., Fukushima, D., Koga, R., Takagi, Y., Ko, K., Kohno, K., Nakaguro, M., Nakamura, S., Hontani, H., Takeuchi, I.: Multi-scale domain-adversarial multiple-instance cnn for cancer subtype classification with unannotated histopathological images. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR). pp. 3852-3861 (2020) +13. Huang, Y., Zhao, W., Wang, S., Fu, Y., Jiang, Y., Yu, L.: Conslide: Asynchronous hierarchical interaction transformer with breakup-reorganize rehearsal for continual whole slide image analysis. In: Proceedings of the IEEE/CVF International Conference on Computer Vision (ICCV). pp. 21349-21360 (2023) +14. Huang, Z., Bianchi, F., Yuksekgonul, M., Montine, T.J., Zou, J.: A visual-language foundation model for pathology image analysis using medical twitter. Nature Medicine pp. 1-10 (2023) +15. Ikezogwo, W., Seyfioglu, S., Ghezloo, F., Geva, D., Sheikh Mohammed, F., Anand, P.K., Krishna, R., Shapiro, L.: Quilt-1m: One million image-text pairs for histopathology. Advances in Neural Information Processing Systems (NeurIPS) 36 (2024) +16. Ilse, M., Tomczak, J., Welling, M.: Attention-based deep multiple instance learning. In: International Conference on Machine Learning (ICML). pp. 2127-2136. PMLR (2018) +17. Jia, M., Tang, L., Chen, B.C., Cardie, C., Belongie, S., Hariharan, B., Lim, S.N.: Visual prompt tuning. In: European Conference on Computer Vision (ECCV). pp. 709-727. Springer (2022) +18. Li, B., Li, Y., Eliceiri, K.W.: Dual-stream multiple instance learning network for whole slide image classification with self-supervised contrastive learning. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR). pp. 14318-14328 (2021) +19. Li, H., Yang, F., Zhao, Y., Xing, X., Zhang, J., Gao, M., Huang, J., Wang, L., Yao, J.: Dt-mil: deformable transformer for multi-instance learning on histopathological image. In: Medical Image Computing and Computer Assisted Intervention (MICCAI). pp. 206-216. Springer (2021) +20. Li, H., Zhu, C., Zhang, Y., Sun, Y., Shui, Z., Kuang, W., Zheng, S., Yang, L.: Task-specific fine-tuning via variational information bottleneck for weakly-supervised pathology whole slide image classification. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR). pp. 7454-7463 (2023) +21. Li, J., Li, D., Xiong, C., Hoi, S.: Blip: Bootstrapping language-image pre-training for unified vision-language understanding and generation. In: International Conference on Machine Learning (ICML). pp. 12888-12900. PMLR (2022) +22. Lin, T., Xu, H., Yang, C., Xu, Y.: Interventional multi-instance learning with deconfounded instance-level prediction. In: Proceedings of the AAAI Conference on Artificial Intelligence (AAAI). vol. 36, pp. 1601-1609 (2022) +23. Lin, T., Yu, Z., Hu, H., Xu, Y., Chen, C.W.: Interventional bag multi-instance learning on whole-slide pathological images. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR). pp. 19830-19839 (2023) +24. Lu, M.Y., Chen, B., Zhang, A., Williamson, D.F., Chen, R.J., Ding, T., Le, L.P., Chuang, Y.S., Mahmood, F.: Visual language pretrained multiple instance zero-shot transfer for histopathology images. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR). pp. 19764-19775 (2023) + +25. Lu, M.Y., Williamson, D.F., Chen, T.Y., Chen, R.J., Barbieri, M., Mahmood, F.: Data-efficient and weakly supervised computational pathology on whole-slide images. Nature Biomedical Engineering 5(6), 555-570 (2021) +26. Qu, L., Liu, S., Liu, X., Wang, M., Song, Z.: Towards label-efficient automatic diagnosis and analysis: a comprehensive survey of advanced deep learning-based weakly-supervised, semi-supervised and self-supervised techniques in histopathological image analysis. Physics in Medicine & Biology (2022) +27. Qu, L., Luo, X., Fu, K., Wang, M., Song, Z.: The rise of ai language pathologists: Exploring two-level prompt learning for few-shot weakly-supervised whole slide image classification. arXiv preprint arXiv:2305.17891 (2023) +28. Qu, L., Luo, X., Liu, S., Wang, M., Song, Z.: Dgmil: Distribution guided multiple instance learning for whole slide image classification. In: Medical Image Computing and Computer Assisted Intervention (MICCAI). pp. 24-34. Springer (2022) +29. Radford, A., Kim, J.W., Hallacy, C., Ramesh, A., Goh, G., Agarwal, S., Sastry, G., Askell, A., Mishkin, P., Clark, J., et al.: Learning transferable visual models from natural language supervision. In: International Conference on Machine Learning (ICML). pp. 8748-8763. PMLR (2021) +30. Rony, J., Belharbi, S., Dolz, J., Ayed, I.B., McCaffrey, L., Granger, E.: Deep weakly-supervised learning methods for classification and localization in histology images: a survey. arXiv preprint arXiv:1909.03354 (2019) +31. Shao, Z., Bian, H., Chen, Y., Wang, Y., Zhang, J., Ji, X., et al.: Transmil: Transformer based correlated multiple instance learning for whole slide image classification. Advances in Neural Information Processing Systems (NeurIPS) 34, 2136-2147 (2021) +32. Shi, X., Xing, F., Xie, Y., Zhang, Z., Cui, L., Yang, L.: Loss-based attention for deep multiple instance learning. In: Proceedings of the AAAI Conference on Artificial Intelligence (AAAI). vol. 34, pp. 5742-5749 (2020) +33. Song, A.H., Jaume, G., Williamson, D.F., Lu, M.Y., Vaidya, A., Miller, T.R., Mahmood, F.: Artificial intelligence for digital and computational pathology. Nature Reviews Bioengineering pp. 1-20 (2023) +34. Tu, C., Zhang, Y., Ning, Z.: Dual-curriculum contrastive multi-instance learning for cancer prognosis analysis with whole slide images. Advances in Neural Information Processing Systems (NeurIPS) 35, 29484-29497 (2022) +35. Wang, X., Xiang, J., Zhang, J., Yang, S., Yang, Z., Wang, M.H., Zhang, J., Yang, W., Huang, J., Han, X.: Scl-wc: Cross-slide contrastive learning for weakly-supervised whole-slide image classification. Advances in Neural Information Processing Systems (NeurIPS) 35, 18009-18021 (2022) +36. Wasim, S.T., Naseer, M., Khan, S., Khan, F.S., Shah, M.: Vita-clip: Video and text adaptive clip via multimodal prompting. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR). pp. 23034-23044 (2023) +37. Xu, G., Song, Z., Sun, Z., Ku, C., Yang, Z., Liu, C., Wang, S., Ma, J., Xu, W.: Camel: A weakly supervised learning framework for histopathology image segmentation. In: Proceedings of the IEEE/CVF International Conference on Computer Vision (ICCV). pp. 10682-10691 (2019) +38. Yao, H., Zhang, R., Xu, C.: Visual-language prompt tuning with knowledge-guided context optimization. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR). pp. 6757-6767 (2023) +39. Yao, J., Zhu, X., Jonnagaddala, J., Hawkins, N., Huang, J.: Whole slide images based cancer survival prediction using attention guided deep multiple instance learning networks. Medical Image Analysis 65, 101789 (2020) + +40. Zhang, H., Meng, Y., Zhao, Y., Qiao, Y., Yang, X., Coupland, S.E., Zheng, Y.: Dtdf-mil: Double-tier feature distillation multiple instance learning for histopathology whole slide image classification. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR). pp. 18802-18812 (2022) +41. Zhang, Y., Gao, J., Zhou, M., Wang, X., Qiao, Y., Zhang, S., Wang, D.: Text-guided foundation model adaptation for pathological image classification. In: International Conference on Medical Image Computing and Computer-Assisted Intervention (MICCAI). pp. 272-282. Springer (2023) +42. Zhou, K., Yang, J., Loy, C.C., Liu, Z.: Learning to prompt for vision-language models. International Journal of Computer Vision 130(9), 2337-2348 (2022) +43. Zhu, X., Yao, J., Zhu, F., Huang, J.: Wsisa: Making survival prediction from whole slide histopathological images. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR). pp. 7234-7242 (2017) \ No newline at end of file diff --git a/2024/Pathology-knowledge Enhanced Multi-instance Prompt Learning for Few-shot Whole Slide Image Classification/images.zip b/2024/Pathology-knowledge Enhanced Multi-instance Prompt Learning for Few-shot Whole Slide Image Classification/images.zip new file mode 100644 index 0000000000000000000000000000000000000000..744129584ee8e6635e104ff6f69efa0478ecfb7e --- /dev/null +++ b/2024/Pathology-knowledge Enhanced Multi-instance Prompt Learning for Few-shot Whole Slide Image Classification/images.zip @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:705bc3aed8437440e0c6dfc1876b95452ecf8d813f12fab0d0144bfba0f62287 +size 451679 diff --git a/2024/Pathology-knowledge Enhanced Multi-instance Prompt Learning for Few-shot Whole Slide Image Classification/layout.json b/2024/Pathology-knowledge Enhanced Multi-instance Prompt Learning for Few-shot Whole Slide Image Classification/layout.json new file mode 100644 index 0000000000000000000000000000000000000000..4b932d7c1af52b47e3727f3de2b7d165ae6af48d --- /dev/null +++ b/2024/Pathology-knowledge Enhanced Multi-instance Prompt Learning for Few-shot Whole Slide Image Classification/layout.json @@ -0,0 +1,8492 @@ +{ + "pdf_info": [ + { + "para_blocks": [ + { + "bbox": [ + 159, + 111, + 455, + 166 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 159, + 111, + 455, + 166 + ], + "spans": [ + { + "bbox": [ + 159, + 111, + 455, + 166 + ], + "type": "text", + "content": "Pathology-knowledge Enhanced Multi-instance Prompt Learning for Few-shot Whole Slide Image Classification" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 133, + 185, + 479, + 217 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 185, + 479, + 217 + ], + "spans": [ + { + "bbox": [ + 133, + 185, + 479, + 217 + ], + "type": "text", + "content": "Linhao Qu" + }, + { + "bbox": [ + 133, + 185, + 479, + 217 + ], + "type": "inline_equation", + "content": "^{1}" + }, + { + "bbox": [ + 133, + 185, + 479, + 217 + ], + "type": "text", + "content": ", Dingkang Yang" + }, + { + "bbox": [ + 133, + 185, + 479, + 217 + ], + "type": "inline_equation", + "content": "^{2}" + }, + { + "bbox": [ + 133, + 185, + 479, + 217 + ], + "type": "text", + "content": ", Dan Huang" + }, + { + "bbox": [ + 133, + 185, + 479, + 217 + ], + "type": "inline_equation", + "content": "^{3}" + }, + { + "bbox": [ + 133, + 185, + 479, + 217 + ], + "type": "text", + "content": ", Qinhao Guo" + }, + { + "bbox": [ + 133, + 185, + 479, + 217 + ], + "type": "inline_equation", + "content": "^{4}" + }, + { + "bbox": [ + 133, + 185, + 479, + 217 + ], + "type": "text", + "content": ", Rongkui Luo" + }, + { + "bbox": [ + 133, + 185, + 479, + 217 + ], + "type": "inline_equation", + "content": "^{5}" + }, + { + "bbox": [ + 133, + 185, + 479, + 217 + ], + "type": "text", + "content": ", Shaoting Zhang" + }, + { + "bbox": [ + 133, + 185, + 479, + 217 + ], + "type": "inline_equation", + "content": "^{1}" + }, + { + "bbox": [ + 133, + 185, + 479, + 217 + ], + "type": "text", + "content": ", and Xiaosong Wang" + }, + { + "bbox": [ + 133, + 185, + 479, + 217 + ], + "type": "inline_equation", + "content": "^{1\\boxtimes}" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 179, + 224, + 434, + 236 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 179, + 224, + 434, + 236 + ], + "spans": [ + { + "bbox": [ + 179, + 224, + 434, + 236 + ], + "type": "text", + "content": "1 Shanghai Artificial Intelligence Laboratory, Shanghai, China" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 144, + 236, + 469, + 247 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 144, + 236, + 469, + 247 + ], + "spans": [ + { + "bbox": [ + 144, + 236, + 469, + 247 + ], + "type": "inline_equation", + "content": "^{2}" + }, + { + "bbox": [ + 144, + 236, + 469, + 247 + ], + "type": "text", + "content": " Academy for Engineering and Technology, Fudan University, Shanghai, China" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 146, + 247, + 466, + 258 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 146, + 247, + 466, + 258 + ], + "spans": [ + { + "bbox": [ + 146, + 247, + 466, + 258 + ], + "type": "inline_equation", + "content": "^{3}" + }, + { + "bbox": [ + 146, + 247, + 466, + 258 + ], + "type": "text", + "content": " Department of Pathology, Fudan University Shanghai Cancer Center, China" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 137, + 258, + 476, + 269 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 137, + 258, + 476, + 269 + ], + "spans": [ + { + "bbox": [ + 137, + 258, + 476, + 269 + ], + "type": "inline_equation", + "content": "^{4}" + }, + { + "bbox": [ + 137, + 258, + 476, + 269 + ], + "type": "text", + "content": " Department of Gynecologic Oncology, Fudan University Shanghai Cancer Center" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 135, + 269, + 479, + 292 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 135, + 269, + 479, + 292 + ], + "spans": [ + { + "bbox": [ + 135, + 269, + 479, + 292 + ], + "type": "inline_equation", + "content": "^{5}" + }, + { + "bbox": [ + 135, + 269, + 479, + 292 + ], + "type": "text", + "content": " Department of Pathology, Zhongshan Hospital, Fudan University, Shanghai, China wangxiaosong@pjlab.org.cn;" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 159, + 316, + 455, + 579 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 159, + 316, + 455, + 579 + ], + "spans": [ + { + "bbox": [ + 159, + 316, + 455, + 579 + ], + "type": "text", + "content": "Abstract. Current multi-instance learning algorithms for pathology image analysis often require a substantial number of Whole Slide Images for effective training but exhibit suboptimal performance in scenarios with limited learning data. In clinical settings, restricted access to pathology slides is inevitable due to patient privacy concerns and the prevalence of rare or emerging diseases. The emergence of the Few-shot Weakly Supervised WSI Classification accommodates the significant challenge of the limited slide data and sparse slide-level labels for diagnosis. Prompt learning based on the pre-trained models (e.g., CLIP) appears to be a promising scheme for this setting; however, current research in this area is limited, and existing algorithms often focus solely on patch-level prompts or confine themselves to language prompts. This paper proposes a multi-instance prompt learning framework enhanced with pathology knowledge, i.e., integrating visual and textual prior knowledge into prompts at both patch and slide levels. The training process employs a combination of static and learnable prompts, effectively guiding the activation of pre-trained models and further facilitating the diagnosis of key pathology patterns. Lightweight Messenger (self-attention) and Summary (attention-pooling) layers are introduced to model relationships between patches and slides within the same patient data. Additionally, alignment-wise contrastive losses ensure the feature-level alignment between visual and textual learnable prompts for both patches and slides. Our method demonstrates superior performance in three challenging clinical tasks, significantly outperforming comparative few-shot methods." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 160, + 589, + 394, + 600 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 160, + 589, + 394, + 600 + ], + "spans": [ + { + "bbox": [ + 160, + 589, + 394, + 600 + ], + "type": "text", + "content": "Keywords: Pathology image analysis " + }, + { + "bbox": [ + 160, + 589, + 394, + 600 + ], + "type": "inline_equation", + "content": "\\cdot" + }, + { + "bbox": [ + 160, + 589, + 394, + 600 + ], + "type": "text", + "content": " Prompt learning" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 132, + 618, + 230, + 631 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 618, + 230, + 631 + ], + "spans": [ + { + "bbox": [ + 132, + 618, + 230, + 631 + ], + "type": "text", + "content": "1 Introduction" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 130, + 641, + 482, + 666 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 641, + 482, + 666 + ], + "spans": [ + { + "bbox": [ + 130, + 641, + 482, + 666 + ], + "type": "text", + "content": "Deep learning for pathology Whole Slide Image (WSI) classification offers a rapid, objective, and precise diagnostic approach, enhancing prognosis predic" + } + ] + } + ], + "index": 10 + } + ], + "discarded_blocks": [], + "page_size": [ + 612, + 792 + ], + "page_idx": 0 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 204, + 116, + 411, + 243 + ], + "blocks": [ + { + "bbox": [ + 204, + 116, + 411, + 243 + ], + "lines": [ + { + "bbox": [ + 204, + 116, + 411, + 243 + ], + "spans": [ + { + "bbox": [ + 204, + 116, + 411, + 243 + ], + "type": "image", + "image_path": "404a3d468aa48cec21122f2dfb34cb762b56fb207372dc71391b81cf728087ec.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 132, + 254, + 482, + 299 + ], + "lines": [ + { + "bbox": [ + 132, + 254, + 482, + 299 + ], + "spans": [ + { + "bbox": [ + 132, + 254, + 482, + 299 + ], + "type": "text", + "content": "Fig. 1: Existing methods such as CoOp [42] do not fully consider task-related pathology visual features and its association with specific terms in FSWC. We utilize aligned task-specific image examples and language descriptions to enhance visual and textual prompt learning at both patch and slide levels." + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_caption" + } + ], + "index": 2 + }, + { + "bbox": [ + 130, + 326, + 479, + 434 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 326, + 479, + 434 + ], + "spans": [ + { + "bbox": [ + 130, + 326, + 479, + 434 + ], + "type": "text", + "content": "tion and treatment response assessment [3, 4, 8, 13, 20, 26]. Given the gigapixel resolution of WSIs, segmentation into multiple patches is essential for processing. However, obtaining detailed annotations for each patch is challenging. Instead, it often results in slide-level ground truth labels. Multiple Instance Learning (MIL), a prominent weakly supervised paradigm, has emerged as a key technique for WSI classification [5, 9, 25, 30]. In MIL, each WSI is treated as a bag, with patches as instances. During training, only slide-level labels are available, necessitating a comprehensive model to capture relationships between patches and slide labels for accurate classification [28]." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 130, + 436, + 480, + 567 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 436, + 480, + 567 + ], + "spans": [ + { + "bbox": [ + 130, + 436, + 480, + 567 + ], + "type": "text", + "content": "MIL algorithms achieve slide-level classification by aggregating patch-level features or predictions. Though effective, existing algorithms typically demand a considerable number of slides for practical training, performing inadequately in scenarios with limited samples. In clinical practice, challenges such as preserving patient privacy, the rarity of certain diseases, and emerging conditions, particularly for prognosis and rare tumor subtypes, hinder the acquisition of a substantial quantity of WSIs [9,26,33]. Recently, the challenging problem of Few-shot weakly Supervised WSI Classification (FSWC) has emerged [27]. FSWC is characterized by the limited availability of training WSIs (e.g., typically 2, 4, 8, 16, or 32 per class) and sparse slide-level annotations, with unknown fine-grained annotations for numerous patches in each WSI." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 130, + 570, + 481, + 665 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 570, + 481, + 665 + ], + "spans": [ + { + "bbox": [ + 130, + 570, + 481, + 665 + ], + "type": "text", + "content": "In image classification, recent advancements highlight the effectiveness of Vision-Language Models (VLMs) like CLIP [29], BLIP [21], and Flamingo [1]. These models undergo training on large-scale language-image pairs using alignment contrastive learning, demonstrating notable zero-shot and few-shot transfer learning capabilities. In the pathology image domain, despite valuable contributions from initiatives like MI-Zero [24] and PLIP [14] in constructing VLMs for computational pathology, challenges including ethical concerns, annotation scarcity, and storage limitations make it impractical to train foundational mod-" + } + ] + } + ], + "index": 6 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "spans": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "type": "text", + "content": "2" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 166, + 91, + 215, + 102 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 166, + 91, + 215, + 102 + ], + "spans": [ + { + "bbox": [ + 166, + 91, + 215, + 102 + ], + "type": "text", + "content": "L. Qu et al." + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 1 + }, + { + "para_blocks": [ + { + "bbox": [ + 130, + 116, + 482, + 224 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 116, + 482, + 224 + ], + "spans": [ + { + "bbox": [ + 130, + 116, + 482, + 224 + ], + "type": "text", + "content": "els for each specific downstream task [27]. Furthermore, existing VLMs exhibit restricted direct zero-shot transfer capabilities for specific downstream tasks [7,36,38]. Considering these challenges, the application of existing VLMs for developing prompt learning algorithms tailored to few-shot downstream tasks appears promising. Prompt learning involves adapting a pre-trained and then fixed VLM to new tasks by integrating task-specific prompts as additional model inputs. In the context of few-shot learning, these prompts can be either fixed or trainable, with trainable prompts enabling contrastive learning between visual and language features [11,42]." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 130, + 224, + 482, + 368 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 224, + 482, + 368 + ], + "spans": [ + { + "bbox": [ + 130, + 224, + 482, + 368 + ], + "type": "text", + "content": "Presently, research on prompt learning for FSWC is limited [27]. Previous efforts, such as MI-Zero [24], PLIP [14], and CITE [41], have predominantly focused on the utilization of \"patch-level\" prompts—textual prompts associated with pathology regions of interest. Nonetheless, visual or textual prompts in domain-specific downstream tasks may not be understandable (often unseen) for the pre-trained model and thus may not properly activate the relevant features of the pre-trained model. The TOP framework [27] represents the initial endeavor to implement prompt learning by introducing language-based prior descriptions in the prompts. However, pathology images encompass highly specialized visual features and corresponding pathology-specific terms, posing a challenge for networks to comprehensively understand and learn pathology-related features and specialized descriptions solely through language explanations for classification." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 130, + 368, + 482, + 511 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 368, + 482, + 511 + ], + "spans": [ + { + "bbox": [ + 130, + 368, + 482, + 511 + ], + "type": "text", + "content": "To address this issue, we propose Pathology-knowledge Enhanced Multi-instance Prompt Learning for FSWC named PEMP, with a concise architecture depicted in Fig. 1. PEMP is a multi-instance prompt learning framework specifically designed for the FSWC problem, incorporating both visual and textual samples (as a form of prior knowledge) at both patch and slide levels. Drawing inspiration from pathologists learning pathology images and language knowledge from textbooks, we introduce task-specific image examples and corresponding language descriptions, injected in both the visual and textual prompts. During prompt learning, the model is fed with a more understandable and effective guide (in image examples and plain words) and then activates more meaningful features for the diagnosis of key patch-level and slide-level pathology patterns in scenarios with limited training samples." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 146, + 512, + 333, + 523 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 146, + 512, + 333, + 523 + ], + "spans": [ + { + "bbox": [ + 146, + 512, + 333, + 523 + ], + "type": "text", + "content": "The contribution of this work is threefold:" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 138, + 533, + 481, + 666 + ], + "type": "list", + "angle": 0, + "index": 9, + "blocks": [ + { + "bbox": [ + 138, + 533, + 481, + 593 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 533, + 481, + 593 + ], + "spans": [ + { + "bbox": [ + 138, + 533, + 481, + 593 + ], + "type": "text", + "content": "- We propose a multi-instance prompt learning framework, PEMP, introducing vision and text prior knowledge in the designed prompts at both patch and slide levels. A combination of static and learnable prompts is employed in the training, effectively adapting the pre-trained model to focus on key pathology patterns in patches and slides." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 138, + 593, + 481, + 641 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 593, + 481, + 641 + ], + "spans": [ + { + "bbox": [ + 138, + 593, + 481, + 641 + ], + "type": "text", + "content": "- We explore the alignment of visual and textual prompts using contrastive losses at both patch and slide levels. We design lightweight Messenger (self-attention) and Summary (attention-pooling) layers to model relationships of patch-to-patch and patch-to-slide within the same patient data." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 138, + 641, + 481, + 666 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 641, + 481, + 666 + ], + "spans": [ + { + "bbox": [ + 138, + 641, + 481, + 666 + ], + "type": "text", + "content": "- We evaluate PEMP's performance in addressing three challenging pathology tasks in clinical practice, including predicting patient survival and lymph" + } + ] + } + ], + "index": 8 + } + ], + "sub_type": "text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 184, + 91, + 447, + 103 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 184, + 91, + 447, + 103 + ], + "spans": [ + { + "bbox": [ + 184, + 91, + 447, + 103 + ], + "type": "text", + "content": "Pathology-knowledge Enhanced Multi-instance Prompt Learning" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 474, + 92, + 480, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 474, + 92, + 480, + 100 + ], + "spans": [ + { + "bbox": [ + 474, + 92, + 480, + 100 + ], + "type": "text", + "content": "3" + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 2 + }, + { + "para_blocks": [ + { + "bbox": [ + 148, + 116, + 482, + 165 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 148, + 116, + 482, + 165 + ], + "spans": [ + { + "bbox": [ + 148, + 116, + 482, + 165 + ], + "type": "text", + "content": "node metastasis in early cervical cancer H&E slides and subtype classification for a rare tumor (round cell tumor). PEMP achieves superior performance in all these tasks under the few-shot setting, notably outperforming all comparative methods by a large margin (4% on average)." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 132, + 183, + 238, + 196 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 183, + 238, + 196 + ], + "spans": [ + { + "bbox": [ + 132, + 183, + 238, + 196 + ], + "type": "text", + "content": "2 Related Work" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 132, + 209, + 297, + 221 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 209, + 297, + 221 + ], + "spans": [ + { + "bbox": [ + 132, + 209, + 297, + 221 + ], + "type": "text", + "content": "2.1 MIL for WSI Classification" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 130, + 231, + 482, + 363 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 231, + 482, + 363 + ], + "spans": [ + { + "bbox": [ + 130, + 231, + 482, + 363 + ], + "type": "text", + "content": "Current MIL algorithms for WSI classification can be broadly categorized as instance-based [2, 10, 22, 28, 37] and bag-based [6, 12, 16, 18, 19, 23, 25, 31, 32, 34, 35, 39, 40, 43] methods. Instance-based methods involve generating pseudo-labels for each patch and training patch classifiers to score all patches within a slide, whereas bag-based methods, the current mainstream, extract patch features and employ techniques like attention weighting [12, 16, 18, 25, 32, 39, 40, 43] to aggregate these features into slide-level features. While existing algorithms perform well with ample training data, they lack specificity for FSWC tasks, struggling with limited samples. Our proposed PEMP, falling under the bag-based MIL paradigm, distinguishes itself by innovatively addressing FSWC through VLMs and prompt learning, deviating from traditional bag-based MIL approaches." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 132, + 380, + 313, + 393 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 380, + 313, + 393 + ], + "spans": [ + { + "bbox": [ + 132, + 380, + 313, + 393 + ], + "type": "text", + "content": "2.2 VLM-Based Prompt Learning" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 130, + 402, + 482, + 521 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 402, + 482, + 521 + ], + "spans": [ + { + "bbox": [ + 130, + 402, + 482, + 521 + ], + "type": "text", + "content": "VLMs, comprising image and text encoders, are trained on large-scale image-text pairs using alignment contrastive learning to establish a shared feature space. Pre-trained VLMs like CLIP [29], BLIP [21], and Flamingo [1] exhibit robust zero-shot and few-shot transferability in image recognition tasks. However, limited accuracy in zero-shot transfer has fueled interest in few-shot learning based on VLMs, where the prompt learning paradigm, demonstrating strong performance [7,36,38], is emerging. Prompt learning leverages a pre-trained VLM in the adaptation to new downstream tasks by incorporating task-specific prompts through enriched model inputs. These prompts could be static with manual designs or learnable vectors [11,42]." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 130, + 522, + 482, + 666 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 522, + 482, + 666 + ], + "spans": [ + { + "bbox": [ + 130, + 522, + 482, + 666 + ], + "type": "text", + "content": "In pathology, recent research emphasizes constructing specialized VLMs, such as MI-Zero [24], PLIP [14], and QUILT [15]. However, privacy concerns, annotation scarcity, and storage demands make training separate foundational models for each downstream task impractical. Existing VLMs exhibit limited zero-shot transferability, especially for rare diseases. Unlike fine-tuning VLMs at the data level or constructing large datasets like these works, we introduce a novel prompt learning paradigm based on VLMs (e.g., CLIP with frozen parameters) to address the FSWC problem. Importantly, our scheme is versatile across various foundation models. Moreover, MI-Zero, PLIP and QUILT mainly contribute by collecting large datasets to construct pathology VLMs but lack effective methods for FSWC. Their simple adapting strategies for FSWC (e.g., LinearProbe and Attention Pooling) limit their performance in challenging clinical tasks." + } + ] + } + ], + "index": 8 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "spans": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "type": "text", + "content": "4" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 166, + 91, + 215, + 102 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 166, + 91, + 215, + 102 + ], + "spans": [ + { + "bbox": [ + 166, + 91, + 215, + 102 + ], + "type": "text", + "content": "L. Qu et al." + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 3 + }, + { + "para_blocks": [ + { + "bbox": [ + 130, + 116, + 482, + 224 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 116, + 482, + 224 + ], + "spans": [ + { + "bbox": [ + 130, + 116, + 482, + 224 + ], + "type": "text", + "content": "In pathology, prompt learning research for FSWC is limited. Prior works like MI-Zero [24], PLIP [14], and CITE [41] only explore \"patch-level\" prompts, limiting slide-level classification performance. TOP [27] introduces prompt learning at both patch and slide levels but lacks task-specific visual knowledge on the vision side. Relying solely on language explanations is challenging for the networks to understand pathology features with limited training samples. In contrast, PEMP's learning process operates at patch and slide levels, covers visual and textual prompt design, and introduces task-specific image examples and descriptions separately as prior knowledge." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 131, + 241, + 318, + 254 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 131, + 241, + 318, + 254 + ], + "spans": [ + { + "bbox": [ + 131, + 241, + 318, + 254 + ], + "type": "text", + "content": "3 FSWC Baseline Framework" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 131, + 265, + 268, + 276 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 131, + 265, + 268, + 276 + ], + "spans": [ + { + "bbox": [ + 131, + 265, + 268, + 276 + ], + "type": "text", + "content": "3.1 Problem Formulation" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 130, + 284, + 482, + 440 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 284, + 482, + 440 + ], + "spans": [ + { + "bbox": [ + 130, + 284, + 482, + 440 + ], + "type": "text", + "content": "We provide a brief introduction to the problem definition of FSWC and the pre-trained VLM used in our study. Given a dataset " + }, + { + "bbox": [ + 130, + 284, + 482, + 440 + ], + "type": "inline_equation", + "content": "W = \\{W_{1}, W_{2}, \\ldots, W_{N}\\}" + }, + { + "bbox": [ + 130, + 284, + 482, + 440 + ], + "type": "text", + "content": " containing " + }, + { + "bbox": [ + 130, + 284, + 482, + 440 + ], + "type": "inline_equation", + "content": "N" + }, + { + "bbox": [ + 130, + 284, + 482, + 440 + ], + "type": "text", + "content": " WSIs, each WSI " + }, + { + "bbox": [ + 130, + 284, + 482, + 440 + ], + "type": "inline_equation", + "content": "W_{i}" + }, + { + "bbox": [ + 130, + 284, + 482, + 440 + ], + "type": "text", + "content": " is divided into non-overlapping small image patches " + }, + { + "bbox": [ + 130, + 284, + 482, + 440 + ], + "type": "inline_equation", + "content": "\\{p_{i,j}, j = 1, 2, \\ldots, n_{i}\\}" + }, + { + "bbox": [ + 130, + 284, + 482, + 440 + ], + "type": "text", + "content": ", where " + }, + { + "bbox": [ + 130, + 284, + 482, + 440 + ], + "type": "inline_equation", + "content": "n_{i}" + }, + { + "bbox": [ + 130, + 284, + 482, + 440 + ], + "type": "text", + "content": " represents the number of patches obtained from " + }, + { + "bbox": [ + 130, + 284, + 482, + 440 + ], + "type": "inline_equation", + "content": "W_{i}" + }, + { + "bbox": [ + 130, + 284, + 482, + 440 + ], + "type": "text", + "content": ". It is important to note that the number of patches obtained from each slide may vary. All patches in " + }, + { + "bbox": [ + 130, + 284, + 482, + 440 + ], + "type": "inline_equation", + "content": "W_{i}" + }, + { + "bbox": [ + 130, + 284, + 482, + 440 + ], + "type": "text", + "content": " collectively form a bag, and each patch serves as an instance of this bag. Each bag is assigned a label " + }, + { + "bbox": [ + 130, + 284, + 482, + 440 + ], + "type": "inline_equation", + "content": "Y_{i} \\in \\{0, 1, \\ldots, U\\}" + }, + { + "bbox": [ + 130, + 284, + 482, + 440 + ], + "type": "text", + "content": ", while the label of each instance is unknown, and " + }, + { + "bbox": [ + 130, + 284, + 482, + 440 + ], + "type": "inline_equation", + "content": "U" + }, + { + "bbox": [ + 130, + 284, + 482, + 440 + ], + "type": "text", + "content": " is the number of classes. FSWC is highly challenging as it allows training with only a limited number of slides (with labels). In the FSWC task, \"shot\" refers to the number of WSIs, e.g., " + }, + { + "bbox": [ + 130, + 284, + 482, + 440 + ], + "type": "inline_equation", + "content": "K" + }, + { + "bbox": [ + 130, + 284, + 482, + 440 + ], + "type": "text", + "content": " pairs of WSIs and labels for training in a " + }, + { + "bbox": [ + 130, + 284, + 482, + 440 + ], + "type": "inline_equation", + "content": "K" + }, + { + "bbox": [ + 130, + 284, + 482, + 440 + ], + "type": "text", + "content": "-shot setting, while the trained model is evaluated on a completely independent test set. Typically, " + }, + { + "bbox": [ + 130, + 284, + 482, + 440 + ], + "type": "inline_equation", + "content": "K" + }, + { + "bbox": [ + 130, + 284, + 482, + 440 + ], + "type": "text", + "content": " can take values such as 2, 4, 8, 16, or 32." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 131, + 456, + 389, + 468 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 131, + 456, + 389, + 468 + ], + "spans": [ + { + "bbox": [ + 131, + 456, + 389, + 468 + ], + "type": "text", + "content": "3.2 Attention Aggregation for WSI Classification" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 130, + 475, + 482, + 536 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 475, + 482, + 536 + ], + "spans": [ + { + "bbox": [ + 130, + 475, + 482, + 536 + ], + "type": "text", + "content": "The attention-based pooling strategy [16] is a classical MIL aggregation method used for WSI classification. Initially, an encoder " + }, + { + "bbox": [ + 130, + 475, + 482, + 536 + ], + "type": "inline_equation", + "content": "\\phi_{img}" + }, + { + "bbox": [ + 130, + 475, + 482, + 536 + ], + "type": "text", + "content": " is used to extract features " + }, + { + "bbox": [ + 130, + 475, + 482, + 536 + ], + "type": "inline_equation", + "content": "f_{i,j}" + }, + { + "bbox": [ + 130, + 475, + 482, + 536 + ], + "type": "text", + "content": " for all patches " + }, + { + "bbox": [ + 130, + 475, + 482, + 536 + ], + "type": "inline_equation", + "content": "\\{p_{i,j}, j = 1,2,\\dots n_i\\}" + }, + { + "bbox": [ + 130, + 475, + 482, + 536 + ], + "type": "text", + "content": " in slide " + }, + { + "bbox": [ + 130, + 475, + 482, + 536 + ], + "type": "inline_equation", + "content": "W_{i}" + }, + { + "bbox": [ + 130, + 475, + 482, + 536 + ], + "type": "text", + "content": ", and then aggregate these patch features using an attention-based trainable aggregation function " + }, + { + "bbox": [ + 130, + 475, + 482, + 536 + ], + "type": "inline_equation", + "content": "A(\\cdot)" + }, + { + "bbox": [ + 130, + 475, + 482, + 536 + ], + "type": "text", + "content": " to obtain the bag feature " + }, + { + "bbox": [ + 130, + 475, + 482, + 536 + ], + "type": "inline_equation", + "content": "F_{i}" + }, + { + "bbox": [ + 130, + 475, + 482, + 536 + ], + "type": "text", + "content": ":" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 200, + 544, + 481, + 576 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 200, + 544, + 481, + 576 + ], + "spans": [ + { + "bbox": [ + 200, + 544, + 481, + 576 + ], + "type": "interline_equation", + "content": "F _ {i} = A \\left(\\phi_ {i m g} \\left(p _ {i, 1}\\right), \\phi_ {i m g} \\left(p _ {i, 2}\\right), \\dots\\right) = \\sum_ {j = 1} ^ {n _ {i}} a _ {i, j} f _ {i, j}, \\tag {1}", + "image_path": "04b363f6c21c61b1dc2212966b09a9c22c3dcc2489a5661141e2e8eb1d74f7c3.jpg" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 228, + 584, + 481, + 615 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 228, + 584, + 481, + 615 + ], + "spans": [ + { + "bbox": [ + 228, + 584, + 481, + 615 + ], + "type": "interline_equation", + "content": "a _ {i, j} = \\frac {\\exp \\left\\{w ^ {\\top} \\tanh \\left(V f _ {i , j} ^ {\\top}\\right) \\right\\}}{\\sum_ {j = 1} ^ {n _ {i}} \\exp \\left\\{w ^ {\\top} \\tanh \\left(V f _ {i , j} ^ {\\top}\\right) \\right\\}}, \\tag {2}", + "image_path": "1f90d95fab5b32e0b1ead7925327113e417b3585dd58fb74d461121ae9d93ed9.jpg" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 130, + 617, + 482, + 666 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 617, + 482, + 666 + ], + "spans": [ + { + "bbox": [ + 130, + 617, + 482, + 666 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 130, + 617, + 482, + 666 + ], + "type": "inline_equation", + "content": "a_{i,j}" + }, + { + "bbox": [ + 130, + 617, + 482, + 666 + ], + "type": "text", + "content": " is the attention score predicted by the self-attention network, which is parameterized by " + }, + { + "bbox": [ + 130, + 617, + 482, + 666 + ], + "type": "inline_equation", + "content": "w" + }, + { + "bbox": [ + 130, + 617, + 482, + 666 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 130, + 617, + 482, + 666 + ], + "type": "inline_equation", + "content": "V" + }, + { + "bbox": [ + 130, + 617, + 482, + 666 + ], + "type": "text", + "content": ". The weights " + }, + { + "bbox": [ + 130, + 617, + 482, + 666 + ], + "type": "inline_equation", + "content": "a_{i,j}" + }, + { + "bbox": [ + 130, + 617, + 482, + 666 + ], + "type": "text", + "content": " reflect the contribution of each patch in making the slide-level prediction. The patch-level encoder " + }, + { + "bbox": [ + 130, + 617, + 482, + 666 + ], + "type": "inline_equation", + "content": "\\phi_{img}" + }, + { + "bbox": [ + 130, + 617, + 482, + 666 + ], + "type": "text", + "content": " is from the pre-trained CLIP [29] and remains fixed during the training process." + } + ] + } + ], + "index": 10 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 184, + 91, + 448, + 103 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 184, + 91, + 448, + 103 + ], + "spans": [ + { + "bbox": [ + 184, + 91, + 448, + 103 + ], + "type": "text", + "content": "Pathology-knowledge Enhanced Multi-instance Prompt Learning" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 474, + 92, + 480, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 474, + 92, + 480, + 100 + ], + "spans": [ + { + "bbox": [ + 474, + 92, + 480, + 100 + ], + "type": "text", + "content": "5" + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 4 + }, + { + "para_blocks": [ + { + "bbox": [ + 132, + 116, + 400, + 129 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 116, + 400, + 129 + ], + "spans": [ + { + "bbox": [ + 132, + 116, + 400, + 129 + ], + "type": "text", + "content": "3.3 VLM backbone and Few-shot Prompt Learning" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 130, + 134, + 479, + 181 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 134, + 479, + 181 + ], + "spans": [ + { + "bbox": [ + 130, + 134, + 479, + 181 + ], + "type": "text", + "content": "We utilize CLIP [29] as the backbone of the proposed framework and also the pretrained VLM for feature extraction, which is trained on a massive dataset of 400 million image-text pairs, including medical data, facilitating efficient knowledge transfer of diverse visual concepts to downstream tasks." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 130, + 182, + 480, + 327 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 182, + 480, + 327 + ], + "spans": [ + { + "bbox": [ + 130, + 182, + 480, + 327 + ], + "type": "text", + "content": "Prompt learning enables the efficient adaptation of a fixed pre-trained CLIP in few-shot downstream classification tasks by introducing task-specific prompts into the model input. These prompts, whether fixed or learnable, can be applied to the text encoder, visual encoder, or both simultaneously. In the few-shot setting, the training could also involve aligning visual and text features through alignment contrastive learning (AC-Loss) in the specific domain. Formally, we denote the visual encoder as " + }, + { + "bbox": [ + 130, + 182, + 480, + 327 + ], + "type": "inline_equation", + "content": "\\phi_{img}" + }, + { + "bbox": [ + 130, + 182, + 480, + 327 + ], + "type": "text", + "content": " and the text encoder as " + }, + { + "bbox": [ + 130, + 182, + 480, + 327 + ], + "type": "inline_equation", + "content": "\\phi_{text}" + }, + { + "bbox": [ + 130, + 182, + 480, + 327 + ], + "type": "text", + "content": ". For a downstream classification task with " + }, + { + "bbox": [ + 130, + 182, + 480, + 327 + ], + "type": "inline_equation", + "content": "U" + }, + { + "bbox": [ + 130, + 182, + 480, + 327 + ], + "type": "text", + "content": " classes, a commonly used fixed hand-crafted prompt, denoted as [Class Prompt], guides classification on the text side. Specifically, assuming the name of the " + }, + { + "bbox": [ + 130, + 182, + 480, + 327 + ], + "type": "inline_equation", + "content": "i" + }, + { + "bbox": [ + 130, + 182, + 480, + 327 + ], + "type": "text", + "content": "-th class is “[class-name],\" the corresponding hand-crafted prompt is typically designed as “a photo of a [class-name].” These [Class Prompts] are pre-encoded into a vector " + }, + { + "bbox": [ + 130, + 182, + 480, + 327 + ], + "type": "inline_equation", + "content": "C = \\{c_i\\}_{i=1}^U" + }, + { + "bbox": [ + 130, + 182, + 480, + 327 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 130, + 327, + 480, + 434 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 327, + 480, + 434 + ], + "spans": [ + { + "bbox": [ + 130, + 327, + 480, + 434 + ], + "type": "text", + "content": "To further enhance generalization to downstream tasks, it is also popular to explore learning a set of continuous context vectors as learnable prompts, denoted as " + }, + { + "bbox": [ + 130, + 327, + 480, + 434 + ], + "type": "inline_equation", + "content": "\\mathbb{V} = \\{v_1,v_2,\\ldots ,v_M\\}" + }, + { + "bbox": [ + 130, + 327, + 480, + 434 + ], + "type": "text", + "content": ". These learnable context vectors " + }, + { + "bbox": [ + 130, + 327, + 480, + 434 + ], + "type": "inline_equation", + "content": "\\mathbb{V}" + }, + { + "bbox": [ + 130, + 327, + 480, + 434 + ], + "type": "text", + "content": ", concatenated with the vectors from [Class Prompt] " + }, + { + "bbox": [ + 130, + 327, + 480, + 434 + ], + "type": "inline_equation", + "content": "C" + }, + { + "bbox": [ + 130, + 327, + 480, + 434 + ], + "type": "text", + "content": ", form the complete text prompt " + }, + { + "bbox": [ + 130, + 327, + 480, + 434 + ], + "type": "inline_equation", + "content": "t_i = \\{v_1,v_2,\\dots ,v_M,c_i\\}" + }, + { + "bbox": [ + 130, + 327, + 480, + 434 + ], + "type": "text", + "content": ". This prompt is then fed into the text encoder " + }, + { + "bbox": [ + 130, + 327, + 480, + 434 + ], + "type": "inline_equation", + "content": "\\phi_{text}" + }, + { + "bbox": [ + 130, + 327, + 480, + 434 + ], + "type": "text", + "content": " to obtain the final textual features for each class, " + }, + { + "bbox": [ + 130, + 327, + 480, + 434 + ], + "type": "inline_equation", + "content": "T_{i} = \\phi_{text}(t_{i})" + }, + { + "bbox": [ + 130, + 327, + 480, + 434 + ], + "type": "text", + "content": ". On the visual side, given an image " + }, + { + "bbox": [ + 130, + 327, + 480, + 434 + ], + "type": "inline_equation", + "content": "W_{i}" + }, + { + "bbox": [ + 130, + 327, + 480, + 434 + ], + "type": "text", + "content": " and its label " + }, + { + "bbox": [ + 130, + 327, + 480, + 434 + ], + "type": "inline_equation", + "content": "y" + }, + { + "bbox": [ + 130, + 327, + 480, + 434 + ], + "type": "text", + "content": ", its slide features " + }, + { + "bbox": [ + 130, + 327, + 480, + 434 + ], + "type": "inline_equation", + "content": "F_{i}" + }, + { + "bbox": [ + 130, + 327, + 480, + 434 + ], + "type": "text", + "content": " is computed via Eq. (1). Similar to the text side, learnable prompts could also be appended, which will be inputted into the visual encoder " + }, + { + "bbox": [ + 130, + 327, + 480, + 434 + ], + "type": "inline_equation", + "content": "\\phi_{img}" + }, + { + "bbox": [ + 130, + 327, + 480, + 434 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 130, + 435, + 481, + 495 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 435, + 481, + 495 + ], + "spans": [ + { + "bbox": [ + 130, + 435, + 481, + 495 + ], + "type": "text", + "content": "With a limited number of image-text pairs, the optimization involves minimizing the negative log-likelihood between the visual slide feature " + }, + { + "bbox": [ + 130, + 435, + 481, + 495 + ], + "type": "inline_equation", + "content": "F_{i}" + }, + { + "bbox": [ + 130, + 435, + 481, + 495 + ], + "type": "text", + "content": " and its class text feature " + }, + { + "bbox": [ + 130, + 435, + 481, + 495 + ], + "type": "inline_equation", + "content": "T_{y}" + }, + { + "bbox": [ + 130, + 435, + 481, + 495 + ], + "type": "text", + "content": ", updating all learnable parameters. During testing, the extracted visual features " + }, + { + "bbox": [ + 130, + 435, + 481, + 495 + ], + "type": "inline_equation", + "content": "F_{i}" + }, + { + "bbox": [ + 130, + 435, + 481, + 495 + ], + "type": "text", + "content": " are used to retrieve the class text features " + }, + { + "bbox": [ + 130, + 435, + 481, + 495 + ], + "type": "inline_equation", + "content": "T_{y}" + }, + { + "bbox": [ + 130, + 435, + 481, + 495 + ], + "type": "text", + "content": " to obtain the probabilities for each class." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 224, + 502, + 481, + 531 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 224, + 502, + 481, + 531 + ], + "spans": [ + { + "bbox": [ + 224, + 502, + 481, + 531 + ], + "type": "interline_equation", + "content": "p (y \\mid F _ {i}) = \\frac {\\exp (\\operatorname {s i m} \\left(F _ {i} , T _ {y}\\right) / \\tau)}{\\sum_ {i = 1} ^ {U} \\exp (\\operatorname {s i m} \\left(F _ {i} , T _ {i}\\right) / \\tau)}, \\tag {3}", + "image_path": "44b35b646dc43510ae1918f1f2edeb0869457cf3f7e96be212b2b0f7c50e9efc.jpg" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 130, + 535, + 482, + 584 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 535, + 482, + 584 + ], + "spans": [ + { + "bbox": [ + 130, + 535, + 482, + 584 + ], + "type": "text", + "content": "where, " + }, + { + "bbox": [ + 130, + 535, + 482, + 584 + ], + "type": "inline_equation", + "content": "y" + }, + { + "bbox": [ + 130, + 535, + 482, + 584 + ], + "type": "text", + "content": " represents the ground truth class, " + }, + { + "bbox": [ + 130, + 535, + 482, + 584 + ], + "type": "inline_equation", + "content": "U" + }, + { + "bbox": [ + 130, + 535, + 482, + 584 + ], + "type": "text", + "content": " is the total number of classes, " + }, + { + "bbox": [ + 130, + 535, + 482, + 584 + ], + "type": "inline_equation", + "content": "\\mathrm{sim}(\\cdot)" + }, + { + "bbox": [ + 130, + 535, + 482, + 584 + ], + "type": "text", + "content": " denotes cosine similarity, and " + }, + { + "bbox": [ + 130, + 535, + 482, + 584 + ], + "type": "inline_equation", + "content": "\\tau" + }, + { + "bbox": [ + 130, + 535, + 482, + 584 + ], + "type": "text", + "content": " is a learnable temperature parameter. It is crucial to note that during the training process, the parameters of CLIP's image encoder and text encoder remain fixed and unchanged." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 132, + 599, + 386, + 615 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 599, + 386, + 615 + ], + "spans": [ + { + "bbox": [ + 132, + 599, + 386, + 615 + ], + "type": "text", + "content": "4 Knowledge-enhanced Prompt Learning" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 132, + 623, + 208, + 635 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 623, + 208, + 635 + ], + "spans": [ + { + "bbox": [ + 132, + 623, + 208, + 635 + ], + "type": "text", + "content": "4.1 Overview" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 130, + 641, + 481, + 666 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 641, + 481, + 666 + ], + "spans": [ + { + "bbox": [ + 130, + 641, + 481, + 666 + ], + "type": "text", + "content": "Fig. 2 illustrates the primary workflow of our proposed PEMP, which comprises three learning processes: Visual prompt learning, Textual prompt learning, and" + } + ] + } + ], + "index": 11 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "spans": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "type": "text", + "content": "6" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 166, + 91, + 215, + 102 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 166, + 91, + 215, + 102 + ], + "spans": [ + { + "bbox": [ + 166, + 91, + 215, + 102 + ], + "type": "text", + "content": "L. Qu et al." + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 5 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 137, + 114, + 479, + 352 + ], + "blocks": [ + { + "bbox": [ + 137, + 114, + 479, + 352 + ], + "lines": [ + { + "bbox": [ + 137, + 114, + 479, + 352 + ], + "spans": [ + { + "bbox": [ + 137, + 114, + 479, + 352 + ], + "type": "image", + "image_path": "f73bce66b1b82e2f233c4d1f4045d8cd1037ffd1aca6f64d147ac22b37c9c289.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 132, + 361, + 482, + 384 + ], + "lines": [ + { + "bbox": [ + 132, + 361, + 482, + 384 + ], + "spans": [ + { + "bbox": [ + 132, + 361, + 482, + 384 + ], + "type": "text", + "content": "Fig. 2: Overview of PEMP, where flames represent optimized parameters, and snowflakes indicate frozen parameters during training." + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_caption" + } + ], + "index": 2 + }, + { + "bbox": [ + 130, + 417, + 482, + 548 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 417, + 482, + 548 + ], + "spans": [ + { + "bbox": [ + 130, + 417, + 482, + 548 + ], + "type": "text", + "content": "Two-level prompt alignment via contrastive learning. In visual prompt learning, we address three key tasks. Firstly, we integrate prior visual pathology knowledge into the prompts at both patch-level and slide-level. This guides the model, particularly with limited training samples, to understand the prompt better and produce more meaningful activations for task-specific key pathology patterns. Additionally, we also include the learnable prompts on the slide level, triggering the model's precise adaptation to specific tasks. Thirdly, we introduce a lightweight Messenger (self-attention) Layer and Summary (attention-pooling) Layer. The former efficiently communicates and models features of different patches within the same slide, while the latter aggregates patch features into slide features using an attention mechanism." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 130, + 553, + 482, + 615 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 553, + 482, + 615 + ], + "spans": [ + { + "bbox": [ + 130, + 553, + 482, + 615 + ], + "type": "text", + "content": "In a similar initiative, textual prompt learning concentrates on modeling language descriptions related to pathology tasks, also incorporating patch-level and slide-level priors. Pathologists are employed to provide straightforward language descriptions in the overall task, task-related slide-level pathology, and task-related patch-level pathology, avoiding obscure professional terms." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 130, + 617, + 482, + 666 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 617, + 482, + 666 + ], + "spans": [ + { + "bbox": [ + 130, + 617, + 482, + 666 + ], + "type": "text", + "content": "Recognizing the challenges faced by the VLM in understanding key visual features and complex pathology terms in visual examples, we establish a close alignment between textual prompt learning (with integrated pathology language descriptions) and visual prompt learning (with injected visual examples) at both" + } + ] + } + ], + "index": 6 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 184, + 91, + 448, + 102 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 184, + 91, + 448, + 102 + ], + "spans": [ + { + "bbox": [ + 184, + 91, + 448, + 102 + ], + "type": "text", + "content": "Pathology-knowledge Enhanced Multi-instance Prompt Learning" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 474, + 92, + 480, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 474, + 92, + 480, + 100 + ], + "spans": [ + { + "bbox": [ + 474, + 92, + 480, + 100 + ], + "type": "text", + "content": "7" + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 6 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 133, + 114, + 297, + 178 + ], + "blocks": [ + { + "bbox": [ + 133, + 114, + 297, + 178 + ], + "lines": [ + { + "bbox": [ + 133, + 114, + 297, + 178 + ], + "spans": [ + { + "bbox": [ + 133, + 114, + 297, + 178 + ], + "type": "image", + "image_path": "6db8b8b77b6bd76bff780a10a8034b7e1e89c17c9ffa2f89887e3cd128f3d194.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 132, + 179, + 291, + 228 + ], + "lines": [ + { + "bbox": [ + 132, + 179, + 291, + 228 + ], + "spans": [ + { + "bbox": [ + 132, + 179, + 291, + 228 + ], + "type": "text", + "content": "(a) Examples for predicting the prognosis of early cervical cancer patients from H&E pathology slides of the primary lesion, derived from authoritative textbooks and pathology experts. Top: visual patch examples. Bottom: visual slide examples." + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_footnote" + }, + { + "bbox": [ + 130, + 245, + 482, + 269 + ], + "lines": [ + { + "bbox": [ + 130, + 245, + 482, + 269 + ], + "spans": [ + { + "bbox": [ + 130, + 245, + 482, + 269 + ], + "type": "text", + "content": "Fig. 3: (A) and (B) examples of visual and textual token constructions related to the prognosis of early cervical cancer patients." + } + ] + } + ], + "index": 6, + "angle": 0, + "type": "image_caption" + } + ], + "index": 2 + }, + { + "type": "image", + "bbox": [ + 324, + 114, + 484, + 178 + ], + "blocks": [ + { + "bbox": [ + 324, + 114, + 484, + 178 + ], + "lines": [ + { + "bbox": [ + 324, + 114, + 484, + 178 + ], + "spans": [ + { + "bbox": [ + 324, + 114, + 484, + 178 + ], + "type": "image", + "image_path": "442e7c81a69bbfc81aba21262bee527b54204d7880347cf5c89af58349d7a4bf.jpg" + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 322, + 179, + 482, + 237 + ], + "lines": [ + { + "bbox": [ + 322, + 179, + 482, + 237 + ], + "spans": [ + { + "bbox": [ + 322, + 179, + 482, + 237 + ], + "type": "text", + "content": "(b) Top: Construction of Patch-level descriptive [Token]. Bottom: Construction of Slide-level descriptive [Token]. Examples for the prognosis of early cervical cancer patients from H&E pathology slides of the primary lesion, serving as task-related patch-level and slide-level pathology priors on the text side." + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "image_footnote" + } + ], + "index": 4 + }, + { + "bbox": [ + 130, + 293, + 481, + 319 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 293, + 481, + 319 + ], + "spans": [ + { + "bbox": [ + 130, + 293, + 481, + 319 + ], + "type": "text", + "content": "slide and patch levels. This alignment enhances the organization and cohesion of knowledge-enhanced prompts on both visual and textual sides." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 132, + 341, + 283, + 354 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 341, + 283, + 354 + ], + "spans": [ + { + "bbox": [ + 132, + 341, + 283, + 354 + ], + "type": "text", + "content": "4.2 Visual Prompt Learning" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 130, + 365, + 482, + 581 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 365, + 482, + 581 + ], + "spans": [ + { + "bbox": [ + 130, + 365, + 482, + 581 + ], + "type": "text", + "content": "Construction of visual prompts. In the FSWC task, limited training samples present challenges in acquiring effective knowledge. Moreover, tasks like survival analysis require finding key factors both locally (e.g., patch-level: necrosis, vascular invasion) and globally (e.g., slide-level: tumor boundary, tumor stroma ratio). To overcome these, we out-source prior visual pathology knowledge at both patch and slide levels (shown in the top part of Fig. 3a), guiding the model with task-specific pathology patterns for diagnosis. For challenging slide-level diagnostic tasks, relevant patterns manifest at both levels. Leveraging expertise from pathology professionals, we construct visual patch and slide examples as fixed prompts, as shown in Fig. 3a. For each classification task, highly relevant visual patches and slides indicating high and low risks are retrieved from authoritative sources or pathology experts. For instance, in the prognosis of early cervical cancer, visual patch examples for poor prognosis include \"high-grade atypia\", \"vascular invasion,\" and \"necrosis,\" while slide examples encompass \"overall infiltration with fuzzy tumor boundaries and low tumor stroma ratio.\" Conversely, low-risk examples include \"low-grade dysplasia,\" \"no vascular invasion,\" and \"overall pushing-type with clear tumor boundaries and high tumor stroma ratio.\"" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 130, + 582, + 482, + 666 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 582, + 482, + 666 + ], + "spans": [ + { + "bbox": [ + 130, + 582, + 482, + 666 + ], + "type": "text", + "content": "In total, 6 slides and 6 patch images are used as visual prompts, including 3 typical slides per prognosis category (good and poor), each with 3 local patches in the prognosis of early cervical cancer. For the Lymph Node Metastasis Prediction Task, we used 4 slides and 4 patches, with two representative slides per category (metastasis and non-metastasis) and two local images per slide. For Round Cell Subtype Diagnosis, 10 slides and 10 patches were used, with two slides per category, each with two local images. Selection of these visual exam" + } + ] + } + ], + "index": 10 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "spans": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "type": "text", + "content": "8" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 166, + 91, + 215, + 102 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 166, + 91, + 215, + 102 + ], + "spans": [ + { + "bbox": [ + 166, + 91, + 215, + 102 + ], + "type": "text", + "content": "L. Qu et al." + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 7 + }, + { + "para_blocks": [ + { + "bbox": [ + 130, + 116, + 479, + 140 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 116, + 479, + 140 + ], + "spans": [ + { + "bbox": [ + 130, + 116, + 479, + 140 + ], + "type": "text", + "content": "samples was guided by clinical experts, with each risk factor represented by one slide of the patch and its local patch." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 130, + 140, + 482, + 289 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 140, + 482, + 289 + ], + "spans": [ + { + "bbox": [ + 130, + 140, + 482, + 289 + ], + "type": "text", + "content": "Learning pipeline. As illustrated in Fig. 2, we assume " + }, + { + "bbox": [ + 130, + 140, + 482, + 289 + ], + "type": "inline_equation", + "content": "W_{i}" + }, + { + "bbox": [ + 130, + 140, + 482, + 289 + ], + "type": "text", + "content": " represents the training slide containing " + }, + { + "bbox": [ + 130, + 140, + 482, + 289 + ], + "type": "inline_equation", + "content": "n_i" + }, + { + "bbox": [ + 130, + 140, + 482, + 289 + ], + "type": "text", + "content": " patches named " + }, + { + "bbox": [ + 130, + 140, + 482, + 289 + ], + "type": "inline_equation", + "content": "p_{i,j}" + }, + { + "bbox": [ + 130, + 140, + 482, + 289 + ], + "type": "text", + "content": " and utilize the fixed CLIP image encoder " + }, + { + "bbox": [ + 130, + 140, + 482, + 289 + ], + "type": "inline_equation", + "content": "\\phi_{img}" + }, + { + "bbox": [ + 130, + 140, + 482, + 289 + ], + "type": "text", + "content": " to extract patch features " + }, + { + "bbox": [ + 130, + 140, + 482, + 289 + ], + "type": "inline_equation", + "content": "f_{i,j} = \\phi_{img}(p_{i,j}) \\in \\mathbb{R}^{1 \\times v}" + }, + { + "bbox": [ + 130, + 140, + 482, + 289 + ], + "type": "text", + "content": ", where " + }, + { + "bbox": [ + 130, + 140, + 482, + 289 + ], + "type": "inline_equation", + "content": "v" + }, + { + "bbox": [ + 130, + 140, + 482, + 289 + ], + "type": "text", + "content": " represents the dimension of the features. Subsequently, the constructed visual patch examples " + }, + { + "bbox": [ + 130, + 140, + 482, + 289 + ], + "type": "inline_equation", + "content": "e_l" + }, + { + "bbox": [ + 130, + 140, + 482, + 289 + ], + "type": "text", + "content": " are also transformed into patch features " + }, + { + "bbox": [ + 130, + 140, + 482, + 289 + ], + "type": "inline_equation", + "content": "z_l = \\phi_{img}(e_l)" + }, + { + "bbox": [ + 130, + 140, + 482, + 289 + ], + "type": "text", + "content": ". " + }, + { + "bbox": [ + 130, + 140, + 482, + 289 + ], + "type": "inline_equation", + "content": "z_l" + }, + { + "bbox": [ + 130, + 140, + 482, + 289 + ], + "type": "text", + "content": " remains unchanged throughout the training process and is combined with the current slide's patch features as a fixed visual prompt. Specifically, we match the cosine similarity between " + }, + { + "bbox": [ + 130, + 140, + 482, + 289 + ], + "type": "inline_equation", + "content": "f_{i,j}" + }, + { + "bbox": [ + 130, + 140, + 482, + 289 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 130, + 140, + 482, + 289 + ], + "type": "inline_equation", + "content": "z_l" + }, + { + "bbox": [ + 130, + 140, + 482, + 289 + ], + "type": "text", + "content": ", concatenate the most matched example patch feature with the patch feature as " + }, + { + "bbox": [ + 130, + 140, + 482, + 289 + ], + "type": "inline_equation", + "content": "f_{i,j}^e" + }, + { + "bbox": [ + 130, + 140, + 482, + 289 + ], + "type": "text", + "content": ", and ultimately concatenate all new patch features for the current slide " + }, + { + "bbox": [ + 130, + 140, + 482, + 289 + ], + "type": "inline_equation", + "content": "F_i^e = \\{f_{i,j}^e\\}_{j=1}^{n_i} \\in \\mathbb{R}^{n_i \\times v \\times 2}" + }, + { + "bbox": [ + 130, + 140, + 482, + 289 + ], + "type": "text", + "content": ". To model all patch features within a slide effectively, we introduce a lightweight self-attention layer named Messenger Layer (ML). Taking " + }, + { + "bbox": [ + 130, + 140, + 482, + 289 + ], + "type": "inline_equation", + "content": "F_i^e" + }, + { + "bbox": [ + 130, + 140, + 482, + 289 + ], + "type": "text", + "content": " as input, it outputs " + }, + { + "bbox": [ + 130, + 140, + 482, + 289 + ], + "type": "inline_equation", + "content": "F_i^{ML}" + }, + { + "bbox": [ + 130, + 140, + 482, + 289 + ], + "type": "text", + "content": ":" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 246, + 297, + 481, + 323 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 246, + 297, + 481, + 323 + ], + "spans": [ + { + "bbox": [ + 246, + 297, + 481, + 323 + ], + "type": "interline_equation", + "content": "F _ {i} ^ {M L} = \\operatorname {s o f t m a x} \\left(\\frac {Q K ^ {\\top}}{\\sqrt {d _ {w}}}\\right) V, \\tag {4}", + "image_path": "d2aeab05c82294178863a3969b0565411734dda14623a3f3d00b4117405e8c0b.jpg" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 130, + 330, + 482, + 498 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 330, + 482, + 498 + ], + "spans": [ + { + "bbox": [ + 130, + 330, + 482, + 498 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 130, + 330, + 482, + 498 + ], + "type": "inline_equation", + "content": "Q" + }, + { + "bbox": [ + 130, + 330, + 482, + 498 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 130, + 330, + 482, + 498 + ], + "type": "inline_equation", + "content": "K" + }, + { + "bbox": [ + 130, + 330, + 482, + 498 + ], + "type": "text", + "content": ", and " + }, + { + "bbox": [ + 130, + 330, + 482, + 498 + ], + "type": "inline_equation", + "content": "V" + }, + { + "bbox": [ + 130, + 330, + 482, + 498 + ], + "type": "text", + "content": " are Query, Key, and Value vectors obtained through linear mapping of " + }, + { + "bbox": [ + 130, + 330, + 482, + 498 + ], + "type": "inline_equation", + "content": "F_{i}^{e}" + }, + { + "bbox": [ + 130, + 330, + 482, + 498 + ], + "type": "text", + "content": ", and " + }, + { + "bbox": [ + 130, + 330, + 482, + 498 + ], + "type": "inline_equation", + "content": "d_{w}" + }, + { + "bbox": [ + 130, + 330, + 482, + 498 + ], + "type": "text", + "content": " is the feature dimension. Next, we construct a lightweight Attention Aggregation Layer to aggregate all patch features " + }, + { + "bbox": [ + 130, + 330, + 482, + 498 + ], + "type": "inline_equation", + "content": "F_{i}^{ML}" + }, + { + "bbox": [ + 130, + 330, + 482, + 498 + ], + "type": "text", + "content": " within a slide into a slide feature " + }, + { + "bbox": [ + 130, + 330, + 482, + 498 + ], + "type": "inline_equation", + "content": "F_{i}^{S}" + }, + { + "bbox": [ + 130, + 330, + 482, + 498 + ], + "type": "text", + "content": ". The attention weights are learnable, and the detailed structure is shown in Section 3.2. Following this, the constructed visual slide examples " + }, + { + "bbox": [ + 130, + 330, + 482, + 498 + ], + "type": "inline_equation", + "content": "E_{m}" + }, + { + "bbox": [ + 130, + 330, + 482, + 498 + ], + "type": "text", + "content": " undergo the aforementioned process like " + }, + { + "bbox": [ + 130, + 330, + 482, + 498 + ], + "type": "inline_equation", + "content": "W_{i}" + }, + { + "bbox": [ + 130, + 330, + 482, + 498 + ], + "type": "text", + "content": " (i.e., patch extraction, patch example matching, Messenger Layer, and Summary Layer) to obtain their slide features " + }, + { + "bbox": [ + 130, + 330, + 482, + 498 + ], + "type": "inline_equation", + "content": "Z_{m}" + }, + { + "bbox": [ + 130, + 330, + 482, + 498 + ], + "type": "text", + "content": ". These slide example features are also integrated as prompts into slide-level training. Similar to the patch-level matching, we match the cosine similarity between " + }, + { + "bbox": [ + 130, + 330, + 482, + 498 + ], + "type": "inline_equation", + "content": "F_{i}^{S}" + }, + { + "bbox": [ + 130, + 330, + 482, + 498 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 130, + 330, + 482, + 498 + ], + "type": "inline_equation", + "content": "Z_{m}" + }, + { + "bbox": [ + 130, + 330, + 482, + 498 + ], + "type": "text", + "content": " and concatenate the most matched slide example feature with the slide feature. Additionally, we introduce learnable vector " + }, + { + "bbox": [ + 130, + 330, + 482, + 498 + ], + "type": "inline_equation", + "content": "F_{i}^{P}" + }, + { + "bbox": [ + 130, + 330, + 482, + 498 + ], + "type": "text", + "content": " as a slide-level learnable prompt, concatenate it together, and input them into a one-layer fully-connected layer named \"projector\" to adjust the feature dimension, yielding the final slide visual features " + }, + { + "bbox": [ + 130, + 330, + 482, + 498 + ], + "type": "inline_equation", + "content": "F_{i} \\in \\mathbb{R}^{1 \\times v}" + }, + { + "bbox": [ + 130, + 330, + 482, + 498 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 132, + 514, + 289, + 528 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 514, + 289, + 528 + ], + "spans": [ + { + "bbox": [ + 132, + 514, + 289, + 528 + ], + "type": "text", + "content": "4.3 Textual Prompt Learning" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 130, + 533, + 482, + 629 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 533, + 482, + 629 + ], + "spans": [ + { + "bbox": [ + 130, + 533, + 482, + 629 + ], + "type": "text", + "content": "Construction of textual prompts. Textual prompt learning primarily focuses on modeling language descriptions related to pathology tasks, patch-level, and slide-level pathology priors on the text side. This knowledge is integrated into the training process through fixed descriptive prompts and learnable prompts (shown in the middle of Fig. 2). The construction of textual prompts could be divided into mainly three parts. The following detailed explanation uses the task \"predicting the prognosis of early cervical cancer patients from H&E pathology slides of the primary lesion\" as an example, also illustrated in Fig. 3b." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 130, + 630, + 481, + 666 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 630, + 481, + 666 + ], + "spans": [ + { + "bbox": [ + 130, + 630, + 481, + 666 + ], + "type": "text", + "content": "First, the \"Slide task [Token]\" primarily serves to prompt the main category of the current task, such as \"A Whole Slide Image of cervical cancer from primary tumors with a [poor/good] prognosis.\" Furthermore, we concatenate" + } + ] + } + ], + "index": 8 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 184, + 91, + 447, + 103 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 184, + 91, + 447, + 103 + ], + "spans": [ + { + "bbox": [ + 184, + 91, + 447, + 103 + ], + "type": "text", + "content": "Pathology-knowledge Enhanced Multi-instance Prompt Learning" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 474, + 92, + 480, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 474, + 92, + 480, + 100 + ], + "spans": [ + { + "bbox": [ + 474, + 92, + 480, + 100 + ], + "type": "text", + "content": "9" + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 8 + }, + { + "para_blocks": [ + { + "bbox": [ + 130, + 115, + 479, + 140 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 115, + 479, + 140 + ], + "spans": [ + { + "bbox": [ + 130, + 115, + 479, + 140 + ], + "type": "text", + "content": "learnable prompts " + }, + { + "bbox": [ + 130, + 115, + 479, + 140 + ], + "type": "inline_equation", + "content": "(([\\alpha]_1[\\alpha]_2\\ldots [\\alpha]_M))" + }, + { + "bbox": [ + 130, + 115, + 479, + 140 + ], + "type": "text", + "content": " with the \"Slide task [Token]\" to adaptively adjust and complement during few-shot training." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 130, + 140, + 481, + 272 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 140, + 481, + 272 + ], + "spans": [ + { + "bbox": [ + 130, + 140, + 481, + 272 + ], + "type": "text", + "content": "Then, the \"Slide-level descriptive [Token]\" is a straightforward language description corresponding to the aforementioned visual slide examples. These descriptions are provided by experienced pathology experts, avoiding obscure professional terminology. For instance, the description corresponding to the \"Poor prognosis\" category is: \"Fuzzy Tumor Boundaries. The edges or borders of the tumor are fuzzy or hard to distinguish on the slide level. Low tumor stroma ratio. Few stroma tissues surrounding the tumors.\" In contrast, the description corresponding to the \"Good prognosis\" category is: \"Clear Tumor Boundaries. The edges or borders of the tumor are easy to see and well-defined on the slide level. High tumor stroma ratio. Lots of stroma tissue surrounding the tumors.\" Similarly, we also append learnable prompts " + }, + { + "bbox": [ + 130, + 140, + 481, + 272 + ], + "type": "inline_equation", + "content": "(([\\beta]_{1}[\\beta]_{2}\\ldots [\\beta]_{M}))" + }, + { + "bbox": [ + 130, + 140, + 481, + 272 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 130, + 274, + 482, + 369 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 274, + 482, + 369 + ], + "spans": [ + { + "bbox": [ + 130, + 274, + 482, + 369 + ], + "type": "text", + "content": "Finally, the \"Patch-level descriptive [Token]\" provides the language description corresponding to the aforementioned visual patch examples. For example, the description corresponding to the \"Poor prognosis\" category is: \"Vascular and lymphatic invasion: The presence of cancer cells within the blood and lymphatic vessels. High-level atypia: Pronounced and abnormal variations in the size, shape, and staining pattern of the cells. Necrosis: Empty spaces or regions with a lack of normal cell structure.\" Similarly, immediately following this, we also include learnable prompts " + }, + { + "bbox": [ + 130, + 274, + 482, + 369 + ], + "type": "inline_equation", + "content": "(([\\gamma]_{1}[\\gamma]_{2}\\ldots [\\gamma]_{M}))" + }, + { + "bbox": [ + 130, + 274, + 482, + 369 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 130, + 369, + 482, + 417 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 369, + 482, + 417 + ], + "spans": [ + { + "bbox": [ + 130, + 369, + 482, + 417 + ], + "type": "text", + "content": "It is crucial to note that the second and third parts correspond closely to the same number of visual slide and patch examples in visual prompt learning. We input each part's visual descriptions and learnable vectors together into the CLIP Text Encoder, obtaining the final slide textual features " + }, + { + "bbox": [ + 130, + 369, + 482, + 417 + ], + "type": "inline_equation", + "content": "T_{i} \\in \\mathbb{R}^{1 \\times v}" + }, + { + "bbox": [ + 130, + 369, + 482, + 417 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 132, + 437, + 373, + 449 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 437, + 373, + 449 + ], + "spans": [ + { + "bbox": [ + 132, + 437, + 373, + 449 + ], + "type": "text", + "content": "4.4 Alignment of Visual and Textual Prompts" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 130, + 459, + 482, + 579 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 459, + 482, + 579 + ], + "spans": [ + { + "bbox": [ + 130, + 459, + 482, + 579 + ], + "type": "text", + "content": "Alignment contrastive learning (AC-Loss) is originally used to align corresponding visual representation and textual representation. Nonetheless, the overall loss function in this paper consists of three parts of AC-Loss, expressed as " + }, + { + "bbox": [ + 130, + 459, + 482, + 579 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_{\\mathrm{total}} = \\mathcal{L}_t + \\lambda_1\\mathcal{L}_s + \\lambda_2\\mathcal{L}_p" + }, + { + "bbox": [ + 130, + 459, + 482, + 579 + ], + "type": "text", + "content": ". Conceptually, " + }, + { + "bbox": [ + 130, + 459, + 482, + 579 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_t" + }, + { + "bbox": [ + 130, + 459, + 482, + 579 + ], + "type": "text", + "content": " aligns the overall slide visual features " + }, + { + "bbox": [ + 130, + 459, + 482, + 579 + ], + "type": "inline_equation", + "content": "F_{i}" + }, + { + "bbox": [ + 130, + 459, + 482, + 579 + ], + "type": "text", + "content": " with the slide textual features " + }, + { + "bbox": [ + 130, + 459, + 482, + 579 + ], + "type": "inline_equation", + "content": "T_{i}" + }, + { + "bbox": [ + 130, + 459, + 482, + 579 + ], + "type": "text", + "content": " corresponding to the task category, thereby completing slide-level classification. " + }, + { + "bbox": [ + 130, + 459, + 482, + 579 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_s" + }, + { + "bbox": [ + 130, + 459, + 482, + 579 + ], + "type": "text", + "content": " aligns the features of slide-level visual examples in visual prompt learning with the slide-level pathological language descriptions in textual prompt learning. " + }, + { + "bbox": [ + 130, + 459, + 482, + 579 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_p" + }, + { + "bbox": [ + 130, + 459, + 482, + 579 + ], + "type": "text", + "content": " aligns the features of patch-level visual examples with patch-level pathology language descriptions (shown in the bottom of Fig. 2)." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 130, + 579, + 482, + 628 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 579, + 482, + 628 + ], + "spans": [ + { + "bbox": [ + 130, + 579, + 482, + 628 + ], + "type": "text", + "content": "Mathematically, the forms of these three parts of AC-Loss are identical. Here, we take " + }, + { + "bbox": [ + 130, + 579, + 482, + 628 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_t" + }, + { + "bbox": [ + 130, + 579, + 482, + 628 + ], + "type": "text", + "content": " as an example. AC-Loss optimizes the learnable parameters by minimizing the negative log-likelihood between the image features " + }, + { + "bbox": [ + 130, + 579, + 482, + 628 + ], + "type": "inline_equation", + "content": "F_{i}" + }, + { + "bbox": [ + 130, + 579, + 482, + 628 + ], + "type": "text", + "content": " and the ground truth class text features " + }, + { + "bbox": [ + 130, + 579, + 482, + 628 + ], + "type": "inline_equation", + "content": "T_{y}" + }, + { + "bbox": [ + 130, + 579, + 482, + 628 + ], + "type": "text", + "content": ". Its basic form is as follows:" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 219, + 638, + 481, + 669 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 219, + 638, + 481, + 669 + ], + "spans": [ + { + "bbox": [ + 219, + 638, + 481, + 669 + ], + "type": "interline_equation", + "content": "\\mathcal {L} = - \\sum_ {F _ {i}} \\log \\frac {\\exp \\left(\\operatorname {s i m} \\left(F _ {i} , T _ {y}\\right) / \\tau\\right)}{\\sum_ {i = 1} ^ {U} \\exp \\left(\\operatorname {s i m} \\left(F _ {i} , T _ {i}\\right) / \\tau\\right)}, \\tag {5}", + "image_path": "d8be104b6f9020b7b2a19a42fcc5fc667e4fcc360ff06e192e6e974108fc7a64.jpg" + } + ] + } + ], + "index": 9 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "spans": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "type": "text", + "content": "10" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 166, + 91, + 215, + 102 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 166, + 91, + 215, + 102 + ], + "spans": [ + { + "bbox": [ + 166, + 91, + 215, + 102 + ], + "type": "text", + "content": "L. Qu et al." + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 9 + }, + { + "para_blocks": [ + { + "bbox": [ + 130, + 116, + 479, + 140 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 116, + 479, + 140 + ], + "spans": [ + { + "bbox": [ + 130, + 116, + 479, + 140 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 130, + 116, + 479, + 140 + ], + "type": "inline_equation", + "content": "y" + }, + { + "bbox": [ + 130, + 116, + 479, + 140 + ], + "type": "text", + "content": " is the ground truth category, " + }, + { + "bbox": [ + 130, + 116, + 479, + 140 + ], + "type": "inline_equation", + "content": "U" + }, + { + "bbox": [ + 130, + 116, + 479, + 140 + ], + "type": "text", + "content": " is the total number of categories, " + }, + { + "bbox": [ + 130, + 116, + 479, + 140 + ], + "type": "inline_equation", + "content": "\\mathrm{sim}(\\cdot)" + }, + { + "bbox": [ + 130, + 116, + 479, + 140 + ], + "type": "text", + "content": " denotes the cosine similarity, and " + }, + { + "bbox": [ + 130, + 116, + 479, + 140 + ], + "type": "inline_equation", + "content": "\\tau" + }, + { + "bbox": [ + 130, + 116, + 479, + 140 + ], + "type": "text", + "content": " is a learnable temperature parameter." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 130, + 140, + 482, + 189 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 140, + 482, + 189 + ], + "spans": [ + { + "bbox": [ + 130, + 140, + 482, + 189 + ], + "type": "text", + "content": "During testing, we input the unseen WSI and obtain its slide visual feature in the same fashion as training. Subsequently, we match it separately with the textual features for \"poor prognosis\" and \"good prognosis\" and use softmax to obtain the predicted risks for the two categories." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 131, + 208, + 293, + 221 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 131, + 208, + 293, + 221 + ], + "spans": [ + { + "bbox": [ + 131, + 208, + 293, + 221 + ], + "type": "text", + "content": "5 Experiment and Result" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 130, + 234, + 482, + 498 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 234, + 482, + 498 + ], + "spans": [ + { + "bbox": [ + 130, + 234, + 482, + 498 + ], + "type": "text", + "content": "Datasets and Evaluation Metrics. This paper addresses clinical tasks in H&E pathology diagnosis, specifically focusing on data label scarcity, the complexity of direct diagnosis by pathologists, and limited sample sizes for rare diseases. This study encompasses three clinical tasks for five tumor types across five datasets: (1) Predicting patient survival prognosis and lymph node metastasis from H&E slides of early-stage cervical cancer primary lesions using an in-house dataset (For testing, 80 cases and 74 cases, individually) and a public TCGA-CESC dataset (For testing, 80 cases and 70 cases, individually). Note that the prompts utilized in the public TCGA datasets are the same as those used in the corresponding in-house datasets for each task. We ensure no overlap between our training data and visual prompts (obtained from external authoritative sources or pathology experts) and testing data, across both in-house and TCGA datasets. Furthermore, the lymph node metastasis detection tasks from primary lesions assessed in this study present a significantly greater challenge compared to those from sentinel lymph nodes, as utilized in prior research, with samples and labels being notably rarer. (2) Classifying round cell tumors, including neuroendocrine tumors, malignant melanoma, lymphohematopoietic tumors, and soft tissue tumors, into subtypes using an in-house dataset (For testing, 112 cases). More details are in the Supplementary Materials. For the prognosis prediction task in cervical cancer, the concordance index (C-index) is used, while for other tasks, the Area Under the Curve (AUC) is employed. Higher values for both metrics indicate improved performance." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 130, + 498, + 482, + 666 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 498, + 482, + 666 + ], + "spans": [ + { + "bbox": [ + 130, + 498, + 482, + 666 + ], + "type": "text", + "content": "Compared Methods. Currently, few methods directly address FSWC tasks through prompt learning, with the SOTA being represented by the recently proposed TOP method [27] in the pathology domain. In the natural domain, prompt learning work based on VLMs is limited to single image patches and does not readily apply to slide-level classification. Following the TOP approach, we adopt advanced prompt learning methods from the natural domain for the MIL problem. These methods include: (1) Linear-Probe [29], (2) VPT [17], (3) CoOp [42], and (4) KgCoOp [38]. The Messenger Layer and Summary Layer, as described in this paper, aggregate patch features into a slide feature. All compared methods use the same Messenger & Summary layers for fair comparisons. For Linear-Probe, a linear layer performs slide-level classification on the aggregated feature. VPT incorporates a learnable prompt on the vision side for slide-level classification. CoOp utilizes CLIP's prompt learning by adding the Slide Task [Token] and learnable textual prompts trained through AC-Loss. Kg-" + } + ] + } + ], + "index": 6 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 184, + 91, + 447, + 103 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 184, + 91, + 447, + 103 + ], + "spans": [ + { + "bbox": [ + 184, + 91, + 447, + 103 + ], + "type": "text", + "content": "Pathology-knowledge Enhanced Multi-instance Prompt Learning" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 470, + 91, + 479, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 470, + 91, + 479, + 100 + ], + "spans": [ + { + "bbox": [ + 470, + 91, + 479, + 100 + ], + "type": "text", + "content": "11" + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 10 + }, + { + "para_blocks": [ + { + "type": "table", + "bbox": [ + 167, + 136, + 445, + 190 + ], + "blocks": [ + { + "bbox": [ + 180, + 114, + 432, + 126 + ], + "lines": [ + { + "bbox": [ + 180, + 114, + 432, + 126 + ], + "spans": [ + { + "bbox": [ + 180, + 114, + 432, + 126 + ], + "type": "text", + "content": "Table 1: Comparison for Survival Prognosis Prediction Task." + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 167, + 136, + 445, + 190 + ], + "lines": [ + { + "bbox": [ + 167, + 136, + 445, + 190 + ], + "spans": [ + { + "bbox": [ + 167, + 136, + 445, + 190 + ], + "type": "table", + "html": "
DatasetIn-house DatasetPublic TCGA-CESC Dataset
Method32-shot16-shot8-shot4-shot2-shot32-shot16-shot8-shot4-shot2-shot
LinearProbe0.6200.5620.5430.5010.4580.5840.5770.5430.5110.485
VPT(ECCV'22)0.6260.5690.5450.5010.4640.5890.5730.5500.5180.482
CoOp(IJCV'22)0.6410.5940.5610.5170.4900.6040.5850.5580.5230.487
KgCoOp(CVPR'23)0.6350.5870.5520.5130.4870.5970.5840.5500.5190.486
TOP(NeurIPS'23)0.6520.6080.5740.5390.5080.6110.5970.5660.5360.518
PEMP (ours)0.6670.6370.6140.5870.5620.6370.6240.6020.5770.551
", + "image_path": "bb1c12fcd4f83d8de30a3854a5d076be72a8eff59c35f726d16bfc153f19c7db.jpg" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "table_body" + } + ], + "index": 3 + }, + { + "type": "table", + "bbox": [ + 167, + 221, + 445, + 276 + ], + "blocks": [ + { + "bbox": [ + 168, + 201, + 443, + 212 + ], + "lines": [ + { + "bbox": [ + 168, + 201, + 443, + 212 + ], + "spans": [ + { + "bbox": [ + 168, + 201, + 443, + 212 + ], + "type": "text", + "content": "Table 2: Comparison for Lymph Node Metastasis Prediction Task." + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 167, + 221, + 445, + 276 + ], + "lines": [ + { + "bbox": [ + 167, + 221, + 445, + 276 + ], + "spans": [ + { + "bbox": [ + 167, + 221, + 445, + 276 + ], + "type": "table", + "html": "
DatasetIn-house DatasetPublic TCGA-CESC Dataset
Method32-shot16-shot8-shot4-shot2-shot32-shot16-shot8-shot4-shot2-shot
LinearProbe0.7810.7350.7280.7160.6870.7630.7190.6960.6710.655
VPT(ECCV'22)0.7850.7420.7280.7190.6890.7690.7240.7120.6780.655
CoOp(IJCV'22)0.7920.7570.7350.7270.6930.7860.7410.7230.6910.659
KgCoOp(CVPR'23)0.7890.7530.7310.7240.6930.7750.7330.7100.6870.652
TOP(NeurIPS'23)0.8250.8190.8010.7870.7620.7990.7610.7440.7080.679
PEMP(ours)0.8490.8380.8240.8010.7830.8180.7950.7600.7260.704
", + "image_path": "1f2dd4f90b6e8026354a48d95b0282ec9b7c5a81fa7ea8260f55b232885ea20c.jpg" + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "table_body" + } + ], + "index": 5 + }, + { + "type": "table", + "bbox": [ + 202, + 308, + 410, + 367 + ], + "blocks": [ + { + "bbox": [ + 178, + 286, + 434, + 298 + ], + "lines": [ + { + "bbox": [ + 178, + 286, + 434, + 298 + ], + "spans": [ + { + "bbox": [ + 178, + 286, + 434, + 298 + ], + "type": "text", + "content": "Table 3: Comparison for Round Cell Subtype Diagnosis Task." + } + ] + } + ], + "index": 6, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 202, + 308, + 410, + 367 + ], + "lines": [ + { + "bbox": [ + 202, + 308, + 410, + 367 + ], + "spans": [ + { + "bbox": [ + 202, + 308, + 410, + 367 + ], + "type": "table", + "html": "
Method32-shot16-shot8-shot4-shot2-shot
LinearProbe0.6540.6060.5490.5170.483
VPT(ECCV'22)0.6590.6130.5560.5150.511
CoOp(IJCV'22)0.6670.6270.5680.5530.549
KgCoOp(CVPR'23)0.6610.6190.5470.5350.534
TOP(NeurIPS'23)0.6820.6520.6330.5840.560
PEMP (ours)0.7510.7180.6850.6430.625
", + "image_path": "5bc3176b43b6e30e152e1cdd4c6b0aebf71af604ecc3417f31e9296a458c09f7.jpg" + } + ] + } + ], + "index": 7, + "angle": 0, + "type": "table_body" + } + ], + "index": 7 + }, + { + "bbox": [ + 130, + 388, + 482, + 437 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 388, + 482, + 437 + ], + "spans": [ + { + "bbox": [ + 130, + 388, + 482, + 437 + ], + "type": "text", + "content": "CoOp imposes constraints on learnable textual prompts. Few-shot experiments are conducted for each category with 2, 4, 8, 16, and 32 slides (with labels). Results are reported as an average over five experimental runs. Standard deviation of our main results are provided in the supplementary materials." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 132, + 454, + 198, + 465 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 454, + 198, + 465 + ], + "spans": [ + { + "bbox": [ + 132, + 454, + 198, + 465 + ], + "type": "text", + "content": "5.1 Results" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 130, + 473, + 482, + 666 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 473, + 482, + 666 + ], + "spans": [ + { + "bbox": [ + 130, + 473, + 482, + 666 + ], + "type": "text", + "content": "Survival prognosis prediction. Predicting survival prognosis, especially directly from H&E slides with a limited number of samples, poses a significant challenge. Table 1 showcases the performance of all methods in the in-house dataset and the public TCGA-CESC dataset, with PEMP consistently demonstrating superior performance across all settings with few samples, outperforming all comparative methods significantly. In scenarios with limited samples, LinearProbe and VPT, relying solely on features extracted by the CLIP Image Encoder on the visual side without incorporating language-based prior knowledge, exhibit suboptimal performance. CoOp and KgCoOp, which combine visual and language aspects, struggle to comprehend complex medical terms related to prognosis effectively, limiting their performance. TOP enhances language descriptions at both slide and patch levels but provides explanations solely on the language side, making it challenging to establish a direct relationship between crucial visual features and language descriptions. In contrast, PEMP introduces typical patch and slide examples related to prognosis on the visual side, along with straightforward language descriptions on the language side. Through AC-" + } + ] + } + ], + "index": 10 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "spans": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "type": "text", + "content": "12" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 166, + 91, + 215, + 102 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 166, + 91, + 215, + 102 + ], + "spans": [ + { + "bbox": [ + 166, + 91, + 215, + 102 + ], + "type": "text", + "content": "L. Qu et al." + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 11 + }, + { + "para_blocks": [ + { + "type": "table", + "bbox": [ + 202, + 135, + 411, + 208 + ], + "blocks": [ + { + "bbox": [ + 219, + 114, + 394, + 125 + ], + "lines": [ + { + "bbox": [ + 219, + 114, + 394, + 125 + ], + "spans": [ + { + "bbox": [ + 219, + 114, + 394, + 125 + ], + "type": "text", + "content": "Table 4: Ablation studies of the modules." + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 202, + 135, + 411, + 208 + ], + "lines": [ + { + "bbox": [ + 202, + 135, + 411, + 208 + ], + "spans": [ + { + "bbox": [ + 202, + 135, + 411, + 208 + ], + "type": "table", + "html": "
Method32-shot16-shot8-shot4-shot2-shot
PEMP w/o v&t em.0.6410.5940.5610.5170.490
PEMP vision only0.6490.6070.5680.5430.527
PEMP w/o vision em.0.6550.6130.5770.5400.511
PEMP w/o text em.0.6580.6190.5750.5570.533
PEMP w/o Summary Layer0.6320.5870.5580.5110.487
PEMP w/o Messenger Layer0.6640.6290.5940.5750.554
PEMP w/o Slide-level Prompts0.6560.6200.5810.5520.525
PEMP w/o AC-Loss0.6600.6260.5890.5680.549
PEMP (ours)0.6670.6370.6140.5870.562
", + "image_path": "1d89706e47051ede058dc8bb2c110bc49acce3e95b6a9a56585bfde6f1fe70e0.jpg" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "table_body" + } + ], + "index": 3 + }, + { + "bbox": [ + 130, + 231, + 479, + 280 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 231, + 479, + 280 + ], + "spans": [ + { + "bbox": [ + 130, + 231, + 479, + 280 + ], + "type": "text", + "content": "Loss, the model efficiently completes the task under the guidance of pathology prior knowledge. Additionally, PEMP excels when the sample size is very small (e.g., 2-shot, 4-shot), showcasing its effective utilization of visual and language prior knowledge related to prognosis in H&E pathology slides for diagnosis." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 130, + 281, + 481, + 376 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 281, + 481, + 376 + ], + "spans": [ + { + "bbox": [ + 130, + 281, + 481, + 376 + ], + "type": "text", + "content": "Lymph node metastasis prediction. Table 2 presents the performance of all methods in the in-house and TCGA-CESC datasets. Across all settings with limited samples, PEMP consistently achieves the best performance and significantly outperforms all comparative methods. This indicates that PEMP is also capable of efficiently understanding and leveraging pathological prior knowledge related to metastasis in primary lesion H&E pathology slides for diagnosis. The results of both tasks highlight the immense potential of PEMP in clinical tasks characterized by challenges in label acquisition and diagnostic complexity." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 130, + 377, + 482, + 472 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 377, + 482, + 472 + ], + "spans": [ + { + "bbox": [ + 130, + 377, + 482, + 472 + ], + "type": "text", + "content": "Round cell subtype diagnosis. The classification diagnosis for rare diseases in the round cell subtype task presents an increasingly challenging yet vital clinical issue. Table 3 illustrates the performance of all methods. Notably, with a limited number of shots, most methods face challenges in achieving satisfactory diagnostic results. However, leveraging visual and pathological expertise from pathology experts, PEMP achieves significantly better results for all the settings (6.2% higher AUC on average). This underscores the substantial potential of PEMP in the diagnosis of rare diseases." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 130, + 473, + 482, + 640 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 473, + 482, + 640 + ], + "spans": [ + { + "bbox": [ + 130, + 473, + 482, + 640 + ], + "type": "text", + "content": "Ablation Study. As shown in Table 4, we designed eight variants and evaluated the roles of each component in PEMP in the prognosis task. (1) PEMP w/o v&t em. represents the removal of all vision and text examples, reducing PEMP to CoOp; (2) PEMP vision only denotes the exclusion of the text end, training solely with vision examples as guidance for prior knowledge; (3) PEMP w/o vision em. indicates the removal of all vision examples; (4) PEMP w/o text em. signifies the exclusion of all text examples; (5) PEMP w/o Summary Layer denotes the replacement from Summary Layer to simple average pooling. Please note that the Summary Layer is essentially an Attention Pooling layer, recognized as a crucial component in MIL problems; (6) PEMP w/o Messenger Layer denotes the removal of the Messenger Layer; (7) PEMP w/o Slide-level Prompt denotes the removal of all slide-level prompts; (8) PEMP w/o AC-Loss denotes the removal of AC-Loss for vision and text examples while retaining the AC-Loss for the visual and text features of the slide task token." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 130, + 641, + 481, + 666 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 641, + 481, + 666 + ], + "spans": [ + { + "bbox": [ + 130, + 641, + 481, + 666 + ], + "type": "text", + "content": "Visualization and Interpretability. To illustrate the robust performance of PEMP, we briefly visualize typical cases where visual patch examples and" + } + ] + } + ], + "index": 8 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 184, + 91, + 447, + 103 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 184, + 91, + 447, + 103 + ], + "spans": [ + { + "bbox": [ + 184, + 91, + 447, + 103 + ], + "type": "text", + "content": "Pathology-knowledge Enhanced Multi-instance Prompt Learning" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 470, + 91, + 481, + 101 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 470, + 91, + 481, + 101 + ], + "spans": [ + { + "bbox": [ + 470, + 91, + 481, + 101 + ], + "type": "text", + "content": "13" + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 12 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 185, + 113, + 430, + 271 + ], + "blocks": [ + { + "bbox": [ + 185, + 113, + 430, + 271 + ], + "lines": [ + { + "bbox": [ + 185, + 113, + 430, + 271 + ], + "spans": [ + { + "bbox": [ + 185, + 113, + 430, + 271 + ], + "type": "image", + "image_path": "cb2d44ba90b186e9cf6faf3a1be08efa2a48624c996022d2b57abe93e26c43d5.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 132, + 279, + 480, + 301 + ], + "lines": [ + { + "bbox": [ + 132, + 279, + 480, + 301 + ], + "spans": [ + { + "bbox": [ + 132, + 279, + 480, + 301 + ], + "type": "text", + "content": "Fig. 4: Visualization of key pathology patterns indicating both good and poor prognosis retrieved by PEMP from the test set." + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_caption" + } + ], + "index": 2 + }, + { + "bbox": [ + 130, + 324, + 482, + 422 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 324, + 482, + 422 + ], + "spans": [ + { + "bbox": [ + 130, + 324, + 482, + 422 + ], + "type": "text", + "content": "slide examples exhibit the highest similarity on the prognosis task's test dataset (Fig. 4). PEMP demonstrates accurate learning of pathology prior knowledge, effectively addressing the challenge of predicting survival outcomes with limited training WSIs. More importantly, PEMP also shows high interpretability. PEMP efficiently integrates the high-risk and low-risk pathology image patterns and textual descriptions that pathologists have summarized over an extended period. PEMP achieves excellent performance under the guidance and facilitation of this prior knowledge." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 132, + 437, + 220, + 450 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 437, + 220, + 450 + ], + "spans": [ + { + "bbox": [ + 132, + 437, + 220, + 450 + ], + "type": "text", + "content": "6 Conclusion" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 130, + 462, + 482, + 668 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 462, + 482, + 668 + ], + "spans": [ + { + "bbox": [ + 130, + 462, + 482, + 668 + ], + "type": "text", + "content": "This paper presents PEMP, a pathology-knowledge enhanced multi-instance prompt learning framework, excelling in challenging tasks such as prognosis prediction with difficult label acquisition and rare tumor classification in Few-shot Weakly Supervised WSI Classification. PEMP integrates vision and text prior knowledge into prompts at both patch and slide levels, using a combination of static and learnable prompts during training to guide the effective diagnosis of key pathology patterns. Alignment-wise contrastive losses ensure feature-level alignment between knowledge-enhanced visual and textual prompts at both patch and slide levels. In specific clinical settings, collecting a large number of WSIs may be challenging due to patient privacy issues, the prevalence of rare diseases, new treatment protocols, or small patient groups. PEMP has shown substantial benefits across five datasets, including prognosis and rare diseases, demonstrating significant diagnostic benefits in fields with limited medical resources. PEMP also demonstrates high interpretability in medical diagnoses, indicating significant potential for future clinical applications. Moreover, such prompt-description-based model adaptation could be widely adopted as a general learning paradigm for specialized domain applications." + } + ] + } + ], + "index": 6 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "spans": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "type": "text", + "content": "14" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 166, + 91, + 215, + 102 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 166, + 91, + 215, + 102 + ], + "spans": [ + { + "bbox": [ + 166, + 91, + 215, + 102 + ], + "type": "text", + "content": "L. Qu et al." + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 13 + }, + { + "para_blocks": [ + { + "bbox": [ + 133, + 114, + 246, + 129 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 114, + 246, + 129 + ], + "spans": [ + { + "bbox": [ + 133, + 114, + 246, + 129 + ], + "type": "text", + "content": "Acknowledgements" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 132, + 141, + 487, + 166 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 141, + 487, + 166 + ], + "spans": [ + { + "bbox": [ + 132, + 141, + 487, + 166 + ], + "type": "text", + "content": "This work is funded by the National Key R&D Program of China (2022ZD0160700) and Shanghai AI Laboratory." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 133, + 186, + 197, + 198 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 186, + 197, + 198 + ], + "spans": [ + { + "bbox": [ + 133, + 186, + 197, + 198 + ], + "type": "text", + "content": "References" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 134, + 212, + 481, + 665 + ], + "type": "list", + "angle": 0, + "index": 16, + "blocks": [ + { + "bbox": [ + 138, + 212, + 481, + 255 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 212, + 481, + 255 + ], + "spans": [ + { + "bbox": [ + 138, + 212, + 481, + 255 + ], + "type": "text", + "content": "1. Alayrac, J.B., Donahue, J., Luc, P., Miech, A., Barr, I., Hasson, Y., Lenc, K., Mensch, A., Millican, K., Reynolds, M., et al.: Flamingo: a visual language model for few-shot learning. Advances in Neural Information Processing Systems (NeurIPS) 35, 23716-23736 (2022)" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 138, + 256, + 481, + 300 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 256, + 481, + 300 + ], + "spans": [ + { + "bbox": [ + 138, + 256, + 481, + 300 + ], + "type": "text", + "content": "2. Campanella, G., Hanna, M.G., Geneslaw, L., Miraflor, A., Werneck Krauss Silva, V., Busam, K.J., Brogi, E., Reuter, V.E., Klimstra, D.S., Fuchs, T.J.: Clinical-grade computational pathology using weakly supervised deep learning on whole slide images. Nature Medicine 25(8), 1301-1309 (2019)" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 138, + 301, + 481, + 344 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 301, + 481, + 344 + ], + "spans": [ + { + "bbox": [ + 138, + 301, + 481, + 344 + ], + "type": "text", + "content": "3. Chan, T.H., Cendra, F.J., Ma, L., Yin, G., Yu, L.: Histopathology whole slide image analysis with heterogeneous graph representation learning. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR). pp. 15661-15670 (2023)" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 138, + 345, + 481, + 388 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 345, + 481, + 388 + ], + "spans": [ + { + "bbox": [ + 138, + 345, + 481, + 388 + ], + "type": "text", + "content": "4. Chen, R.J., Chen, C., Li, Y., Chen, T.Y., Trister, A.D., Krishnan, R.G., Mahmood, F.: Scaling vision transformers to gigapixel images via hierarchical self-supervised learning. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR). pp. 16144-16155 (2022)" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 138, + 388, + 481, + 433 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 388, + 481, + 433 + ], + "spans": [ + { + "bbox": [ + 138, + 388, + 481, + 433 + ], + "type": "text", + "content": "5. Chen, R.J., Lu, M.Y., Wang, J., Williamson, D.F., Rodig, S.J., Lindeman, N.I., Mahmood, F.: Pathomic fusion: an integrated framework for fusing histopathology and genomic features for cancer diagnosis and prognosis. IEEE Transactions on Medical Imaging 41(4), 757-770 (2020)" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 138, + 434, + 481, + 477 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 434, + 481, + 477 + ], + "spans": [ + { + "bbox": [ + 138, + 434, + 481, + 477 + ], + "type": "text", + "content": "6. Chen, R.J., Lu, M.Y., Weng, W.H., Chen, T.Y., Williamson, D.F., Manz, T., Shady, M., Mahmood, F.: Multimodal co-attention transformer for survival prediction in gigapixel whole slide images. In: Proceedings of the IEEE/CVF International Conference on Computer Vision (ICCV). pp. 4015-4025 (2021)" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 138, + 478, + 481, + 510 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 478, + 481, + 510 + ], + "spans": [ + { + "bbox": [ + 138, + 478, + 481, + 510 + ], + "type": "text", + "content": "7. Chen, W., Si, C., Zhang, Z., Wang, L., Wang, Z., Tan, T.: Semantic prompt for few-shot image recognition. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR). pp. 23581-23591 (2023)" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 138, + 510, + 481, + 554 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 510, + 481, + 554 + ], + "spans": [ + { + "bbox": [ + 138, + 510, + 481, + 554 + ], + "type": "text", + "content": "8. Chen, Y.C., Lu, C.S.: Rankmix: Data augmentation for weakly supervised learning of classifying whole slide images with diverse sizes and imbalanced categories. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR). pp. 23936-23945 (2023)" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 138, + 555, + 481, + 587 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 555, + 481, + 587 + ], + "spans": [ + { + "bbox": [ + 138, + 555, + 481, + 587 + ], + "type": "text", + "content": "9. Cheplygina, V., de Bruijne, M., Pluim, J.P.: Not-so-supervised: a survey of semi-supervised, multi-instance, and transfer learning in medical image analysis. Medical Image Analysis 54, 280-296 (2019)" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 134, + 588, + 481, + 632 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 134, + 588, + 481, + 632 + ], + "spans": [ + { + "bbox": [ + 134, + 588, + 481, + 632 + ], + "type": "text", + "content": "10. Chikontwe, P., Kim, M., Nam, S.J., Go, H., Park, S.H.: Multiple instance learning with center embeddings for histopathology classification. In: Medical Image Computing and Computer Assisted Intervention (MICCAI). pp. 519-528. Springer (2020)" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 134, + 632, + 481, + 665 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 134, + 632, + 481, + 665 + ], + "spans": [ + { + "bbox": [ + 134, + 632, + 481, + 665 + ], + "type": "text", + "content": "11. Gu, J., Han, Z., Chen, S., Beirami, A., He, B., Zhang, G., Liao, R., Qin, Y., Tresp, V., Torr, P.: A systematic survey of prompt engineering on vision-language foundation models. arXiv preprint arXiv:2307.12980 (2023)" + } + ] + } + ], + "index": 15 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 184, + 91, + 447, + 102 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 184, + 91, + 447, + 102 + ], + "spans": [ + { + "bbox": [ + 184, + 91, + 447, + 102 + ], + "type": "text", + "content": "Pathology-knowledge Enhanced Multi-instance Prompt Learning" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 470, + 92, + 480, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 470, + 92, + 480, + 100 + ], + "spans": [ + { + "bbox": [ + 470, + 92, + 480, + 100 + ], + "type": "text", + "content": "15" + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 14 + }, + { + "para_blocks": [ + { + "bbox": [ + 132, + 116, + 481, + 666 + ], + "type": "list", + "angle": 0, + "index": 15, + "blocks": [ + { + "bbox": [ + 133, + 116, + 481, + 172 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 116, + 481, + 172 + ], + "spans": [ + { + "bbox": [ + 133, + 116, + 481, + 172 + ], + "type": "text", + "content": "12. Hashimoto, N., Fukushima, D., Koga, R., Takagi, Y., Ko, K., Kohno, K., Nakaguro, M., Nakamura, S., Hontani, H., Takeuchi, I.: Multi-scale domain-adversarial multiple-instance cnn for cancer subtype classification with unannotated histopathological images. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR). pp. 3852-3861 (2020)" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 133, + 172, + 481, + 215 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 172, + 481, + 215 + ], + "spans": [ + { + "bbox": [ + 133, + 172, + 481, + 215 + ], + "type": "text", + "content": "13. Huang, Y., Zhao, W., Wang, S., Fu, Y., Jiang, Y., Yu, L.: Conslide: Asynchronous hierarchical interaction transformer with breakup-reorganize rehearsal for continual whole slide image analysis. In: Proceedings of the IEEE/CVF International Conference on Computer Vision (ICCV). pp. 21349-21360 (2023)" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 133, + 216, + 481, + 248 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 216, + 481, + 248 + ], + "spans": [ + { + "bbox": [ + 133, + 216, + 481, + 248 + ], + "type": "text", + "content": "14. Huang, Z., Bianchi, F., Yuksekgonul, M., Montine, T.J., Zou, J.: A visual-language foundation model for pathology image analysis using medical twitter. Nature Medicine pp. 1-10 (2023)" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 133, + 249, + 481, + 292 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 249, + 481, + 292 + ], + "spans": [ + { + "bbox": [ + 133, + 249, + 481, + 292 + ], + "type": "text", + "content": "15. Ikezogwo, W., Seyfioglu, S., Ghezloo, F., Geva, D., Sheikh Mohammed, F., Anand, P.K., Krishna, R., Shapiro, L.: Quilt-1m: One million image-text pairs for histopathology. Advances in Neural Information Processing Systems (NeurIPS) 36 (2024)" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 132, + 293, + 481, + 324 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 293, + 481, + 324 + ], + "spans": [ + { + "bbox": [ + 132, + 293, + 481, + 324 + ], + "type": "text", + "content": "16. Ilse, M., Tomczak, J., Welling, M.: Attention-based deep multiple instance learning. In: International Conference on Machine Learning (ICML). pp. 2127-2136. PMLR (2018)" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 132, + 325, + 481, + 358 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 325, + 481, + 358 + ], + "spans": [ + { + "bbox": [ + 132, + 325, + 481, + 358 + ], + "type": "text", + "content": "17. Jia, M., Tang, L., Chen, B.C., Cardie, C., Belongie, S., Hariharan, B., Lim, S.N.: Visual prompt tuning. In: European Conference on Computer Vision (ECCV). pp. 709-727. Springer (2022)" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 132, + 358, + 481, + 402 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 358, + 481, + 402 + ], + "spans": [ + { + "bbox": [ + 132, + 358, + 481, + 402 + ], + "type": "text", + "content": "18. Li, B., Li, Y., Eliceiri, K.W.: Dual-stream multiple instance learning network for whole slide image classification with self-supervised contrastive learning. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR). pp. 14318-14328 (2021)" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 132, + 402, + 481, + 446 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 402, + 481, + 446 + ], + "spans": [ + { + "bbox": [ + 132, + 402, + 481, + 446 + ], + "type": "text", + "content": "19. Li, H., Yang, F., Zhao, Y., Xing, X., Zhang, J., Gao, M., Huang, J., Wang, L., Yao, J.: Dt-mil: deformable transformer for multi-instance learning on histopathological image. In: Medical Image Computing and Computer Assisted Intervention (MICCAI). pp. 206-216. Springer (2021)" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 132, + 447, + 481, + 500 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 447, + 481, + 500 + ], + "spans": [ + { + "bbox": [ + 132, + 447, + 481, + 500 + ], + "type": "text", + "content": "20. Li, H., Zhu, C., Zhang, Y., Sun, Y., Shui, Z., Kuang, W., Zheng, S., Yang, L.: Task-specific fine-tuning via variational information bottleneck for weakly-supervised pathology whole slide image classification. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR). pp. 7454-7463 (2023)" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 132, + 501, + 481, + 533 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 501, + 481, + 533 + ], + "spans": [ + { + "bbox": [ + 132, + 501, + 481, + 533 + ], + "type": "text", + "content": "21. Li, J., Li, D., Xiong, C., Hoi, S.: Blip: Bootstrapping language-image pre-training for unified vision-language understanding and generation. In: International Conference on Machine Learning (ICML). pp. 12888-12900. PMLR (2022)" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 132, + 534, + 481, + 566 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 534, + 481, + 566 + ], + "spans": [ + { + "bbox": [ + 132, + 534, + 481, + 566 + ], + "type": "text", + "content": "22. Lin, T., Xu, H., Yang, C., Xu, Y.: Interventional multi-instance learning with deconfounded instance-level prediction. In: Proceedings of the AAAI Conference on Artificial Intelligence (AAAI). vol. 36, pp. 1601-1609 (2022)" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 132, + 567, + 481, + 610 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 567, + 481, + 610 + ], + "spans": [ + { + "bbox": [ + 132, + 567, + 481, + 610 + ], + "type": "text", + "content": "23. Lin, T., Yu, Z., Hu, H., Xu, Y., Chen, C.W.: Interventional bag multi-instance learning on whole-slide pathological images. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR). pp. 19830-19839 (2023)" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 132, + 610, + 481, + 666 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 610, + 481, + 666 + ], + "spans": [ + { + "bbox": [ + 132, + 610, + 481, + 666 + ], + "type": "text", + "content": "24. Lu, M.Y., Chen, B., Zhang, A., Williamson, D.F., Chen, R.J., Ding, T., Le, L.P., Chuang, Y.S., Mahmood, F.: Visual language pretrained multiple instance zero-shot transfer for histopathology images. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR). pp. 19764-19775 (2023)" + } + ] + } + ], + "index": 14 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "spans": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "type": "text", + "content": "16" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 166, + 91, + 215, + 102 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 166, + 91, + 215, + 102 + ], + "spans": [ + { + "bbox": [ + 166, + 91, + 215, + 102 + ], + "type": "text", + "content": "L. Qu et al." + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 15 + }, + { + "para_blocks": [ + { + "bbox": [ + 130, + 116, + 482, + 665 + ], + "type": "list", + "angle": 0, + "index": 17, + "blocks": [ + { + "bbox": [ + 130, + 116, + 482, + 149 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 116, + 482, + 149 + ], + "spans": [ + { + "bbox": [ + 130, + 116, + 482, + 149 + ], + "type": "text", + "content": "25. Lu, M.Y., Williamson, D.F., Chen, T.Y., Chen, R.J., Barbieri, M., Mahmood, F.: Data-efficient and weakly supervised computational pathology on whole-slide images. Nature Biomedical Engineering 5(6), 555-570 (2021)" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 130, + 149, + 482, + 192 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 149, + 482, + 192 + ], + "spans": [ + { + "bbox": [ + 130, + 149, + 482, + 192 + ], + "type": "text", + "content": "26. Qu, L., Liu, S., Liu, X., Wang, M., Song, Z.: Towards label-efficient automatic diagnosis and analysis: a comprehensive survey of advanced deep learning-based weakly-supervised, semi-supervised and self-supervised techniques in histopathological image analysis. Physics in Medicine & Biology (2022)" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 130, + 192, + 482, + 224 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 192, + 482, + 224 + ], + "spans": [ + { + "bbox": [ + 130, + 192, + 482, + 224 + ], + "type": "text", + "content": "27. Qu, L., Luo, X., Fu, K., Wang, M., Song, Z.: The rise of ai language pathologists: Exploring two-level prompt learning for few-shot weakly-supervised whole slide image classification. arXiv preprint arXiv:2305.17891 (2023)" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 130, + 224, + 482, + 257 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 224, + 482, + 257 + ], + "spans": [ + { + "bbox": [ + 130, + 224, + 482, + 257 + ], + "type": "text", + "content": "28. Qu, L., Luo, X., Liu, S., Wang, M., Song, Z.: Dgmil: Distribution guided multiple instance learning for whole slide image classification. In: Medical Image Computing and Computer Assisted Intervention (MICCAI). pp. 24-34. Springer (2022)" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 130, + 257, + 482, + 300 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 257, + 482, + 300 + ], + "spans": [ + { + "bbox": [ + 130, + 257, + 482, + 300 + ], + "type": "text", + "content": "29. Radford, A., Kim, J.W., Hallacy, C., Ramesh, A., Goh, G., Agarwal, S., Sastry, G., Askell, A., Mishkin, P., Clark, J., et al.: Learning transferable visual models from natural language supervision. In: International Conference on Machine Learning (ICML). pp. 8748-8763. PMLR (2021)" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 130, + 300, + 482, + 332 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 300, + 482, + 332 + ], + "spans": [ + { + "bbox": [ + 130, + 300, + 482, + 332 + ], + "type": "text", + "content": "30. Rony, J., Belharbi, S., Dolz, J., Ayed, I.B., McCaffrey, L., Granger, E.: Deep weakly-supervised learning methods for classification and localization in histology images: a survey. arXiv preprint arXiv:1909.03354 (2019)" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 130, + 332, + 482, + 374 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 332, + 482, + 374 + ], + "spans": [ + { + "bbox": [ + 130, + 332, + 482, + 374 + ], + "type": "text", + "content": "31. Shao, Z., Bian, H., Chen, Y., Wang, Y., Zhang, J., Ji, X., et al.: Transmil: Transformer based correlated multiple instance learning for whole slide image classification. Advances in Neural Information Processing Systems (NeurIPS) 34, 2136-2147 (2021)" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 130, + 374, + 482, + 407 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 374, + 482, + 407 + ], + "spans": [ + { + "bbox": [ + 130, + 374, + 482, + 407 + ], + "type": "text", + "content": "32. Shi, X., Xing, F., Xie, Y., Zhang, Z., Cui, L., Yang, L.: Loss-based attention for deep multiple instance learning. In: Proceedings of the AAAI Conference on Artificial Intelligence (AAAI). vol. 34, pp. 5742-5749 (2020)" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 130, + 407, + 482, + 439 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 407, + 482, + 439 + ], + "spans": [ + { + "bbox": [ + 130, + 407, + 482, + 439 + ], + "type": "text", + "content": "33. Song, A.H., Jaume, G., Williamson, D.F., Lu, M.Y., Vaidya, A., Miller, T.R., Mahmood, F.: Artificial intelligence for digital and computational pathology. Nature Reviews Bioengineering pp. 1-20 (2023)" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 130, + 439, + 482, + 472 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 439, + 482, + 472 + ], + "spans": [ + { + "bbox": [ + 130, + 439, + 482, + 472 + ], + "type": "text", + "content": "34. Tu, C., Zhang, Y., Ning, Z.: Dual-curriculum contrastive multi-instance learning for cancer prognosis analysis with whole slide images. Advances in Neural Information Processing Systems (NeurIPS) 35, 29484-29497 (2022)" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 130, + 472, + 482, + 514 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 472, + 482, + 514 + ], + "spans": [ + { + "bbox": [ + 130, + 472, + 482, + 514 + ], + "type": "text", + "content": "35. Wang, X., Xiang, J., Zhang, J., Yang, S., Yang, Z., Wang, M.H., Zhang, J., Yang, W., Huang, J., Han, X.: Scl-wc: Cross-slide contrastive learning for weakly-supervised whole-slide image classification. Advances in Neural Information Processing Systems (NeurIPS) 35, 18009-18021 (2022)" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 130, + 514, + 482, + 557 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 514, + 482, + 557 + ], + "spans": [ + { + "bbox": [ + 130, + 514, + 482, + 557 + ], + "type": "text", + "content": "36. Wasim, S.T., Naseer, M., Khan, S., Khan, F.S., Shah, M.: Vita-clip: Video and text adaptive clip via multimodal prompting. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR). pp. 23034-23044 (2023)" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 130, + 557, + 482, + 601 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 557, + 482, + 601 + ], + "spans": [ + { + "bbox": [ + 130, + 557, + 482, + 601 + ], + "type": "text", + "content": "37. Xu, G., Song, Z., Sun, Z., Ku, C., Yang, Z., Liu, C., Wang, S., Ma, J., Xu, W.: Camel: A weakly supervised learning framework for histopathology image segmentation. In: Proceedings of the IEEE/CVF International Conference on Computer Vision (ICCV). pp. 10682-10691 (2019)" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 130, + 601, + 482, + 633 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 601, + 482, + 633 + ], + "spans": [ + { + "bbox": [ + 130, + 601, + 482, + 633 + ], + "type": "text", + "content": "38. Yao, H., Zhang, R., Xu, C.: Visual-language prompt tuning with knowledge-guided context optimization. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR). pp. 6757-6767 (2023)" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 130, + 633, + 482, + 665 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 633, + 482, + 665 + ], + "spans": [ + { + "bbox": [ + 130, + 633, + 482, + 665 + ], + "type": "text", + "content": "39. Yao, J., Zhu, X., Jonnagaddala, J., Hawkins, N., Huang, J.: Whole slide images based cancer survival prediction using attention guided deep multiple instance learning networks. Medical Image Analysis 65, 101789 (2020)" + } + ] + } + ], + "index": 16 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 184, + 91, + 447, + 102 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 184, + 91, + 447, + 102 + ], + "spans": [ + { + "bbox": [ + 184, + 91, + 447, + 102 + ], + "type": "text", + "content": "Pathology-knowledge Enhanced Multi-instance Prompt Learning" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 470, + 92, + 480, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 470, + 92, + 480, + 100 + ], + "spans": [ + { + "bbox": [ + 470, + 92, + 480, + 100 + ], + "type": "text", + "content": "17" + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 16 + }, + { + "para_blocks": [ + { + "bbox": [ + 132, + 116, + 481, + 259 + ], + "type": "list", + "angle": 0, + "index": 6, + "blocks": [ + { + "bbox": [ + 133, + 116, + 481, + 160 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 116, + 481, + 160 + ], + "spans": [ + { + "bbox": [ + 133, + 116, + 481, + 160 + ], + "type": "text", + "content": "40. Zhang, H., Meng, Y., Zhao, Y., Qiao, Y., Yang, X., Coupland, S.E., Zheng, Y.: Dtdf-mil: Double-tier feature distillation multiple instance learning for histopathology whole slide image classification. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR). pp. 18802-18812 (2022)" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 133, + 161, + 481, + 205 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 161, + 481, + 205 + ], + "spans": [ + { + "bbox": [ + 133, + 161, + 481, + 205 + ], + "type": "text", + "content": "41. Zhang, Y., Gao, J., Zhou, M., Wang, X., Qiao, Y., Zhang, S., Wang, D.: Text-guided foundation model adaptation for pathological image classification. In: International Conference on Medical Image Computing and Computer-Assisted Intervention (MICCAI). pp. 272-282. Springer (2023)" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 132, + 205, + 481, + 226 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 205, + 481, + 226 + ], + "spans": [ + { + "bbox": [ + 132, + 205, + 481, + 226 + ], + "type": "text", + "content": "42. Zhou, K., Yang, J., Loy, C.C., Liu, Z.: Learning to prompt for vision-language models. International Journal of Computer Vision 130(9), 2337-2348 (2022)" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 132, + 227, + 481, + 259 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 227, + 481, + 259 + ], + "spans": [ + { + "bbox": [ + 132, + 227, + 481, + 259 + ], + "type": "text", + "content": "43. Zhu, X., Yao, J., Zhu, F., Huang, J.: Wsisa: Making survival prediction from whole slide histopathological images. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR). pp. 7234-7242 (2017)" + } + ] + } + ], + "index": 5 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "spans": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "type": "text", + "content": "18" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 166, + 91, + 215, + 102 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 166, + 91, + 215, + 102 + ], + "spans": [ + { + "bbox": [ + 166, + 91, + 215, + 102 + ], + "type": "text", + "content": "L. Qu et al." + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 17 + } + ], + "_backend": "vlm", + "_version_name": "2.6.4" +} \ No newline at end of file diff --git a/2024/Paying More Attention to Images_ A Training-Free Method for Alleviating Hallucination in LVLMs/71543d84-0242-4633-b85d-0ad4feeefd05_content_list.json b/2024/Paying More Attention to Images_ A Training-Free Method for Alleviating Hallucination in LVLMs/71543d84-0242-4633-b85d-0ad4feeefd05_content_list.json new file mode 100644 index 0000000000000000000000000000000000000000..a42c5f23f3a50fc5b02d7df8689edbfd1b33c0b0 --- /dev/null +++ b/2024/Paying More Attention to Images_ A Training-Free Method for Alleviating Hallucination in LVLMs/71543d84-0242-4633-b85d-0ad4feeefd05_content_list.json @@ -0,0 +1,1681 @@ +[ + { + "type": "text", + "text": "Paying More Attention to Image: A Training-Free Method for Alleviating Hallucination in LVLMs", + "text_level": 1, + "bbox": [ + 217, + 140, + 785, + 186 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Shi Liu $^{1}$ , Kecheng Zheng $^{1,2\\dagger}$ , and Wei Chen $^{1\\dagger}$", + "bbox": [ + 333, + 210, + 666, + 227 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "$^{1}$ State Key Lab of CAD&CG, Zhejiang University", + "bbox": [ + 328, + 238, + 671, + 253 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "2 Ant Group", + "bbox": [ + 455, + 253, + 545, + 267 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "{liushi0927,zkechengzk}@gmail.com", + "bbox": [ + 372, + 268, + 630, + 281 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "chenvis@zju.edu.cn", + "bbox": [ + 431, + 282, + 571, + 295 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Abstract. Existing Large Vision-Language Models (LVLMs) primarily align image features of vision encoder with Large Language Models (LLMs) to leverage their superior text generation capabilities. However, the scale disparity between vision encoder and language model may led to LLMs assuming a predominant role in multi-modal comprehension. This imbalance in LVLMs may result in the instances of hallucinatory. Concretely, LVLMs may generate consistent descriptions with or without visual input, indicating that certain outputs are influenced solely by context text. We refer to this phenomenon as \"text inertia.\" To counteract this issue, we introduce a training-free algorithm to find an equilibrium point between image comprehension and language inference. Specifically, we adaptively involve adjusting and amplifying the attention weights assigned to image tokens, thereby granting greater prominence to visual elements. Meanwhile, we subtract the logits of multi-modal inputs from ones of pure text input, which can help LVLMs be not biased towards LLMs. By enhancing images tokens and reducing the stubborn output of LLM, we can let LVLM pay more attention to images, towards alleviating text inertia and reducing the hallucination in LVLMs. Our extensive experiments show that this method substantially reduces the frequency of hallucinatory outputs in various LVLMs in terms of different metrics. Project page is available at https://1albj.github.io/projects/PAI/", + "bbox": [ + 259, + 325, + 743, + 619 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Keywords: Vision-Language Models $\\cdot$ Visual Dialogs $\\cdot$ Hallucination Mitigation", + "bbox": [ + 259, + 630, + 740, + 657 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "1 Introduction", + "text_level": 1, + "bbox": [ + 215, + 681, + 375, + 698 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Recently, Large Vision-Language Models (LVLMs) have made significant strides, exhibiting impressive capabilities across a multitude of tasks [11,15,36,42]. However, these models still struggle with the phenomenon of hallucination. Specifically, there is often a mismatch between the textual content generated by the model and the actual visual input it receives [23].", + "bbox": [ + 212, + 710, + 787, + 787 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Hallucination in LVLMs is often attributed to issues with modality alignment, leading to the development of mitigating strategies through alignment", + "bbox": [ + 212, + 787, + 785, + 818 + ], + "page_idx": 0 + }, + { + "type": "page_footnote", + "text": "† Corresponding authors.", + "bbox": [ + 230, + 824, + 403, + 840 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/917107f1f4568805fcb7915f42da9f1fe4a850bc295b6797be7ad9aaba2190b8.jpg", + "image_caption": [ + "Fig. 1: We present an examination of various input settings, with hallucinations specifically highlighted in red. (a) When using LLAVA for image description, it generates a hallucinated description. (b) Even without image input, when only the historical response preceding the hallucinated description is input to LLAVA, it reproduces the same hallucinated description, a phenomenon we refer to as \"Text Inertia\". (c) Our proposed method, PAI, effectively mitigates this text inertia problem and yields accurate descriptions. (d) Utilizing PAI for image description results in a significantly more precise description." + ], + "image_footnote": [], + "bbox": [ + 225, + 146, + 781, + 308 + ], + "page_idx": 1 + }, + { + "type": "image", + "img_path": "images/8ac3098831f766ebecc7681dddac6a072a8bce51a13fa9a9862ff6938a42c29e.jpg", + "image_caption": [ + "Fig.2: Percentage of text inertia hallucination in all hallucination (calculated with 500 samples). For specific calculation processes, please refer to the supplementary material section A.", + "Fig. 3: Visualization of the average attention ratio of different content. The x-axis denotes the sequence length of the history tokens. The lengths of the image, instruction, BOS, and outlier tokens [9] are all fixed as they are part of the model input, with 576 tokens for the image, 21 for the instruction, and 1 each for the BOS and outlier tokens." + ], + "image_footnote": [], + "bbox": [ + 218, + 458, + 490, + 584 + ], + "page_idx": 1 + }, + { + "type": "image", + "img_path": "images/a2dd0958982eb6cd370c48df41b6d22f3e71098f63f8931d166bf9fc1e3a3984.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 508, + 448, + 782, + 532 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "training optimization [7, 20, 31]. However, is hallucination in LVLMs merely a result of the model's capacity, and can it only be alleviated through additional training? We propose a scenario where LVLMs generate a hallucinated object description. Specifically, even when the image input is removed and only the generated text preceding the hallucinated object word is retained, the LVLMs persist in producing the same hallucinated description, as depicted in Fig. 1.", + "bbox": [ + 212, + 702, + 787, + 792 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "In order to empirically investigate this behavior, we conducted tests on three LVLMs within the context of image describing tasks on the COCO dataset. We identified and conducted a statistical analysis on instances where LVLMs", + "bbox": [ + 212, + 795, + 787, + 839 + ], + "page_idx": 1 + }, + { + "type": "page_number", + "text": "2", + "bbox": [ + 217, + 114, + 228, + 126 + ], + "page_idx": 1 + }, + { + "type": "header", + "text": "Liu et al.", + "bbox": [ + 271, + 114, + 334, + 126 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "generated identical hallucinated object descriptions, even when the input was exclusively historical response text without any image. The observation from Fig. 2 clearly indicates that, even with the application of rigorous identification settings, the phenomenon continues to represent a substantial proportion.", + "bbox": [ + 212, + 146, + 782, + 205 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "We referred this phenomenon as \"Text Inertia\". Our hypothesis is that text inertia arises due to the current generative paradigms mapping image representations onto the text representation space as text tokens. In this mechanism, the LLM becomes the dominant character, and the inference process lacks additional handling of image tokens, leading to their neglect during the generation process. To validate this hypothesis, we have analyzed the attention values ratios of the LLaVA model during the inference process in Fig. 3. Our findings show that despite image tokens occupying a significant proportion, they do not receive substantial attention under the current mechanism. This multimodal chat resembles more of an automatic completion based on context rather than a continuous attention to the image for completion.", + "bbox": [ + 212, + 207, + 784, + 372 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "To close this gap, we introduce a method referred to as Pay Attention to Image (PAI). At a high level, PAI intervenes in the inference process to make it more image-centric, following the original image perception direction. To achieve this, we focus on the self-attention heads in the decoder layers of LVLMs. We enhance the attention weights for image tokens in their original directions during inference. This allows us to use the updated attention matrix to calculate the hidden states for the generated token, thereby incorporating more consideration for image representation during the generation process. To further mitigate text inertia, we construct the input using instruction tokens and historical response tokens, and subtract the model logits of this input from the logits of original model with image tokens input. This strategy helps to reduce the influence of language priors during the generation process. Unlike previous methods for mitigating hallucination that require additional training or external tools, our approach is training-free. Moreover, we are the first to propose an inference intervention method for mitigating hallucination in LVLMs.", + "bbox": [ + 212, + 373, + 785, + 598 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "We then evaluate the response accuracy in the image description task from long sequence generation perspective by employing the CHAIR metric [29] and GPT-4V. In addition, we use POPE [19] and MMHal-Bench [32] to more comprehensively evaluate the model's hallucination performance on VQA task. Furthermore, we have constructed single-turn and multi-turn chat evaluations for POPE. Since our model intervenes in the inference process, it can be used for any decoding method. Therefore, we conducted experiments on three decoding methods of the three models. The experimental results proved the effectiveness of our method in mitigating hallucinations.", + "bbox": [ + 212, + 599, + 785, + 734 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "2 Related Work", + "text_level": 1, + "bbox": [ + 215, + 756, + 387, + 771 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "2.1 Large Vision-Language Models", + "text_level": 1, + "bbox": [ + 215, + 786, + 516, + 801 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "The development of pre-training techniques [4, 27] and instruction tuning techniques [26, 37] has rapidly advanced LLMs technology, such as LLaMA [33] and", + "bbox": [ + 212, + 809, + 782, + 839 + ], + "page_idx": 2 + }, + { + "type": "header", + "text": "PAI", + "bbox": [ + 702, + 114, + 730, + 126 + ], + "page_idx": 2 + }, + { + "type": "page_number", + "text": "3", + "bbox": [ + 774, + 114, + 785, + 126 + ], + "page_idx": 2 + }, + { + "type": "image", + "img_path": "images/fadb7576ac2c6adb7a89282889adfa54b732b07b7997bf7c889672fcdf9f7d64.jpg", + "image_caption": [ + "Fig. 4: Architecture of our PAI. To alleviate text inertia, we additionally construct an input without image. Throughout the forward inference process, we amplify the focus on the image token by edit self-attention maps in LLaMA. Ultimately, we subtract the logits distribution of the language prior during decoding to achieve an accurate description." + ], + "image_footnote": [], + "bbox": [ + 228, + 156, + 782, + 282 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Vicuna [28], further leading to the prosperity of LVLMs technology. Early works, like Flamingo [1] and BLIP-2 [17], have successfully adapted LLMs to visual tasks, demonstrating notable generative capabilities and in-context learning abilities. Recently, the capabilities of LVLMs have further advanced under the influence of visual instruction tuning techniques [24,25]. Using different projectors to map images to the text domain, thereby endowing language generation models with image understanding capabilities, is also a hot research topic [8,25,39,44]. Additionally, several studies focus on visual language tasks such as grounding capabilities [6] and reasoning capabilities [15]. However, recent LVLMs still face the issue of hallucination generation [21].", + "bbox": [ + 212, + 416, + 787, + 568 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "2.2 Mitigation of LVLMs Hallucination", + "text_level": 1, + "bbox": [ + 214, + 589, + 553, + 604 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Hallucination in LVLMs refers to contradictions between the image input and the textual output. Various methods have been proposed to mitigate hallucination. The most direct reason for hallucination generation is that hallucination arises from data bias and knowledge gaps between vision and language. Therefore, better data filtering methods [12, 22, 41] and higher quality annotated data [2] are introduced. Simultaneously, these methods also imply the need for more alignment training [31] or adjustments in the model architecture [7, 20]. These methods can achieve good results, but they are time-consuming and require high computational resources.", + "bbox": [ + 212, + 613, + 787, + 747 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Apart from addressing the ability of LVLMs itself, hallucination can also be mitigated through post-processing methods. This method usually involves using an additional module or external tools to edit the response. Recent methods such as LURE [43] utilize additional data to train a state detector and when hallucination issues are detected, content is regenerated by a revisor model. Woodpecker [40] introduces an external visual model to inspect entities extracted from", + "bbox": [ + 212, + 750, + 787, + 840 + ], + "page_idx": 3 + }, + { + "type": "page_number", + "text": "4", + "bbox": [ + 217, + 114, + 228, + 126 + ], + "page_idx": 3 + }, + { + "type": "header", + "text": "Liu et al.", + "bbox": [ + 271, + 114, + 334, + 126 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "the response, and then the detection results are handed over to the generation model to regenerate better answers. These methods also extend the inference chain and increase inference costs.", + "bbox": [ + 212, + 146, + 782, + 190 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Training-free hallucination mitigation methods have so far only been attempted in decoding methods. OPERA [13] discovered an abnormal attention pattern that accompanies model decoding. It was statistically found that this pattern often accompanies hallucination descriptions, and thus a detection and mitigation method was proposed based on this pattern to alleviate the hallucination faced by the model. VCD [16] introduced the notion that visual uncertainty increases hallucination descriptions and, based on this discovery, proposed a contrast decoding method to alleviate hallucination issues.", + "bbox": [ + 212, + 191, + 787, + 313 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "3 Preliminaries", + "text_level": 1, + "bbox": [ + 215, + 335, + 382, + 351 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "The architecture of LVLMs typically comprises of three main components: an image encoder, a projector, and a language decoder. Both the image encoder and language decoder are usually pre-trained. The image encoder is employed to transform images into image tokens, which are subsequently mapped to the text representation space by the projector. This process enables the concatenation of image tokens with text tokens that are then fed into the language decoder. The language decoder subsequently generates corresponding responses based on the provided instructions.", + "bbox": [ + 212, + 366, + 787, + 487 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "The existing projectors. Currently, projectors predominantly fall into two categories: linear projectors and resamplers. A projector takes N visual features from the image encoder and transforms them into M visual tokens. The linear projector employs a multilayer perceptron to transform all visual features, maintaining a one-to-one transformation which meaning that M equals N. In contrast, the resampler does not preserve all visual features but instead samples visual cues (M, where $\\mathrm{M} < \\mathrm{N}$ ). For instance, Q-former [17] utilizes M learnable queries and Bert [10] to extract information from visual features. Given that the knowledge of images during the generation process solely originates from the output image tokens of the projectors, our attention is concentrated on the image tokens post-projection, irrespective of their preceding modeling process.", + "bbox": [ + 212, + 488, + 787, + 654 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Autoregressive language decoders. Nearly all LVLMs adopt LLaMA-family models as their language decoders, which employ the self-attention mechanism. The visual tokens processed by the projector are concatenated with text tokens and fed into the LLaMA, which carries out the forward decoding process. From the perspective of a single attention head in a single layer, each head repeatedly performs the following attention operation with the same input shape:", + "bbox": [ + 212, + 654, + 787, + 746 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\n\\boldsymbol {O} _ {h} = \\boldsymbol {A} _ {h} \\boldsymbol {V} _ {h}, \\quad \\boldsymbol {A} _ {h} = \\operatorname {s o f t m a x} \\left(\\frac {\\boldsymbol {Q} _ {h} \\boldsymbol {K} _ {h} ^ {\\top}}{\\sqrt {d _ {k}}}\\right). \\tag {1}\n$$\n", + "text_format": "latex", + "bbox": [ + 352, + 756, + 785, + 790 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Each attention head $h$ performs an attention operation using its own set of queries $Q_{h} \\in \\mathbb{R}^{n \\times d_{k}}$ , keys $K_{h} \\in \\mathbb{R}^{n \\times d_{k}}$ , and values $V_{h} \\in \\mathbb{R}^{n \\times d_{k}}$ , where $n$ represents the sequence length and $d_{k}$ represents the hidden dimensions. The", + "bbox": [ + 212, + 794, + 787, + 840 + ], + "page_idx": 4 + }, + { + "type": "header", + "text": "PAI", + "bbox": [ + 702, + 114, + 730, + 126 + ], + "page_idx": 4 + }, + { + "type": "page_number", + "text": "5", + "bbox": [ + 774, + 114, + 785, + 126 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "output $O_{h} \\in \\mathbb{R}^{n \\times d_{k}}$ is modeled by multiplying $V_{h}$ and the attention weights $A_{h} \\in \\mathbb{R}^{n \\times n}$ , where each row represents the weights for each token during feature mixing. This operation enables the model to focus on different parts of the input for each head through attention weights, thereby capturing various parts of the information from the sequence token representations. The final output is the current generated token vocabulary conditional probability distribution $y \\in \\mathbb{R}^{v}$ based on the input instruction representations $X_{I}$ , image representations $X_{V}$ and history generated token representations $X_{H}$ , where $v$ is the size of the vocabulary. This process can be formatted as:", + "bbox": [ + 212, + 146, + 787, + 282 + ], + "page_idx": 5 + }, + { + "type": "equation", + "text": "\n$$\n\\begin{array}{l} \\boldsymbol {y} \\sim p _ {\\text {m o d e l}} (\\boldsymbol {y} \\mid \\mathbf {X} _ {I}, \\mathbf {X} _ {V}, \\mathbf {X} _ {H}), \\\\ \\begin{array}{l} p _ {\\text {m o d e r}} (\\boldsymbol {y} \\mid 1, 1, 1, 1, 1, 1, 1, 1, 1), \\\\ \\propto \\operatorname {s o f t m a x} \\left(\\operatorname {l o g i t} _ {\\text {m o d e l}} (\\boldsymbol {y} \\mid \\mathbf {X} _ {I}, \\mathbf {X} _ {V}, \\mathbf {X} _ {H})\\right), \\end{array} \\tag {2} \\\\ \\end{array}\n$$\n", + "text_format": "latex", + "bbox": [ + 349, + 292, + 784, + 327 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "which calculates the distribution of one token and iterates for the entire response. The sequence generation continues until an EOS (End of Sentence) token is produced, marking the end of the generation and resulting in a complete response.", + "bbox": [ + 212, + 335, + 784, + 383 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "4 Method", + "text_level": 1, + "bbox": [ + 214, + 393, + 330, + 409 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "At the core of our method is a solution for image neglect and text inertia, both of which are fundamentally interconnected. Essentially, as paying more attention to the image, there is a corresponding reduction in the reliance on language priors. Intuitively, in a conversation centered around an image, the model should devote more attention to the image, thereby allowing it to have a significant impact on the response. As such, we identify the self-attention map in the token-level generation and augment the image attention in its original directions. This strategy promotes a more image-centric latent representations. Additionally, to further mitigate the influence of text inertia, we devide the logits distribution of pure text input into the model's output.", + "bbox": [ + 212, + 417, + 787, + 568 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "4.1 Pay More Attention to Image", + "text_level": 1, + "bbox": [ + 214, + 589, + 508, + 604 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Extracting the Self-Attention Matrix. We start from a token-level perspective. The response process in LVLMs is fundamentally generated token by token. Each token is generated based on the input image, instruction, and the historically generated response. This process is facilitated through a multi-layer attention decoder architecture. Consequently, this results in a probability distribution of the vocabulary for the currently generated token. Our goal is to extract the attention matrix of each attention head at every layer, indicating the influence of each content during inference.", + "bbox": [ + 212, + 613, + 787, + 733 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "When generating the $k$ -th token in the sequence, the input representation for the attention head in the forward process includes the instruction representation $\\mathbf{X}_I = [\\pmb{x}_{i_1},\\dots,\\pmb{x}_{i_{n_I}}]$ , image representation $\\mathbf{X}_V = [\\pmb{x}_{v_1},\\dots,\\pmb{x}_{v_{n_V}}]$ , and the representation of the historically generated response $\\mathbf{X}_H = [\\pmb{x}_{h_1},\\dots,\\pmb{x}_{h_{n_H}}]$ . Notably, the image representation considered here is the one that has been processed by the projector. Essentially, the hidden states of each input layer are $\\mathbf{X} = \\mathrm{concat}(\\mathbf{X}_I[1:m],\\mathbf{X}_V,\\mathbf{X}_I[m + 1:n_I],\\mathbf{X}_H)$ , where the notation $\\mathbf{X}_I[1:m]$", + "bbox": [ + 212, + 734, + 787, + 842 + ], + "page_idx": 5 + }, + { + "type": "page_number", + "text": "6", + "bbox": [ + 217, + 114, + 228, + 126 + ], + "page_idx": 5 + }, + { + "type": "header", + "text": "Liu et al.", + "bbox": [ + 271, + 114, + 334, + 127 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "indicates the first $m$ elements in the instruction representation. Each attention head assigns different degrees of attention to each element during the current token representation decoding process. Our aim is to enhance the attention paid to the image. Therefore, we extract the attention weight values related to the image token for the current generated token, intervene, and then redistribute the attention values of each element through softmax.", + "bbox": [ + 212, + 146, + 784, + 234 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Excite model in a trustful direction. There have been attempts in some LLMs works to make the answers generated by LLMs more trustworthy by means of intervention [3, 5, 14, 18, 34]. The implemented approach typically involves intervening with the hidden states. As for defining what constitutes a more trustworthy direction, it usually requires additional projection and training to probe this trustful direction. In our case, a response that is more image-based is considered more trustworthy. Since LVLMs have undergone alignment training, the original attention values provide a direction based on image content. As illustrated in Fig. 4, by amplifying the attention values of image tokens based on the original attention values, we can enhance the trustworthiness of our results.", + "bbox": [ + 212, + 238, + 785, + 388 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Another nuance involves our avoidance of choosing the attention head used to shift intervention. In the ITI method [18], it is stated that not all attention heads should be subjected to intervention. Therefore, they introduce a trustful score to rank each head across all layers and select the top-k heads for intervention. In our case, the less trustworthy heads with lower attention values receive less intervention. We first extract the attention weights of the image tokens for the current generated token from the attention weights $\\tilde{\\pmb{A}}$ before softmax operation. We then use the hyper-parameter $\\alpha$ to control the step size for intervention. From a single attention head perspective, our method can be expressed as follows:", + "bbox": [ + 212, + 388, + 785, + 525 + ], + "page_idx": 6 + }, + { + "type": "equation", + "text": "\n$$\n\\tilde {\\boldsymbol {A}} _ {n, j} = \\tilde {\\boldsymbol {A}} _ {n, j} + \\alpha \\cdot | \\tilde {\\boldsymbol {A}} _ {n, j} | \\quad \\text {f o r} j = m + 1 \\text {t o} m + n _ {V}. \\tag {3}\n$$\n", + "text_format": "latex", + "bbox": [ + 316, + 537, + 784, + 554 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "The model's final vocabulary probability distribution is derived from the projection of the hidden states of the last token in the sequence. Therefore, we extract the attention weights of the last token $n$ on the image tokens by indexing $\\tilde{A}_{n,j}$ . Following the intervention, we use the softmax function to redistribute the attention values of each token during the reassignment of encoded hidden states. This procedure is repeated for each subsequent token prediction in an autoregressive manner and is independent of the choice of the decoding algorithm.", + "bbox": [ + 212, + 566, + 785, + 672 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Excite more precisely with an attention mode prior. The presence of the BOS token, an attention sink pattern [38], in a sentence results in higher attention values during the attention computation process, which may seem counterintuitive. The BOS token typically signifies the start of a sentence and, as such, does not carry significant semantic content. However, the generation of tokens is significantly influenced by this particular token, a similar pattern that also manifests itself in visual models [9]. As mentioned in StreamLLM [38], the pattern of the attention sink emerges when redundant attention values are present.", + "bbox": [ + 212, + 672, + 785, + 808 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Naturally, one might infer that when the sink pattern appears, we excite the image token. To further investigate this phenomenon, as depicted in Fig. 5, we", + "bbox": [ + 212, + 809, + 785, + 839 + ], + "page_idx": 6 + }, + { + "type": "header", + "text": "PAI", + "bbox": [ + 702, + 114, + 730, + 126 + ], + "page_idx": 6 + }, + { + "type": "page_number", + "text": "7", + "bbox": [ + 774, + 114, + 785, + 126 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/648d051086d77520a3c5a9c8117931c2f8d35ec709ace9f6e518616a5cab6353.jpg", + "image_caption": [ + "(a) LLAVA" + ], + "image_footnote": [], + "bbox": [ + 240, + 145, + 406, + 255 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/8ae6e4405431b66600a296d0dd9f8126ae0de3dfb721da08a8110a1d8b02fdc1.jpg", + "image_caption": [ + "(b) Minigpt4", + "Fig. 5: The BOS token attention ratios of three model. We calculate the attention weights of the BOS token for each head in each layer and display them in a heat map." + ], + "image_footnote": [], + "bbox": [ + 416, + 145, + 581, + 256 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/e1acd51a8090e93741e905402c511679c0fd2f57a88cf72eb889f7ef1b475fb9.jpg", + "image_caption": [ + "(c) Shikra" + ], + "image_footnote": [], + "bbox": [ + 589, + 145, + 756, + 255 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "find that the sink phenomenon is not overtly evident in the shallow layers. This is because the shallow layers tend to focus more on encoding semantically rich information [35]. When the encoding of semantically rich tokens stabilizes, the attention sink phenomenon arises. Therefore, we build upon the judgement of intervention timing by calculating the similarity of the hidden states.", + "bbox": [ + 212, + 323, + 787, + 398 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "4.2 Image-Centric Logit Refine", + "text_level": 1, + "bbox": [ + 214, + 421, + 485, + 436 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "In Fig. 1, we observe a peculiar phenomenon where LVLMs continue to generate identical hallucinated text even when the image is removed from the input. This observation naturally leads us to the concept of using the output distribution (when no image is in input) as a reference to penalize our initial prediction distribution. Therefore, we update the distribution of the generated token by:", + "bbox": [ + 212, + 446, + 787, + 522 + ], + "page_idx": 7 + }, + { + "type": "equation", + "text": "\n$$\n\\begin{array}{l} p _ {\\text {m o d e l}} = \\gamma \\cdot p _ {\\text {m o d e l}} (\\boldsymbol {y} | \\boldsymbol {X} _ {V}, \\boldsymbol {X} _ {I}, \\boldsymbol {X} _ {H}) \\\\ - (\\gamma - 1) \\cdot p _ {\\text {m o d e l}} (\\boldsymbol {y} | \\boldsymbol {X} _ {I}, \\boldsymbol {X} _ {H}). \\\\ \\end{array}\n$$\n", + "text_format": "latex", + "bbox": [ + 359, + 532, + 638, + 568 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "This equation effectively reduces the predicted probability based on text alone. The weight $\\gamma$ is used to control the degree of penalty applied to the initial prediction distribution.", + "bbox": [ + 212, + 579, + 784, + 622 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "This operation is conceptually similar to LLM-CFG [30]. Essentially, it provides a guided generation mechanism that allows the model to make informed choices between outputs based on image content and those based on language logic. This way, the model can better balance the influence of visual and textual information in its outputs, leading to more contextually accurate and relevant results.", + "bbox": [ + 212, + 623, + 787, + 713 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "5 Experiments", + "text_level": 1, + "bbox": [ + 214, + 737, + 375, + 753 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "5.1 Setup", + "text_level": 1, + "bbox": [ + 214, + 768, + 310, + 784 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Baselines. We evaluate the effectiveness of our method on three different models. To better compare the impact of image feature tokens after different projectors on our method, we selected two models that use linear projectors, LLAVA", + "bbox": [ + 212, + 794, + 785, + 839 + ], + "page_idx": 7 + }, + { + "type": "page_number", + "text": "8", + "bbox": [ + 217, + 114, + 228, + 126 + ], + "page_idx": 7 + }, + { + "type": "header", + "text": "Liu et al.", + "bbox": [ + 271, + 114, + 334, + 126 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "and Shikra, as well as one model that uses resamplers, Minigpt4. Additionally, for a more convincing comparison, we report on three decoding methods for comparison: greedy, beam search, and nucleus sample. We also selected the OPERA [13] method, which is an improvement on beam search, and the VCD [16] method, which is an improvement on nucleus sampling, to compare with our results. We used the default hyperparameters from the open-source versions of these two methods.", + "bbox": [ + 212, + 146, + 782, + 251 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "Implementation Details. As different models have different lengths of image tokens, leading to different degrees of image neglect, to better align with the image sequence length of the model, we set $\\alpha = 0.5$ for LLAVA, $\\alpha = 0.6$ for Shikra with long image token sequence lengths and $\\alpha = 0.2$ for resampler models with short image token sequences. As text inertia is independent of the image token length, we continuously use $\\gamma = 1.1$ . Apart from this, in the beam search tests, the beam number is set to 5 for all methods, and in the nucleus sample tests, all the common parameters are consistent.", + "bbox": [ + 212, + 252, + 784, + 371 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "5.2 Benchmark & Evaluation Metrics", + "text_level": 1, + "bbox": [ + 215, + 395, + 539, + 410 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "CHAIR [29]. Caption Hallucination Assessment with Image Relevance (CHAIR) is a widely used metric in image captioning tasks. CHAIR operates by creating a set of ground-truth object labels for each image. Any object mentioned in the caption that does not exist in the label set is considered a hallucinated object. CHAIR comprises two evaluation dimensions: instance-level and sentence-level, represented as $\\mathrm{CHAIR_I}$ and $\\mathrm{CHAIR_S}$ , respectively. These are calculated in the following manner:", + "bbox": [ + 212, + 420, + 789, + 526 + ], + "page_idx": 8 + }, + { + "type": "equation", + "text": "\n$$\n\\mathrm {C H A I R} _ {\\mathrm {I}} = \\frac {| \\{\\text {h a l l u c i n a t e d o b j e c t s} \\} |}{\\text {a l l m e n t i o n e d o b j e c t s}}, \\tag {5}\n$$\n", + "text_format": "latex", + "bbox": [ + 372, + 537, + 784, + 569 + ], + "page_idx": 8 + }, + { + "type": "equation", + "text": "\n$$\n\\mathrm {C H A I R} _ {\\mathrm {S}} = \\frac {| \\{\\text {c a p t i o n s w i t h h a l l u c i n a t e d o b j e c t s} \\} |}{\\text {a l l c a p t i o n s}}. \\tag {6}\n$$\n", + "text_format": "latex", + "bbox": [ + 320, + 571, + 784, + 603 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "We conducted experiments based on the validation set of MSCOCO 2014. Given \"Please help me describe the image in detail.\" as prompt, we subsequently employ the CHAIR metric to evaluate the generated description. To evaluate hallucinations in long sequence generation, we adopted the same setup as used in LURE [43] and OPERA [13], setting the max_new_tokens parameter to 512 and randomly sampled 500 instances for evaluation.", + "bbox": [ + 212, + 612, + 803, + 703 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "POPE [19]. The Polling-based Object Probing Evaluation (POPE) is a evaluation metric designed in the VQA paradigm. POPE serves as a metric for assessing object hallucination, evaluating hallucinations by asking LVLMs questions such as \"Is there a in the image?\" Here, is replaced with the constructed ground-truth object from three different types of splits. In the \"random\" split, objects are randomly selected from the entire dataset for evaluation. In the \"popular\" split, objects are chosen from those most frequently appearing in the dataset. In the \"adversarial\" split, objects that are highly related to the image objects are selected for evaluation. We conduct our evaluation", + "bbox": [ + 212, + 704, + 785, + 839 + ], + "page_idx": 8 + }, + { + "type": "header", + "text": "PAI", + "bbox": [ + 702, + 114, + 730, + 126 + ], + "page_idx": 8 + }, + { + "type": "page_number", + "text": "9", + "bbox": [ + 774, + 114, + 785, + 126 + ], + "page_idx": 8 + }, + { + "type": "table", + "img_path": "images/0d64e374c2e394ad130407221af0fbec87aa43b5cd00c110c2c7492beb813861.jpg", + "table_caption": [ + "Table 1: CHAIR hallucination evaluation results on three LVLMs. CHAIR [29] is employed as the evaluation metric, where a smaller number indicates less hallucinations. OPERA is a decoding method based on beam search, VCD is a decoding method based on nucleus sampling, and PAI is an inference intervention method that can be collaborated with any decoding method." + ], + "table_footnote": [], + "table_body": "
DecodingMethodLLAVA [24]Minigpt4 [44]Shikra [6]
CHAIRSCHAIRICHAIRSCHAIRICHAIRSCHAIRI
GreedyVanilla46.613.432.811.151.214.4
PAI24.86.926.38.837.610.0
Beam SearchVanilla46.414.346.613.453.014.7
OPERA [13]44.614.430.19.836.812.4
PAI21.85.624.86.935.811.4
NucleusVanilla58.218.232.711.957.916.4
VCD [16]51.815.134.811.557.616.3
PAI43.414.726.710.349.913.2
", + "bbox": [ + 225, + 226, + 772, + 400 + ], + "page_idx": 9 + }, + { + "type": "text", + "text": "on the COCO dataset with 500 images, with each image having 6 questions for each split of POPE. We evaluate the performance of the model in object recognition tasks using both accuracy score and F1 score. Besides, in order to more comprehensively examine hallucinations in multi-turn chats, we have structured the evaluation in both single-turn and multi-turn dialogue forms.", + "bbox": [ + 212, + 430, + 787, + 505 + ], + "page_idx": 9 + }, + { + "type": "text", + "text": "MMHal-Bench [32]. For further evaluation of our method on some challenging datasets, we choose MMHal-Bench, which is designed with 96 image-question pairs, spread across 8 question categories $\\times 12$ object topics. It contains eight types of questions about object attributes, adversarial objects, comparisons, counting, spatial relations, environment, holistic descriptions, and others to comprehensively assess the model's hallucination performance on high-difficulty datasets. Essentially, it is also a VQA-based evaluation, but unlike the existence-based examination in POPE, its questions also include some logical considerations. Therefore, for evaluation on MMHal-Bench, we first need to answer questions and then use GPT-4 to score the answer based on the response and the ground-truth answer. The evaluation results include the model's scores across all question categories, and the overall score represents the average of these scores.", + "bbox": [ + 212, + 506, + 787, + 700 + ], + "page_idx": 9 + }, + { + "type": "text", + "text": "GPT-4v Assisted Evaluation. To further evaluate the model's performance in image description tasks, we can move beyond the CHAIR metric, which is based on information extraction and only considers object hallucination. We can use GPT-4v for open evaluation. As with previous evaluations [13,16], we sample 50 images on the COCO dataset for evaluation. We construct prompts, input images into GPT-4v, along with the descriptions responses from two assistants. GPT-4v evaluation takes into account two dimensions: Accuracy and Detailedness, denoted respectively as C and D. Detailed prompt construction can be found in the appendix.", + "bbox": [ + 212, + 703, + 787, + 839 + ], + "page_idx": 9 + }, + { + "type": "page_number", + "text": "10", + "bbox": [ + 217, + 114, + 235, + 126 + ], + "page_idx": 9 + }, + { + "type": "header", + "text": "Liu et al.", + "bbox": [ + 271, + 114, + 334, + 126 + ], + "page_idx": 9 + }, + { + "type": "table", + "img_path": "images/4bfdbee49f1ca74b2c86eb401c31c1f76d7a2ca8eeb890c526caaf61c83fc393.jpg", + "table_caption": [ + "Table 2: Quantitative comparison on POPE. The best results are in bold." + ], + "table_footnote": [], + "table_body": "
DecodingMethodLLAVA [24]Minigpt4 [44]Shikra [6]
Single-turnMulti-turnSingle-turnMulti-turnSingle-turnMulti-turn
AccF1AccF1AccF1AccF1AccF1AccF1
GreedyVanilla84.7685.5185.6984.2774.3973.7878.077.7781.5682.0578.1678.77
PAI85.8285.9787.5486.8375.276.280.8281.682.381.4778.5779.11
Beam SearchVanilla84.984.983.5781.1372.5467.7371.7867.9281.5281.8978.1878.13
PAI86.3385.8986.4285.2773.2269.3674.1571.7282.2581.0180.7280.02
NucleusVanilla80.2581.3282.8781.6357.5258.1664.4963.8979.5580.2876.5477.12
VCD [16]79.5381.0284.0383.359.9758.769.7269.3880.2280.576.3877.07
PAI81.7282.8785.0383.9759.1960.870.2470.2781.4980.7177.9978.47
", + "bbox": [ + 218, + 170, + 784, + 313 + ], + "page_idx": 10 + }, + { + "type": "text", + "text": "5.3 Experimental Results", + "text_level": 1, + "bbox": [ + 215, + 345, + 439, + 361 + ], + "page_idx": 10 + }, + { + "type": "text", + "text": "In this section, we analyze the performance of PAI across various hallucination evaluation tasks, including long image description, simplified VQA answer, construction of metric evaluation, and leveraging the near-human cognitive capabilities of GPT-4/GPT-4v as evaluation methods. For further analysis, please refer to the appendix.", + "bbox": [ + 212, + 380, + 782, + 455 + ], + "page_idx": 10 + }, + { + "type": "text", + "text": "Results on long sequence hallucination evaluation. The experimental results are presented in Tab. 1. As our approach is an inference intervention method, it differs from previous decoding hallucination mitigation methods that primarily concentrate on improving a single decoding method. We have tested our method on three decoding techniques. Our method has achieved hallucination mitigation on all three decoding methods used by the three models. However, when integrated with the nucleus, a sampling-based method, the hallucination reduction brought about by our method is not significant. This may be because even though our method has increased the priority of trustful tokens, the sample set during nucleus decoding still contains many hallucination tokens.", + "bbox": [ + 212, + 458, + 784, + 609 + ], + "page_idx": 10 + }, + { + "type": "text", + "text": "Moreover, while OPERA significantly mitigates hallucinations, its time efficiency is considerably higher compared to vanilla. In contrast, our method not only has almost the same time efficiency as vanilla, but it also performs better in reducing hallucination issues. Compared to VCD, during the generation process of long sequence tasks, the introduction of visual uncertainty during decoding sometimes leads to more hallucination descriptions. However, our method can reduce the proportion of hallucination words in the sample pool.", + "bbox": [ + 212, + 611, + 784, + 717 + ], + "page_idx": 10 + }, + { + "type": "text", + "text": "Results on single-turn and multi-turn hallucination evaluation. Unlike the CHAIR evaluation, POPE is in a VQA format, so the response is brief, answering only \"Yes\" or \"No\". The phenomena of text inertia and image neglect may not be as noticeable under this setting, especially for single-turn. However, our method still achieved a notable improvement compared to the vanilla decoding method in single-turn. As for multi-turn evaluation, which involves a longer context, our method achieves more significant improvement as presented in Tab. 2.", + "bbox": [ + 212, + 719, + 785, + 838 + ], + "page_idx": 10 + }, + { + "type": "header", + "text": "PAI", + "bbox": [ + 702, + 114, + 730, + 126 + ], + "page_idx": 10 + }, + { + "type": "page_number", + "text": "11", + "bbox": [ + 767, + 114, + 784, + 126 + ], + "page_idx": 10 + }, + { + "type": "image", + "img_path": "images/f1f43b14b4c3879f46e3ab454def14beab521b8b9cc2ecf5f8574b145088733b.jpg", + "image_caption": [ + "(a) LLAVA" + ], + "image_footnote": [], + "bbox": [ + 225, + 143, + 405, + 251 + ], + "page_idx": 11 + }, + { + "type": "image", + "img_path": "images/db8e51032d049fc4b4c221923991eefe4645b95b4e92f9ba524b5b6149be2c43.jpg", + "image_caption": [ + "(b) Minigpt4", + "Fig. 6: Quantitative comparison on the MMHal-Bench. Higher scores indicate better performance." + ], + "image_footnote": [], + "bbox": [ + 410, + 148, + 591, + 252 + ], + "page_idx": 11 + }, + { + "type": "image", + "img_path": "images/0b1940e754152c4bfb9fc2a83f6b1939375e7e116b782821e321fff8a7e32abe.jpg", + "image_caption": [ + "(c) Shikra" + ], + "image_footnote": [], + "bbox": [ + 596, + 148, + 779, + 251 + ], + "page_idx": 11 + }, + { + "type": "text", + "text": "Results on hallucination evaluation in comprehensive general scenarios. The experimental results, as shown in Fig. 6, indicate that for some more image-based question types, such as object attributes, adversarial objects, and holistic questions, the answers are more accurate when inference intervention with PAI is applied, and there is a certain degree of improvement across all models. However, for some logical questions, such as comparisons and relations, there is no noticeable improvement after intervention. In summary, through the overall metric, i.e., the average of the eight evaluation dimensions, there is a certain degree of improvement compared to the baseline after incorporating PAI.", + "bbox": [ + 212, + 335, + 784, + 470 + ], + "page_idx": 11 + }, + { + "type": "text", + "text": "Results on human-like GPT-4v assisted hallucination evaluation. The experimental results, as shown in Tab. 3, indicate that even when more comprehensive hallucination evaluation dimensions are added, our method, compared to the greedy decoding method, can provide more accurate responses on all three models without losing detail in the description. Given that GPT-4v's visual understanding and language logic capabilities have reached a level close to that of humans, it can more comprehensively illustrate the performance improvements brought about by our method.", + "bbox": [ + 212, + 470, + 785, + 592 + ], + "page_idx": 11 + }, + { + "type": "table", + "img_path": "images/6f26198cbd8df5efbbeec2499315474027d5a0dcaa2d395482e1896b1cdbe3bd.jpg", + "table_caption": [ + "Table 3: Results on GPT-4V evaluation. The best results are in bold." + ], + "table_footnote": [], + "table_body": "
MethodLLAVAMinigpt4Shikra
CDCDCD
Greedy5.625.245.85.745.545.25
PAI6.465.367.045.896.045.05
", + "bbox": [ + 357, + 643, + 638, + 723 + ], + "page_idx": 11 + }, + { + "type": "text", + "text": "5.4 Ablation Study", + "text_level": 1, + "bbox": [ + 215, + 771, + 388, + 786 + ], + "page_idx": 11 + }, + { + "type": "text", + "text": "Our method, PAI, consists of two stages of interventions. In the first stage, during forward inference, the hyperparameter $\\alpha$ is utilized to set the scale of intervention. Simultaneously, the layer prior, represented as \"L\", is used to determine the", + "bbox": [ + 212, + 794, + 784, + 839 + ], + "page_idx": 11 + }, + { + "type": "page_number", + "text": "12", + "bbox": [ + 217, + 114, + 235, + 126 + ], + "page_idx": 11 + }, + { + "type": "header", + "text": "Liu et al.", + "bbox": [ + 271, + 114, + 334, + 126 + ], + "page_idx": 11 + }, + { + "type": "table", + "img_path": "images/a21c849f909346b2ffbdb154eb71411642acb4c43012e7f3e6d68b933ab83621.jpg", + "table_caption": [ + "Table 4: Ablation Study of the Hyperparameter $\\alpha$ . When $\\alpha$ becomes excessively large, resulting in an unbalanced response, we terminate the experiment and denote this with a dash (-). F1 values that are considered outliers are highlighted in red." + ], + "table_footnote": [], + "table_body": "
αγLLLaVAMinigpt4Shikra
CHAIRSCHAIRiF1CHAIRSCHAIRiF1CHAIRSCHAIRiF1
---46.213.875.931.610.569.456.215.874.6
0.11.1✓47.413.876.925.08.069.956.415.075.7
0.21.1✓47.413.676.921.27.770.357.215.675.2
0.41.1✓42.412.476.915.16.367.455.214.475.3
0.51.1✓24.66.874.75.42.961.551.613.275.8
0.61.1✓7.84.862.6---37.210.676.7
0.71.1✓------14.46.164.9
", + "bbox": [ + 236, + 198, + 767, + 354 + ], + "page_idx": 12 + }, + { + "type": "text", + "text": "attention layer for intervention. The second stage unfolds during the decoding process, where we mitigate text inertia by subtracting the logits distribution that results from inputs devoid of image information. In this stage, the scale is managed by the parameter $\\gamma$ .", + "bbox": [ + 212, + 383, + 782, + 444 + ], + "page_idx": 12 + }, + { + "type": "text", + "text": "We use LLaVA-1.5 as the representative LVLM baseline and the greedy decoding method as the basic baseline to compare the impact of our hyperparameters on the task of long sequence image description. To evaluate our method, we choose the CHAIR metric. However, since CHAIR only evaluates the hallucination problem, we have incorporated the F1 score to consider information richness and accuracy. This makes the comparison with the CHAIR metric fairer when the F1 scores are similar. In the above, the F1 scores of the various methods differ slightly, so we have not included this somewhat redundant measure. The F1 score is calculated by counting the objects included in the description, the objects in the ground-truth sets, and the hallucinated objects.", + "bbox": [ + 212, + 444, + 787, + 595 + ], + "page_idx": 12 + }, + { + "type": "text", + "text": "Effects of $\\alpha$ in Exciting Image Attention. In the process of exciting the attention values of image tokens, we introduce a parameter $\\alpha$ to control the amplification scale. As shown in Tab. 4, different LVLMs exhibit varying sensitivity to the amplification scale. This sensitivity not only depends on the length of the model's image tokens (e.g., the image token length of LLaVA-1.5 is 576, while that of Minigpt4 is only 32), but also the original attention weights distribution.", + "bbox": [ + 212, + 595, + 787, + 686 + ], + "page_idx": 12 + }, + { + "type": "text", + "text": "However, a commonality across these models is that an appropriate amplification scale can achieve a balance between the number of hallucinated objects in the description and the amount of information conveyed. If the scale is too small, the description may still contain many hallucinated objects. Conversely, if the scale is too large, the amount of information in the response will decrease.", + "bbox": [ + 212, + 688, + 787, + 762 + ], + "page_idx": 12 + }, + { + "type": "text", + "text": "Effects of Layer Prior in Exciting Image Attention. We further investigate the control of the intervention layer for exciting image attention. As observed in Tab. 5, the introduction of this prior does indeed improve the performance of our method. However, when there is no control over the intervention layer and interventions are applied to all layers, the performances of different models ex", + "bbox": [ + 212, + 763, + 787, + 839 + ], + "page_idx": 12 + }, + { + "type": "header", + "text": "PAI", + "bbox": [ + 702, + 114, + 730, + 126 + ], + "page_idx": 12 + }, + { + "type": "page_number", + "text": "13", + "bbox": [ + 767, + 114, + 785, + 126 + ], + "page_idx": 12 + }, + { + "type": "table", + "img_path": "images/6ad52be2daf7b8fe34bc0471f103a2f4fc0b20d3841f4648a166ba8226b4a13d.jpg", + "table_caption": [ + "Table 5: Ablation Study of Hyperparameter $\\gamma$ and Layer Prior \"L\". Results are presented for the models LLAVA (left), Minigpt4 (middle), and Shikra (right)." + ], + "table_footnote": [], + "table_body": "
αγL CHAIRSCHAIRIF1αγL CHAIRSCHAIRIF1αγL CHAIRSCHAIRIF1
--46.213.875.9--31.610.569.4--56.215.874.6
0.51.1✓24.66.875.70.21.1✓21.27.770.30.61.1✓37.2
0.51.2✓23.46.675.70.21.2✓16.46.270.30.61.2✓40.2
0.51.3✓24.67.574.40.21.3✓14.46.969.50.61.3✓40.6
0.51.5✓25.09.674.60.21.5✓12.65.467.10.61.5✓40.2
0.52.0✓24.27.174.00.22.0✓8.44.664.30.62.0✓39.4
0.51.1✓24.66.875.70.21.1✓21.27.770.30.61.1✓37.2
0.51.1×20.66.771.70.21.1×22.38.470.30.61.1×56.4
", + "bbox": [ + 220, + 184, + 781, + 313 + ], + "page_idx": 13 + }, + { + "type": "text", + "text": "hibit some variations. For Shikra, in the absence of the layer prior, our method's results regress to the baseline. For both LLaVA and Minigpt4, the loss of the layer prior causes some unfavorable fluctuations in both the CHAIR metric and the F1 score.", + "bbox": [ + 215, + 345, + 784, + 406 + ], + "page_idx": 13 + }, + { + "type": "text", + "text": "Effects of $\\gamma$ in Mitigating Language Prior. Tab. 5 presents the results of an ablation study focusing on $\\gamma$ , which adjusts the balance between output distributions from conditioned inputs with excited image tokens and pure text inputs. Unlike the other two models, the Minigpt4 model is highly sensitive to $\\gamma$ . When $\\gamma$ is too large, it can lead to uncontrollable model behavior. The experimental results suggest that maintaining $\\gamma$ within a relatively small range, such as 1.1 - 1.2, yields the most stable performance.", + "bbox": [ + 215, + 409, + 785, + 515 + ], + "page_idx": 13 + }, + { + "type": "text", + "text": "6 Conclusion and Limitation", + "text_level": 1, + "bbox": [ + 217, + 550, + 508, + 566 + ], + "page_idx": 13 + }, + { + "type": "text", + "text": "In this paper, we first analyze the causes and manifestations of hallucinations in LVLMs. We propose a phenomenon termed \"text inertia,\" wherein the model continues to produce the same hallucinatory descriptions even when no image input is provided. This issue fundamentally stems from the model's neglect of image tokens. Consequently, we introduce the PAI method to intervene in the model's inference process, steering it towards an image-based and trustworthy direction. This is a training-free method and does not require any external tools. Extensive experiments on multiple benchmarks and LVLMs have validated the effectiveness of PAI in mitigating hallucination issues.", + "bbox": [ + 215, + 595, + 785, + 731 + ], + "page_idx": 13 + }, + { + "type": "text", + "text": "Limitations: (1) The language decoders of existing open-source LVLMs are primarily models from the LLaMA-family. It is worth exploring whether the issues of image neglect and text inertia are introduced by LLaMA. (2) As we described, our method fundamentally alleviates the image neglect issue during model inference. Its upper limit depends on the capabilities of the well-trained model. Therefore, it is worth investigating whether incorporating this issue as a loss during the training process could lead to further performance improvements.", + "bbox": [ + 215, + 733, + 785, + 839 + ], + "page_idx": 13 + }, + { + "type": "page_number", + "text": "14", + "bbox": [ + 217, + 114, + 235, + 126 + ], + "page_idx": 13 + }, + { + "type": "header", + "text": "Liu et al.", + "bbox": [ + 271, + 114, + 334, + 126 + ], + "page_idx": 13 + }, + { + "type": "text", + "text": "Acknowledgements", + "text_level": 1, + "bbox": [ + 217, + 143, + 401, + 162 + ], + "page_idx": 14 + }, + { + "type": "text", + "text": "This work is supported by the National Science Foundation of China (62132017), Zhejiang Provincial Natural Science Foundation of China (LD24F020011).", + "bbox": [ + 215, + 176, + 784, + 205 + ], + "page_idx": 14 + }, + { + "type": "text", + "text": "References", + "text_level": 1, + "bbox": [ + 217, + 229, + 321, + 244 + ], + "page_idx": 14 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "1. Alayrac, J.B., Donahue, J., Luc, P., Miech, A., Barr, I., Hasson, Y., Lenc, K., Mensch, A., Millican, K., Reynolds, M., Ring, R., Rutherford, E., Cabi, S., Han, T., Gong, Z., Samangooei, S., Monteiro, M., Menick, J., Borgeaud, S., Brock, A., Nematzadeh, A., Sharifzadeh, S., Binkowski, M., Barreira, R., Vinyals, O., Zisserman, A., Simonyan, K.: Flamingo: a visual language model for few-shot learning", + "2. Anisha, G., Yin, J., Bas, E.: Detecting and preventing hallucinations in large vision language models (Aug 2023)", + "3. Azaria, A., Mitchell, T.: The internal state of an llm knows when its lying (Apr 2023)", + "4. Brown, T., Mann, B., Ryder, N., Subbiah, M., Kaplan, J., Dhariwal, P., Neelakanthan, A., Shyam, P., Sastry, G., Amanda, A., Agarwal, S., Herbert-Voss, A., Krueger, G., Tom, H., Child, R., Ramesh, A., Ziegler, D., Wu, J., Winter, C., Hesse, C., Chen, M., Sigler, E., Litwin, M., Gray, S., Benjamin, C., Clark, J., Berner, C., Sam, M., Radford, A., Sutskever, I., Amodei, D.: Language models are few-shot learners. arXiv: Computation and Language, arXiv: Computation and Language (May 2020)", + "5. Burns, C., Ye, H., Klein, D., Steinhardt, J.: Discovering latent knowledge in language models without supervision (Dec 2022)", + "6. Chen, K., Zhang, Z., Zeng, W., Zhang, R., Zhu, F., Zhao, R., Research, S., Sklsde, S., Beihang, B., Seiee, S.: Shikra: Unleashing multimodal llm's referential dialogue magic", + "7. Chen, Z., Wu, J., Wang, W., Su, W., Chen, G., Xing, S., Muyan, Z., Zhang, Q., Zhu, X., Lu, L., Li, B., Luo, P., Lu, T., Qiao, Y., Dai, J.: Intervl: Scaling up vision foundation models and aligning for generic visual-linguistic tasks (Dec 2023)", + "8. Dai, W., Li, J., Li, D., Huat, A., Zhao, J., Wang, W., Li, B., Fung, P., Hoi, S.: Instructclip: Towards general-purpose vision-language models with instruction tuning", + "9. Darcet, T., Oquab, M., Mairal, J., Bojanowski, P.: Vision transformers need registers. arXiv preprint arXiv:2309.16588 (2023)", + "10. Devlin, J., Chang, M.W., Lee, K., Toutanova, K.: Bert: Pre-training of deep bidirectional transformers for language understanding. In: Proceedings of the 2019 Conference of the North (Jan 2019). https://doi.org/10.18653/v1/n19-1423, http://dx.doi.org/10.18653/v1/n19-1423", + "11. Han, Y., Zhang, C., Chen, X., Yang, X., Wang, Z., Yu, G., Fu, B., Zhang, H.: Chartlama: A multimodal llm for chart understanding and generation (Nov 2023)", + "12. Hu, H., Zhang, J., Zhao, M., Sun, Z.: Ciem: Contrastive instruction evaluation method for better instruction tuning", + "13. Huang, Q., Dong, X., Zhang, P., Wang, B., He, C., Wang, J., Lin, D., Zhang, W., Yu, N.: Opera: Alleviating hallucination in multi-modal large language models via over-trust penalty and retrospection-allocation (Nov 2023)", + "14. Kai, J., Zhang, T., Hu, H., Lin, Z.: Sh2: Self-highlighted hesitation helps you decode more truthfully (Jan 2024)" + ], + "bbox": [ + 218, + 258, + 785, + 839 + ], + "page_idx": 14 + }, + { + "type": "header", + "text": "PAI", + "bbox": [ + 702, + 114, + 730, + 126 + ], + "page_idx": 14 + }, + { + "type": "page_number", + "text": "15", + "bbox": [ + 767, + 114, + 785, + 126 + ], + "page_idx": 14 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "15. Lai, X., Tian, Z., Chen, Y., Li, Y., Yuan, Y., Liu, S., Jia, J., Kong, H., Research, M.: Lisa: Reasoning segmentation via large language model", + "16. Leng, S., Zhang, H., Chen, G., Li, X., Lu, S., Miao, C., Bing, L.: Mitigating object hallucinations in large vision-language models through visual contrastive decoding", + "17. Li, J., Li, D., Savarese, S., Hoi, S.: Blip-2: Bootstrapping language-image pretraining with frozen image encoders and large language models", + "18. Li, K., Patel, O., Viégas, F., Pfister, H., Wattenberg, M.: Inference-time intervention: Eliciting truthful answers from a language model. Advances in Neural Information Processing Systems 36 (2024)", + "19. Li, Y., Du, Y., Zhou, K., Wang, J., Zhao, W.X., Wen, J.R.: Evaluating object hallucination in large vision-language models. arXiv preprint arXiv:2305.10355 (2023)", + "20. Li, Z., Yang, B., Liu, Q., Ma, Z., Zhang, S., Yang, J., Sun, Y., Liu, Y., Bai, X.: Monkey: Image resolution and text label are important things for large multi-modal models (Nov 2023)", + "21. Liu, F., Lin, K., Li, L., Wang, J., Yacoob, Y., Wang, L.: Mitigating hallucination in large multi-modal models via robust instruction tuning. In: The Twelfth International Conference on Learning Representations (2023)", + "22. Liu, F., Lin, K., Li, L., Wang, J., Yacoob, Y., Wang, L.: Mitigating hallucination in large multi-modal models via robust instruction tuning (Sep 2023)", + "23. Liu, H., Xue, W., Chen, Y., Chen, D., Zhao, X., Wang, K., Hou, L., Li, R., Peng, W.: A survey on hallucination in large vision-language models", + "24. Liu, H., Li, C., Li, Y., Lee, Y.J.: Improved baselines with visual instruction tuning. arXiv preprint arXiv:2310.03744 (2023)", + "25. Liu, H., Li, C., Wu, Q., Lee, Y., Madison, M., Research, M.: Visual instruction tuning", + "26. Mishra, S., Khashabi, D., Baral, C., Hajishirzi, H.: Cross-task generalization via natural language crowdsourcing instructions. In: Proceedings of the 60th Annual Meeting of the Association for Computational Linguistics (Volume 1: Long Papers) (Jan 2022). https://doi.org/10.18653/v1/2022.acl-long.244, http://dx.doi.org/10.18653/v1/2022.acl-long.244", + "27. Ouyang, L., Wu, J., Jiang, X., Almeida, D., Wainwright, C.L., Mishkin, P., Zhang, C., Agarwal, S., Slama, K., Ray, A., Schulman, J., Hilton, J., Kelton, F., Miller, L., Simens, M., Askell, A., Welinder, P., Christiano, P.F., Leike, J., Lowe, R.: Training language models to follow instructions with human feedback (2022)", + "28. Peng, B., Li, C., He, P., Galley, M., Gao, J.: Instruction tuning with gpt-4", + "29. Rohrbach, A., Hendricks, L.A., Burns, K., Darrell, T., Saenko, K.: Object hallucination in image captioning. In: Proceedings of the 2018 Conference on Empirical Methods in Natural Language Processing (Jan 2018). https://doi.org/10.18653/v1/d18-1437, http://dx.doi.org/10.18653/v1/d18-1437", + "30. Sanchez, G., Fan, H., Spangher, A., Levi, E., Ammanamanchi, P.S., Biderman, S.: Stay on topic with classifier-free guidance. arXiv preprint arXiv:2306.17806 (2023)", + "31. Sun, Z., Shen, S., Cao, S., Liu, H., Li, C., Shen, Y., Gan, C., Gui, L.Y., Wang, Y.X., Yang, Y., Keutzer, K., Darrell, T.: Aligning large multimodal models with factually augmented rlhf (Sep 2023)", + "32. Sun, Z., Shen, S., Cao, S., Liu, H., Li, C., Shen, Y., Gan, C., Gui, L.Y., Wang, Y.X., Yang, Y., et al.: Aligning large multimodal models with factually augmented rlhf. arXiv preprint arXiv:2309.14525 (2023)", + "33. Touvron, H., Lavril, T., Izacard, G., Martinet, X., Lachaux, M.A., Lacroix, T., Rozi'ere, B., Goyal, N., Hambro, E., Azhar, F., Rodriguez, A., Joulin, A., Grave, E., Lample, G.: Llama: Open and efficient foundation language models" + ], + "bbox": [ + 215, + 147, + 784, + 839 + ], + "page_idx": 15 + }, + { + "type": "page_number", + "text": "16", + "bbox": [ + 217, + 114, + 235, + 126 + ], + "page_idx": 15 + }, + { + "type": "header", + "text": "Liu et al.", + "bbox": [ + 271, + 114, + 334, + 126 + ], + "page_idx": 15 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "34. Turner, A., Thiergart, L., Udell, D., Leech, G., Mini, U., MacDiarmid, M.: Activation addition: Steering language models without optimization. arXiv preprint arXiv:2308.10248 (2023)", + "35. Wang, L., Li, L., Dai, D., Chen, D., Zhou, H., Meng, F., Zhou, J., Sun, X.: Label words are anchors: An information flow perspective for understanding in-context learning. arXiv preprint arXiv:2305.14160 (2023)", + "36. Wei, F., Zhang, X., Zhang, A., Zhang, B., Chu, X.: Lenna: Language enhanced reasoning detection assistant (Dec 2023)", + "37. Wei, J., Bosma, M., Zhao, V., Guu, K., Yu, A., Lester, B., Du, N., Dai, A., Le, Q.: Finetuned language models are zero-shot learners. Learning, Learning (Sep 2021)", + "38. Xiao, G., Tian, Y., Chen, B., Han, S., Lewis, M.: Efficient streaming language models with attention sinks (Sep 2023)", + "39. Ye, Q., Xu, H., Xu, G., Ye, J., Yan, M., Zhou, Y., Wang, J., Hu, A., Shi, P., Shi, Y., Li, C., Xu, Y., Chen, H., Tian, J., Qi, Q., Zhang, J., Huang, F.: mplug-owl: Modularization empowers large language models with multimodality", + "40. Yin, S., Fu, C., Zhao, S., Xu, T., Wang, H., Sui, D., Shen, Y., Li, K., Sun, X., Chen, E.: Woodpecker: Hallucination correction for multimodal large language models", + "41. You, H., Zhang, H., Gan, Z., Du, X., Zhang, B., Wang, Z., Cao, L., Chang, S.F., Yang, Y.: Ferret: Refer and ground anything anywhere at any granularity (Oct 2023)", + "42. Zhang, Y., Qian, S., Peng, B., Liu, S., Jia, J.: Prompt highlighter: Interactive control for multi-modal llms. arXiv preprint arXiv:2312.04302 (2023)", + "43. Zhou, Y., Cui, C., Yoon, J., Zhang, L., Deng, Z., Finn, C., Bansal, M., Yao, H.: Analyzing and mitigating object hallucination in large vision-language models (Oct 2023)", + "44. Zhu, D., Chen, J., Shen, X., Li, X., Elhoseiny, M.: Minigpt-4: Enhancing vision-language understanding with advanced large language models" + ], + "bbox": [ + 212, + 146, + 787, + 521 + ], + "page_idx": 16 + }, + { + "type": "header", + "text": "PAI", + "bbox": [ + 702, + 114, + 730, + 126 + ], + "page_idx": 16 + }, + { + "type": "page_number", + "text": "17", + "bbox": [ + 767, + 114, + 785, + 126 + ], + "page_idx": 16 + } +] \ No newline at end of file diff --git a/2024/Paying More Attention to Images_ A Training-Free Method for Alleviating Hallucination in LVLMs/71543d84-0242-4633-b85d-0ad4feeefd05_model.json b/2024/Paying More Attention to Images_ A Training-Free Method for Alleviating Hallucination in LVLMs/71543d84-0242-4633-b85d-0ad4feeefd05_model.json new file mode 100644 index 0000000000000000000000000000000000000000..ee530a0ca31dd4d3b2c72a12341294718ad0d211 --- /dev/null +++ b/2024/Paying More Attention to Images_ A Training-Free Method for Alleviating Hallucination in LVLMs/71543d84-0242-4633-b85d-0ad4feeefd05_model.json @@ -0,0 +1,2247 @@ +[ + [ + { + "type": "title", + "bbox": [ + 0.218, + 0.141, + 0.787, + 0.187 + ], + "angle": 0, + "content": "Paying More Attention to Image: A Training-Free Method for Alleviating Hallucination in LVLMs" + }, + { + "type": "text", + "bbox": [ + 0.334, + 0.212, + 0.668, + 0.228 + ], + "angle": 0, + "content": "Shi Liu\\(^{1}\\), Kecheng Zheng\\(^{1,2\\dagger}\\), and Wei Chen\\(^{1\\dagger}\\)" + }, + { + "type": "text", + "bbox": [ + 0.33, + 0.239, + 0.672, + 0.254 + ], + "angle": 0, + "content": "\\(^{1}\\) State Key Lab of CAD&CG, Zhejiang University" + }, + { + "type": "text", + "bbox": [ + 0.456, + 0.254, + 0.546, + 0.268 + ], + "angle": 0, + "content": "2 Ant Group" + }, + { + "type": "text", + "bbox": [ + 0.374, + 0.269, + 0.631, + 0.282 + ], + "angle": 0, + "content": "{liushi0927,zkechengzk}@gmail.com" + }, + { + "type": "text", + "bbox": [ + 0.432, + 0.284, + 0.573, + 0.296 + ], + "angle": 0, + "content": "chenvis@zju.edu.cn" + }, + { + "type": "text", + "bbox": [ + 0.261, + 0.327, + 0.744, + 0.62 + ], + "angle": 0, + "content": "Abstract. Existing Large Vision-Language Models (LVLMs) primarily align image features of vision encoder with Large Language Models (LLMs) to leverage their superior text generation capabilities. However, the scale disparity between vision encoder and language model may led to LLMs assuming a predominant role in multi-modal comprehension. This imbalance in LVLMs may result in the instances of hallucinatory. Concretely, LVLMs may generate consistent descriptions with or without visual input, indicating that certain outputs are influenced solely by context text. We refer to this phenomenon as \"text inertia.\" To counteract this issue, we introduce a training-free algorithm to find an equilibrium point between image comprehension and language inference. Specifically, we adaptively involve adjusting and amplifying the attention weights assigned to image tokens, thereby granting greater prominence to visual elements. Meanwhile, we subtract the logits of multi-modal inputs from ones of pure text input, which can help LVLMs be not biased towards LLMs. By enhancing images tokens and reducing the stubborn output of LLM, we can let LVLM pay more attention to images, towards alleviating text inertia and reducing the hallucination in LVLMs. Our extensive experiments show that this method substantially reduces the frequency of hallucinatory outputs in various LVLMs in terms of different metrics. Project page is available at https://1albj.github.io/projects/PAI/" + }, + { + "type": "text", + "bbox": [ + 0.261, + 0.631, + 0.741, + 0.659 + ], + "angle": 0, + "content": "Keywords: Vision-Language Models \\(\\cdot\\) Visual Dialogs \\(\\cdot\\) Hallucination Mitigation" + }, + { + "type": "title", + "bbox": [ + 0.216, + 0.683, + 0.377, + 0.699 + ], + "angle": 0, + "content": "1 Introduction" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.712, + 0.788, + 0.788 + ], + "angle": 0, + "content": "Recently, Large Vision-Language Models (LVLMs) have made significant strides, exhibiting impressive capabilities across a multitude of tasks [11,15,36,42]. However, these models still struggle with the phenomenon of hallucination. Specifically, there is often a mismatch between the textual content generated by the model and the actual visual input it receives [23]." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.788, + 0.787, + 0.819 + ], + "angle": 0, + "content": "Hallucination in LVLMs is often attributed to issues with modality alignment, leading to the development of mitigating strategies through alignment" + }, + { + "type": "page_footnote", + "bbox": [ + 0.232, + 0.825, + 0.405, + 0.841 + ], + "angle": 0, + "content": "† Corresponding authors." + } + ], + [ + { + "type": "page_number", + "bbox": [ + 0.218, + 0.116, + 0.23, + 0.127 + ], + "angle": 0, + "content": "2" + }, + { + "type": "header", + "bbox": [ + 0.272, + 0.115, + 0.335, + 0.127 + ], + "angle": 0, + "content": "Liu et al." + }, + { + "type": "image", + "bbox": [ + 0.226, + 0.147, + 0.782, + 0.309 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.214, + 0.32, + 0.788, + 0.432 + ], + "angle": 0, + "content": "Fig. 1: We present an examination of various input settings, with hallucinations specifically highlighted in red. (a) When using LLAVA for image description, it generates a hallucinated description. (b) Even without image input, when only the historical response preceding the hallucinated description is input to LLAVA, it reproduces the same hallucinated description, a phenomenon we refer to as \"Text Inertia\". (c) Our proposed method, PAI, effectively mitigates this text inertia problem and yields accurate descriptions. (d) Utilizing PAI for image description results in a significantly more precise description." + }, + { + "type": "image", + "bbox": [ + 0.22, + 0.459, + 0.491, + 0.585 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.214, + 0.6, + 0.499, + 0.669 + ], + "angle": 0, + "content": "Fig.2: Percentage of text inertia hallucination in all hallucination (calculated with 500 samples). For specific calculation processes, please refer to the supplementary material section A." + }, + { + "type": "image", + "bbox": [ + 0.509, + 0.449, + 0.784, + 0.533 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.503, + 0.545, + 0.788, + 0.669 + ], + "angle": 0, + "content": "Fig. 3: Visualization of the average attention ratio of different content. The x-axis denotes the sequence length of the history tokens. The lengths of the image, instruction, BOS, and outlier tokens [9] are all fixed as they are part of the model input, with 576 tokens for the image, 21 for the instruction, and 1 each for the BOS and outlier tokens." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.703, + 0.788, + 0.794 + ], + "angle": 0, + "content": "training optimization [7, 20, 31]. However, is hallucination in LVLMs merely a result of the model's capacity, and can it only be alleviated through additional training? We propose a scenario where LVLMs generate a hallucinated object description. Specifically, even when the image input is removed and only the generated text preceding the hallucinated object word is retained, the LVLMs persist in producing the same hallucinated description, as depicted in Fig. 1." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.796, + 0.788, + 0.84 + ], + "angle": 0, + "content": "In order to empirically investigate this behavior, we conducted tests on three LVLMs within the context of image describing tasks on the COCO dataset. We identified and conducted a statistical analysis on instances where LVLMs" + } + ], + [ + { + "type": "header", + "bbox": [ + 0.703, + 0.115, + 0.732, + 0.127 + ], + "angle": 0, + "content": "PAI" + }, + { + "type": "page_number", + "bbox": [ + 0.775, + 0.116, + 0.787, + 0.127 + ], + "angle": 0, + "content": "3" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.147, + 0.784, + 0.207 + ], + "angle": 0, + "content": "generated identical hallucinated object descriptions, even when the input was exclusively historical response text without any image. The observation from Fig. 2 clearly indicates that, even with the application of rigorous identification settings, the phenomenon continues to represent a substantial proportion." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.208, + 0.785, + 0.373 + ], + "angle": 0, + "content": "We referred this phenomenon as \"Text Inertia\". Our hypothesis is that text inertia arises due to the current generative paradigms mapping image representations onto the text representation space as text tokens. In this mechanism, the LLM becomes the dominant character, and the inference process lacks additional handling of image tokens, leading to their neglect during the generation process. To validate this hypothesis, we have analyzed the attention values ratios of the LLaVA model during the inference process in Fig. 3. Our findings show that despite image tokens occupying a significant proportion, they do not receive substantial attention under the current mechanism. This multimodal chat resembles more of an automatic completion based on context rather than a continuous attention to the image for completion." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.374, + 0.787, + 0.599 + ], + "angle": 0, + "content": "To close this gap, we introduce a method referred to as Pay Attention to Image (PAI). At a high level, PAI intervenes in the inference process to make it more image-centric, following the original image perception direction. To achieve this, we focus on the self-attention heads in the decoder layers of LVLMs. We enhance the attention weights for image tokens in their original directions during inference. This allows us to use the updated attention matrix to calculate the hidden states for the generated token, thereby incorporating more consideration for image representation during the generation process. To further mitigate text inertia, we construct the input using instruction tokens and historical response tokens, and subtract the model logits of this input from the logits of original model with image tokens input. This strategy helps to reduce the influence of language priors during the generation process. Unlike previous methods for mitigating hallucination that require additional training or external tools, our approach is training-free. Moreover, we are the first to propose an inference intervention method for mitigating hallucination in LVLMs." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.6, + 0.787, + 0.735 + ], + "angle": 0, + "content": "We then evaluate the response accuracy in the image description task from long sequence generation perspective by employing the CHAIR metric [29] and GPT-4V. In addition, we use POPE [19] and MMHal-Bench [32] to more comprehensively evaluate the model's hallucination performance on VQA task. Furthermore, we have constructed single-turn and multi-turn chat evaluations for POPE. Since our model intervenes in the inference process, it can be used for any decoding method. Therefore, we conducted experiments on three decoding methods of the three models. The experimental results proved the effectiveness of our method in mitigating hallucinations." + }, + { + "type": "title", + "bbox": [ + 0.216, + 0.757, + 0.388, + 0.772 + ], + "angle": 0, + "content": "2 Related Work" + }, + { + "type": "title", + "bbox": [ + 0.216, + 0.787, + 0.517, + 0.802 + ], + "angle": 0, + "content": "2.1 Large Vision-Language Models" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.81, + 0.784, + 0.84 + ], + "angle": 0, + "content": "The development of pre-training techniques [4, 27] and instruction tuning techniques [26, 37] has rapidly advanced LLMs technology, such as LLaMA [33] and" + } + ], + [ + { + "type": "page_number", + "bbox": [ + 0.218, + 0.116, + 0.23, + 0.127 + ], + "angle": 0, + "content": "4" + }, + { + "type": "header", + "bbox": [ + 0.272, + 0.115, + 0.335, + 0.127 + ], + "angle": 0, + "content": "Liu et al." + }, + { + "type": "image", + "bbox": [ + 0.23, + 0.157, + 0.784, + 0.284 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.214, + 0.318, + 0.788, + 0.388 + ], + "angle": 0, + "content": "Fig. 4: Architecture of our PAI. To alleviate text inertia, we additionally construct an input without image. Throughout the forward inference process, we amplify the focus on the image token by edit self-attention maps in LLaMA. Ultimately, we subtract the logits distribution of the language prior during decoding to achieve an accurate description." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.417, + 0.789, + 0.569 + ], + "angle": 0, + "content": "Vicuna [28], further leading to the prosperity of LVLMs technology. Early works, like Flamingo [1] and BLIP-2 [17], have successfully adapted LLMs to visual tasks, demonstrating notable generative capabilities and in-context learning abilities. Recently, the capabilities of LVLMs have further advanced under the influence of visual instruction tuning techniques [24,25]. Using different projectors to map images to the text domain, thereby endowing language generation models with image understanding capabilities, is also a hot research topic [8,25,39,44]. Additionally, several studies focus on visual language tasks such as grounding capabilities [6] and reasoning capabilities [15]. However, recent LVLMs still face the issue of hallucination generation [21]." + }, + { + "type": "title", + "bbox": [ + 0.215, + 0.59, + 0.554, + 0.605 + ], + "angle": 0, + "content": "2.2 Mitigation of LVLMs Hallucination" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.614, + 0.788, + 0.748 + ], + "angle": 0, + "content": "Hallucination in LVLMs refers to contradictions between the image input and the textual output. Various methods have been proposed to mitigate hallucination. The most direct reason for hallucination generation is that hallucination arises from data bias and knowledge gaps between vision and language. Therefore, better data filtering methods [12, 22, 41] and higher quality annotated data [2] are introduced. Simultaneously, these methods also imply the need for more alignment training [31] or adjustments in the model architecture [7, 20]. These methods can achieve good results, but they are time-consuming and require high computational resources." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.75, + 0.788, + 0.841 + ], + "angle": 0, + "content": "Apart from addressing the ability of LVLMs itself, hallucination can also be mitigated through post-processing methods. This method usually involves using an additional module or external tools to edit the response. Recent methods such as LURE [43] utilize additional data to train a state detector and when hallucination issues are detected, content is regenerated by a revisor model. Woodpecker [40] introduces an external visual model to inspect entities extracted from" + } + ], + [ + { + "type": "header", + "bbox": [ + 0.703, + 0.115, + 0.732, + 0.127 + ], + "angle": 0, + "content": "PAI" + }, + { + "type": "page_number", + "bbox": [ + 0.776, + 0.116, + 0.786, + 0.127 + ], + "angle": 0, + "content": "5" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.147, + 0.784, + 0.191 + ], + "angle": 0, + "content": "the response, and then the detection results are handed over to the generation model to regenerate better answers. These methods also extend the inference chain and increase inference costs." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.192, + 0.788, + 0.314 + ], + "angle": 0, + "content": "Training-free hallucination mitigation methods have so far only been attempted in decoding methods. OPERA [13] discovered an abnormal attention pattern that accompanies model decoding. It was statistically found that this pattern often accompanies hallucination descriptions, and thus a detection and mitigation method was proposed based on this pattern to alleviate the hallucination faced by the model. VCD [16] introduced the notion that visual uncertainty increases hallucination descriptions and, based on this discovery, proposed a contrast decoding method to alleviate hallucination issues." + }, + { + "type": "title", + "bbox": [ + 0.216, + 0.336, + 0.383, + 0.352 + ], + "angle": 0, + "content": "3 Preliminaries" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.367, + 0.788, + 0.488 + ], + "angle": 0, + "content": "The architecture of LVLMs typically comprises of three main components: an image encoder, a projector, and a language decoder. Both the image encoder and language decoder are usually pre-trained. The image encoder is employed to transform images into image tokens, which are subsequently mapped to the text representation space by the projector. This process enables the concatenation of image tokens with text tokens that are then fed into the language decoder. The language decoder subsequently generates corresponding responses based on the provided instructions." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.489, + 0.788, + 0.655 + ], + "angle": 0, + "content": "The existing projectors. Currently, projectors predominantly fall into two categories: linear projectors and resamplers. A projector takes N visual features from the image encoder and transforms them into M visual tokens. The linear projector employs a multilayer perceptron to transform all visual features, maintaining a one-to-one transformation which meaning that M equals N. In contrast, the resampler does not preserve all visual features but instead samples visual cues (M, where \\(\\mathrm{M} < \\mathrm{N}\\)). For instance, Q-former [17] utilizes M learnable queries and Bert [10] to extract information from visual features. Given that the knowledge of images during the generation process solely originates from the output image tokens of the projectors, our attention is concentrated on the image tokens post-projection, irrespective of their preceding modeling process." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.655, + 0.788, + 0.747 + ], + "angle": 0, + "content": "Autoregressive language decoders. Nearly all LVLMs adopt LLaMA-family models as their language decoders, which employ the self-attention mechanism. The visual tokens processed by the projector are concatenated with text tokens and fed into the LLaMA, which carries out the forward decoding process. From the perspective of a single attention head in a single layer, each head repeatedly performs the following attention operation with the same input shape:" + }, + { + "type": "equation", + "bbox": [ + 0.353, + 0.757, + 0.787, + 0.791 + ], + "angle": 0, + "content": "\\[\n\\boldsymbol {O} _ {h} = \\boldsymbol {A} _ {h} \\boldsymbol {V} _ {h}, \\quad \\boldsymbol {A} _ {h} = \\operatorname {s o f t m a x} \\left(\\frac {\\boldsymbol {Q} _ {h} \\boldsymbol {K} _ {h} ^ {\\top}}{\\sqrt {d _ {k}}}\\right). \\tag {1}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.795, + 0.788, + 0.842 + ], + "angle": 0, + "content": "Each attention head \\( h \\) performs an attention operation using its own set of queries \\( Q_{h} \\in \\mathbb{R}^{n \\times d_{k}} \\), keys \\( K_{h} \\in \\mathbb{R}^{n \\times d_{k}} \\), and values \\( V_{h} \\in \\mathbb{R}^{n \\times d_{k}} \\), where \\( n \\) represents the sequence length and \\( d_{k} \\) represents the hidden dimensions. The" + } + ], + [ + { + "type": "page_number", + "bbox": [ + 0.218, + 0.116, + 0.23, + 0.127 + ], + "angle": 0, + "content": "6" + }, + { + "type": "header", + "bbox": [ + 0.272, + 0.115, + 0.335, + 0.128 + ], + "angle": 0, + "content": "Liu et al." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.147, + 0.788, + 0.283 + ], + "angle": 0, + "content": "output \\( O_{h} \\in \\mathbb{R}^{n \\times d_{k}} \\) is modeled by multiplying \\( V_{h} \\) and the attention weights \\( A_{h} \\in \\mathbb{R}^{n \\times n} \\), where each row represents the weights for each token during feature mixing. This operation enables the model to focus on different parts of the input for each head through attention weights, thereby capturing various parts of the information from the sequence token representations. The final output is the current generated token vocabulary conditional probability distribution \\( y \\in \\mathbb{R}^{v} \\) based on the input instruction representations \\( X_{I} \\), image representations \\( X_{V} \\) and history generated token representations \\( X_{H} \\), where \\( v \\) is the size of the vocabulary. This process can be formatted as:" + }, + { + "type": "equation", + "bbox": [ + 0.351, + 0.294, + 0.785, + 0.328 + ], + "angle": 0, + "content": "\\[\n\\begin{array}{l} \\boldsymbol {y} \\sim p _ {\\text {m o d e l}} (\\boldsymbol {y} \\mid \\mathbf {X} _ {I}, \\mathbf {X} _ {V}, \\mathbf {X} _ {H}), \\\\ \\begin{array}{l} p _ {\\text {m o d e r}} (\\boldsymbol {y} \\mid 1, 1, 1, 1, 1, 1, 1, 1, 1), \\\\ \\propto \\operatorname {s o f t m a x} \\left(\\operatorname {l o g i t} _ {\\text {m o d e l}} (\\boldsymbol {y} \\mid \\mathbf {X} _ {I}, \\mathbf {X} _ {V}, \\mathbf {X} _ {H})\\right), \\end{array} \\tag {2} \\\\ \\end{array}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.337, + 0.785, + 0.385 + ], + "angle": 0, + "content": "which calculates the distribution of one token and iterates for the entire response. The sequence generation continues until an EOS (End of Sentence) token is produced, marking the end of the generation and resulting in a complete response." + }, + { + "type": "title", + "bbox": [ + 0.215, + 0.395, + 0.331, + 0.41 + ], + "angle": 0, + "content": "4 Method" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.418, + 0.788, + 0.569 + ], + "angle": 0, + "content": "At the core of our method is a solution for image neglect and text inertia, both of which are fundamentally interconnected. Essentially, as paying more attention to the image, there is a corresponding reduction in the reliance on language priors. Intuitively, in a conversation centered around an image, the model should devote more attention to the image, thereby allowing it to have a significant impact on the response. As such, we identify the self-attention map in the token-level generation and augment the image attention in its original directions. This strategy promotes a more image-centric latent representations. Additionally, to further mitigate the influence of text inertia, we devide the logits distribution of pure text input into the model's output." + }, + { + "type": "title", + "bbox": [ + 0.215, + 0.59, + 0.509, + 0.606 + ], + "angle": 0, + "content": "4.1 Pay More Attention to Image" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.614, + 0.788, + 0.734 + ], + "angle": 0, + "content": "Extracting the Self-Attention Matrix. We start from a token-level perspective. The response process in LVLMs is fundamentally generated token by token. Each token is generated based on the input image, instruction, and the historically generated response. This process is facilitated through a multi-layer attention decoder architecture. Consequently, this results in a probability distribution of the vocabulary for the currently generated token. Our goal is to extract the attention matrix of each attention head at every layer, indicating the influence of each content during inference." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.735, + 0.788, + 0.843 + ], + "angle": 0, + "content": "When generating the \\(k\\)-th token in the sequence, the input representation for the attention head in the forward process includes the instruction representation \\(\\mathbf{X}_I = [\\pmb{x}_{i_1},\\dots,\\pmb{x}_{i_{n_I}}]\\), image representation \\(\\mathbf{X}_V = [\\pmb{x}_{v_1},\\dots,\\pmb{x}_{v_{n_V}}]\\), and the representation of the historically generated response \\(\\mathbf{X}_H = [\\pmb{x}_{h_1},\\dots,\\pmb{x}_{h_{n_H}}]\\). Notably, the image representation considered here is the one that has been processed by the projector. Essentially, the hidden states of each input layer are \\(\\mathbf{X} = \\mathrm{concat}(\\mathbf{X}_I[1:m],\\mathbf{X}_V,\\mathbf{X}_I[m + 1:n_I],\\mathbf{X}_H)\\), where the notation \\(\\mathbf{X}_I[1:m]\\)" + } + ], + [ + { + "type": "header", + "bbox": [ + 0.703, + 0.115, + 0.732, + 0.127 + ], + "angle": 0, + "content": "PAI" + }, + { + "type": "page_number", + "bbox": [ + 0.775, + 0.116, + 0.787, + 0.127 + ], + "angle": 0, + "content": "7" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.147, + 0.785, + 0.236 + ], + "angle": 0, + "content": "indicates the first \\( m \\) elements in the instruction representation. Each attention head assigns different degrees of attention to each element during the current token representation decoding process. Our aim is to enhance the attention paid to the image. Therefore, we extract the attention weight values related to the image token for the current generated token, intervene, and then redistribute the attention values of each element through softmax." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.239, + 0.787, + 0.389 + ], + "angle": 0, + "content": "Excite model in a trustful direction. There have been attempts in some LLMs works to make the answers generated by LLMs more trustworthy by means of intervention [3, 5, 14, 18, 34]. The implemented approach typically involves intervening with the hidden states. As for defining what constitutes a more trustworthy direction, it usually requires additional projection and training to probe this trustful direction. In our case, a response that is more image-based is considered more trustworthy. Since LVLMs have undergone alignment training, the original attention values provide a direction based on image content. As illustrated in Fig. 4, by amplifying the attention values of image tokens based on the original attention values, we can enhance the trustworthiness of our results." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.39, + 0.787, + 0.526 + ], + "angle": 0, + "content": "Another nuance involves our avoidance of choosing the attention head used to shift intervention. In the ITI method [18], it is stated that not all attention heads should be subjected to intervention. Therefore, they introduce a trustful score to rank each head across all layers and select the top-k heads for intervention. In our case, the less trustworthy heads with lower attention values receive less intervention. We first extract the attention weights of the image tokens for the current generated token from the attention weights \\(\\tilde{\\pmb{A}}\\) before softmax operation. We then use the hyper-parameter \\(\\alpha\\) to control the step size for intervention. From a single attention head perspective, our method can be expressed as follows:" + }, + { + "type": "equation", + "bbox": [ + 0.317, + 0.538, + 0.785, + 0.555 + ], + "angle": 0, + "content": "\\[\n\\tilde {\\boldsymbol {A}} _ {n, j} = \\tilde {\\boldsymbol {A}} _ {n, j} + \\alpha \\cdot | \\tilde {\\boldsymbol {A}} _ {n, j} | \\quad \\text {f o r} j = m + 1 \\text {t o} m + n _ {V}. \\tag {3}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.568, + 0.787, + 0.673 + ], + "angle": 0, + "content": "The model's final vocabulary probability distribution is derived from the projection of the hidden states of the last token in the sequence. Therefore, we extract the attention weights of the last token \\( n \\) on the image tokens by indexing \\( \\tilde{A}_{n,j} \\). Following the intervention, we use the softmax function to redistribute the attention values of each token during the reassignment of encoded hidden states. This procedure is repeated for each subsequent token prediction in an autoregressive manner and is independent of the choice of the decoding algorithm." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.674, + 0.787, + 0.809 + ], + "angle": 0, + "content": "Excite more precisely with an attention mode prior. The presence of the BOS token, an attention sink pattern [38], in a sentence results in higher attention values during the attention computation process, which may seem counterintuitive. The BOS token typically signifies the start of a sentence and, as such, does not carry significant semantic content. However, the generation of tokens is significantly influenced by this particular token, a similar pattern that also manifests itself in visual models [9]. As mentioned in StreamLLM [38], the pattern of the attention sink emerges when redundant attention values are present." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.81, + 0.787, + 0.84 + ], + "angle": 0, + "content": "Naturally, one might infer that when the sink pattern appears, we excite the image token. To further investigate this phenomenon, as depicted in Fig. 5, we" + } + ], + [ + { + "type": "page_number", + "bbox": [ + 0.218, + 0.116, + 0.23, + 0.127 + ], + "angle": 0, + "content": "8" + }, + { + "type": "header", + "bbox": [ + 0.272, + 0.115, + 0.335, + 0.127 + ], + "angle": 0, + "content": "Liu et al." + }, + { + "type": "image", + "bbox": [ + 0.241, + 0.146, + 0.408, + 0.256 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.291, + 0.26, + 0.359, + 0.273 + ], + "angle": 0, + "content": "(a) LLAVA" + }, + { + "type": "image", + "bbox": [ + 0.417, + 0.146, + 0.583, + 0.257 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.46, + 0.26, + 0.539, + 0.273 + ], + "angle": 0, + "content": "(b) Minigpt4" + }, + { + "type": "image", + "bbox": [ + 0.59, + 0.146, + 0.758, + 0.256 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.643, + 0.26, + 0.706, + 0.273 + ], + "angle": 0, + "content": "(c) Shikra" + }, + { + "type": "image_caption", + "bbox": [ + 0.215, + 0.277, + 0.785, + 0.305 + ], + "angle": 0, + "content": "Fig. 5: The BOS token attention ratios of three model. We calculate the attention weights of the BOS token for each head in each layer and display them in a heat map." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.324, + 0.788, + 0.4 + ], + "angle": 0, + "content": "find that the sink phenomenon is not overtly evident in the shallow layers. This is because the shallow layers tend to focus more on encoding semantically rich information [35]. When the encoding of semantically rich tokens stabilizes, the attention sink phenomenon arises. Therefore, we build upon the judgement of intervention timing by calculating the similarity of the hidden states." + }, + { + "type": "title", + "bbox": [ + 0.215, + 0.422, + 0.486, + 0.437 + ], + "angle": 0, + "content": "4.2 Image-Centric Logit Refine" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.447, + 0.788, + 0.523 + ], + "angle": 0, + "content": "In Fig. 1, we observe a peculiar phenomenon where LVLMs continue to generate identical hallucinated text even when the image is removed from the input. This observation naturally leads us to the concept of using the output distribution (when no image is in input) as a reference to penalize our initial prediction distribution. Therefore, we update the distribution of the generated token by:" + }, + { + "type": "equation", + "bbox": [ + 0.36, + 0.533, + 0.64, + 0.569 + ], + "angle": 0, + "content": "\\[\n\\begin{array}{l} p _ {\\text {m o d e l}} = \\gamma \\cdot p _ {\\text {m o d e l}} (\\boldsymbol {y} | \\boldsymbol {X} _ {V}, \\boldsymbol {X} _ {I}, \\boldsymbol {X} _ {H}) \\\\ - (\\gamma - 1) \\cdot p _ {\\text {m o d e l}} (\\boldsymbol {y} | \\boldsymbol {X} _ {I}, \\boldsymbol {X} _ {H}). \\\\ \\end{array}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.58, + 0.785, + 0.623 + ], + "angle": 0, + "content": "This equation effectively reduces the predicted probability based on text alone. The weight \\(\\gamma\\) is used to control the degree of penalty applied to the initial prediction distribution." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.624, + 0.788, + 0.714 + ], + "angle": 0, + "content": "This operation is conceptually similar to LLM-CFG [30]. Essentially, it provides a guided generation mechanism that allows the model to make informed choices between outputs based on image content and those based on language logic. This way, the model can better balance the influence of visual and textual information in its outputs, leading to more contextually accurate and relevant results." + }, + { + "type": "title", + "bbox": [ + 0.215, + 0.738, + 0.376, + 0.755 + ], + "angle": 0, + "content": "5 Experiments" + }, + { + "type": "title", + "bbox": [ + 0.215, + 0.77, + 0.312, + 0.785 + ], + "angle": 0, + "content": "5.1 Setup" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.795, + 0.787, + 0.84 + ], + "angle": 0, + "content": "Baselines. We evaluate the effectiveness of our method on three different models. To better compare the impact of image feature tokens after different projectors on our method, we selected two models that use linear projectors, LLAVA" + } + ], + [ + { + "type": "header", + "bbox": [ + 0.703, + 0.115, + 0.732, + 0.127 + ], + "angle": 0, + "content": "PAI" + }, + { + "type": "page_number", + "bbox": [ + 0.776, + 0.116, + 0.786, + 0.127 + ], + "angle": 0, + "content": "9" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.147, + 0.784, + 0.252 + ], + "angle": 0, + "content": "and Shikra, as well as one model that uses resamplers, Minigpt4. Additionally, for a more convincing comparison, we report on three decoding methods for comparison: greedy, beam search, and nucleus sample. We also selected the OPERA [13] method, which is an improvement on beam search, and the VCD [16] method, which is an improvement on nucleus sampling, to compare with our results. We used the default hyperparameters from the open-source versions of these two methods." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.253, + 0.785, + 0.372 + ], + "angle": 0, + "content": "Implementation Details. As different models have different lengths of image tokens, leading to different degrees of image neglect, to better align with the image sequence length of the model, we set \\(\\alpha = 0.5\\) for LLAVA, \\(\\alpha = 0.6\\) for Shikra with long image token sequence lengths and \\(\\alpha = 0.2\\) for resampler models with short image token sequences. As text inertia is independent of the image token length, we continuously use \\(\\gamma = 1.1\\). Apart from this, in the beam search tests, the beam number is set to 5 for all methods, and in the nucleus sample tests, all the common parameters are consistent." + }, + { + "type": "title", + "bbox": [ + 0.216, + 0.396, + 0.54, + 0.411 + ], + "angle": 0, + "content": "5.2 Benchmark & Evaluation Metrics" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.421, + 0.79, + 0.527 + ], + "angle": 0, + "content": "CHAIR [29]. Caption Hallucination Assessment with Image Relevance (CHAIR) is a widely used metric in image captioning tasks. CHAIR operates by creating a set of ground-truth object labels for each image. Any object mentioned in the caption that does not exist in the label set is considered a hallucinated object. CHAIR comprises two evaluation dimensions: instance-level and sentence-level, represented as \\(\\mathrm{CHAIR_I}\\) and \\(\\mathrm{CHAIR_S}\\), respectively. These are calculated in the following manner:" + }, + { + "type": "equation", + "bbox": [ + 0.374, + 0.539, + 0.785, + 0.57 + ], + "angle": 0, + "content": "\\[\n\\mathrm {C H A I R} _ {\\mathrm {I}} = \\frac {| \\{\\text {h a l l u c i n a t e d o b j e c t s} \\} |}{\\text {a l l m e n t i o n e d o b j e c t s}}, \\tag {5}\n\\]" + }, + { + "type": "equation", + "bbox": [ + 0.321, + 0.572, + 0.785, + 0.604 + ], + "angle": 0, + "content": "\\[\n\\mathrm {C H A I R} _ {\\mathrm {S}} = \\frac {| \\{\\text {c a p t i o n s w i t h h a l l u c i n a t e d o b j e c t s} \\} |}{\\text {a l l c a p t i o n s}}. \\tag {6}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.613, + 0.804, + 0.704 + ], + "angle": 0, + "content": "We conducted experiments based on the validation set of MSCOCO 2014. Given \"Please help me describe the image in detail.\" as prompt, we subsequently employ the CHAIR metric to evaluate the generated description. To evaluate hallucinations in long sequence generation, we adopted the same setup as used in LURE [43] and OPERA [13], setting the max_new_tokens parameter to 512 and randomly sampled 500 instances for evaluation." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.705, + 0.786, + 0.84 + ], + "angle": 0, + "content": "POPE [19]. The Polling-based Object Probing Evaluation (POPE) is a evaluation metric designed in the VQA paradigm. POPE serves as a metric for assessing object hallucination, evaluating hallucinations by asking LVLMs questions such as \"Is there a in the image?\" Here, is replaced with the constructed ground-truth object from three different types of splits. In the \"random\" split, objects are randomly selected from the entire dataset for evaluation. In the \"popular\" split, objects are chosen from those most frequently appearing in the dataset. In the \"adversarial\" split, objects that are highly related to the image objects are selected for evaluation. We conduct our evaluation" + } + ], + [ + { + "type": "page_number", + "bbox": [ + 0.218, + 0.116, + 0.236, + 0.127 + ], + "angle": 0, + "content": "10" + }, + { + "type": "header", + "bbox": [ + 0.272, + 0.115, + 0.335, + 0.127 + ], + "angle": 0, + "content": "Liu et al." + }, + { + "type": "table_caption", + "bbox": [ + 0.214, + 0.144, + 0.788, + 0.215 + ], + "angle": 0, + "content": "Table 1: CHAIR hallucination evaluation results on three LVLMs. CHAIR [29] is employed as the evaluation metric, where a smaller number indicates less hallucinations. OPERA is a decoding method based on beam search, VCD is a decoding method based on nucleus sampling, and PAI is an inference intervention method that can be collaborated with any decoding method." + }, + { + "type": "table", + "bbox": [ + 0.227, + 0.227, + 0.773, + 0.401 + ], + "angle": 0, + "content": "
DecodingMethodLLAVA [24]Minigpt4 [44]Shikra [6]
CHAIRSCHAIRICHAIRSCHAIRICHAIRSCHAIRI
GreedyVanilla46.613.432.811.151.214.4
PAI24.86.926.38.837.610.0
Beam SearchVanilla46.414.346.613.453.014.7
OPERA [13]44.614.430.19.836.812.4
PAI21.85.624.86.935.811.4
NucleusVanilla58.218.232.711.957.916.4
VCD [16]51.815.134.811.557.616.3
PAI43.414.726.710.349.913.2
" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.431, + 0.788, + 0.506 + ], + "angle": 0, + "content": "on the COCO dataset with 500 images, with each image having 6 questions for each split of POPE. We evaluate the performance of the model in object recognition tasks using both accuracy score and F1 score. Besides, in order to more comprehensively examine hallucinations in multi-turn chats, we have structured the evaluation in both single-turn and multi-turn dialogue forms." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.507, + 0.788, + 0.702 + ], + "angle": 0, + "content": "MMHal-Bench [32]. For further evaluation of our method on some challenging datasets, we choose MMHal-Bench, which is designed with 96 image-question pairs, spread across 8 question categories \\(\\times 12\\) object topics. It contains eight types of questions about object attributes, adversarial objects, comparisons, counting, spatial relations, environment, holistic descriptions, and others to comprehensively assess the model's hallucination performance on high-difficulty datasets. Essentially, it is also a VQA-based evaluation, but unlike the existence-based examination in POPE, its questions also include some logical considerations. Therefore, for evaluation on MMHal-Bench, we first need to answer questions and then use GPT-4 to score the answer based on the response and the ground-truth answer. The evaluation results include the model's scores across all question categories, and the overall score represents the average of these scores." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.704, + 0.788, + 0.84 + ], + "angle": 0, + "content": "GPT-4v Assisted Evaluation. To further evaluate the model's performance in image description tasks, we can move beyond the CHAIR metric, which is based on information extraction and only considers object hallucination. We can use GPT-4v for open evaluation. As with previous evaluations [13,16], we sample 50 images on the COCO dataset for evaluation. We construct prompts, input images into GPT-4v, along with the descriptions responses from two assistants. GPT-4v evaluation takes into account two dimensions: Accuracy and Detailedness, denoted respectively as C and D. Detailed prompt construction can be found in the appendix." + } + ], + [ + { + "type": "header", + "bbox": [ + 0.703, + 0.115, + 0.732, + 0.127 + ], + "angle": 0, + "content": "PAI" + }, + { + "type": "page_number", + "bbox": [ + 0.769, + 0.116, + 0.785, + 0.127 + ], + "angle": 0, + "content": "11" + }, + { + "type": "table_caption", + "bbox": [ + 0.238, + 0.145, + 0.764, + 0.159 + ], + "angle": 0, + "content": "Table 2: Quantitative comparison on POPE. The best results are in bold." + }, + { + "type": "table", + "bbox": [ + 0.22, + 0.171, + 0.785, + 0.314 + ], + "angle": 0, + "content": "
DecodingMethodLLAVA [24]Minigpt4 [44]Shikra [6]
Single-turnMulti-turnSingle-turnMulti-turnSingle-turnMulti-turn
AccF1AccF1AccF1AccF1AccF1AccF1
GreedyVanilla84.7685.5185.6984.2774.3973.7878.077.7781.5682.0578.1678.77
PAI85.8285.9787.5486.8375.276.280.8281.682.381.4778.5779.11
Beam SearchVanilla84.984.983.5781.1372.5467.7371.7867.9281.5281.8978.1878.13
PAI86.3385.8986.4285.2773.2269.3674.1571.7282.2581.0180.7280.02
NucleusVanilla80.2581.3282.8781.6357.5258.1664.4963.8979.5580.2876.5477.12
VCD [16]79.5381.0284.0383.359.9758.769.7269.3880.2280.576.3877.07
PAI81.7282.8785.0383.9759.1960.870.2470.2781.4980.7177.9978.47
" + }, + { + "type": "title", + "bbox": [ + 0.216, + 0.347, + 0.441, + 0.362 + ], + "angle": 0, + "content": "5.3 Experimental Results" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.381, + 0.784, + 0.457 + ], + "angle": 0, + "content": "In this section, we analyze the performance of PAI across various hallucination evaluation tasks, including long image description, simplified VQA answer, construction of metric evaluation, and leveraging the near-human cognitive capabilities of GPT-4/GPT-4v as evaluation methods. For further analysis, please refer to the appendix." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.459, + 0.785, + 0.61 + ], + "angle": 0, + "content": "Results on long sequence hallucination evaluation. The experimental results are presented in Tab. 1. As our approach is an inference intervention method, it differs from previous decoding hallucination mitigation methods that primarily concentrate on improving a single decoding method. We have tested our method on three decoding techniques. Our method has achieved hallucination mitigation on all three decoding methods used by the three models. However, when integrated with the nucleus, a sampling-based method, the hallucination reduction brought about by our method is not significant. This may be because even though our method has increased the priority of trustful tokens, the sample set during nucleus decoding still contains many hallucination tokens." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.612, + 0.785, + 0.718 + ], + "angle": 0, + "content": "Moreover, while OPERA significantly mitigates hallucinations, its time efficiency is considerably higher compared to vanilla. In contrast, our method not only has almost the same time efficiency as vanilla, but it also performs better in reducing hallucination issues. Compared to VCD, during the generation process of long sequence tasks, the introduction of visual uncertainty during decoding sometimes leads to more hallucination descriptions. However, our method can reduce the proportion of hallucination words in the sample pool." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.72, + 0.787, + 0.839 + ], + "angle": 0, + "content": "Results on single-turn and multi-turn hallucination evaluation. Unlike the CHAIR evaluation, POPE is in a VQA format, so the response is brief, answering only \"Yes\" or \"No\". The phenomena of text inertia and image neglect may not be as noticeable under this setting, especially for single-turn. However, our method still achieved a notable improvement compared to the vanilla decoding method in single-turn. As for multi-turn evaluation, which involves a longer context, our method achieves more significant improvement as presented in Tab. 2." + } + ], + [ + { + "type": "page_number", + "bbox": [ + 0.218, + 0.116, + 0.236, + 0.127 + ], + "angle": 0, + "content": "12" + }, + { + "type": "header", + "bbox": [ + 0.272, + 0.115, + 0.335, + 0.127 + ], + "angle": 0, + "content": "Liu et al." + }, + { + "type": "image", + "bbox": [ + 0.226, + 0.145, + 0.406, + 0.252 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.283, + 0.255, + 0.351, + 0.267 + ], + "angle": 0, + "content": "(a) LLAVA" + }, + { + "type": "image", + "bbox": [ + 0.411, + 0.149, + 0.593, + 0.253 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.463, + 0.256, + 0.542, + 0.268 + ], + "angle": 0, + "content": "(b) Minigpt4" + }, + { + "type": "image", + "bbox": [ + 0.597, + 0.149, + 0.78, + 0.252 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.657, + 0.255, + 0.72, + 0.268 + ], + "angle": 0, + "content": "(c) Shikra" + }, + { + "type": "image_caption", + "bbox": [ + 0.215, + 0.279, + 0.785, + 0.308 + ], + "angle": 0, + "content": "Fig. 6: Quantitative comparison on the MMHal-Bench. Higher scores indicate better performance." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.336, + 0.785, + 0.472 + ], + "angle": 0, + "content": "Results on hallucination evaluation in comprehensive general scenarios. The experimental results, as shown in Fig. 6, indicate that for some more image-based question types, such as object attributes, adversarial objects, and holistic questions, the answers are more accurate when inference intervention with PAI is applied, and there is a certain degree of improvement across all models. However, for some logical questions, such as comparisons and relations, there is no noticeable improvement after intervention. In summary, through the overall metric, i.e., the average of the eight evaluation dimensions, there is a certain degree of improvement compared to the baseline after incorporating PAI." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.472, + 0.787, + 0.593 + ], + "angle": 0, + "content": "Results on human-like GPT-4v assisted hallucination evaluation. The experimental results, as shown in Tab. 3, indicate that even when more comprehensive hallucination evaluation dimensions are added, our method, compared to the greedy decoding method, can provide more accurate responses on all three models without losing detail in the description. Given that GPT-4v's visual understanding and language logic capabilities have reached a level close to that of humans, it can more comprehensively illustrate the performance improvements brought about by our method." + }, + { + "type": "table_caption", + "bbox": [ + 0.248, + 0.618, + 0.754, + 0.632 + ], + "angle": 0, + "content": "Table 3: Results on GPT-4V evaluation. The best results are in bold." + }, + { + "type": "table", + "bbox": [ + 0.358, + 0.644, + 0.64, + 0.724 + ], + "angle": 0, + "content": "
MethodLLAVAMinigpt4Shikra
CDCDCD
Greedy5.625.245.85.745.545.25
PAI6.465.367.045.896.045.05
" + }, + { + "type": "title", + "bbox": [ + 0.216, + 0.772, + 0.389, + 0.787 + ], + "angle": 0, + "content": "5.4 Ablation Study" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.795, + 0.785, + 0.84 + ], + "angle": 0, + "content": "Our method, PAI, consists of two stages of interventions. In the first stage, during forward inference, the hyperparameter \\(\\alpha\\) is utilized to set the scale of intervention. Simultaneously, the layer prior, represented as \"L\", is used to determine the" + } + ], + [ + { + "type": "header", + "bbox": [ + 0.703, + 0.115, + 0.732, + 0.127 + ], + "angle": 0, + "content": "PAI" + }, + { + "type": "page_number", + "bbox": [ + 0.769, + 0.116, + 0.786, + 0.127 + ], + "angle": 0, + "content": "13" + }, + { + "type": "table_caption", + "bbox": [ + 0.215, + 0.145, + 0.788, + 0.187 + ], + "angle": 0, + "content": "Table 4: Ablation Study of the Hyperparameter \\(\\alpha\\). When \\(\\alpha\\) becomes excessively large, resulting in an unbalanced response, we terminate the experiment and denote this with a dash (-). F1 values that are considered outliers are highlighted in red." + }, + { + "type": "table", + "bbox": [ + 0.237, + 0.199, + 0.768, + 0.355 + ], + "angle": 0, + "content": "
αγLLLaVAMinigpt4Shikra
CHAIRSCHAIRiF1CHAIRSCHAIRiF1CHAIRSCHAIRiF1
---46.213.875.931.610.569.456.215.874.6
0.11.1✓47.413.876.925.08.069.956.415.075.7
0.21.1✓47.413.676.921.27.770.357.215.675.2
0.41.1✓42.412.476.915.16.367.455.214.475.3
0.51.1✓24.66.874.75.42.961.551.613.275.8
0.61.1✓7.84.862.6---37.210.676.7
0.71.1✓------14.46.164.9
" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.384, + 0.784, + 0.445 + ], + "angle": 0, + "content": "attention layer for intervention. The second stage unfolds during the decoding process, where we mitigate text inertia by subtracting the logits distribution that results from inputs devoid of image information. In this stage, the scale is managed by the parameter \\(\\gamma\\)." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.445, + 0.788, + 0.596 + ], + "angle": 0, + "content": "We use LLaVA-1.5 as the representative LVLM baseline and the greedy decoding method as the basic baseline to compare the impact of our hyperparameters on the task of long sequence image description. To evaluate our method, we choose the CHAIR metric. However, since CHAIR only evaluates the hallucination problem, we have incorporated the F1 score to consider information richness and accuracy. This makes the comparison with the CHAIR metric fairer when the F1 scores are similar. In the above, the F1 scores of the various methods differ slightly, so we have not included this somewhat redundant measure. The F1 score is calculated by counting the objects included in the description, the objects in the ground-truth sets, and the hallucinated objects." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.597, + 0.788, + 0.688 + ], + "angle": 0, + "content": "Effects of \\(\\alpha\\) in Exciting Image Attention. In the process of exciting the attention values of image tokens, we introduce a parameter \\(\\alpha\\) to control the amplification scale. As shown in Tab. 4, different LVLMs exhibit varying sensitivity to the amplification scale. This sensitivity not only depends on the length of the model's image tokens (e.g., the image token length of LLaVA-1.5 is 576, while that of Minigpt4 is only 32), but also the original attention weights distribution." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.689, + 0.788, + 0.763 + ], + "angle": 0, + "content": "However, a commonality across these models is that an appropriate amplification scale can achieve a balance between the number of hallucinated objects in the description and the amount of information conveyed. If the scale is too small, the description may still contain many hallucinated objects. Conversely, if the scale is too large, the amount of information in the response will decrease." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.765, + 0.788, + 0.84 + ], + "angle": 0, + "content": "Effects of Layer Prior in Exciting Image Attention. We further investigate the control of the intervention layer for exciting image attention. As observed in Tab. 5, the introduction of this prior does indeed improve the performance of our method. However, when there is no control over the intervention layer and interventions are applied to all layers, the performances of different models ex" + } + ], + [ + { + "type": "page_number", + "bbox": [ + 0.218, + 0.116, + 0.236, + 0.127 + ], + "angle": 0, + "content": "14" + }, + { + "type": "header", + "bbox": [ + 0.272, + 0.115, + 0.335, + 0.127 + ], + "angle": 0, + "content": "Liu et al." + }, + { + "type": "table_caption", + "bbox": [ + 0.217, + 0.145, + 0.787, + 0.174 + ], + "angle": 0, + "content": "Table 5: Ablation Study of Hyperparameter \\(\\gamma\\) and Layer Prior \"L\". Results are presented for the models LLAVA (left), Minigpt4 (middle), and Shikra (right)." + }, + { + "type": "table", + "bbox": [ + 0.222, + 0.185, + 0.782, + 0.314 + ], + "angle": 0, + "content": "
αγL CHAIRSCHAIRIF1αγL CHAIRSCHAIRIF1αγL CHAIRSCHAIRIF1
--46.213.875.9--31.610.569.4--56.215.874.6
0.51.1✓24.66.875.70.21.1✓21.27.770.30.61.1✓37.2
0.51.2✓23.46.675.70.21.2✓16.46.270.30.61.2✓40.2
0.51.3✓24.67.574.40.21.3✓14.46.969.50.61.3✓40.6
0.51.5✓25.09.674.60.21.5✓12.65.467.10.61.5✓40.2
0.52.0✓24.27.174.00.22.0✓8.44.664.30.62.0✓39.4
0.51.1✓24.66.875.70.21.1✓21.27.770.30.61.1✓37.2
0.51.1×20.66.771.70.21.1×22.38.470.30.61.1×56.4
" + }, + { + "type": "text", + "bbox": [ + 0.217, + 0.347, + 0.785, + 0.407 + ], + "angle": 0, + "content": "hibit some variations. For Shikra, in the absence of the layer prior, our method's results regress to the baseline. For both LLaVA and Minigpt4, the loss of the layer prior causes some unfavorable fluctuations in both the CHAIR metric and the F1 score." + }, + { + "type": "text", + "bbox": [ + 0.217, + 0.41, + 0.787, + 0.516 + ], + "angle": 0, + "content": "Effects of \\(\\gamma\\) in Mitigating Language Prior. Tab. 5 presents the results of an ablation study focusing on \\(\\gamma\\), which adjusts the balance between output distributions from conditioned inputs with excited image tokens and pure text inputs. Unlike the other two models, the Minigpt4 model is highly sensitive to \\(\\gamma\\). When \\(\\gamma\\) is too large, it can lead to uncontrollable model behavior. The experimental results suggest that maintaining \\(\\gamma\\) within a relatively small range, such as 1.1 - 1.2, yields the most stable performance." + }, + { + "type": "title", + "bbox": [ + 0.218, + 0.551, + 0.509, + 0.568 + ], + "angle": 0, + "content": "6 Conclusion and Limitation" + }, + { + "type": "text", + "bbox": [ + 0.217, + 0.596, + 0.787, + 0.732 + ], + "angle": 0, + "content": "In this paper, we first analyze the causes and manifestations of hallucinations in LVLMs. We propose a phenomenon termed \"text inertia,\" wherein the model continues to produce the same hallucinatory descriptions even when no image input is provided. This issue fundamentally stems from the model's neglect of image tokens. Consequently, we introduce the PAI method to intervene in the model's inference process, steering it towards an image-based and trustworthy direction. This is a training-free method and does not require any external tools. Extensive experiments on multiple benchmarks and LVLMs have validated the effectiveness of PAI in mitigating hallucination issues." + }, + { + "type": "text", + "bbox": [ + 0.217, + 0.734, + 0.787, + 0.84 + ], + "angle": 0, + "content": "Limitations: (1) The language decoders of existing open-source LVLMs are primarily models from the LLaMA-family. It is worth exploring whether the issues of image neglect and text inertia are introduced by LLaMA. (2) As we described, our method fundamentally alleviates the image neglect issue during model inference. Its upper limit depends on the capabilities of the well-trained model. Therefore, it is worth investigating whether incorporating this issue as a loss during the training process could lead to further performance improvements." + } + ], + [ + { + "type": "header", + "bbox": [ + 0.703, + 0.115, + 0.732, + 0.127 + ], + "angle": 0, + "content": "PAI" + }, + { + "type": "page_number", + "bbox": [ + 0.769, + 0.116, + 0.786, + 0.127 + ], + "angle": 0, + "content": "15" + }, + { + "type": "title", + "bbox": [ + 0.218, + 0.145, + 0.403, + 0.163 + ], + "angle": 0, + "content": "Acknowledgements" + }, + { + "type": "text", + "bbox": [ + 0.217, + 0.177, + 0.785, + 0.207 + ], + "angle": 0, + "content": "This work is supported by the National Science Foundation of China (62132017), Zhejiang Provincial Natural Science Foundation of China (LD24F020011)." + }, + { + "type": "title", + "bbox": [ + 0.218, + 0.23, + 0.323, + 0.245 + ], + "angle": 0, + "content": "References" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.26, + 0.786, + 0.329 + ], + "angle": 0, + "content": "1. Alayrac, J.B., Donahue, J., Luc, P., Miech, A., Barr, I., Hasson, Y., Lenc, K., Mensch, A., Millican, K., Reynolds, M., Ring, R., Rutherford, E., Cabi, S., Han, T., Gong, Z., Samangooei, S., Monteiro, M., Menick, J., Borgeaud, S., Brock, A., Nematzadeh, A., Sharifzadeh, S., Binkowski, M., Barreira, R., Vinyals, O., Zisserman, A., Simonyan, K.: Flamingo: a visual language model for few-shot learning" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.331, + 0.786, + 0.357 + ], + "angle": 0, + "content": "2. Anisha, G., Yin, J., Bas, E.: Detecting and preventing hallucinations in large vision language models (Aug 2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.359, + 0.786, + 0.384 + ], + "angle": 0, + "content": "3. Azaria, A., Mitchell, T.: The internal state of an llm knows when its lying (Apr 2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.385, + 0.786, + 0.481 + ], + "angle": 0, + "content": "4. Brown, T., Mann, B., Ryder, N., Subbiah, M., Kaplan, J., Dhariwal, P., Neelakanthan, A., Shyam, P., Sastry, G., Amanda, A., Agarwal, S., Herbert-Voss, A., Krueger, G., Tom, H., Child, R., Ramesh, A., Ziegler, D., Wu, J., Winter, C., Hesse, C., Chen, M., Sigler, E., Litwin, M., Gray, S., Benjamin, C., Clark, J., Berner, C., Sam, M., Radford, A., Sutskever, I., Amodei, D.: Language models are few-shot learners. arXiv: Computation and Language, arXiv: Computation and Language (May 2020)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.483, + 0.786, + 0.509 + ], + "angle": 0, + "content": "5. Burns, C., Ye, H., Klein, D., Steinhardt, J.: Discovering latent knowledge in language models without supervision (Dec 2022)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.511, + 0.786, + 0.55 + ], + "angle": 0, + "content": "6. Chen, K., Zhang, Z., Zeng, W., Zhang, R., Zhu, F., Zhao, R., Research, S., Sklsde, S., Beihang, B., Seiee, S.: Shikra: Unleashing multimodal llm's referential dialogue magic" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.552, + 0.786, + 0.592 + ], + "angle": 0, + "content": "7. Chen, Z., Wu, J., Wang, W., Su, W., Chen, G., Xing, S., Muyan, Z., Zhang, Q., Zhu, X., Lu, L., Li, B., Luo, P., Lu, T., Qiao, Y., Dai, J.: Intervl: Scaling up vision foundation models and aligning for generic visual-linguistic tasks (Dec 2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.594, + 0.786, + 0.633 + ], + "angle": 0, + "content": "8. Dai, W., Li, J., Li, D., Huat, A., Zhao, J., Wang, W., Li, B., Fung, P., Hoi, S.: Instructclip: Towards general-purpose vision-language models with instruction tuning" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.635, + 0.786, + 0.66 + ], + "angle": 0, + "content": "9. Darcet, T., Oquab, M., Mairal, J., Bojanowski, P.: Vision transformers need registers. arXiv preprint arXiv:2309.16588 (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.22, + 0.661, + 0.786, + 0.716 + ], + "angle": 0, + "content": "10. Devlin, J., Chang, M.W., Lee, K., Toutanova, K.: Bert: Pre-training of deep bidirectional transformers for language understanding. In: Proceedings of the 2019 Conference of the North (Jan 2019). https://doi.org/10.18653/v1/n19-1423, http://dx.doi.org/10.18653/v1/n19-1423" + }, + { + "type": "ref_text", + "bbox": [ + 0.22, + 0.717, + 0.786, + 0.744 + ], + "angle": 0, + "content": "11. Han, Y., Zhang, C., Chen, X., Yang, X., Wang, Z., Yu, G., Fu, B., Zhang, H.: Chartlama: A multimodal llm for chart understanding and generation (Nov 2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.22, + 0.745, + 0.786, + 0.771 + ], + "angle": 0, + "content": "12. Hu, H., Zhang, J., Zhao, M., Sun, Z.: Ciem: Contrastive instruction evaluation method for better instruction tuning" + }, + { + "type": "ref_text", + "bbox": [ + 0.22, + 0.772, + 0.786, + 0.812 + ], + "angle": 0, + "content": "13. Huang, Q., Dong, X., Zhang, P., Wang, B., He, C., Wang, J., Lin, D., Zhang, W., Yu, N.: Opera: Alleviating hallucination in multi-modal large language models via over-trust penalty and retrospection-allocation (Nov 2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.22, + 0.813, + 0.786, + 0.84 + ], + "angle": 0, + "content": "14. Kai, J., Zhang, T., Hu, H., Lin, Z.: Sh2: Self-highlighted hesitation helps you decode more truthfully (Jan 2024)" + }, + { + "type": "list", + "bbox": [ + 0.22, + 0.26, + 0.786, + 0.84 + ], + "angle": 0, + "content": null + } + ], + [ + { + "type": "page_number", + "bbox": [ + 0.218, + 0.116, + 0.236, + 0.127 + ], + "angle": 0, + "content": "16" + }, + { + "type": "header", + "bbox": [ + 0.272, + 0.115, + 0.335, + 0.127 + ], + "angle": 0, + "content": "Liu et al." + }, + { + "type": "ref_text", + "bbox": [ + 0.218, + 0.148, + 0.785, + 0.175 + ], + "angle": 0, + "content": "15. Lai, X., Tian, Z., Chen, Y., Li, Y., Yuan, Y., Liu, S., Jia, J., Kong, H., Research, M.: Lisa: Reasoning segmentation via large language model" + }, + { + "type": "ref_text", + "bbox": [ + 0.218, + 0.177, + 0.785, + 0.203 + ], + "angle": 0, + "content": "16. Leng, S., Zhang, H., Chen, G., Li, X., Lu, S., Miao, C., Bing, L.: Mitigating object hallucinations in large vision-language models through visual contrastive decoding" + }, + { + "type": "ref_text", + "bbox": [ + 0.218, + 0.204, + 0.785, + 0.231 + ], + "angle": 0, + "content": "17. Li, J., Li, D., Savarese, S., Hoi, S.: Blip-2: Bootstrapping language-image pretraining with frozen image encoders and large language models" + }, + { + "type": "ref_text", + "bbox": [ + 0.218, + 0.232, + 0.785, + 0.272 + ], + "angle": 0, + "content": "18. Li, K., Patel, O., Viégas, F., Pfister, H., Wattenberg, M.: Inference-time intervention: Eliciting truthful answers from a language model. Advances in Neural Information Processing Systems 36 (2024)" + }, + { + "type": "ref_text", + "bbox": [ + 0.218, + 0.273, + 0.785, + 0.3 + ], + "angle": 0, + "content": "19. Li, Y., Du, Y., Zhou, K., Wang, J., Zhao, W.X., Wen, J.R.: Evaluating object hallucination in large vision-language models. arXiv preprint arXiv:2305.10355 (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.218, + 0.301, + 0.785, + 0.341 + ], + "angle": 0, + "content": "20. Li, Z., Yang, B., Liu, Q., Ma, Z., Zhang, S., Yang, J., Sun, Y., Liu, Y., Bai, X.: Monkey: Image resolution and text label are important things for large multi-modal models (Nov 2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.342, + 0.785, + 0.383 + ], + "angle": 0, + "content": "21. Liu, F., Lin, K., Li, L., Wang, J., Yacoob, Y., Wang, L.: Mitigating hallucination in large multi-modal models via robust instruction tuning. In: The Twelfth International Conference on Learning Representations (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.384, + 0.785, + 0.411 + ], + "angle": 0, + "content": "22. Liu, F., Lin, K., Li, L., Wang, J., Yacoob, Y., Wang, L.: Mitigating hallucination in large multi-modal models via robust instruction tuning (Sep 2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.412, + 0.785, + 0.438 + ], + "angle": 0, + "content": "23. Liu, H., Xue, W., Chen, Y., Chen, D., Zhao, X., Wang, K., Hou, L., Li, R., Peng, W.: A survey on hallucination in large vision-language models" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.44, + 0.785, + 0.466 + ], + "angle": 0, + "content": "24. Liu, H., Li, C., Li, Y., Lee, Y.J.: Improved baselines with visual instruction tuning. arXiv preprint arXiv:2310.03744 (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.467, + 0.785, + 0.493 + ], + "angle": 0, + "content": "25. Liu, H., Li, C., Wu, Q., Lee, Y., Madison, M., Research, M.: Visual instruction tuning" + }, + { + "type": "ref_text", + "bbox": [ + 0.218, + 0.495, + 0.785, + 0.563 + ], + "angle": 0, + "content": "26. Mishra, S., Khashabi, D., Baral, C., Hajishirzi, H.: Cross-task generalization via natural language crowdsourcing instructions. In: Proceedings of the 60th Annual Meeting of the Association for Computational Linguistics (Volume 1: Long Papers) (Jan 2022). https://doi.org/10.18653/v1/2022.acl-long.244, http://dx.doi.org/10.18653/v1/2022.acl-long.244" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.565, + 0.785, + 0.619 + ], + "angle": 0, + "content": "27. Ouyang, L., Wu, J., Jiang, X., Almeida, D., Wainwright, C.L., Mishkin, P., Zhang, C., Agarwal, S., Slama, K., Ray, A., Schulman, J., Hilton, J., Kelton, F., Miller, L., Simens, M., Askell, A., Welinder, P., Christiano, P.F., Leike, J., Lowe, R.: Training language models to follow instructions with human feedback (2022)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.62, + 0.74, + 0.633 + ], + "angle": 0, + "content": "28. Peng, B., Li, C., He, P., Galley, M., Gao, J.: Instruction tuning with gpt-4" + }, + { + "type": "ref_text", + "bbox": [ + 0.218, + 0.634, + 0.785, + 0.688 + ], + "angle": 0, + "content": "29. Rohrbach, A., Hendricks, L.A., Burns, K., Darrell, T., Saenko, K.: Object hallucination in image captioning. In: Proceedings of the 2018 Conference on Empirical Methods in Natural Language Processing (Jan 2018). https://doi.org/10.18653/v1/d18-1437, http://dx.doi.org/10.18653/v1/d18-1437" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.689, + 0.785, + 0.715 + ], + "angle": 0, + "content": "30. Sanchez, G., Fan, H., Spangher, A., Levi, E., Ammanamanchi, P.S., Biderman, S.: Stay on topic with classifier-free guidance. arXiv preprint arXiv:2306.17806 (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.218, + 0.717, + 0.785, + 0.757 + ], + "angle": 0, + "content": "31. Sun, Z., Shen, S., Cao, S., Liu, H., Li, C., Shen, Y., Gan, C., Gui, L.Y., Wang, Y.X., Yang, Y., Keutzer, K., Darrell, T.: Aligning large multimodal models with factually augmented rlhf (Sep 2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.758, + 0.785, + 0.798 + ], + "angle": 0, + "content": "32. Sun, Z., Shen, S., Cao, S., Liu, H., Li, C., Shen, Y., Gan, C., Gui, L.Y., Wang, Y.X., Yang, Y., et al.: Aligning large multimodal models with factually augmented rlhf. arXiv preprint arXiv:2309.14525 (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.8, + 0.785, + 0.84 + ], + "angle": 0, + "content": "33. Touvron, H., Lavril, T., Izacard, G., Martinet, X., Lachaux, M.A., Lacroix, T., Rozi'ere, B., Goyal, N., Hambro, E., Azhar, F., Rodriguez, A., Joulin, A., Grave, E., Lample, G.: Llama: Open and efficient foundation language models" + }, + { + "type": "list", + "bbox": [ + 0.217, + 0.148, + 0.785, + 0.84 + ], + "angle": 0, + "content": null + } + ], + [ + { + "type": "header", + "bbox": [ + 0.703, + 0.115, + 0.732, + 0.127 + ], + "angle": 0, + "content": "PAI" + }, + { + "type": "page_number", + "bbox": [ + 0.769, + 0.116, + 0.786, + 0.127 + ], + "angle": 0, + "content": "17" + }, + { + "type": "ref_text", + "bbox": [ + 0.214, + 0.147, + 0.788, + 0.189 + ], + "angle": 0, + "content": "34. Turner, A., Thiergart, L., Udell, D., Leech, G., Mini, U., MacDiarmid, M.: Activation addition: Steering language models without optimization. arXiv preprint arXiv:2308.10248 (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.214, + 0.189, + 0.788, + 0.231 + ], + "angle": 0, + "content": "35. Wang, L., Li, L., Dai, D., Chen, D., Zhou, H., Meng, F., Zhou, J., Sun, X.: Label words are anchors: An information flow perspective for understanding in-context learning. arXiv preprint arXiv:2305.14160 (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.214, + 0.231, + 0.788, + 0.259 + ], + "angle": 0, + "content": "36. Wei, F., Zhang, X., Zhang, A., Zhang, B., Chu, X.: Lenna: Language enhanced reasoning detection assistant (Dec 2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.214, + 0.259, + 0.788, + 0.286 + ], + "angle": 0, + "content": "37. Wei, J., Bosma, M., Zhao, V., Guu, K., Yu, A., Lester, B., Du, N., Dai, A., Le, Q.: Finetuned language models are zero-shot learners. Learning, Learning (Sep 2021)" + }, + { + "type": "ref_text", + "bbox": [ + 0.214, + 0.286, + 0.788, + 0.314 + ], + "angle": 0, + "content": "38. Xiao, G., Tian, Y., Chen, B., Han, S., Lewis, M.: Efficient streaming language models with attention sinks (Sep 2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.214, + 0.314, + 0.788, + 0.357 + ], + "angle": 0, + "content": "39. Ye, Q., Xu, H., Xu, G., Ye, J., Yan, M., Zhou, Y., Wang, J., Hu, A., Shi, P., Shi, Y., Li, C., Xu, Y., Chen, H., Tian, J., Qi, Q., Zhang, J., Huang, F.: mplug-owl: Modularization empowers large language models with multimodality" + }, + { + "type": "ref_text", + "bbox": [ + 0.214, + 0.356, + 0.788, + 0.384 + ], + "angle": 0, + "content": "40. Yin, S., Fu, C., Zhao, S., Xu, T., Wang, H., Sui, D., Shen, Y., Li, K., Sun, X., Chen, E.: Woodpecker: Hallucination correction for multimodal large language models" + }, + { + "type": "ref_text", + "bbox": [ + 0.214, + 0.384, + 0.788, + 0.424 + ], + "angle": 0, + "content": "41. You, H., Zhang, H., Gan, Z., Du, X., Zhang, B., Wang, Z., Cao, L., Chang, S.F., Yang, Y.: Ferret: Refer and ground anything anywhere at any granularity (Oct 2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.214, + 0.424, + 0.788, + 0.452 + ], + "angle": 0, + "content": "42. Zhang, Y., Qian, S., Peng, B., Liu, S., Jia, J.: Prompt highlighter: Interactive control for multi-modal llms. arXiv preprint arXiv:2312.04302 (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.214, + 0.452, + 0.788, + 0.493 + ], + "angle": 0, + "content": "43. Zhou, Y., Cui, C., Yoon, J., Zhang, L., Deng, Z., Finn, C., Bansal, M., Yao, H.: Analyzing and mitigating object hallucination in large vision-language models (Oct 2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.214, + 0.493, + 0.788, + 0.522 + ], + "angle": 0, + "content": "44. Zhu, D., Chen, J., Shen, X., Li, X., Elhoseiny, M.: Minigpt-4: Enhancing vision-language understanding with advanced large language models" + }, + { + "type": "list", + "bbox": [ + 0.214, + 0.147, + 0.788, + 0.522 + ], + "angle": 0, + "content": null + } + ] +] \ No newline at end of file diff --git a/2024/Paying More Attention to Images_ A Training-Free Method for Alleviating Hallucination in LVLMs/71543d84-0242-4633-b85d-0ad4feeefd05_origin.pdf b/2024/Paying More Attention to Images_ A Training-Free Method for Alleviating Hallucination in LVLMs/71543d84-0242-4633-b85d-0ad4feeefd05_origin.pdf new file mode 100644 index 0000000000000000000000000000000000000000..8c94f0f6c65cfe1c8f3da9f1fd8b8e745ce4f652 --- /dev/null +++ b/2024/Paying More Attention to Images_ A Training-Free Method for Alleviating Hallucination in LVLMs/71543d84-0242-4633-b85d-0ad4feeefd05_origin.pdf @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:458004172953839b99c9efdbf16776bf9a4854b7e616ee62d5ce36c39ce26f16 +size 1008834 diff --git a/2024/Paying More Attention to Images_ A Training-Free Method for Alleviating Hallucination in LVLMs/full.md b/2024/Paying More Attention to Images_ A Training-Free Method for Alleviating Hallucination in LVLMs/full.md new file mode 100644 index 0000000000000000000000000000000000000000..9917c5dd52a1a5488ecdae064343fec3d98c7410 --- /dev/null +++ b/2024/Paying More Attention to Images_ A Training-Free Method for Alleviating Hallucination in LVLMs/full.md @@ -0,0 +1,288 @@ +# Paying More Attention to Image: A Training-Free Method for Alleviating Hallucination in LVLMs + +Shi Liu $^{1}$ , Kecheng Zheng $^{1,2\dagger}$ , and Wei Chen $^{1\dagger}$ + +$^{1}$ State Key Lab of CAD&CG, Zhejiang University + +2 Ant Group + +{liushi0927,zkechengzk}@gmail.com + +chenvis@zju.edu.cn + +Abstract. Existing Large Vision-Language Models (LVLMs) primarily align image features of vision encoder with Large Language Models (LLMs) to leverage their superior text generation capabilities. However, the scale disparity between vision encoder and language model may led to LLMs assuming a predominant role in multi-modal comprehension. This imbalance in LVLMs may result in the instances of hallucinatory. Concretely, LVLMs may generate consistent descriptions with or without visual input, indicating that certain outputs are influenced solely by context text. We refer to this phenomenon as "text inertia." To counteract this issue, we introduce a training-free algorithm to find an equilibrium point between image comprehension and language inference. Specifically, we adaptively involve adjusting and amplifying the attention weights assigned to image tokens, thereby granting greater prominence to visual elements. Meanwhile, we subtract the logits of multi-modal inputs from ones of pure text input, which can help LVLMs be not biased towards LLMs. By enhancing images tokens and reducing the stubborn output of LLM, we can let LVLM pay more attention to images, towards alleviating text inertia and reducing the hallucination in LVLMs. Our extensive experiments show that this method substantially reduces the frequency of hallucinatory outputs in various LVLMs in terms of different metrics. Project page is available at https://1albj.github.io/projects/PAI/ + +Keywords: Vision-Language Models $\cdot$ Visual Dialogs $\cdot$ Hallucination Mitigation + +# 1 Introduction + +Recently, Large Vision-Language Models (LVLMs) have made significant strides, exhibiting impressive capabilities across a multitude of tasks [11,15,36,42]. However, these models still struggle with the phenomenon of hallucination. Specifically, there is often a mismatch between the textual content generated by the model and the actual visual input it receives [23]. + +Hallucination in LVLMs is often attributed to issues with modality alignment, leading to the development of mitigating strategies through alignment + +![](images/917107f1f4568805fcb7915f42da9f1fe4a850bc295b6797be7ad9aaba2190b8.jpg) +Fig. 1: We present an examination of various input settings, with hallucinations specifically highlighted in red. (a) When using LLAVA for image description, it generates a hallucinated description. (b) Even without image input, when only the historical response preceding the hallucinated description is input to LLAVA, it reproduces the same hallucinated description, a phenomenon we refer to as "Text Inertia". (c) Our proposed method, PAI, effectively mitigates this text inertia problem and yields accurate descriptions. (d) Utilizing PAI for image description results in a significantly more precise description. + +![](images/8ac3098831f766ebecc7681dddac6a072a8bce51a13fa9a9862ff6938a42c29e.jpg) +Fig.2: Percentage of text inertia hallucination in all hallucination (calculated with 500 samples). For specific calculation processes, please refer to the supplementary material section A. +Fig. 3: Visualization of the average attention ratio of different content. The x-axis denotes the sequence length of the history tokens. The lengths of the image, instruction, BOS, and outlier tokens [9] are all fixed as they are part of the model input, with 576 tokens for the image, 21 for the instruction, and 1 each for the BOS and outlier tokens. + +![](images/a2dd0958982eb6cd370c48df41b6d22f3e71098f63f8931d166bf9fc1e3a3984.jpg) + +training optimization [7, 20, 31]. However, is hallucination in LVLMs merely a result of the model's capacity, and can it only be alleviated through additional training? We propose a scenario where LVLMs generate a hallucinated object description. Specifically, even when the image input is removed and only the generated text preceding the hallucinated object word is retained, the LVLMs persist in producing the same hallucinated description, as depicted in Fig. 1. + +In order to empirically investigate this behavior, we conducted tests on three LVLMs within the context of image describing tasks on the COCO dataset. We identified and conducted a statistical analysis on instances where LVLMs + +generated identical hallucinated object descriptions, even when the input was exclusively historical response text without any image. The observation from Fig. 2 clearly indicates that, even with the application of rigorous identification settings, the phenomenon continues to represent a substantial proportion. + +We referred this phenomenon as "Text Inertia". Our hypothesis is that text inertia arises due to the current generative paradigms mapping image representations onto the text representation space as text tokens. In this mechanism, the LLM becomes the dominant character, and the inference process lacks additional handling of image tokens, leading to their neglect during the generation process. To validate this hypothesis, we have analyzed the attention values ratios of the LLaVA model during the inference process in Fig. 3. Our findings show that despite image tokens occupying a significant proportion, they do not receive substantial attention under the current mechanism. This multimodal chat resembles more of an automatic completion based on context rather than a continuous attention to the image for completion. + +To close this gap, we introduce a method referred to as Pay Attention to Image (PAI). At a high level, PAI intervenes in the inference process to make it more image-centric, following the original image perception direction. To achieve this, we focus on the self-attention heads in the decoder layers of LVLMs. We enhance the attention weights for image tokens in their original directions during inference. This allows us to use the updated attention matrix to calculate the hidden states for the generated token, thereby incorporating more consideration for image representation during the generation process. To further mitigate text inertia, we construct the input using instruction tokens and historical response tokens, and subtract the model logits of this input from the logits of original model with image tokens input. This strategy helps to reduce the influence of language priors during the generation process. Unlike previous methods for mitigating hallucination that require additional training or external tools, our approach is training-free. Moreover, we are the first to propose an inference intervention method for mitigating hallucination in LVLMs. + +We then evaluate the response accuracy in the image description task from long sequence generation perspective by employing the CHAIR metric [29] and GPT-4V. In addition, we use POPE [19] and MMHal-Bench [32] to more comprehensively evaluate the model's hallucination performance on VQA task. Furthermore, we have constructed single-turn and multi-turn chat evaluations for POPE. Since our model intervenes in the inference process, it can be used for any decoding method. Therefore, we conducted experiments on three decoding methods of the three models. The experimental results proved the effectiveness of our method in mitigating hallucinations. + +# 2 Related Work + +# 2.1 Large Vision-Language Models + +The development of pre-training techniques [4, 27] and instruction tuning techniques [26, 37] has rapidly advanced LLMs technology, such as LLaMA [33] and + +![](images/fadb7576ac2c6adb7a89282889adfa54b732b07b7997bf7c889672fcdf9f7d64.jpg) +Fig. 4: Architecture of our PAI. To alleviate text inertia, we additionally construct an input without image. Throughout the forward inference process, we amplify the focus on the image token by edit self-attention maps in LLaMA. Ultimately, we subtract the logits distribution of the language prior during decoding to achieve an accurate description. + +Vicuna [28], further leading to the prosperity of LVLMs technology. Early works, like Flamingo [1] and BLIP-2 [17], have successfully adapted LLMs to visual tasks, demonstrating notable generative capabilities and in-context learning abilities. Recently, the capabilities of LVLMs have further advanced under the influence of visual instruction tuning techniques [24,25]. Using different projectors to map images to the text domain, thereby endowing language generation models with image understanding capabilities, is also a hot research topic [8,25,39,44]. Additionally, several studies focus on visual language tasks such as grounding capabilities [6] and reasoning capabilities [15]. However, recent LVLMs still face the issue of hallucination generation [21]. + +# 2.2 Mitigation of LVLMs Hallucination + +Hallucination in LVLMs refers to contradictions between the image input and the textual output. Various methods have been proposed to mitigate hallucination. The most direct reason for hallucination generation is that hallucination arises from data bias and knowledge gaps between vision and language. Therefore, better data filtering methods [12, 22, 41] and higher quality annotated data [2] are introduced. Simultaneously, these methods also imply the need for more alignment training [31] or adjustments in the model architecture [7, 20]. These methods can achieve good results, but they are time-consuming and require high computational resources. + +Apart from addressing the ability of LVLMs itself, hallucination can also be mitigated through post-processing methods. This method usually involves using an additional module or external tools to edit the response. Recent methods such as LURE [43] utilize additional data to train a state detector and when hallucination issues are detected, content is regenerated by a revisor model. Woodpecker [40] introduces an external visual model to inspect entities extracted from + +the response, and then the detection results are handed over to the generation model to regenerate better answers. These methods also extend the inference chain and increase inference costs. + +Training-free hallucination mitigation methods have so far only been attempted in decoding methods. OPERA [13] discovered an abnormal attention pattern that accompanies model decoding. It was statistically found that this pattern often accompanies hallucination descriptions, and thus a detection and mitigation method was proposed based on this pattern to alleviate the hallucination faced by the model. VCD [16] introduced the notion that visual uncertainty increases hallucination descriptions and, based on this discovery, proposed a contrast decoding method to alleviate hallucination issues. + +# 3 Preliminaries + +The architecture of LVLMs typically comprises of three main components: an image encoder, a projector, and a language decoder. Both the image encoder and language decoder are usually pre-trained. The image encoder is employed to transform images into image tokens, which are subsequently mapped to the text representation space by the projector. This process enables the concatenation of image tokens with text tokens that are then fed into the language decoder. The language decoder subsequently generates corresponding responses based on the provided instructions. + +The existing projectors. Currently, projectors predominantly fall into two categories: linear projectors and resamplers. A projector takes N visual features from the image encoder and transforms them into M visual tokens. The linear projector employs a multilayer perceptron to transform all visual features, maintaining a one-to-one transformation which meaning that M equals N. In contrast, the resampler does not preserve all visual features but instead samples visual cues (M, where $\mathrm{M} < \mathrm{N}$ ). For instance, Q-former [17] utilizes M learnable queries and Bert [10] to extract information from visual features. Given that the knowledge of images during the generation process solely originates from the output image tokens of the projectors, our attention is concentrated on the image tokens post-projection, irrespective of their preceding modeling process. + +Autoregressive language decoders. Nearly all LVLMs adopt LLaMA-family models as their language decoders, which employ the self-attention mechanism. The visual tokens processed by the projector are concatenated with text tokens and fed into the LLaMA, which carries out the forward decoding process. From the perspective of a single attention head in a single layer, each head repeatedly performs the following attention operation with the same input shape: + +$$ +\boldsymbol {O} _ {h} = \boldsymbol {A} _ {h} \boldsymbol {V} _ {h}, \quad \boldsymbol {A} _ {h} = \operatorname {s o f t m a x} \left(\frac {\boldsymbol {Q} _ {h} \boldsymbol {K} _ {h} ^ {\top}}{\sqrt {d _ {k}}}\right). \tag {1} +$$ + +Each attention head $h$ performs an attention operation using its own set of queries $Q_{h} \in \mathbb{R}^{n \times d_{k}}$ , keys $K_{h} \in \mathbb{R}^{n \times d_{k}}$ , and values $V_{h} \in \mathbb{R}^{n \times d_{k}}$ , where $n$ represents the sequence length and $d_{k}$ represents the hidden dimensions. The + +output $O_{h} \in \mathbb{R}^{n \times d_{k}}$ is modeled by multiplying $V_{h}$ and the attention weights $A_{h} \in \mathbb{R}^{n \times n}$ , where each row represents the weights for each token during feature mixing. This operation enables the model to focus on different parts of the input for each head through attention weights, thereby capturing various parts of the information from the sequence token representations. The final output is the current generated token vocabulary conditional probability distribution $y \in \mathbb{R}^{v}$ based on the input instruction representations $X_{I}$ , image representations $X_{V}$ and history generated token representations $X_{H}$ , where $v$ is the size of the vocabulary. This process can be formatted as: + +$$ +\begin{array}{l} \boldsymbol {y} \sim p _ {\text {m o d e l}} (\boldsymbol {y} \mid \mathbf {X} _ {I}, \mathbf {X} _ {V}, \mathbf {X} _ {H}), \\ \begin{array}{l} p _ {\text {m o d e r}} (\boldsymbol {y} \mid 1, 1, 1, 1, 1, 1, 1, 1, 1), \\ \propto \operatorname {s o f t m a x} \left(\operatorname {l o g i t} _ {\text {m o d e l}} (\boldsymbol {y} \mid \mathbf {X} _ {I}, \mathbf {X} _ {V}, \mathbf {X} _ {H})\right), \end{array} \tag {2} \\ \end{array} +$$ + +which calculates the distribution of one token and iterates for the entire response. The sequence generation continues until an EOS (End of Sentence) token is produced, marking the end of the generation and resulting in a complete response. + +# 4 Method + +At the core of our method is a solution for image neglect and text inertia, both of which are fundamentally interconnected. Essentially, as paying more attention to the image, there is a corresponding reduction in the reliance on language priors. Intuitively, in a conversation centered around an image, the model should devote more attention to the image, thereby allowing it to have a significant impact on the response. As such, we identify the self-attention map in the token-level generation and augment the image attention in its original directions. This strategy promotes a more image-centric latent representations. Additionally, to further mitigate the influence of text inertia, we devide the logits distribution of pure text input into the model's output. + +# 4.1 Pay More Attention to Image + +Extracting the Self-Attention Matrix. We start from a token-level perspective. The response process in LVLMs is fundamentally generated token by token. Each token is generated based on the input image, instruction, and the historically generated response. This process is facilitated through a multi-layer attention decoder architecture. Consequently, this results in a probability distribution of the vocabulary for the currently generated token. Our goal is to extract the attention matrix of each attention head at every layer, indicating the influence of each content during inference. + +When generating the $k$ -th token in the sequence, the input representation for the attention head in the forward process includes the instruction representation $\mathbf{X}_I = [\pmb{x}_{i_1},\dots,\pmb{x}_{i_{n_I}}]$ , image representation $\mathbf{X}_V = [\pmb{x}_{v_1},\dots,\pmb{x}_{v_{n_V}}]$ , and the representation of the historically generated response $\mathbf{X}_H = [\pmb{x}_{h_1},\dots,\pmb{x}_{h_{n_H}}]$ . Notably, the image representation considered here is the one that has been processed by the projector. Essentially, the hidden states of each input layer are $\mathbf{X} = \mathrm{concat}(\mathbf{X}_I[1:m],\mathbf{X}_V,\mathbf{X}_I[m + 1:n_I],\mathbf{X}_H)$ , where the notation $\mathbf{X}_I[1:m]$ + +indicates the first $m$ elements in the instruction representation. Each attention head assigns different degrees of attention to each element during the current token representation decoding process. Our aim is to enhance the attention paid to the image. Therefore, we extract the attention weight values related to the image token for the current generated token, intervene, and then redistribute the attention values of each element through softmax. + +Excite model in a trustful direction. There have been attempts in some LLMs works to make the answers generated by LLMs more trustworthy by means of intervention [3, 5, 14, 18, 34]. The implemented approach typically involves intervening with the hidden states. As for defining what constitutes a more trustworthy direction, it usually requires additional projection and training to probe this trustful direction. In our case, a response that is more image-based is considered more trustworthy. Since LVLMs have undergone alignment training, the original attention values provide a direction based on image content. As illustrated in Fig. 4, by amplifying the attention values of image tokens based on the original attention values, we can enhance the trustworthiness of our results. + +Another nuance involves our avoidance of choosing the attention head used to shift intervention. In the ITI method [18], it is stated that not all attention heads should be subjected to intervention. Therefore, they introduce a trustful score to rank each head across all layers and select the top-k heads for intervention. In our case, the less trustworthy heads with lower attention values receive less intervention. We first extract the attention weights of the image tokens for the current generated token from the attention weights $\tilde{\pmb{A}}$ before softmax operation. We then use the hyper-parameter $\alpha$ to control the step size for intervention. From a single attention head perspective, our method can be expressed as follows: + +$$ +\tilde {\boldsymbol {A}} _ {n, j} = \tilde {\boldsymbol {A}} _ {n, j} + \alpha \cdot | \tilde {\boldsymbol {A}} _ {n, j} | \quad \text {f o r} j = m + 1 \text {t o} m + n _ {V}. \tag {3} +$$ + +The model's final vocabulary probability distribution is derived from the projection of the hidden states of the last token in the sequence. Therefore, we extract the attention weights of the last token $n$ on the image tokens by indexing $\tilde{A}_{n,j}$ . Following the intervention, we use the softmax function to redistribute the attention values of each token during the reassignment of encoded hidden states. This procedure is repeated for each subsequent token prediction in an autoregressive manner and is independent of the choice of the decoding algorithm. + +Excite more precisely with an attention mode prior. The presence of the BOS token, an attention sink pattern [38], in a sentence results in higher attention values during the attention computation process, which may seem counterintuitive. The BOS token typically signifies the start of a sentence and, as such, does not carry significant semantic content. However, the generation of tokens is significantly influenced by this particular token, a similar pattern that also manifests itself in visual models [9]. As mentioned in StreamLLM [38], the pattern of the attention sink emerges when redundant attention values are present. + +Naturally, one might infer that when the sink pattern appears, we excite the image token. To further investigate this phenomenon, as depicted in Fig. 5, we + +![](images/648d051086d77520a3c5a9c8117931c2f8d35ec709ace9f6e518616a5cab6353.jpg) +(a) LLAVA + +![](images/8ae6e4405431b66600a296d0dd9f8126ae0de3dfb721da08a8110a1d8b02fdc1.jpg) +(b) Minigpt4 +Fig. 5: The BOS token attention ratios of three model. We calculate the attention weights of the BOS token for each head in each layer and display them in a heat map. + +![](images/e1acd51a8090e93741e905402c511679c0fd2f57a88cf72eb889f7ef1b475fb9.jpg) +(c) Shikra + +find that the sink phenomenon is not overtly evident in the shallow layers. This is because the shallow layers tend to focus more on encoding semantically rich information [35]. When the encoding of semantically rich tokens stabilizes, the attention sink phenomenon arises. Therefore, we build upon the judgement of intervention timing by calculating the similarity of the hidden states. + +# 4.2 Image-Centric Logit Refine + +In Fig. 1, we observe a peculiar phenomenon where LVLMs continue to generate identical hallucinated text even when the image is removed from the input. This observation naturally leads us to the concept of using the output distribution (when no image is in input) as a reference to penalize our initial prediction distribution. Therefore, we update the distribution of the generated token by: + +$$ +\begin{array}{l} p _ {\text {m o d e l}} = \gamma \cdot p _ {\text {m o d e l}} (\boldsymbol {y} | \boldsymbol {X} _ {V}, \boldsymbol {X} _ {I}, \boldsymbol {X} _ {H}) \\ - (\gamma - 1) \cdot p _ {\text {m o d e l}} (\boldsymbol {y} | \boldsymbol {X} _ {I}, \boldsymbol {X} _ {H}). \\ \end{array} +$$ + +This equation effectively reduces the predicted probability based on text alone. The weight $\gamma$ is used to control the degree of penalty applied to the initial prediction distribution. + +This operation is conceptually similar to LLM-CFG [30]. Essentially, it provides a guided generation mechanism that allows the model to make informed choices between outputs based on image content and those based on language logic. This way, the model can better balance the influence of visual and textual information in its outputs, leading to more contextually accurate and relevant results. + +# 5 Experiments + +# 5.1 Setup + +Baselines. We evaluate the effectiveness of our method on three different models. To better compare the impact of image feature tokens after different projectors on our method, we selected two models that use linear projectors, LLAVA + +and Shikra, as well as one model that uses resamplers, Minigpt4. Additionally, for a more convincing comparison, we report on three decoding methods for comparison: greedy, beam search, and nucleus sample. We also selected the OPERA [13] method, which is an improvement on beam search, and the VCD [16] method, which is an improvement on nucleus sampling, to compare with our results. We used the default hyperparameters from the open-source versions of these two methods. + +Implementation Details. As different models have different lengths of image tokens, leading to different degrees of image neglect, to better align with the image sequence length of the model, we set $\alpha = 0.5$ for LLAVA, $\alpha = 0.6$ for Shikra with long image token sequence lengths and $\alpha = 0.2$ for resampler models with short image token sequences. As text inertia is independent of the image token length, we continuously use $\gamma = 1.1$ . Apart from this, in the beam search tests, the beam number is set to 5 for all methods, and in the nucleus sample tests, all the common parameters are consistent. + +# 5.2 Benchmark & Evaluation Metrics + +CHAIR [29]. Caption Hallucination Assessment with Image Relevance (CHAIR) is a widely used metric in image captioning tasks. CHAIR operates by creating a set of ground-truth object labels for each image. Any object mentioned in the caption that does not exist in the label set is considered a hallucinated object. CHAIR comprises two evaluation dimensions: instance-level and sentence-level, represented as $\mathrm{CHAIR_I}$ and $\mathrm{CHAIR_S}$ , respectively. These are calculated in the following manner: + +$$ +\mathrm {C H A I R} _ {\mathrm {I}} = \frac {| \{\text {h a l l u c i n a t e d o b j e c t s} \} |}{\text {a l l m e n t i o n e d o b j e c t s}}, \tag {5} +$$ + +$$ +\mathrm {C H A I R} _ {\mathrm {S}} = \frac {| \{\text {c a p t i o n s w i t h h a l l u c i n a t e d o b j e c t s} \} |}{\text {a l l c a p t i o n s}}. \tag {6} +$$ + +We conducted experiments based on the validation set of MSCOCO 2014. Given "Please help me describe the image in detail." as prompt, we subsequently employ the CHAIR metric to evaluate the generated description. To evaluate hallucinations in long sequence generation, we adopted the same setup as used in LURE [43] and OPERA [13], setting the max_new_tokens parameter to 512 and randomly sampled 500 instances for evaluation. + +POPE [19]. The Polling-based Object Probing Evaluation (POPE) is a evaluation metric designed in the VQA paradigm. POPE serves as a metric for assessing object hallucination, evaluating hallucinations by asking LVLMs questions such as "Is there a in the image?" Here, is replaced with the constructed ground-truth object from three different types of splits. In the "random" split, objects are randomly selected from the entire dataset for evaluation. In the "popular" split, objects are chosen from those most frequently appearing in the dataset. In the "adversarial" split, objects that are highly related to the image objects are selected for evaluation. We conduct our evaluation + +Table 1: CHAIR hallucination evaluation results on three LVLMs. CHAIR [29] is employed as the evaluation metric, where a smaller number indicates less hallucinations. OPERA is a decoding method based on beam search, VCD is a decoding method based on nucleus sampling, and PAI is an inference intervention method that can be collaborated with any decoding method. + +
DecodingMethodLLAVA [24]Minigpt4 [44]Shikra [6]
CHAIRSCHAIRICHAIRSCHAIRICHAIRSCHAIRI
GreedyVanilla46.613.432.811.151.214.4
PAI24.86.926.38.837.610.0
Beam SearchVanilla46.414.346.613.453.014.7
OPERA [13]44.614.430.19.836.812.4
PAI21.85.624.86.935.811.4
NucleusVanilla58.218.232.711.957.916.4
VCD [16]51.815.134.811.557.616.3
PAI43.414.726.710.349.913.2
+ +on the COCO dataset with 500 images, with each image having 6 questions for each split of POPE. We evaluate the performance of the model in object recognition tasks using both accuracy score and F1 score. Besides, in order to more comprehensively examine hallucinations in multi-turn chats, we have structured the evaluation in both single-turn and multi-turn dialogue forms. + +MMHal-Bench [32]. For further evaluation of our method on some challenging datasets, we choose MMHal-Bench, which is designed with 96 image-question pairs, spread across 8 question categories $\times 12$ object topics. It contains eight types of questions about object attributes, adversarial objects, comparisons, counting, spatial relations, environment, holistic descriptions, and others to comprehensively assess the model's hallucination performance on high-difficulty datasets. Essentially, it is also a VQA-based evaluation, but unlike the existence-based examination in POPE, its questions also include some logical considerations. Therefore, for evaluation on MMHal-Bench, we first need to answer questions and then use GPT-4 to score the answer based on the response and the ground-truth answer. The evaluation results include the model's scores across all question categories, and the overall score represents the average of these scores. + +GPT-4v Assisted Evaluation. To further evaluate the model's performance in image description tasks, we can move beyond the CHAIR metric, which is based on information extraction and only considers object hallucination. We can use GPT-4v for open evaluation. As with previous evaluations [13,16], we sample 50 images on the COCO dataset for evaluation. We construct prompts, input images into GPT-4v, along with the descriptions responses from two assistants. GPT-4v evaluation takes into account two dimensions: Accuracy and Detailedness, denoted respectively as C and D. Detailed prompt construction can be found in the appendix. + +Table 2: Quantitative comparison on POPE. The best results are in bold. + +
DecodingMethodLLAVA [24]Minigpt4 [44]Shikra [6]
Single-turnMulti-turnSingle-turnMulti-turnSingle-turnMulti-turn
AccF1AccF1AccF1AccF1AccF1AccF1
GreedyVanilla84.7685.5185.6984.2774.3973.7878.077.7781.5682.0578.1678.77
PAI85.8285.9787.5486.8375.276.280.8281.682.381.4778.5779.11
Beam SearchVanilla84.984.983.5781.1372.5467.7371.7867.9281.5281.8978.1878.13
PAI86.3385.8986.4285.2773.2269.3674.1571.7282.2581.0180.7280.02
NucleusVanilla80.2581.3282.8781.6357.5258.1664.4963.8979.5580.2876.5477.12
VCD [16]79.5381.0284.0383.359.9758.769.7269.3880.2280.576.3877.07
PAI81.7282.8785.0383.9759.1960.870.2470.2781.4980.7177.9978.47
+ +# 5.3 Experimental Results + +In this section, we analyze the performance of PAI across various hallucination evaluation tasks, including long image description, simplified VQA answer, construction of metric evaluation, and leveraging the near-human cognitive capabilities of GPT-4/GPT-4v as evaluation methods. For further analysis, please refer to the appendix. + +Results on long sequence hallucination evaluation. The experimental results are presented in Tab. 1. As our approach is an inference intervention method, it differs from previous decoding hallucination mitigation methods that primarily concentrate on improving a single decoding method. We have tested our method on three decoding techniques. Our method has achieved hallucination mitigation on all three decoding methods used by the three models. However, when integrated with the nucleus, a sampling-based method, the hallucination reduction brought about by our method is not significant. This may be because even though our method has increased the priority of trustful tokens, the sample set during nucleus decoding still contains many hallucination tokens. + +Moreover, while OPERA significantly mitigates hallucinations, its time efficiency is considerably higher compared to vanilla. In contrast, our method not only has almost the same time efficiency as vanilla, but it also performs better in reducing hallucination issues. Compared to VCD, during the generation process of long sequence tasks, the introduction of visual uncertainty during decoding sometimes leads to more hallucination descriptions. However, our method can reduce the proportion of hallucination words in the sample pool. + +Results on single-turn and multi-turn hallucination evaluation. Unlike the CHAIR evaluation, POPE is in a VQA format, so the response is brief, answering only "Yes" or "No". The phenomena of text inertia and image neglect may not be as noticeable under this setting, especially for single-turn. However, our method still achieved a notable improvement compared to the vanilla decoding method in single-turn. As for multi-turn evaluation, which involves a longer context, our method achieves more significant improvement as presented in Tab. 2. + +![](images/f1f43b14b4c3879f46e3ab454def14beab521b8b9cc2ecf5f8574b145088733b.jpg) +(a) LLAVA + +![](images/db8e51032d049fc4b4c221923991eefe4645b95b4e92f9ba524b5b6149be2c43.jpg) +(b) Minigpt4 +Fig. 6: Quantitative comparison on the MMHal-Bench. Higher scores indicate better performance. + +![](images/0b1940e754152c4bfb9fc2a83f6b1939375e7e116b782821e321fff8a7e32abe.jpg) +(c) Shikra + +Results on hallucination evaluation in comprehensive general scenarios. The experimental results, as shown in Fig. 6, indicate that for some more image-based question types, such as object attributes, adversarial objects, and holistic questions, the answers are more accurate when inference intervention with PAI is applied, and there is a certain degree of improvement across all models. However, for some logical questions, such as comparisons and relations, there is no noticeable improvement after intervention. In summary, through the overall metric, i.e., the average of the eight evaluation dimensions, there is a certain degree of improvement compared to the baseline after incorporating PAI. + +Results on human-like GPT-4v assisted hallucination evaluation. The experimental results, as shown in Tab. 3, indicate that even when more comprehensive hallucination evaluation dimensions are added, our method, compared to the greedy decoding method, can provide more accurate responses on all three models without losing detail in the description. Given that GPT-4v's visual understanding and language logic capabilities have reached a level close to that of humans, it can more comprehensively illustrate the performance improvements brought about by our method. + +Table 3: Results on GPT-4V evaluation. The best results are in bold. + +
MethodLLAVAMinigpt4Shikra
CDCDCD
Greedy5.625.245.85.745.545.25
PAI6.465.367.045.896.045.05
+ +# 5.4 Ablation Study + +Our method, PAI, consists of two stages of interventions. In the first stage, during forward inference, the hyperparameter $\alpha$ is utilized to set the scale of intervention. Simultaneously, the layer prior, represented as "L", is used to determine the + +Table 4: Ablation Study of the Hyperparameter $\alpha$ . When $\alpha$ becomes excessively large, resulting in an unbalanced response, we terminate the experiment and denote this with a dash (-). F1 values that are considered outliers are highlighted in red. + +
αγLLLaVAMinigpt4Shikra
CHAIRSCHAIRiF1CHAIRSCHAIRiF1CHAIRSCHAIRiF1
---46.213.875.931.610.569.456.215.874.6
0.11.1✓47.413.876.925.08.069.956.415.075.7
0.21.1✓47.413.676.921.27.770.357.215.675.2
0.41.1✓42.412.476.915.16.367.455.214.475.3
0.51.1✓24.66.874.75.42.961.551.613.275.8
0.61.1✓7.84.862.6---37.210.676.7
0.71.1✓------14.46.164.9
+ +attention layer for intervention. The second stage unfolds during the decoding process, where we mitigate text inertia by subtracting the logits distribution that results from inputs devoid of image information. In this stage, the scale is managed by the parameter $\gamma$ . + +We use LLaVA-1.5 as the representative LVLM baseline and the greedy decoding method as the basic baseline to compare the impact of our hyperparameters on the task of long sequence image description. To evaluate our method, we choose the CHAIR metric. However, since CHAIR only evaluates the hallucination problem, we have incorporated the F1 score to consider information richness and accuracy. This makes the comparison with the CHAIR metric fairer when the F1 scores are similar. In the above, the F1 scores of the various methods differ slightly, so we have not included this somewhat redundant measure. The F1 score is calculated by counting the objects included in the description, the objects in the ground-truth sets, and the hallucinated objects. + +Effects of $\alpha$ in Exciting Image Attention. In the process of exciting the attention values of image tokens, we introduce a parameter $\alpha$ to control the amplification scale. As shown in Tab. 4, different LVLMs exhibit varying sensitivity to the amplification scale. This sensitivity not only depends on the length of the model's image tokens (e.g., the image token length of LLaVA-1.5 is 576, while that of Minigpt4 is only 32), but also the original attention weights distribution. + +However, a commonality across these models is that an appropriate amplification scale can achieve a balance between the number of hallucinated objects in the description and the amount of information conveyed. If the scale is too small, the description may still contain many hallucinated objects. Conversely, if the scale is too large, the amount of information in the response will decrease. + +Effects of Layer Prior in Exciting Image Attention. We further investigate the control of the intervention layer for exciting image attention. As observed in Tab. 5, the introduction of this prior does indeed improve the performance of our method. However, when there is no control over the intervention layer and interventions are applied to all layers, the performances of different models ex + +Table 5: Ablation Study of Hyperparameter $\gamma$ and Layer Prior "L". Results are presented for the models LLAVA (left), Minigpt4 (middle), and Shikra (right). + +
αγL CHAIRSCHAIRIF1αγL CHAIRSCHAIRIF1αγL CHAIRSCHAIRIF1
--46.213.875.9--31.610.569.4--56.215.874.6
0.51.1✓24.66.875.70.21.1✓21.27.770.30.61.1✓37.2
0.51.2✓23.46.675.70.21.2✓16.46.270.30.61.2✓40.2
0.51.3✓24.67.574.40.21.3✓14.46.969.50.61.3✓40.6
0.51.5✓25.09.674.60.21.5✓12.65.467.10.61.5✓40.2
0.52.0✓24.27.174.00.22.0✓8.44.664.30.62.0✓39.4
0.51.1✓24.66.875.70.21.1✓21.27.770.30.61.1✓37.2
0.51.1×20.66.771.70.21.1×22.38.470.30.61.1×56.4
+ +hibit some variations. For Shikra, in the absence of the layer prior, our method's results regress to the baseline. For both LLaVA and Minigpt4, the loss of the layer prior causes some unfavorable fluctuations in both the CHAIR metric and the F1 score. + +Effects of $\gamma$ in Mitigating Language Prior. Tab. 5 presents the results of an ablation study focusing on $\gamma$ , which adjusts the balance between output distributions from conditioned inputs with excited image tokens and pure text inputs. Unlike the other two models, the Minigpt4 model is highly sensitive to $\gamma$ . When $\gamma$ is too large, it can lead to uncontrollable model behavior. The experimental results suggest that maintaining $\gamma$ within a relatively small range, such as 1.1 - 1.2, yields the most stable performance. + +# 6 Conclusion and Limitation + +In this paper, we first analyze the causes and manifestations of hallucinations in LVLMs. We propose a phenomenon termed "text inertia," wherein the model continues to produce the same hallucinatory descriptions even when no image input is provided. This issue fundamentally stems from the model's neglect of image tokens. Consequently, we introduce the PAI method to intervene in the model's inference process, steering it towards an image-based and trustworthy direction. This is a training-free method and does not require any external tools. Extensive experiments on multiple benchmarks and LVLMs have validated the effectiveness of PAI in mitigating hallucination issues. + +Limitations: (1) The language decoders of existing open-source LVLMs are primarily models from the LLaMA-family. It is worth exploring whether the issues of image neglect and text inertia are introduced by LLaMA. (2) As we described, our method fundamentally alleviates the image neglect issue during model inference. Its upper limit depends on the capabilities of the well-trained model. Therefore, it is worth investigating whether incorporating this issue as a loss during the training process could lead to further performance improvements. + +# Acknowledgements + +This work is supported by the National Science Foundation of China (62132017), Zhejiang Provincial Natural Science Foundation of China (LD24F020011). + +# References + +1. Alayrac, J.B., Donahue, J., Luc, P., Miech, A., Barr, I., Hasson, Y., Lenc, K., Mensch, A., Millican, K., Reynolds, M., Ring, R., Rutherford, E., Cabi, S., Han, T., Gong, Z., Samangooei, S., Monteiro, M., Menick, J., Borgeaud, S., Brock, A., Nematzadeh, A., Sharifzadeh, S., Binkowski, M., Barreira, R., Vinyals, O., Zisserman, A., Simonyan, K.: Flamingo: a visual language model for few-shot learning +2. Anisha, G., Yin, J., Bas, E.: Detecting and preventing hallucinations in large vision language models (Aug 2023) +3. Azaria, A., Mitchell, T.: The internal state of an llm knows when its lying (Apr 2023) +4. Brown, T., Mann, B., Ryder, N., Subbiah, M., Kaplan, J., Dhariwal, P., Neelakanthan, A., Shyam, P., Sastry, G., Amanda, A., Agarwal, S., Herbert-Voss, A., Krueger, G., Tom, H., Child, R., Ramesh, A., Ziegler, D., Wu, J., Winter, C., Hesse, C., Chen, M., Sigler, E., Litwin, M., Gray, S., Benjamin, C., Clark, J., Berner, C., Sam, M., Radford, A., Sutskever, I., Amodei, D.: Language models are few-shot learners. arXiv: Computation and Language, arXiv: Computation and Language (May 2020) +5. Burns, C., Ye, H., Klein, D., Steinhardt, J.: Discovering latent knowledge in language models without supervision (Dec 2022) +6. Chen, K., Zhang, Z., Zeng, W., Zhang, R., Zhu, F., Zhao, R., Research, S., Sklsde, S., Beihang, B., Seiee, S.: Shikra: Unleashing multimodal llm's referential dialogue magic +7. Chen, Z., Wu, J., Wang, W., Su, W., Chen, G., Xing, S., Muyan, Z., Zhang, Q., Zhu, X., Lu, L., Li, B., Luo, P., Lu, T., Qiao, Y., Dai, J.: Intervl: Scaling up vision foundation models and aligning for generic visual-linguistic tasks (Dec 2023) +8. Dai, W., Li, J., Li, D., Huat, A., Zhao, J., Wang, W., Li, B., Fung, P., Hoi, S.: Instructclip: Towards general-purpose vision-language models with instruction tuning +9. Darcet, T., Oquab, M., Mairal, J., Bojanowski, P.: Vision transformers need registers. arXiv preprint arXiv:2309.16588 (2023) +10. Devlin, J., Chang, M.W., Lee, K., Toutanova, K.: Bert: Pre-training of deep bidirectional transformers for language understanding. In: Proceedings of the 2019 Conference of the North (Jan 2019). https://doi.org/10.18653/v1/n19-1423, http://dx.doi.org/10.18653/v1/n19-1423 +11. Han, Y., Zhang, C., Chen, X., Yang, X., Wang, Z., Yu, G., Fu, B., Zhang, H.: Chartlama: A multimodal llm for chart understanding and generation (Nov 2023) +12. Hu, H., Zhang, J., Zhao, M., Sun, Z.: Ciem: Contrastive instruction evaluation method for better instruction tuning +13. Huang, Q., Dong, X., Zhang, P., Wang, B., He, C., Wang, J., Lin, D., Zhang, W., Yu, N.: Opera: Alleviating hallucination in multi-modal large language models via over-trust penalty and retrospection-allocation (Nov 2023) +14. Kai, J., Zhang, T., Hu, H., Lin, Z.: Sh2: Self-highlighted hesitation helps you decode more truthfully (Jan 2024) + +15. Lai, X., Tian, Z., Chen, Y., Li, Y., Yuan, Y., Liu, S., Jia, J., Kong, H., Research, M.: Lisa: Reasoning segmentation via large language model +16. Leng, S., Zhang, H., Chen, G., Li, X., Lu, S., Miao, C., Bing, L.: Mitigating object hallucinations in large vision-language models through visual contrastive decoding +17. Li, J., Li, D., Savarese, S., Hoi, S.: Blip-2: Bootstrapping language-image pretraining with frozen image encoders and large language models +18. Li, K., Patel, O., Viégas, F., Pfister, H., Wattenberg, M.: Inference-time intervention: Eliciting truthful answers from a language model. Advances in Neural Information Processing Systems 36 (2024) +19. Li, Y., Du, Y., Zhou, K., Wang, J., Zhao, W.X., Wen, J.R.: Evaluating object hallucination in large vision-language models. arXiv preprint arXiv:2305.10355 (2023) +20. Li, Z., Yang, B., Liu, Q., Ma, Z., Zhang, S., Yang, J., Sun, Y., Liu, Y., Bai, X.: Monkey: Image resolution and text label are important things for large multi-modal models (Nov 2023) +21. Liu, F., Lin, K., Li, L., Wang, J., Yacoob, Y., Wang, L.: Mitigating hallucination in large multi-modal models via robust instruction tuning. In: The Twelfth International Conference on Learning Representations (2023) +22. Liu, F., Lin, K., Li, L., Wang, J., Yacoob, Y., Wang, L.: Mitigating hallucination in large multi-modal models via robust instruction tuning (Sep 2023) +23. Liu, H., Xue, W., Chen, Y., Chen, D., Zhao, X., Wang, K., Hou, L., Li, R., Peng, W.: A survey on hallucination in large vision-language models +24. Liu, H., Li, C., Li, Y., Lee, Y.J.: Improved baselines with visual instruction tuning. arXiv preprint arXiv:2310.03744 (2023) +25. Liu, H., Li, C., Wu, Q., Lee, Y., Madison, M., Research, M.: Visual instruction tuning +26. Mishra, S., Khashabi, D., Baral, C., Hajishirzi, H.: Cross-task generalization via natural language crowdsourcing instructions. In: Proceedings of the 60th Annual Meeting of the Association for Computational Linguistics (Volume 1: Long Papers) (Jan 2022). https://doi.org/10.18653/v1/2022.acl-long.244, http://dx.doi.org/10.18653/v1/2022.acl-long.244 +27. Ouyang, L., Wu, J., Jiang, X., Almeida, D., Wainwright, C.L., Mishkin, P., Zhang, C., Agarwal, S., Slama, K., Ray, A., Schulman, J., Hilton, J., Kelton, F., Miller, L., Simens, M., Askell, A., Welinder, P., Christiano, P.F., Leike, J., Lowe, R.: Training language models to follow instructions with human feedback (2022) +28. Peng, B., Li, C., He, P., Galley, M., Gao, J.: Instruction tuning with gpt-4 +29. Rohrbach, A., Hendricks, L.A., Burns, K., Darrell, T., Saenko, K.: Object hallucination in image captioning. In: Proceedings of the 2018 Conference on Empirical Methods in Natural Language Processing (Jan 2018). https://doi.org/10.18653/v1/d18-1437, http://dx.doi.org/10.18653/v1/d18-1437 +30. Sanchez, G., Fan, H., Spangher, A., Levi, E., Ammanamanchi, P.S., Biderman, S.: Stay on topic with classifier-free guidance. arXiv preprint arXiv:2306.17806 (2023) +31. Sun, Z., Shen, S., Cao, S., Liu, H., Li, C., Shen, Y., Gan, C., Gui, L.Y., Wang, Y.X., Yang, Y., Keutzer, K., Darrell, T.: Aligning large multimodal models with factually augmented rlhf (Sep 2023) +32. Sun, Z., Shen, S., Cao, S., Liu, H., Li, C., Shen, Y., Gan, C., Gui, L.Y., Wang, Y.X., Yang, Y., et al.: Aligning large multimodal models with factually augmented rlhf. arXiv preprint arXiv:2309.14525 (2023) +33. Touvron, H., Lavril, T., Izacard, G., Martinet, X., Lachaux, M.A., Lacroix, T., Rozi'ere, B., Goyal, N., Hambro, E., Azhar, F., Rodriguez, A., Joulin, A., Grave, E., Lample, G.: Llama: Open and efficient foundation language models + +34. Turner, A., Thiergart, L., Udell, D., Leech, G., Mini, U., MacDiarmid, M.: Activation addition: Steering language models without optimization. arXiv preprint arXiv:2308.10248 (2023) +35. Wang, L., Li, L., Dai, D., Chen, D., Zhou, H., Meng, F., Zhou, J., Sun, X.: Label words are anchors: An information flow perspective for understanding in-context learning. arXiv preprint arXiv:2305.14160 (2023) +36. Wei, F., Zhang, X., Zhang, A., Zhang, B., Chu, X.: Lenna: Language enhanced reasoning detection assistant (Dec 2023) +37. Wei, J., Bosma, M., Zhao, V., Guu, K., Yu, A., Lester, B., Du, N., Dai, A., Le, Q.: Finetuned language models are zero-shot learners. Learning, Learning (Sep 2021) +38. Xiao, G., Tian, Y., Chen, B., Han, S., Lewis, M.: Efficient streaming language models with attention sinks (Sep 2023) +39. Ye, Q., Xu, H., Xu, G., Ye, J., Yan, M., Zhou, Y., Wang, J., Hu, A., Shi, P., Shi, Y., Li, C., Xu, Y., Chen, H., Tian, J., Qi, Q., Zhang, J., Huang, F.: mplug-owl: Modularization empowers large language models with multimodality +40. Yin, S., Fu, C., Zhao, S., Xu, T., Wang, H., Sui, D., Shen, Y., Li, K., Sun, X., Chen, E.: Woodpecker: Hallucination correction for multimodal large language models +41. You, H., Zhang, H., Gan, Z., Du, X., Zhang, B., Wang, Z., Cao, L., Chang, S.F., Yang, Y.: Ferret: Refer and ground anything anywhere at any granularity (Oct 2023) +42. Zhang, Y., Qian, S., Peng, B., Liu, S., Jia, J.: Prompt highlighter: Interactive control for multi-modal llms. arXiv preprint arXiv:2312.04302 (2023) +43. Zhou, Y., Cui, C., Yoon, J., Zhang, L., Deng, Z., Finn, C., Bansal, M., Yao, H.: Analyzing and mitigating object hallucination in large vision-language models (Oct 2023) +44. Zhu, D., Chen, J., Shen, X., Li, X., Elhoseiny, M.: Minigpt-4: Enhancing vision-language understanding with advanced large language models \ No newline at end of file diff --git a/2024/Paying More Attention to Images_ A Training-Free Method for Alleviating Hallucination in LVLMs/images.zip b/2024/Paying More Attention to Images_ A Training-Free Method for Alleviating Hallucination in LVLMs/images.zip new file mode 100644 index 0000000000000000000000000000000000000000..c9ee632f21520b8d1da8ffbecf184b75a954fa56 --- /dev/null +++ b/2024/Paying More Attention to Images_ A Training-Free Method for Alleviating Hallucination in LVLMs/images.zip @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:cf7c7afd88ff1f9b0dd7df04f72e1f735d8a773006c7fed2e119ce975db2c3d4 +size 496631 diff --git a/2024/Paying More Attention to Images_ A Training-Free Method for Alleviating Hallucination in LVLMs/layout.json b/2024/Paying More Attention to Images_ A Training-Free Method for Alleviating Hallucination in LVLMs/layout.json new file mode 100644 index 0000000000000000000000000000000000000000..2521cd8661e6006c8ffa25125651dd82268cb99f --- /dev/null +++ b/2024/Paying More Attention to Images_ A Training-Free Method for Alleviating Hallucination in LVLMs/layout.json @@ -0,0 +1,7970 @@ +{ + "pdf_info": [ + { + "para_blocks": [ + { + "bbox": [ + 133, + 111, + 481, + 148 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 111, + 481, + 148 + ], + "spans": [ + { + "bbox": [ + 133, + 111, + 481, + 148 + ], + "type": "text", + "content": "Paying More Attention to Image: A Training-Free Method for Alleviating Hallucination in LVLMs" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 204, + 167, + 408, + 180 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 204, + 167, + 408, + 180 + ], + "spans": [ + { + "bbox": [ + 204, + 167, + 408, + 180 + ], + "type": "text", + "content": "Shi Liu" + }, + { + "bbox": [ + 204, + 167, + 408, + 180 + ], + "type": "inline_equation", + "content": "^{1}" + }, + { + "bbox": [ + 204, + 167, + 408, + 180 + ], + "type": "text", + "content": ", Kecheng Zheng" + }, + { + "bbox": [ + 204, + 167, + 408, + 180 + ], + "type": "inline_equation", + "content": "^{1,2\\dagger}" + }, + { + "bbox": [ + 204, + 167, + 408, + 180 + ], + "type": "text", + "content": ", and Wei Chen" + }, + { + "bbox": [ + 204, + 167, + 408, + 180 + ], + "type": "inline_equation", + "content": "^{1\\dagger}" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 201, + 189, + 411, + 201 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 201, + 189, + 411, + 201 + ], + "spans": [ + { + "bbox": [ + 201, + 189, + 411, + 201 + ], + "type": "inline_equation", + "content": "^{1}" + }, + { + "bbox": [ + 201, + 189, + 411, + 201 + ], + "type": "text", + "content": " State Key Lab of CAD&CG, Zhejiang University" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 279, + 201, + 334, + 212 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 279, + 201, + 334, + 212 + ], + "spans": [ + { + "bbox": [ + 279, + 201, + 334, + 212 + ], + "type": "text", + "content": "2 Ant Group" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 228, + 213, + 386, + 223 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 228, + 213, + 386, + 223 + ], + "spans": [ + { + "bbox": [ + 228, + 213, + 386, + 223 + ], + "type": "text", + "content": "{liushi0927,zkechengzk}@gmail.com" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 264, + 224, + 350, + 234 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 264, + 224, + 350, + 234 + ], + "spans": [ + { + "bbox": [ + 264, + 224, + 350, + 234 + ], + "type": "text", + "content": "chenvis@zju.edu.cn" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 159, + 258, + 455, + 491 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 159, + 258, + 455, + 491 + ], + "spans": [ + { + "bbox": [ + 159, + 258, + 455, + 491 + ], + "type": "text", + "content": "Abstract. Existing Large Vision-Language Models (LVLMs) primarily align image features of vision encoder with Large Language Models (LLMs) to leverage their superior text generation capabilities. However, the scale disparity between vision encoder and language model may led to LLMs assuming a predominant role in multi-modal comprehension. This imbalance in LVLMs may result in the instances of hallucinatory. Concretely, LVLMs may generate consistent descriptions with or without visual input, indicating that certain outputs are influenced solely by context text. We refer to this phenomenon as \"text inertia.\" To counteract this issue, we introduce a training-free algorithm to find an equilibrium point between image comprehension and language inference. Specifically, we adaptively involve adjusting and amplifying the attention weights assigned to image tokens, thereby granting greater prominence to visual elements. Meanwhile, we subtract the logits of multi-modal inputs from ones of pure text input, which can help LVLMs be not biased towards LLMs. By enhancing images tokens and reducing the stubborn output of LLM, we can let LVLM pay more attention to images, towards alleviating text inertia and reducing the hallucination in LVLMs. Our extensive experiments show that this method substantially reduces the frequency of hallucinatory outputs in various LVLMs in terms of different metrics. Project page is available at https://1albj.github.io/projects/PAI/" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 159, + 499, + 453, + 521 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 159, + 499, + 453, + 521 + ], + "spans": [ + { + "bbox": [ + 159, + 499, + 453, + 521 + ], + "type": "text", + "content": "Keywords: Vision-Language Models " + }, + { + "bbox": [ + 159, + 499, + 453, + 521 + ], + "type": "inline_equation", + "content": "\\cdot" + }, + { + "bbox": [ + 159, + 499, + 453, + 521 + ], + "type": "text", + "content": " Visual Dialogs " + }, + { + "bbox": [ + 159, + 499, + 453, + 521 + ], + "type": "inline_equation", + "content": "\\cdot" + }, + { + "bbox": [ + 159, + 499, + 453, + 521 + ], + "type": "text", + "content": " Hallucination Mitigation" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 132, + 540, + 230, + 553 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 540, + 230, + 553 + ], + "spans": [ + { + "bbox": [ + 132, + 540, + 230, + 553 + ], + "type": "text", + "content": "1 Introduction" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 130, + 563, + 482, + 624 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 563, + 482, + 624 + ], + "spans": [ + { + "bbox": [ + 130, + 563, + 482, + 624 + ], + "type": "text", + "content": "Recently, Large Vision-Language Models (LVLMs) have made significant strides, exhibiting impressive capabilities across a multitude of tasks [11,15,36,42]. However, these models still struggle with the phenomenon of hallucination. Specifically, there is often a mismatch between the textual content generated by the model and the actual visual input it receives [23]." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 130, + 624, + 481, + 648 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 624, + 481, + 648 + ], + "spans": [ + { + "bbox": [ + 130, + 624, + 481, + 648 + ], + "type": "text", + "content": "Hallucination in LVLMs is often attributed to issues with modality alignment, leading to the development of mitigating strategies through alignment" + } + ] + } + ], + "index": 10 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 141, + 653, + 247, + 666 + ], + "type": "page_footnote", + "angle": 0, + "lines": [ + { + "bbox": [ + 141, + 653, + 247, + 666 + ], + "spans": [ + { + "bbox": [ + 141, + 653, + 247, + 666 + ], + "type": "text", + "content": "† Corresponding authors." + } + ] + } + ], + "index": 11 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 0 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 138, + 116, + 478, + 244 + ], + "blocks": [ + { + "bbox": [ + 138, + 116, + 478, + 244 + ], + "lines": [ + { + "bbox": [ + 138, + 116, + 478, + 244 + ], + "spans": [ + { + "bbox": [ + 138, + 116, + 478, + 244 + ], + "type": "image", + "image_path": "917107f1f4568805fcb7915f42da9f1fe4a850bc295b6797be7ad9aaba2190b8.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 130, + 253, + 482, + 342 + ], + "lines": [ + { + "bbox": [ + 130, + 253, + 482, + 342 + ], + "spans": [ + { + "bbox": [ + 130, + 253, + 482, + 342 + ], + "type": "text", + "content": "Fig. 1: We present an examination of various input settings, with hallucinations specifically highlighted in red. (a) When using LLAVA for image description, it generates a hallucinated description. (b) Even without image input, when only the historical response preceding the hallucinated description is input to LLAVA, it reproduces the same hallucinated description, a phenomenon we refer to as \"Text Inertia\". (c) Our proposed method, PAI, effectively mitigates this text inertia problem and yields accurate descriptions. (d) Utilizing PAI for image description results in a significantly more precise description." + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_caption" + } + ], + "index": 2 + }, + { + "type": "image", + "bbox": [ + 134, + 363, + 300, + 463 + ], + "blocks": [ + { + "bbox": [ + 134, + 363, + 300, + 463 + ], + "lines": [ + { + "bbox": [ + 134, + 363, + 300, + 463 + ], + "spans": [ + { + "bbox": [ + 134, + 363, + 300, + 463 + ], + "type": "image", + "image_path": "8ac3098831f766ebecc7681dddac6a072a8bce51a13fa9a9862ff6938a42c29e.jpg" + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 130, + 475, + 305, + 529 + ], + "lines": [ + { + "bbox": [ + 130, + 475, + 305, + 529 + ], + "spans": [ + { + "bbox": [ + 130, + 475, + 305, + 529 + ], + "type": "text", + "content": "Fig.2: Percentage of text inertia hallucination in all hallucination (calculated with 500 samples). For specific calculation processes, please refer to the supplementary material section A." + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 307, + 431, + 482, + 529 + ], + "lines": [ + { + "bbox": [ + 307, + 431, + 482, + 529 + ], + "spans": [ + { + "bbox": [ + 307, + 431, + 482, + 529 + ], + "type": "text", + "content": "Fig. 3: Visualization of the average attention ratio of different content. The x-axis denotes the sequence length of the history tokens. The lengths of the image, instruction, BOS, and outlier tokens [9] are all fixed as they are part of the model input, with 576 tokens for the image, 21 for the instruction, and 1 each for the BOS and outlier tokens." + } + ] + } + ], + "index": 7, + "angle": 0, + "type": "image_caption" + } + ], + "index": 4 + }, + { + "type": "image", + "bbox": [ + 311, + 355, + 479, + 422 + ], + "blocks": [ + { + "bbox": [ + 311, + 355, + 479, + 422 + ], + "lines": [ + { + "bbox": [ + 311, + 355, + 479, + 422 + ], + "spans": [ + { + "bbox": [ + 311, + 355, + 479, + 422 + ], + "type": "image", + "image_path": "a2dd0958982eb6cd370c48df41b6d22f3e71098f63f8931d166bf9fc1e3a3984.jpg" + } + ] + } + ], + "index": 6, + "angle": 0, + "type": "image_body" + } + ], + "index": 6 + }, + { + "bbox": [ + 130, + 556, + 482, + 628 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 556, + 482, + 628 + ], + "spans": [ + { + "bbox": [ + 130, + 556, + 482, + 628 + ], + "type": "text", + "content": "training optimization [7, 20, 31]. However, is hallucination in LVLMs merely a result of the model's capacity, and can it only be alleviated through additional training? We propose a scenario where LVLMs generate a hallucinated object description. Specifically, even when the image input is removed and only the generated text preceding the hallucinated object word is retained, the LVLMs persist in producing the same hallucinated description, as depicted in Fig. 1." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 130, + 630, + 482, + 665 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 630, + 482, + 665 + ], + "spans": [ + { + "bbox": [ + 130, + 630, + 482, + 665 + ], + "type": "text", + "content": "In order to empirically investigate this behavior, we conducted tests on three LVLMs within the context of image describing tasks on the COCO dataset. We identified and conducted a statistical analysis on instances where LVLMs" + } + ] + } + ], + "index": 9 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "spans": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "type": "text", + "content": "2" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 166, + 91, + 205, + 100 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 166, + 91, + 205, + 100 + ], + "spans": [ + { + "bbox": [ + 166, + 91, + 205, + 100 + ], + "type": "text", + "content": "Liu et al." + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 1 + }, + { + "para_blocks": [ + { + "bbox": [ + 130, + 116, + 479, + 163 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 116, + 479, + 163 + ], + "spans": [ + { + "bbox": [ + 130, + 116, + 479, + 163 + ], + "type": "text", + "content": "generated identical hallucinated object descriptions, even when the input was exclusively historical response text without any image. The observation from Fig. 2 clearly indicates that, even with the application of rigorous identification settings, the phenomenon continues to represent a substantial proportion." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 130, + 164, + 480, + 295 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 164, + 480, + 295 + ], + "spans": [ + { + "bbox": [ + 130, + 164, + 480, + 295 + ], + "type": "text", + "content": "We referred this phenomenon as \"Text Inertia\". Our hypothesis is that text inertia arises due to the current generative paradigms mapping image representations onto the text representation space as text tokens. In this mechanism, the LLM becomes the dominant character, and the inference process lacks additional handling of image tokens, leading to their neglect during the generation process. To validate this hypothesis, we have analyzed the attention values ratios of the LLaVA model during the inference process in Fig. 3. Our findings show that despite image tokens occupying a significant proportion, they do not receive substantial attention under the current mechanism. This multimodal chat resembles more of an automatic completion based on context rather than a continuous attention to the image for completion." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 130, + 296, + 481, + 474 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 296, + 481, + 474 + ], + "spans": [ + { + "bbox": [ + 130, + 296, + 481, + 474 + ], + "type": "text", + "content": "To close this gap, we introduce a method referred to as Pay Attention to Image (PAI). At a high level, PAI intervenes in the inference process to make it more image-centric, following the original image perception direction. To achieve this, we focus on the self-attention heads in the decoder layers of LVLMs. We enhance the attention weights for image tokens in their original directions during inference. This allows us to use the updated attention matrix to calculate the hidden states for the generated token, thereby incorporating more consideration for image representation during the generation process. To further mitigate text inertia, we construct the input using instruction tokens and historical response tokens, and subtract the model logits of this input from the logits of original model with image tokens input. This strategy helps to reduce the influence of language priors during the generation process. Unlike previous methods for mitigating hallucination that require additional training or external tools, our approach is training-free. Moreover, we are the first to propose an inference intervention method for mitigating hallucination in LVLMs." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 130, + 475, + 481, + 582 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 475, + 481, + 582 + ], + "spans": [ + { + "bbox": [ + 130, + 475, + 481, + 582 + ], + "type": "text", + "content": "We then evaluate the response accuracy in the image description task from long sequence generation perspective by employing the CHAIR metric [29] and GPT-4V. In addition, we use POPE [19] and MMHal-Bench [32] to more comprehensively evaluate the model's hallucination performance on VQA task. Furthermore, we have constructed single-turn and multi-turn chat evaluations for POPE. Since our model intervenes in the inference process, it can be used for any decoding method. Therefore, we conducted experiments on three decoding methods of the three models. The experimental results proved the effectiveness of our method in mitigating hallucinations." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 132, + 599, + 237, + 611 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 599, + 237, + 611 + ], + "spans": [ + { + "bbox": [ + 132, + 599, + 237, + 611 + ], + "type": "text", + "content": "2 Related Work" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 132, + 623, + 316, + 635 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 623, + 316, + 635 + ], + "spans": [ + { + "bbox": [ + 132, + 623, + 316, + 635 + ], + "type": "text", + "content": "2.1 Large Vision-Language Models" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 130, + 641, + 479, + 665 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 641, + 479, + 665 + ], + "spans": [ + { + "bbox": [ + 130, + 641, + 479, + 665 + ], + "type": "text", + "content": "The development of pre-training techniques [4, 27] and instruction tuning techniques [26, 37] has rapidly advanced LLMs technology, such as LLaMA [33] and" + } + ] + } + ], + "index": 8 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 430, + 91, + 447, + 100 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 430, + 91, + 447, + 100 + ], + "spans": [ + { + "bbox": [ + 430, + 91, + 447, + 100 + ], + "type": "text", + "content": "PAI" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 474, + 91, + 481, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 474, + 91, + 481, + 100 + ], + "spans": [ + { + "bbox": [ + 474, + 91, + 481, + 100 + ], + "type": "text", + "content": "3" + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 2 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 140, + 124, + 479, + 224 + ], + "blocks": [ + { + "bbox": [ + 140, + 124, + 479, + 224 + ], + "lines": [ + { + "bbox": [ + 140, + 124, + 479, + 224 + ], + "spans": [ + { + "bbox": [ + 140, + 124, + 479, + 224 + ], + "type": "image", + "image_path": "fadb7576ac2c6adb7a89282889adfa54b732b07b7997bf7c889672fcdf9f7d64.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 130, + 251, + 482, + 307 + ], + "lines": [ + { + "bbox": [ + 130, + 251, + 482, + 307 + ], + "spans": [ + { + "bbox": [ + 130, + 251, + 482, + 307 + ], + "type": "text", + "content": "Fig. 4: Architecture of our PAI. To alleviate text inertia, we additionally construct an input without image. Throughout the forward inference process, we amplify the focus on the image token by edit self-attention maps in LLaMA. Ultimately, we subtract the logits distribution of the language prior during decoding to achieve an accurate description." + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_caption" + } + ], + "index": 2 + }, + { + "bbox": [ + 130, + 330, + 482, + 450 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 330, + 482, + 450 + ], + "spans": [ + { + "bbox": [ + 130, + 330, + 482, + 450 + ], + "type": "text", + "content": "Vicuna [28], further leading to the prosperity of LVLMs technology. Early works, like Flamingo [1] and BLIP-2 [17], have successfully adapted LLMs to visual tasks, demonstrating notable generative capabilities and in-context learning abilities. Recently, the capabilities of LVLMs have further advanced under the influence of visual instruction tuning techniques [24,25]. Using different projectors to map images to the text domain, thereby endowing language generation models with image understanding capabilities, is also a hot research topic [8,25,39,44]. Additionally, several studies focus on visual language tasks such as grounding capabilities [6] and reasoning capabilities [15]. However, recent LVLMs still face the issue of hallucination generation [21]." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 131, + 467, + 339, + 479 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 131, + 467, + 339, + 479 + ], + "spans": [ + { + "bbox": [ + 131, + 467, + 339, + 479 + ], + "type": "text", + "content": "2.2 Mitigation of LVLMs Hallucination" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 130, + 486, + 482, + 592 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 486, + 482, + 592 + ], + "spans": [ + { + "bbox": [ + 130, + 486, + 482, + 592 + ], + "type": "text", + "content": "Hallucination in LVLMs refers to contradictions between the image input and the textual output. Various methods have been proposed to mitigate hallucination. The most direct reason for hallucination generation is that hallucination arises from data bias and knowledge gaps between vision and language. Therefore, better data filtering methods [12, 22, 41] and higher quality annotated data [2] are introduced. Simultaneously, these methods also imply the need for more alignment training [31] or adjustments in the model architecture [7, 20]. These methods can achieve good results, but they are time-consuming and require high computational resources." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 130, + 594, + 482, + 666 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 594, + 482, + 666 + ], + "spans": [ + { + "bbox": [ + 130, + 594, + 482, + 666 + ], + "type": "text", + "content": "Apart from addressing the ability of LVLMs itself, hallucination can also be mitigated through post-processing methods. This method usually involves using an additional module or external tools to edit the response. Recent methods such as LURE [43] utilize additional data to train a state detector and when hallucination issues are detected, content is regenerated by a revisor model. Woodpecker [40] introduces an external visual model to inspect entities extracted from" + } + ] + } + ], + "index": 7 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "spans": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "type": "text", + "content": "4" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 166, + 91, + 205, + 100 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 166, + 91, + 205, + 100 + ], + "spans": [ + { + "bbox": [ + 166, + 91, + 205, + 100 + ], + "type": "text", + "content": "Liu et al." + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 3 + }, + { + "para_blocks": [ + { + "bbox": [ + 130, + 116, + 479, + 151 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 116, + 479, + 151 + ], + "spans": [ + { + "bbox": [ + 130, + 116, + 479, + 151 + ], + "type": "text", + "content": "the response, and then the detection results are handed over to the generation model to regenerate better answers. These methods also extend the inference chain and increase inference costs." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 130, + 152, + 482, + 248 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 152, + 482, + 248 + ], + "spans": [ + { + "bbox": [ + 130, + 152, + 482, + 248 + ], + "type": "text", + "content": "Training-free hallucination mitigation methods have so far only been attempted in decoding methods. OPERA [13] discovered an abnormal attention pattern that accompanies model decoding. It was statistically found that this pattern often accompanies hallucination descriptions, and thus a detection and mitigation method was proposed based on this pattern to alleviate the hallucination faced by the model. VCD [16] introduced the notion that visual uncertainty increases hallucination descriptions and, based on this discovery, proposed a contrast decoding method to alleviate hallucination issues." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 132, + 266, + 234, + 278 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 266, + 234, + 278 + ], + "spans": [ + { + "bbox": [ + 132, + 266, + 234, + 278 + ], + "type": "text", + "content": "3 Preliminaries" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 130, + 290, + 482, + 386 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 290, + 482, + 386 + ], + "spans": [ + { + "bbox": [ + 130, + 290, + 482, + 386 + ], + "type": "text", + "content": "The architecture of LVLMs typically comprises of three main components: an image encoder, a projector, and a language decoder. Both the image encoder and language decoder are usually pre-trained. The image encoder is employed to transform images into image tokens, which are subsequently mapped to the text representation space by the projector. This process enables the concatenation of image tokens with text tokens that are then fed into the language decoder. The language decoder subsequently generates corresponding responses based on the provided instructions." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 130, + 387, + 482, + 518 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 387, + 482, + 518 + ], + "spans": [ + { + "bbox": [ + 130, + 387, + 482, + 518 + ], + "type": "text", + "content": "The existing projectors. Currently, projectors predominantly fall into two categories: linear projectors and resamplers. A projector takes N visual features from the image encoder and transforms them into M visual tokens. The linear projector employs a multilayer perceptron to transform all visual features, maintaining a one-to-one transformation which meaning that M equals N. In contrast, the resampler does not preserve all visual features but instead samples visual cues (M, where " + }, + { + "bbox": [ + 130, + 387, + 482, + 518 + ], + "type": "inline_equation", + "content": "\\mathrm{M} < \\mathrm{N}" + }, + { + "bbox": [ + 130, + 387, + 482, + 518 + ], + "type": "text", + "content": "). For instance, Q-former [17] utilizes M learnable queries and Bert [10] to extract information from visual features. Given that the knowledge of images during the generation process solely originates from the output image tokens of the projectors, our attention is concentrated on the image tokens post-projection, irrespective of their preceding modeling process." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 130, + 518, + 482, + 591 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 518, + 482, + 591 + ], + "spans": [ + { + "bbox": [ + 130, + 518, + 482, + 591 + ], + "type": "text", + "content": "Autoregressive language decoders. Nearly all LVLMs adopt LLaMA-family models as their language decoders, which employ the self-attention mechanism. The visual tokens processed by the projector are concatenated with text tokens and fed into the LLaMA, which carries out the forward decoding process. From the perspective of a single attention head in a single layer, each head repeatedly performs the following attention operation with the same input shape:" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 216, + 599, + 481, + 626 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 216, + 599, + 481, + 626 + ], + "spans": [ + { + "bbox": [ + 216, + 599, + 481, + 626 + ], + "type": "interline_equation", + "content": "\\boldsymbol {O} _ {h} = \\boldsymbol {A} _ {h} \\boldsymbol {V} _ {h}, \\quad \\boldsymbol {A} _ {h} = \\operatorname {s o f t m a x} \\left(\\frac {\\boldsymbol {Q} _ {h} \\boldsymbol {K} _ {h} ^ {\\top}}{\\sqrt {d _ {k}}}\\right). \\tag {1}", + "image_path": "ffec5ad7d3f221a1b4b2b06d2e296350fd5eac057ac852273d5b8c646271d653.jpg" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 130, + 629, + 482, + 666 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 629, + 482, + 666 + ], + "spans": [ + { + "bbox": [ + 130, + 629, + 482, + 666 + ], + "type": "text", + "content": "Each attention head " + }, + { + "bbox": [ + 130, + 629, + 482, + 666 + ], + "type": "inline_equation", + "content": "h" + }, + { + "bbox": [ + 130, + 629, + 482, + 666 + ], + "type": "text", + "content": " performs an attention operation using its own set of queries " + }, + { + "bbox": [ + 130, + 629, + 482, + 666 + ], + "type": "inline_equation", + "content": "Q_{h} \\in \\mathbb{R}^{n \\times d_{k}}" + }, + { + "bbox": [ + 130, + 629, + 482, + 666 + ], + "type": "text", + "content": ", keys " + }, + { + "bbox": [ + 130, + 629, + 482, + 666 + ], + "type": "inline_equation", + "content": "K_{h} \\in \\mathbb{R}^{n \\times d_{k}}" + }, + { + "bbox": [ + 130, + 629, + 482, + 666 + ], + "type": "text", + "content": ", and values " + }, + { + "bbox": [ + 130, + 629, + 482, + 666 + ], + "type": "inline_equation", + "content": "V_{h} \\in \\mathbb{R}^{n \\times d_{k}}" + }, + { + "bbox": [ + 130, + 629, + 482, + 666 + ], + "type": "text", + "content": ", where " + }, + { + "bbox": [ + 130, + 629, + 482, + 666 + ], + "type": "inline_equation", + "content": "n" + }, + { + "bbox": [ + 130, + 629, + 482, + 666 + ], + "type": "text", + "content": " represents the sequence length and " + }, + { + "bbox": [ + 130, + 629, + 482, + 666 + ], + "type": "inline_equation", + "content": "d_{k}" + }, + { + "bbox": [ + 130, + 629, + 482, + 666 + ], + "type": "text", + "content": " represents the hidden dimensions. The" + } + ] + } + ], + "index": 9 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 430, + 91, + 447, + 100 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 430, + 91, + 447, + 100 + ], + "spans": [ + { + "bbox": [ + 430, + 91, + 447, + 100 + ], + "type": "text", + "content": "PAI" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 474, + 91, + 481, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 474, + 91, + 481, + 100 + ], + "spans": [ + { + "bbox": [ + 474, + 91, + 481, + 100 + ], + "type": "text", + "content": "5" + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 4 + }, + { + "para_blocks": [ + { + "bbox": [ + 130, + 116, + 482, + 224 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 116, + 482, + 224 + ], + "spans": [ + { + "bbox": [ + 130, + 116, + 482, + 224 + ], + "type": "text", + "content": "output " + }, + { + "bbox": [ + 130, + 116, + 482, + 224 + ], + "type": "inline_equation", + "content": "O_{h} \\in \\mathbb{R}^{n \\times d_{k}}" + }, + { + "bbox": [ + 130, + 116, + 482, + 224 + ], + "type": "text", + "content": " is modeled by multiplying " + }, + { + "bbox": [ + 130, + 116, + 482, + 224 + ], + "type": "inline_equation", + "content": "V_{h}" + }, + { + "bbox": [ + 130, + 116, + 482, + 224 + ], + "type": "text", + "content": " and the attention weights " + }, + { + "bbox": [ + 130, + 116, + 482, + 224 + ], + "type": "inline_equation", + "content": "A_{h} \\in \\mathbb{R}^{n \\times n}" + }, + { + "bbox": [ + 130, + 116, + 482, + 224 + ], + "type": "text", + "content": ", where each row represents the weights for each token during feature mixing. This operation enables the model to focus on different parts of the input for each head through attention weights, thereby capturing various parts of the information from the sequence token representations. The final output is the current generated token vocabulary conditional probability distribution " + }, + { + "bbox": [ + 130, + 116, + 482, + 224 + ], + "type": "inline_equation", + "content": "y \\in \\mathbb{R}^{v}" + }, + { + "bbox": [ + 130, + 116, + 482, + 224 + ], + "type": "text", + "content": " based on the input instruction representations " + }, + { + "bbox": [ + 130, + 116, + 482, + 224 + ], + "type": "inline_equation", + "content": "X_{I}" + }, + { + "bbox": [ + 130, + 116, + 482, + 224 + ], + "type": "text", + "content": ", image representations " + }, + { + "bbox": [ + 130, + 116, + 482, + 224 + ], + "type": "inline_equation", + "content": "X_{V}" + }, + { + "bbox": [ + 130, + 116, + 482, + 224 + ], + "type": "text", + "content": " and history generated token representations " + }, + { + "bbox": [ + 130, + 116, + 482, + 224 + ], + "type": "inline_equation", + "content": "X_{H}" + }, + { + "bbox": [ + 130, + 116, + 482, + 224 + ], + "type": "text", + "content": ", where " + }, + { + "bbox": [ + 130, + 116, + 482, + 224 + ], + "type": "inline_equation", + "content": "v" + }, + { + "bbox": [ + 130, + 116, + 482, + 224 + ], + "type": "text", + "content": " is the size of the vocabulary. This process can be formatted as:" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 214, + 232, + 480, + 259 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 214, + 232, + 480, + 259 + ], + "spans": [ + { + "bbox": [ + 214, + 232, + 480, + 259 + ], + "type": "interline_equation", + "content": "\\begin{array}{l} \\boldsymbol {y} \\sim p _ {\\text {m o d e l}} (\\boldsymbol {y} \\mid \\mathbf {X} _ {I}, \\mathbf {X} _ {V}, \\mathbf {X} _ {H}), \\\\ \\begin{array}{l} p _ {\\text {m o d e r}} (\\boldsymbol {y} \\mid 1, 1, 1, 1, 1, 1, 1, 1, 1), \\\\ \\propto \\operatorname {s o f t m a x} \\left(\\operatorname {l o g i t} _ {\\text {m o d e l}} (\\boldsymbol {y} \\mid \\mathbf {X} _ {I}, \\mathbf {X} _ {V}, \\mathbf {X} _ {H})\\right), \\end{array} \\tag {2} \\\\ \\end{array}", + "image_path": "79fffdcfcb8de48b1231e7987b5a5d2221cb949e21eda43473c931fc0d689873.jpg" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 130, + 266, + 480, + 304 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 266, + 480, + 304 + ], + "spans": [ + { + "bbox": [ + 130, + 266, + 480, + 304 + ], + "type": "text", + "content": "which calculates the distribution of one token and iterates for the entire response. The sequence generation continues until an EOS (End of Sentence) token is produced, marking the end of the generation and resulting in a complete response." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 131, + 312, + 202, + 324 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 131, + 312, + 202, + 324 + ], + "spans": [ + { + "bbox": [ + 131, + 312, + 202, + 324 + ], + "type": "text", + "content": "4 Method" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 130, + 331, + 482, + 450 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 331, + 482, + 450 + ], + "spans": [ + { + "bbox": [ + 130, + 331, + 482, + 450 + ], + "type": "text", + "content": "At the core of our method is a solution for image neglect and text inertia, both of which are fundamentally interconnected. Essentially, as paying more attention to the image, there is a corresponding reduction in the reliance on language priors. Intuitively, in a conversation centered around an image, the model should devote more attention to the image, thereby allowing it to have a significant impact on the response. As such, we identify the self-attention map in the token-level generation and augment the image attention in its original directions. This strategy promotes a more image-centric latent representations. Additionally, to further mitigate the influence of text inertia, we devide the logits distribution of pure text input into the model's output." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 131, + 467, + 311, + 479 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 131, + 467, + 311, + 479 + ], + "spans": [ + { + "bbox": [ + 131, + 467, + 311, + 479 + ], + "type": "text", + "content": "4.1 Pay More Attention to Image" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 130, + 486, + 482, + 581 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 486, + 482, + 581 + ], + "spans": [ + { + "bbox": [ + 130, + 486, + 482, + 581 + ], + "type": "text", + "content": "Extracting the Self-Attention Matrix. We start from a token-level perspective. The response process in LVLMs is fundamentally generated token by token. Each token is generated based on the input image, instruction, and the historically generated response. This process is facilitated through a multi-layer attention decoder architecture. Consequently, this results in a probability distribution of the vocabulary for the currently generated token. Our goal is to extract the attention matrix of each attention head at every layer, indicating the influence of each content during inference." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 130, + 582, + 482, + 667 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 582, + 482, + 667 + ], + "spans": [ + { + "bbox": [ + 130, + 582, + 482, + 667 + ], + "type": "text", + "content": "When generating the " + }, + { + "bbox": [ + 130, + 582, + 482, + 667 + ], + "type": "inline_equation", + "content": "k" + }, + { + "bbox": [ + 130, + 582, + 482, + 667 + ], + "type": "text", + "content": "-th token in the sequence, the input representation for the attention head in the forward process includes the instruction representation " + }, + { + "bbox": [ + 130, + 582, + 482, + 667 + ], + "type": "inline_equation", + "content": "\\mathbf{X}_I = [\\pmb{x}_{i_1},\\dots,\\pmb{x}_{i_{n_I}}]" + }, + { + "bbox": [ + 130, + 582, + 482, + 667 + ], + "type": "text", + "content": ", image representation " + }, + { + "bbox": [ + 130, + 582, + 482, + 667 + ], + "type": "inline_equation", + "content": "\\mathbf{X}_V = [\\pmb{x}_{v_1},\\dots,\\pmb{x}_{v_{n_V}}]" + }, + { + "bbox": [ + 130, + 582, + 482, + 667 + ], + "type": "text", + "content": ", and the representation of the historically generated response " + }, + { + "bbox": [ + 130, + 582, + 482, + 667 + ], + "type": "inline_equation", + "content": "\\mathbf{X}_H = [\\pmb{x}_{h_1},\\dots,\\pmb{x}_{h_{n_H}}]" + }, + { + "bbox": [ + 130, + 582, + 482, + 667 + ], + "type": "text", + "content": ". Notably, the image representation considered here is the one that has been processed by the projector. Essentially, the hidden states of each input layer are " + }, + { + "bbox": [ + 130, + 582, + 482, + 667 + ], + "type": "inline_equation", + "content": "\\mathbf{X} = \\mathrm{concat}(\\mathbf{X}_I[1:m],\\mathbf{X}_V,\\mathbf{X}_I[m + 1:n_I],\\mathbf{X}_H)" + }, + { + "bbox": [ + 130, + 582, + 482, + 667 + ], + "type": "text", + "content": ", where the notation " + }, + { + "bbox": [ + 130, + 582, + 482, + 667 + ], + "type": "inline_equation", + "content": "\\mathbf{X}_I[1:m]" + } + ] + } + ], + "index": 9 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "spans": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "type": "text", + "content": "6" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 166, + 91, + 205, + 101 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 166, + 91, + 205, + 101 + ], + "spans": [ + { + "bbox": [ + 166, + 91, + 205, + 101 + ], + "type": "text", + "content": "Liu et al." + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 5 + }, + { + "para_blocks": [ + { + "bbox": [ + 130, + 116, + 480, + 186 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 116, + 480, + 186 + ], + "spans": [ + { + "bbox": [ + 130, + 116, + 480, + 186 + ], + "type": "text", + "content": "indicates the first " + }, + { + "bbox": [ + 130, + 116, + 480, + 186 + ], + "type": "inline_equation", + "content": "m" + }, + { + "bbox": [ + 130, + 116, + 480, + 186 + ], + "type": "text", + "content": " elements in the instruction representation. Each attention head assigns different degrees of attention to each element during the current token representation decoding process. Our aim is to enhance the attention paid to the image. Therefore, we extract the attention weight values related to the image token for the current generated token, intervene, and then redistribute the attention values of each element through softmax." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 130, + 189, + 481, + 308 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 189, + 481, + 308 + ], + "spans": [ + { + "bbox": [ + 130, + 189, + 481, + 308 + ], + "type": "text", + "content": "Excite model in a trustful direction. There have been attempts in some LLMs works to make the answers generated by LLMs more trustworthy by means of intervention [3, 5, 14, 18, 34]. The implemented approach typically involves intervening with the hidden states. As for defining what constitutes a more trustworthy direction, it usually requires additional projection and training to probe this trustful direction. In our case, a response that is more image-based is considered more trustworthy. Since LVLMs have undergone alignment training, the original attention values provide a direction based on image content. As illustrated in Fig. 4, by amplifying the attention values of image tokens based on the original attention values, we can enhance the trustworthiness of our results." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 130, + 308, + 481, + 416 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 308, + 481, + 416 + ], + "spans": [ + { + "bbox": [ + 130, + 308, + 481, + 416 + ], + "type": "text", + "content": "Another nuance involves our avoidance of choosing the attention head used to shift intervention. In the ITI method [18], it is stated that not all attention heads should be subjected to intervention. Therefore, they introduce a trustful score to rank each head across all layers and select the top-k heads for intervention. In our case, the less trustworthy heads with lower attention values receive less intervention. We first extract the attention weights of the image tokens for the current generated token from the attention weights " + }, + { + "bbox": [ + 130, + 308, + 481, + 416 + ], + "type": "inline_equation", + "content": "\\tilde{\\pmb{A}}" + }, + { + "bbox": [ + 130, + 308, + 481, + 416 + ], + "type": "text", + "content": " before softmax operation. We then use the hyper-parameter " + }, + { + "bbox": [ + 130, + 308, + 481, + 416 + ], + "type": "inline_equation", + "content": "\\alpha" + }, + { + "bbox": [ + 130, + 308, + 481, + 416 + ], + "type": "text", + "content": " to control the step size for intervention. From a single attention head perspective, our method can be expressed as follows:" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 194, + 426, + 480, + 439 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 194, + 426, + 480, + 439 + ], + "spans": [ + { + "bbox": [ + 194, + 426, + 480, + 439 + ], + "type": "interline_equation", + "content": "\\tilde {\\boldsymbol {A}} _ {n, j} = \\tilde {\\boldsymbol {A}} _ {n, j} + \\alpha \\cdot | \\tilde {\\boldsymbol {A}} _ {n, j} | \\quad \\text {f o r} j = m + 1 \\text {t o} m + n _ {V}. \\tag {3}", + "image_path": "5c14a866018c11adec47ee06bbbf169b4e346c9e6d3806ca89c8c6fef5f10ff3.jpg" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 130, + 449, + 481, + 533 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 449, + 481, + 533 + ], + "spans": [ + { + "bbox": [ + 130, + 449, + 481, + 533 + ], + "type": "text", + "content": "The model's final vocabulary probability distribution is derived from the projection of the hidden states of the last token in the sequence. Therefore, we extract the attention weights of the last token " + }, + { + "bbox": [ + 130, + 449, + 481, + 533 + ], + "type": "inline_equation", + "content": "n" + }, + { + "bbox": [ + 130, + 449, + 481, + 533 + ], + "type": "text", + "content": " on the image tokens by indexing " + }, + { + "bbox": [ + 130, + 449, + 481, + 533 + ], + "type": "inline_equation", + "content": "\\tilde{A}_{n,j}" + }, + { + "bbox": [ + 130, + 449, + 481, + 533 + ], + "type": "text", + "content": ". Following the intervention, we use the softmax function to redistribute the attention values of each token during the reassignment of encoded hidden states. This procedure is repeated for each subsequent token prediction in an autoregressive manner and is independent of the choice of the decoding algorithm." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 130, + 533, + 481, + 640 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 533, + 481, + 640 + ], + "spans": [ + { + "bbox": [ + 130, + 533, + 481, + 640 + ], + "type": "text", + "content": "Excite more precisely with an attention mode prior. The presence of the BOS token, an attention sink pattern [38], in a sentence results in higher attention values during the attention computation process, which may seem counterintuitive. The BOS token typically signifies the start of a sentence and, as such, does not carry significant semantic content. However, the generation of tokens is significantly influenced by this particular token, a similar pattern that also manifests itself in visual models [9]. As mentioned in StreamLLM [38], the pattern of the attention sink emerges when redundant attention values are present." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 130, + 641, + 481, + 665 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 641, + 481, + 665 + ], + "spans": [ + { + "bbox": [ + 130, + 641, + 481, + 665 + ], + "type": "text", + "content": "Naturally, one might infer that when the sink pattern appears, we excite the image token. To further investigate this phenomenon, as depicted in Fig. 5, we" + } + ] + } + ], + "index": 8 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 430, + 91, + 447, + 100 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 430, + 91, + 447, + 100 + ], + "spans": [ + { + "bbox": [ + 430, + 91, + 447, + 100 + ], + "type": "text", + "content": "PAI" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 474, + 91, + 481, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 474, + 91, + 481, + 100 + ], + "spans": [ + { + "bbox": [ + 474, + 91, + 481, + 100 + ], + "type": "text", + "content": "7" + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 6 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 147, + 115, + 249, + 202 + ], + "blocks": [ + { + "bbox": [ + 147, + 115, + 249, + 202 + ], + "lines": [ + { + "bbox": [ + 147, + 115, + 249, + 202 + ], + "spans": [ + { + "bbox": [ + 147, + 115, + 249, + 202 + ], + "type": "image", + "image_path": "648d051086d77520a3c5a9c8117931c2f8d35ec709ace9f6e518616a5cab6353.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 178, + 205, + 219, + 216 + ], + "lines": [ + { + "bbox": [ + 178, + 205, + 219, + 216 + ], + "spans": [ + { + "bbox": [ + 178, + 205, + 219, + 216 + ], + "type": "text", + "content": "(a) LLAVA" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_caption" + } + ], + "index": 2 + }, + { + "type": "image", + "bbox": [ + 255, + 115, + 356, + 203 + ], + "blocks": [ + { + "bbox": [ + 255, + 115, + 356, + 203 + ], + "lines": [ + { + "bbox": [ + 255, + 115, + 356, + 203 + ], + "spans": [ + { + "bbox": [ + 255, + 115, + 356, + 203 + ], + "type": "image", + "image_path": "8ae6e4405431b66600a296d0dd9f8126ae0de3dfb721da08a8110a1d8b02fdc1.jpg" + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 281, + 205, + 329, + 216 + ], + "lines": [ + { + "bbox": [ + 281, + 205, + 329, + 216 + ], + "spans": [ + { + "bbox": [ + 281, + 205, + 329, + 216 + ], + "type": "text", + "content": "(b) Minigpt4" + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 131, + 219, + 480, + 241 + ], + "lines": [ + { + "bbox": [ + 131, + 219, + 480, + 241 + ], + "spans": [ + { + "bbox": [ + 131, + 219, + 480, + 241 + ], + "type": "text", + "content": "Fig. 5: The BOS token attention ratios of three model. We calculate the attention weights of the BOS token for each head in each layer and display them in a heat map." + } + ] + } + ], + "index": 8, + "angle": 0, + "type": "image_caption" + } + ], + "index": 4 + }, + { + "type": "image", + "bbox": [ + 361, + 115, + 463, + 202 + ], + "blocks": [ + { + "bbox": [ + 361, + 115, + 463, + 202 + ], + "lines": [ + { + "bbox": [ + 361, + 115, + 463, + 202 + ], + "spans": [ + { + "bbox": [ + 361, + 115, + 463, + 202 + ], + "type": "image", + "image_path": "e1acd51a8090e93741e905402c511679c0fd2f57a88cf72eb889f7ef1b475fb9.jpg" + } + ] + } + ], + "index": 6, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 393, + 205, + 432, + 216 + ], + "lines": [ + { + "bbox": [ + 393, + 205, + 432, + 216 + ], + "spans": [ + { + "bbox": [ + 393, + 205, + 432, + 216 + ], + "type": "text", + "content": "(c) Shikra" + } + ] + } + ], + "index": 7, + "angle": 0, + "type": "image_caption" + } + ], + "index": 6 + }, + { + "bbox": [ + 130, + 256, + 482, + 316 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 256, + 482, + 316 + ], + "spans": [ + { + "bbox": [ + 130, + 256, + 482, + 316 + ], + "type": "text", + "content": "find that the sink phenomenon is not overtly evident in the shallow layers. This is because the shallow layers tend to focus more on encoding semantically rich information [35]. When the encoding of semantically rich tokens stabilizes, the attention sink phenomenon arises. Therefore, we build upon the judgement of intervention timing by calculating the similarity of the hidden states." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 131, + 334, + 297, + 346 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 131, + 334, + 297, + 346 + ], + "spans": [ + { + "bbox": [ + 131, + 334, + 297, + 346 + ], + "type": "text", + "content": "4.2 Image-Centric Logit Refine" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 130, + 354, + 482, + 414 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 354, + 482, + 414 + ], + "spans": [ + { + "bbox": [ + 130, + 354, + 482, + 414 + ], + "type": "text", + "content": "In Fig. 1, we observe a peculiar phenomenon where LVLMs continue to generate identical hallucinated text even when the image is removed from the input. This observation naturally leads us to the concept of using the output distribution (when no image is in input) as a reference to penalize our initial prediction distribution. Therefore, we update the distribution of the generated token by:" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 220, + 422, + 391, + 450 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 220, + 422, + 391, + 450 + ], + "spans": [ + { + "bbox": [ + 220, + 422, + 391, + 450 + ], + "type": "interline_equation", + "content": "\\begin{array}{l} p _ {\\text {m o d e l}} = \\gamma \\cdot p _ {\\text {m o d e l}} (\\boldsymbol {y} | \\boldsymbol {X} _ {V}, \\boldsymbol {X} _ {I}, \\boldsymbol {X} _ {H}) \\\\ - (\\gamma - 1) \\cdot p _ {\\text {m o d e l}} (\\boldsymbol {y} | \\boldsymbol {X} _ {I}, \\boldsymbol {X} _ {H}). \\\\ \\end{array}", + "image_path": "742536dda40f4af6dbde1064949afc8c8ac3db6827e65160ed9c1aac0e6fe3ab.jpg" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 130, + 459, + 480, + 493 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 459, + 480, + 493 + ], + "spans": [ + { + "bbox": [ + 130, + 459, + 480, + 493 + ], + "type": "text", + "content": "This equation effectively reduces the predicted probability based on text alone. The weight " + }, + { + "bbox": [ + 130, + 459, + 480, + 493 + ], + "type": "inline_equation", + "content": "\\gamma" + }, + { + "bbox": [ + 130, + 459, + 480, + 493 + ], + "type": "text", + "content": " is used to control the degree of penalty applied to the initial prediction distribution." + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 130, + 494, + 482, + 565 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 494, + 482, + 565 + ], + "spans": [ + { + "bbox": [ + 130, + 494, + 482, + 565 + ], + "type": "text", + "content": "This operation is conceptually similar to LLM-CFG [30]. Essentially, it provides a guided generation mechanism that allows the model to make informed choices between outputs based on image content and those based on language logic. This way, the model can better balance the influence of visual and textual information in its outputs, leading to more contextually accurate and relevant results." + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 131, + 584, + 230, + 597 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 131, + 584, + 230, + 597 + ], + "spans": [ + { + "bbox": [ + 131, + 584, + 230, + 597 + ], + "type": "text", + "content": "5 Experiments" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 131, + 609, + 190, + 621 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 131, + 609, + 190, + 621 + ], + "spans": [ + { + "bbox": [ + 131, + 609, + 190, + 621 + ], + "type": "text", + "content": "5.1 Setup" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 130, + 629, + 481, + 665 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 629, + 481, + 665 + ], + "spans": [ + { + "bbox": [ + 130, + 629, + 481, + 665 + ], + "type": "text", + "content": "Baselines. We evaluate the effectiveness of our method on three different models. To better compare the impact of image feature tokens after different projectors on our method, we selected two models that use linear projectors, LLAVA" + } + ] + } + ], + "index": 17 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "spans": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "type": "text", + "content": "8" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 166, + 91, + 205, + 100 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 166, + 91, + 205, + 100 + ], + "spans": [ + { + "bbox": [ + 166, + 91, + 205, + 100 + ], + "type": "text", + "content": "Liu et al." + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 7 + }, + { + "para_blocks": [ + { + "bbox": [ + 130, + 116, + 479, + 199 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 116, + 479, + 199 + ], + "spans": [ + { + "bbox": [ + 130, + 116, + 479, + 199 + ], + "type": "text", + "content": "and Shikra, as well as one model that uses resamplers, Minigpt4. Additionally, for a more convincing comparison, we report on three decoding methods for comparison: greedy, beam search, and nucleus sample. We also selected the OPERA [13] method, which is an improvement on beam search, and the VCD [16] method, which is an improvement on nucleus sampling, to compare with our results. We used the default hyperparameters from the open-source versions of these two methods." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 130, + 200, + 480, + 294 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 200, + 480, + 294 + ], + "spans": [ + { + "bbox": [ + 130, + 200, + 480, + 294 + ], + "type": "text", + "content": "Implementation Details. As different models have different lengths of image tokens, leading to different degrees of image neglect, to better align with the image sequence length of the model, we set " + }, + { + "bbox": [ + 130, + 200, + 480, + 294 + ], + "type": "inline_equation", + "content": "\\alpha = 0.5" + }, + { + "bbox": [ + 130, + 200, + 480, + 294 + ], + "type": "text", + "content": " for LLAVA, " + }, + { + "bbox": [ + 130, + 200, + 480, + 294 + ], + "type": "inline_equation", + "content": "\\alpha = 0.6" + }, + { + "bbox": [ + 130, + 200, + 480, + 294 + ], + "type": "text", + "content": " for Shikra with long image token sequence lengths and " + }, + { + "bbox": [ + 130, + 200, + 480, + 294 + ], + "type": "inline_equation", + "content": "\\alpha = 0.2" + }, + { + "bbox": [ + 130, + 200, + 480, + 294 + ], + "type": "text", + "content": " for resampler models with short image token sequences. As text inertia is independent of the image token length, we continuously use " + }, + { + "bbox": [ + 130, + 200, + 480, + 294 + ], + "type": "inline_equation", + "content": "\\gamma = 1.1" + }, + { + "bbox": [ + 130, + 200, + 480, + 294 + ], + "type": "text", + "content": ". Apart from this, in the beam search tests, the beam number is set to 5 for all methods, and in the nucleus sample tests, all the common parameters are consistent." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 132, + 313, + 330, + 325 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 313, + 330, + 325 + ], + "spans": [ + { + "bbox": [ + 132, + 313, + 330, + 325 + ], + "type": "text", + "content": "5.2 Benchmark & Evaluation Metrics" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 130, + 333, + 483, + 417 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 333, + 483, + 417 + ], + "spans": [ + { + "bbox": [ + 130, + 333, + 483, + 417 + ], + "type": "text", + "content": "CHAIR [29]. Caption Hallucination Assessment with Image Relevance (CHAIR) is a widely used metric in image captioning tasks. CHAIR operates by creating a set of ground-truth object labels for each image. Any object mentioned in the caption that does not exist in the label set is considered a hallucinated object. CHAIR comprises two evaluation dimensions: instance-level and sentence-level, represented as " + }, + { + "bbox": [ + 130, + 333, + 483, + 417 + ], + "type": "inline_equation", + "content": "\\mathrm{CHAIR_I}" + }, + { + "bbox": [ + 130, + 333, + 483, + 417 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 130, + 333, + 483, + 417 + ], + "type": "inline_equation", + "content": "\\mathrm{CHAIR_S}" + }, + { + "bbox": [ + 130, + 333, + 483, + 417 + ], + "type": "text", + "content": ", respectively. These are calculated in the following manner:" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 228, + 426, + 480, + 451 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 228, + 426, + 480, + 451 + ], + "spans": [ + { + "bbox": [ + 228, + 426, + 480, + 451 + ], + "type": "interline_equation", + "content": "\\mathrm {C H A I R} _ {\\mathrm {I}} = \\frac {| \\{\\text {h a l l u c i n a t e d o b j e c t s} \\} |}{\\text {a l l m e n t i o n e d o b j e c t s}}, \\tag {5}", + "image_path": "2813e59e32bf86265baecda4c4cbc1965e3e20c39bcc18ab7440551faa3465dd.jpg" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 196, + 453, + 480, + 478 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 196, + 453, + 480, + 478 + ], + "spans": [ + { + "bbox": [ + 196, + 453, + 480, + 478 + ], + "type": "interline_equation", + "content": "\\mathrm {C H A I R} _ {\\mathrm {S}} = \\frac {| \\{\\text {c a p t i o n s w i t h h a l l u c i n a t e d o b j e c t s} \\} |}{\\text {a l l c a p t i o n s}}. \\tag {6}", + "image_path": "8592bc3ee9b5e2784afa5ccf3343ef2fef16338e2d4c16e44fe5acbd6dd2dd93.jpg" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 130, + 485, + 492, + 557 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 485, + 492, + 557 + ], + "spans": [ + { + "bbox": [ + 130, + 485, + 492, + 557 + ], + "type": "text", + "content": "We conducted experiments based on the validation set of MSCOCO 2014. Given \"Please help me describe the image in detail.\" as prompt, we subsequently employ the CHAIR metric to evaluate the generated description. To evaluate hallucinations in long sequence generation, we adopted the same setup as used in LURE [43] and OPERA [13], setting the max_new_tokens parameter to 512 and randomly sampled 500 instances for evaluation." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 130, + 558, + 481, + 665 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 558, + 481, + 665 + ], + "spans": [ + { + "bbox": [ + 130, + 558, + 481, + 665 + ], + "type": "text", + "content": "POPE [19]. The Polling-based Object Probing Evaluation (POPE) is a evaluation metric designed in the VQA paradigm. POPE serves as a metric for assessing object hallucination, evaluating hallucinations by asking LVLMs questions such as \"Is there a in the image?\" Here, is replaced with the constructed ground-truth object from three different types of splits. In the \"random\" split, objects are randomly selected from the entire dataset for evaluation. In the \"popular\" split, objects are chosen from those most frequently appearing in the dataset. In the \"adversarial\" split, objects that are highly related to the image objects are selected for evaluation. We conduct our evaluation" + } + ] + } + ], + "index": 9 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 430, + 91, + 447, + 100 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 430, + 91, + 447, + 100 + ], + "spans": [ + { + "bbox": [ + 430, + 91, + 447, + 100 + ], + "type": "text", + "content": "PAI" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 474, + 91, + 481, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 474, + 91, + 481, + 100 + ], + "spans": [ + { + "bbox": [ + 474, + 91, + 481, + 100 + ], + "type": "text", + "content": "9" + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 8 + }, + { + "para_blocks": [ + { + "type": "table", + "bbox": [ + 138, + 179, + 473, + 317 + ], + "blocks": [ + { + "bbox": [ + 130, + 114, + 482, + 170 + ], + "lines": [ + { + "bbox": [ + 130, + 114, + 482, + 170 + ], + "spans": [ + { + "bbox": [ + 130, + 114, + 482, + 170 + ], + "type": "text", + "content": "Table 1: CHAIR hallucination evaluation results on three LVLMs. CHAIR [29] is employed as the evaluation metric, where a smaller number indicates less hallucinations. OPERA is a decoding method based on beam search, VCD is a decoding method based on nucleus sampling, and PAI is an inference intervention method that can be collaborated with any decoding method." + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 138, + 179, + 473, + 317 + ], + "lines": [ + { + "bbox": [ + 138, + 179, + 473, + 317 + ], + "spans": [ + { + "bbox": [ + 138, + 179, + 473, + 317 + ], + "type": "table", + "html": "
DecodingMethodLLAVA [24]Minigpt4 [44]Shikra [6]
CHAIRSCHAIRICHAIRSCHAIRICHAIRSCHAIRI
GreedyVanilla46.613.432.811.151.214.4
PAI24.86.926.38.837.610.0
Beam SearchVanilla46.414.346.613.453.014.7
OPERA [13]44.614.430.19.836.812.4
PAI21.85.624.86.935.811.4
NucleusVanilla58.218.232.711.957.916.4
VCD [16]51.815.134.811.557.616.3
PAI43.414.726.710.349.913.2
", + "image_path": "0d64e374c2e394ad130407221af0fbec87aa43b5cd00c110c2c7492beb813861.jpg" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "table_body" + } + ], + "index": 3 + }, + { + "bbox": [ + 130, + 341, + 482, + 400 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 341, + 482, + 400 + ], + "spans": [ + { + "bbox": [ + 130, + 341, + 482, + 400 + ], + "type": "text", + "content": "on the COCO dataset with 500 images, with each image having 6 questions for each split of POPE. We evaluate the performance of the model in object recognition tasks using both accuracy score and F1 score. Besides, in order to more comprehensively examine hallucinations in multi-turn chats, we have structured the evaluation in both single-turn and multi-turn dialogue forms." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 130, + 401, + 482, + 555 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 401, + 482, + 555 + ], + "spans": [ + { + "bbox": [ + 130, + 401, + 482, + 555 + ], + "type": "text", + "content": "MMHal-Bench [32]. For further evaluation of our method on some challenging datasets, we choose MMHal-Bench, which is designed with 96 image-question pairs, spread across 8 question categories " + }, + { + "bbox": [ + 130, + 401, + 482, + 555 + ], + "type": "inline_equation", + "content": "\\times 12" + }, + { + "bbox": [ + 130, + 401, + 482, + 555 + ], + "type": "text", + "content": " object topics. It contains eight types of questions about object attributes, adversarial objects, comparisons, counting, spatial relations, environment, holistic descriptions, and others to comprehensively assess the model's hallucination performance on high-difficulty datasets. Essentially, it is also a VQA-based evaluation, but unlike the existence-based examination in POPE, its questions also include some logical considerations. Therefore, for evaluation on MMHal-Bench, we first need to answer questions and then use GPT-4 to score the answer based on the response and the ground-truth answer. The evaluation results include the model's scores across all question categories, and the overall score represents the average of these scores." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 130, + 557, + 482, + 665 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 557, + 482, + 665 + ], + "spans": [ + { + "bbox": [ + 130, + 557, + 482, + 665 + ], + "type": "text", + "content": "GPT-4v Assisted Evaluation. To further evaluate the model's performance in image description tasks, we can move beyond the CHAIR metric, which is based on information extraction and only considers object hallucination. We can use GPT-4v for open evaluation. As with previous evaluations [13,16], we sample 50 images on the COCO dataset for evaluation. We construct prompts, input images into GPT-4v, along with the descriptions responses from two assistants. GPT-4v evaluation takes into account two dimensions: Accuracy and Detailedness, denoted respectively as C and D. Detailed prompt construction can be found in the appendix." + } + ] + } + ], + "index": 6 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "spans": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "type": "text", + "content": "10" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 166, + 91, + 205, + 100 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 166, + 91, + 205, + 100 + ], + "spans": [ + { + "bbox": [ + 166, + 91, + 205, + 100 + ], + "type": "text", + "content": "Liu et al." + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 9 + }, + { + "para_blocks": [ + { + "type": "table", + "bbox": [ + 134, + 135, + 480, + 248 + ], + "blocks": [ + { + "bbox": [ + 145, + 114, + 467, + 125 + ], + "lines": [ + { + "bbox": [ + 145, + 114, + 467, + 125 + ], + "spans": [ + { + "bbox": [ + 145, + 114, + 467, + 125 + ], + "type": "text", + "content": "Table 2: Quantitative comparison on POPE. The best results are in bold." + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 134, + 135, + 480, + 248 + ], + "lines": [ + { + "bbox": [ + 134, + 135, + 480, + 248 + ], + "spans": [ + { + "bbox": [ + 134, + 135, + 480, + 248 + ], + "type": "table", + "html": "
DecodingMethodLLAVA [24]Minigpt4 [44]Shikra [6]
Single-turnMulti-turnSingle-turnMulti-turnSingle-turnMulti-turn
AccF1AccF1AccF1AccF1AccF1AccF1
GreedyVanilla84.7685.5185.6984.2774.3973.7878.077.7781.5682.0578.1678.77
PAI85.8285.9787.5486.8375.276.280.8281.682.381.4778.5779.11
Beam SearchVanilla84.984.983.5781.1372.5467.7371.7867.9281.5281.8978.1878.13
PAI86.3385.8986.4285.2773.2269.3674.1571.7282.2581.0180.7280.02
NucleusVanilla80.2581.3282.8781.6357.5258.1664.4963.8979.5580.2876.5477.12
VCD [16]79.5381.0284.0383.359.9758.769.7269.3880.2280.576.3877.07
PAI81.7282.8785.0383.9759.1960.870.2470.2781.4980.7177.9978.47
", + "image_path": "4bfdbee49f1ca74b2c86eb401c31c1f76d7a2ca8eeb890c526caaf61c83fc393.jpg" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "table_body" + } + ], + "index": 3 + }, + { + "bbox": [ + 132, + 274, + 269, + 286 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 274, + 269, + 286 + ], + "spans": [ + { + "bbox": [ + 132, + 274, + 269, + 286 + ], + "type": "text", + "content": "5.3 Experimental Results" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 130, + 301, + 479, + 361 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 301, + 479, + 361 + ], + "spans": [ + { + "bbox": [ + 130, + 301, + 479, + 361 + ], + "type": "text", + "content": "In this section, we analyze the performance of PAI across various hallucination evaluation tasks, including long image description, simplified VQA answer, construction of metric evaluation, and leveraging the near-human cognitive capabilities of GPT-4/GPT-4v as evaluation methods. For further analysis, please refer to the appendix." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 130, + 363, + 480, + 483 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 363, + 480, + 483 + ], + "spans": [ + { + "bbox": [ + 130, + 363, + 480, + 483 + ], + "type": "text", + "content": "Results on long sequence hallucination evaluation. The experimental results are presented in Tab. 1. As our approach is an inference intervention method, it differs from previous decoding hallucination mitigation methods that primarily concentrate on improving a single decoding method. We have tested our method on three decoding techniques. Our method has achieved hallucination mitigation on all three decoding methods used by the three models. However, when integrated with the nucleus, a sampling-based method, the hallucination reduction brought about by our method is not significant. This may be because even though our method has increased the priority of trustful tokens, the sample set during nucleus decoding still contains many hallucination tokens." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 130, + 484, + 480, + 568 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 484, + 480, + 568 + ], + "spans": [ + { + "bbox": [ + 130, + 484, + 480, + 568 + ], + "type": "text", + "content": "Moreover, while OPERA significantly mitigates hallucinations, its time efficiency is considerably higher compared to vanilla. In contrast, our method not only has almost the same time efficiency as vanilla, but it also performs better in reducing hallucination issues. Compared to VCD, during the generation process of long sequence tasks, the introduction of visual uncertainty during decoding sometimes leads to more hallucination descriptions. However, our method can reduce the proportion of hallucination words in the sample pool." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 130, + 570, + 481, + 664 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 570, + 481, + 664 + ], + "spans": [ + { + "bbox": [ + 130, + 570, + 481, + 664 + ], + "type": "text", + "content": "Results on single-turn and multi-turn hallucination evaluation. Unlike the CHAIR evaluation, POPE is in a VQA format, so the response is brief, answering only \"Yes\" or \"No\". The phenomena of text inertia and image neglect may not be as noticeable under this setting, especially for single-turn. However, our method still achieved a notable improvement compared to the vanilla decoding method in single-turn. As for multi-turn evaluation, which involves a longer context, our method achieves more significant improvement as presented in Tab. 2." + } + ] + } + ], + "index": 8 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 430, + 91, + 447, + 100 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 430, + 91, + 447, + 100 + ], + "spans": [ + { + "bbox": [ + 430, + 91, + 447, + 100 + ], + "type": "text", + "content": "PAI" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 470, + 91, + 480, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 470, + 91, + 480, + 100 + ], + "spans": [ + { + "bbox": [ + 470, + 91, + 480, + 100 + ], + "type": "text", + "content": "11" + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 10 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 138, + 114, + 248, + 199 + ], + "blocks": [ + { + "bbox": [ + 138, + 114, + 248, + 199 + ], + "lines": [ + { + "bbox": [ + 138, + 114, + 248, + 199 + ], + "spans": [ + { + "bbox": [ + 138, + 114, + 248, + 199 + ], + "type": "image", + "image_path": "f1f43b14b4c3879f46e3ab454def14beab521b8b9cc2ecf5f8574b145088733b.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 173, + 201, + 214, + 211 + ], + "lines": [ + { + "bbox": [ + 173, + 201, + 214, + 211 + ], + "spans": [ + { + "bbox": [ + 173, + 201, + 214, + 211 + ], + "type": "text", + "content": "(a) LLAVA" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_caption" + } + ], + "index": 2 + }, + { + "type": "image", + "bbox": [ + 251, + 118, + 362, + 200 + ], + "blocks": [ + { + "bbox": [ + 251, + 118, + 362, + 200 + ], + "lines": [ + { + "bbox": [ + 251, + 118, + 362, + 200 + ], + "spans": [ + { + "bbox": [ + 251, + 118, + 362, + 200 + ], + "type": "image", + "image_path": "db8e51032d049fc4b4c221923991eefe4645b95b4e92f9ba524b5b6149be2c43.jpg" + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 283, + 202, + 331, + 212 + ], + "lines": [ + { + "bbox": [ + 283, + 202, + 331, + 212 + ], + "spans": [ + { + "bbox": [ + 283, + 202, + 331, + 212 + ], + "type": "text", + "content": "(b) Minigpt4" + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 131, + 220, + 480, + 243 + ], + "lines": [ + { + "bbox": [ + 131, + 220, + 480, + 243 + ], + "spans": [ + { + "bbox": [ + 131, + 220, + 480, + 243 + ], + "type": "text", + "content": "Fig. 6: Quantitative comparison on the MMHal-Bench. Higher scores indicate better performance." + } + ] + } + ], + "index": 8, + "angle": 0, + "type": "image_caption" + } + ], + "index": 4 + }, + { + "type": "image", + "bbox": [ + 365, + 118, + 477, + 199 + ], + "blocks": [ + { + "bbox": [ + 365, + 118, + 477, + 199 + ], + "lines": [ + { + "bbox": [ + 365, + 118, + 477, + 199 + ], + "spans": [ + { + "bbox": [ + 365, + 118, + 477, + 199 + ], + "type": "image", + "image_path": "0b1940e754152c4bfb9fc2a83f6b1939375e7e116b782821e321fff8a7e32abe.jpg" + } + ] + } + ], + "index": 6, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 402, + 201, + 440, + 212 + ], + "lines": [ + { + "bbox": [ + 402, + 201, + 440, + 212 + ], + "spans": [ + { + "bbox": [ + 402, + 201, + 440, + 212 + ], + "type": "text", + "content": "(c) Shikra" + } + ] + } + ], + "index": 7, + "angle": 0, + "type": "image_caption" + } + ], + "index": 6 + }, + { + "bbox": [ + 130, + 266, + 480, + 373 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 266, + 480, + 373 + ], + "spans": [ + { + "bbox": [ + 130, + 266, + 480, + 373 + ], + "type": "text", + "content": "Results on hallucination evaluation in comprehensive general scenarios. The experimental results, as shown in Fig. 6, indicate that for some more image-based question types, such as object attributes, adversarial objects, and holistic questions, the answers are more accurate when inference intervention with PAI is applied, and there is a certain degree of improvement across all models. However, for some logical questions, such as comparisons and relations, there is no noticeable improvement after intervention. In summary, through the overall metric, i.e., the average of the eight evaluation dimensions, there is a certain degree of improvement compared to the baseline after incorporating PAI." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 130, + 373, + 481, + 469 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 373, + 481, + 469 + ], + "spans": [ + { + "bbox": [ + 130, + 373, + 481, + 469 + ], + "type": "text", + "content": "Results on human-like GPT-4v assisted hallucination evaluation. The experimental results, as shown in Tab. 3, indicate that even when more comprehensive hallucination evaluation dimensions are added, our method, compared to the greedy decoding method, can provide more accurate responses on all three models without losing detail in the description. Given that GPT-4v's visual understanding and language logic capabilities have reached a level close to that of humans, it can more comprehensively illustrate the performance improvements brought about by our method." + } + ] + } + ], + "index": 10 + }, + { + "type": "table", + "bbox": [ + 219, + 510, + 391, + 573 + ], + "blocks": [ + { + "bbox": [ + 151, + 489, + 461, + 500 + ], + "lines": [ + { + "bbox": [ + 151, + 489, + 461, + 500 + ], + "spans": [ + { + "bbox": [ + 151, + 489, + 461, + 500 + ], + "type": "text", + "content": "Table 3: Results on GPT-4V evaluation. The best results are in bold." + } + ] + } + ], + "index": 11, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 219, + 510, + 391, + 573 + ], + "lines": [ + { + "bbox": [ + 219, + 510, + 391, + 573 + ], + "spans": [ + { + "bbox": [ + 219, + 510, + 391, + 573 + ], + "type": "table", + "html": "
MethodLLAVAMinigpt4Shikra
CDCDCD
Greedy5.625.245.85.745.545.25
PAI6.465.367.045.896.045.05
", + "image_path": "6f26198cbd8df5efbbeec2499315474027d5a0dcaa2d395482e1896b1cdbe3bd.jpg" + } + ] + } + ], + "index": 12, + "angle": 0, + "type": "table_body" + } + ], + "index": 12 + }, + { + "bbox": [ + 132, + 611, + 238, + 623 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 611, + 238, + 623 + ], + "spans": [ + { + "bbox": [ + 132, + 611, + 238, + 623 + ], + "type": "text", + "content": "5.4 Ablation Study" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 130, + 629, + 480, + 665 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 629, + 480, + 665 + ], + "spans": [ + { + "bbox": [ + 130, + 629, + 480, + 665 + ], + "type": "text", + "content": "Our method, PAI, consists of two stages of interventions. In the first stage, during forward inference, the hyperparameter " + }, + { + "bbox": [ + 130, + 629, + 480, + 665 + ], + "type": "inline_equation", + "content": "\\alpha" + }, + { + "bbox": [ + 130, + 629, + 480, + 665 + ], + "type": "text", + "content": " is utilized to set the scale of intervention. Simultaneously, the layer prior, represented as \"L\", is used to determine the" + } + ] + } + ], + "index": 14 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "spans": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "type": "text", + "content": "12" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 166, + 91, + 205, + 100 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 166, + 91, + 205, + 100 + ], + "spans": [ + { + "bbox": [ + 166, + 91, + 205, + 100 + ], + "type": "text", + "content": "Liu et al." + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 11 + }, + { + "para_blocks": [ + { + "type": "table", + "bbox": [ + 145, + 157, + 470, + 281 + ], + "blocks": [ + { + "bbox": [ + 131, + 114, + 482, + 148 + ], + "lines": [ + { + "bbox": [ + 131, + 114, + 482, + 148 + ], + "spans": [ + { + "bbox": [ + 131, + 114, + 482, + 148 + ], + "type": "text", + "content": "Table 4: Ablation Study of the Hyperparameter " + }, + { + "bbox": [ + 131, + 114, + 482, + 148 + ], + "type": "inline_equation", + "content": "\\alpha" + }, + { + "bbox": [ + 131, + 114, + 482, + 148 + ], + "type": "text", + "content": ". When " + }, + { + "bbox": [ + 131, + 114, + 482, + 148 + ], + "type": "inline_equation", + "content": "\\alpha" + }, + { + "bbox": [ + 131, + 114, + 482, + 148 + ], + "type": "text", + "content": " becomes excessively large, resulting in an unbalanced response, we terminate the experiment and denote this with a dash (-). F1 values that are considered outliers are highlighted in red." + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 145, + 157, + 470, + 281 + ], + "lines": [ + { + "bbox": [ + 145, + 157, + 470, + 281 + ], + "spans": [ + { + "bbox": [ + 145, + 157, + 470, + 281 + ], + "type": "table", + "html": "
αγLLLaVAMinigpt4Shikra
CHAIRSCHAIRiF1CHAIRSCHAIRiF1CHAIRSCHAIRiF1
---46.213.875.931.610.569.456.215.874.6
0.11.1✓47.413.876.925.08.069.956.415.075.7
0.21.1✓47.413.676.921.27.770.357.215.675.2
0.41.1✓42.412.476.915.16.367.455.214.475.3
0.51.1✓24.66.874.75.42.961.551.613.275.8
0.61.1✓7.84.862.6---37.210.676.7
0.71.1✓------14.46.164.9
", + "image_path": "a21c849f909346b2ffbdb154eb71411642acb4c43012e7f3e6d68b933ab83621.jpg" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "table_body" + } + ], + "index": 3 + }, + { + "bbox": [ + 130, + 304, + 479, + 352 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 304, + 479, + 352 + ], + "spans": [ + { + "bbox": [ + 130, + 304, + 479, + 352 + ], + "type": "text", + "content": "attention layer for intervention. The second stage unfolds during the decoding process, where we mitigate text inertia by subtracting the logits distribution that results from inputs devoid of image information. In this stage, the scale is managed by the parameter " + }, + { + "bbox": [ + 130, + 304, + 479, + 352 + ], + "type": "inline_equation", + "content": "\\gamma" + }, + { + "bbox": [ + 130, + 304, + 479, + 352 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 130, + 352, + 482, + 472 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 352, + 482, + 472 + ], + "spans": [ + { + "bbox": [ + 130, + 352, + 482, + 472 + ], + "type": "text", + "content": "We use LLaVA-1.5 as the representative LVLM baseline and the greedy decoding method as the basic baseline to compare the impact of our hyperparameters on the task of long sequence image description. To evaluate our method, we choose the CHAIR metric. However, since CHAIR only evaluates the hallucination problem, we have incorporated the F1 score to consider information richness and accuracy. This makes the comparison with the CHAIR metric fairer when the F1 scores are similar. In the above, the F1 scores of the various methods differ slightly, so we have not included this somewhat redundant measure. The F1 score is calculated by counting the objects included in the description, the objects in the ground-truth sets, and the hallucinated objects." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 130, + 472, + 482, + 544 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 472, + 482, + 544 + ], + "spans": [ + { + "bbox": [ + 130, + 472, + 482, + 544 + ], + "type": "text", + "content": "Effects of " + }, + { + "bbox": [ + 130, + 472, + 482, + 544 + ], + "type": "inline_equation", + "content": "\\alpha" + }, + { + "bbox": [ + 130, + 472, + 482, + 544 + ], + "type": "text", + "content": " in Exciting Image Attention. In the process of exciting the attention values of image tokens, we introduce a parameter " + }, + { + "bbox": [ + 130, + 472, + 482, + 544 + ], + "type": "inline_equation", + "content": "\\alpha" + }, + { + "bbox": [ + 130, + 472, + 482, + 544 + ], + "type": "text", + "content": " to control the amplification scale. As shown in Tab. 4, different LVLMs exhibit varying sensitivity to the amplification scale. This sensitivity not only depends on the length of the model's image tokens (e.g., the image token length of LLaVA-1.5 is 576, while that of Minigpt4 is only 32), but also the original attention weights distribution." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 130, + 545, + 482, + 604 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 545, + 482, + 604 + ], + "spans": [ + { + "bbox": [ + 130, + 545, + 482, + 604 + ], + "type": "text", + "content": "However, a commonality across these models is that an appropriate amplification scale can achieve a balance between the number of hallucinated objects in the description and the amount of information conveyed. If the scale is too small, the description may still contain many hallucinated objects. Conversely, if the scale is too large, the amount of information in the response will decrease." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 130, + 605, + 482, + 665 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 605, + 482, + 665 + ], + "spans": [ + { + "bbox": [ + 130, + 605, + 482, + 665 + ], + "type": "text", + "content": "Effects of Layer Prior in Exciting Image Attention. We further investigate the control of the intervention layer for exciting image attention. As observed in Tab. 5, the introduction of this prior does indeed improve the performance of our method. However, when there is no control over the intervention layer and interventions are applied to all layers, the performances of different models ex" + } + ] + } + ], + "index": 8 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 430, + 91, + 447, + 100 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 430, + 91, + 447, + 100 + ], + "spans": [ + { + "bbox": [ + 430, + 91, + 447, + 100 + ], + "type": "text", + "content": "PAI" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 470, + 91, + 481, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 470, + 91, + 481, + 100 + ], + "spans": [ + { + "bbox": [ + 470, + 91, + 481, + 100 + ], + "type": "text", + "content": "13" + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 12 + }, + { + "para_blocks": [ + { + "type": "table", + "bbox": [ + 135, + 146, + 478, + 248 + ], + "blocks": [ + { + "bbox": [ + 132, + 114, + 481, + 137 + ], + "lines": [ + { + "bbox": [ + 132, + 114, + 481, + 137 + ], + "spans": [ + { + "bbox": [ + 132, + 114, + 481, + 137 + ], + "type": "text", + "content": "Table 5: Ablation Study of Hyperparameter " + }, + { + "bbox": [ + 132, + 114, + 481, + 137 + ], + "type": "inline_equation", + "content": "\\gamma" + }, + { + "bbox": [ + 132, + 114, + 481, + 137 + ], + "type": "text", + "content": " and Layer Prior \"L\". Results are presented for the models LLAVA (left), Minigpt4 (middle), and Shikra (right)." + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 135, + 146, + 478, + 248 + ], + "lines": [ + { + "bbox": [ + 135, + 146, + 478, + 248 + ], + "spans": [ + { + "bbox": [ + 135, + 146, + 478, + 248 + ], + "type": "table", + "html": "
αγL CHAIRSCHAIRIF1αγL CHAIRSCHAIRIF1αγL CHAIRSCHAIRIF1
--46.213.875.9--31.610.569.4--56.215.874.6
0.51.1✓24.66.875.70.21.1✓21.27.770.30.61.1✓37.2
0.51.2✓23.46.675.70.21.2✓16.46.270.30.61.2✓40.2
0.51.3✓24.67.574.40.21.3✓14.46.969.50.61.3✓40.6
0.51.5✓25.09.674.60.21.5✓12.65.467.10.61.5✓40.2
0.52.0✓24.27.174.00.22.0✓8.44.664.30.62.0✓39.4
0.51.1✓24.66.875.70.21.1✓21.27.770.30.61.1✓37.2
0.51.1×20.66.771.70.21.1×22.38.470.30.61.1×56.4
", + "image_path": "6ad52be2daf7b8fe34bc0471f103a2f4fc0b20d3841f4648a166ba8226b4a13d.jpg" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "table_body" + } + ], + "index": 3 + }, + { + "bbox": [ + 132, + 274, + 480, + 322 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 274, + 480, + 322 + ], + "spans": [ + { + "bbox": [ + 132, + 274, + 480, + 322 + ], + "type": "text", + "content": "hibit some variations. For Shikra, in the absence of the layer prior, our method's results regress to the baseline. For both LLaVA and Minigpt4, the loss of the layer prior causes some unfavorable fluctuations in both the CHAIR metric and the F1 score." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 132, + 324, + 481, + 408 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 324, + 481, + 408 + ], + "spans": [ + { + "bbox": [ + 132, + 324, + 481, + 408 + ], + "type": "text", + "content": "Effects of " + }, + { + "bbox": [ + 132, + 324, + 481, + 408 + ], + "type": "inline_equation", + "content": "\\gamma" + }, + { + "bbox": [ + 132, + 324, + 481, + 408 + ], + "type": "text", + "content": " in Mitigating Language Prior. Tab. 5 presents the results of an ablation study focusing on " + }, + { + "bbox": [ + 132, + 324, + 481, + 408 + ], + "type": "inline_equation", + "content": "\\gamma" + }, + { + "bbox": [ + 132, + 324, + 481, + 408 + ], + "type": "text", + "content": ", which adjusts the balance between output distributions from conditioned inputs with excited image tokens and pure text inputs. Unlike the other two models, the Minigpt4 model is highly sensitive to " + }, + { + "bbox": [ + 132, + 324, + 481, + 408 + ], + "type": "inline_equation", + "content": "\\gamma" + }, + { + "bbox": [ + 132, + 324, + 481, + 408 + ], + "type": "text", + "content": ". When " + }, + { + "bbox": [ + 132, + 324, + 481, + 408 + ], + "type": "inline_equation", + "content": "\\gamma" + }, + { + "bbox": [ + 132, + 324, + 481, + 408 + ], + "type": "text", + "content": " is too large, it can lead to uncontrollable model behavior. The experimental results suggest that maintaining " + }, + { + "bbox": [ + 132, + 324, + 481, + 408 + ], + "type": "inline_equation", + "content": "\\gamma" + }, + { + "bbox": [ + 132, + 324, + 481, + 408 + ], + "type": "text", + "content": " within a relatively small range, such as 1.1 - 1.2, yields the most stable performance." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 133, + 436, + 311, + 449 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 436, + 311, + 449 + ], + "spans": [ + { + "bbox": [ + 133, + 436, + 311, + 449 + ], + "type": "text", + "content": "6 Conclusion and Limitation" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 132, + 472, + 481, + 579 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 472, + 481, + 579 + ], + "spans": [ + { + "bbox": [ + 132, + 472, + 481, + 579 + ], + "type": "text", + "content": "In this paper, we first analyze the causes and manifestations of hallucinations in LVLMs. We propose a phenomenon termed \"text inertia,\" wherein the model continues to produce the same hallucinatory descriptions even when no image input is provided. This issue fundamentally stems from the model's neglect of image tokens. Consequently, we introduce the PAI method to intervene in the model's inference process, steering it towards an image-based and trustworthy direction. This is a training-free method and does not require any external tools. Extensive experiments on multiple benchmarks and LVLMs have validated the effectiveness of PAI in mitigating hallucination issues." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 132, + 581, + 481, + 665 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 581, + 481, + 665 + ], + "spans": [ + { + "bbox": [ + 132, + 581, + 481, + 665 + ], + "type": "text", + "content": "Limitations: (1) The language decoders of existing open-source LVLMs are primarily models from the LLaMA-family. It is worth exploring whether the issues of image neglect and text inertia are introduced by LLaMA. (2) As we described, our method fundamentally alleviates the image neglect issue during model inference. Its upper limit depends on the capabilities of the well-trained model. Therefore, it is worth investigating whether incorporating this issue as a loss during the training process could lead to further performance improvements." + } + ] + } + ], + "index": 8 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "spans": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "type": "text", + "content": "14" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 166, + 91, + 205, + 100 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 166, + 91, + 205, + 100 + ], + "spans": [ + { + "bbox": [ + 166, + 91, + 205, + 100 + ], + "type": "text", + "content": "Liu et al." + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 13 + }, + { + "para_blocks": [ + { + "bbox": [ + 133, + 114, + 246, + 129 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 114, + 246, + 129 + ], + "spans": [ + { + "bbox": [ + 133, + 114, + 246, + 129 + ], + "type": "text", + "content": "Acknowledgements" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 132, + 140, + 480, + 163 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 140, + 480, + 163 + ], + "spans": [ + { + "bbox": [ + 132, + 140, + 480, + 163 + ], + "type": "text", + "content": "This work is supported by the National Science Foundation of China (62132017), Zhejiang Provincial Natural Science Foundation of China (LD24F020011)." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 133, + 182, + 197, + 194 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 182, + 197, + 194 + ], + "spans": [ + { + "bbox": [ + 133, + 182, + 197, + 194 + ], + "type": "text", + "content": "References" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 134, + 205, + 481, + 665 + ], + "type": "list", + "angle": 0, + "index": 19, + "blocks": [ + { + "bbox": [ + 138, + 205, + 481, + 260 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 205, + 481, + 260 + ], + "spans": [ + { + "bbox": [ + 138, + 205, + 481, + 260 + ], + "type": "text", + "content": "1. Alayrac, J.B., Donahue, J., Luc, P., Miech, A., Barr, I., Hasson, Y., Lenc, K., Mensch, A., Millican, K., Reynolds, M., Ring, R., Rutherford, E., Cabi, S., Han, T., Gong, Z., Samangooei, S., Monteiro, M., Menick, J., Borgeaud, S., Brock, A., Nematzadeh, A., Sharifzadeh, S., Binkowski, M., Barreira, R., Vinyals, O., Zisserman, A., Simonyan, K.: Flamingo: a visual language model for few-shot learning" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 138, + 262, + 481, + 282 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 262, + 481, + 282 + ], + "spans": [ + { + "bbox": [ + 138, + 262, + 481, + 282 + ], + "type": "text", + "content": "2. Anisha, G., Yin, J., Bas, E.: Detecting and preventing hallucinations in large vision language models (Aug 2023)" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 138, + 284, + 481, + 304 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 284, + 481, + 304 + ], + "spans": [ + { + "bbox": [ + 138, + 284, + 481, + 304 + ], + "type": "text", + "content": "3. Azaria, A., Mitchell, T.: The internal state of an llm knows when its lying (Apr 2023)" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 138, + 304, + 481, + 380 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 304, + 481, + 380 + ], + "spans": [ + { + "bbox": [ + 138, + 304, + 481, + 380 + ], + "type": "text", + "content": "4. Brown, T., Mann, B., Ryder, N., Subbiah, M., Kaplan, J., Dhariwal, P., Neelakanthan, A., Shyam, P., Sastry, G., Amanda, A., Agarwal, S., Herbert-Voss, A., Krueger, G., Tom, H., Child, R., Ramesh, A., Ziegler, D., Wu, J., Winter, C., Hesse, C., Chen, M., Sigler, E., Litwin, M., Gray, S., Benjamin, C., Clark, J., Berner, C., Sam, M., Radford, A., Sutskever, I., Amodei, D.: Language models are few-shot learners. arXiv: Computation and Language, arXiv: Computation and Language (May 2020)" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 138, + 382, + 481, + 403 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 382, + 481, + 403 + ], + "spans": [ + { + "bbox": [ + 138, + 382, + 481, + 403 + ], + "type": "text", + "content": "5. Burns, C., Ye, H., Klein, D., Steinhardt, J.: Discovering latent knowledge in language models without supervision (Dec 2022)" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 138, + 404, + 481, + 435 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 404, + 481, + 435 + ], + "spans": [ + { + "bbox": [ + 138, + 404, + 481, + 435 + ], + "type": "text", + "content": "6. Chen, K., Zhang, Z., Zeng, W., Zhang, R., Zhu, F., Zhao, R., Research, S., Sklsde, S., Beihang, B., Seiee, S.: Shikra: Unleashing multimodal llm's referential dialogue magic" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 138, + 437, + 481, + 468 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 437, + 481, + 468 + ], + "spans": [ + { + "bbox": [ + 138, + 437, + 481, + 468 + ], + "type": "text", + "content": "7. Chen, Z., Wu, J., Wang, W., Su, W., Chen, G., Xing, S., Muyan, Z., Zhang, Q., Zhu, X., Lu, L., Li, B., Luo, P., Lu, T., Qiao, Y., Dai, J.: Intervl: Scaling up vision foundation models and aligning for generic visual-linguistic tasks (Dec 2023)" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 138, + 470, + 481, + 501 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 470, + 481, + 501 + ], + "spans": [ + { + "bbox": [ + 138, + 470, + 481, + 501 + ], + "type": "text", + "content": "8. Dai, W., Li, J., Li, D., Huat, A., Zhao, J., Wang, W., Li, B., Fung, P., Hoi, S.: Instructclip: Towards general-purpose vision-language models with instruction tuning" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 138, + 502, + 481, + 522 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 502, + 481, + 522 + ], + "spans": [ + { + "bbox": [ + 138, + 502, + 481, + 522 + ], + "type": "text", + "content": "9. Darcet, T., Oquab, M., Mairal, J., Bojanowski, P.: Vision transformers need registers. arXiv preprint arXiv:2309.16588 (2023)" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 134, + 523, + 481, + 567 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 134, + 523, + 481, + 567 + ], + "spans": [ + { + "bbox": [ + 134, + 523, + 481, + 567 + ], + "type": "text", + "content": "10. Devlin, J., Chang, M.W., Lee, K., Toutanova, K.: Bert: Pre-training of deep bidirectional transformers for language understanding. In: Proceedings of the 2019 Conference of the North (Jan 2019). https://doi.org/10.18653/v1/n19-1423, http://dx.doi.org/10.18653/v1/n19-1423" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 134, + 567, + 481, + 589 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 134, + 567, + 481, + 589 + ], + "spans": [ + { + "bbox": [ + 134, + 567, + 481, + 589 + ], + "type": "text", + "content": "11. Han, Y., Zhang, C., Chen, X., Yang, X., Wang, Z., Yu, G., Fu, B., Zhang, H.: Chartlama: A multimodal llm for chart understanding and generation (Nov 2023)" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 134, + 590, + 481, + 610 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 134, + 590, + 481, + 610 + ], + "spans": [ + { + "bbox": [ + 134, + 590, + 481, + 610 + ], + "type": "text", + "content": "12. Hu, H., Zhang, J., Zhao, M., Sun, Z.: Ciem: Contrastive instruction evaluation method for better instruction tuning" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 134, + 611, + 481, + 643 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 134, + 611, + 481, + 643 + ], + "spans": [ + { + "bbox": [ + 134, + 611, + 481, + 643 + ], + "type": "text", + "content": "13. Huang, Q., Dong, X., Zhang, P., Wang, B., He, C., Wang, J., Lin, D., Zhang, W., Yu, N.: Opera: Alleviating hallucination in multi-modal large language models via over-trust penalty and retrospection-allocation (Nov 2023)" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 134, + 643, + 481, + 665 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 134, + 643, + 481, + 665 + ], + "spans": [ + { + "bbox": [ + 134, + 643, + 481, + 665 + ], + "type": "text", + "content": "14. Kai, J., Zhang, T., Hu, H., Lin, Z.: Sh2: Self-highlighted hesitation helps you decode more truthfully (Jan 2024)" + } + ] + } + ], + "index": 18 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 430, + 91, + 447, + 100 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 430, + 91, + 447, + 100 + ], + "spans": [ + { + "bbox": [ + 430, + 91, + 447, + 100 + ], + "type": "text", + "content": "PAI" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 470, + 91, + 481, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 470, + 91, + 481, + 100 + ], + "spans": [ + { + "bbox": [ + 470, + 91, + 481, + 100 + ], + "type": "text", + "content": "15" + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 14 + }, + { + "para_blocks": [ + { + "bbox": [ + 132, + 117, + 480, + 665 + ], + "type": "list", + "angle": 0, + "index": 21, + "blocks": [ + { + "bbox": [ + 133, + 117, + 480, + 138 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 117, + 480, + 138 + ], + "spans": [ + { + "bbox": [ + 133, + 117, + 480, + 138 + ], + "type": "text", + "content": "15. Lai, X., Tian, Z., Chen, Y., Li, Y., Yuan, Y., Liu, S., Jia, J., Kong, H., Research, M.: Lisa: Reasoning segmentation via large language model" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 133, + 140, + 480, + 160 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 140, + 480, + 160 + ], + "spans": [ + { + "bbox": [ + 133, + 140, + 480, + 160 + ], + "type": "text", + "content": "16. Leng, S., Zhang, H., Chen, G., Li, X., Lu, S., Miao, C., Bing, L.: Mitigating object hallucinations in large vision-language models through visual contrastive decoding" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 133, + 161, + 480, + 182 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 161, + 480, + 182 + ], + "spans": [ + { + "bbox": [ + 133, + 161, + 480, + 182 + ], + "type": "text", + "content": "17. Li, J., Li, D., Savarese, S., Hoi, S.: Blip-2: Bootstrapping language-image pretraining with frozen image encoders and large language models" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 133, + 183, + 480, + 215 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 183, + 480, + 215 + ], + "spans": [ + { + "bbox": [ + 133, + 183, + 480, + 215 + ], + "type": "text", + "content": "18. Li, K., Patel, O., Viégas, F., Pfister, H., Wattenberg, M.: Inference-time intervention: Eliciting truthful answers from a language model. Advances in Neural Information Processing Systems 36 (2024)" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 133, + 216, + 480, + 237 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 216, + 480, + 237 + ], + "spans": [ + { + "bbox": [ + 133, + 216, + 480, + 237 + ], + "type": "text", + "content": "19. Li, Y., Du, Y., Zhou, K., Wang, J., Zhao, W.X., Wen, J.R.: Evaluating object hallucination in large vision-language models. arXiv preprint arXiv:2305.10355 (2023)" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 133, + 238, + 480, + 270 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 238, + 480, + 270 + ], + "spans": [ + { + "bbox": [ + 133, + 238, + 480, + 270 + ], + "type": "text", + "content": "20. Li, Z., Yang, B., Liu, Q., Ma, Z., Zhang, S., Yang, J., Sun, Y., Liu, Y., Bai, X.: Monkey: Image resolution and text label are important things for large multi-modal models (Nov 2023)" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 132, + 270, + 480, + 303 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 270, + 480, + 303 + ], + "spans": [ + { + "bbox": [ + 132, + 270, + 480, + 303 + ], + "type": "text", + "content": "21. Liu, F., Lin, K., Li, L., Wang, J., Yacoob, Y., Wang, L.: Mitigating hallucination in large multi-modal models via robust instruction tuning. In: The Twelfth International Conference on Learning Representations (2023)" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 132, + 304, + 480, + 325 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 304, + 480, + 325 + ], + "spans": [ + { + "bbox": [ + 132, + 304, + 480, + 325 + ], + "type": "text", + "content": "22. Liu, F., Lin, K., Li, L., Wang, J., Yacoob, Y., Wang, L.: Mitigating hallucination in large multi-modal models via robust instruction tuning (Sep 2023)" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 132, + 326, + 480, + 346 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 326, + 480, + 346 + ], + "spans": [ + { + "bbox": [ + 132, + 326, + 480, + 346 + ], + "type": "text", + "content": "23. Liu, H., Xue, W., Chen, Y., Chen, D., Zhao, X., Wang, K., Hou, L., Li, R., Peng, W.: A survey on hallucination in large vision-language models" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 132, + 348, + 480, + 369 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 348, + 480, + 369 + ], + "spans": [ + { + "bbox": [ + 132, + 348, + 480, + 369 + ], + "type": "text", + "content": "24. Liu, H., Li, C., Li, Y., Lee, Y.J.: Improved baselines with visual instruction tuning. arXiv preprint arXiv:2310.03744 (2023)" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 132, + 369, + 480, + 390 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 369, + 480, + 390 + ], + "spans": [ + { + "bbox": [ + 132, + 369, + 480, + 390 + ], + "type": "text", + "content": "25. Liu, H., Li, C., Wu, Q., Lee, Y., Madison, M., Research, M.: Visual instruction tuning" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 133, + 392, + 480, + 445 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 392, + 480, + 445 + ], + "spans": [ + { + "bbox": [ + 133, + 392, + 480, + 445 + ], + "type": "text", + "content": "26. Mishra, S., Khashabi, D., Baral, C., Hajishirzi, H.: Cross-task generalization via natural language crowdsourcing instructions. In: Proceedings of the 60th Annual Meeting of the Association for Computational Linguistics (Volume 1: Long Papers) (Jan 2022). https://doi.org/10.18653/v1/2022.acl-long.244, http://dx.doi.org/10.18653/v1/2022.acl-long.244" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 132, + 447, + 480, + 490 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 447, + 480, + 490 + ], + "spans": [ + { + "bbox": [ + 132, + 447, + 480, + 490 + ], + "type": "text", + "content": "27. Ouyang, L., Wu, J., Jiang, X., Almeida, D., Wainwright, C.L., Mishkin, P., Zhang, C., Agarwal, S., Slama, K., Ray, A., Schulman, J., Hilton, J., Kelton, F., Miller, L., Simens, M., Askell, A., Welinder, P., Christiano, P.F., Leike, J., Lowe, R.: Training language models to follow instructions with human feedback (2022)" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 132, + 491, + 452, + 501 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 491, + 452, + 501 + ], + "spans": [ + { + "bbox": [ + 132, + 491, + 452, + 501 + ], + "type": "text", + "content": "28. Peng, B., Li, C., He, P., Galley, M., Gao, J.: Instruction tuning with gpt-4" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 133, + 502, + 480, + 544 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 502, + 480, + 544 + ], + "spans": [ + { + "bbox": [ + 133, + 502, + 480, + 544 + ], + "type": "text", + "content": "29. Rohrbach, A., Hendricks, L.A., Burns, K., Darrell, T., Saenko, K.: Object hallucination in image captioning. In: Proceedings of the 2018 Conference on Empirical Methods in Natural Language Processing (Jan 2018). https://doi.org/10.18653/v1/d18-1437, http://dx.doi.org/10.18653/v1/d18-1437" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 132, + 545, + 480, + 566 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 545, + 480, + 566 + ], + "spans": [ + { + "bbox": [ + 132, + 545, + 480, + 566 + ], + "type": "text", + "content": "30. Sanchez, G., Fan, H., Spangher, A., Levi, E., Ammanamanchi, P.S., Biderman, S.: Stay on topic with classifier-free guidance. arXiv preprint arXiv:2306.17806 (2023)" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 133, + 567, + 480, + 599 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 567, + 480, + 599 + ], + "spans": [ + { + "bbox": [ + 133, + 567, + 480, + 599 + ], + "type": "text", + "content": "31. Sun, Z., Shen, S., Cao, S., Liu, H., Li, C., Shen, Y., Gan, C., Gui, L.Y., Wang, Y.X., Yang, Y., Keutzer, K., Darrell, T.: Aligning large multimodal models with factually augmented rlhf (Sep 2023)" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 132, + 600, + 480, + 632 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 600, + 480, + 632 + ], + "spans": [ + { + "bbox": [ + 132, + 600, + 480, + 632 + ], + "type": "text", + "content": "32. Sun, Z., Shen, S., Cao, S., Liu, H., Li, C., Shen, Y., Gan, C., Gui, L.Y., Wang, Y.X., Yang, Y., et al.: Aligning large multimodal models with factually augmented rlhf. arXiv preprint arXiv:2309.14525 (2023)" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 132, + 633, + 480, + 665 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 633, + 480, + 665 + ], + "spans": [ + { + "bbox": [ + 132, + 633, + 480, + 665 + ], + "type": "text", + "content": "33. Touvron, H., Lavril, T., Izacard, G., Martinet, X., Lachaux, M.A., Lacroix, T., Rozi'ere, B., Goyal, N., Hambro, E., Azhar, F., Rodriguez, A., Joulin, A., Grave, E., Lample, G.: Llama: Open and efficient foundation language models" + } + ] + } + ], + "index": 20 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "spans": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "type": "text", + "content": "16" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 166, + 91, + 205, + 100 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 166, + 91, + 205, + 100 + ], + "spans": [ + { + "bbox": [ + 166, + 91, + 205, + 100 + ], + "type": "text", + "content": "Liu et al." + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 15 + }, + { + "para_blocks": [ + { + "bbox": [ + 130, + 116, + 482, + 413 + ], + "type": "list", + "angle": 0, + "index": 13, + "blocks": [ + { + "bbox": [ + 130, + 116, + 482, + 149 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 116, + 482, + 149 + ], + "spans": [ + { + "bbox": [ + 130, + 116, + 482, + 149 + ], + "type": "text", + "content": "34. Turner, A., Thiergart, L., Udell, D., Leech, G., Mini, U., MacDiarmid, M.: Activation addition: Steering language models without optimization. arXiv preprint arXiv:2308.10248 (2023)" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 130, + 149, + 482, + 182 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 149, + 482, + 182 + ], + "spans": [ + { + "bbox": [ + 130, + 149, + 482, + 182 + ], + "type": "text", + "content": "35. Wang, L., Li, L., Dai, D., Chen, D., Zhou, H., Meng, F., Zhou, J., Sun, X.: Label words are anchors: An information flow perspective for understanding in-context learning. arXiv preprint arXiv:2305.14160 (2023)" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 130, + 182, + 482, + 205 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 182, + 482, + 205 + ], + "spans": [ + { + "bbox": [ + 130, + 182, + 482, + 205 + ], + "type": "text", + "content": "36. Wei, F., Zhang, X., Zhang, A., Zhang, B., Chu, X.: Lenna: Language enhanced reasoning detection assistant (Dec 2023)" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 130, + 205, + 482, + 226 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 205, + 482, + 226 + ], + "spans": [ + { + "bbox": [ + 130, + 205, + 482, + 226 + ], + "type": "text", + "content": "37. Wei, J., Bosma, M., Zhao, V., Guu, K., Yu, A., Lester, B., Du, N., Dai, A., Le, Q.: Finetuned language models are zero-shot learners. Learning, Learning (Sep 2021)" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 130, + 226, + 482, + 248 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 226, + 482, + 248 + ], + "spans": [ + { + "bbox": [ + 130, + 226, + 482, + 248 + ], + "type": "text", + "content": "38. Xiao, G., Tian, Y., Chen, B., Han, S., Lewis, M.: Efficient streaming language models with attention sinks (Sep 2023)" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 130, + 248, + 482, + 282 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 248, + 482, + 282 + ], + "spans": [ + { + "bbox": [ + 130, + 248, + 482, + 282 + ], + "type": "text", + "content": "39. Ye, Q., Xu, H., Xu, G., Ye, J., Yan, M., Zhou, Y., Wang, J., Hu, A., Shi, P., Shi, Y., Li, C., Xu, Y., Chen, H., Tian, J., Qi, Q., Zhang, J., Huang, F.: mplug-owl: Modularization empowers large language models with multimodality" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 130, + 281, + 482, + 304 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 281, + 482, + 304 + ], + "spans": [ + { + "bbox": [ + 130, + 281, + 482, + 304 + ], + "type": "text", + "content": "40. Yin, S., Fu, C., Zhao, S., Xu, T., Wang, H., Sui, D., Shen, Y., Li, K., Sun, X., Chen, E.: Woodpecker: Hallucination correction for multimodal large language models" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 130, + 304, + 482, + 335 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 304, + 482, + 335 + ], + "spans": [ + { + "bbox": [ + 130, + 304, + 482, + 335 + ], + "type": "text", + "content": "41. You, H., Zhang, H., Gan, Z., Du, X., Zhang, B., Wang, Z., Cao, L., Chang, S.F., Yang, Y.: Ferret: Refer and ground anything anywhere at any granularity (Oct 2023)" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 130, + 335, + 482, + 357 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 335, + 482, + 357 + ], + "spans": [ + { + "bbox": [ + 130, + 335, + 482, + 357 + ], + "type": "text", + "content": "42. Zhang, Y., Qian, S., Peng, B., Liu, S., Jia, J.: Prompt highlighter: Interactive control for multi-modal llms. arXiv preprint arXiv:2312.04302 (2023)" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 130, + 357, + 482, + 390 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 357, + 482, + 390 + ], + "spans": [ + { + "bbox": [ + 130, + 357, + 482, + 390 + ], + "type": "text", + "content": "43. Zhou, Y., Cui, C., Yoon, J., Zhang, L., Deng, Z., Finn, C., Bansal, M., Yao, H.: Analyzing and mitigating object hallucination in large vision-language models (Oct 2023)" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 130, + 390, + 482, + 413 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 390, + 482, + 413 + ], + "spans": [ + { + "bbox": [ + 130, + 390, + 482, + 413 + ], + "type": "text", + "content": "44. Zhu, D., Chen, J., Shen, X., Li, X., Elhoseiny, M.: Minigpt-4: Enhancing vision-language understanding with advanced large language models" + } + ] + } + ], + "index": 12 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 430, + 91, + 447, + 100 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 430, + 91, + 447, + 100 + ], + "spans": [ + { + "bbox": [ + 430, + 91, + 447, + 100 + ], + "type": "text", + "content": "PAI" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 470, + 91, + 481, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 470, + 91, + 481, + 100 + ], + "spans": [ + { + "bbox": [ + 470, + 91, + 481, + 100 + ], + "type": "text", + "content": "17" + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 16 + } + ], + "_backend": "vlm", + "_version_name": "2.6.4" +} \ No newline at end of file diff --git a/2024/Per-Gaussian Embedding-Based Deformation for Deformable 3D Gaussian Splatting/8af14cc6-a852-4c45-9f34-dffb20b17a0a_content_list.json b/2024/Per-Gaussian Embedding-Based Deformation for Deformable 3D Gaussian Splatting/8af14cc6-a852-4c45-9f34-dffb20b17a0a_content_list.json new file mode 100644 index 0000000000000000000000000000000000000000..0e3879d0c987fc3deadb5c90933c495696789b76 --- /dev/null +++ b/2024/Per-Gaussian Embedding-Based Deformation for Deformable 3D Gaussian Splatting/8af14cc6-a852-4c45-9f34-dffb20b17a0a_content_list.json @@ -0,0 +1,1718 @@ +[ + { + "type": "text", + "text": "Per-Gaussian Embedding-Based Deformation for Deformable 3D Gaussian Splitting", + "text_level": 1, + "bbox": [ + 223, + 140, + 779, + 186 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Jeongmin Bae1*, Seoha Kim1*, Youngsik Yun1, Hahyun Lee2, Gun Bang2, and Youngjung Uh1", + "bbox": [ + 310, + 210, + 689, + 244 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "1 Yonsei University, Seoul 03722, Korea {jaymin.bae, hailey07, bbangsik, yj.uh}@yonsei.ac.kr", + "bbox": [ + 299, + 253, + 702, + 282 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "$^{2}$ Electronics and Telecommunications Research Institute, Daejeon 34129, Korea {hanilee, gbang}@etri.re.kr", + "bbox": [ + 230, + 282, + 769, + 310 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/6fcce4eccafc285e23868dd70db2ef6658f862c19476727352665328c2371815.jpg", + "image_caption": [ + "Fig. 1: Overview. (a) Existing deformable 3D Gaussian Splitting methods show blurry results in complex dynamic scenes, even with deformation fields using finer feature grids. (b) Our model solves the problem by employing per-Gaussian latent embeddings to predict deformations for each Gaussian and achieves a clearer results." + ], + "image_footnote": [], + "bbox": [ + 220, + 348, + 408, + 463 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/d936d00b55af85cd3fe51ff366a7edc6a05aa8fe639f9a7a1557ecaba53e90b0.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 410, + 348, + 583, + 463 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/71408cd8124ebccd72cbaca806df1fa6aec9b530f4aeb22a3bdc4c309401d958.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 594, + 348, + 767, + 463 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Abstract. As 3D Gaussian Splitting (3DGS) provides fast and high-quality novel view synthesis, it is a natural extension to deform a canonical 3DGS to multiple frames for representing a dynamic scene. However, previous works fail to accurately reconstruct complex dynamic scenes. We attribute the failure to the design of the deformation field, which is built as a coordinate-based function. This approach is problematic because 3DGS is a mixture of multiple fields centered at the Gaussians, not just a single coordinate-based framework. To resolve this problem, we define the deformation as a function of per-Gaussian embeddings and temporal embeddings. Moreover, we decompose deformations as coarse and fine deformations to model slow and fast movements, respectively. Also, we introduce a local smoothness regularization for per-Gaussian embedding to improve the details in dynamic regions.", + "bbox": [ + 259, + 583, + 740, + 763 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Project page: https://jeongminb.github.io/e-d3dns/", + "bbox": [ + 269, + 765, + 614, + 777 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Keywords: Deformable gaussian splatting $\\cdot$ 4D scene reconstruction $\\cdot$ Novel view synthesis", + "bbox": [ + 259, + 791, + 740, + 819 + ], + "page_idx": 0 + }, + { + "type": "page_footnote", + "text": "* Authors contributed equally to this work.", + "bbox": [ + 217, + 825, + 511, + 839 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "1 Introduction", + "text_level": 1, + "bbox": [ + 217, + 143, + 374, + 160 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Dynamic scene reconstruction from multi-view input videos is an important task in computer vision, as it can be extended to various applications and industries such as mixed reality, content production, etc. Neural Radiance Fields (NeRF) [19], which enable photorealistic novel view synthesis from multi-view inputs, can represent dynamic scenes by modeling the scene with an additional time input [12, 22]. However, typical NeRFs require querying multilayer perceptron (MLP) for hundreds of points per camera ray, which limits rendering speed.", + "bbox": [ + 212, + 191, + 787, + 299 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "On the other hand, the recently emerging 3D Gaussian Splatting (3DGS) [10] has the advantage of real-time rendering compared to NeRFs using a differentiable rasterizer for 3D Gaussian primitives. 3DGS directly optimizes the parameters of 3D Gaussians (position, opacity, anisotropic covariance, and spherical harmonics coefficients) and renders them via projection and $\\alpha$ -blending. Since 3DGS has the characteristics of continuous volumetric radiance fields, some recent studies [3,8,14,29,31,32] represent dynamic scenes by defining a canonical 3DGS and deforming it to individual frames as deformable NeRFs [31] do. Specifically, they model the deformation as a function of 4D (x, y, z, t) coordinates with MLPs or grids to predict the change in the 3D Gaussian parameters.", + "bbox": [ + 212, + 301, + 787, + 452 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "However, since 3DGS is a mixTURE of multiple volumetric fields, it is not appropriate to model the deformation of Gaussian parameters with a single coordinate-based network to represent dynamic scenes. In addition, existing field-based approaches are constrained by the resolution of the grid which models the deformation field, the capacity of the model, or the frequencies of the input. As shown in Figure 1, existing study does not properly represent complex dynamic scenes, and even introducing an additional feature grid that is twice the maximum resolution has only a slight improvement in performance (See Appendix for more results). We alleviate this problem by introducing a novel dynamic representation to deform each Gaussian.", + "bbox": [ + 212, + 455, + 787, + 606 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "In this paper, we model the deformation of Gaussians at frames as 1) a function of a product space of per-Gaussian embeddings and temporal embeddings. We expect this rational design to bring quality improvement by precisely modeling different deformations of different Gaussians. Additionally, 2) We decompose temporal variations of the parameters into coarse and fine components, namely coarse-fine deformation. The coarse deformation represents large or slow movements in the scene, while fine deformation learns the fast or detailed movements that coarse deformation does not cover. Finally, we propose 3) a local smoothness regularization for per-Gaussian embedding to ensure the deformations of neighboring Gaussians are similar.", + "bbox": [ + 212, + 609, + 787, + 761 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "In our experiments, we observe that our per-Gaussian embeddings, coarse-fine deformation, and regularization improve the deformation quality. Our approach outperforms baselines in capturing fine details in dynamic regions and excels even under challenging camera settings. Additionally, our method also achieves fast rendering speed and relatively low capacity.", + "bbox": [ + 212, + 763, + 787, + 840 + ], + "page_idx": 1 + }, + { + "type": "page_number", + "text": "2", + "bbox": [ + 217, + 114, + 228, + 126 + ], + "page_idx": 1 + }, + { + "type": "header", + "text": "J. Bae et al.", + "bbox": [ + 271, + 114, + 354, + 126 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "2 Related Work", + "text_level": 1, + "bbox": [ + 215, + 143, + 387, + 160 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "In this section, we review methods for dynamic scene reconstruction that deform 3D canonical space and methods for reconstructing dynamic scenes utilizing dynamic 3D Gaussians. Afterward, we review methods that use embeddings and spatial relationships of Gaussians.", + "bbox": [ + 212, + 178, + 784, + 239 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Deforming 3D Canonical Space D-NeRF [22] reconstructs dynamic scenes by deforming ray samples over time, using the deformation network that takes 3D coordinates and timestamps of the sample as inputs. Nerfies [20] and HyperNeRF [21] use per-frame trainable deformation codes instead of time conditions to deform the canonical space. Instead of deforming from the canonical frame to the entire frames, HyperReel [1] deforms the ray sample of the keyframe to represent the intermediate frame. 4DGaussians [29] and D3DGS [31] reconstruct the dynamic scene with a deformation network which inputs the center position of the canonical 3D Gaussians and timestamps. MoDGS [15] learns the mapping between canonical space and space at specific timestamp through invertible MLP. In contrast, we demonstrate a novel deformation representation as a function of a product space of per-Gaussian latent embeddings and temporal embeddings.", + "bbox": [ + 212, + 262, + 787, + 445 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Dynamic 3D Gaussians To extend the fast rendering speed of 3D Gaussian Splitting [10] into dynamic scene reconstructions. 4D Gaussians [29] decodes features from multi-resolution HexPlanes [2] for temporal deformation of 3D Gaussians. While D3DGS [31] uses an implicit function that processes the time and location of the Gaussian. 4DGS [30] decomposes the 4D Gaussians into a time-conditioned 3D Gaussians and a marginal 1D Gaussians. STG [13] represents changes in 3D Gaussian over time through a temporal opacity and a polynomial function for each Gaussian.", + "bbox": [ + 212, + 468, + 787, + 587 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Our method uses deformable 3D Gaussians as 4D Gaussians [29] and D3DGS [31] do, but does not necessitate the separated feature field to obtain the input feature of the deformation decoder. Our approach uses embeddings allocated to each Gaussian and a temporal embedding shared within a specific frame.", + "bbox": [ + 212, + 589, + 787, + 650 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Latent Embedding on Novel View Synthesis Some studies incorporate latent embeddings to represent different states of the static and dynamic scene. NeRF-W [18] and Block-NeRF [25] employ per-image embeddings to capture different appearances of a scene, representing the scenes from in-the-wild image collections. DyNeRF and MixVoxels [12, 27] employ a temporal embedding for each frame to represent dynamic scenes. Nerfies [20] and HyperNeRF [21] incorporate both per-frame appearance and deformation embeddings. Sync-NeRF [11] introduces time offset to calibrate the misaligned temporal embeddings on dynamic scenes from unsynchronized videos. We introduce per-Gaussian latent embedding to encode the changes over time of each Gaussian and use temporal embeddings to represent different states in each frame of the scene.", + "bbox": [ + 212, + 672, + 787, + 839 + ], + "page_idx": 2 + }, + { + "type": "header", + "text": "E-D3DGS", + "bbox": [ + 661, + 114, + 730, + 126 + ], + "page_idx": 2 + }, + { + "type": "page_number", + "text": "3", + "bbox": [ + 774, + 116, + 784, + 126 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Considering Spatial Relationships of Gaussians Scaffold-GS [16] reconstruct 3D scenes by synthesizing Gaussians from anchors, utilizing that the neighboring Gaussians have similar properties. SAGS [26] creates a graph based on k-nearest neighbors (KNN) so that each Gaussian is optimized while considering its neighboring Gaussians. In dynamic scene reconstruction, SC-GS [9] and GaussianPrediction [34] deform Gaussians by combining the deformations of key point Gaussians. Dynamic 3D Gaussians [17] utilizes regularization to encourage that Gaussians and their neighboring Gaussians deform with local rigidity. Similarly, we propose a local smoothness regularization that encourages neighboring Gaussians to have similar embeddings, resulting in similar deformations.", + "bbox": [ + 212, + 146, + 787, + 297 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "3 Method", + "text_level": 1, + "bbox": [ + 215, + 321, + 330, + 338 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "In this section, we first provide a brief overview of 3D Gaussian Splatting (Section 3.1). Next, we introduce our overall framework, embedding-based deformation for Gaussians (Section 3.2) and coarse-fine deformation scheme consisting of coarse and fine deformation functions (Section 3.3). Finally, we present a local smoothness regularization for per-Gaussian embeddings to achieve better details on dynamic regions (Section 3.4).", + "bbox": [ + 212, + 354, + 787, + 446 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "3.1 Preliminary: 3D Gaussian Splitting", + "text_level": 1, + "bbox": [ + 215, + 468, + 555, + 484 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "3D Gaussian splattering [10] optimizes a set of anisotropic 3D Gaussians through differentiable tile rasterization to reconstruct a static 3D scene. By its efficient rasterization, the optimized model enables real-time rendering of high-quality images. Each 3D Gaussian kernel $G_{i}(x)$ at the point $x$ consists with position $\\mathbf{x}_i$ , rotation $R_i$ , and scale $S_i$ :", + "bbox": [ + 212, + 494, + 785, + 570 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\nG _ {i} (x) = e ^ {- \\frac {1}{2} (x - \\mathbf {x} _ {i}) ^ {T} \\Sigma_ {i} ^ {- 1} (x - \\mathbf {x} _ {i})}, \\quad \\text {w h e r e} \\quad \\Sigma_ {i} = R _ {i} S _ {i} S _ {i} ^ {T} R _ {i} ^ {T}. \\tag {1}\n$$\n", + "text_format": "latex", + "bbox": [ + 287, + 580, + 785, + 601 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "To projecting 3D Gaussians to 2D for rendering, covariance matrix $\\Sigma^{\\prime}$ are calculated by viewing transform $W$ and the Jacobian $J$ of the affine approximation of the projective transformation [35] as follows:", + "bbox": [ + 212, + 613, + 785, + 659 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\n\\Sigma^ {\\prime} = J W \\Sigma W ^ {T} J ^ {T}. \\tag {2}\n$$\n", + "text_format": "latex", + "bbox": [ + 434, + 670, + 784, + 686 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Blending $\\mathcal{N}$ depth-ordered projected points that overlap the pixel, the Gaussian kernel $G_{i}(x)$ is multiplied by the opacity of the Gaussian $\\sigma_{i}$ and calculates the pixel color $C$ with the color of the Gaussian $c_{i}$ :", + "bbox": [ + 212, + 699, + 785, + 744 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\nC = \\sum_ {i \\in \\mathcal {N}} c _ {i} \\alpha_ {i} \\prod_ {j = 1} ^ {i - 1} (1 - \\alpha_ {j}), \\quad \\text {w h e r e} \\quad \\alpha_ {i} = \\sigma_ {i} G _ {i} (x). \\tag {3}\n$$\n", + "text_format": "latex", + "bbox": [ + 316, + 757, + 785, + 799 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "The color of Gaussian $c_{i}$ is determined using the SH coefficient with view direction.", + "bbox": [ + 212, + 809, + 785, + 839 + ], + "page_idx": 3 + }, + { + "type": "page_number", + "text": "4", + "bbox": [ + 217, + 114, + 228, + 126 + ], + "page_idx": 3 + }, + { + "type": "header", + "text": "J. Bae et al.", + "bbox": [ + 271, + 114, + 354, + 126 + ], + "page_idx": 3 + }, + { + "type": "image", + "img_path": "images/5a6ada8e19479be6d76f7158e58487be4d65b99cfd38a791de1ac921bcd6b476.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 233, + 146, + 379, + 210 + ], + "page_idx": 4 + }, + { + "type": "image", + "img_path": "images/b92e21303c10ae6fcf13d58c84f42e4fa8bbf5dfcf03143fce17f66da01e6ef4.jpg", + "image_caption": [ + "(a) Embeddings for deformation" + ], + "image_footnote": [], + "bbox": [ + 238, + 215, + 385, + 290 + ], + "page_idx": 4 + }, + { + "type": "image", + "img_path": "images/0732f9a2f4bbb5c92e44341a0fd45613b07fcc21609fc5458e11d5f0abd3e5cd.jpg", + "image_caption": [ + "(b) Per-Gaussian deformation" + ], + "image_footnote": [], + "bbox": [ + 405, + 145, + 598, + 289 + ], + "page_idx": 4 + }, + { + "type": "image", + "img_path": "images/afd61294da1e2007c76213a87be27107281fbaa0c8329bae0a3df65ca12b0be2.jpg", + "image_caption": [ + "(c) Local smoothness regularization", + "Fig. 2: Framework. Existing coordinate-based network methods struggle to represent complex dynamic scenes. To this end, we define per-Gaussian deformation. (a) Firstly, we assign a latent embedding for each Gaussian. Additionally, we introduce coarse and fine temporal embeddings to represent the slow and fast state of the dynamic scene. (b) By employing two decoders that take per-Gaussian latent embeddings along with coarse and fine temporal embeddings as input, we estimate slow or large changes and fast or detailed changes to model the final deformation, respectively. (c) Finally, we introduce a local smoothness regularization so that the embeddings of neighboring Gaussians are similar." + ], + "image_footnote": [], + "bbox": [ + 622, + 148, + 754, + 289 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "3.2 Embedding-Based Deformation for Gaussians", + "text_level": 1, + "bbox": [ + 215, + 469, + 635, + 484 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Deformable NeRFs consist of a deformation field that predicts displacement $\\Delta \\mathbf{x}$ for a given coordinate $\\mathbf{x}$ from the canonical space to each target frame, and a radiance field that maps color and density from a given coordinate in the canonical space $(\\mathbf{x} + \\Delta \\mathbf{x})$ . Existing deformable Gaussian methods employ the same approach for predicting the deformation of Gaussians, i.e., utilizing a deformation field based on coordinates.", + "bbox": [ + 212, + 491, + 787, + 580 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Unlike previous methods, we start from the design of 3DGS: the 3D scene is represented as a mixture of Gaussians that have individual radiance fields. Accordingly, the deformation should be defined for each Gaussian. Based on this intuition, we introduce a function $\\mathcal{F}_{\\theta}$ that produces deformation from learnable embeddings $\\mathbf{z}_{\\mathrm{g}} \\in \\mathbb{R}^{32}$ belonging to individual Gaussians (Figure 2a), and typical temporal embeddings $\\mathbf{z}_t \\in \\mathbb{R}^{256}$ for different frames:", + "bbox": [ + 212, + 582, + 787, + 672 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\n\\mathcal {F} _ {\\theta}: \\left(\\mathbf {z} _ {\\mathrm {g}}, \\mathbf {z} _ {t}\\right)\\rightarrow \\left(\\Delta \\mathbf {x}, \\Delta \\mathbf {r}, \\Delta \\mathbf {s}, \\Delta \\sigma , \\Delta Y\\right), \\tag {4}\n$$\n", + "text_format": "latex", + "bbox": [ + 362, + 680, + 785, + 698 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "where $\\mathbf{r}$ is a rotation quaternion, $\\mathbf{s}$ is a vector for scaling, $\\sigma$ is an opacity, and $Y$ is SH coefficients for modeling view-dependent color. We implement $\\mathcal{F}_{\\theta}$ as a shallow multi-layer perceptron (MLP) followed by an MLP head for each parameter. As a result, the Gaussian parameters at frame $t$ are determined by adding $\\mathcal{F}_{\\theta}(\\mathbf{z}_{\\mathrm{g}},\\mathbf{z}_t)$ to the canonical Gaussian parameters (Figure 2c).", + "bbox": [ + 212, + 704, + 787, + 779 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "We jointly optimize the per-Gaussian embeddings $\\mathbf{z}_{\\mathrm{g}}$ , the deformation function $\\mathcal{F}_{\\theta}$ , and the canonical Gaussian parameters to minimize the rendering loss. We use the L1 and periodic DSSIM as the rendering loss between the rendered image and the ground truth image.", + "bbox": [ + 212, + 779, + 787, + 839 + ], + "page_idx": 4 + }, + { + "type": "header", + "text": "E-D3DGS", + "bbox": [ + 661, + 114, + 730, + 126 + ], + "page_idx": 4 + }, + { + "type": "page_number", + "text": "5", + "bbox": [ + 774, + 116, + 784, + 126 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "3.3 Coarse-Fine Deformation", + "text_level": 1, + "bbox": [ + 215, + 146, + 470, + 159 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Different parts of a scene may have coarse and fine motions [5]. E.g., a hand swiftly stirs a pan (fine) while a body slowly moves from left to right (coarse). Based on this intuition, we introduce a coarse-fine deformation that produces a summation of coarse and fine deformations.", + "bbox": [ + 212, + 172, + 782, + 231 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Coarse-fine deformation consists of two functions with the same architecture: one for coarse and one for fine deformation (Figure 2c). The functions receive different temporal embeddings as follows:", + "bbox": [ + 212, + 233, + 782, + 277 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Following typical temporal embeddings, we start from a 1D feature grid $Z \\in \\mathbb{R}^{N \\times 256}$ for $N$ frames and use an embedding $\\mathbf{z}_t^{\\mathrm{f}} = \\mathrm{interp}(Z,t)$ for fine deformation. For coarse deformation, we linearly downsample $Z$ by a factor of 5 to remove high-frequencies responsible for fast and detailed deformation. Then we compute $\\mathbf{z}_t^{\\mathrm{c}}$ as a linear interpolation of embeddings at enclosing grid points (Figure 2b).", + "bbox": [ + 214, + 277, + 784, + 369 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "As a result, coarse deformation $\\mathcal{F}_{\\theta_{\\mathrm{c}}}(\\mathbf{z}_{\\mathrm{g}},\\mathbf{z}_{t}^{c})$ is responsible for representing large or slow movements in the scene, while fine deformation $\\mathcal{F}_{\\theta_{\\mathrm{f}}}(\\mathbf{z}_{\\mathrm{g}},\\mathbf{z}_{t}^{\\mathrm{f}})$ learns the fast or detailed movements that coarse deformation does not cover. This improves the deformation quality. Refer to the Ablation study section for more details.", + "bbox": [ + 212, + 369, + 784, + 444 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "3.4 Local Smoothness Regularization", + "text_level": 1, + "bbox": [ + 215, + 468, + 534, + 483 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Neighboring Gaussians constructing dynamic objects tend to exhibit locally similar deformation. Inspired by [17], we introduce a local smoothness regularization for per-Gaussian embedding $\\mathbf{z}_{\\mathrm{g}}$ (Figure 2d) to encourage similar deformations between nearby Gaussians $i$ and $j$ :", + "bbox": [ + 212, + 494, + 782, + 555 + ], + "page_idx": 5 + }, + { + "type": "equation", + "text": "\n$$\n\\mathcal {L} ^ {\\mathrm {e m b} - \\mathrm {r e g}} = \\frac {1}{k | \\mathcal {S} |} \\sum_ {i \\in \\mathcal {S}} \\sum_ {j \\in \\mathrm {K N N} _ {i; k}} \\left(w _ {i, j} \\| \\mathbf {z} _ {\\mathrm {g} i} - \\mathbf {z} _ {\\mathrm {g} j} \\| _ {2}\\right),\n$$\n", + "text_format": "latex", + "bbox": [ + 331, + 565, + 669, + 602 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "where $w_{i,j} = \\exp (-\\lambda_w\\| \\mu_j - \\mu_i\\| _2^2)$ is the weighting factor and $\\mu$ is the Gaussian center. We set $\\lambda_w$ to 2000 and $k$ to 20 following [17]. To reduce the computational cost, we obtain sets of k-nearest-neighbors only when the densification occurs.", + "bbox": [ + 212, + 614, + 784, + 660 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Note that unlike previous approaches that directly constrain physical properties such as rigidity or rotation, We implicitly induce locally similar deformation by ensuring that per-Gaussian embeddings are locally smooth. Our regularization allows better capture of textures and details of dynamic objects.", + "bbox": [ + 212, + 662, + 782, + 722 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "4 Experiment", + "text_level": 1, + "bbox": [ + 215, + 746, + 366, + 763 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "In this section, we first describe the criterion for selection of baselines, and evaluation metrics. We then demonstrate the effectiveness of our method through comparisons with various baselines and datasets (Section 4.1-4.2). Finally, we conduct analysis and ablations of our method (Section 4.3).", + "bbox": [ + 212, + 779, + 784, + 839 + ], + "page_idx": 5 + }, + { + "type": "page_number", + "text": "6", + "bbox": [ + 217, + 114, + 228, + 126 + ], + "page_idx": 5 + }, + { + "type": "header", + "text": "J. Bae et al.", + "bbox": [ + 271, + 114, + 354, + 126 + ], + "page_idx": 5 + }, + { + "type": "table", + "img_path": "images/89503dc1b1a2b278ce5ed3c00e2d39577c28e5acca30f018646205806c32ccd0.jpg", + "table_caption": [ + "Table 1: Average performance in the test view on Neural 3D Video Dataset The computational cost was measured based on flame_salmon_1 on the A6000. ${}^{1}$ flame salmon scene includes only the first segment, comprising 300 frames. ${}^{2}$ reported time and DyNeRF is trained on 8 GPUs and tested only on flame salmon. ${}^{3}$ trained with 90 frames. ${}^{4}$ trained with 50 frames." + ], + "table_footnote": [], + "table_body": "
modelmetriccomputational cost
PSNR↑SSIM↑LPIPS↓Training time↓FPS↑Model size↓
DyNeRF12 [12]29.58-0.0831344 hours0.0156 MB
NeRFPlayer13 [24]30.69-0.1116 hours0.051654 MB
Mix Voxels [27]30.300.9180.1271 hours 40 mins0.93512 MB
K-Planes [6]30.860.9390.0961 hours 40 mins0.13309 MB
HyperReel4 [1]30.370.9210.1069 hours 20 mins1.01362 MB
4DGS [30]31.190.9400.0519 hours 30 mins33.78700 MB
4DGaussians [29]30.710.9350.05650 mins51.959 MB
Ours31.310.9450.0371 hours 52 mins74.535 MB
", + "bbox": [ + 274, + 219, + 722, + 340 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Baselines We choose the start-of-the-art method as a baseline in each dataset. We compared against DyNeRF, NeRFPlayer, MixVoxels, K-Planes, HyperReel, Nerfies, HyperNeRF, and TiNeuVox on the NeRF baseline. In detail, we use the version of NeRFPlayer TensoRF VM, HyperNeRF DF, Mixvoxels-L, K-Planes hybrid. We compared with 4DGaussians, 4DGS, and D3DGS based on the Gaussian baseline. Meanwhile, we have not included STG in our comparison due to its requirement for per-frame Structure from Motion (SfM) points, which makes conducting a fair comparison challenging. Also, STG is not a deformable 3D Gaussian approach. We followed the official code and configuration, except increasing the training iterations for the Technicolor dataset to 1.5 times that of the 4DGaussians, in comparison to the Neural 3D Video dataset.", + "bbox": [ + 212, + 368, + 787, + 534 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Metrics We report the quality of rendered images using PSNR, SSIM, and LPIPS. Peak Signal-to-Noise Ratio (PSNR) quantifies pixel color error between the rendered video and the ground truth. We utilize SSIM [28] to account for the perceived similarity of the rendered image. Additionally, we measure higher-level perceptual similarity using Learned Perceptual Image Patch Similarity (LPIPS) [33] with an AlexNet Backbone. Higher PSNR and SSIM values and lower LPIPS values indicate better visual quality.", + "bbox": [ + 212, + 554, + 787, + 660 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "4.1 Effectiveness on Dynamic Region", + "text_level": 1, + "bbox": [ + 215, + 680, + 535, + 696 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Neural 3D Video Dataset [12] includes 20 multi-view videos, with each scene consisting of either 300 frames, except for the flame_salmon scene, which comprises 1200 frames. These scenes encompass a relatively long duration and various movements, with some featuring multiple objects in motion. We utilized Neural 3D Video dataset to observe the capability to capture dynamic areas. Total six scenes (coffee_martini, cook_s spinach, cut_ roasted_beef, flame_salmon, flame_ steak, sear_ steak) are evaluated in Figure 3 and Table 1. The flame_salmon scene divided into four segments, each containing 300 frames.", + "bbox": [ + 212, + 703, + 787, + 839 + ], + "page_idx": 6 + }, + { + "type": "header", + "text": "E-D3DGS", + "bbox": [ + 661, + 114, + 730, + 126 + ], + "page_idx": 6 + }, + { + "type": "page_number", + "text": "7", + "bbox": [ + 774, + 116, + 784, + 126 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/a175a53e84dca37bd016a27fc78420aa86657301a2861a78a0e1deb900b85278.jpg", + "image_caption": [ + "Fig. 3: Qualitative comparisons on the Neural 3D Video Dataset." + ], + "image_footnote": [], + "bbox": [ + 225, + 148, + 781, + 494 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Table 1 presents quantitative metrics on the average metrics across the test views of all scenes, computational and storage costs on the first fragment of flame_salmon scene. Refer to the Appendix for per-scene details. Our method demonstrates superior reconstruction quality, FPS, and model size across compared to baselines. As the table shows, NeRF baselines generally required longer training and rendering times. While 4DGS shows relatively high reconstruction performance, it demands longer training times and larger VRAM storage capacity compared to other baselines. 4DGaussians requires lower computational and storage cost but it displays low reconstruction quality in some scenes with rapid dynamics, as shown in the teaser and Figure 3.", + "bbox": [ + 212, + 554, + 787, + 705 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Figure 3 reports the rendering quality. Our method successfully reconstructs the fine details in moving areas, outperforming baselines on average metrics across test views. Baselines show blurred dynamic areas or severe artifacts in low-light scenes such as cook_spinach and flame_steak. 4DGS exhibits the disappearance of some static areas. In 4DGaussians, a consistent over-smoothing occurs in dynamic areas. All baselines experienced reduced quality in reflective or thin dynamic areas like clamps, torches, and windows.", + "bbox": [ + 212, + 705, + 787, + 811 + ], + "page_idx": 7 + }, + { + "type": "page_number", + "text": "8", + "bbox": [ + 217, + 116, + 228, + 126 + ], + "page_idx": 7 + }, + { + "type": "header", + "text": "J. Bae et al.", + "bbox": [ + 271, + 114, + 354, + 126 + ], + "page_idx": 7 + }, + { + "type": "table", + "img_path": "images/3aa0e70ff4a26c3fd49461ceb8ceb37cbbfb5423f11426610880f896522f372e.jpg", + "table_caption": [ + "Table 2: Average performance in the test view on Technicolor Dataset" + ], + "table_footnote": [], + "table_body": "
modelmetriccomputational cost
PSNR↑SSIM↑LPIPS↓Training time↓FPS↑Model size↓
DyNeRF31.80-0.140-0.020.6 MB
HyperReel32.320.8990.1182 hours 45 mins0.91289 MB
4DGaussians29.620.8440.17625 mins34.851 MB
Ours33.240.9070.1002 hours 55 mins60.877 MB
", + "bbox": [ + 303, + 171, + 694, + 241 + ], + "page_idx": 8 + }, + { + "type": "image", + "img_path": "images/7b89d6665a098b61b9ff43f633b4ccaaeaac2f65d6bfac8f0a81fe4c2d88105d.jpg", + "image_caption": [ + "Fig. 4: Qualitative comparisons on the Technicolor Dataset." + ], + "image_footnote": [], + "bbox": [ + 263, + 268, + 738, + 516 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "Technicolor Light Field Dataset [23] is a multi-view dataset captured with a time-synchronized $4 \\times 4$ camera rig, containing intricate details. We train ours and the baselines on 50 frames of five commonly used scenes (Birthday, Fabien, Painter, Theater, Trains) using full-resolution videos at $2048 \\times 1088$ pixels, with the second row and second column cameras used as test views.", + "bbox": [ + 212, + 585, + 784, + 661 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "Table 2 reports the average metrics across the test views of all scenes, computational and storage costs on Painter scene. HyperReel demonstrates overall high-quality results but struggles with relatively slow training times and FPS, and a larger model size. 4DGaussians exhibits fast training times and FPS but significantly underperforms in reconstructing fine details compared to other baselines. However, our method demonstrates superior reconstruction quality and faster FPS compared to the baselines.", + "bbox": [ + 212, + 667, + 787, + 773 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "As shown in Figure 4, HyperReel produces noisy artifacts due to incorrect predictions of the displacement vector. 4D Gaussians fails to capture fine details in dynamic areas, exhibiting over-smoothing results. All baselines struggle to accurately reconstruct rapidly moving thin areas like fingers.", + "bbox": [ + 212, + 779, + 787, + 840 + ], + "page_idx": 8 + }, + { + "type": "header", + "text": "E-D3DGS", + "bbox": [ + 661, + 114, + 730, + 126 + ], + "page_idx": 8 + }, + { + "type": "page_number", + "text": "9", + "bbox": [ + 774, + 116, + 785, + 126 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "4.2 Scenes with Challenging Camera Setting", + "text_level": 1, + "bbox": [ + 215, + 146, + 598, + 162 + ], + "page_idx": 9 + }, + { + "type": "table", + "img_path": "images/59206c1fde35444860325d35db55ec4c7d9f13f6cce2f4cd894da96552013a02.jpg", + "table_caption": [ + "Table 3: Average performance in the test view on Hypernerf Dataset" + ], + "table_footnote": [], + "table_body": "
modelmetriccomputational cost
PSNR↑SSIM↑LPIPS↓Training time↓FPS↑Model size↓
Nerfies [20]22.23-0.170~ hours< 1-
HyperNeRF DS [21]22.290.5980.15332 hours< 115 MB
TiNeuVox [4]24.200.6160.3933 hours 30 mins148 MB
D3DGS [31]22.400.5980.27516 mins6.95309 MB
4DGaussians25.030.6820.28150 mins96.360 MB
Ours25.430.6970.2311 hours 15 mins139.333 MB
", + "bbox": [ + 303, + 220, + 697, + 306 + ], + "page_idx": 9 + }, + { + "type": "image", + "img_path": "images/0af1d5d643a009ddb43d392d4dda53c65c6d78129889c44152499a7f00721d7d.jpg", + "image_caption": [ + "Fig. 5: Qualitative comparisons on the HyperNeRF Dataset." + ], + "image_footnote": [], + "bbox": [ + 308, + 358, + 699, + 613 + ], + "page_idx": 9 + }, + { + "type": "text", + "text": "HyperNeRF dataset includes videos captured using two phones rigidly mounted on a handheld rig. We train on all frames of four scenes (3D Printer, Banana, Broom, Chicken) at a resolution downsampled by half to $536 \\times 960$ . Due to memory constraints, D3DGS is trained on images downsampled by a quarter.", + "bbox": [ + 212, + 688, + 797, + 750 + ], + "page_idx": 9 + }, + { + "type": "text", + "text": "The table shows that our method outperforms the reconstruction performance with previous methods along with compact model size and faster FPS. Figure 5 shows that previous methods struggle to reconstruct fast-moving parts such as fingers and broom. Especially D3DGS deteriorates in Broom scene. Table 3 reports the average metrics across the test views of all scenes, computational and storage costs on Broom scene.", + "bbox": [ + 212, + 750, + 787, + 839 + ], + "page_idx": 9 + }, + { + "type": "page_number", + "text": "10", + "bbox": [ + 217, + 114, + 235, + 126 + ], + "page_idx": 9 + }, + { + "type": "header", + "text": "J. Bae et al.", + "bbox": [ + 271, + 114, + 354, + 126 + ], + "page_idx": 9 + }, + { + "type": "text", + "text": "4.3 Analyses and Ablation Study", + "text_level": 1, + "bbox": [ + 215, + 146, + 503, + 161 + ], + "page_idx": 10 + }, + { + "type": "image", + "img_path": "images/554182633416a77d1122a24ee13aca9acf78d3e7386f4b74f0a86321b399cf3f.jpg", + "image_caption": [ + "(a) canonical rendering" + ], + "image_footnote": [], + "bbox": [ + 228, + 202, + 362, + 292 + ], + "page_idx": 10 + }, + { + "type": "image", + "img_path": "images/ef95661573a08819491da6e91681cbeac3e0e8a019056f9fd124f980af81b83d.jpg", + "image_caption": [ + "(b) coarse deform only" + ], + "image_footnote": [], + "bbox": [ + 366, + 202, + 500, + 292 + ], + "page_idx": 10 + }, + { + "type": "image", + "img_path": "images/f50c0b5dffd6ba9cf2bff68b59325df636266e36b8d7c187275bfb9149d0f806.jpg", + "image_caption": [ + "(c) fine deform only", + "Fig. 6: Deformation components. (a) The canonical space contains Gaussians to represent all target frames of the scene. (b) Applying coarse deformation to the canonical space roughly reflects the dynamics of the scene. (c) The rendering without coarse deformation and only with fine deformation looks similar to the canonical rendering, i.e., responsible for fine deformations. (d) Applying both coarse and fine deformation yields natural rendering results." + ], + "image_footnote": [], + "bbox": [ + 504, + 202, + 637, + 292 + ], + "page_idx": 10 + }, + { + "type": "image", + "img_path": "images/18b121f78a54ca89cbabdf980caff35cc0c5327d1255bd859b62f672dc88a3be.jpg", + "image_caption": [ + "(d) full rendering" + ], + "image_footnote": [], + "bbox": [ + 640, + 202, + 774, + 292 + ], + "page_idx": 10 + }, + { + "type": "image", + "img_path": "images/fb76741ca4e880981dcf037fa4e559ba30b0687e2bde18d7bc9053a7b61b7084.jpg", + "image_caption": [ + "Frame 67", + "Fig.7: Visualization of the magnitude of deformation. Coarse deformation (blue) captures large and slow changes, such as the movement of the head and torso, while fine deformation (red) is responsible for the fast and detailed movements of arms, tongs, shadows, etc." + ], + "image_footnote": [], + "bbox": [ + 254, + 465, + 498, + 566 + ], + "page_idx": 10 + }, + { + "type": "image", + "img_path": "images/50d55bb686db8f59500bd3aa1cf5cfb1e87af5d1b32675153dff4ad99b305459.jpg", + "image_caption": [ + "Frame 185" + ], + "image_footnote": [], + "bbox": [ + 506, + 465, + 750, + 566 + ], + "page_idx": 10 + }, + { + "type": "text", + "text": "Deformation components In Figure 6, we present an analysis of the coarse-fine deformation. To achieve this, we render a flame_steak scene by omitting each of our deformation components one by one. Our full rendering results from adding coarse and fine deformation to the canonical space (Figure 6d). When both are removed, rendering yields an image in canonical space (Figure 6a). Rendering with the coarse deformation, which handles large or slow changes in the scene, produces results similar to the full rendering (Figure 6b). On the other hand, fine deformation is responsible for fast or detailed changes in the scene, yielding rendering similar to canonical space (Figure 6c).", + "bbox": [ + 212, + 704, + 787, + 840 + ], + "page_idx": 10 + }, + { + "type": "header", + "text": "E-D3DGS", + "bbox": [ + 661, + 114, + 730, + 126 + ], + "page_idx": 10 + }, + { + "type": "page_number", + "text": "11", + "bbox": [ + 767, + 116, + 784, + 126 + ], + "page_idx": 10 + }, + { + "type": "image", + "img_path": "images/2bca6504c44f69bb82078709672f279a3f0355b578947c07ed64af8cc70a0d42.jpg", + "image_caption": [ + "(a) Ours Full" + ], + "image_footnote": [], + "bbox": [ + 230, + 146, + 362, + 205 + ], + "page_idx": 11 + }, + { + "type": "image", + "img_path": "images/e0e64da9a3a3a1ee90ae7561f2342964b90def5092752639732998512bd40911.jpg", + "image_caption": [ + "(b) coarse deform only", + "Fig. 8: Qualitative ablation results on coarse-fine deformation. (a) Our model achieves clear results with both coarse and fine decoders. (b-c) The quality of dynamic areas decreases if one is missing. (d) Additionally, introducing the coordinates of Gaussian as an additional input into our decoders results in a decrease in the quality of both static and dynamic regions." + ], + "image_footnote": [], + "bbox": [ + 367, + 146, + 500, + 205 + ], + "page_idx": 11 + }, + { + "type": "image", + "img_path": "images/4395843b08f134c7c78a9cdf10b0cf79b6f072e54949ccd9e6aa5029378afb78.jpg", + "image_caption": [ + "(c) fine deform only" + ], + "image_footnote": [], + "bbox": [ + 504, + 146, + 635, + 205 + ], + "page_idx": 11 + }, + { + "type": "image", + "img_path": "images/bf4ba8ceca3ad52abc64a7819dd90ae4f1080d641cecd2f61b3272451caa9f09.jpg", + "image_caption": [ + "(d) Ours $+\\mathbf{x}$ injection" + ], + "image_footnote": [], + "bbox": [ + 640, + 146, + 772, + 205 + ], + "page_idx": 11 + }, + { + "type": "table", + "img_path": "images/8121f6ebd9c0cb9ce637c36f4b3a5ff9de4d9de0d967689cb7b984f00bca8fc1.jpg", + "table_caption": [ + "Table 4: Quantitative ablation results on coarse-fine deformation." + ], + "table_footnote": [], + "table_body": "
MethodPSNR↑SSIM↑LPIPS↓
Ours29.700.9330.041
coarse deformation only29.480.9310.044
fine deformation only29.230.9320.043
Ours + x injection29.600.9310.045
", + "bbox": [ + 303, + 344, + 694, + 419 + ], + "page_idx": 11 + }, + { + "type": "text", + "text": "To examine the roles of the coarse and fine deformation in the coarse-fine deformation, we conduct a visualization on flame_steak scene. First, we compute the Euclidean norm of positional shifts between the current and subsequent frames. We then add the value to the DC components of the SH coefficients proportionally to the magnitude: blue for coarse deformation and red for fine deformation. For visual clarity, we render the original scene in grayscale. As illustrated in Figure 7, coarse deformation models slower changes such as body movement, while fine deformation models faster movements like cooking arms. Thus, we demonstrate that by downsampling the temporal embedding grid $Z$ , we can effectively separate and model slow and fast deformations in the scene.", + "bbox": [ + 212, + 450, + 787, + 602 + ], + "page_idx": 11 + }, + { + "type": "text", + "text": "Ablation study We report the results of an ablation study on the deformation decoder in Figure 8 and Table 4. First, our full method (using both coarse and fine decoders) produces clear rendering results and models dynamic variations well (Figure 8a). Training only with the coarse or fine decoder leads to blurred dynamic areas and a failure to accurately capture detailed motion (Figure 8b-c). Additionally, we demonstrate experiments where the Gaussian center coordinates $\\mathbf{x}$ are injected into the input of each decoder. As shown in Figure 8d, including the Gaussian coordinates degrades the quality of deformation, supporting our argument that coordinate dependency should be removed from the deformation function.", + "bbox": [ + 212, + 627, + 787, + 777 + ], + "page_idx": 11 + }, + { + "type": "text", + "text": "Furthermore, we report the results of an ablation study on the local smoothness regularization for per-Gaussian embeddings. As shown in Figure 9, our regularization improves the details and texture quality of moving objects. In Table 5, we show a performance comparison between the proposed regularization and", + "bbox": [ + 212, + 779, + 787, + 840 + ], + "page_idx": 11 + }, + { + "type": "page_number", + "text": "12", + "bbox": [ + 217, + 114, + 235, + 126 + ], + "page_idx": 11 + }, + { + "type": "header", + "text": "J. Bae et al.", + "bbox": [ + 271, + 114, + 354, + 126 + ], + "page_idx": 11 + }, + { + "type": "table", + "img_path": "images/54606bd904be834a2aca183afcb1f51d7c13b0aca4866e91582c9c7c0eabaf92.jpg", + "table_caption": [ + "Table 5: Quantitative ablation results on local smoothness regularization. We compare the performance of applying our regularization with the physically-based regularization of Dynamic 3D Gaussians [17]. Our regularization better captures details of dynamic objects." + ], + "table_footnote": [], + "table_body": "
MethodPSNR↑SSIM↑LPIPS↓
Ours w/o embedding reg32.260.9510.037
+ our embedding reg32.340.9520.036
+ physically-based reg32.080.9500.036
", + "bbox": [ + 289, + 212, + 709, + 273 + ], + "page_idx": 12 + }, + { + "type": "image", + "img_path": "images/d3bae3efa4e225290e81c2d17779d77a3ca69c066c77d6dc77f115a31d42a4d0.jpg", + "image_caption": [ + "Ours full" + ], + "image_footnote": [], + "bbox": [ + 279, + 290, + 385, + 356 + ], + "page_idx": 12 + }, + { + "type": "image", + "img_path": "images/c36f81fa9ab94dd9ac95aacb559f7ef87d408ee6efb558c2020798abcf59085f.jpg", + "image_caption": [ + "Ours w/o reg", + "Fig. 9: Qualitative ablation results on local smoothness regularization." + ], + "image_footnote": [], + "bbox": [ + 388, + 291, + 493, + 356 + ], + "page_idx": 12 + }, + { + "type": "image", + "img_path": "images/8b07b1a8212952eb21f2114145017091c1187a1b426506870859cc4dab973fbe.jpg", + "image_caption": [ + "Ours full" + ], + "image_footnote": [], + "bbox": [ + 508, + 290, + 614, + 356 + ], + "page_idx": 12 + }, + { + "type": "image", + "img_path": "images/06618c98b077500de0b23326417919b4bc8618a7dc9b5629c67c1a56fee04333.jpg", + "image_caption": [ + "Ours w/o reg" + ], + "image_footnote": [], + "bbox": [ + 616, + 291, + 723, + 356 + ], + "page_idx": 12 + }, + { + "type": "text", + "text": "the existing physically-based regularizations. To apply the method proposed by Luiten et al. [17] to ours, we make some modifications: 1) Like our method, we find the set of k-nearest-neighbours only when the densification occurs to reduce the computational cost. 2) For long-term local-isometry loss, we use the time of video frame used in previous training step instead of using the time of the first frame. Our regularization is simple and shows better performance compared to previous method.", + "bbox": [ + 212, + 429, + 787, + 536 + ], + "page_idx": 12 + }, + { + "type": "text", + "text": "5 Conclusion and Limitation", + "text_level": 1, + "bbox": [ + 214, + 558, + 509, + 574 + ], + "page_idx": 12 + }, + { + "type": "text", + "text": "We propose a per-Gaussian deformation for 3DGS that takes per-Gaussian embeddings as input, instead of using the typical deformation fields from previous deformable 3DGS works, resulting in high performance. We enhance the reconstruction quality by decomposing the dynamic changes into coarse and fine deformation. However, our method learns inappropriate Gaussian deformation with casually captured monocular videos [7], like other baselines. We plan to address it in future work by introducing useful prior for monocular video settings.", + "bbox": [ + 212, + 590, + 787, + 696 + ], + "page_idx": 12 + }, + { + "type": "image", + "img_path": "images/6244d426dee9c97c7a6af5d392b85dc2e62d9930adb0f449e27e8ee2e9bf65d4.jpg", + "image_caption": [ + "4DGaussians", + "Fig. 10: Limitation. Ours struggles with the casually captured monocular videos." + ], + "image_footnote": [], + "bbox": [ + 253, + 738, + 372, + 797 + ], + "page_idx": 12 + }, + { + "type": "image", + "img_path": "images/f4931a4dd7763d923edda247a6db3fd6060e5ba8252a115078cb119a2e7cfd80.jpg", + "image_caption": [ + "Ours" + ], + "image_footnote": [], + "bbox": [ + 375, + 738, + 496, + 797 + ], + "page_idx": 12 + }, + { + "type": "image", + "img_path": "images/6ac9a1640c0e847e7259dda48c198e6168348e5d8b5c2bc45e4371ad1b744732.jpg", + "image_caption": [ + "4DGaussians" + ], + "image_footnote": [], + "bbox": [ + 506, + 738, + 627, + 797 + ], + "page_idx": 12 + }, + { + "type": "image", + "img_path": "images/6097d5b27e045786fd13e8636bf6c77675c427cc4df28eedc04e4c0608ef8c21.jpg", + "image_caption": [ + "Ours" + ], + "image_footnote": [], + "bbox": [ + 629, + 738, + 750, + 797 + ], + "page_idx": 12 + }, + { + "type": "header", + "text": "E-D3DGS", + "bbox": [ + 661, + 114, + 730, + 126 + ], + "page_idx": 12 + }, + { + "type": "page_number", + "text": "13", + "bbox": [ + 767, + 114, + 785, + 126 + ], + "page_idx": 12 + }, + { + "type": "text", + "text": "Acknowledgements", + "text_level": 1, + "bbox": [ + 217, + 143, + 401, + 162 + ], + "page_idx": 13 + }, + { + "type": "text", + "text": "This work is supported by the Institute for Information & Communications Technology Planning & Evaluation (IITP) grant funded by the Korea government(MSIT) (No. 2017-0-00072, Development of Audio/Video Coding and Light Field Media Fundamental Technologies for Ultra Realistic Tera-media)", + "bbox": [ + 215, + 175, + 785, + 234 + ], + "page_idx": 13 + }, + { + "type": "text", + "text": "References", + "text_level": 1, + "bbox": [ + 217, + 258, + 321, + 273 + ], + "page_idx": 13 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "1. Attal, B., Huang, J.B., Richardt, C., Zollhoefer, M., Kopf, J., OaAZToole, M., Kim, C.: Hyperreel: High-fidelity 6-dof video with ray-conditioned sampling. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 16610-16620 (2023)", + "2. Cao, A., Johnson, J.: Hexplane: A fast representation for dynamic scenes. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 130-141 (2023)", + "3. Duisterhof, B.P., Mandi, Z., Yao, Y., Liu, J.W., Shou, M.Z., Song, S., Ichnowski, J.: Md-splatting: Learning metric deformation from 4d gaussians in highly deformable scenes. arXiv preprint arXiv:2312.00583 (2023)", + "4. Fang, J., Yi, T., Wang, X., Xie, L., Zhang, X., Liu, W., Nießner, M., Tian, Q.: Fast dynamic radiance fields with time-aware neural voxels. In: SIGGRAPH Asia 2022 Conference Papers (2022)", + "5. Feichtenhofer, C., Fan, H., Malik, J., He, K.: Slowfast networks for video recognition. In: Proceedings of the IEEE/CVF international conference on computer vision. pp. 6202-6211 (2019)", + "6. Fridovich-Keil, S., Meanti, G., Warburg, F.R., Recht, B., Kanazawa, A.: K-planes: Explicit radiance fields in space, time, and appearance. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 12479-12488 (2023)", + "7. Gao, H., Li, R., Tulsiani, S., Russell, B., Kanazawa, A.: Monocular dynamic view synthesis: A reality check. In: NeurIPS (2022)", + "8. Huang, Y.H., Sun, Y.T., Yang, Z., Lyu, X., Cao, Y.P., Qi, X.: Sc-gs: Sparse-controlled gaussian splatting for editable dynamic scenes. arXiv preprint arXiv:2312.14937 (2023)", + "9. Huang, Y.H., Sun, Y.T., Yang, Z., Lyu, X., Cao, Y.P., Qi, X.: Sc-gs: Sparse-controlled gaussian splatting for editable dynamic scenes. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR). pp. 4220-4230 (June 2024)", + "0. Kerbl, B., Kopanas, G., Leimkuhler, T., Drettakis, G.: 3d gaussian splatting for real-time radiance field rendering. ACM Transactions on Graphics 42(4) (July 2023), https://repo-sam.inria.fr/fungraph/3d-gaussian-splatting/", + "1. Kim, S., Bae, J., Yun, Y., Lee, H., Bang, G., Uh, Y.: Sync-nerf: Generalizing dynamic nerfs to unsynchronized videos. arXiv preprint arXiv:2310.13356 (2023)", + "12. Li, T., Slavcheva, M., Zollhoefer, M., Green, S., Lassner, C., Kim, C., Schmidt, T., Lovegrove, S., Goesele, M., Newcombe, R., et al.: Neural 3d video synthesis from multi-view video. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (2022)", + "13. Li, Z., Chen, Z., Li, Z., Xu, Y.: Spacetime gaussian feature splatting for real-time dynamic view synthesis. arXiv preprint arXiv:2312.16812 (2023)" + ], + "bbox": [ + 218, + 287, + 785, + 839 + ], + "page_idx": 13 + }, + { + "type": "page_number", + "text": "14", + "bbox": [ + 217, + 114, + 235, + 126 + ], + "page_idx": 13 + }, + { + "type": "header", + "text": "J. Bae et al.", + "bbox": [ + 271, + 114, + 354, + 126 + ], + "page_idx": 13 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "14. Liang, Y., Khan, N., Li, Z., Nguyen-Phuoc, T., Lanman, D., Tompkin, J., Xiao, L.: Gaufre: Gaussian deformation fields for real-time dynamic novel view synthesis (2023)", + "15. Liu, Q., Liu, Y., Wang, J., Lv, X., Wang, P., Wang, W., Hou, J.: Modgs: Dynamic gaussian splatting from causally-captured monocular videos (2024), https:// arxiv.org/abs/2406.00434", + "16. Lu, T., Yu, M., Xu, L., Xiangli, Y., Wang, L., Lin, D., Dai, B.: Scaffold-gs: Structured 3d gaussians for view-adaptive rendering. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 20654-20664 (2024)", + "17. Luiten, J., Kopanas, G., Leibe, B., Ramanan, D.: Dynamic 3d gaussians: Tracking by persistent dynamic view synthesis. In: 3DV (2024)", + "18. Martin-Brualla, R., Radwan, N., Sajjadi, M.S.M., Barron, J.T., Dosovitskiy, A., Duckworth, D.: NeRF in the Wild: Neural Radiance Fields for Unconstrained Photo Collections. In: CVPR (2021)", + "19. Mildenhall, B., Srinivasan, P.P., Tancik, M., Barron, J.T., Ramamoorthi, R., Ng, R.: Nerf: Representing scenes as neural radiance fields for view synthesis. Communications of the ACM 65(1), 99-106 (2021)", + "20. Park, K., Sinha, U., Barron, J.T., Bouaziz, S., Goldman, D.B., Seitz, S.M., Martin-Brualla, R.: Nerfies: Deformable neural radiance fields. ICCV (2021)", + "21. Park, K., Sinha, U., Hedman, P., Barron, J.T., Bouaziz, S., Goldman, D.B., Martin-Brualla, R., Seitz, S.M.: Hypernerf: A higher-dimensional representation for topologically varying neural radiance fields. arXiv preprint arXiv:2106.13228 (2021)", + "22. Pumarola, A., Corona, E., Pons-Moll, G., Moreno-Noguer, F.: D-nerf: Neural radiance fields for dynamic scenes. arXiv preprint arXiv:2011.13961 (2020)", + "23. Sabater, N., Boisson, G., Vandame, B., Kerbiriou, P., Babon, F., Hog, M., Gendrot, R., Langlois, T., Bureller, O., Schubert, A., et al.: Dataset and pipeline for multiview light-field video. In: Proceedings of the IEEE conference on computer vision and pattern recognition Workshops. pp. 30-40 (2017)", + "24. Song, L., Chen, A., Li, Z., Chen, Z., Chen, L., Yuan, J., Xu, Y., Geiger, A.: Nerflayer: A streamable dynamic scene representation with decomposed neural radiance fields. IEEE Transactions on Visualization and Computer Graphics 29(5), 2732-2742 (2023)", + "25. Tancik, M., Casser, V., Yan, X., Pradhan, S., Mildenhall, B., Srinivasan, P.P., Barron, J.T., Kretzschmar, H.: Block-nerf: Scalable large scene neural view synthesis (2022)", + "26. Ververas, E., Potamias, R.A., Song, J., Deng, J., Zafeiriou, S.: Sags: Structure-aware 3d gaussian splatting. arXiv:2404.19149 (2024)", + "27. Wang, F., Tan, S., Li, X., Tian, Z., Liu, H.: Mixed neural voxels for fast multi-view video synthesis. arXiv preprint arXiv:2212.00190 (2022)", + "28. Wang, Z., Bovik, A.C., Sheikh, H.R., Simoncelli, E.P.: Image quality assessment: from error visibility to structural similarity. IEEE transactions on image processing 13(4), 600-612 (2004)", + "29. Wu, G., Yi, T., Fang, J., Xie, L., Zhang, X., Wei, W., Liu, W., Tian, Q., Xinggang, W.: 4d gaussian splatting for real-time dynamic scene rendering. arXiv preprint arXiv:2310.08528 (2023)", + "30. Yang, Z., Yang, H., Pan, Z., Zhang, L.: Real-time photorealistic dynamic scene representation and rendering with 4d gaussian splatting. In: International Conference on Learning Representations (ICLR) (2024)", + "31. Yang, Z., Gao, X., Zhou, W., Jiao, S., Zhang, Y., Jin, X.: Deformable 3d gaussians for high-fidelity monocular dynamic scene reconstruction. arXiv preprint arXiv:2309.13101 (2023)" + ], + "bbox": [ + 212, + 146, + 787, + 839 + ], + "page_idx": 14 + }, + { + "type": "header", + "text": "E-D3DGS", + "bbox": [ + 661, + 114, + 730, + 126 + ], + "page_idx": 14 + }, + { + "type": "page_number", + "text": "15", + "bbox": [ + 767, + 116, + 785, + 126 + ], + "page_idx": 14 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "32. Yu, H., Julin, J., Milacski, Z.A., Niinuma, K., Jeni, L.A.: Cogs: Controllable gaussian splatting (2023)", + "33. Zhang, R., Isola, P., Efros, A.A., Shechtman, E., Wang, O.: The unreasonable effectiveness of deep features as a perceptual metric. In: Proceedings of the IEEE conference on computer vision and pattern recognition. pp. 586-595 (2018)", + "34. Zhao, B., Li, Y., Sun, Z., Zeng, L., Shen, Y., Ma, R., Zhang, Y., Bao, H., Cui, Z.: Gaussian prediction: Dynamic 3d gaussian prediction for motion extrapolation and free view synthesis. In: Special Interest Group on Computer Graphics and Interactive Techniques Conference Papers aA24. SIGGRAPH aAZ24, ACM (Jul 2024). https://doi.org/10.1145/3641519.3657417, http://dx.doi.org/10.1145/3641519.3657417", + "35. Zwicker, M., Pfister, H., Van Baar, J., Gross, M.: Surface splattering. In: Proceedings of the 28th annual conference on Computer graphics and interactive techniques. pp. 371-378 (2001)" + ], + "bbox": [ + 215, + 146, + 784, + 340 + ], + "page_idx": 15 + }, + { + "type": "page_number", + "text": "16", + "bbox": [ + 217, + 114, + 235, + 126 + ], + "page_idx": 15 + }, + { + "type": "header", + "text": "J. Bae et al.", + "bbox": [ + 271, + 114, + 354, + 126 + ], + "page_idx": 15 + } +] \ No newline at end of file diff --git a/2024/Per-Gaussian Embedding-Based Deformation for Deformable 3D Gaussian Splatting/8af14cc6-a852-4c45-9f34-dffb20b17a0a_model.json b/2024/Per-Gaussian Embedding-Based Deformation for Deformable 3D Gaussian Splatting/8af14cc6-a852-4c45-9f34-dffb20b17a0a_model.json new file mode 100644 index 0000000000000000000000000000000000000000..99fe80ddfde2e4494452c7433b1441298380e5c8 --- /dev/null +++ b/2024/Per-Gaussian Embedding-Based Deformation for Deformable 3D Gaussian Splatting/8af14cc6-a852-4c45-9f34-dffb20b17a0a_model.json @@ -0,0 +1,2333 @@ +[ + [ + { + "type": "title", + "bbox": [ + 0.224, + 0.141, + 0.78, + 0.187 + ], + "angle": 0, + "content": "Per-Gaussian Embedding-Based Deformation for Deformable 3D Gaussian Splitting" + }, + { + "type": "text", + "bbox": [ + 0.312, + 0.212, + 0.69, + 0.245 + ], + "angle": 0, + "content": "Jeongmin Bae1*, Seoha Kim1*, Youngsik Yun1, Hahyun Lee2, Gun Bang2, and Youngjung Uh1" + }, + { + "type": "text", + "bbox": [ + 0.3, + 0.255, + 0.704, + 0.283 + ], + "angle": 0, + "content": "1 Yonsei University, Seoul 03722, Korea {jaymin.bae, hailey07, bbangsik, yj.uh}@yonsei.ac.kr" + }, + { + "type": "text", + "bbox": [ + 0.232, + 0.283, + 0.771, + 0.311 + ], + "angle": 0, + "content": "\\(^{2}\\) Electronics and Telecommunications Research Institute, Daejeon 34129, Korea {hanilee, gbang}@etri.re.kr" + }, + { + "type": "image", + "bbox": [ + 0.221, + 0.349, + 0.409, + 0.464 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.411, + 0.349, + 0.584, + 0.464 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.595, + 0.349, + 0.768, + 0.464 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.214, + 0.476, + 0.788, + 0.532 + ], + "angle": 0, + "content": "Fig. 1: Overview. (a) Existing deformable 3D Gaussian Splitting methods show blurry results in complex dynamic scenes, even with deformation fields using finer feature grids. (b) Our model solves the problem by employing per-Gaussian latent embeddings to predict deformations for each Gaussian and achieves a clearer results." + }, + { + "type": "text", + "bbox": [ + 0.261, + 0.584, + 0.741, + 0.765 + ], + "angle": 0, + "content": "Abstract. As 3D Gaussian Splitting (3DGS) provides fast and high-quality novel view synthesis, it is a natural extension to deform a canonical 3DGS to multiple frames for representing a dynamic scene. However, previous works fail to accurately reconstruct complex dynamic scenes. We attribute the failure to the design of the deformation field, which is built as a coordinate-based function. This approach is problematic because 3DGS is a mixture of multiple fields centered at the Gaussians, not just a single coordinate-based framework. To resolve this problem, we define the deformation as a function of per-Gaussian embeddings and temporal embeddings. Moreover, we decompose deformations as coarse and fine deformations to model slow and fast movements, respectively. Also, we introduce a local smoothness regularization for per-Gaussian embedding to improve the details in dynamic regions." + }, + { + "type": "text", + "bbox": [ + 0.27, + 0.766, + 0.615, + 0.779 + ], + "angle": 0, + "content": "Project page: https://jeongminb.github.io/e-d3dns/" + }, + { + "type": "text", + "bbox": [ + 0.261, + 0.792, + 0.741, + 0.82 + ], + "angle": 0, + "content": "Keywords: Deformable gaussian splatting \\(\\cdot\\) 4D scene reconstruction \\(\\cdot\\) Novel view synthesis" + }, + { + "type": "page_footnote", + "bbox": [ + 0.218, + 0.826, + 0.512, + 0.84 + ], + "angle": 0, + "content": "* Authors contributed equally to this work." + } + ], + [ + { + "type": "page_number", + "bbox": [ + 0.218, + 0.116, + 0.23, + 0.127 + ], + "angle": 0, + "content": "2" + }, + { + "type": "header", + "bbox": [ + 0.272, + 0.115, + 0.355, + 0.127 + ], + "angle": 0, + "content": "J. Bae et al." + }, + { + "type": "title", + "bbox": [ + 0.218, + 0.145, + 0.375, + 0.161 + ], + "angle": 0, + "content": "1 Introduction" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.193, + 0.788, + 0.3 + ], + "angle": 0, + "content": "Dynamic scene reconstruction from multi-view input videos is an important task in computer vision, as it can be extended to various applications and industries such as mixed reality, content production, etc. Neural Radiance Fields (NeRF) [19], which enable photorealistic novel view synthesis from multi-view inputs, can represent dynamic scenes by modeling the scene with an additional time input [12, 22]. However, typical NeRFs require querying multilayer perceptron (MLP) for hundreds of points per camera ray, which limits rendering speed." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.302, + 0.788, + 0.453 + ], + "angle": 0, + "content": "On the other hand, the recently emerging 3D Gaussian Splatting (3DGS) [10] has the advantage of real-time rendering compared to NeRFs using a differentiable rasterizer for 3D Gaussian primitives. 3DGS directly optimizes the parameters of 3D Gaussians (position, opacity, anisotropic covariance, and spherical harmonics coefficients) and renders them via projection and \\(\\alpha\\)-blending. Since 3DGS has the characteristics of continuous volumetric radiance fields, some recent studies [3,8,14,29,31,32] represent dynamic scenes by defining a canonical 3DGS and deforming it to individual frames as deformable NeRFs [31] do. Specifically, they model the deformation as a function of 4D (x, y, z, t) coordinates with MLPs or grids to predict the change in the 3D Gaussian parameters." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.457, + 0.788, + 0.607 + ], + "angle": 0, + "content": "However, since 3DGS is a mixTURE of multiple volumetric fields, it is not appropriate to model the deformation of Gaussian parameters with a single coordinate-based network to represent dynamic scenes. In addition, existing field-based approaches are constrained by the resolution of the grid which models the deformation field, the capacity of the model, or the frequencies of the input. As shown in Figure 1, existing study does not properly represent complex dynamic scenes, and even introducing an additional feature grid that is twice the maximum resolution has only a slight improvement in performance (See Appendix for more results). We alleviate this problem by introducing a novel dynamic representation to deform each Gaussian." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.611, + 0.788, + 0.762 + ], + "angle": 0, + "content": "In this paper, we model the deformation of Gaussians at frames as 1) a function of a product space of per-Gaussian embeddings and temporal embeddings. We expect this rational design to bring quality improvement by precisely modeling different deformations of different Gaussians. Additionally, 2) We decompose temporal variations of the parameters into coarse and fine components, namely coarse-fine deformation. The coarse deformation represents large or slow movements in the scene, while fine deformation learns the fast or detailed movements that coarse deformation does not cover. Finally, we propose 3) a local smoothness regularization for per-Gaussian embedding to ensure the deformations of neighboring Gaussians are similar." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.765, + 0.788, + 0.842 + ], + "angle": 0, + "content": "In our experiments, we observe that our per-Gaussian embeddings, coarse-fine deformation, and regularization improve the deformation quality. Our approach outperforms baselines in capturing fine details in dynamic regions and excels even under challenging camera settings. Additionally, our method also achieves fast rendering speed and relatively low capacity." + } + ], + [ + { + "type": "header", + "bbox": [ + 0.663, + 0.115, + 0.732, + 0.127 + ], + "angle": 0, + "content": "E-D3DGS" + }, + { + "type": "page_number", + "bbox": [ + 0.775, + 0.117, + 0.785, + 0.127 + ], + "angle": 0, + "content": "3" + }, + { + "type": "title", + "bbox": [ + 0.217, + 0.145, + 0.388, + 0.161 + ], + "angle": 0, + "content": "2 Related Work" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.179, + 0.785, + 0.24 + ], + "angle": 0, + "content": "In this section, we review methods for dynamic scene reconstruction that deform 3D canonical space and methods for reconstructing dynamic scenes utilizing dynamic 3D Gaussians. Afterward, we review methods that use embeddings and spatial relationships of Gaussians." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.263, + 0.788, + 0.446 + ], + "angle": 0, + "content": "Deforming 3D Canonical Space D-NeRF [22] reconstructs dynamic scenes by deforming ray samples over time, using the deformation network that takes 3D coordinates and timestamps of the sample as inputs. Nerfies [20] and HyperNeRF [21] use per-frame trainable deformation codes instead of time conditions to deform the canonical space. Instead of deforming from the canonical frame to the entire frames, HyperReel [1] deforms the ray sample of the keyframe to represent the intermediate frame. 4DGaussians [29] and D3DGS [31] reconstruct the dynamic scene with a deformation network which inputs the center position of the canonical 3D Gaussians and timestamps. MoDGS [15] learns the mapping between canonical space and space at specific timestamp through invertible MLP. In contrast, we demonstrate a novel deformation representation as a function of a product space of per-Gaussian latent embeddings and temporal embeddings." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.469, + 0.788, + 0.588 + ], + "angle": 0, + "content": "Dynamic 3D Gaussians To extend the fast rendering speed of 3D Gaussian Splitting [10] into dynamic scene reconstructions. 4D Gaussians [29] decodes features from multi-resolution HexPlanes [2] for temporal deformation of 3D Gaussians. While D3DGS [31] uses an implicit function that processes the time and location of the Gaussian. 4DGS [30] decomposes the 4D Gaussians into a time-conditioned 3D Gaussians and a marginal 1D Gaussians. STG [13] represents changes in 3D Gaussian over time through a temporal opacity and a polynomial function for each Gaussian." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.59, + 0.788, + 0.651 + ], + "angle": 0, + "content": "Our method uses deformable 3D Gaussians as 4D Gaussians [29] and D3DGS [31] do, but does not necessitate the separated feature field to obtain the input feature of the deformation decoder. Our approach uses embeddings allocated to each Gaussian and a temporal embedding shared within a specific frame." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.674, + 0.788, + 0.84 + ], + "angle": 0, + "content": "Latent Embedding on Novel View Synthesis Some studies incorporate latent embeddings to represent different states of the static and dynamic scene. NeRF-W [18] and Block-NeRF [25] employ per-image embeddings to capture different appearances of a scene, representing the scenes from in-the-wild image collections. DyNeRF and MixVoxels [12, 27] employ a temporal embedding for each frame to represent dynamic scenes. Nerfies [20] and HyperNeRF [21] incorporate both per-frame appearance and deformation embeddings. Sync-NeRF [11] introduces time offset to calibrate the misaligned temporal embeddings on dynamic scenes from unsynchronized videos. We introduce per-Gaussian latent embedding to encode the changes over time of each Gaussian and use temporal embeddings to represent different states in each frame of the scene." + } + ], + [ + { + "type": "page_number", + "bbox": [ + 0.218, + 0.116, + 0.23, + 0.127 + ], + "angle": 0, + "content": "4" + }, + { + "type": "header", + "bbox": [ + 0.272, + 0.115, + 0.355, + 0.127 + ], + "angle": 0, + "content": "J. Bae et al." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.147, + 0.788, + 0.299 + ], + "angle": 0, + "content": "Considering Spatial Relationships of Gaussians Scaffold-GS [16] reconstruct 3D scenes by synthesizing Gaussians from anchors, utilizing that the neighboring Gaussians have similar properties. SAGS [26] creates a graph based on k-nearest neighbors (KNN) so that each Gaussian is optimized while considering its neighboring Gaussians. In dynamic scene reconstruction, SC-GS [9] and GaussianPrediction [34] deform Gaussians by combining the deformations of key point Gaussians. Dynamic 3D Gaussians [17] utilizes regularization to encourage that Gaussians and their neighboring Gaussians deform with local rigidity. Similarly, we propose a local smoothness regularization that encourages neighboring Gaussians to have similar embeddings, resulting in similar deformations." + }, + { + "type": "title", + "bbox": [ + 0.216, + 0.322, + 0.331, + 0.339 + ], + "angle": 0, + "content": "3 Method" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.355, + 0.788, + 0.448 + ], + "angle": 0, + "content": "In this section, we first provide a brief overview of 3D Gaussian Splatting (Section 3.1). Next, we introduce our overall framework, embedding-based deformation for Gaussians (Section 3.2) and coarse-fine deformation scheme consisting of coarse and fine deformation functions (Section 3.3). Finally, we present a local smoothness regularization for per-Gaussian embeddings to achieve better details on dynamic regions (Section 3.4)." + }, + { + "type": "title", + "bbox": [ + 0.216, + 0.469, + 0.557, + 0.485 + ], + "angle": 0, + "content": "3.1 Preliminary: 3D Gaussian Splitting" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.495, + 0.787, + 0.571 + ], + "angle": 0, + "content": "3D Gaussian splattering [10] optimizes a set of anisotropic 3D Gaussians through differentiable tile rasterization to reconstruct a static 3D scene. By its efficient rasterization, the optimized model enables real-time rendering of high-quality images. Each 3D Gaussian kernel \\( G_{i}(x) \\) at the point \\( x \\) consists with position \\( \\mathbf{x}_i \\), rotation \\( R_i \\), and scale \\( S_i \\):" + }, + { + "type": "equation", + "bbox": [ + 0.289, + 0.582, + 0.786, + 0.602 + ], + "angle": 0, + "content": "\\[\nG _ {i} (x) = e ^ {- \\frac {1}{2} (x - \\mathbf {x} _ {i}) ^ {T} \\Sigma_ {i} ^ {- 1} (x - \\mathbf {x} _ {i})}, \\quad \\text {w h e r e} \\quad \\Sigma_ {i} = R _ {i} S _ {i} S _ {i} ^ {T} R _ {i} ^ {T}. \\tag {1}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.614, + 0.787, + 0.66 + ], + "angle": 0, + "content": "To projecting 3D Gaussians to 2D for rendering, covariance matrix \\(\\Sigma^{\\prime}\\) are calculated by viewing transform \\(W\\) and the Jacobian \\(J\\) of the affine approximation of the projective transformation [35] as follows:" + }, + { + "type": "equation", + "bbox": [ + 0.435, + 0.671, + 0.785, + 0.688 + ], + "angle": 0, + "content": "\\[\n\\Sigma^ {\\prime} = J W \\Sigma W ^ {T} J ^ {T}. \\tag {2}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.7, + 0.787, + 0.746 + ], + "angle": 0, + "content": "Blending \\(\\mathcal{N}\\) depth-ordered projected points that overlap the pixel, the Gaussian kernel \\(G_{i}(x)\\) is multiplied by the opacity of the Gaussian \\(\\sigma_{i}\\) and calculates the pixel color \\(C\\) with the color of the Gaussian \\(c_{i}\\):" + }, + { + "type": "equation", + "bbox": [ + 0.317, + 0.758, + 0.786, + 0.8 + ], + "angle": 0, + "content": "\\[\nC = \\sum_ {i \\in \\mathcal {N}} c _ {i} \\alpha_ {i} \\prod_ {j = 1} ^ {i - 1} (1 - \\alpha_ {j}), \\quad \\text {w h e r e} \\quad \\alpha_ {i} = \\sigma_ {i} G _ {i} (x). \\tag {3}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.81, + 0.787, + 0.84 + ], + "angle": 0, + "content": "The color of Gaussian \\( c_{i} \\) is determined using the SH coefficient with view direction." + } + ], + [ + { + "type": "header", + "bbox": [ + 0.663, + 0.115, + 0.732, + 0.127 + ], + "angle": 0, + "content": "E-D3DGS" + }, + { + "type": "page_number", + "bbox": [ + 0.775, + 0.117, + 0.785, + 0.127 + ], + "angle": 0, + "content": "5" + }, + { + "type": "image", + "bbox": [ + 0.235, + 0.147, + 0.38, + 0.211 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.239, + 0.217, + 0.386, + 0.291 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.237, + 0.295, + 0.39, + 0.306 + ], + "angle": 0, + "content": "(a) Embeddings for deformation" + }, + { + "type": "image", + "bbox": [ + 0.406, + 0.146, + 0.599, + 0.29 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.433, + 0.295, + 0.572, + 0.305 + ], + "angle": 0, + "content": "(b) Per-Gaussian deformation" + }, + { + "type": "image", + "bbox": [ + 0.624, + 0.149, + 0.756, + 0.29 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.607, + 0.295, + 0.776, + 0.306 + ], + "angle": 0, + "content": "(c) Local smoothness regularization" + }, + { + "type": "image_caption", + "bbox": [ + 0.215, + 0.319, + 0.789, + 0.443 + ], + "angle": 0, + "content": "Fig. 2: Framework. Existing coordinate-based network methods struggle to represent complex dynamic scenes. To this end, we define per-Gaussian deformation. (a) Firstly, we assign a latent embedding for each Gaussian. Additionally, we introduce coarse and fine temporal embeddings to represent the slow and fast state of the dynamic scene. (b) By employing two decoders that take per-Gaussian latent embeddings along with coarse and fine temporal embeddings as input, we estimate slow or large changes and fast or detailed changes to model the final deformation, respectively. (c) Finally, we introduce a local smoothness regularization so that the embeddings of neighboring Gaussians are similar." + }, + { + "type": "title", + "bbox": [ + 0.216, + 0.47, + 0.636, + 0.485 + ], + "angle": 0, + "content": "3.2 Embedding-Based Deformation for Gaussians" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.492, + 0.788, + 0.581 + ], + "angle": 0, + "content": "Deformable NeRFs consist of a deformation field that predicts displacement \\(\\Delta \\mathbf{x}\\) for a given coordinate \\(\\mathbf{x}\\) from the canonical space to each target frame, and a radiance field that maps color and density from a given coordinate in the canonical space \\((\\mathbf{x} + \\Delta \\mathbf{x})\\). Existing deformable Gaussian methods employ the same approach for predicting the deformation of Gaussians, i.e., utilizing a deformation field based on coordinates." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.583, + 0.788, + 0.674 + ], + "angle": 0, + "content": "Unlike previous methods, we start from the design of 3DGS: the 3D scene is represented as a mixture of Gaussians that have individual radiance fields. Accordingly, the deformation should be defined for each Gaussian. Based on this intuition, we introduce a function \\(\\mathcal{F}_{\\theta}\\) that produces deformation from learnable embeddings \\(\\mathbf{z}_{\\mathrm{g}} \\in \\mathbb{R}^{32}\\) belonging to individual Gaussians (Figure 2a), and typical temporal embeddings \\(\\mathbf{z}_t \\in \\mathbb{R}^{256}\\) for different frames:" + }, + { + "type": "equation", + "bbox": [ + 0.364, + 0.681, + 0.787, + 0.699 + ], + "angle": 0, + "content": "\\[\n\\mathcal {F} _ {\\theta}: \\left(\\mathbf {z} _ {\\mathrm {g}}, \\mathbf {z} _ {t}\\right)\\rightarrow \\left(\\Delta \\mathbf {x}, \\Delta \\mathbf {r}, \\Delta \\mathbf {s}, \\Delta \\sigma , \\Delta Y\\right), \\tag {4}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.705, + 0.788, + 0.78 + ], + "angle": 0, + "content": "where \\(\\mathbf{r}\\) is a rotation quaternion, \\(\\mathbf{s}\\) is a vector for scaling, \\(\\sigma\\) is an opacity, and \\(Y\\) is SH coefficients for modeling view-dependent color. We implement \\(\\mathcal{F}_{\\theta}\\) as a shallow multi-layer perceptron (MLP) followed by an MLP head for each parameter. As a result, the Gaussian parameters at frame \\(t\\) are determined by adding \\(\\mathcal{F}_{\\theta}(\\mathbf{z}_{\\mathrm{g}},\\mathbf{z}_t)\\) to the canonical Gaussian parameters (Figure 2c)." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.78, + 0.788, + 0.84 + ], + "angle": 0, + "content": "We jointly optimize the per-Gaussian embeddings \\(\\mathbf{z}_{\\mathrm{g}}\\), the deformation function \\(\\mathcal{F}_{\\theta}\\), and the canonical Gaussian parameters to minimize the rendering loss. We use the L1 and periodic DSSIM as the rendering loss between the rendered image and the ground truth image." + } + ], + [ + { + "type": "page_number", + "bbox": [ + 0.218, + 0.116, + 0.23, + 0.127 + ], + "angle": 0, + "content": "6" + }, + { + "type": "header", + "bbox": [ + 0.272, + 0.115, + 0.355, + 0.127 + ], + "angle": 0, + "content": "J. Bae et al." + }, + { + "type": "title", + "bbox": [ + 0.216, + 0.147, + 0.471, + 0.16 + ], + "angle": 0, + "content": "3.3 Coarse-Fine Deformation" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.173, + 0.784, + 0.232 + ], + "angle": 0, + "content": "Different parts of a scene may have coarse and fine motions [5]. E.g., a hand swiftly stirs a pan (fine) while a body slowly moves from left to right (coarse). Based on this intuition, we introduce a coarse-fine deformation that produces a summation of coarse and fine deformations." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.234, + 0.784, + 0.278 + ], + "angle": 0, + "content": "Coarse-fine deformation consists of two functions with the same architecture: one for coarse and one for fine deformation (Figure 2c). The functions receive different temporal embeddings as follows:" + }, + { + "type": "text", + "bbox": [ + 0.215, + 0.279, + 0.785, + 0.37 + ], + "angle": 0, + "content": "Following typical temporal embeddings, we start from a 1D feature grid \\( Z \\in \\mathbb{R}^{N \\times 256} \\) for \\( N \\) frames and use an embedding \\( \\mathbf{z}_t^{\\mathrm{f}} = \\mathrm{interp}(Z,t) \\) for fine deformation. For coarse deformation, we linearly downsample \\( Z \\) by a factor of 5 to remove high-frequencies responsible for fast and detailed deformation. Then we compute \\( \\mathbf{z}_t^{\\mathrm{c}} \\) as a linear interpolation of embeddings at enclosing grid points (Figure 2b)." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.37, + 0.785, + 0.445 + ], + "angle": 0, + "content": "As a result, coarse deformation \\(\\mathcal{F}_{\\theta_{\\mathrm{c}}}(\\mathbf{z}_{\\mathrm{g}},\\mathbf{z}_{t}^{c})\\) is responsible for representing large or slow movements in the scene, while fine deformation \\(\\mathcal{F}_{\\theta_{\\mathrm{f}}}(\\mathbf{z}_{\\mathrm{g}},\\mathbf{z}_{t}^{\\mathrm{f}})\\) learns the fast or detailed movements that coarse deformation does not cover. This improves the deformation quality. Refer to the Ablation study section for more details." + }, + { + "type": "title", + "bbox": [ + 0.216, + 0.469, + 0.535, + 0.484 + ], + "angle": 0, + "content": "3.4 Local Smoothness Regularization" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.495, + 0.784, + 0.556 + ], + "angle": 0, + "content": "Neighboring Gaussians constructing dynamic objects tend to exhibit locally similar deformation. Inspired by [17], we introduce a local smoothness regularization for per-Gaussian embedding \\(\\mathbf{z}_{\\mathrm{g}}\\) (Figure 2d) to encourage similar deformations between nearby Gaussians \\(i\\) and \\(j\\):" + }, + { + "type": "equation", + "bbox": [ + 0.333, + 0.566, + 0.671, + 0.603 + ], + "angle": 0, + "content": "\\[\n\\mathcal {L} ^ {\\mathrm {e m b} - \\mathrm {r e g}} = \\frac {1}{k | \\mathcal {S} |} \\sum_ {i \\in \\mathcal {S}} \\sum_ {j \\in \\mathrm {K N N} _ {i; k}} \\left(w _ {i, j} \\| \\mathbf {z} _ {\\mathrm {g} i} - \\mathbf {z} _ {\\mathrm {g} j} \\| _ {2}\\right),\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.616, + 0.785, + 0.661 + ], + "angle": 0, + "content": "where \\( w_{i,j} = \\exp (-\\lambda_w\\| \\mu_j - \\mu_i\\| _2^2) \\) is the weighting factor and \\( \\mu \\) is the Gaussian center. We set \\( \\lambda_w \\) to 2000 and \\( k \\) to 20 following [17]. To reduce the computational cost, we obtain sets of k-nearest-neighbors only when the densification occurs." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.663, + 0.784, + 0.723 + ], + "angle": 0, + "content": "Note that unlike previous approaches that directly constrain physical properties such as rigidity or rotation, We implicitly induce locally similar deformation by ensuring that per-Gaussian embeddings are locally smooth. Our regularization allows better capture of textures and details of dynamic objects." + }, + { + "type": "title", + "bbox": [ + 0.216, + 0.747, + 0.367, + 0.764 + ], + "angle": 0, + "content": "4 Experiment" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.78, + 0.785, + 0.84 + ], + "angle": 0, + "content": "In this section, we first describe the criterion for selection of baselines, and evaluation metrics. We then demonstrate the effectiveness of our method through comparisons with various baselines and datasets (Section 4.1-4.2). Finally, we conduct analysis and ablations of our method (Section 4.3)." + } + ], + [ + { + "type": "header", + "bbox": [ + 0.663, + 0.115, + 0.732, + 0.127 + ], + "angle": 0, + "content": "E-D3DGS" + }, + { + "type": "page_number", + "bbox": [ + 0.775, + 0.117, + 0.785, + 0.127 + ], + "angle": 0, + "content": "7" + }, + { + "type": "table_caption", + "bbox": [ + 0.217, + 0.145, + 0.788, + 0.214 + ], + "angle": 0, + "content": "Table 1: Average performance in the test view on Neural 3D Video Dataset The computational cost was measured based on flame_salmon_1 on the A6000. \\( {}^{1} \\) flame salmon scene includes only the first segment, comprising 300 frames. \\( {}^{2} \\) reported time and DyNeRF is trained on 8 GPUs and tested only on flame salmon. \\( {}^{3} \\) trained with 90 frames. \\( {}^{4} \\) trained with 50 frames." + }, + { + "type": "table", + "bbox": [ + 0.276, + 0.22, + 0.723, + 0.342 + ], + "angle": 0, + "content": "
modelmetriccomputational cost
PSNR↑SSIM↑LPIPS↓Training time↓FPS↑Model size↓
DyNeRF12 [12]29.58-0.0831344 hours0.0156 MB
NeRFPlayer13 [24]30.69-0.1116 hours0.051654 MB
Mix Voxels [27]30.300.9180.1271 hours 40 mins0.93512 MB
K-Planes [6]30.860.9390.0961 hours 40 mins0.13309 MB
HyperReel4 [1]30.370.9210.1069 hours 20 mins1.01362 MB
4DGS [30]31.190.9400.0519 hours 30 mins33.78700 MB
4DGaussians [29]30.710.9350.05650 mins51.959 MB
Ours31.310.9450.0371 hours 52 mins74.535 MB
" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.369, + 0.788, + 0.535 + ], + "angle": 0, + "content": "Baselines We choose the start-of-the-art method as a baseline in each dataset. We compared against DyNeRF, NeRFPlayer, MixVoxels, K-Planes, HyperReel, Nerfies, HyperNeRF, and TiNeuVox on the NeRF baseline. In detail, we use the version of NeRFPlayer TensoRF VM, HyperNeRF DF, Mixvoxels-L, K-Planes hybrid. We compared with 4DGaussians, 4DGS, and D3DGS based on the Gaussian baseline. Meanwhile, we have not included STG in our comparison due to its requirement for per-frame Structure from Motion (SfM) points, which makes conducting a fair comparison challenging. Also, STG is not a deformable 3D Gaussian approach. We followed the official code and configuration, except increasing the training iterations for the Technicolor dataset to 1.5 times that of the 4DGaussians, in comparison to the Neural 3D Video dataset." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.555, + 0.788, + 0.661 + ], + "angle": 0, + "content": "Metrics We report the quality of rendered images using PSNR, SSIM, and LPIPS. Peak Signal-to-Noise Ratio (PSNR) quantifies pixel color error between the rendered video and the ground truth. We utilize SSIM [28] to account for the perceived similarity of the rendered image. Additionally, we measure higher-level perceptual similarity using Learned Perceptual Image Patch Similarity (LPIPS) [33] with an AlexNet Backbone. Higher PSNR and SSIM values and lower LPIPS values indicate better visual quality." + }, + { + "type": "title", + "bbox": [ + 0.216, + 0.681, + 0.536, + 0.697 + ], + "angle": 0, + "content": "4.1 Effectiveness on Dynamic Region" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.704, + 0.788, + 0.84 + ], + "angle": 0, + "content": "Neural 3D Video Dataset [12] includes 20 multi-view videos, with each scene consisting of either 300 frames, except for the flame_salmon scene, which comprises 1200 frames. These scenes encompass a relatively long duration and various movements, with some featuring multiple objects in motion. We utilized Neural 3D Video dataset to observe the capability to capture dynamic areas. Total six scenes (coffee_martini, cook_s spinach, cut_ roasted_beef, flame_salmon, flame_ steak, sear_ steak) are evaluated in Figure 3 and Table 1. The flame_salmon scene divided into four segments, each containing 300 frames." + } + ], + [ + { + "type": "page_number", + "bbox": [ + 0.218, + 0.117, + 0.229, + 0.127 + ], + "angle": 0, + "content": "8" + }, + { + "type": "header", + "bbox": [ + 0.272, + 0.116, + 0.356, + 0.127 + ], + "angle": 0, + "content": "J. Bae et al." + }, + { + "type": "image", + "bbox": [ + 0.226, + 0.149, + 0.782, + 0.495 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.252, + 0.509, + 0.75, + 0.525 + ], + "angle": 0, + "content": "Fig. 3: Qualitative comparisons on the Neural 3D Video Dataset." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.555, + 0.788, + 0.706 + ], + "angle": 0, + "content": "Table 1 presents quantitative metrics on the average metrics across the test views of all scenes, computational and storage costs on the first fragment of flame_salmon scene. Refer to the Appendix for per-scene details. Our method demonstrates superior reconstruction quality, FPS, and model size across compared to baselines. As the table shows, NeRF baselines generally required longer training and rendering times. While 4DGS shows relatively high reconstruction performance, it demands longer training times and larger VRAM storage capacity compared to other baselines. 4DGaussians requires lower computational and storage cost but it displays low reconstruction quality in some scenes with rapid dynamics, as shown in the teaser and Figure 3." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.707, + 0.788, + 0.813 + ], + "angle": 0, + "content": "Figure 3 reports the rendering quality. Our method successfully reconstructs the fine details in moving areas, outperforming baselines on average metrics across test views. Baselines show blurred dynamic areas or severe artifacts in low-light scenes such as cook_spinach and flame_steak. 4DGS exhibits the disappearance of some static areas. In 4DGaussians, a consistent over-smoothing occurs in dynamic areas. All baselines experienced reduced quality in reflective or thin dynamic areas like clamps, torches, and windows." + } + ], + [ + { + "type": "header", + "bbox": [ + 0.663, + 0.115, + 0.732, + 0.127 + ], + "angle": 0, + "content": "E-D3DGS" + }, + { + "type": "page_number", + "bbox": [ + 0.776, + 0.117, + 0.786, + 0.127 + ], + "angle": 0, + "content": "9" + }, + { + "type": "table_caption", + "bbox": [ + 0.233, + 0.145, + 0.769, + 0.159 + ], + "angle": 0, + "content": "Table 2: Average performance in the test view on Technicolor Dataset" + }, + { + "type": "table", + "bbox": [ + 0.304, + 0.172, + 0.695, + 0.242 + ], + "angle": 0, + "content": "
modelmetriccomputational cost
PSNR↑SSIM↑LPIPS↓Training time↓FPS↑Model size↓
DyNeRF31.80-0.140-0.020.6 MB
HyperReel32.320.8990.1182 hours 45 mins0.91289 MB
4DGaussians29.620.8440.17625 mins34.851 MB
Ours33.240.9070.1002 hours 55 mins60.877 MB
" + }, + { + "type": "image", + "bbox": [ + 0.264, + 0.269, + 0.74, + 0.517 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.274, + 0.531, + 0.728, + 0.546 + ], + "angle": 0, + "content": "Fig. 4: Qualitative comparisons on the Technicolor Dataset." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.587, + 0.785, + 0.662 + ], + "angle": 0, + "content": "Technicolor Light Field Dataset [23] is a multi-view dataset captured with a time-synchronized \\(4 \\times 4\\) camera rig, containing intricate details. We train ours and the baselines on 50 frames of five commonly used scenes (Birthday, Fabien, Painter, Theater, Trains) using full-resolution videos at \\(2048 \\times 1088\\) pixels, with the second row and second column cameras used as test views." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.669, + 0.788, + 0.774 + ], + "angle": 0, + "content": "Table 2 reports the average metrics across the test views of all scenes, computational and storage costs on Painter scene. HyperReel demonstrates overall high-quality results but struggles with relatively slow training times and FPS, and a larger model size. 4DGaussians exhibits fast training times and FPS but significantly underperforms in reconstructing fine details compared to other baselines. However, our method demonstrates superior reconstruction quality and faster FPS compared to the baselines." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.78, + 0.788, + 0.841 + ], + "angle": 0, + "content": "As shown in Figure 4, HyperReel produces noisy artifacts due to incorrect predictions of the displacement vector. 4D Gaussians fails to capture fine details in dynamic areas, exhibiting over-smoothing results. All baselines struggle to accurately reconstruct rapidly moving thin areas like fingers." + } + ], + [ + { + "type": "page_number", + "bbox": [ + 0.218, + 0.116, + 0.236, + 0.127 + ], + "angle": 0, + "content": "10" + }, + { + "type": "header", + "bbox": [ + 0.272, + 0.115, + 0.355, + 0.127 + ], + "angle": 0, + "content": "J. Bae et al." + }, + { + "type": "title", + "bbox": [ + 0.216, + 0.147, + 0.599, + 0.163 + ], + "angle": 0, + "content": "4.2 Scenes with Challenging Camera Setting" + }, + { + "type": "table_caption", + "bbox": [ + 0.237, + 0.194, + 0.766, + 0.209 + ], + "angle": 0, + "content": "Table 3: Average performance in the test view on Hypernerf Dataset" + }, + { + "type": "table", + "bbox": [ + 0.304, + 0.221, + 0.699, + 0.307 + ], + "angle": 0, + "content": "
modelmetriccomputational cost
PSNR↑SSIM↑LPIPS↓Training time↓FPS↑Model size↓
Nerfies [20]22.23-0.170~ hours< 1-
HyperNeRF DS [21]22.290.5980.15332 hours< 115 MB
TiNeuVox [4]24.200.6160.3933 hours 30 mins148 MB
D3DGS [31]22.400.5980.27516 mins6.95309 MB
4DGaussians25.030.6820.28150 mins96.360 MB
Ours25.430.6970.2311 hours 15 mins139.333 MB
" + }, + { + "type": "image", + "bbox": [ + 0.309, + 0.359, + 0.7, + 0.614 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.27, + 0.627, + 0.733, + 0.643 + ], + "angle": 0, + "content": "Fig. 5: Qualitative comparisons on the HyperNeRF Dataset." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.689, + 0.799, + 0.75 + ], + "angle": 0, + "content": "HyperNeRF dataset includes videos captured using two phones rigidly mounted on a handheld rig. We train on all frames of four scenes (3D Printer, Banana, Broom, Chicken) at a resolution downsampled by half to \\(536 \\times 960\\). Due to memory constraints, D3DGS is trained on images downsampled by a quarter." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.75, + 0.789, + 0.84 + ], + "angle": 0, + "content": "The table shows that our method outperforms the reconstruction performance with previous methods along with compact model size and faster FPS. Figure 5 shows that previous methods struggle to reconstruct fast-moving parts such as fingers and broom. Especially D3DGS deteriorates in Broom scene. Table 3 reports the average metrics across the test views of all scenes, computational and storage costs on Broom scene." + } + ], + [ + { + "type": "header", + "bbox": [ + 0.663, + 0.115, + 0.732, + 0.127 + ], + "angle": 0, + "content": "E-D3DGS" + }, + { + "type": "page_number", + "bbox": [ + 0.769, + 0.117, + 0.785, + 0.127 + ], + "angle": 0, + "content": "11" + }, + { + "type": "title", + "bbox": [ + 0.216, + 0.147, + 0.504, + 0.162 + ], + "angle": 0, + "content": "4.3 Analyses and Ablation Study" + }, + { + "type": "image", + "bbox": [ + 0.229, + 0.203, + 0.363, + 0.294 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.233, + 0.296, + 0.361, + 0.308 + ], + "angle": 0, + "content": "(a) canonical rendering" + }, + { + "type": "image", + "bbox": [ + 0.367, + 0.203, + 0.501, + 0.294 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.374, + 0.295, + 0.499, + 0.308 + ], + "angle": 0, + "content": "(b) coarse deform only" + }, + { + "type": "image", + "bbox": [ + 0.505, + 0.203, + 0.638, + 0.294 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.515, + 0.295, + 0.627, + 0.308 + ], + "angle": 0, + "content": "(c) fine deform only" + }, + { + "type": "image", + "bbox": [ + 0.642, + 0.203, + 0.776, + 0.294 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.661, + 0.295, + 0.756, + 0.308 + ], + "angle": 0, + "content": "(d) full rendering" + }, + { + "type": "image_caption", + "bbox": [ + 0.214, + 0.322, + 0.788, + 0.405 + ], + "angle": 0, + "content": "Fig. 6: Deformation components. (a) The canonical space contains Gaussians to represent all target frames of the scene. (b) Applying coarse deformation to the canonical space roughly reflects the dynamics of the scene. (c) The rendering without coarse deformation and only with fine deformation looks similar to the canonical rendering, i.e., responsible for fine deformations. (d) Applying both coarse and fine deformation yields natural rendering results." + }, + { + "type": "image", + "bbox": [ + 0.256, + 0.466, + 0.499, + 0.568 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.343, + 0.569, + 0.411, + 0.582 + ], + "angle": 0, + "content": "Frame 67" + }, + { + "type": "image", + "bbox": [ + 0.507, + 0.466, + 0.751, + 0.568 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.59, + 0.569, + 0.668, + 0.582 + ], + "angle": 0, + "content": "Frame 185" + }, + { + "type": "image_caption", + "bbox": [ + 0.214, + 0.596, + 0.788, + 0.651 + ], + "angle": 0, + "content": "Fig.7: Visualization of the magnitude of deformation. Coarse deformation (blue) captures large and slow changes, such as the movement of the head and torso, while fine deformation (red) is responsible for the fast and detailed movements of arms, tongs, shadows, etc." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.705, + 0.789, + 0.841 + ], + "angle": 0, + "content": "Deformation components In Figure 6, we present an analysis of the coarse-fine deformation. To achieve this, we render a flame_steak scene by omitting each of our deformation components one by one. Our full rendering results from adding coarse and fine deformation to the canonical space (Figure 6d). When both are removed, rendering yields an image in canonical space (Figure 6a). Rendering with the coarse deformation, which handles large or slow changes in the scene, produces results similar to the full rendering (Figure 6b). On the other hand, fine deformation is responsible for fast or detailed changes in the scene, yielding rendering similar to canonical space (Figure 6c)." + } + ], + [ + { + "type": "page_number", + "bbox": [ + 0.218, + 0.116, + 0.236, + 0.127 + ], + "angle": 0, + "content": "12" + }, + { + "type": "header", + "bbox": [ + 0.272, + 0.115, + 0.355, + 0.127 + ], + "angle": 0, + "content": "J. Bae et al." + }, + { + "type": "image", + "bbox": [ + 0.232, + 0.147, + 0.363, + 0.207 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.259, + 0.208, + 0.337, + 0.22 + ], + "angle": 0, + "content": "(a) Ours Full" + }, + { + "type": "image", + "bbox": [ + 0.368, + 0.147, + 0.501, + 0.207 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.373, + 0.208, + 0.498, + 0.22 + ], + "angle": 0, + "content": "(b) coarse deform only" + }, + { + "type": "image", + "bbox": [ + 0.505, + 0.147, + 0.636, + 0.207 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.515, + 0.208, + 0.626, + 0.22 + ], + "angle": 0, + "content": "(c) fine deform only" + }, + { + "type": "image", + "bbox": [ + 0.642, + 0.147, + 0.774, + 0.207 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.647, + 0.208, + 0.768, + 0.22 + ], + "angle": 0, + "content": "(d) Ours \\(+\\mathbf{x}\\) injection" + }, + { + "type": "image_caption", + "bbox": [ + 0.214, + 0.232, + 0.788, + 0.304 + ], + "angle": 0, + "content": "Fig. 8: Qualitative ablation results on coarse-fine deformation. (a) Our model achieves clear results with both coarse and fine decoders. (b-c) The quality of dynamic areas decreases if one is missing. (d) Additionally, introducing the coordinates of Gaussian as an additional input into our decoders results in a decrease in the quality of both static and dynamic regions." + }, + { + "type": "table_caption", + "bbox": [ + 0.25, + 0.318, + 0.751, + 0.334 + ], + "angle": 0, + "content": "Table 4: Quantitative ablation results on coarse-fine deformation." + }, + { + "type": "table", + "bbox": [ + 0.305, + 0.345, + 0.695, + 0.42 + ], + "angle": 0, + "content": "
MethodPSNR↑SSIM↑LPIPS↓
Ours29.700.9330.041
coarse deformation only29.480.9310.044
fine deformation only29.230.9320.043
Ours + x injection29.600.9310.045
" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.451, + 0.788, + 0.603 + ], + "angle": 0, + "content": "To examine the roles of the coarse and fine deformation in the coarse-fine deformation, we conduct a visualization on flame_steak scene. First, we compute the Euclidean norm of positional shifts between the current and subsequent frames. We then add the value to the DC components of the SH coefficients proportionally to the magnitude: blue for coarse deformation and red for fine deformation. For visual clarity, we render the original scene in grayscale. As illustrated in Figure 7, coarse deformation models slower changes such as body movement, while fine deformation models faster movements like cooking arms. Thus, we demonstrate that by downsampling the temporal embedding grid \\( Z \\), we can effectively separate and model slow and fast deformations in the scene." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.628, + 0.788, + 0.779 + ], + "angle": 0, + "content": "Ablation study We report the results of an ablation study on the deformation decoder in Figure 8 and Table 4. First, our full method (using both coarse and fine decoders) produces clear rendering results and models dynamic variations well (Figure 8a). Training only with the coarse or fine decoder leads to blurred dynamic areas and a failure to accurately capture detailed motion (Figure 8b-c). Additionally, we demonstrate experiments where the Gaussian center coordinates \\(\\mathbf{x}\\) are injected into the input of each decoder. As shown in Figure 8d, including the Gaussian coordinates degrades the quality of deformation, supporting our argument that coordinate dependency should be removed from the deformation function." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.78, + 0.788, + 0.842 + ], + "angle": 0, + "content": "Furthermore, we report the results of an ablation study on the local smoothness regularization for per-Gaussian embeddings. As shown in Figure 9, our regularization improves the details and texture quality of moving objects. In Table 5, we show a performance comparison between the proposed regularization and" + } + ], + [ + { + "type": "header", + "bbox": [ + 0.663, + 0.115, + 0.732, + 0.127 + ], + "angle": 0, + "content": "E-D3DGS" + }, + { + "type": "page_number", + "bbox": [ + 0.769, + 0.116, + 0.786, + 0.127 + ], + "angle": 0, + "content": "13" + }, + { + "type": "table_caption", + "bbox": [ + 0.215, + 0.145, + 0.788, + 0.201 + ], + "angle": 0, + "content": "Table 5: Quantitative ablation results on local smoothness regularization. We compare the performance of applying our regularization with the physically-based regularization of Dynamic 3D Gaussians [17]. Our regularization better captures details of dynamic objects." + }, + { + "type": "table", + "bbox": [ + 0.29, + 0.213, + 0.71, + 0.274 + ], + "angle": 0, + "content": "
MethodPSNR↑SSIM↑LPIPS↓
Ours w/o embedding reg32.260.9510.037
+ our embedding reg32.340.9520.036
+ physically-based reg32.080.9500.036
" + }, + { + "type": "image", + "bbox": [ + 0.28, + 0.291, + 0.386, + 0.357 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.306, + 0.359, + 0.361, + 0.37 + ], + "angle": 0, + "content": "Ours full" + }, + { + "type": "image", + "bbox": [ + 0.39, + 0.292, + 0.495, + 0.357 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.405, + 0.359, + 0.48, + 0.37 + ], + "angle": 0, + "content": "Ours w/o reg" + }, + { + "type": "image", + "bbox": [ + 0.509, + 0.291, + 0.615, + 0.357 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.538, + 0.359, + 0.594, + 0.37 + ], + "angle": 0, + "content": "Ours full" + }, + { + "type": "image", + "bbox": [ + 0.617, + 0.292, + 0.724, + 0.357 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.634, + 0.359, + 0.708, + 0.37 + ], + "angle": 0, + "content": "Ours w/o reg" + }, + { + "type": "image_caption", + "bbox": [ + 0.233, + 0.386, + 0.768, + 0.401 + ], + "angle": 0, + "content": "Fig. 9: Qualitative ablation results on local smoothness regularization." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.43, + 0.788, + 0.537 + ], + "angle": 0, + "content": "the existing physically-based regularizations. To apply the method proposed by Luiten et al. [17] to ours, we make some modifications: 1) Like our method, we find the set of k-nearest-neighbours only when the densification occurs to reduce the computational cost. 2) For long-term local-isometry loss, we use the time of video frame used in previous training step instead of using the time of the first frame. Our regularization is simple and shows better performance compared to previous method." + }, + { + "type": "title", + "bbox": [ + 0.215, + 0.559, + 0.51, + 0.575 + ], + "angle": 0, + "content": "5 Conclusion and Limitation" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.592, + 0.788, + 0.698 + ], + "angle": 0, + "content": "We propose a per-Gaussian deformation for 3DGS that takes per-Gaussian embeddings as input, instead of using the typical deformation fields from previous deformable 3DGS works, resulting in high performance. We enhance the reconstruction quality by decomposing the dynamic changes into coarse and fine deformation. However, our method learns inappropriate Gaussian deformation with casually captured monocular videos [7], like other baselines. We plan to address it in future work by introducing useful prior for monocular video settings." + }, + { + "type": "image", + "bbox": [ + 0.254, + 0.739, + 0.374, + 0.799 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.279, + 0.799, + 0.351, + 0.81 + ], + "angle": 0, + "content": "4DGaussians" + }, + { + "type": "image", + "bbox": [ + 0.376, + 0.739, + 0.497, + 0.799 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.421, + 0.8, + 0.452, + 0.81 + ], + "angle": 0, + "content": "Ours" + }, + { + "type": "image", + "bbox": [ + 0.507, + 0.739, + 0.628, + 0.799 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.532, + 0.799, + 0.604, + 0.81 + ], + "angle": 0, + "content": "4DGaussians" + }, + { + "type": "image", + "bbox": [ + 0.63, + 0.739, + 0.75, + 0.799 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.674, + 0.8, + 0.705, + 0.81 + ], + "angle": 0, + "content": "Ours" + }, + { + "type": "image_caption", + "bbox": [ + 0.223, + 0.822, + 0.777, + 0.837 + ], + "angle": 0, + "content": "Fig. 10: Limitation. Ours struggles with the casually captured monocular videos." + } + ], + [ + { + "type": "page_number", + "bbox": [ + 0.218, + 0.116, + 0.236, + 0.127 + ], + "angle": 0, + "content": "14" + }, + { + "type": "header", + "bbox": [ + 0.272, + 0.115, + 0.355, + 0.127 + ], + "angle": 0, + "content": "J. Bae et al." + }, + { + "type": "title", + "bbox": [ + 0.218, + 0.145, + 0.403, + 0.163 + ], + "angle": 0, + "content": "Acknowledgements" + }, + { + "type": "text", + "bbox": [ + 0.217, + 0.176, + 0.787, + 0.236 + ], + "angle": 0, + "content": "This work is supported by the Institute for Information & Communications Technology Planning & Evaluation (IITP) grant funded by the Korea government(MSIT) (No. 2017-0-00072, Development of Audio/Video Coding and Light Field Media Fundamental Technologies for Ultra Realistic Tera-media)" + }, + { + "type": "title", + "bbox": [ + 0.218, + 0.259, + 0.323, + 0.274 + ], + "angle": 0, + "content": "References" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.288, + 0.787, + 0.345 + ], + "angle": 0, + "content": "1. Attal, B., Huang, J.B., Richardt, C., Zollhoefer, M., Kopf, J., OaAZToole, M., Kim, C.: Hyperreel: High-fidelity 6-dof video with ray-conditioned sampling. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 16610-16620 (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.346, + 0.787, + 0.386 + ], + "angle": 0, + "content": "2. Cao, A., Johnson, J.: Hexplane: A fast representation for dynamic scenes. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 130-141 (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.387, + 0.787, + 0.428 + ], + "angle": 0, + "content": "3. Duisterhof, B.P., Mandi, Z., Yao, Y., Liu, J.W., Shou, M.Z., Song, S., Ichnowski, J.: Md-splatting: Learning metric deformation from 4d gaussians in highly deformable scenes. arXiv preprint arXiv:2312.00583 (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.429, + 0.787, + 0.469 + ], + "angle": 0, + "content": "4. Fang, J., Yi, T., Wang, X., Xie, L., Zhang, X., Liu, W., Nießner, M., Tian, Q.: Fast dynamic radiance fields with time-aware neural voxels. In: SIGGRAPH Asia 2022 Conference Papers (2022)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.47, + 0.787, + 0.51 + ], + "angle": 0, + "content": "5. Feichtenhofer, C., Fan, H., Malik, J., He, K.: Slowfast networks for video recognition. In: Proceedings of the IEEE/CVF international conference on computer vision. pp. 6202-6211 (2019)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.511, + 0.787, + 0.564 + ], + "angle": 0, + "content": "6. Fridovich-Keil, S., Meanti, G., Warburg, F.R., Recht, B., Kanazawa, A.: K-planes: Explicit radiance fields in space, time, and appearance. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 12479-12488 (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.565, + 0.787, + 0.593 + ], + "angle": 0, + "content": "7. Gao, H., Li, R., Tulsiani, S., Russell, B., Kanazawa, A.: Monocular dynamic view synthesis: A reality check. In: NeurIPS (2022)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.594, + 0.787, + 0.634 + ], + "angle": 0, + "content": "8. Huang, Y.H., Sun, Y.T., Yang, Z., Lyu, X., Cao, Y.P., Qi, X.: Sc-gs: Sparse-controlled gaussian splatting for editable dynamic scenes. arXiv preprint arXiv:2312.14937 (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.635, + 0.787, + 0.689 + ], + "angle": 0, + "content": "9. Huang, Y.H., Sun, Y.T., Yang, Z., Lyu, X., Cao, Y.P., Qi, X.: Sc-gs: Sparse-controlled gaussian splatting for editable dynamic scenes. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR). pp. 4220-4230 (June 2024)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.69, + 0.787, + 0.731 + ], + "angle": 0, + "content": "0. Kerbl, B., Kopanas, G., Leimkuhler, T., Drettakis, G.: 3d gaussian splatting for real-time radiance field rendering. ACM Transactions on Graphics 42(4) (July 2023), https://repo-sam.inria.fr/fungraph/3d-gaussian-splatting/" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.731, + 0.787, + 0.758 + ], + "angle": 0, + "content": "1. Kim, S., Bae, J., Yun, Y., Lee, H., Bang, G., Uh, Y.: Sync-nerf: Generalizing dynamic nerfs to unsynchronized videos. arXiv preprint arXiv:2310.13356 (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.22, + 0.759, + 0.787, + 0.813 + ], + "angle": 0, + "content": "12. Li, T., Slavcheva, M., Zollhoefer, M., Green, S., Lassner, C., Kim, C., Schmidt, T., Lovegrove, S., Goesele, M., Newcombe, R., et al.: Neural 3d video synthesis from multi-view video. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (2022)" + }, + { + "type": "ref_text", + "bbox": [ + 0.22, + 0.814, + 0.787, + 0.84 + ], + "angle": 0, + "content": "13. Li, Z., Chen, Z., Li, Z., Xu, Y.: Spacetime gaussian feature splatting for real-time dynamic view synthesis. arXiv preprint arXiv:2312.16812 (2023)" + }, + { + "type": "list", + "bbox": [ + 0.22, + 0.288, + 0.787, + 0.84 + ], + "angle": 0, + "content": null + } + ], + [ + { + "type": "header", + "bbox": [ + 0.663, + 0.115, + 0.732, + 0.127 + ], + "angle": 0, + "content": "E-D3DGS" + }, + { + "type": "page_number", + "bbox": [ + 0.769, + 0.117, + 0.786, + 0.127 + ], + "angle": 0, + "content": "15" + }, + { + "type": "ref_text", + "bbox": [ + 0.214, + 0.147, + 0.788, + 0.189 + ], + "angle": 0, + "content": "14. Liang, Y., Khan, N., Li, Z., Nguyen-Phuoc, T., Lanman, D., Tompkin, J., Xiao, L.: Gaufre: Gaussian deformation fields for real-time dynamic novel view synthesis (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.214, + 0.189, + 0.788, + 0.23 + ], + "angle": 0, + "content": "15. Liu, Q., Liu, Y., Wang, J., Lv, X., Wang, P., Wang, W., Hou, J.: Modgs: Dynamic gaussian splatting from causally-captured monocular videos (2024), https:// arxiv.org/abs/2406.00434" + }, + { + "type": "ref_text", + "bbox": [ + 0.214, + 0.23, + 0.788, + 0.271 + ], + "angle": 0, + "content": "16. Lu, T., Yu, M., Xu, L., Xiangli, Y., Wang, L., Lin, D., Dai, B.: Scaffold-gs: Structured 3d gaussians for view-adaptive rendering. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 20654-20664 (2024)" + }, + { + "type": "ref_text", + "bbox": [ + 0.214, + 0.271, + 0.788, + 0.298 + ], + "angle": 0, + "content": "17. Luiten, J., Kopanas, G., Leibe, B., Ramanan, D.: Dynamic 3d gaussians: Tracking by persistent dynamic view synthesis. In: 3DV (2024)" + }, + { + "type": "ref_text", + "bbox": [ + 0.214, + 0.298, + 0.788, + 0.338 + ], + "angle": 0, + "content": "18. Martin-Brualla, R., Radwan, N., Sajjadi, M.S.M., Barron, J.T., Dosovitskiy, A., Duckworth, D.: NeRF in the Wild: Neural Radiance Fields for Unconstrained Photo Collections. In: CVPR (2021)" + }, + { + "type": "ref_text", + "bbox": [ + 0.214, + 0.338, + 0.788, + 0.379 + ], + "angle": 0, + "content": "19. Mildenhall, B., Srinivasan, P.P., Tancik, M., Barron, J.T., Ramamoorthi, R., Ng, R.: Nerf: Representing scenes as neural radiance fields for view synthesis. Communications of the ACM 65(1), 99-106 (2021)" + }, + { + "type": "ref_text", + "bbox": [ + 0.214, + 0.379, + 0.788, + 0.406 + ], + "angle": 0, + "content": "20. Park, K., Sinha, U., Barron, J.T., Bouaziz, S., Goldman, D.B., Seitz, S.M., Martin-Brualla, R.: Nerfies: Deformable neural radiance fields. ICCV (2021)" + }, + { + "type": "ref_text", + "bbox": [ + 0.214, + 0.406, + 0.788, + 0.447 + ], + "angle": 0, + "content": "21. Park, K., Sinha, U., Hedman, P., Barron, J.T., Bouaziz, S., Goldman, D.B., Martin-Brualla, R., Seitz, S.M.: Hypernerf: A higher-dimensional representation for topologically varying neural radiance fields. arXiv preprint arXiv:2106.13228 (2021)" + }, + { + "type": "ref_text", + "bbox": [ + 0.214, + 0.447, + 0.788, + 0.473 + ], + "angle": 0, + "content": "22. Pumarola, A., Corona, E., Pons-Moll, G., Moreno-Noguer, F.: D-nerf: Neural radiance fields for dynamic scenes. arXiv preprint arXiv:2011.13961 (2020)" + }, + { + "type": "ref_text", + "bbox": [ + 0.214, + 0.473, + 0.788, + 0.528 + ], + "angle": 0, + "content": "23. Sabater, N., Boisson, G., Vandame, B., Kerbiriou, P., Babon, F., Hog, M., Gendrot, R., Langlois, T., Bureller, O., Schubert, A., et al.: Dataset and pipeline for multiview light-field video. In: Proceedings of the IEEE conference on computer vision and pattern recognition Workshops. pp. 30-40 (2017)" + }, + { + "type": "ref_text", + "bbox": [ + 0.214, + 0.528, + 0.788, + 0.582 + ], + "angle": 0, + "content": "24. Song, L., Chen, A., Li, Z., Chen, Z., Chen, L., Yuan, J., Xu, Y., Geiger, A.: Nerflayer: A streamable dynamic scene representation with decomposed neural radiance fields. IEEE Transactions on Visualization and Computer Graphics 29(5), 2732-2742 (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.214, + 0.582, + 0.788, + 0.624 + ], + "angle": 0, + "content": "25. Tancik, M., Casser, V., Yan, X., Pradhan, S., Mildenhall, B., Srinivasan, P.P., Barron, J.T., Kretzschmar, H.: Block-nerf: Scalable large scene neural view synthesis (2022)" + }, + { + "type": "ref_text", + "bbox": [ + 0.214, + 0.624, + 0.788, + 0.651 + ], + "angle": 0, + "content": "26. Ververas, E., Potamias, R.A., Song, J., Deng, J., Zafeiriou, S.: Sags: Structure-aware 3d gaussian splatting. arXiv:2404.19149 (2024)" + }, + { + "type": "ref_text", + "bbox": [ + 0.214, + 0.651, + 0.788, + 0.677 + ], + "angle": 0, + "content": "27. Wang, F., Tan, S., Li, X., Tian, Z., Liu, H.: Mixed neural voxels for fast multi-view video synthesis. arXiv preprint arXiv:2212.00190 (2022)" + }, + { + "type": "ref_text", + "bbox": [ + 0.214, + 0.677, + 0.788, + 0.718 + ], + "angle": 0, + "content": "28. Wang, Z., Bovik, A.C., Sheikh, H.R., Simoncelli, E.P.: Image quality assessment: from error visibility to structural similarity. IEEE transactions on image processing 13(4), 600-612 (2004)" + }, + { + "type": "ref_text", + "bbox": [ + 0.214, + 0.718, + 0.788, + 0.759 + ], + "angle": 0, + "content": "29. Wu, G., Yi, T., Fang, J., Xie, L., Zhang, X., Wei, W., Liu, W., Tian, Q., Xinggang, W.: 4d gaussian splatting for real-time dynamic scene rendering. arXiv preprint arXiv:2310.08528 (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.214, + 0.759, + 0.788, + 0.8 + ], + "angle": 0, + "content": "30. Yang, Z., Yang, H., Pan, Z., Zhang, L.: Real-time photorealistic dynamic scene representation and rendering with 4d gaussian splatting. In: International Conference on Learning Representations (ICLR) (2024)" + }, + { + "type": "ref_text", + "bbox": [ + 0.214, + 0.8, + 0.788, + 0.84 + ], + "angle": 0, + "content": "31. Yang, Z., Gao, X., Zhou, W., Jiao, S., Zhang, Y., Jin, X.: Deformable 3d gaussians for high-fidelity monocular dynamic scene reconstruction. arXiv preprint arXiv:2309.13101 (2023)" + }, + { + "type": "list", + "bbox": [ + 0.214, + 0.147, + 0.788, + 0.84 + ], + "angle": 0, + "content": null + } + ], + [ + { + "type": "page_number", + "bbox": [ + 0.218, + 0.116, + 0.236, + 0.127 + ], + "angle": 0, + "content": "16" + }, + { + "type": "header", + "bbox": [ + 0.272, + 0.115, + 0.355, + 0.127 + ], + "angle": 0, + "content": "J. Bae et al." + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.147, + 0.785, + 0.175 + ], + "angle": 0, + "content": "32. Yu, H., Julin, J., Milacski, Z.A., Niinuma, K., Jeni, L.A.: Cogs: Controllable gaussian splatting (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.175, + 0.785, + 0.217 + ], + "angle": 0, + "content": "33. Zhang, R., Isola, P., Efros, A.A., Shechtman, E., Wang, O.: The unreasonable effectiveness of deep features as a perceptual metric. In: Proceedings of the IEEE conference on computer vision and pattern recognition. pp. 586-595 (2018)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.218, + 0.785, + 0.299 + ], + "angle": 0, + "content": "34. Zhao, B., Li, Y., Sun, Z., Zeng, L., Shen, Y., Ma, R., Zhang, Y., Bao, H., Cui, Z.: Gaussian prediction: Dynamic 3d gaussian prediction for motion extrapolation and free view synthesis. In: Special Interest Group on Computer Graphics and Interactive Techniques Conference Papers aA24. SIGGRAPH aAZ24, ACM (Jul 2024). https://doi.org/10.1145/3641519.3657417, http://dx.doi.org/10.1145/3641519.3657417" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.3, + 0.785, + 0.342 + ], + "angle": 0, + "content": "35. Zwicker, M., Pfister, H., Van Baar, J., Gross, M.: Surface splattering. In: Proceedings of the 28th annual conference on Computer graphics and interactive techniques. pp. 371-378 (2001)" + }, + { + "type": "list", + "bbox": [ + 0.217, + 0.147, + 0.785, + 0.342 + ], + "angle": 0, + "content": null + } + ] +] \ No newline at end of file diff --git a/2024/Per-Gaussian Embedding-Based Deformation for Deformable 3D Gaussian Splatting/8af14cc6-a852-4c45-9f34-dffb20b17a0a_origin.pdf b/2024/Per-Gaussian Embedding-Based Deformation for Deformable 3D Gaussian Splatting/8af14cc6-a852-4c45-9f34-dffb20b17a0a_origin.pdf new file mode 100644 index 0000000000000000000000000000000000000000..03d4804c46492ff67480df3a39164211655135dd --- /dev/null +++ b/2024/Per-Gaussian Embedding-Based Deformation for Deformable 3D Gaussian Splatting/8af14cc6-a852-4c45-9f34-dffb20b17a0a_origin.pdf @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:9c3e5917abd69c4a01680f52fca0de2156099a90090c2fb834278bd3c4b45bfa +size 2917125 diff --git a/2024/Per-Gaussian Embedding-Based Deformation for Deformable 3D Gaussian Splatting/full.md b/2024/Per-Gaussian Embedding-Based Deformation for Deformable 3D Gaussian Splatting/full.md new file mode 100644 index 0000000000000000000000000000000000000000..e259e1f82ba82b0eb4d438361bbd7a8231ad4951 --- /dev/null +++ b/2024/Per-Gaussian Embedding-Based Deformation for Deformable 3D Gaussian Splatting/full.md @@ -0,0 +1,296 @@ +# Per-Gaussian Embedding-Based Deformation for Deformable 3D Gaussian Splitting + +Jeongmin Bae1*, Seoha Kim1*, Youngsik Yun1, Hahyun Lee2, Gun Bang2, and Youngjung Uh1 + +1 Yonsei University, Seoul 03722, Korea {jaymin.bae, hailey07, bbangsik, yj.uh}@yonsei.ac.kr + +$^{2}$ Electronics and Telecommunications Research Institute, Daejeon 34129, Korea {hanilee, gbang}@etri.re.kr + +![](images/6fcce4eccafc285e23868dd70db2ef6658f862c19476727352665328c2371815.jpg) +Fig. 1: Overview. (a) Existing deformable 3D Gaussian Splitting methods show blurry results in complex dynamic scenes, even with deformation fields using finer feature grids. (b) Our model solves the problem by employing per-Gaussian latent embeddings to predict deformations for each Gaussian and achieves a clearer results. + +![](images/d936d00b55af85cd3fe51ff366a7edc6a05aa8fe639f9a7a1557ecaba53e90b0.jpg) + +![](images/71408cd8124ebccd72cbaca806df1fa6aec9b530f4aeb22a3bdc4c309401d958.jpg) + +Abstract. As 3D Gaussian Splitting (3DGS) provides fast and high-quality novel view synthesis, it is a natural extension to deform a canonical 3DGS to multiple frames for representing a dynamic scene. However, previous works fail to accurately reconstruct complex dynamic scenes. We attribute the failure to the design of the deformation field, which is built as a coordinate-based function. This approach is problematic because 3DGS is a mixture of multiple fields centered at the Gaussians, not just a single coordinate-based framework. To resolve this problem, we define the deformation as a function of per-Gaussian embeddings and temporal embeddings. Moreover, we decompose deformations as coarse and fine deformations to model slow and fast movements, respectively. Also, we introduce a local smoothness regularization for per-Gaussian embedding to improve the details in dynamic regions. + +Project page: https://jeongminb.github.io/e-d3dns/ + +Keywords: Deformable gaussian splatting $\cdot$ 4D scene reconstruction $\cdot$ Novel view synthesis + +# 1 Introduction + +Dynamic scene reconstruction from multi-view input videos is an important task in computer vision, as it can be extended to various applications and industries such as mixed reality, content production, etc. Neural Radiance Fields (NeRF) [19], which enable photorealistic novel view synthesis from multi-view inputs, can represent dynamic scenes by modeling the scene with an additional time input [12, 22]. However, typical NeRFs require querying multilayer perceptron (MLP) for hundreds of points per camera ray, which limits rendering speed. + +On the other hand, the recently emerging 3D Gaussian Splatting (3DGS) [10] has the advantage of real-time rendering compared to NeRFs using a differentiable rasterizer for 3D Gaussian primitives. 3DGS directly optimizes the parameters of 3D Gaussians (position, opacity, anisotropic covariance, and spherical harmonics coefficients) and renders them via projection and $\alpha$ -blending. Since 3DGS has the characteristics of continuous volumetric radiance fields, some recent studies [3,8,14,29,31,32] represent dynamic scenes by defining a canonical 3DGS and deforming it to individual frames as deformable NeRFs [31] do. Specifically, they model the deformation as a function of 4D (x, y, z, t) coordinates with MLPs or grids to predict the change in the 3D Gaussian parameters. + +However, since 3DGS is a mixTURE of multiple volumetric fields, it is not appropriate to model the deformation of Gaussian parameters with a single coordinate-based network to represent dynamic scenes. In addition, existing field-based approaches are constrained by the resolution of the grid which models the deformation field, the capacity of the model, or the frequencies of the input. As shown in Figure 1, existing study does not properly represent complex dynamic scenes, and even introducing an additional feature grid that is twice the maximum resolution has only a slight improvement in performance (See Appendix for more results). We alleviate this problem by introducing a novel dynamic representation to deform each Gaussian. + +In this paper, we model the deformation of Gaussians at frames as 1) a function of a product space of per-Gaussian embeddings and temporal embeddings. We expect this rational design to bring quality improvement by precisely modeling different deformations of different Gaussians. Additionally, 2) We decompose temporal variations of the parameters into coarse and fine components, namely coarse-fine deformation. The coarse deformation represents large or slow movements in the scene, while fine deformation learns the fast or detailed movements that coarse deformation does not cover. Finally, we propose 3) a local smoothness regularization for per-Gaussian embedding to ensure the deformations of neighboring Gaussians are similar. + +In our experiments, we observe that our per-Gaussian embeddings, coarse-fine deformation, and regularization improve the deformation quality. Our approach outperforms baselines in capturing fine details in dynamic regions and excels even under challenging camera settings. Additionally, our method also achieves fast rendering speed and relatively low capacity. + +# 2 Related Work + +In this section, we review methods for dynamic scene reconstruction that deform 3D canonical space and methods for reconstructing dynamic scenes utilizing dynamic 3D Gaussians. Afterward, we review methods that use embeddings and spatial relationships of Gaussians. + +Deforming 3D Canonical Space D-NeRF [22] reconstructs dynamic scenes by deforming ray samples over time, using the deformation network that takes 3D coordinates and timestamps of the sample as inputs. Nerfies [20] and HyperNeRF [21] use per-frame trainable deformation codes instead of time conditions to deform the canonical space. Instead of deforming from the canonical frame to the entire frames, HyperReel [1] deforms the ray sample of the keyframe to represent the intermediate frame. 4DGaussians [29] and D3DGS [31] reconstruct the dynamic scene with a deformation network which inputs the center position of the canonical 3D Gaussians and timestamps. MoDGS [15] learns the mapping between canonical space and space at specific timestamp through invertible MLP. In contrast, we demonstrate a novel deformation representation as a function of a product space of per-Gaussian latent embeddings and temporal embeddings. + +Dynamic 3D Gaussians To extend the fast rendering speed of 3D Gaussian Splitting [10] into dynamic scene reconstructions. 4D Gaussians [29] decodes features from multi-resolution HexPlanes [2] for temporal deformation of 3D Gaussians. While D3DGS [31] uses an implicit function that processes the time and location of the Gaussian. 4DGS [30] decomposes the 4D Gaussians into a time-conditioned 3D Gaussians and a marginal 1D Gaussians. STG [13] represents changes in 3D Gaussian over time through a temporal opacity and a polynomial function for each Gaussian. + +Our method uses deformable 3D Gaussians as 4D Gaussians [29] and D3DGS [31] do, but does not necessitate the separated feature field to obtain the input feature of the deformation decoder. Our approach uses embeddings allocated to each Gaussian and a temporal embedding shared within a specific frame. + +Latent Embedding on Novel View Synthesis Some studies incorporate latent embeddings to represent different states of the static and dynamic scene. NeRF-W [18] and Block-NeRF [25] employ per-image embeddings to capture different appearances of a scene, representing the scenes from in-the-wild image collections. DyNeRF and MixVoxels [12, 27] employ a temporal embedding for each frame to represent dynamic scenes. Nerfies [20] and HyperNeRF [21] incorporate both per-frame appearance and deformation embeddings. Sync-NeRF [11] introduces time offset to calibrate the misaligned temporal embeddings on dynamic scenes from unsynchronized videos. We introduce per-Gaussian latent embedding to encode the changes over time of each Gaussian and use temporal embeddings to represent different states in each frame of the scene. + +Considering Spatial Relationships of Gaussians Scaffold-GS [16] reconstruct 3D scenes by synthesizing Gaussians from anchors, utilizing that the neighboring Gaussians have similar properties. SAGS [26] creates a graph based on k-nearest neighbors (KNN) so that each Gaussian is optimized while considering its neighboring Gaussians. In dynamic scene reconstruction, SC-GS [9] and GaussianPrediction [34] deform Gaussians by combining the deformations of key point Gaussians. Dynamic 3D Gaussians [17] utilizes regularization to encourage that Gaussians and their neighboring Gaussians deform with local rigidity. Similarly, we propose a local smoothness regularization that encourages neighboring Gaussians to have similar embeddings, resulting in similar deformations. + +# 3 Method + +In this section, we first provide a brief overview of 3D Gaussian Splatting (Section 3.1). Next, we introduce our overall framework, embedding-based deformation for Gaussians (Section 3.2) and coarse-fine deformation scheme consisting of coarse and fine deformation functions (Section 3.3). Finally, we present a local smoothness regularization for per-Gaussian embeddings to achieve better details on dynamic regions (Section 3.4). + +# 3.1 Preliminary: 3D Gaussian Splitting + +3D Gaussian splattering [10] optimizes a set of anisotropic 3D Gaussians through differentiable tile rasterization to reconstruct a static 3D scene. By its efficient rasterization, the optimized model enables real-time rendering of high-quality images. Each 3D Gaussian kernel $G_{i}(x)$ at the point $x$ consists with position $\mathbf{x}_i$ , rotation $R_i$ , and scale $S_i$ : + +$$ +G _ {i} (x) = e ^ {- \frac {1}{2} (x - \mathbf {x} _ {i}) ^ {T} \Sigma_ {i} ^ {- 1} (x - \mathbf {x} _ {i})}, \quad \text {w h e r e} \quad \Sigma_ {i} = R _ {i} S _ {i} S _ {i} ^ {T} R _ {i} ^ {T}. \tag {1} +$$ + +To projecting 3D Gaussians to 2D for rendering, covariance matrix $\Sigma^{\prime}$ are calculated by viewing transform $W$ and the Jacobian $J$ of the affine approximation of the projective transformation [35] as follows: + +$$ +\Sigma^ {\prime} = J W \Sigma W ^ {T} J ^ {T}. \tag {2} +$$ + +Blending $\mathcal{N}$ depth-ordered projected points that overlap the pixel, the Gaussian kernel $G_{i}(x)$ is multiplied by the opacity of the Gaussian $\sigma_{i}$ and calculates the pixel color $C$ with the color of the Gaussian $c_{i}$ : + +$$ +C = \sum_ {i \in \mathcal {N}} c _ {i} \alpha_ {i} \prod_ {j = 1} ^ {i - 1} (1 - \alpha_ {j}), \quad \text {w h e r e} \quad \alpha_ {i} = \sigma_ {i} G _ {i} (x). \tag {3} +$$ + +The color of Gaussian $c_{i}$ is determined using the SH coefficient with view direction. + +![](images/5a6ada8e19479be6d76f7158e58487be4d65b99cfd38a791de1ac921bcd6b476.jpg) + +![](images/b92e21303c10ae6fcf13d58c84f42e4fa8bbf5dfcf03143fce17f66da01e6ef4.jpg) +(a) Embeddings for deformation + +![](images/0732f9a2f4bbb5c92e44341a0fd45613b07fcc21609fc5458e11d5f0abd3e5cd.jpg) +(b) Per-Gaussian deformation + +![](images/afd61294da1e2007c76213a87be27107281fbaa0c8329bae0a3df65ca12b0be2.jpg) +(c) Local smoothness regularization +Fig. 2: Framework. Existing coordinate-based network methods struggle to represent complex dynamic scenes. To this end, we define per-Gaussian deformation. (a) Firstly, we assign a latent embedding for each Gaussian. Additionally, we introduce coarse and fine temporal embeddings to represent the slow and fast state of the dynamic scene. (b) By employing two decoders that take per-Gaussian latent embeddings along with coarse and fine temporal embeddings as input, we estimate slow or large changes and fast or detailed changes to model the final deformation, respectively. (c) Finally, we introduce a local smoothness regularization so that the embeddings of neighboring Gaussians are similar. + +# 3.2 Embedding-Based Deformation for Gaussians + +Deformable NeRFs consist of a deformation field that predicts displacement $\Delta \mathbf{x}$ for a given coordinate $\mathbf{x}$ from the canonical space to each target frame, and a radiance field that maps color and density from a given coordinate in the canonical space $(\mathbf{x} + \Delta \mathbf{x})$ . Existing deformable Gaussian methods employ the same approach for predicting the deformation of Gaussians, i.e., utilizing a deformation field based on coordinates. + +Unlike previous methods, we start from the design of 3DGS: the 3D scene is represented as a mixture of Gaussians that have individual radiance fields. Accordingly, the deformation should be defined for each Gaussian. Based on this intuition, we introduce a function $\mathcal{F}_{\theta}$ that produces deformation from learnable embeddings $\mathbf{z}_{\mathrm{g}} \in \mathbb{R}^{32}$ belonging to individual Gaussians (Figure 2a), and typical temporal embeddings $\mathbf{z}_t \in \mathbb{R}^{256}$ for different frames: + +$$ +\mathcal {F} _ {\theta}: \left(\mathbf {z} _ {\mathrm {g}}, \mathbf {z} _ {t}\right)\rightarrow \left(\Delta \mathbf {x}, \Delta \mathbf {r}, \Delta \mathbf {s}, \Delta \sigma , \Delta Y\right), \tag {4} +$$ + +where $\mathbf{r}$ is a rotation quaternion, $\mathbf{s}$ is a vector for scaling, $\sigma$ is an opacity, and $Y$ is SH coefficients for modeling view-dependent color. We implement $\mathcal{F}_{\theta}$ as a shallow multi-layer perceptron (MLP) followed by an MLP head for each parameter. As a result, the Gaussian parameters at frame $t$ are determined by adding $\mathcal{F}_{\theta}(\mathbf{z}_{\mathrm{g}},\mathbf{z}_t)$ to the canonical Gaussian parameters (Figure 2c). + +We jointly optimize the per-Gaussian embeddings $\mathbf{z}_{\mathrm{g}}$ , the deformation function $\mathcal{F}_{\theta}$ , and the canonical Gaussian parameters to minimize the rendering loss. We use the L1 and periodic DSSIM as the rendering loss between the rendered image and the ground truth image. + +# 3.3 Coarse-Fine Deformation + +Different parts of a scene may have coarse and fine motions [5]. E.g., a hand swiftly stirs a pan (fine) while a body slowly moves from left to right (coarse). Based on this intuition, we introduce a coarse-fine deformation that produces a summation of coarse and fine deformations. + +Coarse-fine deformation consists of two functions with the same architecture: one for coarse and one for fine deformation (Figure 2c). The functions receive different temporal embeddings as follows: + +Following typical temporal embeddings, we start from a 1D feature grid $Z \in \mathbb{R}^{N \times 256}$ for $N$ frames and use an embedding $\mathbf{z}_t^{\mathrm{f}} = \mathrm{interp}(Z,t)$ for fine deformation. For coarse deformation, we linearly downsample $Z$ by a factor of 5 to remove high-frequencies responsible for fast and detailed deformation. Then we compute $\mathbf{z}_t^{\mathrm{c}}$ as a linear interpolation of embeddings at enclosing grid points (Figure 2b). + +As a result, coarse deformation $\mathcal{F}_{\theta_{\mathrm{c}}}(\mathbf{z}_{\mathrm{g}},\mathbf{z}_{t}^{c})$ is responsible for representing large or slow movements in the scene, while fine deformation $\mathcal{F}_{\theta_{\mathrm{f}}}(\mathbf{z}_{\mathrm{g}},\mathbf{z}_{t}^{\mathrm{f}})$ learns the fast or detailed movements that coarse deformation does not cover. This improves the deformation quality. Refer to the Ablation study section for more details. + +# 3.4 Local Smoothness Regularization + +Neighboring Gaussians constructing dynamic objects tend to exhibit locally similar deformation. Inspired by [17], we introduce a local smoothness regularization for per-Gaussian embedding $\mathbf{z}_{\mathrm{g}}$ (Figure 2d) to encourage similar deformations between nearby Gaussians $i$ and $j$ : + +$$ +\mathcal {L} ^ {\mathrm {e m b} - \mathrm {r e g}} = \frac {1}{k | \mathcal {S} |} \sum_ {i \in \mathcal {S}} \sum_ {j \in \mathrm {K N N} _ {i; k}} \left(w _ {i, j} \| \mathbf {z} _ {\mathrm {g} i} - \mathbf {z} _ {\mathrm {g} j} \| _ {2}\right), +$$ + +where $w_{i,j} = \exp (-\lambda_w\| \mu_j - \mu_i\| _2^2)$ is the weighting factor and $\mu$ is the Gaussian center. We set $\lambda_w$ to 2000 and $k$ to 20 following [17]. To reduce the computational cost, we obtain sets of k-nearest-neighbors only when the densification occurs. + +Note that unlike previous approaches that directly constrain physical properties such as rigidity or rotation, We implicitly induce locally similar deformation by ensuring that per-Gaussian embeddings are locally smooth. Our regularization allows better capture of textures and details of dynamic objects. + +# 4 Experiment + +In this section, we first describe the criterion for selection of baselines, and evaluation metrics. We then demonstrate the effectiveness of our method through comparisons with various baselines and datasets (Section 4.1-4.2). Finally, we conduct analysis and ablations of our method (Section 4.3). + +Table 1: Average performance in the test view on Neural 3D Video Dataset The computational cost was measured based on flame_salmon_1 on the A6000. ${}^{1}$ flame salmon scene includes only the first segment, comprising 300 frames. ${}^{2}$ reported time and DyNeRF is trained on 8 GPUs and tested only on flame salmon. ${}^{3}$ trained with 90 frames. ${}^{4}$ trained with 50 frames. + +
modelmetriccomputational cost
PSNR↑SSIM↑LPIPS↓Training time↓FPS↑Model size↓
DyNeRF12 [12]29.58-0.0831344 hours0.0156 MB
NeRFPlayer13 [24]30.69-0.1116 hours0.051654 MB
Mix Voxels [27]30.300.9180.1271 hours 40 mins0.93512 MB
K-Planes [6]30.860.9390.0961 hours 40 mins0.13309 MB
HyperReel4 [1]30.370.9210.1069 hours 20 mins1.01362 MB
4DGS [30]31.190.9400.0519 hours 30 mins33.78700 MB
4DGaussians [29]30.710.9350.05650 mins51.959 MB
Ours31.310.9450.0371 hours 52 mins74.535 MB
+ +Baselines We choose the start-of-the-art method as a baseline in each dataset. We compared against DyNeRF, NeRFPlayer, MixVoxels, K-Planes, HyperReel, Nerfies, HyperNeRF, and TiNeuVox on the NeRF baseline. In detail, we use the version of NeRFPlayer TensoRF VM, HyperNeRF DF, Mixvoxels-L, K-Planes hybrid. We compared with 4DGaussians, 4DGS, and D3DGS based on the Gaussian baseline. Meanwhile, we have not included STG in our comparison due to its requirement for per-frame Structure from Motion (SfM) points, which makes conducting a fair comparison challenging. Also, STG is not a deformable 3D Gaussian approach. We followed the official code and configuration, except increasing the training iterations for the Technicolor dataset to 1.5 times that of the 4DGaussians, in comparison to the Neural 3D Video dataset. + +Metrics We report the quality of rendered images using PSNR, SSIM, and LPIPS. Peak Signal-to-Noise Ratio (PSNR) quantifies pixel color error between the rendered video and the ground truth. We utilize SSIM [28] to account for the perceived similarity of the rendered image. Additionally, we measure higher-level perceptual similarity using Learned Perceptual Image Patch Similarity (LPIPS) [33] with an AlexNet Backbone. Higher PSNR and SSIM values and lower LPIPS values indicate better visual quality. + +# 4.1 Effectiveness on Dynamic Region + +Neural 3D Video Dataset [12] includes 20 multi-view videos, with each scene consisting of either 300 frames, except for the flame_salmon scene, which comprises 1200 frames. These scenes encompass a relatively long duration and various movements, with some featuring multiple objects in motion. We utilized Neural 3D Video dataset to observe the capability to capture dynamic areas. Total six scenes (coffee_martini, cook_s spinach, cut_ roasted_beef, flame_salmon, flame_ steak, sear_ steak) are evaluated in Figure 3 and Table 1. The flame_salmon scene divided into four segments, each containing 300 frames. + +![](images/a175a53e84dca37bd016a27fc78420aa86657301a2861a78a0e1deb900b85278.jpg) +Fig. 3: Qualitative comparisons on the Neural 3D Video Dataset. + +Table 1 presents quantitative metrics on the average metrics across the test views of all scenes, computational and storage costs on the first fragment of flame_salmon scene. Refer to the Appendix for per-scene details. Our method demonstrates superior reconstruction quality, FPS, and model size across compared to baselines. As the table shows, NeRF baselines generally required longer training and rendering times. While 4DGS shows relatively high reconstruction performance, it demands longer training times and larger VRAM storage capacity compared to other baselines. 4DGaussians requires lower computational and storage cost but it displays low reconstruction quality in some scenes with rapid dynamics, as shown in the teaser and Figure 3. + +Figure 3 reports the rendering quality. Our method successfully reconstructs the fine details in moving areas, outperforming baselines on average metrics across test views. Baselines show blurred dynamic areas or severe artifacts in low-light scenes such as cook_spinach and flame_steak. 4DGS exhibits the disappearance of some static areas. In 4DGaussians, a consistent over-smoothing occurs in dynamic areas. All baselines experienced reduced quality in reflective or thin dynamic areas like clamps, torches, and windows. + +Table 2: Average performance in the test view on Technicolor Dataset + +
modelmetriccomputational cost
PSNR↑SSIM↑LPIPS↓Training time↓FPS↑Model size↓
DyNeRF31.80-0.140-0.020.6 MB
HyperReel32.320.8990.1182 hours 45 mins0.91289 MB
4DGaussians29.620.8440.17625 mins34.851 MB
Ours33.240.9070.1002 hours 55 mins60.877 MB
+ +![](images/7b89d6665a098b61b9ff43f633b4ccaaeaac2f65d6bfac8f0a81fe4c2d88105d.jpg) +Fig. 4: Qualitative comparisons on the Technicolor Dataset. + +Technicolor Light Field Dataset [23] is a multi-view dataset captured with a time-synchronized $4 \times 4$ camera rig, containing intricate details. We train ours and the baselines on 50 frames of five commonly used scenes (Birthday, Fabien, Painter, Theater, Trains) using full-resolution videos at $2048 \times 1088$ pixels, with the second row and second column cameras used as test views. + +Table 2 reports the average metrics across the test views of all scenes, computational and storage costs on Painter scene. HyperReel demonstrates overall high-quality results but struggles with relatively slow training times and FPS, and a larger model size. 4DGaussians exhibits fast training times and FPS but significantly underperforms in reconstructing fine details compared to other baselines. However, our method demonstrates superior reconstruction quality and faster FPS compared to the baselines. + +As shown in Figure 4, HyperReel produces noisy artifacts due to incorrect predictions of the displacement vector. 4D Gaussians fails to capture fine details in dynamic areas, exhibiting over-smoothing results. All baselines struggle to accurately reconstruct rapidly moving thin areas like fingers. + +# 4.2 Scenes with Challenging Camera Setting + +Table 3: Average performance in the test view on Hypernerf Dataset + +
modelmetriccomputational cost
PSNR↑SSIM↑LPIPS↓Training time↓FPS↑Model size↓
Nerfies [20]22.23-0.170~ hours< 1-
HyperNeRF DS [21]22.290.5980.15332 hours< 115 MB
TiNeuVox [4]24.200.6160.3933 hours 30 mins148 MB
D3DGS [31]22.400.5980.27516 mins6.95309 MB
4DGaussians25.030.6820.28150 mins96.360 MB
Ours25.430.6970.2311 hours 15 mins139.333 MB
+ +![](images/0af1d5d643a009ddb43d392d4dda53c65c6d78129889c44152499a7f00721d7d.jpg) +Fig. 5: Qualitative comparisons on the HyperNeRF Dataset. + +HyperNeRF dataset includes videos captured using two phones rigidly mounted on a handheld rig. We train on all frames of four scenes (3D Printer, Banana, Broom, Chicken) at a resolution downsampled by half to $536 \times 960$ . Due to memory constraints, D3DGS is trained on images downsampled by a quarter. + +The table shows that our method outperforms the reconstruction performance with previous methods along with compact model size and faster FPS. Figure 5 shows that previous methods struggle to reconstruct fast-moving parts such as fingers and broom. Especially D3DGS deteriorates in Broom scene. Table 3 reports the average metrics across the test views of all scenes, computational and storage costs on Broom scene. + +# 4.3 Analyses and Ablation Study + +![](images/554182633416a77d1122a24ee13aca9acf78d3e7386f4b74f0a86321b399cf3f.jpg) +(a) canonical rendering + +![](images/ef95661573a08819491da6e91681cbeac3e0e8a019056f9fd124f980af81b83d.jpg) +(b) coarse deform only + +![](images/f50c0b5dffd6ba9cf2bff68b59325df636266e36b8d7c187275bfb9149d0f806.jpg) +(c) fine deform only +Fig. 6: Deformation components. (a) The canonical space contains Gaussians to represent all target frames of the scene. (b) Applying coarse deformation to the canonical space roughly reflects the dynamics of the scene. (c) The rendering without coarse deformation and only with fine deformation looks similar to the canonical rendering, i.e., responsible for fine deformations. (d) Applying both coarse and fine deformation yields natural rendering results. + +![](images/18b121f78a54ca89cbabdf980caff35cc0c5327d1255bd859b62f672dc88a3be.jpg) +(d) full rendering + +![](images/fb76741ca4e880981dcf037fa4e559ba30b0687e2bde18d7bc9053a7b61b7084.jpg) +Frame 67 +Fig.7: Visualization of the magnitude of deformation. Coarse deformation (blue) captures large and slow changes, such as the movement of the head and torso, while fine deformation (red) is responsible for the fast and detailed movements of arms, tongs, shadows, etc. + +![](images/50d55bb686db8f59500bd3aa1cf5cfb1e87af5d1b32675153dff4ad99b305459.jpg) +Frame 185 + +Deformation components In Figure 6, we present an analysis of the coarse-fine deformation. To achieve this, we render a flame_steak scene by omitting each of our deformation components one by one. Our full rendering results from adding coarse and fine deformation to the canonical space (Figure 6d). When both are removed, rendering yields an image in canonical space (Figure 6a). Rendering with the coarse deformation, which handles large or slow changes in the scene, produces results similar to the full rendering (Figure 6b). On the other hand, fine deformation is responsible for fast or detailed changes in the scene, yielding rendering similar to canonical space (Figure 6c). + +![](images/2bca6504c44f69bb82078709672f279a3f0355b578947c07ed64af8cc70a0d42.jpg) +(a) Ours Full + +![](images/e0e64da9a3a3a1ee90ae7561f2342964b90def5092752639732998512bd40911.jpg) +(b) coarse deform only +Fig. 8: Qualitative ablation results on coarse-fine deformation. (a) Our model achieves clear results with both coarse and fine decoders. (b-c) The quality of dynamic areas decreases if one is missing. (d) Additionally, introducing the coordinates of Gaussian as an additional input into our decoders results in a decrease in the quality of both static and dynamic regions. + +![](images/4395843b08f134c7c78a9cdf10b0cf79b6f072e54949ccd9e6aa5029378afb78.jpg) +(c) fine deform only + +![](images/bf4ba8ceca3ad52abc64a7819dd90ae4f1080d641cecd2f61b3272451caa9f09.jpg) +(d) Ours $+\mathbf{x}$ injection + +Table 4: Quantitative ablation results on coarse-fine deformation. + +
MethodPSNR↑SSIM↑LPIPS↓
Ours29.700.9330.041
coarse deformation only29.480.9310.044
fine deformation only29.230.9320.043
Ours + x injection29.600.9310.045
+ +To examine the roles of the coarse and fine deformation in the coarse-fine deformation, we conduct a visualization on flame_steak scene. First, we compute the Euclidean norm of positional shifts between the current and subsequent frames. We then add the value to the DC components of the SH coefficients proportionally to the magnitude: blue for coarse deformation and red for fine deformation. For visual clarity, we render the original scene in grayscale. As illustrated in Figure 7, coarse deformation models slower changes such as body movement, while fine deformation models faster movements like cooking arms. Thus, we demonstrate that by downsampling the temporal embedding grid $Z$ , we can effectively separate and model slow and fast deformations in the scene. + +Ablation study We report the results of an ablation study on the deformation decoder in Figure 8 and Table 4. First, our full method (using both coarse and fine decoders) produces clear rendering results and models dynamic variations well (Figure 8a). Training only with the coarse or fine decoder leads to blurred dynamic areas and a failure to accurately capture detailed motion (Figure 8b-c). Additionally, we demonstrate experiments where the Gaussian center coordinates $\mathbf{x}$ are injected into the input of each decoder. As shown in Figure 8d, including the Gaussian coordinates degrades the quality of deformation, supporting our argument that coordinate dependency should be removed from the deformation function. + +Furthermore, we report the results of an ablation study on the local smoothness regularization for per-Gaussian embeddings. As shown in Figure 9, our regularization improves the details and texture quality of moving objects. In Table 5, we show a performance comparison between the proposed regularization and + +Table 5: Quantitative ablation results on local smoothness regularization. We compare the performance of applying our regularization with the physically-based regularization of Dynamic 3D Gaussians [17]. Our regularization better captures details of dynamic objects. + +
MethodPSNR↑SSIM↑LPIPS↓
Ours w/o embedding reg32.260.9510.037
+ our embedding reg32.340.9520.036
+ physically-based reg32.080.9500.036
+ +![](images/d3bae3efa4e225290e81c2d17779d77a3ca69c066c77d6dc77f115a31d42a4d0.jpg) +Ours full + +![](images/c36f81fa9ab94dd9ac95aacb559f7ef87d408ee6efb558c2020798abcf59085f.jpg) +Ours w/o reg +Fig. 9: Qualitative ablation results on local smoothness regularization. + +![](images/8b07b1a8212952eb21f2114145017091c1187a1b426506870859cc4dab973fbe.jpg) +Ours full + +![](images/06618c98b077500de0b23326417919b4bc8618a7dc9b5629c67c1a56fee04333.jpg) +Ours w/o reg + +the existing physically-based regularizations. To apply the method proposed by Luiten et al. [17] to ours, we make some modifications: 1) Like our method, we find the set of k-nearest-neighbours only when the densification occurs to reduce the computational cost. 2) For long-term local-isometry loss, we use the time of video frame used in previous training step instead of using the time of the first frame. Our regularization is simple and shows better performance compared to previous method. + +# 5 Conclusion and Limitation + +We propose a per-Gaussian deformation for 3DGS that takes per-Gaussian embeddings as input, instead of using the typical deformation fields from previous deformable 3DGS works, resulting in high performance. We enhance the reconstruction quality by decomposing the dynamic changes into coarse and fine deformation. However, our method learns inappropriate Gaussian deformation with casually captured monocular videos [7], like other baselines. We plan to address it in future work by introducing useful prior for monocular video settings. + +![](images/6244d426dee9c97c7a6af5d392b85dc2e62d9930adb0f449e27e8ee2e9bf65d4.jpg) +4DGaussians +Fig. 10: Limitation. Ours struggles with the casually captured monocular videos. + +![](images/f4931a4dd7763d923edda247a6db3fd6060e5ba8252a115078cb119a2e7cfd80.jpg) +Ours + +![](images/6ac9a1640c0e847e7259dda48c198e6168348e5d8b5c2bc45e4371ad1b744732.jpg) +4DGaussians + +![](images/6097d5b27e045786fd13e8636bf6c77675c427cc4df28eedc04e4c0608ef8c21.jpg) +Ours + +# Acknowledgements + +This work is supported by the Institute for Information & Communications Technology Planning & Evaluation (IITP) grant funded by the Korea government(MSIT) (No. 2017-0-00072, Development of Audio/Video Coding and Light Field Media Fundamental Technologies for Ultra Realistic Tera-media) + +# References + +1. Attal, B., Huang, J.B., Richardt, C., Zollhoefer, M., Kopf, J., OaAZToole, M., Kim, C.: Hyperreel: High-fidelity 6-dof video with ray-conditioned sampling. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 16610-16620 (2023) +2. Cao, A., Johnson, J.: Hexplane: A fast representation for dynamic scenes. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 130-141 (2023) +3. Duisterhof, B.P., Mandi, Z., Yao, Y., Liu, J.W., Shou, M.Z., Song, S., Ichnowski, J.: Md-splatting: Learning metric deformation from 4d gaussians in highly deformable scenes. arXiv preprint arXiv:2312.00583 (2023) +4. Fang, J., Yi, T., Wang, X., Xie, L., Zhang, X., Liu, W., Nießner, M., Tian, Q.: Fast dynamic radiance fields with time-aware neural voxels. In: SIGGRAPH Asia 2022 Conference Papers (2022) +5. Feichtenhofer, C., Fan, H., Malik, J., He, K.: Slowfast networks for video recognition. In: Proceedings of the IEEE/CVF international conference on computer vision. pp. 6202-6211 (2019) +6. Fridovich-Keil, S., Meanti, G., Warburg, F.R., Recht, B., Kanazawa, A.: K-planes: Explicit radiance fields in space, time, and appearance. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 12479-12488 (2023) +7. Gao, H., Li, R., Tulsiani, S., Russell, B., Kanazawa, A.: Monocular dynamic view synthesis: A reality check. In: NeurIPS (2022) +8. Huang, Y.H., Sun, Y.T., Yang, Z., Lyu, X., Cao, Y.P., Qi, X.: Sc-gs: Sparse-controlled gaussian splatting for editable dynamic scenes. arXiv preprint arXiv:2312.14937 (2023) +9. Huang, Y.H., Sun, Y.T., Yang, Z., Lyu, X., Cao, Y.P., Qi, X.: Sc-gs: Sparse-controlled gaussian splatting for editable dynamic scenes. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR). pp. 4220-4230 (June 2024) +0. Kerbl, B., Kopanas, G., Leimkuhler, T., Drettakis, G.: 3d gaussian splatting for real-time radiance field rendering. ACM Transactions on Graphics 42(4) (July 2023), https://repo-sam.inria.fr/fungraph/3d-gaussian-splatting/ +1. Kim, S., Bae, J., Yun, Y., Lee, H., Bang, G., Uh, Y.: Sync-nerf: Generalizing dynamic nerfs to unsynchronized videos. arXiv preprint arXiv:2310.13356 (2023) +12. Li, T., Slavcheva, M., Zollhoefer, M., Green, S., Lassner, C., Kim, C., Schmidt, T., Lovegrove, S., Goesele, M., Newcombe, R., et al.: Neural 3d video synthesis from multi-view video. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (2022) +13. Li, Z., Chen, Z., Li, Z., Xu, Y.: Spacetime gaussian feature splatting for real-time dynamic view synthesis. arXiv preprint arXiv:2312.16812 (2023) + +14. Liang, Y., Khan, N., Li, Z., Nguyen-Phuoc, T., Lanman, D., Tompkin, J., Xiao, L.: Gaufre: Gaussian deformation fields for real-time dynamic novel view synthesis (2023) +15. Liu, Q., Liu, Y., Wang, J., Lv, X., Wang, P., Wang, W., Hou, J.: Modgs: Dynamic gaussian splatting from causally-captured monocular videos (2024), https:// arxiv.org/abs/2406.00434 +16. Lu, T., Yu, M., Xu, L., Xiangli, Y., Wang, L., Lin, D., Dai, B.: Scaffold-gs: Structured 3d gaussians for view-adaptive rendering. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 20654-20664 (2024) +17. Luiten, J., Kopanas, G., Leibe, B., Ramanan, D.: Dynamic 3d gaussians: Tracking by persistent dynamic view synthesis. In: 3DV (2024) +18. Martin-Brualla, R., Radwan, N., Sajjadi, M.S.M., Barron, J.T., Dosovitskiy, A., Duckworth, D.: NeRF in the Wild: Neural Radiance Fields for Unconstrained Photo Collections. In: CVPR (2021) +19. Mildenhall, B., Srinivasan, P.P., Tancik, M., Barron, J.T., Ramamoorthi, R., Ng, R.: Nerf: Representing scenes as neural radiance fields for view synthesis. Communications of the ACM 65(1), 99-106 (2021) +20. Park, K., Sinha, U., Barron, J.T., Bouaziz, S., Goldman, D.B., Seitz, S.M., Martin-Brualla, R.: Nerfies: Deformable neural radiance fields. ICCV (2021) +21. Park, K., Sinha, U., Hedman, P., Barron, J.T., Bouaziz, S., Goldman, D.B., Martin-Brualla, R., Seitz, S.M.: Hypernerf: A higher-dimensional representation for topologically varying neural radiance fields. arXiv preprint arXiv:2106.13228 (2021) +22. Pumarola, A., Corona, E., Pons-Moll, G., Moreno-Noguer, F.: D-nerf: Neural radiance fields for dynamic scenes. arXiv preprint arXiv:2011.13961 (2020) +23. Sabater, N., Boisson, G., Vandame, B., Kerbiriou, P., Babon, F., Hog, M., Gendrot, R., Langlois, T., Bureller, O., Schubert, A., et al.: Dataset and pipeline for multiview light-field video. In: Proceedings of the IEEE conference on computer vision and pattern recognition Workshops. pp. 30-40 (2017) +24. Song, L., Chen, A., Li, Z., Chen, Z., Chen, L., Yuan, J., Xu, Y., Geiger, A.: Nerflayer: A streamable dynamic scene representation with decomposed neural radiance fields. IEEE Transactions on Visualization and Computer Graphics 29(5), 2732-2742 (2023) +25. Tancik, M., Casser, V., Yan, X., Pradhan, S., Mildenhall, B., Srinivasan, P.P., Barron, J.T., Kretzschmar, H.: Block-nerf: Scalable large scene neural view synthesis (2022) +26. Ververas, E., Potamias, R.A., Song, J., Deng, J., Zafeiriou, S.: Sags: Structure-aware 3d gaussian splatting. arXiv:2404.19149 (2024) +27. Wang, F., Tan, S., Li, X., Tian, Z., Liu, H.: Mixed neural voxels for fast multi-view video synthesis. arXiv preprint arXiv:2212.00190 (2022) +28. Wang, Z., Bovik, A.C., Sheikh, H.R., Simoncelli, E.P.: Image quality assessment: from error visibility to structural similarity. IEEE transactions on image processing 13(4), 600-612 (2004) +29. Wu, G., Yi, T., Fang, J., Xie, L., Zhang, X., Wei, W., Liu, W., Tian, Q., Xinggang, W.: 4d gaussian splatting for real-time dynamic scene rendering. arXiv preprint arXiv:2310.08528 (2023) +30. Yang, Z., Yang, H., Pan, Z., Zhang, L.: Real-time photorealistic dynamic scene representation and rendering with 4d gaussian splatting. In: International Conference on Learning Representations (ICLR) (2024) +31. Yang, Z., Gao, X., Zhou, W., Jiao, S., Zhang, Y., Jin, X.: Deformable 3d gaussians for high-fidelity monocular dynamic scene reconstruction. arXiv preprint arXiv:2309.13101 (2023) + +32. Yu, H., Julin, J., Milacski, Z.A., Niinuma, K., Jeni, L.A.: Cogs: Controllable gaussian splatting (2023) +33. Zhang, R., Isola, P., Efros, A.A., Shechtman, E., Wang, O.: The unreasonable effectiveness of deep features as a perceptual metric. In: Proceedings of the IEEE conference on computer vision and pattern recognition. pp. 586-595 (2018) +34. Zhao, B., Li, Y., Sun, Z., Zeng, L., Shen, Y., Ma, R., Zhang, Y., Bao, H., Cui, Z.: Gaussian prediction: Dynamic 3d gaussian prediction for motion extrapolation and free view synthesis. In: Special Interest Group on Computer Graphics and Interactive Techniques Conference Papers aA24. SIGGRAPH aAZ24, ACM (Jul 2024). https://doi.org/10.1145/3641519.3657417, http://dx.doi.org/10.1145/3641519.3657417 +35. Zwicker, M., Pfister, H., Van Baar, J., Gross, M.: Surface splattering. In: Proceedings of the 28th annual conference on Computer graphics and interactive techniques. pp. 371-378 (2001) \ No newline at end of file diff --git a/2024/Per-Gaussian Embedding-Based Deformation for Deformable 3D Gaussian Splatting/images.zip b/2024/Per-Gaussian Embedding-Based Deformation for Deformable 3D Gaussian Splatting/images.zip new file mode 100644 index 0000000000000000000000000000000000000000..8dec8971ce27d32118057e532b94e6f127d2e52b --- /dev/null +++ b/2024/Per-Gaussian Embedding-Based Deformation for Deformable 3D Gaussian Splatting/images.zip @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:1a35ec3830ebf7940440421d40512d33fef46d7f9f9fb4b4172079d2a2f959d4 +size 665561 diff --git a/2024/Per-Gaussian Embedding-Based Deformation for Deformable 3D Gaussian Splatting/layout.json b/2024/Per-Gaussian Embedding-Based Deformation for Deformable 3D Gaussian Splatting/layout.json new file mode 100644 index 0000000000000000000000000000000000000000..5f58bdba391293e4ef622230c9e0fc63c79d9c47 --- /dev/null +++ b/2024/Per-Gaussian Embedding-Based Deformation for Deformable 3D Gaussian Splatting/layout.json @@ -0,0 +1,8558 @@ +{ + "pdf_info": [ + { + "para_blocks": [ + { + "bbox": [ + 137, + 111, + 477, + 148 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 137, + 111, + 477, + 148 + ], + "spans": [ + { + "bbox": [ + 137, + 111, + 477, + 148 + ], + "type": "text", + "content": "Per-Gaussian Embedding-Based Deformation for Deformable 3D Gaussian Splitting" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 190, + 167, + 422, + 194 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 190, + 167, + 422, + 194 + ], + "spans": [ + { + "bbox": [ + 190, + 167, + 422, + 194 + ], + "type": "text", + "content": "Jeongmin Bae1*, Seoha Kim1*, Youngsik Yun1, Hahyun Lee2, Gun Bang2, and Youngjung Uh1" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 183, + 201, + 430, + 224 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 183, + 201, + 430, + 224 + ], + "spans": [ + { + "bbox": [ + 183, + 201, + 430, + 224 + ], + "type": "text", + "content": "1 Yonsei University, Seoul 03722, Korea {jaymin.bae, hailey07, bbangsik, yj.uh}@yonsei.ac.kr" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 141, + 224, + 471, + 246 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 141, + 224, + 471, + 246 + ], + "spans": [ + { + "bbox": [ + 141, + 224, + 471, + 246 + ], + "type": "inline_equation", + "content": "^{2}" + }, + { + "bbox": [ + 141, + 224, + 471, + 246 + ], + "type": "text", + "content": " Electronics and Telecommunications Research Institute, Daejeon 34129, Korea {hanilee, gbang}@etri.re.kr" + } + ] + } + ], + "index": 3 + }, + { + "type": "image", + "bbox": [ + 135, + 276, + 250, + 367 + ], + "blocks": [ + { + "bbox": [ + 135, + 276, + 250, + 367 + ], + "lines": [ + { + "bbox": [ + 135, + 276, + 250, + 367 + ], + "spans": [ + { + "bbox": [ + 135, + 276, + 250, + 367 + ], + "type": "image", + "image_path": "6fcce4eccafc285e23868dd70db2ef6658f862c19476727352665328c2371815.jpg" + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 130, + 376, + 482, + 421 + ], + "lines": [ + { + "bbox": [ + 130, + 376, + 482, + 421 + ], + "spans": [ + { + "bbox": [ + 130, + 376, + 482, + 421 + ], + "type": "text", + "content": "Fig. 1: Overview. (a) Existing deformable 3D Gaussian Splitting methods show blurry results in complex dynamic scenes, even with deformation fields using finer feature grids. (b) Our model solves the problem by employing per-Gaussian latent embeddings to predict deformations for each Gaussian and achieves a clearer results." + } + ] + } + ], + "index": 7, + "angle": 0, + "type": "image_caption" + } + ], + "index": 4 + }, + { + "type": "image", + "bbox": [ + 251, + 276, + 357, + 367 + ], + "blocks": [ + { + "bbox": [ + 251, + 276, + 357, + 367 + ], + "lines": [ + { + "bbox": [ + 251, + 276, + 357, + 367 + ], + "spans": [ + { + "bbox": [ + 251, + 276, + 357, + 367 + ], + "type": "image", + "image_path": "d936d00b55af85cd3fe51ff366a7edc6a05aa8fe639f9a7a1557ecaba53e90b0.jpg" + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "image_body" + } + ], + "index": 5 + }, + { + "type": "image", + "bbox": [ + 364, + 276, + 470, + 367 + ], + "blocks": [ + { + "bbox": [ + 364, + 276, + 470, + 367 + ], + "lines": [ + { + "bbox": [ + 364, + 276, + 470, + 367 + ], + "spans": [ + { + "bbox": [ + 364, + 276, + 470, + 367 + ], + "type": "image", + "image_path": "71408cd8124ebccd72cbaca806df1fa6aec9b530f4aeb22a3bdc4c309401d958.jpg" + } + ] + } + ], + "index": 6, + "angle": 0, + "type": "image_body" + } + ], + "index": 6 + }, + { + "bbox": [ + 159, + 462, + 453, + 605 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 159, + 462, + 453, + 605 + ], + "spans": [ + { + "bbox": [ + 159, + 462, + 453, + 605 + ], + "type": "text", + "content": "Abstract. As 3D Gaussian Splitting (3DGS) provides fast and high-quality novel view synthesis, it is a natural extension to deform a canonical 3DGS to multiple frames for representing a dynamic scene. However, previous works fail to accurately reconstruct complex dynamic scenes. We attribute the failure to the design of the deformation field, which is built as a coordinate-based function. This approach is problematic because 3DGS is a mixture of multiple fields centered at the Gaussians, not just a single coordinate-based framework. To resolve this problem, we define the deformation as a function of per-Gaussian embeddings and temporal embeddings. Moreover, we decompose deformations as coarse and fine deformations to model slow and fast movements, respectively. Also, we introduce a local smoothness regularization for per-Gaussian embedding to improve the details in dynamic regions." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 165, + 606, + 376, + 616 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 165, + 606, + 376, + 616 + ], + "spans": [ + { + "bbox": [ + 165, + 606, + 376, + 616 + ], + "type": "text", + "content": "Project page: https://jeongminb.github.io/e-d3dns/" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 159, + 627, + 453, + 649 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 159, + 627, + 453, + 649 + ], + "spans": [ + { + "bbox": [ + 159, + 627, + 453, + 649 + ], + "type": "text", + "content": "Keywords: Deformable gaussian splatting " + }, + { + "bbox": [ + 159, + 627, + 453, + 649 + ], + "type": "inline_equation", + "content": "\\cdot" + }, + { + "bbox": [ + 159, + 627, + 453, + 649 + ], + "type": "text", + "content": " 4D scene reconstruction " + }, + { + "bbox": [ + 159, + 627, + 453, + 649 + ], + "type": "inline_equation", + "content": "\\cdot" + }, + { + "bbox": [ + 159, + 627, + 453, + 649 + ], + "type": "text", + "content": " Novel view synthesis" + } + ] + } + ], + "index": 10 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 133, + 654, + 313, + 665 + ], + "type": "page_footnote", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 654, + 313, + 665 + ], + "spans": [ + { + "bbox": [ + 133, + 654, + 313, + 665 + ], + "type": "text", + "content": "* Authors contributed equally to this work." + } + ] + } + ], + "index": 11 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 0 + }, + { + "para_blocks": [ + { + "bbox": [ + 133, + 114, + 229, + 127 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 114, + 229, + 127 + ], + "spans": [ + { + "bbox": [ + 133, + 114, + 229, + 127 + ], + "type": "text", + "content": "1 Introduction" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 130, + 152, + 482, + 237 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 152, + 482, + 237 + ], + "spans": [ + { + "bbox": [ + 130, + 152, + 482, + 237 + ], + "type": "text", + "content": "Dynamic scene reconstruction from multi-view input videos is an important task in computer vision, as it can be extended to various applications and industries such as mixed reality, content production, etc. Neural Radiance Fields (NeRF) [19], which enable photorealistic novel view synthesis from multi-view inputs, can represent dynamic scenes by modeling the scene with an additional time input [12, 22]. However, typical NeRFs require querying multilayer perceptron (MLP) for hundreds of points per camera ray, which limits rendering speed." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 130, + 239, + 482, + 358 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 239, + 482, + 358 + ], + "spans": [ + { + "bbox": [ + 130, + 239, + 482, + 358 + ], + "type": "text", + "content": "On the other hand, the recently emerging 3D Gaussian Splatting (3DGS) [10] has the advantage of real-time rendering compared to NeRFs using a differentiable rasterizer for 3D Gaussian primitives. 3DGS directly optimizes the parameters of 3D Gaussians (position, opacity, anisotropic covariance, and spherical harmonics coefficients) and renders them via projection and " + }, + { + "bbox": [ + 130, + 239, + 482, + 358 + ], + "type": "inline_equation", + "content": "\\alpha" + }, + { + "bbox": [ + 130, + 239, + 482, + 358 + ], + "type": "text", + "content": "-blending. Since 3DGS has the characteristics of continuous volumetric radiance fields, some recent studies [3,8,14,29,31,32] represent dynamic scenes by defining a canonical 3DGS and deforming it to individual frames as deformable NeRFs [31] do. Specifically, they model the deformation as a function of 4D (x, y, z, t) coordinates with MLPs or grids to predict the change in the 3D Gaussian parameters." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 130, + 361, + 482, + 480 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 361, + 482, + 480 + ], + "spans": [ + { + "bbox": [ + 130, + 361, + 482, + 480 + ], + "type": "text", + "content": "However, since 3DGS is a mixTURE of multiple volumetric fields, it is not appropriate to model the deformation of Gaussian parameters with a single coordinate-based network to represent dynamic scenes. In addition, existing field-based approaches are constrained by the resolution of the grid which models the deformation field, the capacity of the model, or the frequencies of the input. As shown in Figure 1, existing study does not properly represent complex dynamic scenes, and even introducing an additional feature grid that is twice the maximum resolution has only a slight improvement in performance (See Appendix for more results). We alleviate this problem by introducing a novel dynamic representation to deform each Gaussian." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 130, + 483, + 482, + 603 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 483, + 482, + 603 + ], + "spans": [ + { + "bbox": [ + 130, + 483, + 482, + 603 + ], + "type": "text", + "content": "In this paper, we model the deformation of Gaussians at frames as 1) a function of a product space of per-Gaussian embeddings and temporal embeddings. We expect this rational design to bring quality improvement by precisely modeling different deformations of different Gaussians. Additionally, 2) We decompose temporal variations of the parameters into coarse and fine components, namely coarse-fine deformation. The coarse deformation represents large or slow movements in the scene, while fine deformation learns the fast or detailed movements that coarse deformation does not cover. Finally, we propose 3) a local smoothness regularization for per-Gaussian embedding to ensure the deformations of neighboring Gaussians are similar." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 130, + 605, + 482, + 666 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 605, + 482, + 666 + ], + "spans": [ + { + "bbox": [ + 130, + 605, + 482, + 666 + ], + "type": "text", + "content": "In our experiments, we observe that our per-Gaussian embeddings, coarse-fine deformation, and regularization improve the deformation quality. Our approach outperforms baselines in capturing fine details in dynamic regions and excels even under challenging camera settings. Additionally, our method also achieves fast rendering speed and relatively low capacity." + } + ] + } + ], + "index": 7 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "spans": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "type": "text", + "content": "2" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 166, + 91, + 217, + 100 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 166, + 91, + 217, + 100 + ], + "spans": [ + { + "bbox": [ + 166, + 91, + 217, + 100 + ], + "type": "text", + "content": "J. Bae et al." + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 1 + }, + { + "para_blocks": [ + { + "bbox": [ + 132, + 114, + 237, + 127 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 114, + 237, + 127 + ], + "spans": [ + { + "bbox": [ + 132, + 114, + 237, + 127 + ], + "type": "text", + "content": "2 Related Work" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 130, + 141, + 480, + 190 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 141, + 480, + 190 + ], + "spans": [ + { + "bbox": [ + 130, + 141, + 480, + 190 + ], + "type": "text", + "content": "In this section, we review methods for dynamic scene reconstruction that deform 3D canonical space and methods for reconstructing dynamic scenes utilizing dynamic 3D Gaussians. Afterward, we review methods that use embeddings and spatial relationships of Gaussians." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 130, + 208, + 482, + 353 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 208, + 482, + 353 + ], + "spans": [ + { + "bbox": [ + 130, + 208, + 482, + 353 + ], + "type": "text", + "content": "Deforming 3D Canonical Space D-NeRF [22] reconstructs dynamic scenes by deforming ray samples over time, using the deformation network that takes 3D coordinates and timestamps of the sample as inputs. Nerfies [20] and HyperNeRF [21] use per-frame trainable deformation codes instead of time conditions to deform the canonical space. Instead of deforming from the canonical frame to the entire frames, HyperReel [1] deforms the ray sample of the keyframe to represent the intermediate frame. 4DGaussians [29] and D3DGS [31] reconstruct the dynamic scene with a deformation network which inputs the center position of the canonical 3D Gaussians and timestamps. MoDGS [15] learns the mapping between canonical space and space at specific timestamp through invertible MLP. In contrast, we demonstrate a novel deformation representation as a function of a product space of per-Gaussian latent embeddings and temporal embeddings." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 130, + 371, + 482, + 465 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 371, + 482, + 465 + ], + "spans": [ + { + "bbox": [ + 130, + 371, + 482, + 465 + ], + "type": "text", + "content": "Dynamic 3D Gaussians To extend the fast rendering speed of 3D Gaussian Splitting [10] into dynamic scene reconstructions. 4D Gaussians [29] decodes features from multi-resolution HexPlanes [2] for temporal deformation of 3D Gaussians. While D3DGS [31] uses an implicit function that processes the time and location of the Gaussian. 4DGS [30] decomposes the 4D Gaussians into a time-conditioned 3D Gaussians and a marginal 1D Gaussians. STG [13] represents changes in 3D Gaussian over time through a temporal opacity and a polynomial function for each Gaussian." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 130, + 467, + 482, + 515 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 467, + 482, + 515 + ], + "spans": [ + { + "bbox": [ + 130, + 467, + 482, + 515 + ], + "type": "text", + "content": "Our method uses deformable 3D Gaussians as 4D Gaussians [29] and D3DGS [31] do, but does not necessitate the separated feature field to obtain the input feature of the deformation decoder. Our approach uses embeddings allocated to each Gaussian and a temporal embedding shared within a specific frame." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 130, + 533, + 482, + 665 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 533, + 482, + 665 + ], + "spans": [ + { + "bbox": [ + 130, + 533, + 482, + 665 + ], + "type": "text", + "content": "Latent Embedding on Novel View Synthesis Some studies incorporate latent embeddings to represent different states of the static and dynamic scene. NeRF-W [18] and Block-NeRF [25] employ per-image embeddings to capture different appearances of a scene, representing the scenes from in-the-wild image collections. DyNeRF and MixVoxels [12, 27] employ a temporal embedding for each frame to represent dynamic scenes. Nerfies [20] and HyperNeRF [21] incorporate both per-frame appearance and deformation embeddings. Sync-NeRF [11] introduces time offset to calibrate the misaligned temporal embeddings on dynamic scenes from unsynchronized videos. We introduce per-Gaussian latent embedding to encode the changes over time of each Gaussian and use temporal embeddings to represent different states in each frame of the scene." + } + ] + } + ], + "index": 7 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 405, + 91, + 447, + 100 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 405, + 91, + 447, + 100 + ], + "spans": [ + { + "bbox": [ + 405, + 91, + 447, + 100 + ], + "type": "text", + "content": "E-D3DGS" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 474, + 92, + 480, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 474, + 92, + 480, + 100 + ], + "spans": [ + { + "bbox": [ + 474, + 92, + 480, + 100 + ], + "type": "text", + "content": "3" + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 2 + }, + { + "para_blocks": [ + { + "bbox": [ + 130, + 116, + 482, + 236 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 116, + 482, + 236 + ], + "spans": [ + { + "bbox": [ + 130, + 116, + 482, + 236 + ], + "type": "text", + "content": "Considering Spatial Relationships of Gaussians Scaffold-GS [16] reconstruct 3D scenes by synthesizing Gaussians from anchors, utilizing that the neighboring Gaussians have similar properties. SAGS [26] creates a graph based on k-nearest neighbors (KNN) so that each Gaussian is optimized while considering its neighboring Gaussians. In dynamic scene reconstruction, SC-GS [9] and GaussianPrediction [34] deform Gaussians by combining the deformations of key point Gaussians. Dynamic 3D Gaussians [17] utilizes regularization to encourage that Gaussians and their neighboring Gaussians deform with local rigidity. Similarly, we propose a local smoothness regularization that encourages neighboring Gaussians to have similar embeddings, resulting in similar deformations." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 132, + 255, + 202, + 268 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 255, + 202, + 268 + ], + "spans": [ + { + "bbox": [ + 132, + 255, + 202, + 268 + ], + "type": "text", + "content": "3 Method" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 130, + 281, + 482, + 354 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 281, + 482, + 354 + ], + "spans": [ + { + "bbox": [ + 130, + 281, + 482, + 354 + ], + "type": "text", + "content": "In this section, we first provide a brief overview of 3D Gaussian Splatting (Section 3.1). Next, we introduce our overall framework, embedding-based deformation for Gaussians (Section 3.2) and coarse-fine deformation scheme consisting of coarse and fine deformation functions (Section 3.3). Finally, we present a local smoothness regularization for per-Gaussian embeddings to achieve better details on dynamic regions (Section 3.4)." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 132, + 371, + 340, + 384 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 371, + 340, + 384 + ], + "spans": [ + { + "bbox": [ + 132, + 371, + 340, + 384 + ], + "type": "text", + "content": "3.1 Preliminary: 3D Gaussian Splitting" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 130, + 392, + 481, + 452 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 392, + 481, + 452 + ], + "spans": [ + { + "bbox": [ + 130, + 392, + 481, + 452 + ], + "type": "text", + "content": "3D Gaussian splattering [10] optimizes a set of anisotropic 3D Gaussians through differentiable tile rasterization to reconstruct a static 3D scene. By its efficient rasterization, the optimized model enables real-time rendering of high-quality images. Each 3D Gaussian kernel " + }, + { + "bbox": [ + 130, + 392, + 481, + 452 + ], + "type": "inline_equation", + "content": "G_{i}(x)" + }, + { + "bbox": [ + 130, + 392, + 481, + 452 + ], + "type": "text", + "content": " at the point " + }, + { + "bbox": [ + 130, + 392, + 481, + 452 + ], + "type": "inline_equation", + "content": "x" + }, + { + "bbox": [ + 130, + 392, + 481, + 452 + ], + "type": "text", + "content": " consists with position " + }, + { + "bbox": [ + 130, + 392, + 481, + 452 + ], + "type": "inline_equation", + "content": "\\mathbf{x}_i" + }, + { + "bbox": [ + 130, + 392, + 481, + 452 + ], + "type": "text", + "content": ", rotation " + }, + { + "bbox": [ + 130, + 392, + 481, + 452 + ], + "type": "inline_equation", + "content": "R_i" + }, + { + "bbox": [ + 130, + 392, + 481, + 452 + ], + "type": "text", + "content": ", and scale " + }, + { + "bbox": [ + 130, + 392, + 481, + 452 + ], + "type": "inline_equation", + "content": "S_i" + }, + { + "bbox": [ + 130, + 392, + 481, + 452 + ], + "type": "text", + "content": ":" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 176, + 460, + 481, + 476 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 176, + 460, + 481, + 476 + ], + "spans": [ + { + "bbox": [ + 176, + 460, + 481, + 476 + ], + "type": "interline_equation", + "content": "G _ {i} (x) = e ^ {- \\frac {1}{2} (x - \\mathbf {x} _ {i}) ^ {T} \\Sigma_ {i} ^ {- 1} (x - \\mathbf {x} _ {i})}, \\quad \\text {w h e r e} \\quad \\Sigma_ {i} = R _ {i} S _ {i} S _ {i} ^ {T} R _ {i} ^ {T}. \\tag {1}", + "image_path": "5beb628896b9c5b53c663de707fc8ac21a9507fb1716426a8cb9740f4b7e7431.jpg" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 130, + 486, + 481, + 522 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 486, + 481, + 522 + ], + "spans": [ + { + "bbox": [ + 130, + 486, + 481, + 522 + ], + "type": "text", + "content": "To projecting 3D Gaussians to 2D for rendering, covariance matrix " + }, + { + "bbox": [ + 130, + 486, + 481, + 522 + ], + "type": "inline_equation", + "content": "\\Sigma^{\\prime}" + }, + { + "bbox": [ + 130, + 486, + 481, + 522 + ], + "type": "text", + "content": " are calculated by viewing transform " + }, + { + "bbox": [ + 130, + 486, + 481, + 522 + ], + "type": "inline_equation", + "content": "W" + }, + { + "bbox": [ + 130, + 486, + 481, + 522 + ], + "type": "text", + "content": " and the Jacobian " + }, + { + "bbox": [ + 130, + 486, + 481, + 522 + ], + "type": "inline_equation", + "content": "J" + }, + { + "bbox": [ + 130, + 486, + 481, + 522 + ], + "type": "text", + "content": " of the affine approximation of the projective transformation [35] as follows:" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 266, + 531, + 480, + 544 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 266, + 531, + 480, + 544 + ], + "spans": [ + { + "bbox": [ + 266, + 531, + 480, + 544 + ], + "type": "interline_equation", + "content": "\\Sigma^ {\\prime} = J W \\Sigma W ^ {T} J ^ {T}. \\tag {2}", + "image_path": "ab68f87dee9bb141b9e499db1c45959af88ae016b94c38ca3bf9db0a71d630d1.jpg" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 130, + 554, + 481, + 590 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 554, + 481, + 590 + ], + "spans": [ + { + "bbox": [ + 130, + 554, + 481, + 590 + ], + "type": "text", + "content": "Blending " + }, + { + "bbox": [ + 130, + 554, + 481, + 590 + ], + "type": "inline_equation", + "content": "\\mathcal{N}" + }, + { + "bbox": [ + 130, + 554, + 481, + 590 + ], + "type": "text", + "content": " depth-ordered projected points that overlap the pixel, the Gaussian kernel " + }, + { + "bbox": [ + 130, + 554, + 481, + 590 + ], + "type": "inline_equation", + "content": "G_{i}(x)" + }, + { + "bbox": [ + 130, + 554, + 481, + 590 + ], + "type": "text", + "content": " is multiplied by the opacity of the Gaussian " + }, + { + "bbox": [ + 130, + 554, + 481, + 590 + ], + "type": "inline_equation", + "content": "\\sigma_{i}" + }, + { + "bbox": [ + 130, + 554, + 481, + 590 + ], + "type": "text", + "content": " and calculates the pixel color " + }, + { + "bbox": [ + 130, + 554, + 481, + 590 + ], + "type": "inline_equation", + "content": "C" + }, + { + "bbox": [ + 130, + 554, + 481, + 590 + ], + "type": "text", + "content": " with the color of the Gaussian " + }, + { + "bbox": [ + 130, + 554, + 481, + 590 + ], + "type": "inline_equation", + "content": "c_{i}" + }, + { + "bbox": [ + 130, + 554, + 481, + 590 + ], + "type": "text", + "content": ":" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 194, + 600, + 481, + 633 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 194, + 600, + 481, + 633 + ], + "spans": [ + { + "bbox": [ + 194, + 600, + 481, + 633 + ], + "type": "interline_equation", + "content": "C = \\sum_ {i \\in \\mathcal {N}} c _ {i} \\alpha_ {i} \\prod_ {j = 1} ^ {i - 1} (1 - \\alpha_ {j}), \\quad \\text {w h e r e} \\quad \\alpha_ {i} = \\sigma_ {i} G _ {i} (x). \\tag {3}", + "image_path": "cfe44c0fe76e6353fbf103b7aa23e3861e17ab3c192aadef9939bd79f663eb01.jpg" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 130, + 641, + 481, + 665 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 641, + 481, + 665 + ], + "spans": [ + { + "bbox": [ + 130, + 641, + 481, + 665 + ], + "type": "text", + "content": "The color of Gaussian " + }, + { + "bbox": [ + 130, + 641, + 481, + 665 + ], + "type": "inline_equation", + "content": "c_{i}" + }, + { + "bbox": [ + 130, + 641, + 481, + 665 + ], + "type": "text", + "content": " is determined using the SH coefficient with view direction." + } + ] + } + ], + "index": 12 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "spans": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "type": "text", + "content": "4" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 166, + 91, + 217, + 100 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 166, + 91, + 217, + 100 + ], + "spans": [ + { + "bbox": [ + 166, + 91, + 217, + 100 + ], + "type": "text", + "content": "J. Bae et al." + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 3 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 143, + 116, + 232, + 167 + ], + "blocks": [ + { + "bbox": [ + 143, + 116, + 232, + 167 + ], + "lines": [ + { + "bbox": [ + 143, + 116, + 232, + 167 + ], + "spans": [ + { + "bbox": [ + 143, + 116, + 232, + 167 + ], + "type": "image", + "image_path": "5a6ada8e19479be6d76f7158e58487be4d65b99cfd38a791de1ac921bcd6b476.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_body" + } + ], + "index": 2 + }, + { + "type": "image", + "bbox": [ + 146, + 171, + 236, + 230 + ], + "blocks": [ + { + "bbox": [ + 146, + 171, + 236, + 230 + ], + "lines": [ + { + "bbox": [ + 146, + 171, + 236, + 230 + ], + "spans": [ + { + "bbox": [ + 146, + 171, + 236, + 230 + ], + "type": "image", + "image_path": "b92e21303c10ae6fcf13d58c84f42e4fa8bbf5dfcf03143fce17f66da01e6ef4.jpg" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 145, + 233, + 238, + 242 + ], + "lines": [ + { + "bbox": [ + 145, + 233, + 238, + 242 + ], + "spans": [ + { + "bbox": [ + 145, + 233, + 238, + 242 + ], + "type": "text", + "content": "(a) Embeddings for deformation" + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "image_caption" + } + ], + "index": 3 + }, + { + "type": "image", + "bbox": [ + 248, + 115, + 366, + 229 + ], + "blocks": [ + { + "bbox": [ + 248, + 115, + 366, + 229 + ], + "lines": [ + { + "bbox": [ + 248, + 115, + 366, + 229 + ], + "spans": [ + { + "bbox": [ + 248, + 115, + 366, + 229 + ], + "type": "image", + "image_path": "0732f9a2f4bbb5c92e44341a0fd45613b07fcc21609fc5458e11d5f0abd3e5cd.jpg" + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 264, + 233, + 350, + 241 + ], + "lines": [ + { + "bbox": [ + 264, + 233, + 350, + 241 + ], + "spans": [ + { + "bbox": [ + 264, + 233, + 350, + 241 + ], + "type": "text", + "content": "(b) Per-Gaussian deformation" + } + ] + } + ], + "index": 6, + "angle": 0, + "type": "image_caption" + } + ], + "index": 5 + }, + { + "type": "image", + "bbox": [ + 381, + 118, + 462, + 229 + ], + "blocks": [ + { + "bbox": [ + 381, + 118, + 462, + 229 + ], + "lines": [ + { + "bbox": [ + 381, + 118, + 462, + 229 + ], + "spans": [ + { + "bbox": [ + 381, + 118, + 462, + 229 + ], + "type": "image", + "image_path": "afd61294da1e2007c76213a87be27107281fbaa0c8329bae0a3df65ca12b0be2.jpg" + } + ] + } + ], + "index": 7, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 371, + 233, + 474, + 242 + ], + "lines": [ + { + "bbox": [ + 371, + 233, + 474, + 242 + ], + "spans": [ + { + "bbox": [ + 371, + 233, + 474, + 242 + ], + "type": "text", + "content": "(c) Local smoothness regularization" + } + ] + } + ], + "index": 8, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 131, + 252, + 482, + 350 + ], + "lines": [ + { + "bbox": [ + 131, + 252, + 482, + 350 + ], + "spans": [ + { + "bbox": [ + 131, + 252, + 482, + 350 + ], + "type": "text", + "content": "Fig. 2: Framework. Existing coordinate-based network methods struggle to represent complex dynamic scenes. To this end, we define per-Gaussian deformation. (a) Firstly, we assign a latent embedding for each Gaussian. Additionally, we introduce coarse and fine temporal embeddings to represent the slow and fast state of the dynamic scene. (b) By employing two decoders that take per-Gaussian latent embeddings along with coarse and fine temporal embeddings as input, we estimate slow or large changes and fast or detailed changes to model the final deformation, respectively. (c) Finally, we introduce a local smoothness regularization so that the embeddings of neighboring Gaussians are similar." + } + ] + } + ], + "index": 9, + "angle": 0, + "type": "image_caption" + } + ], + "index": 7 + }, + { + "bbox": [ + 132, + 372, + 389, + 384 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 372, + 389, + 384 + ], + "spans": [ + { + "bbox": [ + 132, + 372, + 389, + 384 + ], + "type": "text", + "content": "3.2 Embedding-Based Deformation for Gaussians" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 130, + 389, + 482, + 460 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 389, + 482, + 460 + ], + "spans": [ + { + "bbox": [ + 130, + 389, + 482, + 460 + ], + "type": "text", + "content": "Deformable NeRFs consist of a deformation field that predicts displacement " + }, + { + "bbox": [ + 130, + 389, + 482, + 460 + ], + "type": "inline_equation", + "content": "\\Delta \\mathbf{x}" + }, + { + "bbox": [ + 130, + 389, + 482, + 460 + ], + "type": "text", + "content": " for a given coordinate " + }, + { + "bbox": [ + 130, + 389, + 482, + 460 + ], + "type": "inline_equation", + "content": "\\mathbf{x}" + }, + { + "bbox": [ + 130, + 389, + 482, + 460 + ], + "type": "text", + "content": " from the canonical space to each target frame, and a radiance field that maps color and density from a given coordinate in the canonical space " + }, + { + "bbox": [ + 130, + 389, + 482, + 460 + ], + "type": "inline_equation", + "content": "(\\mathbf{x} + \\Delta \\mathbf{x})" + }, + { + "bbox": [ + 130, + 389, + 482, + 460 + ], + "type": "text", + "content": ". Existing deformable Gaussian methods employ the same approach for predicting the deformation of Gaussians, i.e., utilizing a deformation field based on coordinates." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 130, + 461, + 482, + 533 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 461, + 482, + 533 + ], + "spans": [ + { + "bbox": [ + 130, + 461, + 482, + 533 + ], + "type": "text", + "content": "Unlike previous methods, we start from the design of 3DGS: the 3D scene is represented as a mixture of Gaussians that have individual radiance fields. Accordingly, the deformation should be defined for each Gaussian. Based on this intuition, we introduce a function " + }, + { + "bbox": [ + 130, + 461, + 482, + 533 + ], + "type": "inline_equation", + "content": "\\mathcal{F}_{\\theta}" + }, + { + "bbox": [ + 130, + 461, + 482, + 533 + ], + "type": "text", + "content": " that produces deformation from learnable embeddings " + }, + { + "bbox": [ + 130, + 461, + 482, + 533 + ], + "type": "inline_equation", + "content": "\\mathbf{z}_{\\mathrm{g}} \\in \\mathbb{R}^{32}" + }, + { + "bbox": [ + 130, + 461, + 482, + 533 + ], + "type": "text", + "content": " belonging to individual Gaussians (Figure 2a), and typical temporal embeddings " + }, + { + "bbox": [ + 130, + 461, + 482, + 533 + ], + "type": "inline_equation", + "content": "\\mathbf{z}_t \\in \\mathbb{R}^{256}" + }, + { + "bbox": [ + 130, + 461, + 482, + 533 + ], + "type": "text", + "content": " for different frames:" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 222, + 539, + 481, + 553 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 222, + 539, + 481, + 553 + ], + "spans": [ + { + "bbox": [ + 222, + 539, + 481, + 553 + ], + "type": "interline_equation", + "content": "\\mathcal {F} _ {\\theta}: \\left(\\mathbf {z} _ {\\mathrm {g}}, \\mathbf {z} _ {t}\\right)\\rightarrow \\left(\\Delta \\mathbf {x}, \\Delta \\mathbf {r}, \\Delta \\mathbf {s}, \\Delta \\sigma , \\Delta Y\\right), \\tag {4}", + "image_path": "0e60c99a05b6b27e508ff3d8a826f4b0b2732f20512b6eb92a8a106724eddd2f.jpg" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 130, + 558, + 482, + 617 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 558, + 482, + 617 + ], + "spans": [ + { + "bbox": [ + 130, + 558, + 482, + 617 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 130, + 558, + 482, + 617 + ], + "type": "inline_equation", + "content": "\\mathbf{r}" + }, + { + "bbox": [ + 130, + 558, + 482, + 617 + ], + "type": "text", + "content": " is a rotation quaternion, " + }, + { + "bbox": [ + 130, + 558, + 482, + 617 + ], + "type": "inline_equation", + "content": "\\mathbf{s}" + }, + { + "bbox": [ + 130, + 558, + 482, + 617 + ], + "type": "text", + "content": " is a vector for scaling, " + }, + { + "bbox": [ + 130, + 558, + 482, + 617 + ], + "type": "inline_equation", + "content": "\\sigma" + }, + { + "bbox": [ + 130, + 558, + 482, + 617 + ], + "type": "text", + "content": " is an opacity, and " + }, + { + "bbox": [ + 130, + 558, + 482, + 617 + ], + "type": "inline_equation", + "content": "Y" + }, + { + "bbox": [ + 130, + 558, + 482, + 617 + ], + "type": "text", + "content": " is SH coefficients for modeling view-dependent color. We implement " + }, + { + "bbox": [ + 130, + 558, + 482, + 617 + ], + "type": "inline_equation", + "content": "\\mathcal{F}_{\\theta}" + }, + { + "bbox": [ + 130, + 558, + 482, + 617 + ], + "type": "text", + "content": " as a shallow multi-layer perceptron (MLP) followed by an MLP head for each parameter. As a result, the Gaussian parameters at frame " + }, + { + "bbox": [ + 130, + 558, + 482, + 617 + ], + "type": "inline_equation", + "content": "t" + }, + { + "bbox": [ + 130, + 558, + 482, + 617 + ], + "type": "text", + "content": " are determined by adding " + }, + { + "bbox": [ + 130, + 558, + 482, + 617 + ], + "type": "inline_equation", + "content": "\\mathcal{F}_{\\theta}(\\mathbf{z}_{\\mathrm{g}},\\mathbf{z}_t)" + }, + { + "bbox": [ + 130, + 558, + 482, + 617 + ], + "type": "text", + "content": " to the canonical Gaussian parameters (Figure 2c)." + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 130, + 617, + 482, + 665 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 617, + 482, + 665 + ], + "spans": [ + { + "bbox": [ + 130, + 617, + 482, + 665 + ], + "type": "text", + "content": "We jointly optimize the per-Gaussian embeddings " + }, + { + "bbox": [ + 130, + 617, + 482, + 665 + ], + "type": "inline_equation", + "content": "\\mathbf{z}_{\\mathrm{g}}" + }, + { + "bbox": [ + 130, + 617, + 482, + 665 + ], + "type": "text", + "content": ", the deformation function " + }, + { + "bbox": [ + 130, + 617, + 482, + 665 + ], + "type": "inline_equation", + "content": "\\mathcal{F}_{\\theta}" + }, + { + "bbox": [ + 130, + 617, + 482, + 665 + ], + "type": "text", + "content": ", and the canonical Gaussian parameters to minimize the rendering loss. We use the L1 and periodic DSSIM as the rendering loss between the rendered image and the ground truth image." + } + ] + } + ], + "index": 15 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 405, + 91, + 447, + 100 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 405, + 91, + 447, + 100 + ], + "spans": [ + { + "bbox": [ + 405, + 91, + 447, + 100 + ], + "type": "text", + "content": "E-D3DGS" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 474, + 92, + 480, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 474, + 92, + 480, + 100 + ], + "spans": [ + { + "bbox": [ + 474, + 92, + 480, + 100 + ], + "type": "text", + "content": "5" + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 4 + }, + { + "para_blocks": [ + { + "bbox": [ + 132, + 116, + 288, + 126 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 116, + 288, + 126 + ], + "spans": [ + { + "bbox": [ + 132, + 116, + 288, + 126 + ], + "type": "text", + "content": "3.3 Coarse-Fine Deformation" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 130, + 137, + 479, + 183 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 137, + 479, + 183 + ], + "spans": [ + { + "bbox": [ + 130, + 137, + 479, + 183 + ], + "type": "text", + "content": "Different parts of a scene may have coarse and fine motions [5]. E.g., a hand swiftly stirs a pan (fine) while a body slowly moves from left to right (coarse). Based on this intuition, we introduce a coarse-fine deformation that produces a summation of coarse and fine deformations." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 130, + 185, + 479, + 220 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 185, + 479, + 220 + ], + "spans": [ + { + "bbox": [ + 130, + 185, + 479, + 220 + ], + "type": "text", + "content": "Coarse-fine deformation consists of two functions with the same architecture: one for coarse and one for fine deformation (Figure 2c). The functions receive different temporal embeddings as follows:" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 131, + 220, + 480, + 293 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 131, + 220, + 480, + 293 + ], + "spans": [ + { + "bbox": [ + 131, + 220, + 480, + 293 + ], + "type": "text", + "content": "Following typical temporal embeddings, we start from a 1D feature grid " + }, + { + "bbox": [ + 131, + 220, + 480, + 293 + ], + "type": "inline_equation", + "content": "Z \\in \\mathbb{R}^{N \\times 256}" + }, + { + "bbox": [ + 131, + 220, + 480, + 293 + ], + "type": "text", + "content": " for " + }, + { + "bbox": [ + 131, + 220, + 480, + 293 + ], + "type": "inline_equation", + "content": "N" + }, + { + "bbox": [ + 131, + 220, + 480, + 293 + ], + "type": "text", + "content": " frames and use an embedding " + }, + { + "bbox": [ + 131, + 220, + 480, + 293 + ], + "type": "inline_equation", + "content": "\\mathbf{z}_t^{\\mathrm{f}} = \\mathrm{interp}(Z,t)" + }, + { + "bbox": [ + 131, + 220, + 480, + 293 + ], + "type": "text", + "content": " for fine deformation. For coarse deformation, we linearly downsample " + }, + { + "bbox": [ + 131, + 220, + 480, + 293 + ], + "type": "inline_equation", + "content": "Z" + }, + { + "bbox": [ + 131, + 220, + 480, + 293 + ], + "type": "text", + "content": " by a factor of 5 to remove high-frequencies responsible for fast and detailed deformation. Then we compute " + }, + { + "bbox": [ + 131, + 220, + 480, + 293 + ], + "type": "inline_equation", + "content": "\\mathbf{z}_t^{\\mathrm{c}}" + }, + { + "bbox": [ + 131, + 220, + 480, + 293 + ], + "type": "text", + "content": " as a linear interpolation of embeddings at enclosing grid points (Figure 2b)." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 130, + 293, + 480, + 352 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 293, + 480, + 352 + ], + "spans": [ + { + "bbox": [ + 130, + 293, + 480, + 352 + ], + "type": "text", + "content": "As a result, coarse deformation " + }, + { + "bbox": [ + 130, + 293, + 480, + 352 + ], + "type": "inline_equation", + "content": "\\mathcal{F}_{\\theta_{\\mathrm{c}}}(\\mathbf{z}_{\\mathrm{g}},\\mathbf{z}_{t}^{c})" + }, + { + "bbox": [ + 130, + 293, + 480, + 352 + ], + "type": "text", + "content": " is responsible for representing large or slow movements in the scene, while fine deformation " + }, + { + "bbox": [ + 130, + 293, + 480, + 352 + ], + "type": "inline_equation", + "content": "\\mathcal{F}_{\\theta_{\\mathrm{f}}}(\\mathbf{z}_{\\mathrm{g}},\\mathbf{z}_{t}^{\\mathrm{f}})" + }, + { + "bbox": [ + 130, + 293, + 480, + 352 + ], + "type": "text", + "content": " learns the fast or detailed movements that coarse deformation does not cover. This improves the deformation quality. Refer to the Ablation study section for more details." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 132, + 371, + 327, + 383 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 371, + 327, + 383 + ], + "spans": [ + { + "bbox": [ + 132, + 371, + 327, + 383 + ], + "type": "text", + "content": "3.4 Local Smoothness Regularization" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 130, + 392, + 479, + 440 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 392, + 479, + 440 + ], + "spans": [ + { + "bbox": [ + 130, + 392, + 479, + 440 + ], + "type": "text", + "content": "Neighboring Gaussians constructing dynamic objects tend to exhibit locally similar deformation. Inspired by [17], we introduce a local smoothness regularization for per-Gaussian embedding " + }, + { + "bbox": [ + 130, + 392, + 479, + 440 + ], + "type": "inline_equation", + "content": "\\mathbf{z}_{\\mathrm{g}}" + }, + { + "bbox": [ + 130, + 392, + 479, + 440 + ], + "type": "text", + "content": " (Figure 2d) to encourage similar deformations between nearby Gaussians " + }, + { + "bbox": [ + 130, + 392, + 479, + 440 + ], + "type": "inline_equation", + "content": "i" + }, + { + "bbox": [ + 130, + 392, + 479, + 440 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 130, + 392, + 479, + 440 + ], + "type": "inline_equation", + "content": "j" + }, + { + "bbox": [ + 130, + 392, + 479, + 440 + ], + "type": "text", + "content": ":" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 203, + 448, + 410, + 477 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 203, + 448, + 410, + 477 + ], + "spans": [ + { + "bbox": [ + 203, + 448, + 410, + 477 + ], + "type": "interline_equation", + "content": "\\mathcal {L} ^ {\\mathrm {e m b} - \\mathrm {r e g}} = \\frac {1}{k | \\mathcal {S} |} \\sum_ {i \\in \\mathcal {S}} \\sum_ {j \\in \\mathrm {K N N} _ {i; k}} \\left(w _ {i, j} \\| \\mathbf {z} _ {\\mathrm {g} i} - \\mathbf {z} _ {\\mathrm {g} j} \\| _ {2}\\right),", + "image_path": "fe97acd29c9a1d44d4b1fda160818fd2f6eb04ddc4a1c57551c43b46c38ef5ad.jpg" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 130, + 487, + 480, + 523 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 487, + 480, + 523 + ], + "spans": [ + { + "bbox": [ + 130, + 487, + 480, + 523 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 130, + 487, + 480, + 523 + ], + "type": "inline_equation", + "content": "w_{i,j} = \\exp (-\\lambda_w\\| \\mu_j - \\mu_i\\| _2^2)" + }, + { + "bbox": [ + 130, + 487, + 480, + 523 + ], + "type": "text", + "content": " is the weighting factor and " + }, + { + "bbox": [ + 130, + 487, + 480, + 523 + ], + "type": "inline_equation", + "content": "\\mu" + }, + { + "bbox": [ + 130, + 487, + 480, + 523 + ], + "type": "text", + "content": " is the Gaussian center. We set " + }, + { + "bbox": [ + 130, + 487, + 480, + 523 + ], + "type": "inline_equation", + "content": "\\lambda_w" + }, + { + "bbox": [ + 130, + 487, + 480, + 523 + ], + "type": "text", + "content": " to 2000 and " + }, + { + "bbox": [ + 130, + 487, + 480, + 523 + ], + "type": "inline_equation", + "content": "k" + }, + { + "bbox": [ + 130, + 487, + 480, + 523 + ], + "type": "text", + "content": " to 20 following [17]. To reduce the computational cost, we obtain sets of k-nearest-neighbors only when the densification occurs." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 130, + 525, + 479, + 572 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 525, + 479, + 572 + ], + "spans": [ + { + "bbox": [ + 130, + 525, + 479, + 572 + ], + "type": "text", + "content": "Note that unlike previous approaches that directly constrain physical properties such as rigidity or rotation, We implicitly induce locally similar deformation by ensuring that per-Gaussian embeddings are locally smooth. Our regularization allows better capture of textures and details of dynamic objects." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 132, + 591, + 224, + 605 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 591, + 224, + 605 + ], + "spans": [ + { + "bbox": [ + 132, + 591, + 224, + 605 + ], + "type": "text", + "content": "4 Experiment" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 130, + 617, + 480, + 665 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 617, + 480, + 665 + ], + "spans": [ + { + "bbox": [ + 130, + 617, + 480, + 665 + ], + "type": "text", + "content": "In this section, we first describe the criterion for selection of baselines, and evaluation metrics. We then demonstrate the effectiveness of our method through comparisons with various baselines and datasets (Section 4.1-4.2). Finally, we conduct analysis and ablations of our method (Section 4.3)." + } + ] + } + ], + "index": 13 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "spans": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "type": "text", + "content": "6" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 166, + 91, + 217, + 100 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 166, + 91, + 217, + 100 + ], + "spans": [ + { + "bbox": [ + 166, + 91, + 217, + 100 + ], + "type": "text", + "content": "J. Bae et al." + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 5 + }, + { + "para_blocks": [ + { + "type": "table", + "bbox": [ + 168, + 174, + 442, + 270 + ], + "blocks": [ + { + "bbox": [ + 132, + 114, + 482, + 169 + ], + "lines": [ + { + "bbox": [ + 132, + 114, + 482, + 169 + ], + "spans": [ + { + "bbox": [ + 132, + 114, + 482, + 169 + ], + "type": "text", + "content": "Table 1: Average performance in the test view on Neural 3D Video Dataset The computational cost was measured based on flame_salmon_1 on the A6000. " + }, + { + "bbox": [ + 132, + 114, + 482, + 169 + ], + "type": "inline_equation", + "content": "{}^{1}" + }, + { + "bbox": [ + 132, + 114, + 482, + 169 + ], + "type": "text", + "content": " flame salmon scene includes only the first segment, comprising 300 frames. " + }, + { + "bbox": [ + 132, + 114, + 482, + 169 + ], + "type": "inline_equation", + "content": "{}^{2}" + }, + { + "bbox": [ + 132, + 114, + 482, + 169 + ], + "type": "text", + "content": " reported time and DyNeRF is trained on 8 GPUs and tested only on flame salmon. " + }, + { + "bbox": [ + 132, + 114, + 482, + 169 + ], + "type": "inline_equation", + "content": "{}^{3}" + }, + { + "bbox": [ + 132, + 114, + 482, + 169 + ], + "type": "text", + "content": " trained with 90 frames. " + }, + { + "bbox": [ + 132, + 114, + 482, + 169 + ], + "type": "inline_equation", + "content": "{}^{4}" + }, + { + "bbox": [ + 132, + 114, + 482, + 169 + ], + "type": "text", + "content": " trained with 50 frames." + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 168, + 174, + 442, + 270 + ], + "lines": [ + { + "bbox": [ + 168, + 174, + 442, + 270 + ], + "spans": [ + { + "bbox": [ + 168, + 174, + 442, + 270 + ], + "type": "table", + "html": "
modelmetriccomputational cost
PSNR↑SSIM↑LPIPS↓Training time↓FPS↑Model size↓
DyNeRF12 [12]29.58-0.0831344 hours0.0156 MB
NeRFPlayer13 [24]30.69-0.1116 hours0.051654 MB
Mix Voxels [27]30.300.9180.1271 hours 40 mins0.93512 MB
K-Planes [6]30.860.9390.0961 hours 40 mins0.13309 MB
HyperReel4 [1]30.370.9210.1069 hours 20 mins1.01362 MB
4DGS [30]31.190.9400.0519 hours 30 mins33.78700 MB
4DGaussians [29]30.710.9350.05650 mins51.959 MB
Ours31.310.9450.0371 hours 52 mins74.535 MB
", + "image_path": "89503dc1b1a2b278ce5ed3c00e2d39577c28e5acca30f018646205806c32ccd0.jpg" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "table_body" + } + ], + "index": 3 + }, + { + "bbox": [ + 130, + 292, + 482, + 423 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 292, + 482, + 423 + ], + "spans": [ + { + "bbox": [ + 130, + 292, + 482, + 423 + ], + "type": "text", + "content": "Baselines We choose the start-of-the-art method as a baseline in each dataset. We compared against DyNeRF, NeRFPlayer, MixVoxels, K-Planes, HyperReel, Nerfies, HyperNeRF, and TiNeuVox on the NeRF baseline. In detail, we use the version of NeRFPlayer TensoRF VM, HyperNeRF DF, Mixvoxels-L, K-Planes hybrid. We compared with 4DGaussians, 4DGS, and D3DGS based on the Gaussian baseline. Meanwhile, we have not included STG in our comparison due to its requirement for per-frame Structure from Motion (SfM) points, which makes conducting a fair comparison challenging. Also, STG is not a deformable 3D Gaussian approach. We followed the official code and configuration, except increasing the training iterations for the Technicolor dataset to 1.5 times that of the 4DGaussians, in comparison to the Neural 3D Video dataset." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 130, + 439, + 482, + 523 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 439, + 482, + 523 + ], + "spans": [ + { + "bbox": [ + 130, + 439, + 482, + 523 + ], + "type": "text", + "content": "Metrics We report the quality of rendered images using PSNR, SSIM, and LPIPS. Peak Signal-to-Noise Ratio (PSNR) quantifies pixel color error between the rendered video and the ground truth. We utilize SSIM [28] to account for the perceived similarity of the rendered image. Additionally, we measure higher-level perceptual similarity using Learned Perceptual Image Patch Similarity (LPIPS) [33] with an AlexNet Backbone. Higher PSNR and SSIM values and lower LPIPS values indicate better visual quality." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 132, + 539, + 328, + 552 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 539, + 328, + 552 + ], + "spans": [ + { + "bbox": [ + 132, + 539, + 328, + 552 + ], + "type": "text", + "content": "4.1 Effectiveness on Dynamic Region" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 130, + 557, + 482, + 665 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 557, + 482, + 665 + ], + "spans": [ + { + "bbox": [ + 130, + 557, + 482, + 665 + ], + "type": "text", + "content": "Neural 3D Video Dataset [12] includes 20 multi-view videos, with each scene consisting of either 300 frames, except for the flame_salmon scene, which comprises 1200 frames. These scenes encompass a relatively long duration and various movements, with some featuring multiple objects in motion. We utilized Neural 3D Video dataset to observe the capability to capture dynamic areas. Total six scenes (coffee_martini, cook_s spinach, cut_ roasted_beef, flame_salmon, flame_ steak, sear_ steak) are evaluated in Figure 3 and Table 1. The flame_salmon scene divided into four segments, each containing 300 frames." + } + ] + } + ], + "index": 7 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 405, + 91, + 447, + 100 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 405, + 91, + 447, + 100 + ], + "spans": [ + { + "bbox": [ + 405, + 91, + 447, + 100 + ], + "type": "text", + "content": "E-D3DGS" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 474, + 92, + 480, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 474, + 92, + 480, + 100 + ], + "spans": [ + { + "bbox": [ + 474, + 92, + 480, + 100 + ], + "type": "text", + "content": "7" + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 6 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 138, + 118, + 478, + 392 + ], + "blocks": [ + { + "bbox": [ + 138, + 118, + 478, + 392 + ], + "lines": [ + { + "bbox": [ + 138, + 118, + 478, + 392 + ], + "spans": [ + { + "bbox": [ + 138, + 118, + 478, + 392 + ], + "type": "image", + "image_path": "a175a53e84dca37bd016a27fc78420aa86657301a2861a78a0e1deb900b85278.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 154, + 403, + 459, + 415 + ], + "lines": [ + { + "bbox": [ + 154, + 403, + 459, + 415 + ], + "spans": [ + { + "bbox": [ + 154, + 403, + 459, + 415 + ], + "type": "text", + "content": "Fig. 3: Qualitative comparisons on the Neural 3D Video Dataset." + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_caption" + } + ], + "index": 2 + }, + { + "bbox": [ + 130, + 439, + 482, + 559 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 439, + 482, + 559 + ], + "spans": [ + { + "bbox": [ + 130, + 439, + 482, + 559 + ], + "type": "text", + "content": "Table 1 presents quantitative metrics on the average metrics across the test views of all scenes, computational and storage costs on the first fragment of flame_salmon scene. Refer to the Appendix for per-scene details. Our method demonstrates superior reconstruction quality, FPS, and model size across compared to baselines. As the table shows, NeRF baselines generally required longer training and rendering times. While 4DGS shows relatively high reconstruction performance, it demands longer training times and larger VRAM storage capacity compared to other baselines. 4DGaussians requires lower computational and storage cost but it displays low reconstruction quality in some scenes with rapid dynamics, as shown in the teaser and Figure 3." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 130, + 559, + 482, + 643 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 559, + 482, + 643 + ], + "spans": [ + { + "bbox": [ + 130, + 559, + 482, + 643 + ], + "type": "text", + "content": "Figure 3 reports the rendering quality. Our method successfully reconstructs the fine details in moving areas, outperforming baselines on average metrics across test views. Baselines show blurred dynamic areas or severe artifacts in low-light scenes such as cook_spinach and flame_steak. 4DGS exhibits the disappearance of some static areas. In 4DGaussians, a consistent over-smoothing occurs in dynamic areas. All baselines experienced reduced quality in reflective or thin dynamic areas like clamps, torches, and windows." + } + ] + } + ], + "index": 5 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 133, + 92, + 140, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 92, + 140, + 100 + ], + "spans": [ + { + "bbox": [ + 133, + 92, + 140, + 100 + ], + "type": "text", + "content": "8" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 166, + 91, + 217, + 100 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 166, + 91, + 217, + 100 + ], + "spans": [ + { + "bbox": [ + 166, + 91, + 217, + 100 + ], + "type": "text", + "content": "J. Bae et al." + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 7 + }, + { + "para_blocks": [ + { + "type": "table", + "bbox": [ + 186, + 136, + 425, + 191 + ], + "blocks": [ + { + "bbox": [ + 142, + 114, + 470, + 125 + ], + "lines": [ + { + "bbox": [ + 142, + 114, + 470, + 125 + ], + "spans": [ + { + "bbox": [ + 142, + 114, + 470, + 125 + ], + "type": "text", + "content": "Table 2: Average performance in the test view on Technicolor Dataset" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 186, + 136, + 425, + 191 + ], + "lines": [ + { + "bbox": [ + 186, + 136, + 425, + 191 + ], + "spans": [ + { + "bbox": [ + 186, + 136, + 425, + 191 + ], + "type": "table", + "html": "
modelmetriccomputational cost
PSNR↑SSIM↑LPIPS↓Training time↓FPS↑Model size↓
DyNeRF31.80-0.140-0.020.6 MB
HyperReel32.320.8990.1182 hours 45 mins0.91289 MB
4DGaussians29.620.8440.17625 mins34.851 MB
Ours33.240.9070.1002 hours 55 mins60.877 MB
", + "image_path": "3aa0e70ff4a26c3fd49461ceb8ceb37cbbfb5423f11426610880f896522f372e.jpg" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "table_body" + } + ], + "index": 3 + }, + { + "type": "image", + "bbox": [ + 161, + 213, + 452, + 409 + ], + "blocks": [ + { + "bbox": [ + 161, + 213, + 452, + 409 + ], + "lines": [ + { + "bbox": [ + 161, + 213, + 452, + 409 + ], + "spans": [ + { + "bbox": [ + 161, + 213, + 452, + 409 + ], + "type": "image", + "image_path": "7b89d6665a098b61b9ff43f633b4ccaaeaac2f65d6bfac8f0a81fe4c2d88105d.jpg" + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 167, + 420, + 445, + 432 + ], + "lines": [ + { + "bbox": [ + 167, + 420, + 445, + 432 + ], + "spans": [ + { + "bbox": [ + 167, + 420, + 445, + 432 + ], + "type": "text", + "content": "Fig. 4: Qualitative comparisons on the Technicolor Dataset." + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "image_caption" + } + ], + "index": 4 + }, + { + "bbox": [ + 130, + 464, + 480, + 524 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 464, + 480, + 524 + ], + "spans": [ + { + "bbox": [ + 130, + 464, + 480, + 524 + ], + "type": "text", + "content": "Technicolor Light Field Dataset [23] is a multi-view dataset captured with a time-synchronized " + }, + { + "bbox": [ + 130, + 464, + 480, + 524 + ], + "type": "inline_equation", + "content": "4 \\times 4" + }, + { + "bbox": [ + 130, + 464, + 480, + 524 + ], + "type": "text", + "content": " camera rig, containing intricate details. We train ours and the baselines on 50 frames of five commonly used scenes (Birthday, Fabien, Painter, Theater, Trains) using full-resolution videos at " + }, + { + "bbox": [ + 130, + 464, + 480, + 524 + ], + "type": "inline_equation", + "content": "2048 \\times 1088" + }, + { + "bbox": [ + 130, + 464, + 480, + 524 + ], + "type": "text", + "content": " pixels, with the second row and second column cameras used as test views." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 130, + 529, + 482, + 613 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 529, + 482, + 613 + ], + "spans": [ + { + "bbox": [ + 130, + 529, + 482, + 613 + ], + "type": "text", + "content": "Table 2 reports the average metrics across the test views of all scenes, computational and storage costs on Painter scene. HyperReel demonstrates overall high-quality results but struggles with relatively slow training times and FPS, and a larger model size. 4DGaussians exhibits fast training times and FPS but significantly underperforms in reconstructing fine details compared to other baselines. However, our method demonstrates superior reconstruction quality and faster FPS compared to the baselines." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 130, + 617, + 482, + 666 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 617, + 482, + 666 + ], + "spans": [ + { + "bbox": [ + 130, + 617, + 482, + 666 + ], + "type": "text", + "content": "As shown in Figure 4, HyperReel produces noisy artifacts due to incorrect predictions of the displacement vector. 4D Gaussians fails to capture fine details in dynamic areas, exhibiting over-smoothing results. All baselines struggle to accurately reconstruct rapidly moving thin areas like fingers." + } + ] + } + ], + "index": 8 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 405, + 91, + 447, + 100 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 405, + 91, + 447, + 100 + ], + "spans": [ + { + "bbox": [ + 405, + 91, + 447, + 100 + ], + "type": "text", + "content": "E-D3DGS" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 474, + 92, + 481, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 474, + 92, + 481, + 100 + ], + "spans": [ + { + "bbox": [ + 474, + 92, + 481, + 100 + ], + "type": "text", + "content": "9" + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 8 + }, + { + "para_blocks": [ + { + "bbox": [ + 132, + 116, + 366, + 129 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 116, + 366, + 129 + ], + "spans": [ + { + "bbox": [ + 132, + 116, + 366, + 129 + ], + "type": "text", + "content": "4.2 Scenes with Challenging Camera Setting" + } + ] + } + ], + "index": 2 + }, + { + "type": "table", + "bbox": [ + 186, + 175, + 427, + 243 + ], + "blocks": [ + { + "bbox": [ + 145, + 153, + 468, + 165 + ], + "lines": [ + { + "bbox": [ + 145, + 153, + 468, + 165 + ], + "spans": [ + { + "bbox": [ + 145, + 153, + 468, + 165 + ], + "type": "text", + "content": "Table 3: Average performance in the test view on Hypernerf Dataset" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 186, + 175, + 427, + 243 + ], + "lines": [ + { + "bbox": [ + 186, + 175, + 427, + 243 + ], + "spans": [ + { + "bbox": [ + 186, + 175, + 427, + 243 + ], + "type": "table", + "html": "
modelmetriccomputational cost
PSNR↑SSIM↑LPIPS↓Training time↓FPS↑Model size↓
Nerfies [20]22.23-0.170~ hours< 1-
HyperNeRF DS [21]22.290.5980.15332 hours< 115 MB
TiNeuVox [4]24.200.6160.3933 hours 30 mins148 MB
D3DGS [31]22.400.5980.27516 mins6.95309 MB
4DGaussians25.030.6820.28150 mins96.360 MB
Ours25.430.6970.2311 hours 15 mins139.333 MB
", + "image_path": "59206c1fde35444860325d35db55ec4c7d9f13f6cce2f4cd894da96552013a02.jpg" + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "table_body" + } + ], + "index": 4 + }, + { + "type": "image", + "bbox": [ + 189, + 284, + 428, + 486 + ], + "blocks": [ + { + "bbox": [ + 189, + 284, + 428, + 486 + ], + "lines": [ + { + "bbox": [ + 189, + 284, + 428, + 486 + ], + "spans": [ + { + "bbox": [ + 189, + 284, + 428, + 486 + ], + "type": "image", + "image_path": "0af1d5d643a009ddb43d392d4dda53c65c6d78129889c44152499a7f00721d7d.jpg" + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 165, + 496, + 448, + 509 + ], + "lines": [ + { + "bbox": [ + 165, + 496, + 448, + 509 + ], + "spans": [ + { + "bbox": [ + 165, + 496, + 448, + 509 + ], + "type": "text", + "content": "Fig. 5: Qualitative comparisons on the HyperNeRF Dataset." + } + ] + } + ], + "index": 6, + "angle": 0, + "type": "image_caption" + } + ], + "index": 5 + }, + { + "bbox": [ + 130, + 545, + 488, + 594 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 545, + 488, + 594 + ], + "spans": [ + { + "bbox": [ + 130, + 545, + 488, + 594 + ], + "type": "text", + "content": "HyperNeRF dataset includes videos captured using two phones rigidly mounted on a handheld rig. We train on all frames of four scenes (3D Printer, Banana, Broom, Chicken) at a resolution downsampled by half to " + }, + { + "bbox": [ + 130, + 545, + 488, + 594 + ], + "type": "inline_equation", + "content": "536 \\times 960" + }, + { + "bbox": [ + 130, + 545, + 488, + 594 + ], + "type": "text", + "content": ". Due to memory constraints, D3DGS is trained on images downsampled by a quarter." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 130, + 594, + 482, + 665 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 594, + 482, + 665 + ], + "spans": [ + { + "bbox": [ + 130, + 594, + 482, + 665 + ], + "type": "text", + "content": "The table shows that our method outperforms the reconstruction performance with previous methods along with compact model size and faster FPS. Figure 5 shows that previous methods struggle to reconstruct fast-moving parts such as fingers and broom. Especially D3DGS deteriorates in Broom scene. Table 3 reports the average metrics across the test views of all scenes, computational and storage costs on Broom scene." + } + ] + } + ], + "index": 8 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "spans": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "type": "text", + "content": "10" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 166, + 91, + 217, + 100 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 166, + 91, + 217, + 100 + ], + "spans": [ + { + "bbox": [ + 166, + 91, + 217, + 100 + ], + "type": "text", + "content": "J. Bae et al." + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 9 + }, + { + "para_blocks": [ + { + "bbox": [ + 132, + 116, + 308, + 128 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 116, + 308, + 128 + ], + "spans": [ + { + "bbox": [ + 132, + 116, + 308, + 128 + ], + "type": "text", + "content": "4.3 Analyses and Ablation Study" + } + ] + } + ], + "index": 2 + }, + { + "type": "image", + "bbox": [ + 140, + 160, + 222, + 232 + ], + "blocks": [ + { + "bbox": [ + 140, + 160, + 222, + 232 + ], + "lines": [ + { + "bbox": [ + 140, + 160, + 222, + 232 + ], + "spans": [ + { + "bbox": [ + 140, + 160, + 222, + 232 + ], + "type": "image", + "image_path": "554182633416a77d1122a24ee13aca9acf78d3e7386f4b74f0a86321b399cf3f.jpg" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 142, + 234, + 220, + 243 + ], + "lines": [ + { + "bbox": [ + 142, + 234, + 220, + 243 + ], + "spans": [ + { + "bbox": [ + 142, + 234, + 220, + 243 + ], + "type": "text", + "content": "(a) canonical rendering" + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "image_caption" + } + ], + "index": 3 + }, + { + "type": "image", + "bbox": [ + 224, + 160, + 306, + 232 + ], + "blocks": [ + { + "bbox": [ + 224, + 160, + 306, + 232 + ], + "lines": [ + { + "bbox": [ + 224, + 160, + 306, + 232 + ], + "spans": [ + { + "bbox": [ + 224, + 160, + 306, + 232 + ], + "type": "image", + "image_path": "ef95661573a08819491da6e91681cbeac3e0e8a019056f9fd124f980af81b83d.jpg" + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 228, + 233, + 305, + 243 + ], + "lines": [ + { + "bbox": [ + 228, + 233, + 305, + 243 + ], + "spans": [ + { + "bbox": [ + 228, + 233, + 305, + 243 + ], + "type": "text", + "content": "(b) coarse deform only" + } + ] + } + ], + "index": 6, + "angle": 0, + "type": "image_caption" + } + ], + "index": 5 + }, + { + "type": "image", + "bbox": [ + 309, + 160, + 390, + 232 + ], + "blocks": [ + { + "bbox": [ + 309, + 160, + 390, + 232 + ], + "lines": [ + { + "bbox": [ + 309, + 160, + 390, + 232 + ], + "spans": [ + { + "bbox": [ + 309, + 160, + 390, + 232 + ], + "type": "image", + "image_path": "f50c0b5dffd6ba9cf2bff68b59325df636266e36b8d7c187275bfb9149d0f806.jpg" + } + ] + } + ], + "index": 7, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 315, + 233, + 383, + 243 + ], + "lines": [ + { + "bbox": [ + 315, + 233, + 383, + 243 + ], + "spans": [ + { + "bbox": [ + 315, + 233, + 383, + 243 + ], + "type": "text", + "content": "(c) fine deform only" + } + ] + } + ], + "index": 8, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 130, + 255, + 482, + 320 + ], + "lines": [ + { + "bbox": [ + 130, + 255, + 482, + 320 + ], + "spans": [ + { + "bbox": [ + 130, + 255, + 482, + 320 + ], + "type": "text", + "content": "Fig. 6: Deformation components. (a) The canonical space contains Gaussians to represent all target frames of the scene. (b) Applying coarse deformation to the canonical space roughly reflects the dynamics of the scene. (c) The rendering without coarse deformation and only with fine deformation looks similar to the canonical rendering, i.e., responsible for fine deformations. (d) Applying both coarse and fine deformation yields natural rendering results." + } + ] + } + ], + "index": 11, + "angle": 0, + "type": "image_caption" + } + ], + "index": 7 + }, + { + "type": "image", + "bbox": [ + 392, + 160, + 474, + 232 + ], + "blocks": [ + { + "bbox": [ + 392, + 160, + 474, + 232 + ], + "lines": [ + { + "bbox": [ + 392, + 160, + 474, + 232 + ], + "spans": [ + { + "bbox": [ + 392, + 160, + 474, + 232 + ], + "type": "image", + "image_path": "18b121f78a54ca89cbabdf980caff35cc0c5327d1255bd859b62f672dc88a3be.jpg" + } + ] + } + ], + "index": 9, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 404, + 233, + 462, + 243 + ], + "lines": [ + { + "bbox": [ + 404, + 233, + 462, + 243 + ], + "spans": [ + { + "bbox": [ + 404, + 233, + 462, + 243 + ], + "type": "text", + "content": "(d) full rendering" + } + ] + } + ], + "index": 10, + "angle": 0, + "type": "image_caption" + } + ], + "index": 9 + }, + { + "type": "image", + "bbox": [ + 156, + 369, + 305, + 449 + ], + "blocks": [ + { + "bbox": [ + 156, + 369, + 305, + 449 + ], + "lines": [ + { + "bbox": [ + 156, + 369, + 305, + 449 + ], + "spans": [ + { + "bbox": [ + 156, + 369, + 305, + 449 + ], + "type": "image", + "image_path": "fb76741ca4e880981dcf037fa4e559ba30b0687e2bde18d7bc9053a7b61b7084.jpg" + } + ] + } + ], + "index": 12, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 209, + 450, + 251, + 460 + ], + "lines": [ + { + "bbox": [ + 209, + 450, + 251, + 460 + ], + "spans": [ + { + "bbox": [ + 209, + 450, + 251, + 460 + ], + "type": "text", + "content": "Frame 67" + } + ] + } + ], + "index": 13, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 130, + 472, + 482, + 515 + ], + "lines": [ + { + "bbox": [ + 130, + 472, + 482, + 515 + ], + "spans": [ + { + "bbox": [ + 130, + 472, + 482, + 515 + ], + "type": "text", + "content": "Fig.7: Visualization of the magnitude of deformation. Coarse deformation (blue) captures large and slow changes, such as the movement of the head and torso, while fine deformation (red) is responsible for the fast and detailed movements of arms, tongs, shadows, etc." + } + ] + } + ], + "index": 16, + "angle": 0, + "type": "image_caption" + } + ], + "index": 12 + }, + { + "type": "image", + "bbox": [ + 310, + 369, + 459, + 449 + ], + "blocks": [ + { + "bbox": [ + 310, + 369, + 459, + 449 + ], + "lines": [ + { + "bbox": [ + 310, + 369, + 459, + 449 + ], + "spans": [ + { + "bbox": [ + 310, + 369, + 459, + 449 + ], + "type": "image", + "image_path": "50d55bb686db8f59500bd3aa1cf5cfb1e87af5d1b32675153dff4ad99b305459.jpg" + } + ] + } + ], + "index": 14, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 361, + 450, + 408, + 460 + ], + "lines": [ + { + "bbox": [ + 361, + 450, + 408, + 460 + ], + "spans": [ + { + "bbox": [ + 361, + 450, + 408, + 460 + ], + "type": "text", + "content": "Frame 185" + } + ] + } + ], + "index": 15, + "angle": 0, + "type": "image_caption" + } + ], + "index": 14 + }, + { + "bbox": [ + 130, + 558, + 482, + 666 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 558, + 482, + 666 + ], + "spans": [ + { + "bbox": [ + 130, + 558, + 482, + 666 + ], + "type": "text", + "content": "Deformation components In Figure 6, we present an analysis of the coarse-fine deformation. To achieve this, we render a flame_steak scene by omitting each of our deformation components one by one. Our full rendering results from adding coarse and fine deformation to the canonical space (Figure 6d). When both are removed, rendering yields an image in canonical space (Figure 6a). Rendering with the coarse deformation, which handles large or slow changes in the scene, produces results similar to the full rendering (Figure 6b). On the other hand, fine deformation is responsible for fast or detailed changes in the scene, yielding rendering similar to canonical space (Figure 6c)." + } + ] + } + ], + "index": 17 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 405, + 91, + 447, + 100 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 405, + 91, + 447, + 100 + ], + "spans": [ + { + "bbox": [ + 405, + 91, + 447, + 100 + ], + "type": "text", + "content": "E-D3DGS" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 470, + 92, + 480, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 470, + 92, + 480, + 100 + ], + "spans": [ + { + "bbox": [ + 470, + 92, + 480, + 100 + ], + "type": "text", + "content": "11" + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 10 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 141, + 116, + 222, + 163 + ], + "blocks": [ + { + "bbox": [ + 141, + 116, + 222, + 163 + ], + "lines": [ + { + "bbox": [ + 141, + 116, + 222, + 163 + ], + "spans": [ + { + "bbox": [ + 141, + 116, + 222, + 163 + ], + "type": "image", + "image_path": "2bca6504c44f69bb82078709672f279a3f0355b578947c07ed64af8cc70a0d42.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 158, + 164, + 206, + 174 + ], + "lines": [ + { + "bbox": [ + 158, + 164, + 206, + 174 + ], + "spans": [ + { + "bbox": [ + 158, + 164, + 206, + 174 + ], + "type": "text", + "content": "(a) Ours Full" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_caption" + } + ], + "index": 2 + }, + { + "type": "image", + "bbox": [ + 225, + 116, + 306, + 163 + ], + "blocks": [ + { + "bbox": [ + 225, + 116, + 306, + 163 + ], + "lines": [ + { + "bbox": [ + 225, + 116, + 306, + 163 + ], + "spans": [ + { + "bbox": [ + 225, + 116, + 306, + 163 + ], + "type": "image", + "image_path": "e0e64da9a3a3a1ee90ae7561f2342964b90def5092752639732998512bd40911.jpg" + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 228, + 164, + 304, + 174 + ], + "lines": [ + { + "bbox": [ + 228, + 164, + 304, + 174 + ], + "spans": [ + { + "bbox": [ + 228, + 164, + 304, + 174 + ], + "type": "text", + "content": "(b) coarse deform only" + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 130, + 183, + 482, + 240 + ], + "lines": [ + { + "bbox": [ + 130, + 183, + 482, + 240 + ], + "spans": [ + { + "bbox": [ + 130, + 183, + 482, + 240 + ], + "type": "text", + "content": "Fig. 8: Qualitative ablation results on coarse-fine deformation. (a) Our model achieves clear results with both coarse and fine decoders. (b-c) The quality of dynamic areas decreases if one is missing. (d) Additionally, introducing the coordinates of Gaussian as an additional input into our decoders results in a decrease in the quality of both static and dynamic regions." + } + ] + } + ], + "index": 10, + "angle": 0, + "type": "image_caption" + } + ], + "index": 4 + }, + { + "type": "image", + "bbox": [ + 309, + 116, + 389, + 163 + ], + "blocks": [ + { + "bbox": [ + 309, + 116, + 389, + 163 + ], + "lines": [ + { + "bbox": [ + 309, + 116, + 389, + 163 + ], + "spans": [ + { + "bbox": [ + 309, + 116, + 389, + 163 + ], + "type": "image", + "image_path": "4395843b08f134c7c78a9cdf10b0cf79b6f072e54949ccd9e6aa5029378afb78.jpg" + } + ] + } + ], + "index": 6, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 315, + 164, + 383, + 174 + ], + "lines": [ + { + "bbox": [ + 315, + 164, + 383, + 174 + ], + "spans": [ + { + "bbox": [ + 315, + 164, + 383, + 174 + ], + "type": "text", + "content": "(c) fine deform only" + } + ] + } + ], + "index": 7, + "angle": 0, + "type": "image_caption" + } + ], + "index": 6 + }, + { + "type": "image", + "bbox": [ + 392, + 116, + 473, + 163 + ], + "blocks": [ + { + "bbox": [ + 392, + 116, + 473, + 163 + ], + "lines": [ + { + "bbox": [ + 392, + 116, + 473, + 163 + ], + "spans": [ + { + "bbox": [ + 392, + 116, + 473, + 163 + ], + "type": "image", + "image_path": "bf4ba8ceca3ad52abc64a7819dd90ae4f1080d641cecd2f61b3272451caa9f09.jpg" + } + ] + } + ], + "index": 8, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 395, + 164, + 470, + 174 + ], + "lines": [ + { + "bbox": [ + 395, + 164, + 470, + 174 + ], + "spans": [ + { + "bbox": [ + 395, + 164, + 470, + 174 + ], + "type": "text", + "content": "(d) Ours " + }, + { + "bbox": [ + 395, + 164, + 470, + 174 + ], + "type": "inline_equation", + "content": "+\\mathbf{x}" + }, + { + "bbox": [ + 395, + 164, + 470, + 174 + ], + "type": "text", + "content": " injection" + } + ] + } + ], + "index": 9, + "angle": 0, + "type": "image_caption" + } + ], + "index": 8 + }, + { + "type": "table", + "bbox": [ + 186, + 273, + 425, + 332 + ], + "blocks": [ + { + "bbox": [ + 153, + 251, + 459, + 264 + ], + "lines": [ + { + "bbox": [ + 153, + 251, + 459, + 264 + ], + "spans": [ + { + "bbox": [ + 153, + 251, + 459, + 264 + ], + "type": "text", + "content": "Table 4: Quantitative ablation results on coarse-fine deformation." + } + ] + } + ], + "index": 11, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 186, + 273, + 425, + 332 + ], + "lines": [ + { + "bbox": [ + 186, + 273, + 425, + 332 + ], + "spans": [ + { + "bbox": [ + 186, + 273, + 425, + 332 + ], + "type": "table", + "html": "
MethodPSNR↑SSIM↑LPIPS↓
Ours29.700.9330.041
coarse deformation only29.480.9310.044
fine deformation only29.230.9320.043
Ours + x injection29.600.9310.045
", + "image_path": "8121f6ebd9c0cb9ce637c36f4b3a5ff9de4d9de0d967689cb7b984f00bca8fc1.jpg" + } + ] + } + ], + "index": 12, + "angle": 0, + "type": "table_body" + } + ], + "index": 12 + }, + { + "bbox": [ + 130, + 357, + 482, + 477 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 357, + 482, + 477 + ], + "spans": [ + { + "bbox": [ + 130, + 357, + 482, + 477 + ], + "type": "text", + "content": "To examine the roles of the coarse and fine deformation in the coarse-fine deformation, we conduct a visualization on flame_steak scene. First, we compute the Euclidean norm of positional shifts between the current and subsequent frames. We then add the value to the DC components of the SH coefficients proportionally to the magnitude: blue for coarse deformation and red for fine deformation. For visual clarity, we render the original scene in grayscale. As illustrated in Figure 7, coarse deformation models slower changes such as body movement, while fine deformation models faster movements like cooking arms. Thus, we demonstrate that by downsampling the temporal embedding grid " + }, + { + "bbox": [ + 130, + 357, + 482, + 477 + ], + "type": "inline_equation", + "content": "Z" + }, + { + "bbox": [ + 130, + 357, + 482, + 477 + ], + "type": "text", + "content": ", we can effectively separate and model slow and fast deformations in the scene." + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 130, + 497, + 482, + 616 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 497, + 482, + 616 + ], + "spans": [ + { + "bbox": [ + 130, + 497, + 482, + 616 + ], + "type": "text", + "content": "Ablation study We report the results of an ablation study on the deformation decoder in Figure 8 and Table 4. First, our full method (using both coarse and fine decoders) produces clear rendering results and models dynamic variations well (Figure 8a). Training only with the coarse or fine decoder leads to blurred dynamic areas and a failure to accurately capture detailed motion (Figure 8b-c). Additionally, we demonstrate experiments where the Gaussian center coordinates " + }, + { + "bbox": [ + 130, + 497, + 482, + 616 + ], + "type": "inline_equation", + "content": "\\mathbf{x}" + }, + { + "bbox": [ + 130, + 497, + 482, + 616 + ], + "type": "text", + "content": " are injected into the input of each decoder. As shown in Figure 8d, including the Gaussian coordinates degrades the quality of deformation, supporting our argument that coordinate dependency should be removed from the deformation function." + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 130, + 617, + 482, + 666 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 617, + 482, + 666 + ], + "spans": [ + { + "bbox": [ + 130, + 617, + 482, + 666 + ], + "type": "text", + "content": "Furthermore, we report the results of an ablation study on the local smoothness regularization for per-Gaussian embeddings. As shown in Figure 9, our regularization improves the details and texture quality of moving objects. In Table 5, we show a performance comparison between the proposed regularization and" + } + ] + } + ], + "index": 15 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "spans": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "type": "text", + "content": "12" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 166, + 91, + 217, + 100 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 166, + 91, + 217, + 100 + ], + "spans": [ + { + "bbox": [ + 166, + 91, + 217, + 100 + ], + "type": "text", + "content": "J. Bae et al." + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 11 + }, + { + "para_blocks": [ + { + "type": "table", + "bbox": [ + 177, + 168, + 434, + 217 + ], + "blocks": [ + { + "bbox": [ + 131, + 114, + 482, + 159 + ], + "lines": [ + { + "bbox": [ + 131, + 114, + 482, + 159 + ], + "spans": [ + { + "bbox": [ + 131, + 114, + 482, + 159 + ], + "type": "text", + "content": "Table 5: Quantitative ablation results on local smoothness regularization. We compare the performance of applying our regularization with the physically-based regularization of Dynamic 3D Gaussians [17]. Our regularization better captures details of dynamic objects." + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 177, + 168, + 434, + 217 + ], + "lines": [ + { + "bbox": [ + 177, + 168, + 434, + 217 + ], + "spans": [ + { + "bbox": [ + 177, + 168, + 434, + 217 + ], + "type": "table", + "html": "
MethodPSNR↑SSIM↑LPIPS↓
Ours w/o embedding reg32.260.9510.037
+ our embedding reg32.340.9520.036
+ physically-based reg32.080.9500.036
", + "image_path": "54606bd904be834a2aca183afcb1f51d7c13b0aca4866e91582c9c7c0eabaf92.jpg" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "table_body" + } + ], + "index": 3 + }, + { + "type": "image", + "bbox": [ + 171, + 230, + 236, + 282 + ], + "blocks": [ + { + "bbox": [ + 171, + 230, + 236, + 282 + ], + "lines": [ + { + "bbox": [ + 171, + 230, + 236, + 282 + ], + "spans": [ + { + "bbox": [ + 171, + 230, + 236, + 282 + ], + "type": "image", + "image_path": "d3bae3efa4e225290e81c2d17779d77a3ca69c066c77d6dc77f115a31d42a4d0.jpg" + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 187, + 284, + 220, + 293 + ], + "lines": [ + { + "bbox": [ + 187, + 284, + 220, + 293 + ], + "spans": [ + { + "bbox": [ + 187, + 284, + 220, + 293 + ], + "type": "text", + "content": "Ours full" + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "image_caption" + } + ], + "index": 4 + }, + { + "type": "image", + "bbox": [ + 238, + 231, + 302, + 282 + ], + "blocks": [ + { + "bbox": [ + 238, + 231, + 302, + 282 + ], + "lines": [ + { + "bbox": [ + 238, + 231, + 302, + 282 + ], + "spans": [ + { + "bbox": [ + 238, + 231, + 302, + 282 + ], + "type": "image", + "image_path": "c36f81fa9ab94dd9ac95aacb559f7ef87d408ee6efb558c2020798abcf59085f.jpg" + } + ] + } + ], + "index": 6, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 247, + 284, + 293, + 293 + ], + "lines": [ + { + "bbox": [ + 247, + 284, + 293, + 293 + ], + "spans": [ + { + "bbox": [ + 247, + 284, + 293, + 293 + ], + "type": "text", + "content": "Ours w/o reg" + } + ] + } + ], + "index": 7, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 142, + 305, + 470, + 317 + ], + "lines": [ + { + "bbox": [ + 142, + 305, + 470, + 317 + ], + "spans": [ + { + "bbox": [ + 142, + 305, + 470, + 317 + ], + "type": "text", + "content": "Fig. 9: Qualitative ablation results on local smoothness regularization." + } + ] + } + ], + "index": 12, + "angle": 0, + "type": "image_caption" + } + ], + "index": 6 + }, + { + "type": "image", + "bbox": [ + 311, + 230, + 376, + 282 + ], + "blocks": [ + { + "bbox": [ + 311, + 230, + 376, + 282 + ], + "lines": [ + { + "bbox": [ + 311, + 230, + 376, + 282 + ], + "spans": [ + { + "bbox": [ + 311, + 230, + 376, + 282 + ], + "type": "image", + "image_path": "8b07b1a8212952eb21f2114145017091c1187a1b426506870859cc4dab973fbe.jpg" + } + ] + } + ], + "index": 8, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 329, + 284, + 363, + 293 + ], + "lines": [ + { + "bbox": [ + 329, + 284, + 363, + 293 + ], + "spans": [ + { + "bbox": [ + 329, + 284, + 363, + 293 + ], + "type": "text", + "content": "Ours full" + } + ] + } + ], + "index": 9, + "angle": 0, + "type": "image_caption" + } + ], + "index": 8 + }, + { + "type": "image", + "bbox": [ + 377, + 231, + 443, + 282 + ], + "blocks": [ + { + "bbox": [ + 377, + 231, + 443, + 282 + ], + "lines": [ + { + "bbox": [ + 377, + 231, + 443, + 282 + ], + "spans": [ + { + "bbox": [ + 377, + 231, + 443, + 282 + ], + "type": "image", + "image_path": "06618c98b077500de0b23326417919b4bc8618a7dc9b5629c67c1a56fee04333.jpg" + } + ] + } + ], + "index": 10, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 388, + 284, + 433, + 293 + ], + "lines": [ + { + "bbox": [ + 388, + 284, + 433, + 293 + ], + "spans": [ + { + "bbox": [ + 388, + 284, + 433, + 293 + ], + "type": "text", + "content": "Ours w/o reg" + } + ] + } + ], + "index": 11, + "angle": 0, + "type": "image_caption" + } + ], + "index": 10 + }, + { + "bbox": [ + 130, + 340, + 482, + 425 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 340, + 482, + 425 + ], + "spans": [ + { + "bbox": [ + 130, + 340, + 482, + 425 + ], + "type": "text", + "content": "the existing physically-based regularizations. To apply the method proposed by Luiten et al. [17] to ours, we make some modifications: 1) Like our method, we find the set of k-nearest-neighbours only when the densification occurs to reduce the computational cost. 2) For long-term local-isometry loss, we use the time of video frame used in previous training step instead of using the time of the first frame. Our regularization is simple and shows better performance compared to previous method." + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 131, + 442, + 312, + 455 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 131, + 442, + 312, + 455 + ], + "spans": [ + { + "bbox": [ + 131, + 442, + 312, + 455 + ], + "type": "text", + "content": "5 Conclusion and Limitation" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 130, + 468, + 482, + 552 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 468, + 482, + 552 + ], + "spans": [ + { + "bbox": [ + 130, + 468, + 482, + 552 + ], + "type": "text", + "content": "We propose a per-Gaussian deformation for 3DGS that takes per-Gaussian embeddings as input, instead of using the typical deformation fields from previous deformable 3DGS works, resulting in high performance. We enhance the reconstruction quality by decomposing the dynamic changes into coarse and fine deformation. However, our method learns inappropriate Gaussian deformation with casually captured monocular videos [7], like other baselines. We plan to address it in future work by introducing useful prior for monocular video settings." + } + ] + } + ], + "index": 15 + }, + { + "type": "image", + "bbox": [ + 155, + 585, + 228, + 632 + ], + "blocks": [ + { + "bbox": [ + 155, + 585, + 228, + 632 + ], + "lines": [ + { + "bbox": [ + 155, + 585, + 228, + 632 + ], + "spans": [ + { + "bbox": [ + 155, + 585, + 228, + 632 + ], + "type": "image", + "image_path": "6244d426dee9c97c7a6af5d392b85dc2e62d9930adb0f449e27e8ee2e9bf65d4.jpg" + } + ] + } + ], + "index": 16, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 170, + 632, + 214, + 641 + ], + "lines": [ + { + "bbox": [ + 170, + 632, + 214, + 641 + ], + "spans": [ + { + "bbox": [ + 170, + 632, + 214, + 641 + ], + "type": "text", + "content": "4DGaussians" + } + ] + } + ], + "index": 17, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 136, + 651, + 475, + 662 + ], + "lines": [ + { + "bbox": [ + 136, + 651, + 475, + 662 + ], + "spans": [ + { + "bbox": [ + 136, + 651, + 475, + 662 + ], + "type": "text", + "content": "Fig. 10: Limitation. Ours struggles with the casually captured monocular videos." + } + ] + } + ], + "index": 24, + "angle": 0, + "type": "image_caption" + } + ], + "index": 16 + }, + { + "type": "image", + "bbox": [ + 230, + 585, + 304, + 632 + ], + "blocks": [ + { + "bbox": [ + 230, + 585, + 304, + 632 + ], + "lines": [ + { + "bbox": [ + 230, + 585, + 304, + 632 + ], + "spans": [ + { + "bbox": [ + 230, + 585, + 304, + 632 + ], + "type": "image", + "image_path": "f4931a4dd7763d923edda247a6db3fd6060e5ba8252a115078cb119a2e7cfd80.jpg" + } + ] + } + ], + "index": 18, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 257, + 633, + 276, + 641 + ], + "lines": [ + { + "bbox": [ + 257, + 633, + 276, + 641 + ], + "spans": [ + { + "bbox": [ + 257, + 633, + 276, + 641 + ], + "type": "text", + "content": "Ours" + } + ] + } + ], + "index": 19, + "angle": 0, + "type": "image_caption" + } + ], + "index": 18 + }, + { + "type": "image", + "bbox": [ + 310, + 585, + 384, + 632 + ], + "blocks": [ + { + "bbox": [ + 310, + 585, + 384, + 632 + ], + "lines": [ + { + "bbox": [ + 310, + 585, + 384, + 632 + ], + "spans": [ + { + "bbox": [ + 310, + 585, + 384, + 632 + ], + "type": "image", + "image_path": "6ac9a1640c0e847e7259dda48c198e6168348e5d8b5c2bc45e4371ad1b744732.jpg" + } + ] + } + ], + "index": 20, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 325, + 632, + 369, + 641 + ], + "lines": [ + { + "bbox": [ + 325, + 632, + 369, + 641 + ], + "spans": [ + { + "bbox": [ + 325, + 632, + 369, + 641 + ], + "type": "text", + "content": "4DGaussians" + } + ] + } + ], + "index": 21, + "angle": 0, + "type": "image_caption" + } + ], + "index": 20 + }, + { + "type": "image", + "bbox": [ + 385, + 585, + 459, + 632 + ], + "blocks": [ + { + "bbox": [ + 385, + 585, + 459, + 632 + ], + "lines": [ + { + "bbox": [ + 385, + 585, + 459, + 632 + ], + "spans": [ + { + "bbox": [ + 385, + 585, + 459, + 632 + ], + "type": "image", + "image_path": "6097d5b27e045786fd13e8636bf6c77675c427cc4df28eedc04e4c0608ef8c21.jpg" + } + ] + } + ], + "index": 22, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 412, + 633, + 431, + 641 + ], + "lines": [ + { + "bbox": [ + 412, + 633, + 431, + 641 + ], + "spans": [ + { + "bbox": [ + 412, + 633, + 431, + 641 + ], + "type": "text", + "content": "Ours" + } + ] + } + ], + "index": 23, + "angle": 0, + "type": "image_caption" + } + ], + "index": 22 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 405, + 91, + 447, + 100 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 405, + 91, + 447, + 100 + ], + "spans": [ + { + "bbox": [ + 405, + 91, + 447, + 100 + ], + "type": "text", + "content": "E-D3DGS" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 470, + 91, + 481, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 470, + 91, + 481, + 100 + ], + "spans": [ + { + "bbox": [ + 470, + 91, + 481, + 100 + ], + "type": "text", + "content": "13" + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 12 + }, + { + "para_blocks": [ + { + "bbox": [ + 133, + 114, + 246, + 129 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 114, + 246, + 129 + ], + "spans": [ + { + "bbox": [ + 133, + 114, + 246, + 129 + ], + "type": "text", + "content": "Acknowledgements" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 132, + 139, + 481, + 186 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 139, + 481, + 186 + ], + "spans": [ + { + "bbox": [ + 132, + 139, + 481, + 186 + ], + "type": "text", + "content": "This work is supported by the Institute for Information & Communications Technology Planning & Evaluation (IITP) grant funded by the Korea government(MSIT) (No. 2017-0-00072, Development of Audio/Video Coding and Light Field Media Fundamental Technologies for Ultra Realistic Tera-media)" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 133, + 205, + 197, + 217 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 205, + 197, + 217 + ], + "spans": [ + { + "bbox": [ + 133, + 205, + 197, + 217 + ], + "type": "text", + "content": "References" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 134, + 228, + 481, + 665 + ], + "type": "list", + "angle": 0, + "index": 18, + "blocks": [ + { + "bbox": [ + 138, + 228, + 481, + 273 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 228, + 481, + 273 + ], + "spans": [ + { + "bbox": [ + 138, + 228, + 481, + 273 + ], + "type": "text", + "content": "1. Attal, B., Huang, J.B., Richardt, C., Zollhoefer, M., Kopf, J., OaAZToole, M., Kim, C.: Hyperreel: High-fidelity 6-dof video with ray-conditioned sampling. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 16610-16620 (2023)" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 138, + 274, + 481, + 305 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 274, + 481, + 305 + ], + "spans": [ + { + "bbox": [ + 138, + 274, + 481, + 305 + ], + "type": "text", + "content": "2. Cao, A., Johnson, J.: Hexplane: A fast representation for dynamic scenes. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 130-141 (2023)" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 138, + 306, + 481, + 338 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 306, + 481, + 338 + ], + "spans": [ + { + "bbox": [ + 138, + 306, + 481, + 338 + ], + "type": "text", + "content": "3. Duisterhof, B.P., Mandi, Z., Yao, Y., Liu, J.W., Shou, M.Z., Song, S., Ichnowski, J.: Md-splatting: Learning metric deformation from 4d gaussians in highly deformable scenes. arXiv preprint arXiv:2312.00583 (2023)" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 138, + 339, + 481, + 371 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 339, + 481, + 371 + ], + "spans": [ + { + "bbox": [ + 138, + 339, + 481, + 371 + ], + "type": "text", + "content": "4. Fang, J., Yi, T., Wang, X., Xie, L., Zhang, X., Liu, W., Nießner, M., Tian, Q.: Fast dynamic radiance fields with time-aware neural voxels. In: SIGGRAPH Asia 2022 Conference Papers (2022)" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 138, + 372, + 481, + 403 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 372, + 481, + 403 + ], + "spans": [ + { + "bbox": [ + 138, + 372, + 481, + 403 + ], + "type": "text", + "content": "5. Feichtenhofer, C., Fan, H., Malik, J., He, K.: Slowfast networks for video recognition. In: Proceedings of the IEEE/CVF international conference on computer vision. pp. 6202-6211 (2019)" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 138, + 404, + 481, + 446 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 404, + 481, + 446 + ], + "spans": [ + { + "bbox": [ + 138, + 404, + 481, + 446 + ], + "type": "text", + "content": "6. Fridovich-Keil, S., Meanti, G., Warburg, F.R., Recht, B., Kanazawa, A.: K-planes: Explicit radiance fields in space, time, and appearance. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 12479-12488 (2023)" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 138, + 447, + 481, + 469 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 447, + 481, + 469 + ], + "spans": [ + { + "bbox": [ + 138, + 447, + 481, + 469 + ], + "type": "text", + "content": "7. Gao, H., Li, R., Tulsiani, S., Russell, B., Kanazawa, A.: Monocular dynamic view synthesis: A reality check. In: NeurIPS (2022)" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 138, + 470, + 481, + 502 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 470, + 481, + 502 + ], + "spans": [ + { + "bbox": [ + 138, + 470, + 481, + 502 + ], + "type": "text", + "content": "8. Huang, Y.H., Sun, Y.T., Yang, Z., Lyu, X., Cao, Y.P., Qi, X.: Sc-gs: Sparse-controlled gaussian splatting for editable dynamic scenes. arXiv preprint arXiv:2312.14937 (2023)" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 138, + 502, + 481, + 545 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 502, + 481, + 545 + ], + "spans": [ + { + "bbox": [ + 138, + 502, + 481, + 545 + ], + "type": "text", + "content": "9. Huang, Y.H., Sun, Y.T., Yang, Z., Lyu, X., Cao, Y.P., Qi, X.: Sc-gs: Sparse-controlled gaussian splatting for editable dynamic scenes. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR). pp. 4220-4230 (June 2024)" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 138, + 546, + 481, + 578 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 546, + 481, + 578 + ], + "spans": [ + { + "bbox": [ + 138, + 546, + 481, + 578 + ], + "type": "text", + "content": "0. Kerbl, B., Kopanas, G., Leimkuhler, T., Drettakis, G.: 3d gaussian splatting for real-time radiance field rendering. ACM Transactions on Graphics 42(4) (July 2023), https://repo-sam.inria.fr/fungraph/3d-gaussian-splatting/" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 138, + 578, + 481, + 600 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 578, + 481, + 600 + ], + "spans": [ + { + "bbox": [ + 138, + 578, + 481, + 600 + ], + "type": "text", + "content": "1. Kim, S., Bae, J., Yun, Y., Lee, H., Bang, G., Uh, Y.: Sync-nerf: Generalizing dynamic nerfs to unsynchronized videos. arXiv preprint arXiv:2310.13356 (2023)" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 134, + 601, + 481, + 643 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 134, + 601, + 481, + 643 + ], + "spans": [ + { + "bbox": [ + 134, + 601, + 481, + 643 + ], + "type": "text", + "content": "12. Li, T., Slavcheva, M., Zollhoefer, M., Green, S., Lassner, C., Kim, C., Schmidt, T., Lovegrove, S., Goesele, M., Newcombe, R., et al.: Neural 3d video synthesis from multi-view video. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (2022)" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 134, + 644, + 481, + 665 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 134, + 644, + 481, + 665 + ], + "spans": [ + { + "bbox": [ + 134, + 644, + 481, + 665 + ], + "type": "text", + "content": "13. Li, Z., Chen, Z., Li, Z., Xu, Y.: Spacetime gaussian feature splatting for real-time dynamic view synthesis. arXiv preprint arXiv:2312.16812 (2023)" + } + ] + } + ], + "index": 17 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "spans": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "type": "text", + "content": "14" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 166, + 91, + 217, + 100 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 166, + 91, + 217, + 100 + ], + "spans": [ + { + "bbox": [ + 166, + 91, + 217, + 100 + ], + "type": "text", + "content": "J. Bae et al." + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 13 + }, + { + "para_blocks": [ + { + "bbox": [ + 130, + 116, + 482, + 665 + ], + "type": "list", + "angle": 0, + "index": 20, + "blocks": [ + { + "bbox": [ + 130, + 116, + 482, + 149 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 116, + 482, + 149 + ], + "spans": [ + { + "bbox": [ + 130, + 116, + 482, + 149 + ], + "type": "text", + "content": "14. Liang, Y., Khan, N., Li, Z., Nguyen-Phuoc, T., Lanman, D., Tompkin, J., Xiao, L.: Gaufre: Gaussian deformation fields for real-time dynamic novel view synthesis (2023)" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 130, + 149, + 482, + 182 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 149, + 482, + 182 + ], + "spans": [ + { + "bbox": [ + 130, + 149, + 482, + 182 + ], + "type": "text", + "content": "15. Liu, Q., Liu, Y., Wang, J., Lv, X., Wang, P., Wang, W., Hou, J.: Modgs: Dynamic gaussian splatting from causally-captured monocular videos (2024), https:// arxiv.org/abs/2406.00434" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 130, + 182, + 482, + 214 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 182, + 482, + 214 + ], + "spans": [ + { + "bbox": [ + 130, + 182, + 482, + 214 + ], + "type": "text", + "content": "16. Lu, T., Yu, M., Xu, L., Xiangli, Y., Wang, L., Lin, D., Dai, B.: Scaffold-gs: Structured 3d gaussians for view-adaptive rendering. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 20654-20664 (2024)" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 130, + 214, + 482, + 236 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 214, + 482, + 236 + ], + "spans": [ + { + "bbox": [ + 130, + 214, + 482, + 236 + ], + "type": "text", + "content": "17. Luiten, J., Kopanas, G., Leibe, B., Ramanan, D.: Dynamic 3d gaussians: Tracking by persistent dynamic view synthesis. In: 3DV (2024)" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 130, + 236, + 482, + 267 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 236, + 482, + 267 + ], + "spans": [ + { + "bbox": [ + 130, + 236, + 482, + 267 + ], + "type": "text", + "content": "18. Martin-Brualla, R., Radwan, N., Sajjadi, M.S.M., Barron, J.T., Dosovitskiy, A., Duckworth, D.: NeRF in the Wild: Neural Radiance Fields for Unconstrained Photo Collections. In: CVPR (2021)" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 130, + 267, + 482, + 300 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 267, + 482, + 300 + ], + "spans": [ + { + "bbox": [ + 130, + 267, + 482, + 300 + ], + "type": "text", + "content": "19. Mildenhall, B., Srinivasan, P.P., Tancik, M., Barron, J.T., Ramamoorthi, R., Ng, R.: Nerf: Representing scenes as neural radiance fields for view synthesis. Communications of the ACM 65(1), 99-106 (2021)" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 130, + 300, + 482, + 321 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 300, + 482, + 321 + ], + "spans": [ + { + "bbox": [ + 130, + 300, + 482, + 321 + ], + "type": "text", + "content": "20. Park, K., Sinha, U., Barron, J.T., Bouaziz, S., Goldman, D.B., Seitz, S.M., Martin-Brualla, R.: Nerfies: Deformable neural radiance fields. ICCV (2021)" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 130, + 321, + 482, + 354 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 321, + 482, + 354 + ], + "spans": [ + { + "bbox": [ + 130, + 321, + 482, + 354 + ], + "type": "text", + "content": "21. Park, K., Sinha, U., Hedman, P., Barron, J.T., Bouaziz, S., Goldman, D.B., Martin-Brualla, R., Seitz, S.M.: Hypernerf: A higher-dimensional representation for topologically varying neural radiance fields. arXiv preprint arXiv:2106.13228 (2021)" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 130, + 354, + 482, + 374 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 354, + 482, + 374 + ], + "spans": [ + { + "bbox": [ + 130, + 354, + 482, + 374 + ], + "type": "text", + "content": "22. Pumarola, A., Corona, E., Pons-Moll, G., Moreno-Noguer, F.: D-nerf: Neural radiance fields for dynamic scenes. arXiv preprint arXiv:2011.13961 (2020)" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 130, + 374, + 482, + 418 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 374, + 482, + 418 + ], + "spans": [ + { + "bbox": [ + 130, + 374, + 482, + 418 + ], + "type": "text", + "content": "23. Sabater, N., Boisson, G., Vandame, B., Kerbiriou, P., Babon, F., Hog, M., Gendrot, R., Langlois, T., Bureller, O., Schubert, A., et al.: Dataset and pipeline for multiview light-field video. In: Proceedings of the IEEE conference on computer vision and pattern recognition Workshops. pp. 30-40 (2017)" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 130, + 418, + 482, + 460 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 418, + 482, + 460 + ], + "spans": [ + { + "bbox": [ + 130, + 418, + 482, + 460 + ], + "type": "text", + "content": "24. Song, L., Chen, A., Li, Z., Chen, Z., Chen, L., Yuan, J., Xu, Y., Geiger, A.: Nerflayer: A streamable dynamic scene representation with decomposed neural radiance fields. IEEE Transactions on Visualization and Computer Graphics 29(5), 2732-2742 (2023)" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 130, + 460, + 482, + 494 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 460, + 482, + 494 + ], + "spans": [ + { + "bbox": [ + 130, + 460, + 482, + 494 + ], + "type": "text", + "content": "25. Tancik, M., Casser, V., Yan, X., Pradhan, S., Mildenhall, B., Srinivasan, P.P., Barron, J.T., Kretzschmar, H.: Block-nerf: Scalable large scene neural view synthesis (2022)" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 130, + 494, + 482, + 515 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 494, + 482, + 515 + ], + "spans": [ + { + "bbox": [ + 130, + 494, + 482, + 515 + ], + "type": "text", + "content": "26. Ververas, E., Potamias, R.A., Song, J., Deng, J., Zafeiriou, S.: Sags: Structure-aware 3d gaussian splatting. arXiv:2404.19149 (2024)" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 130, + 515, + 482, + 536 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 515, + 482, + 536 + ], + "spans": [ + { + "bbox": [ + 130, + 515, + 482, + 536 + ], + "type": "text", + "content": "27. Wang, F., Tan, S., Li, X., Tian, Z., Liu, H.: Mixed neural voxels for fast multi-view video synthesis. arXiv preprint arXiv:2212.00190 (2022)" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 130, + 536, + 482, + 568 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 536, + 482, + 568 + ], + "spans": [ + { + "bbox": [ + 130, + 536, + 482, + 568 + ], + "type": "text", + "content": "28. Wang, Z., Bovik, A.C., Sheikh, H.R., Simoncelli, E.P.: Image quality assessment: from error visibility to structural similarity. IEEE transactions on image processing 13(4), 600-612 (2004)" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 130, + 568, + 482, + 601 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 568, + 482, + 601 + ], + "spans": [ + { + "bbox": [ + 130, + 568, + 482, + 601 + ], + "type": "text", + "content": "29. Wu, G., Yi, T., Fang, J., Xie, L., Zhang, X., Wei, W., Liu, W., Tian, Q., Xinggang, W.: 4d gaussian splatting for real-time dynamic scene rendering. arXiv preprint arXiv:2310.08528 (2023)" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 130, + 601, + 482, + 633 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 601, + 482, + 633 + ], + "spans": [ + { + "bbox": [ + 130, + 601, + 482, + 633 + ], + "type": "text", + "content": "30. Yang, Z., Yang, H., Pan, Z., Zhang, L.: Real-time photorealistic dynamic scene representation and rendering with 4d gaussian splatting. In: International Conference on Learning Representations (ICLR) (2024)" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 130, + 633, + 482, + 665 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 633, + 482, + 665 + ], + "spans": [ + { + "bbox": [ + 130, + 633, + 482, + 665 + ], + "type": "text", + "content": "31. Yang, Z., Gao, X., Zhou, W., Jiao, S., Zhang, Y., Jin, X.: Deformable 3d gaussians for high-fidelity monocular dynamic scene reconstruction. arXiv preprint arXiv:2309.13101 (2023)" + } + ] + } + ], + "index": 19 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 405, + 91, + 447, + 100 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 405, + 91, + 447, + 100 + ], + "spans": [ + { + "bbox": [ + 405, + 91, + 447, + 100 + ], + "type": "text", + "content": "E-D3DGS" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 470, + 92, + 481, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 470, + 92, + 481, + 100 + ], + "spans": [ + { + "bbox": [ + 470, + 92, + 481, + 100 + ], + "type": "text", + "content": "15" + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 14 + }, + { + "para_blocks": [ + { + "bbox": [ + 132, + 116, + 480, + 270 + ], + "type": "list", + "angle": 0, + "index": 6, + "blocks": [ + { + "bbox": [ + 132, + 116, + 480, + 138 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 116, + 480, + 138 + ], + "spans": [ + { + "bbox": [ + 132, + 116, + 480, + 138 + ], + "type": "text", + "content": "32. Yu, H., Julin, J., Milacski, Z.A., Niinuma, K., Jeni, L.A.: Cogs: Controllable gaussian splatting (2023)" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 132, + 138, + 480, + 171 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 138, + 480, + 171 + ], + "spans": [ + { + "bbox": [ + 132, + 138, + 480, + 171 + ], + "type": "text", + "content": "33. Zhang, R., Isola, P., Efros, A.A., Shechtman, E., Wang, O.: The unreasonable effectiveness of deep features as a perceptual metric. In: Proceedings of the IEEE conference on computer vision and pattern recognition. pp. 586-595 (2018)" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 132, + 172, + 480, + 236 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 172, + 480, + 236 + ], + "spans": [ + { + "bbox": [ + 132, + 172, + 480, + 236 + ], + "type": "text", + "content": "34. Zhao, B., Li, Y., Sun, Z., Zeng, L., Shen, Y., Ma, R., Zhang, Y., Bao, H., Cui, Z.: Gaussian prediction: Dynamic 3d gaussian prediction for motion extrapolation and free view synthesis. In: Special Interest Group on Computer Graphics and Interactive Techniques Conference Papers aA24. SIGGRAPH aAZ24, ACM (Jul 2024). https://doi.org/10.1145/3641519.3657417, http://dx.doi.org/10.1145/3641519.3657417" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 132, + 237, + 480, + 270 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 237, + 480, + 270 + ], + "spans": [ + { + "bbox": [ + 132, + 237, + 480, + 270 + ], + "type": "text", + "content": "35. Zwicker, M., Pfister, H., Van Baar, J., Gross, M.: Surface splattering. In: Proceedings of the 28th annual conference on Computer graphics and interactive techniques. pp. 371-378 (2001)" + } + ] + } + ], + "index": 5 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "spans": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "type": "text", + "content": "16" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 166, + 91, + 217, + 100 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 166, + 91, + 217, + 100 + ], + "spans": [ + { + "bbox": [ + 166, + 91, + 217, + 100 + ], + "type": "text", + "content": "J. Bae et al." + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 15 + } + ], + "_backend": "vlm", + "_version_name": "2.6.4" +} \ No newline at end of file diff --git "a/2024/Perceptual Evaluation of Audio-Visual Synchrony Grounded in Viewers\342\200\231 Opinion Scores/59607ef8-05c8-452e-b964-a6375c1e0fbc_content_list.json" "b/2024/Perceptual Evaluation of Audio-Visual Synchrony Grounded in Viewers\342\200\231 Opinion Scores/59607ef8-05c8-452e-b964-a6375c1e0fbc_content_list.json" new file mode 100644 index 0000000000000000000000000000000000000000..fdf9cca12dec210e20d4c8e1d7ce61b81adc6775 --- /dev/null +++ "b/2024/Perceptual Evaluation of Audio-Visual Synchrony Grounded in Viewers\342\200\231 Opinion Scores/59607ef8-05c8-452e-b964-a6375c1e0fbc_content_list.json" @@ -0,0 +1,1711 @@ +[ + { + "type": "text", + "text": "Perceptual Evaluation of Audio-Visual Synchrony Grounded in Viewers' Opinion Scores", + "text_level": 1, + "bbox": [ + 218, + 140, + 782, + 186 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Lucas Goncalves\\*, Prashant Mathur?, Chandrashekhar Lavania?, Metehan Cekic?, Marcello Federico?, and Kyu J. Han?", + "bbox": [ + 228, + 210, + 774, + 244 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "1 The University of Texas at Dallas \n2 Amazon pramathu@amazon.com", + "bbox": [ + 379, + 253, + 625, + 297 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Abstract. Recent advancements in audio-visual generative modeling have been propelled by progress in deep learning and the availability of data-rich benchmarks. However, the growth is not attributed solely to models and benchmarks. Universally accepted evaluation metrics also play an important role in advancing the field. While there are many metrics available to evaluate audio and visual content separately, there is a lack of metrics that offer a quantitative and interpretable measure of audio-visual synchronization for videos 'in the wild'. To address this gap, we first created a large scale human annotated dataset $(100 + \\mathrm{hrs})$ representing nine types of synchronization errors in audio-visual content and how human perceive them. We then developed a PEAVS (Perceptual Evaluation of Audio-Visual Synchrony) score, a novel automatic metric with a 5-point scale that evaluates the quality of audio-visual synchronization. We validate PEAVS using a newly generated dataset, achieving a Pearson correlation of 0.79 at the set level and 0.54 at the clip level when compared to human labels. In our experiments, we observe a relative gain $50\\%$ over a natural extension of Fréchet based metrics for Audio-Visual synchrony, confirming PEAVS' efficacy in objectively modeling subjective perceptions of audio-visual synchronization for videos 'in the wild'.", + "bbox": [ + 261, + 332, + 743, + 609 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "1 Introduction", + "text_level": 1, + "bbox": [ + 215, + 633, + 375, + 648 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Audio-visual (AV) generative modeling has made rapid progress in recent years due to advancements in deep learning models [38] and the availability of large-scale datasets like VGGSound [9] and AudioSet [15]. However, evaluation of these models remains an open challenge. Existing automatic metrics often focus on specific aspects, such as image quality [19,37] or audio fidelity [10,23,32], lacking a satisfactory measure for assessing audio-visual synchronization of 'in the wild' videos. In fact, lack of well-established metrics for AV synchrony issues can already be observed in literature. For instance, recent works in the realm of audio-visual generation, like MM-Diffusion [33], had to solely rely on single-modality metrics to evaluate their outputs. The work on Diff-Foley [26] had", + "bbox": [ + 212, + 664, + 787, + 818 + ], + "page_idx": 0 + }, + { + "type": "page_footnote", + "text": "* Work done during internship at Amazon.", + "bbox": [ + 217, + 824, + 509, + 839 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "to train their own audio-visual synchrony classifier. More advanced evaluation metrics are needed to provide a holistic assessment of audio-visual coherence. The development of such metrics is essential to accurately judge model performance, identify failure cases, and drive further progress in this evolving field.", + "bbox": [ + 212, + 146, + 787, + 205 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Recent studies such as AV Synchrony Transformer (AVST) [8], SparseSync [21] and Diff-Foley [26] have defined the synchronization problem between audio-visual modalities primarily as an issue of temporal offset determination. However, this perspective addresses only one aspect of synchronization challenges. In our work, we extend the scope to a broader range of synchronization issues, encompassing audio/visual speed variations, intermittent muting, fragment shuffling, AV flickering, and temporal shifts. We also delve into developing a perceptual, automatic metric that aligns with human judgment. To achieve this, we initiated a large scale annotation study, gathering human assessments on audio-visual synchrony across over 100 hours of both original and distorted content, featuring the aforementioned synchronization issues.", + "bbox": [ + 212, + 207, + 789, + 372 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "In this endeavor, our objective is to create a metric that accomplishes the following: (a) attains a high correlation with human assessments of synchronization issues, (b) offers interpretable scoring for end users, and (c) operates in a reference-free manner, meaning it does not rely on ground truth for quality prediction. The key innovation in our approach lies in the development of an interpretable scale based on detailed perceptual guidelines provided to human raters. By optimizing for concordance with human-annotated scores during training, our metric produces scores directly aligned with predefined levels of audio-visual (a)synchrony. This emphasis on interpretability and alignment with human perception addresses the need for a more insightful evaluation of AV synchronization.", + "bbox": [ + 212, + 373, + 787, + 537 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Our main contributions are as follows:", + "bbox": [ + 215, + 539, + 493, + 553 + ], + "page_idx": 1 + }, + { + "type": "list", + "sub_type": "text", + "list_items": [ + "- We present a new Audio-Visual Synchrony human perception (AVS) benchmark data set with over 120K annotations and over $100+$ hrs of content. Each video in this benchmark is annotated by three different annotators and we see an agreement of 0.71 Krippendorff's alpha [25].", + "- We propose a new audio-visual synchrony evaluation metric - PEAVS - that is reference-free and has interpretable scoring on a scale of 1 to 5.", + "- PEAVS shows high correlation of 0.79 with human judgements on the benchmark data, significantly outperforming a Fréchet distance based AV synchrony metric." + ], + "bbox": [ + 225, + 561, + 785, + 696 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "2 Related Works", + "text_level": 1, + "bbox": [ + 215, + 719, + 397, + 734 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Unimodal Metrics: Several factors have fueled progress in single modality generation, including advancement in deep learning model architectures and training/optimization methodology, increased computational capabilities, a rich availability of multimodal data, and the development of consistent metrics. Established metrics like Inception Score (ISc) [34], Fréchet Inception Distance (FID) [20], Fréchet Audio Distance (FAD) [23], Fréchet Video Distance (FVD) [37],", + "bbox": [ + 212, + 750, + 787, + 843 + ], + "page_idx": 1 + }, + { + "type": "page_number", + "text": "2", + "bbox": [ + 217, + 114, + 228, + 126 + ], + "page_idx": 1 + }, + { + "type": "header", + "text": "Goncalves et al. 2024", + "bbox": [ + 271, + 114, + 413, + 127 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "inter alia, are frequently employed to gauge the quality of the content generated by unimodal generative models, providing a standard for comparison. However, these metrics are reliant only on one modality - either audio [10,23,27,31,32,35,39], video [5,19,24,37,42,44], or just applied to images [20,34]. In the evaluation of generated multimodal content, ensuring the quality of individual modalities is only half the battle. It is crucial to evaluate the synchronization between audio and visual elements, especially in the context of audio-visual generation.", + "bbox": [ + 212, + 146, + 787, + 253 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Multimodal Metrics: For AV synchrony evaluation, it is important to determine the misalignment between audio and visual inputs. In the past, researchers have explored handling synchrony by statistical approaches and specifically designed features to pinpoint where the sound comes from, like locating the source of a voice in a scene [17]. However, with the rise of deep learning recent studies are harnessing distinct connections between audio and visuals, like those in human speech [11], audio-visual correspondence to check for matching content [3], or musical performances [4]. Some recent research studies, such as those by Chen et al. [8] and Iashin et al. [21], propose methods to determine audio-visual synchronization in a natural diverse video environment. However, their focus is solely on tackling the synchronization problem arising from delays in audio or video streams, specifically addressing differences in timing (offsets). In these existing approaches, the model is trained with the objective of either 1) detecting synchronization errors as a binary classification or 2) predicting the precise offset values through regression. In contrast, our aim is to have the model replicate the way humans perceive synchronization issues across a range of distortion types (see more in Sec 3.1).", + "bbox": [ + 212, + 271, + 789, + 532 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Metrics for \"in the wild\" videos: The assessment of multimodal content has been extensively investigated within the computer vision community however most of the work has been tailored to lip-synchrony for instance, SyncNet [11] quantifies lip-sync errors by comparing mouth movements to the corresponding speech; AlignNet [40] focuses on aligning video content with ground truth audio, and predicts frame-wise offsets. In this work, we focus on evaluation of AV synchrony for \"in the wild\" videos which apart from Chen et al. [8], Iashin et al. [21] has been explored in Diff-Foley [26] where the authors train their own alignment classifier to detect synchrony. However, none of these works directly target modeling of human perception.", + "bbox": [ + 212, + 547, + 789, + 700 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Perceptual Metrics: Previous studies on audio metrics [28,29] demonstrated that a learned metric trained on data collected via crowd-sourced human annotations, particularly on just noticeably different samples, correlates well with mean opinion scores. For video modality, Wang et. al. [41] designed a perceptual metric — learnt over a large scale human annotated data — designed to distinguish user generated video content in different dimensions such as semantic content, technical quality, and compression level. Our work in contrast centers on the synchronization between audio and video than solely on video or audio quality.", + "bbox": [ + 212, + 718, + 787, + 843 + ], + "page_idx": 2 + }, + { + "type": "header", + "text": "Perceptual Evaluation of Audio-Visual Synchrony", + "bbox": [ + 398, + 114, + 730, + 130 + ], + "page_idx": 2 + }, + { + "type": "page_number", + "text": "3", + "bbox": [ + 774, + 116, + 785, + 126 + ], + "page_idx": 2 + }, + { + "type": "table", + "img_path": "images/851cf86a671a07346aa98e9ca3486f069c01e268ef2d78c923b3b4bae9788d58.jpg", + "table_caption": [], + "table_footnote": [], + "table_body": "
IDDistortionParameterLevels
1Temporal Shift Audioshift length (sec)-1, -.5, -.125, .045, .1, .125, .25, .5, 1, 2
2Audio Speed Change upspeed factor %.025, .05, .10, .15, .20, .25, .30, .4, .5, .75
3Video Speed Change upspeed factor %.025, .05, .10, .15, .20, .25, .30, .4, .5, .75
4Audio Speed Change downspeed factor %.025, .05, .10, .15, .20, .25, .30, .4, .5, .75
5Video Speed Change downspeed factor %.025, .05, .10, .15, .20, .25, .30, .4, .5, .75
6Intermittent Mutingduration (sec) of mute for every 1 sec.01, .025, .05, .1, .2, .3, .5, 1, 2.5, 4
7Randomly Sized Gaps Videogap duration (sec) and prob. 40%.01, .025, .05, .1, .2, .3, .5, 1, 2.5, 4
8Fragment Shufflingduration (sec) of segments to shuffle.3, .4, .5, 1, 1.5, 2, 2.5, 3, 3.5, 4
9AV Flickeringgap duration (sec) and prob. 40%.01, .025, .05, .1, .2, .3, .5, 1, 2.5, 4
", + "bbox": [ + 217, + 143, + 785, + 421 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Table 1: Summary of Distortions and their Parameters for Audio-Visual Content.", + "bbox": [ + 225, + 422, + 774, + 434 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "3 AVS Benchmark Dataset", + "text_level": 1, + "bbox": [ + 215, + 481, + 493, + 497 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "To develop an effective metric that aligns closely with human perception, it is important to have a dataset representative of human judgments. With this in mind, our study was designed to create a benchmark for our metric training and to assess its correlation with human perception, especially in the context of AV synchronization.", + "bbox": [ + 215, + 525, + 785, + 601 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "3.1 Data Preparation and Distortions", + "text_level": 1, + "bbox": [ + 215, + 636, + 539, + 651 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Audio-visual synchrony, the alignment of audio and visual elements, is important for an immersive multimedia experience. Even minor disruptions can notably degrade the user's experience, as witnessed in scenarios like watching a video where a mere mismatch in drums playing (visuals) and beats (audio) can become distracting. Various real-world challenges, from unreliable network connection to encoding mishaps, can lead to such anomalies. To produce realistic data representing such scenarios, we first collected 200 videos from the AudioSet corpus' evaluation split [15] ensuring that there is an aspect of synchrony in each of the samples. The samples selected are examples of \"in the wild\" videos, and does not include any samples with faces. We only include videos from actions such as: car driving by, dogs barking, instruments being played, manual labor being", + "bbox": [ + 215, + 672, + 787, + 840 + ], + "page_idx": 3 + }, + { + "type": "page_number", + "text": "4", + "bbox": [ + 217, + 114, + 228, + 126 + ], + "page_idx": 3 + }, + { + "type": "header", + "text": "Goncalves et al. 2024", + "bbox": [ + 271, + 114, + 413, + 126 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "done, etc. Each video then underwent nine synchrony related distortions at ten varying levels as shown in Table 1, totaling in 18,200 distorted videos.", + "bbox": [ + 212, + 145, + 782, + 176 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Following [37], we apply static and temporal noises in the AV content. Temporal noise is where the distortions in either or both modalities cause desynchronization in the AV content:", + "bbox": [ + 212, + 176, + 784, + 220 + ], + "page_idx": 4 + }, + { + "type": "list", + "sub_type": "text", + "list_items": [ + "- Temporal Misalignment: The audio track was offset, either forwards or backwards relative to the video [13], creating de-synchronizations spanning from -1 to 2 seconds.", + "- Audio Speed Change: Similar to FAD [23], the audio playback speed was manipulated, without corresponding changes in the video, resulting in gradual de-synchronization in the AV content. Depending on the speed change (up/down), we either clip the audio or visual frames to keep the same duration as the original video. In this setting, we ensure the sound characteristics is preserved by keeping essential features such as pitch and timbre while altering speed to minimize changes on sound texture.", + "- Video Speed Change: Similar to FVD [37], the speed of the video is adjusted while leaving the audio track untouched. This would cause the visuals to slowly move out of sync with the audio. As in the audio speed change, we clip the content to keep the same duration as the original one.", + "- Fragment Shuffling: Similar to the global swap distortion in FVD [37], both audio and video tracks were segmented and rearranged, preserving segment-level synchrony but disrupting overall synchronization of the video." + ], + "bbox": [ + 223, + 232, + 784, + 486 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Static noise is where the distortions do not necessarily change the synchrony in the AV content but it can be perceived as synchronization errors by humans:", + "bbox": [ + 212, + 500, + 782, + 530 + ], + "page_idx": 4 + }, + { + "type": "list", + "sub_type": "text", + "list_items": [ + "- Intermittent Muting: Intermittent periods of silences with varying duration is introduced in the audio track. This distortion disrupts the continuity, which can pose a challenge for some synchronization metrics. This distortion is similar to Pops as described in Kilgour et. al. [23].", + "- Randomly Sized Gaps: We randomly pick a frame with a certain probability and add black out visuals with varying number of frames in the video track. This distortion disrupts the continuity of the video. This distortion is similar to black rectangle noise as introduced in FVD.", + "- AV Flickering: Periods of silence and blackout were introduced in the audio-visual track. Although flickering by itself is not a synchronization issue, it can however be perceived as synchronization errors by humans and poses a challenging scenario for metrics [22]." + ], + "bbox": [ + 223, + 542, + 782, + 722 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "3.2 Data Annotation", + "text_level": 1, + "bbox": [ + 215, + 746, + 401, + 760 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Our aim in this task is to NOT evaluate the standalone quality of either audio or video. Instead, our focus is on synchrony issues and how humans perceive", + "bbox": [ + 212, + 772, + 782, + 801 + ], + "page_idx": 4 + }, + { + "type": "header", + "text": "Perceptual Evaluation of Audio-Visual Synchrony", + "bbox": [ + 398, + 114, + 732, + 130 + ], + "page_idx": 4 + }, + { + "type": "page_number", + "text": "5", + "bbox": [ + 774, + 116, + 785, + 126 + ], + "page_idx": 4 + }, + { + "type": "page_footnote", + "text": "3 Appendix B.4 contains a brief overview of top-30 label frequencies contained within the videos subset selected for our dataset.", + "bbox": [ + 217, + 808, + 782, + 837 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "them. Since pairwise stimulus experiments have shown to be faster than single-stimuli 12 and have been used in dubbing evaluation to foster relative measures in the rating 2, we also present our participants with two videos side-by-side. They are asked to play both videos, compare them, and rate each one on a scale from 1-5 based on the provided guidelines 4", + "bbox": [ + 212, + 146, + 787, + 223 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "From 18.2K videos, we randomly sampled (with replacement) 20K video pairs for comparison while taking into account the different levels of distortion (c.f. Table[1]). Participants conducted pairwise comparisons on these videos and each pair was rated at least three times by different annotators. This resulted in a total of 60,000 pairwise ratings or 120,000 ratings for individual videos. Each video duration being at least 10 seconds, this amounts to $100+$ hours of annotated videos with each video annotated at least 3 times. When looking at the inter-annotator agreement, we divided our data based on the severity of disagreement between the annotators. If all annotators annotated different scores those samples were put in \"disagreement\" group which constitute nearly $10\\%$ of the total annotated data. Within the disagreement group, we saw about $16\\%$ of cases where disagreements were equal or higher than 3, e.g., one video was rated as $1/4/5$ by three annotators. These high-disagreement samples were then sent back to the annotators for QA, after which we received their revised annotations. We are unable to directly calculate inter-annotator agreement (like Cohen's or Fleiss Kappa), as we had a large pool of annotators $(80+)$ and not all annotators annotated all samples. However, we do observe a Krippendorff's alpha value of 0.71 for the ratings assigned to all the samples, which corresponds to a moderate agreement [25].", + "bbox": [ + 212, + 227, + 787, + 501 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "After removing outliers and dedduplication we are left with 15.2K videos out of $40\\mathrm{K}$ We divide this 15.2K into 10.6K for training, 2.3K for development and 2.3K for evaluation. While creating these splits, we ensured that no original files and their derived (i.e., distorted) files are present in more than one set.", + "bbox": [ + 212, + 505, + 787, + 566 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "To gain insights into which distortion types posed the most perceptual challenges for humans in terms of synchronization, we examined videos that underwent direct comparisons and analyzed the variations in the annotated scores. Discrepancies in these scores (absolute values) offered insights into how differently these distortion types were perceived. Figure highlights that intermittent muting (type 6) is consistently the most noticeable distortion amongst all types (all high peaks). This observation aligns with the score distribution we observed in AVS benchmark (see Appendix Section B.5) and our manual analysis where audio muting was the most disruptive distortion and humans could easily differentiate audio muting from other distortions.", + "bbox": [ + 212, + 571, + 789, + 723 + ], + "page_idx": 5 + }, + { + "type": "page_number", + "text": "6", + "bbox": [ + 217, + 114, + 228, + 126 + ], + "page_idx": 5 + }, + { + "type": "header", + "text": "Goncalves et al. 2024", + "bbox": [ + 271, + 114, + 413, + 127 + ], + "page_idx": 5 + }, + { + "type": "page_footnote", + "text": "4 Annotation guidelines are shown in Appendix B.2. Upon publication, we will release the data set and code for creating distorted samples.", + "bbox": [ + 217, + 753, + 785, + 782 + ], + "page_idx": 5 + }, + { + "type": "page_footnote", + "text": "5 Since we randomly sampled with replacement, one video was annotated approx. 2.4 times by annotators, which led to duplication.", + "bbox": [ + 218, + 782, + 785, + 811 + ], + "page_idx": 5 + }, + { + "type": "page_footnote", + "text": "6 We leave out audio-shift distortion in this analysis as the levels are not linearly increasing or decreasing.", + "bbox": [ + 220, + 811, + 785, + 839 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/773fe6d4d007ca70d0ae7c1efa105f19eedcb71bbd0ab9f0b902f884a357b8c4.jpg", + "image_caption": [ + "Fig. 1: We compare absolute differences in annotation score across the distortion types. In this plot, x-axis shows distortion types that were compared, i.e., '2 4' represents distortion type 2 v/s 4 and y-axis represents the difference in scores across annotation tasks. For ID to distortion type mapping see Table1" + ], + "image_footnote": [], + "bbox": [ + 364, + 152, + 619, + 311 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "4 Preliminary Experiments", + "text_level": 1, + "bbox": [ + 215, + 407, + 496, + 426 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Studies by Unterthiner et al. [37] (Frechet Video Distance) and Kilgour et al. [23] (Frechet Audio Distance) have demonstrated the efficacy of embeddings from I3D [7] and VGGish [18] models in capturing qualitative and temporal characteristics from audio and video inputs, respectively. In the absence of dedicated metrics for assessing AV synchrony of 'in the wild' content and the lack of baseline models, we formulated a new Frechet audio-visual distance (FAVD) metric. [7] Additionally, we want to assess if the combination of embeddings from I3D and VGGish models can be useful for detecting synchrony issues.", + "bbox": [ + 212, + 445, + 787, + 568 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Frechet Distance metrics—FID, FAD, FVD, and now FAVD—compare the statistics of generated samples to those of the ground truth or a large collection of samples. In our preliminary experiment, we aimed to explore how FAD, FVD, and FAVD metrics respond to temporal discrepancies in the input data. We conducted our study on 18.2K videos, generating two sets of audio-visual embeddings: one from disrupted content and another representing aligned content (ground truth). For each type and level of distortion, we created a distinct “evaluation set” to analyze the impact of each specific distortion type and level on these metrics. In our experiments, we deliberately chose not to use FID (Frechet Inception Distance) due to its limitation in evaluating temporal aspects. FID is designed for static images and cannot adequately assess content with time-dependent elements, such as videos, where scene continuity and flow are essential for quality evaluation. By focusing on metrics capable of capturing these temporal dynamics, we aim to provide a more accurate and relevant evaluation of the audio-visual content's quality.", + "bbox": [ + 212, + 569, + 787, + 796 + ], + "page_idx": 6 + }, + { + "type": "header", + "text": "Perceptual Evaluation of Audio-Visual Synchrony", + "bbox": [ + 398, + 114, + 732, + 130 + ], + "page_idx": 6 + }, + { + "type": "page_number", + "text": "7", + "bbox": [ + 774, + 114, + 785, + 126 + ], + "page_idx": 6 + }, + { + "type": "page_footnote", + "text": "$^7$ SparseSync 21 and AVST 8 are only able to detect audio shift but our benchmark contains many more synchrony issues which render these metrics unusable.", + "bbox": [ + 217, + 808, + 785, + 840 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "We use I3D and VGGish to extract features from audio-visual content within the same time span (of 0.96 seconds - a default time span from VGGish).", + "bbox": [ + 212, + 146, + 782, + 176 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "These representations are concatenated time-wise and we calculate multivariate Gaussians for both evaluation set embeddings, $N_{e}(\\mu_{e}, \\Sigma_{e})$ , and the ground truth AV collection embeddings, $N_{av}(\\mu_{av}, \\Sigma_{av})$ .", + "bbox": [ + 212, + 176, + 782, + 222 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "We then compute the Fréchet distance between these two Gaussians, as described by Dowson et al. [14]:", + "bbox": [ + 212, + 222, + 782, + 255 + ], + "page_idx": 7 + }, + { + "type": "equation", + "text": "\n$$\nF \\left(N _ {a v}, N _ {e}\\right) = \\left| \\mu_ {a v} - \\mu_ {e} \\right| ^ {2} + t r \\left(\\Sigma_ {a v} + \\Sigma_ {e} - 2 \\left(\\Sigma_ {a v} \\Sigma_ {e}\\right) ^ {\\frac {1}{2}}\\right) \\tag {1}\n$$\n", + "text_format": "latex", + "bbox": [ + 300, + 263, + 785, + 284 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "where $tr$ is the trace of the matrix. For FAD and FVD, we compute the embeddings from respective modalities.", + "bbox": [ + 212, + 289, + 782, + 319 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/e24f3a313f158fa75cf27bfe3d6742143da122bc14727a4150345d9065c09640.jpg", + "image_caption": [ + "(a) Audio Shift" + ], + "image_footnote": [], + "bbox": [ + 218, + 345, + 398, + 449 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/19b53ca4a5b3a1f64d0729a1c841858a1f974fac72ac23f5297b831e879c774e.jpg", + "image_caption": [ + "(b) Audio Speed Up", + "Fig. 2: In these plots, we show the effect of distortions with varying levels on FAD, FVD, and FAVD for three distortion types. Flat trend line implies that a metric is not able to capture the distortion type. Increasing or decreasing trends show that a metric is susceptible to varying levels in a distortion. Distortion levels are taken from Table Detailed plots for all distortion types are available in Appendix C." + ], + "image_footnote": [], + "bbox": [ + 410, + 345, + 589, + 449 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/b787053c3e13812b46d4fac9920f659be36ebf9c11cf5f7972e046b478d3d119.jpg", + "image_caption": [ + "(c) Audio Speed Down" + ], + "image_footnote": [], + "bbox": [ + 602, + 345, + 781, + 449 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Examining the plots in Figure 2 it becomes evident that only FAVD (depicted in green) is responsive to various temporal distortions associated with AV synchrony across different levels. This is mainly due to use of covariance matrices $(\\Sigma_{av}, \\Sigma_e)$ from audio-visual features (shown in Eq 1) which helps FAVD track the interaction of AV modality. Additionally, in cases like audio shift (a) where neither FAD nor FVD can detect synchronization issues (indicated by flat trend lines), and audio speed-up where FAD shows minimal movement, FAVD formulation exhibits a more pronounced reaction, highlighting that the fusion of features from I3D and VGGish models effectively captures temporality in an audio-visual context. Based on these observations, we chose to use I3D and VGGish features as inputs for our metric training, with FAVD serving as a baseline metric for the remainder of our work.", + "bbox": [ + 212, + 575, + 787, + 756 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "5 PEAVS Score Formulation", + "text_level": 1, + "bbox": [ + 215, + 779, + 509, + 794 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "This section describes the architecture we have used to train our 'PEAVS' score metric and its training details.", + "bbox": [ + 212, + 809, + 782, + 839 + ], + "page_idx": 7 + }, + { + "type": "page_number", + "text": "8", + "bbox": [ + 217, + 114, + 228, + 126 + ], + "page_idx": 7 + }, + { + "type": "header", + "text": "Goncalves et al. 2024", + "bbox": [ + 271, + 114, + 415, + 127 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "5.1 Architecture", + "text_level": 1, + "bbox": [ + 215, + 146, + 367, + 159 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "The input to our proposed model are the video and audio streams, as seen in Figure 3. For feature extraction, we leverage I3D and VGGish. In addition to this, our design draws inspiration from the audio-visual framework presented in [16]. Central to our architecture is the utilization of the cross-modal transformer formulation [36], which excels in capturing temporal information and relationships between different modalities. To ensure temporal consistency across modalities, both audio and visual input features are extracted over matched time spans and are then relayed through cross-modal transformer layers.", + "bbox": [ + 212, + 169, + 787, + 290 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "There are several key components in the our architecture. First, the audio and visual inputs are processed through our frozen backbone feature extractors. Both features are extracted over 0.96 second windows. The I3D extracted inputs $x_{v}$ have dimensionality $x_{v} \\in \\mathbb{R}^{N_{av} \\times 1024}$ , where $N_{av}$ is the dimension of the audiovisual feature sequence, and 1,024 is the feature vector dimension. The acoustic feature vector $x_{a}$ extracted from the VGGish model is $x_{a} \\in \\mathbb{R}^{N_{av} \\times 128}$ , where $N_{av}$ is the dimension of the acoustic feature sequence, and 128 is the feature vector dimension. We ensure that all inputs to the cross-modal layers have same dimensionality, by projecting the visual features from 1,024 to 128 through the use of a 1D convolutional layer to produce $\\bar{x}_{v} \\in \\mathbb{R}^{N_{v} \\times 128}$ .", + "bbox": [ + 212, + 291, + 787, + 441 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "Next, we introduce positional embeddings to the feature vectors, following the step adopted in the original transformer paper [38]. The resultant features are used by the model to compute Query, Key, and Value vectors for each modality, as illustrated in Figure [3]. In our model, the Query vector from one modality is paired in the transformer layer with the Key and Value vectors of the other, ensuring a cross-modal interaction. This design choice allows the cross-modal transformer encoder layers for each modality branch to produce representations for a given modality that are influenced by the other modality. Following this, the output of the cross-modal layers are refined through the self-attention layers to enhance the cross-modal representations. The resultant features are then processed by averaging the outputs from the self-attention layers followed by concatenation of these averaged representations from both modality branches. Finally, a multilayer perceptron (MLP) with three fully-connected layers is integrated to produce a prediction score based on these concatenated features.", + "bbox": [ + 212, + 441, + 787, + 654 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "5.2 Training", + "text_level": 1, + "bbox": [ + 215, + 674, + 333, + 690 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "As seen in Figure 3 our training method consists of a two-stage strategy. In stage 1, we pre-train the model using a contrastive learning task, distinguishing between aligned and non-aligned pairs. Afterwards, in Stage Two, we fine-tune the model on our primary task, which involves predicting scores based on human annotators scores.", + "bbox": [ + 212, + 698, + 787, + 773 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "Stage 1 In stage 1, we pre-train our model using a contrastive learning objective. In this phase, we used AudioSet's Balanced train split [15], which after downloading all videos possible and removing corrupted ones resulted in 17K", + "bbox": [ + 212, + 794, + 787, + 840 + ], + "page_idx": 8 + }, + { + "type": "header", + "text": "Perceptual Evaluation of Audio-Visual Synchrony", + "bbox": [ + 398, + 114, + 730, + 130 + ], + "page_idx": 8 + }, + { + "type": "page_number", + "text": "9", + "bbox": [ + 774, + 116, + 785, + 126 + ], + "page_idx": 8 + }, + { + "type": "image", + "img_path": "images/a4001f608f38d265a3dcc211dedc86cad65cc81f032da7600e6c5e9c9af145ff.jpg", + "image_caption": [ + "Fig. 3: Framework Overview" + ], + "image_footnote": [], + "bbox": [ + 263, + 143, + 738, + 282 + ], + "page_idx": 9 + }, + { + "type": "text", + "text": "videos for our contrastive training stage We then subjected each of these files to ten levels of temporal misalignment (audio shift), as specified in section 3.1. This pre-processing resulted in a total of $\\sim 190\\mathrm{K}$ files for the contrastive learning phase. Similar to SyncNet [11], the training objective ensures that the outputs of both the audio and video branches are similar for aligned pairs and dissimilar for misaligned pairs. Specifically, we either minimize or maximize the Euclidean distance between the network outputs for these cases, respectively. Our objective consists of a contrastive loss (Equation 2). The primary intention behind this pre-training is to embed regularization into the audio and visual features extracted from our frozen backbone networks. This ensures that the model prioritizes temporal attributes of the features over their qualitative characteristics.", + "bbox": [ + 212, + 334, + 787, + 503 + ], + "page_idx": 9 + }, + { + "type": "equation", + "text": "\n$$\n\\mathcal {L} _ {c} = \\frac {1}{2 N} \\sum_ {n = 1} ^ {N} \\left(y _ {n}\\right) D _ {n} ^ {2} + \\left(1 - y _ {n}\\right) \\left(\\max \\left(0, m - D _ {n}\\right)\\right) ^ {2} \\tag {2}\n$$\n", + "text_format": "latex", + "bbox": [ + 318, + 513, + 784, + 554 + ], + "page_idx": 9 + }, + { + "type": "text", + "text": "where $y \\in \\{0,1\\}$ is the label (1 for similar pairs and 0 for dissimilar pairs), $D_w$ is the Euclidean distance between the two data points in the pair after passing them through the network, and $m$ is the margin.", + "bbox": [ + 212, + 556, + 787, + 603 + ], + "page_idx": 9 + }, + { + "type": "text", + "text": "Stage 2 In stage 2 the output modality heads of the pre-trained model are concatenated together to produce audio-visual representations. Three fully-connected layers are added on top of these concatenated heads. We then fine tune the entire model for the downstream task of predicting human-alignment scores. Since the aggregated human annotation scores are non-negative real numbers, we approach model training as a regression task and utilize concordance correlation coefficient (CCC) to measure the agreement between the true and predicted scores. The CCC measurement is illustrated in Equation 3", + "bbox": [ + 212, + 621, + 789, + 743 + ], + "page_idx": 9 + }, + { + "type": "equation", + "text": "\n$$\n\\mathcal {L} _ {C C C} = \\frac {2 \\rho \\sigma_ {x} \\sigma_ {y}}{\\sigma_ {x} ^ {2} + \\sigma_ {y} ^ {2} + (\\mu_ {x} - \\mu_ {y}) ^ {2}} \\tag {3}\n$$\n", + "text_format": "latex", + "bbox": [ + 388, + 748, + 784, + 781 + ], + "page_idx": 9 + }, + { + "type": "text", + "text": "where $\\mu_{x}$ and $\\mu_{y}$ denote the means of the true and predicted scores of a batch, respectively; $\\sigma_{x}$ and $\\sigma_{y}$ represent the standard deviations of the true and predicted", + "bbox": [ + 212, + 786, + 787, + 820 + ], + "page_idx": 9 + }, + { + "type": "page_number", + "text": "10", + "bbox": [ + 217, + 114, + 235, + 126 + ], + "page_idx": 9 + }, + { + "type": "header", + "text": "Goncalves et al. 2024", + "bbox": [ + 271, + 114, + 413, + 126 + ], + "page_idx": 9 + }, + { + "type": "page_footnote", + "text": "Note that this is different from AVS benchmark.", + "bbox": [ + 217, + 824, + 558, + 838 + ], + "page_idx": 9 + }, + { + "type": "text", + "text": "scores of the batch, respectively; and $\\rho$ is their Pearson's correlation coefficient. Our training objective is to maximize CCC, ensuring that the predicted scores closely correlate with the true scores.", + "bbox": [ + 212, + 146, + 787, + 191 + ], + "page_idx": 10 + }, + { + "type": "text", + "text": "Training Details In stage 1, we divided the video files obtained from AudioSet's Balanced train split [15] into an $80\\%$ training set and a $20\\%$ validation set. This $80/20$ split is carried out on the 17K original videos and extrapolated to distorted samples. We pre-trained the model in stage 1 until the validation loss plateaued. In stage 2, we use the train/validation/testing sets provided in the AVS benchmark corpus. We fine-tuned the model on training data for 20 epochs, saving the best-performing model based on the validation set performance [9].", + "bbox": [ + 212, + 214, + 787, + 321 + ], + "page_idx": 10 + }, + { + "type": "text", + "text": "6 Experiments", + "text_level": 1, + "bbox": [ + 214, + 344, + 375, + 363 + ], + "page_idx": 10 + }, + { + "type": "text", + "text": "In this section, we evaluate the performance of our trained model using human-evaluation scores. We compare the model's scores with those provided by human evaluators on the test split of AVS benchmark. To further assess the effectiveness of our cross-modal transformer architecture for metric training, we run two ablation experiments. First, to assess the impact of pre-training in stage 1, we directly train our model from scratch as in stage 2. This variant is termed \"Cross-Modal Base\" as it incorporates the same audiovisual cross-modal framework but excludes the contrastive learning pre-training phase. Secondly, to assess the importance of cross-modal framework, we test a version without cross-modal layers called \"Basic Transformer\". This version employs only the self-attention branches, differently from the cross-modal approach of sharing query vector from one modality to another.", + "bbox": [ + 212, + 377, + 787, + 559 + ], + "page_idx": 10 + }, + { + "type": "text", + "text": "6.1 Human Scores Correlation", + "text_level": 1, + "bbox": [ + 214, + 580, + 482, + 595 + ], + "page_idx": 10 + }, + { + "type": "text", + "text": "For the above models, we compute the correlation of their outputs with the human judgements using pearson correlation coefficient (PCC). Since Fréchet based metrics require a set of videos to calculate distance between multivariate Gaussians we compute set level correlations. As PEAVS and its variants produce a score for each video separately, we average the scores across the test set to compute set level scores[10] We also provide correlations at a much fine-grained clip level. Examining the results presented in Table 2 FAVD shows a relative improvement of $11\\%$ (from 0.475 to 0.527) and $64\\%$ (from 0.321 to 0.527) over FAD and FVD baselines setting a strong baseline for future work. This highlights FAVD's effectiveness in capturing AV synchrony issues. PEAVS model achieves a Pearson correlation of 0.794 at the Set Level, marking a substantial $51\\%$ improvement over the robust FAVD baseline. Notably, when we analyze the", + "bbox": [ + 212, + 607, + 787, + 789 + ], + "page_idx": 10 + }, + { + "type": "header", + "text": "Perceptual Evaluation of Audio-Visual Synchrony", + "bbox": [ + 398, + 114, + 730, + 130 + ], + "page_idx": 10 + }, + { + "type": "page_number", + "text": "11", + "bbox": [ + 767, + 116, + 782, + 126 + ], + "page_idx": 10 + }, + { + "type": "page_footnote", + "text": "For further details on the pre-training and fine-tuning settings, refer to appendix A.", + "bbox": [ + 214, + 796, + 785, + 811 + ], + "page_idx": 10 + }, + { + "type": "page_footnote", + "text": "10 Each set is defined as a collection of videos showing same distortion type and level. In total we had 90 sets, i.e. 9 distortions x 10 levels.", + "bbox": [ + 214, + 811, + 785, + 839 + ], + "page_idx": 10 + }, + { + "type": "image", + "img_path": "images/986ade816d78df75e02b6cfa4103d131b70c0a0bab1c49b40d8d1ba00b13873b.jpg", + "image_caption": [ + "Fig. 4: PEAVS results with Human Evaluation Scores compared side-by-side for each distortion type and at different levels. PEAVS scores are represented by blue bars and Human Evaluation scores are represented by orange bars. The x-axis are marked by the parameter and its levels as shown in Table $\\boxed{1}$ $\\rho$ is the correlation of PEAVS with human judgements." + ], + "image_footnote": [], + "bbox": [ + 271, + 143, + 733, + 356 + ], + "page_idx": 11 + }, + { + "type": "text", + "text": "impact of pre-training on PEAVS (c.f. Cross-Modal Base), performance drops by $7\\%$ & $11\\%$ at set and clip level, respectively, underscoring the significance of pre-training in PEAVS. Furthermore, removing the cross-modal layers (c.f. Base Transformer) results in a performance decrease of $9.4\\%$ & $19\\%$ at set and clip level, respectively, emphasizing the critical role of cross-modality layers.", + "bbox": [ + 212, + 462, + 787, + 540 + ], + "page_idx": 11 + }, + { + "type": "table", + "img_path": "images/10821540b8ee629036b50107e9f8469f426f2801e5d6a7ada44758151c815b92.jpg", + "table_caption": [], + "table_footnote": [], + "table_body": "
Model TypeSet LevelClip Level
FVD370.321-
FAD230.475-
FAVD (Ours)0.527-
Basic Transformer0.7200.434
Cross-Modal Base0.7360.477
PEAVS (Ours)0.7940.536
", + "bbox": [ + 357, + 563, + 643, + 666 + ], + "page_idx": 11 + }, + { + "type": "text", + "text": "Table 2: Correlation scores of different model types at Set and Clip levels.", + "bbox": [ + 250, + 665, + 750, + 679 + ], + "page_idx": 11 + }, + { + "type": "text", + "text": "6.2 Per Distortion Analysis", + "text_level": 1, + "bbox": [ + 214, + 741, + 455, + 756 + ], + "page_idx": 11 + }, + { + "type": "text", + "text": "In this section, we analyze the response of PEAVS to varying distortions at different levels. Figure 4 factorizes the results in Table 2 by each distortion type and level. PEAVS scores are shown side-by-side with the scores from human annotations on AVS benchmark test set. Generally, PEAVS metric aligns well with scores from human judgements across the board except in a few cases.", + "bbox": [ + 212, + 763, + 787, + 839 + ], + "page_idx": 11 + }, + { + "type": "page_number", + "text": "12", + "bbox": [ + 217, + 114, + 235, + 126 + ], + "page_idx": 11 + }, + { + "type": "header", + "text": "Goncalves et al. 2024", + "bbox": [ + 271, + 114, + 413, + 126 + ], + "page_idx": 11 + }, + { + "type": "text", + "text": "In top-right corner of Figure 4 we present the effects of audio shift. For ease of analysis, we display these results in terms of absolute shift values. We start with a shift of 0.045 based on a report from ITU which indicated that the threshold for audio-visual shift detection by human ranges from -125ms to +45ms. [13]. At this level, both the metric's output and human scores peak (>4), and they subsequently decrease as the shift increases.", + "bbox": [ + 212, + 145, + 787, + 236 + ], + "page_idx": 12 + }, + { + "type": "text", + "text": "In contrast to audio-shift, PEAVS shows the highest correlation for intermittent muting distortion which implies it is easier for the metric to capture this distortion. This confirms the notion that we observed in Section 3.2 where intermittent muting stood out as the most noticeable distortion and it was easier for humans to detect it as well.", + "bbox": [ + 212, + 238, + 787, + 311 + ], + "page_idx": 12 + }, + { + "type": "text", + "text": "In case of the fragment shuffling (bottom-right corner), initial levels of perturbation — shuffled fragments of 0.3-0.4 seconds — is significantly more disruptive for humans than the metric, i.e. the metric fails to capture this insight.", + "bbox": [ + 212, + 314, + 785, + 359 + ], + "page_idx": 12 + }, + { + "type": "table", + "img_path": "images/8ae076f098d0fbb44b62f391afa32ba6d1439b87d8db9097a459c6d0a94201e2.jpg", + "table_caption": [], + "table_footnote": [], + "table_body": "
Predicted
Pos.Neg.
ActualPos39227
Neg4130
", + "bbox": [ + 300, + 386, + 421, + 449 + ], + "page_idx": 12 + }, + { + "type": "table", + "img_path": "images/2a144fd31dfa35b7333f52e3ea15f81349ceda3227c900d35a45a2ef978ecbca.jpg", + "table_caption": [], + "table_footnote": [], + "table_body": "
Predicted
Pos.Neg.
ActualPos92174
Neg24110
", + "bbox": [ + 584, + 387, + 702, + 449 + ], + "page_idx": 12 + }, + { + "type": "text", + "text": "Table 3: Confusion matrix for SparseSync (left) and PEAVS (right). Accuracy of SparseSync = 42.3% v/s PEAVS = 50.5%", + "bbox": [ + 212, + 450, + 789, + 478 + ], + "page_idx": 12 + }, + { + "type": "text", + "text": "6.3 PEAVS vs. SparseSync", + "text_level": 1, + "bbox": [ + 214, + 512, + 454, + 527 + ], + "page_idx": 12 + }, + { + "type": "text", + "text": "In this section, we report an experiment comparing PEAVS with SparseSync, which was introduced by Iashin et al. [21] for detecting audio-visual shifts inside 'in the wild' videos. While SparseSync's goal is to objectively quantify specific synchronization issues, by estimating an audio-visual offset shift value, PEAVS evaluates synchronization under multiple facets and according to a perceptual scale. Due to their diverging objectives, a direct comparison is challenging. Moreover, SparseSync is trained on 21 classes representing distinct audio-shift levels from -2.0 sec (left) to 2.0 sec (right) with increments of 0.2, with 0.0 representing no audio shift; while PEAVS is trained to predict float values between 1 and 5.", + "bbox": [ + 212, + 541, + 787, + 676 + ], + "page_idx": 12 + }, + { + "type": "text", + "text": "In this experiment, we assess both model's performance in accurately classifying 400 videos, including 200 ground truth videos and 200 randomly audio-shifted versions of them. We consider distortions of 0.045, 0.1 and $\\pm 0.125$ as positive cases (i.e. ground truth) as ITU recommend an acceptability threshold for AV shift up to $185\\mathrm{ms}$ . This also aligns well with SparseSync buckets where first distortion starts at $\\pm 0.2$ seconds (i.e. SparseSync's step-size). To ensure a fair assessment, we divided the PEAVS scale into 21 bins, matching the output structure of SparseSync. For ground truth videos, we anticipate that", + "bbox": [ + 212, + 679, + 787, + 800 + ], + "page_idx": 12 + }, + { + "type": "header", + "text": "Perceptual Evaluation of Audio-Visual Synchrony", + "bbox": [ + 398, + 114, + 732, + 128 + ], + "page_idx": 12 + }, + { + "type": "page_number", + "text": "13", + "bbox": [ + 767, + 114, + 785, + 126 + ], + "page_idx": 12 + }, + { + "type": "page_footnote", + "text": "11 More details about this evaluation set are in Appendix Section D", + "bbox": [ + 212, + 809, + 669, + 827 + ], + "page_idx": 12 + }, + { + "type": "page_footnote", + "text": "12 We use increments of 0.238 $(= \\frac{5}{21})$ .", + "bbox": [ + 212, + 825, + 464, + 840 + ], + "page_idx": 12 + }, + { + "type": "text", + "text": "SparseSync will classify them as having zero shifts (0.0 class), while PEAVS is expected to assign them to the highest scoring bin (i.e. (4.76, 5]).", + "bbox": [ + 212, + 146, + 782, + 176 + ], + "page_idx": 13 + }, + { + "type": "text", + "text": "In Table3 we show the confusion matrices for SparseSync (left) and PEAVS (right). In terms of accuracy, PEAVS outperforms SparseSync by $19\\%$ , i.e. $50.5\\%$ vs. $42.3\\%$ . In particular, SparseSync shows an $18\\%$ edge in detecting distortions while PEAVS shows a $136\\%$ higher rate in detecting clean content.", + "bbox": [ + 212, + 176, + 784, + 236 + ], + "page_idx": 13 + }, + { + "type": "text", + "text": "7 Limitations", + "text_level": 1, + "bbox": [ + 215, + 258, + 362, + 273 + ], + "page_idx": 13 + }, + { + "type": "text", + "text": "Like any research, our work is not without its limitations, and we have identified a couple of them here:", + "bbox": [ + 212, + 289, + 782, + 318 + ], + "page_idx": 13 + }, + { + "type": "text", + "text": "Dataset: Due to budget constraints (30K USD), we could only source 200 videos and conduct 60K pair-wise annotations. In selecting these 200 videos, we ensured diversity. Our dataset comprises 215 labels out of the 512 present in AudioSet, representing only $40\\%$ of labels. Expanding this dataset for broader video coverage is left for future work.", + "bbox": [ + 212, + 320, + 784, + 393 + ], + "page_idx": 13 + }, + { + "type": "text", + "text": "Metric: The limited size of the AVS benchmark highlights the necessity of pretraining in model-based metrics (in our case, PEAVS). This is further evidenced by the results in Table 2 where pre-training improves correlation from 0.736 (Cross-Modal Base) to 0.794 (PEAVS). In this study, we only pre-trained PEAVS on audio-shift noise, leaving pre-training on other noise types for future work.", + "bbox": [ + 212, + 395, + 784, + 469 + ], + "page_idx": 13 + }, + { + "type": "text", + "text": "\"In the wild\" videos: Due to proprietary issues, we could not experiment with datasets featuring talking faces like VoxCeleb 30 or Lip Reading Sentences [1]. Our study is confined to addressing synchrony issues for \"in the wild\" videos.", + "bbox": [ + 212, + 470, + 782, + 516 + ], + "page_idx": 13 + }, + { + "type": "text", + "text": "8 Conclusions", + "text_level": 1, + "bbox": [ + 215, + 537, + 367, + 553 + ], + "page_idx": 13 + }, + { + "type": "text", + "text": "We introduced PEAVS, a novel metric designed for measuring synchrony in audio-visual content, an important advancement in the field of evaluation of audio-visual content. While the domain of audio-visual generative modeling is ever-evolving, assessing synchronization remains crucial for enhancing user experience. Although numerous metrics cater to either audio or visual content separately, a noticeable gap exists in evaluating their synchronization with a focus on viewers opinions. Our study addresses this gap with PEAVS, a metric based on a vast human-annotated dataset. The strong correlation between PEAVS scores and human evaluations confirms its efficacy in capturing perceptions surrounding audio-visual synchronization. As the scope of audio-visual content generation expands, tools like PEAVS are essential in maintaining a balanced interplay between sight and sound, guiding both creators and researchers.", + "bbox": [ + 212, + 566, + 784, + 748 + ], + "page_idx": 13 + }, + { + "type": "text", + "text": "Furthermore, we introduced the Audio-Visual Synchrony human perception benchmark, which provides a new dataset detailing human-perceptual scores on various audio-visual synchrony challenges for real-world videos. By isolating specific challenges in a controlled audio-visual environment, we aim to simplify the task for researchers to propose, assess, and analyze potential solutions while maintaining agreement with human perception.", + "bbox": [ + 212, + 750, + 784, + 839 + ], + "page_idx": 13 + }, + { + "type": "page_number", + "text": "14", + "bbox": [ + 217, + 114, + 235, + 126 + ], + "page_idx": 13 + }, + { + "type": "header", + "text": "Goncalves et al. 2024", + "bbox": [ + 271, + 114, + 413, + 126 + ], + "page_idx": 13 + }, + { + "type": "text", + "text": "References", + "text_level": 1, + "bbox": [ + 215, + 143, + 323, + 159 + ], + "page_idx": 14 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "1. Afouras, T., Chung, J.S., Zisserman, A.: Lrs3-ted: a large-scale dataset for visual speech recognition (2018)", + "2. Agarwal, M., Agrawal, S., Anastasopoulos, A., Bentivogli, L., Bojar, O., Borg, C., Carpuat, M., Cattoni, R., Cettolo, M., Chen, M., Chen, W., Choukri, K., Chronopoulou, A., Currey, A., Declerck, T., Dong, Q., Duh, K., Esteve, Y., Federico, M., Gahbiche, S., Haddow, B., Hsu, B., Mon Htut, P., Inaguma, H., Javorsky, D., Judge, J., Kano, Y., Ko, T., Kumar, R., Li, P., Ma, X., Mathur, P., Matusov, E., McNamee, P., P. McCrae, J., Murray, K., Nadejde, M., Nakamura, S., Negri, M., Nguyen, H., Niehues, J., Niu, X., Kr. Ojha, A., E. Ortega, J., Pal, P., Pino, J., van der Plas, L., Polak, P., Ripbeth, E., Salesky, E., Shi, J., Sperber, M., Stüker, S., Sudoh, K., Tang, Y., Thompson, B., Tran, K., Turchi, M., Waibel, A., Wang, M., Watanabe, S., Zevallos, R.: FINDINGS OF THE IWSLT 2023 EVALUATION CAMPAIGN. In: Proceedings of the 20th International Conference on Spoken Language Translation (IWSLT 2023). pp. 1-61. Association for Computational Linguistics, Toronto, Canada (in-person and online) (Jul 2023). https://doi.org/10.18653/v1/2023.iwslt-1.1 https://aclanthology.org/2023.iwslt-1.1", + "3. Arandjelović, R., Zisserman, A.: Look, listen and learn. 2017 IEEE International Conference on Computer Vision (ICCV) pp. 609-617 (2017), https://apisemantic scholar.org/CorpusID:10769575", + "4. Arandjelović, R., Zisserman, A.: Objects that sound. In: Computer Vision - ECCV 2018. pp. 451-466. Springer International Publishing, Cham (2018)", + "5. Binkowski, M., Sutherland, D.J., Arbel, M., Gretton, A.: Demystifying MMD gans. In: 6th International Conference on Learning Representations, ICLR 2018, Vancouver, BC, Canada, April 30 - May 3, 2018, Conference Track Proceedings. Open-Review.net (2018), https://openreview.net/forum?id=r1lUOzWCw", + "6. BT.1359, R.I.R.: BT.1359 - Relative timing of sound and vision for broadcasting (1998)", + "7. Carreira, J., Zisserman, A.: Quo vadis, action recognition? a new model and the kinetics dataset. In: 2017 IEEE Conference on Computer Vision and Pattern Recognition (CVPR). pp. 4724-4733. IEEE Computer Society, Los Alamitos, CA, USA (jul 2017). https://doi.org/10.1109/CVPR.2017.502 https://doi.ieeecomputersociety.org/10.1109/CVPR.2017.502", + "8. Chen, H., Xie, W., Afouras, T., Nagrani, A., Vedaldi, A., Zisserman, A.: Audiovisual synchronisation in the wild (2021)", + "9. Chen, H., Xie, W., Vedaldi, A., Zisserman, A.: Vggsound: A large-scale audio-visual dataset (2020)", + "0. Chinen, M., Lim, F.S.C., Skoglund, J., Gureev, N., O'Gorman, F., Hines, A.: Visqol v3: An open source production ready objective speech and audio metric. In: 2020 Twelfth International Conference on Quality of Multimedia Experience (QoMEX). pp. 1-6 (2020). https://doi.org/10.1109/QoMEX48832.2020.9123150", + "1. Chung, J.S., Zisserman, A.: Out of time: automated lip sync in the wild. In: Workshop on Multi-view Lip-reading, ACCV (2016)", + "12. Clark, A.P., Howard, K.L., Woods, A.T., Penton-Voak, I.S., Neumann, C.: Why rate when you could compare? Using the \"EloChoice\" package to assess pairwise comparisons of perceived physical strength. PLOS ONE 13(1), e0190393 (Jan 2018). https://doi.org/10.1371/journal.pone.0190393 https://dx.plos.org/10.1371/journal.pone.0190393" + ], + "bbox": [ + 218, + 175, + 785, + 840 + ], + "page_idx": 14 + }, + { + "type": "header", + "text": "Perceptual Evaluation of Audio-Visual Synchrony", + "bbox": [ + 398, + 114, + 730, + 128 + ], + "page_idx": 14 + }, + { + "type": "page_number", + "text": "15", + "bbox": [ + 767, + 116, + 785, + 126 + ], + "page_idx": 14 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "13. Committee, A.T.S.: Atsc implementation subcommittee finding: Relative timing of sound and vision for broadcast operations. Tech. rep., Advanced Television Systems Committee, 1750 K Street, N.W., Suite 1200, Washington, D.C. 20006 (6 2003), doc. IS-191", + "14. Dowson, D., Landau, B.: The frechet distance between multivariate normal distributions. Journal of Multivariate Analysis 12(3), 450-455 (1982). https://doi.org/https://doi.org/10.1016/0047-259X(82)90077-X, https://www.sciencedirect.com/science/article/pii/0047259X8290077X", + "15. Gemmeke, J.F., Ellis, D.P.W., Freedman, D., Jansen, A., Lawrence, W., Moore, R.C., Plakal, M., Ritter, M.: Audio set: An ontology and human-labeled dataset for audio events. In: 2017 IEEE International Conference on Acoustics, Speech and Signal Processing (ICASSP). pp. 776-780 (2017). https://doi.org/10.1109/ICASSP.2017.7952261", + "16. Goncalves, L., Busso, C.: Auxformer: Robust approach to audiovisual emotion recognition. In: ICASSP 2022 - 2022 IEEE International Conference on Acoustics, Speech and Signal Processing (ICASSP). pp. 7357-7361 (2022). https://doi.org/10.1109/ICASSP43922.2022.9747157", + "17. Hershey, J., Movellan, J.: Audio vision: Using audio-visual synchrony to locate sounds. In: Solla, S., Leen, T., Müller, K. (eds.) Advances in Neural Information Processing Systems. vol. 12. MIT Press (1999), https://proceedings.neurips.cc/paper_files/paper/1999/file/b618c3210e934362ac261db280128c22-Paper.pdf", + "18. Hershey, S., Chaudhuri, S., Ellis, D.P.W., Gemmeke, J.F., Jansen, A., Moore, R.C., Plakal, M., Platt, D., Saurous, R.A., Seybold, B., Slaney, M., Weiss, R.J., Wilson, K.: Cnn architectures for large-scale audio classification (2017)", + "19. Hessel, J., Holtzman, A., Forbes, M., Bras, R.L., Choi, Y.: CLIPScore: a reference-free evaluation metric for image captioning. In: EMNLP (2021)", + "20. Heusel, M., Ramsauer, H., Unterthiner, T., Nessler, B., Hochreiter, S.: Gans trained by a two time-scale update rule converge to a local nash equilibrium. In: Neural Information Processing Systems (2017)", + "21. Iashin, V., Xie, W., Rahtu, E., Zisserman, A.: Sparse in space and time: Audiovisual synchronisation with trainable selectors. In: British Machine Vision Conference (BMVC) (2022)", + "22. ITU-T RECOMMENDATION, P.: Subjective video quality assessment methods for multimedia applications (1999)", + "23. Kilgour, K., Zuluaga, M., Roblek, D., Sharifi, M.: Fréchet audio distance: A reference-free metric for evaluating music enhancement algorithms. In: Interspeech (2019), https://api-semanticscholar.org/CorpusID:202725406", + "24. Korhonen, J., You, J.: Korhonen. In: 2012 Fourth International Workshop on Quality of Multimedia Experience. pp. 37-38 (2012). https://doi.org/10.1109/QoMEX.2012.6263880", + "25. Krippendorff, K.: Content Analysis: An Introduction to Its Methodology (second edition). Sage Publications (2004)", + "26. Luo, S., Yan, C., Hu, C., Zhao, H.: Diff-foley: Synchronized video-to-audio synthesis with latent diffusion models (2023)", + "27. Luo, Y., Mesgarani, N.: Tasnet: Time-domain audio separation network for real-time, single-channel speech separation. In: 2018 IEEE International Conference on Acoustics, Speech and Signal Processing (ICASSP). pp. 696-700 (2018). https://doi.org/10.1109/ICASSP.2018.8462116" + ], + "bbox": [ + 217, + 146, + 785, + 840 + ], + "page_idx": 15 + }, + { + "type": "page_number", + "text": "16", + "bbox": [ + 217, + 114, + 235, + 126 + ], + "page_idx": 15 + }, + { + "type": "header", + "text": "Goncalves et al. 2024", + "bbox": [ + 271, + 114, + 413, + 126 + ], + "page_idx": 15 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "28. Manocha, P., Finkelstein, A., Zhang, R., Bryan, N.J., Mysore, G.J., Jin, Z.: A Differentiable Perceptual Audio Metric Learned from Just Noticeable Differences. In: Interspeech 2020. pp. 2852-2856. ISCA (Oct 2020). https://doi.org/10.21437/Interspeech.2020-1191 https://www.isca-speech.org/archive/interspeech_2020/manocha20_interspeech.html", + "29. Manocha, P., Kumar, A., Xu, B., Menon, A., Gebru, I.D., Ithapu, V.K., Calamia, P.: Dplm: A deep perceptual spatial-audio localization metric. In: 2021 IEEE Workshop on Applications of Signal Processing to Audio and Acoustics (WASPAA). pp. 6-10 (2021). https://doi.org/10.1109/WASPAA52581.2021.9632781", + "30. Nagrani, A., Chung, J.S., Zisserman, A.: Voxceleb: a large-scale speaker identification dataset. In: INTERSPEECH (2017)", + "31. Rix, A., Beerends, J., Hollier, M., Hekstra, A.: Perceptual evaluation of speech quality (pesq)-a new method for speech quality assessment of telephone networks and codescs. In: 2001 IEEE International Conference on Acoustics, Speech, and Signal Processing. Proceedings (Cat. No.01CH37221). vol. 2, pp. 749-752 vol.2 (2001). https://doi.org/10.1109/ICASSP.2001.941023", + "32. Roux, J.L., Wisdom, S., Erdogan, H., Hershey, J.R.: Sdr - half-baked or well done? In: ICASSP 2019 - 2019 IEEE International Conference on Acoustics, Speech and Signal Processing (ICASSP). pp. 626-630 (2019). https://doi.org/10.1109/ICASSP.2019.8683855", + "33. Ruan, L., Ma, Y., Yang, H., He, H., Liu, B., Fu, J., Yuan, N., Jin, Q., Guo, B.: Mm-diffusion: Learning multi-modal diffusion models for joint audio and video generation. In: 2023 IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR). pp. 10219-10228. IEEE Computer Society, Los Alamitos, CA, USA (jun 2023). https://doi.org/10.1109/CVPR52729.2023.00985 https://doi.ieeeccomputersociety.org/10.1109/CVPR52729.2023.00985", + "34. Salimans, T., Goodfellow, I., Zaremba, W., Cheung, V., Radford, A., Chen, X.: Improved techniques for training gans. In: Proceedings of the 30th International Conference on Neural Information Processing Systems. p. 2234-2242. NIPS'16, Curran Associates Inc., Red Hook, NY, USA (2016)", + "35. Taal, C.H., Hendriks, R.C., Heusdens, R., Jensen, J.: An algorithm for intelligibility prediction of time-frequency weighted noisy speech. IEEE Transactions on Audio, Speech, and Language Processing 19(7), 2125-2136 (2011). https://doi.org/10.1109/TASL.2011.2114881", + "36. Tsai, Y.H.H., Bai, S., Liang, P.P., Kolter, J.Z., Morency, L.P., Salakhutdinov, R.: Multimodal transformer for unaligned multimodal language sequences. In: Proceedings of the 57th Annual Meeting of the Association for Computational Linguistics (Volume 1: Long Papers). Association for Computational Linguistics, Florence, Italy (7 2019)", + "37. Unterthiner, T., van Steenkiste, S., Kurach, K., Marinier, R., Michalski, M., Gelly, S.: Towards accurate generative models of video: A new metric & challenges (2019)", + "38. Vaswani, A., Shazeer, N., Parmar, N., Uszkoreit, J., Jones, L., Gomez, A.N., Kaiser, L.u., Polosukhin, I.: Attention is all you need. In: Advances in Neural Information Processing Systems. vol. 30. Curran Associates, Inc. (2017), https://proceedings.neurips.cc/paper_files/paper/2017/file/3f5ee243547dee91fbd053c1c4a845aa-Paper.pdf", + "39. Vincent, E., Gribonval, R., Fevotte, C.: Performance measurement in blind audio source separation. IEEE Transactions on Audio, Speech, and Language Processing 14(4), 1462-1469 (2006). https://doi.org/10.1109/TSA.2005.858005", + "40. Wang, J., Fang, Z., Zhao, H.: Alignnet: A unifying approach to audio-visual alignment. In: WACV (2020)" + ], + "bbox": [ + 215, + 146, + 787, + 839 + ], + "page_idx": 16 + }, + { + "type": "header", + "text": "Perceptual Evaluation of Audio-Visual Synchrony", + "bbox": [ + 398, + 114, + 732, + 128 + ], + "page_idx": 16 + }, + { + "type": "page_number", + "text": "17", + "bbox": [ + 767, + 116, + 785, + 126 + ], + "page_idx": 16 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "41. Wang, Y., Ke, J., Talebi, H., Yim, J.G., Birkbeck, N., Adsumilli, B., Milanfar, P., Yang, F.: Rich features for perceptual quality assessment ofUGC videos. In: 2021 IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR). pp. 13430-13439 (2021). https://doi.org/10.1109/CVPR46437.2021.01323", + "42. Wang, Z., Simoncelli, E., Bovik, A.: Multiscale structural similarity for image quality assessment. In: The Thirty-Seventh Asilomar Conference on Signals, Systems & Computers, 2003. vol. 2, pp. 1398-1402 Vol.2 (2003). https://doi.org/10.1109/ACSSC.2003.1292216", + "43. Wang, Z., Bovik, A.: A universal image quality index. IEEE Signal Processing Letters 9(3), 81-84 (2002). https://doi.org/10.1109/97.995823", + "44. Zhang, R., Isola, P., Efros, A.A., Shechtman, E., Wang, O.: The unreasonable effectiveness of deep features as a perceptual metric. In: 2018 IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 586-595 (2018). https://doi.org/10.1109/CVPR.2018.00068" + ], + "bbox": [ + 215, + 146, + 785, + 342 + ], + "page_idx": 17 + }, + { + "type": "page_number", + "text": "18", + "bbox": [ + 217, + 114, + 235, + 126 + ], + "page_idx": 17 + }, + { + "type": "header", + "text": "Goncalves et al. 2024", + "bbox": [ + 271, + 114, + 413, + 126 + ], + "page_idx": 17 + } +] \ No newline at end of file diff --git "a/2024/Perceptual Evaluation of Audio-Visual Synchrony Grounded in Viewers\342\200\231 Opinion Scores/59607ef8-05c8-452e-b964-a6375c1e0fbc_model.json" "b/2024/Perceptual Evaluation of Audio-Visual Synchrony Grounded in Viewers\342\200\231 Opinion Scores/59607ef8-05c8-452e-b964-a6375c1e0fbc_model.json" new file mode 100644 index 0000000000000000000000000000000000000000..63bd269b417f94f8e32b6c0f9b94a7ad0799e598 --- /dev/null +++ "b/2024/Perceptual Evaluation of Audio-Visual Synchrony Grounded in Viewers\342\200\231 Opinion Scores/59607ef8-05c8-452e-b964-a6375c1e0fbc_model.json" @@ -0,0 +1,2326 @@ +[ + [ + { + "type": "title", + "bbox": [ + 0.219, + 0.141, + 0.784, + 0.187 + ], + "angle": 0, + "content": "Perceptual Evaluation of Audio-Visual Synchrony Grounded in Viewers' Opinion Scores" + }, + { + "type": "text", + "bbox": [ + 0.229, + 0.212, + 0.776, + 0.245 + ], + "angle": 0, + "content": "Lucas Goncalves\\*, Prashant Mathur?, Chandrashekhar Lavania?, Metehan Cekic?, Marcello Federico?, and Kyu J. Han?" + }, + { + "type": "text", + "bbox": [ + 0.38, + 0.255, + 0.626, + 0.298 + ], + "angle": 0, + "content": "1 The University of Texas at Dallas \n2 Amazon pramathu@amazon.com" + }, + { + "type": "text", + "bbox": [ + 0.262, + 0.333, + 0.744, + 0.61 + ], + "angle": 0, + "content": "Abstract. Recent advancements in audio-visual generative modeling have been propelled by progress in deep learning and the availability of data-rich benchmarks. However, the growth is not attributed solely to models and benchmarks. Universally accepted evaluation metrics also play an important role in advancing the field. While there are many metrics available to evaluate audio and visual content separately, there is a lack of metrics that offer a quantitative and interpretable measure of audio-visual synchronization for videos 'in the wild'. To address this gap, we first created a large scale human annotated dataset \\((100 + \\mathrm{hrs})\\) representing nine types of synchronization errors in audio-visual content and how human perceive them. We then developed a PEAVS (Perceptual Evaluation of Audio-Visual Synchrony) score, a novel automatic metric with a 5-point scale that evaluates the quality of audio-visual synchronization. We validate PEAVS using a newly generated dataset, achieving a Pearson correlation of 0.79 at the set level and 0.54 at the clip level when compared to human labels. In our experiments, we observe a relative gain \\(50\\%\\) over a natural extension of Fréchet based metrics for Audio-Visual synchrony, confirming PEAVS' efficacy in objectively modeling subjective perceptions of audio-visual synchronization for videos 'in the wild'." + }, + { + "type": "title", + "bbox": [ + 0.217, + 0.635, + 0.377, + 0.65 + ], + "angle": 0, + "content": "1 Introduction" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.665, + 0.789, + 0.819 + ], + "angle": 0, + "content": "Audio-visual (AV) generative modeling has made rapid progress in recent years due to advancements in deep learning models [38] and the availability of large-scale datasets like VGGSound [9] and AudioSet [15]. However, evaluation of these models remains an open challenge. Existing automatic metrics often focus on specific aspects, such as image quality [19,37] or audio fidelity [10,23,32], lacking a satisfactory measure for assessing audio-visual synchronization of 'in the wild' videos. In fact, lack of well-established metrics for AV synchrony issues can already be observed in literature. For instance, recent works in the realm of audio-visual generation, like MM-Diffusion [33], had to solely rely on single-modality metrics to evaluate their outputs. The work on Diff-Foley [26] had" + }, + { + "type": "page_footnote", + "bbox": [ + 0.218, + 0.825, + 0.51, + 0.84 + ], + "angle": 0, + "content": "* Work done during internship at Amazon." + } + ], + [ + { + "type": "page_number", + "bbox": [ + 0.218, + 0.116, + 0.23, + 0.127 + ], + "angle": 0, + "content": "2" + }, + { + "type": "header", + "bbox": [ + 0.272, + 0.115, + 0.415, + 0.128 + ], + "angle": 0, + "content": "Goncalves et al. 2024" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.147, + 0.788, + 0.207 + ], + "angle": 0, + "content": "to train their own audio-visual synchrony classifier. More advanced evaluation metrics are needed to provide a holistic assessment of audio-visual coherence. The development of such metrics is essential to accurately judge model performance, identify failure cases, and drive further progress in this evolving field." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.208, + 0.79, + 0.373 + ], + "angle": 0, + "content": "Recent studies such as AV Synchrony Transformer (AVST) [8], SparseSync [21] and Diff-Foley [26] have defined the synchronization problem between audio-visual modalities primarily as an issue of temporal offset determination. However, this perspective addresses only one aspect of synchronization challenges. In our work, we extend the scope to a broader range of synchronization issues, encompassing audio/visual speed variations, intermittent muting, fragment shuffling, AV flickering, and temporal shifts. We also delve into developing a perceptual, automatic metric that aligns with human judgment. To achieve this, we initiated a large scale annotation study, gathering human assessments on audio-visual synchrony across over 100 hours of both original and distorted content, featuring the aforementioned synchronization issues." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.374, + 0.789, + 0.539 + ], + "angle": 0, + "content": "In this endeavor, our objective is to create a metric that accomplishes the following: (a) attains a high correlation with human assessments of synchronization issues, (b) offers interpretable scoring for end users, and (c) operates in a reference-free manner, meaning it does not rely on ground truth for quality prediction. The key innovation in our approach lies in the development of an interpretable scale based on detailed perceptual guidelines provided to human raters. By optimizing for concordance with human-annotated scores during training, our metric produces scores directly aligned with predefined levels of audio-visual (a)synchrony. This emphasis on interpretability and alignment with human perception addresses the need for a more insightful evaluation of AV synchronization." + }, + { + "type": "text", + "bbox": [ + 0.216, + 0.54, + 0.495, + 0.554 + ], + "angle": 0, + "content": "Our main contributions are as follows:" + }, + { + "type": "text", + "bbox": [ + 0.226, + 0.563, + 0.787, + 0.625 + ], + "angle": 0, + "content": "- We present a new Audio-Visual Synchrony human perception (AVS) benchmark data set with over 120K annotations and over \\(100+\\) hrs of content. Each video in this benchmark is annotated by three different annotators and we see an agreement of 0.71 Krippendorff's alpha [25]." + }, + { + "type": "text", + "bbox": [ + 0.226, + 0.623, + 0.787, + 0.653 + ], + "angle": 0, + "content": "- We propose a new audio-visual synchrony evaluation metric - PEAVS - that is reference-free and has interpretable scoring on a scale of 1 to 5." + }, + { + "type": "text", + "bbox": [ + 0.226, + 0.653, + 0.787, + 0.698 + ], + "angle": 0, + "content": "- PEAVS shows high correlation of 0.79 with human judgements on the benchmark data, significantly outperforming a Fréchet distance based AV synchrony metric." + }, + { + "type": "list", + "bbox": [ + 0.226, + 0.563, + 0.787, + 0.698 + ], + "angle": 0, + "content": null + }, + { + "type": "title", + "bbox": [ + 0.216, + 0.72, + 0.398, + 0.736 + ], + "angle": 0, + "content": "2 Related Works" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.75, + 0.788, + 0.844 + ], + "angle": 0, + "content": "Unimodal Metrics: Several factors have fueled progress in single modality generation, including advancement in deep learning model architectures and training/optimization methodology, increased computational capabilities, a rich availability of multimodal data, and the development of consistent metrics. Established metrics like Inception Score (ISc) [34], Fréchet Inception Distance (FID) [20], Fréchet Audio Distance (FAD) [23], Fréchet Video Distance (FVD) [37]," + } + ], + [ + { + "type": "header", + "bbox": [ + 0.4, + 0.115, + 0.732, + 0.131 + ], + "angle": 0, + "content": "Perceptual Evaluation of Audio-Visual Synchrony" + }, + { + "type": "page_number", + "bbox": [ + 0.775, + 0.117, + 0.787, + 0.127 + ], + "angle": 0, + "content": "3" + }, + { + "type": "text", + "bbox": [ + 0.213, + 0.147, + 0.788, + 0.254 + ], + "angle": 0, + "content": "inter alia, are frequently employed to gauge the quality of the content generated by unimodal generative models, providing a standard for comparison. However, these metrics are reliant only on one modality - either audio [10,23,27,31,32,35,39], video [5,19,24,37,42,44], or just applied to images [20,34]. In the evaluation of generated multimodal content, ensuring the quality of individual modalities is only half the battle. It is crucial to evaluate the synchronization between audio and visual elements, especially in the context of audio-visual generation." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.272, + 0.79, + 0.533 + ], + "angle": 0, + "content": "Multimodal Metrics: For AV synchrony evaluation, it is important to determine the misalignment between audio and visual inputs. In the past, researchers have explored handling synchrony by statistical approaches and specifically designed features to pinpoint where the sound comes from, like locating the source of a voice in a scene [17]. However, with the rise of deep learning recent studies are harnessing distinct connections between audio and visuals, like those in human speech [11], audio-visual correspondence to check for matching content [3], or musical performances [4]. Some recent research studies, such as those by Chen et al. [8] and Iashin et al. [21], propose methods to determine audio-visual synchronization in a natural diverse video environment. However, their focus is solely on tackling the synchronization problem arising from delays in audio or video streams, specifically addressing differences in timing (offsets). In these existing approaches, the model is trained with the objective of either 1) detecting synchronization errors as a binary classification or 2) predicting the precise offset values through regression. In contrast, our aim is to have the model replicate the way humans perceive synchronization issues across a range of distortion types (see more in Sec 3.1)." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.548, + 0.79, + 0.702 + ], + "angle": 0, + "content": "Metrics for \"in the wild\" videos: The assessment of multimodal content has been extensively investigated within the computer vision community however most of the work has been tailored to lip-synchrony for instance, SyncNet [11] quantifies lip-sync errors by comparing mouth movements to the corresponding speech; AlignNet [40] focuses on aligning video content with ground truth audio, and predicts frame-wise offsets. In this work, we focus on evaluation of AV synchrony for \"in the wild\" videos which apart from Chen et al. [8], Iashin et al. [21] has been explored in Diff-Foley [26] where the authors train their own alignment classifier to detect synchrony. However, none of these works directly target modeling of human perception." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.719, + 0.788, + 0.844 + ], + "angle": 0, + "content": "Perceptual Metrics: Previous studies on audio metrics [28,29] demonstrated that a learned metric trained on data collected via crowd-sourced human annotations, particularly on just noticeably different samples, correlates well with mean opinion scores. For video modality, Wang et. al. [41] designed a perceptual metric — learnt over a large scale human annotated data — designed to distinguish user generated video content in different dimensions such as semantic content, technical quality, and compression level. Our work in contrast centers on the synchronization between audio and video than solely on video or audio quality." + } + ], + [ + { + "type": "page_number", + "bbox": [ + 0.218, + 0.116, + 0.23, + 0.127 + ], + "angle": 0, + "content": "4" + }, + { + "type": "header", + "bbox": [ + 0.272, + 0.115, + 0.415, + 0.127 + ], + "angle": 0, + "content": "Goncalves et al. 2024" + }, + { + "type": "table", + "bbox": [ + 0.218, + 0.145, + 0.787, + 0.422 + ], + "angle": 0, + "content": "
IDDistortionParameterLevels
1Temporal Shift Audioshift length (sec)-1, -.5, -.125, .045, .1, .125, .25, .5, 1, 2
2Audio Speed Change upspeed factor %.025, .05, .10, .15, .20, .25, .30, .4, .5, .75
3Video Speed Change upspeed factor %.025, .05, .10, .15, .20, .25, .30, .4, .5, .75
4Audio Speed Change downspeed factor %.025, .05, .10, .15, .20, .25, .30, .4, .5, .75
5Video Speed Change downspeed factor %.025, .05, .10, .15, .20, .25, .30, .4, .5, .75
6Intermittent Mutingduration (sec) of mute for every 1 sec.01, .025, .05, .1, .2, .3, .5, 1, 2.5, 4
7Randomly Sized Gaps Videogap duration (sec) and prob. 40%.01, .025, .05, .1, .2, .3, .5, 1, 2.5, 4
8Fragment Shufflingduration (sec) of segments to shuffle.3, .4, .5, 1, 1.5, 2, 2.5, 3, 3.5, 4
9AV Flickeringgap duration (sec) and prob. 40%.01, .025, .05, .1, .2, .3, .5, 1, 2.5, 4
" + }, + { + "type": "table_caption", + "bbox": [ + 0.226, + 0.423, + 0.776, + 0.435 + ], + "angle": 0, + "content": "Table 1: Summary of Distortions and their Parameters for Audio-Visual Content." + }, + { + "type": "title", + "bbox": [ + 0.217, + 0.482, + 0.494, + 0.498 + ], + "angle": 0, + "content": "3 AVS Benchmark Dataset" + }, + { + "type": "text", + "bbox": [ + 0.216, + 0.526, + 0.787, + 0.602 + ], + "angle": 0, + "content": "To develop an effective metric that aligns closely with human perception, it is important to have a dataset representative of human judgments. With this in mind, our study was designed to create a benchmark for our metric training and to assess its correlation with human perception, especially in the context of AV synchronization." + }, + { + "type": "title", + "bbox": [ + 0.217, + 0.637, + 0.54, + 0.652 + ], + "angle": 0, + "content": "3.1 Data Preparation and Distortions" + }, + { + "type": "text", + "bbox": [ + 0.216, + 0.674, + 0.788, + 0.841 + ], + "angle": 0, + "content": "Audio-visual synchrony, the alignment of audio and visual elements, is important for an immersive multimedia experience. Even minor disruptions can notably degrade the user's experience, as witnessed in scenarios like watching a video where a mere mismatch in drums playing (visuals) and beats (audio) can become distracting. Various real-world challenges, from unreliable network connection to encoding mishaps, can lead to such anomalies. To produce realistic data representing such scenarios, we first collected 200 videos from the AudioSet corpus' evaluation split [15] ensuring that there is an aspect of synchrony in each of the samples. The samples selected are examples of \"in the wild\" videos, and does not include any samples with faces. We only include videos from actions such as: car driving by, dogs barking, instruments being played, manual labor being" + } + ], + [ + { + "type": "header", + "bbox": [ + 0.4, + 0.115, + 0.733, + 0.131 + ], + "angle": 0, + "content": "Perceptual Evaluation of Audio-Visual Synchrony" + }, + { + "type": "page_number", + "bbox": [ + 0.775, + 0.117, + 0.787, + 0.127 + ], + "angle": 0, + "content": "5" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.146, + 0.784, + 0.177 + ], + "angle": 0, + "content": "done, etc. Each video then underwent nine synchrony related distortions at ten varying levels as shown in Table 1, totaling in 18,200 distorted videos." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.178, + 0.785, + 0.222 + ], + "angle": 0, + "content": "Following [37], we apply static and temporal noises in the AV content. Temporal noise is where the distortions in either or both modalities cause desynchronization in the AV content:" + }, + { + "type": "text", + "bbox": [ + 0.226, + 0.233, + 0.784, + 0.275 + ], + "angle": 0, + "content": "- Temporal Misalignment: The audio track was offset, either forwards or backwards relative to the video [13], creating de-synchronizations spanning from -1 to 2 seconds." + }, + { + "type": "text", + "bbox": [ + 0.225, + 0.278, + 0.785, + 0.381 + ], + "angle": 0, + "content": "- Audio Speed Change: Similar to FAD [23], the audio playback speed was manipulated, without corresponding changes in the video, resulting in gradual de-synchronization in the AV content. Depending on the speed change (up/down), we either clip the audio or visual frames to keep the same duration as the original video. In this setting, we ensure the sound characteristics is preserved by keeping essential features such as pitch and timbre while altering speed to minimize changes on sound texture." + }, + { + "type": "text", + "bbox": [ + 0.225, + 0.384, + 0.784, + 0.441 + ], + "angle": 0, + "content": "- Video Speed Change: Similar to FVD [37], the speed of the video is adjusted while leaving the audio track untouched. This would cause the visuals to slowly move out of sync with the audio. As in the audio speed change, we clip the content to keep the same duration as the original one." + }, + { + "type": "text", + "bbox": [ + 0.225, + 0.445, + 0.784, + 0.487 + ], + "angle": 0, + "content": "- Fragment Shuffling: Similar to the global swap distortion in FVD [37], both audio and video tracks were segmented and rearranged, preserving segment-level synchrony but disrupting overall synchronization of the video." + }, + { + "type": "list", + "bbox": [ + 0.225, + 0.233, + 0.785, + 0.487 + ], + "angle": 0, + "content": null + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.5, + 0.784, + 0.531 + ], + "angle": 0, + "content": "Static noise is where the distortions do not necessarily change the synchrony in the AV content but it can be perceived as synchronization errors by humans:" + }, + { + "type": "text", + "bbox": [ + 0.225, + 0.543, + 0.784, + 0.601 + ], + "angle": 0, + "content": "- Intermittent Muting: Intermittent periods of silences with varying duration is introduced in the audio track. This distortion disrupts the continuity, which can pose a challenge for some synchronization metrics. This distortion is similar to Pops as described in Kilgour et. al. [23]." + }, + { + "type": "text", + "bbox": [ + 0.225, + 0.604, + 0.784, + 0.66 + ], + "angle": 0, + "content": "- Randomly Sized Gaps: We randomly pick a frame with a certain probability and add black out visuals with varying number of frames in the video track. This distortion disrupts the continuity of the video. This distortion is similar to black rectangle noise as introduced in FVD." + }, + { + "type": "text", + "bbox": [ + 0.225, + 0.664, + 0.784, + 0.723 + ], + "angle": 0, + "content": "- AV Flickering: Periods of silence and blackout were introduced in the audio-visual track. Although flickering by itself is not a synchronization issue, it can however be perceived as synchronization errors by humans and poses a challenging scenario for metrics [22]." + }, + { + "type": "list", + "bbox": [ + 0.225, + 0.543, + 0.784, + 0.723 + ], + "angle": 0, + "content": null + }, + { + "type": "title", + "bbox": [ + 0.216, + 0.747, + 0.403, + 0.761 + ], + "angle": 0, + "content": "3.2 Data Annotation" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.773, + 0.784, + 0.803 + ], + "angle": 0, + "content": "Our aim in this task is to NOT evaluate the standalone quality of either audio or video. Instead, our focus is on synchrony issues and how humans perceive" + }, + { + "type": "page_footnote", + "bbox": [ + 0.218, + 0.809, + 0.784, + 0.838 + ], + "angle": 0, + "content": "3 Appendix B.4 contains a brief overview of top-30 label frequencies contained within the videos subset selected for our dataset." + } + ], + [ + { + "type": "page_number", + "bbox": [ + 0.218, + 0.116, + 0.23, + 0.127 + ], + "angle": 0, + "content": "6" + }, + { + "type": "header", + "bbox": [ + 0.272, + 0.115, + 0.415, + 0.128 + ], + "angle": 0, + "content": "Goncalves et al. 2024" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.147, + 0.788, + 0.224 + ], + "angle": 0, + "content": "them. Since pairwise stimulus experiments have shown to be faster than single-stimuli 12 and have been used in dubbing evaluation to foster relative measures in the rating 2, we also present our participants with two videos side-by-side. They are asked to play both videos, compare them, and rate each one on a scale from 1-5 based on the provided guidelines 4" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.228, + 0.789, + 0.502 + ], + "angle": 0, + "content": "From 18.2K videos, we randomly sampled (with replacement) 20K video pairs for comparison while taking into account the different levels of distortion (c.f. Table[1]). Participants conducted pairwise comparisons on these videos and each pair was rated at least three times by different annotators. This resulted in a total of 60,000 pairwise ratings or 120,000 ratings for individual videos. Each video duration being at least 10 seconds, this amounts to \\(100+\\) hours of annotated videos with each video annotated at least 3 times. When looking at the inter-annotator agreement, we divided our data based on the severity of disagreement between the annotators. If all annotators annotated different scores those samples were put in \"disagreement\" group which constitute nearly \\(10\\%\\) of the total annotated data. Within the disagreement group, we saw about \\(16\\%\\) of cases where disagreements were equal or higher than 3, e.g., one video was rated as \\(1/4/5\\) by three annotators. These high-disagreement samples were then sent back to the annotators for QA, after which we received their revised annotations. We are unable to directly calculate inter-annotator agreement (like Cohen's or Fleiss Kappa), as we had a large pool of annotators \\((80+)\\) and not all annotators annotated all samples. However, we do observe a Krippendorff's alpha value of 0.71 for the ratings assigned to all the samples, which corresponds to a moderate agreement [25]." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.506, + 0.789, + 0.567 + ], + "angle": 0, + "content": "After removing outliers and dedduplication we are left with 15.2K videos out of \\(40\\mathrm{K}\\) We divide this 15.2K into 10.6K for training, 2.3K for development and 2.3K for evaluation. While creating these splits, we ensured that no original files and their derived (i.e., distorted) files are present in more than one set." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.573, + 0.79, + 0.724 + ], + "angle": 0, + "content": "To gain insights into which distortion types posed the most perceptual challenges for humans in terms of synchronization, we examined videos that underwent direct comparisons and analyzed the variations in the annotated scores. Discrepancies in these scores (absolute values) offered insights into how differently these distortion types were perceived. Figure highlights that intermittent muting (type 6) is consistently the most noticeable distortion amongst all types (all high peaks). This observation aligns with the score distribution we observed in AVS benchmark (see Appendix Section B.5) and our manual analysis where audio muting was the most disruptive distortion and humans could easily differentiate audio muting from other distortions." + }, + { + "type": "page_footnote", + "bbox": [ + 0.218, + 0.755, + 0.787, + 0.784 + ], + "angle": 0, + "content": "4 Annotation guidelines are shown in Appendix B.2. Upon publication, we will release the data set and code for creating distorted samples." + }, + { + "type": "page_footnote", + "bbox": [ + 0.22, + 0.784, + 0.786, + 0.812 + ], + "angle": 0, + "content": "5 Since we randomly sampled with replacement, one video was annotated approx. 2.4 times by annotators, which led to duplication." + }, + { + "type": "page_footnote", + "bbox": [ + 0.222, + 0.812, + 0.786, + 0.84 + ], + "angle": 0, + "content": "6 We leave out audio-shift distortion in this analysis as the levels are not linearly increasing or decreasing." + }, + { + "type": "list", + "bbox": [ + 0.218, + 0.755, + 0.787, + 0.84 + ], + "angle": 0, + "content": null + } + ], + [ + { + "type": "header", + "bbox": [ + 0.4, + 0.115, + 0.733, + 0.131 + ], + "angle": 0, + "content": "Perceptual Evaluation of Audio-Visual Synchrony" + }, + { + "type": "page_number", + "bbox": [ + 0.775, + 0.116, + 0.787, + 0.127 + ], + "angle": 0, + "content": "7" + }, + { + "type": "image", + "bbox": [ + 0.365, + 0.154, + 0.62, + 0.312 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.214, + 0.321, + 0.788, + 0.381 + ], + "angle": 0, + "content": "Fig. 1: We compare absolute differences in annotation score across the distortion types. In this plot, x-axis shows distortion types that were compared, i.e., '2 4' represents distortion type 2 v/s 4 and y-axis represents the difference in scores across annotation tasks. For ID to distortion type mapping see Table1" + }, + { + "type": "title", + "bbox": [ + 0.216, + 0.409, + 0.498, + 0.427 + ], + "angle": 0, + "content": "4 Preliminary Experiments" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.446, + 0.789, + 0.569 + ], + "angle": 0, + "content": "Studies by Unterthiner et al. [37] (Frechet Video Distance) and Kilgour et al. [23] (Frechet Audio Distance) have demonstrated the efficacy of embeddings from I3D [7] and VGGish [18] models in capturing qualitative and temporal characteristics from audio and video inputs, respectively. In the absence of dedicated metrics for assessing AV synchrony of 'in the wild' content and the lack of baseline models, we formulated a new Frechet audio-visual distance (FAVD) metric. [7] Additionally, we want to assess if the combination of embeddings from I3D and VGGish models can be useful for detecting synchrony issues." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.57, + 0.789, + 0.797 + ], + "angle": 0, + "content": "Frechet Distance metrics—FID, FAD, FVD, and now FAVD—compare the statistics of generated samples to those of the ground truth or a large collection of samples. In our preliminary experiment, we aimed to explore how FAD, FVD, and FAVD metrics respond to temporal discrepancies in the input data. We conducted our study on 18.2K videos, generating two sets of audio-visual embeddings: one from disrupted content and another representing aligned content (ground truth). For each type and level of distortion, we created a distinct “evaluation set” to analyze the impact of each specific distortion type and level on these metrics. In our experiments, we deliberately chose not to use FID (Frechet Inception Distance) due to its limitation in evaluating temporal aspects. FID is designed for static images and cannot adequately assess content with time-dependent elements, such as videos, where scene continuity and flow are essential for quality evaluation. By focusing on metrics capable of capturing these temporal dynamics, we aim to provide a more accurate and relevant evaluation of the audio-visual content's quality." + }, + { + "type": "page_footnote", + "bbox": [ + 0.218, + 0.809, + 0.787, + 0.842 + ], + "angle": 0, + "content": "\\(^7\\) SparseSync 21 and AVST 8 are only able to detect audio shift but our benchmark contains many more synchrony issues which render these metrics unusable." + } + ], + [ + { + "type": "page_number", + "bbox": [ + 0.218, + 0.116, + 0.23, + 0.127 + ], + "angle": 0, + "content": "8" + }, + { + "type": "header", + "bbox": [ + 0.272, + 0.115, + 0.416, + 0.128 + ], + "angle": 0, + "content": "Goncalves et al. 2024" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.147, + 0.784, + 0.177 + ], + "angle": 0, + "content": "We use I3D and VGGish to extract features from audio-visual content within the same time span (of 0.96 seconds - a default time span from VGGish)." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.178, + 0.784, + 0.223 + ], + "angle": 0, + "content": "These representations are concatenated time-wise and we calculate multivariate Gaussians for both evaluation set embeddings, \\( N_{e}(\\mu_{e}, \\Sigma_{e}) \\), and the ground truth AV collection embeddings, \\( N_{av}(\\mu_{av}, \\Sigma_{av}) \\)." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.223, + 0.784, + 0.256 + ], + "angle": 0, + "content": "We then compute the Fréchet distance between these two Gaussians, as described by Dowson et al. [14]:" + }, + { + "type": "equation", + "bbox": [ + 0.301, + 0.265, + 0.786, + 0.285 + ], + "angle": 0, + "content": "\\[\nF \\left(N _ {a v}, N _ {e}\\right) = \\left| \\mu_ {a v} - \\mu_ {e} \\right| ^ {2} + t r \\left(\\Sigma_ {a v} + \\Sigma_ {e} - 2 \\left(\\Sigma_ {a v} \\Sigma_ {e}\\right) ^ {\\frac {1}{2}}\\right) \\tag {1}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.29, + 0.784, + 0.32 + ], + "angle": 0, + "content": "where \\( tr \\) is the trace of the matrix. For FAD and FVD, we compute the embeddings from respective modalities." + }, + { + "type": "image", + "bbox": [ + 0.219, + 0.347, + 0.399, + 0.45 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.264, + 0.453, + 0.356, + 0.465 + ], + "angle": 0, + "content": "(a) Audio Shift" + }, + { + "type": "image", + "bbox": [ + 0.411, + 0.347, + 0.591, + 0.45 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.442, + 0.453, + 0.562, + 0.466 + ], + "angle": 0, + "content": "(b) Audio Speed Up" + }, + { + "type": "image", + "bbox": [ + 0.604, + 0.347, + 0.782, + 0.45 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.627, + 0.453, + 0.761, + 0.466 + ], + "angle": 0, + "content": "(c) Audio Speed Down" + }, + { + "type": "image_caption", + "bbox": [ + 0.214, + 0.477, + 0.785, + 0.552 + ], + "angle": 0, + "content": "Fig. 2: In these plots, we show the effect of distortions with varying levels on FAD, FVD, and FAVD for three distortion types. Flat trend line implies that a metric is not able to capture the distortion type. Increasing or decreasing trends show that a metric is susceptible to varying levels in a distortion. Distortion levels are taken from Table Detailed plots for all distortion types are available in Appendix C." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.576, + 0.788, + 0.757 + ], + "angle": 0, + "content": "Examining the plots in Figure 2 it becomes evident that only FAVD (depicted in green) is responsive to various temporal distortions associated with AV synchrony across different levels. This is mainly due to use of covariance matrices \\((\\Sigma_{av}, \\Sigma_e)\\) from audio-visual features (shown in Eq 1) which helps FAVD track the interaction of AV modality. Additionally, in cases like audio shift (a) where neither FAD nor FVD can detect synchronization issues (indicated by flat trend lines), and audio speed-up where FAD shows minimal movement, FAVD formulation exhibits a more pronounced reaction, highlighting that the fusion of features from I3D and VGGish models effectively captures temporality in an audio-visual context. Based on these observations, we chose to use I3D and VGGish features as inputs for our metric training, with FAVD serving as a baseline metric for the remainder of our work." + }, + { + "type": "title", + "bbox": [ + 0.216, + 0.78, + 0.51, + 0.795 + ], + "angle": 0, + "content": "5 PEAVS Score Formulation" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.81, + 0.784, + 0.84 + ], + "angle": 0, + "content": "This section describes the architecture we have used to train our 'PEAVS' score metric and its training details." + } + ], + [ + { + "type": "header", + "bbox": [ + 0.4, + 0.115, + 0.732, + 0.131 + ], + "angle": 0, + "content": "Perceptual Evaluation of Audio-Visual Synchrony" + }, + { + "type": "page_number", + "bbox": [ + 0.775, + 0.117, + 0.787, + 0.127 + ], + "angle": 0, + "content": "9" + }, + { + "type": "title", + "bbox": [ + 0.216, + 0.147, + 0.368, + 0.16 + ], + "angle": 0, + "content": "5.1 Architecture" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.17, + 0.788, + 0.291 + ], + "angle": 0, + "content": "The input to our proposed model are the video and audio streams, as seen in Figure 3. For feature extraction, we leverage I3D and VGGish. In addition to this, our design draws inspiration from the audio-visual framework presented in [16]. Central to our architecture is the utilization of the cross-modal transformer formulation [36], which excels in capturing temporal information and relationships between different modalities. To ensure temporal consistency across modalities, both audio and visual input features are extracted over matched time spans and are then relayed through cross-modal transformer layers." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.292, + 0.789, + 0.442 + ], + "angle": 0, + "content": "There are several key components in the our architecture. First, the audio and visual inputs are processed through our frozen backbone feature extractors. Both features are extracted over 0.96 second windows. The I3D extracted inputs \\( x_{v} \\) have dimensionality \\( x_{v} \\in \\mathbb{R}^{N_{av} \\times 1024} \\), where \\( N_{av} \\) is the dimension of the audiovisual feature sequence, and 1,024 is the feature vector dimension. The acoustic feature vector \\( x_{a} \\) extracted from the VGGish model is \\( x_{a} \\in \\mathbb{R}^{N_{av} \\times 128} \\), where \\( N_{av} \\) is the dimension of the acoustic feature sequence, and 128 is the feature vector dimension. We ensure that all inputs to the cross-modal layers have same dimensionality, by projecting the visual features from 1,024 to 128 through the use of a 1D convolutional layer to produce \\( \\bar{x}_{v} \\in \\mathbb{R}^{N_{v} \\times 128} \\)." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.442, + 0.789, + 0.655 + ], + "angle": 0, + "content": "Next, we introduce positional embeddings to the feature vectors, following the step adopted in the original transformer paper [38]. The resultant features are used by the model to compute Query, Key, and Value vectors for each modality, as illustrated in Figure [3]. In our model, the Query vector from one modality is paired in the transformer layer with the Key and Value vectors of the other, ensuring a cross-modal interaction. This design choice allows the cross-modal transformer encoder layers for each modality branch to produce representations for a given modality that are influenced by the other modality. Following this, the output of the cross-modal layers are refined through the self-attention layers to enhance the cross-modal representations. The resultant features are then processed by averaging the outputs from the self-attention layers followed by concatenation of these averaged representations from both modality branches. Finally, a multilayer perceptron (MLP) with three fully-connected layers is integrated to produce a prediction score based on these concatenated features." + }, + { + "type": "title", + "bbox": [ + 0.216, + 0.675, + 0.334, + 0.691 + ], + "angle": 0, + "content": "5.2 Training" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.699, + 0.788, + 0.775 + ], + "angle": 0, + "content": "As seen in Figure 3 our training method consists of a two-stage strategy. In stage 1, we pre-train the model using a contrastive learning task, distinguishing between aligned and non-aligned pairs. Afterwards, in Stage Two, we fine-tune the model on our primary task, which involves predicting scores based on human annotators scores." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.795, + 0.789, + 0.842 + ], + "angle": 0, + "content": "Stage 1 In stage 1, we pre-train our model using a contrastive learning objective. In this phase, we used AudioSet's Balanced train split [15], which after downloading all videos possible and removing corrupted ones resulted in 17K" + } + ], + [ + { + "type": "page_number", + "bbox": [ + 0.218, + 0.116, + 0.236, + 0.127 + ], + "angle": 0, + "content": "10" + }, + { + "type": "header", + "bbox": [ + 0.272, + 0.115, + 0.415, + 0.127 + ], + "angle": 0, + "content": "Goncalves et al. 2024" + }, + { + "type": "image", + "bbox": [ + 0.264, + 0.144, + 0.74, + 0.283 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.403, + 0.296, + 0.6, + 0.31 + ], + "angle": 0, + "content": "Fig. 3: Framework Overview" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.335, + 0.788, + 0.504 + ], + "angle": 0, + "content": "videos for our contrastive training stage We then subjected each of these files to ten levels of temporal misalignment (audio shift), as specified in section 3.1. This pre-processing resulted in a total of \\(\\sim 190\\mathrm{K}\\) files for the contrastive learning phase. Similar to SyncNet [11], the training objective ensures that the outputs of both the audio and video branches are similar for aligned pairs and dissimilar for misaligned pairs. Specifically, we either minimize or maximize the Euclidean distance between the network outputs for these cases, respectively. Our objective consists of a contrastive loss (Equation 2). The primary intention behind this pre-training is to embed regularization into the audio and visual features extracted from our frozen backbone networks. This ensures that the model prioritizes temporal attributes of the features over their qualitative characteristics." + }, + { + "type": "equation", + "bbox": [ + 0.319, + 0.514, + 0.785, + 0.555 + ], + "angle": 0, + "content": "\\[\n\\mathcal {L} _ {c} = \\frac {1}{2 N} \\sum_ {n = 1} ^ {N} \\left(y _ {n}\\right) D _ {n} ^ {2} + \\left(1 - y _ {n}\\right) \\left(\\max \\left(0, m - D _ {n}\\right)\\right) ^ {2} \\tag {2}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.558, + 0.788, + 0.604 + ], + "angle": 0, + "content": "where \\( y \\in \\{0,1\\} \\) is the label (1 for similar pairs and 0 for dissimilar pairs), \\( D_w \\) is the Euclidean distance between the two data points in the pair after passing them through the network, and \\( m \\) is the margin." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.622, + 0.79, + 0.744 + ], + "angle": 0, + "content": "Stage 2 In stage 2 the output modality heads of the pre-trained model are concatenated together to produce audio-visual representations. Three fully-connected layers are added on top of these concatenated heads. We then fine tune the entire model for the downstream task of predicting human-alignment scores. Since the aggregated human annotation scores are non-negative real numbers, we approach model training as a regression task and utilize concordance correlation coefficient (CCC) to measure the agreement between the true and predicted scores. The CCC measurement is illustrated in Equation 3" + }, + { + "type": "equation", + "bbox": [ + 0.389, + 0.749, + 0.785, + 0.782 + ], + "angle": 0, + "content": "\\[\n\\mathcal {L} _ {C C C} = \\frac {2 \\rho \\sigma_ {x} \\sigma_ {y}}{\\sigma_ {x} ^ {2} + \\sigma_ {y} ^ {2} + (\\mu_ {x} - \\mu_ {y}) ^ {2}} \\tag {3}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.787, + 0.788, + 0.821 + ], + "angle": 0, + "content": "where \\(\\mu_{x}\\) and \\(\\mu_{y}\\) denote the means of the true and predicted scores of a batch, respectively; \\(\\sigma_{x}\\) and \\(\\sigma_{y}\\) represent the standard deviations of the true and predicted" + }, + { + "type": "page_footnote", + "bbox": [ + 0.218, + 0.825, + 0.559, + 0.839 + ], + "angle": 0, + "content": "Note that this is different from AVS benchmark." + } + ], + [ + { + "type": "header", + "bbox": [ + 0.4, + 0.115, + 0.732, + 0.131 + ], + "angle": 0, + "content": "Perceptual Evaluation of Audio-Visual Synchrony" + }, + { + "type": "page_number", + "bbox": [ + 0.769, + 0.117, + 0.784, + 0.127 + ], + "angle": 0, + "content": "11" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.147, + 0.788, + 0.193 + ], + "angle": 0, + "content": "scores of the batch, respectively; and \\(\\rho\\) is their Pearson's correlation coefficient. Our training objective is to maximize CCC, ensuring that the predicted scores closely correlate with the true scores." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.215, + 0.788, + 0.323 + ], + "angle": 0, + "content": "Training Details In stage 1, we divided the video files obtained from AudioSet's Balanced train split [15] into an \\(80\\%\\) training set and a \\(20\\%\\) validation set. This \\(80/20\\) split is carried out on the 17K original videos and extrapolated to distorted samples. We pre-trained the model in stage 1 until the validation loss plateaued. In stage 2, we use the train/validation/testing sets provided in the AVS benchmark corpus. We fine-tuned the model on training data for 20 epochs, saving the best-performing model based on the validation set performance [9]." + }, + { + "type": "title", + "bbox": [ + 0.215, + 0.345, + 0.377, + 0.364 + ], + "angle": 0, + "content": "6 Experiments" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.378, + 0.788, + 0.56 + ], + "angle": 0, + "content": "In this section, we evaluate the performance of our trained model using human-evaluation scores. We compare the model's scores with those provided by human evaluators on the test split of AVS benchmark. To further assess the effectiveness of our cross-modal transformer architecture for metric training, we run two ablation experiments. First, to assess the impact of pre-training in stage 1, we directly train our model from scratch as in stage 2. This variant is termed \"Cross-Modal Base\" as it incorporates the same audiovisual cross-modal framework but excludes the contrastive learning pre-training phase. Secondly, to assess the importance of cross-modal framework, we test a version without cross-modal layers called \"Basic Transformer\". This version employs only the self-attention branches, differently from the cross-modal approach of sharing query vector from one modality to another." + }, + { + "type": "title", + "bbox": [ + 0.215, + 0.582, + 0.483, + 0.596 + ], + "angle": 0, + "content": "6.1 Human Scores Correlation" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.608, + 0.788, + 0.79 + ], + "angle": 0, + "content": "For the above models, we compute the correlation of their outputs with the human judgements using pearson correlation coefficient (PCC). Since Fréchet based metrics require a set of videos to calculate distance between multivariate Gaussians we compute set level correlations. As PEAVS and its variants produce a score for each video separately, we average the scores across the test set to compute set level scores[10] We also provide correlations at a much fine-grained clip level. Examining the results presented in Table 2 FAVD shows a relative improvement of \\(11\\%\\) (from 0.475 to 0.527) and \\(64\\%\\) (from 0.321 to 0.527) over FAD and FVD baselines setting a strong baseline for future work. This highlights FAVD's effectiveness in capturing AV synchrony issues. PEAVS model achieves a Pearson correlation of 0.794 at the Set Level, marking a substantial \\(51\\%\\) improvement over the robust FAVD baseline. Notably, when we analyze the" + }, + { + "type": "page_footnote", + "bbox": [ + 0.215, + 0.797, + 0.787, + 0.812 + ], + "angle": 0, + "content": "For further details on the pre-training and fine-tuning settings, refer to appendix A." + }, + { + "type": "page_footnote", + "bbox": [ + 0.215, + 0.812, + 0.787, + 0.84 + ], + "angle": 0, + "content": "10 Each set is defined as a collection of videos showing same distortion type and level. In total we had 90 sets, i.e. 9 distortions x 10 levels." + }, + { + "type": "list", + "bbox": [ + 0.215, + 0.797, + 0.787, + 0.84 + ], + "angle": 0, + "content": null + } + ], + [ + { + "type": "page_number", + "bbox": [ + 0.218, + 0.116, + 0.236, + 0.127 + ], + "angle": 0, + "content": "12" + }, + { + "type": "header", + "bbox": [ + 0.272, + 0.115, + 0.415, + 0.127 + ], + "angle": 0, + "content": "Goncalves et al. 2024" + }, + { + "type": "image", + "bbox": [ + 0.272, + 0.145, + 0.734, + 0.357 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.214, + 0.365, + 0.788, + 0.435 + ], + "angle": 0, + "content": "Fig. 4: PEAVS results with Human Evaluation Scores compared side-by-side for each distortion type and at different levels. PEAVS scores are represented by blue bars and Human Evaluation scores are represented by orange bars. The x-axis are marked by the parameter and its levels as shown in Table \\(\\boxed{1}\\) \\(\\rho\\) is the correlation of PEAVS with human judgements." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.463, + 0.788, + 0.541 + ], + "angle": 0, + "content": "impact of pre-training on PEAVS (c.f. Cross-Modal Base), performance drops by \\(7\\%\\) & \\(11\\%\\) at set and clip level, respectively, underscoring the significance of pre-training in PEAVS. Furthermore, removing the cross-modal layers (c.f. Base Transformer) results in a performance decrease of \\(9.4\\%\\) & \\(19\\%\\) at set and clip level, respectively, emphasizing the critical role of cross-modality layers." + }, + { + "type": "table", + "bbox": [ + 0.359, + 0.564, + 0.645, + 0.667 + ], + "angle": 0, + "content": "
Model TypeSet LevelClip Level
FVD370.321-
FAD230.475-
FAVD (Ours)0.527-
Basic Transformer0.7200.434
Cross-Modal Base0.7360.477
PEAVS (Ours)0.7940.536
" + }, + { + "type": "table_caption", + "bbox": [ + 0.25, + 0.666, + 0.751, + 0.68 + ], + "angle": 0, + "content": "Table 2: Correlation scores of different model types at Set and Clip levels." + }, + { + "type": "title", + "bbox": [ + 0.215, + 0.742, + 0.457, + 0.757 + ], + "angle": 0, + "content": "6.2 Per Distortion Analysis" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.765, + 0.788, + 0.84 + ], + "angle": 0, + "content": "In this section, we analyze the response of PEAVS to varying distortions at different levels. Figure 4 factorizes the results in Table 2 by each distortion type and level. PEAVS scores are shown side-by-side with the scores from human annotations on AVS benchmark test set. Generally, PEAVS metric aligns well with scores from human judgements across the board except in a few cases." + } + ], + [ + { + "type": "header", + "bbox": [ + 0.4, + 0.115, + 0.733, + 0.13 + ], + "angle": 0, + "content": "Perceptual Evaluation of Audio-Visual Synchrony" + }, + { + "type": "page_number", + "bbox": [ + 0.769, + 0.116, + 0.786, + 0.127 + ], + "angle": 0, + "content": "13" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.146, + 0.788, + 0.237 + ], + "angle": 0, + "content": "In top-right corner of Figure 4 we present the effects of audio shift. For ease of analysis, we display these results in terms of absolute shift values. We start with a shift of 0.045 based on a report from ITU which indicated that the threshold for audio-visual shift detection by human ranges from -125ms to +45ms. [13]. At this level, both the metric's output and human scores peak (>4), and they subsequently decrease as the shift increases." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.239, + 0.788, + 0.313 + ], + "angle": 0, + "content": "In contrast to audio-shift, PEAVS shows the highest correlation for intermittent muting distortion which implies it is easier for the metric to capture this distortion. This confirms the notion that we observed in Section 3.2 where intermittent muting stood out as the most noticeable distortion and it was easier for humans to detect it as well." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.315, + 0.787, + 0.361 + ], + "angle": 0, + "content": "In case of the fragment shuffling (bottom-right corner), initial levels of perturbation — shuffled fragments of 0.3-0.4 seconds — is significantly more disruptive for humans than the metric, i.e. the metric fails to capture this insight." + }, + { + "type": "table", + "bbox": [ + 0.302, + 0.387, + 0.422, + 0.45 + ], + "angle": 0, + "content": "
Predicted
Pos.Neg.
ActualPos39227
Neg4130
" + }, + { + "type": "table", + "bbox": [ + 0.586, + 0.388, + 0.703, + 0.45 + ], + "angle": 0, + "content": "
Predicted
Pos.Neg.
ActualPos92174
Neg24110
" + }, + { + "type": "table_caption", + "bbox": [ + 0.214, + 0.451, + 0.79, + 0.479 + ], + "angle": 0, + "content": "Table 3: Confusion matrix for SparseSync (left) and PEAVS (right). Accuracy of SparseSync = 42.3% v/s PEAVS = 50.5%" + }, + { + "type": "title", + "bbox": [ + 0.215, + 0.513, + 0.455, + 0.529 + ], + "angle": 0, + "content": "6.3 PEAVS vs. SparseSync" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.542, + 0.788, + 0.678 + ], + "angle": 0, + "content": "In this section, we report an experiment comparing PEAVS with SparseSync, which was introduced by Iashin et al. [21] for detecting audio-visual shifts inside 'in the wild' videos. While SparseSync's goal is to objectively quantify specific synchronization issues, by estimating an audio-visual offset shift value, PEAVS evaluates synchronization under multiple facets and according to a perceptual scale. Due to their diverging objectives, a direct comparison is challenging. Moreover, SparseSync is trained on 21 classes representing distinct audio-shift levels from -2.0 sec (left) to 2.0 sec (right) with increments of 0.2, with 0.0 representing no audio shift; while PEAVS is trained to predict float values between 1 and 5." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.68, + 0.788, + 0.801 + ], + "angle": 0, + "content": "In this experiment, we assess both model's performance in accurately classifying 400 videos, including 200 ground truth videos and 200 randomly audio-shifted versions of them. We consider distortions of 0.045, 0.1 and \\(\\pm 0.125\\) as positive cases (i.e. ground truth) as ITU recommend an acceptability threshold for AV shift up to \\(185\\mathrm{ms}\\). This also aligns well with SparseSync buckets where first distortion starts at \\(\\pm 0.2\\) seconds (i.e. SparseSync's step-size). To ensure a fair assessment, we divided the PEAVS scale into 21 bins, matching the output structure of SparseSync. For ground truth videos, we anticipate that" + }, + { + "type": "page_footnote", + "bbox": [ + 0.214, + 0.81, + 0.671, + 0.828 + ], + "angle": 0, + "content": "11 More details about this evaluation set are in Appendix Section D" + }, + { + "type": "page_footnote", + "bbox": [ + 0.214, + 0.826, + 0.465, + 0.842 + ], + "angle": 0, + "content": "12 We use increments of 0.238 \\((= \\frac{5}{21})\\)." + }, + { + "type": "list", + "bbox": [ + 0.214, + 0.81, + 0.671, + 0.842 + ], + "angle": 0, + "content": null + } + ], + [ + { + "type": "page_number", + "bbox": [ + 0.218, + 0.116, + 0.236, + 0.127 + ], + "angle": 0, + "content": "14" + }, + { + "type": "header", + "bbox": [ + 0.272, + 0.115, + 0.415, + 0.127 + ], + "angle": 0, + "content": "Goncalves et al. 2024" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.147, + 0.784, + 0.177 + ], + "angle": 0, + "content": "SparseSync will classify them as having zero shifts (0.0 class), while PEAVS is expected to assign them to the highest scoring bin (i.e. (4.76, 5])." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.178, + 0.785, + 0.237 + ], + "angle": 0, + "content": "In Table3 we show the confusion matrices for SparseSync (left) and PEAVS (right). In terms of accuracy, PEAVS outperforms SparseSync by \\(19\\%\\), i.e. \\(50.5\\%\\) vs. \\(42.3\\%\\). In particular, SparseSync shows an \\(18\\%\\) edge in detecting distortions while PEAVS shows a \\(136\\%\\) higher rate in detecting clean content." + }, + { + "type": "title", + "bbox": [ + 0.216, + 0.26, + 0.364, + 0.275 + ], + "angle": 0, + "content": "7 Limitations" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.29, + 0.784, + 0.319 + ], + "angle": 0, + "content": "Like any research, our work is not without its limitations, and we have identified a couple of them here:" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.321, + 0.785, + 0.395 + ], + "angle": 0, + "content": "Dataset: Due to budget constraints (30K USD), we could only source 200 videos and conduct 60K pair-wise annotations. In selecting these 200 videos, we ensured diversity. Our dataset comprises 215 labels out of the 512 present in AudioSet, representing only \\(40\\%\\) of labels. Expanding this dataset for broader video coverage is left for future work." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.396, + 0.785, + 0.47 + ], + "angle": 0, + "content": "Metric: The limited size of the AVS benchmark highlights the necessity of pretraining in model-based metrics (in our case, PEAVS). This is further evidenced by the results in Table 2 where pre-training improves correlation from 0.736 (Cross-Modal Base) to 0.794 (PEAVS). In this study, we only pre-trained PEAVS on audio-shift noise, leaving pre-training on other noise types for future work." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.472, + 0.784, + 0.517 + ], + "angle": 0, + "content": "\"In the wild\" videos: Due to proprietary issues, we could not experiment with datasets featuring talking faces like VoxCeleb 30 or Lip Reading Sentences [1]. Our study is confined to addressing synchrony issues for \"in the wild\" videos." + }, + { + "type": "title", + "bbox": [ + 0.216, + 0.538, + 0.368, + 0.554 + ], + "angle": 0, + "content": "8 Conclusions" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.568, + 0.785, + 0.749 + ], + "angle": 0, + "content": "We introduced PEAVS, a novel metric designed for measuring synchrony in audio-visual content, an important advancement in the field of evaluation of audio-visual content. While the domain of audio-visual generative modeling is ever-evolving, assessing synchronization remains crucial for enhancing user experience. Although numerous metrics cater to either audio or visual content separately, a noticeable gap exists in evaluating their synchronization with a focus on viewers opinions. Our study addresses this gap with PEAVS, a metric based on a vast human-annotated dataset. The strong correlation between PEAVS scores and human evaluations confirms its efficacy in capturing perceptions surrounding audio-visual synchronization. As the scope of audio-visual content generation expands, tools like PEAVS are essential in maintaining a balanced interplay between sight and sound, guiding both creators and researchers." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.75, + 0.785, + 0.84 + ], + "angle": 0, + "content": "Furthermore, we introduced the Audio-Visual Synchrony human perception benchmark, which provides a new dataset detailing human-perceptual scores on various audio-visual synchrony challenges for real-world videos. By isolating specific challenges in a controlled audio-visual environment, we aim to simplify the task for researchers to propose, assess, and analyze potential solutions while maintaining agreement with human perception." + } + ], + [ + { + "type": "header", + "bbox": [ + 0.4, + 0.115, + 0.732, + 0.129 + ], + "angle": 0, + "content": "Perceptual Evaluation of Audio-Visual Synchrony" + }, + { + "type": "page_number", + "bbox": [ + 0.769, + 0.117, + 0.786, + 0.127 + ], + "angle": 0, + "content": "15" + }, + { + "type": "title", + "bbox": [ + 0.217, + 0.145, + 0.324, + 0.16 + ], + "angle": 0, + "content": "References" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.176, + 0.785, + 0.204 + ], + "angle": 0, + "content": "1. Afouras, T., Chung, J.S., Zisserman, A.: Lrs3-ted: a large-scale dataset for visual speech recognition (2018)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.205, + 0.787, + 0.412 + ], + "angle": 0, + "content": "2. Agarwal, M., Agrawal, S., Anastasopoulos, A., Bentivogli, L., Bojar, O., Borg, C., Carpuat, M., Cattoni, R., Cettolo, M., Chen, M., Chen, W., Choukri, K., Chronopoulou, A., Currey, A., Declerck, T., Dong, Q., Duh, K., Esteve, Y., Federico, M., Gahbiche, S., Haddow, B., Hsu, B., Mon Htut, P., Inaguma, H., Javorsky, D., Judge, J., Kano, Y., Ko, T., Kumar, R., Li, P., Ma, X., Mathur, P., Matusov, E., McNamee, P., P. McCrae, J., Murray, K., Nadejde, M., Nakamura, S., Negri, M., Nguyen, H., Niehues, J., Niu, X., Kr. Ojha, A., E. Ortega, J., Pal, P., Pino, J., van der Plas, L., Polak, P., Ripbeth, E., Salesky, E., Shi, J., Sperber, M., Stüker, S., Sudoh, K., Tang, Y., Thompson, B., Tran, K., Turchi, M., Waibel, A., Wang, M., Watanabe, S., Zevallos, R.: FINDINGS OF THE IWSLT 2023 EVALUATION CAMPAIGN. In: Proceedings of the 20th International Conference on Spoken Language Translation (IWSLT 2023). pp. 1-61. Association for Computational Linguistics, Toronto, Canada (in-person and online) (Jul 2023). https://doi.org/10.18653/v1/2023.iwslt-1.1 https://aclanthology.org/2023.iwslt-1.1" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.413, + 0.787, + 0.453 + ], + "angle": 0, + "content": "3. Arandjelović, R., Zisserman, A.: Look, listen and learn. 2017 IEEE International Conference on Computer Vision (ICCV) pp. 609-617 (2017), https://apisemantic scholar.org/CorpusID:10769575" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.454, + 0.787, + 0.481 + ], + "angle": 0, + "content": "4. Arandjelović, R., Zisserman, A.: Objects that sound. In: Computer Vision - ECCV 2018. pp. 451-466. Springer International Publishing, Cham (2018)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.482, + 0.787, + 0.536 + ], + "angle": 0, + "content": "5. Binkowski, M., Sutherland, D.J., Arbel, M., Gretton, A.: Demystifying MMD gans. In: 6th International Conference on Learning Representations, ICLR 2018, Vancouver, BC, Canada, April 30 - May 3, 2018, Conference Track Proceedings. Open-Review.net (2018), https://openreview.net/forum?id=r1lUOzWCw" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.537, + 0.787, + 0.564 + ], + "angle": 0, + "content": "6. BT.1359, R.I.R.: BT.1359 - Relative timing of sound and vision for broadcasting (1998)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.565, + 0.787, + 0.633 + ], + "angle": 0, + "content": "7. Carreira, J., Zisserman, A.: Quo vadis, action recognition? a new model and the kinetics dataset. In: 2017 IEEE Conference on Computer Vision and Pattern Recognition (CVPR). pp. 4724-4733. IEEE Computer Society, Los Alamitos, CA, USA (jul 2017). https://doi.org/10.1109/CVPR.2017.502 https://doi.ieeecomputersociety.org/10.1109/CVPR.2017.502" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.634, + 0.787, + 0.66 + ], + "angle": 0, + "content": "8. Chen, H., Xie, W., Afouras, T., Nagrani, A., Vedaldi, A., Zisserman, A.: Audiovisual synchronisation in the wild (2021)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.661, + 0.787, + 0.688 + ], + "angle": 0, + "content": "9. Chen, H., Xie, W., Vedaldi, A., Zisserman, A.: Vggsound: A large-scale audio-visual dataset (2020)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.689, + 0.787, + 0.744 + ], + "angle": 0, + "content": "0. Chinen, M., Lim, F.S.C., Skoglund, J., Gureev, N., O'Gorman, F., Hines, A.: Visqol v3: An open source production ready objective speech and audio metric. In: 2020 Twelfth International Conference on Quality of Multimedia Experience (QoMEX). pp. 1-6 (2020). https://doi.org/10.1109/QoMEX48832.2020.9123150" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.745, + 0.787, + 0.771 + ], + "angle": 0, + "content": "1. Chung, J.S., Zisserman, A.: Out of time: automated lip sync in the wild. In: Workshop on Multi-view Lip-reading, ACCV (2016)" + }, + { + "type": "ref_text", + "bbox": [ + 0.22, + 0.772, + 0.787, + 0.841 + ], + "angle": 0, + "content": "12. Clark, A.P., Howard, K.L., Woods, A.T., Penton-Voak, I.S., Neumann, C.: Why rate when you could compare? Using the \"EloChoice\" package to assess pairwise comparisons of perceived physical strength. PLOS ONE 13(1), e0190393 (Jan 2018). https://doi.org/10.1371/journal.pone.0190393 https://dx.plos.org/10.1371/journal.pone.0190393" + }, + { + "type": "list", + "bbox": [ + 0.22, + 0.176, + 0.787, + 0.841 + ], + "angle": 0, + "content": null + } + ], + [ + { + "type": "page_number", + "bbox": [ + 0.218, + 0.116, + 0.236, + 0.127 + ], + "angle": 0, + "content": "16" + }, + { + "type": "header", + "bbox": [ + 0.272, + 0.115, + 0.415, + 0.127 + ], + "angle": 0, + "content": "Goncalves et al. 2024" + }, + { + "type": "ref_text", + "bbox": [ + 0.218, + 0.147, + 0.787, + 0.203 + ], + "angle": 0, + "content": "13. Committee, A.T.S.: Atsc implementation subcommittee finding: Relative timing of sound and vision for broadcast operations. Tech. rep., Advanced Television Systems Committee, 1750 K Street, N.W., Suite 1200, Washington, D.C. 20006 (6 2003), doc. IS-191" + }, + { + "type": "ref_text", + "bbox": [ + 0.218, + 0.204, + 0.787, + 0.26 + ], + "angle": 0, + "content": "14. Dowson, D., Landau, B.: The frechet distance between multivariate normal distributions. Journal of Multivariate Analysis 12(3), 450-455 (1982). https://doi.org/https://doi.org/10.1016/0047-259X(82)90077-X, https://www.sciencedirect.com/science/article/pii/0047259X8290077X" + }, + { + "type": "ref_text", + "bbox": [ + 0.219, + 0.261, + 0.787, + 0.33 + ], + "angle": 0, + "content": "15. Gemmeke, J.F., Ellis, D.P.W., Freedman, D., Jansen, A., Lawrence, W., Moore, R.C., Plakal, M., Ritter, M.: Audio set: An ontology and human-labeled dataset for audio events. In: 2017 IEEE International Conference on Acoustics, Speech and Signal Processing (ICASSP). pp. 776-780 (2017). https://doi.org/10.1109/ICASSP.2017.7952261" + }, + { + "type": "ref_text", + "bbox": [ + 0.219, + 0.331, + 0.787, + 0.387 + ], + "angle": 0, + "content": "16. Goncalves, L., Busso, C.: Auxformer: Robust approach to audiovisual emotion recognition. In: ICASSP 2022 - 2022 IEEE International Conference on Acoustics, Speech and Signal Processing (ICASSP). pp. 7357-7361 (2022). https://doi.org/10.1109/ICASSP43922.2022.9747157" + }, + { + "type": "ref_text", + "bbox": [ + 0.219, + 0.388, + 0.787, + 0.457 + ], + "angle": 0, + "content": "17. Hershey, J., Movellan, J.: Audio vision: Using audio-visual synchrony to locate sounds. In: Solla, S., Leen, T., Müller, K. (eds.) Advances in Neural Information Processing Systems. vol. 12. MIT Press (1999), https://proceedings.neurips.cc/paper_files/paper/1999/file/b618c3210e934362ac261db280128c22-Paper.pdf" + }, + { + "type": "ref_text", + "bbox": [ + 0.219, + 0.458, + 0.787, + 0.5 + ], + "angle": 0, + "content": "18. Hershey, S., Chaudhuri, S., Ellis, D.P.W., Gemmeke, J.F., Jansen, A., Moore, R.C., Plakal, M., Platt, D., Saurous, R.A., Seybold, B., Slaney, M., Weiss, R.J., Wilson, K.: Cnn architectures for large-scale audio classification (2017)" + }, + { + "type": "ref_text", + "bbox": [ + 0.219, + 0.501, + 0.787, + 0.528 + ], + "angle": 0, + "content": "19. Hessel, J., Holtzman, A., Forbes, M., Bras, R.L., Choi, Y.: CLIPScore: a reference-free evaluation metric for image captioning. In: EMNLP (2021)" + }, + { + "type": "ref_text", + "bbox": [ + 0.219, + 0.529, + 0.787, + 0.571 + ], + "angle": 0, + "content": "20. Heusel, M., Ramsauer, H., Unterthiner, T., Nessler, B., Hochreiter, S.: Gans trained by a two time-scale update rule converge to a local nash equilibrium. In: Neural Information Processing Systems (2017)" + }, + { + "type": "ref_text", + "bbox": [ + 0.219, + 0.571, + 0.787, + 0.613 + ], + "angle": 0, + "content": "21. Iashin, V., Xie, W., Rahtu, E., Zisserman, A.: Sparse in space and time: Audiovisual synchronisation with trainable selectors. In: British Machine Vision Conference (BMVC) (2022)" + }, + { + "type": "ref_text", + "bbox": [ + 0.219, + 0.614, + 0.787, + 0.642 + ], + "angle": 0, + "content": "22. ITU-T RECOMMENDATION, P.: Subjective video quality assessment methods for multimedia applications (1999)" + }, + { + "type": "ref_text", + "bbox": [ + 0.219, + 0.643, + 0.787, + 0.685 + ], + "angle": 0, + "content": "23. Kilgour, K., Zuluaga, M., Roblek, D., Sharifi, M.: Fréchet audio distance: A reference-free metric for evaluating music enhancement algorithms. In: Interspeech (2019), https://api-semanticscholar.org/CorpusID:202725406" + }, + { + "type": "ref_text", + "bbox": [ + 0.219, + 0.686, + 0.787, + 0.727 + ], + "angle": 0, + "content": "24. Korhonen, J., You, J.: Korhonen. In: 2012 Fourth International Workshop on Quality of Multimedia Experience. pp. 37-38 (2012). https://doi.org/10.1109/QoMEX.2012.6263880" + }, + { + "type": "ref_text", + "bbox": [ + 0.219, + 0.728, + 0.787, + 0.755 + ], + "angle": 0, + "content": "25. Krippendorff, K.: Content Analysis: An Introduction to Its Methodology (second edition). Sage Publications (2004)" + }, + { + "type": "ref_text", + "bbox": [ + 0.219, + 0.757, + 0.787, + 0.784 + ], + "angle": 0, + "content": "26. Luo, S., Yan, C., Hu, C., Zhao, H.: Diff-foley: Synchronized video-to-audio synthesis with latent diffusion models (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.219, + 0.785, + 0.787, + 0.842 + ], + "angle": 0, + "content": "27. Luo, Y., Mesgarani, N.: Tasnet: Time-domain audio separation network for real-time, single-channel speech separation. In: 2018 IEEE International Conference on Acoustics, Speech and Signal Processing (ICASSP). pp. 696-700 (2018). https://doi.org/10.1109/ICASSP.2018.8462116" + }, + { + "type": "list", + "bbox": [ + 0.218, + 0.147, + 0.787, + 0.842 + ], + "angle": 0, + "content": null + } + ], + [ + { + "type": "header", + "bbox": [ + 0.4, + 0.115, + 0.733, + 0.13 + ], + "angle": 0, + "content": "Perceptual Evaluation of Audio-Visual Synchrony" + }, + { + "type": "page_number", + "bbox": [ + 0.769, + 0.117, + 0.786, + 0.127 + ], + "angle": 0, + "content": "17" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.147, + 0.788, + 0.217 + ], + "angle": 0, + "content": "28. Manocha, P., Finkelstein, A., Zhang, R., Bryan, N.J., Mysore, G.J., Jin, Z.: A Differentiable Perceptual Audio Metric Learned from Just Noticeable Differences. In: Interspeech 2020. pp. 2852-2856. ISCA (Oct 2020). https://doi.org/10.21437/Interspeech.2020-1191 https://www.isca-speech.org/archive/interspeech_2020/manocha20_interspeech.html" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.217, + 0.787, + 0.271 + ], + "angle": 0, + "content": "29. Manocha, P., Kumar, A., Xu, B., Menon, A., Gebru, I.D., Ithapu, V.K., Calamia, P.: Dplm: A deep perceptual spatial-audio localization metric. In: 2021 IEEE Workshop on Applications of Signal Processing to Audio and Acoustics (WASPAA). pp. 6-10 (2021). https://doi.org/10.1109/WASPAA52581.2021.9632781" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.271, + 0.787, + 0.298 + ], + "angle": 0, + "content": "30. Nagrani, A., Chung, J.S., Zisserman, A.: Voxceleb: a large-scale speaker identification dataset. In: INTERSPEECH (2017)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.298, + 0.787, + 0.366 + ], + "angle": 0, + "content": "31. Rix, A., Beerends, J., Hollier, M., Hekstra, A.: Perceptual evaluation of speech quality (pesq)-a new method for speech quality assessment of telephone networks and codescs. In: 2001 IEEE International Conference on Acoustics, Speech, and Signal Processing. Proceedings (Cat. No.01CH37221). vol. 2, pp. 749-752 vol.2 (2001). https://doi.org/10.1109/ICASSP.2001.941023" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.366, + 0.787, + 0.419 + ], + "angle": 0, + "content": "32. Roux, J.L., Wisdom, S., Erdogan, H., Hershey, J.R.: Sdr - half-baked or well done? In: ICASSP 2019 - 2019 IEEE International Conference on Acoustics, Speech and Signal Processing (ICASSP). pp. 626-630 (2019). https://doi.org/10.1109/ICASSP.2019.8683855" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.42, + 0.787, + 0.502 + ], + "angle": 0, + "content": "33. Ruan, L., Ma, Y., Yang, H., He, H., Liu, B., Fu, J., Yuan, N., Jin, Q., Guo, B.: Mm-diffusion: Learning multi-modal diffusion models for joint audio and video generation. In: 2023 IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR). pp. 10219-10228. IEEE Computer Society, Los Alamitos, CA, USA (jun 2023). https://doi.org/10.1109/CVPR52729.2023.00985 https://doi.ieeeccomputersociety.org/10.1109/CVPR52729.2023.00985" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.502, + 0.787, + 0.556 + ], + "angle": 0, + "content": "34. Salimans, T., Goodfellow, I., Zaremba, W., Cheung, V., Radford, A., Chen, X.: Improved techniques for training gans. In: Proceedings of the 30th International Conference on Neural Information Processing Systems. p. 2234-2242. NIPS'16, Curran Associates Inc., Red Hook, NY, USA (2016)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.556, + 0.787, + 0.609 + ], + "angle": 0, + "content": "35. Taal, C.H., Hendriks, R.C., Heusdens, R., Jensen, J.: An algorithm for intelligibility prediction of time-frequency weighted noisy speech. IEEE Transactions on Audio, Speech, and Language Processing 19(7), 2125-2136 (2011). https://doi.org/10.1109/TASL.2011.2114881" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.61, + 0.787, + 0.678 + ], + "angle": 0, + "content": "36. Tsai, Y.H.H., Bai, S., Liang, P.P., Kolter, J.Z., Morency, L.P., Salakhutdinov, R.: Multimodal transformer for unaligned multimodal language sequences. In: Proceedings of the 57th Annual Meeting of the Association for Computational Linguistics (Volume 1: Long Papers). Association for Computational Linguistics, Florence, Italy (7 2019)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.679, + 0.787, + 0.704 + ], + "angle": 0, + "content": "37. Unterthiner, T., van Steenkiste, S., Kurach, K., Marinier, R., Michalski, M., Gelly, S.: Towards accurate generative models of video: A new metric & challenges (2019)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.705, + 0.787, + 0.773 + ], + "angle": 0, + "content": "38. Vaswani, A., Shazeer, N., Parmar, N., Uszkoreit, J., Jones, L., Gomez, A.N., Kaiser, L.u., Polosukhin, I.: Attention is all you need. In: Advances in Neural Information Processing Systems. vol. 30. Curran Associates, Inc. (2017), https://proceedings.neurips.cc/paper_files/paper/2017/file/3f5ee243547dee91fbd053c1c4a845aa-Paper.pdf" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.773, + 0.787, + 0.814 + ], + "angle": 0, + "content": "39. Vincent, E., Gribonval, R., Fevotte, C.: Performance measurement in blind audio source separation. IEEE Transactions on Audio, Speech, and Language Processing 14(4), 1462-1469 (2006). https://doi.org/10.1109/TSA.2005.858005" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.814, + 0.787, + 0.84 + ], + "angle": 0, + "content": "40. Wang, J., Fang, Z., Zhao, H.: Alignnet: A unifying approach to audio-visual alignment. In: WACV (2020)" + }, + { + "type": "list", + "bbox": [ + 0.217, + 0.147, + 0.788, + 0.84 + ], + "angle": 0, + "content": null + } + ], + [ + { + "type": "page_number", + "bbox": [ + 0.218, + 0.116, + 0.236, + 0.127 + ], + "angle": 0, + "content": "18" + }, + { + "type": "header", + "bbox": [ + 0.272, + 0.115, + 0.415, + 0.127 + ], + "angle": 0, + "content": "Goncalves et al. 2024" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.147, + 0.787, + 0.205 + ], + "angle": 0, + "content": "41. Wang, Y., Ke, J., Talebi, H., Yim, J.G., Birkbeck, N., Adsumilli, B., Milanfar, P., Yang, F.: Rich features for perceptual quality assessment ofUGC videos. In: 2021 IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR). pp. 13430-13439 (2021). https://doi.org/10.1109/CVPR46437.2021.01323" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.205, + 0.787, + 0.258 + ], + "angle": 0, + "content": "42. Wang, Z., Simoncelli, E., Bovik, A.: Multiscale structural similarity for image quality assessment. In: The Thirty-Seventh Asilomar Conference on Signals, Systems & Computers, 2003. vol. 2, pp. 1398-1402 Vol.2 (2003). https://doi.org/10.1109/ACSSC.2003.1292216" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.259, + 0.787, + 0.287 + ], + "angle": 0, + "content": "43. Wang, Z., Bovik, A.: A universal image quality index. IEEE Signal Processing Letters 9(3), 81-84 (2002). https://doi.org/10.1109/97.995823" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.287, + 0.787, + 0.343 + ], + "angle": 0, + "content": "44. Zhang, R., Isola, P., Efros, A.A., Shechtman, E., Wang, O.: The unreasonable effectiveness of deep features as a perceptual metric. In: 2018 IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 586-595 (2018). https://doi.org/10.1109/CVPR.2018.00068" + }, + { + "type": "list", + "bbox": [ + 0.217, + 0.147, + 0.787, + 0.343 + ], + "angle": 0, + "content": null + } + ] +] \ No newline at end of file diff --git "a/2024/Perceptual Evaluation of Audio-Visual Synchrony Grounded in Viewers\342\200\231 Opinion Scores/59607ef8-05c8-452e-b964-a6375c1e0fbc_origin.pdf" "b/2024/Perceptual Evaluation of Audio-Visual Synchrony Grounded in Viewers\342\200\231 Opinion Scores/59607ef8-05c8-452e-b964-a6375c1e0fbc_origin.pdf" new file mode 100644 index 0000000000000000000000000000000000000000..57360288b23a92541b00cc164ad4e2b08a01764b --- /dev/null +++ "b/2024/Perceptual Evaluation of Audio-Visual Synchrony Grounded in Viewers\342\200\231 Opinion Scores/59607ef8-05c8-452e-b964-a6375c1e0fbc_origin.pdf" @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:fe1198f8797ae0ebc7e17b2e63b53d5efa733bfc4370bc3d1d49c1f4301b3951 +size 3378875 diff --git "a/2024/Perceptual Evaluation of Audio-Visual Synchrony Grounded in Viewers\342\200\231 Opinion Scores/full.md" "b/2024/Perceptual Evaluation of Audio-Visual Synchrony Grounded in Viewers\342\200\231 Opinion Scores/full.md" new file mode 100644 index 0000000000000000000000000000000000000000..1bf511b8ca843c4d717057d696a87f6303304a7b --- /dev/null +++ "b/2024/Perceptual Evaluation of Audio-Visual Synchrony Grounded in Viewers\342\200\231 Opinion Scores/full.md" @@ -0,0 +1,258 @@ +# Perceptual Evaluation of Audio-Visual Synchrony Grounded in Viewers' Opinion Scores + +Lucas Goncalves\*, Prashant Mathur?, Chandrashekhar Lavania?, Metehan Cekic?, Marcello Federico?, and Kyu J. Han? + +1 The University of Texas at Dallas +2 Amazon pramathu@amazon.com + +Abstract. Recent advancements in audio-visual generative modeling have been propelled by progress in deep learning and the availability of data-rich benchmarks. However, the growth is not attributed solely to models and benchmarks. Universally accepted evaluation metrics also play an important role in advancing the field. While there are many metrics available to evaluate audio and visual content separately, there is a lack of metrics that offer a quantitative and interpretable measure of audio-visual synchronization for videos 'in the wild'. To address this gap, we first created a large scale human annotated dataset $(100 + \mathrm{hrs})$ representing nine types of synchronization errors in audio-visual content and how human perceive them. We then developed a PEAVS (Perceptual Evaluation of Audio-Visual Synchrony) score, a novel automatic metric with a 5-point scale that evaluates the quality of audio-visual synchronization. We validate PEAVS using a newly generated dataset, achieving a Pearson correlation of 0.79 at the set level and 0.54 at the clip level when compared to human labels. In our experiments, we observe a relative gain $50\%$ over a natural extension of Fréchet based metrics for Audio-Visual synchrony, confirming PEAVS' efficacy in objectively modeling subjective perceptions of audio-visual synchronization for videos 'in the wild'. + +# 1 Introduction + +Audio-visual (AV) generative modeling has made rapid progress in recent years due to advancements in deep learning models [38] and the availability of large-scale datasets like VGGSound [9] and AudioSet [15]. However, evaluation of these models remains an open challenge. Existing automatic metrics often focus on specific aspects, such as image quality [19,37] or audio fidelity [10,23,32], lacking a satisfactory measure for assessing audio-visual synchronization of 'in the wild' videos. In fact, lack of well-established metrics for AV synchrony issues can already be observed in literature. For instance, recent works in the realm of audio-visual generation, like MM-Diffusion [33], had to solely rely on single-modality metrics to evaluate their outputs. The work on Diff-Foley [26] had + +to train their own audio-visual synchrony classifier. More advanced evaluation metrics are needed to provide a holistic assessment of audio-visual coherence. The development of such metrics is essential to accurately judge model performance, identify failure cases, and drive further progress in this evolving field. + +Recent studies such as AV Synchrony Transformer (AVST) [8], SparseSync [21] and Diff-Foley [26] have defined the synchronization problem between audio-visual modalities primarily as an issue of temporal offset determination. However, this perspective addresses only one aspect of synchronization challenges. In our work, we extend the scope to a broader range of synchronization issues, encompassing audio/visual speed variations, intermittent muting, fragment shuffling, AV flickering, and temporal shifts. We also delve into developing a perceptual, automatic metric that aligns with human judgment. To achieve this, we initiated a large scale annotation study, gathering human assessments on audio-visual synchrony across over 100 hours of both original and distorted content, featuring the aforementioned synchronization issues. + +In this endeavor, our objective is to create a metric that accomplishes the following: (a) attains a high correlation with human assessments of synchronization issues, (b) offers interpretable scoring for end users, and (c) operates in a reference-free manner, meaning it does not rely on ground truth for quality prediction. The key innovation in our approach lies in the development of an interpretable scale based on detailed perceptual guidelines provided to human raters. By optimizing for concordance with human-annotated scores during training, our metric produces scores directly aligned with predefined levels of audio-visual (a)synchrony. This emphasis on interpretability and alignment with human perception addresses the need for a more insightful evaluation of AV synchronization. + +Our main contributions are as follows: + +- We present a new Audio-Visual Synchrony human perception (AVS) benchmark data set with over 120K annotations and over $100+$ hrs of content. Each video in this benchmark is annotated by three different annotators and we see an agreement of 0.71 Krippendorff's alpha [25]. +- We propose a new audio-visual synchrony evaluation metric - PEAVS - that is reference-free and has interpretable scoring on a scale of 1 to 5. +- PEAVS shows high correlation of 0.79 with human judgements on the benchmark data, significantly outperforming a Fréchet distance based AV synchrony metric. + +# 2 Related Works + +Unimodal Metrics: Several factors have fueled progress in single modality generation, including advancement in deep learning model architectures and training/optimization methodology, increased computational capabilities, a rich availability of multimodal data, and the development of consistent metrics. Established metrics like Inception Score (ISc) [34], Fréchet Inception Distance (FID) [20], Fréchet Audio Distance (FAD) [23], Fréchet Video Distance (FVD) [37], + +inter alia, are frequently employed to gauge the quality of the content generated by unimodal generative models, providing a standard for comparison. However, these metrics are reliant only on one modality - either audio [10,23,27,31,32,35,39], video [5,19,24,37,42,44], or just applied to images [20,34]. In the evaluation of generated multimodal content, ensuring the quality of individual modalities is only half the battle. It is crucial to evaluate the synchronization between audio and visual elements, especially in the context of audio-visual generation. + +Multimodal Metrics: For AV synchrony evaluation, it is important to determine the misalignment between audio and visual inputs. In the past, researchers have explored handling synchrony by statistical approaches and specifically designed features to pinpoint where the sound comes from, like locating the source of a voice in a scene [17]. However, with the rise of deep learning recent studies are harnessing distinct connections between audio and visuals, like those in human speech [11], audio-visual correspondence to check for matching content [3], or musical performances [4]. Some recent research studies, such as those by Chen et al. [8] and Iashin et al. [21], propose methods to determine audio-visual synchronization in a natural diverse video environment. However, their focus is solely on tackling the synchronization problem arising from delays in audio or video streams, specifically addressing differences in timing (offsets). In these existing approaches, the model is trained with the objective of either 1) detecting synchronization errors as a binary classification or 2) predicting the precise offset values through regression. In contrast, our aim is to have the model replicate the way humans perceive synchronization issues across a range of distortion types (see more in Sec 3.1). + +Metrics for "in the wild" videos: The assessment of multimodal content has been extensively investigated within the computer vision community however most of the work has been tailored to lip-synchrony for instance, SyncNet [11] quantifies lip-sync errors by comparing mouth movements to the corresponding speech; AlignNet [40] focuses on aligning video content with ground truth audio, and predicts frame-wise offsets. In this work, we focus on evaluation of AV synchrony for "in the wild" videos which apart from Chen et al. [8], Iashin et al. [21] has been explored in Diff-Foley [26] where the authors train their own alignment classifier to detect synchrony. However, none of these works directly target modeling of human perception. + +Perceptual Metrics: Previous studies on audio metrics [28,29] demonstrated that a learned metric trained on data collected via crowd-sourced human annotations, particularly on just noticeably different samples, correlates well with mean opinion scores. For video modality, Wang et. al. [41] designed a perceptual metric — learnt over a large scale human annotated data — designed to distinguish user generated video content in different dimensions such as semantic content, technical quality, and compression level. Our work in contrast centers on the synchronization between audio and video than solely on video or audio quality. + +
IDDistortionParameterLevels
1Temporal Shift Audioshift length (sec)-1, -.5, -.125, .045, .1, .125, .25, .5, 1, 2
2Audio Speed Change upspeed factor %.025, .05, .10, .15, .20, .25, .30, .4, .5, .75
3Video Speed Change upspeed factor %.025, .05, .10, .15, .20, .25, .30, .4, .5, .75
4Audio Speed Change downspeed factor %.025, .05, .10, .15, .20, .25, .30, .4, .5, .75
5Video Speed Change downspeed factor %.025, .05, .10, .15, .20, .25, .30, .4, .5, .75
6Intermittent Mutingduration (sec) of mute for every 1 sec.01, .025, .05, .1, .2, .3, .5, 1, 2.5, 4
7Randomly Sized Gaps Videogap duration (sec) and prob. 40%.01, .025, .05, .1, .2, .3, .5, 1, 2.5, 4
8Fragment Shufflingduration (sec) of segments to shuffle.3, .4, .5, 1, 1.5, 2, 2.5, 3, 3.5, 4
9AV Flickeringgap duration (sec) and prob. 40%.01, .025, .05, .1, .2, .3, .5, 1, 2.5, 4
+ +Table 1: Summary of Distortions and their Parameters for Audio-Visual Content. + +# 3 AVS Benchmark Dataset + +To develop an effective metric that aligns closely with human perception, it is important to have a dataset representative of human judgments. With this in mind, our study was designed to create a benchmark for our metric training and to assess its correlation with human perception, especially in the context of AV synchronization. + +# 3.1 Data Preparation and Distortions + +Audio-visual synchrony, the alignment of audio and visual elements, is important for an immersive multimedia experience. Even minor disruptions can notably degrade the user's experience, as witnessed in scenarios like watching a video where a mere mismatch in drums playing (visuals) and beats (audio) can become distracting. Various real-world challenges, from unreliable network connection to encoding mishaps, can lead to such anomalies. To produce realistic data representing such scenarios, we first collected 200 videos from the AudioSet corpus' evaluation split [15] ensuring that there is an aspect of synchrony in each of the samples. The samples selected are examples of "in the wild" videos, and does not include any samples with faces. We only include videos from actions such as: car driving by, dogs barking, instruments being played, manual labor being + +done, etc. Each video then underwent nine synchrony related distortions at ten varying levels as shown in Table 1, totaling in 18,200 distorted videos. + +Following [37], we apply static and temporal noises in the AV content. Temporal noise is where the distortions in either or both modalities cause desynchronization in the AV content: + +- Temporal Misalignment: The audio track was offset, either forwards or backwards relative to the video [13], creating de-synchronizations spanning from -1 to 2 seconds. +- Audio Speed Change: Similar to FAD [23], the audio playback speed was manipulated, without corresponding changes in the video, resulting in gradual de-synchronization in the AV content. Depending on the speed change (up/down), we either clip the audio or visual frames to keep the same duration as the original video. In this setting, we ensure the sound characteristics is preserved by keeping essential features such as pitch and timbre while altering speed to minimize changes on sound texture. +- Video Speed Change: Similar to FVD [37], the speed of the video is adjusted while leaving the audio track untouched. This would cause the visuals to slowly move out of sync with the audio. As in the audio speed change, we clip the content to keep the same duration as the original one. +- Fragment Shuffling: Similar to the global swap distortion in FVD [37], both audio and video tracks were segmented and rearranged, preserving segment-level synchrony but disrupting overall synchronization of the video. + +Static noise is where the distortions do not necessarily change the synchrony in the AV content but it can be perceived as synchronization errors by humans: + +- Intermittent Muting: Intermittent periods of silences with varying duration is introduced in the audio track. This distortion disrupts the continuity, which can pose a challenge for some synchronization metrics. This distortion is similar to Pops as described in Kilgour et. al. [23]. +- Randomly Sized Gaps: We randomly pick a frame with a certain probability and add black out visuals with varying number of frames in the video track. This distortion disrupts the continuity of the video. This distortion is similar to black rectangle noise as introduced in FVD. +- AV Flickering: Periods of silence and blackout were introduced in the audio-visual track. Although flickering by itself is not a synchronization issue, it can however be perceived as synchronization errors by humans and poses a challenging scenario for metrics [22]. + +# 3.2 Data Annotation + +Our aim in this task is to NOT evaluate the standalone quality of either audio or video. Instead, our focus is on synchrony issues and how humans perceive + +them. Since pairwise stimulus experiments have shown to be faster than single-stimuli 12 and have been used in dubbing evaluation to foster relative measures in the rating 2, we also present our participants with two videos side-by-side. They are asked to play both videos, compare them, and rate each one on a scale from 1-5 based on the provided guidelines 4 + +From 18.2K videos, we randomly sampled (with replacement) 20K video pairs for comparison while taking into account the different levels of distortion (c.f. Table[1]). Participants conducted pairwise comparisons on these videos and each pair was rated at least three times by different annotators. This resulted in a total of 60,000 pairwise ratings or 120,000 ratings for individual videos. Each video duration being at least 10 seconds, this amounts to $100+$ hours of annotated videos with each video annotated at least 3 times. When looking at the inter-annotator agreement, we divided our data based on the severity of disagreement between the annotators. If all annotators annotated different scores those samples were put in "disagreement" group which constitute nearly $10\%$ of the total annotated data. Within the disagreement group, we saw about $16\%$ of cases where disagreements were equal or higher than 3, e.g., one video was rated as $1/4/5$ by three annotators. These high-disagreement samples were then sent back to the annotators for QA, after which we received their revised annotations. We are unable to directly calculate inter-annotator agreement (like Cohen's or Fleiss Kappa), as we had a large pool of annotators $(80+)$ and not all annotators annotated all samples. However, we do observe a Krippendorff's alpha value of 0.71 for the ratings assigned to all the samples, which corresponds to a moderate agreement [25]. + +After removing outliers and dedduplication we are left with 15.2K videos out of $40\mathrm{K}$ We divide this 15.2K into 10.6K for training, 2.3K for development and 2.3K for evaluation. While creating these splits, we ensured that no original files and their derived (i.e., distorted) files are present in more than one set. + +To gain insights into which distortion types posed the most perceptual challenges for humans in terms of synchronization, we examined videos that underwent direct comparisons and analyzed the variations in the annotated scores. Discrepancies in these scores (absolute values) offered insights into how differently these distortion types were perceived. Figure highlights that intermittent muting (type 6) is consistently the most noticeable distortion amongst all types (all high peaks). This observation aligns with the score distribution we observed in AVS benchmark (see Appendix Section B.5) and our manual analysis where audio muting was the most disruptive distortion and humans could easily differentiate audio muting from other distortions. + +![](images/773fe6d4d007ca70d0ae7c1efa105f19eedcb71bbd0ab9f0b902f884a357b8c4.jpg) +Fig. 1: We compare absolute differences in annotation score across the distortion types. In this plot, x-axis shows distortion types that were compared, i.e., '2 4' represents distortion type 2 v/s 4 and y-axis represents the difference in scores across annotation tasks. For ID to distortion type mapping see Table1 + +# 4 Preliminary Experiments + +Studies by Unterthiner et al. [37] (Frechet Video Distance) and Kilgour et al. [23] (Frechet Audio Distance) have demonstrated the efficacy of embeddings from I3D [7] and VGGish [18] models in capturing qualitative and temporal characteristics from audio and video inputs, respectively. In the absence of dedicated metrics for assessing AV synchrony of 'in the wild' content and the lack of baseline models, we formulated a new Frechet audio-visual distance (FAVD) metric. [7] Additionally, we want to assess if the combination of embeddings from I3D and VGGish models can be useful for detecting synchrony issues. + +Frechet Distance metrics—FID, FAD, FVD, and now FAVD—compare the statistics of generated samples to those of the ground truth or a large collection of samples. In our preliminary experiment, we aimed to explore how FAD, FVD, and FAVD metrics respond to temporal discrepancies in the input data. We conducted our study on 18.2K videos, generating two sets of audio-visual embeddings: one from disrupted content and another representing aligned content (ground truth). For each type and level of distortion, we created a distinct “evaluation set” to analyze the impact of each specific distortion type and level on these metrics. In our experiments, we deliberately chose not to use FID (Frechet Inception Distance) due to its limitation in evaluating temporal aspects. FID is designed for static images and cannot adequately assess content with time-dependent elements, such as videos, where scene continuity and flow are essential for quality evaluation. By focusing on metrics capable of capturing these temporal dynamics, we aim to provide a more accurate and relevant evaluation of the audio-visual content's quality. + +We use I3D and VGGish to extract features from audio-visual content within the same time span (of 0.96 seconds - a default time span from VGGish). + +These representations are concatenated time-wise and we calculate multivariate Gaussians for both evaluation set embeddings, $N_{e}(\mu_{e}, \Sigma_{e})$ , and the ground truth AV collection embeddings, $N_{av}(\mu_{av}, \Sigma_{av})$ . + +We then compute the Fréchet distance between these two Gaussians, as described by Dowson et al. [14]: + +$$ +F \left(N _ {a v}, N _ {e}\right) = \left| \mu_ {a v} - \mu_ {e} \right| ^ {2} + t r \left(\Sigma_ {a v} + \Sigma_ {e} - 2 \left(\Sigma_ {a v} \Sigma_ {e}\right) ^ {\frac {1}{2}}\right) \tag {1} +$$ + +where $tr$ is the trace of the matrix. For FAD and FVD, we compute the embeddings from respective modalities. + +![](images/e24f3a313f158fa75cf27bfe3d6742143da122bc14727a4150345d9065c09640.jpg) +(a) Audio Shift + +![](images/19b53ca4a5b3a1f64d0729a1c841858a1f974fac72ac23f5297b831e879c774e.jpg) +(b) Audio Speed Up +Fig. 2: In these plots, we show the effect of distortions with varying levels on FAD, FVD, and FAVD for three distortion types. Flat trend line implies that a metric is not able to capture the distortion type. Increasing or decreasing trends show that a metric is susceptible to varying levels in a distortion. Distortion levels are taken from Table Detailed plots for all distortion types are available in Appendix C. + +![](images/b787053c3e13812b46d4fac9920f659be36ebf9c11cf5f7972e046b478d3d119.jpg) +(c) Audio Speed Down + +Examining the plots in Figure 2 it becomes evident that only FAVD (depicted in green) is responsive to various temporal distortions associated with AV synchrony across different levels. This is mainly due to use of covariance matrices $(\Sigma_{av}, \Sigma_e)$ from audio-visual features (shown in Eq 1) which helps FAVD track the interaction of AV modality. Additionally, in cases like audio shift (a) where neither FAD nor FVD can detect synchronization issues (indicated by flat trend lines), and audio speed-up where FAD shows minimal movement, FAVD formulation exhibits a more pronounced reaction, highlighting that the fusion of features from I3D and VGGish models effectively captures temporality in an audio-visual context. Based on these observations, we chose to use I3D and VGGish features as inputs for our metric training, with FAVD serving as a baseline metric for the remainder of our work. + +# 5 PEAVS Score Formulation + +This section describes the architecture we have used to train our 'PEAVS' score metric and its training details. + +# 5.1 Architecture + +The input to our proposed model are the video and audio streams, as seen in Figure 3. For feature extraction, we leverage I3D and VGGish. In addition to this, our design draws inspiration from the audio-visual framework presented in [16]. Central to our architecture is the utilization of the cross-modal transformer formulation [36], which excels in capturing temporal information and relationships between different modalities. To ensure temporal consistency across modalities, both audio and visual input features are extracted over matched time spans and are then relayed through cross-modal transformer layers. + +There are several key components in the our architecture. First, the audio and visual inputs are processed through our frozen backbone feature extractors. Both features are extracted over 0.96 second windows. The I3D extracted inputs $x_{v}$ have dimensionality $x_{v} \in \mathbb{R}^{N_{av} \times 1024}$ , where $N_{av}$ is the dimension of the audiovisual feature sequence, and 1,024 is the feature vector dimension. The acoustic feature vector $x_{a}$ extracted from the VGGish model is $x_{a} \in \mathbb{R}^{N_{av} \times 128}$ , where $N_{av}$ is the dimension of the acoustic feature sequence, and 128 is the feature vector dimension. We ensure that all inputs to the cross-modal layers have same dimensionality, by projecting the visual features from 1,024 to 128 through the use of a 1D convolutional layer to produce $\bar{x}_{v} \in \mathbb{R}^{N_{v} \times 128}$ . + +Next, we introduce positional embeddings to the feature vectors, following the step adopted in the original transformer paper [38]. The resultant features are used by the model to compute Query, Key, and Value vectors for each modality, as illustrated in Figure [3]. In our model, the Query vector from one modality is paired in the transformer layer with the Key and Value vectors of the other, ensuring a cross-modal interaction. This design choice allows the cross-modal transformer encoder layers for each modality branch to produce representations for a given modality that are influenced by the other modality. Following this, the output of the cross-modal layers are refined through the self-attention layers to enhance the cross-modal representations. The resultant features are then processed by averaging the outputs from the self-attention layers followed by concatenation of these averaged representations from both modality branches. Finally, a multilayer perceptron (MLP) with three fully-connected layers is integrated to produce a prediction score based on these concatenated features. + +# 5.2 Training + +As seen in Figure 3 our training method consists of a two-stage strategy. In stage 1, we pre-train the model using a contrastive learning task, distinguishing between aligned and non-aligned pairs. Afterwards, in Stage Two, we fine-tune the model on our primary task, which involves predicting scores based on human annotators scores. + +Stage 1 In stage 1, we pre-train our model using a contrastive learning objective. In this phase, we used AudioSet's Balanced train split [15], which after downloading all videos possible and removing corrupted ones resulted in 17K + +![](images/a4001f608f38d265a3dcc211dedc86cad65cc81f032da7600e6c5e9c9af145ff.jpg) +Fig. 3: Framework Overview + +videos for our contrastive training stage We then subjected each of these files to ten levels of temporal misalignment (audio shift), as specified in section 3.1. This pre-processing resulted in a total of $\sim 190\mathrm{K}$ files for the contrastive learning phase. Similar to SyncNet [11], the training objective ensures that the outputs of both the audio and video branches are similar for aligned pairs and dissimilar for misaligned pairs. Specifically, we either minimize or maximize the Euclidean distance between the network outputs for these cases, respectively. Our objective consists of a contrastive loss (Equation 2). The primary intention behind this pre-training is to embed regularization into the audio and visual features extracted from our frozen backbone networks. This ensures that the model prioritizes temporal attributes of the features over their qualitative characteristics. + +$$ +\mathcal {L} _ {c} = \frac {1}{2 N} \sum_ {n = 1} ^ {N} \left(y _ {n}\right) D _ {n} ^ {2} + \left(1 - y _ {n}\right) \left(\max \left(0, m - D _ {n}\right)\right) ^ {2} \tag {2} +$$ + +where $y \in \{0,1\}$ is the label (1 for similar pairs and 0 for dissimilar pairs), $D_w$ is the Euclidean distance between the two data points in the pair after passing them through the network, and $m$ is the margin. + +Stage 2 In stage 2 the output modality heads of the pre-trained model are concatenated together to produce audio-visual representations. Three fully-connected layers are added on top of these concatenated heads. We then fine tune the entire model for the downstream task of predicting human-alignment scores. Since the aggregated human annotation scores are non-negative real numbers, we approach model training as a regression task and utilize concordance correlation coefficient (CCC) to measure the agreement between the true and predicted scores. The CCC measurement is illustrated in Equation 3 + +$$ +\mathcal {L} _ {C C C} = \frac {2 \rho \sigma_ {x} \sigma_ {y}}{\sigma_ {x} ^ {2} + \sigma_ {y} ^ {2} + (\mu_ {x} - \mu_ {y}) ^ {2}} \tag {3} +$$ + +where $\mu_{x}$ and $\mu_{y}$ denote the means of the true and predicted scores of a batch, respectively; $\sigma_{x}$ and $\sigma_{y}$ represent the standard deviations of the true and predicted + +scores of the batch, respectively; and $\rho$ is their Pearson's correlation coefficient. Our training objective is to maximize CCC, ensuring that the predicted scores closely correlate with the true scores. + +Training Details In stage 1, we divided the video files obtained from AudioSet's Balanced train split [15] into an $80\%$ training set and a $20\%$ validation set. This $80/20$ split is carried out on the 17K original videos and extrapolated to distorted samples. We pre-trained the model in stage 1 until the validation loss plateaued. In stage 2, we use the train/validation/testing sets provided in the AVS benchmark corpus. We fine-tuned the model on training data for 20 epochs, saving the best-performing model based on the validation set performance [9]. + +# 6 Experiments + +In this section, we evaluate the performance of our trained model using human-evaluation scores. We compare the model's scores with those provided by human evaluators on the test split of AVS benchmark. To further assess the effectiveness of our cross-modal transformer architecture for metric training, we run two ablation experiments. First, to assess the impact of pre-training in stage 1, we directly train our model from scratch as in stage 2. This variant is termed "Cross-Modal Base" as it incorporates the same audiovisual cross-modal framework but excludes the contrastive learning pre-training phase. Secondly, to assess the importance of cross-modal framework, we test a version without cross-modal layers called "Basic Transformer". This version employs only the self-attention branches, differently from the cross-modal approach of sharing query vector from one modality to another. + +# 6.1 Human Scores Correlation + +For the above models, we compute the correlation of their outputs with the human judgements using pearson correlation coefficient (PCC). Since Fréchet based metrics require a set of videos to calculate distance between multivariate Gaussians we compute set level correlations. As PEAVS and its variants produce a score for each video separately, we average the scores across the test set to compute set level scores[10] We also provide correlations at a much fine-grained clip level. Examining the results presented in Table 2 FAVD shows a relative improvement of $11\%$ (from 0.475 to 0.527) and $64\%$ (from 0.321 to 0.527) over FAD and FVD baselines setting a strong baseline for future work. This highlights FAVD's effectiveness in capturing AV synchrony issues. PEAVS model achieves a Pearson correlation of 0.794 at the Set Level, marking a substantial $51\%$ improvement over the robust FAVD baseline. Notably, when we analyze the + +![](images/986ade816d78df75e02b6cfa4103d131b70c0a0bab1c49b40d8d1ba00b13873b.jpg) +Fig. 4: PEAVS results with Human Evaluation Scores compared side-by-side for each distortion type and at different levels. PEAVS scores are represented by blue bars and Human Evaluation scores are represented by orange bars. The x-axis are marked by the parameter and its levels as shown in Table $\boxed{1}$ $\rho$ is the correlation of PEAVS with human judgements. + +impact of pre-training on PEAVS (c.f. Cross-Modal Base), performance drops by $7\%$ & $11\%$ at set and clip level, respectively, underscoring the significance of pre-training in PEAVS. Furthermore, removing the cross-modal layers (c.f. Base Transformer) results in a performance decrease of $9.4\%$ & $19\%$ at set and clip level, respectively, emphasizing the critical role of cross-modality layers. + +
Model TypeSet LevelClip Level
FVD370.321-
FAD230.475-
FAVD (Ours)0.527-
Basic Transformer0.7200.434
Cross-Modal Base0.7360.477
PEAVS (Ours)0.7940.536
+ +Table 2: Correlation scores of different model types at Set and Clip levels. + +# 6.2 Per Distortion Analysis + +In this section, we analyze the response of PEAVS to varying distortions at different levels. Figure 4 factorizes the results in Table 2 by each distortion type and level. PEAVS scores are shown side-by-side with the scores from human annotations on AVS benchmark test set. Generally, PEAVS metric aligns well with scores from human judgements across the board except in a few cases. + +In top-right corner of Figure 4 we present the effects of audio shift. For ease of analysis, we display these results in terms of absolute shift values. We start with a shift of 0.045 based on a report from ITU which indicated that the threshold for audio-visual shift detection by human ranges from -125ms to +45ms. [13]. At this level, both the metric's output and human scores peak (>4), and they subsequently decrease as the shift increases. + +In contrast to audio-shift, PEAVS shows the highest correlation for intermittent muting distortion which implies it is easier for the metric to capture this distortion. This confirms the notion that we observed in Section 3.2 where intermittent muting stood out as the most noticeable distortion and it was easier for humans to detect it as well. + +In case of the fragment shuffling (bottom-right corner), initial levels of perturbation — shuffled fragments of 0.3-0.4 seconds — is significantly more disruptive for humans than the metric, i.e. the metric fails to capture this insight. + +
Predicted
Pos.Neg.
ActualPos39227
Neg4130
+ +
Predicted
Pos.Neg.
ActualPos92174
Neg24110
+ +Table 3: Confusion matrix for SparseSync (left) and PEAVS (right). Accuracy of SparseSync = 42.3% v/s PEAVS = 50.5% + +# 6.3 PEAVS vs. SparseSync + +In this section, we report an experiment comparing PEAVS with SparseSync, which was introduced by Iashin et al. [21] for detecting audio-visual shifts inside 'in the wild' videos. While SparseSync's goal is to objectively quantify specific synchronization issues, by estimating an audio-visual offset shift value, PEAVS evaluates synchronization under multiple facets and according to a perceptual scale. Due to their diverging objectives, a direct comparison is challenging. Moreover, SparseSync is trained on 21 classes representing distinct audio-shift levels from -2.0 sec (left) to 2.0 sec (right) with increments of 0.2, with 0.0 representing no audio shift; while PEAVS is trained to predict float values between 1 and 5. + +In this experiment, we assess both model's performance in accurately classifying 400 videos, including 200 ground truth videos and 200 randomly audio-shifted versions of them. We consider distortions of 0.045, 0.1 and $\pm 0.125$ as positive cases (i.e. ground truth) as ITU recommend an acceptability threshold for AV shift up to $185\mathrm{ms}$ . This also aligns well with SparseSync buckets where first distortion starts at $\pm 0.2$ seconds (i.e. SparseSync's step-size). To ensure a fair assessment, we divided the PEAVS scale into 21 bins, matching the output structure of SparseSync. For ground truth videos, we anticipate that + +SparseSync will classify them as having zero shifts (0.0 class), while PEAVS is expected to assign them to the highest scoring bin (i.e. (4.76, 5]). + +In Table3 we show the confusion matrices for SparseSync (left) and PEAVS (right). In terms of accuracy, PEAVS outperforms SparseSync by $19\%$ , i.e. $50.5\%$ vs. $42.3\%$ . In particular, SparseSync shows an $18\%$ edge in detecting distortions while PEAVS shows a $136\%$ higher rate in detecting clean content. + +# 7 Limitations + +Like any research, our work is not without its limitations, and we have identified a couple of them here: + +Dataset: Due to budget constraints (30K USD), we could only source 200 videos and conduct 60K pair-wise annotations. In selecting these 200 videos, we ensured diversity. Our dataset comprises 215 labels out of the 512 present in AudioSet, representing only $40\%$ of labels. Expanding this dataset for broader video coverage is left for future work. + +Metric: The limited size of the AVS benchmark highlights the necessity of pretraining in model-based metrics (in our case, PEAVS). This is further evidenced by the results in Table 2 where pre-training improves correlation from 0.736 (Cross-Modal Base) to 0.794 (PEAVS). In this study, we only pre-trained PEAVS on audio-shift noise, leaving pre-training on other noise types for future work. + +"In the wild" videos: Due to proprietary issues, we could not experiment with datasets featuring talking faces like VoxCeleb 30 or Lip Reading Sentences [1]. Our study is confined to addressing synchrony issues for "in the wild" videos. + +# 8 Conclusions + +We introduced PEAVS, a novel metric designed for measuring synchrony in audio-visual content, an important advancement in the field of evaluation of audio-visual content. While the domain of audio-visual generative modeling is ever-evolving, assessing synchronization remains crucial for enhancing user experience. Although numerous metrics cater to either audio or visual content separately, a noticeable gap exists in evaluating their synchronization with a focus on viewers opinions. Our study addresses this gap with PEAVS, a metric based on a vast human-annotated dataset. The strong correlation between PEAVS scores and human evaluations confirms its efficacy in capturing perceptions surrounding audio-visual synchronization. As the scope of audio-visual content generation expands, tools like PEAVS are essential in maintaining a balanced interplay between sight and sound, guiding both creators and researchers. + +Furthermore, we introduced the Audio-Visual Synchrony human perception benchmark, which provides a new dataset detailing human-perceptual scores on various audio-visual synchrony challenges for real-world videos. By isolating specific challenges in a controlled audio-visual environment, we aim to simplify the task for researchers to propose, assess, and analyze potential solutions while maintaining agreement with human perception. + +# References + +1. Afouras, T., Chung, J.S., Zisserman, A.: Lrs3-ted: a large-scale dataset for visual speech recognition (2018) +2. Agarwal, M., Agrawal, S., Anastasopoulos, A., Bentivogli, L., Bojar, O., Borg, C., Carpuat, M., Cattoni, R., Cettolo, M., Chen, M., Chen, W., Choukri, K., Chronopoulou, A., Currey, A., Declerck, T., Dong, Q., Duh, K., Esteve, Y., Federico, M., Gahbiche, S., Haddow, B., Hsu, B., Mon Htut, P., Inaguma, H., Javorsky, D., Judge, J., Kano, Y., Ko, T., Kumar, R., Li, P., Ma, X., Mathur, P., Matusov, E., McNamee, P., P. McCrae, J., Murray, K., Nadejde, M., Nakamura, S., Negri, M., Nguyen, H., Niehues, J., Niu, X., Kr. Ojha, A., E. Ortega, J., Pal, P., Pino, J., van der Plas, L., Polak, P., Ripbeth, E., Salesky, E., Shi, J., Sperber, M., Stüker, S., Sudoh, K., Tang, Y., Thompson, B., Tran, K., Turchi, M., Waibel, A., Wang, M., Watanabe, S., Zevallos, R.: FINDINGS OF THE IWSLT 2023 EVALUATION CAMPAIGN. In: Proceedings of the 20th International Conference on Spoken Language Translation (IWSLT 2023). pp. 1-61. Association for Computational Linguistics, Toronto, Canada (in-person and online) (Jul 2023). https://doi.org/10.18653/v1/2023.iwslt-1.1 https://aclanthology.org/2023.iwslt-1.1 +3. Arandjelović, R., Zisserman, A.: Look, listen and learn. 2017 IEEE International Conference on Computer Vision (ICCV) pp. 609-617 (2017), https://apisemantic scholar.org/CorpusID:10769575 +4. Arandjelović, R., Zisserman, A.: Objects that sound. In: Computer Vision - ECCV 2018. pp. 451-466. Springer International Publishing, Cham (2018) +5. Binkowski, M., Sutherland, D.J., Arbel, M., Gretton, A.: Demystifying MMD gans. In: 6th International Conference on Learning Representations, ICLR 2018, Vancouver, BC, Canada, April 30 - May 3, 2018, Conference Track Proceedings. Open-Review.net (2018), https://openreview.net/forum?id=r1lUOzWCw +6. BT.1359, R.I.R.: BT.1359 - Relative timing of sound and vision for broadcasting (1998) +7. Carreira, J., Zisserman, A.: Quo vadis, action recognition? a new model and the kinetics dataset. In: 2017 IEEE Conference on Computer Vision and Pattern Recognition (CVPR). pp. 4724-4733. IEEE Computer Society, Los Alamitos, CA, USA (jul 2017). https://doi.org/10.1109/CVPR.2017.502 https://doi.ieeecomputersociety.org/10.1109/CVPR.2017.502 +8. Chen, H., Xie, W., Afouras, T., Nagrani, A., Vedaldi, A., Zisserman, A.: Audiovisual synchronisation in the wild (2021) +9. Chen, H., Xie, W., Vedaldi, A., Zisserman, A.: Vggsound: A large-scale audio-visual dataset (2020) +0. Chinen, M., Lim, F.S.C., Skoglund, J., Gureev, N., O'Gorman, F., Hines, A.: Visqol v3: An open source production ready objective speech and audio metric. In: 2020 Twelfth International Conference on Quality of Multimedia Experience (QoMEX). pp. 1-6 (2020). https://doi.org/10.1109/QoMEX48832.2020.9123150 +1. Chung, J.S., Zisserman, A.: Out of time: automated lip sync in the wild. In: Workshop on Multi-view Lip-reading, ACCV (2016) +12. Clark, A.P., Howard, K.L., Woods, A.T., Penton-Voak, I.S., Neumann, C.: Why rate when you could compare? Using the "EloChoice" package to assess pairwise comparisons of perceived physical strength. PLOS ONE 13(1), e0190393 (Jan 2018). https://doi.org/10.1371/journal.pone.0190393 https://dx.plos.org/10.1371/journal.pone.0190393 + +13. Committee, A.T.S.: Atsc implementation subcommittee finding: Relative timing of sound and vision for broadcast operations. Tech. rep., Advanced Television Systems Committee, 1750 K Street, N.W., Suite 1200, Washington, D.C. 20006 (6 2003), doc. IS-191 +14. Dowson, D., Landau, B.: The frechet distance between multivariate normal distributions. Journal of Multivariate Analysis 12(3), 450-455 (1982). https://doi.org/https://doi.org/10.1016/0047-259X(82)90077-X, https://www.sciencedirect.com/science/article/pii/0047259X8290077X +15. Gemmeke, J.F., Ellis, D.P.W., Freedman, D., Jansen, A., Lawrence, W., Moore, R.C., Plakal, M., Ritter, M.: Audio set: An ontology and human-labeled dataset for audio events. In: 2017 IEEE International Conference on Acoustics, Speech and Signal Processing (ICASSP). pp. 776-780 (2017). https://doi.org/10.1109/ICASSP.2017.7952261 +16. Goncalves, L., Busso, C.: Auxformer: Robust approach to audiovisual emotion recognition. In: ICASSP 2022 - 2022 IEEE International Conference on Acoustics, Speech and Signal Processing (ICASSP). pp. 7357-7361 (2022). https://doi.org/10.1109/ICASSP43922.2022.9747157 +17. Hershey, J., Movellan, J.: Audio vision: Using audio-visual synchrony to locate sounds. In: Solla, S., Leen, T., Müller, K. (eds.) Advances in Neural Information Processing Systems. vol. 12. MIT Press (1999), https://proceedings.neurips.cc/paper_files/paper/1999/file/b618c3210e934362ac261db280128c22-Paper.pdf +18. Hershey, S., Chaudhuri, S., Ellis, D.P.W., Gemmeke, J.F., Jansen, A., Moore, R.C., Plakal, M., Platt, D., Saurous, R.A., Seybold, B., Slaney, M., Weiss, R.J., Wilson, K.: Cnn architectures for large-scale audio classification (2017) +19. Hessel, J., Holtzman, A., Forbes, M., Bras, R.L., Choi, Y.: CLIPScore: a reference-free evaluation metric for image captioning. In: EMNLP (2021) +20. Heusel, M., Ramsauer, H., Unterthiner, T., Nessler, B., Hochreiter, S.: Gans trained by a two time-scale update rule converge to a local nash equilibrium. In: Neural Information Processing Systems (2017) +21. Iashin, V., Xie, W., Rahtu, E., Zisserman, A.: Sparse in space and time: Audiovisual synchronisation with trainable selectors. In: British Machine Vision Conference (BMVC) (2022) +22. ITU-T RECOMMENDATION, P.: Subjective video quality assessment methods for multimedia applications (1999) +23. Kilgour, K., Zuluaga, M., Roblek, D., Sharifi, M.: Fréchet audio distance: A reference-free metric for evaluating music enhancement algorithms. In: Interspeech (2019), https://api-semanticscholar.org/CorpusID:202725406 +24. Korhonen, J., You, J.: Korhonen. In: 2012 Fourth International Workshop on Quality of Multimedia Experience. pp. 37-38 (2012). https://doi.org/10.1109/QoMEX.2012.6263880 +25. Krippendorff, K.: Content Analysis: An Introduction to Its Methodology (second edition). Sage Publications (2004) +26. Luo, S., Yan, C., Hu, C., Zhao, H.: Diff-foley: Synchronized video-to-audio synthesis with latent diffusion models (2023) +27. Luo, Y., Mesgarani, N.: Tasnet: Time-domain audio separation network for real-time, single-channel speech separation. In: 2018 IEEE International Conference on Acoustics, Speech and Signal Processing (ICASSP). pp. 696-700 (2018). https://doi.org/10.1109/ICASSP.2018.8462116 + +28. Manocha, P., Finkelstein, A., Zhang, R., Bryan, N.J., Mysore, G.J., Jin, Z.: A Differentiable Perceptual Audio Metric Learned from Just Noticeable Differences. In: Interspeech 2020. pp. 2852-2856. ISCA (Oct 2020). https://doi.org/10.21437/Interspeech.2020-1191 https://www.isca-speech.org/archive/interspeech_2020/manocha20_interspeech.html +29. Manocha, P., Kumar, A., Xu, B., Menon, A., Gebru, I.D., Ithapu, V.K., Calamia, P.: Dplm: A deep perceptual spatial-audio localization metric. In: 2021 IEEE Workshop on Applications of Signal Processing to Audio and Acoustics (WASPAA). pp. 6-10 (2021). https://doi.org/10.1109/WASPAA52581.2021.9632781 +30. Nagrani, A., Chung, J.S., Zisserman, A.: Voxceleb: a large-scale speaker identification dataset. In: INTERSPEECH (2017) +31. Rix, A., Beerends, J., Hollier, M., Hekstra, A.: Perceptual evaluation of speech quality (pesq)-a new method for speech quality assessment of telephone networks and codescs. In: 2001 IEEE International Conference on Acoustics, Speech, and Signal Processing. Proceedings (Cat. No.01CH37221). vol. 2, pp. 749-752 vol.2 (2001). https://doi.org/10.1109/ICASSP.2001.941023 +32. Roux, J.L., Wisdom, S., Erdogan, H., Hershey, J.R.: Sdr - half-baked or well done? In: ICASSP 2019 - 2019 IEEE International Conference on Acoustics, Speech and Signal Processing (ICASSP). pp. 626-630 (2019). https://doi.org/10.1109/ICASSP.2019.8683855 +33. Ruan, L., Ma, Y., Yang, H., He, H., Liu, B., Fu, J., Yuan, N., Jin, Q., Guo, B.: Mm-diffusion: Learning multi-modal diffusion models for joint audio and video generation. In: 2023 IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR). pp. 10219-10228. IEEE Computer Society, Los Alamitos, CA, USA (jun 2023). https://doi.org/10.1109/CVPR52729.2023.00985 https://doi.ieeeccomputersociety.org/10.1109/CVPR52729.2023.00985 +34. Salimans, T., Goodfellow, I., Zaremba, W., Cheung, V., Radford, A., Chen, X.: Improved techniques for training gans. In: Proceedings of the 30th International Conference on Neural Information Processing Systems. p. 2234-2242. NIPS'16, Curran Associates Inc., Red Hook, NY, USA (2016) +35. Taal, C.H., Hendriks, R.C., Heusdens, R., Jensen, J.: An algorithm for intelligibility prediction of time-frequency weighted noisy speech. IEEE Transactions on Audio, Speech, and Language Processing 19(7), 2125-2136 (2011). https://doi.org/10.1109/TASL.2011.2114881 +36. Tsai, Y.H.H., Bai, S., Liang, P.P., Kolter, J.Z., Morency, L.P., Salakhutdinov, R.: Multimodal transformer for unaligned multimodal language sequences. In: Proceedings of the 57th Annual Meeting of the Association for Computational Linguistics (Volume 1: Long Papers). Association for Computational Linguistics, Florence, Italy (7 2019) +37. Unterthiner, T., van Steenkiste, S., Kurach, K., Marinier, R., Michalski, M., Gelly, S.: Towards accurate generative models of video: A new metric & challenges (2019) +38. Vaswani, A., Shazeer, N., Parmar, N., Uszkoreit, J., Jones, L., Gomez, A.N., Kaiser, L.u., Polosukhin, I.: Attention is all you need. In: Advances in Neural Information Processing Systems. vol. 30. Curran Associates, Inc. (2017), https://proceedings.neurips.cc/paper_files/paper/2017/file/3f5ee243547dee91fbd053c1c4a845aa-Paper.pdf +39. Vincent, E., Gribonval, R., Fevotte, C.: Performance measurement in blind audio source separation. IEEE Transactions on Audio, Speech, and Language Processing 14(4), 1462-1469 (2006). https://doi.org/10.1109/TSA.2005.858005 +40. Wang, J., Fang, Z., Zhao, H.: Alignnet: A unifying approach to audio-visual alignment. In: WACV (2020) + +41. Wang, Y., Ke, J., Talebi, H., Yim, J.G., Birkbeck, N., Adsumilli, B., Milanfar, P., Yang, F.: Rich features for perceptual quality assessment ofUGC videos. In: 2021 IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR). pp. 13430-13439 (2021). https://doi.org/10.1109/CVPR46437.2021.01323 +42. Wang, Z., Simoncelli, E., Bovik, A.: Multiscale structural similarity for image quality assessment. In: The Thirty-Seventh Asilomar Conference on Signals, Systems & Computers, 2003. vol. 2, pp. 1398-1402 Vol.2 (2003). https://doi.org/10.1109/ACSSC.2003.1292216 +43. Wang, Z., Bovik, A.: A universal image quality index. IEEE Signal Processing Letters 9(3), 81-84 (2002). https://doi.org/10.1109/97.995823 +44. Zhang, R., Isola, P., Efros, A.A., Shechtman, E., Wang, O.: The unreasonable effectiveness of deep features as a perceptual metric. In: 2018 IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 586-595 (2018). https://doi.org/10.1109/CVPR.2018.00068 \ No newline at end of file diff --git "a/2024/Perceptual Evaluation of Audio-Visual Synchrony Grounded in Viewers\342\200\231 Opinion Scores/images.zip" "b/2024/Perceptual Evaluation of Audio-Visual Synchrony Grounded in Viewers\342\200\231 Opinion Scores/images.zip" new file mode 100644 index 0000000000000000000000000000000000000000..d54a3ff6c323c2fe5a89d56df01b082d19e9f9f7 --- /dev/null +++ "b/2024/Perceptual Evaluation of Audio-Visual Synchrony Grounded in Viewers\342\200\231 Opinion Scores/images.zip" @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:6bb069d535f1666e1554e7dba71d73ce6d6f651db422c9db36edbbb2274a0db0 +size 311523 diff --git "a/2024/Perceptual Evaluation of Audio-Visual Synchrony Grounded in Viewers\342\200\231 Opinion Scores/layout.json" "b/2024/Perceptual Evaluation of Audio-Visual Synchrony Grounded in Viewers\342\200\231 Opinion Scores/layout.json" new file mode 100644 index 0000000000000000000000000000000000000000..95c68ab18283ea79b16da7abe158ce7900624746 --- /dev/null +++ "b/2024/Perceptual Evaluation of Audio-Visual Synchrony Grounded in Viewers\342\200\231 Opinion Scores/layout.json" @@ -0,0 +1,7983 @@ +{ + "pdf_info": [ + { + "para_blocks": [ + { + "bbox": [ + 134, + 111, + 479, + 148 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 134, + 111, + 479, + 148 + ], + "spans": [ + { + "bbox": [ + 134, + 111, + 479, + 148 + ], + "type": "text", + "content": "Perceptual Evaluation of Audio-Visual Synchrony Grounded in Viewers' Opinion Scores" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 140, + 167, + 474, + 194 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 140, + 167, + 474, + 194 + ], + "spans": [ + { + "bbox": [ + 140, + 167, + 474, + 194 + ], + "type": "text", + "content": "Lucas Goncalves\\*, Prashant Mathur?, Chandrashekhar Lavania?, Metehan Cekic?, Marcello Federico?, and Kyu J. Han?" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 232, + 201, + 383, + 236 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 232, + 201, + 383, + 236 + ], + "spans": [ + { + "bbox": [ + 232, + 201, + 383, + 236 + ], + "type": "text", + "content": "1 The University of Texas at Dallas \n2 Amazon pramathu@amazon.com" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 160, + 263, + 455, + 483 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 160, + 263, + 455, + 483 + ], + "spans": [ + { + "bbox": [ + 160, + 263, + 455, + 483 + ], + "type": "text", + "content": "Abstract. Recent advancements in audio-visual generative modeling have been propelled by progress in deep learning and the availability of data-rich benchmarks. However, the growth is not attributed solely to models and benchmarks. Universally accepted evaluation metrics also play an important role in advancing the field. While there are many metrics available to evaluate audio and visual content separately, there is a lack of metrics that offer a quantitative and interpretable measure of audio-visual synchronization for videos 'in the wild'. To address this gap, we first created a large scale human annotated dataset " + }, + { + "bbox": [ + 160, + 263, + 455, + 483 + ], + "type": "inline_equation", + "content": "(100 + \\mathrm{hrs})" + }, + { + "bbox": [ + 160, + 263, + 455, + 483 + ], + "type": "text", + "content": " representing nine types of synchronization errors in audio-visual content and how human perceive them. We then developed a PEAVS (Perceptual Evaluation of Audio-Visual Synchrony) score, a novel automatic metric with a 5-point scale that evaluates the quality of audio-visual synchronization. We validate PEAVS using a newly generated dataset, achieving a Pearson correlation of 0.79 at the set level and 0.54 at the clip level when compared to human labels. In our experiments, we observe a relative gain " + }, + { + "bbox": [ + 160, + 263, + 455, + 483 + ], + "type": "inline_equation", + "content": "50\\%" + }, + { + "bbox": [ + 160, + 263, + 455, + 483 + ], + "type": "text", + "content": " over a natural extension of Fréchet based metrics for Audio-Visual synchrony, confirming PEAVS' efficacy in objectively modeling subjective perceptions of audio-visual synchronization for videos 'in the wild'." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 132, + 502, + 230, + 514 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 502, + 230, + 514 + ], + "spans": [ + { + "bbox": [ + 132, + 502, + 230, + 514 + ], + "type": "text", + "content": "1 Introduction" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 130, + 526, + 482, + 648 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 526, + 482, + 648 + ], + "spans": [ + { + "bbox": [ + 130, + 526, + 482, + 648 + ], + "type": "text", + "content": "Audio-visual (AV) generative modeling has made rapid progress in recent years due to advancements in deep learning models [38] and the availability of large-scale datasets like VGGSound [9] and AudioSet [15]. However, evaluation of these models remains an open challenge. Existing automatic metrics often focus on specific aspects, such as image quality [19,37] or audio fidelity [10,23,32], lacking a satisfactory measure for assessing audio-visual synchronization of 'in the wild' videos. In fact, lack of well-established metrics for AV synchrony issues can already be observed in literature. For instance, recent works in the realm of audio-visual generation, like MM-Diffusion [33], had to solely rely on single-modality metrics to evaluate their outputs. The work on Diff-Foley [26] had" + } + ] + } + ], + "index": 5 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 133, + 653, + 312, + 665 + ], + "type": "page_footnote", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 653, + 312, + 665 + ], + "spans": [ + { + "bbox": [ + 133, + 653, + 312, + 665 + ], + "type": "text", + "content": "* Work done during internship at Amazon." + } + ] + } + ], + "index": 6 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 0 + }, + { + "para_blocks": [ + { + "bbox": [ + 130, + 116, + 482, + 163 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 116, + 482, + 163 + ], + "spans": [ + { + "bbox": [ + 130, + 116, + 482, + 163 + ], + "type": "text", + "content": "to train their own audio-visual synchrony classifier. More advanced evaluation metrics are needed to provide a holistic assessment of audio-visual coherence. The development of such metrics is essential to accurately judge model performance, identify failure cases, and drive further progress in this evolving field." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 130, + 164, + 483, + 295 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 164, + 483, + 295 + ], + "spans": [ + { + "bbox": [ + 130, + 164, + 483, + 295 + ], + "type": "text", + "content": "Recent studies such as AV Synchrony Transformer (AVST) [8], SparseSync [21] and Diff-Foley [26] have defined the synchronization problem between audio-visual modalities primarily as an issue of temporal offset determination. However, this perspective addresses only one aspect of synchronization challenges. In our work, we extend the scope to a broader range of synchronization issues, encompassing audio/visual speed variations, intermittent muting, fragment shuffling, AV flickering, and temporal shifts. We also delve into developing a perceptual, automatic metric that aligns with human judgment. To achieve this, we initiated a large scale annotation study, gathering human assessments on audio-visual synchrony across over 100 hours of both original and distorted content, featuring the aforementioned synchronization issues." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 130, + 296, + 482, + 426 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 296, + 482, + 426 + ], + "spans": [ + { + "bbox": [ + 130, + 296, + 482, + 426 + ], + "type": "text", + "content": "In this endeavor, our objective is to create a metric that accomplishes the following: (a) attains a high correlation with human assessments of synchronization issues, (b) offers interpretable scoring for end users, and (c) operates in a reference-free manner, meaning it does not rely on ground truth for quality prediction. The key innovation in our approach lies in the development of an interpretable scale based on detailed perceptual guidelines provided to human raters. By optimizing for concordance with human-annotated scores during training, our metric produces scores directly aligned with predefined levels of audio-visual (a)synchrony. This emphasis on interpretability and alignment with human perception addresses the need for a more insightful evaluation of AV synchronization." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 132, + 427, + 302, + 438 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 427, + 302, + 438 + ], + "spans": [ + { + "bbox": [ + 132, + 427, + 302, + 438 + ], + "type": "text", + "content": "Our main contributions are as follows:" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 138, + 445, + 481, + 552 + ], + "type": "list", + "angle": 0, + "index": 9, + "blocks": [ + { + "bbox": [ + 138, + 445, + 481, + 495 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 445, + 481, + 495 + ], + "spans": [ + { + "bbox": [ + 138, + 445, + 481, + 495 + ], + "type": "text", + "content": "- We present a new Audio-Visual Synchrony human perception (AVS) benchmark data set with over 120K annotations and over " + }, + { + "bbox": [ + 138, + 445, + 481, + 495 + ], + "type": "inline_equation", + "content": "100+" + }, + { + "bbox": [ + 138, + 445, + 481, + 495 + ], + "type": "text", + "content": " hrs of content. Each video in this benchmark is annotated by three different annotators and we see an agreement of 0.71 Krippendorff's alpha [25]." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 138, + 493, + 481, + 517 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 493, + 481, + 517 + ], + "spans": [ + { + "bbox": [ + 138, + 493, + 481, + 517 + ], + "type": "text", + "content": "- We propose a new audio-visual synchrony evaluation metric - PEAVS - that is reference-free and has interpretable scoring on a scale of 1 to 5." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 138, + 517, + 481, + 552 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 517, + 481, + 552 + ], + "spans": [ + { + "bbox": [ + 138, + 517, + 481, + 552 + ], + "type": "text", + "content": "- PEAVS shows high correlation of 0.79 with human judgements on the benchmark data, significantly outperforming a Fréchet distance based AV synchrony metric." + } + ] + } + ], + "index": 8 + } + ], + "sub_type": "text" + }, + { + "bbox": [ + 132, + 570, + 243, + 582 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 570, + 243, + 582 + ], + "spans": [ + { + "bbox": [ + 132, + 570, + 243, + 582 + ], + "type": "text", + "content": "2 Related Works" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 130, + 594, + 482, + 668 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 594, + 482, + 668 + ], + "spans": [ + { + "bbox": [ + 130, + 594, + 482, + 668 + ], + "type": "text", + "content": "Unimodal Metrics: Several factors have fueled progress in single modality generation, including advancement in deep learning model architectures and training/optimization methodology, increased computational capabilities, a rich availability of multimodal data, and the development of consistent metrics. Established metrics like Inception Score (ISc) [34], Fréchet Inception Distance (FID) [20], Fréchet Audio Distance (FAD) [23], Fréchet Video Distance (FVD) [37]," + } + ] + } + ], + "index": 11 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "spans": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "type": "text", + "content": "2" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 166, + 91, + 253, + 101 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 166, + 91, + 253, + 101 + ], + "spans": [ + { + "bbox": [ + 166, + 91, + 253, + 101 + ], + "type": "text", + "content": "Goncalves et al. 2024" + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 1 + }, + { + "para_blocks": [ + { + "bbox": [ + 130, + 116, + 482, + 201 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 116, + 482, + 201 + ], + "spans": [ + { + "bbox": [ + 130, + 116, + 482, + 201 + ], + "type": "text", + "content": "inter alia, are frequently employed to gauge the quality of the content generated by unimodal generative models, providing a standard for comparison. However, these metrics are reliant only on one modality - either audio [10,23,27,31,32,35,39], video [5,19,24,37,42,44], or just applied to images [20,34]. In the evaluation of generated multimodal content, ensuring the quality of individual modalities is only half the battle. It is crucial to evaluate the synchronization between audio and visual elements, especially in the context of audio-visual generation." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 130, + 215, + 483, + 422 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 215, + 483, + 422 + ], + "spans": [ + { + "bbox": [ + 130, + 215, + 483, + 422 + ], + "type": "text", + "content": "Multimodal Metrics: For AV synchrony evaluation, it is important to determine the misalignment between audio and visual inputs. In the past, researchers have explored handling synchrony by statistical approaches and specifically designed features to pinpoint where the sound comes from, like locating the source of a voice in a scene [17]. However, with the rise of deep learning recent studies are harnessing distinct connections between audio and visuals, like those in human speech [11], audio-visual correspondence to check for matching content [3], or musical performances [4]. Some recent research studies, such as those by Chen et al. [8] and Iashin et al. [21], propose methods to determine audio-visual synchronization in a natural diverse video environment. However, their focus is solely on tackling the synchronization problem arising from delays in audio or video streams, specifically addressing differences in timing (offsets). In these existing approaches, the model is trained with the objective of either 1) detecting synchronization errors as a binary classification or 2) predicting the precise offset values through regression. In contrast, our aim is to have the model replicate the way humans perceive synchronization issues across a range of distortion types (see more in Sec 3.1)." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 130, + 434, + 483, + 555 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 434, + 483, + 555 + ], + "spans": [ + { + "bbox": [ + 130, + 434, + 483, + 555 + ], + "type": "text", + "content": "Metrics for \"in the wild\" videos: The assessment of multimodal content has been extensively investigated within the computer vision community however most of the work has been tailored to lip-synchrony for instance, SyncNet [11] quantifies lip-sync errors by comparing mouth movements to the corresponding speech; AlignNet [40] focuses on aligning video content with ground truth audio, and predicts frame-wise offsets. In this work, we focus on evaluation of AV synchrony for \"in the wild\" videos which apart from Chen et al. [8], Iashin et al. [21] has been explored in Diff-Foley [26] where the authors train their own alignment classifier to detect synchrony. However, none of these works directly target modeling of human perception." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 130, + 569, + 482, + 668 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 569, + 482, + 668 + ], + "spans": [ + { + "bbox": [ + 130, + 569, + 482, + 668 + ], + "type": "text", + "content": "Perceptual Metrics: Previous studies on audio metrics [28,29] demonstrated that a learned metric trained on data collected via crowd-sourced human annotations, particularly on just noticeably different samples, correlates well with mean opinion scores. For video modality, Wang et. al. [41] designed a perceptual metric — learnt over a large scale human annotated data — designed to distinguish user generated video content in different dimensions such as semantic content, technical quality, and compression level. Our work in contrast centers on the synchronization between audio and video than solely on video or audio quality." + } + ] + } + ], + "index": 5 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 244, + 91, + 447, + 103 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 244, + 91, + 447, + 103 + ], + "spans": [ + { + "bbox": [ + 244, + 91, + 447, + 103 + ], + "type": "text", + "content": "Perceptual Evaluation of Audio-Visual Synchrony" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 474, + 92, + 481, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 474, + 92, + 481, + 100 + ], + "spans": [ + { + "bbox": [ + 474, + 92, + 481, + 100 + ], + "type": "text", + "content": "3" + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 2 + }, + { + "para_blocks": [ + { + "type": "table", + "bbox": [ + 133, + 114, + 481, + 334 + ], + "blocks": [ + { + "bbox": [ + 133, + 114, + 481, + 334 + ], + "lines": [ + { + "bbox": [ + 133, + 114, + 481, + 334 + ], + "spans": [ + { + "bbox": [ + 133, + 114, + 481, + 334 + ], + "type": "table", + "html": "
IDDistortionParameterLevels
1Temporal Shift Audioshift length (sec)-1, -.5, -.125, .045, .1, .125, .25, .5, 1, 2
2Audio Speed Change upspeed factor %.025, .05, .10, .15, .20, .25, .30, .4, .5, .75
3Video Speed Change upspeed factor %.025, .05, .10, .15, .20, .25, .30, .4, .5, .75
4Audio Speed Change downspeed factor %.025, .05, .10, .15, .20, .25, .30, .4, .5, .75
5Video Speed Change downspeed factor %.025, .05, .10, .15, .20, .25, .30, .4, .5, .75
6Intermittent Mutingduration (sec) of mute for every 1 sec.01, .025, .05, .1, .2, .3, .5, 1, 2.5, 4
7Randomly Sized Gaps Videogap duration (sec) and prob. 40%.01, .025, .05, .1, .2, .3, .5, 1, 2.5, 4
8Fragment Shufflingduration (sec) of segments to shuffle.3, .4, .5, 1, 1.5, 2, 2.5, 3, 3.5, 4
9AV Flickeringgap duration (sec) and prob. 40%.01, .025, .05, .1, .2, .3, .5, 1, 2.5, 4
", + "image_path": "851cf86a671a07346aa98e9ca3486f069c01e268ef2d78c923b3b4bae9788d58.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "table_body" + } + ], + "index": 2 + }, + { + "bbox": [ + 138, + 335, + 474, + 344 + ], + "lines": [ + { + "bbox": [ + 138, + 335, + 474, + 344 + ], + "spans": [ + { + "bbox": [ + 138, + 335, + 474, + 344 + ], + "type": "text", + "content": "Table 1: Summary of Distortions and their Parameters for Audio-Visual Content." + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "text" + }, + { + "bbox": [ + 132, + 381, + 302, + 394 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 381, + 302, + 394 + ], + "spans": [ + { + "bbox": [ + 132, + 381, + 302, + 394 + ], + "type": "text", + "content": "3 AVS Benchmark Dataset" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 132, + 416, + 481, + 476 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 416, + 481, + 476 + ], + "spans": [ + { + "bbox": [ + 132, + 416, + 481, + 476 + ], + "type": "text", + "content": "To develop an effective metric that aligns closely with human perception, it is important to have a dataset representative of human judgments. With this in mind, our study was designed to create a benchmark for our metric training and to assess its correlation with human perception, especially in the context of AV synchronization." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 132, + 504, + 330, + 516 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 504, + 330, + 516 + ], + "spans": [ + { + "bbox": [ + 132, + 504, + 330, + 516 + ], + "type": "text", + "content": "3.1 Data Preparation and Distortions" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 132, + 533, + 482, + 666 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 533, + 482, + 666 + ], + "spans": [ + { + "bbox": [ + 132, + 533, + 482, + 666 + ], + "type": "text", + "content": "Audio-visual synchrony, the alignment of audio and visual elements, is important for an immersive multimedia experience. Even minor disruptions can notably degrade the user's experience, as witnessed in scenarios like watching a video where a mere mismatch in drums playing (visuals) and beats (audio) can become distracting. Various real-world challenges, from unreliable network connection to encoding mishaps, can lead to such anomalies. To produce realistic data representing such scenarios, we first collected 200 videos from the AudioSet corpus' evaluation split [15] ensuring that there is an aspect of synchrony in each of the samples. The samples selected are examples of \"in the wild\" videos, and does not include any samples with faces. We only include videos from actions such as: car driving by, dogs barking, instruments being played, manual labor being" + } + ] + } + ], + "index": 7 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "spans": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "type": "text", + "content": "4" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 166, + 91, + 253, + 100 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 166, + 91, + 253, + 100 + ], + "spans": [ + { + "bbox": [ + 166, + 91, + 253, + 100 + ], + "type": "text", + "content": "Goncalves et al. 2024" + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 3 + }, + { + "para_blocks": [ + { + "bbox": [ + 130, + 115, + 479, + 140 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 115, + 479, + 140 + ], + "spans": [ + { + "bbox": [ + 130, + 115, + 479, + 140 + ], + "type": "text", + "content": "done, etc. Each video then underwent nine synchrony related distortions at ten varying levels as shown in Table 1, totaling in 18,200 distorted videos." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 130, + 140, + 480, + 175 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 140, + 480, + 175 + ], + "spans": [ + { + "bbox": [ + 130, + 140, + 480, + 175 + ], + "type": "text", + "content": "Following [37], we apply static and temporal noises in the AV content. Temporal noise is where the distortions in either or both modalities cause desynchronization in the AV content:" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 137, + 184, + 480, + 385 + ], + "type": "list", + "angle": 0, + "index": 8, + "blocks": [ + { + "bbox": [ + 138, + 184, + 479, + 217 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 184, + 479, + 217 + ], + "spans": [ + { + "bbox": [ + 138, + 184, + 479, + 217 + ], + "type": "text", + "content": "- Temporal Misalignment: The audio track was offset, either forwards or backwards relative to the video [13], creating de-synchronizations spanning from -1 to 2 seconds." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 137, + 220, + 480, + 301 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 137, + 220, + 480, + 301 + ], + "spans": [ + { + "bbox": [ + 137, + 220, + 480, + 301 + ], + "type": "text", + "content": "- Audio Speed Change: Similar to FAD [23], the audio playback speed was manipulated, without corresponding changes in the video, resulting in gradual de-synchronization in the AV content. Depending on the speed change (up/down), we either clip the audio or visual frames to keep the same duration as the original video. In this setting, we ensure the sound characteristics is preserved by keeping essential features such as pitch and timbre while altering speed to minimize changes on sound texture." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 137, + 304, + 479, + 349 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 137, + 304, + 479, + 349 + ], + "spans": [ + { + "bbox": [ + 137, + 304, + 479, + 349 + ], + "type": "text", + "content": "- Video Speed Change: Similar to FVD [37], the speed of the video is adjusted while leaving the audio track untouched. This would cause the visuals to slowly move out of sync with the audio. As in the audio speed change, we clip the content to keep the same duration as the original one." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 137, + 352, + 479, + 385 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 137, + 352, + 479, + 385 + ], + "spans": [ + { + "bbox": [ + 137, + 352, + 479, + 385 + ], + "type": "text", + "content": "- Fragment Shuffling: Similar to the global swap distortion in FVD [37], both audio and video tracks were segmented and rearranged, preserving segment-level synchrony but disrupting overall synchronization of the video." + } + ] + } + ], + "index": 7 + } + ], + "sub_type": "text" + }, + { + "bbox": [ + 130, + 396, + 479, + 420 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 396, + 479, + 420 + ], + "spans": [ + { + "bbox": [ + 130, + 396, + 479, + 420 + ], + "type": "text", + "content": "Static noise is where the distortions do not necessarily change the synchrony in the AV content but it can be perceived as synchronization errors by humans:" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 137, + 430, + 479, + 572 + ], + "type": "list", + "angle": 0, + "index": 13, + "blocks": [ + { + "bbox": [ + 137, + 430, + 479, + 475 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 137, + 430, + 479, + 475 + ], + "spans": [ + { + "bbox": [ + 137, + 430, + 479, + 475 + ], + "type": "text", + "content": "- Intermittent Muting: Intermittent periods of silences with varying duration is introduced in the audio track. This distortion disrupts the continuity, which can pose a challenge for some synchronization metrics. This distortion is similar to Pops as described in Kilgour et. al. [23]." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 137, + 478, + 479, + 522 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 137, + 478, + 479, + 522 + ], + "spans": [ + { + "bbox": [ + 137, + 478, + 479, + 522 + ], + "type": "text", + "content": "- Randomly Sized Gaps: We randomly pick a frame with a certain probability and add black out visuals with varying number of frames in the video track. This distortion disrupts the continuity of the video. This distortion is similar to black rectangle noise as introduced in FVD." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 137, + 525, + 479, + 572 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 137, + 525, + 479, + 572 + ], + "spans": [ + { + "bbox": [ + 137, + 525, + 479, + 572 + ], + "type": "text", + "content": "- AV Flickering: Periods of silence and blackout were introduced in the audio-visual track. Although flickering by itself is not a synchronization issue, it can however be perceived as synchronization errors by humans and poses a challenging scenario for metrics [22]." + } + ] + } + ], + "index": 12 + } + ], + "sub_type": "text" + }, + { + "bbox": [ + 132, + 591, + 246, + 602 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 591, + 246, + 602 + ], + "spans": [ + { + "bbox": [ + 132, + 591, + 246, + 602 + ], + "type": "text", + "content": "3.2 Data Annotation" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 130, + 612, + 479, + 635 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 612, + 479, + 635 + ], + "spans": [ + { + "bbox": [ + 130, + 612, + 479, + 635 + ], + "type": "text", + "content": "Our aim in this task is to NOT evaluate the standalone quality of either audio or video. Instead, our focus is on synchrony issues and how humans perceive" + } + ] + } + ], + "index": 15 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 244, + 91, + 448, + 103 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 244, + 91, + 448, + 103 + ], + "spans": [ + { + "bbox": [ + 244, + 91, + 448, + 103 + ], + "type": "text", + "content": "Perceptual Evaluation of Audio-Visual Synchrony" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 474, + 92, + 481, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 474, + 92, + 481, + 100 + ], + "spans": [ + { + "bbox": [ + 474, + 92, + 481, + 100 + ], + "type": "text", + "content": "5" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 133, + 640, + 479, + 663 + ], + "type": "page_footnote", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 640, + 479, + 663 + ], + "spans": [ + { + "bbox": [ + 133, + 640, + 479, + 663 + ], + "type": "text", + "content": "3 Appendix B.4 contains a brief overview of top-30 label frequencies contained within the videos subset selected for our dataset." + } + ] + } + ], + "index": 16 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 4 + }, + { + "para_blocks": [ + { + "bbox": [ + 130, + 116, + 482, + 177 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 116, + 482, + 177 + ], + "spans": [ + { + "bbox": [ + 130, + 116, + 482, + 177 + ], + "type": "text", + "content": "them. Since pairwise stimulus experiments have shown to be faster than single-stimuli 12 and have been used in dubbing evaluation to foster relative measures in the rating 2, we also present our participants with two videos side-by-side. They are asked to play both videos, compare them, and rate each one on a scale from 1-5 based on the provided guidelines 4" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 130, + 180, + 482, + 397 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 180, + 482, + 397 + ], + "spans": [ + { + "bbox": [ + 130, + 180, + 482, + 397 + ], + "type": "text", + "content": "From 18.2K videos, we randomly sampled (with replacement) 20K video pairs for comparison while taking into account the different levels of distortion (c.f. Table[1]). Participants conducted pairwise comparisons on these videos and each pair was rated at least three times by different annotators. This resulted in a total of 60,000 pairwise ratings or 120,000 ratings for individual videos. Each video duration being at least 10 seconds, this amounts to " + }, + { + "bbox": [ + 130, + 180, + 482, + 397 + ], + "type": "inline_equation", + "content": "100+" + }, + { + "bbox": [ + 130, + 180, + 482, + 397 + ], + "type": "text", + "content": " hours of annotated videos with each video annotated at least 3 times. When looking at the inter-annotator agreement, we divided our data based on the severity of disagreement between the annotators. If all annotators annotated different scores those samples were put in \"disagreement\" group which constitute nearly " + }, + { + "bbox": [ + 130, + 180, + 482, + 397 + ], + "type": "inline_equation", + "content": "10\\%" + }, + { + "bbox": [ + 130, + 180, + 482, + 397 + ], + "type": "text", + "content": " of the total annotated data. Within the disagreement group, we saw about " + }, + { + "bbox": [ + 130, + 180, + 482, + 397 + ], + "type": "inline_equation", + "content": "16\\%" + }, + { + "bbox": [ + 130, + 180, + 482, + 397 + ], + "type": "text", + "content": " of cases where disagreements were equal or higher than 3, e.g., one video was rated as " + }, + { + "bbox": [ + 130, + 180, + 482, + 397 + ], + "type": "inline_equation", + "content": "1/4/5" + }, + { + "bbox": [ + 130, + 180, + 482, + 397 + ], + "type": "text", + "content": " by three annotators. These high-disagreement samples were then sent back to the annotators for QA, after which we received their revised annotations. We are unable to directly calculate inter-annotator agreement (like Cohen's or Fleiss Kappa), as we had a large pool of annotators " + }, + { + "bbox": [ + 130, + 180, + 482, + 397 + ], + "type": "inline_equation", + "content": "(80+)" + }, + { + "bbox": [ + 130, + 180, + 482, + 397 + ], + "type": "text", + "content": " and not all annotators annotated all samples. However, we do observe a Krippendorff's alpha value of 0.71 for the ratings assigned to all the samples, which corresponds to a moderate agreement [25]." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 130, + 400, + 482, + 449 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 400, + 482, + 449 + ], + "spans": [ + { + "bbox": [ + 130, + 400, + 482, + 449 + ], + "type": "text", + "content": "After removing outliers and dedduplication we are left with 15.2K videos out of " + }, + { + "bbox": [ + 130, + 400, + 482, + 449 + ], + "type": "inline_equation", + "content": "40\\mathrm{K}" + }, + { + "bbox": [ + 130, + 400, + 482, + 449 + ], + "type": "text", + "content": " We divide this 15.2K into 10.6K for training, 2.3K for development and 2.3K for evaluation. While creating these splits, we ensured that no original files and their derived (i.e., distorted) files are present in more than one set." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 130, + 453, + 483, + 573 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 453, + 483, + 573 + ], + "spans": [ + { + "bbox": [ + 130, + 453, + 483, + 573 + ], + "type": "text", + "content": "To gain insights into which distortion types posed the most perceptual challenges for humans in terms of synchronization, we examined videos that underwent direct comparisons and analyzed the variations in the annotated scores. Discrepancies in these scores (absolute values) offered insights into how differently these distortion types were perceived. Figure highlights that intermittent muting (type 6) is consistently the most noticeable distortion amongst all types (all high peaks). This observation aligns with the score distribution we observed in AVS benchmark (see Appendix Section B.5) and our manual analysis where audio muting was the most disruptive distortion and humans could easily differentiate audio muting from other distortions." + } + ] + } + ], + "index": 5 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "spans": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "type": "text", + "content": "6" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 166, + 91, + 253, + 101 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 166, + 91, + 253, + 101 + ], + "spans": [ + { + "bbox": [ + 166, + 91, + 253, + 101 + ], + "type": "text", + "content": "Goncalves et al. 2024" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 133, + 597, + 481, + 620 + ], + "type": "page_footnote", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 597, + 481, + 620 + ], + "spans": [ + { + "bbox": [ + 133, + 597, + 481, + 620 + ], + "type": "text", + "content": "4 Annotation guidelines are shown in Appendix B.2. Upon publication, we will release the data set and code for creating distorted samples." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 134, + 620, + 481, + 643 + ], + "type": "page_footnote", + "angle": 0, + "lines": [ + { + "bbox": [ + 134, + 620, + 481, + 643 + ], + "spans": [ + { + "bbox": [ + 134, + 620, + 481, + 643 + ], + "type": "text", + "content": "5 Since we randomly sampled with replacement, one video was annotated approx. 2.4 times by annotators, which led to duplication." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 135, + 643, + 481, + 665 + ], + "type": "page_footnote", + "angle": 0, + "lines": [ + { + "bbox": [ + 135, + 643, + 481, + 665 + ], + "spans": [ + { + "bbox": [ + 135, + 643, + 481, + 665 + ], + "type": "text", + "content": "6 We leave out audio-shift distortion in this analysis as the levels are not linearly increasing or decreasing." + } + ] + } + ], + "index": 8 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 5 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 223, + 121, + 379, + 247 + ], + "blocks": [ + { + "bbox": [ + 223, + 121, + 379, + 247 + ], + "lines": [ + { + "bbox": [ + 223, + 121, + 379, + 247 + ], + "spans": [ + { + "bbox": [ + 223, + 121, + 379, + 247 + ], + "type": "image", + "image_path": "773fe6d4d007ca70d0ae7c1efa105f19eedcb71bbd0ab9f0b902f884a357b8c4.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 130, + 254, + 482, + 301 + ], + "lines": [ + { + "bbox": [ + 130, + 254, + 482, + 301 + ], + "spans": [ + { + "bbox": [ + 130, + 254, + 482, + 301 + ], + "type": "text", + "content": "Fig. 1: We compare absolute differences in annotation score across the distortion types. In this plot, x-axis shows distortion types that were compared, i.e., '2 4' represents distortion type 2 v/s 4 and y-axis represents the difference in scores across annotation tasks. For ID to distortion type mapping see Table1" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_caption" + } + ], + "index": 2 + }, + { + "bbox": [ + 132, + 323, + 304, + 338 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 323, + 304, + 338 + ], + "spans": [ + { + "bbox": [ + 132, + 323, + 304, + 338 + ], + "type": "text", + "content": "4 Preliminary Experiments" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 130, + 353, + 482, + 450 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 353, + 482, + 450 + ], + "spans": [ + { + "bbox": [ + 130, + 353, + 482, + 450 + ], + "type": "text", + "content": "Studies by Unterthiner et al. [37] (Frechet Video Distance) and Kilgour et al. [23] (Frechet Audio Distance) have demonstrated the efficacy of embeddings from I3D [7] and VGGish [18] models in capturing qualitative and temporal characteristics from audio and video inputs, respectively. In the absence of dedicated metrics for assessing AV synchrony of 'in the wild' content and the lack of baseline models, we formulated a new Frechet audio-visual distance (FAVD) metric. [7] Additionally, we want to assess if the combination of embeddings from I3D and VGGish models can be useful for detecting synchrony issues." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 130, + 451, + 482, + 631 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 451, + 482, + 631 + ], + "spans": [ + { + "bbox": [ + 130, + 451, + 482, + 631 + ], + "type": "text", + "content": "Frechet Distance metrics—FID, FAD, FVD, and now FAVD—compare the statistics of generated samples to those of the ground truth or a large collection of samples. In our preliminary experiment, we aimed to explore how FAD, FVD, and FAVD metrics respond to temporal discrepancies in the input data. We conducted our study on 18.2K videos, generating two sets of audio-visual embeddings: one from disrupted content and another representing aligned content (ground truth). For each type and level of distortion, we created a distinct “evaluation set” to analyze the impact of each specific distortion type and level on these metrics. In our experiments, we deliberately chose not to use FID (Frechet Inception Distance) due to its limitation in evaluating temporal aspects. FID is designed for static images and cannot adequately assess content with time-dependent elements, such as videos, where scene continuity and flow are essential for quality evaluation. By focusing on metrics capable of capturing these temporal dynamics, we aim to provide a more accurate and relevant evaluation of the audio-visual content's quality." + } + ] + } + ], + "index": 6 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 244, + 91, + 448, + 103 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 244, + 91, + 448, + 103 + ], + "spans": [ + { + "bbox": [ + 244, + 91, + 448, + 103 + ], + "type": "text", + "content": "Perceptual Evaluation of Audio-Visual Synchrony" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 474, + 91, + 481, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 474, + 91, + 481, + 100 + ], + "spans": [ + { + "bbox": [ + 474, + 91, + 481, + 100 + ], + "type": "text", + "content": "7" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 133, + 640, + 481, + 666 + ], + "type": "page_footnote", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 640, + 481, + 666 + ], + "spans": [ + { + "bbox": [ + 133, + 640, + 481, + 666 + ], + "type": "inline_equation", + "content": "^7" + }, + { + "bbox": [ + 133, + 640, + 481, + 666 + ], + "type": "text", + "content": " SparseSync 21 and AVST 8 are only able to detect audio shift but our benchmark contains many more synchrony issues which render these metrics unusable." + } + ] + } + ], + "index": 7 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 6 + }, + { + "para_blocks": [ + { + "bbox": [ + 130, + 116, + 479, + 140 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 116, + 479, + 140 + ], + "spans": [ + { + "bbox": [ + 130, + 116, + 479, + 140 + ], + "type": "text", + "content": "We use I3D and VGGish to extract features from audio-visual content within the same time span (of 0.96 seconds - a default time span from VGGish)." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 130, + 140, + 479, + 176 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 140, + 479, + 176 + ], + "spans": [ + { + "bbox": [ + 130, + 140, + 479, + 176 + ], + "type": "text", + "content": "These representations are concatenated time-wise and we calculate multivariate Gaussians for both evaluation set embeddings, " + }, + { + "bbox": [ + 130, + 140, + 479, + 176 + ], + "type": "inline_equation", + "content": "N_{e}(\\mu_{e}, \\Sigma_{e})" + }, + { + "bbox": [ + 130, + 140, + 479, + 176 + ], + "type": "text", + "content": ", and the ground truth AV collection embeddings, " + }, + { + "bbox": [ + 130, + 140, + 479, + 176 + ], + "type": "inline_equation", + "content": "N_{av}(\\mu_{av}, \\Sigma_{av})" + }, + { + "bbox": [ + 130, + 140, + 479, + 176 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 130, + 176, + 479, + 202 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 176, + 479, + 202 + ], + "spans": [ + { + "bbox": [ + 130, + 176, + 479, + 202 + ], + "type": "text", + "content": "We then compute the Fréchet distance between these two Gaussians, as described by Dowson et al. [14]:" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 184, + 209, + 481, + 225 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 184, + 209, + 481, + 225 + ], + "spans": [ + { + "bbox": [ + 184, + 209, + 481, + 225 + ], + "type": "interline_equation", + "content": "F \\left(N _ {a v}, N _ {e}\\right) = \\left| \\mu_ {a v} - \\mu_ {e} \\right| ^ {2} + t r \\left(\\Sigma_ {a v} + \\Sigma_ {e} - 2 \\left(\\Sigma_ {a v} \\Sigma_ {e}\\right) ^ {\\frac {1}{2}}\\right) \\tag {1}", + "image_path": "cb0bae1ec0307ce07df19f85131ff45f1096924e7f494d8736469097b10d1261.jpg" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 130, + 229, + 479, + 253 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 229, + 479, + 253 + ], + "spans": [ + { + "bbox": [ + 130, + 229, + 479, + 253 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 130, + 229, + 479, + 253 + ], + "type": "inline_equation", + "content": "tr" + }, + { + "bbox": [ + 130, + 229, + 479, + 253 + ], + "type": "text", + "content": " is the trace of the matrix. For FAD and FVD, we compute the embeddings from respective modalities." + } + ] + } + ], + "index": 6 + }, + { + "type": "image", + "bbox": [ + 134, + 274, + 244, + 356 + ], + "blocks": [ + { + "bbox": [ + 134, + 274, + 244, + 356 + ], + "lines": [ + { + "bbox": [ + 134, + 274, + 244, + 356 + ], + "spans": [ + { + "bbox": [ + 134, + 274, + 244, + 356 + ], + "type": "image", + "image_path": "e24f3a313f158fa75cf27bfe3d6742143da122bc14727a4150345d9065c09640.jpg" + } + ] + } + ], + "index": 7, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 161, + 358, + 217, + 368 + ], + "lines": [ + { + "bbox": [ + 161, + 358, + 217, + 368 + ], + "spans": [ + { + "bbox": [ + 161, + 358, + 217, + 368 + ], + "type": "text", + "content": "(a) Audio Shift" + } + ] + } + ], + "index": 8, + "angle": 0, + "type": "image_caption" + } + ], + "index": 7 + }, + { + "type": "image", + "bbox": [ + 251, + 274, + 361, + 356 + ], + "blocks": [ + { + "bbox": [ + 251, + 274, + 361, + 356 + ], + "lines": [ + { + "bbox": [ + 251, + 274, + 361, + 356 + ], + "spans": [ + { + "bbox": [ + 251, + 274, + 361, + 356 + ], + "type": "image", + "image_path": "19b53ca4a5b3a1f64d0729a1c841858a1f974fac72ac23f5297b831e879c774e.jpg" + } + ] + } + ], + "index": 9, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 270, + 358, + 343, + 369 + ], + "lines": [ + { + "bbox": [ + 270, + 358, + 343, + 369 + ], + "spans": [ + { + "bbox": [ + 270, + 358, + 343, + 369 + ], + "type": "text", + "content": "(b) Audio Speed Up" + } + ] + } + ], + "index": 10, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 130, + 377, + 480, + 437 + ], + "lines": [ + { + "bbox": [ + 130, + 377, + 480, + 437 + ], + "spans": [ + { + "bbox": [ + 130, + 377, + 480, + 437 + ], + "type": "text", + "content": "Fig. 2: In these plots, we show the effect of distortions with varying levels on FAD, FVD, and FAVD for three distortion types. Flat trend line implies that a metric is not able to capture the distortion type. Increasing or decreasing trends show that a metric is susceptible to varying levels in a distortion. Distortion levels are taken from Table Detailed plots for all distortion types are available in Appendix C." + } + ] + } + ], + "index": 13, + "angle": 0, + "type": "image_caption" + } + ], + "index": 9 + }, + { + "type": "image", + "bbox": [ + 369, + 274, + 478, + 356 + ], + "blocks": [ + { + "bbox": [ + 369, + 274, + 478, + 356 + ], + "lines": [ + { + "bbox": [ + 369, + 274, + 478, + 356 + ], + "spans": [ + { + "bbox": [ + 369, + 274, + 478, + 356 + ], + "type": "image", + "image_path": "b787053c3e13812b46d4fac9920f659be36ebf9c11cf5f7972e046b478d3d119.jpg" + } + ] + } + ], + "index": 11, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 383, + 358, + 465, + 369 + ], + "lines": [ + { + "bbox": [ + 383, + 358, + 465, + 369 + ], + "spans": [ + { + "bbox": [ + 383, + 358, + 465, + 369 + ], + "type": "text", + "content": "(c) Audio Speed Down" + } + ] + } + ], + "index": 12, + "angle": 0, + "type": "image_caption" + } + ], + "index": 11 + }, + { + "bbox": [ + 130, + 456, + 482, + 599 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 456, + 482, + 599 + ], + "spans": [ + { + "bbox": [ + 130, + 456, + 482, + 599 + ], + "type": "text", + "content": "Examining the plots in Figure 2 it becomes evident that only FAVD (depicted in green) is responsive to various temporal distortions associated with AV synchrony across different levels. This is mainly due to use of covariance matrices " + }, + { + "bbox": [ + 130, + 456, + 482, + 599 + ], + "type": "inline_equation", + "content": "(\\Sigma_{av}, \\Sigma_e)" + }, + { + "bbox": [ + 130, + 456, + 482, + 599 + ], + "type": "text", + "content": " from audio-visual features (shown in Eq 1) which helps FAVD track the interaction of AV modality. Additionally, in cases like audio shift (a) where neither FAD nor FVD can detect synchronization issues (indicated by flat trend lines), and audio speed-up where FAD shows minimal movement, FAVD formulation exhibits a more pronounced reaction, highlighting that the fusion of features from I3D and VGGish models effectively captures temporality in an audio-visual context. Based on these observations, we chose to use I3D and VGGish features as inputs for our metric training, with FAVD serving as a baseline metric for the remainder of our work." + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 132, + 617, + 312, + 629 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 617, + 312, + 629 + ], + "spans": [ + { + "bbox": [ + 132, + 617, + 312, + 629 + ], + "type": "text", + "content": "5 PEAVS Score Formulation" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 130, + 641, + 479, + 665 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 641, + 479, + 665 + ], + "spans": [ + { + "bbox": [ + 130, + 641, + 479, + 665 + ], + "type": "text", + "content": "This section describes the architecture we have used to train our 'PEAVS' score metric and its training details." + } + ] + } + ], + "index": 16 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "spans": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "type": "text", + "content": "8" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 166, + 91, + 254, + 101 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 166, + 91, + 254, + 101 + ], + "spans": [ + { + "bbox": [ + 166, + 91, + 254, + 101 + ], + "type": "text", + "content": "Goncalves et al. 2024" + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 7 + }, + { + "para_blocks": [ + { + "bbox": [ + 132, + 116, + 225, + 126 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 116, + 225, + 126 + ], + "spans": [ + { + "bbox": [ + 132, + 116, + 225, + 126 + ], + "type": "text", + "content": "5.1 Architecture" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 130, + 134, + 482, + 230 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 134, + 482, + 230 + ], + "spans": [ + { + "bbox": [ + 130, + 134, + 482, + 230 + ], + "type": "text", + "content": "The input to our proposed model are the video and audio streams, as seen in Figure 3. For feature extraction, we leverage I3D and VGGish. In addition to this, our design draws inspiration from the audio-visual framework presented in [16]. Central to our architecture is the utilization of the cross-modal transformer formulation [36], which excels in capturing temporal information and relationships between different modalities. To ensure temporal consistency across modalities, both audio and visual input features are extracted over matched time spans and are then relayed through cross-modal transformer layers." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 130, + 231, + 482, + 350 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 231, + 482, + 350 + ], + "spans": [ + { + "bbox": [ + 130, + 231, + 482, + 350 + ], + "type": "text", + "content": "There are several key components in the our architecture. First, the audio and visual inputs are processed through our frozen backbone feature extractors. Both features are extracted over 0.96 second windows. The I3D extracted inputs " + }, + { + "bbox": [ + 130, + 231, + 482, + 350 + ], + "type": "inline_equation", + "content": "x_{v}" + }, + { + "bbox": [ + 130, + 231, + 482, + 350 + ], + "type": "text", + "content": " have dimensionality " + }, + { + "bbox": [ + 130, + 231, + 482, + 350 + ], + "type": "inline_equation", + "content": "x_{v} \\in \\mathbb{R}^{N_{av} \\times 1024}" + }, + { + "bbox": [ + 130, + 231, + 482, + 350 + ], + "type": "text", + "content": ", where " + }, + { + "bbox": [ + 130, + 231, + 482, + 350 + ], + "type": "inline_equation", + "content": "N_{av}" + }, + { + "bbox": [ + 130, + 231, + 482, + 350 + ], + "type": "text", + "content": " is the dimension of the audiovisual feature sequence, and 1,024 is the feature vector dimension. The acoustic feature vector " + }, + { + "bbox": [ + 130, + 231, + 482, + 350 + ], + "type": "inline_equation", + "content": "x_{a}" + }, + { + "bbox": [ + 130, + 231, + 482, + 350 + ], + "type": "text", + "content": " extracted from the VGGish model is " + }, + { + "bbox": [ + 130, + 231, + 482, + 350 + ], + "type": "inline_equation", + "content": "x_{a} \\in \\mathbb{R}^{N_{av} \\times 128}" + }, + { + "bbox": [ + 130, + 231, + 482, + 350 + ], + "type": "text", + "content": ", where " + }, + { + "bbox": [ + 130, + 231, + 482, + 350 + ], + "type": "inline_equation", + "content": "N_{av}" + }, + { + "bbox": [ + 130, + 231, + 482, + 350 + ], + "type": "text", + "content": " is the dimension of the acoustic feature sequence, and 128 is the feature vector dimension. We ensure that all inputs to the cross-modal layers have same dimensionality, by projecting the visual features from 1,024 to 128 through the use of a 1D convolutional layer to produce " + }, + { + "bbox": [ + 130, + 231, + 482, + 350 + ], + "type": "inline_equation", + "content": "\\bar{x}_{v} \\in \\mathbb{R}^{N_{v} \\times 128}" + }, + { + "bbox": [ + 130, + 231, + 482, + 350 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 130, + 350, + 482, + 518 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 350, + 482, + 518 + ], + "spans": [ + { + "bbox": [ + 130, + 350, + 482, + 518 + ], + "type": "text", + "content": "Next, we introduce positional embeddings to the feature vectors, following the step adopted in the original transformer paper [38]. The resultant features are used by the model to compute Query, Key, and Value vectors for each modality, as illustrated in Figure [3]. In our model, the Query vector from one modality is paired in the transformer layer with the Key and Value vectors of the other, ensuring a cross-modal interaction. This design choice allows the cross-modal transformer encoder layers for each modality branch to produce representations for a given modality that are influenced by the other modality. Following this, the output of the cross-modal layers are refined through the self-attention layers to enhance the cross-modal representations. The resultant features are then processed by averaging the outputs from the self-attention layers followed by concatenation of these averaged representations from both modality branches. Finally, a multilayer perceptron (MLP) with three fully-connected layers is integrated to produce a prediction score based on these concatenated features." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 132, + 534, + 204, + 547 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 534, + 204, + 547 + ], + "spans": [ + { + "bbox": [ + 132, + 534, + 204, + 547 + ], + "type": "text", + "content": "5.2 Training" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 130, + 553, + 482, + 613 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 553, + 482, + 613 + ], + "spans": [ + { + "bbox": [ + 130, + 553, + 482, + 613 + ], + "type": "text", + "content": "As seen in Figure 3 our training method consists of a two-stage strategy. In stage 1, we pre-train the model using a contrastive learning task, distinguishing between aligned and non-aligned pairs. Afterwards, in Stage Two, we fine-tune the model on our primary task, which involves predicting scores based on human annotators scores." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 130, + 629, + 482, + 666 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 629, + 482, + 666 + ], + "spans": [ + { + "bbox": [ + 130, + 629, + 482, + 666 + ], + "type": "text", + "content": "Stage 1 In stage 1, we pre-train our model using a contrastive learning objective. In this phase, we used AudioSet's Balanced train split [15], which after downloading all videos possible and removing corrupted ones resulted in 17K" + } + ] + } + ], + "index": 8 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 244, + 91, + 447, + 103 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 244, + 91, + 447, + 103 + ], + "spans": [ + { + "bbox": [ + 244, + 91, + 447, + 103 + ], + "type": "text", + "content": "Perceptual Evaluation of Audio-Visual Synchrony" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 474, + 92, + 481, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 474, + 92, + 481, + 100 + ], + "spans": [ + { + "bbox": [ + 474, + 92, + 481, + 100 + ], + "type": "text", + "content": "9" + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 8 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 161, + 114, + 452, + 224 + ], + "blocks": [ + { + "bbox": [ + 161, + 114, + 452, + 224 + ], + "lines": [ + { + "bbox": [ + 161, + 114, + 452, + 224 + ], + "spans": [ + { + "bbox": [ + 161, + 114, + 452, + 224 + ], + "type": "image", + "image_path": "a4001f608f38d265a3dcc211dedc86cad65cc81f032da7600e6c5e9c9af145ff.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 246, + 234, + 367, + 245 + ], + "lines": [ + { + "bbox": [ + 246, + 234, + 367, + 245 + ], + "spans": [ + { + "bbox": [ + 246, + 234, + 367, + 245 + ], + "type": "text", + "content": "Fig. 3: Framework Overview" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_caption" + } + ], + "index": 2 + }, + { + "bbox": [ + 130, + 265, + 482, + 399 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 265, + 482, + 399 + ], + "spans": [ + { + "bbox": [ + 130, + 265, + 482, + 399 + ], + "type": "text", + "content": "videos for our contrastive training stage We then subjected each of these files to ten levels of temporal misalignment (audio shift), as specified in section 3.1. This pre-processing resulted in a total of " + }, + { + "bbox": [ + 130, + 265, + 482, + 399 + ], + "type": "inline_equation", + "content": "\\sim 190\\mathrm{K}" + }, + { + "bbox": [ + 130, + 265, + 482, + 399 + ], + "type": "text", + "content": " files for the contrastive learning phase. Similar to SyncNet [11], the training objective ensures that the outputs of both the audio and video branches are similar for aligned pairs and dissimilar for misaligned pairs. Specifically, we either minimize or maximize the Euclidean distance between the network outputs for these cases, respectively. Our objective consists of a contrastive loss (Equation 2). The primary intention behind this pre-training is to embed regularization into the audio and visual features extracted from our frozen backbone networks. This ensures that the model prioritizes temporal attributes of the features over their qualitative characteristics." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 195, + 407, + 480, + 439 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 195, + 407, + 480, + 439 + ], + "spans": [ + { + "bbox": [ + 195, + 407, + 480, + 439 + ], + "type": "interline_equation", + "content": "\\mathcal {L} _ {c} = \\frac {1}{2 N} \\sum_ {n = 1} ^ {N} \\left(y _ {n}\\right) D _ {n} ^ {2} + \\left(1 - y _ {n}\\right) \\left(\\max \\left(0, m - D _ {n}\\right)\\right) ^ {2} \\tag {2}", + "image_path": "81b378421651f84426fc906cb676988c814d79f27e6c79bd2973ef64b8747f26.jpg" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 130, + 441, + 482, + 478 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 441, + 482, + 478 + ], + "spans": [ + { + "bbox": [ + 130, + 441, + 482, + 478 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 130, + 441, + 482, + 478 + ], + "type": "inline_equation", + "content": "y \\in \\{0,1\\}" + }, + { + "bbox": [ + 130, + 441, + 482, + 478 + ], + "type": "text", + "content": " is the label (1 for similar pairs and 0 for dissimilar pairs), " + }, + { + "bbox": [ + 130, + 441, + 482, + 478 + ], + "type": "inline_equation", + "content": "D_w" + }, + { + "bbox": [ + 130, + 441, + 482, + 478 + ], + "type": "text", + "content": " is the Euclidean distance between the two data points in the pair after passing them through the network, and " + }, + { + "bbox": [ + 130, + 441, + 482, + 478 + ], + "type": "inline_equation", + "content": "m" + }, + { + "bbox": [ + 130, + 441, + 482, + 478 + ], + "type": "text", + "content": " is the margin." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 130, + 492, + 483, + 589 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 492, + 483, + 589 + ], + "spans": [ + { + "bbox": [ + 130, + 492, + 483, + 589 + ], + "type": "text", + "content": "Stage 2 In stage 2 the output modality heads of the pre-trained model are concatenated together to produce audio-visual representations. Three fully-connected layers are added on top of these concatenated heads. We then fine tune the entire model for the downstream task of predicting human-alignment scores. Since the aggregated human annotation scores are non-negative real numbers, we approach model training as a regression task and utilize concordance correlation coefficient (CCC) to measure the agreement between the true and predicted scores. The CCC measurement is illustrated in Equation 3" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 238, + 593, + 480, + 619 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 238, + 593, + 480, + 619 + ], + "spans": [ + { + "bbox": [ + 238, + 593, + 480, + 619 + ], + "type": "interline_equation", + "content": "\\mathcal {L} _ {C C C} = \\frac {2 \\rho \\sigma_ {x} \\sigma_ {y}}{\\sigma_ {x} ^ {2} + \\sigma_ {y} ^ {2} + (\\mu_ {x} - \\mu_ {y}) ^ {2}} \\tag {3}", + "image_path": "37717495cda242fcbaeba55f4f24cf55fae70bb29a4d754823cb1ef73b1cc9f1.jpg" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 130, + 623, + 482, + 650 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 623, + 482, + 650 + ], + "spans": [ + { + "bbox": [ + 130, + 623, + 482, + 650 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 130, + 623, + 482, + 650 + ], + "type": "inline_equation", + "content": "\\mu_{x}" + }, + { + "bbox": [ + 130, + 623, + 482, + 650 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 130, + 623, + 482, + 650 + ], + "type": "inline_equation", + "content": "\\mu_{y}" + }, + { + "bbox": [ + 130, + 623, + 482, + 650 + ], + "type": "text", + "content": " denote the means of the true and predicted scores of a batch, respectively; " + }, + { + "bbox": [ + 130, + 623, + 482, + 650 + ], + "type": "inline_equation", + "content": "\\sigma_{x}" + }, + { + "bbox": [ + 130, + 623, + 482, + 650 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 130, + 623, + 482, + 650 + ], + "type": "inline_equation", + "content": "\\sigma_{y}" + }, + { + "bbox": [ + 130, + 623, + 482, + 650 + ], + "type": "text", + "content": " represent the standard deviations of the true and predicted" + } + ] + } + ], + "index": 9 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "spans": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "type": "text", + "content": "10" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 166, + 91, + 253, + 100 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 166, + 91, + 253, + 100 + ], + "spans": [ + { + "bbox": [ + 166, + 91, + 253, + 100 + ], + "type": "text", + "content": "Goncalves et al. 2024" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 133, + 653, + 342, + 664 + ], + "type": "page_footnote", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 653, + 342, + 664 + ], + "spans": [ + { + "bbox": [ + 133, + 653, + 342, + 664 + ], + "type": "text", + "content": "Note that this is different from AVS benchmark." + } + ] + } + ], + "index": 10 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 9 + }, + { + "para_blocks": [ + { + "bbox": [ + 130, + 116, + 482, + 152 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 116, + 482, + 152 + ], + "spans": [ + { + "bbox": [ + 130, + 116, + 482, + 152 + ], + "type": "text", + "content": "scores of the batch, respectively; and " + }, + { + "bbox": [ + 130, + 116, + 482, + 152 + ], + "type": "inline_equation", + "content": "\\rho" + }, + { + "bbox": [ + 130, + 116, + 482, + 152 + ], + "type": "text", + "content": " is their Pearson's correlation coefficient. Our training objective is to maximize CCC, ensuring that the predicted scores closely correlate with the true scores." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 130, + 170, + 482, + 255 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 170, + 482, + 255 + ], + "spans": [ + { + "bbox": [ + 130, + 170, + 482, + 255 + ], + "type": "text", + "content": "Training Details In stage 1, we divided the video files obtained from AudioSet's Balanced train split [15] into an " + }, + { + "bbox": [ + 130, + 170, + 482, + 255 + ], + "type": "inline_equation", + "content": "80\\%" + }, + { + "bbox": [ + 130, + 170, + 482, + 255 + ], + "type": "text", + "content": " training set and a " + }, + { + "bbox": [ + 130, + 170, + 482, + 255 + ], + "type": "inline_equation", + "content": "20\\%" + }, + { + "bbox": [ + 130, + 170, + 482, + 255 + ], + "type": "text", + "content": " validation set. This " + }, + { + "bbox": [ + 130, + 170, + 482, + 255 + ], + "type": "inline_equation", + "content": "80/20" + }, + { + "bbox": [ + 130, + 170, + 482, + 255 + ], + "type": "text", + "content": " split is carried out on the 17K original videos and extrapolated to distorted samples. We pre-trained the model in stage 1 until the validation loss plateaued. In stage 2, we use the train/validation/testing sets provided in the AVS benchmark corpus. We fine-tuned the model on training data for 20 epochs, saving the best-performing model based on the validation set performance [9]." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 131, + 273, + 230, + 288 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 131, + 273, + 230, + 288 + ], + "spans": [ + { + "bbox": [ + 131, + 273, + 230, + 288 + ], + "type": "text", + "content": "6 Experiments" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 130, + 299, + 482, + 443 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 299, + 482, + 443 + ], + "spans": [ + { + "bbox": [ + 130, + 299, + 482, + 443 + ], + "type": "text", + "content": "In this section, we evaluate the performance of our trained model using human-evaluation scores. We compare the model's scores with those provided by human evaluators on the test split of AVS benchmark. To further assess the effectiveness of our cross-modal transformer architecture for metric training, we run two ablation experiments. First, to assess the impact of pre-training in stage 1, we directly train our model from scratch as in stage 2. This variant is termed \"Cross-Modal Base\" as it incorporates the same audiovisual cross-modal framework but excludes the contrastive learning pre-training phase. Secondly, to assess the importance of cross-modal framework, we test a version without cross-modal layers called \"Basic Transformer\". This version employs only the self-attention branches, differently from the cross-modal approach of sharing query vector from one modality to another." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 131, + 460, + 295, + 472 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 131, + 460, + 295, + 472 + ], + "spans": [ + { + "bbox": [ + 131, + 460, + 295, + 472 + ], + "type": "text", + "content": "6.1 Human Scores Correlation" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 130, + 481, + 482, + 625 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 481, + 482, + 625 + ], + "spans": [ + { + "bbox": [ + 130, + 481, + 482, + 625 + ], + "type": "text", + "content": "For the above models, we compute the correlation of their outputs with the human judgements using pearson correlation coefficient (PCC). Since Fréchet based metrics require a set of videos to calculate distance between multivariate Gaussians we compute set level correlations. As PEAVS and its variants produce a score for each video separately, we average the scores across the test set to compute set level scores[10] We also provide correlations at a much fine-grained clip level. Examining the results presented in Table 2 FAVD shows a relative improvement of " + }, + { + "bbox": [ + 130, + 481, + 482, + 625 + ], + "type": "inline_equation", + "content": "11\\%" + }, + { + "bbox": [ + 130, + 481, + 482, + 625 + ], + "type": "text", + "content": " (from 0.475 to 0.527) and " + }, + { + "bbox": [ + 130, + 481, + 482, + 625 + ], + "type": "inline_equation", + "content": "64\\%" + }, + { + "bbox": [ + 130, + 481, + 482, + 625 + ], + "type": "text", + "content": " (from 0.321 to 0.527) over FAD and FVD baselines setting a strong baseline for future work. This highlights FAVD's effectiveness in capturing AV synchrony issues. PEAVS model achieves a Pearson correlation of 0.794 at the Set Level, marking a substantial " + }, + { + "bbox": [ + 130, + 481, + 482, + 625 + ], + "type": "inline_equation", + "content": "51\\%" + }, + { + "bbox": [ + 130, + 481, + 482, + 625 + ], + "type": "text", + "content": " improvement over the robust FAVD baseline. Notably, when we analyze the" + } + ] + } + ], + "index": 7 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 244, + 91, + 447, + 103 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 244, + 91, + 447, + 103 + ], + "spans": [ + { + "bbox": [ + 244, + 91, + 447, + 103 + ], + "type": "text", + "content": "Perceptual Evaluation of Audio-Visual Synchrony" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 470, + 92, + 479, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 470, + 92, + 479, + 100 + ], + "spans": [ + { + "bbox": [ + 470, + 92, + 479, + 100 + ], + "type": "text", + "content": "11" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 131, + 631, + 481, + 643 + ], + "type": "page_footnote", + "angle": 0, + "lines": [ + { + "bbox": [ + 131, + 631, + 481, + 643 + ], + "spans": [ + { + "bbox": [ + 131, + 631, + 481, + 643 + ], + "type": "text", + "content": "For further details on the pre-training and fine-tuning settings, refer to appendix A." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 131, + 643, + 481, + 665 + ], + "type": "page_footnote", + "angle": 0, + "lines": [ + { + "bbox": [ + 131, + 643, + 481, + 665 + ], + "spans": [ + { + "bbox": [ + 131, + 643, + 481, + 665 + ], + "type": "text", + "content": "10 Each set is defined as a collection of videos showing same distortion type and level. In total we had 90 sets, i.e. 9 distortions x 10 levels." + } + ] + } + ], + "index": 9 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 10 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 166, + 114, + 449, + 282 + ], + "blocks": [ + { + "bbox": [ + 166, + 114, + 449, + 282 + ], + "lines": [ + { + "bbox": [ + 166, + 114, + 449, + 282 + ], + "spans": [ + { + "bbox": [ + 166, + 114, + 449, + 282 + ], + "type": "image", + "image_path": "986ade816d78df75e02b6cfa4103d131b70c0a0bab1c49b40d8d1ba00b13873b.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 130, + 289, + 482, + 344 + ], + "lines": [ + { + "bbox": [ + 130, + 289, + 482, + 344 + ], + "spans": [ + { + "bbox": [ + 130, + 289, + 482, + 344 + ], + "type": "text", + "content": "Fig. 4: PEAVS results with Human Evaluation Scores compared side-by-side for each distortion type and at different levels. PEAVS scores are represented by blue bars and Human Evaluation scores are represented by orange bars. The x-axis are marked by the parameter and its levels as shown in Table " + }, + { + "bbox": [ + 130, + 289, + 482, + 344 + ], + "type": "inline_equation", + "content": "\\boxed{1}" + }, + { + "bbox": [ + 130, + 289, + 482, + 344 + ], + "type": "inline_equation", + "content": "\\rho" + }, + { + "bbox": [ + 130, + 289, + 482, + 344 + ], + "type": "text", + "content": " is the correlation of PEAVS with human judgements." + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_caption" + } + ], + "index": 2 + }, + { + "bbox": [ + 130, + 366, + 482, + 428 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 366, + 482, + 428 + ], + "spans": [ + { + "bbox": [ + 130, + 366, + 482, + 428 + ], + "type": "text", + "content": "impact of pre-training on PEAVS (c.f. Cross-Modal Base), performance drops by " + }, + { + "bbox": [ + 130, + 366, + 482, + 428 + ], + "type": "inline_equation", + "content": "7\\%" + }, + { + "bbox": [ + 130, + 366, + 482, + 428 + ], + "type": "text", + "content": " & " + }, + { + "bbox": [ + 130, + 366, + 482, + 428 + ], + "type": "inline_equation", + "content": "11\\%" + }, + { + "bbox": [ + 130, + 366, + 482, + 428 + ], + "type": "text", + "content": " at set and clip level, respectively, underscoring the significance of pre-training in PEAVS. Furthermore, removing the cross-modal layers (c.f. Base Transformer) results in a performance decrease of " + }, + { + "bbox": [ + 130, + 366, + 482, + 428 + ], + "type": "inline_equation", + "content": "9.4\\%" + }, + { + "bbox": [ + 130, + 366, + 482, + 428 + ], + "type": "text", + "content": " & " + }, + { + "bbox": [ + 130, + 366, + 482, + 428 + ], + "type": "inline_equation", + "content": "19\\%" + }, + { + "bbox": [ + 130, + 366, + 482, + 428 + ], + "type": "text", + "content": " at set and clip level, respectively, emphasizing the critical role of cross-modality layers." + } + ] + } + ], + "index": 4 + }, + { + "type": "table", + "bbox": [ + 219, + 446, + 394, + 528 + ], + "blocks": [ + { + "bbox": [ + 219, + 446, + 394, + 528 + ], + "lines": [ + { + "bbox": [ + 219, + 446, + 394, + 528 + ], + "spans": [ + { + "bbox": [ + 219, + 446, + 394, + 528 + ], + "type": "table", + "html": "
Model TypeSet LevelClip Level
FVD370.321-
FAD230.475-
FAVD (Ours)0.527-
Basic Transformer0.7200.434
Cross-Modal Base0.7360.477
PEAVS (Ours)0.7940.536
", + "image_path": "10821540b8ee629036b50107e9f8469f426f2801e5d6a7ada44758151c815b92.jpg" + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "table_body" + } + ], + "index": 5 + }, + { + "bbox": [ + 153, + 527, + 459, + 538 + ], + "lines": [ + { + "bbox": [ + 153, + 527, + 459, + 538 + ], + "spans": [ + { + "bbox": [ + 153, + 527, + 459, + 538 + ], + "type": "text", + "content": "Table 2: Correlation scores of different model types at Set and Clip levels." + } + ] + } + ], + "index": 6, + "angle": 0, + "type": "text" + }, + { + "bbox": [ + 131, + 587, + 279, + 599 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 131, + 587, + 279, + 599 + ], + "spans": [ + { + "bbox": [ + 131, + 587, + 279, + 599 + ], + "type": "text", + "content": "6.2 Per Distortion Analysis" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 130, + 605, + 482, + 665 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 605, + 482, + 665 + ], + "spans": [ + { + "bbox": [ + 130, + 605, + 482, + 665 + ], + "type": "text", + "content": "In this section, we analyze the response of PEAVS to varying distortions at different levels. Figure 4 factorizes the results in Table 2 by each distortion type and level. PEAVS scores are shown side-by-side with the scores from human annotations on AVS benchmark test set. Generally, PEAVS metric aligns well with scores from human judgements across the board except in a few cases." + } + ] + } + ], + "index": 8 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "spans": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "type": "text", + "content": "12" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 166, + 91, + 253, + 100 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 166, + 91, + 253, + 100 + ], + "spans": [ + { + "bbox": [ + 166, + 91, + 253, + 100 + ], + "type": "text", + "content": "Goncalves et al. 2024" + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 11 + }, + { + "para_blocks": [ + { + "bbox": [ + 130, + 115, + 482, + 187 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 115, + 482, + 187 + ], + "spans": [ + { + "bbox": [ + 130, + 115, + 482, + 187 + ], + "type": "text", + "content": "In top-right corner of Figure 4 we present the effects of audio shift. For ease of analysis, we display these results in terms of absolute shift values. We start with a shift of 0.045 based on a report from ITU which indicated that the threshold for audio-visual shift detection by human ranges from -125ms to +45ms. [13]. At this level, both the metric's output and human scores peak (>4), and they subsequently decrease as the shift increases." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 130, + 189, + 482, + 247 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 189, + 482, + 247 + ], + "spans": [ + { + "bbox": [ + 130, + 189, + 482, + 247 + ], + "type": "text", + "content": "In contrast to audio-shift, PEAVS shows the highest correlation for intermittent muting distortion which implies it is easier for the metric to capture this distortion. This confirms the notion that we observed in Section 3.2 where intermittent muting stood out as the most noticeable distortion and it was easier for humans to detect it as well." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 130, + 249, + 481, + 285 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 249, + 481, + 285 + ], + "spans": [ + { + "bbox": [ + 130, + 249, + 481, + 285 + ], + "type": "text", + "content": "In case of the fragment shuffling (bottom-right corner), initial levels of perturbation — shuffled fragments of 0.3-0.4 seconds — is significantly more disruptive for humans than the metric, i.e. the metric fails to capture this insight." + } + ] + } + ], + "index": 4 + }, + { + "type": "table", + "bbox": [ + 184, + 306, + 258, + 356 + ], + "blocks": [ + { + "bbox": [ + 184, + 306, + 258, + 356 + ], + "lines": [ + { + "bbox": [ + 184, + 306, + 258, + 356 + ], + "spans": [ + { + "bbox": [ + 184, + 306, + 258, + 356 + ], + "type": "table", + "html": "
Predicted
Pos.Neg.
ActualPos39227
Neg4130
", + "image_path": "8ae076f098d0fbb44b62f391afa32ba6d1439b87d8db9097a459c6d0a94201e2.jpg" + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "table_body" + } + ], + "index": 5 + }, + { + "type": "table", + "bbox": [ + 358, + 307, + 430, + 356 + ], + "blocks": [ + { + "bbox": [ + 358, + 307, + 430, + 356 + ], + "lines": [ + { + "bbox": [ + 358, + 307, + 430, + 356 + ], + "spans": [ + { + "bbox": [ + 358, + 307, + 430, + 356 + ], + "type": "table", + "html": "
Predicted
Pos.Neg.
ActualPos92174
Neg24110
", + "image_path": "2a144fd31dfa35b7333f52e3ea15f81349ceda3227c900d35a45a2ef978ecbca.jpg" + } + ] + } + ], + "index": 6, + "angle": 0, + "type": "table_body" + } + ], + "index": 6 + }, + { + "bbox": [ + 130, + 357, + 483, + 379 + ], + "lines": [ + { + "bbox": [ + 130, + 357, + 483, + 379 + ], + "spans": [ + { + "bbox": [ + 130, + 357, + 483, + 379 + ], + "type": "text", + "content": "Table 3: Confusion matrix for SparseSync (left) and PEAVS (right). Accuracy of SparseSync = 42.3% v/s PEAVS = 50.5%" + } + ] + } + ], + "index": 7, + "angle": 0, + "type": "text" + }, + { + "bbox": [ + 131, + 406, + 278, + 418 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 131, + 406, + 278, + 418 + ], + "spans": [ + { + "bbox": [ + 131, + 406, + 278, + 418 + ], + "type": "text", + "content": "6.3 PEAVS vs. SparseSync" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 130, + 429, + 482, + 536 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 429, + 482, + 536 + ], + "spans": [ + { + "bbox": [ + 130, + 429, + 482, + 536 + ], + "type": "text", + "content": "In this section, we report an experiment comparing PEAVS with SparseSync, which was introduced by Iashin et al. [21] for detecting audio-visual shifts inside 'in the wild' videos. While SparseSync's goal is to objectively quantify specific synchronization issues, by estimating an audio-visual offset shift value, PEAVS evaluates synchronization under multiple facets and according to a perceptual scale. Due to their diverging objectives, a direct comparison is challenging. Moreover, SparseSync is trained on 21 classes representing distinct audio-shift levels from -2.0 sec (left) to 2.0 sec (right) with increments of 0.2, with 0.0 representing no audio shift; while PEAVS is trained to predict float values between 1 and 5." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 130, + 538, + 482, + 634 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 538, + 482, + 634 + ], + "spans": [ + { + "bbox": [ + 130, + 538, + 482, + 634 + ], + "type": "text", + "content": "In this experiment, we assess both model's performance in accurately classifying 400 videos, including 200 ground truth videos and 200 randomly audio-shifted versions of them. We consider distortions of 0.045, 0.1 and " + }, + { + "bbox": [ + 130, + 538, + 482, + 634 + ], + "type": "inline_equation", + "content": "\\pm 0.125" + }, + { + "bbox": [ + 130, + 538, + 482, + 634 + ], + "type": "text", + "content": " as positive cases (i.e. ground truth) as ITU recommend an acceptability threshold for AV shift up to " + }, + { + "bbox": [ + 130, + 538, + 482, + 634 + ], + "type": "inline_equation", + "content": "185\\mathrm{ms}" + }, + { + "bbox": [ + 130, + 538, + 482, + 634 + ], + "type": "text", + "content": ". This also aligns well with SparseSync buckets where first distortion starts at " + }, + { + "bbox": [ + 130, + 538, + 482, + 634 + ], + "type": "inline_equation", + "content": "\\pm 0.2" + }, + { + "bbox": [ + 130, + 538, + 482, + 634 + ], + "type": "text", + "content": " seconds (i.e. SparseSync's step-size). To ensure a fair assessment, we divided the PEAVS scale into 21 bins, matching the output structure of SparseSync. For ground truth videos, we anticipate that" + } + ] + } + ], + "index": 10 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 244, + 91, + 448, + 102 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 244, + 91, + 448, + 102 + ], + "spans": [ + { + "bbox": [ + 244, + 91, + 448, + 102 + ], + "type": "text", + "content": "Perceptual Evaluation of Audio-Visual Synchrony" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 470, + 91, + 481, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 470, + 91, + 481, + 100 + ], + "spans": [ + { + "bbox": [ + 470, + 91, + 481, + 100 + ], + "type": "text", + "content": "13" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 130, + 641, + 410, + 655 + ], + "type": "page_footnote", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 641, + 410, + 655 + ], + "spans": [ + { + "bbox": [ + 130, + 641, + 410, + 655 + ], + "type": "text", + "content": "11 More details about this evaluation set are in Appendix Section D" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 130, + 654, + 284, + 666 + ], + "type": "page_footnote", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 654, + 284, + 666 + ], + "spans": [ + { + "bbox": [ + 130, + 654, + 284, + 666 + ], + "type": "text", + "content": "12 We use increments of 0.238 " + }, + { + "bbox": [ + 130, + 654, + 284, + 666 + ], + "type": "inline_equation", + "content": "(= \\frac{5}{21})" + }, + { + "bbox": [ + 130, + 654, + 284, + 666 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 12 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 12 + }, + { + "para_blocks": [ + { + "bbox": [ + 130, + 116, + 479, + 140 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 116, + 479, + 140 + ], + "spans": [ + { + "bbox": [ + 130, + 116, + 479, + 140 + ], + "type": "text", + "content": "SparseSync will classify them as having zero shifts (0.0 class), while PEAVS is expected to assign them to the highest scoring bin (i.e. (4.76, 5])." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 130, + 140, + 480, + 187 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 140, + 480, + 187 + ], + "spans": [ + { + "bbox": [ + 130, + 140, + 480, + 187 + ], + "type": "text", + "content": "In Table3 we show the confusion matrices for SparseSync (left) and PEAVS (right). In terms of accuracy, PEAVS outperforms SparseSync by " + }, + { + "bbox": [ + 130, + 140, + 480, + 187 + ], + "type": "inline_equation", + "content": "19\\%" + }, + { + "bbox": [ + 130, + 140, + 480, + 187 + ], + "type": "text", + "content": ", i.e. " + }, + { + "bbox": [ + 130, + 140, + 480, + 187 + ], + "type": "inline_equation", + "content": "50.5\\%" + }, + { + "bbox": [ + 130, + 140, + 480, + 187 + ], + "type": "text", + "content": " vs. " + }, + { + "bbox": [ + 130, + 140, + 480, + 187 + ], + "type": "inline_equation", + "content": "42.3\\%" + }, + { + "bbox": [ + 130, + 140, + 480, + 187 + ], + "type": "text", + "content": ". In particular, SparseSync shows an " + }, + { + "bbox": [ + 130, + 140, + 480, + 187 + ], + "type": "inline_equation", + "content": "18\\%" + }, + { + "bbox": [ + 130, + 140, + 480, + 187 + ], + "type": "text", + "content": " edge in detecting distortions while PEAVS shows a " + }, + { + "bbox": [ + 130, + 140, + 480, + 187 + ], + "type": "inline_equation", + "content": "136\\%" + }, + { + "bbox": [ + 130, + 140, + 480, + 187 + ], + "type": "text", + "content": " higher rate in detecting clean content." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 132, + 205, + 222, + 217 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 205, + 222, + 217 + ], + "spans": [ + { + "bbox": [ + 132, + 205, + 222, + 217 + ], + "type": "text", + "content": "7 Limitations" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 130, + 229, + 479, + 252 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 229, + 479, + 252 + ], + "spans": [ + { + "bbox": [ + 130, + 229, + 479, + 252 + ], + "type": "text", + "content": "Like any research, our work is not without its limitations, and we have identified a couple of them here:" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 130, + 254, + 480, + 312 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 254, + 480, + 312 + ], + "spans": [ + { + "bbox": [ + 130, + 254, + 480, + 312 + ], + "type": "text", + "content": "Dataset: Due to budget constraints (30K USD), we could only source 200 videos and conduct 60K pair-wise annotations. In selecting these 200 videos, we ensured diversity. Our dataset comprises 215 labels out of the 512 present in AudioSet, representing only " + }, + { + "bbox": [ + 130, + 254, + 480, + 312 + ], + "type": "inline_equation", + "content": "40\\%" + }, + { + "bbox": [ + 130, + 254, + 480, + 312 + ], + "type": "text", + "content": " of labels. Expanding this dataset for broader video coverage is left for future work." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 130, + 313, + 480, + 372 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 313, + 480, + 372 + ], + "spans": [ + { + "bbox": [ + 130, + 313, + 480, + 372 + ], + "type": "text", + "content": "Metric: The limited size of the AVS benchmark highlights the necessity of pretraining in model-based metrics (in our case, PEAVS). This is further evidenced by the results in Table 2 where pre-training improves correlation from 0.736 (Cross-Modal Base) to 0.794 (PEAVS). In this study, we only pre-trained PEAVS on audio-shift noise, leaving pre-training on other noise types for future work." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 130, + 373, + 479, + 409 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 373, + 479, + 409 + ], + "spans": [ + { + "bbox": [ + 130, + 373, + 479, + 409 + ], + "type": "text", + "content": "\"In the wild\" videos: Due to proprietary issues, we could not experiment with datasets featuring talking faces like VoxCeleb 30 or Lip Reading Sentences [1]. Our study is confined to addressing synchrony issues for \"in the wild\" videos." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 132, + 426, + 225, + 438 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 426, + 225, + 438 + ], + "spans": [ + { + "bbox": [ + 132, + 426, + 225, + 438 + ], + "type": "text", + "content": "8 Conclusions" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 130, + 449, + 480, + 593 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 449, + 480, + 593 + ], + "spans": [ + { + "bbox": [ + 130, + 449, + 480, + 593 + ], + "type": "text", + "content": "We introduced PEAVS, a novel metric designed for measuring synchrony in audio-visual content, an important advancement in the field of evaluation of audio-visual content. While the domain of audio-visual generative modeling is ever-evolving, assessing synchronization remains crucial for enhancing user experience. Although numerous metrics cater to either audio or visual content separately, a noticeable gap exists in evaluating their synchronization with a focus on viewers opinions. Our study addresses this gap with PEAVS, a metric based on a vast human-annotated dataset. The strong correlation between PEAVS scores and human evaluations confirms its efficacy in capturing perceptions surrounding audio-visual synchronization. As the scope of audio-visual content generation expands, tools like PEAVS are essential in maintaining a balanced interplay between sight and sound, guiding both creators and researchers." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 130, + 594, + 480, + 665 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 594, + 480, + 665 + ], + "spans": [ + { + "bbox": [ + 130, + 594, + 480, + 665 + ], + "type": "text", + "content": "Furthermore, we introduced the Audio-Visual Synchrony human perception benchmark, which provides a new dataset detailing human-perceptual scores on various audio-visual synchrony challenges for real-world videos. By isolating specific challenges in a controlled audio-visual environment, we aim to simplify the task for researchers to propose, assess, and analyze potential solutions while maintaining agreement with human perception." + } + ] + } + ], + "index": 11 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "spans": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "type": "text", + "content": "14" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 166, + 91, + 253, + 100 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 166, + 91, + 253, + 100 + ], + "spans": [ + { + "bbox": [ + 166, + 91, + 253, + 100 + ], + "type": "text", + "content": "Goncalves et al. 2024" + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 13 + }, + { + "para_blocks": [ + { + "bbox": [ + 132, + 114, + 198, + 126 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 114, + 198, + 126 + ], + "spans": [ + { + "bbox": [ + 132, + 114, + 198, + 126 + ], + "type": "text", + "content": "References" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 134, + 139, + 481, + 666 + ], + "type": "list", + "angle": 0, + "index": 15, + "blocks": [ + { + "bbox": [ + 138, + 139, + 480, + 161 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 139, + 480, + 161 + ], + "spans": [ + { + "bbox": [ + 138, + 139, + 480, + 161 + ], + "type": "text", + "content": "1. Afouras, T., Chung, J.S., Zisserman, A.: Lrs3-ted: a large-scale dataset for visual speech recognition (2018)" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 138, + 162, + 481, + 326 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 162, + 481, + 326 + ], + "spans": [ + { + "bbox": [ + 138, + 162, + 481, + 326 + ], + "type": "text", + "content": "2. Agarwal, M., Agrawal, S., Anastasopoulos, A., Bentivogli, L., Bojar, O., Borg, C., Carpuat, M., Cattoni, R., Cettolo, M., Chen, M., Chen, W., Choukri, K., Chronopoulou, A., Currey, A., Declerck, T., Dong, Q., Duh, K., Esteve, Y., Federico, M., Gahbiche, S., Haddow, B., Hsu, B., Mon Htut, P., Inaguma, H., Javorsky, D., Judge, J., Kano, Y., Ko, T., Kumar, R., Li, P., Ma, X., Mathur, P., Matusov, E., McNamee, P., P. McCrae, J., Murray, K., Nadejde, M., Nakamura, S., Negri, M., Nguyen, H., Niehues, J., Niu, X., Kr. Ojha, A., E. Ortega, J., Pal, P., Pino, J., van der Plas, L., Polak, P., Ripbeth, E., Salesky, E., Shi, J., Sperber, M., Stüker, S., Sudoh, K., Tang, Y., Thompson, B., Tran, K., Turchi, M., Waibel, A., Wang, M., Watanabe, S., Zevallos, R.: FINDINGS OF THE IWSLT 2023 EVALUATION CAMPAIGN. In: Proceedings of the 20th International Conference on Spoken Language Translation (IWSLT 2023). pp. 1-61. Association for Computational Linguistics, Toronto, Canada (in-person and online) (Jul 2023). https://doi.org/10.18653/v1/2023.iwslt-1.1 https://aclanthology.org/2023.iwslt-1.1" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 138, + 327, + 481, + 358 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 327, + 481, + 358 + ], + "spans": [ + { + "bbox": [ + 138, + 327, + 481, + 358 + ], + "type": "text", + "content": "3. Arandjelović, R., Zisserman, A.: Look, listen and learn. 2017 IEEE International Conference on Computer Vision (ICCV) pp. 609-617 (2017), https://apisemantic scholar.org/CorpusID:10769575" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 138, + 359, + 481, + 380 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 359, + 481, + 380 + ], + "spans": [ + { + "bbox": [ + 138, + 359, + 481, + 380 + ], + "type": "text", + "content": "4. Arandjelović, R., Zisserman, A.: Objects that sound. In: Computer Vision - ECCV 2018. pp. 451-466. Springer International Publishing, Cham (2018)" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 138, + 381, + 481, + 424 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 381, + 481, + 424 + ], + "spans": [ + { + "bbox": [ + 138, + 381, + 481, + 424 + ], + "type": "text", + "content": "5. Binkowski, M., Sutherland, D.J., Arbel, M., Gretton, A.: Demystifying MMD gans. In: 6th International Conference on Learning Representations, ICLR 2018, Vancouver, BC, Canada, April 30 - May 3, 2018, Conference Track Proceedings. Open-Review.net (2018), https://openreview.net/forum?id=r1lUOzWCw" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 138, + 425, + 481, + 446 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 425, + 481, + 446 + ], + "spans": [ + { + "bbox": [ + 138, + 425, + 481, + 446 + ], + "type": "text", + "content": "6. BT.1359, R.I.R.: BT.1359 - Relative timing of sound and vision for broadcasting (1998)" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 138, + 447, + 481, + 501 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 447, + 481, + 501 + ], + "spans": [ + { + "bbox": [ + 138, + 447, + 481, + 501 + ], + "type": "text", + "content": "7. Carreira, J., Zisserman, A.: Quo vadis, action recognition? a new model and the kinetics dataset. In: 2017 IEEE Conference on Computer Vision and Pattern Recognition (CVPR). pp. 4724-4733. IEEE Computer Society, Los Alamitos, CA, USA (jul 2017). https://doi.org/10.1109/CVPR.2017.502 https://doi.ieeecomputersociety.org/10.1109/CVPR.2017.502" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 138, + 502, + 481, + 522 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 502, + 481, + 522 + ], + "spans": [ + { + "bbox": [ + 138, + 502, + 481, + 522 + ], + "type": "text", + "content": "8. Chen, H., Xie, W., Afouras, T., Nagrani, A., Vedaldi, A., Zisserman, A.: Audiovisual synchronisation in the wild (2021)" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 138, + 523, + 481, + 544 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 523, + 481, + 544 + ], + "spans": [ + { + "bbox": [ + 138, + 523, + 481, + 544 + ], + "type": "text", + "content": "9. Chen, H., Xie, W., Vedaldi, A., Zisserman, A.: Vggsound: A large-scale audio-visual dataset (2020)" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 138, + 545, + 481, + 589 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 545, + 481, + 589 + ], + "spans": [ + { + "bbox": [ + 138, + 545, + 481, + 589 + ], + "type": "text", + "content": "0. Chinen, M., Lim, F.S.C., Skoglund, J., Gureev, N., O'Gorman, F., Hines, A.: Visqol v3: An open source production ready objective speech and audio metric. In: 2020 Twelfth International Conference on Quality of Multimedia Experience (QoMEX). pp. 1-6 (2020). https://doi.org/10.1109/QoMEX48832.2020.9123150" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 138, + 590, + 481, + 610 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 590, + 481, + 610 + ], + "spans": [ + { + "bbox": [ + 138, + 590, + 481, + 610 + ], + "type": "text", + "content": "1. Chung, J.S., Zisserman, A.: Out of time: automated lip sync in the wild. In: Workshop on Multi-view Lip-reading, ACCV (2016)" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 134, + 611, + 481, + 666 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 134, + 611, + 481, + 666 + ], + "spans": [ + { + "bbox": [ + 134, + 611, + 481, + 666 + ], + "type": "text", + "content": "12. Clark, A.P., Howard, K.L., Woods, A.T., Penton-Voak, I.S., Neumann, C.: Why rate when you could compare? Using the \"EloChoice\" package to assess pairwise comparisons of perceived physical strength. PLOS ONE 13(1), e0190393 (Jan 2018). https://doi.org/10.1371/journal.pone.0190393 https://dx.plos.org/10.1371/journal.pone.0190393" + } + ] + } + ], + "index": 14 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 244, + 91, + 447, + 102 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 244, + 91, + 447, + 102 + ], + "spans": [ + { + "bbox": [ + 244, + 91, + 447, + 102 + ], + "type": "text", + "content": "Perceptual Evaluation of Audio-Visual Synchrony" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 470, + 92, + 481, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 470, + 92, + 481, + 100 + ], + "spans": [ + { + "bbox": [ + 470, + 92, + 481, + 100 + ], + "type": "text", + "content": "15" + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 14 + }, + { + "para_blocks": [ + { + "bbox": [ + 133, + 116, + 481, + 666 + ], + "type": "list", + "angle": 0, + "index": 17, + "blocks": [ + { + "bbox": [ + 133, + 116, + 481, + 160 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 116, + 481, + 160 + ], + "spans": [ + { + "bbox": [ + 133, + 116, + 481, + 160 + ], + "type": "text", + "content": "13. Committee, A.T.S.: Atsc implementation subcommittee finding: Relative timing of sound and vision for broadcast operations. Tech. rep., Advanced Television Systems Committee, 1750 K Street, N.W., Suite 1200, Washington, D.C. 20006 (6 2003), doc. IS-191" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 133, + 161, + 481, + 205 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 161, + 481, + 205 + ], + "spans": [ + { + "bbox": [ + 133, + 161, + 481, + 205 + ], + "type": "text", + "content": "14. Dowson, D., Landau, B.: The frechet distance between multivariate normal distributions. Journal of Multivariate Analysis 12(3), 450-455 (1982). https://doi.org/https://doi.org/10.1016/0047-259X(82)90077-X, https://www.sciencedirect.com/science/article/pii/0047259X8290077X" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 134, + 206, + 481, + 261 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 134, + 206, + 481, + 261 + ], + "spans": [ + { + "bbox": [ + 134, + 206, + 481, + 261 + ], + "type": "text", + "content": "15. Gemmeke, J.F., Ellis, D.P.W., Freedman, D., Jansen, A., Lawrence, W., Moore, R.C., Plakal, M., Ritter, M.: Audio set: An ontology and human-labeled dataset for audio events. In: 2017 IEEE International Conference on Acoustics, Speech and Signal Processing (ICASSP). pp. 776-780 (2017). https://doi.org/10.1109/ICASSP.2017.7952261" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 134, + 262, + 481, + 306 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 134, + 262, + 481, + 306 + ], + "spans": [ + { + "bbox": [ + 134, + 262, + 481, + 306 + ], + "type": "text", + "content": "16. Goncalves, L., Busso, C.: Auxformer: Robust approach to audiovisual emotion recognition. In: ICASSP 2022 - 2022 IEEE International Conference on Acoustics, Speech and Signal Processing (ICASSP). pp. 7357-7361 (2022). https://doi.org/10.1109/ICASSP43922.2022.9747157" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 134, + 307, + 481, + 361 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 134, + 307, + 481, + 361 + ], + "spans": [ + { + "bbox": [ + 134, + 307, + 481, + 361 + ], + "type": "text", + "content": "17. Hershey, J., Movellan, J.: Audio vision: Using audio-visual synchrony to locate sounds. In: Solla, S., Leen, T., Müller, K. (eds.) Advances in Neural Information Processing Systems. vol. 12. MIT Press (1999), https://proceedings.neurips.cc/paper_files/paper/1999/file/b618c3210e934362ac261db280128c22-Paper.pdf" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 134, + 362, + 481, + 396 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 134, + 362, + 481, + 396 + ], + "spans": [ + { + "bbox": [ + 134, + 362, + 481, + 396 + ], + "type": "text", + "content": "18. Hershey, S., Chaudhuri, S., Ellis, D.P.W., Gemmeke, J.F., Jansen, A., Moore, R.C., Plakal, M., Platt, D., Saurous, R.A., Seybold, B., Slaney, M., Weiss, R.J., Wilson, K.: Cnn architectures for large-scale audio classification (2017)" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 134, + 396, + 481, + 418 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 134, + 396, + 481, + 418 + ], + "spans": [ + { + "bbox": [ + 134, + 396, + 481, + 418 + ], + "type": "text", + "content": "19. Hessel, J., Holtzman, A., Forbes, M., Bras, R.L., Choi, Y.: CLIPScore: a reference-free evaluation metric for image captioning. In: EMNLP (2021)" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 134, + 418, + 481, + 452 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 134, + 418, + 481, + 452 + ], + "spans": [ + { + "bbox": [ + 134, + 418, + 481, + 452 + ], + "type": "text", + "content": "20. Heusel, M., Ramsauer, H., Unterthiner, T., Nessler, B., Hochreiter, S.: Gans trained by a two time-scale update rule converge to a local nash equilibrium. In: Neural Information Processing Systems (2017)" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 134, + 452, + 481, + 485 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 134, + 452, + 481, + 485 + ], + "spans": [ + { + "bbox": [ + 134, + 452, + 481, + 485 + ], + "type": "text", + "content": "21. Iashin, V., Xie, W., Rahtu, E., Zisserman, A.: Sparse in space and time: Audiovisual synchronisation with trainable selectors. In: British Machine Vision Conference (BMVC) (2022)" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 134, + 486, + 481, + 508 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 134, + 486, + 481, + 508 + ], + "spans": [ + { + "bbox": [ + 134, + 486, + 481, + 508 + ], + "type": "text", + "content": "22. ITU-T RECOMMENDATION, P.: Subjective video quality assessment methods for multimedia applications (1999)" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 134, + 509, + 481, + 542 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 134, + 509, + 481, + 542 + ], + "spans": [ + { + "bbox": [ + 134, + 509, + 481, + 542 + ], + "type": "text", + "content": "23. Kilgour, K., Zuluaga, M., Roblek, D., Sharifi, M.: Fréchet audio distance: A reference-free metric for evaluating music enhancement algorithms. In: Interspeech (2019), https://api-semanticscholar.org/CorpusID:202725406" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 134, + 543, + 481, + 575 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 134, + 543, + 481, + 575 + ], + "spans": [ + { + "bbox": [ + 134, + 543, + 481, + 575 + ], + "type": "text", + "content": "24. Korhonen, J., You, J.: Korhonen. In: 2012 Fourth International Workshop on Quality of Multimedia Experience. pp. 37-38 (2012). https://doi.org/10.1109/QoMEX.2012.6263880" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 134, + 576, + 481, + 597 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 134, + 576, + 481, + 597 + ], + "spans": [ + { + "bbox": [ + 134, + 576, + 481, + 597 + ], + "type": "text", + "content": "25. Krippendorff, K.: Content Analysis: An Introduction to Its Methodology (second edition). Sage Publications (2004)" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 134, + 599, + 481, + 620 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 134, + 599, + 481, + 620 + ], + "spans": [ + { + "bbox": [ + 134, + 599, + 481, + 620 + ], + "type": "text", + "content": "26. Luo, S., Yan, C., Hu, C., Zhao, H.: Diff-foley: Synchronized video-to-audio synthesis with latent diffusion models (2023)" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 134, + 621, + 481, + 666 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 134, + 621, + 481, + 666 + ], + "spans": [ + { + "bbox": [ + 134, + 621, + 481, + 666 + ], + "type": "text", + "content": "27. Luo, Y., Mesgarani, N.: Tasnet: Time-domain audio separation network for real-time, single-channel speech separation. In: 2018 IEEE International Conference on Acoustics, Speech and Signal Processing (ICASSP). pp. 696-700 (2018). https://doi.org/10.1109/ICASSP.2018.8462116" + } + ] + } + ], + "index": 16 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "spans": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "type": "text", + "content": "16" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 166, + 91, + 253, + 100 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 166, + 91, + 253, + 100 + ], + "spans": [ + { + "bbox": [ + 166, + 91, + 253, + 100 + ], + "type": "text", + "content": "Goncalves et al. 2024" + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 15 + }, + { + "para_blocks": [ + { + "bbox": [ + 132, + 116, + 482, + 665 + ], + "type": "list", + "angle": 0, + "index": 15, + "blocks": [ + { + "bbox": [ + 132, + 116, + 482, + 171 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 116, + 482, + 171 + ], + "spans": [ + { + "bbox": [ + 132, + 116, + 482, + 171 + ], + "type": "text", + "content": "28. Manocha, P., Finkelstein, A., Zhang, R., Bryan, N.J., Mysore, G.J., Jin, Z.: A Differentiable Perceptual Audio Metric Learned from Just Noticeable Differences. In: Interspeech 2020. pp. 2852-2856. ISCA (Oct 2020). https://doi.org/10.21437/Interspeech.2020-1191 https://www.isca-speech.org/archive/interspeech_2020/manocha20_interspeech.html" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 132, + 171, + 481, + 214 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 171, + 481, + 214 + ], + "spans": [ + { + "bbox": [ + 132, + 171, + 481, + 214 + ], + "type": "text", + "content": "29. Manocha, P., Kumar, A., Xu, B., Menon, A., Gebru, I.D., Ithapu, V.K., Calamia, P.: Dplm: A deep perceptual spatial-audio localization metric. In: 2021 IEEE Workshop on Applications of Signal Processing to Audio and Acoustics (WASPAA). pp. 6-10 (2021). https://doi.org/10.1109/WASPAA52581.2021.9632781" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 132, + 214, + 481, + 236 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 214, + 481, + 236 + ], + "spans": [ + { + "bbox": [ + 132, + 214, + 481, + 236 + ], + "type": "text", + "content": "30. Nagrani, A., Chung, J.S., Zisserman, A.: Voxceleb: a large-scale speaker identification dataset. In: INTERSPEECH (2017)" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 132, + 236, + 481, + 289 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 236, + 481, + 289 + ], + "spans": [ + { + "bbox": [ + 132, + 236, + 481, + 289 + ], + "type": "text", + "content": "31. Rix, A., Beerends, J., Hollier, M., Hekstra, A.: Perceptual evaluation of speech quality (pesq)-a new method for speech quality assessment of telephone networks and codescs. In: 2001 IEEE International Conference on Acoustics, Speech, and Signal Processing. Proceedings (Cat. No.01CH37221). vol. 2, pp. 749-752 vol.2 (2001). https://doi.org/10.1109/ICASSP.2001.941023" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 132, + 289, + 481, + 331 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 289, + 481, + 331 + ], + "spans": [ + { + "bbox": [ + 132, + 289, + 481, + 331 + ], + "type": "text", + "content": "32. Roux, J.L., Wisdom, S., Erdogan, H., Hershey, J.R.: Sdr - half-baked or well done? In: ICASSP 2019 - 2019 IEEE International Conference on Acoustics, Speech and Signal Processing (ICASSP). pp. 626-630 (2019). https://doi.org/10.1109/ICASSP.2019.8683855" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 132, + 332, + 481, + 397 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 332, + 481, + 397 + ], + "spans": [ + { + "bbox": [ + 132, + 332, + 481, + 397 + ], + "type": "text", + "content": "33. Ruan, L., Ma, Y., Yang, H., He, H., Liu, B., Fu, J., Yuan, N., Jin, Q., Guo, B.: Mm-diffusion: Learning multi-modal diffusion models for joint audio and video generation. In: 2023 IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR). pp. 10219-10228. IEEE Computer Society, Los Alamitos, CA, USA (jun 2023). https://doi.org/10.1109/CVPR52729.2023.00985 https://doi.ieeeccomputersociety.org/10.1109/CVPR52729.2023.00985" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 132, + 397, + 481, + 440 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 397, + 481, + 440 + ], + "spans": [ + { + "bbox": [ + 132, + 397, + 481, + 440 + ], + "type": "text", + "content": "34. Salimans, T., Goodfellow, I., Zaremba, W., Cheung, V., Radford, A., Chen, X.: Improved techniques for training gans. In: Proceedings of the 30th International Conference on Neural Information Processing Systems. p. 2234-2242. NIPS'16, Curran Associates Inc., Red Hook, NY, USA (2016)" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 132, + 440, + 481, + 482 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 440, + 481, + 482 + ], + "spans": [ + { + "bbox": [ + 132, + 440, + 481, + 482 + ], + "type": "text", + "content": "35. Taal, C.H., Hendriks, R.C., Heusdens, R., Jensen, J.: An algorithm for intelligibility prediction of time-frequency weighted noisy speech. IEEE Transactions on Audio, Speech, and Language Processing 19(7), 2125-2136 (2011). https://doi.org/10.1109/TASL.2011.2114881" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 132, + 483, + 481, + 536 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 483, + 481, + 536 + ], + "spans": [ + { + "bbox": [ + 132, + 483, + 481, + 536 + ], + "type": "text", + "content": "36. Tsai, Y.H.H., Bai, S., Liang, P.P., Kolter, J.Z., Morency, L.P., Salakhutdinov, R.: Multimodal transformer for unaligned multimodal language sequences. In: Proceedings of the 57th Annual Meeting of the Association for Computational Linguistics (Volume 1: Long Papers). Association for Computational Linguistics, Florence, Italy (7 2019)" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 132, + 537, + 481, + 557 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 537, + 481, + 557 + ], + "spans": [ + { + "bbox": [ + 132, + 537, + 481, + 557 + ], + "type": "text", + "content": "37. Unterthiner, T., van Steenkiste, S., Kurach, K., Marinier, R., Michalski, M., Gelly, S.: Towards accurate generative models of video: A new metric & challenges (2019)" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 132, + 558, + 481, + 612 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 558, + 481, + 612 + ], + "spans": [ + { + "bbox": [ + 132, + 558, + 481, + 612 + ], + "type": "text", + "content": "38. Vaswani, A., Shazeer, N., Parmar, N., Uszkoreit, J., Jones, L., Gomez, A.N., Kaiser, L.u., Polosukhin, I.: Attention is all you need. In: Advances in Neural Information Processing Systems. vol. 30. Curran Associates, Inc. (2017), https://proceedings.neurips.cc/paper_files/paper/2017/file/3f5ee243547dee91fbd053c1c4a845aa-Paper.pdf" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 132, + 612, + 481, + 644 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 612, + 481, + 644 + ], + "spans": [ + { + "bbox": [ + 132, + 612, + 481, + 644 + ], + "type": "text", + "content": "39. Vincent, E., Gribonval, R., Fevotte, C.: Performance measurement in blind audio source separation. IEEE Transactions on Audio, Speech, and Language Processing 14(4), 1462-1469 (2006). https://doi.org/10.1109/TSA.2005.858005" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 132, + 644, + 481, + 665 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 644, + 481, + 665 + ], + "spans": [ + { + "bbox": [ + 132, + 644, + 481, + 665 + ], + "type": "text", + "content": "40. Wang, J., Fang, Z., Zhao, H.: Alignnet: A unifying approach to audio-visual alignment. In: WACV (2020)" + } + ] + } + ], + "index": 14 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 244, + 91, + 448, + 102 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 244, + 91, + 448, + 102 + ], + "spans": [ + { + "bbox": [ + 244, + 91, + 448, + 102 + ], + "type": "text", + "content": "Perceptual Evaluation of Audio-Visual Synchrony" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 470, + 92, + 481, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 470, + 92, + 481, + 100 + ], + "spans": [ + { + "bbox": [ + 470, + 92, + 481, + 100 + ], + "type": "text", + "content": "17" + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 16 + }, + { + "para_blocks": [ + { + "bbox": [ + 132, + 116, + 481, + 271 + ], + "type": "list", + "angle": 0, + "index": 6, + "blocks": [ + { + "bbox": [ + 132, + 116, + 481, + 162 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 116, + 481, + 162 + ], + "spans": [ + { + "bbox": [ + 132, + 116, + 481, + 162 + ], + "type": "text", + "content": "41. Wang, Y., Ke, J., Talebi, H., Yim, J.G., Birkbeck, N., Adsumilli, B., Milanfar, P., Yang, F.: Rich features for perceptual quality assessment ofUGC videos. In: 2021 IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR). pp. 13430-13439 (2021). https://doi.org/10.1109/CVPR46437.2021.01323" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 132, + 162, + 481, + 204 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 162, + 481, + 204 + ], + "spans": [ + { + "bbox": [ + 132, + 162, + 481, + 204 + ], + "type": "text", + "content": "42. Wang, Z., Simoncelli, E., Bovik, A.: Multiscale structural similarity for image quality assessment. In: The Thirty-Seventh Asilomar Conference on Signals, Systems & Computers, 2003. vol. 2, pp. 1398-1402 Vol.2 (2003). https://doi.org/10.1109/ACSSC.2003.1292216" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 132, + 205, + 481, + 227 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 205, + 481, + 227 + ], + "spans": [ + { + "bbox": [ + 132, + 205, + 481, + 227 + ], + "type": "text", + "content": "43. Wang, Z., Bovik, A.: A universal image quality index. IEEE Signal Processing Letters 9(3), 81-84 (2002). https://doi.org/10.1109/97.995823" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 132, + 227, + 481, + 271 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 227, + 481, + 271 + ], + "spans": [ + { + "bbox": [ + 132, + 227, + 481, + 271 + ], + "type": "text", + "content": "44. Zhang, R., Isola, P., Efros, A.A., Shechtman, E., Wang, O.: The unreasonable effectiveness of deep features as a perceptual metric. In: 2018 IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 586-595 (2018). https://doi.org/10.1109/CVPR.2018.00068" + } + ] + } + ], + "index": 5 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "spans": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "type": "text", + "content": "18" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 166, + 91, + 253, + 100 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 166, + 91, + 253, + 100 + ], + "spans": [ + { + "bbox": [ + 166, + 91, + 253, + 100 + ], + "type": "text", + "content": "Goncalves et al. 2024" + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 17 + } + ], + "_backend": "vlm", + "_version_name": "2.6.4" +} \ No newline at end of file diff --git a/2024/Personalized Federated Domain-Incremental Learning based on Adaptive Knowledge Matching/ea77b70c-2acd-46a7-86ea-e1ba20b951f3_content_list.json b/2024/Personalized Federated Domain-Incremental Learning based on Adaptive Knowledge Matching/ea77b70c-2acd-46a7-86ea-e1ba20b951f3_content_list.json new file mode 100644 index 0000000000000000000000000000000000000000..5aceb2167bef340d0326aa1ac6efcbfe3451ce96 --- /dev/null +++ b/2024/Personalized Federated Domain-Incremental Learning based on Adaptive Knowledge Matching/ea77b70c-2acd-46a7-86ea-e1ba20b951f3_content_list.json @@ -0,0 +1,1571 @@ +[ + { + "type": "text", + "text": "Personalized Federated Domain-Incremental Learning based on Adaptive Knowledge Matching", + "text_level": 1, + "bbox": [ + 225, + 140, + 779, + 186 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Yichen Li $^{1}$ , Wenchao Xu $^{2}$ , Haozhao Wang $^{1}$ *, Yining Qi $^{1}$ , Jingcai Guo $^{2}$ , and Ruixuan Li $^{1}$", + "bbox": [ + 333, + 212, + 668, + 244 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "1 School of Computer Science and Technology, Huazhong University of Science and Technology, Wuhan, China", + "bbox": [ + 243, + 253, + 756, + 282 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "{ycli0204,hz_wang,qiyining,rxli}@hust.edu.cn", + "bbox": [ + 320, + 284, + 683, + 296 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "$^{2}$ Department of Computing, The Hong Kong Polytechnic University, Hongkong, China {wenchao.xu, jc-jingcai.guo} @polyu.edu.hk", + "bbox": [ + 238, + 296, + 761, + 325 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Abstract. This paper focuses on Federated Domain-Incremental Learning (FDIL) where each client continues to learn incremental tasks where their domain shifts from each other. We propose a novel adaptive knowledge matching-based personalized FDIL approach (pFedDIL) which allows each client to alternatively utilize appropriate incremental task learning strategy on the correlation with the knowledge from previous tasks. More specifically, when a new task arrives, each client first calculates its local correlations with previous tasks. Then, the client can choose to adopt a new initial model or a previous model with similar knowledge to train the new task and simultaneously migrate knowledge from previous tasks based on these correlations. Furthermore, to identify the correlations between the new task and previous tasks for each client, we separately employ an auxiliary classifier to each target classification model and propose sharing partial parameters between the target classification model and the auxiliary classifier to condense model parameters. We conduct extensive experiments on several datasets of which results demonstrate that pFedDIL outperforms state-of-the-art methods by up to $14.35\\%$ in terms of average accuracy of all tasks.", + "bbox": [ + 259, + 364, + 743, + 589 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "1 Introduction", + "text_level": 1, + "bbox": [ + 215, + 614, + 356, + 630 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Federated Learning (FL) has become a basic paradigm that facilitates collaborative training of deep neural networks (DNNs) from distributed clients while preserving their data privacy [24, 42]. To deploy FL in real-world applications, it is practical that each client can continuously learn new incremental tasks while retaining the accuracy for the previous tasks [40]. However, history data may often be out of reach with privacy constraints like GDPR and many studies have shown that DNNs will forget the knowledge of previous tasks when learning new tasks [35, 36]. Such a phenomenon is known as 'Catastrophic Forgetting', which is a major challenge of incremental learning in the centralized setting. Especially, this challenge is even more formidable to solve in the FL scenario, where the limited previous data are stored in distributed clients and cannot be accessed by each other [1, 43].", + "bbox": [ + 212, + 648, + 787, + 815 + ], + "page_idx": 0 + }, + { + "type": "page_footnote", + "text": "* Haozhao Wang, Yining Qi and Ruixuan Li are corresponding authors.", + "bbox": [ + 218, + 825, + 645, + 839 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Many recent studies have been proposed to solve the catastrophic forgetting challenge in federated incremental learning. For instance, [40] aims to learn a personalized model for each client by decomposing the model parameters into shared parameters and adaptive parameters and transferring common knowledge between similar tasks across clients. FCIL [6] addresses the federated class-incremental learning and trains a global model by computing additional class-imbalance losses. Authors in [22] utilize extra data at both the server and client sides to conduct knowledge distillation to alleviate forgetting. [28] optimize the generative network and reconstruct previous samples for replay. Recently, Re-Fed is proposed in [20] to prevent catastrophic forgetting with synergistic replay by caching important samples which contribute to both the local and global understanding.", + "bbox": [ + 212, + 146, + 787, + 313 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "While these approaches have achieved great success, they are generally designed for the scenario where the new tasks are incremental classes, i.e., Federated Class-Incremental Learning (FCIL). These methods may lose their effectiveness when the new tasks have the same classes but their domain shift with different feature distributions, i.e., Federated Domain-Incremental Learning (FDIL). To elaborate further, the FCIL [6] method relies on changes in the new data category label in incremental tasks to alleviate forgetting of previous tasks; FOT [1] projects different incremental tasks into different orthogonal subspaces, which means feature isolation for different tasks, but if suffers from the intersected feature space boundaries of incremental tasks. The primary challenge in implementing FCIL methods in the context of FDIL is that FCIL methods typically depend on the connection between previous and new classes, which is not applicable in FDIL scenarios where the classes remain consistent.", + "bbox": [ + 212, + 316, + 789, + 498 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "In this paper, we seek to tackle the challenge of catastrophic forgetting in the FDIL scenario. A simple way to address this problem is to train different models for each incremental task. However, this requires extensive storage and time-consuming re-training. Additionally, all these models need to be saved for future use which is laborious because we need to manually select the appropriate mode for each inference data. Instead of training a separate model for each different task, a better solution is to ideally have only a single model that can work in all incremental tasks. However, naively training a model for all tasks can lead to terrible catastrophic forgetting when the unique knowledge in the new task overlaps the ones in the previous tasks.", + "bbox": [ + 212, + 503, + 789, + 638 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Unlike in the FCIL scenario, where incremental tasks involve data from different categories, In the FDIL scenario, the incremental tasks involve data from different domains, i.e., handwritten digit data and cartoon digit data. Compared to the data from different categories in FCIL, i.e., cat data and dog data, there is more shared feature information between the different incremental tasks, which can facilitate the overall FDIL task [21]. The key to efficiently learning new tasks while memorizing previous tasks in the FDIL scenario is to identify the correlation between incremental tasks and then share the useful common knowledge in different tasks. Like [40] in FCIL, we consider each client maintains their personalized models while the number of personalized models is under a given threshold due to the limited capacity of each client. Given the correlation between tasks, each client can maintain several personalized models and share the same model between similar tasks. In this way, the client can choose a model with tasks most similar to the new task, enabling the acquisition of new task knowledge while retain-", + "bbox": [ + 212, + 643, + 787, + 840 + ], + "page_idx": 1 + }, + { + "type": "page_number", + "text": "2", + "bbox": [ + 217, + 114, + 228, + 126 + ], + "page_idx": 1 + }, + { + "type": "header", + "text": "Yichen Li, et al.", + "bbox": [ + 271, + 114, + 367, + 127 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "ing existing task knowledge. This dual benefit permits the new task to also capitalize on previous knowledge, ultimately enhancing performance. Nonetheless, considering the tasks come incrementally, identifying the inter-task correlations and subsequently training the selected model is challenging.", + "bbox": [ + 212, + 146, + 787, + 205 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "To tackle these limitations, we in this paper propose a novel Personalized Federated Domain-Incremental Learning (pFedDIL) approach which adaptively matches and migrates knowledge among different incremental tasks. More specifically, we propose that each client employs a small auxiliary classifier for each personalized model which is trained to distinguish itself from other tasks. Using the auxiliary classifiers of previous tasks, each client can adaptively identify the correlation of the new task to all previous tasks. Based on these correlations, each client can match the new task with a suitable model, i.e., a new initialized model or a previous model but with similar knowledge. To maximize the training efficiency of the matched model over the new task, we further allow each client to migrate knowledge from other not matched but similar personalized models in a weighted manner, where the weights are the correlations obtained previously. Furthermore, to reduce the size of auxiliary classifiers while enhancing their performance, we propose sharing partial parameters between the target personalized model and the auxiliary classifier. Finally, considering the presence of multiple personalized local models within each client that could lead to additional inference costs, we apply a weighted ensemble distillation over these models where the weights are also obtained from the output of the auxiliary classifiers.", + "bbox": [ + 212, + 207, + 787, + 462 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Through extensive experiments on several datasets and different settings, we show that the proposed pFedDIL significantly improves the model accuracy as compared to state-of-the-art approaches. The major contributions of this paper are summarized as follows:", + "bbox": [ + 212, + 464, + 787, + 523 + ], + "page_idx": 2 + }, + { + "type": "list", + "sub_type": "text", + "list_items": [ + "- We are the first to formulate the problem of federated domain-incremental learning. To address this issue, we propose a novel method named pFedDIL which enables each client to match and migrate knowledge from previous tasks to alleviate catastrophic forgetting.", + "- To adaptively employ incremental task learning strategy and migrate knowledge, we define the knowledge matching intensity by the auxiliary classifier to discern the correlation between the new task and previous tasks.", + "- To facilitate the adaptive selection of inference model among a set of personalized local models, we coordinate all personalized local models to predict weighted by normalized output by auxiliary classifiers. Besides, we further propose sharing partial parameters between the auxiliary classifier and the target classification model to condense model parameters.", + "- We carry out extensive experiments on various datasets and different settings. Experimental results illustrate that our proposed model outperforms the state-of-the-art methods by up to $14.35\\%$ in terms of average accuracy on all tasks." + ], + "bbox": [ + 225, + 532, + 782, + 753 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "2 Related Work", + "text_level": 1, + "bbox": [ + 215, + 779, + 367, + 794 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Federated Learning. Federated learning (FL) is a distributed learning approach that trains a global model by merging multiple locally trained models on private datasets", + "bbox": [ + 212, + 809, + 787, + 840 + ], + "page_idx": 2 + }, + { + "type": "header", + "text": "Federated Domain-Incremental Learning", + "bbox": [ + 486, + 114, + 730, + 128 + ], + "page_idx": 2 + }, + { + "type": "page_number", + "text": "3", + "bbox": [ + 774, + 116, + 785, + 126 + ], + "page_idx": 2 + }, + { + "type": "image", + "img_path": "images/7fa8509dae5f4dafbc2482281b406e213d7f24cde0991e0473be9f773dae0213.jpg", + "image_caption": [ + "Fig. 1: Illustration of the pFedDIL framework. In the training stage, each client randomly receives one domain dataset without replacement to form a task, where different colors represent different domains. Clients can adaptively employ an appropriate incremental task learning strategy and migrate the knowledge from previous tasks with the knowledge matching intensity." + ], + "image_footnote": [], + "bbox": [ + 289, + 143, + 712, + 422 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "[18]. A significant challenge in FL is data heterogeneity, where data distribution among clients is non-identically and independently distributed (Non-IID), negatively impacting model performance [14, 33, 38]. FedAvg is a notable FL method that refines the global model by incorporating parameters trained on local private data [24]. [19] propose a proximal term to align closely with the global model. The federated distillation method [37] suggests distilling multiple local models into a global one by only aggregating soft output predictions. However, existing FL approaches typically assume that the data in each client is static, which cannot account for incremental data with domain shift.", + "bbox": [ + 212, + 521, + 787, + 642 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Incremental Learning. Incremental learning (IL) is a learning paradigm that empowers a model to learn sequentially from a series of tasks while preserving the knowledge acquired from previous tasks [39, 44]. IL can be segmented into three primary variants: task-incremental learning [23], class-incremental learning [29] and domain-incremental learning [25]. Existing strategies can further be grouped into three categories: replay-based methods [30], regularization-based methods [16] and parameter isolation methods [2]. Most IL methods are designed for task-IL or class-IL and are typically server-based. This work focuses on the federated domain-incremental learning scenario and proposes the pFedDIL technique.", + "bbox": [ + 212, + 643, + 787, + 779 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Federated Incremental Learning. Federated Incremental Learning (FIL), an approach that targets the dynamic adaptation of streaming tasks within each client's environment, emphasizes the seamless integration of new data into the global model while preserving the acquired knowledge from previous data streams. Despite its crucial role, FIL", + "bbox": [ + 212, + 779, + 787, + 840 + ], + "page_idx": 3 + }, + { + "type": "page_number", + "text": "4", + "bbox": [ + 217, + 114, + 228, + 126 + ], + "page_idx": 3 + }, + { + "type": "header", + "text": "Yichen Li, et al.", + "bbox": [ + 271, + 114, + 369, + 127 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "has garnered attention relatively recently, with [40] pioneering the field by focusing on Task-Incremental Learning (Task-IL) in FL. In parallel, studies such as [41] takes the multi-granularity expression of knowledge into account, promoting the spatial-temporal integration of knowledge in FCIL. Conversely, [4,5] simplify the problem's complexity by presupposing clients possess sufficient storage capacity to retain historical examples and engage in data sharing, a scenario that deviates from the conventional federated learning paradigm. Beyond image classification, investigations like [7, 15] extend the horizons of FIL into diverse domains. Our research focuses on the domain shift between different tasks and the transfer of the common knowledge.", + "bbox": [ + 214, + 146, + 787, + 282 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "3 Methodology", + "text_level": 1, + "bbox": [ + 215, + 305, + 359, + 321 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "We first formulate our problem definition and then present our pFedDIL approach. Figure 1 gives an overview of our method.", + "bbox": [ + 214, + 337, + 784, + 367 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "3.1 Problem Definition", + "text_level": 1, + "bbox": [ + 215, + 388, + 388, + 402 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "In incremental learning, each model learns from a sequence of streaming tasks $\\{\\mathcal{T}_{(1)},\\mathcal{T}_{(2)},\\ldots \\mathcal{T}_{(n)}\\}$ where $\\mathcal{T}_{(t)}$ denotes the dataset of the $t$ -th task; $\\mathcal{T}_{(t)} = \\{x_i^t,y_i^t\\}_{i = 1}^{N^t}$ , which has $N^t$ pairs of sample data $x_{i}^{t}\\in X^{t}$ and corresponding label $y_{i}^{t}\\in Y^{t}$ . $X^{t}$ and $Y^{t}$ represent the domain space and label space for the $t$ -th task. In this paper, we focus on the variant of domain incremental learning, where all tasks share the same numbers of classes i.e., $\\mathcal{V}^1 = \\mathcal{V}^t,\\forall t\\in [1,n]$ . As the sequence of tasks arrives, the client needs to learn the new task while their domain and data distribution changes, i.e., $\\mathcal{X}^1\\neq \\mathcal{X}^t,\\forall t\\in [1,n]$ .", + "bbox": [ + 214, + 414, + 787, + 523 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "We consider the federated domain-incremental learning scenario. We aim to train a global model for $K$ total clients in the federated domain-incremental learning setting. We consider the local client $k$ can only access the local private streaming tasks $\\{\\mathcal{T}_{(1)}^k,\\mathcal{T}_{(2)}^k,\\dots ,\\mathcal{T}_{(n)}^k\\}$ . Now our goal is to train a set of personalized local models $\\widetilde{w}^{k} = [w_{1}^{k},w_{2}^{k},\\ldots ,w_{d}^{k}]$ for each client $k$ over all existing $t$ tasks $\\mathcal{T} = \\sum_{n = 1}^{t}\\sum_{k = 1}^{K}\\mathcal{T}_{(n)}^{k}$ . Here each personalized local model corresponds to at least one task, and possibly multiple tasks (i.e. client $k$ may train the task $\\mathcal{T}_{(1)}^k$ and $\\mathcal{T}_{(2)}^k$ with the model $w_{1}^{k}$ ). Then, we can formulate the goal as:", + "bbox": [ + 214, + 523, + 787, + 650 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\n\\underset {\\widetilde {w}} {\\arg \\min } \\mathcal {L} (\\widetilde {w}) = \\sum_ {n = 1} ^ {t} \\sum_ {k = 1} ^ {K} \\frac {1}{| \\mathcal {T} |} \\mathcal {L} _ {(n)} ^ {k} (\\widetilde {w}). \\tag {1}\n$$\n", + "text_format": "latex", + "bbox": [ + 370, + 660, + 785, + 700 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "where $\\mathcal{L}_{(n)}^k (\\widetilde{w}) = \\mathbb{E}_{(x,y)\\sim \\mathcal{T}_{(n)}^k}\\left[l(\\widetilde{w};(x,y))\\right]$ is the empirical loss of client $k$ on $n$ -th task with the personalized local model.", + "bbox": [ + 214, + 713, + 784, + 746 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "3.2 The Proposed pFedDIL Approach", + "text_level": 1, + "bbox": [ + 215, + 768, + 493, + 784 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "The key idea of the pFedDIL is to identify the correlation between the new and previous tasks such that clients can adaptively migrate knowledge from previous tasks to improve performance. Specifically, in each communication round, the client in pFedDIL trains", + "bbox": [ + 214, + 794, + 785, + 840 + ], + "page_idx": 4 + }, + { + "type": "header", + "text": "Federated Domain-Incremental Learning", + "bbox": [ + 486, + 114, + 730, + 128 + ], + "page_idx": 4 + }, + { + "type": "page_number", + "text": "5", + "bbox": [ + 774, + 116, + 785, + 126 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "the local model with its private sequence of tasks and the server aggregates local models from all participating clients. Then, when it comes to new tasks, each client first employs the existing personalized models' auxiliary classifiers to calculate the correlations denoted as knowledge matching intensity and applies the appropriate incremental task learning strategy: training the new task with a previous personalized local model which has the most similar knowledge or training with a new initial model. Simultaneously, each client can migrate knowledge from other personalized local models with the knowledge matching intensity to assist in training. During the inference stage, the final inference result is aggregated by all personalized local models weighted by the corresponding auxiliary classifiers. Moreover, we propose sharing partial parameters between the auxiliary classifier and the target personalized model to condense model parameters. The workflow of the pFedDIL is shown in Algorithm 1.", + "bbox": [ + 212, + 146, + 787, + 327 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Knowledge Matching Intensity. To assess the similarity between different incremental tasks, where a higher similarity indicates a greater amount of common knowledge, we propose the knowledge matching intensity. When it comes to the new incremental task: $(t + 1)$ -th task, each client first needs to compute the knowledge matching intensity between the $(t + 1)$ -th task and previous $t$ tasks. To achieve the adaptive knowledge matching intensity estimation for each client, we define the auxiliary classifier, a binary classification deep neural network $\\theta$ for the corresponding local task $T$ , which is trained by the discrimination of the sample belonging to the task $T$ . As a new task emerges, the auxiliary classifier can output the similarity between different tasks based on the extracted feature representation. Supposed that the client $k$ has converged on $t$ tasks within $d$ local existing task models when it comes to the incremental $(t + 1)$ -th task, the client needs to calculate the knowledge matching intensity $\\widetilde{\\rho}_k = [\\rho_1^k,\\rho_2^k,\\dots ,\\rho_d^k ]$ with each auxiliary classifier $\\widetilde{\\theta}^{k} = [\\theta_{1}^{k},\\theta_{2}^{k},\\dots ,\\theta_{d}^{k}]$ :", + "bbox": [ + 212, + 328, + 787, + 527 + ], + "page_idx": 5 + }, + { + "type": "equation", + "text": "\n$$\n\\widetilde {\\rho} ^ {k} = \\frac {1}{N ^ {t + 1}} \\sum_ {i = 1} ^ {N ^ {t + 1}} f \\left(x _ {i} ^ {t + 1}; \\widetilde {\\theta} ^ {k}\\right). \\tag {2}\n$$\n", + "text_format": "latex", + "bbox": [ + 398, + 539, + 785, + 580 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "where the $f(\\cdot)$ denotes the output of the auxiliary classifier, $\\widehat{\\theta}^k$ denotes a set of personalized local auxiliary classifiers in client $k$ . Each auxiliary classifier is trained by minimizing the total empirical loss over the corresponding task's local dataset $\\mathcal{T}$ , which can be formulated as:", + "bbox": [ + 212, + 595, + 787, + 656 + ], + "page_idx": 5 + }, + { + "type": "equation", + "text": "\n$$\n\\mathcal {L} ^ {k} (\\theta) = \\frac {1}{T} \\sum_ {i = 1} ^ {\\tau} \\mathcal {L} _ {C E} \\left(\\theta ; x _ {i}, y _ {i}\\right). \\tag {3}\n$$\n", + "text_format": "latex", + "bbox": [ + 393, + 665, + 785, + 705 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "where $\\mathcal{L}^k (\\theta)$ is the local loss in the client $k$ and $\\mathcal{L}_{CE}$ is the cross-entropy loss function that measures the difference between the prediction and the ground truth labels.", + "bbox": [ + 212, + 718, + 785, + 748 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Incremental Task Learning Strategy. After acquiring the knowledge matching intensity $\\widetilde{\\rho}_k$ with equation (2), the client can select the appropriate incremental task learning strategy. An intuition is that the client will prefer to train with previous personalized local models when the knowledge matching intensity exceeds the $\\lambda$ , which means the high-level similarity between the new task and the previous tasks so that the model can converge on the new task with slight adjustments. Besides, the client will sacrifice", + "bbox": [ + 212, + 750, + 787, + 840 + ], + "page_idx": 5 + }, + { + "type": "page_number", + "text": "6", + "bbox": [ + 217, + 114, + 228, + 126 + ], + "page_idx": 5 + }, + { + "type": "header", + "text": "Yichen Li, et al.", + "bbox": [ + 271, + 114, + 367, + 127 + ], + "page_idx": 5 + }, + { + "type": "code", + "sub_type": "algorithm", + "code_caption": [ + "Algorithm 1: pFedDIL" + ], + "code_body": "Input : $T$ : the communication round; $K$ : client number; $C$ : the fraction of active client in each round; $\\{\\mathcal{T}_{(t)}\\}_{t = 1}^{n}$ : the distributed dataset with $n$ tasks; w: the parameter of the target classification model; $\\theta$ : the parameter of the auxiliary classifier. 1 Initialize the parameter w and $\\theta$ . \n2 for $t = 1$ to T do \n3 $S_{t}\\gets$ (random set of $[C\\cdot K]$ clients); for each selected client $k\\in S_t$ in parallel do \n4 receives $w_{t}$ from the server; \n5 calculate knowledge matching intensity $\\widetilde{\\rho}^k$ with (2); \n6 select learning strategy $w_{t + 1}^{k}$ with (4); \n7 set local models $w_{t + 1}^{k}$ and $\\theta_{t + 1}^{k}$ . \n8 for e = 1 to E do \n9 update $\\theta_{t + 1}^{k}$ with (3); \n10 update $w_{t + 1}^{k}$ through adaptive knowledge migration with (5); \n11 end \n12 pushes $w_{t + 1}^{k}$ to the server. \n13 end \n14 $w_{t}\\gets$ ServerAggregation( $\\{w^{k}\\}_{k\\in S_{t}}$ ) \n15 end", + "bbox": [ + 225, + 166, + 764, + 460 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "some storage by training the new task with the new initial model parameters to ensure accuracy for both the new task and previous tasks. Here we define the incremental task learning strategy as follows:", + "bbox": [ + 212, + 493, + 787, + 539 + ], + "page_idx": 6 + }, + { + "type": "equation", + "text": "\n$$\nw _ {t + 1} ^ {k} = O P T \\left\\{ \\begin{array}{l l} \\hat {w} & i f \\max \\left(\\widetilde {\\rho} _ {k}\\right) < \\lambda \\\\ \\widetilde {w} ^ {k} [ m ] & e l s e \\max \\left(\\widetilde {\\rho} _ {k}\\right) = \\rho_ {k} ^ {m} \\geq \\lambda . \\end{array} \\right. \\tag {4}\n$$\n", + "text_format": "latex", + "bbox": [ + 321, + 545, + 785, + 585 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "where $\\lambda$ is a hyper-parameter, $\\widetilde{w}^k = [w_1^k, w_2^k, \\ldots, w_d^k]$ , $\\hat{w}$ denotes new initial model parameters, and $OPT$ means the alternative of existed personalized local models and a new initial model.", + "bbox": [ + 212, + 593, + 787, + 638 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Knowledge Migration. Then, during the training of $(t + 1)$ -th task, each client respectively migrates the knowledge from other existing personalized local models to assist the training with the calculated knowledge matching intensity $\\widetilde{\\rho}_k$ , minimize the following objective:", + "bbox": [ + 212, + 638, + 787, + 700 + ], + "page_idx": 6 + }, + { + "type": "equation", + "text": "\n$$\n\\min _ {w _ {t + 1} ^ {k}} \\mathcal {L} _ {L o c a l} ^ {k} (w _ {t + 1} ^ {k}) = \\mathcal {L} _ {(t + 1)} ^ {k} (w _ {t + 1} ^ {k}) + \\mathcal {L} _ {K M} ^ {k} (w _ {t + 1} ^ {k}),\n$$\n", + "text_format": "latex", + "bbox": [ + 318, + 705, + 684, + 734 + ], + "page_idx": 6 + }, + { + "type": "equation", + "text": "\n$$\n\\begin{array}{l} \\text {w h e r e} \\quad \\mathcal {L} _ {K M} ^ {k} \\left(w _ {t + 1} ^ {k}\\right) = \\widetilde {\\rho} _ {k} \\cdot \\left| \\left| w _ {t + 1} ^ {k} - \\widetilde {w} ^ {k} \\right| \\right| ^ {2}, \\\\ = \\sum_ {i = 1} ^ {d} \\rho^ {i} \\cdot \\left| \\left| w _ {t + 1} ^ {k} - w _ {i} ^ {k} \\right| \\right| ^ {2}. \\tag {5} \\\\ \\end{array}\n$$\n", + "text_format": "latex", + "bbox": [ + 318, + 737, + 785, + 801 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "where $\\mathcal{L}_{t + 1}^k (w_{t + 1}^k)$ is defined in (1) and $\\mathcal{L}_{KM}^{k}(w_{t + 1}^{k})$ refers to the Knowledge Migration (KM) loss.", + "bbox": [ + 212, + 808, + 785, + 839 + ], + "page_idx": 6 + }, + { + "type": "header", + "text": "Federated Domain-Incremental Learning", + "bbox": [ + 486, + 114, + 732, + 128 + ], + "page_idx": 6 + }, + { + "type": "page_number", + "text": "7", + "bbox": [ + 774, + 114, + 784, + 126 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Inference Stage. During the inference stage, taking into account that each client may have multiple personalized models, how to automatically select the appropriate model for inference is an urgent problem to solve. We have observed that ensemble learning, which enhances model performance by integrating knowledge from multiple smaller models, has achieved widespread success [8]. Utilizing the calculated knowledge matching intensity as the weight for the personalized model inference results, we achieve automated ensemble inference for the inference data. For every inference sample $x \\in \\hat{\\mathcal{T}}$ , the client $k$ computes the weights of each personalized local models $\\widetilde{w}^k$ with the auxiliary $\\widetilde{\\theta}^k$ as $f(\\widetilde{\\theta}^k; x)$ and then normalizes it into the probability:", + "bbox": [ + 212, + 146, + 787, + 285 + ], + "page_idx": 7 + }, + { + "type": "equation", + "text": "\n$$\n\\hat {\\alpha} _ {i} ^ {k} = \\frac {f \\left(\\theta_ {i} ^ {k} ; x\\right)}{\\sum_ {i = 1} ^ {d} f \\left(\\theta_ {i} ^ {k} ; x\\right)}. \\tag {6}\n$$\n", + "text_format": "latex", + "bbox": [ + 424, + 292, + 785, + 330 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "which guarantees that $\\sum_{i=1}^{d} \\hat{\\alpha}_i^k = 1$ . Finally, the client inputs the inference sample $x$ into each personalized local models $\\widetilde{w}^k$ to obtain the soft predictions $s(\\widetilde{w}^k; x)$ and final weighed prediction:", + "bbox": [ + 214, + 340, + 785, + 387 + ], + "page_idx": 7 + }, + { + "type": "equation", + "text": "\n$$\nI n f e r e n c e \\left(\\widetilde {w} ^ {k}; x\\right) = \\sum_ {i = 1} ^ {d} \\hat {\\alpha} _ {i} ^ {k} \\cdot s \\left(w _ {i} ^ {k}; x\\right). \\tag {7}\n$$\n", + "text_format": "latex", + "bbox": [ + 361, + 397, + 785, + 438 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Partial Parameters Sharing. Motivated by the idea of multi-task learning where sharing the encoder between different tasks can promote each other [34], we propose sharing partial parameters between the auxiliary classifier $\\theta^k$ and the target classification model $w^k$ to solve this problem. The intuition behind this is that both the auxiliary classifier and the classification model have to distinguish the sample from the extracted features. Besides, another benefit of sharing layers is that the model parameters can be reduced.", + "bbox": [ + 214, + 446, + 787, + 537 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "To this end, we propose sharing the front model layers between the two models which are used to extract features, as illustrated in Figure 1. Specifically, by denoting the $c$ -layer shared extractor by $\\tilde{w}^k = [\\tilde{w}^{k,1},\\dots,\\tilde{w}^{k,c}]$ where $\\tilde{w}^{k,i}$ is the $i$ -th layer of the shared extractor, we can re-write the $n_d$ -layer auxiliary classifier by $\\theta^k = [\\tilde{w}^k,\\theta^{k,c + 1},\\dots,\\theta^{k,n_d}]$ , and the $n$ -layer target classification model $w^k$ by $w^k = [\\tilde{w}^k,w^{k,c + 1},\\dots,w^{k,n}]$ with $w^{k,i}$ as the $i$ -th layer of the classification model $w^{k,i}$ . With the shared layers, the discriminator and the classification model are trained jointly:", + "bbox": [ + 214, + 537, + 787, + 643 + ], + "page_idx": 7 + }, + { + "type": "equation", + "text": "\n$$\n\\min _ {\\theta^ {k}, w ^ {k}} \\mathcal {L} _ {\\mathcal {J}} ^ {k} \\left(\\theta^ {k}, w ^ {k}\\right) = \\mathcal {L} ^ {k} \\left(\\theta^ {k}\\right) + \\mathcal {L} ^ {k} \\left(w ^ {k}\\right). \\tag {8}\n$$\n", + "text_format": "latex", + "bbox": [ + 367, + 651, + 785, + 676 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "where $\\mathcal{L}^k (\\theta^k)$ refers to (3) and $\\mathcal{L}^k (w^k)$ is defined in (1).", + "bbox": [ + 214, + 686, + 588, + 704 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "4 Experiments", + "text_level": 1, + "bbox": [ + 214, + 724, + 356, + 742 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "4.1 Setup", + "text_level": 1, + "bbox": [ + 214, + 755, + 297, + 771 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Dataset: We conduct all our experiments with heterogeneous dataset partition on three datasets. In FDIL scenarios, new domains are introduced gradually. The dataset initially contains samples from a specific domain, and new domains are introduced at later stages, enabling models to adapt and generalize to new unseen domains.", + "bbox": [ + 212, + 779, + 785, + 840 + ], + "page_idx": 7 + }, + { + "type": "page_number", + "text": "8", + "bbox": [ + 217, + 114, + 228, + 126 + ], + "page_idx": 7 + }, + { + "type": "header", + "text": "Yichen Li, et al.", + "bbox": [ + 271, + 114, + 367, + 127 + ], + "page_idx": 7 + }, + { + "type": "table", + "img_path": "images/ebfd1776b333a67be4d19cb7db17eef38008215d2cdea21544a6b22da27f3072.jpg", + "table_caption": [ + "Table 1: Performance comparison by Top-1 test accuracy for our model and baseline methods on Digit-10 (data heterogeneity: $\\alpha = 0.1$ )." + ], + "table_footnote": [], + "table_body": "
Digit-10 (λ = 0.5)
MethodMNIST → EMNIST → USPS → SVHNAvgΔ(↑)
FedAvg [24]92.82 →84.02 →81.02 →71.8782.436.85↑
FedProx [19]93.01 →84.68 →78.30 →76.4283.106.18↑
FedCIL [28]94.64 →87.52 →82.16 →80.9286.312.97↑
DANN [10] + FL96.07 →86.71 →79.11 →72.1483.515.77↑
Source-Only92.82 →82.15 →75.53 →69.0679.899.39↑
Sharing92.67 →82.91 →76.17 →71.9680.938.35↑
pFedDIL -w/o Migration92.82 →87.09 →84.00 →81.8986.452.83↑
pFedDIL -w/o Sharing92.79 →90.13 →87.59 →86.1989.180.11↑
pFedDIL -w/o Correlation92.82 →89.06 →84.72 →84.3087.731.55↑
pFedDIL92.82 →90.39 →87.14 →86.7589.28/
Upper Bound -Disjoint92.82 →90.33 →93.41 →87.9791.131.85↓
", + "bbox": [ + 259, + 184, + 743, + 409 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "text", + "list_items": [ + "- Digit10: Digit-10 dataset contains 10 digit categories in four domains: MNIST [17], EMNIST [3], USPS [13], SVHN [26]. Each dataset is a digit image classification dataset of 10 classes in a specific domain, such as handwriting style.", + "- Office31: [32] A dataset with images from three different domains: Amazon, Webcam, and DSLR. It consists of 31 object categories, with each domain having around 4,100 images.", + "- DomainNet: [27] A large-scale dataset with images from six different domains: Clipart, Painting, Real, Sketch, Quickdraw, and Infograph. It contains over 0.6 million images across 345 categories." + ], + "bbox": [ + 225, + 436, + 785, + 573 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "Baseline: For a fair comparison with other key related works, we follow the same protocols proposed by [24] to set federated learning(FL) task and utilize the domain-incremental learning (DIL) task by [25]. We evaluate all baselines under the above settings, including two representative FL models FedAvg [24] and FedProx [19], which is better at tackling heterogeneity in federated networks than FedAvg, and the following several methods in FDIL scenario:", + "bbox": [ + 212, + 583, + 787, + 672 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "FedCIL [28]: This approach employs the ACGAN backbone to generate synthetic samples to consolidate the global model and align sample features in the output layer. Authors conduct experiments in FCIL scenarios, and here we adopt it to the FDIL setting.", + "bbox": [ + 212, + 672, + 787, + 719 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "Domain Adaptation + FL: Here we adopt the robust adversarial-based method DANN [10]. This baseline mainly follows the domain adaptation paradigm which is different from the incremental learning and is often prone to issues like catastrophic forgetting.", + "bbox": [ + 212, + 719, + 787, + 763 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "Source-Only: Here we maintain the initial model trained on the initial domain data completely fixed throughout all tasks. This baseline does not perform any adaptation, modification or replacement to its parameters and serves as a lower bound.", + "bbox": [ + 212, + 763, + 787, + 808 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "Sharing: Inspired by the multi-task learning scenario [31], we adopt all front layers before the last fully connected layer as shared layers, and we use relevant different", + "bbox": [ + 212, + 809, + 787, + 839 + ], + "page_idx": 8 + }, + { + "type": "header", + "text": "Federated Domain-Incremental Learning", + "bbox": [ + 486, + 114, + 730, + 128 + ], + "page_idx": 8 + }, + { + "type": "page_number", + "text": "9", + "bbox": [ + 774, + 116, + 785, + 126 + ], + "page_idx": 8 + }, + { + "type": "table", + "img_path": "images/3474622a804649e4b514bd5ea5edf4e1235726b04119535b7573202c9095d82c.jpg", + "table_caption": [ + "Table 2: Performance comparison by Top-1 test accuracy for our model and baseline methods on Office-31 (data heterogeneity: $\\alpha = 1$ )." + ], + "table_footnote": [], + "table_body": "
Office-31 (λ = 0.8)
MethodAmazon → Dlsr → WebcamAvgΔ(↑)
FedAvg [24]46.53 → 24.17 → 29.3433.3512.82↑
FedProx [19]45.19 → 25.68 → 32.2334.3711.80↑
FedCIL [28]49.38 → 39.65 → 44.0444.361.81↑
DANN [10] + FL51.97 → 35.96 → 43.0843.672.50↑
Source-Only46.53 → 20.61 → 28.3231.8214.35↑
Sharing46.11 → 24.23 → 34.2534.8611.31↑
pFedDIL -w/o Migration46.53 → 40.17 → 44.5543.752.42↑
pFedDIL -w/o Sharing46.63 → 42.90 → 47.0345.520.65↑
pFedDIL -w/o Correlation46.53 → 41.66 → 47.1945.131.04↑
pFedDIL46.53 → 43.87 → 48.1246.17/
Upper Bound -Disjoint46.53 → 50.17 → 56.2150.974.80↓
", + "bbox": [ + 289, + 184, + 712, + 409 + ], + "page_idx": 9 + }, + { + "type": "text", + "text": "fully-connected layers to get outputs of different domains. During inference, the fully-connected layer that matches the corresponding domain ID will be applied.", + "bbox": [ + 212, + 438, + 782, + 468 + ], + "page_idx": 9 + }, + { + "type": "text", + "text": "Disjoint: Here we train a separate model for each domain. During inference, the model which matches the corresponding domain ID is used. This baseline requires additional space to store parameters and additional cost to obtain inference-task ID. It is a degradation of domain-incremental learning and provides an upper bound here.", + "bbox": [ + 212, + 469, + 784, + 530 + ], + "page_idx": 9 + }, + { + "type": "text", + "text": "Configuration: Unless otherwise mentioned, we set the number of local training epoch $E = 20$ , communication round $T = 180$ for each domain. Here we train each task until convergence before the arrival of a new task with communication rounds $T$ . For local training, the batch size is 32, the learning rate for our models is 0.001 and the weight decay is $1e - 3$ . Like most FL methods on Non-IID data distribution [24], we use the Dirichlet distribution $(\\alpha)$ to simulate the data heterogeneity from all domains and we apply the same data heterogeneity for all domains in the same dataset. For the classifier in all methods, we employ ResNet18 [12] as the basic backbone. For the multi-task learning structure in our approach, we treat all previous layers before the last fully connected layer as share layers, and we use two different fully connected layers to get outputs as the auxiliary classifier result and target classification result. The total clients' number is 20 for Digit-10 and DomainNet with an active ratio $k = 0.4$ , 10 for Office-31 with an active ratio $k = 0.4$ .", + "bbox": [ + 212, + 530, + 787, + 726 + ], + "page_idx": 9 + }, + { + "type": "text", + "text": "4.2 Performance Overview", + "text_level": 1, + "bbox": [ + 215, + 751, + 418, + 766 + ], + "page_idx": 9 + }, + { + "type": "text", + "text": "Test Accuracy. Table 1,2,3 shows the performance of our pFedDIL and other baseline methods on three datasets, where \" $\\Delta$ \" denotes the improvement of our method with other baselines. We carry out experiments against data heterogeneity on each dataset, and pFedDIL achieves the best performance in most cases by a margin of", + "bbox": [ + 212, + 779, + 787, + 840 + ], + "page_idx": 9 + }, + { + "type": "page_number", + "text": "10", + "bbox": [ + 217, + 114, + 235, + 126 + ], + "page_idx": 9 + }, + { + "type": "header", + "text": "Yichen Li, et al.", + "bbox": [ + 271, + 114, + 367, + 127 + ], + "page_idx": 9 + }, + { + "type": "table", + "img_path": "images/fd7ff2eb691aade0653caf3a4038fe3eae456dc9aa0a1c2e8dbebddee899cbaa.jpg", + "table_caption": [ + "Table 3: Performance comparison by Top-10 test accuracy for our model and baseline methods on DomainNet (data heterogeneity: $\\alpha = 10$ )" + ], + "table_footnote": [], + "table_body": "
DomainNet (λ = 0.8)
MethodClipart → Infograph → Painting → Quickdraw → Real → SketchAvgΔ(↑)
FedAvg [24]48.43 → 37.18 → 45.80 → 45.32 → 44.91 → 50.2545.326.38↑
FedProx [19]47.39 → 38.43 → 44.31 → 47.96 → 42.38 → 51.7745.376.32↑
FedCIL [28]52.14 → 43.68 → 47.10 → 48.75 → 42.89 → 51.2647.644.05↑
DANN [10] + FL50.07 → 39.74 → 43.73 → 45.08 → 43.28 → 52.9645.815.88↑
Source-Only48.43 → 31.01 → 33.12 → 38.15 → 37.62 → 42.8538.5313.16↑
Sharing48.43 → 36.03 → 39.58 → 41.80 → 40.41 → 44.7641.849.86↑
pFedDIL -w/o Migration48.43 → 43.83 → 50.05 → 49.44 → 46.10 → 54.9048.792.90↑
pFedDIL -w/o Sharing48.52 → 47.37 → 51.16 → 53.98 → 50.11 → 55.2451.060.63↑
pFedDIL -w/o Correlation48.43 → 46.71 → 51.09 → 52.34 → 48.07 → 56.9550.601.09↑
pFedDIL48.43 → 47.19 → 52.48 → 54.07 → 50.36 → 57.6151.69/
Upper Bound -Disjoint48.43 → 52.33 → 54.21 → 52.89 → 53.62 → 58.0653.261.57↓
", + "bbox": [ + 218, + 185, + 787, + 383 + ], + "page_idx": 10 + }, + { + "type": "text", + "text": "$1.81\\% \\sim 14.35\\%$ in terms of average accuracy. Due to our challenging FDIL settings, there is a notable gap between all baselines and the upper performance bound. However, pFedDIL almost achieves comparable results with less storage and adaptive task-id inference during the inference stage. These results demonstrate that the appropriate incremental task learning strategy and adaptive knowledge migration technique in pFedDIL can alleviate catastrophic forgetting and thus improve performance.", + "bbox": [ + 212, + 412, + 787, + 503 + ], + "page_idx": 10 + }, + { + "type": "image", + "img_path": "images/84b22f2020e652fdb2540cbd3e3a629542b40f397e4fdb82561514878995d3cc.jpg", + "image_caption": [ + "Fig. 2: Visualized performance w.r.t data heterogeneity for three datasets." + ], + "image_footnote": [], + "bbox": [ + 215, + 529, + 787, + 681 + ], + "page_idx": 10 + }, + { + "type": "text", + "text": "Ablation Study. As shown in Tables 1,2,3, we evaluate the effects of each module in our model via ablation studies. pFedDIL -w/o Migration, pFedDIL -w/o Sharing and pFedDIL- w/o Correlation denote the performance of our model without using knowledge migration, parameters sharing and correlation factor in the inference stage. Compared with pFedDIL, the performance of pFedDIL -w/o Migration, pFedDIL -w/o Sharing and pFedDIL- w/o Correlation degrades evidently with a range of $0.11\\% \\sim 2.90\\%$ . Specifically, the adaptive knowledge migration technique in pFedDIL plays a much", + "bbox": [ + 212, + 733, + 787, + 840 + ], + "page_idx": 10 + }, + { + "type": "header", + "text": "Federated Domain-Incremental Learning", + "bbox": [ + 486, + 114, + 732, + 128 + ], + "page_idx": 10 + }, + { + "type": "page_number", + "text": "11", + "bbox": [ + 767, + 114, + 784, + 126 + ], + "page_idx": 10 + }, + { + "type": "text", + "text": "more important role with the obvious improvement in test accuracy, and the partial parameters sharing trick condenses model parameters by sharing the feature extracted layer while not causing a remarkable impact on test accuracy. Experiment results verify the effectiveness of all modules, confirming all modules are essential to train a robust federated domain-incremental model.", + "bbox": [ + 212, + 146, + 787, + 220 + ], + "page_idx": 11 + }, + { + "type": "text", + "text": "Data heterogeneity. Increasing data heterogeneity levels in experiments is significant because it allows for a more realistic evaluation of the system's performance in diverse real-world scenarios and assesses its robustness and generalization capabilities. In particular, in the FDIL scenario, incremental data is less likely to appear in the form of IID data. Therefore, it is essential to simulate different levels of Non-IID data to validate the effectiveness of our method in the FDIL scenario. Figure 2 displays the accuracy with different levels of data heterogeneity on three datasets. As shown in this figure, all methods achieve an improvement in test accuracy with the decline of data heterogeneity. Most notably, pFedDIL always maintains a leading advantage with different levels of data heterogeneity.", + "bbox": [ + 212, + 222, + 787, + 375 + ], + "page_idx": 11 + }, + { + "type": "table", + "img_path": "images/d479f493feb1f8d846281dc3d315fdd26d56b6352dee4f83ae9984d39a20c8b0.jpg", + "table_caption": [ + "Table 4: Evaluation of various methods in terms of the communication rounds. We report the sum of communication rounds required to achieve the best performance on each task and the running time of a training communication round." + ], + "table_footnote": [], + "table_body": "
DatasetFedAvgFedCILDANN+FLpFedDIL
RoundsTimeRoundsTimeRoundsTimeRoundsTime
Digit10480(r)32.14(s)492(r)50.06(s)553(r)46.77(s)379(r)34.91(s)
Office31463(r)44.96(s)482(r)78.32(s)488(r)62.89(s)411(r)47.58(s)
DomainNet835(r)92.22(s)904(r)165.41(s)861(r)158.01(s)725(r)105.38(s)
", + "bbox": [ + 232, + 454, + 772, + 554 + ], + "page_idx": 11 + }, + { + "type": "text", + "text": "Computational complexity and communication cost. We treat all previous layers in ResNet18 before the last fully connected layer as shared layers, and we use two different fully connected layers to get outputs as the auxiliary classifier result and target classification result. Compared with baseline methods, we only train one additional fully connected layer as the auxiliary classifier for the binary classification task - the discrimination of the sample belonging to the current task. The training time cost (each communication round) is shown in Table 4. Therefore, the training of the auxiliary classifier will not increase the overhead or computational complexity significantly and hardly affect the training efficiency.", + "bbox": [ + 212, + 583, + 787, + 718 + ], + "page_idx": 11 + }, + { + "type": "text", + "text": "Table 4 shows the evaluation of various methods in terms of the communication rounds to reach the test accuracy reported in Table 1, 2 and 3 (180 communication rounds each task). Here we show the sum of communication rounds required to achieve the performance on each task. pFedDIL requires the least communication rounds to achieve the reported test accuracy on all datasets with the knowledge matching. Thus, the local model is easier to converge.", + "bbox": [ + 212, + 719, + 787, + 809 + ], + "page_idx": 11 + }, + { + "type": "text", + "text": "Parameter sensitivity analysis. To figure out whether pFedDIL is sensitive to some specific parameters, we first select some general parameters, which are defined in most", + "bbox": [ + 212, + 809, + 787, + 840 + ], + "page_idx": 11 + }, + { + "type": "page_number", + "text": "12", + "bbox": [ + 217, + 114, + 235, + 126 + ], + "page_idx": 11 + }, + { + "type": "header", + "text": "Yichen Li, et al.", + "bbox": [ + 271, + 114, + 367, + 127 + ], + "page_idx": 11 + }, + { + "type": "image", + "img_path": "images/c4f1bd2a9d2a59063f1df45c42fdd4cfb47ab0298990d4bf4c67b88ded6c517d.jpg", + "image_caption": [ + "Fig. 3: Performance of pFedDIL under different configurations. Here, we select three general parameters in the FL setting: (a) local training epoch $E$ , (b) sample size $B$ in the classifier, (c) client selection ratio $r$ of all clients on DomainNet with $\\alpha = 10$ ." + ], + "image_footnote": [], + "bbox": [ + 214, + 142, + 405, + 281 + ], + "page_idx": 12 + }, + { + "type": "image", + "img_path": "images/d9dcc7a26c79361d656c539b7f6c591d3239425791b64d2c5a619ca75067f725.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 413, + 142, + 596, + 280 + ], + "page_idx": 12 + }, + { + "type": "image", + "img_path": "images/f115331d93e254b7f65dcf2c12690719287aaeab05b8e4474a7546320723e626.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 602, + 142, + 787, + 280 + ], + "page_idx": 12 + }, + { + "type": "text", + "text": "FL: local training epoch $E$ ; sample batch size $B$ , and client selection ratio for global aggregation $r$ on DomainNet with $\\alpha = 10$ to carry out experiments. Figure 3 shows the performance of pFedDIL under different configurations. pFedDIL achieves better results when we increase the local training epochs at the beginning. However, pFedDIL has a comparable performance when the $E$ is set to 20 and 40. The variance resulting from $E$ and $r$ being less than 0.6 and 13 percentage points respectively is comparable to that of existing works and is not significant. In addition, pFedDIL achieves similar performance with different sample sizes $B$ .", + "bbox": [ + 212, + 363, + 787, + 484 + ], + "page_idx": 12 + }, + { + "type": "table", + "img_path": "images/1b5cb1b508973e33250fe9422cd978834a2010434ede295ad42e339b8445c9eb.jpg", + "table_caption": [ + "Table 5: Performance of pFedDIL approach on all datasets with different hyper-parameter $\\lambda$ values. Here we measure average accuracy and model size over all tasks on each client (each model including the auxiliary classifier costs about 22Mb storage)." + ], + "table_footnote": [], + "table_body": "
Datasetλ = 0.0λ = 0.2λ = 0.5λ = 0.8λ = 1.0
Digit-10Avg ACC82.4385.0289.2890.7793.94
Model Size22Mb37Mb46Mb71Mb88Mb
Office-31Avg ACC33.3540.5139.9846.1750.38
Model Size22Mb42Mb48Mb53Mb66Mb
DomainNetAvg ACC45.3248.1949.6551.6953.75
Model Size22Mb73Mb84Mb91Mb132Mb
", + "bbox": [ + 289, + 566, + 709, + 719 + ], + "page_idx": 12 + }, + { + "type": "text", + "text": "Then, we conduct more research on the setting of hyper-parameter $\\lambda$ . When $\\lambda = 0.0$ , pFedDIL will degenerate to the normal FedAvg algorithm without any knowledge migration with the single model. When $\\lambda$ is set to 1.0, pFedDIL will separately train models for each domain with knowledge migration. As shown in Table 5, we select five different $\\lambda$ values to conduct experiments and evaluate the average accuracy and average model size on all personalized local models. In Table 1,2, and 3, we utilize ap", + "bbox": [ + 212, + 750, + 787, + 840 + ], + "page_idx": 12 + }, + { + "type": "header", + "text": "Federated Domain-Incremental Learning", + "bbox": [ + 486, + 114, + 732, + 128 + ], + "page_idx": 12 + }, + { + "type": "page_number", + "text": "13", + "bbox": [ + 769, + 114, + 784, + 126 + ], + "page_idx": 12 + }, + { + "type": "text", + "text": "proximately two auxiliary classifiers per client in Digit-10 and Office-31 and about four in DomainNet. For Table 1,2, we carefully select the optimal value of $\\lambda$ for each dataset, striking a balance between accuracy and storage. As shown in Table 5, pFedDIL consistently outperforms other baselines even when $\\lambda$ is simply set to 0.5 across all datasets (around two auxiliary classifiers per client in Digit-10 and Office-31 and approximately three in DomainNet). This suggests that a simple parameter setting can outperform existing methods, and with a more refined parameter setting, we can better accommodate practical considerations, achieving an optimal trade-off between accuracy and storage for each client. The flexibility of pFedDIL allows users to tune this parameter based on their specific requirements and constraints in real-world scenarios.", + "bbox": [ + 212, + 146, + 787, + 297 + ], + "page_idx": 13 + }, + { + "type": "text", + "text": "5 Limitations", + "text_level": 1, + "bbox": [ + 215, + 323, + 346, + 339 + ], + "page_idx": 13 + }, + { + "type": "text", + "text": "Our work is aimed at solving the forgetting problem under the assumption of shared knowledge between tasks in the domain incremental setting. The problem of learning tasks without similar knowledge is rare in real-world scenarios and violates the principle of incremental learning [11]. The ultimate goal of incremental learning is to better learn new knowledge with the assistance of previous knowledge while also mastering both previous and new knowledge [39]. Furthermore, we investigate the situation where the previous data is strictly unseen to clients, while some research on federated incremental learning assumes that each client can cache some samples for replay [9]. Both situations exist and there is a must to separately develop appropriate methods.", + "bbox": [ + 212, + 357, + 787, + 494 + ], + "page_idx": 13 + }, + { + "type": "text", + "text": "6 Conclusion and Future Work", + "text_level": 1, + "bbox": [ + 215, + 518, + 493, + 535 + ], + "page_idx": 13 + }, + { + "type": "text", + "text": "In this paper, we seek to tackle the catastrophic forgetting in the FDIL scenario. We propose a personalized federated domain-incremental learning approach based on adaptive knowledge matching, named pFedDIL. We leverage the auxiliary classifier to calculate the knowledge-matching intensity for the incremental task-learning strategy selection and knowledge migration. Furthermore, we propose sharing partial parameters between the target classification model and the auxiliary classifier to condense model parameters. Extensive experiments conducted on various datasets and settings show that our method achieves significant improvement in accuracy. Although existing works and our method have demonstrated great effectiveness over the FCIL and FDIL scenarios separately, none of them take both scenarios into account. To deploy the FL system in practical settings, it is necessary to simultaneously handle the class incremental tasks and domain incremental tasks. In the future, we seek to move forward in this field.", + "bbox": [ + 212, + 553, + 787, + 734 + ], + "page_idx": 13 + }, + { + "type": "text", + "text": "Acknowledgements", + "text_level": 1, + "bbox": [ + 215, + 760, + 383, + 777 + ], + "page_idx": 13 + }, + { + "type": "text", + "text": "This work is supported by National Natural Science Foundation of China under grants 62376103, 62302184, 62206102, Science and Technology Support Program of Hubei Province under grant 2022BAA046.", + "bbox": [ + 212, + 794, + 785, + 839 + ], + "page_idx": 13 + }, + { + "type": "page_number", + "text": "14", + "bbox": [ + 217, + 114, + 235, + 126 + ], + "page_idx": 13 + }, + { + "type": "header", + "text": "Yichen Li, et al.", + "bbox": [ + 271, + 114, + 367, + 127 + ], + "page_idx": 13 + }, + { + "type": "text", + "text": "References", + "text_level": 1, + "bbox": [ + 217, + 143, + 310, + 160 + ], + "page_idx": 14 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "1. Bakman, Y.F., Yaldiz, D.N., Ezzeldin, Y.H., Avestimehr, S.: Federated orthogonal training: Mitigating global catastrophic forgetting in continual federated learning. arXiv preprint arXiv:2309.01289 (2023) 1, 2", + "2. Chen, H.J., Cheng, A.C., Juan, D.C., Wei, W., Sun, M.: Mitigating forgetting in online continual learning via instance-aware parameterization. Advances in Neural Information Processing Systems 33, 17466-17477 (2020) 4", + "3. Cohen, G., Afshar, S., Tapson, J., Van Schaik, A.: Emmist: Extending mnist to handwritten letters. In: 2017 international joint conference on neural networks (IJCNN). pp. 2921-2926. IEEE (2017) 9", + "4. Dong, J., Li, H., Cong, Y., Sun, G., Zhang, Y., Van Gool, L.: No one left behind: Real-world federated class-incremental learning. IEEE Transactions on Pattern Analysis and Machine Intelligence 46(4), 2054-2070 (2024). https://doi.org/10.1109/TPAMI.2023.33342135", + "5. Dong, J., Liang, W., Cong, Y., Sun, G.: Heterogeneous forgetting compensation for class incremental learning. In: Proceedings of the IEEE/CVF International Conference on Computer Vision (ICCV). pp. 11742-11751 (October 2023) 5", + "6. Dong, J., Wang, L., Fang, Z., Sun, G., Xu, S., Wang, X., Zhu, Q.: Federated class-incremental learning. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 10164-10173 (2022) 2", + "7. Dong, J., Zhang, D., Cong, Y., Cong, W., Ding, H., Dai, D.: Federated incremental semantic segmentation. In: 2023 IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR). pp. 3934-3943 (2023). https://doi.org/10.1109/CVPR52729.2023.003835", + "8. Dong, X., Yu, Z., Cao, W., Shi, Y., Ma, Q.: A survey on ensemble learning. Frontiers of Computer Science 14, 241-258 (2020) 8", + "9. Fini, E., Lathuiliere, S., Sangineto, E., Nabi, M., Ricci, E.: Online continual learning under extreme memory constraints. In: Computer Vision-ECCV 2020: 16th European Conference, Glasgow, UK, August 23-28, 2020, Proceedings, Part XXVIII 16. pp. 720-735. Springer (2020) 14", + "0. Ganin, Y., Ustinova, E., Ajakan, H., Germain, P., Larochelle, H., Laviolette, F., Marchand, M., Lempitsky, V.: Domain-adversarial training of neural networks. The journal of machine learning research 17(1), 2096-2030 (2016) 9, 10, 11", + "1. Hadsell, R., Rao, D., Rusu, A.A., Pascanu, R.: Embracing change: Continual learning in deep neural networks. Trends in cognitive sciences 24(12), 1028-1040 (2020) 14", + "2. He, K., Zhang, X., Ren, S., Sun, J.: Deep residual learning for image recognition. In: Proceedings of the IEEE conference on computer vision and pattern recognition. pp. 770-778 (2016) 10", + "3. Hull, J.J.: A database for handwritten text recognition research. IEEE Transactions on pattern analysis and machine intelligence 16(5), 550-554 (1994) 9", + "4. Jeong, E., Oh, S., Kim, H., Park, J., Bennis, M., Kim, S.L.: Communication-efficient on-device machine learning: Federated distillation and augmentation under non-id private data. ArXiv abs/1811.11479 (2018) 4", + "5. Jiang, Z., Ren, Y., Lei, M., Zhao, Z.: Fedspeech: Federated text-to-speech with continual learning. In: Proceedings of the Thirtieth International Joint Conference on Artificial Intelligence. IJCAI-2021, International Joint Conferences on Artificial Intelligence Organization (Aug 2021). https://doi.org/10.24963/ijcai.2021/527, http://dx.doi.org/10.24963/ijcai.2021/527 5" + ], + "bbox": [ + 225, + 178, + 785, + 839 + ], + "page_idx": 14 + }, + { + "type": "header", + "text": "Federated Domain-Incremental Learning", + "bbox": [ + 486, + 114, + 730, + 128 + ], + "page_idx": 14 + }, + { + "type": "page_number", + "text": "15", + "bbox": [ + 769, + 114, + 784, + 126 + ], + "page_idx": 14 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "16. Jung, S., Ahn, H., Cha, S., Moon, T.: Continual learning with node-importance based adaptive group sparse regularization. arXiv: Learning (2020) 4", + "17. LeCun, Y., Cortes, C., Burges, C.: Mnist handwritten digit database (2010) 9", + "18. Li, L., Fan, Y., Tse, M., Lin, K.Y.: A review of applications in federated learning. Computers & Industrial Engineering 149, 106854 (2020) 4", + "19. Li, T., Sahu, A.K., Talwalkar, A., Smith, V.: Federated learning: Challenges, methods, and future directions. IEEE signal processing magazine 37(3), 50-60 (2020) 4, 9, 10, 11", + "20. Li, Y., Li, Q., Wang, H., Li, R., Zhong, W., Zhang, G.: Towards efficient replay in federated incremental learning. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR). pp. 12820-12829 (June 2024) 2", + "21. Liu, Y., Tian, X., Li, Y., Xiong, Z., Wu, F.: Compact feature learning for multi-domain image classification. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 7193-7201 (2019) 2", + "22. Ma, Y., Xie, Z., Wang, J., Chen, K., Shou, L.: Continual federated learning based on knowledge distillation. In: Raedt, L.D. (ed.) Proceedings of the Thirty-First International Joint Conference on Artificial Intelligence, IJCAI-22. pp. 2182-2188. International Joint Conferences on Artificial Intelligence Organization (7 2022). https://doi.org/10.24963/ijcai.2022/303, https://doi.org/10.24963/ijcai.2022/303, main Track 2", + "23. Maltoni, D., Lomonaco, V.: Continuous learning in single-incremental-task scenarios. Neural networks : the official journal of the International Neural Network Society 116, 56-73 (2018) 4", + "24. McMahan, B., Moore, E., Ramage, D., Hampson, S., y Arcas, B.A.: Communication-efficient learning of deep networks from decentralized data. In: Artificial intelligence and statistics. pp. 1273-1282. PMLR (2017) 1, 4, 9, 10, 11", + "25. Mirza, M.J., Masana, M., Possegger, H., Bischof, H.: An efficient domain-incremental learning approach to drive in all weather conditions. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 3001-3011 (2022) 4, 9", + "26. Netzer, Y., Wang, T., Coates, A., Bissacco, A., Wu, B., Ng, A.Y.: Reading digits in natural images with unsupervised feature learning (2011) 9", + "27. Peng, X., Bai, Q., Xia, X., Huang, Z., Saenko, K., Wang, B.: Moment matching for multi-source domain adaptation. In: Proceedings of the IEEE/CVF international conference on computer vision. pp. 1406-1415 (2019) 9", + "28. Qi, D., Zhao, H., Li, S.: Better generative replay for continual federated learning. arXiv preprint arXiv:2302.13001 (2023) 2, 9, 10, 11", + "29. Rebuffi, S.A., Kolesnikov, A., Sperl, G., Lampert, C.H.: Incremental classifier and representation learning. In: Proceedings of the IEEE conference on Computer Vision and Pattern Recognition. pp. 2001-2010 (2017) 4", + "30. Rolnick, D., Ahuja, A., Schwarz, J., Lillicrap, T., Wayne, G.: Experience replay for continual learning. Advances in Neural Information Processing Systems 32 (2019) 4", + "31. Ruder, S.: An overview of multi-task learning in deep neural networks. arXiv preprint arXiv:1706.05098 (2017) 9", + "32. Saenko, K., Kulis, B., Fritz, M., Darrell, T.: Adapting visual category models to new domains. In: Computer Vision-ECCV 2010: 11th European Conference on Computer Vision, Heraklion, Crete, Greece, September 5-11, 2010, Proceedings, Part IV 11. pp. 213-226. Springer (2010) 9", + "33. Shen, T., Zhang, J., Jia, X., Zhang, F., Huang, G., Zhou, P., Wu, F., Wu, C.: Federated mutual learning. ArXiv abs/2006.16765 (2020) 4", + "34. Standley, T., Zamir, A., Chen, D., Guibas, L., Malik, J., Savarese, S.: Which tasks should be learned together in multi-task learning? In: International Conference on Machine Learning. pp. 9120-9132. PMLR (2020) 8" + ], + "bbox": [ + 215, + 147, + 784, + 839 + ], + "page_idx": 15 + }, + { + "type": "page_number", + "text": "16", + "bbox": [ + 217, + 114, + 235, + 126 + ], + "page_idx": 15 + }, + { + "type": "header", + "text": "Yichen Li, et al.", + "bbox": [ + 271, + 114, + 367, + 127 + ], + "page_idx": 15 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "35. Tankard, C.: What the gdpr means for businesses. Network Security 2016(6), 5-8 (2016) 1", + "36. Van de Ven, G.M., Tuytelaars, T., Tolias, A.S.: Three types of incremental learning. Nature Machine Intelligence 4(12), 1185-1197 (2022) 1", + "37. Wang, H., Li, Y., Xu, W., Li, R., Zhan, Y., Zeng, Z.: Dafkd: Domain-aware federated knowledge distillation. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 20412-20421 (2023) 4", + "38. Wang, H., Xu, H., Li, Y., Xu, Y., Li, R., Zhang, T.: Fedcda: Federated learning with cross-rounds divergence-aware aggregation. In: The Twelfth International Conference on Learning Representations (2023) 4", + "39. Wang, L., Zhang, X., Su, H., Zhu, J.: A comprehensive survey of continual learning: theory, method and application. IEEE Transactions on Pattern Analysis and Machine Intelligence (2024) 4, 14", + "40. Yoon, J., Jeong, W., Lee, G., Yang, E., Hwang, S.J.: Federated continual learning with weighted inter-client transfer. In: International Conference on Machine Learning. pp. 12073-12086. PMLR (2021) 1, 2, 5", + "41. Yu, H., Yang, X., Gao, X., Kang, Y., Wang, H., Zhang, J., Li, T.: Personalized federated continual learning via multi-granularity prompt. arXiv preprint arXiv:2407.00113 (2024) 5", + "42. Zhang, C., Xie, Y., Bai, H., Yu, B., Li, W., Gao, Y.: A survey on federated learning. Knowledge-Based Systems 216, 106775 (2021) 1", + "43. Zhang, J., Chen, C., Zhuang, W., Lyu, L.: Target: Federated class-continual learning via exemplar-free distillation. In: Proceedings of the IEEE/CVF International Conference on Computer Vision. pp. 4782-4793 (2023) 1", + "44. Zhao, H., Zhou, T., Long, G., Jiang, J., Zhang, C.: Does continual learning equally forget all parameters? In: International Conference on Machine Learning. pp. 42280-42303. PMLR (2023) 4" + ], + "bbox": [ + 215, + 146, + 785, + 492 + ], + "page_idx": 16 + }, + { + "type": "header", + "text": "Federated Domain-Incremental Learning", + "bbox": [ + 486, + 114, + 730, + 128 + ], + "page_idx": 16 + }, + { + "type": "page_number", + "text": "17", + "bbox": [ + 767, + 114, + 785, + 126 + ], + "page_idx": 16 + } +] \ No newline at end of file diff --git a/2024/Personalized Federated Domain-Incremental Learning based on Adaptive Knowledge Matching/ea77b70c-2acd-46a7-86ea-e1ba20b951f3_model.json b/2024/Personalized Federated Domain-Incremental Learning based on Adaptive Knowledge Matching/ea77b70c-2acd-46a7-86ea-e1ba20b951f3_model.json new file mode 100644 index 0000000000000000000000000000000000000000..a90860e91f42910cfd47decdf6570ad6ef60517b --- /dev/null +++ b/2024/Personalized Federated Domain-Incremental Learning based on Adaptive Knowledge Matching/ea77b70c-2acd-46a7-86ea-e1ba20b951f3_model.json @@ -0,0 +1,2137 @@ +[ + [ + { + "type": "title", + "bbox": [ + 0.227, + 0.141, + 0.78, + 0.187 + ], + "angle": 0, + "content": "Personalized Federated Domain-Incremental Learning based on Adaptive Knowledge Matching" + }, + { + "type": "text", + "bbox": [ + 0.334, + 0.213, + 0.669, + 0.245 + ], + "angle": 0, + "content": "Yichen Li\\(^{1}\\), Wenchao Xu\\(^{2}\\), Haozhao Wang\\(^{1}\\)*, Yining Qi\\(^{1}\\), Jingcai Guo\\(^{2}\\), and Ruixuan Li\\(^{1}\\)" + }, + { + "type": "text", + "bbox": [ + 0.245, + 0.255, + 0.757, + 0.283 + ], + "angle": 0, + "content": "1 School of Computer Science and Technology, Huazhong University of Science and Technology, Wuhan, China" + }, + { + "type": "text", + "bbox": [ + 0.321, + 0.285, + 0.684, + 0.297 + ], + "angle": 0, + "content": "{ycli0204,hz_wang,qiyining,rxli}@hust.edu.cn" + }, + { + "type": "text", + "bbox": [ + 0.24, + 0.297, + 0.763, + 0.326 + ], + "angle": 0, + "content": "\\(^{2}\\) Department of Computing, The Hong Kong Polytechnic University, Hongkong, China {wenchao.xu, jc-jingcai.guo} @polyu.edu.hk" + }, + { + "type": "text", + "bbox": [ + 0.261, + 0.366, + 0.745, + 0.59 + ], + "angle": 0, + "content": "Abstract. This paper focuses on Federated Domain-Incremental Learning (FDIL) where each client continues to learn incremental tasks where their domain shifts from each other. We propose a novel adaptive knowledge matching-based personalized FDIL approach (pFedDIL) which allows each client to alternatively utilize appropriate incremental task learning strategy on the correlation with the knowledge from previous tasks. More specifically, when a new task arrives, each client first calculates its local correlations with previous tasks. Then, the client can choose to adopt a new initial model or a previous model with similar knowledge to train the new task and simultaneously migrate knowledge from previous tasks based on these correlations. Furthermore, to identify the correlations between the new task and previous tasks for each client, we separately employ an auxiliary classifier to each target classification model and propose sharing partial parameters between the target classification model and the auxiliary classifier to condense model parameters. We conduct extensive experiments on several datasets of which results demonstrate that pFedDIL outperforms state-of-the-art methods by up to \\(14.35\\%\\) in terms of average accuracy of all tasks." + }, + { + "type": "title", + "bbox": [ + 0.217, + 0.615, + 0.357, + 0.631 + ], + "angle": 0, + "content": "1 Introduction" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.649, + 0.788, + 0.816 + ], + "angle": 0, + "content": "Federated Learning (FL) has become a basic paradigm that facilitates collaborative training of deep neural networks (DNNs) from distributed clients while preserving their data privacy [24, 42]. To deploy FL in real-world applications, it is practical that each client can continuously learn new incremental tasks while retaining the accuracy for the previous tasks [40]. However, history data may often be out of reach with privacy constraints like GDPR and many studies have shown that DNNs will forget the knowledge of previous tasks when learning new tasks [35, 36]. Such a phenomenon is known as 'Catastrophic Forgetting', which is a major challenge of incremental learning in the centralized setting. Especially, this challenge is even more formidable to solve in the FL scenario, where the limited previous data are stored in distributed clients and cannot be accessed by each other [1, 43]." + }, + { + "type": "page_footnote", + "bbox": [ + 0.22, + 0.826, + 0.647, + 0.84 + ], + "angle": 0, + "content": "* Haozhao Wang, Yining Qi and Ruixuan Li are corresponding authors." + } + ], + [ + { + "type": "page_number", + "bbox": [ + 0.218, + 0.116, + 0.23, + 0.127 + ], + "angle": 0, + "content": "2" + }, + { + "type": "header", + "bbox": [ + 0.272, + 0.115, + 0.369, + 0.128 + ], + "angle": 0, + "content": "Yichen Li, et al." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.147, + 0.788, + 0.314 + ], + "angle": 0, + "content": "Many recent studies have been proposed to solve the catastrophic forgetting challenge in federated incremental learning. For instance, [40] aims to learn a personalized model for each client by decomposing the model parameters into shared parameters and adaptive parameters and transferring common knowledge between similar tasks across clients. FCIL [6] addresses the federated class-incremental learning and trains a global model by computing additional class-imbalance losses. Authors in [22] utilize extra data at both the server and client sides to conduct knowledge distillation to alleviate forgetting. [28] optimize the generative network and reconstruct previous samples for replay. Recently, Re-Fed is proposed in [20] to prevent catastrophic forgetting with synergistic replay by caching important samples which contribute to both the local and global understanding." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.318, + 0.79, + 0.499 + ], + "angle": 0, + "content": "While these approaches have achieved great success, they are generally designed for the scenario where the new tasks are incremental classes, i.e., Federated Class-Incremental Learning (FCIL). These methods may lose their effectiveness when the new tasks have the same classes but their domain shift with different feature distributions, i.e., Federated Domain-Incremental Learning (FDIL). To elaborate further, the FCIL [6] method relies on changes in the new data category label in incremental tasks to alleviate forgetting of previous tasks; FOT [1] projects different incremental tasks into different orthogonal subspaces, which means feature isolation for different tasks, but if suffers from the intersected feature space boundaries of incremental tasks. The primary challenge in implementing FCIL methods in the context of FDIL is that FCIL methods typically depend on the connection between previous and new classes, which is not applicable in FDIL scenarios where the classes remain consistent." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.504, + 0.79, + 0.64 + ], + "angle": 0, + "content": "In this paper, we seek to tackle the challenge of catastrophic forgetting in the FDIL scenario. A simple way to address this problem is to train different models for each incremental task. However, this requires extensive storage and time-consuming re-training. Additionally, all these models need to be saved for future use which is laborious because we need to manually select the appropriate mode for each inference data. Instead of training a separate model for each different task, a better solution is to ideally have only a single model that can work in all incremental tasks. However, naively training a model for all tasks can lead to terrible catastrophic forgetting when the unique knowledge in the new task overlaps the ones in the previous tasks." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.644, + 0.788, + 0.841 + ], + "angle": 0, + "content": "Unlike in the FCIL scenario, where incremental tasks involve data from different categories, In the FDIL scenario, the incremental tasks involve data from different domains, i.e., handwritten digit data and cartoon digit data. Compared to the data from different categories in FCIL, i.e., cat data and dog data, there is more shared feature information between the different incremental tasks, which can facilitate the overall FDIL task [21]. The key to efficiently learning new tasks while memorizing previous tasks in the FDIL scenario is to identify the correlation between incremental tasks and then share the useful common knowledge in different tasks. Like [40] in FCIL, we consider each client maintains their personalized models while the number of personalized models is under a given threshold due to the limited capacity of each client. Given the correlation between tasks, each client can maintain several personalized models and share the same model between similar tasks. In this way, the client can choose a model with tasks most similar to the new task, enabling the acquisition of new task knowledge while retain-" + } + ], + [ + { + "type": "header", + "bbox": [ + 0.488, + 0.115, + 0.732, + 0.129 + ], + "angle": 0, + "content": "Federated Domain-Incremental Learning" + }, + { + "type": "page_number", + "bbox": [ + 0.776, + 0.117, + 0.786, + 0.127 + ], + "angle": 0, + "content": "3" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.147, + 0.788, + 0.207 + ], + "angle": 0, + "content": "ing existing task knowledge. This dual benefit permits the new task to also capitalize on previous knowledge, ultimately enhancing performance. Nonetheless, considering the tasks come incrementally, identifying the inter-task correlations and subsequently training the selected model is challenging." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.208, + 0.789, + 0.463 + ], + "angle": 0, + "content": "To tackle these limitations, we in this paper propose a novel Personalized Federated Domain-Incremental Learning (pFedDIL) approach which adaptively matches and migrates knowledge among different incremental tasks. More specifically, we propose that each client employs a small auxiliary classifier for each personalized model which is trained to distinguish itself from other tasks. Using the auxiliary classifiers of previous tasks, each client can adaptively identify the correlation of the new task to all previous tasks. Based on these correlations, each client can match the new task with a suitable model, i.e., a new initialized model or a previous model but with similar knowledge. To maximize the training efficiency of the matched model over the new task, we further allow each client to migrate knowledge from other not matched but similar personalized models in a weighted manner, where the weights are the correlations obtained previously. Furthermore, to reduce the size of auxiliary classifiers while enhancing their performance, we propose sharing partial parameters between the target personalized model and the auxiliary classifier. Finally, considering the presence of multiple personalized local models within each client that could lead to additional inference costs, we apply a weighted ensemble distillation over these models where the weights are also obtained from the output of the auxiliary classifiers." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.465, + 0.789, + 0.524 + ], + "angle": 0, + "content": "Through extensive experiments on several datasets and different settings, we show that the proposed pFedDIL significantly improves the model accuracy as compared to state-of-the-art approaches. The major contributions of this paper are summarized as follows:" + }, + { + "type": "text", + "bbox": [ + 0.227, + 0.533, + 0.784, + 0.591 + ], + "angle": 0, + "content": "- We are the first to formulate the problem of federated domain-incremental learning. To address this issue, we propose a novel method named pFedDIL which enables each client to match and migrate knowledge from previous tasks to alleviate catastrophic forgetting." + }, + { + "type": "text", + "bbox": [ + 0.227, + 0.593, + 0.784, + 0.635 + ], + "angle": 0, + "content": "- To adaptively employ incremental task learning strategy and migrate knowledge, we define the knowledge matching intensity by the auxiliary classifier to discern the correlation between the new task and previous tasks." + }, + { + "type": "text", + "bbox": [ + 0.227, + 0.638, + 0.784, + 0.71 + ], + "angle": 0, + "content": "- To facilitate the adaptive selection of inference model among a set of personalized local models, we coordinate all personalized local models to predict weighted by normalized output by auxiliary classifiers. Besides, we further propose sharing partial parameters between the auxiliary classifier and the target classification model to condense model parameters." + }, + { + "type": "text", + "bbox": [ + 0.227, + 0.713, + 0.784, + 0.755 + ], + "angle": 0, + "content": "- We carry out extensive experiments on various datasets and different settings. Experimental results illustrate that our proposed model outperforms the state-of-the-art methods by up to \\(14.35\\%\\) in terms of average accuracy on all tasks." + }, + { + "type": "list", + "bbox": [ + 0.227, + 0.533, + 0.784, + 0.755 + ], + "angle": 0, + "content": null + }, + { + "type": "title", + "bbox": [ + 0.216, + 0.78, + 0.368, + 0.795 + ], + "angle": 0, + "content": "2 Related Work" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.81, + 0.789, + 0.842 + ], + "angle": 0, + "content": "Federated Learning. Federated learning (FL) is a distributed learning approach that trains a global model by merging multiple locally trained models on private datasets" + } + ], + [ + { + "type": "page_number", + "bbox": [ + 0.218, + 0.116, + 0.23, + 0.127 + ], + "angle": 0, + "content": "4" + }, + { + "type": "header", + "bbox": [ + 0.272, + 0.115, + 0.37, + 0.128 + ], + "angle": 0, + "content": "Yichen Li, et al." + }, + { + "type": "image", + "bbox": [ + 0.29, + 0.144, + 0.713, + 0.424 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.214, + 0.435, + 0.788, + 0.491 + ], + "angle": 0, + "content": "Fig. 1: Illustration of the pFedDIL framework. In the training stage, each client randomly receives one domain dataset without replacement to form a task, where different colors represent different domains. Clients can adaptively employ an appropriate incremental task learning strategy and migrate the knowledge from previous tasks with the knowledge matching intensity." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.522, + 0.788, + 0.643 + ], + "angle": 0, + "content": "[18]. A significant challenge in FL is data heterogeneity, where data distribution among clients is non-identically and independently distributed (Non-IID), negatively impacting model performance [14, 33, 38]. FedAvg is a notable FL method that refines the global model by incorporating parameters trained on local private data [24]. [19] propose a proximal term to align closely with the global model. The federated distillation method [37] suggests distilling multiple local models into a global one by only aggregating soft output predictions. However, existing FL approaches typically assume that the data in each client is static, which cannot account for incremental data with domain shift." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.644, + 0.788, + 0.78 + ], + "angle": 0, + "content": "Incremental Learning. Incremental learning (IL) is a learning paradigm that empowers a model to learn sequentially from a series of tasks while preserving the knowledge acquired from previous tasks [39, 44]. IL can be segmented into three primary variants: task-incremental learning [23], class-incremental learning [29] and domain-incremental learning [25]. Existing strategies can further be grouped into three categories: replay-based methods [30], regularization-based methods [16] and parameter isolation methods [2]. Most IL methods are designed for task-IL or class-IL and are typically server-based. This work focuses on the federated domain-incremental learning scenario and proposes the pFedDIL technique." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.78, + 0.788, + 0.842 + ], + "angle": 0, + "content": "Federated Incremental Learning. Federated Incremental Learning (FIL), an approach that targets the dynamic adaptation of streaming tasks within each client's environment, emphasizes the seamless integration of new data into the global model while preserving the acquired knowledge from previous data streams. Despite its crucial role, FIL" + } + ], + [ + { + "type": "header", + "bbox": [ + 0.488, + 0.115, + 0.732, + 0.129 + ], + "angle": 0, + "content": "Federated Domain-Incremental Learning" + }, + { + "type": "page_number", + "bbox": [ + 0.776, + 0.117, + 0.786, + 0.127 + ], + "angle": 0, + "content": "5" + }, + { + "type": "text", + "bbox": [ + 0.215, + 0.147, + 0.788, + 0.283 + ], + "angle": 0, + "content": "has garnered attention relatively recently, with [40] pioneering the field by focusing on Task-Incremental Learning (Task-IL) in FL. In parallel, studies such as [41] takes the multi-granularity expression of knowledge into account, promoting the spatial-temporal integration of knowledge in FCIL. Conversely, [4,5] simplify the problem's complexity by presupposing clients possess sufficient storage capacity to retain historical examples and engage in data sharing, a scenario that deviates from the conventional federated learning paradigm. Beyond image classification, investigations like [7, 15] extend the horizons of FIL into diverse domains. Our research focuses on the domain shift between different tasks and the transfer of the common knowledge." + }, + { + "type": "title", + "bbox": [ + 0.216, + 0.306, + 0.36, + 0.323 + ], + "angle": 0, + "content": "3 Methodology" + }, + { + "type": "text", + "bbox": [ + 0.215, + 0.338, + 0.785, + 0.368 + ], + "angle": 0, + "content": "We first formulate our problem definition and then present our pFedDIL approach. Figure 1 gives an overview of our method." + }, + { + "type": "title", + "bbox": [ + 0.216, + 0.39, + 0.39, + 0.404 + ], + "angle": 0, + "content": "3.1 Problem Definition" + }, + { + "type": "text", + "bbox": [ + 0.215, + 0.415, + 0.789, + 0.524 + ], + "angle": 0, + "content": "In incremental learning, each model learns from a sequence of streaming tasks \\(\\{\\mathcal{T}_{(1)},\\mathcal{T}_{(2)},\\ldots \\mathcal{T}_{(n)}\\}\\) where \\(\\mathcal{T}_{(t)}\\) denotes the dataset of the \\(t\\)-th task; \\(\\mathcal{T}_{(t)} = \\{x_i^t,y_i^t\\}_{i = 1}^{N^t}\\), which has \\(N^t\\) pairs of sample data \\(x_{i}^{t}\\in X^{t}\\) and corresponding label \\(y_{i}^{t}\\in Y^{t}\\). \\(X^{t}\\) and \\(Y^{t}\\) represent the domain space and label space for the \\(t\\)-th task. In this paper, we focus on the variant of domain incremental learning, where all tasks share the same numbers of classes i.e., \\(\\mathcal{V}^1 = \\mathcal{V}^t,\\forall t\\in [1,n]\\). As the sequence of tasks arrives, the client needs to learn the new task while their domain and data distribution changes, i.e., \\(\\mathcal{X}^1\\neq \\mathcal{X}^t,\\forall t\\in [1,n]\\)." + }, + { + "type": "text", + "bbox": [ + 0.215, + 0.525, + 0.788, + 0.651 + ], + "angle": 0, + "content": "We consider the federated domain-incremental learning scenario. We aim to train a global model for \\(K\\) total clients in the federated domain-incremental learning setting. We consider the local client \\(k\\) can only access the local private streaming tasks \\(\\{\\mathcal{T}_{(1)}^k,\\mathcal{T}_{(2)}^k,\\dots ,\\mathcal{T}_{(n)}^k\\}\\). Now our goal is to train a set of personalized local models \\(\\widetilde{w}^{k} = [w_{1}^{k},w_{2}^{k},\\ldots ,w_{d}^{k}]\\) for each client \\(k\\) over all existing \\(t\\) tasks \\(\\mathcal{T} = \\sum_{n = 1}^{t}\\sum_{k = 1}^{K}\\mathcal{T}_{(n)}^{k}\\). Here each personalized local model corresponds to at least one task, and possibly multiple tasks (i.e. client \\(k\\) may train the task \\(\\mathcal{T}_{(1)}^k\\) and \\(\\mathcal{T}_{(2)}^k\\) with the model \\(w_{1}^{k}\\)). Then, we can formulate the goal as:" + }, + { + "type": "equation", + "bbox": [ + 0.371, + 0.661, + 0.787, + 0.702 + ], + "angle": 0, + "content": "\\[\n\\underset {\\widetilde {w}} {\\arg \\min } \\mathcal {L} (\\widetilde {w}) = \\sum_ {n = 1} ^ {t} \\sum_ {k = 1} ^ {K} \\frac {1}{| \\mathcal {T} |} \\mathcal {L} _ {(n)} ^ {k} (\\widetilde {w}). \\tag {1}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.215, + 0.714, + 0.785, + 0.747 + ], + "angle": 0, + "content": "where \\(\\mathcal{L}_{(n)}^k (\\widetilde{w}) = \\mathbb{E}_{(x,y)\\sim \\mathcal{T}_{(n)}^k}\\left[l(\\widetilde{w};(x,y))\\right]\\) is the empirical loss of client \\(k\\) on \\(n\\)-th task with the personalized local model." + }, + { + "type": "title", + "bbox": [ + 0.216, + 0.77, + 0.495, + 0.785 + ], + "angle": 0, + "content": "3.2 The Proposed pFedDIL Approach" + }, + { + "type": "text", + "bbox": [ + 0.215, + 0.795, + 0.787, + 0.841 + ], + "angle": 0, + "content": "The key idea of the pFedDIL is to identify the correlation between the new and previous tasks such that clients can adaptively migrate knowledge from previous tasks to improve performance. Specifically, in each communication round, the client in pFedDIL trains" + } + ], + [ + { + "type": "page_number", + "bbox": [ + 0.218, + 0.116, + 0.23, + 0.127 + ], + "angle": 0, + "content": "6" + }, + { + "type": "header", + "bbox": [ + 0.272, + 0.115, + 0.369, + 0.128 + ], + "angle": 0, + "content": "Yichen Li, et al." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.147, + 0.788, + 0.328 + ], + "angle": 0, + "content": "the local model with its private sequence of tasks and the server aggregates local models from all participating clients. Then, when it comes to new tasks, each client first employs the existing personalized models' auxiliary classifiers to calculate the correlations denoted as knowledge matching intensity and applies the appropriate incremental task learning strategy: training the new task with a previous personalized local model which has the most similar knowledge or training with a new initial model. Simultaneously, each client can migrate knowledge from other personalized local models with the knowledge matching intensity to assist in training. During the inference stage, the final inference result is aggregated by all personalized local models weighted by the corresponding auxiliary classifiers. Moreover, we propose sharing partial parameters between the auxiliary classifier and the target personalized model to condense model parameters. The workflow of the pFedDIL is shown in Algorithm 1." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.329, + 0.788, + 0.529 + ], + "angle": 0, + "content": "Knowledge Matching Intensity. To assess the similarity between different incremental tasks, where a higher similarity indicates a greater amount of common knowledge, we propose the knowledge matching intensity. When it comes to the new incremental task: \\((t + 1)\\)-th task, each client first needs to compute the knowledge matching intensity between the \\((t + 1)\\)-th task and previous \\(t\\) tasks. To achieve the adaptive knowledge matching intensity estimation for each client, we define the auxiliary classifier, a binary classification deep neural network \\(\\theta\\) for the corresponding local task \\(T\\), which is trained by the discrimination of the sample belonging to the task \\(T\\). As a new task emerges, the auxiliary classifier can output the similarity between different tasks based on the extracted feature representation. Supposed that the client \\(k\\) has converged on \\(t\\) tasks within \\(d\\) local existing task models when it comes to the incremental \\((t + 1)\\)-th task, the client needs to calculate the knowledge matching intensity \\(\\widetilde{\\rho}_k = [\\rho_1^k,\\rho_2^k,\\dots ,\\rho_d^k ]\\) with each auxiliary classifier \\(\\widetilde{\\theta}^{k} = [\\theta_{1}^{k},\\theta_{2}^{k},\\dots ,\\theta_{d}^{k}]\\):" + }, + { + "type": "equation", + "bbox": [ + 0.4, + 0.54, + 0.787, + 0.582 + ], + "angle": 0, + "content": "\\[\n\\widetilde {\\rho} ^ {k} = \\frac {1}{N ^ {t + 1}} \\sum_ {i = 1} ^ {N ^ {t + 1}} f \\left(x _ {i} ^ {t + 1}; \\widetilde {\\theta} ^ {k}\\right). \\tag {2}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.596, + 0.788, + 0.657 + ], + "angle": 0, + "content": "where the \\( f(\\cdot) \\) denotes the output of the auxiliary classifier, \\( \\widehat{\\theta}^k \\) denotes a set of personalized local auxiliary classifiers in client \\( k \\). Each auxiliary classifier is trained by minimizing the total empirical loss over the corresponding task's local dataset \\( \\mathcal{T} \\), which can be formulated as:" + }, + { + "type": "equation", + "bbox": [ + 0.394, + 0.666, + 0.787, + 0.707 + ], + "angle": 0, + "content": "\\[\n\\mathcal {L} ^ {k} (\\theta) = \\frac {1}{T} \\sum_ {i = 1} ^ {\\tau} \\mathcal {L} _ {C E} \\left(\\theta ; x _ {i}, y _ {i}\\right). \\tag {3}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.719, + 0.787, + 0.749 + ], + "angle": 0, + "content": "where \\(\\mathcal{L}^k (\\theta)\\) is the local loss in the client \\(k\\) and \\(\\mathcal{L}_{CE}\\) is the cross-entropy loss function that measures the difference between the prediction and the ground truth labels." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.75, + 0.788, + 0.841 + ], + "angle": 0, + "content": "Incremental Task Learning Strategy. After acquiring the knowledge matching intensity \\(\\widetilde{\\rho}_k\\) with equation (2), the client can select the appropriate incremental task learning strategy. An intuition is that the client will prefer to train with previous personalized local models when the knowledge matching intensity exceeds the \\(\\lambda\\), which means the high-level similarity between the new task and the previous tasks so that the model can converge on the new task with slight adjustments. Besides, the client will sacrifice" + } + ], + [ + { + "type": "header", + "bbox": [ + 0.488, + 0.115, + 0.733, + 0.13 + ], + "angle": 0, + "content": "Federated Domain-Incremental Learning" + }, + { + "type": "page_number", + "bbox": [ + 0.775, + 0.116, + 0.785, + 0.127 + ], + "angle": 0, + "content": "7" + }, + { + "type": "code_caption", + "bbox": [ + 0.228, + 0.15, + 0.387, + 0.166 + ], + "angle": 0, + "content": "Algorithm 1: pFedDIL" + }, + { + "type": "algorithm", + "bbox": [ + 0.226, + 0.167, + 0.765, + 0.462 + ], + "angle": 0, + "content": "Input : \\(T\\) : the communication round; \\(K\\) : client number; \\(C\\) : the fraction of active client in each round; \\(\\{\\mathcal{T}_{(t)}\\}_{t = 1}^{n}\\) : the distributed dataset with \\(n\\) tasks; w: the parameter of the target classification model; \\(\\theta\\) : the parameter of the auxiliary classifier. 1 Initialize the parameter w and \\(\\theta\\) . \n2 for \\(t = 1\\) to T do \n3 \\(S_{t}\\gets\\) (random set of \\([C\\cdot K]\\) clients); for each selected client \\(k\\in S_t\\) in parallel do \n4 receives \\(w_{t}\\) from the server; \n5 calculate knowledge matching intensity \\(\\widetilde{\\rho}^k\\) with (2); \n6 select learning strategy \\(w_{t + 1}^{k}\\) with (4); \n7 set local models \\(w_{t + 1}^{k}\\) and \\(\\theta_{t + 1}^{k}\\) . \n8 for e = 1 to E do \n9 update \\(\\theta_{t + 1}^{k}\\) with (3); \n10 update \\(w_{t + 1}^{k}\\) through adaptive knowledge migration with (5); \n11 end \n12 pushes \\(w_{t + 1}^{k}\\) to the server. \n13 end \n14 \\(w_{t}\\gets\\) ServerAggregation(\\(\\{w^{k}\\}_{k\\in S_{t}}\\) ) \n15 end" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.494, + 0.788, + 0.54 + ], + "angle": 0, + "content": "some storage by training the new task with the new initial model parameters to ensure accuracy for both the new task and previous tasks. Here we define the incremental task learning strategy as follows:" + }, + { + "type": "equation", + "bbox": [ + 0.322, + 0.546, + 0.787, + 0.587 + ], + "angle": 0, + "content": "\\[\nw _ {t + 1} ^ {k} = O P T \\left\\{ \\begin{array}{l l} \\hat {w} & i f \\max \\left(\\widetilde {\\rho} _ {k}\\right) < \\lambda \\\\ \\widetilde {w} ^ {k} [ m ] & e l s e \\max \\left(\\widetilde {\\rho} _ {k}\\right) = \\rho_ {k} ^ {m} \\geq \\lambda . \\end{array} \\right. \\tag {4}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.594, + 0.788, + 0.639 + ], + "angle": 0, + "content": "where \\(\\lambda\\) is a hyper-parameter, \\(\\widetilde{w}^k = [w_1^k, w_2^k, \\ldots, w_d^k]\\), \\(\\hat{w}\\) denotes new initial model parameters, and \\(OPT\\) means the alternative of existed personalized local models and a new initial model." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.64, + 0.788, + 0.701 + ], + "angle": 0, + "content": "Knowledge Migration. Then, during the training of \\((t + 1)\\)-th task, each client respectively migrates the knowledge from other existing personalized local models to assist the training with the calculated knowledge matching intensity \\(\\widetilde{\\rho}_k\\), minimize the following objective:" + }, + { + "type": "equation", + "bbox": [ + 0.319, + 0.707, + 0.685, + 0.735 + ], + "angle": 0, + "content": "\\[\n\\min _ {w _ {t + 1} ^ {k}} \\mathcal {L} _ {L o c a l} ^ {k} (w _ {t + 1} ^ {k}) = \\mathcal {L} _ {(t + 1)} ^ {k} (w _ {t + 1} ^ {k}) + \\mathcal {L} _ {K M} ^ {k} (w _ {t + 1} ^ {k}),\n\\]" + }, + { + "type": "equation", + "bbox": [ + 0.319, + 0.738, + 0.786, + 0.802 + ], + "angle": 0, + "content": "\\[\n\\begin{array}{l} \\text {w h e r e} \\quad \\mathcal {L} _ {K M} ^ {k} \\left(w _ {t + 1} ^ {k}\\right) = \\widetilde {\\rho} _ {k} \\cdot \\left| \\left| w _ {t + 1} ^ {k} - \\widetilde {w} ^ {k} \\right| \\right| ^ {2}, \\\\ = \\sum_ {i = 1} ^ {d} \\rho^ {i} \\cdot \\left| \\left| w _ {t + 1} ^ {k} - w _ {i} ^ {k} \\right| \\right| ^ {2}. \\tag {5} \\\\ \\end{array}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.809, + 0.787, + 0.84 + ], + "angle": 0, + "content": "where \\(\\mathcal{L}_{t + 1}^k (w_{t + 1}^k)\\) is defined in (1) and \\(\\mathcal{L}_{KM}^{k}(w_{t + 1}^{k})\\) refers to the Knowledge Migration (KM) loss." + } + ], + [ + { + "type": "page_number", + "bbox": [ + 0.218, + 0.116, + 0.23, + 0.127 + ], + "angle": 0, + "content": "8" + }, + { + "type": "header", + "bbox": [ + 0.272, + 0.115, + 0.369, + 0.128 + ], + "angle": 0, + "content": "Yichen Li, et al." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.147, + 0.788, + 0.286 + ], + "angle": 0, + "content": "Inference Stage. During the inference stage, taking into account that each client may have multiple personalized models, how to automatically select the appropriate model for inference is an urgent problem to solve. We have observed that ensemble learning, which enhances model performance by integrating knowledge from multiple smaller models, has achieved widespread success [8]. Utilizing the calculated knowledge matching intensity as the weight for the personalized model inference results, we achieve automated ensemble inference for the inference data. For every inference sample \\( x \\in \\hat{\\mathcal{T}} \\), the client \\( k \\) computes the weights of each personalized local models \\( \\widetilde{w}^k \\) with the auxiliary \\( \\widetilde{\\theta}^k \\) as \\( f(\\widetilde{\\theta}^k; x) \\) and then normalizes it into the probability:" + }, + { + "type": "equation", + "bbox": [ + 0.425, + 0.294, + 0.787, + 0.332 + ], + "angle": 0, + "content": "\\[\n\\hat {\\alpha} _ {i} ^ {k} = \\frac {f \\left(\\theta_ {i} ^ {k} ; x\\right)}{\\sum_ {i = 1} ^ {d} f \\left(\\theta_ {i} ^ {k} ; x\\right)}. \\tag {6}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.215, + 0.341, + 0.787, + 0.388 + ], + "angle": 0, + "content": "which guarantees that \\(\\sum_{i=1}^{d} \\hat{\\alpha}_i^k = 1\\). Finally, the client inputs the inference sample \\(x\\) into each personalized local models \\(\\widetilde{w}^k\\) to obtain the soft predictions \\(s(\\widetilde{w}^k; x)\\) and final weighed prediction:" + }, + { + "type": "equation", + "bbox": [ + 0.362, + 0.398, + 0.787, + 0.439 + ], + "angle": 0, + "content": "\\[\nI n f e r e n c e \\left(\\widetilde {w} ^ {k}; x\\right) = \\sum_ {i = 1} ^ {d} \\hat {\\alpha} _ {i} ^ {k} \\cdot s \\left(w _ {i} ^ {k}; x\\right). \\tag {7}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.215, + 0.447, + 0.788, + 0.538 + ], + "angle": 0, + "content": "Partial Parameters Sharing. Motivated by the idea of multi-task learning where sharing the encoder between different tasks can promote each other [34], we propose sharing partial parameters between the auxiliary classifier \\(\\theta^k\\) and the target classification model \\(w^k\\) to solve this problem. The intuition behind this is that both the auxiliary classifier and the classification model have to distinguish the sample from the extracted features. Besides, another benefit of sharing layers is that the model parameters can be reduced." + }, + { + "type": "text", + "bbox": [ + 0.215, + 0.538, + 0.788, + 0.645 + ], + "angle": 0, + "content": "To this end, we propose sharing the front model layers between the two models which are used to extract features, as illustrated in Figure 1. Specifically, by denoting the \\(c\\)-layer shared extractor by \\(\\tilde{w}^k = [\\tilde{w}^{k,1},\\dots,\\tilde{w}^{k,c}]\\) where \\(\\tilde{w}^{k,i}\\) is the \\(i\\)-th layer of the shared extractor, we can re-write the \\(n_d\\)-layer auxiliary classifier by \\(\\theta^k = [\\tilde{w}^k,\\theta^{k,c + 1},\\dots,\\theta^{k,n_d}]\\), and the \\(n\\)-layer target classification model \\(w^k\\) by \\(w^k = [\\tilde{w}^k,w^{k,c + 1},\\dots,w^{k,n}]\\) with \\(w^{k,i}\\) as the \\(i\\)-th layer of the classification model \\(w^{k,i}\\). With the shared layers, the discriminator and the classification model are trained jointly:" + }, + { + "type": "equation", + "bbox": [ + 0.368, + 0.652, + 0.787, + 0.678 + ], + "angle": 0, + "content": "\\[\n\\min _ {\\theta^ {k}, w ^ {k}} \\mathcal {L} _ {\\mathcal {J}} ^ {k} \\left(\\theta^ {k}, w ^ {k}\\right) = \\mathcal {L} ^ {k} \\left(\\theta^ {k}\\right) + \\mathcal {L} ^ {k} \\left(w ^ {k}\\right). \\tag {8}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.215, + 0.687, + 0.589, + 0.705 + ], + "angle": 0, + "content": "where \\(\\mathcal{L}^k (\\theta^k)\\) refers to (3) and \\(\\mathcal{L}^k (w^k)\\) is defined in (1)." + }, + { + "type": "title", + "bbox": [ + 0.215, + 0.726, + 0.357, + 0.743 + ], + "angle": 0, + "content": "4 Experiments" + }, + { + "type": "title", + "bbox": [ + 0.215, + 0.756, + 0.299, + 0.772 + ], + "angle": 0, + "content": "4.1 Setup" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.78, + 0.787, + 0.841 + ], + "angle": 0, + "content": "Dataset: We conduct all our experiments with heterogeneous dataset partition on three datasets. In FDIL scenarios, new domains are introduced gradually. The dataset initially contains samples from a specific domain, and new domains are introduced at later stages, enabling models to adapt and generalize to new unseen domains." + } + ], + [ + { + "type": "header", + "bbox": [ + 0.488, + 0.115, + 0.732, + 0.129 + ], + "angle": 0, + "content": "Federated Domain-Incremental Learning" + }, + { + "type": "page_number", + "bbox": [ + 0.776, + 0.117, + 0.786, + 0.127 + ], + "angle": 0, + "content": "9" + }, + { + "type": "table_caption", + "bbox": [ + 0.216, + 0.145, + 0.788, + 0.174 + ], + "angle": 0, + "content": "Table 1: Performance comparison by Top-1 test accuracy for our model and baseline methods on Digit-10 (data heterogeneity: \\(\\alpha = 0.1\\))." + }, + { + "type": "table", + "bbox": [ + 0.261, + 0.185, + 0.744, + 0.41 + ], + "angle": 0, + "content": "
Digit-10 (λ = 0.5)
MethodMNIST → EMNIST → USPS → SVHNAvgΔ(↑)
FedAvg [24]92.82 →84.02 →81.02 →71.8782.436.85↑
FedProx [19]93.01 →84.68 →78.30 →76.4283.106.18↑
FedCIL [28]94.64 →87.52 →82.16 →80.9286.312.97↑
DANN [10] + FL96.07 →86.71 →79.11 →72.1483.515.77↑
Source-Only92.82 →82.15 →75.53 →69.0679.899.39↑
Sharing92.67 →82.91 →76.17 →71.9680.938.35↑
pFedDIL -w/o Migration92.82 →87.09 →84.00 →81.8986.452.83↑
pFedDIL -w/o Sharing92.79 →90.13 →87.59 →86.1989.180.11↑
pFedDIL -w/o Correlation92.82 →89.06 →84.72 →84.3087.731.55↑
pFedDIL92.82 →90.39 →87.14 →86.7589.28/
Upper Bound -Disjoint92.82 →90.33 →93.41 →87.9791.131.85↓
" + }, + { + "type": "text", + "bbox": [ + 0.227, + 0.438, + 0.787, + 0.483 + ], + "angle": 0, + "content": "- Digit10: Digit-10 dataset contains 10 digit categories in four domains: MNIST [17], EMNIST [3], USPS [13], SVHN [26]. Each dataset is a digit image classification dataset of 10 classes in a specific domain, such as handwriting style." + }, + { + "type": "text", + "bbox": [ + 0.227, + 0.484, + 0.787, + 0.528 + ], + "angle": 0, + "content": "- Office31: [32] A dataset with images from three different domains: Amazon, Webcam, and DSLR. It consists of 31 object categories, with each domain having around 4,100 images." + }, + { + "type": "text", + "bbox": [ + 0.227, + 0.529, + 0.787, + 0.574 + ], + "angle": 0, + "content": "- DomainNet: [27] A large-scale dataset with images from six different domains: Clipart, Painting, Real, Sketch, Quickdraw, and Infograph. It contains over 0.6 million images across 345 categories." + }, + { + "type": "list", + "bbox": [ + 0.227, + 0.438, + 0.787, + 0.574 + ], + "angle": 0, + "content": null + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.584, + 0.788, + 0.674 + ], + "angle": 0, + "content": "Baseline: For a fair comparison with other key related works, we follow the same protocols proposed by [24] to set federated learning(FL) task and utilize the domain-incremental learning (DIL) task by [25]. We evaluate all baselines under the above settings, including two representative FL models FedAvg [24] and FedProx [19], which is better at tackling heterogeneity in federated networks than FedAvg, and the following several methods in FDIL scenario:" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.674, + 0.788, + 0.72 + ], + "angle": 0, + "content": "FedCIL [28]: This approach employs the ACGAN backbone to generate synthetic samples to consolidate the global model and align sample features in the output layer. Authors conduct experiments in FCIL scenarios, and here we adopt it to the FDIL setting." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.72, + 0.788, + 0.765 + ], + "angle": 0, + "content": "Domain Adaptation + FL: Here we adopt the robust adversarial-based method DANN [10]. This baseline mainly follows the domain adaptation paradigm which is different from the incremental learning and is often prone to issues like catastrophic forgetting." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.765, + 0.788, + 0.809 + ], + "angle": 0, + "content": "Source-Only: Here we maintain the initial model trained on the initial domain data completely fixed throughout all tasks. This baseline does not perform any adaptation, modification or replacement to its parameters and serves as a lower bound." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.81, + 0.788, + 0.84 + ], + "angle": 0, + "content": "Sharing: Inspired by the multi-task learning scenario [31], we adopt all front layers before the last fully connected layer as shared layers, and we use relevant different" + } + ], + [ + { + "type": "page_number", + "bbox": [ + 0.218, + 0.116, + 0.236, + 0.127 + ], + "angle": 0, + "content": "10" + }, + { + "type": "header", + "bbox": [ + 0.272, + 0.115, + 0.369, + 0.128 + ], + "angle": 0, + "content": "Yichen Li, et al." + }, + { + "type": "table_caption", + "bbox": [ + 0.216, + 0.145, + 0.788, + 0.174 + ], + "angle": 0, + "content": "Table 2: Performance comparison by Top-1 test accuracy for our model and baseline methods on Office-31 (data heterogeneity: \\(\\alpha = 1\\))." + }, + { + "type": "table", + "bbox": [ + 0.29, + 0.185, + 0.714, + 0.41 + ], + "angle": 0, + "content": "
Office-31 (λ = 0.8)
MethodAmazon → Dlsr → WebcamAvgΔ(↑)
FedAvg [24]46.53 → 24.17 → 29.3433.3512.82↑
FedProx [19]45.19 → 25.68 → 32.2334.3711.80↑
FedCIL [28]49.38 → 39.65 → 44.0444.361.81↑
DANN [10] + FL51.97 → 35.96 → 43.0843.672.50↑
Source-Only46.53 → 20.61 → 28.3231.8214.35↑
Sharing46.11 → 24.23 → 34.2534.8611.31↑
pFedDIL -w/o Migration46.53 → 40.17 → 44.5543.752.42↑
pFedDIL -w/o Sharing46.63 → 42.90 → 47.0345.520.65↑
pFedDIL -w/o Correlation46.53 → 41.66 → 47.1945.131.04↑
pFedDIL46.53 → 43.87 → 48.1246.17/
Upper Bound -Disjoint46.53 → 50.17 → 56.2150.974.80↓
" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.439, + 0.784, + 0.469 + ], + "angle": 0, + "content": "fully-connected layers to get outputs of different domains. During inference, the fully-connected layer that matches the corresponding domain ID will be applied." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.47, + 0.785, + 0.531 + ], + "angle": 0, + "content": "Disjoint: Here we train a separate model for each domain. During inference, the model which matches the corresponding domain ID is used. This baseline requires additional space to store parameters and additional cost to obtain inference-task ID. It is a degradation of domain-incremental learning and provides an upper bound here." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.531, + 0.788, + 0.727 + ], + "angle": 0, + "content": "Configuration: Unless otherwise mentioned, we set the number of local training epoch \\( E = 20 \\), communication round \\( T = 180 \\) for each domain. Here we train each task until convergence before the arrival of a new task with communication rounds \\( T \\). For local training, the batch size is 32, the learning rate for our models is 0.001 and the weight decay is \\( 1e - 3 \\). Like most FL methods on Non-IID data distribution [24], we use the Dirichlet distribution \\( (\\alpha) \\) to simulate the data heterogeneity from all domains and we apply the same data heterogeneity for all domains in the same dataset. For the classifier in all methods, we employ ResNet18 [12] as the basic backbone. For the multi-task learning structure in our approach, we treat all previous layers before the last fully connected layer as share layers, and we use two different fully connected layers to get outputs as the auxiliary classifier result and target classification result. The total clients' number is 20 for Digit-10 and DomainNet with an active ratio \\( k = 0.4 \\), 10 for Office-31 with an active ratio \\( k = 0.4 \\)." + }, + { + "type": "title", + "bbox": [ + 0.216, + 0.752, + 0.419, + 0.767 + ], + "angle": 0, + "content": "4.2 Performance Overview" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.78, + 0.788, + 0.841 + ], + "angle": 0, + "content": "Test Accuracy. Table 1,2,3 shows the performance of our pFedDIL and other baseline methods on three datasets, where \" \\(\\Delta\\) \" denotes the improvement of our method with other baselines. We carry out experiments against data heterogeneity on each dataset, and pFedDIL achieves the best performance in most cases by a margin of" + } + ], + [ + { + "type": "header", + "bbox": [ + 0.488, + 0.115, + 0.733, + 0.129 + ], + "angle": 0, + "content": "Federated Domain-Incremental Learning" + }, + { + "type": "page_number", + "bbox": [ + 0.769, + 0.116, + 0.785, + 0.127 + ], + "angle": 0, + "content": "11" + }, + { + "type": "table_caption", + "bbox": [ + 0.215, + 0.145, + 0.788, + 0.174 + ], + "angle": 0, + "content": "Table 3: Performance comparison by Top-10 test accuracy for our model and baseline methods on DomainNet (data heterogeneity: \\(\\alpha = 10\\))" + }, + { + "type": "table", + "bbox": [ + 0.22, + 0.186, + 0.789, + 0.384 + ], + "angle": 0, + "content": "
DomainNet (λ = 0.8)
MethodClipart → Infograph → Painting → Quickdraw → Real → SketchAvgΔ(↑)
FedAvg [24]48.43 → 37.18 → 45.80 → 45.32 → 44.91 → 50.2545.326.38↑
FedProx [19]47.39 → 38.43 → 44.31 → 47.96 → 42.38 → 51.7745.376.32↑
FedCIL [28]52.14 → 43.68 → 47.10 → 48.75 → 42.89 → 51.2647.644.05↑
DANN [10] + FL50.07 → 39.74 → 43.73 → 45.08 → 43.28 → 52.9645.815.88↑
Source-Only48.43 → 31.01 → 33.12 → 38.15 → 37.62 → 42.8538.5313.16↑
Sharing48.43 → 36.03 → 39.58 → 41.80 → 40.41 → 44.7641.849.86↑
pFedDIL -w/o Migration48.43 → 43.83 → 50.05 → 49.44 → 46.10 → 54.9048.792.90↑
pFedDIL -w/o Sharing48.52 → 47.37 → 51.16 → 53.98 → 50.11 → 55.2451.060.63↑
pFedDIL -w/o Correlation48.43 → 46.71 → 51.09 → 52.34 → 48.07 → 56.9550.601.09↑
pFedDIL48.43 → 47.19 → 52.48 → 54.07 → 50.36 → 57.6151.69/
Upper Bound -Disjoint48.43 → 52.33 → 54.21 → 52.89 → 53.62 → 58.0653.261.57↓
" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.413, + 0.789, + 0.504 + ], + "angle": 0, + "content": "\\(1.81\\% \\sim 14.35\\%\\) in terms of average accuracy. Due to our challenging FDIL settings, there is a notable gap between all baselines and the upper performance bound. However, pFedDIL almost achieves comparable results with less storage and adaptive task-id inference during the inference stage. These results demonstrate that the appropriate incremental task learning strategy and adaptive knowledge migration technique in pFedDIL can alleviate catastrophic forgetting and thus improve performance." + }, + { + "type": "image", + "bbox": [ + 0.216, + 0.53, + 0.789, + 0.682 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.283, + 0.69, + 0.72, + 0.705 + ], + "angle": 0, + "content": "Fig. 2: Visualized performance w.r.t data heterogeneity for three datasets." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.734, + 0.789, + 0.842 + ], + "angle": 0, + "content": "Ablation Study. As shown in Tables 1,2,3, we evaluate the effects of each module in our model via ablation studies. pFedDIL -w/o Migration, pFedDIL -w/o Sharing and pFedDIL- w/o Correlation denote the performance of our model without using knowledge migration, parameters sharing and correlation factor in the inference stage. Compared with pFedDIL, the performance of pFedDIL -w/o Migration, pFedDIL -w/o Sharing and pFedDIL- w/o Correlation degrades evidently with a range of \\(0.11\\% \\sim 2.90\\%\\). Specifically, the adaptive knowledge migration technique in pFedDIL plays a much" + } + ], + [ + { + "type": "page_number", + "bbox": [ + 0.218, + 0.116, + 0.236, + 0.127 + ], + "angle": 0, + "content": "12" + }, + { + "type": "header", + "bbox": [ + 0.272, + 0.115, + 0.369, + 0.128 + ], + "angle": 0, + "content": "Yichen Li, et al." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.147, + 0.788, + 0.221 + ], + "angle": 0, + "content": "more important role with the obvious improvement in test accuracy, and the partial parameters sharing trick condenses model parameters by sharing the feature extracted layer while not causing a remarkable impact on test accuracy. Experiment results verify the effectiveness of all modules, confirming all modules are essential to train a robust federated domain-incremental model." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.223, + 0.788, + 0.375 + ], + "angle": 0, + "content": "Data heterogeneity. Increasing data heterogeneity levels in experiments is significant because it allows for a more realistic evaluation of the system's performance in diverse real-world scenarios and assesses its robustness and generalization capabilities. In particular, in the FDIL scenario, incremental data is less likely to appear in the form of IID data. Therefore, it is essential to simulate different levels of Non-IID data to validate the effectiveness of our method in the FDIL scenario. Figure 2 displays the accuracy with different levels of data heterogeneity on three datasets. As shown in this figure, all methods achieve an improvement in test accuracy with the decline of data heterogeneity. Most notably, pFedDIL always maintains a leading advantage with different levels of data heterogeneity." + }, + { + "type": "table_caption", + "bbox": [ + 0.214, + 0.4, + 0.79, + 0.444 + ], + "angle": 0, + "content": "Table 4: Evaluation of various methods in terms of the communication rounds. We report the sum of communication rounds required to achieve the best performance on each task and the running time of a training communication round." + }, + { + "type": "table", + "bbox": [ + 0.233, + 0.455, + 0.773, + 0.555 + ], + "angle": 0, + "content": "
DatasetFedAvgFedCILDANN+FLpFedDIL
RoundsTimeRoundsTimeRoundsTimeRoundsTime
Digit10480(r)32.14(s)492(r)50.06(s)553(r)46.77(s)379(r)34.91(s)
Office31463(r)44.96(s)482(r)78.32(s)488(r)62.89(s)411(r)47.58(s)
DomainNet835(r)92.22(s)904(r)165.41(s)861(r)158.01(s)725(r)105.38(s)
" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.584, + 0.788, + 0.719 + ], + "angle": 0, + "content": "Computational complexity and communication cost. We treat all previous layers in ResNet18 before the last fully connected layer as shared layers, and we use two different fully connected layers to get outputs as the auxiliary classifier result and target classification result. Compared with baseline methods, we only train one additional fully connected layer as the auxiliary classifier for the binary classification task - the discrimination of the sample belonging to the current task. The training time cost (each communication round) is shown in Table 4. Therefore, the training of the auxiliary classifier will not increase the overhead or computational complexity significantly and hardly affect the training efficiency." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.72, + 0.788, + 0.81 + ], + "angle": 0, + "content": "Table 4 shows the evaluation of various methods in terms of the communication rounds to reach the test accuracy reported in Table 1, 2 and 3 (180 communication rounds each task). Here we show the sum of communication rounds required to achieve the performance on each task. pFedDIL requires the least communication rounds to achieve the reported test accuracy on all datasets with the knowledge matching. Thus, the local model is easier to converge." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.81, + 0.788, + 0.842 + ], + "angle": 0, + "content": "Parameter sensitivity analysis. To figure out whether pFedDIL is sensitive to some specific parameters, we first select some general parameters, which are defined in most" + } + ], + [ + { + "type": "header", + "bbox": [ + 0.488, + 0.115, + 0.733, + 0.129 + ], + "angle": 0, + "content": "Federated Domain-Incremental Learning" + }, + { + "type": "page_number", + "bbox": [ + 0.77, + 0.116, + 0.785, + 0.127 + ], + "angle": 0, + "content": "13" + }, + { + "type": "image", + "bbox": [ + 0.215, + 0.143, + 0.406, + 0.282 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.414, + 0.143, + 0.597, + 0.281 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.604, + 0.143, + 0.788, + 0.281 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.214, + 0.29, + 0.788, + 0.332 + ], + "angle": 0, + "content": "Fig. 3: Performance of pFedDIL under different configurations. Here, we select three general parameters in the FL setting: (a) local training epoch \\( E \\), (b) sample size \\( B \\) in the classifier, (c) client selection ratio \\( r \\) of all clients on DomainNet with \\( \\alpha = 10 \\)." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.364, + 0.789, + 0.485 + ], + "angle": 0, + "content": "FL: local training epoch \\( E \\); sample batch size \\( B \\), and client selection ratio for global aggregation \\( r \\) on DomainNet with \\( \\alpha = 10 \\) to carry out experiments. Figure 3 shows the performance of pFedDIL under different configurations. pFedDIL achieves better results when we increase the local training epochs at the beginning. However, pFedDIL has a comparable performance when the \\( E \\) is set to 20 and 40. The variance resulting from \\( E \\) and \\( r \\) being less than 0.6 and 13 percentage points respectively is comparable to that of existing works and is not significant. In addition, pFedDIL achieves similar performance with different sample sizes \\( B \\)." + }, + { + "type": "table_caption", + "bbox": [ + 0.214, + 0.513, + 0.79, + 0.556 + ], + "angle": 0, + "content": "Table 5: Performance of pFedDIL approach on all datasets with different hyper-parameter \\(\\lambda\\) values. Here we measure average accuracy and model size over all tasks on each client (each model including the auxiliary classifier costs about 22Mb storage)." + }, + { + "type": "table", + "bbox": [ + 0.29, + 0.568, + 0.71, + 0.72 + ], + "angle": 0, + "content": "
Datasetλ = 0.0λ = 0.2λ = 0.5λ = 0.8λ = 1.0
Digit-10Avg ACC82.4385.0289.2890.7793.94
Model Size22Mb37Mb46Mb71Mb88Mb
Office-31Avg ACC33.3540.5139.9846.1750.38
Model Size22Mb42Mb48Mb53Mb66Mb
DomainNetAvg ACC45.3248.1949.6551.6953.75
Model Size22Mb73Mb84Mb91Mb132Mb
" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.75, + 0.788, + 0.841 + ], + "angle": 0, + "content": "Then, we conduct more research on the setting of hyper-parameter \\(\\lambda\\). When \\(\\lambda = 0.0\\), pFedDIL will degenerate to the normal FedAvg algorithm without any knowledge migration with the single model. When \\(\\lambda\\) is set to 1.0, pFedDIL will separately train models for each domain with knowledge migration. As shown in Table 5, we select five different \\(\\lambda\\) values to conduct experiments and evaluate the average accuracy and average model size on all personalized local models. In Table 1,2, and 3, we utilize ap" + } + ], + [ + { + "type": "page_number", + "bbox": [ + 0.218, + 0.116, + 0.236, + 0.127 + ], + "angle": 0, + "content": "14" + }, + { + "type": "header", + "bbox": [ + 0.272, + 0.115, + 0.369, + 0.128 + ], + "angle": 0, + "content": "Yichen Li, et al." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.147, + 0.788, + 0.299 + ], + "angle": 0, + "content": "proximately two auxiliary classifiers per client in Digit-10 and Office-31 and about four in DomainNet. For Table 1,2, we carefully select the optimal value of \\(\\lambda\\) for each dataset, striking a balance between accuracy and storage. As shown in Table 5, pFedDIL consistently outperforms other baselines even when \\(\\lambda\\) is simply set to 0.5 across all datasets (around two auxiliary classifiers per client in Digit-10 and Office-31 and approximately three in DomainNet). This suggests that a simple parameter setting can outperform existing methods, and with a more refined parameter setting, we can better accommodate practical considerations, achieving an optimal trade-off between accuracy and storage for each client. The flexibility of pFedDIL allows users to tune this parameter based on their specific requirements and constraints in real-world scenarios." + }, + { + "type": "title", + "bbox": [ + 0.216, + 0.324, + 0.348, + 0.34 + ], + "angle": 0, + "content": "5 Limitations" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.358, + 0.788, + 0.495 + ], + "angle": 0, + "content": "Our work is aimed at solving the forgetting problem under the assumption of shared knowledge between tasks in the domain incremental setting. The problem of learning tasks without similar knowledge is rare in real-world scenarios and violates the principle of incremental learning [11]. The ultimate goal of incremental learning is to better learn new knowledge with the assistance of previous knowledge while also mastering both previous and new knowledge [39]. Furthermore, we investigate the situation where the previous data is strictly unseen to clients, while some research on federated incremental learning assumes that each client can cache some samples for replay [9]. Both situations exist and there is a must to separately develop appropriate methods." + }, + { + "type": "title", + "bbox": [ + 0.216, + 0.52, + 0.495, + 0.536 + ], + "angle": 0, + "content": "6 Conclusion and Future Work" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.554, + 0.788, + 0.736 + ], + "angle": 0, + "content": "In this paper, we seek to tackle the catastrophic forgetting in the FDIL scenario. We propose a personalized federated domain-incremental learning approach based on adaptive knowledge matching, named pFedDIL. We leverage the auxiliary classifier to calculate the knowledge-matching intensity for the incremental task-learning strategy selection and knowledge migration. Furthermore, we propose sharing partial parameters between the target classification model and the auxiliary classifier to condense model parameters. Extensive experiments conducted on various datasets and settings show that our method achieves significant improvement in accuracy. Although existing works and our method have demonstrated great effectiveness over the FCIL and FDIL scenarios separately, none of them take both scenarios into account. To deploy the FL system in practical settings, it is necessary to simultaneously handle the class incremental tasks and domain incremental tasks. In the future, we seek to move forward in this field." + }, + { + "type": "title", + "bbox": [ + 0.216, + 0.761, + 0.384, + 0.779 + ], + "angle": 0, + "content": "Acknowledgements" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.795, + 0.787, + 0.84 + ], + "angle": 0, + "content": "This work is supported by National Natural Science Foundation of China under grants 62376103, 62302184, 62206102, Science and Technology Support Program of Hubei Province under grant 2022BAA046." + } + ], + [ + { + "type": "header", + "bbox": [ + 0.488, + 0.115, + 0.732, + 0.129 + ], + "angle": 0, + "content": "Federated Domain-Incremental Learning" + }, + { + "type": "page_number", + "bbox": [ + 0.77, + 0.116, + 0.785, + 0.127 + ], + "angle": 0, + "content": "15" + }, + { + "type": "title", + "bbox": [ + 0.218, + 0.145, + 0.312, + 0.161 + ], + "angle": 0, + "content": "References" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.179, + 0.787, + 0.221 + ], + "angle": 0, + "content": "1. Bakman, Y.F., Yaldiz, D.N., Ezzeldin, Y.H., Avestimehr, S.: Federated orthogonal training: Mitigating global catastrophic forgetting in continual federated learning. arXiv preprint arXiv:2309.01289 (2023) 1, 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.223, + 0.787, + 0.263 + ], + "angle": 0, + "content": "2. Chen, H.J., Cheng, A.C., Juan, D.C., Wei, W., Sun, M.: Mitigating forgetting in online continual learning via instance-aware parameterization. Advances in Neural Information Processing Systems 33, 17466-17477 (2020) 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.265, + 0.787, + 0.305 + ], + "angle": 0, + "content": "3. Cohen, G., Afshar, S., Tapson, J., Van Schaik, A.: Emmist: Extending mnist to handwritten letters. In: 2017 international joint conference on neural networks (IJCNN). pp. 2921-2926. IEEE (2017) 9" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.307, + 0.787, + 0.36 + ], + "angle": 0, + "content": "4. Dong, J., Li, H., Cong, Y., Sun, G., Zhang, Y., Van Gool, L.: No one left behind: Real-world federated class-incremental learning. IEEE Transactions on Pattern Analysis and Machine Intelligence 46(4), 2054-2070 (2024). https://doi.org/10.1109/TPAMI.2023.33342135" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.362, + 0.787, + 0.404 + ], + "angle": 0, + "content": "5. Dong, J., Liang, W., Cong, Y., Sun, G.: Heterogeneous forgetting compensation for class incremental learning. In: Proceedings of the IEEE/CVF International Conference on Computer Vision (ICCV). pp. 11742-11751 (October 2023) 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.405, + 0.787, + 0.446 + ], + "angle": 0, + "content": "6. Dong, J., Wang, L., Fang, Z., Sun, G., Xu, S., Wang, X., Zhu, Q.: Federated class-incremental learning. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 10164-10173 (2022) 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.448, + 0.787, + 0.5 + ], + "angle": 0, + "content": "7. Dong, J., Zhang, D., Cong, Y., Cong, W., Ding, H., Dai, D.: Federated incremental semantic segmentation. In: 2023 IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR). pp. 3934-3943 (2023). https://doi.org/10.1109/CVPR52729.2023.003835" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.502, + 0.787, + 0.531 + ], + "angle": 0, + "content": "8. Dong, X., Yu, Z., Cao, W., Shi, Y., Ma, Q.: A survey on ensemble learning. Frontiers of Computer Science 14, 241-258 (2020) 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.532, + 0.787, + 0.586 + ], + "angle": 0, + "content": "9. Fini, E., Lathuiliere, S., Sangineto, E., Nabi, M., Ricci, E.: Online continual learning under extreme memory constraints. In: Computer Vision-ECCV 2020: 16th European Conference, Glasgow, UK, August 23-28, 2020, Proceedings, Part XXVIII 16. pp. 720-735. Springer (2020) 14" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.588, + 0.787, + 0.629 + ], + "angle": 0, + "content": "0. Ganin, Y., Ustinova, E., Ajakan, H., Germain, P., Larochelle, H., Laviolette, F., Marchand, M., Lempitsky, V.: Domain-adversarial training of neural networks. The journal of machine learning research 17(1), 2096-2030 (2016) 9, 10, 11" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.63, + 0.787, + 0.657 + ], + "angle": 0, + "content": "1. Hadsell, R., Rao, D., Rusu, A.A., Pascanu, R.: Embracing change: Continual learning in deep neural networks. Trends in cognitive sciences 24(12), 1028-1040 (2020) 14" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.659, + 0.787, + 0.699 + ], + "angle": 0, + "content": "2. He, K., Zhang, X., Ren, S., Sun, J.: Deep residual learning for image recognition. In: Proceedings of the IEEE conference on computer vision and pattern recognition. pp. 770-778 (2016) 10" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.701, + 0.787, + 0.728 + ], + "angle": 0, + "content": "3. Hull, J.J.: A database for handwritten text recognition research. IEEE Transactions on pattern analysis and machine intelligence 16(5), 550-554 (1994) 9" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.729, + 0.787, + 0.77 + ], + "angle": 0, + "content": "4. Jeong, E., Oh, S., Kim, H., Park, J., Bennis, M., Kim, S.L.: Communication-efficient on-device machine learning: Federated distillation and augmentation under non-id private data. ArXiv abs/1811.11479 (2018) 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.772, + 0.787, + 0.84 + ], + "angle": 0, + "content": "5. Jiang, Z., Ren, Y., Lei, M., Zhao, Z.: Fedspeech: Federated text-to-speech with continual learning. In: Proceedings of the Thirtieth International Joint Conference on Artificial Intelligence. IJCAI-2021, International Joint Conferences on Artificial Intelligence Organization (Aug 2021). https://doi.org/10.24963/ijcai.2021/527, http://dx.doi.org/10.24963/ijcai.2021/527 5" + }, + { + "type": "list", + "bbox": [ + 0.226, + 0.179, + 0.787, + 0.84 + ], + "angle": 0, + "content": null + } + ], + [ + { + "type": "page_number", + "bbox": [ + 0.218, + 0.116, + 0.236, + 0.127 + ], + "angle": 0, + "content": "16" + }, + { + "type": "header", + "bbox": [ + 0.272, + 0.115, + 0.369, + 0.128 + ], + "angle": 0, + "content": "Yichen Li, et al." + }, + { + "type": "ref_text", + "bbox": [ + 0.218, + 0.148, + 0.785, + 0.175 + ], + "angle": 0, + "content": "16. Jung, S., Ahn, H., Cha, S., Moon, T.: Continual learning with node-importance based adaptive group sparse regularization. arXiv: Learning (2020) 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.218, + 0.176, + 0.699, + 0.188 + ], + "angle": 0, + "content": "17. LeCun, Y., Cortes, C., Burges, C.: Mnist handwritten digit database (2010) 9" + }, + { + "type": "ref_text", + "bbox": [ + 0.218, + 0.189, + 0.785, + 0.215 + ], + "angle": 0, + "content": "18. Li, L., Fan, Y., Tse, M., Lin, K.Y.: A review of applications in federated learning. Computers & Industrial Engineering 149, 106854 (2020) 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.218, + 0.216, + 0.785, + 0.242 + ], + "angle": 0, + "content": "19. Li, T., Sahu, A.K., Talwalkar, A., Smith, V.: Federated learning: Challenges, methods, and future directions. IEEE signal processing magazine 37(3), 50-60 (2020) 4, 9, 10, 11" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.243, + 0.785, + 0.283 + ], + "angle": 0, + "content": "20. Li, Y., Li, Q., Wang, H., Li, R., Zhong, W., Zhang, G.: Towards efficient replay in federated incremental learning. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR). pp. 12820-12829 (June 2024) 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.284, + 0.785, + 0.323 + ], + "angle": 0, + "content": "21. Liu, Y., Tian, X., Li, Y., Xiong, Z., Wu, F.: Compact feature learning for multi-domain image classification. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 7193-7201 (2019) 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.324, + 0.785, + 0.404 + ], + "angle": 0, + "content": "22. Ma, Y., Xie, Z., Wang, J., Chen, K., Shou, L.: Continual federated learning based on knowledge distillation. In: Raedt, L.D. (ed.) Proceedings of the Thirty-First International Joint Conference on Artificial Intelligence, IJCAI-22. pp. 2182-2188. International Joint Conferences on Artificial Intelligence Organization (7 2022). https://doi.org/10.24963/ijcai.2022/303, https://doi.org/10.24963/ijcai.2022/303, main Track 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.406, + 0.785, + 0.445 + ], + "angle": 0, + "content": "23. Maltoni, D., Lomonaco, V.: Continuous learning in single-incremental-task scenarios. Neural networks : the official journal of the International Neural Network Society 116, 56-73 (2018) 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.447, + 0.785, + 0.488 + ], + "angle": 0, + "content": "24. McMahan, B., Moore, E., Ramage, D., Hampson, S., y Arcas, B.A.: Communication-efficient learning of deep networks from decentralized data. In: Artificial intelligence and statistics. pp. 1273-1282. PMLR (2017) 1, 4, 9, 10, 11" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.488, + 0.785, + 0.528 + ], + "angle": 0, + "content": "25. Mirza, M.J., Masana, M., Possegger, H., Bischof, H.: An efficient domain-incremental learning approach to drive in all weather conditions. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 3001-3011 (2022) 4, 9" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.529, + 0.785, + 0.555 + ], + "angle": 0, + "content": "26. Netzer, Y., Wang, T., Coates, A., Bissacco, A., Wu, B., Ng, A.Y.: Reading digits in natural images with unsupervised feature learning (2011) 9" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.556, + 0.785, + 0.596 + ], + "angle": 0, + "content": "27. Peng, X., Bai, Q., Xia, X., Huang, Z., Saenko, K., Wang, B.: Moment matching for multi-source domain adaptation. In: Proceedings of the IEEE/CVF international conference on computer vision. pp. 1406-1415 (2019) 9" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.597, + 0.785, + 0.623 + ], + "angle": 0, + "content": "28. Qi, D., Zhao, H., Li, S.: Better generative replay for continual federated learning. arXiv preprint arXiv:2302.13001 (2023) 2, 9, 10, 11" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.624, + 0.785, + 0.663 + ], + "angle": 0, + "content": "29. Rebuffi, S.A., Kolesnikov, A., Sperl, G., Lampert, C.H.: Incremental classifier and representation learning. In: Proceedings of the IEEE conference on Computer Vision and Pattern Recognition. pp. 2001-2010 (2017) 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.664, + 0.785, + 0.69 + ], + "angle": 0, + "content": "30. Rolnick, D., Ahuja, A., Schwarz, J., Lillicrap, T., Wayne, G.: Experience replay for continual learning. Advances in Neural Information Processing Systems 32 (2019) 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.691, + 0.785, + 0.717 + ], + "angle": 0, + "content": "31. Ruder, S.: An overview of multi-task learning in deep neural networks. arXiv preprint arXiv:1706.05098 (2017) 9" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.718, + 0.785, + 0.772 + ], + "angle": 0, + "content": "32. Saenko, K., Kulis, B., Fritz, M., Darrell, T.: Adapting visual category models to new domains. In: Computer Vision-ECCV 2010: 11th European Conference on Computer Vision, Heraklion, Crete, Greece, September 5-11, 2010, Proceedings, Part IV 11. pp. 213-226. Springer (2010) 9" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.773, + 0.785, + 0.799 + ], + "angle": 0, + "content": "33. Shen, T., Zhang, J., Jia, X., Zhang, F., Huang, G., Zhou, P., Wu, F., Wu, C.: Federated mutual learning. ArXiv abs/2006.16765 (2020) 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.8, + 0.785, + 0.84 + ], + "angle": 0, + "content": "34. Standley, T., Zamir, A., Chen, D., Guibas, L., Malik, J., Savarese, S.: Which tasks should be learned together in multi-task learning? In: International Conference on Machine Learning. pp. 9120-9132. PMLR (2020) 8" + }, + { + "type": "list", + "bbox": [ + 0.217, + 0.148, + 0.785, + 0.84 + ], + "angle": 0, + "content": null + } + ], + [ + { + "type": "header", + "bbox": [ + 0.488, + 0.115, + 0.732, + 0.129 + ], + "angle": 0, + "content": "Federated Domain-Incremental Learning" + }, + { + "type": "page_number", + "bbox": [ + 0.769, + 0.116, + 0.786, + 0.127 + ], + "angle": 0, + "content": "17" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.147, + 0.785, + 0.161 + ], + "angle": 0, + "content": "35. Tankard, C.: What the gdpr means for businesses. Network Security 2016(6), 5-8 (2016) 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.162, + 0.787, + 0.188 + ], + "angle": 0, + "content": "36. Van de Ven, G.M., Tuytelaars, T., Tolias, A.S.: Three types of incremental learning. Nature Machine Intelligence 4(12), 1185-1197 (2022) 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.189, + 0.787, + 0.231 + ], + "angle": 0, + "content": "37. Wang, H., Li, Y., Xu, W., Li, R., Zhan, Y., Zeng, Z.: Dafkd: Domain-aware federated knowledge distillation. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 20412-20421 (2023) 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.231, + 0.787, + 0.272 + ], + "angle": 0, + "content": "38. Wang, H., Xu, H., Li, Y., Xu, Y., Li, R., Zhang, T.: Fedcda: Federated learning with cross-rounds divergence-aware aggregation. In: The Twelfth International Conference on Learning Representations (2023) 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.273, + 0.787, + 0.313 + ], + "angle": 0, + "content": "39. Wang, L., Zhang, X., Su, H., Zhu, J.: A comprehensive survey of continual learning: theory, method and application. IEEE Transactions on Pattern Analysis and Machine Intelligence (2024) 4, 14" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.314, + 0.787, + 0.355 + ], + "angle": 0, + "content": "40. Yoon, J., Jeong, W., Lee, G., Yang, E., Hwang, S.J.: Federated continual learning with weighted inter-client transfer. In: International Conference on Machine Learning. pp. 12073-12086. PMLR (2021) 1, 2, 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.356, + 0.787, + 0.384 + ], + "angle": 0, + "content": "41. Yu, H., Yang, X., Gao, X., Kang, Y., Wang, H., Zhang, J., Li, T.: Personalized federated continual learning via multi-granularity prompt. arXiv preprint arXiv:2407.00113 (2024) 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.384, + 0.787, + 0.411 + ], + "angle": 0, + "content": "42. Zhang, C., Xie, Y., Bai, H., Yu, B., Li, W., Gao, Y.: A survey on federated learning. Knowledge-Based Systems 216, 106775 (2021) 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.411, + 0.787, + 0.452 + ], + "angle": 0, + "content": "43. Zhang, J., Chen, C., Zhuang, W., Lyu, L.: Target: Federated class-continual learning via exemplar-free distillation. In: Proceedings of the IEEE/CVF International Conference on Computer Vision. pp. 4782-4793 (2023) 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.453, + 0.787, + 0.493 + ], + "angle": 0, + "content": "44. Zhao, H., Zhou, T., Long, G., Jiang, J., Zhang, C.: Does continual learning equally forget all parameters? In: International Conference on Machine Learning. pp. 42280-42303. PMLR (2023) 4" + }, + { + "type": "list", + "bbox": [ + 0.217, + 0.147, + 0.787, + 0.493 + ], + "angle": 0, + "content": null + } + ] +] \ No newline at end of file diff --git a/2024/Personalized Federated Domain-Incremental Learning based on Adaptive Knowledge Matching/ea77b70c-2acd-46a7-86ea-e1ba20b951f3_origin.pdf b/2024/Personalized Federated Domain-Incremental Learning based on Adaptive Knowledge Matching/ea77b70c-2acd-46a7-86ea-e1ba20b951f3_origin.pdf new file mode 100644 index 0000000000000000000000000000000000000000..da577fd4a0f1e4f9d5d5835b52f3530dd0a51ff8 --- /dev/null +++ b/2024/Personalized Federated Domain-Incremental Learning based on Adaptive Knowledge Matching/ea77b70c-2acd-46a7-86ea-e1ba20b951f3_origin.pdf @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:9113df70d9480ecbe6ae377f6e4f4653f8389586205933d84bd248670ce4e7ba +size 778710 diff --git a/2024/Personalized Federated Domain-Incremental Learning based on Adaptive Knowledge Matching/full.md b/2024/Personalized Federated Domain-Incremental Learning based on Adaptive Knowledge Matching/full.md new file mode 100644 index 0000000000000000000000000000000000000000..127960e10eb2ce5e9a4b4b9e914717e384e78102 --- /dev/null +++ b/2024/Personalized Federated Domain-Incremental Learning based on Adaptive Knowledge Matching/full.md @@ -0,0 +1,287 @@ +# Personalized Federated Domain-Incremental Learning based on Adaptive Knowledge Matching + +Yichen Li $^{1}$ , Wenchao Xu $^{2}$ , Haozhao Wang $^{1}$ *, Yining Qi $^{1}$ , Jingcai Guo $^{2}$ , and Ruixuan Li $^{1}$ + +1 School of Computer Science and Technology, Huazhong University of Science and Technology, Wuhan, China + +{ycli0204,hz_wang,qiyining,rxli}@hust.edu.cn + +$^{2}$ Department of Computing, The Hong Kong Polytechnic University, Hongkong, China {wenchao.xu, jc-jingcai.guo} @polyu.edu.hk + +Abstract. This paper focuses on Federated Domain-Incremental Learning (FDIL) where each client continues to learn incremental tasks where their domain shifts from each other. We propose a novel adaptive knowledge matching-based personalized FDIL approach (pFedDIL) which allows each client to alternatively utilize appropriate incremental task learning strategy on the correlation with the knowledge from previous tasks. More specifically, when a new task arrives, each client first calculates its local correlations with previous tasks. Then, the client can choose to adopt a new initial model or a previous model with similar knowledge to train the new task and simultaneously migrate knowledge from previous tasks based on these correlations. Furthermore, to identify the correlations between the new task and previous tasks for each client, we separately employ an auxiliary classifier to each target classification model and propose sharing partial parameters between the target classification model and the auxiliary classifier to condense model parameters. We conduct extensive experiments on several datasets of which results demonstrate that pFedDIL outperforms state-of-the-art methods by up to $14.35\%$ in terms of average accuracy of all tasks. + +# 1 Introduction + +Federated Learning (FL) has become a basic paradigm that facilitates collaborative training of deep neural networks (DNNs) from distributed clients while preserving their data privacy [24, 42]. To deploy FL in real-world applications, it is practical that each client can continuously learn new incremental tasks while retaining the accuracy for the previous tasks [40]. However, history data may often be out of reach with privacy constraints like GDPR and many studies have shown that DNNs will forget the knowledge of previous tasks when learning new tasks [35, 36]. Such a phenomenon is known as 'Catastrophic Forgetting', which is a major challenge of incremental learning in the centralized setting. Especially, this challenge is even more formidable to solve in the FL scenario, where the limited previous data are stored in distributed clients and cannot be accessed by each other [1, 43]. + +Many recent studies have been proposed to solve the catastrophic forgetting challenge in federated incremental learning. For instance, [40] aims to learn a personalized model for each client by decomposing the model parameters into shared parameters and adaptive parameters and transferring common knowledge between similar tasks across clients. FCIL [6] addresses the federated class-incremental learning and trains a global model by computing additional class-imbalance losses. Authors in [22] utilize extra data at both the server and client sides to conduct knowledge distillation to alleviate forgetting. [28] optimize the generative network and reconstruct previous samples for replay. Recently, Re-Fed is proposed in [20] to prevent catastrophic forgetting with synergistic replay by caching important samples which contribute to both the local and global understanding. + +While these approaches have achieved great success, they are generally designed for the scenario where the new tasks are incremental classes, i.e., Federated Class-Incremental Learning (FCIL). These methods may lose their effectiveness when the new tasks have the same classes but their domain shift with different feature distributions, i.e., Federated Domain-Incremental Learning (FDIL). To elaborate further, the FCIL [6] method relies on changes in the new data category label in incremental tasks to alleviate forgetting of previous tasks; FOT [1] projects different incremental tasks into different orthogonal subspaces, which means feature isolation for different tasks, but if suffers from the intersected feature space boundaries of incremental tasks. The primary challenge in implementing FCIL methods in the context of FDIL is that FCIL methods typically depend on the connection between previous and new classes, which is not applicable in FDIL scenarios where the classes remain consistent. + +In this paper, we seek to tackle the challenge of catastrophic forgetting in the FDIL scenario. A simple way to address this problem is to train different models for each incremental task. However, this requires extensive storage and time-consuming re-training. Additionally, all these models need to be saved for future use which is laborious because we need to manually select the appropriate mode for each inference data. Instead of training a separate model for each different task, a better solution is to ideally have only a single model that can work in all incremental tasks. However, naively training a model for all tasks can lead to terrible catastrophic forgetting when the unique knowledge in the new task overlaps the ones in the previous tasks. + +Unlike in the FCIL scenario, where incremental tasks involve data from different categories, In the FDIL scenario, the incremental tasks involve data from different domains, i.e., handwritten digit data and cartoon digit data. Compared to the data from different categories in FCIL, i.e., cat data and dog data, there is more shared feature information between the different incremental tasks, which can facilitate the overall FDIL task [21]. The key to efficiently learning new tasks while memorizing previous tasks in the FDIL scenario is to identify the correlation between incremental tasks and then share the useful common knowledge in different tasks. Like [40] in FCIL, we consider each client maintains their personalized models while the number of personalized models is under a given threshold due to the limited capacity of each client. Given the correlation between tasks, each client can maintain several personalized models and share the same model between similar tasks. In this way, the client can choose a model with tasks most similar to the new task, enabling the acquisition of new task knowledge while retain- + +ing existing task knowledge. This dual benefit permits the new task to also capitalize on previous knowledge, ultimately enhancing performance. Nonetheless, considering the tasks come incrementally, identifying the inter-task correlations and subsequently training the selected model is challenging. + +To tackle these limitations, we in this paper propose a novel Personalized Federated Domain-Incremental Learning (pFedDIL) approach which adaptively matches and migrates knowledge among different incremental tasks. More specifically, we propose that each client employs a small auxiliary classifier for each personalized model which is trained to distinguish itself from other tasks. Using the auxiliary classifiers of previous tasks, each client can adaptively identify the correlation of the new task to all previous tasks. Based on these correlations, each client can match the new task with a suitable model, i.e., a new initialized model or a previous model but with similar knowledge. To maximize the training efficiency of the matched model over the new task, we further allow each client to migrate knowledge from other not matched but similar personalized models in a weighted manner, where the weights are the correlations obtained previously. Furthermore, to reduce the size of auxiliary classifiers while enhancing their performance, we propose sharing partial parameters between the target personalized model and the auxiliary classifier. Finally, considering the presence of multiple personalized local models within each client that could lead to additional inference costs, we apply a weighted ensemble distillation over these models where the weights are also obtained from the output of the auxiliary classifiers. + +Through extensive experiments on several datasets and different settings, we show that the proposed pFedDIL significantly improves the model accuracy as compared to state-of-the-art approaches. The major contributions of this paper are summarized as follows: + +- We are the first to formulate the problem of federated domain-incremental learning. To address this issue, we propose a novel method named pFedDIL which enables each client to match and migrate knowledge from previous tasks to alleviate catastrophic forgetting. +- To adaptively employ incremental task learning strategy and migrate knowledge, we define the knowledge matching intensity by the auxiliary classifier to discern the correlation between the new task and previous tasks. +- To facilitate the adaptive selection of inference model among a set of personalized local models, we coordinate all personalized local models to predict weighted by normalized output by auxiliary classifiers. Besides, we further propose sharing partial parameters between the auxiliary classifier and the target classification model to condense model parameters. +- We carry out extensive experiments on various datasets and different settings. Experimental results illustrate that our proposed model outperforms the state-of-the-art methods by up to $14.35\%$ in terms of average accuracy on all tasks. + +# 2 Related Work + +Federated Learning. Federated learning (FL) is a distributed learning approach that trains a global model by merging multiple locally trained models on private datasets + +![](images/7fa8509dae5f4dafbc2482281b406e213d7f24cde0991e0473be9f773dae0213.jpg) +Fig. 1: Illustration of the pFedDIL framework. In the training stage, each client randomly receives one domain dataset without replacement to form a task, where different colors represent different domains. Clients can adaptively employ an appropriate incremental task learning strategy and migrate the knowledge from previous tasks with the knowledge matching intensity. + +[18]. A significant challenge in FL is data heterogeneity, where data distribution among clients is non-identically and independently distributed (Non-IID), negatively impacting model performance [14, 33, 38]. FedAvg is a notable FL method that refines the global model by incorporating parameters trained on local private data [24]. [19] propose a proximal term to align closely with the global model. The federated distillation method [37] suggests distilling multiple local models into a global one by only aggregating soft output predictions. However, existing FL approaches typically assume that the data in each client is static, which cannot account for incremental data with domain shift. + +Incremental Learning. Incremental learning (IL) is a learning paradigm that empowers a model to learn sequentially from a series of tasks while preserving the knowledge acquired from previous tasks [39, 44]. IL can be segmented into three primary variants: task-incremental learning [23], class-incremental learning [29] and domain-incremental learning [25]. Existing strategies can further be grouped into three categories: replay-based methods [30], regularization-based methods [16] and parameter isolation methods [2]. Most IL methods are designed for task-IL or class-IL and are typically server-based. This work focuses on the federated domain-incremental learning scenario and proposes the pFedDIL technique. + +Federated Incremental Learning. Federated Incremental Learning (FIL), an approach that targets the dynamic adaptation of streaming tasks within each client's environment, emphasizes the seamless integration of new data into the global model while preserving the acquired knowledge from previous data streams. Despite its crucial role, FIL + +has garnered attention relatively recently, with [40] pioneering the field by focusing on Task-Incremental Learning (Task-IL) in FL. In parallel, studies such as [41] takes the multi-granularity expression of knowledge into account, promoting the spatial-temporal integration of knowledge in FCIL. Conversely, [4,5] simplify the problem's complexity by presupposing clients possess sufficient storage capacity to retain historical examples and engage in data sharing, a scenario that deviates from the conventional federated learning paradigm. Beyond image classification, investigations like [7, 15] extend the horizons of FIL into diverse domains. Our research focuses on the domain shift between different tasks and the transfer of the common knowledge. + +# 3 Methodology + +We first formulate our problem definition and then present our pFedDIL approach. Figure 1 gives an overview of our method. + +# 3.1 Problem Definition + +In incremental learning, each model learns from a sequence of streaming tasks $\{\mathcal{T}_{(1)},\mathcal{T}_{(2)},\ldots \mathcal{T}_{(n)}\}$ where $\mathcal{T}_{(t)}$ denotes the dataset of the $t$ -th task; $\mathcal{T}_{(t)} = \{x_i^t,y_i^t\}_{i = 1}^{N^t}$ , which has $N^t$ pairs of sample data $x_{i}^{t}\in X^{t}$ and corresponding label $y_{i}^{t}\in Y^{t}$ . $X^{t}$ and $Y^{t}$ represent the domain space and label space for the $t$ -th task. In this paper, we focus on the variant of domain incremental learning, where all tasks share the same numbers of classes i.e., $\mathcal{V}^1 = \mathcal{V}^t,\forall t\in [1,n]$ . As the sequence of tasks arrives, the client needs to learn the new task while their domain and data distribution changes, i.e., $\mathcal{X}^1\neq \mathcal{X}^t,\forall t\in [1,n]$ . + +We consider the federated domain-incremental learning scenario. We aim to train a global model for $K$ total clients in the federated domain-incremental learning setting. We consider the local client $k$ can only access the local private streaming tasks $\{\mathcal{T}_{(1)}^k,\mathcal{T}_{(2)}^k,\dots ,\mathcal{T}_{(n)}^k\}$ . Now our goal is to train a set of personalized local models $\widetilde{w}^{k} = [w_{1}^{k},w_{2}^{k},\ldots ,w_{d}^{k}]$ for each client $k$ over all existing $t$ tasks $\mathcal{T} = \sum_{n = 1}^{t}\sum_{k = 1}^{K}\mathcal{T}_{(n)}^{k}$ . Here each personalized local model corresponds to at least one task, and possibly multiple tasks (i.e. client $k$ may train the task $\mathcal{T}_{(1)}^k$ and $\mathcal{T}_{(2)}^k$ with the model $w_{1}^{k}$ ). Then, we can formulate the goal as: + +$$ +\underset {\widetilde {w}} {\arg \min } \mathcal {L} (\widetilde {w}) = \sum_ {n = 1} ^ {t} \sum_ {k = 1} ^ {K} \frac {1}{| \mathcal {T} |} \mathcal {L} _ {(n)} ^ {k} (\widetilde {w}). \tag {1} +$$ + +where $\mathcal{L}_{(n)}^k (\widetilde{w}) = \mathbb{E}_{(x,y)\sim \mathcal{T}_{(n)}^k}\left[l(\widetilde{w};(x,y))\right]$ is the empirical loss of client $k$ on $n$ -th task with the personalized local model. + +# 3.2 The Proposed pFedDIL Approach + +The key idea of the pFedDIL is to identify the correlation between the new and previous tasks such that clients can adaptively migrate knowledge from previous tasks to improve performance. Specifically, in each communication round, the client in pFedDIL trains + +the local model with its private sequence of tasks and the server aggregates local models from all participating clients. Then, when it comes to new tasks, each client first employs the existing personalized models' auxiliary classifiers to calculate the correlations denoted as knowledge matching intensity and applies the appropriate incremental task learning strategy: training the new task with a previous personalized local model which has the most similar knowledge or training with a new initial model. Simultaneously, each client can migrate knowledge from other personalized local models with the knowledge matching intensity to assist in training. During the inference stage, the final inference result is aggregated by all personalized local models weighted by the corresponding auxiliary classifiers. Moreover, we propose sharing partial parameters between the auxiliary classifier and the target personalized model to condense model parameters. The workflow of the pFedDIL is shown in Algorithm 1. + +Knowledge Matching Intensity. To assess the similarity between different incremental tasks, where a higher similarity indicates a greater amount of common knowledge, we propose the knowledge matching intensity. When it comes to the new incremental task: $(t + 1)$ -th task, each client first needs to compute the knowledge matching intensity between the $(t + 1)$ -th task and previous $t$ tasks. To achieve the adaptive knowledge matching intensity estimation for each client, we define the auxiliary classifier, a binary classification deep neural network $\theta$ for the corresponding local task $T$ , which is trained by the discrimination of the sample belonging to the task $T$ . As a new task emerges, the auxiliary classifier can output the similarity between different tasks based on the extracted feature representation. Supposed that the client $k$ has converged on $t$ tasks within $d$ local existing task models when it comes to the incremental $(t + 1)$ -th task, the client needs to calculate the knowledge matching intensity $\widetilde{\rho}_k = [\rho_1^k,\rho_2^k,\dots ,\rho_d^k ]$ with each auxiliary classifier $\widetilde{\theta}^{k} = [\theta_{1}^{k},\theta_{2}^{k},\dots ,\theta_{d}^{k}]$ : + +$$ +\widetilde {\rho} ^ {k} = \frac {1}{N ^ {t + 1}} \sum_ {i = 1} ^ {N ^ {t + 1}} f \left(x _ {i} ^ {t + 1}; \widetilde {\theta} ^ {k}\right). \tag {2} +$$ + +where the $f(\cdot)$ denotes the output of the auxiliary classifier, $\widehat{\theta}^k$ denotes a set of personalized local auxiliary classifiers in client $k$ . Each auxiliary classifier is trained by minimizing the total empirical loss over the corresponding task's local dataset $\mathcal{T}$ , which can be formulated as: + +$$ +\mathcal {L} ^ {k} (\theta) = \frac {1}{T} \sum_ {i = 1} ^ {\tau} \mathcal {L} _ {C E} \left(\theta ; x _ {i}, y _ {i}\right). \tag {3} +$$ + +where $\mathcal{L}^k (\theta)$ is the local loss in the client $k$ and $\mathcal{L}_{CE}$ is the cross-entropy loss function that measures the difference between the prediction and the ground truth labels. + +Incremental Task Learning Strategy. After acquiring the knowledge matching intensity $\widetilde{\rho}_k$ with equation (2), the client can select the appropriate incremental task learning strategy. An intuition is that the client will prefer to train with previous personalized local models when the knowledge matching intensity exceeds the $\lambda$ , which means the high-level similarity between the new task and the previous tasks so that the model can converge on the new task with slight adjustments. Besides, the client will sacrifice + +Algorithm 1: pFedDIL +Input : $T$ : the communication round; $K$ : client number; $C$ : the fraction of active client in each round; $\{\mathcal{T}_{(t)}\}_{t = 1}^{n}$ : the distributed dataset with $n$ tasks; w: the parameter of the target classification model; $\theta$ : the parameter of the auxiliary classifier. 1 Initialize the parameter w and $\theta$ . +2 for $t = 1$ to T do +3 $S_{t}\gets$ (random set of $[C\cdot K]$ clients); for each selected client $k\in S_t$ in parallel do +4 receives $w_{t}$ from the server; +5 calculate knowledge matching intensity $\widetilde{\rho}^k$ with (2); +6 select learning strategy $w_{t + 1}^{k}$ with (4); +7 set local models $w_{t + 1}^{k}$ and $\theta_{t + 1}^{k}$ . +8 for e = 1 to E do +9 update $\theta_{t + 1}^{k}$ with (3); +10 update $w_{t + 1}^{k}$ through adaptive knowledge migration with (5); +11 end +12 pushes $w_{t + 1}^{k}$ to the server. +13 end +14 $w_{t}\gets$ ServerAggregation( $\{w^{k}\}_{k\in S_{t}}$ ) +15 end + +some storage by training the new task with the new initial model parameters to ensure accuracy for both the new task and previous tasks. Here we define the incremental task learning strategy as follows: + +$$ +w _ {t + 1} ^ {k} = O P T \left\{ \begin{array}{l l} \hat {w} & i f \max \left(\widetilde {\rho} _ {k}\right) < \lambda \\ \widetilde {w} ^ {k} [ m ] & e l s e \max \left(\widetilde {\rho} _ {k}\right) = \rho_ {k} ^ {m} \geq \lambda . \end{array} \right. \tag {4} +$$ + +where $\lambda$ is a hyper-parameter, $\widetilde{w}^k = [w_1^k, w_2^k, \ldots, w_d^k]$ , $\hat{w}$ denotes new initial model parameters, and $OPT$ means the alternative of existed personalized local models and a new initial model. + +Knowledge Migration. Then, during the training of $(t + 1)$ -th task, each client respectively migrates the knowledge from other existing personalized local models to assist the training with the calculated knowledge matching intensity $\widetilde{\rho}_k$ , minimize the following objective: + +$$ +\min _ {w _ {t + 1} ^ {k}} \mathcal {L} _ {L o c a l} ^ {k} (w _ {t + 1} ^ {k}) = \mathcal {L} _ {(t + 1)} ^ {k} (w _ {t + 1} ^ {k}) + \mathcal {L} _ {K M} ^ {k} (w _ {t + 1} ^ {k}), +$$ + +$$ +\begin{array}{l} \text {w h e r e} \quad \mathcal {L} _ {K M} ^ {k} \left(w _ {t + 1} ^ {k}\right) = \widetilde {\rho} _ {k} \cdot \left| \left| w _ {t + 1} ^ {k} - \widetilde {w} ^ {k} \right| \right| ^ {2}, \\ = \sum_ {i = 1} ^ {d} \rho^ {i} \cdot \left| \left| w _ {t + 1} ^ {k} - w _ {i} ^ {k} \right| \right| ^ {2}. \tag {5} \\ \end{array} +$$ + +where $\mathcal{L}_{t + 1}^k (w_{t + 1}^k)$ is defined in (1) and $\mathcal{L}_{KM}^{k}(w_{t + 1}^{k})$ refers to the Knowledge Migration (KM) loss. + +Inference Stage. During the inference stage, taking into account that each client may have multiple personalized models, how to automatically select the appropriate model for inference is an urgent problem to solve. We have observed that ensemble learning, which enhances model performance by integrating knowledge from multiple smaller models, has achieved widespread success [8]. Utilizing the calculated knowledge matching intensity as the weight for the personalized model inference results, we achieve automated ensemble inference for the inference data. For every inference sample $x \in \hat{\mathcal{T}}$ , the client $k$ computes the weights of each personalized local models $\widetilde{w}^k$ with the auxiliary $\widetilde{\theta}^k$ as $f(\widetilde{\theta}^k; x)$ and then normalizes it into the probability: + +$$ +\hat {\alpha} _ {i} ^ {k} = \frac {f \left(\theta_ {i} ^ {k} ; x\right)}{\sum_ {i = 1} ^ {d} f \left(\theta_ {i} ^ {k} ; x\right)}. \tag {6} +$$ + +which guarantees that $\sum_{i=1}^{d} \hat{\alpha}_i^k = 1$ . Finally, the client inputs the inference sample $x$ into each personalized local models $\widetilde{w}^k$ to obtain the soft predictions $s(\widetilde{w}^k; x)$ and final weighed prediction: + +$$ +I n f e r e n c e \left(\widetilde {w} ^ {k}; x\right) = \sum_ {i = 1} ^ {d} \hat {\alpha} _ {i} ^ {k} \cdot s \left(w _ {i} ^ {k}; x\right). \tag {7} +$$ + +Partial Parameters Sharing. Motivated by the idea of multi-task learning where sharing the encoder between different tasks can promote each other [34], we propose sharing partial parameters between the auxiliary classifier $\theta^k$ and the target classification model $w^k$ to solve this problem. The intuition behind this is that both the auxiliary classifier and the classification model have to distinguish the sample from the extracted features. Besides, another benefit of sharing layers is that the model parameters can be reduced. + +To this end, we propose sharing the front model layers between the two models which are used to extract features, as illustrated in Figure 1. Specifically, by denoting the $c$ -layer shared extractor by $\tilde{w}^k = [\tilde{w}^{k,1},\dots,\tilde{w}^{k,c}]$ where $\tilde{w}^{k,i}$ is the $i$ -th layer of the shared extractor, we can re-write the $n_d$ -layer auxiliary classifier by $\theta^k = [\tilde{w}^k,\theta^{k,c + 1},\dots,\theta^{k,n_d}]$ , and the $n$ -layer target classification model $w^k$ by $w^k = [\tilde{w}^k,w^{k,c + 1},\dots,w^{k,n}]$ with $w^{k,i}$ as the $i$ -th layer of the classification model $w^{k,i}$ . With the shared layers, the discriminator and the classification model are trained jointly: + +$$ +\min _ {\theta^ {k}, w ^ {k}} \mathcal {L} _ {\mathcal {J}} ^ {k} \left(\theta^ {k}, w ^ {k}\right) = \mathcal {L} ^ {k} \left(\theta^ {k}\right) + \mathcal {L} ^ {k} \left(w ^ {k}\right). \tag {8} +$$ + +where $\mathcal{L}^k (\theta^k)$ refers to (3) and $\mathcal{L}^k (w^k)$ is defined in (1). + +# 4 Experiments + +# 4.1 Setup + +Dataset: We conduct all our experiments with heterogeneous dataset partition on three datasets. In FDIL scenarios, new domains are introduced gradually. The dataset initially contains samples from a specific domain, and new domains are introduced at later stages, enabling models to adapt and generalize to new unseen domains. + +Table 1: Performance comparison by Top-1 test accuracy for our model and baseline methods on Digit-10 (data heterogeneity: $\alpha = 0.1$ ). + +
Digit-10 (λ = 0.5)
MethodMNIST → EMNIST → USPS → SVHNAvgΔ(↑)
FedAvg [24]92.82 →84.02 →81.02 →71.8782.436.85↑
FedProx [19]93.01 →84.68 →78.30 →76.4283.106.18↑
FedCIL [28]94.64 →87.52 →82.16 →80.9286.312.97↑
DANN [10] + FL96.07 →86.71 →79.11 →72.1483.515.77↑
Source-Only92.82 →82.15 →75.53 →69.0679.899.39↑
Sharing92.67 →82.91 →76.17 →71.9680.938.35↑
pFedDIL -w/o Migration92.82 →87.09 →84.00 →81.8986.452.83↑
pFedDIL -w/o Sharing92.79 →90.13 →87.59 →86.1989.180.11↑
pFedDIL -w/o Correlation92.82 →89.06 →84.72 →84.3087.731.55↑
pFedDIL92.82 →90.39 →87.14 →86.7589.28/
Upper Bound -Disjoint92.82 →90.33 →93.41 →87.9791.131.85↓
+ +- Digit10: Digit-10 dataset contains 10 digit categories in four domains: MNIST [17], EMNIST [3], USPS [13], SVHN [26]. Each dataset is a digit image classification dataset of 10 classes in a specific domain, such as handwriting style. +- Office31: [32] A dataset with images from three different domains: Amazon, Webcam, and DSLR. It consists of 31 object categories, with each domain having around 4,100 images. +- DomainNet: [27] A large-scale dataset with images from six different domains: Clipart, Painting, Real, Sketch, Quickdraw, and Infograph. It contains over 0.6 million images across 345 categories. + +Baseline: For a fair comparison with other key related works, we follow the same protocols proposed by [24] to set federated learning(FL) task and utilize the domain-incremental learning (DIL) task by [25]. We evaluate all baselines under the above settings, including two representative FL models FedAvg [24] and FedProx [19], which is better at tackling heterogeneity in federated networks than FedAvg, and the following several methods in FDIL scenario: + +FedCIL [28]: This approach employs the ACGAN backbone to generate synthetic samples to consolidate the global model and align sample features in the output layer. Authors conduct experiments in FCIL scenarios, and here we adopt it to the FDIL setting. + +Domain Adaptation + FL: Here we adopt the robust adversarial-based method DANN [10]. This baseline mainly follows the domain adaptation paradigm which is different from the incremental learning and is often prone to issues like catastrophic forgetting. + +Source-Only: Here we maintain the initial model trained on the initial domain data completely fixed throughout all tasks. This baseline does not perform any adaptation, modification or replacement to its parameters and serves as a lower bound. + +Sharing: Inspired by the multi-task learning scenario [31], we adopt all front layers before the last fully connected layer as shared layers, and we use relevant different + +Table 2: Performance comparison by Top-1 test accuracy for our model and baseline methods on Office-31 (data heterogeneity: $\alpha = 1$ ). + +
Office-31 (λ = 0.8)
MethodAmazon → Dlsr → WebcamAvgΔ(↑)
FedAvg [24]46.53 → 24.17 → 29.3433.3512.82↑
FedProx [19]45.19 → 25.68 → 32.2334.3711.80↑
FedCIL [28]49.38 → 39.65 → 44.0444.361.81↑
DANN [10] + FL51.97 → 35.96 → 43.0843.672.50↑
Source-Only46.53 → 20.61 → 28.3231.8214.35↑
Sharing46.11 → 24.23 → 34.2534.8611.31↑
pFedDIL -w/o Migration46.53 → 40.17 → 44.5543.752.42↑
pFedDIL -w/o Sharing46.63 → 42.90 → 47.0345.520.65↑
pFedDIL -w/o Correlation46.53 → 41.66 → 47.1945.131.04↑
pFedDIL46.53 → 43.87 → 48.1246.17/
Upper Bound -Disjoint46.53 → 50.17 → 56.2150.974.80↓
+ +fully-connected layers to get outputs of different domains. During inference, the fully-connected layer that matches the corresponding domain ID will be applied. + +Disjoint: Here we train a separate model for each domain. During inference, the model which matches the corresponding domain ID is used. This baseline requires additional space to store parameters and additional cost to obtain inference-task ID. It is a degradation of domain-incremental learning and provides an upper bound here. + +Configuration: Unless otherwise mentioned, we set the number of local training epoch $E = 20$ , communication round $T = 180$ for each domain. Here we train each task until convergence before the arrival of a new task with communication rounds $T$ . For local training, the batch size is 32, the learning rate for our models is 0.001 and the weight decay is $1e - 3$ . Like most FL methods on Non-IID data distribution [24], we use the Dirichlet distribution $(\alpha)$ to simulate the data heterogeneity from all domains and we apply the same data heterogeneity for all domains in the same dataset. For the classifier in all methods, we employ ResNet18 [12] as the basic backbone. For the multi-task learning structure in our approach, we treat all previous layers before the last fully connected layer as share layers, and we use two different fully connected layers to get outputs as the auxiliary classifier result and target classification result. The total clients' number is 20 for Digit-10 and DomainNet with an active ratio $k = 0.4$ , 10 for Office-31 with an active ratio $k = 0.4$ . + +# 4.2 Performance Overview + +Test Accuracy. Table 1,2,3 shows the performance of our pFedDIL and other baseline methods on three datasets, where " $\Delta$ " denotes the improvement of our method with other baselines. We carry out experiments against data heterogeneity on each dataset, and pFedDIL achieves the best performance in most cases by a margin of + +Table 3: Performance comparison by Top-10 test accuracy for our model and baseline methods on DomainNet (data heterogeneity: $\alpha = 10$ ) + +
DomainNet (λ = 0.8)
MethodClipart → Infograph → Painting → Quickdraw → Real → SketchAvgΔ(↑)
FedAvg [24]48.43 → 37.18 → 45.80 → 45.32 → 44.91 → 50.2545.326.38↑
FedProx [19]47.39 → 38.43 → 44.31 → 47.96 → 42.38 → 51.7745.376.32↑
FedCIL [28]52.14 → 43.68 → 47.10 → 48.75 → 42.89 → 51.2647.644.05↑
DANN [10] + FL50.07 → 39.74 → 43.73 → 45.08 → 43.28 → 52.9645.815.88↑
Source-Only48.43 → 31.01 → 33.12 → 38.15 → 37.62 → 42.8538.5313.16↑
Sharing48.43 → 36.03 → 39.58 → 41.80 → 40.41 → 44.7641.849.86↑
pFedDIL -w/o Migration48.43 → 43.83 → 50.05 → 49.44 → 46.10 → 54.9048.792.90↑
pFedDIL -w/o Sharing48.52 → 47.37 → 51.16 → 53.98 → 50.11 → 55.2451.060.63↑
pFedDIL -w/o Correlation48.43 → 46.71 → 51.09 → 52.34 → 48.07 → 56.9550.601.09↑
pFedDIL48.43 → 47.19 → 52.48 → 54.07 → 50.36 → 57.6151.69/
Upper Bound -Disjoint48.43 → 52.33 → 54.21 → 52.89 → 53.62 → 58.0653.261.57↓
+ +$1.81\% \sim 14.35\%$ in terms of average accuracy. Due to our challenging FDIL settings, there is a notable gap between all baselines and the upper performance bound. However, pFedDIL almost achieves comparable results with less storage and adaptive task-id inference during the inference stage. These results demonstrate that the appropriate incremental task learning strategy and adaptive knowledge migration technique in pFedDIL can alleviate catastrophic forgetting and thus improve performance. + +![](images/84b22f2020e652fdb2540cbd3e3a629542b40f397e4fdb82561514878995d3cc.jpg) +Fig. 2: Visualized performance w.r.t data heterogeneity for three datasets. + +Ablation Study. As shown in Tables 1,2,3, we evaluate the effects of each module in our model via ablation studies. pFedDIL -w/o Migration, pFedDIL -w/o Sharing and pFedDIL- w/o Correlation denote the performance of our model without using knowledge migration, parameters sharing and correlation factor in the inference stage. Compared with pFedDIL, the performance of pFedDIL -w/o Migration, pFedDIL -w/o Sharing and pFedDIL- w/o Correlation degrades evidently with a range of $0.11\% \sim 2.90\%$ . Specifically, the adaptive knowledge migration technique in pFedDIL plays a much + +more important role with the obvious improvement in test accuracy, and the partial parameters sharing trick condenses model parameters by sharing the feature extracted layer while not causing a remarkable impact on test accuracy. Experiment results verify the effectiveness of all modules, confirming all modules are essential to train a robust federated domain-incremental model. + +Data heterogeneity. Increasing data heterogeneity levels in experiments is significant because it allows for a more realistic evaluation of the system's performance in diverse real-world scenarios and assesses its robustness and generalization capabilities. In particular, in the FDIL scenario, incremental data is less likely to appear in the form of IID data. Therefore, it is essential to simulate different levels of Non-IID data to validate the effectiveness of our method in the FDIL scenario. Figure 2 displays the accuracy with different levels of data heterogeneity on three datasets. As shown in this figure, all methods achieve an improvement in test accuracy with the decline of data heterogeneity. Most notably, pFedDIL always maintains a leading advantage with different levels of data heterogeneity. + +Table 4: Evaluation of various methods in terms of the communication rounds. We report the sum of communication rounds required to achieve the best performance on each task and the running time of a training communication round. + +
DatasetFedAvgFedCILDANN+FLpFedDIL
RoundsTimeRoundsTimeRoundsTimeRoundsTime
Digit10480(r)32.14(s)492(r)50.06(s)553(r)46.77(s)379(r)34.91(s)
Office31463(r)44.96(s)482(r)78.32(s)488(r)62.89(s)411(r)47.58(s)
DomainNet835(r)92.22(s)904(r)165.41(s)861(r)158.01(s)725(r)105.38(s)
+ +Computational complexity and communication cost. We treat all previous layers in ResNet18 before the last fully connected layer as shared layers, and we use two different fully connected layers to get outputs as the auxiliary classifier result and target classification result. Compared with baseline methods, we only train one additional fully connected layer as the auxiliary classifier for the binary classification task - the discrimination of the sample belonging to the current task. The training time cost (each communication round) is shown in Table 4. Therefore, the training of the auxiliary classifier will not increase the overhead or computational complexity significantly and hardly affect the training efficiency. + +Table 4 shows the evaluation of various methods in terms of the communication rounds to reach the test accuracy reported in Table 1, 2 and 3 (180 communication rounds each task). Here we show the sum of communication rounds required to achieve the performance on each task. pFedDIL requires the least communication rounds to achieve the reported test accuracy on all datasets with the knowledge matching. Thus, the local model is easier to converge. + +Parameter sensitivity analysis. To figure out whether pFedDIL is sensitive to some specific parameters, we first select some general parameters, which are defined in most + +![](images/c4f1bd2a9d2a59063f1df45c42fdd4cfb47ab0298990d4bf4c67b88ded6c517d.jpg) +Fig. 3: Performance of pFedDIL under different configurations. Here, we select three general parameters in the FL setting: (a) local training epoch $E$ , (b) sample size $B$ in the classifier, (c) client selection ratio $r$ of all clients on DomainNet with $\alpha = 10$ . + +![](images/d9dcc7a26c79361d656c539b7f6c591d3239425791b64d2c5a619ca75067f725.jpg) + +![](images/f115331d93e254b7f65dcf2c12690719287aaeab05b8e4474a7546320723e626.jpg) + +FL: local training epoch $E$ ; sample batch size $B$ , and client selection ratio for global aggregation $r$ on DomainNet with $\alpha = 10$ to carry out experiments. Figure 3 shows the performance of pFedDIL under different configurations. pFedDIL achieves better results when we increase the local training epochs at the beginning. However, pFedDIL has a comparable performance when the $E$ is set to 20 and 40. The variance resulting from $E$ and $r$ being less than 0.6 and 13 percentage points respectively is comparable to that of existing works and is not significant. In addition, pFedDIL achieves similar performance with different sample sizes $B$ . + +Table 5: Performance of pFedDIL approach on all datasets with different hyper-parameter $\lambda$ values. Here we measure average accuracy and model size over all tasks on each client (each model including the auxiliary classifier costs about 22Mb storage). + +
Datasetλ = 0.0λ = 0.2λ = 0.5λ = 0.8λ = 1.0
Digit-10Avg ACC82.4385.0289.2890.7793.94
Model Size22Mb37Mb46Mb71Mb88Mb
Office-31Avg ACC33.3540.5139.9846.1750.38
Model Size22Mb42Mb48Mb53Mb66Mb
DomainNetAvg ACC45.3248.1949.6551.6953.75
Model Size22Mb73Mb84Mb91Mb132Mb
+ +Then, we conduct more research on the setting of hyper-parameter $\lambda$ . When $\lambda = 0.0$ , pFedDIL will degenerate to the normal FedAvg algorithm without any knowledge migration with the single model. When $\lambda$ is set to 1.0, pFedDIL will separately train models for each domain with knowledge migration. As shown in Table 5, we select five different $\lambda$ values to conduct experiments and evaluate the average accuracy and average model size on all personalized local models. In Table 1,2, and 3, we utilize ap + +proximately two auxiliary classifiers per client in Digit-10 and Office-31 and about four in DomainNet. For Table 1,2, we carefully select the optimal value of $\lambda$ for each dataset, striking a balance between accuracy and storage. As shown in Table 5, pFedDIL consistently outperforms other baselines even when $\lambda$ is simply set to 0.5 across all datasets (around two auxiliary classifiers per client in Digit-10 and Office-31 and approximately three in DomainNet). This suggests that a simple parameter setting can outperform existing methods, and with a more refined parameter setting, we can better accommodate practical considerations, achieving an optimal trade-off between accuracy and storage for each client. The flexibility of pFedDIL allows users to tune this parameter based on their specific requirements and constraints in real-world scenarios. + +# 5 Limitations + +Our work is aimed at solving the forgetting problem under the assumption of shared knowledge between tasks in the domain incremental setting. The problem of learning tasks without similar knowledge is rare in real-world scenarios and violates the principle of incremental learning [11]. The ultimate goal of incremental learning is to better learn new knowledge with the assistance of previous knowledge while also mastering both previous and new knowledge [39]. Furthermore, we investigate the situation where the previous data is strictly unseen to clients, while some research on federated incremental learning assumes that each client can cache some samples for replay [9]. Both situations exist and there is a must to separately develop appropriate methods. + +# 6 Conclusion and Future Work + +In this paper, we seek to tackle the catastrophic forgetting in the FDIL scenario. We propose a personalized federated domain-incremental learning approach based on adaptive knowledge matching, named pFedDIL. We leverage the auxiliary classifier to calculate the knowledge-matching intensity for the incremental task-learning strategy selection and knowledge migration. Furthermore, we propose sharing partial parameters between the target classification model and the auxiliary classifier to condense model parameters. Extensive experiments conducted on various datasets and settings show that our method achieves significant improvement in accuracy. Although existing works and our method have demonstrated great effectiveness over the FCIL and FDIL scenarios separately, none of them take both scenarios into account. To deploy the FL system in practical settings, it is necessary to simultaneously handle the class incremental tasks and domain incremental tasks. In the future, we seek to move forward in this field. + +# Acknowledgements + +This work is supported by National Natural Science Foundation of China under grants 62376103, 62302184, 62206102, Science and Technology Support Program of Hubei Province under grant 2022BAA046. + +# References + +1. Bakman, Y.F., Yaldiz, D.N., Ezzeldin, Y.H., Avestimehr, S.: Federated orthogonal training: Mitigating global catastrophic forgetting in continual federated learning. arXiv preprint arXiv:2309.01289 (2023) 1, 2 +2. Chen, H.J., Cheng, A.C., Juan, D.C., Wei, W., Sun, M.: Mitigating forgetting in online continual learning via instance-aware parameterization. Advances in Neural Information Processing Systems 33, 17466-17477 (2020) 4 +3. Cohen, G., Afshar, S., Tapson, J., Van Schaik, A.: Emmist: Extending mnist to handwritten letters. In: 2017 international joint conference on neural networks (IJCNN). pp. 2921-2926. IEEE (2017) 9 +4. Dong, J., Li, H., Cong, Y., Sun, G., Zhang, Y., Van Gool, L.: No one left behind: Real-world federated class-incremental learning. IEEE Transactions on Pattern Analysis and Machine Intelligence 46(4), 2054-2070 (2024). https://doi.org/10.1109/TPAMI.2023.33342135 +5. Dong, J., Liang, W., Cong, Y., Sun, G.: Heterogeneous forgetting compensation for class incremental learning. In: Proceedings of the IEEE/CVF International Conference on Computer Vision (ICCV). pp. 11742-11751 (October 2023) 5 +6. Dong, J., Wang, L., Fang, Z., Sun, G., Xu, S., Wang, X., Zhu, Q.: Federated class-incremental learning. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 10164-10173 (2022) 2 +7. Dong, J., Zhang, D., Cong, Y., Cong, W., Ding, H., Dai, D.: Federated incremental semantic segmentation. In: 2023 IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR). pp. 3934-3943 (2023). https://doi.org/10.1109/CVPR52729.2023.003835 +8. Dong, X., Yu, Z., Cao, W., Shi, Y., Ma, Q.: A survey on ensemble learning. Frontiers of Computer Science 14, 241-258 (2020) 8 +9. Fini, E., Lathuiliere, S., Sangineto, E., Nabi, M., Ricci, E.: Online continual learning under extreme memory constraints. In: Computer Vision-ECCV 2020: 16th European Conference, Glasgow, UK, August 23-28, 2020, Proceedings, Part XXVIII 16. pp. 720-735. Springer (2020) 14 +0. Ganin, Y., Ustinova, E., Ajakan, H., Germain, P., Larochelle, H., Laviolette, F., Marchand, M., Lempitsky, V.: Domain-adversarial training of neural networks. The journal of machine learning research 17(1), 2096-2030 (2016) 9, 10, 11 +1. Hadsell, R., Rao, D., Rusu, A.A., Pascanu, R.: Embracing change: Continual learning in deep neural networks. Trends in cognitive sciences 24(12), 1028-1040 (2020) 14 +2. He, K., Zhang, X., Ren, S., Sun, J.: Deep residual learning for image recognition. In: Proceedings of the IEEE conference on computer vision and pattern recognition. pp. 770-778 (2016) 10 +3. Hull, J.J.: A database for handwritten text recognition research. IEEE Transactions on pattern analysis and machine intelligence 16(5), 550-554 (1994) 9 +4. Jeong, E., Oh, S., Kim, H., Park, J., Bennis, M., Kim, S.L.: Communication-efficient on-device machine learning: Federated distillation and augmentation under non-id private data. ArXiv abs/1811.11479 (2018) 4 +5. Jiang, Z., Ren, Y., Lei, M., Zhao, Z.: Fedspeech: Federated text-to-speech with continual learning. In: Proceedings of the Thirtieth International Joint Conference on Artificial Intelligence. IJCAI-2021, International Joint Conferences on Artificial Intelligence Organization (Aug 2021). https://doi.org/10.24963/ijcai.2021/527, http://dx.doi.org/10.24963/ijcai.2021/527 5 + +16. Jung, S., Ahn, H., Cha, S., Moon, T.: Continual learning with node-importance based adaptive group sparse regularization. arXiv: Learning (2020) 4 +17. LeCun, Y., Cortes, C., Burges, C.: Mnist handwritten digit database (2010) 9 +18. Li, L., Fan, Y., Tse, M., Lin, K.Y.: A review of applications in federated learning. Computers & Industrial Engineering 149, 106854 (2020) 4 +19. Li, T., Sahu, A.K., Talwalkar, A., Smith, V.: Federated learning: Challenges, methods, and future directions. IEEE signal processing magazine 37(3), 50-60 (2020) 4, 9, 10, 11 +20. Li, Y., Li, Q., Wang, H., Li, R., Zhong, W., Zhang, G.: Towards efficient replay in federated incremental learning. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR). pp. 12820-12829 (June 2024) 2 +21. Liu, Y., Tian, X., Li, Y., Xiong, Z., Wu, F.: Compact feature learning for multi-domain image classification. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 7193-7201 (2019) 2 +22. Ma, Y., Xie, Z., Wang, J., Chen, K., Shou, L.: Continual federated learning based on knowledge distillation. In: Raedt, L.D. (ed.) Proceedings of the Thirty-First International Joint Conference on Artificial Intelligence, IJCAI-22. pp. 2182-2188. International Joint Conferences on Artificial Intelligence Organization (7 2022). https://doi.org/10.24963/ijcai.2022/303, https://doi.org/10.24963/ijcai.2022/303, main Track 2 +23. Maltoni, D., Lomonaco, V.: Continuous learning in single-incremental-task scenarios. Neural networks : the official journal of the International Neural Network Society 116, 56-73 (2018) 4 +24. McMahan, B., Moore, E., Ramage, D., Hampson, S., y Arcas, B.A.: Communication-efficient learning of deep networks from decentralized data. In: Artificial intelligence and statistics. pp. 1273-1282. PMLR (2017) 1, 4, 9, 10, 11 +25. Mirza, M.J., Masana, M., Possegger, H., Bischof, H.: An efficient domain-incremental learning approach to drive in all weather conditions. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 3001-3011 (2022) 4, 9 +26. Netzer, Y., Wang, T., Coates, A., Bissacco, A., Wu, B., Ng, A.Y.: Reading digits in natural images with unsupervised feature learning (2011) 9 +27. Peng, X., Bai, Q., Xia, X., Huang, Z., Saenko, K., Wang, B.: Moment matching for multi-source domain adaptation. In: Proceedings of the IEEE/CVF international conference on computer vision. pp. 1406-1415 (2019) 9 +28. Qi, D., Zhao, H., Li, S.: Better generative replay for continual federated learning. arXiv preprint arXiv:2302.13001 (2023) 2, 9, 10, 11 +29. Rebuffi, S.A., Kolesnikov, A., Sperl, G., Lampert, C.H.: Incremental classifier and representation learning. In: Proceedings of the IEEE conference on Computer Vision and Pattern Recognition. pp. 2001-2010 (2017) 4 +30. Rolnick, D., Ahuja, A., Schwarz, J., Lillicrap, T., Wayne, G.: Experience replay for continual learning. Advances in Neural Information Processing Systems 32 (2019) 4 +31. Ruder, S.: An overview of multi-task learning in deep neural networks. arXiv preprint arXiv:1706.05098 (2017) 9 +32. Saenko, K., Kulis, B., Fritz, M., Darrell, T.: Adapting visual category models to new domains. In: Computer Vision-ECCV 2010: 11th European Conference on Computer Vision, Heraklion, Crete, Greece, September 5-11, 2010, Proceedings, Part IV 11. pp. 213-226. Springer (2010) 9 +33. Shen, T., Zhang, J., Jia, X., Zhang, F., Huang, G., Zhou, P., Wu, F., Wu, C.: Federated mutual learning. ArXiv abs/2006.16765 (2020) 4 +34. Standley, T., Zamir, A., Chen, D., Guibas, L., Malik, J., Savarese, S.: Which tasks should be learned together in multi-task learning? In: International Conference on Machine Learning. pp. 9120-9132. PMLR (2020) 8 + +35. Tankard, C.: What the gdpr means for businesses. Network Security 2016(6), 5-8 (2016) 1 +36. Van de Ven, G.M., Tuytelaars, T., Tolias, A.S.: Three types of incremental learning. Nature Machine Intelligence 4(12), 1185-1197 (2022) 1 +37. Wang, H., Li, Y., Xu, W., Li, R., Zhan, Y., Zeng, Z.: Dafkd: Domain-aware federated knowledge distillation. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 20412-20421 (2023) 4 +38. Wang, H., Xu, H., Li, Y., Xu, Y., Li, R., Zhang, T.: Fedcda: Federated learning with cross-rounds divergence-aware aggregation. In: The Twelfth International Conference on Learning Representations (2023) 4 +39. Wang, L., Zhang, X., Su, H., Zhu, J.: A comprehensive survey of continual learning: theory, method and application. IEEE Transactions on Pattern Analysis and Machine Intelligence (2024) 4, 14 +40. Yoon, J., Jeong, W., Lee, G., Yang, E., Hwang, S.J.: Federated continual learning with weighted inter-client transfer. In: International Conference on Machine Learning. pp. 12073-12086. PMLR (2021) 1, 2, 5 +41. Yu, H., Yang, X., Gao, X., Kang, Y., Wang, H., Zhang, J., Li, T.: Personalized federated continual learning via multi-granularity prompt. arXiv preprint arXiv:2407.00113 (2024) 5 +42. Zhang, C., Xie, Y., Bai, H., Yu, B., Li, W., Gao, Y.: A survey on federated learning. Knowledge-Based Systems 216, 106775 (2021) 1 +43. Zhang, J., Chen, C., Zhuang, W., Lyu, L.: Target: Federated class-continual learning via exemplar-free distillation. In: Proceedings of the IEEE/CVF International Conference on Computer Vision. pp. 4782-4793 (2023) 1 +44. Zhao, H., Zhou, T., Long, G., Jiang, J., Zhang, C.: Does continual learning equally forget all parameters? In: International Conference on Machine Learning. pp. 42280-42303. PMLR (2023) 4 \ No newline at end of file diff --git a/2024/Personalized Federated Domain-Incremental Learning based on Adaptive Knowledge Matching/images.zip b/2024/Personalized Federated Domain-Incremental Learning based on Adaptive Knowledge Matching/images.zip new file mode 100644 index 0000000000000000000000000000000000000000..055b8d9e200a4508de54a89c6e47b6fcf98ece37 --- /dev/null +++ b/2024/Personalized Federated Domain-Incremental Learning based on Adaptive Knowledge Matching/images.zip @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:9eb0aafdcfed7ef84901afd321ee1f6a3a1757303a9ca3d24b7817d7230e9797 +size 511011 diff --git a/2024/Personalized Federated Domain-Incremental Learning based on Adaptive Knowledge Matching/layout.json b/2024/Personalized Federated Domain-Incremental Learning based on Adaptive Knowledge Matching/layout.json new file mode 100644 index 0000000000000000000000000000000000000000..5ed93f0ad1a94bebe27d0e7e45bf1f0138c387e6 --- /dev/null +++ b/2024/Personalized Federated Domain-Incremental Learning based on Adaptive Knowledge Matching/layout.json @@ -0,0 +1,9358 @@ +{ + "pdf_info": [ + { + "para_blocks": [ + { + "bbox": [ + 138, + 111, + 477, + 148 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 111, + 477, + 148 + ], + "spans": [ + { + "bbox": [ + 138, + 111, + 477, + 148 + ], + "type": "text", + "content": "Personalized Federated Domain-Incremental Learning based on Adaptive Knowledge Matching" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 204, + 168, + 409, + 194 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 204, + 168, + 409, + 194 + ], + "spans": [ + { + "bbox": [ + 204, + 168, + 409, + 194 + ], + "type": "text", + "content": "Yichen Li" + }, + { + "bbox": [ + 204, + 168, + 409, + 194 + ], + "type": "inline_equation", + "content": "^{1}" + }, + { + "bbox": [ + 204, + 168, + 409, + 194 + ], + "type": "text", + "content": ", Wenchao Xu" + }, + { + "bbox": [ + 204, + 168, + 409, + 194 + ], + "type": "inline_equation", + "content": "^{2}" + }, + { + "bbox": [ + 204, + 168, + 409, + 194 + ], + "type": "text", + "content": ", Haozhao Wang" + }, + { + "bbox": [ + 204, + 168, + 409, + 194 + ], + "type": "inline_equation", + "content": "^{1}" + }, + { + "bbox": [ + 204, + 168, + 409, + 194 + ], + "type": "text", + "content": "*, Yining Qi" + }, + { + "bbox": [ + 204, + 168, + 409, + 194 + ], + "type": "inline_equation", + "content": "^{1}" + }, + { + "bbox": [ + 204, + 168, + 409, + 194 + ], + "type": "text", + "content": ", Jingcai Guo" + }, + { + "bbox": [ + 204, + 168, + 409, + 194 + ], + "type": "inline_equation", + "content": "^{2}" + }, + { + "bbox": [ + 204, + 168, + 409, + 194 + ], + "type": "text", + "content": ", and Ruixuan Li" + }, + { + "bbox": [ + 204, + 168, + 409, + 194 + ], + "type": "inline_equation", + "content": "^{1}" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 149, + 201, + 463, + 224 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 149, + 201, + 463, + 224 + ], + "spans": [ + { + "bbox": [ + 149, + 201, + 463, + 224 + ], + "type": "text", + "content": "1 School of Computer Science and Technology, Huazhong University of Science and Technology, Wuhan, China" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 196, + 225, + 418, + 235 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 196, + 225, + 418, + 235 + ], + "spans": [ + { + "bbox": [ + 196, + 225, + 418, + 235 + ], + "type": "text", + "content": "{ycli0204,hz_wang,qiyining,rxli}@hust.edu.cn" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 146, + 235, + 466, + 258 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 146, + 235, + 466, + 258 + ], + "spans": [ + { + "bbox": [ + 146, + 235, + 466, + 258 + ], + "type": "inline_equation", + "content": "^{2}" + }, + { + "bbox": [ + 146, + 235, + 466, + 258 + ], + "type": "text", + "content": " Department of Computing, The Hong Kong Polytechnic University, Hongkong, China {wenchao.xu, jc-jingcai.guo} @polyu.edu.hk" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 159, + 289, + 455, + 467 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 159, + 289, + 455, + 467 + ], + "spans": [ + { + "bbox": [ + 159, + 289, + 455, + 467 + ], + "type": "text", + "content": "Abstract. This paper focuses on Federated Domain-Incremental Learning (FDIL) where each client continues to learn incremental tasks where their domain shifts from each other. We propose a novel adaptive knowledge matching-based personalized FDIL approach (pFedDIL) which allows each client to alternatively utilize appropriate incremental task learning strategy on the correlation with the knowledge from previous tasks. More specifically, when a new task arrives, each client first calculates its local correlations with previous tasks. Then, the client can choose to adopt a new initial model or a previous model with similar knowledge to train the new task and simultaneously migrate knowledge from previous tasks based on these correlations. Furthermore, to identify the correlations between the new task and previous tasks for each client, we separately employ an auxiliary classifier to each target classification model and propose sharing partial parameters between the target classification model and the auxiliary classifier to condense model parameters. We conduct extensive experiments on several datasets of which results demonstrate that pFedDIL outperforms state-of-the-art methods by up to " + }, + { + "bbox": [ + 159, + 289, + 455, + 467 + ], + "type": "inline_equation", + "content": "14.35\\%" + }, + { + "bbox": [ + 159, + 289, + 455, + 467 + ], + "type": "text", + "content": " in terms of average accuracy of all tasks." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 132, + 487, + 218, + 499 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 487, + 218, + 499 + ], + "spans": [ + { + "bbox": [ + 132, + 487, + 218, + 499 + ], + "type": "text", + "content": "1 Introduction" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 130, + 514, + 482, + 646 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 514, + 482, + 646 + ], + "spans": [ + { + "bbox": [ + 130, + 514, + 482, + 646 + ], + "type": "text", + "content": "Federated Learning (FL) has become a basic paradigm that facilitates collaborative training of deep neural networks (DNNs) from distributed clients while preserving their data privacy [24, 42]. To deploy FL in real-world applications, it is practical that each client can continuously learn new incremental tasks while retaining the accuracy for the previous tasks [40]. However, history data may often be out of reach with privacy constraints like GDPR and many studies have shown that DNNs will forget the knowledge of previous tasks when learning new tasks [35, 36]. Such a phenomenon is known as 'Catastrophic Forgetting', which is a major challenge of incremental learning in the centralized setting. Especially, this challenge is even more formidable to solve in the FL scenario, where the limited previous data are stored in distributed clients and cannot be accessed by each other [1, 43]." + } + ] + } + ], + "index": 7 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 134, + 654, + 395, + 665 + ], + "type": "page_footnote", + "angle": 0, + "lines": [ + { + "bbox": [ + 134, + 654, + 395, + 665 + ], + "spans": [ + { + "bbox": [ + 134, + 654, + 395, + 665 + ], + "type": "text", + "content": "* Haozhao Wang, Yining Qi and Ruixuan Li are corresponding authors." + } + ] + } + ], + "index": 8 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 0 + }, + { + "para_blocks": [ + { + "bbox": [ + 130, + 116, + 482, + 248 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 116, + 482, + 248 + ], + "spans": [ + { + "bbox": [ + 130, + 116, + 482, + 248 + ], + "type": "text", + "content": "Many recent studies have been proposed to solve the catastrophic forgetting challenge in federated incremental learning. For instance, [40] aims to learn a personalized model for each client by decomposing the model parameters into shared parameters and adaptive parameters and transferring common knowledge between similar tasks across clients. FCIL [6] addresses the federated class-incremental learning and trains a global model by computing additional class-imbalance losses. Authors in [22] utilize extra data at both the server and client sides to conduct knowledge distillation to alleviate forgetting. [28] optimize the generative network and reconstruct previous samples for replay. Recently, Re-Fed is proposed in [20] to prevent catastrophic forgetting with synergistic replay by caching important samples which contribute to both the local and global understanding." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 130, + 251, + 483, + 395 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 251, + 483, + 395 + ], + "spans": [ + { + "bbox": [ + 130, + 251, + 483, + 395 + ], + "type": "text", + "content": "While these approaches have achieved great success, they are generally designed for the scenario where the new tasks are incremental classes, i.e., Federated Class-Incremental Learning (FCIL). These methods may lose their effectiveness when the new tasks have the same classes but their domain shift with different feature distributions, i.e., Federated Domain-Incremental Learning (FDIL). To elaborate further, the FCIL [6] method relies on changes in the new data category label in incremental tasks to alleviate forgetting of previous tasks; FOT [1] projects different incremental tasks into different orthogonal subspaces, which means feature isolation for different tasks, but if suffers from the intersected feature space boundaries of incremental tasks. The primary challenge in implementing FCIL methods in the context of FDIL is that FCIL methods typically depend on the connection between previous and new classes, which is not applicable in FDIL scenarios where the classes remain consistent." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 130, + 399, + 483, + 506 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 399, + 483, + 506 + ], + "spans": [ + { + "bbox": [ + 130, + 399, + 483, + 506 + ], + "type": "text", + "content": "In this paper, we seek to tackle the challenge of catastrophic forgetting in the FDIL scenario. A simple way to address this problem is to train different models for each incremental task. However, this requires extensive storage and time-consuming re-training. Additionally, all these models need to be saved for future use which is laborious because we need to manually select the appropriate mode for each inference data. Instead of training a separate model for each different task, a better solution is to ideally have only a single model that can work in all incremental tasks. However, naively training a model for all tasks can lead to terrible catastrophic forgetting when the unique knowledge in the new task overlaps the ones in the previous tasks." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 130, + 510, + 482, + 666 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 510, + 482, + 666 + ], + "spans": [ + { + "bbox": [ + 130, + 510, + 482, + 666 + ], + "type": "text", + "content": "Unlike in the FCIL scenario, where incremental tasks involve data from different categories, In the FDIL scenario, the incremental tasks involve data from different domains, i.e., handwritten digit data and cartoon digit data. Compared to the data from different categories in FCIL, i.e., cat data and dog data, there is more shared feature information between the different incremental tasks, which can facilitate the overall FDIL task [21]. The key to efficiently learning new tasks while memorizing previous tasks in the FDIL scenario is to identify the correlation between incremental tasks and then share the useful common knowledge in different tasks. Like [40] in FCIL, we consider each client maintains their personalized models while the number of personalized models is under a given threshold due to the limited capacity of each client. Given the correlation between tasks, each client can maintain several personalized models and share the same model between similar tasks. In this way, the client can choose a model with tasks most similar to the new task, enabling the acquisition of new task knowledge while retain-" + } + ] + } + ], + "index": 5 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "spans": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "type": "text", + "content": "2" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 166, + 91, + 225, + 101 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 166, + 91, + 225, + 101 + ], + "spans": [ + { + "bbox": [ + 166, + 91, + 225, + 101 + ], + "type": "text", + "content": "Yichen Li, et al." + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 1 + }, + { + "para_blocks": [ + { + "bbox": [ + 130, + 116, + 482, + 163 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 116, + 482, + 163 + ], + "spans": [ + { + "bbox": [ + 130, + 116, + 482, + 163 + ], + "type": "text", + "content": "ing existing task knowledge. This dual benefit permits the new task to also capitalize on previous knowledge, ultimately enhancing performance. Nonetheless, considering the tasks come incrementally, identifying the inter-task correlations and subsequently training the selected model is challenging." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 130, + 164, + 482, + 366 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 164, + 482, + 366 + ], + "spans": [ + { + "bbox": [ + 130, + 164, + 482, + 366 + ], + "type": "text", + "content": "To tackle these limitations, we in this paper propose a novel Personalized Federated Domain-Incremental Learning (pFedDIL) approach which adaptively matches and migrates knowledge among different incremental tasks. More specifically, we propose that each client employs a small auxiliary classifier for each personalized model which is trained to distinguish itself from other tasks. Using the auxiliary classifiers of previous tasks, each client can adaptively identify the correlation of the new task to all previous tasks. Based on these correlations, each client can match the new task with a suitable model, i.e., a new initialized model or a previous model but with similar knowledge. To maximize the training efficiency of the matched model over the new task, we further allow each client to migrate knowledge from other not matched but similar personalized models in a weighted manner, where the weights are the correlations obtained previously. Furthermore, to reduce the size of auxiliary classifiers while enhancing their performance, we propose sharing partial parameters between the target personalized model and the auxiliary classifier. Finally, considering the presence of multiple personalized local models within each client that could lead to additional inference costs, we apply a weighted ensemble distillation over these models where the weights are also obtained from the output of the auxiliary classifiers." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 130, + 368, + 482, + 415 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 368, + 482, + 415 + ], + "spans": [ + { + "bbox": [ + 130, + 368, + 482, + 415 + ], + "type": "text", + "content": "Through extensive experiments on several datasets and different settings, we show that the proposed pFedDIL significantly improves the model accuracy as compared to state-of-the-art approaches. The major contributions of this paper are summarized as follows:" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 138, + 422, + 479, + 597 + ], + "type": "list", + "angle": 0, + "index": 9, + "blocks": [ + { + "bbox": [ + 138, + 422, + 479, + 468 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 422, + 479, + 468 + ], + "spans": [ + { + "bbox": [ + 138, + 422, + 479, + 468 + ], + "type": "text", + "content": "- We are the first to formulate the problem of federated domain-incremental learning. To address this issue, we propose a novel method named pFedDIL which enables each client to match and migrate knowledge from previous tasks to alleviate catastrophic forgetting." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 138, + 469, + 479, + 502 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 469, + 479, + 502 + ], + "spans": [ + { + "bbox": [ + 138, + 469, + 479, + 502 + ], + "type": "text", + "content": "- To adaptively employ incremental task learning strategy and migrate knowledge, we define the knowledge matching intensity by the auxiliary classifier to discern the correlation between the new task and previous tasks." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 138, + 505, + 479, + 562 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 505, + 479, + 562 + ], + "spans": [ + { + "bbox": [ + 138, + 505, + 479, + 562 + ], + "type": "text", + "content": "- To facilitate the adaptive selection of inference model among a set of personalized local models, we coordinate all personalized local models to predict weighted by normalized output by auxiliary classifiers. Besides, we further propose sharing partial parameters between the auxiliary classifier and the target classification model to condense model parameters." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 138, + 564, + 479, + 597 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 564, + 479, + 597 + ], + "spans": [ + { + "bbox": [ + 138, + 564, + 479, + 597 + ], + "type": "text", + "content": "- We carry out extensive experiments on various datasets and different settings. Experimental results illustrate that our proposed model outperforms the state-of-the-art methods by up to " + }, + { + "bbox": [ + 138, + 564, + 479, + 597 + ], + "type": "inline_equation", + "content": "14.35\\%" + }, + { + "bbox": [ + 138, + 564, + 479, + 597 + ], + "type": "text", + "content": " in terms of average accuracy on all tasks." + } + ] + } + ], + "index": 8 + } + ], + "sub_type": "text" + }, + { + "bbox": [ + 132, + 617, + 225, + 629 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 617, + 225, + 629 + ], + "spans": [ + { + "bbox": [ + 132, + 617, + 225, + 629 + ], + "type": "text", + "content": "2 Related Work" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 130, + 641, + 482, + 666 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 641, + 482, + 666 + ], + "spans": [ + { + "bbox": [ + 130, + 641, + 482, + 666 + ], + "type": "text", + "content": "Federated Learning. Federated learning (FL) is a distributed learning approach that trains a global model by merging multiple locally trained models on private datasets" + } + ] + } + ], + "index": 11 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 298, + 91, + 447, + 102 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 298, + 91, + 447, + 102 + ], + "spans": [ + { + "bbox": [ + 298, + 91, + 447, + 102 + ], + "type": "text", + "content": "Federated Domain-Incremental Learning" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 474, + 92, + 481, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 474, + 92, + 481, + 100 + ], + "spans": [ + { + "bbox": [ + 474, + 92, + 481, + 100 + ], + "type": "text", + "content": "3" + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 2 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 177, + 114, + 436, + 335 + ], + "blocks": [ + { + "bbox": [ + 177, + 114, + 436, + 335 + ], + "lines": [ + { + "bbox": [ + 177, + 114, + 436, + 335 + ], + "spans": [ + { + "bbox": [ + 177, + 114, + 436, + 335 + ], + "type": "image", + "image_path": "7fa8509dae5f4dafbc2482281b406e213d7f24cde0991e0473be9f773dae0213.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 130, + 344, + 482, + 388 + ], + "lines": [ + { + "bbox": [ + 130, + 344, + 482, + 388 + ], + "spans": [ + { + "bbox": [ + 130, + 344, + 482, + 388 + ], + "type": "text", + "content": "Fig. 1: Illustration of the pFedDIL framework. In the training stage, each client randomly receives one domain dataset without replacement to form a task, where different colors represent different domains. Clients can adaptively employ an appropriate incremental task learning strategy and migrate the knowledge from previous tasks with the knowledge matching intensity." + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_caption" + } + ], + "index": 2 + }, + { + "bbox": [ + 130, + 413, + 482, + 509 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 413, + 482, + 509 + ], + "spans": [ + { + "bbox": [ + 130, + 413, + 482, + 509 + ], + "type": "text", + "content": "[18]. A significant challenge in FL is data heterogeneity, where data distribution among clients is non-identically and independently distributed (Non-IID), negatively impacting model performance [14, 33, 38]. FedAvg is a notable FL method that refines the global model by incorporating parameters trained on local private data [24]. [19] propose a proximal term to align closely with the global model. The federated distillation method [37] suggests distilling multiple local models into a global one by only aggregating soft output predictions. However, existing FL approaches typically assume that the data in each client is static, which cannot account for incremental data with domain shift." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 130, + 510, + 482, + 617 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 510, + 482, + 617 + ], + "spans": [ + { + "bbox": [ + 130, + 510, + 482, + 617 + ], + "type": "text", + "content": "Incremental Learning. Incremental learning (IL) is a learning paradigm that empowers a model to learn sequentially from a series of tasks while preserving the knowledge acquired from previous tasks [39, 44]. IL can be segmented into three primary variants: task-incremental learning [23], class-incremental learning [29] and domain-incremental learning [25]. Existing strategies can further be grouped into three categories: replay-based methods [30], regularization-based methods [16] and parameter isolation methods [2]. Most IL methods are designed for task-IL or class-IL and are typically server-based. This work focuses on the federated domain-incremental learning scenario and proposes the pFedDIL technique." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 130, + 617, + 482, + 666 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 617, + 482, + 666 + ], + "spans": [ + { + "bbox": [ + 130, + 617, + 482, + 666 + ], + "type": "text", + "content": "Federated Incremental Learning. Federated Incremental Learning (FIL), an approach that targets the dynamic adaptation of streaming tasks within each client's environment, emphasizes the seamless integration of new data into the global model while preserving the acquired knowledge from previous data streams. Despite its crucial role, FIL" + } + ] + } + ], + "index": 6 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "spans": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "type": "text", + "content": "4" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 166, + 91, + 226, + 101 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 166, + 91, + 226, + 101 + ], + "spans": [ + { + "bbox": [ + 166, + 91, + 226, + 101 + ], + "type": "text", + "content": "Yichen Li, et al." + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 3 + }, + { + "para_blocks": [ + { + "bbox": [ + 131, + 116, + 482, + 224 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 131, + 116, + 482, + 224 + ], + "spans": [ + { + "bbox": [ + 131, + 116, + 482, + 224 + ], + "type": "text", + "content": "has garnered attention relatively recently, with [40] pioneering the field by focusing on Task-Incremental Learning (Task-IL) in FL. In parallel, studies such as [41] takes the multi-granularity expression of knowledge into account, promoting the spatial-temporal integration of knowledge in FCIL. Conversely, [4,5] simplify the problem's complexity by presupposing clients possess sufficient storage capacity to retain historical examples and engage in data sharing, a scenario that deviates from the conventional federated learning paradigm. Beyond image classification, investigations like [7, 15] extend the horizons of FIL into diverse domains. Our research focuses on the domain shift between different tasks and the transfer of the common knowledge." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 132, + 242, + 220, + 255 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 242, + 220, + 255 + ], + "spans": [ + { + "bbox": [ + 132, + 242, + 220, + 255 + ], + "type": "text", + "content": "3 Methodology" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 131, + 267, + 480, + 291 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 131, + 267, + 480, + 291 + ], + "spans": [ + { + "bbox": [ + 131, + 267, + 480, + 291 + ], + "type": "text", + "content": "We first formulate our problem definition and then present our pFedDIL approach. Figure 1 gives an overview of our method." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 132, + 308, + 238, + 319 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 308, + 238, + 319 + ], + "spans": [ + { + "bbox": [ + 132, + 308, + 238, + 319 + ], + "type": "text", + "content": "3.1 Problem Definition" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 131, + 328, + 482, + 415 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 131, + 328, + 482, + 415 + ], + "spans": [ + { + "bbox": [ + 131, + 328, + 482, + 415 + ], + "type": "text", + "content": "In incremental learning, each model learns from a sequence of streaming tasks " + }, + { + "bbox": [ + 131, + 328, + 482, + 415 + ], + "type": "inline_equation", + "content": "\\{\\mathcal{T}_{(1)},\\mathcal{T}_{(2)},\\ldots \\mathcal{T}_{(n)}\\}" + }, + { + "bbox": [ + 131, + 328, + 482, + 415 + ], + "type": "text", + "content": " where " + }, + { + "bbox": [ + 131, + 328, + 482, + 415 + ], + "type": "inline_equation", + "content": "\\mathcal{T}_{(t)}" + }, + { + "bbox": [ + 131, + 328, + 482, + 415 + ], + "type": "text", + "content": " denotes the dataset of the " + }, + { + "bbox": [ + 131, + 328, + 482, + 415 + ], + "type": "inline_equation", + "content": "t" + }, + { + "bbox": [ + 131, + 328, + 482, + 415 + ], + "type": "text", + "content": "-th task; " + }, + { + "bbox": [ + 131, + 328, + 482, + 415 + ], + "type": "inline_equation", + "content": "\\mathcal{T}_{(t)} = \\{x_i^t,y_i^t\\}_{i = 1}^{N^t}" + }, + { + "bbox": [ + 131, + 328, + 482, + 415 + ], + "type": "text", + "content": ", which has " + }, + { + "bbox": [ + 131, + 328, + 482, + 415 + ], + "type": "inline_equation", + "content": "N^t" + }, + { + "bbox": [ + 131, + 328, + 482, + 415 + ], + "type": "text", + "content": " pairs of sample data " + }, + { + "bbox": [ + 131, + 328, + 482, + 415 + ], + "type": "inline_equation", + "content": "x_{i}^{t}\\in X^{t}" + }, + { + "bbox": [ + 131, + 328, + 482, + 415 + ], + "type": "text", + "content": " and corresponding label " + }, + { + "bbox": [ + 131, + 328, + 482, + 415 + ], + "type": "inline_equation", + "content": "y_{i}^{t}\\in Y^{t}" + }, + { + "bbox": [ + 131, + 328, + 482, + 415 + ], + "type": "text", + "content": ". " + }, + { + "bbox": [ + 131, + 328, + 482, + 415 + ], + "type": "inline_equation", + "content": "X^{t}" + }, + { + "bbox": [ + 131, + 328, + 482, + 415 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 131, + 328, + 482, + 415 + ], + "type": "inline_equation", + "content": "Y^{t}" + }, + { + "bbox": [ + 131, + 328, + 482, + 415 + ], + "type": "text", + "content": " represent the domain space and label space for the " + }, + { + "bbox": [ + 131, + 328, + 482, + 415 + ], + "type": "inline_equation", + "content": "t" + }, + { + "bbox": [ + 131, + 328, + 482, + 415 + ], + "type": "text", + "content": "-th task. In this paper, we focus on the variant of domain incremental learning, where all tasks share the same numbers of classes i.e., " + }, + { + "bbox": [ + 131, + 328, + 482, + 415 + ], + "type": "inline_equation", + "content": "\\mathcal{V}^1 = \\mathcal{V}^t,\\forall t\\in [1,n]" + }, + { + "bbox": [ + 131, + 328, + 482, + 415 + ], + "type": "text", + "content": ". As the sequence of tasks arrives, the client needs to learn the new task while their domain and data distribution changes, i.e., " + }, + { + "bbox": [ + 131, + 328, + 482, + 415 + ], + "type": "inline_equation", + "content": "\\mathcal{X}^1\\neq \\mathcal{X}^t,\\forall t\\in [1,n]" + }, + { + "bbox": [ + 131, + 328, + 482, + 415 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 131, + 415, + 482, + 515 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 131, + 415, + 482, + 515 + ], + "spans": [ + { + "bbox": [ + 131, + 415, + 482, + 515 + ], + "type": "text", + "content": "We consider the federated domain-incremental learning scenario. We aim to train a global model for " + }, + { + "bbox": [ + 131, + 415, + 482, + 515 + ], + "type": "inline_equation", + "content": "K" + }, + { + "bbox": [ + 131, + 415, + 482, + 515 + ], + "type": "text", + "content": " total clients in the federated domain-incremental learning setting. We consider the local client " + }, + { + "bbox": [ + 131, + 415, + 482, + 515 + ], + "type": "inline_equation", + "content": "k" + }, + { + "bbox": [ + 131, + 415, + 482, + 515 + ], + "type": "text", + "content": " can only access the local private streaming tasks " + }, + { + "bbox": [ + 131, + 415, + 482, + 515 + ], + "type": "inline_equation", + "content": "\\{\\mathcal{T}_{(1)}^k,\\mathcal{T}_{(2)}^k,\\dots ,\\mathcal{T}_{(n)}^k\\}" + }, + { + "bbox": [ + 131, + 415, + 482, + 515 + ], + "type": "text", + "content": ". Now our goal is to train a set of personalized local models " + }, + { + "bbox": [ + 131, + 415, + 482, + 515 + ], + "type": "inline_equation", + "content": "\\widetilde{w}^{k} = [w_{1}^{k},w_{2}^{k},\\ldots ,w_{d}^{k}]" + }, + { + "bbox": [ + 131, + 415, + 482, + 515 + ], + "type": "text", + "content": " for each client " + }, + { + "bbox": [ + 131, + 415, + 482, + 515 + ], + "type": "inline_equation", + "content": "k" + }, + { + "bbox": [ + 131, + 415, + 482, + 515 + ], + "type": "text", + "content": " over all existing " + }, + { + "bbox": [ + 131, + 415, + 482, + 515 + ], + "type": "inline_equation", + "content": "t" + }, + { + "bbox": [ + 131, + 415, + 482, + 515 + ], + "type": "text", + "content": " tasks " + }, + { + "bbox": [ + 131, + 415, + 482, + 515 + ], + "type": "inline_equation", + "content": "\\mathcal{T} = \\sum_{n = 1}^{t}\\sum_{k = 1}^{K}\\mathcal{T}_{(n)}^{k}" + }, + { + "bbox": [ + 131, + 415, + 482, + 515 + ], + "type": "text", + "content": ". Here each personalized local model corresponds to at least one task, and possibly multiple tasks (i.e. client " + }, + { + "bbox": [ + 131, + 415, + 482, + 515 + ], + "type": "inline_equation", + "content": "k" + }, + { + "bbox": [ + 131, + 415, + 482, + 515 + ], + "type": "text", + "content": " may train the task " + }, + { + "bbox": [ + 131, + 415, + 482, + 515 + ], + "type": "inline_equation", + "content": "\\mathcal{T}_{(1)}^k" + }, + { + "bbox": [ + 131, + 415, + 482, + 515 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 131, + 415, + 482, + 515 + ], + "type": "inline_equation", + "content": "\\mathcal{T}_{(2)}^k" + }, + { + "bbox": [ + 131, + 415, + 482, + 515 + ], + "type": "text", + "content": " with the model " + }, + { + "bbox": [ + 131, + 415, + 482, + 515 + ], + "type": "inline_equation", + "content": "w_{1}^{k}" + }, + { + "bbox": [ + 131, + 415, + 482, + 515 + ], + "type": "text", + "content": "). Then, we can formulate the goal as:" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 227, + 523, + 481, + 555 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 227, + 523, + 481, + 555 + ], + "spans": [ + { + "bbox": [ + 227, + 523, + 481, + 555 + ], + "type": "interline_equation", + "content": "\\underset {\\widetilde {w}} {\\arg \\min } \\mathcal {L} (\\widetilde {w}) = \\sum_ {n = 1} ^ {t} \\sum_ {k = 1} ^ {K} \\frac {1}{| \\mathcal {T} |} \\mathcal {L} _ {(n)} ^ {k} (\\widetilde {w}). \\tag {1}", + "image_path": "aa1d8fed080e7513d479c531d475a9249211daed580e383350b6d540639292b0.jpg" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 131, + 565, + 480, + 591 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 131, + 565, + 480, + 591 + ], + "spans": [ + { + "bbox": [ + 131, + 565, + 480, + 591 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 131, + 565, + 480, + 591 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_{(n)}^k (\\widetilde{w}) = \\mathbb{E}_{(x,y)\\sim \\mathcal{T}_{(n)}^k}\\left[l(\\widetilde{w};(x,y))\\right]" + }, + { + "bbox": [ + 131, + 565, + 480, + 591 + ], + "type": "text", + "content": " is the empirical loss of client " + }, + { + "bbox": [ + 131, + 565, + 480, + 591 + ], + "type": "inline_equation", + "content": "k" + }, + { + "bbox": [ + 131, + 565, + 480, + 591 + ], + "type": "text", + "content": " on " + }, + { + "bbox": [ + 131, + 565, + 480, + 591 + ], + "type": "inline_equation", + "content": "n" + }, + { + "bbox": [ + 131, + 565, + 480, + 591 + ], + "type": "text", + "content": "-th task with the personalized local model." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 132, + 609, + 302, + 621 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 609, + 302, + 621 + ], + "spans": [ + { + "bbox": [ + 132, + 609, + 302, + 621 + ], + "type": "text", + "content": "3.2 The Proposed pFedDIL Approach" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 131, + 629, + 481, + 666 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 131, + 629, + 481, + 666 + ], + "spans": [ + { + "bbox": [ + 131, + 629, + 481, + 666 + ], + "type": "text", + "content": "The key idea of the pFedDIL is to identify the correlation between the new and previous tasks such that clients can adaptively migrate knowledge from previous tasks to improve performance. Specifically, in each communication round, the client in pFedDIL trains" + } + ] + } + ], + "index": 11 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 298, + 91, + 447, + 102 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 298, + 91, + 447, + 102 + ], + "spans": [ + { + "bbox": [ + 298, + 91, + 447, + 102 + ], + "type": "text", + "content": "Federated Domain-Incremental Learning" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 474, + 92, + 481, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 474, + 92, + 481, + 100 + ], + "spans": [ + { + "bbox": [ + 474, + 92, + 481, + 100 + ], + "type": "text", + "content": "5" + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 4 + }, + { + "para_blocks": [ + { + "bbox": [ + 130, + 116, + 482, + 259 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 116, + 482, + 259 + ], + "spans": [ + { + "bbox": [ + 130, + 116, + 482, + 259 + ], + "type": "text", + "content": "the local model with its private sequence of tasks and the server aggregates local models from all participating clients. Then, when it comes to new tasks, each client first employs the existing personalized models' auxiliary classifiers to calculate the correlations denoted as knowledge matching intensity and applies the appropriate incremental task learning strategy: training the new task with a previous personalized local model which has the most similar knowledge or training with a new initial model. Simultaneously, each client can migrate knowledge from other personalized local models with the knowledge matching intensity to assist in training. During the inference stage, the final inference result is aggregated by all personalized local models weighted by the corresponding auxiliary classifiers. Moreover, we propose sharing partial parameters between the auxiliary classifier and the target personalized model to condense model parameters. The workflow of the pFedDIL is shown in Algorithm 1." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 130, + 260, + 482, + 418 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 260, + 482, + 418 + ], + "spans": [ + { + "bbox": [ + 130, + 260, + 482, + 418 + ], + "type": "text", + "content": "Knowledge Matching Intensity. To assess the similarity between different incremental tasks, where a higher similarity indicates a greater amount of common knowledge, we propose the knowledge matching intensity. When it comes to the new incremental task: " + }, + { + "bbox": [ + 130, + 260, + 482, + 418 + ], + "type": "inline_equation", + "content": "(t + 1)" + }, + { + "bbox": [ + 130, + 260, + 482, + 418 + ], + "type": "text", + "content": "-th task, each client first needs to compute the knowledge matching intensity between the " + }, + { + "bbox": [ + 130, + 260, + 482, + 418 + ], + "type": "inline_equation", + "content": "(t + 1)" + }, + { + "bbox": [ + 130, + 260, + 482, + 418 + ], + "type": "text", + "content": "-th task and previous " + }, + { + "bbox": [ + 130, + 260, + 482, + 418 + ], + "type": "inline_equation", + "content": "t" + }, + { + "bbox": [ + 130, + 260, + 482, + 418 + ], + "type": "text", + "content": " tasks. To achieve the adaptive knowledge matching intensity estimation for each client, we define the auxiliary classifier, a binary classification deep neural network " + }, + { + "bbox": [ + 130, + 260, + 482, + 418 + ], + "type": "inline_equation", + "content": "\\theta" + }, + { + "bbox": [ + 130, + 260, + 482, + 418 + ], + "type": "text", + "content": " for the corresponding local task " + }, + { + "bbox": [ + 130, + 260, + 482, + 418 + ], + "type": "inline_equation", + "content": "T" + }, + { + "bbox": [ + 130, + 260, + 482, + 418 + ], + "type": "text", + "content": ", which is trained by the discrimination of the sample belonging to the task " + }, + { + "bbox": [ + 130, + 260, + 482, + 418 + ], + "type": "inline_equation", + "content": "T" + }, + { + "bbox": [ + 130, + 260, + 482, + 418 + ], + "type": "text", + "content": ". As a new task emerges, the auxiliary classifier can output the similarity between different tasks based on the extracted feature representation. Supposed that the client " + }, + { + "bbox": [ + 130, + 260, + 482, + 418 + ], + "type": "inline_equation", + "content": "k" + }, + { + "bbox": [ + 130, + 260, + 482, + 418 + ], + "type": "text", + "content": " has converged on " + }, + { + "bbox": [ + 130, + 260, + 482, + 418 + ], + "type": "inline_equation", + "content": "t" + }, + { + "bbox": [ + 130, + 260, + 482, + 418 + ], + "type": "text", + "content": " tasks within " + }, + { + "bbox": [ + 130, + 260, + 482, + 418 + ], + "type": "inline_equation", + "content": "d" + }, + { + "bbox": [ + 130, + 260, + 482, + 418 + ], + "type": "text", + "content": " local existing task models when it comes to the incremental " + }, + { + "bbox": [ + 130, + 260, + 482, + 418 + ], + "type": "inline_equation", + "content": "(t + 1)" + }, + { + "bbox": [ + 130, + 260, + 482, + 418 + ], + "type": "text", + "content": "-th task, the client needs to calculate the knowledge matching intensity " + }, + { + "bbox": [ + 130, + 260, + 482, + 418 + ], + "type": "inline_equation", + "content": "\\widetilde{\\rho}_k = [\\rho_1^k,\\rho_2^k,\\dots ,\\rho_d^k ]" + }, + { + "bbox": [ + 130, + 260, + 482, + 418 + ], + "type": "text", + "content": " with each auxiliary classifier " + }, + { + "bbox": [ + 130, + 260, + 482, + 418 + ], + "type": "inline_equation", + "content": "\\widetilde{\\theta}^{k} = [\\theta_{1}^{k},\\theta_{2}^{k},\\dots ,\\theta_{d}^{k}]" + }, + { + "bbox": [ + 130, + 260, + 482, + 418 + ], + "type": "text", + "content": ":" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 244, + 427, + 481, + 460 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 244, + 427, + 481, + 460 + ], + "spans": [ + { + "bbox": [ + 244, + 427, + 481, + 460 + ], + "type": "interline_equation", + "content": "\\widetilde {\\rho} ^ {k} = \\frac {1}{N ^ {t + 1}} \\sum_ {i = 1} ^ {N ^ {t + 1}} f \\left(x _ {i} ^ {t + 1}; \\widetilde {\\theta} ^ {k}\\right). \\tag {2}", + "image_path": "0719e7cb3b349eabd02b9b7d1446c714b69abaed9159c3bf15d4667a8857dc6d.jpg" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 130, + 472, + 482, + 520 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 472, + 482, + 520 + ], + "spans": [ + { + "bbox": [ + 130, + 472, + 482, + 520 + ], + "type": "text", + "content": "where the " + }, + { + "bbox": [ + 130, + 472, + 482, + 520 + ], + "type": "inline_equation", + "content": "f(\\cdot)" + }, + { + "bbox": [ + 130, + 472, + 482, + 520 + ], + "type": "text", + "content": " denotes the output of the auxiliary classifier, " + }, + { + "bbox": [ + 130, + 472, + 482, + 520 + ], + "type": "inline_equation", + "content": "\\widehat{\\theta}^k" + }, + { + "bbox": [ + 130, + 472, + 482, + 520 + ], + "type": "text", + "content": " denotes a set of personalized local auxiliary classifiers in client " + }, + { + "bbox": [ + 130, + 472, + 482, + 520 + ], + "type": "inline_equation", + "content": "k" + }, + { + "bbox": [ + 130, + 472, + 482, + 520 + ], + "type": "text", + "content": ". Each auxiliary classifier is trained by minimizing the total empirical loss over the corresponding task's local dataset " + }, + { + "bbox": [ + 130, + 472, + 482, + 520 + ], + "type": "inline_equation", + "content": "\\mathcal{T}" + }, + { + "bbox": [ + 130, + 472, + 482, + 520 + ], + "type": "text", + "content": ", which can be formulated as:" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 241, + 527, + 481, + 559 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 241, + 527, + 481, + 559 + ], + "spans": [ + { + "bbox": [ + 241, + 527, + 481, + 559 + ], + "type": "interline_equation", + "content": "\\mathcal {L} ^ {k} (\\theta) = \\frac {1}{T} \\sum_ {i = 1} ^ {\\tau} \\mathcal {L} _ {C E} \\left(\\theta ; x _ {i}, y _ {i}\\right). \\tag {3}", + "image_path": "d75bd49a4044c896cee478a582d9d1f43d45604f058ac85194f2b29bb7331388.jpg" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 130, + 569, + 481, + 593 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 569, + 481, + 593 + ], + "spans": [ + { + "bbox": [ + 130, + 569, + 481, + 593 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 130, + 569, + 481, + 593 + ], + "type": "inline_equation", + "content": "\\mathcal{L}^k (\\theta)" + }, + { + "bbox": [ + 130, + 569, + 481, + 593 + ], + "type": "text", + "content": " is the local loss in the client " + }, + { + "bbox": [ + 130, + 569, + 481, + 593 + ], + "type": "inline_equation", + "content": "k" + }, + { + "bbox": [ + 130, + 569, + 481, + 593 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 130, + 569, + 481, + 593 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_{CE}" + }, + { + "bbox": [ + 130, + 569, + 481, + 593 + ], + "type": "text", + "content": " is the cross-entropy loss function that measures the difference between the prediction and the ground truth labels." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 130, + 594, + 482, + 666 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 594, + 482, + 666 + ], + "spans": [ + { + "bbox": [ + 130, + 594, + 482, + 666 + ], + "type": "text", + "content": "Incremental Task Learning Strategy. After acquiring the knowledge matching intensity " + }, + { + "bbox": [ + 130, + 594, + 482, + 666 + ], + "type": "inline_equation", + "content": "\\widetilde{\\rho}_k" + }, + { + "bbox": [ + 130, + 594, + 482, + 666 + ], + "type": "text", + "content": " with equation (2), the client can select the appropriate incremental task learning strategy. An intuition is that the client will prefer to train with previous personalized local models when the knowledge matching intensity exceeds the " + }, + { + "bbox": [ + 130, + 594, + 482, + 666 + ], + "type": "inline_equation", + "content": "\\lambda" + }, + { + "bbox": [ + 130, + 594, + 482, + 666 + ], + "type": "text", + "content": ", which means the high-level similarity between the new task and the previous tasks so that the model can converge on the new task with slight adjustments. Besides, the client will sacrifice" + } + ] + } + ], + "index": 8 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "spans": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "type": "text", + "content": "6" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 166, + 91, + 225, + 101 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 166, + 91, + 225, + 101 + ], + "spans": [ + { + "bbox": [ + 166, + 91, + 225, + 101 + ], + "type": "text", + "content": "Yichen Li, et al." + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 5 + }, + { + "para_blocks": [ + { + "type": "code", + "bbox": [ + 138, + 132, + 468, + 365 + ], + "blocks": [ + { + "bbox": [ + 139, + 118, + 236, + 131 + ], + "lines": [ + { + "bbox": [ + 139, + 118, + 236, + 131 + ], + "spans": [ + { + "bbox": [ + 139, + 118, + 236, + 131 + ], + "type": "text", + "content": "Algorithm 1: pFedDIL" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "code_caption" + }, + { + "bbox": [ + 138, + 132, + 468, + 365 + ], + "lines": [ + { + "bbox": [ + 138, + 132, + 468, + 365 + ], + "spans": [ + { + "bbox": [ + 138, + 132, + 468, + 365 + ], + "type": "text", + "content": "Input : " + }, + { + "bbox": [ + 138, + 132, + 468, + 365 + ], + "type": "inline_equation", + "content": "T" + }, + { + "bbox": [ + 138, + 132, + 468, + 365 + ], + "type": "text", + "content": " : the communication round; " + }, + { + "bbox": [ + 138, + 132, + 468, + 365 + ], + "type": "inline_equation", + "content": "K" + }, + { + "bbox": [ + 138, + 132, + 468, + 365 + ], + "type": "text", + "content": " : client number; " + }, + { + "bbox": [ + 138, + 132, + 468, + 365 + ], + "type": "inline_equation", + "content": "C" + }, + { + "bbox": [ + 138, + 132, + 468, + 365 + ], + "type": "text", + "content": " : the fraction of active client in each round; " + }, + { + "bbox": [ + 138, + 132, + 468, + 365 + ], + "type": "inline_equation", + "content": "\\{\\mathcal{T}_{(t)}\\}_{t = 1}^{n}" + }, + { + "bbox": [ + 138, + 132, + 468, + 365 + ], + "type": "text", + "content": " : the distributed dataset with " + }, + { + "bbox": [ + 138, + 132, + 468, + 365 + ], + "type": "inline_equation", + "content": "n" + }, + { + "bbox": [ + 138, + 132, + 468, + 365 + ], + "type": "text", + "content": " tasks; w: the parameter of the target classification model; " + }, + { + "bbox": [ + 138, + 132, + 468, + 365 + ], + "type": "inline_equation", + "content": "\\theta" + }, + { + "bbox": [ + 138, + 132, + 468, + 365 + ], + "type": "text", + "content": " : the parameter of the auxiliary classifier. 1 Initialize the parameter w and " + }, + { + "bbox": [ + 138, + 132, + 468, + 365 + ], + "type": "inline_equation", + "content": "\\theta" + }, + { + "bbox": [ + 138, + 132, + 468, + 365 + ], + "type": "text", + "content": " . \n2 for " + }, + { + "bbox": [ + 138, + 132, + 468, + 365 + ], + "type": "inline_equation", + "content": "t = 1" + }, + { + "bbox": [ + 138, + 132, + 468, + 365 + ], + "type": "text", + "content": " to T do \n3 " + }, + { + "bbox": [ + 138, + 132, + 468, + 365 + ], + "type": "inline_equation", + "content": "S_{t}\\gets" + }, + { + "bbox": [ + 138, + 132, + 468, + 365 + ], + "type": "text", + "content": " (random set of " + }, + { + "bbox": [ + 138, + 132, + 468, + 365 + ], + "type": "inline_equation", + "content": "[C\\cdot K]" + }, + { + "bbox": [ + 138, + 132, + 468, + 365 + ], + "type": "text", + "content": " clients); for each selected client " + }, + { + "bbox": [ + 138, + 132, + 468, + 365 + ], + "type": "inline_equation", + "content": "k\\in S_t" + }, + { + "bbox": [ + 138, + 132, + 468, + 365 + ], + "type": "text", + "content": " in parallel do \n4 receives " + }, + { + "bbox": [ + 138, + 132, + 468, + 365 + ], + "type": "inline_equation", + "content": "w_{t}" + }, + { + "bbox": [ + 138, + 132, + 468, + 365 + ], + "type": "text", + "content": " from the server; \n5 calculate knowledge matching intensity " + }, + { + "bbox": [ + 138, + 132, + 468, + 365 + ], + "type": "inline_equation", + "content": "\\widetilde{\\rho}^k" + }, + { + "bbox": [ + 138, + 132, + 468, + 365 + ], + "type": "text", + "content": " with (2); \n6 select learning strategy " + }, + { + "bbox": [ + 138, + 132, + 468, + 365 + ], + "type": "inline_equation", + "content": "w_{t + 1}^{k}" + }, + { + "bbox": [ + 138, + 132, + 468, + 365 + ], + "type": "text", + "content": " with (4); \n7 set local models " + }, + { + "bbox": [ + 138, + 132, + 468, + 365 + ], + "type": "inline_equation", + "content": "w_{t + 1}^{k}" + }, + { + "bbox": [ + 138, + 132, + 468, + 365 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 138, + 132, + 468, + 365 + ], + "type": "inline_equation", + "content": "\\theta_{t + 1}^{k}" + }, + { + "bbox": [ + 138, + 132, + 468, + 365 + ], + "type": "text", + "content": " . \n8 for e = 1 to E do \n9 update " + }, + { + "bbox": [ + 138, + 132, + 468, + 365 + ], + "type": "inline_equation", + "content": "\\theta_{t + 1}^{k}" + }, + { + "bbox": [ + 138, + 132, + 468, + 365 + ], + "type": "text", + "content": " with (3); \n10 update " + }, + { + "bbox": [ + 138, + 132, + 468, + 365 + ], + "type": "inline_equation", + "content": "w_{t + 1}^{k}" + }, + { + "bbox": [ + 138, + 132, + 468, + 365 + ], + "type": "text", + "content": " through adaptive knowledge migration with (5); \n11 end \n12 pushes " + }, + { + "bbox": [ + 138, + 132, + 468, + 365 + ], + "type": "inline_equation", + "content": "w_{t + 1}^{k}" + }, + { + "bbox": [ + 138, + 132, + 468, + 365 + ], + "type": "text", + "content": " to the server. \n13 end \n14 " + }, + { + "bbox": [ + 138, + 132, + 468, + 365 + ], + "type": "inline_equation", + "content": "w_{t}\\gets" + }, + { + "bbox": [ + 138, + 132, + 468, + 365 + ], + "type": "text", + "content": " ServerAggregation(" + }, + { + "bbox": [ + 138, + 132, + 468, + 365 + ], + "type": "inline_equation", + "content": "\\{w^{k}\\}_{k\\in S_{t}}" + }, + { + "bbox": [ + 138, + 132, + 468, + 365 + ], + "type": "text", + "content": " ) \n15 end" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "code_body" + } + ], + "index": 3, + "sub_type": "algorithm" + }, + { + "bbox": [ + 130, + 391, + 482, + 427 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 391, + 482, + 427 + ], + "spans": [ + { + "bbox": [ + 130, + 391, + 482, + 427 + ], + "type": "text", + "content": "some storage by training the new task with the new initial model parameters to ensure accuracy for both the new task and previous tasks. Here we define the incremental task learning strategy as follows:" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 197, + 432, + 481, + 464 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 197, + 432, + 481, + 464 + ], + "spans": [ + { + "bbox": [ + 197, + 432, + 481, + 464 + ], + "type": "interline_equation", + "content": "w _ {t + 1} ^ {k} = O P T \\left\\{ \\begin{array}{l l} \\hat {w} & i f \\max \\left(\\widetilde {\\rho} _ {k}\\right) < \\lambda \\\\ \\widetilde {w} ^ {k} [ m ] & e l s e \\max \\left(\\widetilde {\\rho} _ {k}\\right) = \\rho_ {k} ^ {m} \\geq \\lambda . \\end{array} \\right. \\tag {4}", + "image_path": "f94a0954de9eeaecb1f610666620428ca9812e693d693cfa2f803ed8e1441e8b.jpg" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 130, + 470, + 482, + 506 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 470, + 482, + 506 + ], + "spans": [ + { + "bbox": [ + 130, + 470, + 482, + 506 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 130, + 470, + 482, + 506 + ], + "type": "inline_equation", + "content": "\\lambda" + }, + { + "bbox": [ + 130, + 470, + 482, + 506 + ], + "type": "text", + "content": " is a hyper-parameter, " + }, + { + "bbox": [ + 130, + 470, + 482, + 506 + ], + "type": "inline_equation", + "content": "\\widetilde{w}^k = [w_1^k, w_2^k, \\ldots, w_d^k]" + }, + { + "bbox": [ + 130, + 470, + 482, + 506 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 130, + 470, + 482, + 506 + ], + "type": "inline_equation", + "content": "\\hat{w}" + }, + { + "bbox": [ + 130, + 470, + 482, + 506 + ], + "type": "text", + "content": " denotes new initial model parameters, and " + }, + { + "bbox": [ + 130, + 470, + 482, + 506 + ], + "type": "inline_equation", + "content": "OPT" + }, + { + "bbox": [ + 130, + 470, + 482, + 506 + ], + "type": "text", + "content": " means the alternative of existed personalized local models and a new initial model." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 130, + 506, + 482, + 555 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 506, + 482, + 555 + ], + "spans": [ + { + "bbox": [ + 130, + 506, + 482, + 555 + ], + "type": "text", + "content": "Knowledge Migration. Then, during the training of " + }, + { + "bbox": [ + 130, + 506, + 482, + 555 + ], + "type": "inline_equation", + "content": "(t + 1)" + }, + { + "bbox": [ + 130, + 506, + 482, + 555 + ], + "type": "text", + "content": "-th task, each client respectively migrates the knowledge from other existing personalized local models to assist the training with the calculated knowledge matching intensity " + }, + { + "bbox": [ + 130, + 506, + 482, + 555 + ], + "type": "inline_equation", + "content": "\\widetilde{\\rho}_k" + }, + { + "bbox": [ + 130, + 506, + 482, + 555 + ], + "type": "text", + "content": ", minimize the following objective:" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 195, + 559, + 419, + 582 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 195, + 559, + 419, + 582 + ], + "spans": [ + { + "bbox": [ + 195, + 559, + 419, + 582 + ], + "type": "interline_equation", + "content": "\\min _ {w _ {t + 1} ^ {k}} \\mathcal {L} _ {L o c a l} ^ {k} (w _ {t + 1} ^ {k}) = \\mathcal {L} _ {(t + 1)} ^ {k} (w _ {t + 1} ^ {k}) + \\mathcal {L} _ {K M} ^ {k} (w _ {t + 1} ^ {k}),", + "image_path": "3a5ec211a7e9ee8941f08a6ecd268f15fe8bc9a3fbc2714f0193dc2c66135d0b.jpg" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 195, + 584, + 481, + 635 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 195, + 584, + 481, + 635 + ], + "spans": [ + { + "bbox": [ + 195, + 584, + 481, + 635 + ], + "type": "interline_equation", + "content": "\\begin{array}{l} \\text {w h e r e} \\quad \\mathcal {L} _ {K M} ^ {k} \\left(w _ {t + 1} ^ {k}\\right) = \\widetilde {\\rho} _ {k} \\cdot \\left| \\left| w _ {t + 1} ^ {k} - \\widetilde {w} ^ {k} \\right| \\right| ^ {2}, \\\\ = \\sum_ {i = 1} ^ {d} \\rho^ {i} \\cdot \\left| \\left| w _ {t + 1} ^ {k} - w _ {i} ^ {k} \\right| \\right| ^ {2}. \\tag {5} \\\\ \\end{array}", + "image_path": "0e95f5fc28ccd705d781dfd7af97ebed5082fa5f7a2b5179c01de1080fbee08a.jpg" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 130, + 640, + 481, + 665 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 640, + 481, + 665 + ], + "spans": [ + { + "bbox": [ + 130, + 640, + 481, + 665 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 130, + 640, + 481, + 665 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_{t + 1}^k (w_{t + 1}^k)" + }, + { + "bbox": [ + 130, + 640, + 481, + 665 + ], + "type": "text", + "content": " is defined in (1) and " + }, + { + "bbox": [ + 130, + 640, + 481, + 665 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_{KM}^{k}(w_{t + 1}^{k})" + }, + { + "bbox": [ + 130, + 640, + 481, + 665 + ], + "type": "text", + "content": " refers to the Knowledge Migration (KM) loss." + } + ] + } + ], + "index": 10 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 298, + 91, + 448, + 102 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 298, + 91, + 448, + 102 + ], + "spans": [ + { + "bbox": [ + 298, + 91, + 448, + 102 + ], + "type": "text", + "content": "Federated Domain-Incremental Learning" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 474, + 91, + 480, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 474, + 91, + 480, + 100 + ], + "spans": [ + { + "bbox": [ + 474, + 91, + 480, + 100 + ], + "type": "text", + "content": "7" + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 6 + }, + { + "para_blocks": [ + { + "bbox": [ + 130, + 116, + 482, + 226 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 116, + 482, + 226 + ], + "spans": [ + { + "bbox": [ + 130, + 116, + 482, + 226 + ], + "type": "text", + "content": "Inference Stage. During the inference stage, taking into account that each client may have multiple personalized models, how to automatically select the appropriate model for inference is an urgent problem to solve. We have observed that ensemble learning, which enhances model performance by integrating knowledge from multiple smaller models, has achieved widespread success [8]. Utilizing the calculated knowledge matching intensity as the weight for the personalized model inference results, we achieve automated ensemble inference for the inference data. For every inference sample " + }, + { + "bbox": [ + 130, + 116, + 482, + 226 + ], + "type": "inline_equation", + "content": "x \\in \\hat{\\mathcal{T}}" + }, + { + "bbox": [ + 130, + 116, + 482, + 226 + ], + "type": "text", + "content": ", the client " + }, + { + "bbox": [ + 130, + 116, + 482, + 226 + ], + "type": "inline_equation", + "content": "k" + }, + { + "bbox": [ + 130, + 116, + 482, + 226 + ], + "type": "text", + "content": " computes the weights of each personalized local models " + }, + { + "bbox": [ + 130, + 116, + 482, + 226 + ], + "type": "inline_equation", + "content": "\\widetilde{w}^k" + }, + { + "bbox": [ + 130, + 116, + 482, + 226 + ], + "type": "text", + "content": " with the auxiliary " + }, + { + "bbox": [ + 130, + 116, + 482, + 226 + ], + "type": "inline_equation", + "content": "\\widetilde{\\theta}^k" + }, + { + "bbox": [ + 130, + 116, + 482, + 226 + ], + "type": "text", + "content": " as " + }, + { + "bbox": [ + 130, + 116, + 482, + 226 + ], + "type": "inline_equation", + "content": "f(\\widetilde{\\theta}^k; x)" + }, + { + "bbox": [ + 130, + 116, + 482, + 226 + ], + "type": "text", + "content": " and then normalizes it into the probability:" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 260, + 232, + 481, + 262 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 260, + 232, + 481, + 262 + ], + "spans": [ + { + "bbox": [ + 260, + 232, + 481, + 262 + ], + "type": "interline_equation", + "content": "\\hat {\\alpha} _ {i} ^ {k} = \\frac {f \\left(\\theta_ {i} ^ {k} ; x\\right)}{\\sum_ {i = 1} ^ {d} f \\left(\\theta_ {i} ^ {k} ; x\\right)}. \\tag {6}", + "image_path": "2a182d932e5cac25f7ccb9989412077785d28f20954567cbe3032c1517f88edd.jpg" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 131, + 270, + 481, + 307 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 131, + 270, + 481, + 307 + ], + "spans": [ + { + "bbox": [ + 131, + 270, + 481, + 307 + ], + "type": "text", + "content": "which guarantees that " + }, + { + "bbox": [ + 131, + 270, + 481, + 307 + ], + "type": "inline_equation", + "content": "\\sum_{i=1}^{d} \\hat{\\alpha}_i^k = 1" + }, + { + "bbox": [ + 131, + 270, + 481, + 307 + ], + "type": "text", + "content": ". Finally, the client inputs the inference sample " + }, + { + "bbox": [ + 131, + 270, + 481, + 307 + ], + "type": "inline_equation", + "content": "x" + }, + { + "bbox": [ + 131, + 270, + 481, + 307 + ], + "type": "text", + "content": " into each personalized local models " + }, + { + "bbox": [ + 131, + 270, + 481, + 307 + ], + "type": "inline_equation", + "content": "\\widetilde{w}^k" + }, + { + "bbox": [ + 131, + 270, + 481, + 307 + ], + "type": "text", + "content": " to obtain the soft predictions " + }, + { + "bbox": [ + 131, + 270, + 481, + 307 + ], + "type": "inline_equation", + "content": "s(\\widetilde{w}^k; x)" + }, + { + "bbox": [ + 131, + 270, + 481, + 307 + ], + "type": "text", + "content": " and final weighed prediction:" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 221, + 315, + 481, + 347 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 221, + 315, + 481, + 347 + ], + "spans": [ + { + "bbox": [ + 221, + 315, + 481, + 347 + ], + "type": "interline_equation", + "content": "I n f e r e n c e \\left(\\widetilde {w} ^ {k}; x\\right) = \\sum_ {i = 1} ^ {d} \\hat {\\alpha} _ {i} ^ {k} \\cdot s \\left(w _ {i} ^ {k}; x\\right). \\tag {7}", + "image_path": "2f1d1a64c71d7f8da5c481a637e86583268906c9f456acc04901f20f14f7da64.jpg" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 131, + 354, + 482, + 426 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 131, + 354, + 482, + 426 + ], + "spans": [ + { + "bbox": [ + 131, + 354, + 482, + 426 + ], + "type": "text", + "content": "Partial Parameters Sharing. Motivated by the idea of multi-task learning where sharing the encoder between different tasks can promote each other [34], we propose sharing partial parameters between the auxiliary classifier " + }, + { + "bbox": [ + 131, + 354, + 482, + 426 + ], + "type": "inline_equation", + "content": "\\theta^k" + }, + { + "bbox": [ + 131, + 354, + 482, + 426 + ], + "type": "text", + "content": " and the target classification model " + }, + { + "bbox": [ + 131, + 354, + 482, + 426 + ], + "type": "inline_equation", + "content": "w^k" + }, + { + "bbox": [ + 131, + 354, + 482, + 426 + ], + "type": "text", + "content": " to solve this problem. The intuition behind this is that both the auxiliary classifier and the classification model have to distinguish the sample from the extracted features. Besides, another benefit of sharing layers is that the model parameters can be reduced." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 131, + 426, + 482, + 510 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 131, + 426, + 482, + 510 + ], + "spans": [ + { + "bbox": [ + 131, + 426, + 482, + 510 + ], + "type": "text", + "content": "To this end, we propose sharing the front model layers between the two models which are used to extract features, as illustrated in Figure 1. Specifically, by denoting the " + }, + { + "bbox": [ + 131, + 426, + 482, + 510 + ], + "type": "inline_equation", + "content": "c" + }, + { + "bbox": [ + 131, + 426, + 482, + 510 + ], + "type": "text", + "content": "-layer shared extractor by " + }, + { + "bbox": [ + 131, + 426, + 482, + 510 + ], + "type": "inline_equation", + "content": "\\tilde{w}^k = [\\tilde{w}^{k,1},\\dots,\\tilde{w}^{k,c}]" + }, + { + "bbox": [ + 131, + 426, + 482, + 510 + ], + "type": "text", + "content": " where " + }, + { + "bbox": [ + 131, + 426, + 482, + 510 + ], + "type": "inline_equation", + "content": "\\tilde{w}^{k,i}" + }, + { + "bbox": [ + 131, + 426, + 482, + 510 + ], + "type": "text", + "content": " is the " + }, + { + "bbox": [ + 131, + 426, + 482, + 510 + ], + "type": "inline_equation", + "content": "i" + }, + { + "bbox": [ + 131, + 426, + 482, + 510 + ], + "type": "text", + "content": "-th layer of the shared extractor, we can re-write the " + }, + { + "bbox": [ + 131, + 426, + 482, + 510 + ], + "type": "inline_equation", + "content": "n_d" + }, + { + "bbox": [ + 131, + 426, + 482, + 510 + ], + "type": "text", + "content": "-layer auxiliary classifier by " + }, + { + "bbox": [ + 131, + 426, + 482, + 510 + ], + "type": "inline_equation", + "content": "\\theta^k = [\\tilde{w}^k,\\theta^{k,c + 1},\\dots,\\theta^{k,n_d}]" + }, + { + "bbox": [ + 131, + 426, + 482, + 510 + ], + "type": "text", + "content": ", and the " + }, + { + "bbox": [ + 131, + 426, + 482, + 510 + ], + "type": "inline_equation", + "content": "n" + }, + { + "bbox": [ + 131, + 426, + 482, + 510 + ], + "type": "text", + "content": "-layer target classification model " + }, + { + "bbox": [ + 131, + 426, + 482, + 510 + ], + "type": "inline_equation", + "content": "w^k" + }, + { + "bbox": [ + 131, + 426, + 482, + 510 + ], + "type": "text", + "content": " by " + }, + { + "bbox": [ + 131, + 426, + 482, + 510 + ], + "type": "inline_equation", + "content": "w^k = [\\tilde{w}^k,w^{k,c + 1},\\dots,w^{k,n}]" + }, + { + "bbox": [ + 131, + 426, + 482, + 510 + ], + "type": "text", + "content": " with " + }, + { + "bbox": [ + 131, + 426, + 482, + 510 + ], + "type": "inline_equation", + "content": "w^{k,i}" + }, + { + "bbox": [ + 131, + 426, + 482, + 510 + ], + "type": "text", + "content": " as the " + }, + { + "bbox": [ + 131, + 426, + 482, + 510 + ], + "type": "inline_equation", + "content": "i" + }, + { + "bbox": [ + 131, + 426, + 482, + 510 + ], + "type": "text", + "content": "-th layer of the classification model " + }, + { + "bbox": [ + 131, + 426, + 482, + 510 + ], + "type": "inline_equation", + "content": "w^{k,i}" + }, + { + "bbox": [ + 131, + 426, + 482, + 510 + ], + "type": "text", + "content": ". With the shared layers, the discriminator and the classification model are trained jointly:" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 225, + 516, + 481, + 536 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 225, + 516, + 481, + 536 + ], + "spans": [ + { + "bbox": [ + 225, + 516, + 481, + 536 + ], + "type": "interline_equation", + "content": "\\min _ {\\theta^ {k}, w ^ {k}} \\mathcal {L} _ {\\mathcal {J}} ^ {k} \\left(\\theta^ {k}, w ^ {k}\\right) = \\mathcal {L} ^ {k} \\left(\\theta^ {k}\\right) + \\mathcal {L} ^ {k} \\left(w ^ {k}\\right). \\tag {8}", + "image_path": "1c9529f446be90cfd958237c727c936d4c1bd9011ee64e771582b099e1cb14fb.jpg" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 131, + 544, + 360, + 558 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 131, + 544, + 360, + 558 + ], + "spans": [ + { + "bbox": [ + 131, + 544, + 360, + 558 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 131, + 544, + 360, + 558 + ], + "type": "inline_equation", + "content": "\\mathcal{L}^k (\\theta^k)" + }, + { + "bbox": [ + 131, + 544, + 360, + 558 + ], + "type": "text", + "content": " refers to (3) and " + }, + { + "bbox": [ + 131, + 544, + 360, + 558 + ], + "type": "inline_equation", + "content": "\\mathcal{L}^k (w^k)" + }, + { + "bbox": [ + 131, + 544, + 360, + 558 + ], + "type": "text", + "content": " is defined in (1)." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 131, + 574, + 218, + 588 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 131, + 574, + 218, + 588 + ], + "spans": [ + { + "bbox": [ + 131, + 574, + 218, + 588 + ], + "type": "text", + "content": "4 Experiments" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 131, + 598, + 182, + 611 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 131, + 598, + 182, + 611 + ], + "spans": [ + { + "bbox": [ + 131, + 598, + 182, + 611 + ], + "type": "text", + "content": "4.1 Setup" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 130, + 617, + 481, + 666 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 617, + 481, + 666 + ], + "spans": [ + { + "bbox": [ + 130, + 617, + 481, + 666 + ], + "type": "text", + "content": "Dataset: We conduct all our experiments with heterogeneous dataset partition on three datasets. In FDIL scenarios, new domains are introduced gradually. The dataset initially contains samples from a specific domain, and new domains are introduced at later stages, enabling models to adapt and generalize to new unseen domains." + } + ] + } + ], + "index": 12 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "spans": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "type": "text", + "content": "8" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 166, + 91, + 225, + 101 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 166, + 91, + 225, + 101 + ], + "spans": [ + { + "bbox": [ + 166, + 91, + 225, + 101 + ], + "type": "text", + "content": "Yichen Li, et al." + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 7 + }, + { + "para_blocks": [ + { + "type": "table", + "bbox": [ + 159, + 146, + 455, + 324 + ], + "blocks": [ + { + "bbox": [ + 132, + 114, + 482, + 137 + ], + "lines": [ + { + "bbox": [ + 132, + 114, + 482, + 137 + ], + "spans": [ + { + "bbox": [ + 132, + 114, + 482, + 137 + ], + "type": "text", + "content": "Table 1: Performance comparison by Top-1 test accuracy for our model and baseline methods on Digit-10 (data heterogeneity: " + }, + { + "bbox": [ + 132, + 114, + 482, + 137 + ], + "type": "inline_equation", + "content": "\\alpha = 0.1" + }, + { + "bbox": [ + 132, + 114, + 482, + 137 + ], + "type": "text", + "content": ")." + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 159, + 146, + 455, + 324 + ], + "lines": [ + { + "bbox": [ + 159, + 146, + 455, + 324 + ], + "spans": [ + { + "bbox": [ + 159, + 146, + 455, + 324 + ], + "type": "table", + "html": "
Digit-10 (λ = 0.5)
MethodMNIST → EMNIST → USPS → SVHNAvgΔ(↑)
FedAvg [24]92.82 →84.02 →81.02 →71.8782.436.85↑
FedProx [19]93.01 →84.68 →78.30 →76.4283.106.18↑
FedCIL [28]94.64 →87.52 →82.16 →80.9286.312.97↑
DANN [10] + FL96.07 →86.71 →79.11 →72.1483.515.77↑
Source-Only92.82 →82.15 →75.53 →69.0679.899.39↑
Sharing92.67 →82.91 →76.17 →71.9680.938.35↑
pFedDIL -w/o Migration92.82 →87.09 →84.00 →81.8986.452.83↑
pFedDIL -w/o Sharing92.79 →90.13 →87.59 →86.1989.180.11↑
pFedDIL -w/o Correlation92.82 →89.06 →84.72 →84.3087.731.55↑
pFedDIL92.82 →90.39 →87.14 →86.7589.28/
Upper Bound -Disjoint92.82 →90.33 →93.41 →87.9791.131.85↓
", + "image_path": "ebfd1776b333a67be4d19cb7db17eef38008215d2cdea21544a6b22da27f3072.jpg" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "table_body" + } + ], + "index": 3 + }, + { + "bbox": [ + 138, + 346, + 481, + 454 + ], + "type": "list", + "angle": 0, + "index": 7, + "blocks": [ + { + "bbox": [ + 138, + 346, + 481, + 382 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 346, + 481, + 382 + ], + "spans": [ + { + "bbox": [ + 138, + 346, + 481, + 382 + ], + "type": "text", + "content": "- Digit10: Digit-10 dataset contains 10 digit categories in four domains: MNIST [17], EMNIST [3], USPS [13], SVHN [26]. Each dataset is a digit image classification dataset of 10 classes in a specific domain, such as handwriting style." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 138, + 383, + 481, + 418 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 383, + 481, + 418 + ], + "spans": [ + { + "bbox": [ + 138, + 383, + 481, + 418 + ], + "type": "text", + "content": "- Office31: [32] A dataset with images from three different domains: Amazon, Webcam, and DSLR. It consists of 31 object categories, with each domain having around 4,100 images." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 138, + 418, + 481, + 454 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 418, + 481, + 454 + ], + "spans": [ + { + "bbox": [ + 138, + 418, + 481, + 454 + ], + "type": "text", + "content": "- DomainNet: [27] A large-scale dataset with images from six different domains: Clipart, Painting, Real, Sketch, Quickdraw, and Infograph. It contains over 0.6 million images across 345 categories." + } + ] + } + ], + "index": 6 + } + ], + "sub_type": "text" + }, + { + "bbox": [ + 130, + 462, + 482, + 533 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 462, + 482, + 533 + ], + "spans": [ + { + "bbox": [ + 130, + 462, + 482, + 533 + ], + "type": "text", + "content": "Baseline: For a fair comparison with other key related works, we follow the same protocols proposed by [24] to set federated learning(FL) task and utilize the domain-incremental learning (DIL) task by [25]. We evaluate all baselines under the above settings, including two representative FL models FedAvg [24] and FedProx [19], which is better at tackling heterogeneity in federated networks than FedAvg, and the following several methods in FDIL scenario:" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 130, + 533, + 482, + 570 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 533, + 482, + 570 + ], + "spans": [ + { + "bbox": [ + 130, + 533, + 482, + 570 + ], + "type": "text", + "content": "FedCIL [28]: This approach employs the ACGAN backbone to generate synthetic samples to consolidate the global model and align sample features in the output layer. Authors conduct experiments in FCIL scenarios, and here we adopt it to the FDIL setting." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 130, + 570, + 482, + 605 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 570, + 482, + 605 + ], + "spans": [ + { + "bbox": [ + 130, + 570, + 482, + 605 + ], + "type": "text", + "content": "Domain Adaptation + FL: Here we adopt the robust adversarial-based method DANN [10]. This baseline mainly follows the domain adaptation paradigm which is different from the incremental learning and is often prone to issues like catastrophic forgetting." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 130, + 605, + 482, + 640 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 605, + 482, + 640 + ], + "spans": [ + { + "bbox": [ + 130, + 605, + 482, + 640 + ], + "type": "text", + "content": "Source-Only: Here we maintain the initial model trained on the initial domain data completely fixed throughout all tasks. This baseline does not perform any adaptation, modification or replacement to its parameters and serves as a lower bound." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 130, + 641, + 482, + 665 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 641, + 482, + 665 + ], + "spans": [ + { + "bbox": [ + 130, + 641, + 482, + 665 + ], + "type": "text", + "content": "Sharing: Inspired by the multi-task learning scenario [31], we adopt all front layers before the last fully connected layer as shared layers, and we use relevant different" + } + ] + } + ], + "index": 12 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 298, + 91, + 447, + 102 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 298, + 91, + 447, + 102 + ], + "spans": [ + { + "bbox": [ + 298, + 91, + 447, + 102 + ], + "type": "text", + "content": "Federated Domain-Incremental Learning" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 474, + 92, + 481, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 474, + 92, + 481, + 100 + ], + "spans": [ + { + "bbox": [ + 474, + 92, + 481, + 100 + ], + "type": "text", + "content": "9" + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 8 + }, + { + "para_blocks": [ + { + "type": "table", + "bbox": [ + 177, + 146, + 436, + 324 + ], + "blocks": [ + { + "bbox": [ + 132, + 114, + 482, + 137 + ], + "lines": [ + { + "bbox": [ + 132, + 114, + 482, + 137 + ], + "spans": [ + { + "bbox": [ + 132, + 114, + 482, + 137 + ], + "type": "text", + "content": "Table 2: Performance comparison by Top-1 test accuracy for our model and baseline methods on Office-31 (data heterogeneity: " + }, + { + "bbox": [ + 132, + 114, + 482, + 137 + ], + "type": "inline_equation", + "content": "\\alpha = 1" + }, + { + "bbox": [ + 132, + 114, + 482, + 137 + ], + "type": "text", + "content": ")." + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 177, + 146, + 436, + 324 + ], + "lines": [ + { + "bbox": [ + 177, + 146, + 436, + 324 + ], + "spans": [ + { + "bbox": [ + 177, + 146, + 436, + 324 + ], + "type": "table", + "html": "
Office-31 (λ = 0.8)
MethodAmazon → Dlsr → WebcamAvgΔ(↑)
FedAvg [24]46.53 → 24.17 → 29.3433.3512.82↑
FedProx [19]45.19 → 25.68 → 32.2334.3711.80↑
FedCIL [28]49.38 → 39.65 → 44.0444.361.81↑
DANN [10] + FL51.97 → 35.96 → 43.0843.672.50↑
Source-Only46.53 → 20.61 → 28.3231.8214.35↑
Sharing46.11 → 24.23 → 34.2534.8611.31↑
pFedDIL -w/o Migration46.53 → 40.17 → 44.5543.752.42↑
pFedDIL -w/o Sharing46.63 → 42.90 → 47.0345.520.65↑
pFedDIL -w/o Correlation46.53 → 41.66 → 47.1945.131.04↑
pFedDIL46.53 → 43.87 → 48.1246.17/
Upper Bound -Disjoint46.53 → 50.17 → 56.2150.974.80↓
", + "image_path": "3474622a804649e4b514bd5ea5edf4e1235726b04119535b7573202c9095d82c.jpg" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "table_body" + } + ], + "index": 3 + }, + { + "bbox": [ + 130, + 347, + 479, + 371 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 347, + 479, + 371 + ], + "spans": [ + { + "bbox": [ + 130, + 347, + 479, + 371 + ], + "type": "text", + "content": "fully-connected layers to get outputs of different domains. During inference, the fully-connected layer that matches the corresponding domain ID will be applied." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 130, + 372, + 480, + 420 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 372, + 480, + 420 + ], + "spans": [ + { + "bbox": [ + 130, + 372, + 480, + 420 + ], + "type": "text", + "content": "Disjoint: Here we train a separate model for each domain. During inference, the model which matches the corresponding domain ID is used. This baseline requires additional space to store parameters and additional cost to obtain inference-task ID. It is a degradation of domain-incremental learning and provides an upper bound here." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 130, + 420, + 482, + 575 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 420, + 482, + 575 + ], + "spans": [ + { + "bbox": [ + 130, + 420, + 482, + 575 + ], + "type": "text", + "content": "Configuration: Unless otherwise mentioned, we set the number of local training epoch " + }, + { + "bbox": [ + 130, + 420, + 482, + 575 + ], + "type": "inline_equation", + "content": "E = 20" + }, + { + "bbox": [ + 130, + 420, + 482, + 575 + ], + "type": "text", + "content": ", communication round " + }, + { + "bbox": [ + 130, + 420, + 482, + 575 + ], + "type": "inline_equation", + "content": "T = 180" + }, + { + "bbox": [ + 130, + 420, + 482, + 575 + ], + "type": "text", + "content": " for each domain. Here we train each task until convergence before the arrival of a new task with communication rounds " + }, + { + "bbox": [ + 130, + 420, + 482, + 575 + ], + "type": "inline_equation", + "content": "T" + }, + { + "bbox": [ + 130, + 420, + 482, + 575 + ], + "type": "text", + "content": ". For local training, the batch size is 32, the learning rate for our models is 0.001 and the weight decay is " + }, + { + "bbox": [ + 130, + 420, + 482, + 575 + ], + "type": "inline_equation", + "content": "1e - 3" + }, + { + "bbox": [ + 130, + 420, + 482, + 575 + ], + "type": "text", + "content": ". Like most FL methods on Non-IID data distribution [24], we use the Dirichlet distribution " + }, + { + "bbox": [ + 130, + 420, + 482, + 575 + ], + "type": "inline_equation", + "content": "(\\alpha)" + }, + { + "bbox": [ + 130, + 420, + 482, + 575 + ], + "type": "text", + "content": " to simulate the data heterogeneity from all domains and we apply the same data heterogeneity for all domains in the same dataset. For the classifier in all methods, we employ ResNet18 [12] as the basic backbone. For the multi-task learning structure in our approach, we treat all previous layers before the last fully connected layer as share layers, and we use two different fully connected layers to get outputs as the auxiliary classifier result and target classification result. The total clients' number is 20 for Digit-10 and DomainNet with an active ratio " + }, + { + "bbox": [ + 130, + 420, + 482, + 575 + ], + "type": "inline_equation", + "content": "k = 0.4" + }, + { + "bbox": [ + 130, + 420, + 482, + 575 + ], + "type": "text", + "content": ", 10 for Office-31 with an active ratio " + }, + { + "bbox": [ + 130, + 420, + 482, + 575 + ], + "type": "inline_equation", + "content": "k = 0.4" + }, + { + "bbox": [ + 130, + 420, + 482, + 575 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 132, + 595, + 256, + 607 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 595, + 256, + 607 + ], + "spans": [ + { + "bbox": [ + 132, + 595, + 256, + 607 + ], + "type": "text", + "content": "4.2 Performance Overview" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 130, + 617, + 482, + 666 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 617, + 482, + 666 + ], + "spans": [ + { + "bbox": [ + 130, + 617, + 482, + 666 + ], + "type": "text", + "content": "Test Accuracy. Table 1,2,3 shows the performance of our pFedDIL and other baseline methods on three datasets, where \" " + }, + { + "bbox": [ + 130, + 617, + 482, + 666 + ], + "type": "inline_equation", + "content": "\\Delta" + }, + { + "bbox": [ + 130, + 617, + 482, + 666 + ], + "type": "text", + "content": " \" denotes the improvement of our method with other baselines. We carry out experiments against data heterogeneity on each dataset, and pFedDIL achieves the best performance in most cases by a margin of" + } + ] + } + ], + "index": 8 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "spans": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "type": "text", + "content": "10" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 166, + 91, + 225, + 101 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 166, + 91, + 225, + 101 + ], + "spans": [ + { + "bbox": [ + 166, + 91, + 225, + 101 + ], + "type": "text", + "content": "Yichen Li, et al." + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 9 + }, + { + "para_blocks": [ + { + "type": "table", + "bbox": [ + 134, + 147, + 482, + 304 + ], + "blocks": [ + { + "bbox": [ + 131, + 114, + 482, + 137 + ], + "lines": [ + { + "bbox": [ + 131, + 114, + 482, + 137 + ], + "spans": [ + { + "bbox": [ + 131, + 114, + 482, + 137 + ], + "type": "text", + "content": "Table 3: Performance comparison by Top-10 test accuracy for our model and baseline methods on DomainNet (data heterogeneity: " + }, + { + "bbox": [ + 131, + 114, + 482, + 137 + ], + "type": "inline_equation", + "content": "\\alpha = 10" + }, + { + "bbox": [ + 131, + 114, + 482, + 137 + ], + "type": "text", + "content": ")" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 134, + 147, + 482, + 304 + ], + "lines": [ + { + "bbox": [ + 134, + 147, + 482, + 304 + ], + "spans": [ + { + "bbox": [ + 134, + 147, + 482, + 304 + ], + "type": "table", + "html": "
DomainNet (λ = 0.8)
MethodClipart → Infograph → Painting → Quickdraw → Real → SketchAvgΔ(↑)
FedAvg [24]48.43 → 37.18 → 45.80 → 45.32 → 44.91 → 50.2545.326.38↑
FedProx [19]47.39 → 38.43 → 44.31 → 47.96 → 42.38 → 51.7745.376.32↑
FedCIL [28]52.14 → 43.68 → 47.10 → 48.75 → 42.89 → 51.2647.644.05↑
DANN [10] + FL50.07 → 39.74 → 43.73 → 45.08 → 43.28 → 52.9645.815.88↑
Source-Only48.43 → 31.01 → 33.12 → 38.15 → 37.62 → 42.8538.5313.16↑
Sharing48.43 → 36.03 → 39.58 → 41.80 → 40.41 → 44.7641.849.86↑
pFedDIL -w/o Migration48.43 → 43.83 → 50.05 → 49.44 → 46.10 → 54.9048.792.90↑
pFedDIL -w/o Sharing48.52 → 47.37 → 51.16 → 53.98 → 50.11 → 55.2451.060.63↑
pFedDIL -w/o Correlation48.43 → 46.71 → 51.09 → 52.34 → 48.07 → 56.9550.601.09↑
pFedDIL48.43 → 47.19 → 52.48 → 54.07 → 50.36 → 57.6151.69/
Upper Bound -Disjoint48.43 → 52.33 → 54.21 → 52.89 → 53.62 → 58.0653.261.57↓
", + "image_path": "fd7ff2eb691aade0653caf3a4038fe3eae456dc9aa0a1c2e8dbebddee899cbaa.jpg" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "table_body" + } + ], + "index": 3 + }, + { + "bbox": [ + 130, + 327, + 482, + 399 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 327, + 482, + 399 + ], + "spans": [ + { + "bbox": [ + 130, + 327, + 482, + 399 + ], + "type": "inline_equation", + "content": "1.81\\% \\sim 14.35\\%" + }, + { + "bbox": [ + 130, + 327, + 482, + 399 + ], + "type": "text", + "content": " in terms of average accuracy. Due to our challenging FDIL settings, there is a notable gap between all baselines and the upper performance bound. However, pFedDIL almost achieves comparable results with less storage and adaptive task-id inference during the inference stage. These results demonstrate that the appropriate incremental task learning strategy and adaptive knowledge migration technique in pFedDIL can alleviate catastrophic forgetting and thus improve performance." + } + ] + } + ], + "index": 4 + }, + { + "type": "image", + "bbox": [ + 132, + 419, + 482, + 540 + ], + "blocks": [ + { + "bbox": [ + 132, + 419, + 482, + 540 + ], + "lines": [ + { + "bbox": [ + 132, + 419, + 482, + 540 + ], + "spans": [ + { + "bbox": [ + 132, + 419, + 482, + 540 + ], + "type": "image", + "image_path": "84b22f2020e652fdb2540cbd3e3a629542b40f397e4fdb82561514878995d3cc.jpg" + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 173, + 546, + 440, + 558 + ], + "lines": [ + { + "bbox": [ + 173, + 546, + 440, + 558 + ], + "spans": [ + { + "bbox": [ + 173, + 546, + 440, + 558 + ], + "type": "text", + "content": "Fig. 2: Visualized performance w.r.t data heterogeneity for three datasets." + } + ] + } + ], + "index": 6, + "angle": 0, + "type": "image_caption" + } + ], + "index": 5 + }, + { + "bbox": [ + 130, + 581, + 482, + 666 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 581, + 482, + 666 + ], + "spans": [ + { + "bbox": [ + 130, + 581, + 482, + 666 + ], + "type": "text", + "content": "Ablation Study. As shown in Tables 1,2,3, we evaluate the effects of each module in our model via ablation studies. pFedDIL -w/o Migration, pFedDIL -w/o Sharing and pFedDIL- w/o Correlation denote the performance of our model without using knowledge migration, parameters sharing and correlation factor in the inference stage. Compared with pFedDIL, the performance of pFedDIL -w/o Migration, pFedDIL -w/o Sharing and pFedDIL- w/o Correlation degrades evidently with a range of " + }, + { + "bbox": [ + 130, + 581, + 482, + 666 + ], + "type": "inline_equation", + "content": "0.11\\% \\sim 2.90\\%" + }, + { + "bbox": [ + 130, + 581, + 482, + 666 + ], + "type": "text", + "content": ". Specifically, the adaptive knowledge migration technique in pFedDIL plays a much" + } + ] + } + ], + "index": 7 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 298, + 91, + 448, + 102 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 298, + 91, + 448, + 102 + ], + "spans": [ + { + "bbox": [ + 298, + 91, + 448, + 102 + ], + "type": "text", + "content": "Federated Domain-Incremental Learning" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 470, + 91, + 480, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 470, + 91, + 480, + 100 + ], + "spans": [ + { + "bbox": [ + 470, + 91, + 480, + 100 + ], + "type": "text", + "content": "11" + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 10 + }, + { + "para_blocks": [ + { + "bbox": [ + 130, + 116, + 482, + 175 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 116, + 482, + 175 + ], + "spans": [ + { + "bbox": [ + 130, + 116, + 482, + 175 + ], + "type": "text", + "content": "more important role with the obvious improvement in test accuracy, and the partial parameters sharing trick condenses model parameters by sharing the feature extracted layer while not causing a remarkable impact on test accuracy. Experiment results verify the effectiveness of all modules, confirming all modules are essential to train a robust federated domain-incremental model." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 130, + 176, + 482, + 297 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 176, + 482, + 297 + ], + "spans": [ + { + "bbox": [ + 130, + 176, + 482, + 297 + ], + "type": "text", + "content": "Data heterogeneity. Increasing data heterogeneity levels in experiments is significant because it allows for a more realistic evaluation of the system's performance in diverse real-world scenarios and assesses its robustness and generalization capabilities. In particular, in the FDIL scenario, incremental data is less likely to appear in the form of IID data. Therefore, it is essential to simulate different levels of Non-IID data to validate the effectiveness of our method in the FDIL scenario. Figure 2 displays the accuracy with different levels of data heterogeneity on three datasets. As shown in this figure, all methods achieve an improvement in test accuracy with the decline of data heterogeneity. Most notably, pFedDIL always maintains a leading advantage with different levels of data heterogeneity." + } + ] + } + ], + "index": 3 + }, + { + "type": "table", + "bbox": [ + 142, + 360, + 473, + 439 + ], + "blocks": [ + { + "bbox": [ + 130, + 316, + 483, + 351 + ], + "lines": [ + { + "bbox": [ + 130, + 316, + 483, + 351 + ], + "spans": [ + { + "bbox": [ + 130, + 316, + 483, + 351 + ], + "type": "text", + "content": "Table 4: Evaluation of various methods in terms of the communication rounds. We report the sum of communication rounds required to achieve the best performance on each task and the running time of a training communication round." + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 142, + 360, + 473, + 439 + ], + "lines": [ + { + "bbox": [ + 142, + 360, + 473, + 439 + ], + "spans": [ + { + "bbox": [ + 142, + 360, + 473, + 439 + ], + "type": "table", + "html": "
DatasetFedAvgFedCILDANN+FLpFedDIL
RoundsTimeRoundsTimeRoundsTimeRoundsTime
Digit10480(r)32.14(s)492(r)50.06(s)553(r)46.77(s)379(r)34.91(s)
Office31463(r)44.96(s)482(r)78.32(s)488(r)62.89(s)411(r)47.58(s)
DomainNet835(r)92.22(s)904(r)165.41(s)861(r)158.01(s)725(r)105.38(s)
", + "image_path": "d479f493feb1f8d846281dc3d315fdd26d56b6352dee4f83ae9984d39a20c8b0.jpg" + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "table_body" + } + ], + "index": 5 + }, + { + "bbox": [ + 130, + 462, + 482, + 569 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 462, + 482, + 569 + ], + "spans": [ + { + "bbox": [ + 130, + 462, + 482, + 569 + ], + "type": "text", + "content": "Computational complexity and communication cost. We treat all previous layers in ResNet18 before the last fully connected layer as shared layers, and we use two different fully connected layers to get outputs as the auxiliary classifier result and target classification result. Compared with baseline methods, we only train one additional fully connected layer as the auxiliary classifier for the binary classification task - the discrimination of the sample belonging to the current task. The training time cost (each communication round) is shown in Table 4. Therefore, the training of the auxiliary classifier will not increase the overhead or computational complexity significantly and hardly affect the training efficiency." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 130, + 570, + 482, + 641 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 570, + 482, + 641 + ], + "spans": [ + { + "bbox": [ + 130, + 570, + 482, + 641 + ], + "type": "text", + "content": "Table 4 shows the evaluation of various methods in terms of the communication rounds to reach the test accuracy reported in Table 1, 2 and 3 (180 communication rounds each task). Here we show the sum of communication rounds required to achieve the performance on each task. pFedDIL requires the least communication rounds to achieve the reported test accuracy on all datasets with the knowledge matching. Thus, the local model is easier to converge." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 130, + 641, + 482, + 666 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 641, + 482, + 666 + ], + "spans": [ + { + "bbox": [ + 130, + 641, + 482, + 666 + ], + "type": "text", + "content": "Parameter sensitivity analysis. To figure out whether pFedDIL is sensitive to some specific parameters, we first select some general parameters, which are defined in most" + } + ] + } + ], + "index": 8 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "spans": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "type": "text", + "content": "12" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 166, + 91, + 225, + 101 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 166, + 91, + 225, + 101 + ], + "spans": [ + { + "bbox": [ + 166, + 91, + 225, + 101 + ], + "type": "text", + "content": "Yichen Li, et al." + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 11 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 131, + 113, + 248, + 223 + ], + "blocks": [ + { + "bbox": [ + 131, + 113, + 248, + 223 + ], + "lines": [ + { + "bbox": [ + 131, + 113, + 248, + 223 + ], + "spans": [ + { + "bbox": [ + 131, + 113, + 248, + 223 + ], + "type": "image", + "image_path": "c4f1bd2a9d2a59063f1df45c42fdd4cfb47ab0298990d4bf4c67b88ded6c517d.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 130, + 229, + 482, + 262 + ], + "lines": [ + { + "bbox": [ + 130, + 229, + 482, + 262 + ], + "spans": [ + { + "bbox": [ + 130, + 229, + 482, + 262 + ], + "type": "text", + "content": "Fig. 3: Performance of pFedDIL under different configurations. Here, we select three general parameters in the FL setting: (a) local training epoch " + }, + { + "bbox": [ + 130, + 229, + 482, + 262 + ], + "type": "inline_equation", + "content": "E" + }, + { + "bbox": [ + 130, + 229, + 482, + 262 + ], + "type": "text", + "content": ", (b) sample size " + }, + { + "bbox": [ + 130, + 229, + 482, + 262 + ], + "type": "inline_equation", + "content": "B" + }, + { + "bbox": [ + 130, + 229, + 482, + 262 + ], + "type": "text", + "content": " in the classifier, (c) client selection ratio " + }, + { + "bbox": [ + 130, + 229, + 482, + 262 + ], + "type": "inline_equation", + "content": "r" + }, + { + "bbox": [ + 130, + 229, + 482, + 262 + ], + "type": "text", + "content": " of all clients on DomainNet with " + }, + { + "bbox": [ + 130, + 229, + 482, + 262 + ], + "type": "inline_equation", + "content": "\\alpha = 10" + }, + { + "bbox": [ + 130, + 229, + 482, + 262 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "image_caption" + } + ], + "index": 2 + }, + { + "type": "image", + "bbox": [ + 253, + 113, + 365, + 222 + ], + "blocks": [ + { + "bbox": [ + 253, + 113, + 365, + 222 + ], + "lines": [ + { + "bbox": [ + 253, + 113, + 365, + 222 + ], + "spans": [ + { + "bbox": [ + 253, + 113, + 365, + 222 + ], + "type": "image", + "image_path": "d9dcc7a26c79361d656c539b7f6c591d3239425791b64d2c5a619ca75067f725.jpg" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_body" + } + ], + "index": 3 + }, + { + "type": "image", + "bbox": [ + 369, + 113, + 482, + 222 + ], + "blocks": [ + { + "bbox": [ + 369, + 113, + 482, + 222 + ], + "lines": [ + { + "bbox": [ + 369, + 113, + 482, + 222 + ], + "spans": [ + { + "bbox": [ + 369, + 113, + 482, + 222 + ], + "type": "image", + "image_path": "f115331d93e254b7f65dcf2c12690719287aaeab05b8e4474a7546320723e626.jpg" + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "image_body" + } + ], + "index": 4 + }, + { + "bbox": [ + 130, + 288, + 482, + 384 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 288, + 482, + 384 + ], + "spans": [ + { + "bbox": [ + 130, + 288, + 482, + 384 + ], + "type": "text", + "content": "FL: local training epoch " + }, + { + "bbox": [ + 130, + 288, + 482, + 384 + ], + "type": "inline_equation", + "content": "E" + }, + { + "bbox": [ + 130, + 288, + 482, + 384 + ], + "type": "text", + "content": "; sample batch size " + }, + { + "bbox": [ + 130, + 288, + 482, + 384 + ], + "type": "inline_equation", + "content": "B" + }, + { + "bbox": [ + 130, + 288, + 482, + 384 + ], + "type": "text", + "content": ", and client selection ratio for global aggregation " + }, + { + "bbox": [ + 130, + 288, + 482, + 384 + ], + "type": "inline_equation", + "content": "r" + }, + { + "bbox": [ + 130, + 288, + 482, + 384 + ], + "type": "text", + "content": " on DomainNet with " + }, + { + "bbox": [ + 130, + 288, + 482, + 384 + ], + "type": "inline_equation", + "content": "\\alpha = 10" + }, + { + "bbox": [ + 130, + 288, + 482, + 384 + ], + "type": "text", + "content": " to carry out experiments. Figure 3 shows the performance of pFedDIL under different configurations. pFedDIL achieves better results when we increase the local training epochs at the beginning. However, pFedDIL has a comparable performance when the " + }, + { + "bbox": [ + 130, + 288, + 482, + 384 + ], + "type": "inline_equation", + "content": "E" + }, + { + "bbox": [ + 130, + 288, + 482, + 384 + ], + "type": "text", + "content": " is set to 20 and 40. The variance resulting from " + }, + { + "bbox": [ + 130, + 288, + 482, + 384 + ], + "type": "inline_equation", + "content": "E" + }, + { + "bbox": [ + 130, + 288, + 482, + 384 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 130, + 288, + 482, + 384 + ], + "type": "inline_equation", + "content": "r" + }, + { + "bbox": [ + 130, + 288, + 482, + 384 + ], + "type": "text", + "content": " being less than 0.6 and 13 percentage points respectively is comparable to that of existing works and is not significant. In addition, pFedDIL achieves similar performance with different sample sizes " + }, + { + "bbox": [ + 130, + 288, + 482, + 384 + ], + "type": "inline_equation", + "content": "B" + }, + { + "bbox": [ + 130, + 288, + 482, + 384 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 6 + }, + { + "type": "table", + "bbox": [ + 177, + 449, + 434, + 570 + ], + "blocks": [ + { + "bbox": [ + 130, + 406, + 483, + 440 + ], + "lines": [ + { + "bbox": [ + 130, + 406, + 483, + 440 + ], + "spans": [ + { + "bbox": [ + 130, + 406, + 483, + 440 + ], + "type": "text", + "content": "Table 5: Performance of pFedDIL approach on all datasets with different hyper-parameter " + }, + { + "bbox": [ + 130, + 406, + 483, + 440 + ], + "type": "inline_equation", + "content": "\\lambda" + }, + { + "bbox": [ + 130, + 406, + 483, + 440 + ], + "type": "text", + "content": " values. Here we measure average accuracy and model size over all tasks on each client (each model including the auxiliary classifier costs about 22Mb storage)." + } + ] + } + ], + "index": 7, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 177, + 449, + 434, + 570 + ], + "lines": [ + { + "bbox": [ + 177, + 449, + 434, + 570 + ], + "spans": [ + { + "bbox": [ + 177, + 449, + 434, + 570 + ], + "type": "table", + "html": "
Datasetλ = 0.0λ = 0.2λ = 0.5λ = 0.8λ = 1.0
Digit-10Avg ACC82.4385.0289.2890.7793.94
Model Size22Mb37Mb46Mb71Mb88Mb
Office-31Avg ACC33.3540.5139.9846.1750.38
Model Size22Mb42Mb48Mb53Mb66Mb
DomainNetAvg ACC45.3248.1949.6551.6953.75
Model Size22Mb73Mb84Mb91Mb132Mb
", + "image_path": "1b5cb1b508973e33250fe9422cd978834a2010434ede295ad42e339b8445c9eb.jpg" + } + ] + } + ], + "index": 8, + "angle": 0, + "type": "table_body" + } + ], + "index": 8 + }, + { + "bbox": [ + 130, + 594, + 482, + 666 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 594, + 482, + 666 + ], + "spans": [ + { + "bbox": [ + 130, + 594, + 482, + 666 + ], + "type": "text", + "content": "Then, we conduct more research on the setting of hyper-parameter " + }, + { + "bbox": [ + 130, + 594, + 482, + 666 + ], + "type": "inline_equation", + "content": "\\lambda" + }, + { + "bbox": [ + 130, + 594, + 482, + 666 + ], + "type": "text", + "content": ". When " + }, + { + "bbox": [ + 130, + 594, + 482, + 666 + ], + "type": "inline_equation", + "content": "\\lambda = 0.0" + }, + { + "bbox": [ + 130, + 594, + 482, + 666 + ], + "type": "text", + "content": ", pFedDIL will degenerate to the normal FedAvg algorithm without any knowledge migration with the single model. When " + }, + { + "bbox": [ + 130, + 594, + 482, + 666 + ], + "type": "inline_equation", + "content": "\\lambda" + }, + { + "bbox": [ + 130, + 594, + 482, + 666 + ], + "type": "text", + "content": " is set to 1.0, pFedDIL will separately train models for each domain with knowledge migration. As shown in Table 5, we select five different " + }, + { + "bbox": [ + 130, + 594, + 482, + 666 + ], + "type": "inline_equation", + "content": "\\lambda" + }, + { + "bbox": [ + 130, + 594, + 482, + 666 + ], + "type": "text", + "content": " values to conduct experiments and evaluate the average accuracy and average model size on all personalized local models. In Table 1,2, and 3, we utilize ap" + } + ] + } + ], + "index": 9 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 298, + 91, + 448, + 102 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 298, + 91, + 448, + 102 + ], + "spans": [ + { + "bbox": [ + 298, + 91, + 448, + 102 + ], + "type": "text", + "content": "Federated Domain-Incremental Learning" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 471, + 91, + 480, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 471, + 91, + 480, + 100 + ], + "spans": [ + { + "bbox": [ + 471, + 91, + 480, + 100 + ], + "type": "text", + "content": "13" + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 12 + }, + { + "para_blocks": [ + { + "bbox": [ + 130, + 116, + 482, + 236 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 116, + 482, + 236 + ], + "spans": [ + { + "bbox": [ + 130, + 116, + 482, + 236 + ], + "type": "text", + "content": "proximately two auxiliary classifiers per client in Digit-10 and Office-31 and about four in DomainNet. For Table 1,2, we carefully select the optimal value of " + }, + { + "bbox": [ + 130, + 116, + 482, + 236 + ], + "type": "inline_equation", + "content": "\\lambda" + }, + { + "bbox": [ + 130, + 116, + 482, + 236 + ], + "type": "text", + "content": " for each dataset, striking a balance between accuracy and storage. As shown in Table 5, pFedDIL consistently outperforms other baselines even when " + }, + { + "bbox": [ + 130, + 116, + 482, + 236 + ], + "type": "inline_equation", + "content": "\\lambda" + }, + { + "bbox": [ + 130, + 116, + 482, + 236 + ], + "type": "text", + "content": " is simply set to 0.5 across all datasets (around two auxiliary classifiers per client in Digit-10 and Office-31 and approximately three in DomainNet). This suggests that a simple parameter setting can outperform existing methods, and with a more refined parameter setting, we can better accommodate practical considerations, achieving an optimal trade-off between accuracy and storage for each client. The flexibility of pFedDIL allows users to tune this parameter based on their specific requirements and constraints in real-world scenarios." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 132, + 256, + 212, + 269 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 256, + 212, + 269 + ], + "spans": [ + { + "bbox": [ + 132, + 256, + 212, + 269 + ], + "type": "text", + "content": "5 Limitations" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 130, + 283, + 482, + 392 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 283, + 482, + 392 + ], + "spans": [ + { + "bbox": [ + 130, + 283, + 482, + 392 + ], + "type": "text", + "content": "Our work is aimed at solving the forgetting problem under the assumption of shared knowledge between tasks in the domain incremental setting. The problem of learning tasks without similar knowledge is rare in real-world scenarios and violates the principle of incremental learning [11]. The ultimate goal of incremental learning is to better learn new knowledge with the assistance of previous knowledge while also mastering both previous and new knowledge [39]. Furthermore, we investigate the situation where the previous data is strictly unseen to clients, while some research on federated incremental learning assumes that each client can cache some samples for replay [9]. Both situations exist and there is a must to separately develop appropriate methods." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 132, + 411, + 302, + 424 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 411, + 302, + 424 + ], + "spans": [ + { + "bbox": [ + 132, + 411, + 302, + 424 + ], + "type": "text", + "content": "6 Conclusion and Future Work" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 130, + 438, + 482, + 582 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 438, + 482, + 582 + ], + "spans": [ + { + "bbox": [ + 130, + 438, + 482, + 582 + ], + "type": "text", + "content": "In this paper, we seek to tackle the catastrophic forgetting in the FDIL scenario. We propose a personalized federated domain-incremental learning approach based on adaptive knowledge matching, named pFedDIL. We leverage the auxiliary classifier to calculate the knowledge-matching intensity for the incremental task-learning strategy selection and knowledge migration. Furthermore, we propose sharing partial parameters between the target classification model and the auxiliary classifier to condense model parameters. Extensive experiments conducted on various datasets and settings show that our method achieves significant improvement in accuracy. Although existing works and our method have demonstrated great effectiveness over the FCIL and FDIL scenarios separately, none of them take both scenarios into account. To deploy the FL system in practical settings, it is necessary to simultaneously handle the class incremental tasks and domain incremental tasks. In the future, we seek to move forward in this field." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 132, + 602, + 235, + 616 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 602, + 235, + 616 + ], + "spans": [ + { + "bbox": [ + 132, + 602, + 235, + 616 + ], + "type": "text", + "content": "Acknowledgements" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 130, + 629, + 481, + 665 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 629, + 481, + 665 + ], + "spans": [ + { + "bbox": [ + 130, + 629, + 481, + 665 + ], + "type": "text", + "content": "This work is supported by National Natural Science Foundation of China under grants 62376103, 62302184, 62206102, Science and Technology Support Program of Hubei Province under grant 2022BAA046." + } + ] + } + ], + "index": 8 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "spans": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "type": "text", + "content": "14" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 166, + 91, + 225, + 101 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 166, + 91, + 225, + 101 + ], + "spans": [ + { + "bbox": [ + 166, + 91, + 225, + 101 + ], + "type": "text", + "content": "Yichen Li, et al." + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 13 + }, + { + "para_blocks": [ + { + "bbox": [ + 133, + 114, + 190, + 127 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 114, + 190, + 127 + ], + "spans": [ + { + "bbox": [ + 133, + 114, + 190, + 127 + ], + "type": "text", + "content": "References" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 138, + 141, + 481, + 665 + ], + "type": "list", + "angle": 0, + "index": 18, + "blocks": [ + { + "bbox": [ + 138, + 141, + 481, + 175 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 141, + 481, + 175 + ], + "spans": [ + { + "bbox": [ + 138, + 141, + 481, + 175 + ], + "type": "text", + "content": "1. Bakman, Y.F., Yaldiz, D.N., Ezzeldin, Y.H., Avestimehr, S.: Federated orthogonal training: Mitigating global catastrophic forgetting in continual federated learning. arXiv preprint arXiv:2309.01289 (2023) 1, 2" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 138, + 176, + 481, + 208 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 176, + 481, + 208 + ], + "spans": [ + { + "bbox": [ + 138, + 176, + 481, + 208 + ], + "type": "text", + "content": "2. Chen, H.J., Cheng, A.C., Juan, D.C., Wei, W., Sun, M.: Mitigating forgetting in online continual learning via instance-aware parameterization. Advances in Neural Information Processing Systems 33, 17466-17477 (2020) 4" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 138, + 209, + 481, + 241 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 209, + 481, + 241 + ], + "spans": [ + { + "bbox": [ + 138, + 209, + 481, + 241 + ], + "type": "text", + "content": "3. Cohen, G., Afshar, S., Tapson, J., Van Schaik, A.: Emmist: Extending mnist to handwritten letters. In: 2017 international joint conference on neural networks (IJCNN). pp. 2921-2926. IEEE (2017) 9" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 138, + 243, + 481, + 285 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 243, + 481, + 285 + ], + "spans": [ + { + "bbox": [ + 138, + 243, + 481, + 285 + ], + "type": "text", + "content": "4. Dong, J., Li, H., Cong, Y., Sun, G., Zhang, Y., Van Gool, L.: No one left behind: Real-world federated class-incremental learning. IEEE Transactions on Pattern Analysis and Machine Intelligence 46(4), 2054-2070 (2024). https://doi.org/10.1109/TPAMI.2023.33342135" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 138, + 286, + 481, + 319 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 286, + 481, + 319 + ], + "spans": [ + { + "bbox": [ + 138, + 286, + 481, + 319 + ], + "type": "text", + "content": "5. Dong, J., Liang, W., Cong, Y., Sun, G.: Heterogeneous forgetting compensation for class incremental learning. In: Proceedings of the IEEE/CVF International Conference on Computer Vision (ICCV). pp. 11742-11751 (October 2023) 5" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 138, + 320, + 481, + 353 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 320, + 481, + 353 + ], + "spans": [ + { + "bbox": [ + 138, + 320, + 481, + 353 + ], + "type": "text", + "content": "6. Dong, J., Wang, L., Fang, Z., Sun, G., Xu, S., Wang, X., Zhu, Q.: Federated class-incremental learning. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 10164-10173 (2022) 2" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 138, + 354, + 481, + 396 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 354, + 481, + 396 + ], + "spans": [ + { + "bbox": [ + 138, + 354, + 481, + 396 + ], + "type": "text", + "content": "7. Dong, J., Zhang, D., Cong, Y., Cong, W., Ding, H., Dai, D.: Federated incremental semantic segmentation. In: 2023 IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR). pp. 3934-3943 (2023). https://doi.org/10.1109/CVPR52729.2023.003835" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 138, + 397, + 481, + 420 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 397, + 481, + 420 + ], + "spans": [ + { + "bbox": [ + 138, + 397, + 481, + 420 + ], + "type": "text", + "content": "8. Dong, X., Yu, Z., Cao, W., Shi, Y., Ma, Q.: A survey on ensemble learning. Frontiers of Computer Science 14, 241-258 (2020) 8" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 138, + 421, + 481, + 464 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 421, + 481, + 464 + ], + "spans": [ + { + "bbox": [ + 138, + 421, + 481, + 464 + ], + "type": "text", + "content": "9. Fini, E., Lathuiliere, S., Sangineto, E., Nabi, M., Ricci, E.: Online continual learning under extreme memory constraints. In: Computer Vision-ECCV 2020: 16th European Conference, Glasgow, UK, August 23-28, 2020, Proceedings, Part XXVIII 16. pp. 720-735. Springer (2020) 14" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 138, + 465, + 481, + 498 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 465, + 481, + 498 + ], + "spans": [ + { + "bbox": [ + 138, + 465, + 481, + 498 + ], + "type": "text", + "content": "0. Ganin, Y., Ustinova, E., Ajakan, H., Germain, P., Larochelle, H., Laviolette, F., Marchand, M., Lempitsky, V.: Domain-adversarial training of neural networks. The journal of machine learning research 17(1), 2096-2030 (2016) 9, 10, 11" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 138, + 498, + 481, + 520 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 498, + 481, + 520 + ], + "spans": [ + { + "bbox": [ + 138, + 498, + 481, + 520 + ], + "type": "text", + "content": "1. Hadsell, R., Rao, D., Rusu, A.A., Pascanu, R.: Embracing change: Continual learning in deep neural networks. Trends in cognitive sciences 24(12), 1028-1040 (2020) 14" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 138, + 521, + 481, + 553 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 521, + 481, + 553 + ], + "spans": [ + { + "bbox": [ + 138, + 521, + 481, + 553 + ], + "type": "text", + "content": "2. He, K., Zhang, X., Ren, S., Sun, J.: Deep residual learning for image recognition. In: Proceedings of the IEEE conference on computer vision and pattern recognition. pp. 770-778 (2016) 10" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 138, + 555, + 481, + 576 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 555, + 481, + 576 + ], + "spans": [ + { + "bbox": [ + 138, + 555, + 481, + 576 + ], + "type": "text", + "content": "3. Hull, J.J.: A database for handwritten text recognition research. IEEE Transactions on pattern analysis and machine intelligence 16(5), 550-554 (1994) 9" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 138, + 577, + 481, + 609 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 577, + 481, + 609 + ], + "spans": [ + { + "bbox": [ + 138, + 577, + 481, + 609 + ], + "type": "text", + "content": "4. Jeong, E., Oh, S., Kim, H., Park, J., Bennis, M., Kim, S.L.: Communication-efficient on-device machine learning: Federated distillation and augmentation under non-id private data. ArXiv abs/1811.11479 (2018) 4" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 138, + 611, + 481, + 665 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 611, + 481, + 665 + ], + "spans": [ + { + "bbox": [ + 138, + 611, + 481, + 665 + ], + "type": "text", + "content": "5. Jiang, Z., Ren, Y., Lei, M., Zhao, Z.: Fedspeech: Federated text-to-speech with continual learning. In: Proceedings of the Thirtieth International Joint Conference on Artificial Intelligence. IJCAI-2021, International Joint Conferences on Artificial Intelligence Organization (Aug 2021). https://doi.org/10.24963/ijcai.2021/527, http://dx.doi.org/10.24963/ijcai.2021/527 5" + } + ] + } + ], + "index": 17 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 298, + 91, + 447, + 102 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 298, + 91, + 447, + 102 + ], + "spans": [ + { + "bbox": [ + 298, + 91, + 447, + 102 + ], + "type": "text", + "content": "Federated Domain-Incremental Learning" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 471, + 91, + 480, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 471, + 91, + 480, + 100 + ], + "spans": [ + { + "bbox": [ + 471, + 91, + 480, + 100 + ], + "type": "text", + "content": "15" + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 14 + }, + { + "para_blocks": [ + { + "bbox": [ + 132, + 117, + 480, + 665 + ], + "type": "list", + "angle": 0, + "index": 21, + "blocks": [ + { + "bbox": [ + 133, + 117, + 480, + 138 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 117, + 480, + 138 + ], + "spans": [ + { + "bbox": [ + 133, + 117, + 480, + 138 + ], + "type": "text", + "content": "16. Jung, S., Ahn, H., Cha, S., Moon, T.: Continual learning with node-importance based adaptive group sparse regularization. arXiv: Learning (2020) 4" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 133, + 139, + 427, + 148 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 139, + 427, + 148 + ], + "spans": [ + { + "bbox": [ + 133, + 139, + 427, + 148 + ], + "type": "text", + "content": "17. LeCun, Y., Cortes, C., Burges, C.: Mnist handwritten digit database (2010) 9" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 133, + 149, + 480, + 170 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 149, + 480, + 170 + ], + "spans": [ + { + "bbox": [ + 133, + 149, + 480, + 170 + ], + "type": "text", + "content": "18. Li, L., Fan, Y., Tse, M., Lin, K.Y.: A review of applications in federated learning. Computers & Industrial Engineering 149, 106854 (2020) 4" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 133, + 171, + 480, + 191 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 171, + 480, + 191 + ], + "spans": [ + { + "bbox": [ + 133, + 171, + 480, + 191 + ], + "type": "text", + "content": "19. Li, T., Sahu, A.K., Talwalkar, A., Smith, V.: Federated learning: Challenges, methods, and future directions. IEEE signal processing magazine 37(3), 50-60 (2020) 4, 9, 10, 11" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 132, + 192, + 480, + 224 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 192, + 480, + 224 + ], + "spans": [ + { + "bbox": [ + 132, + 192, + 480, + 224 + ], + "type": "text", + "content": "20. Li, Y., Li, Q., Wang, H., Li, R., Zhong, W., Zhang, G.: Towards efficient replay in federated incremental learning. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR). pp. 12820-12829 (June 2024) 2" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 132, + 224, + 480, + 255 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 224, + 480, + 255 + ], + "spans": [ + { + "bbox": [ + 132, + 224, + 480, + 255 + ], + "type": "text", + "content": "21. Liu, Y., Tian, X., Li, Y., Xiong, Z., Wu, F.: Compact feature learning for multi-domain image classification. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 7193-7201 (2019) 2" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 132, + 256, + 480, + 319 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 256, + 480, + 319 + ], + "spans": [ + { + "bbox": [ + 132, + 256, + 480, + 319 + ], + "type": "text", + "content": "22. Ma, Y., Xie, Z., Wang, J., Chen, K., Shou, L.: Continual federated learning based on knowledge distillation. In: Raedt, L.D. (ed.) Proceedings of the Thirty-First International Joint Conference on Artificial Intelligence, IJCAI-22. pp. 2182-2188. International Joint Conferences on Artificial Intelligence Organization (7 2022). https://doi.org/10.24963/ijcai.2022/303, https://doi.org/10.24963/ijcai.2022/303, main Track 2" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 132, + 321, + 480, + 352 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 321, + 480, + 352 + ], + "spans": [ + { + "bbox": [ + 132, + 321, + 480, + 352 + ], + "type": "text", + "content": "23. Maltoni, D., Lomonaco, V.: Continuous learning in single-incremental-task scenarios. Neural networks : the official journal of the International Neural Network Society 116, 56-73 (2018) 4" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 132, + 354, + 480, + 386 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 354, + 480, + 386 + ], + "spans": [ + { + "bbox": [ + 132, + 354, + 480, + 386 + ], + "type": "text", + "content": "24. McMahan, B., Moore, E., Ramage, D., Hampson, S., y Arcas, B.A.: Communication-efficient learning of deep networks from decentralized data. In: Artificial intelligence and statistics. pp. 1273-1282. PMLR (2017) 1, 4, 9, 10, 11" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 132, + 386, + 480, + 418 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 386, + 480, + 418 + ], + "spans": [ + { + "bbox": [ + 132, + 386, + 480, + 418 + ], + "type": "text", + "content": "25. Mirza, M.J., Masana, M., Possegger, H., Bischof, H.: An efficient domain-incremental learning approach to drive in all weather conditions. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 3001-3011 (2022) 4, 9" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 132, + 418, + 480, + 439 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 418, + 480, + 439 + ], + "spans": [ + { + "bbox": [ + 132, + 418, + 480, + 439 + ], + "type": "text", + "content": "26. Netzer, Y., Wang, T., Coates, A., Bissacco, A., Wu, B., Ng, A.Y.: Reading digits in natural images with unsupervised feature learning (2011) 9" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 132, + 440, + 480, + 472 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 440, + 480, + 472 + ], + "spans": [ + { + "bbox": [ + 132, + 440, + 480, + 472 + ], + "type": "text", + "content": "27. Peng, X., Bai, Q., Xia, X., Huang, Z., Saenko, K., Wang, B.: Moment matching for multi-source domain adaptation. In: Proceedings of the IEEE/CVF international conference on computer vision. pp. 1406-1415 (2019) 9" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 132, + 472, + 480, + 493 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 472, + 480, + 493 + ], + "spans": [ + { + "bbox": [ + 132, + 472, + 480, + 493 + ], + "type": "text", + "content": "28. Qi, D., Zhao, H., Li, S.: Better generative replay for continual federated learning. arXiv preprint arXiv:2302.13001 (2023) 2, 9, 10, 11" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 132, + 494, + 480, + 525 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 494, + 480, + 525 + ], + "spans": [ + { + "bbox": [ + 132, + 494, + 480, + 525 + ], + "type": "text", + "content": "29. Rebuffi, S.A., Kolesnikov, A., Sperl, G., Lampert, C.H.: Incremental classifier and representation learning. In: Proceedings of the IEEE conference on Computer Vision and Pattern Recognition. pp. 2001-2010 (2017) 4" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 132, + 525, + 480, + 546 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 525, + 480, + 546 + ], + "spans": [ + { + "bbox": [ + 132, + 525, + 480, + 546 + ], + "type": "text", + "content": "30. Rolnick, D., Ahuja, A., Schwarz, J., Lillicrap, T., Wayne, G.: Experience replay for continual learning. Advances in Neural Information Processing Systems 32 (2019) 4" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 132, + 547, + 480, + 567 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 547, + 480, + 567 + ], + "spans": [ + { + "bbox": [ + 132, + 547, + 480, + 567 + ], + "type": "text", + "content": "31. Ruder, S.: An overview of multi-task learning in deep neural networks. arXiv preprint arXiv:1706.05098 (2017) 9" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 132, + 568, + 480, + 611 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 568, + 480, + 611 + ], + "spans": [ + { + "bbox": [ + 132, + 568, + 480, + 611 + ], + "type": "text", + "content": "32. Saenko, K., Kulis, B., Fritz, M., Darrell, T.: Adapting visual category models to new domains. In: Computer Vision-ECCV 2010: 11th European Conference on Computer Vision, Heraklion, Crete, Greece, September 5-11, 2010, Proceedings, Part IV 11. pp. 213-226. Springer (2010) 9" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 132, + 612, + 480, + 632 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 612, + 480, + 632 + ], + "spans": [ + { + "bbox": [ + 132, + 612, + 480, + 632 + ], + "type": "text", + "content": "33. Shen, T., Zhang, J., Jia, X., Zhang, F., Huang, G., Zhou, P., Wu, F., Wu, C.: Federated mutual learning. ArXiv abs/2006.16765 (2020) 4" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 132, + 633, + 480, + 665 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 633, + 480, + 665 + ], + "spans": [ + { + "bbox": [ + 132, + 633, + 480, + 665 + ], + "type": "text", + "content": "34. Standley, T., Zamir, A., Chen, D., Guibas, L., Malik, J., Savarese, S.: Which tasks should be learned together in multi-task learning? In: International Conference on Machine Learning. pp. 9120-9132. PMLR (2020) 8" + } + ] + } + ], + "index": 20 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "spans": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "type": "text", + "content": "16" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 166, + 91, + 225, + 101 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 166, + 91, + 225, + 101 + ], + "spans": [ + { + "bbox": [ + 166, + 91, + 225, + 101 + ], + "type": "text", + "content": "Yichen Li, et al." + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 15 + }, + { + "para_blocks": [ + { + "bbox": [ + 132, + 116, + 481, + 390 + ], + "type": "list", + "angle": 0, + "index": 12, + "blocks": [ + { + "bbox": [ + 132, + 116, + 480, + 127 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 116, + 480, + 127 + ], + "spans": [ + { + "bbox": [ + 132, + 116, + 480, + 127 + ], + "type": "text", + "content": "35. Tankard, C.: What the gdpr means for businesses. Network Security 2016(6), 5-8 (2016) 1" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 132, + 128, + 481, + 148 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 128, + 481, + 148 + ], + "spans": [ + { + "bbox": [ + 132, + 128, + 481, + 148 + ], + "type": "text", + "content": "36. Van de Ven, G.M., Tuytelaars, T., Tolias, A.S.: Three types of incremental learning. Nature Machine Intelligence 4(12), 1185-1197 (2022) 1" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 132, + 149, + 481, + 182 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 149, + 481, + 182 + ], + "spans": [ + { + "bbox": [ + 132, + 149, + 481, + 182 + ], + "type": "text", + "content": "37. Wang, H., Li, Y., Xu, W., Li, R., Zhan, Y., Zeng, Z.: Dafkd: Domain-aware federated knowledge distillation. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 20412-20421 (2023) 4" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 132, + 182, + 481, + 215 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 182, + 481, + 215 + ], + "spans": [ + { + "bbox": [ + 132, + 182, + 481, + 215 + ], + "type": "text", + "content": "38. Wang, H., Xu, H., Li, Y., Xu, Y., Li, R., Zhang, T.: Fedcda: Federated learning with cross-rounds divergence-aware aggregation. In: The Twelfth International Conference on Learning Representations (2023) 4" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 132, + 216, + 481, + 247 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 216, + 481, + 247 + ], + "spans": [ + { + "bbox": [ + 132, + 216, + 481, + 247 + ], + "type": "text", + "content": "39. Wang, L., Zhang, X., Su, H., Zhu, J.: A comprehensive survey of continual learning: theory, method and application. IEEE Transactions on Pattern Analysis and Machine Intelligence (2024) 4, 14" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 132, + 248, + 481, + 281 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 248, + 481, + 281 + ], + "spans": [ + { + "bbox": [ + 132, + 248, + 481, + 281 + ], + "type": "text", + "content": "40. Yoon, J., Jeong, W., Lee, G., Yang, E., Hwang, S.J.: Federated continual learning with weighted inter-client transfer. In: International Conference on Machine Learning. pp. 12073-12086. PMLR (2021) 1, 2, 5" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 132, + 281, + 481, + 304 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 281, + 481, + 304 + ], + "spans": [ + { + "bbox": [ + 132, + 281, + 481, + 304 + ], + "type": "text", + "content": "41. Yu, H., Yang, X., Gao, X., Kang, Y., Wang, H., Zhang, J., Li, T.: Personalized federated continual learning via multi-granularity prompt. arXiv preprint arXiv:2407.00113 (2024) 5" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 132, + 304, + 481, + 325 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 304, + 481, + 325 + ], + "spans": [ + { + "bbox": [ + 132, + 304, + 481, + 325 + ], + "type": "text", + "content": "42. Zhang, C., Xie, Y., Bai, H., Yu, B., Li, W., Gao, Y.: A survey on federated learning. Knowledge-Based Systems 216, 106775 (2021) 1" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 132, + 325, + 481, + 357 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 325, + 481, + 357 + ], + "spans": [ + { + "bbox": [ + 132, + 325, + 481, + 357 + ], + "type": "text", + "content": "43. Zhang, J., Chen, C., Zhuang, W., Lyu, L.: Target: Federated class-continual learning via exemplar-free distillation. In: Proceedings of the IEEE/CVF International Conference on Computer Vision. pp. 4782-4793 (2023) 1" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 132, + 358, + 481, + 390 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 358, + 481, + 390 + ], + "spans": [ + { + "bbox": [ + 132, + 358, + 481, + 390 + ], + "type": "text", + "content": "44. Zhao, H., Zhou, T., Long, G., Jiang, J., Zhang, C.: Does continual learning equally forget all parameters? In: International Conference on Machine Learning. pp. 42280-42303. PMLR (2023) 4" + } + ] + } + ], + "index": 11 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 298, + 91, + 447, + 102 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 298, + 91, + 447, + 102 + ], + "spans": [ + { + "bbox": [ + 298, + 91, + 447, + 102 + ], + "type": "text", + "content": "Federated Domain-Incremental Learning" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 470, + 91, + 481, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 470, + 91, + 481, + 100 + ], + "spans": [ + { + "bbox": [ + 470, + 91, + 481, + 100 + ], + "type": "text", + "content": "17" + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 16 + } + ], + "_backend": "vlm", + "_version_name": "2.6.4" +} \ No newline at end of file diff --git a/2024/Personalized Privacy Protection Mask Against Unauthorized Facial Recognition/40580728-0a4c-4c74-a026-5407ddf870b6_content_list.json b/2024/Personalized Privacy Protection Mask Against Unauthorized Facial Recognition/40580728-0a4c-4c74-a026-5407ddf870b6_content_list.json new file mode 100644 index 0000000000000000000000000000000000000000..568eaba794d937a9024ad716690db211a26f22f1 --- /dev/null +++ b/2024/Personalized Privacy Protection Mask Against Unauthorized Facial Recognition/40580728-0a4c-4c74-a026-5407ddf870b6_content_list.json @@ -0,0 +1,1706 @@ +[ + { + "type": "text", + "text": "Personalized Privacy Protection Mask Against Unauthorized Facial Recognition", + "text_level": 1, + "bbox": [ + 235, + 140, + 767, + 186 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Ka-Ho Chow $^{1}$ , Sihao Hu $^{2}$ , Tiansheng Huang $^{2}$ , and Ling Liu $^{2}$", + "bbox": [ + 256, + 210, + 743, + 227 + ], + "page_idx": 0 + }, + { + "type": "list", + "sub_type": "text", + "list_items": [ + "1 The University of Hong Kong", + "2 Georgia Institute of Technology" + ], + "bbox": [ + 388, + 239, + 614, + 267 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "kachow@cs.hku.hk, {sihaohu,thuang374}@gatech.edu, lingliu@cc.gatech.edu", + "bbox": [ + 228, + 268, + 774, + 281 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Abstract. Face recognition (FR) can be abused for privacy intrusion. Governments, private companies, or even individual attackers can collect facial images by web scraping to build an FR system identifying human faces without their consent. This paper introduces Chameleon, which learns to generate a user-centric personalized privacy protection mask, coined as P3-Mask, to protect facial images against unauthorized FR with three salient features. First, we use a cross-image optimization to generate one P3-Mask for each user instead of tailoring facial perturbation for each facial image of a user. It enables efficient and instant protection even for users with limited computing resources. Second, we incorporate a perceptibility optimization to preserve the visual quality of the protected facial images. Third, we strengthen the robustness of P3-Mask against unknown FR models by integrating focal diversity-optimized ensemble learning into the mask generation process. Extensive experiments on two benchmark datasets show that Chameleon outperforms three state-of-the-art methods with instant protection and minimal degradation of image quality. Furthermore, Chameleon enables cost-effective FR authorization using the P3-Mask as a personalized de-obfuscation key, and it demonstrates high resilience against adaptive adversaries.", + "bbox": [ + 259, + 311, + 740, + 575 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "1 Introduction", + "text_level": 1, + "bbox": [ + 217, + 599, + 375, + 614 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Face recognition (FR) has long been investigated due to its potential for enhancing security and convenience in various domains [22,28,29]. Many pretrained FR models are available online [6,7]. Once a face database (a.k.a. the gallery) with facial images for each person of interest is provided, these pretrained models can be used to recognize them [34].", + "bbox": [ + 212, + 628, + 787, + 704 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "While empowering many life-enriching applications, FR can be abused to cause serious privacy issues [2]. Privacy intruders can build a face database of victims of interest from publicly available facial images on the Internet by web scraping (Figure 1 (top)). By utilizing this database, the adversary can perform unauthorized FR of individual users for stalking victims [3], intruding on victims' privacy by flooding targeted ads [1], or facilitating criminals to commit identity fraud [5]. This is a real threat. For example, companies like Clearview [4] and PimEyes [8] have collected billions of online images and can recognize millions of citizens without their consent.", + "bbox": [ + 212, + 704, + 787, + 839 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/1173bc9750a62a9f482119b53580730f40508b1fd6750a03605ce1312da9b1fe.jpg", + "image_caption": [ + "Fig. 1: (Top) Without protection, the privacy intruder can build a face database by web scraping and identify the unknown face. (Bottom) Chameleon learns the facial signature of the user (protectee) to generate a P3-Mask, which can be applied to protect any facial images before sharing them online against unauthorized FR." + ], + "image_footnote": [], + "bbox": [ + 267, + 142, + 743, + 321 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Such a facial privacy threat has motivated the development of anti-FR technologies [36], which allow users (protectees) to preprocess their facial images before sharing them online. We argue that a competitive anti-FR solution should possess the following four properties. First, from the robustness perspective, the protected facial images should not be matched by an FR model to a query (probe) image with a correct identity (e.g., Bradley Cooper is recognized as Anne Hathaway in Figure 1 (bottom)), including those FR models possibly unknown during the preprocessing but are used by the privacy intruder. Second, from the efficiency perspective, the protection should be done instantly on any device with or without AI accelerators to maintain user experience and support equitable access to privacy protection. Third, from the perception usability perspective, the protected image should (i) preserve the visual quality rather than using excessive noise and (ii) be visually recognizable by humans to be the same person as the one in the original unprotected photo. Finally, the service usability for authorized FR providers is another important property. Existing approaches treat all FR models as unauthorized. We argue that anti-FR solutions should also allow users to grant FR permission to authorized models cost-effectively. This user-initiated de-obfuscation capability is critical to those, e.g., who publish their photos on a social media platform with privacy protection and, at the same time, wish to authorize the platform to perform some FR services (e.g., face tagging [32]) without sending and storing multiple versions of the same photo.", + "bbox": [ + 212, + 416, + 787, + 734 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Based on the above objectives, we introduce Chameleon, a user-centric facial privacy protection system with three contributions. First, we develop an optimization algorithm to construct a Personalized Privacy Protection mask, coined as P3-Mask, for each user. The P3-Mask of a user can be used to protect any facial images of the same user with minimal impact on image quality, including those facial images unseen during the P3-Mask generation process. Second, it boosts the robustness of P3-Mask against unknown FR models by a princi", + "bbox": [ + 212, + 734, + 787, + 840 + ], + "page_idx": 1 + }, + { + "type": "page_number", + "text": "2", + "bbox": [ + 217, + 114, + 228, + 126 + ], + "page_idx": 1 + }, + { + "type": "header", + "text": "K.-H. Chow et al.", + "bbox": [ + 271, + 114, + 392, + 127 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "pled approach, leveraging ensemble learning with models of high focal diversity. The per-user P3-Mask provides higher robustness against unknown FR models while preserving perception usability and the service usability for authorized FR providers. Chameleon users can utilize their own P3-Mask to obfuscate their facial images prior to public release, and the same P3-Mask can be used as a personalized de-obfuscation key to grant authorized FR service providers the ability to restore the facial signature of photos for correct recognition. Third, we conduct extensive experiments on two FR benchmark datasets [20,26] to analyze the shortcomings of state-of-the-art anti-FR solutions [31,39,40] and show that Chameleon can better protect against unknown FR models with a high success rate. The protection is lightweight and in real-time. It remains effective under adaptive privacy intruders deploying various strategies to counter Chameleon.", + "bbox": [ + 212, + 146, + 787, + 328 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "2 Related Work", + "text_level": 1, + "bbox": [ + 215, + 349, + 387, + 364 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Existing anti-FR approaches can be broadly classified into two categories. Synthesis-based approaches [11,16,19,21,33,41] use generative adversarial networks (GANs) [25] to synthesize a face to replace the original one for protection. While the synthetic faces look realistic, they appear to be strangers, not recognizable even by the users themselves, failing to meet the perception usability requirement.", + "bbox": [ + 212, + 378, + 815, + 455 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Chameleon falls into the second category, which applies small changes to facial images [12]. Fawkes [31] formulates an untargeted attack to push the facial image away from the original location in the embedding space; TIP-IM [39] improves the image quality with MMD [9]; LowKey [10] improves the robustness under the image processing pipeline in an end-to-end ML system by incorporating it into the optimization process. These techniques can preserve the visual identity of protected faces. However, they both require iterative optimization for each image. Even for a GPU server, it can take over 100 seconds to perturb one facial image (Section 6.3). In contrast, Chameleon provides instant protection on any device, with protection even stronger than those spending minutes to find the best perturbation for each image. OPOM [40] attempts to improve efficiency by finding the embedding subspace enclosing facial images of the same person and generates a privacy protection mask to push any images away from it. However, it fails to maintain protection effectiveness when image processing operations are applied to protected images (Section 6.1), and the mask degrades the image quality much more significantly than Chameleon (Section 6.3).", + "bbox": [ + 212, + 455, + 787, + 698 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "3 Overview", + "text_level": 1, + "bbox": [ + 215, + 719, + 344, + 734 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "To build an FR system with a pretrained model, the owner first collects a gallery dataset (face database) $\\pmb{D}$ , which includes the facial images of individuals of interest. Then, the pretrained FR model $F$ is used to map each gallery image $\\tilde{\\pmb{x}} \\in \\pmb{D}$ to the embedding $F(\\tilde{\\pmb{x}})$ , and those embeddings should form clusters corresponding to different people. When a facial image $\\pmb{x}$ of an unknown identity, called the probe image, is given, the FR system can use the same FR model $F$ to", + "bbox": [ + 212, + 750, + 787, + 840 + ], + "page_idx": 2 + }, + { + "type": "header", + "text": "Personalized Privacy Protection Mask Against Unauthorized FR", + "bbox": [ + 300, + 114, + 730, + 128 + ], + "page_idx": 2 + }, + { + "type": "page_number", + "text": "3", + "bbox": [ + 774, + 116, + 784, + 126 + ], + "page_idx": 2 + }, + { + "type": "image", + "img_path": "images/47589bdc1be02a49d4c4bac2fbfdf6c5df0a7220694b62be80e651b98c27e757.jpg", + "image_caption": [ + "Fig. 2: An overview of Chameleon's two-phase workflow for offline learning to generate P3-Mask and online protection with P3-Mask for personalized facial signature masking. The P3-Mask can protect any facial images of the same person without further learning." + ], + "image_footnote": [], + "bbox": [ + 217, + 143, + 784, + 258 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "map it to an embedding $F(\\pmb{x})$ and use the identity of the nearest gallery image to be the identity of the unknown person. The identity, $\\mathcal{FR}(\\pmb{x};F,\\pmb{D})$ , of the probe image $\\pmb{x}$ given the FR model $F$ and the gallery dataset $\\pmb{D}$ is defined as:", + "bbox": [ + 214, + 344, + 787, + 390 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\n\\mathcal {F R} (\\boldsymbol {x}; F, \\boldsymbol {\\mathcal {D}}) = \\mathcal {I} \\left(\\underset {\\tilde {\\boldsymbol {x}} \\in \\boldsymbol {\\mathcal {D}}} {\\arg \\min } \\operatorname {D I S T} (F (\\boldsymbol {x}), F (\\tilde {\\boldsymbol {x}}))\\right), \\tag {1}\n$$\n", + "text_format": "latex", + "bbox": [ + 346, + 398, + 785, + 422 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "where $\\mathcal{I}(\\tilde{\\boldsymbol{x}})$ denotes the identity of the gallery image $\\tilde{\\boldsymbol{x}}$ , which is known to the FR system, and $\\mathrm{DIST}(\\cdot ,\\cdot)$ is a distance function such as Euclidean distance.", + "bbox": [ + 214, + 430, + 787, + 460 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Threat Model. We consider a threat model where a privacy intruder conducts web scraping to collect images containing citizens' faces. Web scraping is large-scale and untargeted. Many citizens are included in the face database, and the privacy intruder may not know them. Given a probe image taken by, e.g., a stalker's camera, the privacy intruder uses an FR model on the face database to search for matched image(s). Then, the privacy intruder can analyze the associated metadata, such as the web page from where it was downloaded or even the exact identity. The goal of Chameleon is to allow the user to preprocess her facial images before posting them online. Even if scraped and included in the face database, they will not be matched to a probe image of her.", + "bbox": [ + 214, + 460, + 787, + 612 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "3.1 Chameleon Design", + "text_level": 1, + "bbox": [ + 215, + 633, + 416, + 648 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Chameleon performs preprocessing of the user's photos before sharing them online by applying the user-specific P3-Mask. This mask is generated offline for a Chameleon user. Figure 2 provides an overview of Chameleon's workflow.", + "bbox": [ + 214, + 657, + 787, + 703 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Offline Learning Phase. Chameleon learns the unique facial signature of a user from a few facial images of her using the P3-Mask generator (Section 4), as shown in Figure 2 ①. P3-Mask is designed to apply directly to any facial images of the same user, including those unseen during offline learning, and is robust against lossy image processing operations in an end-to-end ML system without compromising the image quality. We use a focal diversity-optimized ensemble learning method to find a team of FR models such that the P3-Mask generated against them has strong robustness and is effective in countering other FR models that are unknown during offline learning.", + "bbox": [ + 214, + 704, + 787, + 840 + ], + "page_idx": 3 + }, + { + "type": "page_number", + "text": "4", + "bbox": [ + 217, + 114, + 228, + 126 + ], + "page_idx": 3 + }, + { + "type": "header", + "text": "K.-H. Chow et al.", + "bbox": [ + 271, + 114, + 392, + 127 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Online Protection Phase. After the offline learning, ② the P3-Mask can be sent to the user's local device (e.g., the mobile phone). For a photo of the user to be protected, ③ a lightweight face detection model, such as MediaPipe [23], can be used by the Chameleon user running on the user's device to locate the face region and ④ instantly protect it with the P3-Mask before the user shares it online ⑤. The user can also authorize some trusted entities to conduct FR on their shared photos protected with P3-Mask. The user-specific de-obfuscation key ⑥ allows the protected photos to be restored for authorized FR.", + "bbox": [ + 212, + 146, + 787, + 268 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "4 P3-Mask Generation", + "text_level": 1, + "bbox": [ + 215, + 289, + 455, + 306 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "For a user $\\mathcal{P}$ , Chameleon generates a P3-Mask by learning the facial signature $\\mathcal{M}_{\\mathcal{P}}$ from a set of facial images the user provides. Several key factors need to be accomplished. (1) We need to promote cross-image protection as an optimization objective to generate the most representative facial signature applicable to protect any facial images of the user $\\mathcal{P}$ , including those unknown ones during the generation process (Section 4.1). (2) We need to control the amount of perturbation introduced by the P3-Mask. When applied to an unprotected image by removing the learned facial signature, it ensures minimal perception loss to preserve the visual quality (Section 4.2). (3) We need to keep the protected image far away from the original image in the embedding space. Multiple FR models should be used to generate alternative embedding representations to enhance generalizability against unknown models (Section 4.3).", + "bbox": [ + 212, + 320, + 787, + 503 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "4.1 Cross-image Protection Optimization", + "text_level": 1, + "bbox": [ + 215, + 523, + 570, + 539 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "The cross-image protection capability comes from iterative learning on a set of training images the user provides. The idea is to keep fine-tuning the P3-Mask so that it can offer protection simultaneously to training images while preserving image quality. At the $t$ -th iteration, it samples a mini-batch $\\mathcal{B}$ from the training dataset $\\Omega$ containing photos of user $\\mathcal{P}$ . Then, we find the modification to the P3-Mask that can lead to better protection on each image in the mini-batch with better visual quality with the following operations:", + "bbox": [ + 212, + 547, + 787, + 654 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\n\\mathcal {M} _ {\\mathcal {P}} ^ {t + 1} = \\operatorname {C L I P} _ {[ - \\epsilon , \\epsilon ]} \\left(\\mathcal {M} _ {\\mathcal {P}} ^ {t} - \\eta \\operatorname {S I G N} \\left(\\frac {1}{| \\boldsymbol {\\mathcal {B}} |} \\sum_ {\\mathcal {X} \\in \\boldsymbol {\\mathcal {B}}} \\nabla_ {\\mathcal {M} _ {\\mathcal {P}} ^ {t}} \\mathcal {L} \\left(\\mathcal {X}, \\mathcal {M} _ {\\mathcal {P}} ^ {t}; \\boldsymbol {\\mathcal {T}}, \\omega\\right)\\right)\\right). \\tag {2}\n$$\n", + "text_format": "latex", + "bbox": [ + 274, + 661, + 785, + 694 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Specifically, for each image $\\mathcal{X} \\in \\mathcal{B}$ , we use the P3-Mask learned up to the current iteration, $\\mathcal{M}_{\\mathcal{P}}^t$ , to protect it and compute the loss designed with dual goals:", + "bbox": [ + 212, + 700, + 785, + 733 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\n\\mathcal {L} (\\mathcal {X}, \\boldsymbol {\\mathcal {M}} _ {\\mathcal {P}} ^ {t}; \\boldsymbol {\\mathcal {T}}, \\omega) = \\mathcal {L} _ {\\text {P r o t e c t}} (\\mathcal {X}, \\boldsymbol {\\mathcal {M}} _ {\\mathcal {P}} ^ {t}; \\boldsymbol {\\mathcal {T}}) + \\mathcal {L} _ {\\text {P e r c e p t}} (\\mathcal {X}, \\boldsymbol {\\mathcal {M}} _ {\\mathcal {P}} ^ {t}; \\omega). \\tag {3}\n$$\n", + "text_format": "latex", + "bbox": [ + 294, + 739, + 785, + 757 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "It uses $\\mathcal{L}_{\\mathrm{Protect}}$ to learn how to adjust the P3-Mask $\\mathcal{M}_{\\mathcal{P}}^{t}$ from the previous iteration to better protect against a team of pre-selected FR models $\\pmb{\\tau}$ and $\\mathcal{L}_{\\mathrm{Percept}}$ to preserve the image quality controlled by $\\omega$ . We use the SIGN of the gradients to update the P3-Mask with a learning rate $\\eta$ . A clipping function $\\mathrm{CLIP}_{[-\\epsilon, \\epsilon]}$ is applied to ensure the changes made by the P3-Mask on the facial", + "bbox": [ + 212, + 763, + 787, + 842 + ], + "page_idx": 4 + }, + { + "type": "header", + "text": "Personalized Privacy Protection Mask Against Unauthorized FR", + "bbox": [ + 300, + 114, + 732, + 130 + ], + "page_idx": 4 + }, + { + "type": "page_number", + "text": "5", + "bbox": [ + 774, + 116, + 784, + 126 + ], + "page_idx": 4 + }, + { + "type": "image", + "img_path": "images/ebe22766176c65b7c394601e6e1b72ddd49e799792f30246dbf238fe9e49d7d5.jpg", + "image_caption": [ + "Fig. 3: The iterative generation of P3-Mask for a user. Chameleon goes through multiple facial images of the same user to optimize a P3-Mask with ML pipeline awareness." + ], + "image_footnote": [], + "bbox": [ + 230, + 142, + 774, + 281 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "image are bounded by $\\epsilon$ . Such cross-image protection is not limited to those in the training dataset $\\Omega$ but also unseen images of the same user. We next discuss the design of $\\mathcal{L}_{\\mathrm{Protect}}$ , and $\\mathcal{L}_{\\mathrm{Percept}}$ will be detailed in Section 4.2.", + "bbox": [ + 212, + 349, + 784, + 393 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "For privacy protection, P3-Mask maximizes the distance between the protected and unprotected images in the embedding spaces of a team $\\mathcal{T}$ of pre-selected FR models. We incorporate image processing operations into the optimization to avoid those lossy operations degrading P3-Mask's protection. Consider the raw training image $\\mathcal{X}$ in the mini-batch in Figure 3. We first conduct face detection to produce the cropped face $\\mathcal{FD}(\\mathcal{X})$ . Quantization and resizing are executed on the P3-Mask to match the resolution of the cropped face, denoted by $\\mathcal{QR}(\\mathcal{M}_{\\mathcal{P}}^t)$ . Then, we can apply the mask to the face for protection:", + "bbox": [ + 212, + 393, + 785, + 516 + ], + "page_idx": 5 + }, + { + "type": "equation", + "text": "\n$$\n\\Theta (\\mathcal {X}; \\boldsymbol {\\mathcal {M}} _ {\\mathcal {P}} ^ {t}) = \\Psi \\left(\\operatorname {C L I P} _ {[ 0, 2 5 5 ]} \\left(\\mathcal {F D} (\\mathcal {X}) - \\mathcal {Q R} \\left(\\boldsymbol {\\mathcal {M}} _ {\\mathcal {P}} ^ {t}\\right)\\right)\\right), \\tag {4}\n$$\n", + "text_format": "latex", + "bbox": [ + 328, + 527, + 785, + 545 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "where $\\varPsi$ is a padding and resizing function to meet the input resolution requirement of an FR model (e.g., $(112\\times 112)$ in ArcFace [17]). Note that to simplify the notation, we removed certain arguments without causing confusion.", + "bbox": [ + 212, + 547, + 784, + 593 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "The protection optimization in P3-Mask maximizes the average arccosine distance [17], $\\mathrm{ArcCos}$ , between the embedding of the protected face $F(\\Theta(\\mathcal{X};\\mathcal{M}_{\\mathcal{P}}^t))$ and the embedding of its unprotected counterpart $F(\\Psi(\\mathcal{FD}(\\mathcal{X})))$ extracted by every FR model $F \\in \\mathcal{T}$ in the team:", + "bbox": [ + 212, + 593, + 785, + 654 + ], + "page_idx": 5 + }, + { + "type": "equation", + "text": "\n$$\n\\mathcal {L} _ {\\mathrm {P r o t e c t}} (\\mathcal {X}, \\boldsymbol {\\mathcal {M}} _ {\\mathcal {P}} ^ {t}; \\boldsymbol {\\mathcal {T}}) = \\frac {- 1}{| \\boldsymbol {\\mathcal {T}} |} \\sum_ {F \\in \\boldsymbol {\\mathcal {T}}} \\operatorname {A r c C o s} (F (\\Psi (\\mathcal {F D} (\\mathcal {X}))), F (\\Theta (\\mathcal {X}; \\boldsymbol {\\mathcal {M}} _ {\\mathcal {P}} ^ {t}))). \\quad (5)\n$$\n", + "text_format": "latex", + "bbox": [ + 253, + 660, + 785, + 691 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "4.2 Perceptibility Optimization", + "text_level": 1, + "bbox": [ + 215, + 710, + 488, + 726 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Preserving the visual quality of the facial image is necessary to maintain the usability of the protected image in practice. Hence, we incorporate a perceptibility optimization term in Equation 3. The idea is to minimize the perceptual difference between unprotected and protected images while learning a user's facial signature, such that removing the facial signature from a given facial image will have minimal impact on the visual quality of the protected version of the original image. We use the structural similarity (SSIM) [35] to capture perceptual", + "bbox": [ + 212, + 734, + 787, + 840 + ], + "page_idx": 5 + }, + { + "type": "page_number", + "text": "6", + "bbox": [ + 217, + 114, + 228, + 126 + ], + "page_idx": 5 + }, + { + "type": "header", + "text": "K.-H. Chow et al.", + "bbox": [ + 271, + 114, + 392, + 127 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/acfa6a0a3d7c3632dd4e74d91c52c2872adda959c3939046218ea3a1ed790434.jpg", + "image_caption": [ + "(a) Two-model: 28 Options" + ], + "image_footnote": [], + "bbox": [ + 241, + 143, + 431, + 207 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/aa4effa16cff7aa7d27ffbb5a95d34b8fd6a0665dae2bd748fceef68b42e3689.jpg", + "image_caption": [ + "(b) Three-model: 56 Options" + ], + "image_footnote": [], + "bbox": [ + 241, + 223, + 421, + 286 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/96640aaf397d553c48391063d68ae3e920c701982d7df828e5bf54eafdd97ce7.jpg", + "image_caption": [ + "Fig. 5: Out of 28 options of two-model ensembles from a pool of eight models, the ensemble selected by our approach (green) leads to much better protection than a randomly selected one (red).", + "Fig.4: Focal diversity provides a strong indicator of protection effectiveness for selecting an ensemble." + ], + "image_footnote": [], + "bbox": [ + 478, + 143, + 767, + 284 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "differences, as it has been shown to align with the human perception system:", + "bbox": [ + 215, + 382, + 767, + 398 + ], + "page_idx": 6 + }, + { + "type": "equation", + "text": "\n$$\n\\mathcal {L} _ {\\text {P e r c e p t}} \\left(\\mathcal {X}, \\boldsymbol {\\mathcal {M}} _ {\\mathcal {P}} ^ {t}; \\omega\\right) = \\lambda_ {\\text {S S I M}} \\max \\left[ \\frac {1 - \\operatorname {S S I M} \\left(\\Psi \\left(\\boldsymbol {\\mathcal {F D}} (\\mathcal {X})\\right) , \\Theta \\left(\\mathcal {X} ; \\boldsymbol {\\mathcal {M}} _ {\\mathcal {P}} ^ {t}\\right)\\right)}{2} - \\omega , 0 \\right]. \\tag {6}\n$$\n", + "text_format": "latex", + "bbox": [ + 227, + 405, + 784, + 436 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "The parameters, $\\omega$ and $\\lambda_{\\mathrm{SSIM}}$ , enable two features. $\\omega$ controls the SSIM degradation. Any SSIM degradation greater than $2\\omega$ will cause the term to be non-zero, making the optimization adjust the P3-Mask reducing its impact on image quality. $\\lambda_{\\mathrm{SSIM}}$ balances the importance of privacy protection and perceptibility. We use dynamic scheduling [31] such that it will be adjusted automatically.", + "bbox": [ + 215, + 444, + 784, + 521 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "4.3 Focal Diversity Ensemble Optimization", + "text_level": 1, + "bbox": [ + 215, + 542, + 584, + 558 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "For each original image and its protected version by removing the facial signature learned up to the current iteration, we use two FR models in Figure 3 to generate two embeddings, which serve as alternative channels to learn a high-quality facial signature. Choosing FR models that best complement each other can achieve better protection than randomly selected FR models. We use an ensemble optimization method to select the best $k$ -model ensemble among a pool of $N$ FR models. The key idea is to find an ensemble with members making decorrelated mistakes. To quantify this behavior, we leverage the focal diversity framework [13,14,37]. As shown in Figure 4, the higher the focal diversity of the ensemble (a green dot), the stronger the P3-Mask will be in protecting against other FR models not used during the P3-Mask generation process.", + "bbox": [ + 212, + 568, + 785, + 733 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Given a collection of $N$ FR models $\\{F_1,\\dots,F_N\\}$ , we first identify the negative samples for each FR model $F$ by locating validation images that $F$ fails to recognize their true identity. To rank ensembles of size $S$ , we enumerate all $\\binom{N}{S}$ combinations. For each combination (ensemble) $\\pmb{\\tau}$ , we consider each member to be the focal model $F_{\\mathrm{focal}}$ and use its negative samples to statistically estimate the level of negative correlation $\\lambda_{\\mathrm{focal}}(\\pmb{\\tau};F_{\\mathrm{focal}})$ between $F_{\\mathrm{focal}}$ and the remaining models in $\\pmb{\\tau}$ , computed by measuring the degree of disagreements using the", + "bbox": [ + 212, + 734, + 787, + 840 + ], + "page_idx": 6 + }, + { + "type": "header", + "text": "Personalized Privacy Protection Mask Against Unauthorized FR", + "bbox": [ + 300, + 114, + 732, + 128 + ], + "page_idx": 6 + }, + { + "type": "page_number", + "text": "7", + "bbox": [ + 774, + 114, + 784, + 126 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "generalized non-pairwise measure [27]. The same procedure is repeated by considering each member in the ensemble as the focal model, and the focal diversity of the ensemble $\\pmb{\\mathcal{T}}$ is finalized as $d_{\\mathrm{focal}}(\\pmb {\\mathcal{T}}) = \\frac{1}{S}\\sum_{F_{\\mathrm{focal}}\\in \\pmb{\\mathcal{T}}}[1 - \\lambda_{\\mathrm{focal}}(\\pmb {\\mathcal{T}};F_{\\mathrm{focal}})]$ . Given a team size $S$ , the ensemble with the highest focal diversity can be deployed. As shown in Figure 5, the selected two-model team (green bar) is indeed the one leading to the strongest protection among all 28 options, which is over $20\\%$ stronger than a randomly selected ensemble (red bar). Note that we only need to analyze the failures of individual FR models, which is significantly more efficient than generating P3-Mask for each option and conducting evaluation.", + "bbox": [ + 212, + 146, + 787, + 282 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "5 Online Image Protection and Refinement", + "text_level": 1, + "bbox": [ + 215, + 305, + 650, + 321 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "The user $\\mathcal{P}$ obtains her P3-Mask $\\mathcal{M}_{\\mathcal{P}}$ from Chameleon. Given an image $\\mathcal{X}$ , she can obfuscate her facial identity on her device by applying the P3-Mask:", + "bbox": [ + 212, + 335, + 785, + 368 + ], + "page_idx": 7 + }, + { + "type": "equation", + "text": "\n$$\n\\operatorname {M A S K} (\\mathcal {X}; \\boldsymbol {\\mathcal {M}} _ {\\mathcal {P}}) = \\operatorname {C L I P} _ {[ 0, 2 5 5 ]} (\\mathcal {F D} (\\mathcal {X}) - \\mathcal {Q R} (\\boldsymbol {\\mathcal {M}} _ {\\mathcal {P}})), \\tag {7}\n$$\n", + "text_format": "latex", + "bbox": [ + 330, + 378, + 785, + 393 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "which is the protected facial region to be put in the original image $\\mathcal{X}$ to produce the protected version $\\mathcal{X}'$ for sharing.", + "bbox": [ + 212, + 402, + 785, + 431 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Chameleon supports authorizing trusted third parties to conduct FR on a user's facial images in a space-time efficient manner without transmitting and storing multiple versions of the same photo (one protected for public sharing and one unprotected for internal FR). By sharing the P3-Mask $\\mathcal{M}_{\\mathcal{P}}$ as the key, the third parties can de-obfuscate the facial signature by unmasking:", + "bbox": [ + 212, + 431, + 785, + 508 + ], + "page_idx": 7 + }, + { + "type": "equation", + "text": "\n$$\n\\operatorname {U N M A S K} \\left(\\mathcal {X} ^ {\\prime}; \\boldsymbol {\\mathcal {M}} _ {\\mathcal {P}}\\right) = \\operatorname {C L I P} _ {[ 0, 2 5 5 ]} \\left(\\mathcal {F D} \\left(\\mathcal {X} ^ {\\prime}\\right) + \\mathcal {Q R} \\left(\\boldsymbol {\\mathcal {M}} _ {\\mathcal {P}}\\right)\\right). \\tag {8}\n$$\n", + "text_format": "latex", + "bbox": [ + 316, + 518, + 785, + 534 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "In Section 6.2, we will show two properties of this process: (i) the signature-restored photos can be used for FR with no accuracy degradation, and (ii) the protection can only be done by the P3-Mask of the same person in the facial image, and the restoration is successful only if the same key is used.", + "bbox": [ + 212, + 542, + 785, + 604 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "6 Experimental Evaluation", + "text_level": 1, + "bbox": [ + 215, + 627, + 493, + 645 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "We conduct experiments on FaceScrub [26] and LFW [20]. To provide a detailed analysis, ten celebrities are randomly selected as users on FaceScrub (Table 1). For each user, we split their facial images into three parts: (i) $10\\%$ are used as probe images (2nd column), (ii) $70\\%$ are used as gallery images and are used by Chameleon to train the mask (3rd column), and (iii) $20\\%$ are also used as gallery images but unseen during the training process (4th column). The last part is crucial to evaluate the protection of unseen images of the user. In total, we have 123 facial images with \"unknown identities.\" Following relevant works [31, 40], all facial images of other people are included in the gallery dataset as noise. The same splitting method is also used for LFW. By default, the results reported in this paper focus on FaceScrub due to similar observations on LFW. The source code is available at https://github.com/git-disl/Chameleon.", + "bbox": [ + 212, + 657, + 787, + 840 + ], + "page_idx": 7 + }, + { + "type": "page_number", + "text": "8", + "bbox": [ + 217, + 114, + 228, + 126 + ], + "page_idx": 7 + }, + { + "type": "header", + "text": "K.-H. Chow et al.", + "bbox": [ + 271, + 114, + 392, + 127 + ], + "page_idx": 7 + }, + { + "type": "table", + "img_path": "images/59053062356c7a22fe029fd45a1e98c3c5df1d8883027e5b9de792afd3cc213f.jpg", + "table_caption": [ + "Table 1: Ten example celebrities (users) on FaceScrub for analysis." + ], + "table_footnote": [], + "table_body": "
NameProbe ImagesGallery Images
SeenUnseenTotal
Morena Baccarin118223105
Bradley Cooper128925114
America Ferrera1611332145
Gerard Butler118224106
Eva Longoria129126117
Melissa Egan128724111
Kim Catrall139627123
Allison Janney128926115
Roma Downey117822100
Steve Carell139628124
Others (Noise)///48368
Total Number of Gallery Images:49528
", + "bbox": [ + 217, + 185, + 496, + 339 + ], + "page_idx": 8 + }, + { + "type": "table", + "img_path": "images/43ebcbe7d0527bf65f2910daeded75f360c748a51ed65d8983553651252f5699.jpg", + "table_caption": [ + "Table 2: The collection of publicly available pretrained FR models." + ], + "table_footnote": [], + "table_body": "
Model IDNeural Arch.Training DatasetFR Acc.
EN-MCEfficientNetMS-Celeb-1M94.94%
RN50-GLResNet50Glint360K96.68%
RN50-MCResNet50MS-Celeb-1M89.25%
RN50-VFResNet50VGG-Face294.30%
RN50-WFResNet50WebFace600K96.72%
RN18-MCResNet18MS-Celeb-1M82.64%
RN34-MCResNet34MS-Celeb-1M84.49%
RN100-MCResNet100MS-Celeb-1M91.85%
", + "bbox": [ + 506, + 185, + 781, + 292 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "Public clouds (e.g., Azure) now require manual approval to use their FR services. Hence, we believe that privacy intruders will opt for deploying pretrained FR models, as many high-quality ones are available on the Internet and can be used out of the box. Due to the large number of possible FR algorithms and architectures, we first focus on the eight FR models listed in Table 2, which are based on ArcFace [17], the state-of-the-art FR algorithm, with varying neural architectures and training datasets. Chameleon can be easily extended to incorporate any FR models by simply adding them to the collection. Nonetheless, in Section 6.4, we will show that thanks to the focal diversity-optimized teaming, the P3-Mask generated from a collection of ArcFace-only models can also be effective in protecting against privacy intruders using models of other FR algorithms, such as FaceNet [30] and MagFace [24]. By default, we use the two-model team (EN-MC, RN50-GL) with the highest focal diversity (Section 4.3). The P3-Mask is trained for 50 epochs on NVIDIA RTX 2080 SUPER GPU with $\\eta = 0.001$ , $|\\mathcal{B}| = 4$ , $\\epsilon = 0.063$ , and $\\omega = 0.03$ . We provide details for reproducibility and additional results in the appendix.", + "bbox": [ + 212, + 364, + 787, + 608 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "6.1 Protection Effectiveness", + "text_level": 1, + "bbox": [ + 215, + 627, + 460, + 641 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "Table 3 shows the personalized mask for seven users on FaceScrub (other users are available in the appendix). Different users have distinct facial signatures (P3-Mask) learned by Chameleon. We apply these masks to their respective gallery", + "bbox": [ + 212, + 648, + 787, + 696 + ], + "page_idx": 8 + }, + { + "type": "table", + "img_path": "images/b51a87f887545091a74583b8414fefb7b6b1ee9e6c9b9c1db1472c0ab2e0cd28.jpg", + "table_caption": [ + "Table 3: The P3-Mask (rescaled) for seven users on FaceScrub. Chameleon learns the distinct facial signature for each of them." + ], + "table_footnote": [], + "table_body": "
M. BaccarinB. CooperA. FerreraG. ButlerE. LongoriaM. EganK. Cattrall
", + "bbox": [ + 217, + 760, + 785, + 835 + ], + "page_idx": 8 + }, + { + "type": "header", + "text": "Personalized Privacy Protection Mask Against Unauthorized FR", + "bbox": [ + 300, + 114, + 732, + 128 + ], + "page_idx": 8 + }, + { + "type": "page_number", + "text": "9", + "bbox": [ + 774, + 116, + 784, + 126 + ], + "page_idx": 8 + }, + { + "type": "image", + "img_path": "images/0b250f4eeca1c84abfbf5b040210e3290995710c7c564eb2c315f8d4100c2cc6.jpg", + "image_caption": [ + "(a) Dataset: FaceScrub", + "Fig. 6: Protection effectiveness on two benchmarks using Chameleon, OPOM [40], TIP-IM [39], and Fawkes [31] against FR models unknown to them." + ], + "image_footnote": [], + "bbox": [ + 222, + 143, + 498, + 215 + ], + "page_idx": 9 + }, + { + "type": "image", + "img_path": "images/9dd95d6eeed8a8388885d77cde22d6fae71ae328457928e787de1b65c03f3a1e.jpg", + "image_caption": [ + "(b) Dataset: LFW" + ], + "image_footnote": [], + "bbox": [ + 506, + 143, + 771, + 215 + ], + "page_idx": 9 + }, + { + "type": "text", + "text": "images and test the FR accuracy using probe images. All probe images can be correctly identified when no protection mechanism is used. Under Chameleon, the probe images should be matched to an incorrect identity, resulting in a low FR accuracy. Hence, we define an evaluation metric, $\\underline{\\text{Protection Success Rate}}$ (PSR), to be $(100 - \\text{FR Accuracy})$ reported in percentages.", + "bbox": [ + 212, + 304, + 784, + 381 + ], + "page_idx": 9 + }, + { + "type": "text", + "text": "Chameleon can protect against unknown FR models. We compare the cross-model protection of Chameleon with OPOM [40], the only anti-FR approach generating one privacy protection mask for each person as Chameleon, and both TIP-IM [39] and Fawkes [31], which conduct per-image optimization. For a fair comparison, all methods are set with the same perturbation budget. Figure 6 summarizes the results on both datasets. All five FR models are unknown to both protection mechanisms. We make two observations. First, Chameleon consistently outperforms OPOM by a large margin. We found that OPOM masks do not transfer well, especially considering the end-to-end ML pipeline. Its PSR can drop by $33\\%$ because of those lossy operations. Second, Chameleon can be as competitive as those methods conducting per-image optimization. TIP-IM and Fawkes spend minutes to optimize the protection for each image. Still, Chameleon outperforms them and can be applied instantly.", + "bbox": [ + 212, + 383, + 785, + 580 + ], + "page_idx": 9 + }, + { + "type": "text", + "text": "Table 4 reports the detailed PSR for each user Chameleon achieves when the privacy intruder uses different FR models (2nd to 9th columns). We make two observations. First, when the privacy intruder uses an FR model known to the generation process, a PSR of $100\\%$ can be consistently achieved (2nd to 3rd columns). Second, Chameleon can protect against unknown FR models (4th to 9th columns). Even when the FR model used by the intruder is trained on a different dataset (i.e., RN50-VF and RN50-WF) or with a different neural architecture (i.e., RN18-MC, RN34-MC, and RN100-MC) than any FR models known by Chameleon, it still provides a PSR over $90.65\\%$ , meaning that facial images of a user are matched to gallery images of a different person. To demonstrate such a mismatch, in Table 5, we show the probe images from two users and their most similar gallery images found by four unknown FR models with two settings: (i) the \"Unprotected\" scenario where no one employs protection and (ii) the \"Chameleon\" scenario where the intruder scraped photos protected by our solution. Taking M. Baccarin as an example, the most similar gallery images found in the unprotected scenario belong to her. In contrast, when Chameleon is used, the same probe image is misidentified, e.g., as L. Hartley by RN50-MC.", + "bbox": [ + 212, + 583, + 785, + 840 + ], + "page_idx": 9 + }, + { + "type": "page_number", + "text": "10", + "bbox": [ + 217, + 114, + 235, + 126 + ], + "page_idx": 9 + }, + { + "type": "header", + "text": "K.-H. Chow et al.", + "bbox": [ + 271, + 114, + 392, + 126 + ], + "page_idx": 9 + }, + { + "type": "table", + "img_path": "images/43c8e098abcb4d1e2d73b1568ee48ff46c218e87058c9cda54d925575a18ad45.jpg", + "table_caption": [ + "Table 4: Using the two-model team with high focal diversity (2nd and 3rd columns), Chameleon can offer privacy protection even when the privacy intruder uses FR models unseen during the mask generation process (4th to 9th columns)." + ], + "table_footnote": [], + "table_body": "
NameProtection Success Rate - PSR (%)
EN-MCRN50-GLRN50-MCRN50-VFRN50-WFRN18-MCRN34-MCRN100-MCMeanStd
M. Baccarin100.00100.00100.0072.7390.91100.0090.91100.0094.329.64
B. Cooper100.00100.0083.33100.00100.00100.0091.6791.6795.836.30
A. Ferrera100.00100.0093.7575.00100.00100.0093.75100.0095.318.68
G. Butler100.00100.0090.91100.00100.0090.9190.91100.0096.594.70
E. Longoria100.00100.00100.00100.00100.00100.00100.00100.00100.000.00
M. Egan100.00100.00100.00100.00100.00100.00100.0091.6798.962.95
K. Cattrall100.00100.00100.0084.62100.00100.00100.00100.0098.085.44
A. Janney100.00100.0091.67100.0091.67100.00100.00100.0097.923.86
R. Downey100.00100.00100.0072.7390.91100.0090.91100.0094.329.64
S. Carell100.00100.0084.6292.3192.3192.3176.92100.0092.318.22
Mean100.00100.0093.5290.6596.5897.4194.4297.42
Std0.000.006.4011.164.434.187.414.15
", + "bbox": [ + 217, + 199, + 785, + 361 + ], + "page_idx": 10 + }, + { + "type": "table", + "img_path": "images/ae5e2d205e2de1759f9d9c692a81167ab63ff8de595a628bf51476a764d1b514.jpg", + "table_caption": [ + "Table 5: The most similar gallery images on FaceScrub found by different FR models unknown to Chameleon." + ], + "table_footnote": [], + "table_body": "
Probe\nImageThe Most Similar Gallery Image & Its Identity Found by Unseen FR Models
RN50-MCRN50-VFRN34-MCRN100-MC
UnprotectedChameleonUnprotectedChameleonUnprotectedChameleonUnprotectedChameleon
M. BaccarinIdentifiedProtectedIdentifiedProtectedIdentifiedProtectedIdentifiedProtected
M. BaccarinL. HartleyM. BaccarinL. HartleyM. BaccarinC. ElectraM. BaccarinL. Hartley
B. CooperIdentifiedProtectedIdentifiedProtectedIdentifiedProtectedIdentifiedProtected
B. CooperP. WalkerB. CooperM. VartanB. CooperM. VartanB. CooperM. Vartan
", + "bbox": [ + 217, + 415, + 794, + 568 + ], + "page_idx": 10 + }, + { + "type": "text", + "text": "6.2 FR Service Usability", + "text_level": 1, + "bbox": [ + 214, + 594, + 433, + 609 + ], + "page_idx": 10 + }, + { + "type": "text", + "text": "Chameleon allows users to grant FR permission to trusted third parties by sharing their P3-Mask to de-obfuscate the protected image. There is no need to transmit and store an unprotected photo for internal use and a protected one for public visibility, doubling network and storage costs. Table 6 shows the results in FR accuracy. First, using the correct mask (i.e., the one used to protect the images) to reverse the protection can restore FR accuracy. Taking M. Baccarin as an example, the FR accuracy increases from $6.82\\%$ before unmasking (b) to $100\\%$ after unmasking (c). Second, using an incorrect mask for unmasking cannot restore the FR accuracy and can lead to even worse performance. It drops from $6.82\\%$ to $1.14\\%$ after unmasking (d).", + "bbox": [ + 212, + 616, + 787, + 768 + ], + "page_idx": 10 + }, + { + "type": "text", + "text": "6.3 Protection Cost Analysis", + "text_level": 1, + "bbox": [ + 214, + 787, + 470, + 801 + ], + "page_idx": 10 + }, + { + "type": "text", + "text": "Speed and Resources. Figure 7 compares Chameleon with OPOM, TIP-IM, and Fawkes regarding the protection time per face, compute, and storage costs.", + "bbox": [ + 212, + 809, + 785, + 839 + ], + "page_idx": 10 + }, + { + "type": "header", + "text": "Personalized Privacy Protection Mask Against Unauthorized FR", + "bbox": [ + 300, + 114, + 732, + 128 + ], + "page_idx": 10 + }, + { + "type": "page_number", + "text": "11", + "bbox": [ + 767, + 116, + 782, + 126 + ], + "page_idx": 10 + }, + { + "type": "table", + "img_path": "images/ccf695ca088ae618891d5511694f1706621df7f842f480e0bf8fc213903d40b1.jpg", + "table_caption": [ + "Table 6: The protection by Chameleon can reduce FR accuracy against FR models (b), but the user can provide the P3-Mask used to protect her images as the key for the authorized third party to reverse the protection process (c), which leads to restored FR performance. However, an incorrect key cannot restore FR and may worsen it (d)." + ], + "table_footnote": [], + "table_body": "
ScenarioFace Recognition Accuracy (%)
EN-MCRN50-GLRN50-MCRN50-VFRN50-WFRN18-MCRN34-MCRN100-MCMeanStd
User: M. Baccarin
(a) No Protection100.00100.00100.00100.00100.00100.00100.00100.00100.000.00
(b) Protected0.000.009.0918.189.099.090.009.096.826.43
(c) Correctly Unmasked100.00100.00100.00100.00100.00100.00100.00100.00100.000.00
(d) Incorrectly Unmasked0.000.000.000.009.090.000.000.001.143.21
User: B. Cooper
(a) No Protection100.00100.00100.00100.00100.00100.00100.00100.00100.000.00
(b) Protected0.000. 0016.670.000.000.008.338.334.176.30
(c) Correctly Unmasked100.00100.00100.00100.00100.00100.00100.00100.00100.000.00
(d) Incorrectly Unmasked0.000.000. 000.000.000.000.000.000.000.00
", + "bbox": [ + 215, + 212, + 785, + 353 + ], + "page_idx": 11 + }, + { + "type": "image", + "img_path": "images/c0ada8d7af1371cac2804844d97f3bc7f896e94d46a492f3084236d029994d0b.jpg", + "image_caption": [ + "Fig. 7: Chameleon offers much better privacy protection than OPOM without sacrificing speed and resource usages. It can protect instantly and require negligible compute resources as OPOM, while TIP-IM and Fawkes are slow and costly." + ], + "image_footnote": [], + "bbox": [ + 251, + 367, + 349, + 449 + ], + "page_idx": 11 + }, + { + "type": "image", + "img_path": "images/f9d6a25330c8829d82e56b24e0174b87015eef8742c9b4d9ced91233ce04b07f.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 352, + 367, + 447, + 449 + ], + "page_idx": 11 + }, + { + "type": "image", + "img_path": "images/d94dc64126842e1ab631d873825ca79fc62ca79a0d7b111bfe091d3bbd045ffa.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 452, + 367, + 547, + 449 + ], + "page_idx": 11 + }, + { + "type": "image", + "img_path": "images/aa18156c72669038ebbb258d2174e3b003517e59b7bf42117c01540a525f1dd8.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 553, + 367, + 650, + 449 + ], + "page_idx": 11 + }, + { + "type": "image", + "img_path": "images/43f16377adc45961eedfc4aa4c8ca73c101a322c1aff2f2c279dbd0c48c09728.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 653, + 367, + 751, + 449 + ], + "page_idx": 11 + }, + { + "type": "text", + "text": "Chameleon and OPOM produce masks that are applicable to any facial image of the same user. The protection can be completed in 0.0076 seconds. Note that Chameleon is significantly more effective than OPOM, as described in Figure 6. Instead, TIP-IM and Fawkes require per-image optimization and take 105.12 seconds to complete the protection with GPUs for acceleration and storage for FR models. Chameleon only needs to conduct simple arithmetic operations. Only the P3-Mask needs to be stored on the user device. Hence, it can be deployed with instant protection even on an edge device with weak computing power.", + "bbox": [ + 212, + 535, + 787, + 657 + ], + "page_idx": 11 + }, + { + "type": "text", + "text": "Image Quality. Chameleon can well preserve image quality. Table 7 provides two examples for four users, contrasting the facial images with no protection (3rd and 5th columns) with the ones with their facial signature removed by Chameleon (4th and 6th columns). They capture different variations of facial images: M. Baccarin's images have different lighting conditions, B. Cooper's have different face sizes, E. Longoria's have different postures, and G. Bulter's have different expressions. The protected images are visually similar to the unprotected counterparts, but they will not be matched to the clean images of the corresponding person. Indeed, Chameleon can generate higher-quality images than existing methods. As reported in Table 8, Chameleon not only offers better protection (Figure 6) with a similar cost (Figure 7) as OPOM but also achieves much better image quality measured in SSIM (0.9493 vs 0.8839).", + "bbox": [ + 212, + 659, + 787, + 840 + ], + "page_idx": 11 + }, + { + "type": "page_number", + "text": "12", + "bbox": [ + 217, + 114, + 235, + 126 + ], + "page_idx": 11 + }, + { + "type": "header", + "text": "K.-H. Chow et al.", + "bbox": [ + 271, + 114, + 392, + 127 + ], + "page_idx": 11 + }, + { + "type": "table", + "img_path": "images/d4feeb1ae43ca7d63837a93041546e9976dc59da9c0f846544ca4e5f61061e4f.jpg", + "table_caption": [ + "Table 7: The P3-Mask can be applied to any facial im- Table 8: Chameleon offers ages of the same person while preserving image quality. much better protection while" + ], + "table_footnote": [], + "table_body": "
UserMaskSample Image 1Sample Image 2
No ProtectionProtectedNo ProtectionProtected
M. Baccarin
B. Cooper
E. Longoria
G. Image
", + "bbox": [ + 222, + 185, + 583, + 392 + ], + "page_idx": 12 + }, + { + "type": "table", + "img_path": "images/b64dc63214ea2c3e635ae630f6018fbd459c3f50f5ac49baa6841ba1a212d690.jpg", + "table_caption": [], + "table_footnote": [], + "table_body": "
Example
Original SSIM: 1.0000
Chameleon SSIM: 0.9493
OPOM SSIM: 0.8839
TIP-IM SSIM: 0.8850
Fawkes SSIM: 0.9612
", + "bbox": [ + 612, + 199, + 764, + 401 + ], + "page_idx": 12 + }, + { + "type": "text", + "text": "6.4 Focal Diversity-based Teaming", + "text_level": 1, + "bbox": [ + 215, + 431, + 514, + 448 + ], + "page_idx": 12 + }, + { + "type": "text", + "text": "Chameleon automatically selects a high-quality team for deployment with a specified budget. Table 9 compares the detailed PSR using the most diverse and the least diverse teams with (a) two and (b) three models. In both cases, the most diverse team outperforms the least diverse one, which is only effective when the privacy intruder uses the FR model known in the team. The selection of high-quality teams is non-trivial because we observe that composing a team of FR models with the highest FR accuracy is not always a good option, and a team of models having different neural architectures is not always the top priority [13].", + "bbox": [ + 212, + 460, + 787, + 583 + ], + "page_idx": 12 + }, + { + "type": "text", + "text": "In Figure 8, we further show that the P3-Mask generated from a carefully-chosen team can protect against FR models of unknown algorithms. Specifically,", + "bbox": [ + 212, + 583, + 785, + 614 + ], + "page_idx": 12 + }, + { + "type": "table", + "img_path": "images/9458f030ebe4eed0eb0e34d235396c19d5705138af6465bdca3a0d0a44de6897.jpg", + "table_caption": [ + "Table 9: The most diverse and the least diverse teams with (a) two or (b) three models identified by our focal diversity-based teaming method. The most diverse teams offer significantly better protection in terms of protection success rates." + ], + "table_footnote": [], + "table_body": "
Protection TeamProtection Success Rate - PSR (%)
EN-MCRN50-GLRN50-MCRN50-VFRN50-WFRN18-MCRN34-MCRN100-MCMeanStd
(a) Two-model Teams
Most Diverse:(EN-MC, RN50-GL)100.00100.0093.5290.6596.5897.4194.4297.4296.253.23
Least Diverse:(RN18-MC, RN34-MC)11.3860.16100.0055.2873.17100.00100.0080.4972.5628.54
(b) Three-model Teams
Most Diverse:(EN-MC, RN50-GL, RN50-WF)100.00100.00100.00100.0092.6893.50100.00100.0098.273.00
Least Diverse:(RN18-MC, RN34-MC, RN50-MC)2.4478.86100.0052.0384.55100.00100.0093.5076.4231.82
", + "bbox": [ + 217, + 696, + 792, + 837 + ], + "page_idx": 12 + }, + { + "type": "header", + "text": "Personalized Privacy Protection Mask Against Unauthorized FR", + "bbox": [ + 300, + 114, + 732, + 128 + ], + "page_idx": 12 + }, + { + "type": "page_number", + "text": "13", + "bbox": [ + 767, + 114, + 785, + 126 + ], + "page_idx": 12 + }, + { + "type": "image", + "img_path": "images/b369f682755ec1966b4de25ceb8421d5d5ec12fd6c27737248f0001a7362c308.jpg", + "image_caption": [ + "Fig. 8: The most diverse teams allow Chameleon to be effective against FR models of unknown algorithms." + ], + "image_footnote": [], + "bbox": [ + 228, + 143, + 450, + 231 + ], + "page_idx": 13 + }, + { + "type": "image", + "img_path": "images/bed3d7bad28822bdaaff5ee932ea111a0ec9135f62865cde53bcad95f29288ce.jpg", + "image_caption": [ + "Fig. 9: The PSR of Chameleon against adaptive intruders using different strategies to wash out the patterns in protected images." + ], + "image_footnote": [], + "bbox": [ + 475, + 150, + 779, + 231 + ], + "page_idx": 13 + }, + { + "type": "text", + "text": "we consider the most and the least diverse three-model teams in Table 9(b). Even though both teams include only FR models based on ArcFace [17], the most diverse one can effectively protect against FR models based on FaceNet [30] or MagFace [24]. The protection effectiveness can be further strengthened by incorporating more FR models into the collection.", + "bbox": [ + 212, + 321, + 787, + 397 + ], + "page_idx": 13 + }, + { + "type": "text", + "text": "6.5 Chameleon Against Adaptive Adversaries", + "text_level": 1, + "bbox": [ + 214, + 421, + 604, + 436 + ], + "page_idx": 13 + }, + { + "type": "text", + "text": "An adaptive adversary may run Chameleon to produce the P3-Mask for each person and de-obfuscate their photos. When probe images need to be identified, they will be matched against a \"clean\" face database. However, it is infeasible, as we show in Section 6.2, that the FR accuracy can only be restored when the masks used for obfuscation and de-obfuscation are identical. It is impossible to reproduce the same P3-Mask used by the user because it requires the same set of clean images and random states.", + "bbox": [ + 212, + 448, + 787, + 553 + ], + "page_idx": 13 + }, + { + "type": "text", + "text": "Alternatively, an adaptive adversary may wash out patterns introduced by Chameleon before performing FR. We consider three popular lightweight and task-agnostic methods studied in the adversarial example domain [18] in Figure 9, including (a) JPEG compression [15], (b) Gaussian Filter, and (d) Median Filter [38]. The results show that their effectiveness is limited on Chameleon. The consideration of the end-to-end ML pipeline increases the robustness of P3-Mask.", + "bbox": [ + 212, + 554, + 787, + 645 + ], + "page_idx": 13 + }, + { + "type": "text", + "text": "7 Conclusions", + "text_level": 1, + "bbox": [ + 215, + 670, + 369, + 686 + ], + "page_idx": 13 + }, + { + "type": "text", + "text": "We have presented Chameleon against unauthorized FR. Chameleon generates a P3-Mask for each user with cross-image and perceptibility optimizations to offer (i) instant and lightweight protection on any facial images of the same user and (ii) preservation of image quality. Also, we have shown that P3-Mask enables cost-effective de-obfuscation for authorizing FR services. Such an authorization process can only be conducted by the one with the identical mask used for protection. Chameleon also features a focal diversity-optimized teaming method to select a high-quality FR team to generate P3-Mask with strong robustness against unknown FR models.", + "bbox": [ + 212, + 703, + 787, + 840 + ], + "page_idx": 13 + }, + { + "type": "page_number", + "text": "14", + "bbox": [ + 217, + 114, + 235, + 126 + ], + "page_idx": 13 + }, + { + "type": "header", + "text": "K.-H. Chow et al.", + "bbox": [ + 271, + 114, + 392, + 126 + ], + "page_idx": 13 + }, + { + "type": "text", + "text": "Acknowledgments", + "text_level": 1, + "bbox": [ + 217, + 143, + 392, + 162 + ], + "page_idx": 14 + }, + { + "type": "text", + "text": "This research is partially sponsored by the NSF CISE grants 2302720, 2312758, 2038029, an IBM faculty award, and a grant from CISCO Edge AI program. It is part of the PhD dissertation of the first author, who graduated from Georgia Tech in Spring 2023. The first author acknowledges the support of the IBM PhD Fellowship in 2022-2023 and the support from the HKU-CS Start-up Fund.", + "bbox": [ + 212, + 179, + 787, + 257 + ], + "page_idx": 14 + }, + { + "type": "text", + "text": "References", + "text_level": 1, + "bbox": [ + 215, + 282, + 323, + 299 + ], + "page_idx": 14 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "1. Brazilian retailer quizzed over facial recognition tech. https://www.zdnet.com/article/brazilian-retailer-quizzed-over-facial-recognition-tech/ (2019)", + "2. Facial recognition in the united states: Privacy concerns and legal developments. https://www.asisonline.org/security-management-magazine/monthlyissues/security-technology/archive/2021/december/facial-recognition-in-the-us-privacy-concerns-and-legal-developments/ (2021)", + "3. The secretive company that might end privacy as we know it. https://www.nytimes.com/2020/01/18/technology/clearview-privacy-facial-recognition.html (2021)", + "4. Clearview AI. https://www.clearview.ai/ (2023)", + "5. Facial recognition and identity risk. https://www.equifax.co.uk/resources/identity-protection/facial-recognition-and-identity-risk.html (2023)", + "6. HuggingFace. https://huggingface.co (2023)", + "7. Insightface: 2d and 3d face analysis project. https://github.com/deepinsight/insightface (2023)", + "8. PimEyes. https://pimeyes.com/ (2023)", + "9. Borgwardt, K.M., Gretton, A., Rasch, M.J., Kriegel, H.P., Scholkopf, B., Smola, A.J.: Integrating structured biological data by kernel maximum mean discrepancy. Bioinformatics 22(14), e49-e57 (2006)", + "10. Cherepanova, V., Goldblum, M., Foley, H., Duan, S., Dickerson, J., Taylor, G., Goldstein, T.: Lowkey: Leveraging adversarial attacks to protect social media users from facial recognition. In: International Conference on Learning Representations (ICLR) (2021)", + "11. Choi, Y., Choi, M., Kim, M., Ha, J.W., Kim, S., Choo, J.: Stargan: Unified generative adversarial networks for multi-domain image-to-image translation. In: Proceedings of the IEEE conference on computer vision and pattern recognition. pp. 8789-8797 (2018)", + "12. Chow, K.H., Hu, S., Huang, T., Ilhan, F., Wei, W., Liu, L.: Diversity-driven privacy protection masks against unauthorized face recognition. Proceedings on Privacy Enhancing Technologies 4, 381-392 (2024)", + "13. Chow, K.H., Liu, L.: Robust object detection fusion against deception. In: Proceedings of the 27th ACM SIGKDD Conference on Knowledge Discovery & Data Mining. pp. 2703-2713 (2021)", + "14. Chow, K.H., Liu, L.: Boosting object detection ensembles with error diversity. In: 2022 IEEE International Conference on Data Mining (ICDM). pp. 903-908. IEEE (2022)" + ], + "bbox": [ + 218, + 316, + 784, + 839 + ], + "page_idx": 14 + }, + { + "type": "header", + "text": "Personalized Privacy Protection Mask Against Unauthorized FR", + "bbox": [ + 300, + 114, + 730, + 128 + ], + "page_idx": 14 + }, + { + "type": "page_number", + "text": "15", + "bbox": [ + 767, + 114, + 785, + 126 + ], + "page_idx": 14 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "15. Das, N., Shanbhogue, M., Chen, S.T., Hohman, F., Li, S., Chen, L., Kounavis, M.E., Chau, D.H.: Shield: Fast, practical defense and vaccination for deep learning usingJPEG compression. In: Proceedings of the 24th ACM SIGKDD International Conference on Knowledge Discovery & Data Mining. pp. 196-204 (2018)", + "16. Deb, D., Zhang, J., Jain, A.K.: Advfaces: Adversarial face synthesis. In: 2020 IEEE International Joint Conference on Biometrics (IJCB). pp. 1-10. IEEE (2020)", + "17. Deng, J., Guo, J., Xue, N., Zafeiriou, S.: Arcface: Additive angular margin loss for deep face recognition. In: Proceedings of the IEEE/CVF conference on computer vision and pattern recognition. pp. 4690-4699 (2019)", + "18. Guo, C., Rana, M., Cisse, M., Van Der Maaten, L.: Countering adversarial images using input transformations. arXiv preprint arXiv:1711.00117 (2017)", + "19. Hu, S., Liu, X., Zhang, Y., Li, M., Zhang, L.Y., Jin, H., Wu, L.: Protecting facial privacy: generating adversarial identity masks via style-robust makeup transfer. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 15014-15023 (2022)", + "20. Huang, G.B., Ramesh, M., Berg, T., Learned-Miller, E.: Labeled faces in the wild: A database for studying face recognition in unconstrained environments. Tech. Rep. 07-49, University of Massachusetts, Amherst (October 2007)", + "21. Li, T., Lin, L.: Anonymousnet: Natural face de-identification with measurable privacy. In: Proceedings of the IEEE/CVF conference on computer vision and pattern recognition workshops. pp. 0–0 (2019)", + "22. Liu, L., Zhou, B., Zou, Z., Yeh, S.C., Zheng, L.: A smart unstaffed retail shop based on artificial intelligence and iot. In: 2018 IEEE 23rd International workshop on computer aided modeling and design of communication links and networks (CAMAD). pp. 1-4. IEEE (2018)", + "23. Lugaresi, C., Tang, J., Nash, H., McClanahan, C., Uboweja, E., Hays, M., Zhang, F., Chang, C.L., Yong, M.G., Lee, J., et al.: Mediapipe: A framework for building perception pipelines. arXiv preprint arXiv:1906.08172 (2019)", + "24. Meng, Q., Zhao, S., Huang, Z., Zhou, F.: Magface: A universal representation for face recognition and quality assessment. In: Proceedings of the IEEE/CVF conference on computer vision and pattern recognition. pp. 14225-14234 (2021)", + "25. Mirza, M., Osindero, S.: Conditional generative adversarial nets. arXiv preprint arXiv:1411.1784 (2014)", + "26. Ng, H.W., Winkler, S.: A data-driven approach to cleaning large face datasets. In: 2014 IEEE international conference on image processing (ICIP). pp. 343-347. IEEE (2014)", + "27. Partridge, D., Krzanowski, W.: Software diversity: practical statistics for its measurement and exploitation. Information and software technology 39(10), 707-717 (1997)", + "28. Pinto, N., Stone, Z., Zickler, T., Cox, D.: Scaling up biologically-inspired computer vision: A case study in unconstrained face recognition on facebook. In: CVPR 2011 WORKSHOPS. pp. 35-42. IEEE (2011)", + "29. Sahani, M., Nanda, C., Sahu, A.K., Pattnaik, B.: Web-based online embedded door access control and home security system based on face recognition. In: 2015 International Conference on Circuits, Power and Computing Technologies [ICCPCT-2015]. pp. 1-6. IEEE (2015)", + "30. Schroff, F., Kalenichenko, D., Philbin, J.: Facenet: A unified embedding for face recognition and clustering. In: Proceedings of the IEEE conference on computer vision and pattern recognition. pp. 815-823 (2015)" + ], + "bbox": [ + 215, + 146, + 785, + 840 + ], + "page_idx": 15 + }, + { + "type": "page_number", + "text": "16", + "bbox": [ + 217, + 114, + 235, + 126 + ], + "page_idx": 15 + }, + { + "type": "header", + "text": "K.-H. Chow et al.", + "bbox": [ + 271, + 114, + 390, + 127 + ], + "page_idx": 15 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "31. Shan, S., Wenger, E., Zhang, J., Li, H., Zheng, H., Zhao, B.Y.: Fawkes: Protecting privacy against unauthorized deep learning models. In: Proceedings of the 29th USENIX Security Symposium (2020)", + "32. Stone, Z., Zickler, T., Darrell, T.: Autotagging facebook: Social network context improves photo annotation. In: 2008 IEEE computer society conference on computer vision and pattern recognition workshops. pp. 1-8. IEEE (2008)", + "33. Sun, Q., Tewari, A., Xu, W., Fritz, M., Theobalt, C., Schiele, B.: A hybrid model for identity obfuscation by face replacement. In: Proceedings of the European conference on computer vision (ECCV). pp. 553-569 (2018)", + "34. Wang, M., Deng, W.: Deep face recognition: A survey. Neurocomputing 429, 215-244 (2021)", + "35. Wang, Z., Bovik, A.C., Sheikh, H.R., Simoncelli, E.P.: Image quality assessment: from error visibility to structural similarity. IEEE transactions on image processing 13(4), 600-612 (2004)", + "36. Wenger, E., Shan, S., Zheng, H., Zhao, B.Y.: Sok: Anti-facial recognition technology. In: 2023 IEEE Symposium on Security and Privacy (SP). pp. 134-151 (2023)", + "37. Wu, Y., Liu, L., Xie, Z., Chow, K.H., Wei, W.: Boosting ensemble accuracy by revisiting ensemble diversity metrics. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 16469-16477 (2021)", + "38. Xu, W., Evans, D., Qi, Y.: Feature squeezing: Detecting adversarial examples in deep neural networks (2018)", + "39. Yang, X., Dong, Y., Pang, T., Su, H., Zhu, J., Chen, Y., Xue, H.: Towards face encryption by generating adversarial identity masks. In: Proceedings of the IEEE/CVF International Conference on Computer Vision. pp. 3897-3907 (2021)", + "40. Zhong, Y., Deng, W.: Opom: Customized invisible cloak towards face privacy protection. IEEE Transactions on Pattern Analysis and Machine Intelligence (2022)", + "41. Zhu, Z.A., Lu, Y.Z., Chiang, C.K.: Generating adversarial examples by makeup attacks on face recognition. In: 2019 IEEE International Conference on Image Processing (ICIP). pp. 2516-2520. IEEE (2019)" + ], + "bbox": [ + 212, + 146, + 787, + 550 + ], + "page_idx": 16 + }, + { + "type": "header", + "text": "Personalized Privacy Protection Mask Against Unauthorized FR", + "bbox": [ + 300, + 114, + 730, + 127 + ], + "page_idx": 16 + }, + { + "type": "page_number", + "text": "17", + "bbox": [ + 767, + 116, + 784, + 126 + ], + "page_idx": 16 + } +] \ No newline at end of file diff --git a/2024/Personalized Privacy Protection Mask Against Unauthorized Facial Recognition/40580728-0a4c-4c74-a026-5407ddf870b6_model.json b/2024/Personalized Privacy Protection Mask Against Unauthorized Facial Recognition/40580728-0a4c-4c74-a026-5407ddf870b6_model.json new file mode 100644 index 0000000000000000000000000000000000000000..641c485d60d2dcabf71114942fe8521ebeb51a12 --- /dev/null +++ b/2024/Personalized Privacy Protection Mask Against Unauthorized Facial Recognition/40580728-0a4c-4c74-a026-5407ddf870b6_model.json @@ -0,0 +1,2269 @@ +[ + [ + { + "type": "title", + "bbox": [ + 0.236, + 0.141, + 0.768, + 0.187 + ], + "angle": 0, + "content": "Personalized Privacy Protection Mask Against Unauthorized Facial Recognition" + }, + { + "type": "text", + "bbox": [ + 0.258, + 0.212, + 0.744, + 0.228 + ], + "angle": 0, + "content": "Ka-Ho Chow\\(^{1}\\), Sihao Hu\\(^{2}\\), Tiansheng Huang\\(^{2}\\), and Ling Liu\\(^{2}\\)" + }, + { + "type": "text", + "bbox": [ + 0.394, + 0.24, + 0.61, + 0.254 + ], + "angle": 0, + "content": "1 The University of Hong Kong" + }, + { + "type": "text", + "bbox": [ + 0.39, + 0.255, + 0.615, + 0.268 + ], + "angle": 0, + "content": "2 Georgia Institute of Technology" + }, + { + "type": "list", + "bbox": [ + 0.39, + 0.24, + 0.615, + 0.268 + ], + "angle": 0, + "content": null + }, + { + "type": "text", + "bbox": [ + 0.229, + 0.27, + 0.776, + 0.282 + ], + "angle": 0, + "content": "kachow@cs.hku.hk, {sihaohu,thuang374}@gatech.edu, lingliu@cc.gatech.edu" + }, + { + "type": "text", + "bbox": [ + 0.261, + 0.313, + 0.741, + 0.577 + ], + "angle": 0, + "content": "Abstract. Face recognition (FR) can be abused for privacy intrusion. Governments, private companies, or even individual attackers can collect facial images by web scraping to build an FR system identifying human faces without their consent. This paper introduces Chameleon, which learns to generate a user-centric personalized privacy protection mask, coined as P3-Mask, to protect facial images against unauthorized FR with three salient features. First, we use a cross-image optimization to generate one P3-Mask for each user instead of tailoring facial perturbation for each facial image of a user. It enables efficient and instant protection even for users with limited computing resources. Second, we incorporate a perceptibility optimization to preserve the visual quality of the protected facial images. Third, we strengthen the robustness of P3-Mask against unknown FR models by integrating focal diversity-optimized ensemble learning into the mask generation process. Extensive experiments on two benchmark datasets show that Chameleon outperforms three state-of-the-art methods with instant protection and minimal degradation of image quality. Furthermore, Chameleon enables cost-effective FR authorization using the P3-Mask as a personalized de-obfuscation key, and it demonstrates high resilience against adaptive adversaries." + }, + { + "type": "title", + "bbox": [ + 0.218, + 0.6, + 0.377, + 0.616 + ], + "angle": 0, + "content": "1 Introduction" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.629, + 0.788, + 0.705 + ], + "angle": 0, + "content": "Face recognition (FR) has long been investigated due to its potential for enhancing security and convenience in various domains [22,28,29]. Many pretrained FR models are available online [6,7]. Once a face database (a.k.a. the gallery) with facial images for each person of interest is provided, these pretrained models can be used to recognize them [34]." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.705, + 0.788, + 0.84 + ], + "angle": 0, + "content": "While empowering many life-enriching applications, FR can be abused to cause serious privacy issues [2]. Privacy intruders can build a face database of victims of interest from publicly available facial images on the Internet by web scraping (Figure 1 (top)). By utilizing this database, the adversary can perform unauthorized FR of individual users for stalking victims [3], intruding on victims' privacy by flooding targeted ads [1], or facilitating criminals to commit identity fraud [5]. This is a real threat. For example, companies like Clearview [4] and PimEyes [8] have collected billions of online images and can recognize millions of citizens without their consent." + } + ], + [ + { + "type": "page_number", + "bbox": [ + 0.218, + 0.116, + 0.23, + 0.127 + ], + "angle": 0, + "content": "2" + }, + { + "type": "header", + "bbox": [ + 0.272, + 0.115, + 0.393, + 0.128 + ], + "angle": 0, + "content": "K.-H. Chow et al." + }, + { + "type": "image", + "bbox": [ + 0.269, + 0.143, + 0.744, + 0.322 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.214, + 0.331, + 0.788, + 0.388 + ], + "angle": 0, + "content": "Fig. 1: (Top) Without protection, the privacy intruder can build a face database by web scraping and identify the unknown face. (Bottom) Chameleon learns the facial signature of the user (protectee) to generate a P3-Mask, which can be applied to protect any facial images before sharing them online against unauthorized FR." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.417, + 0.788, + 0.735 + ], + "angle": 0, + "content": "Such a facial privacy threat has motivated the development of anti-FR technologies [36], which allow users (protectees) to preprocess their facial images before sharing them online. We argue that a competitive anti-FR solution should possess the following four properties. First, from the robustness perspective, the protected facial images should not be matched by an FR model to a query (probe) image with a correct identity (e.g., Bradley Cooper is recognized as Anne Hathaway in Figure 1 (bottom)), including those FR models possibly unknown during the preprocessing but are used by the privacy intruder. Second, from the efficiency perspective, the protection should be done instantly on any device with or without AI accelerators to maintain user experience and support equitable access to privacy protection. Third, from the perception usability perspective, the protected image should (i) preserve the visual quality rather than using excessive noise and (ii) be visually recognizable by humans to be the same person as the one in the original unprotected photo. Finally, the service usability for authorized FR providers is another important property. Existing approaches treat all FR models as unauthorized. We argue that anti-FR solutions should also allow users to grant FR permission to authorized models cost-effectively. This user-initiated de-obfuscation capability is critical to those, e.g., who publish their photos on a social media platform with privacy protection and, at the same time, wish to authorize the platform to perform some FR services (e.g., face tagging [32]) without sending and storing multiple versions of the same photo." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.735, + 0.788, + 0.842 + ], + "angle": 0, + "content": "Based on the above objectives, we introduce Chameleon, a user-centric facial privacy protection system with three contributions. First, we develop an optimization algorithm to construct a Personalized Privacy Protection mask, coined as P3-Mask, for each user. The P3-Mask of a user can be used to protect any facial images of the same user with minimal impact on image quality, including those facial images unseen during the P3-Mask generation process. Second, it boosts the robustness of P3-Mask against unknown FR models by a princi" + } + ], + [ + { + "type": "header", + "bbox": [ + 0.302, + 0.115, + 0.732, + 0.129 + ], + "angle": 0, + "content": "Personalized Privacy Protection Mask Against Unauthorized FR" + }, + { + "type": "page_number", + "bbox": [ + 0.775, + 0.117, + 0.785, + 0.127 + ], + "angle": 0, + "content": "3" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.147, + 0.788, + 0.329 + ], + "angle": 0, + "content": "pled approach, leveraging ensemble learning with models of high focal diversity. The per-user P3-Mask provides higher robustness against unknown FR models while preserving perception usability and the service usability for authorized FR providers. Chameleon users can utilize their own P3-Mask to obfuscate their facial images prior to public release, and the same P3-Mask can be used as a personalized de-obfuscation key to grant authorized FR service providers the ability to restore the facial signature of photos for correct recognition. Third, we conduct extensive experiments on two FR benchmark datasets [20,26] to analyze the shortcomings of state-of-the-art anti-FR solutions [31,39,40] and show that Chameleon can better protect against unknown FR models with a high success rate. The protection is lightweight and in real-time. It remains effective under adaptive privacy intruders deploying various strategies to counter Chameleon." + }, + { + "type": "title", + "bbox": [ + 0.216, + 0.35, + 0.388, + 0.366 + ], + "angle": 0, + "content": "2 Related Work" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.38, + 0.816, + 0.456 + ], + "angle": 0, + "content": "Existing anti-FR approaches can be broadly classified into two categories. Synthesis-based approaches [11,16,19,21,33,41] use generative adversarial networks (GANs) [25] to synthesize a face to replace the original one for protection. While the synthetic faces look realistic, they appear to be strangers, not recognizable even by the users themselves, failing to meet the perception usability requirement." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.457, + 0.789, + 0.699 + ], + "angle": 0, + "content": "Chameleon falls into the second category, which applies small changes to facial images [12]. Fawkes [31] formulates an untargeted attack to push the facial image away from the original location in the embedding space; TIP-IM [39] improves the image quality with MMD [9]; LowKey [10] improves the robustness under the image processing pipeline in an end-to-end ML system by incorporating it into the optimization process. These techniques can preserve the visual identity of protected faces. However, they both require iterative optimization for each image. Even for a GPU server, it can take over 100 seconds to perturb one facial image (Section 6.3). In contrast, Chameleon provides instant protection on any device, with protection even stronger than those spending minutes to find the best perturbation for each image. OPOM [40] attempts to improve efficiency by finding the embedding subspace enclosing facial images of the same person and generates a privacy protection mask to push any images away from it. However, it fails to maintain protection effectiveness when image processing operations are applied to protected images (Section 6.1), and the mask degrades the image quality much more significantly than Chameleon (Section 6.3)." + }, + { + "type": "title", + "bbox": [ + 0.216, + 0.72, + 0.345, + 0.735 + ], + "angle": 0, + "content": "3 Overview" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.75, + 0.788, + 0.841 + ], + "angle": 0, + "content": "To build an FR system with a pretrained model, the owner first collects a gallery dataset (face database) \\(\\pmb{D}\\), which includes the facial images of individuals of interest. Then, the pretrained FR model \\(F\\) is used to map each gallery image \\(\\tilde{\\pmb{x}} \\in \\pmb{D}\\) to the embedding \\(F(\\tilde{\\pmb{x}})\\), and those embeddings should form clusters corresponding to different people. When a facial image \\(\\pmb{x}\\) of an unknown identity, called the probe image, is given, the FR system can use the same FR model \\(F\\) to" + } + ], + [ + { + "type": "page_number", + "bbox": [ + 0.218, + 0.116, + 0.23, + 0.127 + ], + "angle": 0, + "content": "4" + }, + { + "type": "header", + "bbox": [ + 0.272, + 0.115, + 0.393, + 0.128 + ], + "angle": 0, + "content": "K.-H. Chow et al." + }, + { + "type": "image", + "bbox": [ + 0.218, + 0.144, + 0.785, + 0.26 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.215, + 0.272, + 0.788, + 0.318 + ], + "angle": 0, + "content": "Fig. 2: An overview of Chameleon's two-phase workflow for offline learning to generate P3-Mask and online protection with P3-Mask for personalized facial signature masking. The P3-Mask can protect any facial images of the same person without further learning." + }, + { + "type": "text", + "bbox": [ + 0.215, + 0.345, + 0.789, + 0.391 + ], + "angle": 0, + "content": "map it to an embedding \\( F(\\pmb{x}) \\) and use the identity of the nearest gallery image to be the identity of the unknown person. The identity, \\( \\mathcal{FR}(\\pmb{x};F,\\pmb{D}) \\), of the probe image \\( \\pmb{x} \\) given the FR model \\( F \\) and the gallery dataset \\( \\pmb{D} \\) is defined as:" + }, + { + "type": "equation", + "bbox": [ + 0.347, + 0.399, + 0.787, + 0.423 + ], + "angle": 0, + "content": "\\[\n\\mathcal {F R} (\\boldsymbol {x}; F, \\boldsymbol {\\mathcal {D}}) = \\mathcal {I} \\left(\\underset {\\tilde {\\boldsymbol {x}} \\in \\boldsymbol {\\mathcal {D}}} {\\arg \\min } \\operatorname {D I S T} (F (\\boldsymbol {x}), F (\\tilde {\\boldsymbol {x}}))\\right), \\tag {1}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.215, + 0.431, + 0.789, + 0.461 + ], + "angle": 0, + "content": "where \\(\\mathcal{I}(\\tilde{\\boldsymbol{x}})\\) denotes the identity of the gallery image \\(\\tilde{\\boldsymbol{x}}\\), which is known to the FR system, and \\(\\mathrm{DIST}(\\cdot ,\\cdot)\\) is a distance function such as Euclidean distance." + }, + { + "type": "text", + "bbox": [ + 0.215, + 0.462, + 0.789, + 0.613 + ], + "angle": 0, + "content": "Threat Model. We consider a threat model where a privacy intruder conducts web scraping to collect images containing citizens' faces. Web scraping is large-scale and untargeted. Many citizens are included in the face database, and the privacy intruder may not know them. Given a probe image taken by, e.g., a stalker's camera, the privacy intruder uses an FR model on the face database to search for matched image(s). Then, the privacy intruder can analyze the associated metadata, such as the web page from where it was downloaded or even the exact identity. The goal of Chameleon is to allow the user to preprocess her facial images before posting them online. Even if scraped and included in the face database, they will not be matched to a probe image of her." + }, + { + "type": "title", + "bbox": [ + 0.216, + 0.634, + 0.418, + 0.65 + ], + "angle": 0, + "content": "3.1 Chameleon Design" + }, + { + "type": "text", + "bbox": [ + 0.215, + 0.659, + 0.789, + 0.704 + ], + "angle": 0, + "content": "Chameleon performs preprocessing of the user's photos before sharing them online by applying the user-specific P3-Mask. This mask is generated offline for a Chameleon user. Figure 2 provides an overview of Chameleon's workflow." + }, + { + "type": "text", + "bbox": [ + 0.215, + 0.705, + 0.789, + 0.841 + ], + "angle": 0, + "content": "Offline Learning Phase. Chameleon learns the unique facial signature of a user from a few facial images of her using the P3-Mask generator (Section 4), as shown in Figure 2 ①. P3-Mask is designed to apply directly to any facial images of the same user, including those unseen during offline learning, and is robust against lossy image processing operations in an end-to-end ML system without compromising the image quality. We use a focal diversity-optimized ensemble learning method to find a team of FR models such that the P3-Mask generated against them has strong robustness and is effective in countering other FR models that are unknown during offline learning." + } + ], + [ + { + "type": "header", + "bbox": [ + 0.301, + 0.115, + 0.733, + 0.131 + ], + "angle": 0, + "content": "Personalized Privacy Protection Mask Against Unauthorized FR" + }, + { + "type": "page_number", + "bbox": [ + 0.775, + 0.117, + 0.785, + 0.127 + ], + "angle": 0, + "content": "5" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.147, + 0.788, + 0.27 + ], + "angle": 0, + "content": "Online Protection Phase. After the offline learning, ② the P3-Mask can be sent to the user's local device (e.g., the mobile phone). For a photo of the user to be protected, ③ a lightweight face detection model, such as MediaPipe [23], can be used by the Chameleon user running on the user's device to locate the face region and ④ instantly protect it with the P3-Mask before the user shares it online ⑤. The user can also authorize some trusted entities to conduct FR on their shared photos protected with P3-Mask. The user-specific de-obfuscation key ⑥ allows the protected photos to be restored for authorized FR." + }, + { + "type": "title", + "bbox": [ + 0.216, + 0.29, + 0.456, + 0.307 + ], + "angle": 0, + "content": "4 P3-Mask Generation" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.321, + 0.789, + 0.504 + ], + "angle": 0, + "content": "For a user \\(\\mathcal{P}\\), Chameleon generates a P3-Mask by learning the facial signature \\(\\mathcal{M}_{\\mathcal{P}}\\) from a set of facial images the user provides. Several key factors need to be accomplished. (1) We need to promote cross-image protection as an optimization objective to generate the most representative facial signature applicable to protect any facial images of the user \\(\\mathcal{P}\\), including those unknown ones during the generation process (Section 4.1). (2) We need to control the amount of perturbation introduced by the P3-Mask. When applied to an unprotected image by removing the learned facial signature, it ensures minimal perception loss to preserve the visual quality (Section 4.2). (3) We need to keep the protected image far away from the original image in the embedding space. Multiple FR models should be used to generate alternative embedding representations to enhance generalizability against unknown models (Section 4.3)." + }, + { + "type": "title", + "bbox": [ + 0.216, + 0.524, + 0.571, + 0.54 + ], + "angle": 0, + "content": "4.1 Cross-image Protection Optimization" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.548, + 0.788, + 0.655 + ], + "angle": 0, + "content": "The cross-image protection capability comes from iterative learning on a set of training images the user provides. The idea is to keep fine-tuning the P3-Mask so that it can offer protection simultaneously to training images while preserving image quality. At the \\( t \\)-th iteration, it samples a mini-batch \\( \\mathcal{B} \\) from the training dataset \\( \\Omega \\) containing photos of user \\( \\mathcal{P} \\). Then, we find the modification to the P3-Mask that can lead to better protection on each image in the mini-batch with better visual quality with the following operations:" + }, + { + "type": "equation", + "bbox": [ + 0.275, + 0.662, + 0.787, + 0.695 + ], + "angle": 0, + "content": "\\[\n\\mathcal {M} _ {\\mathcal {P}} ^ {t + 1} = \\operatorname {C L I P} _ {[ - \\epsilon , \\epsilon ]} \\left(\\mathcal {M} _ {\\mathcal {P}} ^ {t} - \\eta \\operatorname {S I G N} \\left(\\frac {1}{| \\boldsymbol {\\mathcal {B}} |} \\sum_ {\\mathcal {X} \\in \\boldsymbol {\\mathcal {B}}} \\nabla_ {\\mathcal {M} _ {\\mathcal {P}} ^ {t}} \\mathcal {L} \\left(\\mathcal {X}, \\mathcal {M} _ {\\mathcal {P}} ^ {t}; \\boldsymbol {\\mathcal {T}}, \\omega\\right)\\right)\\right). \\tag {2}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.701, + 0.787, + 0.734 + ], + "angle": 0, + "content": "Specifically, for each image \\(\\mathcal{X} \\in \\mathcal{B}\\), we use the P3-Mask learned up to the current iteration, \\(\\mathcal{M}_{\\mathcal{P}}^t\\), to protect it and compute the loss designed with dual goals:" + }, + { + "type": "equation", + "bbox": [ + 0.295, + 0.741, + 0.787, + 0.758 + ], + "angle": 0, + "content": "\\[\n\\mathcal {L} (\\mathcal {X}, \\boldsymbol {\\mathcal {M}} _ {\\mathcal {P}} ^ {t}; \\boldsymbol {\\mathcal {T}}, \\omega) = \\mathcal {L} _ {\\text {P r o t e c t}} (\\mathcal {X}, \\boldsymbol {\\mathcal {M}} _ {\\mathcal {P}} ^ {t}; \\boldsymbol {\\mathcal {T}}) + \\mathcal {L} _ {\\text {P e r c e p t}} (\\mathcal {X}, \\boldsymbol {\\mathcal {M}} _ {\\mathcal {P}} ^ {t}; \\omega). \\tag {3}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.765, + 0.788, + 0.843 + ], + "angle": 0, + "content": "It uses \\(\\mathcal{L}_{\\mathrm{Protect}}\\) to learn how to adjust the P3-Mask \\(\\mathcal{M}_{\\mathcal{P}}^{t}\\) from the previous iteration to better protect against a team of pre-selected FR models \\(\\pmb{\\tau}\\) and \\(\\mathcal{L}_{\\mathrm{Percept}}\\) to preserve the image quality controlled by \\(\\omega\\). We use the SIGN of the gradients to update the P3-Mask with a learning rate \\(\\eta\\). A clipping function \\(\\mathrm{CLIP}_{[-\\epsilon, \\epsilon]}\\) is applied to ensure the changes made by the P3-Mask on the facial" + } + ], + [ + { + "type": "page_number", + "bbox": [ + 0.218, + 0.116, + 0.23, + 0.127 + ], + "angle": 0, + "content": "6" + }, + { + "type": "header", + "bbox": [ + 0.272, + 0.115, + 0.393, + 0.128 + ], + "angle": 0, + "content": "K.-H. Chow et al." + }, + { + "type": "image", + "bbox": [ + 0.232, + 0.143, + 0.775, + 0.282 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.215, + 0.293, + 0.785, + 0.322 + ], + "angle": 0, + "content": "Fig. 3: The iterative generation of P3-Mask for a user. Chameleon goes through multiple facial images of the same user to optimize a P3-Mask with ML pipeline awareness." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.35, + 0.785, + 0.395 + ], + "angle": 0, + "content": "image are bounded by \\(\\epsilon\\). Such cross-image protection is not limited to those in the training dataset \\(\\Omega\\) but also unseen images of the same user. We next discuss the design of \\(\\mathcal{L}_{\\mathrm{Protect}}\\), and \\(\\mathcal{L}_{\\mathrm{Percept}}\\) will be detailed in Section 4.2." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.395, + 0.787, + 0.517 + ], + "angle": 0, + "content": "For privacy protection, P3-Mask maximizes the distance between the protected and unprotected images in the embedding spaces of a team \\(\\mathcal{T}\\) of pre-selected FR models. We incorporate image processing operations into the optimization to avoid those lossy operations degrading P3-Mask's protection. Consider the raw training image \\(\\mathcal{X}\\) in the mini-batch in Figure 3. We first conduct face detection to produce the cropped face \\(\\mathcal{FD}(\\mathcal{X})\\). Quantization and resizing are executed on the P3-Mask to match the resolution of the cropped face, denoted by \\(\\mathcal{QR}(\\mathcal{M}_{\\mathcal{P}}^t)\\). Then, we can apply the mask to the face for protection:" + }, + { + "type": "equation", + "bbox": [ + 0.33, + 0.529, + 0.786, + 0.546 + ], + "angle": 0, + "content": "\\[\n\\Theta (\\mathcal {X}; \\boldsymbol {\\mathcal {M}} _ {\\mathcal {P}} ^ {t}) = \\Psi \\left(\\operatorname {C L I P} _ {[ 0, 2 5 5 ]} \\left(\\mathcal {F D} (\\mathcal {X}) - \\mathcal {Q R} \\left(\\boldsymbol {\\mathcal {M}} _ {\\mathcal {P}} ^ {t}\\right)\\right)\\right), \\tag {4}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.548, + 0.785, + 0.594 + ], + "angle": 0, + "content": "where \\(\\varPsi\\) is a padding and resizing function to meet the input resolution requirement of an FR model (e.g., \\((112\\times 112)\\) in ArcFace [17]). Note that to simplify the notation, we removed certain arguments without causing confusion." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.594, + 0.787, + 0.655 + ], + "angle": 0, + "content": "The protection optimization in P3-Mask maximizes the average arccosine distance [17], \\(\\mathrm{ArcCos}\\), between the embedding of the protected face \\(F(\\Theta(\\mathcal{X};\\mathcal{M}_{\\mathcal{P}}^t))\\) and the embedding of its unprotected counterpart \\(F(\\Psi(\\mathcal{FD}(\\mathcal{X})))\\) extracted by every FR model \\(F \\in \\mathcal{T}\\) in the team:" + }, + { + "type": "equation", + "bbox": [ + 0.254, + 0.661, + 0.787, + 0.693 + ], + "angle": 0, + "content": "\\[\n\\mathcal {L} _ {\\mathrm {P r o t e c t}} (\\mathcal {X}, \\boldsymbol {\\mathcal {M}} _ {\\mathcal {P}} ^ {t}; \\boldsymbol {\\mathcal {T}}) = \\frac {- 1}{| \\boldsymbol {\\mathcal {T}} |} \\sum_ {F \\in \\boldsymbol {\\mathcal {T}}} \\operatorname {A r c C o s} (F (\\Psi (\\mathcal {F D} (\\mathcal {X}))), F (\\Theta (\\mathcal {X}; \\boldsymbol {\\mathcal {M}} _ {\\mathcal {P}} ^ {t}))). \\quad (5)\n\\]" + }, + { + "type": "title", + "bbox": [ + 0.216, + 0.711, + 0.489, + 0.727 + ], + "angle": 0, + "content": "4.2 Perceptibility Optimization" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.735, + 0.788, + 0.842 + ], + "angle": 0, + "content": "Preserving the visual quality of the facial image is necessary to maintain the usability of the protected image in practice. Hence, we incorporate a perceptibility optimization term in Equation 3. The idea is to minimize the perceptual difference between unprotected and protected images while learning a user's facial signature, such that removing the facial signature from a given facial image will have minimal impact on the visual quality of the protected version of the original image. We use the structural similarity (SSIM) [35] to capture perceptual" + } + ], + [ + { + "type": "header", + "bbox": [ + 0.302, + 0.115, + 0.733, + 0.129 + ], + "angle": 0, + "content": "Personalized Privacy Protection Mask Against Unauthorized FR" + }, + { + "type": "page_number", + "bbox": [ + 0.775, + 0.116, + 0.785, + 0.127 + ], + "angle": 0, + "content": "7" + }, + { + "type": "image", + "bbox": [ + 0.242, + 0.144, + 0.432, + 0.208 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.257, + 0.209, + 0.414, + 0.22 + ], + "angle": 0, + "content": "(a) Two-model: 28 Options" + }, + { + "type": "image", + "bbox": [ + 0.242, + 0.224, + 0.422, + 0.287 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.252, + 0.288, + 0.42, + 0.299 + ], + "angle": 0, + "content": "(b) Three-model: 56 Options" + }, + { + "type": "image", + "bbox": [ + 0.479, + 0.144, + 0.768, + 0.285 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.459, + 0.297, + 0.787, + 0.354 + ], + "angle": 0, + "content": "Fig. 5: Out of 28 options of two-model ensembles from a pool of eight models, the ensemble selected by our approach (green) leads to much better protection than a randomly selected one (red)." + }, + { + "type": "image_caption", + "bbox": [ + 0.216, + 0.311, + 0.453, + 0.354 + ], + "angle": 0, + "content": "Fig.4: Focal diversity provides a strong indicator of protection effectiveness for selecting an ensemble." + }, + { + "type": "text", + "bbox": [ + 0.216, + 0.383, + 0.769, + 0.399 + ], + "angle": 0, + "content": "differences, as it has been shown to align with the human perception system:" + }, + { + "type": "equation", + "bbox": [ + 0.228, + 0.406, + 0.785, + 0.438 + ], + "angle": 0, + "content": "\\[\n\\mathcal {L} _ {\\text {P e r c e p t}} \\left(\\mathcal {X}, \\boldsymbol {\\mathcal {M}} _ {\\mathcal {P}} ^ {t}; \\omega\\right) = \\lambda_ {\\text {S S I M}} \\max \\left[ \\frac {1 - \\operatorname {S S I M} \\left(\\Psi \\left(\\boldsymbol {\\mathcal {F D}} (\\mathcal {X})\\right) , \\Theta \\left(\\mathcal {X} ; \\boldsymbol {\\mathcal {M}} _ {\\mathcal {P}} ^ {t}\\right)\\right)}{2} - \\omega , 0 \\right]. \\tag {6}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.216, + 0.445, + 0.785, + 0.522 + ], + "angle": 0, + "content": "The parameters, \\(\\omega\\) and \\(\\lambda_{\\mathrm{SSIM}}\\), enable two features. \\(\\omega\\) controls the SSIM degradation. Any SSIM degradation greater than \\(2\\omega\\) will cause the term to be non-zero, making the optimization adjust the P3-Mask reducing its impact on image quality. \\(\\lambda_{\\mathrm{SSIM}}\\) balances the importance of privacy protection and perceptibility. We use dynamic scheduling [31] such that it will be adjusted automatically." + }, + { + "type": "title", + "bbox": [ + 0.216, + 0.543, + 0.585, + 0.559 + ], + "angle": 0, + "content": "4.3 Focal Diversity Ensemble Optimization" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.569, + 0.787, + 0.734 + ], + "angle": 0, + "content": "For each original image and its protected version by removing the facial signature learned up to the current iteration, we use two FR models in Figure 3 to generate two embeddings, which serve as alternative channels to learn a high-quality facial signature. Choosing FR models that best complement each other can achieve better protection than randomly selected FR models. We use an ensemble optimization method to select the best \\( k \\)-model ensemble among a pool of \\( N \\) FR models. The key idea is to find an ensemble with members making decorrelated mistakes. To quantify this behavior, we leverage the focal diversity framework [13,14,37]. As shown in Figure 4, the higher the focal diversity of the ensemble (a green dot), the stronger the P3-Mask will be in protecting against other FR models not used during the P3-Mask generation process." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.735, + 0.788, + 0.841 + ], + "angle": 0, + "content": "Given a collection of \\(N\\) FR models \\(\\{F_1,\\dots,F_N\\}\\), we first identify the negative samples for each FR model \\(F\\) by locating validation images that \\(F\\) fails to recognize their true identity. To rank ensembles of size \\(S\\), we enumerate all \\(\\binom{N}{S}\\) combinations. For each combination (ensemble) \\(\\pmb{\\tau}\\), we consider each member to be the focal model \\(F_{\\mathrm{focal}}\\) and use its negative samples to statistically estimate the level of negative correlation \\(\\lambda_{\\mathrm{focal}}(\\pmb{\\tau};F_{\\mathrm{focal}})\\) between \\(F_{\\mathrm{focal}}\\) and the remaining models in \\(\\pmb{\\tau}\\), computed by measuring the degree of disagreements using the" + } + ], + [ + { + "type": "page_number", + "bbox": [ + 0.218, + 0.116, + 0.23, + 0.127 + ], + "angle": 0, + "content": "8" + }, + { + "type": "header", + "bbox": [ + 0.272, + 0.115, + 0.393, + 0.128 + ], + "angle": 0, + "content": "K.-H. Chow et al." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.147, + 0.788, + 0.284 + ], + "angle": 0, + "content": "generalized non-pairwise measure [27]. The same procedure is repeated by considering each member in the ensemble as the focal model, and the focal diversity of the ensemble \\(\\pmb{\\mathcal{T}}\\) is finalized as \\(d_{\\mathrm{focal}}(\\pmb {\\mathcal{T}}) = \\frac{1}{S}\\sum_{F_{\\mathrm{focal}}\\in \\pmb{\\mathcal{T}}}[1 - \\lambda_{\\mathrm{focal}}(\\pmb {\\mathcal{T}};F_{\\mathrm{focal}})]\\). Given a team size \\(S\\), the ensemble with the highest focal diversity can be deployed. As shown in Figure 5, the selected two-model team (green bar) is indeed the one leading to the strongest protection among all 28 options, which is over \\(20\\%\\) stronger than a randomly selected ensemble (red bar). Note that we only need to analyze the failures of individual FR models, which is significantly more efficient than generating P3-Mask for each option and conducting evaluation." + }, + { + "type": "title", + "bbox": [ + 0.216, + 0.306, + 0.651, + 0.323 + ], + "angle": 0, + "content": "5 Online Image Protection and Refinement" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.337, + 0.787, + 0.369 + ], + "angle": 0, + "content": "The user \\(\\mathcal{P}\\) obtains her P3-Mask \\(\\mathcal{M}_{\\mathcal{P}}\\) from Chameleon. Given an image \\(\\mathcal{X}\\), she can obfuscate her facial identity on her device by applying the P3-Mask:" + }, + { + "type": "equation", + "bbox": [ + 0.331, + 0.379, + 0.786, + 0.394 + ], + "angle": 0, + "content": "\\[\n\\operatorname {M A S K} (\\mathcal {X}; \\boldsymbol {\\mathcal {M}} _ {\\mathcal {P}}) = \\operatorname {C L I P} _ {[ 0, 2 5 5 ]} (\\mathcal {F D} (\\mathcal {X}) - \\mathcal {Q R} (\\boldsymbol {\\mathcal {M}} _ {\\mathcal {P}})), \\tag {7}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.403, + 0.787, + 0.433 + ], + "angle": 0, + "content": "which is the protected facial region to be put in the original image \\(\\mathcal{X}\\) to produce the protected version \\(\\mathcal{X}'\\) for sharing." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.433, + 0.787, + 0.509 + ], + "angle": 0, + "content": "Chameleon supports authorizing trusted third parties to conduct FR on a user's facial images in a space-time efficient manner without transmitting and storing multiple versions of the same photo (one protected for public sharing and one unprotected for internal FR). By sharing the P3-Mask \\(\\mathcal{M}_{\\mathcal{P}}\\) as the key, the third parties can de-obfuscate the facial signature by unmasking:" + }, + { + "type": "equation", + "bbox": [ + 0.317, + 0.52, + 0.786, + 0.535 + ], + "angle": 0, + "content": "\\[\n\\operatorname {U N M A S K} \\left(\\mathcal {X} ^ {\\prime}; \\boldsymbol {\\mathcal {M}} _ {\\mathcal {P}}\\right) = \\operatorname {C L I P} _ {[ 0, 2 5 5 ]} \\left(\\mathcal {F D} \\left(\\mathcal {X} ^ {\\prime}\\right) + \\mathcal {Q R} \\left(\\boldsymbol {\\mathcal {M}} _ {\\mathcal {P}}\\right)\\right). \\tag {8}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.544, + 0.787, + 0.605 + ], + "angle": 0, + "content": "In Section 6.2, we will show two properties of this process: (i) the signature-restored photos can be used for FR with no accuracy degradation, and (ii) the protection can only be done by the P3-Mask of the same person in the facial image, and the restoration is successful only if the same key is used." + }, + { + "type": "title", + "bbox": [ + 0.216, + 0.628, + 0.494, + 0.646 + ], + "angle": 0, + "content": "6 Experimental Evaluation" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.659, + 0.788, + 0.841 + ], + "angle": 0, + "content": "We conduct experiments on FaceScrub [26] and LFW [20]. To provide a detailed analysis, ten celebrities are randomly selected as users on FaceScrub (Table 1). For each user, we split their facial images into three parts: (i) \\(10\\%\\) are used as probe images (2nd column), (ii) \\(70\\%\\) are used as gallery images and are used by Chameleon to train the mask (3rd column), and (iii) \\(20\\%\\) are also used as gallery images but unseen during the training process (4th column). The last part is crucial to evaluate the protection of unseen images of the user. In total, we have 123 facial images with \"unknown identities.\" Following relevant works [31, 40], all facial images of other people are included in the gallery dataset as noise. The same splitting method is also used for LFW. By default, the results reported in this paper focus on FaceScrub due to similar observations on LFW. The source code is available at https://github.com/git-disl/Chameleon." + } + ], + [ + { + "type": "header", + "bbox": [ + 0.301, + 0.115, + 0.733, + 0.129 + ], + "angle": 0, + "content": "Personalized Privacy Protection Mask Against Unauthorized FR" + }, + { + "type": "page_number", + "bbox": [ + 0.775, + 0.117, + 0.785, + 0.127 + ], + "angle": 0, + "content": "9" + }, + { + "type": "table_caption", + "bbox": [ + 0.216, + 0.145, + 0.495, + 0.173 + ], + "angle": 0, + "content": "Table 1: Ten example celebrities (users) on FaceScrub for analysis." + }, + { + "type": "table_caption", + "bbox": [ + 0.502, + 0.145, + 0.786, + 0.173 + ], + "angle": 0, + "content": "Table 2: The collection of publicly available pretrained FR models." + }, + { + "type": "table", + "bbox": [ + 0.218, + 0.186, + 0.497, + 0.34 + ], + "angle": 0, + "content": "
NameProbe ImagesGallery Images
SeenUnseenTotal
Morena Baccarin118223105
Bradley Cooper128925114
America Ferrera1611332145
Gerard Butler118224106
Eva Longoria129126117
Melissa Egan128724111
Kim Catrall139627123
Allison Janney128926115
Roma Downey117822100
Steve Carell139628124
Others (Noise)///48368
Total Number of Gallery Images:49528
" + }, + { + "type": "table", + "bbox": [ + 0.507, + 0.186, + 0.782, + 0.294 + ], + "angle": 0, + "content": "
Model IDNeural Arch.Training DatasetFR Acc.
EN-MCEfficientNetMS-Celeb-1M94.94%
RN50-GLResNet50Glint360K96.68%
RN50-MCResNet50MS-Celeb-1M89.25%
RN50-VFResNet50VGG-Face294.30%
RN50-WFResNet50WebFace600K96.72%
RN18-MCResNet18MS-Celeb-1M82.64%
RN34-MCResNet34MS-Celeb-1M84.49%
RN100-MCResNet100MS-Celeb-1M91.85%
" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.366, + 0.788, + 0.609 + ], + "angle": 0, + "content": "Public clouds (e.g., Azure) now require manual approval to use their FR services. Hence, we believe that privacy intruders will opt for deploying pretrained FR models, as many high-quality ones are available on the Internet and can be used out of the box. Due to the large number of possible FR algorithms and architectures, we first focus on the eight FR models listed in Table 2, which are based on ArcFace [17], the state-of-the-art FR algorithm, with varying neural architectures and training datasets. Chameleon can be easily extended to incorporate any FR models by simply adding them to the collection. Nonetheless, in Section 6.4, we will show that thanks to the focal diversity-optimized teaming, the P3-Mask generated from a collection of ArcFace-only models can also be effective in protecting against privacy intruders using models of other FR algorithms, such as FaceNet [30] and MagFace [24]. By default, we use the two-model team (EN-MC, RN50-GL) with the highest focal diversity (Section 4.3). The P3-Mask is trained for 50 epochs on NVIDIA RTX 2080 SUPER GPU with \\(\\eta = 0.001\\), \\(|\\mathcal{B}| = 4\\), \\(\\epsilon = 0.063\\), and \\(\\omega = 0.03\\). We provide details for reproducibility and additional results in the appendix." + }, + { + "type": "title", + "bbox": [ + 0.216, + 0.628, + 0.461, + 0.642 + ], + "angle": 0, + "content": "6.1 Protection Effectiveness" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.65, + 0.788, + 0.698 + ], + "angle": 0, + "content": "Table 3 shows the personalized mask for seven users on FaceScrub (other users are available in the appendix). Different users have distinct facial signatures (P3-Mask) learned by Chameleon. We apply these masks to their respective gallery" + }, + { + "type": "table_caption", + "bbox": [ + 0.216, + 0.72, + 0.787, + 0.748 + ], + "angle": 0, + "content": "Table 3: The P3-Mask (rescaled) for seven users on FaceScrub. Chameleon learns the distinct facial signature for each of them." + }, + { + "type": "table", + "bbox": [ + 0.218, + 0.761, + 0.786, + 0.837 + ], + "angle": 0, + "content": "
M. BaccarinB. CooperA. FerreraG. ButlerE. LongoriaM. EganK. Cattrall
" + } + ], + [ + { + "type": "page_number", + "bbox": [ + 0.218, + 0.116, + 0.236, + 0.127 + ], + "angle": 0, + "content": "10" + }, + { + "type": "header", + "bbox": [ + 0.272, + 0.115, + 0.393, + 0.127 + ], + "angle": 0, + "content": "K.-H. Chow et al." + }, + { + "type": "image", + "bbox": [ + 0.223, + 0.144, + 0.499, + 0.216 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.293, + 0.217, + 0.429, + 0.228 + ], + "angle": 0, + "content": "(a) Dataset: FaceScrub" + }, + { + "type": "image", + "bbox": [ + 0.508, + 0.145, + 0.772, + 0.216 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.59, + 0.217, + 0.697, + 0.228 + ], + "angle": 0, + "content": "(b) Dataset: LFW" + }, + { + "type": "image_caption", + "bbox": [ + 0.215, + 0.239, + 0.785, + 0.268 + ], + "angle": 0, + "content": "Fig. 6: Protection effectiveness on two benchmarks using Chameleon, OPOM [40], TIP-IM [39], and Fawkes [31] against FR models unknown to them." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.305, + 0.785, + 0.382 + ], + "angle": 0, + "content": "images and test the FR accuracy using probe images. All probe images can be correctly identified when no protection mechanism is used. Under Chameleon, the probe images should be matched to an incorrect identity, resulting in a low FR accuracy. Hence, we define an evaluation metric, \\(\\underline{\\text{Protection Success Rate}}\\) (PSR), to be \\((100 - \\text{FR Accuracy})\\) reported in percentages." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.384, + 0.787, + 0.581 + ], + "angle": 0, + "content": "Chameleon can protect against unknown FR models. We compare the cross-model protection of Chameleon with OPOM [40], the only anti-FR approach generating one privacy protection mask for each person as Chameleon, and both TIP-IM [39] and Fawkes [31], which conduct per-image optimization. For a fair comparison, all methods are set with the same perturbation budget. Figure 6 summarizes the results on both datasets. All five FR models are unknown to both protection mechanisms. We make two observations. First, Chameleon consistently outperforms OPOM by a large margin. We found that OPOM masks do not transfer well, especially considering the end-to-end ML pipeline. Its PSR can drop by \\(33\\%\\) because of those lossy operations. Second, Chameleon can be as competitive as those methods conducting per-image optimization. TIP-IM and Fawkes spend minutes to optimize the protection for each image. Still, Chameleon outperforms them and can be applied instantly." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.584, + 0.787, + 0.842 + ], + "angle": 0, + "content": "Table 4 reports the detailed PSR for each user Chameleon achieves when the privacy intruder uses different FR models (2nd to 9th columns). We make two observations. First, when the privacy intruder uses an FR model known to the generation process, a PSR of \\(100\\%\\) can be consistently achieved (2nd to 3rd columns). Second, Chameleon can protect against unknown FR models (4th to 9th columns). Even when the FR model used by the intruder is trained on a different dataset (i.e., RN50-VF and RN50-WF) or with a different neural architecture (i.e., RN18-MC, RN34-MC, and RN100-MC) than any FR models known by Chameleon, it still provides a PSR over \\(90.65\\%\\), meaning that facial images of a user are matched to gallery images of a different person. To demonstrate such a mismatch, in Table 5, we show the probe images from two users and their most similar gallery images found by four unknown FR models with two settings: (i) the \"Unprotected\" scenario where no one employs protection and (ii) the \"Chameleon\" scenario where the intruder scraped photos protected by our solution. Taking M. Baccarin as an example, the most similar gallery images found in the unprotected scenario belong to her. In contrast, when Chameleon is used, the same probe image is misidentified, e.g., as L. Hartley by RN50-MC." + } + ], + [ + { + "type": "header", + "bbox": [ + 0.301, + 0.115, + 0.733, + 0.129 + ], + "angle": 0, + "content": "Personalized Privacy Protection Mask Against Unauthorized FR" + }, + { + "type": "page_number", + "bbox": [ + 0.768, + 0.117, + 0.784, + 0.127 + ], + "angle": 0, + "content": "11" + }, + { + "type": "table_caption", + "bbox": [ + 0.215, + 0.145, + 0.788, + 0.187 + ], + "angle": 0, + "content": "Table 4: Using the two-model team with high focal diversity (2nd and 3rd columns), Chameleon can offer privacy protection even when the privacy intruder uses FR models unseen during the mask generation process (4th to 9th columns)." + }, + { + "type": "table", + "bbox": [ + 0.218, + 0.2, + 0.787, + 0.362 + ], + "angle": 0, + "content": "
NameProtection Success Rate - PSR (%)
EN-MCRN50-GLRN50-MCRN50-VFRN50-WFRN18-MCRN34-MCRN100-MCMeanStd
M. Baccarin100.00100.00100.0072.7390.91100.0090.91100.0094.329.64
B. Cooper100.00100.0083.33100.00100.00100.0091.6791.6795.836.30
A. Ferrera100.00100.0093.7575.00100.00100.0093.75100.0095.318.68
G. Butler100.00100.0090.91100.00100.0090.9190.91100.0096.594.70
E. Longoria100.00100.00100.00100.00100.00100.00100.00100.00100.000.00
M. Egan100.00100.00100.00100.00100.00100.00100.0091.6798.962.95
K. Cattrall100.00100.00100.0084.62100.00100.00100.00100.0098.085.44
A. Janney100.00100.0091.67100.0091.67100.00100.00100.0097.923.86
R. Downey100.00100.00100.0072.7390.91100.0090.91100.0094.329.64
S. Carell100.00100.0084.6292.3192.3192.3176.92100.0092.318.22
Mean100.00100.0093.5290.6596.5897.4194.4297.42
Std0.000.006.4011.164.434.187.414.15
" + }, + { + "type": "table_caption", + "bbox": [ + 0.215, + 0.375, + 0.788, + 0.403 + ], + "angle": 0, + "content": "Table 5: The most similar gallery images on FaceScrub found by different FR models unknown to Chameleon." + }, + { + "type": "table", + "bbox": [ + 0.218, + 0.416, + 0.795, + 0.569 + ], + "angle": 0, + "content": "
Probe\nImageThe Most Similar Gallery Image & Its Identity Found by Unseen FR Models
RN50-MCRN50-VFRN34-MCRN100-MC
UnprotectedChameleonUnprotectedChameleonUnprotectedChameleonUnprotectedChameleon
M. BaccarinIdentifiedProtectedIdentifiedProtectedIdentifiedProtectedIdentifiedProtected
M. BaccarinL. HartleyM. BaccarinL. HartleyM. BaccarinC. ElectraM. BaccarinL. Hartley
B. CooperIdentifiedProtectedIdentifiedProtectedIdentifiedProtectedIdentifiedProtected
B. CooperP. WalkerB. CooperM. VartanB. CooperM. VartanB. CooperM. Vartan
" + }, + { + "type": "title", + "bbox": [ + 0.215, + 0.595, + 0.434, + 0.61 + ], + "angle": 0, + "content": "6.2 FR Service Usability" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.617, + 0.789, + 0.769 + ], + "angle": 0, + "content": "Chameleon allows users to grant FR permission to trusted third parties by sharing their P3-Mask to de-obfuscate the protected image. There is no need to transmit and store an unprotected photo for internal use and a protected one for public visibility, doubling network and storage costs. Table 6 shows the results in FR accuracy. First, using the correct mask (i.e., the one used to protect the images) to reverse the protection can restore FR accuracy. Taking M. Baccarin as an example, the FR accuracy increases from \\(6.82\\%\\) before unmasking (b) to \\(100\\%\\) after unmasking (c). Second, using an incorrect mask for unmasking cannot restore the FR accuracy and can lead to even worse performance. It drops from \\(6.82\\%\\) to \\(1.14\\%\\) after unmasking (d)." + }, + { + "type": "title", + "bbox": [ + 0.215, + 0.788, + 0.471, + 0.803 + ], + "angle": 0, + "content": "6.3 Protection Cost Analysis" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.81, + 0.787, + 0.84 + ], + "angle": 0, + "content": "Speed and Resources. Figure 7 compares Chameleon with OPOM, TIP-IM, and Fawkes regarding the protection time per face, compute, and storage costs." + } + ], + [ + { + "type": "page_number", + "bbox": [ + 0.218, + 0.116, + 0.236, + 0.127 + ], + "angle": 0, + "content": "12" + }, + { + "type": "header", + "bbox": [ + 0.272, + 0.115, + 0.393, + 0.128 + ], + "angle": 0, + "content": "K.-H. Chow et al." + }, + { + "type": "table_caption", + "bbox": [ + 0.216, + 0.145, + 0.788, + 0.203 + ], + "angle": 0, + "content": "Table 6: The protection by Chameleon can reduce FR accuracy against FR models (b), but the user can provide the P3-Mask used to protect her images as the key for the authorized third party to reverse the protection process (c), which leads to restored FR performance. However, an incorrect key cannot restore FR and may worsen it (d)." + }, + { + "type": "table", + "bbox": [ + 0.217, + 0.213, + 0.787, + 0.354 + ], + "angle": 0, + "content": "
ScenarioFace Recognition Accuracy (%)
EN-MCRN50-GLRN50-MCRN50-VFRN50-WFRN18-MCRN34-MCRN100-MCMeanStd
User: M. Baccarin
(a) No Protection100.00100.00100.00100.00100.00100.00100.00100.00100.000.00
(b) Protected0.000.009.0918.189.099.090.009.096.826.43
(c) Correctly Unmasked100.00100.00100.00100.00100.00100.00100.00100.00100.000.00
(d) Incorrectly Unmasked0.000.000.000.009.090.000.000.001.143.21
User: B. Cooper
(a) No Protection100.00100.00100.00100.00100.00100.00100.00100.00100.000.00
(b) Protected0.000. 0016.670.000.000.008.338.334.176.30
(c) Correctly Unmasked100.00100.00100.00100.00100.00100.00100.00100.00100.000.00
(d) Incorrectly Unmasked0.000.000. 000.000.000.000.000.000.000.00
" + }, + { + "type": "image", + "bbox": [ + 0.253, + 0.368, + 0.35, + 0.45 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.353, + 0.368, + 0.449, + 0.45 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.454, + 0.368, + 0.549, + 0.45 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.555, + 0.368, + 0.651, + 0.45 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.655, + 0.368, + 0.752, + 0.45 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.214, + 0.46, + 0.788, + 0.503 + ], + "angle": 0, + "content": "Fig. 7: Chameleon offers much better privacy protection than OPOM without sacrificing speed and resource usages. It can protect instantly and require negligible compute resources as OPOM, while TIP-IM and Fawkes are slow and costly." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.536, + 0.788, + 0.658 + ], + "angle": 0, + "content": "Chameleon and OPOM produce masks that are applicable to any facial image of the same user. The protection can be completed in 0.0076 seconds. Note that Chameleon is significantly more effective than OPOM, as described in Figure 6. Instead, TIP-IM and Fawkes require per-image optimization and take 105.12 seconds to complete the protection with GPUs for acceleration and storage for FR models. Chameleon only needs to conduct simple arithmetic operations. Only the P3-Mask needs to be stored on the user device. Hence, it can be deployed with instant protection even on an edge device with weak computing power." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.66, + 0.789, + 0.841 + ], + "angle": 0, + "content": "Image Quality. Chameleon can well preserve image quality. Table 7 provides two examples for four users, contrasting the facial images with no protection (3rd and 5th columns) with the ones with their facial signature removed by Chameleon (4th and 6th columns). They capture different variations of facial images: M. Baccarin's images have different lighting conditions, B. Cooper's have different face sizes, E. Longoria's have different postures, and G. Bulter's have different expressions. The protected images are visually similar to the unprotected counterparts, but they will not be matched to the clean images of the corresponding person. Indeed, Chameleon can generate higher-quality images than existing methods. As reported in Table 8, Chameleon not only offers better protection (Figure 6) with a similar cost (Figure 7) as OPOM but also achieves much better image quality measured in SSIM (0.9493 vs 0.8839)." + } + ], + [ + { + "type": "header", + "bbox": [ + 0.301, + 0.115, + 0.733, + 0.129 + ], + "angle": 0, + "content": "Personalized Privacy Protection Mask Against Unauthorized FR" + }, + { + "type": "page_number", + "bbox": [ + 0.768, + 0.116, + 0.786, + 0.127 + ], + "angle": 0, + "content": "13" + }, + { + "type": "table_caption", + "bbox": [ + 0.216, + 0.145, + 0.788, + 0.174 + ], + "angle": 0, + "content": "Table 7: The P3-Mask can be applied to any facial im- Table 8: Chameleon offers ages of the same person while preserving image quality. much better protection while" + }, + { + "type": "table", + "bbox": [ + 0.223, + 0.186, + 0.584, + 0.393 + ], + "angle": 0, + "content": "
UserMaskSample Image 1Sample Image 2
No ProtectionProtectedNo ProtectionProtected
M. Baccarin
B. Cooper
E. Longoria
G. Image
" + }, + { + "type": "table", + "bbox": [ + 0.613, + 0.2, + 0.766, + 0.402 + ], + "angle": 0, + "content": "
Example
Original SSIM: 1.0000
Chameleon SSIM: 0.9493
OPOM SSIM: 0.8839
TIP-IM SSIM: 0.8850
Fawkes SSIM: 0.9612
" + }, + { + "type": "title", + "bbox": [ + 0.216, + 0.433, + 0.515, + 0.449 + ], + "angle": 0, + "content": "6.4 Focal Diversity-based Teaming" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.462, + 0.788, + 0.584 + ], + "angle": 0, + "content": "Chameleon automatically selects a high-quality team for deployment with a specified budget. Table 9 compares the detailed PSR using the most diverse and the least diverse teams with (a) two and (b) three models. In both cases, the most diverse team outperforms the least diverse one, which is only effective when the privacy intruder uses the FR model known in the team. The selection of high-quality teams is non-trivial because we observe that composing a team of FR models with the highest FR accuracy is not always a good option, and a team of models having different neural architectures is not always the top priority [13]." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.584, + 0.787, + 0.615 + ], + "angle": 0, + "content": "In Figure 8, we further show that the P3-Mask generated from a carefully-chosen team can protect against FR models of unknown algorithms. Specifically," + }, + { + "type": "table_caption", + "bbox": [ + 0.214, + 0.643, + 0.788, + 0.685 + ], + "angle": 0, + "content": "Table 9: The most diverse and the least diverse teams with (a) two or (b) three models identified by our focal diversity-based teaming method. The most diverse teams offer significantly better protection in terms of protection success rates." + }, + { + "type": "table", + "bbox": [ + 0.218, + 0.697, + 0.794, + 0.838 + ], + "angle": 0, + "content": "
Protection TeamProtection Success Rate - PSR (%)
EN-MCRN50-GLRN50-MCRN50-VFRN50-WFRN18-MCRN34-MCRN100-MCMeanStd
(a) Two-model Teams
Most Diverse:(EN-MC, RN50-GL)100.00100.0093.5290.6596.5897.4194.4297.4296.253.23
Least Diverse:(RN18-MC, RN34-MC)11.3860.16100.0055.2873.17100.00100.0080.4972.5628.54
(b) Three-model Teams
Most Diverse:(EN-MC, RN50-GL, RN50-WF)100.00100.00100.00100.0092.6893.50100.00100.0098.273.00
Least Diverse:(RN18-MC, RN34-MC, RN50-MC)2.4478.86100.0052.0384.55100.00100.0093.5076.4231.82
" + } + ], + [ + { + "type": "page_number", + "bbox": [ + 0.218, + 0.116, + 0.236, + 0.127 + ], + "angle": 0, + "content": "14" + }, + { + "type": "header", + "bbox": [ + 0.272, + 0.115, + 0.393, + 0.127 + ], + "angle": 0, + "content": "K.-H. Chow et al." + }, + { + "type": "image", + "bbox": [ + 0.23, + 0.144, + 0.452, + 0.232 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.216, + 0.237, + 0.465, + 0.279 + ], + "angle": 0, + "content": "Fig. 8: The most diverse teams allow Chameleon to be effective against FR models of unknown algorithms." + }, + { + "type": "image", + "bbox": [ + 0.477, + 0.151, + 0.78, + 0.232 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.47, + 0.237, + 0.788, + 0.279 + ], + "angle": 0, + "content": "Fig. 9: The PSR of Chameleon against adaptive intruders using different strategies to wash out the patterns in protected images." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.322, + 0.788, + 0.398 + ], + "angle": 0, + "content": "we consider the most and the least diverse three-model teams in Table 9(b). Even though both teams include only FR models based on ArcFace [17], the most diverse one can effectively protect against FR models based on FaceNet [30] or MagFace [24]. The protection effectiveness can be further strengthened by incorporating more FR models into the collection." + }, + { + "type": "title", + "bbox": [ + 0.215, + 0.422, + 0.606, + 0.438 + ], + "angle": 0, + "content": "6.5 Chameleon Against Adaptive Adversaries" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.449, + 0.788, + 0.554 + ], + "angle": 0, + "content": "An adaptive adversary may run Chameleon to produce the P3-Mask for each person and de-obfuscate their photos. When probe images need to be identified, they will be matched against a \"clean\" face database. However, it is infeasible, as we show in Section 6.2, that the FR accuracy can only be restored when the masks used for obfuscation and de-obfuscation are identical. It is impossible to reproduce the same P3-Mask used by the user because it requires the same set of clean images and random states." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.555, + 0.788, + 0.646 + ], + "angle": 0, + "content": "Alternatively, an adaptive adversary may wash out patterns introduced by Chameleon before performing FR. We consider three popular lightweight and task-agnostic methods studied in the adversarial example domain [18] in Figure 9, including (a) JPEG compression [15], (b) Gaussian Filter, and (d) Median Filter [38]. The results show that their effectiveness is limited on Chameleon. The consideration of the end-to-end ML pipeline increases the robustness of P3-Mask." + }, + { + "type": "title", + "bbox": [ + 0.216, + 0.671, + 0.37, + 0.687 + ], + "angle": 0, + "content": "7 Conclusions" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.704, + 0.789, + 0.841 + ], + "angle": 0, + "content": "We have presented Chameleon against unauthorized FR. Chameleon generates a P3-Mask for each user with cross-image and perceptibility optimizations to offer (i) instant and lightweight protection on any facial images of the same user and (ii) preservation of image quality. Also, we have shown that P3-Mask enables cost-effective de-obfuscation for authorizing FR services. Such an authorization process can only be conducted by the one with the identical mask used for protection. Chameleon also features a focal diversity-optimized teaming method to select a high-quality FR team to generate P3-Mask with strong robustness against unknown FR models." + } + ], + [ + { + "type": "header", + "bbox": [ + 0.301, + 0.115, + 0.732, + 0.129 + ], + "angle": 0, + "content": "Personalized Privacy Protection Mask Against Unauthorized FR" + }, + { + "type": "page_number", + "bbox": [ + 0.768, + 0.116, + 0.786, + 0.127 + ], + "angle": 0, + "content": "15" + }, + { + "type": "title", + "bbox": [ + 0.218, + 0.145, + 0.393, + 0.163 + ], + "angle": 0, + "content": "Acknowledgments" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.18, + 0.788, + 0.258 + ], + "angle": 0, + "content": "This research is partially sponsored by the NSF CISE grants 2302720, 2312758, 2038029, an IBM faculty award, and a grant from CISCO Edge AI program. It is part of the PhD dissertation of the first author, who graduated from Georgia Tech in Spring 2023. The first author acknowledges the support of the IBM PhD Fellowship in 2022-2023 and the support from the HKU-CS Start-up Fund." + }, + { + "type": "title", + "bbox": [ + 0.216, + 0.283, + 0.324, + 0.3 + ], + "angle": 0, + "content": "References" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.318, + 0.785, + 0.36 + ], + "angle": 0, + "content": "1. Brazilian retailer quizzed over facial recognition tech. https://www.zdnet.com/article/brazilian-retailer-quizzed-over-facial-recognition-tech/ (2019)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.361, + 0.785, + 0.416 + ], + "angle": 0, + "content": "2. Facial recognition in the united states: Privacy concerns and legal developments. https://www.asisonline.org/security-management-magazine/monthlyissues/security-technology/archive/2021/december/facial-recognition-in-the-us-privacy-concerns-and-legal-developments/ (2021)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.418, + 0.785, + 0.459 + ], + "angle": 0, + "content": "3. The secretive company that might end privacy as we know it. https://www.nytimes.com/2020/01/18/technology/clearview-privacy-facial-recognition.html (2021)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.46, + 0.564, + 0.473 + ], + "angle": 0, + "content": "4. Clearview AI. https://www.clearview.ai/ (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.475, + 0.785, + 0.502 + ], + "angle": 0, + "content": "5. Facial recognition and identity risk. https://www.equifax.co.uk/resources/identity-protection/facial-recognition-and-identity-risk.html (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.503, + 0.535, + 0.517 + ], + "angle": 0, + "content": "6. HuggingFace. https://huggingface.co (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.518, + 0.785, + 0.545 + ], + "angle": 0, + "content": "7. Insightface: 2d and 3d face analysis project. https://github.com/deepinsight/insightface (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.546, + 0.504, + 0.559 + ], + "angle": 0, + "content": "8. PimEyes. https://pimeyes.com/ (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.56, + 0.785, + 0.601 + ], + "angle": 0, + "content": "9. Borgwardt, K.M., Gretton, A., Rasch, M.J., Kriegel, H.P., Scholkopf, B., Smola, A.J.: Integrating structured biological data by kernel maximum mean discrepancy. Bioinformatics 22(14), e49-e57 (2006)" + }, + { + "type": "ref_text", + "bbox": [ + 0.22, + 0.602, + 0.785, + 0.657 + ], + "angle": 0, + "content": "10. Cherepanova, V., Goldblum, M., Foley, H., Duan, S., Dickerson, J., Taylor, G., Goldstein, T.: Lowkey: Leveraging adversarial attacks to protect social media users from facial recognition. In: International Conference on Learning Representations (ICLR) (2021)" + }, + { + "type": "ref_text", + "bbox": [ + 0.22, + 0.658, + 0.785, + 0.713 + ], + "angle": 0, + "content": "11. Choi, Y., Choi, M., Kim, M., Ha, J.W., Kim, S., Choo, J.: Stargan: Unified generative adversarial networks for multi-domain image-to-image translation. In: Proceedings of the IEEE conference on computer vision and pattern recognition. pp. 8789-8797 (2018)" + }, + { + "type": "ref_text", + "bbox": [ + 0.22, + 0.714, + 0.785, + 0.756 + ], + "angle": 0, + "content": "12. Chow, K.H., Hu, S., Huang, T., Ilhan, F., Wei, W., Liu, L.: Diversity-driven privacy protection masks against unauthorized face recognition. Proceedings on Privacy Enhancing Technologies 4, 381-392 (2024)" + }, + { + "type": "ref_text", + "bbox": [ + 0.22, + 0.757, + 0.785, + 0.799 + ], + "angle": 0, + "content": "13. Chow, K.H., Liu, L.: Robust object detection fusion against deception. In: Proceedings of the 27th ACM SIGKDD Conference on Knowledge Discovery & Data Mining. pp. 2703-2713 (2021)" + }, + { + "type": "ref_text", + "bbox": [ + 0.22, + 0.799, + 0.785, + 0.84 + ], + "angle": 0, + "content": "14. Chow, K.H., Liu, L.: Boosting object detection ensembles with error diversity. In: 2022 IEEE International Conference on Data Mining (ICDM). pp. 903-908. IEEE (2022)" + }, + { + "type": "list", + "bbox": [ + 0.22, + 0.318, + 0.785, + 0.84 + ], + "angle": 0, + "content": null + } + ], + [ + { + "type": "page_number", + "bbox": [ + 0.218, + 0.116, + 0.236, + 0.127 + ], + "angle": 0, + "content": "16" + }, + { + "type": "header", + "bbox": [ + 0.272, + 0.115, + 0.392, + 0.128 + ], + "angle": 0, + "content": "K.-H. Chow et al." + }, + { + "type": "ref_text", + "bbox": [ + 0.218, + 0.147, + 0.787, + 0.204 + ], + "angle": 0, + "content": "15. Das, N., Shanbhogue, M., Chen, S.T., Hohman, F., Li, S., Chen, L., Kounavis, M.E., Chau, D.H.: Shield: Fast, practical defense and vaccination for deep learning usingJPEG compression. In: Proceedings of the 24th ACM SIGKDD International Conference on Knowledge Discovery & Data Mining. pp. 196-204 (2018)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.205, + 0.786, + 0.233 + ], + "angle": 0, + "content": "16. Deb, D., Zhang, J., Jain, A.K.: Advfaces: Adversarial face synthesis. In: 2020 IEEE International Joint Conference on Biometrics (IJCB). pp. 1-10. IEEE (2020)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.234, + 0.787, + 0.275 + ], + "angle": 0, + "content": "17. Deng, J., Guo, J., Xue, N., Zafeiriou, S.: Arcface: Additive angular margin loss for deep face recognition. In: Proceedings of the IEEE/CVF conference on computer vision and pattern recognition. pp. 4690-4699 (2019)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.276, + 0.787, + 0.303 + ], + "angle": 0, + "content": "18. Guo, C., Rana, M., Cisse, M., Van Der Maaten, L.: Countering adversarial images using input transformations. arXiv preprint arXiv:1711.00117 (2017)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.304, + 0.787, + 0.359 + ], + "angle": 0, + "content": "19. Hu, S., Liu, X., Zhang, Y., Li, M., Zhang, L.Y., Jin, H., Wu, L.: Protecting facial privacy: generating adversarial identity masks via style-robust makeup transfer. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 15014-15023 (2022)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.36, + 0.787, + 0.402 + ], + "angle": 0, + "content": "20. Huang, G.B., Ramesh, M., Berg, T., Learned-Miller, E.: Labeled faces in the wild: A database for studying face recognition in unconstrained environments. Tech. Rep. 07-49, University of Massachusetts, Amherst (October 2007)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.403, + 0.787, + 0.445 + ], + "angle": 0, + "content": "21. Li, T., Lin, L.: Anonymousnet: Natural face de-identification with measurable privacy. In: Proceedings of the IEEE/CVF conference on computer vision and pattern recognition workshops. pp. 0–0 (2019)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.445, + 0.787, + 0.5 + ], + "angle": 0, + "content": "22. Liu, L., Zhou, B., Zou, Z., Yeh, S.C., Zheng, L.: A smart unstaffed retail shop based on artificial intelligence and iot. In: 2018 IEEE 23rd International workshop on computer aided modeling and design of communication links and networks (CAMAD). pp. 1-4. IEEE (2018)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.501, + 0.787, + 0.543 + ], + "angle": 0, + "content": "23. Lugaresi, C., Tang, J., Nash, H., McClanahan, C., Uboweja, E., Hays, M., Zhang, F., Chang, C.L., Yong, M.G., Lee, J., et al.: Mediapipe: A framework for building perception pipelines. arXiv preprint arXiv:1906.08172 (2019)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.544, + 0.787, + 0.586 + ], + "angle": 0, + "content": "24. Meng, Q., Zhao, S., Huang, Z., Zhou, F.: Magface: A universal representation for face recognition and quality assessment. In: Proceedings of the IEEE/CVF conference on computer vision and pattern recognition. pp. 14225-14234 (2021)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.587, + 0.787, + 0.614 + ], + "angle": 0, + "content": "25. Mirza, M., Osindero, S.: Conditional generative adversarial nets. arXiv preprint arXiv:1411.1784 (2014)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.615, + 0.787, + 0.657 + ], + "angle": 0, + "content": "26. Ng, H.W., Winkler, S.: A data-driven approach to cleaning large face datasets. In: 2014 IEEE international conference on image processing (ICIP). pp. 343-347. IEEE (2014)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.658, + 0.787, + 0.699 + ], + "angle": 0, + "content": "27. Partridge, D., Krzanowski, W.: Software diversity: practical statistics for its measurement and exploitation. Information and software technology 39(10), 707-717 (1997)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.7, + 0.787, + 0.741 + ], + "angle": 0, + "content": "28. Pinto, N., Stone, Z., Zickler, T., Cox, D.: Scaling up biologically-inspired computer vision: A case study in unconstrained face recognition on facebook. In: CVPR 2011 WORKSHOPS. pp. 35-42. IEEE (2011)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.742, + 0.787, + 0.798 + ], + "angle": 0, + "content": "29. Sahani, M., Nanda, C., Sahu, A.K., Pattnaik, B.: Web-based online embedded door access control and home security system based on face recognition. In: 2015 International Conference on Circuits, Power and Computing Technologies [ICCPCT-2015]. pp. 1-6. IEEE (2015)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.799, + 0.787, + 0.841 + ], + "angle": 0, + "content": "30. Schroff, F., Kalenichenko, D., Philbin, J.: Facenet: A unified embedding for face recognition and clustering. In: Proceedings of the IEEE conference on computer vision and pattern recognition. pp. 815-823 (2015)" + }, + { + "type": "list", + "bbox": [ + 0.217, + 0.147, + 0.787, + 0.841 + ], + "angle": 0, + "content": null + } + ], + [ + { + "type": "header", + "bbox": [ + 0.302, + 0.115, + 0.732, + 0.128 + ], + "angle": 0, + "content": "Personalized Privacy Protection Mask Against Unauthorized FR" + }, + { + "type": "page_number", + "bbox": [ + 0.769, + 0.117, + 0.785, + 0.127 + ], + "angle": 0, + "content": "17" + }, + { + "type": "ref_text", + "bbox": [ + 0.214, + 0.147, + 0.788, + 0.189 + ], + "angle": 0, + "content": "31. Shan, S., Wenger, E., Zhang, J., Li, H., Zheng, H., Zhao, B.Y.: Fawkes: Protecting privacy against unauthorized deep learning models. In: Proceedings of the 29th USENIX Security Symposium (2020)" + }, + { + "type": "ref_text", + "bbox": [ + 0.214, + 0.19, + 0.788, + 0.232 + ], + "angle": 0, + "content": "32. Stone, Z., Zickler, T., Darrell, T.: Autotagging facebook: Social network context improves photo annotation. In: 2008 IEEE computer society conference on computer vision and pattern recognition workshops. pp. 1-8. IEEE (2008)" + }, + { + "type": "ref_text", + "bbox": [ + 0.214, + 0.232, + 0.788, + 0.273 + ], + "angle": 0, + "content": "33. Sun, Q., Tewari, A., Xu, W., Fritz, M., Theobalt, C., Schiele, B.: A hybrid model for identity obfuscation by face replacement. In: Proceedings of the European conference on computer vision (ECCV). pp. 553-569 (2018)" + }, + { + "type": "ref_text", + "bbox": [ + 0.214, + 0.273, + 0.788, + 0.3 + ], + "angle": 0, + "content": "34. Wang, M., Deng, W.: Deep face recognition: A survey. Neurocomputing 429, 215-244 (2021)" + }, + { + "type": "ref_text", + "bbox": [ + 0.214, + 0.3, + 0.788, + 0.342 + ], + "angle": 0, + "content": "35. Wang, Z., Bovik, A.C., Sheikh, H.R., Simoncelli, E.P.: Image quality assessment: from error visibility to structural similarity. IEEE transactions on image processing 13(4), 600-612 (2004)" + }, + { + "type": "ref_text", + "bbox": [ + 0.214, + 0.342, + 0.788, + 0.37 + ], + "angle": 0, + "content": "36. Wenger, E., Shan, S., Zheng, H., Zhao, B.Y.: Sok: Anti-facial recognition technology. In: 2023 IEEE Symposium on Security and Privacy (SP). pp. 134-151 (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.214, + 0.371, + 0.788, + 0.412 + ], + "angle": 0, + "content": "37. Wu, Y., Liu, L., Xie, Z., Chow, K.H., Wei, W.: Boosting ensemble accuracy by revisiting ensemble diversity metrics. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 16469-16477 (2021)" + }, + { + "type": "ref_text", + "bbox": [ + 0.214, + 0.412, + 0.788, + 0.438 + ], + "angle": 0, + "content": "38. Xu, W., Evans, D., Qi, Y.: Feature squeezing: Detecting adversarial examples in deep neural networks (2018)" + }, + { + "type": "ref_text", + "bbox": [ + 0.214, + 0.439, + 0.788, + 0.481 + ], + "angle": 0, + "content": "39. Yang, X., Dong, Y., Pang, T., Su, H., Zhu, J., Chen, Y., Xue, H.: Towards face encryption by generating adversarial identity masks. In: Proceedings of the IEEE/CVF International Conference on Computer Vision. pp. 3897-3907 (2021)" + }, + { + "type": "ref_text", + "bbox": [ + 0.214, + 0.481, + 0.788, + 0.508 + ], + "angle": 0, + "content": "40. Zhong, Y., Deng, W.: Opom: Customized invisible cloak towards face privacy protection. IEEE Transactions on Pattern Analysis and Machine Intelligence (2022)" + }, + { + "type": "ref_text", + "bbox": [ + 0.214, + 0.508, + 0.788, + 0.551 + ], + "angle": 0, + "content": "41. Zhu, Z.A., Lu, Y.Z., Chiang, C.K.: Generating adversarial examples by makeup attacks on face recognition. In: 2019 IEEE International Conference on Image Processing (ICIP). pp. 2516-2520. IEEE (2019)" + }, + { + "type": "list", + "bbox": [ + 0.214, + 0.147, + 0.788, + 0.551 + ], + "angle": 0, + "content": null + } + ] +] \ No newline at end of file diff --git a/2024/Personalized Privacy Protection Mask Against Unauthorized Facial Recognition/40580728-0a4c-4c74-a026-5407ddf870b6_origin.pdf b/2024/Personalized Privacy Protection Mask Against Unauthorized Facial Recognition/40580728-0a4c-4c74-a026-5407ddf870b6_origin.pdf new file mode 100644 index 0000000000000000000000000000000000000000..483612f9f902537d4c889370238fb7aea160b2f0 --- /dev/null +++ b/2024/Personalized Privacy Protection Mask Against Unauthorized Facial Recognition/40580728-0a4c-4c74-a026-5407ddf870b6_origin.pdf @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:1a699d3810ee9cbe77f0895b8d198f0b4a4be91bce2b3925b5bf741db0a69ec8 +size 2969023 diff --git a/2024/Personalized Privacy Protection Mask Against Unauthorized Facial Recognition/full.md b/2024/Personalized Privacy Protection Mask Against Unauthorized Facial Recognition/full.md new file mode 100644 index 0000000000000000000000000000000000000000..5d495e8d3280a0793b366fd05e94d6dd2925789b --- /dev/null +++ b/2024/Personalized Privacy Protection Mask Against Unauthorized Facial Recognition/full.md @@ -0,0 +1,298 @@ +# Personalized Privacy Protection Mask Against Unauthorized Facial Recognition + +Ka-Ho Chow $^{1}$ , Sihao Hu $^{2}$ , Tiansheng Huang $^{2}$ , and Ling Liu $^{2}$ + +1 The University of Hong Kong +2 Georgia Institute of Technology + +kachow@cs.hku.hk, {sihaohu,thuang374}@gatech.edu, lingliu@cc.gatech.edu + +Abstract. Face recognition (FR) can be abused for privacy intrusion. Governments, private companies, or even individual attackers can collect facial images by web scraping to build an FR system identifying human faces without their consent. This paper introduces Chameleon, which learns to generate a user-centric personalized privacy protection mask, coined as P3-Mask, to protect facial images against unauthorized FR with three salient features. First, we use a cross-image optimization to generate one P3-Mask for each user instead of tailoring facial perturbation for each facial image of a user. It enables efficient and instant protection even for users with limited computing resources. Second, we incorporate a perceptibility optimization to preserve the visual quality of the protected facial images. Third, we strengthen the robustness of P3-Mask against unknown FR models by integrating focal diversity-optimized ensemble learning into the mask generation process. Extensive experiments on two benchmark datasets show that Chameleon outperforms three state-of-the-art methods with instant protection and minimal degradation of image quality. Furthermore, Chameleon enables cost-effective FR authorization using the P3-Mask as a personalized de-obfuscation key, and it demonstrates high resilience against adaptive adversaries. + +# 1 Introduction + +Face recognition (FR) has long been investigated due to its potential for enhancing security and convenience in various domains [22,28,29]. Many pretrained FR models are available online [6,7]. Once a face database (a.k.a. the gallery) with facial images for each person of interest is provided, these pretrained models can be used to recognize them [34]. + +While empowering many life-enriching applications, FR can be abused to cause serious privacy issues [2]. Privacy intruders can build a face database of victims of interest from publicly available facial images on the Internet by web scraping (Figure 1 (top)). By utilizing this database, the adversary can perform unauthorized FR of individual users for stalking victims [3], intruding on victims' privacy by flooding targeted ads [1], or facilitating criminals to commit identity fraud [5]. This is a real threat. For example, companies like Clearview [4] and PimEyes [8] have collected billions of online images and can recognize millions of citizens without their consent. + +![](images/1173bc9750a62a9f482119b53580730f40508b1fd6750a03605ce1312da9b1fe.jpg) +Fig. 1: (Top) Without protection, the privacy intruder can build a face database by web scraping and identify the unknown face. (Bottom) Chameleon learns the facial signature of the user (protectee) to generate a P3-Mask, which can be applied to protect any facial images before sharing them online against unauthorized FR. + +Such a facial privacy threat has motivated the development of anti-FR technologies [36], which allow users (protectees) to preprocess their facial images before sharing them online. We argue that a competitive anti-FR solution should possess the following four properties. First, from the robustness perspective, the protected facial images should not be matched by an FR model to a query (probe) image with a correct identity (e.g., Bradley Cooper is recognized as Anne Hathaway in Figure 1 (bottom)), including those FR models possibly unknown during the preprocessing but are used by the privacy intruder. Second, from the efficiency perspective, the protection should be done instantly on any device with or without AI accelerators to maintain user experience and support equitable access to privacy protection. Third, from the perception usability perspective, the protected image should (i) preserve the visual quality rather than using excessive noise and (ii) be visually recognizable by humans to be the same person as the one in the original unprotected photo. Finally, the service usability for authorized FR providers is another important property. Existing approaches treat all FR models as unauthorized. We argue that anti-FR solutions should also allow users to grant FR permission to authorized models cost-effectively. This user-initiated de-obfuscation capability is critical to those, e.g., who publish their photos on a social media platform with privacy protection and, at the same time, wish to authorize the platform to perform some FR services (e.g., face tagging [32]) without sending and storing multiple versions of the same photo. + +Based on the above objectives, we introduce Chameleon, a user-centric facial privacy protection system with three contributions. First, we develop an optimization algorithm to construct a Personalized Privacy Protection mask, coined as P3-Mask, for each user. The P3-Mask of a user can be used to protect any facial images of the same user with minimal impact on image quality, including those facial images unseen during the P3-Mask generation process. Second, it boosts the robustness of P3-Mask against unknown FR models by a princi + +pled approach, leveraging ensemble learning with models of high focal diversity. The per-user P3-Mask provides higher robustness against unknown FR models while preserving perception usability and the service usability for authorized FR providers. Chameleon users can utilize their own P3-Mask to obfuscate their facial images prior to public release, and the same P3-Mask can be used as a personalized de-obfuscation key to grant authorized FR service providers the ability to restore the facial signature of photos for correct recognition. Third, we conduct extensive experiments on two FR benchmark datasets [20,26] to analyze the shortcomings of state-of-the-art anti-FR solutions [31,39,40] and show that Chameleon can better protect against unknown FR models with a high success rate. The protection is lightweight and in real-time. It remains effective under adaptive privacy intruders deploying various strategies to counter Chameleon. + +# 2 Related Work + +Existing anti-FR approaches can be broadly classified into two categories. Synthesis-based approaches [11,16,19,21,33,41] use generative adversarial networks (GANs) [25] to synthesize a face to replace the original one for protection. While the synthetic faces look realistic, they appear to be strangers, not recognizable even by the users themselves, failing to meet the perception usability requirement. + +Chameleon falls into the second category, which applies small changes to facial images [12]. Fawkes [31] formulates an untargeted attack to push the facial image away from the original location in the embedding space; TIP-IM [39] improves the image quality with MMD [9]; LowKey [10] improves the robustness under the image processing pipeline in an end-to-end ML system by incorporating it into the optimization process. These techniques can preserve the visual identity of protected faces. However, they both require iterative optimization for each image. Even for a GPU server, it can take over 100 seconds to perturb one facial image (Section 6.3). In contrast, Chameleon provides instant protection on any device, with protection even stronger than those spending minutes to find the best perturbation for each image. OPOM [40] attempts to improve efficiency by finding the embedding subspace enclosing facial images of the same person and generates a privacy protection mask to push any images away from it. However, it fails to maintain protection effectiveness when image processing operations are applied to protected images (Section 6.1), and the mask degrades the image quality much more significantly than Chameleon (Section 6.3). + +# 3 Overview + +To build an FR system with a pretrained model, the owner first collects a gallery dataset (face database) $\pmb{D}$ , which includes the facial images of individuals of interest. Then, the pretrained FR model $F$ is used to map each gallery image $\tilde{\pmb{x}} \in \pmb{D}$ to the embedding $F(\tilde{\pmb{x}})$ , and those embeddings should form clusters corresponding to different people. When a facial image $\pmb{x}$ of an unknown identity, called the probe image, is given, the FR system can use the same FR model $F$ to + +![](images/47589bdc1be02a49d4c4bac2fbfdf6c5df0a7220694b62be80e651b98c27e757.jpg) +Fig. 2: An overview of Chameleon's two-phase workflow for offline learning to generate P3-Mask and online protection with P3-Mask for personalized facial signature masking. The P3-Mask can protect any facial images of the same person without further learning. + +map it to an embedding $F(\pmb{x})$ and use the identity of the nearest gallery image to be the identity of the unknown person. The identity, $\mathcal{FR}(\pmb{x};F,\pmb{D})$ , of the probe image $\pmb{x}$ given the FR model $F$ and the gallery dataset $\pmb{D}$ is defined as: + +$$ +\mathcal {F R} (\boldsymbol {x}; F, \boldsymbol {\mathcal {D}}) = \mathcal {I} \left(\underset {\tilde {\boldsymbol {x}} \in \boldsymbol {\mathcal {D}}} {\arg \min } \operatorname {D I S T} (F (\boldsymbol {x}), F (\tilde {\boldsymbol {x}}))\right), \tag {1} +$$ + +where $\mathcal{I}(\tilde{\boldsymbol{x}})$ denotes the identity of the gallery image $\tilde{\boldsymbol{x}}$ , which is known to the FR system, and $\mathrm{DIST}(\cdot ,\cdot)$ is a distance function such as Euclidean distance. + +Threat Model. We consider a threat model where a privacy intruder conducts web scraping to collect images containing citizens' faces. Web scraping is large-scale and untargeted. Many citizens are included in the face database, and the privacy intruder may not know them. Given a probe image taken by, e.g., a stalker's camera, the privacy intruder uses an FR model on the face database to search for matched image(s). Then, the privacy intruder can analyze the associated metadata, such as the web page from where it was downloaded or even the exact identity. The goal of Chameleon is to allow the user to preprocess her facial images before posting them online. Even if scraped and included in the face database, they will not be matched to a probe image of her. + +# 3.1 Chameleon Design + +Chameleon performs preprocessing of the user's photos before sharing them online by applying the user-specific P3-Mask. This mask is generated offline for a Chameleon user. Figure 2 provides an overview of Chameleon's workflow. + +Offline Learning Phase. Chameleon learns the unique facial signature of a user from a few facial images of her using the P3-Mask generator (Section 4), as shown in Figure 2 ①. P3-Mask is designed to apply directly to any facial images of the same user, including those unseen during offline learning, and is robust against lossy image processing operations in an end-to-end ML system without compromising the image quality. We use a focal diversity-optimized ensemble learning method to find a team of FR models such that the P3-Mask generated against them has strong robustness and is effective in countering other FR models that are unknown during offline learning. + +Online Protection Phase. After the offline learning, ② the P3-Mask can be sent to the user's local device (e.g., the mobile phone). For a photo of the user to be protected, ③ a lightweight face detection model, such as MediaPipe [23], can be used by the Chameleon user running on the user's device to locate the face region and ④ instantly protect it with the P3-Mask before the user shares it online ⑤. The user can also authorize some trusted entities to conduct FR on their shared photos protected with P3-Mask. The user-specific de-obfuscation key ⑥ allows the protected photos to be restored for authorized FR. + +# 4 P3-Mask Generation + +For a user $\mathcal{P}$ , Chameleon generates a P3-Mask by learning the facial signature $\mathcal{M}_{\mathcal{P}}$ from a set of facial images the user provides. Several key factors need to be accomplished. (1) We need to promote cross-image protection as an optimization objective to generate the most representative facial signature applicable to protect any facial images of the user $\mathcal{P}$ , including those unknown ones during the generation process (Section 4.1). (2) We need to control the amount of perturbation introduced by the P3-Mask. When applied to an unprotected image by removing the learned facial signature, it ensures minimal perception loss to preserve the visual quality (Section 4.2). (3) We need to keep the protected image far away from the original image in the embedding space. Multiple FR models should be used to generate alternative embedding representations to enhance generalizability against unknown models (Section 4.3). + +# 4.1 Cross-image Protection Optimization + +The cross-image protection capability comes from iterative learning on a set of training images the user provides. The idea is to keep fine-tuning the P3-Mask so that it can offer protection simultaneously to training images while preserving image quality. At the $t$ -th iteration, it samples a mini-batch $\mathcal{B}$ from the training dataset $\Omega$ containing photos of user $\mathcal{P}$ . Then, we find the modification to the P3-Mask that can lead to better protection on each image in the mini-batch with better visual quality with the following operations: + +$$ +\mathcal {M} _ {\mathcal {P}} ^ {t + 1} = \operatorname {C L I P} _ {[ - \epsilon , \epsilon ]} \left(\mathcal {M} _ {\mathcal {P}} ^ {t} - \eta \operatorname {S I G N} \left(\frac {1}{| \boldsymbol {\mathcal {B}} |} \sum_ {\mathcal {X} \in \boldsymbol {\mathcal {B}}} \nabla_ {\mathcal {M} _ {\mathcal {P}} ^ {t}} \mathcal {L} \left(\mathcal {X}, \mathcal {M} _ {\mathcal {P}} ^ {t}; \boldsymbol {\mathcal {T}}, \omega\right)\right)\right). \tag {2} +$$ + +Specifically, for each image $\mathcal{X} \in \mathcal{B}$ , we use the P3-Mask learned up to the current iteration, $\mathcal{M}_{\mathcal{P}}^t$ , to protect it and compute the loss designed with dual goals: + +$$ +\mathcal {L} (\mathcal {X}, \boldsymbol {\mathcal {M}} _ {\mathcal {P}} ^ {t}; \boldsymbol {\mathcal {T}}, \omega) = \mathcal {L} _ {\text {P r o t e c t}} (\mathcal {X}, \boldsymbol {\mathcal {M}} _ {\mathcal {P}} ^ {t}; \boldsymbol {\mathcal {T}}) + \mathcal {L} _ {\text {P e r c e p t}} (\mathcal {X}, \boldsymbol {\mathcal {M}} _ {\mathcal {P}} ^ {t}; \omega). \tag {3} +$$ + +It uses $\mathcal{L}_{\mathrm{Protect}}$ to learn how to adjust the P3-Mask $\mathcal{M}_{\mathcal{P}}^{t}$ from the previous iteration to better protect against a team of pre-selected FR models $\pmb{\tau}$ and $\mathcal{L}_{\mathrm{Percept}}$ to preserve the image quality controlled by $\omega$ . We use the SIGN of the gradients to update the P3-Mask with a learning rate $\eta$ . A clipping function $\mathrm{CLIP}_{[-\epsilon, \epsilon]}$ is applied to ensure the changes made by the P3-Mask on the facial + +![](images/ebe22766176c65b7c394601e6e1b72ddd49e799792f30246dbf238fe9e49d7d5.jpg) +Fig. 3: The iterative generation of P3-Mask for a user. Chameleon goes through multiple facial images of the same user to optimize a P3-Mask with ML pipeline awareness. + +image are bounded by $\epsilon$ . Such cross-image protection is not limited to those in the training dataset $\Omega$ but also unseen images of the same user. We next discuss the design of $\mathcal{L}_{\mathrm{Protect}}$ , and $\mathcal{L}_{\mathrm{Percept}}$ will be detailed in Section 4.2. + +For privacy protection, P3-Mask maximizes the distance between the protected and unprotected images in the embedding spaces of a team $\mathcal{T}$ of pre-selected FR models. We incorporate image processing operations into the optimization to avoid those lossy operations degrading P3-Mask's protection. Consider the raw training image $\mathcal{X}$ in the mini-batch in Figure 3. We first conduct face detection to produce the cropped face $\mathcal{FD}(\mathcal{X})$ . Quantization and resizing are executed on the P3-Mask to match the resolution of the cropped face, denoted by $\mathcal{QR}(\mathcal{M}_{\mathcal{P}}^t)$ . Then, we can apply the mask to the face for protection: + +$$ +\Theta (\mathcal {X}; \boldsymbol {\mathcal {M}} _ {\mathcal {P}} ^ {t}) = \Psi \left(\operatorname {C L I P} _ {[ 0, 2 5 5 ]} \left(\mathcal {F D} (\mathcal {X}) - \mathcal {Q R} \left(\boldsymbol {\mathcal {M}} _ {\mathcal {P}} ^ {t}\right)\right)\right), \tag {4} +$$ + +where $\varPsi$ is a padding and resizing function to meet the input resolution requirement of an FR model (e.g., $(112\times 112)$ in ArcFace [17]). Note that to simplify the notation, we removed certain arguments without causing confusion. + +The protection optimization in P3-Mask maximizes the average arccosine distance [17], $\mathrm{ArcCos}$ , between the embedding of the protected face $F(\Theta(\mathcal{X};\mathcal{M}_{\mathcal{P}}^t))$ and the embedding of its unprotected counterpart $F(\Psi(\mathcal{FD}(\mathcal{X})))$ extracted by every FR model $F \in \mathcal{T}$ in the team: + +$$ +\mathcal {L} _ {\mathrm {P r o t e c t}} (\mathcal {X}, \boldsymbol {\mathcal {M}} _ {\mathcal {P}} ^ {t}; \boldsymbol {\mathcal {T}}) = \frac {- 1}{| \boldsymbol {\mathcal {T}} |} \sum_ {F \in \boldsymbol {\mathcal {T}}} \operatorname {A r c C o s} (F (\Psi (\mathcal {F D} (\mathcal {X}))), F (\Theta (\mathcal {X}; \boldsymbol {\mathcal {M}} _ {\mathcal {P}} ^ {t}))). \quad (5) +$$ + +# 4.2 Perceptibility Optimization + +Preserving the visual quality of the facial image is necessary to maintain the usability of the protected image in practice. Hence, we incorporate a perceptibility optimization term in Equation 3. The idea is to minimize the perceptual difference between unprotected and protected images while learning a user's facial signature, such that removing the facial signature from a given facial image will have minimal impact on the visual quality of the protected version of the original image. We use the structural similarity (SSIM) [35] to capture perceptual + +![](images/acfa6a0a3d7c3632dd4e74d91c52c2872adda959c3939046218ea3a1ed790434.jpg) +(a) Two-model: 28 Options + +![](images/aa4effa16cff7aa7d27ffbb5a95d34b8fd6a0665dae2bd748fceef68b42e3689.jpg) +(b) Three-model: 56 Options + +![](images/96640aaf397d553c48391063d68ae3e920c701982d7df828e5bf54eafdd97ce7.jpg) +Fig. 5: Out of 28 options of two-model ensembles from a pool of eight models, the ensemble selected by our approach (green) leads to much better protection than a randomly selected one (red). +Fig.4: Focal diversity provides a strong indicator of protection effectiveness for selecting an ensemble. + +differences, as it has been shown to align with the human perception system: + +$$ +\mathcal {L} _ {\text {P e r c e p t}} \left(\mathcal {X}, \boldsymbol {\mathcal {M}} _ {\mathcal {P}} ^ {t}; \omega\right) = \lambda_ {\text {S S I M}} \max \left[ \frac {1 - \operatorname {S S I M} \left(\Psi \left(\boldsymbol {\mathcal {F D}} (\mathcal {X})\right) , \Theta \left(\mathcal {X} ; \boldsymbol {\mathcal {M}} _ {\mathcal {P}} ^ {t}\right)\right)}{2} - \omega , 0 \right]. \tag {6} +$$ + +The parameters, $\omega$ and $\lambda_{\mathrm{SSIM}}$ , enable two features. $\omega$ controls the SSIM degradation. Any SSIM degradation greater than $2\omega$ will cause the term to be non-zero, making the optimization adjust the P3-Mask reducing its impact on image quality. $\lambda_{\mathrm{SSIM}}$ balances the importance of privacy protection and perceptibility. We use dynamic scheduling [31] such that it will be adjusted automatically. + +# 4.3 Focal Diversity Ensemble Optimization + +For each original image and its protected version by removing the facial signature learned up to the current iteration, we use two FR models in Figure 3 to generate two embeddings, which serve as alternative channels to learn a high-quality facial signature. Choosing FR models that best complement each other can achieve better protection than randomly selected FR models. We use an ensemble optimization method to select the best $k$ -model ensemble among a pool of $N$ FR models. The key idea is to find an ensemble with members making decorrelated mistakes. To quantify this behavior, we leverage the focal diversity framework [13,14,37]. As shown in Figure 4, the higher the focal diversity of the ensemble (a green dot), the stronger the P3-Mask will be in protecting against other FR models not used during the P3-Mask generation process. + +Given a collection of $N$ FR models $\{F_1,\dots,F_N\}$ , we first identify the negative samples for each FR model $F$ by locating validation images that $F$ fails to recognize their true identity. To rank ensembles of size $S$ , we enumerate all $\binom{N}{S}$ combinations. For each combination (ensemble) $\pmb{\tau}$ , we consider each member to be the focal model $F_{\mathrm{focal}}$ and use its negative samples to statistically estimate the level of negative correlation $\lambda_{\mathrm{focal}}(\pmb{\tau};F_{\mathrm{focal}})$ between $F_{\mathrm{focal}}$ and the remaining models in $\pmb{\tau}$ , computed by measuring the degree of disagreements using the + +generalized non-pairwise measure [27]. The same procedure is repeated by considering each member in the ensemble as the focal model, and the focal diversity of the ensemble $\pmb{\mathcal{T}}$ is finalized as $d_{\mathrm{focal}}(\pmb {\mathcal{T}}) = \frac{1}{S}\sum_{F_{\mathrm{focal}}\in \pmb{\mathcal{T}}}[1 - \lambda_{\mathrm{focal}}(\pmb {\mathcal{T}};F_{\mathrm{focal}})]$ . Given a team size $S$ , the ensemble with the highest focal diversity can be deployed. As shown in Figure 5, the selected two-model team (green bar) is indeed the one leading to the strongest protection among all 28 options, which is over $20\%$ stronger than a randomly selected ensemble (red bar). Note that we only need to analyze the failures of individual FR models, which is significantly more efficient than generating P3-Mask for each option and conducting evaluation. + +# 5 Online Image Protection and Refinement + +The user $\mathcal{P}$ obtains her P3-Mask $\mathcal{M}_{\mathcal{P}}$ from Chameleon. Given an image $\mathcal{X}$ , she can obfuscate her facial identity on her device by applying the P3-Mask: + +$$ +\operatorname {M A S K} (\mathcal {X}; \boldsymbol {\mathcal {M}} _ {\mathcal {P}}) = \operatorname {C L I P} _ {[ 0, 2 5 5 ]} (\mathcal {F D} (\mathcal {X}) - \mathcal {Q R} (\boldsymbol {\mathcal {M}} _ {\mathcal {P}})), \tag {7} +$$ + +which is the protected facial region to be put in the original image $\mathcal{X}$ to produce the protected version $\mathcal{X}'$ for sharing. + +Chameleon supports authorizing trusted third parties to conduct FR on a user's facial images in a space-time efficient manner without transmitting and storing multiple versions of the same photo (one protected for public sharing and one unprotected for internal FR). By sharing the P3-Mask $\mathcal{M}_{\mathcal{P}}$ as the key, the third parties can de-obfuscate the facial signature by unmasking: + +$$ +\operatorname {U N M A S K} \left(\mathcal {X} ^ {\prime}; \boldsymbol {\mathcal {M}} _ {\mathcal {P}}\right) = \operatorname {C L I P} _ {[ 0, 2 5 5 ]} \left(\mathcal {F D} \left(\mathcal {X} ^ {\prime}\right) + \mathcal {Q R} \left(\boldsymbol {\mathcal {M}} _ {\mathcal {P}}\right)\right). \tag {8} +$$ + +In Section 6.2, we will show two properties of this process: (i) the signature-restored photos can be used for FR with no accuracy degradation, and (ii) the protection can only be done by the P3-Mask of the same person in the facial image, and the restoration is successful only if the same key is used. + +# 6 Experimental Evaluation + +We conduct experiments on FaceScrub [26] and LFW [20]. To provide a detailed analysis, ten celebrities are randomly selected as users on FaceScrub (Table 1). For each user, we split their facial images into three parts: (i) $10\%$ are used as probe images (2nd column), (ii) $70\%$ are used as gallery images and are used by Chameleon to train the mask (3rd column), and (iii) $20\%$ are also used as gallery images but unseen during the training process (4th column). The last part is crucial to evaluate the protection of unseen images of the user. In total, we have 123 facial images with "unknown identities." Following relevant works [31, 40], all facial images of other people are included in the gallery dataset as noise. The same splitting method is also used for LFW. By default, the results reported in this paper focus on FaceScrub due to similar observations on LFW. The source code is available at https://github.com/git-disl/Chameleon. + +Table 1: Ten example celebrities (users) on FaceScrub for analysis. + +
NameProbe ImagesGallery Images
SeenUnseenTotal
Morena Baccarin118223105
Bradley Cooper128925114
America Ferrera1611332145
Gerard Butler118224106
Eva Longoria129126117
Melissa Egan128724111
Kim Catrall139627123
Allison Janney128926115
Roma Downey117822100
Steve Carell139628124
Others (Noise)///48368
Total Number of Gallery Images:49528
+ +Table 2: The collection of publicly available pretrained FR models. + +
Model IDNeural Arch.Training DatasetFR Acc.
EN-MCEfficientNetMS-Celeb-1M94.94%
RN50-GLResNet50Glint360K96.68%
RN50-MCResNet50MS-Celeb-1M89.25%
RN50-VFResNet50VGG-Face294.30%
RN50-WFResNet50WebFace600K96.72%
RN18-MCResNet18MS-Celeb-1M82.64%
RN34-MCResNet34MS-Celeb-1M84.49%
RN100-MCResNet100MS-Celeb-1M91.85%
+ +Public clouds (e.g., Azure) now require manual approval to use their FR services. Hence, we believe that privacy intruders will opt for deploying pretrained FR models, as many high-quality ones are available on the Internet and can be used out of the box. Due to the large number of possible FR algorithms and architectures, we first focus on the eight FR models listed in Table 2, which are based on ArcFace [17], the state-of-the-art FR algorithm, with varying neural architectures and training datasets. Chameleon can be easily extended to incorporate any FR models by simply adding them to the collection. Nonetheless, in Section 6.4, we will show that thanks to the focal diversity-optimized teaming, the P3-Mask generated from a collection of ArcFace-only models can also be effective in protecting against privacy intruders using models of other FR algorithms, such as FaceNet [30] and MagFace [24]. By default, we use the two-model team (EN-MC, RN50-GL) with the highest focal diversity (Section 4.3). The P3-Mask is trained for 50 epochs on NVIDIA RTX 2080 SUPER GPU with $\eta = 0.001$ , $|\mathcal{B}| = 4$ , $\epsilon = 0.063$ , and $\omega = 0.03$ . We provide details for reproducibility and additional results in the appendix. + +# 6.1 Protection Effectiveness + +Table 3 shows the personalized mask for seven users on FaceScrub (other users are available in the appendix). Different users have distinct facial signatures (P3-Mask) learned by Chameleon. We apply these masks to their respective gallery + +Table 3: The P3-Mask (rescaled) for seven users on FaceScrub. Chameleon learns the distinct facial signature for each of them. + +
M. BaccarinB. CooperA. FerreraG. ButlerE. LongoriaM. EganK. Cattrall
+ +![](images/0b250f4eeca1c84abfbf5b040210e3290995710c7c564eb2c315f8d4100c2cc6.jpg) +(a) Dataset: FaceScrub +Fig. 6: Protection effectiveness on two benchmarks using Chameleon, OPOM [40], TIP-IM [39], and Fawkes [31] against FR models unknown to them. + +![](images/9dd95d6eeed8a8388885d77cde22d6fae71ae328457928e787de1b65c03f3a1e.jpg) +(b) Dataset: LFW + +images and test the FR accuracy using probe images. All probe images can be correctly identified when no protection mechanism is used. Under Chameleon, the probe images should be matched to an incorrect identity, resulting in a low FR accuracy. Hence, we define an evaluation metric, $\underline{\text{Protection Success Rate}}$ (PSR), to be $(100 - \text{FR Accuracy})$ reported in percentages. + +Chameleon can protect against unknown FR models. We compare the cross-model protection of Chameleon with OPOM [40], the only anti-FR approach generating one privacy protection mask for each person as Chameleon, and both TIP-IM [39] and Fawkes [31], which conduct per-image optimization. For a fair comparison, all methods are set with the same perturbation budget. Figure 6 summarizes the results on both datasets. All five FR models are unknown to both protection mechanisms. We make two observations. First, Chameleon consistently outperforms OPOM by a large margin. We found that OPOM masks do not transfer well, especially considering the end-to-end ML pipeline. Its PSR can drop by $33\%$ because of those lossy operations. Second, Chameleon can be as competitive as those methods conducting per-image optimization. TIP-IM and Fawkes spend minutes to optimize the protection for each image. Still, Chameleon outperforms them and can be applied instantly. + +Table 4 reports the detailed PSR for each user Chameleon achieves when the privacy intruder uses different FR models (2nd to 9th columns). We make two observations. First, when the privacy intruder uses an FR model known to the generation process, a PSR of $100\%$ can be consistently achieved (2nd to 3rd columns). Second, Chameleon can protect against unknown FR models (4th to 9th columns). Even when the FR model used by the intruder is trained on a different dataset (i.e., RN50-VF and RN50-WF) or with a different neural architecture (i.e., RN18-MC, RN34-MC, and RN100-MC) than any FR models known by Chameleon, it still provides a PSR over $90.65\%$ , meaning that facial images of a user are matched to gallery images of a different person. To demonstrate such a mismatch, in Table 5, we show the probe images from two users and their most similar gallery images found by four unknown FR models with two settings: (i) the "Unprotected" scenario where no one employs protection and (ii) the "Chameleon" scenario where the intruder scraped photos protected by our solution. Taking M. Baccarin as an example, the most similar gallery images found in the unprotected scenario belong to her. In contrast, when Chameleon is used, the same probe image is misidentified, e.g., as L. Hartley by RN50-MC. + +Table 4: Using the two-model team with high focal diversity (2nd and 3rd columns), Chameleon can offer privacy protection even when the privacy intruder uses FR models unseen during the mask generation process (4th to 9th columns). + +
NameProtection Success Rate - PSR (%)
EN-MCRN50-GLRN50-MCRN50-VFRN50-WFRN18-MCRN34-MCRN100-MCMeanStd
M. Baccarin100.00100.00100.0072.7390.91100.0090.91100.0094.329.64
B. Cooper100.00100.0083.33100.00100.00100.0091.6791.6795.836.30
A. Ferrera100.00100.0093.7575.00100.00100.0093.75100.0095.318.68
G. Butler100.00100.0090.91100.00100.0090.9190.91100.0096.594.70
E. Longoria100.00100.00100.00100.00100.00100.00100.00100.00100.000.00
M. Egan100.00100.00100.00100.00100.00100.00100.0091.6798.962.95
K. Cattrall100.00100.00100.0084.62100.00100.00100.00100.0098.085.44
A. Janney100.00100.0091.67100.0091.67100.00100.00100.0097.923.86
R. Downey100.00100.00100.0072.7390.91100.0090.91100.0094.329.64
S. Carell100.00100.0084.6292.3192.3192.3176.92100.0092.318.22
Mean100.00100.0093.5290.6596.5897.4194.4297.42
Std0.000.006.4011.164.434.187.414.15
+ +Table 5: The most similar gallery images on FaceScrub found by different FR models unknown to Chameleon. + +
Probe +ImageThe Most Similar Gallery Image & Its Identity Found by Unseen FR Models
RN50-MCRN50-VFRN34-MCRN100-MC
UnprotectedChameleonUnprotectedChameleonUnprotectedChameleonUnprotectedChameleon
M. BaccarinIdentifiedProtectedIdentifiedProtectedIdentifiedProtectedIdentifiedProtected
M. BaccarinL. HartleyM. BaccarinL. HartleyM. BaccarinC. ElectraM. BaccarinL. Hartley
B. CooperIdentifiedProtectedIdentifiedProtectedIdentifiedProtectedIdentifiedProtected
B. CooperP. WalkerB. CooperM. VartanB. CooperM. VartanB. CooperM. Vartan
+ +# 6.2 FR Service Usability + +Chameleon allows users to grant FR permission to trusted third parties by sharing their P3-Mask to de-obfuscate the protected image. There is no need to transmit and store an unprotected photo for internal use and a protected one for public visibility, doubling network and storage costs. Table 6 shows the results in FR accuracy. First, using the correct mask (i.e., the one used to protect the images) to reverse the protection can restore FR accuracy. Taking M. Baccarin as an example, the FR accuracy increases from $6.82\%$ before unmasking (b) to $100\%$ after unmasking (c). Second, using an incorrect mask for unmasking cannot restore the FR accuracy and can lead to even worse performance. It drops from $6.82\%$ to $1.14\%$ after unmasking (d). + +# 6.3 Protection Cost Analysis + +Speed and Resources. Figure 7 compares Chameleon with OPOM, TIP-IM, and Fawkes regarding the protection time per face, compute, and storage costs. + +Table 6: The protection by Chameleon can reduce FR accuracy against FR models (b), but the user can provide the P3-Mask used to protect her images as the key for the authorized third party to reverse the protection process (c), which leads to restored FR performance. However, an incorrect key cannot restore FR and may worsen it (d). + +
ScenarioFace Recognition Accuracy (%)
EN-MCRN50-GLRN50-MCRN50-VFRN50-WFRN18-MCRN34-MCRN100-MCMeanStd
User: M. Baccarin
(a) No Protection100.00100.00100.00100.00100.00100.00100.00100.00100.000.00
(b) Protected0.000.009.0918.189.099.090.009.096.826.43
(c) Correctly Unmasked100.00100.00100.00100.00100.00100.00100.00100.00100.000.00
(d) Incorrectly Unmasked0.000.000.000.009.090.000.000.001.143.21
User: B. Cooper
(a) No Protection100.00100.00100.00100.00100.00100.00100.00100.00100.000.00
(b) Protected0.000. 0016.670.000.000.008.338.334.176.30
(c) Correctly Unmasked100.00100.00100.00100.00100.00100.00100.00100.00100.000.00
(d) Incorrectly Unmasked0.000.000. 000.000.000.000.000.000.000.00
+ +![](images/c0ada8d7af1371cac2804844d97f3bc7f896e94d46a492f3084236d029994d0b.jpg) +Fig. 7: Chameleon offers much better privacy protection than OPOM without sacrificing speed and resource usages. It can protect instantly and require negligible compute resources as OPOM, while TIP-IM and Fawkes are slow and costly. + +![](images/f9d6a25330c8829d82e56b24e0174b87015eef8742c9b4d9ced91233ce04b07f.jpg) + +![](images/d94dc64126842e1ab631d873825ca79fc62ca79a0d7b111bfe091d3bbd045ffa.jpg) + +![](images/aa18156c72669038ebbb258d2174e3b003517e59b7bf42117c01540a525f1dd8.jpg) + +![](images/43f16377adc45961eedfc4aa4c8ca73c101a322c1aff2f2c279dbd0c48c09728.jpg) + +Chameleon and OPOM produce masks that are applicable to any facial image of the same user. The protection can be completed in 0.0076 seconds. Note that Chameleon is significantly more effective than OPOM, as described in Figure 6. Instead, TIP-IM and Fawkes require per-image optimization and take 105.12 seconds to complete the protection with GPUs for acceleration and storage for FR models. Chameleon only needs to conduct simple arithmetic operations. Only the P3-Mask needs to be stored on the user device. Hence, it can be deployed with instant protection even on an edge device with weak computing power. + +Image Quality. Chameleon can well preserve image quality. Table 7 provides two examples for four users, contrasting the facial images with no protection (3rd and 5th columns) with the ones with their facial signature removed by Chameleon (4th and 6th columns). They capture different variations of facial images: M. Baccarin's images have different lighting conditions, B. Cooper's have different face sizes, E. Longoria's have different postures, and G. Bulter's have different expressions. The protected images are visually similar to the unprotected counterparts, but they will not be matched to the clean images of the corresponding person. Indeed, Chameleon can generate higher-quality images than existing methods. As reported in Table 8, Chameleon not only offers better protection (Figure 6) with a similar cost (Figure 7) as OPOM but also achieves much better image quality measured in SSIM (0.9493 vs 0.8839). + +Table 7: The P3-Mask can be applied to any facial im- Table 8: Chameleon offers ages of the same person while preserving image quality. much better protection while + +
UserMaskSample Image 1Sample Image 2
No ProtectionProtectedNo ProtectionProtected
M. Baccarin
B. Cooper
E. Longoria
G. Image
+ +
Example
Original SSIM: 1.0000
Chameleon SSIM: 0.9493
OPOM SSIM: 0.8839
TIP-IM SSIM: 0.8850
Fawkes SSIM: 0.9612
+ +# 6.4 Focal Diversity-based Teaming + +Chameleon automatically selects a high-quality team for deployment with a specified budget. Table 9 compares the detailed PSR using the most diverse and the least diverse teams with (a) two and (b) three models. In both cases, the most diverse team outperforms the least diverse one, which is only effective when the privacy intruder uses the FR model known in the team. The selection of high-quality teams is non-trivial because we observe that composing a team of FR models with the highest FR accuracy is not always a good option, and a team of models having different neural architectures is not always the top priority [13]. + +In Figure 8, we further show that the P3-Mask generated from a carefully-chosen team can protect against FR models of unknown algorithms. Specifically, + +Table 9: The most diverse and the least diverse teams with (a) two or (b) three models identified by our focal diversity-based teaming method. The most diverse teams offer significantly better protection in terms of protection success rates. + +
Protection TeamProtection Success Rate - PSR (%)
EN-MCRN50-GLRN50-MCRN50-VFRN50-WFRN18-MCRN34-MCRN100-MCMeanStd
(a) Two-model Teams
Most Diverse:(EN-MC, RN50-GL)100.00100.0093.5290.6596.5897.4194.4297.4296.253.23
Least Diverse:(RN18-MC, RN34-MC)11.3860.16100.0055.2873.17100.00100.0080.4972.5628.54
(b) Three-model Teams
Most Diverse:(EN-MC, RN50-GL, RN50-WF)100.00100.00100.00100.0092.6893.50100.00100.0098.273.00
Least Diverse:(RN18-MC, RN34-MC, RN50-MC)2.4478.86100.0052.0384.55100.00100.0093.5076.4231.82
+ +![](images/b369f682755ec1966b4de25ceb8421d5d5ec12fd6c27737248f0001a7362c308.jpg) +Fig. 8: The most diverse teams allow Chameleon to be effective against FR models of unknown algorithms. + +![](images/bed3d7bad28822bdaaff5ee932ea111a0ec9135f62865cde53bcad95f29288ce.jpg) +Fig. 9: The PSR of Chameleon against adaptive intruders using different strategies to wash out the patterns in protected images. + +we consider the most and the least diverse three-model teams in Table 9(b). Even though both teams include only FR models based on ArcFace [17], the most diverse one can effectively protect against FR models based on FaceNet [30] or MagFace [24]. The protection effectiveness can be further strengthened by incorporating more FR models into the collection. + +# 6.5 Chameleon Against Adaptive Adversaries + +An adaptive adversary may run Chameleon to produce the P3-Mask for each person and de-obfuscate their photos. When probe images need to be identified, they will be matched against a "clean" face database. However, it is infeasible, as we show in Section 6.2, that the FR accuracy can only be restored when the masks used for obfuscation and de-obfuscation are identical. It is impossible to reproduce the same P3-Mask used by the user because it requires the same set of clean images and random states. + +Alternatively, an adaptive adversary may wash out patterns introduced by Chameleon before performing FR. We consider three popular lightweight and task-agnostic methods studied in the adversarial example domain [18] in Figure 9, including (a) JPEG compression [15], (b) Gaussian Filter, and (d) Median Filter [38]. The results show that their effectiveness is limited on Chameleon. The consideration of the end-to-end ML pipeline increases the robustness of P3-Mask. + +# 7 Conclusions + +We have presented Chameleon against unauthorized FR. Chameleon generates a P3-Mask for each user with cross-image and perceptibility optimizations to offer (i) instant and lightweight protection on any facial images of the same user and (ii) preservation of image quality. Also, we have shown that P3-Mask enables cost-effective de-obfuscation for authorizing FR services. Such an authorization process can only be conducted by the one with the identical mask used for protection. Chameleon also features a focal diversity-optimized teaming method to select a high-quality FR team to generate P3-Mask with strong robustness against unknown FR models. + +# Acknowledgments + +This research is partially sponsored by the NSF CISE grants 2302720, 2312758, 2038029, an IBM faculty award, and a grant from CISCO Edge AI program. It is part of the PhD dissertation of the first author, who graduated from Georgia Tech in Spring 2023. The first author acknowledges the support of the IBM PhD Fellowship in 2022-2023 and the support from the HKU-CS Start-up Fund. + +# References + +1. Brazilian retailer quizzed over facial recognition tech. https://www.zdnet.com/article/brazilian-retailer-quizzed-over-facial-recognition-tech/ (2019) +2. Facial recognition in the united states: Privacy concerns and legal developments. https://www.asisonline.org/security-management-magazine/monthlyissues/security-technology/archive/2021/december/facial-recognition-in-the-us-privacy-concerns-and-legal-developments/ (2021) +3. The secretive company that might end privacy as we know it. https://www.nytimes.com/2020/01/18/technology/clearview-privacy-facial-recognition.html (2021) +4. Clearview AI. https://www.clearview.ai/ (2023) +5. Facial recognition and identity risk. https://www.equifax.co.uk/resources/identity-protection/facial-recognition-and-identity-risk.html (2023) +6. HuggingFace. https://huggingface.co (2023) +7. Insightface: 2d and 3d face analysis project. https://github.com/deepinsight/insightface (2023) +8. PimEyes. https://pimeyes.com/ (2023) +9. Borgwardt, K.M., Gretton, A., Rasch, M.J., Kriegel, H.P., Scholkopf, B., Smola, A.J.: Integrating structured biological data by kernel maximum mean discrepancy. Bioinformatics 22(14), e49-e57 (2006) +10. Cherepanova, V., Goldblum, M., Foley, H., Duan, S., Dickerson, J., Taylor, G., Goldstein, T.: Lowkey: Leveraging adversarial attacks to protect social media users from facial recognition. In: International Conference on Learning Representations (ICLR) (2021) +11. Choi, Y., Choi, M., Kim, M., Ha, J.W., Kim, S., Choo, J.: Stargan: Unified generative adversarial networks for multi-domain image-to-image translation. In: Proceedings of the IEEE conference on computer vision and pattern recognition. pp. 8789-8797 (2018) +12. Chow, K.H., Hu, S., Huang, T., Ilhan, F., Wei, W., Liu, L.: Diversity-driven privacy protection masks against unauthorized face recognition. Proceedings on Privacy Enhancing Technologies 4, 381-392 (2024) +13. Chow, K.H., Liu, L.: Robust object detection fusion against deception. In: Proceedings of the 27th ACM SIGKDD Conference on Knowledge Discovery & Data Mining. pp. 2703-2713 (2021) +14. Chow, K.H., Liu, L.: Boosting object detection ensembles with error diversity. In: 2022 IEEE International Conference on Data Mining (ICDM). pp. 903-908. IEEE (2022) + +15. Das, N., Shanbhogue, M., Chen, S.T., Hohman, F., Li, S., Chen, L., Kounavis, M.E., Chau, D.H.: Shield: Fast, practical defense and vaccination for deep learning usingJPEG compression. In: Proceedings of the 24th ACM SIGKDD International Conference on Knowledge Discovery & Data Mining. pp. 196-204 (2018) +16. Deb, D., Zhang, J., Jain, A.K.: Advfaces: Adversarial face synthesis. In: 2020 IEEE International Joint Conference on Biometrics (IJCB). pp. 1-10. IEEE (2020) +17. Deng, J., Guo, J., Xue, N., Zafeiriou, S.: Arcface: Additive angular margin loss for deep face recognition. In: Proceedings of the IEEE/CVF conference on computer vision and pattern recognition. pp. 4690-4699 (2019) +18. Guo, C., Rana, M., Cisse, M., Van Der Maaten, L.: Countering adversarial images using input transformations. arXiv preprint arXiv:1711.00117 (2017) +19. Hu, S., Liu, X., Zhang, Y., Li, M., Zhang, L.Y., Jin, H., Wu, L.: Protecting facial privacy: generating adversarial identity masks via style-robust makeup transfer. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 15014-15023 (2022) +20. Huang, G.B., Ramesh, M., Berg, T., Learned-Miller, E.: Labeled faces in the wild: A database for studying face recognition in unconstrained environments. Tech. Rep. 07-49, University of Massachusetts, Amherst (October 2007) +21. Li, T., Lin, L.: Anonymousnet: Natural face de-identification with measurable privacy. In: Proceedings of the IEEE/CVF conference on computer vision and pattern recognition workshops. pp. 0–0 (2019) +22. Liu, L., Zhou, B., Zou, Z., Yeh, S.C., Zheng, L.: A smart unstaffed retail shop based on artificial intelligence and iot. In: 2018 IEEE 23rd International workshop on computer aided modeling and design of communication links and networks (CAMAD). pp. 1-4. IEEE (2018) +23. Lugaresi, C., Tang, J., Nash, H., McClanahan, C., Uboweja, E., Hays, M., Zhang, F., Chang, C.L., Yong, M.G., Lee, J., et al.: Mediapipe: A framework for building perception pipelines. arXiv preprint arXiv:1906.08172 (2019) +24. Meng, Q., Zhao, S., Huang, Z., Zhou, F.: Magface: A universal representation for face recognition and quality assessment. In: Proceedings of the IEEE/CVF conference on computer vision and pattern recognition. pp. 14225-14234 (2021) +25. Mirza, M., Osindero, S.: Conditional generative adversarial nets. arXiv preprint arXiv:1411.1784 (2014) +26. Ng, H.W., Winkler, S.: A data-driven approach to cleaning large face datasets. In: 2014 IEEE international conference on image processing (ICIP). pp. 343-347. IEEE (2014) +27. Partridge, D., Krzanowski, W.: Software diversity: practical statistics for its measurement and exploitation. Information and software technology 39(10), 707-717 (1997) +28. Pinto, N., Stone, Z., Zickler, T., Cox, D.: Scaling up biologically-inspired computer vision: A case study in unconstrained face recognition on facebook. In: CVPR 2011 WORKSHOPS. pp. 35-42. IEEE (2011) +29. Sahani, M., Nanda, C., Sahu, A.K., Pattnaik, B.: Web-based online embedded door access control and home security system based on face recognition. In: 2015 International Conference on Circuits, Power and Computing Technologies [ICCPCT-2015]. pp. 1-6. IEEE (2015) +30. Schroff, F., Kalenichenko, D., Philbin, J.: Facenet: A unified embedding for face recognition and clustering. In: Proceedings of the IEEE conference on computer vision and pattern recognition. pp. 815-823 (2015) + +31. Shan, S., Wenger, E., Zhang, J., Li, H., Zheng, H., Zhao, B.Y.: Fawkes: Protecting privacy against unauthorized deep learning models. In: Proceedings of the 29th USENIX Security Symposium (2020) +32. Stone, Z., Zickler, T., Darrell, T.: Autotagging facebook: Social network context improves photo annotation. In: 2008 IEEE computer society conference on computer vision and pattern recognition workshops. pp. 1-8. IEEE (2008) +33. Sun, Q., Tewari, A., Xu, W., Fritz, M., Theobalt, C., Schiele, B.: A hybrid model for identity obfuscation by face replacement. In: Proceedings of the European conference on computer vision (ECCV). pp. 553-569 (2018) +34. Wang, M., Deng, W.: Deep face recognition: A survey. Neurocomputing 429, 215-244 (2021) +35. Wang, Z., Bovik, A.C., Sheikh, H.R., Simoncelli, E.P.: Image quality assessment: from error visibility to structural similarity. IEEE transactions on image processing 13(4), 600-612 (2004) +36. Wenger, E., Shan, S., Zheng, H., Zhao, B.Y.: Sok: Anti-facial recognition technology. In: 2023 IEEE Symposium on Security and Privacy (SP). pp. 134-151 (2023) +37. Wu, Y., Liu, L., Xie, Z., Chow, K.H., Wei, W.: Boosting ensemble accuracy by revisiting ensemble diversity metrics. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 16469-16477 (2021) +38. Xu, W., Evans, D., Qi, Y.: Feature squeezing: Detecting adversarial examples in deep neural networks (2018) +39. Yang, X., Dong, Y., Pang, T., Su, H., Zhu, J., Chen, Y., Xue, H.: Towards face encryption by generating adversarial identity masks. In: Proceedings of the IEEE/CVF International Conference on Computer Vision. pp. 3897-3907 (2021) +40. Zhong, Y., Deng, W.: Opom: Customized invisible cloak towards face privacy protection. IEEE Transactions on Pattern Analysis and Machine Intelligence (2022) +41. Zhu, Z.A., Lu, Y.Z., Chiang, C.K.: Generating adversarial examples by makeup attacks on face recognition. In: 2019 IEEE International Conference on Image Processing (ICIP). pp. 2516-2520. IEEE (2019) \ No newline at end of file diff --git a/2024/Personalized Privacy Protection Mask Against Unauthorized Facial Recognition/images.zip b/2024/Personalized Privacy Protection Mask Against Unauthorized Facial Recognition/images.zip new file mode 100644 index 0000000000000000000000000000000000000000..e42964096fbc3f702dd360f0da540a63283b3c59 --- /dev/null +++ b/2024/Personalized Privacy Protection Mask Against Unauthorized Facial Recognition/images.zip @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:6ae71dbd517946e7614030545a469a860556b3e099f4f2e37da1f55f4e1de525 +size 883005 diff --git a/2024/Personalized Privacy Protection Mask Against Unauthorized Facial Recognition/layout.json b/2024/Personalized Privacy Protection Mask Against Unauthorized Facial Recognition/layout.json new file mode 100644 index 0000000000000000000000000000000000000000..e02c02fc158a17b6b5ae2a186fcbbb93ba94d709 --- /dev/null +++ b/2024/Personalized Privacy Protection Mask Against Unauthorized Facial Recognition/layout.json @@ -0,0 +1,8950 @@ +{ + "pdf_info": [ + { + "para_blocks": [ + { + "bbox": [ + 144, + 111, + 470, + 148 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 144, + 111, + 470, + 148 + ], + "spans": [ + { + "bbox": [ + 144, + 111, + 470, + 148 + ], + "type": "text", + "content": "Personalized Privacy Protection Mask Against Unauthorized Facial Recognition" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 157, + 167, + 455, + 180 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 157, + 167, + 455, + 180 + ], + "spans": [ + { + "bbox": [ + 157, + 167, + 455, + 180 + ], + "type": "text", + "content": "Ka-Ho Chow" + }, + { + "bbox": [ + 157, + 167, + 455, + 180 + ], + "type": "inline_equation", + "content": "^{1}" + }, + { + "bbox": [ + 157, + 167, + 455, + 180 + ], + "type": "text", + "content": ", Sihao Hu" + }, + { + "bbox": [ + 157, + 167, + 455, + 180 + ], + "type": "inline_equation", + "content": "^{2}" + }, + { + "bbox": [ + 157, + 167, + 455, + 180 + ], + "type": "text", + "content": ", Tiansheng Huang" + }, + { + "bbox": [ + 157, + 167, + 455, + 180 + ], + "type": "inline_equation", + "content": "^{2}" + }, + { + "bbox": [ + 157, + 167, + 455, + 180 + ], + "type": "text", + "content": ", and Ling Liu" + }, + { + "bbox": [ + 157, + 167, + 455, + 180 + ], + "type": "inline_equation", + "content": "^{2}" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 238, + 190, + 376, + 212 + ], + "type": "list", + "angle": 0, + "index": 4, + "blocks": [ + { + "bbox": [ + 241, + 190, + 373, + 201 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 241, + 190, + 373, + 201 + ], + "spans": [ + { + "bbox": [ + 241, + 190, + 373, + 201 + ], + "type": "text", + "content": "1 The University of Hong Kong" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 238, + 201, + 376, + 212 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 238, + 201, + 376, + 212 + ], + "spans": [ + { + "bbox": [ + 238, + 201, + 376, + 212 + ], + "type": "text", + "content": "2 Georgia Institute of Technology" + } + ] + } + ], + "index": 3 + } + ], + "sub_type": "text" + }, + { + "bbox": [ + 140, + 213, + 474, + 223 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 140, + 213, + 474, + 223 + ], + "spans": [ + { + "bbox": [ + 140, + 213, + 474, + 223 + ], + "type": "text", + "content": "kachow@cs.hku.hk, {sihaohu,thuang374}@gatech.edu, lingliu@cc.gatech.edu" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 159, + 247, + 453, + 456 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 159, + 247, + 453, + 456 + ], + "spans": [ + { + "bbox": [ + 159, + 247, + 453, + 456 + ], + "type": "text", + "content": "Abstract. Face recognition (FR) can be abused for privacy intrusion. Governments, private companies, or even individual attackers can collect facial images by web scraping to build an FR system identifying human faces without their consent. This paper introduces Chameleon, which learns to generate a user-centric personalized privacy protection mask, coined as P3-Mask, to protect facial images against unauthorized FR with three salient features. First, we use a cross-image optimization to generate one P3-Mask for each user instead of tailoring facial perturbation for each facial image of a user. It enables efficient and instant protection even for users with limited computing resources. Second, we incorporate a perceptibility optimization to preserve the visual quality of the protected facial images. Third, we strengthen the robustness of P3-Mask against unknown FR models by integrating focal diversity-optimized ensemble learning into the mask generation process. Extensive experiments on two benchmark datasets show that Chameleon outperforms three state-of-the-art methods with instant protection and minimal degradation of image quality. Furthermore, Chameleon enables cost-effective FR authorization using the P3-Mask as a personalized de-obfuscation key, and it demonstrates high resilience against adaptive adversaries." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 133, + 475, + 230, + 487 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 475, + 230, + 487 + ], + "spans": [ + { + "bbox": [ + 133, + 475, + 230, + 487 + ], + "type": "text", + "content": "1 Introduction" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 130, + 498, + 482, + 558 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 498, + 482, + 558 + ], + "spans": [ + { + "bbox": [ + 130, + 498, + 482, + 558 + ], + "type": "text", + "content": "Face recognition (FR) has long been investigated due to its potential for enhancing security and convenience in various domains [22,28,29]. Many pretrained FR models are available online [6,7]. Once a face database (a.k.a. the gallery) with facial images for each person of interest is provided, these pretrained models can be used to recognize them [34]." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 130, + 558, + 482, + 665 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 558, + 482, + 665 + ], + "spans": [ + { + "bbox": [ + 130, + 558, + 482, + 665 + ], + "type": "text", + "content": "While empowering many life-enriching applications, FR can be abused to cause serious privacy issues [2]. Privacy intruders can build a face database of victims of interest from publicly available facial images on the Internet by web scraping (Figure 1 (top)). By utilizing this database, the adversary can perform unauthorized FR of individual users for stalking victims [3], intruding on victims' privacy by flooding targeted ads [1], or facilitating criminals to commit identity fraud [5]. This is a real threat. For example, companies like Clearview [4] and PimEyes [8] have collected billions of online images and can recognize millions of citizens without their consent." + } + ] + } + ], + "index": 9 + } + ], + "discarded_blocks": [], + "page_size": [ + 612, + 792 + ], + "page_idx": 0 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 164, + 113, + 455, + 255 + ], + "blocks": [ + { + "bbox": [ + 164, + 113, + 455, + 255 + ], + "lines": [ + { + "bbox": [ + 164, + 113, + 455, + 255 + ], + "spans": [ + { + "bbox": [ + 164, + 113, + 455, + 255 + ], + "type": "image", + "image_path": "1173bc9750a62a9f482119b53580730f40508b1fd6750a03605ce1312da9b1fe.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 130, + 262, + 482, + 307 + ], + "lines": [ + { + "bbox": [ + 130, + 262, + 482, + 307 + ], + "spans": [ + { + "bbox": [ + 130, + 262, + 482, + 307 + ], + "type": "text", + "content": "Fig. 1: (Top) Without protection, the privacy intruder can build a face database by web scraping and identify the unknown face. (Bottom) Chameleon learns the facial signature of the user (protectee) to generate a P3-Mask, which can be applied to protect any facial images before sharing them online against unauthorized FR." + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_caption" + } + ], + "index": 2 + }, + { + "bbox": [ + 130, + 330, + 482, + 582 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 330, + 482, + 582 + ], + "spans": [ + { + "bbox": [ + 130, + 330, + 482, + 582 + ], + "type": "text", + "content": "Such a facial privacy threat has motivated the development of anti-FR technologies [36], which allow users (protectees) to preprocess their facial images before sharing them online. We argue that a competitive anti-FR solution should possess the following four properties. First, from the robustness perspective, the protected facial images should not be matched by an FR model to a query (probe) image with a correct identity (e.g., Bradley Cooper is recognized as Anne Hathaway in Figure 1 (bottom)), including those FR models possibly unknown during the preprocessing but are used by the privacy intruder. Second, from the efficiency perspective, the protection should be done instantly on any device with or without AI accelerators to maintain user experience and support equitable access to privacy protection. Third, from the perception usability perspective, the protected image should (i) preserve the visual quality rather than using excessive noise and (ii) be visually recognizable by humans to be the same person as the one in the original unprotected photo. Finally, the service usability for authorized FR providers is another important property. Existing approaches treat all FR models as unauthorized. We argue that anti-FR solutions should also allow users to grant FR permission to authorized models cost-effectively. This user-initiated de-obfuscation capability is critical to those, e.g., who publish their photos on a social media platform with privacy protection and, at the same time, wish to authorize the platform to perform some FR services (e.g., face tagging [32]) without sending and storing multiple versions of the same photo." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 130, + 582, + 482, + 666 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 582, + 482, + 666 + ], + "spans": [ + { + "bbox": [ + 130, + 582, + 482, + 666 + ], + "type": "text", + "content": "Based on the above objectives, we introduce Chameleon, a user-centric facial privacy protection system with three contributions. First, we develop an optimization algorithm to construct a Personalized Privacy Protection mask, coined as P3-Mask, for each user. The P3-Mask of a user can be used to protect any facial images of the same user with minimal impact on image quality, including those facial images unseen during the P3-Mask generation process. Second, it boosts the robustness of P3-Mask against unknown FR models by a princi" + } + ] + } + ], + "index": 5 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "spans": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "type": "text", + "content": "2" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 166, + 91, + 240, + 101 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 166, + 91, + 240, + 101 + ], + "spans": [ + { + "bbox": [ + 166, + 91, + 240, + 101 + ], + "type": "text", + "content": "K.-H. Chow et al." + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 1 + }, + { + "para_blocks": [ + { + "bbox": [ + 130, + 116, + 482, + 260 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 116, + 482, + 260 + ], + "spans": [ + { + "bbox": [ + 130, + 116, + 482, + 260 + ], + "type": "text", + "content": "pled approach, leveraging ensemble learning with models of high focal diversity. The per-user P3-Mask provides higher robustness against unknown FR models while preserving perception usability and the service usability for authorized FR providers. Chameleon users can utilize their own P3-Mask to obfuscate their facial images prior to public release, and the same P3-Mask can be used as a personalized de-obfuscation key to grant authorized FR service providers the ability to restore the facial signature of photos for correct recognition. Third, we conduct extensive experiments on two FR benchmark datasets [20,26] to analyze the shortcomings of state-of-the-art anti-FR solutions [31,39,40] and show that Chameleon can better protect against unknown FR models with a high success rate. The protection is lightweight and in real-time. It remains effective under adaptive privacy intruders deploying various strategies to counter Chameleon." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 132, + 277, + 237, + 289 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 277, + 237, + 289 + ], + "spans": [ + { + "bbox": [ + 132, + 277, + 237, + 289 + ], + "type": "text", + "content": "2 Related Work" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 130, + 300, + 499, + 361 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 300, + 499, + 361 + ], + "spans": [ + { + "bbox": [ + 130, + 300, + 499, + 361 + ], + "type": "text", + "content": "Existing anti-FR approaches can be broadly classified into two categories. Synthesis-based approaches [11,16,19,21,33,41] use generative adversarial networks (GANs) [25] to synthesize a face to replace the original one for protection. While the synthetic faces look realistic, they appear to be strangers, not recognizable even by the users themselves, failing to meet the perception usability requirement." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 130, + 361, + 482, + 553 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 361, + 482, + 553 + ], + "spans": [ + { + "bbox": [ + 130, + 361, + 482, + 553 + ], + "type": "text", + "content": "Chameleon falls into the second category, which applies small changes to facial images [12]. Fawkes [31] formulates an untargeted attack to push the facial image away from the original location in the embedding space; TIP-IM [39] improves the image quality with MMD [9]; LowKey [10] improves the robustness under the image processing pipeline in an end-to-end ML system by incorporating it into the optimization process. These techniques can preserve the visual identity of protected faces. However, they both require iterative optimization for each image. Even for a GPU server, it can take over 100 seconds to perturb one facial image (Section 6.3). In contrast, Chameleon provides instant protection on any device, with protection even stronger than those spending minutes to find the best perturbation for each image. OPOM [40] attempts to improve efficiency by finding the embedding subspace enclosing facial images of the same person and generates a privacy protection mask to push any images away from it. However, it fails to maintain protection effectiveness when image processing operations are applied to protected images (Section 6.1), and the mask degrades the image quality much more significantly than Chameleon (Section 6.3)." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 132, + 570, + 211, + 582 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 570, + 211, + 582 + ], + "spans": [ + { + "bbox": [ + 132, + 570, + 211, + 582 + ], + "type": "text", + "content": "3 Overview" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 130, + 594, + 482, + 666 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 594, + 482, + 666 + ], + "spans": [ + { + "bbox": [ + 130, + 594, + 482, + 666 + ], + "type": "text", + "content": "To build an FR system with a pretrained model, the owner first collects a gallery dataset (face database) " + }, + { + "bbox": [ + 130, + 594, + 482, + 666 + ], + "type": "inline_equation", + "content": "\\pmb{D}" + }, + { + "bbox": [ + 130, + 594, + 482, + 666 + ], + "type": "text", + "content": ", which includes the facial images of individuals of interest. Then, the pretrained FR model " + }, + { + "bbox": [ + 130, + 594, + 482, + 666 + ], + "type": "inline_equation", + "content": "F" + }, + { + "bbox": [ + 130, + 594, + 482, + 666 + ], + "type": "text", + "content": " is used to map each gallery image " + }, + { + "bbox": [ + 130, + 594, + 482, + 666 + ], + "type": "inline_equation", + "content": "\\tilde{\\pmb{x}} \\in \\pmb{D}" + }, + { + "bbox": [ + 130, + 594, + 482, + 666 + ], + "type": "text", + "content": " to the embedding " + }, + { + "bbox": [ + 130, + 594, + 482, + 666 + ], + "type": "inline_equation", + "content": "F(\\tilde{\\pmb{x}})" + }, + { + "bbox": [ + 130, + 594, + 482, + 666 + ], + "type": "text", + "content": ", and those embeddings should form clusters corresponding to different people. When a facial image " + }, + { + "bbox": [ + 130, + 594, + 482, + 666 + ], + "type": "inline_equation", + "content": "\\pmb{x}" + }, + { + "bbox": [ + 130, + 594, + 482, + 666 + ], + "type": "text", + "content": " of an unknown identity, called the probe image, is given, the FR system can use the same FR model " + }, + { + "bbox": [ + 130, + 594, + 482, + 666 + ], + "type": "inline_equation", + "content": "F" + }, + { + "bbox": [ + 130, + 594, + 482, + 666 + ], + "type": "text", + "content": " to" + } + ] + } + ], + "index": 7 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 184, + 91, + 447, + 102 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 184, + 91, + 447, + 102 + ], + "spans": [ + { + "bbox": [ + 184, + 91, + 447, + 102 + ], + "type": "text", + "content": "Personalized Privacy Protection Mask Against Unauthorized FR" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 474, + 92, + 480, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 474, + 92, + 480, + 100 + ], + "spans": [ + { + "bbox": [ + 474, + 92, + 480, + 100 + ], + "type": "text", + "content": "3" + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 2 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 133, + 114, + 480, + 205 + ], + "blocks": [ + { + "bbox": [ + 133, + 114, + 480, + 205 + ], + "lines": [ + { + "bbox": [ + 133, + 114, + 480, + 205 + ], + "spans": [ + { + "bbox": [ + 133, + 114, + 480, + 205 + ], + "type": "image", + "image_path": "47589bdc1be02a49d4c4bac2fbfdf6c5df0a7220694b62be80e651b98c27e757.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 131, + 215, + 482, + 251 + ], + "lines": [ + { + "bbox": [ + 131, + 215, + 482, + 251 + ], + "spans": [ + { + "bbox": [ + 131, + 215, + 482, + 251 + ], + "type": "text", + "content": "Fig. 2: An overview of Chameleon's two-phase workflow for offline learning to generate P3-Mask and online protection with P3-Mask for personalized facial signature masking. The P3-Mask can protect any facial images of the same person without further learning." + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_caption" + } + ], + "index": 2 + }, + { + "bbox": [ + 131, + 273, + 482, + 309 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 131, + 273, + 482, + 309 + ], + "spans": [ + { + "bbox": [ + 131, + 273, + 482, + 309 + ], + "type": "text", + "content": "map it to an embedding " + }, + { + "bbox": [ + 131, + 273, + 482, + 309 + ], + "type": "inline_equation", + "content": "F(\\pmb{x})" + }, + { + "bbox": [ + 131, + 273, + 482, + 309 + ], + "type": "text", + "content": " and use the identity of the nearest gallery image to be the identity of the unknown person. The identity, " + }, + { + "bbox": [ + 131, + 273, + 482, + 309 + ], + "type": "inline_equation", + "content": "\\mathcal{FR}(\\pmb{x};F,\\pmb{D})" + }, + { + "bbox": [ + 131, + 273, + 482, + 309 + ], + "type": "text", + "content": ", of the probe image " + }, + { + "bbox": [ + 131, + 273, + 482, + 309 + ], + "type": "inline_equation", + "content": "\\pmb{x}" + }, + { + "bbox": [ + 131, + 273, + 482, + 309 + ], + "type": "text", + "content": " given the FR model " + }, + { + "bbox": [ + 131, + 273, + 482, + 309 + ], + "type": "inline_equation", + "content": "F" + }, + { + "bbox": [ + 131, + 273, + 482, + 309 + ], + "type": "text", + "content": " and the gallery dataset " + }, + { + "bbox": [ + 131, + 273, + 482, + 309 + ], + "type": "inline_equation", + "content": "\\pmb{D}" + }, + { + "bbox": [ + 131, + 273, + 482, + 309 + ], + "type": "text", + "content": " is defined as:" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 212, + 316, + 481, + 335 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 212, + 316, + 481, + 335 + ], + "spans": [ + { + "bbox": [ + 212, + 316, + 481, + 335 + ], + "type": "interline_equation", + "content": "\\mathcal {F R} (\\boldsymbol {x}; F, \\boldsymbol {\\mathcal {D}}) = \\mathcal {I} \\left(\\underset {\\tilde {\\boldsymbol {x}} \\in \\boldsymbol {\\mathcal {D}}} {\\arg \\min } \\operatorname {D I S T} (F (\\boldsymbol {x}), F (\\tilde {\\boldsymbol {x}}))\\right), \\tag {1}", + "image_path": "2ef87694abc060a34d5de58a285f59e0164529167cde5b030ca34fbad776ca2f.jpg" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 131, + 341, + 482, + 365 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 131, + 341, + 482, + 365 + ], + "spans": [ + { + "bbox": [ + 131, + 341, + 482, + 365 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 131, + 341, + 482, + 365 + ], + "type": "inline_equation", + "content": "\\mathcal{I}(\\tilde{\\boldsymbol{x}})" + }, + { + "bbox": [ + 131, + 341, + 482, + 365 + ], + "type": "text", + "content": " denotes the identity of the gallery image " + }, + { + "bbox": [ + 131, + 341, + 482, + 365 + ], + "type": "inline_equation", + "content": "\\tilde{\\boldsymbol{x}}" + }, + { + "bbox": [ + 131, + 341, + 482, + 365 + ], + "type": "text", + "content": ", which is known to the FR system, and " + }, + { + "bbox": [ + 131, + 341, + 482, + 365 + ], + "type": "inline_equation", + "content": "\\mathrm{DIST}(\\cdot ,\\cdot)" + }, + { + "bbox": [ + 131, + 341, + 482, + 365 + ], + "type": "text", + "content": " is a distance function such as Euclidean distance." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 131, + 365, + 482, + 485 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 131, + 365, + 482, + 485 + ], + "spans": [ + { + "bbox": [ + 131, + 365, + 482, + 485 + ], + "type": "text", + "content": "Threat Model. We consider a threat model where a privacy intruder conducts web scraping to collect images containing citizens' faces. Web scraping is large-scale and untargeted. Many citizens are included in the face database, and the privacy intruder may not know them. Given a probe image taken by, e.g., a stalker's camera, the privacy intruder uses an FR model on the face database to search for matched image(s). Then, the privacy intruder can analyze the associated metadata, such as the web page from where it was downloaded or even the exact identity. The goal of Chameleon is to allow the user to preprocess her facial images before posting them online. Even if scraped and included in the face database, they will not be matched to a probe image of her." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 132, + 502, + 255, + 514 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 502, + 255, + 514 + ], + "spans": [ + { + "bbox": [ + 132, + 502, + 255, + 514 + ], + "type": "text", + "content": "3.1 Chameleon Design" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 131, + 521, + 482, + 557 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 131, + 521, + 482, + 557 + ], + "spans": [ + { + "bbox": [ + 131, + 521, + 482, + 557 + ], + "type": "text", + "content": "Chameleon performs preprocessing of the user's photos before sharing them online by applying the user-specific P3-Mask. This mask is generated offline for a Chameleon user. Figure 2 provides an overview of Chameleon's workflow." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 131, + 558, + 482, + 666 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 131, + 558, + 482, + 666 + ], + "spans": [ + { + "bbox": [ + 131, + 558, + 482, + 666 + ], + "type": "text", + "content": "Offline Learning Phase. Chameleon learns the unique facial signature of a user from a few facial images of her using the P3-Mask generator (Section 4), as shown in Figure 2 ①. P3-Mask is designed to apply directly to any facial images of the same user, including those unseen during offline learning, and is robust against lossy image processing operations in an end-to-end ML system without compromising the image quality. We use a focal diversity-optimized ensemble learning method to find a team of FR models such that the P3-Mask generated against them has strong robustness and is effective in countering other FR models that are unknown during offline learning." + } + ] + } + ], + "index": 10 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "spans": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "type": "text", + "content": "4" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 166, + 91, + 240, + 101 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 166, + 91, + 240, + 101 + ], + "spans": [ + { + "bbox": [ + 166, + 91, + 240, + 101 + ], + "type": "text", + "content": "K.-H. Chow et al." + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 3 + }, + { + "para_blocks": [ + { + "bbox": [ + 130, + 116, + 482, + 213 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 116, + 482, + 213 + ], + "spans": [ + { + "bbox": [ + 130, + 116, + 482, + 213 + ], + "type": "text", + "content": "Online Protection Phase. After the offline learning, ② the P3-Mask can be sent to the user's local device (e.g., the mobile phone). For a photo of the user to be protected, ③ a lightweight face detection model, such as MediaPipe [23], can be used by the Chameleon user running on the user's device to locate the face region and ④ instantly protect it with the P3-Mask before the user shares it online ⑤. The user can also authorize some trusted entities to conduct FR on their shared photos protected with P3-Mask. The user-specific de-obfuscation key ⑥ allows the protected photos to be restored for authorized FR." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 132, + 229, + 279, + 243 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 229, + 279, + 243 + ], + "spans": [ + { + "bbox": [ + 132, + 229, + 279, + 243 + ], + "type": "text", + "content": "4 P3-Mask Generation" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 130, + 254, + 482, + 399 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 254, + 482, + 399 + ], + "spans": [ + { + "bbox": [ + 130, + 254, + 482, + 399 + ], + "type": "text", + "content": "For a user " + }, + { + "bbox": [ + 130, + 254, + 482, + 399 + ], + "type": "inline_equation", + "content": "\\mathcal{P}" + }, + { + "bbox": [ + 130, + 254, + 482, + 399 + ], + "type": "text", + "content": ", Chameleon generates a P3-Mask by learning the facial signature " + }, + { + "bbox": [ + 130, + 254, + 482, + 399 + ], + "type": "inline_equation", + "content": "\\mathcal{M}_{\\mathcal{P}}" + }, + { + "bbox": [ + 130, + 254, + 482, + 399 + ], + "type": "text", + "content": " from a set of facial images the user provides. Several key factors need to be accomplished. (1) We need to promote cross-image protection as an optimization objective to generate the most representative facial signature applicable to protect any facial images of the user " + }, + { + "bbox": [ + 130, + 254, + 482, + 399 + ], + "type": "inline_equation", + "content": "\\mathcal{P}" + }, + { + "bbox": [ + 130, + 254, + 482, + 399 + ], + "type": "text", + "content": ", including those unknown ones during the generation process (Section 4.1). (2) We need to control the amount of perturbation introduced by the P3-Mask. When applied to an unprotected image by removing the learned facial signature, it ensures minimal perception loss to preserve the visual quality (Section 4.2). (3) We need to keep the protected image far away from the original image in the embedding space. Multiple FR models should be used to generate alternative embedding representations to enhance generalizability against unknown models (Section 4.3)." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 132, + 415, + 349, + 427 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 415, + 349, + 427 + ], + "spans": [ + { + "bbox": [ + 132, + 415, + 349, + 427 + ], + "type": "text", + "content": "4.1 Cross-image Protection Optimization" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 130, + 434, + 482, + 518 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 434, + 482, + 518 + ], + "spans": [ + { + "bbox": [ + 130, + 434, + 482, + 518 + ], + "type": "text", + "content": "The cross-image protection capability comes from iterative learning on a set of training images the user provides. The idea is to keep fine-tuning the P3-Mask so that it can offer protection simultaneously to training images while preserving image quality. At the " + }, + { + "bbox": [ + 130, + 434, + 482, + 518 + ], + "type": "inline_equation", + "content": "t" + }, + { + "bbox": [ + 130, + 434, + 482, + 518 + ], + "type": "text", + "content": "-th iteration, it samples a mini-batch " + }, + { + "bbox": [ + 130, + 434, + 482, + 518 + ], + "type": "inline_equation", + "content": "\\mathcal{B}" + }, + { + "bbox": [ + 130, + 434, + 482, + 518 + ], + "type": "text", + "content": " from the training dataset " + }, + { + "bbox": [ + 130, + 434, + 482, + 518 + ], + "type": "inline_equation", + "content": "\\Omega" + }, + { + "bbox": [ + 130, + 434, + 482, + 518 + ], + "type": "text", + "content": " containing photos of user " + }, + { + "bbox": [ + 130, + 434, + 482, + 518 + ], + "type": "inline_equation", + "content": "\\mathcal{P}" + }, + { + "bbox": [ + 130, + 434, + 482, + 518 + ], + "type": "text", + "content": ". Then, we find the modification to the P3-Mask that can lead to better protection on each image in the mini-batch with better visual quality with the following operations:" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 168, + 524, + 481, + 550 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 168, + 524, + 481, + 550 + ], + "spans": [ + { + "bbox": [ + 168, + 524, + 481, + 550 + ], + "type": "interline_equation", + "content": "\\mathcal {M} _ {\\mathcal {P}} ^ {t + 1} = \\operatorname {C L I P} _ {[ - \\epsilon , \\epsilon ]} \\left(\\mathcal {M} _ {\\mathcal {P}} ^ {t} - \\eta \\operatorname {S I G N} \\left(\\frac {1}{| \\boldsymbol {\\mathcal {B}} |} \\sum_ {\\mathcal {X} \\in \\boldsymbol {\\mathcal {B}}} \\nabla_ {\\mathcal {M} _ {\\mathcal {P}} ^ {t}} \\mathcal {L} \\left(\\mathcal {X}, \\mathcal {M} _ {\\mathcal {P}} ^ {t}; \\boldsymbol {\\mathcal {T}}, \\omega\\right)\\right)\\right). \\tag {2}", + "image_path": "a6642ab7c0bfb3e5da23f9f1721833980cb16b637f8f00f4b90d6aaafd9c6c07.jpg" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 130, + 555, + 481, + 581 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 555, + 481, + 581 + ], + "spans": [ + { + "bbox": [ + 130, + 555, + 481, + 581 + ], + "type": "text", + "content": "Specifically, for each image " + }, + { + "bbox": [ + 130, + 555, + 481, + 581 + ], + "type": "inline_equation", + "content": "\\mathcal{X} \\in \\mathcal{B}" + }, + { + "bbox": [ + 130, + 555, + 481, + 581 + ], + "type": "text", + "content": ", we use the P3-Mask learned up to the current iteration, " + }, + { + "bbox": [ + 130, + 555, + 481, + 581 + ], + "type": "inline_equation", + "content": "\\mathcal{M}_{\\mathcal{P}}^t" + }, + { + "bbox": [ + 130, + 555, + 481, + 581 + ], + "type": "text", + "content": ", to protect it and compute the loss designed with dual goals:" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 180, + 586, + 481, + 600 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 180, + 586, + 481, + 600 + ], + "spans": [ + { + "bbox": [ + 180, + 586, + 481, + 600 + ], + "type": "interline_equation", + "content": "\\mathcal {L} (\\mathcal {X}, \\boldsymbol {\\mathcal {M}} _ {\\mathcal {P}} ^ {t}; \\boldsymbol {\\mathcal {T}}, \\omega) = \\mathcal {L} _ {\\text {P r o t e c t}} (\\mathcal {X}, \\boldsymbol {\\mathcal {M}} _ {\\mathcal {P}} ^ {t}; \\boldsymbol {\\mathcal {T}}) + \\mathcal {L} _ {\\text {P e r c e p t}} (\\mathcal {X}, \\boldsymbol {\\mathcal {M}} _ {\\mathcal {P}} ^ {t}; \\omega). \\tag {3}", + "image_path": "88072d9cd72ca2d87cd41fe49dee419340482e21ea7ff9191bd2cd50afa2c5a2.jpg" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 130, + 605, + 482, + 667 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 605, + 482, + 667 + ], + "spans": [ + { + "bbox": [ + 130, + 605, + 482, + 667 + ], + "type": "text", + "content": "It uses " + }, + { + "bbox": [ + 130, + 605, + 482, + 667 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_{\\mathrm{Protect}}" + }, + { + "bbox": [ + 130, + 605, + 482, + 667 + ], + "type": "text", + "content": " to learn how to adjust the P3-Mask " + }, + { + "bbox": [ + 130, + 605, + 482, + 667 + ], + "type": "inline_equation", + "content": "\\mathcal{M}_{\\mathcal{P}}^{t}" + }, + { + "bbox": [ + 130, + 605, + 482, + 667 + ], + "type": "text", + "content": " from the previous iteration to better protect against a team of pre-selected FR models " + }, + { + "bbox": [ + 130, + 605, + 482, + 667 + ], + "type": "inline_equation", + "content": "\\pmb{\\tau}" + }, + { + "bbox": [ + 130, + 605, + 482, + 667 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 130, + 605, + 482, + 667 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_{\\mathrm{Percept}}" + }, + { + "bbox": [ + 130, + 605, + 482, + 667 + ], + "type": "text", + "content": " to preserve the image quality controlled by " + }, + { + "bbox": [ + 130, + 605, + 482, + 667 + ], + "type": "inline_equation", + "content": "\\omega" + }, + { + "bbox": [ + 130, + 605, + 482, + 667 + ], + "type": "text", + "content": ". We use the SIGN of the gradients to update the P3-Mask with a learning rate " + }, + { + "bbox": [ + 130, + 605, + 482, + 667 + ], + "type": "inline_equation", + "content": "\\eta" + }, + { + "bbox": [ + 130, + 605, + 482, + 667 + ], + "type": "text", + "content": ". A clipping function " + }, + { + "bbox": [ + 130, + 605, + 482, + 667 + ], + "type": "inline_equation", + "content": "\\mathrm{CLIP}_{[-\\epsilon, \\epsilon]}" + }, + { + "bbox": [ + 130, + 605, + 482, + 667 + ], + "type": "text", + "content": " is applied to ensure the changes made by the P3-Mask on the facial" + } + ] + } + ], + "index": 10 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 184, + 91, + 448, + 103 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 184, + 91, + 448, + 103 + ], + "spans": [ + { + "bbox": [ + 184, + 91, + 448, + 103 + ], + "type": "text", + "content": "Personalized Privacy Protection Mask Against Unauthorized FR" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 474, + 92, + 480, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 474, + 92, + 480, + 100 + ], + "spans": [ + { + "bbox": [ + 474, + 92, + 480, + 100 + ], + "type": "text", + "content": "5" + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 4 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 141, + 113, + 474, + 223 + ], + "blocks": [ + { + "bbox": [ + 141, + 113, + 474, + 223 + ], + "lines": [ + { + "bbox": [ + 141, + 113, + 474, + 223 + ], + "spans": [ + { + "bbox": [ + 141, + 113, + 474, + 223 + ], + "type": "image", + "image_path": "ebe22766176c65b7c394601e6e1b72ddd49e799792f30246dbf238fe9e49d7d5.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 131, + 232, + 480, + 255 + ], + "lines": [ + { + "bbox": [ + 131, + 232, + 480, + 255 + ], + "spans": [ + { + "bbox": [ + 131, + 232, + 480, + 255 + ], + "type": "text", + "content": "Fig. 3: The iterative generation of P3-Mask for a user. Chameleon goes through multiple facial images of the same user to optimize a P3-Mask with ML pipeline awareness." + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_caption" + } + ], + "index": 2 + }, + { + "bbox": [ + 130, + 277, + 480, + 312 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 277, + 480, + 312 + ], + "spans": [ + { + "bbox": [ + 130, + 277, + 480, + 312 + ], + "type": "text", + "content": "image are bounded by " + }, + { + "bbox": [ + 130, + 277, + 480, + 312 + ], + "type": "inline_equation", + "content": "\\epsilon" + }, + { + "bbox": [ + 130, + 277, + 480, + 312 + ], + "type": "text", + "content": ". Such cross-image protection is not limited to those in the training dataset " + }, + { + "bbox": [ + 130, + 277, + 480, + 312 + ], + "type": "inline_equation", + "content": "\\Omega" + }, + { + "bbox": [ + 130, + 277, + 480, + 312 + ], + "type": "text", + "content": " but also unseen images of the same user. We next discuss the design of " + }, + { + "bbox": [ + 130, + 277, + 480, + 312 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_{\\mathrm{Protect}}" + }, + { + "bbox": [ + 130, + 277, + 480, + 312 + ], + "type": "text", + "content": ", and " + }, + { + "bbox": [ + 130, + 277, + 480, + 312 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_{\\mathrm{Percept}}" + }, + { + "bbox": [ + 130, + 277, + 480, + 312 + ], + "type": "text", + "content": " will be detailed in Section 4.2." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 130, + 312, + 481, + 409 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 312, + 481, + 409 + ], + "spans": [ + { + "bbox": [ + 130, + 312, + 481, + 409 + ], + "type": "text", + "content": "For privacy protection, P3-Mask maximizes the distance between the protected and unprotected images in the embedding spaces of a team " + }, + { + "bbox": [ + 130, + 312, + 481, + 409 + ], + "type": "inline_equation", + "content": "\\mathcal{T}" + }, + { + "bbox": [ + 130, + 312, + 481, + 409 + ], + "type": "text", + "content": " of pre-selected FR models. We incorporate image processing operations into the optimization to avoid those lossy operations degrading P3-Mask's protection. Consider the raw training image " + }, + { + "bbox": [ + 130, + 312, + 481, + 409 + ], + "type": "inline_equation", + "content": "\\mathcal{X}" + }, + { + "bbox": [ + 130, + 312, + 481, + 409 + ], + "type": "text", + "content": " in the mini-batch in Figure 3. We first conduct face detection to produce the cropped face " + }, + { + "bbox": [ + 130, + 312, + 481, + 409 + ], + "type": "inline_equation", + "content": "\\mathcal{FD}(\\mathcal{X})" + }, + { + "bbox": [ + 130, + 312, + 481, + 409 + ], + "type": "text", + "content": ". Quantization and resizing are executed on the P3-Mask to match the resolution of the cropped face, denoted by " + }, + { + "bbox": [ + 130, + 312, + 481, + 409 + ], + "type": "inline_equation", + "content": "\\mathcal{QR}(\\mathcal{M}_{\\mathcal{P}}^t)" + }, + { + "bbox": [ + 130, + 312, + 481, + 409 + ], + "type": "text", + "content": ". Then, we can apply the mask to the face for protection:" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 201, + 418, + 481, + 432 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 201, + 418, + 481, + 432 + ], + "spans": [ + { + "bbox": [ + 201, + 418, + 481, + 432 + ], + "type": "interline_equation", + "content": "\\Theta (\\mathcal {X}; \\boldsymbol {\\mathcal {M}} _ {\\mathcal {P}} ^ {t}) = \\Psi \\left(\\operatorname {C L I P} _ {[ 0, 2 5 5 ]} \\left(\\mathcal {F D} (\\mathcal {X}) - \\mathcal {Q R} \\left(\\boldsymbol {\\mathcal {M}} _ {\\mathcal {P}} ^ {t}\\right)\\right)\\right), \\tag {4}", + "image_path": "1af4b441de5216981bec6fb2ca095277ce6fd90f8bd284a64b638dde849a7c4b.jpg" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 130, + 434, + 480, + 470 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 434, + 480, + 470 + ], + "spans": [ + { + "bbox": [ + 130, + 434, + 480, + 470 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 130, + 434, + 480, + 470 + ], + "type": "inline_equation", + "content": "\\varPsi" + }, + { + "bbox": [ + 130, + 434, + 480, + 470 + ], + "type": "text", + "content": " is a padding and resizing function to meet the input resolution requirement of an FR model (e.g., " + }, + { + "bbox": [ + 130, + 434, + 480, + 470 + ], + "type": "inline_equation", + "content": "(112\\times 112)" + }, + { + "bbox": [ + 130, + 434, + 480, + 470 + ], + "type": "text", + "content": " in ArcFace [17]). Note that to simplify the notation, we removed certain arguments without causing confusion." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 130, + 470, + 481, + 518 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 470, + 481, + 518 + ], + "spans": [ + { + "bbox": [ + 130, + 470, + 481, + 518 + ], + "type": "text", + "content": "The protection optimization in P3-Mask maximizes the average arccosine distance [17], " + }, + { + "bbox": [ + 130, + 470, + 481, + 518 + ], + "type": "inline_equation", + "content": "\\mathrm{ArcCos}" + }, + { + "bbox": [ + 130, + 470, + 481, + 518 + ], + "type": "text", + "content": ", between the embedding of the protected face " + }, + { + "bbox": [ + 130, + 470, + 481, + 518 + ], + "type": "inline_equation", + "content": "F(\\Theta(\\mathcal{X};\\mathcal{M}_{\\mathcal{P}}^t))" + }, + { + "bbox": [ + 130, + 470, + 481, + 518 + ], + "type": "text", + "content": " and the embedding of its unprotected counterpart " + }, + { + "bbox": [ + 130, + 470, + 481, + 518 + ], + "type": "inline_equation", + "content": "F(\\Psi(\\mathcal{FD}(\\mathcal{X})))" + }, + { + "bbox": [ + 130, + 470, + 481, + 518 + ], + "type": "text", + "content": " extracted by every FR model " + }, + { + "bbox": [ + 130, + 470, + 481, + 518 + ], + "type": "inline_equation", + "content": "F \\in \\mathcal{T}" + }, + { + "bbox": [ + 130, + 470, + 481, + 518 + ], + "type": "text", + "content": " in the team:" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 155, + 523, + 481, + 548 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 155, + 523, + 481, + 548 + ], + "spans": [ + { + "bbox": [ + 155, + 523, + 481, + 548 + ], + "type": "interline_equation", + "content": "\\mathcal {L} _ {\\mathrm {P r o t e c t}} (\\mathcal {X}, \\boldsymbol {\\mathcal {M}} _ {\\mathcal {P}} ^ {t}; \\boldsymbol {\\mathcal {T}}) = \\frac {- 1}{| \\boldsymbol {\\mathcal {T}} |} \\sum_ {F \\in \\boldsymbol {\\mathcal {T}}} \\operatorname {A r c C o s} (F (\\Psi (\\mathcal {F D} (\\mathcal {X}))), F (\\Theta (\\mathcal {X}; \\boldsymbol {\\mathcal {M}} _ {\\mathcal {P}} ^ {t}))). \\quad (5)", + "image_path": "15beb0bb041071ad0fa1de485141a1fe2dbfdb8444e284b2489c1a6d9181dfd2.jpg" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 132, + 563, + 299, + 575 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 563, + 299, + 575 + ], + "spans": [ + { + "bbox": [ + 132, + 563, + 299, + 575 + ], + "type": "text", + "content": "4.2 Perceptibility Optimization" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 130, + 582, + 482, + 666 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 582, + 482, + 666 + ], + "spans": [ + { + "bbox": [ + 130, + 582, + 482, + 666 + ], + "type": "text", + "content": "Preserving the visual quality of the facial image is necessary to maintain the usability of the protected image in practice. Hence, we incorporate a perceptibility optimization term in Equation 3. The idea is to minimize the perceptual difference between unprotected and protected images while learning a user's facial signature, such that removing the facial signature from a given facial image will have minimal impact on the visual quality of the protected version of the original image. We use the structural similarity (SSIM) [35] to capture perceptual" + } + ] + } + ], + "index": 11 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "spans": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "type": "text", + "content": "6" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 166, + 91, + 240, + 101 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 166, + 91, + 240, + 101 + ], + "spans": [ + { + "bbox": [ + 166, + 91, + 240, + 101 + ], + "type": "text", + "content": "K.-H. Chow et al." + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 5 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 148, + 114, + 264, + 164 + ], + "blocks": [ + { + "bbox": [ + 148, + 114, + 264, + 164 + ], + "lines": [ + { + "bbox": [ + 148, + 114, + 264, + 164 + ], + "spans": [ + { + "bbox": [ + 148, + 114, + 264, + 164 + ], + "type": "image", + "image_path": "acfa6a0a3d7c3632dd4e74d91c52c2872adda959c3939046218ea3a1ed790434.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 157, + 165, + 253, + 174 + ], + "lines": [ + { + "bbox": [ + 157, + 165, + 253, + 174 + ], + "spans": [ + { + "bbox": [ + 157, + 165, + 253, + 174 + ], + "type": "text", + "content": "(a) Two-model: 28 Options" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_caption" + } + ], + "index": 2 + }, + { + "type": "image", + "bbox": [ + 148, + 177, + 258, + 227 + ], + "blocks": [ + { + "bbox": [ + 148, + 177, + 258, + 227 + ], + "lines": [ + { + "bbox": [ + 148, + 177, + 258, + 227 + ], + "spans": [ + { + "bbox": [ + 148, + 177, + 258, + 227 + ], + "type": "image", + "image_path": "aa4effa16cff7aa7d27ffbb5a95d34b8fd6a0665dae2bd748fceef68b42e3689.jpg" + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 154, + 228, + 257, + 236 + ], + "lines": [ + { + "bbox": [ + 154, + 228, + 257, + 236 + ], + "spans": [ + { + "bbox": [ + 154, + 228, + 257, + 236 + ], + "type": "text", + "content": "(b) Three-model: 56 Options" + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "image_caption" + } + ], + "index": 4 + }, + { + "type": "image", + "bbox": [ + 293, + 114, + 470, + 225 + ], + "blocks": [ + { + "bbox": [ + 293, + 114, + 470, + 225 + ], + "lines": [ + { + "bbox": [ + 293, + 114, + 470, + 225 + ], + "spans": [ + { + "bbox": [ + 293, + 114, + 470, + 225 + ], + "type": "image", + "image_path": "96640aaf397d553c48391063d68ae3e920c701982d7df828e5bf54eafdd97ce7.jpg" + } + ] + } + ], + "index": 6, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 280, + 235, + 481, + 280 + ], + "lines": [ + { + "bbox": [ + 280, + 235, + 481, + 280 + ], + "spans": [ + { + "bbox": [ + 280, + 235, + 481, + 280 + ], + "type": "text", + "content": "Fig. 5: Out of 28 options of two-model ensembles from a pool of eight models, the ensemble selected by our approach (green) leads to much better protection than a randomly selected one (red)." + } + ] + } + ], + "index": 7, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 132, + 246, + 277, + 280 + ], + "lines": [ + { + "bbox": [ + 132, + 246, + 277, + 280 + ], + "spans": [ + { + "bbox": [ + 132, + 246, + 277, + 280 + ], + "type": "text", + "content": "Fig.4: Focal diversity provides a strong indicator of protection effectiveness for selecting an ensemble." + } + ] + } + ], + "index": 8, + "angle": 0, + "type": "image_caption" + } + ], + "index": 6 + }, + { + "bbox": [ + 132, + 303, + 470, + 316 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 303, + 470, + 316 + ], + "spans": [ + { + "bbox": [ + 132, + 303, + 470, + 316 + ], + "type": "text", + "content": "differences, as it has been shown to align with the human perception system:" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 139, + 321, + 480, + 346 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 139, + 321, + 480, + 346 + ], + "spans": [ + { + "bbox": [ + 139, + 321, + 480, + 346 + ], + "type": "interline_equation", + "content": "\\mathcal {L} _ {\\text {P e r c e p t}} \\left(\\mathcal {X}, \\boldsymbol {\\mathcal {M}} _ {\\mathcal {P}} ^ {t}; \\omega\\right) = \\lambda_ {\\text {S S I M}} \\max \\left[ \\frac {1 - \\operatorname {S S I M} \\left(\\Psi \\left(\\boldsymbol {\\mathcal {F D}} (\\mathcal {X})\\right) , \\Theta \\left(\\mathcal {X} ; \\boldsymbol {\\mathcal {M}} _ {\\mathcal {P}} ^ {t}\\right)\\right)}{2} - \\omega , 0 \\right]. \\tag {6}", + "image_path": "41bbf6195ca5c8d5722f6ecc84df9cca967da66c77558fe755738f7c79ae9b91.jpg" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 132, + 352, + 480, + 413 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 352, + 480, + 413 + ], + "spans": [ + { + "bbox": [ + 132, + 352, + 480, + 413 + ], + "type": "text", + "content": "The parameters, " + }, + { + "bbox": [ + 132, + 352, + 480, + 413 + ], + "type": "inline_equation", + "content": "\\omega" + }, + { + "bbox": [ + 132, + 352, + 480, + 413 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 132, + 352, + 480, + 413 + ], + "type": "inline_equation", + "content": "\\lambda_{\\mathrm{SSIM}}" + }, + { + "bbox": [ + 132, + 352, + 480, + 413 + ], + "type": "text", + "content": ", enable two features. " + }, + { + "bbox": [ + 132, + 352, + 480, + 413 + ], + "type": "inline_equation", + "content": "\\omega" + }, + { + "bbox": [ + 132, + 352, + 480, + 413 + ], + "type": "text", + "content": " controls the SSIM degradation. Any SSIM degradation greater than " + }, + { + "bbox": [ + 132, + 352, + 480, + 413 + ], + "type": "inline_equation", + "content": "2\\omega" + }, + { + "bbox": [ + 132, + 352, + 480, + 413 + ], + "type": "text", + "content": " will cause the term to be non-zero, making the optimization adjust the P3-Mask reducing its impact on image quality. " + }, + { + "bbox": [ + 132, + 352, + 480, + 413 + ], + "type": "inline_equation", + "content": "\\lambda_{\\mathrm{SSIM}}" + }, + { + "bbox": [ + 132, + 352, + 480, + 413 + ], + "type": "text", + "content": " balances the importance of privacy protection and perceptibility. We use dynamic scheduling [31] such that it will be adjusted automatically." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 132, + 430, + 358, + 442 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 430, + 358, + 442 + ], + "spans": [ + { + "bbox": [ + 132, + 430, + 358, + 442 + ], + "type": "text", + "content": "4.3 Focal Diversity Ensemble Optimization" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 130, + 450, + 481, + 581 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 450, + 481, + 581 + ], + "spans": [ + { + "bbox": [ + 130, + 450, + 481, + 581 + ], + "type": "text", + "content": "For each original image and its protected version by removing the facial signature learned up to the current iteration, we use two FR models in Figure 3 to generate two embeddings, which serve as alternative channels to learn a high-quality facial signature. Choosing FR models that best complement each other can achieve better protection than randomly selected FR models. We use an ensemble optimization method to select the best " + }, + { + "bbox": [ + 130, + 450, + 481, + 581 + ], + "type": "inline_equation", + "content": "k" + }, + { + "bbox": [ + 130, + 450, + 481, + 581 + ], + "type": "text", + "content": "-model ensemble among a pool of " + }, + { + "bbox": [ + 130, + 450, + 481, + 581 + ], + "type": "inline_equation", + "content": "N" + }, + { + "bbox": [ + 130, + 450, + 481, + 581 + ], + "type": "text", + "content": " FR models. The key idea is to find an ensemble with members making decorrelated mistakes. To quantify this behavior, we leverage the focal diversity framework [13,14,37]. As shown in Figure 4, the higher the focal diversity of the ensemble (a green dot), the stronger the P3-Mask will be in protecting against other FR models not used during the P3-Mask generation process." + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 130, + 582, + 482, + 666 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 582, + 482, + 666 + ], + "spans": [ + { + "bbox": [ + 130, + 582, + 482, + 666 + ], + "type": "text", + "content": "Given a collection of " + }, + { + "bbox": [ + 130, + 582, + 482, + 666 + ], + "type": "inline_equation", + "content": "N" + }, + { + "bbox": [ + 130, + 582, + 482, + 666 + ], + "type": "text", + "content": " FR models " + }, + { + "bbox": [ + 130, + 582, + 482, + 666 + ], + "type": "inline_equation", + "content": "\\{F_1,\\dots,F_N\\}" + }, + { + "bbox": [ + 130, + 582, + 482, + 666 + ], + "type": "text", + "content": ", we first identify the negative samples for each FR model " + }, + { + "bbox": [ + 130, + 582, + 482, + 666 + ], + "type": "inline_equation", + "content": "F" + }, + { + "bbox": [ + 130, + 582, + 482, + 666 + ], + "type": "text", + "content": " by locating validation images that " + }, + { + "bbox": [ + 130, + 582, + 482, + 666 + ], + "type": "inline_equation", + "content": "F" + }, + { + "bbox": [ + 130, + 582, + 482, + 666 + ], + "type": "text", + "content": " fails to recognize their true identity. To rank ensembles of size " + }, + { + "bbox": [ + 130, + 582, + 482, + 666 + ], + "type": "inline_equation", + "content": "S" + }, + { + "bbox": [ + 130, + 582, + 482, + 666 + ], + "type": "text", + "content": ", we enumerate all " + }, + { + "bbox": [ + 130, + 582, + 482, + 666 + ], + "type": "inline_equation", + "content": "\\binom{N}{S}" + }, + { + "bbox": [ + 130, + 582, + 482, + 666 + ], + "type": "text", + "content": " combinations. For each combination (ensemble) " + }, + { + "bbox": [ + 130, + 582, + 482, + 666 + ], + "type": "inline_equation", + "content": "\\pmb{\\tau}" + }, + { + "bbox": [ + 130, + 582, + 482, + 666 + ], + "type": "text", + "content": ", we consider each member to be the focal model " + }, + { + "bbox": [ + 130, + 582, + 482, + 666 + ], + "type": "inline_equation", + "content": "F_{\\mathrm{focal}}" + }, + { + "bbox": [ + 130, + 582, + 482, + 666 + ], + "type": "text", + "content": " and use its negative samples to statistically estimate the level of negative correlation " + }, + { + "bbox": [ + 130, + 582, + 482, + 666 + ], + "type": "inline_equation", + "content": "\\lambda_{\\mathrm{focal}}(\\pmb{\\tau};F_{\\mathrm{focal}})" + }, + { + "bbox": [ + 130, + 582, + 482, + 666 + ], + "type": "text", + "content": " between " + }, + { + "bbox": [ + 130, + 582, + 482, + 666 + ], + "type": "inline_equation", + "content": "F_{\\mathrm{focal}}" + }, + { + "bbox": [ + 130, + 582, + 482, + 666 + ], + "type": "text", + "content": " and the remaining models in " + }, + { + "bbox": [ + 130, + 582, + 482, + 666 + ], + "type": "inline_equation", + "content": "\\pmb{\\tau}" + }, + { + "bbox": [ + 130, + 582, + 482, + 666 + ], + "type": "text", + "content": ", computed by measuring the degree of disagreements using the" + } + ] + } + ], + "index": 14 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 184, + 91, + 448, + 102 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 184, + 91, + 448, + 102 + ], + "spans": [ + { + "bbox": [ + 184, + 91, + 448, + 102 + ], + "type": "text", + "content": "Personalized Privacy Protection Mask Against Unauthorized FR" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 474, + 91, + 480, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 474, + 91, + 480, + 100 + ], + "spans": [ + { + "bbox": [ + 474, + 91, + 480, + 100 + ], + "type": "text", + "content": "7" + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 6 + }, + { + "para_blocks": [ + { + "bbox": [ + 130, + 116, + 482, + 224 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 116, + 482, + 224 + ], + "spans": [ + { + "bbox": [ + 130, + 116, + 482, + 224 + ], + "type": "text", + "content": "generalized non-pairwise measure [27]. The same procedure is repeated by considering each member in the ensemble as the focal model, and the focal diversity of the ensemble " + }, + { + "bbox": [ + 130, + 116, + 482, + 224 + ], + "type": "inline_equation", + "content": "\\pmb{\\mathcal{T}}" + }, + { + "bbox": [ + 130, + 116, + 482, + 224 + ], + "type": "text", + "content": " is finalized as " + }, + { + "bbox": [ + 130, + 116, + 482, + 224 + ], + "type": "inline_equation", + "content": "d_{\\mathrm{focal}}(\\pmb {\\mathcal{T}}) = \\frac{1}{S}\\sum_{F_{\\mathrm{focal}}\\in \\pmb{\\mathcal{T}}}[1 - \\lambda_{\\mathrm{focal}}(\\pmb {\\mathcal{T}};F_{\\mathrm{focal}})]" + }, + { + "bbox": [ + 130, + 116, + 482, + 224 + ], + "type": "text", + "content": ". Given a team size " + }, + { + "bbox": [ + 130, + 116, + 482, + 224 + ], + "type": "inline_equation", + "content": "S" + }, + { + "bbox": [ + 130, + 116, + 482, + 224 + ], + "type": "text", + "content": ", the ensemble with the highest focal diversity can be deployed. As shown in Figure 5, the selected two-model team (green bar) is indeed the one leading to the strongest protection among all 28 options, which is over " + }, + { + "bbox": [ + 130, + 116, + 482, + 224 + ], + "type": "inline_equation", + "content": "20\\%" + }, + { + "bbox": [ + 130, + 116, + 482, + 224 + ], + "type": "text", + "content": " stronger than a randomly selected ensemble (red bar). Note that we only need to analyze the failures of individual FR models, which is significantly more efficient than generating P3-Mask for each option and conducting evaluation." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 132, + 242, + 398, + 255 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 242, + 398, + 255 + ], + "spans": [ + { + "bbox": [ + 132, + 242, + 398, + 255 + ], + "type": "text", + "content": "5 Online Image Protection and Refinement" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 130, + 266, + 481, + 292 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 266, + 481, + 292 + ], + "spans": [ + { + "bbox": [ + 130, + 266, + 481, + 292 + ], + "type": "text", + "content": "The user " + }, + { + "bbox": [ + 130, + 266, + 481, + 292 + ], + "type": "inline_equation", + "content": "\\mathcal{P}" + }, + { + "bbox": [ + 130, + 266, + 481, + 292 + ], + "type": "text", + "content": " obtains her P3-Mask " + }, + { + "bbox": [ + 130, + 266, + 481, + 292 + ], + "type": "inline_equation", + "content": "\\mathcal{M}_{\\mathcal{P}}" + }, + { + "bbox": [ + 130, + 266, + 481, + 292 + ], + "type": "text", + "content": " from Chameleon. Given an image " + }, + { + "bbox": [ + 130, + 266, + 481, + 292 + ], + "type": "inline_equation", + "content": "\\mathcal{X}" + }, + { + "bbox": [ + 130, + 266, + 481, + 292 + ], + "type": "text", + "content": ", she can obfuscate her facial identity on her device by applying the P3-Mask:" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 202, + 300, + 481, + 312 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 202, + 300, + 481, + 312 + ], + "spans": [ + { + "bbox": [ + 202, + 300, + 481, + 312 + ], + "type": "interline_equation", + "content": "\\operatorname {M A S K} (\\mathcal {X}; \\boldsymbol {\\mathcal {M}} _ {\\mathcal {P}}) = \\operatorname {C L I P} _ {[ 0, 2 5 5 ]} (\\mathcal {F D} (\\mathcal {X}) - \\mathcal {Q R} (\\boldsymbol {\\mathcal {M}} _ {\\mathcal {P}})), \\tag {7}", + "image_path": "c5906c7950fe1b7f246a19ad8518e433f17af5956a423b7d6510041e5eb27347.jpg" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 130, + 319, + 481, + 342 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 319, + 481, + 342 + ], + "spans": [ + { + "bbox": [ + 130, + 319, + 481, + 342 + ], + "type": "text", + "content": "which is the protected facial region to be put in the original image " + }, + { + "bbox": [ + 130, + 319, + 481, + 342 + ], + "type": "inline_equation", + "content": "\\mathcal{X}" + }, + { + "bbox": [ + 130, + 319, + 481, + 342 + ], + "type": "text", + "content": " to produce the protected version " + }, + { + "bbox": [ + 130, + 319, + 481, + 342 + ], + "type": "inline_equation", + "content": "\\mathcal{X}'" + }, + { + "bbox": [ + 130, + 319, + 481, + 342 + ], + "type": "text", + "content": " for sharing." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 130, + 342, + 481, + 403 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 342, + 481, + 403 + ], + "spans": [ + { + "bbox": [ + 130, + 342, + 481, + 403 + ], + "type": "text", + "content": "Chameleon supports authorizing trusted third parties to conduct FR on a user's facial images in a space-time efficient manner without transmitting and storing multiple versions of the same photo (one protected for public sharing and one unprotected for internal FR). By sharing the P3-Mask " + }, + { + "bbox": [ + 130, + 342, + 481, + 403 + ], + "type": "inline_equation", + "content": "\\mathcal{M}_{\\mathcal{P}}" + }, + { + "bbox": [ + 130, + 342, + 481, + 403 + ], + "type": "text", + "content": " as the key, the third parties can de-obfuscate the facial signature by unmasking:" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 194, + 411, + 481, + 423 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 194, + 411, + 481, + 423 + ], + "spans": [ + { + "bbox": [ + 194, + 411, + 481, + 423 + ], + "type": "interline_equation", + "content": "\\operatorname {U N M A S K} \\left(\\mathcal {X} ^ {\\prime}; \\boldsymbol {\\mathcal {M}} _ {\\mathcal {P}}\\right) = \\operatorname {C L I P} _ {[ 0, 2 5 5 ]} \\left(\\mathcal {F D} \\left(\\mathcal {X} ^ {\\prime}\\right) + \\mathcal {Q R} \\left(\\boldsymbol {\\mathcal {M}} _ {\\mathcal {P}}\\right)\\right). \\tag {8}", + "image_path": "c9cdab331a150df480dcdad605564f2e8733c73c40100dc0a448599a5656deb2.jpg" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 130, + 430, + 481, + 479 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 430, + 481, + 479 + ], + "spans": [ + { + "bbox": [ + 130, + 430, + 481, + 479 + ], + "type": "text", + "content": "In Section 6.2, we will show two properties of this process: (i) the signature-restored photos can be used for FR with no accuracy degradation, and (ii) the protection can only be done by the P3-Mask of the same person in the facial image, and the restoration is successful only if the same key is used." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 132, + 497, + 302, + 511 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 497, + 302, + 511 + ], + "spans": [ + { + "bbox": [ + 132, + 497, + 302, + 511 + ], + "type": "text", + "content": "6 Experimental Evaluation" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 130, + 521, + 482, + 666 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 521, + 482, + 666 + ], + "spans": [ + { + "bbox": [ + 130, + 521, + 482, + 666 + ], + "type": "text", + "content": "We conduct experiments on FaceScrub [26] and LFW [20]. To provide a detailed analysis, ten celebrities are randomly selected as users on FaceScrub (Table 1). For each user, we split their facial images into three parts: (i) " + }, + { + "bbox": [ + 130, + 521, + 482, + 666 + ], + "type": "inline_equation", + "content": "10\\%" + }, + { + "bbox": [ + 130, + 521, + 482, + 666 + ], + "type": "text", + "content": " are used as probe images (2nd column), (ii) " + }, + { + "bbox": [ + 130, + 521, + 482, + 666 + ], + "type": "inline_equation", + "content": "70\\%" + }, + { + "bbox": [ + 130, + 521, + 482, + 666 + ], + "type": "text", + "content": " are used as gallery images and are used by Chameleon to train the mask (3rd column), and (iii) " + }, + { + "bbox": [ + 130, + 521, + 482, + 666 + ], + "type": "inline_equation", + "content": "20\\%" + }, + { + "bbox": [ + 130, + 521, + 482, + 666 + ], + "type": "text", + "content": " are also used as gallery images but unseen during the training process (4th column). The last part is crucial to evaluate the protection of unseen images of the user. In total, we have 123 facial images with \"unknown identities.\" Following relevant works [31, 40], all facial images of other people are included in the gallery dataset as noise. The same splitting method is also used for LFW. By default, the results reported in this paper focus on FaceScrub due to similar observations on LFW. The source code is available at https://github.com/git-disl/Chameleon." + } + ] + } + ], + "index": 11 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "spans": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "type": "text", + "content": "8" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 166, + 91, + 240, + 101 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 166, + 91, + 240, + 101 + ], + "spans": [ + { + "bbox": [ + 166, + 91, + 240, + 101 + ], + "type": "text", + "content": "K.-H. Chow et al." + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 7 + }, + { + "para_blocks": [ + { + "type": "table", + "bbox": [ + 133, + 147, + 304, + 269 + ], + "blocks": [ + { + "bbox": [ + 132, + 114, + 302, + 137 + ], + "lines": [ + { + "bbox": [ + 132, + 114, + 302, + 137 + ], + "spans": [ + { + "bbox": [ + 132, + 114, + 302, + 137 + ], + "type": "text", + "content": "Table 1: Ten example celebrities (users) on FaceScrub for analysis." + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 133, + 147, + 304, + 269 + ], + "lines": [ + { + "bbox": [ + 133, + 147, + 304, + 269 + ], + "spans": [ + { + "bbox": [ + 133, + 147, + 304, + 269 + ], + "type": "table", + "html": "
NameProbe ImagesGallery Images
SeenUnseenTotal
Morena Baccarin118223105
Bradley Cooper128925114
America Ferrera1611332145
Gerard Butler118224106
Eva Longoria129126117
Melissa Egan128724111
Kim Catrall139627123
Allison Janney128926115
Roma Downey117822100
Steve Carell139628124
Others (Noise)///48368
Total Number of Gallery Images:49528
", + "image_path": "59053062356c7a22fe029fd45a1e98c3c5df1d8883027e5b9de792afd3cc213f.jpg" + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "table_body" + } + ], + "index": 4 + }, + { + "type": "table", + "bbox": [ + 310, + 147, + 478, + 232 + ], + "blocks": [ + { + "bbox": [ + 307, + 114, + 481, + 137 + ], + "lines": [ + { + "bbox": [ + 307, + 114, + 481, + 137 + ], + "spans": [ + { + "bbox": [ + 307, + 114, + 481, + 137 + ], + "type": "text", + "content": "Table 2: The collection of publicly available pretrained FR models." + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 310, + 147, + 478, + 232 + ], + "lines": [ + { + "bbox": [ + 310, + 147, + 478, + 232 + ], + "spans": [ + { + "bbox": [ + 310, + 147, + 478, + 232 + ], + "type": "table", + "html": "
Model IDNeural Arch.Training DatasetFR Acc.
EN-MCEfficientNetMS-Celeb-1M94.94%
RN50-GLResNet50Glint360K96.68%
RN50-MCResNet50MS-Celeb-1M89.25%
RN50-VFResNet50VGG-Face294.30%
RN50-WFResNet50WebFace600K96.72%
RN18-MCResNet18MS-Celeb-1M82.64%
RN34-MCResNet34MS-Celeb-1M84.49%
RN100-MCResNet100MS-Celeb-1M91.85%
", + "image_path": "43ebcbe7d0527bf65f2910daeded75f360c748a51ed65d8983553651252f5699.jpg" + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "table_body" + } + ], + "index": 5 + }, + { + "bbox": [ + 130, + 289, + 482, + 482 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 289, + 482, + 482 + ], + "spans": [ + { + "bbox": [ + 130, + 289, + 482, + 482 + ], + "type": "text", + "content": "Public clouds (e.g., Azure) now require manual approval to use their FR services. Hence, we believe that privacy intruders will opt for deploying pretrained FR models, as many high-quality ones are available on the Internet and can be used out of the box. Due to the large number of possible FR algorithms and architectures, we first focus on the eight FR models listed in Table 2, which are based on ArcFace [17], the state-of-the-art FR algorithm, with varying neural architectures and training datasets. Chameleon can be easily extended to incorporate any FR models by simply adding them to the collection. Nonetheless, in Section 6.4, we will show that thanks to the focal diversity-optimized teaming, the P3-Mask generated from a collection of ArcFace-only models can also be effective in protecting against privacy intruders using models of other FR algorithms, such as FaceNet [30] and MagFace [24]. By default, we use the two-model team (EN-MC, RN50-GL) with the highest focal diversity (Section 4.3). The P3-Mask is trained for 50 epochs on NVIDIA RTX 2080 SUPER GPU with " + }, + { + "bbox": [ + 130, + 289, + 482, + 482 + ], + "type": "inline_equation", + "content": "\\eta = 0.001" + }, + { + "bbox": [ + 130, + 289, + 482, + 482 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 130, + 289, + 482, + 482 + ], + "type": "inline_equation", + "content": "|\\mathcal{B}| = 4" + }, + { + "bbox": [ + 130, + 289, + 482, + 482 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 130, + 289, + 482, + 482 + ], + "type": "inline_equation", + "content": "\\epsilon = 0.063" + }, + { + "bbox": [ + 130, + 289, + 482, + 482 + ], + "type": "text", + "content": ", and " + }, + { + "bbox": [ + 130, + 289, + 482, + 482 + ], + "type": "inline_equation", + "content": "\\omega = 0.03" + }, + { + "bbox": [ + 130, + 289, + 482, + 482 + ], + "type": "text", + "content": ". We provide details for reproducibility and additional results in the appendix." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 132, + 497, + 282, + 508 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 497, + 282, + 508 + ], + "spans": [ + { + "bbox": [ + 132, + 497, + 282, + 508 + ], + "type": "text", + "content": "6.1 Protection Effectiveness" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 130, + 514, + 482, + 552 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 514, + 482, + 552 + ], + "spans": [ + { + "bbox": [ + 130, + 514, + 482, + 552 + ], + "type": "text", + "content": "Table 3 shows the personalized mask for seven users on FaceScrub (other users are available in the appendix). Different users have distinct facial signatures (P3-Mask) learned by Chameleon. We apply these masks to their respective gallery" + } + ] + } + ], + "index": 8 + }, + { + "type": "table", + "bbox": [ + 133, + 602, + 481, + 662 + ], + "blocks": [ + { + "bbox": [ + 132, + 570, + 481, + 592 + ], + "lines": [ + { + "bbox": [ + 132, + 570, + 481, + 592 + ], + "spans": [ + { + "bbox": [ + 132, + 570, + 481, + 592 + ], + "type": "text", + "content": "Table 3: The P3-Mask (rescaled) for seven users on FaceScrub. Chameleon learns the distinct facial signature for each of them." + } + ] + } + ], + "index": 9, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 133, + 602, + 481, + 662 + ], + "lines": [ + { + "bbox": [ + 133, + 602, + 481, + 662 + ], + "spans": [ + { + "bbox": [ + 133, + 602, + 481, + 662 + ], + "type": "table", + "html": "
M. BaccarinB. CooperA. FerreraG. ButlerE. LongoriaM. EganK. Cattrall
", + "image_path": "b51a87f887545091a74583b8414fefb7b6b1ee9e6c9b9c1db1472c0ab2e0cd28.jpg" + } + ] + } + ], + "index": 10, + "angle": 0, + "type": "table_body" + } + ], + "index": 10 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 184, + 91, + 448, + 102 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 184, + 91, + 448, + 102 + ], + "spans": [ + { + "bbox": [ + 184, + 91, + 448, + 102 + ], + "type": "text", + "content": "Personalized Privacy Protection Mask Against Unauthorized FR" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 474, + 92, + 480, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 474, + 92, + 480, + 100 + ], + "spans": [ + { + "bbox": [ + 474, + 92, + 480, + 100 + ], + "type": "text", + "content": "9" + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 8 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 136, + 114, + 305, + 171 + ], + "blocks": [ + { + "bbox": [ + 136, + 114, + 305, + 171 + ], + "lines": [ + { + "bbox": [ + 136, + 114, + 305, + 171 + ], + "spans": [ + { + "bbox": [ + 136, + 114, + 305, + 171 + ], + "type": "image", + "image_path": "0b250f4eeca1c84abfbf5b040210e3290995710c7c564eb2c315f8d4100c2cc6.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 179, + 171, + 262, + 180 + ], + "lines": [ + { + "bbox": [ + 179, + 171, + 262, + 180 + ], + "spans": [ + { + "bbox": [ + 179, + 171, + 262, + 180 + ], + "type": "text", + "content": "(a) Dataset: FaceScrub" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 131, + 189, + 480, + 212 + ], + "lines": [ + { + "bbox": [ + 131, + 189, + 480, + 212 + ], + "spans": [ + { + "bbox": [ + 131, + 189, + 480, + 212 + ], + "type": "text", + "content": "Fig. 6: Protection effectiveness on two benchmarks using Chameleon, OPOM [40], TIP-IM [39], and Fawkes [31] against FR models unknown to them." + } + ] + } + ], + "index": 6, + "angle": 0, + "type": "image_caption" + } + ], + "index": 2 + }, + { + "type": "image", + "bbox": [ + 310, + 114, + 472, + 171 + ], + "blocks": [ + { + "bbox": [ + 310, + 114, + 472, + 171 + ], + "lines": [ + { + "bbox": [ + 310, + 114, + 472, + 171 + ], + "spans": [ + { + "bbox": [ + 310, + 114, + 472, + 171 + ], + "type": "image", + "image_path": "9dd95d6eeed8a8388885d77cde22d6fae71ae328457928e787de1b65c03f3a1e.jpg" + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 361, + 171, + 426, + 180 + ], + "lines": [ + { + "bbox": [ + 361, + 171, + 426, + 180 + ], + "spans": [ + { + "bbox": [ + 361, + 171, + 426, + 180 + ], + "type": "text", + "content": "(b) Dataset: LFW" + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "image_caption" + } + ], + "index": 4 + }, + { + "bbox": [ + 130, + 241, + 480, + 302 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 241, + 480, + 302 + ], + "spans": [ + { + "bbox": [ + 130, + 241, + 480, + 302 + ], + "type": "text", + "content": "images and test the FR accuracy using probe images. All probe images can be correctly identified when no protection mechanism is used. Under Chameleon, the probe images should be matched to an incorrect identity, resulting in a low FR accuracy. Hence, we define an evaluation metric, " + }, + { + "bbox": [ + 130, + 241, + 480, + 302 + ], + "type": "inline_equation", + "content": "\\underline{\\text{Protection Success Rate}}" + }, + { + "bbox": [ + 130, + 241, + 480, + 302 + ], + "type": "text", + "content": " (PSR), to be " + }, + { + "bbox": [ + 130, + 241, + 480, + 302 + ], + "type": "inline_equation", + "content": "(100 - \\text{FR Accuracy})" + }, + { + "bbox": [ + 130, + 241, + 480, + 302 + ], + "type": "text", + "content": " reported in percentages." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 130, + 304, + 481, + 460 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 304, + 481, + 460 + ], + "spans": [ + { + "bbox": [ + 130, + 304, + 481, + 460 + ], + "type": "text", + "content": "Chameleon can protect against unknown FR models. We compare the cross-model protection of Chameleon with OPOM [40], the only anti-FR approach generating one privacy protection mask for each person as Chameleon, and both TIP-IM [39] and Fawkes [31], which conduct per-image optimization. For a fair comparison, all methods are set with the same perturbation budget. Figure 6 summarizes the results on both datasets. All five FR models are unknown to both protection mechanisms. We make two observations. First, Chameleon consistently outperforms OPOM by a large margin. We found that OPOM masks do not transfer well, especially considering the end-to-end ML pipeline. Its PSR can drop by " + }, + { + "bbox": [ + 130, + 304, + 481, + 460 + ], + "type": "inline_equation", + "content": "33\\%" + }, + { + "bbox": [ + 130, + 304, + 481, + 460 + ], + "type": "text", + "content": " because of those lossy operations. Second, Chameleon can be as competitive as those methods conducting per-image optimization. TIP-IM and Fawkes spend minutes to optimize the protection for each image. Still, Chameleon outperforms them and can be applied instantly." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 130, + 462, + 481, + 666 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 462, + 481, + 666 + ], + "spans": [ + { + "bbox": [ + 130, + 462, + 481, + 666 + ], + "type": "text", + "content": "Table 4 reports the detailed PSR for each user Chameleon achieves when the privacy intruder uses different FR models (2nd to 9th columns). We make two observations. First, when the privacy intruder uses an FR model known to the generation process, a PSR of " + }, + { + "bbox": [ + 130, + 462, + 481, + 666 + ], + "type": "inline_equation", + "content": "100\\%" + }, + { + "bbox": [ + 130, + 462, + 481, + 666 + ], + "type": "text", + "content": " can be consistently achieved (2nd to 3rd columns). Second, Chameleon can protect against unknown FR models (4th to 9th columns). Even when the FR model used by the intruder is trained on a different dataset (i.e., RN50-VF and RN50-WF) or with a different neural architecture (i.e., RN18-MC, RN34-MC, and RN100-MC) than any FR models known by Chameleon, it still provides a PSR over " + }, + { + "bbox": [ + 130, + 462, + 481, + 666 + ], + "type": "inline_equation", + "content": "90.65\\%" + }, + { + "bbox": [ + 130, + 462, + 481, + 666 + ], + "type": "text", + "content": ", meaning that facial images of a user are matched to gallery images of a different person. To demonstrate such a mismatch, in Table 5, we show the probe images from two users and their most similar gallery images found by four unknown FR models with two settings: (i) the \"Unprotected\" scenario where no one employs protection and (ii) the \"Chameleon\" scenario where the intruder scraped photos protected by our solution. Taking M. Baccarin as an example, the most similar gallery images found in the unprotected scenario belong to her. In contrast, when Chameleon is used, the same probe image is misidentified, e.g., as L. Hartley by RN50-MC." + } + ] + } + ], + "index": 9 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "spans": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "type": "text", + "content": "10" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 166, + 91, + 240, + 100 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 166, + 91, + 240, + 100 + ], + "spans": [ + { + "bbox": [ + 166, + 91, + 240, + 100 + ], + "type": "text", + "content": "K.-H. Chow et al." + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 9 + }, + { + "para_blocks": [ + { + "type": "table", + "bbox": [ + 133, + 158, + 481, + 286 + ], + "blocks": [ + { + "bbox": [ + 131, + 114, + 482, + 148 + ], + "lines": [ + { + "bbox": [ + 131, + 114, + 482, + 148 + ], + "spans": [ + { + "bbox": [ + 131, + 114, + 482, + 148 + ], + "type": "text", + "content": "Table 4: Using the two-model team with high focal diversity (2nd and 3rd columns), Chameleon can offer privacy protection even when the privacy intruder uses FR models unseen during the mask generation process (4th to 9th columns)." + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 133, + 158, + 481, + 286 + ], + "lines": [ + { + "bbox": [ + 133, + 158, + 481, + 286 + ], + "spans": [ + { + "bbox": [ + 133, + 158, + 481, + 286 + ], + "type": "table", + "html": "
NameProtection Success Rate - PSR (%)
EN-MCRN50-GLRN50-MCRN50-VFRN50-WFRN18-MCRN34-MCRN100-MCMeanStd
M. Baccarin100.00100.00100.0072.7390.91100.0090.91100.0094.329.64
B. Cooper100.00100.0083.33100.00100.00100.0091.6791.6795.836.30
A. Ferrera100.00100.0093.7575.00100.00100.0093.75100.0095.318.68
G. Butler100.00100.0090.91100.00100.0090.9190.91100.0096.594.70
E. Longoria100.00100.00100.00100.00100.00100.00100.00100.00100.000.00
M. Egan100.00100.00100.00100.00100.00100.00100.0091.6798.962.95
K. Cattrall100.00100.00100.0084.62100.00100.00100.00100.0098.085.44
A. Janney100.00100.0091.67100.0091.67100.00100.00100.0097.923.86
R. Downey100.00100.00100.0072.7390.91100.0090.91100.0094.329.64
S. Carell100.00100.0084.6292.3192.3192.3176.92100.0092.318.22
Mean100.00100.0093.5290.6596.5897.4194.4297.42
Std0.000.006.4011.164.434.187.414.15
", + "image_path": "43c8e098abcb4d1e2d73b1568ee48ff46c218e87058c9cda54d925575a18ad45.jpg" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "table_body" + } + ], + "index": 3 + }, + { + "type": "table", + "bbox": [ + 133, + 329, + 486, + 450 + ], + "blocks": [ + { + "bbox": [ + 131, + 297, + 482, + 319 + ], + "lines": [ + { + "bbox": [ + 131, + 297, + 482, + 319 + ], + "spans": [ + { + "bbox": [ + 131, + 297, + 482, + 319 + ], + "type": "text", + "content": "Table 5: The most similar gallery images on FaceScrub found by different FR models unknown to Chameleon." + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 133, + 329, + 486, + 450 + ], + "lines": [ + { + "bbox": [ + 133, + 329, + 486, + 450 + ], + "spans": [ + { + "bbox": [ + 133, + 329, + 486, + 450 + ], + "type": "table", + "html": "
Probe\nImageThe Most Similar Gallery Image & Its Identity Found by Unseen FR Models
RN50-MCRN50-VFRN34-MCRN100-MC
UnprotectedChameleonUnprotectedChameleonUnprotectedChameleonUnprotectedChameleon
M. BaccarinIdentifiedProtectedIdentifiedProtectedIdentifiedProtectedIdentifiedProtected
M. BaccarinL. HartleyM. BaccarinL. HartleyM. BaccarinC. ElectraM. BaccarinL. Hartley
B. CooperIdentifiedProtectedIdentifiedProtectedIdentifiedProtectedIdentifiedProtected
B. CooperP. WalkerB. CooperM. VartanB. CooperM. VartanB. CooperM. Vartan
", + "image_path": "ae5e2d205e2de1759f9d9c692a81167ab63ff8de595a628bf51476a764d1b514.jpg" + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "table_body" + } + ], + "index": 5 + }, + { + "bbox": [ + 131, + 471, + 265, + 483 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 131, + 471, + 265, + 483 + ], + "spans": [ + { + "bbox": [ + 131, + 471, + 265, + 483 + ], + "type": "text", + "content": "6.2 FR Service Usability" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 130, + 488, + 482, + 609 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 488, + 482, + 609 + ], + "spans": [ + { + "bbox": [ + 130, + 488, + 482, + 609 + ], + "type": "text", + "content": "Chameleon allows users to grant FR permission to trusted third parties by sharing their P3-Mask to de-obfuscate the protected image. There is no need to transmit and store an unprotected photo for internal use and a protected one for public visibility, doubling network and storage costs. Table 6 shows the results in FR accuracy. First, using the correct mask (i.e., the one used to protect the images) to reverse the protection can restore FR accuracy. Taking M. Baccarin as an example, the FR accuracy increases from " + }, + { + "bbox": [ + 130, + 488, + 482, + 609 + ], + "type": "inline_equation", + "content": "6.82\\%" + }, + { + "bbox": [ + 130, + 488, + 482, + 609 + ], + "type": "text", + "content": " before unmasking (b) to " + }, + { + "bbox": [ + 130, + 488, + 482, + 609 + ], + "type": "inline_equation", + "content": "100\\%" + }, + { + "bbox": [ + 130, + 488, + 482, + 609 + ], + "type": "text", + "content": " after unmasking (c). Second, using an incorrect mask for unmasking cannot restore the FR accuracy and can lead to even worse performance. It drops from " + }, + { + "bbox": [ + 130, + 488, + 482, + 609 + ], + "type": "inline_equation", + "content": "6.82\\%" + }, + { + "bbox": [ + 130, + 488, + 482, + 609 + ], + "type": "text", + "content": " to " + }, + { + "bbox": [ + 130, + 488, + 482, + 609 + ], + "type": "inline_equation", + "content": "1.14\\%" + }, + { + "bbox": [ + 130, + 488, + 482, + 609 + ], + "type": "text", + "content": " after unmasking (d)." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 131, + 624, + 288, + 635 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 131, + 624, + 288, + 635 + ], + "spans": [ + { + "bbox": [ + 131, + 624, + 288, + 635 + ], + "type": "text", + "content": "6.3 Protection Cost Analysis" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 130, + 641, + 481, + 665 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 641, + 481, + 665 + ], + "spans": [ + { + "bbox": [ + 130, + 641, + 481, + 665 + ], + "type": "text", + "content": "Speed and Resources. Figure 7 compares Chameleon with OPOM, TIP-IM, and Fawkes regarding the protection time per face, compute, and storage costs." + } + ] + } + ], + "index": 9 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 184, + 91, + 448, + 102 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 184, + 91, + 448, + 102 + ], + "spans": [ + { + "bbox": [ + 184, + 91, + 448, + 102 + ], + "type": "text", + "content": "Personalized Privacy Protection Mask Against Unauthorized FR" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 470, + 92, + 479, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 470, + 92, + 479, + 100 + ], + "spans": [ + { + "bbox": [ + 470, + 92, + 479, + 100 + ], + "type": "text", + "content": "11" + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 10 + }, + { + "para_blocks": [ + { + "type": "table", + "bbox": [ + 132, + 168, + 481, + 280 + ], + "blocks": [ + { + "bbox": [ + 132, + 114, + 482, + 160 + ], + "lines": [ + { + "bbox": [ + 132, + 114, + 482, + 160 + ], + "spans": [ + { + "bbox": [ + 132, + 114, + 482, + 160 + ], + "type": "text", + "content": "Table 6: The protection by Chameleon can reduce FR accuracy against FR models (b), but the user can provide the P3-Mask used to protect her images as the key for the authorized third party to reverse the protection process (c), which leads to restored FR performance. However, an incorrect key cannot restore FR and may worsen it (d)." + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 132, + 168, + 481, + 280 + ], + "lines": [ + { + "bbox": [ + 132, + 168, + 481, + 280 + ], + "spans": [ + { + "bbox": [ + 132, + 168, + 481, + 280 + ], + "type": "table", + "html": "
ScenarioFace Recognition Accuracy (%)
EN-MCRN50-GLRN50-MCRN50-VFRN50-WFRN18-MCRN34-MCRN100-MCMeanStd
User: M. Baccarin
(a) No Protection100.00100.00100.00100.00100.00100.00100.00100.00100.000.00
(b) Protected0.000.009.0918.189.099.090.009.096.826.43
(c) Correctly Unmasked100.00100.00100.00100.00100.00100.00100.00100.00100.000.00
(d) Incorrectly Unmasked0.000.000.000.009.090.000.000.001.143.21
User: B. Cooper
(a) No Protection100.00100.00100.00100.00100.00100.00100.00100.00100.000.00
(b) Protected0.000. 0016.670.000.000.008.338.334.176.30
(c) Correctly Unmasked100.00100.00100.00100.00100.00100.00100.00100.00100.000.00
(d) Incorrectly Unmasked0.000.000. 000.000.000.000.000.000.000.00
", + "image_path": "ccf695ca088ae618891d5511694f1706621df7f842f480e0bf8fc213903d40b1.jpg" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "table_body" + } + ], + "index": 3 + }, + { + "type": "image", + "bbox": [ + 154, + 291, + 214, + 356 + ], + "blocks": [ + { + "bbox": [ + 154, + 291, + 214, + 356 + ], + "lines": [ + { + "bbox": [ + 154, + 291, + 214, + 356 + ], + "spans": [ + { + "bbox": [ + 154, + 291, + 214, + 356 + ], + "type": "image", + "image_path": "c0ada8d7af1371cac2804844d97f3bc7f896e94d46a492f3084236d029994d0b.jpg" + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 130, + 364, + 482, + 398 + ], + "lines": [ + { + "bbox": [ + 130, + 364, + 482, + 398 + ], + "spans": [ + { + "bbox": [ + 130, + 364, + 482, + 398 + ], + "type": "text", + "content": "Fig. 7: Chameleon offers much better privacy protection than OPOM without sacrificing speed and resource usages. It can protect instantly and require negligible compute resources as OPOM, while TIP-IM and Fawkes are slow and costly." + } + ] + } + ], + "index": 9, + "angle": 0, + "type": "image_caption" + } + ], + "index": 4 + }, + { + "type": "image", + "bbox": [ + 216, + 291, + 274, + 356 + ], + "blocks": [ + { + "bbox": [ + 216, + 291, + 274, + 356 + ], + "lines": [ + { + "bbox": [ + 216, + 291, + 274, + 356 + ], + "spans": [ + { + "bbox": [ + 216, + 291, + 274, + 356 + ], + "type": "image", + "image_path": "f9d6a25330c8829d82e56b24e0174b87015eef8742c9b4d9ced91233ce04b07f.jpg" + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "image_body" + } + ], + "index": 5 + }, + { + "type": "image", + "bbox": [ + 277, + 291, + 335, + 356 + ], + "blocks": [ + { + "bbox": [ + 277, + 291, + 335, + 356 + ], + "lines": [ + { + "bbox": [ + 277, + 291, + 335, + 356 + ], + "spans": [ + { + "bbox": [ + 277, + 291, + 335, + 356 + ], + "type": "image", + "image_path": "d94dc64126842e1ab631d873825ca79fc62ca79a0d7b111bfe091d3bbd045ffa.jpg" + } + ] + } + ], + "index": 6, + "angle": 0, + "type": "image_body" + } + ], + "index": 6 + }, + { + "type": "image", + "bbox": [ + 339, + 291, + 398, + 356 + ], + "blocks": [ + { + "bbox": [ + 339, + 291, + 398, + 356 + ], + "lines": [ + { + "bbox": [ + 339, + 291, + 398, + 356 + ], + "spans": [ + { + "bbox": [ + 339, + 291, + 398, + 356 + ], + "type": "image", + "image_path": "aa18156c72669038ebbb258d2174e3b003517e59b7bf42117c01540a525f1dd8.jpg" + } + ] + } + ], + "index": 7, + "angle": 0, + "type": "image_body" + } + ], + "index": 7 + }, + { + "type": "image", + "bbox": [ + 400, + 291, + 460, + 356 + ], + "blocks": [ + { + "bbox": [ + 400, + 291, + 460, + 356 + ], + "lines": [ + { + "bbox": [ + 400, + 291, + 460, + 356 + ], + "spans": [ + { + "bbox": [ + 400, + 291, + 460, + 356 + ], + "type": "image", + "image_path": "43f16377adc45961eedfc4aa4c8ca73c101a322c1aff2f2c279dbd0c48c09728.jpg" + } + ] + } + ], + "index": 8, + "angle": 0, + "type": "image_body" + } + ], + "index": 8 + }, + { + "bbox": [ + 130, + 424, + 482, + 521 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 424, + 482, + 521 + ], + "spans": [ + { + "bbox": [ + 130, + 424, + 482, + 521 + ], + "type": "text", + "content": "Chameleon and OPOM produce masks that are applicable to any facial image of the same user. The protection can be completed in 0.0076 seconds. Note that Chameleon is significantly more effective than OPOM, as described in Figure 6. Instead, TIP-IM and Fawkes require per-image optimization and take 105.12 seconds to complete the protection with GPUs for acceleration and storage for FR models. Chameleon only needs to conduct simple arithmetic operations. Only the P3-Mask needs to be stored on the user device. Hence, it can be deployed with instant protection even on an edge device with weak computing power." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 130, + 522, + 482, + 666 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 522, + 482, + 666 + ], + "spans": [ + { + "bbox": [ + 130, + 522, + 482, + 666 + ], + "type": "text", + "content": "Image Quality. Chameleon can well preserve image quality. Table 7 provides two examples for four users, contrasting the facial images with no protection (3rd and 5th columns) with the ones with their facial signature removed by Chameleon (4th and 6th columns). They capture different variations of facial images: M. Baccarin's images have different lighting conditions, B. Cooper's have different face sizes, E. Longoria's have different postures, and G. Bulter's have different expressions. The protected images are visually similar to the unprotected counterparts, but they will not be matched to the clean images of the corresponding person. Indeed, Chameleon can generate higher-quality images than existing methods. As reported in Table 8, Chameleon not only offers better protection (Figure 6) with a similar cost (Figure 7) as OPOM but also achieves much better image quality measured in SSIM (0.9493 vs 0.8839)." + } + ] + } + ], + "index": 11 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "spans": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "type": "text", + "content": "12" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 166, + 91, + 240, + 101 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 166, + 91, + 240, + 101 + ], + "spans": [ + { + "bbox": [ + 166, + 91, + 240, + 101 + ], + "type": "text", + "content": "K.-H. Chow et al." + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 11 + }, + { + "para_blocks": [ + { + "type": "table", + "bbox": [ + 136, + 147, + 357, + 311 + ], + "blocks": [ + { + "bbox": [ + 132, + 114, + 482, + 137 + ], + "lines": [ + { + "bbox": [ + 132, + 114, + 482, + 137 + ], + "spans": [ + { + "bbox": [ + 132, + 114, + 482, + 137 + ], + "type": "text", + "content": "Table 7: The P3-Mask can be applied to any facial im- Table 8: Chameleon offers ages of the same person while preserving image quality. much better protection while" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 136, + 147, + 357, + 311 + ], + "lines": [ + { + "bbox": [ + 136, + 147, + 357, + 311 + ], + "spans": [ + { + "bbox": [ + 136, + 147, + 357, + 311 + ], + "type": "table", + "html": "
UserMaskSample Image 1Sample Image 2
No ProtectionProtectedNo ProtectionProtected
M. Baccarin
B. Cooper
E. Longoria
G. Image
", + "image_path": "d4feeb1ae43ca7d63837a93041546e9976dc59da9c0f846544ca4e5f61061e4f.jpg" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "table_body" + } + ], + "index": 3 + }, + { + "type": "table", + "bbox": [ + 375, + 158, + 468, + 318 + ], + "blocks": [ + { + "bbox": [ + 375, + 158, + 468, + 318 + ], + "lines": [ + { + "bbox": [ + 375, + 158, + 468, + 318 + ], + "spans": [ + { + "bbox": [ + 375, + 158, + 468, + 318 + ], + "type": "table", + "html": "
Example
Original SSIM: 1.0000
Chameleon SSIM: 0.9493
OPOM SSIM: 0.8839
TIP-IM SSIM: 0.8850
Fawkes SSIM: 0.9612
", + "image_path": "b64dc63214ea2c3e635ae630f6018fbd459c3f50f5ac49baa6841ba1a212d690.jpg" + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "table_body" + } + ], + "index": 4 + }, + { + "bbox": [ + 132, + 342, + 315, + 355 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 342, + 315, + 355 + ], + "spans": [ + { + "bbox": [ + 132, + 342, + 315, + 355 + ], + "type": "text", + "content": "6.4 Focal Diversity-based Teaming" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 130, + 365, + 482, + 462 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 365, + 482, + 462 + ], + "spans": [ + { + "bbox": [ + 130, + 365, + 482, + 462 + ], + "type": "text", + "content": "Chameleon automatically selects a high-quality team for deployment with a specified budget. Table 9 compares the detailed PSR using the most diverse and the least diverse teams with (a) two and (b) three models. In both cases, the most diverse team outperforms the least diverse one, which is only effective when the privacy intruder uses the FR model known in the team. The selection of high-quality teams is non-trivial because we observe that composing a team of FR models with the highest FR accuracy is not always a good option, and a team of models having different neural architectures is not always the top priority [13]." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 130, + 462, + 481, + 487 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 462, + 481, + 487 + ], + "spans": [ + { + "bbox": [ + 130, + 462, + 481, + 487 + ], + "type": "text", + "content": "In Figure 8, we further show that the P3-Mask generated from a carefully-chosen team can protect against FR models of unknown algorithms. Specifically," + } + ] + } + ], + "index": 7 + }, + { + "type": "table", + "bbox": [ + 133, + 552, + 485, + 663 + ], + "blocks": [ + { + "bbox": [ + 130, + 509, + 482, + 542 + ], + "lines": [ + { + "bbox": [ + 130, + 509, + 482, + 542 + ], + "spans": [ + { + "bbox": [ + 130, + 509, + 482, + 542 + ], + "type": "text", + "content": "Table 9: The most diverse and the least diverse teams with (a) two or (b) three models identified by our focal diversity-based teaming method. The most diverse teams offer significantly better protection in terms of protection success rates." + } + ] + } + ], + "index": 8, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 133, + 552, + 485, + 663 + ], + "lines": [ + { + "bbox": [ + 133, + 552, + 485, + 663 + ], + "spans": [ + { + "bbox": [ + 133, + 552, + 485, + 663 + ], + "type": "table", + "html": "
Protection TeamProtection Success Rate - PSR (%)
EN-MCRN50-GLRN50-MCRN50-VFRN50-WFRN18-MCRN34-MCRN100-MCMeanStd
(a) Two-model Teams
Most Diverse:(EN-MC, RN50-GL)100.00100.0093.5290.6596.5897.4194.4297.4296.253.23
Least Diverse:(RN18-MC, RN34-MC)11.3860.16100.0055.2873.17100.00100.0080.4972.5628.54
(b) Three-model Teams
Most Diverse:(EN-MC, RN50-GL, RN50-WF)100.00100.00100.00100.0092.6893.50100.00100.0098.273.00
Least Diverse:(RN18-MC, RN34-MC, RN50-MC)2.4478.86100.0052.0384.55100.00100.0093.5076.4231.82
", + "image_path": "9458f030ebe4eed0eb0e34d235396c19d5705138af6465bdca3a0d0a44de6897.jpg" + } + ] + } + ], + "index": 9, + "angle": 0, + "type": "table_body" + } + ], + "index": 9 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 184, + 91, + 448, + 102 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 184, + 91, + 448, + 102 + ], + "spans": [ + { + "bbox": [ + 184, + 91, + 448, + 102 + ], + "type": "text", + "content": "Personalized Privacy Protection Mask Against Unauthorized FR" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 470, + 91, + 481, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 470, + 91, + 481, + 100 + ], + "spans": [ + { + "bbox": [ + 470, + 91, + 481, + 100 + ], + "type": "text", + "content": "13" + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 12 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 140, + 114, + 276, + 183 + ], + "blocks": [ + { + "bbox": [ + 140, + 114, + 276, + 183 + ], + "lines": [ + { + "bbox": [ + 140, + 114, + 276, + 183 + ], + "spans": [ + { + "bbox": [ + 140, + 114, + 276, + 183 + ], + "type": "image", + "image_path": "b369f682755ec1966b4de25ceb8421d5d5ec12fd6c27737248f0001a7362c308.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 132, + 187, + 284, + 220 + ], + "lines": [ + { + "bbox": [ + 132, + 187, + 284, + 220 + ], + "spans": [ + { + "bbox": [ + 132, + 187, + 284, + 220 + ], + "type": "text", + "content": "Fig. 8: The most diverse teams allow Chameleon to be effective against FR models of unknown algorithms." + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_caption" + } + ], + "index": 2 + }, + { + "type": "image", + "bbox": [ + 291, + 119, + 477, + 183 + ], + "blocks": [ + { + "bbox": [ + 291, + 119, + 477, + 183 + ], + "lines": [ + { + "bbox": [ + 291, + 119, + 477, + 183 + ], + "spans": [ + { + "bbox": [ + 291, + 119, + 477, + 183 + ], + "type": "image", + "image_path": "bed3d7bad28822bdaaff5ee932ea111a0ec9135f62865cde53bcad95f29288ce.jpg" + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 287, + 187, + 482, + 220 + ], + "lines": [ + { + "bbox": [ + 287, + 187, + 482, + 220 + ], + "spans": [ + { + "bbox": [ + 287, + 187, + 482, + 220 + ], + "type": "text", + "content": "Fig. 9: The PSR of Chameleon against adaptive intruders using different strategies to wash out the patterns in protected images." + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "image_caption" + } + ], + "index": 4 + }, + { + "bbox": [ + 130, + 255, + 482, + 315 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 255, + 482, + 315 + ], + "spans": [ + { + "bbox": [ + 130, + 255, + 482, + 315 + ], + "type": "text", + "content": "we consider the most and the least diverse three-model teams in Table 9(b). Even though both teams include only FR models based on ArcFace [17], the most diverse one can effectively protect against FR models based on FaceNet [30] or MagFace [24]. The protection effectiveness can be further strengthened by incorporating more FR models into the collection." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 131, + 334, + 370, + 346 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 131, + 334, + 370, + 346 + ], + "spans": [ + { + "bbox": [ + 131, + 334, + 370, + 346 + ], + "type": "text", + "content": "6.5 Chameleon Against Adaptive Adversaries" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 130, + 355, + 482, + 438 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 355, + 482, + 438 + ], + "spans": [ + { + "bbox": [ + 130, + 355, + 482, + 438 + ], + "type": "text", + "content": "An adaptive adversary may run Chameleon to produce the P3-Mask for each person and de-obfuscate their photos. When probe images need to be identified, they will be matched against a \"clean\" face database. However, it is infeasible, as we show in Section 6.2, that the FR accuracy can only be restored when the masks used for obfuscation and de-obfuscation are identical. It is impossible to reproduce the same P3-Mask used by the user because it requires the same set of clean images and random states." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 130, + 439, + 482, + 511 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 439, + 482, + 511 + ], + "spans": [ + { + "bbox": [ + 130, + 439, + 482, + 511 + ], + "type": "text", + "content": "Alternatively, an adaptive adversary may wash out patterns introduced by Chameleon before performing FR. We consider three popular lightweight and task-agnostic methods studied in the adversarial example domain [18] in Figure 9, including (a) JPEG compression [15], (b) Gaussian Filter, and (d) Median Filter [38]. The results show that their effectiveness is limited on Chameleon. The consideration of the end-to-end ML pipeline increases the robustness of P3-Mask." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 132, + 531, + 226, + 544 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 531, + 226, + 544 + ], + "spans": [ + { + "bbox": [ + 132, + 531, + 226, + 544 + ], + "type": "text", + "content": "7 Conclusions" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 130, + 557, + 482, + 666 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 557, + 482, + 666 + ], + "spans": [ + { + "bbox": [ + 130, + 557, + 482, + 666 + ], + "type": "text", + "content": "We have presented Chameleon against unauthorized FR. Chameleon generates a P3-Mask for each user with cross-image and perceptibility optimizations to offer (i) instant and lightweight protection on any facial images of the same user and (ii) preservation of image quality. Also, we have shown that P3-Mask enables cost-effective de-obfuscation for authorizing FR services. Such an authorization process can only be conducted by the one with the identical mask used for protection. Chameleon also features a focal diversity-optimized teaming method to select a high-quality FR team to generate P3-Mask with strong robustness against unknown FR models." + } + ] + } + ], + "index": 11 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "spans": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "type": "text", + "content": "14" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 166, + 91, + 240, + 100 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 166, + 91, + 240, + 100 + ], + "spans": [ + { + "bbox": [ + 166, + 91, + 240, + 100 + ], + "type": "text", + "content": "K.-H. Chow et al." + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 13 + }, + { + "para_blocks": [ + { + "bbox": [ + 133, + 114, + 240, + 129 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 114, + 240, + 129 + ], + "spans": [ + { + "bbox": [ + 133, + 114, + 240, + 129 + ], + "type": "text", + "content": "Acknowledgments" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 130, + 142, + 482, + 204 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 142, + 482, + 204 + ], + "spans": [ + { + "bbox": [ + 130, + 142, + 482, + 204 + ], + "type": "text", + "content": "This research is partially sponsored by the NSF CISE grants 2302720, 2312758, 2038029, an IBM faculty award, and a grant from CISCO Edge AI program. It is part of the PhD dissertation of the first author, who graduated from Georgia Tech in Spring 2023. The first author acknowledges the support of the IBM PhD Fellowship in 2022-2023 and the support from the HKU-CS Start-up Fund." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 132, + 224, + 198, + 237 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 224, + 198, + 237 + ], + "spans": [ + { + "bbox": [ + 132, + 224, + 198, + 237 + ], + "type": "text", + "content": "References" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 134, + 251, + 480, + 665 + ], + "type": "list", + "angle": 0, + "index": 19, + "blocks": [ + { + "bbox": [ + 138, + 251, + 480, + 285 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 251, + 480, + 285 + ], + "spans": [ + { + "bbox": [ + 138, + 251, + 480, + 285 + ], + "type": "text", + "content": "1. Brazilian retailer quizzed over facial recognition tech. https://www.zdnet.com/article/brazilian-retailer-quizzed-over-facial-recognition-tech/ (2019)" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 138, + 285, + 480, + 329 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 285, + 480, + 329 + ], + "spans": [ + { + "bbox": [ + 138, + 285, + 480, + 329 + ], + "type": "text", + "content": "2. Facial recognition in the united states: Privacy concerns and legal developments. https://www.asisonline.org/security-management-magazine/monthlyissues/security-technology/archive/2021/december/facial-recognition-in-the-us-privacy-concerns-and-legal-developments/ (2021)" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 138, + 331, + 480, + 363 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 331, + 480, + 363 + ], + "spans": [ + { + "bbox": [ + 138, + 331, + 480, + 363 + ], + "type": "text", + "content": "3. The secretive company that might end privacy as we know it. https://www.nytimes.com/2020/01/18/technology/clearview-privacy-facial-recognition.html (2021)" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 138, + 364, + 345, + 374 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 364, + 345, + 374 + ], + "spans": [ + { + "bbox": [ + 138, + 364, + 345, + 374 + ], + "type": "text", + "content": "4. Clearview AI. https://www.clearview.ai/ (2023)" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 138, + 376, + 480, + 397 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 376, + 480, + 397 + ], + "spans": [ + { + "bbox": [ + 138, + 376, + 480, + 397 + ], + "type": "text", + "content": "5. Facial recognition and identity risk. https://www.equifax.co.uk/resources/identity-protection/facial-recognition-and-identity-risk.html (2023)" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 138, + 398, + 327, + 409 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 398, + 327, + 409 + ], + "spans": [ + { + "bbox": [ + 138, + 398, + 327, + 409 + ], + "type": "text", + "content": "6. HuggingFace. https://huggingface.co (2023)" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 138, + 410, + 480, + 431 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 410, + 480, + 431 + ], + "spans": [ + { + "bbox": [ + 138, + 410, + 480, + 431 + ], + "type": "text", + "content": "7. Insightface: 2d and 3d face analysis project. https://github.com/deepinsight/insightface (2023)" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 138, + 432, + 308, + 442 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 432, + 308, + 442 + ], + "spans": [ + { + "bbox": [ + 138, + 432, + 308, + 442 + ], + "type": "text", + "content": "8. PimEyes. https://pimeyes.com/ (2023)" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 138, + 443, + 480, + 475 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 443, + 480, + 475 + ], + "spans": [ + { + "bbox": [ + 138, + 443, + 480, + 475 + ], + "type": "text", + "content": "9. Borgwardt, K.M., Gretton, A., Rasch, M.J., Kriegel, H.P., Scholkopf, B., Smola, A.J.: Integrating structured biological data by kernel maximum mean discrepancy. Bioinformatics 22(14), e49-e57 (2006)" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 134, + 476, + 480, + 520 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 134, + 476, + 480, + 520 + ], + "spans": [ + { + "bbox": [ + 134, + 476, + 480, + 520 + ], + "type": "text", + "content": "10. Cherepanova, V., Goldblum, M., Foley, H., Duan, S., Dickerson, J., Taylor, G., Goldstein, T.: Lowkey: Leveraging adversarial attacks to protect social media users from facial recognition. In: International Conference on Learning Representations (ICLR) (2021)" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 134, + 521, + 480, + 564 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 134, + 521, + 480, + 564 + ], + "spans": [ + { + "bbox": [ + 134, + 521, + 480, + 564 + ], + "type": "text", + "content": "11. Choi, Y., Choi, M., Kim, M., Ha, J.W., Kim, S., Choo, J.: Stargan: Unified generative adversarial networks for multi-domain image-to-image translation. In: Proceedings of the IEEE conference on computer vision and pattern recognition. pp. 8789-8797 (2018)" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 134, + 565, + 480, + 598 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 134, + 565, + 480, + 598 + ], + "spans": [ + { + "bbox": [ + 134, + 565, + 480, + 598 + ], + "type": "text", + "content": "12. Chow, K.H., Hu, S., Huang, T., Ilhan, F., Wei, W., Liu, L.: Diversity-driven privacy protection masks against unauthorized face recognition. Proceedings on Privacy Enhancing Technologies 4, 381-392 (2024)" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 134, + 599, + 480, + 632 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 134, + 599, + 480, + 632 + ], + "spans": [ + { + "bbox": [ + 134, + 599, + 480, + 632 + ], + "type": "text", + "content": "13. Chow, K.H., Liu, L.: Robust object detection fusion against deception. In: Proceedings of the 27th ACM SIGKDD Conference on Knowledge Discovery & Data Mining. pp. 2703-2713 (2021)" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 134, + 632, + 480, + 665 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 134, + 632, + 480, + 665 + ], + "spans": [ + { + "bbox": [ + 134, + 632, + 480, + 665 + ], + "type": "text", + "content": "14. Chow, K.H., Liu, L.: Boosting object detection ensembles with error diversity. In: 2022 IEEE International Conference on Data Mining (ICDM). pp. 903-908. IEEE (2022)" + } + ] + } + ], + "index": 18 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 184, + 91, + 447, + 102 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 184, + 91, + 447, + 102 + ], + "spans": [ + { + "bbox": [ + 184, + 91, + 447, + 102 + ], + "type": "text", + "content": "Personalized Privacy Protection Mask Against Unauthorized FR" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 470, + 91, + 481, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 470, + 91, + 481, + 100 + ], + "spans": [ + { + "bbox": [ + 470, + 91, + 481, + 100 + ], + "type": "text", + "content": "15" + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 14 + }, + { + "para_blocks": [ + { + "bbox": [ + 132, + 116, + 481, + 666 + ], + "type": "list", + "angle": 0, + "index": 18, + "blocks": [ + { + "bbox": [ + 133, + 116, + 481, + 161 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 116, + 481, + 161 + ], + "spans": [ + { + "bbox": [ + 133, + 116, + 481, + 161 + ], + "type": "text", + "content": "15. Das, N., Shanbhogue, M., Chen, S.T., Hohman, F., Li, S., Chen, L., Kounavis, M.E., Chau, D.H.: Shield: Fast, practical defense and vaccination for deep learning usingJPEG compression. In: Proceedings of the 24th ACM SIGKDD International Conference on Knowledge Discovery & Data Mining. pp. 196-204 (2018)" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 132, + 162, + 481, + 184 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 162, + 481, + 184 + ], + "spans": [ + { + "bbox": [ + 132, + 162, + 481, + 184 + ], + "type": "text", + "content": "16. Deb, D., Zhang, J., Jain, A.K.: Advfaces: Adversarial face synthesis. In: 2020 IEEE International Joint Conference on Biometrics (IJCB). pp. 1-10. IEEE (2020)" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 132, + 185, + 481, + 217 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 185, + 481, + 217 + ], + "spans": [ + { + "bbox": [ + 132, + 185, + 481, + 217 + ], + "type": "text", + "content": "17. Deng, J., Guo, J., Xue, N., Zafeiriou, S.: Arcface: Additive angular margin loss for deep face recognition. In: Proceedings of the IEEE/CVF conference on computer vision and pattern recognition. pp. 4690-4699 (2019)" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 132, + 218, + 481, + 239 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 218, + 481, + 239 + ], + "spans": [ + { + "bbox": [ + 132, + 218, + 481, + 239 + ], + "type": "text", + "content": "18. Guo, C., Rana, M., Cisse, M., Van Der Maaten, L.: Countering adversarial images using input transformations. arXiv preprint arXiv:1711.00117 (2017)" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 132, + 240, + 481, + 284 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 240, + 481, + 284 + ], + "spans": [ + { + "bbox": [ + 132, + 240, + 481, + 284 + ], + "type": "text", + "content": "19. Hu, S., Liu, X., Zhang, Y., Li, M., Zhang, L.Y., Jin, H., Wu, L.: Protecting facial privacy: generating adversarial identity masks via style-robust makeup transfer. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 15014-15023 (2022)" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 132, + 285, + 481, + 318 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 285, + 481, + 318 + ], + "spans": [ + { + "bbox": [ + 132, + 285, + 481, + 318 + ], + "type": "text", + "content": "20. Huang, G.B., Ramesh, M., Berg, T., Learned-Miller, E.: Labeled faces in the wild: A database for studying face recognition in unconstrained environments. Tech. Rep. 07-49, University of Massachusetts, Amherst (October 2007)" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 132, + 319, + 481, + 352 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 319, + 481, + 352 + ], + "spans": [ + { + "bbox": [ + 132, + 319, + 481, + 352 + ], + "type": "text", + "content": "21. Li, T., Lin, L.: Anonymousnet: Natural face de-identification with measurable privacy. In: Proceedings of the IEEE/CVF conference on computer vision and pattern recognition workshops. pp. 0–0 (2019)" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 132, + 352, + 481, + 396 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 352, + 481, + 396 + ], + "spans": [ + { + "bbox": [ + 132, + 352, + 481, + 396 + ], + "type": "text", + "content": "22. Liu, L., Zhou, B., Zou, Z., Yeh, S.C., Zheng, L.: A smart unstaffed retail shop based on artificial intelligence and iot. In: 2018 IEEE 23rd International workshop on computer aided modeling and design of communication links and networks (CAMAD). pp. 1-4. IEEE (2018)" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 132, + 396, + 481, + 430 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 396, + 481, + 430 + ], + "spans": [ + { + "bbox": [ + 132, + 396, + 481, + 430 + ], + "type": "text", + "content": "23. Lugaresi, C., Tang, J., Nash, H., McClanahan, C., Uboweja, E., Hays, M., Zhang, F., Chang, C.L., Yong, M.G., Lee, J., et al.: Mediapipe: A framework for building perception pipelines. arXiv preprint arXiv:1906.08172 (2019)" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 132, + 430, + 481, + 464 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 430, + 481, + 464 + ], + "spans": [ + { + "bbox": [ + 132, + 430, + 481, + 464 + ], + "type": "text", + "content": "24. Meng, Q., Zhao, S., Huang, Z., Zhou, F.: Magface: A universal representation for face recognition and quality assessment. In: Proceedings of the IEEE/CVF conference on computer vision and pattern recognition. pp. 14225-14234 (2021)" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 132, + 464, + 481, + 486 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 464, + 481, + 486 + ], + "spans": [ + { + "bbox": [ + 132, + 464, + 481, + 486 + ], + "type": "text", + "content": "25. Mirza, M., Osindero, S.: Conditional generative adversarial nets. arXiv preprint arXiv:1411.1784 (2014)" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 132, + 487, + 481, + 520 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 487, + 481, + 520 + ], + "spans": [ + { + "bbox": [ + 132, + 487, + 481, + 520 + ], + "type": "text", + "content": "26. Ng, H.W., Winkler, S.: A data-driven approach to cleaning large face datasets. In: 2014 IEEE international conference on image processing (ICIP). pp. 343-347. IEEE (2014)" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 132, + 521, + 481, + 553 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 521, + 481, + 553 + ], + "spans": [ + { + "bbox": [ + 132, + 521, + 481, + 553 + ], + "type": "text", + "content": "27. Partridge, D., Krzanowski, W.: Software diversity: practical statistics for its measurement and exploitation. Information and software technology 39(10), 707-717 (1997)" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 132, + 554, + 481, + 586 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 554, + 481, + 586 + ], + "spans": [ + { + "bbox": [ + 132, + 554, + 481, + 586 + ], + "type": "text", + "content": "28. Pinto, N., Stone, Z., Zickler, T., Cox, D.: Scaling up biologically-inspired computer vision: A case study in unconstrained face recognition on facebook. In: CVPR 2011 WORKSHOPS. pp. 35-42. IEEE (2011)" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 132, + 587, + 481, + 632 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 587, + 481, + 632 + ], + "spans": [ + { + "bbox": [ + 132, + 587, + 481, + 632 + ], + "type": "text", + "content": "29. Sahani, M., Nanda, C., Sahu, A.K., Pattnaik, B.: Web-based online embedded door access control and home security system based on face recognition. In: 2015 International Conference on Circuits, Power and Computing Technologies [ICCPCT-2015]. pp. 1-6. IEEE (2015)" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 132, + 632, + 481, + 666 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 632, + 481, + 666 + ], + "spans": [ + { + "bbox": [ + 132, + 632, + 481, + 666 + ], + "type": "text", + "content": "30. Schroff, F., Kalenichenko, D., Philbin, J.: Facenet: A unified embedding for face recognition and clustering. In: Proceedings of the IEEE conference on computer vision and pattern recognition. pp. 815-823 (2015)" + } + ] + } + ], + "index": 17 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "spans": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "type": "text", + "content": "16" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 166, + 91, + 239, + 101 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 166, + 91, + 239, + 101 + ], + "spans": [ + { + "bbox": [ + 166, + 91, + 239, + 101 + ], + "type": "text", + "content": "K.-H. Chow et al." + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 15 + }, + { + "para_blocks": [ + { + "bbox": [ + 130, + 116, + 482, + 436 + ], + "type": "list", + "angle": 0, + "index": 13, + "blocks": [ + { + "bbox": [ + 130, + 116, + 482, + 149 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 116, + 482, + 149 + ], + "spans": [ + { + "bbox": [ + 130, + 116, + 482, + 149 + ], + "type": "text", + "content": "31. Shan, S., Wenger, E., Zhang, J., Li, H., Zheng, H., Zhao, B.Y.: Fawkes: Protecting privacy against unauthorized deep learning models. In: Proceedings of the 29th USENIX Security Symposium (2020)" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 130, + 150, + 482, + 183 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 150, + 482, + 183 + ], + "spans": [ + { + "bbox": [ + 130, + 150, + 482, + 183 + ], + "type": "text", + "content": "32. Stone, Z., Zickler, T., Darrell, T.: Autotagging facebook: Social network context improves photo annotation. In: 2008 IEEE computer society conference on computer vision and pattern recognition workshops. pp. 1-8. IEEE (2008)" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 130, + 183, + 482, + 216 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 183, + 482, + 216 + ], + "spans": [ + { + "bbox": [ + 130, + 183, + 482, + 216 + ], + "type": "text", + "content": "33. Sun, Q., Tewari, A., Xu, W., Fritz, M., Theobalt, C., Schiele, B.: A hybrid model for identity obfuscation by face replacement. In: Proceedings of the European conference on computer vision (ECCV). pp. 553-569 (2018)" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 130, + 216, + 482, + 237 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 216, + 482, + 237 + ], + "spans": [ + { + "bbox": [ + 130, + 216, + 482, + 237 + ], + "type": "text", + "content": "34. Wang, M., Deng, W.: Deep face recognition: A survey. Neurocomputing 429, 215-244 (2021)" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 130, + 237, + 482, + 270 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 237, + 482, + 270 + ], + "spans": [ + { + "bbox": [ + 130, + 237, + 482, + 270 + ], + "type": "text", + "content": "35. Wang, Z., Bovik, A.C., Sheikh, H.R., Simoncelli, E.P.: Image quality assessment: from error visibility to structural similarity. IEEE transactions on image processing 13(4), 600-612 (2004)" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 130, + 270, + 482, + 293 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 270, + 482, + 293 + ], + "spans": [ + { + "bbox": [ + 130, + 270, + 482, + 293 + ], + "type": "text", + "content": "36. Wenger, E., Shan, S., Zheng, H., Zhao, B.Y.: Sok: Anti-facial recognition technology. In: 2023 IEEE Symposium on Security and Privacy (SP). pp. 134-151 (2023)" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 130, + 293, + 482, + 326 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 293, + 482, + 326 + ], + "spans": [ + { + "bbox": [ + 130, + 293, + 482, + 326 + ], + "type": "text", + "content": "37. Wu, Y., Liu, L., Xie, Z., Chow, K.H., Wei, W.: Boosting ensemble accuracy by revisiting ensemble diversity metrics. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 16469-16477 (2021)" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 130, + 326, + 482, + 346 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 326, + 482, + 346 + ], + "spans": [ + { + "bbox": [ + 130, + 326, + 482, + 346 + ], + "type": "text", + "content": "38. Xu, W., Evans, D., Qi, Y.: Feature squeezing: Detecting adversarial examples in deep neural networks (2018)" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 130, + 347, + 482, + 380 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 347, + 482, + 380 + ], + "spans": [ + { + "bbox": [ + 130, + 347, + 482, + 380 + ], + "type": "text", + "content": "39. Yang, X., Dong, Y., Pang, T., Su, H., Zhu, J., Chen, Y., Xue, H.: Towards face encryption by generating adversarial identity masks. In: Proceedings of the IEEE/CVF International Conference on Computer Vision. pp. 3897-3907 (2021)" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 130, + 380, + 482, + 402 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 380, + 482, + 402 + ], + "spans": [ + { + "bbox": [ + 130, + 380, + 482, + 402 + ], + "type": "text", + "content": "40. Zhong, Y., Deng, W.: Opom: Customized invisible cloak towards face privacy protection. IEEE Transactions on Pattern Analysis and Machine Intelligence (2022)" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 130, + 402, + 482, + 436 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 402, + 482, + 436 + ], + "spans": [ + { + "bbox": [ + 130, + 402, + 482, + 436 + ], + "type": "text", + "content": "41. Zhu, Z.A., Lu, Y.Z., Chiang, C.K.: Generating adversarial examples by makeup attacks on face recognition. In: 2019 IEEE International Conference on Image Processing (ICIP). pp. 2516-2520. IEEE (2019)" + } + ] + } + ], + "index": 12 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 184, + 91, + 447, + 101 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 184, + 91, + 447, + 101 + ], + "spans": [ + { + "bbox": [ + 184, + 91, + 447, + 101 + ], + "type": "text", + "content": "Personalized Privacy Protection Mask Against Unauthorized FR" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 470, + 92, + 480, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 470, + 92, + 480, + 100 + ], + "spans": [ + { + "bbox": [ + 470, + 92, + 480, + 100 + ], + "type": "text", + "content": "17" + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 16 + } + ], + "_backend": "vlm", + "_version_name": "2.6.4" +} \ No newline at end of file diff --git a/2024/Personalized Video Relighting With an At-Home Light Stage/0751f00d-698a-4afe-b1e5-471a9991f11a_content_list.json b/2024/Personalized Video Relighting With an At-Home Light Stage/0751f00d-698a-4afe-b1e5-471a9991f11a_content_list.json new file mode 100644 index 0000000000000000000000000000000000000000..d8b86647a061db8a1e9eab0bab30a6f7cb811c59 --- /dev/null +++ b/2024/Personalized Video Relighting With an At-Home Light Stage/0751f00d-698a-4afe-b1e5-471a9991f11a_content_list.json @@ -0,0 +1,1726 @@ +[ + { + "type": "text", + "text": "Personalized Video Relighting With an At-Home Light Stage", + "text_level": 1, + "bbox": [ + 217, + 138, + 785, + 186 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Jun Myeong Choi, Max Christman, and Roni Sengupta", + "bbox": [ + 282, + 212, + 715, + 227 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "University of North Carolina at Chapel Hill, NC 27514, USA {chedgekr,mrlc,ronisen}@cs.unc.edu", + "bbox": [ + 297, + 239, + 702, + 268 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Abstract. In this paper, we develop a personalized video relighting algorithm that produces high-quality and temporally consistent relit videos under any pose, expression, and lighting condition in real-time. Existing relighting algorithms typically rely either on publicly available synthetic data, which yields poor relighting results, or on actual light stage data which is difficult to acquire. We show that by just capturing recordings of a user watching YouTube videos on a monitor we can train a personalized algorithm capable of performing high-quality relighting under any condition. Our key contribution is a novel image-based neural relighting architecture that effectively separates the intrinsic appearance features - the geometry and reflectance of the face - from the source lighting and then combines them with the target lighting to generate a relit image. This neural architecture enables smoothing of intrinsic appearance features leading to temporally stable video relighting. Both qualitative and quantitative evaluations show that our architecture improves portrait image relighting quality and temporal consistency over state-of-the-art approaches on both casually captured 'Light Stage at Your Desk' (LSYD) and light-stage-captured 'One Light At a Time' (OLAT) datasets. Source code is available at https://github.com/chedgekorea/relighting", + "bbox": [ + 261, + 306, + 738, + 570 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Keywords: Image-based Relighting $\\cdot$ 'At Home' Light Stage", + "bbox": [ + 261, + 584, + 671, + 598 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "1 Introduction", + "text_level": 1, + "bbox": [ + 217, + 625, + 375, + 641 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "With the recent rise in popularity of video conferencing for business, educational, and personal activities, there is a significant demand for improving facial lighting. Virtually relighting our images and videos helps us improve the appearance of our faces without requiring explicit studio-quality lighting in a dedicated space or any specialized lighting expertise. Recent advances in deep neural networks have renewed interest in the problem of virtual relighting.", + "bbox": [ + 212, + 657, + 785, + 748 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Training a deep neural network for relighting requires extensive training data that includes source images paired with relit target images. One way of acquiring this data is by using a large spherical rig with numerous lights and cameras, known as a light stage [5]. While light stage data has been shown to produce high-quality relighting results [36,44,23,38,22,48], the limited availability of datasets, trained models, and access to the light stage itself has impeded further research.", + "bbox": [ + 212, + 750, + 785, + 840 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/1951539ed1a5559b57e7ce4e301eaaefd57d9678d1ca1e0db360061607e81ca4.jpg", + "image_caption": [ + "(a) LSYD" + ], + "image_footnote": [], + "bbox": [ + 217, + 143, + 313, + 218 + ], + "page_idx": 1 + }, + { + "type": "image", + "img_path": "images/79068ed969abfda3cc3af4a60fde3d37c4d59d765d1610e5f6b912f45b7224e3.jpg", + "image_caption": [ + "(b) Input Image", + "(c) Relit-Net" + ], + "image_footnote": [], + "bbox": [ + 316, + 143, + 424, + 218 + ], + "page_idx": 1 + }, + { + "type": "image", + "img_path": "images/e5ed7a9ff0b984b3fc910c98baafa9bbed2a936230c25cfc8842529b713f165d.jpg", + "image_caption": [ + "Fig. 1. We learn a personalized relighting algorithm that generates temporally consistent and high-quality portrait videos under different lighting. We train the network using recordings of users watching YouTube on a monitor, thereby creating a Light Stage at Your Desk (LSYD). We project a portion of the LDR environment map with a $180^{\\circ}$ FoV as the monitor light, while a portion of the remaining $90^{\\circ}$ FoV is mapped as the background. We can achieve a harmonization effect with the virtual background." + ], + "image_footnote": [], + "bbox": [ + 429, + 143, + 568, + 217 + ], + "page_idx": 1 + }, + { + "type": "image", + "img_path": "images/498cf8b11c53efdcb57fd436e2dac13f0e565b5e8446a7fa2bfdc4ac4ab73ee0.jpg", + "image_caption": [ + "(d) Output" + ], + "image_footnote": [], + "bbox": [ + 570, + 143, + 674, + 217 + ], + "page_idx": 1 + }, + { + "type": "image", + "img_path": "images/b92e180af5d3093e1ef4b72359ac9bb55874352b69ba10f3d37ad269ef4190ab.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 676, + 143, + 784, + 217 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "For example, One Light At a Time (Dynamic-OLAT) [44] is the only publicly available light stage relighting dataset consisting of four individuals only. As a result, researchers have often turned to synthetic data to train their relighting algorithms [31,49,13,35]. Unfortunately, existing synthetic data compromises the quality of relit images.", + "bbox": [ + 212, + 356, + 787, + 431 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "We draw inspiration from recent work by Sengupta et al. [30,26] and develop a personalized relighting model by capturing a single user's appearance while being lit by a computer monitor. However, the results and applications of Sengupta et al. [30] have several limitations. First, it requires capturing users with fixed poses and expressions, an unrealistic requirement for actual users that impedes casual capture. Second, the relighting algorithm requires a dark room with negligible ambient light, limiting the environment in which this can be applied. Third, the resulting relit video is temporally unstable, exhibiting significant flickering artifacts, making it unsuitable for relighting Zoom calls. Lastly, it requires knowledge of source lighting and thus is unable to relight arbitrary portrait images captured in the wild.", + "bbox": [ + 212, + 431, + 787, + 598 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "In this paper, we show that casually captured light stage data is sufficient to develop a high-quality temporally consistent video portrait relighting algorithm that works under arbitrary conditions (i.e. pose, expression, and ambient lighting) in real-time ( $\\sim 45$ fps). To that end, we create our own casually captured light stage dataset with varying pose, expression, and lighting, called Light Stage at Your Desk (LSYD). Our key contribution is a neural image-based relighting architecture, based on the commonly used U-Net [36,49,22,40,38], that better disentangles the source lighting from the user's intrinsic facial appearance (shape and reflectance) and then adds back the target lighting to generate a relit image. Existing image-based relighting architectures [36,30] fail to accurately separate source lighting information from intrinsic appearance features in the encoder, leading to inconsistent and temporally unstable video relighting. To this end, we introduce the light-conditioned feature normalization (LCFN) module, which performs relighting and also predicts the source lighting from an input image. The LCFN module also enables temporal stability by performing exponential smoothing of de-lit intrinsic appearance features and facilitates relighting", + "bbox": [ + 212, + 598, + 787, + 840 + ], + "page_idx": 1 + }, + { + "type": "page_number", + "text": "2", + "bbox": [ + 217, + 114, + 228, + 126 + ], + "page_idx": 1 + }, + { + "type": "header", + "text": "J. Choi et al.", + "bbox": [ + 271, + 114, + 359, + 126 + ], + "page_idx": 1 + }, + { + "type": "image", + "img_path": "images/d784e968271e27145566bd56bcd0d2a19d76403025d4ab61a5c09277c144379d.jpg", + "image_caption": [ + "Fig. 2. We highlight the key structural differences between our relighting architecture and that of [36,30]. Our approach removes source lighting information from input image features and only propagates intrinsic appearance (geometry and reflectance) features from the encoder to decoder, which results in better relighting quality and more temporal stability. In contrast [36,30] propagates entire image features from the encoder to the decoder without 'de-light', and expects the decoder to remove source lighting and add target lighting information." + ], + "image_footnote": [], + "bbox": [ + 215, + 141, + 377, + 251 + ], + "page_idx": 2 + }, + { + "type": "image", + "img_path": "images/1ced49ce92679efc4604fff228850a52d0086db1ef5eb137770013b58345d242.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 388, + 141, + 580, + 252 + ], + "page_idx": 2 + }, + { + "type": "image", + "img_path": "images/630e0bb2e4047e9f8bb240ca08bcb1f2163bdc72ee37c2f48a3fc373b40cbf58.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 591, + 141, + 787, + 251 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "of any arbitrary portrait image with unknown source lighting. We also improve the data pre-processing pipeline from Sengupta et al. [30] to make the relighting algorithm more robust to pose, expression, and ambient lighting conditions.", + "bbox": [ + 212, + 396, + 787, + 441 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "We compare our relighting network with two other algorithms: Sun et al. [36], which was originally trained on light stage data (OLAT), and Sengupta et al. [30], which was originally trained on casually captured data (albeit with fixed pose, expression, and no ambient lighting). For a fair comparison, we train all algorithms for personalized relighting using the same data pre-processing steps and loss functions on 5 individuals from our LSYD dataset and 4 individuals from OLAT [44]. Our network outperforms Sun et al. [36] and Sengupta et al. [30] by $22.3\\%$ and $23.6\\%$ respectively on the LSYD dataset and by $23.5\\%$ and $25.6\\%$ on the OLAT dataset, in terms of LPIPS. Qualitatively our method produces superior relighting in terms of color and quality. We further show that our approach is more temporally consistent, leading to less flickering than Sun et al. [36] or Sengupta et al. [30]. Detailed ablation studies show that LCFN and source monitor prediction improves relighting quality, feature, and source monitor smoothing improves temporal consistency, and data pre-processing improves robustness to pose and expression.", + "bbox": [ + 212, + 443, + 787, + 669 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "To summarize, our main contributions are as follows:", + "bbox": [ + 215, + 670, + 599, + 684 + ], + "page_idx": 2 + }, + { + "type": "list", + "sub_type": "text", + "list_items": [ + "- We show that casually captured Light Stage at Your Desk (LSYD) data can be used to build a high-quality temporally consistent personalized video relighting algorithm without requiring access to an expensive light stage setup.", + "- We introduce a novel video relighting architecture that separates the source lighting from the user's intrinsic appearance features and then adds back the target lighting, leading to improved relighting and temporal consistency for videos.", + "- While our relighting network focuses on 'at home' Light Stage (LSYD dataset) outperforming state-of-the-art algorithms, we also perform equally well on actual Light Stage captured OLAT [44] datasets, and any arbitrary portrait image captured 'in the wild'." + ], + "bbox": [ + 212, + 686, + 787, + 839 + ], + "page_idx": 2 + }, + { + "type": "header", + "text": "Personalized Video Relighting With an At-Home Light Stage", + "bbox": [ + 323, + 114, + 732, + 128 + ], + "page_idx": 2 + }, + { + "type": "page_number", + "text": "3", + "bbox": [ + 774, + 116, + 784, + 126 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "2 Related Work", + "text_level": 1, + "bbox": [ + 215, + 143, + 387, + 160 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Portrait relighting methods change the appearance of the face to match a target lighting condition. This can be expressed through lighting parameters (e.g., an environment map, spherical harmonics, directional lighting, etc.) or through a reference image of another person. Our approach relights a portrait image to a lighting condition expressed through a low dynamic range (LDR) image representing the image on the monitor.", + "bbox": [ + 212, + 175, + 785, + 265 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Image based relighting. Before the rise of deep learning, attempts were made at non deep learning image based relighting [34,24,33]. Shu et al. [34] introduced a face relighting approach that uses a mass-transport formulation for the transfer of illumination between images. Peers et al. [24] demonstrated a method for relighting portrait images with flat lighting to match specific target environments, incorporating a reference subject database for approximation. Shih et al. [33] adopted a multiscale technique to transfer local image statistics from reference portraits onto new ones, facilitating the matching of attributes like local contrast and overall lighting direction. Recent advancements in deep learning have caused significant shifts to the landscape of portrait relighting [25,11,20]. Research on ratio images for relighting [49,13] was conducted, utilizing public datasets and employing methods based on ratio images. However, this is limited to synthetic data, resulting in a significant domain gap with real data. The widespread application of light stages [5] in gathering data has enabled numerous groundbreaking research endeavors[36,44,38,28]. However, these methods rely on capturing data with a light stage, which are expensive and inaccessible. Instead, our approach builds a personalized relighting algorithm using casually captured videos from the desk recording setup introduced in Sengupta et al. [30]. In contrast to Sengupta et al. [30], we can collect data during daily computer usage by minimizing numerous constraints, eliminating the need for specific efforts in data collection.", + "bbox": [ + 212, + 266, + 787, + 568 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Relighting with explicit decomposition. Creating a virtual relighting dataset through the utilization of synthetic human models to train their networks has been carried out in some studies [31,18,42]. However, when using synthetic data to train a neural network, the large domain gap between synthetic and real data impacts the model's performance on real data. In contrast, others [12,4,27,2,37] generate relit images by using public datasets and employ methods based on 3D model rendering. Specifically, Hou et al. [12] takes a more advanced approach by introducing explicit components, where rays originating from the face intersect with other parts of the facial geometry to create relit images. Moreover, the extensive use of light stages [5] has enabled numerous innovative studies [22,40,23,10,43,7,21,47,38,41,19] in this domain. Some researchers have incorporated explicit elements such as albedo, normals, specular maps, and diffuse maps into their methodologies [22,40,23,7]. Others have taken a physics-based rendering approach [10,43] to resolve these issues. Other papers [21,47] aim to manipulate lighting conditions and generate images under different lighting scenarios using texture information. However, explicit decomposition methods require ground truth (GT) data for these intrinsic components to train, either from synthetic data or real light stage data. Obtaining GT data for 'at-home'", + "bbox": [ + 212, + 568, + 789, + 840 + ], + "page_idx": 3 + }, + { + "type": "page_number", + "text": "4", + "bbox": [ + 217, + 114, + 230, + 126 + ], + "page_idx": 3 + }, + { + "type": "header", + "text": "J. Choi et al.", + "bbox": [ + 271, + 114, + 359, + 127 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "captures is impossible, thus we focus on image-based relighting. In contrast, recent studies aim to streamline the capture process, using a mobile phone camera [32] or a sun stage [39] instead of a light stage. Nevertheless, due to their reliance on per-scene optimization, both of these papers lack the capability for real-time relighting and to generalize to unseen appearances of the individual and are only limited to the particular capture. Instead, our approach performs image-based relighting enabling real-time temporally consistent video relighting, at $\\sim 45$ fps.", + "bbox": [ + 212, + 146, + 782, + 252 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "3 Method", + "text_level": 1, + "bbox": [ + 215, + 273, + 330, + 289 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Our setup is similar to Gerstner et al. [9] and Sengupta et al. [30], where a user's face is captured while illuminated by their monitor. By capturing multiple videos of the user's face along with the video on their monitor, we build our 'at home' light stage dataset. We then use these data to train a personalized portrait relighting algorithm that can render the user's face under arbitrary lighting conditions. Specifically, given a portrait image $I_{\\mathrm{src}}$ , corresponding source monitor lighting $L_{\\mathrm{src}}$ , and target monitor lighting $L_{\\mathrm{trg}}$ , our aim is to learn a function $G$ that relights $I_{\\mathrm{src}}$ under $L_{\\mathrm{trg}}$ :", + "bbox": [ + 212, + 304, + 784, + 425 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\n\\hat {I} _ {\\mathrm {t r g}}, \\hat {L} _ {\\mathrm {s r c}} = G \\left(I _ {\\mathrm {s r c}}, L _ {\\mathrm {s r c}}, L _ {\\mathrm {t r g}}; \\theta_ {G}\\right). \\tag {1}\n$$\n", + "text_format": "latex", + "bbox": [ + 385, + 436, + 784, + 454 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Note that our formulation can be used for scenarios where the source lighting is unknown by simply replacing the input source lighting with the predicted source lighting, unlike previous approaches [30].", + "bbox": [ + 212, + 463, + 782, + 508 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "In the following sections, we outline our methodology for portrait video relighting using a monitor as a light stage. Sec. 3.1 outlines strategies for constructing training data pairs from casually captured videos that allow flexibility in facial expression, pose, and ambient lighting. In Sec. 3.2, we introduce our relighting network architecture that disentangles lighting from intrinsic appearance using light-conditioned feature normalization, leading to high-quality relit images. In Sec. 3.3, we propose additional techniques which enforce temporal consistency and eliminate flickering, also using LCFN. Finally, in Sec. 3.4, we discuss how to train our relighting network.", + "bbox": [ + 212, + 508, + 784, + 643 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "3.1 Constructing training data pairs", + "text_level": 1, + "bbox": [ + 215, + 665, + 529, + 681 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "While past work [30] imposed requirements of a neutral pose, expression, and dark room, we loosen these constraints to allow subjects in any conditions. The only constraint we maintain is that the room lighting shall not overpower the light emitted from the monitor. For example, if the capture occurs in front of a window with bright sunlight, the light from the monitor will have minimal impact on the face. As in Sengupta et al. [30], we aim to generate source and monitor image pairs $(I_{\\mathrm{src}}, L_{\\mathrm{src}})$ , as well as target image and target monitor pairs $(I_{\\mathrm{trg}}, L_{\\mathrm{trg}})$ , such that we can train our network to produce $\\tilde{I}_{\\mathrm{trg}}$ where $I_{\\mathrm{trg}}$ is the pseudo ground truth. However, due to unrestricted subject movement during data collection, there is a lack of pixel-aligned data, making random pairs", + "bbox": [ + 212, + 688, + 785, + 840 + ], + "page_idx": 4 + }, + { + "type": "header", + "text": "Personalized Video Relighting With an At-Home Light Stage", + "bbox": [ + 323, + 114, + 732, + 128 + ], + "page_idx": 4 + }, + { + "type": "page_number", + "text": "5", + "bbox": [ + 774, + 116, + 784, + 126 + ], + "page_idx": 4 + }, + { + "type": "image", + "img_path": "images/92c0bb905334e053dd6310840d1766b1017349265b63ae1111b14933f7e37fcf.jpg", + "image_caption": [ + "Fig. 3. We first de-light the input image features extracted by the U-Net encoder using Adaptive Instance Normalization (AdaIN) guided by the lighting features extracted from the source lighting with a Light Encoder. We then pass these light-normalized encoder features to the decoder of the U-Net and apply another set of AdaIN guided by the features extracted from the target lighting with the Light Encoder. We additionally predict source lighting from the U-Net encoder using a Light Decoder." + ], + "image_footnote": [], + "bbox": [ + 217, + 143, + 787, + 324 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "unsuitable. Previous work [30] utilized segmentation for pairing. However, we observed that segmentation is ineffective at finding pairs of images with the same pose and expression. Thus, we instead use facial keypoint detection [16] to obtain source and target image pairs.", + "bbox": [ + 212, + 446, + 787, + 508 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "3.2 Relighting network architecture", + "text_level": 1, + "bbox": [ + 214, + 527, + 524, + 544 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Our network architecture, as illustrated in Fig. 3, is built upon the well-established U-Net [29]. This architecture is comprised of an encoder and a decoder with skip connections, which are commonly used in existing portrait relighting algorithms [36,30,49,13,38]. Our U-Net's encoder, similar to Sengupta et al. [30] and Sun et al. [36], processes the source portrait $I_{\\mathrm{src}}$ by applying multiple convolutional layers of varying strides (1 or 2). This process progressively reduces spatial resolution while increasing the number of channels, yielding a latent feature space. The decoder performs the inverse function of the encoder by upsampling from the latent features and simultaneously skip-connecting to intermediate features from the encoder. These skip connections transport high-frequency shape and appearance information from the encoder to the decoder, ultimately resulting in the generation of a realistic relit image. However, they also carry source illumination features from the encoder to the decoder, leading to subpar relighting quality and temporal flickering. To address this issue, we introduce light-conditioned feature normalization (LCFN) for the skip-connected features to better disentangle lighting features from intrinsic appearance features.", + "bbox": [ + 212, + 551, + 787, + 792 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "To disentangle lighting and intrinsic appearance components from the encoded features - i.e. to de-light - we first predict the source lighting from the encoder features. In contrast to Sun et al. [36], which predicts the illumination", + "bbox": [ + 212, + 795, + 787, + 840 + ], + "page_idx": 5 + }, + { + "type": "page_number", + "text": "6", + "bbox": [ + 217, + 114, + 228, + 126 + ], + "page_idx": 5 + }, + { + "type": "header", + "text": "J. Choi et al.", + "bbox": [ + 271, + 114, + 359, + 126 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "$\\hat{L}_{\\mathrm{src}}$ corresponding to the source image $I_{\\mathrm{src}}$ using the final encoded features, we take a different approach. We extract features at intermediate steps within the encoder, downsample them, and concatenate them using a confidence learning approach [14] to predict the illumination $\\hat{L}_{\\mathrm{src}}$ .", + "bbox": [ + 212, + 145, + 784, + 205 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "The LCFN module uses the lighting features generated by the lighting encoder to perform Adaptive Instance Normalization (AdaIN) [17] on the encoder features. We begin by using a multi-layer perceptron (MLP) to encode lighting features, transforming the lighting information of $L_{\\mathrm{src}}$ and $L_{\\mathrm{trg}}$ into a compact, low-dimensional representation $(\\mathrm{d} = 256)$ . We apply AdaIN to encoder features using the source lighting features, producing normalized features $f^l$ (for $l = 1,\\dots ,7$ ). Through this normalization process, we induce de-lighting, effectively removing the lighting information present in the encoder features. Starting from the de-lit latent features $f^7$ , we perform progressive bi-linear upsampling. At each upsampling step, we apply AdaIN to the concatenated feature, incorporating the target lighting features encoded by the lighting encoder. This construction using the LCFN module and source lighting prediction allows us to effectively remove source lighting features from the input and only propagate intrinsic appearance features from the encoder to the decoder. We then add target lighting features in the decoder. The LCFN module also contributes towards temporal consistency (see Sec. 3.3). See Fig. 2 for a comparison between our architecture and those of Sun et al. [36] and Sengupta et al. [30].", + "bbox": [ + 212, + 207, + 787, + 465 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "3.3 Enforcing temporal consistency", + "text_level": 1, + "bbox": [ + 214, + 483, + 521, + 498 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Temporal consistency is vital in making relit videos stable, realistic, and aesthetically pleasing. Previous single-image portrait relighting techniques [30,36] do not incorporate explicit temporal modeling, leading to undesirable flickering artifacts when applied to videos. Accuracy in single-image portrait relighting can often be uncorrelated to temporal flickering. Inconsistencies across frames are even more noticeable when the source lighting $L_{\\mathrm{src}}$ changes continuously.", + "bbox": [ + 212, + 505, + 784, + 595 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "When applied to skip-connected features, LCFN provides a natural defense against temporal flickering by removing source lighting features from the input image. However, it cannot ensure temporal consistency on its own. We notice two further problems: (1) when the source lighting gradually changes, LCFN often leaks small amounts of source lighting information to the decoder, leading to flickering; (2) when source lighting changes abruptly, undesirable fading effects can be observed.", + "bbox": [ + 212, + 595, + 785, + 700 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "To address this issue, we propose a skip-connected feature smoothing technique that assumes neighboring frames share the same intrinsic appearance features, obtained after de-lighting input image features with LCFN. We apply a simple exponential smoothing of de-lit features generated by LCFN, denoted as $f^l$ , using all the previous frames:", + "bbox": [ + 212, + 700, + 785, + 777 + ], + "page_idx": 6 + }, + { + "type": "equation", + "text": "\n$$\nf _ {t} ^ {l} := \\alpha \\cdot f _ {t} ^ {l} + (1 - \\alpha) \\cdot f _ {t - 1} ^ {l} \\quad (\\text {f o r} l = 1, \\dots , 7) \\tag {2}\n$$\n", + "text_format": "latex", + "bbox": [ + 339, + 784, + 784, + 801 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "with $\\alpha = 0.7$ . Note that exponential smoothing does not work without de-lit LCFN features, which removes time-varying source lighting.", + "bbox": [ + 212, + 809, + 784, + 839 + ], + "page_idx": 6 + }, + { + "type": "header", + "text": "Personalized Video Relighting With an At-Home Light Stage", + "bbox": [ + 323, + 114, + 732, + 130 + ], + "page_idx": 6 + }, + { + "type": "page_number", + "text": "7", + "bbox": [ + 774, + 114, + 785, + 126 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "We further notice that when the monitor light changes abruptly the relighting effect is delayed by a few frames, mainly due to the limited refresh rate of the monitor and frame rate of the camera. We thus propose doing a weighted average of source monitor lighting $L_{\\mathrm{src}}$ from a sequence of previous and current frames to achieve smoother and more natural results:", + "bbox": [ + 212, + 146, + 787, + 220 + ], + "page_idx": 7 + }, + { + "type": "equation", + "text": "\n$$\nL _ {\\text {s r c a v g}} ^ {t} = \\frac {\\sum_ {i = 0} ^ {N - 1} \\beta^ {i} L _ {\\text {s r c}} ^ {t - i}}{\\sum_ {i = 0} ^ {N - 1} \\beta^ {i}}, \\tag {3}\n$$\n", + "text_format": "latex", + "bbox": [ + 413, + 224, + 785, + 263 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "where $\\beta = 0.6$ and $N = 3$ .", + "bbox": [ + 215, + 267, + 406, + 280 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "3.4 Training relighting network", + "text_level": 1, + "bbox": [ + 215, + 300, + 488, + 315 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Our model is trained through minimizing a weighted combination of three loss functions: generator loss, discriminator loss, and monitor loss. The first loss aims to minimize the discrepancies between the true target image $I_{\\mathrm{trg}}$ in our dataset and the predicted target relit image $\\hat{I}_{\\mathrm{trg}}$ , leading to accurately relit images. We adopted our generator loss (Eq. 4) and our discriminator loss (Eq. 5) from Sengupta et al. [30]:", + "bbox": [ + 212, + 320, + 782, + 412 + ], + "page_idx": 7 + }, + { + "type": "equation", + "text": "\n$$\n\\mathcal {L} _ {G} ^ {I} = \\lambda_ {L 1} \\mathcal {L} _ {L 1} \\left(I _ {\\mathrm {t r g}}, \\hat {I} _ {\\mathrm {t r g}}\\right) + \\lambda_ {P} \\mathcal {L} _ {P} \\left(I _ {\\mathrm {t r g}}, \\hat {I} _ {\\mathrm {t r g}}\\right) \\tag {4}\n$$\n", + "text_format": "latex", + "bbox": [ + 339, + 419, + 784, + 444 + ], + "page_idx": 7 + }, + { + "type": "equation", + "text": "\n$$\n+ \\lambda_ {C} \\mathcal {L} _ {C} \\left(I _ {\\mathrm {s r c}}, \\hat {I} _ {\\mathrm {s r c}} ^ {C}\\right) + \\lambda_ {D} \\left(D \\left(\\hat {I} _ {\\mathrm {t r g}}; \\theta_ {D}\\right) - 1\\right) ^ {2},\n$$\n", + "text_format": "latex", + "bbox": [ + 364, + 440, + 663, + 457 + ], + "page_idx": 7 + }, + { + "type": "equation", + "text": "\n$$\n\\mathcal {L} _ {D} = \\left(D \\left(I _ {\\mathrm {t r g}}; \\theta_ {D}\\right) - 1\\right) ^ {2} + \\left(D \\left(\\hat {I} _ {\\mathrm {t r g}}; \\theta_ {D}\\right)\\right) ^ {2}, \\tag {5}\n$$\n", + "text_format": "latex", + "bbox": [ + 354, + 468, + 784, + 484 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "where $\\mathcal{L}_{L1}$ denotes L1 loss, $\\mathcal{L}_P$ denotes perceptual loss [45], $\\mathcal{L}_C$ denotes cycle consistency loss [50], and $D$ is the discriminator [15]. $\\hat{I}_{\\mathrm{src}}^C$ and $\\hat{L}_{\\mathrm{trg}}^C$ are the outputs from $G(\\hat{I}_{\\mathrm{trg}}, L_{\\mathrm{trg}}, L_{\\mathrm{src}},; \\theta_G)$", + "bbox": [ + 212, + 488, + 782, + 537 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "The monitor reconstruction loss focuses on minimizing the errors between the predicted source light $\\hat{L}_{\\mathrm{src}}$ and the true source light $L_{\\mathrm{src}}$ and is expected to enforce improved disentanglement of lighting information from intrinsic appearance features.", + "bbox": [ + 212, + 537, + 782, + 595 + ], + "page_idx": 7 + }, + { + "type": "equation", + "text": "\n$$\n\\mathcal {L} _ {G} ^ {M} = \\lambda_ {L 1} \\mathcal {L} _ {L 1} \\left(L _ {\\mathrm {s r c}}, \\hat {L} _ {\\mathrm {s r c}}\\right) + \\lambda_ {P} \\mathcal {L} _ {P} \\left(L _ {\\mathrm {s r c}}, \\hat {L} _ {\\mathrm {s r c}}\\right) + \\lambda_ {C} \\mathcal {L} _ {C} \\left(L _ {\\mathrm {t r g}}, \\hat {L} _ {C} ^ {\\text {t r g}}\\right). \\tag {6}\n$$\n", + "text_format": "latex", + "bbox": [ + 272, + 602, + 784, + 619 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Finally, we minimize the image generator loss $\\mathcal{L}_G^I$ , the discriminator loss $\\mathcal{L}_D$ , and the illumination loss $\\mathcal{L}_G^M$ together:", + "bbox": [ + 212, + 625, + 782, + 655 + ], + "page_idx": 7 + }, + { + "type": "equation", + "text": "\n$$\n\\min _ {G, D} \\mathcal {L} _ {G} ^ {I} + \\mathcal {L} _ {D} + \\lambda_ {G} ^ {M} \\mathcal {L} _ {G} ^ {M}. \\tag {7}\n$$\n", + "text_format": "latex", + "bbox": [ + 416, + 660, + 784, + 681 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "4 Experiments", + "text_level": 1, + "bbox": [ + 215, + 707, + 375, + 723 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "In Sec. 4.1, we first discuss our data collection process, which was based on the approach used to compile the Light Stage at Your Desk (LSYD) dataset. In Sec. 4.2, we perform quantitative and qualitative comparisons with existing single-image portrait relighting algorithms and evaluate their temporal consistency. In Sec. 4.3, we demonstrate predicting, at a low-level, what the user is looking at on the monitor screen. Finally, in Sec. 4.4, we perform ablation studies, evaluating the network architecture's impact on relighting performance.", + "bbox": [ + 212, + 734, + 787, + 840 + ], + "page_idx": 7 + }, + { + "type": "page_number", + "text": "8", + "bbox": [ + 217, + 114, + 228, + 126 + ], + "page_idx": 7 + }, + { + "type": "header", + "text": "J. Choi et al.", + "bbox": [ + 271, + 114, + 359, + 127 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/3d129e60991e7581d7e9075533df2730fdc6e734021c7fe4e525a995406d229e.jpg", + "image_caption": [ + "Fig. 4. We perform a qualitative comparison with existing techniques [30,36] on the LSYD dataset. Source and target (images & lighting) were unseen during training. All models are personalized, i.e. trained on images of that individual only. We (Col. 3) produce significantly better results compared to existing approaches (Cols 4 and 5)." + ], + "image_footnote": [], + "bbox": [ + 218, + 142, + 787, + 503 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "4.1 Data", + "text_level": 1, + "bbox": [ + 215, + 602, + 303, + 616 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "We recorded data from 5 users of diverse ethnicities and genders to ensure a wide range of skin types. Each participant wore a variety of outfits, and we used 4 different ambient lighting conditions per person to mimic the conditions of real-life online meetings. We directed the participants to continuously change their facial expression and pose during the capture sessions. Each user's face was captured while watching 8 different videos, each 8 minutes long, on different days with varying appearances. We randomly hold out 1 video for testing and use the remaining 7 videos for training. We use this testing sequence only for qualitative evaluation, not for any quantitative metrics. This is because quantitative evaluation requires a pair of source and target images of the same person in the same pose but under different lighting conditions. This is difficult to obtain accurately for the aforementioned test video sequence since the participants naturally vary their pose and expression over the course of the video. Instead, we capture an additional test sequence, used only for numerical evaluation in", + "bbox": [ + 212, + 628, + 787, + 840 + ], + "page_idx": 8 + }, + { + "type": "header", + "text": "Personalized Video Relighting With an At-Home Light Stage", + "bbox": [ + 323, + 114, + 732, + 128 + ], + "page_idx": 8 + }, + { + "type": "page_number", + "text": "9", + "bbox": [ + 774, + 116, + 784, + 126 + ], + "page_idx": 8 + }, + { + "type": "image", + "img_path": "images/7ae403f850a9bf0fb1465764a59d13c479f036723f1f9049c48e584246657292.jpg", + "image_caption": [ + "Fig. 5. We perform a qualitative comparison on the OLAT dataset [44]. Our approach outperforms [36,30] and can render strong directional lighting and specular highlights without any explicit modeling of geometry and reflectance." + ], + "image_footnote": [], + "bbox": [ + 217, + 143, + 785, + 417 + ], + "page_idx": 9 + }, + { + "type": "text", + "text": "which the participant is captured in 9 different pose-expression combinations, each with a distinct monitor light. For each pose, we can create $\\binom{9}{2} = 36$ source and target pairs as input and pseudo ground-truth, resulting in a total of 324 test data pairs per user. Additionally, we compared methods using Dynamic OLAT Dataset [44] with environment lighting maps [1,8]. During this comparison, we converted HDR environment maps to LDR and utilized a $270^{\\circ}$ FoV as a monitor light to assess the relighting results. We want to note that all the test data is composed of unseen portraits and lightings for both LSYD and OLAT datasets.", + "bbox": [ + 212, + 502, + 787, + 623 + ], + "page_idx": 9 + }, + { + "type": "text", + "text": "4.2 Comparison with existing approaches", + "text_level": 1, + "bbox": [ + 215, + 647, + 570, + 662 + ], + "page_idx": 9 + }, + { + "type": "text", + "text": "We employ three error metrics to assess relighting performance: RMSE, LPIPS [45], and DISTS [6]. LPIPS and DISTS are more robust to slight differences in pose between the relit image and the pseudo ground truth and detect perceptual differences more effectively than RMSE.", + "bbox": [ + 212, + 672, + 785, + 733 + ], + "page_idx": 9 + }, + { + "type": "text", + "text": "**Portrait image relighting.** We compared our approach with existing portrait relighting neural architectures — Sun et al. [36] and Sengupta et al. [30] — by training on our captured LSYD dataset using the same pre-processing for all three architectures (see Sec. 3.1). Our training loss, given in Sec. 3.4, can handle misalignment in source-target pairs in training data, similar to the loss proposed in Sengupta et al. [30] (we use an additional loss on source monitor lighting prediction). For Sun et al. [36], we train both with their original loss", + "bbox": [ + 212, + 734, + 787, + 840 + ], + "page_idx": 9 + }, + { + "type": "page_number", + "text": "10", + "bbox": [ + 217, + 114, + 235, + 126 + ], + "page_idx": 9 + }, + { + "type": "header", + "text": "J. Choi et al.", + "bbox": [ + 271, + 114, + 359, + 126 + ], + "page_idx": 9 + }, + { + "type": "table", + "img_path": "images/2e792c25fbccdfab1e58e91ab85c71bf777faa4a190227951cee6e0fe344644c.jpg", + "table_caption": [ + "Table 1. We train a personalized relighting model on 5 users from our LSYD dataset and 4 users from the OLAT dataset [44]. We evaluate these models using source & target portrait images, as well as lighting, that were not encountered during the training phase, and report average RMSE, LPIPS [46], and DISTS[6] scores both on LSYD and OLAT dataset. Our method can perform relighting without source lighting $L_{\\mathrm{src}}$ , by simply using the predicted light source from our model as input lighting. Our method significantly outperforms Sun et al. [36] and Sengupta et al. [30]." + ], + "table_footnote": [], + "table_body": "
Known source lightingLSYD dataOLAT data [44]
LsrcLPIPS ↓DISTS ↓RMSE ↓LPIPS ↓DISTS ↓RMSE ↓
Sun et al. [36] w/ LSun-0.17120.16298.59580.22730.17456.2692
Sun et al. [36] w/ LOurs-0.10290.11528.44760.22670.15696.1898
Ours-0.08390.09538.32220.18120.13366.0931
Sengupta et al. [30]✓0.10180.11058.28260.22370.16756.1751
Ours✓0.08320.09538.19390.18090.13345.9548
", + "bbox": [ + 215, + 250, + 787, + 347 + ], + "page_idx": 10 + }, + { + "type": "text", + "text": "function $L_{\\mathrm{Sun}}$ (which expects perfect source-target pose alignment obtained in OLAT data) and with our proposed loss function $L_{\\mathrm{ours}}$ to specifically handle misalignment in LSYD data. We train personalized models on 5 users from the LSYD dataset and on 4 users from the publicly available Dynamic OLAT Dataset [44] with 2361 indoor environment lighting maps [1,8].", + "bbox": [ + 212, + 380, + 787, + 455 + ], + "page_idx": 10 + }, + { + "type": "text", + "text": "For our quantitative evaluation, we test our model on 1620 test images across 5 users with unseen appearance and lighting conditions on the LSYD dataset and on 7172 test images from the Dynamic OLAT dataset. We present the result in Tab. 1. We observe that our proposed approach outperforms Sengupta et al. [30] and Sun et al. [36] by $22.3\\%$ and $23.6\\%$ respectively on the LSYD dataset and by $23.5\\%$ and $25.6\\%$ on the OLAT dataset, when comparing LPIPS score. Our qualitative comparison, as presented in Fig. 4 and Fig. 5, shows that our model performs superior re", + "bbox": [ + 212, + 459, + 447, + 744 + ], + "page_idx": 10 + }, + { + "type": "image", + "img_path": "images/829b6f38fc6832be1caaaef5fd422d37ecbc21eaa7971f639fa05e93b4f24e65.jpg", + "image_caption": [ + "Fig.6. Our method can also relight portrait images captured with unknown light sources. We relight a source image (Col. 1) with target light shown in the inset in Col. 2. We add a reference of how the face appears under that target light." + ], + "image_footnote": [], + "bbox": [ + 460, + 459, + 782, + 654 + ], + "page_idx": 10 + }, + { + "type": "text", + "text": "lighting in terms of color, quality, and consistency, and can render strong directional lighting and specular highlights without any explicit modeling.", + "bbox": [ + 212, + 746, + 787, + 776 + ], + "page_idx": 10 + }, + { + "type": "text", + "text": "Note that Sun et al. [36] does not require the source lighting $L_{\\mathrm{src}}$ during test time. Our proposed approach can also perform relighting without prior knowledge of source lighting $L_{\\mathrm{src}}$ by simply predicting $\\hat{L}_{\\mathrm{src}}$ and using it for light-conditioned feature normalization. We show that even in the absence of", + "bbox": [ + 212, + 779, + 787, + 840 + ], + "page_idx": 10 + }, + { + "type": "header", + "text": "Personalized Video Relighting With an At-Home Light Stage", + "bbox": [ + 323, + 113, + 732, + 128 + ], + "page_idx": 10 + }, + { + "type": "page_number", + "text": "11", + "bbox": [ + 767, + 114, + 782, + 126 + ], + "page_idx": 10 + }, + { + "type": "text", + "text": "$L_{\\mathrm{src}}$ , our method outperforms Sun et al. [36] by $22.6\\%$ on the LSYD data and by $25.5\\%$ on the OLAT data, in terms of LPIPS.", + "bbox": [ + 212, + 145, + 782, + 176 + ], + "page_idx": 11 + }, + { + "type": "text", + "text": "In Fig. 6, we demonstrate that our approach can relight any portrait image captured 'in-the-wild' without requiring the knowledge of source lighting, and outperforms Sun et al. [36]. Our approach first predicts a proxy source monitor lighting, imagining the portrait image to be captured under a monitor lighting, and uses this predicted source lighting in the LCFN module to 'de-lit' the input image and relight it with a target lighting.", + "bbox": [ + 212, + 181, + 784, + 272 + ], + "page_idx": 11 + }, + { + "type": "text", + "text": "In Fig. 7 1st-row, we show the results of relighting an input image with directional lights by moving specific bright areas on the monitor screen. Our method learns to render directional lighting effects and cast shadows as needed. To better illustrate the effect of moving lighting, we first relight the input image with no light reflected from the monitor (row-2 col-1), which produces a relief image under ambient room", + "bbox": [ + 212, + 276, + 444, + 487 + ], + "page_idx": 11 + }, + { + "type": "image", + "img_path": "images/766dae490665a6cdf8c35774e60d851e62c027862619c7e6fc1c5c69f13aab8d.jpg", + "image_caption": [ + "Fig. 7. We relight the input image with directional lights (1st row) and without any target light ('Relight A'). We then subtract the 'Relight A' from directional images (row-2 col-2:5). Our method learns to render the effects of directional lighting and decouple ambient and dominant frontal lighting (from monitor)." + ], + "image_footnote": [], + "bbox": [ + 460, + 273, + 527, + 372 + ], + "page_idx": 11 + }, + { + "type": "image", + "img_path": "images/73735b90d02b709dfef115d3bca1386ca66be993dcfedc53fc48db8e666923a6.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 527, + 273, + 589, + 372 + ], + "page_idx": 11 + }, + { + "type": "image", + "img_path": "images/833ec48b41a5f287af20c0dbca34ef64febc2c348d0c5a2df04626e2c89da384.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 589, + 273, + 653, + 372 + ], + "page_idx": 11 + }, + { + "type": "image", + "img_path": "images/b26c96fe6105b580185f4398c11bfe3564db9ff5b0580854bc0f24c3ccbf2c12.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 653, + 273, + 717, + 372 + ], + "page_idx": 11 + }, + { + "type": "image", + "img_path": "images/42ebe8fb772bfb7547b07bc29344a51be6b13c15d1f84a9046870973fa7ef54e.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 717, + 273, + 781, + 372 + ], + "page_idx": 11 + }, + { + "type": "text", + "text": "lighting only. Subtracting this 'de-lit' image from the relit images under directional lighting (row-2 col 2-5) highlights the ability of our method to decouple ambient room illumination from dominant frontal lighting. However, due to the constraint that our lighting is limited to the illumination emitted from the monitor, we may not accurately depict extreme lighting effects. To this end, we show the strong directional light effects using the OLAT dataset in Fig. 5.", + "bbox": [ + 212, + 488, + 784, + 580 + ], + "page_idx": 11 + }, + { + "type": "text", + "text": "Portrait video relighting. Next, we evaluate the temporal consistency of each portrait video relighting algorithm. For each user in the LSYD data, we relit the held-out test video with 50 different target lighting conditions, creating 50 relit videos. We then computed the RMSE between adjacent frames in relit videos as a measure of temporal consistency. Since the pose is almost identical between adjacent frames, lower RMSE error indicates temporally consistent relighting. We then report the average temporal RMSE across all such adjacent frames. In practice,", + "bbox": [ + 212, + 583, + 464, + 840 + ], + "page_idx": 11 + }, + { + "type": "table", + "img_path": "images/8376becaae33b62fc258e719efff9f718f176a0e660b472453b8490a7e6d87ec.jpg", + "table_caption": [ + "Table 2. We evaluate temporal consistency by relighting a test video captured with varying source lighting with the same target lighting and calculating RMSE between adjacent frames. We then report the average RMSE across all adjacent frames and compute an error rate to indicate the percent of adjacent frames with RMSE higher than a threshold." + ], + "table_footnote": [], + "table_body": "
ThresholdRMSE ↓Error Rate (%)
>0.2>0.3>0.4
Sun et al. [36]5.8613.532.611.06
Sengupta et al. [30]6.3721.835.401.75
+Lsrc_avg5.7613.312.340.98
Ours (base)6.0116.223.611.08
+Lsrc_avg5.7313.092.310.83
+LCFN5.6813.042.280.71
+Lsrc_avg + LCFN5.5512.892.220.65
", + "bbox": [ + 475, + 696, + 785, + 830 + ], + "page_idx": 11 + }, + { + "type": "page_number", + "text": "12", + "bbox": [ + 217, + 114, + 235, + 126 + ], + "page_idx": 11 + }, + { + "type": "header", + "text": "J. Choi et al.", + "bbox": [ + 271, + 114, + 359, + 127 + ], + "page_idx": 11 + }, + { + "type": "text", + "text": "however, a significant fraction of adjacent frame pairs have extremely similar lighting between the two frames, making their relit frames naturally consistent anyway. Only in a small percentage of adjacent frames does the source lighting significantly change, leading to obvious flickering in the relit video if temporal consistency is not maintained. Thus, in addition to average temporal RMSE, we also compute the error rate for three different thresholds: 0.2 (low), 0.3 (medium), and 0.4 (high), which indicate the percentage of adjacent frames where RMSE", + "bbox": [ + 212, + 146, + 787, + 252 + ], + "page_idx": 12 + }, + { + "type": "text", + "text": "error is more than the threshold.", + "bbox": [ + 212, + 252, + 452, + 266 + ], + "page_idx": 12 + }, + { + "type": "text", + "text": "In Tab. 2 and Fig. 8, we compare our approach with and without skip-connected feature smoothing to past works [36,30]. Note that this temporal smoothing of the skip-connected features can only be applied in our framework since we de-light encoder features from the source lighting with LCFN. Both for our approach and for Sengupta et al. [30], we can further apply smoothing of input source lighting to handle abrupt changes. We observe that our method produces the most temporally consistent relighting while also being the most accurate (see Tab. 1). We further note that both skip-connected feature smoothing and smoothing of source lighting improve temporal consistency.", + "bbox": [ + 212, + 266, + 447, + 630 + ], + "page_idx": 12 + }, + { + "type": "image", + "img_path": "images/807de019c307e9311014a483b329950f97e590d5be83fe3c49ea9433a007255e.jpg", + "image_caption": [ + "Fig.8. We show temporal consistency between adjacent frames separated by 0.33s by relighting a test video captured with varying source lighting with the same target lighting. Note that [36,30] both exhibit abrupt changes in lighting between frames $t$ and $t + 10$ , while our approach produces a more stable result." + ], + "image_footnote": [], + "bbox": [ + 459, + 265, + 784, + 518 + ], + "page_idx": 12 + }, + { + "type": "text", + "text": "4.3 Monitor prediction", + "text_level": 1, + "bbox": [ + 215, + 650, + 419, + 666 + ], + "page_idx": 12 + }, + { + "type": "text", + "text": "Our network can also be used to predict the source monitor lighting of any input image, as shown in Fig. 9. For the LSYD dataset, we do have images of the source monitor for every input image, so we can calculate monitor prediction accuracy numerically. We observe that our method slightly out", + "bbox": [ + 212, + 672, + 446, + 823 + ], + "page_idx": 12 + }, + { + "type": "text", + "text": "performs [36] with a Mean Absolute Error of 0.15 vs 0.17, also qualitatively", + "bbox": [ + 212, + 824, + 785, + 840 + ], + "page_idx": 12 + }, + { + "type": "image", + "img_path": "images/c359b5fab8d4a90308e8904d8d440b06fda78369c1aef82172ac4752626fb693.jpg", + "image_caption": [ + "Fig. 9. Monitor prediction comparison on LSYD data (Col. 1 and 2) and \"in the wild\" setting (Col. 3 and 4). The second row represents the monitor prediction by [36] and Ours." + ], + "image_footnote": [], + "bbox": [ + 459, + 671, + 782, + 756 + ], + "page_idx": 12 + }, + { + "type": "header", + "text": "Personalized Video Relighting With an At-Home Light Stage", + "bbox": [ + 323, + 114, + 732, + 128 + ], + "page_idx": 12 + }, + { + "type": "page_number", + "text": "13", + "bbox": [ + 767, + 114, + 784, + 126 + ], + "page_idx": 12 + }, + { + "type": "text", + "text": "producing more meaningful visualizations. For portrait images, we do not have ground-truth source lighting, but visualizations show meaningful predictions.", + "bbox": [ + 212, + 146, + 782, + 176 + ], + "page_idx": 13 + }, + { + "type": "text", + "text": "The ability to predict monitor lighting from images captured by our monitor has many implications. This technique can be used to detect deep fake avatars during live video calls by purposefully projecting specific images via screen sharing and observing if we can detect the same image from the webcam feed of other attendees in the call. If there is a mismatch between the projected monitor image and the predicted one, this likely indicates a deep fake avatar in the video call.", + "bbox": [ + 212, + 176, + 784, + 267 + ], + "page_idx": 13 + }, + { + "type": "text", + "text": "4.4 Ablation studies", + "text_level": 1, + "bbox": [ + 214, + 289, + 397, + 304 + ], + "page_idx": 13 + }, + { + "type": "text", + "text": "We report the removal of various components from our relighting network in Tab. 3, specifically LCFN and source monitor lighting prediction using intermediate encoder features. We observe that both improve final relighting performance, which shows their effectiveness in disentangling source", + "bbox": [ + 212, + 314, + 464, + 450 + ], + "page_idx": 13 + }, + { + "type": "table", + "img_path": "images/c58fece9fc9f1ef686e4767e6a28bbed2cfca27cfa9736ed99e18e8c678ce97a.jpg", + "table_caption": [ + "Table 3. Both LCFN and source monitor prediction $\\mathbf{L}_{\\mathrm{src}}$ improve relighting performance by effectively disentangling source lighting information from intrinsic appearance features." + ], + "table_footnote": [], + "table_body": "
Lsrcde-lightingRMSE ↓LPIPS ↓DISTS ↓
--8.42300.09640.1071
✓-8.25960.09150.1013
-✓8.19070.09040.0966
✓✓8.07460.08530.0963
", + "bbox": [ + 475, + 372, + 785, + 443 + ], + "page_idx": 13 + }, + { + "type": "text", + "text": "lighting information from intrinsic appearance features.", + "bbox": [ + 214, + 450, + 612, + 465 + ], + "page_idx": 13 + }, + { + "type": "text", + "text": "5 Conclusion", + "text_level": 1, + "bbox": [ + 214, + 488, + 357, + 503 + ], + "page_idx": 13 + }, + { + "type": "text", + "text": "We propose a personalized video relighting algorithm that leverages casually captured LSYD data to generate real-time high-quality temporally consistent relit videos under any pose, expression, and lighting conditions. We present a novel network architecture that can perform high-quality relighting on both the LSYD and OLAT datasets, rendering challenging lighting conditions like directional lights, shadows, and specularities. We achieve this without using any 3D information or performing explicit decomposition, simply by achieving better image-based relighting enabled by our proposed neural architecture with LCFN module. Our method enables better lighting quality during live video calls and in portrait images, and produces better harmonization with virtual backgrounds.", + "bbox": [ + 212, + 520, + 784, + 671 + ], + "page_idx": 13 + }, + { + "type": "text", + "text": "Limitation. Since we utilize only the front-facing monitor light as the source lighting, we do not account for scenarios where the light source is located on the sides $(90^{\\circ})$ or behind $(180^{\\circ})$ . However, it does not necessarily mean that our model cannot learn directional lighting effects, as demonstrated by training on the OLAT dataset which contains full $360^{\\circ}$ lighting, see Fig. 5.", + "bbox": [ + 212, + 671, + 784, + 746 + ], + "page_idx": 13 + }, + { + "type": "text", + "text": "Ethical considerations. While our primary goal is to allow people to improve their facial appearance with virtual relighting, we note that it is also a form of image manipulation and can be used for malicious purposes [3]. Furthermore, we emphasize that while currently we can only predict the user's direct line-of-sight monitor lighting in low resolution, the potential for high-resolution monitor prediction in the future could raise significant privacy concerns.", + "bbox": [ + 212, + 746, + 785, + 838 + ], + "page_idx": 13 + }, + { + "type": "page_number", + "text": "14", + "bbox": [ + 217, + 114, + 235, + 126 + ], + "page_idx": 13 + }, + { + "type": "header", + "text": "J. Choi et al.", + "bbox": [ + 271, + 114, + 359, + 126 + ], + "page_idx": 13 + }, + { + "type": "text", + "text": "References", + "text_level": 1, + "bbox": [ + 217, + 143, + 321, + 159 + ], + "page_idx": 14 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "1. Bolduc, C., Giroux, J., Hébert, M., Demers, C., Lalonde, J.F.: Beyond the pixel: a photometrically calibrated hdr dataset for luminance and color prediction (2023)", + "2. Chen, Z., Liu, Z.: Relighting4d: Neural relightable human from videos (2022)", + "3. Choi, J.M., Leung, J., Frahm, N., Christman, M., Bertasius, G., Sengupta, R.: Building secure and engaging video communication by using monitor illumination. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 4377-4386 (2024)", + "4. Daichi Tajima, Yoshihiro Kanamori, Y.E.: Relighting humans in the wild: Monocular full-body human relighting with domain adaptation. Computer Graphics Forum (Proc. of Pacific Graphics 2021) 40(7), 205-216 (2021)", + "5. Debevec, P., Wenger, A., Tchou, C., Gardner, A., Waese, J., Hawkins, T.: A lighting reproduction approach to live-action compositing. ACM Trans. Graph. 21(3), 547-556 (jul 2002). https://doi.org/10.1145/566654.566614, https://doi.org/10.1145/566654.566614", + "6. Ding, K., Ma, K., Wang, S., Simoncelli, E.P.: Image quality assessment: Unifying structure and texture similarity. CoRR abs/2004.07728 (2020), https://arxiv.org/abs/2004.07728", + "7. Futschik, D., Ritland, K., Vecore, J., Fanello, S., Orts-Escolano, S., Curless, B., Sykora, D., Pandey, R.: Controllable light diffusion for portraits (2023)", + "8. Gardner, M.A., Sunkavalli, K., Yumer, E., Shen, X., Gambaretto, E., Gagné, C., Lalonde, J.F.: Learning to predict indoor illumination from a single image (2017)", + "9. Gerstner, C.R., Farid, H.: Detecting real-time deep-fake videos using active illumination. In: 2022 IEEE/CVF Conference on Computer Vision and Pattern Recognition Workshops (CVPRW). pp. 53-60 (2022). https://doi.org/10.1109/CVPRW56347.2022.00015", + "0. Guo, K., Lincoln, P., Davidson, P., Busch, J., Yu, X., Whalen, M., Harvey, G., Orts-Escalano, S., Pandey, R., Dourgarian, J., Tang, D., Tkach, A., Kowdle, A., Cooper, E., Dou, M., Fanello, S., Fyffe, G., Rhemann, C., Taylor, J., Debevec, P., Izadi, S.: The relightables: Volumetric performance capture of humans with realistic relighting (2019)", + "1. He, Y., Xing, Y., Zhang, T., Chen, Q.: Unsupervised portrait shadow removal via generative priors. In: Proceedings of the 29th ACM International Conference on Multimedia. MM '21, ACM (Oct 2021). https://doi.org/10.1145/3474085.3475663, http://dx.doi.org/10.1145/3474085.3475663", + "2. Hou, A., Sarkis, M., Bi, N., Tong, Y., Liu, X.: Face relighting with geometrically consistent shadows. In: In Proceeding of IEEE Computer Vision and Pattern Recognition. New Orleans, LA (June 2022)", + "3. Hou, A., Zhang, Z., Sarkis, M., Bi, N., Tong, Y., Liu, X.: Towards high fidelity face relighting with realistic shadows. In: IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR) (2021)", + "4. Hu, Y., Wang, B., Lin, S.: Fc 4: Fully convolutional color constancy with confidence-weighted pooling. In: Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition. pp. 4085-4094 (2017)", + "5. Isola, P., Zhu, J.Y., Zhou, T., Efros, A.A.: Image-to-image translation with conditional adversarial networks. CVPR (2017)", + "6. Jin, H., Liao, S., Shao, L.: Pixel-in-pixel net: Towards efficient facial landmark detection in the wild. International Journal of Computer Vision (Sep 2021). https://doi.org/10.1007/s11263-021-01521-4, http://dx.doi.org/10.1007/s11263-021-01521-4" + ], + "bbox": [ + 225, + 172, + 784, + 838 + ], + "page_idx": 14 + }, + { + "type": "header", + "text": "Personalized Video Relighting With an At-Home Light Stage", + "bbox": [ + 323, + 114, + 730, + 128 + ], + "page_idx": 14 + }, + { + "type": "page_number", + "text": "15", + "bbox": [ + 767, + 116, + 784, + 126 + ], + "page_idx": 14 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "17. Karras, T., Laine, S., Aittala, M., Hellsten, J., Lehtinen, J., Aila, T.: Analyzing and improving the image quality of StyleGAN. In: Proc. CVPR (2020)", + "18. Lagunas, M., Sun, X., Yang, J., Villegas, R., Zhang, J., Shu, Z., Masia, B., Gutierrez, D.: Single-image full-body human relighting. CoRR abs/2107.07259 (2021), https://arxiv.org/abs/2107.07259", + "19. LeGendre, C., Ma, W.C., Pandey, R., Fanello, S., Rhemann, C., Dourgarian, J., Busch, J., Debevec, P.: Learning illumination from diverse portraits (2020)", + "20. Liu*, Y., Hou*, A., Huang, X., Ren, L., Liu, X.: Blind removal of facial foreign shadows. In: In Proceedings of British Machine Vision Conference (BMVC). London, UK (November 2022)", + "21. Meka, A., Pandey, R., Haene, C., Orts-Escolano, S., Barnum, P., Davidson, P., Erickson, D., Zhang, Y., Taylor, J., Bouaziz, S., Legendre, C., Ma, W.C., Overbeck, R., Beeler, T., Debevec, P., Izadi, S., Theobalt, C., Rhemann, C., Fanello, S.: Deep relightable textures - volumetric performance capture with neural rendering. vol. 39 (December 2020). https://doi.org/10.1145/3414685.3417814, http://gvv.mpi-inf.mpg.de/projects/DeepRelightableTextures/", + "22. Nestmeyer, T., Lalonde, J., Matthews, I., Lehrmann, A.: Learning physics-guided face relighting under directional light. In: 2020 IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR). pp. 5123-5132. IEEE Computer Society, Los Alamitos, CA, USA (jun 2020). https://doi.org/10.1109/CVPR42600.2020.00517, https://doi.ieeecomputersociety.org/10.1109/CVPR42600.2020.00517", + "23. Pandey, R., Orts-Escolano, S., LeGendre, C., Haene, C., Bouaziz, S., Rhemann, C., Debevec, P., Fanello, S.: Total relighting: Learning to relight portraits for background replacement. vol. 40 (August 2021). https://doi.org/10.1145/3450626.3459872", + "24. Peers, P., Tamura, N., Matusik, W., Debevec, P.: Post-production facial performance relighting using reflectance transfer. ACM Trans. Graph. 26(3), 52-es (jul 2007). https://doi.org/10.1145/1276377.1276442, https://doi.org/10.1145/1276377.1276442", + "25. Ponglertnapakorn, P., Tritrong, N., Suwajanakorn, S.: Difareli: Diffusion face relighting (2023), https://arxiv.org/abs/2304.09479", + "26. Qi, L., Wu, J., Wang, A.N., Wang, S., Sengupta, R.: My3dgen: A scalable personalized 3d generative model (2023)", + "27. Qiu, H., Chen, Z., Jiang, Y., Zhou, H., Fan, X., Yang, L., Wu, W., Liu, Z.: Relitalk: Relightable talking portrait generation from a single video (2023)", + "28. Ren, M., Xiong, W., Yoon, J.S., Shu, Z., Zhang, J., Jung, H., Gerig, G., Zhang, H.: Relightful harmonization: Lighting-aware portrait background replacement (2023)", + "29. Ronneberger, O., Fischer, P., Brox, T.: U-net: Convolutional networks for biomedical image segmentation. CoRR abs/1505.04597 (2015), http://arxiv.org/abs/1505.04597", + "30. Sengupta, S., Curless, B., Kemelmacher-Shlizerman, I., Seitz, S.M.: A light stage on every desk. CoRR abs/2105.08051 (2021), https://arxiv.org/abs/2105.08051", + "31. Sengupta, S., Kanazawa, A., Castillo, C.D., Jacobs, D.W.: Sfsnet: Learning shape, reflectance and illuminance of faces in the wild'. In: Proceedings of the IEEE conference on computer vision and pattern recognition. pp. 6296-6305 (2018)", + "32. Sevastopolsky, A., Ignatiev, S., Ferrer, G., Burnaev, E., Lempitsky, V.: Relightable 3d head portraits from a smartphone video (2020)", + "33. Shih, Y., Paris, S., Barnes, C., Freeman, W.T., Durand, F.: Style transfer for headshot portraits. ACM Trans. Graph. 33(4) (jul 2014). https://doi.org/10.1145/2601097.2601137, https://doi.org/10.1145/2601097.2601137" + ], + "bbox": [ + 215, + 147, + 784, + 839 + ], + "page_idx": 15 + }, + { + "type": "page_number", + "text": "16", + "bbox": [ + 217, + 114, + 235, + 126 + ], + "page_idx": 15 + }, + { + "type": "header", + "text": "J. Choi et al.", + "bbox": [ + 271, + 114, + 359, + 126 + ], + "page_idx": 15 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "34. Shu, Z., Hadap, S., Shechtman, E., Sunkavalli, K., Paris, S., Samaras, D.: Portrait lighting transfer using a mass transport approach. ACM Trans. Graph. 37(1) (oct 2017). https://doi.org/10.1145/3095816, https://doi.org/10.1145/3095816", + "35. Song, G., Cham, T.J., Cai, J., Zheng, J.: Half-body portrait relighting with overcomplete lighting representation (06 2021)", + "36. Sun, T., Barron, J.T., Tsai, Y., Xu, Z., Yu, X., Fyffe, G., Rhemann, C., Busch, J., Debevec, P.E., Ramamoorthi, R.: Single image portrait relighting. CoRR abs/1905.00824 (2019), http://arxiv.org/abs/1905.00824", + "37. Sun, T., Lin, K.E., Bi, S., Xu, Z., Ramamoorthi, R.: Nelf: Neural light-transport field for portrait view synthesis and relighting (2021)", + "38. Sun, T., Xu, Z., Zhang, X., Fanello, S.R., Rhemann, C., Debevec, P.E., Tsai, Y., Barron, J.T., Ramamoorthi, R.: Light stage super-resolution: Continuous high-frequency relighting. CoRR abs/2010.08888 (2020), https://arxiv.org/abs/2010.08888", + "39. Wang, Y., Holynski, A., Zhang, X., Zhang, X.: Sunstage: Portrait reconstruction and relighting using the sun as a light stage. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 20792-20802 (2023)", + "40. Wang, Z., Yu, X., Lu, M., Wang, Q., Qian, C., Xu, F.: Single image portrait relighting via explicit multiple reflectance channel modeling. ACM Trans. Graph. 39(6) (Nov 2020). https://doi.org/10.1145/3414685.3417824", + "41. Yang, H., Zheng, M., Feng, W., Huang, H., Lai, Y.K., Wan, P., Wang, Z., Ma, C.: Towards practical capture of high-fidelity relightable avatars (2023)", + "42. Yeh, Y.Y., Nagano, K., Khamis, S., Kautz, J., Liu, M.Y., Wang, T.C.: 41(6), 1-21 (nov 2022). https://doi.org/10.1145/3550454.3555442, https://doi.org/10. 1145", + "43. Yeh, Y.Y., Nagano, K., Khamis, S., Kautz, J., Liu, M.Y., Wang, T.C.: Learning to relight portrait images via a virtual light stage and synthetic-to-real adaptation. ACM Transactions on Graphics (TOG) (2022)", + "44. Zhang, L., Zhang, Q., Wu, M., Yu, J., Xu, L.: Neural video portrait relighting in real-time via consistency modeling (2021)", + "45. Zhang, R., Isola, P., Efros, A.A., Shechtman, E., Wang, O.: The unreasonable effectiveness of deep features as a perceptual metric. In: 2018 IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR). pp. 586-595. IEEE Computer Society, Los Alamitos, CA, USA (jun 2018). https://doi.org/10.1109/CVPR.2018.00068, https://doi.ieeecomputersociety.org/10.1109/CVPR.2018.00068", + "46. Zhang, R., Isola, P., Efros, A.A., Shechtman, E., Wang, O.: The unreasonable effectiveness of deep features as a perceptual metric. CoRR abs/1801.03924 (2018), http://arxiv.org/abs/1801.03924", + "47. Zhang, X., Fanello, S.R., Tsai, Y., Sun, T., Xue, T., Pandey, R., Orts-Escalano, S., Davidson, P.L., Rhemann, C., Debevec, P.E., Barron, J.T., Ramamoorthi, R., Freeman, W.T.: Neural light transport for relighting and view synthesis. CoRR abs/2008.03806 (2020), https://arxiv.org/abs/2008.03806", + "48. Zhang, X., Barron, J.T., Tsai, Y.T., Pandey, R., Zhang, X., Ng, R., Jacobs, D.E.: Portrait shadow manipulation. vol. 39 (2020)", + "49. Zhou, H., Hadap, S., Sunkavalli, K., Jacobs, D.: Deep single-image portrait relighting. In: 2019 IEEE/CVF International Conference on Computer Vision (ICCV). pp. 7193-7201 (2019). https://doi.org/10.1109/ICCV.2019.00729", + "50. Zhu, J.Y., Park, T., Isola, P., Efros, A.A.: Unpaired image-to-image translation using cycle-consistent adversarial networks. In: Computer Vision (ICCV), 2017 IEEE International Conference on (2017)" + ], + "bbox": [ + 212, + 146, + 787, + 834 + ], + "page_idx": 16 + }, + { + "type": "header", + "text": "Personalized Video Relighting With an At-Home Light Stage", + "bbox": [ + 323, + 114, + 730, + 128 + ], + "page_idx": 16 + }, + { + "type": "page_number", + "text": "17", + "bbox": [ + 767, + 116, + 784, + 126 + ], + "page_idx": 16 + } +] \ No newline at end of file diff --git a/2024/Personalized Video Relighting With an At-Home Light Stage/0751f00d-698a-4afe-b1e5-471a9991f11a_model.json b/2024/Personalized Video Relighting With an At-Home Light Stage/0751f00d-698a-4afe-b1e5-471a9991f11a_model.json new file mode 100644 index 0000000000000000000000000000000000000000..556bdd40a281ae9e3ccb45610439b5ed28367002 --- /dev/null +++ b/2024/Personalized Video Relighting With an At-Home Light Stage/0751f00d-698a-4afe-b1e5-471a9991f11a_model.json @@ -0,0 +1,2357 @@ +[ + [ + { + "type": "title", + "bbox": [ + 0.218, + 0.14, + 0.787, + 0.187 + ], + "angle": 0, + "content": "Personalized Video Relighting With an At-Home Light Stage" + }, + { + "type": "text", + "bbox": [ + 0.284, + 0.213, + 0.717, + 0.228 + ], + "angle": 0, + "content": "Jun Myeong Choi, Max Christman, and Roni Sengupta" + }, + { + "type": "text", + "bbox": [ + 0.298, + 0.24, + 0.704, + 0.269 + ], + "angle": 0, + "content": "University of North Carolina at Chapel Hill, NC 27514, USA {chedgekr,mrlc,ronisen}@cs.unc.edu" + }, + { + "type": "text", + "bbox": [ + 0.263, + 0.308, + 0.74, + 0.571 + ], + "angle": 0, + "content": "Abstract. In this paper, we develop a personalized video relighting algorithm that produces high-quality and temporally consistent relit videos under any pose, expression, and lighting condition in real-time. Existing relighting algorithms typically rely either on publicly available synthetic data, which yields poor relighting results, or on actual light stage data which is difficult to acquire. We show that by just capturing recordings of a user watching YouTube videos on a monitor we can train a personalized algorithm capable of performing high-quality relighting under any condition. Our key contribution is a novel image-based neural relighting architecture that effectively separates the intrinsic appearance features - the geometry and reflectance of the face - from the source lighting and then combines them with the target lighting to generate a relit image. This neural architecture enables smoothing of intrinsic appearance features leading to temporally stable video relighting. Both qualitative and quantitative evaluations show that our architecture improves portrait image relighting quality and temporal consistency over state-of-the-art approaches on both casually captured 'Light Stage at Your Desk' (LSYD) and light-stage-captured 'One Light At a Time' (OLAT) datasets. Source code is available at https://github.com/chedgekorea/relighting" + }, + { + "type": "text", + "bbox": [ + 0.263, + 0.585, + 0.672, + 0.599 + ], + "angle": 0, + "content": "Keywords: Image-based Relighting \\(\\cdot\\) 'At Home' Light Stage" + }, + { + "type": "title", + "bbox": [ + 0.218, + 0.626, + 0.377, + 0.642 + ], + "angle": 0, + "content": "1 Introduction" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.659, + 0.787, + 0.749 + ], + "angle": 0, + "content": "With the recent rise in popularity of video conferencing for business, educational, and personal activities, there is a significant demand for improving facial lighting. Virtually relighting our images and videos helps us improve the appearance of our faces without requiring explicit studio-quality lighting in a dedicated space or any specialized lighting expertise. Recent advances in deep neural networks have renewed interest in the problem of virtual relighting." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.75, + 0.787, + 0.841 + ], + "angle": 0, + "content": "Training a deep neural network for relighting requires extensive training data that includes source images paired with relit target images. One way of acquiring this data is by using a large spherical rig with numerous lights and cameras, known as a light stage [5]. While light stage data has been shown to produce high-quality relighting results [36,44,23,38,22,48], the limited availability of datasets, trained models, and access to the light stage itself has impeded further research." + } + ], + [ + { + "type": "page_number", + "bbox": [ + 0.218, + 0.116, + 0.23, + 0.127 + ], + "angle": 0, + "content": "2" + }, + { + "type": "header", + "bbox": [ + 0.272, + 0.115, + 0.361, + 0.127 + ], + "angle": 0, + "content": "J. Choi et al." + }, + { + "type": "image", + "bbox": [ + 0.218, + 0.144, + 0.315, + 0.219 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.245, + 0.22, + 0.287, + 0.23 + ], + "angle": 0, + "content": "(a) LSYD" + }, + { + "type": "image", + "bbox": [ + 0.317, + 0.144, + 0.425, + 0.219 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.338, + 0.22, + 0.402, + 0.23 + ], + "angle": 0, + "content": "(b) Input Image" + }, + { + "type": "image", + "bbox": [ + 0.431, + 0.144, + 0.569, + 0.218 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.416, + 0.22, + 0.468, + 0.23 + ], + "angle": 0, + "content": "(c) Relit-Net" + }, + { + "type": "image", + "bbox": [ + 0.571, + 0.144, + 0.676, + 0.218 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.601, + 0.22, + 0.645, + 0.23 + ], + "angle": 0, + "content": "(d) Output" + }, + { + "type": "image", + "bbox": [ + 0.678, + 0.145, + 0.785, + 0.218 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.214, + 0.245, + 0.788, + 0.33 + ], + "angle": 0, + "content": "Fig. 1. We learn a personalized relighting algorithm that generates temporally consistent and high-quality portrait videos under different lighting. We train the network using recordings of users watching YouTube on a monitor, thereby creating a Light Stage at Your Desk (LSYD). We project a portion of the LDR environment map with a \\(180^{\\circ}\\) FoV as the monitor light, while a portion of the remaining \\(90^{\\circ}\\) FoV is mapped as the background. We can achieve a harmonization effect with the virtual background." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.357, + 0.788, + 0.432 + ], + "angle": 0, + "content": "For example, One Light At a Time (Dynamic-OLAT) [44] is the only publicly available light stage relighting dataset consisting of four individuals only. As a result, researchers have often turned to synthetic data to train their relighting algorithms [31,49,13,35]. Unfortunately, existing synthetic data compromises the quality of relit images." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.433, + 0.788, + 0.599 + ], + "angle": 0, + "content": "We draw inspiration from recent work by Sengupta et al. [30,26] and develop a personalized relighting model by capturing a single user's appearance while being lit by a computer monitor. However, the results and applications of Sengupta et al. [30] have several limitations. First, it requires capturing users with fixed poses and expressions, an unrealistic requirement for actual users that impedes casual capture. Second, the relighting algorithm requires a dark room with negligible ambient light, limiting the environment in which this can be applied. Third, the resulting relit video is temporally unstable, exhibiting significant flickering artifacts, making it unsuitable for relighting Zoom calls. Lastly, it requires knowledge of source lighting and thus is unable to relight arbitrary portrait images captured in the wild." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.599, + 0.789, + 0.842 + ], + "angle": 0, + "content": "In this paper, we show that casually captured light stage data is sufficient to develop a high-quality temporally consistent video portrait relighting algorithm that works under arbitrary conditions (i.e. pose, expression, and ambient lighting) in real-time (\\(\\sim 45\\) fps). To that end, we create our own casually captured light stage dataset with varying pose, expression, and lighting, called Light Stage at Your Desk (LSYD). Our key contribution is a neural image-based relighting architecture, based on the commonly used U-Net [36,49,22,40,38], that better disentangles the source lighting from the user's intrinsic facial appearance (shape and reflectance) and then adds back the target lighting to generate a relit image. Existing image-based relighting architectures [36,30] fail to accurately separate source lighting information from intrinsic appearance features in the encoder, leading to inconsistent and temporally unstable video relighting. To this end, we introduce the light-conditioned feature normalization (LCFN) module, which performs relighting and also predicts the source lighting from an input image. The LCFN module also enables temporal stability by performing exponential smoothing of de-lit intrinsic appearance features and facilitates relighting" + } + ], + [ + { + "type": "header", + "bbox": [ + 0.325, + 0.115, + 0.733, + 0.13 + ], + "angle": 0, + "content": "Personalized Video Relighting With an At-Home Light Stage" + }, + { + "type": "page_number", + "bbox": [ + 0.775, + 0.117, + 0.785, + 0.127 + ], + "angle": 0, + "content": "3" + }, + { + "type": "image", + "bbox": [ + 0.216, + 0.142, + 0.378, + 0.252 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.389, + 0.142, + 0.581, + 0.253 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.593, + 0.142, + 0.789, + 0.252 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.214, + 0.269, + 0.788, + 0.368 + ], + "angle": 0, + "content": "Fig. 2. We highlight the key structural differences between our relighting architecture and that of [36,30]. Our approach removes source lighting information from input image features and only propagates intrinsic appearance (geometry and reflectance) features from the encoder to decoder, which results in better relighting quality and more temporal stability. In contrast [36,30] propagates entire image features from the encoder to the decoder without 'de-light', and expects the decoder to remove source lighting and add target lighting information." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.397, + 0.789, + 0.443 + ], + "angle": 0, + "content": "of any arbitrary portrait image with unknown source lighting. We also improve the data pre-processing pipeline from Sengupta et al. [30] to make the relighting algorithm more robust to pose, expression, and ambient lighting conditions." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.444, + 0.789, + 0.67 + ], + "angle": 0, + "content": "We compare our relighting network with two other algorithms: Sun et al. [36], which was originally trained on light stage data (OLAT), and Sengupta et al. [30], which was originally trained on casually captured data (albeit with fixed pose, expression, and no ambient lighting). For a fair comparison, we train all algorithms for personalized relighting using the same data pre-processing steps and loss functions on 5 individuals from our LSYD dataset and 4 individuals from OLAT [44]. Our network outperforms Sun et al. [36] and Sengupta et al. [30] by \\(22.3\\%\\) and \\(23.6\\%\\) respectively on the LSYD dataset and by \\(23.5\\%\\) and \\(25.6\\%\\) on the OLAT dataset, in terms of LPIPS. Qualitatively our method produces superior relighting in terms of color and quality. We further show that our approach is more temporally consistent, leading to less flickering than Sun et al. [36] or Sengupta et al. [30]. Detailed ablation studies show that LCFN and source monitor prediction improves relighting quality, feature, and source monitor smoothing improves temporal consistency, and data pre-processing improves robustness to pose and expression." + }, + { + "type": "text", + "bbox": [ + 0.216, + 0.671, + 0.6, + 0.685 + ], + "angle": 0, + "content": "To summarize, our main contributions are as follows:" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.687, + 0.789, + 0.734 + ], + "angle": 0, + "content": "- We show that casually captured Light Stage at Your Desk (LSYD) data can be used to build a high-quality temporally consistent personalized video relighting algorithm without requiring access to an expensive light stage setup." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.734, + 0.789, + 0.779 + ], + "angle": 0, + "content": "- We introduce a novel video relighting architecture that separates the source lighting from the user's intrinsic appearance features and then adds back the target lighting, leading to improved relighting and temporal consistency for videos." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.78, + 0.789, + 0.84 + ], + "angle": 0, + "content": "- While our relighting network focuses on 'at home' Light Stage (LSYD dataset) outperforming state-of-the-art algorithms, we also perform equally well on actual Light Stage captured OLAT [44] datasets, and any arbitrary portrait image captured 'in the wild'." + }, + { + "type": "list", + "bbox": [ + 0.214, + 0.687, + 0.789, + 0.84 + ], + "angle": 0, + "content": null + } + ], + [ + { + "type": "page_number", + "bbox": [ + 0.218, + 0.116, + 0.232, + 0.127 + ], + "angle": 0, + "content": "4" + }, + { + "type": "header", + "bbox": [ + 0.272, + 0.115, + 0.361, + 0.128 + ], + "angle": 0, + "content": "J. Choi et al." + }, + { + "type": "title", + "bbox": [ + 0.217, + 0.145, + 0.388, + 0.161 + ], + "angle": 0, + "content": "2 Related Work" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.176, + 0.787, + 0.266 + ], + "angle": 0, + "content": "Portrait relighting methods change the appearance of the face to match a target lighting condition. This can be expressed through lighting parameters (e.g., an environment map, spherical harmonics, directional lighting, etc.) or through a reference image of another person. Our approach relights a portrait image to a lighting condition expressed through a low dynamic range (LDR) image representing the image on the monitor." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.267, + 0.789, + 0.569 + ], + "angle": 0, + "content": "Image based relighting. Before the rise of deep learning, attempts were made at non deep learning image based relighting [34,24,33]. Shu et al. [34] introduced a face relighting approach that uses a mass-transport formulation for the transfer of illumination between images. Peers et al. [24] demonstrated a method for relighting portrait images with flat lighting to match specific target environments, incorporating a reference subject database for approximation. Shih et al. [33] adopted a multiscale technique to transfer local image statistics from reference portraits onto new ones, facilitating the matching of attributes like local contrast and overall lighting direction. Recent advancements in deep learning have caused significant shifts to the landscape of portrait relighting [25,11,20]. Research on ratio images for relighting [49,13] was conducted, utilizing public datasets and employing methods based on ratio images. However, this is limited to synthetic data, resulting in a significant domain gap with real data. The widespread application of light stages [5] in gathering data has enabled numerous groundbreaking research endeavors[36,44,38,28]. However, these methods rely on capturing data with a light stage, which are expensive and inaccessible. Instead, our approach builds a personalized relighting algorithm using casually captured videos from the desk recording setup introduced in Sengupta et al. [30]. In contrast to Sengupta et al. [30], we can collect data during daily computer usage by minimizing numerous constraints, eliminating the need for specific efforts in data collection." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.569, + 0.79, + 0.841 + ], + "angle": 0, + "content": "Relighting with explicit decomposition. Creating a virtual relighting dataset through the utilization of synthetic human models to train their networks has been carried out in some studies [31,18,42]. However, when using synthetic data to train a neural network, the large domain gap between synthetic and real data impacts the model's performance on real data. In contrast, others [12,4,27,2,37] generate relit images by using public datasets and employ methods based on 3D model rendering. Specifically, Hou et al. [12] takes a more advanced approach by introducing explicit components, where rays originating from the face intersect with other parts of the facial geometry to create relit images. Moreover, the extensive use of light stages [5] has enabled numerous innovative studies [22,40,23,10,43,7,21,47,38,41,19] in this domain. Some researchers have incorporated explicit elements such as albedo, normals, specular maps, and diffuse maps into their methodologies [22,40,23,7]. Others have taken a physics-based rendering approach [10,43] to resolve these issues. Other papers [21,47] aim to manipulate lighting conditions and generate images under different lighting scenarios using texture information. However, explicit decomposition methods require ground truth (GT) data for these intrinsic components to train, either from synthetic data or real light stage data. Obtaining GT data for 'at-home'" + } + ], + [ + { + "type": "header", + "bbox": [ + 0.325, + 0.115, + 0.733, + 0.13 + ], + "angle": 0, + "content": "Personalized Video Relighting With an At-Home Light Stage" + }, + { + "type": "page_number", + "bbox": [ + 0.775, + 0.117, + 0.785, + 0.127 + ], + "angle": 0, + "content": "5" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.147, + 0.784, + 0.253 + ], + "angle": 0, + "content": "captures is impossible, thus we focus on image-based relighting. In contrast, recent studies aim to streamline the capture process, using a mobile phone camera [32] or a sun stage [39] instead of a light stage. Nevertheless, due to their reliance on per-scene optimization, both of these papers lack the capability for real-time relighting and to generalize to unseen appearances of the individual and are only limited to the particular capture. Instead, our approach performs image-based relighting enabling real-time temporally consistent video relighting, at \\(\\sim 45\\) fps." + }, + { + "type": "title", + "bbox": [ + 0.216, + 0.275, + 0.331, + 0.29 + ], + "angle": 0, + "content": "3 Method" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.305, + 0.785, + 0.426 + ], + "angle": 0, + "content": "Our setup is similar to Gerstner et al. [9] and Sengupta et al. [30], where a user's face is captured while illuminated by their monitor. By capturing multiple videos of the user's face along with the video on their monitor, we build our 'at home' light stage dataset. We then use these data to train a personalized portrait relighting algorithm that can render the user's face under arbitrary lighting conditions. Specifically, given a portrait image \\( I_{\\mathrm{src}} \\), corresponding source monitor lighting \\( L_{\\mathrm{src}} \\), and target monitor lighting \\( L_{\\mathrm{trg}} \\), our aim is to learn a function \\( G \\) that relights \\( I_{\\mathrm{src}} \\) under \\( L_{\\mathrm{trg}} \\):" + }, + { + "type": "equation", + "bbox": [ + 0.386, + 0.437, + 0.785, + 0.455 + ], + "angle": 0, + "content": "\\[\n\\hat {I} _ {\\mathrm {t r g}}, \\hat {L} _ {\\mathrm {s r c}} = G \\left(I _ {\\mathrm {s r c}}, L _ {\\mathrm {s r c}}, L _ {\\mathrm {t r g}}; \\theta_ {G}\\right). \\tag {1}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.464, + 0.784, + 0.509 + ], + "angle": 0, + "content": "Note that our formulation can be used for scenarios where the source lighting is unknown by simply replacing the input source lighting with the predicted source lighting, unlike previous approaches [30]." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.51, + 0.785, + 0.644 + ], + "angle": 0, + "content": "In the following sections, we outline our methodology for portrait video relighting using a monitor as a light stage. Sec. 3.1 outlines strategies for constructing training data pairs from casually captured videos that allow flexibility in facial expression, pose, and ambient lighting. In Sec. 3.2, we introduce our relighting network architecture that disentangles lighting from intrinsic appearance using light-conditioned feature normalization, leading to high-quality relit images. In Sec. 3.3, we propose additional techniques which enforce temporal consistency and eliminate flickering, also using LCFN. Finally, in Sec. 3.4, we discuss how to train our relighting network." + }, + { + "type": "title", + "bbox": [ + 0.216, + 0.666, + 0.53, + 0.682 + ], + "angle": 0, + "content": "3.1 Constructing training data pairs" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.689, + 0.787, + 0.841 + ], + "angle": 0, + "content": "While past work [30] imposed requirements of a neutral pose, expression, and dark room, we loosen these constraints to allow subjects in any conditions. The only constraint we maintain is that the room lighting shall not overpower the light emitted from the monitor. For example, if the capture occurs in front of a window with bright sunlight, the light from the monitor will have minimal impact on the face. As in Sengupta et al. [30], we aim to generate source and monitor image pairs \\((I_{\\mathrm{src}}, L_{\\mathrm{src}})\\), as well as target image and target monitor pairs \\((I_{\\mathrm{trg}}, L_{\\mathrm{trg}})\\), such that we can train our network to produce \\(\\tilde{I}_{\\mathrm{trg}}\\) where \\(I_{\\mathrm{trg}}\\) is the pseudo ground truth. However, due to unrestricted subject movement during data collection, there is a lack of pixel-aligned data, making random pairs" + } + ], + [ + { + "type": "page_number", + "bbox": [ + 0.218, + 0.116, + 0.23, + 0.127 + ], + "angle": 0, + "content": "6" + }, + { + "type": "header", + "bbox": [ + 0.272, + 0.115, + 0.361, + 0.127 + ], + "angle": 0, + "content": "J. Choi et al." + }, + { + "type": "image", + "bbox": [ + 0.218, + 0.144, + 0.789, + 0.325 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.214, + 0.336, + 0.788, + 0.42 + ], + "angle": 0, + "content": "Fig. 3. We first de-light the input image features extracted by the U-Net encoder using Adaptive Instance Normalization (AdaIN) guided by the lighting features extracted from the source lighting with a Light Encoder. We then pass these light-normalized encoder features to the decoder of the U-Net and apply another set of AdaIN guided by the features extracted from the target lighting with the Light Encoder. We additionally predict source lighting from the U-Net encoder using a Light Decoder." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.447, + 0.788, + 0.509 + ], + "angle": 0, + "content": "unsuitable. Previous work [30] utilized segmentation for pairing. However, we observed that segmentation is ineffective at finding pairs of images with the same pose and expression. Thus, we instead use facial keypoint detection [16] to obtain source and target image pairs." + }, + { + "type": "title", + "bbox": [ + 0.215, + 0.529, + 0.525, + 0.545 + ], + "angle": 0, + "content": "3.2 Relighting network architecture" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.553, + 0.789, + 0.794 + ], + "angle": 0, + "content": "Our network architecture, as illustrated in Fig. 3, is built upon the well-established U-Net [29]. This architecture is comprised of an encoder and a decoder with skip connections, which are commonly used in existing portrait relighting algorithms [36,30,49,13,38]. Our U-Net's encoder, similar to Sengupta et al. [30] and Sun et al. [36], processes the source portrait \\( I_{\\mathrm{src}} \\) by applying multiple convolutional layers of varying strides (1 or 2). This process progressively reduces spatial resolution while increasing the number of channels, yielding a latent feature space. The decoder performs the inverse function of the encoder by upsampling from the latent features and simultaneously skip-connecting to intermediate features from the encoder. These skip connections transport high-frequency shape and appearance information from the encoder to the decoder, ultimately resulting in the generation of a realistic relit image. However, they also carry source illumination features from the encoder to the decoder, leading to subpar relighting quality and temporal flickering. To address this issue, we introduce light-conditioned feature normalization (LCFN) for the skip-connected features to better disentangle lighting features from intrinsic appearance features." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.796, + 0.788, + 0.842 + ], + "angle": 0, + "content": "To disentangle lighting and intrinsic appearance components from the encoded features - i.e. to de-light - we first predict the source lighting from the encoder features. In contrast to Sun et al. [36], which predicts the illumination" + } + ], + [ + { + "type": "header", + "bbox": [ + 0.324, + 0.115, + 0.733, + 0.131 + ], + "angle": 0, + "content": "Personalized Video Relighting With an At-Home Light Stage" + }, + { + "type": "page_number", + "bbox": [ + 0.775, + 0.116, + 0.787, + 0.127 + ], + "angle": 0, + "content": "7" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.146, + 0.785, + 0.207 + ], + "angle": 0, + "content": "\\(\\hat{L}_{\\mathrm{src}}\\) corresponding to the source image \\(I_{\\mathrm{src}}\\) using the final encoded features, we take a different approach. We extract features at intermediate steps within the encoder, downsample them, and concatenate them using a confidence learning approach [14] to predict the illumination \\(\\hat{L}_{\\mathrm{src}}\\)." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.208, + 0.788, + 0.466 + ], + "angle": 0, + "content": "The LCFN module uses the lighting features generated by the lighting encoder to perform Adaptive Instance Normalization (AdaIN) [17] on the encoder features. We begin by using a multi-layer perceptron (MLP) to encode lighting features, transforming the lighting information of \\(L_{\\mathrm{src}}\\) and \\(L_{\\mathrm{trg}}\\) into a compact, low-dimensional representation \\((\\mathrm{d} = 256)\\). We apply AdaIN to encoder features using the source lighting features, producing normalized features \\(f^l\\) (for \\(l = 1,\\dots ,7\\)). Through this normalization process, we induce de-lighting, effectively removing the lighting information present in the encoder features. Starting from the de-lit latent features \\(f^7\\), we perform progressive bi-linear upsampling. At each upsampling step, we apply AdaIN to the concatenated feature, incorporating the target lighting features encoded by the lighting encoder. This construction using the LCFN module and source lighting prediction allows us to effectively remove source lighting features from the input and only propagate intrinsic appearance features from the encoder to the decoder. We then add target lighting features in the decoder. The LCFN module also contributes towards temporal consistency (see Sec. 3.3). See Fig. 2 for a comparison between our architecture and those of Sun et al. [36] and Sengupta et al. [30]." + }, + { + "type": "title", + "bbox": [ + 0.215, + 0.484, + 0.522, + 0.499 + ], + "angle": 0, + "content": "3.3 Enforcing temporal consistency" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.506, + 0.785, + 0.596 + ], + "angle": 0, + "content": "Temporal consistency is vital in making relit videos stable, realistic, and aesthetically pleasing. Previous single-image portrait relighting techniques [30,36] do not incorporate explicit temporal modeling, leading to undesirable flickering artifacts when applied to videos. Accuracy in single-image portrait relighting can often be uncorrelated to temporal flickering. Inconsistencies across frames are even more noticeable when the source lighting \\( L_{\\mathrm{src}} \\) changes continuously." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.597, + 0.786, + 0.701 + ], + "angle": 0, + "content": "When applied to skip-connected features, LCFN provides a natural defense against temporal flickering by removing source lighting features from the input image. However, it cannot ensure temporal consistency on its own. We notice two further problems: (1) when the source lighting gradually changes, LCFN often leaks small amounts of source lighting information to the decoder, leading to flickering; (2) when source lighting changes abruptly, undesirable fading effects can be observed." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.702, + 0.786, + 0.779 + ], + "angle": 0, + "content": "To address this issue, we propose a skip-connected feature smoothing technique that assumes neighboring frames share the same intrinsic appearance features, obtained after de-lighting input image features with LCFN. We apply a simple exponential smoothing of de-lit features generated by LCFN, denoted as \\( f^l \\), using all the previous frames:" + }, + { + "type": "equation", + "bbox": [ + 0.341, + 0.785, + 0.785, + 0.803 + ], + "angle": 0, + "content": "\\[\nf _ {t} ^ {l} := \\alpha \\cdot f _ {t} ^ {l} + (1 - \\alpha) \\cdot f _ {t - 1} ^ {l} \\quad (\\text {f o r} l = 1, \\dots , 7) \\tag {2}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.81, + 0.785, + 0.84 + ], + "angle": 0, + "content": "with \\(\\alpha = 0.7\\). Note that exponential smoothing does not work without de-lit LCFN features, which removes time-varying source lighting." + } + ], + [ + { + "type": "page_number", + "bbox": [ + 0.218, + 0.116, + 0.23, + 0.127 + ], + "angle": 0, + "content": "8" + }, + { + "type": "header", + "bbox": [ + 0.272, + 0.115, + 0.361, + 0.128 + ], + "angle": 0, + "content": "J. Choi et al." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.147, + 0.788, + 0.222 + ], + "angle": 0, + "content": "We further notice that when the monitor light changes abruptly the relighting effect is delayed by a few frames, mainly due to the limited refresh rate of the monitor and frame rate of the camera. We thus propose doing a weighted average of source monitor lighting \\( L_{\\mathrm{src}} \\) from a sequence of previous and current frames to achieve smoother and more natural results:" + }, + { + "type": "equation", + "bbox": [ + 0.414, + 0.226, + 0.786, + 0.264 + ], + "angle": 0, + "content": "\\[\nL _ {\\text {s r c a v g}} ^ {t} = \\frac {\\sum_ {i = 0} ^ {N - 1} \\beta^ {i} L _ {\\text {s r c}} ^ {t - i}}{\\sum_ {i = 0} ^ {N - 1} \\beta^ {i}}, \\tag {3}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.216, + 0.268, + 0.407, + 0.281 + ], + "angle": 0, + "content": "where \\(\\beta = 0.6\\) and \\(N = 3\\)." + }, + { + "type": "title", + "bbox": [ + 0.216, + 0.301, + 0.49, + 0.316 + ], + "angle": 0, + "content": "3.4 Training relighting network" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.321, + 0.784, + 0.413 + ], + "angle": 0, + "content": "Our model is trained through minimizing a weighted combination of three loss functions: generator loss, discriminator loss, and monitor loss. The first loss aims to minimize the discrepancies between the true target image \\( I_{\\mathrm{trg}} \\) in our dataset and the predicted target relit image \\( \\hat{I}_{\\mathrm{trg}} \\), leading to accurately relit images. We adopted our generator loss (Eq. 4) and our discriminator loss (Eq. 5) from Sengupta et al. [30]:" + }, + { + "type": "equation", + "bbox": [ + 0.34, + 0.42, + 0.785, + 0.445 + ], + "angle": 0, + "content": "\\[\n\\mathcal {L} _ {G} ^ {I} = \\lambda_ {L 1} \\mathcal {L} _ {L 1} \\left(I _ {\\mathrm {t r g}}, \\hat {I} _ {\\mathrm {t r g}}\\right) + \\lambda_ {P} \\mathcal {L} _ {P} \\left(I _ {\\mathrm {t r g}}, \\hat {I} _ {\\mathrm {t r g}}\\right) \\tag {4}\n\\]" + }, + { + "type": "equation", + "bbox": [ + 0.365, + 0.441, + 0.664, + 0.458 + ], + "angle": 0, + "content": "\\[\n+ \\lambda_ {C} \\mathcal {L} _ {C} \\left(I _ {\\mathrm {s r c}}, \\hat {I} _ {\\mathrm {s r c}} ^ {C}\\right) + \\lambda_ {D} \\left(D \\left(\\hat {I} _ {\\mathrm {t r g}}; \\theta_ {D}\\right) - 1\\right) ^ {2},\n\\]" + }, + { + "type": "equation", + "bbox": [ + 0.355, + 0.469, + 0.785, + 0.485 + ], + "angle": 0, + "content": "\\[\n\\mathcal {L} _ {D} = \\left(D \\left(I _ {\\mathrm {t r g}}; \\theta_ {D}\\right) - 1\\right) ^ {2} + \\left(D \\left(\\hat {I} _ {\\mathrm {t r g}}; \\theta_ {D}\\right)\\right) ^ {2}, \\tag {5}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.489, + 0.784, + 0.538 + ], + "angle": 0, + "content": "where \\(\\mathcal{L}_{L1}\\) denotes L1 loss, \\(\\mathcal{L}_P\\) denotes perceptual loss [45], \\(\\mathcal{L}_C\\) denotes cycle consistency loss [50], and \\(D\\) is the discriminator [15]. \\(\\hat{I}_{\\mathrm{src}}^C\\) and \\(\\hat{L}_{\\mathrm{trg}}^C\\) are the outputs from \\(G(\\hat{I}_{\\mathrm{trg}}, L_{\\mathrm{trg}}, L_{\\mathrm{src}},; \\theta_G)\\)" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.538, + 0.784, + 0.596 + ], + "angle": 0, + "content": "The monitor reconstruction loss focuses on minimizing the errors between the predicted source light \\(\\hat{L}_{\\mathrm{src}}\\) and the true source light \\(L_{\\mathrm{src}}\\) and is expected to enforce improved disentanglement of lighting information from intrinsic appearance features." + }, + { + "type": "equation", + "bbox": [ + 0.273, + 0.603, + 0.785, + 0.62 + ], + "angle": 0, + "content": "\\[\n\\mathcal {L} _ {G} ^ {M} = \\lambda_ {L 1} \\mathcal {L} _ {L 1} \\left(L _ {\\mathrm {s r c}}, \\hat {L} _ {\\mathrm {s r c}}\\right) + \\lambda_ {P} \\mathcal {L} _ {P} \\left(L _ {\\mathrm {s r c}}, \\hat {L} _ {\\mathrm {s r c}}\\right) + \\lambda_ {C} \\mathcal {L} _ {C} \\left(L _ {\\mathrm {t r g}}, \\hat {L} _ {C} ^ {\\text {t r g}}\\right). \\tag {6}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.625, + 0.784, + 0.656 + ], + "angle": 0, + "content": "Finally, we minimize the image generator loss \\(\\mathcal{L}_G^I\\), the discriminator loss \\(\\mathcal{L}_D\\), and the illumination loss \\(\\mathcal{L}_G^M\\) together:" + }, + { + "type": "equation", + "bbox": [ + 0.418, + 0.661, + 0.785, + 0.683 + ], + "angle": 0, + "content": "\\[\n\\min _ {G, D} \\mathcal {L} _ {G} ^ {I} + \\mathcal {L} _ {D} + \\lambda_ {G} ^ {M} \\mathcal {L} _ {G} ^ {M}. \\tag {7}\n\\]" + }, + { + "type": "title", + "bbox": [ + 0.216, + 0.708, + 0.376, + 0.724 + ], + "angle": 0, + "content": "4 Experiments" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.735, + 0.788, + 0.841 + ], + "angle": 0, + "content": "In Sec. 4.1, we first discuss our data collection process, which was based on the approach used to compile the Light Stage at Your Desk (LSYD) dataset. In Sec. 4.2, we perform quantitative and qualitative comparisons with existing single-image portrait relighting algorithms and evaluate their temporal consistency. In Sec. 4.3, we demonstrate predicting, at a low-level, what the user is looking at on the monitor screen. Finally, in Sec. 4.4, we perform ablation studies, evaluating the network architecture's impact on relighting performance." + } + ], + [ + { + "type": "header", + "bbox": [ + 0.325, + 0.115, + 0.733, + 0.129 + ], + "angle": 0, + "content": "Personalized Video Relighting With an At-Home Light Stage" + }, + { + "type": "page_number", + "bbox": [ + 0.775, + 0.117, + 0.785, + 0.127 + ], + "angle": 0, + "content": "9" + }, + { + "type": "image", + "bbox": [ + 0.219, + 0.143, + 0.788, + 0.504 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.214, + 0.518, + 0.788, + 0.575 + ], + "angle": 0, + "content": "Fig. 4. We perform a qualitative comparison with existing techniques [30,36] on the LSYD dataset. Source and target (images & lighting) were unseen during training. All models are personalized, i.e. trained on images of that individual only. We (Col. 3) produce significantly better results compared to existing approaches (Cols 4 and 5)." + }, + { + "type": "title", + "bbox": [ + 0.216, + 0.603, + 0.304, + 0.617 + ], + "angle": 0, + "content": "4.1 Data" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.629, + 0.789, + 0.842 + ], + "angle": 0, + "content": "We recorded data from 5 users of diverse ethnicities and genders to ensure a wide range of skin types. Each participant wore a variety of outfits, and we used 4 different ambient lighting conditions per person to mimic the conditions of real-life online meetings. We directed the participants to continuously change their facial expression and pose during the capture sessions. Each user's face was captured while watching 8 different videos, each 8 minutes long, on different days with varying appearances. We randomly hold out 1 video for testing and use the remaining 7 videos for training. We use this testing sequence only for qualitative evaluation, not for any quantitative metrics. This is because quantitative evaluation requires a pair of source and target images of the same person in the same pose but under different lighting conditions. This is difficult to obtain accurately for the aforementioned test video sequence since the participants naturally vary their pose and expression over the course of the video. Instead, we capture an additional test sequence, used only for numerical evaluation in" + } + ], + [ + { + "type": "page_number", + "bbox": [ + 0.218, + 0.116, + 0.236, + 0.127 + ], + "angle": 0, + "content": "10" + }, + { + "type": "header", + "bbox": [ + 0.272, + 0.115, + 0.361, + 0.127 + ], + "angle": 0, + "content": "J. Choi et al." + }, + { + "type": "image", + "bbox": [ + 0.218, + 0.145, + 0.787, + 0.419 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.214, + 0.432, + 0.788, + 0.475 + ], + "angle": 0, + "content": "Fig. 5. We perform a qualitative comparison on the OLAT dataset [44]. Our approach outperforms [36,30] and can render strong directional lighting and specular highlights without any explicit modeling of geometry and reflectance." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.503, + 0.788, + 0.624 + ], + "angle": 0, + "content": "which the participant is captured in 9 different pose-expression combinations, each with a distinct monitor light. For each pose, we can create \\(\\binom{9}{2} = 36\\) source and target pairs as input and pseudo ground-truth, resulting in a total of 324 test data pairs per user. Additionally, we compared methods using Dynamic OLAT Dataset [44] with environment lighting maps [1,8]. During this comparison, we converted HDR environment maps to LDR and utilized a \\(270^{\\circ}\\) FoV as a monitor light to assess the relighting results. We want to note that all the test data is composed of unseen portraits and lightings for both LSYD and OLAT datasets." + }, + { + "type": "title", + "bbox": [ + 0.216, + 0.648, + 0.571, + 0.663 + ], + "angle": 0, + "content": "4.2 Comparison with existing approaches" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.674, + 0.787, + 0.734 + ], + "angle": 0, + "content": "We employ three error metrics to assess relighting performance: RMSE, LPIPS [45], and DISTS [6]. LPIPS and DISTS are more robust to slight differences in pose between the relit image and the pseudo ground truth and detect perceptual differences more effectively than RMSE." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.735, + 0.788, + 0.841 + ], + "angle": 0, + "content": "**Portrait image relighting.** We compared our approach with existing portrait relighting neural architectures — Sun et al. [36] and Sengupta et al. [30] — by training on our captured LSYD dataset using the same pre-processing for all three architectures (see Sec. 3.1). Our training loss, given in Sec. 3.4, can handle misalignment in source-target pairs in training data, similar to the loss proposed in Sengupta et al. [30] (we use an additional loss on source monitor lighting prediction). For Sun et al. [36], we train both with their original loss" + } + ], + [ + { + "type": "header", + "bbox": [ + 0.324, + 0.114, + 0.733, + 0.13 + ], + "angle": 0, + "content": "Personalized Video Relighting With an At-Home Light Stage" + }, + { + "type": "page_number", + "bbox": [ + 0.768, + 0.116, + 0.784, + 0.127 + ], + "angle": 0, + "content": "11" + }, + { + "type": "table_caption", + "bbox": [ + 0.214, + 0.143, + 0.788, + 0.242 + ], + "angle": 0, + "content": "Table 1. We train a personalized relighting model on 5 users from our LSYD dataset and 4 users from the OLAT dataset [44]. We evaluate these models using source & target portrait images, as well as lighting, that were not encountered during the training phase, and report average RMSE, LPIPS [46], and DISTS[6] scores both on LSYD and OLAT dataset. Our method can perform relighting without source lighting \\(L_{\\mathrm{src}}\\), by simply using the predicted light source from our model as input lighting. Our method significantly outperforms Sun et al. [36] and Sengupta et al. [30]." + }, + { + "type": "table", + "bbox": [ + 0.216, + 0.251, + 0.789, + 0.348 + ], + "angle": 0, + "content": "
Known source lightingLSYD dataOLAT data [44]
LsrcLPIPS ↓DISTS ↓RMSE ↓LPIPS ↓DISTS ↓RMSE ↓
Sun et al. [36] w/ LSun-0.17120.16298.59580.22730.17456.2692
Sun et al. [36] w/ LOurs-0.10290.11528.44760.22670.15696.1898
Ours-0.08390.09538.32220.18120.13366.0931
Sengupta et al. [30]✓0.10180.11058.28260.22370.16756.1751
Ours✓0.08320.09538.19390.18090.13345.9548
" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.381, + 0.789, + 0.457 + ], + "angle": 0, + "content": "function \\( L_{\\mathrm{Sun}} \\) (which expects perfect source-target pose alignment obtained in OLAT data) and with our proposed loss function \\( L_{\\mathrm{ours}} \\) to specifically handle misalignment in LSYD data. We train personalized models on 5 users from the LSYD dataset and on 4 users from the publicly available Dynamic OLAT Dataset [44] with 2361 indoor environment lighting maps [1,8]." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.46, + 0.449, + 0.746 + ], + "angle": 0, + "content": "For our quantitative evaluation, we test our model on 1620 test images across 5 users with unseen appearance and lighting conditions on the LSYD dataset and on 7172 test images from the Dynamic OLAT dataset. We present the result in Tab. 1. We observe that our proposed approach outperforms Sengupta et al. [30] and Sun et al. [36] by \\(22.3\\%\\) and \\(23.6\\%\\) respectively on the LSYD dataset and by \\(23.5\\%\\) and \\(25.6\\%\\) on the OLAT dataset, when comparing LPIPS score. Our qualitative comparison, as presented in Fig. 4 and Fig. 5, shows that our model performs superior re" + }, + { + "type": "image", + "bbox": [ + 0.461, + 0.46, + 0.784, + 0.655 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.456, + 0.668, + 0.788, + 0.74 + ], + "angle": 0, + "content": "Fig.6. Our method can also relight portrait images captured with unknown light sources. We relight a source image (Col. 1) with target light shown in the inset in Col. 2. We add a reference of how the face appears under that target light." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.747, + 0.788, + 0.777 + ], + "angle": 0, + "content": "lighting in terms of color, quality, and consistency, and can render strong directional lighting and specular highlights without any explicit modeling." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.78, + 0.789, + 0.841 + ], + "angle": 0, + "content": "Note that Sun et al. [36] does not require the source lighting \\( L_{\\mathrm{src}} \\) during test time. Our proposed approach can also perform relighting without prior knowledge of source lighting \\( L_{\\mathrm{src}} \\) by simply predicting \\( \\hat{L}_{\\mathrm{src}} \\) and using it for light-conditioned feature normalization. We show that even in the absence of" + } + ], + [ + { + "type": "page_number", + "bbox": [ + 0.218, + 0.116, + 0.236, + 0.127 + ], + "angle": 0, + "content": "12" + }, + { + "type": "header", + "bbox": [ + 0.272, + 0.115, + 0.361, + 0.128 + ], + "angle": 0, + "content": "J. Choi et al." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.146, + 0.784, + 0.178 + ], + "angle": 0, + "content": "\\(L_{\\mathrm{src}}\\), our method outperforms Sun et al. [36] by \\(22.6\\%\\) on the LSYD data and by \\(25.5\\%\\) on the OLAT data, in terms of LPIPS." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.182, + 0.785, + 0.273 + ], + "angle": 0, + "content": "In Fig. 6, we demonstrate that our approach can relight any portrait image captured 'in-the-wild' without requiring the knowledge of source lighting, and outperforms Sun et al. [36]. Our approach first predicts a proxy source monitor lighting, imagining the portrait image to be captured under a monitor lighting, and uses this predicted source lighting in the LCFN module to 'de-lit' the input image and relight it with a target lighting." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.277, + 0.446, + 0.488 + ], + "angle": 0, + "content": "In Fig. 7 1st-row, we show the results of relighting an input image with directional lights by moving specific bright areas on the monitor screen. Our method learns to render directional lighting effects and cast shadows as needed. To better illustrate the effect of moving lighting, we first relight the input image with no light reflected from the monitor (row-2 col-1), which produces a relief image under ambient room" + }, + { + "type": "image", + "bbox": [ + 0.462, + 0.275, + 0.528, + 0.373 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.528, + 0.275, + 0.59, + 0.373 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.59, + 0.275, + 0.655, + 0.373 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.655, + 0.275, + 0.718, + 0.373 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.718, + 0.275, + 0.782, + 0.373 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.456, + 0.389, + 0.788, + 0.486 + ], + "angle": 0, + "content": "Fig. 7. We relight the input image with directional lights (1st row) and without any target light ('Relight A'). We then subtract the 'Relight A' from directional images (row-2 col-2:5). Our method learns to render the effects of directional lighting and decouple ambient and dominant frontal lighting (from monitor)." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.489, + 0.785, + 0.581 + ], + "angle": 0, + "content": "lighting only. Subtracting this 'de-lit' image from the relit images under directional lighting (row-2 col 2-5) highlights the ability of our method to decouple ambient room illumination from dominant frontal lighting. However, due to the constraint that our lighting is limited to the illumination emitted from the monitor, we may not accurately depict extreme lighting effects. To this end, we show the strong directional light effects using the OLAT dataset in Fig. 5." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.584, + 0.465, + 0.841 + ], + "angle": 0, + "content": "Portrait video relighting. Next, we evaluate the temporal consistency of each portrait video relighting algorithm. For each user in the LSYD data, we relit the held-out test video with 50 different target lighting conditions, creating 50 relit videos. We then computed the RMSE between adjacent frames in relit videos as a measure of temporal consistency. Since the pose is almost identical between adjacent frames, lower RMSE error indicates temporally consistent relighting. We then report the average temporal RMSE across all such adjacent frames. In practice," + }, + { + "type": "table_caption", + "bbox": [ + 0.473, + 0.585, + 0.788, + 0.696 + ], + "angle": 0, + "content": "Table 2. We evaluate temporal consistency by relighting a test video captured with varying source lighting with the same target lighting and calculating RMSE between adjacent frames. We then report the average RMSE across all adjacent frames and compute an error rate to indicate the percent of adjacent frames with RMSE higher than a threshold." + }, + { + "type": "table", + "bbox": [ + 0.476, + 0.698, + 0.787, + 0.832 + ], + "angle": 0, + "content": "
ThresholdRMSE ↓Error Rate (%)
>0.2>0.3>0.4
Sun et al. [36]5.8613.532.611.06
Sengupta et al. [30]6.3721.835.401.75
+Lsrc_avg5.7613.312.340.98
Ours (base)6.0116.223.611.08
+Lsrc_avg5.7313.092.310.83
+LCFN5.6813.042.280.71
+Lsrc_avg + LCFN5.5512.892.220.65
" + } + ], + [ + { + "type": "header", + "bbox": [ + 0.325, + 0.115, + 0.733, + 0.13 + ], + "angle": 0, + "content": "Personalized Video Relighting With an At-Home Light Stage" + }, + { + "type": "page_number", + "bbox": [ + 0.769, + 0.116, + 0.785, + 0.127 + ], + "angle": 0, + "content": "13" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.147, + 0.788, + 0.253 + ], + "angle": 0, + "content": "however, a significant fraction of adjacent frame pairs have extremely similar lighting between the two frames, making their relit frames naturally consistent anyway. Only in a small percentage of adjacent frames does the source lighting significantly change, leading to obvious flickering in the relit video if temporal consistency is not maintained. Thus, in addition to average temporal RMSE, we also compute the error rate for three different thresholds: 0.2 (low), 0.3 (medium), and 0.4 (high), which indicate the percentage of adjacent frames where RMSE" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.253, + 0.454, + 0.267 + ], + "angle": 0, + "content": "error is more than the threshold." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.267, + 0.449, + 0.631 + ], + "angle": 0, + "content": "In Tab. 2 and Fig. 8, we compare our approach with and without skip-connected feature smoothing to past works [36,30]. Note that this temporal smoothing of the skip-connected features can only be applied in our framework since we de-light encoder features from the source lighting with LCFN. Both for our approach and for Sengupta et al. [30], we can further apply smoothing of input source lighting to handle abrupt changes. We observe that our method produces the most temporally consistent relighting while also being the most accurate (see Tab. 1). We further note that both skip-connected feature smoothing and smoothing of source lighting improve temporal consistency." + }, + { + "type": "image", + "bbox": [ + 0.46, + 0.266, + 0.785, + 0.519 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.456, + 0.533, + 0.788, + 0.629 + ], + "angle": 0, + "content": "Fig.8. We show temporal consistency between adjacent frames separated by 0.33s by relighting a test video captured with varying source lighting with the same target lighting. Note that [36,30] both exhibit abrupt changes in lighting between frames \\( t \\) and \\( t + 10 \\), while our approach produces a more stable result." + }, + { + "type": "title", + "bbox": [ + 0.216, + 0.651, + 0.421, + 0.667 + ], + "angle": 0, + "content": "4.3 Monitor prediction" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.674, + 0.447, + 0.824 + ], + "angle": 0, + "content": "Our network can also be used to predict the source monitor lighting of any input image, as shown in Fig. 9. For the LSYD dataset, we do have images of the source monitor for every input image, so we can calculate monitor prediction accuracy numerically. We observe that our method slightly out" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.825, + 0.787, + 0.842 + ], + "angle": 0, + "content": "performs [36] with a Mean Absolute Error of 0.15 vs 0.17, also qualitatively" + }, + { + "type": "image", + "bbox": [ + 0.46, + 0.672, + 0.784, + 0.757 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.456, + 0.77, + 0.788, + 0.826 + ], + "angle": 0, + "content": "Fig. 9. Monitor prediction comparison on LSYD data (Col. 1 and 2) and \"in the wild\" setting (Col. 3 and 4). The second row represents the monitor prediction by [36] and Ours." + } + ], + [ + { + "type": "page_number", + "bbox": [ + 0.218, + 0.116, + 0.236, + 0.127 + ], + "angle": 0, + "content": "14" + }, + { + "type": "header", + "bbox": [ + 0.272, + 0.115, + 0.361, + 0.127 + ], + "angle": 0, + "content": "J. Choi et al." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.147, + 0.784, + 0.177 + ], + "angle": 0, + "content": "producing more meaningful visualizations. For portrait images, we do not have ground-truth source lighting, but visualizations show meaningful predictions." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.178, + 0.785, + 0.268 + ], + "angle": 0, + "content": "The ability to predict monitor lighting from images captured by our monitor has many implications. This technique can be used to detect deep fake avatars during live video calls by purposefully projecting specific images via screen sharing and observing if we can detect the same image from the webcam feed of other attendees in the call. If there is a mismatch between the projected monitor image and the predicted one, this likely indicates a deep fake avatar in the video call." + }, + { + "type": "title", + "bbox": [ + 0.215, + 0.29, + 0.398, + 0.305 + ], + "angle": 0, + "content": "4.4 Ablation studies" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.315, + 0.465, + 0.451 + ], + "angle": 0, + "content": "We report the removal of various components from our relighting network in Tab. 3, specifically LCFN and source monitor lighting prediction using intermediate encoder features. We observe that both improve final relighting performance, which shows their effectiveness in disentangling source" + }, + { + "type": "table_caption", + "bbox": [ + 0.474, + 0.316, + 0.787, + 0.372 + ], + "angle": 0, + "content": "Table 3. Both LCFN and source monitor prediction \\( \\mathbf{L}_{\\mathrm{src}} \\) improve relighting performance by effectively disentangling source lighting information from intrinsic appearance features." + }, + { + "type": "table", + "bbox": [ + 0.477, + 0.373, + 0.787, + 0.444 + ], + "angle": 0, + "content": "
Lsrcde-lightingRMSE ↓LPIPS ↓DISTS ↓
--8.42300.09640.1071
✓-8.25960.09150.1013
-✓8.19070.09040.0966
✓✓8.07460.08530.0963
" + }, + { + "type": "text", + "bbox": [ + 0.215, + 0.452, + 0.614, + 0.466 + ], + "angle": 0, + "content": "lighting information from intrinsic appearance features." + }, + { + "type": "title", + "bbox": [ + 0.215, + 0.489, + 0.359, + 0.505 + ], + "angle": 0, + "content": "5 Conclusion" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.521, + 0.785, + 0.672 + ], + "angle": 0, + "content": "We propose a personalized video relighting algorithm that leverages casually captured LSYD data to generate real-time high-quality temporally consistent relit videos under any pose, expression, and lighting conditions. We present a novel network architecture that can perform high-quality relighting on both the LSYD and OLAT datasets, rendering challenging lighting conditions like directional lights, shadows, and specularities. We achieve this without using any 3D information or performing explicit decomposition, simply by achieving better image-based relighting enabled by our proposed neural architecture with LCFN module. Our method enables better lighting quality during live video calls and in portrait images, and produces better harmonization with virtual backgrounds." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.672, + 0.785, + 0.747 + ], + "angle": 0, + "content": "Limitation. Since we utilize only the front-facing monitor light as the source lighting, we do not account for scenarios where the light source is located on the sides \\((90^{\\circ})\\) or behind \\((180^{\\circ})\\). However, it does not necessarily mean that our model cannot learn directional lighting effects, as demonstrated by training on the OLAT dataset which contains full \\(360^{\\circ}\\) lighting, see Fig. 5." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.747, + 0.787, + 0.839 + ], + "angle": 0, + "content": "Ethical considerations. While our primary goal is to allow people to improve their facial appearance with virtual relighting, we note that it is also a form of image manipulation and can be used for malicious purposes [3]. Furthermore, we emphasize that while currently we can only predict the user's direct line-of-sight monitor lighting in low resolution, the potential for high-resolution monitor prediction in the future could raise significant privacy concerns." + } + ], + [ + { + "type": "header", + "bbox": [ + 0.325, + 0.115, + 0.732, + 0.129 + ], + "angle": 0, + "content": "Personalized Video Relighting With an At-Home Light Stage" + }, + { + "type": "page_number", + "bbox": [ + 0.769, + 0.117, + 0.785, + 0.127 + ], + "angle": 0, + "content": "15" + }, + { + "type": "title", + "bbox": [ + 0.218, + 0.145, + 0.323, + 0.16 + ], + "angle": 0, + "content": "References" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.173, + 0.785, + 0.201 + ], + "angle": 0, + "content": "1. Bolduc, C., Giroux, J., Hébert, M., Demers, C., Lalonde, J.F.: Beyond the pixel: a photometrically calibrated hdr dataset for luminance and color prediction (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.201, + 0.756, + 0.214 + ], + "angle": 0, + "content": "2. Chen, Z., Liu, Z.: Relighting4d: Neural relightable human from videos (2022)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.214, + 0.785, + 0.268 + ], + "angle": 0, + "content": "3. Choi, J.M., Leung, J., Frahm, N., Christman, M., Bertasius, G., Sengupta, R.: Building secure and engaging video communication by using monitor illumination. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 4377-4386 (2024)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.268, + 0.785, + 0.31 + ], + "angle": 0, + "content": "4. Daichi Tajima, Yoshihiro Kanamori, Y.E.: Relighting humans in the wild: Monocular full-body human relighting with domain adaptation. Computer Graphics Forum (Proc. of Pacific Graphics 2021) 40(7), 205-216 (2021)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.31, + 0.785, + 0.363 + ], + "angle": 0, + "content": "5. Debevec, P., Wenger, A., Tchou, C., Gardner, A., Waese, J., Hawkins, T.: A lighting reproduction approach to live-action compositing. ACM Trans. Graph. 21(3), 547-556 (jul 2002). https://doi.org/10.1145/566654.566614, https://doi.org/10.1145/566654.566614" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.364, + 0.785, + 0.404 + ], + "angle": 0, + "content": "6. Ding, K., Ma, K., Wang, S., Simoncelli, E.P.: Image quality assessment: Unifying structure and texture similarity. CoRR abs/2004.07728 (2020), https://arxiv.org/abs/2004.07728" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.405, + 0.785, + 0.432 + ], + "angle": 0, + "content": "7. Futschik, D., Ritland, K., Vecore, J., Fanello, S., Orts-Escolano, S., Curless, B., Sykora, D., Pandey, R.: Controllable light diffusion for portraits (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.432, + 0.785, + 0.459 + ], + "angle": 0, + "content": "8. Gardner, M.A., Sunkavalli, K., Yumer, E., Shen, X., Gambaretto, E., Gagné, C., Lalonde, J.F.: Learning to predict indoor illumination from a single image (2017)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.459, + 0.785, + 0.512 + ], + "angle": 0, + "content": "9. Gerstner, C.R., Farid, H.: Detecting real-time deep-fake videos using active illumination. In: 2022 IEEE/CVF Conference on Computer Vision and Pattern Recognition Workshops (CVPRW). pp. 53-60 (2022). https://doi.org/10.1109/CVPRW56347.2022.00015" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.512, + 0.785, + 0.581 + ], + "angle": 0, + "content": "0. Guo, K., Lincoln, P., Davidson, P., Busch, J., Yu, X., Whalen, M., Harvey, G., Orts-Escalano, S., Pandey, R., Dourgarian, J., Tang, D., Tkach, A., Kowdle, A., Cooper, E., Dou, M., Fanello, S., Fyffe, G., Rhemann, C., Taylor, J., Debevec, P., Izadi, S.: The relightables: Volumetric performance capture of humans with realistic relighting (2019)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.581, + 0.785, + 0.636 + ], + "angle": 0, + "content": "1. He, Y., Xing, Y., Zhang, T., Chen, Q.: Unsupervised portrait shadow removal via generative priors. In: Proceedings of the 29th ACM International Conference on Multimedia. MM '21, ACM (Oct 2021). https://doi.org/10.1145/3474085.3475663, http://dx.doi.org/10.1145/3474085.3475663" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.636, + 0.785, + 0.678 + ], + "angle": 0, + "content": "2. Hou, A., Sarkis, M., Bi, N., Tong, Y., Liu, X.: Face relighting with geometrically consistent shadows. In: In Proceeding of IEEE Computer Vision and Pattern Recognition. New Orleans, LA (June 2022)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.678, + 0.785, + 0.718 + ], + "angle": 0, + "content": "3. Hou, A., Zhang, Z., Sarkis, M., Bi, N., Tong, Y., Liu, X.: Towards high fidelity face relighting with realistic shadows. In: IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR) (2021)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.718, + 0.785, + 0.759 + ], + "angle": 0, + "content": "4. Hu, Y., Wang, B., Lin, S.: Fc 4: Fully convolutional color constancy with confidence-weighted pooling. In: Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition. pp. 4085-4094 (2017)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.759, + 0.785, + 0.785 + ], + "angle": 0, + "content": "5. Isola, P., Zhu, J.Y., Zhou, T., Efros, A.A.: Image-to-image translation with conditional adversarial networks. CVPR (2017)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.785, + 0.785, + 0.839 + ], + "angle": 0, + "content": "6. Jin, H., Liao, S., Shao, L.: Pixel-in-pixel net: Towards efficient facial landmark detection in the wild. International Journal of Computer Vision (Sep 2021). https://doi.org/10.1007/s11263-021-01521-4, http://dx.doi.org/10.1007/s11263-021-01521-4" + }, + { + "type": "list", + "bbox": [ + 0.226, + 0.173, + 0.785, + 0.839 + ], + "angle": 0, + "content": null + } + ], + [ + { + "type": "page_number", + "bbox": [ + 0.218, + 0.116, + 0.236, + 0.127 + ], + "angle": 0, + "content": "16" + }, + { + "type": "header", + "bbox": [ + 0.272, + 0.115, + 0.361, + 0.127 + ], + "angle": 0, + "content": "J. Choi et al." + }, + { + "type": "ref_text", + "bbox": [ + 0.218, + 0.148, + 0.785, + 0.175 + ], + "angle": 0, + "content": "17. Karras, T., Laine, S., Aittala, M., Hellsten, J., Lehtinen, J., Aila, T.: Analyzing and improving the image quality of StyleGAN. In: Proc. CVPR (2020)" + }, + { + "type": "ref_text", + "bbox": [ + 0.218, + 0.176, + 0.785, + 0.215 + ], + "angle": 0, + "content": "18. Lagunas, M., Sun, X., Yang, J., Villegas, R., Zhang, J., Shu, Z., Masia, B., Gutierrez, D.: Single-image full-body human relighting. CoRR abs/2107.07259 (2021), https://arxiv.org/abs/2107.07259" + }, + { + "type": "ref_text", + "bbox": [ + 0.218, + 0.217, + 0.785, + 0.242 + ], + "angle": 0, + "content": "19. LeGendre, C., Ma, W.C., Pandey, R., Fanello, S., Rhemann, C., Dourgarian, J., Busch, J., Debevec, P.: Learning illumination from diverse portraits (2020)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.243, + 0.785, + 0.283 + ], + "angle": 0, + "content": "20. Liu*, Y., Hou*, A., Huang, X., Ren, L., Liu, X.: Blind removal of facial foreign shadows. In: In Proceedings of British Machine Vision Conference (BMVC). London, UK (November 2022)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.284, + 0.785, + 0.365 + ], + "angle": 0, + "content": "21. Meka, A., Pandey, R., Haene, C., Orts-Escolano, S., Barnum, P., Davidson, P., Erickson, D., Zhang, Y., Taylor, J., Bouaziz, S., Legendre, C., Ma, W.C., Overbeck, R., Beeler, T., Debevec, P., Izadi, S., Theobalt, C., Rhemann, C., Fanello, S.: Deep relightable textures - volumetric performance capture with neural rendering. vol. 39 (December 2020). https://doi.org/10.1145/3414685.3417814, http://gvv.mpi-inf.mpg.de/projects/DeepRelightableTextures/" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.367, + 0.785, + 0.446 + ], + "angle": 0, + "content": "22. Nestmeyer, T., Lalonde, J., Matthews, I., Lehrmann, A.: Learning physics-guided face relighting under directional light. In: 2020 IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR). pp. 5123-5132. IEEE Computer Society, Los Alamitos, CA, USA (jun 2020). https://doi.org/10.1109/CVPR42600.2020.00517, https://doi.ieeecomputersociety.org/10.1109/CVPR42600.2020.00517" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.448, + 0.785, + 0.5 + ], + "angle": 0, + "content": "23. Pandey, R., Orts-Escolano, S., LeGendre, C., Haene, C., Bouaziz, S., Rhemann, C., Debevec, P., Fanello, S.: Total relighting: Learning to relight portraits for background replacement. vol. 40 (August 2021). https://doi.org/10.1145/3450626.3459872" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.502, + 0.785, + 0.555 + ], + "angle": 0, + "content": "24. Peers, P., Tamura, N., Matusik, W., Debevec, P.: Post-production facial performance relighting using reflectance transfer. ACM Trans. Graph. 26(3), 52-es (jul 2007). https://doi.org/10.1145/1276377.1276442, https://doi.org/10.1145/1276377.1276442" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.557, + 0.785, + 0.584 + ], + "angle": 0, + "content": "25. Ponglertnapakorn, P., Tritrong, N., Suwajanakorn, S.: Difareli: Diffusion face relighting (2023), https://arxiv.org/abs/2304.09479" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.585, + 0.785, + 0.61 + ], + "angle": 0, + "content": "26. Qi, L., Wu, J., Wang, A.N., Wang, S., Sengupta, R.: My3dgen: A scalable personalized 3d generative model (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.611, + 0.785, + 0.637 + ], + "angle": 0, + "content": "27. Qiu, H., Chen, Z., Jiang, Y., Zhou, H., Fan, X., Yang, L., Wu, W., Liu, Z.: Relitalk: Relightable talking portrait generation from a single video (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.638, + 0.785, + 0.664 + ], + "angle": 0, + "content": "28. Ren, M., Xiong, W., Yoon, J.S., Shu, Z., Zhang, J., Jung, H., Gerig, G., Zhang, H.: Relightful harmonization: Lighting-aware portrait background replacement (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.665, + 0.785, + 0.703 + ], + "angle": 0, + "content": "29. Ronneberger, O., Fischer, P., Brox, T.: U-net: Convolutional networks for biomedical image segmentation. CoRR abs/1505.04597 (2015), http://arxiv.org/abs/1505.04597" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.705, + 0.785, + 0.731 + ], + "angle": 0, + "content": "30. Sengupta, S., Curless, B., Kemelmacher-Shlizerman, I., Seitz, S.M.: A light stage on every desk. CoRR abs/2105.08051 (2021), https://arxiv.org/abs/2105.08051" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.732, + 0.785, + 0.772 + ], + "angle": 0, + "content": "31. Sengupta, S., Kanazawa, A., Castillo, C.D., Jacobs, D.W.: Sfsnet: Learning shape, reflectance and illuminance of faces in the wild'. In: Proceedings of the IEEE conference on computer vision and pattern recognition. pp. 6296-6305 (2018)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.773, + 0.785, + 0.799 + ], + "angle": 0, + "content": "32. Sevastopolsky, A., Ignatiev, S., Ferrer, G., Burnaev, E., Lempitsky, V.: Relightable 3d head portraits from a smartphone video (2020)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.8, + 0.785, + 0.84 + ], + "angle": 0, + "content": "33. Shih, Y., Paris, S., Barnes, C., Freeman, W.T., Durand, F.: Style transfer for headshot portraits. ACM Trans. Graph. 33(4) (jul 2014). https://doi.org/10.1145/2601097.2601137, https://doi.org/10.1145/2601097.2601137" + }, + { + "type": "list", + "bbox": [ + 0.217, + 0.148, + 0.785, + 0.84 + ], + "angle": 0, + "content": null + } + ], + [ + { + "type": "header", + "bbox": [ + 0.325, + 0.115, + 0.732, + 0.129 + ], + "angle": 0, + "content": "Personalized Video Relighting With an At-Home Light Stage" + }, + { + "type": "page_number", + "bbox": [ + 0.769, + 0.117, + 0.785, + 0.127 + ], + "angle": 0, + "content": "17" + }, + { + "type": "ref_text", + "bbox": [ + 0.214, + 0.147, + 0.788, + 0.189 + ], + "angle": 0, + "content": "34. Shu, Z., Hadap, S., Shechtman, E., Sunkavalli, K., Paris, S., Samaras, D.: Portrait lighting transfer using a mass transport approach. ACM Trans. Graph. 37(1) (oct 2017). https://doi.org/10.1145/3095816, https://doi.org/10.1145/3095816" + }, + { + "type": "ref_text", + "bbox": [ + 0.214, + 0.189, + 0.788, + 0.215 + ], + "angle": 0, + "content": "35. Song, G., Cham, T.J., Cai, J., Zheng, J.: Half-body portrait relighting with overcomplete lighting representation (06 2021)" + }, + { + "type": "ref_text", + "bbox": [ + 0.214, + 0.215, + 0.788, + 0.257 + ], + "angle": 0, + "content": "36. Sun, T., Barron, J.T., Tsai, Y., Xu, Z., Yu, X., Fyffe, G., Rhemann, C., Busch, J., Debevec, P.E., Ramamoorthi, R.: Single image portrait relighting. CoRR abs/1905.00824 (2019), http://arxiv.org/abs/1905.00824" + }, + { + "type": "ref_text", + "bbox": [ + 0.214, + 0.257, + 0.788, + 0.282 + ], + "angle": 0, + "content": "37. Sun, T., Lin, K.E., Bi, S., Xu, Z., Ramamoorthi, R.: Nelf: Neural light-transport field for portrait view synthesis and relighting (2021)" + }, + { + "type": "ref_text", + "bbox": [ + 0.214, + 0.282, + 0.788, + 0.335 + ], + "angle": 0, + "content": "38. Sun, T., Xu, Z., Zhang, X., Fanello, S.R., Rhemann, C., Debevec, P.E., Tsai, Y., Barron, J.T., Ramamoorthi, R.: Light stage super-resolution: Continuous high-frequency relighting. CoRR abs/2010.08888 (2020), https://arxiv.org/abs/2010.08888" + }, + { + "type": "ref_text", + "bbox": [ + 0.214, + 0.335, + 0.788, + 0.377 + ], + "angle": 0, + "content": "39. Wang, Y., Holynski, A., Zhang, X., Zhang, X.: Sunstage: Portrait reconstruction and relighting using the sun as a light stage. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 20792-20802 (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.214, + 0.377, + 0.788, + 0.417 + ], + "angle": 0, + "content": "40. Wang, Z., Yu, X., Lu, M., Wang, Q., Qian, C., Xu, F.: Single image portrait relighting via explicit multiple reflectance channel modeling. ACM Trans. Graph. 39(6) (Nov 2020). https://doi.org/10.1145/3414685.3417824" + }, + { + "type": "ref_text", + "bbox": [ + 0.214, + 0.417, + 0.788, + 0.443 + ], + "angle": 0, + "content": "41. Yang, H., Zheng, M., Feng, W., Huang, H., Lai, Y.K., Wan, P., Wang, Z., Ma, C.: Towards practical capture of high-fidelity relightable avatars (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.214, + 0.443, + 0.788, + 0.482 + ], + "angle": 0, + "content": "42. Yeh, Y.Y., Nagano, K., Khamis, S., Kautz, J., Liu, M.Y., Wang, T.C.: 41(6), 1-21 (nov 2022). https://doi.org/10.1145/3550454.3555442, https://doi.org/10. 1145" + }, + { + "type": "ref_text", + "bbox": [ + 0.214, + 0.482, + 0.788, + 0.524 + ], + "angle": 0, + "content": "43. Yeh, Y.Y., Nagano, K., Khamis, S., Kautz, J., Liu, M.Y., Wang, T.C.: Learning to relight portrait images via a virtual light stage and synthetic-to-real adaptation. ACM Transactions on Graphics (TOG) (2022)" + }, + { + "type": "ref_text", + "bbox": [ + 0.214, + 0.524, + 0.788, + 0.551 + ], + "angle": 0, + "content": "44. Zhang, L., Zhang, Q., Wu, M., Yu, J., Xu, L.: Neural video portrait relighting in real-time via consistency modeling (2021)" + }, + { + "type": "ref_text", + "bbox": [ + 0.214, + 0.551, + 0.788, + 0.632 + ], + "angle": 0, + "content": "45. Zhang, R., Isola, P., Efros, A.A., Shechtman, E., Wang, O.: The unreasonable effectiveness of deep features as a perceptual metric. In: 2018 IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR). pp. 586-595. IEEE Computer Society, Los Alamitos, CA, USA (jun 2018). https://doi.org/10.1109/CVPR.2018.00068, https://doi.ieeecomputersociety.org/10.1109/CVPR.2018.00068" + }, + { + "type": "ref_text", + "bbox": [ + 0.214, + 0.632, + 0.788, + 0.672 + ], + "angle": 0, + "content": "46. Zhang, R., Isola, P., Efros, A.A., Shechtman, E., Wang, O.: The unreasonable effectiveness of deep features as a perceptual metric. CoRR abs/1801.03924 (2018), http://arxiv.org/abs/1801.03924" + }, + { + "type": "ref_text", + "bbox": [ + 0.214, + 0.672, + 0.788, + 0.727 + ], + "angle": 0, + "content": "47. Zhang, X., Fanello, S.R., Tsai, Y., Sun, T., Xue, T., Pandey, R., Orts-Escalano, S., Davidson, P.L., Rhemann, C., Debevec, P.E., Barron, J.T., Ramamoorthi, R., Freeman, W.T.: Neural light transport for relighting and view synthesis. CoRR abs/2008.03806 (2020), https://arxiv.org/abs/2008.03806" + }, + { + "type": "ref_text", + "bbox": [ + 0.214, + 0.727, + 0.788, + 0.753 + ], + "angle": 0, + "content": "48. Zhang, X., Barron, J.T., Tsai, Y.T., Pandey, R., Zhang, X., Ng, R., Jacobs, D.E.: Portrait shadow manipulation. vol. 39 (2020)" + }, + { + "type": "ref_text", + "bbox": [ + 0.214, + 0.753, + 0.788, + 0.793 + ], + "angle": 0, + "content": "49. Zhou, H., Hadap, S., Sunkavalli, K., Jacobs, D.: Deep single-image portrait relighting. In: 2019 IEEE/CVF International Conference on Computer Vision (ICCV). pp. 7193-7201 (2019). https://doi.org/10.1109/ICCV.2019.00729" + }, + { + "type": "ref_text", + "bbox": [ + 0.214, + 0.793, + 0.788, + 0.835 + ], + "angle": 0, + "content": "50. Zhu, J.Y., Park, T., Isola, P., Efros, A.A.: Unpaired image-to-image translation using cycle-consistent adversarial networks. In: Computer Vision (ICCV), 2017 IEEE International Conference on (2017)" + }, + { + "type": "list", + "bbox": [ + 0.214, + 0.147, + 0.788, + 0.835 + ], + "angle": 0, + "content": null + } + ] +] \ No newline at end of file diff --git a/2024/Personalized Video Relighting With an At-Home Light Stage/0751f00d-698a-4afe-b1e5-471a9991f11a_origin.pdf b/2024/Personalized Video Relighting With an At-Home Light Stage/0751f00d-698a-4afe-b1e5-471a9991f11a_origin.pdf new file mode 100644 index 0000000000000000000000000000000000000000..cdd7f59fb10787359b076c04513edb030fe8ccdf --- /dev/null +++ b/2024/Personalized Video Relighting With an At-Home Light Stage/0751f00d-698a-4afe-b1e5-471a9991f11a_origin.pdf @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:b79f8a23f5be02b26f087a2b193685bad5603aa1400d12edc44a32869468f4dd +size 8906615 diff --git a/2024/Personalized Video Relighting With an At-Home Light Stage/full.md b/2024/Personalized Video Relighting With an At-Home Light Stage/full.md new file mode 100644 index 0000000000000000000000000000000000000000..696861760a9fe59ccd2181bbac499b595ba74ca3 --- /dev/null +++ b/2024/Personalized Video Relighting With an At-Home Light Stage/full.md @@ -0,0 +1,303 @@ +# Personalized Video Relighting With an At-Home Light Stage + +Jun Myeong Choi, Max Christman, and Roni Sengupta + +University of North Carolina at Chapel Hill, NC 27514, USA {chedgekr,mrlc,ronisen}@cs.unc.edu + +Abstract. In this paper, we develop a personalized video relighting algorithm that produces high-quality and temporally consistent relit videos under any pose, expression, and lighting condition in real-time. Existing relighting algorithms typically rely either on publicly available synthetic data, which yields poor relighting results, or on actual light stage data which is difficult to acquire. We show that by just capturing recordings of a user watching YouTube videos on a monitor we can train a personalized algorithm capable of performing high-quality relighting under any condition. Our key contribution is a novel image-based neural relighting architecture that effectively separates the intrinsic appearance features - the geometry and reflectance of the face - from the source lighting and then combines them with the target lighting to generate a relit image. This neural architecture enables smoothing of intrinsic appearance features leading to temporally stable video relighting. Both qualitative and quantitative evaluations show that our architecture improves portrait image relighting quality and temporal consistency over state-of-the-art approaches on both casually captured 'Light Stage at Your Desk' (LSYD) and light-stage-captured 'One Light At a Time' (OLAT) datasets. Source code is available at https://github.com/chedgekorea/relighting + +Keywords: Image-based Relighting $\cdot$ 'At Home' Light Stage + +# 1 Introduction + +With the recent rise in popularity of video conferencing for business, educational, and personal activities, there is a significant demand for improving facial lighting. Virtually relighting our images and videos helps us improve the appearance of our faces without requiring explicit studio-quality lighting in a dedicated space or any specialized lighting expertise. Recent advances in deep neural networks have renewed interest in the problem of virtual relighting. + +Training a deep neural network for relighting requires extensive training data that includes source images paired with relit target images. One way of acquiring this data is by using a large spherical rig with numerous lights and cameras, known as a light stage [5]. While light stage data has been shown to produce high-quality relighting results [36,44,23,38,22,48], the limited availability of datasets, trained models, and access to the light stage itself has impeded further research. + +![](images/1951539ed1a5559b57e7ce4e301eaaefd57d9678d1ca1e0db360061607e81ca4.jpg) +(a) LSYD + +![](images/79068ed969abfda3cc3af4a60fde3d37c4d59d765d1610e5f6b912f45b7224e3.jpg) +(b) Input Image +(c) Relit-Net + +![](images/e5ed7a9ff0b984b3fc910c98baafa9bbed2a936230c25cfc8842529b713f165d.jpg) +Fig. 1. We learn a personalized relighting algorithm that generates temporally consistent and high-quality portrait videos under different lighting. We train the network using recordings of users watching YouTube on a monitor, thereby creating a Light Stage at Your Desk (LSYD). We project a portion of the LDR environment map with a $180^{\circ}$ FoV as the monitor light, while a portion of the remaining $90^{\circ}$ FoV is mapped as the background. We can achieve a harmonization effect with the virtual background. + +![](images/498cf8b11c53efdcb57fd436e2dac13f0e565b5e8446a7fa2bfdc4ac4ab73ee0.jpg) +(d) Output + +![](images/b92e180af5d3093e1ef4b72359ac9bb55874352b69ba10f3d37ad269ef4190ab.jpg) + +For example, One Light At a Time (Dynamic-OLAT) [44] is the only publicly available light stage relighting dataset consisting of four individuals only. As a result, researchers have often turned to synthetic data to train their relighting algorithms [31,49,13,35]. Unfortunately, existing synthetic data compromises the quality of relit images. + +We draw inspiration from recent work by Sengupta et al. [30,26] and develop a personalized relighting model by capturing a single user's appearance while being lit by a computer monitor. However, the results and applications of Sengupta et al. [30] have several limitations. First, it requires capturing users with fixed poses and expressions, an unrealistic requirement for actual users that impedes casual capture. Second, the relighting algorithm requires a dark room with negligible ambient light, limiting the environment in which this can be applied. Third, the resulting relit video is temporally unstable, exhibiting significant flickering artifacts, making it unsuitable for relighting Zoom calls. Lastly, it requires knowledge of source lighting and thus is unable to relight arbitrary portrait images captured in the wild. + +In this paper, we show that casually captured light stage data is sufficient to develop a high-quality temporally consistent video portrait relighting algorithm that works under arbitrary conditions (i.e. pose, expression, and ambient lighting) in real-time ( $\sim 45$ fps). To that end, we create our own casually captured light stage dataset with varying pose, expression, and lighting, called Light Stage at Your Desk (LSYD). Our key contribution is a neural image-based relighting architecture, based on the commonly used U-Net [36,49,22,40,38], that better disentangles the source lighting from the user's intrinsic facial appearance (shape and reflectance) and then adds back the target lighting to generate a relit image. Existing image-based relighting architectures [36,30] fail to accurately separate source lighting information from intrinsic appearance features in the encoder, leading to inconsistent and temporally unstable video relighting. To this end, we introduce the light-conditioned feature normalization (LCFN) module, which performs relighting and also predicts the source lighting from an input image. The LCFN module also enables temporal stability by performing exponential smoothing of de-lit intrinsic appearance features and facilitates relighting + +![](images/d784e968271e27145566bd56bcd0d2a19d76403025d4ab61a5c09277c144379d.jpg) +Fig. 2. We highlight the key structural differences between our relighting architecture and that of [36,30]. Our approach removes source lighting information from input image features and only propagates intrinsic appearance (geometry and reflectance) features from the encoder to decoder, which results in better relighting quality and more temporal stability. In contrast [36,30] propagates entire image features from the encoder to the decoder without 'de-light', and expects the decoder to remove source lighting and add target lighting information. + +![](images/1ced49ce92679efc4604fff228850a52d0086db1ef5eb137770013b58345d242.jpg) + +![](images/630e0bb2e4047e9f8bb240ca08bcb1f2163bdc72ee37c2f48a3fc373b40cbf58.jpg) + +of any arbitrary portrait image with unknown source lighting. We also improve the data pre-processing pipeline from Sengupta et al. [30] to make the relighting algorithm more robust to pose, expression, and ambient lighting conditions. + +We compare our relighting network with two other algorithms: Sun et al. [36], which was originally trained on light stage data (OLAT), and Sengupta et al. [30], which was originally trained on casually captured data (albeit with fixed pose, expression, and no ambient lighting). For a fair comparison, we train all algorithms for personalized relighting using the same data pre-processing steps and loss functions on 5 individuals from our LSYD dataset and 4 individuals from OLAT [44]. Our network outperforms Sun et al. [36] and Sengupta et al. [30] by $22.3\%$ and $23.6\%$ respectively on the LSYD dataset and by $23.5\%$ and $25.6\%$ on the OLAT dataset, in terms of LPIPS. Qualitatively our method produces superior relighting in terms of color and quality. We further show that our approach is more temporally consistent, leading to less flickering than Sun et al. [36] or Sengupta et al. [30]. Detailed ablation studies show that LCFN and source monitor prediction improves relighting quality, feature, and source monitor smoothing improves temporal consistency, and data pre-processing improves robustness to pose and expression. + +To summarize, our main contributions are as follows: + +- We show that casually captured Light Stage at Your Desk (LSYD) data can be used to build a high-quality temporally consistent personalized video relighting algorithm without requiring access to an expensive light stage setup. +- We introduce a novel video relighting architecture that separates the source lighting from the user's intrinsic appearance features and then adds back the target lighting, leading to improved relighting and temporal consistency for videos. +- While our relighting network focuses on 'at home' Light Stage (LSYD dataset) outperforming state-of-the-art algorithms, we also perform equally well on actual Light Stage captured OLAT [44] datasets, and any arbitrary portrait image captured 'in the wild'. + +# 2 Related Work + +Portrait relighting methods change the appearance of the face to match a target lighting condition. This can be expressed through lighting parameters (e.g., an environment map, spherical harmonics, directional lighting, etc.) or through a reference image of another person. Our approach relights a portrait image to a lighting condition expressed through a low dynamic range (LDR) image representing the image on the monitor. + +Image based relighting. Before the rise of deep learning, attempts were made at non deep learning image based relighting [34,24,33]. Shu et al. [34] introduced a face relighting approach that uses a mass-transport formulation for the transfer of illumination between images. Peers et al. [24] demonstrated a method for relighting portrait images with flat lighting to match specific target environments, incorporating a reference subject database for approximation. Shih et al. [33] adopted a multiscale technique to transfer local image statistics from reference portraits onto new ones, facilitating the matching of attributes like local contrast and overall lighting direction. Recent advancements in deep learning have caused significant shifts to the landscape of portrait relighting [25,11,20]. Research on ratio images for relighting [49,13] was conducted, utilizing public datasets and employing methods based on ratio images. However, this is limited to synthetic data, resulting in a significant domain gap with real data. The widespread application of light stages [5] in gathering data has enabled numerous groundbreaking research endeavors[36,44,38,28]. However, these methods rely on capturing data with a light stage, which are expensive and inaccessible. Instead, our approach builds a personalized relighting algorithm using casually captured videos from the desk recording setup introduced in Sengupta et al. [30]. In contrast to Sengupta et al. [30], we can collect data during daily computer usage by minimizing numerous constraints, eliminating the need for specific efforts in data collection. + +Relighting with explicit decomposition. Creating a virtual relighting dataset through the utilization of synthetic human models to train their networks has been carried out in some studies [31,18,42]. However, when using synthetic data to train a neural network, the large domain gap between synthetic and real data impacts the model's performance on real data. In contrast, others [12,4,27,2,37] generate relit images by using public datasets and employ methods based on 3D model rendering. Specifically, Hou et al. [12] takes a more advanced approach by introducing explicit components, where rays originating from the face intersect with other parts of the facial geometry to create relit images. Moreover, the extensive use of light stages [5] has enabled numerous innovative studies [22,40,23,10,43,7,21,47,38,41,19] in this domain. Some researchers have incorporated explicit elements such as albedo, normals, specular maps, and diffuse maps into their methodologies [22,40,23,7]. Others have taken a physics-based rendering approach [10,43] to resolve these issues. Other papers [21,47] aim to manipulate lighting conditions and generate images under different lighting scenarios using texture information. However, explicit decomposition methods require ground truth (GT) data for these intrinsic components to train, either from synthetic data or real light stage data. Obtaining GT data for 'at-home' + +captures is impossible, thus we focus on image-based relighting. In contrast, recent studies aim to streamline the capture process, using a mobile phone camera [32] or a sun stage [39] instead of a light stage. Nevertheless, due to their reliance on per-scene optimization, both of these papers lack the capability for real-time relighting and to generalize to unseen appearances of the individual and are only limited to the particular capture. Instead, our approach performs image-based relighting enabling real-time temporally consistent video relighting, at $\sim 45$ fps. + +# 3 Method + +Our setup is similar to Gerstner et al. [9] and Sengupta et al. [30], where a user's face is captured while illuminated by their monitor. By capturing multiple videos of the user's face along with the video on their monitor, we build our 'at home' light stage dataset. We then use these data to train a personalized portrait relighting algorithm that can render the user's face under arbitrary lighting conditions. Specifically, given a portrait image $I_{\mathrm{src}}$ , corresponding source monitor lighting $L_{\mathrm{src}}$ , and target monitor lighting $L_{\mathrm{trg}}$ , our aim is to learn a function $G$ that relights $I_{\mathrm{src}}$ under $L_{\mathrm{trg}}$ : + +$$ +\hat {I} _ {\mathrm {t r g}}, \hat {L} _ {\mathrm {s r c}} = G \left(I _ {\mathrm {s r c}}, L _ {\mathrm {s r c}}, L _ {\mathrm {t r g}}; \theta_ {G}\right). \tag {1} +$$ + +Note that our formulation can be used for scenarios where the source lighting is unknown by simply replacing the input source lighting with the predicted source lighting, unlike previous approaches [30]. + +In the following sections, we outline our methodology for portrait video relighting using a monitor as a light stage. Sec. 3.1 outlines strategies for constructing training data pairs from casually captured videos that allow flexibility in facial expression, pose, and ambient lighting. In Sec. 3.2, we introduce our relighting network architecture that disentangles lighting from intrinsic appearance using light-conditioned feature normalization, leading to high-quality relit images. In Sec. 3.3, we propose additional techniques which enforce temporal consistency and eliminate flickering, also using LCFN. Finally, in Sec. 3.4, we discuss how to train our relighting network. + +# 3.1 Constructing training data pairs + +While past work [30] imposed requirements of a neutral pose, expression, and dark room, we loosen these constraints to allow subjects in any conditions. The only constraint we maintain is that the room lighting shall not overpower the light emitted from the monitor. For example, if the capture occurs in front of a window with bright sunlight, the light from the monitor will have minimal impact on the face. As in Sengupta et al. [30], we aim to generate source and monitor image pairs $(I_{\mathrm{src}}, L_{\mathrm{src}})$ , as well as target image and target monitor pairs $(I_{\mathrm{trg}}, L_{\mathrm{trg}})$ , such that we can train our network to produce $\tilde{I}_{\mathrm{trg}}$ where $I_{\mathrm{trg}}$ is the pseudo ground truth. However, due to unrestricted subject movement during data collection, there is a lack of pixel-aligned data, making random pairs + +![](images/92c0bb905334e053dd6310840d1766b1017349265b63ae1111b14933f7e37fcf.jpg) +Fig. 3. We first de-light the input image features extracted by the U-Net encoder using Adaptive Instance Normalization (AdaIN) guided by the lighting features extracted from the source lighting with a Light Encoder. We then pass these light-normalized encoder features to the decoder of the U-Net and apply another set of AdaIN guided by the features extracted from the target lighting with the Light Encoder. We additionally predict source lighting from the U-Net encoder using a Light Decoder. + +unsuitable. Previous work [30] utilized segmentation for pairing. However, we observed that segmentation is ineffective at finding pairs of images with the same pose and expression. Thus, we instead use facial keypoint detection [16] to obtain source and target image pairs. + +# 3.2 Relighting network architecture + +Our network architecture, as illustrated in Fig. 3, is built upon the well-established U-Net [29]. This architecture is comprised of an encoder and a decoder with skip connections, which are commonly used in existing portrait relighting algorithms [36,30,49,13,38]. Our U-Net's encoder, similar to Sengupta et al. [30] and Sun et al. [36], processes the source portrait $I_{\mathrm{src}}$ by applying multiple convolutional layers of varying strides (1 or 2). This process progressively reduces spatial resolution while increasing the number of channels, yielding a latent feature space. The decoder performs the inverse function of the encoder by upsampling from the latent features and simultaneously skip-connecting to intermediate features from the encoder. These skip connections transport high-frequency shape and appearance information from the encoder to the decoder, ultimately resulting in the generation of a realistic relit image. However, they also carry source illumination features from the encoder to the decoder, leading to subpar relighting quality and temporal flickering. To address this issue, we introduce light-conditioned feature normalization (LCFN) for the skip-connected features to better disentangle lighting features from intrinsic appearance features. + +To disentangle lighting and intrinsic appearance components from the encoded features - i.e. to de-light - we first predict the source lighting from the encoder features. In contrast to Sun et al. [36], which predicts the illumination + +$\hat{L}_{\mathrm{src}}$ corresponding to the source image $I_{\mathrm{src}}$ using the final encoded features, we take a different approach. We extract features at intermediate steps within the encoder, downsample them, and concatenate them using a confidence learning approach [14] to predict the illumination $\hat{L}_{\mathrm{src}}$ . + +The LCFN module uses the lighting features generated by the lighting encoder to perform Adaptive Instance Normalization (AdaIN) [17] on the encoder features. We begin by using a multi-layer perceptron (MLP) to encode lighting features, transforming the lighting information of $L_{\mathrm{src}}$ and $L_{\mathrm{trg}}$ into a compact, low-dimensional representation $(\mathrm{d} = 256)$ . We apply AdaIN to encoder features using the source lighting features, producing normalized features $f^l$ (for $l = 1,\dots ,7$ ). Through this normalization process, we induce de-lighting, effectively removing the lighting information present in the encoder features. Starting from the de-lit latent features $f^7$ , we perform progressive bi-linear upsampling. At each upsampling step, we apply AdaIN to the concatenated feature, incorporating the target lighting features encoded by the lighting encoder. This construction using the LCFN module and source lighting prediction allows us to effectively remove source lighting features from the input and only propagate intrinsic appearance features from the encoder to the decoder. We then add target lighting features in the decoder. The LCFN module also contributes towards temporal consistency (see Sec. 3.3). See Fig. 2 for a comparison between our architecture and those of Sun et al. [36] and Sengupta et al. [30]. + +# 3.3 Enforcing temporal consistency + +Temporal consistency is vital in making relit videos stable, realistic, and aesthetically pleasing. Previous single-image portrait relighting techniques [30,36] do not incorporate explicit temporal modeling, leading to undesirable flickering artifacts when applied to videos. Accuracy in single-image portrait relighting can often be uncorrelated to temporal flickering. Inconsistencies across frames are even more noticeable when the source lighting $L_{\mathrm{src}}$ changes continuously. + +When applied to skip-connected features, LCFN provides a natural defense against temporal flickering by removing source lighting features from the input image. However, it cannot ensure temporal consistency on its own. We notice two further problems: (1) when the source lighting gradually changes, LCFN often leaks small amounts of source lighting information to the decoder, leading to flickering; (2) when source lighting changes abruptly, undesirable fading effects can be observed. + +To address this issue, we propose a skip-connected feature smoothing technique that assumes neighboring frames share the same intrinsic appearance features, obtained after de-lighting input image features with LCFN. We apply a simple exponential smoothing of de-lit features generated by LCFN, denoted as $f^l$ , using all the previous frames: + +$$ +f _ {t} ^ {l} := \alpha \cdot f _ {t} ^ {l} + (1 - \alpha) \cdot f _ {t - 1} ^ {l} \quad (\text {f o r} l = 1, \dots , 7) \tag {2} +$$ + +with $\alpha = 0.7$ . Note that exponential smoothing does not work without de-lit LCFN features, which removes time-varying source lighting. + +We further notice that when the monitor light changes abruptly the relighting effect is delayed by a few frames, mainly due to the limited refresh rate of the monitor and frame rate of the camera. We thus propose doing a weighted average of source monitor lighting $L_{\mathrm{src}}$ from a sequence of previous and current frames to achieve smoother and more natural results: + +$$ +L _ {\text {s r c a v g}} ^ {t} = \frac {\sum_ {i = 0} ^ {N - 1} \beta^ {i} L _ {\text {s r c}} ^ {t - i}}{\sum_ {i = 0} ^ {N - 1} \beta^ {i}}, \tag {3} +$$ + +where $\beta = 0.6$ and $N = 3$ . + +# 3.4 Training relighting network + +Our model is trained through minimizing a weighted combination of three loss functions: generator loss, discriminator loss, and monitor loss. The first loss aims to minimize the discrepancies between the true target image $I_{\mathrm{trg}}$ in our dataset and the predicted target relit image $\hat{I}_{\mathrm{trg}}$ , leading to accurately relit images. We adopted our generator loss (Eq. 4) and our discriminator loss (Eq. 5) from Sengupta et al. [30]: + +$$ +\mathcal {L} _ {G} ^ {I} = \lambda_ {L 1} \mathcal {L} _ {L 1} \left(I _ {\mathrm {t r g}}, \hat {I} _ {\mathrm {t r g}}\right) + \lambda_ {P} \mathcal {L} _ {P} \left(I _ {\mathrm {t r g}}, \hat {I} _ {\mathrm {t r g}}\right) \tag {4} +$$ + +$$ ++ \lambda_ {C} \mathcal {L} _ {C} \left(I _ {\mathrm {s r c}}, \hat {I} _ {\mathrm {s r c}} ^ {C}\right) + \lambda_ {D} \left(D \left(\hat {I} _ {\mathrm {t r g}}; \theta_ {D}\right) - 1\right) ^ {2}, +$$ + +$$ +\mathcal {L} _ {D} = \left(D \left(I _ {\mathrm {t r g}}; \theta_ {D}\right) - 1\right) ^ {2} + \left(D \left(\hat {I} _ {\mathrm {t r g}}; \theta_ {D}\right)\right) ^ {2}, \tag {5} +$$ + +where $\mathcal{L}_{L1}$ denotes L1 loss, $\mathcal{L}_P$ denotes perceptual loss [45], $\mathcal{L}_C$ denotes cycle consistency loss [50], and $D$ is the discriminator [15]. $\hat{I}_{\mathrm{src}}^C$ and $\hat{L}_{\mathrm{trg}}^C$ are the outputs from $G(\hat{I}_{\mathrm{trg}}, L_{\mathrm{trg}}, L_{\mathrm{src}},; \theta_G)$ + +The monitor reconstruction loss focuses on minimizing the errors between the predicted source light $\hat{L}_{\mathrm{src}}$ and the true source light $L_{\mathrm{src}}$ and is expected to enforce improved disentanglement of lighting information from intrinsic appearance features. + +$$ +\mathcal {L} _ {G} ^ {M} = \lambda_ {L 1} \mathcal {L} _ {L 1} \left(L _ {\mathrm {s r c}}, \hat {L} _ {\mathrm {s r c}}\right) + \lambda_ {P} \mathcal {L} _ {P} \left(L _ {\mathrm {s r c}}, \hat {L} _ {\mathrm {s r c}}\right) + \lambda_ {C} \mathcal {L} _ {C} \left(L _ {\mathrm {t r g}}, \hat {L} _ {C} ^ {\text {t r g}}\right). \tag {6} +$$ + +Finally, we minimize the image generator loss $\mathcal{L}_G^I$ , the discriminator loss $\mathcal{L}_D$ , and the illumination loss $\mathcal{L}_G^M$ together: + +$$ +\min _ {G, D} \mathcal {L} _ {G} ^ {I} + \mathcal {L} _ {D} + \lambda_ {G} ^ {M} \mathcal {L} _ {G} ^ {M}. \tag {7} +$$ + +# 4 Experiments + +In Sec. 4.1, we first discuss our data collection process, which was based on the approach used to compile the Light Stage at Your Desk (LSYD) dataset. In Sec. 4.2, we perform quantitative and qualitative comparisons with existing single-image portrait relighting algorithms and evaluate their temporal consistency. In Sec. 4.3, we demonstrate predicting, at a low-level, what the user is looking at on the monitor screen. Finally, in Sec. 4.4, we perform ablation studies, evaluating the network architecture's impact on relighting performance. + +![](images/3d129e60991e7581d7e9075533df2730fdc6e734021c7fe4e525a995406d229e.jpg) +Fig. 4. We perform a qualitative comparison with existing techniques [30,36] on the LSYD dataset. Source and target (images & lighting) were unseen during training. All models are personalized, i.e. trained on images of that individual only. We (Col. 3) produce significantly better results compared to existing approaches (Cols 4 and 5). + +# 4.1 Data + +We recorded data from 5 users of diverse ethnicities and genders to ensure a wide range of skin types. Each participant wore a variety of outfits, and we used 4 different ambient lighting conditions per person to mimic the conditions of real-life online meetings. We directed the participants to continuously change their facial expression and pose during the capture sessions. Each user's face was captured while watching 8 different videos, each 8 minutes long, on different days with varying appearances. We randomly hold out 1 video for testing and use the remaining 7 videos for training. We use this testing sequence only for qualitative evaluation, not for any quantitative metrics. This is because quantitative evaluation requires a pair of source and target images of the same person in the same pose but under different lighting conditions. This is difficult to obtain accurately for the aforementioned test video sequence since the participants naturally vary their pose and expression over the course of the video. Instead, we capture an additional test sequence, used only for numerical evaluation in + +![](images/7ae403f850a9bf0fb1465764a59d13c479f036723f1f9049c48e584246657292.jpg) +Fig. 5. We perform a qualitative comparison on the OLAT dataset [44]. Our approach outperforms [36,30] and can render strong directional lighting and specular highlights without any explicit modeling of geometry and reflectance. + +which the participant is captured in 9 different pose-expression combinations, each with a distinct monitor light. For each pose, we can create $\binom{9}{2} = 36$ source and target pairs as input and pseudo ground-truth, resulting in a total of 324 test data pairs per user. Additionally, we compared methods using Dynamic OLAT Dataset [44] with environment lighting maps [1,8]. During this comparison, we converted HDR environment maps to LDR and utilized a $270^{\circ}$ FoV as a monitor light to assess the relighting results. We want to note that all the test data is composed of unseen portraits and lightings for both LSYD and OLAT datasets. + +# 4.2 Comparison with existing approaches + +We employ three error metrics to assess relighting performance: RMSE, LPIPS [45], and DISTS [6]. LPIPS and DISTS are more robust to slight differences in pose between the relit image and the pseudo ground truth and detect perceptual differences more effectively than RMSE. + +**Portrait image relighting.** We compared our approach with existing portrait relighting neural architectures — Sun et al. [36] and Sengupta et al. [30] — by training on our captured LSYD dataset using the same pre-processing for all three architectures (see Sec. 3.1). Our training loss, given in Sec. 3.4, can handle misalignment in source-target pairs in training data, similar to the loss proposed in Sengupta et al. [30] (we use an additional loss on source monitor lighting prediction). For Sun et al. [36], we train both with their original loss + +Table 1. We train a personalized relighting model on 5 users from our LSYD dataset and 4 users from the OLAT dataset [44]. We evaluate these models using source & target portrait images, as well as lighting, that were not encountered during the training phase, and report average RMSE, LPIPS [46], and DISTS[6] scores both on LSYD and OLAT dataset. Our method can perform relighting without source lighting $L_{\mathrm{src}}$ , by simply using the predicted light source from our model as input lighting. Our method significantly outperforms Sun et al. [36] and Sengupta et al. [30]. + +
Known source lightingLSYD dataOLAT data [44]
LsrcLPIPS ↓DISTS ↓RMSE ↓LPIPS ↓DISTS ↓RMSE ↓
Sun et al. [36] w/ LSun-0.17120.16298.59580.22730.17456.2692
Sun et al. [36] w/ LOurs-0.10290.11528.44760.22670.15696.1898
Ours-0.08390.09538.32220.18120.13366.0931
Sengupta et al. [30]✓0.10180.11058.28260.22370.16756.1751
Ours✓0.08320.09538.19390.18090.13345.9548
+ +function $L_{\mathrm{Sun}}$ (which expects perfect source-target pose alignment obtained in OLAT data) and with our proposed loss function $L_{\mathrm{ours}}$ to specifically handle misalignment in LSYD data. We train personalized models on 5 users from the LSYD dataset and on 4 users from the publicly available Dynamic OLAT Dataset [44] with 2361 indoor environment lighting maps [1,8]. + +For our quantitative evaluation, we test our model on 1620 test images across 5 users with unseen appearance and lighting conditions on the LSYD dataset and on 7172 test images from the Dynamic OLAT dataset. We present the result in Tab. 1. We observe that our proposed approach outperforms Sengupta et al. [30] and Sun et al. [36] by $22.3\%$ and $23.6\%$ respectively on the LSYD dataset and by $23.5\%$ and $25.6\%$ on the OLAT dataset, when comparing LPIPS score. Our qualitative comparison, as presented in Fig. 4 and Fig. 5, shows that our model performs superior re + +![](images/829b6f38fc6832be1caaaef5fd422d37ecbc21eaa7971f639fa05e93b4f24e65.jpg) +Fig.6. Our method can also relight portrait images captured with unknown light sources. We relight a source image (Col. 1) with target light shown in the inset in Col. 2. We add a reference of how the face appears under that target light. + +lighting in terms of color, quality, and consistency, and can render strong directional lighting and specular highlights without any explicit modeling. + +Note that Sun et al. [36] does not require the source lighting $L_{\mathrm{src}}$ during test time. Our proposed approach can also perform relighting without prior knowledge of source lighting $L_{\mathrm{src}}$ by simply predicting $\hat{L}_{\mathrm{src}}$ and using it for light-conditioned feature normalization. We show that even in the absence of + +$L_{\mathrm{src}}$ , our method outperforms Sun et al. [36] by $22.6\%$ on the LSYD data and by $25.5\%$ on the OLAT data, in terms of LPIPS. + +In Fig. 6, we demonstrate that our approach can relight any portrait image captured 'in-the-wild' without requiring the knowledge of source lighting, and outperforms Sun et al. [36]. Our approach first predicts a proxy source monitor lighting, imagining the portrait image to be captured under a monitor lighting, and uses this predicted source lighting in the LCFN module to 'de-lit' the input image and relight it with a target lighting. + +In Fig. 7 1st-row, we show the results of relighting an input image with directional lights by moving specific bright areas on the monitor screen. Our method learns to render directional lighting effects and cast shadows as needed. To better illustrate the effect of moving lighting, we first relight the input image with no light reflected from the monitor (row-2 col-1), which produces a relief image under ambient room + +![](images/766dae490665a6cdf8c35774e60d851e62c027862619c7e6fc1c5c69f13aab8d.jpg) +Fig. 7. We relight the input image with directional lights (1st row) and without any target light ('Relight A'). We then subtract the 'Relight A' from directional images (row-2 col-2:5). Our method learns to render the effects of directional lighting and decouple ambient and dominant frontal lighting (from monitor). + +![](images/73735b90d02b709dfef115d3bca1386ca66be993dcfedc53fc48db8e666923a6.jpg) + +![](images/833ec48b41a5f287af20c0dbca34ef64febc2c348d0c5a2df04626e2c89da384.jpg) + +![](images/b26c96fe6105b580185f4398c11bfe3564db9ff5b0580854bc0f24c3ccbf2c12.jpg) + +![](images/42ebe8fb772bfb7547b07bc29344a51be6b13c15d1f84a9046870973fa7ef54e.jpg) + +lighting only. Subtracting this 'de-lit' image from the relit images under directional lighting (row-2 col 2-5) highlights the ability of our method to decouple ambient room illumination from dominant frontal lighting. However, due to the constraint that our lighting is limited to the illumination emitted from the monitor, we may not accurately depict extreme lighting effects. To this end, we show the strong directional light effects using the OLAT dataset in Fig. 5. + +Portrait video relighting. Next, we evaluate the temporal consistency of each portrait video relighting algorithm. For each user in the LSYD data, we relit the held-out test video with 50 different target lighting conditions, creating 50 relit videos. We then computed the RMSE between adjacent frames in relit videos as a measure of temporal consistency. Since the pose is almost identical between adjacent frames, lower RMSE error indicates temporally consistent relighting. We then report the average temporal RMSE across all such adjacent frames. In practice, + +Table 2. We evaluate temporal consistency by relighting a test video captured with varying source lighting with the same target lighting and calculating RMSE between adjacent frames. We then report the average RMSE across all adjacent frames and compute an error rate to indicate the percent of adjacent frames with RMSE higher than a threshold. + +
ThresholdRMSE ↓Error Rate (%)
>0.2>0.3>0.4
Sun et al. [36]5.8613.532.611.06
Sengupta et al. [30]6.3721.835.401.75
+Lsrc_avg5.7613.312.340.98
Ours (base)6.0116.223.611.08
+Lsrc_avg5.7313.092.310.83
+LCFN5.6813.042.280.71
+Lsrc_avg + LCFN5.5512.892.220.65
+ +however, a significant fraction of adjacent frame pairs have extremely similar lighting between the two frames, making their relit frames naturally consistent anyway. Only in a small percentage of adjacent frames does the source lighting significantly change, leading to obvious flickering in the relit video if temporal consistency is not maintained. Thus, in addition to average temporal RMSE, we also compute the error rate for three different thresholds: 0.2 (low), 0.3 (medium), and 0.4 (high), which indicate the percentage of adjacent frames where RMSE + +error is more than the threshold. + +In Tab. 2 and Fig. 8, we compare our approach with and without skip-connected feature smoothing to past works [36,30]. Note that this temporal smoothing of the skip-connected features can only be applied in our framework since we de-light encoder features from the source lighting with LCFN. Both for our approach and for Sengupta et al. [30], we can further apply smoothing of input source lighting to handle abrupt changes. We observe that our method produces the most temporally consistent relighting while also being the most accurate (see Tab. 1). We further note that both skip-connected feature smoothing and smoothing of source lighting improve temporal consistency. + +![](images/807de019c307e9311014a483b329950f97e590d5be83fe3c49ea9433a007255e.jpg) +Fig.8. We show temporal consistency between adjacent frames separated by 0.33s by relighting a test video captured with varying source lighting with the same target lighting. Note that [36,30] both exhibit abrupt changes in lighting between frames $t$ and $t + 10$ , while our approach produces a more stable result. + +# 4.3 Monitor prediction + +Our network can also be used to predict the source monitor lighting of any input image, as shown in Fig. 9. For the LSYD dataset, we do have images of the source monitor for every input image, so we can calculate monitor prediction accuracy numerically. We observe that our method slightly out + +performs [36] with a Mean Absolute Error of 0.15 vs 0.17, also qualitatively + +![](images/c359b5fab8d4a90308e8904d8d440b06fda78369c1aef82172ac4752626fb693.jpg) +Fig. 9. Monitor prediction comparison on LSYD data (Col. 1 and 2) and "in the wild" setting (Col. 3 and 4). The second row represents the monitor prediction by [36] and Ours. + +producing more meaningful visualizations. For portrait images, we do not have ground-truth source lighting, but visualizations show meaningful predictions. + +The ability to predict monitor lighting from images captured by our monitor has many implications. This technique can be used to detect deep fake avatars during live video calls by purposefully projecting specific images via screen sharing and observing if we can detect the same image from the webcam feed of other attendees in the call. If there is a mismatch between the projected monitor image and the predicted one, this likely indicates a deep fake avatar in the video call. + +# 4.4 Ablation studies + +We report the removal of various components from our relighting network in Tab. 3, specifically LCFN and source monitor lighting prediction using intermediate encoder features. We observe that both improve final relighting performance, which shows their effectiveness in disentangling source + +Table 3. Both LCFN and source monitor prediction $\mathbf{L}_{\mathrm{src}}$ improve relighting performance by effectively disentangling source lighting information from intrinsic appearance features. + +
Lsrcde-lightingRMSE ↓LPIPS ↓DISTS ↓
--8.42300.09640.1071
✓-8.25960.09150.1013
-✓8.19070.09040.0966
✓✓8.07460.08530.0963
+ +lighting information from intrinsic appearance features. + +# 5 Conclusion + +We propose a personalized video relighting algorithm that leverages casually captured LSYD data to generate real-time high-quality temporally consistent relit videos under any pose, expression, and lighting conditions. We present a novel network architecture that can perform high-quality relighting on both the LSYD and OLAT datasets, rendering challenging lighting conditions like directional lights, shadows, and specularities. We achieve this without using any 3D information or performing explicit decomposition, simply by achieving better image-based relighting enabled by our proposed neural architecture with LCFN module. Our method enables better lighting quality during live video calls and in portrait images, and produces better harmonization with virtual backgrounds. + +Limitation. Since we utilize only the front-facing monitor light as the source lighting, we do not account for scenarios where the light source is located on the sides $(90^{\circ})$ or behind $(180^{\circ})$ . However, it does not necessarily mean that our model cannot learn directional lighting effects, as demonstrated by training on the OLAT dataset which contains full $360^{\circ}$ lighting, see Fig. 5. + +Ethical considerations. While our primary goal is to allow people to improve their facial appearance with virtual relighting, we note that it is also a form of image manipulation and can be used for malicious purposes [3]. Furthermore, we emphasize that while currently we can only predict the user's direct line-of-sight monitor lighting in low resolution, the potential for high-resolution monitor prediction in the future could raise significant privacy concerns. + +# References + +1. Bolduc, C., Giroux, J., Hébert, M., Demers, C., Lalonde, J.F.: Beyond the pixel: a photometrically calibrated hdr dataset for luminance and color prediction (2023) +2. Chen, Z., Liu, Z.: Relighting4d: Neural relightable human from videos (2022) +3. Choi, J.M., Leung, J., Frahm, N., Christman, M., Bertasius, G., Sengupta, R.: Building secure and engaging video communication by using monitor illumination. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 4377-4386 (2024) +4. Daichi Tajima, Yoshihiro Kanamori, Y.E.: Relighting humans in the wild: Monocular full-body human relighting with domain adaptation. Computer Graphics Forum (Proc. of Pacific Graphics 2021) 40(7), 205-216 (2021) +5. Debevec, P., Wenger, A., Tchou, C., Gardner, A., Waese, J., Hawkins, T.: A lighting reproduction approach to live-action compositing. ACM Trans. Graph. 21(3), 547-556 (jul 2002). https://doi.org/10.1145/566654.566614, https://doi.org/10.1145/566654.566614 +6. Ding, K., Ma, K., Wang, S., Simoncelli, E.P.: Image quality assessment: Unifying structure and texture similarity. CoRR abs/2004.07728 (2020), https://arxiv.org/abs/2004.07728 +7. Futschik, D., Ritland, K., Vecore, J., Fanello, S., Orts-Escolano, S., Curless, B., Sykora, D., Pandey, R.: Controllable light diffusion for portraits (2023) +8. Gardner, M.A., Sunkavalli, K., Yumer, E., Shen, X., Gambaretto, E., Gagné, C., Lalonde, J.F.: Learning to predict indoor illumination from a single image (2017) +9. Gerstner, C.R., Farid, H.: Detecting real-time deep-fake videos using active illumination. In: 2022 IEEE/CVF Conference on Computer Vision and Pattern Recognition Workshops (CVPRW). pp. 53-60 (2022). https://doi.org/10.1109/CVPRW56347.2022.00015 +0. Guo, K., Lincoln, P., Davidson, P., Busch, J., Yu, X., Whalen, M., Harvey, G., Orts-Escalano, S., Pandey, R., Dourgarian, J., Tang, D., Tkach, A., Kowdle, A., Cooper, E., Dou, M., Fanello, S., Fyffe, G., Rhemann, C., Taylor, J., Debevec, P., Izadi, S.: The relightables: Volumetric performance capture of humans with realistic relighting (2019) +1. He, Y., Xing, Y., Zhang, T., Chen, Q.: Unsupervised portrait shadow removal via generative priors. In: Proceedings of the 29th ACM International Conference on Multimedia. MM '21, ACM (Oct 2021). https://doi.org/10.1145/3474085.3475663, http://dx.doi.org/10.1145/3474085.3475663 +2. Hou, A., Sarkis, M., Bi, N., Tong, Y., Liu, X.: Face relighting with geometrically consistent shadows. In: In Proceeding of IEEE Computer Vision and Pattern Recognition. New Orleans, LA (June 2022) +3. Hou, A., Zhang, Z., Sarkis, M., Bi, N., Tong, Y., Liu, X.: Towards high fidelity face relighting with realistic shadows. In: IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR) (2021) +4. Hu, Y., Wang, B., Lin, S.: Fc 4: Fully convolutional color constancy with confidence-weighted pooling. In: Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition. pp. 4085-4094 (2017) +5. Isola, P., Zhu, J.Y., Zhou, T., Efros, A.A.: Image-to-image translation with conditional adversarial networks. CVPR (2017) +6. Jin, H., Liao, S., Shao, L.: Pixel-in-pixel net: Towards efficient facial landmark detection in the wild. International Journal of Computer Vision (Sep 2021). https://doi.org/10.1007/s11263-021-01521-4, http://dx.doi.org/10.1007/s11263-021-01521-4 + +17. Karras, T., Laine, S., Aittala, M., Hellsten, J., Lehtinen, J., Aila, T.: Analyzing and improving the image quality of StyleGAN. In: Proc. CVPR (2020) +18. Lagunas, M., Sun, X., Yang, J., Villegas, R., Zhang, J., Shu, Z., Masia, B., Gutierrez, D.: Single-image full-body human relighting. CoRR abs/2107.07259 (2021), https://arxiv.org/abs/2107.07259 +19. LeGendre, C., Ma, W.C., Pandey, R., Fanello, S., Rhemann, C., Dourgarian, J., Busch, J., Debevec, P.: Learning illumination from diverse portraits (2020) +20. Liu*, Y., Hou*, A., Huang, X., Ren, L., Liu, X.: Blind removal of facial foreign shadows. In: In Proceedings of British Machine Vision Conference (BMVC). London, UK (November 2022) +21. Meka, A., Pandey, R., Haene, C., Orts-Escolano, S., Barnum, P., Davidson, P., Erickson, D., Zhang, Y., Taylor, J., Bouaziz, S., Legendre, C., Ma, W.C., Overbeck, R., Beeler, T., Debevec, P., Izadi, S., Theobalt, C., Rhemann, C., Fanello, S.: Deep relightable textures - volumetric performance capture with neural rendering. vol. 39 (December 2020). https://doi.org/10.1145/3414685.3417814, http://gvv.mpi-inf.mpg.de/projects/DeepRelightableTextures/ +22. Nestmeyer, T., Lalonde, J., Matthews, I., Lehrmann, A.: Learning physics-guided face relighting under directional light. In: 2020 IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR). pp. 5123-5132. IEEE Computer Society, Los Alamitos, CA, USA (jun 2020). https://doi.org/10.1109/CVPR42600.2020.00517, https://doi.ieeecomputersociety.org/10.1109/CVPR42600.2020.00517 +23. Pandey, R., Orts-Escolano, S., LeGendre, C., Haene, C., Bouaziz, S., Rhemann, C., Debevec, P., Fanello, S.: Total relighting: Learning to relight portraits for background replacement. vol. 40 (August 2021). https://doi.org/10.1145/3450626.3459872 +24. Peers, P., Tamura, N., Matusik, W., Debevec, P.: Post-production facial performance relighting using reflectance transfer. ACM Trans. Graph. 26(3), 52-es (jul 2007). https://doi.org/10.1145/1276377.1276442, https://doi.org/10.1145/1276377.1276442 +25. Ponglertnapakorn, P., Tritrong, N., Suwajanakorn, S.: Difareli: Diffusion face relighting (2023), https://arxiv.org/abs/2304.09479 +26. Qi, L., Wu, J., Wang, A.N., Wang, S., Sengupta, R.: My3dgen: A scalable personalized 3d generative model (2023) +27. Qiu, H., Chen, Z., Jiang, Y., Zhou, H., Fan, X., Yang, L., Wu, W., Liu, Z.: Relitalk: Relightable talking portrait generation from a single video (2023) +28. Ren, M., Xiong, W., Yoon, J.S., Shu, Z., Zhang, J., Jung, H., Gerig, G., Zhang, H.: Relightful harmonization: Lighting-aware portrait background replacement (2023) +29. Ronneberger, O., Fischer, P., Brox, T.: U-net: Convolutional networks for biomedical image segmentation. CoRR abs/1505.04597 (2015), http://arxiv.org/abs/1505.04597 +30. Sengupta, S., Curless, B., Kemelmacher-Shlizerman, I., Seitz, S.M.: A light stage on every desk. CoRR abs/2105.08051 (2021), https://arxiv.org/abs/2105.08051 +31. Sengupta, S., Kanazawa, A., Castillo, C.D., Jacobs, D.W.: Sfsnet: Learning shape, reflectance and illuminance of faces in the wild'. In: Proceedings of the IEEE conference on computer vision and pattern recognition. pp. 6296-6305 (2018) +32. Sevastopolsky, A., Ignatiev, S., Ferrer, G., Burnaev, E., Lempitsky, V.: Relightable 3d head portraits from a smartphone video (2020) +33. Shih, Y., Paris, S., Barnes, C., Freeman, W.T., Durand, F.: Style transfer for headshot portraits. ACM Trans. Graph. 33(4) (jul 2014). https://doi.org/10.1145/2601097.2601137, https://doi.org/10.1145/2601097.2601137 + +34. Shu, Z., Hadap, S., Shechtman, E., Sunkavalli, K., Paris, S., Samaras, D.: Portrait lighting transfer using a mass transport approach. ACM Trans. Graph. 37(1) (oct 2017). https://doi.org/10.1145/3095816, https://doi.org/10.1145/3095816 +35. Song, G., Cham, T.J., Cai, J., Zheng, J.: Half-body portrait relighting with overcomplete lighting representation (06 2021) +36. Sun, T., Barron, J.T., Tsai, Y., Xu, Z., Yu, X., Fyffe, G., Rhemann, C., Busch, J., Debevec, P.E., Ramamoorthi, R.: Single image portrait relighting. CoRR abs/1905.00824 (2019), http://arxiv.org/abs/1905.00824 +37. Sun, T., Lin, K.E., Bi, S., Xu, Z., Ramamoorthi, R.: Nelf: Neural light-transport field for portrait view synthesis and relighting (2021) +38. Sun, T., Xu, Z., Zhang, X., Fanello, S.R., Rhemann, C., Debevec, P.E., Tsai, Y., Barron, J.T., Ramamoorthi, R.: Light stage super-resolution: Continuous high-frequency relighting. CoRR abs/2010.08888 (2020), https://arxiv.org/abs/2010.08888 +39. Wang, Y., Holynski, A., Zhang, X., Zhang, X.: Sunstage: Portrait reconstruction and relighting using the sun as a light stage. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 20792-20802 (2023) +40. Wang, Z., Yu, X., Lu, M., Wang, Q., Qian, C., Xu, F.: Single image portrait relighting via explicit multiple reflectance channel modeling. ACM Trans. Graph. 39(6) (Nov 2020). https://doi.org/10.1145/3414685.3417824 +41. Yang, H., Zheng, M., Feng, W., Huang, H., Lai, Y.K., Wan, P., Wang, Z., Ma, C.: Towards practical capture of high-fidelity relightable avatars (2023) +42. Yeh, Y.Y., Nagano, K., Khamis, S., Kautz, J., Liu, M.Y., Wang, T.C.: 41(6), 1-21 (nov 2022). https://doi.org/10.1145/3550454.3555442, https://doi.org/10. 1145 +43. Yeh, Y.Y., Nagano, K., Khamis, S., Kautz, J., Liu, M.Y., Wang, T.C.: Learning to relight portrait images via a virtual light stage and synthetic-to-real adaptation. ACM Transactions on Graphics (TOG) (2022) +44. Zhang, L., Zhang, Q., Wu, M., Yu, J., Xu, L.: Neural video portrait relighting in real-time via consistency modeling (2021) +45. Zhang, R., Isola, P., Efros, A.A., Shechtman, E., Wang, O.: The unreasonable effectiveness of deep features as a perceptual metric. In: 2018 IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR). pp. 586-595. IEEE Computer Society, Los Alamitos, CA, USA (jun 2018). https://doi.org/10.1109/CVPR.2018.00068, https://doi.ieeecomputersociety.org/10.1109/CVPR.2018.00068 +46. Zhang, R., Isola, P., Efros, A.A., Shechtman, E., Wang, O.: The unreasonable effectiveness of deep features as a perceptual metric. CoRR abs/1801.03924 (2018), http://arxiv.org/abs/1801.03924 +47. Zhang, X., Fanello, S.R., Tsai, Y., Sun, T., Xue, T., Pandey, R., Orts-Escalano, S., Davidson, P.L., Rhemann, C., Debevec, P.E., Barron, J.T., Ramamoorthi, R., Freeman, W.T.: Neural light transport for relighting and view synthesis. CoRR abs/2008.03806 (2020), https://arxiv.org/abs/2008.03806 +48. Zhang, X., Barron, J.T., Tsai, Y.T., Pandey, R., Zhang, X., Ng, R., Jacobs, D.E.: Portrait shadow manipulation. vol. 39 (2020) +49. Zhou, H., Hadap, S., Sunkavalli, K., Jacobs, D.: Deep single-image portrait relighting. In: 2019 IEEE/CVF International Conference on Computer Vision (ICCV). pp. 7193-7201 (2019). https://doi.org/10.1109/ICCV.2019.00729 +50. Zhu, J.Y., Park, T., Isola, P., Efros, A.A.: Unpaired image-to-image translation using cycle-consistent adversarial networks. In: Computer Vision (ICCV), 2017 IEEE International Conference on (2017) \ No newline at end of file diff --git a/2024/Personalized Video Relighting With an At-Home Light Stage/images.zip b/2024/Personalized Video Relighting With an At-Home Light Stage/images.zip new file mode 100644 index 0000000000000000000000000000000000000000..ac4e7793b7a66bbb982f4ffae2a349dbd9b6e645 --- /dev/null +++ b/2024/Personalized Video Relighting With an At-Home Light Stage/images.zip @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:d2b0af3ac2672af4aa90f307337ecb61d1539f98b32bb22ca606de1859f1d7b7 +size 577832 diff --git a/2024/Personalized Video Relighting With an At-Home Light Stage/layout.json b/2024/Personalized Video Relighting With an At-Home Light Stage/layout.json new file mode 100644 index 0000000000000000000000000000000000000000..2567c80d58cb1521b99ea7e60c3c4627aa060c55 --- /dev/null +++ b/2024/Personalized Video Relighting With an At-Home Light Stage/layout.json @@ -0,0 +1,8724 @@ +{ + "pdf_info": [ + { + "para_blocks": [ + { + "bbox": [ + 133, + 110, + 481, + 148 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 110, + 481, + 148 + ], + "spans": [ + { + "bbox": [ + 133, + 110, + 481, + 148 + ], + "type": "text", + "content": "Personalized Video Relighting With an At-Home Light Stage" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 173, + 168, + 438, + 180 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 173, + 168, + 438, + 180 + ], + "spans": [ + { + "bbox": [ + 173, + 168, + 438, + 180 + ], + "type": "text", + "content": "Jun Myeong Choi, Max Christman, and Roni Sengupta" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 182, + 190, + 430, + 213 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 182, + 190, + 430, + 213 + ], + "spans": [ + { + "bbox": [ + 182, + 190, + 430, + 213 + ], + "type": "text", + "content": "University of North Carolina at Chapel Hill, NC 27514, USA {chedgekr,mrlc,ronisen}@cs.unc.edu" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 160, + 243, + 452, + 452 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 160, + 243, + 452, + 452 + ], + "spans": [ + { + "bbox": [ + 160, + 243, + 452, + 452 + ], + "type": "text", + "content": "Abstract. In this paper, we develop a personalized video relighting algorithm that produces high-quality and temporally consistent relit videos under any pose, expression, and lighting condition in real-time. Existing relighting algorithms typically rely either on publicly available synthetic data, which yields poor relighting results, or on actual light stage data which is difficult to acquire. We show that by just capturing recordings of a user watching YouTube videos on a monitor we can train a personalized algorithm capable of performing high-quality relighting under any condition. Our key contribution is a novel image-based neural relighting architecture that effectively separates the intrinsic appearance features - the geometry and reflectance of the face - from the source lighting and then combines them with the target lighting to generate a relit image. This neural architecture enables smoothing of intrinsic appearance features leading to temporally stable video relighting. Both qualitative and quantitative evaluations show that our architecture improves portrait image relighting quality and temporal consistency over state-of-the-art approaches on both casually captured 'Light Stage at Your Desk' (LSYD) and light-stage-captured 'One Light At a Time' (OLAT) datasets. Source code is available at https://github.com/chedgekorea/relighting" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 160, + 463, + 411, + 474 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 160, + 463, + 411, + 474 + ], + "spans": [ + { + "bbox": [ + 160, + 463, + 411, + 474 + ], + "type": "text", + "content": "Keywords: Image-based Relighting " + }, + { + "bbox": [ + 160, + 463, + 411, + 474 + ], + "type": "inline_equation", + "content": "\\cdot" + }, + { + "bbox": [ + 160, + 463, + 411, + 474 + ], + "type": "text", + "content": " 'At Home' Light Stage" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 133, + 495, + 230, + 508 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 495, + 230, + 508 + ], + "spans": [ + { + "bbox": [ + 133, + 495, + 230, + 508 + ], + "type": "text", + "content": "1 Introduction" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 130, + 521, + 481, + 593 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 521, + 481, + 593 + ], + "spans": [ + { + "bbox": [ + 130, + 521, + 481, + 593 + ], + "type": "text", + "content": "With the recent rise in popularity of video conferencing for business, educational, and personal activities, there is a significant demand for improving facial lighting. Virtually relighting our images and videos helps us improve the appearance of our faces without requiring explicit studio-quality lighting in a dedicated space or any specialized lighting expertise. Recent advances in deep neural networks have renewed interest in the problem of virtual relighting." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 130, + 594, + 481, + 666 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 594, + 481, + 666 + ], + "spans": [ + { + "bbox": [ + 130, + 594, + 481, + 666 + ], + "type": "text", + "content": "Training a deep neural network for relighting requires extensive training data that includes source images paired with relit target images. One way of acquiring this data is by using a large spherical rig with numerous lights and cameras, known as a light stage [5]. While light stage data has been shown to produce high-quality relighting results [36,44,23,38,22,48], the limited availability of datasets, trained models, and access to the light stage itself has impeded further research." + } + ] + } + ], + "index": 7 + } + ], + "discarded_blocks": [], + "page_size": [ + 612, + 792 + ], + "page_idx": 0 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 133, + 114, + 192, + 173 + ], + "blocks": [ + { + "bbox": [ + 133, + 114, + 192, + 173 + ], + "lines": [ + { + "bbox": [ + 133, + 114, + 192, + 173 + ], + "spans": [ + { + "bbox": [ + 133, + 114, + 192, + 173 + ], + "type": "image", + "image_path": "1951539ed1a5559b57e7ce4e301eaaefd57d9678d1ca1e0db360061607e81ca4.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 149, + 174, + 175, + 182 + ], + "lines": [ + { + "bbox": [ + 149, + 174, + 175, + 182 + ], + "spans": [ + { + "bbox": [ + 149, + 174, + 175, + 182 + ], + "type": "text", + "content": "(a) LSYD" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_caption" + } + ], + "index": 2 + }, + { + "type": "image", + "bbox": [ + 194, + 114, + 260, + 173 + ], + "blocks": [ + { + "bbox": [ + 194, + 114, + 260, + 173 + ], + "lines": [ + { + "bbox": [ + 194, + 114, + 260, + 173 + ], + "spans": [ + { + "bbox": [ + 194, + 114, + 260, + 173 + ], + "type": "image", + "image_path": "79068ed969abfda3cc3af4a60fde3d37c4d59d765d1610e5f6b912f45b7224e3.jpg" + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 206, + 174, + 246, + 182 + ], + "lines": [ + { + "bbox": [ + 206, + 174, + 246, + 182 + ], + "spans": [ + { + "bbox": [ + 206, + 174, + 246, + 182 + ], + "type": "text", + "content": "(b) Input Image" + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 254, + 174, + 286, + 182 + ], + "lines": [ + { + "bbox": [ + 254, + 174, + 286, + 182 + ], + "spans": [ + { + "bbox": [ + 254, + 174, + 286, + 182 + ], + "type": "text", + "content": "(c) Relit-Net" + } + ] + } + ], + "index": 7, + "angle": 0, + "type": "image_caption" + } + ], + "index": 4 + }, + { + "type": "image", + "bbox": [ + 263, + 114, + 348, + 172 + ], + "blocks": [ + { + "bbox": [ + 263, + 114, + 348, + 172 + ], + "lines": [ + { + "bbox": [ + 263, + 114, + 348, + 172 + ], + "spans": [ + { + "bbox": [ + 263, + 114, + 348, + 172 + ], + "type": "image", + "image_path": "e5ed7a9ff0b984b3fc910c98baafa9bbed2a936230c25cfc8842529b713f165d.jpg" + } + ] + } + ], + "index": 6, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 130, + 194, + 482, + 261 + ], + "lines": [ + { + "bbox": [ + 130, + 194, + 482, + 261 + ], + "spans": [ + { + "bbox": [ + 130, + 194, + 482, + 261 + ], + "type": "text", + "content": "Fig. 1. We learn a personalized relighting algorithm that generates temporally consistent and high-quality portrait videos under different lighting. We train the network using recordings of users watching YouTube on a monitor, thereby creating a Light Stage at Your Desk (LSYD). We project a portion of the LDR environment map with a " + }, + { + "bbox": [ + 130, + 194, + 482, + 261 + ], + "type": "inline_equation", + "content": "180^{\\circ}" + }, + { + "bbox": [ + 130, + 194, + 482, + 261 + ], + "type": "text", + "content": " FoV as the monitor light, while a portion of the remaining " + }, + { + "bbox": [ + 130, + 194, + 482, + 261 + ], + "type": "inline_equation", + "content": "90^{\\circ}" + }, + { + "bbox": [ + 130, + 194, + 482, + 261 + ], + "type": "text", + "content": " FoV is mapped as the background. We can achieve a harmonization effect with the virtual background." + } + ] + } + ], + "index": 11, + "angle": 0, + "type": "image_caption" + } + ], + "index": 6 + }, + { + "type": "image", + "bbox": [ + 349, + 114, + 413, + 172 + ], + "blocks": [ + { + "bbox": [ + 349, + 114, + 413, + 172 + ], + "lines": [ + { + "bbox": [ + 349, + 114, + 413, + 172 + ], + "spans": [ + { + "bbox": [ + 349, + 114, + 413, + 172 + ], + "type": "image", + "image_path": "498cf8b11c53efdcb57fd436e2dac13f0e565b5e8446a7fa2bfdc4ac4ab73ee0.jpg" + } + ] + } + ], + "index": 8, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 367, + 174, + 394, + 182 + ], + "lines": [ + { + "bbox": [ + 367, + 174, + 394, + 182 + ], + "spans": [ + { + "bbox": [ + 367, + 174, + 394, + 182 + ], + "type": "text", + "content": "(d) Output" + } + ] + } + ], + "index": 9, + "angle": 0, + "type": "image_caption" + } + ], + "index": 8 + }, + { + "type": "image", + "bbox": [ + 414, + 114, + 480, + 172 + ], + "blocks": [ + { + "bbox": [ + 414, + 114, + 480, + 172 + ], + "lines": [ + { + "bbox": [ + 414, + 114, + 480, + 172 + ], + "spans": [ + { + "bbox": [ + 414, + 114, + 480, + 172 + ], + "type": "image", + "image_path": "b92e180af5d3093e1ef4b72359ac9bb55874352b69ba10f3d37ad269ef4190ab.jpg" + } + ] + } + ], + "index": 10, + "angle": 0, + "type": "image_body" + } + ], + "index": 10 + }, + { + "bbox": [ + 130, + 282, + 482, + 342 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 282, + 482, + 342 + ], + "spans": [ + { + "bbox": [ + 130, + 282, + 482, + 342 + ], + "type": "text", + "content": "For example, One Light At a Time (Dynamic-OLAT) [44] is the only publicly available light stage relighting dataset consisting of four individuals only. As a result, researchers have often turned to synthetic data to train their relighting algorithms [31,49,13,35]. Unfortunately, existing synthetic data compromises the quality of relit images." + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 130, + 342, + 482, + 474 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 342, + 482, + 474 + ], + "spans": [ + { + "bbox": [ + 130, + 342, + 482, + 474 + ], + "type": "text", + "content": "We draw inspiration from recent work by Sengupta et al. [30,26] and develop a personalized relighting model by capturing a single user's appearance while being lit by a computer monitor. However, the results and applications of Sengupta et al. [30] have several limitations. First, it requires capturing users with fixed poses and expressions, an unrealistic requirement for actual users that impedes casual capture. Second, the relighting algorithm requires a dark room with negligible ambient light, limiting the environment in which this can be applied. Third, the resulting relit video is temporally unstable, exhibiting significant flickering artifacts, making it unsuitable for relighting Zoom calls. Lastly, it requires knowledge of source lighting and thus is unable to relight arbitrary portrait images captured in the wild." + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 130, + 474, + 482, + 666 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 474, + 482, + 666 + ], + "spans": [ + { + "bbox": [ + 130, + 474, + 482, + 666 + ], + "type": "text", + "content": "In this paper, we show that casually captured light stage data is sufficient to develop a high-quality temporally consistent video portrait relighting algorithm that works under arbitrary conditions (i.e. pose, expression, and ambient lighting) in real-time (" + }, + { + "bbox": [ + 130, + 474, + 482, + 666 + ], + "type": "inline_equation", + "content": "\\sim 45" + }, + { + "bbox": [ + 130, + 474, + 482, + 666 + ], + "type": "text", + "content": " fps). To that end, we create our own casually captured light stage dataset with varying pose, expression, and lighting, called Light Stage at Your Desk (LSYD). Our key contribution is a neural image-based relighting architecture, based on the commonly used U-Net [36,49,22,40,38], that better disentangles the source lighting from the user's intrinsic facial appearance (shape and reflectance) and then adds back the target lighting to generate a relit image. Existing image-based relighting architectures [36,30] fail to accurately separate source lighting information from intrinsic appearance features in the encoder, leading to inconsistent and temporally unstable video relighting. To this end, we introduce the light-conditioned feature normalization (LCFN) module, which performs relighting and also predicts the source lighting from an input image. The LCFN module also enables temporal stability by performing exponential smoothing of de-lit intrinsic appearance features and facilitates relighting" + } + ] + } + ], + "index": 14 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "spans": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "type": "text", + "content": "2" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 166, + 91, + 220, + 100 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 166, + 91, + 220, + 100 + ], + "spans": [ + { + "bbox": [ + 166, + 91, + 220, + 100 + ], + "type": "text", + "content": "J. Choi et al." + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 1 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 132, + 112, + 231, + 199 + ], + "blocks": [ + { + "bbox": [ + 132, + 112, + 231, + 199 + ], + "lines": [ + { + "bbox": [ + 132, + 112, + 231, + 199 + ], + "spans": [ + { + "bbox": [ + 132, + 112, + 231, + 199 + ], + "type": "image", + "image_path": "d784e968271e27145566bd56bcd0d2a19d76403025d4ab61a5c09277c144379d.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 130, + 213, + 482, + 291 + ], + "lines": [ + { + "bbox": [ + 130, + 213, + 482, + 291 + ], + "spans": [ + { + "bbox": [ + 130, + 213, + 482, + 291 + ], + "type": "text", + "content": "Fig. 2. We highlight the key structural differences between our relighting architecture and that of [36,30]. Our approach removes source lighting information from input image features and only propagates intrinsic appearance (geometry and reflectance) features from the encoder to decoder, which results in better relighting quality and more temporal stability. In contrast [36,30] propagates entire image features from the encoder to the decoder without 'de-light', and expects the decoder to remove source lighting and add target lighting information." + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "image_caption" + } + ], + "index": 2 + }, + { + "type": "image", + "bbox": [ + 238, + 112, + 355, + 200 + ], + "blocks": [ + { + "bbox": [ + 238, + 112, + 355, + 200 + ], + "lines": [ + { + "bbox": [ + 238, + 112, + 355, + 200 + ], + "spans": [ + { + "bbox": [ + 238, + 112, + 355, + 200 + ], + "type": "image", + "image_path": "1ced49ce92679efc4604fff228850a52d0086db1ef5eb137770013b58345d242.jpg" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_body" + } + ], + "index": 3 + }, + { + "type": "image", + "bbox": [ + 362, + 112, + 482, + 199 + ], + "blocks": [ + { + "bbox": [ + 362, + 112, + 482, + 199 + ], + "lines": [ + { + "bbox": [ + 362, + 112, + 482, + 199 + ], + "spans": [ + { + "bbox": [ + 362, + 112, + 482, + 199 + ], + "type": "image", + "image_path": "630e0bb2e4047e9f8bb240ca08bcb1f2163bdc72ee37c2f48a3fc373b40cbf58.jpg" + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "image_body" + } + ], + "index": 4 + }, + { + "bbox": [ + 130, + 314, + 482, + 350 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 314, + 482, + 350 + ], + "spans": [ + { + "bbox": [ + 130, + 314, + 482, + 350 + ], + "type": "text", + "content": "of any arbitrary portrait image with unknown source lighting. We also improve the data pre-processing pipeline from Sengupta et al. [30] to make the relighting algorithm more robust to pose, expression, and ambient lighting conditions." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 130, + 351, + 482, + 530 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 351, + 482, + 530 + ], + "spans": [ + { + "bbox": [ + 130, + 351, + 482, + 530 + ], + "type": "text", + "content": "We compare our relighting network with two other algorithms: Sun et al. [36], which was originally trained on light stage data (OLAT), and Sengupta et al. [30], which was originally trained on casually captured data (albeit with fixed pose, expression, and no ambient lighting). For a fair comparison, we train all algorithms for personalized relighting using the same data pre-processing steps and loss functions on 5 individuals from our LSYD dataset and 4 individuals from OLAT [44]. Our network outperforms Sun et al. [36] and Sengupta et al. [30] by " + }, + { + "bbox": [ + 130, + 351, + 482, + 530 + ], + "type": "inline_equation", + "content": "22.3\\%" + }, + { + "bbox": [ + 130, + 351, + 482, + 530 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 130, + 351, + 482, + 530 + ], + "type": "inline_equation", + "content": "23.6\\%" + }, + { + "bbox": [ + 130, + 351, + 482, + 530 + ], + "type": "text", + "content": " respectively on the LSYD dataset and by " + }, + { + "bbox": [ + 130, + 351, + 482, + 530 + ], + "type": "inline_equation", + "content": "23.5\\%" + }, + { + "bbox": [ + 130, + 351, + 482, + 530 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 130, + 351, + 482, + 530 + ], + "type": "inline_equation", + "content": "25.6\\%" + }, + { + "bbox": [ + 130, + 351, + 482, + 530 + ], + "type": "text", + "content": " on the OLAT dataset, in terms of LPIPS. Qualitatively our method produces superior relighting in terms of color and quality. We further show that our approach is more temporally consistent, leading to less flickering than Sun et al. [36] or Sengupta et al. [30]. Detailed ablation studies show that LCFN and source monitor prediction improves relighting quality, feature, and source monitor smoothing improves temporal consistency, and data pre-processing improves robustness to pose and expression." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 132, + 531, + 367, + 542 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 531, + 367, + 542 + ], + "spans": [ + { + "bbox": [ + 132, + 531, + 367, + 542 + ], + "type": "text", + "content": "To summarize, our main contributions are as follows:" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 130, + 544, + 482, + 665 + ], + "type": "list", + "angle": 0, + "index": 12, + "blocks": [ + { + "bbox": [ + 130, + 544, + 482, + 581 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 544, + 482, + 581 + ], + "spans": [ + { + "bbox": [ + 130, + 544, + 482, + 581 + ], + "type": "text", + "content": "- We show that casually captured Light Stage at Your Desk (LSYD) data can be used to build a high-quality temporally consistent personalized video relighting algorithm without requiring access to an expensive light stage setup." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 130, + 581, + 482, + 616 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 581, + 482, + 616 + ], + "spans": [ + { + "bbox": [ + 130, + 581, + 482, + 616 + ], + "type": "text", + "content": "- We introduce a novel video relighting architecture that separates the source lighting from the user's intrinsic appearance features and then adds back the target lighting, leading to improved relighting and temporal consistency for videos." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 130, + 617, + 482, + 665 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 617, + 482, + 665 + ], + "spans": [ + { + "bbox": [ + 130, + 617, + 482, + 665 + ], + "type": "text", + "content": "- While our relighting network focuses on 'at home' Light Stage (LSYD dataset) outperforming state-of-the-art algorithms, we also perform equally well on actual Light Stage captured OLAT [44] datasets, and any arbitrary portrait image captured 'in the wild'." + } + ] + } + ], + "index": 11 + } + ], + "sub_type": "text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 198, + 91, + 448, + 102 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 198, + 91, + 448, + 102 + ], + "spans": [ + { + "bbox": [ + 198, + 91, + 448, + 102 + ], + "type": "text", + "content": "Personalized Video Relighting With an At-Home Light Stage" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 474, + 92, + 480, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 474, + 92, + 480, + 100 + ], + "spans": [ + { + "bbox": [ + 474, + 92, + 480, + 100 + ], + "type": "text", + "content": "3" + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 2 + }, + { + "para_blocks": [ + { + "bbox": [ + 132, + 114, + 237, + 127 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 114, + 237, + 127 + ], + "spans": [ + { + "bbox": [ + 132, + 114, + 237, + 127 + ], + "type": "text", + "content": "2 Related Work" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 130, + 139, + 481, + 210 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 139, + 481, + 210 + ], + "spans": [ + { + "bbox": [ + 130, + 139, + 481, + 210 + ], + "type": "text", + "content": "Portrait relighting methods change the appearance of the face to match a target lighting condition. This can be expressed through lighting parameters (e.g., an environment map, spherical harmonics, directional lighting, etc.) or through a reference image of another person. Our approach relights a portrait image to a lighting condition expressed through a low dynamic range (LDR) image representing the image on the monitor." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 130, + 211, + 482, + 450 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 211, + 482, + 450 + ], + "spans": [ + { + "bbox": [ + 130, + 211, + 482, + 450 + ], + "type": "text", + "content": "Image based relighting. Before the rise of deep learning, attempts were made at non deep learning image based relighting [34,24,33]. Shu et al. [34] introduced a face relighting approach that uses a mass-transport formulation for the transfer of illumination between images. Peers et al. [24] demonstrated a method for relighting portrait images with flat lighting to match specific target environments, incorporating a reference subject database for approximation. Shih et al. [33] adopted a multiscale technique to transfer local image statistics from reference portraits onto new ones, facilitating the matching of attributes like local contrast and overall lighting direction. Recent advancements in deep learning have caused significant shifts to the landscape of portrait relighting [25,11,20]. Research on ratio images for relighting [49,13] was conducted, utilizing public datasets and employing methods based on ratio images. However, this is limited to synthetic data, resulting in a significant domain gap with real data. The widespread application of light stages [5] in gathering data has enabled numerous groundbreaking research endeavors[36,44,38,28]. However, these methods rely on capturing data with a light stage, which are expensive and inaccessible. Instead, our approach builds a personalized relighting algorithm using casually captured videos from the desk recording setup introduced in Sengupta et al. [30]. In contrast to Sengupta et al. [30], we can collect data during daily computer usage by minimizing numerous constraints, eliminating the need for specific efforts in data collection." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 130, + 450, + 483, + 666 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 450, + 483, + 666 + ], + "spans": [ + { + "bbox": [ + 130, + 450, + 483, + 666 + ], + "type": "text", + "content": "Relighting with explicit decomposition. Creating a virtual relighting dataset through the utilization of synthetic human models to train their networks has been carried out in some studies [31,18,42]. However, when using synthetic data to train a neural network, the large domain gap between synthetic and real data impacts the model's performance on real data. In contrast, others [12,4,27,2,37] generate relit images by using public datasets and employ methods based on 3D model rendering. Specifically, Hou et al. [12] takes a more advanced approach by introducing explicit components, where rays originating from the face intersect with other parts of the facial geometry to create relit images. Moreover, the extensive use of light stages [5] has enabled numerous innovative studies [22,40,23,10,43,7,21,47,38,41,19] in this domain. Some researchers have incorporated explicit elements such as albedo, normals, specular maps, and diffuse maps into their methodologies [22,40,23,7]. Others have taken a physics-based rendering approach [10,43] to resolve these issues. Other papers [21,47] aim to manipulate lighting conditions and generate images under different lighting scenarios using texture information. However, explicit decomposition methods require ground truth (GT) data for these intrinsic components to train, either from synthetic data or real light stage data. Obtaining GT data for 'at-home'" + } + ] + } + ], + "index": 5 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 133, + 91, + 141, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 91, + 141, + 100 + ], + "spans": [ + { + "bbox": [ + 133, + 91, + 141, + 100 + ], + "type": "text", + "content": "4" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 166, + 91, + 220, + 101 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 166, + 91, + 220, + 101 + ], + "spans": [ + { + "bbox": [ + 166, + 91, + 220, + 101 + ], + "type": "text", + "content": "J. Choi et al." + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 3 + }, + { + "para_blocks": [ + { + "bbox": [ + 130, + 116, + 479, + 200 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 116, + 479, + 200 + ], + "spans": [ + { + "bbox": [ + 130, + 116, + 479, + 200 + ], + "type": "text", + "content": "captures is impossible, thus we focus on image-based relighting. In contrast, recent studies aim to streamline the capture process, using a mobile phone camera [32] or a sun stage [39] instead of a light stage. Nevertheless, due to their reliance on per-scene optimization, both of these papers lack the capability for real-time relighting and to generalize to unseen appearances of the individual and are only limited to the particular capture. Instead, our approach performs image-based relighting enabling real-time temporally consistent video relighting, at " + }, + { + "bbox": [ + 130, + 116, + 479, + 200 + ], + "type": "inline_equation", + "content": "\\sim 45" + }, + { + "bbox": [ + 130, + 116, + 479, + 200 + ], + "type": "text", + "content": " fps." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 132, + 217, + 202, + 229 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 217, + 202, + 229 + ], + "spans": [ + { + "bbox": [ + 132, + 217, + 202, + 229 + ], + "type": "text", + "content": "3 Method" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 130, + 241, + 480, + 337 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 241, + 480, + 337 + ], + "spans": [ + { + "bbox": [ + 130, + 241, + 480, + 337 + ], + "type": "text", + "content": "Our setup is similar to Gerstner et al. [9] and Sengupta et al. [30], where a user's face is captured while illuminated by their monitor. By capturing multiple videos of the user's face along with the video on their monitor, we build our 'at home' light stage dataset. We then use these data to train a personalized portrait relighting algorithm that can render the user's face under arbitrary lighting conditions. Specifically, given a portrait image " + }, + { + "bbox": [ + 130, + 241, + 480, + 337 + ], + "type": "inline_equation", + "content": "I_{\\mathrm{src}}" + }, + { + "bbox": [ + 130, + 241, + 480, + 337 + ], + "type": "text", + "content": ", corresponding source monitor lighting " + }, + { + "bbox": [ + 130, + 241, + 480, + 337 + ], + "type": "inline_equation", + "content": "L_{\\mathrm{src}}" + }, + { + "bbox": [ + 130, + 241, + 480, + 337 + ], + "type": "text", + "content": ", and target monitor lighting " + }, + { + "bbox": [ + 130, + 241, + 480, + 337 + ], + "type": "inline_equation", + "content": "L_{\\mathrm{trg}}" + }, + { + "bbox": [ + 130, + 241, + 480, + 337 + ], + "type": "text", + "content": ", our aim is to learn a function " + }, + { + "bbox": [ + 130, + 241, + 480, + 337 + ], + "type": "inline_equation", + "content": "G" + }, + { + "bbox": [ + 130, + 241, + 480, + 337 + ], + "type": "text", + "content": " that relights " + }, + { + "bbox": [ + 130, + 241, + 480, + 337 + ], + "type": "inline_equation", + "content": "I_{\\mathrm{src}}" + }, + { + "bbox": [ + 130, + 241, + 480, + 337 + ], + "type": "text", + "content": " under " + }, + { + "bbox": [ + 130, + 241, + 480, + 337 + ], + "type": "inline_equation", + "content": "L_{\\mathrm{trg}}" + }, + { + "bbox": [ + 130, + 241, + 480, + 337 + ], + "type": "text", + "content": ":" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 236, + 346, + 480, + 360 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 236, + 346, + 480, + 360 + ], + "spans": [ + { + "bbox": [ + 236, + 346, + 480, + 360 + ], + "type": "interline_equation", + "content": "\\hat {I} _ {\\mathrm {t r g}}, \\hat {L} _ {\\mathrm {s r c}} = G \\left(I _ {\\mathrm {s r c}}, L _ {\\mathrm {s r c}}, L _ {\\mathrm {t r g}}; \\theta_ {G}\\right). \\tag {1}", + "image_path": "d28de1d2651bc3f917561977cf4896fe5ee5de302063237ca10a1433fa6d6cef.jpg" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 130, + 367, + 479, + 403 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 367, + 479, + 403 + ], + "spans": [ + { + "bbox": [ + 130, + 367, + 479, + 403 + ], + "type": "text", + "content": "Note that our formulation can be used for scenarios where the source lighting is unknown by simply replacing the input source lighting with the predicted source lighting, unlike previous approaches [30]." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 130, + 403, + 480, + 510 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 403, + 480, + 510 + ], + "spans": [ + { + "bbox": [ + 130, + 403, + 480, + 510 + ], + "type": "text", + "content": "In the following sections, we outline our methodology for portrait video relighting using a monitor as a light stage. Sec. 3.1 outlines strategies for constructing training data pairs from casually captured videos that allow flexibility in facial expression, pose, and ambient lighting. In Sec. 3.2, we introduce our relighting network architecture that disentangles lighting from intrinsic appearance using light-conditioned feature normalization, leading to high-quality relit images. In Sec. 3.3, we propose additional techniques which enforce temporal consistency and eliminate flickering, also using LCFN. Finally, in Sec. 3.4, we discuss how to train our relighting network." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 132, + 527, + 324, + 540 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 527, + 324, + 540 + ], + "spans": [ + { + "bbox": [ + 132, + 527, + 324, + 540 + ], + "type": "text", + "content": "3.1 Constructing training data pairs" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 130, + 545, + 481, + 666 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 545, + 481, + 666 + ], + "spans": [ + { + "bbox": [ + 130, + 545, + 481, + 666 + ], + "type": "text", + "content": "While past work [30] imposed requirements of a neutral pose, expression, and dark room, we loosen these constraints to allow subjects in any conditions. The only constraint we maintain is that the room lighting shall not overpower the light emitted from the monitor. For example, if the capture occurs in front of a window with bright sunlight, the light from the monitor will have minimal impact on the face. As in Sengupta et al. [30], we aim to generate source and monitor image pairs " + }, + { + "bbox": [ + 130, + 545, + 481, + 666 + ], + "type": "inline_equation", + "content": "(I_{\\mathrm{src}}, L_{\\mathrm{src}})" + }, + { + "bbox": [ + 130, + 545, + 481, + 666 + ], + "type": "text", + "content": ", as well as target image and target monitor pairs " + }, + { + "bbox": [ + 130, + 545, + 481, + 666 + ], + "type": "inline_equation", + "content": "(I_{\\mathrm{trg}}, L_{\\mathrm{trg}})" + }, + { + "bbox": [ + 130, + 545, + 481, + 666 + ], + "type": "text", + "content": ", such that we can train our network to produce " + }, + { + "bbox": [ + 130, + 545, + 481, + 666 + ], + "type": "inline_equation", + "content": "\\tilde{I}_{\\mathrm{trg}}" + }, + { + "bbox": [ + 130, + 545, + 481, + 666 + ], + "type": "text", + "content": " where " + }, + { + "bbox": [ + 130, + 545, + 481, + 666 + ], + "type": "inline_equation", + "content": "I_{\\mathrm{trg}}" + }, + { + "bbox": [ + 130, + 545, + 481, + 666 + ], + "type": "text", + "content": " is the pseudo ground truth. However, due to unrestricted subject movement during data collection, there is a lack of pixel-aligned data, making random pairs" + } + ] + } + ], + "index": 9 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 198, + 91, + 448, + 102 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 198, + 91, + 448, + 102 + ], + "spans": [ + { + "bbox": [ + 198, + 91, + 448, + 102 + ], + "type": "text", + "content": "Personalized Video Relighting With an At-Home Light Stage" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 474, + 92, + 480, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 474, + 92, + 480, + 100 + ], + "spans": [ + { + "bbox": [ + 474, + 92, + 480, + 100 + ], + "type": "text", + "content": "5" + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 4 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 133, + 114, + 482, + 257 + ], + "blocks": [ + { + "bbox": [ + 133, + 114, + 482, + 257 + ], + "lines": [ + { + "bbox": [ + 133, + 114, + 482, + 257 + ], + "spans": [ + { + "bbox": [ + 133, + 114, + 482, + 257 + ], + "type": "image", + "image_path": "92c0bb905334e053dd6310840d1766b1017349265b63ae1111b14933f7e37fcf.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 130, + 266, + 482, + 332 + ], + "lines": [ + { + "bbox": [ + 130, + 266, + 482, + 332 + ], + "spans": [ + { + "bbox": [ + 130, + 266, + 482, + 332 + ], + "type": "text", + "content": "Fig. 3. We first de-light the input image features extracted by the U-Net encoder using Adaptive Instance Normalization (AdaIN) guided by the lighting features extracted from the source lighting with a Light Encoder. We then pass these light-normalized encoder features to the decoder of the U-Net and apply another set of AdaIN guided by the features extracted from the target lighting with the Light Encoder. We additionally predict source lighting from the U-Net encoder using a Light Decoder." + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_caption" + } + ], + "index": 2 + }, + { + "bbox": [ + 130, + 354, + 482, + 403 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 354, + 482, + 403 + ], + "spans": [ + { + "bbox": [ + 130, + 354, + 482, + 403 + ], + "type": "text", + "content": "unsuitable. Previous work [30] utilized segmentation for pairing. However, we observed that segmentation is ineffective at finding pairs of images with the same pose and expression. Thus, we instead use facial keypoint detection [16] to obtain source and target image pairs." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 131, + 418, + 321, + 431 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 131, + 418, + 321, + 431 + ], + "spans": [ + { + "bbox": [ + 131, + 418, + 321, + 431 + ], + "type": "text", + "content": "3.2 Relighting network architecture" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 130, + 437, + 482, + 628 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 437, + 482, + 628 + ], + "spans": [ + { + "bbox": [ + 130, + 437, + 482, + 628 + ], + "type": "text", + "content": "Our network architecture, as illustrated in Fig. 3, is built upon the well-established U-Net [29]. This architecture is comprised of an encoder and a decoder with skip connections, which are commonly used in existing portrait relighting algorithms [36,30,49,13,38]. Our U-Net's encoder, similar to Sengupta et al. [30] and Sun et al. [36], processes the source portrait " + }, + { + "bbox": [ + 130, + 437, + 482, + 628 + ], + "type": "inline_equation", + "content": "I_{\\mathrm{src}}" + }, + { + "bbox": [ + 130, + 437, + 482, + 628 + ], + "type": "text", + "content": " by applying multiple convolutional layers of varying strides (1 or 2). This process progressively reduces spatial resolution while increasing the number of channels, yielding a latent feature space. The decoder performs the inverse function of the encoder by upsampling from the latent features and simultaneously skip-connecting to intermediate features from the encoder. These skip connections transport high-frequency shape and appearance information from the encoder to the decoder, ultimately resulting in the generation of a realistic relit image. However, they also carry source illumination features from the encoder to the decoder, leading to subpar relighting quality and temporal flickering. To address this issue, we introduce light-conditioned feature normalization (LCFN) for the skip-connected features to better disentangle lighting features from intrinsic appearance features." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 130, + 630, + 482, + 666 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 630, + 482, + 666 + ], + "spans": [ + { + "bbox": [ + 130, + 630, + 482, + 666 + ], + "type": "text", + "content": "To disentangle lighting and intrinsic appearance components from the encoded features - i.e. to de-light - we first predict the source lighting from the encoder features. In contrast to Sun et al. [36], which predicts the illumination" + } + ] + } + ], + "index": 7 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "spans": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "type": "text", + "content": "6" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 166, + 91, + 220, + 100 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 166, + 91, + 220, + 100 + ], + "spans": [ + { + "bbox": [ + 166, + 91, + 220, + 100 + ], + "type": "text", + "content": "J. Choi et al." + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 5 + }, + { + "para_blocks": [ + { + "bbox": [ + 130, + 115, + 480, + 163 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 115, + 480, + 163 + ], + "spans": [ + { + "bbox": [ + 130, + 115, + 480, + 163 + ], + "type": "inline_equation", + "content": "\\hat{L}_{\\mathrm{src}}" + }, + { + "bbox": [ + 130, + 115, + 480, + 163 + ], + "type": "text", + "content": " corresponding to the source image " + }, + { + "bbox": [ + 130, + 115, + 480, + 163 + ], + "type": "inline_equation", + "content": "I_{\\mathrm{src}}" + }, + { + "bbox": [ + 130, + 115, + 480, + 163 + ], + "type": "text", + "content": " using the final encoded features, we take a different approach. We extract features at intermediate steps within the encoder, downsample them, and concatenate them using a confidence learning approach [14] to predict the illumination " + }, + { + "bbox": [ + 130, + 115, + 480, + 163 + ], + "type": "inline_equation", + "content": "\\hat{L}_{\\mathrm{src}}" + }, + { + "bbox": [ + 130, + 115, + 480, + 163 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 130, + 164, + 482, + 369 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 164, + 482, + 369 + ], + "spans": [ + { + "bbox": [ + 130, + 164, + 482, + 369 + ], + "type": "text", + "content": "The LCFN module uses the lighting features generated by the lighting encoder to perform Adaptive Instance Normalization (AdaIN) [17] on the encoder features. We begin by using a multi-layer perceptron (MLP) to encode lighting features, transforming the lighting information of " + }, + { + "bbox": [ + 130, + 164, + 482, + 369 + ], + "type": "inline_equation", + "content": "L_{\\mathrm{src}}" + }, + { + "bbox": [ + 130, + 164, + 482, + 369 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 130, + 164, + 482, + 369 + ], + "type": "inline_equation", + "content": "L_{\\mathrm{trg}}" + }, + { + "bbox": [ + 130, + 164, + 482, + 369 + ], + "type": "text", + "content": " into a compact, low-dimensional representation " + }, + { + "bbox": [ + 130, + 164, + 482, + 369 + ], + "type": "inline_equation", + "content": "(\\mathrm{d} = 256)" + }, + { + "bbox": [ + 130, + 164, + 482, + 369 + ], + "type": "text", + "content": ". We apply AdaIN to encoder features using the source lighting features, producing normalized features " + }, + { + "bbox": [ + 130, + 164, + 482, + 369 + ], + "type": "inline_equation", + "content": "f^l" + }, + { + "bbox": [ + 130, + 164, + 482, + 369 + ], + "type": "text", + "content": " (for " + }, + { + "bbox": [ + 130, + 164, + 482, + 369 + ], + "type": "inline_equation", + "content": "l = 1,\\dots ,7" + }, + { + "bbox": [ + 130, + 164, + 482, + 369 + ], + "type": "text", + "content": "). Through this normalization process, we induce de-lighting, effectively removing the lighting information present in the encoder features. Starting from the de-lit latent features " + }, + { + "bbox": [ + 130, + 164, + 482, + 369 + ], + "type": "inline_equation", + "content": "f^7" + }, + { + "bbox": [ + 130, + 164, + 482, + 369 + ], + "type": "text", + "content": ", we perform progressive bi-linear upsampling. At each upsampling step, we apply AdaIN to the concatenated feature, incorporating the target lighting features encoded by the lighting encoder. This construction using the LCFN module and source lighting prediction allows us to effectively remove source lighting features from the input and only propagate intrinsic appearance features from the encoder to the decoder. We then add target lighting features in the decoder. The LCFN module also contributes towards temporal consistency (see Sec. 3.3). See Fig. 2 for a comparison between our architecture and those of Sun et al. [36] and Sengupta et al. [30]." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 131, + 383, + 319, + 395 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 131, + 383, + 319, + 395 + ], + "spans": [ + { + "bbox": [ + 131, + 383, + 319, + 395 + ], + "type": "text", + "content": "3.3 Enforcing temporal consistency" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 130, + 400, + 480, + 472 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 400, + 480, + 472 + ], + "spans": [ + { + "bbox": [ + 130, + 400, + 480, + 472 + ], + "type": "text", + "content": "Temporal consistency is vital in making relit videos stable, realistic, and aesthetically pleasing. Previous single-image portrait relighting techniques [30,36] do not incorporate explicit temporal modeling, leading to undesirable flickering artifacts when applied to videos. Accuracy in single-image portrait relighting can often be uncorrelated to temporal flickering. Inconsistencies across frames are even more noticeable when the source lighting " + }, + { + "bbox": [ + 130, + 400, + 480, + 472 + ], + "type": "inline_equation", + "content": "L_{\\mathrm{src}}" + }, + { + "bbox": [ + 130, + 400, + 480, + 472 + ], + "type": "text", + "content": " changes continuously." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 130, + 472, + 481, + 555 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 472, + 481, + 555 + ], + "spans": [ + { + "bbox": [ + 130, + 472, + 481, + 555 + ], + "type": "text", + "content": "When applied to skip-connected features, LCFN provides a natural defense against temporal flickering by removing source lighting features from the input image. However, it cannot ensure temporal consistency on its own. We notice two further problems: (1) when the source lighting gradually changes, LCFN often leaks small amounts of source lighting information to the decoder, leading to flickering; (2) when source lighting changes abruptly, undesirable fading effects can be observed." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 130, + 555, + 481, + 616 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 555, + 481, + 616 + ], + "spans": [ + { + "bbox": [ + 130, + 555, + 481, + 616 + ], + "type": "text", + "content": "To address this issue, we propose a skip-connected feature smoothing technique that assumes neighboring frames share the same intrinsic appearance features, obtained after de-lighting input image features with LCFN. We apply a simple exponential smoothing of de-lit features generated by LCFN, denoted as " + }, + { + "bbox": [ + 130, + 555, + 481, + 616 + ], + "type": "inline_equation", + "content": "f^l" + }, + { + "bbox": [ + 130, + 555, + 481, + 616 + ], + "type": "text", + "content": ", using all the previous frames:" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 208, + 621, + 480, + 635 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 208, + 621, + 480, + 635 + ], + "spans": [ + { + "bbox": [ + 208, + 621, + 480, + 635 + ], + "type": "interline_equation", + "content": "f _ {t} ^ {l} := \\alpha \\cdot f _ {t} ^ {l} + (1 - \\alpha) \\cdot f _ {t - 1} ^ {l} \\quad (\\text {f o r} l = 1, \\dots , 7) \\tag {2}", + "image_path": "723ead59ff206995d4136844373b94655bec2295e2e03b9800246109236ac0d8.jpg" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 130, + 641, + 480, + 665 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 641, + 480, + 665 + ], + "spans": [ + { + "bbox": [ + 130, + 641, + 480, + 665 + ], + "type": "text", + "content": "with " + }, + { + "bbox": [ + 130, + 641, + 480, + 665 + ], + "type": "inline_equation", + "content": "\\alpha = 0.7" + }, + { + "bbox": [ + 130, + 641, + 480, + 665 + ], + "type": "text", + "content": ". Note that exponential smoothing does not work without de-lit LCFN features, which removes time-varying source lighting." + } + ] + } + ], + "index": 9 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 198, + 91, + 448, + 103 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 198, + 91, + 448, + 103 + ], + "spans": [ + { + "bbox": [ + 198, + 91, + 448, + 103 + ], + "type": "text", + "content": "Personalized Video Relighting With an At-Home Light Stage" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 474, + 91, + 481, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 474, + 91, + 481, + 100 + ], + "spans": [ + { + "bbox": [ + 474, + 91, + 481, + 100 + ], + "type": "text", + "content": "7" + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 6 + }, + { + "para_blocks": [ + { + "bbox": [ + 130, + 116, + 482, + 175 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 116, + 482, + 175 + ], + "spans": [ + { + "bbox": [ + 130, + 116, + 482, + 175 + ], + "type": "text", + "content": "We further notice that when the monitor light changes abruptly the relighting effect is delayed by a few frames, mainly due to the limited refresh rate of the monitor and frame rate of the camera. We thus propose doing a weighted average of source monitor lighting " + }, + { + "bbox": [ + 130, + 116, + 482, + 175 + ], + "type": "inline_equation", + "content": "L_{\\mathrm{src}}" + }, + { + "bbox": [ + 130, + 116, + 482, + 175 + ], + "type": "text", + "content": " from a sequence of previous and current frames to achieve smoother and more natural results:" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 253, + 178, + 481, + 209 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 253, + 178, + 481, + 209 + ], + "spans": [ + { + "bbox": [ + 253, + 178, + 481, + 209 + ], + "type": "interline_equation", + "content": "L _ {\\text {s r c a v g}} ^ {t} = \\frac {\\sum_ {i = 0} ^ {N - 1} \\beta^ {i} L _ {\\text {s r c}} ^ {t - i}}{\\sum_ {i = 0} ^ {N - 1} \\beta^ {i}}, \\tag {3}", + "image_path": "529bc30ab465e7e30f3ef390e6e59443e88ceab5c8918a923f21dcc1974e3f35.jpg" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 132, + 212, + 249, + 222 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 212, + 249, + 222 + ], + "spans": [ + { + "bbox": [ + 132, + 212, + 249, + 222 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 132, + 212, + 249, + 222 + ], + "type": "inline_equation", + "content": "\\beta = 0.6" + }, + { + "bbox": [ + 132, + 212, + 249, + 222 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 132, + 212, + 249, + 222 + ], + "type": "inline_equation", + "content": "N = 3" + }, + { + "bbox": [ + 132, + 212, + 249, + 222 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 132, + 238, + 299, + 250 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 238, + 299, + 250 + ], + "spans": [ + { + "bbox": [ + 132, + 238, + 299, + 250 + ], + "type": "text", + "content": "3.4 Training relighting network" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 130, + 254, + 479, + 327 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 254, + 479, + 327 + ], + "spans": [ + { + "bbox": [ + 130, + 254, + 479, + 327 + ], + "type": "text", + "content": "Our model is trained through minimizing a weighted combination of three loss functions: generator loss, discriminator loss, and monitor loss. The first loss aims to minimize the discrepancies between the true target image " + }, + { + "bbox": [ + 130, + 254, + 479, + 327 + ], + "type": "inline_equation", + "content": "I_{\\mathrm{trg}}" + }, + { + "bbox": [ + 130, + 254, + 479, + 327 + ], + "type": "text", + "content": " in our dataset and the predicted target relit image " + }, + { + "bbox": [ + 130, + 254, + 479, + 327 + ], + "type": "inline_equation", + "content": "\\hat{I}_{\\mathrm{trg}}" + }, + { + "bbox": [ + 130, + 254, + 479, + 327 + ], + "type": "text", + "content": ", leading to accurately relit images. We adopted our generator loss (Eq. 4) and our discriminator loss (Eq. 5) from Sengupta et al. [30]:" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 208, + 332, + 480, + 352 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 208, + 332, + 480, + 352 + ], + "spans": [ + { + "bbox": [ + 208, + 332, + 480, + 352 + ], + "type": "interline_equation", + "content": "\\mathcal {L} _ {G} ^ {I} = \\lambda_ {L 1} \\mathcal {L} _ {L 1} \\left(I _ {\\mathrm {t r g}}, \\hat {I} _ {\\mathrm {t r g}}\\right) + \\lambda_ {P} \\mathcal {L} _ {P} \\left(I _ {\\mathrm {t r g}}, \\hat {I} _ {\\mathrm {t r g}}\\right) \\tag {4}", + "image_path": "6d21465784c25d3802325cd7a6c4843516fd43a0261e99a4650902f7ae43db17.jpg" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 223, + 349, + 406, + 362 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 223, + 349, + 406, + 362 + ], + "spans": [ + { + "bbox": [ + 223, + 349, + 406, + 362 + ], + "type": "interline_equation", + "content": "+ \\lambda_ {C} \\mathcal {L} _ {C} \\left(I _ {\\mathrm {s r c}}, \\hat {I} _ {\\mathrm {s r c}} ^ {C}\\right) + \\lambda_ {D} \\left(D \\left(\\hat {I} _ {\\mathrm {t r g}}; \\theta_ {D}\\right) - 1\\right) ^ {2},", + "image_path": "cfcbacde1c85c9c49c81541ce7cca154042bfd2a4c5c86cffe532ae47de7e39a.jpg" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 217, + 371, + 480, + 384 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 217, + 371, + 480, + 384 + ], + "spans": [ + { + "bbox": [ + 217, + 371, + 480, + 384 + ], + "type": "interline_equation", + "content": "\\mathcal {L} _ {D} = \\left(D \\left(I _ {\\mathrm {t r g}}; \\theta_ {D}\\right) - 1\\right) ^ {2} + \\left(D \\left(\\hat {I} _ {\\mathrm {t r g}}; \\theta_ {D}\\right)\\right) ^ {2}, \\tag {5}", + "image_path": "08f90095da4ab0f4955abe8263500dd779bd6f5bb921baad8ffbbd31ac4d310e.jpg" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 130, + 387, + 479, + 426 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 387, + 479, + 426 + ], + "spans": [ + { + "bbox": [ + 130, + 387, + 479, + 426 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 130, + 387, + 479, + 426 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_{L1}" + }, + { + "bbox": [ + 130, + 387, + 479, + 426 + ], + "type": "text", + "content": " denotes L1 loss, " + }, + { + "bbox": [ + 130, + 387, + 479, + 426 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_P" + }, + { + "bbox": [ + 130, + 387, + 479, + 426 + ], + "type": "text", + "content": " denotes perceptual loss [45], " + }, + { + "bbox": [ + 130, + 387, + 479, + 426 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_C" + }, + { + "bbox": [ + 130, + 387, + 479, + 426 + ], + "type": "text", + "content": " denotes cycle consistency loss [50], and " + }, + { + "bbox": [ + 130, + 387, + 479, + 426 + ], + "type": "inline_equation", + "content": "D" + }, + { + "bbox": [ + 130, + 387, + 479, + 426 + ], + "type": "text", + "content": " is the discriminator [15]. " + }, + { + "bbox": [ + 130, + 387, + 479, + 426 + ], + "type": "inline_equation", + "content": "\\hat{I}_{\\mathrm{src}}^C" + }, + { + "bbox": [ + 130, + 387, + 479, + 426 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 130, + 387, + 479, + 426 + ], + "type": "inline_equation", + "content": "\\hat{L}_{\\mathrm{trg}}^C" + }, + { + "bbox": [ + 130, + 387, + 479, + 426 + ], + "type": "text", + "content": " are the outputs from " + }, + { + "bbox": [ + 130, + 387, + 479, + 426 + ], + "type": "inline_equation", + "content": "G(\\hat{I}_{\\mathrm{trg}}, L_{\\mathrm{trg}}, L_{\\mathrm{src}},; \\theta_G)" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 130, + 426, + 479, + 472 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 426, + 479, + 472 + ], + "spans": [ + { + "bbox": [ + 130, + 426, + 479, + 472 + ], + "type": "text", + "content": "The monitor reconstruction loss focuses on minimizing the errors between the predicted source light " + }, + { + "bbox": [ + 130, + 426, + 479, + 472 + ], + "type": "inline_equation", + "content": "\\hat{L}_{\\mathrm{src}}" + }, + { + "bbox": [ + 130, + 426, + 479, + 472 + ], + "type": "text", + "content": " and the true source light " + }, + { + "bbox": [ + 130, + 426, + 479, + 472 + ], + "type": "inline_equation", + "content": "L_{\\mathrm{src}}" + }, + { + "bbox": [ + 130, + 426, + 479, + 472 + ], + "type": "text", + "content": " and is expected to enforce improved disentanglement of lighting information from intrinsic appearance features." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 167, + 477, + 480, + 491 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 167, + 477, + 480, + 491 + ], + "spans": [ + { + "bbox": [ + 167, + 477, + 480, + 491 + ], + "type": "interline_equation", + "content": "\\mathcal {L} _ {G} ^ {M} = \\lambda_ {L 1} \\mathcal {L} _ {L 1} \\left(L _ {\\mathrm {s r c}}, \\hat {L} _ {\\mathrm {s r c}}\\right) + \\lambda_ {P} \\mathcal {L} _ {P} \\left(L _ {\\mathrm {s r c}}, \\hat {L} _ {\\mathrm {s r c}}\\right) + \\lambda_ {C} \\mathcal {L} _ {C} \\left(L _ {\\mathrm {t r g}}, \\hat {L} _ {C} ^ {\\text {t r g}}\\right). \\tag {6}", + "image_path": "d51711339ecbf38b3b4c3a72b68554782159da97266c7e61c5cbc8c25df693f0.jpg" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 130, + 495, + 479, + 519 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 495, + 479, + 519 + ], + "spans": [ + { + "bbox": [ + 130, + 495, + 479, + 519 + ], + "type": "text", + "content": "Finally, we minimize the image generator loss " + }, + { + "bbox": [ + 130, + 495, + 479, + 519 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_G^I" + }, + { + "bbox": [ + 130, + 495, + 479, + 519 + ], + "type": "text", + "content": ", the discriminator loss " + }, + { + "bbox": [ + 130, + 495, + 479, + 519 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_D" + }, + { + "bbox": [ + 130, + 495, + 479, + 519 + ], + "type": "text", + "content": ", and the illumination loss " + }, + { + "bbox": [ + 130, + 495, + 479, + 519 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_G^M" + }, + { + "bbox": [ + 130, + 495, + 479, + 519 + ], + "type": "text", + "content": " together:" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 255, + 523, + 480, + 540 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 255, + 523, + 480, + 540 + ], + "spans": [ + { + "bbox": [ + 255, + 523, + 480, + 540 + ], + "type": "interline_equation", + "content": "\\min _ {G, D} \\mathcal {L} _ {G} ^ {I} + \\mathcal {L} _ {D} + \\lambda_ {G} ^ {M} \\mathcal {L} _ {G} ^ {M}. \\tag {7}", + "image_path": "cc47435099bcfbfbb54084b22d3a053f85e80d9d0689302cb589f0fe132f8700.jpg" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 132, + 560, + 230, + 573 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 560, + 230, + 573 + ], + "spans": [ + { + "bbox": [ + 132, + 560, + 230, + 573 + ], + "type": "text", + "content": "4 Experiments" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 130, + 582, + 482, + 666 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 582, + 482, + 666 + ], + "spans": [ + { + "bbox": [ + 130, + 582, + 482, + 666 + ], + "type": "text", + "content": "In Sec. 4.1, we first discuss our data collection process, which was based on the approach used to compile the Light Stage at Your Desk (LSYD) dataset. In Sec. 4.2, we perform quantitative and qualitative comparisons with existing single-image portrait relighting algorithms and evaluate their temporal consistency. In Sec. 4.3, we demonstrate predicting, at a low-level, what the user is looking at on the monitor screen. Finally, in Sec. 4.4, we perform ablation studies, evaluating the network architecture's impact on relighting performance." + } + ] + } + ], + "index": 16 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "spans": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "type": "text", + "content": "8" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 166, + 91, + 220, + 101 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 166, + 91, + 220, + 101 + ], + "spans": [ + { + "bbox": [ + 166, + 91, + 220, + 101 + ], + "type": "text", + "content": "J. Choi et al." + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 7 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 134, + 113, + 482, + 399 + ], + "blocks": [ + { + "bbox": [ + 134, + 113, + 482, + 399 + ], + "lines": [ + { + "bbox": [ + 134, + 113, + 482, + 399 + ], + "spans": [ + { + "bbox": [ + 134, + 113, + 482, + 399 + ], + "type": "image", + "image_path": "3d129e60991e7581d7e9075533df2730fdc6e734021c7fe4e525a995406d229e.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 130, + 410, + 482, + 455 + ], + "lines": [ + { + "bbox": [ + 130, + 410, + 482, + 455 + ], + "spans": [ + { + "bbox": [ + 130, + 410, + 482, + 455 + ], + "type": "text", + "content": "Fig. 4. We perform a qualitative comparison with existing techniques [30,36] on the LSYD dataset. Source and target (images & lighting) were unseen during training. All models are personalized, i.e. trained on images of that individual only. We (Col. 3) produce significantly better results compared to existing approaches (Cols 4 and 5)." + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_caption" + } + ], + "index": 2 + }, + { + "bbox": [ + 132, + 477, + 186, + 488 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 477, + 186, + 488 + ], + "spans": [ + { + "bbox": [ + 132, + 477, + 186, + 488 + ], + "type": "text", + "content": "4.1 Data" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 130, + 498, + 482, + 666 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 498, + 482, + 666 + ], + "spans": [ + { + "bbox": [ + 130, + 498, + 482, + 666 + ], + "type": "text", + "content": "We recorded data from 5 users of diverse ethnicities and genders to ensure a wide range of skin types. Each participant wore a variety of outfits, and we used 4 different ambient lighting conditions per person to mimic the conditions of real-life online meetings. We directed the participants to continuously change their facial expression and pose during the capture sessions. Each user's face was captured while watching 8 different videos, each 8 minutes long, on different days with varying appearances. We randomly hold out 1 video for testing and use the remaining 7 videos for training. We use this testing sequence only for qualitative evaluation, not for any quantitative metrics. This is because quantitative evaluation requires a pair of source and target images of the same person in the same pose but under different lighting conditions. This is difficult to obtain accurately for the aforementioned test video sequence since the participants naturally vary their pose and expression over the course of the video. Instead, we capture an additional test sequence, used only for numerical evaluation in" + } + ] + } + ], + "index": 5 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 198, + 91, + 448, + 102 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 198, + 91, + 448, + 102 + ], + "spans": [ + { + "bbox": [ + 198, + 91, + 448, + 102 + ], + "type": "text", + "content": "Personalized Video Relighting With an At-Home Light Stage" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 474, + 92, + 480, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 474, + 92, + 480, + 100 + ], + "spans": [ + { + "bbox": [ + 474, + 92, + 480, + 100 + ], + "type": "text", + "content": "9" + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 8 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 133, + 114, + 481, + 331 + ], + "blocks": [ + { + "bbox": [ + 133, + 114, + 481, + 331 + ], + "lines": [ + { + "bbox": [ + 133, + 114, + 481, + 331 + ], + "spans": [ + { + "bbox": [ + 133, + 114, + 481, + 331 + ], + "type": "image", + "image_path": "7ae403f850a9bf0fb1465764a59d13c479f036723f1f9049c48e584246657292.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 130, + 342, + 482, + 376 + ], + "lines": [ + { + "bbox": [ + 130, + 342, + 482, + 376 + ], + "spans": [ + { + "bbox": [ + 130, + 342, + 482, + 376 + ], + "type": "text", + "content": "Fig. 5. We perform a qualitative comparison on the OLAT dataset [44]. Our approach outperforms [36,30] and can render strong directional lighting and specular highlights without any explicit modeling of geometry and reflectance." + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_caption" + } + ], + "index": 2 + }, + { + "bbox": [ + 130, + 398, + 482, + 494 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 398, + 482, + 494 + ], + "spans": [ + { + "bbox": [ + 130, + 398, + 482, + 494 + ], + "type": "text", + "content": "which the participant is captured in 9 different pose-expression combinations, each with a distinct monitor light. For each pose, we can create " + }, + { + "bbox": [ + 130, + 398, + 482, + 494 + ], + "type": "inline_equation", + "content": "\\binom{9}{2} = 36" + }, + { + "bbox": [ + 130, + 398, + 482, + 494 + ], + "type": "text", + "content": " source and target pairs as input and pseudo ground-truth, resulting in a total of 324 test data pairs per user. Additionally, we compared methods using Dynamic OLAT Dataset [44] with environment lighting maps [1,8]. During this comparison, we converted HDR environment maps to LDR and utilized a " + }, + { + "bbox": [ + 130, + 398, + 482, + 494 + ], + "type": "inline_equation", + "content": "270^{\\circ}" + }, + { + "bbox": [ + 130, + 398, + 482, + 494 + ], + "type": "text", + "content": " FoV as a monitor light to assess the relighting results. We want to note that all the test data is composed of unseen portraits and lightings for both LSYD and OLAT datasets." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 132, + 513, + 349, + 525 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 513, + 349, + 525 + ], + "spans": [ + { + "bbox": [ + 132, + 513, + 349, + 525 + ], + "type": "text", + "content": "4.2 Comparison with existing approaches" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 130, + 533, + 481, + 581 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 533, + 481, + 581 + ], + "spans": [ + { + "bbox": [ + 130, + 533, + 481, + 581 + ], + "type": "text", + "content": "We employ three error metrics to assess relighting performance: RMSE, LPIPS [45], and DISTS [6]. LPIPS and DISTS are more robust to slight differences in pose between the relit image and the pseudo ground truth and detect perceptual differences more effectively than RMSE." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 130, + 582, + 482, + 666 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 582, + 482, + 666 + ], + "spans": [ + { + "bbox": [ + 130, + 582, + 482, + 666 + ], + "type": "text", + "content": "**Portrait image relighting.** We compared our approach with existing portrait relighting neural architectures — Sun et al. [36] and Sengupta et al. [30] — by training on our captured LSYD dataset using the same pre-processing for all three architectures (see Sec. 3.1). Our training loss, given in Sec. 3.4, can handle misalignment in source-target pairs in training data, similar to the loss proposed in Sengupta et al. [30] (we use an additional loss on source monitor lighting prediction). For Sun et al. [36], we train both with their original loss" + } + ] + } + ], + "index": 7 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "spans": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "type": "text", + "content": "10" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 166, + 91, + 220, + 100 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 166, + 91, + 220, + 100 + ], + "spans": [ + { + "bbox": [ + 166, + 91, + 220, + 100 + ], + "type": "text", + "content": "J. Choi et al." + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 9 + }, + { + "para_blocks": [ + { + "type": "table", + "bbox": [ + 132, + 198, + 482, + 275 + ], + "blocks": [ + { + "bbox": [ + 130, + 113, + 482, + 191 + ], + "lines": [ + { + "bbox": [ + 130, + 113, + 482, + 191 + ], + "spans": [ + { + "bbox": [ + 130, + 113, + 482, + 191 + ], + "type": "text", + "content": "Table 1. We train a personalized relighting model on 5 users from our LSYD dataset and 4 users from the OLAT dataset [44]. We evaluate these models using source & target portrait images, as well as lighting, that were not encountered during the training phase, and report average RMSE, LPIPS [46], and DISTS[6] scores both on LSYD and OLAT dataset. Our method can perform relighting without source lighting " + }, + { + "bbox": [ + 130, + 113, + 482, + 191 + ], + "type": "inline_equation", + "content": "L_{\\mathrm{src}}" + }, + { + "bbox": [ + 130, + 113, + 482, + 191 + ], + "type": "text", + "content": ", by simply using the predicted light source from our model as input lighting. Our method significantly outperforms Sun et al. [36] and Sengupta et al. [30]." + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 132, + 198, + 482, + 275 + ], + "lines": [ + { + "bbox": [ + 132, + 198, + 482, + 275 + ], + "spans": [ + { + "bbox": [ + 132, + 198, + 482, + 275 + ], + "type": "table", + "html": "
Known source lightingLSYD dataOLAT data [44]
LsrcLPIPS ↓DISTS ↓RMSE ↓LPIPS ↓DISTS ↓RMSE ↓
Sun et al. [36] w/ LSun-0.17120.16298.59580.22730.17456.2692
Sun et al. [36] w/ LOurs-0.10290.11528.44760.22670.15696.1898
Ours-0.08390.09538.32220.18120.13366.0931
Sengupta et al. [30]✓0.10180.11058.28260.22370.16756.1751
Ours✓0.08320.09538.19390.18090.13345.9548
", + "image_path": "2e792c25fbccdfab1e58e91ab85c71bf777faa4a190227951cee6e0fe344644c.jpg" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "table_body" + } + ], + "index": 3 + }, + { + "bbox": [ + 130, + 301, + 482, + 361 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 301, + 482, + 361 + ], + "spans": [ + { + "bbox": [ + 130, + 301, + 482, + 361 + ], + "type": "text", + "content": "function " + }, + { + "bbox": [ + 130, + 301, + 482, + 361 + ], + "type": "inline_equation", + "content": "L_{\\mathrm{Sun}}" + }, + { + "bbox": [ + 130, + 301, + 482, + 361 + ], + "type": "text", + "content": " (which expects perfect source-target pose alignment obtained in OLAT data) and with our proposed loss function " + }, + { + "bbox": [ + 130, + 301, + 482, + 361 + ], + "type": "inline_equation", + "content": "L_{\\mathrm{ours}}" + }, + { + "bbox": [ + 130, + 301, + 482, + 361 + ], + "type": "text", + "content": " to specifically handle misalignment in LSYD data. We train personalized models on 5 users from the LSYD dataset and on 4 users from the publicly available Dynamic OLAT Dataset [44] with 2361 indoor environment lighting maps [1,8]." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 130, + 364, + 274, + 590 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 364, + 274, + 590 + ], + "spans": [ + { + "bbox": [ + 130, + 364, + 274, + 590 + ], + "type": "text", + "content": "For our quantitative evaluation, we test our model on 1620 test images across 5 users with unseen appearance and lighting conditions on the LSYD dataset and on 7172 test images from the Dynamic OLAT dataset. We present the result in Tab. 1. We observe that our proposed approach outperforms Sengupta et al. [30] and Sun et al. [36] by " + }, + { + "bbox": [ + 130, + 364, + 274, + 590 + ], + "type": "inline_equation", + "content": "22.3\\%" + }, + { + "bbox": [ + 130, + 364, + 274, + 590 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 130, + 364, + 274, + 590 + ], + "type": "inline_equation", + "content": "23.6\\%" + }, + { + "bbox": [ + 130, + 364, + 274, + 590 + ], + "type": "text", + "content": " respectively on the LSYD dataset and by " + }, + { + "bbox": [ + 130, + 364, + 274, + 590 + ], + "type": "inline_equation", + "content": "23.5\\%" + }, + { + "bbox": [ + 130, + 364, + 274, + 590 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 130, + 364, + 274, + 590 + ], + "type": "inline_equation", + "content": "25.6\\%" + }, + { + "bbox": [ + 130, + 364, + 274, + 590 + ], + "type": "text", + "content": " on the OLAT dataset, when comparing LPIPS score. Our qualitative comparison, as presented in Fig. 4 and Fig. 5, shows that our model performs superior re" + } + ] + } + ], + "index": 5 + }, + { + "type": "image", + "bbox": [ + 282, + 364, + 479, + 518 + ], + "blocks": [ + { + "bbox": [ + 282, + 364, + 479, + 518 + ], + "lines": [ + { + "bbox": [ + 282, + 364, + 479, + 518 + ], + "spans": [ + { + "bbox": [ + 282, + 364, + 479, + 518 + ], + "type": "image", + "image_path": "829b6f38fc6832be1caaaef5fd422d37ecbc21eaa7971f639fa05e93b4f24e65.jpg" + } + ] + } + ], + "index": 6, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 279, + 529, + 482, + 586 + ], + "lines": [ + { + "bbox": [ + 279, + 529, + 482, + 586 + ], + "spans": [ + { + "bbox": [ + 279, + 529, + 482, + 586 + ], + "type": "text", + "content": "Fig.6. Our method can also relight portrait images captured with unknown light sources. We relight a source image (Col. 1) with target light shown in the inset in Col. 2. We add a reference of how the face appears under that target light." + } + ] + } + ], + "index": 7, + "angle": 0, + "type": "image_caption" + } + ], + "index": 6 + }, + { + "bbox": [ + 130, + 591, + 482, + 615 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 591, + 482, + 615 + ], + "spans": [ + { + "bbox": [ + 130, + 591, + 482, + 615 + ], + "type": "text", + "content": "lighting in terms of color, quality, and consistency, and can render strong directional lighting and specular highlights without any explicit modeling." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 130, + 617, + 482, + 666 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 617, + 482, + 666 + ], + "spans": [ + { + "bbox": [ + 130, + 617, + 482, + 666 + ], + "type": "text", + "content": "Note that Sun et al. [36] does not require the source lighting " + }, + { + "bbox": [ + 130, + 617, + 482, + 666 + ], + "type": "inline_equation", + "content": "L_{\\mathrm{src}}" + }, + { + "bbox": [ + 130, + 617, + 482, + 666 + ], + "type": "text", + "content": " during test time. Our proposed approach can also perform relighting without prior knowledge of source lighting " + }, + { + "bbox": [ + 130, + 617, + 482, + 666 + ], + "type": "inline_equation", + "content": "L_{\\mathrm{src}}" + }, + { + "bbox": [ + 130, + 617, + 482, + 666 + ], + "type": "text", + "content": " by simply predicting " + }, + { + "bbox": [ + 130, + 617, + 482, + 666 + ], + "type": "inline_equation", + "content": "\\hat{L}_{\\mathrm{src}}" + }, + { + "bbox": [ + 130, + 617, + 482, + 666 + ], + "type": "text", + "content": " and using it for light-conditioned feature normalization. We show that even in the absence of" + } + ] + } + ], + "index": 9 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 198, + 90, + 448, + 102 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 198, + 90, + 448, + 102 + ], + "spans": [ + { + "bbox": [ + 198, + 90, + 448, + 102 + ], + "type": "text", + "content": "Personalized Video Relighting With an At-Home Light Stage" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 470, + 91, + 479, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 470, + 91, + 479, + 100 + ], + "spans": [ + { + "bbox": [ + 470, + 91, + 479, + 100 + ], + "type": "text", + "content": "11" + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 10 + }, + { + "para_blocks": [ + { + "bbox": [ + 130, + 115, + 479, + 140 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 115, + 479, + 140 + ], + "spans": [ + { + "bbox": [ + 130, + 115, + 479, + 140 + ], + "type": "inline_equation", + "content": "L_{\\mathrm{src}}" + }, + { + "bbox": [ + 130, + 115, + 479, + 140 + ], + "type": "text", + "content": ", our method outperforms Sun et al. [36] by " + }, + { + "bbox": [ + 130, + 115, + 479, + 140 + ], + "type": "inline_equation", + "content": "22.6\\%" + }, + { + "bbox": [ + 130, + 115, + 479, + 140 + ], + "type": "text", + "content": " on the LSYD data and by " + }, + { + "bbox": [ + 130, + 115, + 479, + 140 + ], + "type": "inline_equation", + "content": "25.5\\%" + }, + { + "bbox": [ + 130, + 115, + 479, + 140 + ], + "type": "text", + "content": " on the OLAT data, in terms of LPIPS." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 130, + 144, + 480, + 216 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 144, + 480, + 216 + ], + "spans": [ + { + "bbox": [ + 130, + 144, + 480, + 216 + ], + "type": "text", + "content": "In Fig. 6, we demonstrate that our approach can relight any portrait image captured 'in-the-wild' without requiring the knowledge of source lighting, and outperforms Sun et al. [36]. Our approach first predicts a proxy source monitor lighting, imagining the portrait image to be captured under a monitor lighting, and uses this predicted source lighting in the LCFN module to 'de-lit' the input image and relight it with a target lighting." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 130, + 219, + 272, + 386 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 219, + 272, + 386 + ], + "spans": [ + { + "bbox": [ + 130, + 219, + 272, + 386 + ], + "type": "text", + "content": "In Fig. 7 1st-row, we show the results of relighting an input image with directional lights by moving specific bright areas on the monitor screen. Our method learns to render directional lighting effects and cast shadows as needed. To better illustrate the effect of moving lighting, we first relight the input image with no light reflected from the monitor (row-2 col-1), which produces a relief image under ambient room" + } + ] + } + ], + "index": 4 + }, + { + "type": "image", + "bbox": [ + 282, + 217, + 323, + 295 + ], + "blocks": [ + { + "bbox": [ + 282, + 217, + 323, + 295 + ], + "lines": [ + { + "bbox": [ + 282, + 217, + 323, + 295 + ], + "spans": [ + { + "bbox": [ + 282, + 217, + 323, + 295 + ], + "type": "image", + "image_path": "766dae490665a6cdf8c35774e60d851e62c027862619c7e6fc1c5c69f13aab8d.jpg" + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 279, + 308, + 482, + 384 + ], + "lines": [ + { + "bbox": [ + 279, + 308, + 482, + 384 + ], + "spans": [ + { + "bbox": [ + 279, + 308, + 482, + 384 + ], + "type": "text", + "content": "Fig. 7. We relight the input image with directional lights (1st row) and without any target light ('Relight A'). We then subtract the 'Relight A' from directional images (row-2 col-2:5). Our method learns to render the effects of directional lighting and decouple ambient and dominant frontal lighting (from monitor)." + } + ] + } + ], + "index": 10, + "angle": 0, + "type": "image_caption" + } + ], + "index": 5 + }, + { + "type": "image", + "bbox": [ + 323, + 217, + 361, + 295 + ], + "blocks": [ + { + "bbox": [ + 323, + 217, + 361, + 295 + ], + "lines": [ + { + "bbox": [ + 323, + 217, + 361, + 295 + ], + "spans": [ + { + "bbox": [ + 323, + 217, + 361, + 295 + ], + "type": "image", + "image_path": "73735b90d02b709dfef115d3bca1386ca66be993dcfedc53fc48db8e666923a6.jpg" + } + ] + } + ], + "index": 6, + "angle": 0, + "type": "image_body" + } + ], + "index": 6 + }, + { + "type": "image", + "bbox": [ + 361, + 217, + 400, + 295 + ], + "blocks": [ + { + "bbox": [ + 361, + 217, + 400, + 295 + ], + "lines": [ + { + "bbox": [ + 361, + 217, + 400, + 295 + ], + "spans": [ + { + "bbox": [ + 361, + 217, + 400, + 295 + ], + "type": "image", + "image_path": "833ec48b41a5f287af20c0dbca34ef64febc2c348d0c5a2df04626e2c89da384.jpg" + } + ] + } + ], + "index": 7, + "angle": 0, + "type": "image_body" + } + ], + "index": 7 + }, + { + "type": "image", + "bbox": [ + 400, + 217, + 439, + 295 + ], + "blocks": [ + { + "bbox": [ + 400, + 217, + 439, + 295 + ], + "lines": [ + { + "bbox": [ + 400, + 217, + 439, + 295 + ], + "spans": [ + { + "bbox": [ + 400, + 217, + 439, + 295 + ], + "type": "image", + "image_path": "b26c96fe6105b580185f4398c11bfe3564db9ff5b0580854bc0f24c3ccbf2c12.jpg" + } + ] + } + ], + "index": 8, + "angle": 0, + "type": "image_body" + } + ], + "index": 8 + }, + { + "type": "image", + "bbox": [ + 439, + 217, + 478, + 295 + ], + "blocks": [ + { + "bbox": [ + 439, + 217, + 478, + 295 + ], + "lines": [ + { + "bbox": [ + 439, + 217, + 478, + 295 + ], + "spans": [ + { + "bbox": [ + 439, + 217, + 478, + 295 + ], + "type": "image", + "image_path": "42ebe8fb772bfb7547b07bc29344a51be6b13c15d1f84a9046870973fa7ef54e.jpg" + } + ] + } + ], + "index": 9, + "angle": 0, + "type": "image_body" + } + ], + "index": 9 + }, + { + "bbox": [ + 130, + 387, + 480, + 460 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 387, + 480, + 460 + ], + "spans": [ + { + "bbox": [ + 130, + 387, + 480, + 460 + ], + "type": "text", + "content": "lighting only. Subtracting this 'de-lit' image from the relit images under directional lighting (row-2 col 2-5) highlights the ability of our method to decouple ambient room illumination from dominant frontal lighting. However, due to the constraint that our lighting is limited to the illumination emitted from the monitor, we may not accurately depict extreme lighting effects. To this end, we show the strong directional light effects using the OLAT dataset in Fig. 5." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 130, + 462, + 284, + 666 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 462, + 284, + 666 + ], + "spans": [ + { + "bbox": [ + 130, + 462, + 284, + 666 + ], + "type": "text", + "content": "Portrait video relighting. Next, we evaluate the temporal consistency of each portrait video relighting algorithm. For each user in the LSYD data, we relit the held-out test video with 50 different target lighting conditions, creating 50 relit videos. We then computed the RMSE between adjacent frames in relit videos as a measure of temporal consistency. Since the pose is almost identical between adjacent frames, lower RMSE error indicates temporally consistent relighting. We then report the average temporal RMSE across all such adjacent frames. In practice," + } + ] + } + ], + "index": 12 + }, + { + "type": "table", + "bbox": [ + 291, + 552, + 481, + 658 + ], + "blocks": [ + { + "bbox": [ + 289, + 463, + 482, + 551 + ], + "lines": [ + { + "bbox": [ + 289, + 463, + 482, + 551 + ], + "spans": [ + { + "bbox": [ + 289, + 463, + 482, + 551 + ], + "type": "text", + "content": "Table 2. We evaluate temporal consistency by relighting a test video captured with varying source lighting with the same target lighting and calculating RMSE between adjacent frames. We then report the average RMSE across all adjacent frames and compute an error rate to indicate the percent of adjacent frames with RMSE higher than a threshold." + } + ] + } + ], + "index": 13, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 291, + 552, + 481, + 658 + ], + "lines": [ + { + "bbox": [ + 291, + 552, + 481, + 658 + ], + "spans": [ + { + "bbox": [ + 291, + 552, + 481, + 658 + ], + "type": "table", + "html": "
ThresholdRMSE ↓Error Rate (%)
>0.2>0.3>0.4
Sun et al. [36]5.8613.532.611.06
Sengupta et al. [30]6.3721.835.401.75
+Lsrc_avg5.7613.312.340.98
Ours (base)6.0116.223.611.08
+Lsrc_avg5.7313.092.310.83
+LCFN5.6813.042.280.71
+Lsrc_avg + LCFN5.5512.892.220.65
", + "image_path": "8376becaae33b62fc258e719efff9f718f176a0e660b472453b8490a7e6d87ec.jpg" + } + ] + } + ], + "index": 14, + "angle": 0, + "type": "table_body" + } + ], + "index": 14 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "spans": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "type": "text", + "content": "12" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 166, + 91, + 220, + 101 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 166, + 91, + 220, + 101 + ], + "spans": [ + { + "bbox": [ + 166, + 91, + 220, + 101 + ], + "type": "text", + "content": "J. Choi et al." + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 11 + }, + { + "para_blocks": [ + { + "bbox": [ + 130, + 116, + 482, + 200 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 116, + 482, + 200 + ], + "spans": [ + { + "bbox": [ + 130, + 116, + 482, + 200 + ], + "type": "text", + "content": "however, a significant fraction of adjacent frame pairs have extremely similar lighting between the two frames, making their relit frames naturally consistent anyway. Only in a small percentage of adjacent frames does the source lighting significantly change, leading to obvious flickering in the relit video if temporal consistency is not maintained. Thus, in addition to average temporal RMSE, we also compute the error rate for three different thresholds: 0.2 (low), 0.3 (medium), and 0.4 (high), which indicate the percentage of adjacent frames where RMSE" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 130, + 200, + 277, + 211 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 200, + 277, + 211 + ], + "spans": [ + { + "bbox": [ + 130, + 200, + 277, + 211 + ], + "type": "text", + "content": "error is more than the threshold." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 130, + 211, + 274, + 499 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 211, + 274, + 499 + ], + "spans": [ + { + "bbox": [ + 130, + 211, + 274, + 499 + ], + "type": "text", + "content": "In Tab. 2 and Fig. 8, we compare our approach with and without skip-connected feature smoothing to past works [36,30]. Note that this temporal smoothing of the skip-connected features can only be applied in our framework since we de-light encoder features from the source lighting with LCFN. Both for our approach and for Sengupta et al. [30], we can further apply smoothing of input source lighting to handle abrupt changes. We observe that our method produces the most temporally consistent relighting while also being the most accurate (see Tab. 1). We further note that both skip-connected feature smoothing and smoothing of source lighting improve temporal consistency." + } + ] + } + ], + "index": 4 + }, + { + "type": "image", + "bbox": [ + 281, + 210, + 480, + 411 + ], + "blocks": [ + { + "bbox": [ + 281, + 210, + 480, + 411 + ], + "lines": [ + { + "bbox": [ + 281, + 210, + 480, + 411 + ], + "spans": [ + { + "bbox": [ + 281, + 210, + 480, + 411 + ], + "type": "image", + "image_path": "807de019c307e9311014a483b329950f97e590d5be83fe3c49ea9433a007255e.jpg" + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 279, + 422, + 482, + 498 + ], + "lines": [ + { + "bbox": [ + 279, + 422, + 482, + 498 + ], + "spans": [ + { + "bbox": [ + 279, + 422, + 482, + 498 + ], + "type": "text", + "content": "Fig.8. We show temporal consistency between adjacent frames separated by 0.33s by relighting a test video captured with varying source lighting with the same target lighting. Note that [36,30] both exhibit abrupt changes in lighting between frames " + }, + { + "bbox": [ + 279, + 422, + 482, + 498 + ], + "type": "inline_equation", + "content": "t" + }, + { + "bbox": [ + 279, + 422, + 482, + 498 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 279, + 422, + 482, + 498 + ], + "type": "inline_equation", + "content": "t + 10" + }, + { + "bbox": [ + 279, + 422, + 482, + 498 + ], + "type": "text", + "content": ", while our approach produces a more stable result." + } + ] + } + ], + "index": 6, + "angle": 0, + "type": "image_caption" + } + ], + "index": 5 + }, + { + "bbox": [ + 132, + 515, + 257, + 528 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 515, + 257, + 528 + ], + "spans": [ + { + "bbox": [ + 132, + 515, + 257, + 528 + ], + "type": "text", + "content": "4.3 Monitor prediction" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 130, + 533, + 273, + 652 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 533, + 273, + 652 + ], + "spans": [ + { + "bbox": [ + 130, + 533, + 273, + 652 + ], + "type": "text", + "content": "Our network can also be used to predict the source monitor lighting of any input image, as shown in Fig. 9. For the LSYD dataset, we do have images of the source monitor for every input image, so we can calculate monitor prediction accuracy numerically. We observe that our method slightly out" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 130, + 653, + 481, + 666 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 653, + 481, + 666 + ], + "spans": [ + { + "bbox": [ + 130, + 653, + 481, + 666 + ], + "type": "text", + "content": "performs [36] with a Mean Absolute Error of 0.15 vs 0.17, also qualitatively" + } + ] + } + ], + "index": 9 + }, + { + "type": "image", + "bbox": [ + 281, + 532, + 479, + 599 + ], + "blocks": [ + { + "bbox": [ + 281, + 532, + 479, + 599 + ], + "lines": [ + { + "bbox": [ + 281, + 532, + 479, + 599 + ], + "spans": [ + { + "bbox": [ + 281, + 532, + 479, + 599 + ], + "type": "image", + "image_path": "c359b5fab8d4a90308e8904d8d440b06fda78369c1aef82172ac4752626fb693.jpg" + } + ] + } + ], + "index": 10, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 279, + 609, + 482, + 654 + ], + "lines": [ + { + "bbox": [ + 279, + 609, + 482, + 654 + ], + "spans": [ + { + "bbox": [ + 279, + 609, + 482, + 654 + ], + "type": "text", + "content": "Fig. 9. Monitor prediction comparison on LSYD data (Col. 1 and 2) and \"in the wild\" setting (Col. 3 and 4). The second row represents the monitor prediction by [36] and Ours." + } + ] + } + ], + "index": 11, + "angle": 0, + "type": "image_caption" + } + ], + "index": 10 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 198, + 91, + 448, + 102 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 198, + 91, + 448, + 102 + ], + "spans": [ + { + "bbox": [ + 198, + 91, + 448, + 102 + ], + "type": "text", + "content": "Personalized Video Relighting With an At-Home Light Stage" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 470, + 91, + 480, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 470, + 91, + 480, + 100 + ], + "spans": [ + { + "bbox": [ + 470, + 91, + 480, + 100 + ], + "type": "text", + "content": "13" + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 12 + }, + { + "para_blocks": [ + { + "bbox": [ + 130, + 116, + 479, + 140 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 116, + 479, + 140 + ], + "spans": [ + { + "bbox": [ + 130, + 116, + 479, + 140 + ], + "type": "text", + "content": "producing more meaningful visualizations. For portrait images, we do not have ground-truth source lighting, but visualizations show meaningful predictions." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 130, + 140, + 480, + 212 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 140, + 480, + 212 + ], + "spans": [ + { + "bbox": [ + 130, + 140, + 480, + 212 + ], + "type": "text", + "content": "The ability to predict monitor lighting from images captured by our monitor has many implications. This technique can be used to detect deep fake avatars during live video calls by purposefully projecting specific images via screen sharing and observing if we can detect the same image from the webcam feed of other attendees in the call. If there is a mismatch between the projected monitor image and the predicted one, this likely indicates a deep fake avatar in the video call." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 131, + 229, + 243, + 241 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 131, + 229, + 243, + 241 + ], + "spans": [ + { + "bbox": [ + 131, + 229, + 243, + 241 + ], + "type": "text", + "content": "4.4 Ablation studies" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 130, + 249, + 284, + 357 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 249, + 284, + 357 + ], + "spans": [ + { + "bbox": [ + 130, + 249, + 284, + 357 + ], + "type": "text", + "content": "We report the removal of various components from our relighting network in Tab. 3, specifically LCFN and source monitor lighting prediction using intermediate encoder features. We observe that both improve final relighting performance, which shows their effectiveness in disentangling source" + } + ] + } + ], + "index": 5 + }, + { + "type": "table", + "bbox": [ + 291, + 295, + 481, + 351 + ], + "blocks": [ + { + "bbox": [ + 290, + 250, + 481, + 294 + ], + "lines": [ + { + "bbox": [ + 290, + 250, + 481, + 294 + ], + "spans": [ + { + "bbox": [ + 290, + 250, + 481, + 294 + ], + "type": "text", + "content": "Table 3. Both LCFN and source monitor prediction " + }, + { + "bbox": [ + 290, + 250, + 481, + 294 + ], + "type": "inline_equation", + "content": "\\mathbf{L}_{\\mathrm{src}}" + }, + { + "bbox": [ + 290, + 250, + 481, + 294 + ], + "type": "text", + "content": " improve relighting performance by effectively disentangling source lighting information from intrinsic appearance features." + } + ] + } + ], + "index": 6, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 291, + 295, + 481, + 351 + ], + "lines": [ + { + "bbox": [ + 291, + 295, + 481, + 351 + ], + "spans": [ + { + "bbox": [ + 291, + 295, + 481, + 351 + ], + "type": "table", + "html": "
Lsrcde-lightingRMSE ↓LPIPS ↓DISTS ↓
--8.42300.09640.1071
✓-8.25960.09150.1013
-✓8.19070.09040.0966
✓✓8.07460.08530.0963
", + "image_path": "c58fece9fc9f1ef686e4767e6a28bbed2cfca27cfa9736ed99e18e8c678ce97a.jpg" + } + ] + } + ], + "index": 7, + "angle": 0, + "type": "table_body" + } + ], + "index": 7 + }, + { + "bbox": [ + 131, + 357, + 375, + 369 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 131, + 357, + 375, + 369 + ], + "spans": [ + { + "bbox": [ + 131, + 357, + 375, + 369 + ], + "type": "text", + "content": "lighting information from intrinsic appearance features." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 131, + 387, + 219, + 399 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 131, + 387, + 219, + 399 + ], + "spans": [ + { + "bbox": [ + 131, + 387, + 219, + 399 + ], + "type": "text", + "content": "5 Conclusion" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 130, + 412, + 480, + 532 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 412, + 480, + 532 + ], + "spans": [ + { + "bbox": [ + 130, + 412, + 480, + 532 + ], + "type": "text", + "content": "We propose a personalized video relighting algorithm that leverages casually captured LSYD data to generate real-time high-quality temporally consistent relit videos under any pose, expression, and lighting conditions. We present a novel network architecture that can perform high-quality relighting on both the LSYD and OLAT datasets, rendering challenging lighting conditions like directional lights, shadows, and specularities. We achieve this without using any 3D information or performing explicit decomposition, simply by achieving better image-based relighting enabled by our proposed neural architecture with LCFN module. Our method enables better lighting quality during live video calls and in portrait images, and produces better harmonization with virtual backgrounds." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 130, + 532, + 480, + 591 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 532, + 480, + 591 + ], + "spans": [ + { + "bbox": [ + 130, + 532, + 480, + 591 + ], + "type": "text", + "content": "Limitation. Since we utilize only the front-facing monitor light as the source lighting, we do not account for scenarios where the light source is located on the sides " + }, + { + "bbox": [ + 130, + 532, + 480, + 591 + ], + "type": "inline_equation", + "content": "(90^{\\circ})" + }, + { + "bbox": [ + 130, + 532, + 480, + 591 + ], + "type": "text", + "content": " or behind " + }, + { + "bbox": [ + 130, + 532, + 480, + 591 + ], + "type": "inline_equation", + "content": "(180^{\\circ})" + }, + { + "bbox": [ + 130, + 532, + 480, + 591 + ], + "type": "text", + "content": ". However, it does not necessarily mean that our model cannot learn directional lighting effects, as demonstrated by training on the OLAT dataset which contains full " + }, + { + "bbox": [ + 130, + 532, + 480, + 591 + ], + "type": "inline_equation", + "content": "360^{\\circ}" + }, + { + "bbox": [ + 130, + 532, + 480, + 591 + ], + "type": "text", + "content": " lighting, see Fig. 5." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 130, + 591, + 481, + 664 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 591, + 481, + 664 + ], + "spans": [ + { + "bbox": [ + 130, + 591, + 481, + 664 + ], + "type": "text", + "content": "Ethical considerations. While our primary goal is to allow people to improve their facial appearance with virtual relighting, we note that it is also a form of image manipulation and can be used for malicious purposes [3]. Furthermore, we emphasize that while currently we can only predict the user's direct line-of-sight monitor lighting in low resolution, the potential for high-resolution monitor prediction in the future could raise significant privacy concerns." + } + ] + } + ], + "index": 12 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "spans": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "type": "text", + "content": "14" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 166, + 91, + 220, + 100 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 166, + 91, + 220, + 100 + ], + "spans": [ + { + "bbox": [ + 166, + 91, + 220, + 100 + ], + "type": "text", + "content": "J. Choi et al." + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 13 + }, + { + "para_blocks": [ + { + "bbox": [ + 133, + 114, + 197, + 126 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 114, + 197, + 126 + ], + "spans": [ + { + "bbox": [ + 133, + 114, + 197, + 126 + ], + "type": "text", + "content": "References" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 138, + 137, + 480, + 664 + ], + "type": "list", + "angle": 0, + "index": 19, + "blocks": [ + { + "bbox": [ + 138, + 137, + 480, + 159 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 137, + 480, + 159 + ], + "spans": [ + { + "bbox": [ + 138, + 137, + 480, + 159 + ], + "type": "text", + "content": "1. Bolduc, C., Giroux, J., Hébert, M., Demers, C., Lalonde, J.F.: Beyond the pixel: a photometrically calibrated hdr dataset for luminance and color prediction (2023)" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 138, + 159, + 462, + 169 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 159, + 462, + 169 + ], + "spans": [ + { + "bbox": [ + 138, + 159, + 462, + 169 + ], + "type": "text", + "content": "2. Chen, Z., Liu, Z.: Relighting4d: Neural relightable human from videos (2022)" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 138, + 169, + 480, + 212 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 169, + 480, + 212 + ], + "spans": [ + { + "bbox": [ + 138, + 169, + 480, + 212 + ], + "type": "text", + "content": "3. Choi, J.M., Leung, J., Frahm, N., Christman, M., Bertasius, G., Sengupta, R.: Building secure and engaging video communication by using monitor illumination. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 4377-4386 (2024)" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 138, + 212, + 480, + 245 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 212, + 480, + 245 + ], + "spans": [ + { + "bbox": [ + 138, + 212, + 480, + 245 + ], + "type": "text", + "content": "4. Daichi Tajima, Yoshihiro Kanamori, Y.E.: Relighting humans in the wild: Monocular full-body human relighting with domain adaptation. Computer Graphics Forum (Proc. of Pacific Graphics 2021) 40(7), 205-216 (2021)" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 138, + 245, + 480, + 287 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 245, + 480, + 287 + ], + "spans": [ + { + "bbox": [ + 138, + 245, + 480, + 287 + ], + "type": "text", + "content": "5. Debevec, P., Wenger, A., Tchou, C., Gardner, A., Waese, J., Hawkins, T.: A lighting reproduction approach to live-action compositing. ACM Trans. Graph. 21(3), 547-556 (jul 2002). https://doi.org/10.1145/566654.566614, https://doi.org/10.1145/566654.566614" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 138, + 288, + 480, + 319 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 288, + 480, + 319 + ], + "spans": [ + { + "bbox": [ + 138, + 288, + 480, + 319 + ], + "type": "text", + "content": "6. Ding, K., Ma, K., Wang, S., Simoncelli, E.P.: Image quality assessment: Unifying structure and texture similarity. CoRR abs/2004.07728 (2020), https://arxiv.org/abs/2004.07728" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 138, + 320, + 480, + 342 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 320, + 480, + 342 + ], + "spans": [ + { + "bbox": [ + 138, + 320, + 480, + 342 + ], + "type": "text", + "content": "7. Futschik, D., Ritland, K., Vecore, J., Fanello, S., Orts-Escolano, S., Curless, B., Sykora, D., Pandey, R.: Controllable light diffusion for portraits (2023)" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 138, + 342, + 480, + 363 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 342, + 480, + 363 + ], + "spans": [ + { + "bbox": [ + 138, + 342, + 480, + 363 + ], + "type": "text", + "content": "8. Gardner, M.A., Sunkavalli, K., Yumer, E., Shen, X., Gambaretto, E., Gagné, C., Lalonde, J.F.: Learning to predict indoor illumination from a single image (2017)" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 138, + 363, + 480, + 405 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 363, + 480, + 405 + ], + "spans": [ + { + "bbox": [ + 138, + 363, + 480, + 405 + ], + "type": "text", + "content": "9. Gerstner, C.R., Farid, H.: Detecting real-time deep-fake videos using active illumination. In: 2022 IEEE/CVF Conference on Computer Vision and Pattern Recognition Workshops (CVPRW). pp. 53-60 (2022). https://doi.org/10.1109/CVPRW56347.2022.00015" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 138, + 405, + 480, + 460 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 405, + 480, + 460 + ], + "spans": [ + { + "bbox": [ + 138, + 405, + 480, + 460 + ], + "type": "text", + "content": "0. Guo, K., Lincoln, P., Davidson, P., Busch, J., Yu, X., Whalen, M., Harvey, G., Orts-Escalano, S., Pandey, R., Dourgarian, J., Tang, D., Tkach, A., Kowdle, A., Cooper, E., Dou, M., Fanello, S., Fyffe, G., Rhemann, C., Taylor, J., Debevec, P., Izadi, S.: The relightables: Volumetric performance capture of humans with realistic relighting (2019)" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 138, + 460, + 480, + 503 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 460, + 480, + 503 + ], + "spans": [ + { + "bbox": [ + 138, + 460, + 480, + 503 + ], + "type": "text", + "content": "1. He, Y., Xing, Y., Zhang, T., Chen, Q.: Unsupervised portrait shadow removal via generative priors. In: Proceedings of the 29th ACM International Conference on Multimedia. MM '21, ACM (Oct 2021). https://doi.org/10.1145/3474085.3475663, http://dx.doi.org/10.1145/3474085.3475663" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 138, + 503, + 480, + 536 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 503, + 480, + 536 + ], + "spans": [ + { + "bbox": [ + 138, + 503, + 480, + 536 + ], + "type": "text", + "content": "2. Hou, A., Sarkis, M., Bi, N., Tong, Y., Liu, X.: Face relighting with geometrically consistent shadows. In: In Proceeding of IEEE Computer Vision and Pattern Recognition. New Orleans, LA (June 2022)" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 138, + 536, + 480, + 568 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 536, + 480, + 568 + ], + "spans": [ + { + "bbox": [ + 138, + 536, + 480, + 568 + ], + "type": "text", + "content": "3. Hou, A., Zhang, Z., Sarkis, M., Bi, N., Tong, Y., Liu, X.: Towards high fidelity face relighting with realistic shadows. In: IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR) (2021)" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 138, + 568, + 480, + 601 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 568, + 480, + 601 + ], + "spans": [ + { + "bbox": [ + 138, + 568, + 480, + 601 + ], + "type": "text", + "content": "4. Hu, Y., Wang, B., Lin, S.: Fc 4: Fully convolutional color constancy with confidence-weighted pooling. In: Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition. pp. 4085-4094 (2017)" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 138, + 601, + 480, + 621 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 601, + 480, + 621 + ], + "spans": [ + { + "bbox": [ + 138, + 601, + 480, + 621 + ], + "type": "text", + "content": "5. Isola, P., Zhu, J.Y., Zhou, T., Efros, A.A.: Image-to-image translation with conditional adversarial networks. CVPR (2017)" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 138, + 621, + 480, + 664 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 621, + 480, + 664 + ], + "spans": [ + { + "bbox": [ + 138, + 621, + 480, + 664 + ], + "type": "text", + "content": "6. Jin, H., Liao, S., Shao, L.: Pixel-in-pixel net: Towards efficient facial landmark detection in the wild. International Journal of Computer Vision (Sep 2021). https://doi.org/10.1007/s11263-021-01521-4, http://dx.doi.org/10.1007/s11263-021-01521-4" + } + ] + } + ], + "index": 18 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 198, + 91, + 447, + 102 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 198, + 91, + 447, + 102 + ], + "spans": [ + { + "bbox": [ + 198, + 91, + 447, + 102 + ], + "type": "text", + "content": "Personalized Video Relighting With an At-Home Light Stage" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 470, + 92, + 480, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 470, + 92, + 480, + 100 + ], + "spans": [ + { + "bbox": [ + 470, + 92, + 480, + 100 + ], + "type": "text", + "content": "15" + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 14 + }, + { + "para_blocks": [ + { + "bbox": [ + 132, + 117, + 480, + 665 + ], + "type": "list", + "angle": 0, + "index": 19, + "blocks": [ + { + "bbox": [ + 133, + 117, + 480, + 138 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 117, + 480, + 138 + ], + "spans": [ + { + "bbox": [ + 133, + 117, + 480, + 138 + ], + "type": "text", + "content": "17. Karras, T., Laine, S., Aittala, M., Hellsten, J., Lehtinen, J., Aila, T.: Analyzing and improving the image quality of StyleGAN. In: Proc. CVPR (2020)" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 133, + 139, + 480, + 170 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 139, + 480, + 170 + ], + "spans": [ + { + "bbox": [ + 133, + 139, + 480, + 170 + ], + "type": "text", + "content": "18. Lagunas, M., Sun, X., Yang, J., Villegas, R., Zhang, J., Shu, Z., Masia, B., Gutierrez, D.: Single-image full-body human relighting. CoRR abs/2107.07259 (2021), https://arxiv.org/abs/2107.07259" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 133, + 171, + 480, + 191 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 171, + 480, + 191 + ], + "spans": [ + { + "bbox": [ + 133, + 171, + 480, + 191 + ], + "type": "text", + "content": "19. LeGendre, C., Ma, W.C., Pandey, R., Fanello, S., Rhemann, C., Dourgarian, J., Busch, J., Debevec, P.: Learning illumination from diverse portraits (2020)" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 132, + 192, + 480, + 224 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 192, + 480, + 224 + ], + "spans": [ + { + "bbox": [ + 132, + 192, + 480, + 224 + ], + "type": "text", + "content": "20. Liu*, Y., Hou*, A., Huang, X., Ren, L., Liu, X.: Blind removal of facial foreign shadows. In: In Proceedings of British Machine Vision Conference (BMVC). London, UK (November 2022)" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 132, + 224, + 480, + 289 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 224, + 480, + 289 + ], + "spans": [ + { + "bbox": [ + 132, + 224, + 480, + 289 + ], + "type": "text", + "content": "21. Meka, A., Pandey, R., Haene, C., Orts-Escolano, S., Barnum, P., Davidson, P., Erickson, D., Zhang, Y., Taylor, J., Bouaziz, S., Legendre, C., Ma, W.C., Overbeck, R., Beeler, T., Debevec, P., Izadi, S., Theobalt, C., Rhemann, C., Fanello, S.: Deep relightable textures - volumetric performance capture with neural rendering. vol. 39 (December 2020). https://doi.org/10.1145/3414685.3417814, http://gvv.mpi-inf.mpg.de/projects/DeepRelightableTextures/" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 132, + 290, + 480, + 353 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 290, + 480, + 353 + ], + "spans": [ + { + "bbox": [ + 132, + 290, + 480, + 353 + ], + "type": "text", + "content": "22. Nestmeyer, T., Lalonde, J., Matthews, I., Lehrmann, A.: Learning physics-guided face relighting under directional light. In: 2020 IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR). pp. 5123-5132. IEEE Computer Society, Los Alamitos, CA, USA (jun 2020). https://doi.org/10.1109/CVPR42600.2020.00517, https://doi.ieeecomputersociety.org/10.1109/CVPR42600.2020.00517" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 132, + 354, + 480, + 396 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 354, + 480, + 396 + ], + "spans": [ + { + "bbox": [ + 132, + 354, + 480, + 396 + ], + "type": "text", + "content": "23. Pandey, R., Orts-Escolano, S., LeGendre, C., Haene, C., Bouaziz, S., Rhemann, C., Debevec, P., Fanello, S.: Total relighting: Learning to relight portraits for background replacement. vol. 40 (August 2021). https://doi.org/10.1145/3450626.3459872" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 132, + 397, + 480, + 439 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 397, + 480, + 439 + ], + "spans": [ + { + "bbox": [ + 132, + 397, + 480, + 439 + ], + "type": "text", + "content": "24. Peers, P., Tamura, N., Matusik, W., Debevec, P.: Post-production facial performance relighting using reflectance transfer. ACM Trans. Graph. 26(3), 52-es (jul 2007). https://doi.org/10.1145/1276377.1276442, https://doi.org/10.1145/1276377.1276442" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 132, + 441, + 480, + 462 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 441, + 480, + 462 + ], + "spans": [ + { + "bbox": [ + 132, + 441, + 480, + 462 + ], + "type": "text", + "content": "25. Ponglertnapakorn, P., Tritrong, N., Suwajanakorn, S.: Difareli: Diffusion face relighting (2023), https://arxiv.org/abs/2304.09479" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 132, + 463, + 480, + 483 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 463, + 480, + 483 + ], + "spans": [ + { + "bbox": [ + 132, + 463, + 480, + 483 + ], + "type": "text", + "content": "26. Qi, L., Wu, J., Wang, A.N., Wang, S., Sengupta, R.: My3dgen: A scalable personalized 3d generative model (2023)" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 132, + 483, + 480, + 504 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 483, + 480, + 504 + ], + "spans": [ + { + "bbox": [ + 132, + 483, + 480, + 504 + ], + "type": "text", + "content": "27. Qiu, H., Chen, Z., Jiang, Y., Zhou, H., Fan, X., Yang, L., Wu, W., Liu, Z.: Relitalk: Relightable talking portrait generation from a single video (2023)" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 132, + 505, + 480, + 525 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 505, + 480, + 525 + ], + "spans": [ + { + "bbox": [ + 132, + 505, + 480, + 525 + ], + "type": "text", + "content": "28. Ren, M., Xiong, W., Yoon, J.S., Shu, Z., Zhang, J., Jung, H., Gerig, G., Zhang, H.: Relightful harmonization: Lighting-aware portrait background replacement (2023)" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 132, + 526, + 480, + 556 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 526, + 480, + 556 + ], + "spans": [ + { + "bbox": [ + 132, + 526, + 480, + 556 + ], + "type": "text", + "content": "29. Ronneberger, O., Fischer, P., Brox, T.: U-net: Convolutional networks for biomedical image segmentation. CoRR abs/1505.04597 (2015), http://arxiv.org/abs/1505.04597" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 132, + 558, + 480, + 578 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 558, + 480, + 578 + ], + "spans": [ + { + "bbox": [ + 132, + 558, + 480, + 578 + ], + "type": "text", + "content": "30. Sengupta, S., Curless, B., Kemelmacher-Shlizerman, I., Seitz, S.M.: A light stage on every desk. CoRR abs/2105.08051 (2021), https://arxiv.org/abs/2105.08051" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 132, + 579, + 480, + 611 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 579, + 480, + 611 + ], + "spans": [ + { + "bbox": [ + 132, + 579, + 480, + 611 + ], + "type": "text", + "content": "31. Sengupta, S., Kanazawa, A., Castillo, C.D., Jacobs, D.W.: Sfsnet: Learning shape, reflectance and illuminance of faces in the wild'. In: Proceedings of the IEEE conference on computer vision and pattern recognition. pp. 6296-6305 (2018)" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 132, + 612, + 480, + 632 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 612, + 480, + 632 + ], + "spans": [ + { + "bbox": [ + 132, + 612, + 480, + 632 + ], + "type": "text", + "content": "32. Sevastopolsky, A., Ignatiev, S., Ferrer, G., Burnaev, E., Lempitsky, V.: Relightable 3d head portraits from a smartphone video (2020)" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 132, + 633, + 480, + 665 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 633, + 480, + 665 + ], + "spans": [ + { + "bbox": [ + 132, + 633, + 480, + 665 + ], + "type": "text", + "content": "33. Shih, Y., Paris, S., Barnes, C., Freeman, W.T., Durand, F.: Style transfer for headshot portraits. ACM Trans. Graph. 33(4) (jul 2014). https://doi.org/10.1145/2601097.2601137, https://doi.org/10.1145/2601097.2601137" + } + ] + } + ], + "index": 18 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "spans": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "type": "text", + "content": "16" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 166, + 91, + 220, + 100 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 166, + 91, + 220, + 100 + ], + "spans": [ + { + "bbox": [ + 166, + 91, + 220, + 100 + ], + "type": "text", + "content": "J. Choi et al." + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 15 + }, + { + "para_blocks": [ + { + "bbox": [ + 130, + 116, + 482, + 661 + ], + "type": "list", + "angle": 0, + "index": 19, + "blocks": [ + { + "bbox": [ + 130, + 116, + 482, + 149 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 116, + 482, + 149 + ], + "spans": [ + { + "bbox": [ + 130, + 116, + 482, + 149 + ], + "type": "text", + "content": "34. Shu, Z., Hadap, S., Shechtman, E., Sunkavalli, K., Paris, S., Samaras, D.: Portrait lighting transfer using a mass transport approach. ACM Trans. Graph. 37(1) (oct 2017). https://doi.org/10.1145/3095816, https://doi.org/10.1145/3095816" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 130, + 149, + 482, + 170 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 149, + 482, + 170 + ], + "spans": [ + { + "bbox": [ + 130, + 149, + 482, + 170 + ], + "type": "text", + "content": "35. Song, G., Cham, T.J., Cai, J., Zheng, J.: Half-body portrait relighting with overcomplete lighting representation (06 2021)" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 130, + 170, + 482, + 203 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 170, + 482, + 203 + ], + "spans": [ + { + "bbox": [ + 130, + 170, + 482, + 203 + ], + "type": "text", + "content": "36. Sun, T., Barron, J.T., Tsai, Y., Xu, Z., Yu, X., Fyffe, G., Rhemann, C., Busch, J., Debevec, P.E., Ramamoorthi, R.: Single image portrait relighting. CoRR abs/1905.00824 (2019), http://arxiv.org/abs/1905.00824" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 130, + 203, + 482, + 223 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 203, + 482, + 223 + ], + "spans": [ + { + "bbox": [ + 130, + 203, + 482, + 223 + ], + "type": "text", + "content": "37. Sun, T., Lin, K.E., Bi, S., Xu, Z., Ramamoorthi, R.: Nelf: Neural light-transport field for portrait view synthesis and relighting (2021)" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 130, + 223, + 482, + 265 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 223, + 482, + 265 + ], + "spans": [ + { + "bbox": [ + 130, + 223, + 482, + 265 + ], + "type": "text", + "content": "38. Sun, T., Xu, Z., Zhang, X., Fanello, S.R., Rhemann, C., Debevec, P.E., Tsai, Y., Barron, J.T., Ramamoorthi, R.: Light stage super-resolution: Continuous high-frequency relighting. CoRR abs/2010.08888 (2020), https://arxiv.org/abs/2010.08888" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 130, + 265, + 482, + 298 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 265, + 482, + 298 + ], + "spans": [ + { + "bbox": [ + 130, + 265, + 482, + 298 + ], + "type": "text", + "content": "39. Wang, Y., Holynski, A., Zhang, X., Zhang, X.: Sunstage: Portrait reconstruction and relighting using the sun as a light stage. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 20792-20802 (2023)" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 130, + 298, + 482, + 330 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 298, + 482, + 330 + ], + "spans": [ + { + "bbox": [ + 130, + 298, + 482, + 330 + ], + "type": "text", + "content": "40. Wang, Z., Yu, X., Lu, M., Wang, Q., Qian, C., Xu, F.: Single image portrait relighting via explicit multiple reflectance channel modeling. ACM Trans. Graph. 39(6) (Nov 2020). https://doi.org/10.1145/3414685.3417824" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 130, + 330, + 482, + 350 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 330, + 482, + 350 + ], + "spans": [ + { + "bbox": [ + 130, + 330, + 482, + 350 + ], + "type": "text", + "content": "41. Yang, H., Zheng, M., Feng, W., Huang, H., Lai, Y.K., Wan, P., Wang, Z., Ma, C.: Towards practical capture of high-fidelity relightable avatars (2023)" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 130, + 350, + 482, + 381 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 350, + 482, + 381 + ], + "spans": [ + { + "bbox": [ + 130, + 350, + 482, + 381 + ], + "type": "text", + "content": "42. Yeh, Y.Y., Nagano, K., Khamis, S., Kautz, J., Liu, M.Y., Wang, T.C.: 41(6), 1-21 (nov 2022). https://doi.org/10.1145/3550454.3555442, https://doi.org/10. 1145" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 130, + 381, + 482, + 415 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 381, + 482, + 415 + ], + "spans": [ + { + "bbox": [ + 130, + 381, + 482, + 415 + ], + "type": "text", + "content": "43. Yeh, Y.Y., Nagano, K., Khamis, S., Kautz, J., Liu, M.Y., Wang, T.C.: Learning to relight portrait images via a virtual light stage and synthetic-to-real adaptation. ACM Transactions on Graphics (TOG) (2022)" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 130, + 415, + 482, + 436 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 415, + 482, + 436 + ], + "spans": [ + { + "bbox": [ + 130, + 415, + 482, + 436 + ], + "type": "text", + "content": "44. Zhang, L., Zhang, Q., Wu, M., Yu, J., Xu, L.: Neural video portrait relighting in real-time via consistency modeling (2021)" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 130, + 436, + 482, + 500 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 436, + 482, + 500 + ], + "spans": [ + { + "bbox": [ + 130, + 436, + 482, + 500 + ], + "type": "text", + "content": "45. Zhang, R., Isola, P., Efros, A.A., Shechtman, E., Wang, O.: The unreasonable effectiveness of deep features as a perceptual metric. In: 2018 IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR). pp. 586-595. IEEE Computer Society, Los Alamitos, CA, USA (jun 2018). https://doi.org/10.1109/CVPR.2018.00068, https://doi.ieeecomputersociety.org/10.1109/CVPR.2018.00068" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 130, + 500, + 482, + 532 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 500, + 482, + 532 + ], + "spans": [ + { + "bbox": [ + 130, + 500, + 482, + 532 + ], + "type": "text", + "content": "46. Zhang, R., Isola, P., Efros, A.A., Shechtman, E., Wang, O.: The unreasonable effectiveness of deep features as a perceptual metric. CoRR abs/1801.03924 (2018), http://arxiv.org/abs/1801.03924" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 130, + 532, + 482, + 575 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 532, + 482, + 575 + ], + "spans": [ + { + "bbox": [ + 130, + 532, + 482, + 575 + ], + "type": "text", + "content": "47. Zhang, X., Fanello, S.R., Tsai, Y., Sun, T., Xue, T., Pandey, R., Orts-Escalano, S., Davidson, P.L., Rhemann, C., Debevec, P.E., Barron, J.T., Ramamoorthi, R., Freeman, W.T.: Neural light transport for relighting and view synthesis. CoRR abs/2008.03806 (2020), https://arxiv.org/abs/2008.03806" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 130, + 575, + 482, + 596 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 575, + 482, + 596 + ], + "spans": [ + { + "bbox": [ + 130, + 575, + 482, + 596 + ], + "type": "text", + "content": "48. Zhang, X., Barron, J.T., Tsai, Y.T., Pandey, R., Zhang, X., Ng, R., Jacobs, D.E.: Portrait shadow manipulation. vol. 39 (2020)" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 130, + 596, + 482, + 628 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 596, + 482, + 628 + ], + "spans": [ + { + "bbox": [ + 130, + 596, + 482, + 628 + ], + "type": "text", + "content": "49. Zhou, H., Hadap, S., Sunkavalli, K., Jacobs, D.: Deep single-image portrait relighting. In: 2019 IEEE/CVF International Conference on Computer Vision (ICCV). pp. 7193-7201 (2019). https://doi.org/10.1109/ICCV.2019.00729" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 130, + 628, + 482, + 661 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 628, + 482, + 661 + ], + "spans": [ + { + "bbox": [ + 130, + 628, + 482, + 661 + ], + "type": "text", + "content": "50. Zhu, J.Y., Park, T., Isola, P., Efros, A.A.: Unpaired image-to-image translation using cycle-consistent adversarial networks. In: Computer Vision (ICCV), 2017 IEEE International Conference on (2017)" + } + ] + } + ], + "index": 18 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 198, + 91, + 447, + 102 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 198, + 91, + 447, + 102 + ], + "spans": [ + { + "bbox": [ + 198, + 91, + 447, + 102 + ], + "type": "text", + "content": "Personalized Video Relighting With an At-Home Light Stage" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 470, + 92, + 480, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 470, + 92, + 480, + 100 + ], + "spans": [ + { + "bbox": [ + 470, + 92, + 480, + 100 + ], + "type": "text", + "content": "17" + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 16 + } + ], + "_backend": "vlm", + "_version_name": "2.6.4" +} \ No newline at end of file diff --git a/2024/PetFace_ A Large-Scale Dataset and Benchmark for Animal Identification/79b48743-23c5-4624-b76c-ce49a285a1a7_content_list.json b/2024/PetFace_ A Large-Scale Dataset and Benchmark for Animal Identification/79b48743-23c5-4624-b76c-ce49a285a1a7_content_list.json new file mode 100644 index 0000000000000000000000000000000000000000..a4c0ff62bb3add5cb081568eccf40d738291aad6 --- /dev/null +++ b/2024/PetFace_ A Large-Scale Dataset and Benchmark for Animal Identification/79b48743-23c5-4624-b76c-ce49a285a1a7_content_list.json @@ -0,0 +1,1801 @@ +[ + { + "type": "text", + "text": "PetFace: A Large-Scale Dataset and Benchmark for Animal Identification", + "text_level": 1, + "bbox": [ + 228, + 140, + 776, + 184 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Risa Shinoda $^{1*}$ and Kaede Shiohara $^{2*}$", + "bbox": [ + 344, + 212, + 656, + 227 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "$^{*}$ Equal contribution \n $^{1}$ Kyoto University, Japan $^{2}$ The University of Tokyo, Japan", + "bbox": [ + 297, + 239, + 704, + 268 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Abstract. Automated animal face identification plays a crucial role in the monitoring of behaviors, conducting of surveys, and finding of lost animals. Despite the advancements in human face identification, the lack of datasets and benchmarks in the animal domain has impeded progress. In this paper, we introduce the PetFace dataset, a comprehensive resource for animal face identification encompassing 257,484 unique individuals across 13 animal families and 319 breed categories, including both experimental and pet animals. This large-scale collection of individuals facilitates the investigation of unseen animal face verification, an area that has not been sufficiently explored in existing datasets due to the limited number of individuals. Moreover, PetFace also has fine-grained annotations such as sex, breed, color, and pattern. We provide multiple benchmarks including re-identification for seen individuals and verification for unseen individuals. The models trained on our dataset outperform those trained on prior datasets, even for detailed breed variations and unseen animal families. Our result also indicates that there is some room to improve the performance of integrated identification on multiple animal families. We hope the PetFace dataset will facilitate animal face identification and encourage the development of non-invasive animal automatic identification methods. Our dataset and code are available at https://dahlian00.github.io/PetFacePage/.", + "bbox": [ + 261, + 308, + 738, + 599 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Keywords: Animals $\\cdot$ Re-identification $\\cdot$ Face Recognition", + "bbox": [ + 261, + 613, + 656, + 627 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "1 Introduction", + "text_level": 1, + "bbox": [ + 217, + 655, + 374, + 670 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Animal identification plays a crucial role in animal studies and applications such as monitoring animal behavior, conducting habitat surveys, locating missing animals, and performing health checks. Traditional identification techniques, including ear tags, tattoos, ear punching, and toe clipping, continue to be utilized mainly for experimental animals and livestock [13]. However, given that these methods have the potential to cause stress and pain [37,45,56], their use should be minimized to prioritize animal welfare. Therefore, there is a pressing need for the development and adoption of identification technologies that are not only effective and efficient but also minimally invasive, thereby mitigating the ethical concerns associated with traditional methods. While advanced tools such as", + "bbox": [ + 212, + 688, + 787, + 839 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "digital IDs [34, 55] have been introduced, their applications involve a laborious process, i.e., attaching the devices to each animal individually. The process is costly and potentially stressful for the animals. Moreover, these physical tags can identify only pre-defined individuals, which makes them impractical for use in real world scenarios.", + "bbox": [ + 212, + 146, + 782, + 220 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "In the human domain, digital face recognition is one of the effective approaches for the identification. It has been developed for use in smartphones, airport security, and systems for finding missing people. Therefore, the research community made great efforts to develop sophisticated deep learning-based face recognition models [15,26,33,40,63,65], empowered by large-scale datasets and benchmarks, e.g., [21,29,48,57].", + "bbox": [ + 212, + 222, + 784, + 311 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Despite the promise of human face recognition, the research progress towards automatic animal face individual recognition has been impeded, primarily because of the lack of extensive datasets and benchmarks for animal face recognition. Previous openly available datasets mostly include less than 100 individuals [10, 19, 30, 47, 66], which makes it far from generalized and discriminative identification models and precise evaluation for unseen individuals.", + "bbox": [ + 212, + 313, + 784, + 402 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "In this paper, we introduce a large-scale animal face recognition dataset called PetFace that contains 257,484 individuals in total across 13 species with 319 breeds with 1,012,934 images. We show the example images of our PetFace in Fig 1. The number of individuals in our dataset is over 110 times that in the previous largest animal face dataset [49]. We sourced images and related information from the internet, with automated and manual filtering processes applied to ensure the dataset is not only large but also finely detailed and clean. Moreover, PetFace has fine-grained annotations including sex, breeds, and colors and patterns of their skin, which allows further investigation for fine-grained recognition and evaluation.", + "bbox": [ + 212, + 404, + 784, + 551 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "PetFace offers two benchmarks: one for recognizing known (seen) individuals and the other for recognizing unknown (unseen) ones. We also conduct the verification of the fine-grade breeds and unseen animal categories.", + "bbox": [ + 212, + 555, + 784, + 599 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Our main contributions are as follows: (i) We establish a new dataset for animal face recognition called PetFace, which contains a total of 257,484 individuals across 13 types of animal families and 319 breeds with fine-grade annotations including sex, breed, color of animals. (ii) We set the benchmarks on recognizing known (seen) individuals and unknown (unseen) ones. (iii) We show that the model trained using our dataset shows the generalization capabilities for unseen individuals and even for unseen animal categories.", + "bbox": [ + 212, + 601, + 784, + 705 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "2 Related Work", + "text_level": 1, + "bbox": [ + 215, + 729, + 387, + 747 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Building datasets and benchmarks is an important step in advancing animal re-identification through deep models. While earlier efforts have established the basics of animal face recognition, there is considerable potential for further development. Compared to the dataset for human face recognition [21, 29, 48, 57, 67], those for animal faces [10, 19, 49, 66] have much fewer individuals. The evaluation", + "bbox": [ + 212, + 763, + 784, + 839 + ], + "page_idx": 1 + }, + { + "type": "page_number", + "text": "2", + "bbox": [ + 217, + 114, + 228, + 126 + ], + "page_idx": 1 + }, + { + "type": "header", + "text": "R. Shinoda and K. Shiohara", + "bbox": [ + 271, + 114, + 460, + 127 + ], + "page_idx": 1 + }, + { + "type": "image", + "img_path": "images/e278c278bd4ea741a38a64d8b494fa4b99446d2f4e496c78560e96235f07c450.jpg", + "image_caption": [ + "Fig. 1: Example images of our PetFace. We introduce a large-scale animal face re-identification dataset PetFace that include 257,484 unique individuals across 13 families and 319 breeds. From the left, the images represent Cat, Chimpanzee, Chinchilla, Degus, Dog, Ferret, Guinea pig, Hamster, Hedgehog, Parakeet, Java sparrow, Pig, and Rabbit. The four images enclosed within the white square's grid lines represent the same identity." + ], + "image_footnote": [], + "bbox": [ + 225, + 146, + 764, + 318 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "scenarios represented in these datasets often fall short of real-world applicability; most of the datasets focus on closed-set re-identification, rather than recognizing unseen individuals that is a critical requirement for practical applications. Our work enables the training and evaluation on a large number of individuals across a wider range of animal families and breeds.", + "bbox": [ + 212, + 431, + 787, + 505 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Human Face Identification is the process of identifying an individual's identity using their unique facial characteristics. The huge demands for individual identification have grown mainly in the human domain. In the identification of human faces, which is known as face recognition, considerable efforts have been expended in the research community [15,26,33,40,63,65]. Because of their domain-agnostic frameworks, most of the state-of-the-art methods in face recognition can be exported into other domains, such as animal identification. Furthermore, these advancements have been supported by the introduction of large-scale datasets and benchmarks [21, 29, 48, 57, 67] that have played a crucial role in facilitating the exploration of powerful models. Motivated by this, we create the PetFace dataset to fill the gap between human face recognition and animal face recognition.", + "bbox": [ + 212, + 507, + 787, + 686 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Animal Identification is the process of identifying an individual's identity using their unique body or facial characteristics, which is an important task across various scenarios, including monitoring animals, conducting habitat surveys, and finding missing animals. With the advance of computer vision technology, various openly available datasets contribute to computer vision for animals [6-8, 12, 18, 20, 24, 31, 38, 46, 50, 59, 60, 64]. Various methods are used to create the dataset, such as recording images [12, 18, 24, 53, 60] and videos [22, 38, 46, 50, 64, 68], and using aerial image [6, 7, 20]. Recording individual information is labor intensive; therefore, most of the previous datasets contain only a limited number of individuals and are focused on one species. Recently, WildlifeDatasets [61], which", + "bbox": [ + 212, + 688, + 787, + 840 + ], + "page_idx": 2 + }, + { + "type": "header", + "text": "PetFace: A Large-Scale Dataset and Benchmark for Animal Identification", + "bbox": [ + 241, + 114, + 730, + 127 + ], + "page_idx": 2 + }, + { + "type": "page_number", + "text": "3", + "bbox": [ + 774, + 116, + 784, + 126 + ], + "page_idx": 2 + }, + { + "type": "table", + "img_path": "images/c74f428db1b7be9cd10d348d0f8adc4fd18c86fecf66ffc04df04957401cc25d.jpg", + "table_caption": [], + "table_footnote": [], + "table_body": "
DatasetFamily#Individuals#ImagesFine-grained annotation
CTai [19]Chimp785078Age, Sex
CZoo [19]Chimp242109Age, Sex
MacaqueFaces [66]Monkey346280-
BristolGorillas2020 [10]Gorilla75428-
DogFaceNet [49]Dog13938363-
Flickr-dog [47]Dog42374-
THODBRL2015 [30]Horse472820Angle
ZindiTurtleRecall [4]Turtle226512803Angle
SeaTurtleID2022 [5]Turtle4388729Timestamp
Ours13 families257,4841,012,934Sex, Breed, Color & Pattern
", + "bbox": [ + 215, + 143, + 802, + 311 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Table 1: Openly available animal face identification datasets. Compared to other datasets, PetFace has the largest number of individuals, and a wider range of families and breeds.", + "bbox": [ + 215, + 311, + 789, + 351 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "gathers previously openly available datasets [1-4,6-9,11,12,16,18-20,22,24,27, 28,36,38,46,50-53,60,62,64,66,68,69] combined, was introduced. This research creates the benchmarks using existing available datasets.", + "bbox": [ + 212, + 380, + 785, + 425 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "The advances in human face recognition raise an interest in animal face recognition. One notable challenge is the variation in facial structures between animals and humans. This has led to studies like AnimalWeb [32] and CatFLW [44] that propose specialized methods for animal facial key points detection. Automatic animal face identification has been studied, which can help humans to monitor animals [25, 39, 42, 43, 58]. We review several publicly available face identification datasets in Table 1. Apes, such as Chimpanzees [19] and Gorillas [10], are one of the species that have been studied. However, as the dataset size is relatively limited, the datasets are insufficient to evaluate new unseen individual faces. In addition to primates, research on facial (head) identification has also extended to turtles, with the ZindiTurtleRecall [4] dataset offering a substantial number of images of individual turtles in a controlled environment. To address the limitations of in-the-wild data collection, the SeaTurtleID2022 [5] dataset includes wild data, albeit with a reduced number of individuals, owing to the extensive effort required for data collection. Among datasets containing mammals, the DogFaceNet dataset [49] featuring 1,393 individual dogs stands out for its size. Nevertheless, the significant variation in appearance across dog breeds suggests that even this larger dataset may not be sufficiently comprehensive. In addition to the limited number of individuals, previous animal face datasets often focus on only single animal families. They can contribute to specific animal family research but can not explore animal identification across many families and breeds.", + "bbox": [ + 212, + 426, + 787, + 758 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "3 PetFace Dataset", + "text_level": 1, + "bbox": [ + 215, + 779, + 410, + 794 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "PetFace is a large animal face identification dataset that expands research on animal face recognition, which has been impeded by a scarcity of suitable datasets", + "bbox": [ + 212, + 809, + 785, + 839 + ], + "page_idx": 3 + }, + { + "type": "page_number", + "text": "4", + "bbox": [ + 217, + 114, + 228, + 126 + ], + "page_idx": 3 + }, + { + "type": "header", + "text": "R. Shinoda and K. Shiohara", + "bbox": [ + 271, + 114, + 460, + 127 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "and benchmarks. This section details the construction of the PetFace dataset, including our labor-efficient methods for collecting animal face images and a semi-automated filtering process to ensure quality fine-grained categorization and statistics of the dataset.", + "bbox": [ + 212, + 146, + 787, + 205 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "3.1 Dataset Statistics", + "text_level": 1, + "bbox": [ + 215, + 229, + 408, + 243 + ], + "page_idx": 4 + }, + { + "type": "image", + "img_path": "images/0f27eb20c8632dfa7797be3242c655f47e0a7052738fd419adc8a929646ecaae.jpg", + "image_caption": [ + "(a) Individual distribution" + ], + "image_footnote": [], + "bbox": [ + 222, + 282, + 506, + 422 + ], + "page_idx": 4 + }, + { + "type": "image", + "img_path": "images/1a0ecff7a365b6e7cc328c4333ae9d327b94127160c2ae3f53ec2fb55a39d122.jpg", + "image_caption": [ + "(b) Sex distribution" + ], + "image_footnote": [], + "bbox": [ + 511, + 282, + 800, + 425 + ], + "page_idx": 4 + }, + { + "type": "image", + "img_path": "images/f9693d6e47880e64254c39c50e81fe1f6d914d08f00356bc1442edddf64bf47d.jpg", + "image_caption": [ + "(c) Breed annotation examples" + ], + "image_footnote": [], + "bbox": [ + 254, + 452, + 418, + 589 + ], + "page_idx": 4 + }, + { + "type": "image", + "img_path": "images/49c5cf42f668e584119b5044185027d61611b7fe37ba5d2c479a6baac7bba7a5.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 433, + 452, + 575, + 589 + ], + "page_idx": 4 + }, + { + "type": "image", + "img_path": "images/bc92ab628140b13619e1be092caa67ee40660ca78e28dd92c19d9ee71e912ca9.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 602, + 452, + 743, + 589 + ], + "page_idx": 4 + }, + { + "type": "image", + "img_path": "images/40cb436f1e5bab23e3850bf21b40418f4b39fdd502fc000d83295300609535f6.jpg", + "image_caption": [ + "Fig. 2: Dataset distribution. Each figure represents (a) the number of individuals in each animal family. (b) the sex distribution percentage by animal family. (c) examples of breed annotations. (d) examples of color annotations." + ], + "image_footnote": [], + "bbox": [ + 230, + 625, + 398, + 760 + ], + "page_idx": 4 + }, + { + "type": "image", + "img_path": "images/1d2a7a5b2af52220fb7036713f277dd2e7db3a3e525510caa810ff8e8e0932ad.jpg", + "image_caption": [ + "(d) Color annotation examples" + ], + "image_footnote": [], + "bbox": [ + 426, + 626, + 576, + 758 + ], + "page_idx": 4 + }, + { + "type": "image", + "img_path": "images/a57b023c7b588e33d9f61b895fd0c949f60a2c49f419acf2a24b0e289e81784f.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 601, + 626, + 756, + 757 + ], + "page_idx": 4 + }, + { + "type": "header", + "text": "PetFace: A Large-Scale Dataset and Benchmark for Animal Identification", + "bbox": [ + 241, + 114, + 730, + 128 + ], + "page_idx": 4 + }, + { + "type": "page_number", + "text": "5", + "bbox": [ + 774, + 116, + 784, + 126 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "The PetFace dataset encompasses 1,012,934 images spanning 257,484 unique individuals. Detailed distributions of animal families are illustrated in Fig 2(a). Images are cropped to $224 \\times 224$ pixels around their faces. Our PetFace also has fine-grained annotations. Sex distribution across different animal families is depicted in Fig 2(b), with sex information available for 240,861 individuals, accounting for $94\\%$ of the dataset. The dataset includes annotations for 319 breeds, with examples of breed annotations shown in Fig 2(c). Furthermore, as detailed in Fig 2(d), the dataset provides in-depth color information through two-tier hierarchical annotations. Please see the supplemental for the detailed information per each animal family.", + "bbox": [ + 212, + 146, + 787, + 297 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "3.2 Data Sourcing", + "text_level": 1, + "bbox": [ + 215, + 318, + 382, + 333 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Collecting images of animal faces via photography is labor and time intensive, which impedes the creation of large-scale datasets. In contrast, the human face recognition field benefits greatly from the availability of images sourced from the Internet. In this section, we outline our approach to assembling the collection of animal face images through the Internet, where each is associated with unique individual identifiers.", + "bbox": [ + 212, + 342, + 787, + 431 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Curation of images. In contrast to the human domain, acquiring multiple images for individual animals is more challenging. Unlike human datasets where a large number of celebrity images can be readily sourced, animal images require alternative approaches for curation. We utilized two primary sources: (i) pet shops' websites and (ii) animal adoption websites. The advantage of pet shops lies in their provision of high-quality, diverse images, capturing individuals from various angles and offering detailed information about each animal, including color, sex, and specific breed details. On the other hand, animal adoption sites offer images set against a variety of backgrounds and conditions that are often provided by pet owners, thus ensuring each individual is presented in a unique setting. These sources provide images that are highly suitable for animal recognition tasks and that are especially useful for recognizing animals in varied wild environments. For Chimps, we additionally use images from the webpage of a collaborative research institution.", + "bbox": [ + 212, + 433, + 787, + 643 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "To ensure the dataset's quality, we were selective in choosing websites for curation. We only use the websites introducing each animal on one page to gain the individual IDs. Aware of the potential for pet owners to upload the same images to multiple websites, we chose a single pet adoption website from each region (e.g., one per country) to minimize duplicates. For pet shops, we confirmed that the animals listed were unique to a certain shop to ensure the uniqueness of our dataset entries. Using these sources, we collected 1,443,737 images from 325,420 individuals.", + "bbox": [ + 212, + 645, + 787, + 765 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "3.3 Face Alignment and Filtering", + "text_level": 1, + "bbox": [ + 215, + 785, + 506, + 801 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Face Detection. We detect facial landmarks to align and crop the images. We adopt the AnyFace [35] that is trained on mixed face datasets including", + "bbox": [ + 212, + 809, + 785, + 840 + ], + "page_idx": 5 + }, + { + "type": "page_number", + "text": "6", + "bbox": [ + 217, + 114, + 228, + 126 + ], + "page_idx": 5 + }, + { + "type": "header", + "text": "R. Shinoda and K. Shiohara", + "bbox": [ + 271, + 114, + 460, + 127 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/ffc1f5b9fd830e0de23a28d0bc1c7af993bade2c59de8b24208e9d8c074a9703.jpg", + "image_caption": [ + "Fig. 3: Data filtering process. We adopt a two-stage data filtering approach. Initially, images containing multiple faces are automatically removed. Subsequently, any images that do not depict animal faces or fail to meet alignment criteria are manually eliminated." + ], + "image_footnote": [], + "bbox": [ + 218, + 142, + 802, + 333 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "AnimalWeb [32]. Because the positions of animal facial parts are sometimes very different from those of humans, we use different reference points for different species in face alignment. After landmark detection, we select one frontal image as a reference for each species. We compute the average landmarks over all the landmarks aligned with the reference. Then, we define the target positions of the landmarks in the images; all the images are aligned to the target positions.", + "bbox": [ + 212, + 431, + 787, + 521 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Data Filtering. Fig 3 shows the overview of our data filtering process. Because the fully automatic face detection above sometimes fails, we adopt a two-stage data filtering process to filter out the following cases: 1) Images contain multiple animals simultaneously. 2) Images are unrelated to animals, such as advertisements. 3) Images focus on non-animal elements, such as random patterns in backgrounds, toys, or people, rather than the animals themselves. First, we automatically remove images where multiple faces are detected in an image. Secondly, annotators manually assess all the images and remove those that do not depict the target animals or where the face alignment differs from our intended criteria. Because checking whether the faces are properly aligned or not requires some expertise, all the images are filtered by the authors to ensure the quality of the dataset. This manual filtering process took about 100 man hours. After this stage, we keep 1,012,934 images, which is approximately $70\\%$ of the initial image number. Detailed distributions of animal families are illustrated in Fig 2(a).", + "bbox": [ + 212, + 522, + 787, + 734 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "3.4 Fine-Grained Annotations", + "text_level": 1, + "bbox": [ + 215, + 755, + 478, + 768 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Fine-grained categorization enhances its usefulness for downstream applications, e.g., the creation of challenging datasets based on animal's individual attributes. However, collecting extensive individual animal data poses a major challenge in assembling large datasetse owing to the significant effort involved. In response,", + "bbox": [ + 212, + 779, + 787, + 840 + ], + "page_idx": 6 + }, + { + "type": "header", + "text": "PetFace: A Large-Scale Dataset and Benchmark for Animal Identification", + "bbox": [ + 241, + 114, + 730, + 128 + ], + "page_idx": 6 + }, + { + "type": "page_number", + "text": "7", + "bbox": [ + 774, + 116, + 784, + 126 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "we collect as much individual animal information corresponding to the images as possible. Our annotations include sex, color, and specific breed details, depending on their availability on the source websites. The annotation process involves: 1) extracting individual information displayed on websites alongside images, 2) refining raw text data to develop individual data tables, 3) manually verifying that all attributes are refined and free of unrelated information, and 4) manually annotating individuals with missing attributes on the website.", + "bbox": [ + 212, + 146, + 782, + 251 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "As a result, we include a sex category for all classes, with breed annotations for eight animal classes (Cat, Chimp, Dog, Guinea pig, Hamster, Parakeet, Pig, and Rabbit), and color annotations for eleven animal classes (Cat, Chinchilla, Degus, Ferret, Guinea pig, Hamster, Hedgehog, Parakeet, Java sparrow, Pig, and Rabbit). Note that the manual annotations are performed only on colors and patterns because images alone do not allow us to accurately determine an animal's sex or breed. Detailed distributions of sex per animal families are illustrated in Fig 2(b). In Fig 2(c) and (d), we illustrate the examples of breed annotations and color and pattern annotations.", + "bbox": [ + 212, + 252, + 784, + 388 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "4 Experimental Methodology", + "text_level": 1, + "bbox": [ + 215, + 412, + 516, + 430 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "We introduce two principal evaluation protocols: 1) re-identification for seen faces, and 2) verification for unseen faces. To establish these benchmarks, we have curated two distinct types of test sets. Specifically, for unseen face verification, we carefully separate the dataset into training, validation, and test sets to prevent data leakage that can arise from shared backgrounds.", + "bbox": [ + 212, + 445, + 784, + 521 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "4.1 Evaluation Protocols", + "text_level": 1, + "bbox": [ + 215, + 544, + 433, + 558 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "We adopt two types of evaluation.", + "bbox": [ + 215, + 570, + 464, + 585 + ], + "page_idx": 7 + }, + { + "type": "list", + "sub_type": "text", + "list_items": [ + "1. Re-identification: This evaluation selects images from the test data that match the identities present in the training data. This procedure involves identifying test images that correspond to the same identities used during model training. The objective is to assess the model's ability to accurately recognize and associate test images with the correct identities from the training dataset.", + "2. Verification: This evaluation checks if the model can identify unseen faces. This means that we use different identities during the training and testing phases. For each identity, we select one image that matches the identity and one image randomly chosen from a different identity. This process creates pairs of images where one pair consists of images from the same identity, and another pair consists of images from different identities. We then task the model with predicting whether the faces in each pair belong to the same identity or different identities. This approach allows us to comprehensively evaluate the model's ability to recognize and differentiate between individual identities." + ], + "bbox": [ + 222, + 598, + 785, + 838 + ], + "page_idx": 7 + }, + { + "type": "page_number", + "text": "8", + "bbox": [ + 217, + 114, + 228, + 126 + ], + "page_idx": 7 + }, + { + "type": "header", + "text": "R. Shinoda and K. Shiohara", + "bbox": [ + 271, + 114, + 460, + 127 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "4.2 Data Split", + "text_level": 1, + "bbox": [ + 215, + 146, + 346, + 161 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "To establish benchmarks for seen individuals re-identification and unseen individuals verification, we use specific split protocols, creating two test sets designed for each task. For seen individuals re-identification, the process involves verifying images of the same identities used in training; thus, we select test images from the same individuals for the training set. We, therefore, ensure that each individual is represented by at least three photos i.e., a minimum of two for training and one for testing. For unseen individuals verification dataset, we ensure that it contains no images from sources that are also represented in the training or validation sets per animal family to prevent bias from similar environmental conditions. We divide the dataset into training, validation, and each testing sets following a 7:1:2 ratio as closely as possible, given the outlined criteria.", + "bbox": [ + 212, + 176, + 787, + 343 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "4.3 Models", + "text_level": 1, + "bbox": [ + 215, + 369, + 323, + 383 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "We train state-of-the-art models based on deep neural networks on our PetFace to build the benchmarks. For the re-identification task, we approach the training and testing phases as classification problems; an individual is assigned to a class. On the other hand, for the verification task, we train models in the same manner as for the re-identification task but evaluate the models by computing the cosine similarity between pairs of images to be identified if they are the same individual. Specifically, we focus more on the loss functions that are crucial for identification tasks than network architectures. We refer to three important loss functions in addition to the basic Softmax-based classification model:", + "bbox": [ + 212, + 398, + 787, + 534 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "Triplet loss [26] is designed to take a triplet of samples, i.e., an anchor $x_{a}$ , a positive $x_{p}$ (another image of the same identity as the anchor), and a negative $x_{n}$ (an image of a different identity) - and learn embeddings in such a way that the $x_{p}$ is closer to the $x_{a}$ than the $x_{n}$ by a margin. This loss function is particularly beneficial for face Re-ID as it directly targets the relative distances between different and same identity pairs, encouraging the model to learn a feature space where embeddings of the same identity are clustered together while being far from other identity clusters.", + "bbox": [ + 212, + 536, + 787, + 656 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "Center loss [65] works alongside Softmax Loss to enhance the discriminative power of the learned features. While Softmax Loss focuses on inter-class separability, Center Loss aims to minimize the intra-class variations. It does this by penalizing the distance between the deep features of each class and their corresponding class center. Center Loss ensures that the embeddings of the faces of the same individual are closer together, thus making the feature distribution more compact for each identity.", + "bbox": [ + 212, + 657, + 787, + 763 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "ArcFace loss [15] introduces an angular margin between classes to enforce a discriminative feature space. It modifies the Softmax loss by adding a margin penalty to the angle between the feature vector and the corresponding class center in the angular space. This angular margin encourages models to learn more distinguishable embeddings to separate classes.", + "bbox": [ + 212, + 763, + 787, + 839 + ], + "page_idx": 8 + }, + { + "type": "header", + "text": "PetFace: A Large-Scale Dataset and Benchmark for Animal Identification", + "bbox": [ + 241, + 114, + 730, + 128 + ], + "page_idx": 8 + }, + { + "type": "page_number", + "text": "9", + "bbox": [ + 774, + 116, + 784, + 126 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "We use ResNet-50 [23], which we found performs better than recent Transformer-based models in Table 5, as our base backbone for all applied loss functions to simplify our experiments and make them easier to grasp.", + "bbox": [ + 212, + 146, + 782, + 191 + ], + "page_idx": 9 + }, + { + "type": "text", + "text": "In addition to the comparison on the models trained on PetFace, we evaluate state-of-the-art models trained on other datasets including:", + "bbox": [ + 212, + 191, + 782, + 222 + ], + "page_idx": 9 + }, + { + "type": "text", + "text": "ImageNet [14] is a conventional image classification dataset including 1000 general object classes. We use the ResNet-50 trained on the dataset.", + "bbox": [ + 212, + 222, + 782, + 252 + ], + "page_idx": 9 + }, + { + "type": "text", + "text": "CLIP [54] learns semantic relationships between images and texts in a cross-modal contrastive learning manner empowered by web-scale image-caption datasets. We use ResNet-50 backbone.", + "bbox": [ + 212, + 252, + 802, + 297 + ], + "page_idx": 9 + }, + { + "type": "text", + "text": "MegaDescriptor [61] is a state-of-the-art model for animal identification trained on a unified animal identification dataset including 33 existing animal re-identification datasets [1-4, 6-9, 11, 12, 16, 18-20, 22, 24, 27, 28, 36, 38, 46, 50-53, 60, 62, 64, 66, 68, 69]. We use the officially provided SwinTransformer-B [41] backbone.", + "bbox": [ + 212, + 297, + 787, + 359 + ], + "page_idx": 9 + }, + { + "type": "text", + "text": "5 Experimental Result", + "text_level": 1, + "bbox": [ + 214, + 383, + 450, + 400 + ], + "page_idx": 9 + }, + { + "type": "text", + "text": "5.1 Benchmark on Animal Face Re-identification", + "text_level": 1, + "bbox": [ + 214, + 416, + 632, + 430 + ], + "page_idx": 9 + }, + { + "type": "text", + "text": "We show the re-identification results in terms AUC in Table 2. We train models with the baseline loss functions independently on each class. We can see that ArcFace performs consistent results, $51.23\\%$ of average accuracy, compared to the other loss functions (41.88% for Softmax and 9.81% for Center). On the other hand, Center loss does not learn sufficient discriminative features for animal face re-identification, especially on Cat and Dog where the number of the test classes are 113,592 and 46,755, respectively. These results encourage the community to explore more effective representation learning methods for re-identification tasks.", + "bbox": [ + 212, + 441, + 782, + 575 + ], + "page_idx": 9 + }, + { + "type": "text", + "text": "Moreover, motivated by MegaDescriptor, we train an ArcFace model on the entire PetFace, denoted as Joint-Trained on PetFace in the table. We observe that the joint-training strategy provides some improvements, e.g., from $54.29\\%$ to $70.30\\%$ for Cat and from $29.08\\%$ to $41.49\\%$ for Pig although in some classes the results get worse, e.g., from $43.27\\%$ to $34.30\\%$ for Chimp and from $62.19\\%$ to $44.78\\%$ to Parakeet. This result indicates that there is some room to improve the performance of integrated identification on imbalance and wide-range datasets.", + "bbox": [ + 212, + 579, + 782, + 683 + ], + "page_idx": 9 + }, + { + "type": "text", + "text": "In Fig 4, we display the Top- $k$ ( $k = 1,3,5$ ) accuracy metrics. While accuracy naturally increases with larger $k$ values, the relative ranking of accuracy across different animal families remains largely consistent. ArcFace maintains the best performance across these evaluations.", + "bbox": [ + 212, + 684, + 782, + 744 + ], + "page_idx": 9 + }, + { + "type": "text", + "text": "5.2 Benchmark on Animal Face Verification", + "text_level": 1, + "bbox": [ + 214, + 768, + 589, + 782 + ], + "page_idx": 9 + }, + { + "type": "text", + "text": "Next, we show the verification results in Table 3. We use the trained models in Table 2 and evaluate them on unseen individuals. Similar to our findings in the re-identification task, ArcFace proved to be effective, achieving the best", + "bbox": [ + 212, + 794, + 782, + 839 + ], + "page_idx": 9 + }, + { + "type": "page_number", + "text": "10", + "bbox": [ + 217, + 114, + 235, + 126 + ], + "page_idx": 9 + }, + { + "type": "header", + "text": "R. Shinoda and K. Shiohara", + "bbox": [ + 271, + 114, + 460, + 127 + ], + "page_idx": 9 + }, + { + "type": "table", + "img_path": "images/9497bce529f21be67ff21894144ca61164219dfb7034b8dbd7415c946ffd99b7.jpg", + "table_caption": [], + "table_footnote": [], + "table_body": "
MethodTop-1 Accuracy (%)Avg
CatChimpChinchillaDegusDogFerretGuineaHamsterHedgehogParakeetJava sparrowPigRabbit
Trained on PetFace
Softmax30.4641.7058.1337.8459.1425.8860.0738.2727.8150.8833.5521.6459.0541.88
Center0.005.3829.7612.030.007.9131.779.4613.761.684.867.443.519.81
Arcface54.2943.2767.3445.4177.8628.9267.9047.3730.9062.1942.2729.0869.1351.23
Joint-Trained on PetFace
ArcFace70.3034.3069.8656.0868.7546.1268.6654.3344.3844.7834.5541.4965.7553.80
", + "bbox": [ + 215, + 143, + 785, + 237 + ], + "page_idx": 10 + }, + { + "type": "table", + "img_path": "images/a742a8c1775f673c63046ff44bb0c93656fda4c529395098795b05c88b9db41c.jpg", + "table_caption": [ + "Table 2: Result on animal re-identification. ArcFace outperforms the other loss functions. The jointly trained model on our PetFace shows the best average top-1 accuracy." + ], + "table_footnote": [], + "table_body": "
MethodAUC (%)Avg
CatChimpChinchillaDegusDogFerretGuinea pigHamsterHedgehogParakeetJava sparrowPigRabbit
Trained on PetFace
Softmax97.9785.2284.4486.1998.9883.1595.3087.7387.2798.3984.6787.0998.5390.38
Center50.2081.6080.9681.3459.6478.7792.1083.5985.2696.9180.5482.5897.0180.81
Triplet96.9477.1076.1272.8097.9779.4883.3780.0881.4791.5677.8675.7294.7483.48
Arcface97.7183.7687.7087.6199.4586.2496.0390.3286.1398.6384.4589.8899.0191.30
Joint-Trained on PetFace
ArcFace98.0483.2089.6386.7999.0188.4494.1492.5087.8698.4089.3892.3898.4592.17
Trained on Other Dataset
ImageNet81.5873.6770.8070.6997.1867.9585.1974.1375.0190.4579.4982.4286.2379.60
CLIP85.1070.7073.4274.6691.8670.8980.6773.9177.6583.1976.5280.0283.5578.63
MegaDescriptor87.3083.0177.8577.7993.7576.5388.3677.9178.7489.7678.6488.0890.3883.70
", + "bbox": [ + 215, + 305, + 785, + 440 + ], + "page_idx": 10 + }, + { + "type": "text", + "text": "Table 3: Result on animal verification. Jointly trained on our PetFace shows the best performance. We also compared to the trained models on other dataset.", + "bbox": [ + 215, + 441, + 782, + 467 + ], + "page_idx": 10 + }, + { + "type": "text", + "text": "results with an average AUC of $92.17\\%$ when jointly trained on the PetFace dataset. Similar to face re-identification in Sec 5.1, in some classes, the results get worse when jointly trained across animal families e.g., Chimp, Degus, Dog, Guinea pig, Parakeet, and Rabbit. We also examined models pre-trained on other datasets including ImageNet, CLIP and MegaDescriptor in order to make more comparisons. Among these models, MegaDescriptor, which trained on the animal re-identification dataset, showed the highest AUC of $83.70\\%$ . We show the similarity distribution on the our cat dataset in Fig 5. Comparing the performance of CLIP, MegaDescriptor, and our joint-trained models, our findings indicate that our model achieves the most distinct separation between positive and negative samples. In contrast, the MegaDescriptor model displays a closer distribution of these samples, and with CLIP, the overlap between positive and negative distributions is significantly more pronounced.", + "bbox": [ + 212, + 508, + 787, + 705 + ], + "page_idx": 10 + }, + { + "type": "text", + "text": "5.3 Comparison with Previous Datasets", + "text_level": 1, + "bbox": [ + 215, + 726, + 558, + 739 + ], + "page_idx": 10 + }, + { + "type": "text", + "text": "We conduct cross-dataset evaluations where models are tested on different datasets other than the training datasets to demonstrate the generality of our dataset. Here, we compare our dataset with 1) CTai [19] and CZoo [19] for Chimpanzee and 2) DogFaceNet [49] and Flickr-dog [47] for Dog. All images of the compared datasets are aligned in the same manner as ours for fair comparisons. We split the identities of the compared datasets into training and test sets in a ratio of 7:3", + "bbox": [ + 212, + 750, + 792, + 839 + ], + "page_idx": 10 + }, + { + "type": "header", + "text": "PetFace: A Large-Scale Dataset and Benchmark for Animal Identification", + "bbox": [ + 241, + 114, + 730, + 128 + ], + "page_idx": 10 + }, + { + "type": "page_number", + "text": "11", + "bbox": [ + 767, + 114, + 782, + 126 + ], + "page_idx": 10 + }, + { + "type": "image", + "img_path": "images/44020877bc6b8dd93cd55b2f47107c5f99c014fca4ac624e1d4666e65beab0b7.jpg", + "image_caption": [ + "(a) Top-1" + ], + "image_footnote": [], + "bbox": [ + 228, + 143, + 403, + 267 + ], + "page_idx": 11 + }, + { + "type": "image", + "img_path": "images/e5f826bc121e1fe340ddeb5183888139cd1573fe3a93abf6e57e39ba39db0628.jpg", + "image_caption": [ + "(b) Top-3" + ], + "image_footnote": [], + "bbox": [ + 415, + 143, + 589, + 267 + ], + "page_idx": 11 + }, + { + "type": "image", + "img_path": "images/a49caaff41a088cdbda4f8dfc47c5565a5c2d9ba5b00fb28d5430cceb1d6f843.jpg", + "image_caption": [ + "(c) Top-5" + ], + "image_footnote": [], + "bbox": [ + 601, + 143, + 774, + 268 + ], + "page_idx": 11 + }, + { + "type": "image", + "img_path": "images/2a1279156c9c47525e6c7ed96d217e4d0b0ee20df2164fe74b7b171889496cbd.jpg", + "image_caption": [ + "Fig. 4: Top-k accuracy ( $k = 1,3,5$ ). ArcFace consistently shows the best performance consistently." + ], + "image_footnote": [], + "bbox": [ + 227, + 340, + 406, + 462 + ], + "page_idx": 11 + }, + { + "type": "image", + "img_path": "images/b4c81e03b3539375dffe769f4499a77e3fdaf41b6ea5e8e1cc9a14c9298569dd.jpg", + "image_caption": [ + "Fig. 5: Similarity distributions of (a) CLIP, (b) MegaDescriptor, and (c) our joint-trained model on cats. The horizontal axis represents similarity and the vertical axis represents frequency. The model trained on PetFace shows the most distinct separation between positive and negative samples." + ], + "image_footnote": [], + "bbox": [ + 415, + 342, + 589, + 463 + ], + "page_idx": 11 + }, + { + "type": "image", + "img_path": "images/69aff722169cfbce354425ea5aa91d16cbde62671cd78d45e99f89e52a0c2ca3.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 601, + 342, + 776, + 463 + ], + "page_idx": 11 + }, + { + "type": "text", + "text": "and then uniformly pick the positive and negative pairs. Because CTai, CZoo, and Flickr-dog have a few test identities, we each pick five positive and negative pairs for each test identity for robust evaluations.", + "bbox": [ + 212, + 560, + 784, + 604 + ], + "page_idx": 11 + }, + { + "type": "text", + "text": "We show the verification results in Tables 4(a) and (b). We exclude the results trained on the same dataset as the test set denoted in gray for fair comparisons. For the Chimpanzee datasets, the model trained on our dataset outperforms previous datasets. For case tested on CTai, our dataset outperforms CZoo (67.33% vs. 66.76%). For the other case tested on CZoo, our dataset also surpasses CTai (71.27% vs. 69.31%). This result demonstrates that our Chimpanzee dataset is more effective and general than the previous datasets. For the Dog datasets, owing to the significant scale of our dataset, the model trained on our dog dataset outperforms not only cross-dataset results but also in-dataset ones denoted in gray. These results support the quality of our PetFace dataset.", + "bbox": [ + 212, + 606, + 785, + 758 + ], + "page_idx": 11 + }, + { + "type": "text", + "text": "5.4 Analysis", + "text_level": 1, + "bbox": [ + 215, + 782, + 333, + 797 + ], + "page_idx": 11 + }, + { + "type": "text", + "text": "Different networks. To examine the impact of different network architectures on the performance on PetFace, we additionally train two backbones,", + "bbox": [ + 214, + 809, + 784, + 839 + ], + "page_idx": 11 + }, + { + "type": "page_number", + "text": "12", + "bbox": [ + 217, + 114, + 235, + 126 + ], + "page_idx": 11 + }, + { + "type": "header", + "text": "R. Shinoda and K. Shiohara", + "bbox": [ + 271, + 114, + 460, + 127 + ], + "page_idx": 11 + }, + { + "type": "table", + "img_path": "images/7c287fbebfbfc9bce903783b460204a850d1345dcee4835c4fc977f51c0796f9.jpg", + "table_caption": [], + "table_footnote": [], + "table_body": "
Train SetTest Set AUC (%)
Database#Ind.#ImagesCTaiCZoo
CTai49285775.8069.31
CZoo17132366.7680.49
Ours-Chimp446267967.3371.27
", + "bbox": [ + 215, + 143, + 475, + 215 + ], + "page_idx": 12 + }, + { + "type": "table", + "img_path": "images/22bfbee7a3c934a731546140d22eaf3e0a7096571792bc65244dab715771aecd.jpg", + "table_caption": [ + "(a) Chimp datasets." + ], + "table_footnote": [], + "table_body": "
Train SetTest Set AUC (%)
Database#Ind.#ImagesDogFaceNetFlickr-dog
DogFaceNet975587998.7285.51
Flickr-dog3026594.9887.77
Ours-Dog4675516834899.6595.54
", + "bbox": [ + 486, + 143, + 769, + 215 + ], + "page_idx": 12 + }, + { + "type": "table", + "img_path": "images/1afcd088dbb284886443328389b94404fd3e72310a052090eb0c7d0833b29ff2.jpg", + "table_caption": [ + "(b) Dog datasets.", + "Table 4: Comparisons with previous datasets for Chimp and Dog. The results when the test sets are the same as the training datasets are excluded for fair comparisons. The models trained on our dataset outperform those trained on previous datasets." + ], + "table_footnote": [], + "table_body": "
MethodAUC (%)Avg
CatChimpChinchillaDegusDogFerretGuinea pigHamsterHedgehogParakeetJava sparrowPigRabbit
ResNet-5097.7183.7687.7087.6199.4586.2496.0390.3286.1398.6384.4589.8899.0191.30
ViT-32-B94.4478.6071.1382.6993.3968.1783.2177.8987.7795.0780.6275.3195.1883.35
Swin-B87.3574.2376.2184.0485.7571.6091.5688.3386.2294.1284.6666.5690.5683.17
", + "bbox": [ + 215, + 324, + 785, + 383 + ], + "page_idx": 12 + }, + { + "type": "text", + "text": "Table 5: Verification result on different backbones. ResNet-50 achieves the best average AUC.", + "bbox": [ + 212, + 383, + 782, + 410 + ], + "page_idx": 12 + }, + { + "type": "text", + "text": "VisionTransformer-32-B [17] and SwinTransformer-B [41], with the same training configuration as our base model (ResNet-50) and report the verification results in Table 5. We observe that ResNet-50 outperforms the additional transformer-based models in most cases and achieves the best average AUC of $91.30\\%$ .", + "bbox": [ + 212, + 457, + 784, + 518 + ], + "page_idx": 12 + }, + { + "type": "text", + "text": "Fine-grained verification. In real world scenarios, animal verification is sometimes conducted within a specific breed rather than different breeds. To meet this demand, we introduce fine-grained verification where positive and negative pairs are chosen only from the same breeds. We refer to the top-10 dog breeds of individuals in our datasets: Chihuahua, Dachshund, French bulldog, Golden retriever, Miniature dachshund, Pomeranian, Shiba inu, Shih tzu, Toy poodle, and Yorkshire terrier. We construct the fine-grained test sets and evaluate the models on each breed. The results are visualized in Fig. 6. It can be seen that our model achieves a good performance on all the breeds (98.30% on average). The results achieved for the baseline models are lower in comparison (84.99%, 80.78%, and 86.25% for ImageNet, CLIP, and MegaDescriptor, respectively). Moreover, we find that the results achieved with MegaDescriptor show significantly more degradation than the cross-breed verification result in Table 3 from 93.75% to 86.25%. while our model maintains the results at a high-level (from 99.01% to 98.39%). This result indicates that our dataset overcomes the limitation of the previous dog dataset [69] that suffers from the poor performance for specific breed verification owing to its insufficient number of individuals (192 individuals).", + "bbox": [ + 212, + 520, + 785, + 792 + ], + "page_idx": 12 + }, + { + "type": "text", + "text": "Generality to Unseen Families. Lastly, we evaluate the verification generality to unseen animal families. To achieve this, we additionally collect 100 identities each of Parrot, Lacertilia, and Squirrel, which are not contained in either our", + "bbox": [ + 212, + 794, + 784, + 839 + ], + "page_idx": 12 + }, + { + "type": "header", + "text": "PetFace: A Large-Scale Dataset and Benchmark for Animal Identification", + "bbox": [ + 241, + 114, + 730, + 128 + ], + "page_idx": 12 + }, + { + "type": "page_number", + "text": "13", + "bbox": [ + 767, + 114, + 784, + 126 + ], + "page_idx": 12 + }, + { + "type": "image", + "img_path": "images/195502cd50d0b78702522deedc82f87355ae467472e0daed69ce4edba877a1d2.jpg", + "image_caption": [ + "Fig.7: Generality to unseen animal families. Trained on our PetFace achieves the best performance even for unseen animal families." + ], + "image_footnote": [], + "bbox": [ + 223, + 150, + 500, + 325 + ], + "page_idx": 13 + }, + { + "type": "image", + "img_path": "images/412009a568fb4dc11aaa8acf98199e26604771e45bd9ebc126e625874b61dd96.jpg", + "image_caption": [ + "Fig. 6: Fine-grained dog verification. Trained on our PetFace Dog achieves the best performance consistently." + ], + "image_footnote": [], + "bbox": [ + 511, + 147, + 771, + 321 + ], + "page_idx": 13 + }, + { + "type": "text", + "text": "dataset or WildLifeDataset [61]. We compare our model that is jointly trained on our dataset across families with the baseline models in Fig. 7. Our model outperforms the baselines on all the species. For Parrot, particularly, our model achieves an AUC of $88.99\\%$ , revealing its superiority over MegaDescritor with a large margin. This result supports the generality and variety of our PetFace dataset.", + "bbox": [ + 212, + 420, + 787, + 510 + ], + "page_idx": 13 + }, + { + "type": "text", + "text": "6 Conclusion", + "text_level": 1, + "bbox": [ + 215, + 535, + 359, + 550 + ], + "page_idx": 13 + }, + { + "type": "text", + "text": "We introduced the PetFace, a comprehensive animal identification dataset encompassing 13 families, featuring 257,484 unique individuals across 319 breed categories. We have collected images and their information and conducted automated and manual filtering to ensure the quality of the large-scale dataset. This dataset also includes detailed annotations of sex, breed, colors, and patterns to facilitate more investigation and application in a real-world scenario. We establish two main benchmarks: 1) re-identification of seen individuals and 2) verification of unseen individuals.", + "bbox": [ + 212, + 566, + 787, + 686 + ], + "page_idx": 13 + }, + { + "type": "text", + "text": "Our experiments show the generality of the models trained on our dataset to verification on other datasets or unseen animal families. We also found that there is still room for improvement in the integrated identification of multiple animal families. To promote further research, we will make this dataset, experiment code, and models available to the research communities. This dataset will enable computer vision researchers to tackle animal face identification across a wide range of breeds and push forward progress on animal face identification tasks.", + "bbox": [ + 212, + 686, + 787, + 792 + ], + "page_idx": 13 + }, + { + "type": "page_number", + "text": "14", + "bbox": [ + 217, + 114, + 235, + 126 + ], + "page_idx": 13 + }, + { + "type": "header", + "text": "R. Shinoda and K. Shiohara", + "bbox": [ + 271, + 114, + 460, + 127 + ], + "page_idx": 13 + }, + { + "type": "text", + "text": "Acknowledgements", + "text_level": 1, + "bbox": [ + 217, + 143, + 401, + 162 + ], + "page_idx": 14 + }, + { + "type": "text", + "text": "This work was partially supported by the Cooperative Research Program of the Primate Research Institute, Kyoto University. We thank Dr. Takuma Yagi from AIST for comments on paper proofreading. We also thank Prof. Toshihiko Yamasaki from the University of Tokyo for providing computation resources.", + "bbox": [ + 215, + 176, + 787, + 239 + ], + "page_idx": 14 + }, + { + "type": "text", + "text": "References", + "text_level": 1, + "bbox": [ + 215, + 263, + 323, + 277 + ], + "page_idx": 14 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "1. Beluga id 2022 (2022), https://lila.science/datasets/beluga-id-2022/", + "2. Hyena id 2022 botswana predator conservation trust (2022). panthera pardus csv custom export. (2022), https://lila.science/datasets/hyena-id-2022/", + "3. Leopard id 2022 botswana predator conservation trust (2022). panthera pardus csv custom export. (2022), https://lila.science/datasets/leopard-id-2022/", + "4. Turtle recall: Conservation challenge (2022), https://zindi.africa/competitions/turtle-recall-conservation-challenge", + "5. Adam, L., Čermák, V., Papafitsoros, K., Picek, L.: Seaturtleid2022: A long-span dataset for reliable sea turtle re-identification. In: IEEE/CVF Winter Conference on Applications of Computer Vision (WACV). pp. 7146-7156 (2024)", + "6. Andrew, W., Gao, J., Mullan, S., Campbell, N., Dowsey, A.W., Burghardt, T.: Visual identification of individual holstein-friesian cattle via deep metric learning. In: Computers and Electronics in Agriculture. vol. 185, p. 106133 (2021)", + "7. Andrew, W., Greatwood, C., Burghardt, T.: Visual localisation and individual identification of holstein friesian cattle via deep learning. In: IEEE/CVF International Conference on Computer Vision (ICCV) Workshops. pp. 2850-2859 (2017)", + "8. Andrew, W., Greatwood, C., Burghardt, T.: Visual localisation and individual identification of holstein friesian cattle via deep learning. In: IEEE/CVF International Conference on Computer Vision (ICCV) Workshops. pp. 2850-2859 (2017)", + "9. Andrew, W., Hannuna, S., Campbell, N., Burghardt, T.: Automatic individual holstein friesian cattle identification via selective local coat pattern matching in rgb-d imagery. In: International Conference on Image Processing (ICIP). pp. 484-488 (2016)", + "10. Brookes, O., Burghardt, T.: A dataset and application for facial recognition of individual gorillas in zoo environments. In: International Conference on Pattern Recognition(ICPR) (2020)", + "11. Cheeseman, T., Southerland, K., Park, J., Olio, M., Flynn, K., Calambokidis, J., Jones, L., Garrigue, C., Frisch-Jordán, A., Howard, A., Reade, W., Neilson, J., Gabriele, C., Clapham, P.: Advanced image recognition: a fully automated, high-accuracy photo-identification matching system for humpback whales. vol. 102, p. 915–929 (2021)", + "12. Christin B. Khan, Shashank, W.K.: Right whale recognition (2015), https://kaggle.com/competitions/noaa-right-whale-recognition", + "13. Dahlborn, K., Bugnon, P., Nevalainen, T., Raspa, M., Verbost, P.M., Spangenberg, E.M.F.: Report of the federation of european laboratory animal science associations working group on animal identification. vol. 47, pp. 2-11 (2013)", + "14. Deng, J., Dong, W., Socher, R., Li, L.J., Li, K., Fei-Fei, L.: Imagenet: A large-scale hierarchical image database. In: IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR). pp. 248-255 (2009)" + ], + "bbox": [ + 218, + 295, + 784, + 839 + ], + "page_idx": 14 + }, + { + "type": "header", + "text": "PetFace: A Large-Scale Dataset and Benchmark for Animal Identification", + "bbox": [ + 241, + 114, + 730, + 128 + ], + "page_idx": 14 + }, + { + "type": "page_number", + "text": "15", + "bbox": [ + 767, + 114, + 785, + 126 + ], + "page_idx": 14 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "15. Deng, J., Guo, J., Xue, N., Zafeiriou, S.: Arcface: Additive angular margin loss for deep face recognition. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR). pp. 4690-4699 (2019)", + "16. Dlamini, N., Zyl, T.L.v.: Automated identification of individuals in wildlife population using siamese neural networks. In: 2020 7th International Conference on Soft Computing Machine Intelligence (ISCMI). pp. 224-228 (2020)", + "17. Dosovitskiy, A., Beyer, L., Kolesnikov, A., Weissenborn, D., Zhai, X., Unterthiner, T., Dehghani, M., Minderer, M., Heigold, G., Gelly, S., Uszkoreit, J., Houlsby, N.: An image is worth 16x16 words: Transformers for image recognition at scale. The International Conference on Learning Representations (ICLR) (2021)", + "18. Ferreira, A.C., Silva, L.R., Renna, F., Brandl, H.B., Renoult, J., Farine, D.R., Covas, R., Doutrelant, C.: Deep learning-based methods for individual recognition in small birds. In: Methods in Ecology and Evolution. vol. 11, pp. 1072-1085 (2020)", + "19. Freytag, A., Rodner, E., Simon, M., Loos, A., Kuhl, H.S., Denzler, J.: Chimpanzee faces in the wild: Log-euclidean cnns for predicting identities and attributes of primates. In: Rosenhahn, B., Andres, B. (eds.) Pattern Recognition. pp. 51-63. Springer International Publishing, Cham (2016)", + "20. Gao, J., Burghardt, T., Andrew, W., Dowsey, A.W., Campbell, N.W.: Towards self-supervision for video identification of individual holstein-friesian cattle: The cows2021 dataset. arXiv preprint arXiv:2105.01938 (2021)", + "21. Guo, Y., Zhang, L., Hu, Y., He, X., Gao, J.: Ms-celeb-1m: A dataset and benchmark for large-scale face recognition. In: Leibe, B., Matas, J., Sebe, N., Welling, M. (eds.) European Conference on Computer Vision (ECCV). pp. 87-102. Springer International Publishing, Cham (2016)", + "22. Haurum, J.B., Karpova, A., Pedersen, M., Bengtson, S.H., Moeslund, T.B.: Identification of zebrafish using metric learning. In: IEEE Winter Conference on Applications of Computer Vision (WACV) Workshops. pp. 1-11 (2020)", + "23. He, K., Zhang, X., Ren, S., Sun, J.: Deep residual learning for image recognition. In: IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR). pp. 770-778 (2016)", + "24. He, Z., Qian, J., Yan, D., Wang, C., Xin, Y.: Animal re-identification algorithm for posture diversity. In: ICASSP 2023 - 2023 IEEE International Conference on Acoustics, Speech and Signal Processing (ICASSP). pp. 1-5 (2023)", + "25. Hitelman, A., Edan, Y., Godo, A., Berenstein, R., Lepar, J., Halachmi, I.: Biometric identification of sheep via a machine-vision system. vol. 194, p. 106713 (2022)", + "26. Hoffer, E., Ailon, N.: Deep metric learning using triplet network. In: arXiv (2018)", + "27. Holmberg, J., Norman, B., Arzoumanian, Z.: Estimating population size, structure, and residency time for whale sharks rhincodon typus through collaborative photo-identification. vol. 7, pp. 39 - 53. Inter-Research Science Center (2009)", + "28. Howard, A., inversion, Southerland, K., Cheeseman, T.: Humpback whale identification challenge (2018), https://kaggle.com/competitions/whale-categorization-playground", + "29. Huang, G.B., Ramesh, M., Berg, T., Learned-Miller, E.: Labeled faces in the wild: A database for studying face recognition in unconstrained environments. No. 07-49 (2007)", + "30. Jarraya, I., Ouarda, W., Alimi, A.M.: A preliminary investigation on horses recognition using facial texture features. In: IEEE International Conference on Systems, Man, and Cybernetics. pp. 2803-2808 (2015)", + "31. Khan, F.F., Li, X., Temple, A.J., Elhoseiny, M.: Fishnet: A large-scale dataset and benchmark for fish recognition, detection, and functional trait prediction. In: Pro" + ], + "bbox": [ + 215, + 147, + 784, + 839 + ], + "page_idx": 15 + }, + { + "type": "page_number", + "text": "16", + "bbox": [ + 217, + 114, + 235, + 126 + ], + "page_idx": 15 + }, + { + "type": "header", + "text": "R. Shinoda and K. Shiohara", + "bbox": [ + 271, + 114, + 460, + 127 + ], + "page_idx": 15 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "ceedings of the IEEE/CVF International Conference on Computer Vision (ICCV). pp. 20496-20506 (October 2023)", + "32. Khan, M.H., McDonagh, J., Khan, S., Shahabuddin, M., Arora, A., Khan, F.S., Shao, L., Tzimiropoulos, G.: Animalweb: A large-scale hierarchical dataset of annotated animal faces. In: IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR). pp. 6937-6946 (2020)", + "33. Kim, M., Jain, A.K., Liu, X.: Adaface: Quality adaptive margin for face recognition. pp. 18729-18738 (2022)", + "34. Krishnan, S., S., S.: Cloud iot based novel livestock monitoring and identification system using uid. In: Sensor Review. vol. 38, pp. 21-33 (2017)", + "35. Kuzdeuov, A., Koishigarina, D., Varol, H.A.: Anyface: A data-centric approach for input-agnostic face detection. In: IEEE International Conference on Big Data and Smart Computing (BigComp). pp. 211-218 (2023)", + "36. Lahiri, M., Tantipathananandh, C., Warungu, R., Rubenstein, D.I., Berger-Wolf, T.Y.: Biometric animal databases from field photographs: identification of individual zebra in the wild. In: Proceedings of the ACM International Conference on Multimedia Retrieval. Association for Computing Machinery (2011)", + "37. Leslie, E., Hernández-Jover, M., Newman, R., Holyoake, P.: Assessment of acute pain experienced by piglets from ear tagging, ear notching and intraperitoneal injectable transponders. In: Applied Animal Behaviour Science. vol. 127, pp. 86-95 (2010)", + "38. Li, S., Li, J., Tang, H., Qian, R., Lin, W.: Atrw: A benchmark for amur tiger re-identification in the wild. In: Proceedings of the 28th ACM International Conference on Multimedia. p. 2590–2598. MM '20, Association for Computing Machinery, New York, NY, USA (2020)", + "39. Li, X., Xiang, Y., Li, S.: Combining convolutional and vision transformer structures for sheep face recognition. In: Computers and Electronics in Agriculture. vol. 205, p. 107651 (2023)", + "40. Liu, W., Wen, Y., Yu, Z., Li, M., Raj, B., Song, L.: Spherface: Deep hypersphere embedding for face recognition. In: IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR). pp. 6738-6746 (2017)", + "41. Liu, Z., Lin, Y., Cao, Y., Hu, H., Wei, Y., Zhang, Z., Lin, S., Guo, B.: Swin transformer: Hierarchical vision transformer using shifted windows. In: IEEE/CVF International Conference on Computer Vision (ICCV). pp. 9992-10002", + "42. Loos, A., Ernst, A.: An automated chimpanzee identification system using face detection and recognition. In: EURASIP Journal on Image and Video Processing. vol. 2013 (2013)", + "43. Marsot, M., Mei, J., Shan, X., Ye, L., Feng, P., Yan, X., Li, C., Zhao, Y.: An adaptive pig face recognition approach using convolutional neural networks. In: Computers and Electronics in Agriculture. vol. 173, p. 105386 (2020)", + "44. Martvel, G., Shimshoni, I., Zamansky, A.: Automated detection of cat facial landmarks. In: International Journal of Computer Vision (IJCV). pp. 1-16 (2023)", + "45. Mazlan, N.H., Lopez Salesansky, N., Burn, C., Wells, D.: Mouse identification methods and potential welfare issues: A survey of current practice in the uk. In: Animal Technology and Welfare. vol. 13, pp. 1-10 (2014)", + "46. Miele, V., Dussert, G., Spataro, B., Chamaillé-Jammes, S., Allainé, D., Bonenfant, C.: Revisiting animal photo-identification using deep metric learning and network analysis. In: Methods in Ecology and Evolution. vol. 12, pp. 863-873. Wiley (2021)", + "47. Moreira, T., Perez, M., Werneck, R., Valle, E.: Where is my puppy? retrieving lost dogs by facial features. In: Multimedia Tools and Applications. vol. 76, p. 15325-15340 (2017)" + ], + "bbox": [ + 215, + 146, + 785, + 839 + ], + "page_idx": 16 + }, + { + "type": "header", + "text": "PetFace: A Large-Scale Dataset and Benchmark for Animal Identification", + "bbox": [ + 241, + 114, + 730, + 128 + ], + "page_idx": 16 + }, + { + "type": "page_number", + "text": "17", + "bbox": [ + 767, + 114, + 785, + 126 + ], + "page_idx": 16 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "48. Moschoglou, S., Papaioannou, A., Sagonas, C., Deng, J., Kotsia, I., Zafeiriou, S.: Agedb: The first manually collected, in-the-wild age database. In: IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR) Workshops. pp. 1997-2005 (2017)", + "49. Mougeot, G., Li, D., Jia, S.: A deep learning approach for dog face verification and recognition. In: Nayak, A.C., Sharma, A. (eds.) PRICAI 2019: Trends in Artificial Intelligence. pp. 418-430. Springer International Publishing, Cham (2019)", + "50. Murali, N., Taylor, G., Levine, J.: Can drosophila melanogaster tell who's who? In: PLOS ONE. vol. 13, p. e0205043 (2018)", + "51. Nepovinnykh, E., Eerola, T., Biard, V., Mutka, P., Niemi, M., Kunnasranta, M., Kalviainen, H.: Sealid: Saimaa ringed seal re-identification dataset. Sensors 22(19) (2022)", + "52. Papafitsoros, K., Adam, L., Čermák, V., Picek, L.: Seaturtleid: A novel long-span dataset highlighting the importance of timestamps in wildlife re-identification. In: arXiv (2023)", + "53. Parham, J., Crall, J., Stewart, C., Berger-Wolf, T., Rubenstein, D.: Animal population censusing at scale with citizen science and photographic identification. In: SS-17-01. pp. 37-44. AAAI Spring Symposium (2017)", + "54. Radford, A., Kim, J.W., Hallacy, C., Ramesh, A., Goh, G., Agarwal, S., Sastry, G., Askell, A., Mishkin, P., Clark, J., Krueger, G., Sutskever, I.: Learning transferable visual models from natural language supervision. In: arXiv (2021)", + "55. Roberts, C.: Radio frequency identification (rfid). In: Computers Security. vol. 25, pp. 18-26 (2006)", + "56. Roughan, J., Sevenoaks, T.: Welfare and scientific considerations of tattooing and ear-tagging for mouse identification. In: Journal of the American Association for Laboratory Animal Science: JAALAS. pp. 142-153 (2018)", + "57. Sengupta, S., Chen, J.C., Castillo, C., Patel, V.M., Chellappa, R., Jacobs, D.W.: Frontal to profile face verification in the wild. In: IEEE Winter Conference on Applications of Computer Vision (WACV). pp. 1-9 (2016)", + "58. Shi, X., Yang, C., Xia, X., Chai, X.: Deep cross-species feature learning for animal face recognition via residual interspecies equivariant network. In: European Conference on Computer Vision (ECCV). p. 667-682 (2020)", + "59. Shinoda, R., Shiohara, K.: Openanimaltracks: A dataset for animal track recognition (2024)", + "60. Trotter, C., Atkinson, G., Sharpe, M., Richardson, K., McGough, A.S., Wright, N., Burville, B., Berggren, P.: Ndd20: A large-scale few-shot dolphin dataset for coarse and fine-grained categorisation. In: arXiv (2020)", + "61. Čermák, V., Picek, L., Adam, L., Papafitsoros, K.: Wildlifedatasets: An open-source toolkit for animal re-identification. In: IEEE/CVF Winter Conference on Applications of Computer Vision (WACV). pp. 5953-5963 (2024)", + "62. Wahltinez, O., Wahltinez, S.: An open-source general purpose machine learning framework for individual animal re-identification using few-shot learning. In: Methods in Ecology and Evolution. vol. 15, pp. 373-387 (2024)", + "63. Wang, H., Wang, Y., Zhou, Z., Ji, X., Gong, D., Zhou, J., Li, Z., Liu, W.: Cosface: Large margin cosine loss for deep face recognition. In: IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR). pp. 5265-5274 (2018)", + "64. Wang, L., Ding, R., Zhai, Y., Zhang, Q., Tang, W., Zheng, N., Hua, G.: Giant panda identification. IEEE Transactions on Image Processing 30, 2837-2849 (2021)", + "65. Wen, Y., Zhang, K., Li, Z., Qiao, Y.: A discriminative feature learning approach for deep face recognition. In: Leibe, B., Matas, J., Sebe, N., Welling, M. (eds.) European Conference on Computer Vision (ECCV). pp. 499-515 (2016)" + ], + "bbox": [ + 215, + 146, + 785, + 840 + ], + "page_idx": 17 + }, + { + "type": "page_number", + "text": "18", + "bbox": [ + 217, + 114, + 235, + 126 + ], + "page_idx": 17 + }, + { + "type": "header", + "text": "R. Shinoda and K. Shiohara", + "bbox": [ + 271, + 114, + 460, + 127 + ], + "page_idx": 17 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "66. Witham, C.L.: Automated face recognition of rhesus macaques. In: Journal of Neuroscience Methods. vol. 300, pp. 157-165 (2018)", + "67. Zhu, Z., Huang, G., Deng, J., Ye, Y., Huang, J., Chen, X., Zhu, J., Yang, T., Lu, J., Du, D., Zhou, J.: Webface260m: A benchmark unveiling the power of million-scale deep face recognition. In: IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR). pp. 10492-10502 (2021)", + "68. Zuerl, M., Dirauf, R., Koeferl, F., Steinlein, N., Sueskind, J., Zanca, D., Brehm, I., Fersen, L.v., Eskofier, B.: Polarbearvidid: A video-based re-identification benchmark dataset for polar bears. In: Animals. vol. 13 (2023)", + "69. Zuffi, S., Kanazawa, A., Berger-Wolf, T., Black, M.J.: Three-d safari: Learning to estimate zebra pose, shape, and texture from images \"in the wild\". In: IEEE/CVF International Conference on Computer Vision (ICCV). pp. 5358-5367 (2019)" + ], + "bbox": [ + 215, + 146, + 787, + 314 + ], + "page_idx": 18 + }, + { + "type": "header", + "text": "PetFace: A Large-Scale Dataset and Benchmark for Animal Identification", + "bbox": [ + 241, + 114, + 730, + 128 + ], + "page_idx": 18 + }, + { + "type": "page_number", + "text": "19", + "bbox": [ + 767, + 114, + 785, + 126 + ], + "page_idx": 18 + } +] \ No newline at end of file diff --git a/2024/PetFace_ A Large-Scale Dataset and Benchmark for Animal Identification/79b48743-23c5-4624-b76c-ce49a285a1a7_model.json b/2024/PetFace_ A Large-Scale Dataset and Benchmark for Animal Identification/79b48743-23c5-4624-b76c-ce49a285a1a7_model.json new file mode 100644 index 0000000000000000000000000000000000000000..dc517adad7d09c442a41279c914b08712d504d5f --- /dev/null +++ b/2024/PetFace_ A Large-Scale Dataset and Benchmark for Animal Identification/79b48743-23c5-4624-b76c-ce49a285a1a7_model.json @@ -0,0 +1,2636 @@ +[ + [ + { + "type": "title", + "bbox": [ + 0.229, + 0.141, + 0.777, + 0.185 + ], + "angle": 0, + "content": "PetFace: A Large-Scale Dataset and Benchmark for Animal Identification" + }, + { + "type": "text", + "bbox": [ + 0.346, + 0.213, + 0.658, + 0.228 + ], + "angle": 0, + "content": "Risa Shinoda\\(^{1*}\\) and Kaede Shiohara\\(^{2*}\\)" + }, + { + "type": "text", + "bbox": [ + 0.298, + 0.241, + 0.705, + 0.27 + ], + "angle": 0, + "content": "\\(^{*}\\)Equal contribution \n\\(^{1}\\)Kyoto University, Japan \\(^{2}\\)The University of Tokyo, Japan" + }, + { + "type": "text", + "bbox": [ + 0.263, + 0.309, + 0.74, + 0.6 + ], + "angle": 0, + "content": "Abstract. Automated animal face identification plays a crucial role in the monitoring of behaviors, conducting of surveys, and finding of lost animals. Despite the advancements in human face identification, the lack of datasets and benchmarks in the animal domain has impeded progress. In this paper, we introduce the PetFace dataset, a comprehensive resource for animal face identification encompassing 257,484 unique individuals across 13 animal families and 319 breed categories, including both experimental and pet animals. This large-scale collection of individuals facilitates the investigation of unseen animal face verification, an area that has not been sufficiently explored in existing datasets due to the limited number of individuals. Moreover, PetFace also has fine-grained annotations such as sex, breed, color, and pattern. We provide multiple benchmarks including re-identification for seen individuals and verification for unseen individuals. The models trained on our dataset outperform those trained on prior datasets, even for detailed breed variations and unseen animal families. Our result also indicates that there is some room to improve the performance of integrated identification on multiple animal families. We hope the PetFace dataset will facilitate animal face identification and encourage the development of non-invasive animal automatic identification methods. Our dataset and code are available at https://dahlian00.github.io/PetFacePage/." + }, + { + "type": "text", + "bbox": [ + 0.263, + 0.614, + 0.658, + 0.628 + ], + "angle": 0, + "content": "Keywords: Animals \\(\\cdot\\) Re-identification \\(\\cdot\\) Face Recognition" + }, + { + "type": "title", + "bbox": [ + 0.218, + 0.656, + 0.375, + 0.671 + ], + "angle": 0, + "content": "1 Introduction" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.689, + 0.788, + 0.84 + ], + "angle": 0, + "content": "Animal identification plays a crucial role in animal studies and applications such as monitoring animal behavior, conducting habitat surveys, locating missing animals, and performing health checks. Traditional identification techniques, including ear tags, tattoos, ear punching, and toe clipping, continue to be utilized mainly for experimental animals and livestock [13]. However, given that these methods have the potential to cause stress and pain [37,45,56], their use should be minimized to prioritize animal welfare. Therefore, there is a pressing need for the development and adoption of identification technologies that are not only effective and efficient but also minimally invasive, thereby mitigating the ethical concerns associated with traditional methods. While advanced tools such as" + } + ], + [ + { + "type": "page_number", + "bbox": [ + 0.218, + 0.116, + 0.23, + 0.127 + ], + "angle": 0, + "content": "2" + }, + { + "type": "header", + "bbox": [ + 0.272, + 0.115, + 0.462, + 0.128 + ], + "angle": 0, + "content": "R. Shinoda and K. Shiohara" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.147, + 0.784, + 0.221 + ], + "angle": 0, + "content": "digital IDs [34, 55] have been introduced, their applications involve a laborious process, i.e., attaching the devices to each animal individually. The process is costly and potentially stressful for the animals. Moreover, these physical tags can identify only pre-defined individuals, which makes them impractical for use in real world scenarios." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.223, + 0.785, + 0.313 + ], + "angle": 0, + "content": "In the human domain, digital face recognition is one of the effective approaches for the identification. It has been developed for use in smartphones, airport security, and systems for finding missing people. Therefore, the research community made great efforts to develop sophisticated deep learning-based face recognition models [15,26,33,40,63,65], empowered by large-scale datasets and benchmarks, e.g., [21,29,48,57]." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.314, + 0.785, + 0.403 + ], + "angle": 0, + "content": "Despite the promise of human face recognition, the research progress towards automatic animal face individual recognition has been impeded, primarily because of the lack of extensive datasets and benchmarks for animal face recognition. Previous openly available datasets mostly include less than 100 individuals [10, 19, 30, 47, 66], which makes it far from generalized and discriminative identification models and precise evaluation for unseen individuals." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.405, + 0.785, + 0.553 + ], + "angle": 0, + "content": "In this paper, we introduce a large-scale animal face recognition dataset called PetFace that contains 257,484 individuals in total across 13 species with 319 breeds with 1,012,934 images. We show the example images of our PetFace in Fig 1. The number of individuals in our dataset is over 110 times that in the previous largest animal face dataset [49]. We sourced images and related information from the internet, with automated and manual filtering processes applied to ensure the dataset is not only large but also finely detailed and clean. Moreover, PetFace has fine-grained annotations including sex, breeds, and colors and patterns of their skin, which allows further investigation for fine-grained recognition and evaluation." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.556, + 0.785, + 0.601 + ], + "angle": 0, + "content": "PetFace offers two benchmarks: one for recognizing known (seen) individuals and the other for recognizing unknown (unseen) ones. We also conduct the verification of the fine-grade breeds and unseen animal categories." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.602, + 0.785, + 0.707 + ], + "angle": 0, + "content": "Our main contributions are as follows: (i) We establish a new dataset for animal face recognition called PetFace, which contains a total of 257,484 individuals across 13 types of animal families and 319 breeds with fine-grade annotations including sex, breed, color of animals. (ii) We set the benchmarks on recognizing known (seen) individuals and unknown (unseen) ones. (iii) We show that the model trained using our dataset shows the generalization capabilities for unseen individuals and even for unseen animal categories." + }, + { + "type": "title", + "bbox": [ + 0.216, + 0.731, + 0.388, + 0.748 + ], + "angle": 0, + "content": "2 Related Work" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.765, + 0.785, + 0.84 + ], + "angle": 0, + "content": "Building datasets and benchmarks is an important step in advancing animal re-identification through deep models. While earlier efforts have established the basics of animal face recognition, there is considerable potential for further development. Compared to the dataset for human face recognition [21, 29, 48, 57, 67], those for animal faces [10, 19, 49, 66] have much fewer individuals. The evaluation" + } + ], + [ + { + "type": "header", + "bbox": [ + 0.242, + 0.115, + 0.731, + 0.128 + ], + "angle": 0, + "content": "PetFace: A Large-Scale Dataset and Benchmark for Animal Identification" + }, + { + "type": "page_number", + "bbox": [ + 0.775, + 0.117, + 0.785, + 0.127 + ], + "angle": 0, + "content": "3" + }, + { + "type": "image", + "bbox": [ + 0.227, + 0.147, + 0.765, + 0.319 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.214, + 0.33, + 0.788, + 0.412 + ], + "angle": 0, + "content": "Fig. 1: Example images of our PetFace. We introduce a large-scale animal face re-identification dataset PetFace that include 257,484 unique individuals across 13 families and 319 breeds. From the left, the images represent Cat, Chimpanzee, Chinchilla, Degus, Dog, Ferret, Guinea pig, Hamster, Hedgehog, Parakeet, Java sparrow, Pig, and Rabbit. The four images enclosed within the white square's grid lines represent the same identity." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.432, + 0.788, + 0.506 + ], + "angle": 0, + "content": "scenarios represented in these datasets often fall short of real-world applicability; most of the datasets focus on closed-set re-identification, rather than recognizing unseen individuals that is a critical requirement for practical applications. Our work enables the training and evaluation on a large number of individuals across a wider range of animal families and breeds." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.508, + 0.788, + 0.688 + ], + "angle": 0, + "content": "Human Face Identification is the process of identifying an individual's identity using their unique facial characteristics. The huge demands for individual identification have grown mainly in the human domain. In the identification of human faces, which is known as face recognition, considerable efforts have been expended in the research community [15,26,33,40,63,65]. Because of their domain-agnostic frameworks, most of the state-of-the-art methods in face recognition can be exported into other domains, such as animal identification. Furthermore, these advancements have been supported by the introduction of large-scale datasets and benchmarks [21, 29, 48, 57, 67] that have played a crucial role in facilitating the exploration of powerful models. Motivated by this, we create the PetFace dataset to fill the gap between human face recognition and animal face recognition." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.689, + 0.788, + 0.841 + ], + "angle": 0, + "content": "Animal Identification is the process of identifying an individual's identity using their unique body or facial characteristics, which is an important task across various scenarios, including monitoring animals, conducting habitat surveys, and finding missing animals. With the advance of computer vision technology, various openly available datasets contribute to computer vision for animals [6-8, 12, 18, 20, 24, 31, 38, 46, 50, 59, 60, 64]. Various methods are used to create the dataset, such as recording images [12, 18, 24, 53, 60] and videos [22, 38, 46, 50, 64, 68], and using aerial image [6, 7, 20]. Recording individual information is labor intensive; therefore, most of the previous datasets contain only a limited number of individuals and are focused on one species. Recently, WildlifeDatasets [61], which" + } + ], + [ + { + "type": "page_number", + "bbox": [ + 0.218, + 0.116, + 0.23, + 0.127 + ], + "angle": 0, + "content": "4" + }, + { + "type": "header", + "bbox": [ + 0.272, + 0.115, + 0.462, + 0.128 + ], + "angle": 0, + "content": "R. Shinoda and K. Shiohara" + }, + { + "type": "table", + "bbox": [ + 0.217, + 0.144, + 0.803, + 0.312 + ], + "angle": 0, + "content": "
DatasetFamily#Individuals#ImagesFine-grained annotation
CTai [19]Chimp785078Age, Sex
CZoo [19]Chimp242109Age, Sex
MacaqueFaces [66]Monkey346280-
BristolGorillas2020 [10]Gorilla75428-
DogFaceNet [49]Dog13938363-
Flickr-dog [47]Dog42374-
THODBRL2015 [30]Horse472820Angle
ZindiTurtleRecall [4]Turtle226512803Angle
SeaTurtleID2022 [5]Turtle4388729Timestamp
Ours13 families257,4841,012,934Sex, Breed, Color & Pattern
" + }, + { + "type": "table_caption", + "bbox": [ + 0.216, + 0.312, + 0.79, + 0.352 + ], + "angle": 0, + "content": "Table 1: Openly available animal face identification datasets. Compared to other datasets, PetFace has the largest number of individuals, and a wider range of families and breeds." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.381, + 0.786, + 0.426 + ], + "angle": 0, + "content": "gathers previously openly available datasets [1-4,6-9,11,12,16,18-20,22,24,27, 28,36,38,46,50-53,60,62,64,66,68,69] combined, was introduced. This research creates the benchmarks using existing available datasets." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.428, + 0.788, + 0.759 + ], + "angle": 0, + "content": "The advances in human face recognition raise an interest in animal face recognition. One notable challenge is the variation in facial structures between animals and humans. This has led to studies like AnimalWeb [32] and CatFLW [44] that propose specialized methods for animal facial key points detection. Automatic animal face identification has been studied, which can help humans to monitor animals [25, 39, 42, 43, 58]. We review several publicly available face identification datasets in Table 1. Apes, such as Chimpanzees [19] and Gorillas [10], are one of the species that have been studied. However, as the dataset size is relatively limited, the datasets are insufficient to evaluate new unseen individual faces. In addition to primates, research on facial (head) identification has also extended to turtles, with the ZindiTurtleRecall [4] dataset offering a substantial number of images of individual turtles in a controlled environment. To address the limitations of in-the-wild data collection, the SeaTurtleID2022 [5] dataset includes wild data, albeit with a reduced number of individuals, owing to the extensive effort required for data collection. Among datasets containing mammals, the DogFaceNet dataset [49] featuring 1,393 individual dogs stands out for its size. Nevertheless, the significant variation in appearance across dog breeds suggests that even this larger dataset may not be sufficiently comprehensive. In addition to the limited number of individuals, previous animal face datasets often focus on only single animal families. They can contribute to specific animal family research but can not explore animal identification across many families and breeds." + }, + { + "type": "title", + "bbox": [ + 0.216, + 0.78, + 0.411, + 0.795 + ], + "angle": 0, + "content": "3 PetFace Dataset" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.81, + 0.787, + 0.84 + ], + "angle": 0, + "content": "PetFace is a large animal face identification dataset that expands research on animal face recognition, which has been impeded by a scarcity of suitable datasets" + } + ], + [ + { + "type": "header", + "bbox": [ + 0.242, + 0.115, + 0.732, + 0.129 + ], + "angle": 0, + "content": "PetFace: A Large-Scale Dataset and Benchmark for Animal Identification" + }, + { + "type": "page_number", + "bbox": [ + 0.775, + 0.117, + 0.785, + 0.127 + ], + "angle": 0, + "content": "5" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.147, + 0.788, + 0.207 + ], + "angle": 0, + "content": "and benchmarks. This section details the construction of the PetFace dataset, including our labor-efficient methods for collecting animal face images and a semi-automated filtering process to ensure quality fine-grained categorization and statistics of the dataset." + }, + { + "type": "title", + "bbox": [ + 0.216, + 0.23, + 0.409, + 0.244 + ], + "angle": 0, + "content": "3.1 Dataset Statistics" + }, + { + "type": "image", + "bbox": [ + 0.223, + 0.283, + 0.508, + 0.424 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.305, + 0.426, + 0.469, + 0.439 + ], + "angle": 0, + "content": "(a) Individual distribution" + }, + { + "type": "image", + "bbox": [ + 0.512, + 0.283, + 0.801, + 0.426 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.595, + 0.426, + 0.72, + 0.44 + ], + "angle": 0, + "content": "(b) Sex distribution" + }, + { + "type": "image", + "bbox": [ + 0.255, + 0.453, + 0.419, + 0.59 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.434, + 0.453, + 0.576, + 0.59 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.604, + 0.453, + 0.744, + 0.59 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.407, + 0.595, + 0.602, + 0.61 + ], + "angle": 0, + "content": "(c) Breed annotation examples" + }, + { + "type": "image", + "bbox": [ + 0.231, + 0.626, + 0.4, + 0.761 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.427, + 0.627, + 0.577, + 0.76 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.602, + 0.627, + 0.758, + 0.758 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.416, + 0.762, + 0.609, + 0.777 + ], + "angle": 0, + "content": "(d) Color annotation examples" + }, + { + "type": "image_caption", + "bbox": [ + 0.214, + 0.794, + 0.788, + 0.837 + ], + "angle": 0, + "content": "Fig. 2: Dataset distribution. Each figure represents (a) the number of individuals in each animal family. (b) the sex distribution percentage by animal family. (c) examples of breed annotations. (d) examples of color annotations." + } + ], + [ + { + "type": "page_number", + "bbox": [ + 0.218, + 0.116, + 0.23, + 0.127 + ], + "angle": 0, + "content": "6" + }, + { + "type": "header", + "bbox": [ + 0.272, + 0.115, + 0.462, + 0.128 + ], + "angle": 0, + "content": "R. Shinoda and K. Shiohara" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.147, + 0.788, + 0.299 + ], + "angle": 0, + "content": "The PetFace dataset encompasses 1,012,934 images spanning 257,484 unique individuals. Detailed distributions of animal families are illustrated in Fig 2(a). Images are cropped to \\(224 \\times 224\\) pixels around their faces. Our PetFace also has fine-grained annotations. Sex distribution across different animal families is depicted in Fig 2(b), with sex information available for 240,861 individuals, accounting for \\(94\\%\\) of the dataset. The dataset includes annotations for 319 breeds, with examples of breed annotations shown in Fig 2(c). Furthermore, as detailed in Fig 2(d), the dataset provides in-depth color information through two-tier hierarchical annotations. Please see the supplemental for the detailed information per each animal family." + }, + { + "type": "title", + "bbox": [ + 0.216, + 0.319, + 0.383, + 0.334 + ], + "angle": 0, + "content": "3.2 Data Sourcing" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.343, + 0.788, + 0.432 + ], + "angle": 0, + "content": "Collecting images of animal faces via photography is labor and time intensive, which impedes the creation of large-scale datasets. In contrast, the human face recognition field benefits greatly from the availability of images sourced from the Internet. In this section, we outline our approach to assembling the collection of animal face images through the Internet, where each is associated with unique individual identifiers." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.434, + 0.788, + 0.644 + ], + "angle": 0, + "content": "Curation of images. In contrast to the human domain, acquiring multiple images for individual animals is more challenging. Unlike human datasets where a large number of celebrity images can be readily sourced, animal images require alternative approaches for curation. We utilized two primary sources: (i) pet shops' websites and (ii) animal adoption websites. The advantage of pet shops lies in their provision of high-quality, diverse images, capturing individuals from various angles and offering detailed information about each animal, including color, sex, and specific breed details. On the other hand, animal adoption sites offer images set against a variety of backgrounds and conditions that are often provided by pet owners, thus ensuring each individual is presented in a unique setting. These sources provide images that are highly suitable for animal recognition tasks and that are especially useful for recognizing animals in varied wild environments. For Chimps, we additionally use images from the webpage of a collaborative research institution." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.646, + 0.788, + 0.766 + ], + "angle": 0, + "content": "To ensure the dataset's quality, we were selective in choosing websites for curation. We only use the websites introducing each animal on one page to gain the individual IDs. Aware of the potential for pet owners to upload the same images to multiple websites, we chose a single pet adoption website from each region (e.g., one per country) to minimize duplicates. For pet shops, we confirmed that the animals listed were unique to a certain shop to ensure the uniqueness of our dataset entries. Using these sources, we collected 1,443,737 images from 325,420 individuals." + }, + { + "type": "title", + "bbox": [ + 0.216, + 0.786, + 0.507, + 0.803 + ], + "angle": 0, + "content": "3.3 Face Alignment and Filtering" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.81, + 0.787, + 0.842 + ], + "angle": 0, + "content": "Face Detection. We detect facial landmarks to align and crop the images. We adopt the AnyFace [35] that is trained on mixed face datasets including" + } + ], + [ + { + "type": "header", + "bbox": [ + 0.242, + 0.115, + 0.732, + 0.129 + ], + "angle": 0, + "content": "PetFace: A Large-Scale Dataset and Benchmark for Animal Identification" + }, + { + "type": "page_number", + "bbox": [ + 0.775, + 0.117, + 0.785, + 0.127 + ], + "angle": 0, + "content": "7" + }, + { + "type": "image", + "bbox": [ + 0.219, + 0.143, + 0.803, + 0.334 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.214, + 0.347, + 0.788, + 0.403 + ], + "angle": 0, + "content": "Fig. 3: Data filtering process. We adopt a two-stage data filtering approach. Initially, images containing multiple faces are automatically removed. Subsequently, any images that do not depict animal faces or fail to meet alignment criteria are manually eliminated." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.432, + 0.788, + 0.522 + ], + "angle": 0, + "content": "AnimalWeb [32]. Because the positions of animal facial parts are sometimes very different from those of humans, we use different reference points for different species in face alignment. After landmark detection, we select one frontal image as a reference for each species. We compute the average landmarks over all the landmarks aligned with the reference. Then, we define the target positions of the landmarks in the images; all the images are aligned to the target positions." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.523, + 0.789, + 0.735 + ], + "angle": 0, + "content": "Data Filtering. Fig 3 shows the overview of our data filtering process. Because the fully automatic face detection above sometimes fails, we adopt a two-stage data filtering process to filter out the following cases: 1) Images contain multiple animals simultaneously. 2) Images are unrelated to animals, such as advertisements. 3) Images focus on non-animal elements, such as random patterns in backgrounds, toys, or people, rather than the animals themselves. First, we automatically remove images where multiple faces are detected in an image. Secondly, annotators manually assess all the images and remove those that do not depict the target animals or where the face alignment differs from our intended criteria. Because checking whether the faces are properly aligned or not requires some expertise, all the images are filtered by the authors to ensure the quality of the dataset. This manual filtering process took about 100 man hours. After this stage, we keep 1,012,934 images, which is approximately \\(70\\%\\) of the initial image number. Detailed distributions of animal families are illustrated in Fig 2(a)." + }, + { + "type": "title", + "bbox": [ + 0.216, + 0.756, + 0.48, + 0.77 + ], + "angle": 0, + "content": "3.4 Fine-Grained Annotations" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.78, + 0.788, + 0.842 + ], + "angle": 0, + "content": "Fine-grained categorization enhances its usefulness for downstream applications, e.g., the creation of challenging datasets based on animal's individual attributes. However, collecting extensive individual animal data poses a major challenge in assembling large datasetse owing to the significant effort involved. In response," + } + ], + [ + { + "type": "page_number", + "bbox": [ + 0.218, + 0.116, + 0.23, + 0.127 + ], + "angle": 0, + "content": "8" + }, + { + "type": "header", + "bbox": [ + 0.272, + 0.115, + 0.462, + 0.128 + ], + "angle": 0, + "content": "R. Shinoda and K. Shiohara" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.147, + 0.783, + 0.252 + ], + "angle": 0, + "content": "we collect as much individual animal information corresponding to the images as possible. Our annotations include sex, color, and specific breed details, depending on their availability on the source websites. The annotation process involves: 1) extracting individual information displayed on websites alongside images, 2) refining raw text data to develop individual data tables, 3) manually verifying that all attributes are refined and free of unrelated information, and 4) manually annotating individuals with missing attributes on the website." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.253, + 0.785, + 0.389 + ], + "angle": 0, + "content": "As a result, we include a sex category for all classes, with breed annotations for eight animal classes (Cat, Chimp, Dog, Guinea pig, Hamster, Parakeet, Pig, and Rabbit), and color annotations for eleven animal classes (Cat, Chinchilla, Degus, Ferret, Guinea pig, Hamster, Hedgehog, Parakeet, Java sparrow, Pig, and Rabbit). Note that the manual annotations are performed only on colors and patterns because images alone do not allow us to accurately determine an animal's sex or breed. Detailed distributions of sex per animal families are illustrated in Fig 2(b). In Fig 2(c) and (d), we illustrate the examples of breed annotations and color and pattern annotations." + }, + { + "type": "title", + "bbox": [ + 0.216, + 0.413, + 0.517, + 0.431 + ], + "angle": 0, + "content": "4 Experimental Methodology" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.446, + 0.785, + 0.522 + ], + "angle": 0, + "content": "We introduce two principal evaluation protocols: 1) re-identification for seen faces, and 2) verification for unseen faces. To establish these benchmarks, we have curated two distinct types of test sets. Specifically, for unseen face verification, we carefully separate the dataset into training, validation, and test sets to prevent data leakage that can arise from shared backgrounds." + }, + { + "type": "title", + "bbox": [ + 0.216, + 0.545, + 0.434, + 0.559 + ], + "angle": 0, + "content": "4.1 Evaluation Protocols" + }, + { + "type": "text", + "bbox": [ + 0.216, + 0.571, + 0.465, + 0.586 + ], + "angle": 0, + "content": "We adopt two types of evaluation." + }, + { + "type": "text", + "bbox": [ + 0.223, + 0.599, + 0.787, + 0.688 + ], + "angle": 0, + "content": "1. Re-identification: This evaluation selects images from the test data that match the identities present in the training data. This procedure involves identifying test images that correspond to the same identities used during model training. The objective is to assess the model's ability to accurately recognize and associate test images with the correct identities from the training dataset." + }, + { + "type": "text", + "bbox": [ + 0.223, + 0.69, + 0.787, + 0.839 + ], + "angle": 0, + "content": "2. Verification: This evaluation checks if the model can identify unseen faces. This means that we use different identities during the training and testing phases. For each identity, we select one image that matches the identity and one image randomly chosen from a different identity. This process creates pairs of images where one pair consists of images from the same identity, and another pair consists of images from different identities. We then task the model with predicting whether the faces in each pair belong to the same identity or different identities. This approach allows us to comprehensively evaluate the model's ability to recognize and differentiate between individual identities." + }, + { + "type": "list", + "bbox": [ + 0.223, + 0.599, + 0.787, + 0.839 + ], + "angle": 0, + "content": null + } + ], + [ + { + "type": "header", + "bbox": [ + 0.242, + 0.115, + 0.732, + 0.129 + ], + "angle": 0, + "content": "PetFace: A Large-Scale Dataset and Benchmark for Animal Identification" + }, + { + "type": "page_number", + "bbox": [ + 0.775, + 0.117, + 0.785, + 0.127 + ], + "angle": 0, + "content": "9" + }, + { + "type": "title", + "bbox": [ + 0.217, + 0.147, + 0.348, + 0.162 + ], + "angle": 0, + "content": "4.2 Data Split" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.177, + 0.788, + 0.344 + ], + "angle": 0, + "content": "To establish benchmarks for seen individuals re-identification and unseen individuals verification, we use specific split protocols, creating two test sets designed for each task. For seen individuals re-identification, the process involves verifying images of the same identities used in training; thus, we select test images from the same individuals for the training set. We, therefore, ensure that each individual is represented by at least three photos i.e., a minimum of two for training and one for testing. For unseen individuals verification dataset, we ensure that it contains no images from sources that are also represented in the training or validation sets per animal family to prevent bias from similar environmental conditions. We divide the dataset into training, validation, and each testing sets following a 7:1:2 ratio as closely as possible, given the outlined criteria." + }, + { + "type": "title", + "bbox": [ + 0.216, + 0.37, + 0.324, + 0.384 + ], + "angle": 0, + "content": "4.3 Models" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.4, + 0.788, + 0.535 + ], + "angle": 0, + "content": "We train state-of-the-art models based on deep neural networks on our PetFace to build the benchmarks. For the re-identification task, we approach the training and testing phases as classification problems; an individual is assigned to a class. On the other hand, for the verification task, we train models in the same manner as for the re-identification task but evaluate the models by computing the cosine similarity between pairs of images to be identified if they are the same individual. Specifically, we focus more on the loss functions that are crucial for identification tasks than network architectures. We refer to three important loss functions in addition to the basic Softmax-based classification model:" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.537, + 0.788, + 0.657 + ], + "angle": 0, + "content": "Triplet loss [26] is designed to take a triplet of samples, i.e., an anchor \\( x_{a} \\), a positive \\( x_{p} \\) (another image of the same identity as the anchor), and a negative \\( x_{n} \\) (an image of a different identity) - and learn embeddings in such a way that the \\( x_{p} \\) is closer to the \\( x_{a} \\) than the \\( x_{n} \\) by a margin. This loss function is particularly beneficial for face Re-ID as it directly targets the relative distances between different and same identity pairs, encouraging the model to learn a feature space where embeddings of the same identity are clustered together while being far from other identity clusters." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.659, + 0.788, + 0.764 + ], + "angle": 0, + "content": "Center loss [65] works alongside Softmax Loss to enhance the discriminative power of the learned features. While Softmax Loss focuses on inter-class separability, Center Loss aims to minimize the intra-class variations. It does this by penalizing the distance between the deep features of each class and their corresponding class center. Center Loss ensures that the embeddings of the faces of the same individual are closer together, thus making the feature distribution more compact for each identity." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.765, + 0.788, + 0.84 + ], + "angle": 0, + "content": "ArcFace loss [15] introduces an angular margin between classes to enforce a discriminative feature space. It modifies the Softmax loss by adding a margin penalty to the angle between the feature vector and the corresponding class center in the angular space. This angular margin encourages models to learn more distinguishable embeddings to separate classes." + } + ], + [ + { + "type": "page_number", + "bbox": [ + 0.218, + 0.116, + 0.236, + 0.127 + ], + "angle": 0, + "content": "10" + }, + { + "type": "header", + "bbox": [ + 0.272, + 0.115, + 0.461, + 0.128 + ], + "angle": 0, + "content": "R. Shinoda and K. Shiohara" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.147, + 0.784, + 0.192 + ], + "angle": 0, + "content": "We use ResNet-50 [23], which we found performs better than recent Transformer-based models in Table 5, as our base backbone for all applied loss functions to simplify our experiments and make them easier to grasp." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.193, + 0.784, + 0.223 + ], + "angle": 0, + "content": "In addition to the comparison on the models trained on PetFace, we evaluate state-of-the-art models trained on other datasets including:" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.223, + 0.784, + 0.253 + ], + "angle": 0, + "content": "ImageNet [14] is a conventional image classification dataset including 1000 general object classes. We use the ResNet-50 trained on the dataset." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.253, + 0.803, + 0.298 + ], + "angle": 0, + "content": "CLIP [54] learns semantic relationships between images and texts in a cross-modal contrastive learning manner empowered by web-scale image-caption datasets. We use ResNet-50 backbone." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.299, + 0.789, + 0.36 + ], + "angle": 0, + "content": "MegaDescriptor [61] is a state-of-the-art model for animal identification trained on a unified animal identification dataset including 33 existing animal re-identification datasets [1-4, 6-9, 11, 12, 16, 18-20, 22, 24, 27, 28, 36, 38, 46, 50-53, 60, 62, 64, 66, 68, 69]. We use the officially provided SwinTransformer-B [41] backbone." + }, + { + "type": "title", + "bbox": [ + 0.215, + 0.384, + 0.452, + 0.401 + ], + "angle": 0, + "content": "5 Experimental Result" + }, + { + "type": "title", + "bbox": [ + 0.215, + 0.417, + 0.633, + 0.431 + ], + "angle": 0, + "content": "5.1 Benchmark on Animal Face Re-identification" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.443, + 0.784, + 0.577 + ], + "angle": 0, + "content": "We show the re-identification results in terms AUC in Table 2. We train models with the baseline loss functions independently on each class. We can see that ArcFace performs consistent results, \\(51.23\\%\\) of average accuracy, compared to the other loss functions (41.88% for Softmax and 9.81% for Center). On the other hand, Center loss does not learn sufficient discriminative features for animal face re-identification, especially on Cat and Dog where the number of the test classes are 113,592 and 46,755, respectively. These results encourage the community to explore more effective representation learning methods for re-identification tasks." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.58, + 0.784, + 0.684 + ], + "angle": 0, + "content": "Moreover, motivated by MegaDescriptor, we train an ArcFace model on the entire PetFace, denoted as Joint-Trained on PetFace in the table. We observe that the joint-training strategy provides some improvements, e.g., from \\(54.29\\%\\) to \\(70.30\\%\\) for Cat and from \\(29.08\\%\\) to \\(41.49\\%\\) for Pig although in some classes the results get worse, e.g., from \\(43.27\\%\\) to \\(34.30\\%\\) for Chimp and from \\(62.19\\%\\) to \\(44.78\\%\\) to Parakeet. This result indicates that there is some room to improve the performance of integrated identification on imbalance and wide-range datasets." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.685, + 0.784, + 0.745 + ], + "angle": 0, + "content": "In Fig 4, we display the Top-\\(k\\) (\\(k = 1,3,5\\)) accuracy metrics. While accuracy naturally increases with larger \\(k\\) values, the relative ranking of accuracy across different animal families remains largely consistent. ArcFace maintains the best performance across these evaluations." + }, + { + "type": "title", + "bbox": [ + 0.215, + 0.769, + 0.591, + 0.783 + ], + "angle": 0, + "content": "5.2 Benchmark on Animal Face Verification" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.795, + 0.784, + 0.84 + ], + "angle": 0, + "content": "Next, we show the verification results in Table 3. We use the trained models in Table 2 and evaluate them on unseen individuals. Similar to our findings in the re-identification task, ArcFace proved to be effective, achieving the best" + } + ], + [ + { + "type": "header", + "bbox": [ + 0.242, + 0.115, + 0.731, + 0.129 + ], + "angle": 0, + "content": "PetFace: A Large-Scale Dataset and Benchmark for Animal Identification" + }, + { + "type": "page_number", + "bbox": [ + 0.768, + 0.116, + 0.784, + 0.127 + ], + "angle": 0, + "content": "11" + }, + { + "type": "table", + "bbox": [ + 0.217, + 0.144, + 0.787, + 0.238 + ], + "angle": 0, + "content": "
MethodTop-1 Accuracy (%)Avg
CatChimpChinchillaDegusDogFerretGuineaHamsterHedgehogParakeetJava sparrowPigRabbit
Trained on PetFace
Softmax30.4641.7058.1337.8459.1425.8860.0738.2727.8150.8833.5521.6459.0541.88
Center0.005.3829.7612.030.007.9131.779.4613.761.684.867.443.519.81
Arcface54.2943.2767.3445.4177.8628.9267.9047.3730.9062.1942.2729.0869.1351.23
Joint-Trained on PetFace
ArcFace70.3034.3069.8656.0868.7546.1268.6654.3344.3844.7834.5541.4965.7553.80
" + }, + { + "type": "table_caption", + "bbox": [ + 0.216, + 0.24, + 0.784, + 0.28 + ], + "angle": 0, + "content": "Table 2: Result on animal re-identification. ArcFace outperforms the other loss functions. The jointly trained model on our PetFace shows the best average top-1 accuracy." + }, + { + "type": "table", + "bbox": [ + 0.217, + 0.306, + 0.787, + 0.441 + ], + "angle": 0, + "content": "
MethodAUC (%)Avg
CatChimpChinchillaDegusDogFerretGuinea pigHamsterHedgehogParakeetJava sparrowPigRabbit
Trained on PetFace
Softmax97.9785.2284.4486.1998.9883.1595.3087.7387.2798.3984.6787.0998.5390.38
Center50.2081.6080.9681.3459.6478.7792.1083.5985.2696.9180.5482.5897.0180.81
Triplet96.9477.1076.1272.8097.9779.4883.3780.0881.4791.5677.8675.7294.7483.48
Arcface97.7183.7687.7087.6199.4586.2496.0390.3286.1398.6384.4589.8899.0191.30
Joint-Trained on PetFace
ArcFace98.0483.2089.6386.7999.0188.4494.1492.5087.8698.4089.3892.3898.4592.17
Trained on Other Dataset
ImageNet81.5873.6770.8070.6997.1867.9585.1974.1375.0190.4579.4982.4286.2379.60
CLIP85.1070.7073.4274.6691.8670.8980.6773.9177.6583.1976.5280.0283.5578.63
MegaDescriptor87.3083.0177.8577.7993.7576.5388.3677.9178.7489.7678.6488.0890.3883.70
" + }, + { + "type": "table_caption", + "bbox": [ + 0.216, + 0.442, + 0.784, + 0.468 + ], + "angle": 0, + "content": "Table 3: Result on animal verification. Jointly trained on our PetFace shows the best performance. We also compared to the trained models on other dataset." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.509, + 0.788, + 0.706 + ], + "angle": 0, + "content": "results with an average AUC of \\(92.17\\%\\) when jointly trained on the PetFace dataset. Similar to face re-identification in Sec 5.1, in some classes, the results get worse when jointly trained across animal families e.g., Chimp, Degus, Dog, Guinea pig, Parakeet, and Rabbit. We also examined models pre-trained on other datasets including ImageNet, CLIP and MegaDescriptor in order to make more comparisons. Among these models, MegaDescriptor, which trained on the animal re-identification dataset, showed the highest AUC of \\(83.70\\%\\). We show the similarity distribution on the our cat dataset in Fig 5. Comparing the performance of CLIP, MegaDescriptor, and our joint-trained models, our findings indicate that our model achieves the most distinct separation between positive and negative samples. In contrast, the MegaDescriptor model displays a closer distribution of these samples, and with CLIP, the overlap between positive and negative distributions is significantly more pronounced." + }, + { + "type": "title", + "bbox": [ + 0.216, + 0.727, + 0.559, + 0.741 + ], + "angle": 0, + "content": "5.3 Comparison with Previous Datasets" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.75, + 0.794, + 0.84 + ], + "angle": 0, + "content": "We conduct cross-dataset evaluations where models are tested on different datasets other than the training datasets to demonstrate the generality of our dataset. Here, we compare our dataset with 1) CTai [19] and CZoo [19] for Chimpanzee and 2) DogFaceNet [49] and Flickr-dog [47] for Dog. All images of the compared datasets are aligned in the same manner as ours for fair comparisons. We split the identities of the compared datasets into training and test sets in a ratio of 7:3" + } + ], + [ + { + "type": "page_number", + "bbox": [ + 0.218, + 0.116, + 0.236, + 0.127 + ], + "angle": 0, + "content": "12" + }, + { + "type": "header", + "bbox": [ + 0.272, + 0.115, + 0.462, + 0.128 + ], + "angle": 0, + "content": "R. Shinoda and K. Shiohara" + }, + { + "type": "image", + "bbox": [ + 0.229, + 0.145, + 0.404, + 0.268 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.287, + 0.273, + 0.345, + 0.285 + ], + "angle": 0, + "content": "(a) Top-1" + }, + { + "type": "image", + "bbox": [ + 0.416, + 0.145, + 0.59, + 0.268 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.472, + 0.273, + 0.532, + 0.285 + ], + "angle": 0, + "content": "(b) Top-3" + }, + { + "type": "image", + "bbox": [ + 0.602, + 0.145, + 0.776, + 0.269 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.659, + 0.273, + 0.717, + 0.285 + ], + "angle": 0, + "content": "(c) Top-5" + }, + { + "type": "image_caption", + "bbox": [ + 0.215, + 0.297, + 0.785, + 0.325 + ], + "angle": 0, + "content": "Fig. 4: Top-k accuracy (\\(k = 1,3,5\\)). ArcFace consistently shows the best performance consistently." + }, + { + "type": "image", + "bbox": [ + 0.228, + 0.342, + 0.407, + 0.463 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.416, + 0.343, + 0.591, + 0.464 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.602, + 0.343, + 0.777, + 0.464 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.214, + 0.474, + 0.785, + 0.531 + ], + "angle": 0, + "content": "Fig. 5: Similarity distributions of (a) CLIP, (b) MegaDescriptor, and (c) our joint-trained model on cats. The horizontal axis represents similarity and the vertical axis represents frequency. The model trained on PetFace shows the most distinct separation between positive and negative samples." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.561, + 0.785, + 0.606 + ], + "angle": 0, + "content": "and then uniformly pick the positive and negative pairs. Because CTai, CZoo, and Flickr-dog have a few test identities, we each pick five positive and negative pairs for each test identity for robust evaluations." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.607, + 0.787, + 0.759 + ], + "angle": 0, + "content": "We show the verification results in Tables 4(a) and (b). We exclude the results trained on the same dataset as the test set denoted in gray for fair comparisons. For the Chimpanzee datasets, the model trained on our dataset outperforms previous datasets. For case tested on CTai, our dataset outperforms CZoo (67.33% vs. 66.76%). For the other case tested on CZoo, our dataset also surpasses CTai (71.27% vs. 69.31%). This result demonstrates that our Chimpanzee dataset is more effective and general than the previous datasets. For the Dog datasets, owing to the significant scale of our dataset, the model trained on our dog dataset outperforms not only cross-dataset results but also in-dataset ones denoted in gray. These results support the quality of our PetFace dataset." + }, + { + "type": "title", + "bbox": [ + 0.216, + 0.783, + 0.334, + 0.798 + ], + "angle": 0, + "content": "5.4 Analysis" + }, + { + "type": "text", + "bbox": [ + 0.215, + 0.81, + 0.785, + 0.84 + ], + "angle": 0, + "content": "Different networks. To examine the impact of different network architectures on the performance on PetFace, we additionally train two backbones," + } + ], + [ + { + "type": "header", + "bbox": [ + 0.242, + 0.115, + 0.731, + 0.129 + ], + "angle": 0, + "content": "PetFace: A Large-Scale Dataset and Benchmark for Animal Identification" + }, + { + "type": "page_number", + "bbox": [ + 0.769, + 0.116, + 0.785, + 0.127 + ], + "angle": 0, + "content": "13" + }, + { + "type": "table", + "bbox": [ + 0.217, + 0.144, + 0.477, + 0.217 + ], + "angle": 0, + "content": "
Train SetTest Set AUC (%)
Database#Ind.#ImagesCTaiCZoo
CTai49285775.8069.31
CZoo17132366.7680.49
Ours-Chimp446267967.3371.27
" + }, + { + "type": "table_caption", + "bbox": [ + 0.307, + 0.219, + 0.404, + 0.229 + ], + "angle": 0, + "content": "(a) Chimp datasets." + }, + { + "type": "table", + "bbox": [ + 0.488, + 0.144, + 0.77, + 0.217 + ], + "angle": 0, + "content": "
Train SetTest Set AUC (%)
Database#Ind.#ImagesDogFaceNetFlickr-dog
DogFaceNet975587998.7285.51
Flickr-dog3026594.9887.77
Ours-Dog4675516834899.6595.54
" + }, + { + "type": "table_caption", + "bbox": [ + 0.595, + 0.219, + 0.68, + 0.229 + ], + "angle": 0, + "content": "(b) Dog datasets." + }, + { + "type": "table_caption", + "bbox": [ + 0.214, + 0.238, + 0.785, + 0.292 + ], + "angle": 0, + "content": "Table 4: Comparisons with previous datasets for Chimp and Dog. The results when the test sets are the same as the training datasets are excluded for fair comparisons. The models trained on our dataset outperform those trained on previous datasets." + }, + { + "type": "table", + "bbox": [ + 0.216, + 0.325, + 0.787, + 0.384 + ], + "angle": 0, + "content": "
MethodAUC (%)Avg
CatChimpChinchillaDegusDogFerretGuinea pigHamsterHedgehogParakeetJava sparrowPigRabbit
ResNet-5097.7183.7687.7087.6199.4586.2496.0390.3286.1398.6384.4589.8899.0191.30
ViT-32-B94.4478.6071.1382.6993.3968.1783.2177.8987.7795.0780.6275.3195.1883.35
Swin-B87.3574.2376.2184.0485.7571.6091.5688.3386.2294.1284.6666.5690.5683.17
" + }, + { + "type": "table_caption", + "bbox": [ + 0.214, + 0.384, + 0.784, + 0.411 + ], + "angle": 0, + "content": "Table 5: Verification result on different backbones. ResNet-50 achieves the best average AUC." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.458, + 0.785, + 0.519 + ], + "angle": 0, + "content": "VisionTransformer-32-B [17] and SwinTransformer-B [41], with the same training configuration as our base model (ResNet-50) and report the verification results in Table 5. We observe that ResNet-50 outperforms the additional transformer-based models in most cases and achieves the best average AUC of \\(91.30\\%\\)." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.521, + 0.787, + 0.793 + ], + "angle": 0, + "content": "Fine-grained verification. In real world scenarios, animal verification is sometimes conducted within a specific breed rather than different breeds. To meet this demand, we introduce fine-grained verification where positive and negative pairs are chosen only from the same breeds. We refer to the top-10 dog breeds of individuals in our datasets: Chihuahua, Dachshund, French bulldog, Golden retriever, Miniature dachshund, Pomeranian, Shiba inu, Shih tzu, Toy poodle, and Yorkshire terrier. We construct the fine-grained test sets and evaluate the models on each breed. The results are visualized in Fig. 6. It can be seen that our model achieves a good performance on all the breeds (98.30% on average). The results achieved for the baseline models are lower in comparison (84.99%, 80.78%, and 86.25% for ImageNet, CLIP, and MegaDescriptor, respectively). Moreover, we find that the results achieved with MegaDescriptor show significantly more degradation than the cross-breed verification result in Table 3 from 93.75% to 86.25%. while our model maintains the results at a high-level (from 99.01% to 98.39%). This result indicates that our dataset overcomes the limitation of the previous dog dataset [69] that suffers from the poor performance for specific breed verification owing to its insufficient number of individuals (192 individuals)." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.795, + 0.785, + 0.84 + ], + "angle": 0, + "content": "Generality to Unseen Families. Lastly, we evaluate the verification generality to unseen animal families. To achieve this, we additionally collect 100 identities each of Parrot, Lacertilia, and Squirrel, which are not contained in either our" + } + ], + [ + { + "type": "page_number", + "bbox": [ + 0.218, + 0.116, + 0.236, + 0.127 + ], + "angle": 0, + "content": "14" + }, + { + "type": "header", + "bbox": [ + 0.272, + 0.115, + 0.462, + 0.128 + ], + "angle": 0, + "content": "R. Shinoda and K. Shiohara" + }, + { + "type": "image", + "bbox": [ + 0.224, + 0.151, + 0.501, + 0.326 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.513, + 0.148, + 0.772, + 0.323 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.218, + 0.338, + 0.505, + 0.381 + ], + "angle": 0, + "content": "Fig. 6: Fine-grained dog verification. Trained on our PetFace Dog achieves the best performance consistently." + }, + { + "type": "image_caption", + "bbox": [ + 0.508, + 0.335, + 0.779, + 0.39 + ], + "angle": 0, + "content": "Fig.7: Generality to unseen animal families. Trained on our PetFace achieves the best performance even for unseen animal families." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.421, + 0.788, + 0.511 + ], + "angle": 0, + "content": "dataset or WildLifeDataset [61]. We compare our model that is jointly trained on our dataset across families with the baseline models in Fig. 7. Our model outperforms the baselines on all the species. For Parrot, particularly, our model achieves an AUC of \\(88.99\\%\\), revealing its superiority over MegaDescritor with a large margin. This result supports the generality and variety of our PetFace dataset." + }, + { + "type": "title", + "bbox": [ + 0.216, + 0.536, + 0.36, + 0.551 + ], + "angle": 0, + "content": "6 Conclusion" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.567, + 0.788, + 0.687 + ], + "angle": 0, + "content": "We introduced the PetFace, a comprehensive animal identification dataset encompassing 13 families, featuring 257,484 unique individuals across 319 breed categories. We have collected images and their information and conducted automated and manual filtering to ensure the quality of the large-scale dataset. This dataset also includes detailed annotations of sex, breed, colors, and patterns to facilitate more investigation and application in a real-world scenario. We establish two main benchmarks: 1) re-identification of seen individuals and 2) verification of unseen individuals." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.688, + 0.788, + 0.794 + ], + "angle": 0, + "content": "Our experiments show the generality of the models trained on our dataset to verification on other datasets or unseen animal families. We also found that there is still room for improvement in the integrated identification of multiple animal families. To promote further research, we will make this dataset, experiment code, and models available to the research communities. This dataset will enable computer vision researchers to tackle animal face identification across a wide range of breeds and push forward progress on animal face identification tasks." + } + ], + [ + { + "type": "header", + "bbox": [ + 0.242, + 0.115, + 0.732, + 0.129 + ], + "angle": 0, + "content": "PetFace: A Large-Scale Dataset and Benchmark for Animal Identification" + }, + { + "type": "page_number", + "bbox": [ + 0.768, + 0.116, + 0.786, + 0.127 + ], + "angle": 0, + "content": "15" + }, + { + "type": "title", + "bbox": [ + 0.218, + 0.145, + 0.403, + 0.163 + ], + "angle": 0, + "content": "Acknowledgements" + }, + { + "type": "text", + "bbox": [ + 0.216, + 0.178, + 0.788, + 0.24 + ], + "angle": 0, + "content": "This work was partially supported by the Cooperative Research Program of the Primate Research Institute, Kyoto University. We thank Dr. Takuma Yagi from AIST for comments on paper proofreading. We also thank Prof. Toshihiko Yamasaki from the University of Tokyo for providing computation resources." + }, + { + "type": "title", + "bbox": [ + 0.217, + 0.264, + 0.324, + 0.279 + ], + "angle": 0, + "content": "References" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.296, + 0.751, + 0.311 + ], + "angle": 0, + "content": "1. Beluga id 2022 (2022), https://lila.science/datasets/beluga-id-2022/" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.312, + 0.785, + 0.338 + ], + "angle": 0, + "content": "2. Hyena id 2022 botswana predator conservation trust (2022). panthera pardus csv custom export. (2022), https://lila.science/datasets/hyena-id-2022/" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.34, + 0.785, + 0.366 + ], + "angle": 0, + "content": "3. Leopard id 2022 botswana predator conservation trust (2022). panthera pardus csv custom export. (2022), https://lila.science/datasets/leopard-id-2022/" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.368, + 0.785, + 0.394 + ], + "angle": 0, + "content": "4. Turtle recall: Conservation challenge (2022), https://zindi.africa/competitions/turtle-recall-conservation-challenge" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.396, + 0.785, + 0.435 + ], + "angle": 0, + "content": "5. Adam, L., Čermák, V., Papafitsoros, K., Picek, L.: Seaturtleid2022: A long-span dataset for reliable sea turtle re-identification. In: IEEE/CVF Winter Conference on Applications of Computer Vision (WACV). pp. 7146-7156 (2024)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.437, + 0.785, + 0.478 + ], + "angle": 0, + "content": "6. Andrew, W., Gao, J., Mullan, S., Campbell, N., Dowsey, A.W., Burghardt, T.: Visual identification of individual holstein-friesian cattle via deep metric learning. In: Computers and Electronics in Agriculture. vol. 185, p. 106133 (2021)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.479, + 0.785, + 0.52 + ], + "angle": 0, + "content": "7. Andrew, W., Greatwood, C., Burghardt, T.: Visual localisation and individual identification of holstein friesian cattle via deep learning. In: IEEE/CVF International Conference on Computer Vision (ICCV) Workshops. pp. 2850-2859 (2017)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.522, + 0.785, + 0.561 + ], + "angle": 0, + "content": "8. Andrew, W., Greatwood, C., Burghardt, T.: Visual localisation and individual identification of holstein friesian cattle via deep learning. In: IEEE/CVF International Conference on Computer Vision (ICCV) Workshops. pp. 2850-2859 (2017)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.563, + 0.785, + 0.616 + ], + "angle": 0, + "content": "9. Andrew, W., Hannuna, S., Campbell, N., Burghardt, T.: Automatic individual holstein friesian cattle identification via selective local coat pattern matching in rgb-d imagery. In: International Conference on Image Processing (ICIP). pp. 484-488 (2016)" + }, + { + "type": "ref_text", + "bbox": [ + 0.221, + 0.618, + 0.785, + 0.658 + ], + "angle": 0, + "content": "10. Brookes, O., Burghardt, T.: A dataset and application for facial recognition of individual gorillas in zoo environments. In: International Conference on Pattern Recognition(ICPR) (2020)" + }, + { + "type": "ref_text", + "bbox": [ + 0.22, + 0.66, + 0.785, + 0.728 + ], + "angle": 0, + "content": "11. Cheeseman, T., Southerland, K., Park, J., Olio, M., Flynn, K., Calambokidis, J., Jones, L., Garrigue, C., Frisch-Jordán, A., Howard, A., Reade, W., Neilson, J., Gabriele, C., Clapham, P.: Advanced image recognition: a fully automated, high-accuracy photo-identification matching system for humpback whales. vol. 102, p. 915–929 (2021)" + }, + { + "type": "ref_text", + "bbox": [ + 0.22, + 0.729, + 0.785, + 0.756 + ], + "angle": 0, + "content": "12. Christin B. Khan, Shashank, W.K.: Right whale recognition (2015), https://kaggle.com/competitions/noaa-right-whale-recognition" + }, + { + "type": "ref_text", + "bbox": [ + 0.22, + 0.758, + 0.785, + 0.798 + ], + "angle": 0, + "content": "13. Dahlborn, K., Bugnon, P., Nevalainen, T., Raspa, M., Verbost, P.M., Spangenberg, E.M.F.: Report of the federation of european laboratory animal science associations working group on animal identification. vol. 47, pp. 2-11 (2013)" + }, + { + "type": "ref_text", + "bbox": [ + 0.22, + 0.8, + 0.785, + 0.84 + ], + "angle": 0, + "content": "14. Deng, J., Dong, W., Socher, R., Li, L.J., Li, K., Fei-Fei, L.: Imagenet: A large-scale hierarchical image database. In: IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR). pp. 248-255 (2009)" + }, + { + "type": "list", + "bbox": [ + 0.22, + 0.296, + 0.785, + 0.84 + ], + "angle": 0, + "content": null + } + ], + [ + { + "type": "page_number", + "bbox": [ + 0.218, + 0.116, + 0.236, + 0.127 + ], + "angle": 0, + "content": "16" + }, + { + "type": "header", + "bbox": [ + 0.272, + 0.115, + 0.462, + 0.128 + ], + "angle": 0, + "content": "R. Shinoda and K. Shiohara" + }, + { + "type": "ref_text", + "bbox": [ + 0.218, + 0.148, + 0.785, + 0.189 + ], + "angle": 0, + "content": "15. Deng, J., Guo, J., Xue, N., Zafeiriou, S.: Arcface: Additive angular margin loss for deep face recognition. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR). pp. 4690-4699 (2019)" + }, + { + "type": "ref_text", + "bbox": [ + 0.218, + 0.19, + 0.785, + 0.232 + ], + "angle": 0, + "content": "16. Dlamini, N., Zyl, T.L.v.: Automated identification of individuals in wildlife population using siamese neural networks. In: 2020 7th International Conference on Soft Computing Machine Intelligence (ISCMI). pp. 224-228 (2020)" + }, + { + "type": "ref_text", + "bbox": [ + 0.218, + 0.232, + 0.785, + 0.286 + ], + "angle": 0, + "content": "17. Dosovitskiy, A., Beyer, L., Kolesnikov, A., Weissenborn, D., Zhai, X., Unterthiner, T., Dehghani, M., Minderer, M., Heigold, G., Gelly, S., Uszkoreit, J., Houlsby, N.: An image is worth 16x16 words: Transformers for image recognition at scale. The International Conference on Learning Representations (ICLR) (2021)" + }, + { + "type": "ref_text", + "bbox": [ + 0.218, + 0.287, + 0.785, + 0.328 + ], + "angle": 0, + "content": "18. Ferreira, A.C., Silva, L.R., Renna, F., Brandl, H.B., Renoult, J., Farine, D.R., Covas, R., Doutrelant, C.: Deep learning-based methods for individual recognition in small birds. In: Methods in Ecology and Evolution. vol. 11, pp. 1072-1085 (2020)" + }, + { + "type": "ref_text", + "bbox": [ + 0.218, + 0.328, + 0.785, + 0.384 + ], + "angle": 0, + "content": "19. Freytag, A., Rodner, E., Simon, M., Loos, A., Kuhl, H.S., Denzler, J.: Chimpanzee faces in the wild: Log-euclidean cnns for predicting identities and attributes of primates. In: Rosenhahn, B., Andres, B. (eds.) Pattern Recognition. pp. 51-63. Springer International Publishing, Cham (2016)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.384, + 0.785, + 0.425 + ], + "angle": 0, + "content": "20. Gao, J., Burghardt, T., Andrew, W., Dowsey, A.W., Campbell, N.W.: Towards self-supervision for video identification of individual holstein-friesian cattle: The cows2021 dataset. arXiv preprint arXiv:2105.01938 (2021)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.425, + 0.785, + 0.48 + ], + "angle": 0, + "content": "21. Guo, Y., Zhang, L., Hu, Y., He, X., Gao, J.: Ms-celeb-1m: A dataset and benchmark for large-scale face recognition. In: Leibe, B., Matas, J., Sebe, N., Welling, M. (eds.) European Conference on Computer Vision (ECCV). pp. 87-102. Springer International Publishing, Cham (2016)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.481, + 0.785, + 0.522 + ], + "angle": 0, + "content": "22. Haurum, J.B., Karpova, A., Pedersen, M., Bengtson, S.H., Moeslund, T.B.: Identification of zebrafish using metric learning. In: IEEE Winter Conference on Applications of Computer Vision (WACV) Workshops. pp. 1-11 (2020)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.522, + 0.785, + 0.564 + ], + "angle": 0, + "content": "23. He, K., Zhang, X., Ren, S., Sun, J.: Deep residual learning for image recognition. In: IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR). pp. 770-778 (2016)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.564, + 0.785, + 0.605 + ], + "angle": 0, + "content": "24. He, Z., Qian, J., Yan, D., Wang, C., Xin, Y.: Animal re-identification algorithm for posture diversity. In: ICASSP 2023 - 2023 IEEE International Conference on Acoustics, Speech and Signal Processing (ICASSP). pp. 1-5 (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.606, + 0.785, + 0.633 + ], + "angle": 0, + "content": "25. Hitelman, A., Edan, Y., Godo, A., Berenstein, R., Lepar, J., Halachmi, I.: Biometric identification of sheep via a machine-vision system. vol. 194, p. 106713 (2022)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.633, + 0.785, + 0.647 + ], + "angle": 0, + "content": "26. Hoffer, E., Ailon, N.: Deep metric learning using triplet network. In: arXiv (2018)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.647, + 0.785, + 0.688 + ], + "angle": 0, + "content": "27. Holmberg, J., Norman, B., Arzoumanian, Z.: Estimating population size, structure, and residency time for whale sharks rhincodon typus through collaborative photo-identification. vol. 7, pp. 39 - 53. Inter-Research Science Center (2009)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.689, + 0.785, + 0.73 + ], + "angle": 0, + "content": "28. Howard, A., inversion, Southerland, K., Cheeseman, T.: Humpback whale identification challenge (2018), https://kaggle.com/competitions/whale-categorization-playground" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.73, + 0.785, + 0.771 + ], + "angle": 0, + "content": "29. Huang, G.B., Ramesh, M., Berg, T., Learned-Miller, E.: Labeled faces in the wild: A database for studying face recognition in unconstrained environments. No. 07-49 (2007)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.772, + 0.785, + 0.813 + ], + "angle": 0, + "content": "30. Jarraya, I., Ouarda, W., Alimi, A.M.: A preliminary investigation on horses recognition using facial texture features. In: IEEE International Conference on Systems, Man, and Cybernetics. pp. 2803-2808 (2015)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.813, + 0.785, + 0.84 + ], + "angle": 0, + "content": "31. Khan, F.F., Li, X., Temple, A.J., Elhoseiny, M.: Fishnet: A large-scale dataset and benchmark for fish recognition, detection, and functional trait prediction. In: Pro" + }, + { + "type": "list", + "bbox": [ + 0.217, + 0.148, + 0.785, + 0.84 + ], + "angle": 0, + "content": null + } + ], + [ + { + "type": "header", + "bbox": [ + 0.242, + 0.115, + 0.732, + 0.129 + ], + "angle": 0, + "content": "PetFace: A Large-Scale Dataset and Benchmark for Animal Identification" + }, + { + "type": "page_number", + "bbox": [ + 0.768, + 0.116, + 0.786, + 0.127 + ], + "angle": 0, + "content": "17" + }, + { + "type": "ref_text", + "bbox": [ + 0.245, + 0.147, + 0.785, + 0.175 + ], + "angle": 0, + "content": "ceedings of the IEEE/CVF International Conference on Computer Vision (ICCV). pp. 20496-20506 (October 2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.218, + 0.175, + 0.786, + 0.23 + ], + "angle": 0, + "content": "32. Khan, M.H., McDonagh, J., Khan, S., Shahabuddin, M., Arora, A., Khan, F.S., Shao, L., Tzimiropoulos, G.: Animalweb: A large-scale hierarchical dataset of annotated animal faces. In: IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR). pp. 6937-6946 (2020)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.23, + 0.785, + 0.257 + ], + "angle": 0, + "content": "33. Kim, M., Jain, A.K., Liu, X.: Adaface: Quality adaptive margin for face recognition. pp. 18729-18738 (2022)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.257, + 0.785, + 0.284 + ], + "angle": 0, + "content": "34. Krishnan, S., S., S.: Cloud iot based novel livestock monitoring and identification system using uid. In: Sensor Review. vol. 38, pp. 21-33 (2017)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.284, + 0.785, + 0.324 + ], + "angle": 0, + "content": "35. Kuzdeuov, A., Koishigarina, D., Varol, H.A.: Anyface: A data-centric approach for input-agnostic face detection. In: IEEE International Conference on Big Data and Smart Computing (BigComp). pp. 211-218 (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.324, + 0.785, + 0.379 + ], + "angle": 0, + "content": "36. Lahiri, M., Tantipathananandh, C., Warungu, R., Rubenstein, D.I., Berger-Wolf, T.Y.: Biometric animal databases from field photographs: identification of individual zebra in the wild. In: Proceedings of the ACM International Conference on Multimedia Retrieval. Association for Computing Machinery (2011)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.379, + 0.785, + 0.433 + ], + "angle": 0, + "content": "37. Leslie, E., Hernández-Jover, M., Newman, R., Holyoake, P.: Assessment of acute pain experienced by piglets from ear tagging, ear notching and intraperitoneal injectable transponders. In: Applied Animal Behaviour Science. vol. 127, pp. 86-95 (2010)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.433, + 0.785, + 0.488 + ], + "angle": 0, + "content": "38. Li, S., Li, J., Tang, H., Qian, R., Lin, W.: Atrw: A benchmark for amur tiger re-identification in the wild. In: Proceedings of the 28th ACM International Conference on Multimedia. p. 2590–2598. MM '20, Association for Computing Machinery, New York, NY, USA (2020)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.488, + 0.785, + 0.529 + ], + "angle": 0, + "content": "39. Li, X., Xiang, Y., Li, S.: Combining convolutional and vision transformer structures for sheep face recognition. In: Computers and Electronics in Agriculture. vol. 205, p. 107651 (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.529, + 0.785, + 0.57 + ], + "angle": 0, + "content": "40. Liu, W., Wen, Y., Yu, Z., Li, M., Raj, B., Song, L.: Spherface: Deep hypersphere embedding for face recognition. In: IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR). pp. 6738-6746 (2017)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.569, + 0.785, + 0.61 + ], + "angle": 0, + "content": "41. Liu, Z., Lin, Y., Cao, Y., Hu, H., Wei, Y., Zhang, Z., Lin, S., Guo, B.: Swin transformer: Hierarchical vision transformer using shifted windows. In: IEEE/CVF International Conference on Computer Vision (ICCV). pp. 9992-10002" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.61, + 0.785, + 0.651 + ], + "angle": 0, + "content": "42. Loos, A., Ernst, A.: An automated chimpanzee identification system using face detection and recognition. In: EURASIP Journal on Image and Video Processing. vol. 2013 (2013)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.651, + 0.785, + 0.692 + ], + "angle": 0, + "content": "43. Marsot, M., Mei, J., Shan, X., Ye, L., Feng, P., Yan, X., Li, C., Zhao, Y.: An adaptive pig face recognition approach using convolutional neural networks. In: Computers and Electronics in Agriculture. vol. 173, p. 105386 (2020)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.692, + 0.785, + 0.718 + ], + "angle": 0, + "content": "44. Martvel, G., Shimshoni, I., Zamansky, A.: Automated detection of cat facial landmarks. In: International Journal of Computer Vision (IJCV). pp. 1-16 (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.718, + 0.785, + 0.759 + ], + "angle": 0, + "content": "45. Mazlan, N.H., Lopez Salesansky, N., Burn, C., Wells, D.: Mouse identification methods and potential welfare issues: A survey of current practice in the uk. In: Animal Technology and Welfare. vol. 13, pp. 1-10 (2014)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.759, + 0.785, + 0.8 + ], + "angle": 0, + "content": "46. Miele, V., Dussert, G., Spataro, B., Chamaillé-Jammes, S., Allainé, D., Bonenfant, C.: Revisiting animal photo-identification using deep metric learning and network analysis. In: Methods in Ecology and Evolution. vol. 12, pp. 863-873. Wiley (2021)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.8, + 0.785, + 0.84 + ], + "angle": 0, + "content": "47. Moreira, T., Perez, M., Werneck, R., Valle, E.: Where is my puppy? retrieving lost dogs by facial features. In: Multimedia Tools and Applications. vol. 76, p. 15325-15340 (2017)" + }, + { + "type": "list", + "bbox": [ + 0.217, + 0.147, + 0.786, + 0.84 + ], + "angle": 0, + "content": null + } + ], + [ + { + "type": "page_number", + "bbox": [ + 0.218, + 0.116, + 0.236, + 0.127 + ], + "angle": 0, + "content": "18" + }, + { + "type": "header", + "bbox": [ + 0.272, + 0.115, + 0.462, + 0.128 + ], + "angle": 0, + "content": "R. Shinoda and K. Shiohara" + }, + { + "type": "ref_text", + "bbox": [ + 0.218, + 0.147, + 0.787, + 0.203 + ], + "angle": 0, + "content": "48. Moschoglou, S., Papaioannou, A., Sagonas, C., Deng, J., Kotsia, I., Zafeiriou, S.: Agedb: The first manually collected, in-the-wild age database. In: IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR) Workshops. pp. 1997-2005 (2017)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.203, + 0.787, + 0.243 + ], + "angle": 0, + "content": "49. Mougeot, G., Li, D., Jia, S.: A deep learning approach for dog face verification and recognition. In: Nayak, A.C., Sharma, A. (eds.) PRICAI 2019: Trends in Artificial Intelligence. pp. 418-430. Springer International Publishing, Cham (2019)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.243, + 0.787, + 0.271 + ], + "angle": 0, + "content": "50. Murali, N., Taylor, G., Levine, J.: Can drosophila melanogaster tell who's who? In: PLOS ONE. vol. 13, p. e0205043 (2018)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.271, + 0.787, + 0.311 + ], + "angle": 0, + "content": "51. Nepovinnykh, E., Eerola, T., Biard, V., Mutka, P., Niemi, M., Kunnasranta, M., Kalviainen, H.: Sealid: Saimaa ringed seal re-identification dataset. Sensors 22(19) (2022)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.311, + 0.787, + 0.352 + ], + "angle": 0, + "content": "52. Papafitsoros, K., Adam, L., Čermák, V., Picek, L.: Seaturtleid: A novel long-span dataset highlighting the importance of timestamps in wildlife re-identification. In: arXiv (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.352, + 0.787, + 0.393 + ], + "angle": 0, + "content": "53. Parham, J., Crall, J., Stewart, C., Berger-Wolf, T., Rubenstein, D.: Animal population censusing at scale with citizen science and photographic identification. In: SS-17-01. pp. 37-44. AAAI Spring Symposium (2017)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.393, + 0.787, + 0.434 + ], + "angle": 0, + "content": "54. Radford, A., Kim, J.W., Hallacy, C., Ramesh, A., Goh, G., Agarwal, S., Sastry, G., Askell, A., Mishkin, P., Clark, J., Krueger, G., Sutskever, I.: Learning transferable visual models from natural language supervision. In: arXiv (2021)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.434, + 0.787, + 0.461 + ], + "angle": 0, + "content": "55. Roberts, C.: Radio frequency identification (rfid). In: Computers Security. vol. 25, pp. 18-26 (2006)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.461, + 0.787, + 0.501 + ], + "angle": 0, + "content": "56. Roughan, J., Sevenoaks, T.: Welfare and scientific considerations of tattooing and ear-tagging for mouse identification. In: Journal of the American Association for Laboratory Animal Science: JAALAS. pp. 142-153 (2018)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.501, + 0.787, + 0.542 + ], + "angle": 0, + "content": "57. Sengupta, S., Chen, J.C., Castillo, C., Patel, V.M., Chellappa, R., Jacobs, D.W.: Frontal to profile face verification in the wild. In: IEEE Winter Conference on Applications of Computer Vision (WACV). pp. 1-9 (2016)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.542, + 0.787, + 0.582 + ], + "angle": 0, + "content": "58. Shi, X., Yang, C., Xia, X., Chai, X.: Deep cross-species feature learning for animal face recognition via residual interspecies equivariant network. In: European Conference on Computer Vision (ECCV). p. 667-682 (2020)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.582, + 0.787, + 0.609 + ], + "angle": 0, + "content": "59. Shinoda, R., Shiohara, K.: Openanimaltracks: A dataset for animal track recognition (2024)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.609, + 0.787, + 0.65 + ], + "angle": 0, + "content": "60. Trotter, C., Atkinson, G., Sharpe, M., Richardson, K., McGough, A.S., Wright, N., Burville, B., Berggren, P.: Ndd20: A large-scale few-shot dolphin dataset for coarse and fine-grained categorisation. In: arXiv (2020)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.65, + 0.787, + 0.691 + ], + "angle": 0, + "content": "61. Čermák, V., Picek, L., Adam, L., Papafitsoros, K.: Wildlifedatasets: An open-source toolkit for animal re-identification. In: IEEE/CVF Winter Conference on Applications of Computer Vision (WACV). pp. 5953-5963 (2024)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.691, + 0.787, + 0.732 + ], + "angle": 0, + "content": "62. Wahltinez, O., Wahltinez, S.: An open-source general purpose machine learning framework for individual animal re-identification using few-shot learning. In: Methods in Ecology and Evolution. vol. 15, pp. 373-387 (2024)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.732, + 0.787, + 0.772 + ], + "angle": 0, + "content": "63. Wang, H., Wang, Y., Zhou, Z., Ji, X., Gong, D., Zhou, J., Li, Z., Liu, W.: Cosface: Large margin cosine loss for deep face recognition. In: IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR). pp. 5265-5274 (2018)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.772, + 0.787, + 0.799 + ], + "angle": 0, + "content": "64. Wang, L., Ding, R., Zhai, Y., Zhang, Q., Tang, W., Zheng, N., Hua, G.: Giant panda identification. IEEE Transactions on Image Processing 30, 2837-2849 (2021)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.799, + 0.787, + 0.841 + ], + "angle": 0, + "content": "65. Wen, Y., Zhang, K., Li, Z., Qiao, Y.: A discriminative feature learning approach for deep face recognition. In: Leibe, B., Matas, J., Sebe, N., Welling, M. (eds.) European Conference on Computer Vision (ECCV). pp. 499-515 (2016)" + }, + { + "type": "list", + "bbox": [ + 0.217, + 0.147, + 0.787, + 0.841 + ], + "angle": 0, + "content": null + } + ], + [ + { + "type": "header", + "bbox": [ + 0.242, + 0.115, + 0.732, + 0.129 + ], + "angle": 0, + "content": "PetFace: A Large-Scale Dataset and Benchmark for Animal Identification" + }, + { + "type": "page_number", + "bbox": [ + 0.768, + 0.116, + 0.786, + 0.127 + ], + "angle": 0, + "content": "19" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.147, + 0.788, + 0.175 + ], + "angle": 0, + "content": "66. Witham, C.L.: Automated face recognition of rhesus macaques. In: Journal of Neuroscience Methods. vol. 300, pp. 157-165 (2018)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.176, + 0.788, + 0.231 + ], + "angle": 0, + "content": "67. Zhu, Z., Huang, G., Deng, J., Ye, Y., Huang, J., Chen, X., Zhu, J., Yang, T., Lu, J., Du, D., Zhou, J.: Webface260m: A benchmark unveiling the power of million-scale deep face recognition. In: IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR). pp. 10492-10502 (2021)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.232, + 0.787, + 0.273 + ], + "angle": 0, + "content": "68. Zuerl, M., Dirauf, R., Koeferl, F., Steinlein, N., Sueskind, J., Zanca, D., Brehm, I., Fersen, L.v., Eskofier, B.: Polarbearvidid: A video-based re-identification benchmark dataset for polar bears. In: Animals. vol. 13 (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.273, + 0.787, + 0.315 + ], + "angle": 0, + "content": "69. Zuffi, S., Kanazawa, A., Berger-Wolf, T., Black, M.J.: Three-d safari: Learning to estimate zebra pose, shape, and texture from images \"in the wild\". In: IEEE/CVF International Conference on Computer Vision (ICCV). pp. 5358-5367 (2019)" + }, + { + "type": "list", + "bbox": [ + 0.217, + 0.147, + 0.788, + 0.315 + ], + "angle": 0, + "content": null + } + ] +] \ No newline at end of file diff --git a/2024/PetFace_ A Large-Scale Dataset and Benchmark for Animal Identification/79b48743-23c5-4624-b76c-ce49a285a1a7_origin.pdf b/2024/PetFace_ A Large-Scale Dataset and Benchmark for Animal Identification/79b48743-23c5-4624-b76c-ce49a285a1a7_origin.pdf new file mode 100644 index 0000000000000000000000000000000000000000..c429e7048fd891091988002227ee0fd099c4b50c --- /dev/null +++ b/2024/PetFace_ A Large-Scale Dataset and Benchmark for Animal Identification/79b48743-23c5-4624-b76c-ce49a285a1a7_origin.pdf @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:57390c2b3beaeaaba2c269c3f66e1acedbf7b13167c3c2a46469c76c27f91450 +size 4588891 diff --git a/2024/PetFace_ A Large-Scale Dataset and Benchmark for Animal Identification/full.md b/2024/PetFace_ A Large-Scale Dataset and Benchmark for Animal Identification/full.md new file mode 100644 index 0000000000000000000000000000000000000000..e6c86899b03beb9718be189bc74553257cc09e89 --- /dev/null +++ b/2024/PetFace_ A Large-Scale Dataset and Benchmark for Animal Identification/full.md @@ -0,0 +1,307 @@ +# PetFace: A Large-Scale Dataset and Benchmark for Animal Identification + +Risa Shinoda $^{1*}$ and Kaede Shiohara $^{2*}$ + +$^{*}$ Equal contribution + $^{1}$ Kyoto University, Japan $^{2}$ The University of Tokyo, Japan + +Abstract. Automated animal face identification plays a crucial role in the monitoring of behaviors, conducting of surveys, and finding of lost animals. Despite the advancements in human face identification, the lack of datasets and benchmarks in the animal domain has impeded progress. In this paper, we introduce the PetFace dataset, a comprehensive resource for animal face identification encompassing 257,484 unique individuals across 13 animal families and 319 breed categories, including both experimental and pet animals. This large-scale collection of individuals facilitates the investigation of unseen animal face verification, an area that has not been sufficiently explored in existing datasets due to the limited number of individuals. Moreover, PetFace also has fine-grained annotations such as sex, breed, color, and pattern. We provide multiple benchmarks including re-identification for seen individuals and verification for unseen individuals. The models trained on our dataset outperform those trained on prior datasets, even for detailed breed variations and unseen animal families. Our result also indicates that there is some room to improve the performance of integrated identification on multiple animal families. We hope the PetFace dataset will facilitate animal face identification and encourage the development of non-invasive animal automatic identification methods. Our dataset and code are available at https://dahlian00.github.io/PetFacePage/. + +Keywords: Animals $\cdot$ Re-identification $\cdot$ Face Recognition + +# 1 Introduction + +Animal identification plays a crucial role in animal studies and applications such as monitoring animal behavior, conducting habitat surveys, locating missing animals, and performing health checks. Traditional identification techniques, including ear tags, tattoos, ear punching, and toe clipping, continue to be utilized mainly for experimental animals and livestock [13]. However, given that these methods have the potential to cause stress and pain [37,45,56], their use should be minimized to prioritize animal welfare. Therefore, there is a pressing need for the development and adoption of identification technologies that are not only effective and efficient but also minimally invasive, thereby mitigating the ethical concerns associated with traditional methods. While advanced tools such as + +digital IDs [34, 55] have been introduced, their applications involve a laborious process, i.e., attaching the devices to each animal individually. The process is costly and potentially stressful for the animals. Moreover, these physical tags can identify only pre-defined individuals, which makes them impractical for use in real world scenarios. + +In the human domain, digital face recognition is one of the effective approaches for the identification. It has been developed for use in smartphones, airport security, and systems for finding missing people. Therefore, the research community made great efforts to develop sophisticated deep learning-based face recognition models [15,26,33,40,63,65], empowered by large-scale datasets and benchmarks, e.g., [21,29,48,57]. + +Despite the promise of human face recognition, the research progress towards automatic animal face individual recognition has been impeded, primarily because of the lack of extensive datasets and benchmarks for animal face recognition. Previous openly available datasets mostly include less than 100 individuals [10, 19, 30, 47, 66], which makes it far from generalized and discriminative identification models and precise evaluation for unseen individuals. + +In this paper, we introduce a large-scale animal face recognition dataset called PetFace that contains 257,484 individuals in total across 13 species with 319 breeds with 1,012,934 images. We show the example images of our PetFace in Fig 1. The number of individuals in our dataset is over 110 times that in the previous largest animal face dataset [49]. We sourced images and related information from the internet, with automated and manual filtering processes applied to ensure the dataset is not only large but also finely detailed and clean. Moreover, PetFace has fine-grained annotations including sex, breeds, and colors and patterns of their skin, which allows further investigation for fine-grained recognition and evaluation. + +PetFace offers two benchmarks: one for recognizing known (seen) individuals and the other for recognizing unknown (unseen) ones. We also conduct the verification of the fine-grade breeds and unseen animal categories. + +Our main contributions are as follows: (i) We establish a new dataset for animal face recognition called PetFace, which contains a total of 257,484 individuals across 13 types of animal families and 319 breeds with fine-grade annotations including sex, breed, color of animals. (ii) We set the benchmarks on recognizing known (seen) individuals and unknown (unseen) ones. (iii) We show that the model trained using our dataset shows the generalization capabilities for unseen individuals and even for unseen animal categories. + +# 2 Related Work + +Building datasets and benchmarks is an important step in advancing animal re-identification through deep models. While earlier efforts have established the basics of animal face recognition, there is considerable potential for further development. Compared to the dataset for human face recognition [21, 29, 48, 57, 67], those for animal faces [10, 19, 49, 66] have much fewer individuals. The evaluation + +![](images/e278c278bd4ea741a38a64d8b494fa4b99446d2f4e496c78560e96235f07c450.jpg) +Fig. 1: Example images of our PetFace. We introduce a large-scale animal face re-identification dataset PetFace that include 257,484 unique individuals across 13 families and 319 breeds. From the left, the images represent Cat, Chimpanzee, Chinchilla, Degus, Dog, Ferret, Guinea pig, Hamster, Hedgehog, Parakeet, Java sparrow, Pig, and Rabbit. The four images enclosed within the white square's grid lines represent the same identity. + +scenarios represented in these datasets often fall short of real-world applicability; most of the datasets focus on closed-set re-identification, rather than recognizing unseen individuals that is a critical requirement for practical applications. Our work enables the training and evaluation on a large number of individuals across a wider range of animal families and breeds. + +Human Face Identification is the process of identifying an individual's identity using their unique facial characteristics. The huge demands for individual identification have grown mainly in the human domain. In the identification of human faces, which is known as face recognition, considerable efforts have been expended in the research community [15,26,33,40,63,65]. Because of their domain-agnostic frameworks, most of the state-of-the-art methods in face recognition can be exported into other domains, such as animal identification. Furthermore, these advancements have been supported by the introduction of large-scale datasets and benchmarks [21, 29, 48, 57, 67] that have played a crucial role in facilitating the exploration of powerful models. Motivated by this, we create the PetFace dataset to fill the gap between human face recognition and animal face recognition. + +Animal Identification is the process of identifying an individual's identity using their unique body or facial characteristics, which is an important task across various scenarios, including monitoring animals, conducting habitat surveys, and finding missing animals. With the advance of computer vision technology, various openly available datasets contribute to computer vision for animals [6-8, 12, 18, 20, 24, 31, 38, 46, 50, 59, 60, 64]. Various methods are used to create the dataset, such as recording images [12, 18, 24, 53, 60] and videos [22, 38, 46, 50, 64, 68], and using aerial image [6, 7, 20]. Recording individual information is labor intensive; therefore, most of the previous datasets contain only a limited number of individuals and are focused on one species. Recently, WildlifeDatasets [61], which + +
DatasetFamily#Individuals#ImagesFine-grained annotation
CTai [19]Chimp785078Age, Sex
CZoo [19]Chimp242109Age, Sex
MacaqueFaces [66]Monkey346280-
BristolGorillas2020 [10]Gorilla75428-
DogFaceNet [49]Dog13938363-
Flickr-dog [47]Dog42374-
THODBRL2015 [30]Horse472820Angle
ZindiTurtleRecall [4]Turtle226512803Angle
SeaTurtleID2022 [5]Turtle4388729Timestamp
Ours13 families257,4841,012,934Sex, Breed, Color & Pattern
+ +Table 1: Openly available animal face identification datasets. Compared to other datasets, PetFace has the largest number of individuals, and a wider range of families and breeds. + +gathers previously openly available datasets [1-4,6-9,11,12,16,18-20,22,24,27, 28,36,38,46,50-53,60,62,64,66,68,69] combined, was introduced. This research creates the benchmarks using existing available datasets. + +The advances in human face recognition raise an interest in animal face recognition. One notable challenge is the variation in facial structures between animals and humans. This has led to studies like AnimalWeb [32] and CatFLW [44] that propose specialized methods for animal facial key points detection. Automatic animal face identification has been studied, which can help humans to monitor animals [25, 39, 42, 43, 58]. We review several publicly available face identification datasets in Table 1. Apes, such as Chimpanzees [19] and Gorillas [10], are one of the species that have been studied. However, as the dataset size is relatively limited, the datasets are insufficient to evaluate new unseen individual faces. In addition to primates, research on facial (head) identification has also extended to turtles, with the ZindiTurtleRecall [4] dataset offering a substantial number of images of individual turtles in a controlled environment. To address the limitations of in-the-wild data collection, the SeaTurtleID2022 [5] dataset includes wild data, albeit with a reduced number of individuals, owing to the extensive effort required for data collection. Among datasets containing mammals, the DogFaceNet dataset [49] featuring 1,393 individual dogs stands out for its size. Nevertheless, the significant variation in appearance across dog breeds suggests that even this larger dataset may not be sufficiently comprehensive. In addition to the limited number of individuals, previous animal face datasets often focus on only single animal families. They can contribute to specific animal family research but can not explore animal identification across many families and breeds. + +# 3 PetFace Dataset + +PetFace is a large animal face identification dataset that expands research on animal face recognition, which has been impeded by a scarcity of suitable datasets + +and benchmarks. This section details the construction of the PetFace dataset, including our labor-efficient methods for collecting animal face images and a semi-automated filtering process to ensure quality fine-grained categorization and statistics of the dataset. + +# 3.1 Dataset Statistics + +![](images/0f27eb20c8632dfa7797be3242c655f47e0a7052738fd419adc8a929646ecaae.jpg) +(a) Individual distribution + +![](images/1a0ecff7a365b6e7cc328c4333ae9d327b94127160c2ae3f53ec2fb55a39d122.jpg) +(b) Sex distribution + +![](images/f9693d6e47880e64254c39c50e81fe1f6d914d08f00356bc1442edddf64bf47d.jpg) +(c) Breed annotation examples + +![](images/49c5cf42f668e584119b5044185027d61611b7fe37ba5d2c479a6baac7bba7a5.jpg) + +![](images/bc92ab628140b13619e1be092caa67ee40660ca78e28dd92c19d9ee71e912ca9.jpg) + +![](images/40cb436f1e5bab23e3850bf21b40418f4b39fdd502fc000d83295300609535f6.jpg) +Fig. 2: Dataset distribution. Each figure represents (a) the number of individuals in each animal family. (b) the sex distribution percentage by animal family. (c) examples of breed annotations. (d) examples of color annotations. + +![](images/1d2a7a5b2af52220fb7036713f277dd2e7db3a3e525510caa810ff8e8e0932ad.jpg) +(d) Color annotation examples + +![](images/a57b023c7b588e33d9f61b895fd0c949f60a2c49f419acf2a24b0e289e81784f.jpg) + +The PetFace dataset encompasses 1,012,934 images spanning 257,484 unique individuals. Detailed distributions of animal families are illustrated in Fig 2(a). Images are cropped to $224 \times 224$ pixels around their faces. Our PetFace also has fine-grained annotations. Sex distribution across different animal families is depicted in Fig 2(b), with sex information available for 240,861 individuals, accounting for $94\%$ of the dataset. The dataset includes annotations for 319 breeds, with examples of breed annotations shown in Fig 2(c). Furthermore, as detailed in Fig 2(d), the dataset provides in-depth color information through two-tier hierarchical annotations. Please see the supplemental for the detailed information per each animal family. + +# 3.2 Data Sourcing + +Collecting images of animal faces via photography is labor and time intensive, which impedes the creation of large-scale datasets. In contrast, the human face recognition field benefits greatly from the availability of images sourced from the Internet. In this section, we outline our approach to assembling the collection of animal face images through the Internet, where each is associated with unique individual identifiers. + +Curation of images. In contrast to the human domain, acquiring multiple images for individual animals is more challenging. Unlike human datasets where a large number of celebrity images can be readily sourced, animal images require alternative approaches for curation. We utilized two primary sources: (i) pet shops' websites and (ii) animal adoption websites. The advantage of pet shops lies in their provision of high-quality, diverse images, capturing individuals from various angles and offering detailed information about each animal, including color, sex, and specific breed details. On the other hand, animal adoption sites offer images set against a variety of backgrounds and conditions that are often provided by pet owners, thus ensuring each individual is presented in a unique setting. These sources provide images that are highly suitable for animal recognition tasks and that are especially useful for recognizing animals in varied wild environments. For Chimps, we additionally use images from the webpage of a collaborative research institution. + +To ensure the dataset's quality, we were selective in choosing websites for curation. We only use the websites introducing each animal on one page to gain the individual IDs. Aware of the potential for pet owners to upload the same images to multiple websites, we chose a single pet adoption website from each region (e.g., one per country) to minimize duplicates. For pet shops, we confirmed that the animals listed were unique to a certain shop to ensure the uniqueness of our dataset entries. Using these sources, we collected 1,443,737 images from 325,420 individuals. + +# 3.3 Face Alignment and Filtering + +Face Detection. We detect facial landmarks to align and crop the images. We adopt the AnyFace [35] that is trained on mixed face datasets including + +![](images/ffc1f5b9fd830e0de23a28d0bc1c7af993bade2c59de8b24208e9d8c074a9703.jpg) +Fig. 3: Data filtering process. We adopt a two-stage data filtering approach. Initially, images containing multiple faces are automatically removed. Subsequently, any images that do not depict animal faces or fail to meet alignment criteria are manually eliminated. + +AnimalWeb [32]. Because the positions of animal facial parts are sometimes very different from those of humans, we use different reference points for different species in face alignment. After landmark detection, we select one frontal image as a reference for each species. We compute the average landmarks over all the landmarks aligned with the reference. Then, we define the target positions of the landmarks in the images; all the images are aligned to the target positions. + +Data Filtering. Fig 3 shows the overview of our data filtering process. Because the fully automatic face detection above sometimes fails, we adopt a two-stage data filtering process to filter out the following cases: 1) Images contain multiple animals simultaneously. 2) Images are unrelated to animals, such as advertisements. 3) Images focus on non-animal elements, such as random patterns in backgrounds, toys, or people, rather than the animals themselves. First, we automatically remove images where multiple faces are detected in an image. Secondly, annotators manually assess all the images and remove those that do not depict the target animals or where the face alignment differs from our intended criteria. Because checking whether the faces are properly aligned or not requires some expertise, all the images are filtered by the authors to ensure the quality of the dataset. This manual filtering process took about 100 man hours. After this stage, we keep 1,012,934 images, which is approximately $70\%$ of the initial image number. Detailed distributions of animal families are illustrated in Fig 2(a). + +# 3.4 Fine-Grained Annotations + +Fine-grained categorization enhances its usefulness for downstream applications, e.g., the creation of challenging datasets based on animal's individual attributes. However, collecting extensive individual animal data poses a major challenge in assembling large datasetse owing to the significant effort involved. In response, + +we collect as much individual animal information corresponding to the images as possible. Our annotations include sex, color, and specific breed details, depending on their availability on the source websites. The annotation process involves: 1) extracting individual information displayed on websites alongside images, 2) refining raw text data to develop individual data tables, 3) manually verifying that all attributes are refined and free of unrelated information, and 4) manually annotating individuals with missing attributes on the website. + +As a result, we include a sex category for all classes, with breed annotations for eight animal classes (Cat, Chimp, Dog, Guinea pig, Hamster, Parakeet, Pig, and Rabbit), and color annotations for eleven animal classes (Cat, Chinchilla, Degus, Ferret, Guinea pig, Hamster, Hedgehog, Parakeet, Java sparrow, Pig, and Rabbit). Note that the manual annotations are performed only on colors and patterns because images alone do not allow us to accurately determine an animal's sex or breed. Detailed distributions of sex per animal families are illustrated in Fig 2(b). In Fig 2(c) and (d), we illustrate the examples of breed annotations and color and pattern annotations. + +# 4 Experimental Methodology + +We introduce two principal evaluation protocols: 1) re-identification for seen faces, and 2) verification for unseen faces. To establish these benchmarks, we have curated two distinct types of test sets. Specifically, for unseen face verification, we carefully separate the dataset into training, validation, and test sets to prevent data leakage that can arise from shared backgrounds. + +# 4.1 Evaluation Protocols + +We adopt two types of evaluation. + +1. Re-identification: This evaluation selects images from the test data that match the identities present in the training data. This procedure involves identifying test images that correspond to the same identities used during model training. The objective is to assess the model's ability to accurately recognize and associate test images with the correct identities from the training dataset. +2. Verification: This evaluation checks if the model can identify unseen faces. This means that we use different identities during the training and testing phases. For each identity, we select one image that matches the identity and one image randomly chosen from a different identity. This process creates pairs of images where one pair consists of images from the same identity, and another pair consists of images from different identities. We then task the model with predicting whether the faces in each pair belong to the same identity or different identities. This approach allows us to comprehensively evaluate the model's ability to recognize and differentiate between individual identities. + +# 4.2 Data Split + +To establish benchmarks for seen individuals re-identification and unseen individuals verification, we use specific split protocols, creating two test sets designed for each task. For seen individuals re-identification, the process involves verifying images of the same identities used in training; thus, we select test images from the same individuals for the training set. We, therefore, ensure that each individual is represented by at least three photos i.e., a minimum of two for training and one for testing. For unseen individuals verification dataset, we ensure that it contains no images from sources that are also represented in the training or validation sets per animal family to prevent bias from similar environmental conditions. We divide the dataset into training, validation, and each testing sets following a 7:1:2 ratio as closely as possible, given the outlined criteria. + +# 4.3 Models + +We train state-of-the-art models based on deep neural networks on our PetFace to build the benchmarks. For the re-identification task, we approach the training and testing phases as classification problems; an individual is assigned to a class. On the other hand, for the verification task, we train models in the same manner as for the re-identification task but evaluate the models by computing the cosine similarity between pairs of images to be identified if they are the same individual. Specifically, we focus more on the loss functions that are crucial for identification tasks than network architectures. We refer to three important loss functions in addition to the basic Softmax-based classification model: + +Triplet loss [26] is designed to take a triplet of samples, i.e., an anchor $x_{a}$ , a positive $x_{p}$ (another image of the same identity as the anchor), and a negative $x_{n}$ (an image of a different identity) - and learn embeddings in such a way that the $x_{p}$ is closer to the $x_{a}$ than the $x_{n}$ by a margin. This loss function is particularly beneficial for face Re-ID as it directly targets the relative distances between different and same identity pairs, encouraging the model to learn a feature space where embeddings of the same identity are clustered together while being far from other identity clusters. + +Center loss [65] works alongside Softmax Loss to enhance the discriminative power of the learned features. While Softmax Loss focuses on inter-class separability, Center Loss aims to minimize the intra-class variations. It does this by penalizing the distance between the deep features of each class and their corresponding class center. Center Loss ensures that the embeddings of the faces of the same individual are closer together, thus making the feature distribution more compact for each identity. + +ArcFace loss [15] introduces an angular margin between classes to enforce a discriminative feature space. It modifies the Softmax loss by adding a margin penalty to the angle between the feature vector and the corresponding class center in the angular space. This angular margin encourages models to learn more distinguishable embeddings to separate classes. + +We use ResNet-50 [23], which we found performs better than recent Transformer-based models in Table 5, as our base backbone for all applied loss functions to simplify our experiments and make them easier to grasp. + +In addition to the comparison on the models trained on PetFace, we evaluate state-of-the-art models trained on other datasets including: + +ImageNet [14] is a conventional image classification dataset including 1000 general object classes. We use the ResNet-50 trained on the dataset. + +CLIP [54] learns semantic relationships between images and texts in a cross-modal contrastive learning manner empowered by web-scale image-caption datasets. We use ResNet-50 backbone. + +MegaDescriptor [61] is a state-of-the-art model for animal identification trained on a unified animal identification dataset including 33 existing animal re-identification datasets [1-4, 6-9, 11, 12, 16, 18-20, 22, 24, 27, 28, 36, 38, 46, 50-53, 60, 62, 64, 66, 68, 69]. We use the officially provided SwinTransformer-B [41] backbone. + +# 5 Experimental Result + +# 5.1 Benchmark on Animal Face Re-identification + +We show the re-identification results in terms AUC in Table 2. We train models with the baseline loss functions independently on each class. We can see that ArcFace performs consistent results, $51.23\%$ of average accuracy, compared to the other loss functions (41.88% for Softmax and 9.81% for Center). On the other hand, Center loss does not learn sufficient discriminative features for animal face re-identification, especially on Cat and Dog where the number of the test classes are 113,592 and 46,755, respectively. These results encourage the community to explore more effective representation learning methods for re-identification tasks. + +Moreover, motivated by MegaDescriptor, we train an ArcFace model on the entire PetFace, denoted as Joint-Trained on PetFace in the table. We observe that the joint-training strategy provides some improvements, e.g., from $54.29\%$ to $70.30\%$ for Cat and from $29.08\%$ to $41.49\%$ for Pig although in some classes the results get worse, e.g., from $43.27\%$ to $34.30\%$ for Chimp and from $62.19\%$ to $44.78\%$ to Parakeet. This result indicates that there is some room to improve the performance of integrated identification on imbalance and wide-range datasets. + +In Fig 4, we display the Top- $k$ ( $k = 1,3,5$ ) accuracy metrics. While accuracy naturally increases with larger $k$ values, the relative ranking of accuracy across different animal families remains largely consistent. ArcFace maintains the best performance across these evaluations. + +# 5.2 Benchmark on Animal Face Verification + +Next, we show the verification results in Table 3. We use the trained models in Table 2 and evaluate them on unseen individuals. Similar to our findings in the re-identification task, ArcFace proved to be effective, achieving the best + +
MethodTop-1 Accuracy (%)Avg
CatChimpChinchillaDegusDogFerretGuineaHamsterHedgehogParakeetJava sparrowPigRabbit
Trained on PetFace
Softmax30.4641.7058.1337.8459.1425.8860.0738.2727.8150.8833.5521.6459.0541.88
Center0.005.3829.7612.030.007.9131.779.4613.761.684.867.443.519.81
Arcface54.2943.2767.3445.4177.8628.9267.9047.3730.9062.1942.2729.0869.1351.23
Joint-Trained on PetFace
ArcFace70.3034.3069.8656.0868.7546.1268.6654.3344.3844.7834.5541.4965.7553.80
+ +Table 2: Result on animal re-identification. ArcFace outperforms the other loss functions. The jointly trained model on our PetFace shows the best average top-1 accuracy. + +
MethodAUC (%)Avg
CatChimpChinchillaDegusDogFerretGuinea pigHamsterHedgehogParakeetJava sparrowPigRabbit
Trained on PetFace
Softmax97.9785.2284.4486.1998.9883.1595.3087.7387.2798.3984.6787.0998.5390.38
Center50.2081.6080.9681.3459.6478.7792.1083.5985.2696.9180.5482.5897.0180.81
Triplet96.9477.1076.1272.8097.9779.4883.3780.0881.4791.5677.8675.7294.7483.48
Arcface97.7183.7687.7087.6199.4586.2496.0390.3286.1398.6384.4589.8899.0191.30
Joint-Trained on PetFace
ArcFace98.0483.2089.6386.7999.0188.4494.1492.5087.8698.4089.3892.3898.4592.17
Trained on Other Dataset
ImageNet81.5873.6770.8070.6997.1867.9585.1974.1375.0190.4579.4982.4286.2379.60
CLIP85.1070.7073.4274.6691.8670.8980.6773.9177.6583.1976.5280.0283.5578.63
MegaDescriptor87.3083.0177.8577.7993.7576.5388.3677.9178.7489.7678.6488.0890.3883.70
+ +Table 3: Result on animal verification. Jointly trained on our PetFace shows the best performance. We also compared to the trained models on other dataset. + +results with an average AUC of $92.17\%$ when jointly trained on the PetFace dataset. Similar to face re-identification in Sec 5.1, in some classes, the results get worse when jointly trained across animal families e.g., Chimp, Degus, Dog, Guinea pig, Parakeet, and Rabbit. We also examined models pre-trained on other datasets including ImageNet, CLIP and MegaDescriptor in order to make more comparisons. Among these models, MegaDescriptor, which trained on the animal re-identification dataset, showed the highest AUC of $83.70\%$ . We show the similarity distribution on the our cat dataset in Fig 5. Comparing the performance of CLIP, MegaDescriptor, and our joint-trained models, our findings indicate that our model achieves the most distinct separation between positive and negative samples. In contrast, the MegaDescriptor model displays a closer distribution of these samples, and with CLIP, the overlap between positive and negative distributions is significantly more pronounced. + +# 5.3 Comparison with Previous Datasets + +We conduct cross-dataset evaluations where models are tested on different datasets other than the training datasets to demonstrate the generality of our dataset. Here, we compare our dataset with 1) CTai [19] and CZoo [19] for Chimpanzee and 2) DogFaceNet [49] and Flickr-dog [47] for Dog. All images of the compared datasets are aligned in the same manner as ours for fair comparisons. We split the identities of the compared datasets into training and test sets in a ratio of 7:3 + +![](images/44020877bc6b8dd93cd55b2f47107c5f99c014fca4ac624e1d4666e65beab0b7.jpg) +(a) Top-1 + +![](images/e5f826bc121e1fe340ddeb5183888139cd1573fe3a93abf6e57e39ba39db0628.jpg) +(b) Top-3 + +![](images/a49caaff41a088cdbda4f8dfc47c5565a5c2d9ba5b00fb28d5430cceb1d6f843.jpg) +(c) Top-5 + +![](images/2a1279156c9c47525e6c7ed96d217e4d0b0ee20df2164fe74b7b171889496cbd.jpg) +Fig. 4: Top-k accuracy ( $k = 1,3,5$ ). ArcFace consistently shows the best performance consistently. + +![](images/b4c81e03b3539375dffe769f4499a77e3fdaf41b6ea5e8e1cc9a14c9298569dd.jpg) +Fig. 5: Similarity distributions of (a) CLIP, (b) MegaDescriptor, and (c) our joint-trained model on cats. The horizontal axis represents similarity and the vertical axis represents frequency. The model trained on PetFace shows the most distinct separation between positive and negative samples. + +![](images/69aff722169cfbce354425ea5aa91d16cbde62671cd78d45e99f89e52a0c2ca3.jpg) + +and then uniformly pick the positive and negative pairs. Because CTai, CZoo, and Flickr-dog have a few test identities, we each pick five positive and negative pairs for each test identity for robust evaluations. + +We show the verification results in Tables 4(a) and (b). We exclude the results trained on the same dataset as the test set denoted in gray for fair comparisons. For the Chimpanzee datasets, the model trained on our dataset outperforms previous datasets. For case tested on CTai, our dataset outperforms CZoo (67.33% vs. 66.76%). For the other case tested on CZoo, our dataset also surpasses CTai (71.27% vs. 69.31%). This result demonstrates that our Chimpanzee dataset is more effective and general than the previous datasets. For the Dog datasets, owing to the significant scale of our dataset, the model trained on our dog dataset outperforms not only cross-dataset results but also in-dataset ones denoted in gray. These results support the quality of our PetFace dataset. + +# 5.4 Analysis + +Different networks. To examine the impact of different network architectures on the performance on PetFace, we additionally train two backbones, + +
Train SetTest Set AUC (%)
Database#Ind.#ImagesCTaiCZoo
CTai49285775.8069.31
CZoo17132366.7680.49
Ours-Chimp446267967.3371.27
+ +(a) Chimp datasets. + +
Train SetTest Set AUC (%)
Database#Ind.#ImagesDogFaceNetFlickr-dog
DogFaceNet975587998.7285.51
Flickr-dog3026594.9887.77
Ours-Dog4675516834899.6595.54
+ +(b) Dog datasets. +Table 4: Comparisons with previous datasets for Chimp and Dog. The results when the test sets are the same as the training datasets are excluded for fair comparisons. The models trained on our dataset outperform those trained on previous datasets. + +
MethodAUC (%)Avg
CatChimpChinchillaDegusDogFerretGuinea pigHamsterHedgehogParakeetJava sparrowPigRabbit
ResNet-5097.7183.7687.7087.6199.4586.2496.0390.3286.1398.6384.4589.8899.0191.30
ViT-32-B94.4478.6071.1382.6993.3968.1783.2177.8987.7795.0780.6275.3195.1883.35
Swin-B87.3574.2376.2184.0485.7571.6091.5688.3386.2294.1284.6666.5690.5683.17
+ +Table 5: Verification result on different backbones. ResNet-50 achieves the best average AUC. + +VisionTransformer-32-B [17] and SwinTransformer-B [41], with the same training configuration as our base model (ResNet-50) and report the verification results in Table 5. We observe that ResNet-50 outperforms the additional transformer-based models in most cases and achieves the best average AUC of $91.30\%$ . + +Fine-grained verification. In real world scenarios, animal verification is sometimes conducted within a specific breed rather than different breeds. To meet this demand, we introduce fine-grained verification where positive and negative pairs are chosen only from the same breeds. We refer to the top-10 dog breeds of individuals in our datasets: Chihuahua, Dachshund, French bulldog, Golden retriever, Miniature dachshund, Pomeranian, Shiba inu, Shih tzu, Toy poodle, and Yorkshire terrier. We construct the fine-grained test sets and evaluate the models on each breed. The results are visualized in Fig. 6. It can be seen that our model achieves a good performance on all the breeds (98.30% on average). The results achieved for the baseline models are lower in comparison (84.99%, 80.78%, and 86.25% for ImageNet, CLIP, and MegaDescriptor, respectively). Moreover, we find that the results achieved with MegaDescriptor show significantly more degradation than the cross-breed verification result in Table 3 from 93.75% to 86.25%. while our model maintains the results at a high-level (from 99.01% to 98.39%). This result indicates that our dataset overcomes the limitation of the previous dog dataset [69] that suffers from the poor performance for specific breed verification owing to its insufficient number of individuals (192 individuals). + +Generality to Unseen Families. Lastly, we evaluate the verification generality to unseen animal families. To achieve this, we additionally collect 100 identities each of Parrot, Lacertilia, and Squirrel, which are not contained in either our + +![](images/195502cd50d0b78702522deedc82f87355ae467472e0daed69ce4edba877a1d2.jpg) +Fig.7: Generality to unseen animal families. Trained on our PetFace achieves the best performance even for unseen animal families. + +![](images/412009a568fb4dc11aaa8acf98199e26604771e45bd9ebc126e625874b61dd96.jpg) +Fig. 6: Fine-grained dog verification. Trained on our PetFace Dog achieves the best performance consistently. + +dataset or WildLifeDataset [61]. We compare our model that is jointly trained on our dataset across families with the baseline models in Fig. 7. Our model outperforms the baselines on all the species. For Parrot, particularly, our model achieves an AUC of $88.99\%$ , revealing its superiority over MegaDescritor with a large margin. This result supports the generality and variety of our PetFace dataset. + +# 6 Conclusion + +We introduced the PetFace, a comprehensive animal identification dataset encompassing 13 families, featuring 257,484 unique individuals across 319 breed categories. We have collected images and their information and conducted automated and manual filtering to ensure the quality of the large-scale dataset. This dataset also includes detailed annotations of sex, breed, colors, and patterns to facilitate more investigation and application in a real-world scenario. We establish two main benchmarks: 1) re-identification of seen individuals and 2) verification of unseen individuals. + +Our experiments show the generality of the models trained on our dataset to verification on other datasets or unseen animal families. We also found that there is still room for improvement in the integrated identification of multiple animal families. To promote further research, we will make this dataset, experiment code, and models available to the research communities. This dataset will enable computer vision researchers to tackle animal face identification across a wide range of breeds and push forward progress on animal face identification tasks. + +# Acknowledgements + +This work was partially supported by the Cooperative Research Program of the Primate Research Institute, Kyoto University. We thank Dr. Takuma Yagi from AIST for comments on paper proofreading. We also thank Prof. Toshihiko Yamasaki from the University of Tokyo for providing computation resources. + +# References + +1. Beluga id 2022 (2022), https://lila.science/datasets/beluga-id-2022/ +2. Hyena id 2022 botswana predator conservation trust (2022). panthera pardus csv custom export. (2022), https://lila.science/datasets/hyena-id-2022/ +3. Leopard id 2022 botswana predator conservation trust (2022). panthera pardus csv custom export. (2022), https://lila.science/datasets/leopard-id-2022/ +4. Turtle recall: Conservation challenge (2022), https://zindi.africa/competitions/turtle-recall-conservation-challenge +5. Adam, L., Čermák, V., Papafitsoros, K., Picek, L.: Seaturtleid2022: A long-span dataset for reliable sea turtle re-identification. In: IEEE/CVF Winter Conference on Applications of Computer Vision (WACV). pp. 7146-7156 (2024) +6. Andrew, W., Gao, J., Mullan, S., Campbell, N., Dowsey, A.W., Burghardt, T.: Visual identification of individual holstein-friesian cattle via deep metric learning. In: Computers and Electronics in Agriculture. vol. 185, p. 106133 (2021) +7. Andrew, W., Greatwood, C., Burghardt, T.: Visual localisation and individual identification of holstein friesian cattle via deep learning. In: IEEE/CVF International Conference on Computer Vision (ICCV) Workshops. pp. 2850-2859 (2017) +8. Andrew, W., Greatwood, C., Burghardt, T.: Visual localisation and individual identification of holstein friesian cattle via deep learning. In: IEEE/CVF International Conference on Computer Vision (ICCV) Workshops. pp. 2850-2859 (2017) +9. Andrew, W., Hannuna, S., Campbell, N., Burghardt, T.: Automatic individual holstein friesian cattle identification via selective local coat pattern matching in rgb-d imagery. In: International Conference on Image Processing (ICIP). pp. 484-488 (2016) +10. Brookes, O., Burghardt, T.: A dataset and application for facial recognition of individual gorillas in zoo environments. In: International Conference on Pattern Recognition(ICPR) (2020) +11. Cheeseman, T., Southerland, K., Park, J., Olio, M., Flynn, K., Calambokidis, J., Jones, L., Garrigue, C., Frisch-Jordán, A., Howard, A., Reade, W., Neilson, J., Gabriele, C., Clapham, P.: Advanced image recognition: a fully automated, high-accuracy photo-identification matching system for humpback whales. vol. 102, p. 915–929 (2021) +12. Christin B. Khan, Shashank, W.K.: Right whale recognition (2015), https://kaggle.com/competitions/noaa-right-whale-recognition +13. Dahlborn, K., Bugnon, P., Nevalainen, T., Raspa, M., Verbost, P.M., Spangenberg, E.M.F.: Report of the federation of european laboratory animal science associations working group on animal identification. vol. 47, pp. 2-11 (2013) +14. Deng, J., Dong, W., Socher, R., Li, L.J., Li, K., Fei-Fei, L.: Imagenet: A large-scale hierarchical image database. In: IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR). pp. 248-255 (2009) + +15. Deng, J., Guo, J., Xue, N., Zafeiriou, S.: Arcface: Additive angular margin loss for deep face recognition. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR). pp. 4690-4699 (2019) +16. Dlamini, N., Zyl, T.L.v.: Automated identification of individuals in wildlife population using siamese neural networks. In: 2020 7th International Conference on Soft Computing Machine Intelligence (ISCMI). pp. 224-228 (2020) +17. Dosovitskiy, A., Beyer, L., Kolesnikov, A., Weissenborn, D., Zhai, X., Unterthiner, T., Dehghani, M., Minderer, M., Heigold, G., Gelly, S., Uszkoreit, J., Houlsby, N.: An image is worth 16x16 words: Transformers for image recognition at scale. The International Conference on Learning Representations (ICLR) (2021) +18. Ferreira, A.C., Silva, L.R., Renna, F., Brandl, H.B., Renoult, J., Farine, D.R., Covas, R., Doutrelant, C.: Deep learning-based methods for individual recognition in small birds. In: Methods in Ecology and Evolution. vol. 11, pp. 1072-1085 (2020) +19. Freytag, A., Rodner, E., Simon, M., Loos, A., Kuhl, H.S., Denzler, J.: Chimpanzee faces in the wild: Log-euclidean cnns for predicting identities and attributes of primates. In: Rosenhahn, B., Andres, B. (eds.) Pattern Recognition. pp. 51-63. Springer International Publishing, Cham (2016) +20. Gao, J., Burghardt, T., Andrew, W., Dowsey, A.W., Campbell, N.W.: Towards self-supervision for video identification of individual holstein-friesian cattle: The cows2021 dataset. arXiv preprint arXiv:2105.01938 (2021) +21. Guo, Y., Zhang, L., Hu, Y., He, X., Gao, J.: Ms-celeb-1m: A dataset and benchmark for large-scale face recognition. In: Leibe, B., Matas, J., Sebe, N., Welling, M. (eds.) European Conference on Computer Vision (ECCV). pp. 87-102. Springer International Publishing, Cham (2016) +22. Haurum, J.B., Karpova, A., Pedersen, M., Bengtson, S.H., Moeslund, T.B.: Identification of zebrafish using metric learning. In: IEEE Winter Conference on Applications of Computer Vision (WACV) Workshops. pp. 1-11 (2020) +23. He, K., Zhang, X., Ren, S., Sun, J.: Deep residual learning for image recognition. In: IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR). pp. 770-778 (2016) +24. He, Z., Qian, J., Yan, D., Wang, C., Xin, Y.: Animal re-identification algorithm for posture diversity. In: ICASSP 2023 - 2023 IEEE International Conference on Acoustics, Speech and Signal Processing (ICASSP). pp. 1-5 (2023) +25. Hitelman, A., Edan, Y., Godo, A., Berenstein, R., Lepar, J., Halachmi, I.: Biometric identification of sheep via a machine-vision system. vol. 194, p. 106713 (2022) +26. Hoffer, E., Ailon, N.: Deep metric learning using triplet network. In: arXiv (2018) +27. Holmberg, J., Norman, B., Arzoumanian, Z.: Estimating population size, structure, and residency time for whale sharks rhincodon typus through collaborative photo-identification. vol. 7, pp. 39 - 53. Inter-Research Science Center (2009) +28. Howard, A., inversion, Southerland, K., Cheeseman, T.: Humpback whale identification challenge (2018), https://kaggle.com/competitions/whale-categorization-playground +29. Huang, G.B., Ramesh, M., Berg, T., Learned-Miller, E.: Labeled faces in the wild: A database for studying face recognition in unconstrained environments. No. 07-49 (2007) +30. Jarraya, I., Ouarda, W., Alimi, A.M.: A preliminary investigation on horses recognition using facial texture features. In: IEEE International Conference on Systems, Man, and Cybernetics. pp. 2803-2808 (2015) +31. Khan, F.F., Li, X., Temple, A.J., Elhoseiny, M.: Fishnet: A large-scale dataset and benchmark for fish recognition, detection, and functional trait prediction. In: Pro + +ceedings of the IEEE/CVF International Conference on Computer Vision (ICCV). pp. 20496-20506 (October 2023) +32. Khan, M.H., McDonagh, J., Khan, S., Shahabuddin, M., Arora, A., Khan, F.S., Shao, L., Tzimiropoulos, G.: Animalweb: A large-scale hierarchical dataset of annotated animal faces. In: IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR). pp. 6937-6946 (2020) +33. Kim, M., Jain, A.K., Liu, X.: Adaface: Quality adaptive margin for face recognition. pp. 18729-18738 (2022) +34. Krishnan, S., S., S.: Cloud iot based novel livestock monitoring and identification system using uid. In: Sensor Review. vol. 38, pp. 21-33 (2017) +35. Kuzdeuov, A., Koishigarina, D., Varol, H.A.: Anyface: A data-centric approach for input-agnostic face detection. In: IEEE International Conference on Big Data and Smart Computing (BigComp). pp. 211-218 (2023) +36. Lahiri, M., Tantipathananandh, C., Warungu, R., Rubenstein, D.I., Berger-Wolf, T.Y.: Biometric animal databases from field photographs: identification of individual zebra in the wild. In: Proceedings of the ACM International Conference on Multimedia Retrieval. Association for Computing Machinery (2011) +37. Leslie, E., Hernández-Jover, M., Newman, R., Holyoake, P.: Assessment of acute pain experienced by piglets from ear tagging, ear notching and intraperitoneal injectable transponders. In: Applied Animal Behaviour Science. vol. 127, pp. 86-95 (2010) +38. Li, S., Li, J., Tang, H., Qian, R., Lin, W.: Atrw: A benchmark for amur tiger re-identification in the wild. In: Proceedings of the 28th ACM International Conference on Multimedia. p. 2590–2598. MM '20, Association for Computing Machinery, New York, NY, USA (2020) +39. Li, X., Xiang, Y., Li, S.: Combining convolutional and vision transformer structures for sheep face recognition. In: Computers and Electronics in Agriculture. vol. 205, p. 107651 (2023) +40. Liu, W., Wen, Y., Yu, Z., Li, M., Raj, B., Song, L.: Spherface: Deep hypersphere embedding for face recognition. In: IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR). pp. 6738-6746 (2017) +41. Liu, Z., Lin, Y., Cao, Y., Hu, H., Wei, Y., Zhang, Z., Lin, S., Guo, B.: Swin transformer: Hierarchical vision transformer using shifted windows. In: IEEE/CVF International Conference on Computer Vision (ICCV). pp. 9992-10002 +42. Loos, A., Ernst, A.: An automated chimpanzee identification system using face detection and recognition. In: EURASIP Journal on Image and Video Processing. vol. 2013 (2013) +43. Marsot, M., Mei, J., Shan, X., Ye, L., Feng, P., Yan, X., Li, C., Zhao, Y.: An adaptive pig face recognition approach using convolutional neural networks. In: Computers and Electronics in Agriculture. vol. 173, p. 105386 (2020) +44. Martvel, G., Shimshoni, I., Zamansky, A.: Automated detection of cat facial landmarks. In: International Journal of Computer Vision (IJCV). pp. 1-16 (2023) +45. Mazlan, N.H., Lopez Salesansky, N., Burn, C., Wells, D.: Mouse identification methods and potential welfare issues: A survey of current practice in the uk. In: Animal Technology and Welfare. vol. 13, pp. 1-10 (2014) +46. Miele, V., Dussert, G., Spataro, B., Chamaillé-Jammes, S., Allainé, D., Bonenfant, C.: Revisiting animal photo-identification using deep metric learning and network analysis. In: Methods in Ecology and Evolution. vol. 12, pp. 863-873. Wiley (2021) +47. Moreira, T., Perez, M., Werneck, R., Valle, E.: Where is my puppy? retrieving lost dogs by facial features. In: Multimedia Tools and Applications. vol. 76, p. 15325-15340 (2017) + +48. Moschoglou, S., Papaioannou, A., Sagonas, C., Deng, J., Kotsia, I., Zafeiriou, S.: Agedb: The first manually collected, in-the-wild age database. In: IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR) Workshops. pp. 1997-2005 (2017) +49. Mougeot, G., Li, D., Jia, S.: A deep learning approach for dog face verification and recognition. In: Nayak, A.C., Sharma, A. (eds.) PRICAI 2019: Trends in Artificial Intelligence. pp. 418-430. Springer International Publishing, Cham (2019) +50. Murali, N., Taylor, G., Levine, J.: Can drosophila melanogaster tell who's who? In: PLOS ONE. vol. 13, p. e0205043 (2018) +51. Nepovinnykh, E., Eerola, T., Biard, V., Mutka, P., Niemi, M., Kunnasranta, M., Kalviainen, H.: Sealid: Saimaa ringed seal re-identification dataset. Sensors 22(19) (2022) +52. Papafitsoros, K., Adam, L., Čermák, V., Picek, L.: Seaturtleid: A novel long-span dataset highlighting the importance of timestamps in wildlife re-identification. In: arXiv (2023) +53. Parham, J., Crall, J., Stewart, C., Berger-Wolf, T., Rubenstein, D.: Animal population censusing at scale with citizen science and photographic identification. In: SS-17-01. pp. 37-44. AAAI Spring Symposium (2017) +54. Radford, A., Kim, J.W., Hallacy, C., Ramesh, A., Goh, G., Agarwal, S., Sastry, G., Askell, A., Mishkin, P., Clark, J., Krueger, G., Sutskever, I.: Learning transferable visual models from natural language supervision. In: arXiv (2021) +55. Roberts, C.: Radio frequency identification (rfid). In: Computers Security. vol. 25, pp. 18-26 (2006) +56. Roughan, J., Sevenoaks, T.: Welfare and scientific considerations of tattooing and ear-tagging for mouse identification. In: Journal of the American Association for Laboratory Animal Science: JAALAS. pp. 142-153 (2018) +57. Sengupta, S., Chen, J.C., Castillo, C., Patel, V.M., Chellappa, R., Jacobs, D.W.: Frontal to profile face verification in the wild. In: IEEE Winter Conference on Applications of Computer Vision (WACV). pp. 1-9 (2016) +58. Shi, X., Yang, C., Xia, X., Chai, X.: Deep cross-species feature learning for animal face recognition via residual interspecies equivariant network. In: European Conference on Computer Vision (ECCV). p. 667-682 (2020) +59. Shinoda, R., Shiohara, K.: Openanimaltracks: A dataset for animal track recognition (2024) +60. Trotter, C., Atkinson, G., Sharpe, M., Richardson, K., McGough, A.S., Wright, N., Burville, B., Berggren, P.: Ndd20: A large-scale few-shot dolphin dataset for coarse and fine-grained categorisation. In: arXiv (2020) +61. Čermák, V., Picek, L., Adam, L., Papafitsoros, K.: Wildlifedatasets: An open-source toolkit for animal re-identification. In: IEEE/CVF Winter Conference on Applications of Computer Vision (WACV). pp. 5953-5963 (2024) +62. Wahltinez, O., Wahltinez, S.: An open-source general purpose machine learning framework for individual animal re-identification using few-shot learning. In: Methods in Ecology and Evolution. vol. 15, pp. 373-387 (2024) +63. Wang, H., Wang, Y., Zhou, Z., Ji, X., Gong, D., Zhou, J., Li, Z., Liu, W.: Cosface: Large margin cosine loss for deep face recognition. In: IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR). pp. 5265-5274 (2018) +64. Wang, L., Ding, R., Zhai, Y., Zhang, Q., Tang, W., Zheng, N., Hua, G.: Giant panda identification. IEEE Transactions on Image Processing 30, 2837-2849 (2021) +65. Wen, Y., Zhang, K., Li, Z., Qiao, Y.: A discriminative feature learning approach for deep face recognition. In: Leibe, B., Matas, J., Sebe, N., Welling, M. (eds.) European Conference on Computer Vision (ECCV). pp. 499-515 (2016) + +66. Witham, C.L.: Automated face recognition of rhesus macaques. In: Journal of Neuroscience Methods. vol. 300, pp. 157-165 (2018) +67. Zhu, Z., Huang, G., Deng, J., Ye, Y., Huang, J., Chen, X., Zhu, J., Yang, T., Lu, J., Du, D., Zhou, J.: Webface260m: A benchmark unveiling the power of million-scale deep face recognition. In: IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR). pp. 10492-10502 (2021) +68. Zuerl, M., Dirauf, R., Koeferl, F., Steinlein, N., Sueskind, J., Zanca, D., Brehm, I., Fersen, L.v., Eskofier, B.: Polarbearvidid: A video-based re-identification benchmark dataset for polar bears. In: Animals. vol. 13 (2023) +69. Zuffi, S., Kanazawa, A., Berger-Wolf, T., Black, M.J.: Three-d safari: Learning to estimate zebra pose, shape, and texture from images "in the wild". In: IEEE/CVF International Conference on Computer Vision (ICCV). pp. 5358-5367 (2019) \ No newline at end of file diff --git a/2024/PetFace_ A Large-Scale Dataset and Benchmark for Animal Identification/images.zip b/2024/PetFace_ A Large-Scale Dataset and Benchmark for Animal Identification/images.zip new file mode 100644 index 0000000000000000000000000000000000000000..031a905379e4f55b6d0ac6a94656547fd4015c94 --- /dev/null +++ b/2024/PetFace_ A Large-Scale Dataset and Benchmark for Animal Identification/images.zip @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:64dd449b53d93eae067904026321f826cfbc872c1912e2f4f810b213cc2f1715 +size 586828 diff --git a/2024/PetFace_ A Large-Scale Dataset and Benchmark for Animal Identification/layout.json b/2024/PetFace_ A Large-Scale Dataset and Benchmark for Animal Identification/layout.json new file mode 100644 index 0000000000000000000000000000000000000000..0ec7f6020fd614c9fbfbc3ff421d890518be915c --- /dev/null +++ b/2024/PetFace_ A Large-Scale Dataset and Benchmark for Animal Identification/layout.json @@ -0,0 +1,8822 @@ +{ + "pdf_info": [ + { + "para_blocks": [ + { + "bbox": [ + 140, + 111, + 475, + 146 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 140, + 111, + 475, + 146 + ], + "spans": [ + { + "bbox": [ + 140, + 111, + 475, + 146 + ], + "type": "text", + "content": "PetFace: A Large-Scale Dataset and Benchmark for Animal Identification" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 211, + 168, + 402, + 180 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 211, + 168, + 402, + 180 + ], + "spans": [ + { + "bbox": [ + 211, + 168, + 402, + 180 + ], + "type": "text", + "content": "Risa Shinoda" + }, + { + "bbox": [ + 211, + 168, + 402, + 180 + ], + "type": "inline_equation", + "content": "^{1*}" + }, + { + "bbox": [ + 211, + 168, + 402, + 180 + ], + "type": "text", + "content": " and Kaede Shiohara" + }, + { + "bbox": [ + 211, + 168, + 402, + 180 + ], + "type": "inline_equation", + "content": "^{2*}" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 182, + 190, + 431, + 213 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 182, + 190, + 431, + 213 + ], + "spans": [ + { + "bbox": [ + 182, + 190, + 431, + 213 + ], + "type": "inline_equation", + "content": "^{*}" + }, + { + "bbox": [ + 182, + 190, + 431, + 213 + ], + "type": "text", + "content": "Equal contribution \n" + }, + { + "bbox": [ + 182, + 190, + 431, + 213 + ], + "type": "inline_equation", + "content": "^{1}" + }, + { + "bbox": [ + 182, + 190, + 431, + 213 + ], + "type": "text", + "content": "Kyoto University, Japan " + }, + { + "bbox": [ + 182, + 190, + 431, + 213 + ], + "type": "inline_equation", + "content": "^{2}" + }, + { + "bbox": [ + 182, + 190, + 431, + 213 + ], + "type": "text", + "content": "The University of Tokyo, Japan" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 160, + 244, + 452, + 475 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 160, + 244, + 452, + 475 + ], + "spans": [ + { + "bbox": [ + 160, + 244, + 452, + 475 + ], + "type": "text", + "content": "Abstract. Automated animal face identification plays a crucial role in the monitoring of behaviors, conducting of surveys, and finding of lost animals. Despite the advancements in human face identification, the lack of datasets and benchmarks in the animal domain has impeded progress. In this paper, we introduce the PetFace dataset, a comprehensive resource for animal face identification encompassing 257,484 unique individuals across 13 animal families and 319 breed categories, including both experimental and pet animals. This large-scale collection of individuals facilitates the investigation of unseen animal face verification, an area that has not been sufficiently explored in existing datasets due to the limited number of individuals. Moreover, PetFace also has fine-grained annotations such as sex, breed, color, and pattern. We provide multiple benchmarks including re-identification for seen individuals and verification for unseen individuals. The models trained on our dataset outperform those trained on prior datasets, even for detailed breed variations and unseen animal families. Our result also indicates that there is some room to improve the performance of integrated identification on multiple animal families. We hope the PetFace dataset will facilitate animal face identification and encourage the development of non-invasive animal automatic identification methods. Our dataset and code are available at https://dahlian00.github.io/PetFacePage/." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 160, + 486, + 402, + 497 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 160, + 486, + 402, + 497 + ], + "spans": [ + { + "bbox": [ + 160, + 486, + 402, + 497 + ], + "type": "text", + "content": "Keywords: Animals " + }, + { + "bbox": [ + 160, + 486, + 402, + 497 + ], + "type": "inline_equation", + "content": "\\cdot" + }, + { + "bbox": [ + 160, + 486, + 402, + 497 + ], + "type": "text", + "content": " Re-identification " + }, + { + "bbox": [ + 160, + 486, + 402, + 497 + ], + "type": "inline_equation", + "content": "\\cdot" + }, + { + "bbox": [ + 160, + 486, + 402, + 497 + ], + "type": "text", + "content": " Face Recognition" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 133, + 519, + 229, + 531 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 519, + 229, + 531 + ], + "spans": [ + { + "bbox": [ + 133, + 519, + 229, + 531 + ], + "type": "text", + "content": "1 Introduction" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 130, + 545, + 482, + 665 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 545, + 482, + 665 + ], + "spans": [ + { + "bbox": [ + 130, + 545, + 482, + 665 + ], + "type": "text", + "content": "Animal identification plays a crucial role in animal studies and applications such as monitoring animal behavior, conducting habitat surveys, locating missing animals, and performing health checks. Traditional identification techniques, including ear tags, tattoos, ear punching, and toe clipping, continue to be utilized mainly for experimental animals and livestock [13]. However, given that these methods have the potential to cause stress and pain [37,45,56], their use should be minimized to prioritize animal welfare. Therefore, there is a pressing need for the development and adoption of identification technologies that are not only effective and efficient but also minimally invasive, thereby mitigating the ethical concerns associated with traditional methods. While advanced tools such as" + } + ] + } + ], + "index": 6 + } + ], + "discarded_blocks": [], + "page_size": [ + 612, + 792 + ], + "page_idx": 0 + }, + { + "para_blocks": [ + { + "bbox": [ + 130, + 116, + 479, + 175 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 116, + 479, + 175 + ], + "spans": [ + { + "bbox": [ + 130, + 116, + 479, + 175 + ], + "type": "text", + "content": "digital IDs [34, 55] have been introduced, their applications involve a laborious process, i.e., attaching the devices to each animal individually. The process is costly and potentially stressful for the animals. Moreover, these physical tags can identify only pre-defined individuals, which makes them impractical for use in real world scenarios." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 130, + 176, + 480, + 247 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 176, + 480, + 247 + ], + "spans": [ + { + "bbox": [ + 130, + 176, + 480, + 247 + ], + "type": "text", + "content": "In the human domain, digital face recognition is one of the effective approaches for the identification. It has been developed for use in smartphones, airport security, and systems for finding missing people. Therefore, the research community made great efforts to develop sophisticated deep learning-based face recognition models [15,26,33,40,63,65], empowered by large-scale datasets and benchmarks, e.g., [21,29,48,57]." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 130, + 248, + 480, + 319 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 248, + 480, + 319 + ], + "spans": [ + { + "bbox": [ + 130, + 248, + 480, + 319 + ], + "type": "text", + "content": "Despite the promise of human face recognition, the research progress towards automatic animal face individual recognition has been impeded, primarily because of the lack of extensive datasets and benchmarks for animal face recognition. Previous openly available datasets mostly include less than 100 individuals [10, 19, 30, 47, 66], which makes it far from generalized and discriminative identification models and precise evaluation for unseen individuals." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 130, + 320, + 480, + 437 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 320, + 480, + 437 + ], + "spans": [ + { + "bbox": [ + 130, + 320, + 480, + 437 + ], + "type": "text", + "content": "In this paper, we introduce a large-scale animal face recognition dataset called PetFace that contains 257,484 individuals in total across 13 species with 319 breeds with 1,012,934 images. We show the example images of our PetFace in Fig 1. The number of individuals in our dataset is over 110 times that in the previous largest animal face dataset [49]. We sourced images and related information from the internet, with automated and manual filtering processes applied to ensure the dataset is not only large but also finely detailed and clean. Moreover, PetFace has fine-grained annotations including sex, breeds, and colors and patterns of their skin, which allows further investigation for fine-grained recognition and evaluation." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 130, + 440, + 480, + 475 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 440, + 480, + 475 + ], + "spans": [ + { + "bbox": [ + 130, + 440, + 480, + 475 + ], + "type": "text", + "content": "PetFace offers two benchmarks: one for recognizing known (seen) individuals and the other for recognizing unknown (unseen) ones. We also conduct the verification of the fine-grade breeds and unseen animal categories." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 130, + 476, + 480, + 559 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 476, + 480, + 559 + ], + "spans": [ + { + "bbox": [ + 130, + 476, + 480, + 559 + ], + "type": "text", + "content": "Our main contributions are as follows: (i) We establish a new dataset for animal face recognition called PetFace, which contains a total of 257,484 individuals across 13 types of animal families and 319 breeds with fine-grade annotations including sex, breed, color of animals. (ii) We set the benchmarks on recognizing known (seen) individuals and unknown (unseen) ones. (iii) We show that the model trained using our dataset shows the generalization capabilities for unseen individuals and even for unseen animal categories." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 132, + 578, + 237, + 592 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 578, + 237, + 592 + ], + "spans": [ + { + "bbox": [ + 132, + 578, + 237, + 592 + ], + "type": "text", + "content": "2 Related Work" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 130, + 605, + 480, + 665 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 605, + 480, + 665 + ], + "spans": [ + { + "bbox": [ + 130, + 605, + 480, + 665 + ], + "type": "text", + "content": "Building datasets and benchmarks is an important step in advancing animal re-identification through deep models. While earlier efforts have established the basics of animal face recognition, there is considerable potential for further development. Compared to the dataset for human face recognition [21, 29, 48, 57, 67], those for animal faces [10, 19, 49, 66] have much fewer individuals. The evaluation" + } + ] + } + ], + "index": 9 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "spans": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "type": "text", + "content": "2" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 166, + 91, + 282, + 101 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 166, + 91, + 282, + 101 + ], + "spans": [ + { + "bbox": [ + 166, + 91, + 282, + 101 + ], + "type": "text", + "content": "R. Shinoda and K. Shiohara" + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 1 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 138, + 116, + 468, + 252 + ], + "blocks": [ + { + "bbox": [ + 138, + 116, + 468, + 252 + ], + "lines": [ + { + "bbox": [ + 138, + 116, + 468, + 252 + ], + "spans": [ + { + "bbox": [ + 138, + 116, + 468, + 252 + ], + "type": "image", + "image_path": "e278c278bd4ea741a38a64d8b494fa4b99446d2f4e496c78560e96235f07c450.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 130, + 261, + 482, + 326 + ], + "lines": [ + { + "bbox": [ + 130, + 261, + 482, + 326 + ], + "spans": [ + { + "bbox": [ + 130, + 261, + 482, + 326 + ], + "type": "text", + "content": "Fig. 1: Example images of our PetFace. We introduce a large-scale animal face re-identification dataset PetFace that include 257,484 unique individuals across 13 families and 319 breeds. From the left, the images represent Cat, Chimpanzee, Chinchilla, Degus, Dog, Ferret, Guinea pig, Hamster, Hedgehog, Parakeet, Java sparrow, Pig, and Rabbit. The four images enclosed within the white square's grid lines represent the same identity." + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_caption" + } + ], + "index": 2 + }, + { + "bbox": [ + 130, + 342, + 482, + 400 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 342, + 482, + 400 + ], + "spans": [ + { + "bbox": [ + 130, + 342, + 482, + 400 + ], + "type": "text", + "content": "scenarios represented in these datasets often fall short of real-world applicability; most of the datasets focus on closed-set re-identification, rather than recognizing unseen individuals that is a critical requirement for practical applications. Our work enables the training and evaluation on a large number of individuals across a wider range of animal families and breeds." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 130, + 402, + 482, + 544 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 402, + 482, + 544 + ], + "spans": [ + { + "bbox": [ + 130, + 402, + 482, + 544 + ], + "type": "text", + "content": "Human Face Identification is the process of identifying an individual's identity using their unique facial characteristics. The huge demands for individual identification have grown mainly in the human domain. In the identification of human faces, which is known as face recognition, considerable efforts have been expended in the research community [15,26,33,40,63,65]. Because of their domain-agnostic frameworks, most of the state-of-the-art methods in face recognition can be exported into other domains, such as animal identification. Furthermore, these advancements have been supported by the introduction of large-scale datasets and benchmarks [21, 29, 48, 57, 67] that have played a crucial role in facilitating the exploration of powerful models. Motivated by this, we create the PetFace dataset to fill the gap between human face recognition and animal face recognition." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 130, + 545, + 482, + 666 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 545, + 482, + 666 + ], + "spans": [ + { + "bbox": [ + 130, + 545, + 482, + 666 + ], + "type": "text", + "content": "Animal Identification is the process of identifying an individual's identity using their unique body or facial characteristics, which is an important task across various scenarios, including monitoring animals, conducting habitat surveys, and finding missing animals. With the advance of computer vision technology, various openly available datasets contribute to computer vision for animals [6-8, 12, 18, 20, 24, 31, 38, 46, 50, 59, 60, 64]. Various methods are used to create the dataset, such as recording images [12, 18, 24, 53, 60] and videos [22, 38, 46, 50, 64, 68], and using aerial image [6, 7, 20]. Recording individual information is labor intensive; therefore, most of the previous datasets contain only a limited number of individuals and are focused on one species. Recently, WildlifeDatasets [61], which" + } + ] + } + ], + "index": 6 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 148, + 91, + 447, + 101 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 148, + 91, + 447, + 101 + ], + "spans": [ + { + "bbox": [ + 148, + 91, + 447, + 101 + ], + "type": "text", + "content": "PetFace: A Large-Scale Dataset and Benchmark for Animal Identification" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 474, + 92, + 480, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 474, + 92, + 480, + 100 + ], + "spans": [ + { + "bbox": [ + 474, + 92, + 480, + 100 + ], + "type": "text", + "content": "3" + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 2 + }, + { + "para_blocks": [ + { + "type": "table", + "bbox": [ + 132, + 114, + 491, + 247 + ], + "blocks": [ + { + "bbox": [ + 132, + 114, + 491, + 247 + ], + "lines": [ + { + "bbox": [ + 132, + 114, + 491, + 247 + ], + "spans": [ + { + "bbox": [ + 132, + 114, + 491, + 247 + ], + "type": "table", + "html": "
DatasetFamily#Individuals#ImagesFine-grained annotation
CTai [19]Chimp785078Age, Sex
CZoo [19]Chimp242109Age, Sex
MacaqueFaces [66]Monkey346280-
BristolGorillas2020 [10]Gorilla75428-
DogFaceNet [49]Dog13938363-
Flickr-dog [47]Dog42374-
THODBRL2015 [30]Horse472820Angle
ZindiTurtleRecall [4]Turtle226512803Angle
SeaTurtleID2022 [5]Turtle4388729Timestamp
Ours13 families257,4841,012,934Sex, Breed, Color & Pattern
", + "image_path": "c74f428db1b7be9cd10d348d0f8adc4fd18c86fecf66ffc04df04957401cc25d.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "table_body" + } + ], + "index": 2 + }, + { + "bbox": [ + 132, + 247, + 483, + 278 + ], + "lines": [ + { + "bbox": [ + 132, + 247, + 483, + 278 + ], + "spans": [ + { + "bbox": [ + 132, + 247, + 483, + 278 + ], + "type": "text", + "content": "Table 1: Openly available animal face identification datasets. Compared to other datasets, PetFace has the largest number of individuals, and a wider range of families and breeds." + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "text" + }, + { + "bbox": [ + 130, + 301, + 481, + 337 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 301, + 481, + 337 + ], + "spans": [ + { + "bbox": [ + 130, + 301, + 481, + 337 + ], + "type": "text", + "content": "gathers previously openly available datasets [1-4,6-9,11,12,16,18-20,22,24,27, 28,36,38,46,50-53,60,62,64,66,68,69] combined, was introduced. This research creates the benchmarks using existing available datasets." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 130, + 338, + 482, + 601 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 338, + 482, + 601 + ], + "spans": [ + { + "bbox": [ + 130, + 338, + 482, + 601 + ], + "type": "text", + "content": "The advances in human face recognition raise an interest in animal face recognition. One notable challenge is the variation in facial structures between animals and humans. This has led to studies like AnimalWeb [32] and CatFLW [44] that propose specialized methods for animal facial key points detection. Automatic animal face identification has been studied, which can help humans to monitor animals [25, 39, 42, 43, 58]. We review several publicly available face identification datasets in Table 1. Apes, such as Chimpanzees [19] and Gorillas [10], are one of the species that have been studied. However, as the dataset size is relatively limited, the datasets are insufficient to evaluate new unseen individual faces. In addition to primates, research on facial (head) identification has also extended to turtles, with the ZindiTurtleRecall [4] dataset offering a substantial number of images of individual turtles in a controlled environment. To address the limitations of in-the-wild data collection, the SeaTurtleID2022 [5] dataset includes wild data, albeit with a reduced number of individuals, owing to the extensive effort required for data collection. Among datasets containing mammals, the DogFaceNet dataset [49] featuring 1,393 individual dogs stands out for its size. Nevertheless, the significant variation in appearance across dog breeds suggests that even this larger dataset may not be sufficiently comprehensive. In addition to the limited number of individuals, previous animal face datasets often focus on only single animal families. They can contribute to specific animal family research but can not explore animal identification across many families and breeds." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 132, + 617, + 251, + 629 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 617, + 251, + 629 + ], + "spans": [ + { + "bbox": [ + 132, + 617, + 251, + 629 + ], + "type": "text", + "content": "3 PetFace Dataset" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 130, + 641, + 481, + 665 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 641, + 481, + 665 + ], + "spans": [ + { + "bbox": [ + 130, + 641, + 481, + 665 + ], + "type": "text", + "content": "PetFace is a large animal face identification dataset that expands research on animal face recognition, which has been impeded by a scarcity of suitable datasets" + } + ] + } + ], + "index": 7 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "spans": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "type": "text", + "content": "4" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 166, + 91, + 282, + 101 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 166, + 91, + 282, + 101 + ], + "spans": [ + { + "bbox": [ + 166, + 91, + 282, + 101 + ], + "type": "text", + "content": "R. Shinoda and K. Shiohara" + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 3 + }, + { + "para_blocks": [ + { + "bbox": [ + 130, + 116, + 482, + 163 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 116, + 482, + 163 + ], + "spans": [ + { + "bbox": [ + 130, + 116, + 482, + 163 + ], + "type": "text", + "content": "and benchmarks. This section details the construction of the PetFace dataset, including our labor-efficient methods for collecting animal face images and a semi-automated filtering process to ensure quality fine-grained categorization and statistics of the dataset." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 132, + 182, + 250, + 193 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 182, + 250, + 193 + ], + "spans": [ + { + "bbox": [ + 132, + 182, + 250, + 193 + ], + "type": "text", + "content": "3.1 Dataset Statistics" + } + ] + } + ], + "index": 3 + }, + { + "type": "image", + "bbox": [ + 136, + 224, + 310, + 335 + ], + "blocks": [ + { + "bbox": [ + 136, + 224, + 310, + 335 + ], + "lines": [ + { + "bbox": [ + 136, + 224, + 310, + 335 + ], + "spans": [ + { + "bbox": [ + 136, + 224, + 310, + 335 + ], + "type": "image", + "image_path": "0f27eb20c8632dfa7797be3242c655f47e0a7052738fd419adc8a929646ecaae.jpg" + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 186, + 337, + 287, + 347 + ], + "lines": [ + { + "bbox": [ + 186, + 337, + 287, + 347 + ], + "spans": [ + { + "bbox": [ + 186, + 337, + 287, + 347 + ], + "type": "text", + "content": "(a) Individual distribution" + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "image_caption" + } + ], + "index": 4 + }, + { + "type": "image", + "bbox": [ + 313, + 224, + 490, + 337 + ], + "blocks": [ + { + "bbox": [ + 313, + 224, + 490, + 337 + ], + "lines": [ + { + "bbox": [ + 313, + 224, + 490, + 337 + ], + "spans": [ + { + "bbox": [ + 313, + 224, + 490, + 337 + ], + "type": "image", + "image_path": "1a0ecff7a365b6e7cc328c4333ae9d327b94127160c2ae3f53ec2fb55a39d122.jpg" + } + ] + } + ], + "index": 6, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 364, + 337, + 440, + 348 + ], + "lines": [ + { + "bbox": [ + 364, + 337, + 440, + 348 + ], + "spans": [ + { + "bbox": [ + 364, + 337, + 440, + 348 + ], + "type": "text", + "content": "(b) Sex distribution" + } + ] + } + ], + "index": 7, + "angle": 0, + "type": "image_caption" + } + ], + "index": 6 + }, + { + "type": "image", + "bbox": [ + 156, + 358, + 256, + 467 + ], + "blocks": [ + { + "bbox": [ + 156, + 358, + 256, + 467 + ], + "lines": [ + { + "bbox": [ + 156, + 358, + 256, + 467 + ], + "spans": [ + { + "bbox": [ + 156, + 358, + 256, + 467 + ], + "type": "image", + "image_path": "f9693d6e47880e64254c39c50e81fe1f6d914d08f00356bc1442edddf64bf47d.jpg" + } + ] + } + ], + "index": 8, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 249, + 471, + 368, + 483 + ], + "lines": [ + { + "bbox": [ + 249, + 471, + 368, + 483 + ], + "spans": [ + { + "bbox": [ + 249, + 471, + 368, + 483 + ], + "type": "text", + "content": "(c) Breed annotation examples" + } + ] + } + ], + "index": 11, + "angle": 0, + "type": "image_caption" + } + ], + "index": 8 + }, + { + "type": "image", + "bbox": [ + 265, + 358, + 352, + 467 + ], + "blocks": [ + { + "bbox": [ + 265, + 358, + 352, + 467 + ], + "lines": [ + { + "bbox": [ + 265, + 358, + 352, + 467 + ], + "spans": [ + { + "bbox": [ + 265, + 358, + 352, + 467 + ], + "type": "image", + "image_path": "49c5cf42f668e584119b5044185027d61611b7fe37ba5d2c479a6baac7bba7a5.jpg" + } + ] + } + ], + "index": 9, + "angle": 0, + "type": "image_body" + } + ], + "index": 9 + }, + { + "type": "image", + "bbox": [ + 369, + 358, + 455, + 467 + ], + "blocks": [ + { + "bbox": [ + 369, + 358, + 455, + 467 + ], + "lines": [ + { + "bbox": [ + 369, + 358, + 455, + 467 + ], + "spans": [ + { + "bbox": [ + 369, + 358, + 455, + 467 + ], + "type": "image", + "image_path": "bc92ab628140b13619e1be092caa67ee40660ca78e28dd92c19d9ee71e912ca9.jpg" + } + ] + } + ], + "index": 10, + "angle": 0, + "type": "image_body" + } + ], + "index": 10 + }, + { + "type": "image", + "bbox": [ + 141, + 495, + 244, + 602 + ], + "blocks": [ + { + "bbox": [ + 141, + 495, + 244, + 602 + ], + "lines": [ + { + "bbox": [ + 141, + 495, + 244, + 602 + ], + "spans": [ + { + "bbox": [ + 141, + 495, + 244, + 602 + ], + "type": "image", + "image_path": "40cb436f1e5bab23e3850bf21b40418f4b39fdd502fc000d83295300609535f6.jpg" + } + ] + } + ], + "index": 12, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 130, + 628, + 482, + 662 + ], + "lines": [ + { + "bbox": [ + 130, + 628, + 482, + 662 + ], + "spans": [ + { + "bbox": [ + 130, + 628, + 482, + 662 + ], + "type": "text", + "content": "Fig. 2: Dataset distribution. Each figure represents (a) the number of individuals in each animal family. (b) the sex distribution percentage by animal family. (c) examples of breed annotations. (d) examples of color annotations." + } + ] + } + ], + "index": 16, + "angle": 0, + "type": "image_caption" + } + ], + "index": 12 + }, + { + "type": "image", + "bbox": [ + 261, + 496, + 353, + 601 + ], + "blocks": [ + { + "bbox": [ + 261, + 496, + 353, + 601 + ], + "lines": [ + { + "bbox": [ + 261, + 496, + 353, + 601 + ], + "spans": [ + { + "bbox": [ + 261, + 496, + 353, + 601 + ], + "type": "image", + "image_path": "1d2a7a5b2af52220fb7036713f277dd2e7db3a3e525510caa810ff8e8e0932ad.jpg" + } + ] + } + ], + "index": 13, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 254, + 603, + 372, + 615 + ], + "lines": [ + { + "bbox": [ + 254, + 603, + 372, + 615 + ], + "spans": [ + { + "bbox": [ + 254, + 603, + 372, + 615 + ], + "type": "text", + "content": "(d) Color annotation examples" + } + ] + } + ], + "index": 15, + "angle": 0, + "type": "image_caption" + } + ], + "index": 13 + }, + { + "type": "image", + "bbox": [ + 368, + 496, + 463, + 600 + ], + "blocks": [ + { + "bbox": [ + 368, + 496, + 463, + 600 + ], + "lines": [ + { + "bbox": [ + 368, + 496, + 463, + 600 + ], + "spans": [ + { + "bbox": [ + 368, + 496, + 463, + 600 + ], + "type": "image", + "image_path": "a57b023c7b588e33d9f61b895fd0c949f60a2c49f419acf2a24b0e289e81784f.jpg" + } + ] + } + ], + "index": 14, + "angle": 0, + "type": "image_body" + } + ], + "index": 14 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 148, + 91, + 447, + 102 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 148, + 91, + 447, + 102 + ], + "spans": [ + { + "bbox": [ + 148, + 91, + 447, + 102 + ], + "type": "text", + "content": "PetFace: A Large-Scale Dataset and Benchmark for Animal Identification" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 474, + 92, + 480, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 474, + 92, + 480, + 100 + ], + "spans": [ + { + "bbox": [ + 474, + 92, + 480, + 100 + ], + "type": "text", + "content": "5" + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 4 + }, + { + "para_blocks": [ + { + "bbox": [ + 130, + 116, + 482, + 236 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 116, + 482, + 236 + ], + "spans": [ + { + "bbox": [ + 130, + 116, + 482, + 236 + ], + "type": "text", + "content": "The PetFace dataset encompasses 1,012,934 images spanning 257,484 unique individuals. Detailed distributions of animal families are illustrated in Fig 2(a). Images are cropped to " + }, + { + "bbox": [ + 130, + 116, + 482, + 236 + ], + "type": "inline_equation", + "content": "224 \\times 224" + }, + { + "bbox": [ + 130, + 116, + 482, + 236 + ], + "type": "text", + "content": " pixels around their faces. Our PetFace also has fine-grained annotations. Sex distribution across different animal families is depicted in Fig 2(b), with sex information available for 240,861 individuals, accounting for " + }, + { + "bbox": [ + 130, + 116, + 482, + 236 + ], + "type": "inline_equation", + "content": "94\\%" + }, + { + "bbox": [ + 130, + 116, + 482, + 236 + ], + "type": "text", + "content": " of the dataset. The dataset includes annotations for 319 breeds, with examples of breed annotations shown in Fig 2(c). Furthermore, as detailed in Fig 2(d), the dataset provides in-depth color information through two-tier hierarchical annotations. Please see the supplemental for the detailed information per each animal family." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 132, + 252, + 234, + 264 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 252, + 234, + 264 + ], + "spans": [ + { + "bbox": [ + 132, + 252, + 234, + 264 + ], + "type": "text", + "content": "3.2 Data Sourcing" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 130, + 271, + 482, + 342 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 271, + 482, + 342 + ], + "spans": [ + { + "bbox": [ + 130, + 271, + 482, + 342 + ], + "type": "text", + "content": "Collecting images of animal faces via photography is labor and time intensive, which impedes the creation of large-scale datasets. In contrast, the human face recognition field benefits greatly from the availability of images sourced from the Internet. In this section, we outline our approach to assembling the collection of animal face images through the Internet, where each is associated with unique individual identifiers." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 130, + 343, + 482, + 510 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 343, + 482, + 510 + ], + "spans": [ + { + "bbox": [ + 130, + 343, + 482, + 510 + ], + "type": "text", + "content": "Curation of images. In contrast to the human domain, acquiring multiple images for individual animals is more challenging. Unlike human datasets where a large number of celebrity images can be readily sourced, animal images require alternative approaches for curation. We utilized two primary sources: (i) pet shops' websites and (ii) animal adoption websites. The advantage of pet shops lies in their provision of high-quality, diverse images, capturing individuals from various angles and offering detailed information about each animal, including color, sex, and specific breed details. On the other hand, animal adoption sites offer images set against a variety of backgrounds and conditions that are often provided by pet owners, thus ensuring each individual is presented in a unique setting. These sources provide images that are highly suitable for animal recognition tasks and that are especially useful for recognizing animals in varied wild environments. For Chimps, we additionally use images from the webpage of a collaborative research institution." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 130, + 511, + 482, + 606 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 511, + 482, + 606 + ], + "spans": [ + { + "bbox": [ + 130, + 511, + 482, + 606 + ], + "type": "text", + "content": "To ensure the dataset's quality, we were selective in choosing websites for curation. We only use the websites introducing each animal on one page to gain the individual IDs. Aware of the potential for pet owners to upload the same images to multiple websites, we chose a single pet adoption website from each region (e.g., one per country) to minimize duplicates. For pet shops, we confirmed that the animals listed were unique to a certain shop to ensure the uniqueness of our dataset entries. Using these sources, we collected 1,443,737 images from 325,420 individuals." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 132, + 622, + 310, + 635 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 622, + 310, + 635 + ], + "spans": [ + { + "bbox": [ + 132, + 622, + 310, + 635 + ], + "type": "text", + "content": "3.3 Face Alignment and Filtering" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 130, + 641, + 481, + 666 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 641, + 481, + 666 + ], + "spans": [ + { + "bbox": [ + 130, + 641, + 481, + 666 + ], + "type": "text", + "content": "Face Detection. We detect facial landmarks to align and crop the images. We adopt the AnyFace [35] that is trained on mixed face datasets including" + } + ] + } + ], + "index": 8 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "spans": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "type": "text", + "content": "6" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 166, + 91, + 282, + 101 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 166, + 91, + 282, + 101 + ], + "spans": [ + { + "bbox": [ + 166, + 91, + 282, + 101 + ], + "type": "text", + "content": "R. Shinoda and K. Shiohara" + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 5 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 134, + 113, + 491, + 264 + ], + "blocks": [ + { + "bbox": [ + 134, + 113, + 491, + 264 + ], + "lines": [ + { + "bbox": [ + 134, + 113, + 491, + 264 + ], + "spans": [ + { + "bbox": [ + 134, + 113, + 491, + 264 + ], + "type": "image", + "image_path": "ffc1f5b9fd830e0de23a28d0bc1c7af993bade2c59de8b24208e9d8c074a9703.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 130, + 274, + 482, + 319 + ], + "lines": [ + { + "bbox": [ + 130, + 274, + 482, + 319 + ], + "spans": [ + { + "bbox": [ + 130, + 274, + 482, + 319 + ], + "type": "text", + "content": "Fig. 3: Data filtering process. We adopt a two-stage data filtering approach. Initially, images containing multiple faces are automatically removed. Subsequently, any images that do not depict animal faces or fail to meet alignment criteria are manually eliminated." + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_caption" + } + ], + "index": 2 + }, + { + "bbox": [ + 130, + 342, + 482, + 413 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 342, + 482, + 413 + ], + "spans": [ + { + "bbox": [ + 130, + 342, + 482, + 413 + ], + "type": "text", + "content": "AnimalWeb [32]. Because the positions of animal facial parts are sometimes very different from those of humans, we use different reference points for different species in face alignment. After landmark detection, we select one frontal image as a reference for each species. We compute the average landmarks over all the landmarks aligned with the reference. Then, we define the target positions of the landmarks in the images; all the images are aligned to the target positions." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 130, + 414, + 482, + 582 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 414, + 482, + 582 + ], + "spans": [ + { + "bbox": [ + 130, + 414, + 482, + 582 + ], + "type": "text", + "content": "Data Filtering. Fig 3 shows the overview of our data filtering process. Because the fully automatic face detection above sometimes fails, we adopt a two-stage data filtering process to filter out the following cases: 1) Images contain multiple animals simultaneously. 2) Images are unrelated to animals, such as advertisements. 3) Images focus on non-animal elements, such as random patterns in backgrounds, toys, or people, rather than the animals themselves. First, we automatically remove images where multiple faces are detected in an image. Secondly, annotators manually assess all the images and remove those that do not depict the target animals or where the face alignment differs from our intended criteria. Because checking whether the faces are properly aligned or not requires some expertise, all the images are filtered by the authors to ensure the quality of the dataset. This manual filtering process took about 100 man hours. After this stage, we keep 1,012,934 images, which is approximately " + }, + { + "bbox": [ + 130, + 414, + 482, + 582 + ], + "type": "inline_equation", + "content": "70\\%" + }, + { + "bbox": [ + 130, + 414, + 482, + 582 + ], + "type": "text", + "content": " of the initial image number. Detailed distributions of animal families are illustrated in Fig 2(a)." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 132, + 598, + 293, + 609 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 598, + 293, + 609 + ], + "spans": [ + { + "bbox": [ + 132, + 598, + 293, + 609 + ], + "type": "text", + "content": "3.4 Fine-Grained Annotations" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 130, + 617, + 482, + 666 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 617, + 482, + 666 + ], + "spans": [ + { + "bbox": [ + 130, + 617, + 482, + 666 + ], + "type": "text", + "content": "Fine-grained categorization enhances its usefulness for downstream applications, e.g., the creation of challenging datasets based on animal's individual attributes. However, collecting extensive individual animal data poses a major challenge in assembling large datasetse owing to the significant effort involved. In response," + } + ] + } + ], + "index": 7 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 148, + 91, + 447, + 102 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 148, + 91, + 447, + 102 + ], + "spans": [ + { + "bbox": [ + 148, + 91, + 447, + 102 + ], + "type": "text", + "content": "PetFace: A Large-Scale Dataset and Benchmark for Animal Identification" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 474, + 92, + 480, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 474, + 92, + 480, + 100 + ], + "spans": [ + { + "bbox": [ + 474, + 92, + 480, + 100 + ], + "type": "text", + "content": "7" + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 6 + }, + { + "para_blocks": [ + { + "bbox": [ + 130, + 116, + 479, + 199 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 116, + 479, + 199 + ], + "spans": [ + { + "bbox": [ + 130, + 116, + 479, + 199 + ], + "type": "text", + "content": "we collect as much individual animal information corresponding to the images as possible. Our annotations include sex, color, and specific breed details, depending on their availability on the source websites. The annotation process involves: 1) extracting individual information displayed on websites alongside images, 2) refining raw text data to develop individual data tables, 3) manually verifying that all attributes are refined and free of unrelated information, and 4) manually annotating individuals with missing attributes on the website." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 130, + 200, + 480, + 308 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 200, + 480, + 308 + ], + "spans": [ + { + "bbox": [ + 130, + 200, + 480, + 308 + ], + "type": "text", + "content": "As a result, we include a sex category for all classes, with breed annotations for eight animal classes (Cat, Chimp, Dog, Guinea pig, Hamster, Parakeet, Pig, and Rabbit), and color annotations for eleven animal classes (Cat, Chinchilla, Degus, Ferret, Guinea pig, Hamster, Hedgehog, Parakeet, Java sparrow, Pig, and Rabbit). Note that the manual annotations are performed only on colors and patterns because images alone do not allow us to accurately determine an animal's sex or breed. Detailed distributions of sex per animal families are illustrated in Fig 2(b). In Fig 2(c) and (d), we illustrate the examples of breed annotations and color and pattern annotations." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 132, + 327, + 316, + 341 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 327, + 316, + 341 + ], + "spans": [ + { + "bbox": [ + 132, + 327, + 316, + 341 + ], + "type": "text", + "content": "4 Experimental Methodology" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 130, + 353, + 480, + 413 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 353, + 480, + 413 + ], + "spans": [ + { + "bbox": [ + 130, + 353, + 480, + 413 + ], + "type": "text", + "content": "We introduce two principal evaluation protocols: 1) re-identification for seen faces, and 2) verification for unseen faces. To establish these benchmarks, we have curated two distinct types of test sets. Specifically, for unseen face verification, we carefully separate the dataset into training, validation, and test sets to prevent data leakage that can arise from shared backgrounds." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 132, + 431, + 265, + 442 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 431, + 265, + 442 + ], + "spans": [ + { + "bbox": [ + 132, + 431, + 265, + 442 + ], + "type": "text", + "content": "4.1 Evaluation Protocols" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 132, + 452, + 284, + 464 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 452, + 284, + 464 + ], + "spans": [ + { + "bbox": [ + 132, + 452, + 284, + 464 + ], + "type": "text", + "content": "We adopt two types of evaluation." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 136, + 474, + 481, + 664 + ], + "type": "list", + "angle": 0, + "index": 10, + "blocks": [ + { + "bbox": [ + 136, + 474, + 481, + 544 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 136, + 474, + 481, + 544 + ], + "spans": [ + { + "bbox": [ + 136, + 474, + 481, + 544 + ], + "type": "text", + "content": "1. Re-identification: This evaluation selects images from the test data that match the identities present in the training data. This procedure involves identifying test images that correspond to the same identities used during model training. The objective is to assess the model's ability to accurately recognize and associate test images with the correct identities from the training dataset." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 136, + 546, + 481, + 664 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 136, + 546, + 481, + 664 + ], + "spans": [ + { + "bbox": [ + 136, + 546, + 481, + 664 + ], + "type": "text", + "content": "2. Verification: This evaluation checks if the model can identify unseen faces. This means that we use different identities during the training and testing phases. For each identity, we select one image that matches the identity and one image randomly chosen from a different identity. This process creates pairs of images where one pair consists of images from the same identity, and another pair consists of images from different identities. We then task the model with predicting whether the faces in each pair belong to the same identity or different identities. This approach allows us to comprehensively evaluate the model's ability to recognize and differentiate between individual identities." + } + ] + } + ], + "index": 9 + } + ], + "sub_type": "text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "spans": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "type": "text", + "content": "8" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 166, + 91, + 282, + 101 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 166, + 91, + 282, + 101 + ], + "spans": [ + { + "bbox": [ + 166, + 91, + 282, + 101 + ], + "type": "text", + "content": "R. Shinoda and K. Shiohara" + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 7 + }, + { + "para_blocks": [ + { + "bbox": [ + 132, + 116, + 212, + 128 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 116, + 212, + 128 + ], + "spans": [ + { + "bbox": [ + 132, + 116, + 212, + 128 + ], + "type": "text", + "content": "4.2 Data Split" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 130, + 140, + 482, + 272 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 140, + 482, + 272 + ], + "spans": [ + { + "bbox": [ + 130, + 140, + 482, + 272 + ], + "type": "text", + "content": "To establish benchmarks for seen individuals re-identification and unseen individuals verification, we use specific split protocols, creating two test sets designed for each task. For seen individuals re-identification, the process involves verifying images of the same identities used in training; thus, we select test images from the same individuals for the training set. We, therefore, ensure that each individual is represented by at least three photos i.e., a minimum of two for training and one for testing. For unseen individuals verification dataset, we ensure that it contains no images from sources that are also represented in the training or validation sets per animal family to prevent bias from similar environmental conditions. We divide the dataset into training, validation, and each testing sets following a 7:1:2 ratio as closely as possible, given the outlined criteria." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 132, + 293, + 198, + 304 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 293, + 198, + 304 + ], + "spans": [ + { + "bbox": [ + 132, + 293, + 198, + 304 + ], + "type": "text", + "content": "4.3 Models" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 130, + 316, + 482, + 423 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 316, + 482, + 423 + ], + "spans": [ + { + "bbox": [ + 130, + 316, + 482, + 423 + ], + "type": "text", + "content": "We train state-of-the-art models based on deep neural networks on our PetFace to build the benchmarks. For the re-identification task, we approach the training and testing phases as classification problems; an individual is assigned to a class. On the other hand, for the verification task, we train models in the same manner as for the re-identification task but evaluate the models by computing the cosine similarity between pairs of images to be identified if they are the same individual. Specifically, we focus more on the loss functions that are crucial for identification tasks than network architectures. We refer to three important loss functions in addition to the basic Softmax-based classification model:" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 130, + 425, + 482, + 520 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 425, + 482, + 520 + ], + "spans": [ + { + "bbox": [ + 130, + 425, + 482, + 520 + ], + "type": "text", + "content": "Triplet loss [26] is designed to take a triplet of samples, i.e., an anchor " + }, + { + "bbox": [ + 130, + 425, + 482, + 520 + ], + "type": "inline_equation", + "content": "x_{a}" + }, + { + "bbox": [ + 130, + 425, + 482, + 520 + ], + "type": "text", + "content": ", a positive " + }, + { + "bbox": [ + 130, + 425, + 482, + 520 + ], + "type": "inline_equation", + "content": "x_{p}" + }, + { + "bbox": [ + 130, + 425, + 482, + 520 + ], + "type": "text", + "content": " (another image of the same identity as the anchor), and a negative " + }, + { + "bbox": [ + 130, + 425, + 482, + 520 + ], + "type": "inline_equation", + "content": "x_{n}" + }, + { + "bbox": [ + 130, + 425, + 482, + 520 + ], + "type": "text", + "content": " (an image of a different identity) - and learn embeddings in such a way that the " + }, + { + "bbox": [ + 130, + 425, + 482, + 520 + ], + "type": "inline_equation", + "content": "x_{p}" + }, + { + "bbox": [ + 130, + 425, + 482, + 520 + ], + "type": "text", + "content": " is closer to the " + }, + { + "bbox": [ + 130, + 425, + 482, + 520 + ], + "type": "inline_equation", + "content": "x_{a}" + }, + { + "bbox": [ + 130, + 425, + 482, + 520 + ], + "type": "text", + "content": " than the " + }, + { + "bbox": [ + 130, + 425, + 482, + 520 + ], + "type": "inline_equation", + "content": "x_{n}" + }, + { + "bbox": [ + 130, + 425, + 482, + 520 + ], + "type": "text", + "content": " by a margin. This loss function is particularly beneficial for face Re-ID as it directly targets the relative distances between different and same identity pairs, encouraging the model to learn a feature space where embeddings of the same identity are clustered together while being far from other identity clusters." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 130, + 521, + 482, + 605 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 521, + 482, + 605 + ], + "spans": [ + { + "bbox": [ + 130, + 521, + 482, + 605 + ], + "type": "text", + "content": "Center loss [65] works alongside Softmax Loss to enhance the discriminative power of the learned features. While Softmax Loss focuses on inter-class separability, Center Loss aims to minimize the intra-class variations. It does this by penalizing the distance between the deep features of each class and their corresponding class center. Center Loss ensures that the embeddings of the faces of the same individual are closer together, thus making the feature distribution more compact for each identity." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 130, + 605, + 482, + 665 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 605, + 482, + 665 + ], + "spans": [ + { + "bbox": [ + 130, + 605, + 482, + 665 + ], + "type": "text", + "content": "ArcFace loss [15] introduces an angular margin between classes to enforce a discriminative feature space. It modifies the Softmax loss by adding a margin penalty to the angle between the feature vector and the corresponding class center in the angular space. This angular margin encourages models to learn more distinguishable embeddings to separate classes." + } + ] + } + ], + "index": 8 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 148, + 91, + 447, + 102 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 148, + 91, + 447, + 102 + ], + "spans": [ + { + "bbox": [ + 148, + 91, + 447, + 102 + ], + "type": "text", + "content": "PetFace: A Large-Scale Dataset and Benchmark for Animal Identification" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 474, + 92, + 480, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 474, + 92, + 480, + 100 + ], + "spans": [ + { + "bbox": [ + 474, + 92, + 480, + 100 + ], + "type": "text", + "content": "9" + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 8 + }, + { + "para_blocks": [ + { + "bbox": [ + 130, + 116, + 479, + 152 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 116, + 479, + 152 + ], + "spans": [ + { + "bbox": [ + 130, + 116, + 479, + 152 + ], + "type": "text", + "content": "We use ResNet-50 [23], which we found performs better than recent Transformer-based models in Table 5, as our base backbone for all applied loss functions to simplify our experiments and make them easier to grasp." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 130, + 152, + 479, + 176 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 152, + 479, + 176 + ], + "spans": [ + { + "bbox": [ + 130, + 152, + 479, + 176 + ], + "type": "text", + "content": "In addition to the comparison on the models trained on PetFace, we evaluate state-of-the-art models trained on other datasets including:" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 130, + 176, + 479, + 200 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 176, + 479, + 200 + ], + "spans": [ + { + "bbox": [ + 130, + 176, + 479, + 200 + ], + "type": "text", + "content": "ImageNet [14] is a conventional image classification dataset including 1000 general object classes. We use the ResNet-50 trained on the dataset." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 130, + 200, + 491, + 236 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 200, + 491, + 236 + ], + "spans": [ + { + "bbox": [ + 130, + 200, + 491, + 236 + ], + "type": "text", + "content": "CLIP [54] learns semantic relationships between images and texts in a cross-modal contrastive learning manner empowered by web-scale image-caption datasets. We use ResNet-50 backbone." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 130, + 236, + 482, + 285 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 236, + 482, + 285 + ], + "spans": [ + { + "bbox": [ + 130, + 236, + 482, + 285 + ], + "type": "text", + "content": "MegaDescriptor [61] is a state-of-the-art model for animal identification trained on a unified animal identification dataset including 33 existing animal re-identification datasets [1-4, 6-9, 11, 12, 16, 18-20, 22, 24, 27, 28, 36, 38, 46, 50-53, 60, 62, 64, 66, 68, 69]. We use the officially provided SwinTransformer-B [41] backbone." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 131, + 304, + 276, + 317 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 131, + 304, + 276, + 317 + ], + "spans": [ + { + "bbox": [ + 131, + 304, + 276, + 317 + ], + "type": "text", + "content": "5 Experimental Result" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 131, + 330, + 387, + 341 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 131, + 330, + 387, + 341 + ], + "spans": [ + { + "bbox": [ + 131, + 330, + 387, + 341 + ], + "type": "text", + "content": "5.1 Benchmark on Animal Face Re-identification" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 130, + 350, + 479, + 456 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 350, + 479, + 456 + ], + "spans": [ + { + "bbox": [ + 130, + 350, + 479, + 456 + ], + "type": "text", + "content": "We show the re-identification results in terms AUC in Table 2. We train models with the baseline loss functions independently on each class. We can see that ArcFace performs consistent results, " + }, + { + "bbox": [ + 130, + 350, + 479, + 456 + ], + "type": "inline_equation", + "content": "51.23\\%" + }, + { + "bbox": [ + 130, + 350, + 479, + 456 + ], + "type": "text", + "content": " of average accuracy, compared to the other loss functions (41.88% for Softmax and 9.81% for Center). On the other hand, Center loss does not learn sufficient discriminative features for animal face re-identification, especially on Cat and Dog where the number of the test classes are 113,592 and 46,755, respectively. These results encourage the community to explore more effective representation learning methods for re-identification tasks." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 130, + 459, + 479, + 541 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 459, + 479, + 541 + ], + "spans": [ + { + "bbox": [ + 130, + 459, + 479, + 541 + ], + "type": "text", + "content": "Moreover, motivated by MegaDescriptor, we train an ArcFace model on the entire PetFace, denoted as Joint-Trained on PetFace in the table. We observe that the joint-training strategy provides some improvements, e.g., from " + }, + { + "bbox": [ + 130, + 459, + 479, + 541 + ], + "type": "inline_equation", + "content": "54.29\\%" + }, + { + "bbox": [ + 130, + 459, + 479, + 541 + ], + "type": "text", + "content": " to " + }, + { + "bbox": [ + 130, + 459, + 479, + 541 + ], + "type": "inline_equation", + "content": "70.30\\%" + }, + { + "bbox": [ + 130, + 459, + 479, + 541 + ], + "type": "text", + "content": " for Cat and from " + }, + { + "bbox": [ + 130, + 459, + 479, + 541 + ], + "type": "inline_equation", + "content": "29.08\\%" + }, + { + "bbox": [ + 130, + 459, + 479, + 541 + ], + "type": "text", + "content": " to " + }, + { + "bbox": [ + 130, + 459, + 479, + 541 + ], + "type": "inline_equation", + "content": "41.49\\%" + }, + { + "bbox": [ + 130, + 459, + 479, + 541 + ], + "type": "text", + "content": " for Pig although in some classes the results get worse, e.g., from " + }, + { + "bbox": [ + 130, + 459, + 479, + 541 + ], + "type": "inline_equation", + "content": "43.27\\%" + }, + { + "bbox": [ + 130, + 459, + 479, + 541 + ], + "type": "text", + "content": " to " + }, + { + "bbox": [ + 130, + 459, + 479, + 541 + ], + "type": "inline_equation", + "content": "34.30\\%" + }, + { + "bbox": [ + 130, + 459, + 479, + 541 + ], + "type": "text", + "content": " for Chimp and from " + }, + { + "bbox": [ + 130, + 459, + 479, + 541 + ], + "type": "inline_equation", + "content": "62.19\\%" + }, + { + "bbox": [ + 130, + 459, + 479, + 541 + ], + "type": "text", + "content": " to " + }, + { + "bbox": [ + 130, + 459, + 479, + 541 + ], + "type": "inline_equation", + "content": "44.78\\%" + }, + { + "bbox": [ + 130, + 459, + 479, + 541 + ], + "type": "text", + "content": " to Parakeet. This result indicates that there is some room to improve the performance of integrated identification on imbalance and wide-range datasets." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 130, + 542, + 479, + 590 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 542, + 479, + 590 + ], + "spans": [ + { + "bbox": [ + 130, + 542, + 479, + 590 + ], + "type": "text", + "content": "In Fig 4, we display the Top-" + }, + { + "bbox": [ + 130, + 542, + 479, + 590 + ], + "type": "inline_equation", + "content": "k" + }, + { + "bbox": [ + 130, + 542, + 479, + 590 + ], + "type": "text", + "content": " (" + }, + { + "bbox": [ + 130, + 542, + 479, + 590 + ], + "type": "inline_equation", + "content": "k = 1,3,5" + }, + { + "bbox": [ + 130, + 542, + 479, + 590 + ], + "type": "text", + "content": ") accuracy metrics. While accuracy naturally increases with larger " + }, + { + "bbox": [ + 130, + 542, + 479, + 590 + ], + "type": "inline_equation", + "content": "k" + }, + { + "bbox": [ + 130, + 542, + 479, + 590 + ], + "type": "text", + "content": " values, the relative ranking of accuracy across different animal families remains largely consistent. ArcFace maintains the best performance across these evaluations." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 131, + 609, + 361, + 620 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 131, + 609, + 361, + 620 + ], + "spans": [ + { + "bbox": [ + 131, + 609, + 361, + 620 + ], + "type": "text", + "content": "5.2 Benchmark on Animal Face Verification" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 130, + 629, + 479, + 665 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 629, + 479, + 665 + ], + "spans": [ + { + "bbox": [ + 130, + 629, + 479, + 665 + ], + "type": "text", + "content": "Next, we show the verification results in Table 3. We use the trained models in Table 2 and evaluate them on unseen individuals. Similar to our findings in the re-identification task, ArcFace proved to be effective, achieving the best" + } + ] + } + ], + "index": 13 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "spans": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "type": "text", + "content": "10" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 166, + 91, + 282, + 101 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 166, + 91, + 282, + 101 + ], + "spans": [ + { + "bbox": [ + 166, + 91, + 282, + 101 + ], + "type": "text", + "content": "R. Shinoda and K. Shiohara" + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 9 + }, + { + "para_blocks": [ + { + "type": "table", + "bbox": [ + 132, + 114, + 481, + 188 + ], + "blocks": [ + { + "bbox": [ + 132, + 114, + 481, + 188 + ], + "lines": [ + { + "bbox": [ + 132, + 114, + 481, + 188 + ], + "spans": [ + { + "bbox": [ + 132, + 114, + 481, + 188 + ], + "type": "table", + "html": "
MethodTop-1 Accuracy (%)Avg
CatChimpChinchillaDegusDogFerretGuineaHamsterHedgehogParakeetJava sparrowPigRabbit
Trained on PetFace
Softmax30.4641.7058.1337.8459.1425.8860.0738.2727.8150.8833.5521.6459.0541.88
Center0.005.3829.7612.030.007.9131.779.4613.761.684.867.443.519.81
Arcface54.2943.2767.3445.4177.8628.9267.9047.3730.9062.1942.2729.0869.1351.23
Joint-Trained on PetFace
ArcFace70.3034.3069.8656.0868.7546.1268.6654.3344.3844.7834.5541.4965.7553.80
", + "image_path": "9497bce529f21be67ff21894144ca61164219dfb7034b8dbd7415c946ffd99b7.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "table_body" + } + ], + "index": 2 + }, + { + "type": "table", + "bbox": [ + 132, + 242, + 481, + 349 + ], + "blocks": [ + { + "bbox": [ + 132, + 190, + 479, + 221 + ], + "lines": [ + { + "bbox": [ + 132, + 190, + 479, + 221 + ], + "spans": [ + { + "bbox": [ + 132, + 190, + 479, + 221 + ], + "type": "text", + "content": "Table 2: Result on animal re-identification. ArcFace outperforms the other loss functions. The jointly trained model on our PetFace shows the best average top-1 accuracy." + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 132, + 242, + 481, + 349 + ], + "lines": [ + { + "bbox": [ + 132, + 242, + 481, + 349 + ], + "spans": [ + { + "bbox": [ + 132, + 242, + 481, + 349 + ], + "type": "table", + "html": "
MethodAUC (%)Avg
CatChimpChinchillaDegusDogFerretGuinea pigHamsterHedgehogParakeetJava sparrowPigRabbit
Trained on PetFace
Softmax97.9785.2284.4486.1998.9883.1595.3087.7387.2798.3984.6787.0998.5390.38
Center50.2081.6080.9681.3459.6478.7792.1083.5985.2696.9180.5482.5897.0180.81
Triplet96.9477.1076.1272.8097.9779.4883.3780.0881.4791.5677.8675.7294.7483.48
Arcface97.7183.7687.7087.6199.4586.2496.0390.3286.1398.6384.4589.8899.0191.30
Joint-Trained on PetFace
ArcFace98.0483.2089.6386.7999.0188.4494.1492.5087.8698.4089.3892.3898.4592.17
Trained on Other Dataset
ImageNet81.5873.6770.8070.6997.1867.9585.1974.1375.0190.4579.4982.4286.2379.60
CLIP85.1070.7073.4274.6691.8670.8980.6773.9177.6583.1976.5280.0283.5578.63
MegaDescriptor87.3083.0177.8577.7993.7576.5388.3677.9178.7489.7678.6488.0890.3883.70
", + "image_path": "a742a8c1775f673c63046ff44bb0c93656fda4c529395098795b05c88b9db41c.jpg" + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "table_body" + } + ], + "index": 4 + }, + { + "bbox": [ + 132, + 350, + 479, + 370 + ], + "lines": [ + { + "bbox": [ + 132, + 350, + 479, + 370 + ], + "spans": [ + { + "bbox": [ + 132, + 350, + 479, + 370 + ], + "type": "text", + "content": "Table 3: Result on animal verification. Jointly trained on our PetFace shows the best performance. We also compared to the trained models on other dataset." + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "text" + }, + { + "bbox": [ + 130, + 403, + 482, + 559 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 403, + 482, + 559 + ], + "spans": [ + { + "bbox": [ + 130, + 403, + 482, + 559 + ], + "type": "text", + "content": "results with an average AUC of " + }, + { + "bbox": [ + 130, + 403, + 482, + 559 + ], + "type": "inline_equation", + "content": "92.17\\%" + }, + { + "bbox": [ + 130, + 403, + 482, + 559 + ], + "type": "text", + "content": " when jointly trained on the PetFace dataset. Similar to face re-identification in Sec 5.1, in some classes, the results get worse when jointly trained across animal families e.g., Chimp, Degus, Dog, Guinea pig, Parakeet, and Rabbit. We also examined models pre-trained on other datasets including ImageNet, CLIP and MegaDescriptor in order to make more comparisons. Among these models, MegaDescriptor, which trained on the animal re-identification dataset, showed the highest AUC of " + }, + { + "bbox": [ + 130, + 403, + 482, + 559 + ], + "type": "inline_equation", + "content": "83.70\\%" + }, + { + "bbox": [ + 130, + 403, + 482, + 559 + ], + "type": "text", + "content": ". We show the similarity distribution on the our cat dataset in Fig 5. Comparing the performance of CLIP, MegaDescriptor, and our joint-trained models, our findings indicate that our model achieves the most distinct separation between positive and negative samples. In contrast, the MegaDescriptor model displays a closer distribution of these samples, and with CLIP, the overlap between positive and negative distributions is significantly more pronounced." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 132, + 575, + 342, + 586 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 575, + 342, + 586 + ], + "spans": [ + { + "bbox": [ + 132, + 575, + 342, + 586 + ], + "type": "text", + "content": "5.3 Comparison with Previous Datasets" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 130, + 594, + 485, + 665 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 594, + 485, + 665 + ], + "spans": [ + { + "bbox": [ + 130, + 594, + 485, + 665 + ], + "type": "text", + "content": "We conduct cross-dataset evaluations where models are tested on different datasets other than the training datasets to demonstrate the generality of our dataset. Here, we compare our dataset with 1) CTai [19] and CZoo [19] for Chimpanzee and 2) DogFaceNet [49] and Flickr-dog [47] for Dog. All images of the compared datasets are aligned in the same manner as ours for fair comparisons. We split the identities of the compared datasets into training and test sets in a ratio of 7:3" + } + ] + } + ], + "index": 8 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 148, + 91, + 447, + 102 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 148, + 91, + 447, + 102 + ], + "spans": [ + { + "bbox": [ + 148, + 91, + 447, + 102 + ], + "type": "text", + "content": "PetFace: A Large-Scale Dataset and Benchmark for Animal Identification" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 470, + 91, + 479, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 470, + 91, + 479, + 100 + ], + "spans": [ + { + "bbox": [ + 470, + 91, + 479, + 100 + ], + "type": "text", + "content": "11" + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 10 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 140, + 114, + 247, + 212 + ], + "blocks": [ + { + "bbox": [ + 140, + 114, + 247, + 212 + ], + "lines": [ + { + "bbox": [ + 140, + 114, + 247, + 212 + ], + "spans": [ + { + "bbox": [ + 140, + 114, + 247, + 212 + ], + "type": "image", + "image_path": "44020877bc6b8dd93cd55b2f47107c5f99c014fca4ac624e1d4666e65beab0b7.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 175, + 216, + 211, + 225 + ], + "lines": [ + { + "bbox": [ + 175, + 216, + 211, + 225 + ], + "spans": [ + { + "bbox": [ + 175, + 216, + 211, + 225 + ], + "type": "text", + "content": "(a) Top-1" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_caption" + } + ], + "index": 2 + }, + { + "type": "image", + "bbox": [ + 254, + 114, + 361, + 212 + ], + "blocks": [ + { + "bbox": [ + 254, + 114, + 361, + 212 + ], + "lines": [ + { + "bbox": [ + 254, + 114, + 361, + 212 + ], + "spans": [ + { + "bbox": [ + 254, + 114, + 361, + 212 + ], + "type": "image", + "image_path": "e5f826bc121e1fe340ddeb5183888139cd1573fe3a93abf6e57e39ba39db0628.jpg" + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 288, + 216, + 325, + 225 + ], + "lines": [ + { + "bbox": [ + 288, + 216, + 325, + 225 + ], + "spans": [ + { + "bbox": [ + 288, + 216, + 325, + 225 + ], + "type": "text", + "content": "(b) Top-3" + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "image_caption" + } + ], + "index": 4 + }, + { + "type": "image", + "bbox": [ + 368, + 114, + 474, + 213 + ], + "blocks": [ + { + "bbox": [ + 368, + 114, + 474, + 213 + ], + "lines": [ + { + "bbox": [ + 368, + 114, + 474, + 213 + ], + "spans": [ + { + "bbox": [ + 368, + 114, + 474, + 213 + ], + "type": "image", + "image_path": "a49caaff41a088cdbda4f8dfc47c5565a5c2d9ba5b00fb28d5430cceb1d6f843.jpg" + } + ] + } + ], + "index": 6, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 403, + 216, + 438, + 225 + ], + "lines": [ + { + "bbox": [ + 403, + 216, + 438, + 225 + ], + "spans": [ + { + "bbox": [ + 403, + 216, + 438, + 225 + ], + "type": "text", + "content": "(c) Top-5" + } + ] + } + ], + "index": 7, + "angle": 0, + "type": "image_caption" + } + ], + "index": 6 + }, + { + "type": "image", + "bbox": [ + 139, + 270, + 249, + 366 + ], + "blocks": [ + { + "bbox": [ + 131, + 235, + 480, + 257 + ], + "lines": [ + { + "bbox": [ + 131, + 235, + 480, + 257 + ], + "spans": [ + { + "bbox": [ + 131, + 235, + 480, + 257 + ], + "type": "text", + "content": "Fig. 4: Top-k accuracy (" + }, + { + "bbox": [ + 131, + 235, + 480, + 257 + ], + "type": "inline_equation", + "content": "k = 1,3,5" + }, + { + "bbox": [ + 131, + 235, + 480, + 257 + ], + "type": "text", + "content": "). ArcFace consistently shows the best performance consistently." + } + ] + } + ], + "index": 8, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 139, + 270, + 249, + 366 + ], + "lines": [ + { + "bbox": [ + 139, + 270, + 249, + 366 + ], + "spans": [ + { + "bbox": [ + 139, + 270, + 249, + 366 + ], + "type": "image", + "image_path": "2a1279156c9c47525e6c7ed96d217e4d0b0ee20df2164fe74b7b171889496cbd.jpg" + } + ] + } + ], + "index": 9, + "angle": 0, + "type": "image_body" + } + ], + "index": 9 + }, + { + "type": "image", + "bbox": [ + 254, + 271, + 361, + 367 + ], + "blocks": [ + { + "bbox": [ + 254, + 271, + 361, + 367 + ], + "lines": [ + { + "bbox": [ + 254, + 271, + 361, + 367 + ], + "spans": [ + { + "bbox": [ + 254, + 271, + 361, + 367 + ], + "type": "image", + "image_path": "b4c81e03b3539375dffe769f4499a77e3fdaf41b6ea5e8e1cc9a14c9298569dd.jpg" + } + ] + } + ], + "index": 10, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 130, + 375, + 480, + 420 + ], + "lines": [ + { + "bbox": [ + 130, + 375, + 480, + 420 + ], + "spans": [ + { + "bbox": [ + 130, + 375, + 480, + 420 + ], + "type": "text", + "content": "Fig. 5: Similarity distributions of (a) CLIP, (b) MegaDescriptor, and (c) our joint-trained model on cats. The horizontal axis represents similarity and the vertical axis represents frequency. The model trained on PetFace shows the most distinct separation between positive and negative samples." + } + ] + } + ], + "index": 12, + "angle": 0, + "type": "image_caption" + } + ], + "index": 10 + }, + { + "type": "image", + "bbox": [ + 368, + 271, + 475, + 367 + ], + "blocks": [ + { + "bbox": [ + 368, + 271, + 475, + 367 + ], + "lines": [ + { + "bbox": [ + 368, + 271, + 475, + 367 + ], + "spans": [ + { + "bbox": [ + 368, + 271, + 475, + 367 + ], + "type": "image", + "image_path": "69aff722169cfbce354425ea5aa91d16cbde62671cd78d45e99f89e52a0c2ca3.jpg" + } + ] + } + ], + "index": 11, + "angle": 0, + "type": "image_body" + } + ], + "index": 11 + }, + { + "bbox": [ + 130, + 444, + 480, + 479 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 444, + 480, + 479 + ], + "spans": [ + { + "bbox": [ + 130, + 444, + 480, + 479 + ], + "type": "text", + "content": "and then uniformly pick the positive and negative pairs. Because CTai, CZoo, and Flickr-dog have a few test identities, we each pick five positive and negative pairs for each test identity for robust evaluations." + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 130, + 480, + 481, + 601 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 480, + 481, + 601 + ], + "spans": [ + { + "bbox": [ + 130, + 480, + 481, + 601 + ], + "type": "text", + "content": "We show the verification results in Tables 4(a) and (b). We exclude the results trained on the same dataset as the test set denoted in gray for fair comparisons. For the Chimpanzee datasets, the model trained on our dataset outperforms previous datasets. For case tested on CTai, our dataset outperforms CZoo (67.33% vs. 66.76%). For the other case tested on CZoo, our dataset also surpasses CTai (71.27% vs. 69.31%). This result demonstrates that our Chimpanzee dataset is more effective and general than the previous datasets. For the Dog datasets, owing to the significant scale of our dataset, the model trained on our dog dataset outperforms not only cross-dataset results but also in-dataset ones denoted in gray. These results support the quality of our PetFace dataset." + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 132, + 620, + 204, + 632 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 620, + 204, + 632 + ], + "spans": [ + { + "bbox": [ + 132, + 620, + 204, + 632 + ], + "type": "text", + "content": "5.4 Analysis" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 131, + 641, + 480, + 665 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 131, + 641, + 480, + 665 + ], + "spans": [ + { + "bbox": [ + 131, + 641, + 480, + 665 + ], + "type": "text", + "content": "Different networks. To examine the impact of different network architectures on the performance on PetFace, we additionally train two backbones," + } + ] + } + ], + "index": 16 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "spans": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "type": "text", + "content": "12" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 166, + 91, + 282, + 101 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 166, + 91, + 282, + 101 + ], + "spans": [ + { + "bbox": [ + 166, + 91, + 282, + 101 + ], + "type": "text", + "content": "R. Shinoda and K. Shiohara" + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 11 + }, + { + "para_blocks": [ + { + "type": "table", + "bbox": [ + 132, + 114, + 291, + 171 + ], + "blocks": [ + { + "bbox": [ + 132, + 114, + 291, + 171 + ], + "lines": [ + { + "bbox": [ + 132, + 114, + 291, + 171 + ], + "spans": [ + { + "bbox": [ + 132, + 114, + 291, + 171 + ], + "type": "table", + "html": "
Train SetTest Set AUC (%)
Database#Ind.#ImagesCTaiCZoo
CTai49285775.8069.31
CZoo17132366.7680.49
Ours-Chimp446267967.3371.27
", + "image_path": "7c287fbebfbfc9bce903783b460204a850d1345dcee4835c4fc977f51c0796f9.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "table_body" + } + ], + "index": 2 + }, + { + "type": "table", + "bbox": [ + 298, + 114, + 471, + 171 + ], + "blocks": [ + { + "bbox": [ + 187, + 173, + 247, + 181 + ], + "lines": [ + { + "bbox": [ + 187, + 173, + 247, + 181 + ], + "spans": [ + { + "bbox": [ + 187, + 173, + 247, + 181 + ], + "type": "text", + "content": "(a) Chimp datasets." + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 298, + 114, + 471, + 171 + ], + "lines": [ + { + "bbox": [ + 298, + 114, + 471, + 171 + ], + "spans": [ + { + "bbox": [ + 298, + 114, + 471, + 171 + ], + "type": "table", + "html": "
Train SetTest Set AUC (%)
Database#Ind.#ImagesDogFaceNetFlickr-dog
DogFaceNet975587998.7285.51
Flickr-dog3026594.9887.77
Ours-Dog4675516834899.6595.54
", + "image_path": "22bfbee7a3c934a731546140d22eaf3e0a7096571792bc65244dab715771aecd.jpg" + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "table_body" + } + ], + "index": 4 + }, + { + "type": "table", + "bbox": [ + 132, + 257, + 481, + 304 + ], + "blocks": [ + { + "bbox": [ + 364, + 173, + 416, + 181 + ], + "lines": [ + { + "bbox": [ + 364, + 173, + 416, + 181 + ], + "spans": [ + { + "bbox": [ + 364, + 173, + 416, + 181 + ], + "type": "text", + "content": "(b) Dog datasets." + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 130, + 188, + 480, + 231 + ], + "lines": [ + { + "bbox": [ + 130, + 188, + 480, + 231 + ], + "spans": [ + { + "bbox": [ + 130, + 188, + 480, + 231 + ], + "type": "text", + "content": "Table 4: Comparisons with previous datasets for Chimp and Dog. The results when the test sets are the same as the training datasets are excluded for fair comparisons. The models trained on our dataset outperform those trained on previous datasets." + } + ] + } + ], + "index": 6, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 132, + 257, + 481, + 304 + ], + "lines": [ + { + "bbox": [ + 132, + 257, + 481, + 304 + ], + "spans": [ + { + "bbox": [ + 132, + 257, + 481, + 304 + ], + "type": "table", + "html": "
MethodAUC (%)Avg
CatChimpChinchillaDegusDogFerretGuinea pigHamsterHedgehogParakeetJava sparrowPigRabbit
ResNet-5097.7183.7687.7087.6199.4586.2496.0390.3286.1398.6384.4589.8899.0191.30
ViT-32-B94.4478.6071.1382.6993.3968.1783.2177.8987.7795.0780.6275.3195.1883.35
Swin-B87.3574.2376.2184.0485.7571.6091.5688.3386.2294.1284.6666.5690.5683.17
", + "image_path": "1afcd088dbb284886443328389b94404fd3e72310a052090eb0c7d0833b29ff2.jpg" + } + ] + } + ], + "index": 7, + "angle": 0, + "type": "table_body" + } + ], + "index": 7 + }, + { + "bbox": [ + 130, + 304, + 479, + 325 + ], + "lines": [ + { + "bbox": [ + 130, + 304, + 479, + 325 + ], + "spans": [ + { + "bbox": [ + 130, + 304, + 479, + 325 + ], + "type": "text", + "content": "Table 5: Verification result on different backbones. ResNet-50 achieves the best average AUC." + } + ] + } + ], + "index": 8, + "angle": 0, + "type": "text" + }, + { + "bbox": [ + 130, + 362, + 480, + 411 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 362, + 480, + 411 + ], + "spans": [ + { + "bbox": [ + 130, + 362, + 480, + 411 + ], + "type": "text", + "content": "VisionTransformer-32-B [17] and SwinTransformer-B [41], with the same training configuration as our base model (ResNet-50) and report the verification results in Table 5. We observe that ResNet-50 outperforms the additional transformer-based models in most cases and achieves the best average AUC of " + }, + { + "bbox": [ + 130, + 362, + 480, + 411 + ], + "type": "inline_equation", + "content": "91.30\\%" + }, + { + "bbox": [ + 130, + 362, + 480, + 411 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 130, + 412, + 481, + 628 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 412, + 481, + 628 + ], + "spans": [ + { + "bbox": [ + 130, + 412, + 481, + 628 + ], + "type": "text", + "content": "Fine-grained verification. In real world scenarios, animal verification is sometimes conducted within a specific breed rather than different breeds. To meet this demand, we introduce fine-grained verification where positive and negative pairs are chosen only from the same breeds. We refer to the top-10 dog breeds of individuals in our datasets: Chihuahua, Dachshund, French bulldog, Golden retriever, Miniature dachshund, Pomeranian, Shiba inu, Shih tzu, Toy poodle, and Yorkshire terrier. We construct the fine-grained test sets and evaluate the models on each breed. The results are visualized in Fig. 6. It can be seen that our model achieves a good performance on all the breeds (98.30% on average). The results achieved for the baseline models are lower in comparison (84.99%, 80.78%, and 86.25% for ImageNet, CLIP, and MegaDescriptor, respectively). Moreover, we find that the results achieved with MegaDescriptor show significantly more degradation than the cross-breed verification result in Table 3 from 93.75% to 86.25%. while our model maintains the results at a high-level (from 99.01% to 98.39%). This result indicates that our dataset overcomes the limitation of the previous dog dataset [69] that suffers from the poor performance for specific breed verification owing to its insufficient number of individuals (192 individuals)." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 130, + 629, + 480, + 665 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 629, + 480, + 665 + ], + "spans": [ + { + "bbox": [ + 130, + 629, + 480, + 665 + ], + "type": "text", + "content": "Generality to Unseen Families. Lastly, we evaluate the verification generality to unseen animal families. To achieve this, we additionally collect 100 identities each of Parrot, Lacertilia, and Squirrel, which are not contained in either our" + } + ] + } + ], + "index": 11 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 148, + 91, + 447, + 102 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 148, + 91, + 447, + 102 + ], + "spans": [ + { + "bbox": [ + 148, + 91, + 447, + 102 + ], + "type": "text", + "content": "PetFace: A Large-Scale Dataset and Benchmark for Animal Identification" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 470, + 91, + 480, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 470, + 91, + 480, + 100 + ], + "spans": [ + { + "bbox": [ + 470, + 91, + 480, + 100 + ], + "type": "text", + "content": "13" + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 12 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 137, + 119, + 306, + 258 + ], + "blocks": [ + { + "bbox": [ + 137, + 119, + 306, + 258 + ], + "lines": [ + { + "bbox": [ + 137, + 119, + 306, + 258 + ], + "spans": [ + { + "bbox": [ + 137, + 119, + 306, + 258 + ], + "type": "image", + "image_path": "195502cd50d0b78702522deedc82f87355ae467472e0daed69ce4edba877a1d2.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 310, + 265, + 476, + 308 + ], + "lines": [ + { + "bbox": [ + 310, + 265, + 476, + 308 + ], + "spans": [ + { + "bbox": [ + 310, + 265, + 476, + 308 + ], + "type": "text", + "content": "Fig.7: Generality to unseen animal families. Trained on our PetFace achieves the best performance even for unseen animal families." + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "image_caption" + } + ], + "index": 2 + }, + { + "type": "image", + "bbox": [ + 313, + 117, + 472, + 255 + ], + "blocks": [ + { + "bbox": [ + 313, + 117, + 472, + 255 + ], + "lines": [ + { + "bbox": [ + 313, + 117, + 472, + 255 + ], + "spans": [ + { + "bbox": [ + 313, + 117, + 472, + 255 + ], + "type": "image", + "image_path": "412009a568fb4dc11aaa8acf98199e26604771e45bd9ebc126e625874b61dd96.jpg" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 133, + 267, + 309, + 301 + ], + "lines": [ + { + "bbox": [ + 133, + 267, + 309, + 301 + ], + "spans": [ + { + "bbox": [ + 133, + 267, + 309, + 301 + ], + "type": "text", + "content": "Fig. 6: Fine-grained dog verification. Trained on our PetFace Dog achieves the best performance consistently." + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "image_caption" + } + ], + "index": 3 + }, + { + "bbox": [ + 130, + 333, + 482, + 404 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 333, + 482, + 404 + ], + "spans": [ + { + "bbox": [ + 130, + 333, + 482, + 404 + ], + "type": "text", + "content": "dataset or WildLifeDataset [61]. We compare our model that is jointly trained on our dataset across families with the baseline models in Fig. 7. Our model outperforms the baselines on all the species. For Parrot, particularly, our model achieves an AUC of " + }, + { + "bbox": [ + 130, + 333, + 482, + 404 + ], + "type": "inline_equation", + "content": "88.99\\%" + }, + { + "bbox": [ + 130, + 333, + 482, + 404 + ], + "type": "text", + "content": ", revealing its superiority over MegaDescritor with a large margin. This result supports the generality and variety of our PetFace dataset." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 132, + 424, + 220, + 436 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 424, + 220, + 436 + ], + "spans": [ + { + "bbox": [ + 132, + 424, + 220, + 436 + ], + "type": "text", + "content": "6 Conclusion" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 130, + 449, + 482, + 544 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 449, + 482, + 544 + ], + "spans": [ + { + "bbox": [ + 130, + 449, + 482, + 544 + ], + "type": "text", + "content": "We introduced the PetFace, a comprehensive animal identification dataset encompassing 13 families, featuring 257,484 unique individuals across 319 breed categories. We have collected images and their information and conducted automated and manual filtering to ensure the quality of the large-scale dataset. This dataset also includes detailed annotations of sex, breed, colors, and patterns to facilitate more investigation and application in a real-world scenario. We establish two main benchmarks: 1) re-identification of seen individuals and 2) verification of unseen individuals." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 130, + 544, + 482, + 628 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 544, + 482, + 628 + ], + "spans": [ + { + "bbox": [ + 130, + 544, + 482, + 628 + ], + "type": "text", + "content": "Our experiments show the generality of the models trained on our dataset to verification on other datasets or unseen animal families. We also found that there is still room for improvement in the integrated identification of multiple animal families. To promote further research, we will make this dataset, experiment code, and models available to the research communities. This dataset will enable computer vision researchers to tackle animal face identification across a wide range of breeds and push forward progress on animal face identification tasks." + } + ] + } + ], + "index": 9 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "spans": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "type": "text", + "content": "14" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 166, + 91, + 282, + 101 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 166, + 91, + 282, + 101 + ], + "spans": [ + { + "bbox": [ + 166, + 91, + 282, + 101 + ], + "type": "text", + "content": "R. Shinoda and K. Shiohara" + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 13 + }, + { + "para_blocks": [ + { + "bbox": [ + 133, + 114, + 246, + 129 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 114, + 246, + 129 + ], + "spans": [ + { + "bbox": [ + 133, + 114, + 246, + 129 + ], + "type": "text", + "content": "Acknowledgements" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 132, + 140, + 482, + 190 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 140, + 482, + 190 + ], + "spans": [ + { + "bbox": [ + 132, + 140, + 482, + 190 + ], + "type": "text", + "content": "This work was partially supported by the Cooperative Research Program of the Primate Research Institute, Kyoto University. We thank Dr. Takuma Yagi from AIST for comments on paper proofreading. We also thank Prof. Toshihiko Yamasaki from the University of Tokyo for providing computation resources." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 132, + 209, + 198, + 220 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 209, + 198, + 220 + ], + "spans": [ + { + "bbox": [ + 132, + 209, + 198, + 220 + ], + "type": "text", + "content": "References" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 134, + 234, + 480, + 665 + ], + "type": "list", + "angle": 0, + "index": 19, + "blocks": [ + { + "bbox": [ + 138, + 234, + 459, + 246 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 234, + 459, + 246 + ], + "spans": [ + { + "bbox": [ + 138, + 234, + 459, + 246 + ], + "type": "text", + "content": "1. Beluga id 2022 (2022), https://lila.science/datasets/beluga-id-2022/" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 138, + 247, + 480, + 267 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 247, + 480, + 267 + ], + "spans": [ + { + "bbox": [ + 138, + 247, + 480, + 267 + ], + "type": "text", + "content": "2. Hyena id 2022 botswana predator conservation trust (2022). panthera pardus csv custom export. (2022), https://lila.science/datasets/hyena-id-2022/" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 138, + 269, + 480, + 289 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 269, + 480, + 289 + ], + "spans": [ + { + "bbox": [ + 138, + 269, + 480, + 289 + ], + "type": "text", + "content": "3. Leopard id 2022 botswana predator conservation trust (2022). panthera pardus csv custom export. (2022), https://lila.science/datasets/leopard-id-2022/" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 138, + 291, + 480, + 312 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 291, + 480, + 312 + ], + "spans": [ + { + "bbox": [ + 138, + 291, + 480, + 312 + ], + "type": "text", + "content": "4. Turtle recall: Conservation challenge (2022), https://zindi.africa/competitions/turtle-recall-conservation-challenge" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 138, + 313, + 480, + 344 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 313, + 480, + 344 + ], + "spans": [ + { + "bbox": [ + 138, + 313, + 480, + 344 + ], + "type": "text", + "content": "5. Adam, L., Čermák, V., Papafitsoros, K., Picek, L.: Seaturtleid2022: A long-span dataset for reliable sea turtle re-identification. In: IEEE/CVF Winter Conference on Applications of Computer Vision (WACV). pp. 7146-7156 (2024)" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 138, + 346, + 480, + 378 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 346, + 480, + 378 + ], + "spans": [ + { + "bbox": [ + 138, + 346, + 480, + 378 + ], + "type": "text", + "content": "6. Andrew, W., Gao, J., Mullan, S., Campbell, N., Dowsey, A.W., Burghardt, T.: Visual identification of individual holstein-friesian cattle via deep metric learning. In: Computers and Electronics in Agriculture. vol. 185, p. 106133 (2021)" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 138, + 379, + 480, + 411 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 379, + 480, + 411 + ], + "spans": [ + { + "bbox": [ + 138, + 379, + 480, + 411 + ], + "type": "text", + "content": "7. Andrew, W., Greatwood, C., Burghardt, T.: Visual localisation and individual identification of holstein friesian cattle via deep learning. In: IEEE/CVF International Conference on Computer Vision (ICCV) Workshops. pp. 2850-2859 (2017)" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 138, + 413, + 480, + 444 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 413, + 480, + 444 + ], + "spans": [ + { + "bbox": [ + 138, + 413, + 480, + 444 + ], + "type": "text", + "content": "8. Andrew, W., Greatwood, C., Burghardt, T.: Visual localisation and individual identification of holstein friesian cattle via deep learning. In: IEEE/CVF International Conference on Computer Vision (ICCV) Workshops. pp. 2850-2859 (2017)" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 138, + 445, + 480, + 487 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 445, + 480, + 487 + ], + "spans": [ + { + "bbox": [ + 138, + 445, + 480, + 487 + ], + "type": "text", + "content": "9. Andrew, W., Hannuna, S., Campbell, N., Burghardt, T.: Automatic individual holstein friesian cattle identification via selective local coat pattern matching in rgb-d imagery. In: International Conference on Image Processing (ICIP). pp. 484-488 (2016)" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 135, + 489, + 480, + 521 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 135, + 489, + 480, + 521 + ], + "spans": [ + { + "bbox": [ + 135, + 489, + 480, + 521 + ], + "type": "text", + "content": "10. Brookes, O., Burghardt, T.: A dataset and application for facial recognition of individual gorillas in zoo environments. In: International Conference on Pattern Recognition(ICPR) (2020)" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 134, + 522, + 480, + 576 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 134, + 522, + 480, + 576 + ], + "spans": [ + { + "bbox": [ + 134, + 522, + 480, + 576 + ], + "type": "text", + "content": "11. Cheeseman, T., Southerland, K., Park, J., Olio, M., Flynn, K., Calambokidis, J., Jones, L., Garrigue, C., Frisch-Jordán, A., Howard, A., Reade, W., Neilson, J., Gabriele, C., Clapham, P.: Advanced image recognition: a fully automated, high-accuracy photo-identification matching system for humpback whales. vol. 102, p. 915–929 (2021)" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 134, + 577, + 480, + 598 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 134, + 577, + 480, + 598 + ], + "spans": [ + { + "bbox": [ + 134, + 577, + 480, + 598 + ], + "type": "text", + "content": "12. Christin B. Khan, Shashank, W.K.: Right whale recognition (2015), https://kaggle.com/competitions/noaa-right-whale-recognition" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 134, + 600, + 480, + 632 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 134, + 600, + 480, + 632 + ], + "spans": [ + { + "bbox": [ + 134, + 600, + 480, + 632 + ], + "type": "text", + "content": "13. Dahlborn, K., Bugnon, P., Nevalainen, T., Raspa, M., Verbost, P.M., Spangenberg, E.M.F.: Report of the federation of european laboratory animal science associations working group on animal identification. vol. 47, pp. 2-11 (2013)" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 134, + 633, + 480, + 665 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 134, + 633, + 480, + 665 + ], + "spans": [ + { + "bbox": [ + 134, + 633, + 480, + 665 + ], + "type": "text", + "content": "14. Deng, J., Dong, W., Socher, R., Li, L.J., Li, K., Fei-Fei, L.: Imagenet: A large-scale hierarchical image database. In: IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR). pp. 248-255 (2009)" + } + ] + } + ], + "index": 18 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 148, + 91, + 447, + 102 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 148, + 91, + 447, + 102 + ], + "spans": [ + { + "bbox": [ + 148, + 91, + 447, + 102 + ], + "type": "text", + "content": "PetFace: A Large-Scale Dataset and Benchmark for Animal Identification" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 470, + 91, + 481, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 470, + 91, + 481, + 100 + ], + "spans": [ + { + "bbox": [ + 470, + 91, + 481, + 100 + ], + "type": "text", + "content": "15" + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 14 + }, + { + "para_blocks": [ + { + "bbox": [ + 132, + 117, + 480, + 665 + ], + "type": "list", + "angle": 0, + "index": 19, + "blocks": [ + { + "bbox": [ + 133, + 117, + 480, + 149 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 117, + 480, + 149 + ], + "spans": [ + { + "bbox": [ + 133, + 117, + 480, + 149 + ], + "type": "text", + "content": "15. Deng, J., Guo, J., Xue, N., Zafeiriou, S.: Arcface: Additive angular margin loss for deep face recognition. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR). pp. 4690-4699 (2019)" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 133, + 150, + 480, + 183 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 150, + 480, + 183 + ], + "spans": [ + { + "bbox": [ + 133, + 150, + 480, + 183 + ], + "type": "text", + "content": "16. Dlamini, N., Zyl, T.L.v.: Automated identification of individuals in wildlife population using siamese neural networks. In: 2020 7th International Conference on Soft Computing Machine Intelligence (ISCMI). pp. 224-228 (2020)" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 133, + 183, + 480, + 226 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 183, + 480, + 226 + ], + "spans": [ + { + "bbox": [ + 133, + 183, + 480, + 226 + ], + "type": "text", + "content": "17. Dosovitskiy, A., Beyer, L., Kolesnikov, A., Weissenborn, D., Zhai, X., Unterthiner, T., Dehghani, M., Minderer, M., Heigold, G., Gelly, S., Uszkoreit, J., Houlsby, N.: An image is worth 16x16 words: Transformers for image recognition at scale. The International Conference on Learning Representations (ICLR) (2021)" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 133, + 227, + 480, + 259 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 227, + 480, + 259 + ], + "spans": [ + { + "bbox": [ + 133, + 227, + 480, + 259 + ], + "type": "text", + "content": "18. Ferreira, A.C., Silva, L.R., Renna, F., Brandl, H.B., Renoult, J., Farine, D.R., Covas, R., Doutrelant, C.: Deep learning-based methods for individual recognition in small birds. In: Methods in Ecology and Evolution. vol. 11, pp. 1072-1085 (2020)" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 133, + 259, + 480, + 304 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 259, + 480, + 304 + ], + "spans": [ + { + "bbox": [ + 133, + 259, + 480, + 304 + ], + "type": "text", + "content": "19. Freytag, A., Rodner, E., Simon, M., Loos, A., Kuhl, H.S., Denzler, J.: Chimpanzee faces in the wild: Log-euclidean cnns for predicting identities and attributes of primates. In: Rosenhahn, B., Andres, B. (eds.) Pattern Recognition. pp. 51-63. Springer International Publishing, Cham (2016)" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 132, + 304, + 480, + 336 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 304, + 480, + 336 + ], + "spans": [ + { + "bbox": [ + 132, + 304, + 480, + 336 + ], + "type": "text", + "content": "20. Gao, J., Burghardt, T., Andrew, W., Dowsey, A.W., Campbell, N.W.: Towards self-supervision for video identification of individual holstein-friesian cattle: The cows2021 dataset. arXiv preprint arXiv:2105.01938 (2021)" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 132, + 336, + 480, + 380 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 336, + 480, + 380 + ], + "spans": [ + { + "bbox": [ + 132, + 336, + 480, + 380 + ], + "type": "text", + "content": "21. Guo, Y., Zhang, L., Hu, Y., He, X., Gao, J.: Ms-celeb-1m: A dataset and benchmark for large-scale face recognition. In: Leibe, B., Matas, J., Sebe, N., Welling, M. (eds.) European Conference on Computer Vision (ECCV). pp. 87-102. Springer International Publishing, Cham (2016)" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 132, + 380, + 480, + 413 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 380, + 480, + 413 + ], + "spans": [ + { + "bbox": [ + 132, + 380, + 480, + 413 + ], + "type": "text", + "content": "22. Haurum, J.B., Karpova, A., Pedersen, M., Bengtson, S.H., Moeslund, T.B.: Identification of zebrafish using metric learning. In: IEEE Winter Conference on Applications of Computer Vision (WACV) Workshops. pp. 1-11 (2020)" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 132, + 413, + 480, + 446 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 413, + 480, + 446 + ], + "spans": [ + { + "bbox": [ + 132, + 413, + 480, + 446 + ], + "type": "text", + "content": "23. He, K., Zhang, X., Ren, S., Sun, J.: Deep residual learning for image recognition. In: IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR). pp. 770-778 (2016)" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 132, + 446, + 480, + 479 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 446, + 480, + 479 + ], + "spans": [ + { + "bbox": [ + 132, + 446, + 480, + 479 + ], + "type": "text", + "content": "24. He, Z., Qian, J., Yan, D., Wang, C., Xin, Y.: Animal re-identification algorithm for posture diversity. In: ICASSP 2023 - 2023 IEEE International Conference on Acoustics, Speech and Signal Processing (ICASSP). pp. 1-5 (2023)" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 132, + 479, + 480, + 501 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 479, + 480, + 501 + ], + "spans": [ + { + "bbox": [ + 132, + 479, + 480, + 501 + ], + "type": "text", + "content": "25. Hitelman, A., Edan, Y., Godo, A., Berenstein, R., Lepar, J., Halachmi, I.: Biometric identification of sheep via a machine-vision system. vol. 194, p. 106713 (2022)" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 132, + 501, + 480, + 512 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 501, + 480, + 512 + ], + "spans": [ + { + "bbox": [ + 132, + 501, + 480, + 512 + ], + "type": "text", + "content": "26. Hoffer, E., Ailon, N.: Deep metric learning using triplet network. In: arXiv (2018)" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 132, + 512, + 480, + 544 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 512, + 480, + 544 + ], + "spans": [ + { + "bbox": [ + 132, + 512, + 480, + 544 + ], + "type": "text", + "content": "27. Holmberg, J., Norman, B., Arzoumanian, Z.: Estimating population size, structure, and residency time for whale sharks rhincodon typus through collaborative photo-identification. vol. 7, pp. 39 - 53. Inter-Research Science Center (2009)" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 132, + 545, + 480, + 578 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 545, + 480, + 578 + ], + "spans": [ + { + "bbox": [ + 132, + 545, + 480, + 578 + ], + "type": "text", + "content": "28. Howard, A., inversion, Southerland, K., Cheeseman, T.: Humpback whale identification challenge (2018), https://kaggle.com/competitions/whale-categorization-playground" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 132, + 578, + 480, + 610 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 578, + 480, + 610 + ], + "spans": [ + { + "bbox": [ + 132, + 578, + 480, + 610 + ], + "type": "text", + "content": "29. Huang, G.B., Ramesh, M., Berg, T., Learned-Miller, E.: Labeled faces in the wild: A database for studying face recognition in unconstrained environments. No. 07-49 (2007)" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 132, + 611, + 480, + 643 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 611, + 480, + 643 + ], + "spans": [ + { + "bbox": [ + 132, + 611, + 480, + 643 + ], + "type": "text", + "content": "30. Jarraya, I., Ouarda, W., Alimi, A.M.: A preliminary investigation on horses recognition using facial texture features. In: IEEE International Conference on Systems, Man, and Cybernetics. pp. 2803-2808 (2015)" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 132, + 643, + 480, + 665 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 643, + 480, + 665 + ], + "spans": [ + { + "bbox": [ + 132, + 643, + 480, + 665 + ], + "type": "text", + "content": "31. Khan, F.F., Li, X., Temple, A.J., Elhoseiny, M.: Fishnet: A large-scale dataset and benchmark for fish recognition, detection, and functional trait prediction. In: Pro" + } + ] + } + ], + "index": 18 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "spans": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "type": "text", + "content": "16" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 166, + 91, + 282, + 101 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 166, + 91, + 282, + 101 + ], + "spans": [ + { + "bbox": [ + 166, + 91, + 282, + 101 + ], + "type": "text", + "content": "R. Shinoda and K. Shiohara" + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 15 + }, + { + "para_blocks": [ + { + "bbox": [ + 132, + 116, + 481, + 665 + ], + "type": "list", + "angle": 0, + "index": 19, + "blocks": [ + { + "bbox": [ + 149, + 116, + 480, + 138 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 149, + 116, + 480, + 138 + ], + "spans": [ + { + "bbox": [ + 149, + 116, + 480, + 138 + ], + "type": "text", + "content": "ceedings of the IEEE/CVF International Conference on Computer Vision (ICCV). pp. 20496-20506 (October 2023)" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 133, + 138, + 481, + 182 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 138, + 481, + 182 + ], + "spans": [ + { + "bbox": [ + 133, + 138, + 481, + 182 + ], + "type": "text", + "content": "32. Khan, M.H., McDonagh, J., Khan, S., Shahabuddin, M., Arora, A., Khan, F.S., Shao, L., Tzimiropoulos, G.: Animalweb: A large-scale hierarchical dataset of annotated animal faces. In: IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR). pp. 6937-6946 (2020)" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 132, + 182, + 480, + 203 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 182, + 480, + 203 + ], + "spans": [ + { + "bbox": [ + 132, + 182, + 480, + 203 + ], + "type": "text", + "content": "33. Kim, M., Jain, A.K., Liu, X.: Adaface: Quality adaptive margin for face recognition. pp. 18729-18738 (2022)" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 132, + 203, + 480, + 224 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 203, + 480, + 224 + ], + "spans": [ + { + "bbox": [ + 132, + 203, + 480, + 224 + ], + "type": "text", + "content": "34. Krishnan, S., S., S.: Cloud iot based novel livestock monitoring and identification system using uid. In: Sensor Review. vol. 38, pp. 21-33 (2017)" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 132, + 224, + 480, + 256 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 224, + 480, + 256 + ], + "spans": [ + { + "bbox": [ + 132, + 224, + 480, + 256 + ], + "type": "text", + "content": "35. Kuzdeuov, A., Koishigarina, D., Varol, H.A.: Anyface: A data-centric approach for input-agnostic face detection. In: IEEE International Conference on Big Data and Smart Computing (BigComp). pp. 211-218 (2023)" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 132, + 256, + 480, + 300 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 256, + 480, + 300 + ], + "spans": [ + { + "bbox": [ + 132, + 256, + 480, + 300 + ], + "type": "text", + "content": "36. Lahiri, M., Tantipathananandh, C., Warungu, R., Rubenstein, D.I., Berger-Wolf, T.Y.: Biometric animal databases from field photographs: identification of individual zebra in the wild. In: Proceedings of the ACM International Conference on Multimedia Retrieval. Association for Computing Machinery (2011)" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 132, + 300, + 480, + 342 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 300, + 480, + 342 + ], + "spans": [ + { + "bbox": [ + 132, + 300, + 480, + 342 + ], + "type": "text", + "content": "37. Leslie, E., Hernández-Jover, M., Newman, R., Holyoake, P.: Assessment of acute pain experienced by piglets from ear tagging, ear notching and intraperitoneal injectable transponders. In: Applied Animal Behaviour Science. vol. 127, pp. 86-95 (2010)" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 132, + 342, + 480, + 386 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 342, + 480, + 386 + ], + "spans": [ + { + "bbox": [ + 132, + 342, + 480, + 386 + ], + "type": "text", + "content": "38. Li, S., Li, J., Tang, H., Qian, R., Lin, W.: Atrw: A benchmark for amur tiger re-identification in the wild. In: Proceedings of the 28th ACM International Conference on Multimedia. p. 2590–2598. MM '20, Association for Computing Machinery, New York, NY, USA (2020)" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 132, + 386, + 480, + 418 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 386, + 480, + 418 + ], + "spans": [ + { + "bbox": [ + 132, + 386, + 480, + 418 + ], + "type": "text", + "content": "39. Li, X., Xiang, Y., Li, S.: Combining convolutional and vision transformer structures for sheep face recognition. In: Computers and Electronics in Agriculture. vol. 205, p. 107651 (2023)" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 132, + 418, + 480, + 451 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 418, + 480, + 451 + ], + "spans": [ + { + "bbox": [ + 132, + 418, + 480, + 451 + ], + "type": "text", + "content": "40. Liu, W., Wen, Y., Yu, Z., Li, M., Raj, B., Song, L.: Spherface: Deep hypersphere embedding for face recognition. In: IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR). pp. 6738-6746 (2017)" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 132, + 450, + 480, + 483 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 450, + 480, + 483 + ], + "spans": [ + { + "bbox": [ + 132, + 450, + 480, + 483 + ], + "type": "text", + "content": "41. Liu, Z., Lin, Y., Cao, Y., Hu, H., Wei, Y., Zhang, Z., Lin, S., Guo, B.: Swin transformer: Hierarchical vision transformer using shifted windows. In: IEEE/CVF International Conference on Computer Vision (ICCV). pp. 9992-10002" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 132, + 483, + 480, + 515 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 483, + 480, + 515 + ], + "spans": [ + { + "bbox": [ + 132, + 483, + 480, + 515 + ], + "type": "text", + "content": "42. Loos, A., Ernst, A.: An automated chimpanzee identification system using face detection and recognition. In: EURASIP Journal on Image and Video Processing. vol. 2013 (2013)" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 132, + 515, + 480, + 548 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 515, + 480, + 548 + ], + "spans": [ + { + "bbox": [ + 132, + 515, + 480, + 548 + ], + "type": "text", + "content": "43. Marsot, M., Mei, J., Shan, X., Ye, L., Feng, P., Yan, X., Li, C., Zhao, Y.: An adaptive pig face recognition approach using convolutional neural networks. In: Computers and Electronics in Agriculture. vol. 173, p. 105386 (2020)" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 132, + 548, + 480, + 568 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 548, + 480, + 568 + ], + "spans": [ + { + "bbox": [ + 132, + 548, + 480, + 568 + ], + "type": "text", + "content": "44. Martvel, G., Shimshoni, I., Zamansky, A.: Automated detection of cat facial landmarks. In: International Journal of Computer Vision (IJCV). pp. 1-16 (2023)" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 132, + 568, + 480, + 601 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 568, + 480, + 601 + ], + "spans": [ + { + "bbox": [ + 132, + 568, + 480, + 601 + ], + "type": "text", + "content": "45. Mazlan, N.H., Lopez Salesansky, N., Burn, C., Wells, D.: Mouse identification methods and potential welfare issues: A survey of current practice in the uk. In: Animal Technology and Welfare. vol. 13, pp. 1-10 (2014)" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 132, + 601, + 480, + 633 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 601, + 480, + 633 + ], + "spans": [ + { + "bbox": [ + 132, + 601, + 480, + 633 + ], + "type": "text", + "content": "46. Miele, V., Dussert, G., Spataro, B., Chamaillé-Jammes, S., Allainé, D., Bonenfant, C.: Revisiting animal photo-identification using deep metric learning and network analysis. In: Methods in Ecology and Evolution. vol. 12, pp. 863-873. Wiley (2021)" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 132, + 633, + 480, + 665 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 633, + 480, + 665 + ], + "spans": [ + { + "bbox": [ + 132, + 633, + 480, + 665 + ], + "type": "text", + "content": "47. Moreira, T., Perez, M., Werneck, R., Valle, E.: Where is my puppy? retrieving lost dogs by facial features. In: Multimedia Tools and Applications. vol. 76, p. 15325-15340 (2017)" + } + ] + } + ], + "index": 18 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 148, + 91, + 447, + 102 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 148, + 91, + 447, + 102 + ], + "spans": [ + { + "bbox": [ + 148, + 91, + 447, + 102 + ], + "type": "text", + "content": "PetFace: A Large-Scale Dataset and Benchmark for Animal Identification" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 470, + 91, + 481, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 470, + 91, + 481, + 100 + ], + "spans": [ + { + "bbox": [ + 470, + 91, + 481, + 100 + ], + "type": "text", + "content": "17" + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 16 + }, + { + "para_blocks": [ + { + "bbox": [ + 132, + 116, + 481, + 666 + ], + "type": "list", + "angle": 0, + "index": 20, + "blocks": [ + { + "bbox": [ + 133, + 116, + 481, + 160 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 116, + 481, + 160 + ], + "spans": [ + { + "bbox": [ + 133, + 116, + 481, + 160 + ], + "type": "text", + "content": "48. Moschoglou, S., Papaioannou, A., Sagonas, C., Deng, J., Kotsia, I., Zafeiriou, S.: Agedb: The first manually collected, in-the-wild age database. In: IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR) Workshops. pp. 1997-2005 (2017)" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 132, + 160, + 481, + 192 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 160, + 481, + 192 + ], + "spans": [ + { + "bbox": [ + 132, + 160, + 481, + 192 + ], + "type": "text", + "content": "49. Mougeot, G., Li, D., Jia, S.: A deep learning approach for dog face verification and recognition. In: Nayak, A.C., Sharma, A. (eds.) PRICAI 2019: Trends in Artificial Intelligence. pp. 418-430. Springer International Publishing, Cham (2019)" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 132, + 192, + 481, + 214 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 192, + 481, + 214 + ], + "spans": [ + { + "bbox": [ + 132, + 192, + 481, + 214 + ], + "type": "text", + "content": "50. Murali, N., Taylor, G., Levine, J.: Can drosophila melanogaster tell who's who? In: PLOS ONE. vol. 13, p. e0205043 (2018)" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 132, + 214, + 481, + 246 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 214, + 481, + 246 + ], + "spans": [ + { + "bbox": [ + 132, + 214, + 481, + 246 + ], + "type": "text", + "content": "51. Nepovinnykh, E., Eerola, T., Biard, V., Mutka, P., Niemi, M., Kunnasranta, M., Kalviainen, H.: Sealid: Saimaa ringed seal re-identification dataset. Sensors 22(19) (2022)" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 132, + 246, + 481, + 278 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 246, + 481, + 278 + ], + "spans": [ + { + "bbox": [ + 132, + 246, + 481, + 278 + ], + "type": "text", + "content": "52. Papafitsoros, K., Adam, L., Čermák, V., Picek, L.: Seaturtleid: A novel long-span dataset highlighting the importance of timestamps in wildlife re-identification. In: arXiv (2023)" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 132, + 278, + 481, + 311 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 278, + 481, + 311 + ], + "spans": [ + { + "bbox": [ + 132, + 278, + 481, + 311 + ], + "type": "text", + "content": "53. Parham, J., Crall, J., Stewart, C., Berger-Wolf, T., Rubenstein, D.: Animal population censusing at scale with citizen science and photographic identification. In: SS-17-01. pp. 37-44. AAAI Spring Symposium (2017)" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 132, + 311, + 481, + 343 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 311, + 481, + 343 + ], + "spans": [ + { + "bbox": [ + 132, + 311, + 481, + 343 + ], + "type": "text", + "content": "54. Radford, A., Kim, J.W., Hallacy, C., Ramesh, A., Goh, G., Agarwal, S., Sastry, G., Askell, A., Mishkin, P., Clark, J., Krueger, G., Sutskever, I.: Learning transferable visual models from natural language supervision. In: arXiv (2021)" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 132, + 343, + 481, + 365 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 343, + 481, + 365 + ], + "spans": [ + { + "bbox": [ + 132, + 343, + 481, + 365 + ], + "type": "text", + "content": "55. Roberts, C.: Radio frequency identification (rfid). In: Computers Security. vol. 25, pp. 18-26 (2006)" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 132, + 365, + 481, + 396 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 365, + 481, + 396 + ], + "spans": [ + { + "bbox": [ + 132, + 365, + 481, + 396 + ], + "type": "text", + "content": "56. Roughan, J., Sevenoaks, T.: Welfare and scientific considerations of tattooing and ear-tagging for mouse identification. In: Journal of the American Association for Laboratory Animal Science: JAALAS. pp. 142-153 (2018)" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 132, + 396, + 481, + 429 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 396, + 481, + 429 + ], + "spans": [ + { + "bbox": [ + 132, + 396, + 481, + 429 + ], + "type": "text", + "content": "57. Sengupta, S., Chen, J.C., Castillo, C., Patel, V.M., Chellappa, R., Jacobs, D.W.: Frontal to profile face verification in the wild. In: IEEE Winter Conference on Applications of Computer Vision (WACV). pp. 1-9 (2016)" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 132, + 429, + 481, + 460 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 429, + 481, + 460 + ], + "spans": [ + { + "bbox": [ + 132, + 429, + 481, + 460 + ], + "type": "text", + "content": "58. Shi, X., Yang, C., Xia, X., Chai, X.: Deep cross-species feature learning for animal face recognition via residual interspecies equivariant network. In: European Conference on Computer Vision (ECCV). p. 667-682 (2020)" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 132, + 460, + 481, + 482 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 460, + 481, + 482 + ], + "spans": [ + { + "bbox": [ + 132, + 460, + 481, + 482 + ], + "type": "text", + "content": "59. Shinoda, R., Shiohara, K.: Openanimaltracks: A dataset for animal track recognition (2024)" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 132, + 482, + 481, + 514 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 482, + 481, + 514 + ], + "spans": [ + { + "bbox": [ + 132, + 482, + 481, + 514 + ], + "type": "text", + "content": "60. Trotter, C., Atkinson, G., Sharpe, M., Richardson, K., McGough, A.S., Wright, N., Burville, B., Berggren, P.: Ndd20: A large-scale few-shot dolphin dataset for coarse and fine-grained categorisation. In: arXiv (2020)" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 132, + 514, + 481, + 547 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 514, + 481, + 547 + ], + "spans": [ + { + "bbox": [ + 132, + 514, + 481, + 547 + ], + "type": "text", + "content": "61. Čermák, V., Picek, L., Adam, L., Papafitsoros, K.: Wildlifedatasets: An open-source toolkit for animal re-identification. In: IEEE/CVF Winter Conference on Applications of Computer Vision (WACV). pp. 5953-5963 (2024)" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 132, + 547, + 481, + 579 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 547, + 481, + 579 + ], + "spans": [ + { + "bbox": [ + 132, + 547, + 481, + 579 + ], + "type": "text", + "content": "62. Wahltinez, O., Wahltinez, S.: An open-source general purpose machine learning framework for individual animal re-identification using few-shot learning. In: Methods in Ecology and Evolution. vol. 15, pp. 373-387 (2024)" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 132, + 579, + 481, + 611 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 579, + 481, + 611 + ], + "spans": [ + { + "bbox": [ + 132, + 579, + 481, + 611 + ], + "type": "text", + "content": "63. Wang, H., Wang, Y., Zhou, Z., Ji, X., Gong, D., Zhou, J., Li, Z., Liu, W.: Cosface: Large margin cosine loss for deep face recognition. In: IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR). pp. 5265-5274 (2018)" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 132, + 611, + 481, + 632 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 611, + 481, + 632 + ], + "spans": [ + { + "bbox": [ + 132, + 611, + 481, + 632 + ], + "type": "text", + "content": "64. Wang, L., Ding, R., Zhai, Y., Zhang, Q., Tang, W., Zheng, N., Hua, G.: Giant panda identification. IEEE Transactions on Image Processing 30, 2837-2849 (2021)" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 132, + 632, + 481, + 666 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 632, + 481, + 666 + ], + "spans": [ + { + "bbox": [ + 132, + 632, + 481, + 666 + ], + "type": "text", + "content": "65. Wen, Y., Zhang, K., Li, Z., Qiao, Y.: A discriminative feature learning approach for deep face recognition. In: Leibe, B., Matas, J., Sebe, N., Welling, M. (eds.) European Conference on Computer Vision (ECCV). pp. 499-515 (2016)" + } + ] + } + ], + "index": 19 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "spans": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "type": "text", + "content": "18" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 166, + 91, + 282, + 101 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 166, + 91, + 282, + 101 + ], + "spans": [ + { + "bbox": [ + 166, + 91, + 282, + 101 + ], + "type": "text", + "content": "R. Shinoda and K. Shiohara" + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 17 + }, + { + "para_blocks": [ + { + "bbox": [ + 132, + 116, + 482, + 249 + ], + "type": "list", + "angle": 0, + "index": 6, + "blocks": [ + { + "bbox": [ + 132, + 116, + 482, + 138 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 116, + 482, + 138 + ], + "spans": [ + { + "bbox": [ + 132, + 116, + 482, + 138 + ], + "type": "text", + "content": "66. Witham, C.L.: Automated face recognition of rhesus macaques. In: Journal of Neuroscience Methods. vol. 300, pp. 157-165 (2018)" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 132, + 139, + 482, + 182 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 139, + 482, + 182 + ], + "spans": [ + { + "bbox": [ + 132, + 139, + 482, + 182 + ], + "type": "text", + "content": "67. Zhu, Z., Huang, G., Deng, J., Ye, Y., Huang, J., Chen, X., Zhu, J., Yang, T., Lu, J., Du, D., Zhou, J.: Webface260m: A benchmark unveiling the power of million-scale deep face recognition. In: IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR). pp. 10492-10502 (2021)" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 132, + 183, + 481, + 216 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 183, + 481, + 216 + ], + "spans": [ + { + "bbox": [ + 132, + 183, + 481, + 216 + ], + "type": "text", + "content": "68. Zuerl, M., Dirauf, R., Koeferl, F., Steinlein, N., Sueskind, J., Zanca, D., Brehm, I., Fersen, L.v., Eskofier, B.: Polarbearvidid: A video-based re-identification benchmark dataset for polar bears. In: Animals. vol. 13 (2023)" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 132, + 216, + 481, + 249 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 216, + 481, + 249 + ], + "spans": [ + { + "bbox": [ + 132, + 216, + 481, + 249 + ], + "type": "text", + "content": "69. Zuffi, S., Kanazawa, A., Berger-Wolf, T., Black, M.J.: Three-d safari: Learning to estimate zebra pose, shape, and texture from images \"in the wild\". In: IEEE/CVF International Conference on Computer Vision (ICCV). pp. 5358-5367 (2019)" + } + ] + } + ], + "index": 5 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 148, + 91, + 447, + 102 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 148, + 91, + 447, + 102 + ], + "spans": [ + { + "bbox": [ + 148, + 91, + 447, + 102 + ], + "type": "text", + "content": "PetFace: A Large-Scale Dataset and Benchmark for Animal Identification" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 470, + 91, + 481, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 470, + 91, + 481, + 100 + ], + "spans": [ + { + "bbox": [ + 470, + 91, + 481, + 100 + ], + "type": "text", + "content": "19" + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 18 + } + ], + "_backend": "vlm", + "_version_name": "2.6.4" +} \ No newline at end of file diff --git a/2024/Phase Concentration and Shortcut Suppression for Weakly Supervised Semantic Segmentation/8ce12d89-9c27-46eb-bd86-41483594c43e_content_list.json b/2024/Phase Concentration and Shortcut Suppression for Weakly Supervised Semantic Segmentation/8ce12d89-9c27-46eb-bd86-41483594c43e_content_list.json new file mode 100644 index 0000000000000000000000000000000000000000..6275be3cdb6591995af68e6a7c0a62dcbb928445 --- /dev/null +++ b/2024/Phase Concentration and Shortcut Suppression for Weakly Supervised Semantic Segmentation/8ce12d89-9c27-46eb-bd86-41483594c43e_content_list.json @@ -0,0 +1,1887 @@ +[ + { + "type": "text", + "text": "Phase Concentration and Shortcut Suppression for Weakly Supervised Semantic Segmentation", + "text_level": 1, + "bbox": [ + 232, + 140, + 772, + 186 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Hoyong Kwon, Jaeseok Jeong, Sung-Hoon Yoon, and Kuk-Jin Yoon", + "bbox": [ + 233, + 212, + 767, + 228 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "KAIST {kwonhoyong3, jason.jeong,yoon307,kjyoon}@kaist.ac.kr", + "bbox": [ + 299, + 239, + 702, + 268 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Abstract. Weakly Supervised Semantic Segmentation (WSSS) with image-level supervision typically acquires object localization information from Class Activation Maps (CAMs). While Vision Transformers (ViTs) in WSSS have been increasingly explored for their superior performance in understanding global context, CAMs from ViT still show imprecise localization in boundary areas and false positive activation. This paper proposes a novel WSSS framework that targets these issues based on the information from the frequency domain. In our framework, we introduce the Magnitude-mixing-based Phase Concentration (MPC) module, which guides the classifier to prioritize phase information containing high-level semantic details. By perturbing and mixing the magnitude, MPC guides the classifier to accentuate and concentrate on the shape information in the phase, thereby leading to finer distinctions in CAMs boundary regions. Additionally, inspired by empirical observations that the classification \"shortcut\" in the frequency domain can induce false positives in CAMs, we introduce a Frequency Shortcut Suppression (FSS) module. This module aims to discourage the formation of such shortcuts, thereby mitigating false positives. The effectiveness of our approach is demonstrated by achieving new state-of-the-art performance on both PASCAL VOC 2012 and MS COCO 2014 datasets. The code is available at https://github.com/kwonhoyong3/PCSS-WSSS.", + "bbox": [ + 259, + 301, + 759, + 593 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Keywords: Weakly Supervised Semantic Segmentation $\\cdot$ Fourier Transform $\\cdot$ Shortcut Learning", + "bbox": [ + 259, + 604, + 740, + 633 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "1 Introduction", + "text_level": 1, + "bbox": [ + 215, + 657, + 375, + 674 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Weakly Supervised Semantic Segmentation (WSSS) aims to enhance its utility by reducing the dependency on human-annotated pixel-level labels in semantic segmentation. With this goal, WSSS mainly explores methods to generate accurate pseudo-labels using only easily obtainable weak labels such as image-level labels [1-3,25,26,29,51,60,68,69], scribbles [37,56], and bounding boxes [8,30,43]. Among these methods, leveraging image-level labels has been extensively investigated, primarily due to the availability of large existing datasets.", + "bbox": [ + 212, + 688, + 787, + 794 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Since image-level labels do not provide spatial information of objects, image-level based WSSS typically involves multiple stages: 1) extract Class Activation Maps (CAMs) to localize the objects, 2) refine CAMs to generate pseudo", + "bbox": [ + 212, + 795, + 787, + 840 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/65cd16063131316a619f9010443bac3b5880b73cd85d2a8e0b275ee06695f42d.jpg", + "image_caption": [ + "(a) Image" + ], + "image_footnote": [], + "bbox": [ + 217, + 143, + 330, + 231 + ], + "page_idx": 1 + }, + { + "type": "image", + "img_path": "images/62e0460bd3241426ade0e6b23fa700c754c555d98809bc34264fd2b646d0a901.jpg", + "image_caption": [ + "(b) Freq. MaskBird" + ], + "image_footnote": [], + "bbox": [ + 331, + 143, + 442, + 231 + ], + "page_idx": 1 + }, + { + "type": "image", + "img_path": "images/5bcad55ed0d73d1c44790b8765f1d2b9e15fcbe8c5bbab80f3f5d41426b03f87.jpg", + "image_caption": [ + "(c) Filtered Image", + "Fig. 1: Experimental results verifying the relationship between CAMs and Frequency Shortcut. The Filtered Image (c) was obtained by removing the Principal Frequency components (b) related to the class of the Image (a). (d) and (e) are CAMs of images (a) and (c), respectively, derived from the same ViT trained on clean images." + ], + "image_footnote": [], + "bbox": [ + 444, + 145, + 555, + 231 + ], + "page_idx": 1 + }, + { + "type": "image", + "img_path": "images/c54e4234d4e86271e2e923e46a5aac5424f519644f519bf3a0cbe5d63e1e8de6.jpg", + "image_caption": [ + "(d) CAM" + ], + "image_footnote": [], + "bbox": [ + 558, + 145, + 671, + 231 + ], + "page_idx": 1 + }, + { + "type": "image", + "img_path": "images/36b2cb78ba651ca7d5e8e6119fc18cf31b0870b71ec92e652c641f6da6d8e1b9.jpg", + "image_caption": [ + "(e) CAM Filtered" + ], + "image_footnote": [], + "bbox": [ + 671, + 145, + 784, + 231 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "ground truth (Pseudo-GT), and 3) train a semantic segmentation model using the Pseudo-GT. As the second and third stages are dependent on the CAMs from the first stage, our work targets to improve the quality of CAMs. In the first stage, unlike Convolutional Neural Networks (CNNs) that focus on discriminative regions, Vision Transformer (ViT) [9] stands out in WSSS for its ability to understand global context through the Multi-Head Self-Attention (MHSA) mechanism [55].", + "bbox": [ + 212, + 340, + 787, + 446 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Despite these advantages over CNN-based CAMs, CAMs from ViT still face several issues; 1) CAMs exhibit imprecise object boundaries due to lack of spatial supervision, and 2) CAMs often localize object-unrelated (false positive) regions. In this paper, we propose two methodologies to address the issues with CAMs.", + "bbox": [ + 212, + 446, + 787, + 508 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "When humans and ViTs visually assess an object, semantics is predominantly influenced by shape over texture [4, 24]. While shape and texture information are intertwined in images, applying Discrete Fourier Transform (DFT) [54] can help extract the two information into magnitude and phase spectral components. Inspired by the fact that the phase spectrum encapsulates shape and high-level semantic information [16, 42], we propose Magnitude-mixing-based Phase Concentration (MPC) to encourage models to focus more on the phase and improve object boundaries in generated CAMs. In MPC, we select a random pair of images consisting of an anchor image and another arbitrary image. Within the intermediate layers of a ViT, we merge the magnitude information from the arbitrary image into the magnitude from the anchor image. This results in the creation of two anchor image features with identical phases but different magnitudes. By ensuring that both features are guided by the same weak supervision and that the generated CAMs are consistent regardless of changes in magnitude, we induce the model to focus on the phase. Consequently, MPC enables the model to concentrate on the crucial information within the phase, leading to CAMs with more precise boundaries.", + "bbox": [ + 212, + 508, + 787, + 763 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "While MPC facilitates precise activation near the boundaries, the issue of activation in object-unrelated regions persists. Interestingly, certain frequency components have been shown to significantly impact classification during the training process of ViT, acting as Frequency Shortcut [58]. A frequency shortcut is a simplicity bias in classification networks, where the network relies on a", + "bbox": [ + 212, + 763, + 787, + 839 + ], + "page_idx": 1 + }, + { + "type": "page_number", + "text": "2", + "bbox": [ + 217, + 114, + 228, + 126 + ], + "page_idx": 1 + }, + { + "type": "header", + "text": "H. Kwon et al.", + "bbox": [ + 271, + 114, + 370, + 127 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "specific frequency rather than semantic content for classification. Additionally, we empirically observed that models tend to exhibit false positives in the CAMs of classes attempting to find shortcuts in the frequency domain. To explore the effects of principal frequency and shortcuts in CAMs, we compare the CAMs from a clean image and a corresponding image with the principal frequency removed (i.e., filtered image), as shown in Fig. 1. While the CAMs from clean image (Fig. 1(d)) suffers from false activations, false activations are relieved in the CAMs from filtered image (Fig. 1(e)). Here, we can verify that certain principal frequencies influence the entire image and cause widespread false activation regardless of the semantics; these can be regarded as shortcut frequencies.", + "bbox": [ + 212, + 146, + 787, + 297 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "To resolve the false positives induced by frequency shortcuts, we propose the Frequency Shortcut Suppression (FSS). Initially, we measured the impact of each frequency on the classification of each class to aggregate a Frequency Shortcut Potential Map (FSPM). Based on FSPM, we removed the frequency shortcut components from the image to generate the filtered image. Through classification on the filtered image and by comparing CAMS generated from the original image and the filtered image, FSS reduces the reliance of the generated CAMs to false positive inducing shortcut frequency components. This removal of shortcut frequency components happens iteratively, with the FSPM constantly adjusted.", + "bbox": [ + 212, + 297, + 787, + 446 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "The main contributions of this paper are summarized as follows:", + "bbox": [ + 240, + 449, + 702, + 462 + ], + "page_idx": 2 + }, + { + "type": "list", + "sub_type": "text", + "list_items": [ + "- Through Magnitude-mixing-based Phase Concentration (MPC), we guide ViT to focus on the meaningful phase information, resulting in precise activation around boundaries.", + "- We proposed a Frequency Shortcut Suppression (FSS) module that effectively reduces false positive activations in CAMs by suppressing shortcut learning in the Frequency Domain." + ], + "bbox": [ + 215, + 474, + 784, + 563 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "To demonstrate the effectiveness of our methodology, we conducted comparisons with state-of-the-art WSSS methods on the PASCAL VOC 2012 [11] and MS COCO 2014 datasets [38], achieving new SoTA performance.", + "bbox": [ + 212, + 575, + 787, + 621 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "2 Related Works", + "text_level": 1, + "bbox": [ + 215, + 645, + 395, + 659 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "2.1 Weakly Supervised Semantic Segmentation", + "text_level": 1, + "bbox": [ + 215, + 676, + 616, + 693 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Most existing WSSS approaches utilize Class Activation Maps (CAMs) [72] obtained from Convolutional Neural Networks (CNNs) to localize objects when only image-level labels are provided. However, raw CAMs (i.e. seeds) fail to localize the whole object with imprecise boundaries. In pursuit of generating more accurate pixel-level pseudo-labels, research in WSSS aims to propose a method to improve the quality of raw CAMs or further refine/post-processing the CAMs. CAMs Quality Improvement To guide CAMs to localize the less discriminative regions, various methods are proposed including online CAMs aggregation [21], sub-categories exploration [3], information bottleneck reduction [27]", + "bbox": [ + 212, + 703, + 787, + 842 + ], + "page_idx": 2 + }, + { + "type": "header", + "text": "Phase Concentration and Shortcut Suppression for WSSS", + "bbox": [ + 346, + 114, + 730, + 128 + ], + "page_idx": 2 + }, + { + "type": "page_number", + "text": "3", + "bbox": [ + 774, + 116, + 784, + 126 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "and mining cross-image relations [12, 34, 52]. Also, [22, 59, 70] introduce various auxiliary tasks to regularize CAMs training. Along with these methods, Adversarial Erasing (AE) [25, 33, 53, 67, 71] methods, which are based on the erase-and-find mechanism, are also actively researched to expand CAMs. Apart from techniques aimed at expanding CAMs, contrastive learning-based WSSS [6, 62, 73] and adversarial learning between classifier-reconstructor [26] are proposed to produce CAMs with precise object boundaries. Though CNN-based WSSS methods show promising performance, with the success of Vision Transformer (ViT) in various vision tasks, several methods [5, 13, 39, 44, 48, 49, 63, 64, 69] are proposed to migrate the ViT for WSSS. MCTformer [63] introduces multi-class token to extract class-specific attention map from ViT by improving the TS-CAM [13], which is designed for weakly-supervised object localization. In the same line, ViT-PCM [47] proposed a method to extract CAMs based on patch-class mapping. Along with these methods, large language model-based WSSS [39, 61, 64] and token-level contrasting [49] are also introduced. Unlike the prior methods, our work is the first to utilize frequency domain information to address the issues with CAMs in ViT-based WSSS.", + "bbox": [ + 212, + 145, + 787, + 402 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "CAMs Post-processing To obtain high-quality pseudo labels for semantic segmentation, various approaches are introduced in WSSS based on affinity learning [1, 2], removing intra-object edges in contours [32], iterative refinements [35, 36]. Besides these methods, Image-matting based refinements [57] and utilizing unsupervised semantic segmentation model [23] greatly boost the quality of pseudo labels. As the post-processing methods can be synergistically applied with methods that generate precise initial seeds (CAMs), this paper focuses on improving the quality of the CAMs.", + "bbox": [ + 212, + 402, + 787, + 525 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "2.2 Fourier Transform in Computer Vision", + "text_level": 1, + "bbox": [ + 214, + 544, + 583, + 559 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Transforming images from the spatial domain to the frequency domain via the Fourier Transform enables the separation of information into phase and amplitude components. Early studies [16,42] have shown that while amplitude carries low-level statistics and texture information, phase encapsulates high-level semantic and shape information. Leveraging this advantage, Fourier Transform has been applied to a variety of vision tasks such as Exposure Correction [20], Semantic Segmentation [18], and Learning with noisy label [17]. Along with the successful application of frequency components, the fields of Domain Adaptation [19, 66] and Generalization [15, 31, 65] have emphasized the importance of phase for its high-level semantic and structural information. FDA [66] performed Domain Adaptation by transforming low-level amplitude in the Fourier Domain, maintaining high-level semantics and structure, building on the understanding that low-level amplitude contains image style information. [65] highlighted the significance of the phase component, which carries semantic information, and proposed Fourier-based data augmentation through Image-level amplitude to perform Domain Generalization with an emphasis on phase information.", + "bbox": [ + 212, + 568, + 787, + 809 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "These studies underscore the significant role of the frequency domain in neural networks. Intriguingly, classification networks engage in shortcut learning", + "bbox": [ + 212, + 809, + 785, + 840 + ], + "page_idx": 3 + }, + { + "type": "page_number", + "text": "4", + "bbox": [ + 217, + 114, + 228, + 126 + ], + "page_idx": 3 + }, + { + "type": "header", + "text": "H. Kwon et al.", + "bbox": [ + 271, + 114, + 370, + 127 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "in the frequency domain [58], identifying specific frequencies for classification. Since shortcut learning [14] can disrupt semantic capture, research efforts have been directed towards reducing such effects [40, 46]. SGT [41] analyzed how vanilla ViT activates areas unrelated to the object due to shortcut learning in the background regions during the classification process and proposed rectifying this through saliency maps.", + "bbox": [ + 212, + 146, + 782, + 238 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "3 Methods", + "text_level": 1, + "bbox": [ + 215, + 256, + 336, + 272 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "3.1 Preliminary", + "text_level": 1, + "bbox": [ + 215, + 285, + 359, + 300 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Vision Transformer for CAMs To propagate the image $\\mathbf{I}$ into Vision Transformer (ViT), it is split into $N\\times N$ patches and forms the patch tokens $\\mathbf{T}_{patch}^{0}\\in \\mathbb{R}^{D\\times N^2}$ , where $D$ is the embedding dimension. Along with $C$ class tokens $\\mathbf{T}_{cls}^{0}\\in \\mathbb{R}^{D\\times C}$ , the patch tokens $\\mathbf{T}_{patch}^{0}$ pass through the Vision Transformer (ViT). Consequently, the final class tokens $\\mathbf{T}_{cls}^{L}\\in \\mathbb{R}^{D\\times C}$ and patch tokens $\\mathbf{T}_{patch}^{L}\\in \\mathbb{R}^{D\\times N^2}$ are derived as outputs, where $L$ represents the number of ViT layers. This procedure can be formally defined as follows:", + "bbox": [ + 212, + 306, + 784, + 421 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\n\\mathbf {T} _ {c l s} ^ {i}, \\mathbf {T} _ {p a t c h} ^ {i} = \\mathbf {V} \\mathrm {i} \\mathbf {T} _ {i} \\left(\\mathbf {T} _ {c l s} ^ {i - 1}, \\mathbf {T} _ {p a t c h} ^ {i - 1}\\right), 1 \\leq i \\leq L \\tag {1}\n$$\n", + "text_format": "latex", + "bbox": [ + 349, + 428, + 784, + 446 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "where $\\mathbf{V}\\mathbf{i}\\mathbf{T}_i$ is $i^{th}$ layer of ViT. The class tokens $\\mathbf{T}_{cls}^L$ are processed through average pooling to derive the classification logits $y_{c} \\in \\mathbb{R}^{C}$ . Meanwhile, the patch tokens $\\mathbf{T}_{patch}^L$ , arranged according to their spatial positions, pass through a convolution layer to produce the Class Activation Maps (CAMs) $\\mathbf{M} \\in \\mathbb{R}^{C \\times N \\times N}$ . By applying Global Average Pooling (GAP) to $\\mathbf{M}$ , we can obtain classification logits $y_{p} \\in \\mathbb{R}^{C}$ in a patch-level. These two classification logits are supervised using a multi-label soft margin loss $\\mathcal{L}_{cls}$ with the given classification labels $\\mathbf{y}$ . Additionally, following [63], we extract the token-to-token attention $\\mathbf{A}_{t2t} \\in \\mathbb{R}^{(C + N^2) \\times (C + N^2)}$ based on multi-head self-attention. From this token-to-token attention $\\mathbf{A}_{t2t}$ , the class-to-patch attention $\\mathbf{A}_{c2p} \\in \\mathbb{R}^{C \\times N^2}$ and patch-to-patch attention $\\mathbf{A}_{p2p} \\in \\mathbb{R}^{N^2 \\times N^2}$ can be extracted. These two types of attention maps are used to refine the CAMs, forming a refined CAMs $\\mathbf{M}_{ref} \\in \\mathbb{R}^{C \\times N \\times N}$ . The aforementioned process can be noted as follows:", + "bbox": [ + 212, + 454, + 787, + 657 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\n\\mathbf {M} _ {r e f} = \\left(\\mathbf {M} \\odot \\mathbf {A} _ {c 2 p}\\right) \\times \\mathbf {A} _ {p 2 p} \\tag {2}\n$$\n", + "text_format": "latex", + "bbox": [ + 406, + 667, + 784, + 683 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "where $\\odot$ and $\\times$ are Hadamard product and matrix multiplication, respectively. Discrete Fourier Transform Let $\\mathcal{F}$ denote the Discrete Fourier Transform (DFT) [54] applied across the spatial dimensions. For a feature map $\\mathbf{X} \\in \\mathbb{R}^{D \\times H \\times W}$ the result of applying $\\mathcal{F}$ to a single channel can be represented as follows:", + "bbox": [ + 212, + 689, + 782, + 750 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\n\\mathcal {F} (\\mathbf {X}) (m, n) = \\sum_ {h, w} \\mathbf {X} (h, w) e ^ {- j 2 \\pi \\left(\\frac {h}{H} m + \\frac {w}{W} n\\right)}, j ^ {2} = - 1 \\tag {3}\n$$\n", + "text_format": "latex", + "bbox": [ + 325, + 758, + 784, + 785 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Since the absolute value and the angle component of $\\mathcal{F}(\\mathbf{X})$ respectively represent magnitude and phase components, we denote the magnitude spectrum as $|\\mathcal{F}(\\mathbf{X})|$ and the phase spectrum as $\\angle \\mathcal{F}(\\mathbf{X})$ throughout the paper.", + "bbox": [ + 212, + 794, + 782, + 840 + ], + "page_idx": 4 + }, + { + "type": "header", + "text": "Phase Concentration and Shortcut Suppression for WSSS", + "bbox": [ + 346, + 114, + 732, + 128 + ], + "page_idx": 4 + }, + { + "type": "page_number", + "text": "5", + "bbox": [ + 774, + 116, + 784, + 126 + ], + "page_idx": 4 + }, + { + "type": "image", + "img_path": "images/e8121360ec3292f87495d4a534d49c1a53760efad3dfb1475008820b494d8053.jpg", + "image_caption": [ + "Fig. 2: Visualization of the proposed framework. The Magnitude-mixing-based Phase Concentration (MPC) is executed at an intermediate layer. Here, the magnitude of features from the anchor image $\\mathbf{I}$ and another arbitrary image $\\hat{\\mathbf{I}}$ are mixed to make a feature with identical phase, but different magnitude. The model is guided to perform correct classification and derive consistent CAMs $\\mathbf{M}_{ref}$ from a pair of features. In the Frequency Shortcut Suppression (FSS), based on the Frequency Shortcut Potential Map (FSPM), frequency components that have a significant impact on the classification of classes in $\\mathbf{I}$ are reduced before being inputted to the ViT. The model is encouraged to classify correctly and minimize the difference between $\\mathbf{M}$ and $\\hat{\\mathbf{M}}$ ." + ], + "image_footnote": [], + "bbox": [ + 217, + 146, + 787, + 372 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "3.2 Magnitude-mixing-based Phase Concentration", + "text_level": 1, + "bbox": [ + 215, + 541, + 643, + 556 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "With the aid of multi-head self-attention, CAMs from ViT are more proficient in capturing the less-discriminative regions than CNN-based methods. Nonetheless, they still show imprecise localization along object boundaries due to a lack of spatial constraint. To improve localization of CAMs, we design the Magnitude-mixing-based Phase Concentration (MPC) module to extract high-level semantic and boundary information shown to be present in the phase component of the frequency domain [16, 42].", + "bbox": [ + 212, + 566, + 787, + 672 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "We consider two ideas in MPC to guide the model to prioritize the phase component: 1) The model should correctly classify the image even when the magnitude component is altered. Similarly, 2) the CAMs generated after the magnitude change should be similar to those obtained before the change. The key to these two ideas is to change the magnitude while keeping the phase component intact. Since the magnitude needs to be 'feasible' yet different, we designed a magnitude-mixing method. Here, we form a random image pair consisting of an anchor image $\\mathbf{I}$ and an arbitrary image $\\dot{\\mathbf{I}}$ for perturbation. Let the input class and patch tokens of $\\mathbf{I}$ be denoted as $[\\mathbf{T}_{cls}^{0},\\mathbf{T}_{patch}^{0}]$ , and those of $\\dot{\\mathbf{I}}$ as $[\\dot{\\mathbf{T}}_{cls}^{0},\\dot{\\mathbf{T}}_{patch}^{0}]$ . We then process each set of tokens through the ViT layers, as outlined in Eq. 1, for $1\\leq i\\leq \\ell$ , to obtain $[\\mathbf{T}_{cls}^{\\ell},\\mathbf{T}_{patch}^{\\ell}]$ and $[\\dot{\\mathbf{T}}_{cls}^{\\ell},\\dot{\\mathbf{T}}_{patch}^{\\ell}]$ respectively. After reshaping", + "bbox": [ + 212, + 674, + 789, + 843 + ], + "page_idx": 5 + }, + { + "type": "page_number", + "text": "6", + "bbox": [ + 217, + 114, + 228, + 126 + ], + "page_idx": 5 + }, + { + "type": "header", + "text": "H. Kwon et al.", + "bbox": [ + 271, + 114, + 370, + 127 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/923a377c43b81f8725a8592a94044bef4ef0331b6b6b973ed1302bd41ba38967.jpg", + "image_caption": [ + "Fig. 3: Visualization of Frequency Influence Measurement. For an image $\\mathbf{I}$ containing class $c$ , a set of images that each masked at different frequency components using $M_{f}$ in the frequency domain of $\\mathbf{I}$ are inputted into ViT. The classification loss for class $c$ is calculated as the influence of frequency $f$ on class $c$ . Accumulate the process across the image set $\\mathbb{I}_c$ to produce $\\mathbf{P}(c)$ , and repeat for each $c$ . Finally, $\\mathbf{P}$ is normalized to obtain the Frequency Shortcut Potential Map (FSPM) for each class." + ], + "image_footnote": [], + "bbox": [ + 233, + 143, + 767, + 258 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "and aligning $\\mathbf{T}_{patch}^{\\ell}$ and $\\bar{\\mathbf{T}}_{patch}^{\\ell}$ according to their spatial positions, the spectral magnitude is obtained and perturbed using the following formula:", + "bbox": [ + 214, + 378, + 784, + 409 + ], + "page_idx": 6 + }, + { + "type": "equation", + "text": "\n$$\n\\bar {\\mathbf {T}} _ {p a t c h} ^ {\\ell} = \\mathcal {F} ^ {- 1} \\left( \\right.\\left\\langle \\right.\\left\\langle \\right. r \\mid \\mathcal {F} \\left(\\mathbf {T} _ {p a t c h} ^ {\\ell}\\right)\\left. \\right| + (1 - r) \\mid \\mathcal {F} \\left(\\dot {\\mathbf {T}} _ {p a t c h} ^ {\\ell}\\right)\\left. \\right|, \\angle \\mathcal {F} \\left(\\mathbf {T} _ {p a t c h} ^ {\\ell}\\right)\\left. \\right\\rangle \\rangle\\left. \\right) \\tag {4}\n$$\n", + "text_format": "latex", + "bbox": [ + 269, + 415, + 785, + 435 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "where $r$ is the perturbation ratio between the anchor image and arbitrary image, and $\\mathcal{F}^{-1}$ is the inverse DFT. Tokens and maps that have been affected by magnitude-mixing are denoted by a bar on top (i.e. $\\bar{\\mathbf{T}}_{patch}^{\\ell}$ ). $\\langle \\langle A,B\\rangle \\rangle$ is coupling of magnitude $A$ and phase $B$ . As shown in Eq. 4, the phase component of $\\bar{\\mathbf{T}}_{patch}^{\\ell}$ remains unchanged as the magnitude is perturbed. After iDFT, the set of tokens $[\\mathbf{T}_{cls}^{\\ell},\\bar{\\mathbf{T}}_{patch}^{\\ell}]$ are reshaped and passed through the remaining layers of ViT, following Eq. 1 from $\\ell < i\\leq L$ , to produce the output tokens $[\\bar{\\mathbf{T}}_{cls}^{L},\\bar{\\mathbf{T}}_{patch}^{L}]$ . The class prediction $\\bar{y}_c$ is obtained from $\\bar{\\mathbf{T}}_{cls}^{L}$ , while the class prediction $\\bar{y}_p$ and refined CAMs $\\bar{\\mathbf{M}}_{ref}$ are generated from $\\bar{\\mathbf{T}}_{patch}^{L}$ . $\\bar{y}_c$ and $\\bar{y}_p$ are used as class predictions in multi-label soft margin loss to obtain $\\mathcal{L}_{mpc - cls}$ . Following the second idea, the difference between the refined CAMs $\\bar{\\mathbf{M}}_{ref}$ and $\\bar{\\mathbf{M}}_{ref}$ are minimized using the following loss:", + "bbox": [ + 214, + 441, + 787, + 630 + ], + "page_idx": 6 + }, + { + "type": "equation", + "text": "\n$$\n\\mathcal {L} _ {\\text {m p c - c a m s}} = \\left| \\mathbf {M} _ {\\text {r e f}} - \\bar {\\mathbf {M}} _ {\\text {r e f}} \\right| _ {1} \\tag {5}\n$$\n", + "text_format": "latex", + "bbox": [ + 398, + 637, + 784, + 655 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Ultimately, the MPC is supervised through the following loss:", + "bbox": [ + 215, + 661, + 661, + 676 + ], + "page_idx": 6 + }, + { + "type": "equation", + "text": "\n$$\n\\mathcal {L} _ {m p c} = \\mathcal {L} _ {m p c - c l s} + \\lambda_ {1} \\mathcal {L} _ {m p c - c a m s} \\tag {6}\n$$\n", + "text_format": "latex", + "bbox": [ + 390, + 685, + 784, + 700 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "where $\\lambda_{1}$ is a hyperparameter to balance the losses. Additionally, the gradient to $\\mathbf{M}_{ref}$ from $\\mathcal{L}_{mpc-cams}$ is detached to guide $\\bar{\\mathbf{M}}_{ref}$ to follow $\\mathbf{M}_{ref}$ . This increases the dependency of the model on the phase information, enabling the model to better understand the boundary information in the phase component.", + "bbox": [ + 214, + 708, + 784, + 768 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "3.3 Frequency Shortcut Suppression", + "text_level": 1, + "bbox": [ + 215, + 787, + 529, + 803 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Through the MPC module, we have addressed the issues of activation around the boundary in CAMs. However, the problem of false activations in CAMs", + "bbox": [ + 214, + 809, + 785, + 839 + ], + "page_idx": 6 + }, + { + "type": "header", + "text": "Phase Concentration and Shortcut Suppression for WSSS", + "bbox": [ + 346, + 114, + 732, + 128 + ], + "page_idx": 6 + }, + { + "type": "page_number", + "text": "7", + "bbox": [ + 774, + 116, + 784, + 126 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "persists. Various factors contribute to false activations in ViT, such as GAP for generating classification logits from CAMs and the over-smoothing problem caused by self-attention of ViT [49]. Interestingly, we experimentally found that false activations can also arise due to Shortcut Learning in the frequency domain [14], which we aim to address. Shortcut Learning during the classification training process in ViT can negatively impact genuine semantic understanding [40], and such shortcuts can lead to wrong activations [41]. Furthermore, it has been revealed that Neural Networks can create shortcuts in the frequency domain [58]. We empirically found a correlation between frequency shortcuts and false positives in CAMs M. Since Frequency Shortcuts are unrelated to objectness, activations caused by frequency shortcuts occur in locations unrelated to objects. We have resolved over-activation by disrupting the creation of frequency shortcuts through the Frequency Shortcut Suppression(FSS) module.", + "bbox": [ + 212, + 146, + 787, + 343 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Frequency Shortcut Potential Map To implement the FSS, the potential of specific frequency components to be used as shortcuts in classification needs to be measured. Inspired by Wang et al. [58], we utilized a Frequency Influence Measurement (FIM) framework, as depicted in Fig. 3, to assess the impact of each component. If the classification loss of the model increases significantly for an image $\\mathbf{I}$ of class $c$ when information about a certain frequency component $f$ is removed, it indicates that $f$ has an important role in classifying class $c$ for the model. Therefore, for each class $c$ , the influence of $f$ on the classification of the image set $\\mathbb{I}_c$ corresponding to that class is calculated by accumulating classification loss. To remove a specific frequency component $f$ from an image, a Masking map $M_f \\in \\mathbb{R}^{H \\times W}$ is used, where the value for $f$ is set to 0, and others are set to 1. For an image $\\mathbf{I}$ , let $\\mathbf{I}_f = \\mathcal{F}^{-1}(\\mathcal{F}(\\mathbf{I}) \\odot M_f)$ represent the image with the $f$ component masked where $\\odot$ is element-wise multiplication. The classification logit derived after passing $\\mathbf{I}_f$ through a ViT is denoted by $y_{\\mathbf{I}_f}$ . When $\\mathbf{P} \\in \\mathbb{R}^{C \\times H \\times W}$ is frequency influence accumulated map, the influence of the frequency component $f$ on the classification of class $c$ is calculated as follows:", + "bbox": [ + 212, + 343, + 789, + 587 + ], + "page_idx": 7 + }, + { + "type": "equation", + "text": "\n$$\n\\mathbf {P} (c, f) = \\sum_ {\\mathbf {I} \\in \\mathbb {I} _ {c}} \\mathcal {L} _ {b c e} \\left(y _ {\\mathbf {I} _ {f}} (c), 1\\right) \\tag {7}\n$$\n", + "text_format": "latex", + "bbox": [ + 403, + 595, + 787, + 625 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "where $y(c)$ and $\\mathcal{L}_{bce}$ are probability of class $c$ and binary cross-entropy loss, respectively. $\\mathbf{P}(c)\\in \\mathbb{R}^{H\\times W}$ ultimately holds the classification influence information of each frequency for class $c$ , with higher values indicating a significant impact on classifying class $c$ . $\\mathbf{P}$ is normalized for each class to create the Frequency Shortcut Potential Map (FSPM) $\\mathbf{FS}\\in \\mathbb{R}^{C\\times H\\times W}$ :", + "bbox": [ + 212, + 633, + 787, + 710 + ], + "page_idx": 7 + }, + { + "type": "equation", + "text": "\n$$\n\\mathbf {F S} (c) = \\operatorname {N o r m} (\\mathbf {P} (c)), c \\in \\{1, 2, \\dots , C \\} \\tag {8}\n$$\n", + "text_format": "latex", + "bbox": [ + 367, + 723, + 785, + 739 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "where Norm is the normalization function that scales each value to fall between 0 and 1.", + "bbox": [ + 212, + 750, + 787, + 777 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Frequency Shortcut Suppression We propose the Frequency Shortcut Suppression (FSS) utilizing the FSPM generated through the FIM process. The FSPM encapsulates information on which frequency components significantly impact classification at the time of FIM execution, indicating frequencies that", + "bbox": [ + 212, + 779, + 787, + 840 + ], + "page_idx": 7 + }, + { + "type": "page_number", + "text": "8", + "bbox": [ + 217, + 114, + 228, + 126 + ], + "page_idx": 7 + }, + { + "type": "header", + "text": "H. Kwon et al.", + "bbox": [ + 271, + 114, + 370, + 127 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "the model is likely to use as shortcuts. The image without shortcut frequency can be generated by reducing the magnitude of frequency components based on the FSPM. Subsequently, the model can be prevented from forming shortcuts by encouraging it to classify correctly and generate consistent CAMs from the original and filtered images.", + "bbox": [ + 212, + 146, + 782, + 222 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "Let an image $\\mathbf{I}$ contains classes $c\\in \\mathbb{C}_{\\mathbf{I}}$ . The filtered image $\\hat{\\mathbf{I}}$ is synthesized by decreasing the frequency components from $\\mathbf{I}$ based on the FSPM, making it challenging for the model to classify through shortcuts:", + "bbox": [ + 212, + 222, + 784, + 268 + ], + "page_idx": 8 + }, + { + "type": "equation", + "text": "\n$$\n\\hat {\\mathbf {I}} = \\frac {1}{N} \\sum_ {c \\in \\mathbb {C} _ {\\mathbf {I}}} \\mathcal {F} ^ {- 1} \\left(\\mathcal {F} (\\mathbf {I}) \\odot (1 - \\mathbf {F S} (c))\\right) \\tag {9}\n$$\n", + "text_format": "latex", + "bbox": [ + 375, + 280, + 785, + 311 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "where $N$ is the number of classes in $\\mathbb{C}_{\\mathbf{I}}$ , and $\\odot$ denotes the Hadamard product. The predictions $\\hat{y}_c$ and $\\hat{y}_p$ derived from $\\hat{\\mathbf{I}}$ via the ViT are trained through a multi-label soft margin loss $\\mathcal{L}_{fss-cls}$ . Additionally, the CAMs $\\hat{\\mathbf{M}}$ generated from $\\hat{\\mathbf{I}}$ and the CAMs $\\mathbf{M}$ derived from the original image $\\mathbf{I}$ are supervised to minimize their difference:", + "bbox": [ + 214, + 323, + 782, + 400 + ], + "page_idx": 8 + }, + { + "type": "equation", + "text": "\n$$\n\\mathcal {L} _ {f s s - c a m s} = | \\mathbf {M} - \\hat {\\mathbf {M}} | _ {1} \\tag {10}\n$$\n", + "text_format": "latex", + "bbox": [ + 426, + 402, + 784, + 420 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "Assuming the model has learned to use a frequency component $f$ as a shortcut for classifying class $c$ , the model cannot use this shortcut $f$ for classification when $f$ is reduced in $\\hat{\\mathbf{I}}$ . In the absence of the frequency shortcut $f$ , the model performs classification based on the semantic understanding of class $c$ , resulting in activation within areas related to the object. Simultaneously, by reducing the difference between the two CAMs $\\mathbf{M}$ and $\\hat{\\mathbf{M}}$ , false positives caused by shortcuts can be resolved.", + "bbox": [ + 212, + 428, + 784, + 531 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "The FSS is ultimately supervised through the following loss:", + "bbox": [ + 238, + 534, + 673, + 547 + ], + "page_idx": 8 + }, + { + "type": "equation", + "text": "\n$$\n\\mathcal {L} _ {f s s} = \\mathcal {L} _ {f s s - c l s} + \\lambda_ {2} \\mathcal {L} _ {f s s - c a m s} \\tag {11}\n$$\n", + "text_format": "latex", + "bbox": [ + 401, + 563, + 784, + 577 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "where $\\lambda_{2}$ is a hyperparameter to balance between the losses.", + "bbox": [ + 214, + 590, + 648, + 604 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "The final loss of the proposed overall framework is as follows:", + "bbox": [ + 238, + 604, + 681, + 619 + ], + "page_idx": 8 + }, + { + "type": "equation", + "text": "\n$$\n\\mathcal {L} _ {\\text {t o t a l}} = \\mathcal {L} _ {\\text {c l s}} + \\mathcal {L} _ {\\text {m p c}} + \\mathcal {L} _ {\\text {f s s}} \\tag {12}\n$$\n", + "text_format": "latex", + "bbox": [ + 408, + 633, + 784, + 650 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "4 Experiments", + "text_level": 1, + "bbox": [ + 214, + 672, + 375, + 691 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "4.1 Experimental Settings", + "text_level": 1, + "bbox": [ + 214, + 707, + 446, + 722 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "Datasets Our proposed method was evaluated using the PASCAL VOC 2012 and MS-COCO 2014 datasets, widely utilized for Weakly Supervised Semantic Segmentation (WSSS). The PASCAL VOC 2012 includes 20 classes of foreground objects along with a background class, and consists of three subsets (train, val, test) containing 10582, 1449, and 1456 images, respectively. Meanwhile, the MS-COCO 2014 is made up of 80 foreground object classes and a background class and is composed of two subsets (82k train set and 40k val set).", + "bbox": [ + 212, + 733, + 784, + 839 + ], + "page_idx": 8 + }, + { + "type": "header", + "text": "Phase Concentration and Shortcut Suppression for WSSS", + "bbox": [ + 346, + 114, + 732, + 128 + ], + "page_idx": 8 + }, + { + "type": "page_number", + "text": "9", + "bbox": [ + 774, + 116, + 784, + 126 + ], + "page_idx": 8 + }, + { + "type": "table", + "img_path": "images/12fcda77fd3ed751eecd40a15ee33d3744580a68da5bceb8f5620a7015ec5a14.jpg", + "table_caption": [ + "Table 1: Ablation study of components on the PASCAL VOC 2012 train set. P: Precision, R: Recall. Values from M are measured under similar R for a fair comparison of P. The best result is represented in Bold." + ], + "table_footnote": [], + "table_body": "
LclsLfssLmpc(x) M(y) Mrf
P(%)R(%)mIoU(%)P(%)R(%)mIoU(%)
Baseline✓53.279.645.374.580.963.2
(a)✓✓71.379.859.476.582.165.4
(b)✓✓67.180.655.978.882.267.2
(c)✓✓✓76.280.763.979.584.869.5
", + "bbox": [ + 308, + 198, + 702, + 272 + ], + "page_idx": 9 + }, + { + "type": "text", + "text": "Evaluation Metric Following prior research [44,63], the mean Intersection over Union (mIoU) metric is used to assess the quality of generated CAMs and the performance of semantic segmentation models trained on CAMs-based pseudolabels. To measure the quality of CAMs, the mIoU with the most optimal threshold is employed by default. Typically, CAMs are evaluated on the train set, whereas the performance of the semantic segmentation model is assessed on the val set. While ground-truth pixel-level labels are accessible for both datasets on the val set and are analyzed locally, the evaluation for the PASCAL VOC 2012 test set is carried out via the official website.", + "bbox": [ + 212, + 297, + 785, + 434 + ], + "page_idx": 9 + }, + { + "type": "text", + "text": "Implementation Details Our framework utilizes DeiT-S pre-trained on ImageNet [50] as the classification backbone for a fair comparison with ViT-based WSSS methods [44, 63]. The classification network is trained for 60 epochs with a batch size of 64 and an initial learning rate of 5e-4, using the Adam optimizer. We adopted the same data augmentation technique as in MCTformer [63], but with different image resize scales. We crop images to a size of $224 \\times 224$ for equal comparison. In the MPC module, $r = 0.3$ was used as the mixing factor, with MPC applied at ViT layer $\\ell = 8$ . To measure the influence of each frequency for FSPM, 100 images were used per class for all classes. Additionally, the FSPM was updated every three epochs through the FIM process. Min-max normalization was utilized in FIM to normalize $\\mathbf{P}$ , with the highest measured loss per class serving as the normalization maximum. For balancing each loss relative to $\\mathcal{L}_{cls}$ , $\\lambda_{1}$ and $\\lambda_{2}$ were both set to 2. For the semantic segmentation model training on the PASCAL VOC 2012 dataset, Deeplab-V1 with ResNet38 was employed, while Deeplab-V2 with ResNet101 was used for the MS COCO 2014 dataset. Additional training details are in Supplementary Materials.", + "bbox": [ + 212, + 434, + 787, + 676 + ], + "page_idx": 9 + }, + { + "type": "text", + "text": "4.2 Ablation Studies", + "text_level": 1, + "bbox": [ + 214, + 696, + 401, + 709 + ], + "page_idx": 9 + }, + { + "type": "text", + "text": "Component Analysis To demonstrate the significance of each module proposed in our framework, we ablated each component as shown in Tab. 1. Observing the performance of refined CAMs in Tab. 1-y, introducing the Frequency Shortcut Suppression (FSS) and Magnitude-mixing-based Phase Concentration (MPC) led to the gain of $2.2\\% \\mathrm{p}$ and $4.0\\% \\mathrm{p}$ respectively over baseline. Our method achieved $69.5\\%$ mIoU by adopting both components. FSS still resulted in a gain of $2.3\\% \\mathrm{p}$ in Tab. 1-c compared to Tab. 1-b, implies that FSS and MPC offer distinct advantages.", + "bbox": [ + 212, + 719, + 787, + 840 + ], + "page_idx": 9 + }, + { + "type": "page_number", + "text": "10", + "bbox": [ + 217, + 114, + 235, + 126 + ], + "page_idx": 9 + }, + { + "type": "header", + "text": "H. Kwon et al.", + "bbox": [ + 271, + 114, + 370, + 127 + ], + "page_idx": 9 + }, + { + "type": "table", + "img_path": "images/472d285ea044979b436ca3a7ce31a7c9f6c5fb96d1ca5f1304ac8f1e68c50196.jpg", + "table_caption": [ + "Table 2: Ablation study of ViT layer index that MPC performed. Peak performance is observed at layer 8, which is represented in Bold." + ], + "table_footnote": [], + "table_body": "
earlylayermiddlelayerlatelayer
l34781011
mIoU(%)68.068.168.669.567.866.4
", + "bbox": [ + 375, + 185, + 630, + 226 + ], + "page_idx": 10 + }, + { + "type": "image", + "img_path": "images/c52d65ecf433beb16c76c62194ab85df54e55dc62e0a42bcf36bf900b83b12b2.jpg", + "image_caption": [ + "(a) Image" + ], + "image_footnote": [], + "bbox": [ + 220, + 244, + 316, + 340 + ], + "page_idx": 10 + }, + { + "type": "image", + "img_path": "images/ea24099a9a5b5aaa78c2359f0e2f578d7b7038092b8bf68feb47d21df69e93f5.jpg", + "image_caption": [ + "(b) GT" + ], + "image_footnote": [], + "bbox": [ + 316, + 246, + 406, + 340 + ], + "page_idx": 10 + }, + { + "type": "image", + "img_path": "images/5a97f56eeb671efab321c91047875e6a1ae3b2671eb72fe28032b4c42429cd9a.jpg", + "image_caption": [ + "(c) Baseline", + "Fig. 4: Comparison of refined CAMs $\\mathbf{M}_{ref}$ between baseline and components in ours. (a) Image, (b) Ground Truth, (c) Baseline, (d) only FSS, (e) only MPC, (f) ours. The yellow and red boxes imply effect of FSS and MPC, respectively." + ], + "image_footnote": [], + "bbox": [ + 408, + 246, + 500, + 340 + ], + "page_idx": 10 + }, + { + "type": "image", + "img_path": "images/1c957814bf72dc6c15c910f5d2235b07d0778e491f45283960ba3d267c0180d8.jpg", + "image_caption": [ + "(d) FSS" + ], + "image_footnote": [], + "bbox": [ + 501, + 246, + 593, + 340 + ], + "page_idx": 10 + }, + { + "type": "image", + "img_path": "images/cc0e49ab30b8e12bc3346679c4dec22a023e2b10017cfdcad830cfec465f31d5.jpg", + "image_caption": [ + "(e) MPC" + ], + "image_footnote": [], + "bbox": [ + 594, + 246, + 686, + 340 + ], + "page_idx": 10 + }, + { + "type": "image", + "img_path": "images/97dbb2141e54829bfcee6eeed5fa8a80a5ccdda2625b4a38501729d331a8b7c6.jpg", + "image_caption": [ + "(f) FSS+MPC" + ], + "image_footnote": [], + "bbox": [ + 687, + 246, + 779, + 340 + ], + "page_idx": 10 + }, + { + "type": "text", + "text": "To understand the benefits of FSS in CAMs, we examined the results of ViT CAMs that are not refined by the attention-map in Tab. 1-x. To conduct a comparative analysis considering the Precision-Recall tradeoff, thresholds corresponding to similar Recall levels were used. Low precision of baseline $\\mathbf{M}$ indicates a significant number of over-activation. Through FSS, we observed a remarkable improvement in precision $(18.1\\%)$ and mIoU $(14.1\\%)$ compared to the baseline, confirming that false positives were successfully resolved (Tab. 1-b).", + "bbox": [ + 212, + 441, + 784, + 547 + ], + "page_idx": 10 + }, + { + "type": "text", + "text": "Ablation Studies of MPC The qualitative comparison between the refined CAMs of baseline and MPC in Fig. 4-e demonstrates the efficacy of the MPC by clearly delineating the boundary of the object. Meanwhile, we investigate the impact of the ratio $r$ in MPC, which is used to preserve the magnitude information of the anchor feature. In experiments conducted within a search space below 0.5, there was a minimum gain of $2.3\\% p$ when compared to applying FSS alone, with the best performance observed at $r = 0.3$", + "bbox": [ + 212, + 550, + 784, + 655 + ], + "page_idx": 10 + }, + { + "type": "text", + "text": "The influence of the layer index $\\ell$ where the MPC is performed was analyzed by conducting experiments on early, middle, and late layers in Tab. 2. The performance was highest when $\\ell$ was set to 8, with a noticeable decline observed in layers beyond this point. This tendency supports the knowledge that later layers of the ViT play a crucial role in capturing high-level semantic information [9]. Introducing changes to the magnitude while preserving the phase information at these later stages could disrupt the semantic analysis of the model. This disruption may stem from the difference between the coupled information used in previous layers and the re-coupled information with the magnitude alteration. This result shows that MPC applied before when ViT features highly condense to high-level semantics can encourage the model to capture boundary information well in the phase spectrum.", + "bbox": [ + 212, + 657, + 785, + 839 + ], + "page_idx": 10 + }, + { + "type": "header", + "text": "Phase Concentration and Shortcut Suppression for WSSS", + "bbox": [ + 346, + 114, + 732, + 128 + ], + "page_idx": 10 + }, + { + "type": "page_number", + "text": "11", + "bbox": [ + 767, + 114, + 782, + 126 + ], + "page_idx": 10 + }, + { + "type": "table", + "img_path": "images/6d0508f924ec085b4dc5a3916bfb59fc9354a639041c7a6d76bbc0366c0d43f7.jpg", + "table_caption": [ + "Table 3: (a) True/False Positive Rate (TPR/FPR) of classification where the high-frequency components following FSPM are retained in images. Metrics are measured under the early training stage of the Baseline. **Bold** indicates the shortcut occurred, and **Underline** represents that model is trying to make a shortcut. (b) Comparative analysis of the Precision of CAMs M between baseline and only with FSS. Precision is evaluated under a similar Recall considering the Precision-Recall trade-off." + ], + "table_footnote": [], + "table_body": "
bkgaerobikebirdboatbottlebuscarcatchaircow
(a)BaselineTPR(%)-58.9636.4388.4753.3668.4650.8379.2668.3555.4817.87
FPR(%)-00.2300.4115.7800.0533.7400.0001.4200.5307.0400.41
(b)BaselineP(%)91.939.224.817.318.958.176.542.880.944.959.7
w/ FSSP(%)92.552.942.259.233.371.187.562.994.143.290.7
tabledoghorsembkpersonplantsheepsofatraintvtotal
(a)BaselineTPR(%)04.7661.2479.4060.2675.4741.8843.6145.4592.5157.23-
FPR(%)00.0001.9700.6400.1605.5600.1500.3800.6300.6800.13-
(b)BaselineP(%)68.666.050.468.658.039.956.479.039.237.053.2
w/ FSSP(%)89.687.579.778.582.764.792.581.254.556.671.3
", + "bbox": [ + 220, + 241, + 787, + 377 + ], + "page_idx": 11 + }, + { + "type": "image", + "img_path": "images/789d710b9f35af13364ccd140b16cb4fc65804fd456cb4ec390943ddb952c958.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 243, + 395, + 318, + 525 + ], + "page_idx": 11 + }, + { + "type": "image", + "img_path": "images/1349e719e53c3d6dc26311c39edc6a225670db56e187656a8c438f96f007088e.jpg", + "image_caption": [ + "PASCAL VOC 2012" + ], + "image_footnote": [], + "bbox": [ + 320, + 396, + 372, + 525 + ], + "page_idx": 11 + }, + { + "type": "image", + "img_path": "images/213cb39012f09625bbe9c2d815cece94a905d22498d187bd838ea4be36e96bb8.jpg", + "image_caption": [ + "Fig.5: Semantic segmentation results on PASCAL VOC 2012 (left) and MS COCO 2014 (right) datasets. From top to bottom: Image, Ours, GT." + ], + "image_footnote": [], + "bbox": [ + 374, + 396, + 452, + 525 + ], + "page_idx": 11 + }, + { + "type": "image", + "img_path": "images/dfd509dcb24401bf676857c83fdfbb838c1f9ae7fc4652ab7cd67b17be60e035.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 452, + 396, + 535, + 525 + ], + "page_idx": 11 + }, + { + "type": "image", + "img_path": "images/4da045b8515eccf717596a93ae711c5d1c7c0d32d5b7a63bd841aa6d01495b9c.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 542, + 396, + 584, + 525 + ], + "page_idx": 11 + }, + { + "type": "image", + "img_path": "images/1aaad2b1fe3c7f8d416a631027eef01bd95d1a7b5dc1dc036c96f8978e44a7c2.jpg", + "image_caption": [ + "MS COCO 2014" + ], + "image_footnote": [], + "bbox": [ + 584, + 396, + 648, + 525 + ], + "page_idx": 11 + }, + { + "type": "image", + "img_path": "images/21735237b7d0dbff15ba83edc10a170472f82434af30023ff1050f256d09c081.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 647, + 396, + 756, + 525 + ], + "page_idx": 11 + }, + { + "type": "text", + "text": "Frequency Shortcut Analysis and Result of Suppression In Tab. 3, we investigated whether shortcut learning occurs during the model training process in early steps and analyzed the correlation with the improvement in CAMs precision facilitated by the FSS. To conduct this analysis, classification influence according to frequency component for each class was first measured using Eq. 7. Based on this measurement, the classification behavior of the model was analyzed when only some high frequencies of high-influence components were retained.", + "bbox": [ + 212, + 609, + 787, + 715 + ], + "page_idx": 11 + }, + { + "type": "text", + "text": "A high TPR indicates that class $c$ possesses the principal frequency components utilized in the experiment (i.e., bird will be correctly classified as bird only with the high-influence frequency of bird). This suggests that the model has identified a straightforward method for classifying class $c$ and is attempting to find frequency shortcuts. To consider a principal frequency as a shortcut, it should direct images to the respective class regardless of the actual semantic content in the image (i.e. car with frequency shortcut of bird will falsely classify as bird). This is expressed in Tab. 3 through high TPR and FPR, which repre", + "bbox": [ + 212, + 719, + 787, + 840 + ], + "page_idx": 11 + }, + { + "type": "page_number", + "text": "12", + "bbox": [ + 217, + 114, + 235, + 126 + ], + "page_idx": 11 + }, + { + "type": "header", + "text": "H. Kwon et al.", + "bbox": [ + 271, + 114, + 370, + 127 + ], + "page_idx": 11 + }, + { + "type": "table", + "img_path": "images/9eecf686b1f57214a57faaa2c4ba2c925cf03a71603d1af0554645b068d416e3.jpg", + "table_caption": [ + "Table 4: Comparison between ours and multi-stage WSSS methods. mIoU(%) is evaluated on PASCAL VOC 2012 train set in both CAMs (seed) and Pseudo-GT (Mask). The backbone and applied post-processing methods (PSA [2], IRN [1]) are listed for a fair comparison. Bold numbers represent the best performance." + ], + "table_footnote": [], + "table_body": "
MethodBackboneSeedPostMask
AdvCAM [28] CVPR'21RN5055.6IRN69.9
OC-CSE [25] ICCV'21WRN3856.0PSA66.9
ECS [53] ICCV'21WRN3856.6PSA67.8
CPN [70] ICCV'21WRN3857.4PSA67.8
AMR [45] AAAI'22RN5056.8PSA69.7
ReCAM [7] CVPR'22RN5054.8IRN70.5
SIPE [6] CVPR'22RN5058.6IRN-
PPC [10] CVPR'22WRN3861.5IRN70.1
AEFT [67] ECCV'22WRN3856.0PSA71.0
ACR [26] CVPR'23WRN3860.3IRN72.3
MCT [63] CVPR'22DeiT-S61.7PSA69.1
USAGE [44] ICCV'23DeiT-S67.7PSA72.8
OursDeiT-S69.5PSA73.2
", + "bbox": [ + 333, + 212, + 674, + 386 + ], + "page_idx": 12 + }, + { + "type": "text", + "text": "sent the rates in which principal frequencies of images are correctly and falsely classified to the shortcut respective class $c$ . Examining Tab. 3-a, we can observe that class bird and bottle exhibit high TPR and FPR, indicating the occurrence of shortcuts. Meanwhile, classes such as car, horse, and train show high TPR, suggesting attempts at creating shortcuts.", + "bbox": [ + 212, + 426, + 787, + 501 + ], + "page_idx": 12 + }, + { + "type": "text", + "text": "Tab. 3-b records the class-specific precision of CAMs $\\mathbf{M}$ for the baseline model and after the introduction of the FSS. To ensure a fair comparison while considering the Precision-Recall trade-off, precision was measured at similar levels of Recall. The introduction of FSS to disrupt shortcut creation resulted in an average increase of $24.0\\%$ in precision for $\\mathbf{M}$ across classes where the baseline model had created or attempted to create shortcuts. This increase confirms that the FSS effectively reduces false positives. Notably, for bird, a class where shortcuts had occurred, the precision of CAMs surged by $42.1\\%$ , showcasing the resolve of over-activation in areas unrelated to the object by FSS. Qualitative comparison of refined CAMs $\\mathbf{M}_{ref}$ before and after the introduction of FSS in Fig. 4-d further validates the effectiveness of FSS.", + "bbox": [ + 212, + 502, + 787, + 667 + ], + "page_idx": 12 + }, + { + "type": "text", + "text": "4.3 Comparison with State-of-The-Arts", + "text_level": 1, + "bbox": [ + 214, + 691, + 555, + 708 + ], + "page_idx": 12 + }, + { + "type": "text", + "text": "To train the semantic segmentation model, the CAMs generated by our method were refined using PSA [2] as performed in previous WSSS works [44, 63, 67]. In Tab. 4, we compared the performance of CAMs (Seed) and pseudo-ground-truth (Mask) on the PASCAL VOC 2012 train set, where our method achieved the best performance in both categories. Utilizing our high-quality pseudo labels for training the semantic segmentation model, we compared the results with other WSSS methods in Tab. 5. Our framework demonstrated superiority by achieving the highest performance on the PASCAL VOC 2012 val and test sets.", + "bbox": [ + 212, + 719, + 787, + 840 + ], + "page_idx": 12 + }, + { + "type": "header", + "text": "Phase Concentration and Shortcut Suppression for WSSS", + "bbox": [ + 346, + 114, + 730, + 128 + ], + "page_idx": 12 + }, + { + "type": "page_number", + "text": "13", + "bbox": [ + 767, + 114, + 785, + 126 + ], + "page_idx": 12 + }, + { + "type": "table", + "img_path": "images/755cbf765516c5c987ac0238169adf5a0f5daedead7f03bd3a0a05d0b4aa9e82.jpg", + "table_caption": [ + "Table 5: Comparison of semantic segmentation performance with multi-stage WSSS methods. Performance (mIoU) is evaluated in PASCAL VOC 2012 and MS COCO 2014 datasets. CNN-based methods are represented above the horizontal line, and ViT-based methods are depicted below the line. Bold represents the best performance." + ], + "table_footnote": [], + "table_body": "
MethodBackboneVOC valVOC testCOCO val
AdvCAM [28] CVPR'21RN10168.168.0-
OC-CSE [25] ICCV'21WRN3868.468.236.4
ECS [53] ICCV'21WRN3866.667.6-
CPN [70] ICCV'21WRN3867.868.5-
AMR [45] AAAT'22RN10168.869.1-
ReCAM [7] CVPR'22RN10168.568.442.9
SIPE [6] CVPR'22RN10168.869.7-
SIPE [6] CVPR'22WRN38--43.6
AEFT [67] ECCV'22WRN3870.971.744.8
ACR [26] CVPR'23WRN3871.971.945.3
MCT [63] CVPR'22WRN3871.971.642.0
USAGE [44] ICCV'23WRN3871.972.842.7
USAGE [44] ICCV'23RN101--44.3
OursWRN3873.273.0-
OursRN101--45.7
", + "bbox": [ + 299, + 212, + 704, + 406 + ], + "page_idx": 13 + }, + { + "type": "text", + "text": "While ViT-based WSSS works tend to outperform CNN-based works on the PASCAL VOC 2012 dataset, they showed lower performance on the MS COCO 2014 dataset due to activation overlapping between classes. Nonetheless, our method achieved the SoTA performance on the MS COCO 2014 dataset by effectively reducing false positives. Additional CAMs and semantic segmentation results are in the Supplementary Materials.", + "bbox": [ + 212, + 436, + 787, + 529 + ], + "page_idx": 13 + }, + { + "type": "text", + "text": "5 Conclusion", + "text_level": 1, + "bbox": [ + 215, + 558, + 359, + 574 + ], + "page_idx": 13 + }, + { + "type": "text", + "text": "Weakly Supervised Semantic Segmentation suffers from CAMs that have imprecise activation around boundary areas and object-unrelated activations. In this paper, we aim to address these challenges by applying a novel perspective in the frequency domain, leveraging the Fourier Transform. Firstly, we propose Magnitude-mixing-based Phase Concentration (MPC), which generates features with identical phase but different magnitude at the intermediate layer, encouraging the model to capture the same semantics from these features. This approach leads the model to pay more attention to shape information in the phase, thereby learning to activate boundaries more clearly. Additionally, we identify the occurrence of frequency shortcuts and their association with the over-activation of CAMs. To resolve this, we introduce Frequency Shortcut Suppression (FSS), which discourages the creation of frequency shortcuts, successfully resolving object-unrelated activations. Our experimental results support the utility of the proposed method, demonstrating its effectiveness in addressing the aforementioned issues in WSSS. Furthermore, we have achieved new state-of-the-art performance on the PASCAL VOC and MS COCO datasets.", + "bbox": [ + 212, + 597, + 787, + 840 + ], + "page_idx": 13 + }, + { + "type": "page_number", + "text": "14", + "bbox": [ + 217, + 114, + 235, + 126 + ], + "page_idx": 13 + }, + { + "type": "header", + "text": "H. Kwon et al.", + "bbox": [ + 271, + 114, + 370, + 127 + ], + "page_idx": 13 + }, + { + "type": "text", + "text": "Acknowledgement This work was supported by the Technology Innovation Program (1415187329,20024355, Development of autonomous driving connectivity technology based on sensor-infrastructure cooperation) funded by the Ministry of Trade, Industry & Energy(MOTIE, Korea) and the National Research Foundation of Korea(NRF) grant funded by the Korea government(MSIT) (NRF2022R1A2B5B03002636).", + "bbox": [ + 212, + 146, + 787, + 238 + ], + "page_idx": 14 + }, + { + "type": "text", + "text": "References", + "text_level": 1, + "bbox": [ + 215, + 256, + 321, + 272 + ], + "page_idx": 14 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "1. Ahn, J., Cho, S., Kwak, S.: Weakly supervised learning of instance segmentation with inter-pixel relations. In: Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition. pp. 2209-2218 (2019)", + "2. Ahn, J., Kwak, S.: Learning pixel-level semantic affinity with image-level supervision for weakly supervised semantic segmentation. In: Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition. pp. 4981-4990 (2018)", + "3. Chang, Y.T., Wang, Q., Hung, W.C., Piramuthu, R., Tsai, Y.H., Yang, M.H.: Weakly-supervised semantic segmentation via sub-category exploration. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 8991-9000 (2020)", + "4. Chen, G., Peng, P., Ma, L., Li, J., Du, L., Tian, Y.: Amplitude-phase recombination: Rethinking robustness of convolutional neural networks in frequency domain. In: Proceedings of the IEEE/CVF International Conference on Computer Vision (ICCV). pp. 458-467 (October 2021)", + "5. Chen, L., Lei, C., Li, R., Li, S., Zhang, Z., Zhang, L.: Fpr: False positive rectification for weakly supervised semantic segmentation. In: Proceedings of the IEEE/CVF International Conference on Computer Vision. pp. 1108-1118 (2023)", + "6. Chen, Q., Yang, L., Lai, J.H., Xie, X.: Self-supervised image-specific prototype exploration for weakly supervised semantic segmentation. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 4288-4298 (2022)", + "7. Chen, Z., Wang, T., Wu, X., Hua, X.S., Zhang, H., Sun, Q.: Class reactivation maps for weakly-supervised semantic segmentation. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 969-978 (2022)", + "8. Dai, J., He, K., Sun, J.: Boxsup: Exploiting bounding boxes to supervise convolutional networks for semantic segmentation. In: Proceedings of the IEEE international conference on computer vision. pp. 1635-1643 (2015)", + "9. Dosovitskiy, A., Beyer, L., Kolesnikov, A., Weissenborn, D., Zhai, X., Unterthiner, T., Dehghani, M., Minderer, M., Heigold, G., Gelly, S., et al.: An image is worth 16x16 words: Transformers for image recognition at scale. arXiv preprint arXiv:2010.11929 (2020)", + "10. Du, Y., Fu, Z., Liu, Q., Wang, Y.: Weakly supervised semantic segmentation by pixel-to-prototype contrast. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 4320-4329 (2022)", + "11. Everingham, M., Van Gool, L., Williams, C.K., Winn, J., Zisserman, A.: The pascal visual object classes (voc) challenge. International journal of computer vision 88(2), 303-338 (2010)", + "12. Fan, J., Zhang, Z., Tan, T., Song, C., Xiao, J.: Cian: Cross-image affinity net for weakly supervised semantic segmentation. In: Proceedings of the AAAI Conference on Artificial Intelligence. vol. 34, pp. 10762-10769 (2020)" + ], + "bbox": [ + 218, + 282, + 784, + 839 + ], + "page_idx": 14 + }, + { + "type": "header", + "text": "Phase Concentration and Shortcut Suppression for WSSS", + "bbox": [ + 346, + 114, + 730, + 128 + ], + "page_idx": 14 + }, + { + "type": "page_number", + "text": "15", + "bbox": [ + 767, + 116, + 785, + 126 + ], + "page_idx": 14 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "13. Gao, W., Wan, F., Pan, X., Peng, Z., Tian, Q., Han, Z., Zhou, B., Ye, Q.: Ts-cam: Token semantic coupled attention map for weakly supervised object localization. In: Proceedings of the IEEE/CVF International Conference on Computer Vision. pp. 2886-2895 (2021)", + "14. Geirhos, R., Jacobsen, J.H., Michaelis, C., Zemel, R., Brendel, W., Bethge, M., Wichmann, F.A.: Shortcut learning in deep neural networks. Nature Machine Intelligence 2(11), 665-673 (2020)", + "15. Guo, J., Wang, N., Qi, L., Shi, Y.: Aloft: A lightweight mlp-like architecture with dynamic low-frequency transform for domain generalization. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 24132-24141 (2023)", + "16. Hansen, B.C., Hess, R.F.: Structural sparseness and spatial phase alignment in natural scenes. JOSA A 24(7), 1873-1885 (2007)", + "17. Huang, H., Kang, H., Liu, S., Salvado, O., Rakotoarivelo, T., Wang, D., Liu, T.: Paddles: Phase-amplitude spectrum disentangled early stopping for learning with noisy labels. In: Proceedings of the IEEE/CVF International Conference on Computer Vision. pp. 16719-16730 (2023)", + "18. Huang, H., Xie, S., Lin, L., Tong, R., Chen, Y.W., Li, Y., Wang, H., Huang, Y., Zheng, Y.: Semicvt: Semi-supervised convolutional vision transformer for semantic segmentation. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 11340-11349 (2023)", + "19. Huang, J., Guan, D., Xiao, A., Lu, S.: Rda: Robust domain adaptation via fourier adversarial attacking. In: Proceedings of the IEEE/CVF international conference on computer vision. pp. 8988-8999 (2021)", + "20. Huang, J., Liu, Y., Zhao, F., Yan, K., Zhang, J., Huang, Y., Zhou, M., Xiong, Z.: Deep fourier-based exposure correction network with spatial-frequency interaction. In: European Conference on Computer Vision. pp. 163-180. Springer (2022)", + "21. Jiang, P.T., Han, L.H., Hou, Q., Cheng, M.M., Wei, Y.: Online attention accumulation for weakly supervised semantic segmentation. IEEE Transactions on Pattern Analysis and Machine Intelligence 44(10), 7062-7077 (2021)", + "22. Jiang, P.T., Yang, Y., Hou, Q., Wei, Y.: L2g: A simple local-to-global knowledge transfer framework for weakly supervised semantic segmentation. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 16886-16896 (2022)", + "23. Jo, S., Yu, I.J., Kim, K.: Mars: Model-agnostic biased object removal without additional supervision for weakly-supervised semantic segmentation. arXiv preprint arXiv:2304.09913 (2023)", + "24. Kim, G., Kim, J., Lee, J.S.: Exploring adversarial robustness of vision transformers in the spectral perspective. In: Proceedings of the IEEE/CVF Winter Conference on Applications of Computer Vision. pp. 3976-3985 (2024)", + "25. Kweon, H., Yoon, S.H., Kim, H., Park, D., Yoon, K.J.: Unlocking the potential of ordinary classifier: Class-specific adversarial erasing framework for weakly supervised semantic segmentation. In: Proceedings of the IEEE/CVF International Conference on Computer Vision. pp. 6994-7003 (2021)", + "26. Kweon, H., Yoon, S.H., Yoon, K.J.: Weakly supervised semantic segmentation via adversarial learning of classifier and reconstructor. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 11329-11339 (2023)", + "27. Lee, J., Choi, J., Mok, J., Yoon, S.: Reducing information bottleneck for weakly supervised semantic segmentation. Advances in Neural Information Processing Systems 34, 27408-27421 (2021)" + ], + "bbox": [ + 217, + 146, + 785, + 840 + ], + "page_idx": 15 + }, + { + "type": "page_number", + "text": "16", + "bbox": [ + 217, + 114, + 235, + 126 + ], + "page_idx": 15 + }, + { + "type": "header", + "text": "H. Kwon et al.", + "bbox": [ + 271, + 114, + 370, + 127 + ], + "page_idx": 15 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "28. Lee, J., Kim, E., Yoon, S.: Anti-adversarily manipulated attributions for weakly and semi-supervised semantic segmentation. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 4071-4080 (2021)", + "29. Lee, J., Oh, S.J., Yun, S., Choe, J., Kim, E., Yoon, S.: Weakly supervised semantic segmentation using out-of-distribution data. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 16897-16906 (2022)", + "30. Lee, J., Yi, J., Shin, C., Yoon, S.: Bbam: Bounding box attribution map for weakly supervised semantic and instance segmentation. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 2643-2652 (2021)", + "31. Lee, S., Bae, J., Kim, H.Y.: Decompose, adjust, compose: Effective normalization by playing with frequency for domain generalization. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 11776-11785 (2023)", + "32. Li, J., Fan, J., Zhang, Z.: Towards noiseless object contours for weakly supervised semantic segmentation. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 16856-16865 (2022)", + "33. Li, K., Wu, Z., Peng, K.C., Ernst, J., Fu, Y.: Tell me where to look: Guided attention inference network. In: Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition. pp. 9215-9223 (2018)", + "34. Li, X., Zhou, T., Li, J., Zhou, Y., Zhang, Z.: Group-wise semantic mining for weakly supervised semantic segmentation. arXiv preprint arXiv:2012.05007 (2020)", + "35. Li, Y., Duan, Y., Kuang, Z., Chen, Y., Zhang, W., Li, X.: Uncertainty estimation via response scaling for pseudo-mask noise mitigation in weakly-supervised semantic segmentation. arXiv preprint arXiv:2112.07431 (2021)", + "36. Li, Y., Kuang, Z., Liu, L., Chen, Y., Zhang, W.: Pseudo-mask matters in weakly-supervised semantic segmentation. In: Proceedings of the IEEE/CVF International Conference on Computer Vision. pp. 6964-6973 (2021)", + "37. Lin, D., Dai, J., Jia, J., He, K., Sun, J.: Scribblesup: Scribble-supervised convolutional networks for semantic segmentation. In: Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition. pp. 3159-3167 (2016)", + "38. Lin, T.Y., Maire, M., Belongie, S., Hays, J., Perona, P., Ramanan, D., Dollar, P., Zitnick, C.L.: Microsoft coco: Common objects in context. In: European conference on computer vision. pp. 740-755. Springer (2014)", + "39. Lin, Y., Chen, M., Wang, W., Wu, B., Li, K., Lin, B., Liu, H., He, X.: Clip is also an efficient segmenter: A text-driven approach for weakly supervised semantic segmentation. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 15305-15314 (2023)", + "40. Luo, X., Wei, L., Wen, L., Yang, J., Xie, L., Xu, Z., Tian, Q.: Rectifying the short-cut learning of background for few-shot learning. Advances in Neural Information Processing Systems 34, 13073-13085 (2021)", + "41. Ma, C., Zhao, L., Chen, Y., Guo, L., Zhang, T., Hu, X., Shen, D., Jiang, X., Liu, T.: Rectify vit shortcut learning by visual saliency. IEEE Transactions on Neural Networks and Learning Systems (2023)", + "42. Oppenheim, A.V., Lim, J.S.: The importance of phase in signals. Proceedings of the IEEE 69(5), 529-541 (1981)", + "43. Papandreou, G., Chen, L.C., Murphy, K.P., Yuille, A.L.: Weakly-and semi-supervised learning of a deep convolutional network for semantic image segmentation. In: Proceedings of the IEEE international conference on computer vision. pp. 1742-1750 (2015)" + ], + "bbox": [ + 212, + 146, + 787, + 839 + ], + "page_idx": 16 + }, + { + "type": "header", + "text": "Phase Concentration and Shortcut Suppression for WSSS", + "bbox": [ + 348, + 114, + 730, + 128 + ], + "page_idx": 16 + }, + { + "type": "page_number", + "text": "17", + "bbox": [ + 767, + 116, + 785, + 126 + ], + "page_idx": 16 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "44. Peng, Z., Wang, G., Xie, L., Jiang, D., Shen, W., Tian, Q.: Usage: A unified seed area generation paradigm for weakly supervised semantic segmentation. arXiv preprint arXiv:2303.07806 (2023)", + "45. Qin, J., Wu, J., Xiao, X., Li, L., Wang, X.: Activation modulation and recalibration scheme for weakly supervised semantic segmentation. arXiv preprint arXiv:2112.08996 (2021)", + "46. Robinson, J., Sun, L., Yu, K., Batmanghelich, K., Jegelka, S., Sra, S.: Can contrastive learning avoid shortcut solutions? Advances in neural information processing systems 34, 4974-4986 (2021)", + "47. Rossetti, S., Zappia, D., Sanzari, M., Schaerf, M., Pirri, F.: Max pooling with vision transformers reconciles class and shape in weakly supervised semantic segmentation. In: European Conference on Computer Vision. pp. 446-463. Springer (2022)", + "48. Ru, L., Zhan, Y., Yu, B., Du, B.: Learning affinity from attention: End-to-end weakly-supervised semantic segmentation with transformers. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 16846-16855 (2022)", + "49. Ru, L., Zheng, H., Zhan, Y., Du, B.: Token contrast for weakly-supervised semantic segmentation. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 3093-3102 (2023)", + "50. Russakovsky, O., Deng, J., Su, H., Krause, J., Satheesh, S., Ma, S., Huang, Z., Karpathy, A., Khosla, A., Bernstein, M., et al.: Imagenet large scale visual recognition challenge. International journal of computer vision 115(3), 211-252 (2015)", + "51. Su, Y., Sun, R., Lin, G., Wu, Q.: Context decoupling augmentation for weakly supervised semantic segmentation. arXiv preprint arXiv:2103.01795 (2021)", + "52. Sun, G., Wang, W., Dai, J., Van Gool, L.: Mining cross-image semantics for weakly supervised semantic segmentation. arXiv preprint arXiv:2007.01947 (2020)", + "53. Sun, K., Shi, H., Zhang, Z., Huang, Y.: ECS-net: Improving weakly supervised semantic segmentation by using connections between class activation maps. In: Proceedings of the IEEE/CVF International Conference on Computer Vision. pp. 7283-7292 (2021)", + "54. Sundararajan, D.: The discrete Fourier transform: theory, algorithms and applications. World Scientific (2001)", + "55. Vaswani, A., Shazeer, N., Parmar, N., Uszkoreit, J., Jones, L., Gomez, A.N., Kaiser, L., Polosukhin, I.: Attention is all you need. Advances in neural information processing systems 30 (2017)", + "56. Vernaza, P., Chandraker, M.: Learning random-walk label propagation for weakly-supervised semantic segmentation. In: Proceedings of the IEEE conference on computer vision and pattern recognition. pp. 7158-7166 (2017)", + "57. Wang, C., Xu, R., Xu, S., Meng, W., Zhang, X.: Treating pseudo-labels generation as image matting for weakly supervised semantic segmentation. In: Proceedings of the IEEE/CVF International Conference on Computer Vision. pp. 755-765 (2023)", + "58. Wang, S., Veldhuis, R., Brune, C., Strisciuglio, N.: What do neural networks learn in image classification? a frequency shortcut perspective. In: Proceedings of the IEEE/CVF International Conference on Computer Vision (ICCV). pp. 1433-1442 (October 2023)", + "59. Wang, Y., Zhang, J., Kan, M., Shan, S., Chen, X.: Self-supervised equivariant attention mechanism for weakly supervised semantic segmentation. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 12275-12284 (2020)" + ], + "bbox": [ + 215, + 146, + 785, + 839 + ], + "page_idx": 17 + }, + { + "type": "page_number", + "text": "18", + "bbox": [ + 217, + 114, + 235, + 126 + ], + "page_idx": 17 + }, + { + "type": "header", + "text": "H. Kwon et al.", + "bbox": [ + 271, + 114, + 370, + 127 + ], + "page_idx": 17 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "60. Wu, T., Huang, J., Gao, G., Wei, X., Wei, X., Luo, X., Liu, C.H.: Embedded discriminative attention mechanism for weakly supervised semantic segmentation. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 16765-16774 (2021)", + "61. Xie, J., Hou, X., Ye, K., Shen, L.: Clims: Cross language image matching for weakly supervised semantic segmentation. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 4483-4492 (2022)", + "62. Xie, J., Xiang, J., Chen, J., Hou, X., Zhao, X., Shen, L.: C2am: Contrastive learning of class-agnostic activation map for weakly supervised object localization and semantic segmentation. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 989–998 (2022)", + "63. Xu, L., Ouyang, W., Bennamoun, M., Boussaid, F., Xu, D.: Multi-class token transformer for weakly supervised semantic segmentation. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 4310-4319 (2022)", + "64. Xu, L., Ouyang, W., Bennamoun, M., Boussaid, F., Xu, D.: Learning multi-modal class-specific tokens for weakly supervised dense object localization. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 19596-19605 (2023)", + "65. Xu, Q., Zhang, R., Zhang, Y., Wang, Y., Tian, Q.: A fourier-based framework for domain generalization. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 14383-14392 (2021)", + "66. Yang, Y., Soatto, S.: Fda: Fourier domain adaptation for semantic segmentation. In: Proceedings of the IEEE/CVF conference on computer vision and pattern recognition. pp. 4085-4095 (2020)", + "67. Yoon, S.H., Kweon, H., Cho, J., Kim, S., Yoon, K.J.: Adversarial erasing framework via triplet with gated pyramid pooling layer for weakly supervised semantic segmentation. In: Computer Vision-ECCV 2022: 17th European Conference, Tel Aviv, Israel, October 23-27, 2022, Proceedings, Part XXIX. pp. 326-344. Springer Nature Switzerland Cham (2022)", + "68. Yoon, S.H., Kweon, H., Jeong, J., Kim, H., Kim, S., Yoon, K.J.: Exploring pixel-level self-supervision for weakly supervised semantic segmentation. arXiv preprint arXiv:2112.05351 (2021)", + "69. Yoon, S.H., Kwon, H., Kim, H., Yoon, K.J.: Class tokens infusion for weakly supervised semantic segmentation. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 3595-3605 (2024)", + "70. Zhang, F., Gu, C., Zhang, C., Dai, Y.: Complementary patch for weakly supervised semantic segmentation. In: Proceedings of the IEEE/CVF International Conference on Computer Vision. pp. 7242-7251 (2021)", + "71. Zhang, X., Wei, Y., Feng, J., Yang, Y., Huang, T.S.: Adversarial complementary learning for weakly supervised object localization. In: Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition. pp. 1325-1334 (2018)", + "72. Zhou, B., Khosla, A., Lapedriza, A., Oliva, A., Torralba, A.: Learning deep features for discriminative localization. In: Proceedings of the IEEE conference on computer vision and pattern recognition. pp. 2921-2929 (2016)", + "73. Zhou, T., Zhang, M., Zhao, F., Li, J.: Regional semantic contrast and aggregation for weakly supervised semantic segmentation. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 4299-4309 (2022)" + ], + "bbox": [ + 212, + 146, + 787, + 811 + ], + "page_idx": 18 + }, + { + "type": "header", + "text": "Phase Concentration and Shortcut Suppression for WSSS", + "bbox": [ + 348, + 114, + 730, + 128 + ], + "page_idx": 18 + }, + { + "type": "page_number", + "text": "19", + "bbox": [ + 767, + 116, + 784, + 126 + ], + "page_idx": 18 + } +] \ No newline at end of file diff --git a/2024/Phase Concentration and Shortcut Suppression for Weakly Supervised Semantic Segmentation/8ce12d89-9c27-46eb-bd86-41483594c43e_model.json b/2024/Phase Concentration and Shortcut Suppression for Weakly Supervised Semantic Segmentation/8ce12d89-9c27-46eb-bd86-41483594c43e_model.json new file mode 100644 index 0000000000000000000000000000000000000000..1d110fa334cc0dbb4fa819f9a0e07173a1caf95b --- /dev/null +++ b/2024/Phase Concentration and Shortcut Suppression for Weakly Supervised Semantic Segmentation/8ce12d89-9c27-46eb-bd86-41483594c43e_model.json @@ -0,0 +1,2790 @@ +[ + [ + { + "type": "title", + "bbox": [ + 0.233, + 0.141, + 0.773, + 0.187 + ], + "angle": 0, + "content": "Phase Concentration and Shortcut Suppression for Weakly Supervised Semantic Segmentation" + }, + { + "type": "text", + "bbox": [ + 0.235, + 0.213, + 0.768, + 0.229 + ], + "angle": 0, + "content": "Hoyong Kwon, Jaeseok Jeong, Sung-Hoon Yoon, and Kuk-Jin Yoon" + }, + { + "type": "text", + "bbox": [ + 0.3, + 0.24, + 0.704, + 0.269 + ], + "angle": 0, + "content": "KAIST {kwonhoyong3, jason.jeong,yoon307,kjyoon}@kaist.ac.kr" + }, + { + "type": "text", + "bbox": [ + 0.261, + 0.302, + 0.761, + 0.594 + ], + "angle": 0, + "content": "Abstract. Weakly Supervised Semantic Segmentation (WSSS) with image-level supervision typically acquires object localization information from Class Activation Maps (CAMs). While Vision Transformers (ViTs) in WSSS have been increasingly explored for their superior performance in understanding global context, CAMs from ViT still show imprecise localization in boundary areas and false positive activation. This paper proposes a novel WSSS framework that targets these issues based on the information from the frequency domain. In our framework, we introduce the Magnitude-mixing-based Phase Concentration (MPC) module, which guides the classifier to prioritize phase information containing high-level semantic details. By perturbing and mixing the magnitude, MPC guides the classifier to accentuate and concentrate on the shape information in the phase, thereby leading to finer distinctions in CAMs boundary regions. Additionally, inspired by empirical observations that the classification \"shortcut\" in the frequency domain can induce false positives in CAMs, we introduce a Frequency Shortcut Suppression (FSS) module. This module aims to discourage the formation of such shortcuts, thereby mitigating false positives. The effectiveness of our approach is demonstrated by achieving new state-of-the-art performance on both PASCAL VOC 2012 and MS COCO 2014 datasets. The code is available at https://github.com/kwonhoyong3/PCSS-WSSS." + }, + { + "type": "text", + "bbox": [ + 0.261, + 0.606, + 0.741, + 0.634 + ], + "angle": 0, + "content": "Keywords: Weakly Supervised Semantic Segmentation \\(\\cdot\\) Fourier Transform \\(\\cdot\\) Shortcut Learning" + }, + { + "type": "title", + "bbox": [ + 0.216, + 0.659, + 0.377, + 0.675 + ], + "angle": 0, + "content": "1 Introduction" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.689, + 0.788, + 0.795 + ], + "angle": 0, + "content": "Weakly Supervised Semantic Segmentation (WSSS) aims to enhance its utility by reducing the dependency on human-annotated pixel-level labels in semantic segmentation. With this goal, WSSS mainly explores methods to generate accurate pseudo-labels using only easily obtainable weak labels such as image-level labels [1-3,25,26,29,51,60,68,69], scribbles [37,56], and bounding boxes [8,30,43]. Among these methods, leveraging image-level labels has been extensively investigated, primarily due to the availability of large existing datasets." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.796, + 0.788, + 0.842 + ], + "angle": 0, + "content": "Since image-level labels do not provide spatial information of objects, image-level based WSSS typically involves multiple stages: 1) extract Class Activation Maps (CAMs) to localize the objects, 2) refine CAMs to generate pseudo" + } + ], + [ + { + "type": "page_number", + "bbox": [ + 0.218, + 0.116, + 0.23, + 0.127 + ], + "angle": 0, + "content": "2" + }, + { + "type": "header", + "bbox": [ + 0.272, + 0.115, + 0.371, + 0.128 + ], + "angle": 0, + "content": "H. Kwon et al." + }, + { + "type": "image", + "bbox": [ + 0.218, + 0.145, + 0.331, + 0.232 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.248, + 0.233, + 0.302, + 0.245 + ], + "angle": 0, + "content": "(a) Image" + }, + { + "type": "image", + "bbox": [ + 0.332, + 0.145, + 0.443, + 0.232 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.334, + 0.233, + 0.443, + 0.245 + ], + "angle": 0, + "content": "(b) Freq. MaskBird" + }, + { + "type": "image", + "bbox": [ + 0.445, + 0.146, + 0.557, + 0.232 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.451, + 0.233, + 0.551, + 0.244 + ], + "angle": 0, + "content": "(c) Filtered Image" + }, + { + "type": "image", + "bbox": [ + 0.559, + 0.146, + 0.672, + 0.232 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.592, + 0.233, + 0.642, + 0.243 + ], + "angle": 0, + "content": "(d) CAM" + }, + { + "type": "image", + "bbox": [ + 0.673, + 0.146, + 0.785, + 0.232 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.681, + 0.233, + 0.776, + 0.243 + ], + "angle": 0, + "content": "(e) CAM Filtered" + }, + { + "type": "image_caption", + "bbox": [ + 0.214, + 0.256, + 0.788, + 0.314 + ], + "angle": 0, + "content": "Fig. 1: Experimental results verifying the relationship between CAMs and Frequency Shortcut. The Filtered Image (c) was obtained by removing the Principal Frequency components (b) related to the class of the Image (a). (d) and (e) are CAMs of images (a) and (c), respectively, derived from the same ViT trained on clean images." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.342, + 0.788, + 0.447 + ], + "angle": 0, + "content": "ground truth (Pseudo-GT), and 3) train a semantic segmentation model using the Pseudo-GT. As the second and third stages are dependent on the CAMs from the first stage, our work targets to improve the quality of CAMs. In the first stage, unlike Convolutional Neural Networks (CNNs) that focus on discriminative regions, Vision Transformer (ViT) [9] stands out in WSSS for its ability to understand global context through the Multi-Head Self-Attention (MHSA) mechanism [55]." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.448, + 0.788, + 0.509 + ], + "angle": 0, + "content": "Despite these advantages over CNN-based CAMs, CAMs from ViT still face several issues; 1) CAMs exhibit imprecise object boundaries due to lack of spatial supervision, and 2) CAMs often localize object-unrelated (false positive) regions. In this paper, we propose two methodologies to address the issues with CAMs." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.509, + 0.788, + 0.765 + ], + "angle": 0, + "content": "When humans and ViTs visually assess an object, semantics is predominantly influenced by shape over texture [4, 24]. While shape and texture information are intertwined in images, applying Discrete Fourier Transform (DFT) [54] can help extract the two information into magnitude and phase spectral components. Inspired by the fact that the phase spectrum encapsulates shape and high-level semantic information [16, 42], we propose Magnitude-mixing-based Phase Concentration (MPC) to encourage models to focus more on the phase and improve object boundaries in generated CAMs. In MPC, we select a random pair of images consisting of an anchor image and another arbitrary image. Within the intermediate layers of a ViT, we merge the magnitude information from the arbitrary image into the magnitude from the anchor image. This results in the creation of two anchor image features with identical phases but different magnitudes. By ensuring that both features are guided by the same weak supervision and that the generated CAMs are consistent regardless of changes in magnitude, we induce the model to focus on the phase. Consequently, MPC enables the model to concentrate on the crucial information within the phase, leading to CAMs with more precise boundaries." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.765, + 0.788, + 0.84 + ], + "angle": 0, + "content": "While MPC facilitates precise activation near the boundaries, the issue of activation in object-unrelated regions persists. Interestingly, certain frequency components have been shown to significantly impact classification during the training process of ViT, acting as Frequency Shortcut [58]. A frequency shortcut is a simplicity bias in classification networks, where the network relies on a" + } + ], + [ + { + "type": "header", + "bbox": [ + 0.348, + 0.115, + 0.732, + 0.129 + ], + "angle": 0, + "content": "Phase Concentration and Shortcut Suppression for WSSS" + }, + { + "type": "page_number", + "bbox": [ + 0.775, + 0.117, + 0.785, + 0.127 + ], + "angle": 0, + "content": "3" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.147, + 0.788, + 0.298 + ], + "angle": 0, + "content": "specific frequency rather than semantic content for classification. Additionally, we empirically observed that models tend to exhibit false positives in the CAMs of classes attempting to find shortcuts in the frequency domain. To explore the effects of principal frequency and shortcuts in CAMs, we compare the CAMs from a clean image and a corresponding image with the principal frequency removed (i.e., filtered image), as shown in Fig. 1. While the CAMs from clean image (Fig. 1(d)) suffers from false activations, false activations are relieved in the CAMs from filtered image (Fig. 1(e)). Here, we can verify that certain principal frequencies influence the entire image and cause widespread false activation regardless of the semantics; these can be regarded as shortcut frequencies." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.299, + 0.788, + 0.448 + ], + "angle": 0, + "content": "To resolve the false positives induced by frequency shortcuts, we propose the Frequency Shortcut Suppression (FSS). Initially, we measured the impact of each frequency on the classification of each class to aggregate a Frequency Shortcut Potential Map (FSPM). Based on FSPM, we removed the frequency shortcut components from the image to generate the filtered image. Through classification on the filtered image and by comparing CAMS generated from the original image and the filtered image, FSS reduces the reliance of the generated CAMs to false positive inducing shortcut frequency components. This removal of shortcut frequency components happens iteratively, with the FSPM constantly adjusted." + }, + { + "type": "text", + "bbox": [ + 0.241, + 0.45, + 0.704, + 0.463 + ], + "angle": 0, + "content": "The main contributions of this paper are summarized as follows:" + }, + { + "type": "text", + "bbox": [ + 0.216, + 0.475, + 0.785, + 0.517 + ], + "angle": 0, + "content": "- Through Magnitude-mixing-based Phase Concentration (MPC), we guide ViT to focus on the meaningful phase information, resulting in precise activation around boundaries." + }, + { + "type": "text", + "bbox": [ + 0.216, + 0.52, + 0.785, + 0.564 + ], + "angle": 0, + "content": "- We proposed a Frequency Shortcut Suppression (FSS) module that effectively reduces false positive activations in CAMs by suppressing shortcut learning in the Frequency Domain." + }, + { + "type": "list", + "bbox": [ + 0.216, + 0.475, + 0.785, + 0.564 + ], + "angle": 0, + "content": null + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.577, + 0.788, + 0.622 + ], + "angle": 0, + "content": "To demonstrate the effectiveness of our methodology, we conducted comparisons with state-of-the-art WSSS methods on the PASCAL VOC 2012 [11] and MS COCO 2014 datasets [38], achieving new SoTA performance." + }, + { + "type": "title", + "bbox": [ + 0.216, + 0.646, + 0.396, + 0.66 + ], + "angle": 0, + "content": "2 Related Works" + }, + { + "type": "title", + "bbox": [ + 0.216, + 0.678, + 0.617, + 0.694 + ], + "angle": 0, + "content": "2.1 Weakly Supervised Semantic Segmentation" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.704, + 0.788, + 0.843 + ], + "angle": 0, + "content": "Most existing WSSS approaches utilize Class Activation Maps (CAMs) [72] obtained from Convolutional Neural Networks (CNNs) to localize objects when only image-level labels are provided. However, raw CAMs (i.e. seeds) fail to localize the whole object with imprecise boundaries. In pursuit of generating more accurate pixel-level pseudo-labels, research in WSSS aims to propose a method to improve the quality of raw CAMs or further refine/post-processing the CAMs. CAMs Quality Improvement To guide CAMs to localize the less discriminative regions, various methods are proposed including online CAMs aggregation [21], sub-categories exploration [3], information bottleneck reduction [27]" + } + ], + [ + { + "type": "page_number", + "bbox": [ + 0.218, + 0.116, + 0.23, + 0.127 + ], + "angle": 0, + "content": "4" + }, + { + "type": "header", + "bbox": [ + 0.272, + 0.115, + 0.371, + 0.128 + ], + "angle": 0, + "content": "H. Kwon et al." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.146, + 0.788, + 0.403 + ], + "angle": 0, + "content": "and mining cross-image relations [12, 34, 52]. Also, [22, 59, 70] introduce various auxiliary tasks to regularize CAMs training. Along with these methods, Adversarial Erasing (AE) [25, 33, 53, 67, 71] methods, which are based on the erase-and-find mechanism, are also actively researched to expand CAMs. Apart from techniques aimed at expanding CAMs, contrastive learning-based WSSS [6, 62, 73] and adversarial learning between classifier-reconstructor [26] are proposed to produce CAMs with precise object boundaries. Though CNN-based WSSS methods show promising performance, with the success of Vision Transformer (ViT) in various vision tasks, several methods [5, 13, 39, 44, 48, 49, 63, 64, 69] are proposed to migrate the ViT for WSSS. MCTformer [63] introduces multi-class token to extract class-specific attention map from ViT by improving the TS-CAM [13], which is designed for weakly-supervised object localization. In the same line, ViT-PCM [47] proposed a method to extract CAMs based on patch-class mapping. Along with these methods, large language model-based WSSS [39, 61, 64] and token-level contrasting [49] are also introduced. Unlike the prior methods, our work is the first to utilize frequency domain information to address the issues with CAMs in ViT-based WSSS." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.403, + 0.788, + 0.526 + ], + "angle": 0, + "content": "CAMs Post-processing To obtain high-quality pseudo labels for semantic segmentation, various approaches are introduced in WSSS based on affinity learning [1, 2], removing intra-object edges in contours [32], iterative refinements [35, 36]. Besides these methods, Image-matting based refinements [57] and utilizing unsupervised semantic segmentation model [23] greatly boost the quality of pseudo labels. As the post-processing methods can be synergistically applied with methods that generate precise initial seeds (CAMs), this paper focuses on improving the quality of the CAMs." + }, + { + "type": "title", + "bbox": [ + 0.215, + 0.545, + 0.584, + 0.56 + ], + "angle": 0, + "content": "2.2 Fourier Transform in Computer Vision" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.569, + 0.788, + 0.81 + ], + "angle": 0, + "content": "Transforming images from the spatial domain to the frequency domain via the Fourier Transform enables the separation of information into phase and amplitude components. Early studies [16,42] have shown that while amplitude carries low-level statistics and texture information, phase encapsulates high-level semantic and shape information. Leveraging this advantage, Fourier Transform has been applied to a variety of vision tasks such as Exposure Correction [20], Semantic Segmentation [18], and Learning with noisy label [17]. Along with the successful application of frequency components, the fields of Domain Adaptation [19, 66] and Generalization [15, 31, 65] have emphasized the importance of phase for its high-level semantic and structural information. FDA [66] performed Domain Adaptation by transforming low-level amplitude in the Fourier Domain, maintaining high-level semantics and structure, building on the understanding that low-level amplitude contains image style information. [65] highlighted the significance of the phase component, which carries semantic information, and proposed Fourier-based data augmentation through Image-level amplitude to perform Domain Generalization with an emphasis on phase information." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.81, + 0.787, + 0.842 + ], + "angle": 0, + "content": "These studies underscore the significant role of the frequency domain in neural networks. Intriguingly, classification networks engage in shortcut learning" + } + ], + [ + { + "type": "header", + "bbox": [ + 0.348, + 0.115, + 0.733, + 0.13 + ], + "angle": 0, + "content": "Phase Concentration and Shortcut Suppression for WSSS" + }, + { + "type": "page_number", + "bbox": [ + 0.775, + 0.117, + 0.785, + 0.127 + ], + "angle": 0, + "content": "5" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.147, + 0.784, + 0.239 + ], + "angle": 0, + "content": "in the frequency domain [58], identifying specific frequencies for classification. Since shortcut learning [14] can disrupt semantic capture, research efforts have been directed towards reducing such effects [40, 46]. SGT [41] analyzed how vanilla ViT activates areas unrelated to the object due to shortcut learning in the background regions during the classification process and proposed rectifying this through saliency maps." + }, + { + "type": "title", + "bbox": [ + 0.216, + 0.257, + 0.338, + 0.273 + ], + "angle": 0, + "content": "3 Methods" + }, + { + "type": "title", + "bbox": [ + 0.216, + 0.286, + 0.36, + 0.301 + ], + "angle": 0, + "content": "3.1 Preliminary" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.307, + 0.785, + 0.422 + ], + "angle": 0, + "content": "Vision Transformer for CAMs To propagate the image \\(\\mathbf{I}\\) into Vision Transformer (ViT), it is split into \\(N\\times N\\) patches and forms the patch tokens \\(\\mathbf{T}_{patch}^{0}\\in \\mathbb{R}^{D\\times N^2}\\), where \\(D\\) is the embedding dimension. Along with \\(C\\) class tokens \\(\\mathbf{T}_{cls}^{0}\\in \\mathbb{R}^{D\\times C}\\), the patch tokens \\(\\mathbf{T}_{patch}^{0}\\) pass through the Vision Transformer (ViT). Consequently, the final class tokens \\(\\mathbf{T}_{cls}^{L}\\in \\mathbb{R}^{D\\times C}\\) and patch tokens \\(\\mathbf{T}_{patch}^{L}\\in \\mathbb{R}^{D\\times N^2}\\) are derived as outputs, where \\(L\\) represents the number of ViT layers. This procedure can be formally defined as follows:" + }, + { + "type": "equation", + "bbox": [ + 0.351, + 0.429, + 0.785, + 0.448 + ], + "angle": 0, + "content": "\\[\n\\mathbf {T} _ {c l s} ^ {i}, \\mathbf {T} _ {p a t c h} ^ {i} = \\mathbf {V} \\mathrm {i} \\mathbf {T} _ {i} \\left(\\mathbf {T} _ {c l s} ^ {i - 1}, \\mathbf {T} _ {p a t c h} ^ {i - 1}\\right), 1 \\leq i \\leq L \\tag {1}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.455, + 0.788, + 0.659 + ], + "angle": 0, + "content": "where \\(\\mathbf{V}\\mathbf{i}\\mathbf{T}_i\\) is \\(i^{th}\\) layer of ViT. The class tokens \\(\\mathbf{T}_{cls}^L\\) are processed through average pooling to derive the classification logits \\(y_{c} \\in \\mathbb{R}^{C}\\). Meanwhile, the patch tokens \\(\\mathbf{T}_{patch}^L\\), arranged according to their spatial positions, pass through a convolution layer to produce the Class Activation Maps (CAMs) \\(\\mathbf{M} \\in \\mathbb{R}^{C \\times N \\times N}\\). By applying Global Average Pooling (GAP) to \\(\\mathbf{M}\\), we can obtain classification logits \\(y_{p} \\in \\mathbb{R}^{C}\\) in a patch-level. These two classification logits are supervised using a multi-label soft margin loss \\(\\mathcal{L}_{cls}\\) with the given classification labels \\(\\mathbf{y}\\). Additionally, following [63], we extract the token-to-token attention \\(\\mathbf{A}_{t2t} \\in \\mathbb{R}^{(C + N^2) \\times (C + N^2)}\\) based on multi-head self-attention. From this token-to-token attention \\(\\mathbf{A}_{t2t}\\), the class-to-patch attention \\(\\mathbf{A}_{c2p} \\in \\mathbb{R}^{C \\times N^2}\\) and patch-to-patch attention \\(\\mathbf{A}_{p2p} \\in \\mathbb{R}^{N^2 \\times N^2}\\) can be extracted. These two types of attention maps are used to refine the CAMs, forming a refined CAMs \\(\\mathbf{M}_{ref} \\in \\mathbb{R}^{C \\times N \\times N}\\). The aforementioned process can be noted as follows:" + }, + { + "type": "equation", + "bbox": [ + 0.407, + 0.668, + 0.785, + 0.684 + ], + "angle": 0, + "content": "\\[\n\\mathbf {M} _ {r e f} = \\left(\\mathbf {M} \\odot \\mathbf {A} _ {c 2 p}\\right) \\times \\mathbf {A} _ {p 2 p} \\tag {2}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.69, + 0.784, + 0.751 + ], + "angle": 0, + "content": "where \\(\\odot\\) and \\(\\times\\) are Hadamard product and matrix multiplication, respectively. Discrete Fourier Transform Let \\(\\mathcal{F}\\) denote the Discrete Fourier Transform (DFT) [54] applied across the spatial dimensions. For a feature map \\(\\mathbf{X} \\in \\mathbb{R}^{D \\times H \\times W}\\) the result of applying \\(\\mathcal{F}\\) to a single channel can be represented as follows:" + }, + { + "type": "equation", + "bbox": [ + 0.326, + 0.759, + 0.785, + 0.786 + ], + "angle": 0, + "content": "\\[\n\\mathcal {F} (\\mathbf {X}) (m, n) = \\sum_ {h, w} \\mathbf {X} (h, w) e ^ {- j 2 \\pi \\left(\\frac {h}{H} m + \\frac {w}{W} n\\right)}, j ^ {2} = - 1 \\tag {3}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.795, + 0.784, + 0.841 + ], + "angle": 0, + "content": "Since the absolute value and the angle component of \\(\\mathcal{F}(\\mathbf{X})\\) respectively represent magnitude and phase components, we denote the magnitude spectrum as \\(|\\mathcal{F}(\\mathbf{X})|\\) and the phase spectrum as \\(\\angle \\mathcal{F}(\\mathbf{X})\\) throughout the paper." + } + ], + [ + { + "type": "page_number", + "bbox": [ + 0.218, + 0.116, + 0.23, + 0.127 + ], + "angle": 0, + "content": "6" + }, + { + "type": "header", + "bbox": [ + 0.272, + 0.115, + 0.371, + 0.128 + ], + "angle": 0, + "content": "H. Kwon et al." + }, + { + "type": "image", + "bbox": [ + 0.218, + 0.147, + 0.789, + 0.373 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.214, + 0.386, + 0.79, + 0.511 + ], + "angle": 0, + "content": "Fig. 2: Visualization of the proposed framework. The Magnitude-mixing-based Phase Concentration (MPC) is executed at an intermediate layer. Here, the magnitude of features from the anchor image \\(\\mathbf{I}\\) and another arbitrary image \\(\\hat{\\mathbf{I}}\\) are mixed to make a feature with identical phase, but different magnitude. The model is guided to perform correct classification and derive consistent CAMs \\(\\mathbf{M}_{ref}\\) from a pair of features. In the Frequency Shortcut Suppression (FSS), based on the Frequency Shortcut Potential Map (FSPM), frequency components that have a significant impact on the classification of classes in \\(\\mathbf{I}\\) are reduced before being inputted to the ViT. The model is encouraged to classify correctly and minimize the difference between \\(\\mathbf{M}\\) and \\(\\hat{\\mathbf{M}}\\)." + }, + { + "type": "title", + "bbox": [ + 0.216, + 0.542, + 0.644, + 0.558 + ], + "angle": 0, + "content": "3.2 Magnitude-mixing-based Phase Concentration" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.568, + 0.789, + 0.674 + ], + "angle": 0, + "content": "With the aid of multi-head self-attention, CAMs from ViT are more proficient in capturing the less-discriminative regions than CNN-based methods. Nonetheless, they still show imprecise localization along object boundaries due to a lack of spatial constraint. To improve localization of CAMs, we design the Magnitude-mixing-based Phase Concentration (MPC) module to extract high-level semantic and boundary information shown to be present in the phase component of the frequency domain [16, 42]." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.675, + 0.79, + 0.844 + ], + "angle": 0, + "content": "We consider two ideas in MPC to guide the model to prioritize the phase component: 1) The model should correctly classify the image even when the magnitude component is altered. Similarly, 2) the CAMs generated after the magnitude change should be similar to those obtained before the change. The key to these two ideas is to change the magnitude while keeping the phase component intact. Since the magnitude needs to be 'feasible' yet different, we designed a magnitude-mixing method. Here, we form a random image pair consisting of an anchor image \\(\\mathbf{I}\\) and an arbitrary image \\(\\dot{\\mathbf{I}}\\) for perturbation. Let the input class and patch tokens of \\(\\mathbf{I}\\) be denoted as \\([\\mathbf{T}_{cls}^{0},\\mathbf{T}_{patch}^{0}]\\), and those of \\(\\dot{\\mathbf{I}}\\) as \\([\\dot{\\mathbf{T}}_{cls}^{0},\\dot{\\mathbf{T}}_{patch}^{0}]\\). We then process each set of tokens through the ViT layers, as outlined in Eq. 1, for \\(1\\leq i\\leq \\ell\\), to obtain \\([\\mathbf{T}_{cls}^{\\ell},\\mathbf{T}_{patch}^{\\ell}]\\) and \\([\\dot{\\mathbf{T}}_{cls}^{\\ell},\\dot{\\mathbf{T}}_{patch}^{\\ell}]\\) respectively. After reshaping" + } + ], + [ + { + "type": "header", + "bbox": [ + 0.348, + 0.115, + 0.733, + 0.129 + ], + "angle": 0, + "content": "Phase Concentration and Shortcut Suppression for WSSS" + }, + { + "type": "page_number", + "bbox": [ + 0.775, + 0.117, + 0.785, + 0.127 + ], + "angle": 0, + "content": "7" + }, + { + "type": "image", + "bbox": [ + 0.235, + 0.145, + 0.769, + 0.259 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.215, + 0.269, + 0.788, + 0.354 + ], + "angle": 0, + "content": "Fig. 3: Visualization of Frequency Influence Measurement. For an image \\(\\mathbf{I}\\) containing class \\(c\\), a set of images that each masked at different frequency components using \\(M_{f}\\) in the frequency domain of \\(\\mathbf{I}\\) are inputted into ViT. The classification loss for class \\(c\\) is calculated as the influence of frequency \\(f\\) on class \\(c\\). Accumulate the process across the image set \\(\\mathbb{I}_c\\) to produce \\(\\mathbf{P}(c)\\), and repeat for each \\(c\\). Finally, \\(\\mathbf{P}\\) is normalized to obtain the Frequency Shortcut Potential Map (FSPM) for each class." + }, + { + "type": "text", + "bbox": [ + 0.215, + 0.379, + 0.785, + 0.41 + ], + "angle": 0, + "content": "and aligning \\(\\mathbf{T}_{patch}^{\\ell}\\) and \\(\\bar{\\mathbf{T}}_{patch}^{\\ell}\\) according to their spatial positions, the spectral magnitude is obtained and perturbed using the following formula:" + }, + { + "type": "equation", + "bbox": [ + 0.27, + 0.416, + 0.786, + 0.436 + ], + "angle": 0, + "content": "\\[\n\\bar {\\mathbf {T}} _ {p a t c h} ^ {\\ell} = \\mathcal {F} ^ {- 1} \\left( \\right.\\left\\langle \\right.\\left\\langle \\right. r \\mid \\mathcal {F} \\left(\\mathbf {T} _ {p a t c h} ^ {\\ell}\\right)\\left. \\right| + (1 - r) \\mid \\mathcal {F} \\left(\\dot {\\mathbf {T}} _ {p a t c h} ^ {\\ell}\\right)\\left. \\right|, \\angle \\mathcal {F} \\left(\\mathbf {T} _ {p a t c h} ^ {\\ell}\\right)\\left. \\right\\rangle \\rangle\\left. \\right) \\tag {4}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.215, + 0.442, + 0.788, + 0.631 + ], + "angle": 0, + "content": "where \\(r\\) is the perturbation ratio between the anchor image and arbitrary image, and \\(\\mathcal{F}^{-1}\\) is the inverse DFT. Tokens and maps that have been affected by magnitude-mixing are denoted by a bar on top (i.e. \\(\\bar{\\mathbf{T}}_{patch}^{\\ell}\\)). \\(\\langle \\langle A,B\\rangle \\rangle\\) is coupling of magnitude \\(A\\) and phase \\(B\\). As shown in Eq. 4, the phase component of \\(\\bar{\\mathbf{T}}_{patch}^{\\ell}\\) remains unchanged as the magnitude is perturbed. After iDFT, the set of tokens \\([\\mathbf{T}_{cls}^{\\ell},\\bar{\\mathbf{T}}_{patch}^{\\ell}]\\) are reshaped and passed through the remaining layers of ViT, following Eq. 1 from \\(\\ell < i\\leq L\\), to produce the output tokens \\([\\bar{\\mathbf{T}}_{cls}^{L},\\bar{\\mathbf{T}}_{patch}^{L}]\\). The class prediction \\(\\bar{y}_c\\) is obtained from \\(\\bar{\\mathbf{T}}_{cls}^{L}\\), while the class prediction \\(\\bar{y}_p\\) and refined CAMs \\(\\bar{\\mathbf{M}}_{ref}\\) are generated from \\(\\bar{\\mathbf{T}}_{patch}^{L}\\). \\(\\bar{y}_c\\) and \\(\\bar{y}_p\\) are used as class predictions in multi-label soft margin loss to obtain \\(\\mathcal{L}_{mpc - cls}\\). Following the second idea, the difference between the refined CAMs \\(\\bar{\\mathbf{M}}_{ref}\\) and \\(\\bar{\\mathbf{M}}_{ref}\\) are minimized using the following loss:" + }, + { + "type": "equation", + "bbox": [ + 0.4, + 0.638, + 0.785, + 0.656 + ], + "angle": 0, + "content": "\\[\n\\mathcal {L} _ {\\text {m p c - c a m s}} = \\left| \\mathbf {M} _ {\\text {r e f}} - \\bar {\\mathbf {M}} _ {\\text {r e f}} \\right| _ {1} \\tag {5}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.216, + 0.662, + 0.662, + 0.677 + ], + "angle": 0, + "content": "Ultimately, the MPC is supervised through the following loss:" + }, + { + "type": "equation", + "bbox": [ + 0.391, + 0.686, + 0.785, + 0.702 + ], + "angle": 0, + "content": "\\[\n\\mathcal {L} _ {m p c} = \\mathcal {L} _ {m p c - c l s} + \\lambda_ {1} \\mathcal {L} _ {m p c - c a m s} \\tag {6}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.215, + 0.709, + 0.785, + 0.77 + ], + "angle": 0, + "content": "where \\(\\lambda_{1}\\) is a hyperparameter to balance the losses. Additionally, the gradient to \\(\\mathbf{M}_{ref}\\) from \\(\\mathcal{L}_{mpc-cams}\\) is detached to guide \\(\\bar{\\mathbf{M}}_{ref}\\) to follow \\(\\mathbf{M}_{ref}\\). This increases the dependency of the model on the phase information, enabling the model to better understand the boundary information in the phase component." + }, + { + "type": "title", + "bbox": [ + 0.216, + 0.788, + 0.53, + 0.804 + ], + "angle": 0, + "content": "3.3 Frequency Shortcut Suppression" + }, + { + "type": "text", + "bbox": [ + 0.215, + 0.81, + 0.787, + 0.84 + ], + "angle": 0, + "content": "Through the MPC module, we have addressed the issues of activation around the boundary in CAMs. However, the problem of false activations in CAMs" + } + ], + [ + { + "type": "page_number", + "bbox": [ + 0.218, + 0.116, + 0.23, + 0.127 + ], + "angle": 0, + "content": "8" + }, + { + "type": "header", + "bbox": [ + 0.272, + 0.115, + 0.372, + 0.128 + ], + "angle": 0, + "content": "H. Kwon et al." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.147, + 0.788, + 0.344 + ], + "angle": 0, + "content": "persists. Various factors contribute to false activations in ViT, such as GAP for generating classification logits from CAMs and the over-smoothing problem caused by self-attention of ViT [49]. Interestingly, we experimentally found that false activations can also arise due to Shortcut Learning in the frequency domain [14], which we aim to address. Shortcut Learning during the classification training process in ViT can negatively impact genuine semantic understanding [40], and such shortcuts can lead to wrong activations [41]. Furthermore, it has been revealed that Neural Networks can create shortcuts in the frequency domain [58]. We empirically found a correlation between frequency shortcuts and false positives in CAMs M. Since Frequency Shortcuts are unrelated to objectness, activations caused by frequency shortcuts occur in locations unrelated to objects. We have resolved over-activation by disrupting the creation of frequency shortcuts through the Frequency Shortcut Suppression(FSS) module." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.344, + 0.79, + 0.588 + ], + "angle": 0, + "content": "Frequency Shortcut Potential Map To implement the FSS, the potential of specific frequency components to be used as shortcuts in classification needs to be measured. Inspired by Wang et al. [58], we utilized a Frequency Influence Measurement (FIM) framework, as depicted in Fig. 3, to assess the impact of each component. If the classification loss of the model increases significantly for an image \\(\\mathbf{I}\\) of class \\(c\\) when information about a certain frequency component \\(f\\) is removed, it indicates that \\(f\\) has an important role in classifying class \\(c\\) for the model. Therefore, for each class \\(c\\), the influence of \\(f\\) on the classification of the image set \\(\\mathbb{I}_c\\) corresponding to that class is calculated by accumulating classification loss. To remove a specific frequency component \\(f\\) from an image, a Masking map \\(M_f \\in \\mathbb{R}^{H \\times W}\\) is used, where the value for \\(f\\) is set to 0, and others are set to 1. For an image \\(\\mathbf{I}\\), let \\(\\mathbf{I}_f = \\mathcal{F}^{-1}(\\mathcal{F}(\\mathbf{I}) \\odot M_f)\\) represent the image with the \\(f\\) component masked where \\(\\odot\\) is element-wise multiplication. The classification logit derived after passing \\(\\mathbf{I}_f\\) through a ViT is denoted by \\(y_{\\mathbf{I}_f}\\). When \\(\\mathbf{P} \\in \\mathbb{R}^{C \\times H \\times W}\\) is frequency influence accumulated map, the influence of the frequency component \\(f\\) on the classification of class \\(c\\) is calculated as follows:" + }, + { + "type": "equation", + "bbox": [ + 0.405, + 0.597, + 0.788, + 0.625 + ], + "angle": 0, + "content": "\\[\n\\mathbf {P} (c, f) = \\sum_ {\\mathbf {I} \\in \\mathbb {I} _ {c}} \\mathcal {L} _ {b c e} \\left(y _ {\\mathbf {I} _ {f}} (c), 1\\right) \\tag {7}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.635, + 0.788, + 0.711 + ], + "angle": 0, + "content": "where \\(y(c)\\) and \\(\\mathcal{L}_{bce}\\) are probability of class \\(c\\) and binary cross-entropy loss, respectively. \\(\\mathbf{P}(c)\\in \\mathbb{R}^{H\\times W}\\) ultimately holds the classification influence information of each frequency for class \\(c\\), with higher values indicating a significant impact on classifying class \\(c\\). \\(\\mathbf{P}\\) is normalized for each class to create the Frequency Shortcut Potential Map (FSPM) \\(\\mathbf{FS}\\in \\mathbb{R}^{C\\times H\\times W}\\):" + }, + { + "type": "equation", + "bbox": [ + 0.368, + 0.724, + 0.786, + 0.74 + ], + "angle": 0, + "content": "\\[\n\\mathbf {F S} (c) = \\operatorname {N o r m} (\\mathbf {P} (c)), c \\in \\{1, 2, \\dots , C \\} \\tag {8}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.75, + 0.788, + 0.779 + ], + "angle": 0, + "content": "where Norm is the normalization function that scales each value to fall between 0 and 1." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.78, + 0.788, + 0.841 + ], + "angle": 0, + "content": "Frequency Shortcut Suppression We propose the Frequency Shortcut Suppression (FSS) utilizing the FSPM generated through the FIM process. The FSPM encapsulates information on which frequency components significantly impact classification at the time of FIM execution, indicating frequencies that" + } + ], + [ + { + "type": "header", + "bbox": [ + 0.348, + 0.115, + 0.733, + 0.13 + ], + "angle": 0, + "content": "Phase Concentration and Shortcut Suppression for WSSS" + }, + { + "type": "page_number", + "bbox": [ + 0.775, + 0.117, + 0.785, + 0.127 + ], + "angle": 0, + "content": "9" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.147, + 0.784, + 0.223 + ], + "angle": 0, + "content": "the model is likely to use as shortcuts. The image without shortcut frequency can be generated by reducing the magnitude of frequency components based on the FSPM. Subsequently, the model can be prevented from forming shortcuts by encouraging it to classify correctly and generate consistent CAMs from the original and filtered images." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.223, + 0.785, + 0.269 + ], + "angle": 0, + "content": "Let an image \\(\\mathbf{I}\\) contains classes \\(c\\in \\mathbb{C}_{\\mathbf{I}}\\). The filtered image \\(\\hat{\\mathbf{I}}\\) is synthesized by decreasing the frequency components from \\(\\mathbf{I}\\) based on the FSPM, making it challenging for the model to classify through shortcuts:" + }, + { + "type": "equation", + "bbox": [ + 0.377, + 0.281, + 0.786, + 0.313 + ], + "angle": 0, + "content": "\\[\n\\hat {\\mathbf {I}} = \\frac {1}{N} \\sum_ {c \\in \\mathbb {C} _ {\\mathbf {I}}} \\mathcal {F} ^ {- 1} \\left(\\mathcal {F} (\\mathbf {I}) \\odot (1 - \\mathbf {F S} (c))\\right) \\tag {9}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.215, + 0.324, + 0.784, + 0.401 + ], + "angle": 0, + "content": "where \\(N\\) is the number of classes in \\(\\mathbb{C}_{\\mathbf{I}}\\), and \\(\\odot\\) denotes the Hadamard product. The predictions \\(\\hat{y}_c\\) and \\(\\hat{y}_p\\) derived from \\(\\hat{\\mathbf{I}}\\) via the ViT are trained through a multi-label soft margin loss \\(\\mathcal{L}_{fss-cls}\\). Additionally, the CAMs \\(\\hat{\\mathbf{M}}\\) generated from \\(\\hat{\\mathbf{I}}\\) and the CAMs \\(\\mathbf{M}\\) derived from the original image \\(\\mathbf{I}\\) are supervised to minimize their difference:" + }, + { + "type": "equation", + "bbox": [ + 0.428, + 0.404, + 0.785, + 0.421 + ], + "angle": 0, + "content": "\\[\n\\mathcal {L} _ {f s s - c a m s} = | \\mathbf {M} - \\hat {\\mathbf {M}} | _ {1} \\tag {10}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.429, + 0.785, + 0.532 + ], + "angle": 0, + "content": "Assuming the model has learned to use a frequency component \\( f \\) as a shortcut for classifying class \\( c \\), the model cannot use this shortcut \\( f \\) for classification when \\( f \\) is reduced in \\( \\hat{\\mathbf{I}} \\). In the absence of the frequency shortcut \\( f \\), the model performs classification based on the semantic understanding of class \\( c \\), resulting in activation within areas related to the object. Simultaneously, by reducing the difference between the two CAMs \\( \\mathbf{M} \\) and \\( \\hat{\\mathbf{M}} \\), false positives caused by shortcuts can be resolved." + }, + { + "type": "text", + "bbox": [ + 0.24, + 0.535, + 0.674, + 0.549 + ], + "angle": 0, + "content": "The FSS is ultimately supervised through the following loss:" + }, + { + "type": "equation", + "bbox": [ + 0.402, + 0.564, + 0.785, + 0.578 + ], + "angle": 0, + "content": "\\[\n\\mathcal {L} _ {f s s} = \\mathcal {L} _ {f s s - c l s} + \\lambda_ {2} \\mathcal {L} _ {f s s - c a m s} \\tag {11}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.215, + 0.591, + 0.65, + 0.605 + ], + "angle": 0, + "content": "where \\(\\lambda_{2}\\) is a hyperparameter to balance between the losses." + }, + { + "type": "text", + "bbox": [ + 0.24, + 0.606, + 0.682, + 0.621 + ], + "angle": 0, + "content": "The final loss of the proposed overall framework is as follows:" + }, + { + "type": "equation", + "bbox": [ + 0.41, + 0.635, + 0.785, + 0.651 + ], + "angle": 0, + "content": "\\[\n\\mathcal {L} _ {\\text {t o t a l}} = \\mathcal {L} _ {\\text {c l s}} + \\mathcal {L} _ {\\text {m p c}} + \\mathcal {L} _ {\\text {f s s}} \\tag {12}\n\\]" + }, + { + "type": "title", + "bbox": [ + 0.215, + 0.674, + 0.376, + 0.692 + ], + "angle": 0, + "content": "4 Experiments" + }, + { + "type": "title", + "bbox": [ + 0.215, + 0.708, + 0.447, + 0.723 + ], + "angle": 0, + "content": "4.1 Experimental Settings" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.734, + 0.785, + 0.84 + ], + "angle": 0, + "content": "Datasets Our proposed method was evaluated using the PASCAL VOC 2012 and MS-COCO 2014 datasets, widely utilized for Weakly Supervised Semantic Segmentation (WSSS). The PASCAL VOC 2012 includes 20 classes of foreground objects along with a background class, and consists of three subsets (train, val, test) containing 10582, 1449, and 1456 images, respectively. Meanwhile, the MS-COCO 2014 is made up of 80 foreground object classes and a background class and is composed of two subsets (82k train set and 40k val set)." + } + ], + [ + { + "type": "page_number", + "bbox": [ + 0.218, + 0.116, + 0.236, + 0.127 + ], + "angle": 0, + "content": "10" + }, + { + "type": "header", + "bbox": [ + 0.272, + 0.115, + 0.371, + 0.128 + ], + "angle": 0, + "content": "H. Kwon et al." + }, + { + "type": "table_caption", + "bbox": [ + 0.214, + 0.145, + 0.788, + 0.187 + ], + "angle": 0, + "content": "Table 1: Ablation study of components on the PASCAL VOC 2012 train set. P: Precision, R: Recall. Values from M are measured under similar R for a fair comparison of P. The best result is represented in Bold." + }, + { + "type": "table", + "bbox": [ + 0.31, + 0.199, + 0.703, + 0.273 + ], + "angle": 0, + "content": "
LclsLfssLmpc(x) M(y) Mrf
P(%)R(%)mIoU(%)P(%)R(%)mIoU(%)
Baseline✓53.279.645.374.580.963.2
(a)✓✓71.379.859.476.582.165.4
(b)✓✓67.180.655.978.882.267.2
(c)✓✓✓76.280.763.979.584.869.5
" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.299, + 0.787, + 0.435 + ], + "angle": 0, + "content": "Evaluation Metric Following prior research [44,63], the mean Intersection over Union (mIoU) metric is used to assess the quality of generated CAMs and the performance of semantic segmentation models trained on CAMs-based pseudolabels. To measure the quality of CAMs, the mIoU with the most optimal threshold is employed by default. Typically, CAMs are evaluated on the train set, whereas the performance of the semantic segmentation model is assessed on the val set. While ground-truth pixel-level labels are accessible for both datasets on the val set and are analyzed locally, the evaluation for the PASCAL VOC 2012 test set is carried out via the official website." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.435, + 0.788, + 0.678 + ], + "angle": 0, + "content": "Implementation Details Our framework utilizes DeiT-S pre-trained on ImageNet [50] as the classification backbone for a fair comparison with ViT-based WSSS methods [44, 63]. The classification network is trained for 60 epochs with a batch size of 64 and an initial learning rate of 5e-4, using the Adam optimizer. We adopted the same data augmentation technique as in MCTformer [63], but with different image resize scales. We crop images to a size of \\(224 \\times 224\\) for equal comparison. In the MPC module, \\(r = 0.3\\) was used as the mixing factor, with MPC applied at ViT layer \\(\\ell = 8\\). To measure the influence of each frequency for FSPM, 100 images were used per class for all classes. Additionally, the FSPM was updated every three epochs through the FIM process. Min-max normalization was utilized in FIM to normalize \\(\\mathbf{P}\\), with the highest measured loss per class serving as the normalization maximum. For balancing each loss relative to \\(\\mathcal{L}_{cls}\\), \\(\\lambda_{1}\\) and \\(\\lambda_{2}\\) were both set to 2. For the semantic segmentation model training on the PASCAL VOC 2012 dataset, Deeplab-V1 with ResNet38 was employed, while Deeplab-V2 with ResNet101 was used for the MS COCO 2014 dataset. Additional training details are in Supplementary Materials." + }, + { + "type": "title", + "bbox": [ + 0.215, + 0.697, + 0.403, + 0.71 + ], + "angle": 0, + "content": "4.2 Ablation Studies" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.72, + 0.788, + 0.841 + ], + "angle": 0, + "content": "Component Analysis To demonstrate the significance of each module proposed in our framework, we ablated each component as shown in Tab. 1. Observing the performance of refined CAMs in Tab. 1-y, introducing the Frequency Shortcut Suppression (FSS) and Magnitude-mixing-based Phase Concentration (MPC) led to the gain of \\(2.2\\% \\mathrm{p}\\) and \\(4.0\\% \\mathrm{p}\\) respectively over baseline. Our method achieved \\(69.5\\%\\) mIoU by adopting both components. FSS still resulted in a gain of \\(2.3\\% \\mathrm{p}\\) in Tab. 1-c compared to Tab. 1-b, implies that FSS and MPC offer distinct advantages." + } + ], + [ + { + "type": "header", + "bbox": [ + 0.348, + 0.115, + 0.733, + 0.129 + ], + "angle": 0, + "content": "Phase Concentration and Shortcut Suppression for WSSS" + }, + { + "type": "page_number", + "bbox": [ + 0.768, + 0.116, + 0.784, + 0.127 + ], + "angle": 0, + "content": "11" + }, + { + "type": "table_caption", + "bbox": [ + 0.215, + 0.145, + 0.785, + 0.173 + ], + "angle": 0, + "content": "Table 2: Ablation study of ViT layer index that MPC performed. Peak performance is observed at layer 8, which is represented in Bold." + }, + { + "type": "table", + "bbox": [ + 0.377, + 0.186, + 0.632, + 0.227 + ], + "angle": 0, + "content": "
earlylayermiddlelayerlatelayer
l34781011
mIoU(%)68.068.168.669.567.866.4
" + }, + { + "type": "image", + "bbox": [ + 0.222, + 0.246, + 0.317, + 0.342 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.241, + 0.343, + 0.297, + 0.354 + ], + "angle": 0, + "content": "(a) Image" + }, + { + "type": "image", + "bbox": [ + 0.318, + 0.247, + 0.408, + 0.342 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.342, + 0.343, + 0.382, + 0.353 + ], + "angle": 0, + "content": "(b) GT" + }, + { + "type": "image", + "bbox": [ + 0.41, + 0.247, + 0.5, + 0.342 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.422, + 0.343, + 0.489, + 0.353 + ], + "angle": 0, + "content": "(c) Baseline" + }, + { + "type": "image", + "bbox": [ + 0.502, + 0.247, + 0.594, + 0.342 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.527, + 0.343, + 0.571, + 0.353 + ], + "angle": 0, + "content": "(d) FSS" + }, + { + "type": "image", + "bbox": [ + 0.595, + 0.247, + 0.687, + 0.342 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.618, + 0.343, + 0.666, + 0.353 + ], + "angle": 0, + "content": "(e) MPC" + }, + { + "type": "image", + "bbox": [ + 0.688, + 0.247, + 0.781, + 0.342 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.697, + 0.343, + 0.774, + 0.353 + ], + "angle": 0, + "content": "(f) FSS+MPC" + }, + { + "type": "image_caption", + "bbox": [ + 0.215, + 0.366, + 0.785, + 0.407 + ], + "angle": 0, + "content": "Fig. 4: Comparison of refined CAMs \\(\\mathbf{M}_{ref}\\) between baseline and components in ours. (a) Image, (b) Ground Truth, (c) Baseline, (d) only FSS, (e) only MPC, (f) ours. The yellow and red boxes imply effect of FSS and MPC, respectively." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.443, + 0.785, + 0.548 + ], + "angle": 0, + "content": "To understand the benefits of FSS in CAMs, we examined the results of ViT CAMs that are not refined by the attention-map in Tab. 1-x. To conduct a comparative analysis considering the Precision-Recall tradeoff, thresholds corresponding to similar Recall levels were used. Low precision of baseline \\(\\mathbf{M}\\) indicates a significant number of over-activation. Through FSS, we observed a remarkable improvement in precision \\((18.1\\%)\\) and mIoU \\((14.1\\%)\\) compared to the baseline, confirming that false positives were successfully resolved (Tab. 1-b)." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.551, + 0.785, + 0.656 + ], + "angle": 0, + "content": "Ablation Studies of MPC The qualitative comparison between the refined CAMs of baseline and MPC in Fig. 4-e demonstrates the efficacy of the MPC by clearly delineating the boundary of the object. Meanwhile, we investigate the impact of the ratio \\( r \\) in MPC, which is used to preserve the magnitude information of the anchor feature. In experiments conducted within a search space below 0.5, there was a minimum gain of \\( 2.3\\% p \\) when compared to applying FSS alone, with the best performance observed at \\( r = 0.3 \\)" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.659, + 0.787, + 0.84 + ], + "angle": 0, + "content": "The influence of the layer index \\(\\ell\\) where the MPC is performed was analyzed by conducting experiments on early, middle, and late layers in Tab. 2. The performance was highest when \\(\\ell\\) was set to 8, with a noticeable decline observed in layers beyond this point. This tendency supports the knowledge that later layers of the ViT play a crucial role in capturing high-level semantic information [9]. Introducing changes to the magnitude while preserving the phase information at these later stages could disrupt the semantic analysis of the model. This disruption may stem from the difference between the coupled information used in previous layers and the re-coupled information with the magnitude alteration. This result shows that MPC applied before when ViT features highly condense to high-level semantics can encourage the model to capture boundary information well in the phase spectrum." + } + ], + [ + { + "type": "page_number", + "bbox": [ + 0.218, + 0.116, + 0.236, + 0.127 + ], + "angle": 0, + "content": "12" + }, + { + "type": "header", + "bbox": [ + 0.272, + 0.115, + 0.371, + 0.128 + ], + "angle": 0, + "content": "H. Kwon et al." + }, + { + "type": "table_caption", + "bbox": [ + 0.214, + 0.145, + 0.788, + 0.228 + ], + "angle": 0, + "content": "Table 3: (a) True/False Positive Rate (TPR/FPR) of classification where the high-frequency components following FSPM are retained in images. Metrics are measured under the early training stage of the Baseline. **Bold** indicates the shortcut occurred, and **Underline** represents that model is trying to make a shortcut. (b) Comparative analysis of the Precision of CAMs M between baseline and only with FSS. Precision is evaluated under a similar Recall considering the Precision-Recall trade-off." + }, + { + "type": "table", + "bbox": [ + 0.221, + 0.242, + 0.788, + 0.378 + ], + "angle": 0, + "content": "
bkgaerobikebirdboatbottlebuscarcatchaircow
(a)BaselineTPR(%)-58.9636.4388.4753.3668.4650.8379.2668.3555.4817.87
FPR(%)-00.2300.4115.7800.0533.7400.0001.4200.5307.0400.41
(b)BaselineP(%)91.939.224.817.318.958.176.542.880.944.959.7
w/ FSSP(%)92.552.942.259.233.371.187.562.994.143.290.7
tabledoghorsembkpersonplantsheepsofatraintvtotal
(a)BaselineTPR(%)04.7661.2479.4060.2675.4741.8843.6145.4592.5157.23-
FPR(%)00.0001.9700.6400.1605.5600.1500.3800.6300.6800.13-
(b)BaselineP(%)68.666.050.468.658.039.956.479.039.237.053.2
w/ FSSP(%)89.687.579.778.582.764.792.581.254.556.671.3
" + }, + { + "type": "image", + "bbox": [ + 0.245, + 0.396, + 0.32, + 0.526 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.321, + 0.397, + 0.374, + 0.526 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.375, + 0.397, + 0.453, + 0.526 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.453, + 0.397, + 0.537, + 0.526 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.544, + 0.397, + 0.585, + 0.526 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.585, + 0.397, + 0.649, + 0.526 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.648, + 0.397, + 0.757, + 0.526 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.344, + 0.527, + 0.445, + 0.536 + ], + "angle": 0, + "content": "PASCAL VOC 2012" + }, + { + "type": "image_caption", + "bbox": [ + 0.611, + 0.527, + 0.694, + 0.536 + ], + "angle": 0, + "content": "MS COCO 2014" + }, + { + "type": "image_caption", + "bbox": [ + 0.214, + 0.548, + 0.788, + 0.577 + ], + "angle": 0, + "content": "Fig.5: Semantic segmentation results on PASCAL VOC 2012 (left) and MS COCO 2014 (right) datasets. From top to bottom: Image, Ours, GT." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.611, + 0.788, + 0.717 + ], + "angle": 0, + "content": "Frequency Shortcut Analysis and Result of Suppression In Tab. 3, we investigated whether shortcut learning occurs during the model training process in early steps and analyzed the correlation with the improvement in CAMs precision facilitated by the FSS. To conduct this analysis, classification influence according to frequency component for each class was first measured using Eq. 7. Based on this measurement, the classification behavior of the model was analyzed when only some high frequencies of high-influence components were retained." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.72, + 0.788, + 0.841 + ], + "angle": 0, + "content": "A high TPR indicates that class \\( c \\) possesses the principal frequency components utilized in the experiment (i.e., bird will be correctly classified as bird only with the high-influence frequency of bird). This suggests that the model has identified a straightforward method for classifying class \\( c \\) and is attempting to find frequency shortcuts. To consider a principal frequency as a shortcut, it should direct images to the respective class regardless of the actual semantic content in the image (i.e. car with frequency shortcut of bird will falsely classify as bird). This is expressed in Tab. 3 through high TPR and FPR, which repre" + } + ], + [ + { + "type": "header", + "bbox": [ + 0.348, + 0.115, + 0.732, + 0.129 + ], + "angle": 0, + "content": "Phase Concentration and Shortcut Suppression for WSSS" + }, + { + "type": "page_number", + "bbox": [ + 0.768, + 0.116, + 0.786, + 0.127 + ], + "angle": 0, + "content": "13" + }, + { + "type": "table_caption", + "bbox": [ + 0.214, + 0.145, + 0.788, + 0.201 + ], + "angle": 0, + "content": "Table 4: Comparison between ours and multi-stage WSSS methods. mIoU(%) is evaluated on PASCAL VOC 2012 train set in both CAMs (seed) and Pseudo-GT (Mask). The backbone and applied post-processing methods (PSA [2], IRN [1]) are listed for a fair comparison. Bold numbers represent the best performance." + }, + { + "type": "table", + "bbox": [ + 0.334, + 0.213, + 0.675, + 0.387 + ], + "angle": 0, + "content": "
MethodBackboneSeedPostMask
AdvCAM [28] CVPR'21RN5055.6IRN69.9
OC-CSE [25] ICCV'21WRN3856.0PSA66.9
ECS [53] ICCV'21WRN3856.6PSA67.8
CPN [70] ICCV'21WRN3857.4PSA67.8
AMR [45] AAAI'22RN5056.8PSA69.7
ReCAM [7] CVPR'22RN5054.8IRN70.5
SIPE [6] CVPR'22RN5058.6IRN-
PPC [10] CVPR'22WRN3861.5IRN70.1
AEFT [67] ECCV'22WRN3856.0PSA71.0
ACR [26] CVPR'23WRN3860.3IRN72.3
MCT [63] CVPR'22DeiT-S61.7PSA69.1
USAGE [44] ICCV'23DeiT-S67.7PSA72.8
OursDeiT-S69.5PSA73.2
" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.427, + 0.788, + 0.502 + ], + "angle": 0, + "content": "sent the rates in which principal frequencies of images are correctly and falsely classified to the shortcut respective class \\( c \\). Examining Tab. 3-a, we can observe that class bird and bottle exhibit high TPR and FPR, indicating the occurrence of shortcuts. Meanwhile, classes such as car, horse, and train show high TPR, suggesting attempts at creating shortcuts." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.503, + 0.789, + 0.669 + ], + "angle": 0, + "content": "Tab. 3-b records the class-specific precision of CAMs \\(\\mathbf{M}\\) for the baseline model and after the introduction of the FSS. To ensure a fair comparison while considering the Precision-Recall trade-off, precision was measured at similar levels of Recall. The introduction of FSS to disrupt shortcut creation resulted in an average increase of \\(24.0\\%\\) in precision for \\(\\mathbf{M}\\) across classes where the baseline model had created or attempted to create shortcuts. This increase confirms that the FSS effectively reduces false positives. Notably, for bird, a class where shortcuts had occurred, the precision of CAMs surged by \\(42.1\\%\\) , showcasing the resolve of over-activation in areas unrelated to the object by FSS. Qualitative comparison of refined CAMs \\(\\mathbf{M}_{ref}\\) before and after the introduction of FSS in Fig. 4-d further validates the effectiveness of FSS." + }, + { + "type": "title", + "bbox": [ + 0.215, + 0.693, + 0.557, + 0.709 + ], + "angle": 0, + "content": "4.3 Comparison with State-of-The-Arts" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.72, + 0.788, + 0.841 + ], + "angle": 0, + "content": "To train the semantic segmentation model, the CAMs generated by our method were refined using PSA [2] as performed in previous WSSS works [44, 63, 67]. In Tab. 4, we compared the performance of CAMs (Seed) and pseudo-ground-truth (Mask) on the PASCAL VOC 2012 train set, where our method achieved the best performance in both categories. Utilizing our high-quality pseudo labels for training the semantic segmentation model, we compared the results with other WSSS methods in Tab. 5. Our framework demonstrated superiority by achieving the highest performance on the PASCAL VOC 2012 val and test sets." + } + ], + [ + { + "type": "page_number", + "bbox": [ + 0.218, + 0.116, + 0.236, + 0.127 + ], + "angle": 0, + "content": "14" + }, + { + "type": "header", + "bbox": [ + 0.272, + 0.115, + 0.371, + 0.128 + ], + "angle": 0, + "content": "H. Kwon et al." + }, + { + "type": "table_caption", + "bbox": [ + 0.216, + 0.145, + 0.788, + 0.201 + ], + "angle": 0, + "content": "Table 5: Comparison of semantic segmentation performance with multi-stage WSSS methods. Performance (mIoU) is evaluated in PASCAL VOC 2012 and MS COCO 2014 datasets. CNN-based methods are represented above the horizontal line, and ViT-based methods are depicted below the line. Bold represents the best performance." + }, + { + "type": "table", + "bbox": [ + 0.3, + 0.213, + 0.705, + 0.407 + ], + "angle": 0, + "content": "
MethodBackboneVOC valVOC testCOCO val
AdvCAM [28] CVPR'21RN10168.168.0-
OC-CSE [25] ICCV'21WRN3868.468.236.4
ECS [53] ICCV'21WRN3866.667.6-
CPN [70] ICCV'21WRN3867.868.5-
AMR [45] AAAT'22RN10168.869.1-
ReCAM [7] CVPR'22RN10168.568.442.9
SIPE [6] CVPR'22RN10168.869.7-
SIPE [6] CVPR'22WRN38--43.6
AEFT [67] ECCV'22WRN3870.971.744.8
ACR [26] CVPR'23WRN3871.971.945.3
MCT [63] CVPR'22WRN3871.971.642.0
USAGE [44] ICCV'23WRN3871.972.842.7
USAGE [44] ICCV'23RN101--44.3
OursWRN3873.273.0-
OursRN101--45.7
" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.437, + 0.788, + 0.53 + ], + "angle": 0, + "content": "While ViT-based WSSS works tend to outperform CNN-based works on the PASCAL VOC 2012 dataset, they showed lower performance on the MS COCO 2014 dataset due to activation overlapping between classes. Nonetheless, our method achieved the SoTA performance on the MS COCO 2014 dataset by effectively reducing false positives. Additional CAMs and semantic segmentation results are in the Supplementary Materials." + }, + { + "type": "title", + "bbox": [ + 0.216, + 0.559, + 0.36, + 0.575 + ], + "angle": 0, + "content": "5 Conclusion" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.598, + 0.789, + 0.841 + ], + "angle": 0, + "content": "Weakly Supervised Semantic Segmentation suffers from CAMs that have imprecise activation around boundary areas and object-unrelated activations. In this paper, we aim to address these challenges by applying a novel perspective in the frequency domain, leveraging the Fourier Transform. Firstly, we propose Magnitude-mixing-based Phase Concentration (MPC), which generates features with identical phase but different magnitude at the intermediate layer, encouraging the model to capture the same semantics from these features. This approach leads the model to pay more attention to shape information in the phase, thereby learning to activate boundaries more clearly. Additionally, we identify the occurrence of frequency shortcuts and their association with the over-activation of CAMs. To resolve this, we introduce Frequency Shortcut Suppression (FSS), which discourages the creation of frequency shortcuts, successfully resolving object-unrelated activations. Our experimental results support the utility of the proposed method, demonstrating its effectiveness in addressing the aforementioned issues in WSSS. Furthermore, we have achieved new state-of-the-art performance on the PASCAL VOC and MS COCO datasets." + } + ], + [ + { + "type": "header", + "bbox": [ + 0.348, + 0.115, + 0.732, + 0.129 + ], + "angle": 0, + "content": "Phase Concentration and Shortcut Suppression for WSSS" + }, + { + "type": "page_number", + "bbox": [ + 0.768, + 0.117, + 0.786, + 0.127 + ], + "angle": 0, + "content": "15" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.147, + 0.788, + 0.239 + ], + "angle": 0, + "content": "Acknowledgement This work was supported by the Technology Innovation Program (1415187329,20024355, Development of autonomous driving connectivity technology based on sensor-infrastructure cooperation) funded by the Ministry of Trade, Industry & Energy(MOTIE, Korea) and the National Research Foundation of Korea(NRF) grant funded by the Korea government(MSIT) (NRF2022R1A2B5B03002636)." + }, + { + "type": "title", + "bbox": [ + 0.216, + 0.257, + 0.323, + 0.273 + ], + "angle": 0, + "content": "References" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.283, + 0.785, + 0.325 + ], + "angle": 0, + "content": "1. Ahn, J., Cho, S., Kwak, S.: Weakly supervised learning of instance segmentation with inter-pixel relations. In: Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition. pp. 2209-2218 (2019)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.326, + 0.785, + 0.365 + ], + "angle": 0, + "content": "2. Ahn, J., Kwak, S.: Learning pixel-level semantic affinity with image-level supervision for weakly supervised semantic segmentation. In: Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition. pp. 4981-4990 (2018)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.365, + 0.785, + 0.42 + ], + "angle": 0, + "content": "3. Chang, Y.T., Wang, Q., Hung, W.C., Piramuthu, R., Tsai, Y.H., Yang, M.H.: Weakly-supervised semantic segmentation via sub-category exploration. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 8991-9000 (2020)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.421, + 0.785, + 0.473 + ], + "angle": 0, + "content": "4. Chen, G., Peng, P., Ma, L., Li, J., Du, L., Tian, Y.: Amplitude-phase recombination: Rethinking robustness of convolutional neural networks in frequency domain. In: Proceedings of the IEEE/CVF International Conference on Computer Vision (ICCV). pp. 458-467 (October 2021)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.475, + 0.785, + 0.514 + ], + "angle": 0, + "content": "5. Chen, L., Lei, C., Li, R., Li, S., Zhang, Z., Zhang, L.: Fpr: False positive rectification for weakly supervised semantic segmentation. In: Proceedings of the IEEE/CVF International Conference on Computer Vision. pp. 1108-1118 (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.515, + 0.785, + 0.569 + ], + "angle": 0, + "content": "6. Chen, Q., Yang, L., Lai, J.H., Xie, X.: Self-supervised image-specific prototype exploration for weakly supervised semantic segmentation. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 4288-4298 (2022)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.569, + 0.785, + 0.623 + ], + "angle": 0, + "content": "7. Chen, Z., Wang, T., Wu, X., Hua, X.S., Zhang, H., Sun, Q.: Class reactivation maps for weakly-supervised semantic segmentation. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 969-978 (2022)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.624, + 0.785, + 0.663 + ], + "angle": 0, + "content": "8. Dai, J., He, K., Sun, J.: Boxsup: Exploiting bounding boxes to supervise convolutional networks for semantic segmentation. In: Proceedings of the IEEE international conference on computer vision. pp. 1635-1643 (2015)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.664, + 0.785, + 0.718 + ], + "angle": 0, + "content": "9. Dosovitskiy, A., Beyer, L., Kolesnikov, A., Weissenborn, D., Zhai, X., Unterthiner, T., Dehghani, M., Minderer, M., Heigold, G., Gelly, S., et al.: An image is worth 16x16 words: Transformers for image recognition at scale. arXiv preprint arXiv:2010.11929 (2020)" + }, + { + "type": "ref_text", + "bbox": [ + 0.22, + 0.719, + 0.785, + 0.759 + ], + "angle": 0, + "content": "10. Du, Y., Fu, Z., Liu, Q., Wang, Y.: Weakly supervised semantic segmentation by pixel-to-prototype contrast. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 4320-4329 (2022)" + }, + { + "type": "ref_text", + "bbox": [ + 0.22, + 0.759, + 0.785, + 0.799 + ], + "angle": 0, + "content": "11. Everingham, M., Van Gool, L., Williams, C.K., Winn, J., Zisserman, A.: The pascal visual object classes (voc) challenge. International journal of computer vision 88(2), 303-338 (2010)" + }, + { + "type": "ref_text", + "bbox": [ + 0.22, + 0.799, + 0.785, + 0.84 + ], + "angle": 0, + "content": "12. Fan, J., Zhang, Z., Tan, T., Song, C., Xiao, J.: Cian: Cross-image affinity net for weakly supervised semantic segmentation. In: Proceedings of the AAAI Conference on Artificial Intelligence. vol. 34, pp. 10762-10769 (2020)" + }, + { + "type": "list", + "bbox": [ + 0.22, + 0.283, + 0.785, + 0.84 + ], + "angle": 0, + "content": null + } + ], + [ + { + "type": "page_number", + "bbox": [ + 0.218, + 0.116, + 0.236, + 0.127 + ], + "angle": 0, + "content": "16" + }, + { + "type": "header", + "bbox": [ + 0.272, + 0.115, + 0.371, + 0.128 + ], + "angle": 0, + "content": "H. Kwon et al." + }, + { + "type": "ref_text", + "bbox": [ + 0.218, + 0.147, + 0.787, + 0.203 + ], + "angle": 0, + "content": "13. Gao, W., Wan, F., Pan, X., Peng, Z., Tian, Q., Han, Z., Zhou, B., Ye, Q.: Ts-cam: Token semantic coupled attention map for weakly supervised object localization. In: Proceedings of the IEEE/CVF International Conference on Computer Vision. pp. 2886-2895 (2021)" + }, + { + "type": "ref_text", + "bbox": [ + 0.218, + 0.203, + 0.787, + 0.243 + ], + "angle": 0, + "content": "14. Geirhos, R., Jacobsen, J.H., Michaelis, C., Zemel, R., Brendel, W., Bethge, M., Wichmann, F.A.: Shortcut learning in deep neural networks. Nature Machine Intelligence 2(11), 665-673 (2020)" + }, + { + "type": "ref_text", + "bbox": [ + 0.218, + 0.243, + 0.787, + 0.298 + ], + "angle": 0, + "content": "15. Guo, J., Wang, N., Qi, L., Shi, Y.: Aloft: A lightweight mlp-like architecture with dynamic low-frequency transform for domain generalization. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 24132-24141 (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.218, + 0.298, + 0.787, + 0.325 + ], + "angle": 0, + "content": "16. Hansen, B.C., Hess, R.F.: Structural sparseness and spatial phase alignment in natural scenes. JOSA A 24(7), 1873-1885 (2007)" + }, + { + "type": "ref_text", + "bbox": [ + 0.218, + 0.325, + 0.787, + 0.379 + ], + "angle": 0, + "content": "17. Huang, H., Kang, H., Liu, S., Salvado, O., Rakotoarivelo, T., Wang, D., Liu, T.: Paddles: Phase-amplitude spectrum disentangled early stopping for learning with noisy labels. In: Proceedings of the IEEE/CVF International Conference on Computer Vision. pp. 16719-16730 (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.218, + 0.379, + 0.787, + 0.434 + ], + "angle": 0, + "content": "18. Huang, H., Xie, S., Lin, L., Tong, R., Chen, Y.W., Li, Y., Wang, H., Huang, Y., Zheng, Y.: Semicvt: Semi-supervised convolutional vision transformer for semantic segmentation. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 11340-11349 (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.218, + 0.434, + 0.787, + 0.473 + ], + "angle": 0, + "content": "19. Huang, J., Guan, D., Xiao, A., Lu, S.: Rda: Robust domain adaptation via fourier adversarial attacking. In: Proceedings of the IEEE/CVF international conference on computer vision. pp. 8988-8999 (2021)" + }, + { + "type": "ref_text", + "bbox": [ + 0.218, + 0.473, + 0.787, + 0.515 + ], + "angle": 0, + "content": "20. Huang, J., Liu, Y., Zhao, F., Yan, K., Zhang, J., Huang, Y., Zhou, M., Xiong, Z.: Deep fourier-based exposure correction network with spatial-frequency interaction. In: European Conference on Computer Vision. pp. 163-180. Springer (2022)" + }, + { + "type": "ref_text", + "bbox": [ + 0.218, + 0.515, + 0.787, + 0.555 + ], + "angle": 0, + "content": "21. Jiang, P.T., Han, L.H., Hou, Q., Cheng, M.M., Wei, Y.: Online attention accumulation for weakly supervised semantic segmentation. IEEE Transactions on Pattern Analysis and Machine Intelligence 44(10), 7062-7077 (2021)" + }, + { + "type": "ref_text", + "bbox": [ + 0.218, + 0.555, + 0.787, + 0.609 + ], + "angle": 0, + "content": "22. Jiang, P.T., Yang, Y., Hou, Q., Wei, Y.: L2g: A simple local-to-global knowledge transfer framework for weakly supervised semantic segmentation. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 16886-16896 (2022)" + }, + { + "type": "ref_text", + "bbox": [ + 0.218, + 0.609, + 0.787, + 0.65 + ], + "angle": 0, + "content": "23. Jo, S., Yu, I.J., Kim, K.: Mars: Model-agnostic biased object removal without additional supervision for weakly-supervised semantic segmentation. arXiv preprint arXiv:2304.09913 (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.218, + 0.65, + 0.787, + 0.69 + ], + "angle": 0, + "content": "24. Kim, G., Kim, J., Lee, J.S.: Exploring adversarial robustness of vision transformers in the spectral perspective. In: Proceedings of the IEEE/CVF Winter Conference on Applications of Computer Vision. pp. 3976-3985 (2024)" + }, + { + "type": "ref_text", + "bbox": [ + 0.218, + 0.69, + 0.787, + 0.746 + ], + "angle": 0, + "content": "25. Kweon, H., Yoon, S.H., Kim, H., Park, D., Yoon, K.J.: Unlocking the potential of ordinary classifier: Class-specific adversarial erasing framework for weakly supervised semantic segmentation. In: Proceedings of the IEEE/CVF International Conference on Computer Vision. pp. 6994-7003 (2021)" + }, + { + "type": "ref_text", + "bbox": [ + 0.218, + 0.746, + 0.787, + 0.799 + ], + "angle": 0, + "content": "26. Kweon, H., Yoon, S.H., Yoon, K.J.: Weakly supervised semantic segmentation via adversarial learning of classifier and reconstructor. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 11329-11339 (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.218, + 0.799, + 0.787, + 0.841 + ], + "angle": 0, + "content": "27. Lee, J., Choi, J., Mok, J., Yoon, S.: Reducing information bottleneck for weakly supervised semantic segmentation. Advances in Neural Information Processing Systems 34, 27408-27421 (2021)" + }, + { + "type": "list", + "bbox": [ + 0.218, + 0.147, + 0.787, + 0.841 + ], + "angle": 0, + "content": null + } + ], + [ + { + "type": "header", + "bbox": [ + 0.349, + 0.115, + 0.732, + 0.129 + ], + "angle": 0, + "content": "Phase Concentration and Shortcut Suppression for WSSS" + }, + { + "type": "page_number", + "bbox": [ + 0.768, + 0.117, + 0.786, + 0.127 + ], + "angle": 0, + "content": "17" + }, + { + "type": "ref_text", + "bbox": [ + 0.214, + 0.147, + 0.788, + 0.19 + ], + "angle": 0, + "content": "28. Lee, J., Kim, E., Yoon, S.: Anti-adversarily manipulated attributions for weakly and semi-supervised semantic segmentation. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 4071-4080 (2021)" + }, + { + "type": "ref_text", + "bbox": [ + 0.214, + 0.19, + 0.788, + 0.233 + ], + "angle": 0, + "content": "29. Lee, J., Oh, S.J., Yun, S., Choe, J., Kim, E., Yoon, S.: Weakly supervised semantic segmentation using out-of-distribution data. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 16897-16906 (2022)" + }, + { + "type": "ref_text", + "bbox": [ + 0.214, + 0.233, + 0.788, + 0.275 + ], + "angle": 0, + "content": "30. Lee, J., Yi, J., Shin, C., Yoon, S.: Bbam: Bounding box attribution map for weakly supervised semantic and instance segmentation. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 2643-2652 (2021)" + }, + { + "type": "ref_text", + "bbox": [ + 0.214, + 0.276, + 0.788, + 0.33 + ], + "angle": 0, + "content": "31. Lee, S., Bae, J., Kim, H.Y.: Decompose, adjust, compose: Effective normalization by playing with frequency for domain generalization. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 11776-11785 (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.214, + 0.33, + 0.788, + 0.374 + ], + "angle": 0, + "content": "32. Li, J., Fan, J., Zhang, Z.: Towards noiseless object contours for weakly supervised semantic segmentation. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 16856-16865 (2022)" + }, + { + "type": "ref_text", + "bbox": [ + 0.214, + 0.374, + 0.788, + 0.416 + ], + "angle": 0, + "content": "33. Li, K., Wu, Z., Peng, K.C., Ernst, J., Fu, Y.: Tell me where to look: Guided attention inference network. In: Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition. pp. 9215-9223 (2018)" + }, + { + "type": "ref_text", + "bbox": [ + 0.214, + 0.416, + 0.788, + 0.445 + ], + "angle": 0, + "content": "34. Li, X., Zhou, T., Li, J., Zhou, Y., Zhang, Z.: Group-wise semantic mining for weakly supervised semantic segmentation. arXiv preprint arXiv:2012.05007 (2020)" + }, + { + "type": "ref_text", + "bbox": [ + 0.214, + 0.445, + 0.788, + 0.487 + ], + "angle": 0, + "content": "35. Li, Y., Duan, Y., Kuang, Z., Chen, Y., Zhang, W., Li, X.: Uncertainty estimation via response scaling for pseudo-mask noise mitigation in weakly-supervised semantic segmentation. arXiv preprint arXiv:2112.07431 (2021)" + }, + { + "type": "ref_text", + "bbox": [ + 0.214, + 0.487, + 0.788, + 0.53 + ], + "angle": 0, + "content": "36. Li, Y., Kuang, Z., Liu, L., Chen, Y., Zhang, W.: Pseudo-mask matters in weakly-supervised semantic segmentation. In: Proceedings of the IEEE/CVF International Conference on Computer Vision. pp. 6964-6973 (2021)" + }, + { + "type": "ref_text", + "bbox": [ + 0.214, + 0.53, + 0.788, + 0.572 + ], + "angle": 0, + "content": "37. Lin, D., Dai, J., Jia, J., He, K., Sun, J.: Scribblesup: Scribble-supervised convolutional networks for semantic segmentation. In: Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition. pp. 3159-3167 (2016)" + }, + { + "type": "ref_text", + "bbox": [ + 0.214, + 0.572, + 0.788, + 0.614 + ], + "angle": 0, + "content": "38. Lin, T.Y., Maire, M., Belongie, S., Hays, J., Perona, P., Ramanan, D., Dollar, P., Zitnick, C.L.: Microsoft coco: Common objects in context. In: European conference on computer vision. pp. 740-755. Springer (2014)" + }, + { + "type": "ref_text", + "bbox": [ + 0.214, + 0.614, + 0.788, + 0.671 + ], + "angle": 0, + "content": "39. Lin, Y., Chen, M., Wang, W., Wu, B., Li, K., Lin, B., Liu, H., He, X.: Clip is also an efficient segmenter: A text-driven approach for weakly supervised semantic segmentation. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 15305-15314 (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.214, + 0.671, + 0.788, + 0.713 + ], + "angle": 0, + "content": "40. Luo, X., Wei, L., Wen, L., Yang, J., Xie, L., Xu, Z., Tian, Q.: Rectifying the short-cut learning of background for few-shot learning. Advances in Neural Information Processing Systems 34, 13073-13085 (2021)" + }, + { + "type": "ref_text", + "bbox": [ + 0.214, + 0.713, + 0.788, + 0.756 + ], + "angle": 0, + "content": "41. Ma, C., Zhao, L., Chen, Y., Guo, L., Zhang, T., Hu, X., Shen, D., Jiang, X., Liu, T.: Rectify vit shortcut learning by visual saliency. IEEE Transactions on Neural Networks and Learning Systems (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.214, + 0.756, + 0.788, + 0.785 + ], + "angle": 0, + "content": "42. Oppenheim, A.V., Lim, J.S.: The importance of phase in signals. Proceedings of the IEEE 69(5), 529-541 (1981)" + }, + { + "type": "ref_text", + "bbox": [ + 0.214, + 0.785, + 0.788, + 0.84 + ], + "angle": 0, + "content": "43. Papandreou, G., Chen, L.C., Murphy, K.P., Yuille, A.L.: Weakly-and semi-supervised learning of a deep convolutional network for semantic image segmentation. In: Proceedings of the IEEE international conference on computer vision. pp. 1742-1750 (2015)" + }, + { + "type": "list", + "bbox": [ + 0.214, + 0.147, + 0.788, + 0.84 + ], + "angle": 0, + "content": null + } + ], + [ + { + "type": "page_number", + "bbox": [ + 0.218, + 0.116, + 0.236, + 0.127 + ], + "angle": 0, + "content": "18" + }, + { + "type": "header", + "bbox": [ + 0.272, + 0.115, + 0.371, + 0.128 + ], + "angle": 0, + "content": "H. Kwon et al." + }, + { + "type": "ref_text", + "bbox": [ + 0.218, + 0.147, + 0.787, + 0.189 + ], + "angle": 0, + "content": "44. Peng, Z., Wang, G., Xie, L., Jiang, D., Shen, W., Tian, Q.: Usage: A unified seed area generation paradigm for weakly supervised semantic segmentation. arXiv preprint arXiv:2303.07806 (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.218, + 0.19, + 0.786, + 0.232 + ], + "angle": 0, + "content": "45. Qin, J., Wu, J., Xiao, X., Li, L., Wang, X.: Activation modulation and recalibration scheme for weakly supervised semantic segmentation. arXiv preprint arXiv:2112.08996 (2021)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.232, + 0.787, + 0.272 + ], + "angle": 0, + "content": "46. Robinson, J., Sun, L., Yu, K., Batmanghelich, K., Jegelka, S., Sra, S.: Can contrastive learning avoid shortcut solutions? Advances in neural information processing systems 34, 4974-4986 (2021)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.273, + 0.787, + 0.327 + ], + "angle": 0, + "content": "47. Rossetti, S., Zappia, D., Sanzari, M., Schaerf, M., Pirri, F.: Max pooling with vision transformers reconciles class and shape in weakly supervised semantic segmentation. In: European Conference on Computer Vision. pp. 446-463. Springer (2022)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.328, + 0.787, + 0.383 + ], + "angle": 0, + "content": "48. Ru, L., Zhan, Y., Yu, B., Du, B.: Learning affinity from attention: End-to-end weakly-supervised semantic segmentation with transformers. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 16846-16855 (2022)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.384, + 0.787, + 0.425 + ], + "angle": 0, + "content": "49. Ru, L., Zheng, H., Zhan, Y., Du, B.: Token contrast for weakly-supervised semantic segmentation. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 3093-3102 (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.425, + 0.787, + 0.467 + ], + "angle": 0, + "content": "50. Russakovsky, O., Deng, J., Su, H., Krause, J., Satheesh, S., Ma, S., Huang, Z., Karpathy, A., Khosla, A., Bernstein, M., et al.: Imagenet large scale visual recognition challenge. International journal of computer vision 115(3), 211-252 (2015)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.467, + 0.787, + 0.494 + ], + "angle": 0, + "content": "51. Su, Y., Sun, R., Lin, G., Wu, Q.: Context decoupling augmentation for weakly supervised semantic segmentation. arXiv preprint arXiv:2103.01795 (2021)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.495, + 0.787, + 0.522 + ], + "angle": 0, + "content": "52. Sun, G., Wang, W., Dai, J., Van Gool, L.: Mining cross-image semantics for weakly supervised semantic segmentation. arXiv preprint arXiv:2007.01947 (2020)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.522, + 0.787, + 0.577 + ], + "angle": 0, + "content": "53. Sun, K., Shi, H., Zhang, Z., Huang, Y.: ECS-net: Improving weakly supervised semantic segmentation by using connections between class activation maps. In: Proceedings of the IEEE/CVF International Conference on Computer Vision. pp. 7283-7292 (2021)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.578, + 0.787, + 0.605 + ], + "angle": 0, + "content": "54. Sundararajan, D.: The discrete Fourier transform: theory, algorithms and applications. World Scientific (2001)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.606, + 0.787, + 0.646 + ], + "angle": 0, + "content": "55. Vaswani, A., Shazeer, N., Parmar, N., Uszkoreit, J., Jones, L., Gomez, A.N., Kaiser, L., Polosukhin, I.: Attention is all you need. Advances in neural information processing systems 30 (2017)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.647, + 0.787, + 0.688 + ], + "angle": 0, + "content": "56. Vernaza, P., Chandraker, M.: Learning random-walk label propagation for weakly-supervised semantic segmentation. In: Proceedings of the IEEE conference on computer vision and pattern recognition. pp. 7158-7166 (2017)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.689, + 0.787, + 0.73 + ], + "angle": 0, + "content": "57. Wang, C., Xu, R., Xu, S., Meng, W., Zhang, X.: Treating pseudo-labels generation as image matting for weakly supervised semantic segmentation. In: Proceedings of the IEEE/CVF International Conference on Computer Vision. pp. 755-765 (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.73, + 0.787, + 0.785 + ], + "angle": 0, + "content": "58. Wang, S., Veldhuis, R., Brune, C., Strisciuglio, N.: What do neural networks learn in image classification? a frequency shortcut perspective. In: Proceedings of the IEEE/CVF International Conference on Computer Vision (ICCV). pp. 1433-1442 (October 2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.785, + 0.787, + 0.84 + ], + "angle": 0, + "content": "59. Wang, Y., Zhang, J., Kan, M., Shan, S., Chen, X.: Self-supervised equivariant attention mechanism for weakly supervised semantic segmentation. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 12275-12284 (2020)" + }, + { + "type": "list", + "bbox": [ + 0.217, + 0.147, + 0.787, + 0.84 + ], + "angle": 0, + "content": null + } + ], + [ + { + "type": "header", + "bbox": [ + 0.349, + 0.115, + 0.732, + 0.129 + ], + "angle": 0, + "content": "Phase Concentration and Shortcut Suppression for WSSS" + }, + { + "type": "page_number", + "bbox": [ + 0.769, + 0.117, + 0.785, + 0.127 + ], + "angle": 0, + "content": "19" + }, + { + "type": "ref_text", + "bbox": [ + 0.214, + 0.147, + 0.788, + 0.203 + ], + "angle": 0, + "content": "60. Wu, T., Huang, J., Gao, G., Wei, X., Wei, X., Luo, X., Liu, C.H.: Embedded discriminative attention mechanism for weakly supervised semantic segmentation. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 16765-16774 (2021)" + }, + { + "type": "ref_text", + "bbox": [ + 0.214, + 0.204, + 0.787, + 0.245 + ], + "angle": 0, + "content": "61. Xie, J., Hou, X., Ye, K., Shen, L.: Clims: Cross language image matching for weakly supervised semantic segmentation. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 4483-4492 (2022)" + }, + { + "type": "ref_text", + "bbox": [ + 0.216, + 0.246, + 0.787, + 0.3 + ], + "angle": 0, + "content": "62. Xie, J., Xiang, J., Chen, J., Hou, X., Zhao, X., Shen, L.: C2am: Contrastive learning of class-agnostic activation map for weakly supervised object localization and semantic segmentation. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 989–998 (2022)" + }, + { + "type": "ref_text", + "bbox": [ + 0.216, + 0.301, + 0.787, + 0.356 + ], + "angle": 0, + "content": "63. Xu, L., Ouyang, W., Bennamoun, M., Boussaid, F., Xu, D.: Multi-class token transformer for weakly supervised semantic segmentation. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 4310-4319 (2022)" + }, + { + "type": "ref_text", + "bbox": [ + 0.216, + 0.357, + 0.787, + 0.412 + ], + "angle": 0, + "content": "64. Xu, L., Ouyang, W., Bennamoun, M., Boussaid, F., Xu, D.: Learning multi-modal class-specific tokens for weakly supervised dense object localization. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 19596-19605 (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.216, + 0.412, + 0.787, + 0.453 + ], + "angle": 0, + "content": "65. Xu, Q., Zhang, R., Zhang, Y., Wang, Y., Tian, Q.: A fourier-based framework for domain generalization. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 14383-14392 (2021)" + }, + { + "type": "ref_text", + "bbox": [ + 0.216, + 0.453, + 0.787, + 0.495 + ], + "angle": 0, + "content": "66. Yang, Y., Soatto, S.: Fda: Fourier domain adaptation for semantic segmentation. In: Proceedings of the IEEE/CVF conference on computer vision and pattern recognition. pp. 4085-4095 (2020)" + }, + { + "type": "ref_text", + "bbox": [ + 0.216, + 0.495, + 0.787, + 0.563 + ], + "angle": 0, + "content": "67. Yoon, S.H., Kweon, H., Cho, J., Kim, S., Yoon, K.J.: Adversarial erasing framework via triplet with gated pyramid pooling layer for weakly supervised semantic segmentation. In: Computer Vision-ECCV 2022: 17th European Conference, Tel Aviv, Israel, October 23-27, 2022, Proceedings, Part XXIX. pp. 326-344. Springer Nature Switzerland Cham (2022)" + }, + { + "type": "ref_text", + "bbox": [ + 0.216, + 0.565, + 0.787, + 0.605 + ], + "angle": 0, + "content": "68. Yoon, S.H., Kweon, H., Jeong, J., Kim, H., Kim, S., Yoon, K.J.: Exploring pixel-level self-supervision for weakly supervised semantic segmentation. arXiv preprint arXiv:2112.05351 (2021)" + }, + { + "type": "ref_text", + "bbox": [ + 0.216, + 0.606, + 0.787, + 0.647 + ], + "angle": 0, + "content": "69. Yoon, S.H., Kwon, H., Kim, H., Yoon, K.J.: Class tokens infusion for weakly supervised semantic segmentation. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 3595-3605 (2024)" + }, + { + "type": "ref_text", + "bbox": [ + 0.216, + 0.647, + 0.787, + 0.688 + ], + "angle": 0, + "content": "70. Zhang, F., Gu, C., Zhang, C., Dai, Y.: Complementary patch for weakly supervised semantic segmentation. In: Proceedings of the IEEE/CVF International Conference on Computer Vision. pp. 7242-7251 (2021)" + }, + { + "type": "ref_text", + "bbox": [ + 0.216, + 0.689, + 0.787, + 0.731 + ], + "angle": 0, + "content": "71. Zhang, X., Wei, Y., Feng, J., Yang, Y., Huang, T.S.: Adversarial complementary learning for weakly supervised object localization. In: Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition. pp. 1325-1334 (2018)" + }, + { + "type": "ref_text", + "bbox": [ + 0.216, + 0.731, + 0.787, + 0.772 + ], + "angle": 0, + "content": "72. Zhou, B., Khosla, A., Lapedriza, A., Oliva, A., Torralba, A.: Learning deep features for discriminative localization. In: Proceedings of the IEEE conference on computer vision and pattern recognition. pp. 2921-2929 (2016)" + }, + { + "type": "ref_text", + "bbox": [ + 0.216, + 0.772, + 0.787, + 0.813 + ], + "angle": 0, + "content": "73. Zhou, T., Zhang, M., Zhao, F., Li, J.: Regional semantic contrast and aggregation for weakly supervised semantic segmentation. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 4299-4309 (2022)" + }, + { + "type": "list", + "bbox": [ + 0.214, + 0.147, + 0.788, + 0.813 + ], + "angle": 0, + "content": null + } + ] +] \ No newline at end of file diff --git a/2024/Phase Concentration and Shortcut Suppression for Weakly Supervised Semantic Segmentation/8ce12d89-9c27-46eb-bd86-41483594c43e_origin.pdf b/2024/Phase Concentration and Shortcut Suppression for Weakly Supervised Semantic Segmentation/8ce12d89-9c27-46eb-bd86-41483594c43e_origin.pdf new file mode 100644 index 0000000000000000000000000000000000000000..e77f8f2814362c1ae879824a348b009ec1d8a0dc --- /dev/null +++ b/2024/Phase Concentration and Shortcut Suppression for Weakly Supervised Semantic Segmentation/8ce12d89-9c27-46eb-bd86-41483594c43e_origin.pdf @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:9eb5ef12fc1c7e2df9c3240886d8531d552cc9705808d851d8ff372c164c4ba1 +size 1287049 diff --git a/2024/Phase Concentration and Shortcut Suppression for Weakly Supervised Semantic Segmentation/full.md b/2024/Phase Concentration and Shortcut Suppression for Weakly Supervised Semantic Segmentation/full.md new file mode 100644 index 0000000000000000000000000000000000000000..6fceeb16b3ff4690ffe27a6014cb0e35fc6f62bf --- /dev/null +++ b/2024/Phase Concentration and Shortcut Suppression for Weakly Supervised Semantic Segmentation/full.md @@ -0,0 +1,352 @@ +# Phase Concentration and Shortcut Suppression for Weakly Supervised Semantic Segmentation + +Hoyong Kwon, Jaeseok Jeong, Sung-Hoon Yoon, and Kuk-Jin Yoon + +KAIST {kwonhoyong3, jason.jeong,yoon307,kjyoon}@kaist.ac.kr + +Abstract. Weakly Supervised Semantic Segmentation (WSSS) with image-level supervision typically acquires object localization information from Class Activation Maps (CAMs). While Vision Transformers (ViTs) in WSSS have been increasingly explored for their superior performance in understanding global context, CAMs from ViT still show imprecise localization in boundary areas and false positive activation. This paper proposes a novel WSSS framework that targets these issues based on the information from the frequency domain. In our framework, we introduce the Magnitude-mixing-based Phase Concentration (MPC) module, which guides the classifier to prioritize phase information containing high-level semantic details. By perturbing and mixing the magnitude, MPC guides the classifier to accentuate and concentrate on the shape information in the phase, thereby leading to finer distinctions in CAMs boundary regions. Additionally, inspired by empirical observations that the classification "shortcut" in the frequency domain can induce false positives in CAMs, we introduce a Frequency Shortcut Suppression (FSS) module. This module aims to discourage the formation of such shortcuts, thereby mitigating false positives. The effectiveness of our approach is demonstrated by achieving new state-of-the-art performance on both PASCAL VOC 2012 and MS COCO 2014 datasets. The code is available at https://github.com/kwonhoyong3/PCSS-WSSS. + +Keywords: Weakly Supervised Semantic Segmentation $\cdot$ Fourier Transform $\cdot$ Shortcut Learning + +# 1 Introduction + +Weakly Supervised Semantic Segmentation (WSSS) aims to enhance its utility by reducing the dependency on human-annotated pixel-level labels in semantic segmentation. With this goal, WSSS mainly explores methods to generate accurate pseudo-labels using only easily obtainable weak labels such as image-level labels [1-3,25,26,29,51,60,68,69], scribbles [37,56], and bounding boxes [8,30,43]. Among these methods, leveraging image-level labels has been extensively investigated, primarily due to the availability of large existing datasets. + +Since image-level labels do not provide spatial information of objects, image-level based WSSS typically involves multiple stages: 1) extract Class Activation Maps (CAMs) to localize the objects, 2) refine CAMs to generate pseudo + +![](images/65cd16063131316a619f9010443bac3b5880b73cd85d2a8e0b275ee06695f42d.jpg) +(a) Image + +![](images/62e0460bd3241426ade0e6b23fa700c754c555d98809bc34264fd2b646d0a901.jpg) +(b) Freq. MaskBird + +![](images/5bcad55ed0d73d1c44790b8765f1d2b9e15fcbe8c5bbab80f3f5d41426b03f87.jpg) +(c) Filtered Image +Fig. 1: Experimental results verifying the relationship between CAMs and Frequency Shortcut. The Filtered Image (c) was obtained by removing the Principal Frequency components (b) related to the class of the Image (a). (d) and (e) are CAMs of images (a) and (c), respectively, derived from the same ViT trained on clean images. + +![](images/c54e4234d4e86271e2e923e46a5aac5424f519644f519bf3a0cbe5d63e1e8de6.jpg) +(d) CAM + +![](images/36b2cb78ba651ca7d5e8e6119fc18cf31b0870b71ec92e652c641f6da6d8e1b9.jpg) +(e) CAM Filtered + +ground truth (Pseudo-GT), and 3) train a semantic segmentation model using the Pseudo-GT. As the second and third stages are dependent on the CAMs from the first stage, our work targets to improve the quality of CAMs. In the first stage, unlike Convolutional Neural Networks (CNNs) that focus on discriminative regions, Vision Transformer (ViT) [9] stands out in WSSS for its ability to understand global context through the Multi-Head Self-Attention (MHSA) mechanism [55]. + +Despite these advantages over CNN-based CAMs, CAMs from ViT still face several issues; 1) CAMs exhibit imprecise object boundaries due to lack of spatial supervision, and 2) CAMs often localize object-unrelated (false positive) regions. In this paper, we propose two methodologies to address the issues with CAMs. + +When humans and ViTs visually assess an object, semantics is predominantly influenced by shape over texture [4, 24]. While shape and texture information are intertwined in images, applying Discrete Fourier Transform (DFT) [54] can help extract the two information into magnitude and phase spectral components. Inspired by the fact that the phase spectrum encapsulates shape and high-level semantic information [16, 42], we propose Magnitude-mixing-based Phase Concentration (MPC) to encourage models to focus more on the phase and improve object boundaries in generated CAMs. In MPC, we select a random pair of images consisting of an anchor image and another arbitrary image. Within the intermediate layers of a ViT, we merge the magnitude information from the arbitrary image into the magnitude from the anchor image. This results in the creation of two anchor image features with identical phases but different magnitudes. By ensuring that both features are guided by the same weak supervision and that the generated CAMs are consistent regardless of changes in magnitude, we induce the model to focus on the phase. Consequently, MPC enables the model to concentrate on the crucial information within the phase, leading to CAMs with more precise boundaries. + +While MPC facilitates precise activation near the boundaries, the issue of activation in object-unrelated regions persists. Interestingly, certain frequency components have been shown to significantly impact classification during the training process of ViT, acting as Frequency Shortcut [58]. A frequency shortcut is a simplicity bias in classification networks, where the network relies on a + +specific frequency rather than semantic content for classification. Additionally, we empirically observed that models tend to exhibit false positives in the CAMs of classes attempting to find shortcuts in the frequency domain. To explore the effects of principal frequency and shortcuts in CAMs, we compare the CAMs from a clean image and a corresponding image with the principal frequency removed (i.e., filtered image), as shown in Fig. 1. While the CAMs from clean image (Fig. 1(d)) suffers from false activations, false activations are relieved in the CAMs from filtered image (Fig. 1(e)). Here, we can verify that certain principal frequencies influence the entire image and cause widespread false activation regardless of the semantics; these can be regarded as shortcut frequencies. + +To resolve the false positives induced by frequency shortcuts, we propose the Frequency Shortcut Suppression (FSS). Initially, we measured the impact of each frequency on the classification of each class to aggregate a Frequency Shortcut Potential Map (FSPM). Based on FSPM, we removed the frequency shortcut components from the image to generate the filtered image. Through classification on the filtered image and by comparing CAMS generated from the original image and the filtered image, FSS reduces the reliance of the generated CAMs to false positive inducing shortcut frequency components. This removal of shortcut frequency components happens iteratively, with the FSPM constantly adjusted. + +The main contributions of this paper are summarized as follows: + +- Through Magnitude-mixing-based Phase Concentration (MPC), we guide ViT to focus on the meaningful phase information, resulting in precise activation around boundaries. +- We proposed a Frequency Shortcut Suppression (FSS) module that effectively reduces false positive activations in CAMs by suppressing shortcut learning in the Frequency Domain. + +To demonstrate the effectiveness of our methodology, we conducted comparisons with state-of-the-art WSSS methods on the PASCAL VOC 2012 [11] and MS COCO 2014 datasets [38], achieving new SoTA performance. + +# 2 Related Works + +# 2.1 Weakly Supervised Semantic Segmentation + +Most existing WSSS approaches utilize Class Activation Maps (CAMs) [72] obtained from Convolutional Neural Networks (CNNs) to localize objects when only image-level labels are provided. However, raw CAMs (i.e. seeds) fail to localize the whole object with imprecise boundaries. In pursuit of generating more accurate pixel-level pseudo-labels, research in WSSS aims to propose a method to improve the quality of raw CAMs or further refine/post-processing the CAMs. CAMs Quality Improvement To guide CAMs to localize the less discriminative regions, various methods are proposed including online CAMs aggregation [21], sub-categories exploration [3], information bottleneck reduction [27] + +and mining cross-image relations [12, 34, 52]. Also, [22, 59, 70] introduce various auxiliary tasks to regularize CAMs training. Along with these methods, Adversarial Erasing (AE) [25, 33, 53, 67, 71] methods, which are based on the erase-and-find mechanism, are also actively researched to expand CAMs. Apart from techniques aimed at expanding CAMs, contrastive learning-based WSSS [6, 62, 73] and adversarial learning between classifier-reconstructor [26] are proposed to produce CAMs with precise object boundaries. Though CNN-based WSSS methods show promising performance, with the success of Vision Transformer (ViT) in various vision tasks, several methods [5, 13, 39, 44, 48, 49, 63, 64, 69] are proposed to migrate the ViT for WSSS. MCTformer [63] introduces multi-class token to extract class-specific attention map from ViT by improving the TS-CAM [13], which is designed for weakly-supervised object localization. In the same line, ViT-PCM [47] proposed a method to extract CAMs based on patch-class mapping. Along with these methods, large language model-based WSSS [39, 61, 64] and token-level contrasting [49] are also introduced. Unlike the prior methods, our work is the first to utilize frequency domain information to address the issues with CAMs in ViT-based WSSS. + +CAMs Post-processing To obtain high-quality pseudo labels for semantic segmentation, various approaches are introduced in WSSS based on affinity learning [1, 2], removing intra-object edges in contours [32], iterative refinements [35, 36]. Besides these methods, Image-matting based refinements [57] and utilizing unsupervised semantic segmentation model [23] greatly boost the quality of pseudo labels. As the post-processing methods can be synergistically applied with methods that generate precise initial seeds (CAMs), this paper focuses on improving the quality of the CAMs. + +# 2.2 Fourier Transform in Computer Vision + +Transforming images from the spatial domain to the frequency domain via the Fourier Transform enables the separation of information into phase and amplitude components. Early studies [16,42] have shown that while amplitude carries low-level statistics and texture information, phase encapsulates high-level semantic and shape information. Leveraging this advantage, Fourier Transform has been applied to a variety of vision tasks such as Exposure Correction [20], Semantic Segmentation [18], and Learning with noisy label [17]. Along with the successful application of frequency components, the fields of Domain Adaptation [19, 66] and Generalization [15, 31, 65] have emphasized the importance of phase for its high-level semantic and structural information. FDA [66] performed Domain Adaptation by transforming low-level amplitude in the Fourier Domain, maintaining high-level semantics and structure, building on the understanding that low-level amplitude contains image style information. [65] highlighted the significance of the phase component, which carries semantic information, and proposed Fourier-based data augmentation through Image-level amplitude to perform Domain Generalization with an emphasis on phase information. + +These studies underscore the significant role of the frequency domain in neural networks. Intriguingly, classification networks engage in shortcut learning + +in the frequency domain [58], identifying specific frequencies for classification. Since shortcut learning [14] can disrupt semantic capture, research efforts have been directed towards reducing such effects [40, 46]. SGT [41] analyzed how vanilla ViT activates areas unrelated to the object due to shortcut learning in the background regions during the classification process and proposed rectifying this through saliency maps. + +# 3 Methods + +# 3.1 Preliminary + +Vision Transformer for CAMs To propagate the image $\mathbf{I}$ into Vision Transformer (ViT), it is split into $N\times N$ patches and forms the patch tokens $\mathbf{T}_{patch}^{0}\in \mathbb{R}^{D\times N^2}$ , where $D$ is the embedding dimension. Along with $C$ class tokens $\mathbf{T}_{cls}^{0}\in \mathbb{R}^{D\times C}$ , the patch tokens $\mathbf{T}_{patch}^{0}$ pass through the Vision Transformer (ViT). Consequently, the final class tokens $\mathbf{T}_{cls}^{L}\in \mathbb{R}^{D\times C}$ and patch tokens $\mathbf{T}_{patch}^{L}\in \mathbb{R}^{D\times N^2}$ are derived as outputs, where $L$ represents the number of ViT layers. This procedure can be formally defined as follows: + +$$ +\mathbf {T} _ {c l s} ^ {i}, \mathbf {T} _ {p a t c h} ^ {i} = \mathbf {V} \mathrm {i} \mathbf {T} _ {i} \left(\mathbf {T} _ {c l s} ^ {i - 1}, \mathbf {T} _ {p a t c h} ^ {i - 1}\right), 1 \leq i \leq L \tag {1} +$$ + +where $\mathbf{V}\mathbf{i}\mathbf{T}_i$ is $i^{th}$ layer of ViT. The class tokens $\mathbf{T}_{cls}^L$ are processed through average pooling to derive the classification logits $y_{c} \in \mathbb{R}^{C}$ . Meanwhile, the patch tokens $\mathbf{T}_{patch}^L$ , arranged according to their spatial positions, pass through a convolution layer to produce the Class Activation Maps (CAMs) $\mathbf{M} \in \mathbb{R}^{C \times N \times N}$ . By applying Global Average Pooling (GAP) to $\mathbf{M}$ , we can obtain classification logits $y_{p} \in \mathbb{R}^{C}$ in a patch-level. These two classification logits are supervised using a multi-label soft margin loss $\mathcal{L}_{cls}$ with the given classification labels $\mathbf{y}$ . Additionally, following [63], we extract the token-to-token attention $\mathbf{A}_{t2t} \in \mathbb{R}^{(C + N^2) \times (C + N^2)}$ based on multi-head self-attention. From this token-to-token attention $\mathbf{A}_{t2t}$ , the class-to-patch attention $\mathbf{A}_{c2p} \in \mathbb{R}^{C \times N^2}$ and patch-to-patch attention $\mathbf{A}_{p2p} \in \mathbb{R}^{N^2 \times N^2}$ can be extracted. These two types of attention maps are used to refine the CAMs, forming a refined CAMs $\mathbf{M}_{ref} \in \mathbb{R}^{C \times N \times N}$ . The aforementioned process can be noted as follows: + +$$ +\mathbf {M} _ {r e f} = \left(\mathbf {M} \odot \mathbf {A} _ {c 2 p}\right) \times \mathbf {A} _ {p 2 p} \tag {2} +$$ + +where $\odot$ and $\times$ are Hadamard product and matrix multiplication, respectively. Discrete Fourier Transform Let $\mathcal{F}$ denote the Discrete Fourier Transform (DFT) [54] applied across the spatial dimensions. For a feature map $\mathbf{X} \in \mathbb{R}^{D \times H \times W}$ the result of applying $\mathcal{F}$ to a single channel can be represented as follows: + +$$ +\mathcal {F} (\mathbf {X}) (m, n) = \sum_ {h, w} \mathbf {X} (h, w) e ^ {- j 2 \pi \left(\frac {h}{H} m + \frac {w}{W} n\right)}, j ^ {2} = - 1 \tag {3} +$$ + +Since the absolute value and the angle component of $\mathcal{F}(\mathbf{X})$ respectively represent magnitude and phase components, we denote the magnitude spectrum as $|\mathcal{F}(\mathbf{X})|$ and the phase spectrum as $\angle \mathcal{F}(\mathbf{X})$ throughout the paper. + +![](images/e8121360ec3292f87495d4a534d49c1a53760efad3dfb1475008820b494d8053.jpg) +Fig. 2: Visualization of the proposed framework. The Magnitude-mixing-based Phase Concentration (MPC) is executed at an intermediate layer. Here, the magnitude of features from the anchor image $\mathbf{I}$ and another arbitrary image $\hat{\mathbf{I}}$ are mixed to make a feature with identical phase, but different magnitude. The model is guided to perform correct classification and derive consistent CAMs $\mathbf{M}_{ref}$ from a pair of features. In the Frequency Shortcut Suppression (FSS), based on the Frequency Shortcut Potential Map (FSPM), frequency components that have a significant impact on the classification of classes in $\mathbf{I}$ are reduced before being inputted to the ViT. The model is encouraged to classify correctly and minimize the difference between $\mathbf{M}$ and $\hat{\mathbf{M}}$ . + +# 3.2 Magnitude-mixing-based Phase Concentration + +With the aid of multi-head self-attention, CAMs from ViT are more proficient in capturing the less-discriminative regions than CNN-based methods. Nonetheless, they still show imprecise localization along object boundaries due to a lack of spatial constraint. To improve localization of CAMs, we design the Magnitude-mixing-based Phase Concentration (MPC) module to extract high-level semantic and boundary information shown to be present in the phase component of the frequency domain [16, 42]. + +We consider two ideas in MPC to guide the model to prioritize the phase component: 1) The model should correctly classify the image even when the magnitude component is altered. Similarly, 2) the CAMs generated after the magnitude change should be similar to those obtained before the change. The key to these two ideas is to change the magnitude while keeping the phase component intact. Since the magnitude needs to be 'feasible' yet different, we designed a magnitude-mixing method. Here, we form a random image pair consisting of an anchor image $\mathbf{I}$ and an arbitrary image $\dot{\mathbf{I}}$ for perturbation. Let the input class and patch tokens of $\mathbf{I}$ be denoted as $[\mathbf{T}_{cls}^{0},\mathbf{T}_{patch}^{0}]$ , and those of $\dot{\mathbf{I}}$ as $[\dot{\mathbf{T}}_{cls}^{0},\dot{\mathbf{T}}_{patch}^{0}]$ . We then process each set of tokens through the ViT layers, as outlined in Eq. 1, for $1\leq i\leq \ell$ , to obtain $[\mathbf{T}_{cls}^{\ell},\mathbf{T}_{patch}^{\ell}]$ and $[\dot{\mathbf{T}}_{cls}^{\ell},\dot{\mathbf{T}}_{patch}^{\ell}]$ respectively. After reshaping + +![](images/923a377c43b81f8725a8592a94044bef4ef0331b6b6b973ed1302bd41ba38967.jpg) +Fig. 3: Visualization of Frequency Influence Measurement. For an image $\mathbf{I}$ containing class $c$ , a set of images that each masked at different frequency components using $M_{f}$ in the frequency domain of $\mathbf{I}$ are inputted into ViT. The classification loss for class $c$ is calculated as the influence of frequency $f$ on class $c$ . Accumulate the process across the image set $\mathbb{I}_c$ to produce $\mathbf{P}(c)$ , and repeat for each $c$ . Finally, $\mathbf{P}$ is normalized to obtain the Frequency Shortcut Potential Map (FSPM) for each class. + +and aligning $\mathbf{T}_{patch}^{\ell}$ and $\bar{\mathbf{T}}_{patch}^{\ell}$ according to their spatial positions, the spectral magnitude is obtained and perturbed using the following formula: + +$$ +\bar {\mathbf {T}} _ {p a t c h} ^ {\ell} = \mathcal {F} ^ {- 1} \left( \right.\left\langle \right.\left\langle \right. r \mid \mathcal {F} \left(\mathbf {T} _ {p a t c h} ^ {\ell}\right)\left. \right| + (1 - r) \mid \mathcal {F} \left(\dot {\mathbf {T}} _ {p a t c h} ^ {\ell}\right)\left. \right|, \angle \mathcal {F} \left(\mathbf {T} _ {p a t c h} ^ {\ell}\right)\left. \right\rangle \rangle\left. \right) \tag {4} +$$ + +where $r$ is the perturbation ratio between the anchor image and arbitrary image, and $\mathcal{F}^{-1}$ is the inverse DFT. Tokens and maps that have been affected by magnitude-mixing are denoted by a bar on top (i.e. $\bar{\mathbf{T}}_{patch}^{\ell}$ ). $\langle \langle A,B\rangle \rangle$ is coupling of magnitude $A$ and phase $B$ . As shown in Eq. 4, the phase component of $\bar{\mathbf{T}}_{patch}^{\ell}$ remains unchanged as the magnitude is perturbed. After iDFT, the set of tokens $[\mathbf{T}_{cls}^{\ell},\bar{\mathbf{T}}_{patch}^{\ell}]$ are reshaped and passed through the remaining layers of ViT, following Eq. 1 from $\ell < i\leq L$ , to produce the output tokens $[\bar{\mathbf{T}}_{cls}^{L},\bar{\mathbf{T}}_{patch}^{L}]$ . The class prediction $\bar{y}_c$ is obtained from $\bar{\mathbf{T}}_{cls}^{L}$ , while the class prediction $\bar{y}_p$ and refined CAMs $\bar{\mathbf{M}}_{ref}$ are generated from $\bar{\mathbf{T}}_{patch}^{L}$ . $\bar{y}_c$ and $\bar{y}_p$ are used as class predictions in multi-label soft margin loss to obtain $\mathcal{L}_{mpc - cls}$ . Following the second idea, the difference between the refined CAMs $\bar{\mathbf{M}}_{ref}$ and $\bar{\mathbf{M}}_{ref}$ are minimized using the following loss: + +$$ +\mathcal {L} _ {\text {m p c - c a m s}} = \left| \mathbf {M} _ {\text {r e f}} - \bar {\mathbf {M}} _ {\text {r e f}} \right| _ {1} \tag {5} +$$ + +Ultimately, the MPC is supervised through the following loss: + +$$ +\mathcal {L} _ {m p c} = \mathcal {L} _ {m p c - c l s} + \lambda_ {1} \mathcal {L} _ {m p c - c a m s} \tag {6} +$$ + +where $\lambda_{1}$ is a hyperparameter to balance the losses. Additionally, the gradient to $\mathbf{M}_{ref}$ from $\mathcal{L}_{mpc-cams}$ is detached to guide $\bar{\mathbf{M}}_{ref}$ to follow $\mathbf{M}_{ref}$ . This increases the dependency of the model on the phase information, enabling the model to better understand the boundary information in the phase component. + +# 3.3 Frequency Shortcut Suppression + +Through the MPC module, we have addressed the issues of activation around the boundary in CAMs. However, the problem of false activations in CAMs + +persists. Various factors contribute to false activations in ViT, such as GAP for generating classification logits from CAMs and the over-smoothing problem caused by self-attention of ViT [49]. Interestingly, we experimentally found that false activations can also arise due to Shortcut Learning in the frequency domain [14], which we aim to address. Shortcut Learning during the classification training process in ViT can negatively impact genuine semantic understanding [40], and such shortcuts can lead to wrong activations [41]. Furthermore, it has been revealed that Neural Networks can create shortcuts in the frequency domain [58]. We empirically found a correlation between frequency shortcuts and false positives in CAMs M. Since Frequency Shortcuts are unrelated to objectness, activations caused by frequency shortcuts occur in locations unrelated to objects. We have resolved over-activation by disrupting the creation of frequency shortcuts through the Frequency Shortcut Suppression(FSS) module. + +Frequency Shortcut Potential Map To implement the FSS, the potential of specific frequency components to be used as shortcuts in classification needs to be measured. Inspired by Wang et al. [58], we utilized a Frequency Influence Measurement (FIM) framework, as depicted in Fig. 3, to assess the impact of each component. If the classification loss of the model increases significantly for an image $\mathbf{I}$ of class $c$ when information about a certain frequency component $f$ is removed, it indicates that $f$ has an important role in classifying class $c$ for the model. Therefore, for each class $c$ , the influence of $f$ on the classification of the image set $\mathbb{I}_c$ corresponding to that class is calculated by accumulating classification loss. To remove a specific frequency component $f$ from an image, a Masking map $M_f \in \mathbb{R}^{H \times W}$ is used, where the value for $f$ is set to 0, and others are set to 1. For an image $\mathbf{I}$ , let $\mathbf{I}_f = \mathcal{F}^{-1}(\mathcal{F}(\mathbf{I}) \odot M_f)$ represent the image with the $f$ component masked where $\odot$ is element-wise multiplication. The classification logit derived after passing $\mathbf{I}_f$ through a ViT is denoted by $y_{\mathbf{I}_f}$ . When $\mathbf{P} \in \mathbb{R}^{C \times H \times W}$ is frequency influence accumulated map, the influence of the frequency component $f$ on the classification of class $c$ is calculated as follows: + +$$ +\mathbf {P} (c, f) = \sum_ {\mathbf {I} \in \mathbb {I} _ {c}} \mathcal {L} _ {b c e} \left(y _ {\mathbf {I} _ {f}} (c), 1\right) \tag {7} +$$ + +where $y(c)$ and $\mathcal{L}_{bce}$ are probability of class $c$ and binary cross-entropy loss, respectively. $\mathbf{P}(c)\in \mathbb{R}^{H\times W}$ ultimately holds the classification influence information of each frequency for class $c$ , with higher values indicating a significant impact on classifying class $c$ . $\mathbf{P}$ is normalized for each class to create the Frequency Shortcut Potential Map (FSPM) $\mathbf{FS}\in \mathbb{R}^{C\times H\times W}$ : + +$$ +\mathbf {F S} (c) = \operatorname {N o r m} (\mathbf {P} (c)), c \in \{1, 2, \dots , C \} \tag {8} +$$ + +where Norm is the normalization function that scales each value to fall between 0 and 1. + +Frequency Shortcut Suppression We propose the Frequency Shortcut Suppression (FSS) utilizing the FSPM generated through the FIM process. The FSPM encapsulates information on which frequency components significantly impact classification at the time of FIM execution, indicating frequencies that + +the model is likely to use as shortcuts. The image without shortcut frequency can be generated by reducing the magnitude of frequency components based on the FSPM. Subsequently, the model can be prevented from forming shortcuts by encouraging it to classify correctly and generate consistent CAMs from the original and filtered images. + +Let an image $\mathbf{I}$ contains classes $c\in \mathbb{C}_{\mathbf{I}}$ . The filtered image $\hat{\mathbf{I}}$ is synthesized by decreasing the frequency components from $\mathbf{I}$ based on the FSPM, making it challenging for the model to classify through shortcuts: + +$$ +\hat {\mathbf {I}} = \frac {1}{N} \sum_ {c \in \mathbb {C} _ {\mathbf {I}}} \mathcal {F} ^ {- 1} \left(\mathcal {F} (\mathbf {I}) \odot (1 - \mathbf {F S} (c))\right) \tag {9} +$$ + +where $N$ is the number of classes in $\mathbb{C}_{\mathbf{I}}$ , and $\odot$ denotes the Hadamard product. The predictions $\hat{y}_c$ and $\hat{y}_p$ derived from $\hat{\mathbf{I}}$ via the ViT are trained through a multi-label soft margin loss $\mathcal{L}_{fss-cls}$ . Additionally, the CAMs $\hat{\mathbf{M}}$ generated from $\hat{\mathbf{I}}$ and the CAMs $\mathbf{M}$ derived from the original image $\mathbf{I}$ are supervised to minimize their difference: + +$$ +\mathcal {L} _ {f s s - c a m s} = | \mathbf {M} - \hat {\mathbf {M}} | _ {1} \tag {10} +$$ + +Assuming the model has learned to use a frequency component $f$ as a shortcut for classifying class $c$ , the model cannot use this shortcut $f$ for classification when $f$ is reduced in $\hat{\mathbf{I}}$ . In the absence of the frequency shortcut $f$ , the model performs classification based on the semantic understanding of class $c$ , resulting in activation within areas related to the object. Simultaneously, by reducing the difference between the two CAMs $\mathbf{M}$ and $\hat{\mathbf{M}}$ , false positives caused by shortcuts can be resolved. + +The FSS is ultimately supervised through the following loss: + +$$ +\mathcal {L} _ {f s s} = \mathcal {L} _ {f s s - c l s} + \lambda_ {2} \mathcal {L} _ {f s s - c a m s} \tag {11} +$$ + +where $\lambda_{2}$ is a hyperparameter to balance between the losses. + +The final loss of the proposed overall framework is as follows: + +$$ +\mathcal {L} _ {\text {t o t a l}} = \mathcal {L} _ {\text {c l s}} + \mathcal {L} _ {\text {m p c}} + \mathcal {L} _ {\text {f s s}} \tag {12} +$$ + +# 4 Experiments + +# 4.1 Experimental Settings + +Datasets Our proposed method was evaluated using the PASCAL VOC 2012 and MS-COCO 2014 datasets, widely utilized for Weakly Supervised Semantic Segmentation (WSSS). The PASCAL VOC 2012 includes 20 classes of foreground objects along with a background class, and consists of three subsets (train, val, test) containing 10582, 1449, and 1456 images, respectively. Meanwhile, the MS-COCO 2014 is made up of 80 foreground object classes and a background class and is composed of two subsets (82k train set and 40k val set). + +Table 1: Ablation study of components on the PASCAL VOC 2012 train set. P: Precision, R: Recall. Values from M are measured under similar R for a fair comparison of P. The best result is represented in Bold. + +
LclsLfssLmpc(x) M(y) Mrf
P(%)R(%)mIoU(%)P(%)R(%)mIoU(%)
Baseline✓53.279.645.374.580.963.2
(a)✓✓71.379.859.476.582.165.4
(b)✓✓67.180.655.978.882.267.2
(c)✓✓✓76.280.763.979.584.869.5
+ +Evaluation Metric Following prior research [44,63], the mean Intersection over Union (mIoU) metric is used to assess the quality of generated CAMs and the performance of semantic segmentation models trained on CAMs-based pseudolabels. To measure the quality of CAMs, the mIoU with the most optimal threshold is employed by default. Typically, CAMs are evaluated on the train set, whereas the performance of the semantic segmentation model is assessed on the val set. While ground-truth pixel-level labels are accessible for both datasets on the val set and are analyzed locally, the evaluation for the PASCAL VOC 2012 test set is carried out via the official website. + +Implementation Details Our framework utilizes DeiT-S pre-trained on ImageNet [50] as the classification backbone for a fair comparison with ViT-based WSSS methods [44, 63]. The classification network is trained for 60 epochs with a batch size of 64 and an initial learning rate of 5e-4, using the Adam optimizer. We adopted the same data augmentation technique as in MCTformer [63], but with different image resize scales. We crop images to a size of $224 \times 224$ for equal comparison. In the MPC module, $r = 0.3$ was used as the mixing factor, with MPC applied at ViT layer $\ell = 8$ . To measure the influence of each frequency for FSPM, 100 images were used per class for all classes. Additionally, the FSPM was updated every three epochs through the FIM process. Min-max normalization was utilized in FIM to normalize $\mathbf{P}$ , with the highest measured loss per class serving as the normalization maximum. For balancing each loss relative to $\mathcal{L}_{cls}$ , $\lambda_{1}$ and $\lambda_{2}$ were both set to 2. For the semantic segmentation model training on the PASCAL VOC 2012 dataset, Deeplab-V1 with ResNet38 was employed, while Deeplab-V2 with ResNet101 was used for the MS COCO 2014 dataset. Additional training details are in Supplementary Materials. + +# 4.2 Ablation Studies + +Component Analysis To demonstrate the significance of each module proposed in our framework, we ablated each component as shown in Tab. 1. Observing the performance of refined CAMs in Tab. 1-y, introducing the Frequency Shortcut Suppression (FSS) and Magnitude-mixing-based Phase Concentration (MPC) led to the gain of $2.2\% \mathrm{p}$ and $4.0\% \mathrm{p}$ respectively over baseline. Our method achieved $69.5\%$ mIoU by adopting both components. FSS still resulted in a gain of $2.3\% \mathrm{p}$ in Tab. 1-c compared to Tab. 1-b, implies that FSS and MPC offer distinct advantages. + +Table 2: Ablation study of ViT layer index that MPC performed. Peak performance is observed at layer 8, which is represented in Bold. + +
earlylayermiddlelayerlatelayer
l34781011
mIoU(%)68.068.168.669.567.866.4
+ +![](images/c52d65ecf433beb16c76c62194ab85df54e55dc62e0a42bcf36bf900b83b12b2.jpg) +(a) Image + +![](images/ea24099a9a5b5aaa78c2359f0e2f578d7b7038092b8bf68feb47d21df69e93f5.jpg) +(b) GT + +![](images/5a97f56eeb671efab321c91047875e6a1ae3b2671eb72fe28032b4c42429cd9a.jpg) +(c) Baseline +Fig. 4: Comparison of refined CAMs $\mathbf{M}_{ref}$ between baseline and components in ours. (a) Image, (b) Ground Truth, (c) Baseline, (d) only FSS, (e) only MPC, (f) ours. The yellow and red boxes imply effect of FSS and MPC, respectively. + +![](images/1c957814bf72dc6c15c910f5d2235b07d0778e491f45283960ba3d267c0180d8.jpg) +(d) FSS + +![](images/cc0e49ab30b8e12bc3346679c4dec22a023e2b10017cfdcad830cfec465f31d5.jpg) +(e) MPC + +![](images/97dbb2141e54829bfcee6eeed5fa8a80a5ccdda2625b4a38501729d331a8b7c6.jpg) +(f) FSS+MPC + +To understand the benefits of FSS in CAMs, we examined the results of ViT CAMs that are not refined by the attention-map in Tab. 1-x. To conduct a comparative analysis considering the Precision-Recall tradeoff, thresholds corresponding to similar Recall levels were used. Low precision of baseline $\mathbf{M}$ indicates a significant number of over-activation. Through FSS, we observed a remarkable improvement in precision $(18.1\%)$ and mIoU $(14.1\%)$ compared to the baseline, confirming that false positives were successfully resolved (Tab. 1-b). + +Ablation Studies of MPC The qualitative comparison between the refined CAMs of baseline and MPC in Fig. 4-e demonstrates the efficacy of the MPC by clearly delineating the boundary of the object. Meanwhile, we investigate the impact of the ratio $r$ in MPC, which is used to preserve the magnitude information of the anchor feature. In experiments conducted within a search space below 0.5, there was a minimum gain of $2.3\% p$ when compared to applying FSS alone, with the best performance observed at $r = 0.3$ + +The influence of the layer index $\ell$ where the MPC is performed was analyzed by conducting experiments on early, middle, and late layers in Tab. 2. The performance was highest when $\ell$ was set to 8, with a noticeable decline observed in layers beyond this point. This tendency supports the knowledge that later layers of the ViT play a crucial role in capturing high-level semantic information [9]. Introducing changes to the magnitude while preserving the phase information at these later stages could disrupt the semantic analysis of the model. This disruption may stem from the difference between the coupled information used in previous layers and the re-coupled information with the magnitude alteration. This result shows that MPC applied before when ViT features highly condense to high-level semantics can encourage the model to capture boundary information well in the phase spectrum. + +Table 3: (a) True/False Positive Rate (TPR/FPR) of classification where the high-frequency components following FSPM are retained in images. Metrics are measured under the early training stage of the Baseline. **Bold** indicates the shortcut occurred, and **Underline** represents that model is trying to make a shortcut. (b) Comparative analysis of the Precision of CAMs M between baseline and only with FSS. Precision is evaluated under a similar Recall considering the Precision-Recall trade-off. + +
bkgaerobikebirdboatbottlebuscarcatchaircow
(a)BaselineTPR(%)-58.9636.4388.4753.3668.4650.8379.2668.3555.4817.87
FPR(%)-00.2300.4115.7800.0533.7400.0001.4200.5307.0400.41
(b)BaselineP(%)91.939.224.817.318.958.176.542.880.944.959.7
w/ FSSP(%)92.552.942.259.233.371.187.562.994.143.290.7
tabledoghorsembkpersonplantsheepsofatraintvtotal
(a)BaselineTPR(%)04.7661.2479.4060.2675.4741.8843.6145.4592.5157.23-
FPR(%)00.0001.9700.6400.1605.5600.1500.3800.6300.6800.13-
(b)BaselineP(%)68.666.050.468.658.039.956.479.039.237.053.2
w/ FSSP(%)89.687.579.778.582.764.792.581.254.556.671.3
+ +![](images/789d710b9f35af13364ccd140b16cb4fc65804fd456cb4ec390943ddb952c958.jpg) + +![](images/1349e719e53c3d6dc26311c39edc6a225670db56e187656a8c438f96f007088e.jpg) +PASCAL VOC 2012 + +![](images/213cb39012f09625bbe9c2d815cece94a905d22498d187bd838ea4be36e96bb8.jpg) +Fig.5: Semantic segmentation results on PASCAL VOC 2012 (left) and MS COCO 2014 (right) datasets. From top to bottom: Image, Ours, GT. + +![](images/dfd509dcb24401bf676857c83fdfbb838c1f9ae7fc4652ab7cd67b17be60e035.jpg) + +![](images/4da045b8515eccf717596a93ae711c5d1c7c0d32d5b7a63bd841aa6d01495b9c.jpg) + +![](images/1aaad2b1fe3c7f8d416a631027eef01bd95d1a7b5dc1dc036c96f8978e44a7c2.jpg) +MS COCO 2014 + +![](images/21735237b7d0dbff15ba83edc10a170472f82434af30023ff1050f256d09c081.jpg) + +Frequency Shortcut Analysis and Result of Suppression In Tab. 3, we investigated whether shortcut learning occurs during the model training process in early steps and analyzed the correlation with the improvement in CAMs precision facilitated by the FSS. To conduct this analysis, classification influence according to frequency component for each class was first measured using Eq. 7. Based on this measurement, the classification behavior of the model was analyzed when only some high frequencies of high-influence components were retained. + +A high TPR indicates that class $c$ possesses the principal frequency components utilized in the experiment (i.e., bird will be correctly classified as bird only with the high-influence frequency of bird). This suggests that the model has identified a straightforward method for classifying class $c$ and is attempting to find frequency shortcuts. To consider a principal frequency as a shortcut, it should direct images to the respective class regardless of the actual semantic content in the image (i.e. car with frequency shortcut of bird will falsely classify as bird). This is expressed in Tab. 3 through high TPR and FPR, which repre + +Table 4: Comparison between ours and multi-stage WSSS methods. mIoU(%) is evaluated on PASCAL VOC 2012 train set in both CAMs (seed) and Pseudo-GT (Mask). The backbone and applied post-processing methods (PSA [2], IRN [1]) are listed for a fair comparison. Bold numbers represent the best performance. + +
MethodBackboneSeedPostMask
AdvCAM [28] CVPR'21RN5055.6IRN69.9
OC-CSE [25] ICCV'21WRN3856.0PSA66.9
ECS [53] ICCV'21WRN3856.6PSA67.8
CPN [70] ICCV'21WRN3857.4PSA67.8
AMR [45] AAAI'22RN5056.8PSA69.7
ReCAM [7] CVPR'22RN5054.8IRN70.5
SIPE [6] CVPR'22RN5058.6IRN-
PPC [10] CVPR'22WRN3861.5IRN70.1
AEFT [67] ECCV'22WRN3856.0PSA71.0
ACR [26] CVPR'23WRN3860.3IRN72.3
MCT [63] CVPR'22DeiT-S61.7PSA69.1
USAGE [44] ICCV'23DeiT-S67.7PSA72.8
OursDeiT-S69.5PSA73.2
+ +sent the rates in which principal frequencies of images are correctly and falsely classified to the shortcut respective class $c$ . Examining Tab. 3-a, we can observe that class bird and bottle exhibit high TPR and FPR, indicating the occurrence of shortcuts. Meanwhile, classes such as car, horse, and train show high TPR, suggesting attempts at creating shortcuts. + +Tab. 3-b records the class-specific precision of CAMs $\mathbf{M}$ for the baseline model and after the introduction of the FSS. To ensure a fair comparison while considering the Precision-Recall trade-off, precision was measured at similar levels of Recall. The introduction of FSS to disrupt shortcut creation resulted in an average increase of $24.0\%$ in precision for $\mathbf{M}$ across classes where the baseline model had created or attempted to create shortcuts. This increase confirms that the FSS effectively reduces false positives. Notably, for bird, a class where shortcuts had occurred, the precision of CAMs surged by $42.1\%$ , showcasing the resolve of over-activation in areas unrelated to the object by FSS. Qualitative comparison of refined CAMs $\mathbf{M}_{ref}$ before and after the introduction of FSS in Fig. 4-d further validates the effectiveness of FSS. + +# 4.3 Comparison with State-of-The-Arts + +To train the semantic segmentation model, the CAMs generated by our method were refined using PSA [2] as performed in previous WSSS works [44, 63, 67]. In Tab. 4, we compared the performance of CAMs (Seed) and pseudo-ground-truth (Mask) on the PASCAL VOC 2012 train set, where our method achieved the best performance in both categories. Utilizing our high-quality pseudo labels for training the semantic segmentation model, we compared the results with other WSSS methods in Tab. 5. Our framework demonstrated superiority by achieving the highest performance on the PASCAL VOC 2012 val and test sets. + +Table 5: Comparison of semantic segmentation performance with multi-stage WSSS methods. Performance (mIoU) is evaluated in PASCAL VOC 2012 and MS COCO 2014 datasets. CNN-based methods are represented above the horizontal line, and ViT-based methods are depicted below the line. Bold represents the best performance. + +
MethodBackboneVOC valVOC testCOCO val
AdvCAM [28] CVPR'21RN10168.168.0-
OC-CSE [25] ICCV'21WRN3868.468.236.4
ECS [53] ICCV'21WRN3866.667.6-
CPN [70] ICCV'21WRN3867.868.5-
AMR [45] AAAT'22RN10168.869.1-
ReCAM [7] CVPR'22RN10168.568.442.9
SIPE [6] CVPR'22RN10168.869.7-
SIPE [6] CVPR'22WRN38--43.6
AEFT [67] ECCV'22WRN3870.971.744.8
ACR [26] CVPR'23WRN3871.971.945.3
MCT [63] CVPR'22WRN3871.971.642.0
USAGE [44] ICCV'23WRN3871.972.842.7
USAGE [44] ICCV'23RN101--44.3
OursWRN3873.273.0-
OursRN101--45.7
+ +While ViT-based WSSS works tend to outperform CNN-based works on the PASCAL VOC 2012 dataset, they showed lower performance on the MS COCO 2014 dataset due to activation overlapping between classes. Nonetheless, our method achieved the SoTA performance on the MS COCO 2014 dataset by effectively reducing false positives. Additional CAMs and semantic segmentation results are in the Supplementary Materials. + +# 5 Conclusion + +Weakly Supervised Semantic Segmentation suffers from CAMs that have imprecise activation around boundary areas and object-unrelated activations. In this paper, we aim to address these challenges by applying a novel perspective in the frequency domain, leveraging the Fourier Transform. Firstly, we propose Magnitude-mixing-based Phase Concentration (MPC), which generates features with identical phase but different magnitude at the intermediate layer, encouraging the model to capture the same semantics from these features. This approach leads the model to pay more attention to shape information in the phase, thereby learning to activate boundaries more clearly. Additionally, we identify the occurrence of frequency shortcuts and their association with the over-activation of CAMs. To resolve this, we introduce Frequency Shortcut Suppression (FSS), which discourages the creation of frequency shortcuts, successfully resolving object-unrelated activations. Our experimental results support the utility of the proposed method, demonstrating its effectiveness in addressing the aforementioned issues in WSSS. Furthermore, we have achieved new state-of-the-art performance on the PASCAL VOC and MS COCO datasets. + +Acknowledgement This work was supported by the Technology Innovation Program (1415187329,20024355, Development of autonomous driving connectivity technology based on sensor-infrastructure cooperation) funded by the Ministry of Trade, Industry & Energy(MOTIE, Korea) and the National Research Foundation of Korea(NRF) grant funded by the Korea government(MSIT) (NRF2022R1A2B5B03002636). + +# References + +1. Ahn, J., Cho, S., Kwak, S.: Weakly supervised learning of instance segmentation with inter-pixel relations. In: Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition. pp. 2209-2218 (2019) +2. Ahn, J., Kwak, S.: Learning pixel-level semantic affinity with image-level supervision for weakly supervised semantic segmentation. In: Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition. pp. 4981-4990 (2018) +3. Chang, Y.T., Wang, Q., Hung, W.C., Piramuthu, R., Tsai, Y.H., Yang, M.H.: Weakly-supervised semantic segmentation via sub-category exploration. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 8991-9000 (2020) +4. Chen, G., Peng, P., Ma, L., Li, J., Du, L., Tian, Y.: Amplitude-phase recombination: Rethinking robustness of convolutional neural networks in frequency domain. In: Proceedings of the IEEE/CVF International Conference on Computer Vision (ICCV). pp. 458-467 (October 2021) +5. Chen, L., Lei, C., Li, R., Li, S., Zhang, Z., Zhang, L.: Fpr: False positive rectification for weakly supervised semantic segmentation. In: Proceedings of the IEEE/CVF International Conference on Computer Vision. pp. 1108-1118 (2023) +6. Chen, Q., Yang, L., Lai, J.H., Xie, X.: Self-supervised image-specific prototype exploration for weakly supervised semantic segmentation. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 4288-4298 (2022) +7. Chen, Z., Wang, T., Wu, X., Hua, X.S., Zhang, H., Sun, Q.: Class reactivation maps for weakly-supervised semantic segmentation. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 969-978 (2022) +8. Dai, J., He, K., Sun, J.: Boxsup: Exploiting bounding boxes to supervise convolutional networks for semantic segmentation. In: Proceedings of the IEEE international conference on computer vision. pp. 1635-1643 (2015) +9. Dosovitskiy, A., Beyer, L., Kolesnikov, A., Weissenborn, D., Zhai, X., Unterthiner, T., Dehghani, M., Minderer, M., Heigold, G., Gelly, S., et al.: An image is worth 16x16 words: Transformers for image recognition at scale. arXiv preprint arXiv:2010.11929 (2020) +10. Du, Y., Fu, Z., Liu, Q., Wang, Y.: Weakly supervised semantic segmentation by pixel-to-prototype contrast. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 4320-4329 (2022) +11. Everingham, M., Van Gool, L., Williams, C.K., Winn, J., Zisserman, A.: The pascal visual object classes (voc) challenge. International journal of computer vision 88(2), 303-338 (2010) +12. Fan, J., Zhang, Z., Tan, T., Song, C., Xiao, J.: Cian: Cross-image affinity net for weakly supervised semantic segmentation. In: Proceedings of the AAAI Conference on Artificial Intelligence. vol. 34, pp. 10762-10769 (2020) + +13. Gao, W., Wan, F., Pan, X., Peng, Z., Tian, Q., Han, Z., Zhou, B., Ye, Q.: Ts-cam: Token semantic coupled attention map for weakly supervised object localization. In: Proceedings of the IEEE/CVF International Conference on Computer Vision. pp. 2886-2895 (2021) +14. Geirhos, R., Jacobsen, J.H., Michaelis, C., Zemel, R., Brendel, W., Bethge, M., Wichmann, F.A.: Shortcut learning in deep neural networks. Nature Machine Intelligence 2(11), 665-673 (2020) +15. Guo, J., Wang, N., Qi, L., Shi, Y.: Aloft: A lightweight mlp-like architecture with dynamic low-frequency transform for domain generalization. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 24132-24141 (2023) +16. Hansen, B.C., Hess, R.F.: Structural sparseness and spatial phase alignment in natural scenes. JOSA A 24(7), 1873-1885 (2007) +17. Huang, H., Kang, H., Liu, S., Salvado, O., Rakotoarivelo, T., Wang, D., Liu, T.: Paddles: Phase-amplitude spectrum disentangled early stopping for learning with noisy labels. In: Proceedings of the IEEE/CVF International Conference on Computer Vision. pp. 16719-16730 (2023) +18. Huang, H., Xie, S., Lin, L., Tong, R., Chen, Y.W., Li, Y., Wang, H., Huang, Y., Zheng, Y.: Semicvt: Semi-supervised convolutional vision transformer for semantic segmentation. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 11340-11349 (2023) +19. Huang, J., Guan, D., Xiao, A., Lu, S.: Rda: Robust domain adaptation via fourier adversarial attacking. In: Proceedings of the IEEE/CVF international conference on computer vision. pp. 8988-8999 (2021) +20. Huang, J., Liu, Y., Zhao, F., Yan, K., Zhang, J., Huang, Y., Zhou, M., Xiong, Z.: Deep fourier-based exposure correction network with spatial-frequency interaction. In: European Conference on Computer Vision. pp. 163-180. Springer (2022) +21. Jiang, P.T., Han, L.H., Hou, Q., Cheng, M.M., Wei, Y.: Online attention accumulation for weakly supervised semantic segmentation. IEEE Transactions on Pattern Analysis and Machine Intelligence 44(10), 7062-7077 (2021) +22. Jiang, P.T., Yang, Y., Hou, Q., Wei, Y.: L2g: A simple local-to-global knowledge transfer framework for weakly supervised semantic segmentation. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 16886-16896 (2022) +23. Jo, S., Yu, I.J., Kim, K.: Mars: Model-agnostic biased object removal without additional supervision for weakly-supervised semantic segmentation. arXiv preprint arXiv:2304.09913 (2023) +24. Kim, G., Kim, J., Lee, J.S.: Exploring adversarial robustness of vision transformers in the spectral perspective. In: Proceedings of the IEEE/CVF Winter Conference on Applications of Computer Vision. pp. 3976-3985 (2024) +25. Kweon, H., Yoon, S.H., Kim, H., Park, D., Yoon, K.J.: Unlocking the potential of ordinary classifier: Class-specific adversarial erasing framework for weakly supervised semantic segmentation. In: Proceedings of the IEEE/CVF International Conference on Computer Vision. pp. 6994-7003 (2021) +26. Kweon, H., Yoon, S.H., Yoon, K.J.: Weakly supervised semantic segmentation via adversarial learning of classifier and reconstructor. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 11329-11339 (2023) +27. Lee, J., Choi, J., Mok, J., Yoon, S.: Reducing information bottleneck for weakly supervised semantic segmentation. Advances in Neural Information Processing Systems 34, 27408-27421 (2021) + +28. Lee, J., Kim, E., Yoon, S.: Anti-adversarily manipulated attributions for weakly and semi-supervised semantic segmentation. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 4071-4080 (2021) +29. Lee, J., Oh, S.J., Yun, S., Choe, J., Kim, E., Yoon, S.: Weakly supervised semantic segmentation using out-of-distribution data. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 16897-16906 (2022) +30. Lee, J., Yi, J., Shin, C., Yoon, S.: Bbam: Bounding box attribution map for weakly supervised semantic and instance segmentation. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 2643-2652 (2021) +31. Lee, S., Bae, J., Kim, H.Y.: Decompose, adjust, compose: Effective normalization by playing with frequency for domain generalization. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 11776-11785 (2023) +32. Li, J., Fan, J., Zhang, Z.: Towards noiseless object contours for weakly supervised semantic segmentation. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 16856-16865 (2022) +33. Li, K., Wu, Z., Peng, K.C., Ernst, J., Fu, Y.: Tell me where to look: Guided attention inference network. In: Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition. pp. 9215-9223 (2018) +34. Li, X., Zhou, T., Li, J., Zhou, Y., Zhang, Z.: Group-wise semantic mining for weakly supervised semantic segmentation. arXiv preprint arXiv:2012.05007 (2020) +35. Li, Y., Duan, Y., Kuang, Z., Chen, Y., Zhang, W., Li, X.: Uncertainty estimation via response scaling for pseudo-mask noise mitigation in weakly-supervised semantic segmentation. arXiv preprint arXiv:2112.07431 (2021) +36. Li, Y., Kuang, Z., Liu, L., Chen, Y., Zhang, W.: Pseudo-mask matters in weakly-supervised semantic segmentation. In: Proceedings of the IEEE/CVF International Conference on Computer Vision. pp. 6964-6973 (2021) +37. Lin, D., Dai, J., Jia, J., He, K., Sun, J.: Scribblesup: Scribble-supervised convolutional networks for semantic segmentation. In: Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition. pp. 3159-3167 (2016) +38. Lin, T.Y., Maire, M., Belongie, S., Hays, J., Perona, P., Ramanan, D., Dollar, P., Zitnick, C.L.: Microsoft coco: Common objects in context. In: European conference on computer vision. pp. 740-755. Springer (2014) +39. Lin, Y., Chen, M., Wang, W., Wu, B., Li, K., Lin, B., Liu, H., He, X.: Clip is also an efficient segmenter: A text-driven approach for weakly supervised semantic segmentation. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 15305-15314 (2023) +40. Luo, X., Wei, L., Wen, L., Yang, J., Xie, L., Xu, Z., Tian, Q.: Rectifying the short-cut learning of background for few-shot learning. Advances in Neural Information Processing Systems 34, 13073-13085 (2021) +41. Ma, C., Zhao, L., Chen, Y., Guo, L., Zhang, T., Hu, X., Shen, D., Jiang, X., Liu, T.: Rectify vit shortcut learning by visual saliency. IEEE Transactions on Neural Networks and Learning Systems (2023) +42. Oppenheim, A.V., Lim, J.S.: The importance of phase in signals. Proceedings of the IEEE 69(5), 529-541 (1981) +43. Papandreou, G., Chen, L.C., Murphy, K.P., Yuille, A.L.: Weakly-and semi-supervised learning of a deep convolutional network for semantic image segmentation. In: Proceedings of the IEEE international conference on computer vision. pp. 1742-1750 (2015) + +44. Peng, Z., Wang, G., Xie, L., Jiang, D., Shen, W., Tian, Q.: Usage: A unified seed area generation paradigm for weakly supervised semantic segmentation. arXiv preprint arXiv:2303.07806 (2023) +45. Qin, J., Wu, J., Xiao, X., Li, L., Wang, X.: Activation modulation and recalibration scheme for weakly supervised semantic segmentation. arXiv preprint arXiv:2112.08996 (2021) +46. Robinson, J., Sun, L., Yu, K., Batmanghelich, K., Jegelka, S., Sra, S.: Can contrastive learning avoid shortcut solutions? Advances in neural information processing systems 34, 4974-4986 (2021) +47. Rossetti, S., Zappia, D., Sanzari, M., Schaerf, M., Pirri, F.: Max pooling with vision transformers reconciles class and shape in weakly supervised semantic segmentation. In: European Conference on Computer Vision. pp. 446-463. Springer (2022) +48. Ru, L., Zhan, Y., Yu, B., Du, B.: Learning affinity from attention: End-to-end weakly-supervised semantic segmentation with transformers. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 16846-16855 (2022) +49. Ru, L., Zheng, H., Zhan, Y., Du, B.: Token contrast for weakly-supervised semantic segmentation. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 3093-3102 (2023) +50. Russakovsky, O., Deng, J., Su, H., Krause, J., Satheesh, S., Ma, S., Huang, Z., Karpathy, A., Khosla, A., Bernstein, M., et al.: Imagenet large scale visual recognition challenge. International journal of computer vision 115(3), 211-252 (2015) +51. Su, Y., Sun, R., Lin, G., Wu, Q.: Context decoupling augmentation for weakly supervised semantic segmentation. arXiv preprint arXiv:2103.01795 (2021) +52. Sun, G., Wang, W., Dai, J., Van Gool, L.: Mining cross-image semantics for weakly supervised semantic segmentation. arXiv preprint arXiv:2007.01947 (2020) +53. Sun, K., Shi, H., Zhang, Z., Huang, Y.: ECS-net: Improving weakly supervised semantic segmentation by using connections between class activation maps. In: Proceedings of the IEEE/CVF International Conference on Computer Vision. pp. 7283-7292 (2021) +54. Sundararajan, D.: The discrete Fourier transform: theory, algorithms and applications. World Scientific (2001) +55. Vaswani, A., Shazeer, N., Parmar, N., Uszkoreit, J., Jones, L., Gomez, A.N., Kaiser, L., Polosukhin, I.: Attention is all you need. Advances in neural information processing systems 30 (2017) +56. Vernaza, P., Chandraker, M.: Learning random-walk label propagation for weakly-supervised semantic segmentation. In: Proceedings of the IEEE conference on computer vision and pattern recognition. pp. 7158-7166 (2017) +57. Wang, C., Xu, R., Xu, S., Meng, W., Zhang, X.: Treating pseudo-labels generation as image matting for weakly supervised semantic segmentation. In: Proceedings of the IEEE/CVF International Conference on Computer Vision. pp. 755-765 (2023) +58. Wang, S., Veldhuis, R., Brune, C., Strisciuglio, N.: What do neural networks learn in image classification? a frequency shortcut perspective. In: Proceedings of the IEEE/CVF International Conference on Computer Vision (ICCV). pp. 1433-1442 (October 2023) +59. Wang, Y., Zhang, J., Kan, M., Shan, S., Chen, X.: Self-supervised equivariant attention mechanism for weakly supervised semantic segmentation. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 12275-12284 (2020) + +60. Wu, T., Huang, J., Gao, G., Wei, X., Wei, X., Luo, X., Liu, C.H.: Embedded discriminative attention mechanism for weakly supervised semantic segmentation. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 16765-16774 (2021) +61. Xie, J., Hou, X., Ye, K., Shen, L.: Clims: Cross language image matching for weakly supervised semantic segmentation. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 4483-4492 (2022) +62. Xie, J., Xiang, J., Chen, J., Hou, X., Zhao, X., Shen, L.: C2am: Contrastive learning of class-agnostic activation map for weakly supervised object localization and semantic segmentation. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 989–998 (2022) +63. Xu, L., Ouyang, W., Bennamoun, M., Boussaid, F., Xu, D.: Multi-class token transformer for weakly supervised semantic segmentation. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 4310-4319 (2022) +64. Xu, L., Ouyang, W., Bennamoun, M., Boussaid, F., Xu, D.: Learning multi-modal class-specific tokens for weakly supervised dense object localization. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 19596-19605 (2023) +65. Xu, Q., Zhang, R., Zhang, Y., Wang, Y., Tian, Q.: A fourier-based framework for domain generalization. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 14383-14392 (2021) +66. Yang, Y., Soatto, S.: Fda: Fourier domain adaptation for semantic segmentation. In: Proceedings of the IEEE/CVF conference on computer vision and pattern recognition. pp. 4085-4095 (2020) +67. Yoon, S.H., Kweon, H., Cho, J., Kim, S., Yoon, K.J.: Adversarial erasing framework via triplet with gated pyramid pooling layer for weakly supervised semantic segmentation. In: Computer Vision-ECCV 2022: 17th European Conference, Tel Aviv, Israel, October 23-27, 2022, Proceedings, Part XXIX. pp. 326-344. Springer Nature Switzerland Cham (2022) +68. Yoon, S.H., Kweon, H., Jeong, J., Kim, H., Kim, S., Yoon, K.J.: Exploring pixel-level self-supervision for weakly supervised semantic segmentation. arXiv preprint arXiv:2112.05351 (2021) +69. Yoon, S.H., Kwon, H., Kim, H., Yoon, K.J.: Class tokens infusion for weakly supervised semantic segmentation. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 3595-3605 (2024) +70. Zhang, F., Gu, C., Zhang, C., Dai, Y.: Complementary patch for weakly supervised semantic segmentation. In: Proceedings of the IEEE/CVF International Conference on Computer Vision. pp. 7242-7251 (2021) +71. Zhang, X., Wei, Y., Feng, J., Yang, Y., Huang, T.S.: Adversarial complementary learning for weakly supervised object localization. In: Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition. pp. 1325-1334 (2018) +72. Zhou, B., Khosla, A., Lapedriza, A., Oliva, A., Torralba, A.: Learning deep features for discriminative localization. In: Proceedings of the IEEE conference on computer vision and pattern recognition. pp. 2921-2929 (2016) +73. Zhou, T., Zhang, M., Zhao, F., Li, J.: Regional semantic contrast and aggregation for weakly supervised semantic segmentation. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 4299-4309 (2022) \ No newline at end of file diff --git a/2024/Phase Concentration and Shortcut Suppression for Weakly Supervised Semantic Segmentation/images.zip b/2024/Phase Concentration and Shortcut Suppression for Weakly Supervised Semantic Segmentation/images.zip new file mode 100644 index 0000000000000000000000000000000000000000..01bc07766f9e8b2bbcb35f725db2bdb17ec8e917 --- /dev/null +++ b/2024/Phase Concentration and Shortcut Suppression for Weakly Supervised Semantic Segmentation/images.zip @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:d166aa8c51e6bef991b53f056858e6bdb3d2a11b653c8cf94cc5f73722599dbd +size 479227 diff --git a/2024/Phase Concentration and Shortcut Suppression for Weakly Supervised Semantic Segmentation/layout.json b/2024/Phase Concentration and Shortcut Suppression for Weakly Supervised Semantic Segmentation/layout.json new file mode 100644 index 0000000000000000000000000000000000000000..23441a3d8a66ef6b6069f3fc92909679785454fa --- /dev/null +++ b/2024/Phase Concentration and Shortcut Suppression for Weakly Supervised Semantic Segmentation/layout.json @@ -0,0 +1,12037 @@ +{ + "pdf_info": [ + { + "para_blocks": [ + { + "bbox": [ + 142, + 111, + 473, + 148 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 142, + 111, + 473, + 148 + ], + "spans": [ + { + "bbox": [ + 142, + 111, + 473, + 148 + ], + "type": "text", + "content": "Phase Concentration and Shortcut Suppression for Weakly Supervised Semantic Segmentation" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 143, + 168, + 470, + 181 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 143, + 168, + 470, + 181 + ], + "spans": [ + { + "bbox": [ + 143, + 168, + 470, + 181 + ], + "type": "text", + "content": "Hoyong Kwon, Jaeseok Jeong, Sung-Hoon Yoon, and Kuk-Jin Yoon" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 183, + 190, + 430, + 213 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 183, + 190, + 430, + 213 + ], + "spans": [ + { + "bbox": [ + 183, + 190, + 430, + 213 + ], + "type": "text", + "content": "KAIST {kwonhoyong3, jason.jeong,yoon307,kjyoon}@kaist.ac.kr" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 159, + 239, + 465, + 470 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 159, + 239, + 465, + 470 + ], + "spans": [ + { + "bbox": [ + 159, + 239, + 465, + 470 + ], + "type": "text", + "content": "Abstract. Weakly Supervised Semantic Segmentation (WSSS) with image-level supervision typically acquires object localization information from Class Activation Maps (CAMs). While Vision Transformers (ViTs) in WSSS have been increasingly explored for their superior performance in understanding global context, CAMs from ViT still show imprecise localization in boundary areas and false positive activation. This paper proposes a novel WSSS framework that targets these issues based on the information from the frequency domain. In our framework, we introduce the Magnitude-mixing-based Phase Concentration (MPC) module, which guides the classifier to prioritize phase information containing high-level semantic details. By perturbing and mixing the magnitude, MPC guides the classifier to accentuate and concentrate on the shape information in the phase, thereby leading to finer distinctions in CAMs boundary regions. Additionally, inspired by empirical observations that the classification \"shortcut\" in the frequency domain can induce false positives in CAMs, we introduce a Frequency Shortcut Suppression (FSS) module. This module aims to discourage the formation of such shortcuts, thereby mitigating false positives. The effectiveness of our approach is demonstrated by achieving new state-of-the-art performance on both PASCAL VOC 2012 and MS COCO 2014 datasets. The code is available at https://github.com/kwonhoyong3/PCSS-WSSS." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 159, + 479, + 453, + 502 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 159, + 479, + 453, + 502 + ], + "spans": [ + { + "bbox": [ + 159, + 479, + 453, + 502 + ], + "type": "text", + "content": "Keywords: Weakly Supervised Semantic Segmentation " + }, + { + "bbox": [ + 159, + 479, + 453, + 502 + ], + "type": "inline_equation", + "content": "\\cdot" + }, + { + "bbox": [ + 159, + 479, + 453, + 502 + ], + "type": "text", + "content": " Fourier Transform " + }, + { + "bbox": [ + 159, + 479, + 453, + 502 + ], + "type": "inline_equation", + "content": "\\cdot" + }, + { + "bbox": [ + 159, + 479, + 453, + 502 + ], + "type": "text", + "content": " Shortcut Learning" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 132, + 521, + 230, + 534 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 521, + 230, + 534 + ], + "spans": [ + { + "bbox": [ + 132, + 521, + 230, + 534 + ], + "type": "text", + "content": "1 Introduction" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 130, + 545, + 482, + 629 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 545, + 482, + 629 + ], + "spans": [ + { + "bbox": [ + 130, + 545, + 482, + 629 + ], + "type": "text", + "content": "Weakly Supervised Semantic Segmentation (WSSS) aims to enhance its utility by reducing the dependency on human-annotated pixel-level labels in semantic segmentation. With this goal, WSSS mainly explores methods to generate accurate pseudo-labels using only easily obtainable weak labels such as image-level labels [1-3,25,26,29,51,60,68,69], scribbles [37,56], and bounding boxes [8,30,43]. Among these methods, leveraging image-level labels has been extensively investigated, primarily due to the availability of large existing datasets." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 130, + 630, + 482, + 666 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 630, + 482, + 666 + ], + "spans": [ + { + "bbox": [ + 130, + 630, + 482, + 666 + ], + "type": "text", + "content": "Since image-level labels do not provide spatial information of objects, image-level based WSSS typically involves multiple stages: 1) extract Class Activation Maps (CAMs) to localize the objects, 2) refine CAMs to generate pseudo" + } + ] + } + ], + "index": 7 + } + ], + "discarded_blocks": [], + "page_size": [ + 612, + 792 + ], + "page_idx": 0 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 133, + 114, + 202, + 183 + ], + "blocks": [ + { + "bbox": [ + 133, + 114, + 202, + 183 + ], + "lines": [ + { + "bbox": [ + 133, + 114, + 202, + 183 + ], + "spans": [ + { + "bbox": [ + 133, + 114, + 202, + 183 + ], + "type": "image", + "image_path": "65cd16063131316a619f9010443bac3b5880b73cd85d2a8e0b275ee06695f42d.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 151, + 184, + 184, + 194 + ], + "lines": [ + { + "bbox": [ + 151, + 184, + 184, + 194 + ], + "spans": [ + { + "bbox": [ + 151, + 184, + 184, + 194 + ], + "type": "text", + "content": "(a) Image" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_caption" + } + ], + "index": 2 + }, + { + "type": "image", + "bbox": [ + 203, + 114, + 271, + 183 + ], + "blocks": [ + { + "bbox": [ + 203, + 114, + 271, + 183 + ], + "lines": [ + { + "bbox": [ + 203, + 114, + 271, + 183 + ], + "spans": [ + { + "bbox": [ + 203, + 114, + 271, + 183 + ], + "type": "image", + "image_path": "62e0460bd3241426ade0e6b23fa700c754c555d98809bc34264fd2b646d0a901.jpg" + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 204, + 184, + 271, + 194 + ], + "lines": [ + { + "bbox": [ + 204, + 184, + 271, + 194 + ], + "spans": [ + { + "bbox": [ + 204, + 184, + 271, + 194 + ], + "type": "text", + "content": "(b) Freq. MaskBird" + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "image_caption" + } + ], + "index": 4 + }, + { + "type": "image", + "bbox": [ + 272, + 115, + 340, + 183 + ], + "blocks": [ + { + "bbox": [ + 272, + 115, + 340, + 183 + ], + "lines": [ + { + "bbox": [ + 272, + 115, + 340, + 183 + ], + "spans": [ + { + "bbox": [ + 272, + 115, + 340, + 183 + ], + "type": "image", + "image_path": "5bcad55ed0d73d1c44790b8765f1d2b9e15fcbe8c5bbab80f3f5d41426b03f87.jpg" + } + ] + } + ], + "index": 6, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 276, + 184, + 337, + 193 + ], + "lines": [ + { + "bbox": [ + 276, + 184, + 337, + 193 + ], + "spans": [ + { + "bbox": [ + 276, + 184, + 337, + 193 + ], + "type": "text", + "content": "(c) Filtered Image" + } + ] + } + ], + "index": 7, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 130, + 202, + 482, + 248 + ], + "lines": [ + { + "bbox": [ + 130, + 202, + 482, + 248 + ], + "spans": [ + { + "bbox": [ + 130, + 202, + 482, + 248 + ], + "type": "text", + "content": "Fig. 1: Experimental results verifying the relationship between CAMs and Frequency Shortcut. The Filtered Image (c) was obtained by removing the Principal Frequency components (b) related to the class of the Image (a). (d) and (e) are CAMs of images (a) and (c), respectively, derived from the same ViT trained on clean images." + } + ] + } + ], + "index": 12, + "angle": 0, + "type": "image_caption" + } + ], + "index": 6 + }, + { + "type": "image", + "bbox": [ + 342, + 115, + 411, + 183 + ], + "blocks": [ + { + "bbox": [ + 342, + 115, + 411, + 183 + ], + "lines": [ + { + "bbox": [ + 342, + 115, + 411, + 183 + ], + "spans": [ + { + "bbox": [ + 342, + 115, + 411, + 183 + ], + "type": "image", + "image_path": "c54e4234d4e86271e2e923e46a5aac5424f519644f519bf3a0cbe5d63e1e8de6.jpg" + } + ] + } + ], + "index": 8, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 362, + 184, + 392, + 192 + ], + "lines": [ + { + "bbox": [ + 362, + 184, + 392, + 192 + ], + "spans": [ + { + "bbox": [ + 362, + 184, + 392, + 192 + ], + "type": "text", + "content": "(d) CAM" + } + ] + } + ], + "index": 9, + "angle": 0, + "type": "image_caption" + } + ], + "index": 8 + }, + { + "type": "image", + "bbox": [ + 411, + 115, + 480, + 183 + ], + "blocks": [ + { + "bbox": [ + 411, + 115, + 480, + 183 + ], + "lines": [ + { + "bbox": [ + 411, + 115, + 480, + 183 + ], + "spans": [ + { + "bbox": [ + 411, + 115, + 480, + 183 + ], + "type": "image", + "image_path": "36b2cb78ba651ca7d5e8e6119fc18cf31b0870b71ec92e652c641f6da6d8e1b9.jpg" + } + ] + } + ], + "index": 10, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 416, + 184, + 474, + 192 + ], + "lines": [ + { + "bbox": [ + 416, + 184, + 474, + 192 + ], + "spans": [ + { + "bbox": [ + 416, + 184, + 474, + 192 + ], + "type": "text", + "content": "(e) CAM Filtered" + } + ] + } + ], + "index": 11, + "angle": 0, + "type": "image_caption" + } + ], + "index": 10 + }, + { + "bbox": [ + 130, + 270, + 482, + 354 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 270, + 482, + 354 + ], + "spans": [ + { + "bbox": [ + 130, + 270, + 482, + 354 + ], + "type": "text", + "content": "ground truth (Pseudo-GT), and 3) train a semantic segmentation model using the Pseudo-GT. As the second and third stages are dependent on the CAMs from the first stage, our work targets to improve the quality of CAMs. In the first stage, unlike Convolutional Neural Networks (CNNs) that focus on discriminative regions, Vision Transformer (ViT) [9] stands out in WSSS for its ability to understand global context through the Multi-Head Self-Attention (MHSA) mechanism [55]." + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 130, + 354, + 482, + 403 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 354, + 482, + 403 + ], + "spans": [ + { + "bbox": [ + 130, + 354, + 482, + 403 + ], + "type": "text", + "content": "Despite these advantages over CNN-based CAMs, CAMs from ViT still face several issues; 1) CAMs exhibit imprecise object boundaries due to lack of spatial supervision, and 2) CAMs often localize object-unrelated (false positive) regions. In this paper, we propose two methodologies to address the issues with CAMs." + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 130, + 403, + 482, + 605 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 403, + 482, + 605 + ], + "spans": [ + { + "bbox": [ + 130, + 403, + 482, + 605 + ], + "type": "text", + "content": "When humans and ViTs visually assess an object, semantics is predominantly influenced by shape over texture [4, 24]. While shape and texture information are intertwined in images, applying Discrete Fourier Transform (DFT) [54] can help extract the two information into magnitude and phase spectral components. Inspired by the fact that the phase spectrum encapsulates shape and high-level semantic information [16, 42], we propose Magnitude-mixing-based Phase Concentration (MPC) to encourage models to focus more on the phase and improve object boundaries in generated CAMs. In MPC, we select a random pair of images consisting of an anchor image and another arbitrary image. Within the intermediate layers of a ViT, we merge the magnitude information from the arbitrary image into the magnitude from the anchor image. This results in the creation of two anchor image features with identical phases but different magnitudes. By ensuring that both features are guided by the same weak supervision and that the generated CAMs are consistent regardless of changes in magnitude, we induce the model to focus on the phase. Consequently, MPC enables the model to concentrate on the crucial information within the phase, leading to CAMs with more precise boundaries." + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 130, + 605, + 482, + 665 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 605, + 482, + 665 + ], + "spans": [ + { + "bbox": [ + 130, + 605, + 482, + 665 + ], + "type": "text", + "content": "While MPC facilitates precise activation near the boundaries, the issue of activation in object-unrelated regions persists. Interestingly, certain frequency components have been shown to significantly impact classification during the training process of ViT, acting as Frequency Shortcut [58]. A frequency shortcut is a simplicity bias in classification networks, where the network relies on a" + } + ] + } + ], + "index": 16 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "spans": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "type": "text", + "content": "2" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 166, + 91, + 227, + 101 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 166, + 91, + 227, + 101 + ], + "spans": [ + { + "bbox": [ + 166, + 91, + 227, + 101 + ], + "type": "text", + "content": "H. Kwon et al." + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 1 + }, + { + "para_blocks": [ + { + "bbox": [ + 130, + 116, + 482, + 236 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 116, + 482, + 236 + ], + "spans": [ + { + "bbox": [ + 130, + 116, + 482, + 236 + ], + "type": "text", + "content": "specific frequency rather than semantic content for classification. Additionally, we empirically observed that models tend to exhibit false positives in the CAMs of classes attempting to find shortcuts in the frequency domain. To explore the effects of principal frequency and shortcuts in CAMs, we compare the CAMs from a clean image and a corresponding image with the principal frequency removed (i.e., filtered image), as shown in Fig. 1. While the CAMs from clean image (Fig. 1(d)) suffers from false activations, false activations are relieved in the CAMs from filtered image (Fig. 1(e)). Here, we can verify that certain principal frequencies influence the entire image and cause widespread false activation regardless of the semantics; these can be regarded as shortcut frequencies." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 130, + 236, + 482, + 354 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 236, + 482, + 354 + ], + "spans": [ + { + "bbox": [ + 130, + 236, + 482, + 354 + ], + "type": "text", + "content": "To resolve the false positives induced by frequency shortcuts, we propose the Frequency Shortcut Suppression (FSS). Initially, we measured the impact of each frequency on the classification of each class to aggregate a Frequency Shortcut Potential Map (FSPM). Based on FSPM, we removed the frequency shortcut components from the image to generate the filtered image. Through classification on the filtered image and by comparing CAMS generated from the original image and the filtered image, FSS reduces the reliance of the generated CAMs to false positive inducing shortcut frequency components. This removal of shortcut frequency components happens iteratively, with the FSPM constantly adjusted." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 147, + 356, + 430, + 366 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 147, + 356, + 430, + 366 + ], + "spans": [ + { + "bbox": [ + 147, + 356, + 430, + 366 + ], + "type": "text", + "content": "The main contributions of this paper are summarized as follows:" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 132, + 376, + 480, + 446 + ], + "type": "list", + "angle": 0, + "index": 7, + "blocks": [ + { + "bbox": [ + 132, + 376, + 480, + 409 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 376, + 480, + 409 + ], + "spans": [ + { + "bbox": [ + 132, + 376, + 480, + 409 + ], + "type": "text", + "content": "- Through Magnitude-mixing-based Phase Concentration (MPC), we guide ViT to focus on the meaningful phase information, resulting in precise activation around boundaries." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 132, + 411, + 480, + 446 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 411, + 480, + 446 + ], + "spans": [ + { + "bbox": [ + 132, + 411, + 480, + 446 + ], + "type": "text", + "content": "- We proposed a Frequency Shortcut Suppression (FSS) module that effectively reduces false positive activations in CAMs by suppressing shortcut learning in the Frequency Domain." + } + ] + } + ], + "index": 6 + } + ], + "sub_type": "text" + }, + { + "bbox": [ + 130, + 456, + 482, + 492 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 456, + 482, + 492 + ], + "spans": [ + { + "bbox": [ + 130, + 456, + 482, + 492 + ], + "type": "text", + "content": "To demonstrate the effectiveness of our methodology, we conducted comparisons with state-of-the-art WSSS methods on the PASCAL VOC 2012 [11] and MS COCO 2014 datasets [38], achieving new SoTA performance." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 132, + 511, + 242, + 522 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 511, + 242, + 522 + ], + "spans": [ + { + "bbox": [ + 132, + 511, + 242, + 522 + ], + "type": "text", + "content": "2 Related Works" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 132, + 536, + 377, + 549 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 536, + 377, + 549 + ], + "spans": [ + { + "bbox": [ + 132, + 536, + 377, + 549 + ], + "type": "text", + "content": "2.1 Weakly Supervised Semantic Segmentation" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 130, + 557, + 482, + 667 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 557, + 482, + 667 + ], + "spans": [ + { + "bbox": [ + 130, + 557, + 482, + 667 + ], + "type": "text", + "content": "Most existing WSSS approaches utilize Class Activation Maps (CAMs) [72] obtained from Convolutional Neural Networks (CNNs) to localize objects when only image-level labels are provided. However, raw CAMs (i.e. seeds) fail to localize the whole object with imprecise boundaries. In pursuit of generating more accurate pixel-level pseudo-labels, research in WSSS aims to propose a method to improve the quality of raw CAMs or further refine/post-processing the CAMs. CAMs Quality Improvement To guide CAMs to localize the less discriminative regions, various methods are proposed including online CAMs aggregation [21], sub-categories exploration [3], information bottleneck reduction [27]" + } + ] + } + ], + "index": 11 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 212, + 91, + 447, + 102 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 212, + 91, + 447, + 102 + ], + "spans": [ + { + "bbox": [ + 212, + 91, + 447, + 102 + ], + "type": "text", + "content": "Phase Concentration and Shortcut Suppression for WSSS" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 474, + 92, + 480, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 474, + 92, + 480, + 100 + ], + "spans": [ + { + "bbox": [ + 474, + 92, + 480, + 100 + ], + "type": "text", + "content": "3" + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 2 + }, + { + "para_blocks": [ + { + "bbox": [ + 130, + 115, + 482, + 319 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 115, + 482, + 319 + ], + "spans": [ + { + "bbox": [ + 130, + 115, + 482, + 319 + ], + "type": "text", + "content": "and mining cross-image relations [12, 34, 52]. Also, [22, 59, 70] introduce various auxiliary tasks to regularize CAMs training. Along with these methods, Adversarial Erasing (AE) [25, 33, 53, 67, 71] methods, which are based on the erase-and-find mechanism, are also actively researched to expand CAMs. Apart from techniques aimed at expanding CAMs, contrastive learning-based WSSS [6, 62, 73] and adversarial learning between classifier-reconstructor [26] are proposed to produce CAMs with precise object boundaries. Though CNN-based WSSS methods show promising performance, with the success of Vision Transformer (ViT) in various vision tasks, several methods [5, 13, 39, 44, 48, 49, 63, 64, 69] are proposed to migrate the ViT for WSSS. MCTformer [63] introduces multi-class token to extract class-specific attention map from ViT by improving the TS-CAM [13], which is designed for weakly-supervised object localization. In the same line, ViT-PCM [47] proposed a method to extract CAMs based on patch-class mapping. Along with these methods, large language model-based WSSS [39, 61, 64] and token-level contrasting [49] are also introduced. Unlike the prior methods, our work is the first to utilize frequency domain information to address the issues with CAMs in ViT-based WSSS." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 130, + 319, + 482, + 416 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 319, + 482, + 416 + ], + "spans": [ + { + "bbox": [ + 130, + 319, + 482, + 416 + ], + "type": "text", + "content": "CAMs Post-processing To obtain high-quality pseudo labels for semantic segmentation, various approaches are introduced in WSSS based on affinity learning [1, 2], removing intra-object edges in contours [32], iterative refinements [35, 36]. Besides these methods, Image-matting based refinements [57] and utilizing unsupervised semantic segmentation model [23] greatly boost the quality of pseudo labels. As the post-processing methods can be synergistically applied with methods that generate precise initial seeds (CAMs), this paper focuses on improving the quality of the CAMs." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 131, + 431, + 357, + 443 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 131, + 431, + 357, + 443 + ], + "spans": [ + { + "bbox": [ + 131, + 431, + 357, + 443 + ], + "type": "text", + "content": "2.2 Fourier Transform in Computer Vision" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 130, + 450, + 482, + 641 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 450, + 482, + 641 + ], + "spans": [ + { + "bbox": [ + 130, + 450, + 482, + 641 + ], + "type": "text", + "content": "Transforming images from the spatial domain to the frequency domain via the Fourier Transform enables the separation of information into phase and amplitude components. Early studies [16,42] have shown that while amplitude carries low-level statistics and texture information, phase encapsulates high-level semantic and shape information. Leveraging this advantage, Fourier Transform has been applied to a variety of vision tasks such as Exposure Correction [20], Semantic Segmentation [18], and Learning with noisy label [17]. Along with the successful application of frequency components, the fields of Domain Adaptation [19, 66] and Generalization [15, 31, 65] have emphasized the importance of phase for its high-level semantic and structural information. FDA [66] performed Domain Adaptation by transforming low-level amplitude in the Fourier Domain, maintaining high-level semantics and structure, building on the understanding that low-level amplitude contains image style information. [65] highlighted the significance of the phase component, which carries semantic information, and proposed Fourier-based data augmentation through Image-level amplitude to perform Domain Generalization with an emphasis on phase information." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 130, + 641, + 481, + 666 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 641, + 481, + 666 + ], + "spans": [ + { + "bbox": [ + 130, + 641, + 481, + 666 + ], + "type": "text", + "content": "These studies underscore the significant role of the frequency domain in neural networks. Intriguingly, classification networks engage in shortcut learning" + } + ] + } + ], + "index": 6 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "spans": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "type": "text", + "content": "4" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 166, + 91, + 227, + 101 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 166, + 91, + 227, + 101 + ], + "spans": [ + { + "bbox": [ + 166, + 91, + 227, + 101 + ], + "type": "text", + "content": "H. Kwon et al." + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 3 + }, + { + "para_blocks": [ + { + "bbox": [ + 130, + 116, + 479, + 189 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 116, + 479, + 189 + ], + "spans": [ + { + "bbox": [ + 130, + 116, + 479, + 189 + ], + "type": "text", + "content": "in the frequency domain [58], identifying specific frequencies for classification. Since shortcut learning [14] can disrupt semantic capture, research efforts have been directed towards reducing such effects [40, 46]. SGT [41] analyzed how vanilla ViT activates areas unrelated to the object due to shortcut learning in the background regions during the classification process and proposed rectifying this through saliency maps." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 132, + 203, + 206, + 216 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 203, + 206, + 216 + ], + "spans": [ + { + "bbox": [ + 132, + 203, + 206, + 216 + ], + "type": "text", + "content": "3 Methods" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 132, + 226, + 220, + 238 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 226, + 220, + 238 + ], + "spans": [ + { + "bbox": [ + 132, + 226, + 220, + 238 + ], + "type": "text", + "content": "3.1 Preliminary" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 130, + 243, + 480, + 334 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 243, + 480, + 334 + ], + "spans": [ + { + "bbox": [ + 130, + 243, + 480, + 334 + ], + "type": "text", + "content": "Vision Transformer for CAMs To propagate the image " + }, + { + "bbox": [ + 130, + 243, + 480, + 334 + ], + "type": "inline_equation", + "content": "\\mathbf{I}" + }, + { + "bbox": [ + 130, + 243, + 480, + 334 + ], + "type": "text", + "content": " into Vision Transformer (ViT), it is split into " + }, + { + "bbox": [ + 130, + 243, + 480, + 334 + ], + "type": "inline_equation", + "content": "N\\times N" + }, + { + "bbox": [ + 130, + 243, + 480, + 334 + ], + "type": "text", + "content": " patches and forms the patch tokens " + }, + { + "bbox": [ + 130, + 243, + 480, + 334 + ], + "type": "inline_equation", + "content": "\\mathbf{T}_{patch}^{0}\\in \\mathbb{R}^{D\\times N^2}" + }, + { + "bbox": [ + 130, + 243, + 480, + 334 + ], + "type": "text", + "content": ", where " + }, + { + "bbox": [ + 130, + 243, + 480, + 334 + ], + "type": "inline_equation", + "content": "D" + }, + { + "bbox": [ + 130, + 243, + 480, + 334 + ], + "type": "text", + "content": " is the embedding dimension. Along with " + }, + { + "bbox": [ + 130, + 243, + 480, + 334 + ], + "type": "inline_equation", + "content": "C" + }, + { + "bbox": [ + 130, + 243, + 480, + 334 + ], + "type": "text", + "content": " class tokens " + }, + { + "bbox": [ + 130, + 243, + 480, + 334 + ], + "type": "inline_equation", + "content": "\\mathbf{T}_{cls}^{0}\\in \\mathbb{R}^{D\\times C}" + }, + { + "bbox": [ + 130, + 243, + 480, + 334 + ], + "type": "text", + "content": ", the patch tokens " + }, + { + "bbox": [ + 130, + 243, + 480, + 334 + ], + "type": "inline_equation", + "content": "\\mathbf{T}_{patch}^{0}" + }, + { + "bbox": [ + 130, + 243, + 480, + 334 + ], + "type": "text", + "content": " pass through the Vision Transformer (ViT). Consequently, the final class tokens " + }, + { + "bbox": [ + 130, + 243, + 480, + 334 + ], + "type": "inline_equation", + "content": "\\mathbf{T}_{cls}^{L}\\in \\mathbb{R}^{D\\times C}" + }, + { + "bbox": [ + 130, + 243, + 480, + 334 + ], + "type": "text", + "content": " and patch tokens " + }, + { + "bbox": [ + 130, + 243, + 480, + 334 + ], + "type": "inline_equation", + "content": "\\mathbf{T}_{patch}^{L}\\in \\mathbb{R}^{D\\times N^2}" + }, + { + "bbox": [ + 130, + 243, + 480, + 334 + ], + "type": "text", + "content": " are derived as outputs, where " + }, + { + "bbox": [ + 130, + 243, + 480, + 334 + ], + "type": "inline_equation", + "content": "L" + }, + { + "bbox": [ + 130, + 243, + 480, + 334 + ], + "type": "text", + "content": " represents the number of ViT layers. This procedure can be formally defined as follows:" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 214, + 339, + 480, + 354 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 214, + 339, + 480, + 354 + ], + "spans": [ + { + "bbox": [ + 214, + 339, + 480, + 354 + ], + "type": "interline_equation", + "content": "\\mathbf {T} _ {c l s} ^ {i}, \\mathbf {T} _ {p a t c h} ^ {i} = \\mathbf {V} \\mathrm {i} \\mathbf {T} _ {i} \\left(\\mathbf {T} _ {c l s} ^ {i - 1}, \\mathbf {T} _ {p a t c h} ^ {i - 1}\\right), 1 \\leq i \\leq L \\tag {1}", + "image_path": "fdcad0f176cd1f0ee7ceb778992d0e02bdea347aee076da72a140a6401d0d192.jpg" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 130, + 360, + 482, + 521 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 360, + 482, + 521 + ], + "spans": [ + { + "bbox": [ + 130, + 360, + 482, + 521 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 130, + 360, + 482, + 521 + ], + "type": "inline_equation", + "content": "\\mathbf{V}\\mathbf{i}\\mathbf{T}_i" + }, + { + "bbox": [ + 130, + 360, + 482, + 521 + ], + "type": "text", + "content": " is " + }, + { + "bbox": [ + 130, + 360, + 482, + 521 + ], + "type": "inline_equation", + "content": "i^{th}" + }, + { + "bbox": [ + 130, + 360, + 482, + 521 + ], + "type": "text", + "content": " layer of ViT. The class tokens " + }, + { + "bbox": [ + 130, + 360, + 482, + 521 + ], + "type": "inline_equation", + "content": "\\mathbf{T}_{cls}^L" + }, + { + "bbox": [ + 130, + 360, + 482, + 521 + ], + "type": "text", + "content": " are processed through average pooling to derive the classification logits " + }, + { + "bbox": [ + 130, + 360, + 482, + 521 + ], + "type": "inline_equation", + "content": "y_{c} \\in \\mathbb{R}^{C}" + }, + { + "bbox": [ + 130, + 360, + 482, + 521 + ], + "type": "text", + "content": ". Meanwhile, the patch tokens " + }, + { + "bbox": [ + 130, + 360, + 482, + 521 + ], + "type": "inline_equation", + "content": "\\mathbf{T}_{patch}^L" + }, + { + "bbox": [ + 130, + 360, + 482, + 521 + ], + "type": "text", + "content": ", arranged according to their spatial positions, pass through a convolution layer to produce the Class Activation Maps (CAMs) " + }, + { + "bbox": [ + 130, + 360, + 482, + 521 + ], + "type": "inline_equation", + "content": "\\mathbf{M} \\in \\mathbb{R}^{C \\times N \\times N}" + }, + { + "bbox": [ + 130, + 360, + 482, + 521 + ], + "type": "text", + "content": ". By applying Global Average Pooling (GAP) to " + }, + { + "bbox": [ + 130, + 360, + 482, + 521 + ], + "type": "inline_equation", + "content": "\\mathbf{M}" + }, + { + "bbox": [ + 130, + 360, + 482, + 521 + ], + "type": "text", + "content": ", we can obtain classification logits " + }, + { + "bbox": [ + 130, + 360, + 482, + 521 + ], + "type": "inline_equation", + "content": "y_{p} \\in \\mathbb{R}^{C}" + }, + { + "bbox": [ + 130, + 360, + 482, + 521 + ], + "type": "text", + "content": " in a patch-level. These two classification logits are supervised using a multi-label soft margin loss " + }, + { + "bbox": [ + 130, + 360, + 482, + 521 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_{cls}" + }, + { + "bbox": [ + 130, + 360, + 482, + 521 + ], + "type": "text", + "content": " with the given classification labels " + }, + { + "bbox": [ + 130, + 360, + 482, + 521 + ], + "type": "inline_equation", + "content": "\\mathbf{y}" + }, + { + "bbox": [ + 130, + 360, + 482, + 521 + ], + "type": "text", + "content": ". Additionally, following [63], we extract the token-to-token attention " + }, + { + "bbox": [ + 130, + 360, + 482, + 521 + ], + "type": "inline_equation", + "content": "\\mathbf{A}_{t2t} \\in \\mathbb{R}^{(C + N^2) \\times (C + N^2)}" + }, + { + "bbox": [ + 130, + 360, + 482, + 521 + ], + "type": "text", + "content": " based on multi-head self-attention. From this token-to-token attention " + }, + { + "bbox": [ + 130, + 360, + 482, + 521 + ], + "type": "inline_equation", + "content": "\\mathbf{A}_{t2t}" + }, + { + "bbox": [ + 130, + 360, + 482, + 521 + ], + "type": "text", + "content": ", the class-to-patch attention " + }, + { + "bbox": [ + 130, + 360, + 482, + 521 + ], + "type": "inline_equation", + "content": "\\mathbf{A}_{c2p} \\in \\mathbb{R}^{C \\times N^2}" + }, + { + "bbox": [ + 130, + 360, + 482, + 521 + ], + "type": "text", + "content": " and patch-to-patch attention " + }, + { + "bbox": [ + 130, + 360, + 482, + 521 + ], + "type": "inline_equation", + "content": "\\mathbf{A}_{p2p} \\in \\mathbb{R}^{N^2 \\times N^2}" + }, + { + "bbox": [ + 130, + 360, + 482, + 521 + ], + "type": "text", + "content": " can be extracted. These two types of attention maps are used to refine the CAMs, forming a refined CAMs " + }, + { + "bbox": [ + 130, + 360, + 482, + 521 + ], + "type": "inline_equation", + "content": "\\mathbf{M}_{ref} \\in \\mathbb{R}^{C \\times N \\times N}" + }, + { + "bbox": [ + 130, + 360, + 482, + 521 + ], + "type": "text", + "content": ". The aforementioned process can be noted as follows:" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 249, + 529, + 480, + 541 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 249, + 529, + 480, + 541 + ], + "spans": [ + { + "bbox": [ + 249, + 529, + 480, + 541 + ], + "type": "interline_equation", + "content": "\\mathbf {M} _ {r e f} = \\left(\\mathbf {M} \\odot \\mathbf {A} _ {c 2 p}\\right) \\times \\mathbf {A} _ {p 2 p} \\tag {2}", + "image_path": "4a08e739ac31554065f23128055fef93ee1d0b343c4e0dd744b5941fbca63431.jpg" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 130, + 546, + 479, + 594 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 546, + 479, + 594 + ], + "spans": [ + { + "bbox": [ + 130, + 546, + 479, + 594 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 130, + 546, + 479, + 594 + ], + "type": "inline_equation", + "content": "\\odot" + }, + { + "bbox": [ + 130, + 546, + 479, + 594 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 130, + 546, + 479, + 594 + ], + "type": "inline_equation", + "content": "\\times" + }, + { + "bbox": [ + 130, + 546, + 479, + 594 + ], + "type": "text", + "content": " are Hadamard product and matrix multiplication, respectively. Discrete Fourier Transform Let " + }, + { + "bbox": [ + 130, + 546, + 479, + 594 + ], + "type": "inline_equation", + "content": "\\mathcal{F}" + }, + { + "bbox": [ + 130, + 546, + 479, + 594 + ], + "type": "text", + "content": " denote the Discrete Fourier Transform (DFT) [54] applied across the spatial dimensions. For a feature map " + }, + { + "bbox": [ + 130, + 546, + 479, + 594 + ], + "type": "inline_equation", + "content": "\\mathbf{X} \\in \\mathbb{R}^{D \\times H \\times W}" + }, + { + "bbox": [ + 130, + 546, + 479, + 594 + ], + "type": "text", + "content": " the result of applying " + }, + { + "bbox": [ + 130, + 546, + 479, + 594 + ], + "type": "inline_equation", + "content": "\\mathcal{F}" + }, + { + "bbox": [ + 130, + 546, + 479, + 594 + ], + "type": "text", + "content": " to a single channel can be represented as follows:" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 199, + 601, + 480, + 622 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 199, + 601, + 480, + 622 + ], + "spans": [ + { + "bbox": [ + 199, + 601, + 480, + 622 + ], + "type": "interline_equation", + "content": "\\mathcal {F} (\\mathbf {X}) (m, n) = \\sum_ {h, w} \\mathbf {X} (h, w) e ^ {- j 2 \\pi \\left(\\frac {h}{H} m + \\frac {w}{W} n\\right)}, j ^ {2} = - 1 \\tag {3}", + "image_path": "6d7ab712a5f04d2cd4df709b37c36196dbe549a8177ecce17c368b82d82a1fcf.jpg" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 130, + 629, + 479, + 666 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 629, + 479, + 666 + ], + "spans": [ + { + "bbox": [ + 130, + 629, + 479, + 666 + ], + "type": "text", + "content": "Since the absolute value and the angle component of " + }, + { + "bbox": [ + 130, + 629, + 479, + 666 + ], + "type": "inline_equation", + "content": "\\mathcal{F}(\\mathbf{X})" + }, + { + "bbox": [ + 130, + 629, + 479, + 666 + ], + "type": "text", + "content": " respectively represent magnitude and phase components, we denote the magnitude spectrum as " + }, + { + "bbox": [ + 130, + 629, + 479, + 666 + ], + "type": "inline_equation", + "content": "|\\mathcal{F}(\\mathbf{X})|" + }, + { + "bbox": [ + 130, + 629, + 479, + 666 + ], + "type": "text", + "content": " and the phase spectrum as " + }, + { + "bbox": [ + 130, + 629, + 479, + 666 + ], + "type": "inline_equation", + "content": "\\angle \\mathcal{F}(\\mathbf{X})" + }, + { + "bbox": [ + 130, + 629, + 479, + 666 + ], + "type": "text", + "content": " throughout the paper." + } + ] + } + ], + "index": 11 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 212, + 91, + 448, + 102 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 212, + 91, + 448, + 102 + ], + "spans": [ + { + "bbox": [ + 212, + 91, + 448, + 102 + ], + "type": "text", + "content": "Phase Concentration and Shortcut Suppression for WSSS" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 474, + 92, + 480, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 474, + 92, + 480, + 100 + ], + "spans": [ + { + "bbox": [ + 474, + 92, + 480, + 100 + ], + "type": "text", + "content": "5" + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 4 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 133, + 116, + 482, + 295 + ], + "blocks": [ + { + "bbox": [ + 133, + 116, + 482, + 295 + ], + "lines": [ + { + "bbox": [ + 133, + 116, + 482, + 295 + ], + "spans": [ + { + "bbox": [ + 133, + 116, + 482, + 295 + ], + "type": "image", + "image_path": "e8121360ec3292f87495d4a534d49c1a53760efad3dfb1475008820b494d8053.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 130, + 305, + 483, + 404 + ], + "lines": [ + { + "bbox": [ + 130, + 305, + 483, + 404 + ], + "spans": [ + { + "bbox": [ + 130, + 305, + 483, + 404 + ], + "type": "text", + "content": "Fig. 2: Visualization of the proposed framework. The Magnitude-mixing-based Phase Concentration (MPC) is executed at an intermediate layer. Here, the magnitude of features from the anchor image " + }, + { + "bbox": [ + 130, + 305, + 483, + 404 + ], + "type": "inline_equation", + "content": "\\mathbf{I}" + }, + { + "bbox": [ + 130, + 305, + 483, + 404 + ], + "type": "text", + "content": " and another arbitrary image " + }, + { + "bbox": [ + 130, + 305, + 483, + 404 + ], + "type": "inline_equation", + "content": "\\hat{\\mathbf{I}}" + }, + { + "bbox": [ + 130, + 305, + 483, + 404 + ], + "type": "text", + "content": " are mixed to make a feature with identical phase, but different magnitude. The model is guided to perform correct classification and derive consistent CAMs " + }, + { + "bbox": [ + 130, + 305, + 483, + 404 + ], + "type": "inline_equation", + "content": "\\mathbf{M}_{ref}" + }, + { + "bbox": [ + 130, + 305, + 483, + 404 + ], + "type": "text", + "content": " from a pair of features. In the Frequency Shortcut Suppression (FSS), based on the Frequency Shortcut Potential Map (FSPM), frequency components that have a significant impact on the classification of classes in " + }, + { + "bbox": [ + 130, + 305, + 483, + 404 + ], + "type": "inline_equation", + "content": "\\mathbf{I}" + }, + { + "bbox": [ + 130, + 305, + 483, + 404 + ], + "type": "text", + "content": " are reduced before being inputted to the ViT. The model is encouraged to classify correctly and minimize the difference between " + }, + { + "bbox": [ + 130, + 305, + 483, + 404 + ], + "type": "inline_equation", + "content": "\\mathbf{M}" + }, + { + "bbox": [ + 130, + 305, + 483, + 404 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 130, + 305, + 483, + 404 + ], + "type": "inline_equation", + "content": "\\hat{\\mathbf{M}}" + }, + { + "bbox": [ + 130, + 305, + 483, + 404 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_caption" + } + ], + "index": 2 + }, + { + "bbox": [ + 132, + 429, + 394, + 441 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 429, + 394, + 441 + ], + "spans": [ + { + "bbox": [ + 132, + 429, + 394, + 441 + ], + "type": "text", + "content": "3.2 Magnitude-mixing-based Phase Concentration" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 130, + 449, + 482, + 533 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 449, + 482, + 533 + ], + "spans": [ + { + "bbox": [ + 130, + 449, + 482, + 533 + ], + "type": "text", + "content": "With the aid of multi-head self-attention, CAMs from ViT are more proficient in capturing the less-discriminative regions than CNN-based methods. Nonetheless, they still show imprecise localization along object boundaries due to a lack of spatial constraint. To improve localization of CAMs, we design the Magnitude-mixing-based Phase Concentration (MPC) module to extract high-level semantic and boundary information shown to be present in the phase component of the frequency domain [16, 42]." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 130, + 534, + 483, + 668 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 534, + 483, + 668 + ], + "spans": [ + { + "bbox": [ + 130, + 534, + 483, + 668 + ], + "type": "text", + "content": "We consider two ideas in MPC to guide the model to prioritize the phase component: 1) The model should correctly classify the image even when the magnitude component is altered. Similarly, 2) the CAMs generated after the magnitude change should be similar to those obtained before the change. The key to these two ideas is to change the magnitude while keeping the phase component intact. Since the magnitude needs to be 'feasible' yet different, we designed a magnitude-mixing method. Here, we form a random image pair consisting of an anchor image " + }, + { + "bbox": [ + 130, + 534, + 483, + 668 + ], + "type": "inline_equation", + "content": "\\mathbf{I}" + }, + { + "bbox": [ + 130, + 534, + 483, + 668 + ], + "type": "text", + "content": " and an arbitrary image " + }, + { + "bbox": [ + 130, + 534, + 483, + 668 + ], + "type": "inline_equation", + "content": "\\dot{\\mathbf{I}}" + }, + { + "bbox": [ + 130, + 534, + 483, + 668 + ], + "type": "text", + "content": " for perturbation. Let the input class and patch tokens of " + }, + { + "bbox": [ + 130, + 534, + 483, + 668 + ], + "type": "inline_equation", + "content": "\\mathbf{I}" + }, + { + "bbox": [ + 130, + 534, + 483, + 668 + ], + "type": "text", + "content": " be denoted as " + }, + { + "bbox": [ + 130, + 534, + 483, + 668 + ], + "type": "inline_equation", + "content": "[\\mathbf{T}_{cls}^{0},\\mathbf{T}_{patch}^{0}]" + }, + { + "bbox": [ + 130, + 534, + 483, + 668 + ], + "type": "text", + "content": ", and those of " + }, + { + "bbox": [ + 130, + 534, + 483, + 668 + ], + "type": "inline_equation", + "content": "\\dot{\\mathbf{I}}" + }, + { + "bbox": [ + 130, + 534, + 483, + 668 + ], + "type": "text", + "content": " as " + }, + { + "bbox": [ + 130, + 534, + 483, + 668 + ], + "type": "inline_equation", + "content": "[\\dot{\\mathbf{T}}_{cls}^{0},\\dot{\\mathbf{T}}_{patch}^{0}]" + }, + { + "bbox": [ + 130, + 534, + 483, + 668 + ], + "type": "text", + "content": ". We then process each set of tokens through the ViT layers, as outlined in Eq. 1, for " + }, + { + "bbox": [ + 130, + 534, + 483, + 668 + ], + "type": "inline_equation", + "content": "1\\leq i\\leq \\ell" + }, + { + "bbox": [ + 130, + 534, + 483, + 668 + ], + "type": "text", + "content": ", to obtain " + }, + { + "bbox": [ + 130, + 534, + 483, + 668 + ], + "type": "inline_equation", + "content": "[\\mathbf{T}_{cls}^{\\ell},\\mathbf{T}_{patch}^{\\ell}]" + }, + { + "bbox": [ + 130, + 534, + 483, + 668 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 130, + 534, + 483, + 668 + ], + "type": "inline_equation", + "content": "[\\dot{\\mathbf{T}}_{cls}^{\\ell},\\dot{\\mathbf{T}}_{patch}^{\\ell}]" + }, + { + "bbox": [ + 130, + 534, + 483, + 668 + ], + "type": "text", + "content": " respectively. After reshaping" + } + ] + } + ], + "index": 6 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "spans": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "type": "text", + "content": "6" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 166, + 91, + 227, + 101 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 166, + 91, + 227, + 101 + ], + "spans": [ + { + "bbox": [ + 166, + 91, + 227, + 101 + ], + "type": "text", + "content": "H. Kwon et al." + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 5 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 143, + 114, + 470, + 205 + ], + "blocks": [ + { + "bbox": [ + 143, + 114, + 470, + 205 + ], + "lines": [ + { + "bbox": [ + 143, + 114, + 470, + 205 + ], + "spans": [ + { + "bbox": [ + 143, + 114, + 470, + 205 + ], + "type": "image", + "image_path": "923a377c43b81f8725a8592a94044bef4ef0331b6b6b973ed1302bd41ba38967.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 131, + 213, + 482, + 280 + ], + "lines": [ + { + "bbox": [ + 131, + 213, + 482, + 280 + ], + "spans": [ + { + "bbox": [ + 131, + 213, + 482, + 280 + ], + "type": "text", + "content": "Fig. 3: Visualization of Frequency Influence Measurement. For an image " + }, + { + "bbox": [ + 131, + 213, + 482, + 280 + ], + "type": "inline_equation", + "content": "\\mathbf{I}" + }, + { + "bbox": [ + 131, + 213, + 482, + 280 + ], + "type": "text", + "content": " containing class " + }, + { + "bbox": [ + 131, + 213, + 482, + 280 + ], + "type": "inline_equation", + "content": "c" + }, + { + "bbox": [ + 131, + 213, + 482, + 280 + ], + "type": "text", + "content": ", a set of images that each masked at different frequency components using " + }, + { + "bbox": [ + 131, + 213, + 482, + 280 + ], + "type": "inline_equation", + "content": "M_{f}" + }, + { + "bbox": [ + 131, + 213, + 482, + 280 + ], + "type": "text", + "content": " in the frequency domain of " + }, + { + "bbox": [ + 131, + 213, + 482, + 280 + ], + "type": "inline_equation", + "content": "\\mathbf{I}" + }, + { + "bbox": [ + 131, + 213, + 482, + 280 + ], + "type": "text", + "content": " are inputted into ViT. The classification loss for class " + }, + { + "bbox": [ + 131, + 213, + 482, + 280 + ], + "type": "inline_equation", + "content": "c" + }, + { + "bbox": [ + 131, + 213, + 482, + 280 + ], + "type": "text", + "content": " is calculated as the influence of frequency " + }, + { + "bbox": [ + 131, + 213, + 482, + 280 + ], + "type": "inline_equation", + "content": "f" + }, + { + "bbox": [ + 131, + 213, + 482, + 280 + ], + "type": "text", + "content": " on class " + }, + { + "bbox": [ + 131, + 213, + 482, + 280 + ], + "type": "inline_equation", + "content": "c" + }, + { + "bbox": [ + 131, + 213, + 482, + 280 + ], + "type": "text", + "content": ". Accumulate the process across the image set " + }, + { + "bbox": [ + 131, + 213, + 482, + 280 + ], + "type": "inline_equation", + "content": "\\mathbb{I}_c" + }, + { + "bbox": [ + 131, + 213, + 482, + 280 + ], + "type": "text", + "content": " to produce " + }, + { + "bbox": [ + 131, + 213, + 482, + 280 + ], + "type": "inline_equation", + "content": "\\mathbf{P}(c)" + }, + { + "bbox": [ + 131, + 213, + 482, + 280 + ], + "type": "text", + "content": ", and repeat for each " + }, + { + "bbox": [ + 131, + 213, + 482, + 280 + ], + "type": "inline_equation", + "content": "c" + }, + { + "bbox": [ + 131, + 213, + 482, + 280 + ], + "type": "text", + "content": ". Finally, " + }, + { + "bbox": [ + 131, + 213, + 482, + 280 + ], + "type": "inline_equation", + "content": "\\mathbf{P}" + }, + { + "bbox": [ + 131, + 213, + 482, + 280 + ], + "type": "text", + "content": " is normalized to obtain the Frequency Shortcut Potential Map (FSPM) for each class." + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_caption" + } + ], + "index": 2 + }, + { + "bbox": [ + 131, + 300, + 480, + 324 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 131, + 300, + 480, + 324 + ], + "spans": [ + { + "bbox": [ + 131, + 300, + 480, + 324 + ], + "type": "text", + "content": "and aligning " + }, + { + "bbox": [ + 131, + 300, + 480, + 324 + ], + "type": "inline_equation", + "content": "\\mathbf{T}_{patch}^{\\ell}" + }, + { + "bbox": [ + 131, + 300, + 480, + 324 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 131, + 300, + 480, + 324 + ], + "type": "inline_equation", + "content": "\\bar{\\mathbf{T}}_{patch}^{\\ell}" + }, + { + "bbox": [ + 131, + 300, + 480, + 324 + ], + "type": "text", + "content": " according to their spatial positions, the spectral magnitude is obtained and perturbed using the following formula:" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 165, + 329, + 481, + 345 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 165, + 329, + 481, + 345 + ], + "spans": [ + { + "bbox": [ + 165, + 329, + 481, + 345 + ], + "type": "interline_equation", + "content": "\\bar {\\mathbf {T}} _ {p a t c h} ^ {\\ell} = \\mathcal {F} ^ {- 1} \\left( \\right.\\left\\langle \\right.\\left\\langle \\right. r \\mid \\mathcal {F} \\left(\\mathbf {T} _ {p a t c h} ^ {\\ell}\\right)\\left. \\right| + (1 - r) \\mid \\mathcal {F} \\left(\\dot {\\mathbf {T}} _ {p a t c h} ^ {\\ell}\\right)\\left. \\right|, \\angle \\mathcal {F} \\left(\\mathbf {T} _ {p a t c h} ^ {\\ell}\\right)\\left. \\right\\rangle \\rangle\\left. \\right) \\tag {4}", + "image_path": "a118e49a2c0fd58ef681530c82a4e7642527edae5d53d337fdaa3269b99ab4a4.jpg" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 131, + 350, + 482, + 499 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 131, + 350, + 482, + 499 + ], + "spans": [ + { + "bbox": [ + 131, + 350, + 482, + 499 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 131, + 350, + 482, + 499 + ], + "type": "inline_equation", + "content": "r" + }, + { + "bbox": [ + 131, + 350, + 482, + 499 + ], + "type": "text", + "content": " is the perturbation ratio between the anchor image and arbitrary image, and " + }, + { + "bbox": [ + 131, + 350, + 482, + 499 + ], + "type": "inline_equation", + "content": "\\mathcal{F}^{-1}" + }, + { + "bbox": [ + 131, + 350, + 482, + 499 + ], + "type": "text", + "content": " is the inverse DFT. Tokens and maps that have been affected by magnitude-mixing are denoted by a bar on top (i.e. " + }, + { + "bbox": [ + 131, + 350, + 482, + 499 + ], + "type": "inline_equation", + "content": "\\bar{\\mathbf{T}}_{patch}^{\\ell}" + }, + { + "bbox": [ + 131, + 350, + 482, + 499 + ], + "type": "text", + "content": "). " + }, + { + "bbox": [ + 131, + 350, + 482, + 499 + ], + "type": "inline_equation", + "content": "\\langle \\langle A,B\\rangle \\rangle" + }, + { + "bbox": [ + 131, + 350, + 482, + 499 + ], + "type": "text", + "content": " is coupling of magnitude " + }, + { + "bbox": [ + 131, + 350, + 482, + 499 + ], + "type": "inline_equation", + "content": "A" + }, + { + "bbox": [ + 131, + 350, + 482, + 499 + ], + "type": "text", + "content": " and phase " + }, + { + "bbox": [ + 131, + 350, + 482, + 499 + ], + "type": "inline_equation", + "content": "B" + }, + { + "bbox": [ + 131, + 350, + 482, + 499 + ], + "type": "text", + "content": ". As shown in Eq. 4, the phase component of " + }, + { + "bbox": [ + 131, + 350, + 482, + 499 + ], + "type": "inline_equation", + "content": "\\bar{\\mathbf{T}}_{patch}^{\\ell}" + }, + { + "bbox": [ + 131, + 350, + 482, + 499 + ], + "type": "text", + "content": " remains unchanged as the magnitude is perturbed. After iDFT, the set of tokens " + }, + { + "bbox": [ + 131, + 350, + 482, + 499 + ], + "type": "inline_equation", + "content": "[\\mathbf{T}_{cls}^{\\ell},\\bar{\\mathbf{T}}_{patch}^{\\ell}]" + }, + { + "bbox": [ + 131, + 350, + 482, + 499 + ], + "type": "text", + "content": " are reshaped and passed through the remaining layers of ViT, following Eq. 1 from " + }, + { + "bbox": [ + 131, + 350, + 482, + 499 + ], + "type": "inline_equation", + "content": "\\ell < i\\leq L" + }, + { + "bbox": [ + 131, + 350, + 482, + 499 + ], + "type": "text", + "content": ", to produce the output tokens " + }, + { + "bbox": [ + 131, + 350, + 482, + 499 + ], + "type": "inline_equation", + "content": "[\\bar{\\mathbf{T}}_{cls}^{L},\\bar{\\mathbf{T}}_{patch}^{L}]" + }, + { + "bbox": [ + 131, + 350, + 482, + 499 + ], + "type": "text", + "content": ". The class prediction " + }, + { + "bbox": [ + 131, + 350, + 482, + 499 + ], + "type": "inline_equation", + "content": "\\bar{y}_c" + }, + { + "bbox": [ + 131, + 350, + 482, + 499 + ], + "type": "text", + "content": " is obtained from " + }, + { + "bbox": [ + 131, + 350, + 482, + 499 + ], + "type": "inline_equation", + "content": "\\bar{\\mathbf{T}}_{cls}^{L}" + }, + { + "bbox": [ + 131, + 350, + 482, + 499 + ], + "type": "text", + "content": ", while the class prediction " + }, + { + "bbox": [ + 131, + 350, + 482, + 499 + ], + "type": "inline_equation", + "content": "\\bar{y}_p" + }, + { + "bbox": [ + 131, + 350, + 482, + 499 + ], + "type": "text", + "content": " and refined CAMs " + }, + { + "bbox": [ + 131, + 350, + 482, + 499 + ], + "type": "inline_equation", + "content": "\\bar{\\mathbf{M}}_{ref}" + }, + { + "bbox": [ + 131, + 350, + 482, + 499 + ], + "type": "text", + "content": " are generated from " + }, + { + "bbox": [ + 131, + 350, + 482, + 499 + ], + "type": "inline_equation", + "content": "\\bar{\\mathbf{T}}_{patch}^{L}" + }, + { + "bbox": [ + 131, + 350, + 482, + 499 + ], + "type": "text", + "content": ". " + }, + { + "bbox": [ + 131, + 350, + 482, + 499 + ], + "type": "inline_equation", + "content": "\\bar{y}_c" + }, + { + "bbox": [ + 131, + 350, + 482, + 499 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 131, + 350, + 482, + 499 + ], + "type": "inline_equation", + "content": "\\bar{y}_p" + }, + { + "bbox": [ + 131, + 350, + 482, + 499 + ], + "type": "text", + "content": " are used as class predictions in multi-label soft margin loss to obtain " + }, + { + "bbox": [ + 131, + 350, + 482, + 499 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_{mpc - cls}" + }, + { + "bbox": [ + 131, + 350, + 482, + 499 + ], + "type": "text", + "content": ". Following the second idea, the difference between the refined CAMs " + }, + { + "bbox": [ + 131, + 350, + 482, + 499 + ], + "type": "inline_equation", + "content": "\\bar{\\mathbf{M}}_{ref}" + }, + { + "bbox": [ + 131, + 350, + 482, + 499 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 131, + 350, + 482, + 499 + ], + "type": "inline_equation", + "content": "\\bar{\\mathbf{M}}_{ref}" + }, + { + "bbox": [ + 131, + 350, + 482, + 499 + ], + "type": "text", + "content": " are minimized using the following loss:" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 244, + 505, + 480, + 519 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 244, + 505, + 480, + 519 + ], + "spans": [ + { + "bbox": [ + 244, + 505, + 480, + 519 + ], + "type": "interline_equation", + "content": "\\mathcal {L} _ {\\text {m p c - c a m s}} = \\left| \\mathbf {M} _ {\\text {r e f}} - \\bar {\\mathbf {M}} _ {\\text {r e f}} \\right| _ {1} \\tag {5}", + "image_path": "d4b05ef204d67e75e2df8a652b1b33309e9c6b967a227c8af3171b918ec2e1a9.jpg" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 132, + 524, + 405, + 536 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 524, + 405, + 536 + ], + "spans": [ + { + "bbox": [ + 132, + 524, + 405, + 536 + ], + "type": "text", + "content": "Ultimately, the MPC is supervised through the following loss:" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 239, + 543, + 480, + 555 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 239, + 543, + 480, + 555 + ], + "spans": [ + { + "bbox": [ + 239, + 543, + 480, + 555 + ], + "type": "interline_equation", + "content": "\\mathcal {L} _ {m p c} = \\mathcal {L} _ {m p c - c l s} + \\lambda_ {1} \\mathcal {L} _ {m p c - c a m s} \\tag {6}", + "image_path": "1822ae872cfa42195841d4ac50a1edb81c15cc09a8df92a30c6bb42d8a182803.jpg" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 131, + 561, + 480, + 609 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 131, + 561, + 480, + 609 + ], + "spans": [ + { + "bbox": [ + 131, + 561, + 480, + 609 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 131, + 561, + 480, + 609 + ], + "type": "inline_equation", + "content": "\\lambda_{1}" + }, + { + "bbox": [ + 131, + 561, + 480, + 609 + ], + "type": "text", + "content": " is a hyperparameter to balance the losses. Additionally, the gradient to " + }, + { + "bbox": [ + 131, + 561, + 480, + 609 + ], + "type": "inline_equation", + "content": "\\mathbf{M}_{ref}" + }, + { + "bbox": [ + 131, + 561, + 480, + 609 + ], + "type": "text", + "content": " from " + }, + { + "bbox": [ + 131, + 561, + 480, + 609 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_{mpc-cams}" + }, + { + "bbox": [ + 131, + 561, + 480, + 609 + ], + "type": "text", + "content": " is detached to guide " + }, + { + "bbox": [ + 131, + 561, + 480, + 609 + ], + "type": "inline_equation", + "content": "\\bar{\\mathbf{M}}_{ref}" + }, + { + "bbox": [ + 131, + 561, + 480, + 609 + ], + "type": "text", + "content": " to follow " + }, + { + "bbox": [ + 131, + 561, + 480, + 609 + ], + "type": "inline_equation", + "content": "\\mathbf{M}_{ref}" + }, + { + "bbox": [ + 131, + 561, + 480, + 609 + ], + "type": "text", + "content": ". This increases the dependency of the model on the phase information, enabling the model to better understand the boundary information in the phase component." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 132, + 624, + 324, + 636 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 624, + 324, + 636 + ], + "spans": [ + { + "bbox": [ + 132, + 624, + 324, + 636 + ], + "type": "text", + "content": "3.3 Frequency Shortcut Suppression" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 131, + 641, + 481, + 665 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 131, + 641, + 481, + 665 + ], + "spans": [ + { + "bbox": [ + 131, + 641, + 481, + 665 + ], + "type": "text", + "content": "Through the MPC module, we have addressed the issues of activation around the boundary in CAMs. However, the problem of false activations in CAMs" + } + ] + } + ], + "index": 12 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 212, + 91, + 448, + 102 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 212, + 91, + 448, + 102 + ], + "spans": [ + { + "bbox": [ + 212, + 91, + 448, + 102 + ], + "type": "text", + "content": "Phase Concentration and Shortcut Suppression for WSSS" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 474, + 92, + 480, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 474, + 92, + 480, + 100 + ], + "spans": [ + { + "bbox": [ + 474, + 92, + 480, + 100 + ], + "type": "text", + "content": "7" + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 6 + }, + { + "para_blocks": [ + { + "bbox": [ + 130, + 116, + 482, + 272 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 116, + 482, + 272 + ], + "spans": [ + { + "bbox": [ + 130, + 116, + 482, + 272 + ], + "type": "text", + "content": "persists. Various factors contribute to false activations in ViT, such as GAP for generating classification logits from CAMs and the over-smoothing problem caused by self-attention of ViT [49]. Interestingly, we experimentally found that false activations can also arise due to Shortcut Learning in the frequency domain [14], which we aim to address. Shortcut Learning during the classification training process in ViT can negatively impact genuine semantic understanding [40], and such shortcuts can lead to wrong activations [41]. Furthermore, it has been revealed that Neural Networks can create shortcuts in the frequency domain [58]. We empirically found a correlation between frequency shortcuts and false positives in CAMs M. Since Frequency Shortcuts are unrelated to objectness, activations caused by frequency shortcuts occur in locations unrelated to objects. We have resolved over-activation by disrupting the creation of frequency shortcuts through the Frequency Shortcut Suppression(FSS) module." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 130, + 272, + 483, + 465 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 272, + 483, + 465 + ], + "spans": [ + { + "bbox": [ + 130, + 272, + 483, + 465 + ], + "type": "text", + "content": "Frequency Shortcut Potential Map To implement the FSS, the potential of specific frequency components to be used as shortcuts in classification needs to be measured. Inspired by Wang et al. [58], we utilized a Frequency Influence Measurement (FIM) framework, as depicted in Fig. 3, to assess the impact of each component. If the classification loss of the model increases significantly for an image " + }, + { + "bbox": [ + 130, + 272, + 483, + 465 + ], + "type": "inline_equation", + "content": "\\mathbf{I}" + }, + { + "bbox": [ + 130, + 272, + 483, + 465 + ], + "type": "text", + "content": " of class " + }, + { + "bbox": [ + 130, + 272, + 483, + 465 + ], + "type": "inline_equation", + "content": "c" + }, + { + "bbox": [ + 130, + 272, + 483, + 465 + ], + "type": "text", + "content": " when information about a certain frequency component " + }, + { + "bbox": [ + 130, + 272, + 483, + 465 + ], + "type": "inline_equation", + "content": "f" + }, + { + "bbox": [ + 130, + 272, + 483, + 465 + ], + "type": "text", + "content": " is removed, it indicates that " + }, + { + "bbox": [ + 130, + 272, + 483, + 465 + ], + "type": "inline_equation", + "content": "f" + }, + { + "bbox": [ + 130, + 272, + 483, + 465 + ], + "type": "text", + "content": " has an important role in classifying class " + }, + { + "bbox": [ + 130, + 272, + 483, + 465 + ], + "type": "inline_equation", + "content": "c" + }, + { + "bbox": [ + 130, + 272, + 483, + 465 + ], + "type": "text", + "content": " for the model. Therefore, for each class " + }, + { + "bbox": [ + 130, + 272, + 483, + 465 + ], + "type": "inline_equation", + "content": "c" + }, + { + "bbox": [ + 130, + 272, + 483, + 465 + ], + "type": "text", + "content": ", the influence of " + }, + { + "bbox": [ + 130, + 272, + 483, + 465 + ], + "type": "inline_equation", + "content": "f" + }, + { + "bbox": [ + 130, + 272, + 483, + 465 + ], + "type": "text", + "content": " on the classification of the image set " + }, + { + "bbox": [ + 130, + 272, + 483, + 465 + ], + "type": "inline_equation", + "content": "\\mathbb{I}_c" + }, + { + "bbox": [ + 130, + 272, + 483, + 465 + ], + "type": "text", + "content": " corresponding to that class is calculated by accumulating classification loss. To remove a specific frequency component " + }, + { + "bbox": [ + 130, + 272, + 483, + 465 + ], + "type": "inline_equation", + "content": "f" + }, + { + "bbox": [ + 130, + 272, + 483, + 465 + ], + "type": "text", + "content": " from an image, a Masking map " + }, + { + "bbox": [ + 130, + 272, + 483, + 465 + ], + "type": "inline_equation", + "content": "M_f \\in \\mathbb{R}^{H \\times W}" + }, + { + "bbox": [ + 130, + 272, + 483, + 465 + ], + "type": "text", + "content": " is used, where the value for " + }, + { + "bbox": [ + 130, + 272, + 483, + 465 + ], + "type": "inline_equation", + "content": "f" + }, + { + "bbox": [ + 130, + 272, + 483, + 465 + ], + "type": "text", + "content": " is set to 0, and others are set to 1. For an image " + }, + { + "bbox": [ + 130, + 272, + 483, + 465 + ], + "type": "inline_equation", + "content": "\\mathbf{I}" + }, + { + "bbox": [ + 130, + 272, + 483, + 465 + ], + "type": "text", + "content": ", let " + }, + { + "bbox": [ + 130, + 272, + 483, + 465 + ], + "type": "inline_equation", + "content": "\\mathbf{I}_f = \\mathcal{F}^{-1}(\\mathcal{F}(\\mathbf{I}) \\odot M_f)" + }, + { + "bbox": [ + 130, + 272, + 483, + 465 + ], + "type": "text", + "content": " represent the image with the " + }, + { + "bbox": [ + 130, + 272, + 483, + 465 + ], + "type": "inline_equation", + "content": "f" + }, + { + "bbox": [ + 130, + 272, + 483, + 465 + ], + "type": "text", + "content": " component masked where " + }, + { + "bbox": [ + 130, + 272, + 483, + 465 + ], + "type": "inline_equation", + "content": "\\odot" + }, + { + "bbox": [ + 130, + 272, + 483, + 465 + ], + "type": "text", + "content": " is element-wise multiplication. The classification logit derived after passing " + }, + { + "bbox": [ + 130, + 272, + 483, + 465 + ], + "type": "inline_equation", + "content": "\\mathbf{I}_f" + }, + { + "bbox": [ + 130, + 272, + 483, + 465 + ], + "type": "text", + "content": " through a ViT is denoted by " + }, + { + "bbox": [ + 130, + 272, + 483, + 465 + ], + "type": "inline_equation", + "content": "y_{\\mathbf{I}_f}" + }, + { + "bbox": [ + 130, + 272, + 483, + 465 + ], + "type": "text", + "content": ". When " + }, + { + "bbox": [ + 130, + 272, + 483, + 465 + ], + "type": "inline_equation", + "content": "\\mathbf{P} \\in \\mathbb{R}^{C \\times H \\times W}" + }, + { + "bbox": [ + 130, + 272, + 483, + 465 + ], + "type": "text", + "content": " is frequency influence accumulated map, the influence of the frequency component " + }, + { + "bbox": [ + 130, + 272, + 483, + 465 + ], + "type": "inline_equation", + "content": "f" + }, + { + "bbox": [ + 130, + 272, + 483, + 465 + ], + "type": "text", + "content": " on the classification of class " + }, + { + "bbox": [ + 130, + 272, + 483, + 465 + ], + "type": "inline_equation", + "content": "c" + }, + { + "bbox": [ + 130, + 272, + 483, + 465 + ], + "type": "text", + "content": " is calculated as follows:" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 247, + 472, + 482, + 495 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 247, + 472, + 482, + 495 + ], + "spans": [ + { + "bbox": [ + 247, + 472, + 482, + 495 + ], + "type": "interline_equation", + "content": "\\mathbf {P} (c, f) = \\sum_ {\\mathbf {I} \\in \\mathbb {I} _ {c}} \\mathcal {L} _ {b c e} \\left(y _ {\\mathbf {I} _ {f}} (c), 1\\right) \\tag {7}", + "image_path": "fd46876f5f5c530ef49ae0b57baf30f3a3f5fcfaa3ecbe9b40d53d32be28abc5.jpg" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 130, + 502, + 482, + 563 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 502, + 482, + 563 + ], + "spans": [ + { + "bbox": [ + 130, + 502, + 482, + 563 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 130, + 502, + 482, + 563 + ], + "type": "inline_equation", + "content": "y(c)" + }, + { + "bbox": [ + 130, + 502, + 482, + 563 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 130, + 502, + 482, + 563 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_{bce}" + }, + { + "bbox": [ + 130, + 502, + 482, + 563 + ], + "type": "text", + "content": " are probability of class " + }, + { + "bbox": [ + 130, + 502, + 482, + 563 + ], + "type": "inline_equation", + "content": "c" + }, + { + "bbox": [ + 130, + 502, + 482, + 563 + ], + "type": "text", + "content": " and binary cross-entropy loss, respectively. " + }, + { + "bbox": [ + 130, + 502, + 482, + 563 + ], + "type": "inline_equation", + "content": "\\mathbf{P}(c)\\in \\mathbb{R}^{H\\times W}" + }, + { + "bbox": [ + 130, + 502, + 482, + 563 + ], + "type": "text", + "content": " ultimately holds the classification influence information of each frequency for class " + }, + { + "bbox": [ + 130, + 502, + 482, + 563 + ], + "type": "inline_equation", + "content": "c" + }, + { + "bbox": [ + 130, + 502, + 482, + 563 + ], + "type": "text", + "content": ", with higher values indicating a significant impact on classifying class " + }, + { + "bbox": [ + 130, + 502, + 482, + 563 + ], + "type": "inline_equation", + "content": "c" + }, + { + "bbox": [ + 130, + 502, + 482, + 563 + ], + "type": "text", + "content": ". " + }, + { + "bbox": [ + 130, + 502, + 482, + 563 + ], + "type": "inline_equation", + "content": "\\mathbf{P}" + }, + { + "bbox": [ + 130, + 502, + 482, + 563 + ], + "type": "text", + "content": " is normalized for each class to create the Frequency Shortcut Potential Map (FSPM) " + }, + { + "bbox": [ + 130, + 502, + 482, + 563 + ], + "type": "inline_equation", + "content": "\\mathbf{FS}\\in \\mathbb{R}^{C\\times H\\times W}" + }, + { + "bbox": [ + 130, + 502, + 482, + 563 + ], + "type": "text", + "content": ":" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 225, + 573, + 481, + 586 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 225, + 573, + 481, + 586 + ], + "spans": [ + { + "bbox": [ + 225, + 573, + 481, + 586 + ], + "type": "interline_equation", + "content": "\\mathbf {F S} (c) = \\operatorname {N o r m} (\\mathbf {P} (c)), c \\in \\{1, 2, \\dots , C \\} \\tag {8}", + "image_path": "2bc45392f4186b7a1798cb43e2bb75beaa35633b574ff6caf5579224c4e17381.jpg" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 130, + 594, + 482, + 616 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 594, + 482, + 616 + ], + "spans": [ + { + "bbox": [ + 130, + 594, + 482, + 616 + ], + "type": "text", + "content": "where Norm is the normalization function that scales each value to fall between 0 and 1." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 130, + 617, + 482, + 666 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 617, + 482, + 666 + ], + "spans": [ + { + "bbox": [ + 130, + 617, + 482, + 666 + ], + "type": "text", + "content": "Frequency Shortcut Suppression We propose the Frequency Shortcut Suppression (FSS) utilizing the FSPM generated through the FIM process. The FSPM encapsulates information on which frequency components significantly impact classification at the time of FIM execution, indicating frequencies that" + } + ] + } + ], + "index": 8 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "spans": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "type": "text", + "content": "8" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 166, + 91, + 227, + 101 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 166, + 91, + 227, + 101 + ], + "spans": [ + { + "bbox": [ + 166, + 91, + 227, + 101 + ], + "type": "text", + "content": "H. Kwon et al." + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 7 + }, + { + "para_blocks": [ + { + "bbox": [ + 130, + 116, + 479, + 176 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 116, + 479, + 176 + ], + "spans": [ + { + "bbox": [ + 130, + 116, + 479, + 176 + ], + "type": "text", + "content": "the model is likely to use as shortcuts. The image without shortcut frequency can be generated by reducing the magnitude of frequency components based on the FSPM. Subsequently, the model can be prevented from forming shortcuts by encouraging it to classify correctly and generate consistent CAMs from the original and filtered images." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 130, + 176, + 480, + 213 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 176, + 480, + 213 + ], + "spans": [ + { + "bbox": [ + 130, + 176, + 480, + 213 + ], + "type": "text", + "content": "Let an image " + }, + { + "bbox": [ + 130, + 176, + 480, + 213 + ], + "type": "inline_equation", + "content": "\\mathbf{I}" + }, + { + "bbox": [ + 130, + 176, + 480, + 213 + ], + "type": "text", + "content": " contains classes " + }, + { + "bbox": [ + 130, + 176, + 480, + 213 + ], + "type": "inline_equation", + "content": "c\\in \\mathbb{C}_{\\mathbf{I}}" + }, + { + "bbox": [ + 130, + 176, + 480, + 213 + ], + "type": "text", + "content": ". The filtered image " + }, + { + "bbox": [ + 130, + 176, + 480, + 213 + ], + "type": "inline_equation", + "content": "\\hat{\\mathbf{I}}" + }, + { + "bbox": [ + 130, + 176, + 480, + 213 + ], + "type": "text", + "content": " is synthesized by decreasing the frequency components from " + }, + { + "bbox": [ + 130, + 176, + 480, + 213 + ], + "type": "inline_equation", + "content": "\\mathbf{I}" + }, + { + "bbox": [ + 130, + 176, + 480, + 213 + ], + "type": "text", + "content": " based on the FSPM, making it challenging for the model to classify through shortcuts:" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 230, + 222, + 481, + 247 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 230, + 222, + 481, + 247 + ], + "spans": [ + { + "bbox": [ + 230, + 222, + 481, + 247 + ], + "type": "interline_equation", + "content": "\\hat {\\mathbf {I}} = \\frac {1}{N} \\sum_ {c \\in \\mathbb {C} _ {\\mathbf {I}}} \\mathcal {F} ^ {- 1} \\left(\\mathcal {F} (\\mathbf {I}) \\odot (1 - \\mathbf {F S} (c))\\right) \\tag {9}", + "image_path": "e4ac5cb2db7ab80cb2ec2b9d68ef4b6a25244ead75f0a0fd09e3f90b5f477981.jpg" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 131, + 256, + 479, + 317 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 131, + 256, + 479, + 317 + ], + "spans": [ + { + "bbox": [ + 131, + 256, + 479, + 317 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 131, + 256, + 479, + 317 + ], + "type": "inline_equation", + "content": "N" + }, + { + "bbox": [ + 131, + 256, + 479, + 317 + ], + "type": "text", + "content": " is the number of classes in " + }, + { + "bbox": [ + 131, + 256, + 479, + 317 + ], + "type": "inline_equation", + "content": "\\mathbb{C}_{\\mathbf{I}}" + }, + { + "bbox": [ + 131, + 256, + 479, + 317 + ], + "type": "text", + "content": ", and " + }, + { + "bbox": [ + 131, + 256, + 479, + 317 + ], + "type": "inline_equation", + "content": "\\odot" + }, + { + "bbox": [ + 131, + 256, + 479, + 317 + ], + "type": "text", + "content": " denotes the Hadamard product. The predictions " + }, + { + "bbox": [ + 131, + 256, + 479, + 317 + ], + "type": "inline_equation", + "content": "\\hat{y}_c" + }, + { + "bbox": [ + 131, + 256, + 479, + 317 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 131, + 256, + 479, + 317 + ], + "type": "inline_equation", + "content": "\\hat{y}_p" + }, + { + "bbox": [ + 131, + 256, + 479, + 317 + ], + "type": "text", + "content": " derived from " + }, + { + "bbox": [ + 131, + 256, + 479, + 317 + ], + "type": "inline_equation", + "content": "\\hat{\\mathbf{I}}" + }, + { + "bbox": [ + 131, + 256, + 479, + 317 + ], + "type": "text", + "content": " via the ViT are trained through a multi-label soft margin loss " + }, + { + "bbox": [ + 131, + 256, + 479, + 317 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_{fss-cls}" + }, + { + "bbox": [ + 131, + 256, + 479, + 317 + ], + "type": "text", + "content": ". Additionally, the CAMs " + }, + { + "bbox": [ + 131, + 256, + 479, + 317 + ], + "type": "inline_equation", + "content": "\\hat{\\mathbf{M}}" + }, + { + "bbox": [ + 131, + 256, + 479, + 317 + ], + "type": "text", + "content": " generated from " + }, + { + "bbox": [ + 131, + 256, + 479, + 317 + ], + "type": "inline_equation", + "content": "\\hat{\\mathbf{I}}" + }, + { + "bbox": [ + 131, + 256, + 479, + 317 + ], + "type": "text", + "content": " and the CAMs " + }, + { + "bbox": [ + 131, + 256, + 479, + 317 + ], + "type": "inline_equation", + "content": "\\mathbf{M}" + }, + { + "bbox": [ + 131, + 256, + 479, + 317 + ], + "type": "text", + "content": " derived from the original image " + }, + { + "bbox": [ + 131, + 256, + 479, + 317 + ], + "type": "inline_equation", + "content": "\\mathbf{I}" + }, + { + "bbox": [ + 131, + 256, + 479, + 317 + ], + "type": "text", + "content": " are supervised to minimize their difference:" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 261, + 319, + 480, + 333 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 261, + 319, + 480, + 333 + ], + "spans": [ + { + "bbox": [ + 261, + 319, + 480, + 333 + ], + "type": "interline_equation", + "content": "\\mathcal {L} _ {f s s - c a m s} = | \\mathbf {M} - \\hat {\\mathbf {M}} | _ {1} \\tag {10}", + "image_path": "ed4fc26fee9cfd289a6f1ae283c9847195e2d575404646b4917618b3519d96a8.jpg" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 130, + 339, + 480, + 421 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 339, + 480, + 421 + ], + "spans": [ + { + "bbox": [ + 130, + 339, + 480, + 421 + ], + "type": "text", + "content": "Assuming the model has learned to use a frequency component " + }, + { + "bbox": [ + 130, + 339, + 480, + 421 + ], + "type": "inline_equation", + "content": "f" + }, + { + "bbox": [ + 130, + 339, + 480, + 421 + ], + "type": "text", + "content": " as a shortcut for classifying class " + }, + { + "bbox": [ + 130, + 339, + 480, + 421 + ], + "type": "inline_equation", + "content": "c" + }, + { + "bbox": [ + 130, + 339, + 480, + 421 + ], + "type": "text", + "content": ", the model cannot use this shortcut " + }, + { + "bbox": [ + 130, + 339, + 480, + 421 + ], + "type": "inline_equation", + "content": "f" + }, + { + "bbox": [ + 130, + 339, + 480, + 421 + ], + "type": "text", + "content": " for classification when " + }, + { + "bbox": [ + 130, + 339, + 480, + 421 + ], + "type": "inline_equation", + "content": "f" + }, + { + "bbox": [ + 130, + 339, + 480, + 421 + ], + "type": "text", + "content": " is reduced in " + }, + { + "bbox": [ + 130, + 339, + 480, + 421 + ], + "type": "inline_equation", + "content": "\\hat{\\mathbf{I}}" + }, + { + "bbox": [ + 130, + 339, + 480, + 421 + ], + "type": "text", + "content": ". In the absence of the frequency shortcut " + }, + { + "bbox": [ + 130, + 339, + 480, + 421 + ], + "type": "inline_equation", + "content": "f" + }, + { + "bbox": [ + 130, + 339, + 480, + 421 + ], + "type": "text", + "content": ", the model performs classification based on the semantic understanding of class " + }, + { + "bbox": [ + 130, + 339, + 480, + 421 + ], + "type": "inline_equation", + "content": "c" + }, + { + "bbox": [ + 130, + 339, + 480, + 421 + ], + "type": "text", + "content": ", resulting in activation within areas related to the object. Simultaneously, by reducing the difference between the two CAMs " + }, + { + "bbox": [ + 130, + 339, + 480, + 421 + ], + "type": "inline_equation", + "content": "\\mathbf{M}" + }, + { + "bbox": [ + 130, + 339, + 480, + 421 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 130, + 339, + 480, + 421 + ], + "type": "inline_equation", + "content": "\\hat{\\mathbf{M}}" + }, + { + "bbox": [ + 130, + 339, + 480, + 421 + ], + "type": "text", + "content": ", false positives caused by shortcuts can be resolved." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 146, + 423, + 412, + 434 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 146, + 423, + 412, + 434 + ], + "spans": [ + { + "bbox": [ + 146, + 423, + 412, + 434 + ], + "type": "text", + "content": "The FSS is ultimately supervised through the following loss:" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 246, + 446, + 480, + 457 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 246, + 446, + 480, + 457 + ], + "spans": [ + { + "bbox": [ + 246, + 446, + 480, + 457 + ], + "type": "interline_equation", + "content": "\\mathcal {L} _ {f s s} = \\mathcal {L} _ {f s s - c l s} + \\lambda_ {2} \\mathcal {L} _ {f s s - c a m s} \\tag {11}", + "image_path": "942bcfd04de29542b93bd04602ff45320b46278726a465f3f122165b9037dddd.jpg" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 131, + 468, + 397, + 479 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 131, + 468, + 397, + 479 + ], + "spans": [ + { + "bbox": [ + 131, + 468, + 397, + 479 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 131, + 468, + 397, + 479 + ], + "type": "inline_equation", + "content": "\\lambda_{2}" + }, + { + "bbox": [ + 131, + 468, + 397, + 479 + ], + "type": "text", + "content": " is a hyperparameter to balance between the losses." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 146, + 479, + 417, + 491 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 146, + 479, + 417, + 491 + ], + "spans": [ + { + "bbox": [ + 146, + 479, + 417, + 491 + ], + "type": "text", + "content": "The final loss of the proposed overall framework is as follows:" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 250, + 502, + 480, + 515 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 250, + 502, + 480, + 515 + ], + "spans": [ + { + "bbox": [ + 250, + 502, + 480, + 515 + ], + "type": "interline_equation", + "content": "\\mathcal {L} _ {\\text {t o t a l}} = \\mathcal {L} _ {\\text {c l s}} + \\mathcal {L} _ {\\text {m p c}} + \\mathcal {L} _ {\\text {f s s}} \\tag {12}", + "image_path": "f57120088dcacf3fd8485f836a2c95b4fb700058113d9e84a2f48b63806bdaaa.jpg" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 131, + 533, + 230, + 548 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 131, + 533, + 230, + 548 + ], + "spans": [ + { + "bbox": [ + 131, + 533, + 230, + 548 + ], + "type": "text", + "content": "4 Experiments" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 131, + 560, + 273, + 572 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 131, + 560, + 273, + 572 + ], + "spans": [ + { + "bbox": [ + 131, + 560, + 273, + 572 + ], + "type": "text", + "content": "4.1 Experimental Settings" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 130, + 581, + 480, + 665 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 581, + 480, + 665 + ], + "spans": [ + { + "bbox": [ + 130, + 581, + 480, + 665 + ], + "type": "text", + "content": "Datasets Our proposed method was evaluated using the PASCAL VOC 2012 and MS-COCO 2014 datasets, widely utilized for Weakly Supervised Semantic Segmentation (WSSS). The PASCAL VOC 2012 includes 20 classes of foreground objects along with a background class, and consists of three subsets (train, val, test) containing 10582, 1449, and 1456 images, respectively. Meanwhile, the MS-COCO 2014 is made up of 80 foreground object classes and a background class and is composed of two subsets (82k train set and 40k val set)." + } + ] + } + ], + "index": 15 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 212, + 91, + 448, + 102 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 212, + 91, + 448, + 102 + ], + "spans": [ + { + "bbox": [ + 212, + 91, + 448, + 102 + ], + "type": "text", + "content": "Phase Concentration and Shortcut Suppression for WSSS" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 474, + 92, + 480, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 474, + 92, + 480, + 100 + ], + "spans": [ + { + "bbox": [ + 474, + 92, + 480, + 100 + ], + "type": "text", + "content": "9" + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 8 + }, + { + "para_blocks": [ + { + "type": "table", + "bbox": [ + 189, + 157, + 430, + 216 + ], + "blocks": [ + { + "bbox": [ + 130, + 114, + 482, + 148 + ], + "lines": [ + { + "bbox": [ + 130, + 114, + 482, + 148 + ], + "spans": [ + { + "bbox": [ + 130, + 114, + 482, + 148 + ], + "type": "text", + "content": "Table 1: Ablation study of components on the PASCAL VOC 2012 train set. P: Precision, R: Recall. Values from M are measured under similar R for a fair comparison of P. The best result is represented in Bold." + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 189, + 157, + 430, + 216 + ], + "lines": [ + { + "bbox": [ + 189, + 157, + 430, + 216 + ], + "spans": [ + { + "bbox": [ + 189, + 157, + 430, + 216 + ], + "type": "table", + "html": "
LclsLfssLmpc(x) M(y) Mrf
P(%)R(%)mIoU(%)P(%)R(%)mIoU(%)
Baseline✓53.279.645.374.580.963.2
(a)✓✓71.379.859.476.582.165.4
(b)✓✓67.180.655.978.882.267.2
(c)✓✓✓76.280.763.979.584.869.5
", + "image_path": "12fcda77fd3ed751eecd40a15ee33d3744580a68da5bceb8f5620a7015ec5a14.jpg" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "table_body" + } + ], + "index": 3 + }, + { + "bbox": [ + 130, + 236, + 481, + 344 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 236, + 481, + 344 + ], + "spans": [ + { + "bbox": [ + 130, + 236, + 481, + 344 + ], + "type": "text", + "content": "Evaluation Metric Following prior research [44,63], the mean Intersection over Union (mIoU) metric is used to assess the quality of generated CAMs and the performance of semantic segmentation models trained on CAMs-based pseudolabels. To measure the quality of CAMs, the mIoU with the most optimal threshold is employed by default. Typically, CAMs are evaluated on the train set, whereas the performance of the semantic segmentation model is assessed on the val set. While ground-truth pixel-level labels are accessible for both datasets on the val set and are analyzed locally, the evaluation for the PASCAL VOC 2012 test set is carried out via the official website." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 130, + 344, + 482, + 536 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 344, + 482, + 536 + ], + "spans": [ + { + "bbox": [ + 130, + 344, + 482, + 536 + ], + "type": "text", + "content": "Implementation Details Our framework utilizes DeiT-S pre-trained on ImageNet [50] as the classification backbone for a fair comparison with ViT-based WSSS methods [44, 63]. The classification network is trained for 60 epochs with a batch size of 64 and an initial learning rate of 5e-4, using the Adam optimizer. We adopted the same data augmentation technique as in MCTformer [63], but with different image resize scales. We crop images to a size of " + }, + { + "bbox": [ + 130, + 344, + 482, + 536 + ], + "type": "inline_equation", + "content": "224 \\times 224" + }, + { + "bbox": [ + 130, + 344, + 482, + 536 + ], + "type": "text", + "content": " for equal comparison. In the MPC module, " + }, + { + "bbox": [ + 130, + 344, + 482, + 536 + ], + "type": "inline_equation", + "content": "r = 0.3" + }, + { + "bbox": [ + 130, + 344, + 482, + 536 + ], + "type": "text", + "content": " was used as the mixing factor, with MPC applied at ViT layer " + }, + { + "bbox": [ + 130, + 344, + 482, + 536 + ], + "type": "inline_equation", + "content": "\\ell = 8" + }, + { + "bbox": [ + 130, + 344, + 482, + 536 + ], + "type": "text", + "content": ". To measure the influence of each frequency for FSPM, 100 images were used per class for all classes. Additionally, the FSPM was updated every three epochs through the FIM process. Min-max normalization was utilized in FIM to normalize " + }, + { + "bbox": [ + 130, + 344, + 482, + 536 + ], + "type": "inline_equation", + "content": "\\mathbf{P}" + }, + { + "bbox": [ + 130, + 344, + 482, + 536 + ], + "type": "text", + "content": ", with the highest measured loss per class serving as the normalization maximum. For balancing each loss relative to " + }, + { + "bbox": [ + 130, + 344, + 482, + 536 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_{cls}" + }, + { + "bbox": [ + 130, + 344, + 482, + 536 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 130, + 344, + 482, + 536 + ], + "type": "inline_equation", + "content": "\\lambda_{1}" + }, + { + "bbox": [ + 130, + 344, + 482, + 536 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 130, + 344, + 482, + 536 + ], + "type": "inline_equation", + "content": "\\lambda_{2}" + }, + { + "bbox": [ + 130, + 344, + 482, + 536 + ], + "type": "text", + "content": " were both set to 2. For the semantic segmentation model training on the PASCAL VOC 2012 dataset, Deeplab-V1 with ResNet38 was employed, while Deeplab-V2 with ResNet101 was used for the MS COCO 2014 dataset. Additional training details are in Supplementary Materials." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 131, + 552, + 246, + 562 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 131, + 552, + 246, + 562 + ], + "spans": [ + { + "bbox": [ + 131, + 552, + 246, + 562 + ], + "type": "text", + "content": "4.2 Ablation Studies" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 130, + 570, + 482, + 666 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 570, + 482, + 666 + ], + "spans": [ + { + "bbox": [ + 130, + 570, + 482, + 666 + ], + "type": "text", + "content": "Component Analysis To demonstrate the significance of each module proposed in our framework, we ablated each component as shown in Tab. 1. Observing the performance of refined CAMs in Tab. 1-y, introducing the Frequency Shortcut Suppression (FSS) and Magnitude-mixing-based Phase Concentration (MPC) led to the gain of " + }, + { + "bbox": [ + 130, + 570, + 482, + 666 + ], + "type": "inline_equation", + "content": "2.2\\% \\mathrm{p}" + }, + { + "bbox": [ + 130, + 570, + 482, + 666 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 130, + 570, + 482, + 666 + ], + "type": "inline_equation", + "content": "4.0\\% \\mathrm{p}" + }, + { + "bbox": [ + 130, + 570, + 482, + 666 + ], + "type": "text", + "content": " respectively over baseline. Our method achieved " + }, + { + "bbox": [ + 130, + 570, + 482, + 666 + ], + "type": "inline_equation", + "content": "69.5\\%" + }, + { + "bbox": [ + 130, + 570, + 482, + 666 + ], + "type": "text", + "content": " mIoU by adopting both components. FSS still resulted in a gain of " + }, + { + "bbox": [ + 130, + 570, + 482, + 666 + ], + "type": "inline_equation", + "content": "2.3\\% \\mathrm{p}" + }, + { + "bbox": [ + 130, + 570, + 482, + 666 + ], + "type": "text", + "content": " in Tab. 1-c compared to Tab. 1-b, implies that FSS and MPC offer distinct advantages." + } + ] + } + ], + "index": 7 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "spans": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "type": "text", + "content": "10" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 166, + 91, + 227, + 101 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 166, + 91, + 227, + 101 + ], + "spans": [ + { + "bbox": [ + 166, + 91, + 227, + 101 + ], + "type": "text", + "content": "H. Kwon et al." + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 9 + }, + { + "para_blocks": [ + { + "type": "table", + "bbox": [ + 230, + 147, + 386, + 179 + ], + "blocks": [ + { + "bbox": [ + 131, + 114, + 480, + 137 + ], + "lines": [ + { + "bbox": [ + 131, + 114, + 480, + 137 + ], + "spans": [ + { + "bbox": [ + 131, + 114, + 480, + 137 + ], + "type": "text", + "content": "Table 2: Ablation study of ViT layer index that MPC performed. Peak performance is observed at layer 8, which is represented in Bold." + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 230, + 147, + 386, + 179 + ], + "lines": [ + { + "bbox": [ + 230, + 147, + 386, + 179 + ], + "spans": [ + { + "bbox": [ + 230, + 147, + 386, + 179 + ], + "type": "table", + "html": "
earlylayermiddlelayerlatelayer
l34781011
mIoU(%)68.068.168.669.567.866.4
", + "image_path": "472d285ea044979b436ca3a7ce31a7c9f6c5fb96d1ca5f1304ac8f1e68c50196.jpg" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "table_body" + } + ], + "index": 3 + }, + { + "type": "image", + "bbox": [ + 135, + 194, + 194, + 270 + ], + "blocks": [ + { + "bbox": [ + 135, + 194, + 194, + 270 + ], + "lines": [ + { + "bbox": [ + 135, + 194, + 194, + 270 + ], + "spans": [ + { + "bbox": [ + 135, + 194, + 194, + 270 + ], + "type": "image", + "image_path": "c52d65ecf433beb16c76c62194ab85df54e55dc62e0a42bcf36bf900b83b12b2.jpg" + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 147, + 271, + 181, + 280 + ], + "lines": [ + { + "bbox": [ + 147, + 271, + 181, + 280 + ], + "spans": [ + { + "bbox": [ + 147, + 271, + 181, + 280 + ], + "type": "text", + "content": "(a) Image" + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "image_caption" + } + ], + "index": 4 + }, + { + "type": "image", + "bbox": [ + 194, + 195, + 249, + 270 + ], + "blocks": [ + { + "bbox": [ + 194, + 195, + 249, + 270 + ], + "lines": [ + { + "bbox": [ + 194, + 195, + 249, + 270 + ], + "spans": [ + { + "bbox": [ + 194, + 195, + 249, + 270 + ], + "type": "image", + "image_path": "ea24099a9a5b5aaa78c2359f0e2f578d7b7038092b8bf68feb47d21df69e93f5.jpg" + } + ] + } + ], + "index": 6, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 209, + 271, + 233, + 279 + ], + "lines": [ + { + "bbox": [ + 209, + 271, + 233, + 279 + ], + "spans": [ + { + "bbox": [ + 209, + 271, + 233, + 279 + ], + "type": "text", + "content": "(b) GT" + } + ] + } + ], + "index": 7, + "angle": 0, + "type": "image_caption" + } + ], + "index": 6 + }, + { + "type": "image", + "bbox": [ + 250, + 195, + 306, + 270 + ], + "blocks": [ + { + "bbox": [ + 250, + 195, + 306, + 270 + ], + "lines": [ + { + "bbox": [ + 250, + 195, + 306, + 270 + ], + "spans": [ + { + "bbox": [ + 250, + 195, + 306, + 270 + ], + "type": "image", + "image_path": "5a97f56eeb671efab321c91047875e6a1ae3b2671eb72fe28032b4c42429cd9a.jpg" + } + ] + } + ], + "index": 8, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 258, + 271, + 299, + 279 + ], + "lines": [ + { + "bbox": [ + 258, + 271, + 299, + 279 + ], + "spans": [ + { + "bbox": [ + 258, + 271, + 299, + 279 + ], + "type": "text", + "content": "(c) Baseline" + } + ] + } + ], + "index": 9, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 131, + 289, + 480, + 322 + ], + "lines": [ + { + "bbox": [ + 131, + 289, + 480, + 322 + ], + "spans": [ + { + "bbox": [ + 131, + 289, + 480, + 322 + ], + "type": "text", + "content": "Fig. 4: Comparison of refined CAMs " + }, + { + "bbox": [ + 131, + 289, + 480, + 322 + ], + "type": "inline_equation", + "content": "\\mathbf{M}_{ref}" + }, + { + "bbox": [ + 131, + 289, + 480, + 322 + ], + "type": "text", + "content": " between baseline and components in ours. (a) Image, (b) Ground Truth, (c) Baseline, (d) only FSS, (e) only MPC, (f) ours. The yellow and red boxes imply effect of FSS and MPC, respectively." + } + ] + } + ], + "index": 16, + "angle": 0, + "type": "image_caption" + } + ], + "index": 8 + }, + { + "type": "image", + "bbox": [ + 307, + 195, + 363, + 270 + ], + "blocks": [ + { + "bbox": [ + 307, + 195, + 363, + 270 + ], + "lines": [ + { + "bbox": [ + 307, + 195, + 363, + 270 + ], + "spans": [ + { + "bbox": [ + 307, + 195, + 363, + 270 + ], + "type": "image", + "image_path": "1c957814bf72dc6c15c910f5d2235b07d0778e491f45283960ba3d267c0180d8.jpg" + } + ] + } + ], + "index": 10, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 322, + 271, + 349, + 279 + ], + "lines": [ + { + "bbox": [ + 322, + 271, + 349, + 279 + ], + "spans": [ + { + "bbox": [ + 322, + 271, + 349, + 279 + ], + "type": "text", + "content": "(d) FSS" + } + ] + } + ], + "index": 11, + "angle": 0, + "type": "image_caption" + } + ], + "index": 10 + }, + { + "type": "image", + "bbox": [ + 364, + 195, + 420, + 270 + ], + "blocks": [ + { + "bbox": [ + 364, + 195, + 420, + 270 + ], + "lines": [ + { + "bbox": [ + 364, + 195, + 420, + 270 + ], + "spans": [ + { + "bbox": [ + 364, + 195, + 420, + 270 + ], + "type": "image", + "image_path": "cc0e49ab30b8e12bc3346679c4dec22a023e2b10017cfdcad830cfec465f31d5.jpg" + } + ] + } + ], + "index": 12, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 378, + 271, + 407, + 279 + ], + "lines": [ + { + "bbox": [ + 378, + 271, + 407, + 279 + ], + "spans": [ + { + "bbox": [ + 378, + 271, + 407, + 279 + ], + "type": "text", + "content": "(e) MPC" + } + ] + } + ], + "index": 13, + "angle": 0, + "type": "image_caption" + } + ], + "index": 12 + }, + { + "type": "image", + "bbox": [ + 421, + 195, + 477, + 270 + ], + "blocks": [ + { + "bbox": [ + 421, + 195, + 477, + 270 + ], + "lines": [ + { + "bbox": [ + 421, + 195, + 477, + 270 + ], + "spans": [ + { + "bbox": [ + 421, + 195, + 477, + 270 + ], + "type": "image", + "image_path": "97dbb2141e54829bfcee6eeed5fa8a80a5ccdda2625b4a38501729d331a8b7c6.jpg" + } + ] + } + ], + "index": 14, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 426, + 271, + 473, + 279 + ], + "lines": [ + { + "bbox": [ + 426, + 271, + 473, + 279 + ], + "spans": [ + { + "bbox": [ + 426, + 271, + 473, + 279 + ], + "type": "text", + "content": "(f) FSS+MPC" + } + ] + } + ], + "index": 15, + "angle": 0, + "type": "image_caption" + } + ], + "index": 14 + }, + { + "bbox": [ + 130, + 350, + 480, + 434 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 350, + 480, + 434 + ], + "spans": [ + { + "bbox": [ + 130, + 350, + 480, + 434 + ], + "type": "text", + "content": "To understand the benefits of FSS in CAMs, we examined the results of ViT CAMs that are not refined by the attention-map in Tab. 1-x. To conduct a comparative analysis considering the Precision-Recall tradeoff, thresholds corresponding to similar Recall levels were used. Low precision of baseline " + }, + { + "bbox": [ + 130, + 350, + 480, + 434 + ], + "type": "inline_equation", + "content": "\\mathbf{M}" + }, + { + "bbox": [ + 130, + 350, + 480, + 434 + ], + "type": "text", + "content": " indicates a significant number of over-activation. Through FSS, we observed a remarkable improvement in precision " + }, + { + "bbox": [ + 130, + 350, + 480, + 434 + ], + "type": "inline_equation", + "content": "(18.1\\%)" + }, + { + "bbox": [ + 130, + 350, + 480, + 434 + ], + "type": "text", + "content": " and mIoU " + }, + { + "bbox": [ + 130, + 350, + 480, + 434 + ], + "type": "inline_equation", + "content": "(14.1\\%)" + }, + { + "bbox": [ + 130, + 350, + 480, + 434 + ], + "type": "text", + "content": " compared to the baseline, confirming that false positives were successfully resolved (Tab. 1-b)." + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 130, + 436, + 480, + 519 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 436, + 480, + 519 + ], + "spans": [ + { + "bbox": [ + 130, + 436, + 480, + 519 + ], + "type": "text", + "content": "Ablation Studies of MPC The qualitative comparison between the refined CAMs of baseline and MPC in Fig. 4-e demonstrates the efficacy of the MPC by clearly delineating the boundary of the object. Meanwhile, we investigate the impact of the ratio " + }, + { + "bbox": [ + 130, + 436, + 480, + 519 + ], + "type": "inline_equation", + "content": "r" + }, + { + "bbox": [ + 130, + 436, + 480, + 519 + ], + "type": "text", + "content": " in MPC, which is used to preserve the magnitude information of the anchor feature. In experiments conducted within a search space below 0.5, there was a minimum gain of " + }, + { + "bbox": [ + 130, + 436, + 480, + 519 + ], + "type": "inline_equation", + "content": "2.3\\% p" + }, + { + "bbox": [ + 130, + 436, + 480, + 519 + ], + "type": "text", + "content": " when compared to applying FSS alone, with the best performance observed at " + }, + { + "bbox": [ + 130, + 436, + 480, + 519 + ], + "type": "inline_equation", + "content": "r = 0.3" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 130, + 521, + 481, + 665 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 521, + 481, + 665 + ], + "spans": [ + { + "bbox": [ + 130, + 521, + 481, + 665 + ], + "type": "text", + "content": "The influence of the layer index " + }, + { + "bbox": [ + 130, + 521, + 481, + 665 + ], + "type": "inline_equation", + "content": "\\ell" + }, + { + "bbox": [ + 130, + 521, + 481, + 665 + ], + "type": "text", + "content": " where the MPC is performed was analyzed by conducting experiments on early, middle, and late layers in Tab. 2. The performance was highest when " + }, + { + "bbox": [ + 130, + 521, + 481, + 665 + ], + "type": "inline_equation", + "content": "\\ell" + }, + { + "bbox": [ + 130, + 521, + 481, + 665 + ], + "type": "text", + "content": " was set to 8, with a noticeable decline observed in layers beyond this point. This tendency supports the knowledge that later layers of the ViT play a crucial role in capturing high-level semantic information [9]. Introducing changes to the magnitude while preserving the phase information at these later stages could disrupt the semantic analysis of the model. This disruption may stem from the difference between the coupled information used in previous layers and the re-coupled information with the magnitude alteration. This result shows that MPC applied before when ViT features highly condense to high-level semantics can encourage the model to capture boundary information well in the phase spectrum." + } + ] + } + ], + "index": 19 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 212, + 91, + 448, + 102 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 212, + 91, + 448, + 102 + ], + "spans": [ + { + "bbox": [ + 212, + 91, + 448, + 102 + ], + "type": "text", + "content": "Phase Concentration and Shortcut Suppression for WSSS" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 470, + 91, + 479, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 470, + 91, + 479, + 100 + ], + "spans": [ + { + "bbox": [ + 470, + 91, + 479, + 100 + ], + "type": "text", + "content": "11" + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 10 + }, + { + "para_blocks": [ + { + "type": "table", + "bbox": [ + 135, + 191, + 482, + 299 + ], + "blocks": [ + { + "bbox": [ + 130, + 114, + 482, + 180 + ], + "lines": [ + { + "bbox": [ + 130, + 114, + 482, + 180 + ], + "spans": [ + { + "bbox": [ + 130, + 114, + 482, + 180 + ], + "type": "text", + "content": "Table 3: (a) True/False Positive Rate (TPR/FPR) of classification where the high-frequency components following FSPM are retained in images. Metrics are measured under the early training stage of the Baseline. **Bold** indicates the shortcut occurred, and **Underline** represents that model is trying to make a shortcut. (b) Comparative analysis of the Precision of CAMs M between baseline and only with FSS. Precision is evaluated under a similar Recall considering the Precision-Recall trade-off." + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 135, + 191, + 482, + 299 + ], + "lines": [ + { + "bbox": [ + 135, + 191, + 482, + 299 + ], + "spans": [ + { + "bbox": [ + 135, + 191, + 482, + 299 + ], + "type": "table", + "html": "
bkgaerobikebirdboatbottlebuscarcatchaircow
(a)BaselineTPR(%)-58.9636.4388.4753.3668.4650.8379.2668.3555.4817.87
FPR(%)-00.2300.4115.7800.0533.7400.0001.4200.5307.0400.41
(b)BaselineP(%)91.939.224.817.318.958.176.542.880.944.959.7
w/ FSSP(%)92.552.942.259.233.371.187.562.994.143.290.7
tabledoghorsembkpersonplantsheepsofatraintvtotal
(a)BaselineTPR(%)04.7661.2479.4060.2675.4741.8843.6145.4592.5157.23-
FPR(%)00.0001.9700.6400.1605.5600.1500.3800.6300.6800.13-
(b)BaselineP(%)68.666.050.468.658.039.956.479.039.237.053.2
w/ FSSP(%)89.687.579.778.582.764.792.581.254.556.671.3
", + "image_path": "6d0508f924ec085b4dc5a3916bfb59fc9354a639041c7a6d76bbc0366c0d43f7.jpg" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "table_body" + } + ], + "index": 3 + }, + { + "type": "image", + "bbox": [ + 149, + 313, + 195, + 416 + ], + "blocks": [ + { + "bbox": [ + 149, + 313, + 195, + 416 + ], + "lines": [ + { + "bbox": [ + 149, + 313, + 195, + 416 + ], + "spans": [ + { + "bbox": [ + 149, + 313, + 195, + 416 + ], + "type": "image", + "image_path": "789d710b9f35af13364ccd140b16cb4fc65804fd456cb4ec390943ddb952c958.jpg" + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "image_body" + } + ], + "index": 4 + }, + { + "type": "image", + "bbox": [ + 196, + 314, + 228, + 416 + ], + "blocks": [ + { + "bbox": [ + 196, + 314, + 228, + 416 + ], + "lines": [ + { + "bbox": [ + 196, + 314, + 228, + 416 + ], + "spans": [ + { + "bbox": [ + 196, + 314, + 228, + 416 + ], + "type": "image", + "image_path": "1349e719e53c3d6dc26311c39edc6a225670db56e187656a8c438f96f007088e.jpg" + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 210, + 417, + 272, + 424 + ], + "lines": [ + { + "bbox": [ + 210, + 417, + 272, + 424 + ], + "spans": [ + { + "bbox": [ + 210, + 417, + 272, + 424 + ], + "type": "text", + "content": "PASCAL VOC 2012" + } + ] + } + ], + "index": 11, + "angle": 0, + "type": "image_caption" + } + ], + "index": 5 + }, + { + "type": "image", + "bbox": [ + 229, + 314, + 277, + 416 + ], + "blocks": [ + { + "bbox": [ + 229, + 314, + 277, + 416 + ], + "lines": [ + { + "bbox": [ + 229, + 314, + 277, + 416 + ], + "spans": [ + { + "bbox": [ + 229, + 314, + 277, + 416 + ], + "type": "image", + "image_path": "213cb39012f09625bbe9c2d815cece94a905d22498d187bd838ea4be36e96bb8.jpg" + } + ] + } + ], + "index": 6, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 130, + 434, + 482, + 456 + ], + "lines": [ + { + "bbox": [ + 130, + 434, + 482, + 456 + ], + "spans": [ + { + "bbox": [ + 130, + 434, + 482, + 456 + ], + "type": "text", + "content": "Fig.5: Semantic segmentation results on PASCAL VOC 2012 (left) and MS COCO 2014 (right) datasets. From top to bottom: Image, Ours, GT." + } + ] + } + ], + "index": 13, + "angle": 0, + "type": "image_caption" + } + ], + "index": 6 + }, + { + "type": "image", + "bbox": [ + 277, + 314, + 328, + 416 + ], + "blocks": [ + { + "bbox": [ + 277, + 314, + 328, + 416 + ], + "lines": [ + { + "bbox": [ + 277, + 314, + 328, + 416 + ], + "spans": [ + { + "bbox": [ + 277, + 314, + 328, + 416 + ], + "type": "image", + "image_path": "dfd509dcb24401bf676857c83fdfbb838c1f9ae7fc4652ab7cd67b17be60e035.jpg" + } + ] + } + ], + "index": 7, + "angle": 0, + "type": "image_body" + } + ], + "index": 7 + }, + { + "type": "image", + "bbox": [ + 332, + 314, + 358, + 416 + ], + "blocks": [ + { + "bbox": [ + 332, + 314, + 358, + 416 + ], + "lines": [ + { + "bbox": [ + 332, + 314, + 358, + 416 + ], + "spans": [ + { + "bbox": [ + 332, + 314, + 358, + 416 + ], + "type": "image", + "image_path": "4da045b8515eccf717596a93ae711c5d1c7c0d32d5b7a63bd841aa6d01495b9c.jpg" + } + ] + } + ], + "index": 8, + "angle": 0, + "type": "image_body" + } + ], + "index": 8 + }, + { + "type": "image", + "bbox": [ + 358, + 314, + 397, + 416 + ], + "blocks": [ + { + "bbox": [ + 358, + 314, + 397, + 416 + ], + "lines": [ + { + "bbox": [ + 358, + 314, + 397, + 416 + ], + "spans": [ + { + "bbox": [ + 358, + 314, + 397, + 416 + ], + "type": "image", + "image_path": "1aaad2b1fe3c7f8d416a631027eef01bd95d1a7b5dc1dc036c96f8978e44a7c2.jpg" + } + ] + } + ], + "index": 9, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 373, + 417, + 424, + 424 + ], + "lines": [ + { + "bbox": [ + 373, + 417, + 424, + 424 + ], + "spans": [ + { + "bbox": [ + 373, + 417, + 424, + 424 + ], + "type": "text", + "content": "MS COCO 2014" + } + ] + } + ], + "index": 12, + "angle": 0, + "type": "image_caption" + } + ], + "index": 9 + }, + { + "type": "image", + "bbox": [ + 396, + 314, + 463, + 416 + ], + "blocks": [ + { + "bbox": [ + 396, + 314, + 463, + 416 + ], + "lines": [ + { + "bbox": [ + 396, + 314, + 463, + 416 + ], + "spans": [ + { + "bbox": [ + 396, + 314, + 463, + 416 + ], + "type": "image", + "image_path": "21735237b7d0dbff15ba83edc10a170472f82434af30023ff1050f256d09c081.jpg" + } + ] + } + ], + "index": 10, + "angle": 0, + "type": "image_body" + } + ], + "index": 10 + }, + { + "bbox": [ + 130, + 483, + 482, + 567 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 483, + 482, + 567 + ], + "spans": [ + { + "bbox": [ + 130, + 483, + 482, + 567 + ], + "type": "text", + "content": "Frequency Shortcut Analysis and Result of Suppression In Tab. 3, we investigated whether shortcut learning occurs during the model training process in early steps and analyzed the correlation with the improvement in CAMs precision facilitated by the FSS. To conduct this analysis, classification influence according to frequency component for each class was first measured using Eq. 7. Based on this measurement, the classification behavior of the model was analyzed when only some high frequencies of high-influence components were retained." + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 130, + 570, + 482, + 666 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 570, + 482, + 666 + ], + "spans": [ + { + "bbox": [ + 130, + 570, + 482, + 666 + ], + "type": "text", + "content": "A high TPR indicates that class " + }, + { + "bbox": [ + 130, + 570, + 482, + 666 + ], + "type": "inline_equation", + "content": "c" + }, + { + "bbox": [ + 130, + 570, + 482, + 666 + ], + "type": "text", + "content": " possesses the principal frequency components utilized in the experiment (i.e., bird will be correctly classified as bird only with the high-influence frequency of bird). This suggests that the model has identified a straightforward method for classifying class " + }, + { + "bbox": [ + 130, + 570, + 482, + 666 + ], + "type": "inline_equation", + "content": "c" + }, + { + "bbox": [ + 130, + 570, + 482, + 666 + ], + "type": "text", + "content": " and is attempting to find frequency shortcuts. To consider a principal frequency as a shortcut, it should direct images to the respective class regardless of the actual semantic content in the image (i.e. car with frequency shortcut of bird will falsely classify as bird). This is expressed in Tab. 3 through high TPR and FPR, which repre" + } + ] + } + ], + "index": 15 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "spans": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "type": "text", + "content": "12" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 166, + 91, + 227, + 101 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 166, + 91, + 227, + 101 + ], + "spans": [ + { + "bbox": [ + 166, + 91, + 227, + 101 + ], + "type": "text", + "content": "H. Kwon et al." + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 11 + }, + { + "para_blocks": [ + { + "type": "table", + "bbox": [ + 204, + 168, + 413, + 306 + ], + "blocks": [ + { + "bbox": [ + 130, + 114, + 482, + 159 + ], + "lines": [ + { + "bbox": [ + 130, + 114, + 482, + 159 + ], + "spans": [ + { + "bbox": [ + 130, + 114, + 482, + 159 + ], + "type": "text", + "content": "Table 4: Comparison between ours and multi-stage WSSS methods. mIoU(%) is evaluated on PASCAL VOC 2012 train set in both CAMs (seed) and Pseudo-GT (Mask). The backbone and applied post-processing methods (PSA [2], IRN [1]) are listed for a fair comparison. Bold numbers represent the best performance." + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 204, + 168, + 413, + 306 + ], + "lines": [ + { + "bbox": [ + 204, + 168, + 413, + 306 + ], + "spans": [ + { + "bbox": [ + 204, + 168, + 413, + 306 + ], + "type": "table", + "html": "
MethodBackboneSeedPostMask
AdvCAM [28] CVPR'21RN5055.6IRN69.9
OC-CSE [25] ICCV'21WRN3856.0PSA66.9
ECS [53] ICCV'21WRN3856.6PSA67.8
CPN [70] ICCV'21WRN3857.4PSA67.8
AMR [45] AAAI'22RN5056.8PSA69.7
ReCAM [7] CVPR'22RN5054.8IRN70.5
SIPE [6] CVPR'22RN5058.6IRN-
PPC [10] CVPR'22WRN3861.5IRN70.1
AEFT [67] ECCV'22WRN3856.0PSA71.0
ACR [26] CVPR'23WRN3860.3IRN72.3
MCT [63] CVPR'22DeiT-S61.7PSA69.1
USAGE [44] ICCV'23DeiT-S67.7PSA72.8
OursDeiT-S69.5PSA73.2
", + "image_path": "9eecf686b1f57214a57faaa2c4ba2c925cf03a71603d1af0554645b068d416e3.jpg" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "table_body" + } + ], + "index": 3 + }, + { + "bbox": [ + 130, + 338, + 482, + 397 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 338, + 482, + 397 + ], + "spans": [ + { + "bbox": [ + 130, + 338, + 482, + 397 + ], + "type": "text", + "content": "sent the rates in which principal frequencies of images are correctly and falsely classified to the shortcut respective class " + }, + { + "bbox": [ + 130, + 338, + 482, + 397 + ], + "type": "inline_equation", + "content": "c" + }, + { + "bbox": [ + 130, + 338, + 482, + 397 + ], + "type": "text", + "content": ". Examining Tab. 3-a, we can observe that class bird and bottle exhibit high TPR and FPR, indicating the occurrence of shortcuts. Meanwhile, classes such as car, horse, and train show high TPR, suggesting attempts at creating shortcuts." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 130, + 398, + 482, + 529 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 398, + 482, + 529 + ], + "spans": [ + { + "bbox": [ + 130, + 398, + 482, + 529 + ], + "type": "text", + "content": "Tab. 3-b records the class-specific precision of CAMs " + }, + { + "bbox": [ + 130, + 398, + 482, + 529 + ], + "type": "inline_equation", + "content": "\\mathbf{M}" + }, + { + "bbox": [ + 130, + 398, + 482, + 529 + ], + "type": "text", + "content": " for the baseline model and after the introduction of the FSS. To ensure a fair comparison while considering the Precision-Recall trade-off, precision was measured at similar levels of Recall. The introduction of FSS to disrupt shortcut creation resulted in an average increase of " + }, + { + "bbox": [ + 130, + 398, + 482, + 529 + ], + "type": "inline_equation", + "content": "24.0\\%" + }, + { + "bbox": [ + 130, + 398, + 482, + 529 + ], + "type": "text", + "content": " in precision for " + }, + { + "bbox": [ + 130, + 398, + 482, + 529 + ], + "type": "inline_equation", + "content": "\\mathbf{M}" + }, + { + "bbox": [ + 130, + 398, + 482, + 529 + ], + "type": "text", + "content": " across classes where the baseline model had created or attempted to create shortcuts. This increase confirms that the FSS effectively reduces false positives. Notably, for bird, a class where shortcuts had occurred, the precision of CAMs surged by " + }, + { + "bbox": [ + 130, + 398, + 482, + 529 + ], + "type": "inline_equation", + "content": "42.1\\%" + }, + { + "bbox": [ + 130, + 398, + 482, + 529 + ], + "type": "text", + "content": " , showcasing the resolve of over-activation in areas unrelated to the object by FSS. Qualitative comparison of refined CAMs " + }, + { + "bbox": [ + 130, + 398, + 482, + 529 + ], + "type": "inline_equation", + "content": "\\mathbf{M}_{ref}" + }, + { + "bbox": [ + 130, + 398, + 482, + 529 + ], + "type": "text", + "content": " before and after the introduction of FSS in Fig. 4-d further validates the effectiveness of FSS." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 131, + 548, + 340, + 561 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 131, + 548, + 340, + 561 + ], + "spans": [ + { + "bbox": [ + 131, + 548, + 340, + 561 + ], + "type": "text", + "content": "4.3 Comparison with State-of-The-Arts" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 130, + 570, + 482, + 666 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 570, + 482, + 666 + ], + "spans": [ + { + "bbox": [ + 130, + 570, + 482, + 666 + ], + "type": "text", + "content": "To train the semantic segmentation model, the CAMs generated by our method were refined using PSA [2] as performed in previous WSSS works [44, 63, 67]. In Tab. 4, we compared the performance of CAMs (Seed) and pseudo-ground-truth (Mask) on the PASCAL VOC 2012 train set, where our method achieved the best performance in both categories. Utilizing our high-quality pseudo labels for training the semantic segmentation model, we compared the results with other WSSS methods in Tab. 5. Our framework demonstrated superiority by achieving the highest performance on the PASCAL VOC 2012 val and test sets." + } + ] + } + ], + "index": 7 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 212, + 91, + 447, + 102 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 212, + 91, + 447, + 102 + ], + "spans": [ + { + "bbox": [ + 212, + 91, + 447, + 102 + ], + "type": "text", + "content": "Phase Concentration and Shortcut Suppression for WSSS" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 470, + 91, + 481, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 470, + 91, + 481, + 100 + ], + "spans": [ + { + "bbox": [ + 470, + 91, + 481, + 100 + ], + "type": "text", + "content": "13" + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 12 + }, + { + "para_blocks": [ + { + "type": "table", + "bbox": [ + 183, + 168, + 431, + 322 + ], + "blocks": [ + { + "bbox": [ + 132, + 114, + 482, + 159 + ], + "lines": [ + { + "bbox": [ + 132, + 114, + 482, + 159 + ], + "spans": [ + { + "bbox": [ + 132, + 114, + 482, + 159 + ], + "type": "text", + "content": "Table 5: Comparison of semantic segmentation performance with multi-stage WSSS methods. Performance (mIoU) is evaluated in PASCAL VOC 2012 and MS COCO 2014 datasets. CNN-based methods are represented above the horizontal line, and ViT-based methods are depicted below the line. Bold represents the best performance." + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 183, + 168, + 431, + 322 + ], + "lines": [ + { + "bbox": [ + 183, + 168, + 431, + 322 + ], + "spans": [ + { + "bbox": [ + 183, + 168, + 431, + 322 + ], + "type": "table", + "html": "
MethodBackboneVOC valVOC testCOCO val
AdvCAM [28] CVPR'21RN10168.168.0-
OC-CSE [25] ICCV'21WRN3868.468.236.4
ECS [53] ICCV'21WRN3866.667.6-
CPN [70] ICCV'21WRN3867.868.5-
AMR [45] AAAT'22RN10168.869.1-
ReCAM [7] CVPR'22RN10168.568.442.9
SIPE [6] CVPR'22RN10168.869.7-
SIPE [6] CVPR'22WRN38--43.6
AEFT [67] ECCV'22WRN3870.971.744.8
ACR [26] CVPR'23WRN3871.971.945.3
MCT [63] CVPR'22WRN3871.971.642.0
USAGE [44] ICCV'23WRN3871.972.842.7
USAGE [44] ICCV'23RN101--44.3
OursWRN3873.273.0-
OursRN101--45.7
", + "image_path": "755cbf765516c5c987ac0238169adf5a0f5daedead7f03bd3a0a05d0b4aa9e82.jpg" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "table_body" + } + ], + "index": 3 + }, + { + "bbox": [ + 130, + 346, + 482, + 419 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 346, + 482, + 419 + ], + "spans": [ + { + "bbox": [ + 130, + 346, + 482, + 419 + ], + "type": "text", + "content": "While ViT-based WSSS works tend to outperform CNN-based works on the PASCAL VOC 2012 dataset, they showed lower performance on the MS COCO 2014 dataset due to activation overlapping between classes. Nonetheless, our method achieved the SoTA performance on the MS COCO 2014 dataset by effectively reducing false positives. Additional CAMs and semantic segmentation results are in the Supplementary Materials." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 132, + 442, + 220, + 455 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 442, + 220, + 455 + ], + "spans": [ + { + "bbox": [ + 132, + 442, + 220, + 455 + ], + "type": "text", + "content": "5 Conclusion" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 130, + 473, + 482, + 666 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 473, + 482, + 666 + ], + "spans": [ + { + "bbox": [ + 130, + 473, + 482, + 666 + ], + "type": "text", + "content": "Weakly Supervised Semantic Segmentation suffers from CAMs that have imprecise activation around boundary areas and object-unrelated activations. In this paper, we aim to address these challenges by applying a novel perspective in the frequency domain, leveraging the Fourier Transform. Firstly, we propose Magnitude-mixing-based Phase Concentration (MPC), which generates features with identical phase but different magnitude at the intermediate layer, encouraging the model to capture the same semantics from these features. This approach leads the model to pay more attention to shape information in the phase, thereby learning to activate boundaries more clearly. Additionally, we identify the occurrence of frequency shortcuts and their association with the over-activation of CAMs. To resolve this, we introduce Frequency Shortcut Suppression (FSS), which discourages the creation of frequency shortcuts, successfully resolving object-unrelated activations. Our experimental results support the utility of the proposed method, demonstrating its effectiveness in addressing the aforementioned issues in WSSS. Furthermore, we have achieved new state-of-the-art performance on the PASCAL VOC and MS COCO datasets." + } + ] + } + ], + "index": 6 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "spans": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "type": "text", + "content": "14" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 166, + 91, + 227, + 101 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 166, + 91, + 227, + 101 + ], + "spans": [ + { + "bbox": [ + 166, + 91, + 227, + 101 + ], + "type": "text", + "content": "H. Kwon et al." + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 13 + }, + { + "para_blocks": [ + { + "bbox": [ + 130, + 116, + 482, + 189 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 116, + 482, + 189 + ], + "spans": [ + { + "bbox": [ + 130, + 116, + 482, + 189 + ], + "type": "text", + "content": "Acknowledgement This work was supported by the Technology Innovation Program (1415187329,20024355, Development of autonomous driving connectivity technology based on sensor-infrastructure cooperation) funded by the Ministry of Trade, Industry & Energy(MOTIE, Korea) and the National Research Foundation of Korea(NRF) grant funded by the Korea government(MSIT) (NRF2022R1A2B5B03002636)." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 132, + 203, + 197, + 216 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 203, + 197, + 216 + ], + "spans": [ + { + "bbox": [ + 132, + 203, + 197, + 216 + ], + "type": "text", + "content": "References" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 134, + 224, + 480, + 665 + ], + "type": "list", + "angle": 0, + "index": 16, + "blocks": [ + { + "bbox": [ + 138, + 224, + 480, + 257 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 224, + 480, + 257 + ], + "spans": [ + { + "bbox": [ + 138, + 224, + 480, + 257 + ], + "type": "text", + "content": "1. Ahn, J., Cho, S., Kwak, S.: Weakly supervised learning of instance segmentation with inter-pixel relations. In: Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition. pp. 2209-2218 (2019)" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 138, + 258, + 480, + 289 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 258, + 480, + 289 + ], + "spans": [ + { + "bbox": [ + 138, + 258, + 480, + 289 + ], + "type": "text", + "content": "2. Ahn, J., Kwak, S.: Learning pixel-level semantic affinity with image-level supervision for weakly supervised semantic segmentation. In: Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition. pp. 4981-4990 (2018)" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 138, + 289, + 480, + 332 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 289, + 480, + 332 + ], + "spans": [ + { + "bbox": [ + 138, + 289, + 480, + 332 + ], + "type": "text", + "content": "3. Chang, Y.T., Wang, Q., Hung, W.C., Piramuthu, R., Tsai, Y.H., Yang, M.H.: Weakly-supervised semantic segmentation via sub-category exploration. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 8991-9000 (2020)" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 138, + 333, + 480, + 374 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 333, + 480, + 374 + ], + "spans": [ + { + "bbox": [ + 138, + 333, + 480, + 374 + ], + "type": "text", + "content": "4. Chen, G., Peng, P., Ma, L., Li, J., Du, L., Tian, Y.: Amplitude-phase recombination: Rethinking robustness of convolutional neural networks in frequency domain. In: Proceedings of the IEEE/CVF International Conference on Computer Vision (ICCV). pp. 458-467 (October 2021)" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 138, + 376, + 480, + 407 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 376, + 480, + 407 + ], + "spans": [ + { + "bbox": [ + 138, + 376, + 480, + 407 + ], + "type": "text", + "content": "5. Chen, L., Lei, C., Li, R., Li, S., Zhang, Z., Zhang, L.: Fpr: False positive rectification for weakly supervised semantic segmentation. In: Proceedings of the IEEE/CVF International Conference on Computer Vision. pp. 1108-1118 (2023)" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 138, + 407, + 480, + 450 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 407, + 480, + 450 + ], + "spans": [ + { + "bbox": [ + 138, + 407, + 480, + 450 + ], + "type": "text", + "content": "6. Chen, Q., Yang, L., Lai, J.H., Xie, X.: Self-supervised image-specific prototype exploration for weakly supervised semantic segmentation. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 4288-4298 (2022)" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 138, + 450, + 480, + 493 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 450, + 480, + 493 + ], + "spans": [ + { + "bbox": [ + 138, + 450, + 480, + 493 + ], + "type": "text", + "content": "7. Chen, Z., Wang, T., Wu, X., Hua, X.S., Zhang, H., Sun, Q.: Class reactivation maps for weakly-supervised semantic segmentation. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 969-978 (2022)" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 138, + 494, + 480, + 525 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 494, + 480, + 525 + ], + "spans": [ + { + "bbox": [ + 138, + 494, + 480, + 525 + ], + "type": "text", + "content": "8. Dai, J., He, K., Sun, J.: Boxsup: Exploiting bounding boxes to supervise convolutional networks for semantic segmentation. In: Proceedings of the IEEE international conference on computer vision. pp. 1635-1643 (2015)" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 138, + 525, + 480, + 568 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 525, + 480, + 568 + ], + "spans": [ + { + "bbox": [ + 138, + 525, + 480, + 568 + ], + "type": "text", + "content": "9. Dosovitskiy, A., Beyer, L., Kolesnikov, A., Weissenborn, D., Zhai, X., Unterthiner, T., Dehghani, M., Minderer, M., Heigold, G., Gelly, S., et al.: An image is worth 16x16 words: Transformers for image recognition at scale. arXiv preprint arXiv:2010.11929 (2020)" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 134, + 569, + 480, + 601 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 134, + 569, + 480, + 601 + ], + "spans": [ + { + "bbox": [ + 134, + 569, + 480, + 601 + ], + "type": "text", + "content": "10. Du, Y., Fu, Z., Liu, Q., Wang, Y.: Weakly supervised semantic segmentation by pixel-to-prototype contrast. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 4320-4329 (2022)" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 134, + 601, + 480, + 632 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 134, + 601, + 480, + 632 + ], + "spans": [ + { + "bbox": [ + 134, + 601, + 480, + 632 + ], + "type": "text", + "content": "11. Everingham, M., Van Gool, L., Williams, C.K., Winn, J., Zisserman, A.: The pascal visual object classes (voc) challenge. International journal of computer vision 88(2), 303-338 (2010)" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 134, + 632, + 480, + 665 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 134, + 632, + 480, + 665 + ], + "spans": [ + { + "bbox": [ + 134, + 632, + 480, + 665 + ], + "type": "text", + "content": "12. Fan, J., Zhang, Z., Tan, T., Song, C., Xiao, J.: Cian: Cross-image affinity net for weakly supervised semantic segmentation. In: Proceedings of the AAAI Conference on Artificial Intelligence. vol. 34, pp. 10762-10769 (2020)" + } + ] + } + ], + "index": 15 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 212, + 91, + 447, + 102 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 212, + 91, + 447, + 102 + ], + "spans": [ + { + "bbox": [ + 212, + 91, + 447, + 102 + ], + "type": "text", + "content": "Phase Concentration and Shortcut Suppression for WSSS" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 470, + 92, + 481, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 470, + 92, + 481, + 100 + ], + "spans": [ + { + "bbox": [ + 470, + 92, + 481, + 100 + ], + "type": "text", + "content": "15" + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 14 + }, + { + "para_blocks": [ + { + "bbox": [ + 133, + 116, + 481, + 666 + ], + "type": "list", + "angle": 0, + "index": 17, + "blocks": [ + { + "bbox": [ + 133, + 116, + 481, + 160 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 116, + 481, + 160 + ], + "spans": [ + { + "bbox": [ + 133, + 116, + 481, + 160 + ], + "type": "text", + "content": "13. Gao, W., Wan, F., Pan, X., Peng, Z., Tian, Q., Han, Z., Zhou, B., Ye, Q.: Ts-cam: Token semantic coupled attention map for weakly supervised object localization. In: Proceedings of the IEEE/CVF International Conference on Computer Vision. pp. 2886-2895 (2021)" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 133, + 160, + 481, + 192 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 160, + 481, + 192 + ], + "spans": [ + { + "bbox": [ + 133, + 160, + 481, + 192 + ], + "type": "text", + "content": "14. Geirhos, R., Jacobsen, J.H., Michaelis, C., Zemel, R., Brendel, W., Bethge, M., Wichmann, F.A.: Shortcut learning in deep neural networks. Nature Machine Intelligence 2(11), 665-673 (2020)" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 133, + 192, + 481, + 236 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 192, + 481, + 236 + ], + "spans": [ + { + "bbox": [ + 133, + 192, + 481, + 236 + ], + "type": "text", + "content": "15. Guo, J., Wang, N., Qi, L., Shi, Y.: Aloft: A lightweight mlp-like architecture with dynamic low-frequency transform for domain generalization. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 24132-24141 (2023)" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 133, + 236, + 481, + 257 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 236, + 481, + 257 + ], + "spans": [ + { + "bbox": [ + 133, + 236, + 481, + 257 + ], + "type": "text", + "content": "16. Hansen, B.C., Hess, R.F.: Structural sparseness and spatial phase alignment in natural scenes. JOSA A 24(7), 1873-1885 (2007)" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 133, + 257, + 481, + 300 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 257, + 481, + 300 + ], + "spans": [ + { + "bbox": [ + 133, + 257, + 481, + 300 + ], + "type": "text", + "content": "17. Huang, H., Kang, H., Liu, S., Salvado, O., Rakotoarivelo, T., Wang, D., Liu, T.: Paddles: Phase-amplitude spectrum disentangled early stopping for learning with noisy labels. In: Proceedings of the IEEE/CVF International Conference on Computer Vision. pp. 16719-16730 (2023)" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 133, + 300, + 481, + 343 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 300, + 481, + 343 + ], + "spans": [ + { + "bbox": [ + 133, + 300, + 481, + 343 + ], + "type": "text", + "content": "18. Huang, H., Xie, S., Lin, L., Tong, R., Chen, Y.W., Li, Y., Wang, H., Huang, Y., Zheng, Y.: Semicvt: Semi-supervised convolutional vision transformer for semantic segmentation. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 11340-11349 (2023)" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 133, + 343, + 481, + 374 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 343, + 481, + 374 + ], + "spans": [ + { + "bbox": [ + 133, + 343, + 481, + 374 + ], + "type": "text", + "content": "19. Huang, J., Guan, D., Xiao, A., Lu, S.: Rda: Robust domain adaptation via fourier adversarial attacking. In: Proceedings of the IEEE/CVF international conference on computer vision. pp. 8988-8999 (2021)" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 133, + 374, + 481, + 407 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 374, + 481, + 407 + ], + "spans": [ + { + "bbox": [ + 133, + 374, + 481, + 407 + ], + "type": "text", + "content": "20. Huang, J., Liu, Y., Zhao, F., Yan, K., Zhang, J., Huang, Y., Zhou, M., Xiong, Z.: Deep fourier-based exposure correction network with spatial-frequency interaction. In: European Conference on Computer Vision. pp. 163-180. Springer (2022)" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 133, + 407, + 481, + 439 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 407, + 481, + 439 + ], + "spans": [ + { + "bbox": [ + 133, + 407, + 481, + 439 + ], + "type": "text", + "content": "21. Jiang, P.T., Han, L.H., Hou, Q., Cheng, M.M., Wei, Y.: Online attention accumulation for weakly supervised semantic segmentation. IEEE Transactions on Pattern Analysis and Machine Intelligence 44(10), 7062-7077 (2021)" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 133, + 439, + 481, + 482 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 439, + 481, + 482 + ], + "spans": [ + { + "bbox": [ + 133, + 439, + 481, + 482 + ], + "type": "text", + "content": "22. Jiang, P.T., Yang, Y., Hou, Q., Wei, Y.: L2g: A simple local-to-global knowledge transfer framework for weakly supervised semantic segmentation. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 16886-16896 (2022)" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 133, + 482, + 481, + 514 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 482, + 481, + 514 + ], + "spans": [ + { + "bbox": [ + 133, + 482, + 481, + 514 + ], + "type": "text", + "content": "23. Jo, S., Yu, I.J., Kim, K.: Mars: Model-agnostic biased object removal without additional supervision for weakly-supervised semantic segmentation. arXiv preprint arXiv:2304.09913 (2023)" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 133, + 514, + 481, + 546 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 514, + 481, + 546 + ], + "spans": [ + { + "bbox": [ + 133, + 514, + 481, + 546 + ], + "type": "text", + "content": "24. Kim, G., Kim, J., Lee, J.S.: Exploring adversarial robustness of vision transformers in the spectral perspective. In: Proceedings of the IEEE/CVF Winter Conference on Applications of Computer Vision. pp. 3976-3985 (2024)" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 133, + 546, + 481, + 590 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 546, + 481, + 590 + ], + "spans": [ + { + "bbox": [ + 133, + 546, + 481, + 590 + ], + "type": "text", + "content": "25. Kweon, H., Yoon, S.H., Kim, H., Park, D., Yoon, K.J.: Unlocking the potential of ordinary classifier: Class-specific adversarial erasing framework for weakly supervised semantic segmentation. In: Proceedings of the IEEE/CVF International Conference on Computer Vision. pp. 6994-7003 (2021)" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 133, + 590, + 481, + 632 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 590, + 481, + 632 + ], + "spans": [ + { + "bbox": [ + 133, + 590, + 481, + 632 + ], + "type": "text", + "content": "26. Kweon, H., Yoon, S.H., Yoon, K.J.: Weakly supervised semantic segmentation via adversarial learning of classifier and reconstructor. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 11329-11339 (2023)" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 133, + 632, + 481, + 666 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 632, + 481, + 666 + ], + "spans": [ + { + "bbox": [ + 133, + 632, + 481, + 666 + ], + "type": "text", + "content": "27. Lee, J., Choi, J., Mok, J., Yoon, S.: Reducing information bottleneck for weakly supervised semantic segmentation. Advances in Neural Information Processing Systems 34, 27408-27421 (2021)" + } + ] + } + ], + "index": 16 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "spans": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "type": "text", + "content": "16" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 166, + 91, + 227, + 101 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 166, + 91, + 227, + 101 + ], + "spans": [ + { + "bbox": [ + 166, + 91, + 227, + 101 + ], + "type": "text", + "content": "H. Kwon et al." + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 15 + }, + { + "para_blocks": [ + { + "bbox": [ + 130, + 116, + 482, + 665 + ], + "type": "list", + "angle": 0, + "index": 18, + "blocks": [ + { + "bbox": [ + 130, + 116, + 482, + 150 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 116, + 482, + 150 + ], + "spans": [ + { + "bbox": [ + 130, + 116, + 482, + 150 + ], + "type": "text", + "content": "28. Lee, J., Kim, E., Yoon, S.: Anti-adversarily manipulated attributions for weakly and semi-supervised semantic segmentation. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 4071-4080 (2021)" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 130, + 150, + 482, + 184 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 150, + 482, + 184 + ], + "spans": [ + { + "bbox": [ + 130, + 150, + 482, + 184 + ], + "type": "text", + "content": "29. Lee, J., Oh, S.J., Yun, S., Choe, J., Kim, E., Yoon, S.: Weakly supervised semantic segmentation using out-of-distribution data. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 16897-16906 (2022)" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 130, + 184, + 482, + 217 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 184, + 482, + 217 + ], + "spans": [ + { + "bbox": [ + 130, + 184, + 482, + 217 + ], + "type": "text", + "content": "30. Lee, J., Yi, J., Shin, C., Yoon, S.: Bbam: Bounding box attribution map for weakly supervised semantic and instance segmentation. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 2643-2652 (2021)" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 130, + 218, + 482, + 261 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 218, + 482, + 261 + ], + "spans": [ + { + "bbox": [ + 130, + 218, + 482, + 261 + ], + "type": "text", + "content": "31. Lee, S., Bae, J., Kim, H.Y.: Decompose, adjust, compose: Effective normalization by playing with frequency for domain generalization. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 11776-11785 (2023)" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 130, + 261, + 482, + 296 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 261, + 482, + 296 + ], + "spans": [ + { + "bbox": [ + 130, + 261, + 482, + 296 + ], + "type": "text", + "content": "32. Li, J., Fan, J., Zhang, Z.: Towards noiseless object contours for weakly supervised semantic segmentation. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 16856-16865 (2022)" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 130, + 296, + 482, + 329 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 296, + 482, + 329 + ], + "spans": [ + { + "bbox": [ + 130, + 296, + 482, + 329 + ], + "type": "text", + "content": "33. Li, K., Wu, Z., Peng, K.C., Ernst, J., Fu, Y.: Tell me where to look: Guided attention inference network. In: Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition. pp. 9215-9223 (2018)" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 130, + 329, + 482, + 352 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 329, + 482, + 352 + ], + "spans": [ + { + "bbox": [ + 130, + 329, + 482, + 352 + ], + "type": "text", + "content": "34. Li, X., Zhou, T., Li, J., Zhou, Y., Zhang, Z.: Group-wise semantic mining for weakly supervised semantic segmentation. arXiv preprint arXiv:2012.05007 (2020)" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 130, + 352, + 482, + 385 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 352, + 482, + 385 + ], + "spans": [ + { + "bbox": [ + 130, + 352, + 482, + 385 + ], + "type": "text", + "content": "35. Li, Y., Duan, Y., Kuang, Z., Chen, Y., Zhang, W., Li, X.: Uncertainty estimation via response scaling for pseudo-mask noise mitigation in weakly-supervised semantic segmentation. arXiv preprint arXiv:2112.07431 (2021)" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 130, + 385, + 482, + 419 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 385, + 482, + 419 + ], + "spans": [ + { + "bbox": [ + 130, + 385, + 482, + 419 + ], + "type": "text", + "content": "36. Li, Y., Kuang, Z., Liu, L., Chen, Y., Zhang, W.: Pseudo-mask matters in weakly-supervised semantic segmentation. In: Proceedings of the IEEE/CVF International Conference on Computer Vision. pp. 6964-6973 (2021)" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 130, + 419, + 482, + 453 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 419, + 482, + 453 + ], + "spans": [ + { + "bbox": [ + 130, + 419, + 482, + 453 + ], + "type": "text", + "content": "37. Lin, D., Dai, J., Jia, J., He, K., Sun, J.: Scribblesup: Scribble-supervised convolutional networks for semantic segmentation. In: Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition. pp. 3159-3167 (2016)" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 130, + 453, + 482, + 486 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 453, + 482, + 486 + ], + "spans": [ + { + "bbox": [ + 130, + 453, + 482, + 486 + ], + "type": "text", + "content": "38. Lin, T.Y., Maire, M., Belongie, S., Hays, J., Perona, P., Ramanan, D., Dollar, P., Zitnick, C.L.: Microsoft coco: Common objects in context. In: European conference on computer vision. pp. 740-755. Springer (2014)" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 130, + 486, + 482, + 531 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 486, + 482, + 531 + ], + "spans": [ + { + "bbox": [ + 130, + 486, + 482, + 531 + ], + "type": "text", + "content": "39. Lin, Y., Chen, M., Wang, W., Wu, B., Li, K., Lin, B., Liu, H., He, X.: Clip is also an efficient segmenter: A text-driven approach for weakly supervised semantic segmentation. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 15305-15314 (2023)" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 130, + 531, + 482, + 564 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 531, + 482, + 564 + ], + "spans": [ + { + "bbox": [ + 130, + 531, + 482, + 564 + ], + "type": "text", + "content": "40. Luo, X., Wei, L., Wen, L., Yang, J., Xie, L., Xu, Z., Tian, Q.: Rectifying the short-cut learning of background for few-shot learning. Advances in Neural Information Processing Systems 34, 13073-13085 (2021)" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 130, + 564, + 482, + 598 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 564, + 482, + 598 + ], + "spans": [ + { + "bbox": [ + 130, + 564, + 482, + 598 + ], + "type": "text", + "content": "41. Ma, C., Zhao, L., Chen, Y., Guo, L., Zhang, T., Hu, X., Shen, D., Jiang, X., Liu, T.: Rectify vit shortcut learning by visual saliency. IEEE Transactions on Neural Networks and Learning Systems (2023)" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 130, + 598, + 482, + 621 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 598, + 482, + 621 + ], + "spans": [ + { + "bbox": [ + 130, + 598, + 482, + 621 + ], + "type": "text", + "content": "42. Oppenheim, A.V., Lim, J.S.: The importance of phase in signals. Proceedings of the IEEE 69(5), 529-541 (1981)" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 130, + 621, + 482, + 665 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 621, + 482, + 665 + ], + "spans": [ + { + "bbox": [ + 130, + 621, + 482, + 665 + ], + "type": "text", + "content": "43. Papandreou, G., Chen, L.C., Murphy, K.P., Yuille, A.L.: Weakly-and semi-supervised learning of a deep convolutional network for semantic image segmentation. In: Proceedings of the IEEE international conference on computer vision. pp. 1742-1750 (2015)" + } + ] + } + ], + "index": 17 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 213, + 91, + 447, + 102 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 213, + 91, + 447, + 102 + ], + "spans": [ + { + "bbox": [ + 213, + 91, + 447, + 102 + ], + "type": "text", + "content": "Phase Concentration and Shortcut Suppression for WSSS" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 470, + 92, + 481, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 470, + 92, + 481, + 100 + ], + "spans": [ + { + "bbox": [ + 470, + 92, + 481, + 100 + ], + "type": "text", + "content": "17" + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 16 + }, + { + "para_blocks": [ + { + "bbox": [ + 132, + 116, + 481, + 665 + ], + "type": "list", + "angle": 0, + "index": 18, + "blocks": [ + { + "bbox": [ + 133, + 116, + 481, + 149 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 116, + 481, + 149 + ], + "spans": [ + { + "bbox": [ + 133, + 116, + 481, + 149 + ], + "type": "text", + "content": "44. Peng, Z., Wang, G., Xie, L., Jiang, D., Shen, W., Tian, Q.: Usage: A unified seed area generation paradigm for weakly supervised semantic segmentation. arXiv preprint arXiv:2303.07806 (2023)" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 133, + 150, + 481, + 183 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 150, + 481, + 183 + ], + "spans": [ + { + "bbox": [ + 133, + 150, + 481, + 183 + ], + "type": "text", + "content": "45. Qin, J., Wu, J., Xiao, X., Li, L., Wang, X.: Activation modulation and recalibration scheme for weakly supervised semantic segmentation. arXiv preprint arXiv:2112.08996 (2021)" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 132, + 183, + 481, + 215 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 183, + 481, + 215 + ], + "spans": [ + { + "bbox": [ + 132, + 183, + 481, + 215 + ], + "type": "text", + "content": "46. Robinson, J., Sun, L., Yu, K., Batmanghelich, K., Jegelka, S., Sra, S.: Can contrastive learning avoid shortcut solutions? Advances in neural information processing systems 34, 4974-4986 (2021)" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 132, + 216, + 481, + 258 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 216, + 481, + 258 + ], + "spans": [ + { + "bbox": [ + 132, + 216, + 481, + 258 + ], + "type": "text", + "content": "47. Rossetti, S., Zappia, D., Sanzari, M., Schaerf, M., Pirri, F.: Max pooling with vision transformers reconciles class and shape in weakly supervised semantic segmentation. In: European Conference on Computer Vision. pp. 446-463. Springer (2022)" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 132, + 259, + 481, + 303 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 259, + 481, + 303 + ], + "spans": [ + { + "bbox": [ + 132, + 259, + 481, + 303 + ], + "type": "text", + "content": "48. Ru, L., Zhan, Y., Yu, B., Du, B.: Learning affinity from attention: End-to-end weakly-supervised semantic segmentation with transformers. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 16846-16855 (2022)" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 132, + 304, + 481, + 336 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 304, + 481, + 336 + ], + "spans": [ + { + "bbox": [ + 132, + 304, + 481, + 336 + ], + "type": "text", + "content": "49. Ru, L., Zheng, H., Zhan, Y., Du, B.: Token contrast for weakly-supervised semantic segmentation. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 3093-3102 (2023)" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 132, + 336, + 481, + 369 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 336, + 481, + 369 + ], + "spans": [ + { + "bbox": [ + 132, + 336, + 481, + 369 + ], + "type": "text", + "content": "50. Russakovsky, O., Deng, J., Su, H., Krause, J., Satheesh, S., Ma, S., Huang, Z., Karpathy, A., Khosla, A., Bernstein, M., et al.: Imagenet large scale visual recognition challenge. International journal of computer vision 115(3), 211-252 (2015)" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 132, + 369, + 481, + 391 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 369, + 481, + 391 + ], + "spans": [ + { + "bbox": [ + 132, + 369, + 481, + 391 + ], + "type": "text", + "content": "51. Su, Y., Sun, R., Lin, G., Wu, Q.: Context decoupling augmentation for weakly supervised semantic segmentation. arXiv preprint arXiv:2103.01795 (2021)" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 132, + 392, + 481, + 413 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 392, + 481, + 413 + ], + "spans": [ + { + "bbox": [ + 132, + 392, + 481, + 413 + ], + "type": "text", + "content": "52. Sun, G., Wang, W., Dai, J., Van Gool, L.: Mining cross-image semantics for weakly supervised semantic segmentation. arXiv preprint arXiv:2007.01947 (2020)" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 132, + 413, + 481, + 456 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 413, + 481, + 456 + ], + "spans": [ + { + "bbox": [ + 132, + 413, + 481, + 456 + ], + "type": "text", + "content": "53. Sun, K., Shi, H., Zhang, Z., Huang, Y.: ECS-net: Improving weakly supervised semantic segmentation by using connections between class activation maps. In: Proceedings of the IEEE/CVF International Conference on Computer Vision. pp. 7283-7292 (2021)" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 132, + 457, + 481, + 479 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 457, + 481, + 479 + ], + "spans": [ + { + "bbox": [ + 132, + 457, + 481, + 479 + ], + "type": "text", + "content": "54. Sundararajan, D.: The discrete Fourier transform: theory, algorithms and applications. World Scientific (2001)" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 132, + 479, + 481, + 511 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 479, + 481, + 511 + ], + "spans": [ + { + "bbox": [ + 132, + 479, + 481, + 511 + ], + "type": "text", + "content": "55. Vaswani, A., Shazeer, N., Parmar, N., Uszkoreit, J., Jones, L., Gomez, A.N., Kaiser, L., Polosukhin, I.: Attention is all you need. Advances in neural information processing systems 30 (2017)" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 132, + 512, + 481, + 544 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 512, + 481, + 544 + ], + "spans": [ + { + "bbox": [ + 132, + 512, + 481, + 544 + ], + "type": "text", + "content": "56. Vernaza, P., Chandraker, M.: Learning random-walk label propagation for weakly-supervised semantic segmentation. In: Proceedings of the IEEE conference on computer vision and pattern recognition. pp. 7158-7166 (2017)" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 132, + 545, + 481, + 578 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 545, + 481, + 578 + ], + "spans": [ + { + "bbox": [ + 132, + 545, + 481, + 578 + ], + "type": "text", + "content": "57. Wang, C., Xu, R., Xu, S., Meng, W., Zhang, X.: Treating pseudo-labels generation as image matting for weakly supervised semantic segmentation. In: Proceedings of the IEEE/CVF International Conference on Computer Vision. pp. 755-765 (2023)" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 132, + 578, + 481, + 621 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 578, + 481, + 621 + ], + "spans": [ + { + "bbox": [ + 132, + 578, + 481, + 621 + ], + "type": "text", + "content": "58. Wang, S., Veldhuis, R., Brune, C., Strisciuglio, N.: What do neural networks learn in image classification? a frequency shortcut perspective. In: Proceedings of the IEEE/CVF International Conference on Computer Vision (ICCV). pp. 1433-1442 (October 2023)" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 132, + 621, + 481, + 665 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 621, + 481, + 665 + ], + "spans": [ + { + "bbox": [ + 132, + 621, + 481, + 665 + ], + "type": "text", + "content": "59. Wang, Y., Zhang, J., Kan, M., Shan, S., Chen, X.: Self-supervised equivariant attention mechanism for weakly supervised semantic segmentation. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 12275-12284 (2020)" + } + ] + } + ], + "index": 17 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "spans": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "type": "text", + "content": "18" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 166, + 91, + 227, + 101 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 166, + 91, + 227, + 101 + ], + "spans": [ + { + "bbox": [ + 166, + 91, + 227, + 101 + ], + "type": "text", + "content": "H. Kwon et al." + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 17 + }, + { + "para_blocks": [ + { + "bbox": [ + 130, + 116, + 482, + 643 + ], + "type": "list", + "angle": 0, + "index": 16, + "blocks": [ + { + "bbox": [ + 130, + 116, + 482, + 160 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 116, + 482, + 160 + ], + "spans": [ + { + "bbox": [ + 130, + 116, + 482, + 160 + ], + "type": "text", + "content": "60. Wu, T., Huang, J., Gao, G., Wei, X., Wei, X., Luo, X., Liu, C.H.: Embedded discriminative attention mechanism for weakly supervised semantic segmentation. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 16765-16774 (2021)" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 130, + 161, + 481, + 194 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 161, + 481, + 194 + ], + "spans": [ + { + "bbox": [ + 130, + 161, + 481, + 194 + ], + "type": "text", + "content": "61. Xie, J., Hou, X., Ye, K., Shen, L.: Clims: Cross language image matching for weakly supervised semantic segmentation. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 4483-4492 (2022)" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 132, + 194, + 481, + 237 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 194, + 481, + 237 + ], + "spans": [ + { + "bbox": [ + 132, + 194, + 481, + 237 + ], + "type": "text", + "content": "62. Xie, J., Xiang, J., Chen, J., Hou, X., Zhao, X., Shen, L.: C2am: Contrastive learning of class-agnostic activation map for weakly supervised object localization and semantic segmentation. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 989–998 (2022)" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 132, + 238, + 481, + 281 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 238, + 481, + 281 + ], + "spans": [ + { + "bbox": [ + 132, + 238, + 481, + 281 + ], + "type": "text", + "content": "63. Xu, L., Ouyang, W., Bennamoun, M., Boussaid, F., Xu, D.: Multi-class token transformer for weakly supervised semantic segmentation. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 4310-4319 (2022)" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 132, + 282, + 481, + 326 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 282, + 481, + 326 + ], + "spans": [ + { + "bbox": [ + 132, + 282, + 481, + 326 + ], + "type": "text", + "content": "64. Xu, L., Ouyang, W., Bennamoun, M., Boussaid, F., Xu, D.: Learning multi-modal class-specific tokens for weakly supervised dense object localization. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 19596-19605 (2023)" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 132, + 326, + 481, + 358 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 326, + 481, + 358 + ], + "spans": [ + { + "bbox": [ + 132, + 326, + 481, + 358 + ], + "type": "text", + "content": "65. Xu, Q., Zhang, R., Zhang, Y., Wang, Y., Tian, Q.: A fourier-based framework for domain generalization. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 14383-14392 (2021)" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 132, + 358, + 481, + 392 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 358, + 481, + 392 + ], + "spans": [ + { + "bbox": [ + 132, + 358, + 481, + 392 + ], + "type": "text", + "content": "66. Yang, Y., Soatto, S.: Fda: Fourier domain adaptation for semantic segmentation. In: Proceedings of the IEEE/CVF conference on computer vision and pattern recognition. pp. 4085-4095 (2020)" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 132, + 392, + 481, + 445 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 392, + 481, + 445 + ], + "spans": [ + { + "bbox": [ + 132, + 392, + 481, + 445 + ], + "type": "text", + "content": "67. Yoon, S.H., Kweon, H., Cho, J., Kim, S., Yoon, K.J.: Adversarial erasing framework via triplet with gated pyramid pooling layer for weakly supervised semantic segmentation. In: Computer Vision-ECCV 2022: 17th European Conference, Tel Aviv, Israel, October 23-27, 2022, Proceedings, Part XXIX. pp. 326-344. Springer Nature Switzerland Cham (2022)" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 132, + 447, + 481, + 479 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 447, + 481, + 479 + ], + "spans": [ + { + "bbox": [ + 132, + 447, + 481, + 479 + ], + "type": "text", + "content": "68. Yoon, S.H., Kweon, H., Jeong, J., Kim, H., Kim, S., Yoon, K.J.: Exploring pixel-level self-supervision for weakly supervised semantic segmentation. arXiv preprint arXiv:2112.05351 (2021)" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 132, + 479, + 481, + 512 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 479, + 481, + 512 + ], + "spans": [ + { + "bbox": [ + 132, + 479, + 481, + 512 + ], + "type": "text", + "content": "69. Yoon, S.H., Kwon, H., Kim, H., Yoon, K.J.: Class tokens infusion for weakly supervised semantic segmentation. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 3595-3605 (2024)" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 132, + 512, + 481, + 544 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 512, + 481, + 544 + ], + "spans": [ + { + "bbox": [ + 132, + 512, + 481, + 544 + ], + "type": "text", + "content": "70. Zhang, F., Gu, C., Zhang, C., Dai, Y.: Complementary patch for weakly supervised semantic segmentation. In: Proceedings of the IEEE/CVF International Conference on Computer Vision. pp. 7242-7251 (2021)" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 132, + 545, + 481, + 578 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 545, + 481, + 578 + ], + "spans": [ + { + "bbox": [ + 132, + 545, + 481, + 578 + ], + "type": "text", + "content": "71. Zhang, X., Wei, Y., Feng, J., Yang, Y., Huang, T.S.: Adversarial complementary learning for weakly supervised object localization. In: Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition. pp. 1325-1334 (2018)" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 132, + 578, + 481, + 611 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 578, + 481, + 611 + ], + "spans": [ + { + "bbox": [ + 132, + 578, + 481, + 611 + ], + "type": "text", + "content": "72. Zhou, B., Khosla, A., Lapedriza, A., Oliva, A., Torralba, A.: Learning deep features for discriminative localization. In: Proceedings of the IEEE conference on computer vision and pattern recognition. pp. 2921-2929 (2016)" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 132, + 611, + 481, + 643 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 611, + 481, + 643 + ], + "spans": [ + { + "bbox": [ + 132, + 611, + 481, + 643 + ], + "type": "text", + "content": "73. Zhou, T., Zhang, M., Zhao, F., Li, J.: Regional semantic contrast and aggregation for weakly supervised semantic segmentation. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 4299-4309 (2022)" + } + ] + } + ], + "index": 15 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 213, + 91, + 447, + 102 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 213, + 91, + 447, + 102 + ], + "spans": [ + { + "bbox": [ + 213, + 91, + 447, + 102 + ], + "type": "text", + "content": "Phase Concentration and Shortcut Suppression for WSSS" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 470, + 92, + 480, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 470, + 92, + 480, + 100 + ], + "spans": [ + { + "bbox": [ + 470, + 92, + 480, + 100 + ], + "type": "text", + "content": "19" + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 18 + } + ], + "_backend": "vlm", + "_version_name": "2.6.4" +} \ No newline at end of file diff --git a/2024/Photon Inhibition for Energy-Efficient Single-Photon Imaging/03673fae-a9a6-4975-8cf4-3d5bc33fd48c_content_list.json b/2024/Photon Inhibition for Energy-Efficient Single-Photon Imaging/03673fae-a9a6-4975-8cf4-3d5bc33fd48c_content_list.json new file mode 100644 index 0000000000000000000000000000000000000000..9bf975253e5ec75dab853ca341690d49cec5a197 --- /dev/null +++ b/2024/Photon Inhibition for Energy-Efficient Single-Photon Imaging/03673fae-a9a6-4975-8cf4-3d5bc33fd48c_content_list.json @@ -0,0 +1,1677 @@ +[ + { + "type": "text", + "text": "Photon Inhibition for Energy-Efficient Single-Photon Imaging", + "text_level": 1, + "bbox": [ + 282, + 140, + 718, + 186 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Lucas J. Koerner1, Shantanu Gupta2, Atul Ingle3, and Mohit Gupta2", + "bbox": [ + 225, + 210, + 774, + 228 + ], + "page_idx": 0 + }, + { + "type": "list", + "sub_type": "text", + "list_items": [ + "$^{1}$ Department of Electrical and Computer Engineering, University of St. Thomas, St. Paul, MN, 55105, USA koerner.lucas@stthomas.edu", + "$^{2}$ Department of Computer Sciences, University of Wisconsin-Madison, Madison, WI, 53706, USA {sgupta,mohitg}@cs.wisc.edu", + "$^{3}$ Department of Computer Science, Portland State University, Portland, OR, 97201, USA \ningle2@pdx.edu" + ], + "bbox": [ + 215, + 238, + 782, + 321 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Abstract. Single-photon cameras (SPCs) are emerging as sensors of choice for various challenging imaging applications. One class of SPCs based on the single-photon avalanche diode (SPAD) detects individual photons using an avalanche process; the raw photon data can then be processed to extract scene information under extremely low light, high dynamic range, and rapid motion. Yet, single-photon sensitivity in SPADs comes at a cost — each photon detection consumes more energy than that of a CMOS camera. This avalanche power significantly limits sensor resolution and could restrict widespread adoption of SPAD-based SPCs. We propose a computational-imaging approach called photon inhibition to address this challenge. Photon inhibition strategically allocates detections in space and time based on downstream inference task goals and resource constraints. We develop lightweight, on-sensor computational inhibition policies that use past photon data to disable SPAD pixels in real-time, to select the most informative future photons. As case studies, we design policies tailored for image reconstruction and edge detection, and demonstrate, both via simulations and real SPC captured data, considerable reduction in photon detections (over $90\\%$ of photons) while maintaining task performance metrics. Our work raises the question of \"which photons should be detected?\", and paves the way for future energy-efficient single-photon imaging. Source code for our experiments is available at https://wisionlab.com/project/inhibition.", + "bbox": [ + 261, + 359, + 738, + 664 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Keywords: single-photon imaging $\\cdot$ SPADs $\\cdot$ energy-efficient vision", + "bbox": [ + 261, + 676, + 715, + 691 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "1 Introduction", + "text_level": 1, + "bbox": [ + 215, + 717, + 375, + 734 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "From vacuum tube-based single-photon detectors invented in the early 20th century [25], to 3D depth sensing LiDAR cameras found in today's smartphones [28], single-photon camera (SPC) technology has come a long way in terms of pixel resolution and commercial availability for a variety of applications. Thanks to CMOS-compatible single-photon avalanche diode (SPAD) arrays, SPCs are being increasingly used not only in niche fields such as scientific imaging and biomedical", + "bbox": [ + 212, + 750, + 787, + 840 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/569a1e0f3cf9c3fd9630c153c427ef6c15f59798b072886a2421b02be122130c.jpg", + "image_caption": [ + "(a)", + "(c)", + "(b)" + ], + "image_footnote": [], + "bbox": [ + 215, + 155, + 444, + 268 + ], + "page_idx": 1 + }, + { + "type": "image", + "img_path": "images/085285b6037fa00abba45c15cafcfeaa09725b66253ac9ed2bc26f4bea002c45.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 452, + 164, + 625, + 257 + ], + "page_idx": 1 + }, + { + "type": "image", + "img_path": "images/fbe062150553a0bf3eca8f4afb5fa72dee4aaa1edcf93b24649106b68171e51e.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 640, + 164, + 779, + 243 + ], + "page_idx": 1 + }, + { + "type": "image", + "img_path": "images/83a5d838b33a448d72f6c75b24a21f7f97476f5d7b1264981d0fa51783af66ff.jpg", + "image_caption": [ + "(d)", + "Fig. 1: Photon inhibition for resource-efficient passive SPAD imaging. (a) Unlike conventional CMOS cameras (CIS), the energy consumption in SPAD cameras increases with scene brightness, severely limiting the applicability of high resolution SPAD cameras in resource-constrained applications. (b) We expand the conventional imaging pipeline to incorporate \"inhibition\" that electronically enables or disables individual pixels to limit bandwidth and power consumption. Our method relies on lightweight mathematical operations called \"inhibition policies\" that update the inhibition patterns based on the history of photon detections. Inhibition policies can be optimized for image SNR or for downstream vision tasks.(c,d) Object detection, a high-level vision task, is successful even with a large fraction of photons inhibited." + ], + "image_footnote": [], + "bbox": [ + 222, + 280, + 450, + 404 + ], + "page_idx": 1 + }, + { + "type": "image", + "img_path": "images/dc6dea98ffdf6a07b3aec81f67e04c50b0a005874003987e8430f359261cc3a5.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 452, + 279, + 632, + 392 + ], + "page_idx": 1 + }, + { + "type": "image", + "img_path": "images/afe810e526862fbe033a7bc4a3153b6205ec818e7b3f981da0caea6b64b89de1.jpg", + "image_caption": [ + "YOLOv8 Object Detection" + ], + "image_footnote": [], + "bbox": [ + 638, + 304, + 785, + 391 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "microscopy [49], but more widely for consumer photography [37]. SPAD-based SPCs have recently been fabricated into kilo-to-megapixel format arrays that are now commercially available [28, 37]. The extreme sensitivity and high speed can benefit passive low-light computer vision tasks, particularly in the presence of rapid scene or camera motion [35], enable wide dynamic-range imaging [14, 31] and photon-starved active imaging applications such as 3D imaging (LiDAR) [21] and fluorescence microscopy [49].", + "bbox": [ + 212, + 582, + 787, + 688 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "SPAD camera pixels detect individual photons with extremely high frame rates by exploiting avalanche multiplication. On one hand, being able to detect individual photons opens up new possibilities and capabilities for computer vision systems. On the other hand, this also presents a unique challenge: Every photon-induced avalanche comes with a non-negligible energy cost, which is a challenge exclusive to SPAD-based cameras. This flux-dependent photon detection power is a significant fraction of total power consumption in today's SPAD cameras and impedes further increases in their spatial resolution [24,37,41,42,47]. For example, extrapolating the avalanche power of a recent SPAD sensor [47] to $\\sim 10$ 's of megapixels format predicts a power consumption of several watts in", + "bbox": [ + 212, + 689, + 787, + 840 + ], + "page_idx": 1 + }, + { + "type": "page_number", + "text": "2", + "bbox": [ + 217, + 114, + 228, + 126 + ], + "page_idx": 1 + }, + { + "type": "header", + "text": "L. Koerner et al.", + "bbox": [ + 271, + 114, + 383, + 126 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "bright light $^{4}$ (as illustrated in Fig. 1(a)), far exceeding the power consumption of modern CMOS image sensors (CIS) of around $300\\mathrm{mW}$ [45].", + "bbox": [ + 212, + 145, + 782, + 176 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "We address this problem by being selective about which photons are detected on the SPAD sensor while maintaining good performance on various computer vision tasks. To reduce avalanche power consumption and decouple it from photon flux, we propose a computational imaging technique called photon inhibition where individual SPAD pixels are disabled adaptively based on previous photon detections in their spatio-temporal neighborhoods. Electrically disabling pixels prevents photon detections to inhibit any avalanche power or subsequent processing. We implement lightweight on-sensor computations called inhibition policies (Fig. 1(b)) to determine, in real-time and at single-pixel and single-frame granularity, which SPAD pixels to enable or disable. Our method is inspired by retinal pre-processing of the human visual system where retinal neurons aggregate photon information over small spatio-temporal neighborhoods to cause neighboring retinal cells to become less sensitive to incident light [12, 16].5", + "bbox": [ + 212, + 178, + 787, + 375 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Given that SPADs introduce a new challenge of flux-dependent power consumption, we establish, from first principles, novel energy-aware imaging performance metrics for resource-constrained single-photon imaging. Based on these metrics, we design families of inhibition policies that distribute photon detections in space and time based on imaging / vision task goals and energy consumption constraints. A critical consideration in the design of inhibition policies stems from the observation that these policies are meant to control (enable / disable) the sensor (Fig. 1(b)). Therefore, it is essential for these policies to be extremely lightweight since these need to be implemented on sensor with very limited compute and memory resources. Furthermore, inhibition policies must execute at ultra-low latency to keep up with high-speed photon detections (reaching up to $100\\mathrm{kHz}$ ). Fortunately, since the raw data output from a SPAD-based SPC consists of binary-valued image frames, SPCs are naturally suited to real-time calculations at the image sensor plane under tight compute and memory budgets. The proposed inhibition policies are lightweight, requiring only simple arithmetic and Boolean operations computed over local spatio-temporal neighborhoods, and thus amenable to in-pixel implementation [2, 7].", + "bbox": [ + 212, + 376, + 787, + 632 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "In simulations and real experimental data, we show that our inhibition policies allocate photon detections to sensor pixels in a way that reduces detection energy for a given accuracy level of various vision tasks. Our results show consequential energy savings when compared to a conventional capture scheme for tasks of (i) image reconstruction: $42\\%$ fewer photon detections at equal image quality; (ii) edge detection: an edge sensitive inhibition policy reduces detections by $30\\%$ at equal F-score; and (iii) YOLOv8 object detection: remains success-", + "bbox": [ + 212, + 633, + 787, + 741 + ], + "page_idx": 2 + }, + { + "type": "header", + "text": "Photon Inhibition for Energy-Efficient Single-Photon Imaging", + "bbox": [ + 321, + 114, + 730, + 128 + ], + "page_idx": 2 + }, + { + "type": "page_number", + "text": "3", + "bbox": [ + 774, + 116, + 784, + 126 + ], + "page_idx": 2 + }, + { + "type": "page_footnote", + "text": "It has been shown, perhaps counter-intuitively, that SPADs do not saturate even under extremely bright conditions [26, 27]. Therefore, SPADs are not restricted to low-flux environments, but are being considered for vision applications across a wide dynamic range of lighting conditions (e.g., from a dark tunnel to bright sunlight).", + "bbox": [ + 217, + 755, + 785, + 811 + ], + "page_idx": 2 + }, + { + "type": "page_footnote", + "text": "5 We borrow the term \"inhibition\" from the phenomenon of \"lateral inhibition\" found in biological vision systems [3].", + "bbox": [ + 220, + 811, + 785, + 839 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "ful with $95\\%$ of photons inhibited under camera motion of a real-world SPC (Fig. 1(c,d)). Through experiments with photon streams captured using real SPAD camera hardware over a wide range of illumination conditions, we show that our proposed inhibition policies preserve low-light details and, in brightlight, decouple flux and detection energy.", + "bbox": [ + 212, + 145, + 782, + 220 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Scope and Limitations: There are several competing image sensor technologies today that resolve single-photon while capturing binary-valued raw frames at rates exceeding thousands of frames/second. SPCs based on \"jots\" [32] that do not rely on avalanche multiplication do not suffer from flux-dependent power consumption as SPADs. Jots are an exciting technology, especially in scenarios that require high resolution and high dynamic range imaging [33]. In this paper, we focus on SPADs, due to their rapidly rising availability and commercial interest [6, 28], and benefits over conventional CMOS sensors, both in low-light and bright scenes for a variety of computer vision tasks [10, 35].", + "bbox": [ + 212, + 224, + 784, + 359 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Our goals in this paper are to (a) raise the question of \"which photons should be detected?\" in the context of energy-efficient single-photon imaging, (b) establish a design space and metrics to evaluate various inhibition policies, and (c) propose plausible policies that respect practical hardware limitations for future on-chip implementation. We emphasize that the inhibition policies proposed in this paper are not necessarily optimal. This work is just a first step towards demonstrating that it is possible to achieve high performance with SPADs, while maintaining low power consumption via photon inhibition.", + "bbox": [ + 212, + 361, + 785, + 482 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "2 Related Work", + "text_level": 1, + "bbox": [ + 215, + 505, + 387, + 522 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Hardware approaches for reduced energy consumption: There is a strong dependence of SPAD power consumption on the pixel size — the smaller the pixel, the lower the avalanche energy [39]. Although recent developments in SPAD pixel technology have reduced pixel sizes to below $4\\mu \\mathrm{m}$ [40], avalanche energy still contributes a large fraction of the total power consumption in a SPAD sensor [44]. SPAD design optimizations have reduced the charge per avalanche by RF modulation of the bias voltage [52], minimization of the junction capacitance [42], and smart [5] and fast [54] quenching circuits. Circuit architectures may require spatial and/or temporal co-incidence [22] to reduce energy downstream in the processing chain, but avalanche energy remains. Our work complements existing hardware approaches by preventing avalanches altogether to reduce illumination-dependent energy consumption.", + "bbox": [ + 212, + 537, + 784, + 718 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "SPAD dead-time and clocked recharge: SPADs require recharge after an avalanche-inducing photon detection during which subsequent photons are not recorded. This dead-time inhibits photons at high exposure [24, 27], yet, power consumption remains excessive when the average inter-photon arrival interval is shorter than the SPAD dead time [38, 42]. Power consumption at high photon flux can be reduced by controlling the global rate of SPAD recharge (\"clocked recharge\") [38, 47]. Within a single recharge period a SPAD detects at most a single photon; subsequent photon arrivals do not induce an avalanche, thus power is", + "bbox": [ + 212, + 719, + 785, + 839 + ], + "page_idx": 3 + }, + { + "type": "page_number", + "text": "4", + "bbox": [ + 217, + 114, + 228, + 126 + ], + "page_idx": 3 + }, + { + "type": "header", + "text": "L. Koerner et al.", + "bbox": [ + 271, + 114, + 383, + 126 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "reduced. To further limit avalanches, clocked recharge has been combined with a limit on the number of detections and time to saturation circuitry to measure the intensity of the saturated pixels [41], coarse pixel-wise exposure control [42], and a sequence of recharge periods similar to exposure bracketing [47]. These methods could be considered special cases of inhibition policies that are global and do not adapt to scene content, and therefore, are limited in flexibility to trade power and measurement fidelity. Consequently, in these methods, the power of avalanches remains a considerable fraction of total SPAD sensor power consumption [47]. For example, in this work we show considerable power savings, reaching up to $90\\%$ , for an object detection task.", + "bbox": [ + 212, + 146, + 787, + 297 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Resource-aware imaging: Event-based vision sensors reduce power consumption by only transmitting scene information when an intensity change is detected [17]. This idea has recently been applied to SPAD arrays to reduce power consumption due to data transfer [11, 46]. In contrast, our method reduces power due to detection by selectively disabling photodiodes based on photon history over small spatial and temporal neighborhoods. Miniaturized cameras with constraints on compute energy have transferred processing to the optical domain [30]. While we focus on passive imaging, depth sensing with SPADs and an active pulsed light source has related constraints such as acquisition time and laser power. Adaptive gating reduces acquisition time [43] and optimal allocation of the laser dwell time among pixels improves data quality [36, 48].", + "bbox": [ + 212, + 297, + 787, + 465 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "3 Observation Model", + "text_level": 1, + "bbox": [ + 215, + 491, + 437, + 508 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "During an exposure time $T$ , a photon flux of $\\phi$ results in an average number of photon conversions, or exposure, of $H = \\phi T$ (we fold the sensor's photon detection probability, or PDP, into the definition of $\\phi$ , meaning it represents an effective flux). The distribution of photon conversions, $K$ , follows the Poisson distribution given by $\\mathsf{P}(K = k;H) = \\frac{H^k e^{-H}}{k!}$ . During each binary exposure period, a SPAD pixel records a '1' if at least one photon was incident during that period, and '0' otherwise. The probability of detecting at least one photon is given by $Y := 1 - \\mathsf{P}(K = 0;H) = 1 - e^{-H}$ . Hence, the SPAD pixel readout in each binary frame is a Bernoulli random variable $B \\sim \\mathrm{Bernoulli}(Y)$ . Multiple exposure time windows, or measurements $(W)$ , are recorded to reduce noise with the total number of detections", + "bbox": [ + 212, + 527, + 787, + 696 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\nD := \\left[ \\sum_ {n = 1} ^ {W} B _ {n} \\right] \\sim \\operatorname {B i n o m i a l} (W, 1 - e ^ {- H}). \\tag {1}\n$$\n", + "text_format": "latex", + "bbox": [ + 357, + 705, + 785, + 750 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "We estimate the probability of detection and flux from a measurement $D$ as", + "bbox": [ + 215, + 758, + 754, + 773 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\n\\widehat {Y} = \\frac {D}{W} \\tag {2}\n$$\n", + "text_format": "latex", + "bbox": [ + 439, + 787, + 785, + 815 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\n\\widehat {H} = - \\ln (1 - \\widehat {Y}). \\qquad \\qquad (3)\n$$\n", + "text_format": "latex", + "bbox": [ + 439, + 816, + 785, + 837 + ], + "page_idx": 4 + }, + { + "type": "header", + "text": "Photon Inhibition for Energy-Efficient Single-Photon Imaging", + "bbox": [ + 320, + 114, + 732, + 130 + ], + "page_idx": 4 + }, + { + "type": "page_number", + "text": "5", + "bbox": [ + 774, + 116, + 784, + 126 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Changes under inhibition: Inhibition is represented by a binary state variable $M$ at every pixel, with $M_{n} = 1$ denoting enabled for the $n$ -th measurement period. $M$ is a random variable when an adaptive or data-dependent inhibition policy is used. The total number of measurements changes to $W_{inh.} \\coloneqq \\sum_{n=1}^{W} M_{n} \\leq W$ . The number of detections becomes", + "bbox": [ + 212, + 146, + 787, + 224 + ], + "page_idx": 5 + }, + { + "type": "equation", + "text": "\n$$\nD _ {i n h.} := \\left[ \\sum_ {n = 1} ^ {W} M _ {n} B _ {n} \\right] = \\left[ \\sum_ {\\substack {n = 1 \\\\ M _ {n} = 1}} ^ {W} B _ {n} \\right] \\sim \\operatorname {B i n o m i a l} \\left(W _ {i n h.}, 1 - e ^ {- H}\\right), \\tag{4}\n$$\n", + "text_format": "latex", + "bbox": [ + 269, + 232, + 785, + 287 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "and the flux is estimated similar to Eqs. (2) and (3). The second summation conveys that, with inhibition, the measurements when the pixel is enabled match the original model in Eq. (1) - a result of the memoryless property of the Poisson arrival process. The model above requires that transitions in $M$ are synchronized with the clock signal used to gate the exposure, so that the PDP is not changed by inhibition. A second assumption is that the time for SPAD recharge is small relative to the clock period. This is a desired property for passive SPAD-based imaging, and holds for many state-of-the-art SPADs [42, 44, 47]. It implies that PDP is approximately constant in time and does not depend on prior pixel state.", + "bbox": [ + 212, + 297, + 787, + 435 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "4 Energy-Aware Performance Metrics", + "text_level": 1, + "bbox": [ + 215, + 455, + 599, + 474 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "The exposure-referred signal-to-noise ratio $(\\mathsf{SNR}_H)$ is commonly used to evaluate single-photon sensor performance, and can be computed as the ratio of the true exposure $H$ and the root-mean-squared error in the estimated exposure $\\sqrt{\\mathbb{E}[(\\widehat{H} - H)^2]}$ [15, 55]:", + "bbox": [ + 212, + 487, + 787, + 556 + ], + "page_idx": 5 + }, + { + "type": "equation", + "text": "\n$$\n\\mathsf {S N R} _ {H} = \\frac {H}{\\sqrt {\\mathbb {E} [ (\\widehat {H} - H) ^ {2} ]}} = H \\sqrt {\\frac {W}{e ^ {H} - 1}}. \\tag {5}\n$$\n", + "text_format": "latex", + "bbox": [ + 357, + 566, + 785, + 611 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "At low incident flux, $\\mathsf{SNR}_H$ is low due to shot noise. $\\mathsf{SNR}_H$ improves as the likelihood of a photon detection increases until, in bright light with $H > 1.6$ , the $\\mathsf{SNR}_H$ degrades due to \"soft\" saturation of the response [27,34].", + "bbox": [ + 212, + 619, + 787, + 666 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "We propose two new energy-aware modifications to $\\mathsf{SNR}_H$ to incorporate SPAD energy costs. First, we propose a detection efficiency metric $\\mathsf{SNR}_{H / D}^2$ defined as the square of SNR normalized by the expected number of detections $\\mathbb{E}[D] = W(1 - e^{-H})$ :", + "bbox": [ + 212, + 666, + 787, + 727 + ], + "page_idx": 5 + }, + { + "type": "equation", + "text": "\n$$\n\\mathsf {S N R} _ {H / D} ^ {2} := \\frac {\\mathsf {S N R} _ {H} ^ {2}}{\\mathbb {E} [ D ]} = \\frac {H ^ {2} e ^ {- H}}{(1 - e ^ {- H}) ^ {2}}. \\tag {6}\n$$\n", + "text_format": "latex", + "bbox": [ + 380, + 736, + 785, + 771 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Fig. 3(a) shows the $\\mathsf{SNR}_H$ (black) and the detection efficiency (red) versus the average photon arrivals per period $(H)$ . When $H \\ll 1$ all detections contribute significant information and $\\mathsf{SNR}_H^2$ increases linearly, similar to an ideal non-saturating sensor only limited by Poisson noise. Accordingly, $\\mathsf{SNR}_{H / D}^2 \\cong 1$ , the", + "bbox": [ + 212, + 779, + 787, + 842 + ], + "page_idx": 5 + }, + { + "type": "page_number", + "text": "6", + "bbox": [ + 217, + 114, + 228, + 126 + ], + "page_idx": 5 + }, + { + "type": "header", + "text": "L. Koerner et al.", + "bbox": [ + 271, + 114, + 383, + 127 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "upper bound of this metric. At larger exposure values, beginning around $H \\approx 0.5$ , the sensor begins to saturate which slows the growth of $\\mathrm{SNR}_H$ and thus degrades the detection efficiency.", + "bbox": [ + 212, + 146, + 782, + 191 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "A separate constraint is the total number of recharge periods during which a pixel is enabled and can measure either '0' or '1'. This number of measurements $(W)$ may be limited due to the energy to read out a frame, the depth of an in-pixel counter, and/or the maximum allowable sensing latency due to motion blur. We establish a second metric, measurement efficiency defined as the square of SNR normalized by the number of measurement windows:", + "bbox": [ + 212, + 191, + 784, + 282 + ], + "page_idx": 6 + }, + { + "type": "equation", + "text": "\n$$\n\\mathrm {S N R} _ {H / W} ^ {2} := \\frac {\\mathrm {S N R} _ {H} ^ {2}}{W} = \\frac {H ^ {2} e ^ {- H}}{1 - e ^ {- H}}. \\tag {7}\n$$\n", + "text_format": "latex", + "bbox": [ + 383, + 292, + 785, + 327 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Fig. 3(a) (blue) shows sub-optimal measurement efficiency at both low and high exposures with the best efficiency at $H = 1.59$ , $Y = 0.80$ , as demonstrated in [8].", + "bbox": [ + 212, + 335, + 782, + 366 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "In an oracle setting with a known image, one can analytically derive nonuniform allocations of measurements $W_{i}$ across the pixels $i$ , constraining the total expected detections to a fixed value and optimizing $\\mathsf{SNR}_H$ or mean-squared-error (see the supplement). The loss metric may also be defined relative to the binary rate $Y$ instead of $H$ as above. In that case a useful base metric could be entropy instead of $\\mathsf{SNR}_H$ [19].", + "bbox": [ + 212, + 367, + 784, + 457 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "5 Spatio-temporal Inhibition Policies", + "text_level": 1, + "bbox": [ + 214, + 479, + 589, + 498 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "We now propose policies that calculate a spatio-temporal inhibition pattern for each pixel and each frame based on the history of photon frames and patterns. Following Sec. 3, we define an inhibition pattern using a binary-valued tensor $M$ where $M(i,j,t) = 0$ if pixel $(i,j)$ is disabled in the $t^{\\mathrm{th}}$ frame, and $M(i,j,t) = 1$ otherwise. All pixels are initially enabled, and on-sensor calculations modify $M$ over time. The binary photon cube is defined as $F(i,j,t) = 1$ if the pixel is enabled (i.e., $M(i,j,t) = 1$ ) and a photon is detected at pixel location $(i,j)$ in the $t^{\\mathrm{th}}$ frame, otherwise it is zero.", + "bbox": [ + 212, + 511, + 787, + 632 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/94a9abed29f4a57bd67fe0da72b93450eef4df52c2f03cf44391ec6e9174d8db.jpg", + "image_caption": [ + "Fig. 2: Calculation-based inhibition with dead time. Arrows represent photons with an 'X' for inhibition. $T$ indicates the clocked recharge period. A score, $S$ , calculated from past frames determines if future measurements are enabled or disabled." + ], + "image_footnote": [], + "bbox": [ + 290, + 662, + 707, + 756 + ], + "page_idx": 6 + }, + { + "type": "header", + "text": "Photon Inhibition for Energy-Efficient Single-Photon Imaging", + "bbox": [ + 320, + 114, + 732, + 130 + ], + "page_idx": 6 + }, + { + "type": "page_number", + "text": "7", + "bbox": [ + 774, + 114, + 784, + 126 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/2cae52915ddae7613b5cc833b5b17fb931d9e2f7e5efbb7027946db00bb5e776.jpg", + "image_caption": [ + "Fig. 3: Efficiency metrics and inhibition policies that track the metrics: (a) The $\\mathsf{SNR}_H$ in dB (black), the detection efficiency (red, --), and the measurement efficiency (blue, --) versus the exposure with $W = 100$ measurements. The binary rate $Y = 1 - e^{-H}$ is indicated on the top axis. The vertical dotted line indicates the exposure and the binary rate at which the $\\mathsf{SNR}_H$ degrades by $3\\mathrm{dB}$ from the peak $\\mathsf{SNR}_H$ . (b,c) Monte Carlo simulations of binary images using calculation-based inhibition policies demonstrate how the allocation of measurements versus the pixel exposure depends upon tuning parameters. (b) the inhibition threshold $\\eta$ adjusts the exposure level at which pixels are inhibited to allow the measurement fraction (the ratio of active measurements to total number of frames) to follow the $\\mathsf{SNR}_{H / D}^2$ curve in (a). A smaller threshold more aggressively inhibits photons. (c) demonstrates the impact of the hold-off time, $\\tau_{H}$ , on the number of measurements allocated to the brightest pixels. The legend indicates the total fraction of photons inhibited as $I_F$ . (b,c) show smoothed curves (Lowess filter, fraction of $1 / 5$ ) of the measurement fraction vs. H." + ], + "image_footnote": [], + "bbox": [ + 215, + 143, + 421, + 255 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/3561ef36ff700472374459675ae16d4cbbcce88f9ca78fbc0c2dce44c3b8c89c.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 423, + 152, + 787, + 257 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Fig. 1(b) shows the components of a photon inhibition processing layer, including the binary frames, $F$ , and inhibition pattern $M$ . For ease of on-sensor implementation, we focus on policies that operate on small and local spatiotemporal neighborhoods of fixed sizes. We rely on local arithmetic and Boolean computations and comparison operations, consistent with current in-pixel computational capabilities [2,9,24]. Fig. 2 shows a proposed on-sensor calculation approach that operates in a streaming fashion as frames accumulate to calculate an inhibition score, $S$ , as the result of a spatio-temporal filter of the binary frames and inhibition pattern. The score at each pixel is calculated as", + "bbox": [ + 212, + 489, + 787, + 626 + ], + "page_idx": 7 + }, + { + "type": "equation", + "text": "\n$$\nS (i, j, t) = K * \\left[ \\left(2 F (i, j, t) - \\mathbf {1}\\right) \\cdot M (i, j, t) \\right] \\tag {8}\n$$\n", + "text_format": "latex", + "bbox": [ + 346, + 637, + 785, + 655 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "which applies a spatio-temporal filtering kernel, $K$ , of dimensions $L, H, T$ to a ternary representation of the pixel result (1, 0, or -1 for a detection, a disabled pixel, or a measurement that does not detect a photon, respectively). The kernel $K$ can typically be separated into spatial and temporal components as $K = K_{s} \\otimes K_{t}$ with dimensions $L \\times H \\times 1$ and $1 \\times 1 \\times T$ , respectively. After each binary frame, the score is compared to a threshold $\\eta$ and the pixel is disabled for the subsequent $\\tau_{H}$ frames: $M(i,j,t') = 0$ for $\\{t'|t + 1 \\leq t' \\leq t + 1 + \\tau_{H}\\}$ if $S(i,j,t) > \\eta$ . Observe in Fig. 3(b,c) that decreasing $\\eta$ and increasing $\\tau_{H}$ can be used to attain more aggressive inhibition with a larger fraction of photons being inhibited. The binary rate of each pixel is estimated (using Eq. (4)) as the ratio of detections to (active) measurements: $\\widehat{Y}(i,j,t) = \\sum_{t'} F(i,j,t') / \\sum_{t'} M(i,j,t')$ . This", + "bbox": [ + 212, + 665, + 787, + 843 + ], + "page_idx": 7 + }, + { + "type": "page_number", + "text": "8", + "bbox": [ + 217, + 114, + 228, + 126 + ], + "page_idx": 7 + }, + { + "type": "header", + "text": "L. Koerner et al.", + "bbox": [ + 271, + 114, + 383, + 126 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "calculation requires a record of the inhibition history which could be accumulated by an in-pixel counter or recreated in a downstream processor if all binary frames are read out.", + "bbox": [ + 212, + 146, + 784, + 191 + ], + "page_idx": 8 + }, + { + "type": "image", + "img_path": "images/711ddf144f258a16fb41255be347bd83cd923e9a2ad2c7aafa1b589fde3d9de6.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 295, + 220, + 704, + 309 + ], + "page_idx": 8 + }, + { + "type": "image", + "img_path": "images/b3071da236b4c798e7d1b4eec7e65417dd36ef56e19ebdabd307cf09e731ef3a.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 212, + 316, + 406, + 411 + ], + "page_idx": 8 + }, + { + "type": "image", + "img_path": "images/712b81ee56401f0e953743fd72dc033f08581a329c1b77aa913b339ba81a7efd.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 418, + 328, + 589, + 411 + ], + "page_idx": 8 + }, + { + "type": "image", + "img_path": "images/42cb2d604a71e767307335682a047ec4b5743d8331153e747755af236427e544.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 596, + 319, + 787, + 410 + ], + "page_idx": 8 + }, + { + "type": "image", + "img_path": "images/ea42ce3feb38c87db4188bdd54e4e392605a2a2ba7f28a00a6e56ce83fc2c4bc.jpg", + "image_caption": [ + "Fig. 4: Saturation look-ahead inhibition. (a) When detections $D_{i}$ of cycle $i$ exceed a threshold, $d_{i}$ , the photons arriving during the next recharge cycle, $i + 1$ , are inhibited. For this drawing $[d_{1}, d_{2}] = [3, 3]$ . (b, c, d) Example with three recharge period settings and brackets of $W_{i} = 10$ measurements each. The saturation thresholds are set as $[d_{1}, d_{2}] = [7, 7]$ . The brackets are combined linearly, weighted by $\\mathsf{SNR}_H^2$ [18]. (b, c) With negligible effect on reconstructed SNR relative to conventional clocked-recharge, significantly fewer photon detections occur under the inhibition policy in high-flux settings. (d) The individual brackets are effectively disabled in flux regimes with low detection efficiency, and the detection rate closely tracks the original $\\mathsf{SNR}_H^2$ curve. Simulation details are in the supplement." + ], + "image_footnote": [], + "bbox": [ + 217, + 419, + 787, + 463 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "Example policies: 1) Single-pixel dead time: A spatial kernel $K_{s}$ of dimensions $1 \\times 1$ may mimic the dead time of a passive SPAD without influence from neighboring pixels. Unlike SPAD recharge generated dead time, the temporal kernel may be extended (e.g., $N$ frames long) to establish a rate threshold for inhibition with reduced quantization noise. 2) Local spatio-temporal averaging: Single binary frames are inherently noisy. An inhibition policy that calculates spatiotemporal averages to estimate the local photon rate may reduce the impact of noise on the inhibition pattern. Another benefit, as shown in Fig. 3(b,c), is that the stochastic nature of binary frames smoothes the distribution of measurements versus pixel flux. Discontinuities (\"dips\") in SNR versus photon flux are undesirable due to the potential for artifacts. 3) Edge enhancement: Pixels may be inhibited if a local neighborhood has little spatial variation in photon rate. This can be achieved, for instance, through a spatial filter $K_{s}$ in Eq. (8) which acts", + "bbox": [ + 212, + 643, + 787, + 840 + ], + "page_idx": 8 + }, + { + "type": "header", + "text": "Photon Inhibition for Energy-Efficient Single-Photon Imaging", + "bbox": [ + 321, + 114, + 732, + 128 + ], + "page_idx": 8 + }, + { + "type": "page_number", + "text": "9", + "bbox": [ + 774, + 116, + 784, + 126 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "like a Laplacian filter. Such a strategy may enhance the fidelity of edges in the image while focusing fewer resources on regions with constant illumination.", + "bbox": [ + 212, + 146, + 782, + 176 + ], + "page_idx": 9 + }, + { + "type": "text", + "text": "Fig. 4 presents a second proposed inhibition policy called saturation lookahead. This policy combines exposure brackets and calculation-based inhibition for a light-weight single-pixel inhibition policy. This policy proceeds as a sequence of cycles (indexed by $i$ ) of binary frames where each binary frame within each cycle uses the same exposure time. Cycle exposure times $T_{i}$ progressively increase ( $T_{1} < T_{2} < T_{3} \\ldots$ ) so that measurements taken in an earlier cycle may predict low detection efficiency (near saturation) at longer exposure times and disable the pixel in these subsequent cycles. The exposure level thresholds for inhibition would typically be set such that the number of photons detected at a given flux level tracks the $\\mathsf{SNR}_H^2$ (see Fig. 4(b)), but may be adjusted further based on the relative importance of power consumption, sensing latency, and SNR in an actual application setting.", + "bbox": [ + 212, + 176, + 787, + 358 + ], + "page_idx": 9 + }, + { + "type": "text", + "text": "6 Simulation-based Evaluation", + "text_level": 1, + "bbox": [ + 215, + 378, + 526, + 396 + ], + "page_idx": 9 + }, + { + "type": "text", + "text": "We use Monte Carlo simulations to generate sequences of grayscale binary frames from a dataset of RGB images [1]. The inhibition policies evaluated extend the baseline inhibition generated by clocked recharge. The inhibition score and patterns for various policies and tuning parameters are calculated from these binary frame sequences. Once inhibition patterns are determined, performance is evaluated by tabulating detections, measurements, and image quality or vision task performance for each step in the sequence (see the supplement for details).", + "bbox": [ + 212, + 410, + 782, + 516 + ], + "page_idx": 9 + }, + { + "type": "text", + "text": "Spatio-Temporal Policies for Imaging: Handcrafted spatial kernels (3x3) were combined with an averaging temporal kernel of length 4 to form lightweight inhibition policies that allocate pixel measurements as described in earlier sections for improved image reconstruction. Fig. 5 summarizes the simulation results. Fig. 5(d,e) display reduction in photon detections at equal structural similarity index measure (SSIM) [51], enabled by disabling bright pixels, for an exposure bracketing sequence. Intensity estimates from each bracket are combined using $\\mathsf{SNR}^2$ weighting [18] and then converted to a binary rate estimate at the center exposure level of 1 ppp. The proposed policy demonstrates an average reduction in detections of $42\\%$ as compared to no inhibition. Fig. 5(f,g) evaluates a single exposure level (1.0 ppp) which is a more challenging scenario, yet the proposed policy still reduces detections by $14\\%$ at $\\mathrm{SSIM} = 0.7$ . See supplement S 3.4 for more examples and S 3.5 for simulations of high dynamic range images.", + "bbox": [ + 212, + 518, + 787, + 715 + ], + "page_idx": 9 + }, + { + "type": "text", + "text": "Edge Detection: The BSDS500 dataset with ground truth boundaries [1] was used to study energy-efficient edge detection via photon inhibition. Binary rate images were processed by pre-trained holistically-nested edge detection (HED) [53] with the resulting edge maps compared to ground truth by the structured edge detection toolbox [13]. Fig. 6 shows the optimal image scale (OIS) F-score versus the average detections per pixel. Interpolated curves (not shown) allow for translating along horizontal lines of equal task performance to assess differences in avalanche energy. At low photon counts the proposed edge-enhancing policy", + "bbox": [ + 212, + 718, + 787, + 840 + ], + "page_idx": 9 + }, + { + "type": "page_number", + "text": "10", + "bbox": [ + 217, + 114, + 235, + 126 + ], + "page_idx": 9 + }, + { + "type": "header", + "text": "L. Koerner et al.", + "bbox": [ + 271, + 114, + 383, + 126 + ], + "page_idx": 9 + }, + { + "type": "image", + "img_path": "images/e936f7b8c80d7ef1abc12fc2a2cd0ef2f6ec60c78629b4c18cbb905d34f9b351.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 218, + 146, + 411, + 265 + ], + "page_idx": 10 + }, + { + "type": "image", + "img_path": "images/2c644c6241f580648fde24f36ba2025a5f18aee034b40212bec9a227fc45f8cd.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 410, + 148, + 602, + 265 + ], + "page_idx": 10 + }, + { + "type": "image", + "img_path": "images/37be8622632bf095fe7076254eed99d5c15995de3301d2702bb838cef731556b.jpg", + "image_caption": [ + "Fig. 5: Power-efficient static imaging via inhibition. (a) Images from an exposure bracketing sequence (average exposure of 0.1, 1.0, 10.0 photons per pixel (ppp)) using clocked recharge without (conventional) and with inhibition. (b) the distribution of measurements based on pixel intensity with $H$ at 1.0 ppp. (c) image quality (SSIM) at equal detections/pixel. (d-g) Handcrafted policies are tested over 20 images and assessed by reductions in detections ( $D$ ) at equal SSIM. The top row (d),(e) use exposure bracketing; (f),(g) use a single exposure of 1.0 ppp. (d),(f) at SSIM=0.7 and (e),(g) at SSIM=0.8. The box shows quartiles with the center line at the median. The proposed policy, $P_{cr}$ , is a $3 \\times 3$ spatial kernel that emphasizes the center pixel ( $\\times 8$ ) and includes the 8 neighbors ( $\\times 1$ ) (see the supplement for policy details). (h) average inhibition patterns for each exposure time. The top most pattern inhibits the brightest pixels only (maximum of $\\sim 60\\%$ inhibition, primarily in the sky). For the longest exposure time the inhibition pattern allocates measurements primarily to the darkest pixels." + ], + "image_footnote": [], + "bbox": [ + 217, + 267, + 346, + 396 + ], + "page_idx": 10 + }, + { + "type": "image", + "img_path": "images/8ac98e90c44360cebbc62ff83dd1ec83dae1631a2f739a213827d5e564d21f1d.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 357, + 268, + 504, + 397 + ], + "page_idx": 10 + }, + { + "type": "image", + "img_path": "images/c0cae04302e9f6a845ac9aa869153b185b9068b50cdc3dc96430a6941404f2c3.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 509, + 268, + 616, + 401 + ], + "page_idx": 10 + }, + { + "type": "image", + "img_path": "images/f286949a31cd84d68bfccd8430fe1b23e8343da6f3ad42f7877c170ea4b243d7.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 611, + 148, + 781, + 248 + ], + "page_idx": 10 + }, + { + "type": "image", + "img_path": "images/f1ceba2f6ec65d61dbe7e96d9bb0f0e5aa9832265988c45cbce2bddce4f8a6c5.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 617, + 248, + 782, + 392 + ], + "page_idx": 10 + }, + { + "type": "image", + "img_path": "images/2976340e63e3a778154ddb56dc0a9359323862d7c5fdafe89b3767c77ce2c246.jpg", + "image_caption": [ + "Fig. 6: Energy-efficient edge detection by inhibition: Edge detection F-scores averaged over 19 images versus average detections/pixel. Horizontal arrows show reduction in detections enabled by inhibition at equal task performance. At and beyond $30\\mathrm{D / pix}$ . no inhibition and the proposed policy are nearly equivalent and plateau at $200\\mathrm{D / pix}$ . The edge detector returns $F = 0.813$ on the original images of this set." + ], + "image_footnote": [], + "bbox": [ + 364, + 608, + 640, + 719 + ], + "page_idx": 10 + }, + { + "type": "header", + "text": "Photon Inhibition for Energy-Efficient Single-Photon Imaging", + "bbox": [ + 321, + 114, + 732, + 128 + ], + "page_idx": 10 + }, + { + "type": "page_number", + "text": "11", + "bbox": [ + 767, + 116, + 782, + 126 + ], + "page_idx": 10 + }, + { + "type": "text", + "text": "demonstrates a $30\\%$ reduction in detections on single exposure time captures. This policy extends the on-sensor calculation approach by calculating two scores, the conventional $3 \\times 3$ Laplacian [50] as $S_{1}$ and a $3 \\times 3$ averaging filter as $S_{2}$ . The final inhibition decision is the Boolean operation of these scores as $((\\eta_{1} < S_{1} < \\eta_{2}) \\land (S_{2} > \\eta_{3})) \\lor (S_{2} > \\eta_{4})$ . The Laplacian policy alone performs poorly as photons in dim regions with minimal spatial variation are inhibited - disabling dim pixels is energy inefficient. Complete policy descriptions are in the supplement.", + "bbox": [ + 212, + 146, + 787, + 252 + ], + "page_idx": 11 + }, + { + "type": "text", + "text": "7 Experiments on Dynamic Scenes", + "text_level": 1, + "bbox": [ + 215, + 277, + 568, + 295 + ], + "page_idx": 11 + }, + { + "type": "text", + "text": "Many real-world scenes contain significant motion even with the high frame rate of a SPAD camera. Burst reconstruction algorithms yield high-quality images from sequences of binary frames [34,35] - we investigate their compatibility with the inhibited photon detection data. In particular, we focus on the saturation look-ahead policy of Fig. 4(a), applied independently at every pixel, and therefore an example of an adaptive single-pixel temporal policy.", + "bbox": [ + 212, + 311, + 785, + 402 + ], + "page_idx": 11 + }, + { + "type": "text", + "text": "To avoid losing salient information under motion, the inhibition policy must limit periods of extended dead time. In the context of the saturation look-ahead policy this limits the total exposure length of the bracketing sequence, since we implicitly assume flux to be constant within each sequence. In our experiments, we use a Fibonacci bracketing [20] sequence $T \\coloneqq \\{1, 1, 2, 3, 5, 8, 13, 21\\}^6$ , denoted in the units of a single minimal exposure time. Every measured bracket is converted to a maximum-likelihood estimate of flux, which is then supplied to the quanta burst photography algorithm [34]. The flux estimator from bracketed measurements is described in detail in supplement Sec. S7.2.", + "bbox": [ + 212, + 402, + 785, + 539 + ], + "page_idx": 11 + }, + { + "type": "text", + "text": "Results with the SwissSPAD2 sensor: The SwissSPAD2 sensor [49] is a prototype SPAD pixel array that can produce binary frames at a rate up to 97,700 FPS, with a resolution of $512 \\times 256$ . In our experiments, we use binary frames captured directly (without inhibition) by the SPAD array as reference data and emulate on-sensor saturation look-ahead inhibition in software. As a pre-processing step, measurements at hot pixels are replaced with their nearest neighbors.", + "bbox": [ + 212, + 542, + 785, + 648 + ], + "page_idx": 11 + }, + { + "type": "text", + "text": "Fig. 7 shows the results of burst reconstruction under three lighting conditions. The raw data is a sequence of $>580,000$ binary frames with scene radiance increasing rapidly by orders of magnitude (Fig. 7(b)), from $<1$ lux to $>4,000$ lux, measured separately with a light meter. For each of 47 equally-spaced keyframes, centered windows of 12,000 binary frames are extracted and processed as described above. Results for the full sequence can be found in the supplementary material. A static inhibition policy of regular sub-sampling (dropping 9 out of every 10 frames) is also applied, which yields a fixed $90\\%$ reduction in both measurements and photon detections under all lighting conditions. Other subsampling factors are discussed in the supplement.", + "bbox": [ + 212, + 648, + 785, + 800 + ], + "page_idx": 11 + }, + { + "type": "page_number", + "text": "12", + "bbox": [ + 217, + 114, + 235, + 126 + ], + "page_idx": 11 + }, + { + "type": "header", + "text": "L. Koerner et al.", + "bbox": [ + 271, + 114, + 383, + 126 + ], + "page_idx": 11 + }, + { + "type": "page_footnote", + "text": "Sequence $T' \\coloneqq \\{1, 1, 1, 3, 3, 3, 8, 8, 25\\}$ yielded similar results. No extensive search over the policy space was performed.", + "bbox": [ + 217, + 810, + 785, + 839 + ], + "page_idx": 11 + }, + { + "type": "image", + "img_path": "images/bcf274d6a24deaa3ce2c5d1e440ce97657328e2c77d381e78e05599f0ff2b257.jpg", + "image_caption": [ + "Fig. 7: Adaptive policies on video sequences enable stronger inhibition, preserve low-light details and, in bright-light, decouple flux and detection energy. (a) Burst reconstructions [34] for three keyframes with varying light levels: the top and bottom row differ by $\\approx 7$ stops. [Images show the detection rates $Y$ and are further gamma-compressed ( $\\gamma = 0.4$ ). Complete results included in the supplement.] Left column results are from the original binary frames without inhibition, and the right column after sub-sampling $10 \\times$ (a fixed $90\\%$ inhibition). Middle column represents exposure bracketing combined with saturation look-ahead (see Fig. 4a for description). Under strong light (top row) the results are reasonable with both methods. However, plain sub-sampling loses details in lower light: notice the furniture and a person's outline in the middle & bottom rows, respectively. Inhibition is instead adaptive to flux. (b) Average exposure level for each keyframe in the sequence. (c,d) Per-keyframe and cumulative detection counts - inhibition ultimately results in fewer photons being detected over the whole sequence. (e) Number of measurements taken for each keyframe; reductions may be translated to energy savings during read-out. Plots in (c,d,e) are sub-sampled for clarity, and crossover points are marked by green arrows." + ], + "image_footnote": [], + "bbox": [ + 212, + 141, + 733, + 476 + ], + "page_idx": 12 + }, + { + "type": "text", + "text": "The top row in Fig. 7(a) shows that under strong light a large fraction of photons ( $>90\\%$ ) can be inhibited through the saturation look-ahead policy and still result in good image quality after burst reconstruction, thus spreading photon detections over a longer period of time to reduce avalanche power. Even simple sub-sampling yields good results in bright light, and may work well under controlled illumination conditions. However, the images in the middle and", + "bbox": [ + 212, + 748, + 787, + 839 + ], + "page_idx": 12 + }, + { + "type": "header", + "text": "Photon Inhibition for Energy-Efficient Single-Photon Imaging", + "bbox": [ + 321, + 114, + 732, + 128 + ], + "page_idx": 12 + }, + { + "type": "page_number", + "text": "13", + "bbox": [ + 767, + 114, + 785, + 126 + ], + "page_idx": 12 + }, + { + "type": "text", + "text": "bottom rows illustrate that this static inhibition policy results in excessive signal loss in lower light leading to a loss of details. Sub-sampling may be enhanced by adjusting exposure time and/or the sub-sampling factor in response to global flux but cannot simultaneously optimize for different light levels within a single frame: over-exposed regions may have clipping artifacts and under-exposed regions are prone to motion blur from incorrect burst reconstruction. The saturation look-ahead policy does adapt to local flux, and allocates relatively more detections to dim regions (see Fig. 1(d)). It thus decouples detection energy from flux (Fig. 7(c); cf. Fig. 1(a)), and results in cumulatively fewer detections than sub-sampling (Fig. 7(d)) due to it being more aggressive in strong light.", + "bbox": [ + 212, + 146, + 787, + 297 + ], + "page_idx": 13 + }, + { + "type": "text", + "text": "Apart from photography, the burst-reconstructed images can be used in computer vision applications, where even stronger inhibition is possible depending on the noise- or blur-robustness of the vision algorithm. Fig. 1(d) shows successful object detection with the YOLOv8 algorithm [29] applied to a burst-reconstructed image, with approximately $95\\%$ photon detections inhibited.", + "bbox": [ + 212, + 297, + 787, + 375 + ], + "page_idx": 13 + }, + { + "type": "text", + "text": "8 Limitations and Future Outlook", + "text_level": 1, + "bbox": [ + 214, + 398, + 562, + 415 + ], + "page_idx": 13 + }, + { + "type": "text", + "text": "Implementation costs. Our focus in this paper is on reducing energy consumption due to avalanches. Although avalanches contribute significant energy as compared to on-sensor computations [2, 42, 47], an important next step is to design a holistic model that includes the energy consumption of computations and readout (leveraging Eq. 7). Our in-pixel computations — power-of-two multiplications which simplify to bit shifts, small spatio-temporal kernels no larger than $3 \\times 3 \\times 4$ — are designed to be lightweight. Fortunately, computational SPAD imagers [2] with in-pixel memory and compute have recently been proposed, with $4 \\times 4$ block of pixels having a 32-bit CPU and over 10 kbits of memory. See the supplement for an estimate of the required circuitry.", + "bbox": [ + 212, + 431, + 787, + 583 + ], + "page_idx": 13 + }, + { + "type": "text", + "text": "More complete models. The noise model used for efficiency metrics only accounts for photon Poisson noise and the quantization noise of Bernoulli samples. Expanding the noise model to include uncorrected pixel sensitivity variations, crosstalk, and afterpulsing may improve performance [4] by allocating measurements with an awareness of practical sensor limitations. Inhibition modulates the number of Bernoulli trials. An unbiased estimator for data-dependent stopping of Bernoulli trials is known [23], yet is not applied in this work since it does not precisely match our situation. Our simulations (SSIM) suggest that bias is less significant than image noise, yet future analytical studies are needed over a range of conditions.", + "bbox": [ + 212, + 587, + 787, + 738 + ], + "page_idx": 13 + }, + { + "type": "text", + "text": "Generalization to other tasks. To generalize the proposed approaches to a variety of vision tasks, task-specific quality metrics must be defined to pose optimization problems for each task. As an example, for the image reconstruction task, Suppl. Sec. 3.1 shows an analytical optimization using the MSE metric when constrained by photon detections. Other tasks could be approached similarly, yet may need to be optimized empirically.", + "bbox": [ + 212, + 742, + 787, + 834 + ], + "page_idx": 13 + }, + { + "type": "page_number", + "text": "14", + "bbox": [ + 217, + 114, + 235, + 126 + ], + "page_idx": 13 + }, + { + "type": "header", + "text": "L. Koerner et al.", + "bbox": [ + 271, + 114, + 383, + 126 + ], + "page_idx": 13 + }, + { + "type": "text", + "text": "Acknowledgments", + "text_level": 1, + "bbox": [ + 217, + 143, + 392, + 162 + ], + "page_idx": 14 + }, + { + "type": "text", + "text": "The authors acknowledge the Minnesota Supercomputing Institute (MSI) at the University of Minnesota for providing resources. A.I. was supported in part by NSF ECCS-2138471. S.G and M.G. were supported in part by NSF CAREER award 1943149, NSF award CNS-2107060, and Wisconsin Alumni Research Foundation (WARF).", + "bbox": [ + 212, + 180, + 787, + 257 + ], + "page_idx": 14 + }, + { + "type": "text", + "text": "References", + "text_level": 1, + "bbox": [ + 215, + 282, + 321, + 299 + ], + "page_idx": 14 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "1. Arbeláez, P., Maire, M., Fowlkes, C., Malik, J.: Contour Detection and Hierarchical Image Segmentation. IEEE Transactions on Pattern Analysis and Machine Intelligence 33(5), 898-916 (May 2011). https://doi.org/10.1109/TPAMI.2010.161", + "2. Ardelean, A.: Computational Imaging SPAD Cameras. Ph.D. thesis, EPFL (2023). https://doi.org/10.5075/epfl-thesis-9501", + "3. Barlow, H.B.: Summation and inhibition in the frog's retina. The Journal of physiology 119(1), 69 (1953)", + "4. Bian, L., Song, H., Peng, L., Chang, X., Yang, X., Horstmeyer, R., Ye, L., Zhu, C., Qin, T., Zheng, D., Zhang, J.: High-resolution single-photon imaging with physics-informed deep learning. Nature Communications 14(1), 5902 (Sep 2023). https://doi.org/10.1038/s41467-023-41597-9", + "5. Boso, G., Buttafava, M., Villa, F., Tosi, A.: Low-cost and compact single-photon counter based on a CMOS SPAD smart pixel. IEEE Photonics Technology Letters 27(23), 2504-2507 (2015)", + "6. Canon Inc.: Canon Launches MS-500 - The World's First Ultra-High-Sensitivity Interchangeable-Lens SPAD Sensor Camera. https://www.usa.canon.com/newsroom/2023/20230801-ms500, Canon Press Release 8/1/2023. Accessed 2/25/2024.", + "7. Carey, S.J., Lopich, A., Barr, D.R., Wang, B., Dudek, P.: A 100,000 fps vision sensor with embedded 535 GOPS/W $256 \\times 256$ SIMD processor array. In: 2013 Symposium on VLSI Circuits. pp. C182-C183 (Jun 2013)", + "8. Chan, S.H.: What Does a One-Bit Quanta Image Sensor Offer? IEEE Transactions on Computational Imaging 8, 770-783 (2022). https://doi.org/10.1109/TCI.2022.3202012", + "9. Charbon, E., Bruschini, C., Lee, M.J.: 3D-Stacked CMOS SPAD Image Sensors: Technology and Applications. In: 2018 25th IEEE International Conference on Electronics, Circuits and Systems (ICECS). pp. 1-4 (Dec 2018). https://doi.org/10.1109/ICECS.2018.8617983", + "10. Chi, Y., Gnanasambandam, A., Koltun, V., Chan, S.H.: Dynamic low-light imaging with quanta image sensors. In: Computer Vision-ECCV 2020: 16th European Conference, Glasgow, UK, August 23-28, 2020, Proceedings, Part XXI 16. pp. 122-138. Springer (2020)", + "11. Della Rocca, F.M., Mai, H., Hutchings, S.W., Al Abbas, T., Buckbee, K., Tsiamis, A., Lomax, P., Gyongy, I., Dutton, N.A., Henderson, R.K.: A $128 \\times 128$ SPAD motion-triggered time-of-flight image sensor with in-pixel histogram and column-parallel vision processor. IEEE Journal of Solid-State Circuits 55(7), 1762-1775 (2020). https://doi.org/10.1109/JSSC.2020.2993722" + ], + "bbox": [ + 217, + 316, + 785, + 839 + ], + "page_idx": 14 + }, + { + "type": "header", + "text": "Photon Inhibition for Energy-Efficient Single-Photon Imaging", + "bbox": [ + 321, + 114, + 730, + 128 + ], + "page_idx": 14 + }, + { + "type": "page_number", + "text": "15", + "bbox": [ + 767, + 116, + 784, + 126 + ], + "page_idx": 14 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "12. Diamond, J.S.: Inhibitory interneurons in the retina: Types, circuitry, and function. Annual Review of Vision Science 3(1), 1-24 (2017). https://doi.org/10.1146/annurev-vision-102016-061345", + "13. Dollar, P., Zitnick, C.L.: Structured Forests for Fast Edge Detection. In: Proceedings of the IEEE International Conference on Computer Vision. pp. 1841-1848 (2013)", + "14. Dutton, N.A.W., Al Abbas, T., Gyongy, I., Mattioli Della Rocca, F., Henderson, R.K.: High Dynamic Range Imaging at the Quantum Limit with Single Photon Avalanche Diode-Based Image Sensors. Sensors 18(4), 1166 (Apr 2018). https://doi.org/10.3390/s18041166", + "15. Fossum, E.R.: Modeling the Performance of Single-Bit and Multi-Bit Quanta Image Sensors. IEEE Journal of the Electron Devices Society 1(9), 166-174 (Sep 2013). https://doi.org/10.1109/JEDS.2013.2284054", + "16. Franke, K., Berens, P., Schubert, T., Bethge, M., Euler, T., Baden, T.: Inhibition decorrelates visual feature representations in the inner retina. Nature 542(7642), 439-444 (Feb 2017). https://doi.org/10.1038/nature21394", + "17. Gallego, G., Delbrück, T., Orchard, G., Bartolozzi, C., Taba, B., Censi, A., Leutenegger, S., Davison, A.J., Conradt, J., Daniilidis, K., et al.: Event-based vision: A survey. IEEE Transactions on Pattern Analysis and Machine Intelligence 44(1), 154-180 (2020)", + "18. Gnanasambandam, A., Chan, S.H.: HDR Imaging With Quanta Image Sensors: Theoretical Limits and Optimal Reconstruction. IEEE Transactions on Computational Imaging 6, 1571-1585 (2020). https://doi.org/10.1109/TCI.2020.3041093", + "19. Gnanasambandam, A., Chan, S.H.: Exposure-Referred Signal-to-Noise Ratio for Digital Image Sensors. IEEE Transactions on Computational Imaging 8, 561-575 (Jun 2022). https://doi.org/10.1109/TCI.2022.3187657", + "20. Gupta, M., Iso, D., Nayar, S.K.: Fibonacci Exposure Bracketing for High Dynamic Range Imaging. In: 2013 IEEE International Conference on Computer Vision. pp. 1473-1480. IEEE, Sydney, Australia (Dec 2013). https://doi.org/10.1109/ICCV.2013.186", + "21. Gyongy, I., Dutton, N.A., Henderson, R.K.: Direct time-of-flight single-photon imaging. IEEE Transactions on Electron Devices 69(6), 2794-2805 (2021)", + "22. Gyongy, I., Martin, G.M., Turpin, A., Ruget, A., Halimi, A., Henderson, R., Leach, J.: High-speed vision with a 3D-stacked SPAD image sensor. In: Advanced Photon Counting Techniques XV. vol. 11721, p. 1172105. SPIE (2021)", + "23. Haldane, J.B.S.: On a Method Of Estimating Frequencies. Biometrika 33(3), 222-225 (Nov 1945). https://doi.org/10.1093/biomet/33.3.222", + "24. Henderson, R.K., Johnston, N., Hutchings, S.W., Gyongy, I., Al Abbas, T., Dutton, N., Tyler, M., Chan, S., Leach, J.: $256 \\times 256$ 40nm/90nm CMOS 3D-stacked 120 dB dynamic-range reconfigurable time-resolved SPAD imager. In: 2019 IEEE International Solid-State Circuits Conference-(ISSCC). pp. 106-108. IEEE (2019). https://doi.org/10.1109/ISSCC.2019.8662355", + "25. Iams, H., Salzberg, B.: The secondary emission phototube. Proceedings of the Institute of Radio Engineers 23(1), 55-64 (1935)", + "26. Ingle, A., Seets, T., Buttafava, M., Gupta, S., Tosi, A., Gupta, M., Velten, A.: Passive Inter-Photon Imaging. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 8585–8595 (2021)", + "27. Ingle, A., Velten, A., Gupta, M.: High Flux Passive Imaging With Single-Photon Sensors. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 6760-6769 (2019)" + ], + "bbox": [ + 217, + 147, + 784, + 839 + ], + "page_idx": 15 + }, + { + "type": "page_number", + "text": "16", + "bbox": [ + 217, + 114, + 235, + 126 + ], + "page_idx": 15 + }, + { + "type": "header", + "text": "L. Koerner et al.", + "bbox": [ + 271, + 114, + 383, + 126 + ], + "page_idx": 15 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "28. J Yoshida: Breaking Down iPad Pro 11's LiDAR Scanner. https://www.eetimes.com/breaking-down-ipad-pro-11s-lidar-scanner/, EE Times 6/5/2020. Accessed 5/6/2021.", + "29. Jocher, G., Chaurasia, A., Qiu, J.: Ultralytics YOLOv8 (Jan 2023), https://github.com/ultralytics/ultralytics, Accessed 3/6/2024", + "30. Koppal, S.J., Gkioulekas, I., Young, T., Park, H., Crozier, K.B., Barrows, G.L., Zickler, T.: Toward Wide-Angle Microvision Sensors. IEEE Transactions on Pattern Analysis and Machine Intelligence 35(12), 2982-2996 (Dec 2013). https://doi.org/10.1109/TPAMI.2013.22", + "31. Liu, Y., Gutierrez-Barragan, F., Ingle, A., Gupta, M., Velten, A.: Single-Photon Camera Guided Extreme Dynamic Range Imaging. In: Proceedings of the IEEE/CVF Winter Conference on Applications of Computer Vision. pp. 1575-1585 (2022)", + "32. Ma, J., Masoodian, S., Starkey, D.A., Fossum, E.R.: Photon-number-resolving megapixel image sensor at room temperature without avalanche gain. Optica 4(12), 1474-1481 (Dec 2017). https://doi.org/10.1364/OPTICA.4.001474", + "33. Ma, J., Zhang, D., Elgendy, O.A., Masoodian, S.: A 0.19e- rms Read Noise 16.7Mpixel Stacked Quanta Image Sensor With 1.1 Mm-Pitch Backside Illuminated Pixels. IEEE Electron Device Letters 42(6), 891-894 (Jun 2021). https: //doi.org/10.1109/LED.2021.3072842", + "34. Ma, S., Gupta, S., Ulku, A.C., Bruschini, C., Charbon, E., Gupta, M.: Quanta burst photography. ACM Transactions on Graphics 39(4), 79:1-79:16 (Aug 2020). https://doi.org/10.1145/3386569.3392470", + "35. Ma, S., Mos, P., Charbon, E., Gupta, M.: Burst Vision Using Single-Photon Cameras. In: Proceedings of the IEEE/CVF Winter Conference on Applications of Computer Vision. pp. 5375-5385 (2023)", + "36. Medin, S.C., Murray-Bruce, J., Castanón, D., Goyal, V.K.: Beyond Binomial and Negative Binomial: Adaptation in Bernoulli Parameter Estimation. IEEE Transactions on Computational Imaging 5(4), 570-584 (Dec 2019). https://doi.org/10.1109/TCI.2019.2913108", + "37. Morimoto, K., Iwata, J., Shinohara, M., Sekine, H., Abdelghafar, A., Tsuchiya, H., Kuroda, Y., Tojima, K., Endo, W., Maehashi, Y., Ota, Y., Sasago, T., Maekawa, S., Hikosaka, S., Kanou, T., Kato, A., Tezuka, T., Yoshizaki, S., Ogawa, T., Uehira, K., Ehara, A., Inui, F., Matsuno, Y., Sakurai, K., Ichikawa, T.: 3.2 Megapixel 3D-Stacked Charge Focusing SPAD for Low-Light Imaging and Depth Sensing. In: 2021 IEEE International Electron Devices Meeting (IEDM). pp. 20.2.1–20.2.4 (Dec 2021). https://doi.org/10.1109/IEDM19574.2021.9720605", + "38. Morimoto, K., Ardelean, A., Wu, M.L., Ulku, A.C., Antolovic, I.M., Bruschini, C., Charbon, E.: Megapixel time-gated SPAD image sensor for 2D and 3D imaging applications. Optica 7(4), 346-354 (2020). https://doi.org/10.1364/OPTICA.386574", + "39. Morimoto, K., Charbon, E.: A scaling law for SPAD pixel miniaturization. Sensors 21(10), 3447 (2021)", + "40. Ogi, J., Sano, F., Nakata, T., Kubo, Y., Onishi, W., Koswaththaghe, C., Mochizuki, T., Tashiro, Y., Hizu, K., Takatsuka, T., et al.: A $3.06\\mu \\mathrm{m}$ spad pixel with embedded metal contact and power grid on deep trench pixel isolation for high-resolution photon-counting. In: 2023 International Image Sensor Workshop (IISW) (May 2023)", + "41. Ogi, J., Takatsuka, T., Hizu, K., Inaoka, Y., Zhu, H., Tochigi, Y., Tashiro, Y., Sano, F., Murakawa, Y., Nakamura, M., Oike, Y.: A 250fps 124dB Dynamic-Range" + ], + "bbox": [ + 215, + 146, + 785, + 839 + ], + "page_idx": 16 + }, + { + "type": "header", + "text": "Photon Inhibition for Energy-Efficient Single-Photon Imaging", + "bbox": [ + 321, + 114, + 730, + 128 + ], + "page_idx": 16 + }, + { + "type": "page_number", + "text": "17", + "bbox": [ + 767, + 116, + 784, + 126 + ], + "page_idx": 16 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "SPAD Image Sensor Stacked with Pixel-Parallel Photon Counter Employing SubFrame Extrapolating Architecture for Motion Artifact Suppression. In: 2021 IEEE International Solid-State Circuits Conference (ISSCC). vol. 64, pp. 113-115 (Feb 2021). https://doi.org/10.1109/ISSCC42613.2021.9365977", + "42. Ota, Y., Morimoto, K., Sasago, T., Shinohara, M., Kuroda, Y., Endo, W., Maehashi, Y., Maekawa, S., Tsuchiya, H., Abdelahafar, A., Hikosaka, S., Motoyama, M., Tojima, K., Uehira, K., Iwata, J., Inui, F., Matsuno, Y., Sakurai, K., Ichikawa, T.: A 0.37W 143dB-Dynamic-Range 1Mpixel Backside-Illuminated Charge-Focusing SPAD Image Sensor with Pixel-Wise Exposure Control and Adaptive Clocked Recharging. In: 2022 IEEE International Solid-State Circuits Conference (ISSCC). vol. 65, pp. 94-96 (Feb 2022). https://doi.org/10.1109/ISSCC42614.2022.9731644", + "43. Po, R., Pediredla, A., Gkioulekas, I.: Adaptive gating for single-photon 3d imaging. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 16354-16363 (2022)", + "44. Severini, F., Cusini, I., Berretta, D., Pasquinelli, K., Incoronato, A., Villa, F.: SPAD pixel with sub-ns dead-time for high-count rate applications. IEEE Journal of Selected Topics in Quantum Electronics 28(2: Optical Detectors), 1-8 (2021)", + "45. Shenzhen CM Technology company Ltd: 16 MP MIPI Camera Module with SONY IMX206 sensor. http://www-camera-module.com/product/mipicameramodule/16mp-mipi-camera-module-sony-imx206-sensor.html (Nov 2023)", + "46. Sundar, V., Ardelean, A., Swedish, T., Brusschini, C., Charbon, E., Gupta, M.: Sodacam: Software-defined cameras via single-photon imaging. In: Proceedings of the IEEE International Conference on Computer Vision (ICCV) (2023)", + "47. Takatsuka, T., Ogi, J., Ikeda, Y., Hizu, K., Inaoka, Y., Sakama, S., Watanabe, I., Ishikawa, T., Shimada, S., Suzuki, J., Maeda, H., Toshima, K., Nonaka, Y., Yamamura, A., Ozawa, H., Koga, F., Oike, Y.: A 3.36 $\\mu$ m-pitch SPAD photon-counting image sensor using clustered multi-cycle clocked recharging technique with intermediate most-significant-bit readout. In: 2023 IEEE Symposium on VLSI Technology and Circuits (VLSI Technology and Circuits). pp. 1-2 (Jun 2023). https://doi.org/10.23919/VLSITechnologyandCir57934.2023.10185241", + "48. Tilmon, B., Sun, Z., Koppal, S.J., Wu, Y., Evangelidis, G., Zahreddine, R., Krishnan, G., Ma, S., Wang, J.: Energy-Efficient Adaptive 3D Sensing. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 5054-5063 (2023)", + "49. Ulku, A.C., Bruschini, C., Antolovic, I.M., Kuo, Y., Ankri, R., Weiss, S., Michalet, X., Charbon, E.: A $512 \\times 512$ SPAD Image Sensor With Integrated Gating for Widefield FLIM. IEEE Journal of Selected Topics in Quantum Electronics 25(1), 1-12 (Jan 2019). https://doi.org/10.1109/jstqe.2018.2867439", + "50. Wang, X.: Laplacian operator-based edge detectors. IEEE Transactions on Pattern Analysis and Machine Intelligence 29(5), 886-890 (2007)", + "51. Wang, Z., Bovik, A.C., Sheikh, H.R., Simoncelli, E.P.: Image quality assessment: From error visibility to structural similarity. IEEE Transactions on Image Processing 13(4), 600-612 (2004)", + "52. Wayne, M.A., Bienfang, J.C., Migdall, A.L.: Low-noise photon counting above $100 \\times 106$ counts per second with a high-efficiency reach-through single-photon avalanche diode system. Applied Physics Letters 118(13) (2021)", + "53. Xie, S., Tu, Z.: Holistically-nested edge detection. In: Proceedings of IEEE International Conference on Computer Vision. pp. 1395-1403 (2015)" + ], + "bbox": [ + 215, + 146, + 785, + 840 + ], + "page_idx": 17 + }, + { + "type": "page_number", + "text": "18", + "bbox": [ + 217, + 114, + 235, + 126 + ], + "page_idx": 17 + }, + { + "type": "header", + "text": "L. Koerner et al.", + "bbox": [ + 271, + 114, + 383, + 126 + ], + "page_idx": 17 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "54. Xu, Y., Lu, J., Wu, Z.: A compact high-speed active quenching and recharging circuit for SPAD detectors. IEEE Photonics Journal 12(5), 1-8 (2020)", + "55. Yang, F., Lu, Y.M., Sbaiz, L., Vetterli, M.: Bits from photons: Oversampled image acquisition using binary poisson statistics. IEEE Transactions on image processing 21(4), 1421-1436 (2011)" + ], + "bbox": [ + 215, + 146, + 785, + 217 + ], + "page_idx": 18 + }, + { + "type": "header", + "text": "Photon Inhibition for Energy-Efficient Single-Photon Imaging", + "bbox": [ + 320, + 114, + 730, + 128 + ], + "page_idx": 18 + }, + { + "type": "page_number", + "text": "19", + "bbox": [ + 767, + 116, + 785, + 126 + ], + "page_idx": 18 + } +] \ No newline at end of file diff --git a/2024/Photon Inhibition for Energy-Efficient Single-Photon Imaging/03673fae-a9a6-4975-8cf4-3d5bc33fd48c_model.json b/2024/Photon Inhibition for Energy-Efficient Single-Photon Imaging/03673fae-a9a6-4975-8cf4-3d5bc33fd48c_model.json new file mode 100644 index 0000000000000000000000000000000000000000..3837e31b194a846333a2053c8ab40446bcd92217 --- /dev/null +++ b/2024/Photon Inhibition for Energy-Efficient Single-Photon Imaging/03673fae-a9a6-4975-8cf4-3d5bc33fd48c_model.json @@ -0,0 +1,2350 @@ +[ + [ + { + "type": "title", + "bbox": [ + 0.284, + 0.141, + 0.72, + 0.187 + ], + "angle": 0, + "content": "Photon Inhibition for Energy-Efficient Single-Photon Imaging" + }, + { + "type": "text", + "bbox": [ + 0.227, + 0.212, + 0.776, + 0.229 + ], + "angle": 0, + "content": "Lucas J. Koerner1, Shantanu Gupta2, Atul Ingle3, and Mohit Gupta2" + }, + { + "type": "text", + "bbox": [ + 0.216, + 0.239, + 0.783, + 0.265 + ], + "angle": 0, + "content": "\\(^{1}\\) Department of Electrical and Computer Engineering, University of St. Thomas, St. Paul, MN, 55105, USA koerner.lucas@stthomas.edu" + }, + { + "type": "text", + "bbox": [ + 0.216, + 0.268, + 0.783, + 0.294 + ], + "angle": 0, + "content": "\\(^{2}\\) Department of Computer Sciences, University of Wisconsin-Madison, Madison, WI, 53706, USA {sgupta,mohitg}@cs.wisc.edu" + }, + { + "type": "text", + "bbox": [ + 0.216, + 0.295, + 0.782, + 0.322 + ], + "angle": 0, + "content": "\\(^{3}\\) Department of Computer Science, Portland State University, Portland, OR, 97201, USA \ningle2@pdx.edu" + }, + { + "type": "list", + "bbox": [ + 0.216, + 0.239, + 0.783, + 0.322 + ], + "angle": 0, + "content": null + }, + { + "type": "text", + "bbox": [ + 0.263, + 0.36, + 0.74, + 0.665 + ], + "angle": 0, + "content": "Abstract. Single-photon cameras (SPCs) are emerging as sensors of choice for various challenging imaging applications. One class of SPCs based on the single-photon avalanche diode (SPAD) detects individual photons using an avalanche process; the raw photon data can then be processed to extract scene information under extremely low light, high dynamic range, and rapid motion. Yet, single-photon sensitivity in SPADs comes at a cost — each photon detection consumes more energy than that of a CMOS camera. This avalanche power significantly limits sensor resolution and could restrict widespread adoption of SPAD-based SPCs. We propose a computational-imaging approach called photon inhibition to address this challenge. Photon inhibition strategically allocates detections in space and time based on downstream inference task goals and resource constraints. We develop lightweight, on-sensor computational inhibition policies that use past photon data to disable SPAD pixels in real-time, to select the most informative future photons. As case studies, we design policies tailored for image reconstruction and edge detection, and demonstrate, both via simulations and real SPC captured data, considerable reduction in photon detections (over \\(90\\%\\) of photons) while maintaining task performance metrics. Our work raises the question of \"which photons should be detected?\", and paves the way for future energy-efficient single-photon imaging. Source code for our experiments is available at https://wisionlab.com/project/inhibition." + }, + { + "type": "text", + "bbox": [ + 0.263, + 0.678, + 0.717, + 0.693 + ], + "angle": 0, + "content": "Keywords: single-photon imaging \\(\\cdot\\) SPADs \\(\\cdot\\) energy-efficient vision" + }, + { + "type": "title", + "bbox": [ + 0.216, + 0.718, + 0.377, + 0.735 + ], + "angle": 0, + "content": "1 Introduction" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.75, + 0.788, + 0.842 + ], + "angle": 0, + "content": "From vacuum tube-based single-photon detectors invented in the early 20th century [25], to 3D depth sensing LiDAR cameras found in today's smartphones [28], single-photon camera (SPC) technology has come a long way in terms of pixel resolution and commercial availability for a variety of applications. Thanks to CMOS-compatible single-photon avalanche diode (SPAD) arrays, SPCs are being increasingly used not only in niche fields such as scientific imaging and biomedical" + } + ], + [ + { + "type": "page_number", + "bbox": [ + 0.218, + 0.116, + 0.23, + 0.127 + ], + "angle": 0, + "content": "2" + }, + { + "type": "header", + "bbox": [ + 0.272, + 0.115, + 0.384, + 0.127 + ], + "angle": 0, + "content": "L. Koerner et al." + }, + { + "type": "image_caption", + "bbox": [ + 0.218, + 0.144, + 0.239, + 0.156 + ], + "angle": 0, + "content": "(a)" + }, + { + "type": "image", + "bbox": [ + 0.217, + 0.156, + 0.446, + 0.269 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.451, + 0.144, + 0.471, + 0.157 + ], + "angle": 0, + "content": "(c)" + }, + { + "type": "image", + "bbox": [ + 0.454, + 0.165, + 0.626, + 0.258 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.642, + 0.165, + 0.781, + 0.244 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.218, + 0.269, + 0.236, + 0.281 + ], + "angle": 0, + "content": "(b)" + }, + { + "type": "image", + "bbox": [ + 0.223, + 0.281, + 0.452, + 0.405 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.452, + 0.269, + 0.472, + 0.281 + ], + "angle": 0, + "content": "(d)" + }, + { + "type": "image", + "bbox": [ + 0.453, + 0.28, + 0.633, + 0.393 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.653, + 0.282, + 0.774, + 0.292 + ], + "angle": 0, + "content": "YOLOv8 Object Detection" + }, + { + "type": "image", + "bbox": [ + 0.64, + 0.305, + 0.787, + 0.392 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.214, + 0.413, + 0.788, + 0.552 + ], + "angle": 0, + "content": "Fig. 1: Photon inhibition for resource-efficient passive SPAD imaging. (a) Unlike conventional CMOS cameras (CIS), the energy consumption in SPAD cameras increases with scene brightness, severely limiting the applicability of high resolution SPAD cameras in resource-constrained applications. (b) We expand the conventional imaging pipeline to incorporate \"inhibition\" that electronically enables or disables individual pixels to limit bandwidth and power consumption. Our method relies on lightweight mathematical operations called \"inhibition policies\" that update the inhibition patterns based on the history of photon detections. Inhibition policies can be optimized for image SNR or for downstream vision tasks.(c,d) Object detection, a high-level vision task, is successful even with a large fraction of photons inhibited." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.583, + 0.788, + 0.689 + ], + "angle": 0, + "content": "microscopy [49], but more widely for consumer photography [37]. SPAD-based SPCs have recently been fabricated into kilo-to-megapixel format arrays that are now commercially available [28, 37]. The extreme sensitivity and high speed can benefit passive low-light computer vision tasks, particularly in the presence of rapid scene or camera motion [35], enable wide dynamic-range imaging [14, 31] and photon-starved active imaging applications such as 3D imaging (LiDAR) [21] and fluorescence microscopy [49]." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.69, + 0.789, + 0.841 + ], + "angle": 0, + "content": "SPAD camera pixels detect individual photons with extremely high frame rates by exploiting avalanche multiplication. On one hand, being able to detect individual photons opens up new possibilities and capabilities for computer vision systems. On the other hand, this also presents a unique challenge: Every photon-induced avalanche comes with a non-negligible energy cost, which is a challenge exclusive to SPAD-based cameras. This flux-dependent photon detection power is a significant fraction of total power consumption in today's SPAD cameras and impedes further increases in their spatial resolution [24,37,41,42,47]. For example, extrapolating the avalanche power of a recent SPAD sensor [47] to \\(\\sim 10\\) 's of megapixels format predicts a power consumption of several watts in" + } + ], + [ + { + "type": "header", + "bbox": [ + 0.322, + 0.115, + 0.732, + 0.129 + ], + "angle": 0, + "content": "Photon Inhibition for Energy-Efficient Single-Photon Imaging" + }, + { + "type": "page_number", + "bbox": [ + 0.775, + 0.117, + 0.785, + 0.127 + ], + "angle": 0, + "content": "3" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.146, + 0.784, + 0.177 + ], + "angle": 0, + "content": "bright light\\(^{4}\\) (as illustrated in Fig. 1(a)), far exceeding the power consumption of modern CMOS image sensors (CIS) of around \\(300\\mathrm{mW}\\) [45]." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.179, + 0.788, + 0.375 + ], + "angle": 0, + "content": "We address this problem by being selective about which photons are detected on the SPAD sensor while maintaining good performance on various computer vision tasks. To reduce avalanche power consumption and decouple it from photon flux, we propose a computational imaging technique called photon inhibition where individual SPAD pixels are disabled adaptively based on previous photon detections in their spatio-temporal neighborhoods. Electrically disabling pixels prevents photon detections to inhibit any avalanche power or subsequent processing. We implement lightweight on-sensor computations called inhibition policies (Fig. 1(b)) to determine, in real-time and at single-pixel and single-frame granularity, which SPAD pixels to enable or disable. Our method is inspired by retinal pre-processing of the human visual system where retinal neurons aggregate photon information over small spatio-temporal neighborhoods to cause neighboring retinal cells to become less sensitive to incident light [12, 16].5" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.377, + 0.788, + 0.633 + ], + "angle": 0, + "content": "Given that SPADs introduce a new challenge of flux-dependent power consumption, we establish, from first principles, novel energy-aware imaging performance metrics for resource-constrained single-photon imaging. Based on these metrics, we design families of inhibition policies that distribute photon detections in space and time based on imaging / vision task goals and energy consumption constraints. A critical consideration in the design of inhibition policies stems from the observation that these policies are meant to control (enable / disable) the sensor (Fig. 1(b)). Therefore, it is essential for these policies to be extremely lightweight since these need to be implemented on sensor with very limited compute and memory resources. Furthermore, inhibition policies must execute at ultra-low latency to keep up with high-speed photon detections (reaching up to \\(100\\mathrm{kHz}\\)). Fortunately, since the raw data output from a SPAD-based SPC consists of binary-valued image frames, SPCs are naturally suited to real-time calculations at the image sensor plane under tight compute and memory budgets. The proposed inhibition policies are lightweight, requiring only simple arithmetic and Boolean operations computed over local spatio-temporal neighborhoods, and thus amenable to in-pixel implementation [2, 7]." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.635, + 0.788, + 0.742 + ], + "angle": 0, + "content": "In simulations and real experimental data, we show that our inhibition policies allocate photon detections to sensor pixels in a way that reduces detection energy for a given accuracy level of various vision tasks. Our results show consequential energy savings when compared to a conventional capture scheme for tasks of (i) image reconstruction: \\(42\\%\\) fewer photon detections at equal image quality; (ii) edge detection: an edge sensitive inhibition policy reduces detections by \\(30\\%\\) at equal F-score; and (iii) YOLOv8 object detection: remains success-" + }, + { + "type": "page_footnote", + "bbox": [ + 0.218, + 0.756, + 0.787, + 0.812 + ], + "angle": 0, + "content": "It has been shown, perhaps counter-intuitively, that SPADs do not saturate even under extremely bright conditions [26, 27]. Therefore, SPADs are not restricted to low-flux environments, but are being considered for vision applications across a wide dynamic range of lighting conditions (e.g., from a dark tunnel to bright sunlight)." + }, + { + "type": "page_footnote", + "bbox": [ + 0.222, + 0.812, + 0.787, + 0.84 + ], + "angle": 0, + "content": "5 We borrow the term \"inhibition\" from the phenomenon of \"lateral inhibition\" found in biological vision systems [3]." + }, + { + "type": "list", + "bbox": [ + 0.218, + 0.756, + 0.787, + 0.84 + ], + "angle": 0, + "content": null + } + ], + [ + { + "type": "page_number", + "bbox": [ + 0.218, + 0.116, + 0.23, + 0.127 + ], + "angle": 0, + "content": "4" + }, + { + "type": "header", + "bbox": [ + 0.272, + 0.115, + 0.384, + 0.127 + ], + "angle": 0, + "content": "L. Koerner et al." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.146, + 0.784, + 0.221 + ], + "angle": 0, + "content": "ful with \\(95\\%\\) of photons inhibited under camera motion of a real-world SPC (Fig. 1(c,d)). Through experiments with photon streams captured using real SPAD camera hardware over a wide range of illumination conditions, we show that our proposed inhibition policies preserve low-light details and, in brightlight, decouple flux and detection energy." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.226, + 0.785, + 0.361 + ], + "angle": 0, + "content": "Scope and Limitations: There are several competing image sensor technologies today that resolve single-photon while capturing binary-valued raw frames at rates exceeding thousands of frames/second. SPCs based on \"jots\" [32] that do not rely on avalanche multiplication do not suffer from flux-dependent power consumption as SPADs. Jots are an exciting technology, especially in scenarios that require high resolution and high dynamic range imaging [33]. In this paper, we focus on SPADs, due to their rapidly rising availability and commercial interest [6, 28], and benefits over conventional CMOS sensors, both in low-light and bright scenes for a variety of computer vision tasks [10, 35]." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.362, + 0.787, + 0.483 + ], + "angle": 0, + "content": "Our goals in this paper are to (a) raise the question of \"which photons should be detected?\" in the context of energy-efficient single-photon imaging, (b) establish a design space and metrics to evaluate various inhibition policies, and (c) propose plausible policies that respect practical hardware limitations for future on-chip implementation. We emphasize that the inhibition policies proposed in this paper are not necessarily optimal. This work is just a first step towards demonstrating that it is possible to achieve high performance with SPADs, while maintaining low power consumption via photon inhibition." + }, + { + "type": "title", + "bbox": [ + 0.216, + 0.506, + 0.388, + 0.523 + ], + "angle": 0, + "content": "2 Related Work" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.538, + 0.785, + 0.719 + ], + "angle": 0, + "content": "Hardware approaches for reduced energy consumption: There is a strong dependence of SPAD power consumption on the pixel size — the smaller the pixel, the lower the avalanche energy [39]. Although recent developments in SPAD pixel technology have reduced pixel sizes to below \\(4\\mu \\mathrm{m}\\) [40], avalanche energy still contributes a large fraction of the total power consumption in a SPAD sensor [44]. SPAD design optimizations have reduced the charge per avalanche by RF modulation of the bias voltage [52], minimization of the junction capacitance [42], and smart [5] and fast [54] quenching circuits. Circuit architectures may require spatial and/or temporal co-incidence [22] to reduce energy downstream in the processing chain, but avalanche energy remains. Our work complements existing hardware approaches by preventing avalanches altogether to reduce illumination-dependent energy consumption." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.72, + 0.787, + 0.84 + ], + "angle": 0, + "content": "SPAD dead-time and clocked recharge: SPADs require recharge after an avalanche-inducing photon detection during which subsequent photons are not recorded. This dead-time inhibits photons at high exposure [24, 27], yet, power consumption remains excessive when the average inter-photon arrival interval is shorter than the SPAD dead time [38, 42]. Power consumption at high photon flux can be reduced by controlling the global rate of SPAD recharge (\"clocked recharge\") [38, 47]. Within a single recharge period a SPAD detects at most a single photon; subsequent photon arrivals do not induce an avalanche, thus power is" + } + ], + [ + { + "type": "header", + "bbox": [ + 0.321, + 0.115, + 0.733, + 0.131 + ], + "angle": 0, + "content": "Photon Inhibition for Energy-Efficient Single-Photon Imaging" + }, + { + "type": "page_number", + "bbox": [ + 0.775, + 0.117, + 0.785, + 0.127 + ], + "angle": 0, + "content": "5" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.147, + 0.788, + 0.298 + ], + "angle": 0, + "content": "reduced. To further limit avalanches, clocked recharge has been combined with a limit on the number of detections and time to saturation circuitry to measure the intensity of the saturated pixels [41], coarse pixel-wise exposure control [42], and a sequence of recharge periods similar to exposure bracketing [47]. These methods could be considered special cases of inhibition policies that are global and do not adapt to scene content, and therefore, are limited in flexibility to trade power and measurement fidelity. Consequently, in these methods, the power of avalanches remains a considerable fraction of total SPAD sensor power consumption [47]. For example, in this work we show considerable power savings, reaching up to \\(90\\%\\), for an object detection task." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.299, + 0.788, + 0.466 + ], + "angle": 0, + "content": "Resource-aware imaging: Event-based vision sensors reduce power consumption by only transmitting scene information when an intensity change is detected [17]. This idea has recently been applied to SPAD arrays to reduce power consumption due to data transfer [11, 46]. In contrast, our method reduces power due to detection by selectively disabling photodiodes based on photon history over small spatial and temporal neighborhoods. Miniaturized cameras with constraints on compute energy have transferred processing to the optical domain [30]. While we focus on passive imaging, depth sensing with SPADs and an active pulsed light source has related constraints such as acquisition time and laser power. Adaptive gating reduces acquisition time [43] and optimal allocation of the laser dwell time among pixels improves data quality [36, 48]." + }, + { + "type": "title", + "bbox": [ + 0.216, + 0.492, + 0.439, + 0.509 + ], + "angle": 0, + "content": "3 Observation Model" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.528, + 0.788, + 0.697 + ], + "angle": 0, + "content": "During an exposure time \\( T \\), a photon flux of \\( \\phi \\) results in an average number of photon conversions, or exposure, of \\( H = \\phi T \\) (we fold the sensor's photon detection probability, or PDP, into the definition of \\( \\phi \\), meaning it represents an effective flux). The distribution of photon conversions, \\( K \\), follows the Poisson distribution given by \\( \\mathsf{P}(K = k;H) = \\frac{H^k e^{-H}}{k!} \\). During each binary exposure period, a SPAD pixel records a '1' if at least one photon was incident during that period, and '0' otherwise. The probability of detecting at least one photon is given by \\( Y := 1 - \\mathsf{P}(K = 0;H) = 1 - e^{-H} \\). Hence, the SPAD pixel readout in each binary frame is a Bernoulli random variable \\( B \\sim \\mathrm{Bernoulli}(Y) \\). Multiple exposure time windows, or measurements \\( (W) \\), are recorded to reduce noise with the total number of detections" + }, + { + "type": "equation", + "bbox": [ + 0.358, + 0.707, + 0.787, + 0.75 + ], + "angle": 0, + "content": "\\[\nD := \\left[ \\sum_ {n = 1} ^ {W} B _ {n} \\right] \\sim \\operatorname {B i n o m i a l} (W, 1 - e ^ {- H}). \\tag {1}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.216, + 0.76, + 0.755, + 0.775 + ], + "angle": 0, + "content": "We estimate the probability of detection and flux from a measurement \\( D \\) as" + }, + { + "type": "equation", + "bbox": [ + 0.441, + 0.788, + 0.786, + 0.816 + ], + "angle": 0, + "content": "\\[\n\\widehat {Y} = \\frac {D}{W} \\tag {2}\n\\]" + }, + { + "type": "equation", + "bbox": [ + 0.441, + 0.818, + 0.786, + 0.838 + ], + "angle": 0, + "content": "\\[\n\\widehat {H} = - \\ln (1 - \\widehat {Y}). \\qquad \\qquad (3)\n\\]" + } + ], + [ + { + "type": "page_number", + "bbox": [ + 0.218, + 0.116, + 0.23, + 0.127 + ], + "angle": 0, + "content": "6" + }, + { + "type": "header", + "bbox": [ + 0.272, + 0.115, + 0.384, + 0.128 + ], + "angle": 0, + "content": "L. Koerner et al." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.147, + 0.788, + 0.225 + ], + "angle": 0, + "content": "Changes under inhibition: Inhibition is represented by a binary state variable \\( M \\) at every pixel, with \\( M_{n} = 1 \\) denoting enabled for the \\( n \\)-th measurement period. \\( M \\) is a random variable when an adaptive or data-dependent inhibition policy is used. The total number of measurements changes to \\( W_{inh.} \\coloneqq \\sum_{n=1}^{W} M_{n} \\leq W \\). The number of detections becomes" + }, + { + "type": "equation", + "bbox": [ + 0.271, + 0.233, + 0.786, + 0.289 + ], + "angle": 0, + "content": "\\[\nD _ {i n h.} := \\left[ \\sum_ {n = 1} ^ {W} M _ {n} B _ {n} \\right] = \\left[ \\sum_ {\\substack {n = 1 \\\\ M _ {n} = 1}} ^ {W} B _ {n} \\right] \\sim \\operatorname {B i n o m i a l} \\left(W _ {i n h.}, 1 - e ^ {- H}\\right), \\tag{4}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.299, + 0.789, + 0.436 + ], + "angle": 0, + "content": "and the flux is estimated similar to Eqs. (2) and (3). The second summation conveys that, with inhibition, the measurements when the pixel is enabled match the original model in Eq. (1) - a result of the memoryless property of the Poisson arrival process. The model above requires that transitions in \\( M \\) are synchronized with the clock signal used to gate the exposure, so that the PDP is not changed by inhibition. A second assumption is that the time for SPAD recharge is small relative to the clock period. This is a desired property for passive SPAD-based imaging, and holds for many state-of-the-art SPADs [42, 44, 47]. It implies that PDP is approximately constant in time and does not depend on prior pixel state." + }, + { + "type": "title", + "bbox": [ + 0.216, + 0.457, + 0.6, + 0.475 + ], + "angle": 0, + "content": "4 Energy-Aware Performance Metrics" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.488, + 0.789, + 0.557 + ], + "angle": 0, + "content": "The exposure-referred signal-to-noise ratio \\((\\mathsf{SNR}_H)\\) is commonly used to evaluate single-photon sensor performance, and can be computed as the ratio of the true exposure \\(H\\) and the root-mean-squared error in the estimated exposure \\(\\sqrt{\\mathbb{E}[(\\widehat{H} - H)^2]}\\) [15, 55]:" + }, + { + "type": "equation", + "bbox": [ + 0.359, + 0.568, + 0.787, + 0.612 + ], + "angle": 0, + "content": "\\[\n\\mathsf {S N R} _ {H} = \\frac {H}{\\sqrt {\\mathbb {E} [ (\\widehat {H} - H) ^ {2} ]}} = H \\sqrt {\\frac {W}{e ^ {H} - 1}}. \\tag {5}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.62, + 0.788, + 0.667 + ], + "angle": 0, + "content": "At low incident flux, \\(\\mathsf{SNR}_H\\) is low due to shot noise. \\(\\mathsf{SNR}_H\\) improves as the likelihood of a photon detection increases until, in bright light with \\(H > 1.6\\), the \\(\\mathsf{SNR}_H\\) degrades due to \"soft\" saturation of the response [27,34]." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.667, + 0.788, + 0.728 + ], + "angle": 0, + "content": "We propose two new energy-aware modifications to \\(\\mathsf{SNR}_H\\) to incorporate SPAD energy costs. First, we propose a detection efficiency metric \\(\\mathsf{SNR}_{H / D}^2\\) defined as the square of SNR normalized by the expected number of detections \\(\\mathbb{E}[D] = W(1 - e^{-H})\\):" + }, + { + "type": "equation", + "bbox": [ + 0.381, + 0.737, + 0.786, + 0.772 + ], + "angle": 0, + "content": "\\[\n\\mathsf {S N R} _ {H / D} ^ {2} := \\frac {\\mathsf {S N R} _ {H} ^ {2}}{\\mathbb {E} [ D ]} = \\frac {H ^ {2} e ^ {- H}}{(1 - e ^ {- H}) ^ {2}}. \\tag {6}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.78, + 0.789, + 0.843 + ], + "angle": 0, + "content": "Fig. 3(a) shows the \\(\\mathsf{SNR}_H\\) (black) and the detection efficiency (red) versus the average photon arrivals per period \\((H)\\). When \\(H \\ll 1\\) all detections contribute significant information and \\(\\mathsf{SNR}_H^2\\) increases linearly, similar to an ideal non-saturating sensor only limited by Poisson noise. Accordingly, \\(\\mathsf{SNR}_{H / D}^2 \\cong 1\\), the" + } + ], + [ + { + "type": "header", + "bbox": [ + 0.321, + 0.115, + 0.733, + 0.131 + ], + "angle": 0, + "content": "Photon Inhibition for Energy-Efficient Single-Photon Imaging" + }, + { + "type": "page_number", + "bbox": [ + 0.775, + 0.116, + 0.785, + 0.127 + ], + "angle": 0, + "content": "7" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.147, + 0.784, + 0.192 + ], + "angle": 0, + "content": "upper bound of this metric. At larger exposure values, beginning around \\( H \\approx 0.5 \\), the sensor begins to saturate which slows the growth of \\( \\mathrm{SNR}_H \\) and thus degrades the detection efficiency." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.193, + 0.785, + 0.283 + ], + "angle": 0, + "content": "A separate constraint is the total number of recharge periods during which a pixel is enabled and can measure either '0' or '1'. This number of measurements \\((W)\\) may be limited due to the energy to read out a frame, the depth of an in-pixel counter, and/or the maximum allowable sensing latency due to motion blur. We establish a second metric, measurement efficiency defined as the square of SNR normalized by the number of measurement windows:" + }, + { + "type": "equation", + "bbox": [ + 0.384, + 0.293, + 0.786, + 0.328 + ], + "angle": 0, + "content": "\\[\n\\mathrm {S N R} _ {H / W} ^ {2} := \\frac {\\mathrm {S N R} _ {H} ^ {2}}{W} = \\frac {H ^ {2} e ^ {- H}}{1 - e ^ {- H}}. \\tag {7}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.336, + 0.784, + 0.367 + ], + "angle": 0, + "content": "Fig. 3(a) (blue) shows sub-optimal measurement efficiency at both low and high exposures with the best efficiency at \\( H = 1.59 \\), \\( Y = 0.80 \\), as demonstrated in [8]." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.368, + 0.785, + 0.458 + ], + "angle": 0, + "content": "In an oracle setting with a known image, one can analytically derive nonuniform allocations of measurements \\( W_{i} \\) across the pixels \\( i \\), constraining the total expected detections to a fixed value and optimizing \\( \\mathsf{SNR}_H \\) or mean-squared-error (see the supplement). The loss metric may also be defined relative to the binary rate \\( Y \\) instead of \\( H \\) as above. In that case a useful base metric could be entropy instead of \\( \\mathsf{SNR}_H \\) [19]." + }, + { + "type": "title", + "bbox": [ + 0.215, + 0.481, + 0.591, + 0.499 + ], + "angle": 0, + "content": "5 Spatio-temporal Inhibition Policies" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.512, + 0.788, + 0.633 + ], + "angle": 0, + "content": "We now propose policies that calculate a spatio-temporal inhibition pattern for each pixel and each frame based on the history of photon frames and patterns. Following Sec. 3, we define an inhibition pattern using a binary-valued tensor \\( M \\) where \\( M(i,j,t) = 0 \\) if pixel \\( (i,j) \\) is disabled in the \\( t^{\\mathrm{th}} \\) frame, and \\( M(i,j,t) = 1 \\) otherwise. All pixels are initially enabled, and on-sensor calculations modify \\( M \\) over time. The binary photon cube is defined as \\( F(i,j,t) = 1 \\) if the pixel is enabled (i.e., \\( M(i,j,t) = 1 \\)) and a photon is detected at pixel location \\( (i,j) \\) in the \\( t^{\\mathrm{th}} \\) frame, otherwise it is zero." + }, + { + "type": "image", + "bbox": [ + 0.292, + 0.664, + 0.709, + 0.757 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.214, + 0.766, + 0.788, + 0.809 + ], + "angle": 0, + "content": "Fig. 2: Calculation-based inhibition with dead time. Arrows represent photons with an 'X' for inhibition. \\( T \\) indicates the clocked recharge period. A score, \\( S \\), calculated from past frames determines if future measurements are enabled or disabled." + } + ], + [ + { + "type": "page_number", + "bbox": [ + 0.218, + 0.116, + 0.23, + 0.127 + ], + "angle": 0, + "content": "8" + }, + { + "type": "header", + "bbox": [ + 0.272, + 0.115, + 0.384, + 0.127 + ], + "angle": 0, + "content": "L. Koerner et al." + }, + { + "type": "image", + "bbox": [ + 0.217, + 0.145, + 0.422, + 0.256 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.424, + 0.153, + 0.788, + 0.258 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.214, + 0.267, + 0.789, + 0.462 + ], + "angle": 0, + "content": "Fig. 3: Efficiency metrics and inhibition policies that track the metrics: (a) The \\(\\mathsf{SNR}_H\\) in dB (black), the detection efficiency (red, --), and the measurement efficiency (blue, --) versus the exposure with \\(W = 100\\) measurements. The binary rate \\(Y = 1 - e^{-H}\\) is indicated on the top axis. The vertical dotted line indicates the exposure and the binary rate at which the \\(\\mathsf{SNR}_H\\) degrades by \\(3\\mathrm{dB}\\) from the peak \\(\\mathsf{SNR}_H\\). (b,c) Monte Carlo simulations of binary images using calculation-based inhibition policies demonstrate how the allocation of measurements versus the pixel exposure depends upon tuning parameters. (b) the inhibition threshold \\(\\eta\\) adjusts the exposure level at which pixels are inhibited to allow the measurement fraction (the ratio of active measurements to total number of frames) to follow the \\(\\mathsf{SNR}_{H / D}^2\\) curve in (a). A smaller threshold more aggressively inhibits photons. (c) demonstrates the impact of the hold-off time, \\(\\tau_{H}\\), on the number of measurements allocated to the brightest pixels. The legend indicates the total fraction of photons inhibited as \\(I_F\\). (b,c) show smoothed curves (Lowess filter, fraction of \\(1 / 5\\)) of the measurement fraction vs. H." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.491, + 0.788, + 0.627 + ], + "angle": 0, + "content": "Fig. 1(b) shows the components of a photon inhibition processing layer, including the binary frames, \\( F \\), and inhibition pattern \\( M \\). For ease of on-sensor implementation, we focus on policies that operate on small and local spatiotemporal neighborhoods of fixed sizes. We rely on local arithmetic and Boolean computations and comparison operations, consistent with current in-pixel computational capabilities [2,9,24]. Fig. 2 shows a proposed on-sensor calculation approach that operates in a streaming fashion as frames accumulate to calculate an inhibition score, \\( S \\), as the result of a spatio-temporal filter of the binary frames and inhibition pattern. The score at each pixel is calculated as" + }, + { + "type": "equation", + "bbox": [ + 0.348, + 0.638, + 0.786, + 0.656 + ], + "angle": 0, + "content": "\\[\nS (i, j, t) = K * \\left[ \\left(2 F (i, j, t) - \\mathbf {1}\\right) \\cdot M (i, j, t) \\right] \\tag {8}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.666, + 0.789, + 0.844 + ], + "angle": 0, + "content": "which applies a spatio-temporal filtering kernel, \\( K \\), of dimensions \\( L, H, T \\) to a ternary representation of the pixel result (1, 0, or -1 for a detection, a disabled pixel, or a measurement that does not detect a photon, respectively). The kernel \\( K \\) can typically be separated into spatial and temporal components as \\( K = K_{s} \\otimes K_{t} \\) with dimensions \\( L \\times H \\times 1 \\) and \\( 1 \\times 1 \\times T \\), respectively. After each binary frame, the score is compared to a threshold \\( \\eta \\) and the pixel is disabled for the subsequent \\( \\tau_{H} \\) frames: \\( M(i,j,t') = 0 \\) for \\( \\{t'|t + 1 \\leq t' \\leq t + 1 + \\tau_{H}\\} \\) if \\( S(i,j,t) > \\eta \\). Observe in Fig. 3(b,c) that decreasing \\( \\eta \\) and increasing \\( \\tau_{H} \\) can be used to attain more aggressive inhibition with a larger fraction of photons being inhibited. The binary rate of each pixel is estimated (using Eq. (4)) as the ratio of detections to (active) measurements: \\( \\widehat{Y}(i,j,t) = \\sum_{t'} F(i,j,t') / \\sum_{t'} M(i,j,t') \\). This" + } + ], + [ + { + "type": "header", + "bbox": [ + 0.322, + 0.115, + 0.733, + 0.13 + ], + "angle": 0, + "content": "Photon Inhibition for Energy-Efficient Single-Photon Imaging" + }, + { + "type": "page_number", + "bbox": [ + 0.775, + 0.117, + 0.785, + 0.127 + ], + "angle": 0, + "content": "9" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.147, + 0.785, + 0.193 + ], + "angle": 0, + "content": "calculation requires a record of the inhibition history which could be accumulated by an in-pixel counter or recreated in a downstream processor if all binary frames are read out." + }, + { + "type": "image", + "bbox": [ + 0.296, + 0.221, + 0.705, + 0.31 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.214, + 0.318, + 0.408, + 0.412 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.419, + 0.329, + 0.59, + 0.412 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.597, + 0.32, + 0.789, + 0.411 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.218, + 0.42, + 0.789, + 0.464 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.214, + 0.474, + 0.788, + 0.613 + ], + "angle": 0, + "content": "Fig. 4: Saturation look-ahead inhibition. (a) When detections \\( D_{i} \\) of cycle \\( i \\) exceed a threshold, \\( d_{i} \\), the photons arriving during the next recharge cycle, \\( i + 1 \\), are inhibited. For this drawing \\( [d_{1}, d_{2}] = [3, 3] \\). (b, c, d) Example with three recharge period settings and brackets of \\( W_{i} = 10 \\) measurements each. The saturation thresholds are set as \\( [d_{1}, d_{2}] = [7, 7] \\). The brackets are combined linearly, weighted by \\( \\mathsf{SNR}_H^2 \\) [18]. (b, c) With negligible effect on reconstructed SNR relative to conventional clocked-recharge, significantly fewer photon detections occur under the inhibition policy in high-flux settings. (d) The individual brackets are effectively disabled in flux regimes with low detection efficiency, and the detection rate closely tracks the original \\( \\mathsf{SNR}_H^2 \\) curve. Simulation details are in the supplement." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.644, + 0.789, + 0.842 + ], + "angle": 0, + "content": "Example policies: 1) Single-pixel dead time: A spatial kernel \\( K_{s} \\) of dimensions \\( 1 \\times 1 \\) may mimic the dead time of a passive SPAD without influence from neighboring pixels. Unlike SPAD recharge generated dead time, the temporal kernel may be extended (e.g., \\( N \\) frames long) to establish a rate threshold for inhibition with reduced quantization noise. 2) Local spatio-temporal averaging: Single binary frames are inherently noisy. An inhibition policy that calculates spatiotemporal averages to estimate the local photon rate may reduce the impact of noise on the inhibition pattern. Another benefit, as shown in Fig. 3(b,c), is that the stochastic nature of binary frames smoothes the distribution of measurements versus pixel flux. Discontinuities (\"dips\") in SNR versus photon flux are undesirable due to the potential for artifacts. 3) Edge enhancement: Pixels may be inhibited if a local neighborhood has little spatial variation in photon rate. This can be achieved, for instance, through a spatial filter \\( K_{s} \\) in Eq. (8) which acts" + } + ], + [ + { + "type": "page_number", + "bbox": [ + 0.218, + 0.116, + 0.236, + 0.127 + ], + "angle": 0, + "content": "10" + }, + { + "type": "header", + "bbox": [ + 0.272, + 0.115, + 0.384, + 0.127 + ], + "angle": 0, + "content": "L. Koerner et al." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.147, + 0.784, + 0.177 + ], + "angle": 0, + "content": "like a Laplacian filter. Such a strategy may enhance the fidelity of edges in the image while focusing fewer resources on regions with constant illumination." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.178, + 0.788, + 0.359 + ], + "angle": 0, + "content": "Fig. 4 presents a second proposed inhibition policy called saturation lookahead. This policy combines exposure brackets and calculation-based inhibition for a light-weight single-pixel inhibition policy. This policy proceeds as a sequence of cycles (indexed by \\( i \\)) of binary frames where each binary frame within each cycle uses the same exposure time. Cycle exposure times \\( T_{i} \\) progressively increase (\\( T_{1} < T_{2} < T_{3} \\ldots \\)) so that measurements taken in an earlier cycle may predict low detection efficiency (near saturation) at longer exposure times and disable the pixel in these subsequent cycles. The exposure level thresholds for inhibition would typically be set such that the number of photons detected at a given flux level tracks the \\( \\mathsf{SNR}_H^2 \\) (see Fig. 4(b)), but may be adjusted further based on the relative importance of power consumption, sensing latency, and SNR in an actual application setting." + }, + { + "type": "title", + "bbox": [ + 0.216, + 0.38, + 0.527, + 0.397 + ], + "angle": 0, + "content": "6 Simulation-based Evaluation" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.411, + 0.784, + 0.517 + ], + "angle": 0, + "content": "We use Monte Carlo simulations to generate sequences of grayscale binary frames from a dataset of RGB images [1]. The inhibition policies evaluated extend the baseline inhibition generated by clocked recharge. The inhibition score and patterns for various policies and tuning parameters are calculated from these binary frame sequences. Once inhibition patterns are determined, performance is evaluated by tabulating detections, measurements, and image quality or vision task performance for each step in the sequence (see the supplement for details)." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.52, + 0.788, + 0.716 + ], + "angle": 0, + "content": "Spatio-Temporal Policies for Imaging: Handcrafted spatial kernels (3x3) were combined with an averaging temporal kernel of length 4 to form lightweight inhibition policies that allocate pixel measurements as described in earlier sections for improved image reconstruction. Fig. 5 summarizes the simulation results. Fig. 5(d,e) display reduction in photon detections at equal structural similarity index measure (SSIM) [51], enabled by disabling bright pixels, for an exposure bracketing sequence. Intensity estimates from each bracket are combined using \\(\\mathsf{SNR}^2\\) weighting [18] and then converted to a binary rate estimate at the center exposure level of 1 ppp. The proposed policy demonstrates an average reduction in detections of \\(42\\%\\) as compared to no inhibition. Fig. 5(f,g) evaluates a single exposure level (1.0 ppp) which is a more challenging scenario, yet the proposed policy still reduces detections by \\(14\\%\\) at \\(\\mathrm{SSIM} = 0.7\\). See supplement S 3.4 for more examples and S 3.5 for simulations of high dynamic range images." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.719, + 0.788, + 0.842 + ], + "angle": 0, + "content": "Edge Detection: The BSDS500 dataset with ground truth boundaries [1] was used to study energy-efficient edge detection via photon inhibition. Binary rate images were processed by pre-trained holistically-nested edge detection (HED) [53] with the resulting edge maps compared to ground truth by the structured edge detection toolbox [13]. Fig. 6 shows the optimal image scale (OIS) F-score versus the average detections per pixel. Interpolated curves (not shown) allow for translating along horizontal lines of equal task performance to assess differences in avalanche energy. At low photon counts the proposed edge-enhancing policy" + } + ], + [ + { + "type": "header", + "bbox": [ + 0.322, + 0.115, + 0.733, + 0.13 + ], + "angle": 0, + "content": "Photon Inhibition for Energy-Efficient Single-Photon Imaging" + }, + { + "type": "page_number", + "bbox": [ + 0.768, + 0.117, + 0.784, + 0.127 + ], + "angle": 0, + "content": "11" + }, + { + "type": "image", + "bbox": [ + 0.219, + 0.147, + 0.412, + 0.266 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.411, + 0.149, + 0.603, + 0.266 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.218, + 0.268, + 0.348, + 0.397 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.358, + 0.269, + 0.505, + 0.398 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.51, + 0.269, + 0.617, + 0.402 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.612, + 0.149, + 0.782, + 0.249 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.619, + 0.249, + 0.783, + 0.393 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.214, + 0.417, + 0.788, + 0.597 + ], + "angle": 0, + "content": "Fig. 5: Power-efficient static imaging via inhibition. (a) Images from an exposure bracketing sequence (average exposure of 0.1, 1.0, 10.0 photons per pixel (ppp)) using clocked recharge without (conventional) and with inhibition. (b) the distribution of measurements based on pixel intensity with \\( H \\) at 1.0 ppp. (c) image quality (SSIM) at equal detections/pixel. (d-g) Handcrafted policies are tested over 20 images and assessed by reductions in detections (\\( D \\)) at equal SSIM. The top row (d),(e) use exposure bracketing; (f),(g) use a single exposure of 1.0 ppp. (d),(f) at SSIM=0.7 and (e),(g) at SSIM=0.8. The box shows quartiles with the center line at the median. The proposed policy, \\( P_{cr} \\), is a \\( 3 \\times 3 \\) spatial kernel that emphasizes the center pixel (\\( \\times 8 \\)) and includes the 8 neighbors (\\( \\times 1 \\)) (see the supplement for policy details). (h) average inhibition patterns for each exposure time. The top most pattern inhibits the brightest pixels only (maximum of \\( \\sim 60\\% \\) inhibition, primarily in the sky). For the longest exposure time the inhibition pattern allocates measurements primarily to the darkest pixels." + }, + { + "type": "image", + "bbox": [ + 0.365, + 0.609, + 0.641, + 0.72 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.214, + 0.732, + 0.788, + 0.803 + ], + "angle": 0, + "content": "Fig. 6: Energy-efficient edge detection by inhibition: Edge detection F-scores averaged over 19 images versus average detections/pixel. Horizontal arrows show reduction in detections enabled by inhibition at equal task performance. At and beyond \\(30\\mathrm{D / pix}\\). no inhibition and the proposed policy are nearly equivalent and plateau at \\(200\\mathrm{D / pix}\\). The edge detector returns \\(F = 0.813\\) on the original images of this set." + } + ], + [ + { + "type": "page_number", + "bbox": [ + 0.218, + 0.116, + 0.236, + 0.127 + ], + "angle": 0, + "content": "12" + }, + { + "type": "header", + "bbox": [ + 0.272, + 0.115, + 0.384, + 0.127 + ], + "angle": 0, + "content": "L. Koerner et al." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.147, + 0.788, + 0.253 + ], + "angle": 0, + "content": "demonstrates a \\(30\\%\\) reduction in detections on single exposure time captures. This policy extends the on-sensor calculation approach by calculating two scores, the conventional \\(3 \\times 3\\) Laplacian [50] as \\(S_{1}\\) and a \\(3 \\times 3\\) averaging filter as \\(S_{2}\\). The final inhibition decision is the Boolean operation of these scores as \\(((\\eta_{1} < S_{1} < \\eta_{2}) \\land (S_{2} > \\eta_{3})) \\lor (S_{2} > \\eta_{4})\\). The Laplacian policy alone performs poorly as photons in dim regions with minimal spatial variation are inhibited - disabling dim pixels is energy inefficient. Complete policy descriptions are in the supplement." + }, + { + "type": "title", + "bbox": [ + 0.216, + 0.278, + 0.57, + 0.296 + ], + "angle": 0, + "content": "7 Experiments on Dynamic Scenes" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.313, + 0.787, + 0.403 + ], + "angle": 0, + "content": "Many real-world scenes contain significant motion even with the high frame rate of a SPAD camera. Burst reconstruction algorithms yield high-quality images from sequences of binary frames [34,35] - we investigate their compatibility with the inhibited photon detection data. In particular, we focus on the saturation look-ahead policy of Fig. 4(a), applied independently at every pixel, and therefore an example of an adaptive single-pixel temporal policy." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.404, + 0.787, + 0.54 + ], + "angle": 0, + "content": "To avoid losing salient information under motion, the inhibition policy must limit periods of extended dead time. In the context of the saturation look-ahead policy this limits the total exposure length of the bracketing sequence, since we implicitly assume flux to be constant within each sequence. In our experiments, we use a Fibonacci bracketing [20] sequence \\( T \\coloneqq \\{1, 1, 2, 3, 5, 8, 13, 21\\}^6 \\), denoted in the units of a single minimal exposure time. Every measured bracket is converted to a maximum-likelihood estimate of flux, which is then supplied to the quanta burst photography algorithm [34]. The flux estimator from bracketed measurements is described in detail in supplement Sec. S7.2." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.544, + 0.787, + 0.649 + ], + "angle": 0, + "content": "Results with the SwissSPAD2 sensor: The SwissSPAD2 sensor [49] is a prototype SPAD pixel array that can produce binary frames at a rate up to 97,700 FPS, with a resolution of \\(512 \\times 256\\). In our experiments, we use binary frames captured directly (without inhibition) by the SPAD array as reference data and emulate on-sensor saturation look-ahead inhibition in software. As a pre-processing step, measurements at hot pixels are replaced with their nearest neighbors." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.65, + 0.787, + 0.801 + ], + "angle": 0, + "content": "Fig. 7 shows the results of burst reconstruction under three lighting conditions. The raw data is a sequence of \\(>580,000\\) binary frames with scene radiance increasing rapidly by orders of magnitude (Fig. 7(b)), from \\(<1\\) lux to \\(>4,000\\) lux, measured separately with a light meter. For each of 47 equally-spaced keyframes, centered windows of 12,000 binary frames are extracted and processed as described above. Results for the full sequence can be found in the supplementary material. A static inhibition policy of regular sub-sampling (dropping 9 out of every 10 frames) is also applied, which yields a fixed \\(90\\%\\) reduction in both measurements and photon detections under all lighting conditions. Other subsampling factors are discussed in the supplement." + }, + { + "type": "page_footnote", + "bbox": [ + 0.218, + 0.811, + 0.787, + 0.84 + ], + "angle": 0, + "content": "Sequence \\( T' \\coloneqq \\{1, 1, 1, 3, 3, 3, 8, 8, 25\\} \\) yielded similar results. No extensive search over the policy space was performed." + } + ], + [ + { + "type": "header", + "bbox": [ + 0.322, + 0.115, + 0.733, + 0.13 + ], + "angle": 0, + "content": "Photon Inhibition for Energy-Efficient Single-Photon Imaging" + }, + { + "type": "page_number", + "bbox": [ + 0.768, + 0.116, + 0.786, + 0.127 + ], + "angle": 0, + "content": "13" + }, + { + "type": "image", + "bbox": [ + 0.214, + 0.142, + 0.734, + 0.477 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.214, + 0.487, + 0.788, + 0.71 + ], + "angle": 0, + "content": "Fig. 7: Adaptive policies on video sequences enable stronger inhibition, preserve low-light details and, in bright-light, decouple flux and detection energy. (a) Burst reconstructions [34] for three keyframes with varying light levels: the top and bottom row differ by \\(\\approx 7\\) stops. [Images show the detection rates \\(Y\\) and are further gamma-compressed (\\(\\gamma = 0.4\\)). Complete results included in the supplement.] Left column results are from the original binary frames without inhibition, and the right column after sub-sampling \\(10 \\times\\) (a fixed \\(90\\%\\) inhibition). Middle column represents exposure bracketing combined with saturation look-ahead (see Fig. 4a for description). Under strong light (top row) the results are reasonable with both methods. However, plain sub-sampling loses details in lower light: notice the furniture and a person's outline in the middle & bottom rows, respectively. Inhibition is instead adaptive to flux. (b) Average exposure level for each keyframe in the sequence. (c,d) Per-keyframe and cumulative detection counts - inhibition ultimately results in fewer photons being detected over the whole sequence. (e) Number of measurements taken for each keyframe; reductions may be translated to energy savings during read-out. Plots in (c,d,e) are sub-sampled for clarity, and crossover points are marked by green arrows." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.749, + 0.788, + 0.84 + ], + "angle": 0, + "content": "The top row in Fig. 7(a) shows that under strong light a large fraction of photons (\\(>90\\%\\)) can be inhibited through the saturation look-ahead policy and still result in good image quality after burst reconstruction, thus spreading photon detections over a longer period of time to reduce avalanche power. Even simple sub-sampling yields good results in bright light, and may work well under controlled illumination conditions. However, the images in the middle and" + } + ], + [ + { + "type": "page_number", + "bbox": [ + 0.218, + 0.116, + 0.236, + 0.127 + ], + "angle": 0, + "content": "14" + }, + { + "type": "header", + "bbox": [ + 0.272, + 0.115, + 0.384, + 0.127 + ], + "angle": 0, + "content": "L. Koerner et al." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.147, + 0.788, + 0.298 + ], + "angle": 0, + "content": "bottom rows illustrate that this static inhibition policy results in excessive signal loss in lower light leading to a loss of details. Sub-sampling may be enhanced by adjusting exposure time and/or the sub-sampling factor in response to global flux but cannot simultaneously optimize for different light levels within a single frame: over-exposed regions may have clipping artifacts and under-exposed regions are prone to motion blur from incorrect burst reconstruction. The saturation look-ahead policy does adapt to local flux, and allocates relatively more detections to dim regions (see Fig. 1(d)). It thus decouples detection energy from flux (Fig. 7(c); cf. Fig. 1(a)), and results in cumulatively fewer detections than sub-sampling (Fig. 7(d)) due to it being more aggressive in strong light." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.299, + 0.788, + 0.375 + ], + "angle": 0, + "content": "Apart from photography, the burst-reconstructed images can be used in computer vision applications, where even stronger inhibition is possible depending on the noise- or blur-robustness of the vision algorithm. Fig. 1(d) shows successful object detection with the YOLOv8 algorithm [29] applied to a burst-reconstructed image, with approximately \\(95\\%\\) photon detections inhibited." + }, + { + "type": "title", + "bbox": [ + 0.215, + 0.399, + 0.563, + 0.416 + ], + "angle": 0, + "content": "8 Limitations and Future Outlook" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.432, + 0.788, + 0.584 + ], + "angle": 0, + "content": "Implementation costs. Our focus in this paper is on reducing energy consumption due to avalanches. Although avalanches contribute significant energy as compared to on-sensor computations [2, 42, 47], an important next step is to design a holistic model that includes the energy consumption of computations and readout (leveraging Eq. 7). Our in-pixel computations — power-of-two multiplications which simplify to bit shifts, small spatio-temporal kernels no larger than \\(3 \\times 3 \\times 4\\) — are designed to be lightweight. Fortunately, computational SPAD imagers [2] with in-pixel memory and compute have recently been proposed, with \\(4 \\times 4\\) block of pixels having a 32-bit CPU and over 10 kbits of memory. See the supplement for an estimate of the required circuitry." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.588, + 0.788, + 0.739 + ], + "angle": 0, + "content": "More complete models. The noise model used for efficiency metrics only accounts for photon Poisson noise and the quantization noise of Bernoulli samples. Expanding the noise model to include uncorrected pixel sensitivity variations, crosstalk, and afterpulsing may improve performance [4] by allocating measurements with an awareness of practical sensor limitations. Inhibition modulates the number of Bernoulli trials. An unbiased estimator for data-dependent stopping of Bernoulli trials is known [23], yet is not applied in this work since it does not precisely match our situation. Our simulations (SSIM) suggest that bias is less significant than image noise, yet future analytical studies are needed over a range of conditions." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.743, + 0.788, + 0.835 + ], + "angle": 0, + "content": "Generalization to other tasks. To generalize the proposed approaches to a variety of vision tasks, task-specific quality metrics must be defined to pose optimization problems for each task. As an example, for the image reconstruction task, Suppl. Sec. 3.1 shows an analytical optimization using the MSE metric when constrained by photon detections. Other tasks could be approached similarly, yet may need to be optimized empirically." + } + ], + [ + { + "type": "header", + "bbox": [ + 0.322, + 0.115, + 0.732, + 0.129 + ], + "angle": 0, + "content": "Photon Inhibition for Energy-Efficient Single-Photon Imaging" + }, + { + "type": "page_number", + "bbox": [ + 0.769, + 0.117, + 0.785, + 0.127 + ], + "angle": 0, + "content": "15" + }, + { + "type": "title", + "bbox": [ + 0.218, + 0.145, + 0.393, + 0.163 + ], + "angle": 0, + "content": "Acknowledgments" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.181, + 0.788, + 0.258 + ], + "angle": 0, + "content": "The authors acknowledge the Minnesota Supercomputing Institute (MSI) at the University of Minnesota for providing resources. A.I. was supported in part by NSF ECCS-2138471. S.G and M.G. were supported in part by NSF CAREER award 1943149, NSF award CNS-2107060, and Wisconsin Alumni Research Foundation (WARF)." + }, + { + "type": "title", + "bbox": [ + 0.216, + 0.284, + 0.323, + 0.3 + ], + "angle": 0, + "content": "References" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.318, + 0.787, + 0.362 + ], + "angle": 0, + "content": "1. Arbeláez, P., Maire, M., Fowlkes, C., Malik, J.: Contour Detection and Hierarchical Image Segmentation. IEEE Transactions on Pattern Analysis and Machine Intelligence 33(5), 898-916 (May 2011). https://doi.org/10.1109/TPAMI.2010.161" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.363, + 0.787, + 0.39 + ], + "angle": 0, + "content": "2. Ardelean, A.: Computational Imaging SPAD Cameras. Ph.D. thesis, EPFL (2023). https://doi.org/10.5075/epfl-thesis-9501" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.391, + 0.787, + 0.418 + ], + "angle": 0, + "content": "3. Barlow, H.B.: Summation and inhibition in the frog's retina. The Journal of physiology 119(1), 69 (1953)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.42, + 0.787, + 0.473 + ], + "angle": 0, + "content": "4. Bian, L., Song, H., Peng, L., Chang, X., Yang, X., Horstmeyer, R., Ye, L., Zhu, C., Qin, T., Zheng, D., Zhang, J.: High-resolution single-photon imaging with physics-informed deep learning. Nature Communications 14(1), 5902 (Sep 2023). https://doi.org/10.1038/s41467-023-41597-9" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.476, + 0.787, + 0.517 + ], + "angle": 0, + "content": "5. Boso, G., Buttafava, M., Villa, F., Tosi, A.: Low-cost and compact single-photon counter based on a CMOS SPAD smart pixel. IEEE Photonics Technology Letters 27(23), 2504-2507 (2015)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.518, + 0.787, + 0.572 + ], + "angle": 0, + "content": "6. Canon Inc.: Canon Launches MS-500 - The World's First Ultra-High-Sensitivity Interchangeable-Lens SPAD Sensor Camera. https://www.usa.canon.com/newsroom/2023/20230801-ms500, Canon Press Release 8/1/2023. Accessed 2/25/2024." + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.574, + 0.787, + 0.615 + ], + "angle": 0, + "content": "7. Carey, S.J., Lopich, A., Barr, D.R., Wang, B., Dudek, P.: A 100,000 fps vision sensor with embedded 535 GOPS/W \\(256 \\times 256\\) SIMD processor array. In: 2013 Symposium on VLSI Circuits. pp. C182-C183 (Jun 2013)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.617, + 0.787, + 0.656 + ], + "angle": 0, + "content": "8. Chan, S.H.: What Does a One-Bit Quanta Image Sensor Offer? IEEE Transactions on Computational Imaging 8, 770-783 (2022). https://doi.org/10.1109/TCI.2022.3202012" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.659, + 0.787, + 0.713 + ], + "angle": 0, + "content": "9. Charbon, E., Bruschini, C., Lee, M.J.: 3D-Stacked CMOS SPAD Image Sensors: Technology and Applications. In: 2018 25th IEEE International Conference on Electronics, Circuits and Systems (ICECS). pp. 1-4 (Dec 2018). https://doi.org/10.1109/ICECS.2018.8617983" + }, + { + "type": "ref_text", + "bbox": [ + 0.218, + 0.715, + 0.787, + 0.77 + ], + "angle": 0, + "content": "10. Chi, Y., Gnanasambandam, A., Koltun, V., Chan, S.H.: Dynamic low-light imaging with quanta image sensors. In: Computer Vision-ECCV 2020: 16th European Conference, Glasgow, UK, August 23-28, 2020, Proceedings, Part XXI 16. pp. 122-138. Springer (2020)" + }, + { + "type": "ref_text", + "bbox": [ + 0.218, + 0.772, + 0.787, + 0.84 + ], + "angle": 0, + "content": "11. Della Rocca, F.M., Mai, H., Hutchings, S.W., Al Abbas, T., Buckbee, K., Tsiamis, A., Lomax, P., Gyongy, I., Dutton, N.A., Henderson, R.K.: A \\(128 \\times 128\\) SPAD motion-triggered time-of-flight image sensor with in-pixel histogram and column-parallel vision processor. IEEE Journal of Solid-State Circuits 55(7), 1762-1775 (2020). https://doi.org/10.1109/JSSC.2020.2993722" + }, + { + "type": "list", + "bbox": [ + 0.218, + 0.318, + 0.787, + 0.84 + ], + "angle": 0, + "content": null + } + ], + [ + { + "type": "page_number", + "bbox": [ + 0.218, + 0.116, + 0.236, + 0.127 + ], + "angle": 0, + "content": "16" + }, + { + "type": "header", + "bbox": [ + 0.272, + 0.115, + 0.384, + 0.127 + ], + "angle": 0, + "content": "L. Koerner et al." + }, + { + "type": "ref_text", + "bbox": [ + 0.218, + 0.148, + 0.785, + 0.188 + ], + "angle": 0, + "content": "12. Diamond, J.S.: Inhibitory interneurons in the retina: Types, circuitry, and function. Annual Review of Vision Science 3(1), 1-24 (2017). https://doi.org/10.1146/annurev-vision-102016-061345" + }, + { + "type": "ref_text", + "bbox": [ + 0.218, + 0.189, + 0.785, + 0.229 + ], + "angle": 0, + "content": "13. Dollar, P., Zitnick, C.L.: Structured Forests for Fast Edge Detection. In: Proceedings of the IEEE International Conference on Computer Vision. pp. 1841-1848 (2013)" + }, + { + "type": "ref_text", + "bbox": [ + 0.218, + 0.23, + 0.785, + 0.284 + ], + "angle": 0, + "content": "14. Dutton, N.A.W., Al Abbas, T., Gyongy, I., Mattioli Della Rocca, F., Henderson, R.K.: High Dynamic Range Imaging at the Quantum Limit with Single Photon Avalanche Diode-Based Image Sensors. Sensors 18(4), 1166 (Apr 2018). https://doi.org/10.3390/s18041166" + }, + { + "type": "ref_text", + "bbox": [ + 0.218, + 0.285, + 0.785, + 0.325 + ], + "angle": 0, + "content": "15. Fossum, E.R.: Modeling the Performance of Single-Bit and Multi-Bit Quanta Image Sensors. IEEE Journal of the Electron Devices Society 1(9), 166-174 (Sep 2013). https://doi.org/10.1109/JEDS.2013.2284054" + }, + { + "type": "ref_text", + "bbox": [ + 0.218, + 0.326, + 0.785, + 0.365 + ], + "angle": 0, + "content": "16. Franke, K., Berens, P., Schubert, T., Bethge, M., Euler, T., Baden, T.: Inhibition decorrelates visual feature representations in the inner retina. Nature 542(7642), 439-444 (Feb 2017). https://doi.org/10.1038/nature21394" + }, + { + "type": "ref_text", + "bbox": [ + 0.218, + 0.365, + 0.785, + 0.42 + ], + "angle": 0, + "content": "17. Gallego, G., Delbrück, T., Orchard, G., Bartolozzi, C., Taba, B., Censi, A., Leutenegger, S., Davison, A.J., Conradt, J., Daniilidis, K., et al.: Event-based vision: A survey. IEEE Transactions on Pattern Analysis and Machine Intelligence 44(1), 154-180 (2020)" + }, + { + "type": "ref_text", + "bbox": [ + 0.218, + 0.421, + 0.785, + 0.473 + ], + "angle": 0, + "content": "18. Gnanasambandam, A., Chan, S.H.: HDR Imaging With Quanta Image Sensors: Theoretical Limits and Optimal Reconstruction. IEEE Transactions on Computational Imaging 6, 1571-1585 (2020). https://doi.org/10.1109/TCI.2020.3041093" + }, + { + "type": "ref_text", + "bbox": [ + 0.218, + 0.475, + 0.785, + 0.515 + ], + "angle": 0, + "content": "19. Gnanasambandam, A., Chan, S.H.: Exposure-Referred Signal-to-Noise Ratio for Digital Image Sensors. IEEE Transactions on Computational Imaging 8, 561-575 (Jun 2022). https://doi.org/10.1109/TCI.2022.3187657" + }, + { + "type": "ref_text", + "bbox": [ + 0.218, + 0.516, + 0.785, + 0.568 + ], + "angle": 0, + "content": "20. Gupta, M., Iso, D., Nayar, S.K.: Fibonacci Exposure Bracketing for High Dynamic Range Imaging. In: 2013 IEEE International Conference on Computer Vision. pp. 1473-1480. IEEE, Sydney, Australia (Dec 2013). https://doi.org/10.1109/ICCV.2013.186" + }, + { + "type": "ref_text", + "bbox": [ + 0.218, + 0.569, + 0.785, + 0.597 + ], + "angle": 0, + "content": "21. Gyongy, I., Dutton, N.A., Henderson, R.K.: Direct time-of-flight single-photon imaging. IEEE Transactions on Electron Devices 69(6), 2794-2805 (2021)" + }, + { + "type": "ref_text", + "bbox": [ + 0.218, + 0.597, + 0.785, + 0.637 + ], + "angle": 0, + "content": "22. Gyongy, I., Martin, G.M., Turpin, A., Ruget, A., Halimi, A., Henderson, R., Leach, J.: High-speed vision with a 3D-stacked SPAD image sensor. In: Advanced Photon Counting Techniques XV. vol. 11721, p. 1172105. SPIE (2021)" + }, + { + "type": "ref_text", + "bbox": [ + 0.218, + 0.637, + 0.785, + 0.664 + ], + "angle": 0, + "content": "23. Haldane, J.B.S.: On a Method Of Estimating Frequencies. Biometrika 33(3), 222-225 (Nov 1945). https://doi.org/10.1093/biomet/33.3.222" + }, + { + "type": "ref_text", + "bbox": [ + 0.218, + 0.664, + 0.785, + 0.732 + ], + "angle": 0, + "content": "24. Henderson, R.K., Johnston, N., Hutchings, S.W., Gyongy, I., Al Abbas, T., Dutton, N., Tyler, M., Chan, S., Leach, J.: \\(256 \\times 256\\) 40nm/90nm CMOS 3D-stacked 120 dB dynamic-range reconfigurable time-resolved SPAD imager. In: 2019 IEEE International Solid-State Circuits Conference-(ISSCC). pp. 106-108. IEEE (2019). https://doi.org/10.1109/ISSCC.2019.8662355" + }, + { + "type": "ref_text", + "bbox": [ + 0.218, + 0.733, + 0.785, + 0.759 + ], + "angle": 0, + "content": "25. Iams, H., Salzberg, B.: The secondary emission phototube. Proceedings of the Institute of Radio Engineers 23(1), 55-64 (1935)" + }, + { + "type": "ref_text", + "bbox": [ + 0.218, + 0.759, + 0.785, + 0.8 + ], + "angle": 0, + "content": "26. Ingle, A., Seets, T., Buttafava, M., Gupta, S., Tosi, A., Gupta, M., Velten, A.: Passive Inter-Photon Imaging. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 8585–8595 (2021)" + }, + { + "type": "ref_text", + "bbox": [ + 0.218, + 0.8, + 0.785, + 0.84 + ], + "angle": 0, + "content": "27. Ingle, A., Velten, A., Gupta, M.: High Flux Passive Imaging With Single-Photon Sensors. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 6760-6769 (2019)" + }, + { + "type": "list", + "bbox": [ + 0.218, + 0.148, + 0.785, + 0.84 + ], + "angle": 0, + "content": null + } + ], + [ + { + "type": "header", + "bbox": [ + 0.322, + 0.115, + 0.732, + 0.129 + ], + "angle": 0, + "content": "Photon Inhibition for Energy-Efficient Single-Photon Imaging" + }, + { + "type": "page_number", + "bbox": [ + 0.769, + 0.117, + 0.785, + 0.127 + ], + "angle": 0, + "content": "17" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.147, + 0.786, + 0.188 + ], + "angle": 0, + "content": "28. J Yoshida: Breaking Down iPad Pro 11's LiDAR Scanner. https://www.eetimes.com/breaking-down-ipad-pro-11s-lidar-scanner/, EE Times 6/5/2020. Accessed 5/6/2021." + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.19, + 0.786, + 0.217 + ], + "angle": 0, + "content": "29. Jocher, G., Chaurasia, A., Qiu, J.: Ultralytics YOLOv8 (Jan 2023), https://github.com/ultralytics/ultralytics, Accessed 3/6/2024" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.218, + 0.786, + 0.272 + ], + "angle": 0, + "content": "30. Koppal, S.J., Gkioulekas, I., Young, T., Park, H., Crozier, K.B., Barrows, G.L., Zickler, T.: Toward Wide-Angle Microvision Sensors. IEEE Transactions on Pattern Analysis and Machine Intelligence 35(12), 2982-2996 (Dec 2013). https://doi.org/10.1109/TPAMI.2013.22" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.273, + 0.786, + 0.327 + ], + "angle": 0, + "content": "31. Liu, Y., Gutierrez-Barragan, F., Ingle, A., Gupta, M., Velten, A.: Single-Photon Camera Guided Extreme Dynamic Range Imaging. In: Proceedings of the IEEE/CVF Winter Conference on Applications of Computer Vision. pp. 1575-1585 (2022)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.328, + 0.786, + 0.37 + ], + "angle": 0, + "content": "32. Ma, J., Masoodian, S., Starkey, D.A., Fossum, E.R.: Photon-number-resolving megapixel image sensor at room temperature without avalanche gain. Optica 4(12), 1474-1481 (Dec 2017). https://doi.org/10.1364/OPTICA.4.001474" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.371, + 0.786, + 0.424 + ], + "angle": 0, + "content": "33. Ma, J., Zhang, D., Elgendy, O.A., Masoodian, S.: A 0.19e- rms Read Noise 16.7Mpixel Stacked Quanta Image Sensor With 1.1 Mm-Pitch Backside Illuminated Pixels. IEEE Electron Device Letters 42(6), 891-894 (Jun 2021). https: //doi.org/10.1109/LED.2021.3072842" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.425, + 0.786, + 0.466 + ], + "angle": 0, + "content": "34. Ma, S., Gupta, S., Ulku, A.C., Bruschini, C., Charbon, E., Gupta, M.: Quanta burst photography. ACM Transactions on Graphics 39(4), 79:1-79:16 (Aug 2020). https://doi.org/10.1145/3386569.3392470" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.467, + 0.786, + 0.508 + ], + "angle": 0, + "content": "35. Ma, S., Mos, P., Charbon, E., Gupta, M.: Burst Vision Using Single-Photon Cameras. In: Proceedings of the IEEE/CVF Winter Conference on Applications of Computer Vision. pp. 5375-5385 (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.509, + 0.786, + 0.562 + ], + "angle": 0, + "content": "36. Medin, S.C., Murray-Bruce, J., Castanón, D., Goyal, V.K.: Beyond Binomial and Negative Binomial: Adaptation in Bernoulli Parameter Estimation. IEEE Transactions on Computational Imaging 5(4), 570-584 (Dec 2019). https://doi.org/10.1109/TCI.2019.2913108" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.564, + 0.786, + 0.659 + ], + "angle": 0, + "content": "37. Morimoto, K., Iwata, J., Shinohara, M., Sekine, H., Abdelghafar, A., Tsuchiya, H., Kuroda, Y., Tojima, K., Endo, W., Maehashi, Y., Ota, Y., Sasago, T., Maekawa, S., Hikosaka, S., Kanou, T., Kato, A., Tezuka, T., Yoshizaki, S., Ogawa, T., Uehira, K., Ehara, A., Inui, F., Matsuno, Y., Sakurai, K., Ichikawa, T.: 3.2 Megapixel 3D-Stacked Charge Focusing SPAD for Low-Light Imaging and Depth Sensing. In: 2021 IEEE International Electron Devices Meeting (IEDM). pp. 20.2.1–20.2.4 (Dec 2021). https://doi.org/10.1109/IEDM19574.2021.9720605" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.66, + 0.786, + 0.714 + ], + "angle": 0, + "content": "38. Morimoto, K., Ardelean, A., Wu, M.L., Ulku, A.C., Antolovic, I.M., Bruschini, C., Charbon, E.: Megapixel time-gated SPAD image sensor for 2D and 3D imaging applications. Optica 7(4), 346-354 (2020). https://doi.org/10.1364/OPTICA.386574" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.715, + 0.786, + 0.743 + ], + "angle": 0, + "content": "39. Morimoto, K., Charbon, E.: A scaling law for SPAD pixel miniaturization. Sensors 21(10), 3447 (2021)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.744, + 0.786, + 0.812 + ], + "angle": 0, + "content": "40. Ogi, J., Sano, F., Nakata, T., Kubo, Y., Onishi, W., Koswaththaghe, C., Mochizuki, T., Tashiro, Y., Hizu, K., Takatsuka, T., et al.: A \\(3.06\\mu \\mathrm{m}\\) spad pixel with embedded metal contact and power grid on deep trench pixel isolation for high-resolution photon-counting. In: 2023 International Image Sensor Workshop (IISW) (May 2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.813, + 0.786, + 0.84 + ], + "angle": 0, + "content": "41. Ogi, J., Takatsuka, T., Hizu, K., Inaoka, Y., Zhu, H., Tochigi, Y., Tashiro, Y., Sano, F., Murakawa, Y., Nakamura, M., Oike, Y.: A 250fps 124dB Dynamic-Range" + }, + { + "type": "list", + "bbox": [ + 0.217, + 0.147, + 0.786, + 0.84 + ], + "angle": 0, + "content": null + } + ], + [ + { + "type": "page_number", + "bbox": [ + 0.218, + 0.116, + 0.236, + 0.127 + ], + "angle": 0, + "content": "18" + }, + { + "type": "header", + "bbox": [ + 0.272, + 0.115, + 0.384, + 0.127 + ], + "angle": 0, + "content": "L. Koerner et al." + }, + { + "type": "ref_text", + "bbox": [ + 0.241, + 0.147, + 0.786, + 0.203 + ], + "angle": 0, + "content": "SPAD Image Sensor Stacked with Pixel-Parallel Photon Counter Employing SubFrame Extrapolating Architecture for Motion Artifact Suppression. In: 2021 IEEE International Solid-State Circuits Conference (ISSCC). vol. 64, pp. 113-115 (Feb 2021). https://doi.org/10.1109/ISSCC42613.2021.9365977" + }, + { + "type": "ref_text", + "bbox": [ + 0.218, + 0.204, + 0.787, + 0.313 + ], + "angle": 0, + "content": "42. Ota, Y., Morimoto, K., Sasago, T., Shinohara, M., Kuroda, Y., Endo, W., Maehashi, Y., Maekawa, S., Tsuchiya, H., Abdelahafar, A., Hikosaka, S., Motoyama, M., Tojima, K., Uehira, K., Iwata, J., Inui, F., Matsuno, Y., Sakurai, K., Ichikawa, T.: A 0.37W 143dB-Dynamic-Range 1Mpixel Backside-Illuminated Charge-Focusing SPAD Image Sensor with Pixel-Wise Exposure Control and Adaptive Clocked Recharging. In: 2022 IEEE International Solid-State Circuits Conference (ISSCC). vol. 65, pp. 94-96 (Feb 2022). https://doi.org/10.1109/ISSCC42614.2022.9731644" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.314, + 0.786, + 0.357 + ], + "angle": 0, + "content": "43. Po, R., Pediredla, A., Gkioulekas, I.: Adaptive gating for single-photon 3d imaging. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 16354-16363 (2022)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.357, + 0.786, + 0.398 + ], + "angle": 0, + "content": "44. Severini, F., Cusini, I., Berretta, D., Pasquinelli, K., Incoronato, A., Villa, F.: SPAD pixel with sub-ns dead-time for high-count rate applications. IEEE Journal of Selected Topics in Quantum Electronics 28(2: Optical Detectors), 1-8 (2021)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.398, + 0.786, + 0.453 + ], + "angle": 0, + "content": "45. Shenzhen CM Technology company Ltd: 16 MP MIPI Camera Module with SONY IMX206 sensor. http://www-camera-module.com/product/mipicameramodule/16mp-mipi-camera-module-sony-imx206-sensor.html (Nov 2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.453, + 0.787, + 0.495 + ], + "angle": 0, + "content": "46. Sundar, V., Ardelean, A., Swedish, T., Brusschini, C., Charbon, E., Gupta, M.: Sodacam: Software-defined cameras via single-photon imaging. In: Proceedings of the IEEE International Conference on Computer Vision (ICCV) (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.495, + 0.787, + 0.591 + ], + "angle": 0, + "content": "47. Takatsuka, T., Ogi, J., Ikeda, Y., Hizu, K., Inaoka, Y., Sakama, S., Watanabe, I., Ishikawa, T., Shimada, S., Suzuki, J., Maeda, H., Toshima, K., Nonaka, Y., Yamamura, A., Ozawa, H., Koga, F., Oike, Y.: A 3.36 \\(\\mu\\)m-pitch SPAD photon-counting image sensor using clustered multi-cycle clocked recharging technique with intermediate most-significant-bit readout. In: 2023 IEEE Symposium on VLSI Technology and Circuits (VLSI Technology and Circuits). pp. 1-2 (Jun 2023). https://doi.org/10.23919/VLSITechnologyandCir57934.2023.10185241" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.591, + 0.787, + 0.647 + ], + "angle": 0, + "content": "48. Tilmon, B., Sun, Z., Koppal, S.J., Wu, Y., Evangelidis, G., Zahreddine, R., Krishnan, G., Ma, S., Wang, J.: Energy-Efficient Adaptive 3D Sensing. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 5054-5063 (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.647, + 0.787, + 0.702 + ], + "angle": 0, + "content": "49. Ulku, A.C., Bruschini, C., Antolovic, I.M., Kuo, Y., Ankri, R., Weiss, S., Michalet, X., Charbon, E.: A \\(512 \\times 512\\) SPAD Image Sensor With Integrated Gating for Widefield FLIM. IEEE Journal of Selected Topics in Quantum Electronics 25(1), 1-12 (Jan 2019). https://doi.org/10.1109/jstqe.2018.2867439" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.702, + 0.787, + 0.731 + ], + "angle": 0, + "content": "50. Wang, X.: Laplacian operator-based edge detectors. IEEE Transactions on Pattern Analysis and Machine Intelligence 29(5), 886-890 (2007)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.731, + 0.787, + 0.772 + ], + "angle": 0, + "content": "51. Wang, Z., Bovik, A.C., Sheikh, H.R., Simoncelli, E.P.: Image quality assessment: From error visibility to structural similarity. IEEE Transactions on Image Processing 13(4), 600-612 (2004)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.772, + 0.787, + 0.813 + ], + "angle": 0, + "content": "52. Wayne, M.A., Bienfang, J.C., Migdall, A.L.: Low-noise photon counting above \\(100 \\times 106\\) counts per second with a high-efficiency reach-through single-photon avalanche diode system. Applied Physics Letters 118(13) (2021)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.813, + 0.787, + 0.841 + ], + "angle": 0, + "content": "53. Xie, S., Tu, Z.: Holistically-nested edge detection. In: Proceedings of IEEE International Conference on Computer Vision. pp. 1395-1403 (2015)" + }, + { + "type": "list", + "bbox": [ + 0.217, + 0.147, + 0.787, + 0.841 + ], + "angle": 0, + "content": null + } + ], + [ + { + "type": "header", + "bbox": [ + 0.321, + 0.115, + 0.732, + 0.13 + ], + "angle": 0, + "content": "Photon Inhibition for Energy-Efficient Single-Photon Imaging" + }, + { + "type": "page_number", + "bbox": [ + 0.768, + 0.117, + 0.786, + 0.127 + ], + "angle": 0, + "content": "19" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.147, + 0.786, + 0.175 + ], + "angle": 0, + "content": "54. Xu, Y., Lu, J., Wu, Z.: A compact high-speed active quenching and recharging circuit for SPAD detectors. IEEE Photonics Journal 12(5), 1-8 (2020)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.176, + 0.787, + 0.218 + ], + "angle": 0, + "content": "55. Yang, F., Lu, Y.M., Sbaiz, L., Vetterli, M.: Bits from photons: Oversampled image acquisition using binary poisson statistics. IEEE Transactions on image processing 21(4), 1421-1436 (2011)" + }, + { + "type": "list", + "bbox": [ + 0.217, + 0.147, + 0.787, + 0.218 + ], + "angle": 0, + "content": null + } + ] +] \ No newline at end of file diff --git a/2024/Photon Inhibition for Energy-Efficient Single-Photon Imaging/03673fae-a9a6-4975-8cf4-3d5bc33fd48c_origin.pdf b/2024/Photon Inhibition for Energy-Efficient Single-Photon Imaging/03673fae-a9a6-4975-8cf4-3d5bc33fd48c_origin.pdf new file mode 100644 index 0000000000000000000000000000000000000000..5d1578e185d77740f50cc4e82907867a0549bc8a --- /dev/null +++ b/2024/Photon Inhibition for Energy-Efficient Single-Photon Imaging/03673fae-a9a6-4975-8cf4-3d5bc33fd48c_origin.pdf @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:be1ccca8dcea15941b7476a3b58210c3e21ddee367ebb8c2b81a710205157779 +size 4740557 diff --git a/2024/Photon Inhibition for Energy-Efficient Single-Photon Imaging/full.md b/2024/Photon Inhibition for Energy-Efficient Single-Photon Imaging/full.md new file mode 100644 index 0000000000000000000000000000000000000000..246307ab4a7852f2ee1c3a7fc4d11201e3d1d8ba --- /dev/null +++ b/2024/Photon Inhibition for Energy-Efficient Single-Photon Imaging/full.md @@ -0,0 +1,279 @@ +# Photon Inhibition for Energy-Efficient Single-Photon Imaging + +Lucas J. Koerner1, Shantanu Gupta2, Atul Ingle3, and Mohit Gupta2 + +$^{1}$ Department of Electrical and Computer Engineering, University of St. Thomas, St. Paul, MN, 55105, USA koerner.lucas@stthomas.edu +$^{2}$ Department of Computer Sciences, University of Wisconsin-Madison, Madison, WI, 53706, USA {sgupta,mohitg}@cs.wisc.edu +$^{3}$ Department of Computer Science, Portland State University, Portland, OR, 97201, USA +ingle2@pdx.edu + +Abstract. Single-photon cameras (SPCs) are emerging as sensors of choice for various challenging imaging applications. One class of SPCs based on the single-photon avalanche diode (SPAD) detects individual photons using an avalanche process; the raw photon data can then be processed to extract scene information under extremely low light, high dynamic range, and rapid motion. Yet, single-photon sensitivity in SPADs comes at a cost — each photon detection consumes more energy than that of a CMOS camera. This avalanche power significantly limits sensor resolution and could restrict widespread adoption of SPAD-based SPCs. We propose a computational-imaging approach called photon inhibition to address this challenge. Photon inhibition strategically allocates detections in space and time based on downstream inference task goals and resource constraints. We develop lightweight, on-sensor computational inhibition policies that use past photon data to disable SPAD pixels in real-time, to select the most informative future photons. As case studies, we design policies tailored for image reconstruction and edge detection, and demonstrate, both via simulations and real SPC captured data, considerable reduction in photon detections (over $90\%$ of photons) while maintaining task performance metrics. Our work raises the question of "which photons should be detected?", and paves the way for future energy-efficient single-photon imaging. Source code for our experiments is available at https://wisionlab.com/project/inhibition. + +Keywords: single-photon imaging $\cdot$ SPADs $\cdot$ energy-efficient vision + +# 1 Introduction + +From vacuum tube-based single-photon detectors invented in the early 20th century [25], to 3D depth sensing LiDAR cameras found in today's smartphones [28], single-photon camera (SPC) technology has come a long way in terms of pixel resolution and commercial availability for a variety of applications. Thanks to CMOS-compatible single-photon avalanche diode (SPAD) arrays, SPCs are being increasingly used not only in niche fields such as scientific imaging and biomedical + +![](images/569a1e0f3cf9c3fd9630c153c427ef6c15f59798b072886a2421b02be122130c.jpg) +(a) +(c) +(b) + +![](images/085285b6037fa00abba45c15cafcfeaa09725b66253ac9ed2bc26f4bea002c45.jpg) + +![](images/fbe062150553a0bf3eca8f4afb5fa72dee4aaa1edcf93b24649106b68171e51e.jpg) + +![](images/83a5d838b33a448d72f6c75b24a21f7f97476f5d7b1264981d0fa51783af66ff.jpg) +(d) +Fig. 1: Photon inhibition for resource-efficient passive SPAD imaging. (a) Unlike conventional CMOS cameras (CIS), the energy consumption in SPAD cameras increases with scene brightness, severely limiting the applicability of high resolution SPAD cameras in resource-constrained applications. (b) We expand the conventional imaging pipeline to incorporate "inhibition" that electronically enables or disables individual pixels to limit bandwidth and power consumption. Our method relies on lightweight mathematical operations called "inhibition policies" that update the inhibition patterns based on the history of photon detections. Inhibition policies can be optimized for image SNR or for downstream vision tasks.(c,d) Object detection, a high-level vision task, is successful even with a large fraction of photons inhibited. + +![](images/dc6dea98ffdf6a07b3aec81f67e04c50b0a005874003987e8430f359261cc3a5.jpg) + +![](images/afe810e526862fbe033a7bc4a3153b6205ec818e7b3f981da0caea6b64b89de1.jpg) +YOLOv8 Object Detection + +microscopy [49], but more widely for consumer photography [37]. SPAD-based SPCs have recently been fabricated into kilo-to-megapixel format arrays that are now commercially available [28, 37]. The extreme sensitivity and high speed can benefit passive low-light computer vision tasks, particularly in the presence of rapid scene or camera motion [35], enable wide dynamic-range imaging [14, 31] and photon-starved active imaging applications such as 3D imaging (LiDAR) [21] and fluorescence microscopy [49]. + +SPAD camera pixels detect individual photons with extremely high frame rates by exploiting avalanche multiplication. On one hand, being able to detect individual photons opens up new possibilities and capabilities for computer vision systems. On the other hand, this also presents a unique challenge: Every photon-induced avalanche comes with a non-negligible energy cost, which is a challenge exclusive to SPAD-based cameras. This flux-dependent photon detection power is a significant fraction of total power consumption in today's SPAD cameras and impedes further increases in their spatial resolution [24,37,41,42,47]. For example, extrapolating the avalanche power of a recent SPAD sensor [47] to $\sim 10$ 's of megapixels format predicts a power consumption of several watts in + +bright light $^{4}$ (as illustrated in Fig. 1(a)), far exceeding the power consumption of modern CMOS image sensors (CIS) of around $300\mathrm{mW}$ [45]. + +We address this problem by being selective about which photons are detected on the SPAD sensor while maintaining good performance on various computer vision tasks. To reduce avalanche power consumption and decouple it from photon flux, we propose a computational imaging technique called photon inhibition where individual SPAD pixels are disabled adaptively based on previous photon detections in their spatio-temporal neighborhoods. Electrically disabling pixels prevents photon detections to inhibit any avalanche power or subsequent processing. We implement lightweight on-sensor computations called inhibition policies (Fig. 1(b)) to determine, in real-time and at single-pixel and single-frame granularity, which SPAD pixels to enable or disable. Our method is inspired by retinal pre-processing of the human visual system where retinal neurons aggregate photon information over small spatio-temporal neighborhoods to cause neighboring retinal cells to become less sensitive to incident light [12, 16].5 + +Given that SPADs introduce a new challenge of flux-dependent power consumption, we establish, from first principles, novel energy-aware imaging performance metrics for resource-constrained single-photon imaging. Based on these metrics, we design families of inhibition policies that distribute photon detections in space and time based on imaging / vision task goals and energy consumption constraints. A critical consideration in the design of inhibition policies stems from the observation that these policies are meant to control (enable / disable) the sensor (Fig. 1(b)). Therefore, it is essential for these policies to be extremely lightweight since these need to be implemented on sensor with very limited compute and memory resources. Furthermore, inhibition policies must execute at ultra-low latency to keep up with high-speed photon detections (reaching up to $100\mathrm{kHz}$ ). Fortunately, since the raw data output from a SPAD-based SPC consists of binary-valued image frames, SPCs are naturally suited to real-time calculations at the image sensor plane under tight compute and memory budgets. The proposed inhibition policies are lightweight, requiring only simple arithmetic and Boolean operations computed over local spatio-temporal neighborhoods, and thus amenable to in-pixel implementation [2, 7]. + +In simulations and real experimental data, we show that our inhibition policies allocate photon detections to sensor pixels in a way that reduces detection energy for a given accuracy level of various vision tasks. Our results show consequential energy savings when compared to a conventional capture scheme for tasks of (i) image reconstruction: $42\%$ fewer photon detections at equal image quality; (ii) edge detection: an edge sensitive inhibition policy reduces detections by $30\%$ at equal F-score; and (iii) YOLOv8 object detection: remains success- + +ful with $95\%$ of photons inhibited under camera motion of a real-world SPC (Fig. 1(c,d)). Through experiments with photon streams captured using real SPAD camera hardware over a wide range of illumination conditions, we show that our proposed inhibition policies preserve low-light details and, in brightlight, decouple flux and detection energy. + +Scope and Limitations: There are several competing image sensor technologies today that resolve single-photon while capturing binary-valued raw frames at rates exceeding thousands of frames/second. SPCs based on "jots" [32] that do not rely on avalanche multiplication do not suffer from flux-dependent power consumption as SPADs. Jots are an exciting technology, especially in scenarios that require high resolution and high dynamic range imaging [33]. In this paper, we focus on SPADs, due to their rapidly rising availability and commercial interest [6, 28], and benefits over conventional CMOS sensors, both in low-light and bright scenes for a variety of computer vision tasks [10, 35]. + +Our goals in this paper are to (a) raise the question of "which photons should be detected?" in the context of energy-efficient single-photon imaging, (b) establish a design space and metrics to evaluate various inhibition policies, and (c) propose plausible policies that respect practical hardware limitations for future on-chip implementation. We emphasize that the inhibition policies proposed in this paper are not necessarily optimal. This work is just a first step towards demonstrating that it is possible to achieve high performance with SPADs, while maintaining low power consumption via photon inhibition. + +# 2 Related Work + +Hardware approaches for reduced energy consumption: There is a strong dependence of SPAD power consumption on the pixel size — the smaller the pixel, the lower the avalanche energy [39]. Although recent developments in SPAD pixel technology have reduced pixel sizes to below $4\mu \mathrm{m}$ [40], avalanche energy still contributes a large fraction of the total power consumption in a SPAD sensor [44]. SPAD design optimizations have reduced the charge per avalanche by RF modulation of the bias voltage [52], minimization of the junction capacitance [42], and smart [5] and fast [54] quenching circuits. Circuit architectures may require spatial and/or temporal co-incidence [22] to reduce energy downstream in the processing chain, but avalanche energy remains. Our work complements existing hardware approaches by preventing avalanches altogether to reduce illumination-dependent energy consumption. + +SPAD dead-time and clocked recharge: SPADs require recharge after an avalanche-inducing photon detection during which subsequent photons are not recorded. This dead-time inhibits photons at high exposure [24, 27], yet, power consumption remains excessive when the average inter-photon arrival interval is shorter than the SPAD dead time [38, 42]. Power consumption at high photon flux can be reduced by controlling the global rate of SPAD recharge ("clocked recharge") [38, 47]. Within a single recharge period a SPAD detects at most a single photon; subsequent photon arrivals do not induce an avalanche, thus power is + +reduced. To further limit avalanches, clocked recharge has been combined with a limit on the number of detections and time to saturation circuitry to measure the intensity of the saturated pixels [41], coarse pixel-wise exposure control [42], and a sequence of recharge periods similar to exposure bracketing [47]. These methods could be considered special cases of inhibition policies that are global and do not adapt to scene content, and therefore, are limited in flexibility to trade power and measurement fidelity. Consequently, in these methods, the power of avalanches remains a considerable fraction of total SPAD sensor power consumption [47]. For example, in this work we show considerable power savings, reaching up to $90\%$ , for an object detection task. + +Resource-aware imaging: Event-based vision sensors reduce power consumption by only transmitting scene information when an intensity change is detected [17]. This idea has recently been applied to SPAD arrays to reduce power consumption due to data transfer [11, 46]. In contrast, our method reduces power due to detection by selectively disabling photodiodes based on photon history over small spatial and temporal neighborhoods. Miniaturized cameras with constraints on compute energy have transferred processing to the optical domain [30]. While we focus on passive imaging, depth sensing with SPADs and an active pulsed light source has related constraints such as acquisition time and laser power. Adaptive gating reduces acquisition time [43] and optimal allocation of the laser dwell time among pixels improves data quality [36, 48]. + +# 3 Observation Model + +During an exposure time $T$ , a photon flux of $\phi$ results in an average number of photon conversions, or exposure, of $H = \phi T$ (we fold the sensor's photon detection probability, or PDP, into the definition of $\phi$ , meaning it represents an effective flux). The distribution of photon conversions, $K$ , follows the Poisson distribution given by $\mathsf{P}(K = k;H) = \frac{H^k e^{-H}}{k!}$ . During each binary exposure period, a SPAD pixel records a '1' if at least one photon was incident during that period, and '0' otherwise. The probability of detecting at least one photon is given by $Y := 1 - \mathsf{P}(K = 0;H) = 1 - e^{-H}$ . Hence, the SPAD pixel readout in each binary frame is a Bernoulli random variable $B \sim \mathrm{Bernoulli}(Y)$ . Multiple exposure time windows, or measurements $(W)$ , are recorded to reduce noise with the total number of detections + +$$ +D := \left[ \sum_ {n = 1} ^ {W} B _ {n} \right] \sim \operatorname {B i n o m i a l} (W, 1 - e ^ {- H}). \tag {1} +$$ + +We estimate the probability of detection and flux from a measurement $D$ as + +$$ +\widehat {Y} = \frac {D}{W} \tag {2} +$$ + +$$ +\widehat {H} = - \ln (1 - \widehat {Y}). \qquad \qquad (3) +$$ + +Changes under inhibition: Inhibition is represented by a binary state variable $M$ at every pixel, with $M_{n} = 1$ denoting enabled for the $n$ -th measurement period. $M$ is a random variable when an adaptive or data-dependent inhibition policy is used. The total number of measurements changes to $W_{inh.} \coloneqq \sum_{n=1}^{W} M_{n} \leq W$ . The number of detections becomes + +$$ +D _ {i n h.} := \left[ \sum_ {n = 1} ^ {W} M _ {n} B _ {n} \right] = \left[ \sum_ {\substack {n = 1 \\ M _ {n} = 1}} ^ {W} B _ {n} \right] \sim \operatorname {B i n o m i a l} \left(W _ {i n h.}, 1 - e ^ {- H}\right), \tag{4} +$$ + +and the flux is estimated similar to Eqs. (2) and (3). The second summation conveys that, with inhibition, the measurements when the pixel is enabled match the original model in Eq. (1) - a result of the memoryless property of the Poisson arrival process. The model above requires that transitions in $M$ are synchronized with the clock signal used to gate the exposure, so that the PDP is not changed by inhibition. A second assumption is that the time for SPAD recharge is small relative to the clock period. This is a desired property for passive SPAD-based imaging, and holds for many state-of-the-art SPADs [42, 44, 47]. It implies that PDP is approximately constant in time and does not depend on prior pixel state. + +# 4 Energy-Aware Performance Metrics + +The exposure-referred signal-to-noise ratio $(\mathsf{SNR}_H)$ is commonly used to evaluate single-photon sensor performance, and can be computed as the ratio of the true exposure $H$ and the root-mean-squared error in the estimated exposure $\sqrt{\mathbb{E}[(\widehat{H} - H)^2]}$ [15, 55]: + +$$ +\mathsf {S N R} _ {H} = \frac {H}{\sqrt {\mathbb {E} [ (\widehat {H} - H) ^ {2} ]}} = H \sqrt {\frac {W}{e ^ {H} - 1}}. \tag {5} +$$ + +At low incident flux, $\mathsf{SNR}_H$ is low due to shot noise. $\mathsf{SNR}_H$ improves as the likelihood of a photon detection increases until, in bright light with $H > 1.6$ , the $\mathsf{SNR}_H$ degrades due to "soft" saturation of the response [27,34]. + +We propose two new energy-aware modifications to $\mathsf{SNR}_H$ to incorporate SPAD energy costs. First, we propose a detection efficiency metric $\mathsf{SNR}_{H / D}^2$ defined as the square of SNR normalized by the expected number of detections $\mathbb{E}[D] = W(1 - e^{-H})$ : + +$$ +\mathsf {S N R} _ {H / D} ^ {2} := \frac {\mathsf {S N R} _ {H} ^ {2}}{\mathbb {E} [ D ]} = \frac {H ^ {2} e ^ {- H}}{(1 - e ^ {- H}) ^ {2}}. \tag {6} +$$ + +Fig. 3(a) shows the $\mathsf{SNR}_H$ (black) and the detection efficiency (red) versus the average photon arrivals per period $(H)$ . When $H \ll 1$ all detections contribute significant information and $\mathsf{SNR}_H^2$ increases linearly, similar to an ideal non-saturating sensor only limited by Poisson noise. Accordingly, $\mathsf{SNR}_{H / D}^2 \cong 1$ , the + +upper bound of this metric. At larger exposure values, beginning around $H \approx 0.5$ , the sensor begins to saturate which slows the growth of $\mathrm{SNR}_H$ and thus degrades the detection efficiency. + +A separate constraint is the total number of recharge periods during which a pixel is enabled and can measure either '0' or '1'. This number of measurements $(W)$ may be limited due to the energy to read out a frame, the depth of an in-pixel counter, and/or the maximum allowable sensing latency due to motion blur. We establish a second metric, measurement efficiency defined as the square of SNR normalized by the number of measurement windows: + +$$ +\mathrm {S N R} _ {H / W} ^ {2} := \frac {\mathrm {S N R} _ {H} ^ {2}}{W} = \frac {H ^ {2} e ^ {- H}}{1 - e ^ {- H}}. \tag {7} +$$ + +Fig. 3(a) (blue) shows sub-optimal measurement efficiency at both low and high exposures with the best efficiency at $H = 1.59$ , $Y = 0.80$ , as demonstrated in [8]. + +In an oracle setting with a known image, one can analytically derive nonuniform allocations of measurements $W_{i}$ across the pixels $i$ , constraining the total expected detections to a fixed value and optimizing $\mathsf{SNR}_H$ or mean-squared-error (see the supplement). The loss metric may also be defined relative to the binary rate $Y$ instead of $H$ as above. In that case a useful base metric could be entropy instead of $\mathsf{SNR}_H$ [19]. + +# 5 Spatio-temporal Inhibition Policies + +We now propose policies that calculate a spatio-temporal inhibition pattern for each pixel and each frame based on the history of photon frames and patterns. Following Sec. 3, we define an inhibition pattern using a binary-valued tensor $M$ where $M(i,j,t) = 0$ if pixel $(i,j)$ is disabled in the $t^{\mathrm{th}}$ frame, and $M(i,j,t) = 1$ otherwise. All pixels are initially enabled, and on-sensor calculations modify $M$ over time. The binary photon cube is defined as $F(i,j,t) = 1$ if the pixel is enabled (i.e., $M(i,j,t) = 1$ ) and a photon is detected at pixel location $(i,j)$ in the $t^{\mathrm{th}}$ frame, otherwise it is zero. + +![](images/94a9abed29f4a57bd67fe0da72b93450eef4df52c2f03cf44391ec6e9174d8db.jpg) +Fig. 2: Calculation-based inhibition with dead time. Arrows represent photons with an 'X' for inhibition. $T$ indicates the clocked recharge period. A score, $S$ , calculated from past frames determines if future measurements are enabled or disabled. + +![](images/2cae52915ddae7613b5cc833b5b17fb931d9e2f7e5efbb7027946db00bb5e776.jpg) +Fig. 3: Efficiency metrics and inhibition policies that track the metrics: (a) The $\mathsf{SNR}_H$ in dB (black), the detection efficiency (red, --), and the measurement efficiency (blue, --) versus the exposure with $W = 100$ measurements. The binary rate $Y = 1 - e^{-H}$ is indicated on the top axis. The vertical dotted line indicates the exposure and the binary rate at which the $\mathsf{SNR}_H$ degrades by $3\mathrm{dB}$ from the peak $\mathsf{SNR}_H$ . (b,c) Monte Carlo simulations of binary images using calculation-based inhibition policies demonstrate how the allocation of measurements versus the pixel exposure depends upon tuning parameters. (b) the inhibition threshold $\eta$ adjusts the exposure level at which pixels are inhibited to allow the measurement fraction (the ratio of active measurements to total number of frames) to follow the $\mathsf{SNR}_{H / D}^2$ curve in (a). A smaller threshold more aggressively inhibits photons. (c) demonstrates the impact of the hold-off time, $\tau_{H}$ , on the number of measurements allocated to the brightest pixels. The legend indicates the total fraction of photons inhibited as $I_F$ . (b,c) show smoothed curves (Lowess filter, fraction of $1 / 5$ ) of the measurement fraction vs. H. + +![](images/3561ef36ff700472374459675ae16d4cbbcce88f9ca78fbc0c2dce44c3b8c89c.jpg) + +Fig. 1(b) shows the components of a photon inhibition processing layer, including the binary frames, $F$ , and inhibition pattern $M$ . For ease of on-sensor implementation, we focus on policies that operate on small and local spatiotemporal neighborhoods of fixed sizes. We rely on local arithmetic and Boolean computations and comparison operations, consistent with current in-pixel computational capabilities [2,9,24]. Fig. 2 shows a proposed on-sensor calculation approach that operates in a streaming fashion as frames accumulate to calculate an inhibition score, $S$ , as the result of a spatio-temporal filter of the binary frames and inhibition pattern. The score at each pixel is calculated as + +$$ +S (i, j, t) = K * \left[ \left(2 F (i, j, t) - \mathbf {1}\right) \cdot M (i, j, t) \right] \tag {8} +$$ + +which applies a spatio-temporal filtering kernel, $K$ , of dimensions $L, H, T$ to a ternary representation of the pixel result (1, 0, or -1 for a detection, a disabled pixel, or a measurement that does not detect a photon, respectively). The kernel $K$ can typically be separated into spatial and temporal components as $K = K_{s} \otimes K_{t}$ with dimensions $L \times H \times 1$ and $1 \times 1 \times T$ , respectively. After each binary frame, the score is compared to a threshold $\eta$ and the pixel is disabled for the subsequent $\tau_{H}$ frames: $M(i,j,t') = 0$ for $\{t'|t + 1 \leq t' \leq t + 1 + \tau_{H}\}$ if $S(i,j,t) > \eta$ . Observe in Fig. 3(b,c) that decreasing $\eta$ and increasing $\tau_{H}$ can be used to attain more aggressive inhibition with a larger fraction of photons being inhibited. The binary rate of each pixel is estimated (using Eq. (4)) as the ratio of detections to (active) measurements: $\widehat{Y}(i,j,t) = \sum_{t'} F(i,j,t') / \sum_{t'} M(i,j,t')$ . This + +calculation requires a record of the inhibition history which could be accumulated by an in-pixel counter or recreated in a downstream processor if all binary frames are read out. + +![](images/711ddf144f258a16fb41255be347bd83cd923e9a2ad2c7aafa1b589fde3d9de6.jpg) + +![](images/b3071da236b4c798e7d1b4eec7e65417dd36ef56e19ebdabd307cf09e731ef3a.jpg) + +![](images/712b81ee56401f0e953743fd72dc033f08581a329c1b77aa913b339ba81a7efd.jpg) + +![](images/42cb2d604a71e767307335682a047ec4b5743d8331153e747755af236427e544.jpg) + +![](images/ea42ce3feb38c87db4188bdd54e4e392605a2a2ba7f28a00a6e56ce83fc2c4bc.jpg) +Fig. 4: Saturation look-ahead inhibition. (a) When detections $D_{i}$ of cycle $i$ exceed a threshold, $d_{i}$ , the photons arriving during the next recharge cycle, $i + 1$ , are inhibited. For this drawing $[d_{1}, d_{2}] = [3, 3]$ . (b, c, d) Example with three recharge period settings and brackets of $W_{i} = 10$ measurements each. The saturation thresholds are set as $[d_{1}, d_{2}] = [7, 7]$ . The brackets are combined linearly, weighted by $\mathsf{SNR}_H^2$ [18]. (b, c) With negligible effect on reconstructed SNR relative to conventional clocked-recharge, significantly fewer photon detections occur under the inhibition policy in high-flux settings. (d) The individual brackets are effectively disabled in flux regimes with low detection efficiency, and the detection rate closely tracks the original $\mathsf{SNR}_H^2$ curve. Simulation details are in the supplement. + +Example policies: 1) Single-pixel dead time: A spatial kernel $K_{s}$ of dimensions $1 \times 1$ may mimic the dead time of a passive SPAD without influence from neighboring pixels. Unlike SPAD recharge generated dead time, the temporal kernel may be extended (e.g., $N$ frames long) to establish a rate threshold for inhibition with reduced quantization noise. 2) Local spatio-temporal averaging: Single binary frames are inherently noisy. An inhibition policy that calculates spatiotemporal averages to estimate the local photon rate may reduce the impact of noise on the inhibition pattern. Another benefit, as shown in Fig. 3(b,c), is that the stochastic nature of binary frames smoothes the distribution of measurements versus pixel flux. Discontinuities ("dips") in SNR versus photon flux are undesirable due to the potential for artifacts. 3) Edge enhancement: Pixels may be inhibited if a local neighborhood has little spatial variation in photon rate. This can be achieved, for instance, through a spatial filter $K_{s}$ in Eq. (8) which acts + +like a Laplacian filter. Such a strategy may enhance the fidelity of edges in the image while focusing fewer resources on regions with constant illumination. + +Fig. 4 presents a second proposed inhibition policy called saturation lookahead. This policy combines exposure brackets and calculation-based inhibition for a light-weight single-pixel inhibition policy. This policy proceeds as a sequence of cycles (indexed by $i$ ) of binary frames where each binary frame within each cycle uses the same exposure time. Cycle exposure times $T_{i}$ progressively increase ( $T_{1} < T_{2} < T_{3} \ldots$ ) so that measurements taken in an earlier cycle may predict low detection efficiency (near saturation) at longer exposure times and disable the pixel in these subsequent cycles. The exposure level thresholds for inhibition would typically be set such that the number of photons detected at a given flux level tracks the $\mathsf{SNR}_H^2$ (see Fig. 4(b)), but may be adjusted further based on the relative importance of power consumption, sensing latency, and SNR in an actual application setting. + +# 6 Simulation-based Evaluation + +We use Monte Carlo simulations to generate sequences of grayscale binary frames from a dataset of RGB images [1]. The inhibition policies evaluated extend the baseline inhibition generated by clocked recharge. The inhibition score and patterns for various policies and tuning parameters are calculated from these binary frame sequences. Once inhibition patterns are determined, performance is evaluated by tabulating detections, measurements, and image quality or vision task performance for each step in the sequence (see the supplement for details). + +Spatio-Temporal Policies for Imaging: Handcrafted spatial kernels (3x3) were combined with an averaging temporal kernel of length 4 to form lightweight inhibition policies that allocate pixel measurements as described in earlier sections for improved image reconstruction. Fig. 5 summarizes the simulation results. Fig. 5(d,e) display reduction in photon detections at equal structural similarity index measure (SSIM) [51], enabled by disabling bright pixels, for an exposure bracketing sequence. Intensity estimates from each bracket are combined using $\mathsf{SNR}^2$ weighting [18] and then converted to a binary rate estimate at the center exposure level of 1 ppp. The proposed policy demonstrates an average reduction in detections of $42\%$ as compared to no inhibition. Fig. 5(f,g) evaluates a single exposure level (1.0 ppp) which is a more challenging scenario, yet the proposed policy still reduces detections by $14\%$ at $\mathrm{SSIM} = 0.7$ . See supplement S 3.4 for more examples and S 3.5 for simulations of high dynamic range images. + +Edge Detection: The BSDS500 dataset with ground truth boundaries [1] was used to study energy-efficient edge detection via photon inhibition. Binary rate images were processed by pre-trained holistically-nested edge detection (HED) [53] with the resulting edge maps compared to ground truth by the structured edge detection toolbox [13]. Fig. 6 shows the optimal image scale (OIS) F-score versus the average detections per pixel. Interpolated curves (not shown) allow for translating along horizontal lines of equal task performance to assess differences in avalanche energy. At low photon counts the proposed edge-enhancing policy + +![](images/e936f7b8c80d7ef1abc12fc2a2cd0ef2f6ec60c78629b4c18cbb905d34f9b351.jpg) + +![](images/2c644c6241f580648fde24f36ba2025a5f18aee034b40212bec9a227fc45f8cd.jpg) + +![](images/37be8622632bf095fe7076254eed99d5c15995de3301d2702bb838cef731556b.jpg) +Fig. 5: Power-efficient static imaging via inhibition. (a) Images from an exposure bracketing sequence (average exposure of 0.1, 1.0, 10.0 photons per pixel (ppp)) using clocked recharge without (conventional) and with inhibition. (b) the distribution of measurements based on pixel intensity with $H$ at 1.0 ppp. (c) image quality (SSIM) at equal detections/pixel. (d-g) Handcrafted policies are tested over 20 images and assessed by reductions in detections ( $D$ ) at equal SSIM. The top row (d),(e) use exposure bracketing; (f),(g) use a single exposure of 1.0 ppp. (d),(f) at SSIM=0.7 and (e),(g) at SSIM=0.8. The box shows quartiles with the center line at the median. The proposed policy, $P_{cr}$ , is a $3 \times 3$ spatial kernel that emphasizes the center pixel ( $\times 8$ ) and includes the 8 neighbors ( $\times 1$ ) (see the supplement for policy details). (h) average inhibition patterns for each exposure time. The top most pattern inhibits the brightest pixels only (maximum of $\sim 60\%$ inhibition, primarily in the sky). For the longest exposure time the inhibition pattern allocates measurements primarily to the darkest pixels. + +![](images/8ac98e90c44360cebbc62ff83dd1ec83dae1631a2f739a213827d5e564d21f1d.jpg) + +![](images/c0cae04302e9f6a845ac9aa869153b185b9068b50cdc3dc96430a6941404f2c3.jpg) + +![](images/f286949a31cd84d68bfccd8430fe1b23e8343da6f3ad42f7877c170ea4b243d7.jpg) + +![](images/f1ceba2f6ec65d61dbe7e96d9bb0f0e5aa9832265988c45cbce2bddce4f8a6c5.jpg) + +![](images/2976340e63e3a778154ddb56dc0a9359323862d7c5fdafe89b3767c77ce2c246.jpg) +Fig. 6: Energy-efficient edge detection by inhibition: Edge detection F-scores averaged over 19 images versus average detections/pixel. Horizontal arrows show reduction in detections enabled by inhibition at equal task performance. At and beyond $30\mathrm{D / pix}$ . no inhibition and the proposed policy are nearly equivalent and plateau at $200\mathrm{D / pix}$ . The edge detector returns $F = 0.813$ on the original images of this set. + +demonstrates a $30\%$ reduction in detections on single exposure time captures. This policy extends the on-sensor calculation approach by calculating two scores, the conventional $3 \times 3$ Laplacian [50] as $S_{1}$ and a $3 \times 3$ averaging filter as $S_{2}$ . The final inhibition decision is the Boolean operation of these scores as $((\eta_{1} < S_{1} < \eta_{2}) \land (S_{2} > \eta_{3})) \lor (S_{2} > \eta_{4})$ . The Laplacian policy alone performs poorly as photons in dim regions with minimal spatial variation are inhibited - disabling dim pixels is energy inefficient. Complete policy descriptions are in the supplement. + +# 7 Experiments on Dynamic Scenes + +Many real-world scenes contain significant motion even with the high frame rate of a SPAD camera. Burst reconstruction algorithms yield high-quality images from sequences of binary frames [34,35] - we investigate their compatibility with the inhibited photon detection data. In particular, we focus on the saturation look-ahead policy of Fig. 4(a), applied independently at every pixel, and therefore an example of an adaptive single-pixel temporal policy. + +To avoid losing salient information under motion, the inhibition policy must limit periods of extended dead time. In the context of the saturation look-ahead policy this limits the total exposure length of the bracketing sequence, since we implicitly assume flux to be constant within each sequence. In our experiments, we use a Fibonacci bracketing [20] sequence $T \coloneqq \{1, 1, 2, 3, 5, 8, 13, 21\}^6$ , denoted in the units of a single minimal exposure time. Every measured bracket is converted to a maximum-likelihood estimate of flux, which is then supplied to the quanta burst photography algorithm [34]. The flux estimator from bracketed measurements is described in detail in supplement Sec. S7.2. + +Results with the SwissSPAD2 sensor: The SwissSPAD2 sensor [49] is a prototype SPAD pixel array that can produce binary frames at a rate up to 97,700 FPS, with a resolution of $512 \times 256$ . In our experiments, we use binary frames captured directly (without inhibition) by the SPAD array as reference data and emulate on-sensor saturation look-ahead inhibition in software. As a pre-processing step, measurements at hot pixels are replaced with their nearest neighbors. + +Fig. 7 shows the results of burst reconstruction under three lighting conditions. The raw data is a sequence of $>580,000$ binary frames with scene radiance increasing rapidly by orders of magnitude (Fig. 7(b)), from $<1$ lux to $>4,000$ lux, measured separately with a light meter. For each of 47 equally-spaced keyframes, centered windows of 12,000 binary frames are extracted and processed as described above. Results for the full sequence can be found in the supplementary material. A static inhibition policy of regular sub-sampling (dropping 9 out of every 10 frames) is also applied, which yields a fixed $90\%$ reduction in both measurements and photon detections under all lighting conditions. Other subsampling factors are discussed in the supplement. + +![](images/bcf274d6a24deaa3ce2c5d1e440ce97657328e2c77d381e78e05599f0ff2b257.jpg) +Fig. 7: Adaptive policies on video sequences enable stronger inhibition, preserve low-light details and, in bright-light, decouple flux and detection energy. (a) Burst reconstructions [34] for three keyframes with varying light levels: the top and bottom row differ by $\approx 7$ stops. [Images show the detection rates $Y$ and are further gamma-compressed ( $\gamma = 0.4$ ). Complete results included in the supplement.] Left column results are from the original binary frames without inhibition, and the right column after sub-sampling $10 \times$ (a fixed $90\%$ inhibition). Middle column represents exposure bracketing combined with saturation look-ahead (see Fig. 4a for description). Under strong light (top row) the results are reasonable with both methods. However, plain sub-sampling loses details in lower light: notice the furniture and a person's outline in the middle & bottom rows, respectively. Inhibition is instead adaptive to flux. (b) Average exposure level for each keyframe in the sequence. (c,d) Per-keyframe and cumulative detection counts - inhibition ultimately results in fewer photons being detected over the whole sequence. (e) Number of measurements taken for each keyframe; reductions may be translated to energy savings during read-out. Plots in (c,d,e) are sub-sampled for clarity, and crossover points are marked by green arrows. + +The top row in Fig. 7(a) shows that under strong light a large fraction of photons ( $>90\%$ ) can be inhibited through the saturation look-ahead policy and still result in good image quality after burst reconstruction, thus spreading photon detections over a longer period of time to reduce avalanche power. Even simple sub-sampling yields good results in bright light, and may work well under controlled illumination conditions. However, the images in the middle and + +bottom rows illustrate that this static inhibition policy results in excessive signal loss in lower light leading to a loss of details. Sub-sampling may be enhanced by adjusting exposure time and/or the sub-sampling factor in response to global flux but cannot simultaneously optimize for different light levels within a single frame: over-exposed regions may have clipping artifacts and under-exposed regions are prone to motion blur from incorrect burst reconstruction. The saturation look-ahead policy does adapt to local flux, and allocates relatively more detections to dim regions (see Fig. 1(d)). It thus decouples detection energy from flux (Fig. 7(c); cf. Fig. 1(a)), and results in cumulatively fewer detections than sub-sampling (Fig. 7(d)) due to it being more aggressive in strong light. + +Apart from photography, the burst-reconstructed images can be used in computer vision applications, where even stronger inhibition is possible depending on the noise- or blur-robustness of the vision algorithm. Fig. 1(d) shows successful object detection with the YOLOv8 algorithm [29] applied to a burst-reconstructed image, with approximately $95\%$ photon detections inhibited. + +# 8 Limitations and Future Outlook + +Implementation costs. Our focus in this paper is on reducing energy consumption due to avalanches. Although avalanches contribute significant energy as compared to on-sensor computations [2, 42, 47], an important next step is to design a holistic model that includes the energy consumption of computations and readout (leveraging Eq. 7). Our in-pixel computations — power-of-two multiplications which simplify to bit shifts, small spatio-temporal kernels no larger than $3 \times 3 \times 4$ — are designed to be lightweight. Fortunately, computational SPAD imagers [2] with in-pixel memory and compute have recently been proposed, with $4 \times 4$ block of pixels having a 32-bit CPU and over 10 kbits of memory. See the supplement for an estimate of the required circuitry. + +More complete models. The noise model used for efficiency metrics only accounts for photon Poisson noise and the quantization noise of Bernoulli samples. Expanding the noise model to include uncorrected pixel sensitivity variations, crosstalk, and afterpulsing may improve performance [4] by allocating measurements with an awareness of practical sensor limitations. Inhibition modulates the number of Bernoulli trials. An unbiased estimator for data-dependent stopping of Bernoulli trials is known [23], yet is not applied in this work since it does not precisely match our situation. Our simulations (SSIM) suggest that bias is less significant than image noise, yet future analytical studies are needed over a range of conditions. + +Generalization to other tasks. To generalize the proposed approaches to a variety of vision tasks, task-specific quality metrics must be defined to pose optimization problems for each task. As an example, for the image reconstruction task, Suppl. Sec. 3.1 shows an analytical optimization using the MSE metric when constrained by photon detections. Other tasks could be approached similarly, yet may need to be optimized empirically. + +# Acknowledgments + +The authors acknowledge the Minnesota Supercomputing Institute (MSI) at the University of Minnesota for providing resources. A.I. was supported in part by NSF ECCS-2138471. S.G and M.G. were supported in part by NSF CAREER award 1943149, NSF award CNS-2107060, and Wisconsin Alumni Research Foundation (WARF). + +# References + +1. Arbeláez, P., Maire, M., Fowlkes, C., Malik, J.: Contour Detection and Hierarchical Image Segmentation. IEEE Transactions on Pattern Analysis and Machine Intelligence 33(5), 898-916 (May 2011). https://doi.org/10.1109/TPAMI.2010.161 +2. Ardelean, A.: Computational Imaging SPAD Cameras. Ph.D. thesis, EPFL (2023). https://doi.org/10.5075/epfl-thesis-9501 +3. Barlow, H.B.: Summation and inhibition in the frog's retina. The Journal of physiology 119(1), 69 (1953) +4. Bian, L., Song, H., Peng, L., Chang, X., Yang, X., Horstmeyer, R., Ye, L., Zhu, C., Qin, T., Zheng, D., Zhang, J.: High-resolution single-photon imaging with physics-informed deep learning. Nature Communications 14(1), 5902 (Sep 2023). https://doi.org/10.1038/s41467-023-41597-9 +5. Boso, G., Buttafava, M., Villa, F., Tosi, A.: Low-cost and compact single-photon counter based on a CMOS SPAD smart pixel. IEEE Photonics Technology Letters 27(23), 2504-2507 (2015) +6. Canon Inc.: Canon Launches MS-500 - The World's First Ultra-High-Sensitivity Interchangeable-Lens SPAD Sensor Camera. https://www.usa.canon.com/newsroom/2023/20230801-ms500, Canon Press Release 8/1/2023. Accessed 2/25/2024. +7. Carey, S.J., Lopich, A., Barr, D.R., Wang, B., Dudek, P.: A 100,000 fps vision sensor with embedded 535 GOPS/W $256 \times 256$ SIMD processor array. In: 2013 Symposium on VLSI Circuits. pp. C182-C183 (Jun 2013) +8. Chan, S.H.: What Does a One-Bit Quanta Image Sensor Offer? IEEE Transactions on Computational Imaging 8, 770-783 (2022). https://doi.org/10.1109/TCI.2022.3202012 +9. Charbon, E., Bruschini, C., Lee, M.J.: 3D-Stacked CMOS SPAD Image Sensors: Technology and Applications. In: 2018 25th IEEE International Conference on Electronics, Circuits and Systems (ICECS). pp. 1-4 (Dec 2018). https://doi.org/10.1109/ICECS.2018.8617983 +10. Chi, Y., Gnanasambandam, A., Koltun, V., Chan, S.H.: Dynamic low-light imaging with quanta image sensors. In: Computer Vision-ECCV 2020: 16th European Conference, Glasgow, UK, August 23-28, 2020, Proceedings, Part XXI 16. pp. 122-138. Springer (2020) +11. Della Rocca, F.M., Mai, H., Hutchings, S.W., Al Abbas, T., Buckbee, K., Tsiamis, A., Lomax, P., Gyongy, I., Dutton, N.A., Henderson, R.K.: A $128 \times 128$ SPAD motion-triggered time-of-flight image sensor with in-pixel histogram and column-parallel vision processor. IEEE Journal of Solid-State Circuits 55(7), 1762-1775 (2020). https://doi.org/10.1109/JSSC.2020.2993722 + +12. Diamond, J.S.: Inhibitory interneurons in the retina: Types, circuitry, and function. Annual Review of Vision Science 3(1), 1-24 (2017). https://doi.org/10.1146/annurev-vision-102016-061345 +13. Dollar, P., Zitnick, C.L.: Structured Forests for Fast Edge Detection. In: Proceedings of the IEEE International Conference on Computer Vision. pp. 1841-1848 (2013) +14. Dutton, N.A.W., Al Abbas, T., Gyongy, I., Mattioli Della Rocca, F., Henderson, R.K.: High Dynamic Range Imaging at the Quantum Limit with Single Photon Avalanche Diode-Based Image Sensors. Sensors 18(4), 1166 (Apr 2018). https://doi.org/10.3390/s18041166 +15. Fossum, E.R.: Modeling the Performance of Single-Bit and Multi-Bit Quanta Image Sensors. IEEE Journal of the Electron Devices Society 1(9), 166-174 (Sep 2013). https://doi.org/10.1109/JEDS.2013.2284054 +16. Franke, K., Berens, P., Schubert, T., Bethge, M., Euler, T., Baden, T.: Inhibition decorrelates visual feature representations in the inner retina. Nature 542(7642), 439-444 (Feb 2017). https://doi.org/10.1038/nature21394 +17. Gallego, G., Delbrück, T., Orchard, G., Bartolozzi, C., Taba, B., Censi, A., Leutenegger, S., Davison, A.J., Conradt, J., Daniilidis, K., et al.: Event-based vision: A survey. IEEE Transactions on Pattern Analysis and Machine Intelligence 44(1), 154-180 (2020) +18. Gnanasambandam, A., Chan, S.H.: HDR Imaging With Quanta Image Sensors: Theoretical Limits and Optimal Reconstruction. IEEE Transactions on Computational Imaging 6, 1571-1585 (2020). https://doi.org/10.1109/TCI.2020.3041093 +19. Gnanasambandam, A., Chan, S.H.: Exposure-Referred Signal-to-Noise Ratio for Digital Image Sensors. IEEE Transactions on Computational Imaging 8, 561-575 (Jun 2022). https://doi.org/10.1109/TCI.2022.3187657 +20. Gupta, M., Iso, D., Nayar, S.K.: Fibonacci Exposure Bracketing for High Dynamic Range Imaging. In: 2013 IEEE International Conference on Computer Vision. pp. 1473-1480. IEEE, Sydney, Australia (Dec 2013). https://doi.org/10.1109/ICCV.2013.186 +21. Gyongy, I., Dutton, N.A., Henderson, R.K.: Direct time-of-flight single-photon imaging. IEEE Transactions on Electron Devices 69(6), 2794-2805 (2021) +22. Gyongy, I., Martin, G.M., Turpin, A., Ruget, A., Halimi, A., Henderson, R., Leach, J.: High-speed vision with a 3D-stacked SPAD image sensor. In: Advanced Photon Counting Techniques XV. vol. 11721, p. 1172105. SPIE (2021) +23. Haldane, J.B.S.: On a Method Of Estimating Frequencies. Biometrika 33(3), 222-225 (Nov 1945). https://doi.org/10.1093/biomet/33.3.222 +24. Henderson, R.K., Johnston, N., Hutchings, S.W., Gyongy, I., Al Abbas, T., Dutton, N., Tyler, M., Chan, S., Leach, J.: $256 \times 256$ 40nm/90nm CMOS 3D-stacked 120 dB dynamic-range reconfigurable time-resolved SPAD imager. In: 2019 IEEE International Solid-State Circuits Conference-(ISSCC). pp. 106-108. IEEE (2019). https://doi.org/10.1109/ISSCC.2019.8662355 +25. Iams, H., Salzberg, B.: The secondary emission phototube. Proceedings of the Institute of Radio Engineers 23(1), 55-64 (1935) +26. Ingle, A., Seets, T., Buttafava, M., Gupta, S., Tosi, A., Gupta, M., Velten, A.: Passive Inter-Photon Imaging. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 8585–8595 (2021) +27. Ingle, A., Velten, A., Gupta, M.: High Flux Passive Imaging With Single-Photon Sensors. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 6760-6769 (2019) + +28. J Yoshida: Breaking Down iPad Pro 11's LiDAR Scanner. https://www.eetimes.com/breaking-down-ipad-pro-11s-lidar-scanner/, EE Times 6/5/2020. Accessed 5/6/2021. +29. Jocher, G., Chaurasia, A., Qiu, J.: Ultralytics YOLOv8 (Jan 2023), https://github.com/ultralytics/ultralytics, Accessed 3/6/2024 +30. Koppal, S.J., Gkioulekas, I., Young, T., Park, H., Crozier, K.B., Barrows, G.L., Zickler, T.: Toward Wide-Angle Microvision Sensors. IEEE Transactions on Pattern Analysis and Machine Intelligence 35(12), 2982-2996 (Dec 2013). https://doi.org/10.1109/TPAMI.2013.22 +31. Liu, Y., Gutierrez-Barragan, F., Ingle, A., Gupta, M., Velten, A.: Single-Photon Camera Guided Extreme Dynamic Range Imaging. In: Proceedings of the IEEE/CVF Winter Conference on Applications of Computer Vision. pp. 1575-1585 (2022) +32. Ma, J., Masoodian, S., Starkey, D.A., Fossum, E.R.: Photon-number-resolving megapixel image sensor at room temperature without avalanche gain. Optica 4(12), 1474-1481 (Dec 2017). https://doi.org/10.1364/OPTICA.4.001474 +33. Ma, J., Zhang, D., Elgendy, O.A., Masoodian, S.: A 0.19e- rms Read Noise 16.7Mpixel Stacked Quanta Image Sensor With 1.1 Mm-Pitch Backside Illuminated Pixels. IEEE Electron Device Letters 42(6), 891-894 (Jun 2021). https: //doi.org/10.1109/LED.2021.3072842 +34. Ma, S., Gupta, S., Ulku, A.C., Bruschini, C., Charbon, E., Gupta, M.: Quanta burst photography. ACM Transactions on Graphics 39(4), 79:1-79:16 (Aug 2020). https://doi.org/10.1145/3386569.3392470 +35. Ma, S., Mos, P., Charbon, E., Gupta, M.: Burst Vision Using Single-Photon Cameras. In: Proceedings of the IEEE/CVF Winter Conference on Applications of Computer Vision. pp. 5375-5385 (2023) +36. Medin, S.C., Murray-Bruce, J., Castanón, D., Goyal, V.K.: Beyond Binomial and Negative Binomial: Adaptation in Bernoulli Parameter Estimation. IEEE Transactions on Computational Imaging 5(4), 570-584 (Dec 2019). https://doi.org/10.1109/TCI.2019.2913108 +37. Morimoto, K., Iwata, J., Shinohara, M., Sekine, H., Abdelghafar, A., Tsuchiya, H., Kuroda, Y., Tojima, K., Endo, W., Maehashi, Y., Ota, Y., Sasago, T., Maekawa, S., Hikosaka, S., Kanou, T., Kato, A., Tezuka, T., Yoshizaki, S., Ogawa, T., Uehira, K., Ehara, A., Inui, F., Matsuno, Y., Sakurai, K., Ichikawa, T.: 3.2 Megapixel 3D-Stacked Charge Focusing SPAD for Low-Light Imaging and Depth Sensing. In: 2021 IEEE International Electron Devices Meeting (IEDM). pp. 20.2.1–20.2.4 (Dec 2021). https://doi.org/10.1109/IEDM19574.2021.9720605 +38. Morimoto, K., Ardelean, A., Wu, M.L., Ulku, A.C., Antolovic, I.M., Bruschini, C., Charbon, E.: Megapixel time-gated SPAD image sensor for 2D and 3D imaging applications. Optica 7(4), 346-354 (2020). https://doi.org/10.1364/OPTICA.386574 +39. Morimoto, K., Charbon, E.: A scaling law for SPAD pixel miniaturization. Sensors 21(10), 3447 (2021) +40. Ogi, J., Sano, F., Nakata, T., Kubo, Y., Onishi, W., Koswaththaghe, C., Mochizuki, T., Tashiro, Y., Hizu, K., Takatsuka, T., et al.: A $3.06\mu \mathrm{m}$ spad pixel with embedded metal contact and power grid on deep trench pixel isolation for high-resolution photon-counting. In: 2023 International Image Sensor Workshop (IISW) (May 2023) +41. Ogi, J., Takatsuka, T., Hizu, K., Inaoka, Y., Zhu, H., Tochigi, Y., Tashiro, Y., Sano, F., Murakawa, Y., Nakamura, M., Oike, Y.: A 250fps 124dB Dynamic-Range + +SPAD Image Sensor Stacked with Pixel-Parallel Photon Counter Employing SubFrame Extrapolating Architecture for Motion Artifact Suppression. In: 2021 IEEE International Solid-State Circuits Conference (ISSCC). vol. 64, pp. 113-115 (Feb 2021). https://doi.org/10.1109/ISSCC42613.2021.9365977 +42. Ota, Y., Morimoto, K., Sasago, T., Shinohara, M., Kuroda, Y., Endo, W., Maehashi, Y., Maekawa, S., Tsuchiya, H., Abdelahafar, A., Hikosaka, S., Motoyama, M., Tojima, K., Uehira, K., Iwata, J., Inui, F., Matsuno, Y., Sakurai, K., Ichikawa, T.: A 0.37W 143dB-Dynamic-Range 1Mpixel Backside-Illuminated Charge-Focusing SPAD Image Sensor with Pixel-Wise Exposure Control and Adaptive Clocked Recharging. In: 2022 IEEE International Solid-State Circuits Conference (ISSCC). vol. 65, pp. 94-96 (Feb 2022). https://doi.org/10.1109/ISSCC42614.2022.9731644 +43. Po, R., Pediredla, A., Gkioulekas, I.: Adaptive gating for single-photon 3d imaging. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 16354-16363 (2022) +44. Severini, F., Cusini, I., Berretta, D., Pasquinelli, K., Incoronato, A., Villa, F.: SPAD pixel with sub-ns dead-time for high-count rate applications. IEEE Journal of Selected Topics in Quantum Electronics 28(2: Optical Detectors), 1-8 (2021) +45. Shenzhen CM Technology company Ltd: 16 MP MIPI Camera Module with SONY IMX206 sensor. http://www-camera-module.com/product/mipicameramodule/16mp-mipi-camera-module-sony-imx206-sensor.html (Nov 2023) +46. Sundar, V., Ardelean, A., Swedish, T., Brusschini, C., Charbon, E., Gupta, M.: Sodacam: Software-defined cameras via single-photon imaging. In: Proceedings of the IEEE International Conference on Computer Vision (ICCV) (2023) +47. Takatsuka, T., Ogi, J., Ikeda, Y., Hizu, K., Inaoka, Y., Sakama, S., Watanabe, I., Ishikawa, T., Shimada, S., Suzuki, J., Maeda, H., Toshima, K., Nonaka, Y., Yamamura, A., Ozawa, H., Koga, F., Oike, Y.: A 3.36 $\mu$ m-pitch SPAD photon-counting image sensor using clustered multi-cycle clocked recharging technique with intermediate most-significant-bit readout. In: 2023 IEEE Symposium on VLSI Technology and Circuits (VLSI Technology and Circuits). pp. 1-2 (Jun 2023). https://doi.org/10.23919/VLSITechnologyandCir57934.2023.10185241 +48. Tilmon, B., Sun, Z., Koppal, S.J., Wu, Y., Evangelidis, G., Zahreddine, R., Krishnan, G., Ma, S., Wang, J.: Energy-Efficient Adaptive 3D Sensing. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 5054-5063 (2023) +49. Ulku, A.C., Bruschini, C., Antolovic, I.M., Kuo, Y., Ankri, R., Weiss, S., Michalet, X., Charbon, E.: A $512 \times 512$ SPAD Image Sensor With Integrated Gating for Widefield FLIM. IEEE Journal of Selected Topics in Quantum Electronics 25(1), 1-12 (Jan 2019). https://doi.org/10.1109/jstqe.2018.2867439 +50. Wang, X.: Laplacian operator-based edge detectors. IEEE Transactions on Pattern Analysis and Machine Intelligence 29(5), 886-890 (2007) +51. Wang, Z., Bovik, A.C., Sheikh, H.R., Simoncelli, E.P.: Image quality assessment: From error visibility to structural similarity. IEEE Transactions on Image Processing 13(4), 600-612 (2004) +52. Wayne, M.A., Bienfang, J.C., Migdall, A.L.: Low-noise photon counting above $100 \times 106$ counts per second with a high-efficiency reach-through single-photon avalanche diode system. Applied Physics Letters 118(13) (2021) +53. Xie, S., Tu, Z.: Holistically-nested edge detection. In: Proceedings of IEEE International Conference on Computer Vision. pp. 1395-1403 (2015) + +54. Xu, Y., Lu, J., Wu, Z.: A compact high-speed active quenching and recharging circuit for SPAD detectors. IEEE Photonics Journal 12(5), 1-8 (2020) +55. Yang, F., Lu, Y.M., Sbaiz, L., Vetterli, M.: Bits from photons: Oversampled image acquisition using binary poisson statistics. IEEE Transactions on image processing 21(4), 1421-1436 (2011) \ No newline at end of file diff --git a/2024/Photon Inhibition for Energy-Efficient Single-Photon Imaging/images.zip b/2024/Photon Inhibition for Energy-Efficient Single-Photon Imaging/images.zip new file mode 100644 index 0000000000000000000000000000000000000000..92c1df06ccbb8024ef2465c66759b70e96d8e36c --- /dev/null +++ b/2024/Photon Inhibition for Energy-Efficient Single-Photon Imaging/images.zip @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:7ee3179fedb56219f2a28c6a349c643de3dc1e202fee29fc803188261332b3c0 +size 481858 diff --git a/2024/Photon Inhibition for Energy-Efficient Single-Photon Imaging/layout.json b/2024/Photon Inhibition for Energy-Efficient Single-Photon Imaging/layout.json new file mode 100644 index 0000000000000000000000000000000000000000..55b23bd678cfdd5ff8999200349bfbed518cd795 --- /dev/null +++ b/2024/Photon Inhibition for Energy-Efficient Single-Photon Imaging/layout.json @@ -0,0 +1,10391 @@ +{ + "pdf_info": [ + { + "para_blocks": [ + { + "bbox": [ + 173, + 111, + 440, + 148 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 173, + 111, + 440, + 148 + ], + "spans": [ + { + "bbox": [ + 173, + 111, + 440, + 148 + ], + "type": "text", + "content": "Photon Inhibition for Energy-Efficient Single-Photon Imaging" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 138, + 167, + 474, + 181 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 167, + 474, + 181 + ], + "spans": [ + { + "bbox": [ + 138, + 167, + 474, + 181 + ], + "type": "text", + "content": "Lucas J. Koerner1, Shantanu Gupta2, Atul Ingle3, and Mohit Gupta2" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 132, + 189, + 479, + 255 + ], + "type": "list", + "angle": 0, + "index": 5, + "blocks": [ + { + "bbox": [ + 132, + 189, + 479, + 209 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 189, + 479, + 209 + ], + "spans": [ + { + "bbox": [ + 132, + 189, + 479, + 209 + ], + "type": "inline_equation", + "content": "^{1}" + }, + { + "bbox": [ + 132, + 189, + 479, + 209 + ], + "type": "text", + "content": " Department of Electrical and Computer Engineering, University of St. Thomas, St. Paul, MN, 55105, USA koerner.lucas@stthomas.edu" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 132, + 212, + 479, + 232 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 212, + 479, + 232 + ], + "spans": [ + { + "bbox": [ + 132, + 212, + 479, + 232 + ], + "type": "inline_equation", + "content": "^{2}" + }, + { + "bbox": [ + 132, + 212, + 479, + 232 + ], + "type": "text", + "content": " Department of Computer Sciences, University of Wisconsin-Madison, Madison, WI, 53706, USA {sgupta,mohitg}@cs.wisc.edu" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 132, + 233, + 478, + 255 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 233, + 478, + 255 + ], + "spans": [ + { + "bbox": [ + 132, + 233, + 478, + 255 + ], + "type": "inline_equation", + "content": "^{3}" + }, + { + "bbox": [ + 132, + 233, + 478, + 255 + ], + "type": "text", + "content": " Department of Computer Science, Portland State University, Portland, OR, 97201, USA \ningle2@pdx.edu" + } + ] + } + ], + "index": 4 + } + ], + "sub_type": "text" + }, + { + "bbox": [ + 160, + 285, + 452, + 526 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 160, + 285, + 452, + 526 + ], + "spans": [ + { + "bbox": [ + 160, + 285, + 452, + 526 + ], + "type": "text", + "content": "Abstract. Single-photon cameras (SPCs) are emerging as sensors of choice for various challenging imaging applications. One class of SPCs based on the single-photon avalanche diode (SPAD) detects individual photons using an avalanche process; the raw photon data can then be processed to extract scene information under extremely low light, high dynamic range, and rapid motion. Yet, single-photon sensitivity in SPADs comes at a cost — each photon detection consumes more energy than that of a CMOS camera. This avalanche power significantly limits sensor resolution and could restrict widespread adoption of SPAD-based SPCs. We propose a computational-imaging approach called photon inhibition to address this challenge. Photon inhibition strategically allocates detections in space and time based on downstream inference task goals and resource constraints. We develop lightweight, on-sensor computational inhibition policies that use past photon data to disable SPAD pixels in real-time, to select the most informative future photons. As case studies, we design policies tailored for image reconstruction and edge detection, and demonstrate, both via simulations and real SPC captured data, considerable reduction in photon detections (over " + }, + { + "bbox": [ + 160, + 285, + 452, + 526 + ], + "type": "inline_equation", + "content": "90\\%" + }, + { + "bbox": [ + 160, + 285, + 452, + 526 + ], + "type": "text", + "content": " of photons) while maintaining task performance metrics. Our work raises the question of \"which photons should be detected?\", and paves the way for future energy-efficient single-photon imaging. Source code for our experiments is available at https://wisionlab.com/project/inhibition." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 160, + 536, + 438, + 548 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 160, + 536, + 438, + 548 + ], + "spans": [ + { + "bbox": [ + 160, + 536, + 438, + 548 + ], + "type": "text", + "content": "Keywords: single-photon imaging " + }, + { + "bbox": [ + 160, + 536, + 438, + 548 + ], + "type": "inline_equation", + "content": "\\cdot" + }, + { + "bbox": [ + 160, + 536, + 438, + 548 + ], + "type": "text", + "content": " SPADs " + }, + { + "bbox": [ + 160, + 536, + 438, + 548 + ], + "type": "inline_equation", + "content": "\\cdot" + }, + { + "bbox": [ + 160, + 536, + 438, + 548 + ], + "type": "text", + "content": " energy-efficient vision" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 132, + 568, + 230, + 582 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 568, + 230, + 582 + ], + "spans": [ + { + "bbox": [ + 132, + 568, + 230, + 582 + ], + "type": "text", + "content": "1 Introduction" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 130, + 594, + 482, + 666 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 594, + 482, + 666 + ], + "spans": [ + { + "bbox": [ + 130, + 594, + 482, + 666 + ], + "type": "text", + "content": "From vacuum tube-based single-photon detectors invented in the early 20th century [25], to 3D depth sensing LiDAR cameras found in today's smartphones [28], single-photon camera (SPC) technology has come a long way in terms of pixel resolution and commercial availability for a variety of applications. Thanks to CMOS-compatible single-photon avalanche diode (SPAD) arrays, SPCs are being increasingly used not only in niche fields such as scientific imaging and biomedical" + } + ] + } + ], + "index": 9 + } + ], + "discarded_blocks": [], + "page_size": [ + 612, + 792 + ], + "page_idx": 0 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 132, + 123, + 272, + 213 + ], + "blocks": [ + { + "bbox": [ + 133, + 114, + 146, + 123 + ], + "lines": [ + { + "bbox": [ + 133, + 114, + 146, + 123 + ], + "spans": [ + { + "bbox": [ + 133, + 114, + 146, + 123 + ], + "type": "text", + "content": "(a)" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 132, + 123, + 272, + 213 + ], + "lines": [ + { + "bbox": [ + 132, + 123, + 272, + 213 + ], + "spans": [ + { + "bbox": [ + 132, + 123, + 272, + 213 + ], + "type": "image", + "image_path": "569a1e0f3cf9c3fd9630c153c427ef6c15f59798b072886a2421b02be122130c.jpg" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 276, + 114, + 288, + 124 + ], + "lines": [ + { + "bbox": [ + 276, + 114, + 288, + 124 + ], + "spans": [ + { + "bbox": [ + 276, + 114, + 288, + 124 + ], + "type": "text", + "content": "(c)" + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 133, + 213, + 144, + 222 + ], + "lines": [ + { + "bbox": [ + 133, + 213, + 144, + 222 + ], + "spans": [ + { + "bbox": [ + 133, + 213, + 144, + 222 + ], + "type": "text", + "content": "(b)" + } + ] + } + ], + "index": 7, + "angle": 0, + "type": "image_caption" + } + ], + "index": 3 + }, + { + "type": "image", + "bbox": [ + 277, + 130, + 383, + 204 + ], + "blocks": [ + { + "bbox": [ + 277, + 130, + 383, + 204 + ], + "lines": [ + { + "bbox": [ + 277, + 130, + 383, + 204 + ], + "spans": [ + { + "bbox": [ + 277, + 130, + 383, + 204 + ], + "type": "image", + "image_path": "085285b6037fa00abba45c15cafcfeaa09725b66253ac9ed2bc26f4bea002c45.jpg" + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "image_body" + } + ], + "index": 5 + }, + { + "type": "image", + "bbox": [ + 392, + 130, + 477, + 193 + ], + "blocks": [ + { + "bbox": [ + 392, + 130, + 477, + 193 + ], + "lines": [ + { + "bbox": [ + 392, + 130, + 477, + 193 + ], + "spans": [ + { + "bbox": [ + 392, + 130, + 477, + 193 + ], + "type": "image", + "image_path": "fbe062150553a0bf3eca8f4afb5fa72dee4aaa1edcf93b24649106b68171e51e.jpg" + } + ] + } + ], + "index": 6, + "angle": 0, + "type": "image_body" + } + ], + "index": 6 + }, + { + "type": "image", + "bbox": [ + 136, + 222, + 276, + 320 + ], + "blocks": [ + { + "bbox": [ + 136, + 222, + 276, + 320 + ], + "lines": [ + { + "bbox": [ + 136, + 222, + 276, + 320 + ], + "spans": [ + { + "bbox": [ + 136, + 222, + 276, + 320 + ], + "type": "image", + "image_path": "83a5d838b33a448d72f6c75b24a21f7f97476f5d7b1264981d0fa51783af66ff.jpg" + } + ] + } + ], + "index": 8, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 276, + 213, + 288, + 222 + ], + "lines": [ + { + "bbox": [ + 276, + 213, + 288, + 222 + ], + "spans": [ + { + "bbox": [ + 276, + 213, + 288, + 222 + ], + "type": "text", + "content": "(d)" + } + ] + } + ], + "index": 9, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 130, + 327, + 482, + 437 + ], + "lines": [ + { + "bbox": [ + 130, + 327, + 482, + 437 + ], + "spans": [ + { + "bbox": [ + 130, + 327, + 482, + 437 + ], + "type": "text", + "content": "Fig. 1: Photon inhibition for resource-efficient passive SPAD imaging. (a) Unlike conventional CMOS cameras (CIS), the energy consumption in SPAD cameras increases with scene brightness, severely limiting the applicability of high resolution SPAD cameras in resource-constrained applications. (b) We expand the conventional imaging pipeline to incorporate \"inhibition\" that electronically enables or disables individual pixels to limit bandwidth and power consumption. Our method relies on lightweight mathematical operations called \"inhibition policies\" that update the inhibition patterns based on the history of photon detections. Inhibition policies can be optimized for image SNR or for downstream vision tasks.(c,d) Object detection, a high-level vision task, is successful even with a large fraction of photons inhibited." + } + ] + } + ], + "index": 13, + "angle": 0, + "type": "image_caption" + } + ], + "index": 8 + }, + { + "type": "image", + "bbox": [ + 277, + 221, + 387, + 311 + ], + "blocks": [ + { + "bbox": [ + 277, + 221, + 387, + 311 + ], + "lines": [ + { + "bbox": [ + 277, + 221, + 387, + 311 + ], + "spans": [ + { + "bbox": [ + 277, + 221, + 387, + 311 + ], + "type": "image", + "image_path": "dc6dea98ffdf6a07b3aec81f67e04c50b0a005874003987e8430f359261cc3a5.jpg" + } + ] + } + ], + "index": 10, + "angle": 0, + "type": "image_body" + } + ], + "index": 10 + }, + { + "type": "image", + "bbox": [ + 391, + 241, + 481, + 310 + ], + "blocks": [ + { + "bbox": [ + 399, + 223, + 473, + 231 + ], + "lines": [ + { + "bbox": [ + 399, + 223, + 473, + 231 + ], + "spans": [ + { + "bbox": [ + 399, + 223, + 473, + 231 + ], + "type": "text", + "content": "YOLOv8 Object Detection" + } + ] + } + ], + "index": 11, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 391, + 241, + 481, + 310 + ], + "lines": [ + { + "bbox": [ + 391, + 241, + 481, + 310 + ], + "spans": [ + { + "bbox": [ + 391, + 241, + 481, + 310 + ], + "type": "image", + "image_path": "afe810e526862fbe033a7bc4a3153b6205ec818e7b3f981da0caea6b64b89de1.jpg" + } + ] + } + ], + "index": 12, + "angle": 0, + "type": "image_body" + } + ], + "index": 12 + }, + { + "bbox": [ + 130, + 461, + 482, + 545 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 461, + 482, + 545 + ], + "spans": [ + { + "bbox": [ + 130, + 461, + 482, + 545 + ], + "type": "text", + "content": "microscopy [49], but more widely for consumer photography [37]. SPAD-based SPCs have recently been fabricated into kilo-to-megapixel format arrays that are now commercially available [28, 37]. The extreme sensitivity and high speed can benefit passive low-light computer vision tasks, particularly in the presence of rapid scene or camera motion [35], enable wide dynamic-range imaging [14, 31] and photon-starved active imaging applications such as 3D imaging (LiDAR) [21] and fluorescence microscopy [49]." + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 130, + 546, + 482, + 666 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 546, + 482, + 666 + ], + "spans": [ + { + "bbox": [ + 130, + 546, + 482, + 666 + ], + "type": "text", + "content": "SPAD camera pixels detect individual photons with extremely high frame rates by exploiting avalanche multiplication. On one hand, being able to detect individual photons opens up new possibilities and capabilities for computer vision systems. On the other hand, this also presents a unique challenge: Every photon-induced avalanche comes with a non-negligible energy cost, which is a challenge exclusive to SPAD-based cameras. This flux-dependent photon detection power is a significant fraction of total power consumption in today's SPAD cameras and impedes further increases in their spatial resolution [24,37,41,42,47]. For example, extrapolating the avalanche power of a recent SPAD sensor [47] to " + }, + { + "bbox": [ + 130, + 546, + 482, + 666 + ], + "type": "inline_equation", + "content": "\\sim 10" + }, + { + "bbox": [ + 130, + 546, + 482, + 666 + ], + "type": "text", + "content": " 's of megapixels format predicts a power consumption of several watts in" + } + ] + } + ], + "index": 15 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "spans": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "type": "text", + "content": "2" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 166, + 91, + 235, + 100 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 166, + 91, + 235, + 100 + ], + "spans": [ + { + "bbox": [ + 166, + 91, + 235, + 100 + ], + "type": "text", + "content": "L. Koerner et al." + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 1 + }, + { + "para_blocks": [ + { + "bbox": [ + 130, + 115, + 479, + 140 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 115, + 479, + 140 + ], + "spans": [ + { + "bbox": [ + 130, + 115, + 479, + 140 + ], + "type": "text", + "content": "bright light" + }, + { + "bbox": [ + 130, + 115, + 479, + 140 + ], + "type": "inline_equation", + "content": "^{4}" + }, + { + "bbox": [ + 130, + 115, + 479, + 140 + ], + "type": "text", + "content": " (as illustrated in Fig. 1(a)), far exceeding the power consumption of modern CMOS image sensors (CIS) of around " + }, + { + "bbox": [ + 130, + 115, + 479, + 140 + ], + "type": "inline_equation", + "content": "300\\mathrm{mW}" + }, + { + "bbox": [ + 130, + 115, + 479, + 140 + ], + "type": "text", + "content": " [45]." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 130, + 141, + 482, + 297 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 141, + 482, + 297 + ], + "spans": [ + { + "bbox": [ + 130, + 141, + 482, + 297 + ], + "type": "text", + "content": "We address this problem by being selective about which photons are detected on the SPAD sensor while maintaining good performance on various computer vision tasks. To reduce avalanche power consumption and decouple it from photon flux, we propose a computational imaging technique called photon inhibition where individual SPAD pixels are disabled adaptively based on previous photon detections in their spatio-temporal neighborhoods. Electrically disabling pixels prevents photon detections to inhibit any avalanche power or subsequent processing. We implement lightweight on-sensor computations called inhibition policies (Fig. 1(b)) to determine, in real-time and at single-pixel and single-frame granularity, which SPAD pixels to enable or disable. Our method is inspired by retinal pre-processing of the human visual system where retinal neurons aggregate photon information over small spatio-temporal neighborhoods to cause neighboring retinal cells to become less sensitive to incident light [12, 16].5" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 130, + 298, + 482, + 501 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 298, + 482, + 501 + ], + "spans": [ + { + "bbox": [ + 130, + 298, + 482, + 501 + ], + "type": "text", + "content": "Given that SPADs introduce a new challenge of flux-dependent power consumption, we establish, from first principles, novel energy-aware imaging performance metrics for resource-constrained single-photon imaging. Based on these metrics, we design families of inhibition policies that distribute photon detections in space and time based on imaging / vision task goals and energy consumption constraints. A critical consideration in the design of inhibition policies stems from the observation that these policies are meant to control (enable / disable) the sensor (Fig. 1(b)). Therefore, it is essential for these policies to be extremely lightweight since these need to be implemented on sensor with very limited compute and memory resources. Furthermore, inhibition policies must execute at ultra-low latency to keep up with high-speed photon detections (reaching up to " + }, + { + "bbox": [ + 130, + 298, + 482, + 501 + ], + "type": "inline_equation", + "content": "100\\mathrm{kHz}" + }, + { + "bbox": [ + 130, + 298, + 482, + 501 + ], + "type": "text", + "content": "). Fortunately, since the raw data output from a SPAD-based SPC consists of binary-valued image frames, SPCs are naturally suited to real-time calculations at the image sensor plane under tight compute and memory budgets. The proposed inhibition policies are lightweight, requiring only simple arithmetic and Boolean operations computed over local spatio-temporal neighborhoods, and thus amenable to in-pixel implementation [2, 7]." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 130, + 502, + 482, + 587 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 502, + 482, + 587 + ], + "spans": [ + { + "bbox": [ + 130, + 502, + 482, + 587 + ], + "type": "text", + "content": "In simulations and real experimental data, we show that our inhibition policies allocate photon detections to sensor pixels in a way that reduces detection energy for a given accuracy level of various vision tasks. Our results show consequential energy savings when compared to a conventional capture scheme for tasks of (i) image reconstruction: " + }, + { + "bbox": [ + 130, + 502, + 482, + 587 + ], + "type": "inline_equation", + "content": "42\\%" + }, + { + "bbox": [ + 130, + 502, + 482, + 587 + ], + "type": "text", + "content": " fewer photon detections at equal image quality; (ii) edge detection: an edge sensitive inhibition policy reduces detections by " + }, + { + "bbox": [ + 130, + 502, + 482, + 587 + ], + "type": "inline_equation", + "content": "30\\%" + }, + { + "bbox": [ + 130, + 502, + 482, + 587 + ], + "type": "text", + "content": " at equal F-score; and (iii) YOLOv8 object detection: remains success-" + } + ] + } + ], + "index": 5 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 197, + 91, + 447, + 102 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 197, + 91, + 447, + 102 + ], + "spans": [ + { + "bbox": [ + 197, + 91, + 447, + 102 + ], + "type": "text", + "content": "Photon Inhibition for Energy-Efficient Single-Photon Imaging" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 474, + 92, + 480, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 474, + 92, + 480, + 100 + ], + "spans": [ + { + "bbox": [ + 474, + 92, + 480, + 100 + ], + "type": "text", + "content": "3" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 133, + 598, + 481, + 643 + ], + "type": "page_footnote", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 598, + 481, + 643 + ], + "spans": [ + { + "bbox": [ + 133, + 598, + 481, + 643 + ], + "type": "text", + "content": "It has been shown, perhaps counter-intuitively, that SPADs do not saturate even under extremely bright conditions [26, 27]. Therefore, SPADs are not restricted to low-flux environments, but are being considered for vision applications across a wide dynamic range of lighting conditions (e.g., from a dark tunnel to bright sunlight)." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 135, + 643, + 481, + 665 + ], + "type": "page_footnote", + "angle": 0, + "lines": [ + { + "bbox": [ + 135, + 643, + 481, + 665 + ], + "spans": [ + { + "bbox": [ + 135, + 643, + 481, + 665 + ], + "type": "text", + "content": "5 We borrow the term \"inhibition\" from the phenomenon of \"lateral inhibition\" found in biological vision systems [3]." + } + ] + } + ], + "index": 7 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 2 + }, + { + "para_blocks": [ + { + "bbox": [ + 130, + 115, + 479, + 175 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 115, + 479, + 175 + ], + "spans": [ + { + "bbox": [ + 130, + 115, + 479, + 175 + ], + "type": "text", + "content": "ful with " + }, + { + "bbox": [ + 130, + 115, + 479, + 175 + ], + "type": "inline_equation", + "content": "95\\%" + }, + { + "bbox": [ + 130, + 115, + 479, + 175 + ], + "type": "text", + "content": " of photons inhibited under camera motion of a real-world SPC (Fig. 1(c,d)). Through experiments with photon streams captured using real SPAD camera hardware over a wide range of illumination conditions, we show that our proposed inhibition policies preserve low-light details and, in brightlight, decouple flux and detection energy." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 130, + 178, + 480, + 285 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 178, + 480, + 285 + ], + "spans": [ + { + "bbox": [ + 130, + 178, + 480, + 285 + ], + "type": "text", + "content": "Scope and Limitations: There are several competing image sensor technologies today that resolve single-photon while capturing binary-valued raw frames at rates exceeding thousands of frames/second. SPCs based on \"jots\" [32] that do not rely on avalanche multiplication do not suffer from flux-dependent power consumption as SPADs. Jots are an exciting technology, especially in scenarios that require high resolution and high dynamic range imaging [33]. In this paper, we focus on SPADs, due to their rapidly rising availability and commercial interest [6, 28], and benefits over conventional CMOS sensors, both in low-light and bright scenes for a variety of computer vision tasks [10, 35]." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 130, + 286, + 481, + 382 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 286, + 481, + 382 + ], + "spans": [ + { + "bbox": [ + 130, + 286, + 481, + 382 + ], + "type": "text", + "content": "Our goals in this paper are to (a) raise the question of \"which photons should be detected?\" in the context of energy-efficient single-photon imaging, (b) establish a design space and metrics to evaluate various inhibition policies, and (c) propose plausible policies that respect practical hardware limitations for future on-chip implementation. We emphasize that the inhibition policies proposed in this paper are not necessarily optimal. This work is just a first step towards demonstrating that it is possible to achieve high performance with SPADs, while maintaining low power consumption via photon inhibition." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 132, + 400, + 237, + 414 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 400, + 237, + 414 + ], + "spans": [ + { + "bbox": [ + 132, + 400, + 237, + 414 + ], + "type": "text", + "content": "2 Related Work" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 130, + 426, + 480, + 569 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 426, + 480, + 569 + ], + "spans": [ + { + "bbox": [ + 130, + 426, + 480, + 569 + ], + "type": "text", + "content": "Hardware approaches for reduced energy consumption: There is a strong dependence of SPAD power consumption on the pixel size — the smaller the pixel, the lower the avalanche energy [39]. Although recent developments in SPAD pixel technology have reduced pixel sizes to below " + }, + { + "bbox": [ + 130, + 426, + 480, + 569 + ], + "type": "inline_equation", + "content": "4\\mu \\mathrm{m}" + }, + { + "bbox": [ + 130, + 426, + 480, + 569 + ], + "type": "text", + "content": " [40], avalanche energy still contributes a large fraction of the total power consumption in a SPAD sensor [44]. SPAD design optimizations have reduced the charge per avalanche by RF modulation of the bias voltage [52], minimization of the junction capacitance [42], and smart [5] and fast [54] quenching circuits. Circuit architectures may require spatial and/or temporal co-incidence [22] to reduce energy downstream in the processing chain, but avalanche energy remains. Our work complements existing hardware approaches by preventing avalanches altogether to reduce illumination-dependent energy consumption." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 130, + 570, + 481, + 665 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 570, + 481, + 665 + ], + "spans": [ + { + "bbox": [ + 130, + 570, + 481, + 665 + ], + "type": "text", + "content": "SPAD dead-time and clocked recharge: SPADs require recharge after an avalanche-inducing photon detection during which subsequent photons are not recorded. This dead-time inhibits photons at high exposure [24, 27], yet, power consumption remains excessive when the average inter-photon arrival interval is shorter than the SPAD dead time [38, 42]. Power consumption at high photon flux can be reduced by controlling the global rate of SPAD recharge (\"clocked recharge\") [38, 47]. Within a single recharge period a SPAD detects at most a single photon; subsequent photon arrivals do not induce an avalanche, thus power is" + } + ] + } + ], + "index": 7 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "spans": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "type": "text", + "content": "4" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 166, + 91, + 235, + 100 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 166, + 91, + 235, + 100 + ], + "spans": [ + { + "bbox": [ + 166, + 91, + 235, + 100 + ], + "type": "text", + "content": "L. Koerner et al." + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 3 + }, + { + "para_blocks": [ + { + "bbox": [ + 130, + 116, + 482, + 236 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 116, + 482, + 236 + ], + "spans": [ + { + "bbox": [ + 130, + 116, + 482, + 236 + ], + "type": "text", + "content": "reduced. To further limit avalanches, clocked recharge has been combined with a limit on the number of detections and time to saturation circuitry to measure the intensity of the saturated pixels [41], coarse pixel-wise exposure control [42], and a sequence of recharge periods similar to exposure bracketing [47]. These methods could be considered special cases of inhibition policies that are global and do not adapt to scene content, and therefore, are limited in flexibility to trade power and measurement fidelity. Consequently, in these methods, the power of avalanches remains a considerable fraction of total SPAD sensor power consumption [47]. For example, in this work we show considerable power savings, reaching up to " + }, + { + "bbox": [ + 130, + 116, + 482, + 236 + ], + "type": "inline_equation", + "content": "90\\%" + }, + { + "bbox": [ + 130, + 116, + 482, + 236 + ], + "type": "text", + "content": ", for an object detection task." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 130, + 236, + 482, + 369 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 236, + 482, + 369 + ], + "spans": [ + { + "bbox": [ + 130, + 236, + 482, + 369 + ], + "type": "text", + "content": "Resource-aware imaging: Event-based vision sensors reduce power consumption by only transmitting scene information when an intensity change is detected [17]. This idea has recently been applied to SPAD arrays to reduce power consumption due to data transfer [11, 46]. In contrast, our method reduces power due to detection by selectively disabling photodiodes based on photon history over small spatial and temporal neighborhoods. Miniaturized cameras with constraints on compute energy have transferred processing to the optical domain [30]. While we focus on passive imaging, depth sensing with SPADs and an active pulsed light source has related constraints such as acquisition time and laser power. Adaptive gating reduces acquisition time [43] and optimal allocation of the laser dwell time among pixels improves data quality [36, 48]." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 132, + 389, + 268, + 403 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 389, + 268, + 403 + ], + "spans": [ + { + "bbox": [ + 132, + 389, + 268, + 403 + ], + "type": "text", + "content": "3 Observation Model" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 130, + 418, + 482, + 552 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 418, + 482, + 552 + ], + "spans": [ + { + "bbox": [ + 130, + 418, + 482, + 552 + ], + "type": "text", + "content": "During an exposure time " + }, + { + "bbox": [ + 130, + 418, + 482, + 552 + ], + "type": "inline_equation", + "content": "T" + }, + { + "bbox": [ + 130, + 418, + 482, + 552 + ], + "type": "text", + "content": ", a photon flux of " + }, + { + "bbox": [ + 130, + 418, + 482, + 552 + ], + "type": "inline_equation", + "content": "\\phi" + }, + { + "bbox": [ + 130, + 418, + 482, + 552 + ], + "type": "text", + "content": " results in an average number of photon conversions, or exposure, of " + }, + { + "bbox": [ + 130, + 418, + 482, + 552 + ], + "type": "inline_equation", + "content": "H = \\phi T" + }, + { + "bbox": [ + 130, + 418, + 482, + 552 + ], + "type": "text", + "content": " (we fold the sensor's photon detection probability, or PDP, into the definition of " + }, + { + "bbox": [ + 130, + 418, + 482, + 552 + ], + "type": "inline_equation", + "content": "\\phi" + }, + { + "bbox": [ + 130, + 418, + 482, + 552 + ], + "type": "text", + "content": ", meaning it represents an effective flux). The distribution of photon conversions, " + }, + { + "bbox": [ + 130, + 418, + 482, + 552 + ], + "type": "inline_equation", + "content": "K" + }, + { + "bbox": [ + 130, + 418, + 482, + 552 + ], + "type": "text", + "content": ", follows the Poisson distribution given by " + }, + { + "bbox": [ + 130, + 418, + 482, + 552 + ], + "type": "inline_equation", + "content": "\\mathsf{P}(K = k;H) = \\frac{H^k e^{-H}}{k!}" + }, + { + "bbox": [ + 130, + 418, + 482, + 552 + ], + "type": "text", + "content": ". During each binary exposure period, a SPAD pixel records a '1' if at least one photon was incident during that period, and '0' otherwise. The probability of detecting at least one photon is given by " + }, + { + "bbox": [ + 130, + 418, + 482, + 552 + ], + "type": "inline_equation", + "content": "Y := 1 - \\mathsf{P}(K = 0;H) = 1 - e^{-H}" + }, + { + "bbox": [ + 130, + 418, + 482, + 552 + ], + "type": "text", + "content": ". Hence, the SPAD pixel readout in each binary frame is a Bernoulli random variable " + }, + { + "bbox": [ + 130, + 418, + 482, + 552 + ], + "type": "inline_equation", + "content": "B \\sim \\mathrm{Bernoulli}(Y)" + }, + { + "bbox": [ + 130, + 418, + 482, + 552 + ], + "type": "text", + "content": ". Multiple exposure time windows, or measurements " + }, + { + "bbox": [ + 130, + 418, + 482, + 552 + ], + "type": "inline_equation", + "content": "(W)" + }, + { + "bbox": [ + 130, + 418, + 482, + 552 + ], + "type": "text", + "content": ", are recorded to reduce noise with the total number of detections" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 219, + 559, + 481, + 594 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 219, + 559, + 481, + 594 + ], + "spans": [ + { + "bbox": [ + 219, + 559, + 481, + 594 + ], + "type": "interline_equation", + "content": "D := \\left[ \\sum_ {n = 1} ^ {W} B _ {n} \\right] \\sim \\operatorname {B i n o m i a l} (W, 1 - e ^ {- H}). \\tag {1}", + "image_path": "9f2e3bdc394b545661d25e9635b6cab90a3a684e783076c23ab5b963465e24d6.jpg" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 132, + 601, + 462, + 613 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 601, + 462, + 613 + ], + "spans": [ + { + "bbox": [ + 132, + 601, + 462, + 613 + ], + "type": "text", + "content": "We estimate the probability of detection and flux from a measurement " + }, + { + "bbox": [ + 132, + 601, + 462, + 613 + ], + "type": "inline_equation", + "content": "D" + }, + { + "bbox": [ + 132, + 601, + 462, + 613 + ], + "type": "text", + "content": " as" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 269, + 624, + 481, + 646 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 269, + 624, + 481, + 646 + ], + "spans": [ + { + "bbox": [ + 269, + 624, + 481, + 646 + ], + "type": "interline_equation", + "content": "\\widehat {Y} = \\frac {D}{W} \\tag {2}", + "image_path": "e55693c27a7a02c0937f2039aad616821539e5b1c6a4e021a8e83fac5e9d8d2b.jpg" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 269, + 647, + 481, + 663 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 269, + 647, + 481, + 663 + ], + "spans": [ + { + "bbox": [ + 269, + 647, + 481, + 663 + ], + "type": "interline_equation", + "content": "\\widehat {H} = - \\ln (1 - \\widehat {Y}). \\qquad \\qquad (3)", + "image_path": "d09ed886ff3cb77a3198e00c4730d1bbaf28e6bf577ec904a839818a032b4d83.jpg" + } + ] + } + ], + "index": 9 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 196, + 91, + 448, + 103 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 196, + 91, + 448, + 103 + ], + "spans": [ + { + "bbox": [ + 196, + 91, + 448, + 103 + ], + "type": "text", + "content": "Photon Inhibition for Energy-Efficient Single-Photon Imaging" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 474, + 92, + 480, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 474, + 92, + 480, + 100 + ], + "spans": [ + { + "bbox": [ + 474, + 92, + 480, + 100 + ], + "type": "text", + "content": "5" + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 4 + }, + { + "para_blocks": [ + { + "bbox": [ + 130, + 116, + 482, + 178 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 116, + 482, + 178 + ], + "spans": [ + { + "bbox": [ + 130, + 116, + 482, + 178 + ], + "type": "text", + "content": "Changes under inhibition: Inhibition is represented by a binary state variable " + }, + { + "bbox": [ + 130, + 116, + 482, + 178 + ], + "type": "inline_equation", + "content": "M" + }, + { + "bbox": [ + 130, + 116, + 482, + 178 + ], + "type": "text", + "content": " at every pixel, with " + }, + { + "bbox": [ + 130, + 116, + 482, + 178 + ], + "type": "inline_equation", + "content": "M_{n} = 1" + }, + { + "bbox": [ + 130, + 116, + 482, + 178 + ], + "type": "text", + "content": " denoting enabled for the " + }, + { + "bbox": [ + 130, + 116, + 482, + 178 + ], + "type": "inline_equation", + "content": "n" + }, + { + "bbox": [ + 130, + 116, + 482, + 178 + ], + "type": "text", + "content": "-th measurement period. " + }, + { + "bbox": [ + 130, + 116, + 482, + 178 + ], + "type": "inline_equation", + "content": "M" + }, + { + "bbox": [ + 130, + 116, + 482, + 178 + ], + "type": "text", + "content": " is a random variable when an adaptive or data-dependent inhibition policy is used. The total number of measurements changes to " + }, + { + "bbox": [ + 130, + 116, + 482, + 178 + ], + "type": "inline_equation", + "content": "W_{inh.} \\coloneqq \\sum_{n=1}^{W} M_{n} \\leq W" + }, + { + "bbox": [ + 130, + 116, + 482, + 178 + ], + "type": "text", + "content": ". The number of detections becomes" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 165, + 184, + 481, + 228 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 165, + 184, + 481, + 228 + ], + "spans": [ + { + "bbox": [ + 165, + 184, + 481, + 228 + ], + "type": "interline_equation", + "content": "D _ {i n h.} := \\left[ \\sum_ {n = 1} ^ {W} M _ {n} B _ {n} \\right] = \\left[ \\sum_ {\\substack {n = 1 \\\\ M _ {n} = 1}} ^ {W} B _ {n} \\right] \\sim \\operatorname {B i n o m i a l} \\left(W _ {i n h.}, 1 - e ^ {- H}\\right), \\tag{4}", + "image_path": "b8cfc7aef08ca2a3de3ec692ab1de38cf7896854019b213947c88320d334dc35.jpg" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 130, + 236, + 482, + 345 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 236, + 482, + 345 + ], + "spans": [ + { + "bbox": [ + 130, + 236, + 482, + 345 + ], + "type": "text", + "content": "and the flux is estimated similar to Eqs. (2) and (3). The second summation conveys that, with inhibition, the measurements when the pixel is enabled match the original model in Eq. (1) - a result of the memoryless property of the Poisson arrival process. The model above requires that transitions in " + }, + { + "bbox": [ + 130, + 236, + 482, + 345 + ], + "type": "inline_equation", + "content": "M" + }, + { + "bbox": [ + 130, + 236, + 482, + 345 + ], + "type": "text", + "content": " are synchronized with the clock signal used to gate the exposure, so that the PDP is not changed by inhibition. A second assumption is that the time for SPAD recharge is small relative to the clock period. This is a desired property for passive SPAD-based imaging, and holds for many state-of-the-art SPADs [42, 44, 47]. It implies that PDP is approximately constant in time and does not depend on prior pixel state." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 132, + 361, + 367, + 376 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 361, + 367, + 376 + ], + "spans": [ + { + "bbox": [ + 132, + 361, + 367, + 376 + ], + "type": "text", + "content": "4 Energy-Aware Performance Metrics" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 130, + 386, + 482, + 441 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 386, + 482, + 441 + ], + "spans": [ + { + "bbox": [ + 130, + 386, + 482, + 441 + ], + "type": "text", + "content": "The exposure-referred signal-to-noise ratio " + }, + { + "bbox": [ + 130, + 386, + 482, + 441 + ], + "type": "inline_equation", + "content": "(\\mathsf{SNR}_H)" + }, + { + "bbox": [ + 130, + 386, + 482, + 441 + ], + "type": "text", + "content": " is commonly used to evaluate single-photon sensor performance, and can be computed as the ratio of the true exposure " + }, + { + "bbox": [ + 130, + 386, + 482, + 441 + ], + "type": "inline_equation", + "content": "H" + }, + { + "bbox": [ + 130, + 386, + 482, + 441 + ], + "type": "text", + "content": " and the root-mean-squared error in the estimated exposure " + }, + { + "bbox": [ + 130, + 386, + 482, + 441 + ], + "type": "inline_equation", + "content": "\\sqrt{\\mathbb{E}[(\\widehat{H} - H)^2]}" + }, + { + "bbox": [ + 130, + 386, + 482, + 441 + ], + "type": "text", + "content": " [15, 55]:" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 219, + 449, + 481, + 484 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 219, + 449, + 481, + 484 + ], + "spans": [ + { + "bbox": [ + 219, + 449, + 481, + 484 + ], + "type": "interline_equation", + "content": "\\mathsf {S N R} _ {H} = \\frac {H}{\\sqrt {\\mathbb {E} [ (\\widehat {H} - H) ^ {2} ]}} = H \\sqrt {\\frac {W}{e ^ {H} - 1}}. \\tag {5}", + "image_path": "291a8cee486e9e6fab0c7652f7f89c1c84b7998faf6de74a035d8f228253d42e.jpg" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 130, + 491, + 482, + 528 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 491, + 482, + 528 + ], + "spans": [ + { + "bbox": [ + 130, + 491, + 482, + 528 + ], + "type": "text", + "content": "At low incident flux, " + }, + { + "bbox": [ + 130, + 491, + 482, + 528 + ], + "type": "inline_equation", + "content": "\\mathsf{SNR}_H" + }, + { + "bbox": [ + 130, + 491, + 482, + 528 + ], + "type": "text", + "content": " is low due to shot noise. " + }, + { + "bbox": [ + 130, + 491, + 482, + 528 + ], + "type": "inline_equation", + "content": "\\mathsf{SNR}_H" + }, + { + "bbox": [ + 130, + 491, + 482, + 528 + ], + "type": "text", + "content": " improves as the likelihood of a photon detection increases until, in bright light with " + }, + { + "bbox": [ + 130, + 491, + 482, + 528 + ], + "type": "inline_equation", + "content": "H > 1.6" + }, + { + "bbox": [ + 130, + 491, + 482, + 528 + ], + "type": "text", + "content": ", the " + }, + { + "bbox": [ + 130, + 491, + 482, + 528 + ], + "type": "inline_equation", + "content": "\\mathsf{SNR}_H" + }, + { + "bbox": [ + 130, + 491, + 482, + 528 + ], + "type": "text", + "content": " degrades due to \"soft\" saturation of the response [27,34]." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 130, + 528, + 482, + 576 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 528, + 482, + 576 + ], + "spans": [ + { + "bbox": [ + 130, + 528, + 482, + 576 + ], + "type": "text", + "content": "We propose two new energy-aware modifications to " + }, + { + "bbox": [ + 130, + 528, + 482, + 576 + ], + "type": "inline_equation", + "content": "\\mathsf{SNR}_H" + }, + { + "bbox": [ + 130, + 528, + 482, + 576 + ], + "type": "text", + "content": " to incorporate SPAD energy costs. First, we propose a detection efficiency metric " + }, + { + "bbox": [ + 130, + 528, + 482, + 576 + ], + "type": "inline_equation", + "content": "\\mathsf{SNR}_{H / D}^2" + }, + { + "bbox": [ + 130, + 528, + 482, + 576 + ], + "type": "text", + "content": " defined as the square of SNR normalized by the expected number of detections " + }, + { + "bbox": [ + 130, + 528, + 482, + 576 + ], + "type": "inline_equation", + "content": "\\mathbb{E}[D] = W(1 - e^{-H})" + }, + { + "bbox": [ + 130, + 528, + 482, + 576 + ], + "type": "text", + "content": ":" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 233, + 583, + 481, + 611 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 233, + 583, + 481, + 611 + ], + "spans": [ + { + "bbox": [ + 233, + 583, + 481, + 611 + ], + "type": "interline_equation", + "content": "\\mathsf {S N R} _ {H / D} ^ {2} := \\frac {\\mathsf {S N R} _ {H} ^ {2}}{\\mathbb {E} [ D ]} = \\frac {H ^ {2} e ^ {- H}}{(1 - e ^ {- H}) ^ {2}}. \\tag {6}", + "image_path": "069909b65991730013ed6f71f47d306b4f86fa6612fdf9faddad3f2f8546f487.jpg" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 130, + 617, + 482, + 667 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 617, + 482, + 667 + ], + "spans": [ + { + "bbox": [ + 130, + 617, + 482, + 667 + ], + "type": "text", + "content": "Fig. 3(a) shows the " + }, + { + "bbox": [ + 130, + 617, + 482, + 667 + ], + "type": "inline_equation", + "content": "\\mathsf{SNR}_H" + }, + { + "bbox": [ + 130, + 617, + 482, + 667 + ], + "type": "text", + "content": " (black) and the detection efficiency (red) versus the average photon arrivals per period " + }, + { + "bbox": [ + 130, + 617, + 482, + 667 + ], + "type": "inline_equation", + "content": "(H)" + }, + { + "bbox": [ + 130, + 617, + 482, + 667 + ], + "type": "text", + "content": ". When " + }, + { + "bbox": [ + 130, + 617, + 482, + 667 + ], + "type": "inline_equation", + "content": "H \\ll 1" + }, + { + "bbox": [ + 130, + 617, + 482, + 667 + ], + "type": "text", + "content": " all detections contribute significant information and " + }, + { + "bbox": [ + 130, + 617, + 482, + 667 + ], + "type": "inline_equation", + "content": "\\mathsf{SNR}_H^2" + }, + { + "bbox": [ + 130, + 617, + 482, + 667 + ], + "type": "text", + "content": " increases linearly, similar to an ideal non-saturating sensor only limited by Poisson noise. Accordingly, " + }, + { + "bbox": [ + 130, + 617, + 482, + 667 + ], + "type": "inline_equation", + "content": "\\mathsf{SNR}_{H / D}^2 \\cong 1" + }, + { + "bbox": [ + 130, + 617, + 482, + 667 + ], + "type": "text", + "content": ", the" + } + ] + } + ], + "index": 11 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "spans": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "type": "text", + "content": "6" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 166, + 91, + 235, + 101 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 166, + 91, + 235, + 101 + ], + "spans": [ + { + "bbox": [ + 166, + 91, + 235, + 101 + ], + "type": "text", + "content": "L. Koerner et al." + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 5 + }, + { + "para_blocks": [ + { + "bbox": [ + 130, + 116, + 479, + 152 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 116, + 479, + 152 + ], + "spans": [ + { + "bbox": [ + 130, + 116, + 479, + 152 + ], + "type": "text", + "content": "upper bound of this metric. At larger exposure values, beginning around " + }, + { + "bbox": [ + 130, + 116, + 479, + 152 + ], + "type": "inline_equation", + "content": "H \\approx 0.5" + }, + { + "bbox": [ + 130, + 116, + 479, + 152 + ], + "type": "text", + "content": ", the sensor begins to saturate which slows the growth of " + }, + { + "bbox": [ + 130, + 116, + 479, + 152 + ], + "type": "inline_equation", + "content": "\\mathrm{SNR}_H" + }, + { + "bbox": [ + 130, + 116, + 479, + 152 + ], + "type": "text", + "content": " and thus degrades the detection efficiency." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 130, + 152, + 480, + 224 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 152, + 480, + 224 + ], + "spans": [ + { + "bbox": [ + 130, + 152, + 480, + 224 + ], + "type": "text", + "content": "A separate constraint is the total number of recharge periods during which a pixel is enabled and can measure either '0' or '1'. This number of measurements " + }, + { + "bbox": [ + 130, + 152, + 480, + 224 + ], + "type": "inline_equation", + "content": "(W)" + }, + { + "bbox": [ + 130, + 152, + 480, + 224 + ], + "type": "text", + "content": " may be limited due to the energy to read out a frame, the depth of an in-pixel counter, and/or the maximum allowable sensing latency due to motion blur. We establish a second metric, measurement efficiency defined as the square of SNR normalized by the number of measurement windows:" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 235, + 232, + 481, + 259 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 235, + 232, + 481, + 259 + ], + "spans": [ + { + "bbox": [ + 235, + 232, + 481, + 259 + ], + "type": "interline_equation", + "content": "\\mathrm {S N R} _ {H / W} ^ {2} := \\frac {\\mathrm {S N R} _ {H} ^ {2}}{W} = \\frac {H ^ {2} e ^ {- H}}{1 - e ^ {- H}}. \\tag {7}", + "image_path": "ac0e5bb3324ddbfbfe3b1e2b4c53f5aedbc25f5990ed6e1408dd607e2a3e2cd8.jpg" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 130, + 266, + 479, + 290 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 266, + 479, + 290 + ], + "spans": [ + { + "bbox": [ + 130, + 266, + 479, + 290 + ], + "type": "text", + "content": "Fig. 3(a) (blue) shows sub-optimal measurement efficiency at both low and high exposures with the best efficiency at " + }, + { + "bbox": [ + 130, + 266, + 479, + 290 + ], + "type": "inline_equation", + "content": "H = 1.59" + }, + { + "bbox": [ + 130, + 266, + 479, + 290 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 130, + 266, + 479, + 290 + ], + "type": "inline_equation", + "content": "Y = 0.80" + }, + { + "bbox": [ + 130, + 266, + 479, + 290 + ], + "type": "text", + "content": ", as demonstrated in [8]." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 130, + 291, + 480, + 362 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 291, + 480, + 362 + ], + "spans": [ + { + "bbox": [ + 130, + 291, + 480, + 362 + ], + "type": "text", + "content": "In an oracle setting with a known image, one can analytically derive nonuniform allocations of measurements " + }, + { + "bbox": [ + 130, + 291, + 480, + 362 + ], + "type": "inline_equation", + "content": "W_{i}" + }, + { + "bbox": [ + 130, + 291, + 480, + 362 + ], + "type": "text", + "content": " across the pixels " + }, + { + "bbox": [ + 130, + 291, + 480, + 362 + ], + "type": "inline_equation", + "content": "i" + }, + { + "bbox": [ + 130, + 291, + 480, + 362 + ], + "type": "text", + "content": ", constraining the total expected detections to a fixed value and optimizing " + }, + { + "bbox": [ + 130, + 291, + 480, + 362 + ], + "type": "inline_equation", + "content": "\\mathsf{SNR}_H" + }, + { + "bbox": [ + 130, + 291, + 480, + 362 + ], + "type": "text", + "content": " or mean-squared-error (see the supplement). The loss metric may also be defined relative to the binary rate " + }, + { + "bbox": [ + 130, + 291, + 480, + 362 + ], + "type": "inline_equation", + "content": "Y" + }, + { + "bbox": [ + 130, + 291, + 480, + 362 + ], + "type": "text", + "content": " instead of " + }, + { + "bbox": [ + 130, + 291, + 480, + 362 + ], + "type": "inline_equation", + "content": "H" + }, + { + "bbox": [ + 130, + 291, + 480, + 362 + ], + "type": "text", + "content": " as above. In that case a useful base metric could be entropy instead of " + }, + { + "bbox": [ + 130, + 291, + 480, + 362 + ], + "type": "inline_equation", + "content": "\\mathsf{SNR}_H" + }, + { + "bbox": [ + 130, + 291, + 480, + 362 + ], + "type": "text", + "content": " [19]." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 131, + 380, + 361, + 395 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 131, + 380, + 361, + 395 + ], + "spans": [ + { + "bbox": [ + 131, + 380, + 361, + 395 + ], + "type": "text", + "content": "5 Spatio-temporal Inhibition Policies" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 130, + 405, + 482, + 501 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 405, + 482, + 501 + ], + "spans": [ + { + "bbox": [ + 130, + 405, + 482, + 501 + ], + "type": "text", + "content": "We now propose policies that calculate a spatio-temporal inhibition pattern for each pixel and each frame based on the history of photon frames and patterns. Following Sec. 3, we define an inhibition pattern using a binary-valued tensor " + }, + { + "bbox": [ + 130, + 405, + 482, + 501 + ], + "type": "inline_equation", + "content": "M" + }, + { + "bbox": [ + 130, + 405, + 482, + 501 + ], + "type": "text", + "content": " where " + }, + { + "bbox": [ + 130, + 405, + 482, + 501 + ], + "type": "inline_equation", + "content": "M(i,j,t) = 0" + }, + { + "bbox": [ + 130, + 405, + 482, + 501 + ], + "type": "text", + "content": " if pixel " + }, + { + "bbox": [ + 130, + 405, + 482, + 501 + ], + "type": "inline_equation", + "content": "(i,j)" + }, + { + "bbox": [ + 130, + 405, + 482, + 501 + ], + "type": "text", + "content": " is disabled in the " + }, + { + "bbox": [ + 130, + 405, + 482, + 501 + ], + "type": "inline_equation", + "content": "t^{\\mathrm{th}}" + }, + { + "bbox": [ + 130, + 405, + 482, + 501 + ], + "type": "text", + "content": " frame, and " + }, + { + "bbox": [ + 130, + 405, + 482, + 501 + ], + "type": "inline_equation", + "content": "M(i,j,t) = 1" + }, + { + "bbox": [ + 130, + 405, + 482, + 501 + ], + "type": "text", + "content": " otherwise. All pixels are initially enabled, and on-sensor calculations modify " + }, + { + "bbox": [ + 130, + 405, + 482, + 501 + ], + "type": "inline_equation", + "content": "M" + }, + { + "bbox": [ + 130, + 405, + 482, + 501 + ], + "type": "text", + "content": " over time. The binary photon cube is defined as " + }, + { + "bbox": [ + 130, + 405, + 482, + 501 + ], + "type": "inline_equation", + "content": "F(i,j,t) = 1" + }, + { + "bbox": [ + 130, + 405, + 482, + 501 + ], + "type": "text", + "content": " if the pixel is enabled (i.e., " + }, + { + "bbox": [ + 130, + 405, + 482, + 501 + ], + "type": "inline_equation", + "content": "M(i,j,t) = 1" + }, + { + "bbox": [ + 130, + 405, + 482, + 501 + ], + "type": "text", + "content": ") and a photon is detected at pixel location " + }, + { + "bbox": [ + 130, + 405, + 482, + 501 + ], + "type": "inline_equation", + "content": "(i,j)" + }, + { + "bbox": [ + 130, + 405, + 482, + 501 + ], + "type": "text", + "content": " in the " + }, + { + "bbox": [ + 130, + 405, + 482, + 501 + ], + "type": "inline_equation", + "content": "t^{\\mathrm{th}}" + }, + { + "bbox": [ + 130, + 405, + 482, + 501 + ], + "type": "text", + "content": " frame, otherwise it is zero." + } + ] + } + ], + "index": 8 + }, + { + "type": "image", + "bbox": [ + 178, + 525, + 433, + 599 + ], + "blocks": [ + { + "bbox": [ + 178, + 525, + 433, + 599 + ], + "lines": [ + { + "bbox": [ + 178, + 525, + 433, + 599 + ], + "spans": [ + { + "bbox": [ + 178, + 525, + 433, + 599 + ], + "type": "image", + "image_path": "94a9abed29f4a57bd67fe0da72b93450eef4df52c2f03cf44391ec6e9174d8db.jpg" + } + ] + } + ], + "index": 9, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 130, + 606, + 482, + 640 + ], + "lines": [ + { + "bbox": [ + 130, + 606, + 482, + 640 + ], + "spans": [ + { + "bbox": [ + 130, + 606, + 482, + 640 + ], + "type": "text", + "content": "Fig. 2: Calculation-based inhibition with dead time. Arrows represent photons with an 'X' for inhibition. " + }, + { + "bbox": [ + 130, + 606, + 482, + 640 + ], + "type": "inline_equation", + "content": "T" + }, + { + "bbox": [ + 130, + 606, + 482, + 640 + ], + "type": "text", + "content": " indicates the clocked recharge period. A score, " + }, + { + "bbox": [ + 130, + 606, + 482, + 640 + ], + "type": "inline_equation", + "content": "S" + }, + { + "bbox": [ + 130, + 606, + 482, + 640 + ], + "type": "text", + "content": ", calculated from past frames determines if future measurements are enabled or disabled." + } + ] + } + ], + "index": 10, + "angle": 0, + "type": "image_caption" + } + ], + "index": 9 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 196, + 91, + 448, + 103 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 196, + 91, + 448, + 103 + ], + "spans": [ + { + "bbox": [ + 196, + 91, + 448, + 103 + ], + "type": "text", + "content": "Photon Inhibition for Energy-Efficient Single-Photon Imaging" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 474, + 91, + 480, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 474, + 91, + 480, + 100 + ], + "spans": [ + { + "bbox": [ + 474, + 91, + 480, + 100 + ], + "type": "text", + "content": "7" + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 6 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 132, + 114, + 258, + 202 + ], + "blocks": [ + { + "bbox": [ + 132, + 114, + 258, + 202 + ], + "lines": [ + { + "bbox": [ + 132, + 114, + 258, + 202 + ], + "spans": [ + { + "bbox": [ + 132, + 114, + 258, + 202 + ], + "type": "image", + "image_path": "2cae52915ddae7613b5cc833b5b17fb931d9e2f7e5efbb7027946db00bb5e776.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 130, + 211, + 482, + 365 + ], + "lines": [ + { + "bbox": [ + 130, + 211, + 482, + 365 + ], + "spans": [ + { + "bbox": [ + 130, + 211, + 482, + 365 + ], + "type": "text", + "content": "Fig. 3: Efficiency metrics and inhibition policies that track the metrics: (a) The " + }, + { + "bbox": [ + 130, + 211, + 482, + 365 + ], + "type": "inline_equation", + "content": "\\mathsf{SNR}_H" + }, + { + "bbox": [ + 130, + 211, + 482, + 365 + ], + "type": "text", + "content": " in dB (black), the detection efficiency (red, --), and the measurement efficiency (blue, --) versus the exposure with " + }, + { + "bbox": [ + 130, + 211, + 482, + 365 + ], + "type": "inline_equation", + "content": "W = 100" + }, + { + "bbox": [ + 130, + 211, + 482, + 365 + ], + "type": "text", + "content": " measurements. The binary rate " + }, + { + "bbox": [ + 130, + 211, + 482, + 365 + ], + "type": "inline_equation", + "content": "Y = 1 - e^{-H}" + }, + { + "bbox": [ + 130, + 211, + 482, + 365 + ], + "type": "text", + "content": " is indicated on the top axis. The vertical dotted line indicates the exposure and the binary rate at which the " + }, + { + "bbox": [ + 130, + 211, + 482, + 365 + ], + "type": "inline_equation", + "content": "\\mathsf{SNR}_H" + }, + { + "bbox": [ + 130, + 211, + 482, + 365 + ], + "type": "text", + "content": " degrades by " + }, + { + "bbox": [ + 130, + 211, + 482, + 365 + ], + "type": "inline_equation", + "content": "3\\mathrm{dB}" + }, + { + "bbox": [ + 130, + 211, + 482, + 365 + ], + "type": "text", + "content": " from the peak " + }, + { + "bbox": [ + 130, + 211, + 482, + 365 + ], + "type": "inline_equation", + "content": "\\mathsf{SNR}_H" + }, + { + "bbox": [ + 130, + 211, + 482, + 365 + ], + "type": "text", + "content": ". (b,c) Monte Carlo simulations of binary images using calculation-based inhibition policies demonstrate how the allocation of measurements versus the pixel exposure depends upon tuning parameters. (b) the inhibition threshold " + }, + { + "bbox": [ + 130, + 211, + 482, + 365 + ], + "type": "inline_equation", + "content": "\\eta" + }, + { + "bbox": [ + 130, + 211, + 482, + 365 + ], + "type": "text", + "content": " adjusts the exposure level at which pixels are inhibited to allow the measurement fraction (the ratio of active measurements to total number of frames) to follow the " + }, + { + "bbox": [ + 130, + 211, + 482, + 365 + ], + "type": "inline_equation", + "content": "\\mathsf{SNR}_{H / D}^2" + }, + { + "bbox": [ + 130, + 211, + 482, + 365 + ], + "type": "text", + "content": " curve in (a). A smaller threshold more aggressively inhibits photons. (c) demonstrates the impact of the hold-off time, " + }, + { + "bbox": [ + 130, + 211, + 482, + 365 + ], + "type": "inline_equation", + "content": "\\tau_{H}" + }, + { + "bbox": [ + 130, + 211, + 482, + 365 + ], + "type": "text", + "content": ", on the number of measurements allocated to the brightest pixels. The legend indicates the total fraction of photons inhibited as " + }, + { + "bbox": [ + 130, + 211, + 482, + 365 + ], + "type": "inline_equation", + "content": "I_F" + }, + { + "bbox": [ + 130, + 211, + 482, + 365 + ], + "type": "text", + "content": ". (b,c) show smoothed curves (Lowess filter, fraction of " + }, + { + "bbox": [ + 130, + 211, + 482, + 365 + ], + "type": "inline_equation", + "content": "1 / 5" + }, + { + "bbox": [ + 130, + 211, + 482, + 365 + ], + "type": "text", + "content": ") of the measurement fraction vs. H." + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "image_caption" + } + ], + "index": 2 + }, + { + "type": "image", + "bbox": [ + 259, + 121, + 482, + 204 + ], + "blocks": [ + { + "bbox": [ + 259, + 121, + 482, + 204 + ], + "lines": [ + { + "bbox": [ + 259, + 121, + 482, + 204 + ], + "spans": [ + { + "bbox": [ + 259, + 121, + 482, + 204 + ], + "type": "image", + "image_path": "3561ef36ff700472374459675ae16d4cbbcce88f9ca78fbc0c2dce44c3b8c89c.jpg" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_body" + } + ], + "index": 3 + }, + { + "bbox": [ + 130, + 388, + 482, + 496 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 388, + 482, + 496 + ], + "spans": [ + { + "bbox": [ + 130, + 388, + 482, + 496 + ], + "type": "text", + "content": "Fig. 1(b) shows the components of a photon inhibition processing layer, including the binary frames, " + }, + { + "bbox": [ + 130, + 388, + 482, + 496 + ], + "type": "inline_equation", + "content": "F" + }, + { + "bbox": [ + 130, + 388, + 482, + 496 + ], + "type": "text", + "content": ", and inhibition pattern " + }, + { + "bbox": [ + 130, + 388, + 482, + 496 + ], + "type": "inline_equation", + "content": "M" + }, + { + "bbox": [ + 130, + 388, + 482, + 496 + ], + "type": "text", + "content": ". For ease of on-sensor implementation, we focus on policies that operate on small and local spatiotemporal neighborhoods of fixed sizes. We rely on local arithmetic and Boolean computations and comparison operations, consistent with current in-pixel computational capabilities [2,9,24]. Fig. 2 shows a proposed on-sensor calculation approach that operates in a streaming fashion as frames accumulate to calculate an inhibition score, " + }, + { + "bbox": [ + 130, + 388, + 482, + 496 + ], + "type": "inline_equation", + "content": "S" + }, + { + "bbox": [ + 130, + 388, + 482, + 496 + ], + "type": "text", + "content": ", as the result of a spatio-temporal filter of the binary frames and inhibition pattern. The score at each pixel is calculated as" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 212, + 505, + 481, + 519 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 212, + 505, + 481, + 519 + ], + "spans": [ + { + "bbox": [ + 212, + 505, + 481, + 519 + ], + "type": "interline_equation", + "content": "S (i, j, t) = K * \\left[ \\left(2 F (i, j, t) - \\mathbf {1}\\right) \\cdot M (i, j, t) \\right] \\tag {8}", + "image_path": "cafb8057a2a739da9df9a61e0de04b79c0212250ef4fd3e13b770fc72dab5bbf.jpg" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 130, + 527, + 482, + 668 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 527, + 482, + 668 + ], + "spans": [ + { + "bbox": [ + 130, + 527, + 482, + 668 + ], + "type": "text", + "content": "which applies a spatio-temporal filtering kernel, " + }, + { + "bbox": [ + 130, + 527, + 482, + 668 + ], + "type": "inline_equation", + "content": "K" + }, + { + "bbox": [ + 130, + 527, + 482, + 668 + ], + "type": "text", + "content": ", of dimensions " + }, + { + "bbox": [ + 130, + 527, + 482, + 668 + ], + "type": "inline_equation", + "content": "L, H, T" + }, + { + "bbox": [ + 130, + 527, + 482, + 668 + ], + "type": "text", + "content": " to a ternary representation of the pixel result (1, 0, or -1 for a detection, a disabled pixel, or a measurement that does not detect a photon, respectively). The kernel " + }, + { + "bbox": [ + 130, + 527, + 482, + 668 + ], + "type": "inline_equation", + "content": "K" + }, + { + "bbox": [ + 130, + 527, + 482, + 668 + ], + "type": "text", + "content": " can typically be separated into spatial and temporal components as " + }, + { + "bbox": [ + 130, + 527, + 482, + 668 + ], + "type": "inline_equation", + "content": "K = K_{s} \\otimes K_{t}" + }, + { + "bbox": [ + 130, + 527, + 482, + 668 + ], + "type": "text", + "content": " with dimensions " + }, + { + "bbox": [ + 130, + 527, + 482, + 668 + ], + "type": "inline_equation", + "content": "L \\times H \\times 1" + }, + { + "bbox": [ + 130, + 527, + 482, + 668 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 130, + 527, + 482, + 668 + ], + "type": "inline_equation", + "content": "1 \\times 1 \\times T" + }, + { + "bbox": [ + 130, + 527, + 482, + 668 + ], + "type": "text", + "content": ", respectively. After each binary frame, the score is compared to a threshold " + }, + { + "bbox": [ + 130, + 527, + 482, + 668 + ], + "type": "inline_equation", + "content": "\\eta" + }, + { + "bbox": [ + 130, + 527, + 482, + 668 + ], + "type": "text", + "content": " and the pixel is disabled for the subsequent " + }, + { + "bbox": [ + 130, + 527, + 482, + 668 + ], + "type": "inline_equation", + "content": "\\tau_{H}" + }, + { + "bbox": [ + 130, + 527, + 482, + 668 + ], + "type": "text", + "content": " frames: " + }, + { + "bbox": [ + 130, + 527, + 482, + 668 + ], + "type": "inline_equation", + "content": "M(i,j,t') = 0" + }, + { + "bbox": [ + 130, + 527, + 482, + 668 + ], + "type": "text", + "content": " for " + }, + { + "bbox": [ + 130, + 527, + 482, + 668 + ], + "type": "inline_equation", + "content": "\\{t'|t + 1 \\leq t' \\leq t + 1 + \\tau_{H}\\}" + }, + { + "bbox": [ + 130, + 527, + 482, + 668 + ], + "type": "text", + "content": " if " + }, + { + "bbox": [ + 130, + 527, + 482, + 668 + ], + "type": "inline_equation", + "content": "S(i,j,t) > \\eta" + }, + { + "bbox": [ + 130, + 527, + 482, + 668 + ], + "type": "text", + "content": ". Observe in Fig. 3(b,c) that decreasing " + }, + { + "bbox": [ + 130, + 527, + 482, + 668 + ], + "type": "inline_equation", + "content": "\\eta" + }, + { + "bbox": [ + 130, + 527, + 482, + 668 + ], + "type": "text", + "content": " and increasing " + }, + { + "bbox": [ + 130, + 527, + 482, + 668 + ], + "type": "inline_equation", + "content": "\\tau_{H}" + }, + { + "bbox": [ + 130, + 527, + 482, + 668 + ], + "type": "text", + "content": " can be used to attain more aggressive inhibition with a larger fraction of photons being inhibited. The binary rate of each pixel is estimated (using Eq. (4)) as the ratio of detections to (active) measurements: " + }, + { + "bbox": [ + 130, + 527, + 482, + 668 + ], + "type": "inline_equation", + "content": "\\widehat{Y}(i,j,t) = \\sum_{t'} F(i,j,t') / \\sum_{t'} M(i,j,t')" + }, + { + "bbox": [ + 130, + 527, + 482, + 668 + ], + "type": "text", + "content": ". This" + } + ] + } + ], + "index": 7 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "spans": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "type": "text", + "content": "8" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 166, + 91, + 235, + 100 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 166, + 91, + 235, + 100 + ], + "spans": [ + { + "bbox": [ + 166, + 91, + 235, + 100 + ], + "type": "text", + "content": "L. Koerner et al." + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 7 + }, + { + "para_blocks": [ + { + "bbox": [ + 130, + 116, + 480, + 152 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 116, + 480, + 152 + ], + "spans": [ + { + "bbox": [ + 130, + 116, + 480, + 152 + ], + "type": "text", + "content": "calculation requires a record of the inhibition history which could be accumulated by an in-pixel counter or recreated in a downstream processor if all binary frames are read out." + } + ] + } + ], + "index": 2 + }, + { + "type": "image", + "bbox": [ + 181, + 175, + 431, + 245 + ], + "blocks": [ + { + "bbox": [ + 181, + 175, + 431, + 245 + ], + "lines": [ + { + "bbox": [ + 181, + 175, + 431, + 245 + ], + "spans": [ + { + "bbox": [ + 181, + 175, + 431, + 245 + ], + "type": "image", + "image_path": "711ddf144f258a16fb41255be347bd83cd923e9a2ad2c7aafa1b589fde3d9de6.jpg" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_body" + } + ], + "index": 3 + }, + { + "type": "image", + "bbox": [ + 130, + 251, + 249, + 326 + ], + "blocks": [ + { + "bbox": [ + 130, + 251, + 249, + 326 + ], + "lines": [ + { + "bbox": [ + 130, + 251, + 249, + 326 + ], + "spans": [ + { + "bbox": [ + 130, + 251, + 249, + 326 + ], + "type": "image", + "image_path": "b3071da236b4c798e7d1b4eec7e65417dd36ef56e19ebdabd307cf09e731ef3a.jpg" + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "image_body" + } + ], + "index": 4 + }, + { + "type": "image", + "bbox": [ + 256, + 260, + 361, + 326 + ], + "blocks": [ + { + "bbox": [ + 256, + 260, + 361, + 326 + ], + "lines": [ + { + "bbox": [ + 256, + 260, + 361, + 326 + ], + "spans": [ + { + "bbox": [ + 256, + 260, + 361, + 326 + ], + "type": "image", + "image_path": "712b81ee56401f0e953743fd72dc033f08581a329c1b77aa913b339ba81a7efd.jpg" + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "image_body" + } + ], + "index": 5 + }, + { + "type": "image", + "bbox": [ + 365, + 253, + 482, + 325 + ], + "blocks": [ + { + "bbox": [ + 365, + 253, + 482, + 325 + ], + "lines": [ + { + "bbox": [ + 365, + 253, + 482, + 325 + ], + "spans": [ + { + "bbox": [ + 365, + 253, + 482, + 325 + ], + "type": "image", + "image_path": "42cb2d604a71e767307335682a047ec4b5743d8331153e747755af236427e544.jpg" + } + ] + } + ], + "index": 6, + "angle": 0, + "type": "image_body" + } + ], + "index": 6 + }, + { + "type": "image", + "bbox": [ + 133, + 332, + 482, + 367 + ], + "blocks": [ + { + "bbox": [ + 133, + 332, + 482, + 367 + ], + "lines": [ + { + "bbox": [ + 133, + 332, + 482, + 367 + ], + "spans": [ + { + "bbox": [ + 133, + 332, + 482, + 367 + ], + "type": "image", + "image_path": "ea42ce3feb38c87db4188bdd54e4e392605a2a2ba7f28a00a6e56ce83fc2c4bc.jpg" + } + ] + } + ], + "index": 7, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 130, + 375, + 482, + 485 + ], + "lines": [ + { + "bbox": [ + 130, + 375, + 482, + 485 + ], + "spans": [ + { + "bbox": [ + 130, + 375, + 482, + 485 + ], + "type": "text", + "content": "Fig. 4: Saturation look-ahead inhibition. (a) When detections " + }, + { + "bbox": [ + 130, + 375, + 482, + 485 + ], + "type": "inline_equation", + "content": "D_{i}" + }, + { + "bbox": [ + 130, + 375, + 482, + 485 + ], + "type": "text", + "content": " of cycle " + }, + { + "bbox": [ + 130, + 375, + 482, + 485 + ], + "type": "inline_equation", + "content": "i" + }, + { + "bbox": [ + 130, + 375, + 482, + 485 + ], + "type": "text", + "content": " exceed a threshold, " + }, + { + "bbox": [ + 130, + 375, + 482, + 485 + ], + "type": "inline_equation", + "content": "d_{i}" + }, + { + "bbox": [ + 130, + 375, + 482, + 485 + ], + "type": "text", + "content": ", the photons arriving during the next recharge cycle, " + }, + { + "bbox": [ + 130, + 375, + 482, + 485 + ], + "type": "inline_equation", + "content": "i + 1" + }, + { + "bbox": [ + 130, + 375, + 482, + 485 + ], + "type": "text", + "content": ", are inhibited. For this drawing " + }, + { + "bbox": [ + 130, + 375, + 482, + 485 + ], + "type": "inline_equation", + "content": "[d_{1}, d_{2}] = [3, 3]" + }, + { + "bbox": [ + 130, + 375, + 482, + 485 + ], + "type": "text", + "content": ". (b, c, d) Example with three recharge period settings and brackets of " + }, + { + "bbox": [ + 130, + 375, + 482, + 485 + ], + "type": "inline_equation", + "content": "W_{i} = 10" + }, + { + "bbox": [ + 130, + 375, + 482, + 485 + ], + "type": "text", + "content": " measurements each. The saturation thresholds are set as " + }, + { + "bbox": [ + 130, + 375, + 482, + 485 + ], + "type": "inline_equation", + "content": "[d_{1}, d_{2}] = [7, 7]" + }, + { + "bbox": [ + 130, + 375, + 482, + 485 + ], + "type": "text", + "content": ". The brackets are combined linearly, weighted by " + }, + { + "bbox": [ + 130, + 375, + 482, + 485 + ], + "type": "inline_equation", + "content": "\\mathsf{SNR}_H^2" + }, + { + "bbox": [ + 130, + 375, + 482, + 485 + ], + "type": "text", + "content": " [18]. (b, c) With negligible effect on reconstructed SNR relative to conventional clocked-recharge, significantly fewer photon detections occur under the inhibition policy in high-flux settings. (d) The individual brackets are effectively disabled in flux regimes with low detection efficiency, and the detection rate closely tracks the original " + }, + { + "bbox": [ + 130, + 375, + 482, + 485 + ], + "type": "inline_equation", + "content": "\\mathsf{SNR}_H^2" + }, + { + "bbox": [ + 130, + 375, + 482, + 485 + ], + "type": "text", + "content": " curve. Simulation details are in the supplement." + } + ] + } + ], + "index": 8, + "angle": 0, + "type": "image_caption" + } + ], + "index": 7 + }, + { + "bbox": [ + 130, + 510, + 482, + 666 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 510, + 482, + 666 + ], + "spans": [ + { + "bbox": [ + 130, + 510, + 482, + 666 + ], + "type": "text", + "content": "Example policies: 1) Single-pixel dead time: A spatial kernel " + }, + { + "bbox": [ + 130, + 510, + 482, + 666 + ], + "type": "inline_equation", + "content": "K_{s}" + }, + { + "bbox": [ + 130, + 510, + 482, + 666 + ], + "type": "text", + "content": " of dimensions " + }, + { + "bbox": [ + 130, + 510, + 482, + 666 + ], + "type": "inline_equation", + "content": "1 \\times 1" + }, + { + "bbox": [ + 130, + 510, + 482, + 666 + ], + "type": "text", + "content": " may mimic the dead time of a passive SPAD without influence from neighboring pixels. Unlike SPAD recharge generated dead time, the temporal kernel may be extended (e.g., " + }, + { + "bbox": [ + 130, + 510, + 482, + 666 + ], + "type": "inline_equation", + "content": "N" + }, + { + "bbox": [ + 130, + 510, + 482, + 666 + ], + "type": "text", + "content": " frames long) to establish a rate threshold for inhibition with reduced quantization noise. 2) Local spatio-temporal averaging: Single binary frames are inherently noisy. An inhibition policy that calculates spatiotemporal averages to estimate the local photon rate may reduce the impact of noise on the inhibition pattern. Another benefit, as shown in Fig. 3(b,c), is that the stochastic nature of binary frames smoothes the distribution of measurements versus pixel flux. Discontinuities (\"dips\") in SNR versus photon flux are undesirable due to the potential for artifacts. 3) Edge enhancement: Pixels may be inhibited if a local neighborhood has little spatial variation in photon rate. This can be achieved, for instance, through a spatial filter " + }, + { + "bbox": [ + 130, + 510, + 482, + 666 + ], + "type": "inline_equation", + "content": "K_{s}" + }, + { + "bbox": [ + 130, + 510, + 482, + 666 + ], + "type": "text", + "content": " in Eq. (8) which acts" + } + ] + } + ], + "index": 9 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 197, + 91, + 448, + 102 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 197, + 91, + 448, + 102 + ], + "spans": [ + { + "bbox": [ + 197, + 91, + 448, + 102 + ], + "type": "text", + "content": "Photon Inhibition for Energy-Efficient Single-Photon Imaging" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 474, + 92, + 480, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 474, + 92, + 480, + 100 + ], + "spans": [ + { + "bbox": [ + 474, + 92, + 480, + 100 + ], + "type": "text", + "content": "9" + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 8 + }, + { + "para_blocks": [ + { + "bbox": [ + 130, + 116, + 479, + 140 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 116, + 479, + 140 + ], + "spans": [ + { + "bbox": [ + 130, + 116, + 479, + 140 + ], + "type": "text", + "content": "like a Laplacian filter. Such a strategy may enhance the fidelity of edges in the image while focusing fewer resources on regions with constant illumination." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 130, + 140, + 482, + 284 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 140, + 482, + 284 + ], + "spans": [ + { + "bbox": [ + 130, + 140, + 482, + 284 + ], + "type": "text", + "content": "Fig. 4 presents a second proposed inhibition policy called saturation lookahead. This policy combines exposure brackets and calculation-based inhibition for a light-weight single-pixel inhibition policy. This policy proceeds as a sequence of cycles (indexed by " + }, + { + "bbox": [ + 130, + 140, + 482, + 284 + ], + "type": "inline_equation", + "content": "i" + }, + { + "bbox": [ + 130, + 140, + 482, + 284 + ], + "type": "text", + "content": ") of binary frames where each binary frame within each cycle uses the same exposure time. Cycle exposure times " + }, + { + "bbox": [ + 130, + 140, + 482, + 284 + ], + "type": "inline_equation", + "content": "T_{i}" + }, + { + "bbox": [ + 130, + 140, + 482, + 284 + ], + "type": "text", + "content": " progressively increase (" + }, + { + "bbox": [ + 130, + 140, + 482, + 284 + ], + "type": "inline_equation", + "content": "T_{1} < T_{2} < T_{3} \\ldots" + }, + { + "bbox": [ + 130, + 140, + 482, + 284 + ], + "type": "text", + "content": ") so that measurements taken in an earlier cycle may predict low detection efficiency (near saturation) at longer exposure times and disable the pixel in these subsequent cycles. The exposure level thresholds for inhibition would typically be set such that the number of photons detected at a given flux level tracks the " + }, + { + "bbox": [ + 130, + 140, + 482, + 284 + ], + "type": "inline_equation", + "content": "\\mathsf{SNR}_H^2" + }, + { + "bbox": [ + 130, + 140, + 482, + 284 + ], + "type": "text", + "content": " (see Fig. 4(b)), but may be adjusted further based on the relative importance of power consumption, sensing latency, and SNR in an actual application setting." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 132, + 300, + 322, + 314 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 300, + 322, + 314 + ], + "spans": [ + { + "bbox": [ + 132, + 300, + 322, + 314 + ], + "type": "text", + "content": "6 Simulation-based Evaluation" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 130, + 325, + 479, + 409 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 325, + 479, + 409 + ], + "spans": [ + { + "bbox": [ + 130, + 325, + 479, + 409 + ], + "type": "text", + "content": "We use Monte Carlo simulations to generate sequences of grayscale binary frames from a dataset of RGB images [1]. The inhibition policies evaluated extend the baseline inhibition generated by clocked recharge. The inhibition score and patterns for various policies and tuning parameters are calculated from these binary frame sequences. Once inhibition patterns are determined, performance is evaluated by tabulating detections, measurements, and image quality or vision task performance for each step in the sequence (see the supplement for details)." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 130, + 411, + 482, + 567 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 411, + 482, + 567 + ], + "spans": [ + { + "bbox": [ + 130, + 411, + 482, + 567 + ], + "type": "text", + "content": "Spatio-Temporal Policies for Imaging: Handcrafted spatial kernels (3x3) were combined with an averaging temporal kernel of length 4 to form lightweight inhibition policies that allocate pixel measurements as described in earlier sections for improved image reconstruction. Fig. 5 summarizes the simulation results. Fig. 5(d,e) display reduction in photon detections at equal structural similarity index measure (SSIM) [51], enabled by disabling bright pixels, for an exposure bracketing sequence. Intensity estimates from each bracket are combined using " + }, + { + "bbox": [ + 130, + 411, + 482, + 567 + ], + "type": "inline_equation", + "content": "\\mathsf{SNR}^2" + }, + { + "bbox": [ + 130, + 411, + 482, + 567 + ], + "type": "text", + "content": " weighting [18] and then converted to a binary rate estimate at the center exposure level of 1 ppp. The proposed policy demonstrates an average reduction in detections of " + }, + { + "bbox": [ + 130, + 411, + 482, + 567 + ], + "type": "inline_equation", + "content": "42\\%" + }, + { + "bbox": [ + 130, + 411, + 482, + 567 + ], + "type": "text", + "content": " as compared to no inhibition. Fig. 5(f,g) evaluates a single exposure level (1.0 ppp) which is a more challenging scenario, yet the proposed policy still reduces detections by " + }, + { + "bbox": [ + 130, + 411, + 482, + 567 + ], + "type": "inline_equation", + "content": "14\\%" + }, + { + "bbox": [ + 130, + 411, + 482, + 567 + ], + "type": "text", + "content": " at " + }, + { + "bbox": [ + 130, + 411, + 482, + 567 + ], + "type": "inline_equation", + "content": "\\mathrm{SSIM} = 0.7" + }, + { + "bbox": [ + 130, + 411, + 482, + 567 + ], + "type": "text", + "content": ". See supplement S 3.4 for more examples and S 3.5 for simulations of high dynamic range images." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 130, + 569, + 482, + 666 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 569, + 482, + 666 + ], + "spans": [ + { + "bbox": [ + 130, + 569, + 482, + 666 + ], + "type": "text", + "content": "Edge Detection: The BSDS500 dataset with ground truth boundaries [1] was used to study energy-efficient edge detection via photon inhibition. Binary rate images were processed by pre-trained holistically-nested edge detection (HED) [53] with the resulting edge maps compared to ground truth by the structured edge detection toolbox [13]. Fig. 6 shows the optimal image scale (OIS) F-score versus the average detections per pixel. Interpolated curves (not shown) allow for translating along horizontal lines of equal task performance to assess differences in avalanche energy. At low photon counts the proposed edge-enhancing policy" + } + ] + } + ], + "index": 7 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "spans": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "type": "text", + "content": "10" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 166, + 91, + 235, + 100 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 166, + 91, + 235, + 100 + ], + "spans": [ + { + "bbox": [ + 166, + 91, + 235, + 100 + ], + "type": "text", + "content": "L. Koerner et al." + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 9 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 134, + 116, + 252, + 210 + ], + "blocks": [ + { + "bbox": [ + 134, + 116, + 252, + 210 + ], + "lines": [ + { + "bbox": [ + 134, + 116, + 252, + 210 + ], + "spans": [ + { + "bbox": [ + 134, + 116, + 252, + 210 + ], + "type": "image", + "image_path": "e936f7b8c80d7ef1abc12fc2a2cd0ef2f6ec60c78629b4c18cbb905d34f9b351.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_body" + } + ], + "index": 2 + }, + { + "type": "image", + "bbox": [ + 251, + 118, + 369, + 210 + ], + "blocks": [ + { + "bbox": [ + 251, + 118, + 369, + 210 + ], + "lines": [ + { + "bbox": [ + 251, + 118, + 369, + 210 + ], + "spans": [ + { + "bbox": [ + 251, + 118, + 369, + 210 + ], + "type": "image", + "image_path": "2c644c6241f580648fde24f36ba2025a5f18aee034b40212bec9a227fc45f8cd.jpg" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_body" + } + ], + "index": 3 + }, + { + "type": "image", + "bbox": [ + 133, + 212, + 212, + 314 + ], + "blocks": [ + { + "bbox": [ + 133, + 212, + 212, + 314 + ], + "lines": [ + { + "bbox": [ + 133, + 212, + 212, + 314 + ], + "spans": [ + { + "bbox": [ + 133, + 212, + 212, + 314 + ], + "type": "image", + "image_path": "37be8622632bf095fe7076254eed99d5c15995de3301d2702bb838cef731556b.jpg" + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 130, + 330, + 482, + 472 + ], + "lines": [ + { + "bbox": [ + 130, + 330, + 482, + 472 + ], + "spans": [ + { + "bbox": [ + 130, + 330, + 482, + 472 + ], + "type": "text", + "content": "Fig. 5: Power-efficient static imaging via inhibition. (a) Images from an exposure bracketing sequence (average exposure of 0.1, 1.0, 10.0 photons per pixel (ppp)) using clocked recharge without (conventional) and with inhibition. (b) the distribution of measurements based on pixel intensity with " + }, + { + "bbox": [ + 130, + 330, + 482, + 472 + ], + "type": "inline_equation", + "content": "H" + }, + { + "bbox": [ + 130, + 330, + 482, + 472 + ], + "type": "text", + "content": " at 1.0 ppp. (c) image quality (SSIM) at equal detections/pixel. (d-g) Handcrafted policies are tested over 20 images and assessed by reductions in detections (" + }, + { + "bbox": [ + 130, + 330, + 482, + 472 + ], + "type": "inline_equation", + "content": "D" + }, + { + "bbox": [ + 130, + 330, + 482, + 472 + ], + "type": "text", + "content": ") at equal SSIM. The top row (d),(e) use exposure bracketing; (f),(g) use a single exposure of 1.0 ppp. (d),(f) at SSIM=0.7 and (e),(g) at SSIM=0.8. The box shows quartiles with the center line at the median. The proposed policy, " + }, + { + "bbox": [ + 130, + 330, + 482, + 472 + ], + "type": "inline_equation", + "content": "P_{cr}" + }, + { + "bbox": [ + 130, + 330, + 482, + 472 + ], + "type": "text", + "content": ", is a " + }, + { + "bbox": [ + 130, + 330, + 482, + 472 + ], + "type": "inline_equation", + "content": "3 \\times 3" + }, + { + "bbox": [ + 130, + 330, + 482, + 472 + ], + "type": "text", + "content": " spatial kernel that emphasizes the center pixel (" + }, + { + "bbox": [ + 130, + 330, + 482, + 472 + ], + "type": "inline_equation", + "content": "\\times 8" + }, + { + "bbox": [ + 130, + 330, + 482, + 472 + ], + "type": "text", + "content": ") and includes the 8 neighbors (" + }, + { + "bbox": [ + 130, + 330, + 482, + 472 + ], + "type": "inline_equation", + "content": "\\times 1" + }, + { + "bbox": [ + 130, + 330, + 482, + 472 + ], + "type": "text", + "content": ") (see the supplement for policy details). (h) average inhibition patterns for each exposure time. The top most pattern inhibits the brightest pixels only (maximum of " + }, + { + "bbox": [ + 130, + 330, + 482, + 472 + ], + "type": "inline_equation", + "content": "\\sim 60\\%" + }, + { + "bbox": [ + 130, + 330, + 482, + 472 + ], + "type": "text", + "content": " inhibition, primarily in the sky). For the longest exposure time the inhibition pattern allocates measurements primarily to the darkest pixels." + } + ] + } + ], + "index": 9, + "angle": 0, + "type": "image_caption" + } + ], + "index": 4 + }, + { + "type": "image", + "bbox": [ + 219, + 213, + 309, + 315 + ], + "blocks": [ + { + "bbox": [ + 219, + 213, + 309, + 315 + ], + "lines": [ + { + "bbox": [ + 219, + 213, + 309, + 315 + ], + "spans": [ + { + "bbox": [ + 219, + 213, + 309, + 315 + ], + "type": "image", + "image_path": "8ac98e90c44360cebbc62ff83dd1ec83dae1631a2f739a213827d5e564d21f1d.jpg" + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "image_body" + } + ], + "index": 5 + }, + { + "type": "image", + "bbox": [ + 312, + 213, + 377, + 318 + ], + "blocks": [ + { + "bbox": [ + 312, + 213, + 377, + 318 + ], + "lines": [ + { + "bbox": [ + 312, + 213, + 377, + 318 + ], + "spans": [ + { + "bbox": [ + 312, + 213, + 377, + 318 + ], + "type": "image", + "image_path": "c0cae04302e9f6a845ac9aa869153b185b9068b50cdc3dc96430a6941404f2c3.jpg" + } + ] + } + ], + "index": 6, + "angle": 0, + "type": "image_body" + } + ], + "index": 6 + }, + { + "type": "image", + "bbox": [ + 374, + 118, + 478, + 197 + ], + "blocks": [ + { + "bbox": [ + 374, + 118, + 478, + 197 + ], + "lines": [ + { + "bbox": [ + 374, + 118, + 478, + 197 + ], + "spans": [ + { + "bbox": [ + 374, + 118, + 478, + 197 + ], + "type": "image", + "image_path": "f286949a31cd84d68bfccd8430fe1b23e8343da6f3ad42f7877c170ea4b243d7.jpg" + } + ] + } + ], + "index": 7, + "angle": 0, + "type": "image_body" + } + ], + "index": 7 + }, + { + "type": "image", + "bbox": [ + 378, + 197, + 479, + 311 + ], + "blocks": [ + { + "bbox": [ + 378, + 197, + 479, + 311 + ], + "lines": [ + { + "bbox": [ + 378, + 197, + 479, + 311 + ], + "spans": [ + { + "bbox": [ + 378, + 197, + 479, + 311 + ], + "type": "image", + "image_path": "f1ceba2f6ec65d61dbe7e96d9bb0f0e5aa9832265988c45cbce2bddce4f8a6c5.jpg" + } + ] + } + ], + "index": 8, + "angle": 0, + "type": "image_body" + } + ], + "index": 8 + }, + { + "type": "image", + "bbox": [ + 223, + 482, + 392, + 570 + ], + "blocks": [ + { + "bbox": [ + 223, + 482, + 392, + 570 + ], + "lines": [ + { + "bbox": [ + 223, + 482, + 392, + 570 + ], + "spans": [ + { + "bbox": [ + 223, + 482, + 392, + 570 + ], + "type": "image", + "image_path": "2976340e63e3a778154ddb56dc0a9359323862d7c5fdafe89b3767c77ce2c246.jpg" + } + ] + } + ], + "index": 10, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 130, + 579, + 482, + 635 + ], + "lines": [ + { + "bbox": [ + 130, + 579, + 482, + 635 + ], + "spans": [ + { + "bbox": [ + 130, + 579, + 482, + 635 + ], + "type": "text", + "content": "Fig. 6: Energy-efficient edge detection by inhibition: Edge detection F-scores averaged over 19 images versus average detections/pixel. Horizontal arrows show reduction in detections enabled by inhibition at equal task performance. At and beyond " + }, + { + "bbox": [ + 130, + 579, + 482, + 635 + ], + "type": "inline_equation", + "content": "30\\mathrm{D / pix}" + }, + { + "bbox": [ + 130, + 579, + 482, + 635 + ], + "type": "text", + "content": ". no inhibition and the proposed policy are nearly equivalent and plateau at " + }, + { + "bbox": [ + 130, + 579, + 482, + 635 + ], + "type": "inline_equation", + "content": "200\\mathrm{D / pix}" + }, + { + "bbox": [ + 130, + 579, + 482, + 635 + ], + "type": "text", + "content": ". The edge detector returns " + }, + { + "bbox": [ + 130, + 579, + 482, + 635 + ], + "type": "inline_equation", + "content": "F = 0.813" + }, + { + "bbox": [ + 130, + 579, + 482, + 635 + ], + "type": "text", + "content": " on the original images of this set." + } + ] + } + ], + "index": 11, + "angle": 0, + "type": "image_caption" + } + ], + "index": 10 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 197, + 91, + 448, + 102 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 197, + 91, + 448, + 102 + ], + "spans": [ + { + "bbox": [ + 197, + 91, + 448, + 102 + ], + "type": "text", + "content": "Photon Inhibition for Energy-Efficient Single-Photon Imaging" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 470, + 92, + 479, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 470, + 92, + 479, + 100 + ], + "spans": [ + { + "bbox": [ + 470, + 92, + 479, + 100 + ], + "type": "text", + "content": "11" + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 10 + }, + { + "para_blocks": [ + { + "bbox": [ + 130, + 116, + 482, + 200 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 116, + 482, + 200 + ], + "spans": [ + { + "bbox": [ + 130, + 116, + 482, + 200 + ], + "type": "text", + "content": "demonstrates a " + }, + { + "bbox": [ + 130, + 116, + 482, + 200 + ], + "type": "inline_equation", + "content": "30\\%" + }, + { + "bbox": [ + 130, + 116, + 482, + 200 + ], + "type": "text", + "content": " reduction in detections on single exposure time captures. This policy extends the on-sensor calculation approach by calculating two scores, the conventional " + }, + { + "bbox": [ + 130, + 116, + 482, + 200 + ], + "type": "inline_equation", + "content": "3 \\times 3" + }, + { + "bbox": [ + 130, + 116, + 482, + 200 + ], + "type": "text", + "content": " Laplacian [50] as " + }, + { + "bbox": [ + 130, + 116, + 482, + 200 + ], + "type": "inline_equation", + "content": "S_{1}" + }, + { + "bbox": [ + 130, + 116, + 482, + 200 + ], + "type": "text", + "content": " and a " + }, + { + "bbox": [ + 130, + 116, + 482, + 200 + ], + "type": "inline_equation", + "content": "3 \\times 3" + }, + { + "bbox": [ + 130, + 116, + 482, + 200 + ], + "type": "text", + "content": " averaging filter as " + }, + { + "bbox": [ + 130, + 116, + 482, + 200 + ], + "type": "inline_equation", + "content": "S_{2}" + }, + { + "bbox": [ + 130, + 116, + 482, + 200 + ], + "type": "text", + "content": ". The final inhibition decision is the Boolean operation of these scores as " + }, + { + "bbox": [ + 130, + 116, + 482, + 200 + ], + "type": "inline_equation", + "content": "((\\eta_{1} < S_{1} < \\eta_{2}) \\land (S_{2} > \\eta_{3})) \\lor (S_{2} > \\eta_{4})" + }, + { + "bbox": [ + 130, + 116, + 482, + 200 + ], + "type": "text", + "content": ". The Laplacian policy alone performs poorly as photons in dim regions with minimal spatial variation are inhibited - disabling dim pixels is energy inefficient. Complete policy descriptions are in the supplement." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 132, + 220, + 348, + 234 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 220, + 348, + 234 + ], + "spans": [ + { + "bbox": [ + 132, + 220, + 348, + 234 + ], + "type": "text", + "content": "7 Experiments on Dynamic Scenes" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 130, + 247, + 481, + 319 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 247, + 481, + 319 + ], + "spans": [ + { + "bbox": [ + 130, + 247, + 481, + 319 + ], + "type": "text", + "content": "Many real-world scenes contain significant motion even with the high frame rate of a SPAD camera. Burst reconstruction algorithms yield high-quality images from sequences of binary frames [34,35] - we investigate their compatibility with the inhibited photon detection data. In particular, we focus on the saturation look-ahead policy of Fig. 4(a), applied independently at every pixel, and therefore an example of an adaptive single-pixel temporal policy." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 130, + 319, + 481, + 427 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 319, + 481, + 427 + ], + "spans": [ + { + "bbox": [ + 130, + 319, + 481, + 427 + ], + "type": "text", + "content": "To avoid losing salient information under motion, the inhibition policy must limit periods of extended dead time. In the context of the saturation look-ahead policy this limits the total exposure length of the bracketing sequence, since we implicitly assume flux to be constant within each sequence. In our experiments, we use a Fibonacci bracketing [20] sequence " + }, + { + "bbox": [ + 130, + 319, + 481, + 427 + ], + "type": "inline_equation", + "content": "T \\coloneqq \\{1, 1, 2, 3, 5, 8, 13, 21\\}^6" + }, + { + "bbox": [ + 130, + 319, + 481, + 427 + ], + "type": "text", + "content": ", denoted in the units of a single minimal exposure time. Every measured bracket is converted to a maximum-likelihood estimate of flux, which is then supplied to the quanta burst photography algorithm [34]. The flux estimator from bracketed measurements is described in detail in supplement Sec. S7.2." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 130, + 430, + 481, + 514 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 430, + 481, + 514 + ], + "spans": [ + { + "bbox": [ + 130, + 430, + 481, + 514 + ], + "type": "text", + "content": "Results with the SwissSPAD2 sensor: The SwissSPAD2 sensor [49] is a prototype SPAD pixel array that can produce binary frames at a rate up to 97,700 FPS, with a resolution of " + }, + { + "bbox": [ + 130, + 430, + 481, + 514 + ], + "type": "inline_equation", + "content": "512 \\times 256" + }, + { + "bbox": [ + 130, + 430, + 481, + 514 + ], + "type": "text", + "content": ". In our experiments, we use binary frames captured directly (without inhibition) by the SPAD array as reference data and emulate on-sensor saturation look-ahead inhibition in software. As a pre-processing step, measurements at hot pixels are replaced with their nearest neighbors." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 130, + 514, + 481, + 634 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 514, + 481, + 634 + ], + "spans": [ + { + "bbox": [ + 130, + 514, + 481, + 634 + ], + "type": "text", + "content": "Fig. 7 shows the results of burst reconstruction under three lighting conditions. The raw data is a sequence of " + }, + { + "bbox": [ + 130, + 514, + 481, + 634 + ], + "type": "inline_equation", + "content": ">580,000" + }, + { + "bbox": [ + 130, + 514, + 481, + 634 + ], + "type": "text", + "content": " binary frames with scene radiance increasing rapidly by orders of magnitude (Fig. 7(b)), from " + }, + { + "bbox": [ + 130, + 514, + 481, + 634 + ], + "type": "inline_equation", + "content": "<1" + }, + { + "bbox": [ + 130, + 514, + 481, + 634 + ], + "type": "text", + "content": " lux to " + }, + { + "bbox": [ + 130, + 514, + 481, + 634 + ], + "type": "inline_equation", + "content": ">4,000" + }, + { + "bbox": [ + 130, + 514, + 481, + 634 + ], + "type": "text", + "content": " lux, measured separately with a light meter. For each of 47 equally-spaced keyframes, centered windows of 12,000 binary frames are extracted and processed as described above. Results for the full sequence can be found in the supplementary material. A static inhibition policy of regular sub-sampling (dropping 9 out of every 10 frames) is also applied, which yields a fixed " + }, + { + "bbox": [ + 130, + 514, + 481, + 634 + ], + "type": "inline_equation", + "content": "90\\%" + }, + { + "bbox": [ + 130, + 514, + 481, + 634 + ], + "type": "text", + "content": " reduction in both measurements and photon detections under all lighting conditions. Other subsampling factors are discussed in the supplement." + } + ] + } + ], + "index": 7 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "spans": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "type": "text", + "content": "12" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 166, + 91, + 235, + 100 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 166, + 91, + 235, + 100 + ], + "spans": [ + { + "bbox": [ + 166, + 91, + 235, + 100 + ], + "type": "text", + "content": "L. Koerner et al." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 133, + 642, + 481, + 665 + ], + "type": "page_footnote", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 642, + 481, + 665 + ], + "spans": [ + { + "bbox": [ + 133, + 642, + 481, + 665 + ], + "type": "text", + "content": "Sequence " + }, + { + "bbox": [ + 133, + 642, + 481, + 665 + ], + "type": "inline_equation", + "content": "T' \\coloneqq \\{1, 1, 1, 3, 3, 3, 8, 8, 25\\}" + }, + { + "bbox": [ + 133, + 642, + 481, + 665 + ], + "type": "text", + "content": " yielded similar results. No extensive search over the policy space was performed." + } + ] + } + ], + "index": 8 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 11 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 130, + 112, + 449, + 377 + ], + "blocks": [ + { + "bbox": [ + 130, + 112, + 449, + 377 + ], + "lines": [ + { + "bbox": [ + 130, + 112, + 449, + 377 + ], + "spans": [ + { + "bbox": [ + 130, + 112, + 449, + 377 + ], + "type": "image", + "image_path": "bcf274d6a24deaa3ce2c5d1e440ce97657328e2c77d381e78e05599f0ff2b257.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 130, + 385, + 482, + 562 + ], + "lines": [ + { + "bbox": [ + 130, + 385, + 482, + 562 + ], + "spans": [ + { + "bbox": [ + 130, + 385, + 482, + 562 + ], + "type": "text", + "content": "Fig. 7: Adaptive policies on video sequences enable stronger inhibition, preserve low-light details and, in bright-light, decouple flux and detection energy. (a) Burst reconstructions [34] for three keyframes with varying light levels: the top and bottom row differ by " + }, + { + "bbox": [ + 130, + 385, + 482, + 562 + ], + "type": "inline_equation", + "content": "\\approx 7" + }, + { + "bbox": [ + 130, + 385, + 482, + 562 + ], + "type": "text", + "content": " stops. [Images show the detection rates " + }, + { + "bbox": [ + 130, + 385, + 482, + 562 + ], + "type": "inline_equation", + "content": "Y" + }, + { + "bbox": [ + 130, + 385, + 482, + 562 + ], + "type": "text", + "content": " and are further gamma-compressed (" + }, + { + "bbox": [ + 130, + 385, + 482, + 562 + ], + "type": "inline_equation", + "content": "\\gamma = 0.4" + }, + { + "bbox": [ + 130, + 385, + 482, + 562 + ], + "type": "text", + "content": "). Complete results included in the supplement.] Left column results are from the original binary frames without inhibition, and the right column after sub-sampling " + }, + { + "bbox": [ + 130, + 385, + 482, + 562 + ], + "type": "inline_equation", + "content": "10 \\times" + }, + { + "bbox": [ + 130, + 385, + 482, + 562 + ], + "type": "text", + "content": " (a fixed " + }, + { + "bbox": [ + 130, + 385, + 482, + 562 + ], + "type": "inline_equation", + "content": "90\\%" + }, + { + "bbox": [ + 130, + 385, + 482, + 562 + ], + "type": "text", + "content": " inhibition). Middle column represents exposure bracketing combined with saturation look-ahead (see Fig. 4a for description). Under strong light (top row) the results are reasonable with both methods. However, plain sub-sampling loses details in lower light: notice the furniture and a person's outline in the middle & bottom rows, respectively. Inhibition is instead adaptive to flux. (b) Average exposure level for each keyframe in the sequence. (c,d) Per-keyframe and cumulative detection counts - inhibition ultimately results in fewer photons being detected over the whole sequence. (e) Number of measurements taken for each keyframe; reductions may be translated to energy savings during read-out. Plots in (c,d,e) are sub-sampled for clarity, and crossover points are marked by green arrows." + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_caption" + } + ], + "index": 2 + }, + { + "bbox": [ + 130, + 593, + 482, + 665 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 593, + 482, + 665 + ], + "spans": [ + { + "bbox": [ + 130, + 593, + 482, + 665 + ], + "type": "text", + "content": "The top row in Fig. 7(a) shows that under strong light a large fraction of photons (" + }, + { + "bbox": [ + 130, + 593, + 482, + 665 + ], + "type": "inline_equation", + "content": ">90\\%" + }, + { + "bbox": [ + 130, + 593, + 482, + 665 + ], + "type": "text", + "content": ") can be inhibited through the saturation look-ahead policy and still result in good image quality after burst reconstruction, thus spreading photon detections over a longer period of time to reduce avalanche power. Even simple sub-sampling yields good results in bright light, and may work well under controlled illumination conditions. However, the images in the middle and" + } + ] + } + ], + "index": 4 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 197, + 91, + 448, + 102 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 197, + 91, + 448, + 102 + ], + "spans": [ + { + "bbox": [ + 197, + 91, + 448, + 102 + ], + "type": "text", + "content": "Photon Inhibition for Energy-Efficient Single-Photon Imaging" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 470, + 91, + 481, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 470, + 91, + 481, + 100 + ], + "spans": [ + { + "bbox": [ + 470, + 91, + 481, + 100 + ], + "type": "text", + "content": "13" + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 12 + }, + { + "para_blocks": [ + { + "bbox": [ + 130, + 116, + 482, + 236 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 116, + 482, + 236 + ], + "spans": [ + { + "bbox": [ + 130, + 116, + 482, + 236 + ], + "type": "text", + "content": "bottom rows illustrate that this static inhibition policy results in excessive signal loss in lower light leading to a loss of details. Sub-sampling may be enhanced by adjusting exposure time and/or the sub-sampling factor in response to global flux but cannot simultaneously optimize for different light levels within a single frame: over-exposed regions may have clipping artifacts and under-exposed regions are prone to motion blur from incorrect burst reconstruction. The saturation look-ahead policy does adapt to local flux, and allocates relatively more detections to dim regions (see Fig. 1(d)). It thus decouples detection energy from flux (Fig. 7(c); cf. Fig. 1(a)), and results in cumulatively fewer detections than sub-sampling (Fig. 7(d)) due to it being more aggressive in strong light." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 130, + 236, + 482, + 297 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 236, + 482, + 297 + ], + "spans": [ + { + "bbox": [ + 130, + 236, + 482, + 297 + ], + "type": "text", + "content": "Apart from photography, the burst-reconstructed images can be used in computer vision applications, where even stronger inhibition is possible depending on the noise- or blur-robustness of the vision algorithm. Fig. 1(d) shows successful object detection with the YOLOv8 algorithm [29] applied to a burst-reconstructed image, with approximately " + }, + { + "bbox": [ + 130, + 236, + 482, + 297 + ], + "type": "inline_equation", + "content": "95\\%" + }, + { + "bbox": [ + 130, + 236, + 482, + 297 + ], + "type": "text", + "content": " photon detections inhibited." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 131, + 316, + 344, + 329 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 131, + 316, + 344, + 329 + ], + "spans": [ + { + "bbox": [ + 131, + 316, + 344, + 329 + ], + "type": "text", + "content": "8 Limitations and Future Outlook" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 130, + 342, + 482, + 462 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 342, + 482, + 462 + ], + "spans": [ + { + "bbox": [ + 130, + 342, + 482, + 462 + ], + "type": "text", + "content": "Implementation costs. Our focus in this paper is on reducing energy consumption due to avalanches. Although avalanches contribute significant energy as compared to on-sensor computations [2, 42, 47], an important next step is to design a holistic model that includes the energy consumption of computations and readout (leveraging Eq. 7). Our in-pixel computations — power-of-two multiplications which simplify to bit shifts, small spatio-temporal kernels no larger than " + }, + { + "bbox": [ + 130, + 342, + 482, + 462 + ], + "type": "inline_equation", + "content": "3 \\times 3 \\times 4" + }, + { + "bbox": [ + 130, + 342, + 482, + 462 + ], + "type": "text", + "content": " — are designed to be lightweight. Fortunately, computational SPAD imagers [2] with in-pixel memory and compute have recently been proposed, with " + }, + { + "bbox": [ + 130, + 342, + 482, + 462 + ], + "type": "inline_equation", + "content": "4 \\times 4" + }, + { + "bbox": [ + 130, + 342, + 482, + 462 + ], + "type": "text", + "content": " block of pixels having a 32-bit CPU and over 10 kbits of memory. See the supplement for an estimate of the required circuitry." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 130, + 465, + 482, + 585 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 465, + 482, + 585 + ], + "spans": [ + { + "bbox": [ + 130, + 465, + 482, + 585 + ], + "type": "text", + "content": "More complete models. The noise model used for efficiency metrics only accounts for photon Poisson noise and the quantization noise of Bernoulli samples. Expanding the noise model to include uncorrected pixel sensitivity variations, crosstalk, and afterpulsing may improve performance [4] by allocating measurements with an awareness of practical sensor limitations. Inhibition modulates the number of Bernoulli trials. An unbiased estimator for data-dependent stopping of Bernoulli trials is known [23], yet is not applied in this work since it does not precisely match our situation. Our simulations (SSIM) suggest that bias is less significant than image noise, yet future analytical studies are needed over a range of conditions." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 130, + 588, + 482, + 661 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 588, + 482, + 661 + ], + "spans": [ + { + "bbox": [ + 130, + 588, + 482, + 661 + ], + "type": "text", + "content": "Generalization to other tasks. To generalize the proposed approaches to a variety of vision tasks, task-specific quality metrics must be defined to pose optimization problems for each task. As an example, for the image reconstruction task, Suppl. Sec. 3.1 shows an analytical optimization using the MSE metric when constrained by photon detections. Other tasks could be approached similarly, yet may need to be optimized empirically." + } + ] + } + ], + "index": 7 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "spans": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "type": "text", + "content": "14" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 166, + 91, + 235, + 100 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 166, + 91, + 235, + 100 + ], + "spans": [ + { + "bbox": [ + 166, + 91, + 235, + 100 + ], + "type": "text", + "content": "L. Koerner et al." + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 13 + }, + { + "para_blocks": [ + { + "bbox": [ + 133, + 114, + 240, + 129 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 114, + 240, + 129 + ], + "spans": [ + { + "bbox": [ + 133, + 114, + 240, + 129 + ], + "type": "text", + "content": "Acknowledgments" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 130, + 143, + 482, + 204 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 143, + 482, + 204 + ], + "spans": [ + { + "bbox": [ + 130, + 143, + 482, + 204 + ], + "type": "text", + "content": "The authors acknowledge the Minnesota Supercomputing Institute (MSI) at the University of Minnesota for providing resources. A.I. was supported in part by NSF ECCS-2138471. S.G and M.G. were supported in part by NSF CAREER award 1943149, NSF award CNS-2107060, and Wisconsin Alumni Research Foundation (WARF)." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 132, + 224, + 197, + 237 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 224, + 197, + 237 + ], + "spans": [ + { + "bbox": [ + 132, + 224, + 197, + 237 + ], + "type": "text", + "content": "References" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 133, + 251, + 481, + 665 + ], + "type": "list", + "angle": 0, + "index": 16, + "blocks": [ + { + "bbox": [ + 138, + 251, + 481, + 286 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 251, + 481, + 286 + ], + "spans": [ + { + "bbox": [ + 138, + 251, + 481, + 286 + ], + "type": "text", + "content": "1. Arbeláez, P., Maire, M., Fowlkes, C., Malik, J.: Contour Detection and Hierarchical Image Segmentation. IEEE Transactions on Pattern Analysis and Machine Intelligence 33(5), 898-916 (May 2011). https://doi.org/10.1109/TPAMI.2010.161" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 138, + 287, + 481, + 308 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 287, + 481, + 308 + ], + "spans": [ + { + "bbox": [ + 138, + 287, + 481, + 308 + ], + "type": "text", + "content": "2. Ardelean, A.: Computational Imaging SPAD Cameras. Ph.D. thesis, EPFL (2023). https://doi.org/10.5075/epfl-thesis-9501" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 138, + 309, + 481, + 331 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 309, + 481, + 331 + ], + "spans": [ + { + "bbox": [ + 138, + 309, + 481, + 331 + ], + "type": "text", + "content": "3. Barlow, H.B.: Summation and inhibition in the frog's retina. The Journal of physiology 119(1), 69 (1953)" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 138, + 332, + 481, + 374 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 332, + 481, + 374 + ], + "spans": [ + { + "bbox": [ + 138, + 332, + 481, + 374 + ], + "type": "text", + "content": "4. Bian, L., Song, H., Peng, L., Chang, X., Yang, X., Horstmeyer, R., Ye, L., Zhu, C., Qin, T., Zheng, D., Zhang, J.: High-resolution single-photon imaging with physics-informed deep learning. Nature Communications 14(1), 5902 (Sep 2023). https://doi.org/10.1038/s41467-023-41597-9" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 138, + 376, + 481, + 409 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 376, + 481, + 409 + ], + "spans": [ + { + "bbox": [ + 138, + 376, + 481, + 409 + ], + "type": "text", + "content": "5. Boso, G., Buttafava, M., Villa, F., Tosi, A.: Low-cost and compact single-photon counter based on a CMOS SPAD smart pixel. IEEE Photonics Technology Letters 27(23), 2504-2507 (2015)" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 138, + 410, + 481, + 453 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 410, + 481, + 453 + ], + "spans": [ + { + "bbox": [ + 138, + 410, + 481, + 453 + ], + "type": "text", + "content": "6. Canon Inc.: Canon Launches MS-500 - The World's First Ultra-High-Sensitivity Interchangeable-Lens SPAD Sensor Camera. https://www.usa.canon.com/newsroom/2023/20230801-ms500, Canon Press Release 8/1/2023. Accessed 2/25/2024." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 138, + 454, + 481, + 487 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 454, + 481, + 487 + ], + "spans": [ + { + "bbox": [ + 138, + 454, + 481, + 487 + ], + "type": "text", + "content": "7. Carey, S.J., Lopich, A., Barr, D.R., Wang, B., Dudek, P.: A 100,000 fps vision sensor with embedded 535 GOPS/W " + }, + { + "bbox": [ + 138, + 454, + 481, + 487 + ], + "type": "inline_equation", + "content": "256 \\times 256" + }, + { + "bbox": [ + 138, + 454, + 481, + 487 + ], + "type": "text", + "content": " SIMD processor array. In: 2013 Symposium on VLSI Circuits. pp. C182-C183 (Jun 2013)" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 138, + 488, + 481, + 519 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 488, + 481, + 519 + ], + "spans": [ + { + "bbox": [ + 138, + 488, + 481, + 519 + ], + "type": "text", + "content": "8. Chan, S.H.: What Does a One-Bit Quanta Image Sensor Offer? IEEE Transactions on Computational Imaging 8, 770-783 (2022). https://doi.org/10.1109/TCI.2022.3202012" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 138, + 521, + 481, + 564 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 521, + 481, + 564 + ], + "spans": [ + { + "bbox": [ + 138, + 521, + 481, + 564 + ], + "type": "text", + "content": "9. Charbon, E., Bruschini, C., Lee, M.J.: 3D-Stacked CMOS SPAD Image Sensors: Technology and Applications. In: 2018 25th IEEE International Conference on Electronics, Circuits and Systems (ICECS). pp. 1-4 (Dec 2018). https://doi.org/10.1109/ICECS.2018.8617983" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 133, + 566, + 481, + 609 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 566, + 481, + 609 + ], + "spans": [ + { + "bbox": [ + 133, + 566, + 481, + 609 + ], + "type": "text", + "content": "10. Chi, Y., Gnanasambandam, A., Koltun, V., Chan, S.H.: Dynamic low-light imaging with quanta image sensors. In: Computer Vision-ECCV 2020: 16th European Conference, Glasgow, UK, August 23-28, 2020, Proceedings, Part XXI 16. pp. 122-138. Springer (2020)" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 133, + 611, + 481, + 665 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 611, + 481, + 665 + ], + "spans": [ + { + "bbox": [ + 133, + 611, + 481, + 665 + ], + "type": "text", + "content": "11. Della Rocca, F.M., Mai, H., Hutchings, S.W., Al Abbas, T., Buckbee, K., Tsiamis, A., Lomax, P., Gyongy, I., Dutton, N.A., Henderson, R.K.: A " + }, + { + "bbox": [ + 133, + 611, + 481, + 665 + ], + "type": "inline_equation", + "content": "128 \\times 128" + }, + { + "bbox": [ + 133, + 611, + 481, + 665 + ], + "type": "text", + "content": " SPAD motion-triggered time-of-flight image sensor with in-pixel histogram and column-parallel vision processor. IEEE Journal of Solid-State Circuits 55(7), 1762-1775 (2020). https://doi.org/10.1109/JSSC.2020.2993722" + } + ] + } + ], + "index": 15 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 197, + 91, + 447, + 102 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 197, + 91, + 447, + 102 + ], + "spans": [ + { + "bbox": [ + 197, + 91, + 447, + 102 + ], + "type": "text", + "content": "Photon Inhibition for Energy-Efficient Single-Photon Imaging" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 470, + 92, + 480, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 470, + 92, + 480, + 100 + ], + "spans": [ + { + "bbox": [ + 470, + 92, + 480, + 100 + ], + "type": "text", + "content": "15" + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 14 + }, + { + "para_blocks": [ + { + "bbox": [ + 133, + 117, + 480, + 665 + ], + "type": "list", + "angle": 0, + "index": 18, + "blocks": [ + { + "bbox": [ + 133, + 117, + 480, + 148 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 117, + 480, + 148 + ], + "spans": [ + { + "bbox": [ + 133, + 117, + 480, + 148 + ], + "type": "text", + "content": "12. Diamond, J.S.: Inhibitory interneurons in the retina: Types, circuitry, and function. Annual Review of Vision Science 3(1), 1-24 (2017). https://doi.org/10.1146/annurev-vision-102016-061345" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 133, + 149, + 480, + 181 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 149, + 480, + 181 + ], + "spans": [ + { + "bbox": [ + 133, + 149, + 480, + 181 + ], + "type": "text", + "content": "13. Dollar, P., Zitnick, C.L.: Structured Forests for Fast Edge Detection. In: Proceedings of the IEEE International Conference on Computer Vision. pp. 1841-1848 (2013)" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 133, + 182, + 480, + 224 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 182, + 480, + 224 + ], + "spans": [ + { + "bbox": [ + 133, + 182, + 480, + 224 + ], + "type": "text", + "content": "14. Dutton, N.A.W., Al Abbas, T., Gyongy, I., Mattioli Della Rocca, F., Henderson, R.K.: High Dynamic Range Imaging at the Quantum Limit with Single Photon Avalanche Diode-Based Image Sensors. Sensors 18(4), 1166 (Apr 2018). https://doi.org/10.3390/s18041166" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 133, + 225, + 480, + 257 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 225, + 480, + 257 + ], + "spans": [ + { + "bbox": [ + 133, + 225, + 480, + 257 + ], + "type": "text", + "content": "15. Fossum, E.R.: Modeling the Performance of Single-Bit and Multi-Bit Quanta Image Sensors. IEEE Journal of the Electron Devices Society 1(9), 166-174 (Sep 2013). https://doi.org/10.1109/JEDS.2013.2284054" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 133, + 258, + 480, + 289 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 258, + 480, + 289 + ], + "spans": [ + { + "bbox": [ + 133, + 258, + 480, + 289 + ], + "type": "text", + "content": "16. Franke, K., Berens, P., Schubert, T., Bethge, M., Euler, T., Baden, T.: Inhibition decorrelates visual feature representations in the inner retina. Nature 542(7642), 439-444 (Feb 2017). https://doi.org/10.1038/nature21394" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 133, + 289, + 480, + 332 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 289, + 480, + 332 + ], + "spans": [ + { + "bbox": [ + 133, + 289, + 480, + 332 + ], + "type": "text", + "content": "17. Gallego, G., Delbrück, T., Orchard, G., Bartolozzi, C., Taba, B., Censi, A., Leutenegger, S., Davison, A.J., Conradt, J., Daniilidis, K., et al.: Event-based vision: A survey. IEEE Transactions on Pattern Analysis and Machine Intelligence 44(1), 154-180 (2020)" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 133, + 333, + 480, + 374 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 333, + 480, + 374 + ], + "spans": [ + { + "bbox": [ + 133, + 333, + 480, + 374 + ], + "type": "text", + "content": "18. Gnanasambandam, A., Chan, S.H.: HDR Imaging With Quanta Image Sensors: Theoretical Limits and Optimal Reconstruction. IEEE Transactions on Computational Imaging 6, 1571-1585 (2020). https://doi.org/10.1109/TCI.2020.3041093" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 133, + 376, + 480, + 407 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 376, + 480, + 407 + ], + "spans": [ + { + "bbox": [ + 133, + 376, + 480, + 407 + ], + "type": "text", + "content": "19. Gnanasambandam, A., Chan, S.H.: Exposure-Referred Signal-to-Noise Ratio for Digital Image Sensors. IEEE Transactions on Computational Imaging 8, 561-575 (Jun 2022). https://doi.org/10.1109/TCI.2022.3187657" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 133, + 408, + 480, + 449 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 408, + 480, + 449 + ], + "spans": [ + { + "bbox": [ + 133, + 408, + 480, + 449 + ], + "type": "text", + "content": "20. Gupta, M., Iso, D., Nayar, S.K.: Fibonacci Exposure Bracketing for High Dynamic Range Imaging. In: 2013 IEEE International Conference on Computer Vision. pp. 1473-1480. IEEE, Sydney, Australia (Dec 2013). https://doi.org/10.1109/ICCV.2013.186" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 133, + 450, + 480, + 472 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 450, + 480, + 472 + ], + "spans": [ + { + "bbox": [ + 133, + 450, + 480, + 472 + ], + "type": "text", + "content": "21. Gyongy, I., Dutton, N.A., Henderson, R.K.: Direct time-of-flight single-photon imaging. IEEE Transactions on Electron Devices 69(6), 2794-2805 (2021)" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 133, + 472, + 480, + 504 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 472, + 480, + 504 + ], + "spans": [ + { + "bbox": [ + 133, + 472, + 480, + 504 + ], + "type": "text", + "content": "22. Gyongy, I., Martin, G.M., Turpin, A., Ruget, A., Halimi, A., Henderson, R., Leach, J.: High-speed vision with a 3D-stacked SPAD image sensor. In: Advanced Photon Counting Techniques XV. vol. 11721, p. 1172105. SPIE (2021)" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 133, + 504, + 480, + 525 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 504, + 480, + 525 + ], + "spans": [ + { + "bbox": [ + 133, + 504, + 480, + 525 + ], + "type": "text", + "content": "23. Haldane, J.B.S.: On a Method Of Estimating Frequencies. Biometrika 33(3), 222-225 (Nov 1945). https://doi.org/10.1093/biomet/33.3.222" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 133, + 525, + 480, + 579 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 525, + 480, + 579 + ], + "spans": [ + { + "bbox": [ + 133, + 525, + 480, + 579 + ], + "type": "text", + "content": "24. Henderson, R.K., Johnston, N., Hutchings, S.W., Gyongy, I., Al Abbas, T., Dutton, N., Tyler, M., Chan, S., Leach, J.: " + }, + { + "bbox": [ + 133, + 525, + 480, + 579 + ], + "type": "inline_equation", + "content": "256 \\times 256" + }, + { + "bbox": [ + 133, + 525, + 480, + 579 + ], + "type": "text", + "content": " 40nm/90nm CMOS 3D-stacked 120 dB dynamic-range reconfigurable time-resolved SPAD imager. In: 2019 IEEE International Solid-State Circuits Conference-(ISSCC). pp. 106-108. IEEE (2019). https://doi.org/10.1109/ISSCC.2019.8662355" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 133, + 580, + 480, + 601 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 580, + 480, + 601 + ], + "spans": [ + { + "bbox": [ + 133, + 580, + 480, + 601 + ], + "type": "text", + "content": "25. Iams, H., Salzberg, B.: The secondary emission phototube. Proceedings of the Institute of Radio Engineers 23(1), 55-64 (1935)" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 133, + 601, + 480, + 633 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 601, + 480, + 633 + ], + "spans": [ + { + "bbox": [ + 133, + 601, + 480, + 633 + ], + "type": "text", + "content": "26. Ingle, A., Seets, T., Buttafava, M., Gupta, S., Tosi, A., Gupta, M., Velten, A.: Passive Inter-Photon Imaging. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 8585–8595 (2021)" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 133, + 633, + 480, + 665 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 633, + 480, + 665 + ], + "spans": [ + { + "bbox": [ + 133, + 633, + 480, + 665 + ], + "type": "text", + "content": "27. Ingle, A., Velten, A., Gupta, M.: High Flux Passive Imaging With Single-Photon Sensors. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 6760-6769 (2019)" + } + ] + } + ], + "index": 17 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "spans": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "type": "text", + "content": "16" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 166, + 91, + 235, + 100 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 166, + 91, + 235, + 100 + ], + "spans": [ + { + "bbox": [ + 166, + 91, + 235, + 100 + ], + "type": "text", + "content": "L. Koerner et al." + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 15 + }, + { + "para_blocks": [ + { + "bbox": [ + 132, + 116, + 481, + 665 + ], + "type": "list", + "angle": 0, + "index": 16, + "blocks": [ + { + "bbox": [ + 132, + 116, + 481, + 148 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 116, + 481, + 148 + ], + "spans": [ + { + "bbox": [ + 132, + 116, + 481, + 148 + ], + "type": "text", + "content": "28. J Yoshida: Breaking Down iPad Pro 11's LiDAR Scanner. https://www.eetimes.com/breaking-down-ipad-pro-11s-lidar-scanner/, EE Times 6/5/2020. Accessed 5/6/2021." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 132, + 150, + 481, + 171 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 150, + 481, + 171 + ], + "spans": [ + { + "bbox": [ + 132, + 150, + 481, + 171 + ], + "type": "text", + "content": "29. Jocher, G., Chaurasia, A., Qiu, J.: Ultralytics YOLOv8 (Jan 2023), https://github.com/ultralytics/ultralytics, Accessed 3/6/2024" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 132, + 172, + 481, + 215 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 172, + 481, + 215 + ], + "spans": [ + { + "bbox": [ + 132, + 172, + 481, + 215 + ], + "type": "text", + "content": "30. Koppal, S.J., Gkioulekas, I., Young, T., Park, H., Crozier, K.B., Barrows, G.L., Zickler, T.: Toward Wide-Angle Microvision Sensors. IEEE Transactions on Pattern Analysis and Machine Intelligence 35(12), 2982-2996 (Dec 2013). https://doi.org/10.1109/TPAMI.2013.22" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 132, + 216, + 481, + 258 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 216, + 481, + 258 + ], + "spans": [ + { + "bbox": [ + 132, + 216, + 481, + 258 + ], + "type": "text", + "content": "31. Liu, Y., Gutierrez-Barragan, F., Ingle, A., Gupta, M., Velten, A.: Single-Photon Camera Guided Extreme Dynamic Range Imaging. In: Proceedings of the IEEE/CVF Winter Conference on Applications of Computer Vision. pp. 1575-1585 (2022)" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 132, + 259, + 481, + 293 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 259, + 481, + 293 + ], + "spans": [ + { + "bbox": [ + 132, + 259, + 481, + 293 + ], + "type": "text", + "content": "32. Ma, J., Masoodian, S., Starkey, D.A., Fossum, E.R.: Photon-number-resolving megapixel image sensor at room temperature without avalanche gain. Optica 4(12), 1474-1481 (Dec 2017). https://doi.org/10.1364/OPTICA.4.001474" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 132, + 293, + 481, + 335 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 293, + 481, + 335 + ], + "spans": [ + { + "bbox": [ + 132, + 293, + 481, + 335 + ], + "type": "text", + "content": "33. Ma, J., Zhang, D., Elgendy, O.A., Masoodian, S.: A 0.19e- rms Read Noise 16.7Mpixel Stacked Quanta Image Sensor With 1.1 Mm-Pitch Backside Illuminated Pixels. IEEE Electron Device Letters 42(6), 891-894 (Jun 2021). https: //doi.org/10.1109/LED.2021.3072842" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 132, + 336, + 481, + 369 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 336, + 481, + 369 + ], + "spans": [ + { + "bbox": [ + 132, + 336, + 481, + 369 + ], + "type": "text", + "content": "34. Ma, S., Gupta, S., Ulku, A.C., Bruschini, C., Charbon, E., Gupta, M.: Quanta burst photography. ACM Transactions on Graphics 39(4), 79:1-79:16 (Aug 2020). https://doi.org/10.1145/3386569.3392470" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 132, + 369, + 481, + 402 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 369, + 481, + 402 + ], + "spans": [ + { + "bbox": [ + 132, + 369, + 481, + 402 + ], + "type": "text", + "content": "35. Ma, S., Mos, P., Charbon, E., Gupta, M.: Burst Vision Using Single-Photon Cameras. In: Proceedings of the IEEE/CVF Winter Conference on Applications of Computer Vision. pp. 5375-5385 (2023)" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 132, + 403, + 481, + 445 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 403, + 481, + 445 + ], + "spans": [ + { + "bbox": [ + 132, + 403, + 481, + 445 + ], + "type": "text", + "content": "36. Medin, S.C., Murray-Bruce, J., Castanón, D., Goyal, V.K.: Beyond Binomial and Negative Binomial: Adaptation in Bernoulli Parameter Estimation. IEEE Transactions on Computational Imaging 5(4), 570-584 (Dec 2019). https://doi.org/10.1109/TCI.2019.2913108" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 132, + 446, + 481, + 521 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 446, + 481, + 521 + ], + "spans": [ + { + "bbox": [ + 132, + 446, + 481, + 521 + ], + "type": "text", + "content": "37. Morimoto, K., Iwata, J., Shinohara, M., Sekine, H., Abdelghafar, A., Tsuchiya, H., Kuroda, Y., Tojima, K., Endo, W., Maehashi, Y., Ota, Y., Sasago, T., Maekawa, S., Hikosaka, S., Kanou, T., Kato, A., Tezuka, T., Yoshizaki, S., Ogawa, T., Uehira, K., Ehara, A., Inui, F., Matsuno, Y., Sakurai, K., Ichikawa, T.: 3.2 Megapixel 3D-Stacked Charge Focusing SPAD for Low-Light Imaging and Depth Sensing. In: 2021 IEEE International Electron Devices Meeting (IEDM). pp. 20.2.1–20.2.4 (Dec 2021). https://doi.org/10.1109/IEDM19574.2021.9720605" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 132, + 522, + 481, + 565 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 522, + 481, + 565 + ], + "spans": [ + { + "bbox": [ + 132, + 522, + 481, + 565 + ], + "type": "text", + "content": "38. Morimoto, K., Ardelean, A., Wu, M.L., Ulku, A.C., Antolovic, I.M., Bruschini, C., Charbon, E.: Megapixel time-gated SPAD image sensor for 2D and 3D imaging applications. Optica 7(4), 346-354 (2020). https://doi.org/10.1364/OPTICA.386574" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 132, + 566, + 481, + 588 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 566, + 481, + 588 + ], + "spans": [ + { + "bbox": [ + 132, + 566, + 481, + 588 + ], + "type": "text", + "content": "39. Morimoto, K., Charbon, E.: A scaling law for SPAD pixel miniaturization. Sensors 21(10), 3447 (2021)" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 132, + 589, + 481, + 643 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 589, + 481, + 643 + ], + "spans": [ + { + "bbox": [ + 132, + 589, + 481, + 643 + ], + "type": "text", + "content": "40. Ogi, J., Sano, F., Nakata, T., Kubo, Y., Onishi, W., Koswaththaghe, C., Mochizuki, T., Tashiro, Y., Hizu, K., Takatsuka, T., et al.: A " + }, + { + "bbox": [ + 132, + 589, + 481, + 643 + ], + "type": "inline_equation", + "content": "3.06\\mu \\mathrm{m}" + }, + { + "bbox": [ + 132, + 589, + 481, + 643 + ], + "type": "text", + "content": " spad pixel with embedded metal contact and power grid on deep trench pixel isolation for high-resolution photon-counting. In: 2023 International Image Sensor Workshop (IISW) (May 2023)" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 132, + 643, + 481, + 665 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 643, + 481, + 665 + ], + "spans": [ + { + "bbox": [ + 132, + 643, + 481, + 665 + ], + "type": "text", + "content": "41. Ogi, J., Takatsuka, T., Hizu, K., Inaoka, Y., Zhu, H., Tochigi, Y., Tashiro, Y., Sano, F., Murakawa, Y., Nakamura, M., Oike, Y.: A 250fps 124dB Dynamic-Range" + } + ] + } + ], + "index": 15 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 197, + 91, + 447, + 102 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 197, + 91, + 447, + 102 + ], + "spans": [ + { + "bbox": [ + 197, + 91, + 447, + 102 + ], + "type": "text", + "content": "Photon Inhibition for Energy-Efficient Single-Photon Imaging" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 470, + 92, + 480, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 470, + 92, + 480, + 100 + ], + "spans": [ + { + "bbox": [ + 470, + 92, + 480, + 100 + ], + "type": "text", + "content": "17" + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 16 + }, + { + "para_blocks": [ + { + "bbox": [ + 132, + 116, + 481, + 666 + ], + "type": "list", + "angle": 0, + "index": 15, + "blocks": [ + { + "bbox": [ + 147, + 116, + 481, + 160 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 147, + 116, + 481, + 160 + ], + "spans": [ + { + "bbox": [ + 147, + 116, + 481, + 160 + ], + "type": "text", + "content": "SPAD Image Sensor Stacked with Pixel-Parallel Photon Counter Employing SubFrame Extrapolating Architecture for Motion Artifact Suppression. In: 2021 IEEE International Solid-State Circuits Conference (ISSCC). vol. 64, pp. 113-115 (Feb 2021). https://doi.org/10.1109/ISSCC42613.2021.9365977" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 133, + 161, + 481, + 247 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 161, + 481, + 247 + ], + "spans": [ + { + "bbox": [ + 133, + 161, + 481, + 247 + ], + "type": "text", + "content": "42. Ota, Y., Morimoto, K., Sasago, T., Shinohara, M., Kuroda, Y., Endo, W., Maehashi, Y., Maekawa, S., Tsuchiya, H., Abdelahafar, A., Hikosaka, S., Motoyama, M., Tojima, K., Uehira, K., Iwata, J., Inui, F., Matsuno, Y., Sakurai, K., Ichikawa, T.: A 0.37W 143dB-Dynamic-Range 1Mpixel Backside-Illuminated Charge-Focusing SPAD Image Sensor with Pixel-Wise Exposure Control and Adaptive Clocked Recharging. In: 2022 IEEE International Solid-State Circuits Conference (ISSCC). vol. 65, pp. 94-96 (Feb 2022). https://doi.org/10.1109/ISSCC42614.2022.9731644" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 132, + 248, + 481, + 282 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 248, + 481, + 282 + ], + "spans": [ + { + "bbox": [ + 132, + 248, + 481, + 282 + ], + "type": "text", + "content": "43. Po, R., Pediredla, A., Gkioulekas, I.: Adaptive gating for single-photon 3d imaging. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 16354-16363 (2022)" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 132, + 282, + 481, + 315 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 282, + 481, + 315 + ], + "spans": [ + { + "bbox": [ + 132, + 282, + 481, + 315 + ], + "type": "text", + "content": "44. Severini, F., Cusini, I., Berretta, D., Pasquinelli, K., Incoronato, A., Villa, F.: SPAD pixel with sub-ns dead-time for high-count rate applications. IEEE Journal of Selected Topics in Quantum Electronics 28(2: Optical Detectors), 1-8 (2021)" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 132, + 315, + 481, + 358 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 315, + 481, + 358 + ], + "spans": [ + { + "bbox": [ + 132, + 315, + 481, + 358 + ], + "type": "text", + "content": "45. Shenzhen CM Technology company Ltd: 16 MP MIPI Camera Module with SONY IMX206 sensor. http://www-camera-module.com/product/mipicameramodule/16mp-mipi-camera-module-sony-imx206-sensor.html (Nov 2023)" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 132, + 358, + 481, + 392 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 358, + 481, + 392 + ], + "spans": [ + { + "bbox": [ + 132, + 358, + 481, + 392 + ], + "type": "text", + "content": "46. Sundar, V., Ardelean, A., Swedish, T., Brusschini, C., Charbon, E., Gupta, M.: Sodacam: Software-defined cameras via single-photon imaging. In: Proceedings of the IEEE International Conference on Computer Vision (ICCV) (2023)" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 132, + 392, + 481, + 468 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 392, + 481, + 468 + ], + "spans": [ + { + "bbox": [ + 132, + 392, + 481, + 468 + ], + "type": "text", + "content": "47. Takatsuka, T., Ogi, J., Ikeda, Y., Hizu, K., Inaoka, Y., Sakama, S., Watanabe, I., Ishikawa, T., Shimada, S., Suzuki, J., Maeda, H., Toshima, K., Nonaka, Y., Yamamura, A., Ozawa, H., Koga, F., Oike, Y.: A 3.36 " + }, + { + "bbox": [ + 132, + 392, + 481, + 468 + ], + "type": "inline_equation", + "content": "\\mu" + }, + { + "bbox": [ + 132, + 392, + 481, + 468 + ], + "type": "text", + "content": "m-pitch SPAD photon-counting image sensor using clustered multi-cycle clocked recharging technique with intermediate most-significant-bit readout. In: 2023 IEEE Symposium on VLSI Technology and Circuits (VLSI Technology and Circuits). pp. 1-2 (Jun 2023). https://doi.org/10.23919/VLSITechnologyandCir57934.2023.10185241" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 132, + 468, + 481, + 512 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 468, + 481, + 512 + ], + "spans": [ + { + "bbox": [ + 132, + 468, + 481, + 512 + ], + "type": "text", + "content": "48. Tilmon, B., Sun, Z., Koppal, S.J., Wu, Y., Evangelidis, G., Zahreddine, R., Krishnan, G., Ma, S., Wang, J.: Energy-Efficient Adaptive 3D Sensing. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 5054-5063 (2023)" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 132, + 512, + 481, + 555 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 512, + 481, + 555 + ], + "spans": [ + { + "bbox": [ + 132, + 512, + 481, + 555 + ], + "type": "text", + "content": "49. Ulku, A.C., Bruschini, C., Antolovic, I.M., Kuo, Y., Ankri, R., Weiss, S., Michalet, X., Charbon, E.: A " + }, + { + "bbox": [ + 132, + 512, + 481, + 555 + ], + "type": "inline_equation", + "content": "512 \\times 512" + }, + { + "bbox": [ + 132, + 512, + 481, + 555 + ], + "type": "text", + "content": " SPAD Image Sensor With Integrated Gating for Widefield FLIM. IEEE Journal of Selected Topics in Quantum Electronics 25(1), 1-12 (Jan 2019). https://doi.org/10.1109/jstqe.2018.2867439" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 132, + 555, + 481, + 578 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 555, + 481, + 578 + ], + "spans": [ + { + "bbox": [ + 132, + 555, + 481, + 578 + ], + "type": "text", + "content": "50. Wang, X.: Laplacian operator-based edge detectors. IEEE Transactions on Pattern Analysis and Machine Intelligence 29(5), 886-890 (2007)" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 132, + 578, + 481, + 611 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 578, + 481, + 611 + ], + "spans": [ + { + "bbox": [ + 132, + 578, + 481, + 611 + ], + "type": "text", + "content": "51. Wang, Z., Bovik, A.C., Sheikh, H.R., Simoncelli, E.P.: Image quality assessment: From error visibility to structural similarity. IEEE Transactions on Image Processing 13(4), 600-612 (2004)" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 132, + 611, + 481, + 643 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 611, + 481, + 643 + ], + "spans": [ + { + "bbox": [ + 132, + 611, + 481, + 643 + ], + "type": "text", + "content": "52. Wayne, M.A., Bienfang, J.C., Migdall, A.L.: Low-noise photon counting above " + }, + { + "bbox": [ + 132, + 611, + 481, + 643 + ], + "type": "inline_equation", + "content": "100 \\times 106" + }, + { + "bbox": [ + 132, + 611, + 481, + 643 + ], + "type": "text", + "content": " counts per second with a high-efficiency reach-through single-photon avalanche diode system. Applied Physics Letters 118(13) (2021)" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 132, + 643, + 481, + 666 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 643, + 481, + 666 + ], + "spans": [ + { + "bbox": [ + 132, + 643, + 481, + 666 + ], + "type": "text", + "content": "53. Xie, S., Tu, Z.: Holistically-nested edge detection. In: Proceedings of IEEE International Conference on Computer Vision. pp. 1395-1403 (2015)" + } + ] + } + ], + "index": 14 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "spans": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "type": "text", + "content": "18" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 166, + 91, + 235, + 100 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 166, + 91, + 235, + 100 + ], + "spans": [ + { + "bbox": [ + 166, + 91, + 235, + 100 + ], + "type": "text", + "content": "L. Koerner et al." + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 17 + }, + { + "para_blocks": [ + { + "bbox": [ + 132, + 116, + 481, + 172 + ], + "type": "list", + "angle": 0, + "index": 4, + "blocks": [ + { + "bbox": [ + 132, + 116, + 481, + 138 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 116, + 481, + 138 + ], + "spans": [ + { + "bbox": [ + 132, + 116, + 481, + 138 + ], + "type": "text", + "content": "54. Xu, Y., Lu, J., Wu, Z.: A compact high-speed active quenching and recharging circuit for SPAD detectors. IEEE Photonics Journal 12(5), 1-8 (2020)" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 132, + 139, + 481, + 172 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 139, + 481, + 172 + ], + "spans": [ + { + "bbox": [ + 132, + 139, + 481, + 172 + ], + "type": "text", + "content": "55. Yang, F., Lu, Y.M., Sbaiz, L., Vetterli, M.: Bits from photons: Oversampled image acquisition using binary poisson statistics. IEEE Transactions on image processing 21(4), 1421-1436 (2011)" + } + ] + } + ], + "index": 3 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 196, + 91, + 447, + 102 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 196, + 91, + 447, + 102 + ], + "spans": [ + { + "bbox": [ + 196, + 91, + 447, + 102 + ], + "type": "text", + "content": "Photon Inhibition for Energy-Efficient Single-Photon Imaging" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 470, + 92, + 481, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 470, + 92, + 481, + 100 + ], + "spans": [ + { + "bbox": [ + 470, + 92, + 481, + 100 + ], + "type": "text", + "content": "19" + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 18 + } + ], + "_backend": "vlm", + "_version_name": "2.6.4" +} \ No newline at end of file diff --git a/2024/Photorealistic Object Insertion with Diffusion-Guided Inverse Rendering/27eaedd8-53d6-44df-a63a-fb92723dff8a_content_list.json b/2024/Photorealistic Object Insertion with Diffusion-Guided Inverse Rendering/27eaedd8-53d6-44df-a63a-fb92723dff8a_content_list.json new file mode 100644 index 0000000000000000000000000000000000000000..9e10c5a20b1d69543840a828cc3cf938c7a9333b --- /dev/null +++ b/2024/Photorealistic Object Insertion with Diffusion-Guided Inverse Rendering/27eaedd8-53d6-44df-a63a-fb92723dff8a_content_list.json @@ -0,0 +1,2195 @@ +[ + { + "type": "text", + "text": "Photorealistic Object Insertion with Diffusion-Guided Inverse Rendering", + "text_level": 1, + "bbox": [ + 295, + 140, + 705, + 185 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Ruofan Liang $^{1,2,3}$ , Zan Gojcic $^{1}$ , Merlin Nimier-David $^{1}$ , David Acuna $^{1}$ , Nandita Vijaykumar $^{2,3}$ , Sanja Fidler $^{1,2,3}$ , and Zian Wang $^{1,2,3}$", + "bbox": [ + 251, + 210, + 751, + 243 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "$^{1}$ NVIDIA $^{2}$ University of Toronto $^{3}$ Vector Institute", + "bbox": [ + 316, + 253, + 683, + 268 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Abstract. The correct insertion of virtual objects in images of real-world scenes requires a deep understanding of the scene's lighting, geometry and materials, as well as the image formation process. While recent large-scale diffusion models have shown strong generative and inpainting capabilities, we find that current models do not sufficiently \"understand\" the scene shown in a single picture to generate consistent lighting effects (shadows, bright reflections, etc.) while preserving the identity and details of the composited object. We propose using a personalized large diffusion model as guidance to a physically based inverse rendering process. Our method recovers scene lighting and tone-mapping parameters, allowing the photorealistic composition of arbitrary virtual objects in single frames or videos of indoor or outdoor scenes. Our physically based pipeline further enables automatic materials and tone-mapping refinement.", + "bbox": [ + 261, + 304, + 738, + 484 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Keywords: Inverse rendering $\\cdot$ Diffusion models $\\cdot$ Personalization $\\cdot$ Virtual object insertion $\\cdot$ Physically based rendering", + "bbox": [ + 261, + 498, + 740, + 526 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "1 Introduction", + "text_level": 1, + "bbox": [ + 217, + 551, + 374, + 566 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Virtual object insertion enables a range of applications from virtual production, to interactive gaming and synthetic data generation. To produce photorealistic insertions, the interactions between the virtual objects and the environment need to be modeled faithfully, such as accurate specular highlights and shadows.", + "bbox": [ + 215, + 582, + 785, + 642 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "A standard virtual object insertion pipeline typically includes three key steps: i) lighting estimation from the input image, ii) 3D proxy geometry creation, and iii) composited image rendering in a rendering engine. However, the first and arguably most important step is still an open research question. Lighting estimation is particularly challenging when dealing with limited inputs such as a single image, captured using a consumer device with a low dynamic range. Indeed, inverse rendering is a fundamentally ill-posed problem.", + "bbox": [ + 215, + 643, + 787, + 748 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "To constrain its solution space, prior works either aimed to define hand-crafted priors [9,18,31,75] or to learn them from the data [15-17,22,23,32,36,56,62,63,77]. However, the former often fall short when applied to real-world scenes, while the latter suffer from scarcity of the ground truth data. As a result, these algorithms are often heavily tailored to a specific domain, e.g. indoor [15-17,50,63] or outdoor scenes [22,23,56,62,77].", + "bbox": [ + 215, + 750, + 787, + 840 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/eb9557c42ae5d42da10f49db59147e2ba8d2ce6694f2dfbe520005d7073f2e0e.jpg", + "image_caption": [ + "Virtual Object Insertion", + "A vase with dried flowers" + ], + "image_footnote": [], + "bbox": [ + 217, + 162, + 390, + 252 + ], + "page_idx": 1 + }, + { + "type": "image", + "img_path": "images/036064d38f10febd97136079be0dcbefb40515995ac8bf67bcf6b7592671a527.jpg", + "image_caption": [ + "A black SUV car" + ], + "image_footnote": [], + "bbox": [ + 397, + 162, + 570, + 252 + ], + "page_idx": 1 + }, + { + "type": "image", + "img_path": "images/06be440aeacb2afda4aa9bd81a386182d56eea84dc38e5aae79daaad1061b530.jpg", + "image_caption": [ + "A metal bucket", + "Fig.1: We propose DiPIR, a physically based method to recover lighting from a single image, enabling arbitrary virtual object compositing into indoor and outdoor scenes, as well as material and tone-mapping optimization. Project page: https://research.nvidia.com/labs/toronto-ai/DiPIR/" + ], + "image_footnote": [], + "bbox": [ + 217, + 267, + 390, + 357 + ], + "page_idx": 1 + }, + { + "type": "image", + "img_path": "images/02d163547fd641c2d36da6ca0f8d4e05bdbd5c0d796c839ee7913b8a6cd206c8.jpg", + "image_caption": [ + "A white SUV car" + ], + "image_footnote": [], + "bbox": [ + 397, + 267, + 570, + 357 + ], + "page_idx": 1 + }, + { + "type": "image", + "img_path": "images/8efe3259dafa467b6a2b948151535c18b85b91cadd43cbfebe3ffdc3caf2a4e0.jpg", + "image_caption": [ + "Applications", + "Insert multiple objects" + ], + "image_footnote": [], + "bbox": [ + 602, + 162, + 774, + 252 + ], + "page_idx": 1 + }, + { + "type": "image", + "img_path": "images/036680c1483f31be9a9d096e0efa3ebb30ad26a8be0e2345049567543452e6ae.jpg", + "image_caption": [ + "Optimize material & tone mapping" + ], + "image_footnote": [], + "bbox": [ + 601, + 267, + 774, + 357 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "To address these challenges, we propose to reuse the strong image generation priors learned by large diffusion models (DMs) [46] as a guidance for inverse rendering. Unlike hand-crafted or supervised data-driven priors that are often specific to a domain, DMs are trained on massive datasets and show a remarkable \"understanding\" of the world and the underlying physical concepts. While DMs still often fail to produce accurate lighting effects such as shadows and reflections [49] in their generations, we observe that they can provide valuable guidance when combined with a physically-based renderer and are adapted to the scene.", + "bbox": [ + 212, + 446, + 787, + 566 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Specifically, we present Diffusion Prior for Inverse Rendering (DiPIR), building on three main contributions. First, we use a physically based renderer to accurately simulate the interaction between the light and the 3D asset to generate the final composited image. We also account for the unknown tone-mapping curve to mimic the camera sensor response. Second, we propose a lightweight personalization scheme of the pre-trained DM, based on the input image and the type of inserted asset. Third, we design a variant of the SDS loss [45] which makes use of this personalization and improves training stability.", + "bbox": [ + 212, + 566, + 787, + 686 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "In DiPIR, the DM acts similarly to a human evaluator. It takes the edited image as input and propagates the feedback signal to physically-based scene attributes via differentiable rendering, thus enabling end-to-end optimization. We experimentally show that DiPIR outperforms existing state-of-the-art lighting estimation methods for object insertion across indoor and outdoor datasets.", + "bbox": [ + 212, + 686, + 787, + 763 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "2 Related Work", + "text_level": 1, + "bbox": [ + 215, + 782, + 388, + 800 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Inverse rendering is the task of recovering intrinsic properties of a scene, including materials, shape, and lighting, from a single or multiple images [5].", + "bbox": [ + 212, + 809, + 787, + 840 + ], + "page_idx": 1 + }, + { + "type": "page_number", + "text": "2", + "bbox": [ + 217, + 114, + 228, + 126 + ], + "page_idx": 1 + }, + { + "type": "header", + "text": "R. Liang et al.", + "bbox": [ + 271, + 114, + 370, + 128 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "The main challenge of inverse rendering lies in the ill-posed nature of the task. Prior knowledge about materials and lighting effects is crucial to solving under-constrained inverse rendering problems, but how to best define and incorporate such priors remains an open research question.", + "bbox": [ + 212, + 146, + 787, + 205 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Early methods formulated inverse rendering as an optimization problem and used hand-crafted regularization terms such as local smoothness and sparsity of materials [4,5,9,18,31,75] to constrain the solution space. However, real-world lighting effects are often complex and difficult to describe using hand-crafted priors. The advent of deep learning has facilitated the learning of data-driven priors from ground truth supervision [6,8,30,34,35,38,50,62-64,69]. Yet, acquiring real-world data with accurately annotated intrinsic decomposition necessitates specialized capture devices [18] or expert annotators [6,30]. Collecting real-world datasets at a large scale is therefore extremely challenging. As a result, existing methods often resort to synthetic datasets [17,35,50,56,63,77], or are carefully designed to be training data-efficient [22,23,62]. The resulting algorithms are therefore often tailored to specific domains, such as indoor [15-17, 50, 63] or day-time outdoor scenes [22,23,56,62,77].", + "bbox": [ + 212, + 207, + 787, + 404 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Lighting estimation is a subtask of inverse rendering that specifically focuses on inferring lighting and is a core component of photorealistic virtual object insertion pipelines. Existing lighting estimation methods are often designed as feed-forward neural networks that take a single image as input and directly regress lighting information in form of spherical lobes [17,35,70,76], environment maps [16,50,53,56,77], parametric light sources [15], or volumetric lighting [37, 54,62,63]. The environment lighting can also be recovered through the proxy geometry shown in the single image [32,67]. Recently, lighting estimation was also formulated as a generative task. StyleLight [61] is based on a dual-branch StyleGAN model, which first performs GAN inversion through the LDR branch and then generates the panorama with the HDR branch. EverLight [12] instead regresses an initial HDR environment map from the input image and later refines it with a GAN model [27]. Concurrent to our work, DiffusionLight [44] proposed to inpaint a chrome ball in the center of the given image using an inpainting diffusion model and unwarp it to an environment map. To obtain the HDR environment map, DiffusionLight trains a LoRA [24] for exposure bracketing and merges multiple generated LDR chrome balls with different exposures.", + "bbox": [ + 212, + 409, + 787, + 667 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Note that lighting is fundamentally a High Dynamic Range (HDR) quantity, and reproducing this full range still remains challenging. HDR is crucial for basic effects such as sharp shadows, particularly in outdoor day-time settings where the sun is several orders of magnitudes brighter than the rest of the environment.", + "bbox": [ + 212, + 667, + 789, + 728 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Physically-based differentiable rendering. Rendering algorithms such as path tracing [26] aim to produce realistic images of virtual scenes by accurately simulating the physical processes involved in light transport. Physically-based rendering (PBR) typically uses a geometric representation of the scene, scattering functions modeling the behavior of the scene's surfaces, as well as camera and lighting models. Recent years have seen the development of many differentiable rendering methods and frameworks [25, 33, 42, 43, 60, 71, 72].", + "bbox": [ + 212, + 734, + 787, + 840 + ], + "page_idx": 2 + }, + { + "type": "header", + "text": "Photorealistic Object Insertion with Diffusion-Guided Inverse Rendering", + "bbox": [ + 250, + 114, + 730, + 130 + ], + "page_idx": 2 + }, + { + "type": "page_number", + "text": "3", + "bbox": [ + 774, + 116, + 784, + 126 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "While particular attention has been devoted to the difficult problem of computing derivatives at discontinuities [3, 33, 39, 65, 74], such as due to visibility changes, we focus here on the comparatively simpler lighting, material and tone-mapping derivatives, which generally do not introduce discontinuities.", + "bbox": [ + 212, + 146, + 787, + 207 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Diffusion model priors and personalization. Priors learned by the text-to-image DMs [2,11,46,48], on large-scale datasets can be adapted to various applications such as monocular depth estimation [28], authentic image completion [57], and image restoration [10]. Among various adaptation techniques, DreamBooth [47] proposes to finetune the DM on subject images, while Textual Inversion [14] optimizes new word embeddings for target generation. LoRA [24,51] simplifies adaptation by freezing the pretrained model weights and injecting trainable rank decomposition matrices, thus greatly reducing the number of trainable parameters. Most similar to ours are concurrent works on DM adaptation for lighting estimation [44] and intrinsic image decomposition [29]. The former adapts the DM to exposure bracketing for LDR to HDR lifting, while the latter fine-tunes a pretrained SD model to directly output the albedo and BRDF properties of an image.", + "bbox": [ + 212, + 213, + 789, + 411 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "3 Preliminaries", + "text_level": 1, + "bbox": [ + 215, + 431, + 382, + 449 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Diffusion Models. Diffusion models are a family of generative models built around two key processes. A forward process, which gradually adds noise to data samples $\\mathbf{x} \\sim p(\\mathbf{x})$ removing their structure over time $t$ . This is achieved using a noise schedule determined by $\\alpha_{t}$ and $\\sigma_{t}$ as $\\mathbf{x}_{t} = \\alpha_{t}\\mathbf{x} + \\sigma_{t}\\boldsymbol{\\epsilon}, \\boldsymbol{\\epsilon} \\sim \\mathcal{N}(\\mathbf{0},\\mathbf{I})$ . In contrast, the reverse process gradually removes this noise, restoring the structure. The reverse process is parameterized by a conditional neural network, $\\epsilon_{\\theta}$ , trained to predict the noise $\\epsilon$ at a given timestep $t$ according to the following simplified objective [20]:", + "bbox": [ + 212, + 463, + 787, + 585 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\n\\mathbb {E} _ {\\mathbf {x} \\sim p (\\mathbf {x}), \\boldsymbol {\\epsilon} \\sim \\mathcal {N} (\\mathbf {0}, \\boldsymbol {I}), t \\sim T} \\left[ w (t) | | \\boldsymbol {\\epsilon} _ {\\boldsymbol {\\theta}} (\\mathbf {x} _ {t}, t, \\boldsymbol {c}) - \\boldsymbol {\\epsilon} | | _ {2} ^ {2} \\right], \\tag {1}\n$$\n", + "text_format": "latex", + "bbox": [ + 334, + 594, + 787, + 613 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "where $c$ represents a condition (e.g. text, image, etc.) that allows controlling the generation process, $w(t)$ represents a time-conditional weighting, and $T$ is a set containing a selection of timesteps.", + "bbox": [ + 212, + 621, + 787, + 666 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "In this work, we use a Latent Diffusion Model (LDM) [46] in which the diffusion process is conducted in a lower-dimensional latent space. Specifically, the encoder $\\mathcal{E}$ maps samples from the data distribution $\\mathbf{x} \\sim p(\\mathbf{x})$ into a latent space $\\mathbb{Z}$ . The decoder $\\mathcal{D}$ performs the inverse operation, such that $\\mathcal{D}(\\mathcal{E}(\\mathbf{x})) \\approx \\mathbf{x}$ . As follows, for LDMs $\\mathbf{x}$ in Eq. 1 is replaced by its latent $\\mathbf{z} = \\mathcal{E}(\\mathbf{x})$ .", + "bbox": [ + 212, + 667, + 789, + 744 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Personalization and Fine-tuning. Fine-tuning all parameters of a pretrained DM requires significant computational resources and time. To alleviate this, LoRA [24] injects trainable low-rank decomposition matrices and aims to learn only the variations from the pretrained weights. Specifically, consider a linear layer represented as $\\pmb{h} = \\pmb{W}_0\\pmb{x}$ . Here $\\pmb{W}_0 \\in \\mathbb{R}^{n \\times n}$ and $\\pmb{x} \\in \\mathbb{R}^n$ are the pretrained", + "bbox": [ + 212, + 763, + 787, + 842 + ], + "page_idx": 3 + }, + { + "type": "page_number", + "text": "4", + "bbox": [ + 217, + 114, + 228, + 126 + ], + "page_idx": 3 + }, + { + "type": "header", + "text": "R. Liang et al.", + "bbox": [ + 271, + 114, + 370, + 128 + ], + "page_idx": 3 + }, + { + "type": "image", + "img_path": "images/e37dbec6ed4ca66a42ced518975652ead35e15ea9dfb8476f76d5260b2b2cb5b.jpg", + "image_caption": [ + "Fig. 2: Method overview. Given an input image, we first construct a virtual 3D scene with a virtual object and proxy plane. Our physically-based renderer then differentiably simulates the interactions of the estimizable environment map with the inserted virtual object and its effect on the background scene (shadowing) (left). At each iteration, the rendered image is diffused and passed through a personalized diffusion model (middle). The gradient of the adapted Score Distillation formulation is propagated back to the environment map and the tone-mapping curve through the differentiable renderer. Upon convergence, we recover lighting and tone-mapping parameters, which allow photorealistic compositing of virtual objects from a single image (right)." + ], + "image_footnote": [], + "bbox": [ + 215, + 143, + 658, + 318 + ], + "page_idx": 4 + }, + { + "type": "image", + "img_path": "images/9642069bf67e8f1ea1bde726c288ba8dee1cad210dbc647aee5ec017040eb035.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 669, + 143, + 779, + 318 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "weights and input, respectively. Applying LoRA, this layer is modified to $\\pmb{h} = (W_0 + \\pmb{\\Delta W})\\pmb{x} = W_0\\pmb{x} + AB\\pmb{x}$ . Notably, $\\pmb{\\Delta W} = \\pmb{A}\\pmb{B}$ is a combination of low-rank matrices $\\pmb{A} \\in \\mathbb{R}^{n \\times r}$ and $\\pmb{B} \\in \\mathbb{R}^{r \\times n}$ with $r \\ll n$ . During LoRA fine-tuning, only the added term $\\pmb{\\Delta W}$ is updated while the parameters $W_0$ remain unchanged. The rank $r$ is a hyperparameter that trades efficiency with model capacity.", + "bbox": [ + 212, + 465, + 787, + 542 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "4 Method", + "text_level": 1, + "bbox": [ + 215, + 564, + 330, + 579 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Given a single image as input, DiPIR recovers scene lighting and tone-mapping parameters, with the goal of photorealistic insertion of virtual objects. An overview of our method is shown in Fig. 2. Sec. 4.1 describes our representation and the differentiable rendering process, while Sec. 4.2 and Sec. 4.3 provide details on diffusion model guidance and optimization formulation, respectively.", + "bbox": [ + 212, + 594, + 787, + 672 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "4.1 Physically-based Virtual Object Insertion", + "text_level": 1, + "bbox": [ + 214, + 680, + 606, + 695 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Virtual scene. Inserting a virtual object $\\mathcal{X}$ into an image $\\mathbf{I}_{\\mathrm{bg}} \\in \\mathbb{R}^{h \\times w \\times 3}$ requires creating a 3D proxy virtual scene, viewed from the correct camera pose. Here, we assume that the user provides a specific placement (pose) for $\\mathcal{X}$ , but in some cases, an appropriate pose can also be determined automatically, e.g. by detecting the floor plane and scene scale.", + "bbox": [ + 212, + 700, + 787, + 775 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "To model the effects of the inserted object on the original image, such as shadows cast by $\\mathcal{X}$ , we also assume a known proxy geometry $\\mathcal{P}$ . We found a simple ground plane acting as a shadow catcher underneath the virtual object to be sufficient in all of our experiments. This proxy plane can be easily placed manually or automatically generated based on e.g. depth or LiDAR data.", + "bbox": [ + 212, + 776, + 787, + 852 + ], + "page_idx": 4 + }, + { + "type": "header", + "text": "Photorealistic Object Insertion with Diffusion-Guided Inverse Rendering", + "bbox": [ + 250, + 114, + 732, + 128 + ], + "page_idx": 4 + }, + { + "type": "page_number", + "text": "5", + "bbox": [ + 774, + 116, + 784, + 126 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Light representation. We represent the scene's lighting $\\mathbf{L}$ with a set of $N$ optizable Spherical Gaussian (SG) parameters $\\{\\mathbf{c}_k,\\pmb {\\mu}_k,\\sigma_k\\}_{k = 1}^N\\in \\mathbb{R}^{N\\times 7}$ , where the radiance for one SG lobe at the direction $\\pmb {v}\\in \\mathbb{R}^3$ is defined as", + "bbox": [ + 212, + 146, + 787, + 191 + ], + "page_idx": 5 + }, + { + "type": "equation", + "text": "\n$$\n\\mathbf {G} _ {k} (\\pmb {v}; \\mathbf {c} _ {k}, \\pmb {\\mu} _ {k}, \\sigma_ {k}) = \\mathbf {c} e ^ {- (1 - \\pmb {v} \\cdot \\pmb {\\mu}) / \\sigma^ {2}}, \\quad \\mathrm {w h e r e} \\mathbf {c} _ {k} \\in \\mathbb {R} ^ {3}, \\pmb {\\mu} _ {k} \\in \\mathbb {R} ^ {3}, \\sigma_ {k} \\in \\mathbb {R} _ {+}. \\quad (2)\n$$\n", + "text_format": "latex", + "bbox": [ + 243, + 196, + 787, + 217 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "The overall environment map $\\mathbf{L} \\in \\mathbb{R}^{H \\times W \\times 3}$ is computed as:", + "bbox": [ + 214, + 222, + 650, + 239 + ], + "page_idx": 5 + }, + { + "type": "equation", + "text": "\n$$\n\\mathbf {L} _ {i, j} = \\sum_ {k = 1} ^ {N} \\mathbf {G} _ {k} \\left(\\boldsymbol {v} _ {i, j}; \\mathbf {c} _ {k}, \\boldsymbol {\\mu} _ {k}, \\sigma_ {k}\\right), \\tag {3}\n$$\n", + "text_format": "latex", + "bbox": [ + 390, + 244, + 787, + 287 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "where $\\pmb{v}_{i,j}$ is the direction corresponding to the pixel $(i,j)$ using the standard spherical environment parameterization. Note that we chose a SG-based lighting representation for its good convergence properties and simplicity, but many alternatives exist and might be combined with our method.", + "bbox": [ + 212, + 292, + 787, + 354 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Differentiable rendering. Inserting a virtual object into a scene involves simulating the interactions of the estimizable environment map with the inserted virtual object (foreground), and the inserted object's effect on the background scene (shadow). We describe below the details respectively.", + "bbox": [ + 212, + 359, + 787, + 421 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Foreground image. Given the optimized light representation $\\mathbf{L}$ from above and the virtual object $\\mathcal{X}$ (including its geometry and materials), a foreground image $\\mathbf{I}_{\\mathrm{fg}}$ of the inserted object can be rendered directly using standard path tracing:", + "bbox": [ + 214, + 430, + 787, + 477 + ], + "page_idx": 5 + }, + { + "type": "equation", + "text": "\n$$\n\\mathbf {I} _ {\\mathrm {f g}} = \\operatorname {P a t h T r a c e} (\\mathcal {X}, \\mathbf {L}, D), \\tag {4}\n$$\n", + "text_format": "latex", + "bbox": [ + 406, + 484, + 787, + 501 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "$D$ is the maximum number of interactions along the light path. Since we do not assume that the provided proxy geometry is accurate nor that it has materials, we omit the effect of light reflected by the object into the scene, for example, due to a highly specular inserted object.", + "bbox": [ + 212, + 507, + 787, + 569 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Shadow ratio. As the background image $\\mathbf{I}_{\\mathrm{bg}}$ already faithfully represents the scene, our task is limited to simulating the appearance of the inserted object and its effect on the nearby parts of the scene, i.e. the shadows cast by the object onto the proxy geometry.", + "bbox": [ + 212, + 580, + 787, + 638 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "The shadow ratio $\\beta_{\\mathrm{shadow}} \\in \\mathbb{R}^{h \\times w \\times 3}$ , which accounts for the effect of the object on the surrounding scene, is computed as the ratio between the radiance received by the proxy geometry with and without the inserted object present. A low value of $\\beta_{\\mathrm{shadow}}$ indicates a strongly shadowed region:", + "bbox": [ + 212, + 638, + 787, + 700 + ], + "page_idx": 5 + }, + { + "type": "equation", + "text": "\n$$\n\\boldsymbol {\\beta} _ {\\mathrm {s h a d o w}} = \\frac {\\operatorname {P a t h T r a c e} (\\mathcal {X} \\cup \\mathcal {P} , \\mathbf {L} , 1)}{\\operatorname {P a t h T r a c e} (\\mathcal {P} , \\mathbf {L} , 1)}, \\tag {5}\n$$\n", + "text_format": "latex", + "bbox": [ + 375, + 707, + 787, + 739 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "This is similar to the method of Wang et al. [62], however we use Multiple Importance Sampling (MIS) [59] between lighting and BSDF for better sampling efficiency. The maximum path length has been limited to 1 interaction in order to reduce memory and computational cost. Unless specific material information is provided with the proxy geometry, we use a Lambertian BSDF, in which case the albedo cancels out in the ratio. In practice, the computation of the shadow ratio and foreground image are combined in order to reuse shared terms.", + "bbox": [ + 212, + 744, + 787, + 852 + ], + "page_idx": 5 + }, + { + "type": "page_number", + "text": "6", + "bbox": [ + 217, + 114, + 228, + 126 + ], + "page_idx": 5 + }, + { + "type": "header", + "text": "R. Liang et al.", + "bbox": [ + 271, + 114, + 369, + 128 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Tone-mapping. To compensate for the unknown tone-mapping of the input image, we introduces an estimable tone correction function $f(\\cdot)$ that is applied on the inserted object $\\hat{\\mathbf{I}}_{\\mathrm{fg}} = f(\\mathbf{I}_{\\mathrm{fg}}; \\boldsymbol{\\theta}_{\\mathrm{fg}})$ and shadows $\\hat{\\beta}_{\\mathrm{shadow}} = f(\\boldsymbol{\\beta}_{\\mathrm{shadow}}; \\boldsymbol{\\theta}_{\\mathrm{shadow}})$ . Our tone curve parameterization follows from [13] and employs a monotonic rational-quadratic spline as its basic building block. This spline is composed of $K_{\\mathrm{s}}$ bins, each defined by the quotient of two quadratic polynomials. Notably, these functions are differentiable, and allow direct parameterization of the derivatives and heights at each knot. We apply a single spline for the foreground image, and different correction splines for each RGB channel for the shadow ratio to provide the flexibility to adjust the color of the shadow. We set $K_{\\mathrm{s}} = 5$ for these curves.", + "bbox": [ + 212, + 146, + 787, + 299 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Differentiability. The final output image is an alpha-composite of the foreground object, shadows, and background image:", + "bbox": [ + 214, + 306, + 789, + 338 + ], + "page_idx": 6 + }, + { + "type": "equation", + "text": "\n$$\n\\mathbf {I} _ {\\mathrm {c o m p}} = \\left(1 - \\mathbf {V} (\\mathcal {X})\\right) \\cdot \\hat {\\boldsymbol {\\beta}} _ {\\mathrm {s h a d o w}} \\cdot \\mathbf {I} _ {\\mathrm {b g}} + \\mathbf {V} (\\mathcal {X}) \\cdot \\tilde {\\mathbf {I}} _ {\\mathrm {f g}}, \\tag {6}\n$$\n", + "text_format": "latex", + "bbox": [ + 331, + 347, + 787, + 366 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "where $\\mathbf{V}(\\mathcal{X})$ equals 1 when $\\mathcal{X}$ is directly visible from the camera, and 0 otherwise.", + "bbox": [ + 212, + 375, + 787, + 390 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "As neither foreground rendering, shadow rendering, or the compositing operation require derivatives w.r.t. the object placement or other discontinuous quantities, we can rely on automatic differentiation to obtain gradients of any pixelwise loss w.r.t. the lighting or material properties. To this end, we use the Path Replay Backpropagation [60] integrator of Mitsuba 3 [25]. The estimizable parameters are the Spherical Gaussian coefficients used in the lighting representation, and the tone curves' parameters $\\theta_{\\mathrm{fg}},\\theta_{\\mathrm{shadow}}$", + "bbox": [ + 212, + 391, + 787, + 496 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "While each of the operations above include simplifications (e.g. we do not account for secondary lighting from the proxy geometry), we have found them to be sufficient in practice: the remaining imperfections in the simulation can be sufficiently offset by the lighting optimization.", + "bbox": [ + 212, + 497, + 787, + 556 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "4.2 Diffusion Guidance", + "text_level": 1, + "bbox": [ + 214, + 578, + 419, + 593 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "The composited image produced by our differentiable rendering pipeline serves as the input to a DM that is used to compute a guidance signal, employing an optimization objective similar to Score Distillation Sampling (SDS) [45]. However, while DMs inherently have robust priors for lighting, we found that they do not provide out-of-the-box the necessary guidance for our specific needs. Consequently, we propose an adaptive score distillation loss specifically designed for object insertion tasks that exploits a personalization strategy which we detail in the following section.", + "bbox": [ + 212, + 603, + 787, + 724 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Personalization with concept preservation. Off-the-shelf DMs often do not provide robust guidance for virtual object insertion, especially in out-of-distribution scenes such as outdoor driving environments. A potential solution, inspired by [57], is to adapt the DM using an image from the target scene. In our experience, however, this approach often resulted in too much overfitting to the target scene's content, reducing the model's ability to adapt to the scene with a newly inserted object. This led to artifacts and an unstable optimization process.", + "bbox": [ + 212, + 731, + 787, + 839 + ], + "page_idx": 6 + }, + { + "type": "header", + "text": "Photorealistic Object Insertion with Diffusion-Guided Inverse Rendering", + "bbox": [ + 248, + 114, + 732, + 130 + ], + "page_idx": 6 + }, + { + "type": "page_number", + "text": "7", + "bbox": [ + 774, + 116, + 784, + 126 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "To mitigate this issue, we propose fine-tuning the DM with a focus on preserving the identity of the objects to be inserted (Fig. 3). We specifically achieved this by generating additional synthetic images for the insertable class concept (e.g. car). We sample those images from the off-the-self DM, starting with a base prompt such as \"a photo of a car\" and appending attributes such as color, background, lighting, and size to ensure diversity in the generated data.", + "bbox": [ + 212, + 146, + 488, + 340 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/95c3c69265f135eeb2c9b3cf21fbbb3ce87fd46adddf951bd9781515088ca11a.jpg", + "image_caption": [ + "Input image", + "\"ascene in the style of sks rendering" + ], + "image_footnote": [], + "bbox": [ + 500, + 155, + 602, + 215 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/27d5ec4434c5f4a6e8cfd23876b353c58a69d45a6f4e5ebdf9f6035594d2d943.jpg", + "image_caption": [ + "Generated images", + "\"aphotofacar\"", + "Fig. 3: Personalization with concept preservation." + ], + "image_footnote": [], + "bbox": [ + 500, + 227, + 602, + 304 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/fd243130eba1d645b5aeafdfe5b4fb49c059f81d2820fbef31de7e85cff82d16.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 611, + 183, + 761, + 263 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "We employ LoRA with rank 4 to fine-tune the DM, combining the in-domain target example with this supplemental data. Our training follows the objective described in Eq. 1, where $c$ corresponds to two predefined prompts: \"a scene in the style of sks rendering\" for the target image, and \"a photo of a {concept class}\". We sample approximately 30-40 supplementary images for indoor scenes and 200 for outdoor scenes. The total time spent for fine-tuning is typically less than 15 minutes on one high-end GPU with mixed precision training.", + "bbox": [ + 212, + 342, + 787, + 448 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Score Distillation with adapted guidance. Score Distillation Sampling [45] leverages a pretrained DM to guide the optimization of a differentiable, parametric image-rendering function $g_{\\phi} \\coloneqq \\pmb{x}$ . Specifically, the parameters $\\phi$ in our scenario corresponds to parameters of the Spherical Gaussian lighting and tone-mapping curves, are updated using the gradient:", + "bbox": [ + 212, + 453, + 787, + 529 + ], + "page_idx": 7 + }, + { + "type": "equation", + "text": "\n$$\n\\nabla_ {\\phi} \\mathcal {L} _ {\\mathrm {S D S}} (\\boldsymbol {\\phi}, \\boldsymbol {\\theta}) := \\mathbb {E} _ {\\boldsymbol {\\epsilon} \\sim \\mathcal {N} (\\mathbf {0}, \\boldsymbol {I}), t \\sim T} \\left[ w (t) \\left(\\hat {\\epsilon} _ {\\boldsymbol {\\theta}} (\\boldsymbol {z} _ {t}, t, \\boldsymbol {c}) - \\boldsymbol {\\epsilon}\\right) \\frac {\\partial \\boldsymbol {z} _ {t}}{\\partial \\phi} \\right], \\qquad (7)\n$$\n", + "text_format": "latex", + "bbox": [ + 277, + 535, + 787, + 569 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "where $\\pmb{z} = \\mathcal{E}(g_{\\phi})$ and $\\hat{\\epsilon}_{\\theta}(\\pmb{z}_t, t, \\pmb{c}) \\coloneqq (1 + s)\\pmb{\\epsilon}_{\\theta}(\\pmb{z}_t, t, \\pmb{c}) - s\\pmb{\\epsilon}_{\\theta}(\\pmb{z}_t, t, \\emptyset)$ . $\\hat{\\epsilon}$ denotes the classifier free guidance (CFG) version [21] of $\\pmb{\\epsilon}_{\\theta}$ used in text-conditioned DMs to enable higher quality generation via a guidance scale parameter $s$ .", + "bbox": [ + 212, + 575, + 787, + 619 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Unfortunately, we encountered training instabilities when applying the original formulation of the SDS loss to our problem. Instead, we adopt the following alternative that integrates the LoRA personalization, and we call it LDS loss:", + "bbox": [ + 212, + 621, + 787, + 666 + ], + "page_idx": 7 + }, + { + "type": "equation", + "text": "\n$$\n\\nabla_ {\\phi} \\mathcal {L} _ {\\mathrm {L D S}} (\\phi , \\boldsymbol {\\theta}) := \\mathbb {E} _ {\\epsilon \\sim \\mathcal {N} (\\boldsymbol {0}, \\boldsymbol {I}), t \\sim T} \\left[ w (t) \\left(\\epsilon_ {(\\boldsymbol {\\theta} + \\boldsymbol {\\Delta W})} (\\boldsymbol {z} _ {t}, t, \\boldsymbol {c}) - \\epsilon_ {\\boldsymbol {\\theta}} (\\boldsymbol {z} _ {t}, t, \\varnothing)\\right) \\frac {\\partial \\boldsymbol {z} _ {t}}{\\partial \\phi} \\right], \\tag {8}\n$$\n", + "text_format": "latex", + "bbox": [ + 217, + 672, + 785, + 718 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "where $\\epsilon_{(\\theta +\\Delta W)}$ represents the predicted noise of the LoRA personalized model described in the previous section. Notably, this loss function bears resemblance to the Classifier Score Distillation (CSD) [68]. However, it's important to note that in our case, the delta is calculated between the adapted and non-adapted versions of the diffusion model. Intuitively, this delta guides the optimization process in a direction determined by the personalized model which preserves the concept of the inserted object while also capturing the appearance and semantics of the specific scene.", + "bbox": [ + 212, + 719, + 787, + 839 + ], + "page_idx": 7 + }, + { + "type": "page_number", + "text": "8", + "bbox": [ + 217, + 114, + 228, + 126 + ], + "page_idx": 7 + }, + { + "type": "header", + "text": "R. Liang et al.", + "bbox": [ + 271, + 114, + 370, + 128 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "4.3 Optimization Formulation", + "text_level": 1, + "bbox": [ + 215, + 146, + 477, + 162 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "After completing the diffusion model's personalization, we optimize the lighting and tone-mapping parameters using the following loss function:", + "bbox": [ + 214, + 172, + 784, + 202 + ], + "page_idx": 8 + }, + { + "type": "equation", + "text": "\n$$\n\\mathcal {L} = \\mathcal {L} _ {\\mathrm {L D S}} + \\lambda_ {\\text {c o n s i s t e n c y}} \\mathcal {L} _ {\\text {c o n s i s t e n c y}} + \\lambda_ {\\text {r e g}} \\mathcal {L} _ {\\text {r e g}}. \\tag {9}\n$$\n", + "text_format": "latex", + "bbox": [ + 339, + 215, + 785, + 232 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "The loss' main component is the diffusion guidance $\\mathcal{L}_{\\mathrm{LDS}}$ , which provides a perceptual realism objective on the edited result with object insertion. We use the text prompt \"a photo of a {concept class} in a scene in the style of sks rendering\", which is a natural combination of the two text prompts used in personalization (Fig. 3). Here {concept class}, e.g. car, provides the context of the object to insert, while the personalized token sks provides the lighting style of the input image. The diffusion guidance $\\mathcal{L}_{\\mathrm{LDS}}$ is applied on the edited image $\\mathbf{I}_{\\mathrm{comp}}$ (Eq. 6), and backpropagated to the estimizable parameters through the differentiable rendering process (Eq. 8).", + "bbox": [ + 214, + 243, + 787, + 378 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "The two regularizers $\\mathcal{L}_{\\mathrm{consistency}}$ and $\\mathcal{L}_{\\mathrm{reg}}$ are related to environment map fusion, which we detail below.", + "bbox": [ + 214, + 380, + 785, + 409 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "Environment map initialization and fusion. The personalized DM provides guidance in two ways: (i) encouraging lighting consistency of the foreground object and the scene, such as the reflections and scale of the inserted object; and (ii) encouraging accurate shadows cast by the inserted object onto the background scene, such as the scale, direction, and color of the shadow. However, we empirically observe that these two signals can conflict in the early phase of the optimization.", + "bbox": [ + 214, + 417, + 787, + 523 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "To address this, we initialize two separate environment maps, $\\mathbf{L}^{\\mathrm{fg}},\\mathbf{L}^{\\mathrm{shadow}}\\in$ $\\mathbb{R}^{H\\times W\\times 3}$ , to light the foreground inserted object (Eq. 4) and cast shadows (Eq. 5) respectively. As the optimization progresses, the two environment maps are progressively fused into a single environment map $\\mathbf{L}^{\\mathrm{fused}}$ through scaling the the $\\mathbf{L}^{\\mathrm{fg}}$ by the relative luminance of between $\\mathbf{L}^{\\mathrm{fg}}$ and $\\mathbf{L}^{\\mathrm{shadow}}$ . We also use the following two regularization terms for this fusion process. First, we encourage the consistency between the normalized luminance $\\tilde{\\mathbf{L}}^{\\mathrm{fg}},\\tilde{\\mathbf{L}}^{\\mathrm{shadow}}\\in \\mathbb{R}^{H\\times W}$ of the environment maps by minimizing:", + "bbox": [ + 214, + 523, + 787, + 645 + ], + "page_idx": 8 + }, + { + "type": "equation", + "text": "\n$$\n\\mathcal {L} _ {\\text {c o n s i s t e n c y}} = - \\sum_ {i, j} \\tilde {\\mathbf {L}} _ {i, j} ^ {\\text {s h a d o w}} \\log \\left(\\tilde {\\mathbf {L}} _ {i, j} ^ {\\mathrm {f g}}\\right) \\Delta \\Omega_ {i, j} \\tag {10}\n$$\n", + "text_format": "latex", + "bbox": [ + 346, + 657, + 785, + 691 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "where $\\varDelta\\varOmega_{i,j}$ is the corresponding solid angle at pixel $(i,j)$ and the gradient for $\\tilde{\\mathbf{L}}^{\\mathrm{shadow}}$ is detached. Second, as the shadow environment map $\\mathbf{L}^{\\mathrm{shadow}}$ is supervised mainly via the shadow ratio $\\beta_{\\mathrm{shadow}}\\in \\mathbb{R}^{H\\times W\\times 3}$ of Eq. 5, to encourage concentrated high peaks for the sharp shadows and suppress the ambient light in $\\mathbf{L}^{\\mathrm{shadow}}$ , we also add a L2 regularizer in log-space with the Cauchy loss [7]:", + "bbox": [ + 214, + 702, + 787, + 780 + ], + "page_idx": 8 + }, + { + "type": "equation", + "text": "\n$$\n\\mathcal {L} _ {\\text {r e g}} = \\sum_ {i, j, c} \\log \\left(1 + 2 \\left(\\mathbf {L} _ {i, j, c} ^ {\\text {s h a d o w}}\\right) ^ {2}\\right) \\Delta \\Omega_ {i, j}. \\tag {11}\n$$\n", + "text_format": "latex", + "bbox": [ + 357, + 791, + 785, + 825 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "Please refer to the Supplement for more details.", + "bbox": [ + 215, + 835, + 563, + 851 + ], + "page_idx": 8 + }, + { + "type": "header", + "text": "Photorealistic Object Insertion with Diffusion-Guided Inverse Rendering", + "bbox": [ + 248, + 114, + 732, + 130 + ], + "page_idx": 8 + }, + { + "type": "page_number", + "text": "9", + "bbox": [ + 774, + 116, + 784, + 126 + ], + "page_idx": 8 + }, + { + "type": "table", + "img_path": "images/b1e4052f342c259114472b8d4e1ff65df3464ebd1f9364e7460a0a15923f7e51.jpg", + "table_caption": [ + "Table 1: Quantitative user study on outdoor street scenes. For each scene, users are shown two results—one produced by our method, and another produced by one of the baselines—and select which is more realistic. We report the results averaged across 3 user studies with 9 users each. Our method outperforms all baselines ( $>50\\%$ ) and is preferred in almost all illumination conditions." + ], + "table_footnote": [], + "table_body": "
DaytimeTwilightNightAll scenes
SunnyCloudy
Hold-Geoffroy et al. [22]60.8%66.7%74.1%85.7%68.8%
NLFE [62]80.4%73.3%44.4%52.4%67.4%
StyleLight [61]76.5%91.1%66.7%66.7%77.8%
DiffusionLight [44]80.4%68.9%55.6%71.4%70.8%
", + "bbox": [ + 297, + 215, + 702, + 290 + ], + "page_idx": 9 + }, + { + "type": "table", + "img_path": "images/6a9e834e67d11ecdbefeddebd625a376c2dc748d935d5e40c4453ae48ffad717.jpg", + "table_caption": [ + "Table 2: Quantitative evaluation on PolyHaven scenes. We report user study preference scores similar to Table 1. Metrics are computed w.r.t. to a “reference” image where the virtual object is lit by the ground-truth environment map." + ], + "table_footnote": [], + "table_body": "
MethodsOurs preferred ↓RMSE ↓SSIM ↑LPIPS [73] ↓si-RMSE [18] ↓
Wang et al. [63]84.8%0.0630.9850.01750.037
StyleLight [61]75.8%0.0560.9860.02020.039
DiffusionLight [44]66.7%0.0620.9850.01620.034
Ours/0.0480.9890.01470.027
", + "bbox": [ + 297, + 339, + 702, + 398 + ], + "page_idx": 9 + }, + { + "type": "text", + "text": "Training details. The personalized DM receives a crop of $\\mathbf{I}_{\\mathrm{comp}}$ (Eq. 6) with added noise. Since the inserted object likely does not cover the full background image, we use the visibility mask of the virtual object $V(\\mathcal{X})$ to locate the 2D bounding box of the inserted object, and randomly crop around the inserted object. The size of the cropped image is also randomized, with smaller crops covering local details of the inserted object and larger crops providing more visual cues of the lighting effects within the scene. Each scene is optimized for 600 iterations, and the maximum strength of the diffusion guidance is linearly decreased from 0.6 to 0.3. We use $\\lambda_{\\mathrm{consistency}} = 0.03$ and $\\lambda_{\\mathrm{reg}} = 0.01$ .", + "bbox": [ + 212, + 409, + 787, + 546 + ], + "page_idx": 9 + }, + { + "type": "text", + "text": "5 Experiments", + "text_level": 1, + "bbox": [ + 214, + 561, + 375, + 580 + ], + "page_idx": 9 + }, + { + "type": "text", + "text": "In this section, we evaluate DiPIR on a collection of outdoor and indoor scenes, covering diverse lighting conditions and different application domains. We first outline the experiment settings in Sec. 5.1 and then provide evaluation results on benchmark datasets in Sec. 5.2. Finally, we conduct an ablation study in Sec. 5.3.", + "bbox": [ + 212, + 584, + 787, + 645 + ], + "page_idx": 9 + }, + { + "type": "text", + "text": "5.1 Experiment Settings", + "text_level": 1, + "bbox": [ + 214, + 657, + 433, + 672 + ], + "page_idx": 9 + }, + { + "type": "text", + "text": "Waymo dataset. Following the data split from [66], we use 48 scenes with diverse illumination conditions from the Waymo Open Dataset [55] including 32 daytime (17 sunny and 15 cloudy), 9 twilight, and 7 night scenes. Each scene consists of an input image, a 3D car asset randomly selected from an asset bank, the ground plane, and a location to insert the car. Inspired by applications in synthetic data generation, this process is fully automatic: we use the LiDAR point cloud and semantic segmentation [58] to fit a ground plane and detect the empty space to insert a car.", + "bbox": [ + 212, + 679, + 787, + 800 + ], + "page_idx": 9 + }, + { + "type": "text", + "text": "PolyHaven dataset. We use 11 HDR environment maps from PolyHaven [1] and manually place a known ground plane and virtual object in each scene. Each", + "bbox": [ + 214, + 809, + 785, + 839 + ], + "page_idx": 9 + }, + { + "type": "page_number", + "text": "10", + "bbox": [ + 217, + 114, + 235, + 126 + ], + "page_idx": 9 + }, + { + "type": "header", + "text": "R. Liang et al.", + "bbox": [ + 271, + 114, + 370, + 128 + ], + "page_idx": 9 + }, + { + "type": "image", + "img_path": "images/8524d671d57967f33486f8f65292c31c0d91fa4498a0a052d1e1d0975d57f6e0.jpg", + "image_caption": [ + "Fig. 4: Comparisons on inserting objects into cropped HDRIs from PolyHaven." + ], + "image_footnote": [], + "bbox": [ + 218, + 143, + 784, + 296 + ], + "page_idx": 10 + }, + { + "type": "image", + "img_path": "images/4fc027309d44855e0aa1432d1ac2501f11a0c6f8ff7a4a98afa307ed39ac0e1e.jpg", + "image_caption": [ + "Fig. 5: Inserting car assets into Waymo driving scenes. Note the direction and sharpness of shadows, as well as overall brightness, color and specular highlights on inserted cars." + ], + "image_footnote": [], + "bbox": [ + 218, + 314, + 784, + 580 + ], + "page_idx": 10 + }, + { + "type": "text", + "text": "scenario therefore consists of an LDR background image $\\mathbf{I}_{\\mathrm{bg}}$ rendered directly from the environment map, the posed virtual object, and the proxy plane. A pseudo-ground-truth rendering is created by lighting the inserted object with the environment map itself.", + "bbox": [ + 212, + 618, + 787, + 679 + ], + "page_idx": 10 + }, + { + "type": "text", + "text": "User study. To evaluate the perceptual realism of the virtual object insertion, we conducted a user study where participants received a pair of two object insertion results in a random order - one generated using our proposed method, the other using a baseline approach. The participants were then instructed to compare the differences between the two results, inspect the lighting effects of the inserted object, and choose the image they perceived as more realistic. We invited 9 users to perform a binary selection for each image pair and used majority voting to determine the preferred image for each comparison. In the following, we report the percentage of times that our method was selected over the baseline. This process was repeated three times and a preferred percentage $>50\\%$ indicates", + "bbox": [ + 212, + 689, + 787, + 840 + ], + "page_idx": 10 + }, + { + "type": "header", + "text": "Photorealistic Object Insertion with Diffusion-Guided Inverse Rendering", + "bbox": [ + 250, + 114, + 732, + 128 + ], + "page_idx": 10 + }, + { + "type": "page_number", + "text": "11", + "bbox": [ + 767, + 114, + 782, + 126 + ], + "page_idx": 10 + }, + { + "type": "text", + "text": "Ours outperforming baselines. We include more details about the user study and statistical evaluation of its results in the Supplement.", + "bbox": [ + 212, + 146, + 782, + 176 + ], + "page_idx": 11 + }, + { + "type": "text", + "text": "Baselines. On the Waymo dataset of urban street scenes, we compare our method to the representative outdoor lighting estimation methods of Hold-Geoffroy et al. [22] and NLFE [62]; as well as the generative lighting estimation methods StyleLight [61] and DiffusionLight [44]. On the PolyHaven dataset, we compare with Wang et al. [63], StyleLight [61], DiffusionLight [44], and pseudo-ground-truth rendering.", + "bbox": [ + 212, + 184, + 787, + 275 + ], + "page_idx": 11 + }, + { + "type": "text", + "text": "5.2 Evaluation on Benchmark Datasets", + "text_level": 1, + "bbox": [ + 214, + 296, + 552, + 311 + ], + "page_idx": 11 + }, + { + "type": "text", + "text": "Urban street scenes. In Table 1, we provide quantitative user study results on the Waymo dataset [55], separated into 4 subsets based on illumination conditions, as well as on all scenes. Compared to all baselines, our method outperforms prior state-of-the-art methods (chosen $>50\\%$ ), and is preferred in almost every subset.", + "bbox": [ + 212, + 316, + 785, + 378 + ], + "page_idx": 11 + }, + { + "type": "text", + "text": "StyleLight [61] is primarily trained on indoor HDR panoramas and suffers from domain gap to outdoor scenes. Similarly, the sky model predicted by Hold-Geoffroy et al. [22] is tailored to outdoor daytime scenes and suffers severely from out-of-domain twilight and night scenes. NLFE [62] unprojects the surrounding scene geometry and predicts volumetric lighting, which makes it generalize better to night scenes, but it can fail to estimate an accurate scale and color of the peak for daytime scenes. DiffusionLight [44], a concurrent work, shows impressive results by inpainting a chrome sphere and predicting environment maps with realistic high-frequency details. However, it often struggles to predict a high-intensity peak in sunny daytime scenes or a proper scale for night scenes.", + "bbox": [ + 212, + 378, + 787, + 527 + ], + "page_idx": 11 + }, + { + "type": "text", + "text": "A qualitative comparison is shown in Fig. 5. Our method achieves high-quality object insertion, making it a promising approach for applications such as simulating synthetic data, augmented reality navigation and urban planning.", + "bbox": [ + 212, + 529, + 785, + 575 + ], + "page_idx": 11 + }, + { + "type": "text", + "text": "PolyHaven dataset. We show the quantitative comparison in Table 2 and qualitative results in Fig. 4. Our method is preferred over all baseline methods in the user study, and outperforms baselines on quantitative metrics. Our method produces virtual object insertion that naturally and consistently blend in the scenes, making it promising for virtual production tasks.", + "bbox": [ + 212, + 582, + 785, + 657 + ], + "page_idx": 11 + }, + { + "type": "text", + "text": "5.3 Ablation Study", + "text_level": 1, + "bbox": [ + 214, + 679, + 388, + 695 + ], + "page_idx": 11 + }, + { + "type": "text", + "text": "We extensively ablate our proposed design choices on diffusion guidance and scene representations, and provide the quantitative ablation in Table 3 with user study and qualitative results in Fig. 6. For the diffusion guidance, we compare with simplified or alternative versions: (i) Ours (SDS) removes the adaptive guidance (Eq. 8) and uses the original SDS [45] formulation (Eq. 7); (ii) Ours (SDS w/o LoRA) uses the original SDS loss with non-adapted DM; (iii) Ours (w/o concept preservation) uses our LDS loss with adapted DM but not using any class images for concept preservation. We additionally compare with a baseline of \"dataset update\" which applies a photometric L1 loss on DM repeatedly edited results", + "bbox": [ + 212, + 704, + 787, + 840 + ], + "page_idx": 11 + }, + { + "type": "page_number", + "text": "12", + "bbox": [ + 217, + 114, + 235, + 126 + ], + "page_idx": 11 + }, + { + "type": "header", + "text": "R. Liang et al.", + "bbox": [ + 271, + 114, + 370, + 128 + ], + "page_idx": 11 + }, + { + "type": "table", + "img_path": "images/0eef9cb0d0294e21241cc91251acf1926f02c47427bd6b77274a1b4efa915e08.jpg", + "table_caption": [], + "table_footnote": [], + "table_body": "
MethodsOurs preferred ↓
Ours (dataset update)85.2%
Ours (SDS [45])74.1%
Ours (SDS [45] w/o LoRA)90.7%
Ours (w/o concept preservation)64.8%
Ours (w/o tone curve)68.5%
Ours (w/o env. map fusion)66.7%
", + "bbox": [ + 222, + 143, + 485, + 224 + ], + "page_idx": 12 + }, + { + "type": "text", + "text": "Table 3: Ablation study on outdoor driving scenes [55]. We report the percentage of images that users preferred DiPIR compared to its ablated versions. Our full pipeline produces results that are preferred more often over its ablated versions.", + "bbox": [ + 504, + 143, + 794, + 226 + ], + "page_idx": 12 + }, + { + "type": "image", + "img_path": "images/3e40de3e5434336d40fdd305d84be29c140bc97fddcaf84f801bf9c1598822f7.jpg", + "image_caption": [ + "Ours (SDS w/o LoRA)" + ], + "image_footnote": [], + "bbox": [ + 217, + 234, + 359, + 308 + ], + "page_idx": 12 + }, + { + "type": "image", + "img_path": "images/58d3b97546e9e7ad48a0c84a29763969d058d13933177e10619513613ccec2be.jpg", + "image_caption": [ + "Ours (SDS w/ LoRA)" + ], + "image_footnote": [], + "bbox": [ + 361, + 234, + 500, + 308 + ], + "page_idx": 12 + }, + { + "type": "image", + "img_path": "images/2dac3abc527de255f449c05c8f17e09ee34fad2cc05b5b665f2de952db3469ce.jpg", + "image_caption": [ + "Ours (dataset update)" + ], + "image_footnote": [], + "bbox": [ + 501, + 234, + 640, + 308 + ], + "page_idx": 12 + }, + { + "type": "image", + "img_path": "images/9b12df193fc4464c23fec94403d8ff51f70b51a46ed16e68459ab780aa6943b3.jpg", + "image_caption": [ + "Ours" + ], + "image_footnote": [], + "bbox": [ + 643, + 277, + 784, + 349 + ], + "page_idx": 12 + }, + { + "type": "image", + "img_path": "images/6e016a3a6e3a24546148c073f8c34287a03d2c0cac1fd1932cda4f1d3863970e.jpg", + "image_caption": [ + "Ours (w/o concept preservation)", + "Ours (w/o tone-mapping curve)", + "Fig. 6: Qualitative ablation study of our design choices." + ], + "image_footnote": [], + "bbox": [ + 218, + 321, + 359, + 393 + ], + "page_idx": 12 + }, + { + "type": "image", + "img_path": "images/2eb94103d8cdc037afb39b45888493f7a5fd1a85dec865c18848b8418481a0df.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 361, + 321, + 500, + 393 + ], + "page_idx": 12 + }, + { + "type": "image", + "img_path": "images/ba5e321021181cd0b723f72dd9d7ef58fa2551bffc890b699c73d9c12a731631.jpg", + "image_caption": [ + "Ours (w/o env. fusion)" + ], + "image_footnote": [], + "bbox": [ + 501, + 321, + 640, + 393 + ], + "page_idx": 12 + }, + { + "type": "text", + "text": "curve)", + "bbox": [ + 410, + 405, + 449, + 415 + ], + "page_idx": 12 + }, + { + "type": "text", + "text": "via SDEdit [41], denoted as Ours (dataset update). Although this approach is adopted in text-guided scene editing [19], we found it challenging to produce stable gradients from such a \"discrete\" guidance.", + "bbox": [ + 212, + 449, + 784, + 494 + ], + "page_idx": 12 + }, + { + "type": "text", + "text": "When designing the light and tone representations, we observe that the trainable tone-mapping curve provides additional flexibility to compensate for unknown tone-mapping in the input image, and better match the color and scale of the shadows. The two environment map fusion scheme (Sec. 4.3) allows to robustly recover the high intensity peaks in the early phase of training, and converge to better quality. Their respective effect is ablated as \"Ours (w/o tone curve)\" and \"Ours (w/o env. map fusion)\".", + "bbox": [ + 212, + 494, + 784, + 601 + ], + "page_idx": 12 + }, + { + "type": "text", + "text": "6 Applications", + "text_level": 1, + "bbox": [ + 215, + 619, + 375, + 638 + ], + "page_idx": 12 + }, + { + "type": "text", + "text": "Since our method recovers physically based lighting information, arbitrary new virtual objects can be inserted after the optimization, as shown in Fig. 2. DiPIR can also optimize other scene attributes such as materials and local lighting. We perform preliminary experiments in this direction.", + "bbox": [ + 212, + 646, + 784, + 707 + ], + "page_idx": 12 + }, + { + "type": "text", + "text": "Material optimization. When combined with differentiable rendering, DMs can provide a guidance signal for material attributes, as shown in Fig. 7. Given a purely diffuse car and enabling Metallic and Roughness properties as estimizable parameters, the diffusion guidance can optimize and make the car look more shiny. By additionally changing the text prompt to \"a carmine red car\" and making the base color of the car an estimizable parameter, we show that the DM can propagate the text-condition to the PBR attribute and change the color of the car to red. When enabling local emission as an estimizable parameter, the diffusion model can also turn on the headlights of cars in night scenes.", + "bbox": [ + 212, + 715, + 785, + 852 + ], + "page_idx": 12 + }, + { + "type": "header", + "text": "Photorealistic Object Insertion with Diffusion-Guided Inverse Rendering", + "bbox": [ + 250, + 114, + 732, + 128 + ], + "page_idx": 12 + }, + { + "type": "page_number", + "text": "13", + "bbox": [ + 767, + 114, + 785, + 126 + ], + "page_idx": 12 + }, + { + "type": "image", + "img_path": "images/6364e243daa135c0da173a6e797c14e80dd7297c20f39c9539a6eb0c7a83f193.jpg", + "image_caption": [ + "Optimize tone mapping curve" + ], + "image_footnote": [], + "bbox": [ + 217, + 143, + 361, + 231 + ], + "page_idx": 13 + }, + { + "type": "image", + "img_path": "images/8cc9eb2aea1140c8dce37960fd30b0fe17d36d0415922ef0caea87fd0dfbeaec.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 362, + 145, + 506, + 231 + ], + "page_idx": 13 + }, + { + "type": "image", + "img_path": "images/146dd37cccebf6de47098121eec0b7a18fe18c063977f0dbbddcb99a6903f994.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 509, + 145, + 651, + 231 + ], + "page_idx": 13 + }, + { + "type": "image", + "img_path": "images/658200a4dde3bbe76bae075467cf465d254d93baa8d9c771f33dad7d68a1f73a.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 663, + 143, + 785, + 231 + ], + "page_idx": 13 + }, + { + "type": "image", + "img_path": "images/e50a2cf49f962e438fc1009327adfc9955637f39d2402732e9f5c3df9844526f.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 710, + 229, + 741, + 252 + ], + "page_idx": 13 + }, + { + "type": "image", + "img_path": "images/296bd4293843b9a9820b90be5587f777cf6f1238b5beac1d54d0f2d8732bc6de.jpg", + "image_caption": [ + "Optimize asset materials" + ], + "image_footnote": [], + "bbox": [ + 217, + 252, + 361, + 338 + ], + "page_idx": 13 + }, + { + "type": "image", + "img_path": "images/ee8f975fae2507cf9632e18414a4a9af4b4ab3c75d0da895092b5537498b2f38.jpg", + "image_caption": [ + "Fig. 7: Our physically based inverse rendering pipeline unlocks further applications such as material, local emission and tone-mapping refinement." + ], + "image_footnote": [], + "bbox": [ + 362, + 252, + 506, + 338 + ], + "page_idx": 13 + }, + { + "type": "image", + "img_path": "images/fbd79d06af1fd28af9eb4b9cd3d7c8b6c2c1891c2d46384ee89a4cbb615e7eea.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 509, + 252, + 651, + 338 + ], + "page_idx": 13 + }, + { + "type": "image", + "img_path": "images/5cb04fccb262f75c9525d9e04105a9963b5a95bb91f028caf2349ee806891af2.jpg", + "image_caption": [ + "Optimize emissions" + ], + "image_footnote": [], + "bbox": [ + 663, + 252, + 785, + 338 + ], + "page_idx": 13 + }, + { + "type": "text", + "text": "Tone-mapping adjustment. We use a controlled experiment to further evaluate how well DMs understand tone-mapping. In Fig. 7, we freeze the estimated environment map and apply a manual tone adjustment on the background image. The diffusion guidance is used to optimize the tone curve such that the inserted object matches the surrounding background in the final composited result.", + "bbox": [ + 212, + 404, + 789, + 481 + ], + "page_idx": 13 + }, + { + "type": "text", + "text": "7 Discussion", + "text_level": 1, + "bbox": [ + 215, + 491, + 354, + 507 + ], + "page_idx": 13 + }, + { + "type": "text", + "text": "Our method leverages large diffusion models' inherent scene understanding capabilities as guidance to a physically based inverse rendering pipeline. We design a diffusion guidance signal with scene-specific personalization and a differentiable inverse rendering pipeline to recover lighting and tone-mapping parameters. Our method enables inserting virtual objects into scenes, but also optimizing other scene parameters such as the materials of the inserted object or account for tone-mapping mismatches between cameras. We believe that this combination of the differentiable rendering process and data-driven priors can be used successfully in many other content creation applications such as relighting and animation.", + "bbox": [ + 212, + 515, + 787, + 667 + ], + "page_idx": 13 + }, + { + "type": "text", + "text": "Limitations and future work. Our Spherical Gaussians-based lighting representation is adequate for general objects [35], but might not behave realistically for highly specular materials. For more complex lighting representations adding generative priors on the environment map [40] is a direction worth exploring. The rendering formulation could be extended to account for effects such as reflections from the scene itself onto the inserted object (e.g. color bleeding), but that might introduce more ambiguities and require knowing the materials of the proxy geometry (refer to Supplement C.4 for failure case examples). Finally, while DM personalization significantly improves the quality of the results, it adds overhead and complexity to the pipeline. Recent personalization methods that do not require test-time finetuning [52] could be used to mitigate this overhead.", + "bbox": [ + 212, + 672, + 789, + 840 + ], + "page_idx": 13 + }, + { + "type": "page_number", + "text": "14", + "bbox": [ + 217, + 114, + 235, + 126 + ], + "page_idx": 13 + }, + { + "type": "header", + "text": "R. Liang et al.", + "bbox": [ + 271, + 114, + 370, + 128 + ], + "page_idx": 13 + }, + { + "type": "text", + "text": "Acknowledgements", + "text_level": 1, + "bbox": [ + 217, + 143, + 401, + 162 + ], + "page_idx": 14 + }, + { + "type": "text", + "text": "The authors are grateful for the feedback received from Nicholas Sharp and Huan Ling during the project. We thank the original artists of the 3D assets used in this work: inciprocal, peyman.khaleghi, Kuutti Siitonen, TurboSquid and their artists Hum3D and Amaranthus.", + "bbox": [ + 212, + 178, + 787, + 238 + ], + "page_idx": 14 + }, + { + "type": "text", + "text": "References", + "text_level": 1, + "bbox": [ + 215, + 263, + 323, + 277 + ], + "page_idx": 14 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "1. Poly Haven - The Public 3D Asset Library, https://polyhaven.com", + "2. Balaji, Y., Nah, S., Huang, X., Vahdat, A., Song, J., Zhang, Q., Kreis, K., Aittala, M., Aila, T., Laine, S., Catanzaro, B., Karras, T., Liu, M.Y.: ediff-i: Text-to-image diffusion models with ensemble of expert denoisers. arXiv preprint arXiv:2211.01324 (2022)", + "3. Bangaru, S.P., Li, T.M., Durand, F.: Unbiased warped-area sampling for differentiable rendering. ACM Trans. Graph. 39(6) (nov 2020). https://doi.org/10.1145/3414685.3417833, https://doi.org/10.1145/3414685.3417833", + "4. Barron, J.T., Malik, J.: Shape, illumination, and reflectance from shading. IEEE transactions on pattern analysis and machine intelligence 37(8), 1670-1687 (2014)", + "5. Barrow, H., Tenenbaum, J., Hanson, A., Riseman, E.: Recovering intrinsic scene characteristics. Comput. Vis. Syst 2, 3-26 (1978)", + "6. Bell, S., Bala, K., Snavely, N.: Intrinsic images in the wild. ACM Transactions on Graphics (TOG) 33(4), 159 (2014)", + "7. Black, M.J., Anandan, P.: The robust estimation of multiple motions: Parametric and piecewise-smooth flow fields. Computer vision and image understanding 63(1), 75-104 (1996)", + "8. Boss, M., Jampani, V., Kim, K., Lensch, H.P., Kautz, J.: Two-shot spatially-varying brdf and shape estimation. In: CVPR (2020)", + "9. Bousseau, A., Paris, S., Durand, F.: User-assisted intrinsic images. In: ACM Transactions on Graphics (TOG). vol. 28, p. 130. ACM (2009)", + "10. Chari, P., Ma, S., Ostashev, D., Kadambi, A., Krishnan, G., Wang, J., Aberman, K.: Personalized restoration via dual-pivot tuning. arXiv preprint arXiv:2312.17234 (2023)", + "11. Dai, X., Hou, J., Ma, C.Y., Tsai, S., Wang, J., Wang, R., Zhang, P., Vandenhende, S., Wang, X., Dubey, A., et al.: Emu: Enhancing image generation models using photogenic needles in a haystack. arXiv preprint arXiv:2309.15807 (2023)", + "12. Dastjerdi, M.R.K., Eisenmann, J., Hold-Geoffroy, Y., Lalonde, J.F.: Everlight: Indoor-outdoor editable hdr lighting estimation. In: Proceedings of the IEEE/CVF International Conference on Computer Vision (ICCV). pp. 7420-7429 (October 2023)", + "13. Durkan, C., Bekasov, A., Murray, I., Papamakarios, G.: Neural spline flows. Advances in neural information processing systems 32 (2019)", + "14. Gal, R., Alaluf, Y., Atzmon, Y., Patashnik, O., Bermano, A.H., Chechik, G., Cohen-Or, D.: An image is worth one word: Personalizing text-to-image generation using textual inversion (2022). https://doi.org/10.48550/ARXIV.2208.01618, https://arxiv.org/abs/2208.01618", + "15. Gardner, M.A., Hold-Geoffroy, Y., Sunkavalli, K., Gagné, C., Lalonde, J.F.: Deep parametric indoor lighting estimation. In: ICCV. pp. 7175-7183 (2019)" + ], + "bbox": [ + 218, + 295, + 785, + 839 + ], + "page_idx": 14 + }, + { + "type": "header", + "text": "Photorealistic Object Insertion with Diffusion-Guided Inverse Rendering", + "bbox": [ + 250, + 114, + 730, + 128 + ], + "page_idx": 14 + }, + { + "type": "page_number", + "text": "15", + "bbox": [ + 767, + 116, + 784, + 126 + ], + "page_idx": 14 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "16. Gardner, M.A., Sunkavalli, K., Yumer, E., Shen, X., Gambaretto, E., Gagné, C., Lalonde, J.F.: Learning to predict indoor illumination from a single image. arXiv preprint arXiv:1704.00090 (2017)", + "17. Garon, M., Sunkavalli, K., Hadap, S., Carr, N., Lalonde, J.F.: Fast spatially-varying indoor lighting estimation. In: CVPR. pp. 6908-6917 (2019)", + "18. Grosse, R., Johnson, M.K., Adelson, E.H., Freeman, W.T.: Ground truth dataset and baseline evaluations for intrinsic image algorithms. In: ICCV. pp. 2335-2342. IEEE (2009)", + "19. Haque, A., Tancik, M., Efros, A., Holynski, A., Kanazawa, A.: Instruct-nerf2nerf: Editing 3d scenes with instructions. In: Proceedings of the IEEE/CVF International Conference on Computer Vision (2023)", + "20. Ho, J., Jain, A., Abbeel, P.: Denoising diffusion probabilistic models. arXiv preprint arxiv:2006.11239 (2020)", + "21. Ho, J., Salimans, T.: Classifier-free diffusion guidance. arXiv preprint arXiv:2207.12598 (2022)", + "22. Hold-Geoffroy, Y., Athawale, A., Lalonde, J.F.: Deep sky modeling for single image outdoor lighting estimation. In: CVPR. pp. 6927-6935 (2019)", + "23. Hold-Geoffroy, Y., Sunkavalli, K., Hadap, S., Gambaretto, E., Lalonde, J.F.: Deep outdoor illumination estimation. In: CVPR. pp. 7312-7321 (2017)", + "24. Hu, E.J., yelong shen, Wallis, P., Allen-Zhu, Z., Li, Y., Wang, S., Wang, L., Chen, W.: LoRA: Low-rank adaptation of large language models. In: International Conference on Learning Representations (2022), https://openreview.net/forum?id=nZeVKeeFYf9", + "25. Jakob, W., Speierer, S., Roussel, N., Nimier-David, M., Vicini, D., Zeltner, T., Nicolet, B., Crespo, M., Leroy, V., Zhang, Z.: Mitsuba 3 renderer (2022), https://mitsubarenderer.org", + "26. Kajiya, J.T.: The rendering equation. In: Proceedings of the 13th annual conference on Computer graphics and interactive techniques. pp. 143-150 (1986)", + "27. Karimi Dastjerdi, M.R., Hold-Geoffroy, Y., Eisenmann, J., Khodadadeh, S., Lalonde, J.F.: Guided co-modulated gan for 360 field of view extrapolation. International Conference on 3D Vision (3DV) (2022)", + "28. Ke, B., Obukhov, A., Huang, S., Metzger, N., Daudt, R.C., Schindler, K.: Repurposing diffusion-based image generators for monocular depth estimation. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR) (2024)", + "29. Kocsis, P., Sitzmann, V., Nießner, M.: Intrinsic image diffusion for single-view material estimation. In: arxiv (2023)", + "30. Kovacs, B., Bell, S., Snavely, N., Bala, K.: Shading annotations in the wild. In: CVPR. pp. 6998-7007 (2017)", + "31. Land, E.H., McCann, J.J.: Lightness and retina theory. Josa 61(1), 1-11 (1971)", + "32. LeGendre, C., Ma, W.C., Fyffe, G., Flynn, J., Charbonnel, L., Busch, J., Debevec, P.: Deeplight: Learning illumination for unconstrained mobile mixed reality. In: CVPR. pp. 5918-5928 (2019)", + "33. Li, T.M., Aittala, M., Durand, F., Lehtinen, J.: Differentiable monte carlo ray tracing through edge sampling. ACM Trans. Graph. 37(6) (dec 2018). https://doi.org/10.1145/3272127.3275109, https://doi.org/10.1145/3272127.3275109", + "34. Li, Z., Snavely, N.: Cginintrinsics: Better intrinsic image decomposition through physically-based rendering. In: ECCV. pp. 371-387 (2018)", + "35. Li, Z., Shafiei, M., Ramamoorthi, R., Sunkavalli, K., Chandraker, M.: Inverse rendering for complex indoor scenes: Shape, spatially-varying lighting and svbrdf from a single image. In: CVPR. pp. 2475-2484 (2020)" + ], + "bbox": [ + 215, + 146, + 785, + 839 + ], + "page_idx": 15 + }, + { + "type": "page_number", + "text": "16", + "bbox": [ + 217, + 114, + 235, + 126 + ], + "page_idx": 15 + }, + { + "type": "header", + "text": "R. Liang et al.", + "bbox": [ + 271, + 114, + 369, + 127 + ], + "page_idx": 15 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "36. Li, Z., Xu, Z., Ramamoorthi, R., Sunkavalli, K., Chandraker, M.: Learning to reconstruct shape and spatially-varying reflectance from a single image. ACM Transactions on Graphics (TOG) 37(6), 1-11 (2018)", + "37. Li, Z., Yu, L., Okunev, M., Chandraker, M., Dong, Z.: Spatiotemporally consistent hdr indoor lighting estimation. ACM Trans. Graph. 42(3) (jun 2023). https://doi.org/10.1145/3595921, https://doi.org/10.1145/3595921", + "38. Li, Z., Yu, T.W., Sang, S., Wang, S., Bi, S., Xu, Z., Yu, H.X., Sunkavalli, K., Hasan, M., Ramamoorthi, R., et al.: Openrooms: An end-to-end open framework for photorealistic indoor scene datasets. arXiv preprint arXiv:2007.12868 (2020)", + "39. Loubet, G., Holzschuch, N., Jakob, W.: Reparameterizing discontinuous integrands for differentiable rendering. ACM Trans. Graph. 38(6) (nov 2019). https://doi.org/10.1145/3355089.3356510, https://doi.org/10.1145/3355089.3356510", + "40. Lyu, L., Tewari, A., Habermann, M., Saito, S., Zollhöfer, M., Leimkuehler, T., Theobalt, C.: Diffusion posterior illumination for ambiguity-aware inverse rendering. ACM Transactions on Graphics 42(6) (2023)", + "41. Meng, C., He, Y., Song, Y., Song, J., Wu, J., Zhu, J.Y., Ermon, S.: SDEdit: Guided image synthesis and editing with stochastic differential equations. In: International Conference on Learning Representations (2022)", + "42. Nimier-David, M., Speierer, S., Ruiz, B., Jakob, W.: Radiative backpropagation: an adjoint method for lightning-fast differentiable rendering. ACM Trans. Graph. 39(4) (aug 2020). https://doi.org/10.1145/3386569.3392406, https://doi.org/10.1145/3386569.3392406", + "43. Nimier-David, M., Vicini, D., Zeltner, T., Jakob, W.: Mitsuba 2: a retargetable forward and inverse renderer. ACM Trans. Graph. 38(6) (nov 2019). https://doi.org/10.1145/3355089.3356498, https://doi.org/10.1145/3355089.3356498", + "44. Phongthawee, P., Chinchuthakun, W., Sinsunthithet, N., Raj, A., Jampani, V., Khungurn, P., Suwajanakorn, S.: Diffusionlight: Light probes for free by painting a chrome ball. In: ArXiv (2023)", + "45. Poole, B., Jain, A., Barron, J.T., Mildenhall, B.: Dreamfusion: Text-to-3d using 2d diffusion. arXiv (2022)", + "46. Rombach, R., Blattmann, A., Lorenz, D., Esser, P., Ommer, B.: High-resolution image synthesis with latent diffusion models (2021)", + "47. Ruiz, N., Li, Y., Jampani, V., Pritch, Y., Rubinstein, M., Aberman, K.: Dreambooth: Fine tuning text-to-image diffusion models for subject-driven generation (2022)", + "48. Sahara, C., Chan, W., Saxena, S., Li, L., Whang, J., Denton, E.L., Ghasemipour, K., Gontijo Lopes, R., Karagol Ayan, B., Salimans, T., et al.: Photorealistic text-to-image diffusion models with deep language understanding. Advances in Neural Information Processing Systems 35, 36479-36494 (2022)", + "49. Sarkar, A., Mai, H., Mahapatra, A., Lazebnik, S., Forsyth, D., Bhattad, A.: Shadows don't lie and lines can't bend! generative models don't know projective geometry...for now (2023)", + "50. Sengupta, S., Gu, J., Kim, K., Liu, G., Jacobs, D.W., Kautz, J.: Neural inverse rendering of an indoor scene from a single image. In: ICCV (2019)", + "51. Shah, V., Ruiz, N., Cole, F., Lu, E., Lazebnik, S., Li, Y., Jampani, V.:Ziplora: Any subject in any style by effectively merging loras (2023)", + "52. Shi, J., Xiong, W., Lin, Z., Jung, H.J.: Instantbooth: Personalized text-to-image generation without test-time finetuning. arXiv preprint arXiv:2304.03411 (2023)", + "53. Song, S., Funkhouser, T.: Neural illumination: Lighting prediction for indoor environments. In: CVPR. pp. 6918-6926 (2019)" + ], + "bbox": [ + 215, + 146, + 785, + 839 + ], + "page_idx": 16 + }, + { + "type": "header", + "text": "Photorealistic Object Insertion with Diffusion-Guided Inverse Rendering", + "bbox": [ + 250, + 114, + 730, + 128 + ], + "page_idx": 16 + }, + { + "type": "page_number", + "text": "17", + "bbox": [ + 767, + 114, + 785, + 126 + ], + "page_idx": 16 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "54. Srinivasan, P.P., Mildenhall, B., Tancik, M., Barron, J.T., Tucker, R., Snavely, N.: Lighthouse: Predicting lighting volumes for spatially-coherent illumination. In: CVPR. pp. 8080-8089 (2020)", + "55. Sun, P., Kretzschmar, H., Dotiwalla, X., Chouard, A., Patnaik, V., Tsui, P., Guo, J., Zhou, Y., Chai, Y., Caine, B., et al.: Scalability in perception for autonomous driving: Waymo open dataset. In: CVPR (2020)", + "56. Tang, J., Zhu, Y., Wang, H., Chan, J.H., Li, S., Shi, B.: Estimating spatially-varying lighting in urban scenes with disentangled representation. In: ECCV (2022)", + "57. Tang, L., Ruiz, N., Qinghao, C., Li, Y., Holynski, A., Jacobs, D.E., Hariharan, B., Pritch, Y., Wadhwa, N., Aberman, K., Rubinstein, M.: Realfill: Reference-driven generation for authentic image completion. arXiv preprint arXiv:2309.16668 (2023)", + "58. Tao, A., Sapra, K., Catanzaro, B.: Hierarchical multi-scale attention for semantic segmentation. arXiv preprint arXiv:2005.10821 (2020)", + "59. Veach, E., Guibas, L.J.: Optimally combining sampling techniques for monte carlo rendering. In: Proceedings of the 22nd annual conference on Computer graphics and interactive techniques. pp. 419-428 (1995)", + "60. Vicini, D., Speierer, S., Jakob, W.: Path replay backpropagation: differentiating light paths using constant memory and linear time. ACM Trans. Graph. 40(4) (jul 2021). https://doi.org/10.1145/3450626.3459804, https://doi.org/10.1145/3450626.3459804", + "61. Wang, G., Yang, Y., Loy, C.C., Liu, Z.: Stylelight: HDR panorama generation for lighting estimation and editing. In: European Conference on Computer Vision (ECCV) (2022)", + "62. Wang, Z., Chen, W., Acuna, D., Kautz, J., Fidler, S.: Neural light field estimation for street scenes with differentiable virtual object insertion. In: ECCV (2022)", + "63. Wang, Z., Philion, J., Fidler, S., Kautz, J.: Learning indoor inverse rendering with 3d spatially-varying lighting. In: ICCV (2021)", + "64. Wimbauer, F., Wu, S., Rupprecht, C.: De-rendering 3d objects in the wild. In: CVPR (2022)", + "65. Yan, K., Lassner, C., Budge, B., Dong, Z., Zhao, S.: Efficient estimation of boundary integrals for path-space differentiable rendering. ACM Trans. Graph. 41(4) (jul 2022). https://doi.org/10.1145/3528223.3530080, https://doi.org/10.1145/3528223.3530080", + "66. Yang, J., Ivanovic, B., Litany, O., Weng, X., Kim, S.W., Li, B., Che, T., Xu, D., Fidler, S., Pavone, M., Wang, Y.: Emernerf: Emergent spatial-temporal scene decomposition via self-supervision. arXiv preprint arXiv:2311.02077 (2023)", + "67. Yu, H.X., Agarwala, S., Herrmann, C., Szeliski, R., Snavely, N., Wu, J., Sun, D.: Accidental light probes. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 12521-12530 (2023)", + "68. Yu, X., Guo, Y.C., Li, Y., Liang, D., Zhang, S.H., Qi, X.: Text-to-3d with classifier score distillation (2023)", + "69. Yu, Y., Smith, W.A.: Inverserendernet: Learning single image inverse rendering. In: CVPR (2019)", + "70. Zhan, F., Zhang, C., Yu, Y., Chang, Y., Lu, S., Ma, F., Xie, X.: Emlight: Lighting estimation via spherical distribution approximation. In: Proceedings of the AAAI Conference on Artificial Intelligence (2021)", + "71. Zhang, C., Miller, B., Yan, K., Gkioulekas, I., Zhao, S.: Path-space differentiable rendering. ACM Trans. Graph. 39(4) (aug 2020). https://doi.org/10.1145/3386569.3392383, https://doi.org/10.1145/3386569.3392383" + ], + "bbox": [ + 215, + 146, + 787, + 839 + ], + "page_idx": 17 + }, + { + "type": "page_number", + "text": "18", + "bbox": [ + 217, + 114, + 235, + 126 + ], + "page_idx": 17 + }, + { + "type": "header", + "text": "R. Liang et al.", + "bbox": [ + 271, + 114, + 370, + 128 + ], + "page_idx": 17 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "72. Zhang, C., Yu, Z., Zhao, S.: Path-space differentiable rendering of participating media. ACM Trans. Graph. 40(4) (jul 2021). https://doi.org/10.1145/3450626.3459782, https://doi.org/10.1145/3450626.3459782", + "73. Zhang, R., Isola, P., Efros, A.A., Shechtman, E., Wang, O.: The unreasonable effectiveness of deep features as a perceptual metric. In: Proceedings of the IEEE conference on computer vision and pattern recognition. pp. 586-595 (2018)", + "74. Zhang, Z., Roussel, N., Jakob, W.: Projective sampling for differentiable rendering of geometry. ACM Trans. Graph. 42(6) (dec 2023). https://doi.org/10.1145/3618385, https://doi.org/10.1145/3618385", + "75. Zhao, Q., Tan, P., Dai, Q., Shen, L., Wu, E., Lin, S.: A closed-form solution to retinax with nonlocal texture constraints 34(7), 1437-1444 (2012)", + "76. Zhao, Y., Guo, T.: Pointar: Efficient lighting estimation for mobile augmented reality. arXiv preprint arXiv:2004.00006 (2020)", + "77. Zhu, Y., Zhang, Y., Li, S., Shi, B.: Spatially-varying outdoor lighting estimation from intrinsics. In: CVPR (2021)" + ], + "bbox": [ + 215, + 146, + 787, + 354 + ], + "page_idx": 18 + }, + { + "type": "header", + "text": "Photorealistic Object Insertion with Diffusion-Guided Inverse Rendering", + "bbox": [ + 250, + 114, + 730, + 128 + ], + "page_idx": 18 + }, + { + "type": "page_number", + "text": "19", + "bbox": [ + 767, + 114, + 785, + 126 + ], + "page_idx": 18 + } +] \ No newline at end of file diff --git a/2024/Photorealistic Object Insertion with Diffusion-Guided Inverse Rendering/27eaedd8-53d6-44df-a63a-fb92723dff8a_model.json b/2024/Photorealistic Object Insertion with Diffusion-Guided Inverse Rendering/27eaedd8-53d6-44df-a63a-fb92723dff8a_model.json new file mode 100644 index 0000000000000000000000000000000000000000..fe839a37a4021f53d098a4b5f16765690e30e132 --- /dev/null +++ b/2024/Photorealistic Object Insertion with Diffusion-Guided Inverse Rendering/27eaedd8-53d6-44df-a63a-fb92723dff8a_model.json @@ -0,0 +1,3186 @@ +[ + [ + { + "type": "title", + "bbox": [ + 0.296, + 0.141, + 0.707, + 0.186 + ], + "angle": 0, + "content": "Photorealistic Object Insertion with Diffusion-Guided Inverse Rendering" + }, + { + "type": "text", + "bbox": [ + 0.252, + 0.212, + 0.753, + 0.244 + ], + "angle": 0, + "content": "Ruofan Liang\\(^{1,2,3}\\), Zan Gojcic\\(^{1}\\), Merlin Nimier-David\\(^{1}\\), David Acuna\\(^{1}\\), Nandita Vijaykumar\\(^{2,3}\\), Sanja Fidler\\(^{1,2,3}\\), and Zian Wang\\(^{1,2,3}\\)" + }, + { + "type": "text", + "bbox": [ + 0.318, + 0.254, + 0.684, + 0.269 + ], + "angle": 0, + "content": "\\(^{1}\\)NVIDIA \\(^{2}\\)University of Toronto \\(^{3}\\)Vector Institute" + }, + { + "type": "text", + "bbox": [ + 0.263, + 0.305, + 0.74, + 0.485 + ], + "angle": 0, + "content": "Abstract. The correct insertion of virtual objects in images of real-world scenes requires a deep understanding of the scene's lighting, geometry and materials, as well as the image formation process. While recent large-scale diffusion models have shown strong generative and inpainting capabilities, we find that current models do not sufficiently \"understand\" the scene shown in a single picture to generate consistent lighting effects (shadows, bright reflections, etc.) while preserving the identity and details of the composited object. We propose using a personalized large diffusion model as guidance to a physically based inverse rendering process. Our method recovers scene lighting and tone-mapping parameters, allowing the photorealistic composition of arbitrary virtual objects in single frames or videos of indoor or outdoor scenes. Our physically based pipeline further enables automatic materials and tone-mapping refinement." + }, + { + "type": "text", + "bbox": [ + 0.263, + 0.499, + 0.741, + 0.527 + ], + "angle": 0, + "content": "Keywords: Inverse rendering \\(\\cdot\\) Diffusion models \\(\\cdot\\) Personalization \\(\\cdot\\) Virtual object insertion \\(\\cdot\\) Physically based rendering" + }, + { + "type": "title", + "bbox": [ + 0.218, + 0.552, + 0.375, + 0.568 + ], + "angle": 0, + "content": "1 Introduction" + }, + { + "type": "text", + "bbox": [ + 0.217, + 0.583, + 0.787, + 0.643 + ], + "angle": 0, + "content": "Virtual object insertion enables a range of applications from virtual production, to interactive gaming and synthetic data generation. To produce photorealistic insertions, the interactions between the virtual objects and the environment need to be modeled faithfully, such as accurate specular highlights and shadows." + }, + { + "type": "text", + "bbox": [ + 0.217, + 0.645, + 0.788, + 0.749 + ], + "angle": 0, + "content": "A standard virtual object insertion pipeline typically includes three key steps: i) lighting estimation from the input image, ii) 3D proxy geometry creation, and iii) composited image rendering in a rendering engine. However, the first and arguably most important step is still an open research question. Lighting estimation is particularly challenging when dealing with limited inputs such as a single image, captured using a consumer device with a low dynamic range. Indeed, inverse rendering is a fundamentally ill-posed problem." + }, + { + "type": "text", + "bbox": [ + 0.217, + 0.75, + 0.788, + 0.841 + ], + "angle": 0, + "content": "To constrain its solution space, prior works either aimed to define hand-crafted priors [9,18,31,75] or to learn them from the data [15-17,22,23,32,36,56,62,63,77]. However, the former often fall short when applied to real-world scenes, while the latter suffer from scarcity of the ground truth data. As a result, these algorithms are often heavily tailored to a specific domain, e.g. indoor [15-17,50,63] or outdoor scenes [22,23,56,62,77]." + } + ], + [ + { + "type": "page_number", + "bbox": [ + 0.218, + 0.116, + 0.23, + 0.127 + ], + "angle": 0, + "content": "2" + }, + { + "type": "header", + "bbox": [ + 0.272, + 0.115, + 0.371, + 0.129 + ], + "angle": 0, + "content": "R. Liang et al." + }, + { + "type": "image_caption", + "bbox": [ + 0.315, + 0.15, + 0.471, + 0.163 + ], + "angle": 0, + "content": "Virtual Object Insertion" + }, + { + "type": "image", + "bbox": [ + 0.218, + 0.164, + 0.391, + 0.253 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.242, + 0.254, + 0.375, + 0.266 + ], + "angle": 0, + "content": "A vase with dried flowers" + }, + { + "type": "image", + "bbox": [ + 0.398, + 0.164, + 0.571, + 0.253 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.442, + 0.254, + 0.53, + 0.264 + ], + "angle": 0, + "content": "A black SUV car" + }, + { + "type": "image", + "bbox": [ + 0.218, + 0.268, + 0.391, + 0.358 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.265, + 0.36, + 0.344, + 0.371 + ], + "angle": 0, + "content": "A metal bucket" + }, + { + "type": "image", + "bbox": [ + 0.398, + 0.268, + 0.571, + 0.358 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.44, + 0.36, + 0.528, + 0.371 + ], + "angle": 0, + "content": "A white SUV car" + }, + { + "type": "image_caption", + "bbox": [ + 0.647, + 0.15, + 0.731, + 0.163 + ], + "angle": 0, + "content": "Applications" + }, + { + "type": "image", + "bbox": [ + 0.603, + 0.164, + 0.776, + 0.253 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.631, + 0.254, + 0.747, + 0.266 + ], + "angle": 0, + "content": "Insert multiple objects" + }, + { + "type": "image", + "bbox": [ + 0.602, + 0.268, + 0.776, + 0.358 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.598, + 0.359, + 0.778, + 0.372 + ], + "angle": 0, + "content": "Optimize material & tone mapping" + }, + { + "type": "image_caption", + "bbox": [ + 0.214, + 0.38, + 0.789, + 0.437 + ], + "angle": 0, + "content": "Fig.1: We propose DiPIR, a physically based method to recover lighting from a single image, enabling arbitrary virtual object compositing into indoor and outdoor scenes, as well as material and tone-mapping optimization. Project page: https://research.nvidia.com/labs/toronto-ai/DiPIR/" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.447, + 0.789, + 0.567 + ], + "angle": 0, + "content": "To address these challenges, we propose to reuse the strong image generation priors learned by large diffusion models (DMs) [46] as a guidance for inverse rendering. Unlike hand-crafted or supervised data-driven priors that are often specific to a domain, DMs are trained on massive datasets and show a remarkable \"understanding\" of the world and the underlying physical concepts. While DMs still often fail to produce accurate lighting effects such as shadows and reflections [49] in their generations, we observe that they can provide valuable guidance when combined with a physically-based renderer and are adapted to the scene." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.568, + 0.789, + 0.688 + ], + "angle": 0, + "content": "Specifically, we present Diffusion Prior for Inverse Rendering (DiPIR), building on three main contributions. First, we use a physically based renderer to accurately simulate the interaction between the light and the 3D asset to generate the final composited image. We also account for the unknown tone-mapping curve to mimic the camera sensor response. Second, we propose a lightweight personalization scheme of the pre-trained DM, based on the input image and the type of inserted asset. Third, we design a variant of the SDS loss [45] which makes use of this personalization and improves training stability." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.688, + 0.789, + 0.764 + ], + "angle": 0, + "content": "In DiPIR, the DM acts similarly to a human evaluator. It takes the edited image as input and propagates the feedback signal to physically-based scene attributes via differentiable rendering, thus enabling end-to-end optimization. We experimentally show that DiPIR outperforms existing state-of-the-art lighting estimation methods for object insertion across indoor and outdoor datasets." + }, + { + "type": "title", + "bbox": [ + 0.216, + 0.784, + 0.389, + 0.801 + ], + "angle": 0, + "content": "2 Related Work" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.81, + 0.789, + 0.842 + ], + "angle": 0, + "content": "Inverse rendering is the task of recovering intrinsic properties of a scene, including materials, shape, and lighting, from a single or multiple images [5]." + } + ], + [ + { + "type": "header", + "bbox": [ + 0.25, + 0.115, + 0.732, + 0.131 + ], + "angle": 0, + "content": "Photorealistic Object Insertion with Diffusion-Guided Inverse Rendering" + }, + { + "type": "page_number", + "bbox": [ + 0.775, + 0.117, + 0.785, + 0.127 + ], + "angle": 0, + "content": "3" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.147, + 0.788, + 0.207 + ], + "angle": 0, + "content": "The main challenge of inverse rendering lies in the ill-posed nature of the task. Prior knowledge about materials and lighting effects is crucial to solving under-constrained inverse rendering problems, but how to best define and incorporate such priors remains an open research question." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.208, + 0.789, + 0.405 + ], + "angle": 0, + "content": "Early methods formulated inverse rendering as an optimization problem and used hand-crafted regularization terms such as local smoothness and sparsity of materials [4,5,9,18,31,75] to constrain the solution space. However, real-world lighting effects are often complex and difficult to describe using hand-crafted priors. The advent of deep learning has facilitated the learning of data-driven priors from ground truth supervision [6,8,30,34,35,38,50,62-64,69]. Yet, acquiring real-world data with accurately annotated intrinsic decomposition necessitates specialized capture devices [18] or expert annotators [6,30]. Collecting real-world datasets at a large scale is therefore extremely challenging. As a result, existing methods often resort to synthetic datasets [17,35,50,56,63,77], or are carefully designed to be training data-efficient [22,23,62]. The resulting algorithms are therefore often tailored to specific domains, such as indoor [15-17, 50, 63] or day-time outdoor scenes [22,23,56,62,77]." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.41, + 0.789, + 0.668 + ], + "angle": 0, + "content": "Lighting estimation is a subtask of inverse rendering that specifically focuses on inferring lighting and is a core component of photorealistic virtual object insertion pipelines. Existing lighting estimation methods are often designed as feed-forward neural networks that take a single image as input and directly regress lighting information in form of spherical lobes [17,35,70,76], environment maps [16,50,53,56,77], parametric light sources [15], or volumetric lighting [37, 54,62,63]. The environment lighting can also be recovered through the proxy geometry shown in the single image [32,67]. Recently, lighting estimation was also formulated as a generative task. StyleLight [61] is based on a dual-branch StyleGAN model, which first performs GAN inversion through the LDR branch and then generates the panorama with the HDR branch. EverLight [12] instead regresses an initial HDR environment map from the input image and later refines it with a GAN model [27]. Concurrent to our work, DiffusionLight [44] proposed to inpaint a chrome ball in the center of the given image using an inpainting diffusion model and unwarp it to an environment map. To obtain the HDR environment map, DiffusionLight trains a LoRA [24] for exposure bracketing and merges multiple generated LDR chrome balls with different exposures." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.668, + 0.79, + 0.729 + ], + "angle": 0, + "content": "Note that lighting is fundamentally a High Dynamic Range (HDR) quantity, and reproducing this full range still remains challenging. HDR is crucial for basic effects such as sharp shadows, particularly in outdoor day-time settings where the sun is several orders of magnitudes brighter than the rest of the environment." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.735, + 0.789, + 0.842 + ], + "angle": 0, + "content": "Physically-based differentiable rendering. Rendering algorithms such as path tracing [26] aim to produce realistic images of virtual scenes by accurately simulating the physical processes involved in light transport. Physically-based rendering (PBR) typically uses a geometric representation of the scene, scattering functions modeling the behavior of the scene's surfaces, as well as camera and lighting models. Recent years have seen the development of many differentiable rendering methods and frameworks [25, 33, 42, 43, 60, 71, 72]." + } + ], + [ + { + "type": "page_number", + "bbox": [ + 0.218, + 0.116, + 0.23, + 0.127 + ], + "angle": 0, + "content": "4" + }, + { + "type": "header", + "bbox": [ + 0.272, + 0.115, + 0.371, + 0.129 + ], + "angle": 0, + "content": "R. Liang et al." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.147, + 0.788, + 0.208 + ], + "angle": 0, + "content": "While particular attention has been devoted to the difficult problem of computing derivatives at discontinuities [3, 33, 39, 65, 74], such as due to visibility changes, we focus here on the comparatively simpler lighting, material and tone-mapping derivatives, which generally do not introduce discontinuities." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.214, + 0.79, + 0.412 + ], + "angle": 0, + "content": "Diffusion model priors and personalization. Priors learned by the text-to-image DMs [2,11,46,48], on large-scale datasets can be adapted to various applications such as monocular depth estimation [28], authentic image completion [57], and image restoration [10]. Among various adaptation techniques, DreamBooth [47] proposes to finetune the DM on subject images, while Textual Inversion [14] optimizes new word embeddings for target generation. LoRA [24,51] simplifies adaptation by freezing the pretrained model weights and injecting trainable rank decomposition matrices, thus greatly reducing the number of trainable parameters. Most similar to ours are concurrent works on DM adaptation for lighting estimation [44] and intrinsic image decomposition [29]. The former adapts the DM to exposure bracketing for LDR to HDR lifting, while the latter fine-tunes a pretrained SD model to directly output the albedo and BRDF properties of an image." + }, + { + "type": "title", + "bbox": [ + 0.216, + 0.433, + 0.383, + 0.45 + ], + "angle": 0, + "content": "3 Preliminaries" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.464, + 0.789, + 0.587 + ], + "angle": 0, + "content": "Diffusion Models. Diffusion models are a family of generative models built around two key processes. A forward process, which gradually adds noise to data samples \\(\\mathbf{x} \\sim p(\\mathbf{x})\\) removing their structure over time \\(t\\). This is achieved using a noise schedule determined by \\(\\alpha_{t}\\) and \\(\\sigma_{t}\\) as \\(\\mathbf{x}_{t} = \\alpha_{t}\\mathbf{x} + \\sigma_{t}\\boldsymbol{\\epsilon}, \\boldsymbol{\\epsilon} \\sim \\mathcal{N}(\\mathbf{0},\\mathbf{I})\\). In contrast, the reverse process gradually removes this noise, restoring the structure. The reverse process is parameterized by a conditional neural network, \\(\\epsilon_{\\theta}\\), trained to predict the noise \\(\\epsilon\\) at a given timestep \\(t\\) according to the following simplified objective [20]:" + }, + { + "type": "equation", + "bbox": [ + 0.335, + 0.595, + 0.788, + 0.614 + ], + "angle": 0, + "content": "\\[\n\\mathbb {E} _ {\\mathbf {x} \\sim p (\\mathbf {x}), \\boldsymbol {\\epsilon} \\sim \\mathcal {N} (\\mathbf {0}, \\boldsymbol {I}), t \\sim T} \\left[ w (t) | | \\boldsymbol {\\epsilon} _ {\\boldsymbol {\\theta}} (\\mathbf {x} _ {t}, t, \\boldsymbol {c}) - \\boldsymbol {\\epsilon} | | _ {2} ^ {2} \\right], \\tag {1}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.622, + 0.788, + 0.667 + ], + "angle": 0, + "content": "where \\( c \\) represents a condition (e.g. text, image, etc.) that allows controlling the generation process, \\( w(t) \\) represents a time-conditional weighting, and \\( T \\) is a set containing a selection of timesteps." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.668, + 0.79, + 0.745 + ], + "angle": 0, + "content": "In this work, we use a Latent Diffusion Model (LDM) [46] in which the diffusion process is conducted in a lower-dimensional latent space. Specifically, the encoder \\(\\mathcal{E}\\) maps samples from the data distribution \\(\\mathbf{x} \\sim p(\\mathbf{x})\\) into a latent space \\(\\mathbb{Z}\\). The decoder \\(\\mathcal{D}\\) performs the inverse operation, such that \\(\\mathcal{D}(\\mathcal{E}(\\mathbf{x})) \\approx \\mathbf{x}\\). As follows, for LDMs \\(\\mathbf{x}\\) in Eq. 1 is replaced by its latent \\(\\mathbf{z} = \\mathcal{E}(\\mathbf{x})\\)." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.765, + 0.789, + 0.843 + ], + "angle": 0, + "content": "Personalization and Fine-tuning. Fine-tuning all parameters of a pretrained DM requires significant computational resources and time. To alleviate this, LoRA [24] injects trainable low-rank decomposition matrices and aims to learn only the variations from the pretrained weights. Specifically, consider a linear layer represented as \\( \\pmb{h} = \\pmb{W}_0\\pmb{x} \\). Here \\( \\pmb{W}_0 \\in \\mathbb{R}^{n \\times n} \\) and \\( \\pmb{x} \\in \\mathbb{R}^n \\) are the pretrained" + } + ], + [ + { + "type": "header", + "bbox": [ + 0.25, + 0.115, + 0.733, + 0.129 + ], + "angle": 0, + "content": "Photorealistic Object Insertion with Diffusion-Guided Inverse Rendering" + }, + { + "type": "page_number", + "bbox": [ + 0.775, + 0.117, + 0.785, + 0.127 + ], + "angle": 0, + "content": "5" + }, + { + "type": "image", + "bbox": [ + 0.216, + 0.145, + 0.66, + 0.319 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.671, + 0.144, + 0.781, + 0.319 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.214, + 0.324, + 0.788, + 0.45 + ], + "angle": 0, + "content": "Fig. 2: Method overview. Given an input image, we first construct a virtual 3D scene with a virtual object and proxy plane. Our physically-based renderer then differentiably simulates the interactions of the estimizable environment map with the inserted virtual object and its effect on the background scene (shadowing) (left). At each iteration, the rendered image is diffused and passed through a personalized diffusion model (middle). The gradient of the adapted Score Distillation formulation is propagated back to the environment map and the tone-mapping curve through the differentiable renderer. Upon convergence, we recover lighting and tone-mapping parameters, which allow photorealistic compositing of virtual objects from a single image (right)." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.467, + 0.788, + 0.543 + ], + "angle": 0, + "content": "weights and input, respectively. Applying LoRA, this layer is modified to \\( \\pmb{h} = (W_0 + \\pmb{\\Delta W})\\pmb{x} = W_0\\pmb{x} + AB\\pmb{x} \\). Notably, \\( \\pmb{\\Delta W} = \\pmb{A}\\pmb{B} \\) is a combination of low-rank matrices \\( \\pmb{A} \\in \\mathbb{R}^{n \\times r} \\) and \\( \\pmb{B} \\in \\mathbb{R}^{r \\times n} \\) with \\( r \\ll n \\). During LoRA fine-tuning, only the added term \\( \\pmb{\\Delta W} \\) is updated while the parameters \\( W_0 \\) remain unchanged. The rank \\( r \\) is a hyperparameter that trades efficiency with model capacity." + }, + { + "type": "title", + "bbox": [ + 0.216, + 0.565, + 0.331, + 0.58 + ], + "angle": 0, + "content": "4 Method" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.595, + 0.788, + 0.673 + ], + "angle": 0, + "content": "Given a single image as input, DiPIR recovers scene lighting and tone-mapping parameters, with the goal of photorealistic insertion of virtual objects. An overview of our method is shown in Fig. 2. Sec. 4.1 describes our representation and the differentiable rendering process, while Sec. 4.2 and Sec. 4.3 provide details on diffusion model guidance and optimization formulation, respectively." + }, + { + "type": "title", + "bbox": [ + 0.215, + 0.681, + 0.607, + 0.696 + ], + "angle": 0, + "content": "4.1 Physically-based Virtual Object Insertion" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.701, + 0.788, + 0.776 + ], + "angle": 0, + "content": "Virtual scene. Inserting a virtual object \\(\\mathcal{X}\\) into an image \\(\\mathbf{I}_{\\mathrm{bg}} \\in \\mathbb{R}^{h \\times w \\times 3}\\) requires creating a 3D proxy virtual scene, viewed from the correct camera pose. Here, we assume that the user provides a specific placement (pose) for \\(\\mathcal{X}\\), but in some cases, an appropriate pose can also be determined automatically, e.g. by detecting the floor plane and scene scale." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.777, + 0.788, + 0.853 + ], + "angle": 0, + "content": "To model the effects of the inserted object on the original image, such as shadows cast by \\(\\mathcal{X}\\), we also assume a known proxy geometry \\(\\mathcal{P}\\). We found a simple ground plane acting as a shadow catcher underneath the virtual object to be sufficient in all of our experiments. This proxy plane can be easily placed manually or automatically generated based on e.g. depth or LiDAR data." + } + ], + [ + { + "type": "page_number", + "bbox": [ + 0.218, + 0.116, + 0.23, + 0.127 + ], + "angle": 0, + "content": "6" + }, + { + "type": "header", + "bbox": [ + 0.272, + 0.115, + 0.37, + 0.129 + ], + "angle": 0, + "content": "R. Liang et al." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.147, + 0.788, + 0.193 + ], + "angle": 0, + "content": "Light representation. We represent the scene's lighting \\(\\mathbf{L}\\) with a set of \\(N\\) optizable Spherical Gaussian (SG) parameters \\(\\{\\mathbf{c}_k,\\pmb {\\mu}_k,\\sigma_k\\}_{k = 1}^N\\in \\mathbb{R}^{N\\times 7}\\), where the radiance for one SG lobe at the direction \\(\\pmb {v}\\in \\mathbb{R}^3\\) is defined as" + }, + { + "type": "equation", + "bbox": [ + 0.244, + 0.198, + 0.788, + 0.218 + ], + "angle": 0, + "content": "\\[\n\\mathbf {G} _ {k} (\\pmb {v}; \\mathbf {c} _ {k}, \\pmb {\\mu} _ {k}, \\sigma_ {k}) = \\mathbf {c} e ^ {- (1 - \\pmb {v} \\cdot \\pmb {\\mu}) / \\sigma^ {2}}, \\quad \\mathrm {w h e r e} \\mathbf {c} _ {k} \\in \\mathbb {R} ^ {3}, \\pmb {\\mu} _ {k} \\in \\mathbb {R} ^ {3}, \\sigma_ {k} \\in \\mathbb {R} _ {+}. \\quad (2)\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.215, + 0.223, + 0.651, + 0.24 + ], + "angle": 0, + "content": "The overall environment map \\(\\mathbf{L} \\in \\mathbb{R}^{H \\times W \\times 3}\\) is computed as:" + }, + { + "type": "equation", + "bbox": [ + 0.391, + 0.246, + 0.788, + 0.288 + ], + "angle": 0, + "content": "\\[\n\\mathbf {L} _ {i, j} = \\sum_ {k = 1} ^ {N} \\mathbf {G} _ {k} \\left(\\boldsymbol {v} _ {i, j}; \\mathbf {c} _ {k}, \\boldsymbol {\\mu} _ {k}, \\sigma_ {k}\\right), \\tag {3}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.294, + 0.789, + 0.355 + ], + "angle": 0, + "content": "where \\(\\pmb{v}_{i,j}\\) is the direction corresponding to the pixel \\((i,j)\\) using the standard spherical environment parameterization. Note that we chose a SG-based lighting representation for its good convergence properties and simplicity, but many alternatives exist and might be combined with our method." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.36, + 0.789, + 0.422 + ], + "angle": 0, + "content": "Differentiable rendering. Inserting a virtual object into a scene involves simulating the interactions of the estimizable environment map with the inserted virtual object (foreground), and the inserted object's effect on the background scene (shadow). We describe below the details respectively." + }, + { + "type": "text", + "bbox": [ + 0.215, + 0.431, + 0.789, + 0.478 + ], + "angle": 0, + "content": "Foreground image. Given the optimized light representation \\(\\mathbf{L}\\) from above and the virtual object \\(\\mathcal{X}\\) (including its geometry and materials), a foreground image \\(\\mathbf{I}_{\\mathrm{fg}}\\) of the inserted object can be rendered directly using standard path tracing:" + }, + { + "type": "equation", + "bbox": [ + 0.408, + 0.485, + 0.788, + 0.502 + ], + "angle": 0, + "content": "\\[\n\\mathbf {I} _ {\\mathrm {f g}} = \\operatorname {P a t h T r a c e} (\\mathcal {X}, \\mathbf {L}, D), \\tag {4}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.508, + 0.789, + 0.57 + ], + "angle": 0, + "content": "\\(D\\) is the maximum number of interactions along the light path. Since we do not assume that the provided proxy geometry is accurate nor that it has materials, we omit the effect of light reflected by the object into the scene, for example, due to a highly specular inserted object." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.581, + 0.789, + 0.64 + ], + "angle": 0, + "content": "Shadow ratio. As the background image \\(\\mathbf{I}_{\\mathrm{bg}}\\) already faithfully represents the scene, our task is limited to simulating the appearance of the inserted object and its effect on the nearby parts of the scene, i.e. the shadows cast by the object onto the proxy geometry." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.64, + 0.789, + 0.702 + ], + "angle": 0, + "content": "The shadow ratio \\(\\beta_{\\mathrm{shadow}} \\in \\mathbb{R}^{h \\times w \\times 3}\\), which accounts for the effect of the object on the surrounding scene, is computed as the ratio between the radiance received by the proxy geometry with and without the inserted object present. A low value of \\(\\beta_{\\mathrm{shadow}}\\) indicates a strongly shadowed region:" + }, + { + "type": "equation", + "bbox": [ + 0.377, + 0.708, + 0.788, + 0.741 + ], + "angle": 0, + "content": "\\[\n\\boldsymbol {\\beta} _ {\\mathrm {s h a d o w}} = \\frac {\\operatorname {P a t h T r a c e} (\\mathcal {X} \\cup \\mathcal {P} , \\mathbf {L} , 1)}{\\operatorname {P a t h T r a c e} (\\mathcal {P} , \\mathbf {L} , 1)}, \\tag {5}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.746, + 0.789, + 0.853 + ], + "angle": 0, + "content": "This is similar to the method of Wang et al. [62], however we use Multiple Importance Sampling (MIS) [59] between lighting and BSDF for better sampling efficiency. The maximum path length has been limited to 1 interaction in order to reduce memory and computational cost. Unless specific material information is provided with the proxy geometry, we use a Lambertian BSDF, in which case the albedo cancels out in the ratio. In practice, the computation of the shadow ratio and foreground image are combined in order to reuse shared terms." + } + ], + [ + { + "type": "header", + "bbox": [ + 0.249, + 0.115, + 0.733, + 0.131 + ], + "angle": 0, + "content": "Photorealistic Object Insertion with Diffusion-Guided Inverse Rendering" + }, + { + "type": "page_number", + "bbox": [ + 0.775, + 0.117, + 0.785, + 0.127 + ], + "angle": 0, + "content": "7" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.147, + 0.788, + 0.3 + ], + "angle": 0, + "content": "Tone-mapping. To compensate for the unknown tone-mapping of the input image, we introduces an estimable tone correction function \\( f(\\cdot) \\) that is applied on the inserted object \\( \\hat{\\mathbf{I}}_{\\mathrm{fg}} = f(\\mathbf{I}_{\\mathrm{fg}}; \\boldsymbol{\\theta}_{\\mathrm{fg}}) \\) and shadows \\( \\hat{\\beta}_{\\mathrm{shadow}} = f(\\boldsymbol{\\beta}_{\\mathrm{shadow}}; \\boldsymbol{\\theta}_{\\mathrm{shadow}}) \\). Our tone curve parameterization follows from [13] and employs a monotonic rational-quadratic spline as its basic building block. This spline is composed of \\( K_{\\mathrm{s}} \\) bins, each defined by the quotient of two quadratic polynomials. Notably, these functions are differentiable, and allow direct parameterization of the derivatives and heights at each knot. We apply a single spline for the foreground image, and different correction splines for each RGB channel for the shadow ratio to provide the flexibility to adjust the color of the shadow. We set \\( K_{\\mathrm{s}} = 5 \\) for these curves." + }, + { + "type": "text", + "bbox": [ + 0.215, + 0.307, + 0.79, + 0.339 + ], + "angle": 0, + "content": "Differentiability. The final output image is an alpha-composite of the foreground object, shadows, and background image:" + }, + { + "type": "equation", + "bbox": [ + 0.333, + 0.348, + 0.788, + 0.367 + ], + "angle": 0, + "content": "\\[\n\\mathbf {I} _ {\\mathrm {c o m p}} = \\left(1 - \\mathbf {V} (\\mathcal {X})\\right) \\cdot \\hat {\\boldsymbol {\\beta}} _ {\\mathrm {s h a d o w}} \\cdot \\mathbf {I} _ {\\mathrm {b g}} + \\mathbf {V} (\\mathcal {X}) \\cdot \\tilde {\\mathbf {I}} _ {\\mathrm {f g}}, \\tag {6}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.376, + 0.788, + 0.391 + ], + "angle": 0, + "content": "where \\(\\mathbf{V}(\\mathcal{X})\\) equals 1 when \\(\\mathcal{X}\\) is directly visible from the camera, and 0 otherwise." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.392, + 0.788, + 0.497 + ], + "angle": 0, + "content": "As neither foreground rendering, shadow rendering, or the compositing operation require derivatives w.r.t. the object placement or other discontinuous quantities, we can rely on automatic differentiation to obtain gradients of any pixelwise loss w.r.t. the lighting or material properties. To this end, we use the Path Replay Backpropagation [60] integrator of Mitsuba 3 [25]. The estimizable parameters are the Spherical Gaussian coefficients used in the lighting representation, and the tone curves' parameters \\(\\theta_{\\mathrm{fg}},\\theta_{\\mathrm{shadow}}\\)" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.498, + 0.788, + 0.557 + ], + "angle": 0, + "content": "While each of the operations above include simplifications (e.g. we do not account for secondary lighting from the proxy geometry), we have found them to be sufficient in practice: the remaining imperfections in the simulation can be sufficiently offset by the lighting optimization." + }, + { + "type": "title", + "bbox": [ + 0.215, + 0.579, + 0.421, + 0.594 + ], + "angle": 0, + "content": "4.2 Diffusion Guidance" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.604, + 0.788, + 0.726 + ], + "angle": 0, + "content": "The composited image produced by our differentiable rendering pipeline serves as the input to a DM that is used to compute a guidance signal, employing an optimization objective similar to Score Distillation Sampling (SDS) [45]. However, while DMs inherently have robust priors for lighting, we found that they do not provide out-of-the-box the necessary guidance for our specific needs. Consequently, we propose an adaptive score distillation loss specifically designed for object insertion tasks that exploits a personalization strategy which we detail in the following section." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.732, + 0.788, + 0.84 + ], + "angle": 0, + "content": "Personalization with concept preservation. Off-the-shelf DMs often do not provide robust guidance for virtual object insertion, especially in out-of-distribution scenes such as outdoor driving environments. A potential solution, inspired by [57], is to adapt the DM using an image from the target scene. In our experience, however, this approach often resulted in too much overfitting to the target scene's content, reducing the model's ability to adapt to the scene with a newly inserted object. This led to artifacts and an unstable optimization process." + } + ], + [ + { + "type": "page_number", + "bbox": [ + 0.218, + 0.116, + 0.23, + 0.127 + ], + "angle": 0, + "content": "8" + }, + { + "type": "header", + "bbox": [ + 0.272, + 0.115, + 0.371, + 0.129 + ], + "angle": 0, + "content": "R. Liang et al." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.147, + 0.49, + 0.342 + ], + "angle": 0, + "content": "To mitigate this issue, we propose fine-tuning the DM with a focus on preserving the identity of the objects to be inserted (Fig. 3). We specifically achieved this by generating additional synthetic images for the insertable class concept (e.g. car). We sample those images from the off-the-self DM, starting with a base prompt such as \"a photo of a car\" and appending attributes such as color, background, lighting, and size to ensure diversity in the generated data." + }, + { + "type": "image", + "bbox": [ + 0.501, + 0.156, + 0.603, + 0.216 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.523, + 0.217, + 0.581, + 0.227 + ], + "angle": 0, + "content": "Input image" + }, + { + "type": "image_caption", + "bbox": [ + 0.612, + 0.16, + 0.788, + 0.172 + ], + "angle": 0, + "content": "\"ascene in the style of sks rendering" + }, + { + "type": "image", + "bbox": [ + 0.5, + 0.228, + 0.603, + 0.305 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.51, + 0.306, + 0.593, + 0.314 + ], + "angle": 0, + "content": "Generated images" + }, + { + "type": "image", + "bbox": [ + 0.612, + 0.184, + 0.762, + 0.265 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.613, + 0.284, + 0.699, + 0.295 + ], + "angle": 0, + "content": "\"aphotofacar\"" + }, + { + "type": "image_caption", + "bbox": [ + 0.499, + 0.321, + 0.785, + 0.334 + ], + "angle": 0, + "content": "Fig. 3: Personalization with concept preservation." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.343, + 0.788, + 0.449 + ], + "angle": 0, + "content": "We employ LoRA with rank 4 to fine-tune the DM, combining the in-domain target example with this supplemental data. Our training follows the objective described in Eq. 1, where \\( c \\) corresponds to two predefined prompts: \"a scene in the style of sks rendering\" for the target image, and \"a photo of a {concept class}\". We sample approximately 30-40 supplementary images for indoor scenes and 200 for outdoor scenes. The total time spent for fine-tuning is typically less than 15 minutes on one high-end GPU with mixed precision training." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.454, + 0.788, + 0.53 + ], + "angle": 0, + "content": "Score Distillation with adapted guidance. Score Distillation Sampling [45] leverages a pretrained DM to guide the optimization of a differentiable, parametric image-rendering function \\( g_{\\phi} \\coloneqq \\pmb{x} \\). Specifically, the parameters \\( \\phi \\) in our scenario corresponds to parameters of the Spherical Gaussian lighting and tone-mapping curves, are updated using the gradient:" + }, + { + "type": "equation", + "bbox": [ + 0.278, + 0.536, + 0.788, + 0.57 + ], + "angle": 0, + "content": "\\[\n\\nabla_ {\\phi} \\mathcal {L} _ {\\mathrm {S D S}} (\\boldsymbol {\\phi}, \\boldsymbol {\\theta}) := \\mathbb {E} _ {\\boldsymbol {\\epsilon} \\sim \\mathcal {N} (\\mathbf {0}, \\boldsymbol {I}), t \\sim T} \\left[ w (t) \\left(\\hat {\\epsilon} _ {\\boldsymbol {\\theta}} (\\boldsymbol {z} _ {t}, t, \\boldsymbol {c}) - \\boldsymbol {\\epsilon}\\right) \\frac {\\partial \\boldsymbol {z} _ {t}}{\\partial \\phi} \\right], \\qquad (7)\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.576, + 0.788, + 0.621 + ], + "angle": 0, + "content": "where \\( \\pmb{z} = \\mathcal{E}(g_{\\phi}) \\) and \\( \\hat{\\epsilon}_{\\theta}(\\pmb{z}_t, t, \\pmb{c}) \\coloneqq (1 + s)\\pmb{\\epsilon}_{\\theta}(\\pmb{z}_t, t, \\pmb{c}) - s\\pmb{\\epsilon}_{\\theta}(\\pmb{z}_t, t, \\emptyset) \\). \\( \\hat{\\epsilon} \\) denotes the classifier free guidance (CFG) version [21] of \\( \\pmb{\\epsilon}_{\\theta} \\) used in text-conditioned DMs to enable higher quality generation via a guidance scale parameter \\( s \\)." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.622, + 0.788, + 0.667 + ], + "angle": 0, + "content": "Unfortunately, we encountered training instabilities when applying the original formulation of the SDS loss to our problem. Instead, we adopt the following alternative that integrates the LoRA personalization, and we call it LDS loss:" + }, + { + "type": "equation", + "bbox": [ + 0.218, + 0.673, + 0.787, + 0.719 + ], + "angle": 0, + "content": "\\[\n\\nabla_ {\\phi} \\mathcal {L} _ {\\mathrm {L D S}} (\\phi , \\boldsymbol {\\theta}) := \\mathbb {E} _ {\\epsilon \\sim \\mathcal {N} (\\boldsymbol {0}, \\boldsymbol {I}), t \\sim T} \\left[ w (t) \\left(\\epsilon_ {(\\boldsymbol {\\theta} + \\boldsymbol {\\Delta W})} (\\boldsymbol {z} _ {t}, t, \\boldsymbol {c}) - \\epsilon_ {\\boldsymbol {\\theta}} (\\boldsymbol {z} _ {t}, t, \\varnothing)\\right) \\frac {\\partial \\boldsymbol {z} _ {t}}{\\partial \\phi} \\right], \\tag {8}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.72, + 0.788, + 0.84 + ], + "angle": 0, + "content": "where \\(\\epsilon_{(\\theta +\\Delta W)}\\) represents the predicted noise of the LoRA personalized model described in the previous section. Notably, this loss function bears resemblance to the Classifier Score Distillation (CSD) [68]. However, it's important to note that in our case, the delta is calculated between the adapted and non-adapted versions of the diffusion model. Intuitively, this delta guides the optimization process in a direction determined by the personalized model which preserves the concept of the inserted object while also capturing the appearance and semantics of the specific scene." + } + ], + [ + { + "type": "header", + "bbox": [ + 0.249, + 0.115, + 0.733, + 0.131 + ], + "angle": 0, + "content": "Photorealistic Object Insertion with Diffusion-Guided Inverse Rendering" + }, + { + "type": "page_number", + "bbox": [ + 0.775, + 0.117, + 0.785, + 0.127 + ], + "angle": 0, + "content": "9" + }, + { + "type": "title", + "bbox": [ + 0.216, + 0.147, + 0.478, + 0.163 + ], + "angle": 0, + "content": "4.3 Optimization Formulation" + }, + { + "type": "text", + "bbox": [ + 0.215, + 0.173, + 0.785, + 0.203 + ], + "angle": 0, + "content": "After completing the diffusion model's personalization, we optimize the lighting and tone-mapping parameters using the following loss function:" + }, + { + "type": "equation", + "bbox": [ + 0.341, + 0.217, + 0.787, + 0.233 + ], + "angle": 0, + "content": "\\[\n\\mathcal {L} = \\mathcal {L} _ {\\mathrm {L D S}} + \\lambda_ {\\text {c o n s i s t e n c y}} \\mathcal {L} _ {\\text {c o n s i s t e n c y}} + \\lambda_ {\\text {r e g}} \\mathcal {L} _ {\\text {r e g}}. \\tag {9}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.215, + 0.244, + 0.788, + 0.379 + ], + "angle": 0, + "content": "The loss' main component is the diffusion guidance \\(\\mathcal{L}_{\\mathrm{LDS}}\\), which provides a perceptual realism objective on the edited result with object insertion. We use the text prompt \"a photo of a {concept class} in a scene in the style of sks rendering\", which is a natural combination of the two text prompts used in personalization (Fig. 3). Here {concept class}, e.g. car, provides the context of the object to insert, while the personalized token sks provides the lighting style of the input image. The diffusion guidance \\(\\mathcal{L}_{\\mathrm{LDS}}\\) is applied on the edited image \\(\\mathbf{I}_{\\mathrm{comp}}\\) (Eq. 6), and backpropagated to the estimizable parameters through the differentiable rendering process (Eq. 8)." + }, + { + "type": "text", + "bbox": [ + 0.215, + 0.381, + 0.787, + 0.41 + ], + "angle": 0, + "content": "The two regularizers \\(\\mathcal{L}_{\\mathrm{consistency}}\\) and \\(\\mathcal{L}_{\\mathrm{reg}}\\) are related to environment map fusion, which we detail below." + }, + { + "type": "text", + "bbox": [ + 0.215, + 0.419, + 0.788, + 0.524 + ], + "angle": 0, + "content": "Environment map initialization and fusion. The personalized DM provides guidance in two ways: (i) encouraging lighting consistency of the foreground object and the scene, such as the reflections and scale of the inserted object; and (ii) encouraging accurate shadows cast by the inserted object onto the background scene, such as the scale, direction, and color of the shadow. However, we empirically observe that these two signals can conflict in the early phase of the optimization." + }, + { + "type": "text", + "bbox": [ + 0.215, + 0.525, + 0.788, + 0.646 + ], + "angle": 0, + "content": "To address this, we initialize two separate environment maps, \\(\\mathbf{L}^{\\mathrm{fg}},\\mathbf{L}^{\\mathrm{shadow}}\\in\\) \\(\\mathbb{R}^{H\\times W\\times 3}\\), to light the foreground inserted object (Eq. 4) and cast shadows (Eq. 5) respectively. As the optimization progresses, the two environment maps are progressively fused into a single environment map \\(\\mathbf{L}^{\\mathrm{fused}}\\) through scaling the the \\(\\mathbf{L}^{\\mathrm{fg}}\\) by the relative luminance of between \\(\\mathbf{L}^{\\mathrm{fg}}\\) and \\(\\mathbf{L}^{\\mathrm{shadow}}\\). We also use the following two regularization terms for this fusion process. First, we encourage the consistency between the normalized luminance \\(\\tilde{\\mathbf{L}}^{\\mathrm{fg}},\\tilde{\\mathbf{L}}^{\\mathrm{shadow}}\\in \\mathbb{R}^{H\\times W}\\) of the environment maps by minimizing:" + }, + { + "type": "equation", + "bbox": [ + 0.347, + 0.658, + 0.787, + 0.692 + ], + "angle": 0, + "content": "\\[\n\\mathcal {L} _ {\\text {c o n s i s t e n c y}} = - \\sum_ {i, j} \\tilde {\\mathbf {L}} _ {i, j} ^ {\\text {s h a d o w}} \\log \\left(\\tilde {\\mathbf {L}} _ {i, j} ^ {\\mathrm {f g}}\\right) \\Delta \\Omega_ {i, j} \\tag {10}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.215, + 0.703, + 0.788, + 0.781 + ], + "angle": 0, + "content": "where \\(\\varDelta\\varOmega_{i,j}\\) is the corresponding solid angle at pixel \\((i,j)\\) and the gradient for \\(\\tilde{\\mathbf{L}}^{\\mathrm{shadow}}\\) is detached. Second, as the shadow environment map \\(\\mathbf{L}^{\\mathrm{shadow}}\\) is supervised mainly via the shadow ratio \\(\\beta_{\\mathrm{shadow}}\\in \\mathbb{R}^{H\\times W\\times 3}\\) of Eq. 5, to encourage concentrated high peaks for the sharp shadows and suppress the ambient light in \\(\\mathbf{L}^{\\mathrm{shadow}}\\), we also add a L2 regularizer in log-space with the Cauchy loss [7]:" + }, + { + "type": "equation", + "bbox": [ + 0.359, + 0.792, + 0.787, + 0.827 + ], + "angle": 0, + "content": "\\[\n\\mathcal {L} _ {\\text {r e g}} = \\sum_ {i, j, c} \\log \\left(1 + 2 \\left(\\mathbf {L} _ {i, j, c} ^ {\\text {s h a d o w}}\\right) ^ {2}\\right) \\Delta \\Omega_ {i, j}. \\tag {11}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.216, + 0.837, + 0.564, + 0.852 + ], + "angle": 0, + "content": "Please refer to the Supplement for more details." + } + ], + [ + { + "type": "page_number", + "bbox": [ + 0.218, + 0.116, + 0.236, + 0.127 + ], + "angle": 0, + "content": "10" + }, + { + "type": "header", + "bbox": [ + 0.272, + 0.115, + 0.371, + 0.129 + ], + "angle": 0, + "content": "R. Liang et al." + }, + { + "type": "table_caption", + "bbox": [ + 0.214, + 0.145, + 0.788, + 0.214 + ], + "angle": 0, + "content": "Table 1: Quantitative user study on outdoor street scenes. For each scene, users are shown two results—one produced by our method, and another produced by one of the baselines—and select which is more realistic. We report the results averaged across 3 user studies with 9 users each. Our method outperforms all baselines (\\(>50\\%\\)) and is preferred in almost all illumination conditions." + }, + { + "type": "table", + "bbox": [ + 0.299, + 0.217, + 0.704, + 0.291 + ], + "angle": 0, + "content": "
DaytimeTwilightNightAll scenes
SunnyCloudy
Hold-Geoffroy et al. [22]60.8%66.7%74.1%85.7%68.8%
NLFE [62]80.4%73.3%44.4%52.4%67.4%
StyleLight [61]76.5%91.1%66.7%66.7%77.8%
DiffusionLight [44]80.4%68.9%55.6%71.4%70.8%
" + }, + { + "type": "table_caption", + "bbox": [ + 0.214, + 0.296, + 0.787, + 0.338 + ], + "angle": 0, + "content": "Table 2: Quantitative evaluation on PolyHaven scenes. We report user study preference scores similar to Table 1. Metrics are computed w.r.t. to a “reference” image where the virtual object is lit by the ground-truth environment map." + }, + { + "type": "table", + "bbox": [ + 0.298, + 0.34, + 0.704, + 0.399 + ], + "angle": 0, + "content": "
MethodsOurs preferred ↓RMSE ↓SSIM ↑LPIPS [73] ↓si-RMSE [18] ↓
Wang et al. [63]84.8%0.0630.9850.01750.037
StyleLight [61]75.8%0.0560.9860.02020.039
DiffusionLight [44]66.7%0.0620.9850.01620.034
Ours/0.0480.9890.01470.027
" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.41, + 0.788, + 0.547 + ], + "angle": 0, + "content": "Training details. The personalized DM receives a crop of \\(\\mathbf{I}_{\\mathrm{comp}}\\) (Eq. 6) with added noise. Since the inserted object likely does not cover the full background image, we use the visibility mask of the virtual object \\(V(\\mathcal{X})\\) to locate the 2D bounding box of the inserted object, and randomly crop around the inserted object. The size of the cropped image is also randomized, with smaller crops covering local details of the inserted object and larger crops providing more visual cues of the lighting effects within the scene. Each scene is optimized for 600 iterations, and the maximum strength of the diffusion guidance is linearly decreased from 0.6 to 0.3. We use \\(\\lambda_{\\mathrm{consistency}} = 0.03\\) and \\(\\lambda_{\\mathrm{reg}} = 0.01\\)." + }, + { + "type": "title", + "bbox": [ + 0.215, + 0.563, + 0.377, + 0.581 + ], + "angle": 0, + "content": "5 Experiments" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.585, + 0.788, + 0.646 + ], + "angle": 0, + "content": "In this section, we evaluate DiPIR on a collection of outdoor and indoor scenes, covering diverse lighting conditions and different application domains. We first outline the experiment settings in Sec. 5.1 and then provide evaluation results on benchmark datasets in Sec. 5.2. Finally, we conduct an ablation study in Sec. 5.3." + }, + { + "type": "title", + "bbox": [ + 0.215, + 0.658, + 0.434, + 0.674 + ], + "angle": 0, + "content": "5.1 Experiment Settings" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.68, + 0.788, + 0.801 + ], + "angle": 0, + "content": "Waymo dataset. Following the data split from [66], we use 48 scenes with diverse illumination conditions from the Waymo Open Dataset [55] including 32 daytime (17 sunny and 15 cloudy), 9 twilight, and 7 night scenes. Each scene consists of an input image, a 3D car asset randomly selected from an asset bank, the ground plane, and a location to insert the car. Inspired by applications in synthetic data generation, this process is fully automatic: we use the LiDAR point cloud and semantic segmentation [58] to fit a ground plane and detect the empty space to insert a car." + }, + { + "type": "text", + "bbox": [ + 0.215, + 0.81, + 0.787, + 0.84 + ], + "angle": 0, + "content": "PolyHaven dataset. We use 11 HDR environment maps from PolyHaven [1] and manually place a known ground plane and virtual object in each scene. Each" + } + ], + [ + { + "type": "header", + "bbox": [ + 0.25, + 0.115, + 0.733, + 0.13 + ], + "angle": 0, + "content": "Photorealistic Object Insertion with Diffusion-Guided Inverse Rendering" + }, + { + "type": "page_number", + "bbox": [ + 0.768, + 0.116, + 0.784, + 0.127 + ], + "angle": 0, + "content": "11" + }, + { + "type": "image", + "bbox": [ + 0.219, + 0.144, + 0.785, + 0.297 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.236, + 0.299, + 0.766, + 0.313 + ], + "angle": 0, + "content": "Fig. 4: Comparisons on inserting objects into cropped HDRIs from PolyHaven." + }, + { + "type": "image", + "bbox": [ + 0.219, + 0.315, + 0.785, + 0.581 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.214, + 0.583, + 0.787, + 0.611 + ], + "angle": 0, + "content": "Fig. 5: Inserting car assets into Waymo driving scenes. Note the direction and sharpness of shadows, as well as overall brightness, color and specular highlights on inserted cars." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.619, + 0.788, + 0.68 + ], + "angle": 0, + "content": "scenario therefore consists of an LDR background image \\(\\mathbf{I}_{\\mathrm{bg}}\\) rendered directly from the environment map, the posed virtual object, and the proxy plane. A pseudo-ground-truth rendering is created by lighting the inserted object with the environment map itself." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.69, + 0.789, + 0.841 + ], + "angle": 0, + "content": "User study. To evaluate the perceptual realism of the virtual object insertion, we conducted a user study where participants received a pair of two object insertion results in a random order - one generated using our proposed method, the other using a baseline approach. The participants were then instructed to compare the differences between the two results, inspect the lighting effects of the inserted object, and choose the image they perceived as more realistic. We invited 9 users to perform a binary selection for each image pair and used majority voting to determine the preferred image for each comparison. In the following, we report the percentage of times that our method was selected over the baseline. This process was repeated three times and a preferred percentage \\(>50\\%\\) indicates" + } + ], + [ + { + "type": "page_number", + "bbox": [ + 0.218, + 0.116, + 0.236, + 0.127 + ], + "angle": 0, + "content": "12" + }, + { + "type": "header", + "bbox": [ + 0.272, + 0.115, + 0.371, + 0.129 + ], + "angle": 0, + "content": "R. Liang et al." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.147, + 0.784, + 0.178 + ], + "angle": 0, + "content": "Ours outperforming baselines. We include more details about the user study and statistical evaluation of its results in the Supplement." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.185, + 0.788, + 0.276 + ], + "angle": 0, + "content": "Baselines. On the Waymo dataset of urban street scenes, we compare our method to the representative outdoor lighting estimation methods of Hold-Geoffroy et al. [22] and NLFE [62]; as well as the generative lighting estimation methods StyleLight [61] and DiffusionLight [44]. On the PolyHaven dataset, we compare with Wang et al. [63], StyleLight [61], DiffusionLight [44], and pseudo-ground-truth rendering." + }, + { + "type": "title", + "bbox": [ + 0.215, + 0.297, + 0.553, + 0.312 + ], + "angle": 0, + "content": "5.2 Evaluation on Benchmark Datasets" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.318, + 0.787, + 0.379 + ], + "angle": 0, + "content": "Urban street scenes. In Table 1, we provide quantitative user study results on the Waymo dataset [55], separated into 4 subsets based on illumination conditions, as well as on all scenes. Compared to all baselines, our method outperforms prior state-of-the-art methods (chosen \\(>50\\%\\)), and is preferred in almost every subset." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.379, + 0.788, + 0.529 + ], + "angle": 0, + "content": "StyleLight [61] is primarily trained on indoor HDR panoramas and suffers from domain gap to outdoor scenes. Similarly, the sky model predicted by Hold-Geoffroy et al. [22] is tailored to outdoor daytime scenes and suffers severely from out-of-domain twilight and night scenes. NLFE [62] unprojects the surrounding scene geometry and predicts volumetric lighting, which makes it generalize better to night scenes, but it can fail to estimate an accurate scale and color of the peak for daytime scenes. DiffusionLight [44], a concurrent work, shows impressive results by inpainting a chrome sphere and predicting environment maps with realistic high-frequency details. However, it often struggles to predict a high-intensity peak in sunny daytime scenes or a proper scale for night scenes." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.53, + 0.787, + 0.576 + ], + "angle": 0, + "content": "A qualitative comparison is shown in Fig. 5. Our method achieves high-quality object insertion, making it a promising approach for applications such as simulating synthetic data, augmented reality navigation and urban planning." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.583, + 0.787, + 0.658 + ], + "angle": 0, + "content": "PolyHaven dataset. We show the quantitative comparison in Table 2 and qualitative results in Fig. 4. Our method is preferred over all baseline methods in the user study, and outperforms baselines on quantitative metrics. Our method produces virtual object insertion that naturally and consistently blend in the scenes, making it promising for virtual production tasks." + }, + { + "type": "title", + "bbox": [ + 0.215, + 0.68, + 0.39, + 0.696 + ], + "angle": 0, + "content": "5.3 Ablation Study" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.705, + 0.788, + 0.841 + ], + "angle": 0, + "content": "We extensively ablate our proposed design choices on diffusion guidance and scene representations, and provide the quantitative ablation in Table 3 with user study and qualitative results in Fig. 6. For the diffusion guidance, we compare with simplified or alternative versions: (i) Ours (SDS) removes the adaptive guidance (Eq. 8) and uses the original SDS [45] formulation (Eq. 7); (ii) Ours (SDS w/o LoRA) uses the original SDS loss with non-adapted DM; (iii) Ours (w/o concept preservation) uses our LDS loss with adapted DM but not using any class images for concept preservation. We additionally compare with a baseline of \"dataset update\" which applies a photometric L1 loss on DM repeatedly edited results" + } + ], + [ + { + "type": "header", + "bbox": [ + 0.25, + 0.115, + 0.733, + 0.13 + ], + "angle": 0, + "content": "Photorealistic Object Insertion with Diffusion-Guided Inverse Rendering" + }, + { + "type": "page_number", + "bbox": [ + 0.768, + 0.116, + 0.786, + 0.127 + ], + "angle": 0, + "content": "13" + }, + { + "type": "table", + "bbox": [ + 0.223, + 0.144, + 0.486, + 0.226 + ], + "angle": 0, + "content": "
MethodsOurs preferred ↓
Ours (dataset update)85.2%
Ours (SDS [45])74.1%
Ours (SDS [45] w/o LoRA)90.7%
Ours (w/o concept preservation)64.8%
Ours (w/o tone curve)68.5%
Ours (w/o env. map fusion)66.7%
" + }, + { + "type": "table_caption", + "bbox": [ + 0.506, + 0.145, + 0.795, + 0.227 + ], + "angle": 0, + "content": "Table 3: Ablation study on outdoor driving scenes [55]. We report the percentage of images that users preferred DiPIR compared to its ablated versions. Our full pipeline produces results that are preferred more often over its ablated versions." + }, + { + "type": "image", + "bbox": [ + 0.218, + 0.235, + 0.36, + 0.309 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.223, + 0.31, + 0.356, + 0.32 + ], + "angle": 0, + "content": "Ours (SDS w/o LoRA)" + }, + { + "type": "image", + "bbox": [ + 0.362, + 0.236, + 0.501, + 0.309 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.369, + 0.31, + 0.493, + 0.32 + ], + "angle": 0, + "content": "Ours (SDS w/ LoRA)" + }, + { + "type": "image", + "bbox": [ + 0.503, + 0.236, + 0.642, + 0.309 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.509, + 0.31, + 0.636, + 0.321 + ], + "angle": 0, + "content": "Ours (dataset update)" + }, + { + "type": "image", + "bbox": [ + 0.645, + 0.278, + 0.785, + 0.35 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.701, + 0.351, + 0.731, + 0.36 + ], + "angle": 0, + "content": "Ours" + }, + { + "type": "image", + "bbox": [ + 0.219, + 0.322, + 0.36, + 0.394 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.235, + 0.395, + 0.345, + 0.416 + ], + "angle": 0, + "content": "Ours (w/o concept preservation)" + }, + { + "type": "image", + "bbox": [ + 0.362, + 0.322, + 0.501, + 0.394 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.36, + 0.395, + 0.504, + 0.416 + ], + "angle": 0, + "content": "Ours (w/o tone-mapping curve)" + }, + { + "type": "image", + "bbox": [ + 0.503, + 0.322, + 0.642, + 0.394 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.509, + 0.395, + 0.639, + 0.405 + ], + "angle": 0, + "content": "Ours (w/o env. fusion)" + }, + { + "type": "image_caption", + "bbox": [ + 0.411, + 0.406, + 0.45, + 0.416 + ], + "angle": 0, + "content": "curve)" + }, + { + "type": "image_caption", + "bbox": [ + 0.314, + 0.422, + 0.687, + 0.436 + ], + "angle": 0, + "content": "Fig. 6: Qualitative ablation study of our design choices." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.45, + 0.785, + 0.495 + ], + "angle": 0, + "content": "via SDEdit [41], denoted as Ours (dataset update). Although this approach is adopted in text-guided scene editing [19], we found it challenging to produce stable gradients from such a \"discrete\" guidance." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.496, + 0.785, + 0.602 + ], + "angle": 0, + "content": "When designing the light and tone representations, we observe that the trainable tone-mapping curve provides additional flexibility to compensate for unknown tone-mapping in the input image, and better match the color and scale of the shadows. The two environment map fusion scheme (Sec. 4.3) allows to robustly recover the high intensity peaks in the early phase of training, and converge to better quality. Their respective effect is ablated as \"Ours (w/o tone curve)\" and \"Ours (w/o env. map fusion)\"." + }, + { + "type": "title", + "bbox": [ + 0.216, + 0.621, + 0.376, + 0.64 + ], + "angle": 0, + "content": "6 Applications" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.647, + 0.785, + 0.708 + ], + "angle": 0, + "content": "Since our method recovers physically based lighting information, arbitrary new virtual objects can be inserted after the optimization, as shown in Fig. 2. DiPIR can also optimize other scene attributes such as materials and local lighting. We perform preliminary experiments in this direction." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.716, + 0.787, + 0.853 + ], + "angle": 0, + "content": "Material optimization. When combined with differentiable rendering, DMs can provide a guidance signal for material attributes, as shown in Fig. 7. Given a purely diffuse car and enabling Metallic and Roughness properties as estimizable parameters, the diffusion guidance can optimize and make the car look more shiny. By additionally changing the text prompt to \"a carmine red car\" and making the base color of the car an estimizable parameter, we show that the DM can propagate the text-condition to the PBR attribute and change the color of the car to red. When enabling local emission as an estimizable parameter, the diffusion model can also turn on the headlights of cars in night scenes." + } + ], + [ + { + "type": "page_number", + "bbox": [ + 0.218, + 0.116, + 0.236, + 0.127 + ], + "angle": 0, + "content": "14" + }, + { + "type": "header", + "bbox": [ + 0.272, + 0.115, + 0.371, + 0.129 + ], + "angle": 0, + "content": "R. Liang et al." + }, + { + "type": "image", + "bbox": [ + 0.218, + 0.145, + 0.362, + 0.232 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.364, + 0.146, + 0.508, + 0.232 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.51, + 0.146, + 0.653, + 0.232 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.35, + 0.235, + 0.521, + 0.248 + ], + "angle": 0, + "content": "Optimize tone mapping curve" + }, + { + "type": "image", + "bbox": [ + 0.664, + 0.145, + 0.787, + 0.232 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.711, + 0.23, + 0.742, + 0.253 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.218, + 0.253, + 0.362, + 0.339 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.364, + 0.253, + 0.508, + 0.339 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.51, + 0.253, + 0.653, + 0.339 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.365, + 0.342, + 0.506, + 0.355 + ], + "angle": 0, + "content": "Optimize asset materials" + }, + { + "type": "image", + "bbox": [ + 0.664, + 0.253, + 0.787, + 0.339 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.665, + 0.342, + 0.78, + 0.355 + ], + "angle": 0, + "content": "Optimize emissions" + }, + { + "type": "image_caption", + "bbox": [ + 0.214, + 0.358, + 0.788, + 0.387 + ], + "angle": 0, + "content": "Fig. 7: Our physically based inverse rendering pipeline unlocks further applications such as material, local emission and tone-mapping refinement." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.405, + 0.79, + 0.482 + ], + "angle": 0, + "content": "Tone-mapping adjustment. We use a controlled experiment to further evaluate how well DMs understand tone-mapping. In Fig. 7, we freeze the estimated environment map and apply a manual tone adjustment on the background image. The diffusion guidance is used to optimize the tone curve such that the inserted object matches the surrounding background in the final composited result." + }, + { + "type": "title", + "bbox": [ + 0.216, + 0.492, + 0.356, + 0.508 + ], + "angle": 0, + "content": "7 Discussion" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.516, + 0.789, + 0.668 + ], + "angle": 0, + "content": "Our method leverages large diffusion models' inherent scene understanding capabilities as guidance to a physically based inverse rendering pipeline. We design a diffusion guidance signal with scene-specific personalization and a differentiable inverse rendering pipeline to recover lighting and tone-mapping parameters. Our method enables inserting virtual objects into scenes, but also optimizing other scene parameters such as the materials of the inserted object or account for tone-mapping mismatches between cameras. We believe that this combination of the differentiable rendering process and data-driven priors can be used successfully in many other content creation applications such as relighting and animation." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.674, + 0.79, + 0.842 + ], + "angle": 0, + "content": "Limitations and future work. Our Spherical Gaussians-based lighting representation is adequate for general objects [35], but might not behave realistically for highly specular materials. For more complex lighting representations adding generative priors on the environment map [40] is a direction worth exploring. The rendering formulation could be extended to account for effects such as reflections from the scene itself onto the inserted object (e.g. color bleeding), but that might introduce more ambiguities and require knowing the materials of the proxy geometry (refer to Supplement C.4 for failure case examples). Finally, while DM personalization significantly improves the quality of the results, it adds overhead and complexity to the pipeline. Recent personalization methods that do not require test-time finetuning [52] could be used to mitigate this overhead." + } + ], + [ + { + "type": "header", + "bbox": [ + 0.25, + 0.115, + 0.732, + 0.129 + ], + "angle": 0, + "content": "Photorealistic Object Insertion with Diffusion-Guided Inverse Rendering" + }, + { + "type": "page_number", + "bbox": [ + 0.768, + 0.117, + 0.785, + 0.127 + ], + "angle": 0, + "content": "15" + }, + { + "type": "title", + "bbox": [ + 0.218, + 0.145, + 0.403, + 0.163 + ], + "angle": 0, + "content": "Acknowledgements" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.179, + 0.788, + 0.239 + ], + "angle": 0, + "content": "The authors are grateful for the feedback received from Nicholas Sharp and Huan Ling during the project. We thank the original artists of the 3D assets used in this work: inciprocal, peyman.khaleghi, Kuutti Siitonen, TurboSquid and their artists Hum3D and Amaranthus." + }, + { + "type": "title", + "bbox": [ + 0.217, + 0.264, + 0.324, + 0.279 + ], + "angle": 0, + "content": "References" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.296, + 0.7, + 0.31 + ], + "angle": 0, + "content": "1. Poly Haven - The Public 3D Asset Library, https://polyhaven.com" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.311, + 0.787, + 0.365 + ], + "angle": 0, + "content": "2. Balaji, Y., Nah, S., Huang, X., Vahdat, A., Song, J., Zhang, Q., Kreis, K., Aittala, M., Aila, T., Laine, S., Catanzaro, B., Karras, T., Liu, M.Y.: ediff-i: Text-to-image diffusion models with ensemble of expert denoisers. arXiv preprint arXiv:2211.01324 (2022)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.367, + 0.787, + 0.407 + ], + "angle": 0, + "content": "3. Bangaru, S.P., Li, T.M., Durand, F.: Unbiased warped-area sampling for differentiable rendering. ACM Trans. Graph. 39(6) (nov 2020). https://doi.org/10.1145/3414685.3417833, https://doi.org/10.1145/3414685.3417833" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.409, + 0.787, + 0.435 + ], + "angle": 0, + "content": "4. Barron, J.T., Malik, J.: Shape, illumination, and reflectance from shading. IEEE transactions on pattern analysis and machine intelligence 37(8), 1670-1687 (2014)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.437, + 0.786, + 0.464 + ], + "angle": 0, + "content": "5. Barrow, H., Tenenbaum, J., Hanson, A., Riseman, E.: Recovering intrinsic scene characteristics. Comput. Vis. Syst 2, 3-26 (1978)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.465, + 0.786, + 0.492 + ], + "angle": 0, + "content": "6. Bell, S., Bala, K., Snavely, N.: Intrinsic images in the wild. ACM Transactions on Graphics (TOG) 33(4), 159 (2014)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.493, + 0.787, + 0.533 + ], + "angle": 0, + "content": "7. Black, M.J., Anandan, P.: The robust estimation of multiple motions: Parametric and piecewise-smooth flow fields. Computer vision and image understanding 63(1), 75-104 (1996)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.535, + 0.787, + 0.561 + ], + "angle": 0, + "content": "8. Boss, M., Jampani, V., Kim, K., Lensch, H.P., Kautz, J.: Two-shot spatially-varying brdf and shape estimation. In: CVPR (2020)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.563, + 0.787, + 0.589 + ], + "angle": 0, + "content": "9. Bousseau, A., Paris, S., Durand, F.: User-assisted intrinsic images. In: ACM Transactions on Graphics (TOG). vol. 28, p. 130. ACM (2009)" + }, + { + "type": "ref_text", + "bbox": [ + 0.22, + 0.591, + 0.787, + 0.631 + ], + "angle": 0, + "content": "10. Chari, P., Ma, S., Ostashev, D., Kadambi, A., Krishnan, G., Wang, J., Aberman, K.: Personalized restoration via dual-pivot tuning. arXiv preprint arXiv:2312.17234 (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.22, + 0.633, + 0.787, + 0.673 + ], + "angle": 0, + "content": "11. Dai, X., Hou, J., Ma, C.Y., Tsai, S., Wang, J., Wang, R., Zhang, P., Vandenhende, S., Wang, X., Dubey, A., et al.: Emu: Enhancing image generation models using photogenic needles in a haystack. arXiv preprint arXiv:2309.15807 (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.22, + 0.674, + 0.787, + 0.728 + ], + "angle": 0, + "content": "12. Dastjerdi, M.R.K., Eisenmann, J., Hold-Geoffroy, Y., Lalonde, J.F.: Everlight: Indoor-outdoor editable hdr lighting estimation. In: Proceedings of the IEEE/CVF International Conference on Computer Vision (ICCV). pp. 7420-7429 (October 2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.22, + 0.729, + 0.787, + 0.756 + ], + "angle": 0, + "content": "13. Durkan, C., Bekasov, A., Murray, I., Papamakarios, G.: Neural spline flows. Advances in neural information processing systems 32 (2019)" + }, + { + "type": "ref_text", + "bbox": [ + 0.22, + 0.758, + 0.787, + 0.811 + ], + "angle": 0, + "content": "14. Gal, R., Alaluf, Y., Atzmon, Y., Patashnik, O., Bermano, A.H., Chechik, G., Cohen-Or, D.: An image is worth one word: Personalizing text-to-image generation using textual inversion (2022). https://doi.org/10.48550/ARXIV.2208.01618, https://arxiv.org/abs/2208.01618" + }, + { + "type": "ref_text", + "bbox": [ + 0.22, + 0.813, + 0.787, + 0.84 + ], + "angle": 0, + "content": "15. Gardner, M.A., Hold-Geoffroy, Y., Sunkavalli, K., Gagné, C., Lalonde, J.F.: Deep parametric indoor lighting estimation. In: ICCV. pp. 7175-7183 (2019)" + }, + { + "type": "list", + "bbox": [ + 0.22, + 0.296, + 0.787, + 0.84 + ], + "angle": 0, + "content": null + } + ], + [ + { + "type": "page_number", + "bbox": [ + 0.218, + 0.116, + 0.236, + 0.127 + ], + "angle": 0, + "content": "16" + }, + { + "type": "header", + "bbox": [ + 0.272, + 0.115, + 0.37, + 0.128 + ], + "angle": 0, + "content": "R. Liang et al." + }, + { + "type": "ref_text", + "bbox": [ + 0.218, + 0.147, + 0.787, + 0.189 + ], + "angle": 0, + "content": "16. Gardner, M.A., Sunkavalli, K., Yumer, E., Shen, X., Gambaretto, E., Gagné, C., Lalonde, J.F.: Learning to predict indoor illumination from a single image. arXiv preprint arXiv:1704.00090 (2017)" + }, + { + "type": "ref_text", + "bbox": [ + 0.218, + 0.189, + 0.787, + 0.217 + ], + "angle": 0, + "content": "17. Garon, M., Sunkavalli, K., Hadap, S., Carr, N., Lalonde, J.F.: Fast spatially-varying indoor lighting estimation. In: CVPR. pp. 6908-6917 (2019)" + }, + { + "type": "ref_text", + "bbox": [ + 0.218, + 0.217, + 0.787, + 0.257 + ], + "angle": 0, + "content": "18. Grosse, R., Johnson, M.K., Adelson, E.H., Freeman, W.T.: Ground truth dataset and baseline evaluations for intrinsic image algorithms. In: ICCV. pp. 2335-2342. IEEE (2009)" + }, + { + "type": "ref_text", + "bbox": [ + 0.218, + 0.257, + 0.787, + 0.298 + ], + "angle": 0, + "content": "19. Haque, A., Tancik, M., Efros, A., Holynski, A., Kanazawa, A.: Instruct-nerf2nerf: Editing 3d scenes with instructions. In: Proceedings of the IEEE/CVF International Conference on Computer Vision (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.298, + 0.787, + 0.325 + ], + "angle": 0, + "content": "20. Ho, J., Jain, A., Abbeel, P.: Denoising diffusion probabilistic models. arXiv preprint arxiv:2006.11239 (2020)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.325, + 0.787, + 0.352 + ], + "angle": 0, + "content": "21. Ho, J., Salimans, T.: Classifier-free diffusion guidance. arXiv preprint arXiv:2207.12598 (2022)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.352, + 0.787, + 0.379 + ], + "angle": 0, + "content": "22. Hold-Geoffroy, Y., Athawale, A., Lalonde, J.F.: Deep sky modeling for single image outdoor lighting estimation. In: CVPR. pp. 6927-6935 (2019)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.379, + 0.787, + 0.406 + ], + "angle": 0, + "content": "23. Hold-Geoffroy, Y., Sunkavalli, K., Hadap, S., Gambaretto, E., Lalonde, J.F.: Deep outdoor illumination estimation. In: CVPR. pp. 7312-7321 (2017)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.406, + 0.787, + 0.459 + ], + "angle": 0, + "content": "24. Hu, E.J., yelong shen, Wallis, P., Allen-Zhu, Z., Li, Y., Wang, S., Wang, L., Chen, W.: LoRA: Low-rank adaptation of large language models. In: International Conference on Learning Representations (2022), https://openreview.net/forum?id=nZeVKeeFYf9" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.459, + 0.787, + 0.502 + ], + "angle": 0, + "content": "25. Jakob, W., Speierer, S., Roussel, N., Nimier-David, M., Vicini, D., Zeltner, T., Nicolet, B., Crespo, M., Leroy, V., Zhang, Z.: Mitsuba 3 renderer (2022), https://mitsubarenderer.org" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.502, + 0.787, + 0.529 + ], + "angle": 0, + "content": "26. Kajiya, J.T.: The rendering equation. In: Proceedings of the 13th annual conference on Computer graphics and interactive techniques. pp. 143-150 (1986)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.529, + 0.787, + 0.569 + ], + "angle": 0, + "content": "27. Karimi Dastjerdi, M.R., Hold-Geoffroy, Y., Eisenmann, J., Khodadadeh, S., Lalonde, J.F.: Guided co-modulated gan for 360 field of view extrapolation. International Conference on 3D Vision (3DV) (2022)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.569, + 0.787, + 0.624 + ], + "angle": 0, + "content": "28. Ke, B., Obukhov, A., Huang, S., Metzger, N., Daudt, R.C., Schindler, K.: Repurposing diffusion-based image generators for monocular depth estimation. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR) (2024)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.624, + 0.787, + 0.651 + ], + "angle": 0, + "content": "29. Kocsis, P., Sitzmann, V., Nießner, M.: Intrinsic image diffusion for single-view material estimation. In: arxiv (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.651, + 0.787, + 0.678 + ], + "angle": 0, + "content": "30. Kovacs, B., Bell, S., Snavely, N., Bala, K.: Shading annotations in the wild. In: CVPR. pp. 6998-7007 (2017)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.678, + 0.787, + 0.691 + ], + "angle": 0, + "content": "31. Land, E.H., McCann, J.J.: Lightness and retina theory. Josa 61(1), 1-11 (1971)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.691, + 0.787, + 0.732 + ], + "angle": 0, + "content": "32. LeGendre, C., Ma, W.C., Fyffe, G., Flynn, J., Charbonnel, L., Busch, J., Debevec, P.: Deeplight: Learning illumination for unconstrained mobile mixed reality. In: CVPR. pp. 5918-5928 (2019)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.732, + 0.787, + 0.772 + ], + "angle": 0, + "content": "33. Li, T.M., Aittala, M., Durand, F., Lehtinen, J.: Differentiable monte carlo ray tracing through edge sampling. ACM Trans. Graph. 37(6) (dec 2018). https://doi.org/10.1145/3272127.3275109, https://doi.org/10.1145/3272127.3275109" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.772, + 0.787, + 0.8 + ], + "angle": 0, + "content": "34. Li, Z., Snavely, N.: Cginintrinsics: Better intrinsic image decomposition through physically-based rendering. In: ECCV. pp. 371-387 (2018)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.8, + 0.787, + 0.84 + ], + "angle": 0, + "content": "35. Li, Z., Shafiei, M., Ramamoorthi, R., Sunkavalli, K., Chandraker, M.: Inverse rendering for complex indoor scenes: Shape, spatially-varying lighting and svbrdf from a single image. In: CVPR. pp. 2475-2484 (2020)" + }, + { + "type": "list", + "bbox": [ + 0.217, + 0.147, + 0.787, + 0.84 + ], + "angle": 0, + "content": null + } + ], + [ + { + "type": "header", + "bbox": [ + 0.25, + 0.115, + 0.732, + 0.129 + ], + "angle": 0, + "content": "Photorealistic Object Insertion with Diffusion-Guided Inverse Rendering" + }, + { + "type": "page_number", + "bbox": [ + 0.768, + 0.116, + 0.786, + 0.127 + ], + "angle": 0, + "content": "17" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.147, + 0.785, + 0.189 + ], + "angle": 0, + "content": "36. Li, Z., Xu, Z., Ramamoorthi, R., Sunkavalli, K., Chandraker, M.: Learning to reconstruct shape and spatially-varying reflectance from a single image. ACM Transactions on Graphics (TOG) 37(6), 1-11 (2018)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.191, + 0.787, + 0.232 + ], + "angle": 0, + "content": "37. Li, Z., Yu, L., Okunev, M., Chandraker, M., Dong, Z.: Spatiotemporally consistent hdr indoor lighting estimation. ACM Trans. Graph. 42(3) (jun 2023). https://doi.org/10.1145/3595921, https://doi.org/10.1145/3595921" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.233, + 0.785, + 0.274 + ], + "angle": 0, + "content": "38. Li, Z., Yu, T.W., Sang, S., Wang, S., Bi, S., Xu, Z., Yu, H.X., Sunkavalli, K., Hasan, M., Ramamoorthi, R., et al.: Openrooms: An end-to-end open framework for photorealistic indoor scene datasets. arXiv preprint arXiv:2007.12868 (2020)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.276, + 0.787, + 0.317 + ], + "angle": 0, + "content": "39. Loubet, G., Holzschuch, N., Jakob, W.: Reparameterizing discontinuous integrands for differentiable rendering. ACM Trans. Graph. 38(6) (nov 2019). https://doi.org/10.1145/3355089.3356510, https://doi.org/10.1145/3355089.3356510" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.318, + 0.787, + 0.358 + ], + "angle": 0, + "content": "40. Lyu, L., Tewari, A., Habermann, M., Saito, S., Zollhöfer, M., Leimkuehler, T., Theobalt, C.: Diffusion posterior illumination for ambiguity-aware inverse rendering. ACM Transactions on Graphics 42(6) (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.36, + 0.785, + 0.401 + ], + "angle": 0, + "content": "41. Meng, C., He, Y., Song, Y., Song, J., Wu, J., Zhu, J.Y., Ermon, S.: SDEdit: Guided image synthesis and editing with stochastic differential equations. In: International Conference on Learning Representations (2022)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.402, + 0.787, + 0.457 + ], + "angle": 0, + "content": "42. Nimier-David, M., Speierer, S., Ruiz, B., Jakob, W.: Radiative backpropagation: an adjoint method for lightning-fast differentiable rendering. ACM Trans. Graph. 39(4) (aug 2020). https://doi.org/10.1145/3386569.3392406, https://doi.org/10.1145/3386569.3392406" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.459, + 0.787, + 0.5 + ], + "angle": 0, + "content": "43. Nimier-David, M., Vicini, D., Zeltner, T., Jakob, W.: Mitsuba 2: a retargetable forward and inverse renderer. ACM Trans. Graph. 38(6) (nov 2019). https://doi.org/10.1145/3355089.3356498, https://doi.org/10.1145/3355089.3356498" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.501, + 0.787, + 0.542 + ], + "angle": 0, + "content": "44. Phongthawee, P., Chinchuthakun, W., Sinsunthithet, N., Raj, A., Jampani, V., Khungurn, P., Suwajanakorn, S.: Diffusionlight: Light probes for free by painting a chrome ball. In: ArXiv (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.543, + 0.785, + 0.571 + ], + "angle": 0, + "content": "45. Poole, B., Jain, A., Barron, J.T., Mildenhall, B.: Dreamfusion: Text-to-3d using 2d diffusion. arXiv (2022)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.572, + 0.785, + 0.599 + ], + "angle": 0, + "content": "46. Rombach, R., Blattmann, A., Lorenz, D., Esser, P., Ommer, B.: High-resolution image synthesis with latent diffusion models (2021)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.601, + 0.787, + 0.627 + ], + "angle": 0, + "content": "47. Ruiz, N., Li, Y., Jampani, V., Pritch, Y., Rubinstein, M., Aberman, K.: Dreambooth: Fine tuning text-to-image diffusion models for subject-driven generation (2022)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.629, + 0.787, + 0.684 + ], + "angle": 0, + "content": "48. Sahara, C., Chan, W., Saxena, S., Li, L., Whang, J., Denton, E.L., Ghasemipour, K., Gontijo Lopes, R., Karagol Ayan, B., Salimans, T., et al.: Photorealistic text-to-image diffusion models with deep language understanding. Advances in Neural Information Processing Systems 35, 36479-36494 (2022)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.686, + 0.785, + 0.726 + ], + "angle": 0, + "content": "49. Sarkar, A., Mai, H., Mahapatra, A., Lazebnik, S., Forsyth, D., Bhattad, A.: Shadows don't lie and lines can't bend! generative models don't know projective geometry...for now (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.728, + 0.785, + 0.754 + ], + "angle": 0, + "content": "50. Sengupta, S., Gu, J., Kim, K., Liu, G., Jacobs, D.W., Kautz, J.: Neural inverse rendering of an indoor scene from a single image. In: ICCV (2019)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.756, + 0.785, + 0.783 + ], + "angle": 0, + "content": "51. Shah, V., Ruiz, N., Cole, F., Lu, E., Lazebnik, S., Li, Y., Jampani, V.:Ziplora: Any subject in any style by effectively merging loras (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.785, + 0.785, + 0.812 + ], + "angle": 0, + "content": "52. Shi, J., Xiong, W., Lin, Z., Jung, H.J.: Instantbooth: Personalized text-to-image generation without test-time finetuning. arXiv preprint arXiv:2304.03411 (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.813, + 0.785, + 0.84 + ], + "angle": 0, + "content": "53. Song, S., Funkhouser, T.: Neural illumination: Lighting prediction for indoor environments. In: CVPR. pp. 6918-6926 (2019)" + }, + { + "type": "list", + "bbox": [ + 0.217, + 0.147, + 0.787, + 0.84 + ], + "angle": 0, + "content": null + } + ], + [ + { + "type": "page_number", + "bbox": [ + 0.218, + 0.116, + 0.236, + 0.127 + ], + "angle": 0, + "content": "18" + }, + { + "type": "header", + "bbox": [ + 0.272, + 0.115, + 0.371, + 0.129 + ], + "angle": 0, + "content": "R. Liang et al." + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.147, + 0.788, + 0.189 + ], + "angle": 0, + "content": "54. Srinivasan, P.P., Mildenhall, B., Tancik, M., Barron, J.T., Tucker, R., Snavely, N.: Lighthouse: Predicting lighting volumes for spatially-coherent illumination. In: CVPR. pp. 8080-8089 (2020)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.19, + 0.788, + 0.232 + ], + "angle": 0, + "content": "55. Sun, P., Kretzschmar, H., Dotiwalla, X., Chouard, A., Patnaik, V., Tsui, P., Guo, J., Zhou, Y., Chai, Y., Caine, B., et al.: Scalability in perception for autonomous driving: Waymo open dataset. In: CVPR (2020)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.233, + 0.787, + 0.261 + ], + "angle": 0, + "content": "56. Tang, J., Zhu, Y., Wang, H., Chan, J.H., Li, S., Shi, B.: Estimating spatially-varying lighting in urban scenes with disentangled representation. In: ECCV (2022)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.262, + 0.787, + 0.303 + ], + "angle": 0, + "content": "57. Tang, L., Ruiz, N., Qinghao, C., Li, Y., Holynski, A., Jacobs, D.E., Hariharan, B., Pritch, Y., Wadhwa, N., Aberman, K., Rubinstein, M.: Realfill: Reference-driven generation for authentic image completion. arXiv preprint arXiv:2309.16668 (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.304, + 0.787, + 0.332 + ], + "angle": 0, + "content": "58. Tao, A., Sapra, K., Catanzaro, B.: Hierarchical multi-scale attention for semantic segmentation. arXiv preprint arXiv:2005.10821 (2020)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.333, + 0.787, + 0.374 + ], + "angle": 0, + "content": "59. Veach, E., Guibas, L.J.: Optimally combining sampling techniques for monte carlo rendering. In: Proceedings of the 22nd annual conference on Computer graphics and interactive techniques. pp. 419-428 (1995)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.375, + 0.787, + 0.429 + ], + "angle": 0, + "content": "60. Vicini, D., Speierer, S., Jakob, W.: Path replay backpropagation: differentiating light paths using constant memory and linear time. ACM Trans. Graph. 40(4) (jul 2021). https://doi.org/10.1145/3450626.3459804, https://doi.org/10.1145/3450626.3459804" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.43, + 0.787, + 0.472 + ], + "angle": 0, + "content": "61. Wang, G., Yang, Y., Loy, C.C., Liu, Z.: Stylelight: HDR panorama generation for lighting estimation and editing. In: European Conference on Computer Vision (ECCV) (2022)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.473, + 0.787, + 0.502 + ], + "angle": 0, + "content": "62. Wang, Z., Chen, W., Acuna, D., Kautz, J., Fidler, S.: Neural light field estimation for street scenes with differentiable virtual object insertion. In: ECCV (2022)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.502, + 0.787, + 0.529 + ], + "angle": 0, + "content": "63. Wang, Z., Philion, J., Fidler, S., Kautz, J.: Learning indoor inverse rendering with 3d spatially-varying lighting. In: ICCV (2021)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.53, + 0.787, + 0.558 + ], + "angle": 0, + "content": "64. Wimbauer, F., Wu, S., Rupprecht, C.: De-rendering 3d objects in the wild. In: CVPR (2022)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.559, + 0.787, + 0.613 + ], + "angle": 0, + "content": "65. Yan, K., Lassner, C., Budge, B., Dong, Z., Zhao, S.: Efficient estimation of boundary integrals for path-space differentiable rendering. ACM Trans. Graph. 41(4) (jul 2022). https://doi.org/10.1145/3528223.3530080, https://doi.org/10.1145/3528223.3530080" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.614, + 0.787, + 0.657 + ], + "angle": 0, + "content": "66. Yang, J., Ivanovic, B., Litany, O., Weng, X., Kim, S.W., Li, B., Che, T., Xu, D., Fidler, S., Pavone, M., Wang, Y.: Emernerf: Emergent spatial-temporal scene decomposition via self-supervision. arXiv preprint arXiv:2311.02077 (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.657, + 0.787, + 0.699 + ], + "angle": 0, + "content": "67. Yu, H.X., Agarwala, S., Herrmann, C., Szeliski, R., Snavely, N., Wu, J., Sun, D.: Accidental light probes. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 12521-12530 (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.7, + 0.787, + 0.727 + ], + "angle": 0, + "content": "68. Yu, X., Guo, Y.C., Li, Y., Liang, D., Zhang, S.H., Qi, X.: Text-to-3d with classifier score distillation (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.728, + 0.787, + 0.755 + ], + "angle": 0, + "content": "69. Yu, Y., Smith, W.A.: Inverserendernet: Learning single image inverse rendering. In: CVPR (2019)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.756, + 0.787, + 0.798 + ], + "angle": 0, + "content": "70. Zhan, F., Zhang, C., Yu, Y., Chang, Y., Lu, S., Ma, F., Xie, X.: Emlight: Lighting estimation via spherical distribution approximation. In: Proceedings of the AAAI Conference on Artificial Intelligence (2021)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.799, + 0.787, + 0.84 + ], + "angle": 0, + "content": "71. Zhang, C., Miller, B., Yan, K., Gkioulekas, I., Zhao, S.: Path-space differentiable rendering. ACM Trans. Graph. 39(4) (aug 2020). https://doi.org/10.1145/3386569.3392383, https://doi.org/10.1145/3386569.3392383" + }, + { + "type": "list", + "bbox": [ + 0.217, + 0.147, + 0.788, + 0.84 + ], + "angle": 0, + "content": null + } + ], + [ + { + "type": "header", + "bbox": [ + 0.25, + 0.115, + 0.732, + 0.129 + ], + "angle": 0, + "content": "Photorealistic Object Insertion with Diffusion-Guided Inverse Rendering" + }, + { + "type": "page_number", + "bbox": [ + 0.768, + 0.116, + 0.786, + 0.127 + ], + "angle": 0, + "content": "19" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.147, + 0.788, + 0.189 + ], + "angle": 0, + "content": "72. Zhang, C., Yu, Z., Zhao, S.: Path-space differentiable rendering of participating media. ACM Trans. Graph. 40(4) (jul 2021). https://doi.org/10.1145/3450626.3459782, https://doi.org/10.1145/3450626.3459782" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.19, + 0.787, + 0.232 + ], + "angle": 0, + "content": "73. Zhang, R., Isola, P., Efros, A.A., Shechtman, E., Wang, O.: The unreasonable effectiveness of deep features as a perceptual metric. In: Proceedings of the IEEE conference on computer vision and pattern recognition. pp. 586-595 (2018)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.232, + 0.787, + 0.272 + ], + "angle": 0, + "content": "74. Zhang, Z., Roussel, N., Jakob, W.: Projective sampling for differentiable rendering of geometry. ACM Trans. Graph. 42(6) (dec 2023). https://doi.org/10.1145/3618385, https://doi.org/10.1145/3618385" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.273, + 0.787, + 0.3 + ], + "angle": 0, + "content": "75. Zhao, Q., Tan, P., Dai, Q., Shen, L., Wu, E., Lin, S.: A closed-form solution to retinax with nonlocal texture constraints 34(7), 1437-1444 (2012)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.3, + 0.787, + 0.328 + ], + "angle": 0, + "content": "76. Zhao, Y., Guo, T.: Pointar: Efficient lighting estimation for mobile augmented reality. arXiv preprint arXiv:2004.00006 (2020)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.328, + 0.787, + 0.356 + ], + "angle": 0, + "content": "77. Zhu, Y., Zhang, Y., Li, S., Shi, B.: Spatially-varying outdoor lighting estimation from intrinsics. In: CVPR (2021)" + }, + { + "type": "list", + "bbox": [ + 0.217, + 0.147, + 0.788, + 0.356 + ], + "angle": 0, + "content": null + } + ] +] \ No newline at end of file diff --git a/2024/Photorealistic Object Insertion with Diffusion-Guided Inverse Rendering/27eaedd8-53d6-44df-a63a-fb92723dff8a_origin.pdf b/2024/Photorealistic Object Insertion with Diffusion-Guided Inverse Rendering/27eaedd8-53d6-44df-a63a-fb92723dff8a_origin.pdf new file mode 100644 index 0000000000000000000000000000000000000000..ba4c6e57ddd90effd640773c99961f1e7109a22e --- /dev/null +++ b/2024/Photorealistic Object Insertion with Diffusion-Guided Inverse Rendering/27eaedd8-53d6-44df-a63a-fb92723dff8a_origin.pdf @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:6ecc3ec26c1757582ae76a632d2aba71b3588f1ed12a49da61e22a4edebd461b +size 9150369 diff --git a/2024/Photorealistic Object Insertion with Diffusion-Guided Inverse Rendering/full.md b/2024/Photorealistic Object Insertion with Diffusion-Guided Inverse Rendering/full.md new file mode 100644 index 0000000000000000000000000000000000000000..a522fa22f7f18fb7433abe9800210c67a30af5b8 --- /dev/null +++ b/2024/Photorealistic Object Insertion with Diffusion-Guided Inverse Rendering/full.md @@ -0,0 +1,410 @@ +# Photorealistic Object Insertion with Diffusion-Guided Inverse Rendering + +Ruofan Liang $^{1,2,3}$ , Zan Gojcic $^{1}$ , Merlin Nimier-David $^{1}$ , David Acuna $^{1}$ , Nandita Vijaykumar $^{2,3}$ , Sanja Fidler $^{1,2,3}$ , and Zian Wang $^{1,2,3}$ + +$^{1}$ NVIDIA $^{2}$ University of Toronto $^{3}$ Vector Institute + +Abstract. The correct insertion of virtual objects in images of real-world scenes requires a deep understanding of the scene's lighting, geometry and materials, as well as the image formation process. While recent large-scale diffusion models have shown strong generative and inpainting capabilities, we find that current models do not sufficiently "understand" the scene shown in a single picture to generate consistent lighting effects (shadows, bright reflections, etc.) while preserving the identity and details of the composited object. We propose using a personalized large diffusion model as guidance to a physically based inverse rendering process. Our method recovers scene lighting and tone-mapping parameters, allowing the photorealistic composition of arbitrary virtual objects in single frames or videos of indoor or outdoor scenes. Our physically based pipeline further enables automatic materials and tone-mapping refinement. + +Keywords: Inverse rendering $\cdot$ Diffusion models $\cdot$ Personalization $\cdot$ Virtual object insertion $\cdot$ Physically based rendering + +# 1 Introduction + +Virtual object insertion enables a range of applications from virtual production, to interactive gaming and synthetic data generation. To produce photorealistic insertions, the interactions between the virtual objects and the environment need to be modeled faithfully, such as accurate specular highlights and shadows. + +A standard virtual object insertion pipeline typically includes three key steps: i) lighting estimation from the input image, ii) 3D proxy geometry creation, and iii) composited image rendering in a rendering engine. However, the first and arguably most important step is still an open research question. Lighting estimation is particularly challenging when dealing with limited inputs such as a single image, captured using a consumer device with a low dynamic range. Indeed, inverse rendering is a fundamentally ill-posed problem. + +To constrain its solution space, prior works either aimed to define hand-crafted priors [9,18,31,75] or to learn them from the data [15-17,22,23,32,36,56,62,63,77]. However, the former often fall short when applied to real-world scenes, while the latter suffer from scarcity of the ground truth data. As a result, these algorithms are often heavily tailored to a specific domain, e.g. indoor [15-17,50,63] or outdoor scenes [22,23,56,62,77]. + +![](images/eb9557c42ae5d42da10f49db59147e2ba8d2ce6694f2dfbe520005d7073f2e0e.jpg) +Virtual Object Insertion +A vase with dried flowers + +![](images/036064d38f10febd97136079be0dcbefb40515995ac8bf67bcf6b7592671a527.jpg) +A black SUV car + +![](images/06be440aeacb2afda4aa9bd81a386182d56eea84dc38e5aae79daaad1061b530.jpg) +A metal bucket +Fig.1: We propose DiPIR, a physically based method to recover lighting from a single image, enabling arbitrary virtual object compositing into indoor and outdoor scenes, as well as material and tone-mapping optimization. Project page: https://research.nvidia.com/labs/toronto-ai/DiPIR/ + +![](images/02d163547fd641c2d36da6ca0f8d4e05bdbd5c0d796c839ee7913b8a6cd206c8.jpg) +A white SUV car + +![](images/8efe3259dafa467b6a2b948151535c18b85b91cadd43cbfebe3ffdc3caf2a4e0.jpg) +Applications +Insert multiple objects + +![](images/036680c1483f31be9a9d096e0efa3ebb30ad26a8be0e2345049567543452e6ae.jpg) +Optimize material & tone mapping + +To address these challenges, we propose to reuse the strong image generation priors learned by large diffusion models (DMs) [46] as a guidance for inverse rendering. Unlike hand-crafted or supervised data-driven priors that are often specific to a domain, DMs are trained on massive datasets and show a remarkable "understanding" of the world and the underlying physical concepts. While DMs still often fail to produce accurate lighting effects such as shadows and reflections [49] in their generations, we observe that they can provide valuable guidance when combined with a physically-based renderer and are adapted to the scene. + +Specifically, we present Diffusion Prior for Inverse Rendering (DiPIR), building on three main contributions. First, we use a physically based renderer to accurately simulate the interaction between the light and the 3D asset to generate the final composited image. We also account for the unknown tone-mapping curve to mimic the camera sensor response. Second, we propose a lightweight personalization scheme of the pre-trained DM, based on the input image and the type of inserted asset. Third, we design a variant of the SDS loss [45] which makes use of this personalization and improves training stability. + +In DiPIR, the DM acts similarly to a human evaluator. It takes the edited image as input and propagates the feedback signal to physically-based scene attributes via differentiable rendering, thus enabling end-to-end optimization. We experimentally show that DiPIR outperforms existing state-of-the-art lighting estimation methods for object insertion across indoor and outdoor datasets. + +# 2 Related Work + +Inverse rendering is the task of recovering intrinsic properties of a scene, including materials, shape, and lighting, from a single or multiple images [5]. + +The main challenge of inverse rendering lies in the ill-posed nature of the task. Prior knowledge about materials and lighting effects is crucial to solving under-constrained inverse rendering problems, but how to best define and incorporate such priors remains an open research question. + +Early methods formulated inverse rendering as an optimization problem and used hand-crafted regularization terms such as local smoothness and sparsity of materials [4,5,9,18,31,75] to constrain the solution space. However, real-world lighting effects are often complex and difficult to describe using hand-crafted priors. The advent of deep learning has facilitated the learning of data-driven priors from ground truth supervision [6,8,30,34,35,38,50,62-64,69]. Yet, acquiring real-world data with accurately annotated intrinsic decomposition necessitates specialized capture devices [18] or expert annotators [6,30]. Collecting real-world datasets at a large scale is therefore extremely challenging. As a result, existing methods often resort to synthetic datasets [17,35,50,56,63,77], or are carefully designed to be training data-efficient [22,23,62]. The resulting algorithms are therefore often tailored to specific domains, such as indoor [15-17, 50, 63] or day-time outdoor scenes [22,23,56,62,77]. + +Lighting estimation is a subtask of inverse rendering that specifically focuses on inferring lighting and is a core component of photorealistic virtual object insertion pipelines. Existing lighting estimation methods are often designed as feed-forward neural networks that take a single image as input and directly regress lighting information in form of spherical lobes [17,35,70,76], environment maps [16,50,53,56,77], parametric light sources [15], or volumetric lighting [37, 54,62,63]. The environment lighting can also be recovered through the proxy geometry shown in the single image [32,67]. Recently, lighting estimation was also formulated as a generative task. StyleLight [61] is based on a dual-branch StyleGAN model, which first performs GAN inversion through the LDR branch and then generates the panorama with the HDR branch. EverLight [12] instead regresses an initial HDR environment map from the input image and later refines it with a GAN model [27]. Concurrent to our work, DiffusionLight [44] proposed to inpaint a chrome ball in the center of the given image using an inpainting diffusion model and unwarp it to an environment map. To obtain the HDR environment map, DiffusionLight trains a LoRA [24] for exposure bracketing and merges multiple generated LDR chrome balls with different exposures. + +Note that lighting is fundamentally a High Dynamic Range (HDR) quantity, and reproducing this full range still remains challenging. HDR is crucial for basic effects such as sharp shadows, particularly in outdoor day-time settings where the sun is several orders of magnitudes brighter than the rest of the environment. + +Physically-based differentiable rendering. Rendering algorithms such as path tracing [26] aim to produce realistic images of virtual scenes by accurately simulating the physical processes involved in light transport. Physically-based rendering (PBR) typically uses a geometric representation of the scene, scattering functions modeling the behavior of the scene's surfaces, as well as camera and lighting models. Recent years have seen the development of many differentiable rendering methods and frameworks [25, 33, 42, 43, 60, 71, 72]. + +While particular attention has been devoted to the difficult problem of computing derivatives at discontinuities [3, 33, 39, 65, 74], such as due to visibility changes, we focus here on the comparatively simpler lighting, material and tone-mapping derivatives, which generally do not introduce discontinuities. + +Diffusion model priors and personalization. Priors learned by the text-to-image DMs [2,11,46,48], on large-scale datasets can be adapted to various applications such as monocular depth estimation [28], authentic image completion [57], and image restoration [10]. Among various adaptation techniques, DreamBooth [47] proposes to finetune the DM on subject images, while Textual Inversion [14] optimizes new word embeddings for target generation. LoRA [24,51] simplifies adaptation by freezing the pretrained model weights and injecting trainable rank decomposition matrices, thus greatly reducing the number of trainable parameters. Most similar to ours are concurrent works on DM adaptation for lighting estimation [44] and intrinsic image decomposition [29]. The former adapts the DM to exposure bracketing for LDR to HDR lifting, while the latter fine-tunes a pretrained SD model to directly output the albedo and BRDF properties of an image. + +# 3 Preliminaries + +Diffusion Models. Diffusion models are a family of generative models built around two key processes. A forward process, which gradually adds noise to data samples $\mathbf{x} \sim p(\mathbf{x})$ removing their structure over time $t$ . This is achieved using a noise schedule determined by $\alpha_{t}$ and $\sigma_{t}$ as $\mathbf{x}_{t} = \alpha_{t}\mathbf{x} + \sigma_{t}\boldsymbol{\epsilon}, \boldsymbol{\epsilon} \sim \mathcal{N}(\mathbf{0},\mathbf{I})$ . In contrast, the reverse process gradually removes this noise, restoring the structure. The reverse process is parameterized by a conditional neural network, $\epsilon_{\theta}$ , trained to predict the noise $\epsilon$ at a given timestep $t$ according to the following simplified objective [20]: + +$$ +\mathbb {E} _ {\mathbf {x} \sim p (\mathbf {x}), \boldsymbol {\epsilon} \sim \mathcal {N} (\mathbf {0}, \boldsymbol {I}), t \sim T} \left[ w (t) | | \boldsymbol {\epsilon} _ {\boldsymbol {\theta}} (\mathbf {x} _ {t}, t, \boldsymbol {c}) - \boldsymbol {\epsilon} | | _ {2} ^ {2} \right], \tag {1} +$$ + +where $c$ represents a condition (e.g. text, image, etc.) that allows controlling the generation process, $w(t)$ represents a time-conditional weighting, and $T$ is a set containing a selection of timesteps. + +In this work, we use a Latent Diffusion Model (LDM) [46] in which the diffusion process is conducted in a lower-dimensional latent space. Specifically, the encoder $\mathcal{E}$ maps samples from the data distribution $\mathbf{x} \sim p(\mathbf{x})$ into a latent space $\mathbb{Z}$ . The decoder $\mathcal{D}$ performs the inverse operation, such that $\mathcal{D}(\mathcal{E}(\mathbf{x})) \approx \mathbf{x}$ . As follows, for LDMs $\mathbf{x}$ in Eq. 1 is replaced by its latent $\mathbf{z} = \mathcal{E}(\mathbf{x})$ . + +Personalization and Fine-tuning. Fine-tuning all parameters of a pretrained DM requires significant computational resources and time. To alleviate this, LoRA [24] injects trainable low-rank decomposition matrices and aims to learn only the variations from the pretrained weights. Specifically, consider a linear layer represented as $\pmb{h} = \pmb{W}_0\pmb{x}$ . Here $\pmb{W}_0 \in \mathbb{R}^{n \times n}$ and $\pmb{x} \in \mathbb{R}^n$ are the pretrained + +![](images/e37dbec6ed4ca66a42ced518975652ead35e15ea9dfb8476f76d5260b2b2cb5b.jpg) +Fig. 2: Method overview. Given an input image, we first construct a virtual 3D scene with a virtual object and proxy plane. Our physically-based renderer then differentiably simulates the interactions of the estimizable environment map with the inserted virtual object and its effect on the background scene (shadowing) (left). At each iteration, the rendered image is diffused and passed through a personalized diffusion model (middle). The gradient of the adapted Score Distillation formulation is propagated back to the environment map and the tone-mapping curve through the differentiable renderer. Upon convergence, we recover lighting and tone-mapping parameters, which allow photorealistic compositing of virtual objects from a single image (right). + +![](images/9642069bf67e8f1ea1bde726c288ba8dee1cad210dbc647aee5ec017040eb035.jpg) + +weights and input, respectively. Applying LoRA, this layer is modified to $\pmb{h} = (W_0 + \pmb{\Delta W})\pmb{x} = W_0\pmb{x} + AB\pmb{x}$ . Notably, $\pmb{\Delta W} = \pmb{A}\pmb{B}$ is a combination of low-rank matrices $\pmb{A} \in \mathbb{R}^{n \times r}$ and $\pmb{B} \in \mathbb{R}^{r \times n}$ with $r \ll n$ . During LoRA fine-tuning, only the added term $\pmb{\Delta W}$ is updated while the parameters $W_0$ remain unchanged. The rank $r$ is a hyperparameter that trades efficiency with model capacity. + +# 4 Method + +Given a single image as input, DiPIR recovers scene lighting and tone-mapping parameters, with the goal of photorealistic insertion of virtual objects. An overview of our method is shown in Fig. 2. Sec. 4.1 describes our representation and the differentiable rendering process, while Sec. 4.2 and Sec. 4.3 provide details on diffusion model guidance and optimization formulation, respectively. + +# 4.1 Physically-based Virtual Object Insertion + +Virtual scene. Inserting a virtual object $\mathcal{X}$ into an image $\mathbf{I}_{\mathrm{bg}} \in \mathbb{R}^{h \times w \times 3}$ requires creating a 3D proxy virtual scene, viewed from the correct camera pose. Here, we assume that the user provides a specific placement (pose) for $\mathcal{X}$ , but in some cases, an appropriate pose can also be determined automatically, e.g. by detecting the floor plane and scene scale. + +To model the effects of the inserted object on the original image, such as shadows cast by $\mathcal{X}$ , we also assume a known proxy geometry $\mathcal{P}$ . We found a simple ground plane acting as a shadow catcher underneath the virtual object to be sufficient in all of our experiments. This proxy plane can be easily placed manually or automatically generated based on e.g. depth or LiDAR data. + +Light representation. We represent the scene's lighting $\mathbf{L}$ with a set of $N$ optizable Spherical Gaussian (SG) parameters $\{\mathbf{c}_k,\pmb {\mu}_k,\sigma_k\}_{k = 1}^N\in \mathbb{R}^{N\times 7}$ , where the radiance for one SG lobe at the direction $\pmb {v}\in \mathbb{R}^3$ is defined as + +$$ +\mathbf {G} _ {k} (\pmb {v}; \mathbf {c} _ {k}, \pmb {\mu} _ {k}, \sigma_ {k}) = \mathbf {c} e ^ {- (1 - \pmb {v} \cdot \pmb {\mu}) / \sigma^ {2}}, \quad \mathrm {w h e r e} \mathbf {c} _ {k} \in \mathbb {R} ^ {3}, \pmb {\mu} _ {k} \in \mathbb {R} ^ {3}, \sigma_ {k} \in \mathbb {R} _ {+}. \quad (2) +$$ + +The overall environment map $\mathbf{L} \in \mathbb{R}^{H \times W \times 3}$ is computed as: + +$$ +\mathbf {L} _ {i, j} = \sum_ {k = 1} ^ {N} \mathbf {G} _ {k} \left(\boldsymbol {v} _ {i, j}; \mathbf {c} _ {k}, \boldsymbol {\mu} _ {k}, \sigma_ {k}\right), \tag {3} +$$ + +where $\pmb{v}_{i,j}$ is the direction corresponding to the pixel $(i,j)$ using the standard spherical environment parameterization. Note that we chose a SG-based lighting representation for its good convergence properties and simplicity, but many alternatives exist and might be combined with our method. + +Differentiable rendering. Inserting a virtual object into a scene involves simulating the interactions of the estimizable environment map with the inserted virtual object (foreground), and the inserted object's effect on the background scene (shadow). We describe below the details respectively. + +Foreground image. Given the optimized light representation $\mathbf{L}$ from above and the virtual object $\mathcal{X}$ (including its geometry and materials), a foreground image $\mathbf{I}_{\mathrm{fg}}$ of the inserted object can be rendered directly using standard path tracing: + +$$ +\mathbf {I} _ {\mathrm {f g}} = \operatorname {P a t h T r a c e} (\mathcal {X}, \mathbf {L}, D), \tag {4} +$$ + +$D$ is the maximum number of interactions along the light path. Since we do not assume that the provided proxy geometry is accurate nor that it has materials, we omit the effect of light reflected by the object into the scene, for example, due to a highly specular inserted object. + +Shadow ratio. As the background image $\mathbf{I}_{\mathrm{bg}}$ already faithfully represents the scene, our task is limited to simulating the appearance of the inserted object and its effect on the nearby parts of the scene, i.e. the shadows cast by the object onto the proxy geometry. + +The shadow ratio $\beta_{\mathrm{shadow}} \in \mathbb{R}^{h \times w \times 3}$ , which accounts for the effect of the object on the surrounding scene, is computed as the ratio between the radiance received by the proxy geometry with and without the inserted object present. A low value of $\beta_{\mathrm{shadow}}$ indicates a strongly shadowed region: + +$$ +\boldsymbol {\beta} _ {\mathrm {s h a d o w}} = \frac {\operatorname {P a t h T r a c e} (\mathcal {X} \cup \mathcal {P} , \mathbf {L} , 1)}{\operatorname {P a t h T r a c e} (\mathcal {P} , \mathbf {L} , 1)}, \tag {5} +$$ + +This is similar to the method of Wang et al. [62], however we use Multiple Importance Sampling (MIS) [59] between lighting and BSDF for better sampling efficiency. The maximum path length has been limited to 1 interaction in order to reduce memory and computational cost. Unless specific material information is provided with the proxy geometry, we use a Lambertian BSDF, in which case the albedo cancels out in the ratio. In practice, the computation of the shadow ratio and foreground image are combined in order to reuse shared terms. + +Tone-mapping. To compensate for the unknown tone-mapping of the input image, we introduces an estimable tone correction function $f(\cdot)$ that is applied on the inserted object $\hat{\mathbf{I}}_{\mathrm{fg}} = f(\mathbf{I}_{\mathrm{fg}}; \boldsymbol{\theta}_{\mathrm{fg}})$ and shadows $\hat{\beta}_{\mathrm{shadow}} = f(\boldsymbol{\beta}_{\mathrm{shadow}}; \boldsymbol{\theta}_{\mathrm{shadow}})$ . Our tone curve parameterization follows from [13] and employs a monotonic rational-quadratic spline as its basic building block. This spline is composed of $K_{\mathrm{s}}$ bins, each defined by the quotient of two quadratic polynomials. Notably, these functions are differentiable, and allow direct parameterization of the derivatives and heights at each knot. We apply a single spline for the foreground image, and different correction splines for each RGB channel for the shadow ratio to provide the flexibility to adjust the color of the shadow. We set $K_{\mathrm{s}} = 5$ for these curves. + +Differentiability. The final output image is an alpha-composite of the foreground object, shadows, and background image: + +$$ +\mathbf {I} _ {\mathrm {c o m p}} = \left(1 - \mathbf {V} (\mathcal {X})\right) \cdot \hat {\boldsymbol {\beta}} _ {\mathrm {s h a d o w}} \cdot \mathbf {I} _ {\mathrm {b g}} + \mathbf {V} (\mathcal {X}) \cdot \tilde {\mathbf {I}} _ {\mathrm {f g}}, \tag {6} +$$ + +where $\mathbf{V}(\mathcal{X})$ equals 1 when $\mathcal{X}$ is directly visible from the camera, and 0 otherwise. + +As neither foreground rendering, shadow rendering, or the compositing operation require derivatives w.r.t. the object placement or other discontinuous quantities, we can rely on automatic differentiation to obtain gradients of any pixelwise loss w.r.t. the lighting or material properties. To this end, we use the Path Replay Backpropagation [60] integrator of Mitsuba 3 [25]. The estimizable parameters are the Spherical Gaussian coefficients used in the lighting representation, and the tone curves' parameters $\theta_{\mathrm{fg}},\theta_{\mathrm{shadow}}$ + +While each of the operations above include simplifications (e.g. we do not account for secondary lighting from the proxy geometry), we have found them to be sufficient in practice: the remaining imperfections in the simulation can be sufficiently offset by the lighting optimization. + +# 4.2 Diffusion Guidance + +The composited image produced by our differentiable rendering pipeline serves as the input to a DM that is used to compute a guidance signal, employing an optimization objective similar to Score Distillation Sampling (SDS) [45]. However, while DMs inherently have robust priors for lighting, we found that they do not provide out-of-the-box the necessary guidance for our specific needs. Consequently, we propose an adaptive score distillation loss specifically designed for object insertion tasks that exploits a personalization strategy which we detail in the following section. + +Personalization with concept preservation. Off-the-shelf DMs often do not provide robust guidance for virtual object insertion, especially in out-of-distribution scenes such as outdoor driving environments. A potential solution, inspired by [57], is to adapt the DM using an image from the target scene. In our experience, however, this approach often resulted in too much overfitting to the target scene's content, reducing the model's ability to adapt to the scene with a newly inserted object. This led to artifacts and an unstable optimization process. + +To mitigate this issue, we propose fine-tuning the DM with a focus on preserving the identity of the objects to be inserted (Fig. 3). We specifically achieved this by generating additional synthetic images for the insertable class concept (e.g. car). We sample those images from the off-the-self DM, starting with a base prompt such as "a photo of a car" and appending attributes such as color, background, lighting, and size to ensure diversity in the generated data. + +![](images/95c3c69265f135eeb2c9b3cf21fbbb3ce87fd46adddf951bd9781515088ca11a.jpg) +Input image +"ascene in the style of sks rendering + +![](images/27d5ec4434c5f4a6e8cfd23876b353c58a69d45a6f4e5ebdf9f6035594d2d943.jpg) +Generated images +"aphotofacar" +Fig. 3: Personalization with concept preservation. + +![](images/fd243130eba1d645b5aeafdfe5b4fb49c059f81d2820fbef31de7e85cff82d16.jpg) + +We employ LoRA with rank 4 to fine-tune the DM, combining the in-domain target example with this supplemental data. Our training follows the objective described in Eq. 1, where $c$ corresponds to two predefined prompts: "a scene in the style of sks rendering" for the target image, and "a photo of a {concept class}". We sample approximately 30-40 supplementary images for indoor scenes and 200 for outdoor scenes. The total time spent for fine-tuning is typically less than 15 minutes on one high-end GPU with mixed precision training. + +Score Distillation with adapted guidance. Score Distillation Sampling [45] leverages a pretrained DM to guide the optimization of a differentiable, parametric image-rendering function $g_{\phi} \coloneqq \pmb{x}$ . Specifically, the parameters $\phi$ in our scenario corresponds to parameters of the Spherical Gaussian lighting and tone-mapping curves, are updated using the gradient: + +$$ +\nabla_ {\phi} \mathcal {L} _ {\mathrm {S D S}} (\boldsymbol {\phi}, \boldsymbol {\theta}) := \mathbb {E} _ {\boldsymbol {\epsilon} \sim \mathcal {N} (\mathbf {0}, \boldsymbol {I}), t \sim T} \left[ w (t) \left(\hat {\epsilon} _ {\boldsymbol {\theta}} (\boldsymbol {z} _ {t}, t, \boldsymbol {c}) - \boldsymbol {\epsilon}\right) \frac {\partial \boldsymbol {z} _ {t}}{\partial \phi} \right], \qquad (7) +$$ + +where $\pmb{z} = \mathcal{E}(g_{\phi})$ and $\hat{\epsilon}_{\theta}(\pmb{z}_t, t, \pmb{c}) \coloneqq (1 + s)\pmb{\epsilon}_{\theta}(\pmb{z}_t, t, \pmb{c}) - s\pmb{\epsilon}_{\theta}(\pmb{z}_t, t, \emptyset)$ . $\hat{\epsilon}$ denotes the classifier free guidance (CFG) version [21] of $\pmb{\epsilon}_{\theta}$ used in text-conditioned DMs to enable higher quality generation via a guidance scale parameter $s$ . + +Unfortunately, we encountered training instabilities when applying the original formulation of the SDS loss to our problem. Instead, we adopt the following alternative that integrates the LoRA personalization, and we call it LDS loss: + +$$ +\nabla_ {\phi} \mathcal {L} _ {\mathrm {L D S}} (\phi , \boldsymbol {\theta}) := \mathbb {E} _ {\epsilon \sim \mathcal {N} (\boldsymbol {0}, \boldsymbol {I}), t \sim T} \left[ w (t) \left(\epsilon_ {(\boldsymbol {\theta} + \boldsymbol {\Delta W})} (\boldsymbol {z} _ {t}, t, \boldsymbol {c}) - \epsilon_ {\boldsymbol {\theta}} (\boldsymbol {z} _ {t}, t, \varnothing)\right) \frac {\partial \boldsymbol {z} _ {t}}{\partial \phi} \right], \tag {8} +$$ + +where $\epsilon_{(\theta +\Delta W)}$ represents the predicted noise of the LoRA personalized model described in the previous section. Notably, this loss function bears resemblance to the Classifier Score Distillation (CSD) [68]. However, it's important to note that in our case, the delta is calculated between the adapted and non-adapted versions of the diffusion model. Intuitively, this delta guides the optimization process in a direction determined by the personalized model which preserves the concept of the inserted object while also capturing the appearance and semantics of the specific scene. + +# 4.3 Optimization Formulation + +After completing the diffusion model's personalization, we optimize the lighting and tone-mapping parameters using the following loss function: + +$$ +\mathcal {L} = \mathcal {L} _ {\mathrm {L D S}} + \lambda_ {\text {c o n s i s t e n c y}} \mathcal {L} _ {\text {c o n s i s t e n c y}} + \lambda_ {\text {r e g}} \mathcal {L} _ {\text {r e g}}. \tag {9} +$$ + +The loss' main component is the diffusion guidance $\mathcal{L}_{\mathrm{LDS}}$ , which provides a perceptual realism objective on the edited result with object insertion. We use the text prompt "a photo of a {concept class} in a scene in the style of sks rendering", which is a natural combination of the two text prompts used in personalization (Fig. 3). Here {concept class}, e.g. car, provides the context of the object to insert, while the personalized token sks provides the lighting style of the input image. The diffusion guidance $\mathcal{L}_{\mathrm{LDS}}$ is applied on the edited image $\mathbf{I}_{\mathrm{comp}}$ (Eq. 6), and backpropagated to the estimizable parameters through the differentiable rendering process (Eq. 8). + +The two regularizers $\mathcal{L}_{\mathrm{consistency}}$ and $\mathcal{L}_{\mathrm{reg}}$ are related to environment map fusion, which we detail below. + +Environment map initialization and fusion. The personalized DM provides guidance in two ways: (i) encouraging lighting consistency of the foreground object and the scene, such as the reflections and scale of the inserted object; and (ii) encouraging accurate shadows cast by the inserted object onto the background scene, such as the scale, direction, and color of the shadow. However, we empirically observe that these two signals can conflict in the early phase of the optimization. + +To address this, we initialize two separate environment maps, $\mathbf{L}^{\mathrm{fg}},\mathbf{L}^{\mathrm{shadow}}\in$ $\mathbb{R}^{H\times W\times 3}$ , to light the foreground inserted object (Eq. 4) and cast shadows (Eq. 5) respectively. As the optimization progresses, the two environment maps are progressively fused into a single environment map $\mathbf{L}^{\mathrm{fused}}$ through scaling the the $\mathbf{L}^{\mathrm{fg}}$ by the relative luminance of between $\mathbf{L}^{\mathrm{fg}}$ and $\mathbf{L}^{\mathrm{shadow}}$ . We also use the following two regularization terms for this fusion process. First, we encourage the consistency between the normalized luminance $\tilde{\mathbf{L}}^{\mathrm{fg}},\tilde{\mathbf{L}}^{\mathrm{shadow}}\in \mathbb{R}^{H\times W}$ of the environment maps by minimizing: + +$$ +\mathcal {L} _ {\text {c o n s i s t e n c y}} = - \sum_ {i, j} \tilde {\mathbf {L}} _ {i, j} ^ {\text {s h a d o w}} \log \left(\tilde {\mathbf {L}} _ {i, j} ^ {\mathrm {f g}}\right) \Delta \Omega_ {i, j} \tag {10} +$$ + +where $\varDelta\varOmega_{i,j}$ is the corresponding solid angle at pixel $(i,j)$ and the gradient for $\tilde{\mathbf{L}}^{\mathrm{shadow}}$ is detached. Second, as the shadow environment map $\mathbf{L}^{\mathrm{shadow}}$ is supervised mainly via the shadow ratio $\beta_{\mathrm{shadow}}\in \mathbb{R}^{H\times W\times 3}$ of Eq. 5, to encourage concentrated high peaks for the sharp shadows and suppress the ambient light in $\mathbf{L}^{\mathrm{shadow}}$ , we also add a L2 regularizer in log-space with the Cauchy loss [7]: + +$$ +\mathcal {L} _ {\text {r e g}} = \sum_ {i, j, c} \log \left(1 + 2 \left(\mathbf {L} _ {i, j, c} ^ {\text {s h a d o w}}\right) ^ {2}\right) \Delta \Omega_ {i, j}. \tag {11} +$$ + +Please refer to the Supplement for more details. + +Table 1: Quantitative user study on outdoor street scenes. For each scene, users are shown two results—one produced by our method, and another produced by one of the baselines—and select which is more realistic. We report the results averaged across 3 user studies with 9 users each. Our method outperforms all baselines ( $>50\%$ ) and is preferred in almost all illumination conditions. + +
DaytimeTwilightNightAll scenes
SunnyCloudy
Hold-Geoffroy et al. [22]60.8%66.7%74.1%85.7%68.8%
NLFE [62]80.4%73.3%44.4%52.4%67.4%
StyleLight [61]76.5%91.1%66.7%66.7%77.8%
DiffusionLight [44]80.4%68.9%55.6%71.4%70.8%
+ +Table 2: Quantitative evaluation on PolyHaven scenes. We report user study preference scores similar to Table 1. Metrics are computed w.r.t. to a “reference” image where the virtual object is lit by the ground-truth environment map. + +
MethodsOurs preferred ↓RMSE ↓SSIM ↑LPIPS [73] ↓si-RMSE [18] ↓
Wang et al. [63]84.8%0.0630.9850.01750.037
StyleLight [61]75.8%0.0560.9860.02020.039
DiffusionLight [44]66.7%0.0620.9850.01620.034
Ours/0.0480.9890.01470.027
+ +Training details. The personalized DM receives a crop of $\mathbf{I}_{\mathrm{comp}}$ (Eq. 6) with added noise. Since the inserted object likely does not cover the full background image, we use the visibility mask of the virtual object $V(\mathcal{X})$ to locate the 2D bounding box of the inserted object, and randomly crop around the inserted object. The size of the cropped image is also randomized, with smaller crops covering local details of the inserted object and larger crops providing more visual cues of the lighting effects within the scene. Each scene is optimized for 600 iterations, and the maximum strength of the diffusion guidance is linearly decreased from 0.6 to 0.3. We use $\lambda_{\mathrm{consistency}} = 0.03$ and $\lambda_{\mathrm{reg}} = 0.01$ . + +# 5 Experiments + +In this section, we evaluate DiPIR on a collection of outdoor and indoor scenes, covering diverse lighting conditions and different application domains. We first outline the experiment settings in Sec. 5.1 and then provide evaluation results on benchmark datasets in Sec. 5.2. Finally, we conduct an ablation study in Sec. 5.3. + +# 5.1 Experiment Settings + +Waymo dataset. Following the data split from [66], we use 48 scenes with diverse illumination conditions from the Waymo Open Dataset [55] including 32 daytime (17 sunny and 15 cloudy), 9 twilight, and 7 night scenes. Each scene consists of an input image, a 3D car asset randomly selected from an asset bank, the ground plane, and a location to insert the car. Inspired by applications in synthetic data generation, this process is fully automatic: we use the LiDAR point cloud and semantic segmentation [58] to fit a ground plane and detect the empty space to insert a car. + +PolyHaven dataset. We use 11 HDR environment maps from PolyHaven [1] and manually place a known ground plane and virtual object in each scene. Each + +![](images/8524d671d57967f33486f8f65292c31c0d91fa4498a0a052d1e1d0975d57f6e0.jpg) +Fig. 4: Comparisons on inserting objects into cropped HDRIs from PolyHaven. + +![](images/4fc027309d44855e0aa1432d1ac2501f11a0c6f8ff7a4a98afa307ed39ac0e1e.jpg) +Fig. 5: Inserting car assets into Waymo driving scenes. Note the direction and sharpness of shadows, as well as overall brightness, color and specular highlights on inserted cars. + +scenario therefore consists of an LDR background image $\mathbf{I}_{\mathrm{bg}}$ rendered directly from the environment map, the posed virtual object, and the proxy plane. A pseudo-ground-truth rendering is created by lighting the inserted object with the environment map itself. + +User study. To evaluate the perceptual realism of the virtual object insertion, we conducted a user study where participants received a pair of two object insertion results in a random order - one generated using our proposed method, the other using a baseline approach. The participants were then instructed to compare the differences between the two results, inspect the lighting effects of the inserted object, and choose the image they perceived as more realistic. We invited 9 users to perform a binary selection for each image pair and used majority voting to determine the preferred image for each comparison. In the following, we report the percentage of times that our method was selected over the baseline. This process was repeated three times and a preferred percentage $>50\%$ indicates + +Ours outperforming baselines. We include more details about the user study and statistical evaluation of its results in the Supplement. + +Baselines. On the Waymo dataset of urban street scenes, we compare our method to the representative outdoor lighting estimation methods of Hold-Geoffroy et al. [22] and NLFE [62]; as well as the generative lighting estimation methods StyleLight [61] and DiffusionLight [44]. On the PolyHaven dataset, we compare with Wang et al. [63], StyleLight [61], DiffusionLight [44], and pseudo-ground-truth rendering. + +# 5.2 Evaluation on Benchmark Datasets + +Urban street scenes. In Table 1, we provide quantitative user study results on the Waymo dataset [55], separated into 4 subsets based on illumination conditions, as well as on all scenes. Compared to all baselines, our method outperforms prior state-of-the-art methods (chosen $>50\%$ ), and is preferred in almost every subset. + +StyleLight [61] is primarily trained on indoor HDR panoramas and suffers from domain gap to outdoor scenes. Similarly, the sky model predicted by Hold-Geoffroy et al. [22] is tailored to outdoor daytime scenes and suffers severely from out-of-domain twilight and night scenes. NLFE [62] unprojects the surrounding scene geometry and predicts volumetric lighting, which makes it generalize better to night scenes, but it can fail to estimate an accurate scale and color of the peak for daytime scenes. DiffusionLight [44], a concurrent work, shows impressive results by inpainting a chrome sphere and predicting environment maps with realistic high-frequency details. However, it often struggles to predict a high-intensity peak in sunny daytime scenes or a proper scale for night scenes. + +A qualitative comparison is shown in Fig. 5. Our method achieves high-quality object insertion, making it a promising approach for applications such as simulating synthetic data, augmented reality navigation and urban planning. + +PolyHaven dataset. We show the quantitative comparison in Table 2 and qualitative results in Fig. 4. Our method is preferred over all baseline methods in the user study, and outperforms baselines on quantitative metrics. Our method produces virtual object insertion that naturally and consistently blend in the scenes, making it promising for virtual production tasks. + +# 5.3 Ablation Study + +We extensively ablate our proposed design choices on diffusion guidance and scene representations, and provide the quantitative ablation in Table 3 with user study and qualitative results in Fig. 6. For the diffusion guidance, we compare with simplified or alternative versions: (i) Ours (SDS) removes the adaptive guidance (Eq. 8) and uses the original SDS [45] formulation (Eq. 7); (ii) Ours (SDS w/o LoRA) uses the original SDS loss with non-adapted DM; (iii) Ours (w/o concept preservation) uses our LDS loss with adapted DM but not using any class images for concept preservation. We additionally compare with a baseline of "dataset update" which applies a photometric L1 loss on DM repeatedly edited results + +
MethodsOurs preferred ↓
Ours (dataset update)85.2%
Ours (SDS [45])74.1%
Ours (SDS [45] w/o LoRA)90.7%
Ours (w/o concept preservation)64.8%
Ours (w/o tone curve)68.5%
Ours (w/o env. map fusion)66.7%
+ +Table 3: Ablation study on outdoor driving scenes [55]. We report the percentage of images that users preferred DiPIR compared to its ablated versions. Our full pipeline produces results that are preferred more often over its ablated versions. + +![](images/3e40de3e5434336d40fdd305d84be29c140bc97fddcaf84f801bf9c1598822f7.jpg) +Ours (SDS w/o LoRA) + +![](images/58d3b97546e9e7ad48a0c84a29763969d058d13933177e10619513613ccec2be.jpg) +Ours (SDS w/ LoRA) + +![](images/2dac3abc527de255f449c05c8f17e09ee34fad2cc05b5b665f2de952db3469ce.jpg) +Ours (dataset update) + +![](images/9b12df193fc4464c23fec94403d8ff51f70b51a46ed16e68459ab780aa6943b3.jpg) +Ours + +![](images/6e016a3a6e3a24546148c073f8c34287a03d2c0cac1fd1932cda4f1d3863970e.jpg) +Ours (w/o concept preservation) +Ours (w/o tone-mapping curve) +Fig. 6: Qualitative ablation study of our design choices. + +![](images/2eb94103d8cdc037afb39b45888493f7a5fd1a85dec865c18848b8418481a0df.jpg) + +![](images/ba5e321021181cd0b723f72dd9d7ef58fa2551bffc890b699c73d9c12a731631.jpg) +Ours (w/o env. fusion) + +curve) + +via SDEdit [41], denoted as Ours (dataset update). Although this approach is adopted in text-guided scene editing [19], we found it challenging to produce stable gradients from such a "discrete" guidance. + +When designing the light and tone representations, we observe that the trainable tone-mapping curve provides additional flexibility to compensate for unknown tone-mapping in the input image, and better match the color and scale of the shadows. The two environment map fusion scheme (Sec. 4.3) allows to robustly recover the high intensity peaks in the early phase of training, and converge to better quality. Their respective effect is ablated as "Ours (w/o tone curve)" and "Ours (w/o env. map fusion)". + +# 6 Applications + +Since our method recovers physically based lighting information, arbitrary new virtual objects can be inserted after the optimization, as shown in Fig. 2. DiPIR can also optimize other scene attributes such as materials and local lighting. We perform preliminary experiments in this direction. + +Material optimization. When combined with differentiable rendering, DMs can provide a guidance signal for material attributes, as shown in Fig. 7. Given a purely diffuse car and enabling Metallic and Roughness properties as estimizable parameters, the diffusion guidance can optimize and make the car look more shiny. By additionally changing the text prompt to "a carmine red car" and making the base color of the car an estimizable parameter, we show that the DM can propagate the text-condition to the PBR attribute and change the color of the car to red. When enabling local emission as an estimizable parameter, the diffusion model can also turn on the headlights of cars in night scenes. + +![](images/6364e243daa135c0da173a6e797c14e80dd7297c20f39c9539a6eb0c7a83f193.jpg) +Optimize tone mapping curve + +![](images/8cc9eb2aea1140c8dce37960fd30b0fe17d36d0415922ef0caea87fd0dfbeaec.jpg) + +![](images/146dd37cccebf6de47098121eec0b7a18fe18c063977f0dbbddcb99a6903f994.jpg) + +![](images/658200a4dde3bbe76bae075467cf465d254d93baa8d9c771f33dad7d68a1f73a.jpg) + +![](images/e50a2cf49f962e438fc1009327adfc9955637f39d2402732e9f5c3df9844526f.jpg) + +![](images/296bd4293843b9a9820b90be5587f777cf6f1238b5beac1d54d0f2d8732bc6de.jpg) +Optimize asset materials + +![](images/ee8f975fae2507cf9632e18414a4a9af4b4ab3c75d0da895092b5537498b2f38.jpg) +Fig. 7: Our physically based inverse rendering pipeline unlocks further applications such as material, local emission and tone-mapping refinement. + +![](images/fbd79d06af1fd28af9eb4b9cd3d7c8b6c2c1891c2d46384ee89a4cbb615e7eea.jpg) + +![](images/5cb04fccb262f75c9525d9e04105a9963b5a95bb91f028caf2349ee806891af2.jpg) +Optimize emissions + +Tone-mapping adjustment. We use a controlled experiment to further evaluate how well DMs understand tone-mapping. In Fig. 7, we freeze the estimated environment map and apply a manual tone adjustment on the background image. The diffusion guidance is used to optimize the tone curve such that the inserted object matches the surrounding background in the final composited result. + +# 7 Discussion + +Our method leverages large diffusion models' inherent scene understanding capabilities as guidance to a physically based inverse rendering pipeline. We design a diffusion guidance signal with scene-specific personalization and a differentiable inverse rendering pipeline to recover lighting and tone-mapping parameters. Our method enables inserting virtual objects into scenes, but also optimizing other scene parameters such as the materials of the inserted object or account for tone-mapping mismatches between cameras. We believe that this combination of the differentiable rendering process and data-driven priors can be used successfully in many other content creation applications such as relighting and animation. + +Limitations and future work. Our Spherical Gaussians-based lighting representation is adequate for general objects [35], but might not behave realistically for highly specular materials. For more complex lighting representations adding generative priors on the environment map [40] is a direction worth exploring. The rendering formulation could be extended to account for effects such as reflections from the scene itself onto the inserted object (e.g. color bleeding), but that might introduce more ambiguities and require knowing the materials of the proxy geometry (refer to Supplement C.4 for failure case examples). Finally, while DM personalization significantly improves the quality of the results, it adds overhead and complexity to the pipeline. Recent personalization methods that do not require test-time finetuning [52] could be used to mitigate this overhead. + +# Acknowledgements + +The authors are grateful for the feedback received from Nicholas Sharp and Huan Ling during the project. We thank the original artists of the 3D assets used in this work: inciprocal, peyman.khaleghi, Kuutti Siitonen, TurboSquid and their artists Hum3D and Amaranthus. + +# References + +1. Poly Haven - The Public 3D Asset Library, https://polyhaven.com +2. Balaji, Y., Nah, S., Huang, X., Vahdat, A., Song, J., Zhang, Q., Kreis, K., Aittala, M., Aila, T., Laine, S., Catanzaro, B., Karras, T., Liu, M.Y.: ediff-i: Text-to-image diffusion models with ensemble of expert denoisers. arXiv preprint arXiv:2211.01324 (2022) +3. Bangaru, S.P., Li, T.M., Durand, F.: Unbiased warped-area sampling for differentiable rendering. ACM Trans. Graph. 39(6) (nov 2020). https://doi.org/10.1145/3414685.3417833, https://doi.org/10.1145/3414685.3417833 +4. Barron, J.T., Malik, J.: Shape, illumination, and reflectance from shading. IEEE transactions on pattern analysis and machine intelligence 37(8), 1670-1687 (2014) +5. Barrow, H., Tenenbaum, J., Hanson, A., Riseman, E.: Recovering intrinsic scene characteristics. Comput. Vis. Syst 2, 3-26 (1978) +6. Bell, S., Bala, K., Snavely, N.: Intrinsic images in the wild. ACM Transactions on Graphics (TOG) 33(4), 159 (2014) +7. Black, M.J., Anandan, P.: The robust estimation of multiple motions: Parametric and piecewise-smooth flow fields. Computer vision and image understanding 63(1), 75-104 (1996) +8. Boss, M., Jampani, V., Kim, K., Lensch, H.P., Kautz, J.: Two-shot spatially-varying brdf and shape estimation. In: CVPR (2020) +9. Bousseau, A., Paris, S., Durand, F.: User-assisted intrinsic images. In: ACM Transactions on Graphics (TOG). vol. 28, p. 130. ACM (2009) +10. Chari, P., Ma, S., Ostashev, D., Kadambi, A., Krishnan, G., Wang, J., Aberman, K.: Personalized restoration via dual-pivot tuning. arXiv preprint arXiv:2312.17234 (2023) +11. Dai, X., Hou, J., Ma, C.Y., Tsai, S., Wang, J., Wang, R., Zhang, P., Vandenhende, S., Wang, X., Dubey, A., et al.: Emu: Enhancing image generation models using photogenic needles in a haystack. arXiv preprint arXiv:2309.15807 (2023) +12. Dastjerdi, M.R.K., Eisenmann, J., Hold-Geoffroy, Y., Lalonde, J.F.: Everlight: Indoor-outdoor editable hdr lighting estimation. In: Proceedings of the IEEE/CVF International Conference on Computer Vision (ICCV). pp. 7420-7429 (October 2023) +13. Durkan, C., Bekasov, A., Murray, I., Papamakarios, G.: Neural spline flows. Advances in neural information processing systems 32 (2019) +14. Gal, R., Alaluf, Y., Atzmon, Y., Patashnik, O., Bermano, A.H., Chechik, G., Cohen-Or, D.: An image is worth one word: Personalizing text-to-image generation using textual inversion (2022). https://doi.org/10.48550/ARXIV.2208.01618, https://arxiv.org/abs/2208.01618 +15. Gardner, M.A., Hold-Geoffroy, Y., Sunkavalli, K., Gagné, C., Lalonde, J.F.: Deep parametric indoor lighting estimation. In: ICCV. pp. 7175-7183 (2019) + +16. Gardner, M.A., Sunkavalli, K., Yumer, E., Shen, X., Gambaretto, E., Gagné, C., Lalonde, J.F.: Learning to predict indoor illumination from a single image. arXiv preprint arXiv:1704.00090 (2017) +17. Garon, M., Sunkavalli, K., Hadap, S., Carr, N., Lalonde, J.F.: Fast spatially-varying indoor lighting estimation. In: CVPR. pp. 6908-6917 (2019) +18. Grosse, R., Johnson, M.K., Adelson, E.H., Freeman, W.T.: Ground truth dataset and baseline evaluations for intrinsic image algorithms. In: ICCV. pp. 2335-2342. IEEE (2009) +19. Haque, A., Tancik, M., Efros, A., Holynski, A., Kanazawa, A.: Instruct-nerf2nerf: Editing 3d scenes with instructions. In: Proceedings of the IEEE/CVF International Conference on Computer Vision (2023) +20. Ho, J., Jain, A., Abbeel, P.: Denoising diffusion probabilistic models. arXiv preprint arxiv:2006.11239 (2020) +21. Ho, J., Salimans, T.: Classifier-free diffusion guidance. arXiv preprint arXiv:2207.12598 (2022) +22. Hold-Geoffroy, Y., Athawale, A., Lalonde, J.F.: Deep sky modeling for single image outdoor lighting estimation. In: CVPR. pp. 6927-6935 (2019) +23. Hold-Geoffroy, Y., Sunkavalli, K., Hadap, S., Gambaretto, E., Lalonde, J.F.: Deep outdoor illumination estimation. In: CVPR. pp. 7312-7321 (2017) +24. Hu, E.J., yelong shen, Wallis, P., Allen-Zhu, Z., Li, Y., Wang, S., Wang, L., Chen, W.: LoRA: Low-rank adaptation of large language models. In: International Conference on Learning Representations (2022), https://openreview.net/forum?id=nZeVKeeFYf9 +25. Jakob, W., Speierer, S., Roussel, N., Nimier-David, M., Vicini, D., Zeltner, T., Nicolet, B., Crespo, M., Leroy, V., Zhang, Z.: Mitsuba 3 renderer (2022), https://mitsubarenderer.org +26. Kajiya, J.T.: The rendering equation. In: Proceedings of the 13th annual conference on Computer graphics and interactive techniques. pp. 143-150 (1986) +27. Karimi Dastjerdi, M.R., Hold-Geoffroy, Y., Eisenmann, J., Khodadadeh, S., Lalonde, J.F.: Guided co-modulated gan for 360 field of view extrapolation. International Conference on 3D Vision (3DV) (2022) +28. Ke, B., Obukhov, A., Huang, S., Metzger, N., Daudt, R.C., Schindler, K.: Repurposing diffusion-based image generators for monocular depth estimation. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR) (2024) +29. Kocsis, P., Sitzmann, V., Nießner, M.: Intrinsic image diffusion for single-view material estimation. In: arxiv (2023) +30. Kovacs, B., Bell, S., Snavely, N., Bala, K.: Shading annotations in the wild. In: CVPR. pp. 6998-7007 (2017) +31. Land, E.H., McCann, J.J.: Lightness and retina theory. Josa 61(1), 1-11 (1971) +32. LeGendre, C., Ma, W.C., Fyffe, G., Flynn, J., Charbonnel, L., Busch, J., Debevec, P.: Deeplight: Learning illumination for unconstrained mobile mixed reality. In: CVPR. pp. 5918-5928 (2019) +33. Li, T.M., Aittala, M., Durand, F., Lehtinen, J.: Differentiable monte carlo ray tracing through edge sampling. ACM Trans. Graph. 37(6) (dec 2018). https://doi.org/10.1145/3272127.3275109, https://doi.org/10.1145/3272127.3275109 +34. Li, Z., Snavely, N.: Cginintrinsics: Better intrinsic image decomposition through physically-based rendering. In: ECCV. pp. 371-387 (2018) +35. Li, Z., Shafiei, M., Ramamoorthi, R., Sunkavalli, K., Chandraker, M.: Inverse rendering for complex indoor scenes: Shape, spatially-varying lighting and svbrdf from a single image. In: CVPR. pp. 2475-2484 (2020) + +36. Li, Z., Xu, Z., Ramamoorthi, R., Sunkavalli, K., Chandraker, M.: Learning to reconstruct shape and spatially-varying reflectance from a single image. ACM Transactions on Graphics (TOG) 37(6), 1-11 (2018) +37. Li, Z., Yu, L., Okunev, M., Chandraker, M., Dong, Z.: Spatiotemporally consistent hdr indoor lighting estimation. ACM Trans. Graph. 42(3) (jun 2023). https://doi.org/10.1145/3595921, https://doi.org/10.1145/3595921 +38. Li, Z., Yu, T.W., Sang, S., Wang, S., Bi, S., Xu, Z., Yu, H.X., Sunkavalli, K., Hasan, M., Ramamoorthi, R., et al.: Openrooms: An end-to-end open framework for photorealistic indoor scene datasets. arXiv preprint arXiv:2007.12868 (2020) +39. Loubet, G., Holzschuch, N., Jakob, W.: Reparameterizing discontinuous integrands for differentiable rendering. ACM Trans. Graph. 38(6) (nov 2019). https://doi.org/10.1145/3355089.3356510, https://doi.org/10.1145/3355089.3356510 +40. Lyu, L., Tewari, A., Habermann, M., Saito, S., Zollhöfer, M., Leimkuehler, T., Theobalt, C.: Diffusion posterior illumination for ambiguity-aware inverse rendering. ACM Transactions on Graphics 42(6) (2023) +41. Meng, C., He, Y., Song, Y., Song, J., Wu, J., Zhu, J.Y., Ermon, S.: SDEdit: Guided image synthesis and editing with stochastic differential equations. In: International Conference on Learning Representations (2022) +42. Nimier-David, M., Speierer, S., Ruiz, B., Jakob, W.: Radiative backpropagation: an adjoint method for lightning-fast differentiable rendering. ACM Trans. Graph. 39(4) (aug 2020). https://doi.org/10.1145/3386569.3392406, https://doi.org/10.1145/3386569.3392406 +43. Nimier-David, M., Vicini, D., Zeltner, T., Jakob, W.: Mitsuba 2: a retargetable forward and inverse renderer. ACM Trans. Graph. 38(6) (nov 2019). https://doi.org/10.1145/3355089.3356498, https://doi.org/10.1145/3355089.3356498 +44. Phongthawee, P., Chinchuthakun, W., Sinsunthithet, N., Raj, A., Jampani, V., Khungurn, P., Suwajanakorn, S.: Diffusionlight: Light probes for free by painting a chrome ball. In: ArXiv (2023) +45. Poole, B., Jain, A., Barron, J.T., Mildenhall, B.: Dreamfusion: Text-to-3d using 2d diffusion. arXiv (2022) +46. Rombach, R., Blattmann, A., Lorenz, D., Esser, P., Ommer, B.: High-resolution image synthesis with latent diffusion models (2021) +47. Ruiz, N., Li, Y., Jampani, V., Pritch, Y., Rubinstein, M., Aberman, K.: Dreambooth: Fine tuning text-to-image diffusion models for subject-driven generation (2022) +48. Sahara, C., Chan, W., Saxena, S., Li, L., Whang, J., Denton, E.L., Ghasemipour, K., Gontijo Lopes, R., Karagol Ayan, B., Salimans, T., et al.: Photorealistic text-to-image diffusion models with deep language understanding. Advances in Neural Information Processing Systems 35, 36479-36494 (2022) +49. Sarkar, A., Mai, H., Mahapatra, A., Lazebnik, S., Forsyth, D., Bhattad, A.: Shadows don't lie and lines can't bend! generative models don't know projective geometry...for now (2023) +50. Sengupta, S., Gu, J., Kim, K., Liu, G., Jacobs, D.W., Kautz, J.: Neural inverse rendering of an indoor scene from a single image. In: ICCV (2019) +51. Shah, V., Ruiz, N., Cole, F., Lu, E., Lazebnik, S., Li, Y., Jampani, V.:Ziplora: Any subject in any style by effectively merging loras (2023) +52. Shi, J., Xiong, W., Lin, Z., Jung, H.J.: Instantbooth: Personalized text-to-image generation without test-time finetuning. arXiv preprint arXiv:2304.03411 (2023) +53. Song, S., Funkhouser, T.: Neural illumination: Lighting prediction for indoor environments. In: CVPR. pp. 6918-6926 (2019) + +54. Srinivasan, P.P., Mildenhall, B., Tancik, M., Barron, J.T., Tucker, R., Snavely, N.: Lighthouse: Predicting lighting volumes for spatially-coherent illumination. In: CVPR. pp. 8080-8089 (2020) +55. Sun, P., Kretzschmar, H., Dotiwalla, X., Chouard, A., Patnaik, V., Tsui, P., Guo, J., Zhou, Y., Chai, Y., Caine, B., et al.: Scalability in perception for autonomous driving: Waymo open dataset. In: CVPR (2020) +56. Tang, J., Zhu, Y., Wang, H., Chan, J.H., Li, S., Shi, B.: Estimating spatially-varying lighting in urban scenes with disentangled representation. In: ECCV (2022) +57. Tang, L., Ruiz, N., Qinghao, C., Li, Y., Holynski, A., Jacobs, D.E., Hariharan, B., Pritch, Y., Wadhwa, N., Aberman, K., Rubinstein, M.: Realfill: Reference-driven generation for authentic image completion. arXiv preprint arXiv:2309.16668 (2023) +58. Tao, A., Sapra, K., Catanzaro, B.: Hierarchical multi-scale attention for semantic segmentation. arXiv preprint arXiv:2005.10821 (2020) +59. Veach, E., Guibas, L.J.: Optimally combining sampling techniques for monte carlo rendering. In: Proceedings of the 22nd annual conference on Computer graphics and interactive techniques. pp. 419-428 (1995) +60. Vicini, D., Speierer, S., Jakob, W.: Path replay backpropagation: differentiating light paths using constant memory and linear time. ACM Trans. Graph. 40(4) (jul 2021). https://doi.org/10.1145/3450626.3459804, https://doi.org/10.1145/3450626.3459804 +61. Wang, G., Yang, Y., Loy, C.C., Liu, Z.: Stylelight: HDR panorama generation for lighting estimation and editing. In: European Conference on Computer Vision (ECCV) (2022) +62. Wang, Z., Chen, W., Acuna, D., Kautz, J., Fidler, S.: Neural light field estimation for street scenes with differentiable virtual object insertion. In: ECCV (2022) +63. Wang, Z., Philion, J., Fidler, S., Kautz, J.: Learning indoor inverse rendering with 3d spatially-varying lighting. In: ICCV (2021) +64. Wimbauer, F., Wu, S., Rupprecht, C.: De-rendering 3d objects in the wild. In: CVPR (2022) +65. Yan, K., Lassner, C., Budge, B., Dong, Z., Zhao, S.: Efficient estimation of boundary integrals for path-space differentiable rendering. ACM Trans. Graph. 41(4) (jul 2022). https://doi.org/10.1145/3528223.3530080, https://doi.org/10.1145/3528223.3530080 +66. Yang, J., Ivanovic, B., Litany, O., Weng, X., Kim, S.W., Li, B., Che, T., Xu, D., Fidler, S., Pavone, M., Wang, Y.: Emernerf: Emergent spatial-temporal scene decomposition via self-supervision. arXiv preprint arXiv:2311.02077 (2023) +67. Yu, H.X., Agarwala, S., Herrmann, C., Szeliski, R., Snavely, N., Wu, J., Sun, D.: Accidental light probes. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 12521-12530 (2023) +68. Yu, X., Guo, Y.C., Li, Y., Liang, D., Zhang, S.H., Qi, X.: Text-to-3d with classifier score distillation (2023) +69. Yu, Y., Smith, W.A.: Inverserendernet: Learning single image inverse rendering. In: CVPR (2019) +70. Zhan, F., Zhang, C., Yu, Y., Chang, Y., Lu, S., Ma, F., Xie, X.: Emlight: Lighting estimation via spherical distribution approximation. In: Proceedings of the AAAI Conference on Artificial Intelligence (2021) +71. Zhang, C., Miller, B., Yan, K., Gkioulekas, I., Zhao, S.: Path-space differentiable rendering. ACM Trans. Graph. 39(4) (aug 2020). https://doi.org/10.1145/3386569.3392383, https://doi.org/10.1145/3386569.3392383 + +72. Zhang, C., Yu, Z., Zhao, S.: Path-space differentiable rendering of participating media. ACM Trans. Graph. 40(4) (jul 2021). https://doi.org/10.1145/3450626.3459782, https://doi.org/10.1145/3450626.3459782 +73. Zhang, R., Isola, P., Efros, A.A., Shechtman, E., Wang, O.: The unreasonable effectiveness of deep features as a perceptual metric. In: Proceedings of the IEEE conference on computer vision and pattern recognition. pp. 586-595 (2018) +74. Zhang, Z., Roussel, N., Jakob, W.: Projective sampling for differentiable rendering of geometry. ACM Trans. Graph. 42(6) (dec 2023). https://doi.org/10.1145/3618385, https://doi.org/10.1145/3618385 +75. Zhao, Q., Tan, P., Dai, Q., Shen, L., Wu, E., Lin, S.: A closed-form solution to retinax with nonlocal texture constraints 34(7), 1437-1444 (2012) +76. Zhao, Y., Guo, T.: Pointar: Efficient lighting estimation for mobile augmented reality. arXiv preprint arXiv:2004.00006 (2020) +77. Zhu, Y., Zhang, Y., Li, S., Shi, B.: Spatially-varying outdoor lighting estimation from intrinsics. In: CVPR (2021) \ No newline at end of file diff --git a/2024/Photorealistic Object Insertion with Diffusion-Guided Inverse Rendering/images.zip b/2024/Photorealistic Object Insertion with Diffusion-Guided Inverse Rendering/images.zip new file mode 100644 index 0000000000000000000000000000000000000000..3646386c69cba1eb585a5637bc03c722296e71fd --- /dev/null +++ b/2024/Photorealistic Object Insertion with Diffusion-Guided Inverse Rendering/images.zip @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:f490b35a07862617a87856e8d2a09becefd88bea8e009571ed5ab098d19ae194 +size 540688 diff --git a/2024/Photorealistic Object Insertion with Diffusion-Guided Inverse Rendering/layout.json b/2024/Photorealistic Object Insertion with Diffusion-Guided Inverse Rendering/layout.json new file mode 100644 index 0000000000000000000000000000000000000000..7f24f229516b41ca32bb2df63eb41a3b4ac2fe87 --- /dev/null +++ b/2024/Photorealistic Object Insertion with Diffusion-Guided Inverse Rendering/layout.json @@ -0,0 +1,12005 @@ +{ + "pdf_info": [ + { + "para_blocks": [ + { + "bbox": [ + 181, + 111, + 432, + 147 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 181, + 111, + 432, + 147 + ], + "spans": [ + { + "bbox": [ + 181, + 111, + 432, + 147 + ], + "type": "text", + "content": "Photorealistic Object Insertion with Diffusion-Guided Inverse Rendering" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 154, + 167, + 460, + 193 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 154, + 167, + 460, + 193 + ], + "spans": [ + { + "bbox": [ + 154, + 167, + 460, + 193 + ], + "type": "text", + "content": "Ruofan Liang" + }, + { + "bbox": [ + 154, + 167, + 460, + 193 + ], + "type": "inline_equation", + "content": "^{1,2,3}" + }, + { + "bbox": [ + 154, + 167, + 460, + 193 + ], + "type": "text", + "content": ", Zan Gojcic" + }, + { + "bbox": [ + 154, + 167, + 460, + 193 + ], + "type": "inline_equation", + "content": "^{1}" + }, + { + "bbox": [ + 154, + 167, + 460, + 193 + ], + "type": "text", + "content": ", Merlin Nimier-David" + }, + { + "bbox": [ + 154, + 167, + 460, + 193 + ], + "type": "inline_equation", + "content": "^{1}" + }, + { + "bbox": [ + 154, + 167, + 460, + 193 + ], + "type": "text", + "content": ", David Acuna" + }, + { + "bbox": [ + 154, + 167, + 460, + 193 + ], + "type": "inline_equation", + "content": "^{1}" + }, + { + "bbox": [ + 154, + 167, + 460, + 193 + ], + "type": "text", + "content": ", Nandita Vijaykumar" + }, + { + "bbox": [ + 154, + 167, + 460, + 193 + ], + "type": "inline_equation", + "content": "^{2,3}" + }, + { + "bbox": [ + 154, + 167, + 460, + 193 + ], + "type": "text", + "content": ", Sanja Fidler" + }, + { + "bbox": [ + 154, + 167, + 460, + 193 + ], + "type": "inline_equation", + "content": "^{1,2,3}" + }, + { + "bbox": [ + 154, + 167, + 460, + 193 + ], + "type": "text", + "content": ", and Zian Wang" + }, + { + "bbox": [ + 154, + 167, + 460, + 193 + ], + "type": "inline_equation", + "content": "^{1,2,3}" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 194, + 201, + 418, + 213 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 194, + 201, + 418, + 213 + ], + "spans": [ + { + "bbox": [ + 194, + 201, + 418, + 213 + ], + "type": "inline_equation", + "content": "^{1}" + }, + { + "bbox": [ + 194, + 201, + 418, + 213 + ], + "type": "text", + "content": "NVIDIA " + }, + { + "bbox": [ + 194, + 201, + 418, + 213 + ], + "type": "inline_equation", + "content": "^{2}" + }, + { + "bbox": [ + 194, + 201, + 418, + 213 + ], + "type": "text", + "content": "University of Toronto " + }, + { + "bbox": [ + 194, + 201, + 418, + 213 + ], + "type": "inline_equation", + "content": "^{3}" + }, + { + "bbox": [ + 194, + 201, + 418, + 213 + ], + "type": "text", + "content": "Vector Institute" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 160, + 241, + 452, + 384 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 160, + 241, + 452, + 384 + ], + "spans": [ + { + "bbox": [ + 160, + 241, + 452, + 384 + ], + "type": "text", + "content": "Abstract. The correct insertion of virtual objects in images of real-world scenes requires a deep understanding of the scene's lighting, geometry and materials, as well as the image formation process. While recent large-scale diffusion models have shown strong generative and inpainting capabilities, we find that current models do not sufficiently \"understand\" the scene shown in a single picture to generate consistent lighting effects (shadows, bright reflections, etc.) while preserving the identity and details of the composited object. We propose using a personalized large diffusion model as guidance to a physically based inverse rendering process. Our method recovers scene lighting and tone-mapping parameters, allowing the photorealistic composition of arbitrary virtual objects in single frames or videos of indoor or outdoor scenes. Our physically based pipeline further enables automatic materials and tone-mapping refinement." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 160, + 395, + 453, + 417 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 160, + 395, + 453, + 417 + ], + "spans": [ + { + "bbox": [ + 160, + 395, + 453, + 417 + ], + "type": "text", + "content": "Keywords: Inverse rendering " + }, + { + "bbox": [ + 160, + 395, + 453, + 417 + ], + "type": "inline_equation", + "content": "\\cdot" + }, + { + "bbox": [ + 160, + 395, + 453, + 417 + ], + "type": "text", + "content": " Diffusion models " + }, + { + "bbox": [ + 160, + 395, + 453, + 417 + ], + "type": "inline_equation", + "content": "\\cdot" + }, + { + "bbox": [ + 160, + 395, + 453, + 417 + ], + "type": "text", + "content": " Personalization " + }, + { + "bbox": [ + 160, + 395, + 453, + 417 + ], + "type": "inline_equation", + "content": "\\cdot" + }, + { + "bbox": [ + 160, + 395, + 453, + 417 + ], + "type": "text", + "content": " Virtual object insertion " + }, + { + "bbox": [ + 160, + 395, + 453, + 417 + ], + "type": "inline_equation", + "content": "\\cdot" + }, + { + "bbox": [ + 160, + 395, + 453, + 417 + ], + "type": "text", + "content": " Physically based rendering" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 133, + 437, + 229, + 449 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 437, + 229, + 449 + ], + "spans": [ + { + "bbox": [ + 133, + 437, + 229, + 449 + ], + "type": "text", + "content": "1 Introduction" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 132, + 461, + 481, + 509 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 461, + 481, + 509 + ], + "spans": [ + { + "bbox": [ + 132, + 461, + 481, + 509 + ], + "type": "text", + "content": "Virtual object insertion enables a range of applications from virtual production, to interactive gaming and synthetic data generation. To produce photorealistic insertions, the interactions between the virtual objects and the environment need to be modeled faithfully, such as accurate specular highlights and shadows." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 132, + 510, + 482, + 593 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 510, + 482, + 593 + ], + "spans": [ + { + "bbox": [ + 132, + 510, + 482, + 593 + ], + "type": "text", + "content": "A standard virtual object insertion pipeline typically includes three key steps: i) lighting estimation from the input image, ii) 3D proxy geometry creation, and iii) composited image rendering in a rendering engine. However, the first and arguably most important step is still an open research question. Lighting estimation is particularly challenging when dealing with limited inputs such as a single image, captured using a consumer device with a low dynamic range. Indeed, inverse rendering is a fundamentally ill-posed problem." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 132, + 594, + 482, + 666 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 594, + 482, + 666 + ], + "spans": [ + { + "bbox": [ + 132, + 594, + 482, + 666 + ], + "type": "text", + "content": "To constrain its solution space, prior works either aimed to define hand-crafted priors [9,18,31,75] or to learn them from the data [15-17,22,23,32,36,56,62,63,77]. However, the former often fall short when applied to real-world scenes, while the latter suffer from scarcity of the ground truth data. As a result, these algorithms are often heavily tailored to a specific domain, e.g. indoor [15-17,50,63] or outdoor scenes [22,23,56,62,77]." + } + ] + } + ], + "index": 8 + } + ], + "discarded_blocks": [], + "page_size": [ + 612, + 792 + ], + "page_idx": 0 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 133, + 129, + 239, + 200 + ], + "blocks": [ + { + "bbox": [ + 192, + 118, + 288, + 129 + ], + "lines": [ + { + "bbox": [ + 192, + 118, + 288, + 129 + ], + "spans": [ + { + "bbox": [ + 192, + 118, + 288, + 129 + ], + "type": "text", + "content": "Virtual Object Insertion" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 133, + 129, + 239, + 200 + ], + "lines": [ + { + "bbox": [ + 133, + 129, + 239, + 200 + ], + "spans": [ + { + "bbox": [ + 133, + 129, + 239, + 200 + ], + "type": "image", + "image_path": "eb9557c42ae5d42da10f49db59147e2ba8d2ce6694f2dfbe520005d7073f2e0e.jpg" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 148, + 201, + 229, + 210 + ], + "lines": [ + { + "bbox": [ + 148, + 201, + 229, + 210 + ], + "spans": [ + { + "bbox": [ + 148, + 201, + 229, + 210 + ], + "type": "text", + "content": "A vase with dried flowers" + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "image_caption" + } + ], + "index": 3 + }, + { + "type": "image", + "bbox": [ + 243, + 129, + 349, + 200 + ], + "blocks": [ + { + "bbox": [ + 243, + 129, + 349, + 200 + ], + "lines": [ + { + "bbox": [ + 243, + 129, + 349, + 200 + ], + "spans": [ + { + "bbox": [ + 243, + 129, + 349, + 200 + ], + "type": "image", + "image_path": "036064d38f10febd97136079be0dcbefb40515995ac8bf67bcf6b7592671a527.jpg" + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 270, + 201, + 324, + 209 + ], + "lines": [ + { + "bbox": [ + 270, + 201, + 324, + 209 + ], + "spans": [ + { + "bbox": [ + 270, + 201, + 324, + 209 + ], + "type": "text", + "content": "A black SUV car" + } + ] + } + ], + "index": 6, + "angle": 0, + "type": "image_caption" + } + ], + "index": 5 + }, + { + "type": "image", + "bbox": [ + 133, + 212, + 239, + 283 + ], + "blocks": [ + { + "bbox": [ + 133, + 212, + 239, + 283 + ], + "lines": [ + { + "bbox": [ + 133, + 212, + 239, + 283 + ], + "spans": [ + { + "bbox": [ + 133, + 212, + 239, + 283 + ], + "type": "image", + "image_path": "06be440aeacb2afda4aa9bd81a386182d56eea84dc38e5aae79daaad1061b530.jpg" + } + ] + } + ], + "index": 7, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 162, + 285, + 210, + 293 + ], + "lines": [ + { + "bbox": [ + 162, + 285, + 210, + 293 + ], + "spans": [ + { + "bbox": [ + 162, + 285, + 210, + 293 + ], + "type": "text", + "content": "A metal bucket" + } + ] + } + ], + "index": 8, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 130, + 300, + 482, + 346 + ], + "lines": [ + { + "bbox": [ + 130, + 300, + 482, + 346 + ], + "spans": [ + { + "bbox": [ + 130, + 300, + 482, + 346 + ], + "type": "text", + "content": "Fig.1: We propose DiPIR, a physically based method to recover lighting from a single image, enabling arbitrary virtual object compositing into indoor and outdoor scenes, as well as material and tone-mapping optimization. Project page: https://research.nvidia.com/labs/toronto-ai/DiPIR/" + } + ] + } + ], + "index": 16, + "angle": 0, + "type": "image_caption" + } + ], + "index": 7 + }, + { + "type": "image", + "bbox": [ + 243, + 212, + 349, + 283 + ], + "blocks": [ + { + "bbox": [ + 243, + 212, + 349, + 283 + ], + "lines": [ + { + "bbox": [ + 243, + 212, + 349, + 283 + ], + "spans": [ + { + "bbox": [ + 243, + 212, + 349, + 283 + ], + "type": "image", + "image_path": "02d163547fd641c2d36da6ca0f8d4e05bdbd5c0d796c839ee7913b8a6cd206c8.jpg" + } + ] + } + ], + "index": 9, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 269, + 285, + 323, + 293 + ], + "lines": [ + { + "bbox": [ + 269, + 285, + 323, + 293 + ], + "spans": [ + { + "bbox": [ + 269, + 285, + 323, + 293 + ], + "type": "text", + "content": "A white SUV car" + } + ] + } + ], + "index": 10, + "angle": 0, + "type": "image_caption" + } + ], + "index": 9 + }, + { + "type": "image", + "bbox": [ + 369, + 129, + 474, + 200 + ], + "blocks": [ + { + "bbox": [ + 395, + 118, + 447, + 129 + ], + "lines": [ + { + "bbox": [ + 395, + 118, + 447, + 129 + ], + "spans": [ + { + "bbox": [ + 395, + 118, + 447, + 129 + ], + "type": "text", + "content": "Applications" + } + ] + } + ], + "index": 11, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 369, + 129, + 474, + 200 + ], + "lines": [ + { + "bbox": [ + 369, + 129, + 474, + 200 + ], + "spans": [ + { + "bbox": [ + 369, + 129, + 474, + 200 + ], + "type": "image", + "image_path": "8efe3259dafa467b6a2b948151535c18b85b91cadd43cbfebe3ffdc3caf2a4e0.jpg" + } + ] + } + ], + "index": 12, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 386, + 201, + 457, + 210 + ], + "lines": [ + { + "bbox": [ + 386, + 201, + 457, + 210 + ], + "spans": [ + { + "bbox": [ + 386, + 201, + 457, + 210 + ], + "type": "text", + "content": "Insert multiple objects" + } + ] + } + ], + "index": 13, + "angle": 0, + "type": "image_caption" + } + ], + "index": 12 + }, + { + "type": "image", + "bbox": [ + 368, + 212, + 474, + 283 + ], + "blocks": [ + { + "bbox": [ + 368, + 212, + 474, + 283 + ], + "lines": [ + { + "bbox": [ + 368, + 212, + 474, + 283 + ], + "spans": [ + { + "bbox": [ + 368, + 212, + 474, + 283 + ], + "type": "image", + "image_path": "036680c1483f31be9a9d096e0efa3ebb30ad26a8be0e2345049567543452e6ae.jpg" + } + ] + } + ], + "index": 14, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 365, + 284, + 476, + 294 + ], + "lines": [ + { + "bbox": [ + 365, + 284, + 476, + 294 + ], + "spans": [ + { + "bbox": [ + 365, + 284, + 476, + 294 + ], + "type": "text", + "content": "Optimize material & tone mapping" + } + ] + } + ], + "index": 15, + "angle": 0, + "type": "image_caption" + } + ], + "index": 14 + }, + { + "bbox": [ + 130, + 354, + 482, + 449 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 354, + 482, + 449 + ], + "spans": [ + { + "bbox": [ + 130, + 354, + 482, + 449 + ], + "type": "text", + "content": "To address these challenges, we propose to reuse the strong image generation priors learned by large diffusion models (DMs) [46] as a guidance for inverse rendering. Unlike hand-crafted or supervised data-driven priors that are often specific to a domain, DMs are trained on massive datasets and show a remarkable \"understanding\" of the world and the underlying physical concepts. While DMs still often fail to produce accurate lighting effects such as shadows and reflections [49] in their generations, we observe that they can provide valuable guidance when combined with a physically-based renderer and are adapted to the scene." + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 130, + 449, + 482, + 544 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 449, + 482, + 544 + ], + "spans": [ + { + "bbox": [ + 130, + 449, + 482, + 544 + ], + "type": "text", + "content": "Specifically, we present Diffusion Prior for Inverse Rendering (DiPIR), building on three main contributions. First, we use a physically based renderer to accurately simulate the interaction between the light and the 3D asset to generate the final composited image. We also account for the unknown tone-mapping curve to mimic the camera sensor response. Second, we propose a lightweight personalization scheme of the pre-trained DM, based on the input image and the type of inserted asset. Third, we design a variant of the SDS loss [45] which makes use of this personalization and improves training stability." + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 130, + 544, + 482, + 605 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 544, + 482, + 605 + ], + "spans": [ + { + "bbox": [ + 130, + 544, + 482, + 605 + ], + "type": "text", + "content": "In DiPIR, the DM acts similarly to a human evaluator. It takes the edited image as input and propagates the feedback signal to physically-based scene attributes via differentiable rendering, thus enabling end-to-end optimization. We experimentally show that DiPIR outperforms existing state-of-the-art lighting estimation methods for object insertion across indoor and outdoor datasets." + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 132, + 620, + 238, + 634 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 620, + 238, + 634 + ], + "spans": [ + { + "bbox": [ + 132, + 620, + 238, + 634 + ], + "type": "text", + "content": "2 Related Work" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 130, + 641, + 482, + 666 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 641, + 482, + 666 + ], + "spans": [ + { + "bbox": [ + 130, + 641, + 482, + 666 + ], + "type": "text", + "content": "Inverse rendering is the task of recovering intrinsic properties of a scene, including materials, shape, and lighting, from a single or multiple images [5]." + } + ] + } + ], + "index": 21 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "spans": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "type": "text", + "content": "2" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 166, + 91, + 227, + 102 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 166, + 91, + 227, + 102 + ], + "spans": [ + { + "bbox": [ + 166, + 91, + 227, + 102 + ], + "type": "text", + "content": "R. Liang et al." + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 1 + }, + { + "para_blocks": [ + { + "bbox": [ + 130, + 116, + 482, + 163 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 116, + 482, + 163 + ], + "spans": [ + { + "bbox": [ + 130, + 116, + 482, + 163 + ], + "type": "text", + "content": "The main challenge of inverse rendering lies in the ill-posed nature of the task. Prior knowledge about materials and lighting effects is crucial to solving under-constrained inverse rendering problems, but how to best define and incorporate such priors remains an open research question." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 130, + 164, + 482, + 320 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 164, + 482, + 320 + ], + "spans": [ + { + "bbox": [ + 130, + 164, + 482, + 320 + ], + "type": "text", + "content": "Early methods formulated inverse rendering as an optimization problem and used hand-crafted regularization terms such as local smoothness and sparsity of materials [4,5,9,18,31,75] to constrain the solution space. However, real-world lighting effects are often complex and difficult to describe using hand-crafted priors. The advent of deep learning has facilitated the learning of data-driven priors from ground truth supervision [6,8,30,34,35,38,50,62-64,69]. Yet, acquiring real-world data with accurately annotated intrinsic decomposition necessitates specialized capture devices [18] or expert annotators [6,30]. Collecting real-world datasets at a large scale is therefore extremely challenging. As a result, existing methods often resort to synthetic datasets [17,35,50,56,63,77], or are carefully designed to be training data-efficient [22,23,62]. The resulting algorithms are therefore often tailored to specific domains, such as indoor [15-17, 50, 63] or day-time outdoor scenes [22,23,56,62,77]." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 130, + 324, + 482, + 529 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 324, + 482, + 529 + ], + "spans": [ + { + "bbox": [ + 130, + 324, + 482, + 529 + ], + "type": "text", + "content": "Lighting estimation is a subtask of inverse rendering that specifically focuses on inferring lighting and is a core component of photorealistic virtual object insertion pipelines. Existing lighting estimation methods are often designed as feed-forward neural networks that take a single image as input and directly regress lighting information in form of spherical lobes [17,35,70,76], environment maps [16,50,53,56,77], parametric light sources [15], or volumetric lighting [37, 54,62,63]. The environment lighting can also be recovered through the proxy geometry shown in the single image [32,67]. Recently, lighting estimation was also formulated as a generative task. StyleLight [61] is based on a dual-branch StyleGAN model, which first performs GAN inversion through the LDR branch and then generates the panorama with the HDR branch. EverLight [12] instead regresses an initial HDR environment map from the input image and later refines it with a GAN model [27]. Concurrent to our work, DiffusionLight [44] proposed to inpaint a chrome ball in the center of the given image using an inpainting diffusion model and unwarp it to an environment map. To obtain the HDR environment map, DiffusionLight trains a LoRA [24] for exposure bracketing and merges multiple generated LDR chrome balls with different exposures." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 130, + 529, + 483, + 577 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 529, + 483, + 577 + ], + "spans": [ + { + "bbox": [ + 130, + 529, + 483, + 577 + ], + "type": "text", + "content": "Note that lighting is fundamentally a High Dynamic Range (HDR) quantity, and reproducing this full range still remains challenging. HDR is crucial for basic effects such as sharp shadows, particularly in outdoor day-time settings where the sun is several orders of magnitudes brighter than the rest of the environment." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 130, + 582, + 482, + 666 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 582, + 482, + 666 + ], + "spans": [ + { + "bbox": [ + 130, + 582, + 482, + 666 + ], + "type": "text", + "content": "Physically-based differentiable rendering. Rendering algorithms such as path tracing [26] aim to produce realistic images of virtual scenes by accurately simulating the physical processes involved in light transport. Physically-based rendering (PBR) typically uses a geometric representation of the scene, scattering functions modeling the behavior of the scene's surfaces, as well as camera and lighting models. Recent years have seen the development of many differentiable rendering methods and frameworks [25, 33, 42, 43, 60, 71, 72]." + } + ] + } + ], + "index": 6 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 153, + 91, + 447, + 103 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 153, + 91, + 447, + 103 + ], + "spans": [ + { + "bbox": [ + 153, + 91, + 447, + 103 + ], + "type": "text", + "content": "Photorealistic Object Insertion with Diffusion-Guided Inverse Rendering" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 474, + 92, + 480, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 474, + 92, + 480, + 100 + ], + "spans": [ + { + "bbox": [ + 474, + 92, + 480, + 100 + ], + "type": "text", + "content": "3" + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 2 + }, + { + "para_blocks": [ + { + "bbox": [ + 130, + 116, + 482, + 164 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 116, + 482, + 164 + ], + "spans": [ + { + "bbox": [ + 130, + 116, + 482, + 164 + ], + "type": "text", + "content": "While particular attention has been devoted to the difficult problem of computing derivatives at discontinuities [3, 33, 39, 65, 74], such as due to visibility changes, we focus here on the comparatively simpler lighting, material and tone-mapping derivatives, which generally do not introduce discontinuities." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 130, + 169, + 483, + 326 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 169, + 483, + 326 + ], + "spans": [ + { + "bbox": [ + 130, + 169, + 483, + 326 + ], + "type": "text", + "content": "Diffusion model priors and personalization. Priors learned by the text-to-image DMs [2,11,46,48], on large-scale datasets can be adapted to various applications such as monocular depth estimation [28], authentic image completion [57], and image restoration [10]. Among various adaptation techniques, DreamBooth [47] proposes to finetune the DM on subject images, while Textual Inversion [14] optimizes new word embeddings for target generation. LoRA [24,51] simplifies adaptation by freezing the pretrained model weights and injecting trainable rank decomposition matrices, thus greatly reducing the number of trainable parameters. Most similar to ours are concurrent works on DM adaptation for lighting estimation [44] and intrinsic image decomposition [29]. The former adapts the DM to exposure bracketing for LDR to HDR lifting, while the latter fine-tunes a pretrained SD model to directly output the albedo and BRDF properties of an image." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 132, + 342, + 234, + 356 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 342, + 234, + 356 + ], + "spans": [ + { + "bbox": [ + 132, + 342, + 234, + 356 + ], + "type": "text", + "content": "3 Preliminaries" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 130, + 367, + 482, + 464 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 367, + 482, + 464 + ], + "spans": [ + { + "bbox": [ + 130, + 367, + 482, + 464 + ], + "type": "text", + "content": "Diffusion Models. Diffusion models are a family of generative models built around two key processes. A forward process, which gradually adds noise to data samples " + }, + { + "bbox": [ + 130, + 367, + 482, + 464 + ], + "type": "inline_equation", + "content": "\\mathbf{x} \\sim p(\\mathbf{x})" + }, + { + "bbox": [ + 130, + 367, + 482, + 464 + ], + "type": "text", + "content": " removing their structure over time " + }, + { + "bbox": [ + 130, + 367, + 482, + 464 + ], + "type": "inline_equation", + "content": "t" + }, + { + "bbox": [ + 130, + 367, + 482, + 464 + ], + "type": "text", + "content": ". This is achieved using a noise schedule determined by " + }, + { + "bbox": [ + 130, + 367, + 482, + 464 + ], + "type": "inline_equation", + "content": "\\alpha_{t}" + }, + { + "bbox": [ + 130, + 367, + 482, + 464 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 130, + 367, + 482, + 464 + ], + "type": "inline_equation", + "content": "\\sigma_{t}" + }, + { + "bbox": [ + 130, + 367, + 482, + 464 + ], + "type": "text", + "content": " as " + }, + { + "bbox": [ + 130, + 367, + 482, + 464 + ], + "type": "inline_equation", + "content": "\\mathbf{x}_{t} = \\alpha_{t}\\mathbf{x} + \\sigma_{t}\\boldsymbol{\\epsilon}, \\boldsymbol{\\epsilon} \\sim \\mathcal{N}(\\mathbf{0},\\mathbf{I})" + }, + { + "bbox": [ + 130, + 367, + 482, + 464 + ], + "type": "text", + "content": ". In contrast, the reverse process gradually removes this noise, restoring the structure. The reverse process is parameterized by a conditional neural network, " + }, + { + "bbox": [ + 130, + 367, + 482, + 464 + ], + "type": "inline_equation", + "content": "\\epsilon_{\\theta}" + }, + { + "bbox": [ + 130, + 367, + 482, + 464 + ], + "type": "text", + "content": ", trained to predict the noise " + }, + { + "bbox": [ + 130, + 367, + 482, + 464 + ], + "type": "inline_equation", + "content": "\\epsilon" + }, + { + "bbox": [ + 130, + 367, + 482, + 464 + ], + "type": "text", + "content": " at a given timestep " + }, + { + "bbox": [ + 130, + 367, + 482, + 464 + ], + "type": "inline_equation", + "content": "t" + }, + { + "bbox": [ + 130, + 367, + 482, + 464 + ], + "type": "text", + "content": " according to the following simplified objective [20]:" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 205, + 471, + 482, + 486 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 205, + 471, + 482, + 486 + ], + "spans": [ + { + "bbox": [ + 205, + 471, + 482, + 486 + ], + "type": "interline_equation", + "content": "\\mathbb {E} _ {\\mathbf {x} \\sim p (\\mathbf {x}), \\boldsymbol {\\epsilon} \\sim \\mathcal {N} (\\mathbf {0}, \\boldsymbol {I}), t \\sim T} \\left[ w (t) | | \\boldsymbol {\\epsilon} _ {\\boldsymbol {\\theta}} (\\mathbf {x} _ {t}, t, \\boldsymbol {c}) - \\boldsymbol {\\epsilon} | | _ {2} ^ {2} \\right], \\tag {1}", + "image_path": "63f07189617e0a22e7ed4b6ae707424c24cc1e5903d97a1be1f5fcdda5e1682d.jpg" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 130, + 492, + 482, + 528 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 492, + 482, + 528 + ], + "spans": [ + { + "bbox": [ + 130, + 492, + 482, + 528 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 130, + 492, + 482, + 528 + ], + "type": "inline_equation", + "content": "c" + }, + { + "bbox": [ + 130, + 492, + 482, + 528 + ], + "type": "text", + "content": " represents a condition (e.g. text, image, etc.) that allows controlling the generation process, " + }, + { + "bbox": [ + 130, + 492, + 482, + 528 + ], + "type": "inline_equation", + "content": "w(t)" + }, + { + "bbox": [ + 130, + 492, + 482, + 528 + ], + "type": "text", + "content": " represents a time-conditional weighting, and " + }, + { + "bbox": [ + 130, + 492, + 482, + 528 + ], + "type": "inline_equation", + "content": "T" + }, + { + "bbox": [ + 130, + 492, + 482, + 528 + ], + "type": "text", + "content": " is a set containing a selection of timesteps." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 130, + 529, + 483, + 590 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 529, + 483, + 590 + ], + "spans": [ + { + "bbox": [ + 130, + 529, + 483, + 590 + ], + "type": "text", + "content": "In this work, we use a Latent Diffusion Model (LDM) [46] in which the diffusion process is conducted in a lower-dimensional latent space. Specifically, the encoder " + }, + { + "bbox": [ + 130, + 529, + 483, + 590 + ], + "type": "inline_equation", + "content": "\\mathcal{E}" + }, + { + "bbox": [ + 130, + 529, + 483, + 590 + ], + "type": "text", + "content": " maps samples from the data distribution " + }, + { + "bbox": [ + 130, + 529, + 483, + 590 + ], + "type": "inline_equation", + "content": "\\mathbf{x} \\sim p(\\mathbf{x})" + }, + { + "bbox": [ + 130, + 529, + 483, + 590 + ], + "type": "text", + "content": " into a latent space " + }, + { + "bbox": [ + 130, + 529, + 483, + 590 + ], + "type": "inline_equation", + "content": "\\mathbb{Z}" + }, + { + "bbox": [ + 130, + 529, + 483, + 590 + ], + "type": "text", + "content": ". The decoder " + }, + { + "bbox": [ + 130, + 529, + 483, + 590 + ], + "type": "inline_equation", + "content": "\\mathcal{D}" + }, + { + "bbox": [ + 130, + 529, + 483, + 590 + ], + "type": "text", + "content": " performs the inverse operation, such that " + }, + { + "bbox": [ + 130, + 529, + 483, + 590 + ], + "type": "inline_equation", + "content": "\\mathcal{D}(\\mathcal{E}(\\mathbf{x})) \\approx \\mathbf{x}" + }, + { + "bbox": [ + 130, + 529, + 483, + 590 + ], + "type": "text", + "content": ". As follows, for LDMs " + }, + { + "bbox": [ + 130, + 529, + 483, + 590 + ], + "type": "inline_equation", + "content": "\\mathbf{x}" + }, + { + "bbox": [ + 130, + 529, + 483, + 590 + ], + "type": "text", + "content": " in Eq. 1 is replaced by its latent " + }, + { + "bbox": [ + 130, + 529, + 483, + 590 + ], + "type": "inline_equation", + "content": "\\mathbf{z} = \\mathcal{E}(\\mathbf{x})" + }, + { + "bbox": [ + 130, + 529, + 483, + 590 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 130, + 605, + 482, + 667 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 605, + 482, + 667 + ], + "spans": [ + { + "bbox": [ + 130, + 605, + 482, + 667 + ], + "type": "text", + "content": "Personalization and Fine-tuning. Fine-tuning all parameters of a pretrained DM requires significant computational resources and time. To alleviate this, LoRA [24] injects trainable low-rank decomposition matrices and aims to learn only the variations from the pretrained weights. Specifically, consider a linear layer represented as " + }, + { + "bbox": [ + 130, + 605, + 482, + 667 + ], + "type": "inline_equation", + "content": "\\pmb{h} = \\pmb{W}_0\\pmb{x}" + }, + { + "bbox": [ + 130, + 605, + 482, + 667 + ], + "type": "text", + "content": ". Here " + }, + { + "bbox": [ + 130, + 605, + 482, + 667 + ], + "type": "inline_equation", + "content": "\\pmb{W}_0 \\in \\mathbb{R}^{n \\times n}" + }, + { + "bbox": [ + 130, + 605, + 482, + 667 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 130, + 605, + 482, + 667 + ], + "type": "inline_equation", + "content": "\\pmb{x} \\in \\mathbb{R}^n" + }, + { + "bbox": [ + 130, + 605, + 482, + 667 + ], + "type": "text", + "content": " are the pretrained" + } + ] + } + ], + "index": 9 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "spans": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "type": "text", + "content": "4" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 166, + 91, + 227, + 102 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 166, + 91, + 227, + 102 + ], + "spans": [ + { + "bbox": [ + 166, + 91, + 227, + 102 + ], + "type": "text", + "content": "R. Liang et al." + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 3 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 132, + 114, + 403, + 252 + ], + "blocks": [ + { + "bbox": [ + 132, + 114, + 403, + 252 + ], + "lines": [ + { + "bbox": [ + 132, + 114, + 403, + 252 + ], + "spans": [ + { + "bbox": [ + 132, + 114, + 403, + 252 + ], + "type": "image", + "image_path": "e37dbec6ed4ca66a42ced518975652ead35e15ea9dfb8476f76d5260b2b2cb5b.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 130, + 256, + 482, + 356 + ], + "lines": [ + { + "bbox": [ + 130, + 256, + 482, + 356 + ], + "spans": [ + { + "bbox": [ + 130, + 256, + 482, + 356 + ], + "type": "text", + "content": "Fig. 2: Method overview. Given an input image, we first construct a virtual 3D scene with a virtual object and proxy plane. Our physically-based renderer then differentiably simulates the interactions of the estimizable environment map with the inserted virtual object and its effect on the background scene (shadowing) (left). At each iteration, the rendered image is diffused and passed through a personalized diffusion model (middle). The gradient of the adapted Score Distillation formulation is propagated back to the environment map and the tone-mapping curve through the differentiable renderer. Upon convergence, we recover lighting and tone-mapping parameters, which allow photorealistic compositing of virtual objects from a single image (right)." + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "image_caption" + } + ], + "index": 2 + }, + { + "type": "image", + "bbox": [ + 410, + 114, + 477, + 252 + ], + "blocks": [ + { + "bbox": [ + 410, + 114, + 477, + 252 + ], + "lines": [ + { + "bbox": [ + 410, + 114, + 477, + 252 + ], + "spans": [ + { + "bbox": [ + 410, + 114, + 477, + 252 + ], + "type": "image", + "image_path": "9642069bf67e8f1ea1bde726c288ba8dee1cad210dbc647aee5ec017040eb035.jpg" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_body" + } + ], + "index": 3 + }, + { + "bbox": [ + 130, + 369, + 482, + 430 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 369, + 482, + 430 + ], + "spans": [ + { + "bbox": [ + 130, + 369, + 482, + 430 + ], + "type": "text", + "content": "weights and input, respectively. Applying LoRA, this layer is modified to " + }, + { + "bbox": [ + 130, + 369, + 482, + 430 + ], + "type": "inline_equation", + "content": "\\pmb{h} = (W_0 + \\pmb{\\Delta W})\\pmb{x} = W_0\\pmb{x} + AB\\pmb{x}" + }, + { + "bbox": [ + 130, + 369, + 482, + 430 + ], + "type": "text", + "content": ". Notably, " + }, + { + "bbox": [ + 130, + 369, + 482, + 430 + ], + "type": "inline_equation", + "content": "\\pmb{\\Delta W} = \\pmb{A}\\pmb{B}" + }, + { + "bbox": [ + 130, + 369, + 482, + 430 + ], + "type": "text", + "content": " is a combination of low-rank matrices " + }, + { + "bbox": [ + 130, + 369, + 482, + 430 + ], + "type": "inline_equation", + "content": "\\pmb{A} \\in \\mathbb{R}^{n \\times r}" + }, + { + "bbox": [ + 130, + 369, + 482, + 430 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 130, + 369, + 482, + 430 + ], + "type": "inline_equation", + "content": "\\pmb{B} \\in \\mathbb{R}^{r \\times n}" + }, + { + "bbox": [ + 130, + 369, + 482, + 430 + ], + "type": "text", + "content": " with " + }, + { + "bbox": [ + 130, + 369, + 482, + 430 + ], + "type": "inline_equation", + "content": "r \\ll n" + }, + { + "bbox": [ + 130, + 369, + 482, + 430 + ], + "type": "text", + "content": ". During LoRA fine-tuning, only the added term " + }, + { + "bbox": [ + 130, + 369, + 482, + 430 + ], + "type": "inline_equation", + "content": "\\pmb{\\Delta W}" + }, + { + "bbox": [ + 130, + 369, + 482, + 430 + ], + "type": "text", + "content": " is updated while the parameters " + }, + { + "bbox": [ + 130, + 369, + 482, + 430 + ], + "type": "inline_equation", + "content": "W_0" + }, + { + "bbox": [ + 130, + 369, + 482, + 430 + ], + "type": "text", + "content": " remain unchanged. The rank " + }, + { + "bbox": [ + 130, + 369, + 482, + 430 + ], + "type": "inline_equation", + "content": "r" + }, + { + "bbox": [ + 130, + 369, + 482, + 430 + ], + "type": "text", + "content": " is a hyperparameter that trades efficiency with model capacity." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 132, + 447, + 202, + 459 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 447, + 202, + 459 + ], + "spans": [ + { + "bbox": [ + 132, + 447, + 202, + 459 + ], + "type": "text", + "content": "4 Method" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 130, + 471, + 482, + 533 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 471, + 482, + 533 + ], + "spans": [ + { + "bbox": [ + 130, + 471, + 482, + 533 + ], + "type": "text", + "content": "Given a single image as input, DiPIR recovers scene lighting and tone-mapping parameters, with the goal of photorealistic insertion of virtual objects. An overview of our method is shown in Fig. 2. Sec. 4.1 describes our representation and the differentiable rendering process, while Sec. 4.2 and Sec. 4.3 provide details on diffusion model guidance and optimization formulation, respectively." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 131, + 539, + 371, + 551 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 131, + 539, + 371, + 551 + ], + "spans": [ + { + "bbox": [ + 131, + 539, + 371, + 551 + ], + "type": "text", + "content": "4.1 Physically-based Virtual Object Insertion" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 130, + 555, + 482, + 614 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 555, + 482, + 614 + ], + "spans": [ + { + "bbox": [ + 130, + 555, + 482, + 614 + ], + "type": "text", + "content": "Virtual scene. Inserting a virtual object " + }, + { + "bbox": [ + 130, + 555, + 482, + 614 + ], + "type": "inline_equation", + "content": "\\mathcal{X}" + }, + { + "bbox": [ + 130, + 555, + 482, + 614 + ], + "type": "text", + "content": " into an image " + }, + { + "bbox": [ + 130, + 555, + 482, + 614 + ], + "type": "inline_equation", + "content": "\\mathbf{I}_{\\mathrm{bg}} \\in \\mathbb{R}^{h \\times w \\times 3}" + }, + { + "bbox": [ + 130, + 555, + 482, + 614 + ], + "type": "text", + "content": " requires creating a 3D proxy virtual scene, viewed from the correct camera pose. Here, we assume that the user provides a specific placement (pose) for " + }, + { + "bbox": [ + 130, + 555, + 482, + 614 + ], + "type": "inline_equation", + "content": "\\mathcal{X}" + }, + { + "bbox": [ + 130, + 555, + 482, + 614 + ], + "type": "text", + "content": ", but in some cases, an appropriate pose can also be determined automatically, e.g. by detecting the floor plane and scene scale." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 130, + 615, + 482, + 675 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 615, + 482, + 675 + ], + "spans": [ + { + "bbox": [ + 130, + 615, + 482, + 675 + ], + "type": "text", + "content": "To model the effects of the inserted object on the original image, such as shadows cast by " + }, + { + "bbox": [ + 130, + 615, + 482, + 675 + ], + "type": "inline_equation", + "content": "\\mathcal{X}" + }, + { + "bbox": [ + 130, + 615, + 482, + 675 + ], + "type": "text", + "content": ", we also assume a known proxy geometry " + }, + { + "bbox": [ + 130, + 615, + 482, + 675 + ], + "type": "inline_equation", + "content": "\\mathcal{P}" + }, + { + "bbox": [ + 130, + 615, + 482, + 675 + ], + "type": "text", + "content": ". We found a simple ground plane acting as a shadow catcher underneath the virtual object to be sufficient in all of our experiments. This proxy plane can be easily placed manually or automatically generated based on e.g. depth or LiDAR data." + } + ] + } + ], + "index": 10 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 153, + 91, + 448, + 102 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 153, + 91, + 448, + 102 + ], + "spans": [ + { + "bbox": [ + 153, + 91, + 448, + 102 + ], + "type": "text", + "content": "Photorealistic Object Insertion with Diffusion-Guided Inverse Rendering" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 474, + 92, + 480, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 474, + 92, + 480, + 100 + ], + "spans": [ + { + "bbox": [ + 474, + 92, + 480, + 100 + ], + "type": "text", + "content": "5" + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 4 + }, + { + "para_blocks": [ + { + "bbox": [ + 130, + 116, + 482, + 152 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 116, + 482, + 152 + ], + "spans": [ + { + "bbox": [ + 130, + 116, + 482, + 152 + ], + "type": "text", + "content": "Light representation. We represent the scene's lighting " + }, + { + "bbox": [ + 130, + 116, + 482, + 152 + ], + "type": "inline_equation", + "content": "\\mathbf{L}" + }, + { + "bbox": [ + 130, + 116, + 482, + 152 + ], + "type": "text", + "content": " with a set of " + }, + { + "bbox": [ + 130, + 116, + 482, + 152 + ], + "type": "inline_equation", + "content": "N" + }, + { + "bbox": [ + 130, + 116, + 482, + 152 + ], + "type": "text", + "content": " optizable Spherical Gaussian (SG) parameters " + }, + { + "bbox": [ + 130, + 116, + 482, + 152 + ], + "type": "inline_equation", + "content": "\\{\\mathbf{c}_k,\\pmb {\\mu}_k,\\sigma_k\\}_{k = 1}^N\\in \\mathbb{R}^{N\\times 7}" + }, + { + "bbox": [ + 130, + 116, + 482, + 152 + ], + "type": "text", + "content": ", where the radiance for one SG lobe at the direction " + }, + { + "bbox": [ + 130, + 116, + 482, + 152 + ], + "type": "inline_equation", + "content": "\\pmb {v}\\in \\mathbb{R}^3" + }, + { + "bbox": [ + 130, + 116, + 482, + 152 + ], + "type": "text", + "content": " is defined as" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 149, + 156, + 482, + 172 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 149, + 156, + 482, + 172 + ], + "spans": [ + { + "bbox": [ + 149, + 156, + 482, + 172 + ], + "type": "interline_equation", + "content": "\\mathbf {G} _ {k} (\\pmb {v}; \\mathbf {c} _ {k}, \\pmb {\\mu} _ {k}, \\sigma_ {k}) = \\mathbf {c} e ^ {- (1 - \\pmb {v} \\cdot \\pmb {\\mu}) / \\sigma^ {2}}, \\quad \\mathrm {w h e r e} \\mathbf {c} _ {k} \\in \\mathbb {R} ^ {3}, \\pmb {\\mu} _ {k} \\in \\mathbb {R} ^ {3}, \\sigma_ {k} \\in \\mathbb {R} _ {+}. \\quad (2)", + "image_path": "3197b295eff2b7a8cb367fdcf3858f8ad493e6291ed0f938a5939ccfcf3d2af4.jpg" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 131, + 176, + 398, + 190 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 131, + 176, + 398, + 190 + ], + "spans": [ + { + "bbox": [ + 131, + 176, + 398, + 190 + ], + "type": "text", + "content": "The overall environment map " + }, + { + "bbox": [ + 131, + 176, + 398, + 190 + ], + "type": "inline_equation", + "content": "\\mathbf{L} \\in \\mathbb{R}^{H \\times W \\times 3}" + }, + { + "bbox": [ + 131, + 176, + 398, + 190 + ], + "type": "text", + "content": " is computed as:" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 239, + 194, + 482, + 228 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 239, + 194, + 482, + 228 + ], + "spans": [ + { + "bbox": [ + 239, + 194, + 482, + 228 + ], + "type": "interline_equation", + "content": "\\mathbf {L} _ {i, j} = \\sum_ {k = 1} ^ {N} \\mathbf {G} _ {k} \\left(\\boldsymbol {v} _ {i, j}; \\mathbf {c} _ {k}, \\boldsymbol {\\mu} _ {k}, \\sigma_ {k}\\right), \\tag {3}", + "image_path": "23b223f9a5ff38ea50d131e1d4826821cc2e0113b58db1d9e8937693bc01d4b2.jpg" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 130, + 232, + 482, + 281 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 232, + 482, + 281 + ], + "spans": [ + { + "bbox": [ + 130, + 232, + 482, + 281 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 130, + 232, + 482, + 281 + ], + "type": "inline_equation", + "content": "\\pmb{v}_{i,j}" + }, + { + "bbox": [ + 130, + 232, + 482, + 281 + ], + "type": "text", + "content": " is the direction corresponding to the pixel " + }, + { + "bbox": [ + 130, + 232, + 482, + 281 + ], + "type": "inline_equation", + "content": "(i,j)" + }, + { + "bbox": [ + 130, + 232, + 482, + 281 + ], + "type": "text", + "content": " using the standard spherical environment parameterization. Note that we chose a SG-based lighting representation for its good convergence properties and simplicity, but many alternatives exist and might be combined with our method." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 130, + 285, + 482, + 334 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 285, + 482, + 334 + ], + "spans": [ + { + "bbox": [ + 130, + 285, + 482, + 334 + ], + "type": "text", + "content": "Differentiable rendering. Inserting a virtual object into a scene involves simulating the interactions of the estimizable environment map with the inserted virtual object (foreground), and the inserted object's effect on the background scene (shadow). We describe below the details respectively." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 131, + 341, + 482, + 378 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 131, + 341, + 482, + 378 + ], + "spans": [ + { + "bbox": [ + 131, + 341, + 482, + 378 + ], + "type": "text", + "content": "Foreground image. Given the optimized light representation " + }, + { + "bbox": [ + 131, + 341, + 482, + 378 + ], + "type": "inline_equation", + "content": "\\mathbf{L}" + }, + { + "bbox": [ + 131, + 341, + 482, + 378 + ], + "type": "text", + "content": " from above and the virtual object " + }, + { + "bbox": [ + 131, + 341, + 482, + 378 + ], + "type": "inline_equation", + "content": "\\mathcal{X}" + }, + { + "bbox": [ + 131, + 341, + 482, + 378 + ], + "type": "text", + "content": " (including its geometry and materials), a foreground image " + }, + { + "bbox": [ + 131, + 341, + 482, + 378 + ], + "type": "inline_equation", + "content": "\\mathbf{I}_{\\mathrm{fg}}" + }, + { + "bbox": [ + 131, + 341, + 482, + 378 + ], + "type": "text", + "content": " of the inserted object can be rendered directly using standard path tracing:" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 249, + 384, + 482, + 397 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 249, + 384, + 482, + 397 + ], + "spans": [ + { + "bbox": [ + 249, + 384, + 482, + 397 + ], + "type": "interline_equation", + "content": "\\mathbf {I} _ {\\mathrm {f g}} = \\operatorname {P a t h T r a c e} (\\mathcal {X}, \\mathbf {L}, D), \\tag {4}", + "image_path": "cb4b27237819afcf48560020eec4f7ea9a143641ab2f5a7b595d539a9ccba718.jpg" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 130, + 402, + 482, + 451 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 402, + 482, + 451 + ], + "spans": [ + { + "bbox": [ + 130, + 402, + 482, + 451 + ], + "type": "inline_equation", + "content": "D" + }, + { + "bbox": [ + 130, + 402, + 482, + 451 + ], + "type": "text", + "content": " is the maximum number of interactions along the light path. Since we do not assume that the provided proxy geometry is accurate nor that it has materials, we omit the effect of light reflected by the object into the scene, for example, due to a highly specular inserted object." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 130, + 460, + 482, + 506 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 460, + 482, + 506 + ], + "spans": [ + { + "bbox": [ + 130, + 460, + 482, + 506 + ], + "type": "text", + "content": "Shadow ratio. As the background image " + }, + { + "bbox": [ + 130, + 460, + 482, + 506 + ], + "type": "inline_equation", + "content": "\\mathbf{I}_{\\mathrm{bg}}" + }, + { + "bbox": [ + 130, + 460, + 482, + 506 + ], + "type": "text", + "content": " already faithfully represents the scene, our task is limited to simulating the appearance of the inserted object and its effect on the nearby parts of the scene, i.e. the shadows cast by the object onto the proxy geometry." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 130, + 506, + 482, + 555 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 506, + 482, + 555 + ], + "spans": [ + { + "bbox": [ + 130, + 506, + 482, + 555 + ], + "type": "text", + "content": "The shadow ratio " + }, + { + "bbox": [ + 130, + 506, + 482, + 555 + ], + "type": "inline_equation", + "content": "\\beta_{\\mathrm{shadow}} \\in \\mathbb{R}^{h \\times w \\times 3}" + }, + { + "bbox": [ + 130, + 506, + 482, + 555 + ], + "type": "text", + "content": ", which accounts for the effect of the object on the surrounding scene, is computed as the ratio between the radiance received by the proxy geometry with and without the inserted object present. A low value of " + }, + { + "bbox": [ + 130, + 506, + 482, + 555 + ], + "type": "inline_equation", + "content": "\\beta_{\\mathrm{shadow}}" + }, + { + "bbox": [ + 130, + 506, + 482, + 555 + ], + "type": "text", + "content": " indicates a strongly shadowed region:" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 230, + 560, + 482, + 586 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 230, + 560, + 482, + 586 + ], + "spans": [ + { + "bbox": [ + 230, + 560, + 482, + 586 + ], + "type": "interline_equation", + "content": "\\boldsymbol {\\beta} _ {\\mathrm {s h a d o w}} = \\frac {\\operatorname {P a t h T r a c e} (\\mathcal {X} \\cup \\mathcal {P} , \\mathbf {L} , 1)}{\\operatorname {P a t h T r a c e} (\\mathcal {P} , \\mathbf {L} , 1)}, \\tag {5}", + "image_path": "f6e966044b169254f927275118ddd9dde6f6a9f04b57365dcc08e1dc02ae209a.jpg" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 130, + 590, + 482, + 675 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 590, + 482, + 675 + ], + "spans": [ + { + "bbox": [ + 130, + 590, + 482, + 675 + ], + "type": "text", + "content": "This is similar to the method of Wang et al. [62], however we use Multiple Importance Sampling (MIS) [59] between lighting and BSDF for better sampling efficiency. The maximum path length has been limited to 1 interaction in order to reduce memory and computational cost. Unless specific material information is provided with the proxy geometry, we use a Lambertian BSDF, in which case the albedo cancels out in the ratio. In practice, the computation of the shadow ratio and foreground image are combined in order to reuse shared terms." + } + ] + } + ], + "index": 14 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "spans": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "type": "text", + "content": "6" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 166, + 91, + 226, + 102 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 166, + 91, + 226, + 102 + ], + "spans": [ + { + "bbox": [ + 166, + 91, + 226, + 102 + ], + "type": "text", + "content": "R. Liang et al." + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 5 + }, + { + "para_blocks": [ + { + "bbox": [ + 130, + 116, + 482, + 237 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 116, + 482, + 237 + ], + "spans": [ + { + "bbox": [ + 130, + 116, + 482, + 237 + ], + "type": "text", + "content": "Tone-mapping. To compensate for the unknown tone-mapping of the input image, we introduces an estimable tone correction function " + }, + { + "bbox": [ + 130, + 116, + 482, + 237 + ], + "type": "inline_equation", + "content": "f(\\cdot)" + }, + { + "bbox": [ + 130, + 116, + 482, + 237 + ], + "type": "text", + "content": " that is applied on the inserted object " + }, + { + "bbox": [ + 130, + 116, + 482, + 237 + ], + "type": "inline_equation", + "content": "\\hat{\\mathbf{I}}_{\\mathrm{fg}} = f(\\mathbf{I}_{\\mathrm{fg}}; \\boldsymbol{\\theta}_{\\mathrm{fg}})" + }, + { + "bbox": [ + 130, + 116, + 482, + 237 + ], + "type": "text", + "content": " and shadows " + }, + { + "bbox": [ + 130, + 116, + 482, + 237 + ], + "type": "inline_equation", + "content": "\\hat{\\beta}_{\\mathrm{shadow}} = f(\\boldsymbol{\\beta}_{\\mathrm{shadow}}; \\boldsymbol{\\theta}_{\\mathrm{shadow}})" + }, + { + "bbox": [ + 130, + 116, + 482, + 237 + ], + "type": "text", + "content": ". Our tone curve parameterization follows from [13] and employs a monotonic rational-quadratic spline as its basic building block. This spline is composed of " + }, + { + "bbox": [ + 130, + 116, + 482, + 237 + ], + "type": "inline_equation", + "content": "K_{\\mathrm{s}}" + }, + { + "bbox": [ + 130, + 116, + 482, + 237 + ], + "type": "text", + "content": " bins, each defined by the quotient of two quadratic polynomials. Notably, these functions are differentiable, and allow direct parameterization of the derivatives and heights at each knot. We apply a single spline for the foreground image, and different correction splines for each RGB channel for the shadow ratio to provide the flexibility to adjust the color of the shadow. We set " + }, + { + "bbox": [ + 130, + 116, + 482, + 237 + ], + "type": "inline_equation", + "content": "K_{\\mathrm{s}} = 5" + }, + { + "bbox": [ + 130, + 116, + 482, + 237 + ], + "type": "text", + "content": " for these curves." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 131, + 243, + 483, + 268 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 131, + 243, + 483, + 268 + ], + "spans": [ + { + "bbox": [ + 131, + 243, + 483, + 268 + ], + "type": "text", + "content": "Differentiability. The final output image is an alpha-composite of the foreground object, shadows, and background image:" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 203, + 275, + 482, + 290 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 203, + 275, + 482, + 290 + ], + "spans": [ + { + "bbox": [ + 203, + 275, + 482, + 290 + ], + "type": "interline_equation", + "content": "\\mathbf {I} _ {\\mathrm {c o m p}} = \\left(1 - \\mathbf {V} (\\mathcal {X})\\right) \\cdot \\hat {\\boldsymbol {\\beta}} _ {\\mathrm {s h a d o w}} \\cdot \\mathbf {I} _ {\\mathrm {b g}} + \\mathbf {V} (\\mathcal {X}) \\cdot \\tilde {\\mathbf {I}} _ {\\mathrm {f g}}, \\tag {6}", + "image_path": "b7186cce66d1dd241bd9b811962dea6b7fdd944763de65fd5ab3b375c3138e6e.jpg" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 130, + 297, + 482, + 309 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 297, + 482, + 309 + ], + "spans": [ + { + "bbox": [ + 130, + 297, + 482, + 309 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 130, + 297, + 482, + 309 + ], + "type": "inline_equation", + "content": "\\mathbf{V}(\\mathcal{X})" + }, + { + "bbox": [ + 130, + 297, + 482, + 309 + ], + "type": "text", + "content": " equals 1 when " + }, + { + "bbox": [ + 130, + 297, + 482, + 309 + ], + "type": "inline_equation", + "content": "\\mathcal{X}" + }, + { + "bbox": [ + 130, + 297, + 482, + 309 + ], + "type": "text", + "content": " is directly visible from the camera, and 0 otherwise." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 130, + 310, + 482, + 393 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 310, + 482, + 393 + ], + "spans": [ + { + "bbox": [ + 130, + 310, + 482, + 393 + ], + "type": "text", + "content": "As neither foreground rendering, shadow rendering, or the compositing operation require derivatives w.r.t. the object placement or other discontinuous quantities, we can rely on automatic differentiation to obtain gradients of any pixelwise loss w.r.t. the lighting or material properties. To this end, we use the Path Replay Backpropagation [60] integrator of Mitsuba 3 [25]. The estimizable parameters are the Spherical Gaussian coefficients used in the lighting representation, and the tone curves' parameters " + }, + { + "bbox": [ + 130, + 310, + 482, + 393 + ], + "type": "inline_equation", + "content": "\\theta_{\\mathrm{fg}},\\theta_{\\mathrm{shadow}}" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 130, + 394, + 482, + 441 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 394, + 482, + 441 + ], + "spans": [ + { + "bbox": [ + 130, + 394, + 482, + 441 + ], + "type": "text", + "content": "While each of the operations above include simplifications (e.g. we do not account for secondary lighting from the proxy geometry), we have found them to be sufficient in practice: the remaining imperfections in the simulation can be sufficiently offset by the lighting optimization." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 131, + 458, + 257, + 470 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 131, + 458, + 257, + 470 + ], + "spans": [ + { + "bbox": [ + 131, + 458, + 257, + 470 + ], + "type": "text", + "content": "4.2 Diffusion Guidance" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 130, + 478, + 482, + 574 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 478, + 482, + 574 + ], + "spans": [ + { + "bbox": [ + 130, + 478, + 482, + 574 + ], + "type": "text", + "content": "The composited image produced by our differentiable rendering pipeline serves as the input to a DM that is used to compute a guidance signal, employing an optimization objective similar to Score Distillation Sampling (SDS) [45]. However, while DMs inherently have robust priors for lighting, we found that they do not provide out-of-the-box the necessary guidance for our specific needs. Consequently, we propose an adaptive score distillation loss specifically designed for object insertion tasks that exploits a personalization strategy which we detail in the following section." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 130, + 579, + 482, + 665 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 579, + 482, + 665 + ], + "spans": [ + { + "bbox": [ + 130, + 579, + 482, + 665 + ], + "type": "text", + "content": "Personalization with concept preservation. Off-the-shelf DMs often do not provide robust guidance for virtual object insertion, especially in out-of-distribution scenes such as outdoor driving environments. A potential solution, inspired by [57], is to adapt the DM using an image from the target scene. In our experience, however, this approach often resulted in too much overfitting to the target scene's content, reducing the model's ability to adapt to the scene with a newly inserted object. This led to artifacts and an unstable optimization process." + } + ] + } + ], + "index": 10 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 152, + 91, + 448, + 103 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 152, + 91, + 448, + 103 + ], + "spans": [ + { + "bbox": [ + 152, + 91, + 448, + 103 + ], + "type": "text", + "content": "Photorealistic Object Insertion with Diffusion-Guided Inverse Rendering" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 474, + 92, + 480, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 474, + 92, + 480, + 100 + ], + "spans": [ + { + "bbox": [ + 474, + 92, + 480, + 100 + ], + "type": "text", + "content": "7" + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 6 + }, + { + "para_blocks": [ + { + "bbox": [ + 130, + 116, + 299, + 270 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 116, + 299, + 270 + ], + "spans": [ + { + "bbox": [ + 130, + 116, + 299, + 270 + ], + "type": "text", + "content": "To mitigate this issue, we propose fine-tuning the DM with a focus on preserving the identity of the objects to be inserted (Fig. 3). We specifically achieved this by generating additional synthetic images for the insertable class concept (e.g. car). We sample those images from the off-the-self DM, starting with a base prompt such as \"a photo of a car\" and appending attributes such as color, background, lighting, and size to ensure diversity in the generated data." + } + ] + } + ], + "index": 2 + }, + { + "type": "image", + "bbox": [ + 306, + 123, + 369, + 171 + ], + "blocks": [ + { + "bbox": [ + 306, + 123, + 369, + 171 + ], + "lines": [ + { + "bbox": [ + 306, + 123, + 369, + 171 + ], + "spans": [ + { + "bbox": [ + 306, + 123, + 369, + 171 + ], + "type": "image", + "image_path": "95c3c69265f135eeb2c9b3cf21fbbb3ce87fd46adddf951bd9781515088ca11a.jpg" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 320, + 171, + 355, + 179 + ], + "lines": [ + { + "bbox": [ + 320, + 171, + 355, + 179 + ], + "spans": [ + { + "bbox": [ + 320, + 171, + 355, + 179 + ], + "type": "text", + "content": "Input image" + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 374, + 126, + 482, + 136 + ], + "lines": [ + { + "bbox": [ + 374, + 126, + 482, + 136 + ], + "spans": [ + { + "bbox": [ + 374, + 126, + 482, + 136 + ], + "type": "text", + "content": "\"ascene in the style of sks rendering" + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "image_caption" + } + ], + "index": 3 + }, + { + "type": "image", + "bbox": [ + 306, + 180, + 369, + 241 + ], + "blocks": [ + { + "bbox": [ + 306, + 180, + 369, + 241 + ], + "lines": [ + { + "bbox": [ + 306, + 180, + 369, + 241 + ], + "spans": [ + { + "bbox": [ + 306, + 180, + 369, + 241 + ], + "type": "image", + "image_path": "27d5ec4434c5f4a6e8cfd23876b353c58a69d45a6f4e5ebdf9f6035594d2d943.jpg" + } + ] + } + ], + "index": 6, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 312, + 242, + 362, + 248 + ], + "lines": [ + { + "bbox": [ + 312, + 242, + 362, + 248 + ], + "spans": [ + { + "bbox": [ + 312, + 242, + 362, + 248 + ], + "type": "text", + "content": "Generated images" + } + ] + } + ], + "index": 7, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 375, + 224, + 427, + 233 + ], + "lines": [ + { + "bbox": [ + 375, + 224, + 427, + 233 + ], + "spans": [ + { + "bbox": [ + 375, + 224, + 427, + 233 + ], + "type": "text", + "content": "\"aphotofacar\"" + } + ] + } + ], + "index": 9, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 305, + 254, + 480, + 264 + ], + "lines": [ + { + "bbox": [ + 305, + 254, + 480, + 264 + ], + "spans": [ + { + "bbox": [ + 305, + 254, + 480, + 264 + ], + "type": "text", + "content": "Fig. 3: Personalization with concept preservation." + } + ] + } + ], + "index": 10, + "angle": 0, + "type": "image_caption" + } + ], + "index": 6 + }, + { + "type": "image", + "bbox": [ + 374, + 145, + 466, + 209 + ], + "blocks": [ + { + "bbox": [ + 374, + 145, + 466, + 209 + ], + "lines": [ + { + "bbox": [ + 374, + 145, + 466, + 209 + ], + "spans": [ + { + "bbox": [ + 374, + 145, + 466, + 209 + ], + "type": "image", + "image_path": "fd243130eba1d645b5aeafdfe5b4fb49c059f81d2820fbef31de7e85cff82d16.jpg" + } + ] + } + ], + "index": 8, + "angle": 0, + "type": "image_body" + } + ], + "index": 8 + }, + { + "bbox": [ + 130, + 271, + 482, + 355 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 271, + 482, + 355 + ], + "spans": [ + { + "bbox": [ + 130, + 271, + 482, + 355 + ], + "type": "text", + "content": "We employ LoRA with rank 4 to fine-tune the DM, combining the in-domain target example with this supplemental data. Our training follows the objective described in Eq. 1, where " + }, + { + "bbox": [ + 130, + 271, + 482, + 355 + ], + "type": "inline_equation", + "content": "c" + }, + { + "bbox": [ + 130, + 271, + 482, + 355 + ], + "type": "text", + "content": " corresponds to two predefined prompts: \"a scene in the style of sks rendering\" for the target image, and \"a photo of a {concept class}\". We sample approximately 30-40 supplementary images for indoor scenes and 200 for outdoor scenes. The total time spent for fine-tuning is typically less than 15 minutes on one high-end GPU with mixed precision training." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 130, + 359, + 482, + 419 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 359, + 482, + 419 + ], + "spans": [ + { + "bbox": [ + 130, + 359, + 482, + 419 + ], + "type": "text", + "content": "Score Distillation with adapted guidance. Score Distillation Sampling [45] leverages a pretrained DM to guide the optimization of a differentiable, parametric image-rendering function " + }, + { + "bbox": [ + 130, + 359, + 482, + 419 + ], + "type": "inline_equation", + "content": "g_{\\phi} \\coloneqq \\pmb{x}" + }, + { + "bbox": [ + 130, + 359, + 482, + 419 + ], + "type": "text", + "content": ". Specifically, the parameters " + }, + { + "bbox": [ + 130, + 359, + 482, + 419 + ], + "type": "inline_equation", + "content": "\\phi" + }, + { + "bbox": [ + 130, + 359, + 482, + 419 + ], + "type": "text", + "content": " in our scenario corresponds to parameters of the Spherical Gaussian lighting and tone-mapping curves, are updated using the gradient:" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 170, + 424, + 482, + 451 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 170, + 424, + 482, + 451 + ], + "spans": [ + { + "bbox": [ + 170, + 424, + 482, + 451 + ], + "type": "interline_equation", + "content": "\\nabla_ {\\phi} \\mathcal {L} _ {\\mathrm {S D S}} (\\boldsymbol {\\phi}, \\boldsymbol {\\theta}) := \\mathbb {E} _ {\\boldsymbol {\\epsilon} \\sim \\mathcal {N} (\\mathbf {0}, \\boldsymbol {I}), t \\sim T} \\left[ w (t) \\left(\\hat {\\epsilon} _ {\\boldsymbol {\\theta}} (\\boldsymbol {z} _ {t}, t, \\boldsymbol {c}) - \\boldsymbol {\\epsilon}\\right) \\frac {\\partial \\boldsymbol {z} _ {t}}{\\partial \\phi} \\right], \\qquad (7)", + "image_path": "9e99393b9382ff60c65d2d012c1ff0ab400fc0f7bb23f71a01ece43ed13bf31d.jpg" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 130, + 456, + 482, + 491 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 456, + 482, + 491 + ], + "spans": [ + { + "bbox": [ + 130, + 456, + 482, + 491 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 130, + 456, + 482, + 491 + ], + "type": "inline_equation", + "content": "\\pmb{z} = \\mathcal{E}(g_{\\phi})" + }, + { + "bbox": [ + 130, + 456, + 482, + 491 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 130, + 456, + 482, + 491 + ], + "type": "inline_equation", + "content": "\\hat{\\epsilon}_{\\theta}(\\pmb{z}_t, t, \\pmb{c}) \\coloneqq (1 + s)\\pmb{\\epsilon}_{\\theta}(\\pmb{z}_t, t, \\pmb{c}) - s\\pmb{\\epsilon}_{\\theta}(\\pmb{z}_t, t, \\emptyset)" + }, + { + "bbox": [ + 130, + 456, + 482, + 491 + ], + "type": "text", + "content": ". " + }, + { + "bbox": [ + 130, + 456, + 482, + 491 + ], + "type": "inline_equation", + "content": "\\hat{\\epsilon}" + }, + { + "bbox": [ + 130, + 456, + 482, + 491 + ], + "type": "text", + "content": " denotes the classifier free guidance (CFG) version [21] of " + }, + { + "bbox": [ + 130, + 456, + 482, + 491 + ], + "type": "inline_equation", + "content": "\\pmb{\\epsilon}_{\\theta}" + }, + { + "bbox": [ + 130, + 456, + 482, + 491 + ], + "type": "text", + "content": " used in text-conditioned DMs to enable higher quality generation via a guidance scale parameter " + }, + { + "bbox": [ + 130, + 456, + 482, + 491 + ], + "type": "inline_equation", + "content": "s" + }, + { + "bbox": [ + 130, + 456, + 482, + 491 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 130, + 492, + 482, + 528 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 492, + 482, + 528 + ], + "spans": [ + { + "bbox": [ + 130, + 492, + 482, + 528 + ], + "type": "text", + "content": "Unfortunately, we encountered training instabilities when applying the original formulation of the SDS loss to our problem. Instead, we adopt the following alternative that integrates the LoRA personalization, and we call it LDS loss:" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 133, + 533, + 481, + 569 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 533, + 481, + 569 + ], + "spans": [ + { + "bbox": [ + 133, + 533, + 481, + 569 + ], + "type": "interline_equation", + "content": "\\nabla_ {\\phi} \\mathcal {L} _ {\\mathrm {L D S}} (\\phi , \\boldsymbol {\\theta}) := \\mathbb {E} _ {\\epsilon \\sim \\mathcal {N} (\\boldsymbol {0}, \\boldsymbol {I}), t \\sim T} \\left[ w (t) \\left(\\epsilon_ {(\\boldsymbol {\\theta} + \\boldsymbol {\\Delta W})} (\\boldsymbol {z} _ {t}, t, \\boldsymbol {c}) - \\epsilon_ {\\boldsymbol {\\theta}} (\\boldsymbol {z} _ {t}, t, \\varnothing)\\right) \\frac {\\partial \\boldsymbol {z} _ {t}}{\\partial \\phi} \\right], \\tag {8}", + "image_path": "c4b57f05821f53bca34158528b7c8190cf181bdac79c8b54ec5ad7742feb911a.jpg" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 130, + 570, + 482, + 665 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 570, + 482, + 665 + ], + "spans": [ + { + "bbox": [ + 130, + 570, + 482, + 665 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 130, + 570, + 482, + 665 + ], + "type": "inline_equation", + "content": "\\epsilon_{(\\theta +\\Delta W)}" + }, + { + "bbox": [ + 130, + 570, + 482, + 665 + ], + "type": "text", + "content": " represents the predicted noise of the LoRA personalized model described in the previous section. Notably, this loss function bears resemblance to the Classifier Score Distillation (CSD) [68]. However, it's important to note that in our case, the delta is calculated between the adapted and non-adapted versions of the diffusion model. Intuitively, this delta guides the optimization process in a direction determined by the personalized model which preserves the concept of the inserted object while also capturing the appearance and semantics of the specific scene." + } + ] + } + ], + "index": 17 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "spans": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "type": "text", + "content": "8" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 166, + 91, + 227, + 102 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 166, + 91, + 227, + 102 + ], + "spans": [ + { + "bbox": [ + 166, + 91, + 227, + 102 + ], + "type": "text", + "content": "R. Liang et al." + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 7 + }, + { + "para_blocks": [ + { + "bbox": [ + 132, + 116, + 292, + 129 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 116, + 292, + 129 + ], + "spans": [ + { + "bbox": [ + 132, + 116, + 292, + 129 + ], + "type": "text", + "content": "4.3 Optimization Formulation" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 131, + 137, + 480, + 160 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 131, + 137, + 480, + 160 + ], + "spans": [ + { + "bbox": [ + 131, + 137, + 480, + 160 + ], + "type": "text", + "content": "After completing the diffusion model's personalization, we optimize the lighting and tone-mapping parameters using the following loss function:" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 208, + 171, + 481, + 184 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 208, + 171, + 481, + 184 + ], + "spans": [ + { + "bbox": [ + 208, + 171, + 481, + 184 + ], + "type": "interline_equation", + "content": "\\mathcal {L} = \\mathcal {L} _ {\\mathrm {L D S}} + \\lambda_ {\\text {c o n s i s t e n c y}} \\mathcal {L} _ {\\text {c o n s i s t e n c y}} + \\lambda_ {\\text {r e g}} \\mathcal {L} _ {\\text {r e g}}. \\tag {9}", + "image_path": "29cf6fa091da1adf5b0bc9d18a422ee0a0db74b8ed23bf3fa5f4874e1889315d.jpg" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 131, + 193, + 482, + 300 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 131, + 193, + 482, + 300 + ], + "spans": [ + { + "bbox": [ + 131, + 193, + 482, + 300 + ], + "type": "text", + "content": "The loss' main component is the diffusion guidance " + }, + { + "bbox": [ + 131, + 193, + 482, + 300 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_{\\mathrm{LDS}}" + }, + { + "bbox": [ + 131, + 193, + 482, + 300 + ], + "type": "text", + "content": ", which provides a perceptual realism objective on the edited result with object insertion. We use the text prompt \"a photo of a {concept class} in a scene in the style of sks rendering\", which is a natural combination of the two text prompts used in personalization (Fig. 3). Here {concept class}, e.g. car, provides the context of the object to insert, while the personalized token sks provides the lighting style of the input image. The diffusion guidance " + }, + { + "bbox": [ + 131, + 193, + 482, + 300 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_{\\mathrm{LDS}}" + }, + { + "bbox": [ + 131, + 193, + 482, + 300 + ], + "type": "text", + "content": " is applied on the edited image " + }, + { + "bbox": [ + 131, + 193, + 482, + 300 + ], + "type": "inline_equation", + "content": "\\mathbf{I}_{\\mathrm{comp}}" + }, + { + "bbox": [ + 131, + 193, + 482, + 300 + ], + "type": "text", + "content": " (Eq. 6), and backpropagated to the estimizable parameters through the differentiable rendering process (Eq. 8)." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 131, + 301, + 481, + 324 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 131, + 301, + 481, + 324 + ], + "spans": [ + { + "bbox": [ + 131, + 301, + 481, + 324 + ], + "type": "text", + "content": "The two regularizers " + }, + { + "bbox": [ + 131, + 301, + 481, + 324 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_{\\mathrm{consistency}}" + }, + { + "bbox": [ + 131, + 301, + 481, + 324 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 131, + 301, + 481, + 324 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_{\\mathrm{reg}}" + }, + { + "bbox": [ + 131, + 301, + 481, + 324 + ], + "type": "text", + "content": " are related to environment map fusion, which we detail below." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 131, + 331, + 482, + 415 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 131, + 331, + 482, + 415 + ], + "spans": [ + { + "bbox": [ + 131, + 331, + 482, + 415 + ], + "type": "text", + "content": "Environment map initialization and fusion. The personalized DM provides guidance in two ways: (i) encouraging lighting consistency of the foreground object and the scene, such as the reflections and scale of the inserted object; and (ii) encouraging accurate shadows cast by the inserted object onto the background scene, such as the scale, direction, and color of the shadow. However, we empirically observe that these two signals can conflict in the early phase of the optimization." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 131, + 415, + 482, + 511 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 131, + 415, + 482, + 511 + ], + "spans": [ + { + "bbox": [ + 131, + 415, + 482, + 511 + ], + "type": "text", + "content": "To address this, we initialize two separate environment maps, " + }, + { + "bbox": [ + 131, + 415, + 482, + 511 + ], + "type": "inline_equation", + "content": "\\mathbf{L}^{\\mathrm{fg}},\\mathbf{L}^{\\mathrm{shadow}}\\in" + }, + { + "bbox": [ + 131, + 415, + 482, + 511 + ], + "type": "inline_equation", + "content": "\\mathbb{R}^{H\\times W\\times 3}" + }, + { + "bbox": [ + 131, + 415, + 482, + 511 + ], + "type": "text", + "content": ", to light the foreground inserted object (Eq. 4) and cast shadows (Eq. 5) respectively. As the optimization progresses, the two environment maps are progressively fused into a single environment map " + }, + { + "bbox": [ + 131, + 415, + 482, + 511 + ], + "type": "inline_equation", + "content": "\\mathbf{L}^{\\mathrm{fused}}" + }, + { + "bbox": [ + 131, + 415, + 482, + 511 + ], + "type": "text", + "content": " through scaling the the " + }, + { + "bbox": [ + 131, + 415, + 482, + 511 + ], + "type": "inline_equation", + "content": "\\mathbf{L}^{\\mathrm{fg}}" + }, + { + "bbox": [ + 131, + 415, + 482, + 511 + ], + "type": "text", + "content": " by the relative luminance of between " + }, + { + "bbox": [ + 131, + 415, + 482, + 511 + ], + "type": "inline_equation", + "content": "\\mathbf{L}^{\\mathrm{fg}}" + }, + { + "bbox": [ + 131, + 415, + 482, + 511 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 131, + 415, + 482, + 511 + ], + "type": "inline_equation", + "content": "\\mathbf{L}^{\\mathrm{shadow}}" + }, + { + "bbox": [ + 131, + 415, + 482, + 511 + ], + "type": "text", + "content": ". We also use the following two regularization terms for this fusion process. First, we encourage the consistency between the normalized luminance " + }, + { + "bbox": [ + 131, + 415, + 482, + 511 + ], + "type": "inline_equation", + "content": "\\tilde{\\mathbf{L}}^{\\mathrm{fg}},\\tilde{\\mathbf{L}}^{\\mathrm{shadow}}\\in \\mathbb{R}^{H\\times W}" + }, + { + "bbox": [ + 131, + 415, + 482, + 511 + ], + "type": "text", + "content": " of the environment maps by minimizing:" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 212, + 521, + 481, + 548 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 212, + 521, + 481, + 548 + ], + "spans": [ + { + "bbox": [ + 212, + 521, + 481, + 548 + ], + "type": "interline_equation", + "content": "\\mathcal {L} _ {\\text {c o n s i s t e n c y}} = - \\sum_ {i, j} \\tilde {\\mathbf {L}} _ {i, j} ^ {\\text {s h a d o w}} \\log \\left(\\tilde {\\mathbf {L}} _ {i, j} ^ {\\mathrm {f g}}\\right) \\Delta \\Omega_ {i, j} \\tag {10}", + "image_path": "81b0ddb86a4741cbf70a687907e14c82054d20e5d2d03a816920e6c237791434.jpg" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 131, + 556, + 482, + 618 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 131, + 556, + 482, + 618 + ], + "spans": [ + { + "bbox": [ + 131, + 556, + 482, + 618 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 131, + 556, + 482, + 618 + ], + "type": "inline_equation", + "content": "\\varDelta\\varOmega_{i,j}" + }, + { + "bbox": [ + 131, + 556, + 482, + 618 + ], + "type": "text", + "content": " is the corresponding solid angle at pixel " + }, + { + "bbox": [ + 131, + 556, + 482, + 618 + ], + "type": "inline_equation", + "content": "(i,j)" + }, + { + "bbox": [ + 131, + 556, + 482, + 618 + ], + "type": "text", + "content": " and the gradient for " + }, + { + "bbox": [ + 131, + 556, + 482, + 618 + ], + "type": "inline_equation", + "content": "\\tilde{\\mathbf{L}}^{\\mathrm{shadow}}" + }, + { + "bbox": [ + 131, + 556, + 482, + 618 + ], + "type": "text", + "content": " is detached. Second, as the shadow environment map " + }, + { + "bbox": [ + 131, + 556, + 482, + 618 + ], + "type": "inline_equation", + "content": "\\mathbf{L}^{\\mathrm{shadow}}" + }, + { + "bbox": [ + 131, + 556, + 482, + 618 + ], + "type": "text", + "content": " is supervised mainly via the shadow ratio " + }, + { + "bbox": [ + 131, + 556, + 482, + 618 + ], + "type": "inline_equation", + "content": "\\beta_{\\mathrm{shadow}}\\in \\mathbb{R}^{H\\times W\\times 3}" + }, + { + "bbox": [ + 131, + 556, + 482, + 618 + ], + "type": "text", + "content": " of Eq. 5, to encourage concentrated high peaks for the sharp shadows and suppress the ambient light in " + }, + { + "bbox": [ + 131, + 556, + 482, + 618 + ], + "type": "inline_equation", + "content": "\\mathbf{L}^{\\mathrm{shadow}}" + }, + { + "bbox": [ + 131, + 556, + 482, + 618 + ], + "type": "text", + "content": ", we also add a L2 regularizer in log-space with the Cauchy loss [7]:" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 219, + 627, + 481, + 654 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 219, + 627, + 481, + 654 + ], + "spans": [ + { + "bbox": [ + 219, + 627, + 481, + 654 + ], + "type": "interline_equation", + "content": "\\mathcal {L} _ {\\text {r e g}} = \\sum_ {i, j, c} \\log \\left(1 + 2 \\left(\\mathbf {L} _ {i, j, c} ^ {\\text {s h a d o w}}\\right) ^ {2}\\right) \\Delta \\Omega_ {i, j}. \\tag {11}", + "image_path": "5b8e0a05dfcde754681a82313269ded649ab2a556f4a4359cd354f60f0dd6e50.jpg" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 132, + 662, + 345, + 674 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 662, + 345, + 674 + ], + "spans": [ + { + "bbox": [ + 132, + 662, + 345, + 674 + ], + "type": "text", + "content": "Please refer to the Supplement for more details." + } + ] + } + ], + "index": 12 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 152, + 91, + 448, + 103 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 152, + 91, + 448, + 103 + ], + "spans": [ + { + "bbox": [ + 152, + 91, + 448, + 103 + ], + "type": "text", + "content": "Photorealistic Object Insertion with Diffusion-Guided Inverse Rendering" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 474, + 92, + 480, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 474, + 92, + 480, + 100 + ], + "spans": [ + { + "bbox": [ + 474, + 92, + 480, + 100 + ], + "type": "text", + "content": "9" + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 8 + }, + { + "para_blocks": [ + { + "type": "table", + "bbox": [ + 182, + 171, + 430, + 230 + ], + "blocks": [ + { + "bbox": [ + 130, + 114, + 482, + 169 + ], + "lines": [ + { + "bbox": [ + 130, + 114, + 482, + 169 + ], + "spans": [ + { + "bbox": [ + 130, + 114, + 482, + 169 + ], + "type": "text", + "content": "Table 1: Quantitative user study on outdoor street scenes. For each scene, users are shown two results—one produced by our method, and another produced by one of the baselines—and select which is more realistic. We report the results averaged across 3 user studies with 9 users each. Our method outperforms all baselines (" + }, + { + "bbox": [ + 130, + 114, + 482, + 169 + ], + "type": "inline_equation", + "content": ">50\\%" + }, + { + "bbox": [ + 130, + 114, + 482, + 169 + ], + "type": "text", + "content": ") and is preferred in almost all illumination conditions." + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 182, + 171, + 430, + 230 + ], + "lines": [ + { + "bbox": [ + 182, + 171, + 430, + 230 + ], + "spans": [ + { + "bbox": [ + 182, + 171, + 430, + 230 + ], + "type": "table", + "html": "
DaytimeTwilightNightAll scenes
SunnyCloudy
Hold-Geoffroy et al. [22]60.8%66.7%74.1%85.7%68.8%
NLFE [62]80.4%73.3%44.4%52.4%67.4%
StyleLight [61]76.5%91.1%66.7%66.7%77.8%
DiffusionLight [44]80.4%68.9%55.6%71.4%70.8%
", + "image_path": "b1e4052f342c259114472b8d4e1ff65df3464ebd1f9364e7460a0a15923f7e51.jpg" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "table_body" + } + ], + "index": 3 + }, + { + "type": "table", + "bbox": [ + 182, + 269, + 430, + 316 + ], + "blocks": [ + { + "bbox": [ + 130, + 234, + 481, + 267 + ], + "lines": [ + { + "bbox": [ + 130, + 234, + 481, + 267 + ], + "spans": [ + { + "bbox": [ + 130, + 234, + 481, + 267 + ], + "type": "text", + "content": "Table 2: Quantitative evaluation on PolyHaven scenes. We report user study preference scores similar to Table 1. Metrics are computed w.r.t. to a “reference” image where the virtual object is lit by the ground-truth environment map." + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 182, + 269, + 430, + 316 + ], + "lines": [ + { + "bbox": [ + 182, + 269, + 430, + 316 + ], + "spans": [ + { + "bbox": [ + 182, + 269, + 430, + 316 + ], + "type": "table", + "html": "
MethodsOurs preferred ↓RMSE ↓SSIM ↑LPIPS [73] ↓si-RMSE [18] ↓
Wang et al. [63]84.8%0.0630.9850.01750.037
StyleLight [61]75.8%0.0560.9860.02020.039
DiffusionLight [44]66.7%0.0620.9850.01620.034
Ours/0.0480.9890.01470.027
", + "image_path": "6a9e834e67d11ecdbefeddebd625a376c2dc748d935d5e40c4453ae48ffad717.jpg" + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "table_body" + } + ], + "index": 5 + }, + { + "bbox": [ + 130, + 324, + 482, + 433 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 324, + 482, + 433 + ], + "spans": [ + { + "bbox": [ + 130, + 324, + 482, + 433 + ], + "type": "text", + "content": "Training details. The personalized DM receives a crop of " + }, + { + "bbox": [ + 130, + 324, + 482, + 433 + ], + "type": "inline_equation", + "content": "\\mathbf{I}_{\\mathrm{comp}}" + }, + { + "bbox": [ + 130, + 324, + 482, + 433 + ], + "type": "text", + "content": " (Eq. 6) with added noise. Since the inserted object likely does not cover the full background image, we use the visibility mask of the virtual object " + }, + { + "bbox": [ + 130, + 324, + 482, + 433 + ], + "type": "inline_equation", + "content": "V(\\mathcal{X})" + }, + { + "bbox": [ + 130, + 324, + 482, + 433 + ], + "type": "text", + "content": " to locate the 2D bounding box of the inserted object, and randomly crop around the inserted object. The size of the cropped image is also randomized, with smaller crops covering local details of the inserted object and larger crops providing more visual cues of the lighting effects within the scene. Each scene is optimized for 600 iterations, and the maximum strength of the diffusion guidance is linearly decreased from 0.6 to 0.3. We use " + }, + { + "bbox": [ + 130, + 324, + 482, + 433 + ], + "type": "inline_equation", + "content": "\\lambda_{\\mathrm{consistency}} = 0.03" + }, + { + "bbox": [ + 130, + 324, + 482, + 433 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 130, + 324, + 482, + 433 + ], + "type": "inline_equation", + "content": "\\lambda_{\\mathrm{reg}} = 0.01" + }, + { + "bbox": [ + 130, + 324, + 482, + 433 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 131, + 445, + 230, + 460 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 131, + 445, + 230, + 460 + ], + "spans": [ + { + "bbox": [ + 131, + 445, + 230, + 460 + ], + "type": "text", + "content": "5 Experiments" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 130, + 463, + 482, + 511 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 463, + 482, + 511 + ], + "spans": [ + { + "bbox": [ + 130, + 463, + 482, + 511 + ], + "type": "text", + "content": "In this section, we evaluate DiPIR on a collection of outdoor and indoor scenes, covering diverse lighting conditions and different application domains. We first outline the experiment settings in Sec. 5.1 and then provide evaluation results on benchmark datasets in Sec. 5.2. Finally, we conduct an ablation study in Sec. 5.3." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 131, + 521, + 265, + 533 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 131, + 521, + 265, + 533 + ], + "spans": [ + { + "bbox": [ + 131, + 521, + 265, + 533 + ], + "type": "text", + "content": "5.1 Experiment Settings" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 130, + 538, + 482, + 634 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 538, + 482, + 634 + ], + "spans": [ + { + "bbox": [ + 130, + 538, + 482, + 634 + ], + "type": "text", + "content": "Waymo dataset. Following the data split from [66], we use 48 scenes with diverse illumination conditions from the Waymo Open Dataset [55] including 32 daytime (17 sunny and 15 cloudy), 9 twilight, and 7 night scenes. Each scene consists of an input image, a 3D car asset randomly selected from an asset bank, the ground plane, and a location to insert the car. Inspired by applications in synthetic data generation, this process is fully automatic: we use the LiDAR point cloud and semantic segmentation [58] to fit a ground plane and detect the empty space to insert a car." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 131, + 641, + 481, + 665 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 131, + 641, + 481, + 665 + ], + "spans": [ + { + "bbox": [ + 131, + 641, + 481, + 665 + ], + "type": "text", + "content": "PolyHaven dataset. We use 11 HDR environment maps from PolyHaven [1] and manually place a known ground plane and virtual object in each scene. Each" + } + ] + } + ], + "index": 11 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "spans": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "type": "text", + "content": "10" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 166, + 91, + 227, + 102 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 166, + 91, + 227, + 102 + ], + "spans": [ + { + "bbox": [ + 166, + 91, + 227, + 102 + ], + "type": "text", + "content": "R. Liang et al." + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 9 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 134, + 114, + 480, + 235 + ], + "blocks": [ + { + "bbox": [ + 134, + 114, + 480, + 235 + ], + "lines": [ + { + "bbox": [ + 134, + 114, + 480, + 235 + ], + "spans": [ + { + "bbox": [ + 134, + 114, + 480, + 235 + ], + "type": "image", + "image_path": "8524d671d57967f33486f8f65292c31c0d91fa4498a0a052d1e1d0975d57f6e0.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 144, + 236, + 468, + 247 + ], + "lines": [ + { + "bbox": [ + 144, + 236, + 468, + 247 + ], + "spans": [ + { + "bbox": [ + 144, + 236, + 468, + 247 + ], + "type": "text", + "content": "Fig. 4: Comparisons on inserting objects into cropped HDRIs from PolyHaven." + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_caption" + } + ], + "index": 2 + }, + { + "type": "image", + "bbox": [ + 134, + 249, + 480, + 460 + ], + "blocks": [ + { + "bbox": [ + 134, + 249, + 480, + 460 + ], + "lines": [ + { + "bbox": [ + 134, + 249, + 480, + 460 + ], + "spans": [ + { + "bbox": [ + 134, + 249, + 480, + 460 + ], + "type": "image", + "image_path": "4fc027309d44855e0aa1432d1ac2501f11a0c6f8ff7a4a98afa307ed39ac0e1e.jpg" + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 130, + 461, + 481, + 483 + ], + "lines": [ + { + "bbox": [ + 130, + 461, + 481, + 483 + ], + "spans": [ + { + "bbox": [ + 130, + 461, + 481, + 483 + ], + "type": "text", + "content": "Fig. 5: Inserting car assets into Waymo driving scenes. Note the direction and sharpness of shadows, as well as overall brightness, color and specular highlights on inserted cars." + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "image_caption" + } + ], + "index": 4 + }, + { + "bbox": [ + 130, + 490, + 482, + 538 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 490, + 482, + 538 + ], + "spans": [ + { + "bbox": [ + 130, + 490, + 482, + 538 + ], + "type": "text", + "content": "scenario therefore consists of an LDR background image " + }, + { + "bbox": [ + 130, + 490, + 482, + 538 + ], + "type": "inline_equation", + "content": "\\mathbf{I}_{\\mathrm{bg}}" + }, + { + "bbox": [ + 130, + 490, + 482, + 538 + ], + "type": "text", + "content": " rendered directly from the environment map, the posed virtual object, and the proxy plane. A pseudo-ground-truth rendering is created by lighting the inserted object with the environment map itself." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 130, + 546, + 482, + 666 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 546, + 482, + 666 + ], + "spans": [ + { + "bbox": [ + 130, + 546, + 482, + 666 + ], + "type": "text", + "content": "User study. To evaluate the perceptual realism of the virtual object insertion, we conducted a user study where participants received a pair of two object insertion results in a random order - one generated using our proposed method, the other using a baseline approach. The participants were then instructed to compare the differences between the two results, inspect the lighting effects of the inserted object, and choose the image they perceived as more realistic. We invited 9 users to perform a binary selection for each image pair and used majority voting to determine the preferred image for each comparison. In the following, we report the percentage of times that our method was selected over the baseline. This process was repeated three times and a preferred percentage " + }, + { + "bbox": [ + 130, + 546, + 482, + 666 + ], + "type": "inline_equation", + "content": ">50\\%" + }, + { + "bbox": [ + 130, + 546, + 482, + 666 + ], + "type": "text", + "content": " indicates" + } + ] + } + ], + "index": 7 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 153, + 91, + 448, + 102 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 153, + 91, + 448, + 102 + ], + "spans": [ + { + "bbox": [ + 153, + 91, + 448, + 102 + ], + "type": "text", + "content": "Photorealistic Object Insertion with Diffusion-Guided Inverse Rendering" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 470, + 91, + 479, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 470, + 91, + 479, + 100 + ], + "spans": [ + { + "bbox": [ + 470, + 91, + 479, + 100 + ], + "type": "text", + "content": "11" + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 10 + }, + { + "para_blocks": [ + { + "bbox": [ + 130, + 116, + 479, + 140 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 116, + 479, + 140 + ], + "spans": [ + { + "bbox": [ + 130, + 116, + 479, + 140 + ], + "type": "text", + "content": "Ours outperforming baselines. We include more details about the user study and statistical evaluation of its results in the Supplement." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 130, + 146, + 482, + 218 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 146, + 482, + 218 + ], + "spans": [ + { + "bbox": [ + 130, + 146, + 482, + 218 + ], + "type": "text", + "content": "Baselines. On the Waymo dataset of urban street scenes, we compare our method to the representative outdoor lighting estimation methods of Hold-Geoffroy et al. [22] and NLFE [62]; as well as the generative lighting estimation methods StyleLight [61] and DiffusionLight [44]. On the PolyHaven dataset, we compare with Wang et al. [63], StyleLight [61], DiffusionLight [44], and pseudo-ground-truth rendering." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 131, + 235, + 338, + 247 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 131, + 235, + 338, + 247 + ], + "spans": [ + { + "bbox": [ + 131, + 235, + 338, + 247 + ], + "type": "text", + "content": "5.2 Evaluation on Benchmark Datasets" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 130, + 251, + 481, + 300 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 251, + 481, + 300 + ], + "spans": [ + { + "bbox": [ + 130, + 251, + 481, + 300 + ], + "type": "text", + "content": "Urban street scenes. In Table 1, we provide quantitative user study results on the Waymo dataset [55], separated into 4 subsets based on illumination conditions, as well as on all scenes. Compared to all baselines, our method outperforms prior state-of-the-art methods (chosen " + }, + { + "bbox": [ + 130, + 251, + 481, + 300 + ], + "type": "inline_equation", + "content": ">50\\%" + }, + { + "bbox": [ + 130, + 251, + 481, + 300 + ], + "type": "text", + "content": "), and is preferred in almost every subset." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 130, + 300, + 482, + 418 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 300, + 482, + 418 + ], + "spans": [ + { + "bbox": [ + 130, + 300, + 482, + 418 + ], + "type": "text", + "content": "StyleLight [61] is primarily trained on indoor HDR panoramas and suffers from domain gap to outdoor scenes. Similarly, the sky model predicted by Hold-Geoffroy et al. [22] is tailored to outdoor daytime scenes and suffers severely from out-of-domain twilight and night scenes. NLFE [62] unprojects the surrounding scene geometry and predicts volumetric lighting, which makes it generalize better to night scenes, but it can fail to estimate an accurate scale and color of the peak for daytime scenes. DiffusionLight [44], a concurrent work, shows impressive results by inpainting a chrome sphere and predicting environment maps with realistic high-frequency details. However, it often struggles to predict a high-intensity peak in sunny daytime scenes or a proper scale for night scenes." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 130, + 419, + 481, + 456 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 419, + 481, + 456 + ], + "spans": [ + { + "bbox": [ + 130, + 419, + 481, + 456 + ], + "type": "text", + "content": "A qualitative comparison is shown in Fig. 5. Our method achieves high-quality object insertion, making it a promising approach for applications such as simulating synthetic data, augmented reality navigation and urban planning." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 130, + 461, + 481, + 521 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 461, + 481, + 521 + ], + "spans": [ + { + "bbox": [ + 130, + 461, + 481, + 521 + ], + "type": "text", + "content": "PolyHaven dataset. We show the quantitative comparison in Table 2 and qualitative results in Fig. 4. Our method is preferred over all baseline methods in the user study, and outperforms baselines on quantitative metrics. Our method produces virtual object insertion that naturally and consistently blend in the scenes, making it promising for virtual production tasks." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 131, + 538, + 238, + 551 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 131, + 538, + 238, + 551 + ], + "spans": [ + { + "bbox": [ + 131, + 538, + 238, + 551 + ], + "type": "text", + "content": "5.3 Ablation Study" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 130, + 558, + 482, + 666 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 558, + 482, + 666 + ], + "spans": [ + { + "bbox": [ + 130, + 558, + 482, + 666 + ], + "type": "text", + "content": "We extensively ablate our proposed design choices on diffusion guidance and scene representations, and provide the quantitative ablation in Table 3 with user study and qualitative results in Fig. 6. For the diffusion guidance, we compare with simplified or alternative versions: (i) Ours (SDS) removes the adaptive guidance (Eq. 8) and uses the original SDS [45] formulation (Eq. 7); (ii) Ours (SDS w/o LoRA) uses the original SDS loss with non-adapted DM; (iii) Ours (w/o concept preservation) uses our LDS loss with adapted DM but not using any class images for concept preservation. We additionally compare with a baseline of \"dataset update\" which applies a photometric L1 loss on DM repeatedly edited results" + } + ] + } + ], + "index": 10 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "spans": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "type": "text", + "content": "12" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 166, + 91, + 227, + 102 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 166, + 91, + 227, + 102 + ], + "spans": [ + { + "bbox": [ + 166, + 91, + 227, + 102 + ], + "type": "text", + "content": "R. Liang et al." + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 11 + }, + { + "para_blocks": [ + { + "type": "table", + "bbox": [ + 136, + 114, + 297, + 178 + ], + "blocks": [ + { + "bbox": [ + 136, + 114, + 297, + 178 + ], + "lines": [ + { + "bbox": [ + 136, + 114, + 297, + 178 + ], + "spans": [ + { + "bbox": [ + 136, + 114, + 297, + 178 + ], + "type": "table", + "html": "
MethodsOurs preferred ↓
Ours (dataset update)85.2%
Ours (SDS [45])74.1%
Ours (SDS [45] w/o LoRA)90.7%
Ours (w/o concept preservation)64.8%
Ours (w/o tone curve)68.5%
Ours (w/o env. map fusion)66.7%
", + "image_path": "0eef9cb0d0294e21241cc91251acf1926f02c47427bd6b77274a1b4efa915e08.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "table_body" + } + ], + "index": 2 + }, + { + "bbox": [ + 309, + 114, + 486, + 179 + ], + "lines": [ + { + "bbox": [ + 309, + 114, + 486, + 179 + ], + "spans": [ + { + "bbox": [ + 309, + 114, + 486, + 179 + ], + "type": "text", + "content": "Table 3: Ablation study on outdoor driving scenes [55]. We report the percentage of images that users preferred DiPIR compared to its ablated versions. Our full pipeline produces results that are preferred more often over its ablated versions." + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "text" + }, + { + "type": "image", + "bbox": [ + 133, + 186, + 220, + 244 + ], + "blocks": [ + { + "bbox": [ + 133, + 186, + 220, + 244 + ], + "lines": [ + { + "bbox": [ + 133, + 186, + 220, + 244 + ], + "spans": [ + { + "bbox": [ + 133, + 186, + 220, + 244 + ], + "type": "image", + "image_path": "3e40de3e5434336d40fdd305d84be29c140bc97fddcaf84f801bf9c1598822f7.jpg" + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 136, + 245, + 217, + 253 + ], + "lines": [ + { + "bbox": [ + 136, + 245, + 217, + 253 + ], + "spans": [ + { + "bbox": [ + 136, + 245, + 217, + 253 + ], + "type": "text", + "content": "Ours (SDS w/o LoRA)" + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "image_caption" + } + ], + "index": 4 + }, + { + "type": "image", + "bbox": [ + 221, + 186, + 306, + 244 + ], + "blocks": [ + { + "bbox": [ + 221, + 186, + 306, + 244 + ], + "lines": [ + { + "bbox": [ + 221, + 186, + 306, + 244 + ], + "spans": [ + { + "bbox": [ + 221, + 186, + 306, + 244 + ], + "type": "image", + "image_path": "58d3b97546e9e7ad48a0c84a29763969d058d13933177e10619513613ccec2be.jpg" + } + ] + } + ], + "index": 6, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 225, + 245, + 301, + 253 + ], + "lines": [ + { + "bbox": [ + 225, + 245, + 301, + 253 + ], + "spans": [ + { + "bbox": [ + 225, + 245, + 301, + 253 + ], + "type": "text", + "content": "Ours (SDS w/ LoRA)" + } + ] + } + ], + "index": 7, + "angle": 0, + "type": "image_caption" + } + ], + "index": 6 + }, + { + "type": "image", + "bbox": [ + 307, + 186, + 392, + 244 + ], + "blocks": [ + { + "bbox": [ + 307, + 186, + 392, + 244 + ], + "lines": [ + { + "bbox": [ + 307, + 186, + 392, + 244 + ], + "spans": [ + { + "bbox": [ + 307, + 186, + 392, + 244 + ], + "type": "image", + "image_path": "2dac3abc527de255f449c05c8f17e09ee34fad2cc05b5b665f2de952db3469ce.jpg" + } + ] + } + ], + "index": 8, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 311, + 245, + 389, + 254 + ], + "lines": [ + { + "bbox": [ + 311, + 245, + 389, + 254 + ], + "spans": [ + { + "bbox": [ + 311, + 245, + 389, + 254 + ], + "type": "text", + "content": "Ours (dataset update)" + } + ] + } + ], + "index": 9, + "angle": 0, + "type": "image_caption" + } + ], + "index": 8 + }, + { + "type": "image", + "bbox": [ + 394, + 220, + 480, + 277 + ], + "blocks": [ + { + "bbox": [ + 394, + 220, + 480, + 277 + ], + "lines": [ + { + "bbox": [ + 394, + 220, + 480, + 277 + ], + "spans": [ + { + "bbox": [ + 394, + 220, + 480, + 277 + ], + "type": "image", + "image_path": "9b12df193fc4464c23fec94403d8ff51f70b51a46ed16e68459ab780aa6943b3.jpg" + } + ] + } + ], + "index": 10, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 429, + 277, + 447, + 285 + ], + "lines": [ + { + "bbox": [ + 429, + 277, + 447, + 285 + ], + "spans": [ + { + "bbox": [ + 429, + 277, + 447, + 285 + ], + "type": "text", + "content": "Ours" + } + ] + } + ], + "index": 11, + "angle": 0, + "type": "image_caption" + } + ], + "index": 10 + }, + { + "type": "image", + "bbox": [ + 134, + 255, + 220, + 312 + ], + "blocks": [ + { + "bbox": [ + 134, + 255, + 220, + 312 + ], + "lines": [ + { + "bbox": [ + 134, + 255, + 220, + 312 + ], + "spans": [ + { + "bbox": [ + 134, + 255, + 220, + 312 + ], + "type": "image", + "image_path": "6e016a3a6e3a24546148c073f8c34287a03d2c0cac1fd1932cda4f1d3863970e.jpg" + } + ] + } + ], + "index": 12, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 143, + 312, + 211, + 329 + ], + "lines": [ + { + "bbox": [ + 143, + 312, + 211, + 329 + ], + "spans": [ + { + "bbox": [ + 143, + 312, + 211, + 329 + ], + "type": "text", + "content": "Ours (w/o concept preservation)" + } + ] + } + ], + "index": 13, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 220, + 312, + 308, + 329 + ], + "lines": [ + { + "bbox": [ + 220, + 312, + 308, + 329 + ], + "spans": [ + { + "bbox": [ + 220, + 312, + 308, + 329 + ], + "type": "text", + "content": "Ours (w/o tone-mapping curve)" + } + ] + } + ], + "index": 15, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 192, + 334, + 420, + 345 + ], + "lines": [ + { + "bbox": [ + 192, + 334, + 420, + 345 + ], + "spans": [ + { + "bbox": [ + 192, + 334, + 420, + 345 + ], + "type": "text", + "content": "Fig. 6: Qualitative ablation study of our design choices." + } + ] + } + ], + "index": 19, + "angle": 0, + "type": "image_caption" + } + ], + "index": 12 + }, + { + "type": "image", + "bbox": [ + 221, + 255, + 306, + 312 + ], + "blocks": [ + { + "bbox": [ + 221, + 255, + 306, + 312 + ], + "lines": [ + { + "bbox": [ + 221, + 255, + 306, + 312 + ], + "spans": [ + { + "bbox": [ + 221, + 255, + 306, + 312 + ], + "type": "image", + "image_path": "2eb94103d8cdc037afb39b45888493f7a5fd1a85dec865c18848b8418481a0df.jpg" + } + ] + } + ], + "index": 14, + "angle": 0, + "type": "image_body" + } + ], + "index": 14 + }, + { + "type": "image", + "bbox": [ + 307, + 255, + 392, + 312 + ], + "blocks": [ + { + "bbox": [ + 307, + 255, + 392, + 312 + ], + "lines": [ + { + "bbox": [ + 307, + 255, + 392, + 312 + ], + "spans": [ + { + "bbox": [ + 307, + 255, + 392, + 312 + ], + "type": "image", + "image_path": "ba5e321021181cd0b723f72dd9d7ef58fa2551bffc890b699c73d9c12a731631.jpg" + } + ] + } + ], + "index": 16, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 311, + 312, + 391, + 320 + ], + "lines": [ + { + "bbox": [ + 311, + 312, + 391, + 320 + ], + "spans": [ + { + "bbox": [ + 311, + 312, + 391, + 320 + ], + "type": "text", + "content": "Ours (w/o env. fusion)" + } + ] + } + ], + "index": 17, + "angle": 0, + "type": "image_caption" + } + ], + "index": 16 + }, + { + "bbox": [ + 251, + 321, + 275, + 329 + ], + "angle": 0, + "lines": [ + { + "bbox": [ + 251, + 321, + 275, + 329 + ], + "spans": [ + { + "bbox": [ + 251, + 321, + 275, + 329 + ], + "type": "text", + "content": "curve)" + } + ] + } + ], + "index": 18, + "type": "text" + }, + { + "bbox": [ + 130, + 356, + 480, + 392 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 356, + 480, + 392 + ], + "spans": [ + { + "bbox": [ + 130, + 356, + 480, + 392 + ], + "type": "text", + "content": "via SDEdit [41], denoted as Ours (dataset update). Although this approach is adopted in text-guided scene editing [19], we found it challenging to produce stable gradients from such a \"discrete\" guidance." + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 130, + 392, + 480, + 476 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 392, + 480, + 476 + ], + "spans": [ + { + "bbox": [ + 130, + 392, + 480, + 476 + ], + "type": "text", + "content": "When designing the light and tone representations, we observe that the trainable tone-mapping curve provides additional flexibility to compensate for unknown tone-mapping in the input image, and better match the color and scale of the shadows. The two environment map fusion scheme (Sec. 4.3) allows to robustly recover the high intensity peaks in the early phase of training, and converge to better quality. Their respective effect is ablated as \"Ours (w/o tone curve)\" and \"Ours (w/o env. map fusion)\"." + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 132, + 491, + 230, + 506 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 491, + 230, + 506 + ], + "spans": [ + { + "bbox": [ + 132, + 491, + 230, + 506 + ], + "type": "text", + "content": "6 Applications" + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 130, + 512, + 480, + 560 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 512, + 480, + 560 + ], + "spans": [ + { + "bbox": [ + 130, + 512, + 480, + 560 + ], + "type": "text", + "content": "Since our method recovers physically based lighting information, arbitrary new virtual objects can be inserted after the optimization, as shown in Fig. 2. DiPIR can also optimize other scene attributes such as materials and local lighting. We perform preliminary experiments in this direction." + } + ] + } + ], + "index": 23 + }, + { + "bbox": [ + 130, + 567, + 481, + 675 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 567, + 481, + 675 + ], + "spans": [ + { + "bbox": [ + 130, + 567, + 481, + 675 + ], + "type": "text", + "content": "Material optimization. When combined with differentiable rendering, DMs can provide a guidance signal for material attributes, as shown in Fig. 7. Given a purely diffuse car and enabling Metallic and Roughness properties as estimizable parameters, the diffusion guidance can optimize and make the car look more shiny. By additionally changing the text prompt to \"a carmine red car\" and making the base color of the car an estimizable parameter, we show that the DM can propagate the text-condition to the PBR attribute and change the color of the car to red. When enabling local emission as an estimizable parameter, the diffusion model can also turn on the headlights of cars in night scenes." + } + ] + } + ], + "index": 24 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 153, + 91, + 448, + 102 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 153, + 91, + 448, + 102 + ], + "spans": [ + { + "bbox": [ + 153, + 91, + 448, + 102 + ], + "type": "text", + "content": "Photorealistic Object Insertion with Diffusion-Guided Inverse Rendering" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 470, + 91, + 481, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 470, + 91, + 481, + 100 + ], + "spans": [ + { + "bbox": [ + 470, + 91, + 481, + 100 + ], + "type": "text", + "content": "13" + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 12 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 133, + 114, + 221, + 183 + ], + "blocks": [ + { + "bbox": [ + 133, + 114, + 221, + 183 + ], + "lines": [ + { + "bbox": [ + 133, + 114, + 221, + 183 + ], + "spans": [ + { + "bbox": [ + 133, + 114, + 221, + 183 + ], + "type": "image", + "image_path": "6364e243daa135c0da173a6e797c14e80dd7297c20f39c9539a6eb0c7a83f193.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 214, + 186, + 318, + 196 + ], + "lines": [ + { + "bbox": [ + 214, + 186, + 318, + 196 + ], + "spans": [ + { + "bbox": [ + 214, + 186, + 318, + 196 + ], + "type": "text", + "content": "Optimize tone mapping curve" + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "image_caption" + } + ], + "index": 2 + }, + { + "type": "image", + "bbox": [ + 222, + 115, + 310, + 183 + ], + "blocks": [ + { + "bbox": [ + 222, + 115, + 310, + 183 + ], + "lines": [ + { + "bbox": [ + 222, + 115, + 310, + 183 + ], + "spans": [ + { + "bbox": [ + 222, + 115, + 310, + 183 + ], + "type": "image", + "image_path": "8cc9eb2aea1140c8dce37960fd30b0fe17d36d0415922ef0caea87fd0dfbeaec.jpg" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_body" + } + ], + "index": 3 + }, + { + "type": "image", + "bbox": [ + 312, + 115, + 399, + 183 + ], + "blocks": [ + { + "bbox": [ + 312, + 115, + 399, + 183 + ], + "lines": [ + { + "bbox": [ + 312, + 115, + 399, + 183 + ], + "spans": [ + { + "bbox": [ + 312, + 115, + 399, + 183 + ], + "type": "image", + "image_path": "146dd37cccebf6de47098121eec0b7a18fe18c063977f0dbbddcb99a6903f994.jpg" + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "image_body" + } + ], + "index": 4 + }, + { + "type": "image", + "bbox": [ + 406, + 114, + 481, + 183 + ], + "blocks": [ + { + "bbox": [ + 406, + 114, + 481, + 183 + ], + "lines": [ + { + "bbox": [ + 406, + 114, + 481, + 183 + ], + "spans": [ + { + "bbox": [ + 406, + 114, + 481, + 183 + ], + "type": "image", + "image_path": "658200a4dde3bbe76bae075467cf465d254d93baa8d9c771f33dad7d68a1f73a.jpg" + } + ] + } + ], + "index": 6, + "angle": 0, + "type": "image_body" + } + ], + "index": 6 + }, + { + "type": "image", + "bbox": [ + 435, + 182, + 454, + 200 + ], + "blocks": [ + { + "bbox": [ + 435, + 182, + 454, + 200 + ], + "lines": [ + { + "bbox": [ + 435, + 182, + 454, + 200 + ], + "spans": [ + { + "bbox": [ + 435, + 182, + 454, + 200 + ], + "type": "image", + "image_path": "e50a2cf49f962e438fc1009327adfc9955637f39d2402732e9f5c3df9844526f.jpg" + } + ] + } + ], + "index": 7, + "angle": 0, + "type": "image_body" + } + ], + "index": 7 + }, + { + "type": "image", + "bbox": [ + 133, + 200, + 221, + 268 + ], + "blocks": [ + { + "bbox": [ + 133, + 200, + 221, + 268 + ], + "lines": [ + { + "bbox": [ + 133, + 200, + 221, + 268 + ], + "spans": [ + { + "bbox": [ + 133, + 200, + 221, + 268 + ], + "type": "image", + "image_path": "296bd4293843b9a9820b90be5587f777cf6f1238b5beac1d54d0f2d8732bc6de.jpg" + } + ] + } + ], + "index": 8, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 223, + 270, + 309, + 281 + ], + "lines": [ + { + "bbox": [ + 223, + 270, + 309, + 281 + ], + "spans": [ + { + "bbox": [ + 223, + 270, + 309, + 281 + ], + "type": "text", + "content": "Optimize asset materials" + } + ] + } + ], + "index": 11, + "angle": 0, + "type": "image_caption" + } + ], + "index": 8 + }, + { + "type": "image", + "bbox": [ + 222, + 200, + 310, + 268 + ], + "blocks": [ + { + "bbox": [ + 222, + 200, + 310, + 268 + ], + "lines": [ + { + "bbox": [ + 222, + 200, + 310, + 268 + ], + "spans": [ + { + "bbox": [ + 222, + 200, + 310, + 268 + ], + "type": "image", + "image_path": "ee8f975fae2507cf9632e18414a4a9af4b4ab3c75d0da895092b5537498b2f38.jpg" + } + ] + } + ], + "index": 9, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 130, + 283, + 482, + 306 + ], + "lines": [ + { + "bbox": [ + 130, + 283, + 482, + 306 + ], + "spans": [ + { + "bbox": [ + 130, + 283, + 482, + 306 + ], + "type": "text", + "content": "Fig. 7: Our physically based inverse rendering pipeline unlocks further applications such as material, local emission and tone-mapping refinement." + } + ] + } + ], + "index": 14, + "angle": 0, + "type": "image_caption" + } + ], + "index": 9 + }, + { + "type": "image", + "bbox": [ + 312, + 200, + 399, + 268 + ], + "blocks": [ + { + "bbox": [ + 312, + 200, + 399, + 268 + ], + "lines": [ + { + "bbox": [ + 312, + 200, + 399, + 268 + ], + "spans": [ + { + "bbox": [ + 312, + 200, + 399, + 268 + ], + "type": "image", + "image_path": "fbd79d06af1fd28af9eb4b9cd3d7c8b6c2c1891c2d46384ee89a4cbb615e7eea.jpg" + } + ] + } + ], + "index": 10, + "angle": 0, + "type": "image_body" + } + ], + "index": 10 + }, + { + "type": "image", + "bbox": [ + 406, + 200, + 481, + 268 + ], + "blocks": [ + { + "bbox": [ + 406, + 200, + 481, + 268 + ], + "lines": [ + { + "bbox": [ + 406, + 200, + 481, + 268 + ], + "spans": [ + { + "bbox": [ + 406, + 200, + 481, + 268 + ], + "type": "image", + "image_path": "5cb04fccb262f75c9525d9e04105a9963b5a95bb91f028caf2349ee806891af2.jpg" + } + ] + } + ], + "index": 12, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 406, + 270, + 477, + 281 + ], + "lines": [ + { + "bbox": [ + 406, + 270, + 477, + 281 + ], + "spans": [ + { + "bbox": [ + 406, + 270, + 477, + 281 + ], + "type": "text", + "content": "Optimize emissions" + } + ] + } + ], + "index": 13, + "angle": 0, + "type": "image_caption" + } + ], + "index": 12 + }, + { + "bbox": [ + 130, + 320, + 483, + 381 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 320, + 483, + 381 + ], + "spans": [ + { + "bbox": [ + 130, + 320, + 483, + 381 + ], + "type": "text", + "content": "Tone-mapping adjustment. We use a controlled experiment to further evaluate how well DMs understand tone-mapping. In Fig. 7, we freeze the estimated environment map and apply a manual tone adjustment on the background image. The diffusion guidance is used to optimize the tone curve such that the inserted object matches the surrounding background in the final composited result." + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 132, + 389, + 217, + 402 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 389, + 217, + 402 + ], + "spans": [ + { + "bbox": [ + 132, + 389, + 217, + 402 + ], + "type": "text", + "content": "7 Discussion" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 130, + 408, + 482, + 529 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 408, + 482, + 529 + ], + "spans": [ + { + "bbox": [ + 130, + 408, + 482, + 529 + ], + "type": "text", + "content": "Our method leverages large diffusion models' inherent scene understanding capabilities as guidance to a physically based inverse rendering pipeline. We design a diffusion guidance signal with scene-specific personalization and a differentiable inverse rendering pipeline to recover lighting and tone-mapping parameters. Our method enables inserting virtual objects into scenes, but also optimizing other scene parameters such as the materials of the inserted object or account for tone-mapping mismatches between cameras. We believe that this combination of the differentiable rendering process and data-driven priors can be used successfully in many other content creation applications such as relighting and animation." + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 130, + 533, + 483, + 666 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 533, + 483, + 666 + ], + "spans": [ + { + "bbox": [ + 130, + 533, + 483, + 666 + ], + "type": "text", + "content": "Limitations and future work. Our Spherical Gaussians-based lighting representation is adequate for general objects [35], but might not behave realistically for highly specular materials. For more complex lighting representations adding generative priors on the environment map [40] is a direction worth exploring. The rendering formulation could be extended to account for effects such as reflections from the scene itself onto the inserted object (e.g. color bleeding), but that might introduce more ambiguities and require knowing the materials of the proxy geometry (refer to Supplement C.4 for failure case examples). Finally, while DM personalization significantly improves the quality of the results, it adds overhead and complexity to the pipeline. Recent personalization methods that do not require test-time finetuning [52] could be used to mitigate this overhead." + } + ] + } + ], + "index": 18 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "spans": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "type": "text", + "content": "14" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 166, + 91, + 227, + 102 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 166, + 91, + 227, + 102 + ], + "spans": [ + { + "bbox": [ + 166, + 91, + 227, + 102 + ], + "type": "text", + "content": "R. Liang et al." + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 13 + }, + { + "para_blocks": [ + { + "bbox": [ + 133, + 114, + 246, + 129 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 114, + 246, + 129 + ], + "spans": [ + { + "bbox": [ + 133, + 114, + 246, + 129 + ], + "type": "text", + "content": "Acknowledgements" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 130, + 141, + 482, + 189 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 141, + 482, + 189 + ], + "spans": [ + { + "bbox": [ + 130, + 141, + 482, + 189 + ], + "type": "text", + "content": "The authors are grateful for the feedback received from Nicholas Sharp and Huan Ling during the project. We thank the original artists of the 3D assets used in this work: inciprocal, peyman.khaleghi, Kuutti Siitonen, TurboSquid and their artists Hum3D and Amaranthus." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 132, + 209, + 198, + 220 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 209, + 198, + 220 + ], + "spans": [ + { + "bbox": [ + 132, + 209, + 198, + 220 + ], + "type": "text", + "content": "References" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 134, + 234, + 481, + 665 + ], + "type": "list", + "angle": 0, + "index": 20, + "blocks": [ + { + "bbox": [ + 138, + 234, + 428, + 245 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 234, + 428, + 245 + ], + "spans": [ + { + "bbox": [ + 138, + 234, + 428, + 245 + ], + "type": "text", + "content": "1. Poly Haven - The Public 3D Asset Library, https://polyhaven.com" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 138, + 246, + 481, + 289 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 246, + 481, + 289 + ], + "spans": [ + { + "bbox": [ + 138, + 246, + 481, + 289 + ], + "type": "text", + "content": "2. Balaji, Y., Nah, S., Huang, X., Vahdat, A., Song, J., Zhang, Q., Kreis, K., Aittala, M., Aila, T., Laine, S., Catanzaro, B., Karras, T., Liu, M.Y.: ediff-i: Text-to-image diffusion models with ensemble of expert denoisers. arXiv preprint arXiv:2211.01324 (2022)" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 138, + 290, + 481, + 322 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 290, + 481, + 322 + ], + "spans": [ + { + "bbox": [ + 138, + 290, + 481, + 322 + ], + "type": "text", + "content": "3. Bangaru, S.P., Li, T.M., Durand, F.: Unbiased warped-area sampling for differentiable rendering. ACM Trans. Graph. 39(6) (nov 2020). https://doi.org/10.1145/3414685.3417833, https://doi.org/10.1145/3414685.3417833" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 138, + 323, + 481, + 344 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 323, + 481, + 344 + ], + "spans": [ + { + "bbox": [ + 138, + 323, + 481, + 344 + ], + "type": "text", + "content": "4. Barron, J.T., Malik, J.: Shape, illumination, and reflectance from shading. IEEE transactions on pattern analysis and machine intelligence 37(8), 1670-1687 (2014)" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 138, + 346, + 481, + 367 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 346, + 481, + 367 + ], + "spans": [ + { + "bbox": [ + 138, + 346, + 481, + 367 + ], + "type": "text", + "content": "5. Barrow, H., Tenenbaum, J., Hanson, A., Riseman, E.: Recovering intrinsic scene characteristics. Comput. Vis. Syst 2, 3-26 (1978)" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 138, + 368, + 481, + 389 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 368, + 481, + 389 + ], + "spans": [ + { + "bbox": [ + 138, + 368, + 481, + 389 + ], + "type": "text", + "content": "6. Bell, S., Bala, K., Snavely, N.: Intrinsic images in the wild. ACM Transactions on Graphics (TOG) 33(4), 159 (2014)" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 138, + 390, + 481, + 422 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 390, + 481, + 422 + ], + "spans": [ + { + "bbox": [ + 138, + 390, + 481, + 422 + ], + "type": "text", + "content": "7. Black, M.J., Anandan, P.: The robust estimation of multiple motions: Parametric and piecewise-smooth flow fields. Computer vision and image understanding 63(1), 75-104 (1996)" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 138, + 423, + 481, + 444 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 423, + 481, + 444 + ], + "spans": [ + { + "bbox": [ + 138, + 423, + 481, + 444 + ], + "type": "text", + "content": "8. Boss, M., Jampani, V., Kim, K., Lensch, H.P., Kautz, J.: Two-shot spatially-varying brdf and shape estimation. In: CVPR (2020)" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 138, + 445, + 481, + 466 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 445, + 481, + 466 + ], + "spans": [ + { + "bbox": [ + 138, + 445, + 481, + 466 + ], + "type": "text", + "content": "9. Bousseau, A., Paris, S., Durand, F.: User-assisted intrinsic images. In: ACM Transactions on Graphics (TOG). vol. 28, p. 130. ACM (2009)" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 134, + 468, + 481, + 499 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 134, + 468, + 481, + 499 + ], + "spans": [ + { + "bbox": [ + 134, + 468, + 481, + 499 + ], + "type": "text", + "content": "10. Chari, P., Ma, S., Ostashev, D., Kadambi, A., Krishnan, G., Wang, J., Aberman, K.: Personalized restoration via dual-pivot tuning. arXiv preprint arXiv:2312.17234 (2023)" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 134, + 501, + 481, + 533 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 134, + 501, + 481, + 533 + ], + "spans": [ + { + "bbox": [ + 134, + 501, + 481, + 533 + ], + "type": "text", + "content": "11. Dai, X., Hou, J., Ma, C.Y., Tsai, S., Wang, J., Wang, R., Zhang, P., Vandenhende, S., Wang, X., Dubey, A., et al.: Emu: Enhancing image generation models using photogenic needles in a haystack. arXiv preprint arXiv:2309.15807 (2023)" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 134, + 533, + 481, + 576 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 134, + 533, + 481, + 576 + ], + "spans": [ + { + "bbox": [ + 134, + 533, + 481, + 576 + ], + "type": "text", + "content": "12. Dastjerdi, M.R.K., Eisenmann, J., Hold-Geoffroy, Y., Lalonde, J.F.: Everlight: Indoor-outdoor editable hdr lighting estimation. In: Proceedings of the IEEE/CVF International Conference on Computer Vision (ICCV). pp. 7420-7429 (October 2023)" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 134, + 577, + 481, + 598 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 134, + 577, + 481, + 598 + ], + "spans": [ + { + "bbox": [ + 134, + 577, + 481, + 598 + ], + "type": "text", + "content": "13. Durkan, C., Bekasov, A., Murray, I., Papamakarios, G.: Neural spline flows. Advances in neural information processing systems 32 (2019)" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 134, + 600, + 481, + 642 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 134, + 600, + 481, + 642 + ], + "spans": [ + { + "bbox": [ + 134, + 600, + 481, + 642 + ], + "type": "text", + "content": "14. Gal, R., Alaluf, Y., Atzmon, Y., Patashnik, O., Bermano, A.H., Chechik, G., Cohen-Or, D.: An image is worth one word: Personalizing text-to-image generation using textual inversion (2022). https://doi.org/10.48550/ARXIV.2208.01618, https://arxiv.org/abs/2208.01618" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 134, + 643, + 481, + 665 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 134, + 643, + 481, + 665 + ], + "spans": [ + { + "bbox": [ + 134, + 643, + 481, + 665 + ], + "type": "text", + "content": "15. Gardner, M.A., Hold-Geoffroy, Y., Sunkavalli, K., Gagné, C., Lalonde, J.F.: Deep parametric indoor lighting estimation. In: ICCV. pp. 7175-7183 (2019)" + } + ] + } + ], + "index": 19 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 153, + 91, + 447, + 102 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 153, + 91, + 447, + 102 + ], + "spans": [ + { + "bbox": [ + 153, + 91, + 447, + 102 + ], + "type": "text", + "content": "Photorealistic Object Insertion with Diffusion-Guided Inverse Rendering" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 470, + 92, + 480, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 470, + 92, + 480, + 100 + ], + "spans": [ + { + "bbox": [ + 470, + 92, + 480, + 100 + ], + "type": "text", + "content": "15" + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 14 + }, + { + "para_blocks": [ + { + "bbox": [ + 132, + 116, + 481, + 665 + ], + "type": "list", + "angle": 0, + "index": 22, + "blocks": [ + { + "bbox": [ + 133, + 116, + 481, + 149 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 116, + 481, + 149 + ], + "spans": [ + { + "bbox": [ + 133, + 116, + 481, + 149 + ], + "type": "text", + "content": "16. Gardner, M.A., Sunkavalli, K., Yumer, E., Shen, X., Gambaretto, E., Gagné, C., Lalonde, J.F.: Learning to predict indoor illumination from a single image. arXiv preprint arXiv:1704.00090 (2017)" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 133, + 149, + 481, + 171 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 149, + 481, + 171 + ], + "spans": [ + { + "bbox": [ + 133, + 149, + 481, + 171 + ], + "type": "text", + "content": "17. Garon, M., Sunkavalli, K., Hadap, S., Carr, N., Lalonde, J.F.: Fast spatially-varying indoor lighting estimation. In: CVPR. pp. 6908-6917 (2019)" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 133, + 171, + 481, + 203 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 171, + 481, + 203 + ], + "spans": [ + { + "bbox": [ + 133, + 171, + 481, + 203 + ], + "type": "text", + "content": "18. Grosse, R., Johnson, M.K., Adelson, E.H., Freeman, W.T.: Ground truth dataset and baseline evaluations for intrinsic image algorithms. In: ICCV. pp. 2335-2342. IEEE (2009)" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 133, + 203, + 481, + 236 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 203, + 481, + 236 + ], + "spans": [ + { + "bbox": [ + 133, + 203, + 481, + 236 + ], + "type": "text", + "content": "19. Haque, A., Tancik, M., Efros, A., Holynski, A., Kanazawa, A.: Instruct-nerf2nerf: Editing 3d scenes with instructions. In: Proceedings of the IEEE/CVF International Conference on Computer Vision (2023)" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 132, + 236, + 481, + 257 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 236, + 481, + 257 + ], + "spans": [ + { + "bbox": [ + 132, + 236, + 481, + 257 + ], + "type": "text", + "content": "20. Ho, J., Jain, A., Abbeel, P.: Denoising diffusion probabilistic models. arXiv preprint arxiv:2006.11239 (2020)" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 132, + 257, + 481, + 278 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 257, + 481, + 278 + ], + "spans": [ + { + "bbox": [ + 132, + 257, + 481, + 278 + ], + "type": "text", + "content": "21. Ho, J., Salimans, T.: Classifier-free diffusion guidance. arXiv preprint arXiv:2207.12598 (2022)" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 132, + 278, + 481, + 300 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 278, + 481, + 300 + ], + "spans": [ + { + "bbox": [ + 132, + 278, + 481, + 300 + ], + "type": "text", + "content": "22. Hold-Geoffroy, Y., Athawale, A., Lalonde, J.F.: Deep sky modeling for single image outdoor lighting estimation. In: CVPR. pp. 6927-6935 (2019)" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 132, + 300, + 481, + 321 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 300, + 481, + 321 + ], + "spans": [ + { + "bbox": [ + 132, + 300, + 481, + 321 + ], + "type": "text", + "content": "23. Hold-Geoffroy, Y., Sunkavalli, K., Hadap, S., Gambaretto, E., Lalonde, J.F.: Deep outdoor illumination estimation. In: CVPR. pp. 7312-7321 (2017)" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 132, + 321, + 481, + 363 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 321, + 481, + 363 + ], + "spans": [ + { + "bbox": [ + 132, + 321, + 481, + 363 + ], + "type": "text", + "content": "24. Hu, E.J., yelong shen, Wallis, P., Allen-Zhu, Z., Li, Y., Wang, S., Wang, L., Chen, W.: LoRA: Low-rank adaptation of large language models. In: International Conference on Learning Representations (2022), https://openreview.net/forum?id=nZeVKeeFYf9" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 132, + 363, + 481, + 397 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 363, + 481, + 397 + ], + "spans": [ + { + "bbox": [ + 132, + 363, + 481, + 397 + ], + "type": "text", + "content": "25. Jakob, W., Speierer, S., Roussel, N., Nimier-David, M., Vicini, D., Zeltner, T., Nicolet, B., Crespo, M., Leroy, V., Zhang, Z.: Mitsuba 3 renderer (2022), https://mitsubarenderer.org" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 132, + 397, + 481, + 418 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 397, + 481, + 418 + ], + "spans": [ + { + "bbox": [ + 132, + 397, + 481, + 418 + ], + "type": "text", + "content": "26. Kajiya, J.T.: The rendering equation. In: Proceedings of the 13th annual conference on Computer graphics and interactive techniques. pp. 143-150 (1986)" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 132, + 418, + 481, + 450 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 418, + 481, + 450 + ], + "spans": [ + { + "bbox": [ + 132, + 418, + 481, + 450 + ], + "type": "text", + "content": "27. Karimi Dastjerdi, M.R., Hold-Geoffroy, Y., Eisenmann, J., Khodadadeh, S., Lalonde, J.F.: Guided co-modulated gan for 360 field of view extrapolation. International Conference on 3D Vision (3DV) (2022)" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 132, + 450, + 481, + 494 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 450, + 481, + 494 + ], + "spans": [ + { + "bbox": [ + 132, + 450, + 481, + 494 + ], + "type": "text", + "content": "28. Ke, B., Obukhov, A., Huang, S., Metzger, N., Daudt, R.C., Schindler, K.: Repurposing diffusion-based image generators for monocular depth estimation. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR) (2024)" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 132, + 494, + 481, + 515 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 494, + 481, + 515 + ], + "spans": [ + { + "bbox": [ + 132, + 494, + 481, + 515 + ], + "type": "text", + "content": "29. Kocsis, P., Sitzmann, V., Nießner, M.: Intrinsic image diffusion for single-view material estimation. In: arxiv (2023)" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 132, + 515, + 481, + 536 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 515, + 481, + 536 + ], + "spans": [ + { + "bbox": [ + 132, + 515, + 481, + 536 + ], + "type": "text", + "content": "30. Kovacs, B., Bell, S., Snavely, N., Bala, K.: Shading annotations in the wild. In: CVPR. pp. 6998-7007 (2017)" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 132, + 536, + 481, + 547 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 536, + 481, + 547 + ], + "spans": [ + { + "bbox": [ + 132, + 536, + 481, + 547 + ], + "type": "text", + "content": "31. Land, E.H., McCann, J.J.: Lightness and retina theory. Josa 61(1), 1-11 (1971)" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 132, + 547, + 481, + 579 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 547, + 481, + 579 + ], + "spans": [ + { + "bbox": [ + 132, + 547, + 481, + 579 + ], + "type": "text", + "content": "32. LeGendre, C., Ma, W.C., Fyffe, G., Flynn, J., Charbonnel, L., Busch, J., Debevec, P.: Deeplight: Learning illumination for unconstrained mobile mixed reality. In: CVPR. pp. 5918-5928 (2019)" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 132, + 579, + 481, + 611 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 579, + 481, + 611 + ], + "spans": [ + { + "bbox": [ + 132, + 579, + 481, + 611 + ], + "type": "text", + "content": "33. Li, T.M., Aittala, M., Durand, F., Lehtinen, J.: Differentiable monte carlo ray tracing through edge sampling. ACM Trans. Graph. 37(6) (dec 2018). https://doi.org/10.1145/3272127.3275109, https://doi.org/10.1145/3272127.3275109" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 132, + 611, + 481, + 633 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 611, + 481, + 633 + ], + "spans": [ + { + "bbox": [ + 132, + 611, + 481, + 633 + ], + "type": "text", + "content": "34. Li, Z., Snavely, N.: Cginintrinsics: Better intrinsic image decomposition through physically-based rendering. In: ECCV. pp. 371-387 (2018)" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 132, + 633, + 481, + 665 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 633, + 481, + 665 + ], + "spans": [ + { + "bbox": [ + 132, + 633, + 481, + 665 + ], + "type": "text", + "content": "35. Li, Z., Shafiei, M., Ramamoorthi, R., Sunkavalli, K., Chandraker, M.: Inverse rendering for complex indoor scenes: Shape, spatially-varying lighting and svbrdf from a single image. In: CVPR. pp. 2475-2484 (2020)" + } + ] + } + ], + "index": 21 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "spans": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "type": "text", + "content": "16" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 166, + 91, + 226, + 101 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 166, + 91, + 226, + 101 + ], + "spans": [ + { + "bbox": [ + 166, + 91, + 226, + 101 + ], + "type": "text", + "content": "R. Liang et al." + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 15 + }, + { + "para_blocks": [ + { + "bbox": [ + 132, + 116, + 481, + 665 + ], + "type": "list", + "angle": 0, + "index": 20, + "blocks": [ + { + "bbox": [ + 132, + 116, + 480, + 149 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 116, + 480, + 149 + ], + "spans": [ + { + "bbox": [ + 132, + 116, + 480, + 149 + ], + "type": "text", + "content": "36. Li, Z., Xu, Z., Ramamoorthi, R., Sunkavalli, K., Chandraker, M.: Learning to reconstruct shape and spatially-varying reflectance from a single image. ACM Transactions on Graphics (TOG) 37(6), 1-11 (2018)" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 132, + 151, + 481, + 183 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 151, + 481, + 183 + ], + "spans": [ + { + "bbox": [ + 132, + 151, + 481, + 183 + ], + "type": "text", + "content": "37. Li, Z., Yu, L., Okunev, M., Chandraker, M., Dong, Z.: Spatiotemporally consistent hdr indoor lighting estimation. ACM Trans. Graph. 42(3) (jun 2023). https://doi.org/10.1145/3595921, https://doi.org/10.1145/3595921" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 132, + 184, + 480, + 217 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 184, + 480, + 217 + ], + "spans": [ + { + "bbox": [ + 132, + 184, + 480, + 217 + ], + "type": "text", + "content": "38. Li, Z., Yu, T.W., Sang, S., Wang, S., Bi, S., Xu, Z., Yu, H.X., Sunkavalli, K., Hasan, M., Ramamoorthi, R., et al.: Openrooms: An end-to-end open framework for photorealistic indoor scene datasets. arXiv preprint arXiv:2007.12868 (2020)" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 132, + 218, + 481, + 251 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 218, + 481, + 251 + ], + "spans": [ + { + "bbox": [ + 132, + 218, + 481, + 251 + ], + "type": "text", + "content": "39. Loubet, G., Holzschuch, N., Jakob, W.: Reparameterizing discontinuous integrands for differentiable rendering. ACM Trans. Graph. 38(6) (nov 2019). https://doi.org/10.1145/3355089.3356510, https://doi.org/10.1145/3355089.3356510" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 132, + 251, + 481, + 283 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 251, + 481, + 283 + ], + "spans": [ + { + "bbox": [ + 132, + 251, + 481, + 283 + ], + "type": "text", + "content": "40. Lyu, L., Tewari, A., Habermann, M., Saito, S., Zollhöfer, M., Leimkuehler, T., Theobalt, C.: Diffusion posterior illumination for ambiguity-aware inverse rendering. ACM Transactions on Graphics 42(6) (2023)" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 132, + 285, + 480, + 317 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 285, + 480, + 317 + ], + "spans": [ + { + "bbox": [ + 132, + 285, + 480, + 317 + ], + "type": "text", + "content": "41. Meng, C., He, Y., Song, Y., Song, J., Wu, J., Zhu, J.Y., Ermon, S.: SDEdit: Guided image synthesis and editing with stochastic differential equations. In: International Conference on Learning Representations (2022)" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 132, + 318, + 481, + 361 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 318, + 481, + 361 + ], + "spans": [ + { + "bbox": [ + 132, + 318, + 481, + 361 + ], + "type": "text", + "content": "42. Nimier-David, M., Speierer, S., Ruiz, B., Jakob, W.: Radiative backpropagation: an adjoint method for lightning-fast differentiable rendering. ACM Trans. Graph. 39(4) (aug 2020). https://doi.org/10.1145/3386569.3392406, https://doi.org/10.1145/3386569.3392406" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 132, + 363, + 481, + 396 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 363, + 481, + 396 + ], + "spans": [ + { + "bbox": [ + 132, + 363, + 481, + 396 + ], + "type": "text", + "content": "43. Nimier-David, M., Vicini, D., Zeltner, T., Jakob, W.: Mitsuba 2: a retargetable forward and inverse renderer. ACM Trans. Graph. 38(6) (nov 2019). https://doi.org/10.1145/3355089.3356498, https://doi.org/10.1145/3355089.3356498" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 132, + 396, + 481, + 429 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 396, + 481, + 429 + ], + "spans": [ + { + "bbox": [ + 132, + 396, + 481, + 429 + ], + "type": "text", + "content": "44. Phongthawee, P., Chinchuthakun, W., Sinsunthithet, N., Raj, A., Jampani, V., Khungurn, P., Suwajanakorn, S.: Diffusionlight: Light probes for free by painting a chrome ball. In: ArXiv (2023)" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 132, + 430, + 480, + 452 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 430, + 480, + 452 + ], + "spans": [ + { + "bbox": [ + 132, + 430, + 480, + 452 + ], + "type": "text", + "content": "45. Poole, B., Jain, A., Barron, J.T., Mildenhall, B.: Dreamfusion: Text-to-3d using 2d diffusion. arXiv (2022)" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 132, + 453, + 480, + 474 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 453, + 480, + 474 + ], + "spans": [ + { + "bbox": [ + 132, + 453, + 480, + 474 + ], + "type": "text", + "content": "46. Rombach, R., Blattmann, A., Lorenz, D., Esser, P., Ommer, B.: High-resolution image synthesis with latent diffusion models (2021)" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 132, + 475, + 481, + 496 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 475, + 481, + 496 + ], + "spans": [ + { + "bbox": [ + 132, + 475, + 481, + 496 + ], + "type": "text", + "content": "47. Ruiz, N., Li, Y., Jampani, V., Pritch, Y., Rubinstein, M., Aberman, K.: Dreambooth: Fine tuning text-to-image diffusion models for subject-driven generation (2022)" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 132, + 498, + 481, + 541 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 498, + 481, + 541 + ], + "spans": [ + { + "bbox": [ + 132, + 498, + 481, + 541 + ], + "type": "text", + "content": "48. Sahara, C., Chan, W., Saxena, S., Li, L., Whang, J., Denton, E.L., Ghasemipour, K., Gontijo Lopes, R., Karagol Ayan, B., Salimans, T., et al.: Photorealistic text-to-image diffusion models with deep language understanding. Advances in Neural Information Processing Systems 35, 36479-36494 (2022)" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 132, + 543, + 480, + 574 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 543, + 480, + 574 + ], + "spans": [ + { + "bbox": [ + 132, + 543, + 480, + 574 + ], + "type": "text", + "content": "49. Sarkar, A., Mai, H., Mahapatra, A., Lazebnik, S., Forsyth, D., Bhattad, A.: Shadows don't lie and lines can't bend! generative models don't know projective geometry...for now (2023)" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 132, + 576, + 480, + 597 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 576, + 480, + 597 + ], + "spans": [ + { + "bbox": [ + 132, + 576, + 480, + 597 + ], + "type": "text", + "content": "50. Sengupta, S., Gu, J., Kim, K., Liu, G., Jacobs, D.W., Kautz, J.: Neural inverse rendering of an indoor scene from a single image. In: ICCV (2019)" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 132, + 598, + 480, + 620 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 598, + 480, + 620 + ], + "spans": [ + { + "bbox": [ + 132, + 598, + 480, + 620 + ], + "type": "text", + "content": "51. Shah, V., Ruiz, N., Cole, F., Lu, E., Lazebnik, S., Li, Y., Jampani, V.:Ziplora: Any subject in any style by effectively merging loras (2023)" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 132, + 621, + 480, + 643 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 621, + 480, + 643 + ], + "spans": [ + { + "bbox": [ + 132, + 621, + 480, + 643 + ], + "type": "text", + "content": "52. Shi, J., Xiong, W., Lin, Z., Jung, H.J.: Instantbooth: Personalized text-to-image generation without test-time finetuning. arXiv preprint arXiv:2304.03411 (2023)" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 132, + 643, + 480, + 665 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 643, + 480, + 665 + ], + "spans": [ + { + "bbox": [ + 132, + 643, + 480, + 665 + ], + "type": "text", + "content": "53. Song, S., Funkhouser, T.: Neural illumination: Lighting prediction for indoor environments. In: CVPR. pp. 6918-6926 (2019)" + } + ] + } + ], + "index": 19 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 153, + 91, + 447, + 102 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 153, + 91, + 447, + 102 + ], + "spans": [ + { + "bbox": [ + 153, + 91, + 447, + 102 + ], + "type": "text", + "content": "Photorealistic Object Insertion with Diffusion-Guided Inverse Rendering" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 470, + 91, + 481, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 470, + 91, + 481, + 100 + ], + "spans": [ + { + "bbox": [ + 470, + 91, + 481, + 100 + ], + "type": "text", + "content": "17" + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 16 + }, + { + "para_blocks": [ + { + "bbox": [ + 132, + 116, + 482, + 665 + ], + "type": "list", + "angle": 0, + "index": 20, + "blocks": [ + { + "bbox": [ + 132, + 116, + 482, + 149 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 116, + 482, + 149 + ], + "spans": [ + { + "bbox": [ + 132, + 116, + 482, + 149 + ], + "type": "text", + "content": "54. Srinivasan, P.P., Mildenhall, B., Tancik, M., Barron, J.T., Tucker, R., Snavely, N.: Lighthouse: Predicting lighting volumes for spatially-coherent illumination. In: CVPR. pp. 8080-8089 (2020)" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 132, + 150, + 482, + 183 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 150, + 482, + 183 + ], + "spans": [ + { + "bbox": [ + 132, + 150, + 482, + 183 + ], + "type": "text", + "content": "55. Sun, P., Kretzschmar, H., Dotiwalla, X., Chouard, A., Patnaik, V., Tsui, P., Guo, J., Zhou, Y., Chai, Y., Caine, B., et al.: Scalability in perception for autonomous driving: Waymo open dataset. In: CVPR (2020)" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 132, + 184, + 481, + 206 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 184, + 481, + 206 + ], + "spans": [ + { + "bbox": [ + 132, + 184, + 481, + 206 + ], + "type": "text", + "content": "56. Tang, J., Zhu, Y., Wang, H., Chan, J.H., Li, S., Shi, B.: Estimating spatially-varying lighting in urban scenes with disentangled representation. In: ECCV (2022)" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 132, + 207, + 481, + 239 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 207, + 481, + 239 + ], + "spans": [ + { + "bbox": [ + 132, + 207, + 481, + 239 + ], + "type": "text", + "content": "57. Tang, L., Ruiz, N., Qinghao, C., Li, Y., Holynski, A., Jacobs, D.E., Hariharan, B., Pritch, Y., Wadhwa, N., Aberman, K., Rubinstein, M.: Realfill: Reference-driven generation for authentic image completion. arXiv preprint arXiv:2309.16668 (2023)" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 132, + 240, + 481, + 262 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 240, + 481, + 262 + ], + "spans": [ + { + "bbox": [ + 132, + 240, + 481, + 262 + ], + "type": "text", + "content": "58. Tao, A., Sapra, K., Catanzaro, B.: Hierarchical multi-scale attention for semantic segmentation. arXiv preprint arXiv:2005.10821 (2020)" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 132, + 263, + 481, + 296 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 263, + 481, + 296 + ], + "spans": [ + { + "bbox": [ + 132, + 263, + 481, + 296 + ], + "type": "text", + "content": "59. Veach, E., Guibas, L.J.: Optimally combining sampling techniques for monte carlo rendering. In: Proceedings of the 22nd annual conference on Computer graphics and interactive techniques. pp. 419-428 (1995)" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 132, + 297, + 481, + 339 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 297, + 481, + 339 + ], + "spans": [ + { + "bbox": [ + 132, + 297, + 481, + 339 + ], + "type": "text", + "content": "60. Vicini, D., Speierer, S., Jakob, W.: Path replay backpropagation: differentiating light paths using constant memory and linear time. ACM Trans. Graph. 40(4) (jul 2021). https://doi.org/10.1145/3450626.3459804, https://doi.org/10.1145/3450626.3459804" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 132, + 340, + 481, + 373 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 340, + 481, + 373 + ], + "spans": [ + { + "bbox": [ + 132, + 340, + 481, + 373 + ], + "type": "text", + "content": "61. Wang, G., Yang, Y., Loy, C.C., Liu, Z.: Stylelight: HDR panorama generation for lighting estimation and editing. In: European Conference on Computer Vision (ECCV) (2022)" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 132, + 374, + 481, + 397 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 374, + 481, + 397 + ], + "spans": [ + { + "bbox": [ + 132, + 374, + 481, + 397 + ], + "type": "text", + "content": "62. Wang, Z., Chen, W., Acuna, D., Kautz, J., Fidler, S.: Neural light field estimation for street scenes with differentiable virtual object insertion. In: ECCV (2022)" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 132, + 397, + 481, + 418 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 397, + 481, + 418 + ], + "spans": [ + { + "bbox": [ + 132, + 397, + 481, + 418 + ], + "type": "text", + "content": "63. Wang, Z., Philion, J., Fidler, S., Kautz, J.: Learning indoor inverse rendering with 3d spatially-varying lighting. In: ICCV (2021)" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 132, + 419, + 481, + 441 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 419, + 481, + 441 + ], + "spans": [ + { + "bbox": [ + 132, + 419, + 481, + 441 + ], + "type": "text", + "content": "64. Wimbauer, F., Wu, S., Rupprecht, C.: De-rendering 3d objects in the wild. In: CVPR (2022)" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 132, + 442, + 481, + 485 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 442, + 481, + 485 + ], + "spans": [ + { + "bbox": [ + 132, + 442, + 481, + 485 + ], + "type": "text", + "content": "65. Yan, K., Lassner, C., Budge, B., Dong, Z., Zhao, S.: Efficient estimation of boundary integrals for path-space differentiable rendering. ACM Trans. Graph. 41(4) (jul 2022). https://doi.org/10.1145/3528223.3530080, https://doi.org/10.1145/3528223.3530080" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 132, + 486, + 481, + 520 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 486, + 481, + 520 + ], + "spans": [ + { + "bbox": [ + 132, + 486, + 481, + 520 + ], + "type": "text", + "content": "66. Yang, J., Ivanovic, B., Litany, O., Weng, X., Kim, S.W., Li, B., Che, T., Xu, D., Fidler, S., Pavone, M., Wang, Y.: Emernerf: Emergent spatial-temporal scene decomposition via self-supervision. arXiv preprint arXiv:2311.02077 (2023)" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 132, + 520, + 481, + 553 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 520, + 481, + 553 + ], + "spans": [ + { + "bbox": [ + 132, + 520, + 481, + 553 + ], + "type": "text", + "content": "67. Yu, H.X., Agarwala, S., Herrmann, C., Szeliski, R., Snavely, N., Wu, J., Sun, D.: Accidental light probes. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 12521-12530 (2023)" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 132, + 554, + 481, + 575 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 554, + 481, + 575 + ], + "spans": [ + { + "bbox": [ + 132, + 554, + 481, + 575 + ], + "type": "text", + "content": "68. Yu, X., Guo, Y.C., Li, Y., Liang, D., Zhang, S.H., Qi, X.: Text-to-3d with classifier score distillation (2023)" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 132, + 576, + 481, + 597 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 576, + 481, + 597 + ], + "spans": [ + { + "bbox": [ + 132, + 576, + 481, + 597 + ], + "type": "text", + "content": "69. Yu, Y., Smith, W.A.: Inverserendernet: Learning single image inverse rendering. In: CVPR (2019)" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 132, + 598, + 481, + 632 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 598, + 481, + 632 + ], + "spans": [ + { + "bbox": [ + 132, + 598, + 481, + 632 + ], + "type": "text", + "content": "70. Zhan, F., Zhang, C., Yu, Y., Chang, Y., Lu, S., Ma, F., Xie, X.: Emlight: Lighting estimation via spherical distribution approximation. In: Proceedings of the AAAI Conference on Artificial Intelligence (2021)" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 132, + 632, + 481, + 665 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 632, + 481, + 665 + ], + "spans": [ + { + "bbox": [ + 132, + 632, + 481, + 665 + ], + "type": "text", + "content": "71. Zhang, C., Miller, B., Yan, K., Gkioulekas, I., Zhao, S.: Path-space differentiable rendering. ACM Trans. Graph. 39(4) (aug 2020). https://doi.org/10.1145/3386569.3392383, https://doi.org/10.1145/3386569.3392383" + } + ] + } + ], + "index": 19 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "spans": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "type": "text", + "content": "18" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 166, + 91, + 227, + 102 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 166, + 91, + 227, + 102 + ], + "spans": [ + { + "bbox": [ + 166, + 91, + 227, + 102 + ], + "type": "text", + "content": "R. Liang et al." + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 17 + }, + { + "para_blocks": [ + { + "bbox": [ + 132, + 116, + 482, + 281 + ], + "type": "list", + "angle": 0, + "index": 8, + "blocks": [ + { + "bbox": [ + 132, + 116, + 482, + 149 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 116, + 482, + 149 + ], + "spans": [ + { + "bbox": [ + 132, + 116, + 482, + 149 + ], + "type": "text", + "content": "72. Zhang, C., Yu, Z., Zhao, S.: Path-space differentiable rendering of participating media. ACM Trans. Graph. 40(4) (jul 2021). https://doi.org/10.1145/3450626.3459782, https://doi.org/10.1145/3450626.3459782" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 132, + 150, + 481, + 183 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 150, + 481, + 183 + ], + "spans": [ + { + "bbox": [ + 132, + 150, + 481, + 183 + ], + "type": "text", + "content": "73. Zhang, R., Isola, P., Efros, A.A., Shechtman, E., Wang, O.: The unreasonable effectiveness of deep features as a perceptual metric. In: Proceedings of the IEEE conference on computer vision and pattern recognition. pp. 586-595 (2018)" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 132, + 183, + 481, + 215 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 183, + 481, + 215 + ], + "spans": [ + { + "bbox": [ + 132, + 183, + 481, + 215 + ], + "type": "text", + "content": "74. Zhang, Z., Roussel, N., Jakob, W.: Projective sampling for differentiable rendering of geometry. ACM Trans. Graph. 42(6) (dec 2023). https://doi.org/10.1145/3618385, https://doi.org/10.1145/3618385" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 132, + 216, + 481, + 237 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 216, + 481, + 237 + ], + "spans": [ + { + "bbox": [ + 132, + 216, + 481, + 237 + ], + "type": "text", + "content": "75. Zhao, Q., Tan, P., Dai, Q., Shen, L., Wu, E., Lin, S.: A closed-form solution to retinax with nonlocal texture constraints 34(7), 1437-1444 (2012)" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 132, + 237, + 481, + 259 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 237, + 481, + 259 + ], + "spans": [ + { + "bbox": [ + 132, + 237, + 481, + 259 + ], + "type": "text", + "content": "76. Zhao, Y., Guo, T.: Pointar: Efficient lighting estimation for mobile augmented reality. arXiv preprint arXiv:2004.00006 (2020)" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 132, + 259, + 481, + 281 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 259, + 481, + 281 + ], + "spans": [ + { + "bbox": [ + 132, + 259, + 481, + 281 + ], + "type": "text", + "content": "77. Zhu, Y., Zhang, Y., Li, S., Shi, B.: Spatially-varying outdoor lighting estimation from intrinsics. In: CVPR (2021)" + } + ] + } + ], + "index": 7 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 153, + 91, + 447, + 102 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 153, + 91, + 447, + 102 + ], + "spans": [ + { + "bbox": [ + 153, + 91, + 447, + 102 + ], + "type": "text", + "content": "Photorealistic Object Insertion with Diffusion-Guided Inverse Rendering" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 470, + 91, + 481, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 470, + 91, + 481, + 100 + ], + "spans": [ + { + "bbox": [ + 470, + 91, + 481, + 100 + ], + "type": "text", + "content": "19" + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 18 + } + ], + "_backend": "vlm", + "_version_name": "2.6.4" +} \ No newline at end of file diff --git a/2024/Photorealistic Video Generation with Diffusion Models/327f0bbf-9985-4862-ae62-9cd9ec84c015_content_list.json b/2024/Photorealistic Video Generation with Diffusion Models/327f0bbf-9985-4862-ae62-9cd9ec84c015_content_list.json new file mode 100644 index 0000000000000000000000000000000000000000..58b1836ca5e41186b759b1d41de666dcec04dfba --- /dev/null +++ b/2024/Photorealistic Video Generation with Diffusion Models/327f0bbf-9985-4862-ae62-9cd9ec84c015_content_list.json @@ -0,0 +1,2015 @@ +[ + { + "type": "text", + "text": "Photorealistic Video Generation with Diffusion Models", + "text_level": 1, + "bbox": [ + 232, + 140, + 772, + 184 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Agrim Gupta $^{1,2,*}$ , Lijun Yu $^{2}$ , Kihyuk Sohn $^{2}$ , Xiuye Gu $^{2}$ , Meera Hahn $^{2}$ , Fei-Fei Li $^{1}$ , Irfan Essa $^{2,3}$ , Lu Jiang $^{2}$ , and José Lezama $^{2}$", + "bbox": [ + 217, + 210, + 785, + 244 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "1 Stanford University", + "bbox": [ + 426, + 253, + 575, + 268 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "2 Google Research", + "bbox": [ + 437, + 268, + 563, + 282 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "3 Georgia Institute of Technology", + "bbox": [ + 387, + 282, + 614, + 297 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "https://walt-video-diffusion.github.io/", + "bbox": [ + 349, + 297, + 653, + 310 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/19f0fe423e933f545060e36b9bc5b155b5f95cbd7e4131958491645bc98ed314.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 217, + 325, + 330, + 377 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/5826244976bd7e7d23345decf1df76c8d3b3d26043fad67d77e1210e3b7f2a50.jpg", + "image_caption": [ + "A cute cat riding a yellow surfboard, fast waves." + ], + "image_footnote": [], + "bbox": [ + 330, + 325, + 442, + 377 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/421200607f84cd04c391726528a1354659e5e1d4219e97f2046347ee1947f759.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 442, + 325, + 555, + 377 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/b9fe99a791eb987a86b0054656e3ea8274efa49b4854bb716fbc4d2c3a139fc5.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 555, + 325, + 668, + 377 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/7464fe23b9b93ab68c099ebff01c971496686123e5d797a0d036f1814bfccff3.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 669, + 325, + 782, + 377 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/e6cee2ae0357022c32630f622964095cd5f0cc2c629d0bb0c9e33902137efdef.jpg", + "image_caption": [ + "A cute cordi wearing a blue sweater, walking towards the camera, slow motion." + ], + "image_footnote": [], + "bbox": [ + 217, + 396, + 330, + 448 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/0e881ba7887a6586e6286ca479834a2fb97c0435e64e7e7a78608f5b9dc2e50c.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 330, + 396, + 442, + 448 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/5d495a6f51d47964d0c8865d31471091a93fcfc3c8c547f8a59119d564434fc5.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 442, + 396, + 555, + 448 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/cc33d11addaedb8aef7f3ce72761f04ba5e014dacd947edba35c98be0de4acdc.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 555, + 396, + 668, + 448 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/dce248c507f5b8a4286e0fd3a9cdc05a54ac6ef936246d06624948ddfd7a8b12.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 669, + 396, + 782, + 448 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/9184a5a762781a0348298aef64f01f53d2760718d9f6ebb2dabd9c4c35fc69c9.jpg", + "image_caption": [ + "A raccoon wearing a black jacket, dancing in slow motion in front of the pyramids." + ], + "image_footnote": [], + "bbox": [ + 217, + 470, + 330, + 523 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/eb9aa9125b66742d3426c50f21429f2461b625f90fb6552b7410797b5d3b57ad.jpg", + "image_caption": [ + "Fig. 1: W.A.L.T samples for text-to-video generation. W.A.L.T generates high-resolution, temporally consistent photorealistic videos from text prompts. The samples shown are $512 \\times 896$ resolution over 3.6 seconds duration at 8 frames per second." + ], + "image_footnote": [], + "bbox": [ + 330, + 470, + 442, + 523 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/c69d86db9537d85db7a46a005601f41462aa78e40b73b2c333d0fb77d8393e75.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 442, + 470, + 555, + 523 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/fce371586e81973d39130c746f40e2130e3824b4a98199858d584bde1b879b81.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 555, + 470, + 668, + 523 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/4639bffb1849a85039b93335e4ae288a124be8e418f63e11bf3d13e79f436aee.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 669, + 470, + 782, + 523 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Abstract. We present W.A.L.T, a diffusion transformer for photorealistic video generation from text prompts. Our approach has two key design decisions. First, we use a causal encoder to jointly compress images and videos within a unified latent space, enabling training and generation across modalities. Second, for memory and training efficiency, we use a window attention architecture tailored for joint spatial and spatiotemporal generative modeling. Taken together these design decisions enable us to achieve state-of-the-art performance on established video (UCF-101 and Kinetics-600) and image (ImageNet) generation benchmarks without using classifier free guidance. Finally, we also train a cascade of three models for the task of text-to-video generation consisting of a base latent video diffusion model, and two video super-resolution diffusion models to generate videos of $512 \\times 896$ resolution at 8 frames per second.", + "bbox": [ + 259, + 617, + 743, + 797 + ], + "page_idx": 0 + }, + { + "type": "page_footnote", + "text": "*Work partially done during an internship at Google.", + "bbox": [ + 232, + 825, + 586, + 839 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "1 Introduction", + "text_level": 1, + "bbox": [ + 217, + 143, + 375, + 161 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Modeling the complex and dynamic visual world with a high level of detail is the ultimate goal of visual generative models. To accomplish this goal, we need neural architectures that can efficiently process high dimensional inputs, capture long range dependencies and have powerful scaling and parallelization properties.", + "bbox": [ + 212, + 174, + 785, + 234 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "The transformer [74] is a neural network architecture designed with such desirable properties. This has encouraged the research community to increasingly favor it over domain-specific architectures in diverse fields such as language [28, 56, 58, 59], audio [1], speech [57], vision [19, 32], and robotics [6, 8, 91]. Such a trend towards unification allows researchers to share and build upon advances in traditionally disparate domains. Thus, leading to a virtuous cycle of innovation and improvement in model design favoring transformers.", + "bbox": [ + 212, + 236, + 785, + 340 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Generative modelling of images and videos has been however a notable exception to this trend. Diffusion models [68, 70], which have emerged as a leading paradigm for generative modelling of images [17, 36] and videos [38], predominantly rely on the U-Net architecture [36, 63] as backbone, consisting of a series of convolutional [47] and self-attention [74] layers. This preference stems from the fact that the memory demands of full attention mechanisms in transformers scales quadratically with input sequence length. Such scaling leads to prohibitively high costs when processing high-dimensional signals like video with transformers.", + "bbox": [ + 212, + 342, + 785, + 474 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "The computational requirements of diffusion models can be reduced by operating in a lower-dimensional latent space derived from an autoencoder [21,73,76], giving rise to latent diffusion models (LDMs) [62]. A critical design choice in this context is the type of latent space employed: spatial compression (per frame latents) versus spatiotemporal compression. Spatial compression is often preferred because it enables leveraging pre-trained image autoencoders and LDMs, which are trained on large paired image-text datasets. However, this choice increases network complexity and limits the use of transformers as backbones, especially in generating high-resolution videos due to memory constraints. On the other hand, while spatiotemporal compression can mitigate these issues, it precludes the use of paired image-text datasets, which are much larger and diverse than their video counterparts. Hence, to leverage all sources of visual data and unlock scaling via transformers, it is critical to have a unified compressed representation of images and videos.", + "bbox": [ + 212, + 478, + 785, + 686 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "In this work, we present Window Attention Latent Transformer (W.A.L.T): a novel transformer-based method for latent video diffusion models (LVDMs). Our method consists of two stages. First, an autoencoder maps both videos and images into a unified, lower-dimensional latent space. This design choice enables training a single generative model jointly on image and video datasets and significantly reduces the computational burden for generating high resolution videos. Subsequently, we propose a new design of transformer blocks for latent video diffusion modeling composed of self-attention layers that alternate between non-overlapping, window-restricted spatial and spatiotemporal attention. This design offers two primary benefits: Firstly, the use of local window attention", + "bbox": [ + 212, + 688, + 785, + 839 + ], + "page_idx": 1 + }, + { + "type": "page_number", + "text": "2", + "bbox": [ + 217, + 114, + 228, + 126 + ], + "page_idx": 1 + }, + { + "type": "header", + "text": "A. Gupta et al.", + "bbox": [ + 271, + 114, + 375, + 128 + ], + "page_idx": 1 + }, + { + "type": "image", + "img_path": "images/bfe32dabb32c1ccdeba7d70de9370b595c60b91393d21684e6d5d5b0109dc573.jpg", + "image_caption": [ + "Fig. 2: W.A.L.T. We encode images and videos into a shared latent space. The transformer backbone processes these latents with blocks having two layers of window-restricted attention: spatial layers capture spatial relations in both images and video, while spatiotemporal layers model temporal dynamics in videos and pass through images via identity attention mask. Text conditioning is done via spatial cross-attention." + ], + "image_footnote": [], + "bbox": [ + 215, + 143, + 781, + 276 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "significantly lowers computational demands. Secondly, it facilitates joint training, where the spatial layers independently process images and video frames, while the spatiotemporal layers are dedicated to modeling the temporal relationships in videos.", + "bbox": [ + 212, + 375, + 787, + 433 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "While conceptually simple, our method provides the first empirical evidence of transformers' superior generation quality and parameter efficiency in latent video diffusion on public benchmarks. Specifically, we report state-of-the-art results on class-conditional video generation (UCF-101 [71]), frame prediction (Kinetics-600 [10]) and class conditional image generation (ImageNet [16]) without using classifier free guidance. Finally, to showcase the scalability and efficiency of our method we also demonstrate results on the challenging task of photorealistic text-to-video generation. We train a cascade of three models consisting of a base latent video diffusion model, and two video super-resolution diffusion models to generate videos of $512 \\times 896$ resolution at 8 frames per second and report state-of-the-art zero-shot FVD score on the UCF-101 benchmark.", + "bbox": [ + 212, + 434, + 787, + 599 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Our contributions can be summarized as follows,", + "bbox": [ + 238, + 601, + 594, + 616 + ], + "page_idx": 2 + }, + { + "type": "list", + "sub_type": "text", + "list_items": [ + "- We propose W.A.L.T, a framework for transformer-based latent diffusion in a unified image and video latent space.", + "- We validate W.A.L.T by performing extensive ablations and achieving state-of-the-art video generation on frame prediction (K600), class-conditional video (UCF-101) and image (ImageNet) generation.", + "- We demonstrate the scaling properties of W.A.L.T and achieve state-of-the-art results in zero-shot UCF-101 text-to-video benchmark." + ], + "bbox": [ + 223, + 625, + 784, + 729 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "2 Related Work", + "text_level": 1, + "bbox": [ + 215, + 763, + 387, + 779 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Video Diffusion Models. Diffusion models have shown impressive results in image [36,40,53,62,68,69] and video generation [5,25,31,35,38,67]. Video diffusion models can be categorized into pixel-space [35,38,67] and latent-space [5,25,33,84]", + "bbox": [ + 212, + 794, + 785, + 842 + ], + "page_idx": 2 + }, + { + "type": "header", + "text": "Photorealistic Video Generation with Diffusion Models", + "bbox": [ + 364, + 114, + 732, + 127 + ], + "page_idx": 2 + }, + { + "type": "page_number", + "text": "3", + "bbox": [ + 774, + 116, + 784, + 126 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "approaches, the later bringing important efficiency advantages when modeling videos. Ho et al. [38] demonstrated that the quality of text conditioned video generation can be significantly improved by jointly training on image and video data. Similarly, to leverage image datasets, latent video diffusion models inflate a pre-trained image model, typically a U-Net [63], into a video model by adding temporal layers, and initializing them as the identity function [5,35,67]. Although computationally efficient, this approach couples the design of video and image models, and precludes spatiotemporal compression. In this work, we operate on a unified latent space for images and videos, allowing us to leverage large scale image and video datasets while enjoying computational efficiency gains from spatiotemporal compression of videos.", + "bbox": [ + 212, + 146, + 787, + 313 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Transformers for Generative Modeling. Multiple classes of generative models have utilized Transformers [74] as backbone, such as, Generative adversarial networks [43, 48, 87], autoregressive [11, 12, 21, 22, 29, 60, 75, 78, 79, 81, 82] and diffusion [2, 23, 42, 51, 54, 89] models. Inspired by the success of autoregressive pretraining of large language models [56, 58, 59], Ramesh et al. [60] trained a text-to-image generation model by predicting the next visual token obtained from an image tokenizer. Subsequently, this approach was applied to multiple applications including class-conditional image generation [21, 80], text-to-image [18, 60, 77, 81] or image-to-image translation [22, 78]. Similarly, for video generation, transformer-based models were proposed to predict next tokens using 3D extensions of VQGAN [24, 39, 79, 82] or using per frame image latents [29]. Autoregressive sampling of videos is typically impractical given the very long sequences involved. To alleviate this issue, non-autoregressive sampling [11, 12], i.e. parallel token prediction, has been adopted as a more efficient solution for transformer-based video generation [29, 75, 82]. Recently, the community has started adopting transformers as the denoising backbone for diffusion models in place of U-Net [13, 40, 51, 54, 89]. To the best of our knowledge, our work is the first successful empirical demonstration (§ 5.1) of a transformer-based backbone for jointly training image and video latent diffusion models.", + "bbox": [ + 212, + 325, + 789, + 613 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "3 Background", + "text_level": 1, + "bbox": [ + 215, + 635, + 370, + 652 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Diffusion formulation. Diffusion models [36, 68, 70] are a class of generative models which learn to generate data by iteratively denoising samples drawn from a noise distribution. Gaussian diffusion models assume a forward noising process which gradually applies noise $(\\epsilon)$ to real data $(x_0 \\sim p_{\\mathrm{data}})$ . Concretely,", + "bbox": [ + 212, + 665, + 787, + 727 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\n\\boldsymbol {x} _ {t} = \\sqrt {\\gamma (t)} \\boldsymbol {x} _ {0} + \\sqrt {1 - \\gamma (t)} \\epsilon , \\tag {1}\n$$\n", + "text_format": "latex", + "bbox": [ + 392, + 737, + 785, + 753 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "where $\\epsilon \\sim \\mathcal{N}(\\mathbf{0},\\mathbf{I}), t \\in [0,1]$ , and $\\gamma(t)$ is a monotonically decreasing function (noise schedule) from 1 to 0. Diffusion models are trained to learn the reverse process that inverts the forward corruptions:", + "bbox": [ + 212, + 763, + 787, + 810 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\n\\mathbb {E} _ {\\boldsymbol {x} \\sim p _ {\\text {d a t a}}, t \\sim \\mathcal {U} (0, 1), \\boldsymbol {\\epsilon} \\sim \\mathcal {N} (\\boldsymbol {0}, \\boldsymbol {I})} \\left[ \\| \\boldsymbol {y} - f _ {\\theta} (\\boldsymbol {x} _ {t}; \\boldsymbol {c}, t) \\| ^ {2} \\right], \\tag {2}\n$$\n", + "text_format": "latex", + "bbox": [ + 336, + 819, + 785, + 843 + ], + "page_idx": 3 + }, + { + "type": "page_number", + "text": "4", + "bbox": [ + 217, + 114, + 228, + 126 + ], + "page_idx": 3 + }, + { + "type": "header", + "text": "A. Gupta et al.", + "bbox": [ + 271, + 114, + 375, + 128 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "where $f_{\\theta}$ is the denoiser model parameterized by a neural network, $\\pmb{c}$ is conditioning information e.g., class labels or text prompts, and the target $\\pmb{y}$ can be random noise $\\pmb{\\epsilon}$ , denoised input $\\pmb{x_0}$ or $\\pmb{v} = \\sqrt{1 - \\gamma(t)} \\pmb{\\epsilon} - \\sqrt{\\gamma(t)} \\pmb{x_0}$ . Following [35, 65], we use $\\pmb{v}$ -prediction in all our experiments.", + "bbox": [ + 212, + 146, + 787, + 205 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Latent diffusion models (LDMs). Processing high-resolution images and videos using raw pixels requires considerable computational resources. To address this, LDMs operate on the low dimensional latent space of a VQ-VAE [21,73]. VQ-VAE consists of an encoder $E(\\pmb{x})$ that encodes an input video $\\pmb{x} \\in \\mathbb{R}^{T \\times H \\times W \\times 3}$ into a latent representation $\\pmb{z} \\in \\mathbb{R}^{t \\times h \\times w \\times c}$ . The encoder downsamples the video by a factor of $f_{s} = H / h = W / w$ and $f_{t} = T / t$ , where $T = t = 1$ corresponds to using an image autoencoder. An important distinction from the original VQ-VAE is the absence of a codebook of quantized embeddings as diffusion models can operate on continuous latent spaces. A decoder $D$ is trained to predict a reconstruction of the video, $\\hat{\\pmb{x}}$ , from $\\pmb{z}$ . Following VQ-GAN [21], reconstruction quality can be further improved by adding adversarial [27] and perceptual losses [44,88].", + "bbox": [ + 212, + 207, + 789, + 375 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "4 W.A.L.T", + "text_level": 1, + "bbox": [ + 214, + 393, + 344, + 409 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "4.1 Learning Visual Tokens", + "text_level": 1, + "bbox": [ + 214, + 424, + 455, + 439 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "A key design decision in video generative modeling is the choice of latent space representation. Ideally, we want a shared and unified compressed visual representation that can be used for generative modeling of both images and videos [75,83]. The unified representation is important because joint image-video learning is preferable due to a scarcity of labeled video data [35], such as text-video pairs. Concretely, given a video sequence $\\pmb{x} \\in \\mathbb{R}^{(1 + T) \\times H \\times W \\times C}$ , we aim to learn a low-dimensional representation $\\pmb{z} \\in \\mathbb{R}^{(1 + t) \\times h \\times w \\times c}$ that performs spatial-temporal compression by a factor of $f_{s} = H / h = W / w$ in space and a factor of $f_{t} = T / t$ in time. To enable a unified representation for both videos and static images, the first frame is always encoded independently from the rest of the video. This allows static images $\\pmb{x} \\in \\mathbb{R}^{1 \\times H \\times W \\times C}$ to be treated as videos with a single frame, i.e. $\\pmb{z} \\in \\mathbb{R}^{1 \\times h \\times w \\times c}$ .", + "bbox": [ + 212, + 446, + 789, + 628 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "We instantiate this design with the causal 3D CNN encoder-decoder architecture of the MAGVIT-v2 tokenizer [83]. Typically the encoder-decoder consists of regular 3D convolution layers which cannot process the first frame independently [24,82]. This limitation stems from the fact that a regular convolutional kernel of size $(k_{t},k_{h},k_{w})$ will operate on $\\left\\lfloor \\frac{k_t - 1}{2}\\right\\rfloor$ frames before and $\\left\\lfloor \\frac{k_t}{2}\\right\\rfloor$ frames after the input frames. Causal 3D convolution layers solve this issue as the convolutional kernel operates on only the past $k_{t} - 1$ frames. This ensures that the output for each frame is influenced solely by the preceding frames, enabling the model to tokenize the first frame independently.", + "bbox": [ + 212, + 628, + 787, + 763 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "After this stage, the input to our model is a batch of latent tensors $\\mathbf{z} \\in \\mathbb{R}^{(1 + t) \\times h \\times w \\times c}$ representing a single video or a stack of $1 + t$ independent images (Fig. 2). Different from [83], our latent representation is real-valued and quantization-free. In the section below we describe how our model jointly processes a mixed batch of images and videos.", + "bbox": [ + 212, + 765, + 787, + 839 + ], + "page_idx": 4 + }, + { + "type": "header", + "text": "Photorealistic Video Generation with Diffusion Models", + "bbox": [ + 364, + 113, + 732, + 127 + ], + "page_idx": 4 + }, + { + "type": "page_number", + "text": "5", + "bbox": [ + 774, + 116, + 784, + 126 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "4.2 Learning to Generate Images and Videos", + "text_level": 1, + "bbox": [ + 215, + 146, + 602, + 161 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Patchify. Following the original ViT [19], we \"patchify\" each latent frame independently by converting it into a sequence of non-overlapping $h_p \\times w_p$ patches where $h_p = h/p$ , $w_p = w/p$ and $p$ is the patch size. We use learnable positional embeddings [74], which are the sum of space and time positional embeddings. Position embeddings are added to the linear projections [19] of the patches. Note that for images, we simply add the temporal position embedding corresponding to the first latent frame.", + "bbox": [ + 212, + 176, + 787, + 281 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Window attention. Transformer models composed entirely of global self-attention modules incur significant compute and memory costs, especially for video tasks. For efficiency and for processing images and videos jointly we compute self-attention in windows [29, 74], based on two types of non-overlapping configurations: spatial (S) and spatiotemporal (ST), cf. Fig. 2. Spatial Window (SW) attention is restricted to all the tokens within a latent frame of size $1 \\times h_p \\times w_p$ (the first dimension is time). SW models the spatial relations in images and videos. Spatiotemporal Window (STW) attention is restricted within a 3D window of size $(1 + t) \\times h_p' \\times h_w'$ , modeling the temporal relationships among video latent frames. For images, we simply use an identity attention mask ensuring that the value embeddings corresponding to the image frame latents are passed through the layer as is. Finally, in addition to absolute position embeddings we also use relative position embeddings [50].", + "bbox": [ + 212, + 282, + 787, + 479 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Our design, while conceptually straightforward, achieves computational efficiency and enables joint training on image and video datasets. In contrast to methods based on frame-level autoencoders [5, 25, 29], our approach does not suffer from flickering artifacts, which often result from encoding and decoding video frames independently. However, similar to Blattmann et al. [5], we can also potentially leverage pre-trained image LDMs with transformer backbones by simply interleaving STW layers.", + "bbox": [ + 212, + 479, + 789, + 585 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "4.3 Conditional Generation", + "text_level": 1, + "bbox": [ + 215, + 612, + 457, + 626 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "To enable controllable video generation, in addition to conditioning on timestep $t$ , diffusion models are often conditioned on additional conditional information $\\mathbf{c}$ such as class labels, natural language, past frames or low resolution videos. In our transformer backbone, we incorporate three types of conditioning mechanisms as described in what follows:", + "bbox": [ + 212, + 642, + 787, + 717 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Cross-attention. In addition to self-attention layers in our window transformer blocks, we add a cross-attention layer for text conditioned generation. When training models on just videos, the cross-attention layer employs the same window-restricted attention as the self-attention layer, meaning S/ST blocks will have SW/STW cross-attention layers (Fig. 2). However, for joint training, we only use SW cross-attention layers. For cross-attention we concatenate the input signal (query) with the conditioning signal (key, value) as our early experiments showed this improves performance.", + "bbox": [ + 212, + 719, + 787, + 840 + ], + "page_idx": 5 + }, + { + "type": "page_number", + "text": "6", + "bbox": [ + 217, + 114, + 228, + 126 + ], + "page_idx": 5 + }, + { + "type": "header", + "text": "A. Gupta et al.", + "bbox": [ + 271, + 114, + 375, + 128 + ], + "page_idx": 5 + }, + { + "type": "table", + "img_path": "images/897ebf563cab4cc4cfeace5b178aca3aaa8c22e83939c680504a80f44e942410.jpg", + "table_caption": [], + "table_footnote": [], + "table_body": "
MethodK600 FVD↓UCF FVD↓params.steps
TrIVD-GAN-FP [52]25.7±0.7--1
Video Diffusion [38]16.2±0.3-1.1B256
RIN [42]10.8-411M1000
TATS [24]-332±18321M1024
Phenaki [75]36.4±0.2-227M48
MAGVIT [82]9.9±0.376±2306M12
MAGVITv2 [83]4.3±0.158±2307M24
W.A.L.T-L Ours3.3±0.046±2313M50
W.A.L.T-XL Ours-36±2460M50
", + "bbox": [ + 302, + 143, + 696, + 291 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Table 1: Video generation evaluation on frame prediction on Kinetics-600 and class-conditional generation on UCF-101.", + "bbox": [ + 212, + 299, + 784, + 327 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "AdaLN-LoRA. Adaptive normalization layers are an important component in a broad range of generative and visual synthesis models [17, 20, 45, 53-55]. A simple way to incorporate adaptive layer normalization is to include for each layer $i$ , an MLP layer to regress a vector of conditioning parameters $A^i = \\mathbf{MLP}(c + t)$ , where $A^i = \\text{concat}(\\gamma_1, \\gamma_2, \\beta_1, \\beta_2, \\alpha_1, \\alpha_2)$ , $A^i \\in \\mathbb{R}^{6 \\times d_{\\text{model}}}$ , and $c \\in \\mathbb{R}^{d_{\\text{model}}}, t \\in \\mathbb{R}^{d_{\\text{model}}}$ are the condition and timestep embeddings. In the transformer block, $\\gamma$ and $\\beta$ scale and shift the inputs of the multi-head attention and MLP layers, respectively, while $\\alpha$ scales the output of both the multi-head attention and MLP layers. The parameter count of these additional MLP layers scales linearly with the number of layers and quadratically with the model's dimensional size (num_blocks $\\times$ $d_{\\text{model}} \\times 6 \\times d_{\\text{model}}$ ). For instance, in a ViT-g model with 1B parameters, the MLP layers contribute an additional 475M parameters. Inspired by [41], we propose a simple solution dubbed AdaLN-LoRA, to reduce the model parameters. For each layer, we regress conditioning parameters as", + "bbox": [ + 212, + 376, + 787, + 589 + ], + "page_idx": 6 + }, + { + "type": "equation", + "text": "\n$$\nA ^ {1} = \\operatorname {M L P} (\\boldsymbol {c} + \\boldsymbol {t}), \\quad A ^ {i} = A ^ {1} + W _ {b} ^ {i} W _ {a} ^ {i} (\\boldsymbol {c} + \\boldsymbol {t}) \\quad \\forall i \\neq 1, \\tag {3}\n$$\n", + "text_format": "latex", + "bbox": [ + 292, + 604, + 785, + 622 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "where $W_{b}^{i}\\in \\mathbb{R}^{d_{\\mathrm{model}}\\times r}$ , $W_{a}^{i}\\in \\mathbb{R}^{r\\times (6\\times d_{\\mathrm{model}})}$ . This reduces the number of trainable model parameters significantly when $r\\ll d_{\\mathrm{model}}$ . For example, a ViT-g model with $r = 2$ reduces the MLP parameters from 475M to 12M.", + "bbox": [ + 212, + 638, + 784, + 686 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Self-conditioning. In addition to being conditioned on external inputs, iterative generative algorithms can also be conditioned on their own previously generated samples during inference [3, 14, 66]. Specifically, Chen et al. [14] modify the training process for diffusion models, such that with some probability $p_{\\mathrm{sc}}$ the model first generates a sample $\\tilde{\\pmb{z}}_0 = f_\\theta(\\pmb{z}_t; \\pmb{0}, \\pmb{c}, t)$ and then refines this estimate using another forward pass conditioned on this initial sample: $f_\\theta(\\pmb{z}_t; \\text{stopgrad}(\\tilde{\\pmb{z}}_0), \\pmb{c}, t)$ . With probability $1 - p_{\\mathrm{sc}}$ , only a single forward pass is done. We concatenate the model estimate with the input along the channel dimension and found this simple technique to work well when used in conjunction with $\\pmb{v}$ -prediction.", + "bbox": [ + 212, + 688, + 787, + 840 + ], + "page_idx": 6 + }, + { + "type": "header", + "text": "Photorealistic Video Generation with Diffusion Models", + "bbox": [ + 364, + 114, + 732, + 127 + ], + "page_idx": 6 + }, + { + "type": "page_number", + "text": "7", + "bbox": [ + 774, + 116, + 784, + 126 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "4.4 Autoregressive Generation", + "text_level": 1, + "bbox": [ + 215, + 146, + 482, + 161 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "For generating long videos via autoregressive prediction we also train our model jointly on the task of frame prediction. This is achieved by conditioning the model on past frames with a probability of $p_{\\mathrm{fp}}$ during training. Specifically, the model is conditioned using $c_{\\mathrm{fp}} = \\text{concat}(m_{\\mathrm{fp}} \\circ z_t, m_{\\mathrm{fp}})$ , where $m_{\\mathrm{fp}}$ is a binary mask. The binary mask indicates the number of past frames used for conditioning. We condition on either 1 latent frame (image to video generation) or 2 latent frames (video prediction). This conditioning is integrated into the model through concatenation along the channel dimension of the noisy latent input. During inference, we use standard classifier-free guidance with $c_{\\mathrm{fp}}$ as the conditioning signal.", + "bbox": [ + 212, + 172, + 787, + 324 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "4.5 Video Super Resolution", + "text_level": 1, + "bbox": [ + 215, + 345, + 459, + 361 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Generating high-resolution videos with a single model is computationally prohibitive. Following [37], we use a cascaded approach with three models operating at increasing resolutions. Our base model generates videos at $128 \\times 128$ resolution which are subsequently upsampled twice via two super resolution stages. We first spatially upscale the low resolution input $z^{\\mathrm{lr}}$ (video or image) using a depth-to-space convolution operation. Note that, unlike training where ground truth low-resolution inputs are available, inference relies on latents produced by preceding stages (cf. teaching-forcing). To reduce this discrepancy and improve the robustness of the super-resolution stages in handling artifacts generated by lower resolution stages, we use noise conditioning augmentation [37]. Concretely, noise is added in accordance with $\\gamma(t)$ , by sampling a noise level as $t_{\\mathrm{sr}} \\sim \\mathcal{U}(0, t_{\\max \\text{noise}})$ and is provided as input to our AdaLN-LoRA layers.", + "bbox": [ + 212, + 372, + 787, + 551 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Aspect-ratio finetuning. To simplify training and leverage broad data sources with different aspect ratios, we train our base stage using a square aspect ratio. We fine-tune the base stage on a subset of data to generate videos with a 9:16 aspect ratio by interpolating position embeddings.", + "bbox": [ + 212, + 553, + 787, + 614 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "5 Experiments", + "text_level": 1, + "bbox": [ + 215, + 637, + 375, + 655 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "In this section, we evaluate our method on multiple tasks: class-conditional image and video generation, frame prediction and text conditioned video generation and perform extensive ablation studies of different design choices. For qualitative results, see Fig. 1, Fig. 3, Fig. 4 and videos on our project website. See appendix for additional details.", + "bbox": [ + 212, + 670, + 787, + 744 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "5.1 Visual Generation", + "text_level": 1, + "bbox": [ + 215, + 768, + 413, + 782 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Video generation. We consider two standard video benchmarks, UCF-101 [71] for class-conditional generation and Kinetics-600 [10] for video prediction with 5 conditioning frames. We use FVD [72] as our primary evaluation metric.", + "bbox": [ + 212, + 794, + 787, + 840 + ], + "page_idx": 7 + }, + { + "type": "page_number", + "text": "8", + "bbox": [ + 217, + 114, + 228, + 126 + ], + "page_idx": 7 + }, + { + "type": "header", + "text": "A. Gupta et al.", + "bbox": [ + 271, + 114, + 375, + 128 + ], + "page_idx": 7 + }, + { + "type": "table", + "img_path": "images/eace865a00a02acbec11eaeae07259da3d898262a25d303499ae451eea99efca.jpg", + "table_caption": [], + "table_footnote": [], + "table_body": "
MethodCost (Iter×BS)FID↓IS↑params.steps
BigGAN-deep [7]-6.95171.4160M1
LDM-4 [62]178k×120010.56103.5400M250
DiT-XL/2 [54]7000k×2569.62121.5675M250
ADM [17]-7.49127.5608M2000
MDT [23]6500k×2566.23143.0676M250
MaskDiT [89]1200k×10245.69178.0736M40
RIN [42]600k×10243.42182.0410M1000
simple diffusion [40]500K×20482.77211.82B512
VDM++ [46]-2.40225.32B512
W.A.L.T-L Ours437k×10242.56215.1460M50
", + "bbox": [ + 297, + 143, + 696, + 284 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "Table 2: Class-conditional image generation on ImageNet ${256} \\times {256}$ . We adopt the evaluation protocol and implementation of ADM [17] and report results without classifier free guidance.", + "bbox": [ + 212, + 295, + 787, + 338 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "Across both datasets, W.A.L.T significantly outperforms all prior works (Tab. 1). Compared to prior video diffusion models, we achieve state-of-the-art performance with less model parameters, and require 50 DDIM [69] inference steps.", + "bbox": [ + 212, + 380, + 785, + 425 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "Image generation. To verify the modeling capabilities of $W.A.L.T$ on the image domain, we train a version of $W.A.L.T$ for the standard ImageNet class-conditional setting. For evaluation, we follow ADM [17] and report the FID [34] and Inception [64] scores calculated on 50K samples generated in 50 DDIM steps. We compare (Table 2) $W.A.L.T$ with state-of-the-art image generation methods for $256 \\times 256$ resolution. Our model outperforms prior works without requiring specialized schedules, convolution inductive bias, improved diffusion losses, and classifier free guidance. Although VDM++ [46] has slightly better FID score, the model has significantly more parameters (2B).", + "bbox": [ + 212, + 426, + 787, + 561 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "5.2 Ablation Studies", + "text_level": 1, + "bbox": [ + 214, + 585, + 401, + 599 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "We ablate $W.A.L.T$ to understand the contribution of various design decisions with the default settings: model L, patch size 1, $1 \\times 16 \\times 16$ spatial window, $5 \\times 8 \\times 8$ spatiotemporal window, $p_{\\mathrm{sc}} = 0.9$ , $c = 8$ and $r = 2$ .", + "bbox": [ + 212, + 612, + 785, + 657 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "Patch size. In various computer vision tasks utilizing ViT [19]-based models, a smaller patch size $p$ has been shown to consistently enhance performance [9, 19, 30, 86]. Similarly, our findings also indicate that a reduced patch size improves performance (Table 3a).", + "bbox": [ + 212, + 657, + 785, + 718 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "Window attention. We compare three different STW window configurations with full self-attention (Table 3b). We find that local self-attention can achieve competitive (or better) performance while being significantly faster (up to $2 \\times$ ) and requiring less accelerator memory.", + "bbox": [ + 212, + 719, + 785, + 779 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "Self-conditioning. In Table 3c we study the influence of varying the self-conditioning rate $p_{\\mathrm{sc}}$ on generation quality. We notice a clear trend: increasing the self-conditioning rate from 0.0 (no self-conditioning) to 0.9 improves the FVD score substantially (44%).", + "bbox": [ + 212, + 779, + 785, + 840 + ], + "page_idx": 8 + }, + { + "type": "header", + "text": "Photorealistic Video Generation with Diffusion Models", + "bbox": [ + 364, + 114, + 730, + 127 + ], + "page_idx": 8 + }, + { + "type": "page_number", + "text": "9", + "bbox": [ + 774, + 116, + 784, + 126 + ], + "page_idx": 8 + }, + { + "type": "table", + "img_path": "images/123e19d4947ed99c95dce83ce3a69df7c66c0f1c868ff20604e49086bc265d00.jpg", + "table_caption": [], + "table_footnote": [], + "table_body": "
patch size pFVD↓IS↑
160.787.2
2134.482.2
4461.863.9
", + "bbox": [ + 287, + 162, + 464, + 227 + ], + "page_idx": 9 + }, + { + "type": "table", + "img_path": "images/4f0a18f9f1845217dc8fd74e37d9ef76c7fde2f50520ee967cfad376f6cfbd6f.jpg", + "table_caption": [], + "table_footnote": [], + "table_body": "
st windowFVD↓IS↑sps
5 × 4 × 456.987.32.24
5 × 8 × 859.687.42.00
5 × 16 × 1655.387.41.75
full self attn.59.987.81.20
", + "bbox": [ + 519, + 142, + 743, + 227 + ], + "page_idx": 9 + }, + { + "type": "text", + "text": "(a) Patch size. Lower patch size is significantly better.", + "bbox": [ + 250, + 244, + 495, + 268 + ], + "page_idx": 9 + }, + { + "type": "text", + "text": "(b) Spatiotemporal window size. Full self-attention is not essential for good performance. sps is steps per sec.", + "bbox": [ + 508, + 244, + 751, + 277 + ], + "page_idx": 9 + }, + { + "type": "table", + "img_path": "images/97c77a8e4eae23d72cf0dcc78b48450dfe6f10c1a2d44db228b8055bed69cb90.jpg", + "table_caption": [], + "table_footnote": [], + "table_body": "
pscFVD↓IS↑
0.0109.982.6
0.376.086.5
0.660.086.8
0.961.487.1
", + "bbox": [ + 325, + 291, + 444, + 371 + ], + "page_idx": 9 + }, + { + "type": "table", + "img_path": "images/842a6201c4a43c994dfa2fc9b5e198d2eab341ad9e5bdfc07362fa7389129410.jpg", + "table_caption": [], + "table_footnote": [], + "table_body": "
rFVD↓IS↑params
260.787.2313 M
456.687.3314 M
1655.588.0316 M
6454.487.9324 M
25652.588.5357 M
", + "bbox": [ + 532, + 291, + 712, + 383 + ], + "page_idx": 9 + }, + { + "type": "text", + "text": "(c) Self-conditioning. Higher $p_{\\mathrm{sc}}$ is bet- (d) AdaLN-LoRA. Bigger $r$ is better.", + "bbox": [ + 258, + 387, + 732, + 409 + ], + "page_idx": 9 + }, + { + "type": "table", + "img_path": "images/3d33dc9b09d2a9541b97bb0a7a7e8160696b7e714d831329914c827f1687e825.jpg", + "table_caption": [], + "table_footnote": [], + "table_body": "
FVD↓IS↑
w/o qk norm [15]59.086.8
w/o latent norm67.987.1
w/o zero snr [49]91.084.2
full method60.787.2
", + "bbox": [ + 267, + 422, + 483, + 508 + ], + "page_idx": 9 + }, + { + "type": "text", + "text": "(e) Other improvements. See text for details.", + "bbox": [ + 251, + 511, + 493, + 532 + ], + "page_idx": 9 + }, + { + "type": "table", + "img_path": "images/23b8fd03d08544e187b7772eaa468c02bb881370b5f19a0bd8da85256389f12c.jpg", + "table_caption": [], + "table_footnote": [], + "table_body": "
crFVD↓FVD↓IS↑
437.786.484.9
817.175.486.3
168.267.086.0
323.583.482.9
", + "bbox": [ + 545, + 428, + 717, + 507 + ], + "page_idx": 9 + }, + { + "type": "text", + "text": "(f) Latent dimension $c$ . Higher $c$ is better.", + "bbox": [ + 508, + 511, + 751, + 532 + ], + "page_idx": 9 + }, + { + "type": "text", + "text": "Table 3: Ablation experiments on UCF-101 [71]. We compare FVD and inception scores to ablate important design decisions with the default setting: L model, $1 \\times 16 \\times 16$ spatial window, $5 \\times 8 \\times 8$ saptiotemporal (st) window, $p_{\\mathrm{sc}} = 0.9$ , $c = 8$ and $r = 2$ .", + "bbox": [ + 212, + 544, + 787, + 589 + ], + "page_idx": 9 + }, + { + "type": "text", + "text": "AdaLN-LoRA. An important design decision in diffusion models is the conditioning mechanism. We investigate the effect of increasing the bottleneck dimension $r$ in our proposed AdaLN-LoRA layers (Table 3d). This hyperparameter provides a flexible way to trade off between number of model parameters and generation performance. As shown in Table 3d, increasing $r$ improves performance but also increases model parameters. This highlights an important model design question: given a fixed parameter budget, how should we allocate parameters - either by using separate AdaLN layers, or by increasing base model parameters while using shared AdaLN-LoRA layers? We explore this in Table 4 by comparing two model configurations: W.A.L.T-L with separate AdaLN layers and W.A.L.T-XL with AdaLN-LoRA and $r = 2$ . While both configurations yield similar FVD and Inception scores, W.A.L.T-XL achieves a lower final loss value, suggesting the advantage of allocating more parameters to the base model and choosing an appropriate $r$ value within accelerator memory limits.", + "bbox": [ + 212, + 628, + 787, + 840 + ], + "page_idx": 9 + }, + { + "type": "page_number", + "text": "10", + "bbox": [ + 217, + 114, + 235, + 126 + ], + "page_idx": 9 + }, + { + "type": "header", + "text": "A. Gupta et al.", + "bbox": [ + 271, + 114, + 375, + 128 + ], + "page_idx": 9 + }, + { + "type": "table", + "img_path": "images/feb7c5fc03effc9148f63161e66e38faba4950e4bb65d2ac026c963f2f435101.jpg", + "table_caption": [], + "table_footnote": [], + "table_body": "
ModelAdaLNFVD↓IS↑params.final loss
Lseparate34.690.2458M0.274
XLLoRA-236.789.4460M0.268
", + "bbox": [ + 312, + 143, + 687, + 204 + ], + "page_idx": 10 + }, + { + "type": "text", + "text": "Table 4: Parameter matched comparison between AdaLN-LoRA and per layer adaln layers. See text for details.", + "bbox": [ + 212, + 222, + 784, + 250 + ], + "page_idx": 10 + }, + { + "type": "text", + "text": "Noise schedule. Common latent diffusion noise schedules [62] typically do not ensure a zero signal-to-noise ratio (SNR) at the final timestep, i.e., at $t = 1$ , $\\gamma(t) > 0$ . This leads to a mismatch between training and inference phases. During inference, models are expected to start from purely Gaussian noise, whereas during training, at $t = 1$ , a small amount of signal information remains accessible to the model. This is especially harmful for video generation as videos have high temporal redundancy. Even minimal information leakage at $t = 1$ can reveal substantial information to the model. Addressing this mismatch by enforcing a zero terminal SNR [49] significantly improves performance (Table 3e). Note that this approach was originally proposed to fix over-exposure problems in image generation, but we find it effective for video generation as well.", + "bbox": [ + 212, + 294, + 787, + 460 + ], + "page_idx": 10 + }, + { + "type": "text", + "text": "Autoencoder. Finally, we investigate one critical but often overlooked hyperparameter in the first stage of our model: the channel dimension $c$ of the autoencoder latent $z$ . As shown in Table 3f, increasing $c$ significantly improves the reconstruction quality (lower rFVD) while keeping the same spatial $f_{s}$ and temporal compression $f_{t}$ ratios. Empirically, we found that both lower and higher values of $c$ lead to poor FVD scores in generation, with a sweet spot of $c = 8$ working well across most datasets and tasks we evaluated. We also normalize the latents before processing them via transformer which further improves performance.", + "bbox": [ + 212, + 462, + 787, + 583 + ], + "page_idx": 10 + }, + { + "type": "text", + "text": "In our transformer models, we use query-key normalization [15] as it helps stabilize training for larger models. Finally, we note that some of our default settings are not optimal, as indicated by ablation studies. These defaults were chosen early on for their robustness across datasets, though further tuning may improve performance.", + "bbox": [ + 212, + 584, + 787, + 659 + ], + "page_idx": 10 + }, + { + "type": "text", + "text": "5.3 Text-to-video", + "text_level": 1, + "bbox": [ + 214, + 686, + 375, + 702 + ], + "page_idx": 10 + }, + { + "type": "text", + "text": "We train $W.A.L.T$ for text-to-video jointly on text-image and text-video pairs (Sec. 4.2). We used a dataset of $\\sim 970\\mathrm{M}$ text-image pairs and $\\sim 89\\mathrm{M}$ text-video pairs from the public internet and internal sources. We train our base model at resolution $17 \\times 128 \\times 128$ (3B parameters), and two $2 \\times$ cascaded super-resolution models for $17 \\times 128 \\times 224 \\rightarrow 17 \\times 256 \\times 448$ (L, 1.3B, $p = 2$ ) and $17 \\times 256 \\times 448 \\rightarrow 17 \\times 512 \\times 896$ (L, 419M, $p = 2$ ) respectively. We fine-tune the base stage for the $9:16$ aspect ratio to generate videos at resolution $128 \\times 224$ . We use classifier free guidance for all our text-to-video results.", + "bbox": [ + 212, + 718, + 787, + 839 + ], + "page_idx": 10 + }, + { + "type": "header", + "text": "Photorealistic Video Generation with Diffusion Models", + "bbox": [ + 364, + 113, + 730, + 127 + ], + "page_idx": 10 + }, + { + "type": "page_number", + "text": "11", + "bbox": [ + 767, + 114, + 782, + 126 + ], + "page_idx": 10 + }, + { + "type": "image", + "img_path": "images/3fbda4a8bdb8e31bb1dc4a016cfeac307af56b48312cc596c6ad45f79c814283.jpg", + "image_caption": [ + "A polar bear swimming." + ], + "image_footnote": [], + "bbox": [ + 218, + 205, + 782, + 258 + ], + "page_idx": 11 + }, + { + "type": "image", + "img_path": "images/6cc71d1c985cc4361b40012b480bdab728ca2d6e6f7641a44e221db218f9a558.jpg", + "image_caption": [ + "Pouring beer into an empty glass, low angle shot, bar in the background." + ], + "image_footnote": [], + "bbox": [ + 217, + 275, + 782, + 328 + ], + "page_idx": 11 + }, + { + "type": "image", + "img_path": "images/77d16469cfeda54b85c88371d136e0c211b417ad9369ce87102415db1e8605bf.jpg", + "image_caption": [ + "A cat eating food out of a bowl, in the style of Van Gogh." + ], + "image_footnote": [], + "bbox": [ + 217, + 349, + 782, + 402 + ], + "page_idx": 11 + }, + { + "type": "image", + "img_path": "images/661d6094151bd65db1e7b4c6027bf3930220eb15ec10423e7b8150443ea8964b.jpg", + "image_caption": [ + "Pouring chocolate sauce over vanilla ice cream in a cone, studio lighting." + ], + "image_footnote": [], + "bbox": [ + 217, + 420, + 782, + 472 + ], + "page_idx": 11 + }, + { + "type": "image", + "img_path": "images/a3b932a13c27a28ee5f53b633e465054e881c7e085bb01d3cb2ca3f5c7c64dea.jpg", + "image_caption": [ + "A robot ballerina dancing gracefully, highly detailed, studio lighting." + ], + "image_footnote": [], + "bbox": [ + 217, + 494, + 782, + 546 + ], + "page_idx": 11 + }, + { + "type": "image", + "img_path": "images/adaf4da68cd717ac491e1f84afc9a5f418f9589c42e8f56cfbd7f95c38da5be8.jpg", + "image_caption": [ + "A cute panda skateboarding in the sky, over snow covered mountains, with dreamy whimsical atmosphere." + ], + "image_footnote": [], + "bbox": [ + 217, + 568, + 782, + 619 + ], + "page_idx": 11 + }, + { + "type": "image", + "img_path": "images/b05555d5f782fdeb616ea285f818bd0552a812fb53692b4f84e47397cb382fab.jpg", + "image_caption": [ + "An astronaut riding a horse.", + "Fig. 3: Qualitative evaluation. Example videos generated by $W.A.L.T$ from natural language prompts at $512 \\times 896$ resolution over 3.6 seconds duration at 8 frames per second. The $W.A.L.T$ model is able to generate temporally consistent photorealistic videos that align with the textual prompt. There are more results on our project website." + ], + "image_footnote": [], + "bbox": [ + 217, + 643, + 782, + 696 + ], + "page_idx": 11 + }, + { + "type": "page_number", + "text": "12", + "bbox": [ + 217, + 114, + 235, + 126 + ], + "page_idx": 11 + }, + { + "type": "header", + "text": "A. Gupta et al.", + "bbox": [ + 271, + 114, + 375, + 128 + ], + "page_idx": 11 + }, + { + "type": "table", + "img_path": "images/666b1a4e881dcc3d0e8c38d428b292d2dab85efac9f7280befba5d19e662ff85.jpg", + "table_caption": [], + "table_footnote": [], + "table_body": "
MethodIS (↑)FVD (↓)
CogVideo (Chinese) [39]23.6751.3
CogVideo (English) [39]25.3701.6
MagicVideo [90]-699.0
EMU Video* [26]42.7606.2
Video LDM [5]33.5550.6
Make-A-Video [67]33.0367.2
PYoCo [25]47.8355.2
PixelDance* [85]42.1242.8
SVD* [4]-242.0
W.A.L.T (Ours) 419M (video only)26.8598.8
W.A.L.T (Ours) 419M (video + image)31.7344.5
W.A.L.T (Ours) 3B (video + image)35.1258.1
", + "bbox": [ + 326, + 143, + 666, + 321 + ], + "page_idx": 12 + }, + { + "type": "text", + "text": "Table 5: UCF-101 text-to-video generation. Joint training on image and video datasets in conjunction with scaling the model parameters is essential for high quality video generation. * Denotes concurrent work.", + "bbox": [ + 212, + 330, + 787, + 375 + ], + "page_idx": 12 + }, + { + "type": "text", + "text": "Quantitative Evaluation Evaluating text-conditioned video generation scientifically remains a significant challenge, in part due to the absence of standardized training datasets and benchmarks. So far we have focused our experiments and analyses on the standard academic benchmarks, which use the same training data to ensure controlled and fair comparisons. Nevertheless, to compare with prior work on text-to-video, we also report results on the UCF-101 dataset in the zero-shot evaluation protocol in Table 5 [25,39,67]. Also see supplement.", + "bbox": [ + 212, + 416, + 785, + 521 + ], + "page_idx": 12 + }, + { + "type": "text", + "text": "Joint training. A primary strength of our framework is its ability to train simultaneously on both image and video datasets. In Table 5 we ablate the impact of this joint training. Specifically, we trained two versions of W.A.L.T-L (each with 419M params.) models using the default settings specified in § 5.2. We find that joint training leads to a notable improvement across both metrics. Our results align with the findings of Ho et al. [38], who demonstrated the benefits of joint training for pixel-based video diffusion models with U-Net backbones.", + "bbox": [ + 212, + 523, + 787, + 628 + ], + "page_idx": 12 + }, + { + "type": "text", + "text": "Scaling. Transformers are known for their ability to scale effectively in many tasks [6,15,58]. In Table 5 we show the benefits of scaling our transformer model for video diffusion. Scaling our base model size leads to significant improvements on both the metrics. It is important to note, however, that our base model is considerably smaller than leading text-to-video systems. For instance, Ho et al. [35] trained base model of 5.7B parameters. Hence, we believe scaling our models further is an important direction of future work.", + "bbox": [ + 212, + 628, + 787, + 733 + ], + "page_idx": 12 + }, + { + "type": "text", + "text": "Comparison with prior work. In Table 5, we present a system-level comparison of various text-to-video methods. Our results are promising, surpassing all previous work in the FVD metric, and being competitive with concurrent work. In terms of the IS, W.A.L.T outperforms all previous work but PYoCo [25]. A possible explanation for this discrepancy might be PYoCo's use of stronger text embeddings. Specifically, they utilize both CLIP [56] and T5-XXL [61] encoders, whereas we employ a T5-XL [61] text encoder only.", + "bbox": [ + 212, + 734, + 787, + 840 + ], + "page_idx": 12 + }, + { + "type": "header", + "text": "Photorealistic Video Generation with Diffusion Models", + "bbox": [ + 364, + 113, + 730, + 127 + ], + "page_idx": 12 + }, + { + "type": "page_number", + "text": "13", + "bbox": [ + 767, + 114, + 784, + 126 + ], + "page_idx": 12 + }, + { + "type": "image", + "img_path": "images/b4fc56e49d0c06ee7d076fb75256abb4700ab7f064b86ef37f7efeb8527cb724.jpg", + "image_caption": [ + "Fig. 4: Samples at $512 \\times 896$ resolution. Top: Text-to-video samples of consistent 3D camera motion (5.1 secs). Bottom: Image-to-video samples (3.8 secs). Images are generated using off-shelf models. Visit our project website to view them in video format." + ], + "image_footnote": [], + "bbox": [ + 218, + 143, + 782, + 439 + ], + "page_idx": 13 + }, + { + "type": "text", + "text": "Qualitative Results As mentioned in § 4.4, we jointly train our model on the task of frame prediction conditioned on 1 or 2 latent frames. Hence, our model can be used for animating images (image-to-video) and generating longer videos with consistent camera motion (Fig. 4). See videos on our project website.", + "bbox": [ + 212, + 542, + 787, + 604 + ], + "page_idx": 13 + }, + { + "type": "text", + "text": "6 Conclusion", + "text_level": 1, + "bbox": [ + 215, + 627, + 359, + 643 + ], + "page_idx": 13 + }, + { + "type": "text", + "text": "In this work, we introduce W.A.L.T, a simple, scalable, and efficient transformer-based framework for latent video diffusion models. A key design decision in our framework is to operate on a shared, compressed visual representation that enables generative modeling of both images and videos. We demonstrate state-of-the-art results for image and video generation using a transformer backbone with windowed attention. We also train a cascade of three W.A.L.T models jointly on image and video datasets, to synthesize high-resolution, temporally consistent photorealistic videos from natural language descriptions. While generative modeling has seen tremendous recent advances for images, progress on video generation has lagged behind. We hope that scaling our framework will enable the creation of powerful video generative models which can understand and model the physical world.", + "bbox": [ + 212, + 657, + 787, + 840 + ], + "page_idx": 13 + }, + { + "type": "page_number", + "text": "14", + "bbox": [ + 217, + 114, + 235, + 126 + ], + "page_idx": 13 + }, + { + "type": "header", + "text": "A. Gupta et al.", + "bbox": [ + 271, + 114, + 375, + 128 + ], + "page_idx": 13 + }, + { + "type": "text", + "text": "Acknowledgements", + "text_level": 1, + "bbox": [ + 217, + 143, + 401, + 162 + ], + "page_idx": 14 + }, + { + "type": "text", + "text": "We thank Bryan Seybold, Dan Kondratyuk, David Ross, Hartwig Adam, Huisheng Wang, Jason Baldridge, Mauricio Delbracio and Orly Liba for helpful discussions and feedback.", + "bbox": [ + 212, + 172, + 785, + 218 + ], + "page_idx": 14 + }, + { + "type": "text", + "text": "References", + "text_level": 1, + "bbox": [ + 217, + 239, + 321, + 255 + ], + "page_idx": 14 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "1. Agostinelli, A., Denk, T.I., Borsos, Z., Engel, J., Verzetti, M., Caillon, A., Huang, Q., Jansen, A., Roberts, A., Tagliasacchi, M., et al.: MusicLM: Generating music from text. arXiv:2301.11325 (2023) 2", + "2. Bao, F., Li, C., Cao, Y., Zhu, J.: All are worth words: a vit backbone for score-based diffusion models. In: NeurIPS 2022 Workshop on Score-Based Methods (2022) 4", + "3. Bengio, S., Vinyals, O., Jaitly, N., Shazeer, N.: Scheduled sampling for sequence prediction with recurrent neural networks. Advances in neural information processing systems 28 (2015) 7", + "4. Blattmann, A., Dockhorn, T., Kulal, S., Mendelevitch, D., Kilian, M., Lorenz, D., Levi, Y., English, Z., Voleti, V., Letts, A., et al.: Stable video diffusion: Scaling latent video diffusion models to large datasets. arXiv preprint arXiv:2311.15127 (2023) 13", + "5. Blattmann, A., Rombach, R., Ling, H., Dockhorn, T., Kim, S.W., Fidler, S., Kreis, K.: Align your latents: High-resolution video synthesis with latent diffusion models. In: CVPR (2023) 3, 4, 6, 13", + "6. Bousmalis, K., Vezzani, G., Rao, D., Devin, C., Lee, A.X., Bauza, M., Davchev, T., Zhou, Y., Gupta, A., Raju, A., et al.: Robocat: A self-improving foundation agent for robotic manipulation. arXiv preprint arXiv:2306.11706 (2023) 2, 13", + "7. Brock, A., Donahue, J., Simonyan, K.: Large scale GAN training for high fidelity natural image synthesis. In: ICLR (2018) 9", + "8. Brohan, A., Brown, N., Carbajal, J., Chebotar, Y., Dabis, J., Finn, C., Gopalakrishnan, K., Hausman, K., Herzog, A., Hsu, J., et al.: Rt-1: Robotics transformer for real-world control at scale. arXiv preprint arXiv:2212.06817 (2022) 2", + "9. Caron, M., Touvron, H., Misra, I., Jégou, H., Mairal, J., Bojanowski, P., Joulin, A.: Emerging properties in self-supervised vision transformers. In: Proceedings of the IEEE/CVF international conference on computer vision. pp. 9650-9660 (2021) 9", + "0. Carreira, J., Noland, E., Banki-Horvath, A., Hillier, C., Zisserman, A.: A short note about Kinetics-600. arXiv:1808.01340 (2018) 3, 8", + "1. Chang, H., Zhang, H., Barber, J., Maschinot, A., Lezama, J., Jiang, L., Yang, M.H., Murphy, K., Freeman, W.T., Rubinstein, M., et al.: Muse: Text-to-image generation via masked generative transformers. In: ICML (2023) 4", + "2. Chang, H., Zhang, H., Jiang, L., Liu, C., Freeman, W.T.: MaskGIT: Masked generative image transformer. In: CVPR (2022) 4", + "3. Chen, J., Yu, J., Ge, C., Yao, L., Xie, E., Wu, Y., Wang, Z., Kwok, J., Luo, P., Lu, H., et al.: Pixart- $\\alpha$ beta: Fast training of diffusion transformer for photorealistic text-to-image synthesis. arXiv preprint arXiv:2310.00426 (2023) 4", + "4. Chen, T., Zhang, R., Hinton, G.: Analog bits: Generating discrete data using diffusion models with self-conditioning. arXiv preprint arXiv:2208.04202 (2022) 7", + "5. Dehghani, M., Djolonga, J., Mustafa, B., Padlewski, P., Heek, J., Gilmer, J., Steiner, A.P., Caron, M., Geirhos, R., Alabdulmohsin, I., et al.: Scaling vision transformers to 22 billion parameters. In: International Conference on Machine Learning. pp. 7480-7512. PMLR (2023) 10, 11, 13" + ], + "bbox": [ + 225, + 267, + 785, + 839 + ], + "page_idx": 14 + }, + { + "type": "header", + "text": "Photorealistic Video Generation with Diffusion Models", + "bbox": [ + 364, + 114, + 730, + 127 + ], + "page_idx": 14 + }, + { + "type": "page_number", + "text": "15", + "bbox": [ + 767, + 116, + 784, + 126 + ], + "page_idx": 14 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "16. Deng, J., Dong, W., Socher, R., Li, L.J., Li, K., Fei-Fei, L.: ImageNet: A large-scale hierarchical image database. In: CVPR (2009) 3", + "17. Dhariwal, P., Nichol, A.: Diffusion models beat GANs on image synthesis. In: NeurIPS (2021) 2, 7, 9", + "18. Ding, M., Yang, Z., Hong, W., Zheng, W., Zhou, C., Yin, D., Lin, J., Zou, X., Shao, Z., Yang, H., et al.:CogView: Mastering text-to-image generation via transformers. In:NeurIPS (2021) 4", + "19. Dosovitskiy, A., Beyer, L., Kolesnikov, A., Weissenborn, D., Zhai, X., Unterthiner, T., Dehghani, M., Minderer, M., Heigold, G., Gelly, S., et al.: An image is worth 16x16 words: Transformers for image recognition at scale. In: ICLR (2020) 2, 6, 9", + "20. Dumoulin, V., Shlens, J., Kudlur, M.: A learned representation for artistic style. arXiv preprint arXiv:1610.07629 (2016) 7", + "21. Esser, P., Rombach, R., Ommer, B.: Taming transformers for high-resolution image synthesis. In: CVPR (2021) 2, 4, 5", + "22. Gafni, O., Polyak, A., Ashual, O., Sheynin, S., Parikh, D., Taigman, Y.: Makea-scene: Scene-based text-to-image generation with human priors. In: European Conference on Computer Vision. pp. 89-106. Springer (2022) 4", + "23. Gao, S., Zhou, P., Cheng, M.M., Yan, S.: Masked diffusion transformer is a strong image synthesizer. arXiv:2303.14389 (2023) 4, 9", + "24. Ge, S., Hayes, T., Yang, H., Yin, X., Pang, G., Jacobs, D., Huang, J.B., Parikh, D.: Long video generation with time-agnostic VQGAN and time-sensitive transformer. In: ECCV (2022) 4, 5, 7", + "25. Ge, S., Nah, S., Liu, G., Poon, T., Tao, A., Catanzaro, B., Jacobs, D., Huang, J.B., Liu, M.Y., Balaji, Y.: Preserve your own correlation: A noise prior for video diffusion models. arXiv preprint arXiv:2305.10474 (2023) 3, 6, 13", + "26. Girdhar, R., Singh, M., Brown, A., Duval, Q., Azadi, S., Rambhatla, S.S., Shah, A., Yin, X., Parikh, D., Misra, I.: Emu video: Factorizing text-to-video generation by explicit image conditioning. arXiv preprint arXiv:2311.10709 (2023) 13", + "27. Goodfellow, I., Pouget-Abadie, J., Mirza, M., Xu, B., Warde-Farley, D., Ozair, S., Courville, A., Bengio, Y.: Generative adversarial nets. Advances in neural information processing systems 27 (2014) 5", + "28. Google: PaLM 2 technical report. arXiv:2305.10403 (2023) 2", + "29. Gupta, A., Tian, S., Zhang, Y., Wu, J., Martín-Martín, R., Fei-Fei, L.: MaskViT: Masked visual pre-training for video prediction. In: ICLR (2022) 4, 6", + "30. Gupta, A., Wu, J., Deng, J., Fei-Fei, L.: Siamese masked autoencoders. arXiv preprint arXiv:2305.14344 (2023) 9", + "31. Harvey, W., Naderiparizi, S., Masrani, V., Weilbach, C., Wood, F.: Flexible diffusion modeling of long videos. Advances in Neural Information Processing Systems 35, 27953-27965 (2022) 3", + "32. He, K., Chen, X., Xie, S., Li, Y., Dollar, P., Girshick, R.: Masked autoencoders are scalable vision learners. In: CVPR (2022) 2", + "33. He, Y., Yang, T., Zhang, Y., Shan, Y., Chen, Q.: Latent video diffusion models for high-fidelity long video generation. arXiv preprint arXiv:2211.13221 (2023) 3", + "34. Heusel, M., Ramsauer, H., Unterthiner, T., Nessler, B., Hochreiter, S.: GANs trained by a two time-scale update rule converge to a local nash equilibrium. In: NeurIPS (2017) 9", + "35. Ho, J., Chan, W., Sahara, C., Whang, J., Gao, R., Gritsenko, A., Kingma, D.P., Poole, B., Norouzi, M., Fleet, D.J., et al.: Imagen video: High definition video generation with diffusion models. arXiv:2210.02303 (2022) 3, 4, 5, 13", + "36. Ho, J., Jain, A., Abbeel, P.: Denoising diffusion probabilistic models. In: NeurIPS (2020) 2, 3, 4" + ], + "bbox": [ + 215, + 146, + 785, + 839 + ], + "page_idx": 15 + }, + { + "type": "page_number", + "text": "16", + "bbox": [ + 217, + 114, + 235, + 126 + ], + "page_idx": 15 + }, + { + "type": "header", + "text": "A. Gupta et al.", + "bbox": [ + 271, + 114, + 374, + 128 + ], + "page_idx": 15 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "37. Ho, J., Sahara, C., Chan, W., Fleet, D.J., Norouzi, M., Salimans, T.: Cascaded diffusion models for high fidelity image generation. JMLR 23(1), 2249-2281 (2022) 8", + "38. Ho, J., Salimans, T., Gritsenko, A., Chan, W., Norouzi, M., Fleet, D.J.: Video diffusion models. In: ICLR Workshops (2022) 2, 3, 4, 7, 13", + "39. Hong, W., Ding, M., Zheng, W., Liu, X., Tang, J.: CogVideo: Large-scale pretraining for text-to-video generation via transformers. arXiv:2205.15868 (2022) 4, 13", + "40. Hoogeboom, E., Heek, J., Salimans, T.: simple diffusion: End-to-end diffusion for high resolution images. In: ICML (2023) 3, 4, 9", + "41. Hu, E.J., Shen, Y., Wallis, P., Allen-Zhu, Z., Li, Y., Wang, S., Wang, L., Chen, W.: Lora: Low-rank adaptation of large language models. In: ICLR (2021) 7", + "42. Jabri, A., Fleet, D.J., Chen, T.: Scalable adaptive computation for iterative generation. In: ICML (2023) 4, 7, 9", + "43. Jiang, Y., Chang, S., Wang, Z.: Transgan: Two pure transformers can make one strong gan, and that can scale up. Advances in Neural Information Processing Systems 34, 14745-14758 (2021) 4", + "44. Johnson, J., Alahi, A., Fei-Fei, L.: Perceptual losses for real-time style transfer and super-resolution. In: ECCV (2016) 5", + "45. Karras, T., Laine, S., Aila, T.: A style-based generator architecture for generative adversarial networks. In: CVPR (2019) 7", + "46. Kingma, D.P., Gao, R.: Understanding the diffusion objective as a weighted integral of ellbos. arXiv:2303.00848 (2023) 9", + "47. LeCun, Y., Bottou, L., Bengio, Y., Haffner, P.: Gradient-based learning applied to document recognition. Proceedings of the IEEE 86(11), 2278-2324 (1998) 2", + "48. Lee, K., Chang, H., Jiang, L., Zhang, H., Tu, Z., Liu, C.: Vitgan: Training gans with vision transformers. arXiv preprint arXiv:2107.04589 (2021) 4", + "49. Lin, S., Liu, B., Li, J., Yang, X.: Common diffusion noise schedules and sample steps are flawed. arXiv preprint arXiv:2305.08891 (2023) 10, 11", + "50. Liu, Z., Lin, Y., Cao, Y., Hu, H., Wei, Y., Zhang, Z., Lin, S., Guo, B.: Swin transformer: Hierarchical vision transformer using shifted windows. In: Proceedings of the IEEE/CVF international conference on computer vision. pp. 10012-10022 (2021) 6", + "51. Lu, H., Yang, G., Fei, N., Huo, Y., Lu, Z., Luo, P., Ding, M.: Vdt: General-purpose video diffusion transformers via mask modeling. arXiv preprint arXiv:2305.13311 (2023) 4", + "52. Luc, P., Clark, A., Dieleman, S., Casas, D.d.L., Doron, Y., Cassirer, A., Simonyan, K.: Transformation-based adversarial video prediction on large-scale data. arXiv:2003.04035 (2020) 7", + "53. Nichol, A.Q., Dhariwal, P.: Improved denoising diffusion probabilistic models. In: International Conference on Machine Learning. pp. 8162-8171. PMLR (2021) 3, 7", + "54. Peebles, W., Xie, S.: Scalable diffusion models with transformers. arXiv:2212.09748 (2022) 4, 7, 9", + "55. Perez, E., Strub, F., De Vries, H., Dumoulin, V., Courville, A.: Film: Visual reasoning with a general conditioning layer. In: Proceedings of the AAAI conference on artificial intelligence. vol. 32 (2018) 7", + "56. Radford, A., Kim, J.W., Hallacy, C., Ramesh, A., Goh, G., Agarwal, S., Sastry, G., Askell, A., Mishkin, P., Clark, J., et al.: Learning transferable visual models from natural language supervision. In: ICML (2021) 2, 4, 13", + "57. Radford, A., Kim, J.W., Xu, T., Brockman, G., McLeavey, C., Sutskever, I.: Robust speech recognition via large-scale weak supervision. In: International Conference on Machine Learning. pp. 28492-28518. PMLR (2023) 2" + ], + "bbox": [ + 212, + 146, + 787, + 839 + ], + "page_idx": 16 + }, + { + "type": "header", + "text": "Photorealistic Video Generation with Diffusion Models", + "bbox": [ + 364, + 114, + 730, + 126 + ], + "page_idx": 16 + }, + { + "type": "page_number", + "text": "17", + "bbox": [ + 767, + 116, + 784, + 126 + ], + "page_idx": 16 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "58. Radford, A., Narasimhan, K., Salimans, T., Sutskever, I., et al.: Improving language understanding by generative pre-training (2018) 2, 4, 13", + "59. Radford, A., Wu, J., Child, R., Luan, D., Amodei, D., Sutskever, I., et al.: Language models are unsupervised multitask learners. OpenAI blog 1(8), 9 (2019) 2, 4", + "60. Ramesh, A., Pavlov, M., Goh, G., Gray, S., Voss, C., Radford, A., Chen, M., Sutskever, I.: Zero-shot text-to-image generation. In: ICML (2021) 4", + "61. Roberts, A., Raffel, C., Lee, K., Matena, M., Shazeer, N., Liu, P.J., Narang, S., Li, W., Zhou, Y.: Exploring the limits of transfer learning with a unified text-to-text transformer (2019) 13", + "62. Rombach, R., Blattmann, A., Lorenz, D., Esser, P., Ommer, B.: High-resolution image synthesis with latent diffusion models. In: CVPR (2022) 2, 3, 9, 11", + "63. Ronneberger, O., Fischer, P., Brox, T.: U-Net: Convolutional networks for biomedical image segmentation. In: MICCAI (2015) 2, 4", + "64. Salimans, T., Goodfellow, I., Zaremba, W., Cheung, V., Radford, A., Chen, X.: Improved techniques for training GANs. In: NeurIPS (2016) 9", + "65. Salimans, T., Ho, J.: Progressive distillation for fast sampling of diffusion models arXiv preprint arXiv:2202.00512 (2022) 5", + "66. Savinov, N., Chung, J., Binkowski, M., Elsen, E., Oord, A.v.d.: Step-unrolled denoising autoencoders for text generation. arXiv preprint arXiv:2112.06749 (2021) 7", + "67. Singer, U., Polyak, A., Hayes, T., Yin, X., An, J., Zhang, S., Hu, Q., Yang, H., Ashual, O., Gafni, O., et al.: Make-a-video: Text-to-video generation without text-video data. arXiv:2209.14792 (2022) 3, 4, 13", + "68. Sohl-Dickstein, J., Weiss, E., Maheswaranathan, N., Ganguli, S.: Deep unsupervised learning using nonequilibrium thermodynamics. In: ICML (2015) 2, 3, 4", + "69. Song, J., Meng, C., Ermon, S.: Denoising diffusion implicit models. arXiv preprint arXiv:2010.02502 (2020) 3, 9", + "70. Song, Y., Ermon, S.: Generative modeling by estimating gradients of the data distribution. In: NeurIPS (2019) 2, 4", + "71. Soomro, K., Zamir, A.R., Shah, M.: UCF101: A dataset of 101 human actions classes from videos in the wild. arXiv:1212.0402 (2012) 3, 8, 10", + "72. Unterthiner, T., van Steenkiste, S., Kurach, K., Marinier, R., Michalski, M., Gelly, S.: Towards accurate generative models of video: A new metric & challenges. arXiv:1812.01717 (2018) 8", + "73. Van Den Oord, A., Vinyals, O., et al.: Neural discrete representation learning. In: NeurIPS (2017) 2, 5", + "74. Vaswani, A., Shazeer, N., Parmar, N., Uszkoreit, J., Jones, L., Gomez, A.N., Kaiser, L., Polosukhin, I.: Attention is all you need. In: NeurIPS (2017) 2, 4, 6", + "75. Villegas, R., Babaeizadeh, M., Kindermans, P.J., Moraldo, H., Zhang, H., Saffar, M.T., Castro, S., Kunze, J., Erhan, D.: Phenaki: Variable length video generation from open domain textual description. arXiv:2210.02399 (2022) 4, 5, 7", + "76. Vincent, P., Larochelle, H., Bengio, Y., Manzagol, P.A.: Extracting and composing robust features with denoising autoencoders. In: Proceedings of the 25th international conference on Machine learning. pp. 1096-1103 (2008) 2", + "77. Wu, C., Huang, L., Zhang, Q., Li, B., Ji, L., Yang, F., Sapiro, G., Duan, N.: Godiva: Generating open-domain videos from natural descriptions. arXiv preprint arXiv:2104.14806 (2021) 4", + "78. Wu, C., Liang, J., Ji, L., Yang, F., Fang, Y., Jiang, D., Duan, N.: Niwa: Visual synthesis pre-training for neural visual world creation. In: European conference on computer vision. pp. 720-736. Springer (2022) 4" + ], + "bbox": [ + 215, + 147, + 785, + 839 + ], + "page_idx": 17 + }, + { + "type": "page_number", + "text": "18", + "bbox": [ + 217, + 114, + 235, + 126 + ], + "page_idx": 17 + }, + { + "type": "header", + "text": "A. Gupta et al.", + "bbox": [ + 271, + 114, + 374, + 128 + ], + "page_idx": 17 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "79. Yan, W., Zhang, Y., Abbeel, P., Srinivas, A.: VideoGPT: Video generation using VQ-VAE and transformers. arXiv:2104.10157 (2021) 4", + "80. Yu, J., Li, X., Koh, J.Y., Zhang, H., Pang, R., Qin, J., Ku, A., Xu, Y., Baldridge, J., Wu, Y.: Vector-quantized image modeling with improved VQGAN. In: ICLR (2022) 4", + "81. Yu, J., Xu, Y., Koh, J.Y., Luong, T., Baid, G., Wang, Z., Vasudevan, V., Ku, A., Yang, Y., Ayan, B.K., et al.: Scaling autoregressive models for content-rich text-to-image generation. arXiv:2206.10789 (2022) 4", + "82. Yu, L., Cheng, Y., Sohn, K., Lezama, J., Zhang, H., Chang, H., Hauptmann, A.G., Yang, M.H., Hao, Y., Essa, I., et al.: MAGVIT: Masked generative video transformer. In: CVPR (2023) 4, 5, 7", + "83. Yu, L., Lezama, J., Gundavarapu, N.B., Versari, L., Sohn, K., Minnen, D., Cheng, Y., Gupta, A., Gu, X., Hauptmann, A.G., et al.: Language model beats diffusion–tokenizer is key to visual generation. arXiv preprint arXiv:2310.05737 (2023) 5, 7", + "84. Yu, S., Sohn, K., Kim, S., Shin, J.: Video probabilistic diffusion models in projected latent space. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 18456-18466 (2023) 3", + "85. Zeng, Y., Wei, G., Zheng, J., Zou, J., Wei, Y., Zhang, Y., Li, H.: Make pixels dance: High-dynamic video generation. arXiv preprint arXiv:2311.10982 (2023) 13", + "86. Zhai, X., Kolesnikov, A., Houlsby, N., Beyer, L.: Scaling vision transformers. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 12104-12113 (2022) 9", + "87. Zhang, B., Gu, S., Zhang, B., Bao, J., Chen, D., Wen, F., Wang, Y., Guo, B.: Styleswin: Transformer-based gan for high-resolution image generation. In: Proceedings of the IEEE/CVF conference on computer vision and pattern recognition. pp. 11304-11314 (2022) 4", + "88. Zhang, R., Isola, P., Efros, A.A., Shechtman, E., Wang, O.: The unreasonable effectiveness of deep features as a perceptual metric. In: CVPR (2018) 5", + "89. Zheng, H., Nie, W., Vahdat, A., Anandkumar, A.: Fast training of diffusion models with masked transformers. arXiv:2306.09305 (2023) 4, 9", + "90. Zhou, D., Wang, W., Yan, H., Lv, W., Zhu, Y., Feng, J.: Magicvideo: Efficient video generation with latent diffusion models. arXiv preprint arXiv:2211.11018 (2022) 13", + "91. Zitkovich, B., Yu, T., Xu, S., Xu, P., Xiao, T., Xia, F., Wu, J., Wohlhart, P., Welker, S., Wahid, A., et al.: RT-2: Vision-language-action models transfer web knowledge to robotic control. In: CoRL (2023) 2" + ], + "bbox": [ + 212, + 146, + 787, + 645 + ], + "page_idx": 18 + }, + { + "type": "header", + "text": "Photorealistic Video Generation with Diffusion Models", + "bbox": [ + 364, + 114, + 730, + 127 + ], + "page_idx": 18 + }, + { + "type": "page_number", + "text": "19", + "bbox": [ + 767, + 116, + 784, + 126 + ], + "page_idx": 18 + } +] \ No newline at end of file diff --git a/2024/Photorealistic Video Generation with Diffusion Models/327f0bbf-9985-4862-ae62-9cd9ec84c015_model.json b/2024/Photorealistic Video Generation with Diffusion Models/327f0bbf-9985-4862-ae62-9cd9ec84c015_model.json new file mode 100644 index 0000000000000000000000000000000000000000..2512be141e90057d83e3e103d66bca5a275e2f1a --- /dev/null +++ b/2024/Photorealistic Video Generation with Diffusion Models/327f0bbf-9985-4862-ae62-9cd9ec84c015_model.json @@ -0,0 +1,3010 @@ +[ + [ + { + "type": "title", + "bbox": [ + 0.233, + 0.141, + 0.773, + 0.185 + ], + "angle": 0, + "content": "Photorealistic Video Generation with Diffusion Models" + }, + { + "type": "text", + "bbox": [ + 0.218, + 0.212, + 0.787, + 0.245 + ], + "angle": 0, + "content": "Agrim Gupta\\(^{1,2,*}\\), Lijun Yu\\(^{2}\\), Kihyuk Sohn\\(^{2}\\), Xiuye Gu\\(^{2}\\), Meera Hahn\\(^{2}\\), Fei-Fei Li\\(^{1}\\), Irfan Essa\\(^{2,3}\\), Lu Jiang\\(^{2}\\), and José Lezama\\(^{2}\\)" + }, + { + "type": "text", + "bbox": [ + 0.428, + 0.255, + 0.576, + 0.269 + ], + "angle": 0, + "content": "1 Stanford University" + }, + { + "type": "text", + "bbox": [ + 0.438, + 0.269, + 0.565, + 0.283 + ], + "angle": 0, + "content": "2 Google Research" + }, + { + "type": "text", + "bbox": [ + 0.388, + 0.283, + 0.616, + 0.298 + ], + "angle": 0, + "content": "3 Georgia Institute of Technology" + }, + { + "type": "text", + "bbox": [ + 0.35, + 0.298, + 0.654, + 0.311 + ], + "angle": 0, + "content": "https://walt-video-diffusion.github.io/" + }, + { + "type": "image", + "bbox": [ + 0.218, + 0.326, + 0.331, + 0.378 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.331, + 0.327, + 0.444, + 0.378 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.444, + 0.327, + 0.557, + 0.378 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.557, + 0.327, + 0.669, + 0.378 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.67, + 0.327, + 0.783, + 0.378 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.397, + 0.381, + 0.61, + 0.392 + ], + "angle": 0, + "content": "A cute cat riding a yellow surfboard, fast waves." + }, + { + "type": "image", + "bbox": [ + 0.218, + 0.397, + 0.331, + 0.449 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.331, + 0.397, + 0.444, + 0.449 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.444, + 0.397, + 0.557, + 0.449 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.557, + 0.397, + 0.669, + 0.449 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.67, + 0.397, + 0.783, + 0.449 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.32, + 0.455, + 0.671, + 0.466 + ], + "angle": 0, + "content": "A cute cordi wearing a blue sweater, walking towards the camera, slow motion." + }, + { + "type": "image", + "bbox": [ + 0.218, + 0.472, + 0.331, + 0.524 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.331, + 0.472, + 0.444, + 0.524 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.444, + 0.472, + 0.557, + 0.524 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.557, + 0.472, + 0.669, + 0.524 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.67, + 0.472, + 0.783, + 0.524 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.316, + 0.53, + 0.68, + 0.541 + ], + "angle": 0, + "content": "A raccoon wearing a black jacket, dancing in slow motion in front of the pyramids." + }, + { + "type": "image_caption", + "bbox": [ + 0.214, + 0.557, + 0.788, + 0.599 + ], + "angle": 0, + "content": "Fig. 1: W.A.L.T samples for text-to-video generation. W.A.L.T generates high-resolution, temporally consistent photorealistic videos from text prompts. The samples shown are \\(512 \\times 896\\) resolution over 3.6 seconds duration at 8 frames per second." + }, + { + "type": "text", + "bbox": [ + 0.261, + 0.618, + 0.744, + 0.799 + ], + "angle": 0, + "content": "Abstract. We present W.A.L.T, a diffusion transformer for photorealistic video generation from text prompts. Our approach has two key design decisions. First, we use a causal encoder to jointly compress images and videos within a unified latent space, enabling training and generation across modalities. Second, for memory and training efficiency, we use a window attention architecture tailored for joint spatial and spatiotemporal generative modeling. Taken together these design decisions enable us to achieve state-of-the-art performance on established video (UCF-101 and Kinetics-600) and image (ImageNet) generation benchmarks without using classifier free guidance. Finally, we also train a cascade of three models for the task of text-to-video generation consisting of a base latent video diffusion model, and two video super-resolution diffusion models to generate videos of \\(512 \\times 896\\) resolution at 8 frames per second." + }, + { + "type": "page_footnote", + "bbox": [ + 0.233, + 0.826, + 0.588, + 0.84 + ], + "angle": 0, + "content": "*Work partially done during an internship at Google." + } + ], + [ + { + "type": "page_number", + "bbox": [ + 0.218, + 0.116, + 0.23, + 0.127 + ], + "angle": 0, + "content": "2" + }, + { + "type": "header", + "bbox": [ + 0.272, + 0.115, + 0.377, + 0.129 + ], + "angle": 0, + "content": "A. Gupta et al." + }, + { + "type": "title", + "bbox": [ + 0.218, + 0.145, + 0.377, + 0.162 + ], + "angle": 0, + "content": "1 Introduction" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.175, + 0.787, + 0.236 + ], + "angle": 0, + "content": "Modeling the complex and dynamic visual world with a high level of detail is the ultimate goal of visual generative models. To accomplish this goal, we need neural architectures that can efficiently process high dimensional inputs, capture long range dependencies and have powerful scaling and parallelization properties." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.237, + 0.787, + 0.341 + ], + "angle": 0, + "content": "The transformer [74] is a neural network architecture designed with such desirable properties. This has encouraged the research community to increasingly favor it over domain-specific architectures in diverse fields such as language [28, 56, 58, 59], audio [1], speech [57], vision [19, 32], and robotics [6, 8, 91]. Such a trend towards unification allows researchers to share and build upon advances in traditionally disparate domains. Thus, leading to a virtuous cycle of innovation and improvement in model design favoring transformers." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.343, + 0.787, + 0.476 + ], + "angle": 0, + "content": "Generative modelling of images and videos has been however a notable exception to this trend. Diffusion models [68, 70], which have emerged as a leading paradigm for generative modelling of images [17, 36] and videos [38], predominantly rely on the U-Net architecture [36, 63] as backbone, consisting of a series of convolutional [47] and self-attention [74] layers. This preference stems from the fact that the memory demands of full attention mechanisms in transformers scales quadratically with input sequence length. Such scaling leads to prohibitively high costs when processing high-dimensional signals like video with transformers." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.479, + 0.787, + 0.688 + ], + "angle": 0, + "content": "The computational requirements of diffusion models can be reduced by operating in a lower-dimensional latent space derived from an autoencoder [21,73,76], giving rise to latent diffusion models (LDMs) [62]. A critical design choice in this context is the type of latent space employed: spatial compression (per frame latents) versus spatiotemporal compression. Spatial compression is often preferred because it enables leveraging pre-trained image autoencoders and LDMs, which are trained on large paired image-text datasets. However, this choice increases network complexity and limits the use of transformers as backbones, especially in generating high-resolution videos due to memory constraints. On the other hand, while spatiotemporal compression can mitigate these issues, it precludes the use of paired image-text datasets, which are much larger and diverse than their video counterparts. Hence, to leverage all sources of visual data and unlock scaling via transformers, it is critical to have a unified compressed representation of images and videos." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.689, + 0.787, + 0.84 + ], + "angle": 0, + "content": "In this work, we present Window Attention Latent Transformer (W.A.L.T): a novel transformer-based method for latent video diffusion models (LVDMs). Our method consists of two stages. First, an autoencoder maps both videos and images into a unified, lower-dimensional latent space. This design choice enables training a single generative model jointly on image and video datasets and significantly reduces the computational burden for generating high resolution videos. Subsequently, we propose a new design of transformer blocks for latent video diffusion modeling composed of self-attention layers that alternate between non-overlapping, window-restricted spatial and spatiotemporal attention. This design offers two primary benefits: Firstly, the use of local window attention" + } + ], + [ + { + "type": "header", + "bbox": [ + 0.366, + 0.115, + 0.733, + 0.128 + ], + "angle": 0, + "content": "Photorealistic Video Generation with Diffusion Models" + }, + { + "type": "page_number", + "bbox": [ + 0.775, + 0.117, + 0.785, + 0.127 + ], + "angle": 0, + "content": "3" + }, + { + "type": "image", + "bbox": [ + 0.217, + 0.144, + 0.782, + 0.277 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.214, + 0.286, + 0.788, + 0.357 + ], + "angle": 0, + "content": "Fig. 2: W.A.L.T. We encode images and videos into a shared latent space. The transformer backbone processes these latents with blocks having two layers of window-restricted attention: spatial layers capture spatial relations in both images and video, while spatiotemporal layers model temporal dynamics in videos and pass through images via identity attention mask. Text conditioning is done via spatial cross-attention." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.375, + 0.788, + 0.434 + ], + "angle": 0, + "content": "significantly lowers computational demands. Secondly, it facilitates joint training, where the spatial layers independently process images and video frames, while the spatiotemporal layers are dedicated to modeling the temporal relationships in videos." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.435, + 0.788, + 0.601 + ], + "angle": 0, + "content": "While conceptually simple, our method provides the first empirical evidence of transformers' superior generation quality and parameter efficiency in latent video diffusion on public benchmarks. Specifically, we report state-of-the-art results on class-conditional video generation (UCF-101 [71]), frame prediction (Kinetics-600 [10]) and class conditional image generation (ImageNet [16]) without using classifier free guidance. Finally, to showcase the scalability and efficiency of our method we also demonstrate results on the challenging task of photorealistic text-to-video generation. We train a cascade of three models consisting of a base latent video diffusion model, and two video super-resolution diffusion models to generate videos of \\(512 \\times 896\\) resolution at 8 frames per second and report state-of-the-art zero-shot FVD score on the UCF-101 benchmark." + }, + { + "type": "text", + "bbox": [ + 0.24, + 0.602, + 0.595, + 0.617 + ], + "angle": 0, + "content": "Our contributions can be summarized as follows," + }, + { + "type": "text", + "bbox": [ + 0.226, + 0.626, + 0.784, + 0.655 + ], + "angle": 0, + "content": "- We propose W.A.L.T, a framework for transformer-based latent diffusion in a unified image and video latent space." + }, + { + "type": "text", + "bbox": [ + 0.225, + 0.656, + 0.785, + 0.701 + ], + "angle": 0, + "content": "- We validate W.A.L.T by performing extensive ablations and achieving state-of-the-art video generation on frame prediction (K600), class-conditional video (UCF-101) and image (ImageNet) generation." + }, + { + "type": "text", + "bbox": [ + 0.225, + 0.701, + 0.785, + 0.73 + ], + "angle": 0, + "content": "- We demonstrate the scaling properties of W.A.L.T and achieve state-of-the-art results in zero-shot UCF-101 text-to-video benchmark." + }, + { + "type": "list", + "bbox": [ + 0.225, + 0.626, + 0.785, + 0.73 + ], + "angle": 0, + "content": null + }, + { + "type": "title", + "bbox": [ + 0.216, + 0.764, + 0.388, + 0.78 + ], + "angle": 0, + "content": "2 Related Work" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.795, + 0.787, + 0.843 + ], + "angle": 0, + "content": "Video Diffusion Models. Diffusion models have shown impressive results in image [36,40,53,62,68,69] and video generation [5,25,31,35,38,67]. Video diffusion models can be categorized into pixel-space [35,38,67] and latent-space [5,25,33,84]" + } + ], + [ + { + "type": "page_number", + "bbox": [ + 0.218, + 0.116, + 0.23, + 0.127 + ], + "angle": 0, + "content": "4" + }, + { + "type": "header", + "bbox": [ + 0.272, + 0.115, + 0.377, + 0.129 + ], + "angle": 0, + "content": "A. Gupta et al." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.147, + 0.788, + 0.314 + ], + "angle": 0, + "content": "approaches, the later bringing important efficiency advantages when modeling videos. Ho et al. [38] demonstrated that the quality of text conditioned video generation can be significantly improved by jointly training on image and video data. Similarly, to leverage image datasets, latent video diffusion models inflate a pre-trained image model, typically a U-Net [63], into a video model by adding temporal layers, and initializing them as the identity function [5,35,67]. Although computationally efficient, this approach couples the design of video and image models, and precludes spatiotemporal compression. In this work, we operate on a unified latent space for images and videos, allowing us to leverage large scale image and video datasets while enjoying computational efficiency gains from spatiotemporal compression of videos." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.327, + 0.79, + 0.614 + ], + "angle": 0, + "content": "Transformers for Generative Modeling. Multiple classes of generative models have utilized Transformers [74] as backbone, such as, Generative adversarial networks [43, 48, 87], autoregressive [11, 12, 21, 22, 29, 60, 75, 78, 79, 81, 82] and diffusion [2, 23, 42, 51, 54, 89] models. Inspired by the success of autoregressive pretraining of large language models [56, 58, 59], Ramesh et al. [60] trained a text-to-image generation model by predicting the next visual token obtained from an image tokenizer. Subsequently, this approach was applied to multiple applications including class-conditional image generation [21, 80], text-to-image [18, 60, 77, 81] or image-to-image translation [22, 78]. Similarly, for video generation, transformer-based models were proposed to predict next tokens using 3D extensions of VQGAN [24, 39, 79, 82] or using per frame image latents [29]. Autoregressive sampling of videos is typically impractical given the very long sequences involved. To alleviate this issue, non-autoregressive sampling [11, 12], i.e. parallel token prediction, has been adopted as a more efficient solution for transformer-based video generation [29, 75, 82]. Recently, the community has started adopting transformers as the denoising backbone for diffusion models in place of U-Net [13, 40, 51, 54, 89]. To the best of our knowledge, our work is the first successful empirical demonstration (§ 5.1) of a transformer-based backbone for jointly training image and video latent diffusion models." + }, + { + "type": "title", + "bbox": [ + 0.216, + 0.636, + 0.371, + 0.654 + ], + "angle": 0, + "content": "3 Background" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.666, + 0.788, + 0.728 + ], + "angle": 0, + "content": "Diffusion formulation. Diffusion models [36, 68, 70] are a class of generative models which learn to generate data by iteratively denoising samples drawn from a noise distribution. Gaussian diffusion models assume a forward noising process which gradually applies noise \\((\\epsilon)\\) to real data \\((x_0 \\sim p_{\\mathrm{data}})\\). Concretely," + }, + { + "type": "equation", + "bbox": [ + 0.393, + 0.738, + 0.786, + 0.755 + ], + "angle": 0, + "content": "\\[\n\\boldsymbol {x} _ {t} = \\sqrt {\\gamma (t)} \\boldsymbol {x} _ {0} + \\sqrt {1 - \\gamma (t)} \\epsilon , \\tag {1}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.764, + 0.788, + 0.811 + ], + "angle": 0, + "content": "where \\(\\epsilon \\sim \\mathcal{N}(\\mathbf{0},\\mathbf{I}), t \\in [0,1]\\), and \\(\\gamma(t)\\) is a monotonically decreasing function (noise schedule) from 1 to 0. Diffusion models are trained to learn the reverse process that inverts the forward corruptions:" + }, + { + "type": "equation", + "bbox": [ + 0.337, + 0.82, + 0.786, + 0.844 + ], + "angle": 0, + "content": "\\[\n\\mathbb {E} _ {\\boldsymbol {x} \\sim p _ {\\text {d a t a}}, t \\sim \\mathcal {U} (0, 1), \\boldsymbol {\\epsilon} \\sim \\mathcal {N} (\\boldsymbol {0}, \\boldsymbol {I})} \\left[ \\| \\boldsymbol {y} - f _ {\\theta} (\\boldsymbol {x} _ {t}; \\boldsymbol {c}, t) \\| ^ {2} \\right], \\tag {2}\n\\]" + } + ], + [ + { + "type": "header", + "bbox": [ + 0.366, + 0.114, + 0.733, + 0.128 + ], + "angle": 0, + "content": "Photorealistic Video Generation with Diffusion Models" + }, + { + "type": "page_number", + "bbox": [ + 0.775, + 0.117, + 0.785, + 0.127 + ], + "angle": 0, + "content": "5" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.147, + 0.788, + 0.207 + ], + "angle": 0, + "content": "where \\( f_{\\theta} \\) is the denoiser model parameterized by a neural network, \\( \\pmb{c} \\) is conditioning information e.g., class labels or text prompts, and the target \\( \\pmb{y} \\) can be random noise \\( \\pmb{\\epsilon} \\), denoised input \\( \\pmb{x_0} \\) or \\( \\pmb{v} = \\sqrt{1 - \\gamma(t)} \\pmb{\\epsilon} - \\sqrt{\\gamma(t)} \\pmb{x_0} \\). Following [35, 65], we use \\( \\pmb{v} \\)-prediction in all our experiments." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.208, + 0.79, + 0.375 + ], + "angle": 0, + "content": "Latent diffusion models (LDMs). Processing high-resolution images and videos using raw pixels requires considerable computational resources. To address this, LDMs operate on the low dimensional latent space of a VQ-VAE [21,73]. VQ-VAE consists of an encoder \\( E(\\pmb{x}) \\) that encodes an input video \\( \\pmb{x} \\in \\mathbb{R}^{T \\times H \\times W \\times 3} \\) into a latent representation \\( \\pmb{z} \\in \\mathbb{R}^{t \\times h \\times w \\times c} \\). The encoder downsamples the video by a factor of \\( f_{s} = H / h = W / w \\) and \\( f_{t} = T / t \\), where \\( T = t = 1 \\) corresponds to using an image autoencoder. An important distinction from the original VQ-VAE is the absence of a codebook of quantized embeddings as diffusion models can operate on continuous latent spaces. A decoder \\( D \\) is trained to predict a reconstruction of the video, \\( \\hat{\\pmb{x}} \\), from \\( \\pmb{z} \\). Following VQ-GAN [21], reconstruction quality can be further improved by adding adversarial [27] and perceptual losses [44,88]." + }, + { + "type": "title", + "bbox": [ + 0.215, + 0.395, + 0.345, + 0.41 + ], + "angle": 0, + "content": "4 W.A.L.T" + }, + { + "type": "title", + "bbox": [ + 0.215, + 0.425, + 0.457, + 0.44 + ], + "angle": 0, + "content": "4.1 Learning Visual Tokens" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.448, + 0.79, + 0.629 + ], + "angle": 0, + "content": "A key design decision in video generative modeling is the choice of latent space representation. Ideally, we want a shared and unified compressed visual representation that can be used for generative modeling of both images and videos [75,83]. The unified representation is important because joint image-video learning is preferable due to a scarcity of labeled video data [35], such as text-video pairs. Concretely, given a video sequence \\(\\pmb{x} \\in \\mathbb{R}^{(1 + T) \\times H \\times W \\times C}\\), we aim to learn a low-dimensional representation \\(\\pmb{z} \\in \\mathbb{R}^{(1 + t) \\times h \\times w \\times c}\\) that performs spatial-temporal compression by a factor of \\(f_{s} = H / h = W / w\\) in space and a factor of \\(f_{t} = T / t\\) in time. To enable a unified representation for both videos and static images, the first frame is always encoded independently from the rest of the video. This allows static images \\(\\pmb{x} \\in \\mathbb{R}^{1 \\times H \\times W \\times C}\\) to be treated as videos with a single frame, i.e. \\(\\pmb{z} \\in \\mathbb{R}^{1 \\times h \\times w \\times c}\\)." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.63, + 0.789, + 0.765 + ], + "angle": 0, + "content": "We instantiate this design with the causal 3D CNN encoder-decoder architecture of the MAGVIT-v2 tokenizer [83]. Typically the encoder-decoder consists of regular 3D convolution layers which cannot process the first frame independently [24,82]. This limitation stems from the fact that a regular convolutional kernel of size \\((k_{t},k_{h},k_{w})\\) will operate on \\(\\left\\lfloor \\frac{k_t - 1}{2}\\right\\rfloor\\) frames before and \\(\\left\\lfloor \\frac{k_t}{2}\\right\\rfloor\\) frames after the input frames. Causal 3D convolution layers solve this issue as the convolutional kernel operates on only the past \\(k_{t} - 1\\) frames. This ensures that the output for each frame is influenced solely by the preceding frames, enabling the model to tokenize the first frame independently." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.766, + 0.789, + 0.84 + ], + "angle": 0, + "content": "After this stage, the input to our model is a batch of latent tensors \\( \\mathbf{z} \\in \\mathbb{R}^{(1 + t) \\times h \\times w \\times c} \\) representing a single video or a stack of \\( 1 + t \\) independent images (Fig. 2). Different from [83], our latent representation is real-valued and quantization-free. In the section below we describe how our model jointly processes a mixed batch of images and videos." + } + ], + [ + { + "type": "page_number", + "bbox": [ + 0.218, + 0.116, + 0.23, + 0.127 + ], + "angle": 0, + "content": "6" + }, + { + "type": "header", + "bbox": [ + 0.272, + 0.115, + 0.377, + 0.129 + ], + "angle": 0, + "content": "A. Gupta et al." + }, + { + "type": "title", + "bbox": [ + 0.216, + 0.147, + 0.603, + 0.162 + ], + "angle": 0, + "content": "4.2 Learning to Generate Images and Videos" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.177, + 0.788, + 0.282 + ], + "angle": 0, + "content": "Patchify. Following the original ViT [19], we \"patchify\" each latent frame independently by converting it into a sequence of non-overlapping \\( h_p \\times w_p \\) patches where \\( h_p = h/p \\), \\( w_p = w/p \\) and \\( p \\) is the patch size. We use learnable positional embeddings [74], which are the sum of space and time positional embeddings. Position embeddings are added to the linear projections [19] of the patches. Note that for images, we simply add the temporal position embedding corresponding to the first latent frame." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.283, + 0.789, + 0.48 + ], + "angle": 0, + "content": "Window attention. Transformer models composed entirely of global self-attention modules incur significant compute and memory costs, especially for video tasks. For efficiency and for processing images and videos jointly we compute self-attention in windows [29, 74], based on two types of non-overlapping configurations: spatial (S) and spatiotemporal (ST), cf. Fig. 2. Spatial Window (SW) attention is restricted to all the tokens within a latent frame of size \\(1 \\times h_p \\times w_p\\) (the first dimension is time). SW models the spatial relations in images and videos. Spatiotemporal Window (STW) attention is restricted within a 3D window of size \\((1 + t) \\times h_p' \\times h_w'\\), modeling the temporal relationships among video latent frames. For images, we simply use an identity attention mask ensuring that the value embeddings corresponding to the image frame latents are passed through the layer as is. Finally, in addition to absolute position embeddings we also use relative position embeddings [50]." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.481, + 0.79, + 0.587 + ], + "angle": 0, + "content": "Our design, while conceptually straightforward, achieves computational efficiency and enables joint training on image and video datasets. In contrast to methods based on frame-level autoencoders [5, 25, 29], our approach does not suffer from flickering artifacts, which often result from encoding and decoding video frames independently. However, similar to Blattmann et al. [5], we can also potentially leverage pre-trained image LDMs with transformer backbones by simply interleaving STW layers." + }, + { + "type": "title", + "bbox": [ + 0.216, + 0.613, + 0.458, + 0.627 + ], + "angle": 0, + "content": "4.3 Conditional Generation" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.643, + 0.788, + 0.718 + ], + "angle": 0, + "content": "To enable controllable video generation, in addition to conditioning on timestep \\( t \\), diffusion models are often conditioned on additional conditional information \\( \\mathbf{c} \\) such as class labels, natural language, past frames or low resolution videos. In our transformer backbone, we incorporate three types of conditioning mechanisms as described in what follows:" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.72, + 0.789, + 0.841 + ], + "angle": 0, + "content": "Cross-attention. In addition to self-attention layers in our window transformer blocks, we add a cross-attention layer for text conditioned generation. When training models on just videos, the cross-attention layer employs the same window-restricted attention as the self-attention layer, meaning S/ST blocks will have SW/STW cross-attention layers (Fig. 2). However, for joint training, we only use SW cross-attention layers. For cross-attention we concatenate the input signal (query) with the conditioning signal (key, value) as our early experiments showed this improves performance." + } + ], + [ + { + "type": "header", + "bbox": [ + 0.366, + 0.115, + 0.733, + 0.128 + ], + "angle": 0, + "content": "Photorealistic Video Generation with Diffusion Models" + }, + { + "type": "page_number", + "bbox": [ + 0.775, + 0.117, + 0.785, + 0.127 + ], + "angle": 0, + "content": "7" + }, + { + "type": "table", + "bbox": [ + 0.303, + 0.144, + 0.697, + 0.292 + ], + "angle": 0, + "content": "
MethodK600 FVD↓UCF FVD↓params.steps
TrIVD-GAN-FP [52]25.7±0.7--1
Video Diffusion [38]16.2±0.3-1.1B256
RIN [42]10.8-411M1000
TATS [24]-332±18321M1024
Phenaki [75]36.4±0.2-227M48
MAGVIT [82]9.9±0.376±2306M12
MAGVITv2 [83]4.3±0.158±2307M24
W.A.L.T-L Ours3.3±0.046±2313M50
W.A.L.T-XL Ours-36±2460M50
" + }, + { + "type": "table_caption", + "bbox": [ + 0.214, + 0.3, + 0.785, + 0.328 + ], + "angle": 0, + "content": "Table 1: Video generation evaluation on frame prediction on Kinetics-600 and class-conditional generation on UCF-101." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.377, + 0.788, + 0.59 + ], + "angle": 0, + "content": "AdaLN-LoRA. Adaptive normalization layers are an important component in a broad range of generative and visual synthesis models [17, 20, 45, 53-55]. A simple way to incorporate adaptive layer normalization is to include for each layer \\( i \\), an MLP layer to regress a vector of conditioning parameters \\( A^i = \\mathbf{MLP}(c + t) \\), where \\( A^i = \\text{concat}(\\gamma_1, \\gamma_2, \\beta_1, \\beta_2, \\alpha_1, \\alpha_2) \\), \\( A^i \\in \\mathbb{R}^{6 \\times d_{\\text{model}}} \\), and \\( c \\in \\mathbb{R}^{d_{\\text{model}}}, t \\in \\mathbb{R}^{d_{\\text{model}}} \\) are the condition and timestep embeddings. In the transformer block, \\( \\gamma \\) and \\( \\beta \\) scale and shift the inputs of the multi-head attention and MLP layers, respectively, while \\( \\alpha \\) scales the output of both the multi-head attention and MLP layers. The parameter count of these additional MLP layers scales linearly with the number of layers and quadratically with the model's dimensional size (num_blocks \\(\\times\\) \\( d_{\\text{model}} \\times 6 \\times d_{\\text{model}} \\)). For instance, in a ViT-g model with 1B parameters, the MLP layers contribute an additional 475M parameters. Inspired by [41], we propose a simple solution dubbed AdaLN-LoRA, to reduce the model parameters. For each layer, we regress conditioning parameters as" + }, + { + "type": "equation", + "bbox": [ + 0.293, + 0.605, + 0.786, + 0.623 + ], + "angle": 0, + "content": "\\[\nA ^ {1} = \\operatorname {M L P} (\\boldsymbol {c} + \\boldsymbol {t}), \\quad A ^ {i} = A ^ {1} + W _ {b} ^ {i} W _ {a} ^ {i} (\\boldsymbol {c} + \\boldsymbol {t}) \\quad \\forall i \\neq 1, \\tag {3}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.64, + 0.785, + 0.687 + ], + "angle": 0, + "content": "where \\( W_{b}^{i}\\in \\mathbb{R}^{d_{\\mathrm{model}}\\times r} \\), \\( W_{a}^{i}\\in \\mathbb{R}^{r\\times (6\\times d_{\\mathrm{model}})} \\). This reduces the number of trainable model parameters significantly when \\( r\\ll d_{\\mathrm{model}} \\). For example, a ViT-g model with \\( r = 2 \\) reduces the MLP parameters from 475M to 12M." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.689, + 0.788, + 0.841 + ], + "angle": 0, + "content": "Self-conditioning. In addition to being conditioned on external inputs, iterative generative algorithms can also be conditioned on their own previously generated samples during inference [3, 14, 66]. Specifically, Chen et al. [14] modify the training process for diffusion models, such that with some probability \\( p_{\\mathrm{sc}} \\) the model first generates a sample \\( \\tilde{\\pmb{z}}_0 = f_\\theta(\\pmb{z}_t; \\pmb{0}, \\pmb{c}, t) \\) and then refines this estimate using another forward pass conditioned on this initial sample: \\( f_\\theta(\\pmb{z}_t; \\text{stopgrad}(\\tilde{\\pmb{z}}_0), \\pmb{c}, t) \\). With probability \\( 1 - p_{\\mathrm{sc}} \\), only a single forward pass is done. We concatenate the model estimate with the input along the channel dimension and found this simple technique to work well when used in conjunction with \\( \\pmb{v} \\)-prediction." + } + ], + [ + { + "type": "page_number", + "bbox": [ + 0.218, + 0.116, + 0.23, + 0.127 + ], + "angle": 0, + "content": "8" + }, + { + "type": "header", + "bbox": [ + 0.272, + 0.115, + 0.377, + 0.129 + ], + "angle": 0, + "content": "A. Gupta et al." + }, + { + "type": "title", + "bbox": [ + 0.216, + 0.147, + 0.483, + 0.162 + ], + "angle": 0, + "content": "4.4 Autoregressive Generation" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.173, + 0.788, + 0.325 + ], + "angle": 0, + "content": "For generating long videos via autoregressive prediction we also train our model jointly on the task of frame prediction. This is achieved by conditioning the model on past frames with a probability of \\( p_{\\mathrm{fp}} \\) during training. Specifically, the model is conditioned using \\( c_{\\mathrm{fp}} = \\text{concat}(m_{\\mathrm{fp}} \\circ z_t, m_{\\mathrm{fp}}) \\), where \\( m_{\\mathrm{fp}} \\) is a binary mask. The binary mask indicates the number of past frames used for conditioning. We condition on either 1 latent frame (image to video generation) or 2 latent frames (video prediction). This conditioning is integrated into the model through concatenation along the channel dimension of the noisy latent input. During inference, we use standard classifier-free guidance with \\( c_{\\mathrm{fp}} \\) as the conditioning signal." + }, + { + "type": "title", + "bbox": [ + 0.216, + 0.347, + 0.46, + 0.362 + ], + "angle": 0, + "content": "4.5 Video Super Resolution" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.373, + 0.789, + 0.553 + ], + "angle": 0, + "content": "Generating high-resolution videos with a single model is computationally prohibitive. Following [37], we use a cascaded approach with three models operating at increasing resolutions. Our base model generates videos at \\(128 \\times 128\\) resolution which are subsequently upsampled twice via two super resolution stages. We first spatially upscale the low resolution input \\(z^{\\mathrm{lr}}\\) (video or image) using a depth-to-space convolution operation. Note that, unlike training where ground truth low-resolution inputs are available, inference relies on latents produced by preceding stages (cf. teaching-forcing). To reduce this discrepancy and improve the robustness of the super-resolution stages in handling artifacts generated by lower resolution stages, we use noise conditioning augmentation [37]. Concretely, noise is added in accordance with \\(\\gamma(t)\\), by sampling a noise level as \\(t_{\\mathrm{sr}} \\sim \\mathcal{U}(0, t_{\\max \\text{noise}})\\) and is provided as input to our AdaLN-LoRA layers." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.554, + 0.788, + 0.615 + ], + "angle": 0, + "content": "Aspect-ratio finetuning. To simplify training and leverage broad data sources with different aspect ratios, we train our base stage using a square aspect ratio. We fine-tune the base stage on a subset of data to generate videos with a 9:16 aspect ratio by interpolating position embeddings." + }, + { + "type": "title", + "bbox": [ + 0.216, + 0.638, + 0.377, + 0.656 + ], + "angle": 0, + "content": "5 Experiments" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.671, + 0.788, + 0.746 + ], + "angle": 0, + "content": "In this section, we evaluate our method on multiple tasks: class-conditional image and video generation, frame prediction and text conditioned video generation and perform extensive ablation studies of different design choices. For qualitative results, see Fig. 1, Fig. 3, Fig. 4 and videos on our project website. See appendix for additional details." + }, + { + "type": "title", + "bbox": [ + 0.216, + 0.769, + 0.414, + 0.784 + ], + "angle": 0, + "content": "5.1 Visual Generation" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.795, + 0.788, + 0.842 + ], + "angle": 0, + "content": "Video generation. We consider two standard video benchmarks, UCF-101 [71] for class-conditional generation and Kinetics-600 [10] for video prediction with 5 conditioning frames. We use FVD [72] as our primary evaluation metric." + } + ], + [ + { + "type": "header", + "bbox": [ + 0.366, + 0.115, + 0.732, + 0.128 + ], + "angle": 0, + "content": "Photorealistic Video Generation with Diffusion Models" + }, + { + "type": "page_number", + "bbox": [ + 0.775, + 0.117, + 0.785, + 0.127 + ], + "angle": 0, + "content": "9" + }, + { + "type": "table", + "bbox": [ + 0.298, + 0.144, + 0.697, + 0.285 + ], + "angle": 0, + "content": "
MethodCost (Iter×BS)FID↓IS↑params.steps
BigGAN-deep [7]-6.95171.4160M1
LDM-4 [62]178k×120010.56103.5400M250
DiT-XL/2 [54]7000k×2569.62121.5675M250
ADM [17]-7.49127.5608M2000
MDT [23]6500k×2566.23143.0676M250
MaskDiT [89]1200k×10245.69178.0736M40
RIN [42]600k×10243.42182.0410M1000
simple diffusion [40]500K×20482.77211.82B512
VDM++ [46]-2.40225.32B512
W.A.L.T-L Ours437k×10242.56215.1460M50
" + }, + { + "type": "table_caption", + "bbox": [ + 0.214, + 0.296, + 0.788, + 0.339 + ], + "angle": 0, + "content": "Table 2: Class-conditional image generation on ImageNet \\( {256} \\times {256} \\) . We adopt the evaluation protocol and implementation of ADM [17] and report results without classifier free guidance." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.381, + 0.787, + 0.426 + ], + "angle": 0, + "content": "Across both datasets, W.A.L.T significantly outperforms all prior works (Tab. 1). Compared to prior video diffusion models, we achieve state-of-the-art performance with less model parameters, and require 50 DDIM [69] inference steps." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.428, + 0.788, + 0.563 + ], + "angle": 0, + "content": "Image generation. To verify the modeling capabilities of \\( W.A.L.T \\) on the image domain, we train a version of \\( W.A.L.T \\) for the standard ImageNet class-conditional setting. For evaluation, we follow ADM [17] and report the FID [34] and Inception [64] scores calculated on 50K samples generated in 50 DDIM steps. We compare (Table 2) \\( W.A.L.T \\) with state-of-the-art image generation methods for \\( 256 \\times 256 \\) resolution. Our model outperforms prior works without requiring specialized schedules, convolution inductive bias, improved diffusion losses, and classifier free guidance. Although VDM++ [46] has slightly better FID score, the model has significantly more parameters (2B)." + }, + { + "type": "title", + "bbox": [ + 0.215, + 0.587, + 0.402, + 0.601 + ], + "angle": 0, + "content": "5.2 Ablation Studies" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.613, + 0.787, + 0.658 + ], + "angle": 0, + "content": "We ablate \\( W.A.L.T \\) to understand the contribution of various design decisions with the default settings: model L, patch size 1, \\( 1 \\times 16 \\times 16 \\) spatial window, \\( 5 \\times 8 \\times 8 \\) spatiotemporal window, \\( p_{\\mathrm{sc}} = 0.9 \\), \\( c = 8 \\) and \\( r = 2 \\)." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.659, + 0.787, + 0.719 + ], + "angle": 0, + "content": "Patch size. In various computer vision tasks utilizing ViT [19]-based models, a smaller patch size \\( p \\) has been shown to consistently enhance performance [9, 19, 30, 86]. Similarly, our findings also indicate that a reduced patch size improves performance (Table 3a)." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.72, + 0.787, + 0.78 + ], + "angle": 0, + "content": "Window attention. We compare three different STW window configurations with full self-attention (Table 3b). We find that local self-attention can achieve competitive (or better) performance while being significantly faster (up to \\(2 \\times\\)) and requiring less accelerator memory." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.78, + 0.787, + 0.841 + ], + "angle": 0, + "content": "Self-conditioning. In Table 3c we study the influence of varying the self-conditioning rate \\( p_{\\mathrm{sc}} \\) on generation quality. We notice a clear trend: increasing the self-conditioning rate from 0.0 (no self-conditioning) to 0.9 improves the FVD score substantially (44%)." + } + ], + [ + { + "type": "page_number", + "bbox": [ + 0.218, + 0.116, + 0.236, + 0.127 + ], + "angle": 0, + "content": "10" + }, + { + "type": "header", + "bbox": [ + 0.272, + 0.115, + 0.377, + 0.129 + ], + "angle": 0, + "content": "A. Gupta et al." + }, + { + "type": "table", + "bbox": [ + 0.288, + 0.164, + 0.465, + 0.228 + ], + "angle": 0, + "content": "
patch size pFVD↓IS↑
160.787.2
2134.482.2
4461.863.9
" + }, + { + "type": "table", + "bbox": [ + 0.521, + 0.143, + 0.744, + 0.228 + ], + "angle": 0, + "content": "
st windowFVD↓IS↑sps
5 × 4 × 456.987.32.24
5 × 8 × 859.687.42.00
5 × 16 × 1655.387.41.75
full self attn.59.987.81.20
" + }, + { + "type": "text", + "bbox": [ + 0.251, + 0.246, + 0.496, + 0.269 + ], + "angle": 0, + "content": "(a) Patch size. Lower patch size is significantly better." + }, + { + "type": "text", + "bbox": [ + 0.509, + 0.246, + 0.753, + 0.279 + ], + "angle": 0, + "content": "(b) Spatiotemporal window size. Full self-attention is not essential for good performance. sps is steps per sec." + }, + { + "type": "table", + "bbox": [ + 0.326, + 0.292, + 0.445, + 0.372 + ], + "angle": 0, + "content": "
pscFVD↓IS↑
0.0109.982.6
0.376.086.5
0.660.086.8
0.961.487.1
" + }, + { + "type": "table", + "bbox": [ + 0.534, + 0.292, + 0.714, + 0.385 + ], + "angle": 0, + "content": "
rFVD↓IS↑params
260.787.2313 M
456.687.3314 M
1655.588.0316 M
6454.487.9324 M
25652.588.5357 M
" + }, + { + "type": "text", + "bbox": [ + 0.259, + 0.388, + 0.733, + 0.41 + ], + "angle": 0, + "content": "(c) Self-conditioning. Higher \\( p_{\\mathrm{sc}} \\) is bet- (d) AdaLN-LoRA. Bigger \\( r \\) is better." + }, + { + "type": "table", + "bbox": [ + 0.269, + 0.423, + 0.485, + 0.509 + ], + "angle": 0, + "content": "
FVD↓IS↑
w/o qk norm [15]59.086.8
w/o latent norm67.987.1
w/o zero snr [49]91.084.2
full method60.787.2
" + }, + { + "type": "text", + "bbox": [ + 0.252, + 0.512, + 0.495, + 0.534 + ], + "angle": 0, + "content": "(e) Other improvements. See text for details." + }, + { + "type": "table", + "bbox": [ + 0.547, + 0.429, + 0.718, + 0.508 + ], + "angle": 0, + "content": "
crFVD↓FVD↓IS↑
437.786.484.9
817.175.486.3
168.267.086.0
323.583.482.9
" + }, + { + "type": "text", + "bbox": [ + 0.509, + 0.512, + 0.752, + 0.534 + ], + "angle": 0, + "content": "(f) Latent dimension \\(c\\). Higher \\(c\\) is better." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.545, + 0.788, + 0.59 + ], + "angle": 0, + "content": "Table 3: Ablation experiments on UCF-101 [71]. We compare FVD and inception scores to ablate important design decisions with the default setting: L model, \\(1 \\times 16 \\times 16\\) spatial window, \\(5 \\times 8 \\times 8\\) saptiotemporal (st) window, \\(p_{\\mathrm{sc}} = 0.9\\), \\(c = 8\\) and \\(r = 2\\)." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.629, + 0.789, + 0.841 + ], + "angle": 0, + "content": "AdaLN-LoRA. An important design decision in diffusion models is the conditioning mechanism. We investigate the effect of increasing the bottleneck dimension \\( r \\) in our proposed AdaLN-LoRA layers (Table 3d). This hyperparameter provides a flexible way to trade off between number of model parameters and generation performance. As shown in Table 3d, increasing \\( r \\) improves performance but also increases model parameters. This highlights an important model design question: given a fixed parameter budget, how should we allocate parameters - either by using separate AdaLN layers, or by increasing base model parameters while using shared AdaLN-LoRA layers? We explore this in Table 4 by comparing two model configurations: W.A.L.T-L with separate AdaLN layers and W.A.L.T-XL with AdaLN-LoRA and \\( r = 2 \\). While both configurations yield similar FVD and Inception scores, W.A.L.T-XL achieves a lower final loss value, suggesting the advantage of allocating more parameters to the base model and choosing an appropriate \\( r \\) value within accelerator memory limits." + } + ], + [ + { + "type": "header", + "bbox": [ + 0.366, + 0.114, + 0.732, + 0.128 + ], + "angle": 0, + "content": "Photorealistic Video Generation with Diffusion Models" + }, + { + "type": "page_number", + "bbox": [ + 0.768, + 0.116, + 0.784, + 0.127 + ], + "angle": 0, + "content": "11" + }, + { + "type": "table", + "bbox": [ + 0.313, + 0.144, + 0.688, + 0.205 + ], + "angle": 0, + "content": "
ModelAdaLNFVD↓IS↑params.final loss
Lseparate34.690.2458M0.274
XLLoRA-236.789.4460M0.268
" + }, + { + "type": "table_caption", + "bbox": [ + 0.214, + 0.223, + 0.785, + 0.25 + ], + "angle": 0, + "content": "Table 4: Parameter matched comparison between AdaLN-LoRA and per layer adaln layers. See text for details." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.295, + 0.789, + 0.462 + ], + "angle": 0, + "content": "Noise schedule. Common latent diffusion noise schedules [62] typically do not ensure a zero signal-to-noise ratio (SNR) at the final timestep, i.e., at \\( t = 1 \\), \\( \\gamma(t) > 0 \\). This leads to a mismatch between training and inference phases. During inference, models are expected to start from purely Gaussian noise, whereas during training, at \\( t = 1 \\), a small amount of signal information remains accessible to the model. This is especially harmful for video generation as videos have high temporal redundancy. Even minimal information leakage at \\( t = 1 \\) can reveal substantial information to the model. Addressing this mismatch by enforcing a zero terminal SNR [49] significantly improves performance (Table 3e). Note that this approach was originally proposed to fix over-exposure problems in image generation, but we find it effective for video generation as well." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.463, + 0.789, + 0.584 + ], + "angle": 0, + "content": "Autoencoder. Finally, we investigate one critical but often overlooked hyperparameter in the first stage of our model: the channel dimension \\( c \\) of the autoencoder latent \\( z \\). As shown in Table 3f, increasing \\( c \\) significantly improves the reconstruction quality (lower rFVD) while keeping the same spatial \\( f_{s} \\) and temporal compression \\( f_{t} \\) ratios. Empirically, we found that both lower and higher values of \\( c \\) lead to poor FVD scores in generation, with a sweet spot of \\( c = 8 \\) working well across most datasets and tasks we evaluated. We also normalize the latents before processing them via transformer which further improves performance." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.585, + 0.789, + 0.66 + ], + "angle": 0, + "content": "In our transformer models, we use query-key normalization [15] as it helps stabilize training for larger models. Finally, we note that some of our default settings are not optimal, as indicated by ablation studies. These defaults were chosen early on for their robustness across datasets, though further tuning may improve performance." + }, + { + "type": "title", + "bbox": [ + 0.215, + 0.688, + 0.377, + 0.703 + ], + "angle": 0, + "content": "5.3 Text-to-video" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.719, + 0.789, + 0.84 + ], + "angle": 0, + "content": "We train \\( W.A.L.T \\) for text-to-video jointly on text-image and text-video pairs (Sec. 4.2). We used a dataset of \\( \\sim 970\\mathrm{M} \\) text-image pairs and \\( \\sim 89\\mathrm{M} \\) text-video pairs from the public internet and internal sources. We train our base model at resolution \\( 17 \\times 128 \\times 128 \\) (3B parameters), and two \\( 2 \\times \\) cascaded super-resolution models for \\( 17 \\times 128 \\times 224 \\rightarrow 17 \\times 256 \\times 448 \\) (L, 1.3B, \\( p = 2 \\)) and \\( 17 \\times 256 \\times 448 \\rightarrow 17 \\times 512 \\times 896 \\) (L, 419M, \\( p = 2 \\)) respectively. We fine-tune the base stage for the \\( 9:16 \\) aspect ratio to generate videos at resolution \\( 128 \\times 224 \\). We use classifier free guidance for all our text-to-video results." + } + ], + [ + { + "type": "page_number", + "bbox": [ + 0.218, + 0.116, + 0.236, + 0.127 + ], + "angle": 0, + "content": "12" + }, + { + "type": "header", + "bbox": [ + 0.272, + 0.115, + 0.377, + 0.129 + ], + "angle": 0, + "content": "A. Gupta et al." + }, + { + "type": "image", + "bbox": [ + 0.219, + 0.206, + 0.784, + 0.259 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.446, + 0.261, + 0.555, + 0.272 + ], + "angle": 0, + "content": "A polar bear swimming." + }, + { + "type": "image", + "bbox": [ + 0.218, + 0.276, + 0.784, + 0.329 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.34, + 0.333, + 0.66, + 0.345 + ], + "angle": 0, + "content": "Pouring beer into an empty glass, low angle shot, bar in the background." + }, + { + "type": "image", + "bbox": [ + 0.218, + 0.35, + 0.784, + 0.403 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.377, + 0.406, + 0.623, + 0.417 + ], + "angle": 0, + "content": "A cat eating food out of a bowl, in the style of Van Gogh." + }, + { + "type": "image", + "bbox": [ + 0.218, + 0.421, + 0.784, + 0.473 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.34, + 0.478, + 0.66, + 0.489 + ], + "angle": 0, + "content": "Pouring chocolate sauce over vanilla ice cream in a cone, studio lighting." + }, + { + "type": "image", + "bbox": [ + 0.218, + 0.495, + 0.784, + 0.547 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.349, + 0.552, + 0.651, + 0.563 + ], + "angle": 0, + "content": "A robot ballerina dancing gracefully, highly detailed, studio lighting." + }, + { + "type": "image", + "bbox": [ + 0.218, + 0.569, + 0.784, + 0.621 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.263, + 0.627, + 0.736, + 0.638 + ], + "angle": 0, + "content": "A cute panda skateboarding in the sky, over snow covered mountains, with dreamy whimsical atmosphere." + }, + { + "type": "image", + "bbox": [ + 0.218, + 0.645, + 0.784, + 0.697 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.437, + 0.703, + 0.564, + 0.714 + ], + "angle": 0, + "content": "An astronaut riding a horse." + }, + { + "type": "image_caption", + "bbox": [ + 0.214, + 0.73, + 0.788, + 0.787 + ], + "angle": 0, + "content": "Fig. 3: Qualitative evaluation. Example videos generated by \\( W.A.L.T \\) from natural language prompts at \\( 512 \\times 896 \\) resolution over 3.6 seconds duration at 8 frames per second. The \\( W.A.L.T \\) model is able to generate temporally consistent photorealistic videos that align with the textual prompt. There are more results on our project website." + } + ], + [ + { + "type": "header", + "bbox": [ + 0.366, + 0.114, + 0.732, + 0.128 + ], + "angle": 0, + "content": "Photorealistic Video Generation with Diffusion Models" + }, + { + "type": "page_number", + "bbox": [ + 0.769, + 0.116, + 0.785, + 0.127 + ], + "angle": 0, + "content": "13" + }, + { + "type": "table", + "bbox": [ + 0.328, + 0.144, + 0.668, + 0.322 + ], + "angle": 0, + "content": "
MethodIS (↑)FVD (↓)
CogVideo (Chinese) [39]23.6751.3
CogVideo (English) [39]25.3701.6
MagicVideo [90]-699.0
EMU Video* [26]42.7606.2
Video LDM [5]33.5550.6
Make-A-Video [67]33.0367.2
PYoCo [25]47.8355.2
PixelDance* [85]42.1242.8
SVD* [4]-242.0
W.A.L.T (Ours) 419M (video only)26.8598.8
W.A.L.T (Ours) 419M (video + image)31.7344.5
W.A.L.T (Ours) 3B (video + image)35.1258.1
" + }, + { + "type": "table_caption", + "bbox": [ + 0.214, + 0.332, + 0.788, + 0.375 + ], + "angle": 0, + "content": "Table 5: UCF-101 text-to-video generation. Joint training on image and video datasets in conjunction with scaling the model parameters is essential for high quality video generation. * Denotes concurrent work." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.417, + 0.787, + 0.522 + ], + "angle": 0, + "content": "Quantitative Evaluation Evaluating text-conditioned video generation scientifically remains a significant challenge, in part due to the absence of standardized training datasets and benchmarks. So far we have focused our experiments and analyses on the standard academic benchmarks, which use the same training data to ensure controlled and fair comparisons. Nevertheless, to compare with prior work on text-to-video, we also report results on the UCF-101 dataset in the zero-shot evaluation protocol in Table 5 [25,39,67]. Also see supplement." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.524, + 0.788, + 0.629 + ], + "angle": 0, + "content": "Joint training. A primary strength of our framework is its ability to train simultaneously on both image and video datasets. In Table 5 we ablate the impact of this joint training. Specifically, we trained two versions of W.A.L.T-L (each with 419M params.) models using the default settings specified in § 5.2. We find that joint training leads to a notable improvement across both metrics. Our results align with the findings of Ho et al. [38], who demonstrated the benefits of joint training for pixel-based video diffusion models with U-Net backbones." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.63, + 0.788, + 0.734 + ], + "angle": 0, + "content": "Scaling. Transformers are known for their ability to scale effectively in many tasks [6,15,58]. In Table 5 we show the benefits of scaling our transformer model for video diffusion. Scaling our base model size leads to significant improvements on both the metrics. It is important to note, however, that our base model is considerably smaller than leading text-to-video systems. For instance, Ho et al. [35] trained base model of 5.7B parameters. Hence, we believe scaling our models further is an important direction of future work." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.735, + 0.788, + 0.841 + ], + "angle": 0, + "content": "Comparison with prior work. In Table 5, we present a system-level comparison of various text-to-video methods. Our results are promising, surpassing all previous work in the FVD metric, and being competitive with concurrent work. In terms of the IS, W.A.L.T outperforms all previous work but PYoCo [25]. A possible explanation for this discrepancy might be PYoCo's use of stronger text embeddings. Specifically, they utilize both CLIP [56] and T5-XXL [61] encoders, whereas we employ a T5-XL [61] text encoder only." + } + ], + [ + { + "type": "page_number", + "bbox": [ + 0.218, + 0.116, + 0.236, + 0.127 + ], + "angle": 0, + "content": "14" + }, + { + "type": "header", + "bbox": [ + 0.272, + 0.115, + 0.377, + 0.129 + ], + "angle": 0, + "content": "A. Gupta et al." + }, + { + "type": "image", + "bbox": [ + 0.219, + 0.145, + 0.784, + 0.44 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.214, + 0.454, + 0.788, + 0.496 + ], + "angle": 0, + "content": "Fig. 4: Samples at \\(512 \\times 896\\) resolution. Top: Text-to-video samples of consistent 3D camera motion (5.1 secs). Bottom: Image-to-video samples (3.8 secs). Images are generated using off-shelf models. Visit our project website to view them in video format." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.544, + 0.788, + 0.605 + ], + "angle": 0, + "content": "Qualitative Results As mentioned in § 4.4, we jointly train our model on the task of frame prediction conditioned on 1 or 2 latent frames. Hence, our model can be used for animating images (image-to-video) and generating longer videos with consistent camera motion (Fig. 4). See videos on our project website." + }, + { + "type": "title", + "bbox": [ + 0.216, + 0.628, + 0.36, + 0.644 + ], + "angle": 0, + "content": "6 Conclusion" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.659, + 0.789, + 0.841 + ], + "angle": 0, + "content": "In this work, we introduce W.A.L.T, a simple, scalable, and efficient transformer-based framework for latent video diffusion models. A key design decision in our framework is to operate on a shared, compressed visual representation that enables generative modeling of both images and videos. We demonstrate state-of-the-art results for image and video generation using a transformer backbone with windowed attention. We also train a cascade of three W.A.L.T models jointly on image and video datasets, to synthesize high-resolution, temporally consistent photorealistic videos from natural language descriptions. While generative modeling has seen tremendous recent advances for images, progress on video generation has lagged behind. We hope that scaling our framework will enable the creation of powerful video generative models which can understand and model the physical world." + } + ], + [ + { + "type": "header", + "bbox": [ + 0.366, + 0.115, + 0.732, + 0.128 + ], + "angle": 0, + "content": "Photorealistic Video Generation with Diffusion Models" + }, + { + "type": "page_number", + "bbox": [ + 0.769, + 0.117, + 0.785, + 0.127 + ], + "angle": 0, + "content": "15" + }, + { + "type": "title", + "bbox": [ + 0.218, + 0.145, + 0.403, + 0.163 + ], + "angle": 0, + "content": "Acknowledgements" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.174, + 0.787, + 0.219 + ], + "angle": 0, + "content": "We thank Bryan Seybold, Dan Kondratyuk, David Ross, Hartwig Adam, Huisheng Wang, Jason Baldridge, Mauricio Delbracio and Orly Liba for helpful discussions and feedback." + }, + { + "type": "title", + "bbox": [ + 0.218, + 0.241, + 0.323, + 0.256 + ], + "angle": 0, + "content": "References" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.268, + 0.787, + 0.31 + ], + "angle": 0, + "content": "1. Agostinelli, A., Denk, T.I., Borsos, Z., Engel, J., Verzetti, M., Caillon, A., Huang, Q., Jansen, A., Roberts, A., Tagliasacchi, M., et al.: MusicLM: Generating music from text. arXiv:2301.11325 (2023) 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.31, + 0.787, + 0.337 + ], + "angle": 0, + "content": "2. Bao, F., Li, C., Cao, Y., Zhu, J.: All are worth words: a vit backbone for score-based diffusion models. In: NeurIPS 2022 Workshop on Score-Based Methods (2022) 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.337, + 0.787, + 0.378 + ], + "angle": 0, + "content": "3. Bengio, S., Vinyals, O., Jaitly, N., Shazeer, N.: Scheduled sampling for sequence prediction with recurrent neural networks. Advances in neural information processing systems 28 (2015) 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.378, + 0.787, + 0.432 + ], + "angle": 0, + "content": "4. Blattmann, A., Dockhorn, T., Kulal, S., Mendelevitch, D., Kilian, M., Lorenz, D., Levi, Y., English, Z., Voleti, V., Letts, A., et al.: Stable video diffusion: Scaling latent video diffusion models to large datasets. arXiv preprint arXiv:2311.15127 (2023) 13" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.433, + 0.787, + 0.472 + ], + "angle": 0, + "content": "5. Blattmann, A., Rombach, R., Ling, H., Dockhorn, T., Kim, S.W., Fidler, S., Kreis, K.: Align your latents: High-resolution video synthesis with latent diffusion models. In: CVPR (2023) 3, 4, 6, 13" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.473, + 0.787, + 0.514 + ], + "angle": 0, + "content": "6. Bousmalis, K., Vezzani, G., Rao, D., Devin, C., Lee, A.X., Bauza, M., Davchev, T., Zhou, Y., Gupta, A., Raju, A., et al.: Robocat: A self-improving foundation agent for robotic manipulation. arXiv preprint arXiv:2306.11706 (2023) 2, 13" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.514, + 0.787, + 0.542 + ], + "angle": 0, + "content": "7. Brock, A., Donahue, J., Simonyan, K.: Large scale GAN training for high fidelity natural image synthesis. In: ICLR (2018) 9" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.542, + 0.787, + 0.582 + ], + "angle": 0, + "content": "8. Brohan, A., Brown, N., Carbajal, J., Chebotar, Y., Dabis, J., Finn, C., Gopalakrishnan, K., Hausman, K., Herzog, A., Hsu, J., et al.: Rt-1: Robotics transformer for real-world control at scale. arXiv preprint arXiv:2212.06817 (2022) 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.582, + 0.787, + 0.623 + ], + "angle": 0, + "content": "9. Caron, M., Touvron, H., Misra, I., Jégou, H., Mairal, J., Bojanowski, P., Joulin, A.: Emerging properties in self-supervised vision transformers. In: Proceedings of the IEEE/CVF international conference on computer vision. pp. 9650-9660 (2021) 9" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.623, + 0.787, + 0.65 + ], + "angle": 0, + "content": "0. Carreira, J., Noland, E., Banki-Horvath, A., Hillier, C., Zisserman, A.: A short note about Kinetics-600. arXiv:1808.01340 (2018) 3, 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.65, + 0.787, + 0.69 + ], + "angle": 0, + "content": "1. Chang, H., Zhang, H., Barber, J., Maschinot, A., Lezama, J., Jiang, L., Yang, M.H., Murphy, K., Freeman, W.T., Rubinstein, M., et al.: Muse: Text-to-image generation via masked generative transformers. In: ICML (2023) 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.691, + 0.787, + 0.718 + ], + "angle": 0, + "content": "2. Chang, H., Zhang, H., Jiang, L., Liu, C., Freeman, W.T.: MaskGIT: Masked generative image transformer. In: CVPR (2022) 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.718, + 0.787, + 0.759 + ], + "angle": 0, + "content": "3. Chen, J., Yu, J., Ge, C., Yao, L., Xie, E., Wu, Y., Wang, Z., Kwok, J., Luo, P., Lu, H., et al.: Pixart-\\(\\alpha\\)beta: Fast training of diffusion transformer for photorealistic text-to-image synthesis. arXiv preprint arXiv:2310.00426 (2023) 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.759, + 0.787, + 0.785 + ], + "angle": 0, + "content": "4. Chen, T., Zhang, R., Hinton, G.: Analog bits: Generating discrete data using diffusion models with self-conditioning. arXiv preprint arXiv:2208.04202 (2022) 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.785, + 0.787, + 0.84 + ], + "angle": 0, + "content": "5. Dehghani, M., Djolonga, J., Mustafa, B., Padlewski, P., Heek, J., Gilmer, J., Steiner, A.P., Caron, M., Geirhos, R., Alabdulmohsin, I., et al.: Scaling vision transformers to 22 billion parameters. In: International Conference on Machine Learning. pp. 7480-7512. PMLR (2023) 10, 11, 13" + }, + { + "type": "list", + "bbox": [ + 0.226, + 0.268, + 0.787, + 0.84 + ], + "angle": 0, + "content": null + } + ], + [ + { + "type": "page_number", + "bbox": [ + 0.218, + 0.116, + 0.236, + 0.127 + ], + "angle": 0, + "content": "16" + }, + { + "type": "header", + "bbox": [ + 0.272, + 0.115, + 0.375, + 0.129 + ], + "angle": 0, + "content": "A. Gupta et al." + }, + { + "type": "ref_text", + "bbox": [ + 0.218, + 0.147, + 0.785, + 0.175 + ], + "angle": 0, + "content": "16. Deng, J., Dong, W., Socher, R., Li, L.J., Li, K., Fei-Fei, L.: ImageNet: A large-scale hierarchical image database. In: CVPR (2009) 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.218, + 0.176, + 0.787, + 0.202 + ], + "angle": 0, + "content": "17. Dhariwal, P., Nichol, A.: Diffusion models beat GANs on image synthesis. In: NeurIPS (2021) 2, 7, 9" + }, + { + "type": "ref_text", + "bbox": [ + 0.218, + 0.203, + 0.787, + 0.243 + ], + "angle": 0, + "content": "18. Ding, M., Yang, Z., Hong, W., Zheng, W., Zhou, C., Yin, D., Lin, J., Zou, X., Shao, Z., Yang, H., et al.:CogView: Mastering text-to-image generation via transformers. In:NeurIPS (2021) 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.244, + 0.787, + 0.284 + ], + "angle": 0, + "content": "19. Dosovitskiy, A., Beyer, L., Kolesnikov, A., Weissenborn, D., Zhai, X., Unterthiner, T., Dehghani, M., Minderer, M., Heigold, G., Gelly, S., et al.: An image is worth 16x16 words: Transformers for image recognition at scale. In: ICLR (2020) 2, 6, 9" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.285, + 0.787, + 0.311 + ], + "angle": 0, + "content": "20. Dumoulin, V., Shlens, J., Kudlur, M.: A learned representation for artistic style. arXiv preprint arXiv:1610.07629 (2016) 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.312, + 0.787, + 0.338 + ], + "angle": 0, + "content": "21. Esser, P., Rombach, R., Ommer, B.: Taming transformers for high-resolution image synthesis. In: CVPR (2021) 2, 4, 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.339, + 0.787, + 0.379 + ], + "angle": 0, + "content": "22. Gafni, O., Polyak, A., Ashual, O., Sheynin, S., Parikh, D., Taigman, Y.: Makea-scene: Scene-based text-to-image generation with human priors. In: European Conference on Computer Vision. pp. 89-106. Springer (2022) 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.379, + 0.787, + 0.405 + ], + "angle": 0, + "content": "23. Gao, S., Zhou, P., Cheng, M.M., Yan, S.: Masked diffusion transformer is a strong image synthesizer. arXiv:2303.14389 (2023) 4, 9" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.407, + 0.787, + 0.446 + ], + "angle": 0, + "content": "24. Ge, S., Hayes, T., Yang, H., Yin, X., Pang, G., Jacobs, D., Huang, J.B., Parikh, D.: Long video generation with time-agnostic VQGAN and time-sensitive transformer. In: ECCV (2022) 4, 5, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.447, + 0.787, + 0.488 + ], + "angle": 0, + "content": "25. Ge, S., Nah, S., Liu, G., Poon, T., Tao, A., Catanzaro, B., Jacobs, D., Huang, J.B., Liu, M.Y., Balaji, Y.: Preserve your own correlation: A noise prior for video diffusion models. arXiv preprint arXiv:2305.10474 (2023) 3, 6, 13" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.488, + 0.787, + 0.528 + ], + "angle": 0, + "content": "26. Girdhar, R., Singh, M., Brown, A., Duval, Q., Azadi, S., Rambhatla, S.S., Shah, A., Yin, X., Parikh, D., Misra, I.: Emu video: Factorizing text-to-video generation by explicit image conditioning. arXiv preprint arXiv:2311.10709 (2023) 13" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.529, + 0.787, + 0.569 + ], + "angle": 0, + "content": "27. Goodfellow, I., Pouget-Abadie, J., Mirza, M., Xu, B., Warde-Farley, D., Ozair, S., Courville, A., Bengio, Y.: Generative adversarial nets. Advances in neural information processing systems 27 (2014) 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.57, + 0.65, + 0.582 + ], + "angle": 0, + "content": "28. Google: PaLM 2 technical report. arXiv:2305.10403 (2023) 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.583, + 0.787, + 0.609 + ], + "angle": 0, + "content": "29. Gupta, A., Tian, S., Zhang, Y., Wu, J., Martín-Martín, R., Fei-Fei, L.: MaskViT: Masked visual pre-training for video prediction. In: ICLR (2022) 4, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.61, + 0.787, + 0.637 + ], + "angle": 0, + "content": "30. Gupta, A., Wu, J., Deng, J., Fei-Fei, L.: Siamese masked autoencoders. arXiv preprint arXiv:2305.14344 (2023) 9" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.637, + 0.787, + 0.677 + ], + "angle": 0, + "content": "31. Harvey, W., Naderiparizi, S., Masrani, V., Weilbach, C., Wood, F.: Flexible diffusion modeling of long videos. Advances in Neural Information Processing Systems 35, 27953-27965 (2022) 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.678, + 0.787, + 0.704 + ], + "angle": 0, + "content": "32. He, K., Chen, X., Xie, S., Li, Y., Dollar, P., Girshick, R.: Masked autoencoders are scalable vision learners. In: CVPR (2022) 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.705, + 0.787, + 0.731 + ], + "angle": 0, + "content": "33. He, Y., Yang, T., Zhang, Y., Shan, Y., Chen, Q.: Latent video diffusion models for high-fidelity long video generation. arXiv preprint arXiv:2211.13221 (2023) 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.732, + 0.787, + 0.771 + ], + "angle": 0, + "content": "34. Heusel, M., Ramsauer, H., Unterthiner, T., Nessler, B., Hochreiter, S.: GANs trained by a two time-scale update rule converge to a local nash equilibrium. In: NeurIPS (2017) 9" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.772, + 0.787, + 0.813 + ], + "angle": 0, + "content": "35. Ho, J., Chan, W., Sahara, C., Whang, J., Gao, R., Gritsenko, A., Kingma, D.P., Poole, B., Norouzi, M., Fleet, D.J., et al.: Imagen video: High definition video generation with diffusion models. arXiv:2210.02303 (2022) 3, 4, 5, 13" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.814, + 0.787, + 0.84 + ], + "angle": 0, + "content": "36. Ho, J., Jain, A., Abbeel, P.: Denoising diffusion probabilistic models. In: NeurIPS (2020) 2, 3, 4" + }, + { + "type": "list", + "bbox": [ + 0.217, + 0.147, + 0.787, + 0.84 + ], + "angle": 0, + "content": null + } + ], + [ + { + "type": "header", + "bbox": [ + 0.366, + 0.115, + 0.732, + 0.127 + ], + "angle": 0, + "content": "Photorealistic Video Generation with Diffusion Models" + }, + { + "type": "page_number", + "bbox": [ + 0.768, + 0.117, + 0.785, + 0.127 + ], + "angle": 0, + "content": "17" + }, + { + "type": "ref_text", + "bbox": [ + 0.214, + 0.147, + 0.788, + 0.188 + ], + "angle": 0, + "content": "37. Ho, J., Sahara, C., Chan, W., Fleet, D.J., Norouzi, M., Salimans, T.: Cascaded diffusion models for high fidelity image generation. JMLR 23(1), 2249-2281 (2022) 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.214, + 0.189, + 0.788, + 0.217 + ], + "angle": 0, + "content": "38. Ho, J., Salimans, T., Gritsenko, A., Chan, W., Norouzi, M., Fleet, D.J.: Video diffusion models. In: ICLR Workshops (2022) 2, 3, 4, 7, 13" + }, + { + "type": "ref_text", + "bbox": [ + 0.214, + 0.218, + 0.788, + 0.243 + ], + "angle": 0, + "content": "39. Hong, W., Ding, M., Zheng, W., Liu, X., Tang, J.: CogVideo: Large-scale pretraining for text-to-video generation via transformers. arXiv:2205.15868 (2022) 4, 13" + }, + { + "type": "ref_text", + "bbox": [ + 0.214, + 0.244, + 0.788, + 0.27 + ], + "angle": 0, + "content": "40. Hoogeboom, E., Heek, J., Salimans, T.: simple diffusion: End-to-end diffusion for high resolution images. In: ICML (2023) 3, 4, 9" + }, + { + "type": "ref_text", + "bbox": [ + 0.214, + 0.27, + 0.788, + 0.298 + ], + "angle": 0, + "content": "41. Hu, E.J., Shen, Y., Wallis, P., Allen-Zhu, Z., Li, Y., Wang, S., Wang, L., Chen, W.: Lora: Low-rank adaptation of large language models. In: ICLR (2021) 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.214, + 0.298, + 0.788, + 0.324 + ], + "angle": 0, + "content": "42. Jabri, A., Fleet, D.J., Chen, T.: Scalable adaptive computation for iterative generation. In: ICML (2023) 4, 7, 9" + }, + { + "type": "ref_text", + "bbox": [ + 0.214, + 0.325, + 0.788, + 0.365 + ], + "angle": 0, + "content": "43. Jiang, Y., Chang, S., Wang, Z.: Transgan: Two pure transformers can make one strong gan, and that can scale up. Advances in Neural Information Processing Systems 34, 14745-14758 (2021) 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.214, + 0.365, + 0.788, + 0.392 + ], + "angle": 0, + "content": "44. Johnson, J., Alahi, A., Fei-Fei, L.: Perceptual losses for real-time style transfer and super-resolution. In: ECCV (2016) 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.214, + 0.392, + 0.788, + 0.419 + ], + "angle": 0, + "content": "45. Karras, T., Laine, S., Aila, T.: A style-based generator architecture for generative adversarial networks. In: CVPR (2019) 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.214, + 0.42, + 0.788, + 0.446 + ], + "angle": 0, + "content": "46. Kingma, D.P., Gao, R.: Understanding the diffusion objective as a weighted integral of ellbos. arXiv:2303.00848 (2023) 9" + }, + { + "type": "ref_text", + "bbox": [ + 0.214, + 0.446, + 0.788, + 0.473 + ], + "angle": 0, + "content": "47. LeCun, Y., Bottou, L., Bengio, Y., Haffner, P.: Gradient-based learning applied to document recognition. Proceedings of the IEEE 86(11), 2278-2324 (1998) 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.214, + 0.474, + 0.788, + 0.5 + ], + "angle": 0, + "content": "48. Lee, K., Chang, H., Jiang, L., Zhang, H., Tu, Z., Liu, C.: Vitgan: Training gans with vision transformers. arXiv preprint arXiv:2107.04589 (2021) 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.214, + 0.501, + 0.788, + 0.527 + ], + "angle": 0, + "content": "49. Lin, S., Liu, B., Li, J., Yang, X.: Common diffusion noise schedules and sample steps are flawed. arXiv preprint arXiv:2305.08891 (2023) 10, 11" + }, + { + "type": "ref_text", + "bbox": [ + 0.214, + 0.527, + 0.788, + 0.581 + ], + "angle": 0, + "content": "50. Liu, Z., Lin, Y., Cao, Y., Hu, H., Wei, Y., Zhang, Z., Lin, S., Guo, B.: Swin transformer: Hierarchical vision transformer using shifted windows. In: Proceedings of the IEEE/CVF international conference on computer vision. pp. 10012-10022 (2021) 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.214, + 0.582, + 0.788, + 0.622 + ], + "angle": 0, + "content": "51. Lu, H., Yang, G., Fei, N., Huo, Y., Lu, Z., Luo, P., Ding, M.: Vdt: General-purpose video diffusion transformers via mask modeling. arXiv preprint arXiv:2305.13311 (2023) 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.214, + 0.622, + 0.788, + 0.663 + ], + "angle": 0, + "content": "52. Luc, P., Clark, A., Dieleman, S., Casas, D.d.L., Doron, Y., Cassirer, A., Simonyan, K.: Transformation-based adversarial video prediction on large-scale data. arXiv:2003.04035 (2020) 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.214, + 0.663, + 0.788, + 0.691 + ], + "angle": 0, + "content": "53. Nichol, A.Q., Dhariwal, P.: Improved denoising diffusion probabilistic models. In: International Conference on Machine Learning. pp. 8162-8171. PMLR (2021) 3, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.214, + 0.691, + 0.788, + 0.718 + ], + "angle": 0, + "content": "54. Peebles, W., Xie, S.: Scalable diffusion models with transformers. arXiv:2212.09748 (2022) 4, 7, 9" + }, + { + "type": "ref_text", + "bbox": [ + 0.214, + 0.718, + 0.788, + 0.758 + ], + "angle": 0, + "content": "55. Perez, E., Strub, F., De Vries, H., Dumoulin, V., Courville, A.: Film: Visual reasoning with a general conditioning layer. In: Proceedings of the AAAI conference on artificial intelligence. vol. 32 (2018) 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.214, + 0.759, + 0.788, + 0.799 + ], + "angle": 0, + "content": "56. Radford, A., Kim, J.W., Hallacy, C., Ramesh, A., Goh, G., Agarwal, S., Sastry, G., Askell, A., Mishkin, P., Clark, J., et al.: Learning transferable visual models from natural language supervision. In: ICML (2021) 2, 4, 13" + }, + { + "type": "ref_text", + "bbox": [ + 0.214, + 0.799, + 0.788, + 0.84 + ], + "angle": 0, + "content": "57. Radford, A., Kim, J.W., Xu, T., Brockman, G., McLeavey, C., Sutskever, I.: Robust speech recognition via large-scale weak supervision. In: International Conference on Machine Learning. pp. 28492-28518. PMLR (2023) 2" + }, + { + "type": "list", + "bbox": [ + 0.214, + 0.147, + 0.788, + 0.84 + ], + "angle": 0, + "content": null + } + ], + [ + { + "type": "page_number", + "bbox": [ + 0.218, + 0.116, + 0.236, + 0.127 + ], + "angle": 0, + "content": "18" + }, + { + "type": "header", + "bbox": [ + 0.272, + 0.115, + 0.375, + 0.129 + ], + "angle": 0, + "content": "A. Gupta et al." + }, + { + "type": "ref_text", + "bbox": [ + 0.218, + 0.148, + 0.785, + 0.175 + ], + "angle": 0, + "content": "58. Radford, A., Narasimhan, K., Salimans, T., Sutskever, I., et al.: Improving language understanding by generative pre-training (2018) 2, 4, 13" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.177, + 0.785, + 0.203 + ], + "angle": 0, + "content": "59. Radford, A., Wu, J., Child, R., Luan, D., Amodei, D., Sutskever, I., et al.: Language models are unsupervised multitask learners. OpenAI blog 1(8), 9 (2019) 2, 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.218, + 0.204, + 0.787, + 0.231 + ], + "angle": 0, + "content": "60. Ramesh, A., Pavlov, M., Goh, G., Gray, S., Voss, C., Radford, A., Chen, M., Sutskever, I.: Zero-shot text-to-image generation. In: ICML (2021) 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.232, + 0.786, + 0.272 + ], + "angle": 0, + "content": "61. Roberts, A., Raffel, C., Lee, K., Matena, M., Shazeer, N., Liu, P.J., Narang, S., Li, W., Zhou, Y.: Exploring the limits of transfer learning with a unified text-to-text transformer (2019) 13" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.273, + 0.785, + 0.3 + ], + "angle": 0, + "content": "62. Rombach, R., Blattmann, A., Lorenz, D., Esser, P., Ommer, B.: High-resolution image synthesis with latent diffusion models. In: CVPR (2022) 2, 3, 9, 11" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.301, + 0.785, + 0.327 + ], + "angle": 0, + "content": "63. Ronneberger, O., Fischer, P., Brox, T.: U-Net: Convolutional networks for biomedical image segmentation. In: MICCAI (2015) 2, 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.328, + 0.785, + 0.355 + ], + "angle": 0, + "content": "64. Salimans, T., Goodfellow, I., Zaremba, W., Cheung, V., Radford, A., Chen, X.: Improved techniques for training GANs. In: NeurIPS (2016) 9" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.357, + 0.785, + 0.383 + ], + "angle": 0, + "content": "65. Salimans, T., Ho, J.: Progressive distillation for fast sampling of diffusion models arXiv preprint arXiv:2202.00512 (2022) 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.384, + 0.785, + 0.422 + ], + "angle": 0, + "content": "66. Savinov, N., Chung, J., Binkowski, M., Elsen, E., Oord, A.v.d.: Step-unrolled denoising autoencoders for text generation. arXiv preprint arXiv:2112.06749 (2021) 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.424, + 0.785, + 0.466 + ], + "angle": 0, + "content": "67. Singer, U., Polyak, A., Hayes, T., Yin, X., An, J., Zhang, S., Hu, Q., Yang, H., Ashual, O., Gafni, O., et al.: Make-a-video: Text-to-video generation without text-video data. arXiv:2209.14792 (2022) 3, 4, 13" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.467, + 0.785, + 0.494 + ], + "angle": 0, + "content": "68. Sohl-Dickstein, J., Weiss, E., Maheswaranathan, N., Ganguli, S.: Deep unsupervised learning using nonequilibrium thermodynamics. In: ICML (2015) 2, 3, 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.495, + 0.785, + 0.522 + ], + "angle": 0, + "content": "69. Song, J., Meng, C., Ermon, S.: Denoising diffusion implicit models. arXiv preprint arXiv:2010.02502 (2020) 3, 9" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.523, + 0.785, + 0.549 + ], + "angle": 0, + "content": "70. Song, Y., Ermon, S.: Generative modeling by estimating gradients of the data distribution. In: NeurIPS (2019) 2, 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.55, + 0.785, + 0.577 + ], + "angle": 0, + "content": "71. Soomro, K., Zamir, A.R., Shah, M.: UCF101: A dataset of 101 human actions classes from videos in the wild. arXiv:1212.0402 (2012) 3, 8, 10" + }, + { + "type": "ref_text", + "bbox": [ + 0.218, + 0.578, + 0.787, + 0.618 + ], + "angle": 0, + "content": "72. Unterthiner, T., van Steenkiste, S., Kurach, K., Marinier, R., Michalski, M., Gelly, S.: Towards accurate generative models of video: A new metric & challenges. arXiv:1812.01717 (2018) 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.62, + 0.785, + 0.646 + ], + "angle": 0, + "content": "73. Van Den Oord, A., Vinyals, O., et al.: Neural discrete representation learning. In: NeurIPS (2017) 2, 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.647, + 0.785, + 0.674 + ], + "angle": 0, + "content": "74. Vaswani, A., Shazeer, N., Parmar, N., Uszkoreit, J., Jones, L., Gomez, A.N., Kaiser, L., Polosukhin, I.: Attention is all you need. In: NeurIPS (2017) 2, 4, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.675, + 0.785, + 0.715 + ], + "angle": 0, + "content": "75. Villegas, R., Babaeizadeh, M., Kindermans, P.J., Moraldo, H., Zhang, H., Saffar, M.T., Castro, S., Kunze, J., Erhan, D.: Phenaki: Variable length video generation from open domain textual description. arXiv:2210.02399 (2022) 4, 5, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.716, + 0.785, + 0.757 + ], + "angle": 0, + "content": "76. Vincent, P., Larochelle, H., Bengio, Y., Manzagol, P.A.: Extracting and composing robust features with denoising autoencoders. In: Proceedings of the 25th international conference on Machine learning. pp. 1096-1103 (2008) 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.758, + 0.785, + 0.798 + ], + "angle": 0, + "content": "77. Wu, C., Huang, L., Zhang, Q., Li, B., Ji, L., Yang, F., Sapiro, G., Duan, N.: Godiva: Generating open-domain videos from natural descriptions. arXiv preprint arXiv:2104.14806 (2021) 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.799, + 0.785, + 0.84 + ], + "angle": 0, + "content": "78. Wu, C., Liang, J., Ji, L., Yang, F., Fang, Y., Jiang, D., Duan, N.: Niwa: Visual synthesis pre-training for neural visual world creation. In: European conference on computer vision. pp. 720-736. Springer (2022) 4" + }, + { + "type": "list", + "bbox": [ + 0.217, + 0.148, + 0.787, + 0.84 + ], + "angle": 0, + "content": null + } + ], + [ + { + "type": "header", + "bbox": [ + 0.366, + 0.115, + 0.732, + 0.128 + ], + "angle": 0, + "content": "Photorealistic Video Generation with Diffusion Models" + }, + { + "type": "page_number", + "bbox": [ + 0.769, + 0.117, + 0.785, + 0.127 + ], + "angle": 0, + "content": "19" + }, + { + "type": "ref_text", + "bbox": [ + 0.214, + 0.147, + 0.788, + 0.175 + ], + "angle": 0, + "content": "79. Yan, W., Zhang, Y., Abbeel, P., Srinivas, A.: VideoGPT: Video generation using VQ-VAE and transformers. arXiv:2104.10157 (2021) 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.214, + 0.175, + 0.788, + 0.217 + ], + "angle": 0, + "content": "80. Yu, J., Li, X., Koh, J.Y., Zhang, H., Pang, R., Qin, J., Ku, A., Xu, Y., Baldridge, J., Wu, Y.: Vector-quantized image modeling with improved VQGAN. In: ICLR (2022) 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.216, + 0.218, + 0.788, + 0.259 + ], + "angle": 0, + "content": "81. Yu, J., Xu, Y., Koh, J.Y., Luong, T., Baid, G., Wang, Z., Vasudevan, V., Ku, A., Yang, Y., Ayan, B.K., et al.: Scaling autoregressive models for content-rich text-to-image generation. arXiv:2206.10789 (2022) 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.26, + 0.788, + 0.3 + ], + "angle": 0, + "content": "82. Yu, L., Cheng, Y., Sohn, K., Lezama, J., Zhang, H., Chang, H., Hauptmann, A.G., Yang, M.H., Hao, Y., Essa, I., et al.: MAGVIT: Masked generative video transformer. In: CVPR (2023) 4, 5, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.301, + 0.788, + 0.355 + ], + "angle": 0, + "content": "83. Yu, L., Lezama, J., Gundavarapu, N.B., Versari, L., Sohn, K., Minnen, D., Cheng, Y., Gupta, A., Gu, X., Hauptmann, A.G., et al.: Language model beats diffusion–tokenizer is key to visual generation. arXiv preprint arXiv:2310.05737 (2023) 5, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.356, + 0.788, + 0.397 + ], + "angle": 0, + "content": "84. Yu, S., Sohn, K., Kim, S., Shin, J.: Video probabilistic diffusion models in projected latent space. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 18456-18466 (2023) 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.398, + 0.788, + 0.424 + ], + "angle": 0, + "content": "85. Zeng, Y., Wei, G., Zheng, J., Zou, J., Wei, Y., Zhang, Y., Li, H.: Make pixels dance: High-dynamic video generation. arXiv preprint arXiv:2311.10982 (2023) 13" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.425, + 0.788, + 0.467 + ], + "angle": 0, + "content": "86. Zhai, X., Kolesnikov, A., Houlsby, N., Beyer, L.: Scaling vision transformers. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 12104-12113 (2022) 9" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.468, + 0.788, + 0.522 + ], + "angle": 0, + "content": "87. Zhang, B., Gu, S., Zhang, B., Bao, J., Chen, D., Wen, F., Wang, Y., Guo, B.: Styleswin: Transformer-based gan for high-resolution image generation. In: Proceedings of the IEEE/CVF conference on computer vision and pattern recognition. pp. 11304-11314 (2022) 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.522, + 0.788, + 0.55 + ], + "angle": 0, + "content": "88. Zhang, R., Isola, P., Efros, A.A., Shechtman, E., Wang, O.: The unreasonable effectiveness of deep features as a perceptual metric. In: CVPR (2018) 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.55, + 0.788, + 0.577 + ], + "angle": 0, + "content": "89. Zheng, H., Nie, W., Vahdat, A., Anandkumar, A.: Fast training of diffusion models with masked transformers. arXiv:2306.09305 (2023) 4, 9" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.578, + 0.788, + 0.605 + ], + "angle": 0, + "content": "90. Zhou, D., Wang, W., Yan, H., Lv, W., Zhu, Y., Feng, J.: Magicvideo: Efficient video generation with latent diffusion models. arXiv preprint arXiv:2211.11018 (2022) 13" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.606, + 0.788, + 0.646 + ], + "angle": 0, + "content": "91. Zitkovich, B., Yu, T., Xu, S., Xu, P., Xiao, T., Xia, F., Wu, J., Wohlhart, P., Welker, S., Wahid, A., et al.: RT-2: Vision-language-action models transfer web knowledge to robotic control. In: CoRL (2023) 2" + }, + { + "type": "list", + "bbox": [ + 0.214, + 0.147, + 0.788, + 0.646 + ], + "angle": 0, + "content": null + } + ] +] \ No newline at end of file diff --git a/2024/Photorealistic Video Generation with Diffusion Models/327f0bbf-9985-4862-ae62-9cd9ec84c015_origin.pdf b/2024/Photorealistic Video Generation with Diffusion Models/327f0bbf-9985-4862-ae62-9cd9ec84c015_origin.pdf new file mode 100644 index 0000000000000000000000000000000000000000..88a7f747e2933ca02318fae6d5d766d017817f4f --- /dev/null +++ b/2024/Photorealistic Video Generation with Diffusion Models/327f0bbf-9985-4862-ae62-9cd9ec84c015_origin.pdf @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:a12557cb460cf92ea847a412466bc73825bcd5db378facd509940a7f5bda5d60 +size 3130322 diff --git a/2024/Photorealistic Video Generation with Diffusion Models/full.md b/2024/Photorealistic Video Generation with Diffusion Models/full.md new file mode 100644 index 0000000000000000000000000000000000000000..37ca24acf19d542357cf86600bb72e76ea449712 --- /dev/null +++ b/2024/Photorealistic Video Generation with Diffusion Models/full.md @@ -0,0 +1,357 @@ +# Photorealistic Video Generation with Diffusion Models + +Agrim Gupta $^{1,2,*}$ , Lijun Yu $^{2}$ , Kihyuk Sohn $^{2}$ , Xiuye Gu $^{2}$ , Meera Hahn $^{2}$ , Fei-Fei Li $^{1}$ , Irfan Essa $^{2,3}$ , Lu Jiang $^{2}$ , and José Lezama $^{2}$ + +1 Stanford University + +2 Google Research + +3 Georgia Institute of Technology + +https://walt-video-diffusion.github.io/ + +![](images/19f0fe423e933f545060e36b9bc5b155b5f95cbd7e4131958491645bc98ed314.jpg) + +![](images/5826244976bd7e7d23345decf1df76c8d3b3d26043fad67d77e1210e3b7f2a50.jpg) +A cute cat riding a yellow surfboard, fast waves. + +![](images/421200607f84cd04c391726528a1354659e5e1d4219e97f2046347ee1947f759.jpg) + +![](images/b9fe99a791eb987a86b0054656e3ea8274efa49b4854bb716fbc4d2c3a139fc5.jpg) + +![](images/7464fe23b9b93ab68c099ebff01c971496686123e5d797a0d036f1814bfccff3.jpg) + +![](images/e6cee2ae0357022c32630f622964095cd5f0cc2c629d0bb0c9e33902137efdef.jpg) +A cute cordi wearing a blue sweater, walking towards the camera, slow motion. + +![](images/0e881ba7887a6586e6286ca479834a2fb97c0435e64e7e7a78608f5b9dc2e50c.jpg) + +![](images/5d495a6f51d47964d0c8865d31471091a93fcfc3c8c547f8a59119d564434fc5.jpg) + +![](images/cc33d11addaedb8aef7f3ce72761f04ba5e014dacd947edba35c98be0de4acdc.jpg) + +![](images/dce248c507f5b8a4286e0fd3a9cdc05a54ac6ef936246d06624948ddfd7a8b12.jpg) + +![](images/9184a5a762781a0348298aef64f01f53d2760718d9f6ebb2dabd9c4c35fc69c9.jpg) +A raccoon wearing a black jacket, dancing in slow motion in front of the pyramids. + +![](images/eb9aa9125b66742d3426c50f21429f2461b625f90fb6552b7410797b5d3b57ad.jpg) +Fig. 1: W.A.L.T samples for text-to-video generation. W.A.L.T generates high-resolution, temporally consistent photorealistic videos from text prompts. The samples shown are $512 \times 896$ resolution over 3.6 seconds duration at 8 frames per second. + +![](images/c69d86db9537d85db7a46a005601f41462aa78e40b73b2c333d0fb77d8393e75.jpg) + +![](images/fce371586e81973d39130c746f40e2130e3824b4a98199858d584bde1b879b81.jpg) + +![](images/4639bffb1849a85039b93335e4ae288a124be8e418f63e11bf3d13e79f436aee.jpg) + +Abstract. We present W.A.L.T, a diffusion transformer for photorealistic video generation from text prompts. Our approach has two key design decisions. First, we use a causal encoder to jointly compress images and videos within a unified latent space, enabling training and generation across modalities. Second, for memory and training efficiency, we use a window attention architecture tailored for joint spatial and spatiotemporal generative modeling. Taken together these design decisions enable us to achieve state-of-the-art performance on established video (UCF-101 and Kinetics-600) and image (ImageNet) generation benchmarks without using classifier free guidance. Finally, we also train a cascade of three models for the task of text-to-video generation consisting of a base latent video diffusion model, and two video super-resolution diffusion models to generate videos of $512 \times 896$ resolution at 8 frames per second. + +# 1 Introduction + +Modeling the complex and dynamic visual world with a high level of detail is the ultimate goal of visual generative models. To accomplish this goal, we need neural architectures that can efficiently process high dimensional inputs, capture long range dependencies and have powerful scaling and parallelization properties. + +The transformer [74] is a neural network architecture designed with such desirable properties. This has encouraged the research community to increasingly favor it over domain-specific architectures in diverse fields such as language [28, 56, 58, 59], audio [1], speech [57], vision [19, 32], and robotics [6, 8, 91]. Such a trend towards unification allows researchers to share and build upon advances in traditionally disparate domains. Thus, leading to a virtuous cycle of innovation and improvement in model design favoring transformers. + +Generative modelling of images and videos has been however a notable exception to this trend. Diffusion models [68, 70], which have emerged as a leading paradigm for generative modelling of images [17, 36] and videos [38], predominantly rely on the U-Net architecture [36, 63] as backbone, consisting of a series of convolutional [47] and self-attention [74] layers. This preference stems from the fact that the memory demands of full attention mechanisms in transformers scales quadratically with input sequence length. Such scaling leads to prohibitively high costs when processing high-dimensional signals like video with transformers. + +The computational requirements of diffusion models can be reduced by operating in a lower-dimensional latent space derived from an autoencoder [21,73,76], giving rise to latent diffusion models (LDMs) [62]. A critical design choice in this context is the type of latent space employed: spatial compression (per frame latents) versus spatiotemporal compression. Spatial compression is often preferred because it enables leveraging pre-trained image autoencoders and LDMs, which are trained on large paired image-text datasets. However, this choice increases network complexity and limits the use of transformers as backbones, especially in generating high-resolution videos due to memory constraints. On the other hand, while spatiotemporal compression can mitigate these issues, it precludes the use of paired image-text datasets, which are much larger and diverse than their video counterparts. Hence, to leverage all sources of visual data and unlock scaling via transformers, it is critical to have a unified compressed representation of images and videos. + +In this work, we present Window Attention Latent Transformer (W.A.L.T): a novel transformer-based method for latent video diffusion models (LVDMs). Our method consists of two stages. First, an autoencoder maps both videos and images into a unified, lower-dimensional latent space. This design choice enables training a single generative model jointly on image and video datasets and significantly reduces the computational burden for generating high resolution videos. Subsequently, we propose a new design of transformer blocks for latent video diffusion modeling composed of self-attention layers that alternate between non-overlapping, window-restricted spatial and spatiotemporal attention. This design offers two primary benefits: Firstly, the use of local window attention + +![](images/bfe32dabb32c1ccdeba7d70de9370b595c60b91393d21684e6d5d5b0109dc573.jpg) +Fig. 2: W.A.L.T. We encode images and videos into a shared latent space. The transformer backbone processes these latents with blocks having two layers of window-restricted attention: spatial layers capture spatial relations in both images and video, while spatiotemporal layers model temporal dynamics in videos and pass through images via identity attention mask. Text conditioning is done via spatial cross-attention. + +significantly lowers computational demands. Secondly, it facilitates joint training, where the spatial layers independently process images and video frames, while the spatiotemporal layers are dedicated to modeling the temporal relationships in videos. + +While conceptually simple, our method provides the first empirical evidence of transformers' superior generation quality and parameter efficiency in latent video diffusion on public benchmarks. Specifically, we report state-of-the-art results on class-conditional video generation (UCF-101 [71]), frame prediction (Kinetics-600 [10]) and class conditional image generation (ImageNet [16]) without using classifier free guidance. Finally, to showcase the scalability and efficiency of our method we also demonstrate results on the challenging task of photorealistic text-to-video generation. We train a cascade of three models consisting of a base latent video diffusion model, and two video super-resolution diffusion models to generate videos of $512 \times 896$ resolution at 8 frames per second and report state-of-the-art zero-shot FVD score on the UCF-101 benchmark. + +Our contributions can be summarized as follows, + +- We propose W.A.L.T, a framework for transformer-based latent diffusion in a unified image and video latent space. +- We validate W.A.L.T by performing extensive ablations and achieving state-of-the-art video generation on frame prediction (K600), class-conditional video (UCF-101) and image (ImageNet) generation. +- We demonstrate the scaling properties of W.A.L.T and achieve state-of-the-art results in zero-shot UCF-101 text-to-video benchmark. + +# 2 Related Work + +Video Diffusion Models. Diffusion models have shown impressive results in image [36,40,53,62,68,69] and video generation [5,25,31,35,38,67]. Video diffusion models can be categorized into pixel-space [35,38,67] and latent-space [5,25,33,84] + +approaches, the later bringing important efficiency advantages when modeling videos. Ho et al. [38] demonstrated that the quality of text conditioned video generation can be significantly improved by jointly training on image and video data. Similarly, to leverage image datasets, latent video diffusion models inflate a pre-trained image model, typically a U-Net [63], into a video model by adding temporal layers, and initializing them as the identity function [5,35,67]. Although computationally efficient, this approach couples the design of video and image models, and precludes spatiotemporal compression. In this work, we operate on a unified latent space for images and videos, allowing us to leverage large scale image and video datasets while enjoying computational efficiency gains from spatiotemporal compression of videos. + +Transformers for Generative Modeling. Multiple classes of generative models have utilized Transformers [74] as backbone, such as, Generative adversarial networks [43, 48, 87], autoregressive [11, 12, 21, 22, 29, 60, 75, 78, 79, 81, 82] and diffusion [2, 23, 42, 51, 54, 89] models. Inspired by the success of autoregressive pretraining of large language models [56, 58, 59], Ramesh et al. [60] trained a text-to-image generation model by predicting the next visual token obtained from an image tokenizer. Subsequently, this approach was applied to multiple applications including class-conditional image generation [21, 80], text-to-image [18, 60, 77, 81] or image-to-image translation [22, 78]. Similarly, for video generation, transformer-based models were proposed to predict next tokens using 3D extensions of VQGAN [24, 39, 79, 82] or using per frame image latents [29]. Autoregressive sampling of videos is typically impractical given the very long sequences involved. To alleviate this issue, non-autoregressive sampling [11, 12], i.e. parallel token prediction, has been adopted as a more efficient solution for transformer-based video generation [29, 75, 82]. Recently, the community has started adopting transformers as the denoising backbone for diffusion models in place of U-Net [13, 40, 51, 54, 89]. To the best of our knowledge, our work is the first successful empirical demonstration (§ 5.1) of a transformer-based backbone for jointly training image and video latent diffusion models. + +# 3 Background + +Diffusion formulation. Diffusion models [36, 68, 70] are a class of generative models which learn to generate data by iteratively denoising samples drawn from a noise distribution. Gaussian diffusion models assume a forward noising process which gradually applies noise $(\epsilon)$ to real data $(x_0 \sim p_{\mathrm{data}})$ . Concretely, + +$$ +\boldsymbol {x} _ {t} = \sqrt {\gamma (t)} \boldsymbol {x} _ {0} + \sqrt {1 - \gamma (t)} \epsilon , \tag {1} +$$ + +where $\epsilon \sim \mathcal{N}(\mathbf{0},\mathbf{I}), t \in [0,1]$ , and $\gamma(t)$ is a monotonically decreasing function (noise schedule) from 1 to 0. Diffusion models are trained to learn the reverse process that inverts the forward corruptions: + +$$ +\mathbb {E} _ {\boldsymbol {x} \sim p _ {\text {d a t a}}, t \sim \mathcal {U} (0, 1), \boldsymbol {\epsilon} \sim \mathcal {N} (\boldsymbol {0}, \boldsymbol {I})} \left[ \| \boldsymbol {y} - f _ {\theta} (\boldsymbol {x} _ {t}; \boldsymbol {c}, t) \| ^ {2} \right], \tag {2} +$$ + +where $f_{\theta}$ is the denoiser model parameterized by a neural network, $\pmb{c}$ is conditioning information e.g., class labels or text prompts, and the target $\pmb{y}$ can be random noise $\pmb{\epsilon}$ , denoised input $\pmb{x_0}$ or $\pmb{v} = \sqrt{1 - \gamma(t)} \pmb{\epsilon} - \sqrt{\gamma(t)} \pmb{x_0}$ . Following [35, 65], we use $\pmb{v}$ -prediction in all our experiments. + +Latent diffusion models (LDMs). Processing high-resolution images and videos using raw pixels requires considerable computational resources. To address this, LDMs operate on the low dimensional latent space of a VQ-VAE [21,73]. VQ-VAE consists of an encoder $E(\pmb{x})$ that encodes an input video $\pmb{x} \in \mathbb{R}^{T \times H \times W \times 3}$ into a latent representation $\pmb{z} \in \mathbb{R}^{t \times h \times w \times c}$ . The encoder downsamples the video by a factor of $f_{s} = H / h = W / w$ and $f_{t} = T / t$ , where $T = t = 1$ corresponds to using an image autoencoder. An important distinction from the original VQ-VAE is the absence of a codebook of quantized embeddings as diffusion models can operate on continuous latent spaces. A decoder $D$ is trained to predict a reconstruction of the video, $\hat{\pmb{x}}$ , from $\pmb{z}$ . Following VQ-GAN [21], reconstruction quality can be further improved by adding adversarial [27] and perceptual losses [44,88]. + +# 4 W.A.L.T + +# 4.1 Learning Visual Tokens + +A key design decision in video generative modeling is the choice of latent space representation. Ideally, we want a shared and unified compressed visual representation that can be used for generative modeling of both images and videos [75,83]. The unified representation is important because joint image-video learning is preferable due to a scarcity of labeled video data [35], such as text-video pairs. Concretely, given a video sequence $\pmb{x} \in \mathbb{R}^{(1 + T) \times H \times W \times C}$ , we aim to learn a low-dimensional representation $\pmb{z} \in \mathbb{R}^{(1 + t) \times h \times w \times c}$ that performs spatial-temporal compression by a factor of $f_{s} = H / h = W / w$ in space and a factor of $f_{t} = T / t$ in time. To enable a unified representation for both videos and static images, the first frame is always encoded independently from the rest of the video. This allows static images $\pmb{x} \in \mathbb{R}^{1 \times H \times W \times C}$ to be treated as videos with a single frame, i.e. $\pmb{z} \in \mathbb{R}^{1 \times h \times w \times c}$ . + +We instantiate this design with the causal 3D CNN encoder-decoder architecture of the MAGVIT-v2 tokenizer [83]. Typically the encoder-decoder consists of regular 3D convolution layers which cannot process the first frame independently [24,82]. This limitation stems from the fact that a regular convolutional kernel of size $(k_{t},k_{h},k_{w})$ will operate on $\left\lfloor \frac{k_t - 1}{2}\right\rfloor$ frames before and $\left\lfloor \frac{k_t}{2}\right\rfloor$ frames after the input frames. Causal 3D convolution layers solve this issue as the convolutional kernel operates on only the past $k_{t} - 1$ frames. This ensures that the output for each frame is influenced solely by the preceding frames, enabling the model to tokenize the first frame independently. + +After this stage, the input to our model is a batch of latent tensors $\mathbf{z} \in \mathbb{R}^{(1 + t) \times h \times w \times c}$ representing a single video or a stack of $1 + t$ independent images (Fig. 2). Different from [83], our latent representation is real-valued and quantization-free. In the section below we describe how our model jointly processes a mixed batch of images and videos. + +# 4.2 Learning to Generate Images and Videos + +Patchify. Following the original ViT [19], we "patchify" each latent frame independently by converting it into a sequence of non-overlapping $h_p \times w_p$ patches where $h_p = h/p$ , $w_p = w/p$ and $p$ is the patch size. We use learnable positional embeddings [74], which are the sum of space and time positional embeddings. Position embeddings are added to the linear projections [19] of the patches. Note that for images, we simply add the temporal position embedding corresponding to the first latent frame. + +Window attention. Transformer models composed entirely of global self-attention modules incur significant compute and memory costs, especially for video tasks. For efficiency and for processing images and videos jointly we compute self-attention in windows [29, 74], based on two types of non-overlapping configurations: spatial (S) and spatiotemporal (ST), cf. Fig. 2. Spatial Window (SW) attention is restricted to all the tokens within a latent frame of size $1 \times h_p \times w_p$ (the first dimension is time). SW models the spatial relations in images and videos. Spatiotemporal Window (STW) attention is restricted within a 3D window of size $(1 + t) \times h_p' \times h_w'$ , modeling the temporal relationships among video latent frames. For images, we simply use an identity attention mask ensuring that the value embeddings corresponding to the image frame latents are passed through the layer as is. Finally, in addition to absolute position embeddings we also use relative position embeddings [50]. + +Our design, while conceptually straightforward, achieves computational efficiency and enables joint training on image and video datasets. In contrast to methods based on frame-level autoencoders [5, 25, 29], our approach does not suffer from flickering artifacts, which often result from encoding and decoding video frames independently. However, similar to Blattmann et al. [5], we can also potentially leverage pre-trained image LDMs with transformer backbones by simply interleaving STW layers. + +# 4.3 Conditional Generation + +To enable controllable video generation, in addition to conditioning on timestep $t$ , diffusion models are often conditioned on additional conditional information $\mathbf{c}$ such as class labels, natural language, past frames or low resolution videos. In our transformer backbone, we incorporate three types of conditioning mechanisms as described in what follows: + +Cross-attention. In addition to self-attention layers in our window transformer blocks, we add a cross-attention layer for text conditioned generation. When training models on just videos, the cross-attention layer employs the same window-restricted attention as the self-attention layer, meaning S/ST blocks will have SW/STW cross-attention layers (Fig. 2). However, for joint training, we only use SW cross-attention layers. For cross-attention we concatenate the input signal (query) with the conditioning signal (key, value) as our early experiments showed this improves performance. + +
MethodK600 FVD↓UCF FVD↓params.steps
TrIVD-GAN-FP [52]25.7±0.7--1
Video Diffusion [38]16.2±0.3-1.1B256
RIN [42]10.8-411M1000
TATS [24]-332±18321M1024
Phenaki [75]36.4±0.2-227M48
MAGVIT [82]9.9±0.376±2306M12
MAGVITv2 [83]4.3±0.158±2307M24
W.A.L.T-L Ours3.3±0.046±2313M50
W.A.L.T-XL Ours-36±2460M50
+ +Table 1: Video generation evaluation on frame prediction on Kinetics-600 and class-conditional generation on UCF-101. + +AdaLN-LoRA. Adaptive normalization layers are an important component in a broad range of generative and visual synthesis models [17, 20, 45, 53-55]. A simple way to incorporate adaptive layer normalization is to include for each layer $i$ , an MLP layer to regress a vector of conditioning parameters $A^i = \mathbf{MLP}(c + t)$ , where $A^i = \text{concat}(\gamma_1, \gamma_2, \beta_1, \beta_2, \alpha_1, \alpha_2)$ , $A^i \in \mathbb{R}^{6 \times d_{\text{model}}}$ , and $c \in \mathbb{R}^{d_{\text{model}}}, t \in \mathbb{R}^{d_{\text{model}}}$ are the condition and timestep embeddings. In the transformer block, $\gamma$ and $\beta$ scale and shift the inputs of the multi-head attention and MLP layers, respectively, while $\alpha$ scales the output of both the multi-head attention and MLP layers. The parameter count of these additional MLP layers scales linearly with the number of layers and quadratically with the model's dimensional size (num_blocks $\times$ $d_{\text{model}} \times 6 \times d_{\text{model}}$ ). For instance, in a ViT-g model with 1B parameters, the MLP layers contribute an additional 475M parameters. Inspired by [41], we propose a simple solution dubbed AdaLN-LoRA, to reduce the model parameters. For each layer, we regress conditioning parameters as + +$$ +A ^ {1} = \operatorname {M L P} (\boldsymbol {c} + \boldsymbol {t}), \quad A ^ {i} = A ^ {1} + W _ {b} ^ {i} W _ {a} ^ {i} (\boldsymbol {c} + \boldsymbol {t}) \quad \forall i \neq 1, \tag {3} +$$ + +where $W_{b}^{i}\in \mathbb{R}^{d_{\mathrm{model}}\times r}$ , $W_{a}^{i}\in \mathbb{R}^{r\times (6\times d_{\mathrm{model}})}$ . This reduces the number of trainable model parameters significantly when $r\ll d_{\mathrm{model}}$ . For example, a ViT-g model with $r = 2$ reduces the MLP parameters from 475M to 12M. + +Self-conditioning. In addition to being conditioned on external inputs, iterative generative algorithms can also be conditioned on their own previously generated samples during inference [3, 14, 66]. Specifically, Chen et al. [14] modify the training process for diffusion models, such that with some probability $p_{\mathrm{sc}}$ the model first generates a sample $\tilde{\pmb{z}}_0 = f_\theta(\pmb{z}_t; \pmb{0}, \pmb{c}, t)$ and then refines this estimate using another forward pass conditioned on this initial sample: $f_\theta(\pmb{z}_t; \text{stopgrad}(\tilde{\pmb{z}}_0), \pmb{c}, t)$ . With probability $1 - p_{\mathrm{sc}}$ , only a single forward pass is done. We concatenate the model estimate with the input along the channel dimension and found this simple technique to work well when used in conjunction with $\pmb{v}$ -prediction. + +# 4.4 Autoregressive Generation + +For generating long videos via autoregressive prediction we also train our model jointly on the task of frame prediction. This is achieved by conditioning the model on past frames with a probability of $p_{\mathrm{fp}}$ during training. Specifically, the model is conditioned using $c_{\mathrm{fp}} = \text{concat}(m_{\mathrm{fp}} \circ z_t, m_{\mathrm{fp}})$ , where $m_{\mathrm{fp}}$ is a binary mask. The binary mask indicates the number of past frames used for conditioning. We condition on either 1 latent frame (image to video generation) or 2 latent frames (video prediction). This conditioning is integrated into the model through concatenation along the channel dimension of the noisy latent input. During inference, we use standard classifier-free guidance with $c_{\mathrm{fp}}$ as the conditioning signal. + +# 4.5 Video Super Resolution + +Generating high-resolution videos with a single model is computationally prohibitive. Following [37], we use a cascaded approach with three models operating at increasing resolutions. Our base model generates videos at $128 \times 128$ resolution which are subsequently upsampled twice via two super resolution stages. We first spatially upscale the low resolution input $z^{\mathrm{lr}}$ (video or image) using a depth-to-space convolution operation. Note that, unlike training where ground truth low-resolution inputs are available, inference relies on latents produced by preceding stages (cf. teaching-forcing). To reduce this discrepancy and improve the robustness of the super-resolution stages in handling artifacts generated by lower resolution stages, we use noise conditioning augmentation [37]. Concretely, noise is added in accordance with $\gamma(t)$ , by sampling a noise level as $t_{\mathrm{sr}} \sim \mathcal{U}(0, t_{\max \text{noise}})$ and is provided as input to our AdaLN-LoRA layers. + +Aspect-ratio finetuning. To simplify training and leverage broad data sources with different aspect ratios, we train our base stage using a square aspect ratio. We fine-tune the base stage on a subset of data to generate videos with a 9:16 aspect ratio by interpolating position embeddings. + +# 5 Experiments + +In this section, we evaluate our method on multiple tasks: class-conditional image and video generation, frame prediction and text conditioned video generation and perform extensive ablation studies of different design choices. For qualitative results, see Fig. 1, Fig. 3, Fig. 4 and videos on our project website. See appendix for additional details. + +# 5.1 Visual Generation + +Video generation. We consider two standard video benchmarks, UCF-101 [71] for class-conditional generation and Kinetics-600 [10] for video prediction with 5 conditioning frames. We use FVD [72] as our primary evaluation metric. + +
MethodCost (Iter×BS)FID↓IS↑params.steps
BigGAN-deep [7]-6.95171.4160M1
LDM-4 [62]178k×120010.56103.5400M250
DiT-XL/2 [54]7000k×2569.62121.5675M250
ADM [17]-7.49127.5608M2000
MDT [23]6500k×2566.23143.0676M250
MaskDiT [89]1200k×10245.69178.0736M40
RIN [42]600k×10243.42182.0410M1000
simple diffusion [40]500K×20482.77211.82B512
VDM++ [46]-2.40225.32B512
W.A.L.T-L Ours437k×10242.56215.1460M50
+ +Table 2: Class-conditional image generation on ImageNet ${256} \times {256}$ . We adopt the evaluation protocol and implementation of ADM [17] and report results without classifier free guidance. + +Across both datasets, W.A.L.T significantly outperforms all prior works (Tab. 1). Compared to prior video diffusion models, we achieve state-of-the-art performance with less model parameters, and require 50 DDIM [69] inference steps. + +Image generation. To verify the modeling capabilities of $W.A.L.T$ on the image domain, we train a version of $W.A.L.T$ for the standard ImageNet class-conditional setting. For evaluation, we follow ADM [17] and report the FID [34] and Inception [64] scores calculated on 50K samples generated in 50 DDIM steps. We compare (Table 2) $W.A.L.T$ with state-of-the-art image generation methods for $256 \times 256$ resolution. Our model outperforms prior works without requiring specialized schedules, convolution inductive bias, improved diffusion losses, and classifier free guidance. Although VDM++ [46] has slightly better FID score, the model has significantly more parameters (2B). + +# 5.2 Ablation Studies + +We ablate $W.A.L.T$ to understand the contribution of various design decisions with the default settings: model L, patch size 1, $1 \times 16 \times 16$ spatial window, $5 \times 8 \times 8$ spatiotemporal window, $p_{\mathrm{sc}} = 0.9$ , $c = 8$ and $r = 2$ . + +Patch size. In various computer vision tasks utilizing ViT [19]-based models, a smaller patch size $p$ has been shown to consistently enhance performance [9, 19, 30, 86]. Similarly, our findings also indicate that a reduced patch size improves performance (Table 3a). + +Window attention. We compare three different STW window configurations with full self-attention (Table 3b). We find that local self-attention can achieve competitive (or better) performance while being significantly faster (up to $2 \times$ ) and requiring less accelerator memory. + +Self-conditioning. In Table 3c we study the influence of varying the self-conditioning rate $p_{\mathrm{sc}}$ on generation quality. We notice a clear trend: increasing the self-conditioning rate from 0.0 (no self-conditioning) to 0.9 improves the FVD score substantially (44%). + +
patch size pFVD↓IS↑
160.787.2
2134.482.2
4461.863.9
+ +
st windowFVD↓IS↑sps
5 × 4 × 456.987.32.24
5 × 8 × 859.687.42.00
5 × 16 × 1655.387.41.75
full self attn.59.987.81.20
+ +(a) Patch size. Lower patch size is significantly better. + +(b) Spatiotemporal window size. Full self-attention is not essential for good performance. sps is steps per sec. + +
pscFVD↓IS↑
0.0109.982.6
0.376.086.5
0.660.086.8
0.961.487.1
+ +
rFVD↓IS↑params
260.787.2313 M
456.687.3314 M
1655.588.0316 M
6454.487.9324 M
25652.588.5357 M
+ +(c) Self-conditioning. Higher $p_{\mathrm{sc}}$ is bet- (d) AdaLN-LoRA. Bigger $r$ is better. + +
FVD↓IS↑
w/o qk norm [15]59.086.8
w/o latent norm67.987.1
w/o zero snr [49]91.084.2
full method60.787.2
+ +(e) Other improvements. See text for details. + +
crFVD↓FVD↓IS↑
437.786.484.9
817.175.486.3
168.267.086.0
323.583.482.9
+ +(f) Latent dimension $c$ . Higher $c$ is better. + +Table 3: Ablation experiments on UCF-101 [71]. We compare FVD and inception scores to ablate important design decisions with the default setting: L model, $1 \times 16 \times 16$ spatial window, $5 \times 8 \times 8$ saptiotemporal (st) window, $p_{\mathrm{sc}} = 0.9$ , $c = 8$ and $r = 2$ . + +AdaLN-LoRA. An important design decision in diffusion models is the conditioning mechanism. We investigate the effect of increasing the bottleneck dimension $r$ in our proposed AdaLN-LoRA layers (Table 3d). This hyperparameter provides a flexible way to trade off between number of model parameters and generation performance. As shown in Table 3d, increasing $r$ improves performance but also increases model parameters. This highlights an important model design question: given a fixed parameter budget, how should we allocate parameters - either by using separate AdaLN layers, or by increasing base model parameters while using shared AdaLN-LoRA layers? We explore this in Table 4 by comparing two model configurations: W.A.L.T-L with separate AdaLN layers and W.A.L.T-XL with AdaLN-LoRA and $r = 2$ . While both configurations yield similar FVD and Inception scores, W.A.L.T-XL achieves a lower final loss value, suggesting the advantage of allocating more parameters to the base model and choosing an appropriate $r$ value within accelerator memory limits. + +
ModelAdaLNFVD↓IS↑params.final loss
Lseparate34.690.2458M0.274
XLLoRA-236.789.4460M0.268
+ +Table 4: Parameter matched comparison between AdaLN-LoRA and per layer adaln layers. See text for details. + +Noise schedule. Common latent diffusion noise schedules [62] typically do not ensure a zero signal-to-noise ratio (SNR) at the final timestep, i.e., at $t = 1$ , $\gamma(t) > 0$ . This leads to a mismatch between training and inference phases. During inference, models are expected to start from purely Gaussian noise, whereas during training, at $t = 1$ , a small amount of signal information remains accessible to the model. This is especially harmful for video generation as videos have high temporal redundancy. Even minimal information leakage at $t = 1$ can reveal substantial information to the model. Addressing this mismatch by enforcing a zero terminal SNR [49] significantly improves performance (Table 3e). Note that this approach was originally proposed to fix over-exposure problems in image generation, but we find it effective for video generation as well. + +Autoencoder. Finally, we investigate one critical but often overlooked hyperparameter in the first stage of our model: the channel dimension $c$ of the autoencoder latent $z$ . As shown in Table 3f, increasing $c$ significantly improves the reconstruction quality (lower rFVD) while keeping the same spatial $f_{s}$ and temporal compression $f_{t}$ ratios. Empirically, we found that both lower and higher values of $c$ lead to poor FVD scores in generation, with a sweet spot of $c = 8$ working well across most datasets and tasks we evaluated. We also normalize the latents before processing them via transformer which further improves performance. + +In our transformer models, we use query-key normalization [15] as it helps stabilize training for larger models. Finally, we note that some of our default settings are not optimal, as indicated by ablation studies. These defaults were chosen early on for their robustness across datasets, though further tuning may improve performance. + +# 5.3 Text-to-video + +We train $W.A.L.T$ for text-to-video jointly on text-image and text-video pairs (Sec. 4.2). We used a dataset of $\sim 970\mathrm{M}$ text-image pairs and $\sim 89\mathrm{M}$ text-video pairs from the public internet and internal sources. We train our base model at resolution $17 \times 128 \times 128$ (3B parameters), and two $2 \times$ cascaded super-resolution models for $17 \times 128 \times 224 \rightarrow 17 \times 256 \times 448$ (L, 1.3B, $p = 2$ ) and $17 \times 256 \times 448 \rightarrow 17 \times 512 \times 896$ (L, 419M, $p = 2$ ) respectively. We fine-tune the base stage for the $9:16$ aspect ratio to generate videos at resolution $128 \times 224$ . We use classifier free guidance for all our text-to-video results. + +![](images/3fbda4a8bdb8e31bb1dc4a016cfeac307af56b48312cc596c6ad45f79c814283.jpg) +A polar bear swimming. + +![](images/6cc71d1c985cc4361b40012b480bdab728ca2d6e6f7641a44e221db218f9a558.jpg) +Pouring beer into an empty glass, low angle shot, bar in the background. + +![](images/77d16469cfeda54b85c88371d136e0c211b417ad9369ce87102415db1e8605bf.jpg) +A cat eating food out of a bowl, in the style of Van Gogh. + +![](images/661d6094151bd65db1e7b4c6027bf3930220eb15ec10423e7b8150443ea8964b.jpg) +Pouring chocolate sauce over vanilla ice cream in a cone, studio lighting. + +![](images/a3b932a13c27a28ee5f53b633e465054e881c7e085bb01d3cb2ca3f5c7c64dea.jpg) +A robot ballerina dancing gracefully, highly detailed, studio lighting. + +![](images/adaf4da68cd717ac491e1f84afc9a5f418f9589c42e8f56cfbd7f95c38da5be8.jpg) +A cute panda skateboarding in the sky, over snow covered mountains, with dreamy whimsical atmosphere. + +![](images/b05555d5f782fdeb616ea285f818bd0552a812fb53692b4f84e47397cb382fab.jpg) +An astronaut riding a horse. +Fig. 3: Qualitative evaluation. Example videos generated by $W.A.L.T$ from natural language prompts at $512 \times 896$ resolution over 3.6 seconds duration at 8 frames per second. The $W.A.L.T$ model is able to generate temporally consistent photorealistic videos that align with the textual prompt. There are more results on our project website. + +
MethodIS (↑)FVD (↓)
CogVideo (Chinese) [39]23.6751.3
CogVideo (English) [39]25.3701.6
MagicVideo [90]-699.0
EMU Video* [26]42.7606.2
Video LDM [5]33.5550.6
Make-A-Video [67]33.0367.2
PYoCo [25]47.8355.2
PixelDance* [85]42.1242.8
SVD* [4]-242.0
W.A.L.T (Ours) 419M (video only)26.8598.8
W.A.L.T (Ours) 419M (video + image)31.7344.5
W.A.L.T (Ours) 3B (video + image)35.1258.1
+ +Table 5: UCF-101 text-to-video generation. Joint training on image and video datasets in conjunction with scaling the model parameters is essential for high quality video generation. * Denotes concurrent work. + +Quantitative Evaluation Evaluating text-conditioned video generation scientifically remains a significant challenge, in part due to the absence of standardized training datasets and benchmarks. So far we have focused our experiments and analyses on the standard academic benchmarks, which use the same training data to ensure controlled and fair comparisons. Nevertheless, to compare with prior work on text-to-video, we also report results on the UCF-101 dataset in the zero-shot evaluation protocol in Table 5 [25,39,67]. Also see supplement. + +Joint training. A primary strength of our framework is its ability to train simultaneously on both image and video datasets. In Table 5 we ablate the impact of this joint training. Specifically, we trained two versions of W.A.L.T-L (each with 419M params.) models using the default settings specified in § 5.2. We find that joint training leads to a notable improvement across both metrics. Our results align with the findings of Ho et al. [38], who demonstrated the benefits of joint training for pixel-based video diffusion models with U-Net backbones. + +Scaling. Transformers are known for their ability to scale effectively in many tasks [6,15,58]. In Table 5 we show the benefits of scaling our transformer model for video diffusion. Scaling our base model size leads to significant improvements on both the metrics. It is important to note, however, that our base model is considerably smaller than leading text-to-video systems. For instance, Ho et al. [35] trained base model of 5.7B parameters. Hence, we believe scaling our models further is an important direction of future work. + +Comparison with prior work. In Table 5, we present a system-level comparison of various text-to-video methods. Our results are promising, surpassing all previous work in the FVD metric, and being competitive with concurrent work. In terms of the IS, W.A.L.T outperforms all previous work but PYoCo [25]. A possible explanation for this discrepancy might be PYoCo's use of stronger text embeddings. Specifically, they utilize both CLIP [56] and T5-XXL [61] encoders, whereas we employ a T5-XL [61] text encoder only. + +![](images/b4fc56e49d0c06ee7d076fb75256abb4700ab7f064b86ef37f7efeb8527cb724.jpg) +Fig. 4: Samples at $512 \times 896$ resolution. Top: Text-to-video samples of consistent 3D camera motion (5.1 secs). Bottom: Image-to-video samples (3.8 secs). Images are generated using off-shelf models. Visit our project website to view them in video format. + +Qualitative Results As mentioned in § 4.4, we jointly train our model on the task of frame prediction conditioned on 1 or 2 latent frames. Hence, our model can be used for animating images (image-to-video) and generating longer videos with consistent camera motion (Fig. 4). See videos on our project website. + +# 6 Conclusion + +In this work, we introduce W.A.L.T, a simple, scalable, and efficient transformer-based framework for latent video diffusion models. A key design decision in our framework is to operate on a shared, compressed visual representation that enables generative modeling of both images and videos. We demonstrate state-of-the-art results for image and video generation using a transformer backbone with windowed attention. We also train a cascade of three W.A.L.T models jointly on image and video datasets, to synthesize high-resolution, temporally consistent photorealistic videos from natural language descriptions. While generative modeling has seen tremendous recent advances for images, progress on video generation has lagged behind. We hope that scaling our framework will enable the creation of powerful video generative models which can understand and model the physical world. + +# Acknowledgements + +We thank Bryan Seybold, Dan Kondratyuk, David Ross, Hartwig Adam, Huisheng Wang, Jason Baldridge, Mauricio Delbracio and Orly Liba for helpful discussions and feedback. + +# References + +1. Agostinelli, A., Denk, T.I., Borsos, Z., Engel, J., Verzetti, M., Caillon, A., Huang, Q., Jansen, A., Roberts, A., Tagliasacchi, M., et al.: MusicLM: Generating music from text. arXiv:2301.11325 (2023) 2 +2. Bao, F., Li, C., Cao, Y., Zhu, J.: All are worth words: a vit backbone for score-based diffusion models. In: NeurIPS 2022 Workshop on Score-Based Methods (2022) 4 +3. Bengio, S., Vinyals, O., Jaitly, N., Shazeer, N.: Scheduled sampling for sequence prediction with recurrent neural networks. Advances in neural information processing systems 28 (2015) 7 +4. Blattmann, A., Dockhorn, T., Kulal, S., Mendelevitch, D., Kilian, M., Lorenz, D., Levi, Y., English, Z., Voleti, V., Letts, A., et al.: Stable video diffusion: Scaling latent video diffusion models to large datasets. arXiv preprint arXiv:2311.15127 (2023) 13 +5. Blattmann, A., Rombach, R., Ling, H., Dockhorn, T., Kim, S.W., Fidler, S., Kreis, K.: Align your latents: High-resolution video synthesis with latent diffusion models. In: CVPR (2023) 3, 4, 6, 13 +6. Bousmalis, K., Vezzani, G., Rao, D., Devin, C., Lee, A.X., Bauza, M., Davchev, T., Zhou, Y., Gupta, A., Raju, A., et al.: Robocat: A self-improving foundation agent for robotic manipulation. arXiv preprint arXiv:2306.11706 (2023) 2, 13 +7. Brock, A., Donahue, J., Simonyan, K.: Large scale GAN training for high fidelity natural image synthesis. In: ICLR (2018) 9 +8. Brohan, A., Brown, N., Carbajal, J., Chebotar, Y., Dabis, J., Finn, C., Gopalakrishnan, K., Hausman, K., Herzog, A., Hsu, J., et al.: Rt-1: Robotics transformer for real-world control at scale. arXiv preprint arXiv:2212.06817 (2022) 2 +9. Caron, M., Touvron, H., Misra, I., Jégou, H., Mairal, J., Bojanowski, P., Joulin, A.: Emerging properties in self-supervised vision transformers. In: Proceedings of the IEEE/CVF international conference on computer vision. pp. 9650-9660 (2021) 9 +0. Carreira, J., Noland, E., Banki-Horvath, A., Hillier, C., Zisserman, A.: A short note about Kinetics-600. arXiv:1808.01340 (2018) 3, 8 +1. Chang, H., Zhang, H., Barber, J., Maschinot, A., Lezama, J., Jiang, L., Yang, M.H., Murphy, K., Freeman, W.T., Rubinstein, M., et al.: Muse: Text-to-image generation via masked generative transformers. In: ICML (2023) 4 +2. Chang, H., Zhang, H., Jiang, L., Liu, C., Freeman, W.T.: MaskGIT: Masked generative image transformer. In: CVPR (2022) 4 +3. Chen, J., Yu, J., Ge, C., Yao, L., Xie, E., Wu, Y., Wang, Z., Kwok, J., Luo, P., Lu, H., et al.: Pixart- $\alpha$ beta: Fast training of diffusion transformer for photorealistic text-to-image synthesis. arXiv preprint arXiv:2310.00426 (2023) 4 +4. Chen, T., Zhang, R., Hinton, G.: Analog bits: Generating discrete data using diffusion models with self-conditioning. arXiv preprint arXiv:2208.04202 (2022) 7 +5. Dehghani, M., Djolonga, J., Mustafa, B., Padlewski, P., Heek, J., Gilmer, J., Steiner, A.P., Caron, M., Geirhos, R., Alabdulmohsin, I., et al.: Scaling vision transformers to 22 billion parameters. In: International Conference on Machine Learning. pp. 7480-7512. PMLR (2023) 10, 11, 13 + +16. Deng, J., Dong, W., Socher, R., Li, L.J., Li, K., Fei-Fei, L.: ImageNet: A large-scale hierarchical image database. In: CVPR (2009) 3 +17. Dhariwal, P., Nichol, A.: Diffusion models beat GANs on image synthesis. In: NeurIPS (2021) 2, 7, 9 +18. Ding, M., Yang, Z., Hong, W., Zheng, W., Zhou, C., Yin, D., Lin, J., Zou, X., Shao, Z., Yang, H., et al.:CogView: Mastering text-to-image generation via transformers. In:NeurIPS (2021) 4 +19. Dosovitskiy, A., Beyer, L., Kolesnikov, A., Weissenborn, D., Zhai, X., Unterthiner, T., Dehghani, M., Minderer, M., Heigold, G., Gelly, S., et al.: An image is worth 16x16 words: Transformers for image recognition at scale. In: ICLR (2020) 2, 6, 9 +20. Dumoulin, V., Shlens, J., Kudlur, M.: A learned representation for artistic style. arXiv preprint arXiv:1610.07629 (2016) 7 +21. Esser, P., Rombach, R., Ommer, B.: Taming transformers for high-resolution image synthesis. In: CVPR (2021) 2, 4, 5 +22. Gafni, O., Polyak, A., Ashual, O., Sheynin, S., Parikh, D., Taigman, Y.: Makea-scene: Scene-based text-to-image generation with human priors. In: European Conference on Computer Vision. pp. 89-106. Springer (2022) 4 +23. Gao, S., Zhou, P., Cheng, M.M., Yan, S.: Masked diffusion transformer is a strong image synthesizer. arXiv:2303.14389 (2023) 4, 9 +24. Ge, S., Hayes, T., Yang, H., Yin, X., Pang, G., Jacobs, D., Huang, J.B., Parikh, D.: Long video generation with time-agnostic VQGAN and time-sensitive transformer. In: ECCV (2022) 4, 5, 7 +25. Ge, S., Nah, S., Liu, G., Poon, T., Tao, A., Catanzaro, B., Jacobs, D., Huang, J.B., Liu, M.Y., Balaji, Y.: Preserve your own correlation: A noise prior for video diffusion models. arXiv preprint arXiv:2305.10474 (2023) 3, 6, 13 +26. Girdhar, R., Singh, M., Brown, A., Duval, Q., Azadi, S., Rambhatla, S.S., Shah, A., Yin, X., Parikh, D., Misra, I.: Emu video: Factorizing text-to-video generation by explicit image conditioning. arXiv preprint arXiv:2311.10709 (2023) 13 +27. Goodfellow, I., Pouget-Abadie, J., Mirza, M., Xu, B., Warde-Farley, D., Ozair, S., Courville, A., Bengio, Y.: Generative adversarial nets. Advances in neural information processing systems 27 (2014) 5 +28. Google: PaLM 2 technical report. arXiv:2305.10403 (2023) 2 +29. Gupta, A., Tian, S., Zhang, Y., Wu, J., Martín-Martín, R., Fei-Fei, L.: MaskViT: Masked visual pre-training for video prediction. In: ICLR (2022) 4, 6 +30. Gupta, A., Wu, J., Deng, J., Fei-Fei, L.: Siamese masked autoencoders. arXiv preprint arXiv:2305.14344 (2023) 9 +31. Harvey, W., Naderiparizi, S., Masrani, V., Weilbach, C., Wood, F.: Flexible diffusion modeling of long videos. Advances in Neural Information Processing Systems 35, 27953-27965 (2022) 3 +32. He, K., Chen, X., Xie, S., Li, Y., Dollar, P., Girshick, R.: Masked autoencoders are scalable vision learners. In: CVPR (2022) 2 +33. He, Y., Yang, T., Zhang, Y., Shan, Y., Chen, Q.: Latent video diffusion models for high-fidelity long video generation. arXiv preprint arXiv:2211.13221 (2023) 3 +34. Heusel, M., Ramsauer, H., Unterthiner, T., Nessler, B., Hochreiter, S.: GANs trained by a two time-scale update rule converge to a local nash equilibrium. In: NeurIPS (2017) 9 +35. Ho, J., Chan, W., Sahara, C., Whang, J., Gao, R., Gritsenko, A., Kingma, D.P., Poole, B., Norouzi, M., Fleet, D.J., et al.: Imagen video: High definition video generation with diffusion models. arXiv:2210.02303 (2022) 3, 4, 5, 13 +36. Ho, J., Jain, A., Abbeel, P.: Denoising diffusion probabilistic models. In: NeurIPS (2020) 2, 3, 4 + +37. Ho, J., Sahara, C., Chan, W., Fleet, D.J., Norouzi, M., Salimans, T.: Cascaded diffusion models for high fidelity image generation. JMLR 23(1), 2249-2281 (2022) 8 +38. Ho, J., Salimans, T., Gritsenko, A., Chan, W., Norouzi, M., Fleet, D.J.: Video diffusion models. In: ICLR Workshops (2022) 2, 3, 4, 7, 13 +39. Hong, W., Ding, M., Zheng, W., Liu, X., Tang, J.: CogVideo: Large-scale pretraining for text-to-video generation via transformers. arXiv:2205.15868 (2022) 4, 13 +40. Hoogeboom, E., Heek, J., Salimans, T.: simple diffusion: End-to-end diffusion for high resolution images. In: ICML (2023) 3, 4, 9 +41. Hu, E.J., Shen, Y., Wallis, P., Allen-Zhu, Z., Li, Y., Wang, S., Wang, L., Chen, W.: Lora: Low-rank adaptation of large language models. In: ICLR (2021) 7 +42. Jabri, A., Fleet, D.J., Chen, T.: Scalable adaptive computation for iterative generation. In: ICML (2023) 4, 7, 9 +43. Jiang, Y., Chang, S., Wang, Z.: Transgan: Two pure transformers can make one strong gan, and that can scale up. Advances in Neural Information Processing Systems 34, 14745-14758 (2021) 4 +44. Johnson, J., Alahi, A., Fei-Fei, L.: Perceptual losses for real-time style transfer and super-resolution. In: ECCV (2016) 5 +45. Karras, T., Laine, S., Aila, T.: A style-based generator architecture for generative adversarial networks. In: CVPR (2019) 7 +46. Kingma, D.P., Gao, R.: Understanding the diffusion objective as a weighted integral of ellbos. arXiv:2303.00848 (2023) 9 +47. LeCun, Y., Bottou, L., Bengio, Y., Haffner, P.: Gradient-based learning applied to document recognition. Proceedings of the IEEE 86(11), 2278-2324 (1998) 2 +48. Lee, K., Chang, H., Jiang, L., Zhang, H., Tu, Z., Liu, C.: Vitgan: Training gans with vision transformers. arXiv preprint arXiv:2107.04589 (2021) 4 +49. Lin, S., Liu, B., Li, J., Yang, X.: Common diffusion noise schedules and sample steps are flawed. arXiv preprint arXiv:2305.08891 (2023) 10, 11 +50. Liu, Z., Lin, Y., Cao, Y., Hu, H., Wei, Y., Zhang, Z., Lin, S., Guo, B.: Swin transformer: Hierarchical vision transformer using shifted windows. In: Proceedings of the IEEE/CVF international conference on computer vision. pp. 10012-10022 (2021) 6 +51. Lu, H., Yang, G., Fei, N., Huo, Y., Lu, Z., Luo, P., Ding, M.: Vdt: General-purpose video diffusion transformers via mask modeling. arXiv preprint arXiv:2305.13311 (2023) 4 +52. Luc, P., Clark, A., Dieleman, S., Casas, D.d.L., Doron, Y., Cassirer, A., Simonyan, K.: Transformation-based adversarial video prediction on large-scale data. arXiv:2003.04035 (2020) 7 +53. Nichol, A.Q., Dhariwal, P.: Improved denoising diffusion probabilistic models. In: International Conference on Machine Learning. pp. 8162-8171. PMLR (2021) 3, 7 +54. Peebles, W., Xie, S.: Scalable diffusion models with transformers. arXiv:2212.09748 (2022) 4, 7, 9 +55. Perez, E., Strub, F., De Vries, H., Dumoulin, V., Courville, A.: Film: Visual reasoning with a general conditioning layer. In: Proceedings of the AAAI conference on artificial intelligence. vol. 32 (2018) 7 +56. Radford, A., Kim, J.W., Hallacy, C., Ramesh, A., Goh, G., Agarwal, S., Sastry, G., Askell, A., Mishkin, P., Clark, J., et al.: Learning transferable visual models from natural language supervision. In: ICML (2021) 2, 4, 13 +57. Radford, A., Kim, J.W., Xu, T., Brockman, G., McLeavey, C., Sutskever, I.: Robust speech recognition via large-scale weak supervision. In: International Conference on Machine Learning. pp. 28492-28518. PMLR (2023) 2 + +58. Radford, A., Narasimhan, K., Salimans, T., Sutskever, I., et al.: Improving language understanding by generative pre-training (2018) 2, 4, 13 +59. Radford, A., Wu, J., Child, R., Luan, D., Amodei, D., Sutskever, I., et al.: Language models are unsupervised multitask learners. OpenAI blog 1(8), 9 (2019) 2, 4 +60. Ramesh, A., Pavlov, M., Goh, G., Gray, S., Voss, C., Radford, A., Chen, M., Sutskever, I.: Zero-shot text-to-image generation. In: ICML (2021) 4 +61. Roberts, A., Raffel, C., Lee, K., Matena, M., Shazeer, N., Liu, P.J., Narang, S., Li, W., Zhou, Y.: Exploring the limits of transfer learning with a unified text-to-text transformer (2019) 13 +62. Rombach, R., Blattmann, A., Lorenz, D., Esser, P., Ommer, B.: High-resolution image synthesis with latent diffusion models. In: CVPR (2022) 2, 3, 9, 11 +63. Ronneberger, O., Fischer, P., Brox, T.: U-Net: Convolutional networks for biomedical image segmentation. In: MICCAI (2015) 2, 4 +64. Salimans, T., Goodfellow, I., Zaremba, W., Cheung, V., Radford, A., Chen, X.: Improved techniques for training GANs. In: NeurIPS (2016) 9 +65. Salimans, T., Ho, J.: Progressive distillation for fast sampling of diffusion models arXiv preprint arXiv:2202.00512 (2022) 5 +66. Savinov, N., Chung, J., Binkowski, M., Elsen, E., Oord, A.v.d.: Step-unrolled denoising autoencoders for text generation. arXiv preprint arXiv:2112.06749 (2021) 7 +67. Singer, U., Polyak, A., Hayes, T., Yin, X., An, J., Zhang, S., Hu, Q., Yang, H., Ashual, O., Gafni, O., et al.: Make-a-video: Text-to-video generation without text-video data. arXiv:2209.14792 (2022) 3, 4, 13 +68. Sohl-Dickstein, J., Weiss, E., Maheswaranathan, N., Ganguli, S.: Deep unsupervised learning using nonequilibrium thermodynamics. In: ICML (2015) 2, 3, 4 +69. Song, J., Meng, C., Ermon, S.: Denoising diffusion implicit models. arXiv preprint arXiv:2010.02502 (2020) 3, 9 +70. Song, Y., Ermon, S.: Generative modeling by estimating gradients of the data distribution. In: NeurIPS (2019) 2, 4 +71. Soomro, K., Zamir, A.R., Shah, M.: UCF101: A dataset of 101 human actions classes from videos in the wild. arXiv:1212.0402 (2012) 3, 8, 10 +72. Unterthiner, T., van Steenkiste, S., Kurach, K., Marinier, R., Michalski, M., Gelly, S.: Towards accurate generative models of video: A new metric & challenges. arXiv:1812.01717 (2018) 8 +73. Van Den Oord, A., Vinyals, O., et al.: Neural discrete representation learning. In: NeurIPS (2017) 2, 5 +74. Vaswani, A., Shazeer, N., Parmar, N., Uszkoreit, J., Jones, L., Gomez, A.N., Kaiser, L., Polosukhin, I.: Attention is all you need. In: NeurIPS (2017) 2, 4, 6 +75. Villegas, R., Babaeizadeh, M., Kindermans, P.J., Moraldo, H., Zhang, H., Saffar, M.T., Castro, S., Kunze, J., Erhan, D.: Phenaki: Variable length video generation from open domain textual description. arXiv:2210.02399 (2022) 4, 5, 7 +76. Vincent, P., Larochelle, H., Bengio, Y., Manzagol, P.A.: Extracting and composing robust features with denoising autoencoders. In: Proceedings of the 25th international conference on Machine learning. pp. 1096-1103 (2008) 2 +77. Wu, C., Huang, L., Zhang, Q., Li, B., Ji, L., Yang, F., Sapiro, G., Duan, N.: Godiva: Generating open-domain videos from natural descriptions. arXiv preprint arXiv:2104.14806 (2021) 4 +78. Wu, C., Liang, J., Ji, L., Yang, F., Fang, Y., Jiang, D., Duan, N.: Niwa: Visual synthesis pre-training for neural visual world creation. In: European conference on computer vision. pp. 720-736. Springer (2022) 4 + +79. Yan, W., Zhang, Y., Abbeel, P., Srinivas, A.: VideoGPT: Video generation using VQ-VAE and transformers. arXiv:2104.10157 (2021) 4 +80. Yu, J., Li, X., Koh, J.Y., Zhang, H., Pang, R., Qin, J., Ku, A., Xu, Y., Baldridge, J., Wu, Y.: Vector-quantized image modeling with improved VQGAN. In: ICLR (2022) 4 +81. Yu, J., Xu, Y., Koh, J.Y., Luong, T., Baid, G., Wang, Z., Vasudevan, V., Ku, A., Yang, Y., Ayan, B.K., et al.: Scaling autoregressive models for content-rich text-to-image generation. arXiv:2206.10789 (2022) 4 +82. Yu, L., Cheng, Y., Sohn, K., Lezama, J., Zhang, H., Chang, H., Hauptmann, A.G., Yang, M.H., Hao, Y., Essa, I., et al.: MAGVIT: Masked generative video transformer. In: CVPR (2023) 4, 5, 7 +83. Yu, L., Lezama, J., Gundavarapu, N.B., Versari, L., Sohn, K., Minnen, D., Cheng, Y., Gupta, A., Gu, X., Hauptmann, A.G., et al.: Language model beats diffusion–tokenizer is key to visual generation. arXiv preprint arXiv:2310.05737 (2023) 5, 7 +84. Yu, S., Sohn, K., Kim, S., Shin, J.: Video probabilistic diffusion models in projected latent space. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 18456-18466 (2023) 3 +85. Zeng, Y., Wei, G., Zheng, J., Zou, J., Wei, Y., Zhang, Y., Li, H.: Make pixels dance: High-dynamic video generation. arXiv preprint arXiv:2311.10982 (2023) 13 +86. Zhai, X., Kolesnikov, A., Houlsby, N., Beyer, L.: Scaling vision transformers. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 12104-12113 (2022) 9 +87. Zhang, B., Gu, S., Zhang, B., Bao, J., Chen, D., Wen, F., Wang, Y., Guo, B.: Styleswin: Transformer-based gan for high-resolution image generation. In: Proceedings of the IEEE/CVF conference on computer vision and pattern recognition. pp. 11304-11314 (2022) 4 +88. Zhang, R., Isola, P., Efros, A.A., Shechtman, E., Wang, O.: The unreasonable effectiveness of deep features as a perceptual metric. In: CVPR (2018) 5 +89. Zheng, H., Nie, W., Vahdat, A., Anandkumar, A.: Fast training of diffusion models with masked transformers. arXiv:2306.09305 (2023) 4, 9 +90. Zhou, D., Wang, W., Yan, H., Lv, W., Zhu, Y., Feng, J.: Magicvideo: Efficient video generation with latent diffusion models. arXiv preprint arXiv:2211.11018 (2022) 13 +91. Zitkovich, B., Yu, T., Xu, S., Xu, P., Xiao, T., Xia, F., Wu, J., Wohlhart, P., Welker, S., Wahid, A., et al.: RT-2: Vision-language-action models transfer web knowledge to robotic control. In: CoRL (2023) 2 \ No newline at end of file diff --git a/2024/Photorealistic Video Generation with Diffusion Models/images.zip b/2024/Photorealistic Video Generation with Diffusion Models/images.zip new file mode 100644 index 0000000000000000000000000000000000000000..62b303fd43446bc3a1c108b2a3a8a39204215a29 --- /dev/null +++ b/2024/Photorealistic Video Generation with Diffusion Models/images.zip @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:a1a56b1f018569d2d1fe94b0eef6f2e33dcd8ebb92353890fe4376dface048ea +size 571290 diff --git a/2024/Photorealistic Video Generation with Diffusion Models/layout.json b/2024/Photorealistic Video Generation with Diffusion Models/layout.json new file mode 100644 index 0000000000000000000000000000000000000000..b61b694995dbc4997a3d191fb3e84d7d1c5eea72 --- /dev/null +++ b/2024/Photorealistic Video Generation with Diffusion Models/layout.json @@ -0,0 +1,12072 @@ +{ + "pdf_info": [ + { + "para_blocks": [ + { + "bbox": [ + 142, + 111, + 473, + 146 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 142, + 111, + 473, + 146 + ], + "spans": [ + { + "bbox": [ + 142, + 111, + 473, + 146 + ], + "type": "text", + "content": "Photorealistic Video Generation with Diffusion Models" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 133, + 167, + 481, + 194 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 167, + 481, + 194 + ], + "spans": [ + { + "bbox": [ + 133, + 167, + 481, + 194 + ], + "type": "text", + "content": "Agrim Gupta" + }, + { + "bbox": [ + 133, + 167, + 481, + 194 + ], + "type": "inline_equation", + "content": "^{1,2,*}" + }, + { + "bbox": [ + 133, + 167, + 481, + 194 + ], + "type": "text", + "content": ", Lijun Yu" + }, + { + "bbox": [ + 133, + 167, + 481, + 194 + ], + "type": "inline_equation", + "content": "^{2}" + }, + { + "bbox": [ + 133, + 167, + 481, + 194 + ], + "type": "text", + "content": ", Kihyuk Sohn" + }, + { + "bbox": [ + 133, + 167, + 481, + 194 + ], + "type": "inline_equation", + "content": "^{2}" + }, + { + "bbox": [ + 133, + 167, + 481, + 194 + ], + "type": "text", + "content": ", Xiuye Gu" + }, + { + "bbox": [ + 133, + 167, + 481, + 194 + ], + "type": "inline_equation", + "content": "^{2}" + }, + { + "bbox": [ + 133, + 167, + 481, + 194 + ], + "type": "text", + "content": ", Meera Hahn" + }, + { + "bbox": [ + 133, + 167, + 481, + 194 + ], + "type": "inline_equation", + "content": "^{2}" + }, + { + "bbox": [ + 133, + 167, + 481, + 194 + ], + "type": "text", + "content": ", Fei-Fei Li" + }, + { + "bbox": [ + 133, + 167, + 481, + 194 + ], + "type": "inline_equation", + "content": "^{1}" + }, + { + "bbox": [ + 133, + 167, + 481, + 194 + ], + "type": "text", + "content": ", Irfan Essa" + }, + { + "bbox": [ + 133, + 167, + 481, + 194 + ], + "type": "inline_equation", + "content": "^{2,3}" + }, + { + "bbox": [ + 133, + 167, + 481, + 194 + ], + "type": "text", + "content": ", Lu Jiang" + }, + { + "bbox": [ + 133, + 167, + 481, + 194 + ], + "type": "inline_equation", + "content": "^{2}" + }, + { + "bbox": [ + 133, + 167, + 481, + 194 + ], + "type": "text", + "content": ", and José Lezama" + }, + { + "bbox": [ + 133, + 167, + 481, + 194 + ], + "type": "inline_equation", + "content": "^{2}" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 261, + 201, + 352, + 213 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 261, + 201, + 352, + 213 + ], + "spans": [ + { + "bbox": [ + 261, + 201, + 352, + 213 + ], + "type": "text", + "content": "1 Stanford University" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 268, + 213, + 345, + 224 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 268, + 213, + 345, + 224 + ], + "spans": [ + { + "bbox": [ + 268, + 213, + 345, + 224 + ], + "type": "text", + "content": "2 Google Research" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 237, + 224, + 376, + 236 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 237, + 224, + 376, + 236 + ], + "spans": [ + { + "bbox": [ + 237, + 224, + 376, + 236 + ], + "type": "text", + "content": "3 Georgia Institute of Technology" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 214, + 236, + 400, + 246 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 214, + 236, + 400, + 246 + ], + "spans": [ + { + "bbox": [ + 214, + 236, + 400, + 246 + ], + "type": "text", + "content": "https://walt-video-diffusion.github.io/" + } + ] + } + ], + "index": 5 + }, + { + "type": "image", + "bbox": [ + 133, + 258, + 202, + 299 + ], + "blocks": [ + { + "bbox": [ + 133, + 258, + 202, + 299 + ], + "lines": [ + { + "bbox": [ + 133, + 258, + 202, + 299 + ], + "spans": [ + { + "bbox": [ + 133, + 258, + 202, + 299 + ], + "type": "image", + "image_path": "19f0fe423e933f545060e36b9bc5b155b5f95cbd7e4131958491645bc98ed314.jpg" + } + ] + } + ], + "index": 6, + "angle": 0, + "type": "image_body" + } + ], + "index": 6 + }, + { + "type": "image", + "bbox": [ + 202, + 258, + 271, + 299 + ], + "blocks": [ + { + "bbox": [ + 202, + 258, + 271, + 299 + ], + "lines": [ + { + "bbox": [ + 202, + 258, + 271, + 299 + ], + "spans": [ + { + "bbox": [ + 202, + 258, + 271, + 299 + ], + "type": "image", + "image_path": "5826244976bd7e7d23345decf1df76c8d3b3d26043fad67d77e1210e3b7f2a50.jpg" + } + ] + } + ], + "index": 7, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 242, + 301, + 373, + 310 + ], + "lines": [ + { + "bbox": [ + 242, + 301, + 373, + 310 + ], + "spans": [ + { + "bbox": [ + 242, + 301, + 373, + 310 + ], + "type": "text", + "content": "A cute cat riding a yellow surfboard, fast waves." + } + ] + } + ], + "index": 11, + "angle": 0, + "type": "image_caption" + } + ], + "index": 7 + }, + { + "type": "image", + "bbox": [ + 271, + 258, + 340, + 299 + ], + "blocks": [ + { + "bbox": [ + 271, + 258, + 340, + 299 + ], + "lines": [ + { + "bbox": [ + 271, + 258, + 340, + 299 + ], + "spans": [ + { + "bbox": [ + 271, + 258, + 340, + 299 + ], + "type": "image", + "image_path": "421200607f84cd04c391726528a1354659e5e1d4219e97f2046347ee1947f759.jpg" + } + ] + } + ], + "index": 8, + "angle": 0, + "type": "image_body" + } + ], + "index": 8 + }, + { + "type": "image", + "bbox": [ + 340, + 258, + 409, + 299 + ], + "blocks": [ + { + "bbox": [ + 340, + 258, + 409, + 299 + ], + "lines": [ + { + "bbox": [ + 340, + 258, + 409, + 299 + ], + "spans": [ + { + "bbox": [ + 340, + 258, + 409, + 299 + ], + "type": "image", + "image_path": "b9fe99a791eb987a86b0054656e3ea8274efa49b4854bb716fbc4d2c3a139fc5.jpg" + } + ] + } + ], + "index": 9, + "angle": 0, + "type": "image_body" + } + ], + "index": 9 + }, + { + "type": "image", + "bbox": [ + 410, + 258, + 479, + 299 + ], + "blocks": [ + { + "bbox": [ + 410, + 258, + 479, + 299 + ], + "lines": [ + { + "bbox": [ + 410, + 258, + 479, + 299 + ], + "spans": [ + { + "bbox": [ + 410, + 258, + 479, + 299 + ], + "type": "image", + "image_path": "7464fe23b9b93ab68c099ebff01c971496686123e5d797a0d036f1814bfccff3.jpg" + } + ] + } + ], + "index": 10, + "angle": 0, + "type": "image_body" + } + ], + "index": 10 + }, + { + "type": "image", + "bbox": [ + 133, + 314, + 202, + 355 + ], + "blocks": [ + { + "bbox": [ + 133, + 314, + 202, + 355 + ], + "lines": [ + { + "bbox": [ + 133, + 314, + 202, + 355 + ], + "spans": [ + { + "bbox": [ + 133, + 314, + 202, + 355 + ], + "type": "image", + "image_path": "e6cee2ae0357022c32630f622964095cd5f0cc2c629d0bb0c9e33902137efdef.jpg" + } + ] + } + ], + "index": 12, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 195, + 360, + 410, + 369 + ], + "lines": [ + { + "bbox": [ + 195, + 360, + 410, + 369 + ], + "spans": [ + { + "bbox": [ + 195, + 360, + 410, + 369 + ], + "type": "text", + "content": "A cute cordi wearing a blue sweater, walking towards the camera, slow motion." + } + ] + } + ], + "index": 17, + "angle": 0, + "type": "image_caption" + } + ], + "index": 12 + }, + { + "type": "image", + "bbox": [ + 202, + 314, + 271, + 355 + ], + "blocks": [ + { + "bbox": [ + 202, + 314, + 271, + 355 + ], + "lines": [ + { + "bbox": [ + 202, + 314, + 271, + 355 + ], + "spans": [ + { + "bbox": [ + 202, + 314, + 271, + 355 + ], + "type": "image", + "image_path": "0e881ba7887a6586e6286ca479834a2fb97c0435e64e7e7a78608f5b9dc2e50c.jpg" + } + ] + } + ], + "index": 13, + "angle": 0, + "type": "image_body" + } + ], + "index": 13 + }, + { + "type": "image", + "bbox": [ + 271, + 314, + 340, + 355 + ], + "blocks": [ + { + "bbox": [ + 271, + 314, + 340, + 355 + ], + "lines": [ + { + "bbox": [ + 271, + 314, + 340, + 355 + ], + "spans": [ + { + "bbox": [ + 271, + 314, + 340, + 355 + ], + "type": "image", + "image_path": "5d495a6f51d47964d0c8865d31471091a93fcfc3c8c547f8a59119d564434fc5.jpg" + } + ] + } + ], + "index": 14, + "angle": 0, + "type": "image_body" + } + ], + "index": 14 + }, + { + "type": "image", + "bbox": [ + 340, + 314, + 409, + 355 + ], + "blocks": [ + { + "bbox": [ + 340, + 314, + 409, + 355 + ], + "lines": [ + { + "bbox": [ + 340, + 314, + 409, + 355 + ], + "spans": [ + { + "bbox": [ + 340, + 314, + 409, + 355 + ], + "type": "image", + "image_path": "cc33d11addaedb8aef7f3ce72761f04ba5e014dacd947edba35c98be0de4acdc.jpg" + } + ] + } + ], + "index": 15, + "angle": 0, + "type": "image_body" + } + ], + "index": 15 + }, + { + "type": "image", + "bbox": [ + 410, + 314, + 479, + 355 + ], + "blocks": [ + { + "bbox": [ + 410, + 314, + 479, + 355 + ], + "lines": [ + { + "bbox": [ + 410, + 314, + 479, + 355 + ], + "spans": [ + { + "bbox": [ + 410, + 314, + 479, + 355 + ], + "type": "image", + "image_path": "dce248c507f5b8a4286e0fd3a9cdc05a54ac6ef936246d06624948ddfd7a8b12.jpg" + } + ] + } + ], + "index": 16, + "angle": 0, + "type": "image_body" + } + ], + "index": 16 + }, + { + "type": "image", + "bbox": [ + 133, + 373, + 202, + 415 + ], + "blocks": [ + { + "bbox": [ + 133, + 373, + 202, + 415 + ], + "lines": [ + { + "bbox": [ + 133, + 373, + 202, + 415 + ], + "spans": [ + { + "bbox": [ + 133, + 373, + 202, + 415 + ], + "type": "image", + "image_path": "9184a5a762781a0348298aef64f01f53d2760718d9f6ebb2dabd9c4c35fc69c9.jpg" + } + ] + } + ], + "index": 18, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 193, + 419, + 416, + 428 + ], + "lines": [ + { + "bbox": [ + 193, + 419, + 416, + 428 + ], + "spans": [ + { + "bbox": [ + 193, + 419, + 416, + 428 + ], + "type": "text", + "content": "A raccoon wearing a black jacket, dancing in slow motion in front of the pyramids." + } + ] + } + ], + "index": 23, + "angle": 0, + "type": "image_caption" + } + ], + "index": 18 + }, + { + "type": "image", + "bbox": [ + 202, + 373, + 271, + 415 + ], + "blocks": [ + { + "bbox": [ + 202, + 373, + 271, + 415 + ], + "lines": [ + { + "bbox": [ + 202, + 373, + 271, + 415 + ], + "spans": [ + { + "bbox": [ + 202, + 373, + 271, + 415 + ], + "type": "image", + "image_path": "eb9aa9125b66742d3426c50f21429f2461b625f90fb6552b7410797b5d3b57ad.jpg" + } + ] + } + ], + "index": 19, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 130, + 441, + 482, + 474 + ], + "lines": [ + { + "bbox": [ + 130, + 441, + 482, + 474 + ], + "spans": [ + { + "bbox": [ + 130, + 441, + 482, + 474 + ], + "type": "text", + "content": "Fig. 1: W.A.L.T samples for text-to-video generation. W.A.L.T generates high-resolution, temporally consistent photorealistic videos from text prompts. The samples shown are " + }, + { + "bbox": [ + 130, + 441, + 482, + 474 + ], + "type": "inline_equation", + "content": "512 \\times 896" + }, + { + "bbox": [ + 130, + 441, + 482, + 474 + ], + "type": "text", + "content": " resolution over 3.6 seconds duration at 8 frames per second." + } + ] + } + ], + "index": 24, + "angle": 0, + "type": "image_caption" + } + ], + "index": 19 + }, + { + "type": "image", + "bbox": [ + 271, + 373, + 340, + 415 + ], + "blocks": [ + { + "bbox": [ + 271, + 373, + 340, + 415 + ], + "lines": [ + { + "bbox": [ + 271, + 373, + 340, + 415 + ], + "spans": [ + { + "bbox": [ + 271, + 373, + 340, + 415 + ], + "type": "image", + "image_path": "c69d86db9537d85db7a46a005601f41462aa78e40b73b2c333d0fb77d8393e75.jpg" + } + ] + } + ], + "index": 20, + "angle": 0, + "type": "image_body" + } + ], + "index": 20 + }, + { + "type": "image", + "bbox": [ + 340, + 373, + 409, + 415 + ], + "blocks": [ + { + "bbox": [ + 340, + 373, + 409, + 415 + ], + "lines": [ + { + "bbox": [ + 340, + 373, + 409, + 415 + ], + "spans": [ + { + "bbox": [ + 340, + 373, + 409, + 415 + ], + "type": "image", + "image_path": "fce371586e81973d39130c746f40e2130e3824b4a98199858d584bde1b879b81.jpg" + } + ] + } + ], + "index": 21, + "angle": 0, + "type": "image_body" + } + ], + "index": 21 + }, + { + "type": "image", + "bbox": [ + 410, + 373, + 479, + 415 + ], + "blocks": [ + { + "bbox": [ + 410, + 373, + 479, + 415 + ], + "lines": [ + { + "bbox": [ + 410, + 373, + 479, + 415 + ], + "spans": [ + { + "bbox": [ + 410, + 373, + 479, + 415 + ], + "type": "image", + "image_path": "4639bffb1849a85039b93335e4ae288a124be8e418f63e11bf3d13e79f436aee.jpg" + } + ] + } + ], + "index": 22, + "angle": 0, + "type": "image_body" + } + ], + "index": 22 + }, + { + "bbox": [ + 159, + 489, + 455, + 632 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 159, + 489, + 455, + 632 + ], + "spans": [ + { + "bbox": [ + 159, + 489, + 455, + 632 + ], + "type": "text", + "content": "Abstract. We present W.A.L.T, a diffusion transformer for photorealistic video generation from text prompts. Our approach has two key design decisions. First, we use a causal encoder to jointly compress images and videos within a unified latent space, enabling training and generation across modalities. Second, for memory and training efficiency, we use a window attention architecture tailored for joint spatial and spatiotemporal generative modeling. Taken together these design decisions enable us to achieve state-of-the-art performance on established video (UCF-101 and Kinetics-600) and image (ImageNet) generation benchmarks without using classifier free guidance. Finally, we also train a cascade of three models for the task of text-to-video generation consisting of a base latent video diffusion model, and two video super-resolution diffusion models to generate videos of " + }, + { + "bbox": [ + 159, + 489, + 455, + 632 + ], + "type": "inline_equation", + "content": "512 \\times 896" + }, + { + "bbox": [ + 159, + 489, + 455, + 632 + ], + "type": "text", + "content": " resolution at 8 frames per second." + } + ] + } + ], + "index": 25 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 142, + 654, + 359, + 665 + ], + "type": "page_footnote", + "angle": 0, + "lines": [ + { + "bbox": [ + 142, + 654, + 359, + 665 + ], + "spans": [ + { + "bbox": [ + 142, + 654, + 359, + 665 + ], + "type": "text", + "content": "*Work partially done during an internship at Google." + } + ] + } + ], + "index": 26 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 0 + }, + { + "para_blocks": [ + { + "bbox": [ + 133, + 114, + 230, + 128 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 114, + 230, + 128 + ], + "spans": [ + { + "bbox": [ + 133, + 114, + 230, + 128 + ], + "type": "text", + "content": "1 Introduction" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 130, + 138, + 481, + 186 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 138, + 481, + 186 + ], + "spans": [ + { + "bbox": [ + 130, + 138, + 481, + 186 + ], + "type": "text", + "content": "Modeling the complex and dynamic visual world with a high level of detail is the ultimate goal of visual generative models. To accomplish this goal, we need neural architectures that can efficiently process high dimensional inputs, capture long range dependencies and have powerful scaling and parallelization properties." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 130, + 187, + 481, + 270 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 187, + 481, + 270 + ], + "spans": [ + { + "bbox": [ + 130, + 187, + 481, + 270 + ], + "type": "text", + "content": "The transformer [74] is a neural network architecture designed with such desirable properties. This has encouraged the research community to increasingly favor it over domain-specific architectures in diverse fields such as language [28, 56, 58, 59], audio [1], speech [57], vision [19, 32], and robotics [6, 8, 91]. Such a trend towards unification allows researchers to share and build upon advances in traditionally disparate domains. Thus, leading to a virtuous cycle of innovation and improvement in model design favoring transformers." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 130, + 271, + 481, + 376 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 271, + 481, + 376 + ], + "spans": [ + { + "bbox": [ + 130, + 271, + 481, + 376 + ], + "type": "text", + "content": "Generative modelling of images and videos has been however a notable exception to this trend. Diffusion models [68, 70], which have emerged as a leading paradigm for generative modelling of images [17, 36] and videos [38], predominantly rely on the U-Net architecture [36, 63] as backbone, consisting of a series of convolutional [47] and self-attention [74] layers. This preference stems from the fact that the memory demands of full attention mechanisms in transformers scales quadratically with input sequence length. Such scaling leads to prohibitively high costs when processing high-dimensional signals like video with transformers." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 130, + 379, + 481, + 544 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 379, + 481, + 544 + ], + "spans": [ + { + "bbox": [ + 130, + 379, + 481, + 544 + ], + "type": "text", + "content": "The computational requirements of diffusion models can be reduced by operating in a lower-dimensional latent space derived from an autoencoder [21,73,76], giving rise to latent diffusion models (LDMs) [62]. A critical design choice in this context is the type of latent space employed: spatial compression (per frame latents) versus spatiotemporal compression. Spatial compression is often preferred because it enables leveraging pre-trained image autoencoders and LDMs, which are trained on large paired image-text datasets. However, this choice increases network complexity and limits the use of transformers as backbones, especially in generating high-resolution videos due to memory constraints. On the other hand, while spatiotemporal compression can mitigate these issues, it precludes the use of paired image-text datasets, which are much larger and diverse than their video counterparts. Hence, to leverage all sources of visual data and unlock scaling via transformers, it is critical to have a unified compressed representation of images and videos." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 130, + 545, + 481, + 665 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 545, + 481, + 665 + ], + "spans": [ + { + "bbox": [ + 130, + 545, + 481, + 665 + ], + "type": "text", + "content": "In this work, we present Window Attention Latent Transformer (W.A.L.T): a novel transformer-based method for latent video diffusion models (LVDMs). Our method consists of two stages. First, an autoencoder maps both videos and images into a unified, lower-dimensional latent space. This design choice enables training a single generative model jointly on image and video datasets and significantly reduces the computational burden for generating high resolution videos. Subsequently, we propose a new design of transformer blocks for latent video diffusion modeling composed of self-attention layers that alternate between non-overlapping, window-restricted spatial and spatiotemporal attention. This design offers two primary benefits: Firstly, the use of local window attention" + } + ] + } + ], + "index": 7 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "spans": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "type": "text", + "content": "2" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 166, + 91, + 230, + 102 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 166, + 91, + 230, + 102 + ], + "spans": [ + { + "bbox": [ + 166, + 91, + 230, + 102 + ], + "type": "text", + "content": "A. Gupta et al." + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 1 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 132, + 114, + 478, + 219 + ], + "blocks": [ + { + "bbox": [ + 132, + 114, + 478, + 219 + ], + "lines": [ + { + "bbox": [ + 132, + 114, + 478, + 219 + ], + "spans": [ + { + "bbox": [ + 132, + 114, + 478, + 219 + ], + "type": "image", + "image_path": "bfe32dabb32c1ccdeba7d70de9370b595c60b91393d21684e6d5d5b0109dc573.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 130, + 226, + 482, + 282 + ], + "lines": [ + { + "bbox": [ + 130, + 226, + 482, + 282 + ], + "spans": [ + { + "bbox": [ + 130, + 226, + 482, + 282 + ], + "type": "text", + "content": "Fig. 2: W.A.L.T. We encode images and videos into a shared latent space. The transformer backbone processes these latents with blocks having two layers of window-restricted attention: spatial layers capture spatial relations in both images and video, while spatiotemporal layers model temporal dynamics in videos and pass through images via identity attention mask. Text conditioning is done via spatial cross-attention." + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_caption" + } + ], + "index": 2 + }, + { + "bbox": [ + 130, + 297, + 482, + 343 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 297, + 482, + 343 + ], + "spans": [ + { + "bbox": [ + 130, + 297, + 482, + 343 + ], + "type": "text", + "content": "significantly lowers computational demands. Secondly, it facilitates joint training, where the spatial layers independently process images and video frames, while the spatiotemporal layers are dedicated to modeling the temporal relationships in videos." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 130, + 344, + 482, + 475 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 344, + 482, + 475 + ], + "spans": [ + { + "bbox": [ + 130, + 344, + 482, + 475 + ], + "type": "text", + "content": "While conceptually simple, our method provides the first empirical evidence of transformers' superior generation quality and parameter efficiency in latent video diffusion on public benchmarks. Specifically, we report state-of-the-art results on class-conditional video generation (UCF-101 [71]), frame prediction (Kinetics-600 [10]) and class conditional image generation (ImageNet [16]) without using classifier free guidance. Finally, to showcase the scalability and efficiency of our method we also demonstrate results on the challenging task of photorealistic text-to-video generation. We train a cascade of three models consisting of a base latent video diffusion model, and two video super-resolution diffusion models to generate videos of " + }, + { + "bbox": [ + 130, + 344, + 482, + 475 + ], + "type": "inline_equation", + "content": "512 \\times 896" + }, + { + "bbox": [ + 130, + 344, + 482, + 475 + ], + "type": "text", + "content": " resolution at 8 frames per second and report state-of-the-art zero-shot FVD score on the UCF-101 benchmark." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 146, + 476, + 364, + 488 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 146, + 476, + 364, + 488 + ], + "spans": [ + { + "bbox": [ + 146, + 476, + 364, + 488 + ], + "type": "text", + "content": "Our contributions can be summarized as follows," + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 137, + 495, + 480, + 578 + ], + "type": "list", + "angle": 0, + "index": 10, + "blocks": [ + { + "bbox": [ + 138, + 495, + 479, + 518 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 495, + 479, + 518 + ], + "spans": [ + { + "bbox": [ + 138, + 495, + 479, + 518 + ], + "type": "text", + "content": "- We propose W.A.L.T, a framework for transformer-based latent diffusion in a unified image and video latent space." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 137, + 519, + 480, + 555 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 137, + 519, + 480, + 555 + ], + "spans": [ + { + "bbox": [ + 137, + 519, + 480, + 555 + ], + "type": "text", + "content": "- We validate W.A.L.T by performing extensive ablations and achieving state-of-the-art video generation on frame prediction (K600), class-conditional video (UCF-101) and image (ImageNet) generation." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 137, + 555, + 480, + 578 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 137, + 555, + 480, + 578 + ], + "spans": [ + { + "bbox": [ + 137, + 555, + 480, + 578 + ], + "type": "text", + "content": "- We demonstrate the scaling properties of W.A.L.T and achieve state-of-the-art results in zero-shot UCF-101 text-to-video benchmark." + } + ] + } + ], + "index": 9 + } + ], + "sub_type": "text" + }, + { + "bbox": [ + 132, + 605, + 237, + 617 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 605, + 237, + 617 + ], + "spans": [ + { + "bbox": [ + 132, + 605, + 237, + 617 + ], + "type": "text", + "content": "2 Related Work" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 130, + 629, + 481, + 667 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 629, + 481, + 667 + ], + "spans": [ + { + "bbox": [ + 130, + 629, + 481, + 667 + ], + "type": "text", + "content": "Video Diffusion Models. Diffusion models have shown impressive results in image [36,40,53,62,68,69] and video generation [5,25,31,35,38,67]. Video diffusion models can be categorized into pixel-space [35,38,67] and latent-space [5,25,33,84]" + } + ] + } + ], + "index": 12 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 223, + 91, + 448, + 101 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 223, + 91, + 448, + 101 + ], + "spans": [ + { + "bbox": [ + 223, + 91, + 448, + 101 + ], + "type": "text", + "content": "Photorealistic Video Generation with Diffusion Models" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 474, + 92, + 480, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 474, + 92, + 480, + 100 + ], + "spans": [ + { + "bbox": [ + 474, + 92, + 480, + 100 + ], + "type": "text", + "content": "3" + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 2 + }, + { + "para_blocks": [ + { + "bbox": [ + 130, + 116, + 482, + 248 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 116, + 482, + 248 + ], + "spans": [ + { + "bbox": [ + 130, + 116, + 482, + 248 + ], + "type": "text", + "content": "approaches, the later bringing important efficiency advantages when modeling videos. Ho et al. [38] demonstrated that the quality of text conditioned video generation can be significantly improved by jointly training on image and video data. Similarly, to leverage image datasets, latent video diffusion models inflate a pre-trained image model, typically a U-Net [63], into a video model by adding temporal layers, and initializing them as the identity function [5,35,67]. Although computationally efficient, this approach couples the design of video and image models, and precludes spatiotemporal compression. In this work, we operate on a unified latent space for images and videos, allowing us to leverage large scale image and video datasets while enjoying computational efficiency gains from spatiotemporal compression of videos." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 130, + 258, + 483, + 486 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 258, + 483, + 486 + ], + "spans": [ + { + "bbox": [ + 130, + 258, + 483, + 486 + ], + "type": "text", + "content": "Transformers for Generative Modeling. Multiple classes of generative models have utilized Transformers [74] as backbone, such as, Generative adversarial networks [43, 48, 87], autoregressive [11, 12, 21, 22, 29, 60, 75, 78, 79, 81, 82] and diffusion [2, 23, 42, 51, 54, 89] models. Inspired by the success of autoregressive pretraining of large language models [56, 58, 59], Ramesh et al. [60] trained a text-to-image generation model by predicting the next visual token obtained from an image tokenizer. Subsequently, this approach was applied to multiple applications including class-conditional image generation [21, 80], text-to-image [18, 60, 77, 81] or image-to-image translation [22, 78]. Similarly, for video generation, transformer-based models were proposed to predict next tokens using 3D extensions of VQGAN [24, 39, 79, 82] or using per frame image latents [29]. Autoregressive sampling of videos is typically impractical given the very long sequences involved. To alleviate this issue, non-autoregressive sampling [11, 12], i.e. parallel token prediction, has been adopted as a more efficient solution for transformer-based video generation [29, 75, 82]. Recently, the community has started adopting transformers as the denoising backbone for diffusion models in place of U-Net [13, 40, 51, 54, 89]. To the best of our knowledge, our work is the first successful empirical demonstration (§ 5.1) of a transformer-based backbone for jointly training image and video latent diffusion models." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 132, + 503, + 227, + 517 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 503, + 227, + 517 + ], + "spans": [ + { + "bbox": [ + 132, + 503, + 227, + 517 + ], + "type": "text", + "content": "3 Background" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 130, + 527, + 482, + 576 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 527, + 482, + 576 + ], + "spans": [ + { + "bbox": [ + 130, + 527, + 482, + 576 + ], + "type": "text", + "content": "Diffusion formulation. Diffusion models [36, 68, 70] are a class of generative models which learn to generate data by iteratively denoising samples drawn from a noise distribution. Gaussian diffusion models assume a forward noising process which gradually applies noise " + }, + { + "bbox": [ + 130, + 527, + 482, + 576 + ], + "type": "inline_equation", + "content": "(\\epsilon)" + }, + { + "bbox": [ + 130, + 527, + 482, + 576 + ], + "type": "text", + "content": " to real data " + }, + { + "bbox": [ + 130, + 527, + 482, + 576 + ], + "type": "inline_equation", + "content": "(x_0 \\sim p_{\\mathrm{data}})" + }, + { + "bbox": [ + 130, + 527, + 482, + 576 + ], + "type": "text", + "content": ". Concretely," + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 240, + 584, + 481, + 597 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 240, + 584, + 481, + 597 + ], + "spans": [ + { + "bbox": [ + 240, + 584, + 481, + 597 + ], + "type": "interline_equation", + "content": "\\boldsymbol {x} _ {t} = \\sqrt {\\gamma (t)} \\boldsymbol {x} _ {0} + \\sqrt {1 - \\gamma (t)} \\epsilon , \\tag {1}", + "image_path": "81c82d73683dd0b9e8b8ca61c4e74145737cd7c15e21b1a44a7309916abe91cf.jpg" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 130, + 605, + 482, + 642 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 605, + 482, + 642 + ], + "spans": [ + { + "bbox": [ + 130, + 605, + 482, + 642 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 130, + 605, + 482, + 642 + ], + "type": "inline_equation", + "content": "\\epsilon \\sim \\mathcal{N}(\\mathbf{0},\\mathbf{I}), t \\in [0,1]" + }, + { + "bbox": [ + 130, + 605, + 482, + 642 + ], + "type": "text", + "content": ", and " + }, + { + "bbox": [ + 130, + 605, + 482, + 642 + ], + "type": "inline_equation", + "content": "\\gamma(t)" + }, + { + "bbox": [ + 130, + 605, + 482, + 642 + ], + "type": "text", + "content": " is a monotonically decreasing function (noise schedule) from 1 to 0. Diffusion models are trained to learn the reverse process that inverts the forward corruptions:" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 206, + 649, + 481, + 668 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 206, + 649, + 481, + 668 + ], + "spans": [ + { + "bbox": [ + 206, + 649, + 481, + 668 + ], + "type": "interline_equation", + "content": "\\mathbb {E} _ {\\boldsymbol {x} \\sim p _ {\\text {d a t a}}, t \\sim \\mathcal {U} (0, 1), \\boldsymbol {\\epsilon} \\sim \\mathcal {N} (\\boldsymbol {0}, \\boldsymbol {I})} \\left[ \\| \\boldsymbol {y} - f _ {\\theta} (\\boldsymbol {x} _ {t}; \\boldsymbol {c}, t) \\| ^ {2} \\right], \\tag {2}", + "image_path": "add488d322cc1b1c4c9d39fe680471151329ab02630f51975b4af33e67449d01.jpg" + } + ] + } + ], + "index": 8 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "spans": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "type": "text", + "content": "4" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 166, + 91, + 230, + 102 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 166, + 91, + 230, + 102 + ], + "spans": [ + { + "bbox": [ + 166, + 91, + 230, + 102 + ], + "type": "text", + "content": "A. Gupta et al." + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 3 + }, + { + "para_blocks": [ + { + "bbox": [ + 130, + 116, + 482, + 163 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 116, + 482, + 163 + ], + "spans": [ + { + "bbox": [ + 130, + 116, + 482, + 163 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 130, + 116, + 482, + 163 + ], + "type": "inline_equation", + "content": "f_{\\theta}" + }, + { + "bbox": [ + 130, + 116, + 482, + 163 + ], + "type": "text", + "content": " is the denoiser model parameterized by a neural network, " + }, + { + "bbox": [ + 130, + 116, + 482, + 163 + ], + "type": "inline_equation", + "content": "\\pmb{c}" + }, + { + "bbox": [ + 130, + 116, + 482, + 163 + ], + "type": "text", + "content": " is conditioning information e.g., class labels or text prompts, and the target " + }, + { + "bbox": [ + 130, + 116, + 482, + 163 + ], + "type": "inline_equation", + "content": "\\pmb{y}" + }, + { + "bbox": [ + 130, + 116, + 482, + 163 + ], + "type": "text", + "content": " can be random noise " + }, + { + "bbox": [ + 130, + 116, + 482, + 163 + ], + "type": "inline_equation", + "content": "\\pmb{\\epsilon}" + }, + { + "bbox": [ + 130, + 116, + 482, + 163 + ], + "type": "text", + "content": ", denoised input " + }, + { + "bbox": [ + 130, + 116, + 482, + 163 + ], + "type": "inline_equation", + "content": "\\pmb{x_0}" + }, + { + "bbox": [ + 130, + 116, + 482, + 163 + ], + "type": "text", + "content": " or " + }, + { + "bbox": [ + 130, + 116, + 482, + 163 + ], + "type": "inline_equation", + "content": "\\pmb{v} = \\sqrt{1 - \\gamma(t)} \\pmb{\\epsilon} - \\sqrt{\\gamma(t)} \\pmb{x_0}" + }, + { + "bbox": [ + 130, + 116, + 482, + 163 + ], + "type": "text", + "content": ". Following [35, 65], we use " + }, + { + "bbox": [ + 130, + 116, + 482, + 163 + ], + "type": "inline_equation", + "content": "\\pmb{v}" + }, + { + "bbox": [ + 130, + 116, + 482, + 163 + ], + "type": "text", + "content": "-prediction in all our experiments." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 130, + 164, + 483, + 297 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 164, + 483, + 297 + ], + "spans": [ + { + "bbox": [ + 130, + 164, + 483, + 297 + ], + "type": "text", + "content": "Latent diffusion models (LDMs). Processing high-resolution images and videos using raw pixels requires considerable computational resources. To address this, LDMs operate on the low dimensional latent space of a VQ-VAE [21,73]. VQ-VAE consists of an encoder " + }, + { + "bbox": [ + 130, + 164, + 483, + 297 + ], + "type": "inline_equation", + "content": "E(\\pmb{x})" + }, + { + "bbox": [ + 130, + 164, + 483, + 297 + ], + "type": "text", + "content": " that encodes an input video " + }, + { + "bbox": [ + 130, + 164, + 483, + 297 + ], + "type": "inline_equation", + "content": "\\pmb{x} \\in \\mathbb{R}^{T \\times H \\times W \\times 3}" + }, + { + "bbox": [ + 130, + 164, + 483, + 297 + ], + "type": "text", + "content": " into a latent representation " + }, + { + "bbox": [ + 130, + 164, + 483, + 297 + ], + "type": "inline_equation", + "content": "\\pmb{z} \\in \\mathbb{R}^{t \\times h \\times w \\times c}" + }, + { + "bbox": [ + 130, + 164, + 483, + 297 + ], + "type": "text", + "content": ". The encoder downsamples the video by a factor of " + }, + { + "bbox": [ + 130, + 164, + 483, + 297 + ], + "type": "inline_equation", + "content": "f_{s} = H / h = W / w" + }, + { + "bbox": [ + 130, + 164, + 483, + 297 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 130, + 164, + 483, + 297 + ], + "type": "inline_equation", + "content": "f_{t} = T / t" + }, + { + "bbox": [ + 130, + 164, + 483, + 297 + ], + "type": "text", + "content": ", where " + }, + { + "bbox": [ + 130, + 164, + 483, + 297 + ], + "type": "inline_equation", + "content": "T = t = 1" + }, + { + "bbox": [ + 130, + 164, + 483, + 297 + ], + "type": "text", + "content": " corresponds to using an image autoencoder. An important distinction from the original VQ-VAE is the absence of a codebook of quantized embeddings as diffusion models can operate on continuous latent spaces. A decoder " + }, + { + "bbox": [ + 130, + 164, + 483, + 297 + ], + "type": "inline_equation", + "content": "D" + }, + { + "bbox": [ + 130, + 164, + 483, + 297 + ], + "type": "text", + "content": " is trained to predict a reconstruction of the video, " + }, + { + "bbox": [ + 130, + 164, + 483, + 297 + ], + "type": "inline_equation", + "content": "\\hat{\\pmb{x}}" + }, + { + "bbox": [ + 130, + 164, + 483, + 297 + ], + "type": "text", + "content": ", from " + }, + { + "bbox": [ + 130, + 164, + 483, + 297 + ], + "type": "inline_equation", + "content": "\\pmb{z}" + }, + { + "bbox": [ + 130, + 164, + 483, + 297 + ], + "type": "text", + "content": ". Following VQ-GAN [21], reconstruction quality can be further improved by adding adversarial [27] and perceptual losses [44,88]." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 131, + 312, + 211, + 324 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 131, + 312, + 211, + 324 + ], + "spans": [ + { + "bbox": [ + 131, + 312, + 211, + 324 + ], + "type": "text", + "content": "4 W.A.L.T" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 131, + 336, + 279, + 348 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 131, + 336, + 279, + 348 + ], + "spans": [ + { + "bbox": [ + 131, + 336, + 279, + 348 + ], + "type": "text", + "content": "4.1 Learning Visual Tokens" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 130, + 354, + 483, + 498 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 354, + 483, + 498 + ], + "spans": [ + { + "bbox": [ + 130, + 354, + 483, + 498 + ], + "type": "text", + "content": "A key design decision in video generative modeling is the choice of latent space representation. Ideally, we want a shared and unified compressed visual representation that can be used for generative modeling of both images and videos [75,83]. The unified representation is important because joint image-video learning is preferable due to a scarcity of labeled video data [35], such as text-video pairs. Concretely, given a video sequence " + }, + { + "bbox": [ + 130, + 354, + 483, + 498 + ], + "type": "inline_equation", + "content": "\\pmb{x} \\in \\mathbb{R}^{(1 + T) \\times H \\times W \\times C}" + }, + { + "bbox": [ + 130, + 354, + 483, + 498 + ], + "type": "text", + "content": ", we aim to learn a low-dimensional representation " + }, + { + "bbox": [ + 130, + 354, + 483, + 498 + ], + "type": "inline_equation", + "content": "\\pmb{z} \\in \\mathbb{R}^{(1 + t) \\times h \\times w \\times c}" + }, + { + "bbox": [ + 130, + 354, + 483, + 498 + ], + "type": "text", + "content": " that performs spatial-temporal compression by a factor of " + }, + { + "bbox": [ + 130, + 354, + 483, + 498 + ], + "type": "inline_equation", + "content": "f_{s} = H / h = W / w" + }, + { + "bbox": [ + 130, + 354, + 483, + 498 + ], + "type": "text", + "content": " in space and a factor of " + }, + { + "bbox": [ + 130, + 354, + 483, + 498 + ], + "type": "inline_equation", + "content": "f_{t} = T / t" + }, + { + "bbox": [ + 130, + 354, + 483, + 498 + ], + "type": "text", + "content": " in time. To enable a unified representation for both videos and static images, the first frame is always encoded independently from the rest of the video. This allows static images " + }, + { + "bbox": [ + 130, + 354, + 483, + 498 + ], + "type": "inline_equation", + "content": "\\pmb{x} \\in \\mathbb{R}^{1 \\times H \\times W \\times C}" + }, + { + "bbox": [ + 130, + 354, + 483, + 498 + ], + "type": "text", + "content": " to be treated as videos with a single frame, i.e. " + }, + { + "bbox": [ + 130, + 354, + 483, + 498 + ], + "type": "inline_equation", + "content": "\\pmb{z} \\in \\mathbb{R}^{1 \\times h \\times w \\times c}" + }, + { + "bbox": [ + 130, + 354, + 483, + 498 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 130, + 498, + 482, + 605 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 498, + 482, + 605 + ], + "spans": [ + { + "bbox": [ + 130, + 498, + 482, + 605 + ], + "type": "text", + "content": "We instantiate this design with the causal 3D CNN encoder-decoder architecture of the MAGVIT-v2 tokenizer [83]. Typically the encoder-decoder consists of regular 3D convolution layers which cannot process the first frame independently [24,82]. This limitation stems from the fact that a regular convolutional kernel of size " + }, + { + "bbox": [ + 130, + 498, + 482, + 605 + ], + "type": "inline_equation", + "content": "(k_{t},k_{h},k_{w})" + }, + { + "bbox": [ + 130, + 498, + 482, + 605 + ], + "type": "text", + "content": " will operate on " + }, + { + "bbox": [ + 130, + 498, + 482, + 605 + ], + "type": "inline_equation", + "content": "\\left\\lfloor \\frac{k_t - 1}{2}\\right\\rfloor" + }, + { + "bbox": [ + 130, + 498, + 482, + 605 + ], + "type": "text", + "content": " frames before and " + }, + { + "bbox": [ + 130, + 498, + 482, + 605 + ], + "type": "inline_equation", + "content": "\\left\\lfloor \\frac{k_t}{2}\\right\\rfloor" + }, + { + "bbox": [ + 130, + 498, + 482, + 605 + ], + "type": "text", + "content": " frames after the input frames. Causal 3D convolution layers solve this issue as the convolutional kernel operates on only the past " + }, + { + "bbox": [ + 130, + 498, + 482, + 605 + ], + "type": "inline_equation", + "content": "k_{t} - 1" + }, + { + "bbox": [ + 130, + 498, + 482, + 605 + ], + "type": "text", + "content": " frames. This ensures that the output for each frame is influenced solely by the preceding frames, enabling the model to tokenize the first frame independently." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 130, + 606, + 482, + 665 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 606, + 482, + 665 + ], + "spans": [ + { + "bbox": [ + 130, + 606, + 482, + 665 + ], + "type": "text", + "content": "After this stage, the input to our model is a batch of latent tensors " + }, + { + "bbox": [ + 130, + 606, + 482, + 665 + ], + "type": "inline_equation", + "content": "\\mathbf{z} \\in \\mathbb{R}^{(1 + t) \\times h \\times w \\times c}" + }, + { + "bbox": [ + 130, + 606, + 482, + 665 + ], + "type": "text", + "content": " representing a single video or a stack of " + }, + { + "bbox": [ + 130, + 606, + 482, + 665 + ], + "type": "inline_equation", + "content": "1 + t" + }, + { + "bbox": [ + 130, + 606, + 482, + 665 + ], + "type": "text", + "content": " independent images (Fig. 2). Different from [83], our latent representation is real-valued and quantization-free. In the section below we describe how our model jointly processes a mixed batch of images and videos." + } + ] + } + ], + "index": 8 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 223, + 90, + 448, + 101 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 223, + 90, + 448, + 101 + ], + "spans": [ + { + "bbox": [ + 223, + 90, + 448, + 101 + ], + "type": "text", + "content": "Photorealistic Video Generation with Diffusion Models" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 474, + 92, + 480, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 474, + 92, + 480, + 100 + ], + "spans": [ + { + "bbox": [ + 474, + 92, + 480, + 100 + ], + "type": "text", + "content": "5" + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 4 + }, + { + "para_blocks": [ + { + "bbox": [ + 132, + 116, + 369, + 128 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 116, + 369, + 128 + ], + "spans": [ + { + "bbox": [ + 132, + 116, + 369, + 128 + ], + "type": "text", + "content": "4.2 Learning to Generate Images and Videos" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 130, + 140, + 482, + 223 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 140, + 482, + 223 + ], + "spans": [ + { + "bbox": [ + 130, + 140, + 482, + 223 + ], + "type": "text", + "content": "Patchify. Following the original ViT [19], we \"patchify\" each latent frame independently by converting it into a sequence of non-overlapping " + }, + { + "bbox": [ + 130, + 140, + 482, + 223 + ], + "type": "inline_equation", + "content": "h_p \\times w_p" + }, + { + "bbox": [ + 130, + 140, + 482, + 223 + ], + "type": "text", + "content": " patches where " + }, + { + "bbox": [ + 130, + 140, + 482, + 223 + ], + "type": "inline_equation", + "content": "h_p = h/p" + }, + { + "bbox": [ + 130, + 140, + 482, + 223 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 130, + 140, + 482, + 223 + ], + "type": "inline_equation", + "content": "w_p = w/p" + }, + { + "bbox": [ + 130, + 140, + 482, + 223 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 130, + 140, + 482, + 223 + ], + "type": "inline_equation", + "content": "p" + }, + { + "bbox": [ + 130, + 140, + 482, + 223 + ], + "type": "text", + "content": " is the patch size. We use learnable positional embeddings [74], which are the sum of space and time positional embeddings. Position embeddings are added to the linear projections [19] of the patches. Note that for images, we simply add the temporal position embedding corresponding to the first latent frame." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 130, + 224, + 482, + 380 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 224, + 482, + 380 + ], + "spans": [ + { + "bbox": [ + 130, + 224, + 482, + 380 + ], + "type": "text", + "content": "Window attention. Transformer models composed entirely of global self-attention modules incur significant compute and memory costs, especially for video tasks. For efficiency and for processing images and videos jointly we compute self-attention in windows [29, 74], based on two types of non-overlapping configurations: spatial (S) and spatiotemporal (ST), cf. Fig. 2. Spatial Window (SW) attention is restricted to all the tokens within a latent frame of size " + }, + { + "bbox": [ + 130, + 224, + 482, + 380 + ], + "type": "inline_equation", + "content": "1 \\times h_p \\times w_p" + }, + { + "bbox": [ + 130, + 224, + 482, + 380 + ], + "type": "text", + "content": " (the first dimension is time). SW models the spatial relations in images and videos. Spatiotemporal Window (STW) attention is restricted within a 3D window of size " + }, + { + "bbox": [ + 130, + 224, + 482, + 380 + ], + "type": "inline_equation", + "content": "(1 + t) \\times h_p' \\times h_w'" + }, + { + "bbox": [ + 130, + 224, + 482, + 380 + ], + "type": "text", + "content": ", modeling the temporal relationships among video latent frames. For images, we simply use an identity attention mask ensuring that the value embeddings corresponding to the image frame latents are passed through the layer as is. Finally, in addition to absolute position embeddings we also use relative position embeddings [50]." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 130, + 380, + 483, + 464 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 380, + 483, + 464 + ], + "spans": [ + { + "bbox": [ + 130, + 380, + 483, + 464 + ], + "type": "text", + "content": "Our design, while conceptually straightforward, achieves computational efficiency and enables joint training on image and video datasets. In contrast to methods based on frame-level autoencoders [5, 25, 29], our approach does not suffer from flickering artifacts, which often result from encoding and decoding video frames independently. However, similar to Blattmann et al. [5], we can also potentially leverage pre-trained image LDMs with transformer backbones by simply interleaving STW layers." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 132, + 485, + 280, + 496 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 485, + 280, + 496 + ], + "spans": [ + { + "bbox": [ + 132, + 485, + 280, + 496 + ], + "type": "text", + "content": "4.3 Conditional Generation" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 130, + 509, + 482, + 568 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 509, + 482, + 568 + ], + "spans": [ + { + "bbox": [ + 130, + 509, + 482, + 568 + ], + "type": "text", + "content": "To enable controllable video generation, in addition to conditioning on timestep " + }, + { + "bbox": [ + 130, + 509, + 482, + 568 + ], + "type": "inline_equation", + "content": "t" + }, + { + "bbox": [ + 130, + 509, + 482, + 568 + ], + "type": "text", + "content": ", diffusion models are often conditioned on additional conditional information " + }, + { + "bbox": [ + 130, + 509, + 482, + 568 + ], + "type": "inline_equation", + "content": "\\mathbf{c}" + }, + { + "bbox": [ + 130, + 509, + 482, + 568 + ], + "type": "text", + "content": " such as class labels, natural language, past frames or low resolution videos. In our transformer backbone, we incorporate three types of conditioning mechanisms as described in what follows:" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 130, + 570, + 482, + 666 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 570, + 482, + 666 + ], + "spans": [ + { + "bbox": [ + 130, + 570, + 482, + 666 + ], + "type": "text", + "content": "Cross-attention. In addition to self-attention layers in our window transformer blocks, we add a cross-attention layer for text conditioned generation. When training models on just videos, the cross-attention layer employs the same window-restricted attention as the self-attention layer, meaning S/ST blocks will have SW/STW cross-attention layers (Fig. 2). However, for joint training, we only use SW cross-attention layers. For cross-attention we concatenate the input signal (query) with the conditioning signal (key, value) as our early experiments showed this improves performance." + } + ] + } + ], + "index": 8 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "spans": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "type": "text", + "content": "6" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 166, + 91, + 230, + 102 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 166, + 91, + 230, + 102 + ], + "spans": [ + { + "bbox": [ + 166, + 91, + 230, + 102 + ], + "type": "text", + "content": "A. Gupta et al." + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 5 + }, + { + "para_blocks": [ + { + "type": "table", + "bbox": [ + 185, + 114, + 426, + 231 + ], + "blocks": [ + { + "bbox": [ + 185, + 114, + 426, + 231 + ], + "lines": [ + { + "bbox": [ + 185, + 114, + 426, + 231 + ], + "spans": [ + { + "bbox": [ + 185, + 114, + 426, + 231 + ], + "type": "table", + "html": "
MethodK600 FVD↓UCF FVD↓params.steps
TrIVD-GAN-FP [52]25.7±0.7--1
Video Diffusion [38]16.2±0.3-1.1B256
RIN [42]10.8-411M1000
TATS [24]-332±18321M1024
Phenaki [75]36.4±0.2-227M48
MAGVIT [82]9.9±0.376±2306M12
MAGVITv2 [83]4.3±0.158±2307M24
W.A.L.T-L Ours3.3±0.046±2313M50
W.A.L.T-XL Ours-36±2460M50
", + "image_path": "897ebf563cab4cc4cfeace5b178aca3aaa8c22e83939c680504a80f44e942410.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "table_body" + } + ], + "index": 2 + }, + { + "bbox": [ + 130, + 237, + 480, + 259 + ], + "lines": [ + { + "bbox": [ + 130, + 237, + 480, + 259 + ], + "spans": [ + { + "bbox": [ + 130, + 237, + 480, + 259 + ], + "type": "text", + "content": "Table 1: Video generation evaluation on frame prediction on Kinetics-600 and class-conditional generation on UCF-101." + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "text" + }, + { + "bbox": [ + 130, + 298, + 482, + 467 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 298, + 482, + 467 + ], + "spans": [ + { + "bbox": [ + 130, + 298, + 482, + 467 + ], + "type": "text", + "content": "AdaLN-LoRA. Adaptive normalization layers are an important component in a broad range of generative and visual synthesis models [17, 20, 45, 53-55]. A simple way to incorporate adaptive layer normalization is to include for each layer " + }, + { + "bbox": [ + 130, + 298, + 482, + 467 + ], + "type": "inline_equation", + "content": "i" + }, + { + "bbox": [ + 130, + 298, + 482, + 467 + ], + "type": "text", + "content": ", an MLP layer to regress a vector of conditioning parameters " + }, + { + "bbox": [ + 130, + 298, + 482, + 467 + ], + "type": "inline_equation", + "content": "A^i = \\mathbf{MLP}(c + t)" + }, + { + "bbox": [ + 130, + 298, + 482, + 467 + ], + "type": "text", + "content": ", where " + }, + { + "bbox": [ + 130, + 298, + 482, + 467 + ], + "type": "inline_equation", + "content": "A^i = \\text{concat}(\\gamma_1, \\gamma_2, \\beta_1, \\beta_2, \\alpha_1, \\alpha_2)" + }, + { + "bbox": [ + 130, + 298, + 482, + 467 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 130, + 298, + 482, + 467 + ], + "type": "inline_equation", + "content": "A^i \\in \\mathbb{R}^{6 \\times d_{\\text{model}}}" + }, + { + "bbox": [ + 130, + 298, + 482, + 467 + ], + "type": "text", + "content": ", and " + }, + { + "bbox": [ + 130, + 298, + 482, + 467 + ], + "type": "inline_equation", + "content": "c \\in \\mathbb{R}^{d_{\\text{model}}}, t \\in \\mathbb{R}^{d_{\\text{model}}}" + }, + { + "bbox": [ + 130, + 298, + 482, + 467 + ], + "type": "text", + "content": " are the condition and timestep embeddings. In the transformer block, " + }, + { + "bbox": [ + 130, + 298, + 482, + 467 + ], + "type": "inline_equation", + "content": "\\gamma" + }, + { + "bbox": [ + 130, + 298, + 482, + 467 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 130, + 298, + 482, + 467 + ], + "type": "inline_equation", + "content": "\\beta" + }, + { + "bbox": [ + 130, + 298, + 482, + 467 + ], + "type": "text", + "content": " scale and shift the inputs of the multi-head attention and MLP layers, respectively, while " + }, + { + "bbox": [ + 130, + 298, + 482, + 467 + ], + "type": "inline_equation", + "content": "\\alpha" + }, + { + "bbox": [ + 130, + 298, + 482, + 467 + ], + "type": "text", + "content": " scales the output of both the multi-head attention and MLP layers. The parameter count of these additional MLP layers scales linearly with the number of layers and quadratically with the model's dimensional size (num_blocks " + }, + { + "bbox": [ + 130, + 298, + 482, + 467 + ], + "type": "inline_equation", + "content": "\\times" + }, + { + "bbox": [ + 130, + 298, + 482, + 467 + ], + "type": "inline_equation", + "content": "d_{\\text{model}} \\times 6 \\times d_{\\text{model}}" + }, + { + "bbox": [ + 130, + 298, + 482, + 467 + ], + "type": "text", + "content": "). For instance, in a ViT-g model with 1B parameters, the MLP layers contribute an additional 475M parameters. Inspired by [41], we propose a simple solution dubbed AdaLN-LoRA, to reduce the model parameters. For each layer, we regress conditioning parameters as" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 179, + 479, + 481, + 493 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 179, + 479, + 481, + 493 + ], + "spans": [ + { + "bbox": [ + 179, + 479, + 481, + 493 + ], + "type": "interline_equation", + "content": "A ^ {1} = \\operatorname {M L P} (\\boldsymbol {c} + \\boldsymbol {t}), \\quad A ^ {i} = A ^ {1} + W _ {b} ^ {i} W _ {a} ^ {i} (\\boldsymbol {c} + \\boldsymbol {t}) \\quad \\forall i \\neq 1, \\tag {3}", + "image_path": "6ad72e6176872de2b29a60e47185fd3a33742c56ddb7581f0cc7048d5c855b15.jpg" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 130, + 506, + 480, + 544 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 506, + 480, + 544 + ], + "spans": [ + { + "bbox": [ + 130, + 506, + 480, + 544 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 130, + 506, + 480, + 544 + ], + "type": "inline_equation", + "content": "W_{b}^{i}\\in \\mathbb{R}^{d_{\\mathrm{model}}\\times r}" + }, + { + "bbox": [ + 130, + 506, + 480, + 544 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 130, + 506, + 480, + 544 + ], + "type": "inline_equation", + "content": "W_{a}^{i}\\in \\mathbb{R}^{r\\times (6\\times d_{\\mathrm{model}})}" + }, + { + "bbox": [ + 130, + 506, + 480, + 544 + ], + "type": "text", + "content": ". This reduces the number of trainable model parameters significantly when " + }, + { + "bbox": [ + 130, + 506, + 480, + 544 + ], + "type": "inline_equation", + "content": "r\\ll d_{\\mathrm{model}}" + }, + { + "bbox": [ + 130, + 506, + 480, + 544 + ], + "type": "text", + "content": ". For example, a ViT-g model with " + }, + { + "bbox": [ + 130, + 506, + 480, + 544 + ], + "type": "inline_equation", + "content": "r = 2" + }, + { + "bbox": [ + 130, + 506, + 480, + 544 + ], + "type": "text", + "content": " reduces the MLP parameters from 475M to 12M." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 130, + 545, + 482, + 666 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 545, + 482, + 666 + ], + "spans": [ + { + "bbox": [ + 130, + 545, + 482, + 666 + ], + "type": "text", + "content": "Self-conditioning. In addition to being conditioned on external inputs, iterative generative algorithms can also be conditioned on their own previously generated samples during inference [3, 14, 66]. Specifically, Chen et al. [14] modify the training process for diffusion models, such that with some probability " + }, + { + "bbox": [ + 130, + 545, + 482, + 666 + ], + "type": "inline_equation", + "content": "p_{\\mathrm{sc}}" + }, + { + "bbox": [ + 130, + 545, + 482, + 666 + ], + "type": "text", + "content": " the model first generates a sample " + }, + { + "bbox": [ + 130, + 545, + 482, + 666 + ], + "type": "inline_equation", + "content": "\\tilde{\\pmb{z}}_0 = f_\\theta(\\pmb{z}_t; \\pmb{0}, \\pmb{c}, t)" + }, + { + "bbox": [ + 130, + 545, + 482, + 666 + ], + "type": "text", + "content": " and then refines this estimate using another forward pass conditioned on this initial sample: " + }, + { + "bbox": [ + 130, + 545, + 482, + 666 + ], + "type": "inline_equation", + "content": "f_\\theta(\\pmb{z}_t; \\text{stopgrad}(\\tilde{\\pmb{z}}_0), \\pmb{c}, t)" + }, + { + "bbox": [ + 130, + 545, + 482, + 666 + ], + "type": "text", + "content": ". With probability " + }, + { + "bbox": [ + 130, + 545, + 482, + 666 + ], + "type": "inline_equation", + "content": "1 - p_{\\mathrm{sc}}" + }, + { + "bbox": [ + 130, + 545, + 482, + 666 + ], + "type": "text", + "content": ", only a single forward pass is done. We concatenate the model estimate with the input along the channel dimension and found this simple technique to work well when used in conjunction with " + }, + { + "bbox": [ + 130, + 545, + 482, + 666 + ], + "type": "inline_equation", + "content": "\\pmb{v}" + }, + { + "bbox": [ + 130, + 545, + 482, + 666 + ], + "type": "text", + "content": "-prediction." + } + ] + } + ], + "index": 7 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 223, + 91, + 448, + 101 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 223, + 91, + 448, + 101 + ], + "spans": [ + { + "bbox": [ + 223, + 91, + 448, + 101 + ], + "type": "text", + "content": "Photorealistic Video Generation with Diffusion Models" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 474, + 92, + 480, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 474, + 92, + 480, + 100 + ], + "spans": [ + { + "bbox": [ + 474, + 92, + 480, + 100 + ], + "type": "text", + "content": "7" + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 6 + }, + { + "para_blocks": [ + { + "bbox": [ + 132, + 116, + 295, + 128 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 116, + 295, + 128 + ], + "spans": [ + { + "bbox": [ + 132, + 116, + 295, + 128 + ], + "type": "text", + "content": "4.4 Autoregressive Generation" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 130, + 137, + 482, + 257 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 137, + 482, + 257 + ], + "spans": [ + { + "bbox": [ + 130, + 137, + 482, + 257 + ], + "type": "text", + "content": "For generating long videos via autoregressive prediction we also train our model jointly on the task of frame prediction. This is achieved by conditioning the model on past frames with a probability of " + }, + { + "bbox": [ + 130, + 137, + 482, + 257 + ], + "type": "inline_equation", + "content": "p_{\\mathrm{fp}}" + }, + { + "bbox": [ + 130, + 137, + 482, + 257 + ], + "type": "text", + "content": " during training. Specifically, the model is conditioned using " + }, + { + "bbox": [ + 130, + 137, + 482, + 257 + ], + "type": "inline_equation", + "content": "c_{\\mathrm{fp}} = \\text{concat}(m_{\\mathrm{fp}} \\circ z_t, m_{\\mathrm{fp}})" + }, + { + "bbox": [ + 130, + 137, + 482, + 257 + ], + "type": "text", + "content": ", where " + }, + { + "bbox": [ + 130, + 137, + 482, + 257 + ], + "type": "inline_equation", + "content": "m_{\\mathrm{fp}}" + }, + { + "bbox": [ + 130, + 137, + 482, + 257 + ], + "type": "text", + "content": " is a binary mask. The binary mask indicates the number of past frames used for conditioning. We condition on either 1 latent frame (image to video generation) or 2 latent frames (video prediction). This conditioning is integrated into the model through concatenation along the channel dimension of the noisy latent input. During inference, we use standard classifier-free guidance with " + }, + { + "bbox": [ + 130, + 137, + 482, + 257 + ], + "type": "inline_equation", + "content": "c_{\\mathrm{fp}}" + }, + { + "bbox": [ + 130, + 137, + 482, + 257 + ], + "type": "text", + "content": " as the conditioning signal." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 132, + 274, + 281, + 286 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 274, + 281, + 286 + ], + "spans": [ + { + "bbox": [ + 132, + 274, + 281, + 286 + ], + "type": "text", + "content": "4.5 Video Super Resolution" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 130, + 295, + 482, + 437 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 295, + 482, + 437 + ], + "spans": [ + { + "bbox": [ + 130, + 295, + 482, + 437 + ], + "type": "text", + "content": "Generating high-resolution videos with a single model is computationally prohibitive. Following [37], we use a cascaded approach with three models operating at increasing resolutions. Our base model generates videos at " + }, + { + "bbox": [ + 130, + 295, + 482, + 437 + ], + "type": "inline_equation", + "content": "128 \\times 128" + }, + { + "bbox": [ + 130, + 295, + 482, + 437 + ], + "type": "text", + "content": " resolution which are subsequently upsampled twice via two super resolution stages. We first spatially upscale the low resolution input " + }, + { + "bbox": [ + 130, + 295, + 482, + 437 + ], + "type": "inline_equation", + "content": "z^{\\mathrm{lr}}" + }, + { + "bbox": [ + 130, + 295, + 482, + 437 + ], + "type": "text", + "content": " (video or image) using a depth-to-space convolution operation. Note that, unlike training where ground truth low-resolution inputs are available, inference relies on latents produced by preceding stages (cf. teaching-forcing). To reduce this discrepancy and improve the robustness of the super-resolution stages in handling artifacts generated by lower resolution stages, we use noise conditioning augmentation [37]. Concretely, noise is added in accordance with " + }, + { + "bbox": [ + 130, + 295, + 482, + 437 + ], + "type": "inline_equation", + "content": "\\gamma(t)" + }, + { + "bbox": [ + 130, + 295, + 482, + 437 + ], + "type": "text", + "content": ", by sampling a noise level as " + }, + { + "bbox": [ + 130, + 295, + 482, + 437 + ], + "type": "inline_equation", + "content": "t_{\\mathrm{sr}} \\sim \\mathcal{U}(0, t_{\\max \\text{noise}})" + }, + { + "bbox": [ + 130, + 295, + 482, + 437 + ], + "type": "text", + "content": " and is provided as input to our AdaLN-LoRA layers." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 130, + 438, + 482, + 487 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 438, + 482, + 487 + ], + "spans": [ + { + "bbox": [ + 130, + 438, + 482, + 487 + ], + "type": "text", + "content": "Aspect-ratio finetuning. To simplify training and leverage broad data sources with different aspect ratios, we train our base stage using a square aspect ratio. We fine-tune the base stage on a subset of data to generate videos with a 9:16 aspect ratio by interpolating position embeddings." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 132, + 505, + 230, + 519 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 505, + 230, + 519 + ], + "spans": [ + { + "bbox": [ + 132, + 505, + 230, + 519 + ], + "type": "text", + "content": "5 Experiments" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 130, + 531, + 482, + 590 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 531, + 482, + 590 + ], + "spans": [ + { + "bbox": [ + 130, + 531, + 482, + 590 + ], + "type": "text", + "content": "In this section, we evaluate our method on multiple tasks: class-conditional image and video generation, frame prediction and text conditioned video generation and perform extensive ablation studies of different design choices. For qualitative results, see Fig. 1, Fig. 3, Fig. 4 and videos on our project website. See appendix for additional details." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 132, + 609, + 253, + 620 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 609, + 253, + 620 + ], + "spans": [ + { + "bbox": [ + 132, + 609, + 253, + 620 + ], + "type": "text", + "content": "5.1 Visual Generation" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 130, + 629, + 482, + 666 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 629, + 482, + 666 + ], + "spans": [ + { + "bbox": [ + 130, + 629, + 482, + 666 + ], + "type": "text", + "content": "Video generation. We consider two standard video benchmarks, UCF-101 [71] for class-conditional generation and Kinetics-600 [10] for video prediction with 5 conditioning frames. We use FVD [72] as our primary evaluation metric." + } + ] + } + ], + "index": 10 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "spans": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "type": "text", + "content": "8" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 166, + 91, + 230, + 102 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 166, + 91, + 230, + 102 + ], + "spans": [ + { + "bbox": [ + 166, + 91, + 230, + 102 + ], + "type": "text", + "content": "A. Gupta et al." + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 7 + }, + { + "para_blocks": [ + { + "type": "table", + "bbox": [ + 182, + 114, + 426, + 225 + ], + "blocks": [ + { + "bbox": [ + 182, + 114, + 426, + 225 + ], + "lines": [ + { + "bbox": [ + 182, + 114, + 426, + 225 + ], + "spans": [ + { + "bbox": [ + 182, + 114, + 426, + 225 + ], + "type": "table", + "html": "
MethodCost (Iter×BS)FID↓IS↑params.steps
BigGAN-deep [7]-6.95171.4160M1
LDM-4 [62]178k×120010.56103.5400M250
DiT-XL/2 [54]7000k×2569.62121.5675M250
ADM [17]-7.49127.5608M2000
MDT [23]6500k×2566.23143.0676M250
MaskDiT [89]1200k×10245.69178.0736M40
RIN [42]600k×10243.42182.0410M1000
simple diffusion [40]500K×20482.77211.82B512
VDM++ [46]-2.40225.32B512
W.A.L.T-L Ours437k×10242.56215.1460M50
", + "image_path": "eace865a00a02acbec11eaeae07259da3d898262a25d303499ae451eea99efca.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "table_body" + } + ], + "index": 2 + }, + { + "bbox": [ + 130, + 234, + 482, + 268 + ], + "lines": [ + { + "bbox": [ + 130, + 234, + 482, + 268 + ], + "spans": [ + { + "bbox": [ + 130, + 234, + 482, + 268 + ], + "type": "text", + "content": "Table 2: Class-conditional image generation on ImageNet " + }, + { + "bbox": [ + 130, + 234, + 482, + 268 + ], + "type": "inline_equation", + "content": "{256} \\times {256}" + }, + { + "bbox": [ + 130, + 234, + 482, + 268 + ], + "type": "text", + "content": " . We adopt the evaluation protocol and implementation of ADM [17] and report results without classifier free guidance." + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "text" + }, + { + "bbox": [ + 130, + 301, + 481, + 337 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 301, + 481, + 337 + ], + "spans": [ + { + "bbox": [ + 130, + 301, + 481, + 337 + ], + "type": "text", + "content": "Across both datasets, W.A.L.T significantly outperforms all prior works (Tab. 1). Compared to prior video diffusion models, we achieve state-of-the-art performance with less model parameters, and require 50 DDIM [69] inference steps." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 130, + 338, + 482, + 445 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 338, + 482, + 445 + ], + "spans": [ + { + "bbox": [ + 130, + 338, + 482, + 445 + ], + "type": "text", + "content": "Image generation. To verify the modeling capabilities of " + }, + { + "bbox": [ + 130, + 338, + 482, + 445 + ], + "type": "inline_equation", + "content": "W.A.L.T" + }, + { + "bbox": [ + 130, + 338, + 482, + 445 + ], + "type": "text", + "content": " on the image domain, we train a version of " + }, + { + "bbox": [ + 130, + 338, + 482, + 445 + ], + "type": "inline_equation", + "content": "W.A.L.T" + }, + { + "bbox": [ + 130, + 338, + 482, + 445 + ], + "type": "text", + "content": " for the standard ImageNet class-conditional setting. For evaluation, we follow ADM [17] and report the FID [34] and Inception [64] scores calculated on 50K samples generated in 50 DDIM steps. We compare (Table 2) " + }, + { + "bbox": [ + 130, + 338, + 482, + 445 + ], + "type": "inline_equation", + "content": "W.A.L.T" + }, + { + "bbox": [ + 130, + 338, + 482, + 445 + ], + "type": "text", + "content": " with state-of-the-art image generation methods for " + }, + { + "bbox": [ + 130, + 338, + 482, + 445 + ], + "type": "inline_equation", + "content": "256 \\times 256" + }, + { + "bbox": [ + 130, + 338, + 482, + 445 + ], + "type": "text", + "content": " resolution. Our model outperforms prior works without requiring specialized schedules, convolution inductive bias, improved diffusion losses, and classifier free guidance. Although VDM++ [46] has slightly better FID score, the model has significantly more parameters (2B)." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 131, + 464, + 246, + 475 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 131, + 464, + 246, + 475 + ], + "spans": [ + { + "bbox": [ + 131, + 464, + 246, + 475 + ], + "type": "text", + "content": "5.2 Ablation Studies" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 130, + 485, + 481, + 521 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 485, + 481, + 521 + ], + "spans": [ + { + "bbox": [ + 130, + 485, + 481, + 521 + ], + "type": "text", + "content": "We ablate " + }, + { + "bbox": [ + 130, + 485, + 481, + 521 + ], + "type": "inline_equation", + "content": "W.A.L.T" + }, + { + "bbox": [ + 130, + 485, + 481, + 521 + ], + "type": "text", + "content": " to understand the contribution of various design decisions with the default settings: model L, patch size 1, " + }, + { + "bbox": [ + 130, + 485, + 481, + 521 + ], + "type": "inline_equation", + "content": "1 \\times 16 \\times 16" + }, + { + "bbox": [ + 130, + 485, + 481, + 521 + ], + "type": "text", + "content": " spatial window, " + }, + { + "bbox": [ + 130, + 485, + 481, + 521 + ], + "type": "inline_equation", + "content": "5 \\times 8 \\times 8" + }, + { + "bbox": [ + 130, + 485, + 481, + 521 + ], + "type": "text", + "content": " spatiotemporal window, " + }, + { + "bbox": [ + 130, + 485, + 481, + 521 + ], + "type": "inline_equation", + "content": "p_{\\mathrm{sc}} = 0.9" + }, + { + "bbox": [ + 130, + 485, + 481, + 521 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 130, + 485, + 481, + 521 + ], + "type": "inline_equation", + "content": "c = 8" + }, + { + "bbox": [ + 130, + 485, + 481, + 521 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 130, + 485, + 481, + 521 + ], + "type": "inline_equation", + "content": "r = 2" + }, + { + "bbox": [ + 130, + 485, + 481, + 521 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 130, + 521, + 481, + 569 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 521, + 481, + 569 + ], + "spans": [ + { + "bbox": [ + 130, + 521, + 481, + 569 + ], + "type": "text", + "content": "Patch size. In various computer vision tasks utilizing ViT [19]-based models, a smaller patch size " + }, + { + "bbox": [ + 130, + 521, + 481, + 569 + ], + "type": "inline_equation", + "content": "p" + }, + { + "bbox": [ + 130, + 521, + 481, + 569 + ], + "type": "text", + "content": " has been shown to consistently enhance performance [9, 19, 30, 86]. Similarly, our findings also indicate that a reduced patch size improves performance (Table 3a)." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 130, + 570, + 481, + 617 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 570, + 481, + 617 + ], + "spans": [ + { + "bbox": [ + 130, + 570, + 481, + 617 + ], + "type": "text", + "content": "Window attention. We compare three different STW window configurations with full self-attention (Table 3b). We find that local self-attention can achieve competitive (or better) performance while being significantly faster (up to " + }, + { + "bbox": [ + 130, + 570, + 481, + 617 + ], + "type": "inline_equation", + "content": "2 \\times" + }, + { + "bbox": [ + 130, + 570, + 481, + 617 + ], + "type": "text", + "content": ") and requiring less accelerator memory." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 130, + 617, + 481, + 666 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 617, + 481, + 666 + ], + "spans": [ + { + "bbox": [ + 130, + 617, + 481, + 666 + ], + "type": "text", + "content": "Self-conditioning. In Table 3c we study the influence of varying the self-conditioning rate " + }, + { + "bbox": [ + 130, + 617, + 481, + 666 + ], + "type": "inline_equation", + "content": "p_{\\mathrm{sc}}" + }, + { + "bbox": [ + 130, + 617, + 481, + 666 + ], + "type": "text", + "content": " on generation quality. We notice a clear trend: increasing the self-conditioning rate from 0.0 (no self-conditioning) to 0.9 improves the FVD score substantially (44%)." + } + ] + } + ], + "index": 10 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 223, + 91, + 447, + 101 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 223, + 91, + 447, + 101 + ], + "spans": [ + { + "bbox": [ + 223, + 91, + 447, + 101 + ], + "type": "text", + "content": "Photorealistic Video Generation with Diffusion Models" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 474, + 92, + 480, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 474, + 92, + 480, + 100 + ], + "spans": [ + { + "bbox": [ + 474, + 92, + 480, + 100 + ], + "type": "text", + "content": "9" + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 8 + }, + { + "para_blocks": [ + { + "type": "table", + "bbox": [ + 176, + 129, + 284, + 180 + ], + "blocks": [ + { + "bbox": [ + 176, + 129, + 284, + 180 + ], + "lines": [ + { + "bbox": [ + 176, + 129, + 284, + 180 + ], + "spans": [ + { + "bbox": [ + 176, + 129, + 284, + 180 + ], + "type": "table", + "html": "
patch size pFVD↓IS↑
160.787.2
2134.482.2
4461.863.9
", + "image_path": "123e19d4947ed99c95dce83ce3a69df7c66c0f1c868ff20604e49086bc265d00.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "table_body" + } + ], + "index": 2 + }, + { + "type": "table", + "bbox": [ + 318, + 113, + 455, + 180 + ], + "blocks": [ + { + "bbox": [ + 318, + 113, + 455, + 180 + ], + "lines": [ + { + "bbox": [ + 318, + 113, + 455, + 180 + ], + "spans": [ + { + "bbox": [ + 318, + 113, + 455, + 180 + ], + "type": "table", + "html": "
st windowFVD↓IS↑sps
5 × 4 × 456.987.32.24
5 × 8 × 859.687.42.00
5 × 16 × 1655.387.41.75
full self attn.59.987.81.20
", + "image_path": "4f0a18f9f1845217dc8fd74e37d9ef76c7fde2f50520ee967cfad376f6cfbd6f.jpg" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "table_body" + } + ], + "index": 3 + }, + { + "bbox": [ + 153, + 194, + 303, + 213 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 153, + 194, + 303, + 213 + ], + "spans": [ + { + "bbox": [ + 153, + 194, + 303, + 213 + ], + "type": "text", + "content": "(a) Patch size. Lower patch size is significantly better." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 311, + 194, + 460, + 220 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 311, + 194, + 460, + 220 + ], + "spans": [ + { + "bbox": [ + 311, + 194, + 460, + 220 + ], + "type": "text", + "content": "(b) Spatiotemporal window size. Full self-attention is not essential for good performance. sps is steps per sec." + } + ] + } + ], + "index": 5 + }, + { + "type": "table", + "bbox": [ + 199, + 231, + 272, + 294 + ], + "blocks": [ + { + "bbox": [ + 199, + 231, + 272, + 294 + ], + "lines": [ + { + "bbox": [ + 199, + 231, + 272, + 294 + ], + "spans": [ + { + "bbox": [ + 199, + 231, + 272, + 294 + ], + "type": "table", + "html": "
pscFVD↓IS↑
0.0109.982.6
0.376.086.5
0.660.086.8
0.961.487.1
", + "image_path": "97c77a8e4eae23d72cf0dcc78b48450dfe6f10c1a2d44db228b8055bed69cb90.jpg" + } + ] + } + ], + "index": 6, + "angle": 0, + "type": "table_body" + } + ], + "index": 6 + }, + { + "type": "table", + "bbox": [ + 326, + 231, + 436, + 304 + ], + "blocks": [ + { + "bbox": [ + 326, + 231, + 436, + 304 + ], + "lines": [ + { + "bbox": [ + 326, + 231, + 436, + 304 + ], + "spans": [ + { + "bbox": [ + 326, + 231, + 436, + 304 + ], + "type": "table", + "html": "
rFVD↓IS↑params
260.787.2313 M
456.687.3314 M
1655.588.0316 M
6454.487.9324 M
25652.588.5357 M
", + "image_path": "842a6201c4a43c994dfa2fc9b5e198d2eab341ad9e5bdfc07362fa7389129410.jpg" + } + ] + } + ], + "index": 7, + "angle": 0, + "type": "table_body" + } + ], + "index": 7 + }, + { + "bbox": [ + 158, + 307, + 448, + 324 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 158, + 307, + 448, + 324 + ], + "spans": [ + { + "bbox": [ + 158, + 307, + 448, + 324 + ], + "type": "text", + "content": "(c) Self-conditioning. Higher " + }, + { + "bbox": [ + 158, + 307, + 448, + 324 + ], + "type": "inline_equation", + "content": "p_{\\mathrm{sc}}" + }, + { + "bbox": [ + 158, + 307, + 448, + 324 + ], + "type": "text", + "content": " is bet- (d) AdaLN-LoRA. Bigger " + }, + { + "bbox": [ + 158, + 307, + 448, + 324 + ], + "type": "inline_equation", + "content": "r" + }, + { + "bbox": [ + 158, + 307, + 448, + 324 + ], + "type": "text", + "content": " is better." + } + ] + } + ], + "index": 8 + }, + { + "type": "table", + "bbox": [ + 164, + 335, + 296, + 403 + ], + "blocks": [ + { + "bbox": [ + 164, + 335, + 296, + 403 + ], + "lines": [ + { + "bbox": [ + 164, + 335, + 296, + 403 + ], + "spans": [ + { + "bbox": [ + 164, + 335, + 296, + 403 + ], + "type": "table", + "html": "
FVD↓IS↑
w/o qk norm [15]59.086.8
w/o latent norm67.987.1
w/o zero snr [49]91.084.2
full method60.787.2
", + "image_path": "3d33dc9b09d2a9541b97bb0a7a7e8160696b7e714d831329914c827f1687e825.jpg" + } + ] + } + ], + "index": 9, + "angle": 0, + "type": "table_body" + } + ], + "index": 9 + }, + { + "bbox": [ + 154, + 405, + 302, + 422 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 154, + 405, + 302, + 422 + ], + "spans": [ + { + "bbox": [ + 154, + 405, + 302, + 422 + ], + "type": "text", + "content": "(e) Other improvements. See text for details." + } + ] + } + ], + "index": 10 + }, + { + "type": "table", + "bbox": [ + 334, + 339, + 439, + 402 + ], + "blocks": [ + { + "bbox": [ + 334, + 339, + 439, + 402 + ], + "lines": [ + { + "bbox": [ + 334, + 339, + 439, + 402 + ], + "spans": [ + { + "bbox": [ + 334, + 339, + 439, + 402 + ], + "type": "table", + "html": "
crFVD↓FVD↓IS↑
437.786.484.9
817.175.486.3
168.267.086.0
323.583.482.9
", + "image_path": "23b8fd03d08544e187b7772eaa468c02bb881370b5f19a0bd8da85256389f12c.jpg" + } + ] + } + ], + "index": 11, + "angle": 0, + "type": "table_body" + } + ], + "index": 11 + }, + { + "bbox": [ + 311, + 405, + 460, + 422 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 311, + 405, + 460, + 422 + ], + "spans": [ + { + "bbox": [ + 311, + 405, + 460, + 422 + ], + "type": "text", + "content": "(f) Latent dimension " + }, + { + "bbox": [ + 311, + 405, + 460, + 422 + ], + "type": "inline_equation", + "content": "c" + }, + { + "bbox": [ + 311, + 405, + 460, + 422 + ], + "type": "text", + "content": ". Higher " + }, + { + "bbox": [ + 311, + 405, + 460, + 422 + ], + "type": "inline_equation", + "content": "c" + }, + { + "bbox": [ + 311, + 405, + 460, + 422 + ], + "type": "text", + "content": " is better." + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 130, + 431, + 482, + 467 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 431, + 482, + 467 + ], + "spans": [ + { + "bbox": [ + 130, + 431, + 482, + 467 + ], + "type": "text", + "content": "Table 3: Ablation experiments on UCF-101 [71]. We compare FVD and inception scores to ablate important design decisions with the default setting: L model, " + }, + { + "bbox": [ + 130, + 431, + 482, + 467 + ], + "type": "inline_equation", + "content": "1 \\times 16 \\times 16" + }, + { + "bbox": [ + 130, + 431, + 482, + 467 + ], + "type": "text", + "content": " spatial window, " + }, + { + "bbox": [ + 130, + 431, + 482, + 467 + ], + "type": "inline_equation", + "content": "5 \\times 8 \\times 8" + }, + { + "bbox": [ + 130, + 431, + 482, + 467 + ], + "type": "text", + "content": " saptiotemporal (st) window, " + }, + { + "bbox": [ + 130, + 431, + 482, + 467 + ], + "type": "inline_equation", + "content": "p_{\\mathrm{sc}} = 0.9" + }, + { + "bbox": [ + 130, + 431, + 482, + 467 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 130, + 431, + 482, + 467 + ], + "type": "inline_equation", + "content": "c = 8" + }, + { + "bbox": [ + 130, + 431, + 482, + 467 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 130, + 431, + 482, + 467 + ], + "type": "inline_equation", + "content": "r = 2" + }, + { + "bbox": [ + 130, + 431, + 482, + 467 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 130, + 498, + 482, + 666 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 498, + 482, + 666 + ], + "spans": [ + { + "bbox": [ + 130, + 498, + 482, + 666 + ], + "type": "text", + "content": "AdaLN-LoRA. An important design decision in diffusion models is the conditioning mechanism. We investigate the effect of increasing the bottleneck dimension " + }, + { + "bbox": [ + 130, + 498, + 482, + 666 + ], + "type": "inline_equation", + "content": "r" + }, + { + "bbox": [ + 130, + 498, + 482, + 666 + ], + "type": "text", + "content": " in our proposed AdaLN-LoRA layers (Table 3d). This hyperparameter provides a flexible way to trade off between number of model parameters and generation performance. As shown in Table 3d, increasing " + }, + { + "bbox": [ + 130, + 498, + 482, + 666 + ], + "type": "inline_equation", + "content": "r" + }, + { + "bbox": [ + 130, + 498, + 482, + 666 + ], + "type": "text", + "content": " improves performance but also increases model parameters. This highlights an important model design question: given a fixed parameter budget, how should we allocate parameters - either by using separate AdaLN layers, or by increasing base model parameters while using shared AdaLN-LoRA layers? We explore this in Table 4 by comparing two model configurations: W.A.L.T-L with separate AdaLN layers and W.A.L.T-XL with AdaLN-LoRA and " + }, + { + "bbox": [ + 130, + 498, + 482, + 666 + ], + "type": "inline_equation", + "content": "r = 2" + }, + { + "bbox": [ + 130, + 498, + 482, + 666 + ], + "type": "text", + "content": ". While both configurations yield similar FVD and Inception scores, W.A.L.T-XL achieves a lower final loss value, suggesting the advantage of allocating more parameters to the base model and choosing an appropriate " + }, + { + "bbox": [ + 130, + 498, + 482, + 666 + ], + "type": "inline_equation", + "content": "r" + }, + { + "bbox": [ + 130, + 498, + 482, + 666 + ], + "type": "text", + "content": " value within accelerator memory limits." + } + ] + } + ], + "index": 14 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "spans": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "type": "text", + "content": "10" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 166, + 91, + 230, + 102 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 166, + 91, + 230, + 102 + ], + "spans": [ + { + "bbox": [ + 166, + 91, + 230, + 102 + ], + "type": "text", + "content": "A. Gupta et al." + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 9 + }, + { + "para_blocks": [ + { + "type": "table", + "bbox": [ + 191, + 114, + 421, + 162 + ], + "blocks": [ + { + "bbox": [ + 191, + 114, + 421, + 162 + ], + "lines": [ + { + "bbox": [ + 191, + 114, + 421, + 162 + ], + "spans": [ + { + "bbox": [ + 191, + 114, + 421, + 162 + ], + "type": "table", + "html": "
ModelAdaLNFVD↓IS↑params.final loss
Lseparate34.690.2458M0.274
XLLoRA-236.789.4460M0.268
", + "image_path": "feb7c5fc03effc9148f63161e66e38faba4950e4bb65d2ac026c963f2f435101.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "table_body" + } + ], + "index": 2 + }, + { + "bbox": [ + 130, + 176, + 480, + 198 + ], + "lines": [ + { + "bbox": [ + 130, + 176, + 480, + 198 + ], + "spans": [ + { + "bbox": [ + 130, + 176, + 480, + 198 + ], + "type": "text", + "content": "Table 4: Parameter matched comparison between AdaLN-LoRA and per layer adaln layers. See text for details." + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "text" + }, + { + "bbox": [ + 130, + 233, + 482, + 365 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 233, + 482, + 365 + ], + "spans": [ + { + "bbox": [ + 130, + 233, + 482, + 365 + ], + "type": "text", + "content": "Noise schedule. Common latent diffusion noise schedules [62] typically do not ensure a zero signal-to-noise ratio (SNR) at the final timestep, i.e., at " + }, + { + "bbox": [ + 130, + 233, + 482, + 365 + ], + "type": "inline_equation", + "content": "t = 1" + }, + { + "bbox": [ + 130, + 233, + 482, + 365 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 130, + 233, + 482, + 365 + ], + "type": "inline_equation", + "content": "\\gamma(t) > 0" + }, + { + "bbox": [ + 130, + 233, + 482, + 365 + ], + "type": "text", + "content": ". This leads to a mismatch between training and inference phases. During inference, models are expected to start from purely Gaussian noise, whereas during training, at " + }, + { + "bbox": [ + 130, + 233, + 482, + 365 + ], + "type": "inline_equation", + "content": "t = 1" + }, + { + "bbox": [ + 130, + 233, + 482, + 365 + ], + "type": "text", + "content": ", a small amount of signal information remains accessible to the model. This is especially harmful for video generation as videos have high temporal redundancy. Even minimal information leakage at " + }, + { + "bbox": [ + 130, + 233, + 482, + 365 + ], + "type": "inline_equation", + "content": "t = 1" + }, + { + "bbox": [ + 130, + 233, + 482, + 365 + ], + "type": "text", + "content": " can reveal substantial information to the model. Addressing this mismatch by enforcing a zero terminal SNR [49] significantly improves performance (Table 3e). Note that this approach was originally proposed to fix over-exposure problems in image generation, but we find it effective for video generation as well." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 130, + 366, + 482, + 462 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 366, + 482, + 462 + ], + "spans": [ + { + "bbox": [ + 130, + 366, + 482, + 462 + ], + "type": "text", + "content": "Autoencoder. Finally, we investigate one critical but often overlooked hyperparameter in the first stage of our model: the channel dimension " + }, + { + "bbox": [ + 130, + 366, + 482, + 462 + ], + "type": "inline_equation", + "content": "c" + }, + { + "bbox": [ + 130, + 366, + 482, + 462 + ], + "type": "text", + "content": " of the autoencoder latent " + }, + { + "bbox": [ + 130, + 366, + 482, + 462 + ], + "type": "inline_equation", + "content": "z" + }, + { + "bbox": [ + 130, + 366, + 482, + 462 + ], + "type": "text", + "content": ". As shown in Table 3f, increasing " + }, + { + "bbox": [ + 130, + 366, + 482, + 462 + ], + "type": "inline_equation", + "content": "c" + }, + { + "bbox": [ + 130, + 366, + 482, + 462 + ], + "type": "text", + "content": " significantly improves the reconstruction quality (lower rFVD) while keeping the same spatial " + }, + { + "bbox": [ + 130, + 366, + 482, + 462 + ], + "type": "inline_equation", + "content": "f_{s}" + }, + { + "bbox": [ + 130, + 366, + 482, + 462 + ], + "type": "text", + "content": " and temporal compression " + }, + { + "bbox": [ + 130, + 366, + 482, + 462 + ], + "type": "inline_equation", + "content": "f_{t}" + }, + { + "bbox": [ + 130, + 366, + 482, + 462 + ], + "type": "text", + "content": " ratios. Empirically, we found that both lower and higher values of " + }, + { + "bbox": [ + 130, + 366, + 482, + 462 + ], + "type": "inline_equation", + "content": "c" + }, + { + "bbox": [ + 130, + 366, + 482, + 462 + ], + "type": "text", + "content": " lead to poor FVD scores in generation, with a sweet spot of " + }, + { + "bbox": [ + 130, + 366, + 482, + 462 + ], + "type": "inline_equation", + "content": "c = 8" + }, + { + "bbox": [ + 130, + 366, + 482, + 462 + ], + "type": "text", + "content": " working well across most datasets and tasks we evaluated. We also normalize the latents before processing them via transformer which further improves performance." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 130, + 463, + 482, + 522 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 463, + 482, + 522 + ], + "spans": [ + { + "bbox": [ + 130, + 463, + 482, + 522 + ], + "type": "text", + "content": "In our transformer models, we use query-key normalization [15] as it helps stabilize training for larger models. Finally, we note that some of our default settings are not optimal, as indicated by ablation studies. These defaults were chosen early on for their robustness across datasets, though further tuning may improve performance." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 131, + 544, + 230, + 556 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 131, + 544, + 230, + 556 + ], + "spans": [ + { + "bbox": [ + 131, + 544, + 230, + 556 + ], + "type": "text", + "content": "5.3 Text-to-video" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 130, + 569, + 482, + 665 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 569, + 482, + 665 + ], + "spans": [ + { + "bbox": [ + 130, + 569, + 482, + 665 + ], + "type": "text", + "content": "We train " + }, + { + "bbox": [ + 130, + 569, + 482, + 665 + ], + "type": "inline_equation", + "content": "W.A.L.T" + }, + { + "bbox": [ + 130, + 569, + 482, + 665 + ], + "type": "text", + "content": " for text-to-video jointly on text-image and text-video pairs (Sec. 4.2). We used a dataset of " + }, + { + "bbox": [ + 130, + 569, + 482, + 665 + ], + "type": "inline_equation", + "content": "\\sim 970\\mathrm{M}" + }, + { + "bbox": [ + 130, + 569, + 482, + 665 + ], + "type": "text", + "content": " text-image pairs and " + }, + { + "bbox": [ + 130, + 569, + 482, + 665 + ], + "type": "inline_equation", + "content": "\\sim 89\\mathrm{M}" + }, + { + "bbox": [ + 130, + 569, + 482, + 665 + ], + "type": "text", + "content": " text-video pairs from the public internet and internal sources. We train our base model at resolution " + }, + { + "bbox": [ + 130, + 569, + 482, + 665 + ], + "type": "inline_equation", + "content": "17 \\times 128 \\times 128" + }, + { + "bbox": [ + 130, + 569, + 482, + 665 + ], + "type": "text", + "content": " (3B parameters), and two " + }, + { + "bbox": [ + 130, + 569, + 482, + 665 + ], + "type": "inline_equation", + "content": "2 \\times" + }, + { + "bbox": [ + 130, + 569, + 482, + 665 + ], + "type": "text", + "content": " cascaded super-resolution models for " + }, + { + "bbox": [ + 130, + 569, + 482, + 665 + ], + "type": "inline_equation", + "content": "17 \\times 128 \\times 224 \\rightarrow 17 \\times 256 \\times 448" + }, + { + "bbox": [ + 130, + 569, + 482, + 665 + ], + "type": "text", + "content": " (L, 1.3B, " + }, + { + "bbox": [ + 130, + 569, + 482, + 665 + ], + "type": "inline_equation", + "content": "p = 2" + }, + { + "bbox": [ + 130, + 569, + 482, + 665 + ], + "type": "text", + "content": ") and " + }, + { + "bbox": [ + 130, + 569, + 482, + 665 + ], + "type": "inline_equation", + "content": "17 \\times 256 \\times 448 \\rightarrow 17 \\times 512 \\times 896" + }, + { + "bbox": [ + 130, + 569, + 482, + 665 + ], + "type": "text", + "content": " (L, 419M, " + }, + { + "bbox": [ + 130, + 569, + 482, + 665 + ], + "type": "inline_equation", + "content": "p = 2" + }, + { + "bbox": [ + 130, + 569, + 482, + 665 + ], + "type": "text", + "content": ") respectively. We fine-tune the base stage for the " + }, + { + "bbox": [ + 130, + 569, + 482, + 665 + ], + "type": "inline_equation", + "content": "9:16" + }, + { + "bbox": [ + 130, + 569, + 482, + 665 + ], + "type": "text", + "content": " aspect ratio to generate videos at resolution " + }, + { + "bbox": [ + 130, + 569, + 482, + 665 + ], + "type": "inline_equation", + "content": "128 \\times 224" + }, + { + "bbox": [ + 130, + 569, + 482, + 665 + ], + "type": "text", + "content": ". We use classifier free guidance for all our text-to-video results." + } + ] + } + ], + "index": 8 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 223, + 90, + 447, + 101 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 223, + 90, + 447, + 101 + ], + "spans": [ + { + "bbox": [ + 223, + 90, + 447, + 101 + ], + "type": "text", + "content": "Photorealistic Video Generation with Diffusion Models" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 470, + 91, + 479, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 470, + 91, + 479, + 100 + ], + "spans": [ + { + "bbox": [ + 470, + 91, + 479, + 100 + ], + "type": "text", + "content": "11" + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 10 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 134, + 163, + 479, + 205 + ], + "blocks": [ + { + "bbox": [ + 134, + 163, + 479, + 205 + ], + "lines": [ + { + "bbox": [ + 134, + 163, + 479, + 205 + ], + "spans": [ + { + "bbox": [ + 134, + 163, + 479, + 205 + ], + "type": "image", + "image_path": "3fbda4a8bdb8e31bb1dc4a016cfeac307af56b48312cc596c6ad45f79c814283.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 272, + 206, + 339, + 215 + ], + "lines": [ + { + "bbox": [ + 272, + 206, + 339, + 215 + ], + "spans": [ + { + "bbox": [ + 272, + 206, + 339, + 215 + ], + "type": "text", + "content": "A polar bear swimming." + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_caption" + } + ], + "index": 2 + }, + { + "type": "image", + "bbox": [ + 133, + 218, + 479, + 260 + ], + "blocks": [ + { + "bbox": [ + 133, + 218, + 479, + 260 + ], + "lines": [ + { + "bbox": [ + 133, + 218, + 479, + 260 + ], + "spans": [ + { + "bbox": [ + 133, + 218, + 479, + 260 + ], + "type": "image", + "image_path": "6cc71d1c985cc4361b40012b480bdab728ca2d6e6f7641a44e221db218f9a558.jpg" + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 208, + 263, + 403, + 273 + ], + "lines": [ + { + "bbox": [ + 208, + 263, + 403, + 273 + ], + "spans": [ + { + "bbox": [ + 208, + 263, + 403, + 273 + ], + "type": "text", + "content": "Pouring beer into an empty glass, low angle shot, bar in the background." + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "image_caption" + } + ], + "index": 4 + }, + { + "type": "image", + "bbox": [ + 133, + 277, + 479, + 319 + ], + "blocks": [ + { + "bbox": [ + 133, + 277, + 479, + 319 + ], + "lines": [ + { + "bbox": [ + 133, + 277, + 479, + 319 + ], + "spans": [ + { + "bbox": [ + 133, + 277, + 479, + 319 + ], + "type": "image", + "image_path": "77d16469cfeda54b85c88371d136e0c211b417ad9369ce87102415db1e8605bf.jpg" + } + ] + } + ], + "index": 6, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 230, + 321, + 381, + 330 + ], + "lines": [ + { + "bbox": [ + 230, + 321, + 381, + 330 + ], + "spans": [ + { + "bbox": [ + 230, + 321, + 381, + 330 + ], + "type": "text", + "content": "A cat eating food out of a bowl, in the style of Van Gogh." + } + ] + } + ], + "index": 7, + "angle": 0, + "type": "image_caption" + } + ], + "index": 6 + }, + { + "type": "image", + "bbox": [ + 133, + 333, + 479, + 374 + ], + "blocks": [ + { + "bbox": [ + 133, + 333, + 479, + 374 + ], + "lines": [ + { + "bbox": [ + 133, + 333, + 479, + 374 + ], + "spans": [ + { + "bbox": [ + 133, + 333, + 479, + 374 + ], + "type": "image", + "image_path": "661d6094151bd65db1e7b4c6027bf3930220eb15ec10423e7b8150443ea8964b.jpg" + } + ] + } + ], + "index": 8, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 208, + 378, + 403, + 387 + ], + "lines": [ + { + "bbox": [ + 208, + 378, + 403, + 387 + ], + "spans": [ + { + "bbox": [ + 208, + 378, + 403, + 387 + ], + "type": "text", + "content": "Pouring chocolate sauce over vanilla ice cream in a cone, studio lighting." + } + ] + } + ], + "index": 9, + "angle": 0, + "type": "image_caption" + } + ], + "index": 8 + }, + { + "type": "image", + "bbox": [ + 133, + 392, + 479, + 433 + ], + "blocks": [ + { + "bbox": [ + 133, + 392, + 479, + 433 + ], + "lines": [ + { + "bbox": [ + 133, + 392, + 479, + 433 + ], + "spans": [ + { + "bbox": [ + 133, + 392, + 479, + 433 + ], + "type": "image", + "image_path": "a3b932a13c27a28ee5f53b633e465054e881c7e085bb01d3cb2ca3f5c7c64dea.jpg" + } + ] + } + ], + "index": 10, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 213, + 437, + 398, + 445 + ], + "lines": [ + { + "bbox": [ + 213, + 437, + 398, + 445 + ], + "spans": [ + { + "bbox": [ + 213, + 437, + 398, + 445 + ], + "type": "text", + "content": "A robot ballerina dancing gracefully, highly detailed, studio lighting." + } + ] + } + ], + "index": 11, + "angle": 0, + "type": "image_caption" + } + ], + "index": 10 + }, + { + "type": "image", + "bbox": [ + 133, + 450, + 479, + 491 + ], + "blocks": [ + { + "bbox": [ + 133, + 450, + 479, + 491 + ], + "lines": [ + { + "bbox": [ + 133, + 450, + 479, + 491 + ], + "spans": [ + { + "bbox": [ + 133, + 450, + 479, + 491 + ], + "type": "image", + "image_path": "adaf4da68cd717ac491e1f84afc9a5f418f9589c42e8f56cfbd7f95c38da5be8.jpg" + } + ] + } + ], + "index": 12, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 160, + 496, + 450, + 505 + ], + "lines": [ + { + "bbox": [ + 160, + 496, + 450, + 505 + ], + "spans": [ + { + "bbox": [ + 160, + 496, + 450, + 505 + ], + "type": "text", + "content": "A cute panda skateboarding in the sky, over snow covered mountains, with dreamy whimsical atmosphere." + } + ] + } + ], + "index": 13, + "angle": 0, + "type": "image_caption" + } + ], + "index": 12 + }, + { + "type": "image", + "bbox": [ + 133, + 510, + 479, + 552 + ], + "blocks": [ + { + "bbox": [ + 133, + 510, + 479, + 552 + ], + "lines": [ + { + "bbox": [ + 133, + 510, + 479, + 552 + ], + "spans": [ + { + "bbox": [ + 133, + 510, + 479, + 552 + ], + "type": "image", + "image_path": "b05555d5f782fdeb616ea285f818bd0552a812fb53692b4f84e47397cb382fab.jpg" + } + ] + } + ], + "index": 14, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 267, + 556, + 345, + 565 + ], + "lines": [ + { + "bbox": [ + 267, + 556, + 345, + 565 + ], + "spans": [ + { + "bbox": [ + 267, + 556, + 345, + 565 + ], + "type": "text", + "content": "An astronaut riding a horse." + } + ] + } + ], + "index": 15, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 130, + 578, + 482, + 623 + ], + "lines": [ + { + "bbox": [ + 130, + 578, + 482, + 623 + ], + "spans": [ + { + "bbox": [ + 130, + 578, + 482, + 623 + ], + "type": "text", + "content": "Fig. 3: Qualitative evaluation. Example videos generated by " + }, + { + "bbox": [ + 130, + 578, + 482, + 623 + ], + "type": "inline_equation", + "content": "W.A.L.T" + }, + { + "bbox": [ + 130, + 578, + 482, + 623 + ], + "type": "text", + "content": " from natural language prompts at " + }, + { + "bbox": [ + 130, + 578, + 482, + 623 + ], + "type": "inline_equation", + "content": "512 \\times 896" + }, + { + "bbox": [ + 130, + 578, + 482, + 623 + ], + "type": "text", + "content": " resolution over 3.6 seconds duration at 8 frames per second. The " + }, + { + "bbox": [ + 130, + 578, + 482, + 623 + ], + "type": "inline_equation", + "content": "W.A.L.T" + }, + { + "bbox": [ + 130, + 578, + 482, + 623 + ], + "type": "text", + "content": " model is able to generate temporally consistent photorealistic videos that align with the textual prompt. There are more results on our project website." + } + ] + } + ], + "index": 16, + "angle": 0, + "type": "image_caption" + } + ], + "index": 14 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "spans": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "type": "text", + "content": "12" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 166, + 91, + 230, + 102 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 166, + 91, + 230, + 102 + ], + "spans": [ + { + "bbox": [ + 166, + 91, + 230, + 102 + ], + "type": "text", + "content": "A. Gupta et al." + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 11 + }, + { + "para_blocks": [ + { + "type": "table", + "bbox": [ + 200, + 114, + 408, + 255 + ], + "blocks": [ + { + "bbox": [ + 200, + 114, + 408, + 255 + ], + "lines": [ + { + "bbox": [ + 200, + 114, + 408, + 255 + ], + "spans": [ + { + "bbox": [ + 200, + 114, + 408, + 255 + ], + "type": "table", + "html": "
MethodIS (↑)FVD (↓)
CogVideo (Chinese) [39]23.6751.3
CogVideo (English) [39]25.3701.6
MagicVideo [90]-699.0
EMU Video* [26]42.7606.2
Video LDM [5]33.5550.6
Make-A-Video [67]33.0367.2
PYoCo [25]47.8355.2
PixelDance* [85]42.1242.8
SVD* [4]-242.0
W.A.L.T (Ours) 419M (video only)26.8598.8
W.A.L.T (Ours) 419M (video + image)31.7344.5
W.A.L.T (Ours) 3B (video + image)35.1258.1
", + "image_path": "666b1a4e881dcc3d0e8c38d428b292d2dab85efac9f7280befba5d19e662ff85.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "table_body" + } + ], + "index": 2 + }, + { + "bbox": [ + 130, + 262, + 482, + 297 + ], + "lines": [ + { + "bbox": [ + 130, + 262, + 482, + 297 + ], + "spans": [ + { + "bbox": [ + 130, + 262, + 482, + 297 + ], + "type": "text", + "content": "Table 5: UCF-101 text-to-video generation. Joint training on image and video datasets in conjunction with scaling the model parameters is essential for high quality video generation. * Denotes concurrent work." + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "text" + }, + { + "bbox": [ + 130, + 330, + 481, + 413 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 330, + 481, + 413 + ], + "spans": [ + { + "bbox": [ + 130, + 330, + 481, + 413 + ], + "type": "text", + "content": "Quantitative Evaluation Evaluating text-conditioned video generation scientifically remains a significant challenge, in part due to the absence of standardized training datasets and benchmarks. So far we have focused our experiments and analyses on the standard academic benchmarks, which use the same training data to ensure controlled and fair comparisons. Nevertheless, to compare with prior work on text-to-video, we also report results on the UCF-101 dataset in the zero-shot evaluation protocol in Table 5 [25,39,67]. Also see supplement." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 130, + 415, + 482, + 498 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 415, + 482, + 498 + ], + "spans": [ + { + "bbox": [ + 130, + 415, + 482, + 498 + ], + "type": "text", + "content": "Joint training. A primary strength of our framework is its ability to train simultaneously on both image and video datasets. In Table 5 we ablate the impact of this joint training. Specifically, we trained two versions of W.A.L.T-L (each with 419M params.) models using the default settings specified in § 5.2. We find that joint training leads to a notable improvement across both metrics. Our results align with the findings of Ho et al. [38], who demonstrated the benefits of joint training for pixel-based video diffusion models with U-Net backbones." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 130, + 498, + 482, + 581 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 498, + 482, + 581 + ], + "spans": [ + { + "bbox": [ + 130, + 498, + 482, + 581 + ], + "type": "text", + "content": "Scaling. Transformers are known for their ability to scale effectively in many tasks [6,15,58]. In Table 5 we show the benefits of scaling our transformer model for video diffusion. Scaling our base model size leads to significant improvements on both the metrics. It is important to note, however, that our base model is considerably smaller than leading text-to-video systems. For instance, Ho et al. [35] trained base model of 5.7B parameters. Hence, we believe scaling our models further is an important direction of future work." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 130, + 582, + 482, + 666 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 582, + 482, + 666 + ], + "spans": [ + { + "bbox": [ + 130, + 582, + 482, + 666 + ], + "type": "text", + "content": "Comparison with prior work. In Table 5, we present a system-level comparison of various text-to-video methods. Our results are promising, surpassing all previous work in the FVD metric, and being competitive with concurrent work. In terms of the IS, W.A.L.T outperforms all previous work but PYoCo [25]. A possible explanation for this discrepancy might be PYoCo's use of stronger text embeddings. Specifically, they utilize both CLIP [56] and T5-XXL [61] encoders, whereas we employ a T5-XL [61] text encoder only." + } + ] + } + ], + "index": 7 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 223, + 90, + 447, + 101 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 223, + 90, + 447, + 101 + ], + "spans": [ + { + "bbox": [ + 223, + 90, + 447, + 101 + ], + "type": "text", + "content": "Photorealistic Video Generation with Diffusion Models" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 470, + 91, + 480, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 470, + 91, + 480, + 100 + ], + "spans": [ + { + "bbox": [ + 470, + 91, + 480, + 100 + ], + "type": "text", + "content": "13" + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 12 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 134, + 114, + 479, + 348 + ], + "blocks": [ + { + "bbox": [ + 134, + 114, + 479, + 348 + ], + "lines": [ + { + "bbox": [ + 134, + 114, + 479, + 348 + ], + "spans": [ + { + "bbox": [ + 134, + 114, + 479, + 348 + ], + "type": "image", + "image_path": "b4fc56e49d0c06ee7d076fb75256abb4700ab7f064b86ef37f7efeb8527cb724.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 130, + 359, + 482, + 392 + ], + "lines": [ + { + "bbox": [ + 130, + 359, + 482, + 392 + ], + "spans": [ + { + "bbox": [ + 130, + 359, + 482, + 392 + ], + "type": "text", + "content": "Fig. 4: Samples at " + }, + { + "bbox": [ + 130, + 359, + 482, + 392 + ], + "type": "inline_equation", + "content": "512 \\times 896" + }, + { + "bbox": [ + 130, + 359, + 482, + 392 + ], + "type": "text", + "content": " resolution. Top: Text-to-video samples of consistent 3D camera motion (5.1 secs). Bottom: Image-to-video samples (3.8 secs). Images are generated using off-shelf models. Visit our project website to view them in video format." + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_caption" + } + ], + "index": 2 + }, + { + "bbox": [ + 130, + 430, + 482, + 479 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 430, + 482, + 479 + ], + "spans": [ + { + "bbox": [ + 130, + 430, + 482, + 479 + ], + "type": "text", + "content": "Qualitative Results As mentioned in § 4.4, we jointly train our model on the task of frame prediction conditioned on 1 or 2 latent frames. Hence, our model can be used for animating images (image-to-video) and generating longer videos with consistent camera motion (Fig. 4). See videos on our project website." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 132, + 497, + 220, + 510 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 497, + 220, + 510 + ], + "spans": [ + { + "bbox": [ + 132, + 497, + 220, + 510 + ], + "type": "text", + "content": "6 Conclusion" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 130, + 521, + 482, + 666 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 521, + 482, + 666 + ], + "spans": [ + { + "bbox": [ + 130, + 521, + 482, + 666 + ], + "type": "text", + "content": "In this work, we introduce W.A.L.T, a simple, scalable, and efficient transformer-based framework for latent video diffusion models. A key design decision in our framework is to operate on a shared, compressed visual representation that enables generative modeling of both images and videos. We demonstrate state-of-the-art results for image and video generation using a transformer backbone with windowed attention. We also train a cascade of three W.A.L.T models jointly on image and video datasets, to synthesize high-resolution, temporally consistent photorealistic videos from natural language descriptions. While generative modeling has seen tremendous recent advances for images, progress on video generation has lagged behind. We hope that scaling our framework will enable the creation of powerful video generative models which can understand and model the physical world." + } + ] + } + ], + "index": 6 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "spans": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "type": "text", + "content": "14" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 166, + 91, + 230, + 102 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 166, + 91, + 230, + 102 + ], + "spans": [ + { + "bbox": [ + 166, + 91, + 230, + 102 + ], + "type": "text", + "content": "A. Gupta et al." + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 13 + }, + { + "para_blocks": [ + { + "bbox": [ + 133, + 114, + 246, + 129 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 114, + 246, + 129 + ], + "spans": [ + { + "bbox": [ + 133, + 114, + 246, + 129 + ], + "type": "text", + "content": "Acknowledgements" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 130, + 137, + 481, + 173 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 137, + 481, + 173 + ], + "spans": [ + { + "bbox": [ + 130, + 137, + 481, + 173 + ], + "type": "text", + "content": "We thank Bryan Seybold, Dan Kondratyuk, David Ross, Hartwig Adam, Huisheng Wang, Jason Baldridge, Mauricio Delbracio and Orly Liba for helpful discussions and feedback." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 133, + 190, + 197, + 202 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 190, + 197, + 202 + ], + "spans": [ + { + "bbox": [ + 133, + 190, + 197, + 202 + ], + "type": "text", + "content": "References" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 138, + 212, + 481, + 665 + ], + "type": "list", + "angle": 0, + "index": 20, + "blocks": [ + { + "bbox": [ + 138, + 212, + 481, + 245 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 212, + 481, + 245 + ], + "spans": [ + { + "bbox": [ + 138, + 212, + 481, + 245 + ], + "type": "text", + "content": "1. Agostinelli, A., Denk, T.I., Borsos, Z., Engel, J., Verzetti, M., Caillon, A., Huang, Q., Jansen, A., Roberts, A., Tagliasacchi, M., et al.: MusicLM: Generating music from text. arXiv:2301.11325 (2023) 2" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 138, + 245, + 481, + 266 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 245, + 481, + 266 + ], + "spans": [ + { + "bbox": [ + 138, + 245, + 481, + 266 + ], + "type": "text", + "content": "2. Bao, F., Li, C., Cao, Y., Zhu, J.: All are worth words: a vit backbone for score-based diffusion models. In: NeurIPS 2022 Workshop on Score-Based Methods (2022) 4" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 138, + 266, + 481, + 299 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 266, + 481, + 299 + ], + "spans": [ + { + "bbox": [ + 138, + 266, + 481, + 299 + ], + "type": "text", + "content": "3. Bengio, S., Vinyals, O., Jaitly, N., Shazeer, N.: Scheduled sampling for sequence prediction with recurrent neural networks. Advances in neural information processing systems 28 (2015) 7" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 138, + 299, + 481, + 342 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 299, + 481, + 342 + ], + "spans": [ + { + "bbox": [ + 138, + 299, + 481, + 342 + ], + "type": "text", + "content": "4. Blattmann, A., Dockhorn, T., Kulal, S., Mendelevitch, D., Kilian, M., Lorenz, D., Levi, Y., English, Z., Voleti, V., Letts, A., et al.: Stable video diffusion: Scaling latent video diffusion models to large datasets. arXiv preprint arXiv:2311.15127 (2023) 13" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 138, + 342, + 481, + 373 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 342, + 481, + 373 + ], + "spans": [ + { + "bbox": [ + 138, + 342, + 481, + 373 + ], + "type": "text", + "content": "5. Blattmann, A., Rombach, R., Ling, H., Dockhorn, T., Kim, S.W., Fidler, S., Kreis, K.: Align your latents: High-resolution video synthesis with latent diffusion models. In: CVPR (2023) 3, 4, 6, 13" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 138, + 374, + 481, + 407 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 374, + 481, + 407 + ], + "spans": [ + { + "bbox": [ + 138, + 374, + 481, + 407 + ], + "type": "text", + "content": "6. Bousmalis, K., Vezzani, G., Rao, D., Devin, C., Lee, A.X., Bauza, M., Davchev, T., Zhou, Y., Gupta, A., Raju, A., et al.: Robocat: A self-improving foundation agent for robotic manipulation. arXiv preprint arXiv:2306.11706 (2023) 2, 13" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 138, + 407, + 481, + 429 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 407, + 481, + 429 + ], + "spans": [ + { + "bbox": [ + 138, + 407, + 481, + 429 + ], + "type": "text", + "content": "7. Brock, A., Donahue, J., Simonyan, K.: Large scale GAN training for high fidelity natural image synthesis. In: ICLR (2018) 9" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 138, + 429, + 481, + 460 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 429, + 481, + 460 + ], + "spans": [ + { + "bbox": [ + 138, + 429, + 481, + 460 + ], + "type": "text", + "content": "8. Brohan, A., Brown, N., Carbajal, J., Chebotar, Y., Dabis, J., Finn, C., Gopalakrishnan, K., Hausman, K., Herzog, A., Hsu, J., et al.: Rt-1: Robotics transformer for real-world control at scale. arXiv preprint arXiv:2212.06817 (2022) 2" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 138, + 460, + 481, + 493 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 460, + 481, + 493 + ], + "spans": [ + { + "bbox": [ + 138, + 460, + 481, + 493 + ], + "type": "text", + "content": "9. Caron, M., Touvron, H., Misra, I., Jégou, H., Mairal, J., Bojanowski, P., Joulin, A.: Emerging properties in self-supervised vision transformers. In: Proceedings of the IEEE/CVF international conference on computer vision. pp. 9650-9660 (2021) 9" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 138, + 493, + 481, + 514 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 493, + 481, + 514 + ], + "spans": [ + { + "bbox": [ + 138, + 493, + 481, + 514 + ], + "type": "text", + "content": "0. Carreira, J., Noland, E., Banki-Horvath, A., Hillier, C., Zisserman, A.: A short note about Kinetics-600. arXiv:1808.01340 (2018) 3, 8" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 138, + 514, + 481, + 546 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 514, + 481, + 546 + ], + "spans": [ + { + "bbox": [ + 138, + 514, + 481, + 546 + ], + "type": "text", + "content": "1. Chang, H., Zhang, H., Barber, J., Maschinot, A., Lezama, J., Jiang, L., Yang, M.H., Murphy, K., Freeman, W.T., Rubinstein, M., et al.: Muse: Text-to-image generation via masked generative transformers. In: ICML (2023) 4" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 138, + 547, + 481, + 568 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 547, + 481, + 568 + ], + "spans": [ + { + "bbox": [ + 138, + 547, + 481, + 568 + ], + "type": "text", + "content": "2. Chang, H., Zhang, H., Jiang, L., Liu, C., Freeman, W.T.: MaskGIT: Masked generative image transformer. In: CVPR (2022) 4" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 138, + 568, + 481, + 601 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 568, + 481, + 601 + ], + "spans": [ + { + "bbox": [ + 138, + 568, + 481, + 601 + ], + "type": "text", + "content": "3. Chen, J., Yu, J., Ge, C., Yao, L., Xie, E., Wu, Y., Wang, Z., Kwok, J., Luo, P., Lu, H., et al.: Pixart-" + }, + { + "bbox": [ + 138, + 568, + 481, + 601 + ], + "type": "inline_equation", + "content": "\\alpha" + }, + { + "bbox": [ + 138, + 568, + 481, + 601 + ], + "type": "text", + "content": "beta: Fast training of diffusion transformer for photorealistic text-to-image synthesis. arXiv preprint arXiv:2310.00426 (2023) 4" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 138, + 601, + 481, + 621 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 601, + 481, + 621 + ], + "spans": [ + { + "bbox": [ + 138, + 601, + 481, + 621 + ], + "type": "text", + "content": "4. Chen, T., Zhang, R., Hinton, G.: Analog bits: Generating discrete data using diffusion models with self-conditioning. arXiv preprint arXiv:2208.04202 (2022) 7" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 138, + 621, + 481, + 665 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 621, + 481, + 665 + ], + "spans": [ + { + "bbox": [ + 138, + 621, + 481, + 665 + ], + "type": "text", + "content": "5. Dehghani, M., Djolonga, J., Mustafa, B., Padlewski, P., Heek, J., Gilmer, J., Steiner, A.P., Caron, M., Geirhos, R., Alabdulmohsin, I., et al.: Scaling vision transformers to 22 billion parameters. In: International Conference on Machine Learning. pp. 7480-7512. PMLR (2023) 10, 11, 13" + } + ] + } + ], + "index": 19 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 223, + 91, + 447, + 101 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 223, + 91, + 447, + 101 + ], + "spans": [ + { + "bbox": [ + 223, + 91, + 447, + 101 + ], + "type": "text", + "content": "Photorealistic Video Generation with Diffusion Models" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 470, + 92, + 480, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 470, + 92, + 480, + 100 + ], + "spans": [ + { + "bbox": [ + 470, + 92, + 480, + 100 + ], + "type": "text", + "content": "15" + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 14 + }, + { + "para_blocks": [ + { + "bbox": [ + 132, + 116, + 481, + 665 + ], + "type": "list", + "angle": 0, + "index": 23, + "blocks": [ + { + "bbox": [ + 133, + 116, + 480, + 138 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 116, + 480, + 138 + ], + "spans": [ + { + "bbox": [ + 133, + 116, + 480, + 138 + ], + "type": "text", + "content": "16. Deng, J., Dong, W., Socher, R., Li, L.J., Li, K., Fei-Fei, L.: ImageNet: A large-scale hierarchical image database. In: CVPR (2009) 3" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 133, + 139, + 481, + 159 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 139, + 481, + 159 + ], + "spans": [ + { + "bbox": [ + 133, + 139, + 481, + 159 + ], + "type": "text", + "content": "17. Dhariwal, P., Nichol, A.: Diffusion models beat GANs on image synthesis. In: NeurIPS (2021) 2, 7, 9" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 133, + 160, + 481, + 192 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 160, + 481, + 192 + ], + "spans": [ + { + "bbox": [ + 133, + 160, + 481, + 192 + ], + "type": "text", + "content": "18. Ding, M., Yang, Z., Hong, W., Zheng, W., Zhou, C., Yin, D., Lin, J., Zou, X., Shao, Z., Yang, H., et al.:CogView: Mastering text-to-image generation via transformers. In:NeurIPS (2021) 4" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 132, + 193, + 481, + 224 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 193, + 481, + 224 + ], + "spans": [ + { + "bbox": [ + 132, + 193, + 481, + 224 + ], + "type": "text", + "content": "19. Dosovitskiy, A., Beyer, L., Kolesnikov, A., Weissenborn, D., Zhai, X., Unterthiner, T., Dehghani, M., Minderer, M., Heigold, G., Gelly, S., et al.: An image is worth 16x16 words: Transformers for image recognition at scale. In: ICLR (2020) 2, 6, 9" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 132, + 225, + 481, + 246 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 225, + 481, + 246 + ], + "spans": [ + { + "bbox": [ + 132, + 225, + 481, + 246 + ], + "type": "text", + "content": "20. Dumoulin, V., Shlens, J., Kudlur, M.: A learned representation for artistic style. arXiv preprint arXiv:1610.07629 (2016) 7" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 132, + 247, + 481, + 267 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 247, + 481, + 267 + ], + "spans": [ + { + "bbox": [ + 132, + 247, + 481, + 267 + ], + "type": "text", + "content": "21. Esser, P., Rombach, R., Ommer, B.: Taming transformers for high-resolution image synthesis. In: CVPR (2021) 2, 4, 5" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 132, + 268, + 481, + 300 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 268, + 481, + 300 + ], + "spans": [ + { + "bbox": [ + 132, + 268, + 481, + 300 + ], + "type": "text", + "content": "22. Gafni, O., Polyak, A., Ashual, O., Sheynin, S., Parikh, D., Taigman, Y.: Makea-scene: Scene-based text-to-image generation with human priors. In: European Conference on Computer Vision. pp. 89-106. Springer (2022) 4" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 132, + 300, + 481, + 320 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 300, + 481, + 320 + ], + "spans": [ + { + "bbox": [ + 132, + 300, + 481, + 320 + ], + "type": "text", + "content": "23. Gao, S., Zhou, P., Cheng, M.M., Yan, S.: Masked diffusion transformer is a strong image synthesizer. arXiv:2303.14389 (2023) 4, 9" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 132, + 322, + 481, + 353 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 322, + 481, + 353 + ], + "spans": [ + { + "bbox": [ + 132, + 322, + 481, + 353 + ], + "type": "text", + "content": "24. Ge, S., Hayes, T., Yang, H., Yin, X., Pang, G., Jacobs, D., Huang, J.B., Parikh, D.: Long video generation with time-agnostic VQGAN and time-sensitive transformer. In: ECCV (2022) 4, 5, 7" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 132, + 354, + 481, + 386 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 354, + 481, + 386 + ], + "spans": [ + { + "bbox": [ + 132, + 354, + 481, + 386 + ], + "type": "text", + "content": "25. Ge, S., Nah, S., Liu, G., Poon, T., Tao, A., Catanzaro, B., Jacobs, D., Huang, J.B., Liu, M.Y., Balaji, Y.: Preserve your own correlation: A noise prior for video diffusion models. arXiv preprint arXiv:2305.10474 (2023) 3, 6, 13" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 132, + 386, + 481, + 418 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 386, + 481, + 418 + ], + "spans": [ + { + "bbox": [ + 132, + 386, + 481, + 418 + ], + "type": "text", + "content": "26. Girdhar, R., Singh, M., Brown, A., Duval, Q., Azadi, S., Rambhatla, S.S., Shah, A., Yin, X., Parikh, D., Misra, I.: Emu video: Factorizing text-to-video generation by explicit image conditioning. arXiv preprint arXiv:2311.10709 (2023) 13" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 132, + 418, + 481, + 450 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 418, + 481, + 450 + ], + "spans": [ + { + "bbox": [ + 132, + 418, + 481, + 450 + ], + "type": "text", + "content": "27. Goodfellow, I., Pouget-Abadie, J., Mirza, M., Xu, B., Warde-Farley, D., Ozair, S., Courville, A., Bengio, Y.: Generative adversarial nets. Advances in neural information processing systems 27 (2014) 5" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 132, + 451, + 397, + 460 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 451, + 397, + 460 + ], + "spans": [ + { + "bbox": [ + 132, + 451, + 397, + 460 + ], + "type": "text", + "content": "28. Google: PaLM 2 technical report. arXiv:2305.10403 (2023) 2" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 132, + 461, + 481, + 482 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 461, + 481, + 482 + ], + "spans": [ + { + "bbox": [ + 132, + 461, + 481, + 482 + ], + "type": "text", + "content": "29. Gupta, A., Tian, S., Zhang, Y., Wu, J., Martín-Martín, R., Fei-Fei, L.: MaskViT: Masked visual pre-training for video prediction. In: ICLR (2022) 4, 6" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 132, + 483, + 481, + 504 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 483, + 481, + 504 + ], + "spans": [ + { + "bbox": [ + 132, + 483, + 481, + 504 + ], + "type": "text", + "content": "30. Gupta, A., Wu, J., Deng, J., Fei-Fei, L.: Siamese masked autoencoders. arXiv preprint arXiv:2305.14344 (2023) 9" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 132, + 504, + 481, + 536 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 504, + 481, + 536 + ], + "spans": [ + { + "bbox": [ + 132, + 504, + 481, + 536 + ], + "type": "text", + "content": "31. Harvey, W., Naderiparizi, S., Masrani, V., Weilbach, C., Wood, F.: Flexible diffusion modeling of long videos. Advances in Neural Information Processing Systems 35, 27953-27965 (2022) 3" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 132, + 536, + 481, + 557 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 536, + 481, + 557 + ], + "spans": [ + { + "bbox": [ + 132, + 536, + 481, + 557 + ], + "type": "text", + "content": "32. He, K., Chen, X., Xie, S., Li, Y., Dollar, P., Girshick, R.: Masked autoencoders are scalable vision learners. In: CVPR (2022) 2" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 132, + 558, + 481, + 578 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 558, + 481, + 578 + ], + "spans": [ + { + "bbox": [ + 132, + 558, + 481, + 578 + ], + "type": "text", + "content": "33. He, Y., Yang, T., Zhang, Y., Shan, Y., Chen, Q.: Latent video diffusion models for high-fidelity long video generation. arXiv preprint arXiv:2211.13221 (2023) 3" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 132, + 579, + 481, + 610 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 579, + 481, + 610 + ], + "spans": [ + { + "bbox": [ + 132, + 579, + 481, + 610 + ], + "type": "text", + "content": "34. Heusel, M., Ramsauer, H., Unterthiner, T., Nessler, B., Hochreiter, S.: GANs trained by a two time-scale update rule converge to a local nash equilibrium. In: NeurIPS (2017) 9" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 132, + 611, + 481, + 643 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 611, + 481, + 643 + ], + "spans": [ + { + "bbox": [ + 132, + 611, + 481, + 643 + ], + "type": "text", + "content": "35. Ho, J., Chan, W., Sahara, C., Whang, J., Gao, R., Gritsenko, A., Kingma, D.P., Poole, B., Norouzi, M., Fleet, D.J., et al.: Imagen video: High definition video generation with diffusion models. arXiv:2210.02303 (2022) 3, 4, 5, 13" + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 132, + 644, + 481, + 665 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 644, + 481, + 665 + ], + "spans": [ + { + "bbox": [ + 132, + 644, + 481, + 665 + ], + "type": "text", + "content": "36. Ho, J., Jain, A., Abbeel, P.: Denoising diffusion probabilistic models. In: NeurIPS (2020) 2, 3, 4" + } + ] + } + ], + "index": 22 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "spans": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "type": "text", + "content": "16" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 166, + 91, + 229, + 102 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 166, + 91, + 229, + 102 + ], + "spans": [ + { + "bbox": [ + 166, + 91, + 229, + 102 + ], + "type": "text", + "content": "A. Gupta et al." + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 15 + }, + { + "para_blocks": [ + { + "bbox": [ + 130, + 116, + 482, + 665 + ], + "type": "list", + "angle": 0, + "index": 23, + "blocks": [ + { + "bbox": [ + 130, + 116, + 482, + 148 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 116, + 482, + 148 + ], + "spans": [ + { + "bbox": [ + 130, + 116, + 482, + 148 + ], + "type": "text", + "content": "37. Ho, J., Sahara, C., Chan, W., Fleet, D.J., Norouzi, M., Salimans, T.: Cascaded diffusion models for high fidelity image generation. JMLR 23(1), 2249-2281 (2022) 8" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 130, + 149, + 482, + 171 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 149, + 482, + 171 + ], + "spans": [ + { + "bbox": [ + 130, + 149, + 482, + 171 + ], + "type": "text", + "content": "38. Ho, J., Salimans, T., Gritsenko, A., Chan, W., Norouzi, M., Fleet, D.J.: Video diffusion models. In: ICLR Workshops (2022) 2, 3, 4, 7, 13" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 130, + 172, + 482, + 192 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 172, + 482, + 192 + ], + "spans": [ + { + "bbox": [ + 130, + 172, + 482, + 192 + ], + "type": "text", + "content": "39. Hong, W., Ding, M., Zheng, W., Liu, X., Tang, J.: CogVideo: Large-scale pretraining for text-to-video generation via transformers. arXiv:2205.15868 (2022) 4, 13" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 130, + 193, + 482, + 213 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 193, + 482, + 213 + ], + "spans": [ + { + "bbox": [ + 130, + 193, + 482, + 213 + ], + "type": "text", + "content": "40. Hoogeboom, E., Heek, J., Salimans, T.: simple diffusion: End-to-end diffusion for high resolution images. In: ICML (2023) 3, 4, 9" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 130, + 213, + 482, + 236 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 213, + 482, + 236 + ], + "spans": [ + { + "bbox": [ + 130, + 213, + 482, + 236 + ], + "type": "text", + "content": "41. Hu, E.J., Shen, Y., Wallis, P., Allen-Zhu, Z., Li, Y., Wang, S., Wang, L., Chen, W.: Lora: Low-rank adaptation of large language models. In: ICLR (2021) 7" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 130, + 236, + 482, + 256 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 236, + 482, + 256 + ], + "spans": [ + { + "bbox": [ + 130, + 236, + 482, + 256 + ], + "type": "text", + "content": "42. Jabri, A., Fleet, D.J., Chen, T.: Scalable adaptive computation for iterative generation. In: ICML (2023) 4, 7, 9" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 130, + 257, + 482, + 289 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 257, + 482, + 289 + ], + "spans": [ + { + "bbox": [ + 130, + 257, + 482, + 289 + ], + "type": "text", + "content": "43. Jiang, Y., Chang, S., Wang, Z.: Transgan: Two pure transformers can make one strong gan, and that can scale up. Advances in Neural Information Processing Systems 34, 14745-14758 (2021) 4" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 130, + 289, + 482, + 310 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 289, + 482, + 310 + ], + "spans": [ + { + "bbox": [ + 130, + 289, + 482, + 310 + ], + "type": "text", + "content": "44. Johnson, J., Alahi, A., Fei-Fei, L.: Perceptual losses for real-time style transfer and super-resolution. In: ECCV (2016) 5" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 130, + 310, + 482, + 331 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 310, + 482, + 331 + ], + "spans": [ + { + "bbox": [ + 130, + 310, + 482, + 331 + ], + "type": "text", + "content": "45. Karras, T., Laine, S., Aila, T.: A style-based generator architecture for generative adversarial networks. In: CVPR (2019) 7" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 130, + 332, + 482, + 353 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 332, + 482, + 353 + ], + "spans": [ + { + "bbox": [ + 130, + 332, + 482, + 353 + ], + "type": "text", + "content": "46. Kingma, D.P., Gao, R.: Understanding the diffusion objective as a weighted integral of ellbos. arXiv:2303.00848 (2023) 9" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 130, + 353, + 482, + 374 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 353, + 482, + 374 + ], + "spans": [ + { + "bbox": [ + 130, + 353, + 482, + 374 + ], + "type": "text", + "content": "47. LeCun, Y., Bottou, L., Bengio, Y., Haffner, P.: Gradient-based learning applied to document recognition. Proceedings of the IEEE 86(11), 2278-2324 (1998) 2" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 130, + 375, + 482, + 396 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 375, + 482, + 396 + ], + "spans": [ + { + "bbox": [ + 130, + 375, + 482, + 396 + ], + "type": "text", + "content": "48. Lee, K., Chang, H., Jiang, L., Zhang, H., Tu, Z., Liu, C.: Vitgan: Training gans with vision transformers. arXiv preprint arXiv:2107.04589 (2021) 4" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 130, + 396, + 482, + 417 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 396, + 482, + 417 + ], + "spans": [ + { + "bbox": [ + 130, + 396, + 482, + 417 + ], + "type": "text", + "content": "49. Lin, S., Liu, B., Li, J., Yang, X.: Common diffusion noise schedules and sample steps are flawed. arXiv preprint arXiv:2305.08891 (2023) 10, 11" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 130, + 417, + 482, + 460 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 417, + 482, + 460 + ], + "spans": [ + { + "bbox": [ + 130, + 417, + 482, + 460 + ], + "type": "text", + "content": "50. Liu, Z., Lin, Y., Cao, Y., Hu, H., Wei, Y., Zhang, Z., Lin, S., Guo, B.: Swin transformer: Hierarchical vision transformer using shifted windows. In: Proceedings of the IEEE/CVF international conference on computer vision. pp. 10012-10022 (2021) 6" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 130, + 460, + 482, + 492 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 460, + 482, + 492 + ], + "spans": [ + { + "bbox": [ + 130, + 460, + 482, + 492 + ], + "type": "text", + "content": "51. Lu, H., Yang, G., Fei, N., Huo, Y., Lu, Z., Luo, P., Ding, M.: Vdt: General-purpose video diffusion transformers via mask modeling. arXiv preprint arXiv:2305.13311 (2023) 4" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 130, + 492, + 482, + 525 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 492, + 482, + 525 + ], + "spans": [ + { + "bbox": [ + 130, + 492, + 482, + 525 + ], + "type": "text", + "content": "52. Luc, P., Clark, A., Dieleman, S., Casas, D.d.L., Doron, Y., Cassirer, A., Simonyan, K.: Transformation-based adversarial video prediction on large-scale data. arXiv:2003.04035 (2020) 7" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 130, + 525, + 482, + 547 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 525, + 482, + 547 + ], + "spans": [ + { + "bbox": [ + 130, + 525, + 482, + 547 + ], + "type": "text", + "content": "53. Nichol, A.Q., Dhariwal, P.: Improved denoising diffusion probabilistic models. In: International Conference on Machine Learning. pp. 8162-8171. PMLR (2021) 3, 7" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 130, + 547, + 482, + 568 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 547, + 482, + 568 + ], + "spans": [ + { + "bbox": [ + 130, + 547, + 482, + 568 + ], + "type": "text", + "content": "54. Peebles, W., Xie, S.: Scalable diffusion models with transformers. arXiv:2212.09748 (2022) 4, 7, 9" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 130, + 568, + 482, + 600 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 568, + 482, + 600 + ], + "spans": [ + { + "bbox": [ + 130, + 568, + 482, + 600 + ], + "type": "text", + "content": "55. Perez, E., Strub, F., De Vries, H., Dumoulin, V., Courville, A.: Film: Visual reasoning with a general conditioning layer. In: Proceedings of the AAAI conference on artificial intelligence. vol. 32 (2018) 7" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 130, + 601, + 482, + 632 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 601, + 482, + 632 + ], + "spans": [ + { + "bbox": [ + 130, + 601, + 482, + 632 + ], + "type": "text", + "content": "56. Radford, A., Kim, J.W., Hallacy, C., Ramesh, A., Goh, G., Agarwal, S., Sastry, G., Askell, A., Mishkin, P., Clark, J., et al.: Learning transferable visual models from natural language supervision. In: ICML (2021) 2, 4, 13" + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 130, + 632, + 482, + 665 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 632, + 482, + 665 + ], + "spans": [ + { + "bbox": [ + 130, + 632, + 482, + 665 + ], + "type": "text", + "content": "57. Radford, A., Kim, J.W., Xu, T., Brockman, G., McLeavey, C., Sutskever, I.: Robust speech recognition via large-scale weak supervision. In: International Conference on Machine Learning. pp. 28492-28518. PMLR (2023) 2" + } + ] + } + ], + "index": 22 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 223, + 91, + 447, + 100 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 223, + 91, + 447, + 100 + ], + "spans": [ + { + "bbox": [ + 223, + 91, + 447, + 100 + ], + "type": "text", + "content": "Photorealistic Video Generation with Diffusion Models" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 470, + 92, + 480, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 470, + 92, + 480, + 100 + ], + "spans": [ + { + "bbox": [ + 470, + 92, + 480, + 100 + ], + "type": "text", + "content": "17" + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 16 + }, + { + "para_blocks": [ + { + "bbox": [ + 132, + 117, + 481, + 665 + ], + "type": "list", + "angle": 0, + "index": 23, + "blocks": [ + { + "bbox": [ + 133, + 117, + 480, + 138 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 117, + 480, + 138 + ], + "spans": [ + { + "bbox": [ + 133, + 117, + 480, + 138 + ], + "type": "text", + "content": "58. Radford, A., Narasimhan, K., Salimans, T., Sutskever, I., et al.: Improving language understanding by generative pre-training (2018) 2, 4, 13" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 132, + 140, + 480, + 160 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 140, + 480, + 160 + ], + "spans": [ + { + "bbox": [ + 132, + 140, + 480, + 160 + ], + "type": "text", + "content": "59. Radford, A., Wu, J., Child, R., Luan, D., Amodei, D., Sutskever, I., et al.: Language models are unsupervised multitask learners. OpenAI blog 1(8), 9 (2019) 2, 4" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 133, + 161, + 481, + 182 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 161, + 481, + 182 + ], + "spans": [ + { + "bbox": [ + 133, + 161, + 481, + 182 + ], + "type": "text", + "content": "60. Ramesh, A., Pavlov, M., Goh, G., Gray, S., Voss, C., Radford, A., Chen, M., Sutskever, I.: Zero-shot text-to-image generation. In: ICML (2021) 4" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 132, + 183, + 481, + 215 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 183, + 481, + 215 + ], + "spans": [ + { + "bbox": [ + 132, + 183, + 481, + 215 + ], + "type": "text", + "content": "61. Roberts, A., Raffel, C., Lee, K., Matena, M., Shazeer, N., Liu, P.J., Narang, S., Li, W., Zhou, Y.: Exploring the limits of transfer learning with a unified text-to-text transformer (2019) 13" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 132, + 216, + 480, + 237 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 216, + 480, + 237 + ], + "spans": [ + { + "bbox": [ + 132, + 216, + 480, + 237 + ], + "type": "text", + "content": "62. Rombach, R., Blattmann, A., Lorenz, D., Esser, P., Ommer, B.: High-resolution image synthesis with latent diffusion models. In: CVPR (2022) 2, 3, 9, 11" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 132, + 238, + 480, + 258 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 238, + 480, + 258 + ], + "spans": [ + { + "bbox": [ + 132, + 238, + 480, + 258 + ], + "type": "text", + "content": "63. Ronneberger, O., Fischer, P., Brox, T.: U-Net: Convolutional networks for biomedical image segmentation. In: MICCAI (2015) 2, 4" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 132, + 259, + 480, + 281 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 259, + 480, + 281 + ], + "spans": [ + { + "bbox": [ + 132, + 259, + 480, + 281 + ], + "type": "text", + "content": "64. Salimans, T., Goodfellow, I., Zaremba, W., Cheung, V., Radford, A., Chen, X.: Improved techniques for training GANs. In: NeurIPS (2016) 9" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 132, + 282, + 480, + 303 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 282, + 480, + 303 + ], + "spans": [ + { + "bbox": [ + 132, + 282, + 480, + 303 + ], + "type": "text", + "content": "65. Salimans, T., Ho, J.: Progressive distillation for fast sampling of diffusion models arXiv preprint arXiv:2202.00512 (2022) 5" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 132, + 304, + 480, + 334 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 304, + 480, + 334 + ], + "spans": [ + { + "bbox": [ + 132, + 304, + 480, + 334 + ], + "type": "text", + "content": "66. Savinov, N., Chung, J., Binkowski, M., Elsen, E., Oord, A.v.d.: Step-unrolled denoising autoencoders for text generation. arXiv preprint arXiv:2112.06749 (2021) 7" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 132, + 335, + 480, + 369 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 335, + 480, + 369 + ], + "spans": [ + { + "bbox": [ + 132, + 335, + 480, + 369 + ], + "type": "text", + "content": "67. Singer, U., Polyak, A., Hayes, T., Yin, X., An, J., Zhang, S., Hu, Q., Yang, H., Ashual, O., Gafni, O., et al.: Make-a-video: Text-to-video generation without text-video data. arXiv:2209.14792 (2022) 3, 4, 13" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 132, + 369, + 480, + 391 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 369, + 480, + 391 + ], + "spans": [ + { + "bbox": [ + 132, + 369, + 480, + 391 + ], + "type": "text", + "content": "68. Sohl-Dickstein, J., Weiss, E., Maheswaranathan, N., Ganguli, S.: Deep unsupervised learning using nonequilibrium thermodynamics. In: ICML (2015) 2, 3, 4" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 132, + 392, + 480, + 413 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 392, + 480, + 413 + ], + "spans": [ + { + "bbox": [ + 132, + 392, + 480, + 413 + ], + "type": "text", + "content": "69. Song, J., Meng, C., Ermon, S.: Denoising diffusion implicit models. arXiv preprint arXiv:2010.02502 (2020) 3, 9" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 132, + 414, + 480, + 434 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 414, + 480, + 434 + ], + "spans": [ + { + "bbox": [ + 132, + 414, + 480, + 434 + ], + "type": "text", + "content": "70. Song, Y., Ermon, S.: Generative modeling by estimating gradients of the data distribution. In: NeurIPS (2019) 2, 4" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 132, + 435, + 480, + 456 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 435, + 480, + 456 + ], + "spans": [ + { + "bbox": [ + 132, + 435, + 480, + 456 + ], + "type": "text", + "content": "71. Soomro, K., Zamir, A.R., Shah, M.: UCF101: A dataset of 101 human actions classes from videos in the wild. arXiv:1212.0402 (2012) 3, 8, 10" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 133, + 457, + 481, + 489 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 457, + 481, + 489 + ], + "spans": [ + { + "bbox": [ + 133, + 457, + 481, + 489 + ], + "type": "text", + "content": "72. Unterthiner, T., van Steenkiste, S., Kurach, K., Marinier, R., Michalski, M., Gelly, S.: Towards accurate generative models of video: A new metric & challenges. arXiv:1812.01717 (2018) 8" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 132, + 491, + 480, + 511 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 491, + 480, + 511 + ], + "spans": [ + { + "bbox": [ + 132, + 491, + 480, + 511 + ], + "type": "text", + "content": "73. Van Den Oord, A., Vinyals, O., et al.: Neural discrete representation learning. In: NeurIPS (2017) 2, 5" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 132, + 512, + 480, + 533 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 512, + 480, + 533 + ], + "spans": [ + { + "bbox": [ + 132, + 512, + 480, + 533 + ], + "type": "text", + "content": "74. Vaswani, A., Shazeer, N., Parmar, N., Uszkoreit, J., Jones, L., Gomez, A.N., Kaiser, L., Polosukhin, I.: Attention is all you need. In: NeurIPS (2017) 2, 4, 6" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 132, + 534, + 480, + 566 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 534, + 480, + 566 + ], + "spans": [ + { + "bbox": [ + 132, + 534, + 480, + 566 + ], + "type": "text", + "content": "75. Villegas, R., Babaeizadeh, M., Kindermans, P.J., Moraldo, H., Zhang, H., Saffar, M.T., Castro, S., Kunze, J., Erhan, D.: Phenaki: Variable length video generation from open domain textual description. arXiv:2210.02399 (2022) 4, 5, 7" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 132, + 567, + 480, + 599 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 567, + 480, + 599 + ], + "spans": [ + { + "bbox": [ + 132, + 567, + 480, + 599 + ], + "type": "text", + "content": "76. Vincent, P., Larochelle, H., Bengio, Y., Manzagol, P.A.: Extracting and composing robust features with denoising autoencoders. In: Proceedings of the 25th international conference on Machine learning. pp. 1096-1103 (2008) 2" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 132, + 600, + 480, + 632 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 600, + 480, + 632 + ], + "spans": [ + { + "bbox": [ + 132, + 600, + 480, + 632 + ], + "type": "text", + "content": "77. Wu, C., Huang, L., Zhang, Q., Li, B., Ji, L., Yang, F., Sapiro, G., Duan, N.: Godiva: Generating open-domain videos from natural descriptions. arXiv preprint arXiv:2104.14806 (2021) 4" + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 132, + 632, + 480, + 665 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 632, + 480, + 665 + ], + "spans": [ + { + "bbox": [ + 132, + 632, + 480, + 665 + ], + "type": "text", + "content": "78. Wu, C., Liang, J., Ji, L., Yang, F., Fang, Y., Jiang, D., Duan, N.: Niwa: Visual synthesis pre-training for neural visual world creation. In: European conference on computer vision. pp. 720-736. Springer (2022) 4" + } + ] + } + ], + "index": 22 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "spans": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "type": "text", + "content": "18" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 166, + 91, + 229, + 102 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 166, + 91, + 229, + 102 + ], + "spans": [ + { + "bbox": [ + 166, + 91, + 229, + 102 + ], + "type": "text", + "content": "A. Gupta et al." + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 17 + }, + { + "para_blocks": [ + { + "bbox": [ + 130, + 116, + 482, + 511 + ], + "type": "list", + "angle": 0, + "index": 15, + "blocks": [ + { + "bbox": [ + 130, + 116, + 482, + 138 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 116, + 482, + 138 + ], + "spans": [ + { + "bbox": [ + 130, + 116, + 482, + 138 + ], + "type": "text", + "content": "79. Yan, W., Zhang, Y., Abbeel, P., Srinivas, A.: VideoGPT: Video generation using VQ-VAE and transformers. arXiv:2104.10157 (2021) 4" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 130, + 138, + 482, + 171 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 138, + 482, + 171 + ], + "spans": [ + { + "bbox": [ + 130, + 138, + 482, + 171 + ], + "type": "text", + "content": "80. Yu, J., Li, X., Koh, J.Y., Zhang, H., Pang, R., Qin, J., Ku, A., Xu, Y., Baldridge, J., Wu, Y.: Vector-quantized image modeling with improved VQGAN. In: ICLR (2022) 4" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 132, + 172, + 482, + 205 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 172, + 482, + 205 + ], + "spans": [ + { + "bbox": [ + 132, + 172, + 482, + 205 + ], + "type": "text", + "content": "81. Yu, J., Xu, Y., Koh, J.Y., Luong, T., Baid, G., Wang, Z., Vasudevan, V., Ku, A., Yang, Y., Ayan, B.K., et al.: Scaling autoregressive models for content-rich text-to-image generation. arXiv:2206.10789 (2022) 4" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 132, + 205, + 482, + 237 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 205, + 482, + 237 + ], + "spans": [ + { + "bbox": [ + 132, + 205, + 482, + 237 + ], + "type": "text", + "content": "82. Yu, L., Cheng, Y., Sohn, K., Lezama, J., Zhang, H., Chang, H., Hauptmann, A.G., Yang, M.H., Hao, Y., Essa, I., et al.: MAGVIT: Masked generative video transformer. In: CVPR (2023) 4, 5, 7" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 132, + 238, + 482, + 281 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 238, + 482, + 281 + ], + "spans": [ + { + "bbox": [ + 132, + 238, + 482, + 281 + ], + "type": "text", + "content": "83. Yu, L., Lezama, J., Gundavarapu, N.B., Versari, L., Sohn, K., Minnen, D., Cheng, Y., Gupta, A., Gu, X., Hauptmann, A.G., et al.: Language model beats diffusion–tokenizer is key to visual generation. arXiv preprint arXiv:2310.05737 (2023) 5, 7" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 132, + 281, + 482, + 314 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 281, + 482, + 314 + ], + "spans": [ + { + "bbox": [ + 132, + 281, + 482, + 314 + ], + "type": "text", + "content": "84. Yu, S., Sohn, K., Kim, S., Shin, J.: Video probabilistic diffusion models in projected latent space. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 18456-18466 (2023) 3" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 132, + 315, + 482, + 335 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 315, + 482, + 335 + ], + "spans": [ + { + "bbox": [ + 132, + 315, + 482, + 335 + ], + "type": "text", + "content": "85. Zeng, Y., Wei, G., Zheng, J., Zou, J., Wei, Y., Zhang, Y., Li, H.: Make pixels dance: High-dynamic video generation. arXiv preprint arXiv:2311.10982 (2023) 13" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 132, + 336, + 482, + 369 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 336, + 482, + 369 + ], + "spans": [ + { + "bbox": [ + 132, + 336, + 482, + 369 + ], + "type": "text", + "content": "86. Zhai, X., Kolesnikov, A., Houlsby, N., Beyer, L.: Scaling vision transformers. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 12104-12113 (2022) 9" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 132, + 370, + 482, + 413 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 370, + 482, + 413 + ], + "spans": [ + { + "bbox": [ + 132, + 370, + 482, + 413 + ], + "type": "text", + "content": "87. Zhang, B., Gu, S., Zhang, B., Bao, J., Chen, D., Wen, F., Wang, Y., Guo, B.: Styleswin: Transformer-based gan for high-resolution image generation. In: Proceedings of the IEEE/CVF conference on computer vision and pattern recognition. pp. 11304-11314 (2022) 4" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 132, + 413, + 482, + 435 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 413, + 482, + 435 + ], + "spans": [ + { + "bbox": [ + 132, + 413, + 482, + 435 + ], + "type": "text", + "content": "88. Zhang, R., Isola, P., Efros, A.A., Shechtman, E., Wang, O.: The unreasonable effectiveness of deep features as a perceptual metric. In: CVPR (2018) 5" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 132, + 435, + 482, + 456 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 435, + 482, + 456 + ], + "spans": [ + { + "bbox": [ + 132, + 435, + 482, + 456 + ], + "type": "text", + "content": "89. Zheng, H., Nie, W., Vahdat, A., Anandkumar, A.: Fast training of diffusion models with masked transformers. arXiv:2306.09305 (2023) 4, 9" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 132, + 457, + 482, + 479 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 457, + 482, + 479 + ], + "spans": [ + { + "bbox": [ + 132, + 457, + 482, + 479 + ], + "type": "text", + "content": "90. Zhou, D., Wang, W., Yan, H., Lv, W., Zhu, Y., Feng, J.: Magicvideo: Efficient video generation with latent diffusion models. arXiv preprint arXiv:2211.11018 (2022) 13" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 132, + 479, + 482, + 511 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 479, + 482, + 511 + ], + "spans": [ + { + "bbox": [ + 132, + 479, + 482, + 511 + ], + "type": "text", + "content": "91. Zitkovich, B., Yu, T., Xu, S., Xu, P., Xiao, T., Xia, F., Wu, J., Wohlhart, P., Welker, S., Wahid, A., et al.: RT-2: Vision-language-action models transfer web knowledge to robotic control. In: CoRL (2023) 2" + } + ] + } + ], + "index": 14 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 223, + 91, + 447, + 101 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 223, + 91, + 447, + 101 + ], + "spans": [ + { + "bbox": [ + 223, + 91, + 447, + 101 + ], + "type": "text", + "content": "Photorealistic Video Generation with Diffusion Models" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 470, + 92, + 480, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 470, + 92, + 480, + 100 + ], + "spans": [ + { + "bbox": [ + 470, + 92, + 480, + 100 + ], + "type": "text", + "content": "19" + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 18 + } + ], + "_backend": "vlm", + "_version_name": "2.6.4" +} \ No newline at end of file diff --git a/2024/PhysAvatar_ Learning the Physics of Dressed 3D Avatars from Visual Observations/958c8811-3fcb-45f8-9957-bab313d0628e_content_list.json b/2024/PhysAvatar_ Learning the Physics of Dressed 3D Avatars from Visual Observations/958c8811-3fcb-45f8-9957-bab313d0628e_content_list.json new file mode 100644 index 0000000000000000000000000000000000000000..4607c2a32c3d1b086aa7ff1ca9ad73a38b8bf5f0 --- /dev/null +++ b/2024/PhysAvatar_ Learning the Physics of Dressed 3D Avatars from Visual Observations/958c8811-3fcb-45f8-9957-bab313d0628e_content_list.json @@ -0,0 +1,2153 @@ +[ + { + "type": "text", + "text": "PhysAvatar: Learning the Physics of Dressed 3D Avatars from Visual Observations", + "text_level": 1, + "bbox": [ + 223, + 140, + 779, + 186 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Yang Zheng $^{1\\star}$ , Qingqing Zhao $^{1\\star}$ , Guandao Yang $^{1}$ , Wang Yifan $^{1}$ , Donglai Xiang $^{2}$ , Florian Dubost $^{3}$ , Dmitry Lagun $^{3}$ , Thabo Beeler $^{3}$ , Federico Tombari $^{3,4}$ , Leonidas Guibas $^{1}$ , and Gordon Wetzstein $^{1}$", + "bbox": [ + 215, + 210, + 785, + 258 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "1 Stanford University", + "bbox": [ + 426, + 268, + 575, + 284 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "2 Carnegie Mellon University", + "bbox": [ + 401, + 282, + 599, + 297 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "3 Google", + "bbox": [ + 470, + 297, + 532, + 311 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "4 Technical University of Munich", + "bbox": [ + 390, + 311, + 611, + 324 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Abstract. Modeling and rendering photorealistic avatars is of crucial importance in many applications. Existing methods that build a 3D avatar from visual observations, however, struggle to reconstruct clothed humans. We introduce PhysAvatar, a novel framework that combines inverse rendering with inverse physics to automatically estimate the shape and appearance of a human from multi-view video data along with the physical parameters of the fabric of their clothes. For this purpose, we adopt a mesh-aligned 4D Gaussian technique for spatio-temporal mesh tracking as well as a physically based inverse renderer to estimate the intrinsic material properties. PhysAvatar integrates a physics simulator to estimate the physical parameters of the garments using gradient-based optimization in a principled manner. These novel capabilities enable PhysAvatar to create high-quality novel-view renderings of avatars dressed in loose-fitting clothes under motions and lighting conditions not seen in the training data. This marks a significant advancement towards modeling photorealistic digital humans using physically based inverse rendering with physics in the loop. Our project website is at: https://qingqing-zhao.github.io/PhysAvatar.", + "bbox": [ + 259, + 362, + 743, + 612 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Keywords: neural rendering $\\cdot$ physics $\\cdot$ dynamic modeling $\\cdot$ 3D avatar", + "bbox": [ + 261, + 625, + 736, + 640 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "1 Introduction", + "text_level": 1, + "bbox": [ + 215, + 665, + 375, + 680 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Digital avatars are a vital component in numerous applications, ranging from virtual reality and gaming to telepresence and e-commerce [26,43,101]. A realistic 3D avatar of a person can be readily obtained from visual observations, such as multi-view image [116] and video [6] data. The task of rendering animated 3D avatars from novel viewpoints or when performing unseen motions, however, presents considerable challenges, particularly when the avatar wears loose-fitting garments. Accurately rendering the dynamics of garments in conditions that are not observed in the training data necessitates a holistic approach that not only", + "bbox": [ + 212, + 696, + 787, + 816 + ], + "page_idx": 0 + }, + { + "type": "page_footnote", + "text": "* Equal Contribution", + "bbox": [ + 217, + 824, + 366, + 839 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/9f5e9783a5bafabc9c297493ffa1d29405d2f8b9638b10115540c976aeb1567e.jpg", + "image_caption": [ + "Fig. 1: PhysAvatar is a novel framework that captures the physics of dressed 3D avatars from visual observations, enabling a wide spectrum of applications, such as (a) animation, (b) relighting, and (c) redressing, with high-fidelity rendering results." + ], + "image_footnote": [], + "bbox": [ + 246, + 147, + 524, + 366 + ], + "page_idx": 1 + }, + { + "type": "image", + "img_path": "images/ef0b754e528d2e8ffdf74a4562d29eda1bcd79f861f6206cf04ecf94a9e1a3c6.jpg", + "image_caption": [ + "(b)" + ], + "image_footnote": [], + "bbox": [ + 568, + 152, + 736, + 262 + ], + "page_idx": 1 + }, + { + "type": "image", + "img_path": "images/3c62b42598198dd839a86e1b1569e25ad87a6b2105516b83a9f9c04136db4314.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 563, + 271, + 750, + 364 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "models the shape and appearance of the person but also the physical behavior of their clothes, including friction and collision.", + "bbox": [ + 212, + 445, + 784, + 474 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Learning 3D scene representations from visual data is the core problem addressed by inverse rendering—a field that has recently shown remarkable progress in estimating the geometry and appearance of static and dynamic scenes from multi-view images and videos [105, 106]. In the context of reconstructing 3D avatars, the ability to explicitly control the motions of the avatar in post processing becomes imperative. Most existing methods for reconstructing animatable avatars drive the animation through an underlying skeleton using linear blending skinning (LBS) [64], which adequately models the dynamics of humans dressed in tight-fitting clothes [1,7,23,80,99,119]. With this approach, garments are treated as a rigidly attached part of the body adhering to piece-wise linear transformations, which results in motions that appear rigid and unconvincing in many cases [79, 120, 132]. Several successful attempts to introduce non-rigid deformations through pose-dependent geometry adjustments, such as wrinkles, have recently been made [18,57,59,60,114], although these methods are prone to overfitting to the body poses and motion sequences observed during training. A key problem is that most existing 3D avatar reconstruction approaches neglect to model the dynamics of loose garments in a physically accurate manner, leading to unrealistic fabric behavior and issues like self-penetration. To our knowledge, only the work of Xiang et al. [118] includes a physics-based approach to inverse rendering of digital humans, but their approach requires a tedious manual search process to find the parameters that model the dynamic behavior of their cloth fabrics with reasonable accuracy.", + "bbox": [ + 212, + 477, + 787, + 808 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Here, we introduce PhysAvatar, a novel approach to 3D avatar reconstruction from multi-view video data. PhysAvatar combines inverse rendering with", + "bbox": [ + 214, + 809, + 787, + 839 + ], + "page_idx": 1 + }, + { + "type": "page_number", + "text": "2", + "bbox": [ + 217, + 114, + 228, + 126 + ], + "page_idx": 1 + }, + { + "type": "header", + "text": "Yang Zheng et al.", + "bbox": [ + 271, + 114, + 392, + 128 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "\"inverse physics\" in a principled manner to not only estimate the shape and appearance of the avatar but also physical parameters, including density and stiffness, of the fabrics modeling their (loose) clothes. Our approach includes several parts: Given the reconstructed 3D mesh of the garment in one frame, we first leverage mesh-aligned 4D Gaussians [117] to track surface points of the garment across all frames of the input video sequence. This process establishes dense correspondences on the 3D garment. These data are used as supervision in an inverse physics step, where the physical parameters of the fabric are optimized using a finite-difference approach [50,123]. Finally, we employ a physically based inverse renderer [78] to jointly estimate ambient lighting and the surface material. By leveraging the refined geometry from our simulation step, the inverse renderer can effectively factor in pose-dependent effects, such as shadows caused by self-occlusion, resulting in accurate appearance reconstruction that enables the avatar to be rendered in novel lighting conditions.", + "bbox": [ + 212, + 146, + 787, + 356 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "PhysAvatar offers a comprehensive solution for reconstructing clothed human avatars from multi-view video data to perform novel view and motion synthesis with state-of-the-art realism. The key contributions of our work include:", + "bbox": [ + 212, + 357, + 787, + 402 + ], + "page_idx": 2 + }, + { + "type": "list", + "sub_type": "text", + "list_items": [ + "1. The introduction of a new inverse rendering paradigm for avatars created from real-world captures that incorporates the physics of loose garments in a principled manner;", + "2. A pipeline that includes accurate and efficient mesh reconstruction and tracking using 4D Gaussians; automatic optimization of the garments' physical material properties; and accurate appearance estimation using physically based inverse rendering." + ], + "bbox": [ + 220, + 402, + 784, + 506 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "PhysAvatar demonstrates a novel path to learning physical scene properties from visual observations by incorporating inverse rendering with physics constraints. Code will be available on the project website.", + "bbox": [ + 212, + 508, + 787, + 554 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "2 Related Work", + "text_level": 1, + "bbox": [ + 214, + 571, + 387, + 588 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "2.1 Scene Reconstruction from Visual Observations", + "text_level": 1, + "bbox": [ + 214, + 599, + 653, + 614 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Reconstructing 3D scenes from visual observations is grounded in classic approaches, such as Structure from Motion (SfM) [77, 93, 96]. Recently, this field has witnessed a paradigm shift towards inverse rendering techniques that estimate the shape and appearance of a scene from visual input [105, 106]. Among the many representations developed in this area are neural (radiance) fields [75, 95], neural volumes, surfaces, and signed distance fields [63, 73, 83, 94, 113, 122] as well as differentiable rasterization techniques for meshes [39, 61] and point clouds [40, 41, 48, 124]. Recent inverse rendering approaches are capable of modeling dynamic scenes [25, 84, 89, 108] and enable sophisticated post-capture deformations [36, 126].", + "bbox": [ + 212, + 619, + 787, + 772 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "2.2 Animatable Avatars", + "text_level": 1, + "bbox": [ + 214, + 787, + 426, + 803 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "One central requirement for generating full-body digital humans is to enable animations driven by skeleton-level control [42, 71]. To this end, early works on", + "bbox": [ + 212, + 809, + 785, + 840 + ], + "page_idx": 2 + }, + { + "type": "header", + "text": "PhysAvatar", + "bbox": [ + 651, + 114, + 730, + 128 + ], + "page_idx": 2 + }, + { + "type": "page_number", + "text": "3", + "bbox": [ + 774, + 116, + 784, + 126 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "avatar reconstruction primarily relied on skinning techniques, which successfully model the dynamics of bodies with minimal or tight clothing [3,64]. Clothing is an integral part of everyday human appearance, captured by the saying \"clothes make the (wo)man\". Reconstructing their dynamic behavior from visual data, however, remains a challenge. Several different types of representations have been explored for clothed avatars, including meshes [68] with dynamic textures [4, 33, 130], neural surface [16, 90, 107] and radiance fields [14, 18, 24, 27, 44-46, 86, 87, 99, 131], point sets [66, 67, 69], and 3D Gaussians [34, 58, 82, 132]. Many of these works condition the deformation of clothing on the body pose and some predict future dynamics based on a small number of frames [32, 33, 82].", + "bbox": [ + 212, + 146, + 787, + 297 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "We now discuss recent research most closely related to ours. TAVA [53] learns a deformable neural radiance field of humans or animals using a differentiable skinning method like SNARF [17]. ARAH [115] models 3D avatars using SDF representations and proposes an SDF-based volume rendering method to reconstruct the avatar in canonical space. GS-Avatar [34] achieves real-time animation by leveraging a learned 3D Gaussian predictor. Although these methods show reasonable effectiveness on humans wearing tight garments, they struggle with loose garments because their LBS-based deformation module cannot accommodate the complex dynamics of loose clothing. To address this limitation, Xiang et al. [118] incorporate a cloth simulator based on the eXtended Position Based Dynamics (XPBD) formulation [70] into a deep appearance model to reconstruct realistic dynamics and appearance. However, they require tedious manual adjustments of physical parameters to produce reasonable garment motion, and the code is not publicly available. Concurrent with our work, AniDress [14], while using physics-based simulation, only uses the simulation to produce a garment rigging model. This can cause inaccuracy in cloth dynamics modeling and unrealistic motion artifacts. In contrast, we estimate the physical parameters through a principled gradient-based inverse physics approach and accurately recover the shape and (relightable) appearance using a physically based inverse renderer, achieving high-fidelity results on novel motions.", + "bbox": [ + 212, + 297, + 787, + 599 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "2.3 Physics-Based Simulation", + "text_level": 1, + "bbox": [ + 214, + 619, + 472, + 635 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Physics-based simulation of cloth [5, 62, 88, 103, 109, 111, 125, 127] has been extensively studied and is particularly useful for modeling complicated dynamic effects in interaction with human bodies, such as large deformations, wrinkling, and contact. In this work, we adopt the Codimensional Incremental Potential Contact (C-IPC) solver [52] for its robustness in handling complicated body-cloth collision by the log-barrier penalty function [51]. We refer readers to dedicated surveys [37, 54] for a comprehensive review, and focus on works that solve the inverse parameter estimation problem. One line of work uses specialized devices to simultaneously measure forces and deformation in a controlled setting [47, 74, 112, 128]. Our work falls into the category that estimates physical parameters directly from imagery of garments worn on human subjects [12, 31, 97]. For this purpose, some differentiable simulators have been developed and applied to the optimization problem [55], but only for specific types of solvers,", + "bbox": [ + 212, + 643, + 787, + 840 + ], + "page_idx": 3 + }, + { + "type": "page_number", + "text": "4", + "bbox": [ + 217, + 114, + 228, + 126 + ], + "page_idx": 3 + }, + { + "type": "header", + "text": "Yang Zheng et al.", + "bbox": [ + 271, + 114, + 390, + 128 + ], + "page_idx": 3 + }, + { + "type": "image", + "img_path": "images/3136adbf142bd66ecf1e005f6a2eae2e84bb73f29c49d10e81dd1a0e28a895e6.jpg", + "image_caption": [ + "Fig. 2: Method Overview: (a) PhysAvatar takes multi-view videos and an initial mesh as input. We first perform (b) dynamic mesh tracking (Sec. 3.1). The tacked mesh sequences are then used for (c) garment physics estimation with a physics simulator combined with gradient-based optimization (Sec. 3.2); (d) and appearance estimation through physics-based differentiable rendering (Sec. 3.3). At test time, (e) given a sequence of body poses (f), we simulate garment dynamics with the learned physics parameters and employ physics-based rendering to produce the final images." + ], + "image_footnote": [], + "bbox": [ + 263, + 161, + 733, + 306 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "such as Projective Dynamics [56] and XPBD [98]. The heavy runtime cost of high-quality physics-based simulation motivates research in neural cloth simulation. Early works create datasets of garments simulated on human bodies [8,30], and then train the network to predict garment deformation in a supervised manner [19,81,91,110]. Recently, more attention has been paid to the self-supervised formulation that applies the elasticity energy defined in traditional simulation to predict garments, thereby training the network to act like a simulation solver, including both the (quasi-) static [9,11,22,49,92] and the dynamic [10,28] cases. Most of these works focus on making forward predictions that resemble traditional simulators without solving inverse parameter estimation on real-world garment data, except CaPhy [100]. CaPhy estimates the Saint Venant-Kirchhoff (StVK) elasticity parameters of clothing from 4D scans but only predicts pose-dependent clothing deformation in a quasi-static manner. In addition, all the works mentioned above treat simulation as a separate problem without modeling the photorealistic appearance. By comparison, our work builds avatars with physically based appearance and dynamics that are optimized for faithfulness to the real human capture in a holistic manner.", + "bbox": [ + 212, + 450, + 787, + 708 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "3 Method", + "text_level": 1, + "bbox": [ + 215, + 731, + 330, + 748 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Given multi-view videos as input, our goal is to reconstruct a 3D avatar that can be realistically animated and rendered under novel motions and novel views. The core pipeline (Fig. 2) consists of three primary components: 1. Mesh tracking (Sec. 3.1)—given multi-view videos and a reconstructed mesh at the initial time step as input, we track the deformation of the geometry, which provides the", + "bbox": [ + 212, + 763, + 787, + 840 + ], + "page_idx": 4 + }, + { + "type": "header", + "text": "PhysAvatar", + "bbox": [ + 651, + 114, + 730, + 127 + ], + "page_idx": 4 + }, + { + "type": "page_number", + "text": "5", + "bbox": [ + 774, + 116, + 784, + 126 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "geometry reference for the ensuing physics simulation; 2. Physics-based dynamic modeling (Sec. 3.2)—with the tracked meshes as reference, we estimate the garments' physical properties using a physics simulator; the simulator, together with the estimated physical parameters are used to generate novel motion; 3. Physics-based appearance refinement (Sec. 3.3)—given the tracked geometry, we further enhance the appearance estimation through a physics-based differentiable renderer, which considers self-occlusion and thus eliminate artifacts such as baked-in shadows. In the following, we delve into the details of each component.", + "bbox": [ + 212, + 146, + 787, + 268 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "3.1 Mesh Tracking", + "text_level": 1, + "bbox": [ + 214, + 291, + 387, + 306 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Given multiview videos of $T$ frames and $C$ views, $\\{\\mathbf{I}_{1:T,1:C}\\}$ , and an initial mesh $(\\mathbf{V}_1,\\mathbf{F})$ , with vertices $\\mathbf{V}_1$ and faces $\\mathbf{F}$ , that can be obtained via any existing static scene reconstruction methods, our goal is to accurately track the mesh deformation through the video sequence $\\mathbf{V}_{1:T}$ , which will be used as reference in the subsequent physics simulation. However accurately tracking mesh deformation from visual observation is very challenging. Inspired by the robust and real-time tracking performance from dynamic 3D Gaussians [65], we use 3D Gaussians as a surrogate representation for mesh tracking.", + "bbox": [ + 212, + 316, + 787, + 439 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "We adapt 3D Gaussians to align with the reconstructed mesh surface. Following the definition of dynamic 3D Gaussians, at frame $t$ , each Gaussian is parameterized with position $\\mathbf{p}_t \\in \\mathbb{R}^3$ , rotation quaternion $\\mathbf{q}_t \\in \\mathbb{R}^4$ , color $\\mathbf{c}_t \\in \\mathbb{R}^3$ , scale $\\mathbf{s}_t \\in \\mathbb{R}^3$ and opacity $o_t \\in \\mathbb{R}_+$ . To couple the Gaussians with the mesh, we attach one Gaussian at the barycenter of each face $\\mathbf{f} \\in \\mathbf{F}$ , and determine the Gaussians' quaternions from the attached triangle. Specifically, the rotation of each Gaussian is computed from the rotation of its local frame, whose $x-$ and $z$ -axis are defined using the longest triangle edge at $t = 1$ and the face normal respectively. Furthermore, we set the last scaling factor, which corresponds to the scaling in face normal direction, to be a small constant value, ensuring the Gaussian lies on the mesh surface. The Gaussians can then be rendered in a differentiable manner [40] and we optimize their parameters by minimizing the loss [65] between the rendered images and the reference images:", + "bbox": [ + 212, + 439, + 789, + 636 + ], + "page_idx": 5 + }, + { + "type": "equation", + "text": "\n$$\n\\mathcal {L} _ {\\mathrm {r e n d e r}} = \\lambda \\left\\| \\mathbf {I} _ {i, t} - \\hat {\\mathbf {I}} _ {i, t} \\right\\| _ {1} + (1 - \\lambda) \\cdot \\operatorname {S S I M} \\left(\\mathbf {I} _ {i, t}, \\hat {\\mathbf {I}} _ {i, t}\\right), \\tag {1}\n$$\n", + "text_format": "latex", + "bbox": [ + 316, + 648, + 785, + 676 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "where $\\hat{\\mathbf{I}}_{i,t}$ denotes the image rendered from the perspective of the $i$ -th camera at time $t$ . Note that since the Gaussians' position and rotation are bound to the mesh, the optimization of the mesh vertices will implicitly optimize the parameters of the Gaussians. At the first time step, we optimize color, scale, and opacity. In the following frames $1 < t \\leq T$ , we fix the opacity and scale, and optimize Gaussian colors $\\mathbf{c}_t$ and mesh vertices $\\mathbf{V}_t$ .", + "bbox": [ + 212, + 688, + 787, + 777 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "In addition to the photometric loss, we apply the following regularization terms to preserve the local geometric features: 1) an isometry loss $\\mathcal{L}_{\\mathrm{iso}}$ [65] that preserves edge length through the training sequence, 2) a normal loss $\\mathcal{L}_{\\mathrm{normal}}$ that encourages smooth mesh surfaces, enhancing mesh deformation accuracy", + "bbox": [ + 212, + 779, + 787, + 840 + ], + "page_idx": 5 + }, + { + "type": "page_number", + "text": "6", + "bbox": [ + 217, + 114, + 228, + 126 + ], + "page_idx": 5 + }, + { + "type": "header", + "text": "Yang Zheng et al.", + "bbox": [ + 271, + 114, + 392, + 128 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/840589d5a2624a2f068a0c847a95c5ddd61c20937f953f28d3236fad05acb589.jpg", + "image_caption": [ + "t=12" + ], + "image_footnote": [], + "bbox": [ + 259, + 145, + 362, + 238 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/15269b0c1622e5d39aac4763c334b086b1e9b370a6c84d42c11ad095ddf4250d.jpg", + "image_caption": [ + "Fig. 3: Our method can robustly track a dynamic mesh from input images, providing accurate long-term correspondences. Here we show the rendered images overlaid with the Gaussian trajectories from the previous 12 frames and the optimized meshes." + ], + "image_footnote": [], + "bbox": [ + 388, + 157, + 431, + 238 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/d91cf663f0276150571c76b0686644a18c11dbbfd0cc8dac17e204a228e4ad5f.jpg", + "image_caption": [ + "t=300" + ], + "image_footnote": [], + "bbox": [ + 444, + 162, + 478, + 238 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/2e36ae87752025754e9831984c0cf7856e52707cccf4346ca7cb4a0a4f335113.jpg", + "image_caption": [ + "t=600" + ], + "image_footnote": [], + "bbox": [ + 509, + 159, + 625, + 238 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/cbb42d65500b56dbd675a54f09a977bd57658d40c306806de4dc910d040b0555.jpg", + "image_caption": [ + "t=900" + ], + "image_footnote": [], + "bbox": [ + 643, + 157, + 741, + 238 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "and stability. Denoting the vertex and the triangle of the mesh at frame $t$ as $\\mathbf{v}_t$ and $\\mathbf{f}_t$ , the normal of a face as $\\mathbf{n}(\\cdot)$ , and the 1-ring neighbors $N(\\cdot)$ , the losses are formally defined as:", + "bbox": [ + 214, + 304, + 782, + 349 + ], + "page_idx": 6 + }, + { + "type": "equation", + "text": "\n$$\n\\mathcal {L} _ {\\mathrm {i s o}} = \\frac {1}{| \\mathbf {V} |} \\sum_ {\\mathbf {v} \\in \\mathbf {V}} \\sum_ {\\mathbf {v} ^ {\\prime} \\in N (\\mathbf {v} _ {1})} \\frac {\\left(\\left\\| \\mathbf {v} _ {1} - \\mathbf {v} _ {1} ^ {\\prime} \\right\\| _ {2} - \\left\\| \\mathbf {v} _ {t} - \\mathbf {v} _ {t} ^ {\\prime} \\right\\| _ {2}\\right)}{| N (\\mathbf {v}) |}, \\tag {2}\n$$\n", + "text_format": "latex", + "bbox": [ + 326, + 359, + 784, + 398 + ], + "page_idx": 6 + }, + { + "type": "equation", + "text": "\n$$\n\\mathcal {L} _ {\\text {n o r m a l}} = \\frac {1}{| \\mathbf {F} |} \\sum_ {\\mathbf {f} _ {t} \\in \\mathbf {F}} \\sum_ {\\mathbf {f} _ {t} ^ {\\prime} \\in N (\\mathbf {f} _ {t})} \\left(1 - \\mathbf {n} \\left(\\mathbf {f} _ {t}\\right) \\cdot \\mathbf {n} \\left(\\mathbf {f} _ {t} ^ {\\prime}\\right)\\right). \\tag {3}\n$$\n", + "text_format": "latex", + "bbox": [ + 305, + 402, + 784, + 440 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "The full loss used for mesh tracking is defined as:", + "bbox": [ + 240, + 450, + 594, + 465 + ], + "page_idx": 6 + }, + { + "type": "equation", + "text": "\n$$\n\\mathcal {L} _ {\\text {m e s h}} = \\mathcal {L} _ {\\text {r e n d e r}} + \\lambda_ {i} \\mathcal {L} _ {\\text {i s o}} + \\lambda_ {n} \\mathcal {L} _ {\\text {n o r m a l}}, \\tag {4}\n$$\n", + "text_format": "latex", + "bbox": [ + 364, + 479, + 784, + 494 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "where we set $\\lambda_{i} = 10$ and $\\lambda_{n} = 0.1$ during our training. With the introduced losses, our method can reconstruct a smooth mesh sequence with accurate correspondences, as shown in Fig. 3.", + "bbox": [ + 214, + 505, + 782, + 551 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "3.2 Physics Based Dynamic Modeling", + "text_level": 1, + "bbox": [ + 214, + 573, + 540, + 589 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Given the tracked meshes, $(\\mathbf{V}_{1:T},\\mathbf{F})$ , we estimate the physical properties of the garments such that the simulation of their dynamics matches the reference tracked meshes. Once learned, these parameters can be used to simulate the garment under novel motions and novel views.", + "bbox": [ + 212, + 598, + 784, + 657 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Garment Material Model and Simulation. We model the elastodynamics of the garment by equipping each simulation mesh with the discrete shell hinge bending energy [29,102] and isotropic StVK elastic potential [15,20]. Specifically, we estimate density $\\rho$ , membrane stiffness $\\kappa_{s}$ , and bending stiffness $\\kappa_{b}$ of the garment given the visual observations. For our generic use cases, it is reasonable to assume homogeneous physical properties for each garment piece, i.e., the density, membrane stiffness, and bending stiffness can be parameterized using scalar values, denoted as $\\rho$ , $\\kappa_{s}$ , and $\\kappa_{b}$ . In choosing the simulators, we consider differentiability and quality. Unfortunately, the development of differentiable simulators suitable for our application remains an open research question. Current implementations, such as those documented in [56], cannot be directly applied due to limitations in modeling complex body colliders. Therefore, we propose to use", + "bbox": [ + 214, + 657, + 785, + 840 + ], + "page_idx": 6 + }, + { + "type": "header", + "text": "PhysAvatar", + "bbox": [ + 651, + 114, + 730, + 127 + ], + "page_idx": 6 + }, + { + "type": "page_number", + "text": "7", + "bbox": [ + 774, + 114, + 784, + 126 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "the state-of-the-art robust and stable cloth simulator C-IPC [52] that guarantees no interpenetration with frictional contact. While C-IPC is differentiable, it lacks an implementation of accurate analytical gradients; hence, we use finite difference methods to estimate the gradients for parameter updates. Specifically, we employ an open-source implementation of C-IPC [52] as our cloth simulator, which is a state-of-the-art simulator achieving high-fidelity garment simulation with stability and convergence (see [52] for a comprehensive review of the base algorithm and implementation).", + "bbox": [ + 212, + 146, + 787, + 267 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Garment Material Estimation. From the reconstructed mesh, we segment out the garment vertices to obtain ground truth garment mesh sequences $(\\widehat{\\mathbf{V}}_{1:T}^{g},\\widehat{\\mathbf{F}}^{g})$ . The cloth simulator, $f(\\cdot)$ , takes as input the current garment state—position and velocity $(\\mathbf{V}_t^g,\\dot{\\mathbf{V}}_t^g)$ —the garment's physical properties $(\\rho ,\\kappa_s,\\kappa_b)$ , the body collider $\\{\\mathbf{V}_t^C,\\mathbf{F}^C\\}$ , and garment boundary vertices $\\mathbf{V}_{t + 1}^b$ , and step size $\\Delta t$ . The boundary vertices of the garment are points of the garment that are in permanent contact with the human body, such as the shoulder region for draped garments. We employ the SMPL-X body model [64] as the body collider $\\{\\mathbf{V}_{t + 1}^{C},\\mathbf{F}^{C}\\}$ , extracting SMPL-X body shapes and poses from the video data during the preprocessing (see Sec. 4). The garment and the boundary vertices are manually annotated once on the initial mesh $\\mathbf{V}_1$ for the entire simulation. The simulation function predicting the garment shape in the next frame can be written as below:", + "bbox": [ + 212, + 268, + 805, + 450 + ], + "page_idx": 7 + }, + { + "type": "equation", + "text": "\n$$\n\\mathbf {V} _ {t + 1} ^ {g} = f (\\mathbf {V} _ {t} ^ {g}, \\dot {\\mathbf {V}} _ {t} ^ {g}, \\mathbf {V} _ {t + 1} ^ {b}, \\mathbf {V} _ {t + 1} ^ {C}, \\rho , \\kappa_ {s}, \\kappa_ {b}, \\Delta t). \\tag {5}\n$$\n", + "text_format": "latex", + "bbox": [ + 346, + 464, + 785, + 483 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "We find the optimal physical parameters by minimizing the difference between the simulated and the reconstructed garment meshes. Formally, the optimization objective can be expressed as:", + "bbox": [ + 214, + 488, + 787, + 535 + ], + "page_idx": 7 + }, + { + "type": "equation", + "text": "\n$$\n\\min _ {\\left\\{\\rho , \\kappa_ {s}, \\kappa_ {b} \\right\\}} \\mathcal {L} _ {\\text {s i m}} (\\rho , \\kappa_ {s}, \\kappa_ {b}) = \\sum_ {t = 0} ^ {T} \\| \\mathbf {V} _ {t + 1} ^ {g} - \\widehat {\\mathbf {V}} _ {t + 1} ^ {g} \\| _ {2} ^ {2}, \\tag {6}\n$$\n", + "text_format": "latex", + "bbox": [ + 331, + 547, + 785, + 587 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "where $\\mathbf{V}_1^g = \\widehat{\\mathbf{V}}_1^g$ and $\\mathbf{V}_{t + 1}^g$ is computed by Eq. (5). We estimate gradients via a numerical approach using finite differences [50]. This method is particularly feasible in our context due to the low-dimensional nature of the optimization problem, which involves only three scalar parameters. We employ the forward difference formula for gradient estimation:", + "bbox": [ + 214, + 601, + 787, + 678 + ], + "page_idx": 7 + }, + { + "type": "equation", + "text": "\n$$\n\\frac {\\partial \\mathcal {L} _ {\\mathrm {s i m}}}{\\partial \\rho} \\approx (\\mathcal {L} _ {\\mathrm {s i m}} (\\rho + \\delta_ {\\rho}, \\kappa_ {s}, \\kappa_ {b}) - \\mathcal {L} _ {\\mathrm {s i m}} (\\rho , \\kappa_ {s}, \\kappa_ {b})) / \\delta_ {\\rho}. \\tag {7}\n$$\n", + "text_format": "latex", + "bbox": [ + 320, + 689, + 785, + 722 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Where $\\delta_{\\rho}$ is a hyperparameter for forward difference gradient estimation. The gradients for other parameters are estimated using the same approach, and the parameters are subsequently updated using the Adam optimizer.", + "bbox": [ + 212, + 731, + 785, + 777 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Animation. At test time, we have the initial mesh $\\{\\mathbf{V}_1,\\mathbf{F}_1\\}$ and we are given a sequence of novel human body poses, e.g., provided by skeleton parameters such as SMPL [64]. Using standard LBS, we obtain the boundary locations $\\mathbf{V}_{t + 1}^b$ . We assume the initial garment is at rest, i.e., $\\dot{\\mathbf{V}}_0 = 0$ . We use the same garment", + "bbox": [ + 212, + 777, + 787, + 840 + ], + "page_idx": 7 + }, + { + "type": "page_number", + "text": "8", + "bbox": [ + 217, + 114, + 228, + 126 + ], + "page_idx": 7 + }, + { + "type": "header", + "text": "Yang Zheng et al.", + "bbox": [ + 271, + 114, + 392, + 128 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/99002568d07cf4242785855ef892d1629237b007a2a03966ffc5e73a81dee103.jpg", + "image_caption": [ + "(a)" + ], + "image_footnote": [], + "bbox": [ + 256, + 181, + 379, + 276 + ], + "page_idx": 8 + }, + { + "type": "image", + "img_path": "images/f351861199c87b93e8d1dd80680c976e5c055bce9882210fc89cc0cc9bb7f3fa.jpg", + "image_caption": [ + "Before Optimization", + "After Optimization", + "(b)" + ], + "image_footnote": [], + "bbox": [ + 383, + 181, + 506, + 276 + ], + "page_idx": 8 + }, + { + "type": "image", + "img_path": "images/98753764e95fd30f8733f407016da0016f595eac4b8da8dc62737c5588e1bb27.jpg", + "image_caption": [ + "PBR w/o Opt", + "Fig. 4: Ablation study on appearance estimation: (a) Initial texture map $\\mathbf{T}$ extracted from Gaussian splatting (left) has baked-in shadows highlighted in the red boxes; post-optimization (right), the baked-in shadows are substantially removed. (b) Rendering comparisons demonstrate that our method with the optimized texture map more closely aligns with the ground truth." + ], + "image_footnote": [], + "bbox": [ + 531, + 162, + 586, + 279 + ], + "page_idx": 8 + }, + { + "type": "image", + "img_path": "images/4bd4b31a34cc91e0dfa60e2c1f11cb7b0be21791b202ad017a4675c279128e38.jpg", + "image_caption": [ + "PBR Ours" + ], + "image_footnote": [], + "bbox": [ + 606, + 162, + 663, + 279 + ], + "page_idx": 8 + }, + { + "type": "image", + "img_path": "images/c5b67a8b8e3f2f0dd26d90e77a0c9d7ccde965216f81b6c363acdb4b4bbd9b7f.jpg", + "image_caption": [ + "GT" + ], + "image_footnote": [], + "bbox": [ + 681, + 162, + 740, + 279 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "segmentation, boundary points, and SMPL body collider points as during training, and use Eq. (5) to simulate the deformation of the garment. This is then combined with the body, driven by LBS, to create the holistic animation.", + "bbox": [ + 212, + 422, + 782, + 468 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "3.3 Physics Based Appearance Modeling", + "text_level": 1, + "bbox": [ + 214, + 489, + 565, + 505 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "While tracking with dynamic Gaussians jointly estimates the garments appearance, it does not account for light-surface interactions, consequently the reconstructed appearance contains baked-in shadows, causing uncanny visual artifacts when rendered in novel views and motions (see Fig. 4). In this step, given the mesh sequence $(\\mathbf{V}_{1:T},\\mathbf{F})$ and the per-frame mesh colors $\\mathbf{c}_{1:T}$ reconstructed from Section 3.1, we utilize Mitsuba3 [78]—a physics-based differentiable Monte Carlo renderer—to enhance the appearance reconstruction. For this purpose, we use a time-invariant diffuse texture map $\\mathbf{T}$ to model the appearance. The value $\\mathbf{T}$ is initialized from the average Gaussian colors from the tracking step (Section 3.1), $\\Sigma_{1 < t\\leq T}(\\mathbf{c}(\\mathbf{x})_t) / T$ . For the environment lighting, since the capture studio is typically equipped with uniform point light arrays, we can approximate the environment lighting as an ambient illumination, parameterized with a global $L_{a}\\in \\mathbb{R}^{3}$ .", + "bbox": [ + 212, + 513, + 787, + 709 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "We optimize for $\\mathbf{T}$ and $L_{a}$ using the photometric loss defined in Eq. (8) for all training views and frames $(i,t)$ , where $\\mathcal{R}$ denotes the rendering function and $c_{i,t}$ the camera pose:", + "bbox": [ + 212, + 710, + 785, + 756 + ], + "page_idx": 8 + }, + { + "type": "equation", + "text": "\n$$\n\\mathcal {L} _ {\\mathrm {r g b}} = \\min _ {\\left\\{\\mathbf {T}, L _ {a} \\right\\}} \\sum_ {i, t} \\left\\| \\mathbf {I} _ {i, t} - \\mathcal {R} \\left(\\mathbf {T}, L _ {a}, \\mathbf {V} _ {t}, c _ {i, t}\\right) \\right\\| _ {2} ^ {2}. \\tag {8}\n$$\n", + "text_format": "latex", + "bbox": [ + 334, + 767, + 785, + 800 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "At test time, we can swap the ambient light with more complex illuminations, and produce realistic rendering in arbitrary views and body poses (see Sec. 4.4).", + "bbox": [ + 214, + 809, + 782, + 840 + ], + "page_idx": 8 + }, + { + "type": "header", + "text": "PhysAvatar", + "bbox": [ + 651, + 114, + 730, + 127 + ], + "page_idx": 8 + }, + { + "type": "page_number", + "text": "9", + "bbox": [ + 774, + 116, + 785, + 126 + ], + "page_idx": 8 + }, + { + "type": "image", + "img_path": "images/5570d9eb58ea3748a24e2c2903800d6034251f17e7e56c6cf866db386ed8bf76.jpg", + "image_caption": [ + "Fig. 5: Qualitative results on test poses from the ActorHQ [38] dataset. Our method PhysAvatar achieves state-of-the-art performance in terms of geometry detail and appearance modeling." + ], + "image_footnote": [], + "bbox": [ + 246, + 143, + 759, + 388 + ], + "page_idx": 9 + }, + { + "type": "text", + "text": "4 Experiments", + "text_level": 1, + "bbox": [ + 215, + 459, + 375, + 477 + ], + "page_idx": 9 + }, + { + "type": "text", + "text": "In this section, we explain our experimental setup and results. We recommend watching the supplementary video for better visualization.", + "bbox": [ + 214, + 489, + 785, + 521 + ], + "page_idx": 9 + }, + { + "type": "text", + "text": "4.1 Experimental setup", + "text_level": 1, + "bbox": [ + 215, + 542, + 423, + 559 + ], + "page_idx": 9 + }, + { + "type": "text", + "text": "Data Preparation. We choose four characters from the ActorHQ [35] dataset for our experiments, including two female characters with loose dresses, one female character wearing a tank top and loose-fitting pants, and one male character with a short-sleeved shirt and khaki shorts, as shown in Fig. 1. For each character, we use the ground truth mesh provided by the dataset and re-mesh it to obtain a simulation-ready initial mesh, i.e., one where the triangles have similar areas, to enhance the performance of garment simulation. We then segment the garment, define the boundary vertices of the garment for simulation purposes, and unwrap the whole mesh to get the UV map, all performed in Blender [13]. Since we observe that the human pose estimations from the dataset are quite noisy, we fit the SMPL-X [85] body mesh using multi-view videos by minimizing the distance between the projected SMPL-X joints and 2D keypoint detection results from DWPose [21, 121].", + "bbox": [ + 212, + 566, + 787, + 763 + ], + "page_idx": 9 + }, + { + "type": "text", + "text": "Implementation Details. For training, we utilize the entire set of 160 camera views from the ActorHQ [35] dataset. Although the dataset includes approximately 2000 frames for each character, we selectively use 24 frames with large movements for estimating garment material parameters and 200 frames for optimizing the texture map, which leads to good results while conserving compu", + "bbox": [ + 212, + 765, + 787, + 840 + ], + "page_idx": 9 + }, + { + "type": "page_number", + "text": "10", + "bbox": [ + 217, + 114, + 235, + 126 + ], + "page_idx": 9 + }, + { + "type": "header", + "text": "Yang Zheng et al.", + "bbox": [ + 271, + 114, + 390, + 128 + ], + "page_idx": 9 + }, + { + "type": "table", + "img_path": "images/8571eabe21ef553f54ba1f10d6646325f19039ff79b04b135850430f385cfa36.jpg", + "table_caption": [ + "Table 1: Quantitative comparison with baselines. PhysAvatar outperforms all baselines with respect to geometry accuracy. We achieve the best or competitive results in appearance metrics. See Fig. 5 and Fig. 6 for qualitative comparisons." + ], + "table_footnote": [], + "table_body": "
MethodGeometryAppearance
CD (×103) (↓)F-Score (↑)LPIPS (↓)PSNR (↑)SSIM (↑)
RAAH [115]1.1286.10.05528.60.957
TAVA [53]0.6692.30.05129.60.962
GS-Avatar [34]0.9189.40.04430.60.962
PhysAvatar (ours)0.5592.90.03530.20.957
", + "bbox": [ + 240, + 198, + 759, + 299 + ], + "page_idx": 10 + }, + { + "type": "text", + "text": "tational resources. For evaluation, we choose 200 frames for each character, with novel motions unseen during training. At test time, given a motion sequence defined by SMPL-X [85], we use an LBS weight inpainting algorithm [2] to obtain the skinning weights of the human body with respect to the SMPL-X skeleton. We animate the human body using LBS and simulate the garment, which is driven by the boundary vertices and utilizes the SMPL-X body mesh as the collider. More details are available at supplementary document.", + "bbox": [ + 212, + 329, + 787, + 434 + ], + "page_idx": 10 + }, + { + "type": "text", + "text": "Baselines. We benchmark our method and current open-sourced state-of-the-art methods modeling full-body avatars, including ARAH [115] which leverages SDF-based volume rendering to learn the canonical representation of the avatar, TAVA [53] which learns the canonical radiance field [75] and skinning weight of the character, and GS-Avatar [34] based on a pose-conditioned neural network which predicts the parameters of 3D Gaussians [40]. For all baselines, we use the official public codebase and extend them to our dense-view setting.", + "bbox": [ + 212, + 435, + 789, + 542 + ], + "page_idx": 10 + }, + { + "type": "text", + "text": "4.2 Comparison", + "text_level": 1, + "bbox": [ + 215, + 568, + 362, + 583 + ], + "page_idx": 10 + }, + { + "type": "text", + "text": "We compare PhysAvatar against current state-of-the-art approaches in terms of geometry and appearance quality across test pose sequences, as shown in Table 1, Fig. 5, and Fig. 6.", + "bbox": [ + 212, + 597, + 787, + 641 + ], + "page_idx": 10 + }, + { + "type": "text", + "text": "Geometry Evaluation. For geometry evaluation, we employ the mean Chamfer distance [73] and the mean F-Score [104] at $\\tau = 0.001$ as quantitative metrics. PhysAvatar outperforms all baseline according to these metrics, demonstrating a superior capability in accurately capturing the dynamic nature of garment geometry over time, as shown in Table 1. As shown in Fig. 5, PhysAvatar generates more accurate garment deformation and better captures fine wrinkle details, thanks to the integration of a physics-based simulator. This contrasts with all baseline methods [34,53,115], which do not explicitly model garment dynamics, but instead depend on learning quasi-static, pose-conditioned skinning weights or deformations. Additionally, we show results using motion data from the AMASS dataset [72]. PhysAvatar demonstrates its robustness by consistently synthesizing realistic garment deformation with fine details, while baseline methods sometimes fail to capture large deformations in loose garments (Fig. 6).", + "bbox": [ + 212, + 643, + 787, + 840 + ], + "page_idx": 10 + }, + { + "type": "header", + "text": "PhysAvatar", + "bbox": [ + 651, + 114, + 730, + 128 + ], + "page_idx": 10 + }, + { + "type": "page_number", + "text": "11", + "bbox": [ + 767, + 114, + 784, + 126 + ], + "page_idx": 10 + }, + { + "type": "image", + "img_path": "images/a4694460856dc5457d22f15142cdd2c411dfe3b9168277d8280779a961b74105.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 259, + 145, + 493, + 252 + ], + "page_idx": 11 + }, + { + "type": "image", + "img_path": "images/cf969c606f5cdd7c2d7dcc961e18109834fd4fb66c3d72f6ef0a5ee6d4836203.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 504, + 143, + 743, + 252 + ], + "page_idx": 11 + }, + { + "type": "image", + "img_path": "images/8c1ef83398f094ea9ad4b3222dc39327fce742facd99be65bbf934a3f0cacb69.jpg", + "image_caption": [ + "Fig. 6: Here we show animation results of current state-of-the-art methods including ARAH [115], TAVA [53] and GS-Avatar [34], and our results on test motions from AMASS [72] dataset. Images are rendered from novel views. Please zoom in for details." + ], + "image_footnote": [], + "bbox": [ + 259, + 255, + 493, + 359 + ], + "page_idx": 11 + }, + { + "type": "image", + "img_path": "images/099a6fd63455150aa4c35ad53d2755dad40bebd9895d1b7cd19c880bd9a1788f.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 504, + 255, + 743, + 359 + ], + "page_idx": 11 + }, + { + "type": "text", + "text": "Appearance Evaluation. For appearance evaluation, we utilize Peak Signal-to-Noise Ratio (PSNR), Structural Similarity Index Measure (SSIM), and the Learned Perceptual Image Patch Similarity (LPIPS) [129]. As detailed in Table 1, PhysAvatar surpasses all baselines in LPIPS, a metric shown to be better aligned with human perception [129]. While we achieve the second-best results in PSNR and SSIM, it is important to note that even minor misalignments in images can significantly impact these metrics, potentially not fully reflecting the quality of the renderings. As demonstrated in Fig. 5, PhysAvatar can capture more high-frequency details in facial features and garment textures, thereby creating images that are visually richer and more detailed, compared to the baseline approaches.", + "bbox": [ + 212, + 434, + 787, + 587 + ], + "page_idx": 11 + }, + { + "type": "text", + "text": "4.3 Ablation", + "text_level": 1, + "bbox": [ + 215, + 606, + 334, + 619 + ], + "page_idx": 11 + }, + { + "type": "text", + "text": "We perform ablation studies to assess the impact of each component in our pipeline on the overall performance. For this, we focus on two female characters from the ActorHQ dataset, specifically chosen for their loose garments.", + "bbox": [ + 212, + 628, + 785, + 672 + ], + "page_idx": 11 + }, + { + "type": "text", + "text": "Garment physics estimation and modeling. We ablate our approach against two variants: one employing Linear Blend Skinning (LBS) for garment deformation and another utilizing a physics simulator for garment dynamics without optimized parameters, henceforth referred to as 'w/o physics', which employs random garment parameters. As shown in Table 2, both alternatives exhibit worse performance across all geometric metrics. Further, as shown in Fig. 7, LBS fails to produce plausible deformations for loose garments, such as dresses. While simulation with random parameters yields plausible garment dynamics, they do not accurately reflect real-world observations. In contrast, our method, with learned physics parameters, achieves deformations that closely align with the reference, demonstrating its effectiveness.", + "bbox": [ + 212, + 674, + 787, + 840 + ], + "page_idx": 11 + }, + { + "type": "page_number", + "text": "12", + "bbox": [ + 217, + 114, + 235, + 126 + ], + "page_idx": 11 + }, + { + "type": "header", + "text": "Yang Zheng et al.", + "bbox": [ + 271, + 114, + 390, + 128 + ], + "page_idx": 11 + }, + { + "type": "image", + "img_path": "images/17d588cdd4eb58a72fd1aec8fa865a1c03bca182f33fd92bedea656cc429ec63.jpg", + "image_caption": [ + "LBS" + ], + "image_footnote": [], + "bbox": [ + 230, + 143, + 354, + 242 + ], + "page_idx": 12 + }, + { + "type": "image", + "img_path": "images/1a8a93f90bd59e3c76bb21043bb974199037f1a795c3fd3c63e20a7251ff52f3.jpg", + "image_caption": [ + "w/o Physics" + ], + "image_footnote": [], + "bbox": [ + 367, + 143, + 493, + 242 + ], + "page_idx": 12 + }, + { + "type": "image", + "img_path": "images/a1da5aff1a331a8c48de8c66c5dc38e71c1b6743fc0c1988bb9092a576563771.jpg", + "image_caption": [ + "Ours", + "Fig. 7: Ablation Study on garment physics estimation: LBS-based garment deformation (LBS) fails to generate realistic garment deformation. Physics-based simulation with random parameters (w/o physics) achieves realistic deformation but falls short of matching the ground-truth reference (GT). In contrast, PhysAvatar accurately captures garment dynamics, closely aligning with the reference." + ], + "image_footnote": [], + "bbox": [ + 504, + 143, + 630, + 242 + ], + "page_idx": 12 + }, + { + "type": "image", + "img_path": "images/d3d7c6faf0c6338aa8cf66be7feeb13a0d6a9bf620baf86af22f0e39802858b6.jpg", + "image_caption": [ + "GT" + ], + "image_footnote": [], + "bbox": [ + 645, + 143, + 772, + 242 + ], + "page_idx": 12 + }, + { + "type": "table", + "img_path": "images/04dbd5f39a9da2d3462d2461500925ae2277ce02f975f5be7852ccff89cf2da4.jpg", + "table_caption": [], + "table_footnote": [], + "table_body": "
MethodGeometryAppearance
CD (×10-3) (↓)F-Score (↑)LPIPS (↓)PSNR (↑)SSIM (↑)
LBS0.7191.60.035430.270.951
w/o physics0.6192.20.034730.260.951
w/o appearance——0.040228.260.947
PhysAvatar (ours)0.5693.00.034330.290.951
", + "bbox": [ + 232, + 349, + 769, + 448 + ], + "page_idx": 12 + }, + { + "type": "text", + "text": "Table 2: Ablation Study: PhysAvatar demonstrates improved geometric accuracy compared to both LBS-based animation (LBS) and simulations with random garment parameters (w/o physics), also shown in Fig. 7. PhysAvatar exhibits superior appearance metrics compared to the setting without physics-based appearance estimation (w/o appearance) (see Fig. 4).", + "bbox": [ + 212, + 450, + 784, + 518 + ], + "page_idx": 12 + }, + { + "type": "text", + "text": "Appearance estimation and modeling. We compare rendering outcomes with those obtained without appearance optimization through physics-based inverse rendering [78] (denoted as 'w/o appearance'), where the texture is directly extracted from the Gaussian framework, as detailed in section 3.3. Fig. 4 illustrates that, since the Gaussian framework does not model complex light-surface interactions, shadows are inherently baked into the texture. However, post-optimization, we note a significant removal of these baked shadows. Quantitatively, we see improvements across all appearance metrics post-optimization, shown in Table 2.", + "bbox": [ + 212, + 561, + 787, + 698 + ], + "page_idx": 12 + }, + { + "type": "text", + "text": "4.4 Application", + "text_level": 1, + "bbox": [ + 214, + 722, + 359, + 738 + ], + "page_idx": 12 + }, + { + "type": "text", + "text": "PhysAvatar can facilitate a wide range of applications, such as animation, relighting, and redressing, as demonstrated in Fig. 1. Our method can also benefit from other motion databases in addition to motions defined under the regime of SMPL-X [85] or SMPL [64]. Fig. 1 shows our animation results on challenging motions from Mixamo [76], where we use its auto-rigging algorithm to drive the human body, boundary vertices of the garment, and SMPL-X mesh as the", + "bbox": [ + 212, + 750, + 787, + 840 + ], + "page_idx": 12 + }, + { + "type": "header", + "text": "PhysAvatar", + "bbox": [ + 651, + 114, + 730, + 127 + ], + "page_idx": 12 + }, + { + "type": "page_number", + "text": "13", + "bbox": [ + 767, + 114, + 785, + 126 + ], + "page_idx": 12 + }, + { + "type": "image", + "img_path": "images/7c0335449859589b52503fcf916029c5da5bf9fd2955c2c918c0adf628948c9d.jpg", + "image_caption": [ + "(a) Our results are readily exportable" + ], + "image_footnote": [], + "bbox": [ + 218, + 145, + 357, + 253 + ], + "page_idx": 13 + }, + { + "type": "image", + "img_path": "images/96c922d5af455dc27be21de45f9a56aa895487264f32c38d1ed806da8a3cb514.jpg", + "image_caption": [ + "Fig. 8: Our results can be exported into a traditional computer graphics pipeline, e.g., Blender [13], which enables the avatars performing novel motions to be rendered in unseen environments (a) as well as editing lighting and texture of the garments (b)." + ], + "image_footnote": [], + "bbox": [ + 357, + 146, + 496, + 253 + ], + "page_idx": 13 + }, + { + "type": "image", + "img_path": "images/73ca939b9006627ea05c7464a574a7b69037edf79a3b46a9cbcb42c6465d2379.jpg", + "image_caption": [ + "(b) Results with artistic modifications" + ], + "image_footnote": [], + "bbox": [ + 504, + 145, + 643, + 253 + ], + "page_idx": 13 + }, + { + "type": "image", + "img_path": "images/9d0c9034ce274eda34f5e032698a289f7c8d06e29141aa4506e4210d741a4df4.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 643, + 146, + 782, + 253 + ], + "page_idx": 13 + }, + { + "type": "text", + "text": "collider. Current state-of-the-art approaches cannot handle such settings since they are mainly based on LBS weights from SMPL or SMPL-X. Moreover, while current methods relying on NeRF [75] or 3D Gaussians [40] representation are usually incompatible with legacy graphics pipelines, our results are readily exportable and can be seamlessly integrated with computer graphics software tools (e.g., Blender [13]), enabling post-processing artistic modifications (Fig. 8).", + "bbox": [ + 214, + 345, + 787, + 436 + ], + "page_idx": 13 + }, + { + "type": "text", + "text": "5 Limitations and Future Work", + "text_level": 1, + "bbox": [ + 215, + 457, + 537, + 474 + ], + "page_idx": 13 + }, + { + "type": "text", + "text": "Our method currently depends on manual garment segmentation and mesh UV unwrapping. We are posed to evolve our pipeline towards a more automated system by incorporating neural networks to streamline these data preprocessing tasks. Another limitation of our method is the absence of a perfect underlining human body for garment simulation. Our current use of the SMPL-X [85] model as the body collider introduces potential inaccuracies in collision detection and contact information due to mismatches between the SMPL-X mesh and the actual human body shape. Our method would benefit from using more advanced parametric models that promise a closer approximation to real human body shape. Furthermore, our pipeline leverages dense view captures to model avatars, an approach that, while effective, has its limitations. Adapting our method to a sparse-view setting would enhance the versatility of its applications in environments where capturing dense views is impractical.", + "bbox": [ + 214, + 486, + 787, + 684 + ], + "page_idx": 13 + }, + { + "type": "text", + "text": "6 Conclusion", + "text_level": 1, + "bbox": [ + 215, + 704, + 359, + 720 + ], + "page_idx": 13 + }, + { + "type": "text", + "text": "We introduce PhysAvatar, a comprehensive pipeline to model clothed 3D human avatars with state-of-the-art realism. By integrating a robust and efficient mesh tracking method, and a novel inverse rendering paradigm to capture the physics of loose garments and the appearance of the body, we believe our work represents a significant advancement in digital avatar modeling from visual observations, which also paves the way for innovative applications in entertainment, virtual reality, and digital fashion.", + "bbox": [ + 214, + 733, + 787, + 840 + ], + "page_idx": 13 + }, + { + "type": "page_number", + "text": "14", + "bbox": [ + 217, + 114, + 235, + 126 + ], + "page_idx": 13 + }, + { + "type": "header", + "text": "Yang Zheng et al.", + "bbox": [ + 271, + 114, + 390, + 128 + ], + "page_idx": 13 + }, + { + "type": "text", + "text": "Acknowledgement", + "text_level": 1, + "bbox": [ + 217, + 143, + 392, + 162 + ], + "page_idx": 14 + }, + { + "type": "text", + "text": "We would like to thank Jiayi Eris Zhang for the discussions. This material is based on work that is partially funded by an unrestricted gift from Google, Samsung, an SNF Postdoc Mobility fellowship, ARL grant W911NF-21-2-0104, and a Vannevar Bush Faculty Fellowship.", + "bbox": [ + 212, + 175, + 784, + 231 + ], + "page_idx": 14 + }, + { + "type": "text", + "text": "References", + "text_level": 1, + "bbox": [ + 217, + 253, + 321, + 270 + ], + "page_idx": 14 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "1. Abdal, R., Yifan, W., Shi, Z., Xu, Y., Po, R., Kuang, Z., Chen, Q., Yeung, D.Y., Wetzstein, G.: Gaussian shell maps for efficient 3d human generation. arXiv preprint arXiv:2311.17857 (2023)", + "2. Abdrashitov, R., Raichstat, K., Monsen, J., Hill, D.: Robust skin weights transfer via weight inpainting. In: SIGGRAPH Asia 2023 Technical Communications, pp. 1-4 (2023)", + "3. Anguelov, D., Srinivasan, P., Koller, D., Thrun, S., Rodgers, J., Davis, J.: Scape: shape completion and animation of people. In: ACM SIGGRAPH 2005 Papers, pp. 408-416 (2005)", + "4. Bagautdinov, T., Wu, C., Simon, T., Prada, F., Shiratori, T., Wei, S.E., Xu, W., Sheikh, Y., Saragih, J.: Driving-signal aware full-body avatars. ACM Transactions on Graphics (TOG) 40(4), 1-17 (2021)", + "5. Baraff, D., Witkin, A.: Large steps in cloth simulation. In: Seminal Graphics Papers: Pushing the Boundaries, Volume 2, pp. 767-778 (2023)", + "6. Bashirov, R., Larionov, A., Ustinova, E., Sidorenko, M., Svitov, D., Zakharkin, I., Lempitsky, V.: Morf: Mobile realistic fullbody avatars from a monocular video. In: Proceedings of the IEEE/CVF Winter Conference on Applications of Computer Vision. pp. 3545-3555 (2024)", + "7. Bergman, A., Kellnhofer, P., Yifan, W., Chan, E., Lindell, D., Wetzstein, G.: Generative neural articulated radiance fields. Advances in Neural Information Processing Systems 35, 1990-19916 (2022)", + "8. Bertiche, H., Madadi, M., Escalera, S.: Cloth3d: clothed 3d humans. In: European Conference on Computer Vision. pp. 344-359. Springer (2020)", + "9. Bertiche, H., Madadi, M., Escalera, S.: Pbns: physically based neural simulation for unsupervised garment pose space deformation. ACM Transactions on Graphics (TOG) 40(6), 1-14 (2021)", + "10. Bertiche, H., Madadi, M., Escalera, S.: Neural cloth simulation. ACM Transactions on Graphics (TOG) 41(6), 1-14 (2022)", + "11. Bertiche, H., Madadi, M., Tyson, E., Escalera, S.: Deepsd: Automatic deep skinning and pose space deformation for 3d garment animation. In: Proceedings of the IEEE/CVF International Conference on Computer Vision. pp. 5471-5480 (2021)", + "12. Bhat, K.S., Twigg, C.D., Hodgins, J.K., Khosla, P.K., Popovic, Z., Seitz, S.M.: Estimating cloth simulation parameters from video. In: Proceedings of the 2003 ACM SIGGRAPH/Eurographics symposium on Computer animation. pp. 37-51 (2003)", + "13. Blender Online Coummunity: Blender, blender Foundation, https://www.blender.org/", + "14. Chen, B., Shen, Y., Shuai, Q., Zhou, X., Zhou, K., Zheng, Y.: Anidress: Animatable loose-dressed avatar from sparse views using garment rigging model. arXiv preprint arXiv:2401.15348 (2024)" + ], + "bbox": [ + 225, + 285, + 784, + 839 + ], + "page_idx": 14 + }, + { + "type": "header", + "text": "PhysAvatar", + "bbox": [ + 651, + 114, + 730, + 127 + ], + "page_idx": 14 + }, + { + "type": "page_number", + "text": "15", + "bbox": [ + 767, + 116, + 784, + 126 + ], + "page_idx": 14 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "15. Chen, H.Y., Sastry, A., van Rees, W.M., Vouga, E.: Physical simulation of environmentally induced thin shell deformation. ACM Transactions on Graphics (TOG) 37(4), 1-13 (2018)", + "16. Chen, X., Jiang, T., Song, J., Yang, J., Black, M.J., Geiger, A., Hilliges, O.: gdna: Towards generative detailed neural avatars. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 20427-20437 (2022)", + "17. Chen, X., Zheng, Y., Black, M.J., Hilliges, O., Geiger, A.: Snarf: Differentiable forward skimming for animating non-rigid neural implicit shapes. In: Proceedings of the IEEE/CVF International Conference on Computer Vision. pp. 11594-11604 (2021)", + "18. Chen, Y., Wang, X., Chen, X., Zhang, Q., Li, X., Guo, Y., Wang, J., Wang, F.: Uv volumes for real-time rendering of editable free-view human performance. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 16621-16631 (2023)", + "19. Chentanez, N., Macklin, M., Müller, M., Jeschke, S., Kim, T.Y.: Cloth and skin deformation with a triangle mesh based convolutional neural network. In: Computer Graphics Forum. vol. 39, pp. 123-134. Wiley Online Library (2020)", + "20. Clyde, D., Teran, J., Tamstorf, R.: Modeling and data-driven parameter estimation for woven fabrics. In: Proceedings of the ACM SIGGRAPH/Eurographics Symposium on Computer Animation. pp. 1-11 (2017)", + "21. Contributors, M.: Openmmlab pose estimation toolbox and benchmark. https://github.com/open-mmlab/mmpose (2020)", + "22. De Luigi, L., Li, R., Guillard, B., Salzmann, M., Fua, P.: Drapenet: Garment generation and self-supervised draping. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 1451-1460 (2023)", + "23. Dong, Z., Chen, X., Yang, J., Black, M.J., Hilliges, O., Geiger, A.: Ag3d: Learning to generate 3d avatars from 2d image collections. In: Proceedings of the IEEE/CVF International Conference on Computer Vision (ICCV). pp. 14916-14927 (October 2023)", + "24. Feng, Y., Yang, J., Pollefeys, M., Black, M.J., Bolkart, T.: Capturing and animation of body and clothing from monocular video. In: SIGGRAPH Asia 2022 Conference Papers. pp. 1-9 (2022)", + "25. Fridovich-Keil, S., Meanti, G., Warburg, F.R., Recht, B., Kanazawa, A.: K-planes: Explicit radiance fields in space, time, and appearance. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 12479-12488 (2023)", + "26. Genay, A., Lécuyer, A., Hachet, M.: Being an avatar \"for real\": a survey on virtual embodiment in augmented reality. IEEE Transactions on Visualization and Computer Graphics 28(12), 5071-5090 (2021)", + "27. Geng, C., Peng, S., Xu, Z., Bao, H., Zhou, X.: Learning neural volumetric representations of dynamic humans in minutes. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 8759-8770 (2023)", + "28. Grigorev, A., Black, M.J., Hilliges, O.: Hood: Hierarchical graphs for generalized modelling of clothing dynamics. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 16965-16974 (2023)", + "29. Grinspun, E., Hirani, A.N., Desbrun, M., Schröder, P.: Discrete shells. In: Proceedings of the 2003 ACM SIGGRAPH/Eurographics symposium on Computer animation. pp. 62-67. CiteSeer (2003)", + "30. Gundogdu, E., Constantin, V., Seifoddini, A., Dang, M., Salzmann, M., Fua, P.: Garnet: A two-stream network for fast and accurate 3d cloth draping. In:" + ], + "bbox": [ + 225, + 146, + 785, + 839 + ], + "page_idx": 15 + }, + { + "type": "page_number", + "text": "16", + "bbox": [ + 217, + 114, + 235, + 126 + ], + "page_idx": 15 + }, + { + "type": "header", + "text": "Yang Zheng et al.", + "bbox": [ + 271, + 114, + 390, + 128 + ], + "page_idx": 15 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "Proceedings of the IEEE/CVF International Conference on Computer Vision. pp. 8739-8748 (2019)", + "31. Guo, J., Li, J., Narain, R., Park, H.S.: Inverse simulation: Reconstructing dynamic geometry of clothed humans via optimal control. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 14698-14707 (2021)", + "32. Habermann, M., Liu, L., Xu, W., Pons-Moll, G., Zollhoefer, M., Theobalt, C.: Hdhumans: A hybrid approach for high-fidelity digital humans. Proceedings of the ACM on Computer Graphics and Interactive Techniques 6(3), 1-23 (2023)", + "33. Habermann, M., Liu, L., Xu, W., Zollhoefer, M., Pons-Moll, G., Theobalt, C.: Real-time deep dynamic characters. ACM Transactions on Graphics (ToG) 40(4), 1-16 (2021)", + "34. Hu, L., Zhang, H., Zhang, Y., Zhou, B., Liu, B., Zhang, S., Nie, L.: Gaussian-avatar: Towards realistic human avatar modeling from a single video via animatable 3d gaussians. arXiv preprint arXiv:2312.02134 (2023)", + "35. Isik, M., Rünz, M., Georgopoulos, M., Khakhulin, T., Starck, J., Agapito, L., Nießner, M.: Humanrf: High-fidelity neural radiance fields for humans in motion. ACM Transactions on Graphics (TOG) 42(4), 1-12 (2023). https://doi.org/10.1145/3592415, https://doi.org/10.1145/3592415", + "36. Jambon, C., Kerbl, B., Kopanas, G., Diolatzis, S., Leimkuhler, T., Drettakis, G.: Nerfshop: Interactive editing of neural radiance fields. Proc. ACM Comput. Graph. Interact. Tech. 6, 1:1-1:21 (2023). https://doi.org/10.1145/3585499", + "37. Jiang, Y., Wang, R., Liu, Z.: A survey of cloth simulation and applications. In: 2008 9th International Conference on Computer-Aided Industrial Design and Conceptual Design. pp. 765-769. IEEE (2008)", + "38. Jiang, Y., Shen, Z., Wang, P., Su, Z., Hong, Y., Zhang, Y., Yu, J., Xu, L.: Hifi4g: High-fidelity human performance rendering via compact gaussian splatting. arXiv preprint arXiv:2312.03461 (2023)", + "39. Kato, H., Ushiku, Y., Harada, T.: Neural 3d mesh renderer. 2018 IEEE/CVF Conference on Computer Vision and Pattern Recognition pp. 3907-3916 (2017). https://doi.org/10.1109/CVPR.2018.00411", + "40. Kerbl, B., Kopanas, G., Leimkuhler, T., Drettakis, G.: 3d gaussian splatting for real-time radiance field rendering. ACM Transactions on Graphics 42(4) (2023)", + "41. Keselman, L., Hebert, M.: Approximate differentiable rendering with algebraic surfaces. In: European Conference on Computer Vision. pp. 596-614. Springer (2022)", + "42. Komatsu, K.: Human skin model capable of natural shape variation. The visual computer 3, 265-271 (1988)", + "43. Korban, M., Li, X.: A survey on applications of digital human avatars toward virtual co-presence. arXiv preprint arXiv:2201.04168 (2022)", + "44. Kwon, Y., Kim, D., Ceylan, D., Fuchs, H.: Neural human performer: Learning generalizable radiance fields for human performance rendering. Advances in Neural Information Processing Systems 34, 24741-24752 (2021)", + "45. Kwon, Y., Kim, D., Ceylan, D., Fuchs, H.: Neural image-based avatars: Generalizable radiance fields for human avatar modeling. In: The Eleventh International Conference on Learning Representations (2023)", + "46. Kwon, Y., Liu, L., Fuchs, H., Habermann, M., Theobalt, C.: Deliffas: Deformable light fields for fast avatar synthesis. Advances in Neural Information Processing Systems 36 (2023)" + ], + "bbox": [ + 225, + 146, + 785, + 839 + ], + "page_idx": 16 + }, + { + "type": "header", + "text": "PhysAvatar", + "bbox": [ + 651, + 114, + 730, + 127 + ], + "page_idx": 16 + }, + { + "type": "page_number", + "text": "17", + "bbox": [ + 767, + 114, + 785, + 126 + ], + "page_idx": 16 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "47. Larionov, E., Eckert, M.L., Wolff, K., Stuyck, T.: Estimating cloth elasticity parameters using position-based simulation of compliant constrained dynamics. arXiv preprint arXiv:2212.08790 (2022)", + "48. Lassner, C., Zollhofer, M.: Pulsar: Efficient sphere-based neural rendering. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 1440-1449 (2021)", + "49. Lee, D., Kang, H., Lee, I.K.: Cloth combo: Modeling inter-cloth interaction for draping multi-layered clothes. ACM Transactions on Graphics (TOG) 42(6), 1-13 (2023)", + "50. LeVeque, R.J.: Finite difference methods for ordinary and partial differential equations: steady-state and time-dependent problems. SIAM (2007)", + "51. Li, M., Ferguson, Z., Schneider, T., Langlois, T.R., Zorin, D., Panozzo, D., Jiang, C., Kaufman, D.M.: Incremental potential contact: intersection-and inversion-free, large-deformation dynamics. ACM Trans. Graph. 39(4), 49 (2020)", + "52. Li, M., Kaufman, D.M., Jiang, C.: Codimensional incremental potential contact. ACM Transactions on Graphics (TOG) 40(4), 1-24 (2021)", + "53. Li, R., Tanke, J., Vo, M., Zollhofer, M., Gall, J., Kanazawa, A., Lassner, C.: Tava: Template-free animatable volumetric actors (2022)", + "54. Li, X., Li, X.R., Li, Y., Feng, W.: Review of cloth modeling and simulation for virtual fitting. Textile Research Journal 93(7-8), 1699-1711 (2023)", + "55. Li, Y., Chen, H.y., Larionov, E., Sarafianos, N., Matusik, W., Stuyck, T.: Difavatar: Simulation-ready garment optimization with differentiable simulation. arXiv preprint arXiv:2311.12194 (2023)", + "56. Li, Y., Du, T., Wu, K., Xu, J., Matusik, W.: Diffcloth: Differentiable cloth simulation with dry frictional contact. ACM Transactions on Graphics (TOG) 42(1), 1-20 (2022)", + "57. Li, Z., Zheng, Z., Liu, Y., Zhou, B., Liu, Y.: Posevocab: Learning joint-structured pose embeddings for human avatar modeling. arXiv preprint arXiv:2304.13006 (2023)", + "58. Li, Z., Zheng, Z., Wang, L., Liu, Y.: Animatable gaussians: Learning pose-dependent gaussian maps for high-fidelity human avatar modeling. arXiv preprint arXiv:2311.16096 (2023)", + "59. Liu, L., Habermann, M., Rudnev, V., Sarkar, K., Gu, J., Theobalt, C.: Neural actor: Neural free-view synthesis of human actors with pose control. ACM transactions on graphics (TOG) 40(6), 1-16 (2021)", + "60. Liu, L., Xu, W., Zollhoefer, M., Kim, H., Bernard, F., Habermann, M., Wang, W., Theobalt, C.: Neural rendering and reenactment of human actor videos. ACM Transactions on Graphics (TOG) 38(5), 1-14 (2019)", + "61. Liu, S., Li, T., Chen, W., Li, H.: Soft rasterizer: A differentiable renderer for image-based 3d reasoning. In: Proceedings of the IEEE/CVF International Conference on Computer Vision. pp. 7708-7717 (2019)", + "62. Liu, T., Bouaziz, S., Kavan, L.: Quasi-newton methods for real-time simulation of hyperelastic materials. Acm Transactions on Graphics (TOG) 36(3), 1-16 (2017)", + "63. Lombardi, S., Simon, T., Saragih, J., Schwartz, G., Lehrmann, A., Sheikh, Y.: Neural volumes: Learning dynamic renderable volumes from images. arXiv preprint arXiv:1906.07751 (2019)", + "64. Loper, M., Mahmood, N., Romero, J., Pons-Moll, G., Black, M.J.: Smpl: A skinned multi-person linear model. ACM Transactions on Graphics 34(6) (2015)", + "65. Luiten, J., Kopanas, G., Leibe, B., Ramanan, D.: Dynamic 3d gaussians: Tracking by persistent dynamic view synthesis. arXiv preprint arXiv:2308.09713 (2023)" + ], + "bbox": [ + 225, + 146, + 784, + 839 + ], + "page_idx": 17 + }, + { + "type": "page_number", + "text": "18", + "bbox": [ + 217, + 114, + 235, + 126 + ], + "page_idx": 17 + }, + { + "type": "header", + "text": "Yang Zheng et al.", + "bbox": [ + 271, + 114, + 390, + 128 + ], + "page_idx": 17 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "66. Ma, Q., Saito, S., Yang, J., Tang, S., Black, M.J.: Scale: Modeling clothed humans with a surface codec of articulated local elements. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 16082-16093 (2021)", + "67. Ma, Q., Yang, J., Black, M.J., Tang, S.: Neural point-based shape modeling of humans in challenging clothing. In: 2022 International Conference on 3D Vision (3DV). pp. 679-689. IEEE (2022)", + "68. Ma, Q., Yang, J., Ranjan, A., Pujades, S., Pons-Moll, G., Tang, S., Black, M.J.: Learning to dress 3d people in generative clothing. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 6469-6478 (2020)", + "69. Ma, Q., Yang, J., Tang, S., Black, M.J.: The power of points for modeling humans in clothing. In: Proceedings of the IEEE/CVF International Conference on Computer Vision. pp. 10974-10984 (2021)", + "70. Macklin, M., Müller, M., Chentanez, N.: Xpbd: position-based simulation of compliant constrained dynamics. In: Proceedings of the 9th International Conference on Motion in Games. p. 49-54. MIG '16, Association for Computing Machinery, New York, NY, USA (2016). https://doi.org/10.1145/2994258.2994272, https://doi.org/10.1145/2994258.2994272", + "71. Magnenat, T., Laperrière, R., Thalmann, D.: Joint-dependent local deformations for hand animation and object grasping. Tech. rep., Canadian Inf. Process. Soc (1988)", + "72. Mahmood, N., Ghorbani, N., Troje, N.F., Pons-Moll, G., Black, M.J.: AMASS: Archive of motion capture as surface shapes. In: ICCV (2019)", + "73. Mescheder, L., Oechsle, M., Niemeyer, M., Nowozin, S., Geiger, A.: Occupancy networks: Learning 3d reconstruction in function space. In: Proceedings of the IEEE/CVF conference on computer vision and pattern recognition. pp. 4460-4470 (2019)", + "74. Miguel, E., Bradley, D., Thomaszewski, B., Bickel, B., Matusik, W., Otaduy, M.A., Marschner, S.: Data-driven estimation of cloth simulation models. In: Computer Graphics Forum. vol. 31, pp. 519-528. Wiley Online Library (2012)", + "75. Mildenhall, B., Srinivasan, P.P., Tancik, M., Barron, J.T., Ramamoorthi, R., Ng, R.: Nerf: Representing scenes as neural radiance fields for view synthesis. Communications of the ACM 65(1), 99-106 (2021)", + "76. Mixamo, www MIXamo.com", + "77. Newcombe, R.A., Izadi, S., Hilliges, O., Molyneaux, D., Kim, D., Davison, A.J., Kohi, P., Shotton, J., Hodges, S., Fitzgibbon, A.: Kinectfusion: Real-time dense surface mapping and tracking. In: 2011 10th IEEE international symposium on mixed and augmented reality. pp. 127-136. IEEE (2011)", + "78. Nimier-David, M., Vicini, D., Zeltner, T., Jakob, W.: Mitsuba 2: A retargetable forward and inverse renderer. ACM Transactions on Graphics (TOG) 38(6), 1-17 (2019)", + "79. Noguchi, A., Sun, X., Lin, S., Harada, T.: Neural articulated radiance field. In: Proceedings of the IEEE/CVF International Conference on Computer Vision (ICCV). pp. 5762-5772 (October 2021)", + "80. Noguchi, A., Sun, X., Lin, S., Harada, T.: Unsupervised learning of efficient geometry-aware neural articulated representations. In: European Conference on Computer Vision. pp. 597-614. Springer (2022)", + "81. Pan, X., Mai, J., Jiang, X., Tang, D., Li, J., Shao, T., Zhou, K., Jin, X., Manocha, D.: Predicting loose-fitting garment deformations using bone-driven motion networks. In: ACM SIGGRAPH 2022 Conference Proceedings. pp. 1-10 (2022)" + ], + "bbox": [ + 225, + 146, + 785, + 839 + ], + "page_idx": 18 + }, + { + "type": "header", + "text": "PhysAvatar", + "bbox": [ + 651, + 114, + 730, + 127 + ], + "page_idx": 18 + }, + { + "type": "page_number", + "text": "19", + "bbox": [ + 767, + 114, + 785, + 126 + ], + "page_idx": 18 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "82. Pang, H., Zhu, H., Kortylewski, A., Theobalt, C., Habermann, M.: Ash: Animatable gaussian splats for efficient and photoreal human rendering. arXiv preprint arXiv:2312.05941 (2023)", + "83. Park, J.J., Florence, P., Straub, J., Newcombe, R., Lovegrove, S.: Deepsdf: Learning continuous signed distance functions for shape representation. In: Proceedings of the IEEE/CVF conference on computer vision and pattern recognition. pp. 165-174 (2019)", + "84. Park, K., Sinha, U., Hedman, P., Barron, J.T., Bouaziz, S., Goldman, D.B., Martin-Brualla, R., Seitz, S.M.: Hypernerf: a higher-dimensional representation for topologically varying neural radiance fields. ACM Transactions on Graphics (TOG) 40(6), 1-12 (2021)", + "85. Pavlakos, G., Choutas, V., Ghorbani, N., Bolkart, T., Osman, A.A.A., Tzionas, D., Black, M.J.: Expressive body capture: 3D hands, face, and body from a single image. In: Proceedings IEEE Conf. on Computer Vision and Pattern Recognition (CVPR). pp. 10975-10985 (2019)", + "86. Peng, S., Dong, J., Wang, Q., Zhang, S., Shuai, Q., Zhou, X., Bao, H.: Animatable neural radiance fields for modeling dynamic human bodies. In: Proceedings of the IEEE/CVF International Conference on Computer Vision. pp. 14314-14323 (2021)", + "87. Peng, S., Zhang, Y., Xu, Y., Wang, Q., Shuai, Q., Bao, H., Zhou, X.: Neural body: Implicit neural representations with structured latent codes for novel view synthesis of dynamic humans. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 9054-9063 (2021)", + "88. Peng, T., Wu, W., Liu, J., Li, L., Miao, J., Hu, X., He, R., Li, L.: Pgn-cloth: Physics-based graph network model for 3d cloth animation. Displays 80, 102534 (2023)", + "89. Pumarola, A., Corona, E., Pons-Moll, G., Moreno-Noguer, F.: D-nerf: Neural radiance fields for dynamic scenes. 2021 IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR) pp. 10313-10322 (2020). https://doi.org/10.1109/CVPR46437.2021.01018", + "90. Saito, S., Yang, J., Ma, Q., Black, M.J.: Scanimate: Weakly supervised learning of skinned clothed avatar networks. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 2886-2897 (2021)", + "91. Santesteban, I., Otaduy, M.A., Casas, D.: Learning-based animation of clothing for virtual try-on. In: Computer Graphics Forum. vol. 38, pp. 355-366. Wiley Online Library (2019)", + "92. Santesteban, I., Otaduy, M.A., Casas, D.: Snug: Self-supervised neural dynamic garments. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 8140-8150 (2022)", + "93. Schonberger, J.L., Frahm, J.M.: Structure-from-motion revisited. In: Proceedings of the IEEE conference on computer vision and pattern recognition. pp. 4104-4113 (2016)", + "94. Sitzmann, V., Thies, J., Heide, F., Nießner, M., Wetzstein, G., Zollhöfer, M.: Deepvoxels: Learning persistent 3d feature embeddings. In: Proc. Computer Vision and Pattern Recognition (CVPR), IEEE (2019)", + "95. Sitzmann, V., Zollhöfer, M., Wetzstein, G.: Scene representation networks: Continuous 3d-structure-aware neural scene representations. In: Advances in Neural Information Processing Systems (2019)", + "96. Snavely, N., Seitz, S.M., Szeliski, R.: Photo tourism: exploring photo collections in 3d. In: ACM siggraph 2006 papers, pp. 835-846 (2006)" + ], + "bbox": [ + 223, + 146, + 784, + 839 + ], + "page_idx": 19 + }, + { + "type": "page_number", + "text": "20", + "bbox": [ + 217, + 114, + 235, + 126 + ], + "page_idx": 19 + }, + { + "type": "header", + "text": "Yang Zheng et al.", + "bbox": [ + 271, + 114, + 390, + 128 + ], + "page_idx": 19 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "97. Stoll, C., Gall, J., De Aguiar, E., Thrun, S., Theobalt, C.: Video-based reconstruction of animatable human characters. ACM Transactions on Graphics (TOG) 29(6), 1-10 (2010)", + "98. Stuyck, T., Chen, H.y.: Diffxpbd: Differentiable position-based simulation of compliant constraint dynamics. Proceedings of the ACM on Computer Graphics and Interactive Techniques 6(3), 1-14 (2023)", + "99. Su, S.Y., Yu, F., Zollhöfer, M., Rhodin, H.: A-nerf: Articulated neural radiance fields for learning human shape, appearance, and pose. Advances in Neural Information Processing Systems 34, 12278–12291 (2021)", + "100. Su, Z., Hu, L., Lin, S., Zhang, H., Zhang, S., Thies, J., Liu, Y.: Caphy: Capturing physical properties for animatable human avatars. In: Proceedings of the IEEE/CVF International Conference on Computer Vision. pp. 14150-14160 (2023)", + "101. Sun, M., Yang, D., Kou, D., Jiang, Y., Shan, W., Yan, Z., Zhang, L.: Human 3d avatar modeling with implicit neural representation: A brief survey. In: 2022 14th International Conference on Signal Processing Systems (ICSPS). pp. 818-827. IEEE (2022)", + "102. Tamstorf, R., Grinspun, E.: Discrete bending forces and their jacobians. Graphical models 75(6), 362-370 (2013)", + "103. Tang, M., Tong, R., Narain, R., Meng, C., Manocha, D.: A gpu-based streaming algorithm for high-resolution cloth simulation. In: Computer Graphics Forum. vol. 32, pp. 21-30. Wiley Online Library (2013)", + "104. Tatarchenko, M., Richter, S.R., Ranftl, R., Li, Z., Koltun, V., Brox, T.: What do single-view 3d reconstruction networks learn? In: Proceedings of the IEEE/CVF conference on computer vision and pattern recognition. pp. 3405-3414 (2019)", + "105. Tewari, A., Fried, O., Thies, J., Sitzmann, V., Lombardi, S., Sunkavalli, K., Martin-Brualla, R., Simon, T., Saragih, J.M., Nießner, M., Pandey, R., Fanello, S., Wetzstein, G., Zhu, J.Y., Theobalt, C., Agrawala, M., Shechtman, E., Goldman, D.B., Zollhofer, M.: State of the art on neural rendering. Computer Graphics Forum 39 (2020). https://doi.org/10.1111/cgf.14022", + "106. Tewari, A., Thies, J., Mildenhall, B., Srinivasan, P., Tretschk, E., Yifan, W., Lassner, C., Sitzmann, V., Martin-Brualla, R., Lombardi, S., et al.: Advances in neural rendering. In: Computer Graphics Forum. vol. 41, pp. 703-735. Wiley Online Library (2022)", + "107. Tiwari, G., Sarafianos, N., Tung, T., Pons-Moll, G.: Neural-gif: Neural generalized implicit functions for animating people in clothing. In: Proceedings of the IEEE/CVF International Conference on Computer Vision. pp. 11708-11718 (2021)", + "108. Tretschk, E., Tewari, A., Golyanik, V., Zollhöfer, M., Lassner, C., Theobalt, C.: Non-rigid neural radiance fields: Reconstruction and novel view synthesis of a dynamic scene from monocular video. 2021 IEEE/CVF International Conference on Computer Vision (ICCV) pp. 12939-12950 (2020). https://doi.org/10.1109/ICCV48922.2021.01272", + "109. Vassilev, T., Spanlang, B., Chrysanthou, Y.: Fast cloth animation on walking avatars. In: Computer Graphics Forum. vol. 20, pp. 260-267. Wiley Online Library (2001)", + "110. Vidaurre, R., Santesteban, I., Garces, E., Casas, D.: Fully convolutional graph neural networks for parametric virtual try-on. In: Computer Graphics Forum. vol. 39, pp. 145-156. Wiley Online Library (2020)", + "111. Wang, H.: Gpu-based simulation of cloth wrinkles at submillimeter levels. ACM Transactions on Graphics (TOG) 40(4), 1-14 (2021)" + ], + "bbox": [ + 218, + 146, + 784, + 839 + ], + "page_idx": 20 + }, + { + "type": "header", + "text": "PhysAvatar", + "bbox": [ + 651, + 114, + 730, + 127 + ], + "page_idx": 20 + }, + { + "type": "page_number", + "text": "21", + "bbox": [ + 767, + 114, + 782, + 126 + ], + "page_idx": 20 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "112. Wang, H., O'Brien, J.F., Ramamoorthi, R.: Data-driven elastic models for cloth: modeling and measurement. ACM transactions on graphics (TOG) 30(4), 1-12 (2011)", + "113. Wang, P., Liu, L., Liu, Y., Theobalt, C., Komura, T., Wang, W.: Neus: Learning neural implicit surfaces by volume rendering for multi-view reconstruction. Advances in Neural Information Processing Systems 34, 27171-27183 (2021)", + "114. Wang, S., Schwarz, K., Geiger, A., Tang, S.: Arah: Animatable volume rendering of articulated human sdfs. In: European conference on computer vision. pp. 1-19. Springer (2022)", + "115. Wang, S., Schwarz, K., Geiger, A., Tang, S.: Arah: Animatable volume rendering of articulated human sdfs. In: European Conference on Computer Vision (2022)", + "116. Wang, T., Zhang, B., Zhang, T., Gu, S., Bao, J., Baltrusaitis, T., Shen, J., Chen, D., Wen, F., Chen, Q., et al.: Rodin: A generative model for sculpting 3d digital avatars using diffusion. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 4563-4573 (2023)", + "117. Wu, G., Yi, T., Fang, J., Xie, L., Zhang, X., Wei, W., Liu, W., Tian, Q., Wang, X.: 4d gaussian splatting for real-time dynamic scene rendering. ArXiv abs/2310.08528 (2023). https://doi.org/10.48550/arXiv.2310.08528", + "118. Xiang, D., Bagautdinov, T., Stuyck, T., Prada, F., Romero, J., Xu, W., Saito, S., Guo, J., Smith, B., Shiratori, T., et al.: Dressing avatars: Deep photorealistic appearance for physically simulated clothing. ACM Transactions on Graphics (TOG) 41(6), 1-15 (2022)", + "119. Xu, Y., Yifan, W., Bergman, A.W., Chai, M., Zhou, B., Wetzstein, G.: Efficient 3d articulated human generation with layered surface volumes. arXiv preprint arXiv:2307.05462 (2023)", + "120. Xu, Y., Chen, B., Li, Z., Zhang, H., Wang, L., Zheng, Z., Liu, Y.: Gaussian head avatar: Ultra high-fidelity head avatar via dynamic gaussians. arXiv preprint arXiv:2312.03029 (2023)", + "121. Yang, Z., Zeng, A., Yuan, C., Li, Y.: Effective whole-body pose estimation with two-stages distillation. In: Proceedings of the IEEE/CVF International Conference on Computer Vision. pp. 4210-4220 (2023)", + "122. Yariv, L., Gu, J., Kasten, Y., Lipman, Y.: Volume rendering of neural implicit surfaces. Advances in Neural Information Processing Systems 34, 4805-4815 (2021)", + "123. Yee, K.S., Chen, J.S.: The finite-difference time-domain (fdtd) and the finite-volume time-domain (fvtd) methods in solving maxwell's equations. IEEE Transactions on Antennas and Propagation 45(3), 354-363 (1997)", + "124. Yifan, W., Serena, F., Wu, S., Öztireli, C., Sorkine-Hornung, O.: Differentiable surface splatting for point-based geometry processing. ACM Transactions on Graphics (TOG) 38(6), 1-14 (2019)", + "125. Yu, T., Zheng, Z., Zhong, Y., Zhao, J., Dai, Q., Pons-Moll, G., Liu, Y.: Simulcap: Single-view human performance capture with cloth simulation. In: Proceedings of the IEEE/CVF conference on computer vision and pattern recognition. pp. 5504-5514 (2019)", + "126. Yuan, Y.J., Sun, Y.T., Lai, Y.K., Ma, Y., Jia, R., Gao, L.: Nerf-editing: geometry editing of neural radiance fields. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 18353-18364 (2022)", + "127. Zeller, C.: Cloth simulation on thegpu. In: ACM SIGGRAPH 2005 Sketches, pp. 39-es (2005)", + "128. Zhang, J.X., Lin, G.W.C., Bode, L., Chen, H.y., Stuyck, T., Larionov, E.: Estimating cloth elasticity parameters from homogenized yarn-level models. arXiv preprint arXiv:2401.15169 (2024)" + ], + "bbox": [ + 215, + 146, + 785, + 839 + ], + "page_idx": 21 + }, + { + "type": "page_number", + "text": "22", + "bbox": [ + 217, + 114, + 235, + 126 + ], + "page_idx": 21 + }, + { + "type": "header", + "text": "Yang Zheng et al.", + "bbox": [ + 271, + 114, + 390, + 128 + ], + "page_idx": 21 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "129. Zhang, R., Isola, P., Efros, A.A., Shechtman, E., Wang, O.: The unreasonable effectiveness of deep features as a perceptual metric. In: Proceedings of the IEEE conference on computer vision and pattern recognition. pp. 586-595 (2018)", + "130. Zhao, F., Jiang, Y., Yao, K., Zhang, J., Wang, L., Dai, H., Zhong, Y., Zhang, Y., Wu, M., Xu, L., et al.: Human performance modeling and rendering via neural animated mesh. ACM Transactions on Graphics (TOG) 41(6), 1-17 (2022)", + "131. Zheng, Z., Huang, H., Yu, T., Zhang, H., Guo, Y., Liu, Y.: Structured local radiance fields for human avatar modeling. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 15893-15903 (2022)", + "132. Zielonka, W., Bagautdinov, T., Saito, S., Zollhöfer, M., Thies, J., Romero, J.: Drivable 3d gaussian avatars. arXiv preprint arXiv:2311.08581 (2023)" + ], + "bbox": [ + 217, + 146, + 785, + 299 + ], + "page_idx": 22 + }, + { + "type": "header", + "text": "PhysAvatar", + "bbox": [ + 651, + 114, + 730, + 127 + ], + "page_idx": 22 + }, + { + "type": "page_number", + "text": "23", + "bbox": [ + 767, + 114, + 785, + 126 + ], + "page_idx": 22 + } +] \ No newline at end of file diff --git a/2024/PhysAvatar_ Learning the Physics of Dressed 3D Avatars from Visual Observations/958c8811-3fcb-45f8-9957-bab313d0628e_model.json b/2024/PhysAvatar_ Learning the Physics of Dressed 3D Avatars from Visual Observations/958c8811-3fcb-45f8-9957-bab313d0628e_model.json new file mode 100644 index 0000000000000000000000000000000000000000..93b90e6b1965651ef624202487d53b097547dfa8 --- /dev/null +++ b/2024/PhysAvatar_ Learning the Physics of Dressed 3D Avatars from Visual Observations/958c8811-3fcb-45f8-9957-bab313d0628e_model.json @@ -0,0 +1,3689 @@ +[ + [ + { + "type": "title", + "bbox": [ + 0.225, + 0.141, + 0.78, + 0.187 + ], + "angle": 0, + "content": "PhysAvatar: Learning the Physics of Dressed 3D Avatars from Visual Observations" + }, + { + "type": "text", + "bbox": [ + 0.217, + 0.212, + 0.786, + 0.259 + ], + "angle": 0, + "content": "Yang Zheng\\(^{1\\star}\\), Qingqing Zhao\\(^{1\\star}\\), Guandao Yang\\(^{1}\\), Wang Yifan\\(^{1}\\), Donglai Xiang\\(^{2}\\), Florian Dubost\\(^{3}\\), Dmitry Lagun\\(^{3}\\), Thabo Beeler\\(^{3}\\), Federico Tombari\\(^{3,4}\\), Leonidas Guibas\\(^{1}\\), and Gordon Wetzstein\\(^{1}\\)" + }, + { + "type": "text", + "bbox": [ + 0.428, + 0.27, + 0.576, + 0.285 + ], + "angle": 0, + "content": "1 Stanford University" + }, + { + "type": "text", + "bbox": [ + 0.403, + 0.284, + 0.6, + 0.299 + ], + "angle": 0, + "content": "2 Carnegie Mellon University" + }, + { + "type": "text", + "bbox": [ + 0.471, + 0.298, + 0.533, + 0.312 + ], + "angle": 0, + "content": "3 Google" + }, + { + "type": "text", + "bbox": [ + 0.392, + 0.312, + 0.612, + 0.325 + ], + "angle": 0, + "content": "4 Technical University of Munich" + }, + { + "type": "text", + "bbox": [ + 0.261, + 0.363, + 0.744, + 0.613 + ], + "angle": 0, + "content": "Abstract. Modeling and rendering photorealistic avatars is of crucial importance in many applications. Existing methods that build a 3D avatar from visual observations, however, struggle to reconstruct clothed humans. We introduce PhysAvatar, a novel framework that combines inverse rendering with inverse physics to automatically estimate the shape and appearance of a human from multi-view video data along with the physical parameters of the fabric of their clothes. For this purpose, we adopt a mesh-aligned 4D Gaussian technique for spatio-temporal mesh tracking as well as a physically based inverse renderer to estimate the intrinsic material properties. PhysAvatar integrates a physics simulator to estimate the physical parameters of the garments using gradient-based optimization in a principled manner. These novel capabilities enable PhysAvatar to create high-quality novel-view renderings of avatars dressed in loose-fitting clothes under motions and lighting conditions not seen in the training data. This marks a significant advancement towards modeling photorealistic digital humans using physically based inverse rendering with physics in the loop. Our project website is at: https://qingqing-zhao.github.io/PhysAvatar." + }, + { + "type": "text", + "bbox": [ + 0.262, + 0.626, + 0.738, + 0.641 + ], + "angle": 0, + "content": "Keywords: neural rendering \\(\\cdot\\) physics \\(\\cdot\\) dynamic modeling \\(\\cdot\\) 3D avatar" + }, + { + "type": "title", + "bbox": [ + 0.216, + 0.666, + 0.377, + 0.681 + ], + "angle": 0, + "content": "1 Introduction" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.697, + 0.788, + 0.818 + ], + "angle": 0, + "content": "Digital avatars are a vital component in numerous applications, ranging from virtual reality and gaming to telepresence and e-commerce [26,43,101]. A realistic 3D avatar of a person can be readily obtained from visual observations, such as multi-view image [116] and video [6] data. The task of rendering animated 3D avatars from novel viewpoints or when performing unseen motions, however, presents considerable challenges, particularly when the avatar wears loose-fitting garments. Accurately rendering the dynamics of garments in conditions that are not observed in the training data necessitates a holistic approach that not only" + }, + { + "type": "page_footnote", + "bbox": [ + 0.218, + 0.825, + 0.367, + 0.84 + ], + "angle": 0, + "content": "* Equal Contribution" + } + ], + [ + { + "type": "page_number", + "bbox": [ + 0.218, + 0.116, + 0.23, + 0.127 + ], + "angle": 0, + "content": "2" + }, + { + "type": "header", + "bbox": [ + 0.272, + 0.115, + 0.393, + 0.129 + ], + "angle": 0, + "content": "Yang Zheng et al." + }, + { + "type": "image", + "bbox": [ + 0.248, + 0.148, + 0.526, + 0.367 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.535, + 0.249, + 0.553, + 0.261 + ], + "angle": 0, + "content": "(b)" + }, + { + "type": "image", + "bbox": [ + 0.569, + 0.153, + 0.738, + 0.263 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.564, + 0.272, + 0.75, + 0.365 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.215, + 0.373, + 0.788, + 0.415 + ], + "angle": 0, + "content": "Fig. 1: PhysAvatar is a novel framework that captures the physics of dressed 3D avatars from visual observations, enabling a wide spectrum of applications, such as (a) animation, (b) relighting, and (c) redressing, with high-fidelity rendering results." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.446, + 0.785, + 0.476 + ], + "angle": 0, + "content": "models the shape and appearance of the person but also the physical behavior of their clothes, including friction and collision." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.478, + 0.789, + 0.809 + ], + "angle": 0, + "content": "Learning 3D scene representations from visual data is the core problem addressed by inverse rendering—a field that has recently shown remarkable progress in estimating the geometry and appearance of static and dynamic scenes from multi-view images and videos [105, 106]. In the context of reconstructing 3D avatars, the ability to explicitly control the motions of the avatar in post processing becomes imperative. Most existing methods for reconstructing animatable avatars drive the animation through an underlying skeleton using linear blending skinning (LBS) [64], which adequately models the dynamics of humans dressed in tight-fitting clothes [1,7,23,80,99,119]. With this approach, garments are treated as a rigidly attached part of the body adhering to piece-wise linear transformations, which results in motions that appear rigid and unconvincing in many cases [79, 120, 132]. Several successful attempts to introduce non-rigid deformations through pose-dependent geometry adjustments, such as wrinkles, have recently been made [18,57,59,60,114], although these methods are prone to overfitting to the body poses and motion sequences observed during training. A key problem is that most existing 3D avatar reconstruction approaches neglect to model the dynamics of loose garments in a physically accurate manner, leading to unrealistic fabric behavior and issues like self-penetration. To our knowledge, only the work of Xiang et al. [118] includes a physics-based approach to inverse rendering of digital humans, but their approach requires a tedious manual search process to find the parameters that model the dynamic behavior of their cloth fabrics with reasonable accuracy." + }, + { + "type": "text", + "bbox": [ + 0.215, + 0.81, + 0.789, + 0.84 + ], + "angle": 0, + "content": "Here, we introduce PhysAvatar, a novel approach to 3D avatar reconstruction from multi-view video data. PhysAvatar combines inverse rendering with" + } + ], + [ + { + "type": "header", + "bbox": [ + 0.653, + 0.115, + 0.732, + 0.129 + ], + "angle": 0, + "content": "PhysAvatar" + }, + { + "type": "page_number", + "bbox": [ + 0.775, + 0.117, + 0.785, + 0.127 + ], + "angle": 0, + "content": "3" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.147, + 0.788, + 0.357 + ], + "angle": 0, + "content": "\"inverse physics\" in a principled manner to not only estimate the shape and appearance of the avatar but also physical parameters, including density and stiffness, of the fabrics modeling their (loose) clothes. Our approach includes several parts: Given the reconstructed 3D mesh of the garment in one frame, we first leverage mesh-aligned 4D Gaussians [117] to track surface points of the garment across all frames of the input video sequence. This process establishes dense correspondences on the 3D garment. These data are used as supervision in an inverse physics step, where the physical parameters of the fabric are optimized using a finite-difference approach [50,123]. Finally, we employ a physically based inverse renderer [78] to jointly estimate ambient lighting and the surface material. By leveraging the refined geometry from our simulation step, the inverse renderer can effectively factor in pose-dependent effects, such as shadows caused by self-occlusion, resulting in accurate appearance reconstruction that enables the avatar to be rendered in novel lighting conditions." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.358, + 0.788, + 0.403 + ], + "angle": 0, + "content": "PhysAvatar offers a comprehensive solution for reconstructing clothed human avatars from multi-view video data to perform novel view and motion synthesis with state-of-the-art realism. The key contributions of our work include:" + }, + { + "type": "text", + "bbox": [ + 0.223, + 0.404, + 0.784, + 0.447 + ], + "angle": 0, + "content": "1. The introduction of a new inverse rendering paradigm for avatars created from real-world captures that incorporates the physics of loose garments in a principled manner;" + }, + { + "type": "text", + "bbox": [ + 0.222, + 0.449, + 0.785, + 0.507 + ], + "angle": 0, + "content": "2. A pipeline that includes accurate and efficient mesh reconstruction and tracking using 4D Gaussians; automatic optimization of the garments' physical material properties; and accurate appearance estimation using physically based inverse rendering." + }, + { + "type": "list", + "bbox": [ + 0.222, + 0.404, + 0.785, + 0.507 + ], + "angle": 0, + "content": null + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.509, + 0.788, + 0.555 + ], + "angle": 0, + "content": "PhysAvatar demonstrates a novel path to learning physical scene properties from visual observations by incorporating inverse rendering with physics constraints. Code will be available on the project website." + }, + { + "type": "title", + "bbox": [ + 0.215, + 0.573, + 0.388, + 0.589 + ], + "angle": 0, + "content": "2 Related Work" + }, + { + "type": "title", + "bbox": [ + 0.215, + 0.6, + 0.655, + 0.615 + ], + "angle": 0, + "content": "2.1 Scene Reconstruction from Visual Observations" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.621, + 0.788, + 0.773 + ], + "angle": 0, + "content": "Reconstructing 3D scenes from visual observations is grounded in classic approaches, such as Structure from Motion (SfM) [77, 93, 96]. Recently, this field has witnessed a paradigm shift towards inverse rendering techniques that estimate the shape and appearance of a scene from visual input [105, 106]. Among the many representations developed in this area are neural (radiance) fields [75, 95], neural volumes, surfaces, and signed distance fields [63, 73, 83, 94, 113, 122] as well as differentiable rasterization techniques for meshes [39, 61] and point clouds [40, 41, 48, 124]. Recent inverse rendering approaches are capable of modeling dynamic scenes [25, 84, 89, 108] and enable sophisticated post-capture deformations [36, 126]." + }, + { + "type": "title", + "bbox": [ + 0.215, + 0.789, + 0.427, + 0.804 + ], + "angle": 0, + "content": "2.2 Animatable Avatars" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.81, + 0.787, + 0.842 + ], + "angle": 0, + "content": "One central requirement for generating full-body digital humans is to enable animations driven by skeleton-level control [42, 71]. To this end, early works on" + } + ], + [ + { + "type": "page_number", + "bbox": [ + 0.218, + 0.116, + 0.23, + 0.127 + ], + "angle": 0, + "content": "4" + }, + { + "type": "header", + "bbox": [ + 0.272, + 0.115, + 0.392, + 0.129 + ], + "angle": 0, + "content": "Yang Zheng et al." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.147, + 0.788, + 0.298 + ], + "angle": 0, + "content": "avatar reconstruction primarily relied on skinning techniques, which successfully model the dynamics of bodies with minimal or tight clothing [3,64]. Clothing is an integral part of everyday human appearance, captured by the saying \"clothes make the (wo)man\". Reconstructing their dynamic behavior from visual data, however, remains a challenge. Several different types of representations have been explored for clothed avatars, including meshes [68] with dynamic textures [4, 33, 130], neural surface [16, 90, 107] and radiance fields [14, 18, 24, 27, 44-46, 86, 87, 99, 131], point sets [66, 67, 69], and 3D Gaussians [34, 58, 82, 132]. Many of these works condition the deformation of clothing on the body pose and some predict future dynamics based on a small number of frames [32, 33, 82]." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.298, + 0.788, + 0.601 + ], + "angle": 0, + "content": "We now discuss recent research most closely related to ours. TAVA [53] learns a deformable neural radiance field of humans or animals using a differentiable skinning method like SNARF [17]. ARAH [115] models 3D avatars using SDF representations and proposes an SDF-based volume rendering method to reconstruct the avatar in canonical space. GS-Avatar [34] achieves real-time animation by leveraging a learned 3D Gaussian predictor. Although these methods show reasonable effectiveness on humans wearing tight garments, they struggle with loose garments because their LBS-based deformation module cannot accommodate the complex dynamics of loose clothing. To address this limitation, Xiang et al. [118] incorporate a cloth simulator based on the eXtended Position Based Dynamics (XPBD) formulation [70] into a deep appearance model to reconstruct realistic dynamics and appearance. However, they require tedious manual adjustments of physical parameters to produce reasonable garment motion, and the code is not publicly available. Concurrent with our work, AniDress [14], while using physics-based simulation, only uses the simulation to produce a garment rigging model. This can cause inaccuracy in cloth dynamics modeling and unrealistic motion artifacts. In contrast, we estimate the physical parameters through a principled gradient-based inverse physics approach and accurately recover the shape and (relightable) appearance using a physically based inverse renderer, achieving high-fidelity results on novel motions." + }, + { + "type": "title", + "bbox": [ + 0.215, + 0.621, + 0.473, + 0.636 + ], + "angle": 0, + "content": "2.3 Physics-Based Simulation" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.644, + 0.788, + 0.842 + ], + "angle": 0, + "content": "Physics-based simulation of cloth [5, 62, 88, 103, 109, 111, 125, 127] has been extensively studied and is particularly useful for modeling complicated dynamic effects in interaction with human bodies, such as large deformations, wrinkling, and contact. In this work, we adopt the Codimensional Incremental Potential Contact (C-IPC) solver [52] for its robustness in handling complicated body-cloth collision by the log-barrier penalty function [51]. We refer readers to dedicated surveys [37, 54] for a comprehensive review, and focus on works that solve the inverse parameter estimation problem. One line of work uses specialized devices to simultaneously measure forces and deformation in a controlled setting [47, 74, 112, 128]. Our work falls into the category that estimates physical parameters directly from imagery of garments worn on human subjects [12, 31, 97]. For this purpose, some differentiable simulators have been developed and applied to the optimization problem [55], but only for specific types of solvers," + } + ], + [ + { + "type": "header", + "bbox": [ + 0.653, + 0.115, + 0.732, + 0.128 + ], + "angle": 0, + "content": "PhysAvatar" + }, + { + "type": "page_number", + "bbox": [ + 0.775, + 0.117, + 0.785, + 0.127 + ], + "angle": 0, + "content": "5" + }, + { + "type": "image", + "bbox": [ + 0.264, + 0.162, + 0.734, + 0.307 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.214, + 0.324, + 0.788, + 0.422 + ], + "angle": 0, + "content": "Fig. 2: Method Overview: (a) PhysAvatar takes multi-view videos and an initial mesh as input. We first perform (b) dynamic mesh tracking (Sec. 3.1). The tacked mesh sequences are then used for (c) garment physics estimation with a physics simulator combined with gradient-based optimization (Sec. 3.2); (d) and appearance estimation through physics-based differentiable rendering (Sec. 3.3). At test time, (e) given a sequence of body poses (f), we simulate garment dynamics with the learned physics parameters and employ physics-based rendering to produce the final images." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.451, + 0.789, + 0.709 + ], + "angle": 0, + "content": "such as Projective Dynamics [56] and XPBD [98]. The heavy runtime cost of high-quality physics-based simulation motivates research in neural cloth simulation. Early works create datasets of garments simulated on human bodies [8,30], and then train the network to predict garment deformation in a supervised manner [19,81,91,110]. Recently, more attention has been paid to the self-supervised formulation that applies the elasticity energy defined in traditional simulation to predict garments, thereby training the network to act like a simulation solver, including both the (quasi-) static [9,11,22,49,92] and the dynamic [10,28] cases. Most of these works focus on making forward predictions that resemble traditional simulators without solving inverse parameter estimation on real-world garment data, except CaPhy [100]. CaPhy estimates the Saint Venant-Kirchhoff (StVK) elasticity parameters of clothing from 4D scans but only predicts pose-dependent clothing deformation in a quasi-static manner. In addition, all the works mentioned above treat simulation as a separate problem without modeling the photorealistic appearance. By comparison, our work builds avatars with physically based appearance and dynamics that are optimized for faithfulness to the real human capture in a holistic manner." + }, + { + "type": "title", + "bbox": [ + 0.216, + 0.732, + 0.331, + 0.749 + ], + "angle": 0, + "content": "3 Method" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.765, + 0.788, + 0.841 + ], + "angle": 0, + "content": "Given multi-view videos as input, our goal is to reconstruct a 3D avatar that can be realistically animated and rendered under novel motions and novel views. The core pipeline (Fig. 2) consists of three primary components: 1. Mesh tracking (Sec. 3.1)—given multi-view videos and a reconstructed mesh at the initial time step as input, we track the deformation of the geometry, which provides the" + } + ], + [ + { + "type": "page_number", + "bbox": [ + 0.218, + 0.116, + 0.23, + 0.127 + ], + "angle": 0, + "content": "6" + }, + { + "type": "header", + "bbox": [ + 0.272, + 0.115, + 0.393, + 0.129 + ], + "angle": 0, + "content": "Yang Zheng et al." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.147, + 0.788, + 0.269 + ], + "angle": 0, + "content": "geometry reference for the ensuing physics simulation; 2. Physics-based dynamic modeling (Sec. 3.2)—with the tracked meshes as reference, we estimate the garments' physical properties using a physics simulator; the simulator, together with the estimated physical parameters are used to generate novel motion; 3. Physics-based appearance refinement (Sec. 3.3)—given the tracked geometry, we further enhance the appearance estimation through a physics-based differentiable renderer, which considers self-occlusion and thus eliminate artifacts such as baked-in shadows. In the following, we delve into the details of each component." + }, + { + "type": "title", + "bbox": [ + 0.215, + 0.292, + 0.388, + 0.308 + ], + "angle": 0, + "content": "3.1 Mesh Tracking" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.318, + 0.789, + 0.44 + ], + "angle": 0, + "content": "Given multiview videos of \\(T\\) frames and \\(C\\) views, \\(\\{\\mathbf{I}_{1:T,1:C}\\}\\), and an initial mesh \\((\\mathbf{V}_1,\\mathbf{F})\\), with vertices \\(\\mathbf{V}_1\\) and faces \\(\\mathbf{F}\\), that can be obtained via any existing static scene reconstruction methods, our goal is to accurately track the mesh deformation through the video sequence \\(\\mathbf{V}_{1:T}\\), which will be used as reference in the subsequent physics simulation. However accurately tracking mesh deformation from visual observation is very challenging. Inspired by the robust and real-time tracking performance from dynamic 3D Gaussians [65], we use 3D Gaussians as a surrogate representation for mesh tracking." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.44, + 0.79, + 0.637 + ], + "angle": 0, + "content": "We adapt 3D Gaussians to align with the reconstructed mesh surface. Following the definition of dynamic 3D Gaussians, at frame \\( t \\), each Gaussian is parameterized with position \\( \\mathbf{p}_t \\in \\mathbb{R}^3 \\), rotation quaternion \\( \\mathbf{q}_t \\in \\mathbb{R}^4 \\), color \\( \\mathbf{c}_t \\in \\mathbb{R}^3 \\), scale \\( \\mathbf{s}_t \\in \\mathbb{R}^3 \\) and opacity \\( o_t \\in \\mathbb{R}_+ \\). To couple the Gaussians with the mesh, we attach one Gaussian at the barycenter of each face \\( \\mathbf{f} \\in \\mathbf{F} \\), and determine the Gaussians' quaternions from the attached triangle. Specifically, the rotation of each Gaussian is computed from the rotation of its local frame, whose \\( x- \\) and \\( z \\)-axis are defined using the longest triangle edge at \\( t = 1 \\) and the face normal respectively. Furthermore, we set the last scaling factor, which corresponds to the scaling in face normal direction, to be a small constant value, ensuring the Gaussian lies on the mesh surface. The Gaussians can then be rendered in a differentiable manner [40] and we optimize their parameters by minimizing the loss [65] between the rendered images and the reference images:" + }, + { + "type": "equation", + "bbox": [ + 0.317, + 0.649, + 0.787, + 0.677 + ], + "angle": 0, + "content": "\\[\n\\mathcal {L} _ {\\mathrm {r e n d e r}} = \\lambda \\left\\| \\mathbf {I} _ {i, t} - \\hat {\\mathbf {I}} _ {i, t} \\right\\| _ {1} + (1 - \\lambda) \\cdot \\operatorname {S S I M} \\left(\\mathbf {I} _ {i, t}, \\hat {\\mathbf {I}} _ {i, t}\\right), \\tag {1}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.689, + 0.788, + 0.779 + ], + "angle": 0, + "content": "where \\(\\hat{\\mathbf{I}}_{i,t}\\) denotes the image rendered from the perspective of the \\(i\\)-th camera at time \\(t\\). Note that since the Gaussians' position and rotation are bound to the mesh, the optimization of the mesh vertices will implicitly optimize the parameters of the Gaussians. At the first time step, we optimize color, scale, and opacity. In the following frames \\(1 < t \\leq T\\), we fix the opacity and scale, and optimize Gaussian colors \\(\\mathbf{c}_t\\) and mesh vertices \\(\\mathbf{V}_t\\)." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.78, + 0.788, + 0.841 + ], + "angle": 0, + "content": "In addition to the photometric loss, we apply the following regularization terms to preserve the local geometric features: 1) an isometry loss \\(\\mathcal{L}_{\\mathrm{iso}}\\) [65] that preserves edge length through the training sequence, 2) a normal loss \\(\\mathcal{L}_{\\mathrm{normal}}\\) that encourages smooth mesh surfaces, enhancing mesh deformation accuracy" + } + ], + [ + { + "type": "header", + "bbox": [ + 0.652, + 0.115, + 0.732, + 0.128 + ], + "angle": 0, + "content": "PhysAvatar" + }, + { + "type": "page_number", + "bbox": [ + 0.775, + 0.116, + 0.785, + 0.127 + ], + "angle": 0, + "content": "7" + }, + { + "type": "image", + "bbox": [ + 0.26, + 0.146, + 0.364, + 0.239 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.299, + 0.241, + 0.323, + 0.25 + ], + "angle": 0, + "content": "t=12" + }, + { + "type": "image", + "bbox": [ + 0.389, + 0.159, + 0.432, + 0.239 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.445, + 0.164, + 0.48, + 0.239 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.422, + 0.241, + 0.45, + 0.25 + ], + "angle": 0, + "content": "t=300" + }, + { + "type": "image", + "bbox": [ + 0.51, + 0.16, + 0.627, + 0.239 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.55, + 0.241, + 0.579, + 0.25 + ], + "angle": 0, + "content": "t=600" + }, + { + "type": "image", + "bbox": [ + 0.644, + 0.159, + 0.743, + 0.239 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.676, + 0.241, + 0.706, + 0.25 + ], + "angle": 0, + "content": "t=900" + }, + { + "type": "image_caption", + "bbox": [ + 0.215, + 0.252, + 0.785, + 0.294 + ], + "angle": 0, + "content": "Fig. 3: Our method can robustly track a dynamic mesh from input images, providing accurate long-term correspondences. Here we show the rendered images overlaid with the Gaussian trajectories from the previous 12 frames and the optimized meshes." + }, + { + "type": "text", + "bbox": [ + 0.215, + 0.305, + 0.784, + 0.351 + ], + "angle": 0, + "content": "and stability. Denoting the vertex and the triangle of the mesh at frame \\(t\\) as \\(\\mathbf{v}_t\\) and \\(\\mathbf{f}_t\\), the normal of a face as \\(\\mathbf{n}(\\cdot)\\), and the 1-ring neighbors \\(N(\\cdot)\\), the losses are formally defined as:" + }, + { + "type": "equation", + "bbox": [ + 0.328, + 0.36, + 0.785, + 0.4 + ], + "angle": 0, + "content": "\\[\n\\mathcal {L} _ {\\mathrm {i s o}} = \\frac {1}{| \\mathbf {V} |} \\sum_ {\\mathbf {v} \\in \\mathbf {V}} \\sum_ {\\mathbf {v} ^ {\\prime} \\in N (\\mathbf {v} _ {1})} \\frac {\\left(\\left\\| \\mathbf {v} _ {1} - \\mathbf {v} _ {1} ^ {\\prime} \\right\\| _ {2} - \\left\\| \\mathbf {v} _ {t} - \\mathbf {v} _ {t} ^ {\\prime} \\right\\| _ {2}\\right)}{| N (\\mathbf {v}) |}, \\tag {2}\n\\]" + }, + { + "type": "equation", + "bbox": [ + 0.306, + 0.403, + 0.785, + 0.441 + ], + "angle": 0, + "content": "\\[\n\\mathcal {L} _ {\\text {n o r m a l}} = \\frac {1}{| \\mathbf {F} |} \\sum_ {\\mathbf {f} _ {t} \\in \\mathbf {F}} \\sum_ {\\mathbf {f} _ {t} ^ {\\prime} \\in N (\\mathbf {f} _ {t})} \\left(1 - \\mathbf {n} \\left(\\mathbf {f} _ {t}\\right) \\cdot \\mathbf {n} \\left(\\mathbf {f} _ {t} ^ {\\prime}\\right)\\right). \\tag {3}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.241, + 0.452, + 0.596, + 0.466 + ], + "angle": 0, + "content": "The full loss used for mesh tracking is defined as:" + }, + { + "type": "equation", + "bbox": [ + 0.365, + 0.48, + 0.785, + 0.495 + ], + "angle": 0, + "content": "\\[\n\\mathcal {L} _ {\\text {m e s h}} = \\mathcal {L} _ {\\text {r e n d e r}} + \\lambda_ {i} \\mathcal {L} _ {\\text {i s o}} + \\lambda_ {n} \\mathcal {L} _ {\\text {n o r m a l}}, \\tag {4}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.215, + 0.506, + 0.784, + 0.552 + ], + "angle": 0, + "content": "where we set \\(\\lambda_{i} = 10\\) and \\(\\lambda_{n} = 0.1\\) during our training. With the introduced losses, our method can reconstruct a smooth mesh sequence with accurate correspondences, as shown in Fig. 3." + }, + { + "type": "title", + "bbox": [ + 0.215, + 0.574, + 0.542, + 0.59 + ], + "angle": 0, + "content": "3.2 Physics Based Dynamic Modeling" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.599, + 0.785, + 0.658 + ], + "angle": 0, + "content": "Given the tracked meshes, \\((\\mathbf{V}_{1:T},\\mathbf{F})\\), we estimate the physical properties of the garments such that the simulation of their dynamics matches the reference tracked meshes. Once learned, these parameters can be used to simulate the garment under novel motions and novel views." + }, + { + "type": "text", + "bbox": [ + 0.215, + 0.659, + 0.786, + 0.841 + ], + "angle": 0, + "content": "Garment Material Model and Simulation. We model the elastodynamics of the garment by equipping each simulation mesh with the discrete shell hinge bending energy [29,102] and isotropic StVK elastic potential [15,20]. Specifically, we estimate density \\(\\rho\\), membrane stiffness \\(\\kappa_{s}\\), and bending stiffness \\(\\kappa_{b}\\) of the garment given the visual observations. For our generic use cases, it is reasonable to assume homogeneous physical properties for each garment piece, i.e., the density, membrane stiffness, and bending stiffness can be parameterized using scalar values, denoted as \\(\\rho\\), \\(\\kappa_{s}\\), and \\(\\kappa_{b}\\). In choosing the simulators, we consider differentiability and quality. Unfortunately, the development of differentiable simulators suitable for our application remains an open research question. Current implementations, such as those documented in [56], cannot be directly applied due to limitations in modeling complex body colliders. Therefore, we propose to use" + } + ], + [ + { + "type": "page_number", + "bbox": [ + 0.218, + 0.116, + 0.23, + 0.127 + ], + "angle": 0, + "content": "8" + }, + { + "type": "header", + "bbox": [ + 0.272, + 0.115, + 0.393, + 0.129 + ], + "angle": 0, + "content": "Yang Zheng et al." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.147, + 0.788, + 0.268 + ], + "angle": 0, + "content": "the state-of-the-art robust and stable cloth simulator C-IPC [52] that guarantees no interpenetration with frictional contact. While C-IPC is differentiable, it lacks an implementation of accurate analytical gradients; hence, we use finite difference methods to estimate the gradients for parameter updates. Specifically, we employ an open-source implementation of C-IPC [52] as our cloth simulator, which is a state-of-the-art simulator achieving high-fidelity garment simulation with stability and convergence (see [52] for a comprehensive review of the base algorithm and implementation)." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.269, + 0.807, + 0.451 + ], + "angle": 0, + "content": "Garment Material Estimation. From the reconstructed mesh, we segment out the garment vertices to obtain ground truth garment mesh sequences \\((\\widehat{\\mathbf{V}}_{1:T}^{g},\\widehat{\\mathbf{F}}^{g})\\). The cloth simulator, \\(f(\\cdot)\\), takes as input the current garment state—position and velocity \\((\\mathbf{V}_t^g,\\dot{\\mathbf{V}}_t^g)\\)—the garment's physical properties \\((\\rho ,\\kappa_s,\\kappa_b)\\), the body collider \\(\\{\\mathbf{V}_t^C,\\mathbf{F}^C\\}\\), and garment boundary vertices \\(\\mathbf{V}_{t + 1}^b\\), and step size \\(\\Delta t\\). The boundary vertices of the garment are points of the garment that are in permanent contact with the human body, such as the shoulder region for draped garments. We employ the SMPL-X body model [64] as the body collider \\(\\{\\mathbf{V}_{t + 1}^{C},\\mathbf{F}^{C}\\}\\), extracting SMPL-X body shapes and poses from the video data during the preprocessing (see Sec. 4). The garment and the boundary vertices are manually annotated once on the initial mesh \\(\\mathbf{V}_1\\) for the entire simulation. The simulation function predicting the garment shape in the next frame can be written as below:" + }, + { + "type": "equation", + "bbox": [ + 0.347, + 0.465, + 0.786, + 0.484 + ], + "angle": 0, + "content": "\\[\n\\mathbf {V} _ {t + 1} ^ {g} = f (\\mathbf {V} _ {t} ^ {g}, \\dot {\\mathbf {V}} _ {t} ^ {g}, \\mathbf {V} _ {t + 1} ^ {b}, \\mathbf {V} _ {t + 1} ^ {C}, \\rho , \\kappa_ {s}, \\kappa_ {b}, \\Delta t). \\tag {5}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.215, + 0.489, + 0.788, + 0.536 + ], + "angle": 0, + "content": "We find the optimal physical parameters by minimizing the difference between the simulated and the reconstructed garment meshes. Formally, the optimization objective can be expressed as:" + }, + { + "type": "equation", + "bbox": [ + 0.333, + 0.548, + 0.786, + 0.588 + ], + "angle": 0, + "content": "\\[\n\\min _ {\\left\\{\\rho , \\kappa_ {s}, \\kappa_ {b} \\right\\}} \\mathcal {L} _ {\\text {s i m}} (\\rho , \\kappa_ {s}, \\kappa_ {b}) = \\sum_ {t = 0} ^ {T} \\| \\mathbf {V} _ {t + 1} ^ {g} - \\widehat {\\mathbf {V}} _ {t + 1} ^ {g} \\| _ {2} ^ {2}, \\tag {6}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.215, + 0.602, + 0.788, + 0.679 + ], + "angle": 0, + "content": "where \\(\\mathbf{V}_1^g = \\widehat{\\mathbf{V}}_1^g\\) and \\(\\mathbf{V}_{t + 1}^g\\) is computed by Eq. (5). We estimate gradients via a numerical approach using finite differences [50]. This method is particularly feasible in our context due to the low-dimensional nature of the optimization problem, which involves only three scalar parameters. We employ the forward difference formula for gradient estimation:" + }, + { + "type": "equation", + "bbox": [ + 0.321, + 0.69, + 0.787, + 0.723 + ], + "angle": 0, + "content": "\\[\n\\frac {\\partial \\mathcal {L} _ {\\mathrm {s i m}}}{\\partial \\rho} \\approx (\\mathcal {L} _ {\\mathrm {s i m}} (\\rho + \\delta_ {\\rho}, \\kappa_ {s}, \\kappa_ {b}) - \\mathcal {L} _ {\\mathrm {s i m}} (\\rho , \\kappa_ {s}, \\kappa_ {b})) / \\delta_ {\\rho}. \\tag {7}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.732, + 0.786, + 0.778 + ], + "angle": 0, + "content": "Where \\(\\delta_{\\rho}\\) is a hyperparameter for forward difference gradient estimation. The gradients for other parameters are estimated using the same approach, and the parameters are subsequently updated using the Adam optimizer." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.779, + 0.788, + 0.841 + ], + "angle": 0, + "content": "Animation. At test time, we have the initial mesh \\(\\{\\mathbf{V}_1,\\mathbf{F}_1\\}\\) and we are given a sequence of novel human body poses, e.g., provided by skeleton parameters such as SMPL [64]. Using standard LBS, we obtain the boundary locations \\(\\mathbf{V}_{t + 1}^b\\). We assume the initial garment is at rest, i.e., \\(\\dot{\\mathbf{V}}_0 = 0\\). We use the same garment" + } + ], + [ + { + "type": "header", + "bbox": [ + 0.653, + 0.115, + 0.732, + 0.128 + ], + "angle": 0, + "content": "PhysAvatar" + }, + { + "type": "page_number", + "bbox": [ + 0.776, + 0.117, + 0.786, + 0.127 + ], + "angle": 0, + "content": "9" + }, + { + "type": "image_caption", + "bbox": [ + 0.255, + 0.168, + 0.272, + 0.178 + ], + "angle": 0, + "content": "(a)" + }, + { + "type": "image", + "bbox": [ + 0.258, + 0.182, + 0.38, + 0.277 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.266, + 0.284, + 0.364, + 0.296 + ], + "angle": 0, + "content": "Before Optimization" + }, + { + "type": "image", + "bbox": [ + 0.384, + 0.183, + 0.508, + 0.277 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.402, + 0.284, + 0.491, + 0.296 + ], + "angle": 0, + "content": "After Optimization" + }, + { + "type": "image_caption", + "bbox": [ + 0.509, + 0.168, + 0.525, + 0.179 + ], + "angle": 0, + "content": "(b)" + }, + { + "type": "image", + "bbox": [ + 0.532, + 0.164, + 0.588, + 0.28 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.522, + 0.284, + 0.587, + 0.296 + ], + "angle": 0, + "content": "PBR w/o Opt" + }, + { + "type": "image", + "bbox": [ + 0.607, + 0.164, + 0.665, + 0.28 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.607, + 0.284, + 0.658, + 0.295 + ], + "angle": 0, + "content": "PBR Ours" + }, + { + "type": "image", + "bbox": [ + 0.683, + 0.164, + 0.741, + 0.28 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.697, + 0.284, + 0.715, + 0.295 + ], + "angle": 0, + "content": "GT" + }, + { + "type": "image_caption", + "bbox": [ + 0.214, + 0.324, + 0.785, + 0.394 + ], + "angle": 0, + "content": "Fig. 4: Ablation study on appearance estimation: (a) Initial texture map \\(\\mathbf{T}\\) extracted from Gaussian splatting (left) has baked-in shadows highlighted in the red boxes; post-optimization (right), the baked-in shadows are substantially removed. (b) Rendering comparisons demonstrate that our method with the optimized texture map more closely aligns with the ground truth." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.423, + 0.784, + 0.469 + ], + "angle": 0, + "content": "segmentation, boundary points, and SMPL body collider points as during training, and use Eq. (5) to simulate the deformation of the garment. This is then combined with the body, driven by LBS, to create the holistic animation." + }, + { + "type": "title", + "bbox": [ + 0.215, + 0.49, + 0.566, + 0.506 + ], + "angle": 0, + "content": "3.3 Physics Based Appearance Modeling" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.515, + 0.788, + 0.71 + ], + "angle": 0, + "content": "While tracking with dynamic Gaussians jointly estimates the garments appearance, it does not account for light-surface interactions, consequently the reconstructed appearance contains baked-in shadows, causing uncanny visual artifacts when rendered in novel views and motions (see Fig. 4). In this step, given the mesh sequence \\((\\mathbf{V}_{1:T},\\mathbf{F})\\) and the per-frame mesh colors \\(\\mathbf{c}_{1:T}\\) reconstructed from Section 3.1, we utilize Mitsuba3 [78]—a physics-based differentiable Monte Carlo renderer—to enhance the appearance reconstruction. For this purpose, we use a time-invariant diffuse texture map \\(\\mathbf{T}\\) to model the appearance. The value \\(\\mathbf{T}\\) is initialized from the average Gaussian colors from the tracking step (Section 3.1), \\(\\Sigma_{1 < t\\leq T}(\\mathbf{c}(\\mathbf{x})_t) / T\\). For the environment lighting, since the capture studio is typically equipped with uniform point light arrays, we can approximate the environment lighting as an ambient illumination, parameterized with a global \\(L_{a}\\in \\mathbb{R}^{3}\\)." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.711, + 0.786, + 0.757 + ], + "angle": 0, + "content": "We optimize for \\(\\mathbf{T}\\) and \\(L_{a}\\) using the photometric loss defined in Eq. (8) for all training views and frames \\((i,t)\\), where \\(\\mathcal{R}\\) denotes the rendering function and \\(c_{i,t}\\) the camera pose:" + }, + { + "type": "equation", + "bbox": [ + 0.336, + 0.768, + 0.786, + 0.801 + ], + "angle": 0, + "content": "\\[\n\\mathcal {L} _ {\\mathrm {r g b}} = \\min _ {\\left\\{\\mathbf {T}, L _ {a} \\right\\}} \\sum_ {i, t} \\left\\| \\mathbf {I} _ {i, t} - \\mathcal {R} \\left(\\mathbf {T}, L _ {a}, \\mathbf {V} _ {t}, c _ {i, t}\\right) \\right\\| _ {2} ^ {2}. \\tag {8}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.215, + 0.81, + 0.784, + 0.842 + ], + "angle": 0, + "content": "At test time, we can swap the ambient light with more complex illuminations, and produce realistic rendering in arbitrary views and body poses (see Sec. 4.4)." + } + ], + [ + { + "type": "page_number", + "bbox": [ + 0.218, + 0.116, + 0.236, + 0.127 + ], + "angle": 0, + "content": "10" + }, + { + "type": "header", + "bbox": [ + 0.272, + 0.115, + 0.392, + 0.129 + ], + "angle": 0, + "content": "Yang Zheng et al." + }, + { + "type": "image", + "bbox": [ + 0.248, + 0.144, + 0.761, + 0.389 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.215, + 0.39, + 0.788, + 0.432 + ], + "angle": 0, + "content": "Fig. 5: Qualitative results on test poses from the ActorHQ [38] dataset. Our method PhysAvatar achieves state-of-the-art performance in terms of geometry detail and appearance modeling." + }, + { + "type": "title", + "bbox": [ + 0.216, + 0.46, + 0.376, + 0.478 + ], + "angle": 0, + "content": "4 Experiments" + }, + { + "type": "text", + "bbox": [ + 0.215, + 0.491, + 0.787, + 0.522 + ], + "angle": 0, + "content": "In this section, we explain our experimental setup and results. We recommend watching the supplementary video for better visualization." + }, + { + "type": "title", + "bbox": [ + 0.216, + 0.544, + 0.424, + 0.56 + ], + "angle": 0, + "content": "4.1 Experimental setup" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.568, + 0.788, + 0.765 + ], + "angle": 0, + "content": "Data Preparation. We choose four characters from the ActorHQ [35] dataset for our experiments, including two female characters with loose dresses, one female character wearing a tank top and loose-fitting pants, and one male character with a short-sleeved shirt and khaki shorts, as shown in Fig. 1. For each character, we use the ground truth mesh provided by the dataset and re-mesh it to obtain a simulation-ready initial mesh, i.e., one where the triangles have similar areas, to enhance the performance of garment simulation. We then segment the garment, define the boundary vertices of the garment for simulation purposes, and unwrap the whole mesh to get the UV map, all performed in Blender [13]. Since we observe that the human pose estimations from the dataset are quite noisy, we fit the SMPL-X [85] body mesh using multi-view videos by minimizing the distance between the projected SMPL-X joints and 2D keypoint detection results from DWPose [21, 121]." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.766, + 0.788, + 0.841 + ], + "angle": 0, + "content": "Implementation Details. For training, we utilize the entire set of 160 camera views from the ActorHQ [35] dataset. Although the dataset includes approximately 2000 frames for each character, we selectively use 24 frames with large movements for estimating garment material parameters and 200 frames for optimizing the texture map, which leads to good results while conserving compu" + } + ], + [ + { + "type": "header", + "bbox": [ + 0.652, + 0.115, + 0.732, + 0.129 + ], + "angle": 0, + "content": "PhysAvatar" + }, + { + "type": "page_number", + "bbox": [ + 0.769, + 0.116, + 0.785, + 0.127 + ], + "angle": 0, + "content": "11" + }, + { + "type": "table_caption", + "bbox": [ + 0.215, + 0.145, + 0.788, + 0.188 + ], + "angle": 0, + "content": "Table 1: Quantitative comparison with baselines. PhysAvatar outperforms all baselines with respect to geometry accuracy. We achieve the best or competitive results in appearance metrics. See Fig. 5 and Fig. 6 for qualitative comparisons." + }, + { + "type": "table", + "bbox": [ + 0.241, + 0.199, + 0.76, + 0.3 + ], + "angle": 0, + "content": "
MethodGeometryAppearance
CD (×103) (↓)F-Score (↑)LPIPS (↓)PSNR (↑)SSIM (↑)
RAAH [115]1.1286.10.05528.60.957
TAVA [53]0.6692.30.05129.60.962
GS-Avatar [34]0.9189.40.04430.60.962
PhysAvatar (ours)0.5592.90.03530.20.957
" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.33, + 0.788, + 0.435 + ], + "angle": 0, + "content": "tational resources. For evaluation, we choose 200 frames for each character, with novel motions unseen during training. At test time, given a motion sequence defined by SMPL-X [85], we use an LBS weight inpainting algorithm [2] to obtain the skinning weights of the human body with respect to the SMPL-X skeleton. We animate the human body using LBS and simulate the garment, which is driven by the boundary vertices and utilizes the SMPL-X body mesh as the collider. More details are available at supplementary document." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.436, + 0.79, + 0.543 + ], + "angle": 0, + "content": "Baselines. We benchmark our method and current open-sourced state-of-the-art methods modeling full-body avatars, including ARAH [115] which leverages SDF-based volume rendering to learn the canonical representation of the avatar, TAVA [53] which learns the canonical radiance field [75] and skinning weight of the character, and GS-Avatar [34] based on a pose-conditioned neural network which predicts the parameters of 3D Gaussians [40]. For all baselines, we use the official public codebase and extend them to our dense-view setting." + }, + { + "type": "title", + "bbox": [ + 0.216, + 0.569, + 0.364, + 0.584 + ], + "angle": 0, + "content": "4.2 Comparison" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.598, + 0.789, + 0.642 + ], + "angle": 0, + "content": "We compare PhysAvatar against current state-of-the-art approaches in terms of geometry and appearance quality across test pose sequences, as shown in Table 1, Fig. 5, and Fig. 6." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.644, + 0.789, + 0.841 + ], + "angle": 0, + "content": "Geometry Evaluation. For geometry evaluation, we employ the mean Chamfer distance [73] and the mean F-Score [104] at \\(\\tau = 0.001\\) as quantitative metrics. PhysAvatar outperforms all baseline according to these metrics, demonstrating a superior capability in accurately capturing the dynamic nature of garment geometry over time, as shown in Table 1. As shown in Fig. 5, PhysAvatar generates more accurate garment deformation and better captures fine wrinkle details, thanks to the integration of a physics-based simulator. This contrasts with all baseline methods [34,53,115], which do not explicitly model garment dynamics, but instead depend on learning quasi-static, pose-conditioned skinning weights or deformations. Additionally, we show results using motion data from the AMASS dataset [72]. PhysAvatar demonstrates its robustness by consistently synthesizing realistic garment deformation with fine details, while baseline methods sometimes fail to capture large deformations in loose garments (Fig. 6)." + } + ], + [ + { + "type": "page_number", + "bbox": [ + 0.218, + 0.116, + 0.236, + 0.127 + ], + "angle": 0, + "content": "12" + }, + { + "type": "header", + "bbox": [ + 0.272, + 0.115, + 0.392, + 0.129 + ], + "angle": 0, + "content": "Yang Zheng et al." + }, + { + "type": "image", + "bbox": [ + 0.26, + 0.146, + 0.495, + 0.253 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.506, + 0.145, + 0.744, + 0.253 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.26, + 0.256, + 0.495, + 0.36 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.506, + 0.256, + 0.744, + 0.36 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.214, + 0.365, + 0.788, + 0.409 + ], + "angle": 0, + "content": "Fig. 6: Here we show animation results of current state-of-the-art methods including ARAH [115], TAVA [53] and GS-Avatar [34], and our results on test motions from AMASS [72] dataset. Images are rendered from novel views. Please zoom in for details." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.435, + 0.789, + 0.588 + ], + "angle": 0, + "content": "Appearance Evaluation. For appearance evaluation, we utilize Peak Signal-to-Noise Ratio (PSNR), Structural Similarity Index Measure (SSIM), and the Learned Perceptual Image Patch Similarity (LPIPS) [129]. As detailed in Table 1, PhysAvatar surpasses all baselines in LPIPS, a metric shown to be better aligned with human perception [129]. While we achieve the second-best results in PSNR and SSIM, it is important to note that even minor misalignments in images can significantly impact these metrics, potentially not fully reflecting the quality of the renderings. As demonstrated in Fig. 5, PhysAvatar can capture more high-frequency details in facial features and garment textures, thereby creating images that are visually richer and more detailed, compared to the baseline approaches." + }, + { + "type": "title", + "bbox": [ + 0.216, + 0.607, + 0.336, + 0.62 + ], + "angle": 0, + "content": "4.3 Ablation" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.629, + 0.787, + 0.674 + ], + "angle": 0, + "content": "We perform ablation studies to assess the impact of each component in our pipeline on the overall performance. For this, we focus on two female characters from the ActorHQ dataset, specifically chosen for their loose garments." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.675, + 0.789, + 0.841 + ], + "angle": 0, + "content": "Garment physics estimation and modeling. We ablate our approach against two variants: one employing Linear Blend Skinning (LBS) for garment deformation and another utilizing a physics simulator for garment dynamics without optimized parameters, henceforth referred to as 'w/o physics', which employs random garment parameters. As shown in Table 2, both alternatives exhibit worse performance across all geometric metrics. Further, as shown in Fig. 7, LBS fails to produce plausible deformations for loose garments, such as dresses. While simulation with random parameters yields plausible garment dynamics, they do not accurately reflect real-world observations. In contrast, our method, with learned physics parameters, achieves deformations that closely align with the reference, demonstrating its effectiveness." + } + ], + [ + { + "type": "header", + "bbox": [ + 0.652, + 0.115, + 0.732, + 0.128 + ], + "angle": 0, + "content": "PhysAvatar" + }, + { + "type": "page_number", + "bbox": [ + 0.769, + 0.116, + 0.786, + 0.127 + ], + "angle": 0, + "content": "13" + }, + { + "type": "image", + "bbox": [ + 0.231, + 0.144, + 0.356, + 0.243 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.274, + 0.246, + 0.304, + 0.258 + ], + "angle": 0, + "content": "LBS" + }, + { + "type": "image", + "bbox": [ + 0.368, + 0.144, + 0.495, + 0.243 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.389, + 0.246, + 0.464, + 0.26 + ], + "angle": 0, + "content": "w/o Physics" + }, + { + "type": "image", + "bbox": [ + 0.506, + 0.144, + 0.631, + 0.243 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.548, + 0.246, + 0.579, + 0.258 + ], + "angle": 0, + "content": "Ours" + }, + { + "type": "image", + "bbox": [ + 0.647, + 0.145, + 0.774, + 0.243 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.696, + 0.246, + 0.718, + 0.258 + ], + "angle": 0, + "content": "GT" + }, + { + "type": "image_caption", + "bbox": [ + 0.214, + 0.267, + 0.785, + 0.337 + ], + "angle": 0, + "content": "Fig. 7: Ablation Study on garment physics estimation: LBS-based garment deformation (LBS) fails to generate realistic garment deformation. Physics-based simulation with random parameters (w/o physics) achieves realistic deformation but falls short of matching the ground-truth reference (GT). In contrast, PhysAvatar accurately captures garment dynamics, closely aligning with the reference." + }, + { + "type": "table", + "bbox": [ + 0.233, + 0.351, + 0.77, + 0.449 + ], + "angle": 0, + "content": "
MethodGeometryAppearance
CD (×10-3) (↓)F-Score (↑)LPIPS (↓)PSNR (↑)SSIM (↑)
LBS0.7191.60.035430.270.951
w/o physics0.6192.20.034730.260.951
w/o appearance——0.040228.260.947
PhysAvatar (ours)0.5693.00.034330.290.951
" + }, + { + "type": "table_caption", + "bbox": [ + 0.214, + 0.451, + 0.785, + 0.52 + ], + "angle": 0, + "content": "Table 2: Ablation Study: PhysAvatar demonstrates improved geometric accuracy compared to both LBS-based animation (LBS) and simulations with random garment parameters (w/o physics), also shown in Fig. 7. PhysAvatar exhibits superior appearance metrics compared to the setting without physics-based appearance estimation (w/o appearance) (see Fig. 4)." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.563, + 0.788, + 0.699 + ], + "angle": 0, + "content": "Appearance estimation and modeling. We compare rendering outcomes with those obtained without appearance optimization through physics-based inverse rendering [78] (denoted as 'w/o appearance'), where the texture is directly extracted from the Gaussian framework, as detailed in section 3.3. Fig. 4 illustrates that, since the Gaussian framework does not model complex light-surface interactions, shadows are inherently baked into the texture. However, post-optimization, we note a significant removal of these baked shadows. Quantitatively, we see improvements across all appearance metrics post-optimization, shown in Table 2." + }, + { + "type": "title", + "bbox": [ + 0.215, + 0.723, + 0.36, + 0.739 + ], + "angle": 0, + "content": "4.4 Application" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.75, + 0.788, + 0.841 + ], + "angle": 0, + "content": "PhysAvatar can facilitate a wide range of applications, such as animation, relighting, and redressing, as demonstrated in Fig. 1. Our method can also benefit from other motion databases in addition to motions defined under the regime of SMPL-X [85] or SMPL [64]. Fig. 1 shows our animation results on challenging motions from Mixamo [76], where we use its auto-rigging algorithm to drive the human body, boundary vertices of the garment, and SMPL-X mesh as the" + } + ], + [ + { + "type": "page_number", + "bbox": [ + 0.218, + 0.116, + 0.236, + 0.127 + ], + "angle": 0, + "content": "14" + }, + { + "type": "header", + "bbox": [ + 0.272, + 0.115, + 0.392, + 0.129 + ], + "angle": 0, + "content": "Yang Zheng et al." + }, + { + "type": "image", + "bbox": [ + 0.219, + 0.146, + 0.359, + 0.255 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.259, + 0.259, + 0.458, + 0.272 + ], + "angle": 0, + "content": "(a) Our results are readily exportable" + }, + { + "type": "image", + "bbox": [ + 0.359, + 0.147, + 0.498, + 0.255 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.505, + 0.146, + 0.645, + 0.255 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.645, + 0.147, + 0.784, + 0.255 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.543, + 0.26, + 0.746, + 0.272 + ], + "angle": 0, + "content": "(b) Results with artistic modifications" + }, + { + "type": "image_caption", + "bbox": [ + 0.215, + 0.276, + 0.785, + 0.319 + ], + "angle": 0, + "content": "Fig. 8: Our results can be exported into a traditional computer graphics pipeline, e.g., Blender [13], which enables the avatars performing novel motions to be rendered in unseen environments (a) as well as editing lighting and texture of the garments (b)." + }, + { + "type": "text", + "bbox": [ + 0.215, + 0.346, + 0.788, + 0.438 + ], + "angle": 0, + "content": "collider. Current state-of-the-art approaches cannot handle such settings since they are mainly based on LBS weights from SMPL or SMPL-X. Moreover, while current methods relying on NeRF [75] or 3D Gaussians [40] representation are usually incompatible with legacy graphics pipelines, our results are readily exportable and can be seamlessly integrated with computer graphics software tools (e.g., Blender [13]), enabling post-processing artistic modifications (Fig. 8)." + }, + { + "type": "title", + "bbox": [ + 0.216, + 0.458, + 0.538, + 0.475 + ], + "angle": 0, + "content": "5 Limitations and Future Work" + }, + { + "type": "text", + "bbox": [ + 0.215, + 0.487, + 0.789, + 0.685 + ], + "angle": 0, + "content": "Our method currently depends on manual garment segmentation and mesh UV unwrapping. We are posed to evolve our pipeline towards a more automated system by incorporating neural networks to streamline these data preprocessing tasks. Another limitation of our method is the absence of a perfect underlining human body for garment simulation. Our current use of the SMPL-X [85] model as the body collider introduces potential inaccuracies in collision detection and contact information due to mismatches between the SMPL-X mesh and the actual human body shape. Our method would benefit from using more advanced parametric models that promise a closer approximation to real human body shape. Furthermore, our pipeline leverages dense view captures to model avatars, an approach that, while effective, has its limitations. Adapting our method to a sparse-view setting would enhance the versatility of its applications in environments where capturing dense views is impractical." + }, + { + "type": "title", + "bbox": [ + 0.216, + 0.705, + 0.36, + 0.722 + ], + "angle": 0, + "content": "6 Conclusion" + }, + { + "type": "text", + "bbox": [ + 0.215, + 0.734, + 0.788, + 0.841 + ], + "angle": 0, + "content": "We introduce PhysAvatar, a comprehensive pipeline to model clothed 3D human avatars with state-of-the-art realism. By integrating a robust and efficient mesh tracking method, and a novel inverse rendering paradigm to capture the physics of loose garments and the appearance of the body, we believe our work represents a significant advancement in digital avatar modeling from visual observations, which also paves the way for innovative applications in entertainment, virtual reality, and digital fashion." + } + ], + [ + { + "type": "header", + "bbox": [ + 0.653, + 0.115, + 0.732, + 0.128 + ], + "angle": 0, + "content": "PhysAvatar" + }, + { + "type": "page_number", + "bbox": [ + 0.769, + 0.117, + 0.785, + 0.127 + ], + "angle": 0, + "content": "15" + }, + { + "type": "title", + "bbox": [ + 0.218, + 0.145, + 0.393, + 0.163 + ], + "angle": 0, + "content": "Acknowledgement" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.176, + 0.785, + 0.232 + ], + "angle": 0, + "content": "We would like to thank Jiayi Eris Zhang for the discussions. This material is based on work that is partially funded by an unrestricted gift from Google, Samsung, an SNF Postdoc Mobility fellowship, ARL grant W911NF-21-2-0104, and a Vannevar Bush Faculty Fellowship." + }, + { + "type": "title", + "bbox": [ + 0.218, + 0.255, + 0.322, + 0.271 + ], + "angle": 0, + "content": "References" + }, + { + "type": "ref_text", + "bbox": [ + 0.233, + 0.286, + 0.785, + 0.328 + ], + "angle": 0, + "content": "1. Abdal, R., Yifan, W., Shi, Z., Xu, Y., Po, R., Kuang, Z., Chen, Q., Yeung, D.Y., Wetzstein, G.: Gaussian shell maps for efficient 3d human generation. arXiv preprint arXiv:2311.17857 (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.233, + 0.33, + 0.785, + 0.37 + ], + "angle": 0, + "content": "2. Abdrashitov, R., Raichstat, K., Monsen, J., Hill, D.: Robust skin weights transfer via weight inpainting. In: SIGGRAPH Asia 2023 Technical Communications, pp. 1-4 (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.233, + 0.371, + 0.785, + 0.411 + ], + "angle": 0, + "content": "3. Anguelov, D., Srinivasan, P., Koller, D., Thrun, S., Rodgers, J., Davis, J.: Scape: shape completion and animation of people. In: ACM SIGGRAPH 2005 Papers, pp. 408-416 (2005)" + }, + { + "type": "ref_text", + "bbox": [ + 0.233, + 0.412, + 0.785, + 0.452 + ], + "angle": 0, + "content": "4. Bagautdinov, T., Wu, C., Simon, T., Prada, F., Shiratori, T., Wei, S.E., Xu, W., Sheikh, Y., Saragih, J.: Driving-signal aware full-body avatars. ACM Transactions on Graphics (TOG) 40(4), 1-17 (2021)" + }, + { + "type": "ref_text", + "bbox": [ + 0.233, + 0.453, + 0.785, + 0.48 + ], + "angle": 0, + "content": "5. Baraff, D., Witkin, A.: Large steps in cloth simulation. In: Seminal Graphics Papers: Pushing the Boundaries, Volume 2, pp. 767-778 (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.233, + 0.481, + 0.785, + 0.535 + ], + "angle": 0, + "content": "6. Bashirov, R., Larionov, A., Ustinova, E., Sidorenko, M., Svitov, D., Zakharkin, I., Lempitsky, V.: Morf: Mobile realistic fullbody avatars from a monocular video. In: Proceedings of the IEEE/CVF Winter Conference on Applications of Computer Vision. pp. 3545-3555 (2024)" + }, + { + "type": "ref_text", + "bbox": [ + 0.233, + 0.536, + 0.785, + 0.577 + ], + "angle": 0, + "content": "7. Bergman, A., Kellnhofer, P., Yifan, W., Chan, E., Lindell, D., Wetzstein, G.: Generative neural articulated radiance fields. Advances in Neural Information Processing Systems 35, 1990-19916 (2022)" + }, + { + "type": "ref_text", + "bbox": [ + 0.233, + 0.578, + 0.785, + 0.605 + ], + "angle": 0, + "content": "8. Bertiche, H., Madadi, M., Escalera, S.: Cloth3d: clothed 3d humans. In: European Conference on Computer Vision. pp. 344-359. Springer (2020)" + }, + { + "type": "ref_text", + "bbox": [ + 0.233, + 0.606, + 0.785, + 0.646 + ], + "angle": 0, + "content": "9. Bertiche, H., Madadi, M., Escalera, S.: Pbns: physically based neural simulation for unsupervised garment pose space deformation. ACM Transactions on Graphics (TOG) 40(6), 1-14 (2021)" + }, + { + "type": "ref_text", + "bbox": [ + 0.227, + 0.647, + 0.785, + 0.674 + ], + "angle": 0, + "content": "10. Bertiche, H., Madadi, M., Escalera, S.: Neural cloth simulation. ACM Transactions on Graphics (TOG) 41(6), 1-14 (2022)" + }, + { + "type": "ref_text", + "bbox": [ + 0.227, + 0.675, + 0.785, + 0.715 + ], + "angle": 0, + "content": "11. Bertiche, H., Madadi, M., Tyson, E., Escalera, S.: Deepsd: Automatic deep skinning and pose space deformation for 3d garment animation. In: Proceedings of the IEEE/CVF International Conference on Computer Vision. pp. 5471-5480 (2021)" + }, + { + "type": "ref_text", + "bbox": [ + 0.227, + 0.716, + 0.785, + 0.77 + ], + "angle": 0, + "content": "12. Bhat, K.S., Twigg, C.D., Hodgins, J.K., Khosla, P.K., Popovic, Z., Seitz, S.M.: Estimating cloth simulation parameters from video. In: Proceedings of the 2003 ACM SIGGRAPH/Eurographics symposium on Computer animation. pp. 37-51 (2003)" + }, + { + "type": "ref_text", + "bbox": [ + 0.227, + 0.771, + 0.785, + 0.799 + ], + "angle": 0, + "content": "13. Blender Online Coummunity: Blender, blender Foundation, https://www.blender.org/" + }, + { + "type": "ref_text", + "bbox": [ + 0.227, + 0.799, + 0.785, + 0.84 + ], + "angle": 0, + "content": "14. Chen, B., Shen, Y., Shuai, Q., Zhou, X., Zhou, K., Zheng, Y.: Anidress: Animatable loose-dressed avatar from sparse views using garment rigging model. arXiv preprint arXiv:2401.15348 (2024)" + }, + { + "type": "list", + "bbox": [ + 0.227, + 0.286, + 0.785, + 0.84 + ], + "angle": 0, + "content": null + } + ], + [ + { + "type": "page_number", + "bbox": [ + 0.218, + 0.116, + 0.236, + 0.127 + ], + "angle": 0, + "content": "16" + }, + { + "type": "header", + "bbox": [ + 0.272, + 0.115, + 0.392, + 0.129 + ], + "angle": 0, + "content": "Yang Zheng et al." + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.147, + 0.787, + 0.189 + ], + "angle": 0, + "content": "15. Chen, H.Y., Sastry, A., van Rees, W.M., Vouga, E.: Physical simulation of environmentally induced thin shell deformation. ACM Transactions on Graphics (TOG) 37(4), 1-13 (2018)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.19, + 0.787, + 0.232 + ], + "angle": 0, + "content": "16. Chen, X., Jiang, T., Song, J., Yang, J., Black, M.J., Geiger, A., Hilliges, O.: gdna: Towards generative detailed neural avatars. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 20427-20437 (2022)" + }, + { + "type": "ref_text", + "bbox": [ + 0.227, + 0.232, + 0.787, + 0.286 + ], + "angle": 0, + "content": "17. Chen, X., Zheng, Y., Black, M.J., Hilliges, O., Geiger, A.: Snarf: Differentiable forward skimming for animating non-rigid neural implicit shapes. In: Proceedings of the IEEE/CVF International Conference on Computer Vision. pp. 11594-11604 (2021)" + }, + { + "type": "ref_text", + "bbox": [ + 0.227, + 0.287, + 0.787, + 0.342 + ], + "angle": 0, + "content": "18. Chen, Y., Wang, X., Chen, X., Zhang, Q., Li, X., Guo, Y., Wang, J., Wang, F.: Uv volumes for real-time rendering of editable free-view human performance. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 16621-16631 (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.342, + 0.787, + 0.384 + ], + "angle": 0, + "content": "19. Chentanez, N., Macklin, M., Müller, M., Jeschke, S., Kim, T.Y.: Cloth and skin deformation with a triangle mesh based convolutional neural network. In: Computer Graphics Forum. vol. 39, pp. 123-134. Wiley Online Library (2020)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.384, + 0.787, + 0.425 + ], + "angle": 0, + "content": "20. Clyde, D., Teran, J., Tamstorf, R.: Modeling and data-driven parameter estimation for woven fabrics. In: Proceedings of the ACM SIGGRAPH/Eurographics Symposium on Computer Animation. pp. 1-11 (2017)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.425, + 0.787, + 0.453 + ], + "angle": 0, + "content": "21. Contributors, M.: Openmmlab pose estimation toolbox and benchmark. https://github.com/open-mmlab/mmpose (2020)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.453, + 0.787, + 0.495 + ], + "angle": 0, + "content": "22. De Luigi, L., Li, R., Guillard, B., Salzmann, M., Fua, P.: Drapenet: Garment generation and self-supervised draping. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 1451-1460 (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.495, + 0.787, + 0.55 + ], + "angle": 0, + "content": "23. Dong, Z., Chen, X., Yang, J., Black, M.J., Hilliges, O., Geiger, A.: Ag3d: Learning to generate 3d avatars from 2d image collections. In: Proceedings of the IEEE/CVF International Conference on Computer Vision (ICCV). pp. 14916-14927 (October 2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.55, + 0.787, + 0.592 + ], + "angle": 0, + "content": "24. Feng, Y., Yang, J., Pollefeys, M., Black, M.J., Bolkart, T.: Capturing and animation of body and clothing from monocular video. In: SIGGRAPH Asia 2022 Conference Papers. pp. 1-9 (2022)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.592, + 0.787, + 0.647 + ], + "angle": 0, + "content": "25. Fridovich-Keil, S., Meanti, G., Warburg, F.R., Recht, B., Kanazawa, A.: K-planes: Explicit radiance fields in space, time, and appearance. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 12479-12488 (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.647, + 0.787, + 0.688 + ], + "angle": 0, + "content": "26. Genay, A., Lécuyer, A., Hachet, M.: Being an avatar \"for real\": a survey on virtual embodiment in augmented reality. IEEE Transactions on Visualization and Computer Graphics 28(12), 5071-5090 (2021)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.689, + 0.787, + 0.73 + ], + "angle": 0, + "content": "27. Geng, C., Peng, S., Xu, Z., Bao, H., Zhou, X.: Learning neural volumetric representations of dynamic humans in minutes. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 8759-8770 (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.73, + 0.787, + 0.772 + ], + "angle": 0, + "content": "28. Grigorev, A., Black, M.J., Hilliges, O.: Hood: Hierarchical graphs for generalized modelling of clothing dynamics. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 16965-16974 (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.772, + 0.787, + 0.813 + ], + "angle": 0, + "content": "29. Grinspun, E., Hirani, A.N., Desbrun, M., Schröder, P.: Discrete shells. In: Proceedings of the 2003 ACM SIGGRAPH/Eurographics symposium on Computer animation. pp. 62-67. CiteSeer (2003)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.813, + 0.787, + 0.84 + ], + "angle": 0, + "content": "30. Gundogdu, E., Constantin, V., Seifoddini, A., Dang, M., Salzmann, M., Fua, P.: Garnet: A two-stream network for fast and accurate 3d cloth draping. In:" + }, + { + "type": "list", + "bbox": [ + 0.226, + 0.147, + 0.787, + 0.84 + ], + "angle": 0, + "content": null + } + ], + [ + { + "type": "header", + "bbox": [ + 0.653, + 0.115, + 0.732, + 0.128 + ], + "angle": 0, + "content": "PhysAvatar" + }, + { + "type": "page_number", + "bbox": [ + 0.769, + 0.116, + 0.786, + 0.127 + ], + "angle": 0, + "content": "17" + }, + { + "type": "ref_text", + "bbox": [ + 0.253, + 0.147, + 0.785, + 0.175 + ], + "angle": 0, + "content": "Proceedings of the IEEE/CVF International Conference on Computer Vision. pp. 8739-8748 (2019)" + }, + { + "type": "ref_text", + "bbox": [ + 0.227, + 0.177, + 0.786, + 0.232 + ], + "angle": 0, + "content": "31. Guo, J., Li, J., Narain, R., Park, H.S.: Inverse simulation: Reconstructing dynamic geometry of clothed humans via optimal control. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 14698-14707 (2021)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.233, + 0.787, + 0.274 + ], + "angle": 0, + "content": "32. Habermann, M., Liu, L., Xu, W., Pons-Moll, G., Zollhoefer, M., Theobalt, C.: Hdhumans: A hybrid approach for high-fidelity digital humans. Proceedings of the ACM on Computer Graphics and Interactive Techniques 6(3), 1-23 (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.276, + 0.786, + 0.317 + ], + "angle": 0, + "content": "33. Habermann, M., Liu, L., Xu, W., Zollhoefer, M., Pons-Moll, G., Theobalt, C.: Real-time deep dynamic characters. ACM Transactions on Graphics (ToG) 40(4), 1-16 (2021)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.318, + 0.786, + 0.358 + ], + "angle": 0, + "content": "34. Hu, L., Zhang, H., Zhang, Y., Zhou, B., Liu, B., Zhang, S., Nie, L.: Gaussian-avatar: Towards realistic human avatar modeling from a single video via animatable 3d gaussians. arXiv preprint arXiv:2312.02134 (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.361, + 0.786, + 0.415 + ], + "angle": 0, + "content": "35. Isik, M., Rünz, M., Georgopoulos, M., Khakhulin, T., Starck, J., Agapito, L., Nießner, M.: Humanrf: High-fidelity neural radiance fields for humans in motion. ACM Transactions on Graphics (TOG) 42(4), 1-12 (2023). https://doi.org/10.1145/3592415, https://doi.org/10.1145/3592415" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.417, + 0.786, + 0.457 + ], + "angle": 0, + "content": "36. Jambon, C., Kerbl, B., Kopanas, G., Diolatzis, S., Leimkuhler, T., Drettakis, G.: Nerfshop: Interactive editing of neural radiance fields. Proc. ACM Comput. Graph. Interact. Tech. 6, 1:1-1:21 (2023). https://doi.org/10.1145/3585499" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.459, + 0.786, + 0.5 + ], + "angle": 0, + "content": "37. Jiang, Y., Wang, R., Liu, Z.: A survey of cloth simulation and applications. In: 2008 9th International Conference on Computer-Aided Industrial Design and Conceptual Design. pp. 765-769. IEEE (2008)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.502, + 0.786, + 0.542 + ], + "angle": 0, + "content": "38. Jiang, Y., Shen, Z., Wang, P., Su, Z., Hong, Y., Zhang, Y., Yu, J., Xu, L.: Hifi4g: High-fidelity human performance rendering via compact gaussian splatting. arXiv preprint arXiv:2312.03461 (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.544, + 0.786, + 0.584 + ], + "angle": 0, + "content": "39. Kato, H., Ushiku, Y., Harada, T.: Neural 3d mesh renderer. 2018 IEEE/CVF Conference on Computer Vision and Pattern Recognition pp. 3907-3916 (2017). https://doi.org/10.1109/CVPR.2018.00411" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.586, + 0.786, + 0.613 + ], + "angle": 0, + "content": "40. Kerbl, B., Kopanas, G., Leimkuhler, T., Drettakis, G.: 3d gaussian splatting for real-time radiance field rendering. ACM Transactions on Graphics 42(4) (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.615, + 0.786, + 0.655 + ], + "angle": 0, + "content": "41. Keselman, L., Hebert, M.: Approximate differentiable rendering with algebraic surfaces. In: European Conference on Computer Vision. pp. 596-614. Springer (2022)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.657, + 0.786, + 0.684 + ], + "angle": 0, + "content": "42. Komatsu, K.: Human skin model capable of natural shape variation. The visual computer 3, 265-271 (1988)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.686, + 0.786, + 0.712 + ], + "angle": 0, + "content": "43. Korban, M., Li, X.: A survey on applications of digital human avatars toward virtual co-presence. arXiv preprint arXiv:2201.04168 (2022)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.714, + 0.786, + 0.755 + ], + "angle": 0, + "content": "44. Kwon, Y., Kim, D., Ceylan, D., Fuchs, H.: Neural human performer: Learning generalizable radiance fields for human performance rendering. Advances in Neural Information Processing Systems 34, 24741-24752 (2021)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.757, + 0.786, + 0.797 + ], + "angle": 0, + "content": "45. Kwon, Y., Kim, D., Ceylan, D., Fuchs, H.: Neural image-based avatars: Generalizable radiance fields for human avatar modeling. In: The Eleventh International Conference on Learning Representations (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.799, + 0.786, + 0.84 + ], + "angle": 0, + "content": "46. Kwon, Y., Liu, L., Fuchs, H., Habermann, M., Theobalt, C.: Deliffas: Deformable light fields for fast avatar synthesis. Advances in Neural Information Processing Systems 36 (2023)" + }, + { + "type": "list", + "bbox": [ + 0.226, + 0.147, + 0.787, + 0.84 + ], + "angle": 0, + "content": null + } + ], + [ + { + "type": "page_number", + "bbox": [ + 0.218, + 0.116, + 0.236, + 0.127 + ], + "angle": 0, + "content": "18" + }, + { + "type": "header", + "bbox": [ + 0.272, + 0.115, + 0.392, + 0.129 + ], + "angle": 0, + "content": "Yang Zheng et al." + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.147, + 0.785, + 0.189 + ], + "angle": 0, + "content": "47. Larionov, E., Eckert, M.L., Wolff, K., Stuyck, T.: Estimating cloth elasticity parameters using position-based simulation of compliant constrained dynamics. arXiv preprint arXiv:2212.08790 (2022)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.19, + 0.785, + 0.232 + ], + "angle": 0, + "content": "48. Lassner, C., Zollhofer, M.: Pulsar: Efficient sphere-based neural rendering. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 1440-1449 (2021)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.232, + 0.785, + 0.272 + ], + "angle": 0, + "content": "49. Lee, D., Kang, H., Lee, I.K.: Cloth combo: Modeling inter-cloth interaction for draping multi-layered clothes. ACM Transactions on Graphics (TOG) 42(6), 1-13 (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.273, + 0.785, + 0.3 + ], + "angle": 0, + "content": "50. LeVeque, R.J.: Finite difference methods for ordinary and partial differential equations: steady-state and time-dependent problems. SIAM (2007)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.3, + 0.785, + 0.342 + ], + "angle": 0, + "content": "51. Li, M., Ferguson, Z., Schneider, T., Langlois, T.R., Zorin, D., Panozzo, D., Jiang, C., Kaufman, D.M.: Incremental potential contact: intersection-and inversion-free, large-deformation dynamics. ACM Trans. Graph. 39(4), 49 (2020)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.342, + 0.785, + 0.37 + ], + "angle": 0, + "content": "52. Li, M., Kaufman, D.M., Jiang, C.: Codimensional incremental potential contact. ACM Transactions on Graphics (TOG) 40(4), 1-24 (2021)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.371, + 0.785, + 0.397 + ], + "angle": 0, + "content": "53. Li, R., Tanke, J., Vo, M., Zollhofer, M., Gall, J., Kanazawa, A., Lassner, C.: Tava: Template-free animatable volumetric actors (2022)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.398, + 0.785, + 0.424 + ], + "angle": 0, + "content": "54. Li, X., Li, X.R., Li, Y., Feng, W.: Review of cloth modeling and simulation for virtual fitting. Textile Research Journal 93(7-8), 1699-1711 (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.425, + 0.785, + 0.466 + ], + "angle": 0, + "content": "55. Li, Y., Chen, H.y., Larionov, E., Sarafianos, N., Matusik, W., Stuyck, T.: Difavatar: Simulation-ready garment optimization with differentiable simulation. arXiv preprint arXiv:2311.12194 (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.467, + 0.785, + 0.508 + ], + "angle": 0, + "content": "56. Li, Y., Du, T., Wu, K., Xu, J., Matusik, W.: Diffcloth: Differentiable cloth simulation with dry frictional contact. ACM Transactions on Graphics (TOG) 42(1), 1-20 (2022)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.508, + 0.785, + 0.55 + ], + "angle": 0, + "content": "57. Li, Z., Zheng, Z., Liu, Y., Zhou, B., Liu, Y.: Posevocab: Learning joint-structured pose embeddings for human avatar modeling. arXiv preprint arXiv:2304.13006 (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.55, + 0.785, + 0.591 + ], + "angle": 0, + "content": "58. Li, Z., Zheng, Z., Wang, L., Liu, Y.: Animatable gaussians: Learning pose-dependent gaussian maps for high-fidelity human avatar modeling. arXiv preprint arXiv:2311.16096 (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.591, + 0.785, + 0.633 + ], + "angle": 0, + "content": "59. Liu, L., Habermann, M., Rudnev, V., Sarkar, K., Gu, J., Theobalt, C.: Neural actor: Neural free-view synthesis of human actors with pose control. ACM transactions on graphics (TOG) 40(6), 1-16 (2021)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.633, + 0.785, + 0.675 + ], + "angle": 0, + "content": "60. Liu, L., Xu, W., Zollhoefer, M., Kim, H., Bernard, F., Habermann, M., Wang, W., Theobalt, C.: Neural rendering and reenactment of human actor videos. ACM Transactions on Graphics (TOG) 38(5), 1-14 (2019)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.675, + 0.785, + 0.716 + ], + "angle": 0, + "content": "61. Liu, S., Li, T., Chen, W., Li, H.: Soft rasterizer: A differentiable renderer for image-based 3d reasoning. In: Proceedings of the IEEE/CVF International Conference on Computer Vision. pp. 7708-7717 (2019)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.716, + 0.785, + 0.744 + ], + "angle": 0, + "content": "62. Liu, T., Bouaziz, S., Kavan, L.: Quasi-newton methods for real-time simulation of hyperelastic materials. Acm Transactions on Graphics (TOG) 36(3), 1-16 (2017)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.744, + 0.785, + 0.785 + ], + "angle": 0, + "content": "63. Lombardi, S., Simon, T., Saragih, J., Schwartz, G., Lehrmann, A., Sheikh, Y.: Neural volumes: Learning dynamic renderable volumes from images. arXiv preprint arXiv:1906.07751 (2019)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.785, + 0.785, + 0.813 + ], + "angle": 0, + "content": "64. Loper, M., Mahmood, N., Romero, J., Pons-Moll, G., Black, M.J.: Smpl: A skinned multi-person linear model. ACM Transactions on Graphics 34(6) (2015)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.813, + 0.785, + 0.84 + ], + "angle": 0, + "content": "65. Luiten, J., Kopanas, G., Leibe, B., Ramanan, D.: Dynamic 3d gaussians: Tracking by persistent dynamic view synthesis. arXiv preprint arXiv:2308.09713 (2023)" + }, + { + "type": "list", + "bbox": [ + 0.226, + 0.147, + 0.785, + 0.84 + ], + "angle": 0, + "content": null + } + ], + [ + { + "type": "header", + "bbox": [ + 0.653, + 0.115, + 0.732, + 0.128 + ], + "angle": 0, + "content": "PhysAvatar" + }, + { + "type": "page_number", + "bbox": [ + 0.769, + 0.116, + 0.786, + 0.127 + ], + "angle": 0, + "content": "19" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.147, + 0.786, + 0.202 + ], + "angle": 0, + "content": "66. Ma, Q., Saito, S., Yang, J., Tang, S., Black, M.J.: Scale: Modeling clothed humans with a surface codec of articulated local elements. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 16082-16093 (2021)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.203, + 0.787, + 0.243 + ], + "angle": 0, + "content": "67. Ma, Q., Yang, J., Black, M.J., Tang, S.: Neural point-based shape modeling of humans in challenging clothing. In: 2022 International Conference on 3D Vision (3DV). pp. 679-689. IEEE (2022)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.243, + 0.786, + 0.298 + ], + "angle": 0, + "content": "68. Ma, Q., Yang, J., Ranjan, A., Pujades, S., Pons-Moll, G., Tang, S., Black, M.J.: Learning to dress 3d people in generative clothing. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 6469-6478 (2020)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.298, + 0.786, + 0.338 + ], + "angle": 0, + "content": "69. Ma, Q., Yang, J., Tang, S., Black, M.J.: The power of points for modeling humans in clothing. In: Proceedings of the IEEE/CVF International Conference on Computer Vision. pp. 10974-10984 (2021)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.339, + 0.786, + 0.406 + ], + "angle": 0, + "content": "70. Macklin, M., Müller, M., Chentanez, N.: Xpbd: position-based simulation of compliant constrained dynamics. In: Proceedings of the 9th International Conference on Motion in Games. p. 49-54. MIG '16, Association for Computing Machinery, New York, NY, USA (2016). https://doi.org/10.1145/2994258.2994272, https://doi.org/10.1145/2994258.2994272" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.407, + 0.786, + 0.447 + ], + "angle": 0, + "content": "71. Magnenat, T., Laperrière, R., Thalmann, D.: Joint-dependent local deformations for hand animation and object grasping. Tech. rep., Canadian Inf. Process. Soc (1988)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.447, + 0.786, + 0.473 + ], + "angle": 0, + "content": "72. Mahmood, N., Ghorbani, N., Troje, N.F., Pons-Moll, G., Black, M.J.: AMASS: Archive of motion capture as surface shapes. In: ICCV (2019)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.473, + 0.786, + 0.528 + ], + "angle": 0, + "content": "73. Mescheder, L., Oechsle, M., Niemeyer, M., Nowozin, S., Geiger, A.: Occupancy networks: Learning 3d reconstruction in function space. In: Proceedings of the IEEE/CVF conference on computer vision and pattern recognition. pp. 4460-4470 (2019)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.529, + 0.786, + 0.57 + ], + "angle": 0, + "content": "74. Miguel, E., Bradley, D., Thomaszewski, B., Bickel, B., Matusik, W., Otaduy, M.A., Marschner, S.: Data-driven estimation of cloth simulation models. In: Computer Graphics Forum. vol. 31, pp. 519-528. Wiley Online Library (2012)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.569, + 0.786, + 0.61 + ], + "angle": 0, + "content": "75. Mildenhall, B., Srinivasan, P.P., Tancik, M., Barron, J.T., Ramamoorthi, R., Ng, R.: Nerf: Representing scenes as neural radiance fields for view synthesis. Communications of the ACM 65(1), 99-106 (2021)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.61, + 0.438, + 0.622 + ], + "angle": 0, + "content": "76. Mixamo, www MIXamo.com" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.622, + 0.786, + 0.677 + ], + "angle": 0, + "content": "77. Newcombe, R.A., Izadi, S., Hilliges, O., Molyneaux, D., Kim, D., Davison, A.J., Kohi, P., Shotton, J., Hodges, S., Fitzgibbon, A.: Kinectfusion: Real-time dense surface mapping and tracking. In: 2011 10th IEEE international symposium on mixed and augmented reality. pp. 127-136. IEEE (2011)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.677, + 0.786, + 0.717 + ], + "angle": 0, + "content": "78. Nimier-David, M., Vicini, D., Zeltner, T., Jakob, W.: Mitsuba 2: A retargetable forward and inverse renderer. ACM Transactions on Graphics (TOG) 38(6), 1-17 (2019)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.718, + 0.786, + 0.759 + ], + "angle": 0, + "content": "79. Noguchi, A., Sun, X., Lin, S., Harada, T.: Neural articulated radiance field. In: Proceedings of the IEEE/CVF International Conference on Computer Vision (ICCV). pp. 5762-5772 (October 2021)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.759, + 0.786, + 0.799 + ], + "angle": 0, + "content": "80. Noguchi, A., Sun, X., Lin, S., Harada, T.: Unsupervised learning of efficient geometry-aware neural articulated representations. In: European Conference on Computer Vision. pp. 597-614. Springer (2022)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.799, + 0.786, + 0.84 + ], + "angle": 0, + "content": "81. Pan, X., Mai, J., Jiang, X., Tang, D., Li, J., Shao, T., Zhou, K., Jin, X., Manocha, D.: Predicting loose-fitting garment deformations using bone-driven motion networks. In: ACM SIGGRAPH 2022 Conference Proceedings. pp. 1-10 (2022)" + }, + { + "type": "list", + "bbox": [ + 0.226, + 0.147, + 0.787, + 0.84 + ], + "angle": 0, + "content": null + } + ], + [ + { + "type": "page_number", + "bbox": [ + 0.218, + 0.116, + 0.236, + 0.127 + ], + "angle": 0, + "content": "20" + }, + { + "type": "header", + "bbox": [ + 0.272, + 0.115, + 0.392, + 0.129 + ], + "angle": 0, + "content": "Yang Zheng et al." + }, + { + "type": "ref_text", + "bbox": [ + 0.225, + 0.147, + 0.784, + 0.189 + ], + "angle": 0, + "content": "82. Pang, H., Zhu, H., Kortylewski, A., Theobalt, C., Habermann, M.: Ash: Animatable gaussian splats for efficient and photoreal human rendering. arXiv preprint arXiv:2312.05941 (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.225, + 0.19, + 0.785, + 0.245 + ], + "angle": 0, + "content": "83. Park, J.J., Florence, P., Straub, J., Newcombe, R., Lovegrove, S.: Deepsdf: Learning continuous signed distance functions for shape representation. In: Proceedings of the IEEE/CVF conference on computer vision and pattern recognition. pp. 165-174 (2019)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.246, + 0.785, + 0.3 + ], + "angle": 0, + "content": "84. Park, K., Sinha, U., Hedman, P., Barron, J.T., Bouaziz, S., Goldman, D.B., Martin-Brualla, R., Seitz, S.M.: Hypernerf: a higher-dimensional representation for topologically varying neural radiance fields. ACM Transactions on Graphics (TOG) 40(6), 1-12 (2021)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.301, + 0.785, + 0.357 + ], + "angle": 0, + "content": "85. Pavlakos, G., Choutas, V., Ghorbani, N., Bolkart, T., Osman, A.A.A., Tzionas, D., Black, M.J.: Expressive body capture: 3D hands, face, and body from a single image. In: Proceedings IEEE Conf. on Computer Vision and Pattern Recognition (CVPR). pp. 10975-10985 (2019)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.357, + 0.785, + 0.411 + ], + "angle": 0, + "content": "86. Peng, S., Dong, J., Wang, Q., Zhang, S., Shuai, Q., Zhou, X., Bao, H.: Animatable neural radiance fields for modeling dynamic human bodies. In: Proceedings of the IEEE/CVF International Conference on Computer Vision. pp. 14314-14323 (2021)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.411, + 0.785, + 0.467 + ], + "angle": 0, + "content": "87. Peng, S., Zhang, Y., Xu, Y., Wang, Q., Shuai, Q., Bao, H., Zhou, X.: Neural body: Implicit neural representations with structured latent codes for novel view synthesis of dynamic humans. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 9054-9063 (2021)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.467, + 0.785, + 0.508 + ], + "angle": 0, + "content": "88. Peng, T., Wu, W., Liu, J., Li, L., Miao, J., Hu, X., He, R., Li, L.: Pgn-cloth: Physics-based graph network model for 3d cloth animation. Displays 80, 102534 (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.508, + 0.785, + 0.563 + ], + "angle": 0, + "content": "89. Pumarola, A., Corona, E., Pons-Moll, G., Moreno-Noguer, F.: D-nerf: Neural radiance fields for dynamic scenes. 2021 IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR) pp. 10313-10322 (2020). https://doi.org/10.1109/CVPR46437.2021.01018" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.564, + 0.785, + 0.606 + ], + "angle": 0, + "content": "90. Saito, S., Yang, J., Ma, Q., Black, M.J.: Scanimate: Weakly supervised learning of skinned clothed avatar networks. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 2886-2897 (2021)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.606, + 0.785, + 0.647 + ], + "angle": 0, + "content": "91. Santesteban, I., Otaduy, M.A., Casas, D.: Learning-based animation of clothing for virtual try-on. In: Computer Graphics Forum. vol. 38, pp. 355-366. Wiley Online Library (2019)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.647, + 0.785, + 0.688 + ], + "angle": 0, + "content": "92. Santesteban, I., Otaduy, M.A., Casas, D.: Snug: Self-supervised neural dynamic garments. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 8140-8150 (2022)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.689, + 0.785, + 0.73 + ], + "angle": 0, + "content": "93. Schonberger, J.L., Frahm, J.M.: Structure-from-motion revisited. In: Proceedings of the IEEE conference on computer vision and pattern recognition. pp. 4104-4113 (2016)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.73, + 0.785, + 0.772 + ], + "angle": 0, + "content": "94. Sitzmann, V., Thies, J., Heide, F., Nießner, M., Wetzstein, G., Zollhöfer, M.: Deepvoxels: Learning persistent 3d feature embeddings. In: Proc. Computer Vision and Pattern Recognition (CVPR), IEEE (2019)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.772, + 0.785, + 0.813 + ], + "angle": 0, + "content": "95. Sitzmann, V., Zollhöfer, M., Wetzstein, G.: Scene representation networks: Continuous 3d-structure-aware neural scene representations. In: Advances in Neural Information Processing Systems (2019)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.813, + 0.785, + 0.84 + ], + "angle": 0, + "content": "96. Snavely, N., Seitz, S.M., Szeliski, R.: Photo tourism: exploring photo collections in 3d. In: ACM siggraph 2006 papers, pp. 835-846 (2006)" + }, + { + "type": "list", + "bbox": [ + 0.225, + 0.147, + 0.785, + 0.84 + ], + "angle": 0, + "content": null + } + ], + [ + { + "type": "header", + "bbox": [ + 0.653, + 0.115, + 0.732, + 0.128 + ], + "angle": 0, + "content": "PhysAvatar" + }, + { + "type": "page_number", + "bbox": [ + 0.768, + 0.116, + 0.784, + 0.127 + ], + "angle": 0, + "content": "21" + }, + { + "type": "ref_text", + "bbox": [ + 0.223, + 0.147, + 0.785, + 0.188 + ], + "angle": 0, + "content": "97. Stoll, C., Gall, J., De Aguiar, E., Thrun, S., Theobalt, C.: Video-based reconstruction of animatable human characters. ACM Transactions on Graphics (TOG) 29(6), 1-10 (2010)" + }, + { + "type": "ref_text", + "bbox": [ + 0.225, + 0.189, + 0.785, + 0.229 + ], + "angle": 0, + "content": "98. Stuyck, T., Chen, H.y.: Diffxpbd: Differentiable position-based simulation of compliant constraint dynamics. Proceedings of the ACM on Computer Graphics and Interactive Techniques 6(3), 1-14 (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.231, + 0.785, + 0.27 + ], + "angle": 0, + "content": "99. Su, S.Y., Yu, F., Zollhöfer, M., Rhodin, H.: A-nerf: Articulated neural radiance fields for learning human shape, appearance, and pose. Advances in Neural Information Processing Systems 34, 12278–12291 (2021)" + }, + { + "type": "ref_text", + "bbox": [ + 0.22, + 0.271, + 0.785, + 0.323 + ], + "angle": 0, + "content": "100. Su, Z., Hu, L., Lin, S., Zhang, H., Zhang, S., Thies, J., Liu, Y.: Caphy: Capturing physical properties for animatable human avatars. In: Proceedings of the IEEE/CVF International Conference on Computer Vision. pp. 14150-14160 (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.22, + 0.325, + 0.785, + 0.377 + ], + "angle": 0, + "content": "101. Sun, M., Yang, D., Kou, D., Jiang, Y., Shan, W., Yan, Z., Zhang, L.: Human 3d avatar modeling with implicit neural representation: A brief survey. In: 2022 14th International Conference on Signal Processing Systems (ICSPS). pp. 818-827. IEEE (2022)" + }, + { + "type": "ref_text", + "bbox": [ + 0.22, + 0.379, + 0.785, + 0.405 + ], + "angle": 0, + "content": "102. Tamstorf, R., Grinspun, E.: Discrete bending forces and their jacobians. Graphical models 75(6), 362-370 (2013)" + }, + { + "type": "ref_text", + "bbox": [ + 0.22, + 0.407, + 0.785, + 0.445 + ], + "angle": 0, + "content": "103. Tang, M., Tong, R., Narain, R., Meng, C., Manocha, D.: A gpu-based streaming algorithm for high-resolution cloth simulation. In: Computer Graphics Forum. vol. 32, pp. 21-30. Wiley Online Library (2013)" + }, + { + "type": "ref_text", + "bbox": [ + 0.22, + 0.447, + 0.785, + 0.486 + ], + "angle": 0, + "content": "104. Tatarchenko, M., Richter, S.R., Ranftl, R., Li, Z., Koltun, V., Brox, T.: What do single-view 3d reconstruction networks learn? In: Proceedings of the IEEE/CVF conference on computer vision and pattern recognition. pp. 3405-3414 (2019)" + }, + { + "type": "ref_text", + "bbox": [ + 0.22, + 0.487, + 0.785, + 0.554 + ], + "angle": 0, + "content": "105. Tewari, A., Fried, O., Thies, J., Sitzmann, V., Lombardi, S., Sunkavalli, K., Martin-Brualla, R., Simon, T., Saragih, J.M., Nießner, M., Pandey, R., Fanello, S., Wetzstein, G., Zhu, J.Y., Theobalt, C., Agrawala, M., Shechtman, E., Goldman, D.B., Zollhofer, M.: State of the art on neural rendering. Computer Graphics Forum 39 (2020). https://doi.org/10.1111/cgf.14022" + }, + { + "type": "ref_text", + "bbox": [ + 0.22, + 0.555, + 0.785, + 0.609 + ], + "angle": 0, + "content": "106. Tewari, A., Thies, J., Mildenhall, B., Srinivasan, P., Tretschk, E., Yifan, W., Lassner, C., Sitzmann, V., Martin-Brualla, R., Lombardi, S., et al.: Advances in neural rendering. In: Computer Graphics Forum. vol. 41, pp. 703-735. Wiley Online Library (2022)" + }, + { + "type": "ref_text", + "bbox": [ + 0.22, + 0.61, + 0.785, + 0.663 + ], + "angle": 0, + "content": "107. Tiwari, G., Sarafianos, N., Tung, T., Pons-Moll, G.: Neural-gif: Neural generalized implicit functions for animating people in clothing. In: Proceedings of the IEEE/CVF International Conference on Computer Vision. pp. 11708-11718 (2021)" + }, + { + "type": "ref_text", + "bbox": [ + 0.22, + 0.664, + 0.785, + 0.731 + ], + "angle": 0, + "content": "108. Tretschk, E., Tewari, A., Golyanik, V., Zollhöfer, M., Lassner, C., Theobalt, C.: Non-rigid neural radiance fields: Reconstruction and novel view synthesis of a dynamic scene from monocular video. 2021 IEEE/CVF International Conference on Computer Vision (ICCV) pp. 12939-12950 (2020). https://doi.org/10.1109/ICCV48922.2021.01272" + }, + { + "type": "ref_text", + "bbox": [ + 0.22, + 0.732, + 0.785, + 0.772 + ], + "angle": 0, + "content": "109. Vassilev, T., Spanlang, B., Chrysanthou, Y.: Fast cloth animation on walking avatars. In: Computer Graphics Forum. vol. 20, pp. 260-267. Wiley Online Library (2001)" + }, + { + "type": "ref_text", + "bbox": [ + 0.22, + 0.773, + 0.785, + 0.813 + ], + "angle": 0, + "content": "110. Vidaurre, R., Santesteban, I., Garces, E., Casas, D.: Fully convolutional graph neural networks for parametric virtual try-on. In: Computer Graphics Forum. vol. 39, pp. 145-156. Wiley Online Library (2020)" + }, + { + "type": "ref_text", + "bbox": [ + 0.22, + 0.814, + 0.785, + 0.84 + ], + "angle": 0, + "content": "111. Wang, H.: Gpu-based simulation of cloth wrinkles at submillimeter levels. ACM Transactions on Graphics (TOG) 40(4), 1-14 (2021)" + }, + { + "type": "list", + "bbox": [ + 0.22, + 0.147, + 0.785, + 0.84 + ], + "angle": 0, + "content": null + } + ], + [ + { + "type": "page_number", + "bbox": [ + 0.218, + 0.116, + 0.236, + 0.127 + ], + "angle": 0, + "content": "22" + }, + { + "type": "header", + "bbox": [ + 0.272, + 0.115, + 0.392, + 0.129 + ], + "angle": 0, + "content": "Yang Zheng et al." + }, + { + "type": "ref_text", + "bbox": [ + 0.218, + 0.147, + 0.787, + 0.189 + ], + "angle": 0, + "content": "112. Wang, H., O'Brien, J.F., Ramamoorthi, R.: Data-driven elastic models for cloth: modeling and measurement. ACM transactions on graphics (TOG) 30(4), 1-12 (2011)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.189, + 0.787, + 0.23 + ], + "angle": 0, + "content": "113. Wang, P., Liu, L., Liu, Y., Theobalt, C., Komura, T., Wang, W.: Neus: Learning neural implicit surfaces by volume rendering for multi-view reconstruction. Advances in Neural Information Processing Systems 34, 27171-27183 (2021)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.23, + 0.787, + 0.271 + ], + "angle": 0, + "content": "114. Wang, S., Schwarz, K., Geiger, A., Tang, S.: Arah: Animatable volume rendering of articulated human sdfs. In: European conference on computer vision. pp. 1-19. Springer (2022)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.271, + 0.787, + 0.298 + ], + "angle": 0, + "content": "115. Wang, S., Schwarz, K., Geiger, A., Tang, S.: Arah: Animatable volume rendering of articulated human sdfs. In: European Conference on Computer Vision (2022)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.298, + 0.787, + 0.352 + ], + "angle": 0, + "content": "116. Wang, T., Zhang, B., Zhang, T., Gu, S., Bao, J., Baltrusaitis, T., Shen, J., Chen, D., Wen, F., Chen, Q., et al.: Rodin: A generative model for sculpting 3d digital avatars using diffusion. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 4563-4573 (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.352, + 0.787, + 0.393 + ], + "angle": 0, + "content": "117. Wu, G., Yi, T., Fang, J., Xie, L., Zhang, X., Wei, W., Liu, W., Tian, Q., Wang, X.: 4d gaussian splatting for real-time dynamic scene rendering. ArXiv abs/2310.08528 (2023). https://doi.org/10.48550/arXiv.2310.08528" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.393, + 0.787, + 0.447 + ], + "angle": 0, + "content": "118. Xiang, D., Bagautdinov, T., Stuyck, T., Prada, F., Romero, J., Xu, W., Saito, S., Guo, J., Smith, B., Shiratori, T., et al.: Dressing avatars: Deep photorealistic appearance for physically simulated clothing. ACM Transactions on Graphics (TOG) 41(6), 1-15 (2022)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.447, + 0.787, + 0.488 + ], + "angle": 0, + "content": "119. Xu, Y., Yifan, W., Bergman, A.W., Chai, M., Zhou, B., Wetzstein, G.: Efficient 3d articulated human generation with layered surface volumes. arXiv preprint arXiv:2307.05462 (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.488, + 0.787, + 0.528 + ], + "angle": 0, + "content": "120. Xu, Y., Chen, B., Li, Z., Zhang, H., Wang, L., Zheng, Z., Liu, Y.: Gaussian head avatar: Ultra high-fidelity head avatar via dynamic gaussians. arXiv preprint arXiv:2312.03029 (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.528, + 0.787, + 0.57 + ], + "angle": 0, + "content": "121. Yang, Z., Zeng, A., Yuan, C., Li, Y.: Effective whole-body pose estimation with two-stages distillation. In: Proceedings of the IEEE/CVF International Conference on Computer Vision. pp. 4210-4220 (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.569, + 0.787, + 0.596 + ], + "angle": 0, + "content": "122. Yariv, L., Gu, J., Kasten, Y., Lipman, Y.: Volume rendering of neural implicit surfaces. Advances in Neural Information Processing Systems 34, 4805-4815 (2021)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.596, + 0.787, + 0.637 + ], + "angle": 0, + "content": "123. Yee, K.S., Chen, J.S.: The finite-difference time-domain (fdtd) and the finite-volume time-domain (fvtd) methods in solving maxwell's equations. IEEE Transactions on Antennas and Propagation 45(3), 354-363 (1997)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.637, + 0.787, + 0.678 + ], + "angle": 0, + "content": "124. Yifan, W., Serena, F., Wu, S., Öztireli, C., Sorkine-Hornung, O.: Differentiable surface splatting for point-based geometry processing. ACM Transactions on Graphics (TOG) 38(6), 1-14 (2019)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.678, + 0.787, + 0.733 + ], + "angle": 0, + "content": "125. Yu, T., Zheng, Z., Zhong, Y., Zhao, J., Dai, Q., Pons-Moll, G., Liu, Y.: Simulcap: Single-view human performance capture with cloth simulation. In: Proceedings of the IEEE/CVF conference on computer vision and pattern recognition. pp. 5504-5514 (2019)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.733, + 0.787, + 0.773 + ], + "angle": 0, + "content": "126. Yuan, Y.J., Sun, Y.T., Lai, Y.K., Ma, Y., Jia, R., Gao, L.: Nerf-editing: geometry editing of neural radiance fields. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 18353-18364 (2022)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.773, + 0.787, + 0.8 + ], + "angle": 0, + "content": "127. Zeller, C.: Cloth simulation on thegpu. In: ACM SIGGRAPH 2005 Sketches, pp. 39-es (2005)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.8, + 0.787, + 0.84 + ], + "angle": 0, + "content": "128. Zhang, J.X., Lin, G.W.C., Bode, L., Chen, H.y., Stuyck, T., Larionov, E.: Estimating cloth elasticity parameters from homogenized yarn-level models. arXiv preprint arXiv:2401.15169 (2024)" + }, + { + "type": "list", + "bbox": [ + 0.217, + 0.147, + 0.787, + 0.84 + ], + "angle": 0, + "content": null + } + ], + [ + { + "type": "header", + "bbox": [ + 0.653, + 0.115, + 0.732, + 0.128 + ], + "angle": 0, + "content": "PhysAvatar" + }, + { + "type": "page_number", + "bbox": [ + 0.768, + 0.116, + 0.786, + 0.127 + ], + "angle": 0, + "content": "23" + }, + { + "type": "ref_text", + "bbox": [ + 0.218, + 0.147, + 0.787, + 0.189 + ], + "angle": 0, + "content": "129. Zhang, R., Isola, P., Efros, A.A., Shechtman, E., Wang, O.: The unreasonable effectiveness of deep features as a perceptual metric. In: Proceedings of the IEEE conference on computer vision and pattern recognition. pp. 586-595 (2018)" + }, + { + "type": "ref_text", + "bbox": [ + 0.218, + 0.19, + 0.787, + 0.232 + ], + "angle": 0, + "content": "130. Zhao, F., Jiang, Y., Yao, K., Zhang, J., Wang, L., Dai, H., Zhong, Y., Zhang, Y., Wu, M., Xu, L., et al.: Human performance modeling and rendering via neural animated mesh. ACM Transactions on Graphics (TOG) 41(6), 1-17 (2022)" + }, + { + "type": "ref_text", + "bbox": [ + 0.218, + 0.232, + 0.787, + 0.272 + ], + "angle": 0, + "content": "131. Zheng, Z., Huang, H., Yu, T., Zhang, H., Guo, Y., Liu, Y.: Structured local radiance fields for human avatar modeling. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 15893-15903 (2022)" + }, + { + "type": "ref_text", + "bbox": [ + 0.218, + 0.273, + 0.787, + 0.3 + ], + "angle": 0, + "content": "132. Zielonka, W., Bagautdinov, T., Saito, S., Zollhöfer, M., Thies, J., Romero, J.: Drivable 3d gaussian avatars. arXiv preprint arXiv:2311.08581 (2023)" + }, + { + "type": "list", + "bbox": [ + 0.218, + 0.147, + 0.787, + 0.3 + ], + "angle": 0, + "content": null + } + ] +] \ No newline at end of file diff --git a/2024/PhysAvatar_ Learning the Physics of Dressed 3D Avatars from Visual Observations/958c8811-3fcb-45f8-9957-bab313d0628e_origin.pdf b/2024/PhysAvatar_ Learning the Physics of Dressed 3D Avatars from Visual Observations/958c8811-3fcb-45f8-9957-bab313d0628e_origin.pdf new file mode 100644 index 0000000000000000000000000000000000000000..248dd4907e4be1796b722c4d48af23177e104449 --- /dev/null +++ b/2024/PhysAvatar_ Learning the Physics of Dressed 3D Avatars from Visual Observations/958c8811-3fcb-45f8-9957-bab313d0628e_origin.pdf @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:6a3c317f2272ffe8dbb5f4d97b0503b6e0c39310a67ba925abd897fa53f1de48 +size 7712163 diff --git a/2024/PhysAvatar_ Learning the Physics of Dressed 3D Avatars from Visual Observations/full.md b/2024/PhysAvatar_ Learning the Physics of Dressed 3D Avatars from Visual Observations/full.md new file mode 100644 index 0000000000000000000000000000000000000000..cec66f0048647ab7b6082a2718a0935af6c43cee --- /dev/null +++ b/2024/PhysAvatar_ Learning the Physics of Dressed 3D Avatars from Visual Observations/full.md @@ -0,0 +1,416 @@ +# PhysAvatar: Learning the Physics of Dressed 3D Avatars from Visual Observations + +Yang Zheng $^{1\star}$ , Qingqing Zhao $^{1\star}$ , Guandao Yang $^{1}$ , Wang Yifan $^{1}$ , Donglai Xiang $^{2}$ , Florian Dubost $^{3}$ , Dmitry Lagun $^{3}$ , Thabo Beeler $^{3}$ , Federico Tombari $^{3,4}$ , Leonidas Guibas $^{1}$ , and Gordon Wetzstein $^{1}$ + +1 Stanford University + +2 Carnegie Mellon University + +3 Google + +4 Technical University of Munich + +Abstract. Modeling and rendering photorealistic avatars is of crucial importance in many applications. Existing methods that build a 3D avatar from visual observations, however, struggle to reconstruct clothed humans. We introduce PhysAvatar, a novel framework that combines inverse rendering with inverse physics to automatically estimate the shape and appearance of a human from multi-view video data along with the physical parameters of the fabric of their clothes. For this purpose, we adopt a mesh-aligned 4D Gaussian technique for spatio-temporal mesh tracking as well as a physically based inverse renderer to estimate the intrinsic material properties. PhysAvatar integrates a physics simulator to estimate the physical parameters of the garments using gradient-based optimization in a principled manner. These novel capabilities enable PhysAvatar to create high-quality novel-view renderings of avatars dressed in loose-fitting clothes under motions and lighting conditions not seen in the training data. This marks a significant advancement towards modeling photorealistic digital humans using physically based inverse rendering with physics in the loop. Our project website is at: https://qingqing-zhao.github.io/PhysAvatar. + +Keywords: neural rendering $\cdot$ physics $\cdot$ dynamic modeling $\cdot$ 3D avatar + +# 1 Introduction + +Digital avatars are a vital component in numerous applications, ranging from virtual reality and gaming to telepresence and e-commerce [26,43,101]. A realistic 3D avatar of a person can be readily obtained from visual observations, such as multi-view image [116] and video [6] data. The task of rendering animated 3D avatars from novel viewpoints or when performing unseen motions, however, presents considerable challenges, particularly when the avatar wears loose-fitting garments. Accurately rendering the dynamics of garments in conditions that are not observed in the training data necessitates a holistic approach that not only + +![](images/9f5e9783a5bafabc9c297493ffa1d29405d2f8b9638b10115540c976aeb1567e.jpg) +Fig. 1: PhysAvatar is a novel framework that captures the physics of dressed 3D avatars from visual observations, enabling a wide spectrum of applications, such as (a) animation, (b) relighting, and (c) redressing, with high-fidelity rendering results. + +![](images/ef0b754e528d2e8ffdf74a4562d29eda1bcd79f861f6206cf04ecf94a9e1a3c6.jpg) +(b) + +![](images/3c62b42598198dd839a86e1b1569e25ad87a6b2105516b83a9f9c04136db4314.jpg) + +models the shape and appearance of the person but also the physical behavior of their clothes, including friction and collision. + +Learning 3D scene representations from visual data is the core problem addressed by inverse rendering—a field that has recently shown remarkable progress in estimating the geometry and appearance of static and dynamic scenes from multi-view images and videos [105, 106]. In the context of reconstructing 3D avatars, the ability to explicitly control the motions of the avatar in post processing becomes imperative. Most existing methods for reconstructing animatable avatars drive the animation through an underlying skeleton using linear blending skinning (LBS) [64], which adequately models the dynamics of humans dressed in tight-fitting clothes [1,7,23,80,99,119]. With this approach, garments are treated as a rigidly attached part of the body adhering to piece-wise linear transformations, which results in motions that appear rigid and unconvincing in many cases [79, 120, 132]. Several successful attempts to introduce non-rigid deformations through pose-dependent geometry adjustments, such as wrinkles, have recently been made [18,57,59,60,114], although these methods are prone to overfitting to the body poses and motion sequences observed during training. A key problem is that most existing 3D avatar reconstruction approaches neglect to model the dynamics of loose garments in a physically accurate manner, leading to unrealistic fabric behavior and issues like self-penetration. To our knowledge, only the work of Xiang et al. [118] includes a physics-based approach to inverse rendering of digital humans, but their approach requires a tedious manual search process to find the parameters that model the dynamic behavior of their cloth fabrics with reasonable accuracy. + +Here, we introduce PhysAvatar, a novel approach to 3D avatar reconstruction from multi-view video data. PhysAvatar combines inverse rendering with + +"inverse physics" in a principled manner to not only estimate the shape and appearance of the avatar but also physical parameters, including density and stiffness, of the fabrics modeling their (loose) clothes. Our approach includes several parts: Given the reconstructed 3D mesh of the garment in one frame, we first leverage mesh-aligned 4D Gaussians [117] to track surface points of the garment across all frames of the input video sequence. This process establishes dense correspondences on the 3D garment. These data are used as supervision in an inverse physics step, where the physical parameters of the fabric are optimized using a finite-difference approach [50,123]. Finally, we employ a physically based inverse renderer [78] to jointly estimate ambient lighting and the surface material. By leveraging the refined geometry from our simulation step, the inverse renderer can effectively factor in pose-dependent effects, such as shadows caused by self-occlusion, resulting in accurate appearance reconstruction that enables the avatar to be rendered in novel lighting conditions. + +PhysAvatar offers a comprehensive solution for reconstructing clothed human avatars from multi-view video data to perform novel view and motion synthesis with state-of-the-art realism. The key contributions of our work include: + +1. The introduction of a new inverse rendering paradigm for avatars created from real-world captures that incorporates the physics of loose garments in a principled manner; +2. A pipeline that includes accurate and efficient mesh reconstruction and tracking using 4D Gaussians; automatic optimization of the garments' physical material properties; and accurate appearance estimation using physically based inverse rendering. + +PhysAvatar demonstrates a novel path to learning physical scene properties from visual observations by incorporating inverse rendering with physics constraints. Code will be available on the project website. + +# 2 Related Work + +# 2.1 Scene Reconstruction from Visual Observations + +Reconstructing 3D scenes from visual observations is grounded in classic approaches, such as Structure from Motion (SfM) [77, 93, 96]. Recently, this field has witnessed a paradigm shift towards inverse rendering techniques that estimate the shape and appearance of a scene from visual input [105, 106]. Among the many representations developed in this area are neural (radiance) fields [75, 95], neural volumes, surfaces, and signed distance fields [63, 73, 83, 94, 113, 122] as well as differentiable rasterization techniques for meshes [39, 61] and point clouds [40, 41, 48, 124]. Recent inverse rendering approaches are capable of modeling dynamic scenes [25, 84, 89, 108] and enable sophisticated post-capture deformations [36, 126]. + +# 2.2 Animatable Avatars + +One central requirement for generating full-body digital humans is to enable animations driven by skeleton-level control [42, 71]. To this end, early works on + +avatar reconstruction primarily relied on skinning techniques, which successfully model the dynamics of bodies with minimal or tight clothing [3,64]. Clothing is an integral part of everyday human appearance, captured by the saying "clothes make the (wo)man". Reconstructing their dynamic behavior from visual data, however, remains a challenge. Several different types of representations have been explored for clothed avatars, including meshes [68] with dynamic textures [4, 33, 130], neural surface [16, 90, 107] and radiance fields [14, 18, 24, 27, 44-46, 86, 87, 99, 131], point sets [66, 67, 69], and 3D Gaussians [34, 58, 82, 132]. Many of these works condition the deformation of clothing on the body pose and some predict future dynamics based on a small number of frames [32, 33, 82]. + +We now discuss recent research most closely related to ours. TAVA [53] learns a deformable neural radiance field of humans or animals using a differentiable skinning method like SNARF [17]. ARAH [115] models 3D avatars using SDF representations and proposes an SDF-based volume rendering method to reconstruct the avatar in canonical space. GS-Avatar [34] achieves real-time animation by leveraging a learned 3D Gaussian predictor. Although these methods show reasonable effectiveness on humans wearing tight garments, they struggle with loose garments because their LBS-based deformation module cannot accommodate the complex dynamics of loose clothing. To address this limitation, Xiang et al. [118] incorporate a cloth simulator based on the eXtended Position Based Dynamics (XPBD) formulation [70] into a deep appearance model to reconstruct realistic dynamics and appearance. However, they require tedious manual adjustments of physical parameters to produce reasonable garment motion, and the code is not publicly available. Concurrent with our work, AniDress [14], while using physics-based simulation, only uses the simulation to produce a garment rigging model. This can cause inaccuracy in cloth dynamics modeling and unrealistic motion artifacts. In contrast, we estimate the physical parameters through a principled gradient-based inverse physics approach and accurately recover the shape and (relightable) appearance using a physically based inverse renderer, achieving high-fidelity results on novel motions. + +# 2.3 Physics-Based Simulation + +Physics-based simulation of cloth [5, 62, 88, 103, 109, 111, 125, 127] has been extensively studied and is particularly useful for modeling complicated dynamic effects in interaction with human bodies, such as large deformations, wrinkling, and contact. In this work, we adopt the Codimensional Incremental Potential Contact (C-IPC) solver [52] for its robustness in handling complicated body-cloth collision by the log-barrier penalty function [51]. We refer readers to dedicated surveys [37, 54] for a comprehensive review, and focus on works that solve the inverse parameter estimation problem. One line of work uses specialized devices to simultaneously measure forces and deformation in a controlled setting [47, 74, 112, 128]. Our work falls into the category that estimates physical parameters directly from imagery of garments worn on human subjects [12, 31, 97]. For this purpose, some differentiable simulators have been developed and applied to the optimization problem [55], but only for specific types of solvers, + +![](images/3136adbf142bd66ecf1e005f6a2eae2e84bb73f29c49d10e81dd1a0e28a895e6.jpg) +Fig. 2: Method Overview: (a) PhysAvatar takes multi-view videos and an initial mesh as input. We first perform (b) dynamic mesh tracking (Sec. 3.1). The tacked mesh sequences are then used for (c) garment physics estimation with a physics simulator combined with gradient-based optimization (Sec. 3.2); (d) and appearance estimation through physics-based differentiable rendering (Sec. 3.3). At test time, (e) given a sequence of body poses (f), we simulate garment dynamics with the learned physics parameters and employ physics-based rendering to produce the final images. + +such as Projective Dynamics [56] and XPBD [98]. The heavy runtime cost of high-quality physics-based simulation motivates research in neural cloth simulation. Early works create datasets of garments simulated on human bodies [8,30], and then train the network to predict garment deformation in a supervised manner [19,81,91,110]. Recently, more attention has been paid to the self-supervised formulation that applies the elasticity energy defined in traditional simulation to predict garments, thereby training the network to act like a simulation solver, including both the (quasi-) static [9,11,22,49,92] and the dynamic [10,28] cases. Most of these works focus on making forward predictions that resemble traditional simulators without solving inverse parameter estimation on real-world garment data, except CaPhy [100]. CaPhy estimates the Saint Venant-Kirchhoff (StVK) elasticity parameters of clothing from 4D scans but only predicts pose-dependent clothing deformation in a quasi-static manner. In addition, all the works mentioned above treat simulation as a separate problem without modeling the photorealistic appearance. By comparison, our work builds avatars with physically based appearance and dynamics that are optimized for faithfulness to the real human capture in a holistic manner. + +# 3 Method + +Given multi-view videos as input, our goal is to reconstruct a 3D avatar that can be realistically animated and rendered under novel motions and novel views. The core pipeline (Fig. 2) consists of three primary components: 1. Mesh tracking (Sec. 3.1)—given multi-view videos and a reconstructed mesh at the initial time step as input, we track the deformation of the geometry, which provides the + +geometry reference for the ensuing physics simulation; 2. Physics-based dynamic modeling (Sec. 3.2)—with the tracked meshes as reference, we estimate the garments' physical properties using a physics simulator; the simulator, together with the estimated physical parameters are used to generate novel motion; 3. Physics-based appearance refinement (Sec. 3.3)—given the tracked geometry, we further enhance the appearance estimation through a physics-based differentiable renderer, which considers self-occlusion and thus eliminate artifacts such as baked-in shadows. In the following, we delve into the details of each component. + +# 3.1 Mesh Tracking + +Given multiview videos of $T$ frames and $C$ views, $\{\mathbf{I}_{1:T,1:C}\}$ , and an initial mesh $(\mathbf{V}_1,\mathbf{F})$ , with vertices $\mathbf{V}_1$ and faces $\mathbf{F}$ , that can be obtained via any existing static scene reconstruction methods, our goal is to accurately track the mesh deformation through the video sequence $\mathbf{V}_{1:T}$ , which will be used as reference in the subsequent physics simulation. However accurately tracking mesh deformation from visual observation is very challenging. Inspired by the robust and real-time tracking performance from dynamic 3D Gaussians [65], we use 3D Gaussians as a surrogate representation for mesh tracking. + +We adapt 3D Gaussians to align with the reconstructed mesh surface. Following the definition of dynamic 3D Gaussians, at frame $t$ , each Gaussian is parameterized with position $\mathbf{p}_t \in \mathbb{R}^3$ , rotation quaternion $\mathbf{q}_t \in \mathbb{R}^4$ , color $\mathbf{c}_t \in \mathbb{R}^3$ , scale $\mathbf{s}_t \in \mathbb{R}^3$ and opacity $o_t \in \mathbb{R}_+$ . To couple the Gaussians with the mesh, we attach one Gaussian at the barycenter of each face $\mathbf{f} \in \mathbf{F}$ , and determine the Gaussians' quaternions from the attached triangle. Specifically, the rotation of each Gaussian is computed from the rotation of its local frame, whose $x-$ and $z$ -axis are defined using the longest triangle edge at $t = 1$ and the face normal respectively. Furthermore, we set the last scaling factor, which corresponds to the scaling in face normal direction, to be a small constant value, ensuring the Gaussian lies on the mesh surface. The Gaussians can then be rendered in a differentiable manner [40] and we optimize their parameters by minimizing the loss [65] between the rendered images and the reference images: + +$$ +\mathcal {L} _ {\mathrm {r e n d e r}} = \lambda \left\| \mathbf {I} _ {i, t} - \hat {\mathbf {I}} _ {i, t} \right\| _ {1} + (1 - \lambda) \cdot \operatorname {S S I M} \left(\mathbf {I} _ {i, t}, \hat {\mathbf {I}} _ {i, t}\right), \tag {1} +$$ + +where $\hat{\mathbf{I}}_{i,t}$ denotes the image rendered from the perspective of the $i$ -th camera at time $t$ . Note that since the Gaussians' position and rotation are bound to the mesh, the optimization of the mesh vertices will implicitly optimize the parameters of the Gaussians. At the first time step, we optimize color, scale, and opacity. In the following frames $1 < t \leq T$ , we fix the opacity and scale, and optimize Gaussian colors $\mathbf{c}_t$ and mesh vertices $\mathbf{V}_t$ . + +In addition to the photometric loss, we apply the following regularization terms to preserve the local geometric features: 1) an isometry loss $\mathcal{L}_{\mathrm{iso}}$ [65] that preserves edge length through the training sequence, 2) a normal loss $\mathcal{L}_{\mathrm{normal}}$ that encourages smooth mesh surfaces, enhancing mesh deformation accuracy + +![](images/840589d5a2624a2f068a0c847a95c5ddd61c20937f953f28d3236fad05acb589.jpg) +t=12 + +![](images/15269b0c1622e5d39aac4763c334b086b1e9b370a6c84d42c11ad095ddf4250d.jpg) +Fig. 3: Our method can robustly track a dynamic mesh from input images, providing accurate long-term correspondences. Here we show the rendered images overlaid with the Gaussian trajectories from the previous 12 frames and the optimized meshes. + +![](images/d91cf663f0276150571c76b0686644a18c11dbbfd0cc8dac17e204a228e4ad5f.jpg) +t=300 + +![](images/2e36ae87752025754e9831984c0cf7856e52707cccf4346ca7cb4a0a4f335113.jpg) +t=600 + +![](images/cbb42d65500b56dbd675a54f09a977bd57658d40c306806de4dc910d040b0555.jpg) +t=900 + +and stability. Denoting the vertex and the triangle of the mesh at frame $t$ as $\mathbf{v}_t$ and $\mathbf{f}_t$ , the normal of a face as $\mathbf{n}(\cdot)$ , and the 1-ring neighbors $N(\cdot)$ , the losses are formally defined as: + +$$ +\mathcal {L} _ {\mathrm {i s o}} = \frac {1}{| \mathbf {V} |} \sum_ {\mathbf {v} \in \mathbf {V}} \sum_ {\mathbf {v} ^ {\prime} \in N (\mathbf {v} _ {1})} \frac {\left(\left\| \mathbf {v} _ {1} - \mathbf {v} _ {1} ^ {\prime} \right\| _ {2} - \left\| \mathbf {v} _ {t} - \mathbf {v} _ {t} ^ {\prime} \right\| _ {2}\right)}{| N (\mathbf {v}) |}, \tag {2} +$$ + +$$ +\mathcal {L} _ {\text {n o r m a l}} = \frac {1}{| \mathbf {F} |} \sum_ {\mathbf {f} _ {t} \in \mathbf {F}} \sum_ {\mathbf {f} _ {t} ^ {\prime} \in N (\mathbf {f} _ {t})} \left(1 - \mathbf {n} \left(\mathbf {f} _ {t}\right) \cdot \mathbf {n} \left(\mathbf {f} _ {t} ^ {\prime}\right)\right). \tag {3} +$$ + +The full loss used for mesh tracking is defined as: + +$$ +\mathcal {L} _ {\text {m e s h}} = \mathcal {L} _ {\text {r e n d e r}} + \lambda_ {i} \mathcal {L} _ {\text {i s o}} + \lambda_ {n} \mathcal {L} _ {\text {n o r m a l}}, \tag {4} +$$ + +where we set $\lambda_{i} = 10$ and $\lambda_{n} = 0.1$ during our training. With the introduced losses, our method can reconstruct a smooth mesh sequence with accurate correspondences, as shown in Fig. 3. + +# 3.2 Physics Based Dynamic Modeling + +Given the tracked meshes, $(\mathbf{V}_{1:T},\mathbf{F})$ , we estimate the physical properties of the garments such that the simulation of their dynamics matches the reference tracked meshes. Once learned, these parameters can be used to simulate the garment under novel motions and novel views. + +Garment Material Model and Simulation. We model the elastodynamics of the garment by equipping each simulation mesh with the discrete shell hinge bending energy [29,102] and isotropic StVK elastic potential [15,20]. Specifically, we estimate density $\rho$ , membrane stiffness $\kappa_{s}$ , and bending stiffness $\kappa_{b}$ of the garment given the visual observations. For our generic use cases, it is reasonable to assume homogeneous physical properties for each garment piece, i.e., the density, membrane stiffness, and bending stiffness can be parameterized using scalar values, denoted as $\rho$ , $\kappa_{s}$ , and $\kappa_{b}$ . In choosing the simulators, we consider differentiability and quality. Unfortunately, the development of differentiable simulators suitable for our application remains an open research question. Current implementations, such as those documented in [56], cannot be directly applied due to limitations in modeling complex body colliders. Therefore, we propose to use + +the state-of-the-art robust and stable cloth simulator C-IPC [52] that guarantees no interpenetration with frictional contact. While C-IPC is differentiable, it lacks an implementation of accurate analytical gradients; hence, we use finite difference methods to estimate the gradients for parameter updates. Specifically, we employ an open-source implementation of C-IPC [52] as our cloth simulator, which is a state-of-the-art simulator achieving high-fidelity garment simulation with stability and convergence (see [52] for a comprehensive review of the base algorithm and implementation). + +Garment Material Estimation. From the reconstructed mesh, we segment out the garment vertices to obtain ground truth garment mesh sequences $(\widehat{\mathbf{V}}_{1:T}^{g},\widehat{\mathbf{F}}^{g})$ . The cloth simulator, $f(\cdot)$ , takes as input the current garment state—position and velocity $(\mathbf{V}_t^g,\dot{\mathbf{V}}_t^g)$ —the garment's physical properties $(\rho ,\kappa_s,\kappa_b)$ , the body collider $\{\mathbf{V}_t^C,\mathbf{F}^C\}$ , and garment boundary vertices $\mathbf{V}_{t + 1}^b$ , and step size $\Delta t$ . The boundary vertices of the garment are points of the garment that are in permanent contact with the human body, such as the shoulder region for draped garments. We employ the SMPL-X body model [64] as the body collider $\{\mathbf{V}_{t + 1}^{C},\mathbf{F}^{C}\}$ , extracting SMPL-X body shapes and poses from the video data during the preprocessing (see Sec. 4). The garment and the boundary vertices are manually annotated once on the initial mesh $\mathbf{V}_1$ for the entire simulation. The simulation function predicting the garment shape in the next frame can be written as below: + +$$ +\mathbf {V} _ {t + 1} ^ {g} = f (\mathbf {V} _ {t} ^ {g}, \dot {\mathbf {V}} _ {t} ^ {g}, \mathbf {V} _ {t + 1} ^ {b}, \mathbf {V} _ {t + 1} ^ {C}, \rho , \kappa_ {s}, \kappa_ {b}, \Delta t). \tag {5} +$$ + +We find the optimal physical parameters by minimizing the difference between the simulated and the reconstructed garment meshes. Formally, the optimization objective can be expressed as: + +$$ +\min _ {\left\{\rho , \kappa_ {s}, \kappa_ {b} \right\}} \mathcal {L} _ {\text {s i m}} (\rho , \kappa_ {s}, \kappa_ {b}) = \sum_ {t = 0} ^ {T} \| \mathbf {V} _ {t + 1} ^ {g} - \widehat {\mathbf {V}} _ {t + 1} ^ {g} \| _ {2} ^ {2}, \tag {6} +$$ + +where $\mathbf{V}_1^g = \widehat{\mathbf{V}}_1^g$ and $\mathbf{V}_{t + 1}^g$ is computed by Eq. (5). We estimate gradients via a numerical approach using finite differences [50]. This method is particularly feasible in our context due to the low-dimensional nature of the optimization problem, which involves only three scalar parameters. We employ the forward difference formula for gradient estimation: + +$$ +\frac {\partial \mathcal {L} _ {\mathrm {s i m}}}{\partial \rho} \approx (\mathcal {L} _ {\mathrm {s i m}} (\rho + \delta_ {\rho}, \kappa_ {s}, \kappa_ {b}) - \mathcal {L} _ {\mathrm {s i m}} (\rho , \kappa_ {s}, \kappa_ {b})) / \delta_ {\rho}. \tag {7} +$$ + +Where $\delta_{\rho}$ is a hyperparameter for forward difference gradient estimation. The gradients for other parameters are estimated using the same approach, and the parameters are subsequently updated using the Adam optimizer. + +Animation. At test time, we have the initial mesh $\{\mathbf{V}_1,\mathbf{F}_1\}$ and we are given a sequence of novel human body poses, e.g., provided by skeleton parameters such as SMPL [64]. Using standard LBS, we obtain the boundary locations $\mathbf{V}_{t + 1}^b$ . We assume the initial garment is at rest, i.e., $\dot{\mathbf{V}}_0 = 0$ . We use the same garment + +![](images/99002568d07cf4242785855ef892d1629237b007a2a03966ffc5e73a81dee103.jpg) +(a) + +![](images/f351861199c87b93e8d1dd80680c976e5c055bce9882210fc89cc0cc9bb7f3fa.jpg) +Before Optimization +After Optimization +(b) + +![](images/98753764e95fd30f8733f407016da0016f595eac4b8da8dc62737c5588e1bb27.jpg) +PBR w/o Opt +Fig. 4: Ablation study on appearance estimation: (a) Initial texture map $\mathbf{T}$ extracted from Gaussian splatting (left) has baked-in shadows highlighted in the red boxes; post-optimization (right), the baked-in shadows are substantially removed. (b) Rendering comparisons demonstrate that our method with the optimized texture map more closely aligns with the ground truth. + +![](images/4bd4b31a34cc91e0dfa60e2c1f11cb7b0be21791b202ad017a4675c279128e38.jpg) +PBR Ours + +![](images/c5b67a8b8e3f2f0dd26d90e77a0c9d7ccde965216f81b6c363acdb4b4bbd9b7f.jpg) +GT + +segmentation, boundary points, and SMPL body collider points as during training, and use Eq. (5) to simulate the deformation of the garment. This is then combined with the body, driven by LBS, to create the holistic animation. + +# 3.3 Physics Based Appearance Modeling + +While tracking with dynamic Gaussians jointly estimates the garments appearance, it does not account for light-surface interactions, consequently the reconstructed appearance contains baked-in shadows, causing uncanny visual artifacts when rendered in novel views and motions (see Fig. 4). In this step, given the mesh sequence $(\mathbf{V}_{1:T},\mathbf{F})$ and the per-frame mesh colors $\mathbf{c}_{1:T}$ reconstructed from Section 3.1, we utilize Mitsuba3 [78]—a physics-based differentiable Monte Carlo renderer—to enhance the appearance reconstruction. For this purpose, we use a time-invariant diffuse texture map $\mathbf{T}$ to model the appearance. The value $\mathbf{T}$ is initialized from the average Gaussian colors from the tracking step (Section 3.1), $\Sigma_{1 < t\leq T}(\mathbf{c}(\mathbf{x})_t) / T$ . For the environment lighting, since the capture studio is typically equipped with uniform point light arrays, we can approximate the environment lighting as an ambient illumination, parameterized with a global $L_{a}\in \mathbb{R}^{3}$ . + +We optimize for $\mathbf{T}$ and $L_{a}$ using the photometric loss defined in Eq. (8) for all training views and frames $(i,t)$ , where $\mathcal{R}$ denotes the rendering function and $c_{i,t}$ the camera pose: + +$$ +\mathcal {L} _ {\mathrm {r g b}} = \min _ {\left\{\mathbf {T}, L _ {a} \right\}} \sum_ {i, t} \left\| \mathbf {I} _ {i, t} - \mathcal {R} \left(\mathbf {T}, L _ {a}, \mathbf {V} _ {t}, c _ {i, t}\right) \right\| _ {2} ^ {2}. \tag {8} +$$ + +At test time, we can swap the ambient light with more complex illuminations, and produce realistic rendering in arbitrary views and body poses (see Sec. 4.4). + +![](images/5570d9eb58ea3748a24e2c2903800d6034251f17e7e56c6cf866db386ed8bf76.jpg) +Fig. 5: Qualitative results on test poses from the ActorHQ [38] dataset. Our method PhysAvatar achieves state-of-the-art performance in terms of geometry detail and appearance modeling. + +# 4 Experiments + +In this section, we explain our experimental setup and results. We recommend watching the supplementary video for better visualization. + +# 4.1 Experimental setup + +Data Preparation. We choose four characters from the ActorHQ [35] dataset for our experiments, including two female characters with loose dresses, one female character wearing a tank top and loose-fitting pants, and one male character with a short-sleeved shirt and khaki shorts, as shown in Fig. 1. For each character, we use the ground truth mesh provided by the dataset and re-mesh it to obtain a simulation-ready initial mesh, i.e., one where the triangles have similar areas, to enhance the performance of garment simulation. We then segment the garment, define the boundary vertices of the garment for simulation purposes, and unwrap the whole mesh to get the UV map, all performed in Blender [13]. Since we observe that the human pose estimations from the dataset are quite noisy, we fit the SMPL-X [85] body mesh using multi-view videos by minimizing the distance between the projected SMPL-X joints and 2D keypoint detection results from DWPose [21, 121]. + +Implementation Details. For training, we utilize the entire set of 160 camera views from the ActorHQ [35] dataset. Although the dataset includes approximately 2000 frames for each character, we selectively use 24 frames with large movements for estimating garment material parameters and 200 frames for optimizing the texture map, which leads to good results while conserving compu + +Table 1: Quantitative comparison with baselines. PhysAvatar outperforms all baselines with respect to geometry accuracy. We achieve the best or competitive results in appearance metrics. See Fig. 5 and Fig. 6 for qualitative comparisons. + +
MethodGeometryAppearance
CD (×103) (↓)F-Score (↑)LPIPS (↓)PSNR (↑)SSIM (↑)
RAAH [115]1.1286.10.05528.60.957
TAVA [53]0.6692.30.05129.60.962
GS-Avatar [34]0.9189.40.04430.60.962
PhysAvatar (ours)0.5592.90.03530.20.957
+ +tational resources. For evaluation, we choose 200 frames for each character, with novel motions unseen during training. At test time, given a motion sequence defined by SMPL-X [85], we use an LBS weight inpainting algorithm [2] to obtain the skinning weights of the human body with respect to the SMPL-X skeleton. We animate the human body using LBS and simulate the garment, which is driven by the boundary vertices and utilizes the SMPL-X body mesh as the collider. More details are available at supplementary document. + +Baselines. We benchmark our method and current open-sourced state-of-the-art methods modeling full-body avatars, including ARAH [115] which leverages SDF-based volume rendering to learn the canonical representation of the avatar, TAVA [53] which learns the canonical radiance field [75] and skinning weight of the character, and GS-Avatar [34] based on a pose-conditioned neural network which predicts the parameters of 3D Gaussians [40]. For all baselines, we use the official public codebase and extend them to our dense-view setting. + +# 4.2 Comparison + +We compare PhysAvatar against current state-of-the-art approaches in terms of geometry and appearance quality across test pose sequences, as shown in Table 1, Fig. 5, and Fig. 6. + +Geometry Evaluation. For geometry evaluation, we employ the mean Chamfer distance [73] and the mean F-Score [104] at $\tau = 0.001$ as quantitative metrics. PhysAvatar outperforms all baseline according to these metrics, demonstrating a superior capability in accurately capturing the dynamic nature of garment geometry over time, as shown in Table 1. As shown in Fig. 5, PhysAvatar generates more accurate garment deformation and better captures fine wrinkle details, thanks to the integration of a physics-based simulator. This contrasts with all baseline methods [34,53,115], which do not explicitly model garment dynamics, but instead depend on learning quasi-static, pose-conditioned skinning weights or deformations. Additionally, we show results using motion data from the AMASS dataset [72]. PhysAvatar demonstrates its robustness by consistently synthesizing realistic garment deformation with fine details, while baseline methods sometimes fail to capture large deformations in loose garments (Fig. 6). + +![](images/a4694460856dc5457d22f15142cdd2c411dfe3b9168277d8280779a961b74105.jpg) + +![](images/cf969c606f5cdd7c2d7dcc961e18109834fd4fb66c3d72f6ef0a5ee6d4836203.jpg) + +![](images/8c1ef83398f094ea9ad4b3222dc39327fce742facd99be65bbf934a3f0cacb69.jpg) +Fig. 6: Here we show animation results of current state-of-the-art methods including ARAH [115], TAVA [53] and GS-Avatar [34], and our results on test motions from AMASS [72] dataset. Images are rendered from novel views. Please zoom in for details. + +![](images/099a6fd63455150aa4c35ad53d2755dad40bebd9895d1b7cd19c880bd9a1788f.jpg) + +Appearance Evaluation. For appearance evaluation, we utilize Peak Signal-to-Noise Ratio (PSNR), Structural Similarity Index Measure (SSIM), and the Learned Perceptual Image Patch Similarity (LPIPS) [129]. As detailed in Table 1, PhysAvatar surpasses all baselines in LPIPS, a metric shown to be better aligned with human perception [129]. While we achieve the second-best results in PSNR and SSIM, it is important to note that even minor misalignments in images can significantly impact these metrics, potentially not fully reflecting the quality of the renderings. As demonstrated in Fig. 5, PhysAvatar can capture more high-frequency details in facial features and garment textures, thereby creating images that are visually richer and more detailed, compared to the baseline approaches. + +# 4.3 Ablation + +We perform ablation studies to assess the impact of each component in our pipeline on the overall performance. For this, we focus on two female characters from the ActorHQ dataset, specifically chosen for their loose garments. + +Garment physics estimation and modeling. We ablate our approach against two variants: one employing Linear Blend Skinning (LBS) for garment deformation and another utilizing a physics simulator for garment dynamics without optimized parameters, henceforth referred to as 'w/o physics', which employs random garment parameters. As shown in Table 2, both alternatives exhibit worse performance across all geometric metrics. Further, as shown in Fig. 7, LBS fails to produce plausible deformations for loose garments, such as dresses. While simulation with random parameters yields plausible garment dynamics, they do not accurately reflect real-world observations. In contrast, our method, with learned physics parameters, achieves deformations that closely align with the reference, demonstrating its effectiveness. + +![](images/17d588cdd4eb58a72fd1aec8fa865a1c03bca182f33fd92bedea656cc429ec63.jpg) +LBS + +![](images/1a8a93f90bd59e3c76bb21043bb974199037f1a795c3fd3c63e20a7251ff52f3.jpg) +w/o Physics + +![](images/a1da5aff1a331a8c48de8c66c5dc38e71c1b6743fc0c1988bb9092a576563771.jpg) +Ours +Fig. 7: Ablation Study on garment physics estimation: LBS-based garment deformation (LBS) fails to generate realistic garment deformation. Physics-based simulation with random parameters (w/o physics) achieves realistic deformation but falls short of matching the ground-truth reference (GT). In contrast, PhysAvatar accurately captures garment dynamics, closely aligning with the reference. + +![](images/d3d7c6faf0c6338aa8cf66be7feeb13a0d6a9bf620baf86af22f0e39802858b6.jpg) +GT + +
MethodGeometryAppearance
CD (×10-3) (↓)F-Score (↑)LPIPS (↓)PSNR (↑)SSIM (↑)
LBS0.7191.60.035430.270.951
w/o physics0.6192.20.034730.260.951
w/o appearance——0.040228.260.947
PhysAvatar (ours)0.5693.00.034330.290.951
+ +Table 2: Ablation Study: PhysAvatar demonstrates improved geometric accuracy compared to both LBS-based animation (LBS) and simulations with random garment parameters (w/o physics), also shown in Fig. 7. PhysAvatar exhibits superior appearance metrics compared to the setting without physics-based appearance estimation (w/o appearance) (see Fig. 4). + +Appearance estimation and modeling. We compare rendering outcomes with those obtained without appearance optimization through physics-based inverse rendering [78] (denoted as 'w/o appearance'), where the texture is directly extracted from the Gaussian framework, as detailed in section 3.3. Fig. 4 illustrates that, since the Gaussian framework does not model complex light-surface interactions, shadows are inherently baked into the texture. However, post-optimization, we note a significant removal of these baked shadows. Quantitatively, we see improvements across all appearance metrics post-optimization, shown in Table 2. + +# 4.4 Application + +PhysAvatar can facilitate a wide range of applications, such as animation, relighting, and redressing, as demonstrated in Fig. 1. Our method can also benefit from other motion databases in addition to motions defined under the regime of SMPL-X [85] or SMPL [64]. Fig. 1 shows our animation results on challenging motions from Mixamo [76], where we use its auto-rigging algorithm to drive the human body, boundary vertices of the garment, and SMPL-X mesh as the + +![](images/7c0335449859589b52503fcf916029c5da5bf9fd2955c2c918c0adf628948c9d.jpg) +(a) Our results are readily exportable + +![](images/96c922d5af455dc27be21de45f9a56aa895487264f32c38d1ed806da8a3cb514.jpg) +Fig. 8: Our results can be exported into a traditional computer graphics pipeline, e.g., Blender [13], which enables the avatars performing novel motions to be rendered in unseen environments (a) as well as editing lighting and texture of the garments (b). + +![](images/73ca939b9006627ea05c7464a574a7b69037edf79a3b46a9cbcb42c6465d2379.jpg) +(b) Results with artistic modifications + +![](images/9d0c9034ce274eda34f5e032698a289f7c8d06e29141aa4506e4210d741a4df4.jpg) + +collider. Current state-of-the-art approaches cannot handle such settings since they are mainly based on LBS weights from SMPL or SMPL-X. Moreover, while current methods relying on NeRF [75] or 3D Gaussians [40] representation are usually incompatible with legacy graphics pipelines, our results are readily exportable and can be seamlessly integrated with computer graphics software tools (e.g., Blender [13]), enabling post-processing artistic modifications (Fig. 8). + +# 5 Limitations and Future Work + +Our method currently depends on manual garment segmentation and mesh UV unwrapping. We are posed to evolve our pipeline towards a more automated system by incorporating neural networks to streamline these data preprocessing tasks. Another limitation of our method is the absence of a perfect underlining human body for garment simulation. Our current use of the SMPL-X [85] model as the body collider introduces potential inaccuracies in collision detection and contact information due to mismatches between the SMPL-X mesh and the actual human body shape. Our method would benefit from using more advanced parametric models that promise a closer approximation to real human body shape. Furthermore, our pipeline leverages dense view captures to model avatars, an approach that, while effective, has its limitations. Adapting our method to a sparse-view setting would enhance the versatility of its applications in environments where capturing dense views is impractical. + +# 6 Conclusion + +We introduce PhysAvatar, a comprehensive pipeline to model clothed 3D human avatars with state-of-the-art realism. By integrating a robust and efficient mesh tracking method, and a novel inverse rendering paradigm to capture the physics of loose garments and the appearance of the body, we believe our work represents a significant advancement in digital avatar modeling from visual observations, which also paves the way for innovative applications in entertainment, virtual reality, and digital fashion. + +# Acknowledgement + +We would like to thank Jiayi Eris Zhang for the discussions. This material is based on work that is partially funded by an unrestricted gift from Google, Samsung, an SNF Postdoc Mobility fellowship, ARL grant W911NF-21-2-0104, and a Vannevar Bush Faculty Fellowship. + +# References + +1. Abdal, R., Yifan, W., Shi, Z., Xu, Y., Po, R., Kuang, Z., Chen, Q., Yeung, D.Y., Wetzstein, G.: Gaussian shell maps for efficient 3d human generation. arXiv preprint arXiv:2311.17857 (2023) +2. Abdrashitov, R., Raichstat, K., Monsen, J., Hill, D.: Robust skin weights transfer via weight inpainting. In: SIGGRAPH Asia 2023 Technical Communications, pp. 1-4 (2023) +3. Anguelov, D., Srinivasan, P., Koller, D., Thrun, S., Rodgers, J., Davis, J.: Scape: shape completion and animation of people. In: ACM SIGGRAPH 2005 Papers, pp. 408-416 (2005) +4. Bagautdinov, T., Wu, C., Simon, T., Prada, F., Shiratori, T., Wei, S.E., Xu, W., Sheikh, Y., Saragih, J.: Driving-signal aware full-body avatars. ACM Transactions on Graphics (TOG) 40(4), 1-17 (2021) +5. Baraff, D., Witkin, A.: Large steps in cloth simulation. In: Seminal Graphics Papers: Pushing the Boundaries, Volume 2, pp. 767-778 (2023) +6. Bashirov, R., Larionov, A., Ustinova, E., Sidorenko, M., Svitov, D., Zakharkin, I., Lempitsky, V.: Morf: Mobile realistic fullbody avatars from a monocular video. In: Proceedings of the IEEE/CVF Winter Conference on Applications of Computer Vision. pp. 3545-3555 (2024) +7. Bergman, A., Kellnhofer, P., Yifan, W., Chan, E., Lindell, D., Wetzstein, G.: Generative neural articulated radiance fields. Advances in Neural Information Processing Systems 35, 1990-19916 (2022) +8. Bertiche, H., Madadi, M., Escalera, S.: Cloth3d: clothed 3d humans. In: European Conference on Computer Vision. pp. 344-359. Springer (2020) +9. Bertiche, H., Madadi, M., Escalera, S.: Pbns: physically based neural simulation for unsupervised garment pose space deformation. ACM Transactions on Graphics (TOG) 40(6), 1-14 (2021) +10. Bertiche, H., Madadi, M., Escalera, S.: Neural cloth simulation. ACM Transactions on Graphics (TOG) 41(6), 1-14 (2022) +11. Bertiche, H., Madadi, M., Tyson, E., Escalera, S.: Deepsd: Automatic deep skinning and pose space deformation for 3d garment animation. In: Proceedings of the IEEE/CVF International Conference on Computer Vision. pp. 5471-5480 (2021) +12. Bhat, K.S., Twigg, C.D., Hodgins, J.K., Khosla, P.K., Popovic, Z., Seitz, S.M.: Estimating cloth simulation parameters from video. In: Proceedings of the 2003 ACM SIGGRAPH/Eurographics symposium on Computer animation. pp. 37-51 (2003) +13. Blender Online Coummunity: Blender, blender Foundation, https://www.blender.org/ +14. Chen, B., Shen, Y., Shuai, Q., Zhou, X., Zhou, K., Zheng, Y.: Anidress: Animatable loose-dressed avatar from sparse views using garment rigging model. arXiv preprint arXiv:2401.15348 (2024) + +15. Chen, H.Y., Sastry, A., van Rees, W.M., Vouga, E.: Physical simulation of environmentally induced thin shell deformation. ACM Transactions on Graphics (TOG) 37(4), 1-13 (2018) +16. Chen, X., Jiang, T., Song, J., Yang, J., Black, M.J., Geiger, A., Hilliges, O.: gdna: Towards generative detailed neural avatars. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 20427-20437 (2022) +17. Chen, X., Zheng, Y., Black, M.J., Hilliges, O., Geiger, A.: Snarf: Differentiable forward skimming for animating non-rigid neural implicit shapes. In: Proceedings of the IEEE/CVF International Conference on Computer Vision. pp. 11594-11604 (2021) +18. Chen, Y., Wang, X., Chen, X., Zhang, Q., Li, X., Guo, Y., Wang, J., Wang, F.: Uv volumes for real-time rendering of editable free-view human performance. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 16621-16631 (2023) +19. Chentanez, N., Macklin, M., Müller, M., Jeschke, S., Kim, T.Y.: Cloth and skin deformation with a triangle mesh based convolutional neural network. In: Computer Graphics Forum. vol. 39, pp. 123-134. Wiley Online Library (2020) +20. Clyde, D., Teran, J., Tamstorf, R.: Modeling and data-driven parameter estimation for woven fabrics. In: Proceedings of the ACM SIGGRAPH/Eurographics Symposium on Computer Animation. pp. 1-11 (2017) +21. Contributors, M.: Openmmlab pose estimation toolbox and benchmark. https://github.com/open-mmlab/mmpose (2020) +22. De Luigi, L., Li, R., Guillard, B., Salzmann, M., Fua, P.: Drapenet: Garment generation and self-supervised draping. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 1451-1460 (2023) +23. Dong, Z., Chen, X., Yang, J., Black, M.J., Hilliges, O., Geiger, A.: Ag3d: Learning to generate 3d avatars from 2d image collections. In: Proceedings of the IEEE/CVF International Conference on Computer Vision (ICCV). pp. 14916-14927 (October 2023) +24. Feng, Y., Yang, J., Pollefeys, M., Black, M.J., Bolkart, T.: Capturing and animation of body and clothing from monocular video. In: SIGGRAPH Asia 2022 Conference Papers. pp. 1-9 (2022) +25. Fridovich-Keil, S., Meanti, G., Warburg, F.R., Recht, B., Kanazawa, A.: K-planes: Explicit radiance fields in space, time, and appearance. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 12479-12488 (2023) +26. Genay, A., Lécuyer, A., Hachet, M.: Being an avatar "for real": a survey on virtual embodiment in augmented reality. IEEE Transactions on Visualization and Computer Graphics 28(12), 5071-5090 (2021) +27. Geng, C., Peng, S., Xu, Z., Bao, H., Zhou, X.: Learning neural volumetric representations of dynamic humans in minutes. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 8759-8770 (2023) +28. Grigorev, A., Black, M.J., Hilliges, O.: Hood: Hierarchical graphs for generalized modelling of clothing dynamics. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 16965-16974 (2023) +29. Grinspun, E., Hirani, A.N., Desbrun, M., Schröder, P.: Discrete shells. In: Proceedings of the 2003 ACM SIGGRAPH/Eurographics symposium on Computer animation. pp. 62-67. CiteSeer (2003) +30. Gundogdu, E., Constantin, V., Seifoddini, A., Dang, M., Salzmann, M., Fua, P.: Garnet: A two-stream network for fast and accurate 3d cloth draping. In: + +Proceedings of the IEEE/CVF International Conference on Computer Vision. pp. 8739-8748 (2019) +31. Guo, J., Li, J., Narain, R., Park, H.S.: Inverse simulation: Reconstructing dynamic geometry of clothed humans via optimal control. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 14698-14707 (2021) +32. Habermann, M., Liu, L., Xu, W., Pons-Moll, G., Zollhoefer, M., Theobalt, C.: Hdhumans: A hybrid approach for high-fidelity digital humans. Proceedings of the ACM on Computer Graphics and Interactive Techniques 6(3), 1-23 (2023) +33. Habermann, M., Liu, L., Xu, W., Zollhoefer, M., Pons-Moll, G., Theobalt, C.: Real-time deep dynamic characters. ACM Transactions on Graphics (ToG) 40(4), 1-16 (2021) +34. Hu, L., Zhang, H., Zhang, Y., Zhou, B., Liu, B., Zhang, S., Nie, L.: Gaussian-avatar: Towards realistic human avatar modeling from a single video via animatable 3d gaussians. arXiv preprint arXiv:2312.02134 (2023) +35. Isik, M., Rünz, M., Georgopoulos, M., Khakhulin, T., Starck, J., Agapito, L., Nießner, M.: Humanrf: High-fidelity neural radiance fields for humans in motion. ACM Transactions on Graphics (TOG) 42(4), 1-12 (2023). https://doi.org/10.1145/3592415, https://doi.org/10.1145/3592415 +36. Jambon, C., Kerbl, B., Kopanas, G., Diolatzis, S., Leimkuhler, T., Drettakis, G.: Nerfshop: Interactive editing of neural radiance fields. Proc. ACM Comput. Graph. Interact. Tech. 6, 1:1-1:21 (2023). https://doi.org/10.1145/3585499 +37. Jiang, Y., Wang, R., Liu, Z.: A survey of cloth simulation and applications. In: 2008 9th International Conference on Computer-Aided Industrial Design and Conceptual Design. pp. 765-769. IEEE (2008) +38. Jiang, Y., Shen, Z., Wang, P., Su, Z., Hong, Y., Zhang, Y., Yu, J., Xu, L.: Hifi4g: High-fidelity human performance rendering via compact gaussian splatting. arXiv preprint arXiv:2312.03461 (2023) +39. Kato, H., Ushiku, Y., Harada, T.: Neural 3d mesh renderer. 2018 IEEE/CVF Conference on Computer Vision and Pattern Recognition pp. 3907-3916 (2017). https://doi.org/10.1109/CVPR.2018.00411 +40. Kerbl, B., Kopanas, G., Leimkuhler, T., Drettakis, G.: 3d gaussian splatting for real-time radiance field rendering. ACM Transactions on Graphics 42(4) (2023) +41. Keselman, L., Hebert, M.: Approximate differentiable rendering with algebraic surfaces. In: European Conference on Computer Vision. pp. 596-614. Springer (2022) +42. Komatsu, K.: Human skin model capable of natural shape variation. The visual computer 3, 265-271 (1988) +43. Korban, M., Li, X.: A survey on applications of digital human avatars toward virtual co-presence. arXiv preprint arXiv:2201.04168 (2022) +44. Kwon, Y., Kim, D., Ceylan, D., Fuchs, H.: Neural human performer: Learning generalizable radiance fields for human performance rendering. Advances in Neural Information Processing Systems 34, 24741-24752 (2021) +45. Kwon, Y., Kim, D., Ceylan, D., Fuchs, H.: Neural image-based avatars: Generalizable radiance fields for human avatar modeling. In: The Eleventh International Conference on Learning Representations (2023) +46. Kwon, Y., Liu, L., Fuchs, H., Habermann, M., Theobalt, C.: Deliffas: Deformable light fields for fast avatar synthesis. Advances in Neural Information Processing Systems 36 (2023) + +47. Larionov, E., Eckert, M.L., Wolff, K., Stuyck, T.: Estimating cloth elasticity parameters using position-based simulation of compliant constrained dynamics. arXiv preprint arXiv:2212.08790 (2022) +48. Lassner, C., Zollhofer, M.: Pulsar: Efficient sphere-based neural rendering. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 1440-1449 (2021) +49. Lee, D., Kang, H., Lee, I.K.: Cloth combo: Modeling inter-cloth interaction for draping multi-layered clothes. ACM Transactions on Graphics (TOG) 42(6), 1-13 (2023) +50. LeVeque, R.J.: Finite difference methods for ordinary and partial differential equations: steady-state and time-dependent problems. SIAM (2007) +51. Li, M., Ferguson, Z., Schneider, T., Langlois, T.R., Zorin, D., Panozzo, D., Jiang, C., Kaufman, D.M.: Incremental potential contact: intersection-and inversion-free, large-deformation dynamics. ACM Trans. Graph. 39(4), 49 (2020) +52. Li, M., Kaufman, D.M., Jiang, C.: Codimensional incremental potential contact. ACM Transactions on Graphics (TOG) 40(4), 1-24 (2021) +53. Li, R., Tanke, J., Vo, M., Zollhofer, M., Gall, J., Kanazawa, A., Lassner, C.: Tava: Template-free animatable volumetric actors (2022) +54. Li, X., Li, X.R., Li, Y., Feng, W.: Review of cloth modeling and simulation for virtual fitting. Textile Research Journal 93(7-8), 1699-1711 (2023) +55. Li, Y., Chen, H.y., Larionov, E., Sarafianos, N., Matusik, W., Stuyck, T.: Difavatar: Simulation-ready garment optimization with differentiable simulation. arXiv preprint arXiv:2311.12194 (2023) +56. Li, Y., Du, T., Wu, K., Xu, J., Matusik, W.: Diffcloth: Differentiable cloth simulation with dry frictional contact. ACM Transactions on Graphics (TOG) 42(1), 1-20 (2022) +57. Li, Z., Zheng, Z., Liu, Y., Zhou, B., Liu, Y.: Posevocab: Learning joint-structured pose embeddings for human avatar modeling. arXiv preprint arXiv:2304.13006 (2023) +58. Li, Z., Zheng, Z., Wang, L., Liu, Y.: Animatable gaussians: Learning pose-dependent gaussian maps for high-fidelity human avatar modeling. arXiv preprint arXiv:2311.16096 (2023) +59. Liu, L., Habermann, M., Rudnev, V., Sarkar, K., Gu, J., Theobalt, C.: Neural actor: Neural free-view synthesis of human actors with pose control. ACM transactions on graphics (TOG) 40(6), 1-16 (2021) +60. Liu, L., Xu, W., Zollhoefer, M., Kim, H., Bernard, F., Habermann, M., Wang, W., Theobalt, C.: Neural rendering and reenactment of human actor videos. ACM Transactions on Graphics (TOG) 38(5), 1-14 (2019) +61. Liu, S., Li, T., Chen, W., Li, H.: Soft rasterizer: A differentiable renderer for image-based 3d reasoning. In: Proceedings of the IEEE/CVF International Conference on Computer Vision. pp. 7708-7717 (2019) +62. Liu, T., Bouaziz, S., Kavan, L.: Quasi-newton methods for real-time simulation of hyperelastic materials. Acm Transactions on Graphics (TOG) 36(3), 1-16 (2017) +63. Lombardi, S., Simon, T., Saragih, J., Schwartz, G., Lehrmann, A., Sheikh, Y.: Neural volumes: Learning dynamic renderable volumes from images. arXiv preprint arXiv:1906.07751 (2019) +64. Loper, M., Mahmood, N., Romero, J., Pons-Moll, G., Black, M.J.: Smpl: A skinned multi-person linear model. ACM Transactions on Graphics 34(6) (2015) +65. Luiten, J., Kopanas, G., Leibe, B., Ramanan, D.: Dynamic 3d gaussians: Tracking by persistent dynamic view synthesis. arXiv preprint arXiv:2308.09713 (2023) + +66. Ma, Q., Saito, S., Yang, J., Tang, S., Black, M.J.: Scale: Modeling clothed humans with a surface codec of articulated local elements. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 16082-16093 (2021) +67. Ma, Q., Yang, J., Black, M.J., Tang, S.: Neural point-based shape modeling of humans in challenging clothing. In: 2022 International Conference on 3D Vision (3DV). pp. 679-689. IEEE (2022) +68. Ma, Q., Yang, J., Ranjan, A., Pujades, S., Pons-Moll, G., Tang, S., Black, M.J.: Learning to dress 3d people in generative clothing. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 6469-6478 (2020) +69. Ma, Q., Yang, J., Tang, S., Black, M.J.: The power of points for modeling humans in clothing. In: Proceedings of the IEEE/CVF International Conference on Computer Vision. pp. 10974-10984 (2021) +70. Macklin, M., Müller, M., Chentanez, N.: Xpbd: position-based simulation of compliant constrained dynamics. In: Proceedings of the 9th International Conference on Motion in Games. p. 49-54. MIG '16, Association for Computing Machinery, New York, NY, USA (2016). https://doi.org/10.1145/2994258.2994272, https://doi.org/10.1145/2994258.2994272 +71. Magnenat, T., Laperrière, R., Thalmann, D.: Joint-dependent local deformations for hand animation and object grasping. Tech. rep., Canadian Inf. Process. Soc (1988) +72. Mahmood, N., Ghorbani, N., Troje, N.F., Pons-Moll, G., Black, M.J.: AMASS: Archive of motion capture as surface shapes. In: ICCV (2019) +73. Mescheder, L., Oechsle, M., Niemeyer, M., Nowozin, S., Geiger, A.: Occupancy networks: Learning 3d reconstruction in function space. In: Proceedings of the IEEE/CVF conference on computer vision and pattern recognition. pp. 4460-4470 (2019) +74. Miguel, E., Bradley, D., Thomaszewski, B., Bickel, B., Matusik, W., Otaduy, M.A., Marschner, S.: Data-driven estimation of cloth simulation models. In: Computer Graphics Forum. vol. 31, pp. 519-528. Wiley Online Library (2012) +75. Mildenhall, B., Srinivasan, P.P., Tancik, M., Barron, J.T., Ramamoorthi, R., Ng, R.: Nerf: Representing scenes as neural radiance fields for view synthesis. Communications of the ACM 65(1), 99-106 (2021) +76. Mixamo, www MIXamo.com +77. Newcombe, R.A., Izadi, S., Hilliges, O., Molyneaux, D., Kim, D., Davison, A.J., Kohi, P., Shotton, J., Hodges, S., Fitzgibbon, A.: Kinectfusion: Real-time dense surface mapping and tracking. In: 2011 10th IEEE international symposium on mixed and augmented reality. pp. 127-136. IEEE (2011) +78. Nimier-David, M., Vicini, D., Zeltner, T., Jakob, W.: Mitsuba 2: A retargetable forward and inverse renderer. ACM Transactions on Graphics (TOG) 38(6), 1-17 (2019) +79. Noguchi, A., Sun, X., Lin, S., Harada, T.: Neural articulated radiance field. In: Proceedings of the IEEE/CVF International Conference on Computer Vision (ICCV). pp. 5762-5772 (October 2021) +80. Noguchi, A., Sun, X., Lin, S., Harada, T.: Unsupervised learning of efficient geometry-aware neural articulated representations. In: European Conference on Computer Vision. pp. 597-614. Springer (2022) +81. Pan, X., Mai, J., Jiang, X., Tang, D., Li, J., Shao, T., Zhou, K., Jin, X., Manocha, D.: Predicting loose-fitting garment deformations using bone-driven motion networks. In: ACM SIGGRAPH 2022 Conference Proceedings. pp. 1-10 (2022) + +82. Pang, H., Zhu, H., Kortylewski, A., Theobalt, C., Habermann, M.: Ash: Animatable gaussian splats for efficient and photoreal human rendering. arXiv preprint arXiv:2312.05941 (2023) +83. Park, J.J., Florence, P., Straub, J., Newcombe, R., Lovegrove, S.: Deepsdf: Learning continuous signed distance functions for shape representation. In: Proceedings of the IEEE/CVF conference on computer vision and pattern recognition. pp. 165-174 (2019) +84. Park, K., Sinha, U., Hedman, P., Barron, J.T., Bouaziz, S., Goldman, D.B., Martin-Brualla, R., Seitz, S.M.: Hypernerf: a higher-dimensional representation for topologically varying neural radiance fields. ACM Transactions on Graphics (TOG) 40(6), 1-12 (2021) +85. Pavlakos, G., Choutas, V., Ghorbani, N., Bolkart, T., Osman, A.A.A., Tzionas, D., Black, M.J.: Expressive body capture: 3D hands, face, and body from a single image. In: Proceedings IEEE Conf. on Computer Vision and Pattern Recognition (CVPR). pp. 10975-10985 (2019) +86. Peng, S., Dong, J., Wang, Q., Zhang, S., Shuai, Q., Zhou, X., Bao, H.: Animatable neural radiance fields for modeling dynamic human bodies. In: Proceedings of the IEEE/CVF International Conference on Computer Vision. pp. 14314-14323 (2021) +87. Peng, S., Zhang, Y., Xu, Y., Wang, Q., Shuai, Q., Bao, H., Zhou, X.: Neural body: Implicit neural representations with structured latent codes for novel view synthesis of dynamic humans. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 9054-9063 (2021) +88. Peng, T., Wu, W., Liu, J., Li, L., Miao, J., Hu, X., He, R., Li, L.: Pgn-cloth: Physics-based graph network model for 3d cloth animation. Displays 80, 102534 (2023) +89. Pumarola, A., Corona, E., Pons-Moll, G., Moreno-Noguer, F.: D-nerf: Neural radiance fields for dynamic scenes. 2021 IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR) pp. 10313-10322 (2020). https://doi.org/10.1109/CVPR46437.2021.01018 +90. Saito, S., Yang, J., Ma, Q., Black, M.J.: Scanimate: Weakly supervised learning of skinned clothed avatar networks. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 2886-2897 (2021) +91. Santesteban, I., Otaduy, M.A., Casas, D.: Learning-based animation of clothing for virtual try-on. In: Computer Graphics Forum. vol. 38, pp. 355-366. Wiley Online Library (2019) +92. Santesteban, I., Otaduy, M.A., Casas, D.: Snug: Self-supervised neural dynamic garments. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 8140-8150 (2022) +93. Schonberger, J.L., Frahm, J.M.: Structure-from-motion revisited. In: Proceedings of the IEEE conference on computer vision and pattern recognition. pp. 4104-4113 (2016) +94. Sitzmann, V., Thies, J., Heide, F., Nießner, M., Wetzstein, G., Zollhöfer, M.: Deepvoxels: Learning persistent 3d feature embeddings. In: Proc. Computer Vision and Pattern Recognition (CVPR), IEEE (2019) +95. Sitzmann, V., Zollhöfer, M., Wetzstein, G.: Scene representation networks: Continuous 3d-structure-aware neural scene representations. In: Advances in Neural Information Processing Systems (2019) +96. Snavely, N., Seitz, S.M., Szeliski, R.: Photo tourism: exploring photo collections in 3d. In: ACM siggraph 2006 papers, pp. 835-846 (2006) + +97. Stoll, C., Gall, J., De Aguiar, E., Thrun, S., Theobalt, C.: Video-based reconstruction of animatable human characters. ACM Transactions on Graphics (TOG) 29(6), 1-10 (2010) +98. Stuyck, T., Chen, H.y.: Diffxpbd: Differentiable position-based simulation of compliant constraint dynamics. Proceedings of the ACM on Computer Graphics and Interactive Techniques 6(3), 1-14 (2023) +99. Su, S.Y., Yu, F., Zollhöfer, M., Rhodin, H.: A-nerf: Articulated neural radiance fields for learning human shape, appearance, and pose. Advances in Neural Information Processing Systems 34, 12278–12291 (2021) +100. Su, Z., Hu, L., Lin, S., Zhang, H., Zhang, S., Thies, J., Liu, Y.: Caphy: Capturing physical properties for animatable human avatars. In: Proceedings of the IEEE/CVF International Conference on Computer Vision. pp. 14150-14160 (2023) +101. Sun, M., Yang, D., Kou, D., Jiang, Y., Shan, W., Yan, Z., Zhang, L.: Human 3d avatar modeling with implicit neural representation: A brief survey. In: 2022 14th International Conference on Signal Processing Systems (ICSPS). pp. 818-827. IEEE (2022) +102. Tamstorf, R., Grinspun, E.: Discrete bending forces and their jacobians. Graphical models 75(6), 362-370 (2013) +103. Tang, M., Tong, R., Narain, R., Meng, C., Manocha, D.: A gpu-based streaming algorithm for high-resolution cloth simulation. In: Computer Graphics Forum. vol. 32, pp. 21-30. Wiley Online Library (2013) +104. Tatarchenko, M., Richter, S.R., Ranftl, R., Li, Z., Koltun, V., Brox, T.: What do single-view 3d reconstruction networks learn? In: Proceedings of the IEEE/CVF conference on computer vision and pattern recognition. pp. 3405-3414 (2019) +105. Tewari, A., Fried, O., Thies, J., Sitzmann, V., Lombardi, S., Sunkavalli, K., Martin-Brualla, R., Simon, T., Saragih, J.M., Nießner, M., Pandey, R., Fanello, S., Wetzstein, G., Zhu, J.Y., Theobalt, C., Agrawala, M., Shechtman, E., Goldman, D.B., Zollhofer, M.: State of the art on neural rendering. Computer Graphics Forum 39 (2020). https://doi.org/10.1111/cgf.14022 +106. Tewari, A., Thies, J., Mildenhall, B., Srinivasan, P., Tretschk, E., Yifan, W., Lassner, C., Sitzmann, V., Martin-Brualla, R., Lombardi, S., et al.: Advances in neural rendering. In: Computer Graphics Forum. vol. 41, pp. 703-735. Wiley Online Library (2022) +107. Tiwari, G., Sarafianos, N., Tung, T., Pons-Moll, G.: Neural-gif: Neural generalized implicit functions for animating people in clothing. In: Proceedings of the IEEE/CVF International Conference on Computer Vision. pp. 11708-11718 (2021) +108. Tretschk, E., Tewari, A., Golyanik, V., Zollhöfer, M., Lassner, C., Theobalt, C.: Non-rigid neural radiance fields: Reconstruction and novel view synthesis of a dynamic scene from monocular video. 2021 IEEE/CVF International Conference on Computer Vision (ICCV) pp. 12939-12950 (2020). https://doi.org/10.1109/ICCV48922.2021.01272 +109. Vassilev, T., Spanlang, B., Chrysanthou, Y.: Fast cloth animation on walking avatars. In: Computer Graphics Forum. vol. 20, pp. 260-267. Wiley Online Library (2001) +110. Vidaurre, R., Santesteban, I., Garces, E., Casas, D.: Fully convolutional graph neural networks for parametric virtual try-on. In: Computer Graphics Forum. vol. 39, pp. 145-156. Wiley Online Library (2020) +111. Wang, H.: Gpu-based simulation of cloth wrinkles at submillimeter levels. ACM Transactions on Graphics (TOG) 40(4), 1-14 (2021) + +112. Wang, H., O'Brien, J.F., Ramamoorthi, R.: Data-driven elastic models for cloth: modeling and measurement. ACM transactions on graphics (TOG) 30(4), 1-12 (2011) +113. Wang, P., Liu, L., Liu, Y., Theobalt, C., Komura, T., Wang, W.: Neus: Learning neural implicit surfaces by volume rendering for multi-view reconstruction. Advances in Neural Information Processing Systems 34, 27171-27183 (2021) +114. Wang, S., Schwarz, K., Geiger, A., Tang, S.: Arah: Animatable volume rendering of articulated human sdfs. In: European conference on computer vision. pp. 1-19. Springer (2022) +115. Wang, S., Schwarz, K., Geiger, A., Tang, S.: Arah: Animatable volume rendering of articulated human sdfs. In: European Conference on Computer Vision (2022) +116. Wang, T., Zhang, B., Zhang, T., Gu, S., Bao, J., Baltrusaitis, T., Shen, J., Chen, D., Wen, F., Chen, Q., et al.: Rodin: A generative model for sculpting 3d digital avatars using diffusion. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 4563-4573 (2023) +117. Wu, G., Yi, T., Fang, J., Xie, L., Zhang, X., Wei, W., Liu, W., Tian, Q., Wang, X.: 4d gaussian splatting for real-time dynamic scene rendering. ArXiv abs/2310.08528 (2023). https://doi.org/10.48550/arXiv.2310.08528 +118. Xiang, D., Bagautdinov, T., Stuyck, T., Prada, F., Romero, J., Xu, W., Saito, S., Guo, J., Smith, B., Shiratori, T., et al.: Dressing avatars: Deep photorealistic appearance for physically simulated clothing. ACM Transactions on Graphics (TOG) 41(6), 1-15 (2022) +119. Xu, Y., Yifan, W., Bergman, A.W., Chai, M., Zhou, B., Wetzstein, G.: Efficient 3d articulated human generation with layered surface volumes. arXiv preprint arXiv:2307.05462 (2023) +120. Xu, Y., Chen, B., Li, Z., Zhang, H., Wang, L., Zheng, Z., Liu, Y.: Gaussian head avatar: Ultra high-fidelity head avatar via dynamic gaussians. arXiv preprint arXiv:2312.03029 (2023) +121. Yang, Z., Zeng, A., Yuan, C., Li, Y.: Effective whole-body pose estimation with two-stages distillation. In: Proceedings of the IEEE/CVF International Conference on Computer Vision. pp. 4210-4220 (2023) +122. Yariv, L., Gu, J., Kasten, Y., Lipman, Y.: Volume rendering of neural implicit surfaces. Advances in Neural Information Processing Systems 34, 4805-4815 (2021) +123. Yee, K.S., Chen, J.S.: The finite-difference time-domain (fdtd) and the finite-volume time-domain (fvtd) methods in solving maxwell's equations. IEEE Transactions on Antennas and Propagation 45(3), 354-363 (1997) +124. Yifan, W., Serena, F., Wu, S., Öztireli, C., Sorkine-Hornung, O.: Differentiable surface splatting for point-based geometry processing. ACM Transactions on Graphics (TOG) 38(6), 1-14 (2019) +125. Yu, T., Zheng, Z., Zhong, Y., Zhao, J., Dai, Q., Pons-Moll, G., Liu, Y.: Simulcap: Single-view human performance capture with cloth simulation. In: Proceedings of the IEEE/CVF conference on computer vision and pattern recognition. pp. 5504-5514 (2019) +126. Yuan, Y.J., Sun, Y.T., Lai, Y.K., Ma, Y., Jia, R., Gao, L.: Nerf-editing: geometry editing of neural radiance fields. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 18353-18364 (2022) +127. Zeller, C.: Cloth simulation on thegpu. In: ACM SIGGRAPH 2005 Sketches, pp. 39-es (2005) +128. Zhang, J.X., Lin, G.W.C., Bode, L., Chen, H.y., Stuyck, T., Larionov, E.: Estimating cloth elasticity parameters from homogenized yarn-level models. arXiv preprint arXiv:2401.15169 (2024) + +129. Zhang, R., Isola, P., Efros, A.A., Shechtman, E., Wang, O.: The unreasonable effectiveness of deep features as a perceptual metric. In: Proceedings of the IEEE conference on computer vision and pattern recognition. pp. 586-595 (2018) +130. Zhao, F., Jiang, Y., Yao, K., Zhang, J., Wang, L., Dai, H., Zhong, Y., Zhang, Y., Wu, M., Xu, L., et al.: Human performance modeling and rendering via neural animated mesh. ACM Transactions on Graphics (TOG) 41(6), 1-17 (2022) +131. Zheng, Z., Huang, H., Yu, T., Zhang, H., Guo, Y., Liu, Y.: Structured local radiance fields for human avatar modeling. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 15893-15903 (2022) +132. Zielonka, W., Bagautdinov, T., Saito, S., Zollhöfer, M., Thies, J., Romero, J.: Drivable 3d gaussian avatars. arXiv preprint arXiv:2311.08581 (2023) \ No newline at end of file diff --git a/2024/PhysAvatar_ Learning the Physics of Dressed 3D Avatars from Visual Observations/images.zip b/2024/PhysAvatar_ Learning the Physics of Dressed 3D Avatars from Visual Observations/images.zip new file mode 100644 index 0000000000000000000000000000000000000000..98fd954339295c9f481034fd69b09cc92b1c0e08 --- /dev/null +++ b/2024/PhysAvatar_ Learning the Physics of Dressed 3D Avatars from Visual Observations/images.zip @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:9b6e7efd87f585c3aaae60176ed66c2b98cfcd555ab2b28fd0e7f78f1cb1a72a +size 424539 diff --git a/2024/PhysAvatar_ Learning the Physics of Dressed 3D Avatars from Visual Observations/layout.json b/2024/PhysAvatar_ Learning the Physics of Dressed 3D Avatars from Visual Observations/layout.json new file mode 100644 index 0000000000000000000000000000000000000000..46089679685aba4b4b9997f834483787a5a05b70 --- /dev/null +++ b/2024/PhysAvatar_ Learning the Physics of Dressed 3D Avatars from Visual Observations/layout.json @@ -0,0 +1,12960 @@ +{ + "pdf_info": [ + { + "para_blocks": [ + { + "bbox": [ + 137, + 111, + 477, + 148 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 137, + 111, + 477, + 148 + ], + "spans": [ + { + "bbox": [ + 137, + 111, + 477, + 148 + ], + "type": "text", + "content": "PhysAvatar: Learning the Physics of Dressed 3D Avatars from Visual Observations" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 132, + 167, + 481, + 205 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 167, + 481, + 205 + ], + "spans": [ + { + "bbox": [ + 132, + 167, + 481, + 205 + ], + "type": "text", + "content": "Yang Zheng" + }, + { + "bbox": [ + 132, + 167, + 481, + 205 + ], + "type": "inline_equation", + "content": "^{1\\star}" + }, + { + "bbox": [ + 132, + 167, + 481, + 205 + ], + "type": "text", + "content": ", Qingqing Zhao" + }, + { + "bbox": [ + 132, + 167, + 481, + 205 + ], + "type": "inline_equation", + "content": "^{1\\star}" + }, + { + "bbox": [ + 132, + 167, + 481, + 205 + ], + "type": "text", + "content": ", Guandao Yang" + }, + { + "bbox": [ + 132, + 167, + 481, + 205 + ], + "type": "inline_equation", + "content": "^{1}" + }, + { + "bbox": [ + 132, + 167, + 481, + 205 + ], + "type": "text", + "content": ", Wang Yifan" + }, + { + "bbox": [ + 132, + 167, + 481, + 205 + ], + "type": "inline_equation", + "content": "^{1}" + }, + { + "bbox": [ + 132, + 167, + 481, + 205 + ], + "type": "text", + "content": ", Donglai Xiang" + }, + { + "bbox": [ + 132, + 167, + 481, + 205 + ], + "type": "inline_equation", + "content": "^{2}" + }, + { + "bbox": [ + 132, + 167, + 481, + 205 + ], + "type": "text", + "content": ", Florian Dubost" + }, + { + "bbox": [ + 132, + 167, + 481, + 205 + ], + "type": "inline_equation", + "content": "^{3}" + }, + { + "bbox": [ + 132, + 167, + 481, + 205 + ], + "type": "text", + "content": ", Dmitry Lagun" + }, + { + "bbox": [ + 132, + 167, + 481, + 205 + ], + "type": "inline_equation", + "content": "^{3}" + }, + { + "bbox": [ + 132, + 167, + 481, + 205 + ], + "type": "text", + "content": ", Thabo Beeler" + }, + { + "bbox": [ + 132, + 167, + 481, + 205 + ], + "type": "inline_equation", + "content": "^{3}" + }, + { + "bbox": [ + 132, + 167, + 481, + 205 + ], + "type": "text", + "content": ", Federico Tombari" + }, + { + "bbox": [ + 132, + 167, + 481, + 205 + ], + "type": "inline_equation", + "content": "^{3,4}" + }, + { + "bbox": [ + 132, + 167, + 481, + 205 + ], + "type": "text", + "content": ", Leonidas Guibas" + }, + { + "bbox": [ + 132, + 167, + 481, + 205 + ], + "type": "inline_equation", + "content": "^{1}" + }, + { + "bbox": [ + 132, + 167, + 481, + 205 + ], + "type": "text", + "content": ", and Gordon Wetzstein" + }, + { + "bbox": [ + 132, + 167, + 481, + 205 + ], + "type": "inline_equation", + "content": "^{1}" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 261, + 213, + 352, + 225 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 261, + 213, + 352, + 225 + ], + "spans": [ + { + "bbox": [ + 261, + 213, + 352, + 225 + ], + "type": "text", + "content": "1 Stanford University" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 246, + 224, + 367, + 236 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 246, + 224, + 367, + 236 + ], + "spans": [ + { + "bbox": [ + 246, + 224, + 367, + 236 + ], + "type": "text", + "content": "2 Carnegie Mellon University" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 288, + 236, + 326, + 247 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 288, + 236, + 326, + 247 + ], + "spans": [ + { + "bbox": [ + 288, + 236, + 326, + 247 + ], + "type": "text", + "content": "3 Google" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 239, + 247, + 374, + 257 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 239, + 247, + 374, + 257 + ], + "spans": [ + { + "bbox": [ + 239, + 247, + 374, + 257 + ], + "type": "text", + "content": "4 Technical University of Munich" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 159, + 287, + 455, + 485 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 159, + 287, + 455, + 485 + ], + "spans": [ + { + "bbox": [ + 159, + 287, + 455, + 485 + ], + "type": "text", + "content": "Abstract. Modeling and rendering photorealistic avatars is of crucial importance in many applications. Existing methods that build a 3D avatar from visual observations, however, struggle to reconstruct clothed humans. We introduce PhysAvatar, a novel framework that combines inverse rendering with inverse physics to automatically estimate the shape and appearance of a human from multi-view video data along with the physical parameters of the fabric of their clothes. For this purpose, we adopt a mesh-aligned 4D Gaussian technique for spatio-temporal mesh tracking as well as a physically based inverse renderer to estimate the intrinsic material properties. PhysAvatar integrates a physics simulator to estimate the physical parameters of the garments using gradient-based optimization in a principled manner. These novel capabilities enable PhysAvatar to create high-quality novel-view renderings of avatars dressed in loose-fitting clothes under motions and lighting conditions not seen in the training data. This marks a significant advancement towards modeling photorealistic digital humans using physically based inverse rendering with physics in the loop. Our project website is at: https://qingqing-zhao.github.io/PhysAvatar." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 160, + 495, + 451, + 507 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 160, + 495, + 451, + 507 + ], + "spans": [ + { + "bbox": [ + 160, + 495, + 451, + 507 + ], + "type": "text", + "content": "Keywords: neural rendering " + }, + { + "bbox": [ + 160, + 495, + 451, + 507 + ], + "type": "inline_equation", + "content": "\\cdot" + }, + { + "bbox": [ + 160, + 495, + 451, + 507 + ], + "type": "text", + "content": " physics " + }, + { + "bbox": [ + 160, + 495, + 451, + 507 + ], + "type": "inline_equation", + "content": "\\cdot" + }, + { + "bbox": [ + 160, + 495, + 451, + 507 + ], + "type": "text", + "content": " dynamic modeling " + }, + { + "bbox": [ + 160, + 495, + 451, + 507 + ], + "type": "inline_equation", + "content": "\\cdot" + }, + { + "bbox": [ + 160, + 495, + 451, + 507 + ], + "type": "text", + "content": " 3D avatar" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 132, + 527, + 230, + 539 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 527, + 230, + 539 + ], + "spans": [ + { + "bbox": [ + 132, + 527, + 230, + 539 + ], + "type": "text", + "content": "1 Introduction" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 130, + 552, + 482, + 647 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 552, + 482, + 647 + ], + "spans": [ + { + "bbox": [ + 130, + 552, + 482, + 647 + ], + "type": "text", + "content": "Digital avatars are a vital component in numerous applications, ranging from virtual reality and gaming to telepresence and e-commerce [26,43,101]. A realistic 3D avatar of a person can be readily obtained from visual observations, such as multi-view image [116] and video [6] data. The task of rendering animated 3D avatars from novel viewpoints or when performing unseen motions, however, presents considerable challenges, particularly when the avatar wears loose-fitting garments. Accurately rendering the dynamics of garments in conditions that are not observed in the training data necessitates a holistic approach that not only" + } + ] + } + ], + "index": 9 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 133, + 653, + 224, + 665 + ], + "type": "page_footnote", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 653, + 224, + 665 + ], + "spans": [ + { + "bbox": [ + 133, + 653, + 224, + 665 + ], + "type": "text", + "content": "* Equal Contribution" + } + ] + } + ], + "index": 10 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 0 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 151, + 117, + 321, + 290 + ], + "blocks": [ + { + "bbox": [ + 151, + 117, + 321, + 290 + ], + "lines": [ + { + "bbox": [ + 151, + 117, + 321, + 290 + ], + "spans": [ + { + "bbox": [ + 151, + 117, + 321, + 290 + ], + "type": "image", + "image_path": "9f5e9783a5bafabc9c297493ffa1d29405d2f8b9638b10115540c976aeb1567e.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 131, + 295, + 482, + 328 + ], + "lines": [ + { + "bbox": [ + 131, + 295, + 482, + 328 + ], + "spans": [ + { + "bbox": [ + 131, + 295, + 482, + 328 + ], + "type": "text", + "content": "Fig. 1: PhysAvatar is a novel framework that captures the physics of dressed 3D avatars from visual observations, enabling a wide spectrum of applications, such as (a) animation, (b) relighting, and (c) redressing, with high-fidelity rendering results." + } + ] + } + ], + "index": 6, + "angle": 0, + "type": "image_caption" + } + ], + "index": 2 + }, + { + "type": "image", + "bbox": [ + 348, + 121, + 451, + 208 + ], + "blocks": [ + { + "bbox": [ + 327, + 197, + 338, + 206 + ], + "lines": [ + { + "bbox": [ + 327, + 197, + 338, + 206 + ], + "spans": [ + { + "bbox": [ + 327, + 197, + 338, + 206 + ], + "type": "text", + "content": "(b)" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 348, + 121, + 451, + 208 + ], + "lines": [ + { + "bbox": [ + 348, + 121, + 451, + 208 + ], + "spans": [ + { + "bbox": [ + 348, + 121, + 451, + 208 + ], + "type": "image", + "image_path": "ef0b754e528d2e8ffdf74a4562d29eda1bcd79f861f6206cf04ecf94a9e1a3c6.jpg" + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "image_body" + } + ], + "index": 4 + }, + { + "type": "image", + "bbox": [ + 345, + 215, + 459, + 289 + ], + "blocks": [ + { + "bbox": [ + 345, + 215, + 459, + 289 + ], + "lines": [ + { + "bbox": [ + 345, + 215, + 459, + 289 + ], + "spans": [ + { + "bbox": [ + 345, + 215, + 459, + 289 + ], + "type": "image", + "image_path": "3c62b42598198dd839a86e1b1569e25ad87a6b2105516b83a9f9c04136db4314.jpg" + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "image_body" + } + ], + "index": 5 + }, + { + "bbox": [ + 130, + 353, + 480, + 376 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 353, + 480, + 376 + ], + "spans": [ + { + "bbox": [ + 130, + 353, + 480, + 376 + ], + "type": "text", + "content": "models the shape and appearance of the person but also the physical behavior of their clothes, including friction and collision." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 130, + 378, + 482, + 640 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 378, + 482, + 640 + ], + "spans": [ + { + "bbox": [ + 130, + 378, + 482, + 640 + ], + "type": "text", + "content": "Learning 3D scene representations from visual data is the core problem addressed by inverse rendering—a field that has recently shown remarkable progress in estimating the geometry and appearance of static and dynamic scenes from multi-view images and videos [105, 106]. In the context of reconstructing 3D avatars, the ability to explicitly control the motions of the avatar in post processing becomes imperative. Most existing methods for reconstructing animatable avatars drive the animation through an underlying skeleton using linear blending skinning (LBS) [64], which adequately models the dynamics of humans dressed in tight-fitting clothes [1,7,23,80,99,119]. With this approach, garments are treated as a rigidly attached part of the body adhering to piece-wise linear transformations, which results in motions that appear rigid and unconvincing in many cases [79, 120, 132]. Several successful attempts to introduce non-rigid deformations through pose-dependent geometry adjustments, such as wrinkles, have recently been made [18,57,59,60,114], although these methods are prone to overfitting to the body poses and motion sequences observed during training. A key problem is that most existing 3D avatar reconstruction approaches neglect to model the dynamics of loose garments in a physically accurate manner, leading to unrealistic fabric behavior and issues like self-penetration. To our knowledge, only the work of Xiang et al. [118] includes a physics-based approach to inverse rendering of digital humans, but their approach requires a tedious manual search process to find the parameters that model the dynamic behavior of their cloth fabrics with reasonable accuracy." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 131, + 641, + 482, + 665 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 131, + 641, + 482, + 665 + ], + "spans": [ + { + "bbox": [ + 131, + 641, + 482, + 665 + ], + "type": "text", + "content": "Here, we introduce PhysAvatar, a novel approach to 3D avatar reconstruction from multi-view video data. PhysAvatar combines inverse rendering with" + } + ] + } + ], + "index": 9 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "spans": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "type": "text", + "content": "2" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 166, + 91, + 240, + 102 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 166, + 91, + 240, + 102 + ], + "spans": [ + { + "bbox": [ + 166, + 91, + 240, + 102 + ], + "type": "text", + "content": "Yang Zheng et al." + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 1 + }, + { + "para_blocks": [ + { + "bbox": [ + 130, + 116, + 482, + 282 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 116, + 482, + 282 + ], + "spans": [ + { + "bbox": [ + 130, + 116, + 482, + 282 + ], + "type": "text", + "content": "\"inverse physics\" in a principled manner to not only estimate the shape and appearance of the avatar but also physical parameters, including density and stiffness, of the fabrics modeling their (loose) clothes. Our approach includes several parts: Given the reconstructed 3D mesh of the garment in one frame, we first leverage mesh-aligned 4D Gaussians [117] to track surface points of the garment across all frames of the input video sequence. This process establishes dense correspondences on the 3D garment. These data are used as supervision in an inverse physics step, where the physical parameters of the fabric are optimized using a finite-difference approach [50,123]. Finally, we employ a physically based inverse renderer [78] to jointly estimate ambient lighting and the surface material. By leveraging the refined geometry from our simulation step, the inverse renderer can effectively factor in pose-dependent effects, such as shadows caused by self-occlusion, resulting in accurate appearance reconstruction that enables the avatar to be rendered in novel lighting conditions." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 130, + 283, + 482, + 319 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 283, + 482, + 319 + ], + "spans": [ + { + "bbox": [ + 130, + 283, + 482, + 319 + ], + "type": "text", + "content": "PhysAvatar offers a comprehensive solution for reconstructing clothed human avatars from multi-view video data to perform novel view and motion synthesis with state-of-the-art realism. The key contributions of our work include:" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 135, + 319, + 480, + 401 + ], + "type": "list", + "angle": 0, + "index": 6, + "blocks": [ + { + "bbox": [ + 136, + 319, + 479, + 354 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 136, + 319, + 479, + 354 + ], + "spans": [ + { + "bbox": [ + 136, + 319, + 479, + 354 + ], + "type": "text", + "content": "1. The introduction of a new inverse rendering paradigm for avatars created from real-world captures that incorporates the physics of loose garments in a principled manner;" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 135, + 355, + 480, + 401 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 135, + 355, + 480, + 401 + ], + "spans": [ + { + "bbox": [ + 135, + 355, + 480, + 401 + ], + "type": "text", + "content": "2. A pipeline that includes accurate and efficient mesh reconstruction and tracking using 4D Gaussians; automatic optimization of the garments' physical material properties; and accurate appearance estimation using physically based inverse rendering." + } + ] + } + ], + "index": 5 + } + ], + "sub_type": "text" + }, + { + "bbox": [ + 130, + 403, + 482, + 439 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 403, + 482, + 439 + ], + "spans": [ + { + "bbox": [ + 130, + 403, + 482, + 439 + ], + "type": "text", + "content": "PhysAvatar demonstrates a novel path to learning physical scene properties from visual observations by incorporating inverse rendering with physics constraints. Code will be available on the project website." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 131, + 453, + 237, + 466 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 131, + 453, + 237, + 466 + ], + "spans": [ + { + "bbox": [ + 131, + 453, + 237, + 466 + ], + "type": "text", + "content": "2 Related Work" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 131, + 475, + 400, + 487 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 131, + 475, + 400, + 487 + ], + "spans": [ + { + "bbox": [ + 131, + 475, + 400, + 487 + ], + "type": "text", + "content": "2.1 Scene Reconstruction from Visual Observations" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 130, + 491, + 482, + 612 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 491, + 482, + 612 + ], + "spans": [ + { + "bbox": [ + 130, + 491, + 482, + 612 + ], + "type": "text", + "content": "Reconstructing 3D scenes from visual observations is grounded in classic approaches, such as Structure from Motion (SfM) [77, 93, 96]. Recently, this field has witnessed a paradigm shift towards inverse rendering techniques that estimate the shape and appearance of a scene from visual input [105, 106]. Among the many representations developed in this area are neural (radiance) fields [75, 95], neural volumes, surfaces, and signed distance fields [63, 73, 83, 94, 113, 122] as well as differentiable rasterization techniques for meshes [39, 61] and point clouds [40, 41, 48, 124]. Recent inverse rendering approaches are capable of modeling dynamic scenes [25, 84, 89, 108] and enable sophisticated post-capture deformations [36, 126]." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 131, + 624, + 261, + 636 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 131, + 624, + 261, + 636 + ], + "spans": [ + { + "bbox": [ + 131, + 624, + 261, + 636 + ], + "type": "text", + "content": "2.2 Animatable Avatars" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 130, + 641, + 481, + 666 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 641, + 481, + 666 + ], + "spans": [ + { + "bbox": [ + 130, + 641, + 481, + 666 + ], + "type": "text", + "content": "One central requirement for generating full-body digital humans is to enable animations driven by skeleton-level control [42, 71]. To this end, early works on" + } + ] + } + ], + "index": 12 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 399, + 91, + 447, + 102 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 399, + 91, + 447, + 102 + ], + "spans": [ + { + "bbox": [ + 399, + 91, + 447, + 102 + ], + "type": "text", + "content": "PhysAvatar" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 474, + 92, + 480, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 474, + 92, + 480, + 100 + ], + "spans": [ + { + "bbox": [ + 474, + 92, + 480, + 100 + ], + "type": "text", + "content": "3" + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 2 + }, + { + "para_blocks": [ + { + "bbox": [ + 130, + 116, + 482, + 236 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 116, + 482, + 236 + ], + "spans": [ + { + "bbox": [ + 130, + 116, + 482, + 236 + ], + "type": "text", + "content": "avatar reconstruction primarily relied on skinning techniques, which successfully model the dynamics of bodies with minimal or tight clothing [3,64]. Clothing is an integral part of everyday human appearance, captured by the saying \"clothes make the (wo)man\". Reconstructing their dynamic behavior from visual data, however, remains a challenge. Several different types of representations have been explored for clothed avatars, including meshes [68] with dynamic textures [4, 33, 130], neural surface [16, 90, 107] and radiance fields [14, 18, 24, 27, 44-46, 86, 87, 99, 131], point sets [66, 67, 69], and 3D Gaussians [34, 58, 82, 132]. Many of these works condition the deformation of clothing on the body pose and some predict future dynamics based on a small number of frames [32, 33, 82]." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 130, + 236, + 482, + 475 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 236, + 482, + 475 + ], + "spans": [ + { + "bbox": [ + 130, + 236, + 482, + 475 + ], + "type": "text", + "content": "We now discuss recent research most closely related to ours. TAVA [53] learns a deformable neural radiance field of humans or animals using a differentiable skinning method like SNARF [17]. ARAH [115] models 3D avatars using SDF representations and proposes an SDF-based volume rendering method to reconstruct the avatar in canonical space. GS-Avatar [34] achieves real-time animation by leveraging a learned 3D Gaussian predictor. Although these methods show reasonable effectiveness on humans wearing tight garments, they struggle with loose garments because their LBS-based deformation module cannot accommodate the complex dynamics of loose clothing. To address this limitation, Xiang et al. [118] incorporate a cloth simulator based on the eXtended Position Based Dynamics (XPBD) formulation [70] into a deep appearance model to reconstruct realistic dynamics and appearance. However, they require tedious manual adjustments of physical parameters to produce reasonable garment motion, and the code is not publicly available. Concurrent with our work, AniDress [14], while using physics-based simulation, only uses the simulation to produce a garment rigging model. This can cause inaccuracy in cloth dynamics modeling and unrealistic motion artifacts. In contrast, we estimate the physical parameters through a principled gradient-based inverse physics approach and accurately recover the shape and (relightable) appearance using a physically based inverse renderer, achieving high-fidelity results on novel motions." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 131, + 491, + 289, + 503 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 131, + 491, + 289, + 503 + ], + "spans": [ + { + "bbox": [ + 131, + 491, + 289, + 503 + ], + "type": "text", + "content": "2.3 Physics-Based Simulation" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 130, + 510, + 482, + 666 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 510, + 482, + 666 + ], + "spans": [ + { + "bbox": [ + 130, + 510, + 482, + 666 + ], + "type": "text", + "content": "Physics-based simulation of cloth [5, 62, 88, 103, 109, 111, 125, 127] has been extensively studied and is particularly useful for modeling complicated dynamic effects in interaction with human bodies, such as large deformations, wrinkling, and contact. In this work, we adopt the Codimensional Incremental Potential Contact (C-IPC) solver [52] for its robustness in handling complicated body-cloth collision by the log-barrier penalty function [51]. We refer readers to dedicated surveys [37, 54] for a comprehensive review, and focus on works that solve the inverse parameter estimation problem. One line of work uses specialized devices to simultaneously measure forces and deformation in a controlled setting [47, 74, 112, 128]. Our work falls into the category that estimates physical parameters directly from imagery of garments worn on human subjects [12, 31, 97]. For this purpose, some differentiable simulators have been developed and applied to the optimization problem [55], but only for specific types of solvers," + } + ] + } + ], + "index": 5 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "spans": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "type": "text", + "content": "4" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 166, + 91, + 239, + 102 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 166, + 91, + 239, + 102 + ], + "spans": [ + { + "bbox": [ + 166, + 91, + 239, + 102 + ], + "type": "text", + "content": "Yang Zheng et al." + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 3 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 161, + 128, + 449, + 243 + ], + "blocks": [ + { + "bbox": [ + 161, + 128, + 449, + 243 + ], + "lines": [ + { + "bbox": [ + 161, + 128, + 449, + 243 + ], + "spans": [ + { + "bbox": [ + 161, + 128, + 449, + 243 + ], + "type": "image", + "image_path": "3136adbf142bd66ecf1e005f6a2eae2e84bb73f29c49d10e81dd1a0e28a895e6.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 130, + 256, + 482, + 334 + ], + "lines": [ + { + "bbox": [ + 130, + 256, + 482, + 334 + ], + "spans": [ + { + "bbox": [ + 130, + 256, + 482, + 334 + ], + "type": "text", + "content": "Fig. 2: Method Overview: (a) PhysAvatar takes multi-view videos and an initial mesh as input. We first perform (b) dynamic mesh tracking (Sec. 3.1). The tacked mesh sequences are then used for (c) garment physics estimation with a physics simulator combined with gradient-based optimization (Sec. 3.2); (d) and appearance estimation through physics-based differentiable rendering (Sec. 3.3). At test time, (e) given a sequence of body poses (f), we simulate garment dynamics with the learned physics parameters and employ physics-based rendering to produce the final images." + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_caption" + } + ], + "index": 2 + }, + { + "bbox": [ + 130, + 357, + 482, + 561 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 357, + 482, + 561 + ], + "spans": [ + { + "bbox": [ + 130, + 357, + 482, + 561 + ], + "type": "text", + "content": "such as Projective Dynamics [56] and XPBD [98]. The heavy runtime cost of high-quality physics-based simulation motivates research in neural cloth simulation. Early works create datasets of garments simulated on human bodies [8,30], and then train the network to predict garment deformation in a supervised manner [19,81,91,110]. Recently, more attention has been paid to the self-supervised formulation that applies the elasticity energy defined in traditional simulation to predict garments, thereby training the network to act like a simulation solver, including both the (quasi-) static [9,11,22,49,92] and the dynamic [10,28] cases. Most of these works focus on making forward predictions that resemble traditional simulators without solving inverse parameter estimation on real-world garment data, except CaPhy [100]. CaPhy estimates the Saint Venant-Kirchhoff (StVK) elasticity parameters of clothing from 4D scans but only predicts pose-dependent clothing deformation in a quasi-static manner. In addition, all the works mentioned above treat simulation as a separate problem without modeling the photorealistic appearance. By comparison, our work builds avatars with physically based appearance and dynamics that are optimized for faithfulness to the real human capture in a holistic manner." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 132, + 579, + 202, + 593 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 579, + 202, + 593 + ], + "spans": [ + { + "bbox": [ + 132, + 579, + 202, + 593 + ], + "type": "text", + "content": "3 Method" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 130, + 605, + 482, + 666 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 605, + 482, + 666 + ], + "spans": [ + { + "bbox": [ + 130, + 605, + 482, + 666 + ], + "type": "text", + "content": "Given multi-view videos as input, our goal is to reconstruct a 3D avatar that can be realistically animated and rendered under novel motions and novel views. The core pipeline (Fig. 2) consists of three primary components: 1. Mesh tracking (Sec. 3.1)—given multi-view videos and a reconstructed mesh at the initial time step as input, we track the deformation of the geometry, which provides the" + } + ] + } + ], + "index": 6 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 399, + 91, + 447, + 101 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 399, + 91, + 447, + 101 + ], + "spans": [ + { + "bbox": [ + 399, + 91, + 447, + 101 + ], + "type": "text", + "content": "PhysAvatar" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 474, + 92, + 480, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 474, + 92, + 480, + 100 + ], + "spans": [ + { + "bbox": [ + 474, + 92, + 480, + 100 + ], + "type": "text", + "content": "5" + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 4 + }, + { + "para_blocks": [ + { + "bbox": [ + 130, + 116, + 482, + 213 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 116, + 482, + 213 + ], + "spans": [ + { + "bbox": [ + 130, + 116, + 482, + 213 + ], + "type": "text", + "content": "geometry reference for the ensuing physics simulation; 2. Physics-based dynamic modeling (Sec. 3.2)—with the tracked meshes as reference, we estimate the garments' physical properties using a physics simulator; the simulator, together with the estimated physical parameters are used to generate novel motion; 3. Physics-based appearance refinement (Sec. 3.3)—given the tracked geometry, we further enhance the appearance estimation through a physics-based differentiable renderer, which considers self-occlusion and thus eliminate artifacts such as baked-in shadows. In the following, we delve into the details of each component." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 131, + 231, + 237, + 243 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 131, + 231, + 237, + 243 + ], + "spans": [ + { + "bbox": [ + 131, + 231, + 237, + 243 + ], + "type": "text", + "content": "3.1 Mesh Tracking" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 130, + 251, + 482, + 348 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 251, + 482, + 348 + ], + "spans": [ + { + "bbox": [ + 130, + 251, + 482, + 348 + ], + "type": "text", + "content": "Given multiview videos of " + }, + { + "bbox": [ + 130, + 251, + 482, + 348 + ], + "type": "inline_equation", + "content": "T" + }, + { + "bbox": [ + 130, + 251, + 482, + 348 + ], + "type": "text", + "content": " frames and " + }, + { + "bbox": [ + 130, + 251, + 482, + 348 + ], + "type": "inline_equation", + "content": "C" + }, + { + "bbox": [ + 130, + 251, + 482, + 348 + ], + "type": "text", + "content": " views, " + }, + { + "bbox": [ + 130, + 251, + 482, + 348 + ], + "type": "inline_equation", + "content": "\\{\\mathbf{I}_{1:T,1:C}\\}" + }, + { + "bbox": [ + 130, + 251, + 482, + 348 + ], + "type": "text", + "content": ", and an initial mesh " + }, + { + "bbox": [ + 130, + 251, + 482, + 348 + ], + "type": "inline_equation", + "content": "(\\mathbf{V}_1,\\mathbf{F})" + }, + { + "bbox": [ + 130, + 251, + 482, + 348 + ], + "type": "text", + "content": ", with vertices " + }, + { + "bbox": [ + 130, + 251, + 482, + 348 + ], + "type": "inline_equation", + "content": "\\mathbf{V}_1" + }, + { + "bbox": [ + 130, + 251, + 482, + 348 + ], + "type": "text", + "content": " and faces " + }, + { + "bbox": [ + 130, + 251, + 482, + 348 + ], + "type": "inline_equation", + "content": "\\mathbf{F}" + }, + { + "bbox": [ + 130, + 251, + 482, + 348 + ], + "type": "text", + "content": ", that can be obtained via any existing static scene reconstruction methods, our goal is to accurately track the mesh deformation through the video sequence " + }, + { + "bbox": [ + 130, + 251, + 482, + 348 + ], + "type": "inline_equation", + "content": "\\mathbf{V}_{1:T}" + }, + { + "bbox": [ + 130, + 251, + 482, + 348 + ], + "type": "text", + "content": ", which will be used as reference in the subsequent physics simulation. However accurately tracking mesh deformation from visual observation is very challenging. Inspired by the robust and real-time tracking performance from dynamic 3D Gaussians [65], we use 3D Gaussians as a surrogate representation for mesh tracking." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 130, + 348, + 483, + 504 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 348, + 483, + 504 + ], + "spans": [ + { + "bbox": [ + 130, + 348, + 483, + 504 + ], + "type": "text", + "content": "We adapt 3D Gaussians to align with the reconstructed mesh surface. Following the definition of dynamic 3D Gaussians, at frame " + }, + { + "bbox": [ + 130, + 348, + 483, + 504 + ], + "type": "inline_equation", + "content": "t" + }, + { + "bbox": [ + 130, + 348, + 483, + 504 + ], + "type": "text", + "content": ", each Gaussian is parameterized with position " + }, + { + "bbox": [ + 130, + 348, + 483, + 504 + ], + "type": "inline_equation", + "content": "\\mathbf{p}_t \\in \\mathbb{R}^3" + }, + { + "bbox": [ + 130, + 348, + 483, + 504 + ], + "type": "text", + "content": ", rotation quaternion " + }, + { + "bbox": [ + 130, + 348, + 483, + 504 + ], + "type": "inline_equation", + "content": "\\mathbf{q}_t \\in \\mathbb{R}^4" + }, + { + "bbox": [ + 130, + 348, + 483, + 504 + ], + "type": "text", + "content": ", color " + }, + { + "bbox": [ + 130, + 348, + 483, + 504 + ], + "type": "inline_equation", + "content": "\\mathbf{c}_t \\in \\mathbb{R}^3" + }, + { + "bbox": [ + 130, + 348, + 483, + 504 + ], + "type": "text", + "content": ", scale " + }, + { + "bbox": [ + 130, + 348, + 483, + 504 + ], + "type": "inline_equation", + "content": "\\mathbf{s}_t \\in \\mathbb{R}^3" + }, + { + "bbox": [ + 130, + 348, + 483, + 504 + ], + "type": "text", + "content": " and opacity " + }, + { + "bbox": [ + 130, + 348, + 483, + 504 + ], + "type": "inline_equation", + "content": "o_t \\in \\mathbb{R}_+" + }, + { + "bbox": [ + 130, + 348, + 483, + 504 + ], + "type": "text", + "content": ". To couple the Gaussians with the mesh, we attach one Gaussian at the barycenter of each face " + }, + { + "bbox": [ + 130, + 348, + 483, + 504 + ], + "type": "inline_equation", + "content": "\\mathbf{f} \\in \\mathbf{F}" + }, + { + "bbox": [ + 130, + 348, + 483, + 504 + ], + "type": "text", + "content": ", and determine the Gaussians' quaternions from the attached triangle. Specifically, the rotation of each Gaussian is computed from the rotation of its local frame, whose " + }, + { + "bbox": [ + 130, + 348, + 483, + 504 + ], + "type": "inline_equation", + "content": "x-" + }, + { + "bbox": [ + 130, + 348, + 483, + 504 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 130, + 348, + 483, + 504 + ], + "type": "inline_equation", + "content": "z" + }, + { + "bbox": [ + 130, + 348, + 483, + 504 + ], + "type": "text", + "content": "-axis are defined using the longest triangle edge at " + }, + { + "bbox": [ + 130, + 348, + 483, + 504 + ], + "type": "inline_equation", + "content": "t = 1" + }, + { + "bbox": [ + 130, + 348, + 483, + 504 + ], + "type": "text", + "content": " and the face normal respectively. Furthermore, we set the last scaling factor, which corresponds to the scaling in face normal direction, to be a small constant value, ensuring the Gaussian lies on the mesh surface. The Gaussians can then be rendered in a differentiable manner [40] and we optimize their parameters by minimizing the loss [65] between the rendered images and the reference images:" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 194, + 514, + 481, + 536 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 194, + 514, + 481, + 536 + ], + "spans": [ + { + "bbox": [ + 194, + 514, + 481, + 536 + ], + "type": "interline_equation", + "content": "\\mathcal {L} _ {\\mathrm {r e n d e r}} = \\lambda \\left\\| \\mathbf {I} _ {i, t} - \\hat {\\mathbf {I}} _ {i, t} \\right\\| _ {1} + (1 - \\lambda) \\cdot \\operatorname {S S I M} \\left(\\mathbf {I} _ {i, t}, \\hat {\\mathbf {I}} _ {i, t}\\right), \\tag {1}", + "image_path": "57dfaed8880c307cd8f5e2fcfa99ce2327cc720ce3818d40f3cf8d032bc37ec0.jpg" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 130, + 545, + 482, + 616 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 545, + 482, + 616 + ], + "spans": [ + { + "bbox": [ + 130, + 545, + 482, + 616 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 130, + 545, + 482, + 616 + ], + "type": "inline_equation", + "content": "\\hat{\\mathbf{I}}_{i,t}" + }, + { + "bbox": [ + 130, + 545, + 482, + 616 + ], + "type": "text", + "content": " denotes the image rendered from the perspective of the " + }, + { + "bbox": [ + 130, + 545, + 482, + 616 + ], + "type": "inline_equation", + "content": "i" + }, + { + "bbox": [ + 130, + 545, + 482, + 616 + ], + "type": "text", + "content": "-th camera at time " + }, + { + "bbox": [ + 130, + 545, + 482, + 616 + ], + "type": "inline_equation", + "content": "t" + }, + { + "bbox": [ + 130, + 545, + 482, + 616 + ], + "type": "text", + "content": ". Note that since the Gaussians' position and rotation are bound to the mesh, the optimization of the mesh vertices will implicitly optimize the parameters of the Gaussians. At the first time step, we optimize color, scale, and opacity. In the following frames " + }, + { + "bbox": [ + 130, + 545, + 482, + 616 + ], + "type": "inline_equation", + "content": "1 < t \\leq T" + }, + { + "bbox": [ + 130, + 545, + 482, + 616 + ], + "type": "text", + "content": ", we fix the opacity and scale, and optimize Gaussian colors " + }, + { + "bbox": [ + 130, + 545, + 482, + 616 + ], + "type": "inline_equation", + "content": "\\mathbf{c}_t" + }, + { + "bbox": [ + 130, + 545, + 482, + 616 + ], + "type": "text", + "content": " and mesh vertices " + }, + { + "bbox": [ + 130, + 545, + 482, + 616 + ], + "type": "inline_equation", + "content": "\\mathbf{V}_t" + }, + { + "bbox": [ + 130, + 545, + 482, + 616 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 130, + 617, + 482, + 666 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 617, + 482, + 666 + ], + "spans": [ + { + "bbox": [ + 130, + 617, + 482, + 666 + ], + "type": "text", + "content": "In addition to the photometric loss, we apply the following regularization terms to preserve the local geometric features: 1) an isometry loss " + }, + { + "bbox": [ + 130, + 617, + 482, + 666 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_{\\mathrm{iso}}" + }, + { + "bbox": [ + 130, + 617, + 482, + 666 + ], + "type": "text", + "content": " [65] that preserves edge length through the training sequence, 2) a normal loss " + }, + { + "bbox": [ + 130, + 617, + 482, + 666 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_{\\mathrm{normal}}" + }, + { + "bbox": [ + 130, + 617, + 482, + 666 + ], + "type": "text", + "content": " that encourages smooth mesh surfaces, enhancing mesh deformation accuracy" + } + ] + } + ], + "index": 8 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "spans": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "type": "text", + "content": "6" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 166, + 91, + 240, + 102 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 166, + 91, + 240, + 102 + ], + "spans": [ + { + "bbox": [ + 166, + 91, + 240, + 102 + ], + "type": "text", + "content": "Yang Zheng et al." + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 5 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 159, + 115, + 222, + 189 + ], + "blocks": [ + { + "bbox": [ + 159, + 115, + 222, + 189 + ], + "lines": [ + { + "bbox": [ + 159, + 115, + 222, + 189 + ], + "spans": [ + { + "bbox": [ + 159, + 115, + 222, + 189 + ], + "type": "image", + "image_path": "840589d5a2624a2f068a0c847a95c5ddd61c20937f953f28d3236fad05acb589.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 182, + 190, + 197, + 198 + ], + "lines": [ + { + "bbox": [ + 182, + 190, + 197, + 198 + ], + "spans": [ + { + "bbox": [ + 182, + 190, + 197, + 198 + ], + "type": "text", + "content": "t=12" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_caption" + } + ], + "index": 2 + }, + { + "type": "image", + "bbox": [ + 238, + 125, + 264, + 189 + ], + "blocks": [ + { + "bbox": [ + 238, + 125, + 264, + 189 + ], + "lines": [ + { + "bbox": [ + 238, + 125, + 264, + 189 + ], + "spans": [ + { + "bbox": [ + 238, + 125, + 264, + 189 + ], + "type": "image", + "image_path": "15269b0c1622e5d39aac4763c334b086b1e9b370a6c84d42c11ad095ddf4250d.jpg" + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 131, + 199, + 480, + 232 + ], + "lines": [ + { + "bbox": [ + 131, + 199, + 480, + 232 + ], + "spans": [ + { + "bbox": [ + 131, + 199, + 480, + 232 + ], + "type": "text", + "content": "Fig. 3: Our method can robustly track a dynamic mesh from input images, providing accurate long-term correspondences. Here we show the rendered images overlaid with the Gaussian trajectories from the previous 12 frames and the optimized meshes." + } + ] + } + ], + "index": 11, + "angle": 0, + "type": "image_caption" + } + ], + "index": 4 + }, + { + "type": "image", + "bbox": [ + 272, + 129, + 293, + 189 + ], + "blocks": [ + { + "bbox": [ + 272, + 129, + 293, + 189 + ], + "lines": [ + { + "bbox": [ + 272, + 129, + 293, + 189 + ], + "spans": [ + { + "bbox": [ + 272, + 129, + 293, + 189 + ], + "type": "image", + "image_path": "d91cf663f0276150571c76b0686644a18c11dbbfd0cc8dac17e204a228e4ad5f.jpg" + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 258, + 190, + 275, + 198 + ], + "lines": [ + { + "bbox": [ + 258, + 190, + 275, + 198 + ], + "spans": [ + { + "bbox": [ + 258, + 190, + 275, + 198 + ], + "type": "text", + "content": "t=300" + } + ] + } + ], + "index": 6, + "angle": 0, + "type": "image_caption" + } + ], + "index": 5 + }, + { + "type": "image", + "bbox": [ + 312, + 126, + 383, + 189 + ], + "blocks": [ + { + "bbox": [ + 312, + 126, + 383, + 189 + ], + "lines": [ + { + "bbox": [ + 312, + 126, + 383, + 189 + ], + "spans": [ + { + "bbox": [ + 312, + 126, + 383, + 189 + ], + "type": "image", + "image_path": "2e36ae87752025754e9831984c0cf7856e52707cccf4346ca7cb4a0a4f335113.jpg" + } + ] + } + ], + "index": 7, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 336, + 190, + 354, + 198 + ], + "lines": [ + { + "bbox": [ + 336, + 190, + 354, + 198 + ], + "spans": [ + { + "bbox": [ + 336, + 190, + 354, + 198 + ], + "type": "text", + "content": "t=600" + } + ] + } + ], + "index": 8, + "angle": 0, + "type": "image_caption" + } + ], + "index": 7 + }, + { + "type": "image", + "bbox": [ + 394, + 125, + 454, + 189 + ], + "blocks": [ + { + "bbox": [ + 394, + 125, + 454, + 189 + ], + "lines": [ + { + "bbox": [ + 394, + 125, + 454, + 189 + ], + "spans": [ + { + "bbox": [ + 394, + 125, + 454, + 189 + ], + "type": "image", + "image_path": "cbb42d65500b56dbd675a54f09a977bd57658d40c306806de4dc910d040b0555.jpg" + } + ] + } + ], + "index": 9, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 413, + 190, + 432, + 198 + ], + "lines": [ + { + "bbox": [ + 413, + 190, + 432, + 198 + ], + "spans": [ + { + "bbox": [ + 413, + 190, + 432, + 198 + ], + "type": "text", + "content": "t=900" + } + ] + } + ], + "index": 10, + "angle": 0, + "type": "image_caption" + } + ], + "index": 9 + }, + { + "bbox": [ + 131, + 241, + 479, + 277 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 131, + 241, + 479, + 277 + ], + "spans": [ + { + "bbox": [ + 131, + 241, + 479, + 277 + ], + "type": "text", + "content": "and stability. Denoting the vertex and the triangle of the mesh at frame " + }, + { + "bbox": [ + 131, + 241, + 479, + 277 + ], + "type": "inline_equation", + "content": "t" + }, + { + "bbox": [ + 131, + 241, + 479, + 277 + ], + "type": "text", + "content": " as " + }, + { + "bbox": [ + 131, + 241, + 479, + 277 + ], + "type": "inline_equation", + "content": "\\mathbf{v}_t" + }, + { + "bbox": [ + 131, + 241, + 479, + 277 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 131, + 241, + 479, + 277 + ], + "type": "inline_equation", + "content": "\\mathbf{f}_t" + }, + { + "bbox": [ + 131, + 241, + 479, + 277 + ], + "type": "text", + "content": ", the normal of a face as " + }, + { + "bbox": [ + 131, + 241, + 479, + 277 + ], + "type": "inline_equation", + "content": "\\mathbf{n}(\\cdot)" + }, + { + "bbox": [ + 131, + 241, + 479, + 277 + ], + "type": "text", + "content": ", and the 1-ring neighbors " + }, + { + "bbox": [ + 131, + 241, + 479, + 277 + ], + "type": "inline_equation", + "content": "N(\\cdot)" + }, + { + "bbox": [ + 131, + 241, + 479, + 277 + ], + "type": "text", + "content": ", the losses are formally defined as:" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 200, + 285, + 480, + 316 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 200, + 285, + 480, + 316 + ], + "spans": [ + { + "bbox": [ + 200, + 285, + 480, + 316 + ], + "type": "interline_equation", + "content": "\\mathcal {L} _ {\\mathrm {i s o}} = \\frac {1}{| \\mathbf {V} |} \\sum_ {\\mathbf {v} \\in \\mathbf {V}} \\sum_ {\\mathbf {v} ^ {\\prime} \\in N (\\mathbf {v} _ {1})} \\frac {\\left(\\left\\| \\mathbf {v} _ {1} - \\mathbf {v} _ {1} ^ {\\prime} \\right\\| _ {2} - \\left\\| \\mathbf {v} _ {t} - \\mathbf {v} _ {t} ^ {\\prime} \\right\\| _ {2}\\right)}{| N (\\mathbf {v}) |}, \\tag {2}", + "image_path": "94e81c6c7ae5dee56cc5807ecb64fbb2eff0aa823e9cd25652567f2516b62af2.jpg" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 187, + 319, + 480, + 349 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 187, + 319, + 480, + 349 + ], + "spans": [ + { + "bbox": [ + 187, + 319, + 480, + 349 + ], + "type": "interline_equation", + "content": "\\mathcal {L} _ {\\text {n o r m a l}} = \\frac {1}{| \\mathbf {F} |} \\sum_ {\\mathbf {f} _ {t} \\in \\mathbf {F}} \\sum_ {\\mathbf {f} _ {t} ^ {\\prime} \\in N (\\mathbf {f} _ {t})} \\left(1 - \\mathbf {n} \\left(\\mathbf {f} _ {t}\\right) \\cdot \\mathbf {n} \\left(\\mathbf {f} _ {t} ^ {\\prime}\\right)\\right). \\tag {3}", + "image_path": "826a52609de525f0541e0e5ba371e894029074b62ef3ff6d6969c7e2daed6888.jpg" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 147, + 357, + 364, + 369 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 147, + 357, + 364, + 369 + ], + "spans": [ + { + "bbox": [ + 147, + 357, + 364, + 369 + ], + "type": "text", + "content": "The full loss used for mesh tracking is defined as:" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 223, + 380, + 480, + 392 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 223, + 380, + 480, + 392 + ], + "spans": [ + { + "bbox": [ + 223, + 380, + 480, + 392 + ], + "type": "interline_equation", + "content": "\\mathcal {L} _ {\\text {m e s h}} = \\mathcal {L} _ {\\text {r e n d e r}} + \\lambda_ {i} \\mathcal {L} _ {\\text {i s o}} + \\lambda_ {n} \\mathcal {L} _ {\\text {n o r m a l}}, \\tag {4}", + "image_path": "3b40c0f3e43c105e7451d58d4d4054cc6197b3d4502058a6c5c7caf786e99146.jpg" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 131, + 400, + 479, + 437 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 131, + 400, + 479, + 437 + ], + "spans": [ + { + "bbox": [ + 131, + 400, + 479, + 437 + ], + "type": "text", + "content": "where we set " + }, + { + "bbox": [ + 131, + 400, + 479, + 437 + ], + "type": "inline_equation", + "content": "\\lambda_{i} = 10" + }, + { + "bbox": [ + 131, + 400, + 479, + 437 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 131, + 400, + 479, + 437 + ], + "type": "inline_equation", + "content": "\\lambda_{n} = 0.1" + }, + { + "bbox": [ + 131, + 400, + 479, + 437 + ], + "type": "text", + "content": " during our training. With the introduced losses, our method can reconstruct a smooth mesh sequence with accurate correspondences, as shown in Fig. 3." + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 131, + 454, + 331, + 467 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 131, + 454, + 331, + 467 + ], + "spans": [ + { + "bbox": [ + 131, + 454, + 331, + 467 + ], + "type": "text", + "content": "3.2 Physics Based Dynamic Modeling" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 130, + 474, + 480, + 521 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 474, + 480, + 521 + ], + "spans": [ + { + "bbox": [ + 130, + 474, + 480, + 521 + ], + "type": "text", + "content": "Given the tracked meshes, " + }, + { + "bbox": [ + 130, + 474, + 480, + 521 + ], + "type": "inline_equation", + "content": "(\\mathbf{V}_{1:T},\\mathbf{F})" + }, + { + "bbox": [ + 130, + 474, + 480, + 521 + ], + "type": "text", + "content": ", we estimate the physical properties of the garments such that the simulation of their dynamics matches the reference tracked meshes. Once learned, these parameters can be used to simulate the garment under novel motions and novel views." + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 131, + 521, + 481, + 666 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 131, + 521, + 481, + 666 + ], + "spans": [ + { + "bbox": [ + 131, + 521, + 481, + 666 + ], + "type": "text", + "content": "Garment Material Model and Simulation. We model the elastodynamics of the garment by equipping each simulation mesh with the discrete shell hinge bending energy [29,102] and isotropic StVK elastic potential [15,20]. Specifically, we estimate density " + }, + { + "bbox": [ + 131, + 521, + 481, + 666 + ], + "type": "inline_equation", + "content": "\\rho" + }, + { + "bbox": [ + 131, + 521, + 481, + 666 + ], + "type": "text", + "content": ", membrane stiffness " + }, + { + "bbox": [ + 131, + 521, + 481, + 666 + ], + "type": "inline_equation", + "content": "\\kappa_{s}" + }, + { + "bbox": [ + 131, + 521, + 481, + 666 + ], + "type": "text", + "content": ", and bending stiffness " + }, + { + "bbox": [ + 131, + 521, + 481, + 666 + ], + "type": "inline_equation", + "content": "\\kappa_{b}" + }, + { + "bbox": [ + 131, + 521, + 481, + 666 + ], + "type": "text", + "content": " of the garment given the visual observations. For our generic use cases, it is reasonable to assume homogeneous physical properties for each garment piece, i.e., the density, membrane stiffness, and bending stiffness can be parameterized using scalar values, denoted as " + }, + { + "bbox": [ + 131, + 521, + 481, + 666 + ], + "type": "inline_equation", + "content": "\\rho" + }, + { + "bbox": [ + 131, + 521, + 481, + 666 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 131, + 521, + 481, + 666 + ], + "type": "inline_equation", + "content": "\\kappa_{s}" + }, + { + "bbox": [ + 131, + 521, + 481, + 666 + ], + "type": "text", + "content": ", and " + }, + { + "bbox": [ + 131, + 521, + 481, + 666 + ], + "type": "inline_equation", + "content": "\\kappa_{b}" + }, + { + "bbox": [ + 131, + 521, + 481, + 666 + ], + "type": "text", + "content": ". In choosing the simulators, we consider differentiability and quality. Unfortunately, the development of differentiable simulators suitable for our application remains an open research question. Current implementations, such as those documented in [56], cannot be directly applied due to limitations in modeling complex body colliders. Therefore, we propose to use" + } + ] + } + ], + "index": 20 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 399, + 91, + 447, + 101 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 399, + 91, + 447, + 101 + ], + "spans": [ + { + "bbox": [ + 399, + 91, + 447, + 101 + ], + "type": "text", + "content": "PhysAvatar" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 474, + 91, + 480, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 474, + 91, + 480, + 100 + ], + "spans": [ + { + "bbox": [ + 474, + 91, + 480, + 100 + ], + "type": "text", + "content": "7" + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 6 + }, + { + "para_blocks": [ + { + "bbox": [ + 130, + 116, + 482, + 212 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 116, + 482, + 212 + ], + "spans": [ + { + "bbox": [ + 130, + 116, + 482, + 212 + ], + "type": "text", + "content": "the state-of-the-art robust and stable cloth simulator C-IPC [52] that guarantees no interpenetration with frictional contact. While C-IPC is differentiable, it lacks an implementation of accurate analytical gradients; hence, we use finite difference methods to estimate the gradients for parameter updates. Specifically, we employ an open-source implementation of C-IPC [52] as our cloth simulator, which is a state-of-the-art simulator achieving high-fidelity garment simulation with stability and convergence (see [52] for a comprehensive review of the base algorithm and implementation)." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 130, + 213, + 493, + 357 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 213, + 493, + 357 + ], + "spans": [ + { + "bbox": [ + 130, + 213, + 493, + 357 + ], + "type": "text", + "content": "Garment Material Estimation. From the reconstructed mesh, we segment out the garment vertices to obtain ground truth garment mesh sequences " + }, + { + "bbox": [ + 130, + 213, + 493, + 357 + ], + "type": "inline_equation", + "content": "(\\widehat{\\mathbf{V}}_{1:T}^{g},\\widehat{\\mathbf{F}}^{g})" + }, + { + "bbox": [ + 130, + 213, + 493, + 357 + ], + "type": "text", + "content": ". The cloth simulator, " + }, + { + "bbox": [ + 130, + 213, + 493, + 357 + ], + "type": "inline_equation", + "content": "f(\\cdot)" + }, + { + "bbox": [ + 130, + 213, + 493, + 357 + ], + "type": "text", + "content": ", takes as input the current garment state—position and velocity " + }, + { + "bbox": [ + 130, + 213, + 493, + 357 + ], + "type": "inline_equation", + "content": "(\\mathbf{V}_t^g,\\dot{\\mathbf{V}}_t^g)" + }, + { + "bbox": [ + 130, + 213, + 493, + 357 + ], + "type": "text", + "content": "—the garment's physical properties " + }, + { + "bbox": [ + 130, + 213, + 493, + 357 + ], + "type": "inline_equation", + "content": "(\\rho ,\\kappa_s,\\kappa_b)" + }, + { + "bbox": [ + 130, + 213, + 493, + 357 + ], + "type": "text", + "content": ", the body collider " + }, + { + "bbox": [ + 130, + 213, + 493, + 357 + ], + "type": "inline_equation", + "content": "\\{\\mathbf{V}_t^C,\\mathbf{F}^C\\}" + }, + { + "bbox": [ + 130, + 213, + 493, + 357 + ], + "type": "text", + "content": ", and garment boundary vertices " + }, + { + "bbox": [ + 130, + 213, + 493, + 357 + ], + "type": "inline_equation", + "content": "\\mathbf{V}_{t + 1}^b" + }, + { + "bbox": [ + 130, + 213, + 493, + 357 + ], + "type": "text", + "content": ", and step size " + }, + { + "bbox": [ + 130, + 213, + 493, + 357 + ], + "type": "inline_equation", + "content": "\\Delta t" + }, + { + "bbox": [ + 130, + 213, + 493, + 357 + ], + "type": "text", + "content": ". The boundary vertices of the garment are points of the garment that are in permanent contact with the human body, such as the shoulder region for draped garments. We employ the SMPL-X body model [64] as the body collider " + }, + { + "bbox": [ + 130, + 213, + 493, + 357 + ], + "type": "inline_equation", + "content": "\\{\\mathbf{V}_{t + 1}^{C},\\mathbf{F}^{C}\\}" + }, + { + "bbox": [ + 130, + 213, + 493, + 357 + ], + "type": "text", + "content": ", extracting SMPL-X body shapes and poses from the video data during the preprocessing (see Sec. 4). The garment and the boundary vertices are manually annotated once on the initial mesh " + }, + { + "bbox": [ + 130, + 213, + 493, + 357 + ], + "type": "inline_equation", + "content": "\\mathbf{V}_1" + }, + { + "bbox": [ + 130, + 213, + 493, + 357 + ], + "type": "text", + "content": " for the entire simulation. The simulation function predicting the garment shape in the next frame can be written as below:" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 212, + 368, + 481, + 383 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 212, + 368, + 481, + 383 + ], + "spans": [ + { + "bbox": [ + 212, + 368, + 481, + 383 + ], + "type": "interline_equation", + "content": "\\mathbf {V} _ {t + 1} ^ {g} = f (\\mathbf {V} _ {t} ^ {g}, \\dot {\\mathbf {V}} _ {t} ^ {g}, \\mathbf {V} _ {t + 1} ^ {b}, \\mathbf {V} _ {t + 1} ^ {C}, \\rho , \\kappa_ {s}, \\kappa_ {b}, \\Delta t). \\tag {5}", + "image_path": "3261209df5853ce4c9c18c21968c481dd4107426b48ae4b17a22868cc91c2bfa.jpg" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 131, + 387, + 482, + 424 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 131, + 387, + 482, + 424 + ], + "spans": [ + { + "bbox": [ + 131, + 387, + 482, + 424 + ], + "type": "text", + "content": "We find the optimal physical parameters by minimizing the difference between the simulated and the reconstructed garment meshes. Formally, the optimization objective can be expressed as:" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 203, + 434, + 481, + 465 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 203, + 434, + 481, + 465 + ], + "spans": [ + { + "bbox": [ + 203, + 434, + 481, + 465 + ], + "type": "interline_equation", + "content": "\\min _ {\\left\\{\\rho , \\kappa_ {s}, \\kappa_ {b} \\right\\}} \\mathcal {L} _ {\\text {s i m}} (\\rho , \\kappa_ {s}, \\kappa_ {b}) = \\sum_ {t = 0} ^ {T} \\| \\mathbf {V} _ {t + 1} ^ {g} - \\widehat {\\mathbf {V}} _ {t + 1} ^ {g} \\| _ {2} ^ {2}, \\tag {6}", + "image_path": "864cfdc56296ec126b3df69df14ae1e4a3dcfe718ae3309abc0e8c2d25474693.jpg" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 131, + 476, + 482, + 537 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 131, + 476, + 482, + 537 + ], + "spans": [ + { + "bbox": [ + 131, + 476, + 482, + 537 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 131, + 476, + 482, + 537 + ], + "type": "inline_equation", + "content": "\\mathbf{V}_1^g = \\widehat{\\mathbf{V}}_1^g" + }, + { + "bbox": [ + 131, + 476, + 482, + 537 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 131, + 476, + 482, + 537 + ], + "type": "inline_equation", + "content": "\\mathbf{V}_{t + 1}^g" + }, + { + "bbox": [ + 131, + 476, + 482, + 537 + ], + "type": "text", + "content": " is computed by Eq. (5). We estimate gradients via a numerical approach using finite differences [50]. This method is particularly feasible in our context due to the low-dimensional nature of the optimization problem, which involves only three scalar parameters. We employ the forward difference formula for gradient estimation:" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 196, + 546, + 481, + 572 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 196, + 546, + 481, + 572 + ], + "spans": [ + { + "bbox": [ + 196, + 546, + 481, + 572 + ], + "type": "interline_equation", + "content": "\\frac {\\partial \\mathcal {L} _ {\\mathrm {s i m}}}{\\partial \\rho} \\approx (\\mathcal {L} _ {\\mathrm {s i m}} (\\rho + \\delta_ {\\rho}, \\kappa_ {s}, \\kappa_ {b}) - \\mathcal {L} _ {\\mathrm {s i m}} (\\rho , \\kappa_ {s}, \\kappa_ {b})) / \\delta_ {\\rho}. \\tag {7}", + "image_path": "af47e3f48026822c5237715c3afc6002cb233a23adf456f3ec11a864751e2596.jpg" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 130, + 579, + 481, + 616 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 579, + 481, + 616 + ], + "spans": [ + { + "bbox": [ + 130, + 579, + 481, + 616 + ], + "type": "text", + "content": "Where " + }, + { + "bbox": [ + 130, + 579, + 481, + 616 + ], + "type": "inline_equation", + "content": "\\delta_{\\rho}" + }, + { + "bbox": [ + 130, + 579, + 481, + 616 + ], + "type": "text", + "content": " is a hyperparameter for forward difference gradient estimation. The gradients for other parameters are estimated using the same approach, and the parameters are subsequently updated using the Adam optimizer." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 130, + 616, + 482, + 666 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 616, + 482, + 666 + ], + "spans": [ + { + "bbox": [ + 130, + 616, + 482, + 666 + ], + "type": "text", + "content": "Animation. At test time, we have the initial mesh " + }, + { + "bbox": [ + 130, + 616, + 482, + 666 + ], + "type": "inline_equation", + "content": "\\{\\mathbf{V}_1,\\mathbf{F}_1\\}" + }, + { + "bbox": [ + 130, + 616, + 482, + 666 + ], + "type": "text", + "content": " and we are given a sequence of novel human body poses, e.g., provided by skeleton parameters such as SMPL [64]. Using standard LBS, we obtain the boundary locations " + }, + { + "bbox": [ + 130, + 616, + 482, + 666 + ], + "type": "inline_equation", + "content": "\\mathbf{V}_{t + 1}^b" + }, + { + "bbox": [ + 130, + 616, + 482, + 666 + ], + "type": "text", + "content": ". We assume the initial garment is at rest, i.e., " + }, + { + "bbox": [ + 130, + 616, + 482, + 666 + ], + "type": "inline_equation", + "content": "\\dot{\\mathbf{V}}_0 = 0" + }, + { + "bbox": [ + 130, + 616, + 482, + 666 + ], + "type": "text", + "content": ". We use the same garment" + } + ] + } + ], + "index": 10 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "spans": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "type": "text", + "content": "8" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 166, + 91, + 240, + 102 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 166, + 91, + 240, + 102 + ], + "spans": [ + { + "bbox": [ + 166, + 91, + 240, + 102 + ], + "type": "text", + "content": "Yang Zheng et al." + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 7 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 157, + 144, + 232, + 219 + ], + "blocks": [ + { + "bbox": [ + 156, + 133, + 166, + 140 + ], + "lines": [ + { + "bbox": [ + 156, + 133, + 166, + 140 + ], + "spans": [ + { + "bbox": [ + 156, + 133, + 166, + 140 + ], + "type": "text", + "content": "(a)" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 157, + 144, + 232, + 219 + ], + "lines": [ + { + "bbox": [ + 157, + 144, + 232, + 219 + ], + "spans": [ + { + "bbox": [ + 157, + 144, + 232, + 219 + ], + "type": "image", + "image_path": "99002568d07cf4242785855ef892d1629237b007a2a03966ffc5e73a81dee103.jpg" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_body" + } + ], + "index": 3 + }, + { + "type": "image", + "bbox": [ + 235, + 144, + 310, + 219 + ], + "blocks": [ + { + "bbox": [ + 162, + 224, + 222, + 234 + ], + "lines": [ + { + "bbox": [ + 162, + 224, + 222, + 234 + ], + "spans": [ + { + "bbox": [ + 162, + 224, + 222, + 234 + ], + "type": "text", + "content": "Before Optimization" + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 235, + 144, + 310, + 219 + ], + "lines": [ + { + "bbox": [ + 235, + 144, + 310, + 219 + ], + "spans": [ + { + "bbox": [ + 235, + 144, + 310, + 219 + ], + "type": "image", + "image_path": "f351861199c87b93e8d1dd80680c976e5c055bce9882210fc89cc0cc9bb7f3fa.jpg" + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 246, + 224, + 300, + 234 + ], + "lines": [ + { + "bbox": [ + 246, + 224, + 300, + 234 + ], + "spans": [ + { + "bbox": [ + 246, + 224, + 300, + 234 + ], + "type": "text", + "content": "After Optimization" + } + ] + } + ], + "index": 6, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 311, + 133, + 321, + 141 + ], + "lines": [ + { + "bbox": [ + 311, + 133, + 321, + 141 + ], + "spans": [ + { + "bbox": [ + 311, + 133, + 321, + 141 + ], + "type": "text", + "content": "(b)" + } + ] + } + ], + "index": 7, + "angle": 0, + "type": "image_caption" + } + ], + "index": 5 + }, + { + "type": "image", + "bbox": [ + 325, + 129, + 359, + 221 + ], + "blocks": [ + { + "bbox": [ + 325, + 129, + 359, + 221 + ], + "lines": [ + { + "bbox": [ + 325, + 129, + 359, + 221 + ], + "spans": [ + { + "bbox": [ + 325, + 129, + 359, + 221 + ], + "type": "image", + "image_path": "98753764e95fd30f8733f407016da0016f595eac4b8da8dc62737c5588e1bb27.jpg" + } + ] + } + ], + "index": 8, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 319, + 224, + 359, + 234 + ], + "lines": [ + { + "bbox": [ + 319, + 224, + 359, + 234 + ], + "spans": [ + { + "bbox": [ + 319, + 224, + 359, + 234 + ], + "type": "text", + "content": "PBR w/o Opt" + } + ] + } + ], + "index": 9, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 130, + 256, + 480, + 312 + ], + "lines": [ + { + "bbox": [ + 130, + 256, + 480, + 312 + ], + "spans": [ + { + "bbox": [ + 130, + 256, + 480, + 312 + ], + "type": "text", + "content": "Fig. 4: Ablation study on appearance estimation: (a) Initial texture map " + }, + { + "bbox": [ + 130, + 256, + 480, + 312 + ], + "type": "inline_equation", + "content": "\\mathbf{T}" + }, + { + "bbox": [ + 130, + 256, + 480, + 312 + ], + "type": "text", + "content": " extracted from Gaussian splatting (left) has baked-in shadows highlighted in the red boxes; post-optimization (right), the baked-in shadows are substantially removed. (b) Rendering comparisons demonstrate that our method with the optimized texture map more closely aligns with the ground truth." + } + ] + } + ], + "index": 14, + "angle": 0, + "type": "image_caption" + } + ], + "index": 8 + }, + { + "type": "image", + "bbox": [ + 371, + 129, + 406, + 221 + ], + "blocks": [ + { + "bbox": [ + 371, + 129, + 406, + 221 + ], + "lines": [ + { + "bbox": [ + 371, + 129, + 406, + 221 + ], + "spans": [ + { + "bbox": [ + 371, + 129, + 406, + 221 + ], + "type": "image", + "image_path": "4bd4b31a34cc91e0dfa60e2c1f11cb7b0be21791b202ad017a4675c279128e38.jpg" + } + ] + } + ], + "index": 10, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 371, + 224, + 402, + 233 + ], + "lines": [ + { + "bbox": [ + 371, + 224, + 402, + 233 + ], + "spans": [ + { + "bbox": [ + 371, + 224, + 402, + 233 + ], + "type": "text", + "content": "PBR Ours" + } + ] + } + ], + "index": 11, + "angle": 0, + "type": "image_caption" + } + ], + "index": 10 + }, + { + "type": "image", + "bbox": [ + 417, + 129, + 453, + 221 + ], + "blocks": [ + { + "bbox": [ + 417, + 129, + 453, + 221 + ], + "lines": [ + { + "bbox": [ + 417, + 129, + 453, + 221 + ], + "spans": [ + { + "bbox": [ + 417, + 129, + 453, + 221 + ], + "type": "image", + "image_path": "c5b67a8b8e3f2f0dd26d90e77a0c9d7ccde965216f81b6c363acdb4b4bbd9b7f.jpg" + } + ] + } + ], + "index": 12, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 426, + 224, + 437, + 233 + ], + "lines": [ + { + "bbox": [ + 426, + 224, + 437, + 233 + ], + "spans": [ + { + "bbox": [ + 426, + 224, + 437, + 233 + ], + "type": "text", + "content": "GT" + } + ] + } + ], + "index": 13, + "angle": 0, + "type": "image_caption" + } + ], + "index": 12 + }, + { + "bbox": [ + 130, + 335, + 479, + 371 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 335, + 479, + 371 + ], + "spans": [ + { + "bbox": [ + 130, + 335, + 479, + 371 + ], + "type": "text", + "content": "segmentation, boundary points, and SMPL body collider points as during training, and use Eq. (5) to simulate the deformation of the garment. This is then combined with the body, driven by LBS, to create the holistic animation." + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 131, + 388, + 346, + 400 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 131, + 388, + 346, + 400 + ], + "spans": [ + { + "bbox": [ + 131, + 388, + 346, + 400 + ], + "type": "text", + "content": "3.3 Physics Based Appearance Modeling" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 130, + 407, + 482, + 562 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 407, + 482, + 562 + ], + "spans": [ + { + "bbox": [ + 130, + 407, + 482, + 562 + ], + "type": "text", + "content": "While tracking with dynamic Gaussians jointly estimates the garments appearance, it does not account for light-surface interactions, consequently the reconstructed appearance contains baked-in shadows, causing uncanny visual artifacts when rendered in novel views and motions (see Fig. 4). In this step, given the mesh sequence " + }, + { + "bbox": [ + 130, + 407, + 482, + 562 + ], + "type": "inline_equation", + "content": "(\\mathbf{V}_{1:T},\\mathbf{F})" + }, + { + "bbox": [ + 130, + 407, + 482, + 562 + ], + "type": "text", + "content": " and the per-frame mesh colors " + }, + { + "bbox": [ + 130, + 407, + 482, + 562 + ], + "type": "inline_equation", + "content": "\\mathbf{c}_{1:T}" + }, + { + "bbox": [ + 130, + 407, + 482, + 562 + ], + "type": "text", + "content": " reconstructed from Section 3.1, we utilize Mitsuba3 [78]—a physics-based differentiable Monte Carlo renderer—to enhance the appearance reconstruction. For this purpose, we use a time-invariant diffuse texture map " + }, + { + "bbox": [ + 130, + 407, + 482, + 562 + ], + "type": "inline_equation", + "content": "\\mathbf{T}" + }, + { + "bbox": [ + 130, + 407, + 482, + 562 + ], + "type": "text", + "content": " to model the appearance. The value " + }, + { + "bbox": [ + 130, + 407, + 482, + 562 + ], + "type": "inline_equation", + "content": "\\mathbf{T}" + }, + { + "bbox": [ + 130, + 407, + 482, + 562 + ], + "type": "text", + "content": " is initialized from the average Gaussian colors from the tracking step (Section 3.1), " + }, + { + "bbox": [ + 130, + 407, + 482, + 562 + ], + "type": "inline_equation", + "content": "\\Sigma_{1 < t\\leq T}(\\mathbf{c}(\\mathbf{x})_t) / T" + }, + { + "bbox": [ + 130, + 407, + 482, + 562 + ], + "type": "text", + "content": ". For the environment lighting, since the capture studio is typically equipped with uniform point light arrays, we can approximate the environment lighting as an ambient illumination, parameterized with a global " + }, + { + "bbox": [ + 130, + 407, + 482, + 562 + ], + "type": "inline_equation", + "content": "L_{a}\\in \\mathbb{R}^{3}" + }, + { + "bbox": [ + 130, + 407, + 482, + 562 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 130, + 563, + 481, + 599 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 563, + 481, + 599 + ], + "spans": [ + { + "bbox": [ + 130, + 563, + 481, + 599 + ], + "type": "text", + "content": "We optimize for " + }, + { + "bbox": [ + 130, + 563, + 481, + 599 + ], + "type": "inline_equation", + "content": "\\mathbf{T}" + }, + { + "bbox": [ + 130, + 563, + 481, + 599 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 130, + 563, + 481, + 599 + ], + "type": "inline_equation", + "content": "L_{a}" + }, + { + "bbox": [ + 130, + 563, + 481, + 599 + ], + "type": "text", + "content": " using the photometric loss defined in Eq. (8) for all training views and frames " + }, + { + "bbox": [ + 130, + 563, + 481, + 599 + ], + "type": "inline_equation", + "content": "(i,t)" + }, + { + "bbox": [ + 130, + 563, + 481, + 599 + ], + "type": "text", + "content": ", where " + }, + { + "bbox": [ + 130, + 563, + 481, + 599 + ], + "type": "inline_equation", + "content": "\\mathcal{R}" + }, + { + "bbox": [ + 130, + 563, + 481, + 599 + ], + "type": "text", + "content": " denotes the rendering function and " + }, + { + "bbox": [ + 130, + 563, + 481, + 599 + ], + "type": "inline_equation", + "content": "c_{i,t}" + }, + { + "bbox": [ + 130, + 563, + 481, + 599 + ], + "type": "text", + "content": " the camera pose:" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 205, + 608, + 481, + 634 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 205, + 608, + 481, + 634 + ], + "spans": [ + { + "bbox": [ + 205, + 608, + 481, + 634 + ], + "type": "interline_equation", + "content": "\\mathcal {L} _ {\\mathrm {r g b}} = \\min _ {\\left\\{\\mathbf {T}, L _ {a} \\right\\}} \\sum_ {i, t} \\left\\| \\mathbf {I} _ {i, t} - \\mathcal {R} \\left(\\mathbf {T}, L _ {a}, \\mathbf {V} _ {t}, c _ {i, t}\\right) \\right\\| _ {2} ^ {2}. \\tag {8}", + "image_path": "31fd92d6c577108272fe0a5eb9b320a02e59f502590fc2f016436ab13a2e1262.jpg" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 131, + 641, + 479, + 666 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 131, + 641, + 479, + 666 + ], + "spans": [ + { + "bbox": [ + 131, + 641, + 479, + 666 + ], + "type": "text", + "content": "At test time, we can swap the ambient light with more complex illuminations, and produce realistic rendering in arbitrary views and body poses (see Sec. 4.4)." + } + ] + } + ], + "index": 20 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 399, + 91, + 447, + 101 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 399, + 91, + 447, + 101 + ], + "spans": [ + { + "bbox": [ + 399, + 91, + 447, + 101 + ], + "type": "text", + "content": "PhysAvatar" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 474, + 92, + 481, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 474, + 92, + 481, + 100 + ], + "spans": [ + { + "bbox": [ + 474, + 92, + 481, + 100 + ], + "type": "text", + "content": "9" + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 8 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 151, + 114, + 465, + 308 + ], + "blocks": [ + { + "bbox": [ + 151, + 114, + 465, + 308 + ], + "lines": [ + { + "bbox": [ + 151, + 114, + 465, + 308 + ], + "spans": [ + { + "bbox": [ + 151, + 114, + 465, + 308 + ], + "type": "image", + "image_path": "5570d9eb58ea3748a24e2c2903800d6034251f17e7e56c6cf866db386ed8bf76.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 131, + 308, + 482, + 342 + ], + "lines": [ + { + "bbox": [ + 131, + 308, + 482, + 342 + ], + "spans": [ + { + "bbox": [ + 131, + 308, + 482, + 342 + ], + "type": "text", + "content": "Fig. 5: Qualitative results on test poses from the ActorHQ [38] dataset. Our method PhysAvatar achieves state-of-the-art performance in terms of geometry detail and appearance modeling." + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_caption" + } + ], + "index": 2 + }, + { + "bbox": [ + 132, + 364, + 230, + 378 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 364, + 230, + 378 + ], + "spans": [ + { + "bbox": [ + 132, + 364, + 230, + 378 + ], + "type": "text", + "content": "4 Experiments" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 131, + 388, + 481, + 413 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 131, + 388, + 481, + 413 + ], + "spans": [ + { + "bbox": [ + 131, + 388, + 481, + 413 + ], + "type": "text", + "content": "In this section, we explain our experimental setup and results. We recommend watching the supplementary video for better visualization." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 132, + 430, + 259, + 443 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 430, + 259, + 443 + ], + "spans": [ + { + "bbox": [ + 132, + 430, + 259, + 443 + ], + "type": "text", + "content": "4.1 Experimental setup" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 130, + 449, + 482, + 605 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 449, + 482, + 605 + ], + "spans": [ + { + "bbox": [ + 130, + 449, + 482, + 605 + ], + "type": "text", + "content": "Data Preparation. We choose four characters from the ActorHQ [35] dataset for our experiments, including two female characters with loose dresses, one female character wearing a tank top and loose-fitting pants, and one male character with a short-sleeved shirt and khaki shorts, as shown in Fig. 1. For each character, we use the ground truth mesh provided by the dataset and re-mesh it to obtain a simulation-ready initial mesh, i.e., one where the triangles have similar areas, to enhance the performance of garment simulation. We then segment the garment, define the boundary vertices of the garment for simulation purposes, and unwrap the whole mesh to get the UV map, all performed in Blender [13]. Since we observe that the human pose estimations from the dataset are quite noisy, we fit the SMPL-X [85] body mesh using multi-view videos by minimizing the distance between the projected SMPL-X joints and 2D keypoint detection results from DWPose [21, 121]." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 130, + 606, + 482, + 666 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 606, + 482, + 666 + ], + "spans": [ + { + "bbox": [ + 130, + 606, + 482, + 666 + ], + "type": "text", + "content": "Implementation Details. For training, we utilize the entire set of 160 camera views from the ActorHQ [35] dataset. Although the dataset includes approximately 2000 frames for each character, we selectively use 24 frames with large movements for estimating garment material parameters and 200 frames for optimizing the texture map, which leads to good results while conserving compu" + } + ] + } + ], + "index": 8 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "spans": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "type": "text", + "content": "10" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 166, + 91, + 239, + 102 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 166, + 91, + 239, + 102 + ], + "spans": [ + { + "bbox": [ + 166, + 91, + 239, + 102 + ], + "type": "text", + "content": "Yang Zheng et al." + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 9 + }, + { + "para_blocks": [ + { + "type": "table", + "bbox": [ + 147, + 157, + 465, + 237 + ], + "blocks": [ + { + "bbox": [ + 131, + 114, + 482, + 148 + ], + "lines": [ + { + "bbox": [ + 131, + 114, + 482, + 148 + ], + "spans": [ + { + "bbox": [ + 131, + 114, + 482, + 148 + ], + "type": "text", + "content": "Table 1: Quantitative comparison with baselines. PhysAvatar outperforms all baselines with respect to geometry accuracy. We achieve the best or competitive results in appearance metrics. See Fig. 5 and Fig. 6 for qualitative comparisons." + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 147, + 157, + 465, + 237 + ], + "lines": [ + { + "bbox": [ + 147, + 157, + 465, + 237 + ], + "spans": [ + { + "bbox": [ + 147, + 157, + 465, + 237 + ], + "type": "table", + "html": "
MethodGeometryAppearance
CD (×103) (↓)F-Score (↑)LPIPS (↓)PSNR (↑)SSIM (↑)
RAAH [115]1.1286.10.05528.60.957
TAVA [53]0.6692.30.05129.60.962
GS-Avatar [34]0.9189.40.04430.60.962
PhysAvatar (ours)0.5592.90.03530.20.957
", + "image_path": "8571eabe21ef553f54ba1f10d6646325f19039ff79b04b135850430f385cfa36.jpg" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "table_body" + } + ], + "index": 3 + }, + { + "bbox": [ + 130, + 261, + 482, + 344 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 261, + 482, + 344 + ], + "spans": [ + { + "bbox": [ + 130, + 261, + 482, + 344 + ], + "type": "text", + "content": "tational resources. For evaluation, we choose 200 frames for each character, with novel motions unseen during training. At test time, given a motion sequence defined by SMPL-X [85], we use an LBS weight inpainting algorithm [2] to obtain the skinning weights of the human body with respect to the SMPL-X skeleton. We animate the human body using LBS and simulate the garment, which is driven by the boundary vertices and utilizes the SMPL-X body mesh as the collider. More details are available at supplementary document." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 130, + 345, + 483, + 430 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 345, + 483, + 430 + ], + "spans": [ + { + "bbox": [ + 130, + 345, + 483, + 430 + ], + "type": "text", + "content": "Baselines. We benchmark our method and current open-sourced state-of-the-art methods modeling full-body avatars, including ARAH [115] which leverages SDF-based volume rendering to learn the canonical representation of the avatar, TAVA [53] which learns the canonical radiance field [75] and skinning weight of the character, and GS-Avatar [34] based on a pose-conditioned neural network which predicts the parameters of 3D Gaussians [40]. For all baselines, we use the official public codebase and extend them to our dense-view setting." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 132, + 450, + 222, + 462 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 450, + 222, + 462 + ], + "spans": [ + { + "bbox": [ + 132, + 450, + 222, + 462 + ], + "type": "text", + "content": "4.2 Comparison" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 130, + 473, + 482, + 508 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 473, + 482, + 508 + ], + "spans": [ + { + "bbox": [ + 130, + 473, + 482, + 508 + ], + "type": "text", + "content": "We compare PhysAvatar against current state-of-the-art approaches in terms of geometry and appearance quality across test pose sequences, as shown in Table 1, Fig. 5, and Fig. 6." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 130, + 510, + 482, + 666 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 510, + 482, + 666 + ], + "spans": [ + { + "bbox": [ + 130, + 510, + 482, + 666 + ], + "type": "text", + "content": "Geometry Evaluation. For geometry evaluation, we employ the mean Chamfer distance [73] and the mean F-Score [104] at " + }, + { + "bbox": [ + 130, + 510, + 482, + 666 + ], + "type": "inline_equation", + "content": "\\tau = 0.001" + }, + { + "bbox": [ + 130, + 510, + 482, + 666 + ], + "type": "text", + "content": " as quantitative metrics. PhysAvatar outperforms all baseline according to these metrics, demonstrating a superior capability in accurately capturing the dynamic nature of garment geometry over time, as shown in Table 1. As shown in Fig. 5, PhysAvatar generates more accurate garment deformation and better captures fine wrinkle details, thanks to the integration of a physics-based simulator. This contrasts with all baseline methods [34,53,115], which do not explicitly model garment dynamics, but instead depend on learning quasi-static, pose-conditioned skinning weights or deformations. Additionally, we show results using motion data from the AMASS dataset [72]. PhysAvatar demonstrates its robustness by consistently synthesizing realistic garment deformation with fine details, while baseline methods sometimes fail to capture large deformations in loose garments (Fig. 6)." + } + ] + } + ], + "index": 8 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 399, + 91, + 447, + 102 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 399, + 91, + 447, + 102 + ], + "spans": [ + { + "bbox": [ + 399, + 91, + 447, + 102 + ], + "type": "text", + "content": "PhysAvatar" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 470, + 91, + 480, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 470, + 91, + 480, + 100 + ], + "spans": [ + { + "bbox": [ + 470, + 91, + 480, + 100 + ], + "type": "text", + "content": "11" + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 10 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 159, + 115, + 302, + 200 + ], + "blocks": [ + { + "bbox": [ + 159, + 115, + 302, + 200 + ], + "lines": [ + { + "bbox": [ + 159, + 115, + 302, + 200 + ], + "spans": [ + { + "bbox": [ + 159, + 115, + 302, + 200 + ], + "type": "image", + "image_path": "a4694460856dc5457d22f15142cdd2c411dfe3b9168277d8280779a961b74105.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_body" + } + ], + "index": 2 + }, + { + "type": "image", + "bbox": [ + 309, + 114, + 455, + 200 + ], + "blocks": [ + { + "bbox": [ + 309, + 114, + 455, + 200 + ], + "lines": [ + { + "bbox": [ + 309, + 114, + 455, + 200 + ], + "spans": [ + { + "bbox": [ + 309, + 114, + 455, + 200 + ], + "type": "image", + "image_path": "cf969c606f5cdd7c2d7dcc961e18109834fd4fb66c3d72f6ef0a5ee6d4836203.jpg" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_body" + } + ], + "index": 3 + }, + { + "type": "image", + "bbox": [ + 159, + 202, + 302, + 285 + ], + "blocks": [ + { + "bbox": [ + 159, + 202, + 302, + 285 + ], + "lines": [ + { + "bbox": [ + 159, + 202, + 302, + 285 + ], + "spans": [ + { + "bbox": [ + 159, + 202, + 302, + 285 + ], + "type": "image", + "image_path": "8c1ef83398f094ea9ad4b3222dc39327fce742facd99be65bbf934a3f0cacb69.jpg" + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 130, + 289, + 482, + 323 + ], + "lines": [ + { + "bbox": [ + 130, + 289, + 482, + 323 + ], + "spans": [ + { + "bbox": [ + 130, + 289, + 482, + 323 + ], + "type": "text", + "content": "Fig. 6: Here we show animation results of current state-of-the-art methods including ARAH [115], TAVA [53] and GS-Avatar [34], and our results on test motions from AMASS [72] dataset. Images are rendered from novel views. Please zoom in for details." + } + ] + } + ], + "index": 6, + "angle": 0, + "type": "image_caption" + } + ], + "index": 4 + }, + { + "type": "image", + "bbox": [ + 309, + 202, + 455, + 285 + ], + "blocks": [ + { + "bbox": [ + 309, + 202, + 455, + 285 + ], + "lines": [ + { + "bbox": [ + 309, + 202, + 455, + 285 + ], + "spans": [ + { + "bbox": [ + 309, + 202, + 455, + 285 + ], + "type": "image", + "image_path": "099a6fd63455150aa4c35ad53d2755dad40bebd9895d1b7cd19c880bd9a1788f.jpg" + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "image_body" + } + ], + "index": 5 + }, + { + "bbox": [ + 130, + 344, + 482, + 465 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 344, + 482, + 465 + ], + "spans": [ + { + "bbox": [ + 130, + 344, + 482, + 465 + ], + "type": "text", + "content": "Appearance Evaluation. For appearance evaluation, we utilize Peak Signal-to-Noise Ratio (PSNR), Structural Similarity Index Measure (SSIM), and the Learned Perceptual Image Patch Similarity (LPIPS) [129]. As detailed in Table 1, PhysAvatar surpasses all baselines in LPIPS, a metric shown to be better aligned with human perception [129]. While we achieve the second-best results in PSNR and SSIM, it is important to note that even minor misalignments in images can significantly impact these metrics, potentially not fully reflecting the quality of the renderings. As demonstrated in Fig. 5, PhysAvatar can capture more high-frequency details in facial features and garment textures, thereby creating images that are visually richer and more detailed, compared to the baseline approaches." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 132, + 480, + 205, + 491 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 480, + 205, + 491 + ], + "spans": [ + { + "bbox": [ + 132, + 480, + 205, + 491 + ], + "type": "text", + "content": "4.3 Ablation" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 130, + 498, + 481, + 533 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 498, + 481, + 533 + ], + "spans": [ + { + "bbox": [ + 130, + 498, + 481, + 533 + ], + "type": "text", + "content": "We perform ablation studies to assess the impact of each component in our pipeline on the overall performance. For this, we focus on two female characters from the ActorHQ dataset, specifically chosen for their loose garments." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 130, + 534, + 482, + 666 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 534, + 482, + 666 + ], + "spans": [ + { + "bbox": [ + 130, + 534, + 482, + 666 + ], + "type": "text", + "content": "Garment physics estimation and modeling. We ablate our approach against two variants: one employing Linear Blend Skinning (LBS) for garment deformation and another utilizing a physics simulator for garment dynamics without optimized parameters, henceforth referred to as 'w/o physics', which employs random garment parameters. As shown in Table 2, both alternatives exhibit worse performance across all geometric metrics. Further, as shown in Fig. 7, LBS fails to produce plausible deformations for loose garments, such as dresses. While simulation with random parameters yields plausible garment dynamics, they do not accurately reflect real-world observations. In contrast, our method, with learned physics parameters, achieves deformations that closely align with the reference, demonstrating its effectiveness." + } + ] + } + ], + "index": 10 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "spans": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "type": "text", + "content": "12" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 166, + 91, + 239, + 102 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 166, + 91, + 239, + 102 + ], + "spans": [ + { + "bbox": [ + 166, + 91, + 239, + 102 + ], + "type": "text", + "content": "Yang Zheng et al." + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 11 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 141, + 114, + 217, + 192 + ], + "blocks": [ + { + "bbox": [ + 141, + 114, + 217, + 192 + ], + "lines": [ + { + "bbox": [ + 141, + 114, + 217, + 192 + ], + "spans": [ + { + "bbox": [ + 141, + 114, + 217, + 192 + ], + "type": "image", + "image_path": "17d588cdd4eb58a72fd1aec8fa865a1c03bca182f33fd92bedea656cc429ec63.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 167, + 194, + 186, + 204 + ], + "lines": [ + { + "bbox": [ + 167, + 194, + 186, + 204 + ], + "spans": [ + { + "bbox": [ + 167, + 194, + 186, + 204 + ], + "type": "text", + "content": "LBS" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_caption" + } + ], + "index": 2 + }, + { + "type": "image", + "bbox": [ + 225, + 114, + 302, + 192 + ], + "blocks": [ + { + "bbox": [ + 225, + 114, + 302, + 192 + ], + "lines": [ + { + "bbox": [ + 225, + 114, + 302, + 192 + ], + "spans": [ + { + "bbox": [ + 225, + 114, + 302, + 192 + ], + "type": "image", + "image_path": "1a8a93f90bd59e3c76bb21043bb974199037f1a795c3fd3c63e20a7251ff52f3.jpg" + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 238, + 194, + 283, + 205 + ], + "lines": [ + { + "bbox": [ + 238, + 194, + 283, + 205 + ], + "spans": [ + { + "bbox": [ + 238, + 194, + 283, + 205 + ], + "type": "text", + "content": "w/o Physics" + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "image_caption" + } + ], + "index": 4 + }, + { + "type": "image", + "bbox": [ + 309, + 114, + 386, + 192 + ], + "blocks": [ + { + "bbox": [ + 309, + 114, + 386, + 192 + ], + "lines": [ + { + "bbox": [ + 309, + 114, + 386, + 192 + ], + "spans": [ + { + "bbox": [ + 309, + 114, + 386, + 192 + ], + "type": "image", + "image_path": "a1da5aff1a331a8c48de8c66c5dc38e71c1b6743fc0c1988bb9092a576563771.jpg" + } + ] + } + ], + "index": 6, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 335, + 194, + 354, + 204 + ], + "lines": [ + { + "bbox": [ + 335, + 194, + 354, + 204 + ], + "spans": [ + { + "bbox": [ + 335, + 194, + 354, + 204 + ], + "type": "text", + "content": "Ours" + } + ] + } + ], + "index": 7, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 130, + 211, + 480, + 266 + ], + "lines": [ + { + "bbox": [ + 130, + 211, + 480, + 266 + ], + "spans": [ + { + "bbox": [ + 130, + 211, + 480, + 266 + ], + "type": "text", + "content": "Fig. 7: Ablation Study on garment physics estimation: LBS-based garment deformation (LBS) fails to generate realistic garment deformation. Physics-based simulation with random parameters (w/o physics) achieves realistic deformation but falls short of matching the ground-truth reference (GT). In contrast, PhysAvatar accurately captures garment dynamics, closely aligning with the reference." + } + ] + } + ], + "index": 10, + "angle": 0, + "type": "image_caption" + } + ], + "index": 6 + }, + { + "type": "image", + "bbox": [ + 395, + 114, + 473, + 192 + ], + "blocks": [ + { + "bbox": [ + 395, + 114, + 473, + 192 + ], + "lines": [ + { + "bbox": [ + 395, + 114, + 473, + 192 + ], + "spans": [ + { + "bbox": [ + 395, + 114, + 473, + 192 + ], + "type": "image", + "image_path": "d3d7c6faf0c6338aa8cf66be7feeb13a0d6a9bf620baf86af22f0e39802858b6.jpg" + } + ] + } + ], + "index": 8, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 425, + 194, + 439, + 204 + ], + "lines": [ + { + "bbox": [ + 425, + 194, + 439, + 204 + ], + "spans": [ + { + "bbox": [ + 425, + 194, + 439, + 204 + ], + "type": "text", + "content": "GT" + } + ] + } + ], + "index": 9, + "angle": 0, + "type": "image_caption" + } + ], + "index": 8 + }, + { + "type": "table", + "bbox": [ + 142, + 277, + 471, + 355 + ], + "blocks": [ + { + "bbox": [ + 142, + 277, + 471, + 355 + ], + "lines": [ + { + "bbox": [ + 142, + 277, + 471, + 355 + ], + "spans": [ + { + "bbox": [ + 142, + 277, + 471, + 355 + ], + "type": "table", + "html": "
MethodGeometryAppearance
CD (×10-3) (↓)F-Score (↑)LPIPS (↓)PSNR (↑)SSIM (↑)
LBS0.7191.60.035430.270.951
w/o physics0.6192.20.034730.260.951
w/o appearance——0.040228.260.947
PhysAvatar (ours)0.5693.00.034330.290.951
", + "image_path": "04dbd5f39a9da2d3462d2461500925ae2277ce02f975f5be7852ccff89cf2da4.jpg" + } + ] + } + ], + "index": 11, + "angle": 0, + "type": "table_body" + } + ], + "index": 11 + }, + { + "bbox": [ + 130, + 357, + 480, + 411 + ], + "lines": [ + { + "bbox": [ + 130, + 357, + 480, + 411 + ], + "spans": [ + { + "bbox": [ + 130, + 357, + 480, + 411 + ], + "type": "text", + "content": "Table 2: Ablation Study: PhysAvatar demonstrates improved geometric accuracy compared to both LBS-based animation (LBS) and simulations with random garment parameters (w/o physics), also shown in Fig. 7. PhysAvatar exhibits superior appearance metrics compared to the setting without physics-based appearance estimation (w/o appearance) (see Fig. 4)." + } + ] + } + ], + "index": 12, + "angle": 0, + "type": "text" + }, + { + "bbox": [ + 130, + 445, + 482, + 553 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 445, + 482, + 553 + ], + "spans": [ + { + "bbox": [ + 130, + 445, + 482, + 553 + ], + "type": "text", + "content": "Appearance estimation and modeling. We compare rendering outcomes with those obtained without appearance optimization through physics-based inverse rendering [78] (denoted as 'w/o appearance'), where the texture is directly extracted from the Gaussian framework, as detailed in section 3.3. Fig. 4 illustrates that, since the Gaussian framework does not model complex light-surface interactions, shadows are inherently baked into the texture. However, post-optimization, we note a significant removal of these baked shadows. Quantitatively, we see improvements across all appearance metrics post-optimization, shown in Table 2." + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 131, + 572, + 220, + 585 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 131, + 572, + 220, + 585 + ], + "spans": [ + { + "bbox": [ + 131, + 572, + 220, + 585 + ], + "type": "text", + "content": "4.4 Application" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 130, + 594, + 482, + 666 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 594, + 482, + 666 + ], + "spans": [ + { + "bbox": [ + 130, + 594, + 482, + 666 + ], + "type": "text", + "content": "PhysAvatar can facilitate a wide range of applications, such as animation, relighting, and redressing, as demonstrated in Fig. 1. Our method can also benefit from other motion databases in addition to motions defined under the regime of SMPL-X [85] or SMPL [64]. Fig. 1 shows our animation results on challenging motions from Mixamo [76], where we use its auto-rigging algorithm to drive the human body, boundary vertices of the garment, and SMPL-X mesh as the" + } + ] + } + ], + "index": 15 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 399, + 91, + 447, + 101 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 399, + 91, + 447, + 101 + ], + "spans": [ + { + "bbox": [ + 399, + 91, + 447, + 101 + ], + "type": "text", + "content": "PhysAvatar" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 470, + 91, + 481, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 470, + 91, + 481, + 100 + ], + "spans": [ + { + "bbox": [ + 470, + 91, + 481, + 100 + ], + "type": "text", + "content": "13" + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 12 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 134, + 115, + 219, + 201 + ], + "blocks": [ + { + "bbox": [ + 134, + 115, + 219, + 201 + ], + "lines": [ + { + "bbox": [ + 134, + 115, + 219, + 201 + ], + "spans": [ + { + "bbox": [ + 134, + 115, + 219, + 201 + ], + "type": "image", + "image_path": "7c0335449859589b52503fcf916029c5da5bf9fd2955c2c918c0adf628948c9d.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 158, + 205, + 280, + 215 + ], + "lines": [ + { + "bbox": [ + 158, + 205, + 280, + 215 + ], + "spans": [ + { + "bbox": [ + 158, + 205, + 280, + 215 + ], + "type": "text", + "content": "(a) Our results are readily exportable" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_caption" + } + ], + "index": 2 + }, + { + "type": "image", + "bbox": [ + 219, + 116, + 304, + 201 + ], + "blocks": [ + { + "bbox": [ + 219, + 116, + 304, + 201 + ], + "lines": [ + { + "bbox": [ + 219, + 116, + 304, + 201 + ], + "spans": [ + { + "bbox": [ + 219, + 116, + 304, + 201 + ], + "type": "image", + "image_path": "96c922d5af455dc27be21de45f9a56aa895487264f32c38d1ed806da8a3cb514.jpg" + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 131, + 218, + 480, + 252 + ], + "lines": [ + { + "bbox": [ + 131, + 218, + 480, + 252 + ], + "spans": [ + { + "bbox": [ + 131, + 218, + 480, + 252 + ], + "type": "text", + "content": "Fig. 8: Our results can be exported into a traditional computer graphics pipeline, e.g., Blender [13], which enables the avatars performing novel motions to be rendered in unseen environments (a) as well as editing lighting and texture of the garments (b)." + } + ] + } + ], + "index": 8, + "angle": 0, + "type": "image_caption" + } + ], + "index": 4 + }, + { + "type": "image", + "bbox": [ + 309, + 115, + 394, + 201 + ], + "blocks": [ + { + "bbox": [ + 309, + 115, + 394, + 201 + ], + "lines": [ + { + "bbox": [ + 309, + 115, + 394, + 201 + ], + "spans": [ + { + "bbox": [ + 309, + 115, + 394, + 201 + ], + "type": "image", + "image_path": "73ca939b9006627ea05c7464a574a7b69037edf79a3b46a9cbcb42c6465d2379.jpg" + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 332, + 205, + 456, + 215 + ], + "lines": [ + { + "bbox": [ + 332, + 205, + 456, + 215 + ], + "spans": [ + { + "bbox": [ + 332, + 205, + 456, + 215 + ], + "type": "text", + "content": "(b) Results with artistic modifications" + } + ] + } + ], + "index": 7, + "angle": 0, + "type": "image_caption" + } + ], + "index": 5 + }, + { + "type": "image", + "bbox": [ + 394, + 116, + 479, + 201 + ], + "blocks": [ + { + "bbox": [ + 394, + 116, + 479, + 201 + ], + "lines": [ + { + "bbox": [ + 394, + 116, + 479, + 201 + ], + "spans": [ + { + "bbox": [ + 394, + 116, + 479, + 201 + ], + "type": "image", + "image_path": "9d0c9034ce274eda34f5e032698a289f7c8d06e29141aa4506e4210d741a4df4.jpg" + } + ] + } + ], + "index": 6, + "angle": 0, + "type": "image_body" + } + ], + "index": 6 + }, + { + "bbox": [ + 131, + 274, + 482, + 346 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 131, + 274, + 482, + 346 + ], + "spans": [ + { + "bbox": [ + 131, + 274, + 482, + 346 + ], + "type": "text", + "content": "collider. Current state-of-the-art approaches cannot handle such settings since they are mainly based on LBS weights from SMPL or SMPL-X. Moreover, while current methods relying on NeRF [75] or 3D Gaussians [40] representation are usually incompatible with legacy graphics pipelines, our results are readily exportable and can be seamlessly integrated with computer graphics software tools (e.g., Blender [13]), enabling post-processing artistic modifications (Fig. 8)." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 132, + 362, + 329, + 376 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 362, + 329, + 376 + ], + "spans": [ + { + "bbox": [ + 132, + 362, + 329, + 376 + ], + "type": "text", + "content": "5 Limitations and Future Work" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 131, + 385, + 482, + 542 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 131, + 385, + 482, + 542 + ], + "spans": [ + { + "bbox": [ + 131, + 385, + 482, + 542 + ], + "type": "text", + "content": "Our method currently depends on manual garment segmentation and mesh UV unwrapping. We are posed to evolve our pipeline towards a more automated system by incorporating neural networks to streamline these data preprocessing tasks. Another limitation of our method is the absence of a perfect underlining human body for garment simulation. Our current use of the SMPL-X [85] model as the body collider introduces potential inaccuracies in collision detection and contact information due to mismatches between the SMPL-X mesh and the actual human body shape. Our method would benefit from using more advanced parametric models that promise a closer approximation to real human body shape. Furthermore, our pipeline leverages dense view captures to model avatars, an approach that, while effective, has its limitations. Adapting our method to a sparse-view setting would enhance the versatility of its applications in environments where capturing dense views is impractical." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 132, + 558, + 220, + 571 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 558, + 220, + 571 + ], + "spans": [ + { + "bbox": [ + 132, + 558, + 220, + 571 + ], + "type": "text", + "content": "6 Conclusion" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 131, + 581, + 482, + 666 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 131, + 581, + 482, + 666 + ], + "spans": [ + { + "bbox": [ + 131, + 581, + 482, + 666 + ], + "type": "text", + "content": "We introduce PhysAvatar, a comprehensive pipeline to model clothed 3D human avatars with state-of-the-art realism. By integrating a robust and efficient mesh tracking method, and a novel inverse rendering paradigm to capture the physics of loose garments and the appearance of the body, we believe our work represents a significant advancement in digital avatar modeling from visual observations, which also paves the way for innovative applications in entertainment, virtual reality, and digital fashion." + } + ] + } + ], + "index": 13 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "spans": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "type": "text", + "content": "14" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 166, + 91, + 239, + 102 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 166, + 91, + 239, + 102 + ], + "spans": [ + { + "bbox": [ + 166, + 91, + 239, + 102 + ], + "type": "text", + "content": "Yang Zheng et al." + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 13 + }, + { + "para_blocks": [ + { + "bbox": [ + 133, + 114, + 240, + 129 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 114, + 240, + 129 + ], + "spans": [ + { + "bbox": [ + 133, + 114, + 240, + 129 + ], + "type": "text", + "content": "Acknowledgement" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 130, + 139, + 480, + 183 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 139, + 480, + 183 + ], + "spans": [ + { + "bbox": [ + 130, + 139, + 480, + 183 + ], + "type": "text", + "content": "We would like to thank Jiayi Eris Zhang for the discussions. This material is based on work that is partially funded by an unrestricted gift from Google, Samsung, an SNF Postdoc Mobility fellowship, ARL grant W911NF-21-2-0104, and a Vannevar Bush Faculty Fellowship." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 133, + 201, + 197, + 214 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 201, + 197, + 214 + ], + "spans": [ + { + "bbox": [ + 133, + 201, + 197, + 214 + ], + "type": "text", + "content": "References" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 138, + 226, + 480, + 665 + ], + "type": "list", + "angle": 0, + "index": 19, + "blocks": [ + { + "bbox": [ + 142, + 226, + 480, + 259 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 142, + 226, + 480, + 259 + ], + "spans": [ + { + "bbox": [ + 142, + 226, + 480, + 259 + ], + "type": "text", + "content": "1. Abdal, R., Yifan, W., Shi, Z., Xu, Y., Po, R., Kuang, Z., Chen, Q., Yeung, D.Y., Wetzstein, G.: Gaussian shell maps for efficient 3d human generation. arXiv preprint arXiv:2311.17857 (2023)" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 142, + 261, + 480, + 293 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 142, + 261, + 480, + 293 + ], + "spans": [ + { + "bbox": [ + 142, + 261, + 480, + 293 + ], + "type": "text", + "content": "2. Abdrashitov, R., Raichstat, K., Monsen, J., Hill, D.: Robust skin weights transfer via weight inpainting. In: SIGGRAPH Asia 2023 Technical Communications, pp. 1-4 (2023)" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 142, + 293, + 480, + 325 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 142, + 293, + 480, + 325 + ], + "spans": [ + { + "bbox": [ + 142, + 293, + 480, + 325 + ], + "type": "text", + "content": "3. Anguelov, D., Srinivasan, P., Koller, D., Thrun, S., Rodgers, J., Davis, J.: Scape: shape completion and animation of people. In: ACM SIGGRAPH 2005 Papers, pp. 408-416 (2005)" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 142, + 326, + 480, + 357 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 142, + 326, + 480, + 357 + ], + "spans": [ + { + "bbox": [ + 142, + 326, + 480, + 357 + ], + "type": "text", + "content": "4. Bagautdinov, T., Wu, C., Simon, T., Prada, F., Shiratori, T., Wei, S.E., Xu, W., Sheikh, Y., Saragih, J.: Driving-signal aware full-body avatars. ACM Transactions on Graphics (TOG) 40(4), 1-17 (2021)" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 142, + 358, + 480, + 380 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 142, + 358, + 480, + 380 + ], + "spans": [ + { + "bbox": [ + 142, + 358, + 480, + 380 + ], + "type": "text", + "content": "5. Baraff, D., Witkin, A.: Large steps in cloth simulation. In: Seminal Graphics Papers: Pushing the Boundaries, Volume 2, pp. 767-778 (2023)" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 142, + 380, + 480, + 423 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 142, + 380, + 480, + 423 + ], + "spans": [ + { + "bbox": [ + 142, + 380, + 480, + 423 + ], + "type": "text", + "content": "6. Bashirov, R., Larionov, A., Ustinova, E., Sidorenko, M., Svitov, D., Zakharkin, I., Lempitsky, V.: Morf: Mobile realistic fullbody avatars from a monocular video. In: Proceedings of the IEEE/CVF Winter Conference on Applications of Computer Vision. pp. 3545-3555 (2024)" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 142, + 424, + 480, + 456 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 142, + 424, + 480, + 456 + ], + "spans": [ + { + "bbox": [ + 142, + 424, + 480, + 456 + ], + "type": "text", + "content": "7. Bergman, A., Kellnhofer, P., Yifan, W., Chan, E., Lindell, D., Wetzstein, G.: Generative neural articulated radiance fields. Advances in Neural Information Processing Systems 35, 1990-19916 (2022)" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 142, + 457, + 480, + 479 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 142, + 457, + 480, + 479 + ], + "spans": [ + { + "bbox": [ + 142, + 457, + 480, + 479 + ], + "type": "text", + "content": "8. Bertiche, H., Madadi, M., Escalera, S.: Cloth3d: clothed 3d humans. In: European Conference on Computer Vision. pp. 344-359. Springer (2020)" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 142, + 479, + 480, + 511 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 142, + 479, + 480, + 511 + ], + "spans": [ + { + "bbox": [ + 142, + 479, + 480, + 511 + ], + "type": "text", + "content": "9. Bertiche, H., Madadi, M., Escalera, S.: Pbns: physically based neural simulation for unsupervised garment pose space deformation. ACM Transactions on Graphics (TOG) 40(6), 1-14 (2021)" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 138, + 512, + 480, + 533 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 512, + 480, + 533 + ], + "spans": [ + { + "bbox": [ + 138, + 512, + 480, + 533 + ], + "type": "text", + "content": "10. Bertiche, H., Madadi, M., Escalera, S.: Neural cloth simulation. ACM Transactions on Graphics (TOG) 41(6), 1-14 (2022)" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 138, + 534, + 480, + 566 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 534, + 480, + 566 + ], + "spans": [ + { + "bbox": [ + 138, + 534, + 480, + 566 + ], + "type": "text", + "content": "11. Bertiche, H., Madadi, M., Tyson, E., Escalera, S.: Deepsd: Automatic deep skinning and pose space deformation for 3d garment animation. In: Proceedings of the IEEE/CVF International Conference on Computer Vision. pp. 5471-5480 (2021)" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 138, + 567, + 480, + 609 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 567, + 480, + 609 + ], + "spans": [ + { + "bbox": [ + 138, + 567, + 480, + 609 + ], + "type": "text", + "content": "12. Bhat, K.S., Twigg, C.D., Hodgins, J.K., Khosla, P.K., Popovic, Z., Seitz, S.M.: Estimating cloth simulation parameters from video. In: Proceedings of the 2003 ACM SIGGRAPH/Eurographics symposium on Computer animation. pp. 37-51 (2003)" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 138, + 610, + 480, + 632 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 610, + 480, + 632 + ], + "spans": [ + { + "bbox": [ + 138, + 610, + 480, + 632 + ], + "type": "text", + "content": "13. Blender Online Coummunity: Blender, blender Foundation, https://www.blender.org/" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 138, + 632, + 480, + 665 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 632, + 480, + 665 + ], + "spans": [ + { + "bbox": [ + 138, + 632, + 480, + 665 + ], + "type": "text", + "content": "14. Chen, B., Shen, Y., Shuai, Q., Zhou, X., Zhou, K., Zheng, Y.: Anidress: Animatable loose-dressed avatar from sparse views using garment rigging model. arXiv preprint arXiv:2401.15348 (2024)" + } + ] + } + ], + "index": 18 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 399, + 91, + 447, + 101 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 399, + 91, + 447, + 101 + ], + "spans": [ + { + "bbox": [ + 399, + 91, + 447, + 101 + ], + "type": "text", + "content": "PhysAvatar" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 470, + 92, + 480, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 470, + 92, + 480, + 100 + ], + "spans": [ + { + "bbox": [ + 470, + 92, + 480, + 100 + ], + "type": "text", + "content": "15" + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 14 + }, + { + "para_blocks": [ + { + "bbox": [ + 138, + 116, + 481, + 665 + ], + "type": "list", + "angle": 0, + "index": 18, + "blocks": [ + { + "bbox": [ + 138, + 116, + 481, + 149 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 116, + 481, + 149 + ], + "spans": [ + { + "bbox": [ + 138, + 116, + 481, + 149 + ], + "type": "text", + "content": "15. Chen, H.Y., Sastry, A., van Rees, W.M., Vouga, E.: Physical simulation of environmentally induced thin shell deformation. ACM Transactions on Graphics (TOG) 37(4), 1-13 (2018)" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 138, + 150, + 481, + 183 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 150, + 481, + 183 + ], + "spans": [ + { + "bbox": [ + 138, + 150, + 481, + 183 + ], + "type": "text", + "content": "16. Chen, X., Jiang, T., Song, J., Yang, J., Black, M.J., Geiger, A., Hilliges, O.: gdna: Towards generative detailed neural avatars. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 20427-20437 (2022)" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 138, + 183, + 481, + 226 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 183, + 481, + 226 + ], + "spans": [ + { + "bbox": [ + 138, + 183, + 481, + 226 + ], + "type": "text", + "content": "17. Chen, X., Zheng, Y., Black, M.J., Hilliges, O., Geiger, A.: Snarf: Differentiable forward skimming for animating non-rigid neural implicit shapes. In: Proceedings of the IEEE/CVF International Conference on Computer Vision. pp. 11594-11604 (2021)" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 138, + 227, + 481, + 270 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 227, + 481, + 270 + ], + "spans": [ + { + "bbox": [ + 138, + 227, + 481, + 270 + ], + "type": "text", + "content": "18. Chen, Y., Wang, X., Chen, X., Zhang, Q., Li, X., Guo, Y., Wang, J., Wang, F.: Uv volumes for real-time rendering of editable free-view human performance. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 16621-16631 (2023)" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 138, + 270, + 481, + 304 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 270, + 481, + 304 + ], + "spans": [ + { + "bbox": [ + 138, + 270, + 481, + 304 + ], + "type": "text", + "content": "19. Chentanez, N., Macklin, M., Müller, M., Jeschke, S., Kim, T.Y.: Cloth and skin deformation with a triangle mesh based convolutional neural network. In: Computer Graphics Forum. vol. 39, pp. 123-134. Wiley Online Library (2020)" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 138, + 304, + 481, + 336 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 304, + 481, + 336 + ], + "spans": [ + { + "bbox": [ + 138, + 304, + 481, + 336 + ], + "type": "text", + "content": "20. Clyde, D., Teran, J., Tamstorf, R.: Modeling and data-driven parameter estimation for woven fabrics. In: Proceedings of the ACM SIGGRAPH/Eurographics Symposium on Computer Animation. pp. 1-11 (2017)" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 138, + 336, + 481, + 358 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 336, + 481, + 358 + ], + "spans": [ + { + "bbox": [ + 138, + 336, + 481, + 358 + ], + "type": "text", + "content": "21. Contributors, M.: Openmmlab pose estimation toolbox and benchmark. https://github.com/open-mmlab/mmpose (2020)" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 138, + 358, + 481, + 392 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 358, + 481, + 392 + ], + "spans": [ + { + "bbox": [ + 138, + 358, + 481, + 392 + ], + "type": "text", + "content": "22. De Luigi, L., Li, R., Guillard, B., Salzmann, M., Fua, P.: Drapenet: Garment generation and self-supervised draping. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 1451-1460 (2023)" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 138, + 392, + 481, + 435 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 392, + 481, + 435 + ], + "spans": [ + { + "bbox": [ + 138, + 392, + 481, + 435 + ], + "type": "text", + "content": "23. Dong, Z., Chen, X., Yang, J., Black, M.J., Hilliges, O., Geiger, A.: Ag3d: Learning to generate 3d avatars from 2d image collections. In: Proceedings of the IEEE/CVF International Conference on Computer Vision (ICCV). pp. 14916-14927 (October 2023)" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 138, + 435, + 481, + 468 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 435, + 481, + 468 + ], + "spans": [ + { + "bbox": [ + 138, + 435, + 481, + 468 + ], + "type": "text", + "content": "24. Feng, Y., Yang, J., Pollefeys, M., Black, M.J., Bolkart, T.: Capturing and animation of body and clothing from monocular video. In: SIGGRAPH Asia 2022 Conference Papers. pp. 1-9 (2022)" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 138, + 468, + 481, + 512 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 468, + 481, + 512 + ], + "spans": [ + { + "bbox": [ + 138, + 468, + 481, + 512 + ], + "type": "text", + "content": "25. Fridovich-Keil, S., Meanti, G., Warburg, F.R., Recht, B., Kanazawa, A.: K-planes: Explicit radiance fields in space, time, and appearance. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 12479-12488 (2023)" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 138, + 512, + 481, + 544 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 512, + 481, + 544 + ], + "spans": [ + { + "bbox": [ + 138, + 512, + 481, + 544 + ], + "type": "text", + "content": "26. Genay, A., Lécuyer, A., Hachet, M.: Being an avatar \"for real\": a survey on virtual embodiment in augmented reality. IEEE Transactions on Visualization and Computer Graphics 28(12), 5071-5090 (2021)" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 138, + 545, + 481, + 578 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 545, + 481, + 578 + ], + "spans": [ + { + "bbox": [ + 138, + 545, + 481, + 578 + ], + "type": "text", + "content": "27. Geng, C., Peng, S., Xu, Z., Bao, H., Zhou, X.: Learning neural volumetric representations of dynamic humans in minutes. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 8759-8770 (2023)" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 138, + 578, + 481, + 611 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 578, + 481, + 611 + ], + "spans": [ + { + "bbox": [ + 138, + 578, + 481, + 611 + ], + "type": "text", + "content": "28. Grigorev, A., Black, M.J., Hilliges, O.: Hood: Hierarchical graphs for generalized modelling of clothing dynamics. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 16965-16974 (2023)" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 138, + 611, + 481, + 643 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 611, + 481, + 643 + ], + "spans": [ + { + "bbox": [ + 138, + 611, + 481, + 643 + ], + "type": "text", + "content": "29. Grinspun, E., Hirani, A.N., Desbrun, M., Schröder, P.: Discrete shells. In: Proceedings of the 2003 ACM SIGGRAPH/Eurographics symposium on Computer animation. pp. 62-67. CiteSeer (2003)" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 138, + 643, + 481, + 665 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 643, + 481, + 665 + ], + "spans": [ + { + "bbox": [ + 138, + 643, + 481, + 665 + ], + "type": "text", + "content": "30. Gundogdu, E., Constantin, V., Seifoddini, A., Dang, M., Salzmann, M., Fua, P.: Garnet: A two-stream network for fast and accurate 3d cloth draping. In:" + } + ] + } + ], + "index": 17 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "spans": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "type": "text", + "content": "16" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 166, + 91, + 239, + 102 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 166, + 91, + 239, + 102 + ], + "spans": [ + { + "bbox": [ + 166, + 91, + 239, + 102 + ], + "type": "text", + "content": "Yang Zheng et al." + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 15 + }, + { + "para_blocks": [ + { + "bbox": [ + 138, + 116, + 481, + 665 + ], + "type": "list", + "angle": 0, + "index": 19, + "blocks": [ + { + "bbox": [ + 154, + 116, + 480, + 138 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 154, + 116, + 480, + 138 + ], + "spans": [ + { + "bbox": [ + 154, + 116, + 480, + 138 + ], + "type": "text", + "content": "Proceedings of the IEEE/CVF International Conference on Computer Vision. pp. 8739-8748 (2019)" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 138, + 140, + 481, + 183 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 140, + 481, + 183 + ], + "spans": [ + { + "bbox": [ + 138, + 140, + 481, + 183 + ], + "type": "text", + "content": "31. Guo, J., Li, J., Narain, R., Park, H.S.: Inverse simulation: Reconstructing dynamic geometry of clothed humans via optimal control. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 14698-14707 (2021)" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 138, + 184, + 481, + 217 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 184, + 481, + 217 + ], + "spans": [ + { + "bbox": [ + 138, + 184, + 481, + 217 + ], + "type": "text", + "content": "32. Habermann, M., Liu, L., Xu, W., Pons-Moll, G., Zollhoefer, M., Theobalt, C.: Hdhumans: A hybrid approach for high-fidelity digital humans. Proceedings of the ACM on Computer Graphics and Interactive Techniques 6(3), 1-23 (2023)" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 138, + 218, + 481, + 251 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 218, + 481, + 251 + ], + "spans": [ + { + "bbox": [ + 138, + 218, + 481, + 251 + ], + "type": "text", + "content": "33. Habermann, M., Liu, L., Xu, W., Zollhoefer, M., Pons-Moll, G., Theobalt, C.: Real-time deep dynamic characters. ACM Transactions on Graphics (ToG) 40(4), 1-16 (2021)" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 138, + 251, + 481, + 283 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 251, + 481, + 283 + ], + "spans": [ + { + "bbox": [ + 138, + 251, + 481, + 283 + ], + "type": "text", + "content": "34. Hu, L., Zhang, H., Zhang, Y., Zhou, B., Liu, B., Zhang, S., Nie, L.: Gaussian-avatar: Towards realistic human avatar modeling from a single video via animatable 3d gaussians. arXiv preprint arXiv:2312.02134 (2023)" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 138, + 285, + 481, + 328 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 285, + 481, + 328 + ], + "spans": [ + { + "bbox": [ + 138, + 285, + 481, + 328 + ], + "type": "text", + "content": "35. Isik, M., Rünz, M., Georgopoulos, M., Khakhulin, T., Starck, J., Agapito, L., Nießner, M.: Humanrf: High-fidelity neural radiance fields for humans in motion. ACM Transactions on Graphics (TOG) 42(4), 1-12 (2023). https://doi.org/10.1145/3592415, https://doi.org/10.1145/3592415" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 138, + 330, + 481, + 361 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 330, + 481, + 361 + ], + "spans": [ + { + "bbox": [ + 138, + 330, + 481, + 361 + ], + "type": "text", + "content": "36. Jambon, C., Kerbl, B., Kopanas, G., Diolatzis, S., Leimkuhler, T., Drettakis, G.: Nerfshop: Interactive editing of neural radiance fields. Proc. ACM Comput. Graph. Interact. Tech. 6, 1:1-1:21 (2023). https://doi.org/10.1145/3585499" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 138, + 363, + 481, + 396 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 363, + 481, + 396 + ], + "spans": [ + { + "bbox": [ + 138, + 363, + 481, + 396 + ], + "type": "text", + "content": "37. Jiang, Y., Wang, R., Liu, Z.: A survey of cloth simulation and applications. In: 2008 9th International Conference on Computer-Aided Industrial Design and Conceptual Design. pp. 765-769. IEEE (2008)" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 138, + 397, + 481, + 429 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 397, + 481, + 429 + ], + "spans": [ + { + "bbox": [ + 138, + 397, + 481, + 429 + ], + "type": "text", + "content": "38. Jiang, Y., Shen, Z., Wang, P., Su, Z., Hong, Y., Zhang, Y., Yu, J., Xu, L.: Hifi4g: High-fidelity human performance rendering via compact gaussian splatting. arXiv preprint arXiv:2312.03461 (2023)" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 138, + 430, + 481, + 462 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 430, + 481, + 462 + ], + "spans": [ + { + "bbox": [ + 138, + 430, + 481, + 462 + ], + "type": "text", + "content": "39. Kato, H., Ushiku, Y., Harada, T.: Neural 3d mesh renderer. 2018 IEEE/CVF Conference on Computer Vision and Pattern Recognition pp. 3907-3916 (2017). https://doi.org/10.1109/CVPR.2018.00411" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 138, + 464, + 481, + 485 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 464, + 481, + 485 + ], + "spans": [ + { + "bbox": [ + 138, + 464, + 481, + 485 + ], + "type": "text", + "content": "40. Kerbl, B., Kopanas, G., Leimkuhler, T., Drettakis, G.: 3d gaussian splatting for real-time radiance field rendering. ACM Transactions on Graphics 42(4) (2023)" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 138, + 487, + 481, + 518 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 487, + 481, + 518 + ], + "spans": [ + { + "bbox": [ + 138, + 487, + 481, + 518 + ], + "type": "text", + "content": "41. Keselman, L., Hebert, M.: Approximate differentiable rendering with algebraic surfaces. In: European Conference on Computer Vision. pp. 596-614. Springer (2022)" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 138, + 520, + 481, + 541 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 520, + 481, + 541 + ], + "spans": [ + { + "bbox": [ + 138, + 520, + 481, + 541 + ], + "type": "text", + "content": "42. Komatsu, K.: Human skin model capable of natural shape variation. The visual computer 3, 265-271 (1988)" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 138, + 543, + 481, + 563 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 543, + 481, + 563 + ], + "spans": [ + { + "bbox": [ + 138, + 543, + 481, + 563 + ], + "type": "text", + "content": "43. Korban, M., Li, X.: A survey on applications of digital human avatars toward virtual co-presence. arXiv preprint arXiv:2201.04168 (2022)" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 138, + 565, + 481, + 597 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 565, + 481, + 597 + ], + "spans": [ + { + "bbox": [ + 138, + 565, + 481, + 597 + ], + "type": "text", + "content": "44. Kwon, Y., Kim, D., Ceylan, D., Fuchs, H.: Neural human performer: Learning generalizable radiance fields for human performance rendering. Advances in Neural Information Processing Systems 34, 24741-24752 (2021)" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 138, + 599, + 481, + 631 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 599, + 481, + 631 + ], + "spans": [ + { + "bbox": [ + 138, + 599, + 481, + 631 + ], + "type": "text", + "content": "45. Kwon, Y., Kim, D., Ceylan, D., Fuchs, H.: Neural image-based avatars: Generalizable radiance fields for human avatar modeling. In: The Eleventh International Conference on Learning Representations (2023)" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 138, + 632, + 481, + 665 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 632, + 481, + 665 + ], + "spans": [ + { + "bbox": [ + 138, + 632, + 481, + 665 + ], + "type": "text", + "content": "46. Kwon, Y., Liu, L., Fuchs, H., Habermann, M., Theobalt, C.: Deliffas: Deformable light fields for fast avatar synthesis. Advances in Neural Information Processing Systems 36 (2023)" + } + ] + } + ], + "index": 18 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 399, + 91, + 447, + 101 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 399, + 91, + 447, + 101 + ], + "spans": [ + { + "bbox": [ + 399, + 91, + 447, + 101 + ], + "type": "text", + "content": "PhysAvatar" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 470, + 91, + 481, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 470, + 91, + 481, + 100 + ], + "spans": [ + { + "bbox": [ + 470, + 91, + 481, + 100 + ], + "type": "text", + "content": "17" + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 16 + }, + { + "para_blocks": [ + { + "bbox": [ + 138, + 116, + 480, + 665 + ], + "type": "list", + "angle": 0, + "index": 21, + "blocks": [ + { + "bbox": [ + 138, + 116, + 480, + 149 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 116, + 480, + 149 + ], + "spans": [ + { + "bbox": [ + 138, + 116, + 480, + 149 + ], + "type": "text", + "content": "47. Larionov, E., Eckert, M.L., Wolff, K., Stuyck, T.: Estimating cloth elasticity parameters using position-based simulation of compliant constrained dynamics. arXiv preprint arXiv:2212.08790 (2022)" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 138, + 150, + 480, + 183 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 150, + 480, + 183 + ], + "spans": [ + { + "bbox": [ + 138, + 150, + 480, + 183 + ], + "type": "text", + "content": "48. Lassner, C., Zollhofer, M.: Pulsar: Efficient sphere-based neural rendering. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 1440-1449 (2021)" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 138, + 183, + 480, + 215 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 183, + 480, + 215 + ], + "spans": [ + { + "bbox": [ + 138, + 183, + 480, + 215 + ], + "type": "text", + "content": "49. Lee, D., Kang, H., Lee, I.K.: Cloth combo: Modeling inter-cloth interaction for draping multi-layered clothes. ACM Transactions on Graphics (TOG) 42(6), 1-13 (2023)" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 138, + 216, + 480, + 237 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 216, + 480, + 237 + ], + "spans": [ + { + "bbox": [ + 138, + 216, + 480, + 237 + ], + "type": "text", + "content": "50. LeVeque, R.J.: Finite difference methods for ordinary and partial differential equations: steady-state and time-dependent problems. SIAM (2007)" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 138, + 237, + 480, + 270 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 237, + 480, + 270 + ], + "spans": [ + { + "bbox": [ + 138, + 237, + 480, + 270 + ], + "type": "text", + "content": "51. Li, M., Ferguson, Z., Schneider, T., Langlois, T.R., Zorin, D., Panozzo, D., Jiang, C., Kaufman, D.M.: Incremental potential contact: intersection-and inversion-free, large-deformation dynamics. ACM Trans. Graph. 39(4), 49 (2020)" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 138, + 270, + 480, + 293 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 270, + 480, + 293 + ], + "spans": [ + { + "bbox": [ + 138, + 270, + 480, + 293 + ], + "type": "text", + "content": "52. Li, M., Kaufman, D.M., Jiang, C.: Codimensional incremental potential contact. ACM Transactions on Graphics (TOG) 40(4), 1-24 (2021)" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 138, + 293, + 480, + 314 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 293, + 480, + 314 + ], + "spans": [ + { + "bbox": [ + 138, + 293, + 480, + 314 + ], + "type": "text", + "content": "53. Li, R., Tanke, J., Vo, M., Zollhofer, M., Gall, J., Kanazawa, A., Lassner, C.: Tava: Template-free animatable volumetric actors (2022)" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 138, + 315, + 480, + 335 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 315, + 480, + 335 + ], + "spans": [ + { + "bbox": [ + 138, + 315, + 480, + 335 + ], + "type": "text", + "content": "54. Li, X., Li, X.R., Li, Y., Feng, W.: Review of cloth modeling and simulation for virtual fitting. Textile Research Journal 93(7-8), 1699-1711 (2023)" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 138, + 336, + 480, + 369 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 336, + 480, + 369 + ], + "spans": [ + { + "bbox": [ + 138, + 336, + 480, + 369 + ], + "type": "text", + "content": "55. Li, Y., Chen, H.y., Larionov, E., Sarafianos, N., Matusik, W., Stuyck, T.: Difavatar: Simulation-ready garment optimization with differentiable simulation. arXiv preprint arXiv:2311.12194 (2023)" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 138, + 369, + 480, + 402 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 369, + 480, + 402 + ], + "spans": [ + { + "bbox": [ + 138, + 369, + 480, + 402 + ], + "type": "text", + "content": "56. Li, Y., Du, T., Wu, K., Xu, J., Matusik, W.: Diffcloth: Differentiable cloth simulation with dry frictional contact. ACM Transactions on Graphics (TOG) 42(1), 1-20 (2022)" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 138, + 402, + 480, + 435 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 402, + 480, + 435 + ], + "spans": [ + { + "bbox": [ + 138, + 402, + 480, + 435 + ], + "type": "text", + "content": "57. Li, Z., Zheng, Z., Liu, Y., Zhou, B., Liu, Y.: Posevocab: Learning joint-structured pose embeddings for human avatar modeling. arXiv preprint arXiv:2304.13006 (2023)" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 138, + 435, + 480, + 468 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 435, + 480, + 468 + ], + "spans": [ + { + "bbox": [ + 138, + 435, + 480, + 468 + ], + "type": "text", + "content": "58. Li, Z., Zheng, Z., Wang, L., Liu, Y.: Animatable gaussians: Learning pose-dependent gaussian maps for high-fidelity human avatar modeling. arXiv preprint arXiv:2311.16096 (2023)" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 138, + 468, + 480, + 501 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 468, + 480, + 501 + ], + "spans": [ + { + "bbox": [ + 138, + 468, + 480, + 501 + ], + "type": "text", + "content": "59. Liu, L., Habermann, M., Rudnev, V., Sarkar, K., Gu, J., Theobalt, C.: Neural actor: Neural free-view synthesis of human actors with pose control. ACM transactions on graphics (TOG) 40(6), 1-16 (2021)" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 138, + 501, + 480, + 534 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 501, + 480, + 534 + ], + "spans": [ + { + "bbox": [ + 138, + 501, + 480, + 534 + ], + "type": "text", + "content": "60. Liu, L., Xu, W., Zollhoefer, M., Kim, H., Bernard, F., Habermann, M., Wang, W., Theobalt, C.: Neural rendering and reenactment of human actor videos. ACM Transactions on Graphics (TOG) 38(5), 1-14 (2019)" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 138, + 534, + 480, + 567 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 534, + 480, + 567 + ], + "spans": [ + { + "bbox": [ + 138, + 534, + 480, + 567 + ], + "type": "text", + "content": "61. Liu, S., Li, T., Chen, W., Li, H.: Soft rasterizer: A differentiable renderer for image-based 3d reasoning. In: Proceedings of the IEEE/CVF International Conference on Computer Vision. pp. 7708-7717 (2019)" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 138, + 567, + 480, + 589 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 567, + 480, + 589 + ], + "spans": [ + { + "bbox": [ + 138, + 567, + 480, + 589 + ], + "type": "text", + "content": "62. Liu, T., Bouaziz, S., Kavan, L.: Quasi-newton methods for real-time simulation of hyperelastic materials. Acm Transactions on Graphics (TOG) 36(3), 1-16 (2017)" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 138, + 589, + 480, + 621 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 589, + 480, + 621 + ], + "spans": [ + { + "bbox": [ + 138, + 589, + 480, + 621 + ], + "type": "text", + "content": "63. Lombardi, S., Simon, T., Saragih, J., Schwartz, G., Lehrmann, A., Sheikh, Y.: Neural volumes: Learning dynamic renderable volumes from images. arXiv preprint arXiv:1906.07751 (2019)" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 138, + 621, + 480, + 643 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 621, + 480, + 643 + ], + "spans": [ + { + "bbox": [ + 138, + 621, + 480, + 643 + ], + "type": "text", + "content": "64. Loper, M., Mahmood, N., Romero, J., Pons-Moll, G., Black, M.J.: Smpl: A skinned multi-person linear model. ACM Transactions on Graphics 34(6) (2015)" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 138, + 643, + 480, + 665 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 643, + 480, + 665 + ], + "spans": [ + { + "bbox": [ + 138, + 643, + 480, + 665 + ], + "type": "text", + "content": "65. Luiten, J., Kopanas, G., Leibe, B., Ramanan, D.: Dynamic 3d gaussians: Tracking by persistent dynamic view synthesis. arXiv preprint arXiv:2308.09713 (2023)" + } + ] + } + ], + "index": 20 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "spans": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "type": "text", + "content": "18" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 166, + 91, + 239, + 102 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 166, + 91, + 239, + 102 + ], + "spans": [ + { + "bbox": [ + 166, + 91, + 239, + 102 + ], + "type": "text", + "content": "Yang Zheng et al." + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 17 + }, + { + "para_blocks": [ + { + "bbox": [ + 138, + 116, + 481, + 665 + ], + "type": "list", + "angle": 0, + "index": 18, + "blocks": [ + { + "bbox": [ + 138, + 116, + 481, + 159 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 116, + 481, + 159 + ], + "spans": [ + { + "bbox": [ + 138, + 116, + 481, + 159 + ], + "type": "text", + "content": "66. Ma, Q., Saito, S., Yang, J., Tang, S., Black, M.J.: Scale: Modeling clothed humans with a surface codec of articulated local elements. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 16082-16093 (2021)" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 138, + 160, + 481, + 192 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 160, + 481, + 192 + ], + "spans": [ + { + "bbox": [ + 138, + 160, + 481, + 192 + ], + "type": "text", + "content": "67. Ma, Q., Yang, J., Black, M.J., Tang, S.: Neural point-based shape modeling of humans in challenging clothing. In: 2022 International Conference on 3D Vision (3DV). pp. 679-689. IEEE (2022)" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 138, + 192, + 481, + 236 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 192, + 481, + 236 + ], + "spans": [ + { + "bbox": [ + 138, + 192, + 481, + 236 + ], + "type": "text", + "content": "68. Ma, Q., Yang, J., Ranjan, A., Pujades, S., Pons-Moll, G., Tang, S., Black, M.J.: Learning to dress 3d people in generative clothing. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 6469-6478 (2020)" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 138, + 236, + 481, + 267 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 236, + 481, + 267 + ], + "spans": [ + { + "bbox": [ + 138, + 236, + 481, + 267 + ], + "type": "text", + "content": "69. Ma, Q., Yang, J., Tang, S., Black, M.J.: The power of points for modeling humans in clothing. In: Proceedings of the IEEE/CVF International Conference on Computer Vision. pp. 10974-10984 (2021)" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 138, + 268, + 481, + 321 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 268, + 481, + 321 + ], + "spans": [ + { + "bbox": [ + 138, + 268, + 481, + 321 + ], + "type": "text", + "content": "70. Macklin, M., Müller, M., Chentanez, N.: Xpbd: position-based simulation of compliant constrained dynamics. In: Proceedings of the 9th International Conference on Motion in Games. p. 49-54. MIG '16, Association for Computing Machinery, New York, NY, USA (2016). https://doi.org/10.1145/2994258.2994272, https://doi.org/10.1145/2994258.2994272" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 138, + 322, + 481, + 354 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 322, + 481, + 354 + ], + "spans": [ + { + "bbox": [ + 138, + 322, + 481, + 354 + ], + "type": "text", + "content": "71. Magnenat, T., Laperrière, R., Thalmann, D.: Joint-dependent local deformations for hand animation and object grasping. Tech. rep., Canadian Inf. Process. Soc (1988)" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 138, + 354, + 481, + 374 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 354, + 481, + 374 + ], + "spans": [ + { + "bbox": [ + 138, + 354, + 481, + 374 + ], + "type": "text", + "content": "72. Mahmood, N., Ghorbani, N., Troje, N.F., Pons-Moll, G., Black, M.J.: AMASS: Archive of motion capture as surface shapes. In: ICCV (2019)" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 138, + 374, + 481, + 418 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 374, + 481, + 418 + ], + "spans": [ + { + "bbox": [ + 138, + 374, + 481, + 418 + ], + "type": "text", + "content": "73. Mescheder, L., Oechsle, M., Niemeyer, M., Nowozin, S., Geiger, A.: Occupancy networks: Learning 3d reconstruction in function space. In: Proceedings of the IEEE/CVF conference on computer vision and pattern recognition. pp. 4460-4470 (2019)" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 138, + 418, + 481, + 451 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 418, + 481, + 451 + ], + "spans": [ + { + "bbox": [ + 138, + 418, + 481, + 451 + ], + "type": "text", + "content": "74. Miguel, E., Bradley, D., Thomaszewski, B., Bickel, B., Matusik, W., Otaduy, M.A., Marschner, S.: Data-driven estimation of cloth simulation models. In: Computer Graphics Forum. vol. 31, pp. 519-528. Wiley Online Library (2012)" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 138, + 450, + 481, + 483 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 450, + 481, + 483 + ], + "spans": [ + { + "bbox": [ + 138, + 450, + 481, + 483 + ], + "type": "text", + "content": "75. Mildenhall, B., Srinivasan, P.P., Tancik, M., Barron, J.T., Ramamoorthi, R., Ng, R.: Nerf: Representing scenes as neural radiance fields for view synthesis. Communications of the ACM 65(1), 99-106 (2021)" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 138, + 483, + 268, + 492 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 483, + 268, + 492 + ], + "spans": [ + { + "bbox": [ + 138, + 483, + 268, + 492 + ], + "type": "text", + "content": "76. Mixamo, www MIXamo.com" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 138, + 492, + 481, + 536 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 492, + 481, + 536 + ], + "spans": [ + { + "bbox": [ + 138, + 492, + 481, + 536 + ], + "type": "text", + "content": "77. Newcombe, R.A., Izadi, S., Hilliges, O., Molyneaux, D., Kim, D., Davison, A.J., Kohi, P., Shotton, J., Hodges, S., Fitzgibbon, A.: Kinectfusion: Real-time dense surface mapping and tracking. In: 2011 10th IEEE international symposium on mixed and augmented reality. pp. 127-136. IEEE (2011)" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 138, + 536, + 481, + 567 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 536, + 481, + 567 + ], + "spans": [ + { + "bbox": [ + 138, + 536, + 481, + 567 + ], + "type": "text", + "content": "78. Nimier-David, M., Vicini, D., Zeltner, T., Jakob, W.: Mitsuba 2: A retargetable forward and inverse renderer. ACM Transactions on Graphics (TOG) 38(6), 1-17 (2019)" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 138, + 568, + 481, + 601 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 568, + 481, + 601 + ], + "spans": [ + { + "bbox": [ + 138, + 568, + 481, + 601 + ], + "type": "text", + "content": "79. Noguchi, A., Sun, X., Lin, S., Harada, T.: Neural articulated radiance field. In: Proceedings of the IEEE/CVF International Conference on Computer Vision (ICCV). pp. 5762-5772 (October 2021)" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 138, + 601, + 481, + 632 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 601, + 481, + 632 + ], + "spans": [ + { + "bbox": [ + 138, + 601, + 481, + 632 + ], + "type": "text", + "content": "80. Noguchi, A., Sun, X., Lin, S., Harada, T.: Unsupervised learning of efficient geometry-aware neural articulated representations. In: European Conference on Computer Vision. pp. 597-614. Springer (2022)" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 138, + 632, + 481, + 665 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 632, + 481, + 665 + ], + "spans": [ + { + "bbox": [ + 138, + 632, + 481, + 665 + ], + "type": "text", + "content": "81. Pan, X., Mai, J., Jiang, X., Tang, D., Li, J., Shao, T., Zhou, K., Jin, X., Manocha, D.: Predicting loose-fitting garment deformations using bone-driven motion networks. In: ACM SIGGRAPH 2022 Conference Proceedings. pp. 1-10 (2022)" + } + ] + } + ], + "index": 17 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 399, + 91, + 447, + 101 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 399, + 91, + 447, + 101 + ], + "spans": [ + { + "bbox": [ + 399, + 91, + 447, + 101 + ], + "type": "text", + "content": "PhysAvatar" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 470, + 91, + 481, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 470, + 91, + 481, + 100 + ], + "spans": [ + { + "bbox": [ + 470, + 91, + 481, + 100 + ], + "type": "text", + "content": "19" + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 18 + }, + { + "para_blocks": [ + { + "bbox": [ + 137, + 116, + 480, + 665 + ], + "type": "list", + "angle": 0, + "index": 17, + "blocks": [ + { + "bbox": [ + 137, + 116, + 479, + 149 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 137, + 116, + 479, + 149 + ], + "spans": [ + { + "bbox": [ + 137, + 116, + 479, + 149 + ], + "type": "text", + "content": "82. Pang, H., Zhu, H., Kortylewski, A., Theobalt, C., Habermann, M.: Ash: Animatable gaussian splats for efficient and photoreal human rendering. arXiv preprint arXiv:2312.05941 (2023)" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 137, + 150, + 480, + 194 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 137, + 150, + 480, + 194 + ], + "spans": [ + { + "bbox": [ + 137, + 150, + 480, + 194 + ], + "type": "text", + "content": "83. Park, J.J., Florence, P., Straub, J., Newcombe, R., Lovegrove, S.: Deepsdf: Learning continuous signed distance functions for shape representation. In: Proceedings of the IEEE/CVF conference on computer vision and pattern recognition. pp. 165-174 (2019)" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 138, + 194, + 480, + 237 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 194, + 480, + 237 + ], + "spans": [ + { + "bbox": [ + 138, + 194, + 480, + 237 + ], + "type": "text", + "content": "84. Park, K., Sinha, U., Hedman, P., Barron, J.T., Bouaziz, S., Goldman, D.B., Martin-Brualla, R., Seitz, S.M.: Hypernerf: a higher-dimensional representation for topologically varying neural radiance fields. ACM Transactions on Graphics (TOG) 40(6), 1-12 (2021)" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 138, + 238, + 480, + 282 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 238, + 480, + 282 + ], + "spans": [ + { + "bbox": [ + 138, + 238, + 480, + 282 + ], + "type": "text", + "content": "85. Pavlakos, G., Choutas, V., Ghorbani, N., Bolkart, T., Osman, A.A.A., Tzionas, D., Black, M.J.: Expressive body capture: 3D hands, face, and body from a single image. In: Proceedings IEEE Conf. on Computer Vision and Pattern Recognition (CVPR). pp. 10975-10985 (2019)" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 138, + 282, + 480, + 325 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 282, + 480, + 325 + ], + "spans": [ + { + "bbox": [ + 138, + 282, + 480, + 325 + ], + "type": "text", + "content": "86. Peng, S., Dong, J., Wang, Q., Zhang, S., Shuai, Q., Zhou, X., Bao, H.: Animatable neural radiance fields for modeling dynamic human bodies. In: Proceedings of the IEEE/CVF International Conference on Computer Vision. pp. 14314-14323 (2021)" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 138, + 325, + 480, + 369 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 325, + 480, + 369 + ], + "spans": [ + { + "bbox": [ + 138, + 325, + 480, + 369 + ], + "type": "text", + "content": "87. Peng, S., Zhang, Y., Xu, Y., Wang, Q., Shuai, Q., Bao, H., Zhou, X.: Neural body: Implicit neural representations with structured latent codes for novel view synthesis of dynamic humans. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 9054-9063 (2021)" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 138, + 369, + 480, + 402 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 369, + 480, + 402 + ], + "spans": [ + { + "bbox": [ + 138, + 369, + 480, + 402 + ], + "type": "text", + "content": "88. Peng, T., Wu, W., Liu, J., Li, L., Miao, J., Hu, X., He, R., Li, L.: Pgn-cloth: Physics-based graph network model for 3d cloth animation. Displays 80, 102534 (2023)" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 138, + 402, + 480, + 445 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 402, + 480, + 445 + ], + "spans": [ + { + "bbox": [ + 138, + 402, + 480, + 445 + ], + "type": "text", + "content": "89. Pumarola, A., Corona, E., Pons-Moll, G., Moreno-Noguer, F.: D-nerf: Neural radiance fields for dynamic scenes. 2021 IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR) pp. 10313-10322 (2020). https://doi.org/10.1109/CVPR46437.2021.01018" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 138, + 446, + 480, + 479 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 446, + 480, + 479 + ], + "spans": [ + { + "bbox": [ + 138, + 446, + 480, + 479 + ], + "type": "text", + "content": "90. Saito, S., Yang, J., Ma, Q., Black, M.J.: Scanimate: Weakly supervised learning of skinned clothed avatar networks. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 2886-2897 (2021)" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 138, + 479, + 480, + 512 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 479, + 480, + 512 + ], + "spans": [ + { + "bbox": [ + 138, + 479, + 480, + 512 + ], + "type": "text", + "content": "91. Santesteban, I., Otaduy, M.A., Casas, D.: Learning-based animation of clothing for virtual try-on. In: Computer Graphics Forum. vol. 38, pp. 355-366. Wiley Online Library (2019)" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 138, + 512, + 480, + 544 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 512, + 480, + 544 + ], + "spans": [ + { + "bbox": [ + 138, + 512, + 480, + 544 + ], + "type": "text", + "content": "92. Santesteban, I., Otaduy, M.A., Casas, D.: Snug: Self-supervised neural dynamic garments. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 8140-8150 (2022)" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 138, + 545, + 480, + 578 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 545, + 480, + 578 + ], + "spans": [ + { + "bbox": [ + 138, + 545, + 480, + 578 + ], + "type": "text", + "content": "93. Schonberger, J.L., Frahm, J.M.: Structure-from-motion revisited. In: Proceedings of the IEEE conference on computer vision and pattern recognition. pp. 4104-4113 (2016)" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 138, + 578, + 480, + 611 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 578, + 480, + 611 + ], + "spans": [ + { + "bbox": [ + 138, + 578, + 480, + 611 + ], + "type": "text", + "content": "94. Sitzmann, V., Thies, J., Heide, F., Nießner, M., Wetzstein, G., Zollhöfer, M.: Deepvoxels: Learning persistent 3d feature embeddings. In: Proc. Computer Vision and Pattern Recognition (CVPR), IEEE (2019)" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 138, + 611, + 480, + 643 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 611, + 480, + 643 + ], + "spans": [ + { + "bbox": [ + 138, + 611, + 480, + 643 + ], + "type": "text", + "content": "95. Sitzmann, V., Zollhöfer, M., Wetzstein, G.: Scene representation networks: Continuous 3d-structure-aware neural scene representations. In: Advances in Neural Information Processing Systems (2019)" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 138, + 643, + 480, + 665 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 643, + 480, + 665 + ], + "spans": [ + { + "bbox": [ + 138, + 643, + 480, + 665 + ], + "type": "text", + "content": "96. Snavely, N., Seitz, S.M., Szeliski, R.: Photo tourism: exploring photo collections in 3d. In: ACM siggraph 2006 papers, pp. 835-846 (2006)" + } + ] + } + ], + "index": 16 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "spans": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "type": "text", + "content": "20" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 166, + 91, + 239, + 102 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 166, + 91, + 239, + 102 + ], + "spans": [ + { + "bbox": [ + 166, + 91, + 239, + 102 + ], + "type": "text", + "content": "Yang Zheng et al." + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 19 + }, + { + "para_blocks": [ + { + "bbox": [ + 134, + 116, + 480, + 665 + ], + "type": "list", + "angle": 0, + "index": 17, + "blocks": [ + { + "bbox": [ + 136, + 116, + 480, + 148 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 136, + 116, + 480, + 148 + ], + "spans": [ + { + "bbox": [ + 136, + 116, + 480, + 148 + ], + "type": "text", + "content": "97. Stoll, C., Gall, J., De Aguiar, E., Thrun, S., Theobalt, C.: Video-based reconstruction of animatable human characters. ACM Transactions on Graphics (TOG) 29(6), 1-10 (2010)" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 137, + 149, + 480, + 181 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 137, + 149, + 480, + 181 + ], + "spans": [ + { + "bbox": [ + 137, + 149, + 480, + 181 + ], + "type": "text", + "content": "98. Stuyck, T., Chen, H.y.: Diffxpbd: Differentiable position-based simulation of compliant constraint dynamics. Proceedings of the ACM on Computer Graphics and Interactive Techniques 6(3), 1-14 (2023)" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 138, + 182, + 480, + 213 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 182, + 480, + 213 + ], + "spans": [ + { + "bbox": [ + 138, + 182, + 480, + 213 + ], + "type": "text", + "content": "99. Su, S.Y., Yu, F., Zollhöfer, M., Rhodin, H.: A-nerf: Articulated neural radiance fields for learning human shape, appearance, and pose. Advances in Neural Information Processing Systems 34, 12278–12291 (2021)" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 134, + 214, + 480, + 255 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 134, + 214, + 480, + 255 + ], + "spans": [ + { + "bbox": [ + 134, + 214, + 480, + 255 + ], + "type": "text", + "content": "100. Su, Z., Hu, L., Lin, S., Zhang, H., Zhang, S., Thies, J., Liu, Y.: Caphy: Capturing physical properties for animatable human avatars. In: Proceedings of the IEEE/CVF International Conference on Computer Vision. pp. 14150-14160 (2023)" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 134, + 257, + 480, + 298 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 134, + 257, + 480, + 298 + ], + "spans": [ + { + "bbox": [ + 134, + 257, + 480, + 298 + ], + "type": "text", + "content": "101. Sun, M., Yang, D., Kou, D., Jiang, Y., Shan, W., Yan, Z., Zhang, L.: Human 3d avatar modeling with implicit neural representation: A brief survey. In: 2022 14th International Conference on Signal Processing Systems (ICSPS). pp. 818-827. IEEE (2022)" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 134, + 300, + 480, + 320 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 134, + 300, + 480, + 320 + ], + "spans": [ + { + "bbox": [ + 134, + 300, + 480, + 320 + ], + "type": "text", + "content": "102. Tamstorf, R., Grinspun, E.: Discrete bending forces and their jacobians. Graphical models 75(6), 362-370 (2013)" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 134, + 322, + 480, + 352 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 134, + 322, + 480, + 352 + ], + "spans": [ + { + "bbox": [ + 134, + 322, + 480, + 352 + ], + "type": "text", + "content": "103. Tang, M., Tong, R., Narain, R., Meng, C., Manocha, D.: A gpu-based streaming algorithm for high-resolution cloth simulation. In: Computer Graphics Forum. vol. 32, pp. 21-30. Wiley Online Library (2013)" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 134, + 354, + 480, + 384 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 134, + 354, + 480, + 384 + ], + "spans": [ + { + "bbox": [ + 134, + 354, + 480, + 384 + ], + "type": "text", + "content": "104. Tatarchenko, M., Richter, S.R., Ranftl, R., Li, Z., Koltun, V., Brox, T.: What do single-view 3d reconstruction networks learn? In: Proceedings of the IEEE/CVF conference on computer vision and pattern recognition. pp. 3405-3414 (2019)" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 134, + 385, + 480, + 438 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 134, + 385, + 480, + 438 + ], + "spans": [ + { + "bbox": [ + 134, + 385, + 480, + 438 + ], + "type": "text", + "content": "105. Tewari, A., Fried, O., Thies, J., Sitzmann, V., Lombardi, S., Sunkavalli, K., Martin-Brualla, R., Simon, T., Saragih, J.M., Nießner, M., Pandey, R., Fanello, S., Wetzstein, G., Zhu, J.Y., Theobalt, C., Agrawala, M., Shechtman, E., Goldman, D.B., Zollhofer, M.: State of the art on neural rendering. Computer Graphics Forum 39 (2020). https://doi.org/10.1111/cgf.14022" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 134, + 439, + 480, + 482 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 134, + 439, + 480, + 482 + ], + "spans": [ + { + "bbox": [ + 134, + 439, + 480, + 482 + ], + "type": "text", + "content": "106. Tewari, A., Thies, J., Mildenhall, B., Srinivasan, P., Tretschk, E., Yifan, W., Lassner, C., Sitzmann, V., Martin-Brualla, R., Lombardi, S., et al.: Advances in neural rendering. In: Computer Graphics Forum. vol. 41, pp. 703-735. Wiley Online Library (2022)" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 134, + 483, + 480, + 525 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 134, + 483, + 480, + 525 + ], + "spans": [ + { + "bbox": [ + 134, + 483, + 480, + 525 + ], + "type": "text", + "content": "107. Tiwari, G., Sarafianos, N., Tung, T., Pons-Moll, G.: Neural-gif: Neural generalized implicit functions for animating people in clothing. In: Proceedings of the IEEE/CVF International Conference on Computer Vision. pp. 11708-11718 (2021)" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 134, + 525, + 480, + 578 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 134, + 525, + 480, + 578 + ], + "spans": [ + { + "bbox": [ + 134, + 525, + 480, + 578 + ], + "type": "text", + "content": "108. Tretschk, E., Tewari, A., Golyanik, V., Zollhöfer, M., Lassner, C., Theobalt, C.: Non-rigid neural radiance fields: Reconstruction and novel view synthesis of a dynamic scene from monocular video. 2021 IEEE/CVF International Conference on Computer Vision (ICCV) pp. 12939-12950 (2020). https://doi.org/10.1109/ICCV48922.2021.01272" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 134, + 579, + 480, + 611 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 134, + 579, + 480, + 611 + ], + "spans": [ + { + "bbox": [ + 134, + 579, + 480, + 611 + ], + "type": "text", + "content": "109. Vassilev, T., Spanlang, B., Chrysanthou, Y.: Fast cloth animation on walking avatars. In: Computer Graphics Forum. vol. 20, pp. 260-267. Wiley Online Library (2001)" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 134, + 612, + 480, + 643 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 134, + 612, + 480, + 643 + ], + "spans": [ + { + "bbox": [ + 134, + 612, + 480, + 643 + ], + "type": "text", + "content": "110. Vidaurre, R., Santesteban, I., Garces, E., Casas, D.: Fully convolutional graph neural networks for parametric virtual try-on. In: Computer Graphics Forum. vol. 39, pp. 145-156. Wiley Online Library (2020)" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 134, + 644, + 480, + 665 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 134, + 644, + 480, + 665 + ], + "spans": [ + { + "bbox": [ + 134, + 644, + 480, + 665 + ], + "type": "text", + "content": "111. Wang, H.: Gpu-based simulation of cloth wrinkles at submillimeter levels. ACM Transactions on Graphics (TOG) 40(4), 1-14 (2021)" + } + ] + } + ], + "index": 16 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 399, + 91, + 447, + 101 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 399, + 91, + 447, + 101 + ], + "spans": [ + { + "bbox": [ + 399, + 91, + 447, + 101 + ], + "type": "text", + "content": "PhysAvatar" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 470, + 91, + 479, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 470, + 91, + 479, + 100 + ], + "spans": [ + { + "bbox": [ + 470, + 91, + 479, + 100 + ], + "type": "text", + "content": "21" + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 20 + }, + { + "para_blocks": [ + { + "bbox": [ + 132, + 116, + 481, + 665 + ], + "type": "list", + "angle": 0, + "index": 19, + "blocks": [ + { + "bbox": [ + 133, + 116, + 481, + 149 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 116, + 481, + 149 + ], + "spans": [ + { + "bbox": [ + 133, + 116, + 481, + 149 + ], + "type": "text", + "content": "112. Wang, H., O'Brien, J.F., Ramamoorthi, R.: Data-driven elastic models for cloth: modeling and measurement. ACM transactions on graphics (TOG) 30(4), 1-12 (2011)" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 132, + 149, + 481, + 182 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 149, + 481, + 182 + ], + "spans": [ + { + "bbox": [ + 132, + 149, + 481, + 182 + ], + "type": "text", + "content": "113. Wang, P., Liu, L., Liu, Y., Theobalt, C., Komura, T., Wang, W.: Neus: Learning neural implicit surfaces by volume rendering for multi-view reconstruction. Advances in Neural Information Processing Systems 34, 27171-27183 (2021)" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 132, + 182, + 481, + 214 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 182, + 481, + 214 + ], + "spans": [ + { + "bbox": [ + 132, + 182, + 481, + 214 + ], + "type": "text", + "content": "114. Wang, S., Schwarz, K., Geiger, A., Tang, S.: Arah: Animatable volume rendering of articulated human sdfs. In: European conference on computer vision. pp. 1-19. Springer (2022)" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 132, + 214, + 481, + 236 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 214, + 481, + 236 + ], + "spans": [ + { + "bbox": [ + 132, + 214, + 481, + 236 + ], + "type": "text", + "content": "115. Wang, S., Schwarz, K., Geiger, A., Tang, S.: Arah: Animatable volume rendering of articulated human sdfs. In: European Conference on Computer Vision (2022)" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 132, + 236, + 481, + 278 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 236, + 481, + 278 + ], + "spans": [ + { + "bbox": [ + 132, + 236, + 481, + 278 + ], + "type": "text", + "content": "116. Wang, T., Zhang, B., Zhang, T., Gu, S., Bao, J., Baltrusaitis, T., Shen, J., Chen, D., Wen, F., Chen, Q., et al.: Rodin: A generative model for sculpting 3d digital avatars using diffusion. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 4563-4573 (2023)" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 132, + 278, + 481, + 311 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 278, + 481, + 311 + ], + "spans": [ + { + "bbox": [ + 132, + 278, + 481, + 311 + ], + "type": "text", + "content": "117. Wu, G., Yi, T., Fang, J., Xie, L., Zhang, X., Wei, W., Liu, W., Tian, Q., Wang, X.: 4d gaussian splatting for real-time dynamic scene rendering. ArXiv abs/2310.08528 (2023). https://doi.org/10.48550/arXiv.2310.08528" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 132, + 311, + 481, + 354 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 311, + 481, + 354 + ], + "spans": [ + { + "bbox": [ + 132, + 311, + 481, + 354 + ], + "type": "text", + "content": "118. Xiang, D., Bagautdinov, T., Stuyck, T., Prada, F., Romero, J., Xu, W., Saito, S., Guo, J., Smith, B., Shiratori, T., et al.: Dressing avatars: Deep photorealistic appearance for physically simulated clothing. ACM Transactions on Graphics (TOG) 41(6), 1-15 (2022)" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 132, + 354, + 481, + 386 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 354, + 481, + 386 + ], + "spans": [ + { + "bbox": [ + 132, + 354, + 481, + 386 + ], + "type": "text", + "content": "119. Xu, Y., Yifan, W., Bergman, A.W., Chai, M., Zhou, B., Wetzstein, G.: Efficient 3d articulated human generation with layered surface volumes. arXiv preprint arXiv:2307.05462 (2023)" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 132, + 386, + 481, + 418 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 386, + 481, + 418 + ], + "spans": [ + { + "bbox": [ + 132, + 386, + 481, + 418 + ], + "type": "text", + "content": "120. Xu, Y., Chen, B., Li, Z., Zhang, H., Wang, L., Zheng, Z., Liu, Y.: Gaussian head avatar: Ultra high-fidelity head avatar via dynamic gaussians. arXiv preprint arXiv:2312.03029 (2023)" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 132, + 418, + 481, + 451 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 418, + 481, + 451 + ], + "spans": [ + { + "bbox": [ + 132, + 418, + 481, + 451 + ], + "type": "text", + "content": "121. Yang, Z., Zeng, A., Yuan, C., Li, Y.: Effective whole-body pose estimation with two-stages distillation. In: Proceedings of the IEEE/CVF International Conference on Computer Vision. pp. 4210-4220 (2023)" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 132, + 450, + 481, + 472 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 450, + 481, + 472 + ], + "spans": [ + { + "bbox": [ + 132, + 450, + 481, + 472 + ], + "type": "text", + "content": "122. Yariv, L., Gu, J., Kasten, Y., Lipman, Y.: Volume rendering of neural implicit surfaces. Advances in Neural Information Processing Systems 34, 4805-4815 (2021)" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 132, + 472, + 481, + 504 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 472, + 481, + 504 + ], + "spans": [ + { + "bbox": [ + 132, + 472, + 481, + 504 + ], + "type": "text", + "content": "123. Yee, K.S., Chen, J.S.: The finite-difference time-domain (fdtd) and the finite-volume time-domain (fvtd) methods in solving maxwell's equations. IEEE Transactions on Antennas and Propagation 45(3), 354-363 (1997)" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 132, + 504, + 481, + 536 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 504, + 481, + 536 + ], + "spans": [ + { + "bbox": [ + 132, + 504, + 481, + 536 + ], + "type": "text", + "content": "124. Yifan, W., Serena, F., Wu, S., Öztireli, C., Sorkine-Hornung, O.: Differentiable surface splatting for point-based geometry processing. ACM Transactions on Graphics (TOG) 38(6), 1-14 (2019)" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 132, + 536, + 481, + 580 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 536, + 481, + 580 + ], + "spans": [ + { + "bbox": [ + 132, + 536, + 481, + 580 + ], + "type": "text", + "content": "125. Yu, T., Zheng, Z., Zhong, Y., Zhao, J., Dai, Q., Pons-Moll, G., Liu, Y.: Simulcap: Single-view human performance capture with cloth simulation. In: Proceedings of the IEEE/CVF conference on computer vision and pattern recognition. pp. 5504-5514 (2019)" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 132, + 580, + 481, + 612 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 580, + 481, + 612 + ], + "spans": [ + { + "bbox": [ + 132, + 580, + 481, + 612 + ], + "type": "text", + "content": "126. Yuan, Y.J., Sun, Y.T., Lai, Y.K., Ma, Y., Jia, R., Gao, L.: Nerf-editing: geometry editing of neural radiance fields. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 18353-18364 (2022)" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 132, + 612, + 481, + 633 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 612, + 481, + 633 + ], + "spans": [ + { + "bbox": [ + 132, + 612, + 481, + 633 + ], + "type": "text", + "content": "127. Zeller, C.: Cloth simulation on thegpu. In: ACM SIGGRAPH 2005 Sketches, pp. 39-es (2005)" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 132, + 633, + 481, + 665 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 633, + 481, + 665 + ], + "spans": [ + { + "bbox": [ + 132, + 633, + 481, + 665 + ], + "type": "text", + "content": "128. Zhang, J.X., Lin, G.W.C., Bode, L., Chen, H.y., Stuyck, T., Larionov, E.: Estimating cloth elasticity parameters from homogenized yarn-level models. arXiv preprint arXiv:2401.15169 (2024)" + } + ] + } + ], + "index": 18 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "spans": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "type": "text", + "content": "22" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 166, + 91, + 239, + 102 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 166, + 91, + 239, + 102 + ], + "spans": [ + { + "bbox": [ + 166, + 91, + 239, + 102 + ], + "type": "text", + "content": "Yang Zheng et al." + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 21 + }, + { + "para_blocks": [ + { + "bbox": [ + 133, + 116, + 481, + 237 + ], + "type": "list", + "angle": 0, + "index": 6, + "blocks": [ + { + "bbox": [ + 133, + 116, + 481, + 149 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 116, + 481, + 149 + ], + "spans": [ + { + "bbox": [ + 133, + 116, + 481, + 149 + ], + "type": "text", + "content": "129. Zhang, R., Isola, P., Efros, A.A., Shechtman, E., Wang, O.: The unreasonable effectiveness of deep features as a perceptual metric. In: Proceedings of the IEEE conference on computer vision and pattern recognition. pp. 586-595 (2018)" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 133, + 150, + 481, + 183 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 150, + 481, + 183 + ], + "spans": [ + { + "bbox": [ + 133, + 150, + 481, + 183 + ], + "type": "text", + "content": "130. Zhao, F., Jiang, Y., Yao, K., Zhang, J., Wang, L., Dai, H., Zhong, Y., Zhang, Y., Wu, M., Xu, L., et al.: Human performance modeling and rendering via neural animated mesh. ACM Transactions on Graphics (TOG) 41(6), 1-17 (2022)" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 133, + 183, + 481, + 215 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 183, + 481, + 215 + ], + "spans": [ + { + "bbox": [ + 133, + 183, + 481, + 215 + ], + "type": "text", + "content": "131. Zheng, Z., Huang, H., Yu, T., Zhang, H., Guo, Y., Liu, Y.: Structured local radiance fields for human avatar modeling. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 15893-15903 (2022)" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 133, + 216, + 481, + 237 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 216, + 481, + 237 + ], + "spans": [ + { + "bbox": [ + 133, + 216, + 481, + 237 + ], + "type": "text", + "content": "132. Zielonka, W., Bagautdinov, T., Saito, S., Zollhöfer, M., Thies, J., Romero, J.: Drivable 3d gaussian avatars. arXiv preprint arXiv:2311.08581 (2023)" + } + ] + } + ], + "index": 5 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 399, + 91, + 447, + 101 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 399, + 91, + 447, + 101 + ], + "spans": [ + { + "bbox": [ + 399, + 91, + 447, + 101 + ], + "type": "text", + "content": "PhysAvatar" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 470, + 91, + 481, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 470, + 91, + 481, + 100 + ], + "spans": [ + { + "bbox": [ + 470, + 91, + 481, + 100 + ], + "type": "text", + "content": "23" + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 22 + } + ], + "_backend": "vlm", + "_version_name": "2.6.4" +} \ No newline at end of file diff --git a/2024/PhysGen_ Rigid-Body Physics-Grounded Image-to-Video Generation/63954133-cd49-423f-97fe-3962651db98c_content_list.json b/2024/PhysGen_ Rigid-Body Physics-Grounded Image-to-Video Generation/63954133-cd49-423f-97fe-3962651db98c_content_list.json new file mode 100644 index 0000000000000000000000000000000000000000..a91b52ac0d1fb98d6da0227584eb5f986485bd7d --- /dev/null +++ b/2024/PhysGen_ Rigid-Body Physics-Grounded Image-to-Video Generation/63954133-cd49-423f-97fe-3962651db98c_content_list.json @@ -0,0 +1,1852 @@ +[ + { + "type": "text", + "text": "PhysGen: Rigid-Body Physics-Grounded Image-to-Video Generation", + "text_level": 1, + "bbox": [ + 269, + 140, + 733, + 186 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Shaowei Liu, Zhongzheng Ren, Saurabh Gupta*, and Shenlong Wang*", + "bbox": [ + 246, + 212, + 754, + 228 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "University of Illinois Urbana-Champaign https://stevenlsw.github.io/physgen/", + "bbox": [ + 343, + 239, + 656, + 270 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Abstract. We present PhysGen, a novel image-to-video generation method that converts a single image and an input condition (e.g., force and torque applied to an object in the image) to produce a realistic, physically plausible, and temporally consistent video. Our key insight is to integrate model-based physical simulation with a data-driven video generation process, enabling plausible image-space dynamics. At the heart of our system are three core components: (i) an image understanding module that effectively captures the geometry, materials, and physical parameters of the image; (ii) an image-space dynamics simulation model that utilizes rigid-body physics and inferred parameters to simulate realistic behaviors; and (iii) an image-based rendering and refinement module that leverages generative video diffusion to produce realistic video footage featuring the simulated motion. The resulting videos are realistic in both physics and appearance and are even precisely controllable, showcasing superior results over existing data-driven image-to-video generation works through quantitative comparison and comprehensive user study. PhysGen's resulting videos can be used for various downstream applications, such as turning an image into a realistic animation or allowing users to interact with the image and create various dynamics.", + "bbox": [ + 261, + 301, + 759, + 566 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "1 Introduction", + "text_level": 1, + "bbox": [ + 215, + 590, + 375, + 606 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Looking at the images in Fig. 1, we can effortlessly predict or visualize the potential outcomes of various physical effects applied to the car, the curling stone, and the stack of dominos. This understanding of physics empowers our imagination of the counterfactual: we can mentally imagine various consequences of an action from an image without experiencing them. We aim to provide computers with similar capabilities - understanding and simulating physics from a single image and creating realistic animations. We expect the resulting video to produce realistic animations with physically plausible dynamics and interactions.", + "bbox": [ + 212, + 619, + 787, + 739 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Tremendous progress has been made in generating realistic and physically plausible video footage. Conventional graphics leverage model-based dynamics to simulate plausible motions and utilize a graphics renderer to simulate images. Such methods provide realistic and controllable dynamics. However, both the dynamics of physics and lighting physics are predefined, making it hard to achieve", + "bbox": [ + 212, + 741, + 787, + 816 + ], + "page_idx": 0 + }, + { + "type": "page_footnote", + "text": "* Equal advising", + "bbox": [ + 230, + 825, + 346, + 840 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/1ffc6d1de36ee10d710eb7d62fbced194ca15a6efa6ad2e5bbe333392789cf6a.jpg", + "image_caption": [ + "Input", + "Fig. 1: Given a single image, training-free PhysGen generates future frames controlled by physics and initial state. Semantics, geometry and dynamics are reasoned during generation and the result videos are both physics-grounded and photo-realistic." + ], + "image_footnote": [], + "bbox": [ + 228, + 157, + 781, + 372 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "our goal of animating an image captured in the real world. On the other hand, data-driven image-to-video (I2V) generation techniques [6, 8, 57] learn to create realistic videos from a single image through training diffusion models over internet-scale data. However, despite advancements in video generative models [6, 8, 57], the incorporation of real-world physics principles into the video generation process remains largely unexplored and unsolved. Consequently, the synthesized videos often lack temporal coherence and fail to replicate authentic object motions observed in reality. Furthermore, such text-driven methods also lack fine-grained controllability. For instance, they cannot simulate the consequences of different forces and torques applied to an object.", + "bbox": [ + 212, + 460, + 787, + 611 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "In light of this gap, we propose a paradigm shift in video generation through the introduction of model-based video generative models. In contrast to existing purely generative methods, which are trained in a data-driven manner and rely on diffusion models to learn image space dynamics via scaling laws, we propose grounding object dynamics explicitly using rigid body physics, thereby integrating fundamental physical principles into the generation process. As classical mechanics theory reveals, the motion of an object is determined by its physical properties (e.g., mass, elasticity, surface roughness) and external factors (e.g., external forces, environmental conditions, boundary conditions).", + "bbox": [ + 212, + 612, + 787, + 748 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "We tackle our image-to-video generation problem via a computational framework, PhysGen. PhysGen consists of three stages: 1) image-based physics understanding; 2) model-based dynamics simulation; and 3) generative video rendering with dynamics guidance. Our method first infers object compositions and physical parameters from the input image through large visual foundation model-based reasoning (Sec. 3.1). Given an input force, we then utilize the in", + "bbox": [ + 212, + 750, + 787, + 840 + ], + "page_idx": 1 + }, + { + "type": "page_number", + "text": "2", + "bbox": [ + 217, + 114, + 228, + 126 + ], + "page_idx": 1 + }, + { + "type": "header", + "text": "S. Liu et al.", + "bbox": [ + 271, + 114, + 352, + 126 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "ferred physical parameters to perform realistic dynamic simulations for each object, accounting for their rigid body dynamics, collisions, frictions, and elasticity (Sec. 3.2). Finally, guided by motion and the input image, we present a novel generative video diffusion-based renderer that outputs the final realistic video footage (Sec. 3.3). Several illustrative examples of our framework are shown in Fig. 1. As shown in this figure, our approach combines the best of the worlds of data-driven generative modeling's appearance realism and model-based dynamics's verified physical plausibility. Notably, our generation process is highly controllable, allowing users to specify underlying physics parameters and initial conditions. This capability facilitates a range of interactive applications. Moreover, our proposed generation pipeline operates solely during inference time, eliminating the need for any training.", + "bbox": [ + 212, + 145, + 787, + 327 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "We evaluate PhysGen generation ability on multiple data source including the web and self-captured images. We compare against both state-of-the-art image-to-video models [18, 87, 95] as well as image editing method [28] quantitatively and qualitatively. We also perform user-study to evaluate the physical-realism and photo-realism of the generated video. Our method demonstrates physics-informed results for controllable image-to-video generation, producing realistic video sequences with grounded rigid-body dynamics and interaction. PhysGen combines learning-based generative approaches with traditional model-based physics to deliver visually appealing results without any training. Our approach harnesses a plethora of ingredients from previous works, particularly recent progress in large foundational visual models for segmentation [40, 48, 61, 93], physical understanding [2, 56], normal estimation [23, 26], relighting, and generative rendering [8, 18]. While these individual ingredients exist, our system combines them in a novel manner to enable an exciting new capacity, featuring physical plausibility for video generation at an unprecedented level.", + "bbox": [ + 212, + 329, + 787, + 554 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Our contributions can be summarized as follows: 1) We propose a novel image-space dynamics model that can understand physical parameters from a single image and simulate realistic, controllable, and interactive dynamics. 2) We present a novel image-to-video system by integrating our presented dynamics model with generative diffusion-based video priors. The model is training-free, and the resulting videos are highly realistic, physically plausible, and controllable, consistently surpassing existing state-of-the-art video generative models.", + "bbox": [ + 212, + 555, + 787, + 662 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "2 Related work", + "text_level": 1, + "bbox": [ + 214, + 685, + 383, + 700 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Image-based physics reasoning and simulation. Our image-based dynamics model is built upon physical simulation. Symbolic physics simulators or physics engines have been extensively studied for various physical processes, including rigid body dynamics, fluid simulation, and deformable objects [36-38, 41, 49, 55, 72]. Such methods have been widely used across graphics, scientific discovery, robotics, and video effects. However, these rule-based or solver-based simulators face limitations in terms of expressiveness, efficiency, generalizability, and parameter tuning. To overcome these challenges, neural physics, which combines", + "bbox": [ + 212, + 719, + 787, + 840 + ], + "page_idx": 2 + }, + { + "type": "header", + "text": "PhysGen: Rigid Physics-Grounded Img2Vid", + "bbox": [ + 437, + 114, + 730, + 128 + ], + "page_idx": 2 + }, + { + "type": "page_number", + "text": "3", + "bbox": [ + 774, + 116, + 784, + 126 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "traditional physics simulators/engines with deep neural networks, has been developed to model dynamic processes such as planar pushing and bouncing [4], as well as object interactions [3]. Domain-specific network architectures [7,45,54,63] have been developed to address various physical problems. Additionally, the integration of neural scene representations or large language models with physics simulators [43,50,86] has also been investigated recently. Nevertheless, most forward physics models, whether neural or not, have not resolved the dependency on pre-defined physical parameters, which are often specified by users.", + "bbox": [ + 212, + 146, + 787, + 267 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Reasoning physical parameters from visual data allows us to simulate without manually defined physics, further enhancing convenience and realism for image-based physics simulation. To enable this, subsequent research efforts [21, 43, 44, 80-82, 84, 88] have introduced pipelines that first extract scene representations of the physical world from images or videos using neural networks, and then utilize these representations for either physics reasoning or simulation. Despite the promise of these approaches, they primarily rely on synthetic data for training, which might suffer from generalization issues, or on inverse physics, which requires inference from observed physical phenomena, often requiring video observations instead of a single image. In our work, we aim to generate physics-grounded, high-fidelity predictive videos from static input images of real-world objects with complex backgrounds. Contrary to existing work, we investigate leveraging the power of large pretrained visual foundational models [2, 40, 56] for physics reasoning in a zero-shot manner.", + "bbox": [ + 212, + 268, + 787, + 479 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Video generative model. Video generative models have experienced remarkable advancements in recent years [6,8,12,18,27,29,31,33,42,57,65,74,76,87,92, 95], with the state-of-the-art framework [57] achieving the capability to generate photo-realistic and coherent imaginative videos from text instructions using diffusion models [34,58,62,66,68]. Despite progress, challenges remain, such as the need for extensive training data, object permanence, realistic lighting, and reducing unrealistic motion and physics violations. Overcoming these is key for advancing video generative modeling and its applications, especially in physics-accurate areas like scientific discovery and robotics. Our work addresses some issues by proposing a novel image-to-video framework that merges a model-based approach with a pretrained video prior, grounding the generated content with real-world physics.", + "bbox": [ + 212, + 493, + 787, + 675 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Image animation. Our method is heavily inspired by image-based animation [19, 22, 35, 64, 70, 77, 89] and cinematographs. One common way to improve quality is to incorporate class-specific heuristics [19, 39], which can be physically simulated and integrated into the generation process. Recent advancements in deep learning have led to data-driven solutions in this domain, where temporal neural networks are trained on video datasets to directly predict consecutive video frames from an input image [10, 17, 24, 30, 35, 75]. Moreover, there is growing interest in interactive [9, 11, 46] and controllable [1, 22] image-to-video synthesis. To further improve temporal coherence, more priors such as motion fields [24, 35, 51, 52, 69, 89] or flows [14, 28, 96], 3D geometry information [60, 79],", + "bbox": [ + 212, + 688, + 787, + 842 + ], + "page_idx": 3 + }, + { + "type": "page_number", + "text": "4", + "bbox": [ + 217, + 114, + 228, + 126 + ], + "page_idx": 3 + }, + { + "type": "header", + "text": "S. Liu et al.", + "bbox": [ + 271, + 114, + 352, + 126 + ], + "page_idx": 3 + }, + { + "type": "image", + "img_path": "images/6e8d796f456560f66846d4cb4bd2805cf5c761bfe7fdda1da505e68b384f37f9.jpg", + "image_caption": [ + "Fig. 2: Method overview. Our framework consists of three interleaved components: the perception module, the dynamics simulation module, and the rendering module. The perception module interprets the semantics, geometry, and physical parameters in the given image. The dynamics simulation module simulates the rigid-body motion and interactions of each instance in the scene, governed by Newton's Laws and physical constraints. The rendering module renders the final outcome, leveraging an off-the-shelf relighting model and diffusion-model-based video priors." + ], + "image_footnote": [], + "bbox": [ + 243, + 143, + 759, + 294 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "and user annotations [46] have been subsequently introduced. In this work, we propose a model-based solution and tackle the image animation task from a new perspective, where our method takes a single image and user-specific initial force or torque as input, and models object dynamics explicitly via rigid physics simulation. Our key insight is that physical parameters can be inferred from a single image automatically through large foundation models such as GPT-4V [2, 56].", + "bbox": [ + 212, + 430, + 787, + 521 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "3 Approach", + "text_level": 1, + "bbox": [ + 215, + 563, + 346, + 579 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Given a single input image, our goal is to generate a realistic $T$ -frame video featuring rigid body dynamics and interactions. Our generation can be conditioned on user-specific inputs in the form of an initial force $\\mathbf{F}_0$ and/or torque $\\pmb{\\tau}_{0}$ . Our key insight is to incorporate physics-informed simulation into the generation process, ensuring the dynamics are realistic and interactive, thus producing large, plausible object motion. To achieve this, we present a novel framework consisting of three stages: physical understanding, dynamics generation, and generative rendering. The perception module aims to reason about the semantics, collision geometry, and physics of the objects presented in the image (Sec. 3.1). The dynamics module then leverages the input force/torque, as well as the inferred shape and geometry, to simulate the rigid-body motions of the objects, as well as object-object and object-scene interactions, such as friction and collisions (Sec. 3.2). Finally, we convert the simulated dynamics into pixel-level motion and combine image-based warping and generative video to render the final video (Sec. 3.3). Fig. 2 depicts the overall framework.", + "bbox": [ + 212, + 613, + 787, + 840 + ], + "page_idx": 4 + }, + { + "type": "header", + "text": "PhysGen: Rigid Physics-Grounded Img2Vid", + "bbox": [ + 437, + 114, + 730, + 128 + ], + "page_idx": 4 + }, + { + "type": "page_number", + "text": "5", + "bbox": [ + 774, + 116, + 784, + 126 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "3.1 Perception", + "text_level": 1, + "bbox": [ + 215, + 146, + 352, + 161 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Simulating dynamics on an image requires a holistic understanding of object compositions, materials, geometry, and physics. Toward this goal, we designed our perception module to infer such properties from a single image. Our key insight is to harness readily available, large pretrained models [2, 15, 26, 56, 61] to achieve this goal, as shown in Fig. 2.", + "bbox": [ + 212, + 172, + 785, + 248 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Segmentation. Identifying and segmenting each individual physical entity lays the foundation for image-based dynamics. Inspired by the recent success of large pretrained segmentation models, we incorporate GPT-4V [2, 56] to recognize all image categories and send to Grounded-SAM [61] to detect and segment each individual instance-level objects $\\{\\mathbf{o}^i\\in \\mathbb{R}^{W\\times H}\\}_{i}^{N}$ , where $\\mathbf{o}^i$ is the binary mask for $i$ -th object. We also query the GPT-4V to classify the objects into foreground and background based on their movability. The foreground objects are send to the physical simulation once user inputs are applied. We extract collision boundaries and supporting edges (e.g., ground, walls, etc.) from the background objects. The details are presented in the supplementary.", + "bbox": [ + 212, + 263, + 787, + 415 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Physical properties reasoning. Realistic physics depends on accurate physical parameters, such as surface friction, mass, and elasticity. Unlike prior physics-based image dynamics work [19], our work leverages visual foundational models to reason physical properties directly. Our solution is simple yet effective. Inspired by the success of mask-based prompting [2,56,90], we directly ask GPT-4V [56] for certain physical properties, providing an object mask overlaid on the input image. Following [85,94], we send the crafted prompt to GPT-4V, which contains GPT-4V instructions to return a quantitative measure of each queried property in a metric unit. For each object, we query its mass, elasticity, and friction coefficient $(M_i,E_i,\\mu_i)$ . We use the Coulomb friction model [59] for friction modeling. The elasticity coefficient is a scalar, where 0.0 results in no bounce, and 1.0 results in a perfect bounce. Given each object mass $M_{i}$ in grams and its shape primitive, we compute the rotational inertia $\\mathbf{I}_i$ accordingly.", + "bbox": [ + 212, + 431, + 787, + 627 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Geometry primitives. Rasterized instance masks are not suitable for physical simulation; hence, we need to convert each object into vectorized shape primitives. We fit two types of primitive shapes: a circle and generic polygons. For each object, we choose the primitive type with the maximum coverage of its segmentation. For circles, we fit the center and radius to cover the segmentation mask. Circles allow us to realistically simulate rolling motions. For non-circle objects, we instead fit the generic polygons. Specifically, we perform contour extraction on the segmentation masks to obtain the polygon vertices.", + "bbox": [ + 212, + 642, + 787, + 763 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Intrinsic decomposition. The dynamic movement of the objects will also result in comprehensive changes in shading, as their position wrt the lighting environment changes. To compensate for this effect during rendering, we must perform image decomposition to infer albedo, normal as well as lighting from", + "bbox": [ + 212, + 779, + 787, + 840 + ], + "page_idx": 5 + }, + { + "type": "page_number", + "text": "6", + "bbox": [ + 217, + 114, + 228, + 126 + ], + "page_idx": 5 + }, + { + "type": "header", + "text": "S. Liu et al.", + "bbox": [ + 271, + 114, + 352, + 127 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "the single image for each object $(\\mathbf{A}_0^i,\\mathbf{N}_0^i,\\mathbf{L}_0)$ . To achieve this we leverage off-the-shelf intrinsic decomposition model [15] to compute $\\mathbf{A}_0^i$ and $\\mathbf{L}_0$ and exploits pretrained surface normal estimator for the computation of object normal $\\mathbf{N}_0^i$ .", + "bbox": [ + 214, + 145, + 782, + 193 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Background inpainting. Once the foreground objects move, they will leave holes in the background. To produce realistic and complete video footage, we leverage an off-the-shelf generative image inpainting model [93] to recover the complete background scene without foreground objects. The inpainted image $\\mathbf{B} \\in \\mathbb{R}^{W \\times H \\times 3}$ will be used for composition in the rendering stage.", + "bbox": [ + 214, + 208, + 782, + 284 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "3.2 Image space dynamics simulation", + "text_level": 1, + "bbox": [ + 214, + 309, + 535, + 324 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Given the foreground objects with physical properties, we use rigid-body physics for dynamics simulation in image spaces. We choose to perform simulations in image spaces for three reasons: 1) image-space dynamics are better coupled with our output video; 2) a full 3D simulation requires complete 3D scene reconstruction and understanding, which remains an unsolved problem from a single image; 3) image-space dynamics can already cover various object dynamics and have been widely used in prior work [4, 19, 35, 46, 80]. Specifically, at a given time $t$ , each rigid object $i$ is characterized by its 2D pose and velocity at its center of mass. The position includes a translation $\\mathbf{t}^i(t) \\in \\mathbb{R}^2$ and a rotation $\\mathbf{R}^i(t) \\in \\mathbb{S}\\mathbb{O}(2)$ specified in a world coordinate. The velocity includes linear velocity $\\pmb{\\nu}^i(t) \\in \\mathbb{R}^2$ and angular velocity $\\pmb{\\omega}^i(t) \\in \\mathbb{R}^2$ . Hence, the state of the each object $i$ at time $t$ can be represented as $\\mathbf{q}^i(t) = [\\mathbf{t}^i(t), \\mathbf{R}^i(t), \\pmb{\\nu}^i(t), \\pmb{\\omega}^i(t)]$ .", + "bbox": [ + 212, + 335, + 784, + 518 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Following [25], the rigid body motion dynamics is given by Eq. (1) for each object. For simplicity, we omit the object index $i$ in the following:", + "bbox": [ + 214, + 518, + 782, + 550 + ], + "page_idx": 6 + }, + { + "type": "equation", + "text": "\n$$\n\\frac {d}{d t} \\mathbf {q} (t) = \\frac {d}{d t} \\left[ \\begin{array}{l} \\mathbf {t} (t) \\\\ \\mathbf {R} (t) \\\\ \\boldsymbol {\\nu} (t) \\\\ \\boldsymbol {\\omega} (t) \\end{array} \\right] = \\left[ \\begin{array}{c} \\mathbf {v} (t) \\\\ \\boldsymbol {\\omega} (t) \\times \\mathbf {R} (t) \\\\ \\frac {\\mathbf {F} (t)}{M} \\\\ \\mathbf {I} (t) ^ {- 1} (\\boldsymbol {\\tau} - \\boldsymbol {\\omega} (t) \\times \\mathbf {I} (t) \\boldsymbol {\\omega} (t)) \\end{array} \\right] \\tag {1}\n$$\n", + "text_format": "latex", + "bbox": [ + 307, + 560, + 784, + 626 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "where $\\mathbf{F}$ is the force, $\\pmb{\\tau}$ is the torque, $M$ is the mass of the object, $\\mathbf{I}(t) = \\mathbf{R}(t)\\mathbf{I}\\mathbf{R}(t)^T$ is the rotation inertia in world coordinates, $\\mathbf{I} \\in \\mathbb{R}^{2 \\times 2}$ is the inertia matrix in its body coordinates. This inertia indicates its resistance against rotational motion. The state $\\mathbf{q}(t)$ is given by integrating Eq. (1):", + "bbox": [ + 214, + 637, + 782, + 699 + ], + "page_idx": 6 + }, + { + "type": "equation", + "text": "\n$$\n\\mathbf {q} (t) = \\mathbf {q} (0) + \\int_ {0} ^ {t} \\frac {d}{d t} \\mathbf {q} (t) d t = \\mathbf {q} (0) + \\sum_ {i = 1} ^ {T} \\frac {d}{d t} \\mathbf {q} (t) | _ {t = t _ {i}} \\Delta_ {t} \\tag {2}\n$$\n", + "text_format": "latex", + "bbox": [ + 305, + 710, + 784, + 752 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "where $\\mathbf{q}(0)$ is the initial condition specified in the input image. We compute the integral using numerical ODE integrations, e.g. Euler method [20]. Given the initial state of the objects, the physical motion simulation module synthesis each foreground object future motion. The location of each object is updated by the affine transformation $\\mathbf{T}(t) = [\\mathbf{R}(t),\\mathbf{t}(t)]$ from the initial location.", + "bbox": [ + 214, + 763, + 782, + 840 + ], + "page_idx": 6 + }, + { + "type": "header", + "text": "PhysGen: Rigid Physics-Grounded Img2Vid", + "bbox": [ + 437, + 113, + 730, + 128 + ], + "page_idx": 6 + }, + { + "type": "page_number", + "text": "7", + "bbox": [ + 774, + 114, + 784, + 126 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/36ea1b241fe3a56297e00bc470356239e8a695cb84b3f184cf954e3b8380d2b6.jpg", + "image_caption": [ + "Input" + ], + "image_footnote": [], + "bbox": [ + 267, + 143, + 385, + 234 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/37c392b76552cd52195b85e02498cf7e45a5809b6650bed56d0f4ac515e1c776.jpg", + "image_caption": [ + "Composed (Sec. 3.3)", + "Fig. 3: Rendered video comparison. We show a toy example of piggy bank. The left shows the input frame, and the rest 3 are future frame generations. The composited piggy bank hasn't been aware of the light change. The relighted output synthesized no shadows beneath. The rendered output from diffusion model is most photo realistic." + ], + "image_footnote": [], + "bbox": [ + 385, + 143, + 504, + 234 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/313a06508e7cc87561a7dc9dc12beab66d4812a8c921ade0b5e94f845501af37.jpg", + "image_caption": [ + "Relight (Sec. 3.3)" + ], + "image_footnote": [], + "bbox": [ + 504, + 143, + 620, + 234 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/574ab8361d0e985d1af4576d717f32c5fb8f8bfdef45407afb47f7810c62980b.jpg", + "image_caption": [ + "Render (Sec. 3.4)" + ], + "image_footnote": [], + "bbox": [ + 620, + 143, + 738, + 234 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "External force and torque. For each object, the external forces $\\mathbf{F}(t)$ and torques $\\pmb{\\tau}$ include gravity, friction, and elasticity between the object's surface and the environment. Specifically, we consider the initial external force and torque from user input, as well as gravity, rolling friction, and sliding friction.", + "bbox": [ + 212, + 342, + 782, + 402 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Collision. When objects move, they might collide with each other or with the scene boundary (e.g., hitting a wall). Therefore, collision checking is necessary at every time step. A collision will result in offset forces being applied from its center of mass, producing $\\tau$ and causing the object to rotate. In collision reactions, changes in energy and momentum are minimized to adhere to Newton's conservation laws. In the case of an object falling, it could bounce back or start rolling on the ground due to the elasticity of both the object and the ground.", + "bbox": [ + 212, + 417, + 784, + 523 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "3.3 Rendering", + "text_level": 1, + "bbox": [ + 214, + 545, + 348, + 560 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Given the simulated motion sequence $\\{\\mathbf{q}^i(0),\\dots,\\mathbf{q}^i(t),\\dots\\}_i^N$ and the input image $\\mathbf{C}$ , the rendering module outputs the rendered video $\\{\\dots,\\mathbf{C}(t),\\dots\\}$ . There are a few desiderata for generating realistic video: it needs to be temporally consistent, complete, reflect the lighting changes as the object moves, and accurately represent the simulated image space dynamics. To achieve this goal, we design a novel motion-guided rendering pipeline consisting of an image-based transformation and composition module to ensure motion awareness, a relighting module to ensure the plausibility of lighting physics and a generative video editing module that retouches the composed and relit video to ensure realism.", + "bbox": [ + 212, + 570, + 787, + 708 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Composed video. Given the object state from the physical motion module, we compose an initial video by alpha-blending the foreground scene with the static inpainted background from Sec. 3.1. The foreground scene is rendered by performing forward-warping from the input image to future frames using the affine transformation $\\mathbf{T}(t)$ . The alpha channel is computed using the same procedure, with the input being the segmentation mask in Sec. 3.1:", + "bbox": [ + 212, + 720, + 787, + 811 + ], + "page_idx": 7 + }, + { + "type": "equation", + "text": "\n$$\n\\hat {\\mathbf {X}} (t) = \\operatorname {c o m p o s i t e} (\\mathbf {B}, \\operatorname {w a r p} \\left(\\mathbf {X} ^ {0}, \\mathbf {T} ^ {0} (t)\\right) \\dots \\operatorname {w a r p} \\left(\\mathbf {X} ^ {i}, \\mathbf {T} ^ {i} (t)\\right) \\dots \\dots)\n$$\n", + "text_format": "latex", + "bbox": [ + 272, + 823, + 725, + 840 + ], + "page_idx": 7 + }, + { + "type": "page_number", + "text": "8", + "bbox": [ + 217, + 114, + 228, + 126 + ], + "page_idx": 7 + }, + { + "type": "header", + "text": "S. Liu et al.", + "bbox": [ + 271, + 114, + 352, + 126 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "where $\\mathbf{X}^i$ are segmented input image of $i$ -th object. Apart from the RGB sequence $\\hat{\\mathbf{V}} = \\{\\dots \\hat{\\mathbf{X}}(t) \\dots\\}$ , we also use the same affine warping and composition to compute the blended albedo map $\\hat{\\mathbf{A}}(t)$ , which is later used for relighting.", + "bbox": [ + 212, + 146, + 782, + 193 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "Relighting. Our relighting module simulates the changes in shading due to object movement. Specifically, it takes as input the transformed albedo $\\hat{\\mathbf{A}}(t)$ and RGB image $\\hat{\\mathbf{X}}(t)$ at each time $t$ , as well as the estimated lighting $\\mathbf{L}$ . We first compute the surface normal $\\hat{\\mathbf{N}}(t)$ from the composited RGB image $\\hat{\\mathbf{X}}(t)$ using an off-the-shelf normal estimator [23]. We then perform reshading using the parametric illumination model [15]: $\\tilde{\\mathbf{X}}(t) = f(\\hat{\\mathbf{X}}(t), \\hat{\\mathbf{A}}(t), \\hat{\\mathbf{N}}(t), \\mathbf{L})$ , which returns the relit image $\\hat{\\mathbf{X}}(t)$ for each frame $t$ .", + "bbox": [ + 212, + 207, + 784, + 315 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "3.4 Generative refinement with latent diffusion models", + "text_level": 1, + "bbox": [ + 214, + 338, + 679, + 352 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "One limitation of the single-image based relighting model is that it neglects the background and cannot handle complicated lighting effects due to complex object-object interactions, such as ambient occlusion and cast shadows. Moreover, the composition results in unnatural boundaries. To address these issues and enhance realism, we incorporated a diffusion-based video to refine the relit video, denoted as $\\tilde{\\mathbf{V}} = \\{\\ldots ,\\tilde{\\mathbf{X}} (t),\\ldots \\}$ , and obtain the final video output.", + "bbox": [ + 212, + 362, + 784, + 452 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "Specifically, we incorporate a pretrained video diffusion model to refine our video $\\tilde{\\mathbf{V}}$ . We first use the pretrained latent diffusion encoder [62] to encode the guidance video $\\tilde{\\mathbf{V}}$ into a latent code $\\mathbf{z}$ . Inspired by SDEdit [53], we add noise to the guided latent code and gradually denoise it using the pretrained video diffusion. Given that the content in the guided latent code is already satisfactory, we do not perform the denoising process from scratch. Instead, we define an initial noise strength $s$ where $s \\in [0,1]$ controls the amount of noise added to $\\mathbf{z}$ . In practice, we find that a certain approach finds a good trade-off between fidelity and realism. At each denoising step, we ensure the foreground objects are as consistent as possible with the guidance (copied from the guided latent code) while synthesizing new content (e.g., shadows) in the background (using the denoised latent code). To this end, we use a fusion weight $w$ to control how much of the generated latent to inject into the foreground; the fusion weight is gradually increased during the denoising as the perturbation of noise decreases. We also define a fusion timestamp $\\delta$ as the signal to stop the fusion. The final denoised latent code $\\mathbf{z}_{*}$ is then decoded to our final video output $\\mathbf{V}$ . Please see the supplementary material for a detailed algorithm.", + "bbox": [ + 212, + 453, + 785, + 708 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "Fig. 3 depicts the composed video $\\hat{\\mathbf{V}}$ , the relit video $\\tilde{\\mathbf{V}}$ , and our final output $\\mathbf{V}$ , illustrating the effectiveness of each rendering component.", + "bbox": [ + 214, + 709, + 782, + 739 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "4 Experiments", + "text_level": 1, + "bbox": [ + 214, + 763, + 375, + 780 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "Implementation details. For physical body simulation, we use a 2D rigid body physics simulator Pymunk [13]. For each experiment, we run 120 simulation steps and uniformly sample 16 frames to form the videos. We set the", + "bbox": [ + 212, + 794, + 784, + 839 + ], + "page_idx": 8 + }, + { + "type": "header", + "text": "PhysGen: Rigid Physics-Grounded Img2Vid", + "bbox": [ + 437, + 114, + 730, + 128 + ], + "page_idx": 8 + }, + { + "type": "page_number", + "text": "9", + "bbox": [ + 774, + 116, + 784, + 126 + ], + "page_idx": 8 + }, + { + "type": "image", + "img_path": "images/cb2e40043b69b091db787a08619c2f98e58996c50081e24ce8cae7a305a2026a.jpg", + "image_caption": [ + "Fig. 4: Video generation results. Left: Input initial frame with red arrows representing the initial force or speed on the object. Right: Generated future frames." + ], + "image_footnote": [], + "bbox": [ + 228, + 143, + 781, + 505 + ], + "page_idx": 9 + }, + { + "type": "text", + "text": "generation resolution at $512 \\times 512$ so that frames can easily fit into the diffusion models. For video diffusion prior model, we use SEINE [18]. For inference, DDIM sampling [67] is used and the noise strength $i$ set to $s = 0.5$ , i.e., executing 25/50 steps for denoising. The latents fusion stops at timestamp $\\delta = 5$ .", + "bbox": [ + 212, + 574, + 787, + 636 + ], + "page_idx": 9 + }, + { + "type": "text", + "text": "Speed. PhysGen's run-time speed is 3 minutes for one image-to-video gen on a Nvidia A40 GPU. Image understanding and relighting takes the most time.", + "bbox": [ + 215, + 648, + 787, + 679 + ], + "page_idx": 9 + }, + { + "type": "text", + "text": "Data. To show the generalizability and robustness of our method, we use both internet data and self-captured indoor images from a cell phone. The dataset is diverse, with variations in lighting, object count, geometries, physical attributes, and environmental boundaries. We compare different approaches on 15 photos that require physical reasoning. Additionally, we collect 50 recorded videos of a given scene as ground truth by varying input conditions for quantitative comparison. Randomly selected sequences are shown in the supplementary material.", + "bbox": [ + 212, + 691, + 787, + 797 + ], + "page_idx": 9 + }, + { + "type": "text", + "text": "Baselines. We compare against two streams of approaches. The first stream is current state-of-the-art I2V models: SEINE [18], I2VGen-XL [95] and Dynam-", + "bbox": [ + 215, + 809, + 785, + 840 + ], + "page_idx": 9 + }, + { + "type": "page_number", + "text": "10", + "bbox": [ + 217, + 114, + 235, + 126 + ], + "page_idx": 9 + }, + { + "type": "header", + "text": "S. Liu et al.", + "bbox": [ + 271, + 114, + 351, + 126 + ], + "page_idx": 9 + }, + { + "type": "table", + "img_path": "images/9ce64eef3463ca38a956bcad7c55ca7ddcf10ffd0b5a889f872bdfde2e50357c.jpg", + "table_caption": [], + "table_footnote": [], + "table_body": "
MethodsPhysical-realism↑Photo-realism↑
SEINE [18]1.391.86
DynamiCrafter [87]1.681.81
I2VGen-XL [95]2.112.25
Ours4.143.86
", + "bbox": [ + 299, + 143, + 707, + 219 + ], + "page_idx": 10 + }, + { + "type": "table", + "img_path": "images/afe00fb08c0a22f845ba4658f1e3226cb49a1fd98ab56830e94e21be0a00a4e8.jpg", + "table_caption": [ + "Table 1: Human evaluation. We evaluate both physical-realism and photo-realism of the three competing image-to-video models. The user is asked to evaluate the generated video using a 5-point scale from strongly disagree (1) to strongly agree(5) that the video is physical-realistic and photo-realistic. Our method ranks the first in both terms." + ], + "table_footnote": [], + "table_body": "
MethodsImage-FID↓Motion-FID↓
SEINE [18]138.8938.76
DynamiCrafter [87]99.64109.61
I2VGen-XL [95]104.6282.04
Ours105.7030.20
", + "bbox": [ + 305, + 303, + 697, + 378 + ], + "page_idx": 10 + }, + { + "type": "text", + "text": "Table 2: Quantitative evaluation. We measure the Image-FID and Motion-FID of videos generated by four methods against the collected GT videos for the given scene. Our method achieves both low Image-FID and Motion-FID.", + "bbox": [ + 212, + 380, + 787, + 421 + ], + "page_idx": 10 + }, + { + "type": "text", + "text": "iCrafter [87]. These methods take both an image and a text prompt as input for generation. The text prompt is generated by GPT-4V [56] to describe the future events. Another stream we compare is the image-based manipulation approach Motion Guidance [28]. This method uses an image and optical flow as inputs to predict future movement and generate corresponding images. We employ computed motion flow from our physical motion module as the target input for Motion Guidance to optimize each video frame.", + "bbox": [ + 212, + 465, + 787, + 571 + ], + "page_idx": 10 + }, + { + "type": "text", + "text": "4.1 Results", + "text_level": 1, + "bbox": [ + 214, + 597, + 323, + 611 + ], + "page_idx": 10 + }, + { + "type": "text", + "text": "We show our generated results on 8 different physical procedures from multiple sources in Fig. 4. Our method could simulate reasonable complex physical procedures which involves object-object interaction, object-scene interaction with different physical properties and initial conditions.", + "bbox": [ + 212, + 625, + 785, + 686 + ], + "page_idx": 10 + }, + { + "type": "text", + "text": "Visual comparisons. The comparison results are shown in Fig. 5. As can be seen, image-video generation methods could not produce any physically plausible futures. For each method, even we carefully tune the input prompt, the models couldn't understand the direction, physical meanings and interactions. All 3 models fail to synthesize realistic videos. For motion guidance with accurate motion flow provided, the results is better but brings a lot inconsistency. When there are multiple objects interaction, the method suffers from introducing new contents. For the domino case, the generation is not smooth and contain artifacts as it could not accurately model the physical procedure.", + "bbox": [ + 212, + 703, + 787, + 840 + ], + "page_idx": 10 + }, + { + "type": "header", + "text": "PhysGen: Rigid Physics-Grounded Img2Vid", + "bbox": [ + 437, + 114, + 730, + 128 + ], + "page_idx": 10 + }, + { + "type": "page_number", + "text": "11", + "bbox": [ + 767, + 116, + 782, + 126 + ], + "page_idx": 10 + }, + { + "type": "image", + "img_path": "images/86f9d79b7cd6780ddd6e334eefc6976736fa3d27fbfb61d66444fb372b94826a.jpg", + "image_caption": [ + "Fig.5: Qualitative comparison. against DynamiciCafter(DC) [87], I2VGenXL(VGen) [95], SEINE [18], Motion Guidance(MG) [28]." + ], + "image_footnote": [], + "bbox": [ + 225, + 142, + 781, + 824 + ], + "page_idx": 11 + }, + { + "type": "page_number", + "text": "12", + "bbox": [ + 217, + 114, + 235, + 126 + ], + "page_idx": 11 + }, + { + "type": "header", + "text": "S. Liu et al.", + "bbox": [ + 271, + 114, + 351, + 126 + ], + "page_idx": 11 + }, + { + "type": "image", + "img_path": "images/75290295add3e73aeb644eb061bdea4d415f0c229eaccfb873e5d6f42931b8e2.jpg", + "image_caption": [ + "Input" + ], + "image_footnote": [], + "bbox": [ + 230, + 157, + 316, + 224 + ], + "page_idx": 12 + }, + { + "type": "image", + "img_path": "images/0ada98210e6442fdb327fc985cd08265f97727fa9f72c81e59fc8a07354ba93f.jpg", + "image_caption": [ + "Fig. 6: Controllability. PhysGen generates diverse results reflecting initial forces." + ], + "image_footnote": [], + "bbox": [ + 230, + 224, + 316, + 294 + ], + "page_idx": 12 + }, + { + "type": "image", + "img_path": "images/bb0941cae9800451da6ab728f2ac94f5d0b17b8c65dd8a81f9a082ace7a692af.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 318, + 157, + 408, + 224 + ], + "page_idx": 12 + }, + { + "type": "image", + "img_path": "images/fece2730527167282bfab2ae3f6c4009f48e92632b0f3064b62b7e511991bda8.jpg", + "image_caption": [ + "FG Seg" + ], + "image_footnote": [], + "bbox": [ + 320, + 224, + 408, + 294 + ], + "page_idx": 12 + }, + { + "type": "image", + "img_path": "images/d5729ad87430b019747f7872a6a90ad394ee13475c6e17744e039b1c45073683.jpg", + "image_caption": [ + "Generation (left $\\rightarrow$ right: time steps)" + ], + "image_footnote": [], + "bbox": [ + 410, + 157, + 498, + 224 + ], + "page_idx": 12 + }, + { + "type": "image", + "img_path": "images/eecd168f4f0ae2d75c09dd900b64a0669d1ce990e352a2f3f7ac60bbe09e2984.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 410, + 224, + 498, + 294 + ], + "page_idx": 12 + }, + { + "type": "image", + "img_path": "images/350894b1c5b746a2627a025fba7843bb188198ae686cfe046087138af8db936d.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 501, + 157, + 588, + 224 + ], + "page_idx": 12 + }, + { + "type": "image", + "img_path": "images/76a57d3144cd17a6649d739b23543b32212590e7c1cba0ba8e0057184f9270d2.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 501, + 224, + 588, + 294 + ], + "page_idx": 12 + }, + { + "type": "image", + "img_path": "images/85d759259bf531df7be2066b72aba3fd3c2516399599fd455ff1a8912585c0c5.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 589, + 157, + 678, + 224 + ], + "page_idx": 12 + }, + { + "type": "image", + "img_path": "images/c7133f9ef0c8294c7ee20ac3488e9463d06447bc5630dd661ff692df556b28c2.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 589, + 224, + 678, + 294 + ], + "page_idx": 12 + }, + { + "type": "image", + "img_path": "images/bfa0631fb8f6a922ab56424f6a2056b93b29da966aa546d2fee99b1634ac0e03.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 681, + 157, + 767, + 224 + ], + "page_idx": 12 + }, + { + "type": "image", + "img_path": "images/7fe317890fe4aa7a590abb8a2ebb9997b74dfba92fd73df248cec690370f7926.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 681, + 224, + 767, + 294 + ], + "page_idx": 12 + }, + { + "type": "image", + "img_path": "images/97f771c717e7a716e883b0480a19f8c40ff0cd23b88dd83aab83eb23b47e53d0.jpg", + "image_caption": [ + "Input", + "Fig. 7: Qualitative results from the perception module. The results include foreground and background mask, inpainting output, normal map and inferred physics." + ], + "image_footnote": [], + "bbox": [ + 222, + 349, + 316, + 422 + ], + "page_idx": 12 + }, + { + "type": "image", + "img_path": "images/21ff867b20ac2c770e2e4a9292597255d91a9082e96d6e8c83c3f52b8ce4f8f7.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 344, + 375, + 370, + 393 + ], + "page_idx": 12 + }, + { + "type": "image", + "img_path": "images/9770c9b00fde4df2fe55bff534a1a120add4a9282721b8f57215f3b116d9fd60.jpg", + "image_caption": [ + "BG Seg" + ], + "image_footnote": [], + "bbox": [ + 413, + 349, + 506, + 422 + ], + "page_idx": 12 + }, + { + "type": "image", + "img_path": "images/36d369f2ffd32d1ed180a681022124479dad95114b84f666b395b79e3a4f1802.jpg", + "image_caption": [ + "Inpainting" + ], + "image_footnote": [], + "bbox": [ + 516, + 349, + 602, + 422 + ], + "page_idx": 12 + }, + { + "type": "image", + "img_path": "images/f8dce019dcc5a58c1eb711628fa0e9f7c0f9db3f935dd834c77cab1f759d1eb7.jpg", + "image_caption": [ + "Normal" + ], + "image_footnote": [], + "bbox": [ + 606, + 358, + 689, + 422 + ], + "page_idx": 12 + }, + { + "type": "image", + "img_path": "images/13f7e5e97e3cc966293330450a174d2c7a586dcc4baec7e322f96e6e1f7ddde2.jpg", + "image_caption": [ + "inferred Physics" + ], + "image_footnote": [], + "bbox": [ + 705, + 358, + 787, + 422 + ], + "page_idx": 12 + }, + { + "type": "text", + "text": "4.2 Human evaluation", + "text_level": 1, + "bbox": [ + 215, + 496, + 413, + 510 + ], + "page_idx": 12 + }, + { + "type": "text", + "text": "To comprehensively assess both the physical-realism and photo-realism of the generated videos, we conducted human evaluations following a similar methodology to prior work [16,78,83,91]. Here by physical-realism we mean whether the content of the generated video is realistic in its dynamics and object interactions and accurately reflects the instructed movements.", + "bbox": [ + 212, + 521, + 792, + 595 + ], + "page_idx": 12 + }, + { + "type": "text", + "text": "Specifically, we include 15 videos with diverse objects and conditions for evaluation, compared against other three I2V model generations (SEINE [18], DynamiCrafter [87], I2VGen-XL [95]). For each compared baseline method, we meticulously adjusted the input prompts and conducted multiple runs to select the most plausible output. In contrast, our method did not undergo specific tuning for each video. We randomized the presentation order and asked 14 participants to rate the videos on a five-point scale, from strongly disagree (1) to strongly agree (5) for physical-realism and photo-realism. The results in Tab. 1 show that our method achieves the highest ratings. Our average scores for physical-realism (4.14) and photo-realism (3.86) fall within the Agree level, while other methods perform poorly.", + "bbox": [ + 212, + 597, + 787, + 763 + ], + "page_idx": 12 + }, + { + "type": "text", + "text": "4.3 Quantitative evaluation", + "text_level": 1, + "bbox": [ + 214, + 785, + 455, + 800 + ], + "page_idx": 12 + }, + { + "type": "text", + "text": "We quantitatively evaluate the collected GT videos of a given scene. The GT data includes 833 images. For each method, we generate 50 videos with random", + "bbox": [ + 212, + 809, + 785, + 839 + ], + "page_idx": 12 + }, + { + "type": "header", + "text": "PhysGen: Rigid Physics-Grounded Img2Vid", + "bbox": [ + 437, + 114, + 730, + 128 + ], + "page_idx": 12 + }, + { + "type": "page_number", + "text": "13", + "bbox": [ + 767, + 114, + 784, + 126 + ], + "page_idx": 12 + }, + { + "type": "text", + "text": "samples, collecting 800 images per method. Our method varies initial conditions to ensure all generated videos are different. Following [12], we adopt the Fréchet Inception Distance (FID) [32] as evaluation metric. We didn't use Fréchet Video Distance (FVD) [73] due to the small sample size and the goal to disentangle evaluation of appearance and motion. We evaluate appearance using Image-FID and motion using Motion-FID. For Motion-FID, we use RAFT [71] to extract and colorize optical flow [5], then compute FID from these rendered images. The results are shown in Tab. 2. We notice other methods which either generate minimal motion, leading to low Image-FID and high Motion-FID (e.g., DynamiCrafter [87]), or produce realistic motion but fail to maintain image consistency, resulting in low Motion-FID and high Image-FID (e.g., SEINE [18]). Our approach achieves both low Image-FID and Motion-FID compared to others.", + "bbox": [ + 212, + 146, + 787, + 328 + ], + "page_idx": 13 + }, + { + "type": "text", + "text": "4.4 Analysis", + "text_level": 1, + "bbox": [ + 215, + 348, + 333, + 363 + ], + "page_idx": 13 + }, + { + "type": "text", + "text": "Perception evaluation. We evaluate the perception module using 10 complex open-world images with 118 annotated movable instances (COCO [47] format). Our system achieves 0.93 precision, 0.82 recall at 0.5 IoU. More details are shown in the supplementary. Fig. 7 shows intermediate outputs from our perception module and simulation module.", + "bbox": [ + 212, + 371, + 787, + 446 + ], + "page_idx": 13 + }, + { + "type": "text", + "text": "Physical reasoning. We compare the physical parameters estimated by GPT-4V to random values sampled from human-estimated ranges. Without GPT-4V, Image-FID increases from 105.70 to 111.01, Motion-FID from 30.20 to 36.60. It shows the physical reasoning are crucial for appearance and motion realism.", + "bbox": [ + 214, + 459, + 787, + 521 + ], + "page_idx": 13 + }, + { + "type": "text", + "text": "Controllability. We showcase the controllability and diversity of our method in Fig. 6 by varying the initial conditions. For different rows, we adjust the direction and magnitude of initial speed and forces. The results show diverse physical-plausible trajectories. This demonstrates the potential of applying our method for diverse physical world generations.", + "bbox": [ + 214, + 532, + 787, + 609 + ], + "page_idx": 13 + }, + { + "type": "text", + "text": "Limitations. Our method is an initial step towards physics-based video generation but has two main limitations: it focuses mainly on rigid objects, limiting its application to non-rigid ones, and it lacks a comprehensive 3D understanding, making it unable to handle out-of-plane motions. We leave leveraging deformable physics and full 3D understanding as future work.", + "bbox": [ + 214, + 622, + 787, + 699 + ], + "page_idx": 13 + }, + { + "type": "text", + "text": "5 Conclusion", + "text_level": 1, + "bbox": [ + 215, + 719, + 359, + 734 + ], + "page_idx": 13 + }, + { + "type": "text", + "text": "We present PhysGen, a novel image-to-video generation method that turns a static image into a high-fidelity, physics-grounded, and temporally coherent videos. The main module is a model-based physical simulator with a data-driven video generation process. PhysGen enables a series of controllable and interactive downstream applications. PhysGen provide a new image-to-video generation paradigm and hopefully can inspire more follow-up works.", + "bbox": [ + 212, + 750, + 787, + 840 + ], + "page_idx": 13 + }, + { + "type": "page_number", + "text": "14", + "bbox": [ + 217, + 114, + 235, + 126 + ], + "page_idx": 13 + }, + { + "type": "header", + "text": "S. Liu et al.", + "bbox": [ + 271, + 114, + 352, + 127 + ], + "page_idx": 13 + }, + { + "type": "text", + "text": "Acknowledgements", + "text_level": 1, + "bbox": [ + 217, + 143, + 401, + 162 + ], + "page_idx": 14 + }, + { + "type": "text", + "text": "This project is supported by NSF Awards #2331878, #2340254, and #2312102, the IBM IIDAI Grant, and an Intel Research Gift. We greatly appreciate the NCSA for providing computing resources. We thank Tianhang cheng for helpful discussions. We thank Emily Chen and Gloria Wang for proofreading.", + "bbox": [ + 215, + 176, + 785, + 237 + ], + "page_idx": 14 + }, + { + "type": "header", + "text": "PhysGen: Rigid Physics-Grounded Img2Vid", + "bbox": [ + 437, + 113, + 730, + 128 + ], + "page_idx": 14 + }, + { + "type": "page_number", + "text": "15", + "bbox": [ + 767, + 114, + 785, + 126 + ], + "page_idx": 14 + }, + { + "type": "text", + "text": "References", + "text_level": 1, + "bbox": [ + 217, + 143, + 321, + 159 + ], + "page_idx": 15 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "1. Aberman, K., Weng, Y., Lischinski, D., Cohen-Or, D., Chen, B.: Unpaired motion style transfer from video to animation. ACM TOG (2020)", + "2. Achiam, J., Adler, S., Agarwal, S., Ahmad, L., Akkaya, I., Aleman, F.L., Almeida, D., Altenschmidt, J., Altman, S., Anadkat, S., et al.: Gpt-4 technical report. arXiv preprint arXiv:2303.08774 (2023)", + "3. Ajay, A., Bauza, M., Wu, J., Fazeli, N., Tenenbaum, J.B., Rodriguez, A., Kaelbling, L.P.: Combining Physical Simulators and Object-Based Networks for Control. In: ICRA (2019)", + "4. Ajay, A., Wu, J., Fazeli, N., Bauza, M., Kaelbling, L.P., Tenenbaum, J.B., Rodriguez, A.: Augmenting Physical Simulators with Stochastic Neural Networks: Case Study of Planar Pushing and Bouncing. In: IROS (2018)", + "5. Baker, S., Scharstein, D., Lewis, J.P., Roth, S., Black, M.J., Szeliski, R.: A database and evaluation methodology for optical flow. IJCV (2011)", + "6. Bar-Tal, O., Chefer, H., Tov, O., Herrmann, C., Paiss, R., Zada, S., Ephrat, A., Hur, J., Li, Y., Michaeli, T., et al.: Lumiere: A space-time diffusion model for video generation. arXiv preprint arXiv:2401.12945 (2024)", + "7. Battaglia, P., Pascanu, R., Lai, M., Jimenez Rezende, D., et al.: Interaction networks for learning about objects, relations and physics. NeurIPS (2016)", + "8. Blattmann, A., Dockhorn, T., Kulal, S., Mendelevitch, D., Kilian, M., Lorenz, D., Levi, Y., English, Z., Voleti, V., Letts, A., et al.: Stable video diffusion: Scaling latent video diffusion models to large datasets. arXiv preprint arXiv:2311.15127 (2023)", + "9. Blattmann, A., Milbich, T., Dorkenwald, M., Ommer, B.: ipoke: Poking a still image for controlled stochastic video synthesis. In: ICCV (2021)", + "10. Blattmann, A., Milbich, T., Dorkenwald, M., Ommer, B.: Understanding object dynamics for interactive image-to-video synthesis. In: CVPR (2021)", + "1. Blattmann, A., Milbich, T., Dorkenwald, M., Ommer, B.: Understanding object dynamics for interactive image-to-video synthesis. In: CVPR (2021)", + "2. Blattmann, A., Rombach, R., Ling, H., Dockhorn, T., Kim, S.W., Fidler, S., Kreis, K.: Align your latents: High-resolution video synthesis with latent diffusion models. In: CVPR (2023)", + "3. Blomqvist, V.: Pymunk. https://pymunk.org (2023)", + "4. Bowen, R.S., Tucker, R., Zabih, R., Snavely, N.: Dimensions of motion: Monocular prediction through flow subspaces. In: 3DV (2022)", + "5. Careaga, C., Miangoleh, S.M.H., Aksoy, Y.: Intrinsic harmonization for illumination-aware compositing. arXiv preprint arXiv:2312.03698 (2023)", + "6. Chen, H., Xia, M., He, Y., Zhang, Y., Cun, X., Yang, S., Xing, J., Liu, Y., Chen, Q., Wang, X., et al.: Videocrafter1: Open diffusion models for high-quality video generation. arXiv preprint arXiv:2310.19512 (2023)", + "7. Chen, X., Liu, Z., Chen, M., Feng, Y., Liu, Y., Shen, Y., Zhao, H.: Livephoto: Real image animation with text-guided motion control. arXiv preprint arXiv:2312.02928 (2023)", + "8. Chen, X., Wang, Y., Zhang, L., Zhuang, S., Ma, X., Yu, J., Wang, Y., Lin, D., Qiao, Y., Liu, Z.: Seine: Short-to-long video diffusion model for generative transition and prediction. arXiv preprint arXiv:2310.20700 (2023)", + "9. Chuang, Y.Y., Goldman, D.B., Zheng, K.C., Curless, B., Salesin, D.H., Szeliski, R.: Animating pictures with stochastic motion textures. ACM TOG (2005)" + ], + "bbox": [ + 225, + 178, + 784, + 839 + ], + "page_idx": 15 + }, + { + "type": "page_number", + "text": "16", + "bbox": [ + 217, + 114, + 235, + 126 + ], + "page_idx": 15 + }, + { + "type": "header", + "text": "S. Liu et al.", + "bbox": [ + 271, + 114, + 351, + 126 + ], + "page_idx": 15 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "20. Ciarlet, P.G., Lions, J.L.: Handbook of numerical analysis. Gulf Professional Publishing (1990)", + "21. Davis, A., Bouman, K.L., Chen, J.G., Rubinstein, M., Durand, F., Freeman, W.T.: Visual vibrometry: Estimating material properties from small motion in video. In: CVPR (2015)", + "22. Davis, A., Chen, J.G., Durand, F.: Image-space modal bases for plausible manipulation of objects in video. ACM TOG (2015)", + "23. Eftekhar, A., Sax, A., Malik, J., Zamir, A.: Omnidata: A scalable pipeline for making multi-task mid-level vision datasets from 3d scans. In: ICCV (2021)", + "24. Endo, Y., Kanamori, Y., Kuriyama, S.: Animating landscape: self-supervised learning of decoupled motion and appearance for single-image video synthesis. arXiv preprint arXiv:1910.07192 (2019)", + "25. Friedland, B.: Control system design: an introduction to state-space methods. Courier Corporation (2012)", + "26. Fu, X., Yin, W., Hu, M., Wang, K., Ma, Y., Tan, P., Shen, S., Lin, D., Long, X.: Geowizard: Unleashing the diffusion priors for 3d geometry estimation from a single image. arXiv preprint arXiv:2403.12013 (2024)", + "27. Ge, S., Nah, S., Liu, G., Poon, T., Tao, A., Catanzaro, B., Jacobs, D., Huang, J.B., Liu, M.Y., Balaji, Y.: Preserve your own correlation: A noise prior for video diffusion models. In: ICCV (2023)", + "28. Geng, D., Owens, A.: Motion guidance: Diffusion-based image editing with differentiable motion estimators. In: ICLR (2024)", + "29. Girdhar, R., Singh, M., Brown, A., Duval, Q., Azadi, S., Rambhatla, S.S., Shah, A., Yin, X., Parikh, D., Misra, I.: Emu video: Factorizing text-to-video generation by explicit image conditioning. arXiv preprint arXiv:2311.10709 (2023)", + "30. Guo, Y., Yang, C., Rao, A., Liang, Z., Wang, Y., Qiao, Y., Agrawala, M., Lin, D., Dai, B.: Animatediff: Animate your personalized text-to-image diffusion models without specific tuning (2023)", + "31. Gupta, A., Yu, L., Sohn, K., Gu, X., Hahn, M., Fei-Fei, L., Essa, I., Jiang, L., Lezama, J.: Photorealistic video generation with diffusion models. arXiv preprint arXiv:2312.06662 (2023)", + "32. Heusel, M., Ramsauer, H., Unterthiner, T., Nessler, B., Hochreiter, S.: Gans trained by a two time-scale update rule converge to a local nash equilibrium. NeurIPS (2017)", + "33. Ho, J., Chan, W., Sahara, C., Whang, J., Gao, R., Gritsenko, A., Kingma, D.P., Poole, B., Norouzi, M., Fleet, D.J., et al.: Imagen video: High definition video generation with diffusion models. arXiv preprint arXiv:2210.02303 (2022)", + "34. Ho, J., Jain, A., Abbeel, P.: Denoising diffusion probabilistic models. NeurIPS (2020)", + "35. Holynski, A., Curless, B.L., Seitz, S.M., Szeliski, R.: Animating pictures with eulerian motion fields. In: CVPR (2021)", + "36. Hu, Y., Anderson, L., Li, T.M., Sun, Q., Carr, N., Ragan-Kelley, J., Durand, F.: Difftaichi: Differentiable programming for physical simulation. ICLR (2020)", + "37. Hu, Y., Fang, Y., Ge, Z., Qu, Z., Zhu, Y., Pradhana, A., Jiang, C.: A moving least squares material point method with displacement discontinuity and two-way rigid body coupling. ACM TOG (2018)", + "38. Hu, Y., Li, T.M., Anderson, L., Ragan-Kelley, J., Durand, F.: Taichi: a language for high-performance computation on spatially sparse data structures. ACM TOG (2019)", + "39. Jhou, W.C., Cheng, W.H.: Animating still landscape photographs through cloud motion creation. IEEE Transactions on Multimedia (2015)" + ], + "bbox": [ + 212, + 146, + 784, + 839 + ], + "page_idx": 16 + }, + { + "type": "header", + "text": "PhysGen: Rigid Physics-Grounded Img2Vid", + "bbox": [ + 437, + 114, + 730, + 128 + ], + "page_idx": 16 + }, + { + "type": "page_number", + "text": "17", + "bbox": [ + 767, + 116, + 784, + 126 + ], + "page_idx": 16 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "40. Kirillov, A., Mintun, E., Ravi, N., Mao, H., Rolland, C., Gustafson, L., Xiao, T., Whitehead, S., Berg, A.C., Lo, W.Y., Dollár, P., Girshick, R.: Segment anything. arXiv:2304.02643 (2023)", + "41. Koenig, N., Howard, A.: Design and use paradigms for gazebo, an open-source multi-robot simulator. In: IROS (2004)", + "42. Kondratyuk, D., Yu, L., Gu, X., Lezama, J., Huang, J., Hornung, R., Adam, H., Akbari, H., Alon, Y., Birodkar, V., et al.: Videopoet: A large language model for zero-shot video generation. arXiv preprint arXiv:2312.14125 (2023)", + "43. Li, X., Qiao, Y.L., Chen, P.Y., Jatavallabhula, K.M., Lin, M., Jiang, C., Gan, C.: Pac-nerf: Physics augmented continuum neural radiance fields for geometry-agnostic system identification. In: ICLR (2023)", + "44. Li, Y., Lin, T., Yi, K., Bear, D., Yamins, D.L., Wu, J., Tenenbaum, J.B., Torralba, A.: Visual grounding of learned physical models. In: ICML (2020)", + "45. Li, Y., Wu, J., Tedrake, R., Tenenbaum, J.B., Torralba, A.: Learning particle dynamics for manipulating rigid bodies, deformable objects, and fluids. In: ICLR (2019)", + "46. Li, Z., Tucker, R., Snavely, N., Holynski, A.: Generative image dynamics. arXiv preprint arXiv:2309.07906 (2023)", + "47. Lin, T.Y., Maire, M., Belongie, S., Hays, J., Perona, P., Ramanan, D., Dólar, P., Zitnick, C.L.: Microsoft coco: Common objects in context. In: ECCV (2014)", + "48. Liu, S., Zeng, Z., Ren, T., Li, F., Zhang, H., Yang, J., Li, C., Yang, J., Su, H., Zhu, J., et al.: Grounding dino: Marrying dino with grounded pre-training for open-set object detection. arXiv preprint arXiv:2303.05499 (2023)", + "49. Liu, T., Bargteil, A.W., O'Brien, J.F., Kavan, L.: Fast simulation of mass-spring systems. ACM TOG (2013)", + "50. Lv, J., Huang, Y., Yan, M., Huang, J., Liu, J., Liu, Y., Wen, Y., Chen, X., Chen, S.: Gpt4motion: Scripting physical motions in text-to-video generation via blender-oriented gpt planning. In: CVPRW. pp. 1430-1440 (2024)", + "51. Mahapatra, A., Kulkarni, K.: Controllable animation of fluid elements in still images. In: CVPR (2022)", + "52. Mallya, A., Wang, T.C., Liu, M.Y.: Implicit warping for animation with image sets. In: NeurIPS (2022)", + "53. Meng, C., He, Y., Song, Y., Song, J., Wu, J., Zhu, J.Y., Ermon, S.: SDEdit: Guided image synthesis and editing with stochastic differential equations. In: International Conference on Learning Representations (2022)", + "54. Mrowca, D., Zhuang, C., Wang, E., Haber, N., Fei-Fei, L.F., Tenenbaum, J., Yamins, D.L.: Flexible neural representation for physics prediction. In: NeurIPS (2018)", + "55. NVIDIA: Nvidia physx. https://developer.nvidia.com/physx-sdk (2019)", + "56. OpenAI: Gpt-4v(ision) system card (2023)", + "57. OpenAI: Creating video from text. https://openai.com/sora (2024)", + "58. Peebles, W., Xie, S.: Scalable diffusion models with transformers. arXiv preprint arXiv:2212.09748 (2022)", + "59. Popova, E., Popov, V.L.: The research works of coulomb and amontons and generalized laws of friction. Friction (2015)", + "60. Qiu, H., Chen, Z., Jiang, Y., Zhou, H., Fan, X., Yang, L., Wu, W., Liu, Z.: Relitalk: Relightable talking portrait generation from a single video (2023)", + "61. Ren, T., Liu, S., Zeng, A., Lin, J., Li, K., Cao, H., Chen, J., Huang, X., Chen, Y., Yan, F., Zeng, Z., Zhang, H., Li, F., Yang, J., Li, H., Jiang, Q., Zhang, L.: Grounded sam: Assembling open-world models for diverse visual tasks (2024)" + ], + "bbox": [ + 215, + 146, + 784, + 840 + ], + "page_idx": 17 + }, + { + "type": "page_number", + "text": "18", + "bbox": [ + 217, + 114, + 235, + 126 + ], + "page_idx": 17 + }, + { + "type": "header", + "text": "S. Liu et al.", + "bbox": [ + 271, + 114, + 351, + 126 + ], + "page_idx": 17 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "62. Rombach, R., Blattmann, A., Lorenz, D., Esser, P., Ommer, B.: High-resolution image synthesis with latent diffusion models. In: CVPR (2022)", + "63. Sanchez-Gonzalez, A., Heess, N., Springenberg, J.T., Merel, J., Riedmiller, M., Hadsell, R., Battaglia, P.: Graph networks as learnable physics engines for inference and control. In: ICML (2018)", + "64. Schödl, A., Szeliski, R., Salesin, D.H., Essa, I.: Video textures. In: PACMCGIT (2000)", + "65. Singer, U., Polyak, A., Hayes, T., Yin, X., An, J., Zhang, S., Hu, Q., Yang, H., Ashual, O., Gafni, O., et al.: Make-a-video: Text-to-video generation without text-video data. arXiv preprint arXiv:2209.14792 (2022)", + "66. Sohl-Dickstein, J., Weiss, E., Maheswaranathan, N., Ganguli, S.: Deep unsupervised learning using nonequilibrium thermodynamics. In: ICML (2015)", + "67. Song, J., Meng, C., Ermon, S.: Denoising diffusion implicit models. arXiv preprint arXiv:2010.02502 (2020)", + "68. Song, Y., Ermon, S.: Generative modeling by estimating gradients of the data distribution. NeurIPS (2019)", + "69. Sugimoto, R., He, M., Liao, J., Sander, P.V.: Water simulation and rendering from a still photograph. In: SIGGRAPH Asia (2022)", + "70. Szummer, M., Picard, R.W.: Temporal texture modeling. In: ICIP (1996)", + "71. Teed, Z., Deng, J.: Raft: Recurrent all-pairs field transforms for optical flow. In: ECCV. pp. 402-419 (2020)", + "72. Todorov, E., Erez, T., Tassa, Y.: Mujoco: A physics engine for model-based control. In: IROS (2012)", + "73. Unterthiner, T., Van Steenkiste, S., Kurach, K., Marinier, R., Michalski, M., Gelly, S.: Towards accurate generative models of video: A new metric & challenges. arXiv preprint arXiv:1812.01717 (2018)", + "74. Villegas, R., Babaeizadeh, M., Kindermans, P.J., Moraldo, H., Zhang, H., Saffar, M.T., Castro, S., Kunze, J., Erhan, D.: Phenaki: Variable length video generation from open domain textual description. arXiv preprint arXiv:2210.02399 (2022)", + "75. Wang, X., Yuan, H., Zhang, S., Chen, D., Wang, J., Zhang, Y., Shen, Y., Zhao, D., Zhou, J.: Videocomposer: Compositional video synthesis with motion controllability. arXiv preprint arXiv:2306.02018 (2023)", + "76. Wang, Y., Chen, X., Ma, X., Zhou, S., Huang, Z., Wang, Y., Yang, C., He, Y., Yu, J., Yang, P., et al.: Lavie: High-quality video generation with cascaded latent diffusion models. arXiv preprint arXiv:2309.15103 (2023)", + "77. Wei, L.Y., Levoy, M.: Fast texture synthesis using tree-structured vector quantization. In: PACMCGIT (2000)", + "78. Wei, Y., Zhang, S., Qing, Z., Yuan, H., Liu, Z., Liu, Y., Zhang, Y., Zhou, J., Shan, H.: Dreamvideo: Composing your dream videos with customized subject and motion. arXiv preprint arXiv:2312.04433 (2023)", + "79. Weng, C.Y., Curless, B., Kemelmacher-Shlizerman, I.: Photo wake-up: 3d character animation from a single photo. In: CVPR (2019)", + "80. Wu, J., Lim, J.J., Zhang, H., Tenenbaum, J.B., Freeman, W.T.: Physics 101: Learning physical object properties from unlabeled videos. In: BMVC (2016)", + "81. Wu, J., Lu, E., Kohli, P., Freeman, W.T., Tenenbaum, J.B.: Learning to see physics via visual de-animation. In: NeurIPS (2017)", + "82. Wu, J., Yildirim, I., Lim, J.J., Freeman, W.T., Tenenbaum, J.B.: Galileo: Perceiving physical object properties by integrating a physics engine with deep learning. In: NeurIPS (2015)", + "83. Wu, R., Chen, L., Yang, T., Guo, C., Li, C., Zhang, X.: Lamp: Learn a motion pattern for few-shot-based video generation. arXiv preprint arXiv:2310.10769 (2023)" + ], + "bbox": [ + 212, + 146, + 787, + 840 + ], + "page_idx": 18 + }, + { + "type": "header", + "text": "PhysGen: Rigid Physics-Grounded Img2Vid", + "bbox": [ + 437, + 114, + 730, + 128 + ], + "page_idx": 18 + }, + { + "type": "page_number", + "text": "19", + "bbox": [ + 767, + 116, + 785, + 126 + ], + "page_idx": 18 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "84. Xia, H., Lin, Z.H., Ma, W.C., Wang, S.: Video2game: Real-time, interactive, realistic and browser-compatible environment from a single video. In: CVPR (2024)", + "85. Xia, H., Lin, Z.H., Ma, W.C., Wang, S.: Video2game: Real-time, interactive, realistic and browser-compatible environment from a single video. In: CVPR (2024)", + "86. Xie, T., Zong, Z., Qiu, Y., Li, X., Feng, Y., Yang, Y., Jiang, C.: Physgaussian: Physics-integrated 3d gaussians for generative dynamics. In: CVPR (2024)", + "87. Xing, J., Xia, M., Zhang, Y., Chen, H., Yu, W., Liu, H., Wang, X., Wong, T.T., Shan, Y.: Dynamiccafter: Animating open-domain images with video diffusion priors. arXiv preprint arXiv:2310.12190 (2023)", + "88. Xu, Z., Wu, J., Zeng, A., Tenenbaum, J.B., Song, S.: Densephysnet: Learning dense physical object representations via multi-step dynamic interactions. In: RSS (2019)", + "89. Xue, T., Wu, J., Bouman, K.L., Freeman, W.T.: Visual dynamics: Stochastic future generation via layered cross convolutional networks. T-PAMI (2018)", + "90. Yang, J., Zhang, H., Li, F., Zou, X., Li, C., Gao, J.: Set-of-mark prompting unleashes extraordinary visual grounding in gpt-4v. arXiv preprint arXiv:2310.11441 (2023)", + "91. Yu, J., Cun, X., Qi, C., Zhang, Y., Wang, X., Shan, Y., Zhang, J.: Animatezero: Video diffusion models are zero-shot image animators. arXiv preprint arXiv:2312.03793 (2023)", + "92. Yu, L., Lezama, J., Gundavarapu, N.B., Versari, L., Sohn, K., Minnen, D., Cheng, Y., Gupta, A., Gu, X., Hauptmann, A.G., et al.: Language model beats diffusion-tokenizer is key to visual generation. arXiv preprint arXiv:2310.05737 (2023)", + "93. Yu, T., Feng, R., Feng, R., Liu, J., Jin, X., Zeng, W., Chen, Z.: Inpaint anything: Segment anything meets image inpainting. arXiv preprint arXiv:2304.06790 (2023)", + "94. Zhai, A.J., Shen, Y., Chen, E., Wang, G., Wang, X., Wang, S., Wang, X., Guan, K., Wang, S.: Physical property understanding from language-embedded feature fields. In: CVPR (2024)", + "95. Zhang, S., Wang, J., Zhang, Y., Zhao, K., Yuan, H., Qing, Z., Wang, X., Zhao, D., Zhou, J.: I2vgen-xl: High-quality image-to-video synthesis via cascaded diffusion models. arXiv preprint arXiv:2311.04145 (2023)", + "96. Zhao, J., Zhang, H.: Thin-plate spline motion model for image animation. In: CVPR (2022)" + ], + "bbox": [ + 215, + 146, + 784, + 590 + ], + "page_idx": 19 + }, + { + "type": "page_number", + "text": "20", + "bbox": [ + 217, + 114, + 235, + 126 + ], + "page_idx": 19 + }, + { + "type": "header", + "text": "S. Liu et al.", + "bbox": [ + 271, + 114, + 351, + 126 + ], + "page_idx": 19 + } +] \ No newline at end of file diff --git a/2024/PhysGen_ Rigid-Body Physics-Grounded Image-to-Video Generation/63954133-cd49-423f-97fe-3962651db98c_model.json b/2024/PhysGen_ Rigid-Body Physics-Grounded Image-to-Video Generation/63954133-cd49-423f-97fe-3962651db98c_model.json new file mode 100644 index 0000000000000000000000000000000000000000..8c453f2f0fdf664a5cdf82922b6cba411769bd9a --- /dev/null +++ b/2024/PhysGen_ Rigid-Body Physics-Grounded Image-to-Video Generation/63954133-cd49-423f-97fe-3962651db98c_model.json @@ -0,0 +1,2957 @@ +[ + [ + { + "type": "title", + "bbox": [ + 0.27, + 0.141, + 0.734, + 0.187 + ], + "angle": 0, + "content": "PhysGen: Rigid-Body Physics-Grounded Image-to-Video Generation" + }, + { + "type": "text", + "bbox": [ + 0.247, + 0.213, + 0.756, + 0.229 + ], + "angle": 0, + "content": "Shaowei Liu, Zhongzheng Ren, Saurabh Gupta*, and Shenlong Wang*" + }, + { + "type": "text", + "bbox": [ + 0.344, + 0.24, + 0.658, + 0.271 + ], + "angle": 0, + "content": "University of Illinois Urbana-Champaign https://stevenlsw.github.io/physgen/" + }, + { + "type": "text", + "bbox": [ + 0.262, + 0.303, + 0.761, + 0.567 + ], + "angle": 0, + "content": "Abstract. We present PhysGen, a novel image-to-video generation method that converts a single image and an input condition (e.g., force and torque applied to an object in the image) to produce a realistic, physically plausible, and temporally consistent video. Our key insight is to integrate model-based physical simulation with a data-driven video generation process, enabling plausible image-space dynamics. At the heart of our system are three core components: (i) an image understanding module that effectively captures the geometry, materials, and physical parameters of the image; (ii) an image-space dynamics simulation model that utilizes rigid-body physics and inferred parameters to simulate realistic behaviors; and (iii) an image-based rendering and refinement module that leverages generative video diffusion to produce realistic video footage featuring the simulated motion. The resulting videos are realistic in both physics and appearance and are even precisely controllable, showcasing superior results over existing data-driven image-to-video generation works through quantitative comparison and comprehensive user study. PhysGen's resulting videos can be used for various downstream applications, such as turning an image into a realistic animation or allowing users to interact with the image and create various dynamics." + }, + { + "type": "title", + "bbox": [ + 0.216, + 0.591, + 0.377, + 0.607 + ], + "angle": 0, + "content": "1 Introduction" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.621, + 0.788, + 0.741 + ], + "angle": 0, + "content": "Looking at the images in Fig. 1, we can effortlessly predict or visualize the potential outcomes of various physical effects applied to the car, the curling stone, and the stack of dominos. This understanding of physics empowers our imagination of the counterfactual: we can mentally imagine various consequences of an action from an image without experiencing them. We aim to provide computers with similar capabilities - understanding and simulating physics from a single image and creating realistic animations. We expect the resulting video to produce realistic animations with physically plausible dynamics and interactions." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.742, + 0.788, + 0.817 + ], + "angle": 0, + "content": "Tremendous progress has been made in generating realistic and physically plausible video footage. Conventional graphics leverage model-based dynamics to simulate plausible motions and utilize a graphics renderer to simulate images. Such methods provide realistic and controllable dynamics. However, both the dynamics of physics and lighting physics are predefined, making it hard to achieve" + }, + { + "type": "page_footnote", + "bbox": [ + 0.232, + 0.826, + 0.348, + 0.841 + ], + "angle": 0, + "content": "* Equal advising" + } + ], + [ + { + "type": "page_number", + "bbox": [ + 0.218, + 0.116, + 0.23, + 0.127 + ], + "angle": 0, + "content": "2" + }, + { + "type": "header", + "bbox": [ + 0.272, + 0.115, + 0.353, + 0.127 + ], + "angle": 0, + "content": "S. Liu et al." + }, + { + "type": "image_caption", + "bbox": [ + 0.255, + 0.146, + 0.294, + 0.157 + ], + "angle": 0, + "content": "Input" + }, + { + "type": "image", + "bbox": [ + 0.229, + 0.158, + 0.782, + 0.373 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.214, + 0.375, + 0.788, + 0.416 + ], + "angle": 0, + "content": "Fig. 1: Given a single image, training-free PhysGen generates future frames controlled by physics and initial state. Semantics, geometry and dynamics are reasoned during generation and the result videos are both physics-grounded and photo-realistic." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.461, + 0.788, + 0.612 + ], + "angle": 0, + "content": "our goal of animating an image captured in the real world. On the other hand, data-driven image-to-video (I2V) generation techniques [6, 8, 57] learn to create realistic videos from a single image through training diffusion models over internet-scale data. However, despite advancements in video generative models [6, 8, 57], the incorporation of real-world physics principles into the video generation process remains largely unexplored and unsolved. Consequently, the synthesized videos often lack temporal coherence and fail to replicate authentic object motions observed in reality. Furthermore, such text-driven methods also lack fine-grained controllability. For instance, they cannot simulate the consequences of different forces and torques applied to an object." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.613, + 0.788, + 0.749 + ], + "angle": 0, + "content": "In light of this gap, we propose a paradigm shift in video generation through the introduction of model-based video generative models. In contrast to existing purely generative methods, which are trained in a data-driven manner and rely on diffusion models to learn image space dynamics via scaling laws, we propose grounding object dynamics explicitly using rigid body physics, thereby integrating fundamental physical principles into the generation process. As classical mechanics theory reveals, the motion of an object is determined by its physical properties (e.g., mass, elasticity, surface roughness) and external factors (e.g., external forces, environmental conditions, boundary conditions)." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.75, + 0.788, + 0.841 + ], + "angle": 0, + "content": "We tackle our image-to-video generation problem via a computational framework, PhysGen. PhysGen consists of three stages: 1) image-based physics understanding; 2) model-based dynamics simulation; and 3) generative video rendering with dynamics guidance. Our method first infers object compositions and physical parameters from the input image through large visual foundation model-based reasoning (Sec. 3.1). Given an input force, we then utilize the in" + } + ], + [ + { + "type": "header", + "bbox": [ + 0.439, + 0.115, + 0.732, + 0.129 + ], + "angle": 0, + "content": "PhysGen: Rigid Physics-Grounded Img2Vid" + }, + { + "type": "page_number", + "bbox": [ + 0.775, + 0.117, + 0.785, + 0.127 + ], + "angle": 0, + "content": "3" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.146, + 0.788, + 0.328 + ], + "angle": 0, + "content": "ferred physical parameters to perform realistic dynamic simulations for each object, accounting for their rigid body dynamics, collisions, frictions, and elasticity (Sec. 3.2). Finally, guided by motion and the input image, we present a novel generative video diffusion-based renderer that outputs the final realistic video footage (Sec. 3.3). Several illustrative examples of our framework are shown in Fig. 1. As shown in this figure, our approach combines the best of the worlds of data-driven generative modeling's appearance realism and model-based dynamics's verified physical plausibility. Notably, our generation process is highly controllable, allowing users to specify underlying physics parameters and initial conditions. This capability facilitates a range of interactive applications. Moreover, our proposed generation pipeline operates solely during inference time, eliminating the need for any training." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.33, + 0.788, + 0.555 + ], + "angle": 0, + "content": "We evaluate PhysGen generation ability on multiple data source including the web and self-captured images. We compare against both state-of-the-art image-to-video models [18, 87, 95] as well as image editing method [28] quantitatively and qualitatively. We also perform user-study to evaluate the physical-realism and photo-realism of the generated video. Our method demonstrates physics-informed results for controllable image-to-video generation, producing realistic video sequences with grounded rigid-body dynamics and interaction. PhysGen combines learning-based generative approaches with traditional model-based physics to deliver visually appealing results without any training. Our approach harnesses a plethora of ingredients from previous works, particularly recent progress in large foundational visual models for segmentation [40, 48, 61, 93], physical understanding [2, 56], normal estimation [23, 26], relighting, and generative rendering [8, 18]. While these individual ingredients exist, our system combines them in a novel manner to enable an exciting new capacity, featuring physical plausibility for video generation at an unprecedented level." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.556, + 0.788, + 0.663 + ], + "angle": 0, + "content": "Our contributions can be summarized as follows: 1) We propose a novel image-space dynamics model that can understand physical parameters from a single image and simulate realistic, controllable, and interactive dynamics. 2) We present a novel image-to-video system by integrating our presented dynamics model with generative diffusion-based video priors. The model is training-free, and the resulting videos are highly realistic, physically plausible, and controllable, consistently surpassing existing state-of-the-art video generative models." + }, + { + "type": "title", + "bbox": [ + 0.215, + 0.686, + 0.384, + 0.701 + ], + "angle": 0, + "content": "2 Related work" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.72, + 0.788, + 0.842 + ], + "angle": 0, + "content": "Image-based physics reasoning and simulation. Our image-based dynamics model is built upon physical simulation. Symbolic physics simulators or physics engines have been extensively studied for various physical processes, including rigid body dynamics, fluid simulation, and deformable objects [36-38, 41, 49, 55, 72]. Such methods have been widely used across graphics, scientific discovery, robotics, and video effects. However, these rule-based or solver-based simulators face limitations in terms of expressiveness, efficiency, generalizability, and parameter tuning. To overcome these challenges, neural physics, which combines" + } + ], + [ + { + "type": "page_number", + "bbox": [ + 0.218, + 0.116, + 0.23, + 0.127 + ], + "angle": 0, + "content": "4" + }, + { + "type": "header", + "bbox": [ + 0.272, + 0.115, + 0.353, + 0.127 + ], + "angle": 0, + "content": "S. Liu et al." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.147, + 0.788, + 0.268 + ], + "angle": 0, + "content": "traditional physics simulators/engines with deep neural networks, has been developed to model dynamic processes such as planar pushing and bouncing [4], as well as object interactions [3]. Domain-specific network architectures [7,45,54,63] have been developed to address various physical problems. Additionally, the integration of neural scene representations or large language models with physics simulators [43,50,86] has also been investigated recently. Nevertheless, most forward physics models, whether neural or not, have not resolved the dependency on pre-defined physical parameters, which are often specified by users." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.269, + 0.788, + 0.48 + ], + "angle": 0, + "content": "Reasoning physical parameters from visual data allows us to simulate without manually defined physics, further enhancing convenience and realism for image-based physics simulation. To enable this, subsequent research efforts [21, 43, 44, 80-82, 84, 88] have introduced pipelines that first extract scene representations of the physical world from images or videos using neural networks, and then utilize these representations for either physics reasoning or simulation. Despite the promise of these approaches, they primarily rely on synthetic data for training, which might suffer from generalization issues, or on inverse physics, which requires inference from observed physical phenomena, often requiring video observations instead of a single image. In our work, we aim to generate physics-grounded, high-fidelity predictive videos from static input images of real-world objects with complex backgrounds. Contrary to existing work, we investigate leveraging the power of large pretrained visual foundational models [2, 40, 56] for physics reasoning in a zero-shot manner." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.494, + 0.788, + 0.676 + ], + "angle": 0, + "content": "Video generative model. Video generative models have experienced remarkable advancements in recent years [6,8,12,18,27,29,31,33,42,57,65,74,76,87,92, 95], with the state-of-the-art framework [57] achieving the capability to generate photo-realistic and coherent imaginative videos from text instructions using diffusion models [34,58,62,66,68]. Despite progress, challenges remain, such as the need for extensive training data, object permanence, realistic lighting, and reducing unrealistic motion and physics violations. Overcoming these is key for advancing video generative modeling and its applications, especially in physics-accurate areas like scientific discovery and robotics. Our work addresses some issues by proposing a novel image-to-video framework that merges a model-based approach with a pretrained video prior, grounding the generated content with real-world physics." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.689, + 0.788, + 0.843 + ], + "angle": 0, + "content": "Image animation. Our method is heavily inspired by image-based animation [19, 22, 35, 64, 70, 77, 89] and cinematographs. One common way to improve quality is to incorporate class-specific heuristics [19, 39], which can be physically simulated and integrated into the generation process. Recent advancements in deep learning have led to data-driven solutions in this domain, where temporal neural networks are trained on video datasets to directly predict consecutive video frames from an input image [10, 17, 24, 30, 35, 75]. Moreover, there is growing interest in interactive [9, 11, 46] and controllable [1, 22] image-to-video synthesis. To further improve temporal coherence, more priors such as motion fields [24, 35, 51, 52, 69, 89] or flows [14, 28, 96], 3D geometry information [60, 79]," + } + ], + [ + { + "type": "header", + "bbox": [ + 0.439, + 0.115, + 0.732, + 0.129 + ], + "angle": 0, + "content": "PhysGen: Rigid Physics-Grounded Img2Vid" + }, + { + "type": "page_number", + "bbox": [ + 0.775, + 0.117, + 0.785, + 0.127 + ], + "angle": 0, + "content": "5" + }, + { + "type": "image", + "bbox": [ + 0.245, + 0.144, + 0.761, + 0.295 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.214, + 0.296, + 0.788, + 0.394 + ], + "angle": 0, + "content": "Fig. 2: Method overview. Our framework consists of three interleaved components: the perception module, the dynamics simulation module, and the rendering module. The perception module interprets the semantics, geometry, and physical parameters in the given image. The dynamics simulation module simulates the rigid-body motion and interactions of each instance in the scene, governed by Newton's Laws and physical constraints. The rendering module renders the final outcome, leveraging an off-the-shelf relighting model and diffusion-model-based video priors." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.431, + 0.788, + 0.522 + ], + "angle": 0, + "content": "and user annotations [46] have been subsequently introduced. In this work, we propose a model-based solution and tackle the image animation task from a new perspective, where our method takes a single image and user-specific initial force or torque as input, and models object dynamics explicitly via rigid physics simulation. Our key insight is that physical parameters can be inferred from a single image automatically through large foundation models such as GPT-4V [2, 56]." + }, + { + "type": "title", + "bbox": [ + 0.216, + 0.564, + 0.347, + 0.58 + ], + "angle": 0, + "content": "3 Approach" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.614, + 0.789, + 0.841 + ], + "angle": 0, + "content": "Given a single input image, our goal is to generate a realistic \\(T\\)-frame video featuring rigid body dynamics and interactions. Our generation can be conditioned on user-specific inputs in the form of an initial force \\(\\mathbf{F}_0\\) and/or torque \\(\\pmb{\\tau}_{0}\\). Our key insight is to incorporate physics-informed simulation into the generation process, ensuring the dynamics are realistic and interactive, thus producing large, plausible object motion. To achieve this, we present a novel framework consisting of three stages: physical understanding, dynamics generation, and generative rendering. The perception module aims to reason about the semantics, collision geometry, and physics of the objects presented in the image (Sec. 3.1). The dynamics module then leverages the input force/torque, as well as the inferred shape and geometry, to simulate the rigid-body motions of the objects, as well as object-object and object-scene interactions, such as friction and collisions (Sec. 3.2). Finally, we convert the simulated dynamics into pixel-level motion and combine image-based warping and generative video to render the final video (Sec. 3.3). Fig. 2 depicts the overall framework." + } + ], + [ + { + "type": "page_number", + "bbox": [ + 0.218, + 0.116, + 0.23, + 0.127 + ], + "angle": 0, + "content": "6" + }, + { + "type": "header", + "bbox": [ + 0.272, + 0.115, + 0.353, + 0.128 + ], + "angle": 0, + "content": "S. Liu et al." + }, + { + "type": "title", + "bbox": [ + 0.217, + 0.147, + 0.353, + 0.162 + ], + "angle": 0, + "content": "3.1 Perception" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.173, + 0.787, + 0.249 + ], + "angle": 0, + "content": "Simulating dynamics on an image requires a holistic understanding of object compositions, materials, geometry, and physics. Toward this goal, we designed our perception module to infer such properties from a single image. Our key insight is to harness readily available, large pretrained models [2, 15, 26, 56, 61] to achieve this goal, as shown in Fig. 2." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.265, + 0.788, + 0.416 + ], + "angle": 0, + "content": "Segmentation. Identifying and segmenting each individual physical entity lays the foundation for image-based dynamics. Inspired by the recent success of large pretrained segmentation models, we incorporate GPT-4V [2, 56] to recognize all image categories and send to Grounded-SAM [61] to detect and segment each individual instance-level objects \\(\\{\\mathbf{o}^i\\in \\mathbb{R}^{W\\times H}\\}_{i}^{N}\\), where \\(\\mathbf{o}^i\\) is the binary mask for \\(i\\)-th object. We also query the GPT-4V to classify the objects into foreground and background based on their movability. The foreground objects are send to the physical simulation once user inputs are applied. We extract collision boundaries and supporting edges (e.g., ground, walls, etc.) from the background objects. The details are presented in the supplementary." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.432, + 0.788, + 0.628 + ], + "angle": 0, + "content": "Physical properties reasoning. Realistic physics depends on accurate physical parameters, such as surface friction, mass, and elasticity. Unlike prior physics-based image dynamics work [19], our work leverages visual foundational models to reason physical properties directly. Our solution is simple yet effective. Inspired by the success of mask-based prompting [2,56,90], we directly ask GPT-4V [56] for certain physical properties, providing an object mask overlaid on the input image. Following [85,94], we send the crafted prompt to GPT-4V, which contains GPT-4V instructions to return a quantitative measure of each queried property in a metric unit. For each object, we query its mass, elasticity, and friction coefficient \\((M_i,E_i,\\mu_i)\\). We use the Coulomb friction model [59] for friction modeling. The elasticity coefficient is a scalar, where 0.0 results in no bounce, and 1.0 results in a perfect bounce. Given each object mass \\(M_{i}\\) in grams and its shape primitive, we compute the rotational inertia \\(\\mathbf{I}_i\\) accordingly." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.643, + 0.788, + 0.765 + ], + "angle": 0, + "content": "Geometry primitives. Rasterized instance masks are not suitable for physical simulation; hence, we need to convert each object into vectorized shape primitives. We fit two types of primitive shapes: a circle and generic polygons. For each object, we choose the primitive type with the maximum coverage of its segmentation. For circles, we fit the center and radius to cover the segmentation mask. Circles allow us to realistically simulate rolling motions. For non-circle objects, we instead fit the generic polygons. Specifically, we perform contour extraction on the segmentation masks to obtain the polygon vertices." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.78, + 0.788, + 0.842 + ], + "angle": 0, + "content": "Intrinsic decomposition. The dynamic movement of the objects will also result in comprehensive changes in shading, as their position wrt the lighting environment changes. To compensate for this effect during rendering, we must perform image decomposition to infer albedo, normal as well as lighting from" + } + ], + [ + { + "type": "header", + "bbox": [ + 0.438, + 0.114, + 0.732, + 0.129 + ], + "angle": 0, + "content": "PhysGen: Rigid Physics-Grounded Img2Vid" + }, + { + "type": "page_number", + "bbox": [ + 0.775, + 0.116, + 0.785, + 0.127 + ], + "angle": 0, + "content": "7" + }, + { + "type": "text", + "bbox": [ + 0.215, + 0.146, + 0.784, + 0.194 + ], + "angle": 0, + "content": "the single image for each object \\((\\mathbf{A}_0^i,\\mathbf{N}_0^i,\\mathbf{L}_0)\\). To achieve this we leverage off-the-shelf intrinsic decomposition model [15] to compute \\(\\mathbf{A}_0^i\\) and \\(\\mathbf{L}_0\\) and exploits pretrained surface normal estimator for the computation of object normal \\(\\mathbf{N}_0^i\\)." + }, + { + "type": "text", + "bbox": [ + 0.215, + 0.209, + 0.784, + 0.285 + ], + "angle": 0, + "content": "Background inpainting. Once the foreground objects move, they will leave holes in the background. To produce realistic and complete video footage, we leverage an off-the-shelf generative image inpainting model [93] to recover the complete background scene without foreground objects. The inpainted image \\(\\mathbf{B} \\in \\mathbb{R}^{W \\times H \\times 3}\\) will be used for composition in the rendering stage." + }, + { + "type": "title", + "bbox": [ + 0.215, + 0.31, + 0.536, + 0.325 + ], + "angle": 0, + "content": "3.2 Image space dynamics simulation" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.337, + 0.785, + 0.519 + ], + "angle": 0, + "content": "Given the foreground objects with physical properties, we use rigid-body physics for dynamics simulation in image spaces. We choose to perform simulations in image spaces for three reasons: 1) image-space dynamics are better coupled with our output video; 2) a full 3D simulation requires complete 3D scene reconstruction and understanding, which remains an unsolved problem from a single image; 3) image-space dynamics can already cover various object dynamics and have been widely used in prior work [4, 19, 35, 46, 80]. Specifically, at a given time \\( t \\), each rigid object \\( i \\) is characterized by its 2D pose and velocity at its center of mass. The position includes a translation \\( \\mathbf{t}^i(t) \\in \\mathbb{R}^2 \\) and a rotation \\( \\mathbf{R}^i(t) \\in \\mathbb{S}\\mathbb{O}(2) \\) specified in a world coordinate. The velocity includes linear velocity \\( \\pmb{\\nu}^i(t) \\in \\mathbb{R}^2 \\) and angular velocity \\( \\pmb{\\omega}^i(t) \\in \\mathbb{R}^2 \\). Hence, the state of the each object \\( i \\) at time \\( t \\) can be represented as \\( \\mathbf{q}^i(t) = [\\mathbf{t}^i(t), \\mathbf{R}^i(t), \\pmb{\\nu}^i(t), \\pmb{\\omega}^i(t)] \\)." + }, + { + "type": "text", + "bbox": [ + 0.215, + 0.519, + 0.784, + 0.551 + ], + "angle": 0, + "content": "Following [25], the rigid body motion dynamics is given by Eq. (1) for each object. For simplicity, we omit the object index \\(i\\) in the following:" + }, + { + "type": "equation", + "bbox": [ + 0.308, + 0.561, + 0.785, + 0.627 + ], + "angle": 0, + "content": "\\[\n\\frac {d}{d t} \\mathbf {q} (t) = \\frac {d}{d t} \\left[ \\begin{array}{l} \\mathbf {t} (t) \\\\ \\mathbf {R} (t) \\\\ \\boldsymbol {\\nu} (t) \\\\ \\boldsymbol {\\omega} (t) \\end{array} \\right] = \\left[ \\begin{array}{c} \\mathbf {v} (t) \\\\ \\boldsymbol {\\omega} (t) \\times \\mathbf {R} (t) \\\\ \\frac {\\mathbf {F} (t)}{M} \\\\ \\mathbf {I} (t) ^ {- 1} (\\boldsymbol {\\tau} - \\boldsymbol {\\omega} (t) \\times \\mathbf {I} (t) \\boldsymbol {\\omega} (t)) \\end{array} \\right] \\tag {1}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.215, + 0.638, + 0.784, + 0.7 + ], + "angle": 0, + "content": "where \\(\\mathbf{F}\\) is the force, \\(\\pmb{\\tau}\\) is the torque, \\(M\\) is the mass of the object, \\(\\mathbf{I}(t) = \\mathbf{R}(t)\\mathbf{I}\\mathbf{R}(t)^T\\) is the rotation inertia in world coordinates, \\(\\mathbf{I} \\in \\mathbb{R}^{2 \\times 2}\\) is the inertia matrix in its body coordinates. This inertia indicates its resistance against rotational motion. The state \\(\\mathbf{q}(t)\\) is given by integrating Eq. (1):" + }, + { + "type": "equation", + "bbox": [ + 0.307, + 0.712, + 0.785, + 0.753 + ], + "angle": 0, + "content": "\\[\n\\mathbf {q} (t) = \\mathbf {q} (0) + \\int_ {0} ^ {t} \\frac {d}{d t} \\mathbf {q} (t) d t = \\mathbf {q} (0) + \\sum_ {i = 1} ^ {T} \\frac {d}{d t} \\mathbf {q} (t) | _ {t = t _ {i}} \\Delta_ {t} \\tag {2}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.215, + 0.765, + 0.784, + 0.841 + ], + "angle": 0, + "content": "where \\(\\mathbf{q}(0)\\) is the initial condition specified in the input image. We compute the integral using numerical ODE integrations, e.g. Euler method [20]. Given the initial state of the objects, the physical motion simulation module synthesis each foreground object future motion. The location of each object is updated by the affine transformation \\(\\mathbf{T}(t) = [\\mathbf{R}(t),\\mathbf{t}(t)]\\) from the initial location." + } + ], + [ + { + "type": "page_number", + "bbox": [ + 0.218, + 0.116, + 0.23, + 0.127 + ], + "angle": 0, + "content": "8" + }, + { + "type": "header", + "bbox": [ + 0.272, + 0.115, + 0.353, + 0.127 + ], + "angle": 0, + "content": "S. Liu et al." + }, + { + "type": "image", + "bbox": [ + 0.269, + 0.144, + 0.387, + 0.235 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.313, + 0.236, + 0.341, + 0.245 + ], + "angle": 0, + "content": "Input" + }, + { + "type": "image", + "bbox": [ + 0.387, + 0.145, + 0.505, + 0.235 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.391, + 0.236, + 0.498, + 0.245 + ], + "angle": 0, + "content": "Composed (Sec. 3.3)" + }, + { + "type": "image", + "bbox": [ + 0.505, + 0.145, + 0.622, + 0.235 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.521, + 0.236, + 0.603, + 0.245 + ], + "angle": 0, + "content": "Relight (Sec. 3.3)" + }, + { + "type": "image", + "bbox": [ + 0.622, + 0.145, + 0.74, + 0.235 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.638, + 0.236, + 0.721, + 0.245 + ], + "angle": 0, + "content": "Render (Sec. 3.4)" + }, + { + "type": "image_caption", + "bbox": [ + 0.214, + 0.245, + 0.787, + 0.3 + ], + "angle": 0, + "content": "Fig. 3: Rendered video comparison. We show a toy example of piggy bank. The left shows the input frame, and the rest 3 are future frame generations. The composited piggy bank hasn't been aware of the light change. The relighted output synthesized no shadows beneath. The rendered output from diffusion model is most photo realistic." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.343, + 0.784, + 0.404 + ], + "angle": 0, + "content": "External force and torque. For each object, the external forces \\(\\mathbf{F}(t)\\) and torques \\(\\pmb{\\tau}\\) include gravity, friction, and elasticity between the object's surface and the environment. Specifically, we consider the initial external force and torque from user input, as well as gravity, rolling friction, and sliding friction." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.418, + 0.785, + 0.525 + ], + "angle": 0, + "content": "Collision. When objects move, they might collide with each other or with the scene boundary (e.g., hitting a wall). Therefore, collision checking is necessary at every time step. A collision will result in offset forces being applied from its center of mass, producing \\(\\tau\\) and causing the object to rotate. In collision reactions, changes in energy and momentum are minimized to adhere to Newton's conservation laws. In the case of an object falling, it could bounce back or start rolling on the ground due to the elasticity of both the object and the ground." + }, + { + "type": "title", + "bbox": [ + 0.215, + 0.546, + 0.349, + 0.561 + ], + "angle": 0, + "content": "3.3 Rendering" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.571, + 0.788, + 0.709 + ], + "angle": 0, + "content": "Given the simulated motion sequence \\(\\{\\mathbf{q}^i(0),\\dots,\\mathbf{q}^i(t),\\dots\\}_i^N\\) and the input image \\(\\mathbf{C}\\), the rendering module outputs the rendered video \\(\\{\\dots,\\mathbf{C}(t),\\dots\\}\\). There are a few desiderata for generating realistic video: it needs to be temporally consistent, complete, reflect the lighting changes as the object moves, and accurately represent the simulated image space dynamics. To achieve this goal, we design a novel motion-guided rendering pipeline consisting of an image-based transformation and composition module to ensure motion awareness, a relighting module to ensure the plausibility of lighting physics and a generative video editing module that retouches the composed and relit video to ensure realism." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.722, + 0.788, + 0.813 + ], + "angle": 0, + "content": "Composed video. Given the object state from the physical motion module, we compose an initial video by alpha-blending the foreground scene with the static inpainted background from Sec. 3.1. The foreground scene is rendered by performing forward-warping from the input image to future frames using the affine transformation \\(\\mathbf{T}(t)\\). The alpha channel is computed using the same procedure, with the input being the segmentation mask in Sec. 3.1:" + }, + { + "type": "equation", + "bbox": [ + 0.273, + 0.824, + 0.727, + 0.842 + ], + "angle": 0, + "content": "\\[\n\\hat {\\mathbf {X}} (t) = \\operatorname {c o m p o s i t e} (\\mathbf {B}, \\operatorname {w a r p} \\left(\\mathbf {X} ^ {0}, \\mathbf {T} ^ {0} (t)\\right) \\dots \\operatorname {w a r p} \\left(\\mathbf {X} ^ {i}, \\mathbf {T} ^ {i} (t)\\right) \\dots \\dots)\n\\]" + } + ], + [ + { + "type": "header", + "bbox": [ + 0.439, + 0.115, + 0.732, + 0.129 + ], + "angle": 0, + "content": "PhysGen: Rigid Physics-Grounded Img2Vid" + }, + { + "type": "page_number", + "bbox": [ + 0.775, + 0.117, + 0.785, + 0.127 + ], + "angle": 0, + "content": "9" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.147, + 0.784, + 0.194 + ], + "angle": 0, + "content": "where \\(\\mathbf{X}^i\\) are segmented input image of \\(i\\)-th object. Apart from the RGB sequence \\(\\hat{\\mathbf{V}} = \\{\\dots \\hat{\\mathbf{X}}(t) \\dots\\}\\), we also use the same affine warping and composition to compute the blended albedo map \\(\\hat{\\mathbf{A}}(t)\\), which is later used for relighting." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.208, + 0.785, + 0.316 + ], + "angle": 0, + "content": "Relighting. Our relighting module simulates the changes in shading due to object movement. Specifically, it takes as input the transformed albedo \\(\\hat{\\mathbf{A}}(t)\\) and RGB image \\(\\hat{\\mathbf{X}}(t)\\) at each time \\(t\\), as well as the estimated lighting \\(\\mathbf{L}\\). We first compute the surface normal \\(\\hat{\\mathbf{N}}(t)\\) from the composited RGB image \\(\\hat{\\mathbf{X}}(t)\\) using an off-the-shelf normal estimator [23]. We then perform reshading using the parametric illumination model [15]: \\(\\tilde{\\mathbf{X}}(t) = f(\\hat{\\mathbf{X}}(t), \\hat{\\mathbf{A}}(t), \\hat{\\mathbf{N}}(t), \\mathbf{L})\\), which returns the relit image \\(\\hat{\\mathbf{X}}(t)\\) for each frame \\(t\\)." + }, + { + "type": "title", + "bbox": [ + 0.215, + 0.339, + 0.681, + 0.353 + ], + "angle": 0, + "content": "3.4 Generative refinement with latent diffusion models" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.363, + 0.785, + 0.453 + ], + "angle": 0, + "content": "One limitation of the single-image based relighting model is that it neglects the background and cannot handle complicated lighting effects due to complex object-object interactions, such as ambient occlusion and cast shadows. Moreover, the composition results in unnatural boundaries. To address these issues and enhance realism, we incorporated a diffusion-based video to refine the relit video, denoted as \\(\\tilde{\\mathbf{V}} = \\{\\ldots ,\\tilde{\\mathbf{X}} (t),\\ldots \\}\\), and obtain the final video output." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.454, + 0.786, + 0.709 + ], + "angle": 0, + "content": "Specifically, we incorporate a pretrained video diffusion model to refine our video \\(\\tilde{\\mathbf{V}}\\). We first use the pretrained latent diffusion encoder [62] to encode the guidance video \\(\\tilde{\\mathbf{V}}\\) into a latent code \\(\\mathbf{z}\\). Inspired by SDEdit [53], we add noise to the guided latent code and gradually denoise it using the pretrained video diffusion. Given that the content in the guided latent code is already satisfactory, we do not perform the denoising process from scratch. Instead, we define an initial noise strength \\(s\\) where \\(s \\in [0,1]\\) controls the amount of noise added to \\(\\mathbf{z}\\). In practice, we find that a certain approach finds a good trade-off between fidelity and realism. At each denoising step, we ensure the foreground objects are as consistent as possible with the guidance (copied from the guided latent code) while synthesizing new content (e.g., shadows) in the background (using the denoised latent code). To this end, we use a fusion weight \\(w\\) to control how much of the generated latent to inject into the foreground; the fusion weight is gradually increased during the denoising as the perturbation of noise decreases. We also define a fusion timestamp \\(\\delta\\) as the signal to stop the fusion. The final denoised latent code \\(\\mathbf{z}_{*}\\) is then decoded to our final video output \\(\\mathbf{V}\\). Please see the supplementary material for a detailed algorithm." + }, + { + "type": "text", + "bbox": [ + 0.215, + 0.71, + 0.784, + 0.74 + ], + "angle": 0, + "content": "Fig. 3 depicts the composed video \\(\\hat{\\mathbf{V}}\\), the relit video \\(\\tilde{\\mathbf{V}}\\), and our final output \\(\\mathbf{V}\\), illustrating the effectiveness of each rendering component." + }, + { + "type": "title", + "bbox": [ + 0.215, + 0.764, + 0.376, + 0.781 + ], + "angle": 0, + "content": "4 Experiments" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.795, + 0.785, + 0.84 + ], + "angle": 0, + "content": "Implementation details. For physical body simulation, we use a 2D rigid body physics simulator Pymunk [13]. For each experiment, we run 120 simulation steps and uniformly sample 16 frames to form the videos. We set the" + } + ], + [ + { + "type": "page_number", + "bbox": [ + 0.218, + 0.116, + 0.236, + 0.127 + ], + "angle": 0, + "content": "10" + }, + { + "type": "header", + "bbox": [ + 0.272, + 0.115, + 0.352, + 0.127 + ], + "angle": 0, + "content": "S. Liu et al." + }, + { + "type": "image", + "bbox": [ + 0.229, + 0.145, + 0.782, + 0.506 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.216, + 0.507, + 0.785, + 0.535 + ], + "angle": 0, + "content": "Fig. 4: Video generation results. Left: Input initial frame with red arrows representing the initial force or speed on the object. Right: Generated future frames." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.575, + 0.788, + 0.637 + ], + "angle": 0, + "content": "generation resolution at \\(512 \\times 512\\) so that frames can easily fit into the diffusion models. For video diffusion prior model, we use SEINE [18]. For inference, DDIM sampling [67] is used and the noise strength \\(i\\) set to \\(s = 0.5\\), i.e., executing 25/50 steps for denoising. The latents fusion stops at timestamp \\(\\delta = 5\\)." + }, + { + "type": "text", + "bbox": [ + 0.216, + 0.649, + 0.788, + 0.68 + ], + "angle": 0, + "content": "Speed. PhysGen's run-time speed is 3 minutes for one image-to-video gen on a Nvidia A40 GPU. Image understanding and relighting takes the most time." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.692, + 0.788, + 0.798 + ], + "angle": 0, + "content": "Data. To show the generalizability and robustness of our method, we use both internet data and self-captured indoor images from a cell phone. The dataset is diverse, with variations in lighting, object count, geometries, physical attributes, and environmental boundaries. We compare different approaches on 15 photos that require physical reasoning. Additionally, we collect 50 recorded videos of a given scene as ground truth by varying input conditions for quantitative comparison. Randomly selected sequences are shown in the supplementary material." + }, + { + "type": "text", + "bbox": [ + 0.216, + 0.81, + 0.787, + 0.841 + ], + "angle": 0, + "content": "Baselines. We compare against two streams of approaches. The first stream is current state-of-the-art I2V models: SEINE [18], I2VGen-XL [95] and Dynam-" + } + ], + [ + { + "type": "header", + "bbox": [ + 0.439, + 0.115, + 0.732, + 0.129 + ], + "angle": 0, + "content": "PhysGen: Rigid Physics-Grounded Img2Vid" + }, + { + "type": "page_number", + "bbox": [ + 0.769, + 0.117, + 0.784, + 0.127 + ], + "angle": 0, + "content": "11" + }, + { + "type": "table", + "bbox": [ + 0.3, + 0.144, + 0.709, + 0.22 + ], + "angle": 0, + "content": "
MethodsPhysical-realism↑Photo-realism↑
SEINE [18]1.391.86
DynamiCrafter [87]1.681.81
I2VGen-XL [95]2.112.25
Ours4.143.86
" + }, + { + "type": "table_caption", + "bbox": [ + 0.214, + 0.221, + 0.788, + 0.276 + ], + "angle": 0, + "content": "Table 1: Human evaluation. We evaluate both physical-realism and photo-realism of the three competing image-to-video models. The user is asked to evaluate the generated video using a 5-point scale from strongly disagree (1) to strongly agree(5) that the video is physical-realistic and photo-realistic. Our method ranks the first in both terms." + }, + { + "type": "table", + "bbox": [ + 0.306, + 0.304, + 0.699, + 0.38 + ], + "angle": 0, + "content": "
MethodsImage-FID↓Motion-FID↓
SEINE [18]138.8938.76
DynamiCrafter [87]99.64109.61
I2VGen-XL [95]104.6282.04
Ours105.7030.20
" + }, + { + "type": "table_caption", + "bbox": [ + 0.214, + 0.381, + 0.788, + 0.422 + ], + "angle": 0, + "content": "Table 2: Quantitative evaluation. We measure the Image-FID and Motion-FID of videos generated by four methods against the collected GT videos for the given scene. Our method achieves both low Image-FID and Motion-FID." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.466, + 0.788, + 0.573 + ], + "angle": 0, + "content": "iCrafter [87]. These methods take both an image and a text prompt as input for generation. The text prompt is generated by GPT-4V [56] to describe the future events. Another stream we compare is the image-based manipulation approach Motion Guidance [28]. This method uses an image and optical flow as inputs to predict future movement and generate corresponding images. We employ computed motion flow from our physical motion module as the target input for Motion Guidance to optimize each video frame." + }, + { + "type": "title", + "bbox": [ + 0.215, + 0.598, + 0.325, + 0.612 + ], + "angle": 0, + "content": "4.1 Results" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.626, + 0.787, + 0.687 + ], + "angle": 0, + "content": "We show our generated results on 8 different physical procedures from multiple sources in Fig. 4. Our method could simulate reasonable complex physical procedures which involves object-object interaction, object-scene interaction with different physical properties and initial conditions." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.704, + 0.788, + 0.841 + ], + "angle": 0, + "content": "Visual comparisons. The comparison results are shown in Fig. 5. As can be seen, image-video generation methods could not produce any physically plausible futures. For each method, even we carefully tune the input prompt, the models couldn't understand the direction, physical meanings and interactions. All 3 models fail to synthesize realistic videos. For motion guidance with accurate motion flow provided, the results is better but brings a lot inconsistency. When there are multiple objects interaction, the method suffers from introducing new contents. For the domino case, the generation is not smooth and contain artifacts as it could not accurately model the physical procedure." + } + ], + [ + { + "type": "page_number", + "bbox": [ + 0.218, + 0.116, + 0.236, + 0.127 + ], + "angle": 0, + "content": "12" + }, + { + "type": "header", + "bbox": [ + 0.272, + 0.115, + 0.352, + 0.127 + ], + "angle": 0, + "content": "S. Liu et al." + }, + { + "type": "image", + "bbox": [ + 0.226, + 0.143, + 0.782, + 0.825 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.216, + 0.827, + 0.787, + 0.855 + ], + "angle": 0, + "content": "Fig.5: Qualitative comparison. against DynamiciCafter(DC) [87], I2VGenXL(VGen) [95], SEINE [18], Motion Guidance(MG) [28]." + } + ], + [ + { + "type": "header", + "bbox": [ + 0.439, + 0.115, + 0.732, + 0.129 + ], + "angle": 0, + "content": "PhysGen: Rigid Physics-Grounded Img2Vid" + }, + { + "type": "page_number", + "bbox": [ + 0.769, + 0.116, + 0.785, + 0.127 + ], + "angle": 0, + "content": "13" + }, + { + "type": "image_caption", + "bbox": [ + 0.255, + 0.146, + 0.292, + 0.157 + ], + "angle": 0, + "content": "Input" + }, + { + "type": "image", + "bbox": [ + 0.231, + 0.158, + 0.318, + 0.225 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.231, + 0.226, + 0.318, + 0.295 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.32, + 0.158, + 0.409, + 0.225 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.321, + 0.226, + 0.409, + 0.295 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.439, + 0.146, + 0.66, + 0.158 + ], + "angle": 0, + "content": "Generation (left \\(\\rightarrow\\) right: time steps)" + }, + { + "type": "image", + "bbox": [ + 0.411, + 0.158, + 0.499, + 0.225 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.411, + 0.226, + 0.499, + 0.295 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.502, + 0.158, + 0.589, + 0.225 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.502, + 0.226, + 0.589, + 0.295 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.591, + 0.158, + 0.679, + 0.225 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.591, + 0.226, + 0.679, + 0.295 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.682, + 0.158, + 0.769, + 0.225 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.682, + 0.226, + 0.769, + 0.295 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.224, + 0.296, + 0.778, + 0.311 + ], + "angle": 0, + "content": "Fig. 6: Controllability. PhysGen generates diverse results reflecting initial forces." + }, + { + "type": "image_caption", + "bbox": [ + 0.248, + 0.338, + 0.287, + 0.35 + ], + "angle": 0, + "content": "Input" + }, + { + "type": "image", + "bbox": [ + 0.223, + 0.351, + 0.317, + 0.424 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.339, + 0.338, + 0.391, + 0.351 + ], + "angle": 0, + "content": "FG Seg" + }, + { + "type": "image", + "bbox": [ + 0.345, + 0.375, + 0.372, + 0.394 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.435, + 0.338, + 0.487, + 0.351 + ], + "angle": 0, + "content": "BG Seg" + }, + { + "type": "image", + "bbox": [ + 0.414, + 0.351, + 0.508, + 0.424 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.523, + 0.339, + 0.593, + 0.351 + ], + "angle": 0, + "content": "Inpainting" + }, + { + "type": "image", + "bbox": [ + 0.517, + 0.351, + 0.604, + 0.424 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.623, + 0.339, + 0.674, + 0.35 + ], + "angle": 0, + "content": "Normal" + }, + { + "type": "image", + "bbox": [ + 0.607, + 0.359, + 0.69, + 0.423 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.693, + 0.339, + 0.801, + 0.351 + ], + "angle": 0, + "content": "inferred Physics" + }, + { + "type": "image", + "bbox": [ + 0.706, + 0.359, + 0.788, + 0.423 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.215, + 0.427, + 0.785, + 0.455 + ], + "angle": 0, + "content": "Fig. 7: Qualitative results from the perception module. The results include foreground and background mask, inpainting output, normal map and inferred physics." + }, + { + "type": "title", + "bbox": [ + 0.216, + 0.497, + 0.414, + 0.511 + ], + "angle": 0, + "content": "4.2 Human evaluation" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.522, + 0.793, + 0.596 + ], + "angle": 0, + "content": "To comprehensively assess both the physical-realism and photo-realism of the generated videos, we conducted human evaluations following a similar methodology to prior work [16,78,83,91]. Here by physical-realism we mean whether the content of the generated video is realistic in its dynamics and object interactions and accurately reflects the instructed movements." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.598, + 0.789, + 0.765 + ], + "angle": 0, + "content": "Specifically, we include 15 videos with diverse objects and conditions for evaluation, compared against other three I2V model generations (SEINE [18], DynamiCrafter [87], I2VGen-XL [95]). For each compared baseline method, we meticulously adjusted the input prompts and conducted multiple runs to select the most plausible output. In contrast, our method did not undergo specific tuning for each video. We randomized the presentation order and asked 14 participants to rate the videos on a five-point scale, from strongly disagree (1) to strongly agree (5) for physical-realism and photo-realism. The results in Tab. 1 show that our method achieves the highest ratings. Our average scores for physical-realism (4.14) and photo-realism (3.86) fall within the Agree level, while other methods perform poorly." + }, + { + "type": "title", + "bbox": [ + 0.215, + 0.786, + 0.456, + 0.801 + ], + "angle": 0, + "content": "4.3 Quantitative evaluation" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.81, + 0.787, + 0.84 + ], + "angle": 0, + "content": "We quantitatively evaluate the collected GT videos of a given scene. The GT data includes 833 images. For each method, we generate 50 videos with random" + } + ], + [ + { + "type": "page_number", + "bbox": [ + 0.218, + 0.116, + 0.236, + 0.127 + ], + "angle": 0, + "content": "14" + }, + { + "type": "header", + "bbox": [ + 0.272, + 0.115, + 0.353, + 0.128 + ], + "angle": 0, + "content": "S. Liu et al." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.147, + 0.788, + 0.329 + ], + "angle": 0, + "content": "samples, collecting 800 images per method. Our method varies initial conditions to ensure all generated videos are different. Following [12], we adopt the Fréchet Inception Distance (FID) [32] as evaluation metric. We didn't use Fréchet Video Distance (FVD) [73] due to the small sample size and the goal to disentangle evaluation of appearance and motion. We evaluate appearance using Image-FID and motion using Motion-FID. For Motion-FID, we use RAFT [71] to extract and colorize optical flow [5], then compute FID from these rendered images. The results are shown in Tab. 2. We notice other methods which either generate minimal motion, leading to low Image-FID and high Motion-FID (e.g., DynamiCrafter [87]), or produce realistic motion but fail to maintain image consistency, resulting in low Motion-FID and high Image-FID (e.g., SEINE [18]). Our approach achieves both low Image-FID and Motion-FID compared to others." + }, + { + "type": "title", + "bbox": [ + 0.216, + 0.349, + 0.334, + 0.364 + ], + "angle": 0, + "content": "4.4 Analysis" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.372, + 0.788, + 0.448 + ], + "angle": 0, + "content": "Perception evaluation. We evaluate the perception module using 10 complex open-world images with 118 annotated movable instances (COCO [47] format). Our system achieves 0.93 precision, 0.82 recall at 0.5 IoU. More details are shown in the supplementary. Fig. 7 shows intermediate outputs from our perception module and simulation module." + }, + { + "type": "text", + "bbox": [ + 0.215, + 0.46, + 0.788, + 0.522 + ], + "angle": 0, + "content": "Physical reasoning. We compare the physical parameters estimated by GPT-4V to random values sampled from human-estimated ranges. Without GPT-4V, Image-FID increases from 105.70 to 111.01, Motion-FID from 30.20 to 36.60. It shows the physical reasoning are crucial for appearance and motion realism." + }, + { + "type": "text", + "bbox": [ + 0.215, + 0.534, + 0.788, + 0.61 + ], + "angle": 0, + "content": "Controllability. We showcase the controllability and diversity of our method in Fig. 6 by varying the initial conditions. For different rows, we adjust the direction and magnitude of initial speed and forces. The results show diverse physical-plausible trajectories. This demonstrates the potential of applying our method for diverse physical world generations." + }, + { + "type": "text", + "bbox": [ + 0.215, + 0.623, + 0.788, + 0.7 + ], + "angle": 0, + "content": "Limitations. Our method is an initial step towards physics-based video generation but has two main limitations: it focuses mainly on rigid objects, limiting its application to non-rigid ones, and it lacks a comprehensive 3D understanding, making it unable to handle out-of-plane motions. We leave leveraging deformable physics and full 3D understanding as future work." + }, + { + "type": "title", + "bbox": [ + 0.216, + 0.72, + 0.36, + 0.736 + ], + "angle": 0, + "content": "5 Conclusion" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.75, + 0.788, + 0.841 + ], + "angle": 0, + "content": "We present PhysGen, a novel image-to-video generation method that turns a static image into a high-fidelity, physics-grounded, and temporally coherent videos. The main module is a model-based physical simulator with a data-driven video generation process. PhysGen enables a series of controllable and interactive downstream applications. PhysGen provide a new image-to-video generation paradigm and hopefully can inspire more follow-up works." + } + ], + [ + { + "type": "header", + "bbox": [ + 0.438, + 0.114, + 0.732, + 0.129 + ], + "angle": 0, + "content": "PhysGen: Rigid Physics-Grounded Img2Vid" + }, + { + "type": "page_number", + "bbox": [ + 0.768, + 0.116, + 0.786, + 0.127 + ], + "angle": 0, + "content": "15" + }, + { + "type": "title", + "bbox": [ + 0.218, + 0.145, + 0.403, + 0.163 + ], + "angle": 0, + "content": "Acknowledgements" + }, + { + "type": "text", + "bbox": [ + 0.217, + 0.177, + 0.787, + 0.238 + ], + "angle": 0, + "content": "This project is supported by NSF Awards #2331878, #2340254, and #2312102, the IBM IIDAI Grant, and an Intel Research Gift. We greatly appreciate the NCSA for providing computing resources. We thank Tianhang cheng for helpful discussions. We thank Emily Chen and Gloria Wang for proofreading." + } + ], + [ + { + "type": "page_number", + "bbox": [ + 0.218, + 0.116, + 0.236, + 0.127 + ], + "angle": 0, + "content": "16" + }, + { + "type": "header", + "bbox": [ + 0.272, + 0.115, + 0.352, + 0.127 + ], + "angle": 0, + "content": "S. Liu et al." + }, + { + "type": "title", + "bbox": [ + 0.218, + 0.145, + 0.323, + 0.16 + ], + "angle": 0, + "content": "References" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.179, + 0.785, + 0.208 + ], + "angle": 0, + "content": "1. Aberman, K., Weng, Y., Lischinski, D., Cohen-Or, D., Chen, B.: Unpaired motion style transfer from video to animation. ACM TOG (2020)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.209, + 0.785, + 0.249 + ], + "angle": 0, + "content": "2. Achiam, J., Adler, S., Agarwal, S., Ahmad, L., Akkaya, I., Aleman, F.L., Almeida, D., Altenschmidt, J., Altman, S., Anadkat, S., et al.: Gpt-4 technical report. arXiv preprint arXiv:2303.08774 (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.25, + 0.785, + 0.292 + ], + "angle": 0, + "content": "3. Ajay, A., Bauza, M., Wu, J., Fazeli, N., Tenenbaum, J.B., Rodriguez, A., Kaelbling, L.P.: Combining Physical Simulators and Object-Based Networks for Control. In: ICRA (2019)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.293, + 0.785, + 0.333 + ], + "angle": 0, + "content": "4. Ajay, A., Wu, J., Fazeli, N., Bauza, M., Kaelbling, L.P., Tenenbaum, J.B., Rodriguez, A.: Augmenting Physical Simulators with Stochastic Neural Networks: Case Study of Planar Pushing and Bouncing. In: IROS (2018)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.335, + 0.785, + 0.362 + ], + "angle": 0, + "content": "5. Baker, S., Scharstein, D., Lewis, J.P., Roth, S., Black, M.J., Szeliski, R.: A database and evaluation methodology for optical flow. IJCV (2011)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.363, + 0.785, + 0.404 + ], + "angle": 0, + "content": "6. Bar-Tal, O., Chefer, H., Tov, O., Herrmann, C., Paiss, R., Zada, S., Ephrat, A., Hur, J., Li, Y., Michaeli, T., et al.: Lumiere: A space-time diffusion model for video generation. arXiv preprint arXiv:2401.12945 (2024)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.406, + 0.785, + 0.433 + ], + "angle": 0, + "content": "7. Battaglia, P., Pascanu, R., Lai, M., Jimenez Rezende, D., et al.: Interaction networks for learning about objects, relations and physics. NeurIPS (2016)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.434, + 0.785, + 0.489 + ], + "angle": 0, + "content": "8. Blattmann, A., Dockhorn, T., Kulal, S., Mendelevitch, D., Kilian, M., Lorenz, D., Levi, Y., English, Z., Voleti, V., Letts, A., et al.: Stable video diffusion: Scaling latent video diffusion models to large datasets. arXiv preprint arXiv:2311.15127 (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.49, + 0.785, + 0.516 + ], + "angle": 0, + "content": "9. Blattmann, A., Milbich, T., Dorkenwald, M., Ommer, B.: ipoke: Poking a still image for controlled stochastic video synthesis. In: ICCV (2021)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.518, + 0.785, + 0.544 + ], + "angle": 0, + "content": "10. Blattmann, A., Milbich, T., Dorkenwald, M., Ommer, B.: Understanding object dynamics for interactive image-to-video synthesis. In: CVPR (2021)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.546, + 0.785, + 0.572 + ], + "angle": 0, + "content": "1. Blattmann, A., Milbich, T., Dorkenwald, M., Ommer, B.: Understanding object dynamics for interactive image-to-video synthesis. In: CVPR (2021)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.574, + 0.785, + 0.614 + ], + "angle": 0, + "content": "2. Blattmann, A., Rombach, R., Ling, H., Dockhorn, T., Kim, S.W., Fidler, S., Kreis, K.: Align your latents: High-resolution video synthesis with latent diffusion models. In: CVPR (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.616, + 0.597, + 0.63 + ], + "angle": 0, + "content": "3. Blomqvist, V.: Pymunk. https://pymunk.org (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.631, + 0.785, + 0.657 + ], + "angle": 0, + "content": "4. Bowen, R.S., Tucker, R., Zabih, R., Snavely, N.: Dimensions of motion: Monocular prediction through flow subspaces. In: 3DV (2022)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.659, + 0.785, + 0.685 + ], + "angle": 0, + "content": "5. Careaga, C., Miangoleh, S.M.H., Aksoy, Y.: Intrinsic harmonization for illumination-aware compositing. arXiv preprint arXiv:2312.03698 (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.687, + 0.785, + 0.728 + ], + "angle": 0, + "content": "6. Chen, H., Xia, M., He, Y., Zhang, Y., Cun, X., Yang, S., Xing, J., Liu, Y., Chen, Q., Wang, X., et al.: Videocrafter1: Open diffusion models for high-quality video generation. arXiv preprint arXiv:2310.19512 (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.729, + 0.785, + 0.77 + ], + "angle": 0, + "content": "7. Chen, X., Liu, Z., Chen, M., Feng, Y., Liu, Y., Shen, Y., Zhao, H.: Livephoto: Real image animation with text-guided motion control. arXiv preprint arXiv:2312.02928 (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.771, + 0.785, + 0.812 + ], + "angle": 0, + "content": "8. Chen, X., Wang, Y., Zhang, L., Zhuang, S., Ma, X., Yu, J., Wang, Y., Lin, D., Qiao, Y., Liu, Z.: Seine: Short-to-long video diffusion model for generative transition and prediction. arXiv preprint arXiv:2310.20700 (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.813, + 0.785, + 0.84 + ], + "angle": 0, + "content": "9. Chuang, Y.Y., Goldman, D.B., Zheng, K.C., Curless, B., Salesin, D.H., Szeliski, R.: Animating pictures with stochastic motion textures. ACM TOG (2005)" + }, + { + "type": "list", + "bbox": [ + 0.226, + 0.179, + 0.785, + 0.84 + ], + "angle": 0, + "content": null + } + ], + [ + { + "type": "header", + "bbox": [ + 0.439, + 0.115, + 0.732, + 0.129 + ], + "angle": 0, + "content": "PhysGen: Rigid Physics-Grounded Img2Vid" + }, + { + "type": "page_number", + "bbox": [ + 0.769, + 0.117, + 0.785, + 0.127 + ], + "angle": 0, + "content": "17" + }, + { + "type": "ref_text", + "bbox": [ + 0.214, + 0.147, + 0.785, + 0.175 + ], + "angle": 0, + "content": "20. Ciarlet, P.G., Lions, J.L.: Handbook of numerical analysis. Gulf Professional Publishing (1990)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.176, + 0.785, + 0.215 + ], + "angle": 0, + "content": "21. Davis, A., Bouman, K.L., Chen, J.G., Rubinstein, M., Durand, F., Freeman, W.T.: Visual vibrometry: Estimating material properties from small motion in video. In: CVPR (2015)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.216, + 0.785, + 0.242 + ], + "angle": 0, + "content": "22. Davis, A., Chen, J.G., Durand, F.: Image-space modal bases for plausible manipulation of objects in video. ACM TOG (2015)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.243, + 0.785, + 0.27 + ], + "angle": 0, + "content": "23. Eftekhar, A., Sax, A., Malik, J., Zamir, A.: Omnidata: A scalable pipeline for making multi-task mid-level vision datasets from 3d scans. In: ICCV (2021)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.271, + 0.785, + 0.311 + ], + "angle": 0, + "content": "24. Endo, Y., Kanamori, Y., Kuriyama, S.: Animating landscape: self-supervised learning of decoupled motion and appearance for single-image video synthesis. arXiv preprint arXiv:1910.07192 (2019)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.312, + 0.785, + 0.338 + ], + "angle": 0, + "content": "25. Friedland, B.: Control system design: an introduction to state-space methods. Courier Corporation (2012)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.338, + 0.785, + 0.378 + ], + "angle": 0, + "content": "26. Fu, X., Yin, W., Hu, M., Wang, K., Ma, Y., Tan, P., Shen, S., Lin, D., Long, X.: Geowizard: Unleashing the diffusion priors for 3d geometry estimation from a single image. arXiv preprint arXiv:2403.12013 (2024)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.379, + 0.785, + 0.419 + ], + "angle": 0, + "content": "27. Ge, S., Nah, S., Liu, G., Poon, T., Tao, A., Catanzaro, B., Jacobs, D., Huang, J.B., Liu, M.Y., Balaji, Y.: Preserve your own correlation: A noise prior for video diffusion models. In: ICCV (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.42, + 0.785, + 0.446 + ], + "angle": 0, + "content": "28. Geng, D., Owens, A.: Motion guidance: Diffusion-based image editing with differentiable motion estimators. In: ICLR (2024)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.447, + 0.785, + 0.487 + ], + "angle": 0, + "content": "29. Girdhar, R., Singh, M., Brown, A., Duval, Q., Azadi, S., Rambhatla, S.S., Shah, A., Yin, X., Parikh, D., Misra, I.: Emu video: Factorizing text-to-video generation by explicit image conditioning. arXiv preprint arXiv:2311.10709 (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.488, + 0.785, + 0.528 + ], + "angle": 0, + "content": "30. Guo, Y., Yang, C., Rao, A., Liang, Z., Wang, Y., Qiao, Y., Agrawala, M., Lin, D., Dai, B.: Animatediff: Animate your personalized text-to-image diffusion models without specific tuning (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.529, + 0.785, + 0.568 + ], + "angle": 0, + "content": "31. Gupta, A., Yu, L., Sohn, K., Gu, X., Hahn, M., Fei-Fei, L., Essa, I., Jiang, L., Lezama, J.: Photorealistic video generation with diffusion models. arXiv preprint arXiv:2312.06662 (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.569, + 0.785, + 0.609 + ], + "angle": 0, + "content": "32. Heusel, M., Ramsauer, H., Unterthiner, T., Nessler, B., Hochreiter, S.: Gans trained by a two time-scale update rule converge to a local nash equilibrium. NeurIPS (2017)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.61, + 0.785, + 0.651 + ], + "angle": 0, + "content": "33. Ho, J., Chan, W., Sahara, C., Whang, J., Gao, R., Gritsenko, A., Kingma, D.P., Poole, B., Norouzi, M., Fleet, D.J., et al.: Imagen video: High definition video generation with diffusion models. arXiv preprint arXiv:2210.02303 (2022)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.651, + 0.785, + 0.677 + ], + "angle": 0, + "content": "34. Ho, J., Jain, A., Abbeel, P.: Denoising diffusion probabilistic models. NeurIPS (2020)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.678, + 0.785, + 0.704 + ], + "angle": 0, + "content": "35. Holynski, A., Curless, B.L., Seitz, S.M., Szeliski, R.: Animating pictures with eulerian motion fields. In: CVPR (2021)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.705, + 0.785, + 0.732 + ], + "angle": 0, + "content": "36. Hu, Y., Anderson, L., Li, T.M., Sun, Q., Carr, N., Ragan-Kelley, J., Durand, F.: Difftaichi: Differentiable programming for physical simulation. ICLR (2020)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.733, + 0.785, + 0.772 + ], + "angle": 0, + "content": "37. Hu, Y., Fang, Y., Ge, Z., Qu, Z., Zhu, Y., Pradhana, A., Jiang, C.: A moving least squares material point method with displacement discontinuity and two-way rigid body coupling. ACM TOG (2018)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.773, + 0.785, + 0.812 + ], + "angle": 0, + "content": "38. Hu, Y., Li, T.M., Anderson, L., Ragan-Kelley, J., Durand, F.: Taichi: a language for high-performance computation on spatially sparse data structures. ACM TOG (2019)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.813, + 0.785, + 0.84 + ], + "angle": 0, + "content": "39. Jhou, W.C., Cheng, W.H.: Animating still landscape photographs through cloud motion creation. IEEE Transactions on Multimedia (2015)" + }, + { + "type": "list", + "bbox": [ + 0.214, + 0.147, + 0.785, + 0.84 + ], + "angle": 0, + "content": null + } + ], + [ + { + "type": "page_number", + "bbox": [ + 0.218, + 0.116, + 0.236, + 0.127 + ], + "angle": 0, + "content": "18" + }, + { + "type": "header", + "bbox": [ + 0.272, + 0.115, + 0.352, + 0.127 + ], + "angle": 0, + "content": "S. Liu et al." + }, + { + "type": "ref_text", + "bbox": [ + 0.218, + 0.147, + 0.785, + 0.189 + ], + "angle": 0, + "content": "40. Kirillov, A., Mintun, E., Ravi, N., Mao, H., Rolland, C., Gustafson, L., Xiao, T., Whitehead, S., Berg, A.C., Lo, W.Y., Dollár, P., Girshick, R.: Segment anything. arXiv:2304.02643 (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.19, + 0.785, + 0.218 + ], + "angle": 0, + "content": "41. Koenig, N., Howard, A.: Design and use paradigms for gazebo, an open-source multi-robot simulator. In: IROS (2004)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.218, + 0.785, + 0.259 + ], + "angle": 0, + "content": "42. Kondratyuk, D., Yu, L., Gu, X., Lezama, J., Huang, J., Hornung, R., Adam, H., Akbari, H., Alon, Y., Birodkar, V., et al.: Videopoet: A large language model for zero-shot video generation. arXiv preprint arXiv:2312.14125 (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.259, + 0.785, + 0.3 + ], + "angle": 0, + "content": "43. Li, X., Qiao, Y.L., Chen, P.Y., Jatavallabhula, K.M., Lin, M., Jiang, C., Gan, C.: Pac-nerf: Physics augmented continuum neural radiance fields for geometry-agnostic system identification. In: ICLR (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.3, + 0.785, + 0.328 + ], + "angle": 0, + "content": "44. Li, Y., Lin, T., Yi, K., Bear, D., Yamins, D.L., Wu, J., Tenenbaum, J.B., Torralba, A.: Visual grounding of learned physical models. In: ICML (2020)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.328, + 0.785, + 0.37 + ], + "angle": 0, + "content": "45. Li, Y., Wu, J., Tedrake, R., Tenenbaum, J.B., Torralba, A.: Learning particle dynamics for manipulating rigid bodies, deformable objects, and fluids. In: ICLR (2019)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.37, + 0.785, + 0.398 + ], + "angle": 0, + "content": "46. Li, Z., Tucker, R., Snavely, N., Holynski, A.: Generative image dynamics. arXiv preprint arXiv:2309.07906 (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.398, + 0.785, + 0.425 + ], + "angle": 0, + "content": "47. Lin, T.Y., Maire, M., Belongie, S., Hays, J., Perona, P., Ramanan, D., Dólar, P., Zitnick, C.L.: Microsoft coco: Common objects in context. In: ECCV (2014)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.425, + 0.785, + 0.467 + ], + "angle": 0, + "content": "48. Liu, S., Zeng, Z., Ren, T., Li, F., Zhang, H., Yang, J., Li, C., Yang, J., Su, H., Zhu, J., et al.: Grounding dino: Marrying dino with grounded pre-training for open-set object detection. arXiv preprint arXiv:2303.05499 (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.467, + 0.785, + 0.494 + ], + "angle": 0, + "content": "49. Liu, T., Bargteil, A.W., O'Brien, J.F., Kavan, L.: Fast simulation of mass-spring systems. ACM TOG (2013)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.494, + 0.785, + 0.536 + ], + "angle": 0, + "content": "50. Lv, J., Huang, Y., Yan, M., Huang, J., Liu, J., Liu, Y., Wen, Y., Chen, X., Chen, S.: Gpt4motion: Scripting physical motions in text-to-video generation via blender-oriented gpt planning. In: CVPRW. pp. 1430-1440 (2024)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.536, + 0.785, + 0.563 + ], + "angle": 0, + "content": "51. Mahapatra, A., Kulkarni, K.: Controllable animation of fluid elements in still images. In: CVPR (2022)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.564, + 0.785, + 0.591 + ], + "angle": 0, + "content": "52. Mallya, A., Wang, T.C., Liu, M.Y.: Implicit warping for animation with image sets. In: NeurIPS (2022)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.591, + 0.785, + 0.633 + ], + "angle": 0, + "content": "53. Meng, C., He, Y., Song, Y., Song, J., Wu, J., Zhu, J.Y., Ermon, S.: SDEdit: Guided image synthesis and editing with stochastic differential equations. In: International Conference on Learning Representations (2022)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.633, + 0.785, + 0.674 + ], + "angle": 0, + "content": "54. Mrowca, D., Zhuang, C., Wang, E., Haber, N., Fei-Fei, L.F., Tenenbaum, J., Yamins, D.L.: Flexible neural representation for physics prediction. In: NeurIPS (2018)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.674, + 0.749, + 0.688 + ], + "angle": 0, + "content": "55. NVIDIA: Nvidia physx. https://developer.nvidia.com/physx-sdk (2019)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.688, + 0.528, + 0.702 + ], + "angle": 0, + "content": "56. OpenAI: Gpt-4v(ision) system card (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.702, + 0.706, + 0.716 + ], + "angle": 0, + "content": "57. OpenAI: Creating video from text. https://openai.com/sora (2024)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.716, + 0.785, + 0.743 + ], + "angle": 0, + "content": "58. Peebles, W., Xie, S.: Scalable diffusion models with transformers. arXiv preprint arXiv:2212.09748 (2022)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.743, + 0.785, + 0.771 + ], + "angle": 0, + "content": "59. Popova, E., Popov, V.L.: The research works of coulomb and amontons and generalized laws of friction. Friction (2015)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.771, + 0.785, + 0.799 + ], + "angle": 0, + "content": "60. Qiu, H., Chen, Z., Jiang, Y., Zhou, H., Fan, X., Yang, L., Wu, W., Liu, Z.: Relitalk: Relightable talking portrait generation from a single video (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.799, + 0.785, + 0.841 + ], + "angle": 0, + "content": "61. Ren, T., Liu, S., Zeng, A., Lin, J., Li, K., Cao, H., Chen, J., Huang, X., Chen, Y., Yan, F., Zeng, Z., Zhang, H., Li, F., Yang, J., Li, H., Jiang, Q., Zhang, L.: Grounded sam: Assembling open-world models for diverse visual tasks (2024)" + }, + { + "type": "list", + "bbox": [ + 0.217, + 0.147, + 0.785, + 0.841 + ], + "angle": 0, + "content": null + } + ], + [ + { + "type": "header", + "bbox": [ + 0.439, + 0.115, + 0.732, + 0.129 + ], + "angle": 0, + "content": "PhysGen: Rigid Physics-Grounded Img2Vid" + }, + { + "type": "page_number", + "bbox": [ + 0.769, + 0.117, + 0.786, + 0.127 + ], + "angle": 0, + "content": "19" + }, + { + "type": "ref_text", + "bbox": [ + 0.214, + 0.147, + 0.788, + 0.175 + ], + "angle": 0, + "content": "62. Rombach, R., Blattmann, A., Lorenz, D., Esser, P., Ommer, B.: High-resolution image synthesis with latent diffusion models. In: CVPR (2022)" + }, + { + "type": "ref_text", + "bbox": [ + 0.214, + 0.175, + 0.788, + 0.217 + ], + "angle": 0, + "content": "63. Sanchez-Gonzalez, A., Heess, N., Springenberg, J.T., Merel, J., Riedmiller, M., Hadsell, R., Battaglia, P.: Graph networks as learnable physics engines for inference and control. In: ICML (2018)" + }, + { + "type": "ref_text", + "bbox": [ + 0.214, + 0.217, + 0.788, + 0.242 + ], + "angle": 0, + "content": "64. Schödl, A., Szeliski, R., Salesin, D.H., Essa, I.: Video textures. In: PACMCGIT (2000)" + }, + { + "type": "ref_text", + "bbox": [ + 0.214, + 0.242, + 0.788, + 0.284 + ], + "angle": 0, + "content": "65. Singer, U., Polyak, A., Hayes, T., Yin, X., An, J., Zhang, S., Hu, Q., Yang, H., Ashual, O., Gafni, O., et al.: Make-a-video: Text-to-video generation without text-video data. arXiv preprint arXiv:2209.14792 (2022)" + }, + { + "type": "ref_text", + "bbox": [ + 0.214, + 0.284, + 0.788, + 0.311 + ], + "angle": 0, + "content": "66. Sohl-Dickstein, J., Weiss, E., Maheswaranathan, N., Ganguli, S.: Deep unsupervised learning using nonequilibrium thermodynamics. In: ICML (2015)" + }, + { + "type": "ref_text", + "bbox": [ + 0.214, + 0.311, + 0.788, + 0.338 + ], + "angle": 0, + "content": "67. Song, J., Meng, C., Ermon, S.: Denoising diffusion implicit models. arXiv preprint arXiv:2010.02502 (2020)" + }, + { + "type": "ref_text", + "bbox": [ + 0.214, + 0.338, + 0.788, + 0.365 + ], + "angle": 0, + "content": "68. Song, Y., Ermon, S.: Generative modeling by estimating gradients of the data distribution. NeurIPS (2019)" + }, + { + "type": "ref_text", + "bbox": [ + 0.214, + 0.365, + 0.788, + 0.392 + ], + "angle": 0, + "content": "69. Sugimoto, R., He, M., Liao, J., Sander, P.V.: Water simulation and rendering from a still photograph. In: SIGGRAPH Asia (2022)" + }, + { + "type": "ref_text", + "bbox": [ + 0.214, + 0.392, + 0.732, + 0.405 + ], + "angle": 0, + "content": "70. Szummer, M., Picard, R.W.: Temporal texture modeling. In: ICIP (1996)" + }, + { + "type": "ref_text", + "bbox": [ + 0.214, + 0.405, + 0.788, + 0.432 + ], + "angle": 0, + "content": "71. Teed, Z., Deng, J.: Raft: Recurrent all-pairs field transforms for optical flow. In: ECCV. pp. 402-419 (2020)" + }, + { + "type": "ref_text", + "bbox": [ + 0.214, + 0.432, + 0.788, + 0.459 + ], + "angle": 0, + "content": "72. Todorov, E., Erez, T., Tassa, Y.: Mujoco: A physics engine for model-based control. In: IROS (2012)" + }, + { + "type": "ref_text", + "bbox": [ + 0.214, + 0.459, + 0.788, + 0.5 + ], + "angle": 0, + "content": "73. Unterthiner, T., Van Steenkiste, S., Kurach, K., Marinier, R., Michalski, M., Gelly, S.: Towards accurate generative models of video: A new metric & challenges. arXiv preprint arXiv:1812.01717 (2018)" + }, + { + "type": "ref_text", + "bbox": [ + 0.214, + 0.5, + 0.788, + 0.542 + ], + "angle": 0, + "content": "74. Villegas, R., Babaeizadeh, M., Kindermans, P.J., Moraldo, H., Zhang, H., Saffar, M.T., Castro, S., Kunze, J., Erhan, D.: Phenaki: Variable length video generation from open domain textual description. arXiv preprint arXiv:2210.02399 (2022)" + }, + { + "type": "ref_text", + "bbox": [ + 0.214, + 0.542, + 0.788, + 0.582 + ], + "angle": 0, + "content": "75. Wang, X., Yuan, H., Zhang, S., Chen, D., Wang, J., Zhang, Y., Shen, Y., Zhao, D., Zhou, J.: Videocomposer: Compositional video synthesis with motion controllability. arXiv preprint arXiv:2306.02018 (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.214, + 0.582, + 0.788, + 0.623 + ], + "angle": 0, + "content": "76. Wang, Y., Chen, X., Ma, X., Zhou, S., Huang, Z., Wang, Y., Yang, C., He, Y., Yu, J., Yang, P., et al.: Lavie: High-quality video generation with cascaded latent diffusion models. arXiv preprint arXiv:2309.15103 (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.214, + 0.623, + 0.788, + 0.65 + ], + "angle": 0, + "content": "77. Wei, L.Y., Levoy, M.: Fast texture synthesis using tree-structured vector quantization. In: PACMCGIT (2000)" + }, + { + "type": "ref_text", + "bbox": [ + 0.214, + 0.65, + 0.788, + 0.691 + ], + "angle": 0, + "content": "78. Wei, Y., Zhang, S., Qing, Z., Yuan, H., Liu, Z., Liu, Y., Zhang, Y., Zhou, J., Shan, H.: Dreamvideo: Composing your dream videos with customized subject and motion. arXiv preprint arXiv:2312.04433 (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.214, + 0.691, + 0.788, + 0.718 + ], + "angle": 0, + "content": "79. Weng, C.Y., Curless, B., Kemelmacher-Shlizerman, I.: Photo wake-up: 3d character animation from a single photo. In: CVPR (2019)" + }, + { + "type": "ref_text", + "bbox": [ + 0.214, + 0.718, + 0.788, + 0.746 + ], + "angle": 0, + "content": "80. Wu, J., Lim, J.J., Zhang, H., Tenenbaum, J.B., Freeman, W.T.: Physics 101: Learning physical object properties from unlabeled videos. In: BMVC (2016)" + }, + { + "type": "ref_text", + "bbox": [ + 0.214, + 0.746, + 0.788, + 0.772 + ], + "angle": 0, + "content": "81. Wu, J., Lu, E., Kohli, P., Freeman, W.T., Tenenbaum, J.B.: Learning to see physics via visual de-animation. In: NeurIPS (2017)" + }, + { + "type": "ref_text", + "bbox": [ + 0.214, + 0.772, + 0.788, + 0.813 + ], + "angle": 0, + "content": "82. Wu, J., Yildirim, I., Lim, J.J., Freeman, W.T., Tenenbaum, J.B.: Galileo: Perceiving physical object properties by integrating a physics engine with deep learning. In: NeurIPS (2015)" + }, + { + "type": "ref_text", + "bbox": [ + 0.214, + 0.813, + 0.788, + 0.841 + ], + "angle": 0, + "content": "83. Wu, R., Chen, L., Yang, T., Guo, C., Li, C., Zhang, X.: Lamp: Learn a motion pattern for few-shot-based video generation. arXiv preprint arXiv:2310.10769 (2023)" + }, + { + "type": "list", + "bbox": [ + 0.214, + 0.147, + 0.788, + 0.841 + ], + "angle": 0, + "content": null + } + ], + [ + { + "type": "page_number", + "bbox": [ + 0.218, + 0.116, + 0.236, + 0.127 + ], + "angle": 0, + "content": "20" + }, + { + "type": "header", + "bbox": [ + 0.272, + 0.115, + 0.352, + 0.127 + ], + "angle": 0, + "content": "S. Liu et al." + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.147, + 0.785, + 0.175 + ], + "angle": 0, + "content": "84. Xia, H., Lin, Z.H., Ma, W.C., Wang, S.: Video2game: Real-time, interactive, realistic and browser-compatible environment from a single video. In: CVPR (2024)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.177, + 0.785, + 0.203 + ], + "angle": 0, + "content": "85. Xia, H., Lin, Z.H., Ma, W.C., Wang, S.: Video2game: Real-time, interactive, realistic and browser-compatible environment from a single video. In: CVPR (2024)" + }, + { + "type": "ref_text", + "bbox": [ + 0.218, + 0.204, + 0.785, + 0.231 + ], + "angle": 0, + "content": "86. Xie, T., Zong, Z., Qiu, Y., Li, X., Feng, Y., Yang, Y., Jiang, C.: Physgaussian: Physics-integrated 3d gaussians for generative dynamics. In: CVPR (2024)" + }, + { + "type": "ref_text", + "bbox": [ + 0.218, + 0.232, + 0.785, + 0.272 + ], + "angle": 0, + "content": "87. Xing, J., Xia, M., Zhang, Y., Chen, H., Yu, W., Liu, H., Wang, X., Wong, T.T., Shan, Y.: Dynamiccafter: Animating open-domain images with video diffusion priors. arXiv preprint arXiv:2310.12190 (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.218, + 0.273, + 0.785, + 0.3 + ], + "angle": 0, + "content": "88. Xu, Z., Wu, J., Zeng, A., Tenenbaum, J.B., Song, S.: Densephysnet: Learning dense physical object representations via multi-step dynamic interactions. In: RSS (2019)" + }, + { + "type": "ref_text", + "bbox": [ + 0.218, + 0.301, + 0.785, + 0.327 + ], + "angle": 0, + "content": "89. Xue, T., Wu, J., Bouman, K.L., Freeman, W.T.: Visual dynamics: Stochastic future generation via layered cross convolutional networks. T-PAMI (2018)" + }, + { + "type": "ref_text", + "bbox": [ + 0.218, + 0.328, + 0.785, + 0.368 + ], + "angle": 0, + "content": "90. Yang, J., Zhang, H., Li, F., Zou, X., Li, C., Gao, J.: Set-of-mark prompting unleashes extraordinary visual grounding in gpt-4v. arXiv preprint arXiv:2310.11441 (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.37, + 0.785, + 0.41 + ], + "angle": 0, + "content": "91. Yu, J., Cun, X., Qi, C., Zhang, Y., Wang, X., Shan, Y., Zhang, J.: Animatezero: Video diffusion models are zero-shot image animators. arXiv preprint arXiv:2312.03793 (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.411, + 0.785, + 0.452 + ], + "angle": 0, + "content": "92. Yu, L., Lezama, J., Gundavarapu, N.B., Versari, L., Sohn, K., Minnen, D., Cheng, Y., Gupta, A., Gu, X., Hauptmann, A.G., et al.: Language model beats diffusion-tokenizer is key to visual generation. arXiv preprint arXiv:2310.05737 (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.453, + 0.785, + 0.48 + ], + "angle": 0, + "content": "93. Yu, T., Feng, R., Feng, R., Liu, J., Jin, X., Zeng, W., Chen, Z.: Inpaint anything: Segment anything meets image inpainting. arXiv preprint arXiv:2304.06790 (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.218, + 0.481, + 0.785, + 0.521 + ], + "angle": 0, + "content": "94. Zhai, A.J., Shen, Y., Chen, E., Wang, G., Wang, X., Wang, S., Wang, X., Guan, K., Wang, S.: Physical property understanding from language-embedded feature fields. In: CVPR (2024)" + }, + { + "type": "ref_text", + "bbox": [ + 0.218, + 0.522, + 0.785, + 0.563 + ], + "angle": 0, + "content": "95. Zhang, S., Wang, J., Zhang, Y., Zhao, K., Yuan, H., Qing, Z., Wang, X., Zhao, D., Zhou, J.: I2vgen-xl: High-quality image-to-video synthesis via cascaded diffusion models. arXiv preprint arXiv:2311.04145 (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.565, + 0.785, + 0.591 + ], + "angle": 0, + "content": "96. Zhao, J., Zhang, H.: Thin-plate spline motion model for image animation. In: CVPR (2022)" + }, + { + "type": "list", + "bbox": [ + 0.217, + 0.147, + 0.785, + 0.591 + ], + "angle": 0, + "content": null + } + ] +] \ No newline at end of file diff --git a/2024/PhysGen_ Rigid-Body Physics-Grounded Image-to-Video Generation/63954133-cd49-423f-97fe-3962651db98c_origin.pdf b/2024/PhysGen_ Rigid-Body Physics-Grounded Image-to-Video Generation/63954133-cd49-423f-97fe-3962651db98c_origin.pdf new file mode 100644 index 0000000000000000000000000000000000000000..0a1338b333bcdde39496cdc16a37f7e0981b5fd3 --- /dev/null +++ b/2024/PhysGen_ Rigid-Body Physics-Grounded Image-to-Video Generation/63954133-cd49-423f-97fe-3962651db98c_origin.pdf @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:a2b90e3828bd250e1ac7bcfcd3fe3da31ae6a7e64974e0a70468934016ca07b9 +size 3440777 diff --git a/2024/PhysGen_ Rigid-Body Physics-Grounded Image-to-Video Generation/full.md b/2024/PhysGen_ Rigid-Body Physics-Grounded Image-to-Video Generation/full.md new file mode 100644 index 0000000000000000000000000000000000000000..30d4966ab9d317883379a01fd743f0cd1b837abd --- /dev/null +++ b/2024/PhysGen_ Rigid-Body Physics-Grounded Image-to-Video Generation/full.md @@ -0,0 +1,336 @@ +# PhysGen: Rigid-Body Physics-Grounded Image-to-Video Generation + +Shaowei Liu, Zhongzheng Ren, Saurabh Gupta*, and Shenlong Wang* + +University of Illinois Urbana-Champaign https://stevenlsw.github.io/physgen/ + +Abstract. We present PhysGen, a novel image-to-video generation method that converts a single image and an input condition (e.g., force and torque applied to an object in the image) to produce a realistic, physically plausible, and temporally consistent video. Our key insight is to integrate model-based physical simulation with a data-driven video generation process, enabling plausible image-space dynamics. At the heart of our system are three core components: (i) an image understanding module that effectively captures the geometry, materials, and physical parameters of the image; (ii) an image-space dynamics simulation model that utilizes rigid-body physics and inferred parameters to simulate realistic behaviors; and (iii) an image-based rendering and refinement module that leverages generative video diffusion to produce realistic video footage featuring the simulated motion. The resulting videos are realistic in both physics and appearance and are even precisely controllable, showcasing superior results over existing data-driven image-to-video generation works through quantitative comparison and comprehensive user study. PhysGen's resulting videos can be used for various downstream applications, such as turning an image into a realistic animation or allowing users to interact with the image and create various dynamics. + +# 1 Introduction + +Looking at the images in Fig. 1, we can effortlessly predict or visualize the potential outcomes of various physical effects applied to the car, the curling stone, and the stack of dominos. This understanding of physics empowers our imagination of the counterfactual: we can mentally imagine various consequences of an action from an image without experiencing them. We aim to provide computers with similar capabilities - understanding and simulating physics from a single image and creating realistic animations. We expect the resulting video to produce realistic animations with physically plausible dynamics and interactions. + +Tremendous progress has been made in generating realistic and physically plausible video footage. Conventional graphics leverage model-based dynamics to simulate plausible motions and utilize a graphics renderer to simulate images. Such methods provide realistic and controllable dynamics. However, both the dynamics of physics and lighting physics are predefined, making it hard to achieve + +![](images/1ffc6d1de36ee10d710eb7d62fbced194ca15a6efa6ad2e5bbe333392789cf6a.jpg) +Input +Fig. 1: Given a single image, training-free PhysGen generates future frames controlled by physics and initial state. Semantics, geometry and dynamics are reasoned during generation and the result videos are both physics-grounded and photo-realistic. + +our goal of animating an image captured in the real world. On the other hand, data-driven image-to-video (I2V) generation techniques [6, 8, 57] learn to create realistic videos from a single image through training diffusion models over internet-scale data. However, despite advancements in video generative models [6, 8, 57], the incorporation of real-world physics principles into the video generation process remains largely unexplored and unsolved. Consequently, the synthesized videos often lack temporal coherence and fail to replicate authentic object motions observed in reality. Furthermore, such text-driven methods also lack fine-grained controllability. For instance, they cannot simulate the consequences of different forces and torques applied to an object. + +In light of this gap, we propose a paradigm shift in video generation through the introduction of model-based video generative models. In contrast to existing purely generative methods, which are trained in a data-driven manner and rely on diffusion models to learn image space dynamics via scaling laws, we propose grounding object dynamics explicitly using rigid body physics, thereby integrating fundamental physical principles into the generation process. As classical mechanics theory reveals, the motion of an object is determined by its physical properties (e.g., mass, elasticity, surface roughness) and external factors (e.g., external forces, environmental conditions, boundary conditions). + +We tackle our image-to-video generation problem via a computational framework, PhysGen. PhysGen consists of three stages: 1) image-based physics understanding; 2) model-based dynamics simulation; and 3) generative video rendering with dynamics guidance. Our method first infers object compositions and physical parameters from the input image through large visual foundation model-based reasoning (Sec. 3.1). Given an input force, we then utilize the in + +ferred physical parameters to perform realistic dynamic simulations for each object, accounting for their rigid body dynamics, collisions, frictions, and elasticity (Sec. 3.2). Finally, guided by motion and the input image, we present a novel generative video diffusion-based renderer that outputs the final realistic video footage (Sec. 3.3). Several illustrative examples of our framework are shown in Fig. 1. As shown in this figure, our approach combines the best of the worlds of data-driven generative modeling's appearance realism and model-based dynamics's verified physical plausibility. Notably, our generation process is highly controllable, allowing users to specify underlying physics parameters and initial conditions. This capability facilitates a range of interactive applications. Moreover, our proposed generation pipeline operates solely during inference time, eliminating the need for any training. + +We evaluate PhysGen generation ability on multiple data source including the web and self-captured images. We compare against both state-of-the-art image-to-video models [18, 87, 95] as well as image editing method [28] quantitatively and qualitatively. We also perform user-study to evaluate the physical-realism and photo-realism of the generated video. Our method demonstrates physics-informed results for controllable image-to-video generation, producing realistic video sequences with grounded rigid-body dynamics and interaction. PhysGen combines learning-based generative approaches with traditional model-based physics to deliver visually appealing results without any training. Our approach harnesses a plethora of ingredients from previous works, particularly recent progress in large foundational visual models for segmentation [40, 48, 61, 93], physical understanding [2, 56], normal estimation [23, 26], relighting, and generative rendering [8, 18]. While these individual ingredients exist, our system combines them in a novel manner to enable an exciting new capacity, featuring physical plausibility for video generation at an unprecedented level. + +Our contributions can be summarized as follows: 1) We propose a novel image-space dynamics model that can understand physical parameters from a single image and simulate realistic, controllable, and interactive dynamics. 2) We present a novel image-to-video system by integrating our presented dynamics model with generative diffusion-based video priors. The model is training-free, and the resulting videos are highly realistic, physically plausible, and controllable, consistently surpassing existing state-of-the-art video generative models. + +# 2 Related work + +Image-based physics reasoning and simulation. Our image-based dynamics model is built upon physical simulation. Symbolic physics simulators or physics engines have been extensively studied for various physical processes, including rigid body dynamics, fluid simulation, and deformable objects [36-38, 41, 49, 55, 72]. Such methods have been widely used across graphics, scientific discovery, robotics, and video effects. However, these rule-based or solver-based simulators face limitations in terms of expressiveness, efficiency, generalizability, and parameter tuning. To overcome these challenges, neural physics, which combines + +traditional physics simulators/engines with deep neural networks, has been developed to model dynamic processes such as planar pushing and bouncing [4], as well as object interactions [3]. Domain-specific network architectures [7,45,54,63] have been developed to address various physical problems. Additionally, the integration of neural scene representations or large language models with physics simulators [43,50,86] has also been investigated recently. Nevertheless, most forward physics models, whether neural or not, have not resolved the dependency on pre-defined physical parameters, which are often specified by users. + +Reasoning physical parameters from visual data allows us to simulate without manually defined physics, further enhancing convenience and realism for image-based physics simulation. To enable this, subsequent research efforts [21, 43, 44, 80-82, 84, 88] have introduced pipelines that first extract scene representations of the physical world from images or videos using neural networks, and then utilize these representations for either physics reasoning or simulation. Despite the promise of these approaches, they primarily rely on synthetic data for training, which might suffer from generalization issues, or on inverse physics, which requires inference from observed physical phenomena, often requiring video observations instead of a single image. In our work, we aim to generate physics-grounded, high-fidelity predictive videos from static input images of real-world objects with complex backgrounds. Contrary to existing work, we investigate leveraging the power of large pretrained visual foundational models [2, 40, 56] for physics reasoning in a zero-shot manner. + +Video generative model. Video generative models have experienced remarkable advancements in recent years [6,8,12,18,27,29,31,33,42,57,65,74,76,87,92, 95], with the state-of-the-art framework [57] achieving the capability to generate photo-realistic and coherent imaginative videos from text instructions using diffusion models [34,58,62,66,68]. Despite progress, challenges remain, such as the need for extensive training data, object permanence, realistic lighting, and reducing unrealistic motion and physics violations. Overcoming these is key for advancing video generative modeling and its applications, especially in physics-accurate areas like scientific discovery and robotics. Our work addresses some issues by proposing a novel image-to-video framework that merges a model-based approach with a pretrained video prior, grounding the generated content with real-world physics. + +Image animation. Our method is heavily inspired by image-based animation [19, 22, 35, 64, 70, 77, 89] and cinematographs. One common way to improve quality is to incorporate class-specific heuristics [19, 39], which can be physically simulated and integrated into the generation process. Recent advancements in deep learning have led to data-driven solutions in this domain, where temporal neural networks are trained on video datasets to directly predict consecutive video frames from an input image [10, 17, 24, 30, 35, 75]. Moreover, there is growing interest in interactive [9, 11, 46] and controllable [1, 22] image-to-video synthesis. To further improve temporal coherence, more priors such as motion fields [24, 35, 51, 52, 69, 89] or flows [14, 28, 96], 3D geometry information [60, 79], + +![](images/6e8d796f456560f66846d4cb4bd2805cf5c761bfe7fdda1da505e68b384f37f9.jpg) +Fig. 2: Method overview. Our framework consists of three interleaved components: the perception module, the dynamics simulation module, and the rendering module. The perception module interprets the semantics, geometry, and physical parameters in the given image. The dynamics simulation module simulates the rigid-body motion and interactions of each instance in the scene, governed by Newton's Laws and physical constraints. The rendering module renders the final outcome, leveraging an off-the-shelf relighting model and diffusion-model-based video priors. + +and user annotations [46] have been subsequently introduced. In this work, we propose a model-based solution and tackle the image animation task from a new perspective, where our method takes a single image and user-specific initial force or torque as input, and models object dynamics explicitly via rigid physics simulation. Our key insight is that physical parameters can be inferred from a single image automatically through large foundation models such as GPT-4V [2, 56]. + +# 3 Approach + +Given a single input image, our goal is to generate a realistic $T$ -frame video featuring rigid body dynamics and interactions. Our generation can be conditioned on user-specific inputs in the form of an initial force $\mathbf{F}_0$ and/or torque $\pmb{\tau}_{0}$ . Our key insight is to incorporate physics-informed simulation into the generation process, ensuring the dynamics are realistic and interactive, thus producing large, plausible object motion. To achieve this, we present a novel framework consisting of three stages: physical understanding, dynamics generation, and generative rendering. The perception module aims to reason about the semantics, collision geometry, and physics of the objects presented in the image (Sec. 3.1). The dynamics module then leverages the input force/torque, as well as the inferred shape and geometry, to simulate the rigid-body motions of the objects, as well as object-object and object-scene interactions, such as friction and collisions (Sec. 3.2). Finally, we convert the simulated dynamics into pixel-level motion and combine image-based warping and generative video to render the final video (Sec. 3.3). Fig. 2 depicts the overall framework. + +# 3.1 Perception + +Simulating dynamics on an image requires a holistic understanding of object compositions, materials, geometry, and physics. Toward this goal, we designed our perception module to infer such properties from a single image. Our key insight is to harness readily available, large pretrained models [2, 15, 26, 56, 61] to achieve this goal, as shown in Fig. 2. + +Segmentation. Identifying and segmenting each individual physical entity lays the foundation for image-based dynamics. Inspired by the recent success of large pretrained segmentation models, we incorporate GPT-4V [2, 56] to recognize all image categories and send to Grounded-SAM [61] to detect and segment each individual instance-level objects $\{\mathbf{o}^i\in \mathbb{R}^{W\times H}\}_{i}^{N}$ , where $\mathbf{o}^i$ is the binary mask for $i$ -th object. We also query the GPT-4V to classify the objects into foreground and background based on their movability. The foreground objects are send to the physical simulation once user inputs are applied. We extract collision boundaries and supporting edges (e.g., ground, walls, etc.) from the background objects. The details are presented in the supplementary. + +Physical properties reasoning. Realistic physics depends on accurate physical parameters, such as surface friction, mass, and elasticity. Unlike prior physics-based image dynamics work [19], our work leverages visual foundational models to reason physical properties directly. Our solution is simple yet effective. Inspired by the success of mask-based prompting [2,56,90], we directly ask GPT-4V [56] for certain physical properties, providing an object mask overlaid on the input image. Following [85,94], we send the crafted prompt to GPT-4V, which contains GPT-4V instructions to return a quantitative measure of each queried property in a metric unit. For each object, we query its mass, elasticity, and friction coefficient $(M_i,E_i,\mu_i)$ . We use the Coulomb friction model [59] for friction modeling. The elasticity coefficient is a scalar, where 0.0 results in no bounce, and 1.0 results in a perfect bounce. Given each object mass $M_{i}$ in grams and its shape primitive, we compute the rotational inertia $\mathbf{I}_i$ accordingly. + +Geometry primitives. Rasterized instance masks are not suitable for physical simulation; hence, we need to convert each object into vectorized shape primitives. We fit two types of primitive shapes: a circle and generic polygons. For each object, we choose the primitive type with the maximum coverage of its segmentation. For circles, we fit the center and radius to cover the segmentation mask. Circles allow us to realistically simulate rolling motions. For non-circle objects, we instead fit the generic polygons. Specifically, we perform contour extraction on the segmentation masks to obtain the polygon vertices. + +Intrinsic decomposition. The dynamic movement of the objects will also result in comprehensive changes in shading, as their position wrt the lighting environment changes. To compensate for this effect during rendering, we must perform image decomposition to infer albedo, normal as well as lighting from + +the single image for each object $(\mathbf{A}_0^i,\mathbf{N}_0^i,\mathbf{L}_0)$ . To achieve this we leverage off-the-shelf intrinsic decomposition model [15] to compute $\mathbf{A}_0^i$ and $\mathbf{L}_0$ and exploits pretrained surface normal estimator for the computation of object normal $\mathbf{N}_0^i$ . + +Background inpainting. Once the foreground objects move, they will leave holes in the background. To produce realistic and complete video footage, we leverage an off-the-shelf generative image inpainting model [93] to recover the complete background scene without foreground objects. The inpainted image $\mathbf{B} \in \mathbb{R}^{W \times H \times 3}$ will be used for composition in the rendering stage. + +# 3.2 Image space dynamics simulation + +Given the foreground objects with physical properties, we use rigid-body physics for dynamics simulation in image spaces. We choose to perform simulations in image spaces for three reasons: 1) image-space dynamics are better coupled with our output video; 2) a full 3D simulation requires complete 3D scene reconstruction and understanding, which remains an unsolved problem from a single image; 3) image-space dynamics can already cover various object dynamics and have been widely used in prior work [4, 19, 35, 46, 80]. Specifically, at a given time $t$ , each rigid object $i$ is characterized by its 2D pose and velocity at its center of mass. The position includes a translation $\mathbf{t}^i(t) \in \mathbb{R}^2$ and a rotation $\mathbf{R}^i(t) \in \mathbb{S}\mathbb{O}(2)$ specified in a world coordinate. The velocity includes linear velocity $\pmb{\nu}^i(t) \in \mathbb{R}^2$ and angular velocity $\pmb{\omega}^i(t) \in \mathbb{R}^2$ . Hence, the state of the each object $i$ at time $t$ can be represented as $\mathbf{q}^i(t) = [\mathbf{t}^i(t), \mathbf{R}^i(t), \pmb{\nu}^i(t), \pmb{\omega}^i(t)]$ . + +Following [25], the rigid body motion dynamics is given by Eq. (1) for each object. For simplicity, we omit the object index $i$ in the following: + +$$ +\frac {d}{d t} \mathbf {q} (t) = \frac {d}{d t} \left[ \begin{array}{l} \mathbf {t} (t) \\ \mathbf {R} (t) \\ \boldsymbol {\nu} (t) \\ \boldsymbol {\omega} (t) \end{array} \right] = \left[ \begin{array}{c} \mathbf {v} (t) \\ \boldsymbol {\omega} (t) \times \mathbf {R} (t) \\ \frac {\mathbf {F} (t)}{M} \\ \mathbf {I} (t) ^ {- 1} (\boldsymbol {\tau} - \boldsymbol {\omega} (t) \times \mathbf {I} (t) \boldsymbol {\omega} (t)) \end{array} \right] \tag {1} +$$ + +where $\mathbf{F}$ is the force, $\pmb{\tau}$ is the torque, $M$ is the mass of the object, $\mathbf{I}(t) = \mathbf{R}(t)\mathbf{I}\mathbf{R}(t)^T$ is the rotation inertia in world coordinates, $\mathbf{I} \in \mathbb{R}^{2 \times 2}$ is the inertia matrix in its body coordinates. This inertia indicates its resistance against rotational motion. The state $\mathbf{q}(t)$ is given by integrating Eq. (1): + +$$ +\mathbf {q} (t) = \mathbf {q} (0) + \int_ {0} ^ {t} \frac {d}{d t} \mathbf {q} (t) d t = \mathbf {q} (0) + \sum_ {i = 1} ^ {T} \frac {d}{d t} \mathbf {q} (t) | _ {t = t _ {i}} \Delta_ {t} \tag {2} +$$ + +where $\mathbf{q}(0)$ is the initial condition specified in the input image. We compute the integral using numerical ODE integrations, e.g. Euler method [20]. Given the initial state of the objects, the physical motion simulation module synthesis each foreground object future motion. The location of each object is updated by the affine transformation $\mathbf{T}(t) = [\mathbf{R}(t),\mathbf{t}(t)]$ from the initial location. + +![](images/36ea1b241fe3a56297e00bc470356239e8a695cb84b3f184cf954e3b8380d2b6.jpg) +Input + +![](images/37c392b76552cd52195b85e02498cf7e45a5809b6650bed56d0f4ac515e1c776.jpg) +Composed (Sec. 3.3) +Fig. 3: Rendered video comparison. We show a toy example of piggy bank. The left shows the input frame, and the rest 3 are future frame generations. The composited piggy bank hasn't been aware of the light change. The relighted output synthesized no shadows beneath. The rendered output from diffusion model is most photo realistic. + +![](images/313a06508e7cc87561a7dc9dc12beab66d4812a8c921ade0b5e94f845501af37.jpg) +Relight (Sec. 3.3) + +![](images/574ab8361d0e985d1af4576d717f32c5fb8f8bfdef45407afb47f7810c62980b.jpg) +Render (Sec. 3.4) + +External force and torque. For each object, the external forces $\mathbf{F}(t)$ and torques $\pmb{\tau}$ include gravity, friction, and elasticity between the object's surface and the environment. Specifically, we consider the initial external force and torque from user input, as well as gravity, rolling friction, and sliding friction. + +Collision. When objects move, they might collide with each other or with the scene boundary (e.g., hitting a wall). Therefore, collision checking is necessary at every time step. A collision will result in offset forces being applied from its center of mass, producing $\tau$ and causing the object to rotate. In collision reactions, changes in energy and momentum are minimized to adhere to Newton's conservation laws. In the case of an object falling, it could bounce back or start rolling on the ground due to the elasticity of both the object and the ground. + +# 3.3 Rendering + +Given the simulated motion sequence $\{\mathbf{q}^i(0),\dots,\mathbf{q}^i(t),\dots\}_i^N$ and the input image $\mathbf{C}$ , the rendering module outputs the rendered video $\{\dots,\mathbf{C}(t),\dots\}$ . There are a few desiderata for generating realistic video: it needs to be temporally consistent, complete, reflect the lighting changes as the object moves, and accurately represent the simulated image space dynamics. To achieve this goal, we design a novel motion-guided rendering pipeline consisting of an image-based transformation and composition module to ensure motion awareness, a relighting module to ensure the plausibility of lighting physics and a generative video editing module that retouches the composed and relit video to ensure realism. + +Composed video. Given the object state from the physical motion module, we compose an initial video by alpha-blending the foreground scene with the static inpainted background from Sec. 3.1. The foreground scene is rendered by performing forward-warping from the input image to future frames using the affine transformation $\mathbf{T}(t)$ . The alpha channel is computed using the same procedure, with the input being the segmentation mask in Sec. 3.1: + +$$ +\hat {\mathbf {X}} (t) = \operatorname {c o m p o s i t e} (\mathbf {B}, \operatorname {w a r p} \left(\mathbf {X} ^ {0}, \mathbf {T} ^ {0} (t)\right) \dots \operatorname {w a r p} \left(\mathbf {X} ^ {i}, \mathbf {T} ^ {i} (t)\right) \dots \dots) +$$ + +where $\mathbf{X}^i$ are segmented input image of $i$ -th object. Apart from the RGB sequence $\hat{\mathbf{V}} = \{\dots \hat{\mathbf{X}}(t) \dots\}$ , we also use the same affine warping and composition to compute the blended albedo map $\hat{\mathbf{A}}(t)$ , which is later used for relighting. + +Relighting. Our relighting module simulates the changes in shading due to object movement. Specifically, it takes as input the transformed albedo $\hat{\mathbf{A}}(t)$ and RGB image $\hat{\mathbf{X}}(t)$ at each time $t$ , as well as the estimated lighting $\mathbf{L}$ . We first compute the surface normal $\hat{\mathbf{N}}(t)$ from the composited RGB image $\hat{\mathbf{X}}(t)$ using an off-the-shelf normal estimator [23]. We then perform reshading using the parametric illumination model [15]: $\tilde{\mathbf{X}}(t) = f(\hat{\mathbf{X}}(t), \hat{\mathbf{A}}(t), \hat{\mathbf{N}}(t), \mathbf{L})$ , which returns the relit image $\hat{\mathbf{X}}(t)$ for each frame $t$ . + +# 3.4 Generative refinement with latent diffusion models + +One limitation of the single-image based relighting model is that it neglects the background and cannot handle complicated lighting effects due to complex object-object interactions, such as ambient occlusion and cast shadows. Moreover, the composition results in unnatural boundaries. To address these issues and enhance realism, we incorporated a diffusion-based video to refine the relit video, denoted as $\tilde{\mathbf{V}} = \{\ldots ,\tilde{\mathbf{X}} (t),\ldots \}$ , and obtain the final video output. + +Specifically, we incorporate a pretrained video diffusion model to refine our video $\tilde{\mathbf{V}}$ . We first use the pretrained latent diffusion encoder [62] to encode the guidance video $\tilde{\mathbf{V}}$ into a latent code $\mathbf{z}$ . Inspired by SDEdit [53], we add noise to the guided latent code and gradually denoise it using the pretrained video diffusion. Given that the content in the guided latent code is already satisfactory, we do not perform the denoising process from scratch. Instead, we define an initial noise strength $s$ where $s \in [0,1]$ controls the amount of noise added to $\mathbf{z}$ . In practice, we find that a certain approach finds a good trade-off between fidelity and realism. At each denoising step, we ensure the foreground objects are as consistent as possible with the guidance (copied from the guided latent code) while synthesizing new content (e.g., shadows) in the background (using the denoised latent code). To this end, we use a fusion weight $w$ to control how much of the generated latent to inject into the foreground; the fusion weight is gradually increased during the denoising as the perturbation of noise decreases. We also define a fusion timestamp $\delta$ as the signal to stop the fusion. The final denoised latent code $\mathbf{z}_{*}$ is then decoded to our final video output $\mathbf{V}$ . Please see the supplementary material for a detailed algorithm. + +Fig. 3 depicts the composed video $\hat{\mathbf{V}}$ , the relit video $\tilde{\mathbf{V}}$ , and our final output $\mathbf{V}$ , illustrating the effectiveness of each rendering component. + +# 4 Experiments + +Implementation details. For physical body simulation, we use a 2D rigid body physics simulator Pymunk [13]. For each experiment, we run 120 simulation steps and uniformly sample 16 frames to form the videos. We set the + +![](images/cb2e40043b69b091db787a08619c2f98e58996c50081e24ce8cae7a305a2026a.jpg) +Fig. 4: Video generation results. Left: Input initial frame with red arrows representing the initial force or speed on the object. Right: Generated future frames. + +generation resolution at $512 \times 512$ so that frames can easily fit into the diffusion models. For video diffusion prior model, we use SEINE [18]. For inference, DDIM sampling [67] is used and the noise strength $i$ set to $s = 0.5$ , i.e., executing 25/50 steps for denoising. The latents fusion stops at timestamp $\delta = 5$ . + +Speed. PhysGen's run-time speed is 3 minutes for one image-to-video gen on a Nvidia A40 GPU. Image understanding and relighting takes the most time. + +Data. To show the generalizability and robustness of our method, we use both internet data and self-captured indoor images from a cell phone. The dataset is diverse, with variations in lighting, object count, geometries, physical attributes, and environmental boundaries. We compare different approaches on 15 photos that require physical reasoning. Additionally, we collect 50 recorded videos of a given scene as ground truth by varying input conditions for quantitative comparison. Randomly selected sequences are shown in the supplementary material. + +Baselines. We compare against two streams of approaches. The first stream is current state-of-the-art I2V models: SEINE [18], I2VGen-XL [95] and Dynam- + +
MethodsPhysical-realism↑Photo-realism↑
SEINE [18]1.391.86
DynamiCrafter [87]1.681.81
I2VGen-XL [95]2.112.25
Ours4.143.86
+ +Table 1: Human evaluation. We evaluate both physical-realism and photo-realism of the three competing image-to-video models. The user is asked to evaluate the generated video using a 5-point scale from strongly disagree (1) to strongly agree(5) that the video is physical-realistic and photo-realistic. Our method ranks the first in both terms. + +
MethodsImage-FID↓Motion-FID↓
SEINE [18]138.8938.76
DynamiCrafter [87]99.64109.61
I2VGen-XL [95]104.6282.04
Ours105.7030.20
+ +Table 2: Quantitative evaluation. We measure the Image-FID and Motion-FID of videos generated by four methods against the collected GT videos for the given scene. Our method achieves both low Image-FID and Motion-FID. + +iCrafter [87]. These methods take both an image and a text prompt as input for generation. The text prompt is generated by GPT-4V [56] to describe the future events. Another stream we compare is the image-based manipulation approach Motion Guidance [28]. This method uses an image and optical flow as inputs to predict future movement and generate corresponding images. We employ computed motion flow from our physical motion module as the target input for Motion Guidance to optimize each video frame. + +# 4.1 Results + +We show our generated results on 8 different physical procedures from multiple sources in Fig. 4. Our method could simulate reasonable complex physical procedures which involves object-object interaction, object-scene interaction with different physical properties and initial conditions. + +Visual comparisons. The comparison results are shown in Fig. 5. As can be seen, image-video generation methods could not produce any physically plausible futures. For each method, even we carefully tune the input prompt, the models couldn't understand the direction, physical meanings and interactions. All 3 models fail to synthesize realistic videos. For motion guidance with accurate motion flow provided, the results is better but brings a lot inconsistency. When there are multiple objects interaction, the method suffers from introducing new contents. For the domino case, the generation is not smooth and contain artifacts as it could not accurately model the physical procedure. + +![](images/86f9d79b7cd6780ddd6e334eefc6976736fa3d27fbfb61d66444fb372b94826a.jpg) +Fig.5: Qualitative comparison. against DynamiciCafter(DC) [87], I2VGenXL(VGen) [95], SEINE [18], Motion Guidance(MG) [28]. + +![](images/75290295add3e73aeb644eb061bdea4d415f0c229eaccfb873e5d6f42931b8e2.jpg) +Input + +![](images/0ada98210e6442fdb327fc985cd08265f97727fa9f72c81e59fc8a07354ba93f.jpg) +Fig. 6: Controllability. PhysGen generates diverse results reflecting initial forces. + +![](images/bb0941cae9800451da6ab728f2ac94f5d0b17b8c65dd8a81f9a082ace7a692af.jpg) + +![](images/fece2730527167282bfab2ae3f6c4009f48e92632b0f3064b62b7e511991bda8.jpg) +FG Seg + +![](images/d5729ad87430b019747f7872a6a90ad394ee13475c6e17744e039b1c45073683.jpg) +Generation (left $\rightarrow$ right: time steps) + +![](images/eecd168f4f0ae2d75c09dd900b64a0669d1ce990e352a2f3f7ac60bbe09e2984.jpg) + +![](images/350894b1c5b746a2627a025fba7843bb188198ae686cfe046087138af8db936d.jpg) + +![](images/76a57d3144cd17a6649d739b23543b32212590e7c1cba0ba8e0057184f9270d2.jpg) + +![](images/85d759259bf531df7be2066b72aba3fd3c2516399599fd455ff1a8912585c0c5.jpg) + +![](images/c7133f9ef0c8294c7ee20ac3488e9463d06447bc5630dd661ff692df556b28c2.jpg) + +![](images/bfa0631fb8f6a922ab56424f6a2056b93b29da966aa546d2fee99b1634ac0e03.jpg) + +![](images/7fe317890fe4aa7a590abb8a2ebb9997b74dfba92fd73df248cec690370f7926.jpg) + +![](images/97f771c717e7a716e883b0480a19f8c40ff0cd23b88dd83aab83eb23b47e53d0.jpg) +Input +Fig. 7: Qualitative results from the perception module. The results include foreground and background mask, inpainting output, normal map and inferred physics. + +![](images/21ff867b20ac2c770e2e4a9292597255d91a9082e96d6e8c83c3f52b8ce4f8f7.jpg) + +![](images/9770c9b00fde4df2fe55bff534a1a120add4a9282721b8f57215f3b116d9fd60.jpg) +BG Seg + +![](images/36d369f2ffd32d1ed180a681022124479dad95114b84f666b395b79e3a4f1802.jpg) +Inpainting + +![](images/f8dce019dcc5a58c1eb711628fa0e9f7c0f9db3f935dd834c77cab1f759d1eb7.jpg) +Normal + +![](images/13f7e5e97e3cc966293330450a174d2c7a586dcc4baec7e322f96e6e1f7ddde2.jpg) +inferred Physics + +# 4.2 Human evaluation + +To comprehensively assess both the physical-realism and photo-realism of the generated videos, we conducted human evaluations following a similar methodology to prior work [16,78,83,91]. Here by physical-realism we mean whether the content of the generated video is realistic in its dynamics and object interactions and accurately reflects the instructed movements. + +Specifically, we include 15 videos with diverse objects and conditions for evaluation, compared against other three I2V model generations (SEINE [18], DynamiCrafter [87], I2VGen-XL [95]). For each compared baseline method, we meticulously adjusted the input prompts and conducted multiple runs to select the most plausible output. In contrast, our method did not undergo specific tuning for each video. We randomized the presentation order and asked 14 participants to rate the videos on a five-point scale, from strongly disagree (1) to strongly agree (5) for physical-realism and photo-realism. The results in Tab. 1 show that our method achieves the highest ratings. Our average scores for physical-realism (4.14) and photo-realism (3.86) fall within the Agree level, while other methods perform poorly. + +# 4.3 Quantitative evaluation + +We quantitatively evaluate the collected GT videos of a given scene. The GT data includes 833 images. For each method, we generate 50 videos with random + +samples, collecting 800 images per method. Our method varies initial conditions to ensure all generated videos are different. Following [12], we adopt the Fréchet Inception Distance (FID) [32] as evaluation metric. We didn't use Fréchet Video Distance (FVD) [73] due to the small sample size and the goal to disentangle evaluation of appearance and motion. We evaluate appearance using Image-FID and motion using Motion-FID. For Motion-FID, we use RAFT [71] to extract and colorize optical flow [5], then compute FID from these rendered images. The results are shown in Tab. 2. We notice other methods which either generate minimal motion, leading to low Image-FID and high Motion-FID (e.g., DynamiCrafter [87]), or produce realistic motion but fail to maintain image consistency, resulting in low Motion-FID and high Image-FID (e.g., SEINE [18]). Our approach achieves both low Image-FID and Motion-FID compared to others. + +# 4.4 Analysis + +Perception evaluation. We evaluate the perception module using 10 complex open-world images with 118 annotated movable instances (COCO [47] format). Our system achieves 0.93 precision, 0.82 recall at 0.5 IoU. More details are shown in the supplementary. Fig. 7 shows intermediate outputs from our perception module and simulation module. + +Physical reasoning. We compare the physical parameters estimated by GPT-4V to random values sampled from human-estimated ranges. Without GPT-4V, Image-FID increases from 105.70 to 111.01, Motion-FID from 30.20 to 36.60. It shows the physical reasoning are crucial for appearance and motion realism. + +Controllability. We showcase the controllability and diversity of our method in Fig. 6 by varying the initial conditions. For different rows, we adjust the direction and magnitude of initial speed and forces. The results show diverse physical-plausible trajectories. This demonstrates the potential of applying our method for diverse physical world generations. + +Limitations. Our method is an initial step towards physics-based video generation but has two main limitations: it focuses mainly on rigid objects, limiting its application to non-rigid ones, and it lacks a comprehensive 3D understanding, making it unable to handle out-of-plane motions. We leave leveraging deformable physics and full 3D understanding as future work. + +# 5 Conclusion + +We present PhysGen, a novel image-to-video generation method that turns a static image into a high-fidelity, physics-grounded, and temporally coherent videos. The main module is a model-based physical simulator with a data-driven video generation process. PhysGen enables a series of controllable and interactive downstream applications. PhysGen provide a new image-to-video generation paradigm and hopefully can inspire more follow-up works. + +# Acknowledgements + +This project is supported by NSF Awards #2331878, #2340254, and #2312102, the IBM IIDAI Grant, and an Intel Research Gift. We greatly appreciate the NCSA for providing computing resources. We thank Tianhang cheng for helpful discussions. We thank Emily Chen and Gloria Wang for proofreading. + +# References + +1. Aberman, K., Weng, Y., Lischinski, D., Cohen-Or, D., Chen, B.: Unpaired motion style transfer from video to animation. ACM TOG (2020) +2. Achiam, J., Adler, S., Agarwal, S., Ahmad, L., Akkaya, I., Aleman, F.L., Almeida, D., Altenschmidt, J., Altman, S., Anadkat, S., et al.: Gpt-4 technical report. arXiv preprint arXiv:2303.08774 (2023) +3. Ajay, A., Bauza, M., Wu, J., Fazeli, N., Tenenbaum, J.B., Rodriguez, A., Kaelbling, L.P.: Combining Physical Simulators and Object-Based Networks for Control. In: ICRA (2019) +4. Ajay, A., Wu, J., Fazeli, N., Bauza, M., Kaelbling, L.P., Tenenbaum, J.B., Rodriguez, A.: Augmenting Physical Simulators with Stochastic Neural Networks: Case Study of Planar Pushing and Bouncing. In: IROS (2018) +5. Baker, S., Scharstein, D., Lewis, J.P., Roth, S., Black, M.J., Szeliski, R.: A database and evaluation methodology for optical flow. IJCV (2011) +6. Bar-Tal, O., Chefer, H., Tov, O., Herrmann, C., Paiss, R., Zada, S., Ephrat, A., Hur, J., Li, Y., Michaeli, T., et al.: Lumiere: A space-time diffusion model for video generation. arXiv preprint arXiv:2401.12945 (2024) +7. Battaglia, P., Pascanu, R., Lai, M., Jimenez Rezende, D., et al.: Interaction networks for learning about objects, relations and physics. NeurIPS (2016) +8. Blattmann, A., Dockhorn, T., Kulal, S., Mendelevitch, D., Kilian, M., Lorenz, D., Levi, Y., English, Z., Voleti, V., Letts, A., et al.: Stable video diffusion: Scaling latent video diffusion models to large datasets. arXiv preprint arXiv:2311.15127 (2023) +9. Blattmann, A., Milbich, T., Dorkenwald, M., Ommer, B.: ipoke: Poking a still image for controlled stochastic video synthesis. In: ICCV (2021) +10. Blattmann, A., Milbich, T., Dorkenwald, M., Ommer, B.: Understanding object dynamics for interactive image-to-video synthesis. In: CVPR (2021) +1. Blattmann, A., Milbich, T., Dorkenwald, M., Ommer, B.: Understanding object dynamics for interactive image-to-video synthesis. In: CVPR (2021) +2. Blattmann, A., Rombach, R., Ling, H., Dockhorn, T., Kim, S.W., Fidler, S., Kreis, K.: Align your latents: High-resolution video synthesis with latent diffusion models. In: CVPR (2023) +3. Blomqvist, V.: Pymunk. https://pymunk.org (2023) +4. Bowen, R.S., Tucker, R., Zabih, R., Snavely, N.: Dimensions of motion: Monocular prediction through flow subspaces. In: 3DV (2022) +5. Careaga, C., Miangoleh, S.M.H., Aksoy, Y.: Intrinsic harmonization for illumination-aware compositing. arXiv preprint arXiv:2312.03698 (2023) +6. Chen, H., Xia, M., He, Y., Zhang, Y., Cun, X., Yang, S., Xing, J., Liu, Y., Chen, Q., Wang, X., et al.: Videocrafter1: Open diffusion models for high-quality video generation. arXiv preprint arXiv:2310.19512 (2023) +7. Chen, X., Liu, Z., Chen, M., Feng, Y., Liu, Y., Shen, Y., Zhao, H.: Livephoto: Real image animation with text-guided motion control. arXiv preprint arXiv:2312.02928 (2023) +8. Chen, X., Wang, Y., Zhang, L., Zhuang, S., Ma, X., Yu, J., Wang, Y., Lin, D., Qiao, Y., Liu, Z.: Seine: Short-to-long video diffusion model for generative transition and prediction. arXiv preprint arXiv:2310.20700 (2023) +9. Chuang, Y.Y., Goldman, D.B., Zheng, K.C., Curless, B., Salesin, D.H., Szeliski, R.: Animating pictures with stochastic motion textures. ACM TOG (2005) + +20. Ciarlet, P.G., Lions, J.L.: Handbook of numerical analysis. Gulf Professional Publishing (1990) +21. Davis, A., Bouman, K.L., Chen, J.G., Rubinstein, M., Durand, F., Freeman, W.T.: Visual vibrometry: Estimating material properties from small motion in video. In: CVPR (2015) +22. Davis, A., Chen, J.G., Durand, F.: Image-space modal bases for plausible manipulation of objects in video. ACM TOG (2015) +23. Eftekhar, A., Sax, A., Malik, J., Zamir, A.: Omnidata: A scalable pipeline for making multi-task mid-level vision datasets from 3d scans. In: ICCV (2021) +24. Endo, Y., Kanamori, Y., Kuriyama, S.: Animating landscape: self-supervised learning of decoupled motion and appearance for single-image video synthesis. arXiv preprint arXiv:1910.07192 (2019) +25. Friedland, B.: Control system design: an introduction to state-space methods. Courier Corporation (2012) +26. Fu, X., Yin, W., Hu, M., Wang, K., Ma, Y., Tan, P., Shen, S., Lin, D., Long, X.: Geowizard: Unleashing the diffusion priors for 3d geometry estimation from a single image. arXiv preprint arXiv:2403.12013 (2024) +27. Ge, S., Nah, S., Liu, G., Poon, T., Tao, A., Catanzaro, B., Jacobs, D., Huang, J.B., Liu, M.Y., Balaji, Y.: Preserve your own correlation: A noise prior for video diffusion models. In: ICCV (2023) +28. Geng, D., Owens, A.: Motion guidance: Diffusion-based image editing with differentiable motion estimators. In: ICLR (2024) +29. Girdhar, R., Singh, M., Brown, A., Duval, Q., Azadi, S., Rambhatla, S.S., Shah, A., Yin, X., Parikh, D., Misra, I.: Emu video: Factorizing text-to-video generation by explicit image conditioning. arXiv preprint arXiv:2311.10709 (2023) +30. Guo, Y., Yang, C., Rao, A., Liang, Z., Wang, Y., Qiao, Y., Agrawala, M., Lin, D., Dai, B.: Animatediff: Animate your personalized text-to-image diffusion models without specific tuning (2023) +31. Gupta, A., Yu, L., Sohn, K., Gu, X., Hahn, M., Fei-Fei, L., Essa, I., Jiang, L., Lezama, J.: Photorealistic video generation with diffusion models. arXiv preprint arXiv:2312.06662 (2023) +32. Heusel, M., Ramsauer, H., Unterthiner, T., Nessler, B., Hochreiter, S.: Gans trained by a two time-scale update rule converge to a local nash equilibrium. NeurIPS (2017) +33. Ho, J., Chan, W., Sahara, C., Whang, J., Gao, R., Gritsenko, A., Kingma, D.P., Poole, B., Norouzi, M., Fleet, D.J., et al.: Imagen video: High definition video generation with diffusion models. arXiv preprint arXiv:2210.02303 (2022) +34. Ho, J., Jain, A., Abbeel, P.: Denoising diffusion probabilistic models. NeurIPS (2020) +35. Holynski, A., Curless, B.L., Seitz, S.M., Szeliski, R.: Animating pictures with eulerian motion fields. In: CVPR (2021) +36. Hu, Y., Anderson, L., Li, T.M., Sun, Q., Carr, N., Ragan-Kelley, J., Durand, F.: Difftaichi: Differentiable programming for physical simulation. ICLR (2020) +37. Hu, Y., Fang, Y., Ge, Z., Qu, Z., Zhu, Y., Pradhana, A., Jiang, C.: A moving least squares material point method with displacement discontinuity and two-way rigid body coupling. ACM TOG (2018) +38. Hu, Y., Li, T.M., Anderson, L., Ragan-Kelley, J., Durand, F.: Taichi: a language for high-performance computation on spatially sparse data structures. ACM TOG (2019) +39. Jhou, W.C., Cheng, W.H.: Animating still landscape photographs through cloud motion creation. IEEE Transactions on Multimedia (2015) + +40. Kirillov, A., Mintun, E., Ravi, N., Mao, H., Rolland, C., Gustafson, L., Xiao, T., Whitehead, S., Berg, A.C., Lo, W.Y., Dollár, P., Girshick, R.: Segment anything. arXiv:2304.02643 (2023) +41. Koenig, N., Howard, A.: Design and use paradigms for gazebo, an open-source multi-robot simulator. In: IROS (2004) +42. Kondratyuk, D., Yu, L., Gu, X., Lezama, J., Huang, J., Hornung, R., Adam, H., Akbari, H., Alon, Y., Birodkar, V., et al.: Videopoet: A large language model for zero-shot video generation. arXiv preprint arXiv:2312.14125 (2023) +43. Li, X., Qiao, Y.L., Chen, P.Y., Jatavallabhula, K.M., Lin, M., Jiang, C., Gan, C.: Pac-nerf: Physics augmented continuum neural radiance fields for geometry-agnostic system identification. In: ICLR (2023) +44. Li, Y., Lin, T., Yi, K., Bear, D., Yamins, D.L., Wu, J., Tenenbaum, J.B., Torralba, A.: Visual grounding of learned physical models. In: ICML (2020) +45. Li, Y., Wu, J., Tedrake, R., Tenenbaum, J.B., Torralba, A.: Learning particle dynamics for manipulating rigid bodies, deformable objects, and fluids. In: ICLR (2019) +46. Li, Z., Tucker, R., Snavely, N., Holynski, A.: Generative image dynamics. arXiv preprint arXiv:2309.07906 (2023) +47. Lin, T.Y., Maire, M., Belongie, S., Hays, J., Perona, P., Ramanan, D., Dólar, P., Zitnick, C.L.: Microsoft coco: Common objects in context. In: ECCV (2014) +48. Liu, S., Zeng, Z., Ren, T., Li, F., Zhang, H., Yang, J., Li, C., Yang, J., Su, H., Zhu, J., et al.: Grounding dino: Marrying dino with grounded pre-training for open-set object detection. arXiv preprint arXiv:2303.05499 (2023) +49. Liu, T., Bargteil, A.W., O'Brien, J.F., Kavan, L.: Fast simulation of mass-spring systems. ACM TOG (2013) +50. Lv, J., Huang, Y., Yan, M., Huang, J., Liu, J., Liu, Y., Wen, Y., Chen, X., Chen, S.: Gpt4motion: Scripting physical motions in text-to-video generation via blender-oriented gpt planning. In: CVPRW. pp. 1430-1440 (2024) +51. Mahapatra, A., Kulkarni, K.: Controllable animation of fluid elements in still images. In: CVPR (2022) +52. Mallya, A., Wang, T.C., Liu, M.Y.: Implicit warping for animation with image sets. In: NeurIPS (2022) +53. Meng, C., He, Y., Song, Y., Song, J., Wu, J., Zhu, J.Y., Ermon, S.: SDEdit: Guided image synthesis and editing with stochastic differential equations. In: International Conference on Learning Representations (2022) +54. Mrowca, D., Zhuang, C., Wang, E., Haber, N., Fei-Fei, L.F., Tenenbaum, J., Yamins, D.L.: Flexible neural representation for physics prediction. In: NeurIPS (2018) +55. NVIDIA: Nvidia physx. https://developer.nvidia.com/physx-sdk (2019) +56. OpenAI: Gpt-4v(ision) system card (2023) +57. OpenAI: Creating video from text. https://openai.com/sora (2024) +58. Peebles, W., Xie, S.: Scalable diffusion models with transformers. arXiv preprint arXiv:2212.09748 (2022) +59. Popova, E., Popov, V.L.: The research works of coulomb and amontons and generalized laws of friction. Friction (2015) +60. Qiu, H., Chen, Z., Jiang, Y., Zhou, H., Fan, X., Yang, L., Wu, W., Liu, Z.: Relitalk: Relightable talking portrait generation from a single video (2023) +61. Ren, T., Liu, S., Zeng, A., Lin, J., Li, K., Cao, H., Chen, J., Huang, X., Chen, Y., Yan, F., Zeng, Z., Zhang, H., Li, F., Yang, J., Li, H., Jiang, Q., Zhang, L.: Grounded sam: Assembling open-world models for diverse visual tasks (2024) + +62. Rombach, R., Blattmann, A., Lorenz, D., Esser, P., Ommer, B.: High-resolution image synthesis with latent diffusion models. In: CVPR (2022) +63. Sanchez-Gonzalez, A., Heess, N., Springenberg, J.T., Merel, J., Riedmiller, M., Hadsell, R., Battaglia, P.: Graph networks as learnable physics engines for inference and control. In: ICML (2018) +64. Schödl, A., Szeliski, R., Salesin, D.H., Essa, I.: Video textures. In: PACMCGIT (2000) +65. Singer, U., Polyak, A., Hayes, T., Yin, X., An, J., Zhang, S., Hu, Q., Yang, H., Ashual, O., Gafni, O., et al.: Make-a-video: Text-to-video generation without text-video data. arXiv preprint arXiv:2209.14792 (2022) +66. Sohl-Dickstein, J., Weiss, E., Maheswaranathan, N., Ganguli, S.: Deep unsupervised learning using nonequilibrium thermodynamics. In: ICML (2015) +67. Song, J., Meng, C., Ermon, S.: Denoising diffusion implicit models. arXiv preprint arXiv:2010.02502 (2020) +68. Song, Y., Ermon, S.: Generative modeling by estimating gradients of the data distribution. NeurIPS (2019) +69. Sugimoto, R., He, M., Liao, J., Sander, P.V.: Water simulation and rendering from a still photograph. In: SIGGRAPH Asia (2022) +70. Szummer, M., Picard, R.W.: Temporal texture modeling. In: ICIP (1996) +71. Teed, Z., Deng, J.: Raft: Recurrent all-pairs field transforms for optical flow. In: ECCV. pp. 402-419 (2020) +72. Todorov, E., Erez, T., Tassa, Y.: Mujoco: A physics engine for model-based control. In: IROS (2012) +73. Unterthiner, T., Van Steenkiste, S., Kurach, K., Marinier, R., Michalski, M., Gelly, S.: Towards accurate generative models of video: A new metric & challenges. arXiv preprint arXiv:1812.01717 (2018) +74. Villegas, R., Babaeizadeh, M., Kindermans, P.J., Moraldo, H., Zhang, H., Saffar, M.T., Castro, S., Kunze, J., Erhan, D.: Phenaki: Variable length video generation from open domain textual description. arXiv preprint arXiv:2210.02399 (2022) +75. Wang, X., Yuan, H., Zhang, S., Chen, D., Wang, J., Zhang, Y., Shen, Y., Zhao, D., Zhou, J.: Videocomposer: Compositional video synthesis with motion controllability. arXiv preprint arXiv:2306.02018 (2023) +76. Wang, Y., Chen, X., Ma, X., Zhou, S., Huang, Z., Wang, Y., Yang, C., He, Y., Yu, J., Yang, P., et al.: Lavie: High-quality video generation with cascaded latent diffusion models. arXiv preprint arXiv:2309.15103 (2023) +77. Wei, L.Y., Levoy, M.: Fast texture synthesis using tree-structured vector quantization. In: PACMCGIT (2000) +78. Wei, Y., Zhang, S., Qing, Z., Yuan, H., Liu, Z., Liu, Y., Zhang, Y., Zhou, J., Shan, H.: Dreamvideo: Composing your dream videos with customized subject and motion. arXiv preprint arXiv:2312.04433 (2023) +79. Weng, C.Y., Curless, B., Kemelmacher-Shlizerman, I.: Photo wake-up: 3d character animation from a single photo. In: CVPR (2019) +80. Wu, J., Lim, J.J., Zhang, H., Tenenbaum, J.B., Freeman, W.T.: Physics 101: Learning physical object properties from unlabeled videos. In: BMVC (2016) +81. Wu, J., Lu, E., Kohli, P., Freeman, W.T., Tenenbaum, J.B.: Learning to see physics via visual de-animation. In: NeurIPS (2017) +82. Wu, J., Yildirim, I., Lim, J.J., Freeman, W.T., Tenenbaum, J.B.: Galileo: Perceiving physical object properties by integrating a physics engine with deep learning. In: NeurIPS (2015) +83. Wu, R., Chen, L., Yang, T., Guo, C., Li, C., Zhang, X.: Lamp: Learn a motion pattern for few-shot-based video generation. arXiv preprint arXiv:2310.10769 (2023) + +84. Xia, H., Lin, Z.H., Ma, W.C., Wang, S.: Video2game: Real-time, interactive, realistic and browser-compatible environment from a single video. In: CVPR (2024) +85. Xia, H., Lin, Z.H., Ma, W.C., Wang, S.: Video2game: Real-time, interactive, realistic and browser-compatible environment from a single video. In: CVPR (2024) +86. Xie, T., Zong, Z., Qiu, Y., Li, X., Feng, Y., Yang, Y., Jiang, C.: Physgaussian: Physics-integrated 3d gaussians for generative dynamics. In: CVPR (2024) +87. Xing, J., Xia, M., Zhang, Y., Chen, H., Yu, W., Liu, H., Wang, X., Wong, T.T., Shan, Y.: Dynamiccafter: Animating open-domain images with video diffusion priors. arXiv preprint arXiv:2310.12190 (2023) +88. Xu, Z., Wu, J., Zeng, A., Tenenbaum, J.B., Song, S.: Densephysnet: Learning dense physical object representations via multi-step dynamic interactions. In: RSS (2019) +89. Xue, T., Wu, J., Bouman, K.L., Freeman, W.T.: Visual dynamics: Stochastic future generation via layered cross convolutional networks. T-PAMI (2018) +90. Yang, J., Zhang, H., Li, F., Zou, X., Li, C., Gao, J.: Set-of-mark prompting unleashes extraordinary visual grounding in gpt-4v. arXiv preprint arXiv:2310.11441 (2023) +91. Yu, J., Cun, X., Qi, C., Zhang, Y., Wang, X., Shan, Y., Zhang, J.: Animatezero: Video diffusion models are zero-shot image animators. arXiv preprint arXiv:2312.03793 (2023) +92. Yu, L., Lezama, J., Gundavarapu, N.B., Versari, L., Sohn, K., Minnen, D., Cheng, Y., Gupta, A., Gu, X., Hauptmann, A.G., et al.: Language model beats diffusion-tokenizer is key to visual generation. arXiv preprint arXiv:2310.05737 (2023) +93. Yu, T., Feng, R., Feng, R., Liu, J., Jin, X., Zeng, W., Chen, Z.: Inpaint anything: Segment anything meets image inpainting. arXiv preprint arXiv:2304.06790 (2023) +94. Zhai, A.J., Shen, Y., Chen, E., Wang, G., Wang, X., Wang, S., Wang, X., Guan, K., Wang, S.: Physical property understanding from language-embedded feature fields. In: CVPR (2024) +95. Zhang, S., Wang, J., Zhang, Y., Zhao, K., Yuan, H., Qing, Z., Wang, X., Zhao, D., Zhou, J.: I2vgen-xl: High-quality image-to-video synthesis via cascaded diffusion models. arXiv preprint arXiv:2311.04145 (2023) +96. Zhao, J., Zhang, H.: Thin-plate spline motion model for image animation. In: CVPR (2022) \ No newline at end of file diff --git a/2024/PhysGen_ Rigid-Body Physics-Grounded Image-to-Video Generation/images.zip b/2024/PhysGen_ Rigid-Body Physics-Grounded Image-to-Video Generation/images.zip new file mode 100644 index 0000000000000000000000000000000000000000..4cfbfb011f63df266e9d9350010bb64e9b93f746 --- /dev/null +++ b/2024/PhysGen_ Rigid-Body Physics-Grounded Image-to-Video Generation/images.zip @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:9be2391193182765de577cb164d171866eab23f9f28279224a6923083a823857 +size 536142 diff --git a/2024/PhysGen_ Rigid-Body Physics-Grounded Image-to-Video Generation/layout.json b/2024/PhysGen_ Rigid-Body Physics-Grounded Image-to-Video Generation/layout.json new file mode 100644 index 0000000000000000000000000000000000000000..db5ebac1c083a781deb3f1e3603f3429e00e9fee --- /dev/null +++ b/2024/PhysGen_ Rigid-Body Physics-Grounded Image-to-Video Generation/layout.json @@ -0,0 +1,10637 @@ +{ + "pdf_info": [ + { + "para_blocks": [ + { + "bbox": [ + 165, + 111, + 449, + 148 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 165, + 111, + 449, + 148 + ], + "spans": [ + { + "bbox": [ + 165, + 111, + 449, + 148 + ], + "type": "text", + "content": "PhysGen: Rigid-Body Physics-Grounded Image-to-Video Generation" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 151, + 168, + 462, + 181 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 151, + 168, + 462, + 181 + ], + "spans": [ + { + "bbox": [ + 151, + 168, + 462, + 181 + ], + "type": "text", + "content": "Shaowei Liu, Zhongzheng Ren, Saurabh Gupta*, and Shenlong Wang*" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 210, + 190, + 402, + 214 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 210, + 190, + 402, + 214 + ], + "spans": [ + { + "bbox": [ + 210, + 190, + 402, + 214 + ], + "type": "text", + "content": "University of Illinois Urbana-Champaign https://stevenlsw.github.io/physgen/" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 160, + 239, + 465, + 449 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 160, + 239, + 465, + 449 + ], + "spans": [ + { + "bbox": [ + 160, + 239, + 465, + 449 + ], + "type": "text", + "content": "Abstract. We present PhysGen, a novel image-to-video generation method that converts a single image and an input condition (e.g., force and torque applied to an object in the image) to produce a realistic, physically plausible, and temporally consistent video. Our key insight is to integrate model-based physical simulation with a data-driven video generation process, enabling plausible image-space dynamics. At the heart of our system are three core components: (i) an image understanding module that effectively captures the geometry, materials, and physical parameters of the image; (ii) an image-space dynamics simulation model that utilizes rigid-body physics and inferred parameters to simulate realistic behaviors; and (iii) an image-based rendering and refinement module that leverages generative video diffusion to produce realistic video footage featuring the simulated motion. The resulting videos are realistic in both physics and appearance and are even precisely controllable, showcasing superior results over existing data-driven image-to-video generation works through quantitative comparison and comprehensive user study. PhysGen's resulting videos can be used for various downstream applications, such as turning an image into a realistic animation or allowing users to interact with the image and create various dynamics." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 132, + 468, + 230, + 480 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 468, + 230, + 480 + ], + "spans": [ + { + "bbox": [ + 132, + 468, + 230, + 480 + ], + "type": "text", + "content": "1 Introduction" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 130, + 491, + 482, + 586 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 491, + 482, + 586 + ], + "spans": [ + { + "bbox": [ + 130, + 491, + 482, + 586 + ], + "type": "text", + "content": "Looking at the images in Fig. 1, we can effortlessly predict or visualize the potential outcomes of various physical effects applied to the car, the curling stone, and the stack of dominos. This understanding of physics empowers our imagination of the counterfactual: we can mentally imagine various consequences of an action from an image without experiencing them. We aim to provide computers with similar capabilities - understanding and simulating physics from a single image and creating realistic animations. We expect the resulting video to produce realistic animations with physically plausible dynamics and interactions." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 130, + 587, + 482, + 647 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 587, + 482, + 647 + ], + "spans": [ + { + "bbox": [ + 130, + 587, + 482, + 647 + ], + "type": "text", + "content": "Tremendous progress has been made in generating realistic and physically plausible video footage. Conventional graphics leverage model-based dynamics to simulate plausible motions and utilize a graphics renderer to simulate images. Such methods provide realistic and controllable dynamics. However, both the dynamics of physics and lighting physics are predefined, making it hard to achieve" + } + ] + } + ], + "index": 6 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 141, + 654, + 212, + 666 + ], + "type": "page_footnote", + "angle": 0, + "lines": [ + { + "bbox": [ + 141, + 654, + 212, + 666 + ], + "spans": [ + { + "bbox": [ + 141, + 654, + 212, + 666 + ], + "type": "text", + "content": "* Equal advising" + } + ] + } + ], + "index": 7 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 0 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 140, + 125, + 478, + 295 + ], + "blocks": [ + { + "bbox": [ + 156, + 115, + 179, + 124 + ], + "lines": [ + { + "bbox": [ + 156, + 115, + 179, + 124 + ], + "spans": [ + { + "bbox": [ + 156, + 115, + 179, + 124 + ], + "type": "text", + "content": "Input" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 140, + 125, + 478, + 295 + ], + "lines": [ + { + "bbox": [ + 140, + 125, + 478, + 295 + ], + "spans": [ + { + "bbox": [ + 140, + 125, + 478, + 295 + ], + "type": "image", + "image_path": "1ffc6d1de36ee10d710eb7d62fbced194ca15a6efa6ad2e5bbe333392789cf6a.jpg" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 130, + 297, + 482, + 329 + ], + "lines": [ + { + "bbox": [ + 130, + 297, + 482, + 329 + ], + "spans": [ + { + "bbox": [ + 130, + 297, + 482, + 329 + ], + "type": "text", + "content": "Fig. 1: Given a single image, training-free PhysGen generates future frames controlled by physics and initial state. Semantics, geometry and dynamics are reasoned during generation and the result videos are both physics-grounded and photo-realistic." + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "image_caption" + } + ], + "index": 3 + }, + { + "bbox": [ + 130, + 365, + 482, + 484 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 365, + 482, + 484 + ], + "spans": [ + { + "bbox": [ + 130, + 365, + 482, + 484 + ], + "type": "text", + "content": "our goal of animating an image captured in the real world. On the other hand, data-driven image-to-video (I2V) generation techniques [6, 8, 57] learn to create realistic videos from a single image through training diffusion models over internet-scale data. However, despite advancements in video generative models [6, 8, 57], the incorporation of real-world physics principles into the video generation process remains largely unexplored and unsolved. Consequently, the synthesized videos often lack temporal coherence and fail to replicate authentic object motions observed in reality. Furthermore, such text-driven methods also lack fine-grained controllability. For instance, they cannot simulate the consequences of different forces and torques applied to an object." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 130, + 485, + 482, + 593 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 485, + 482, + 593 + ], + "spans": [ + { + "bbox": [ + 130, + 485, + 482, + 593 + ], + "type": "text", + "content": "In light of this gap, we propose a paradigm shift in video generation through the introduction of model-based video generative models. In contrast to existing purely generative methods, which are trained in a data-driven manner and rely on diffusion models to learn image space dynamics via scaling laws, we propose grounding object dynamics explicitly using rigid body physics, thereby integrating fundamental physical principles into the generation process. As classical mechanics theory reveals, the motion of an object is determined by its physical properties (e.g., mass, elasticity, surface roughness) and external factors (e.g., external forces, environmental conditions, boundary conditions)." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 130, + 594, + 482, + 666 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 594, + 482, + 666 + ], + "spans": [ + { + "bbox": [ + 130, + 594, + 482, + 666 + ], + "type": "text", + "content": "We tackle our image-to-video generation problem via a computational framework, PhysGen. PhysGen consists of three stages: 1) image-based physics understanding; 2) model-based dynamics simulation; and 3) generative video rendering with dynamics guidance. Our method first infers object compositions and physical parameters from the input image through large visual foundation model-based reasoning (Sec. 3.1). Given an input force, we then utilize the in" + } + ] + } + ], + "index": 7 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "spans": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "type": "text", + "content": "2" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 166, + 91, + 216, + 100 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 166, + 91, + 216, + 100 + ], + "spans": [ + { + "bbox": [ + 166, + 91, + 216, + 100 + ], + "type": "text", + "content": "S. Liu et al." + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 1 + }, + { + "para_blocks": [ + { + "bbox": [ + 130, + 115, + 482, + 259 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 115, + 482, + 259 + ], + "spans": [ + { + "bbox": [ + 130, + 115, + 482, + 259 + ], + "type": "text", + "content": "ferred physical parameters to perform realistic dynamic simulations for each object, accounting for their rigid body dynamics, collisions, frictions, and elasticity (Sec. 3.2). Finally, guided by motion and the input image, we present a novel generative video diffusion-based renderer that outputs the final realistic video footage (Sec. 3.3). Several illustrative examples of our framework are shown in Fig. 1. As shown in this figure, our approach combines the best of the worlds of data-driven generative modeling's appearance realism and model-based dynamics's verified physical plausibility. Notably, our generation process is highly controllable, allowing users to specify underlying physics parameters and initial conditions. This capability facilitates a range of interactive applications. Moreover, our proposed generation pipeline operates solely during inference time, eliminating the need for any training." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 130, + 261, + 482, + 439 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 261, + 482, + 439 + ], + "spans": [ + { + "bbox": [ + 130, + 261, + 482, + 439 + ], + "type": "text", + "content": "We evaluate PhysGen generation ability on multiple data source including the web and self-captured images. We compare against both state-of-the-art image-to-video models [18, 87, 95] as well as image editing method [28] quantitatively and qualitatively. We also perform user-study to evaluate the physical-realism and photo-realism of the generated video. Our method demonstrates physics-informed results for controllable image-to-video generation, producing realistic video sequences with grounded rigid-body dynamics and interaction. PhysGen combines learning-based generative approaches with traditional model-based physics to deliver visually appealing results without any training. Our approach harnesses a plethora of ingredients from previous works, particularly recent progress in large foundational visual models for segmentation [40, 48, 61, 93], physical understanding [2, 56], normal estimation [23, 26], relighting, and generative rendering [8, 18]. While these individual ingredients exist, our system combines them in a novel manner to enable an exciting new capacity, featuring physical plausibility for video generation at an unprecedented level." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 130, + 440, + 482, + 525 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 440, + 482, + 525 + ], + "spans": [ + { + "bbox": [ + 130, + 440, + 482, + 525 + ], + "type": "text", + "content": "Our contributions can be summarized as follows: 1) We propose a novel image-space dynamics model that can understand physical parameters from a single image and simulate realistic, controllable, and interactive dynamics. 2) We present a novel image-to-video system by integrating our presented dynamics model with generative diffusion-based video priors. The model is training-free, and the resulting videos are highly realistic, physically plausible, and controllable, consistently surpassing existing state-of-the-art video generative models." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 131, + 543, + 235, + 555 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 131, + 543, + 235, + 555 + ], + "spans": [ + { + "bbox": [ + 131, + 543, + 235, + 555 + ], + "type": "text", + "content": "2 Related work" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 130, + 570, + 482, + 666 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 570, + 482, + 666 + ], + "spans": [ + { + "bbox": [ + 130, + 570, + 482, + 666 + ], + "type": "text", + "content": "Image-based physics reasoning and simulation. Our image-based dynamics model is built upon physical simulation. Symbolic physics simulators or physics engines have been extensively studied for various physical processes, including rigid body dynamics, fluid simulation, and deformable objects [36-38, 41, 49, 55, 72]. Such methods have been widely used across graphics, scientific discovery, robotics, and video effects. However, these rule-based or solver-based simulators face limitations in terms of expressiveness, efficiency, generalizability, and parameter tuning. To overcome these challenges, neural physics, which combines" + } + ] + } + ], + "index": 6 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 268, + 91, + 447, + 102 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 268, + 91, + 447, + 102 + ], + "spans": [ + { + "bbox": [ + 268, + 91, + 447, + 102 + ], + "type": "text", + "content": "PhysGen: Rigid Physics-Grounded Img2Vid" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 474, + 92, + 480, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 474, + 92, + 480, + 100 + ], + "spans": [ + { + "bbox": [ + 474, + 92, + 480, + 100 + ], + "type": "text", + "content": "3" + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 2 + }, + { + "para_blocks": [ + { + "bbox": [ + 130, + 116, + 482, + 212 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 116, + 482, + 212 + ], + "spans": [ + { + "bbox": [ + 130, + 116, + 482, + 212 + ], + "type": "text", + "content": "traditional physics simulators/engines with deep neural networks, has been developed to model dynamic processes such as planar pushing and bouncing [4], as well as object interactions [3]. Domain-specific network architectures [7,45,54,63] have been developed to address various physical problems. Additionally, the integration of neural scene representations or large language models with physics simulators [43,50,86] has also been investigated recently. Nevertheless, most forward physics models, whether neural or not, have not resolved the dependency on pre-defined physical parameters, which are often specified by users." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 130, + 213, + 482, + 380 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 213, + 482, + 380 + ], + "spans": [ + { + "bbox": [ + 130, + 213, + 482, + 380 + ], + "type": "text", + "content": "Reasoning physical parameters from visual data allows us to simulate without manually defined physics, further enhancing convenience and realism for image-based physics simulation. To enable this, subsequent research efforts [21, 43, 44, 80-82, 84, 88] have introduced pipelines that first extract scene representations of the physical world from images or videos using neural networks, and then utilize these representations for either physics reasoning or simulation. Despite the promise of these approaches, they primarily rely on synthetic data for training, which might suffer from generalization issues, or on inverse physics, which requires inference from observed physical phenomena, often requiring video observations instead of a single image. In our work, we aim to generate physics-grounded, high-fidelity predictive videos from static input images of real-world objects with complex backgrounds. Contrary to existing work, we investigate leveraging the power of large pretrained visual foundational models [2, 40, 56] for physics reasoning in a zero-shot manner." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 130, + 391, + 482, + 535 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 391, + 482, + 535 + ], + "spans": [ + { + "bbox": [ + 130, + 391, + 482, + 535 + ], + "type": "text", + "content": "Video generative model. Video generative models have experienced remarkable advancements in recent years [6,8,12,18,27,29,31,33,42,57,65,74,76,87,92, 95], with the state-of-the-art framework [57] achieving the capability to generate photo-realistic and coherent imaginative videos from text instructions using diffusion models [34,58,62,66,68]. Despite progress, challenges remain, such as the need for extensive training data, object permanence, realistic lighting, and reducing unrealistic motion and physics violations. Overcoming these is key for advancing video generative modeling and its applications, especially in physics-accurate areas like scientific discovery and robotics. Our work addresses some issues by proposing a novel image-to-video framework that merges a model-based approach with a pretrained video prior, grounding the generated content with real-world physics." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 130, + 545, + 482, + 667 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 545, + 482, + 667 + ], + "spans": [ + { + "bbox": [ + 130, + 545, + 482, + 667 + ], + "type": "text", + "content": "Image animation. Our method is heavily inspired by image-based animation [19, 22, 35, 64, 70, 77, 89] and cinematographs. One common way to improve quality is to incorporate class-specific heuristics [19, 39], which can be physically simulated and integrated into the generation process. Recent advancements in deep learning have led to data-driven solutions in this domain, where temporal neural networks are trained on video datasets to directly predict consecutive video frames from an input image [10, 17, 24, 30, 35, 75]. Moreover, there is growing interest in interactive [9, 11, 46] and controllable [1, 22] image-to-video synthesis. To further improve temporal coherence, more priors such as motion fields [24, 35, 51, 52, 69, 89] or flows [14, 28, 96], 3D geometry information [60, 79]," + } + ] + } + ], + "index": 5 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "spans": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "type": "text", + "content": "4" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 166, + 91, + 216, + 100 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 166, + 91, + 216, + 100 + ], + "spans": [ + { + "bbox": [ + 166, + 91, + 216, + 100 + ], + "type": "text", + "content": "S. Liu et al." + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 3 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 149, + 114, + 465, + 233 + ], + "blocks": [ + { + "bbox": [ + 149, + 114, + 465, + 233 + ], + "lines": [ + { + "bbox": [ + 149, + 114, + 465, + 233 + ], + "spans": [ + { + "bbox": [ + 149, + 114, + 465, + 233 + ], + "type": "image", + "image_path": "6e8d796f456560f66846d4cb4bd2805cf5c761bfe7fdda1da505e68b384f37f9.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 130, + 234, + 482, + 312 + ], + "lines": [ + { + "bbox": [ + 130, + 234, + 482, + 312 + ], + "spans": [ + { + "bbox": [ + 130, + 234, + 482, + 312 + ], + "type": "text", + "content": "Fig. 2: Method overview. Our framework consists of three interleaved components: the perception module, the dynamics simulation module, and the rendering module. The perception module interprets the semantics, geometry, and physical parameters in the given image. The dynamics simulation module simulates the rigid-body motion and interactions of each instance in the scene, governed by Newton's Laws and physical constraints. The rendering module renders the final outcome, leveraging an off-the-shelf relighting model and diffusion-model-based video priors." + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_caption" + } + ], + "index": 2 + }, + { + "bbox": [ + 130, + 341, + 482, + 413 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 341, + 482, + 413 + ], + "spans": [ + { + "bbox": [ + 130, + 341, + 482, + 413 + ], + "type": "text", + "content": "and user annotations [46] have been subsequently introduced. In this work, we propose a model-based solution and tackle the image animation task from a new perspective, where our method takes a single image and user-specific initial force or torque as input, and models object dynamics explicitly via rigid physics simulation. Our key insight is that physical parameters can be inferred from a single image automatically through large foundation models such as GPT-4V [2, 56]." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 132, + 446, + 212, + 459 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 446, + 212, + 459 + ], + "spans": [ + { + "bbox": [ + 132, + 446, + 212, + 459 + ], + "type": "text", + "content": "3 Approach" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 130, + 486, + 482, + 666 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 486, + 482, + 666 + ], + "spans": [ + { + "bbox": [ + 130, + 486, + 482, + 666 + ], + "type": "text", + "content": "Given a single input image, our goal is to generate a realistic " + }, + { + "bbox": [ + 130, + 486, + 482, + 666 + ], + "type": "inline_equation", + "content": "T" + }, + { + "bbox": [ + 130, + 486, + 482, + 666 + ], + "type": "text", + "content": "-frame video featuring rigid body dynamics and interactions. Our generation can be conditioned on user-specific inputs in the form of an initial force " + }, + { + "bbox": [ + 130, + 486, + 482, + 666 + ], + "type": "inline_equation", + "content": "\\mathbf{F}_0" + }, + { + "bbox": [ + 130, + 486, + 482, + 666 + ], + "type": "text", + "content": " and/or torque " + }, + { + "bbox": [ + 130, + 486, + 482, + 666 + ], + "type": "inline_equation", + "content": "\\pmb{\\tau}_{0}" + }, + { + "bbox": [ + 130, + 486, + 482, + 666 + ], + "type": "text", + "content": ". Our key insight is to incorporate physics-informed simulation into the generation process, ensuring the dynamics are realistic and interactive, thus producing large, plausible object motion. To achieve this, we present a novel framework consisting of three stages: physical understanding, dynamics generation, and generative rendering. The perception module aims to reason about the semantics, collision geometry, and physics of the objects presented in the image (Sec. 3.1). The dynamics module then leverages the input force/torque, as well as the inferred shape and geometry, to simulate the rigid-body motions of the objects, as well as object-object and object-scene interactions, such as friction and collisions (Sec. 3.2). Finally, we convert the simulated dynamics into pixel-level motion and combine image-based warping and generative video to render the final video (Sec. 3.3). Fig. 2 depicts the overall framework." + } + ] + } + ], + "index": 6 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 268, + 91, + 447, + 102 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 268, + 91, + 447, + 102 + ], + "spans": [ + { + "bbox": [ + 268, + 91, + 447, + 102 + ], + "type": "text", + "content": "PhysGen: Rigid Physics-Grounded Img2Vid" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 474, + 92, + 480, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 474, + 92, + 480, + 100 + ], + "spans": [ + { + "bbox": [ + 474, + 92, + 480, + 100 + ], + "type": "text", + "content": "5" + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 4 + }, + { + "para_blocks": [ + { + "bbox": [ + 132, + 116, + 216, + 128 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 116, + 216, + 128 + ], + "spans": [ + { + "bbox": [ + 132, + 116, + 216, + 128 + ], + "type": "text", + "content": "3.1 Perception" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 130, + 137, + 481, + 197 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 137, + 481, + 197 + ], + "spans": [ + { + "bbox": [ + 130, + 137, + 481, + 197 + ], + "type": "text", + "content": "Simulating dynamics on an image requires a holistic understanding of object compositions, materials, geometry, and physics. Toward this goal, we designed our perception module to infer such properties from a single image. Our key insight is to harness readily available, large pretrained models [2, 15, 26, 56, 61] to achieve this goal, as shown in Fig. 2." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 130, + 209, + 482, + 329 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 209, + 482, + 329 + ], + "spans": [ + { + "bbox": [ + 130, + 209, + 482, + 329 + ], + "type": "text", + "content": "Segmentation. Identifying and segmenting each individual physical entity lays the foundation for image-based dynamics. Inspired by the recent success of large pretrained segmentation models, we incorporate GPT-4V [2, 56] to recognize all image categories and send to Grounded-SAM [61] to detect and segment each individual instance-level objects " + }, + { + "bbox": [ + 130, + 209, + 482, + 329 + ], + "type": "inline_equation", + "content": "\\{\\mathbf{o}^i\\in \\mathbb{R}^{W\\times H}\\}_{i}^{N}" + }, + { + "bbox": [ + 130, + 209, + 482, + 329 + ], + "type": "text", + "content": ", where " + }, + { + "bbox": [ + 130, + 209, + 482, + 329 + ], + "type": "inline_equation", + "content": "\\mathbf{o}^i" + }, + { + "bbox": [ + 130, + 209, + 482, + 329 + ], + "type": "text", + "content": " is the binary mask for " + }, + { + "bbox": [ + 130, + 209, + 482, + 329 + ], + "type": "inline_equation", + "content": "i" + }, + { + "bbox": [ + 130, + 209, + 482, + 329 + ], + "type": "text", + "content": "-th object. We also query the GPT-4V to classify the objects into foreground and background based on their movability. The foreground objects are send to the physical simulation once user inputs are applied. We extract collision boundaries and supporting edges (e.g., ground, walls, etc.) from the background objects. The details are presented in the supplementary." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 130, + 342, + 482, + 497 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 342, + 482, + 497 + ], + "spans": [ + { + "bbox": [ + 130, + 342, + 482, + 497 + ], + "type": "text", + "content": "Physical properties reasoning. Realistic physics depends on accurate physical parameters, such as surface friction, mass, and elasticity. Unlike prior physics-based image dynamics work [19], our work leverages visual foundational models to reason physical properties directly. Our solution is simple yet effective. Inspired by the success of mask-based prompting [2,56,90], we directly ask GPT-4V [56] for certain physical properties, providing an object mask overlaid on the input image. Following [85,94], we send the crafted prompt to GPT-4V, which contains GPT-4V instructions to return a quantitative measure of each queried property in a metric unit. For each object, we query its mass, elasticity, and friction coefficient " + }, + { + "bbox": [ + 130, + 342, + 482, + 497 + ], + "type": "inline_equation", + "content": "(M_i,E_i,\\mu_i)" + }, + { + "bbox": [ + 130, + 342, + 482, + 497 + ], + "type": "text", + "content": ". We use the Coulomb friction model [59] for friction modeling. The elasticity coefficient is a scalar, where 0.0 results in no bounce, and 1.0 results in a perfect bounce. Given each object mass " + }, + { + "bbox": [ + 130, + 342, + 482, + 497 + ], + "type": "inline_equation", + "content": "M_{i}" + }, + { + "bbox": [ + 130, + 342, + 482, + 497 + ], + "type": "text", + "content": " in grams and its shape primitive, we compute the rotational inertia " + }, + { + "bbox": [ + 130, + 342, + 482, + 497 + ], + "type": "inline_equation", + "content": "\\mathbf{I}_i" + }, + { + "bbox": [ + 130, + 342, + 482, + 497 + ], + "type": "text", + "content": " accordingly." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 130, + 509, + 482, + 605 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 509, + 482, + 605 + ], + "spans": [ + { + "bbox": [ + 130, + 509, + 482, + 605 + ], + "type": "text", + "content": "Geometry primitives. Rasterized instance masks are not suitable for physical simulation; hence, we need to convert each object into vectorized shape primitives. We fit two types of primitive shapes: a circle and generic polygons. For each object, we choose the primitive type with the maximum coverage of its segmentation. For circles, we fit the center and radius to cover the segmentation mask. Circles allow us to realistically simulate rolling motions. For non-circle objects, we instead fit the generic polygons. Specifically, we perform contour extraction on the segmentation masks to obtain the polygon vertices." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 130, + 617, + 482, + 666 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 617, + 482, + 666 + ], + "spans": [ + { + "bbox": [ + 130, + 617, + 482, + 666 + ], + "type": "text", + "content": "Intrinsic decomposition. The dynamic movement of the objects will also result in comprehensive changes in shading, as their position wrt the lighting environment changes. To compensate for this effect during rendering, we must perform image decomposition to infer albedo, normal as well as lighting from" + } + ] + } + ], + "index": 7 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "spans": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "type": "text", + "content": "6" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 166, + 91, + 216, + 101 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 166, + 91, + 216, + 101 + ], + "spans": [ + { + "bbox": [ + 166, + 91, + 216, + 101 + ], + "type": "text", + "content": "S. Liu et al." + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 5 + }, + { + "para_blocks": [ + { + "bbox": [ + 131, + 115, + 479, + 153 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 131, + 115, + 479, + 153 + ], + "spans": [ + { + "bbox": [ + 131, + 115, + 479, + 153 + ], + "type": "text", + "content": "the single image for each object " + }, + { + "bbox": [ + 131, + 115, + 479, + 153 + ], + "type": "inline_equation", + "content": "(\\mathbf{A}_0^i,\\mathbf{N}_0^i,\\mathbf{L}_0)" + }, + { + "bbox": [ + 131, + 115, + 479, + 153 + ], + "type": "text", + "content": ". To achieve this we leverage off-the-shelf intrinsic decomposition model [15] to compute " + }, + { + "bbox": [ + 131, + 115, + 479, + 153 + ], + "type": "inline_equation", + "content": "\\mathbf{A}_0^i" + }, + { + "bbox": [ + 131, + 115, + 479, + 153 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 131, + 115, + 479, + 153 + ], + "type": "inline_equation", + "content": "\\mathbf{L}_0" + }, + { + "bbox": [ + 131, + 115, + 479, + 153 + ], + "type": "text", + "content": " and exploits pretrained surface normal estimator for the computation of object normal " + }, + { + "bbox": [ + 131, + 115, + 479, + 153 + ], + "type": "inline_equation", + "content": "\\mathbf{N}_0^i" + }, + { + "bbox": [ + 131, + 115, + 479, + 153 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 131, + 165, + 479, + 225 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 131, + 165, + 479, + 225 + ], + "spans": [ + { + "bbox": [ + 131, + 165, + 479, + 225 + ], + "type": "text", + "content": "Background inpainting. Once the foreground objects move, they will leave holes in the background. To produce realistic and complete video footage, we leverage an off-the-shelf generative image inpainting model [93] to recover the complete background scene without foreground objects. The inpainted image " + }, + { + "bbox": [ + 131, + 165, + 479, + 225 + ], + "type": "inline_equation", + "content": "\\mathbf{B} \\in \\mathbb{R}^{W \\times H \\times 3}" + }, + { + "bbox": [ + 131, + 165, + 479, + 225 + ], + "type": "text", + "content": " will be used for composition in the rendering stage." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 131, + 245, + 328, + 257 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 131, + 245, + 328, + 257 + ], + "spans": [ + { + "bbox": [ + 131, + 245, + 328, + 257 + ], + "type": "text", + "content": "3.2 Image space dynamics simulation" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 130, + 266, + 480, + 411 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 266, + 480, + 411 + ], + "spans": [ + { + "bbox": [ + 130, + 266, + 480, + 411 + ], + "type": "text", + "content": "Given the foreground objects with physical properties, we use rigid-body physics for dynamics simulation in image spaces. We choose to perform simulations in image spaces for three reasons: 1) image-space dynamics are better coupled with our output video; 2) a full 3D simulation requires complete 3D scene reconstruction and understanding, which remains an unsolved problem from a single image; 3) image-space dynamics can already cover various object dynamics and have been widely used in prior work [4, 19, 35, 46, 80]. Specifically, at a given time " + }, + { + "bbox": [ + 130, + 266, + 480, + 411 + ], + "type": "inline_equation", + "content": "t" + }, + { + "bbox": [ + 130, + 266, + 480, + 411 + ], + "type": "text", + "content": ", each rigid object " + }, + { + "bbox": [ + 130, + 266, + 480, + 411 + ], + "type": "inline_equation", + "content": "i" + }, + { + "bbox": [ + 130, + 266, + 480, + 411 + ], + "type": "text", + "content": " is characterized by its 2D pose and velocity at its center of mass. The position includes a translation " + }, + { + "bbox": [ + 130, + 266, + 480, + 411 + ], + "type": "inline_equation", + "content": "\\mathbf{t}^i(t) \\in \\mathbb{R}^2" + }, + { + "bbox": [ + 130, + 266, + 480, + 411 + ], + "type": "text", + "content": " and a rotation " + }, + { + "bbox": [ + 130, + 266, + 480, + 411 + ], + "type": "inline_equation", + "content": "\\mathbf{R}^i(t) \\in \\mathbb{S}\\mathbb{O}(2)" + }, + { + "bbox": [ + 130, + 266, + 480, + 411 + ], + "type": "text", + "content": " specified in a world coordinate. The velocity includes linear velocity " + }, + { + "bbox": [ + 130, + 266, + 480, + 411 + ], + "type": "inline_equation", + "content": "\\pmb{\\nu}^i(t) \\in \\mathbb{R}^2" + }, + { + "bbox": [ + 130, + 266, + 480, + 411 + ], + "type": "text", + "content": " and angular velocity " + }, + { + "bbox": [ + 130, + 266, + 480, + 411 + ], + "type": "inline_equation", + "content": "\\pmb{\\omega}^i(t) \\in \\mathbb{R}^2" + }, + { + "bbox": [ + 130, + 266, + 480, + 411 + ], + "type": "text", + "content": ". Hence, the state of the each object " + }, + { + "bbox": [ + 130, + 266, + 480, + 411 + ], + "type": "inline_equation", + "content": "i" + }, + { + "bbox": [ + 130, + 266, + 480, + 411 + ], + "type": "text", + "content": " at time " + }, + { + "bbox": [ + 130, + 266, + 480, + 411 + ], + "type": "inline_equation", + "content": "t" + }, + { + "bbox": [ + 130, + 266, + 480, + 411 + ], + "type": "text", + "content": " can be represented as " + }, + { + "bbox": [ + 130, + 266, + 480, + 411 + ], + "type": "inline_equation", + "content": "\\mathbf{q}^i(t) = [\\mathbf{t}^i(t), \\mathbf{R}^i(t), \\pmb{\\nu}^i(t), \\pmb{\\omega}^i(t)]" + }, + { + "bbox": [ + 130, + 266, + 480, + 411 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 131, + 411, + 479, + 436 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 131, + 411, + 479, + 436 + ], + "spans": [ + { + "bbox": [ + 131, + 411, + 479, + 436 + ], + "type": "text", + "content": "Following [25], the rigid body motion dynamics is given by Eq. (1) for each object. For simplicity, we omit the object index " + }, + { + "bbox": [ + 131, + 411, + 479, + 436 + ], + "type": "inline_equation", + "content": "i" + }, + { + "bbox": [ + 131, + 411, + 479, + 436 + ], + "type": "text", + "content": " in the following:" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 188, + 444, + 480, + 496 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 188, + 444, + 480, + 496 + ], + "spans": [ + { + "bbox": [ + 188, + 444, + 480, + 496 + ], + "type": "interline_equation", + "content": "\\frac {d}{d t} \\mathbf {q} (t) = \\frac {d}{d t} \\left[ \\begin{array}{l} \\mathbf {t} (t) \\\\ \\mathbf {R} (t) \\\\ \\boldsymbol {\\nu} (t) \\\\ \\boldsymbol {\\omega} (t) \\end{array} \\right] = \\left[ \\begin{array}{c} \\mathbf {v} (t) \\\\ \\boldsymbol {\\omega} (t) \\times \\mathbf {R} (t) \\\\ \\frac {\\mathbf {F} (t)}{M} \\\\ \\mathbf {I} (t) ^ {- 1} (\\boldsymbol {\\tau} - \\boldsymbol {\\omega} (t) \\times \\mathbf {I} (t) \\boldsymbol {\\omega} (t)) \\end{array} \\right] \\tag {1}", + "image_path": "2514aa5ee80ceb07020eaf1dfcb598e44e6cd524317137d31c8cc02f95c145a6.jpg" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 131, + 505, + 479, + 554 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 131, + 505, + 479, + 554 + ], + "spans": [ + { + "bbox": [ + 131, + 505, + 479, + 554 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 131, + 505, + 479, + 554 + ], + "type": "inline_equation", + "content": "\\mathbf{F}" + }, + { + "bbox": [ + 131, + 505, + 479, + 554 + ], + "type": "text", + "content": " is the force, " + }, + { + "bbox": [ + 131, + 505, + 479, + 554 + ], + "type": "inline_equation", + "content": "\\pmb{\\tau}" + }, + { + "bbox": [ + 131, + 505, + 479, + 554 + ], + "type": "text", + "content": " is the torque, " + }, + { + "bbox": [ + 131, + 505, + 479, + 554 + ], + "type": "inline_equation", + "content": "M" + }, + { + "bbox": [ + 131, + 505, + 479, + 554 + ], + "type": "text", + "content": " is the mass of the object, " + }, + { + "bbox": [ + 131, + 505, + 479, + 554 + ], + "type": "inline_equation", + "content": "\\mathbf{I}(t) = \\mathbf{R}(t)\\mathbf{I}\\mathbf{R}(t)^T" + }, + { + "bbox": [ + 131, + 505, + 479, + 554 + ], + "type": "text", + "content": " is the rotation inertia in world coordinates, " + }, + { + "bbox": [ + 131, + 505, + 479, + 554 + ], + "type": "inline_equation", + "content": "\\mathbf{I} \\in \\mathbb{R}^{2 \\times 2}" + }, + { + "bbox": [ + 131, + 505, + 479, + 554 + ], + "type": "text", + "content": " is the inertia matrix in its body coordinates. This inertia indicates its resistance against rotational motion. The state " + }, + { + "bbox": [ + 131, + 505, + 479, + 554 + ], + "type": "inline_equation", + "content": "\\mathbf{q}(t)" + }, + { + "bbox": [ + 131, + 505, + 479, + 554 + ], + "type": "text", + "content": " is given by integrating Eq. (1):" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 187, + 563, + 480, + 596 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 187, + 563, + 480, + 596 + ], + "spans": [ + { + "bbox": [ + 187, + 563, + 480, + 596 + ], + "type": "interline_equation", + "content": "\\mathbf {q} (t) = \\mathbf {q} (0) + \\int_ {0} ^ {t} \\frac {d}{d t} \\mathbf {q} (t) d t = \\mathbf {q} (0) + \\sum_ {i = 1} ^ {T} \\frac {d}{d t} \\mathbf {q} (t) | _ {t = t _ {i}} \\Delta_ {t} \\tag {2}", + "image_path": "526ff98248ee2753d786caeedfa7f3f51a14f68973cc2dc86aa9a228604c6592.jpg" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 131, + 605, + 479, + 666 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 131, + 605, + 479, + 666 + ], + "spans": [ + { + "bbox": [ + 131, + 605, + 479, + 666 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 131, + 605, + 479, + 666 + ], + "type": "inline_equation", + "content": "\\mathbf{q}(0)" + }, + { + "bbox": [ + 131, + 605, + 479, + 666 + ], + "type": "text", + "content": " is the initial condition specified in the input image. We compute the integral using numerical ODE integrations, e.g. Euler method [20]. Given the initial state of the objects, the physical motion simulation module synthesis each foreground object future motion. The location of each object is updated by the affine transformation " + }, + { + "bbox": [ + 131, + 605, + 479, + 666 + ], + "type": "inline_equation", + "content": "\\mathbf{T}(t) = [\\mathbf{R}(t),\\mathbf{t}(t)]" + }, + { + "bbox": [ + 131, + 605, + 479, + 666 + ], + "type": "text", + "content": " from the initial location." + } + ] + } + ], + "index": 10 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 268, + 90, + 447, + 102 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 268, + 90, + 447, + 102 + ], + "spans": [ + { + "bbox": [ + 268, + 90, + 447, + 102 + ], + "type": "text", + "content": "PhysGen: Rigid Physics-Grounded Img2Vid" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 474, + 91, + 480, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 474, + 91, + 480, + 100 + ], + "spans": [ + { + "bbox": [ + 474, + 91, + 480, + 100 + ], + "type": "text", + "content": "7" + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 6 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 164, + 114, + 236, + 186 + ], + "blocks": [ + { + "bbox": [ + 164, + 114, + 236, + 186 + ], + "lines": [ + { + "bbox": [ + 164, + 114, + 236, + 186 + ], + "spans": [ + { + "bbox": [ + 164, + 114, + 236, + 186 + ], + "type": "image", + "image_path": "36ea1b241fe3a56297e00bc470356239e8a695cb84b3f184cf954e3b8380d2b6.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 191, + 186, + 208, + 194 + ], + "lines": [ + { + "bbox": [ + 191, + 186, + 208, + 194 + ], + "spans": [ + { + "bbox": [ + 191, + 186, + 208, + 194 + ], + "type": "text", + "content": "Input" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_caption" + } + ], + "index": 2 + }, + { + "type": "image", + "bbox": [ + 236, + 114, + 309, + 186 + ], + "blocks": [ + { + "bbox": [ + 236, + 114, + 309, + 186 + ], + "lines": [ + { + "bbox": [ + 236, + 114, + 309, + 186 + ], + "spans": [ + { + "bbox": [ + 236, + 114, + 309, + 186 + ], + "type": "image", + "image_path": "37c392b76552cd52195b85e02498cf7e45a5809b6650bed56d0f4ac515e1c776.jpg" + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 239, + 186, + 304, + 194 + ], + "lines": [ + { + "bbox": [ + 239, + 186, + 304, + 194 + ], + "spans": [ + { + "bbox": [ + 239, + 186, + 304, + 194 + ], + "type": "text", + "content": "Composed (Sec. 3.3)" + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 130, + 194, + 481, + 237 + ], + "lines": [ + { + "bbox": [ + 130, + 194, + 481, + 237 + ], + "spans": [ + { + "bbox": [ + 130, + 194, + 481, + 237 + ], + "type": "text", + "content": "Fig. 3: Rendered video comparison. We show a toy example of piggy bank. The left shows the input frame, and the rest 3 are future frame generations. The composited piggy bank hasn't been aware of the light change. The relighted output synthesized no shadows beneath. The rendered output from diffusion model is most photo realistic." + } + ] + } + ], + "index": 10, + "angle": 0, + "type": "image_caption" + } + ], + "index": 4 + }, + { + "type": "image", + "bbox": [ + 309, + 114, + 380, + 186 + ], + "blocks": [ + { + "bbox": [ + 309, + 114, + 380, + 186 + ], + "lines": [ + { + "bbox": [ + 309, + 114, + 380, + 186 + ], + "spans": [ + { + "bbox": [ + 309, + 114, + 380, + 186 + ], + "type": "image", + "image_path": "313a06508e7cc87561a7dc9dc12beab66d4812a8c921ade0b5e94f845501af37.jpg" + } + ] + } + ], + "index": 6, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 318, + 186, + 369, + 194 + ], + "lines": [ + { + "bbox": [ + 318, + 186, + 369, + 194 + ], + "spans": [ + { + "bbox": [ + 318, + 186, + 369, + 194 + ], + "type": "text", + "content": "Relight (Sec. 3.3)" + } + ] + } + ], + "index": 7, + "angle": 0, + "type": "image_caption" + } + ], + "index": 6 + }, + { + "type": "image", + "bbox": [ + 380, + 114, + 452, + 186 + ], + "blocks": [ + { + "bbox": [ + 380, + 114, + 452, + 186 + ], + "lines": [ + { + "bbox": [ + 380, + 114, + 452, + 186 + ], + "spans": [ + { + "bbox": [ + 380, + 114, + 452, + 186 + ], + "type": "image", + "image_path": "574ab8361d0e985d1af4576d717f32c5fb8f8bfdef45407afb47f7810c62980b.jpg" + } + ] + } + ], + "index": 8, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 390, + 186, + 441, + 194 + ], + "lines": [ + { + "bbox": [ + 390, + 186, + 441, + 194 + ], + "spans": [ + { + "bbox": [ + 390, + 186, + 441, + 194 + ], + "type": "text", + "content": "Render (Sec. 3.4)" + } + ] + } + ], + "index": 9, + "angle": 0, + "type": "image_caption" + } + ], + "index": 8 + }, + { + "bbox": [ + 130, + 271, + 479, + 319 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 271, + 479, + 319 + ], + "spans": [ + { + "bbox": [ + 130, + 271, + 479, + 319 + ], + "type": "text", + "content": "External force and torque. For each object, the external forces " + }, + { + "bbox": [ + 130, + 271, + 479, + 319 + ], + "type": "inline_equation", + "content": "\\mathbf{F}(t)" + }, + { + "bbox": [ + 130, + 271, + 479, + 319 + ], + "type": "text", + "content": " and torques " + }, + { + "bbox": [ + 130, + 271, + 479, + 319 + ], + "type": "inline_equation", + "content": "\\pmb{\\tau}" + }, + { + "bbox": [ + 130, + 271, + 479, + 319 + ], + "type": "text", + "content": " include gravity, friction, and elasticity between the object's surface and the environment. Specifically, we consider the initial external force and torque from user input, as well as gravity, rolling friction, and sliding friction." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 130, + 331, + 480, + 415 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 331, + 480, + 415 + ], + "spans": [ + { + "bbox": [ + 130, + 331, + 480, + 415 + ], + "type": "text", + "content": "Collision. When objects move, they might collide with each other or with the scene boundary (e.g., hitting a wall). Therefore, collision checking is necessary at every time step. A collision will result in offset forces being applied from its center of mass, producing " + }, + { + "bbox": [ + 130, + 331, + 480, + 415 + ], + "type": "inline_equation", + "content": "\\tau" + }, + { + "bbox": [ + 130, + 331, + 480, + 415 + ], + "type": "text", + "content": " and causing the object to rotate. In collision reactions, changes in energy and momentum are minimized to adhere to Newton's conservation laws. In the case of an object falling, it could bounce back or start rolling on the ground due to the elasticity of both the object and the ground." + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 131, + 432, + 213, + 444 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 131, + 432, + 213, + 444 + ], + "spans": [ + { + "bbox": [ + 131, + 432, + 213, + 444 + ], + "type": "text", + "content": "3.3 Rendering" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 130, + 452, + 482, + 561 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 452, + 482, + 561 + ], + "spans": [ + { + "bbox": [ + 130, + 452, + 482, + 561 + ], + "type": "text", + "content": "Given the simulated motion sequence " + }, + { + "bbox": [ + 130, + 452, + 482, + 561 + ], + "type": "inline_equation", + "content": "\\{\\mathbf{q}^i(0),\\dots,\\mathbf{q}^i(t),\\dots\\}_i^N" + }, + { + "bbox": [ + 130, + 452, + 482, + 561 + ], + "type": "text", + "content": " and the input image " + }, + { + "bbox": [ + 130, + 452, + 482, + 561 + ], + "type": "inline_equation", + "content": "\\mathbf{C}" + }, + { + "bbox": [ + 130, + 452, + 482, + 561 + ], + "type": "text", + "content": ", the rendering module outputs the rendered video " + }, + { + "bbox": [ + 130, + 452, + 482, + 561 + ], + "type": "inline_equation", + "content": "\\{\\dots,\\mathbf{C}(t),\\dots\\}" + }, + { + "bbox": [ + 130, + 452, + 482, + 561 + ], + "type": "text", + "content": ". There are a few desiderata for generating realistic video: it needs to be temporally consistent, complete, reflect the lighting changes as the object moves, and accurately represent the simulated image space dynamics. To achieve this goal, we design a novel motion-guided rendering pipeline consisting of an image-based transformation and composition module to ensure motion awareness, a relighting module to ensure the plausibility of lighting physics and a generative video editing module that retouches the composed and relit video to ensure realism." + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 130, + 571, + 482, + 643 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 571, + 482, + 643 + ], + "spans": [ + { + "bbox": [ + 130, + 571, + 482, + 643 + ], + "type": "text", + "content": "Composed video. Given the object state from the physical motion module, we compose an initial video by alpha-blending the foreground scene with the static inpainted background from Sec. 3.1. The foreground scene is rendered by performing forward-warping from the input image to future frames using the affine transformation " + }, + { + "bbox": [ + 130, + 571, + 482, + 643 + ], + "type": "inline_equation", + "content": "\\mathbf{T}(t)" + }, + { + "bbox": [ + 130, + 571, + 482, + 643 + ], + "type": "text", + "content": ". The alpha channel is computed using the same procedure, with the input being the segmentation mask in Sec. 3.1:" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 167, + 652, + 444, + 666 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 167, + 652, + 444, + 666 + ], + "spans": [ + { + "bbox": [ + 167, + 652, + 444, + 666 + ], + "type": "interline_equation", + "content": "\\hat {\\mathbf {X}} (t) = \\operatorname {c o m p o s i t e} (\\mathbf {B}, \\operatorname {w a r p} \\left(\\mathbf {X} ^ {0}, \\mathbf {T} ^ {0} (t)\\right) \\dots \\operatorname {w a r p} \\left(\\mathbf {X} ^ {i}, \\mathbf {T} ^ {i} (t)\\right) \\dots \\dots)", + "image_path": "8672bd735d7fb325650d88e5709727f60cfbb6682b01d16d03b008ba64f70ccb.jpg" + } + ] + } + ], + "index": 16 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "spans": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "type": "text", + "content": "8" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 166, + 91, + 216, + 100 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 166, + 91, + 216, + 100 + ], + "spans": [ + { + "bbox": [ + 166, + 91, + 216, + 100 + ], + "type": "text", + "content": "S. Liu et al." + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 7 + }, + { + "para_blocks": [ + { + "bbox": [ + 130, + 116, + 479, + 153 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 116, + 479, + 153 + ], + "spans": [ + { + "bbox": [ + 130, + 116, + 479, + 153 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 130, + 116, + 479, + 153 + ], + "type": "inline_equation", + "content": "\\mathbf{X}^i" + }, + { + "bbox": [ + 130, + 116, + 479, + 153 + ], + "type": "text", + "content": " are segmented input image of " + }, + { + "bbox": [ + 130, + 116, + 479, + 153 + ], + "type": "inline_equation", + "content": "i" + }, + { + "bbox": [ + 130, + 116, + 479, + 153 + ], + "type": "text", + "content": "-th object. Apart from the RGB sequence " + }, + { + "bbox": [ + 130, + 116, + 479, + 153 + ], + "type": "inline_equation", + "content": "\\hat{\\mathbf{V}} = \\{\\dots \\hat{\\mathbf{X}}(t) \\dots\\}" + }, + { + "bbox": [ + 130, + 116, + 479, + 153 + ], + "type": "text", + "content": ", we also use the same affine warping and composition to compute the blended albedo map " + }, + { + "bbox": [ + 130, + 116, + 479, + 153 + ], + "type": "inline_equation", + "content": "\\hat{\\mathbf{A}}(t)" + }, + { + "bbox": [ + 130, + 116, + 479, + 153 + ], + "type": "text", + "content": ", which is later used for relighting." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 130, + 164, + 480, + 250 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 164, + 480, + 250 + ], + "spans": [ + { + "bbox": [ + 130, + 164, + 480, + 250 + ], + "type": "text", + "content": "Relighting. Our relighting module simulates the changes in shading due to object movement. Specifically, it takes as input the transformed albedo " + }, + { + "bbox": [ + 130, + 164, + 480, + 250 + ], + "type": "inline_equation", + "content": "\\hat{\\mathbf{A}}(t)" + }, + { + "bbox": [ + 130, + 164, + 480, + 250 + ], + "type": "text", + "content": " and RGB image " + }, + { + "bbox": [ + 130, + 164, + 480, + 250 + ], + "type": "inline_equation", + "content": "\\hat{\\mathbf{X}}(t)" + }, + { + "bbox": [ + 130, + 164, + 480, + 250 + ], + "type": "text", + "content": " at each time " + }, + { + "bbox": [ + 130, + 164, + 480, + 250 + ], + "type": "inline_equation", + "content": "t" + }, + { + "bbox": [ + 130, + 164, + 480, + 250 + ], + "type": "text", + "content": ", as well as the estimated lighting " + }, + { + "bbox": [ + 130, + 164, + 480, + 250 + ], + "type": "inline_equation", + "content": "\\mathbf{L}" + }, + { + "bbox": [ + 130, + 164, + 480, + 250 + ], + "type": "text", + "content": ". We first compute the surface normal " + }, + { + "bbox": [ + 130, + 164, + 480, + 250 + ], + "type": "inline_equation", + "content": "\\hat{\\mathbf{N}}(t)" + }, + { + "bbox": [ + 130, + 164, + 480, + 250 + ], + "type": "text", + "content": " from the composited RGB image " + }, + { + "bbox": [ + 130, + 164, + 480, + 250 + ], + "type": "inline_equation", + "content": "\\hat{\\mathbf{X}}(t)" + }, + { + "bbox": [ + 130, + 164, + 480, + 250 + ], + "type": "text", + "content": " using an off-the-shelf normal estimator [23]. We then perform reshading using the parametric illumination model [15]: " + }, + { + "bbox": [ + 130, + 164, + 480, + 250 + ], + "type": "inline_equation", + "content": "\\tilde{\\mathbf{X}}(t) = f(\\hat{\\mathbf{X}}(t), \\hat{\\mathbf{A}}(t), \\hat{\\mathbf{N}}(t), \\mathbf{L})" + }, + { + "bbox": [ + 130, + 164, + 480, + 250 + ], + "type": "text", + "content": ", which returns the relit image " + }, + { + "bbox": [ + 130, + 164, + 480, + 250 + ], + "type": "inline_equation", + "content": "\\hat{\\mathbf{X}}(t)" + }, + { + "bbox": [ + 130, + 164, + 480, + 250 + ], + "type": "text", + "content": " for each frame " + }, + { + "bbox": [ + 130, + 164, + 480, + 250 + ], + "type": "inline_equation", + "content": "t" + }, + { + "bbox": [ + 130, + 164, + 480, + 250 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 131, + 268, + 416, + 279 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 131, + 268, + 416, + 279 + ], + "spans": [ + { + "bbox": [ + 131, + 268, + 416, + 279 + ], + "type": "text", + "content": "3.4 Generative refinement with latent diffusion models" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 130, + 287, + 480, + 358 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 287, + 480, + 358 + ], + "spans": [ + { + "bbox": [ + 130, + 287, + 480, + 358 + ], + "type": "text", + "content": "One limitation of the single-image based relighting model is that it neglects the background and cannot handle complicated lighting effects due to complex object-object interactions, such as ambient occlusion and cast shadows. Moreover, the composition results in unnatural boundaries. To address these issues and enhance realism, we incorporated a diffusion-based video to refine the relit video, denoted as " + }, + { + "bbox": [ + 130, + 287, + 480, + 358 + ], + "type": "inline_equation", + "content": "\\tilde{\\mathbf{V}} = \\{\\ldots ,\\tilde{\\mathbf{X}} (t),\\ldots \\}" + }, + { + "bbox": [ + 130, + 287, + 480, + 358 + ], + "type": "text", + "content": ", and obtain the final video output." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 130, + 359, + 481, + 561 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 359, + 481, + 561 + ], + "spans": [ + { + "bbox": [ + 130, + 359, + 481, + 561 + ], + "type": "text", + "content": "Specifically, we incorporate a pretrained video diffusion model to refine our video " + }, + { + "bbox": [ + 130, + 359, + 481, + 561 + ], + "type": "inline_equation", + "content": "\\tilde{\\mathbf{V}}" + }, + { + "bbox": [ + 130, + 359, + 481, + 561 + ], + "type": "text", + "content": ". We first use the pretrained latent diffusion encoder [62] to encode the guidance video " + }, + { + "bbox": [ + 130, + 359, + 481, + 561 + ], + "type": "inline_equation", + "content": "\\tilde{\\mathbf{V}}" + }, + { + "bbox": [ + 130, + 359, + 481, + 561 + ], + "type": "text", + "content": " into a latent code " + }, + { + "bbox": [ + 130, + 359, + 481, + 561 + ], + "type": "inline_equation", + "content": "\\mathbf{z}" + }, + { + "bbox": [ + 130, + 359, + 481, + 561 + ], + "type": "text", + "content": ". Inspired by SDEdit [53], we add noise to the guided latent code and gradually denoise it using the pretrained video diffusion. Given that the content in the guided latent code is already satisfactory, we do not perform the denoising process from scratch. Instead, we define an initial noise strength " + }, + { + "bbox": [ + 130, + 359, + 481, + 561 + ], + "type": "inline_equation", + "content": "s" + }, + { + "bbox": [ + 130, + 359, + 481, + 561 + ], + "type": "text", + "content": " where " + }, + { + "bbox": [ + 130, + 359, + 481, + 561 + ], + "type": "inline_equation", + "content": "s \\in [0,1]" + }, + { + "bbox": [ + 130, + 359, + 481, + 561 + ], + "type": "text", + "content": " controls the amount of noise added to " + }, + { + "bbox": [ + 130, + 359, + 481, + 561 + ], + "type": "inline_equation", + "content": "\\mathbf{z}" + }, + { + "bbox": [ + 130, + 359, + 481, + 561 + ], + "type": "text", + "content": ". In practice, we find that a certain approach finds a good trade-off between fidelity and realism. At each denoising step, we ensure the foreground objects are as consistent as possible with the guidance (copied from the guided latent code) while synthesizing new content (e.g., shadows) in the background (using the denoised latent code). To this end, we use a fusion weight " + }, + { + "bbox": [ + 130, + 359, + 481, + 561 + ], + "type": "inline_equation", + "content": "w" + }, + { + "bbox": [ + 130, + 359, + 481, + 561 + ], + "type": "text", + "content": " to control how much of the generated latent to inject into the foreground; the fusion weight is gradually increased during the denoising as the perturbation of noise decreases. We also define a fusion timestamp " + }, + { + "bbox": [ + 130, + 359, + 481, + 561 + ], + "type": "inline_equation", + "content": "\\delta" + }, + { + "bbox": [ + 130, + 359, + 481, + 561 + ], + "type": "text", + "content": " as the signal to stop the fusion. The final denoised latent code " + }, + { + "bbox": [ + 130, + 359, + 481, + 561 + ], + "type": "inline_equation", + "content": "\\mathbf{z}_{*}" + }, + { + "bbox": [ + 130, + 359, + 481, + 561 + ], + "type": "text", + "content": " is then decoded to our final video output " + }, + { + "bbox": [ + 130, + 359, + 481, + 561 + ], + "type": "inline_equation", + "content": "\\mathbf{V}" + }, + { + "bbox": [ + 130, + 359, + 481, + 561 + ], + "type": "text", + "content": ". Please see the supplementary material for a detailed algorithm." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 131, + 562, + 479, + 586 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 131, + 562, + 479, + 586 + ], + "spans": [ + { + "bbox": [ + 131, + 562, + 479, + 586 + ], + "type": "text", + "content": "Fig. 3 depicts the composed video " + }, + { + "bbox": [ + 131, + 562, + 479, + 586 + ], + "type": "inline_equation", + "content": "\\hat{\\mathbf{V}}" + }, + { + "bbox": [ + 131, + 562, + 479, + 586 + ], + "type": "text", + "content": ", the relit video " + }, + { + "bbox": [ + 131, + 562, + 479, + 586 + ], + "type": "inline_equation", + "content": "\\tilde{\\mathbf{V}}" + }, + { + "bbox": [ + 131, + 562, + 479, + 586 + ], + "type": "text", + "content": ", and our final output " + }, + { + "bbox": [ + 131, + 562, + 479, + 586 + ], + "type": "inline_equation", + "content": "\\mathbf{V}" + }, + { + "bbox": [ + 131, + 562, + 479, + 586 + ], + "type": "text", + "content": ", illustrating the effectiveness of each rendering component." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 131, + 605, + 230, + 618 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 131, + 605, + 230, + 618 + ], + "spans": [ + { + "bbox": [ + 131, + 605, + 230, + 618 + ], + "type": "text", + "content": "4 Experiments" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 130, + 629, + 480, + 665 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 629, + 480, + 665 + ], + "spans": [ + { + "bbox": [ + 130, + 629, + 480, + 665 + ], + "type": "text", + "content": "Implementation details. For physical body simulation, we use a 2D rigid body physics simulator Pymunk [13]. For each experiment, we run 120 simulation steps and uniformly sample 16 frames to form the videos. We set the" + } + ] + } + ], + "index": 9 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 268, + 91, + 447, + 102 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 268, + 91, + 447, + 102 + ], + "spans": [ + { + "bbox": [ + 268, + 91, + 447, + 102 + ], + "type": "text", + "content": "PhysGen: Rigid Physics-Grounded Img2Vid" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 474, + 92, + 480, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 474, + 92, + 480, + 100 + ], + "spans": [ + { + "bbox": [ + 474, + 92, + 480, + 100 + ], + "type": "text", + "content": "9" + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 8 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 140, + 114, + 478, + 400 + ], + "blocks": [ + { + "bbox": [ + 140, + 114, + 478, + 400 + ], + "lines": [ + { + "bbox": [ + 140, + 114, + 478, + 400 + ], + "spans": [ + { + "bbox": [ + 140, + 114, + 478, + 400 + ], + "type": "image", + "image_path": "cb2e40043b69b091db787a08619c2f98e58996c50081e24ce8cae7a305a2026a.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 132, + 401, + 480, + 423 + ], + "lines": [ + { + "bbox": [ + 132, + 401, + 480, + 423 + ], + "spans": [ + { + "bbox": [ + 132, + 401, + 480, + 423 + ], + "type": "text", + "content": "Fig. 4: Video generation results. Left: Input initial frame with red arrows representing the initial force or speed on the object. Right: Generated future frames." + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_caption" + } + ], + "index": 2 + }, + { + "bbox": [ + 130, + 455, + 482, + 504 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 455, + 482, + 504 + ], + "spans": [ + { + "bbox": [ + 130, + 455, + 482, + 504 + ], + "type": "text", + "content": "generation resolution at " + }, + { + "bbox": [ + 130, + 455, + 482, + 504 + ], + "type": "inline_equation", + "content": "512 \\times 512" + }, + { + "bbox": [ + 130, + 455, + 482, + 504 + ], + "type": "text", + "content": " so that frames can easily fit into the diffusion models. For video diffusion prior model, we use SEINE [18]. For inference, DDIM sampling [67] is used and the noise strength " + }, + { + "bbox": [ + 130, + 455, + 482, + 504 + ], + "type": "inline_equation", + "content": "i" + }, + { + "bbox": [ + 130, + 455, + 482, + 504 + ], + "type": "text", + "content": " set to " + }, + { + "bbox": [ + 130, + 455, + 482, + 504 + ], + "type": "inline_equation", + "content": "s = 0.5" + }, + { + "bbox": [ + 130, + 455, + 482, + 504 + ], + "type": "text", + "content": ", i.e., executing 25/50 steps for denoising. The latents fusion stops at timestamp " + }, + { + "bbox": [ + 130, + 455, + 482, + 504 + ], + "type": "inline_equation", + "content": "\\delta = 5" + }, + { + "bbox": [ + 130, + 455, + 482, + 504 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 132, + 514, + 482, + 538 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 514, + 482, + 538 + ], + "spans": [ + { + "bbox": [ + 132, + 514, + 482, + 538 + ], + "type": "text", + "content": "Speed. PhysGen's run-time speed is 3 minutes for one image-to-video gen on a Nvidia A40 GPU. Image understanding and relighting takes the most time." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 130, + 548, + 482, + 632 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 548, + 482, + 632 + ], + "spans": [ + { + "bbox": [ + 130, + 548, + 482, + 632 + ], + "type": "text", + "content": "Data. To show the generalizability and robustness of our method, we use both internet data and self-captured indoor images from a cell phone. The dataset is diverse, with variations in lighting, object count, geometries, physical attributes, and environmental boundaries. We compare different approaches on 15 photos that require physical reasoning. Additionally, we collect 50 recorded videos of a given scene as ground truth by varying input conditions for quantitative comparison. Randomly selected sequences are shown in the supplementary material." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 132, + 641, + 481, + 666 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 641, + 481, + 666 + ], + "spans": [ + { + "bbox": [ + 132, + 641, + 481, + 666 + ], + "type": "text", + "content": "Baselines. We compare against two streams of approaches. The first stream is current state-of-the-art I2V models: SEINE [18], I2VGen-XL [95] and Dynam-" + } + ] + } + ], + "index": 7 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "spans": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "type": "text", + "content": "10" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 166, + 91, + 215, + 100 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 166, + 91, + 215, + 100 + ], + "spans": [ + { + "bbox": [ + 166, + 91, + 215, + 100 + ], + "type": "text", + "content": "S. Liu et al." + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 9 + }, + { + "para_blocks": [ + { + "type": "table", + "bbox": [ + 183, + 114, + 433, + 174 + ], + "blocks": [ + { + "bbox": [ + 183, + 114, + 433, + 174 + ], + "lines": [ + { + "bbox": [ + 183, + 114, + 433, + 174 + ], + "spans": [ + { + "bbox": [ + 183, + 114, + 433, + 174 + ], + "type": "table", + "html": "
MethodsPhysical-realism↑Photo-realism↑
SEINE [18]1.391.86
DynamiCrafter [87]1.681.81
I2VGen-XL [95]2.112.25
Ours4.143.86
", + "image_path": "9ce64eef3463ca38a956bcad7c55ca7ddcf10ffd0b5a889f872bdfde2e50357c.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "table_body" + } + ], + "index": 2 + }, + { + "type": "table", + "bbox": [ + 187, + 240, + 427, + 300 + ], + "blocks": [ + { + "bbox": [ + 130, + 175, + 482, + 218 + ], + "lines": [ + { + "bbox": [ + 130, + 175, + 482, + 218 + ], + "spans": [ + { + "bbox": [ + 130, + 175, + 482, + 218 + ], + "type": "text", + "content": "Table 1: Human evaluation. We evaluate both physical-realism and photo-realism of the three competing image-to-video models. The user is asked to evaluate the generated video using a 5-point scale from strongly disagree (1) to strongly agree(5) that the video is physical-realistic and photo-realistic. Our method ranks the first in both terms." + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 187, + 240, + 427, + 300 + ], + "lines": [ + { + "bbox": [ + 187, + 240, + 427, + 300 + ], + "spans": [ + { + "bbox": [ + 187, + 240, + 427, + 300 + ], + "type": "table", + "html": "
MethodsImage-FID↓Motion-FID↓
SEINE [18]138.8938.76
DynamiCrafter [87]99.64109.61
I2VGen-XL [95]104.6282.04
Ours105.7030.20
", + "image_path": "afe00fb08c0a22f845ba4658f1e3226cb49a1fd98ab56830e94e21be0a00a4e8.jpg" + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "table_body" + } + ], + "index": 4 + }, + { + "bbox": [ + 130, + 301, + 482, + 334 + ], + "lines": [ + { + "bbox": [ + 130, + 301, + 482, + 334 + ], + "spans": [ + { + "bbox": [ + 130, + 301, + 482, + 334 + ], + "type": "text", + "content": "Table 2: Quantitative evaluation. We measure the Image-FID and Motion-FID of videos generated by four methods against the collected GT videos for the given scene. Our method achieves both low Image-FID and Motion-FID." + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "text" + }, + { + "bbox": [ + 130, + 369, + 482, + 453 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 369, + 482, + 453 + ], + "spans": [ + { + "bbox": [ + 130, + 369, + 482, + 453 + ], + "type": "text", + "content": "iCrafter [87]. These methods take both an image and a text prompt as input for generation. The text prompt is generated by GPT-4V [56] to describe the future events. Another stream we compare is the image-based manipulation approach Motion Guidance [28]. This method uses an image and optical flow as inputs to predict future movement and generate corresponding images. We employ computed motion flow from our physical motion module as the target input for Motion Guidance to optimize each video frame." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 131, + 473, + 198, + 484 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 131, + 473, + 198, + 484 + ], + "spans": [ + { + "bbox": [ + 131, + 473, + 198, + 484 + ], + "type": "text", + "content": "4.1 Results" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 130, + 495, + 481, + 544 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 495, + 481, + 544 + ], + "spans": [ + { + "bbox": [ + 130, + 495, + 481, + 544 + ], + "type": "text", + "content": "We show our generated results on 8 different physical procedures from multiple sources in Fig. 4. Our method could simulate reasonable complex physical procedures which involves object-object interaction, object-scene interaction with different physical properties and initial conditions." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 130, + 557, + 482, + 666 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 557, + 482, + 666 + ], + "spans": [ + { + "bbox": [ + 130, + 557, + 482, + 666 + ], + "type": "text", + "content": "Visual comparisons. The comparison results are shown in Fig. 5. As can be seen, image-video generation methods could not produce any physically plausible futures. For each method, even we carefully tune the input prompt, the models couldn't understand the direction, physical meanings and interactions. All 3 models fail to synthesize realistic videos. For motion guidance with accurate motion flow provided, the results is better but brings a lot inconsistency. When there are multiple objects interaction, the method suffers from introducing new contents. For the domino case, the generation is not smooth and contain artifacts as it could not accurately model the physical procedure." + } + ] + } + ], + "index": 9 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 268, + 91, + 447, + 102 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 268, + 91, + 447, + 102 + ], + "spans": [ + { + "bbox": [ + 268, + 91, + 447, + 102 + ], + "type": "text", + "content": "PhysGen: Rigid Physics-Grounded Img2Vid" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 470, + 92, + 479, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 470, + 92, + 479, + 100 + ], + "spans": [ + { + "bbox": [ + 470, + 92, + 479, + 100 + ], + "type": "text", + "content": "11" + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 10 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 138, + 113, + 478, + 653 + ], + "blocks": [ + { + "bbox": [ + 138, + 113, + 478, + 653 + ], + "lines": [ + { + "bbox": [ + 138, + 113, + 478, + 653 + ], + "spans": [ + { + "bbox": [ + 138, + 113, + 478, + 653 + ], + "type": "image", + "image_path": "86f9d79b7cd6780ddd6e334eefc6976736fa3d27fbfb61d66444fb372b94826a.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 132, + 654, + 481, + 677 + ], + "lines": [ + { + "bbox": [ + 132, + 654, + 481, + 677 + ], + "spans": [ + { + "bbox": [ + 132, + 654, + 481, + 677 + ], + "type": "text", + "content": "Fig.5: Qualitative comparison. against DynamiciCafter(DC) [87], I2VGenXL(VGen) [95], SEINE [18], Motion Guidance(MG) [28]." + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_caption" + } + ], + "index": 2 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "spans": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "type": "text", + "content": "12" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 166, + 91, + 215, + 100 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 166, + 91, + 215, + 100 + ], + "spans": [ + { + "bbox": [ + 166, + 91, + 215, + 100 + ], + "type": "text", + "content": "S. Liu et al." + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 11 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 141, + 125, + 194, + 178 + ], + "blocks": [ + { + "bbox": [ + 156, + 115, + 178, + 124 + ], + "lines": [ + { + "bbox": [ + 156, + 115, + 178, + 124 + ], + "spans": [ + { + "bbox": [ + 156, + 115, + 178, + 124 + ], + "type": "text", + "content": "Input" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 141, + 125, + 194, + 178 + ], + "lines": [ + { + "bbox": [ + 141, + 125, + 194, + 178 + ], + "spans": [ + { + "bbox": [ + 141, + 125, + 194, + 178 + ], + "type": "image", + "image_path": "75290295add3e73aeb644eb061bdea4d415f0c229eaccfb873e5d6f42931b8e2.jpg" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_body" + } + ], + "index": 3 + }, + { + "type": "image", + "bbox": [ + 141, + 178, + 194, + 233 + ], + "blocks": [ + { + "bbox": [ + 141, + 178, + 194, + 233 + ], + "lines": [ + { + "bbox": [ + 141, + 178, + 194, + 233 + ], + "spans": [ + { + "bbox": [ + 141, + 178, + 194, + 233 + ], + "type": "image", + "image_path": "0ada98210e6442fdb327fc985cd08265f97727fa9f72c81e59fc8a07354ba93f.jpg" + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 137, + 234, + 476, + 246 + ], + "lines": [ + { + "bbox": [ + 137, + 234, + 476, + 246 + ], + "spans": [ + { + "bbox": [ + 137, + 234, + 476, + 246 + ], + "type": "text", + "content": "Fig. 6: Controllability. PhysGen generates diverse results reflecting initial forces." + } + ] + } + ], + "index": 16, + "angle": 0, + "type": "image_caption" + } + ], + "index": 4 + }, + { + "type": "image", + "bbox": [ + 195, + 125, + 250, + 178 + ], + "blocks": [ + { + "bbox": [ + 195, + 125, + 250, + 178 + ], + "lines": [ + { + "bbox": [ + 195, + 125, + 250, + 178 + ], + "spans": [ + { + "bbox": [ + 195, + 125, + 250, + 178 + ], + "type": "image", + "image_path": "bb0941cae9800451da6ab728f2ac94f5d0b17b8c65dd8a81f9a082ace7a692af.jpg" + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "image_body" + } + ], + "index": 5 + }, + { + "type": "image", + "bbox": [ + 196, + 178, + 250, + 233 + ], + "blocks": [ + { + "bbox": [ + 196, + 178, + 250, + 233 + ], + "lines": [ + { + "bbox": [ + 196, + 178, + 250, + 233 + ], + "spans": [ + { + "bbox": [ + 196, + 178, + 250, + 233 + ], + "type": "image", + "image_path": "fece2730527167282bfab2ae3f6c4009f48e92632b0f3064b62b7e511991bda8.jpg" + } + ] + } + ], + "index": 6, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 207, + 267, + 239, + 277 + ], + "lines": [ + { + "bbox": [ + 207, + 267, + 239, + 277 + ], + "spans": [ + { + "bbox": [ + 207, + 267, + 239, + 277 + ], + "type": "text", + "content": "FG Seg" + } + ] + } + ], + "index": 19, + "angle": 0, + "type": "image_caption" + } + ], + "index": 6 + }, + { + "type": "image", + "bbox": [ + 251, + 125, + 305, + 178 + ], + "blocks": [ + { + "bbox": [ + 268, + 115, + 403, + 125 + ], + "lines": [ + { + "bbox": [ + 268, + 115, + 403, + 125 + ], + "spans": [ + { + "bbox": [ + 268, + 115, + 403, + 125 + ], + "type": "text", + "content": "Generation (left " + }, + { + "bbox": [ + 268, + 115, + 403, + 125 + ], + "type": "inline_equation", + "content": "\\rightarrow" + }, + { + "bbox": [ + 268, + 115, + 403, + 125 + ], + "type": "text", + "content": " right: time steps)" + } + ] + } + ], + "index": 7, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 251, + 125, + 305, + 178 + ], + "lines": [ + { + "bbox": [ + 251, + 125, + 305, + 178 + ], + "spans": [ + { + "bbox": [ + 251, + 125, + 305, + 178 + ], + "type": "image", + "image_path": "d5729ad87430b019747f7872a6a90ad394ee13475c6e17744e039b1c45073683.jpg" + } + ] + } + ], + "index": 8, + "angle": 0, + "type": "image_body" + } + ], + "index": 8 + }, + { + "type": "image", + "bbox": [ + 251, + 178, + 305, + 233 + ], + "blocks": [ + { + "bbox": [ + 251, + 178, + 305, + 233 + ], + "lines": [ + { + "bbox": [ + 251, + 178, + 305, + 233 + ], + "spans": [ + { + "bbox": [ + 251, + 178, + 305, + 233 + ], + "type": "image", + "image_path": "eecd168f4f0ae2d75c09dd900b64a0669d1ce990e352a2f3f7ac60bbe09e2984.jpg" + } + ] + } + ], + "index": 9, + "angle": 0, + "type": "image_body" + } + ], + "index": 9 + }, + { + "type": "image", + "bbox": [ + 307, + 125, + 360, + 178 + ], + "blocks": [ + { + "bbox": [ + 307, + 125, + 360, + 178 + ], + "lines": [ + { + "bbox": [ + 307, + 125, + 360, + 178 + ], + "spans": [ + { + "bbox": [ + 307, + 125, + 360, + 178 + ], + "type": "image", + "image_path": "350894b1c5b746a2627a025fba7843bb188198ae686cfe046087138af8db936d.jpg" + } + ] + } + ], + "index": 10, + "angle": 0, + "type": "image_body" + } + ], + "index": 10 + }, + { + "type": "image", + "bbox": [ + 307, + 178, + 360, + 233 + ], + "blocks": [ + { + "bbox": [ + 307, + 178, + 360, + 233 + ], + "lines": [ + { + "bbox": [ + 307, + 178, + 360, + 233 + ], + "spans": [ + { + "bbox": [ + 307, + 178, + 360, + 233 + ], + "type": "image", + "image_path": "76a57d3144cd17a6649d739b23543b32212590e7c1cba0ba8e0057184f9270d2.jpg" + } + ] + } + ], + "index": 11, + "angle": 0, + "type": "image_body" + } + ], + "index": 11 + }, + { + "type": "image", + "bbox": [ + 361, + 125, + 415, + 178 + ], + "blocks": [ + { + "bbox": [ + 361, + 125, + 415, + 178 + ], + "lines": [ + { + "bbox": [ + 361, + 125, + 415, + 178 + ], + "spans": [ + { + "bbox": [ + 361, + 125, + 415, + 178 + ], + "type": "image", + "image_path": "85d759259bf531df7be2066b72aba3fd3c2516399599fd455ff1a8912585c0c5.jpg" + } + ] + } + ], + "index": 12, + "angle": 0, + "type": "image_body" + } + ], + "index": 12 + }, + { + "type": "image", + "bbox": [ + 361, + 178, + 415, + 233 + ], + "blocks": [ + { + "bbox": [ + 361, + 178, + 415, + 233 + ], + "lines": [ + { + "bbox": [ + 361, + 178, + 415, + 233 + ], + "spans": [ + { + "bbox": [ + 361, + 178, + 415, + 233 + ], + "type": "image", + "image_path": "c7133f9ef0c8294c7ee20ac3488e9463d06447bc5630dd661ff692df556b28c2.jpg" + } + ] + } + ], + "index": 13, + "angle": 0, + "type": "image_body" + } + ], + "index": 13 + }, + { + "type": "image", + "bbox": [ + 417, + 125, + 470, + 178 + ], + "blocks": [ + { + "bbox": [ + 417, + 125, + 470, + 178 + ], + "lines": [ + { + "bbox": [ + 417, + 125, + 470, + 178 + ], + "spans": [ + { + "bbox": [ + 417, + 125, + 470, + 178 + ], + "type": "image", + "image_path": "bfa0631fb8f6a922ab56424f6a2056b93b29da966aa546d2fee99b1634ac0e03.jpg" + } + ] + } + ], + "index": 14, + "angle": 0, + "type": "image_body" + } + ], + "index": 14 + }, + { + "type": "image", + "bbox": [ + 417, + 178, + 470, + 233 + ], + "blocks": [ + { + "bbox": [ + 417, + 178, + 470, + 233 + ], + "lines": [ + { + "bbox": [ + 417, + 178, + 470, + 233 + ], + "spans": [ + { + "bbox": [ + 417, + 178, + 470, + 233 + ], + "type": "image", + "image_path": "7fe317890fe4aa7a590abb8a2ebb9997b74dfba92fd73df248cec690370f7926.jpg" + } + ] + } + ], + "index": 15, + "angle": 0, + "type": "image_body" + } + ], + "index": 15 + }, + { + "type": "image", + "bbox": [ + 136, + 277, + 194, + 335 + ], + "blocks": [ + { + "bbox": [ + 151, + 267, + 175, + 277 + ], + "lines": [ + { + "bbox": [ + 151, + 267, + 175, + 277 + ], + "spans": [ + { + "bbox": [ + 151, + 267, + 175, + 277 + ], + "type": "text", + "content": "Input" + } + ] + } + ], + "index": 17, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 136, + 277, + 194, + 335 + ], + "lines": [ + { + "bbox": [ + 136, + 277, + 194, + 335 + ], + "spans": [ + { + "bbox": [ + 136, + 277, + 194, + 335 + ], + "type": "image", + "image_path": "97f771c717e7a716e883b0480a19f8c40ff0cd23b88dd83aab83eb23b47e53d0.jpg" + } + ] + } + ], + "index": 18, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 131, + 338, + 480, + 360 + ], + "lines": [ + { + "bbox": [ + 131, + 338, + 480, + 360 + ], + "spans": [ + { + "bbox": [ + 131, + 338, + 480, + 360 + ], + "type": "text", + "content": "Fig. 7: Qualitative results from the perception module. The results include foreground and background mask, inpainting output, normal map and inferred physics." + } + ] + } + ], + "index": 29, + "angle": 0, + "type": "image_caption" + } + ], + "index": 18 + }, + { + "type": "image", + "bbox": [ + 211, + 297, + 227, + 312 + ], + "blocks": [ + { + "bbox": [ + 211, + 297, + 227, + 312 + ], + "lines": [ + { + "bbox": [ + 211, + 297, + 227, + 312 + ], + "spans": [ + { + "bbox": [ + 211, + 297, + 227, + 312 + ], + "type": "image", + "image_path": "21ff867b20ac2c770e2e4a9292597255d91a9082e96d6e8c83c3f52b8ce4f8f7.jpg" + } + ] + } + ], + "index": 20, + "angle": 0, + "type": "image_body" + } + ], + "index": 20 + }, + { + "type": "image", + "bbox": [ + 253, + 277, + 310, + 335 + ], + "blocks": [ + { + "bbox": [ + 266, + 267, + 298, + 277 + ], + "lines": [ + { + "bbox": [ + 266, + 267, + 298, + 277 + ], + "spans": [ + { + "bbox": [ + 266, + 267, + 298, + 277 + ], + "type": "text", + "content": "BG Seg" + } + ] + } + ], + "index": 21, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 253, + 277, + 310, + 335 + ], + "lines": [ + { + "bbox": [ + 253, + 277, + 310, + 335 + ], + "spans": [ + { + "bbox": [ + 253, + 277, + 310, + 335 + ], + "type": "image", + "image_path": "9770c9b00fde4df2fe55bff534a1a120add4a9282721b8f57215f3b116d9fd60.jpg" + } + ] + } + ], + "index": 22, + "angle": 0, + "type": "image_body" + } + ], + "index": 22 + }, + { + "type": "image", + "bbox": [ + 316, + 277, + 369, + 335 + ], + "blocks": [ + { + "bbox": [ + 320, + 268, + 362, + 277 + ], + "lines": [ + { + "bbox": [ + 320, + 268, + 362, + 277 + ], + "spans": [ + { + "bbox": [ + 320, + 268, + 362, + 277 + ], + "type": "text", + "content": "Inpainting" + } + ] + } + ], + "index": 23, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 316, + 277, + 369, + 335 + ], + "lines": [ + { + "bbox": [ + 316, + 277, + 369, + 335 + ], + "spans": [ + { + "bbox": [ + 316, + 277, + 369, + 335 + ], + "type": "image", + "image_path": "36d369f2ffd32d1ed180a681022124479dad95114b84f666b395b79e3a4f1802.jpg" + } + ] + } + ], + "index": 24, + "angle": 0, + "type": "image_body" + } + ], + "index": 24 + }, + { + "type": "image", + "bbox": [ + 371, + 284, + 422, + 335 + ], + "blocks": [ + { + "bbox": [ + 381, + 268, + 412, + 277 + ], + "lines": [ + { + "bbox": [ + 381, + 268, + 412, + 277 + ], + "spans": [ + { + "bbox": [ + 381, + 268, + 412, + 277 + ], + "type": "text", + "content": "Normal" + } + ] + } + ], + "index": 25, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 371, + 284, + 422, + 335 + ], + "lines": [ + { + "bbox": [ + 371, + 284, + 422, + 335 + ], + "spans": [ + { + "bbox": [ + 371, + 284, + 422, + 335 + ], + "type": "image", + "image_path": "f8dce019dcc5a58c1eb711628fa0e9f7c0f9db3f935dd834c77cab1f759d1eb7.jpg" + } + ] + } + ], + "index": 26, + "angle": 0, + "type": "image_body" + } + ], + "index": 26 + }, + { + "type": "image", + "bbox": [ + 432, + 284, + 482, + 335 + ], + "blocks": [ + { + "bbox": [ + 424, + 268, + 490, + 277 + ], + "lines": [ + { + "bbox": [ + 424, + 268, + 490, + 277 + ], + "spans": [ + { + "bbox": [ + 424, + 268, + 490, + 277 + ], + "type": "text", + "content": "inferred Physics" + } + ] + } + ], + "index": 27, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 432, + 284, + 482, + 335 + ], + "lines": [ + { + "bbox": [ + 432, + 284, + 482, + 335 + ], + "spans": [ + { + "bbox": [ + 432, + 284, + 482, + 335 + ], + "type": "image", + "image_path": "13f7e5e97e3cc966293330450a174d2c7a586dcc4baec7e322f96e6e1f7ddde2.jpg" + } + ] + } + ], + "index": 28, + "angle": 0, + "type": "image_body" + } + ], + "index": 28 + }, + { + "bbox": [ + 132, + 393, + 253, + 404 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 393, + 253, + 404 + ], + "spans": [ + { + "bbox": [ + 132, + 393, + 253, + 404 + ], + "type": "text", + "content": "4.2 Human evaluation" + } + ] + } + ], + "index": 30 + }, + { + "bbox": [ + 130, + 413, + 485, + 472 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 413, + 485, + 472 + ], + "spans": [ + { + "bbox": [ + 130, + 413, + 485, + 472 + ], + "type": "text", + "content": "To comprehensively assess both the physical-realism and photo-realism of the generated videos, we conducted human evaluations following a similar methodology to prior work [16,78,83,91]. Here by physical-realism we mean whether the content of the generated video is realistic in its dynamics and object interactions and accurately reflects the instructed movements." + } + ] + } + ], + "index": 31 + }, + { + "bbox": [ + 130, + 473, + 482, + 605 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 473, + 482, + 605 + ], + "spans": [ + { + "bbox": [ + 130, + 473, + 482, + 605 + ], + "type": "text", + "content": "Specifically, we include 15 videos with diverse objects and conditions for evaluation, compared against other three I2V model generations (SEINE [18], DynamiCrafter [87], I2VGen-XL [95]). For each compared baseline method, we meticulously adjusted the input prompts and conducted multiple runs to select the most plausible output. In contrast, our method did not undergo specific tuning for each video. We randomized the presentation order and asked 14 participants to rate the videos on a five-point scale, from strongly disagree (1) to strongly agree (5) for physical-realism and photo-realism. The results in Tab. 1 show that our method achieves the highest ratings. Our average scores for physical-realism (4.14) and photo-realism (3.86) fall within the Agree level, while other methods perform poorly." + } + ] + } + ], + "index": 32 + }, + { + "bbox": [ + 131, + 622, + 279, + 634 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 131, + 622, + 279, + 634 + ], + "spans": [ + { + "bbox": [ + 131, + 622, + 279, + 634 + ], + "type": "text", + "content": "4.3 Quantitative evaluation" + } + ] + } + ], + "index": 33 + }, + { + "bbox": [ + 130, + 641, + 481, + 665 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 641, + 481, + 665 + ], + "spans": [ + { + "bbox": [ + 130, + 641, + 481, + 665 + ], + "type": "text", + "content": "We quantitatively evaluate the collected GT videos of a given scene. The GT data includes 833 images. For each method, we generate 50 videos with random" + } + ] + } + ], + "index": 34 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 268, + 91, + 447, + 102 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 268, + 91, + 447, + 102 + ], + "spans": [ + { + "bbox": [ + 268, + 91, + 447, + 102 + ], + "type": "text", + "content": "PhysGen: Rigid Physics-Grounded Img2Vid" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 470, + 91, + 480, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 470, + 91, + 480, + 100 + ], + "spans": [ + { + "bbox": [ + 470, + 91, + 480, + 100 + ], + "type": "text", + "content": "13" + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 12 + }, + { + "para_blocks": [ + { + "bbox": [ + 130, + 116, + 482, + 260 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 116, + 482, + 260 + ], + "spans": [ + { + "bbox": [ + 130, + 116, + 482, + 260 + ], + "type": "text", + "content": "samples, collecting 800 images per method. Our method varies initial conditions to ensure all generated videos are different. Following [12], we adopt the Fréchet Inception Distance (FID) [32] as evaluation metric. We didn't use Fréchet Video Distance (FVD) [73] due to the small sample size and the goal to disentangle evaluation of appearance and motion. We evaluate appearance using Image-FID and motion using Motion-FID. For Motion-FID, we use RAFT [71] to extract and colorize optical flow [5], then compute FID from these rendered images. The results are shown in Tab. 2. We notice other methods which either generate minimal motion, leading to low Image-FID and high Motion-FID (e.g., DynamiCrafter [87]), or produce realistic motion but fail to maintain image consistency, resulting in low Motion-FID and high Image-FID (e.g., SEINE [18]). Our approach achieves both low Image-FID and Motion-FID compared to others." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 132, + 276, + 204, + 288 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 276, + 204, + 288 + ], + "spans": [ + { + "bbox": [ + 132, + 276, + 204, + 288 + ], + "type": "text", + "content": "4.4 Analysis" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 130, + 294, + 482, + 354 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 294, + 482, + 354 + ], + "spans": [ + { + "bbox": [ + 130, + 294, + 482, + 354 + ], + "type": "text", + "content": "Perception evaluation. We evaluate the perception module using 10 complex open-world images with 118 annotated movable instances (COCO [47] format). Our system achieves 0.93 precision, 0.82 recall at 0.5 IoU. More details are shown in the supplementary. Fig. 7 shows intermediate outputs from our perception module and simulation module." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 131, + 364, + 482, + 413 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 131, + 364, + 482, + 413 + ], + "spans": [ + { + "bbox": [ + 131, + 364, + 482, + 413 + ], + "type": "text", + "content": "Physical reasoning. We compare the physical parameters estimated by GPT-4V to random values sampled from human-estimated ranges. Without GPT-4V, Image-FID increases from 105.70 to 111.01, Motion-FID from 30.20 to 36.60. It shows the physical reasoning are crucial for appearance and motion realism." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 131, + 422, + 482, + 483 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 131, + 422, + 482, + 483 + ], + "spans": [ + { + "bbox": [ + 131, + 422, + 482, + 483 + ], + "type": "text", + "content": "Controllability. We showcase the controllability and diversity of our method in Fig. 6 by varying the initial conditions. For different rows, we adjust the direction and magnitude of initial speed and forces. The results show diverse physical-plausible trajectories. This demonstrates the potential of applying our method for diverse physical world generations." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 131, + 493, + 482, + 554 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 131, + 493, + 482, + 554 + ], + "spans": [ + { + "bbox": [ + 131, + 493, + 482, + 554 + ], + "type": "text", + "content": "Limitations. Our method is an initial step towards physics-based video generation but has two main limitations: it focuses mainly on rigid objects, limiting its application to non-rigid ones, and it lacks a comprehensive 3D understanding, making it unable to handle out-of-plane motions. We leave leveraging deformable physics and full 3D understanding as future work." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 132, + 570, + 220, + 582 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 570, + 220, + 582 + ], + "spans": [ + { + "bbox": [ + 132, + 570, + 220, + 582 + ], + "type": "text", + "content": "5 Conclusion" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 130, + 594, + 482, + 666 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 594, + 482, + 666 + ], + "spans": [ + { + "bbox": [ + 130, + 594, + 482, + 666 + ], + "type": "text", + "content": "We present PhysGen, a novel image-to-video generation method that turns a static image into a high-fidelity, physics-grounded, and temporally coherent videos. The main module is a model-based physical simulator with a data-driven video generation process. PhysGen enables a series of controllable and interactive downstream applications. PhysGen provide a new image-to-video generation paradigm and hopefully can inspire more follow-up works." + } + ] + } + ], + "index": 9 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "spans": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "type": "text", + "content": "14" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 166, + 91, + 216, + 101 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 166, + 91, + 216, + 101 + ], + "spans": [ + { + "bbox": [ + 166, + 91, + 216, + 101 + ], + "type": "text", + "content": "S. Liu et al." + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 13 + }, + { + "para_blocks": [ + { + "bbox": [ + 133, + 114, + 246, + 129 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 114, + 246, + 129 + ], + "spans": [ + { + "bbox": [ + 133, + 114, + 246, + 129 + ], + "type": "text", + "content": "Acknowledgements" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 132, + 140, + 481, + 188 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 140, + 481, + 188 + ], + "spans": [ + { + "bbox": [ + 132, + 140, + 481, + 188 + ], + "type": "text", + "content": "This project is supported by NSF Awards #2331878, #2340254, and #2312102, the IBM IIDAI Grant, and an Intel Research Gift. We greatly appreciate the NCSA for providing computing resources. We thank Tianhang cheng for helpful discussions. We thank Emily Chen and Gloria Wang for proofreading." + } + ] + } + ], + "index": 3 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 268, + 90, + 447, + 102 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 268, + 90, + 447, + 102 + ], + "spans": [ + { + "bbox": [ + 268, + 90, + 447, + 102 + ], + "type": "text", + "content": "PhysGen: Rigid Physics-Grounded Img2Vid" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 470, + 91, + 481, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 470, + 91, + 481, + 100 + ], + "spans": [ + { + "bbox": [ + 470, + 91, + 481, + 100 + ], + "type": "text", + "content": "15" + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 14 + }, + { + "para_blocks": [ + { + "bbox": [ + 133, + 114, + 197, + 126 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 114, + 197, + 126 + ], + "spans": [ + { + "bbox": [ + 133, + 114, + 197, + 126 + ], + "type": "text", + "content": "References" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 138, + 141, + 480, + 665 + ], + "type": "list", + "angle": 0, + "index": 22, + "blocks": [ + { + "bbox": [ + 138, + 141, + 480, + 164 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 141, + 480, + 164 + ], + "spans": [ + { + "bbox": [ + 138, + 141, + 480, + 164 + ], + "type": "text", + "content": "1. Aberman, K., Weng, Y., Lischinski, D., Cohen-Or, D., Chen, B.: Unpaired motion style transfer from video to animation. ACM TOG (2020)" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 138, + 165, + 480, + 197 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 165, + 480, + 197 + ], + "spans": [ + { + "bbox": [ + 138, + 165, + 480, + 197 + ], + "type": "text", + "content": "2. Achiam, J., Adler, S., Agarwal, S., Ahmad, L., Akkaya, I., Aleman, F.L., Almeida, D., Altenschmidt, J., Altman, S., Anadkat, S., et al.: Gpt-4 technical report. arXiv preprint arXiv:2303.08774 (2023)" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 138, + 198, + 480, + 231 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 198, + 480, + 231 + ], + "spans": [ + { + "bbox": [ + 138, + 198, + 480, + 231 + ], + "type": "text", + "content": "3. Ajay, A., Bauza, M., Wu, J., Fazeli, N., Tenenbaum, J.B., Rodriguez, A., Kaelbling, L.P.: Combining Physical Simulators and Object-Based Networks for Control. In: ICRA (2019)" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 138, + 232, + 480, + 263 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 232, + 480, + 263 + ], + "spans": [ + { + "bbox": [ + 138, + 232, + 480, + 263 + ], + "type": "text", + "content": "4. Ajay, A., Wu, J., Fazeli, N., Bauza, M., Kaelbling, L.P., Tenenbaum, J.B., Rodriguez, A.: Augmenting Physical Simulators with Stochastic Neural Networks: Case Study of Planar Pushing and Bouncing. In: IROS (2018)" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 138, + 265, + 480, + 286 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 265, + 480, + 286 + ], + "spans": [ + { + "bbox": [ + 138, + 265, + 480, + 286 + ], + "type": "text", + "content": "5. Baker, S., Scharstein, D., Lewis, J.P., Roth, S., Black, M.J., Szeliski, R.: A database and evaluation methodology for optical flow. IJCV (2011)" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 138, + 287, + 480, + 319 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 287, + 480, + 319 + ], + "spans": [ + { + "bbox": [ + 138, + 287, + 480, + 319 + ], + "type": "text", + "content": "6. Bar-Tal, O., Chefer, H., Tov, O., Herrmann, C., Paiss, R., Zada, S., Ephrat, A., Hur, J., Li, Y., Michaeli, T., et al.: Lumiere: A space-time diffusion model for video generation. arXiv preprint arXiv:2401.12945 (2024)" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 138, + 321, + 480, + 342 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 321, + 480, + 342 + ], + "spans": [ + { + "bbox": [ + 138, + 321, + 480, + 342 + ], + "type": "text", + "content": "7. Battaglia, P., Pascanu, R., Lai, M., Jimenez Rezende, D., et al.: Interaction networks for learning about objects, relations and physics. NeurIPS (2016)" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 138, + 343, + 480, + 387 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 343, + 480, + 387 + ], + "spans": [ + { + "bbox": [ + 138, + 343, + 480, + 387 + ], + "type": "text", + "content": "8. Blattmann, A., Dockhorn, T., Kulal, S., Mendelevitch, D., Kilian, M., Lorenz, D., Levi, Y., English, Z., Voleti, V., Letts, A., et al.: Stable video diffusion: Scaling latent video diffusion models to large datasets. arXiv preprint arXiv:2311.15127 (2023)" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 138, + 388, + 480, + 408 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 388, + 480, + 408 + ], + "spans": [ + { + "bbox": [ + 138, + 388, + 480, + 408 + ], + "type": "text", + "content": "9. Blattmann, A., Milbich, T., Dorkenwald, M., Ommer, B.: ipoke: Poking a still image for controlled stochastic video synthesis. In: ICCV (2021)" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 138, + 410, + 480, + 430 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 410, + 480, + 430 + ], + "spans": [ + { + "bbox": [ + 138, + 410, + 480, + 430 + ], + "type": "text", + "content": "10. Blattmann, A., Milbich, T., Dorkenwald, M., Ommer, B.: Understanding object dynamics for interactive image-to-video synthesis. In: CVPR (2021)" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 138, + 432, + 480, + 453 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 432, + 480, + 453 + ], + "spans": [ + { + "bbox": [ + 138, + 432, + 480, + 453 + ], + "type": "text", + "content": "1. Blattmann, A., Milbich, T., Dorkenwald, M., Ommer, B.: Understanding object dynamics for interactive image-to-video synthesis. In: CVPR (2021)" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 138, + 454, + 480, + 486 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 454, + 480, + 486 + ], + "spans": [ + { + "bbox": [ + 138, + 454, + 480, + 486 + ], + "type": "text", + "content": "2. Blattmann, A., Rombach, R., Ling, H., Dockhorn, T., Kim, S.W., Fidler, S., Kreis, K.: Align your latents: High-resolution video synthesis with latent diffusion models. In: CVPR (2023)" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 138, + 487, + 365, + 498 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 487, + 365, + 498 + ], + "spans": [ + { + "bbox": [ + 138, + 487, + 365, + 498 + ], + "type": "text", + "content": "3. Blomqvist, V.: Pymunk. https://pymunk.org (2023)" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 138, + 499, + 480, + 520 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 499, + 480, + 520 + ], + "spans": [ + { + "bbox": [ + 138, + 499, + 480, + 520 + ], + "type": "text", + "content": "4. Bowen, R.S., Tucker, R., Zabih, R., Snavely, N.: Dimensions of motion: Monocular prediction through flow subspaces. In: 3DV (2022)" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 138, + 521, + 480, + 542 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 521, + 480, + 542 + ], + "spans": [ + { + "bbox": [ + 138, + 521, + 480, + 542 + ], + "type": "text", + "content": "5. Careaga, C., Miangoleh, S.M.H., Aksoy, Y.: Intrinsic harmonization for illumination-aware compositing. arXiv preprint arXiv:2312.03698 (2023)" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 138, + 544, + 480, + 576 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 544, + 480, + 576 + ], + "spans": [ + { + "bbox": [ + 138, + 544, + 480, + 576 + ], + "type": "text", + "content": "6. Chen, H., Xia, M., He, Y., Zhang, Y., Cun, X., Yang, S., Xing, J., Liu, Y., Chen, Q., Wang, X., et al.: Videocrafter1: Open diffusion models for high-quality video generation. arXiv preprint arXiv:2310.19512 (2023)" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 138, + 577, + 480, + 609 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 577, + 480, + 609 + ], + "spans": [ + { + "bbox": [ + 138, + 577, + 480, + 609 + ], + "type": "text", + "content": "7. Chen, X., Liu, Z., Chen, M., Feng, Y., Liu, Y., Shen, Y., Zhao, H.: Livephoto: Real image animation with text-guided motion control. arXiv preprint arXiv:2312.02928 (2023)" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 138, + 610, + 480, + 643 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 610, + 480, + 643 + ], + "spans": [ + { + "bbox": [ + 138, + 610, + 480, + 643 + ], + "type": "text", + "content": "8. Chen, X., Wang, Y., Zhang, L., Zhuang, S., Ma, X., Yu, J., Wang, Y., Lin, D., Qiao, Y., Liu, Z.: Seine: Short-to-long video diffusion model for generative transition and prediction. arXiv preprint arXiv:2310.20700 (2023)" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 138, + 643, + 480, + 665 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 643, + 480, + 665 + ], + "spans": [ + { + "bbox": [ + 138, + 643, + 480, + 665 + ], + "type": "text", + "content": "9. Chuang, Y.Y., Goldman, D.B., Zheng, K.C., Curless, B., Salesin, D.H., Szeliski, R.: Animating pictures with stochastic motion textures. ACM TOG (2005)" + } + ] + } + ], + "index": 21 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "spans": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "type": "text", + "content": "16" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 166, + 91, + 215, + 100 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 166, + 91, + 215, + 100 + ], + "spans": [ + { + "bbox": [ + 166, + 91, + 215, + 100 + ], + "type": "text", + "content": "S. Liu et al." + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 15 + }, + { + "para_blocks": [ + { + "bbox": [ + 130, + 116, + 480, + 665 + ], + "type": "list", + "angle": 0, + "index": 22, + "blocks": [ + { + "bbox": [ + 130, + 116, + 480, + 138 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 116, + 480, + 138 + ], + "spans": [ + { + "bbox": [ + 130, + 116, + 480, + 138 + ], + "type": "text", + "content": "20. Ciarlet, P.G., Lions, J.L.: Handbook of numerical analysis. Gulf Professional Publishing (1990)" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 132, + 139, + 480, + 170 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 139, + 480, + 170 + ], + "spans": [ + { + "bbox": [ + 132, + 139, + 480, + 170 + ], + "type": "text", + "content": "21. Davis, A., Bouman, K.L., Chen, J.G., Rubinstein, M., Durand, F., Freeman, W.T.: Visual vibrometry: Estimating material properties from small motion in video. In: CVPR (2015)" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 132, + 171, + 480, + 191 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 171, + 480, + 191 + ], + "spans": [ + { + "bbox": [ + 132, + 171, + 480, + 191 + ], + "type": "text", + "content": "22. Davis, A., Chen, J.G., Durand, F.: Image-space modal bases for plausible manipulation of objects in video. ACM TOG (2015)" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 132, + 192, + 480, + 213 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 192, + 480, + 213 + ], + "spans": [ + { + "bbox": [ + 132, + 192, + 480, + 213 + ], + "type": "text", + "content": "23. Eftekhar, A., Sax, A., Malik, J., Zamir, A.: Omnidata: A scalable pipeline for making multi-task mid-level vision datasets from 3d scans. In: ICCV (2021)" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 132, + 214, + 480, + 246 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 214, + 480, + 246 + ], + "spans": [ + { + "bbox": [ + 132, + 214, + 480, + 246 + ], + "type": "text", + "content": "24. Endo, Y., Kanamori, Y., Kuriyama, S.: Animating landscape: self-supervised learning of decoupled motion and appearance for single-image video synthesis. arXiv preprint arXiv:1910.07192 (2019)" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 132, + 247, + 480, + 267 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 247, + 480, + 267 + ], + "spans": [ + { + "bbox": [ + 132, + 247, + 480, + 267 + ], + "type": "text", + "content": "25. Friedland, B.: Control system design: an introduction to state-space methods. Courier Corporation (2012)" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 132, + 267, + 480, + 299 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 267, + 480, + 299 + ], + "spans": [ + { + "bbox": [ + 132, + 267, + 480, + 299 + ], + "type": "text", + "content": "26. Fu, X., Yin, W., Hu, M., Wang, K., Ma, Y., Tan, P., Shen, S., Lin, D., Long, X.: Geowizard: Unleashing the diffusion priors for 3d geometry estimation from a single image. arXiv preprint arXiv:2403.12013 (2024)" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 132, + 300, + 480, + 331 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 300, + 480, + 331 + ], + "spans": [ + { + "bbox": [ + 132, + 300, + 480, + 331 + ], + "type": "text", + "content": "27. Ge, S., Nah, S., Liu, G., Poon, T., Tao, A., Catanzaro, B., Jacobs, D., Huang, J.B., Liu, M.Y., Balaji, Y.: Preserve your own correlation: A noise prior for video diffusion models. In: ICCV (2023)" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 132, + 332, + 480, + 353 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 332, + 480, + 353 + ], + "spans": [ + { + "bbox": [ + 132, + 332, + 480, + 353 + ], + "type": "text", + "content": "28. Geng, D., Owens, A.: Motion guidance: Diffusion-based image editing with differentiable motion estimators. In: ICLR (2024)" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 132, + 354, + 480, + 385 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 354, + 480, + 385 + ], + "spans": [ + { + "bbox": [ + 132, + 354, + 480, + 385 + ], + "type": "text", + "content": "29. Girdhar, R., Singh, M., Brown, A., Duval, Q., Azadi, S., Rambhatla, S.S., Shah, A., Yin, X., Parikh, D., Misra, I.: Emu video: Factorizing text-to-video generation by explicit image conditioning. arXiv preprint arXiv:2311.10709 (2023)" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 132, + 386, + 480, + 418 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 386, + 480, + 418 + ], + "spans": [ + { + "bbox": [ + 132, + 386, + 480, + 418 + ], + "type": "text", + "content": "30. Guo, Y., Yang, C., Rao, A., Liang, Z., Wang, Y., Qiao, Y., Agrawala, M., Lin, D., Dai, B.: Animatediff: Animate your personalized text-to-image diffusion models without specific tuning (2023)" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 132, + 418, + 480, + 449 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 418, + 480, + 449 + ], + "spans": [ + { + "bbox": [ + 132, + 418, + 480, + 449 + ], + "type": "text", + "content": "31. Gupta, A., Yu, L., Sohn, K., Gu, X., Hahn, M., Fei-Fei, L., Essa, I., Jiang, L., Lezama, J.: Photorealistic video generation with diffusion models. arXiv preprint arXiv:2312.06662 (2023)" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 132, + 450, + 480, + 482 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 450, + 480, + 482 + ], + "spans": [ + { + "bbox": [ + 132, + 450, + 480, + 482 + ], + "type": "text", + "content": "32. Heusel, M., Ramsauer, H., Unterthiner, T., Nessler, B., Hochreiter, S.: Gans trained by a two time-scale update rule converge to a local nash equilibrium. NeurIPS (2017)" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 132, + 483, + 480, + 515 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 483, + 480, + 515 + ], + "spans": [ + { + "bbox": [ + 132, + 483, + 480, + 515 + ], + "type": "text", + "content": "33. Ho, J., Chan, W., Sahara, C., Whang, J., Gao, R., Gritsenko, A., Kingma, D.P., Poole, B., Norouzi, M., Fleet, D.J., et al.: Imagen video: High definition video generation with diffusion models. arXiv preprint arXiv:2210.02303 (2022)" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 132, + 515, + 480, + 536 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 515, + 480, + 536 + ], + "spans": [ + { + "bbox": [ + 132, + 515, + 480, + 536 + ], + "type": "text", + "content": "34. Ho, J., Jain, A., Abbeel, P.: Denoising diffusion probabilistic models. NeurIPS (2020)" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 132, + 536, + 480, + 557 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 536, + 480, + 557 + ], + "spans": [ + { + "bbox": [ + 132, + 536, + 480, + 557 + ], + "type": "text", + "content": "35. Holynski, A., Curless, B.L., Seitz, S.M., Szeliski, R.: Animating pictures with eulerian motion fields. In: CVPR (2021)" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 132, + 558, + 480, + 579 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 558, + 480, + 579 + ], + "spans": [ + { + "bbox": [ + 132, + 558, + 480, + 579 + ], + "type": "text", + "content": "36. Hu, Y., Anderson, L., Li, T.M., Sun, Q., Carr, N., Ragan-Kelley, J., Durand, F.: Difftaichi: Differentiable programming for physical simulation. ICLR (2020)" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 132, + 580, + 480, + 611 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 580, + 480, + 611 + ], + "spans": [ + { + "bbox": [ + 132, + 580, + 480, + 611 + ], + "type": "text", + "content": "37. Hu, Y., Fang, Y., Ge, Z., Qu, Z., Zhu, Y., Pradhana, A., Jiang, C.: A moving least squares material point method with displacement discontinuity and two-way rigid body coupling. ACM TOG (2018)" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 132, + 612, + 480, + 643 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 612, + 480, + 643 + ], + "spans": [ + { + "bbox": [ + 132, + 612, + 480, + 643 + ], + "type": "text", + "content": "38. Hu, Y., Li, T.M., Anderson, L., Ragan-Kelley, J., Durand, F.: Taichi: a language for high-performance computation on spatially sparse data structures. ACM TOG (2019)" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 132, + 643, + 480, + 665 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 643, + 480, + 665 + ], + "spans": [ + { + "bbox": [ + 132, + 643, + 480, + 665 + ], + "type": "text", + "content": "39. Jhou, W.C., Cheng, W.H.: Animating still landscape photographs through cloud motion creation. IEEE Transactions on Multimedia (2015)" + } + ] + } + ], + "index": 21 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 268, + 91, + 447, + 102 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 268, + 91, + 447, + 102 + ], + "spans": [ + { + "bbox": [ + 268, + 91, + 447, + 102 + ], + "type": "text", + "content": "PhysGen: Rigid Physics-Grounded Img2Vid" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 470, + 92, + 480, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 470, + 92, + 480, + 100 + ], + "spans": [ + { + "bbox": [ + 470, + 92, + 480, + 100 + ], + "type": "text", + "content": "17" + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 16 + }, + { + "para_blocks": [ + { + "bbox": [ + 132, + 116, + 480, + 666 + ], + "type": "list", + "angle": 0, + "index": 24, + "blocks": [ + { + "bbox": [ + 133, + 116, + 480, + 149 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 116, + 480, + 149 + ], + "spans": [ + { + "bbox": [ + 133, + 116, + 480, + 149 + ], + "type": "text", + "content": "40. Kirillov, A., Mintun, E., Ravi, N., Mao, H., Rolland, C., Gustafson, L., Xiao, T., Whitehead, S., Berg, A.C., Lo, W.Y., Dollár, P., Girshick, R.: Segment anything. arXiv:2304.02643 (2023)" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 132, + 150, + 480, + 172 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 150, + 480, + 172 + ], + "spans": [ + { + "bbox": [ + 132, + 150, + 480, + 172 + ], + "type": "text", + "content": "41. Koenig, N., Howard, A.: Design and use paradigms for gazebo, an open-source multi-robot simulator. In: IROS (2004)" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 132, + 172, + 480, + 205 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 172, + 480, + 205 + ], + "spans": [ + { + "bbox": [ + 132, + 172, + 480, + 205 + ], + "type": "text", + "content": "42. Kondratyuk, D., Yu, L., Gu, X., Lezama, J., Huang, J., Hornung, R., Adam, H., Akbari, H., Alon, Y., Birodkar, V., et al.: Videopoet: A large language model for zero-shot video generation. arXiv preprint arXiv:2312.14125 (2023)" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 132, + 205, + 480, + 237 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 205, + 480, + 237 + ], + "spans": [ + { + "bbox": [ + 132, + 205, + 480, + 237 + ], + "type": "text", + "content": "43. Li, X., Qiao, Y.L., Chen, P.Y., Jatavallabhula, K.M., Lin, M., Jiang, C., Gan, C.: Pac-nerf: Physics augmented continuum neural radiance fields for geometry-agnostic system identification. In: ICLR (2023)" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 132, + 237, + 480, + 259 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 237, + 480, + 259 + ], + "spans": [ + { + "bbox": [ + 132, + 237, + 480, + 259 + ], + "type": "text", + "content": "44. Li, Y., Lin, T., Yi, K., Bear, D., Yamins, D.L., Wu, J., Tenenbaum, J.B., Torralba, A.: Visual grounding of learned physical models. In: ICML (2020)" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 132, + 259, + 480, + 293 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 259, + 480, + 293 + ], + "spans": [ + { + "bbox": [ + 132, + 259, + 480, + 293 + ], + "type": "text", + "content": "45. Li, Y., Wu, J., Tedrake, R., Tenenbaum, J.B., Torralba, A.: Learning particle dynamics for manipulating rigid bodies, deformable objects, and fluids. In: ICLR (2019)" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 132, + 293, + 480, + 315 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 293, + 480, + 315 + ], + "spans": [ + { + "bbox": [ + 132, + 293, + 480, + 315 + ], + "type": "text", + "content": "46. Li, Z., Tucker, R., Snavely, N., Holynski, A.: Generative image dynamics. arXiv preprint arXiv:2309.07906 (2023)" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 132, + 315, + 480, + 336 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 315, + 480, + 336 + ], + "spans": [ + { + "bbox": [ + 132, + 315, + 480, + 336 + ], + "type": "text", + "content": "47. Lin, T.Y., Maire, M., Belongie, S., Hays, J., Perona, P., Ramanan, D., Dólar, P., Zitnick, C.L.: Microsoft coco: Common objects in context. In: ECCV (2014)" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 132, + 336, + 480, + 369 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 336, + 480, + 369 + ], + "spans": [ + { + "bbox": [ + 132, + 336, + 480, + 369 + ], + "type": "text", + "content": "48. Liu, S., Zeng, Z., Ren, T., Li, F., Zhang, H., Yang, J., Li, C., Yang, J., Su, H., Zhu, J., et al.: Grounding dino: Marrying dino with grounded pre-training for open-set object detection. arXiv preprint arXiv:2303.05499 (2023)" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 132, + 369, + 480, + 391 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 369, + 480, + 391 + ], + "spans": [ + { + "bbox": [ + 132, + 369, + 480, + 391 + ], + "type": "text", + "content": "49. Liu, T., Bargteil, A.W., O'Brien, J.F., Kavan, L.: Fast simulation of mass-spring systems. ACM TOG (2013)" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 132, + 391, + 480, + 424 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 391, + 480, + 424 + ], + "spans": [ + { + "bbox": [ + 132, + 391, + 480, + 424 + ], + "type": "text", + "content": "50. Lv, J., Huang, Y., Yan, M., Huang, J., Liu, J., Liu, Y., Wen, Y., Chen, X., Chen, S.: Gpt4motion: Scripting physical motions in text-to-video generation via blender-oriented gpt planning. In: CVPRW. pp. 1430-1440 (2024)" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 132, + 424, + 480, + 445 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 424, + 480, + 445 + ], + "spans": [ + { + "bbox": [ + 132, + 424, + 480, + 445 + ], + "type": "text", + "content": "51. Mahapatra, A., Kulkarni, K.: Controllable animation of fluid elements in still images. In: CVPR (2022)" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 132, + 446, + 480, + 468 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 446, + 480, + 468 + ], + "spans": [ + { + "bbox": [ + 132, + 446, + 480, + 468 + ], + "type": "text", + "content": "52. Mallya, A., Wang, T.C., Liu, M.Y.: Implicit warping for animation with image sets. In: NeurIPS (2022)" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 132, + 468, + 480, + 501 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 468, + 480, + 501 + ], + "spans": [ + { + "bbox": [ + 132, + 468, + 480, + 501 + ], + "type": "text", + "content": "53. Meng, C., He, Y., Song, Y., Song, J., Wu, J., Zhu, J.Y., Ermon, S.: SDEdit: Guided image synthesis and editing with stochastic differential equations. In: International Conference on Learning Representations (2022)" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 132, + 501, + 480, + 533 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 501, + 480, + 533 + ], + "spans": [ + { + "bbox": [ + 132, + 501, + 480, + 533 + ], + "type": "text", + "content": "54. Mrowca, D., Zhuang, C., Wang, E., Haber, N., Fei-Fei, L.F., Tenenbaum, J., Yamins, D.L.: Flexible neural representation for physics prediction. In: NeurIPS (2018)" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 132, + 533, + 458, + 544 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 533, + 458, + 544 + ], + "spans": [ + { + "bbox": [ + 132, + 533, + 458, + 544 + ], + "type": "text", + "content": "55. NVIDIA: Nvidia physx. https://developer.nvidia.com/physx-sdk (2019)" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 132, + 544, + 323, + 555 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 544, + 323, + 555 + ], + "spans": [ + { + "bbox": [ + 132, + 544, + 323, + 555 + ], + "type": "text", + "content": "56. OpenAI: Gpt-4v(ision) system card (2023)" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 132, + 555, + 432, + 567 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 555, + 432, + 567 + ], + "spans": [ + { + "bbox": [ + 132, + 555, + 432, + 567 + ], + "type": "text", + "content": "57. OpenAI: Creating video from text. https://openai.com/sora (2024)" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 132, + 567, + 480, + 588 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 567, + 480, + 588 + ], + "spans": [ + { + "bbox": [ + 132, + 567, + 480, + 588 + ], + "type": "text", + "content": "58. Peebles, W., Xie, S.: Scalable diffusion models with transformers. arXiv preprint arXiv:2212.09748 (2022)" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 132, + 588, + 480, + 610 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 588, + 480, + 610 + ], + "spans": [ + { + "bbox": [ + 132, + 588, + 480, + 610 + ], + "type": "text", + "content": "59. Popova, E., Popov, V.L.: The research works of coulomb and amontons and generalized laws of friction. Friction (2015)" + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 132, + 610, + 480, + 632 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 610, + 480, + 632 + ], + "spans": [ + { + "bbox": [ + 132, + 610, + 480, + 632 + ], + "type": "text", + "content": "60. Qiu, H., Chen, Z., Jiang, Y., Zhou, H., Fan, X., Yang, L., Wu, W., Liu, Z.: Relitalk: Relightable talking portrait generation from a single video (2023)" + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 132, + 632, + 480, + 666 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 632, + 480, + 666 + ], + "spans": [ + { + "bbox": [ + 132, + 632, + 480, + 666 + ], + "type": "text", + "content": "61. Ren, T., Liu, S., Zeng, A., Lin, J., Li, K., Cao, H., Chen, J., Huang, X., Chen, Y., Yan, F., Zeng, Z., Zhang, H., Li, F., Yang, J., Li, H., Jiang, Q., Zhang, L.: Grounded sam: Assembling open-world models for diverse visual tasks (2024)" + } + ] + } + ], + "index": 23 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "spans": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "type": "text", + "content": "18" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 166, + 91, + 215, + 100 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 166, + 91, + 215, + 100 + ], + "spans": [ + { + "bbox": [ + 166, + 91, + 215, + 100 + ], + "type": "text", + "content": "S. Liu et al." + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 17 + }, + { + "para_blocks": [ + { + "bbox": [ + 130, + 116, + 482, + 666 + ], + "type": "list", + "angle": 0, + "index": 24, + "blocks": [ + { + "bbox": [ + 130, + 116, + 482, + 138 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 116, + 482, + 138 + ], + "spans": [ + { + "bbox": [ + 130, + 116, + 482, + 138 + ], + "type": "text", + "content": "62. Rombach, R., Blattmann, A., Lorenz, D., Esser, P., Ommer, B.: High-resolution image synthesis with latent diffusion models. In: CVPR (2022)" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 130, + 138, + 482, + 171 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 138, + 482, + 171 + ], + "spans": [ + { + "bbox": [ + 130, + 138, + 482, + 171 + ], + "type": "text", + "content": "63. Sanchez-Gonzalez, A., Heess, N., Springenberg, J.T., Merel, J., Riedmiller, M., Hadsell, R., Battaglia, P.: Graph networks as learnable physics engines for inference and control. In: ICML (2018)" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 130, + 171, + 482, + 191 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 171, + 482, + 191 + ], + "spans": [ + { + "bbox": [ + 130, + 171, + 482, + 191 + ], + "type": "text", + "content": "64. Schödl, A., Szeliski, R., Salesin, D.H., Essa, I.: Video textures. In: PACMCGIT (2000)" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 130, + 191, + 482, + 224 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 191, + 482, + 224 + ], + "spans": [ + { + "bbox": [ + 130, + 191, + 482, + 224 + ], + "type": "text", + "content": "65. Singer, U., Polyak, A., Hayes, T., Yin, X., An, J., Zhang, S., Hu, Q., Yang, H., Ashual, O., Gafni, O., et al.: Make-a-video: Text-to-video generation without text-video data. arXiv preprint arXiv:2209.14792 (2022)" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 130, + 224, + 482, + 246 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 224, + 482, + 246 + ], + "spans": [ + { + "bbox": [ + 130, + 224, + 482, + 246 + ], + "type": "text", + "content": "66. Sohl-Dickstein, J., Weiss, E., Maheswaranathan, N., Ganguli, S.: Deep unsupervised learning using nonequilibrium thermodynamics. In: ICML (2015)" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 130, + 246, + 482, + 267 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 246, + 482, + 267 + ], + "spans": [ + { + "bbox": [ + 130, + 246, + 482, + 267 + ], + "type": "text", + "content": "67. Song, J., Meng, C., Ermon, S.: Denoising diffusion implicit models. arXiv preprint arXiv:2010.02502 (2020)" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 130, + 267, + 482, + 289 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 267, + 482, + 289 + ], + "spans": [ + { + "bbox": [ + 130, + 267, + 482, + 289 + ], + "type": "text", + "content": "68. Song, Y., Ermon, S.: Generative modeling by estimating gradients of the data distribution. NeurIPS (2019)" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 130, + 289, + 482, + 310 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 289, + 482, + 310 + ], + "spans": [ + { + "bbox": [ + 130, + 289, + 482, + 310 + ], + "type": "text", + "content": "69. Sugimoto, R., He, M., Liao, J., Sander, P.V.: Water simulation and rendering from a still photograph. In: SIGGRAPH Asia (2022)" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 130, + 310, + 447, + 320 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 310, + 447, + 320 + ], + "spans": [ + { + "bbox": [ + 130, + 310, + 447, + 320 + ], + "type": "text", + "content": "70. Szummer, M., Picard, R.W.: Temporal texture modeling. In: ICIP (1996)" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 130, + 320, + 482, + 342 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 320, + 482, + 342 + ], + "spans": [ + { + "bbox": [ + 130, + 320, + 482, + 342 + ], + "type": "text", + "content": "71. Teed, Z., Deng, J.: Raft: Recurrent all-pairs field transforms for optical flow. In: ECCV. pp. 402-419 (2020)" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 130, + 342, + 482, + 363 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 342, + 482, + 363 + ], + "spans": [ + { + "bbox": [ + 130, + 342, + 482, + 363 + ], + "type": "text", + "content": "72. Todorov, E., Erez, T., Tassa, Y.: Mujoco: A physics engine for model-based control. In: IROS (2012)" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 130, + 363, + 482, + 396 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 363, + 482, + 396 + ], + "spans": [ + { + "bbox": [ + 130, + 363, + 482, + 396 + ], + "type": "text", + "content": "73. Unterthiner, T., Van Steenkiste, S., Kurach, K., Marinier, R., Michalski, M., Gelly, S.: Towards accurate generative models of video: A new metric & challenges. arXiv preprint arXiv:1812.01717 (2018)" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 130, + 396, + 482, + 429 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 396, + 482, + 429 + ], + "spans": [ + { + "bbox": [ + 130, + 396, + 482, + 429 + ], + "type": "text", + "content": "74. Villegas, R., Babaeizadeh, M., Kindermans, P.J., Moraldo, H., Zhang, H., Saffar, M.T., Castro, S., Kunze, J., Erhan, D.: Phenaki: Variable length video generation from open domain textual description. arXiv preprint arXiv:2210.02399 (2022)" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 130, + 429, + 482, + 460 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 429, + 482, + 460 + ], + "spans": [ + { + "bbox": [ + 130, + 429, + 482, + 460 + ], + "type": "text", + "content": "75. Wang, X., Yuan, H., Zhang, S., Chen, D., Wang, J., Zhang, Y., Shen, Y., Zhao, D., Zhou, J.: Videocomposer: Compositional video synthesis with motion controllability. arXiv preprint arXiv:2306.02018 (2023)" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 130, + 460, + 482, + 493 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 460, + 482, + 493 + ], + "spans": [ + { + "bbox": [ + 130, + 460, + 482, + 493 + ], + "type": "text", + "content": "76. Wang, Y., Chen, X., Ma, X., Zhou, S., Huang, Z., Wang, Y., Yang, C., He, Y., Yu, J., Yang, P., et al.: Lavie: High-quality video generation with cascaded latent diffusion models. arXiv preprint arXiv:2309.15103 (2023)" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 130, + 493, + 482, + 514 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 493, + 482, + 514 + ], + "spans": [ + { + "bbox": [ + 130, + 493, + 482, + 514 + ], + "type": "text", + "content": "77. Wei, L.Y., Levoy, M.: Fast texture synthesis using tree-structured vector quantization. In: PACMCGIT (2000)" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 130, + 514, + 482, + 547 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 514, + 482, + 547 + ], + "spans": [ + { + "bbox": [ + 130, + 514, + 482, + 547 + ], + "type": "text", + "content": "78. Wei, Y., Zhang, S., Qing, Z., Yuan, H., Liu, Z., Liu, Y., Zhang, Y., Zhou, J., Shan, H.: Dreamvideo: Composing your dream videos with customized subject and motion. arXiv preprint arXiv:2312.04433 (2023)" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 130, + 547, + 482, + 568 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 547, + 482, + 568 + ], + "spans": [ + { + "bbox": [ + 130, + 547, + 482, + 568 + ], + "type": "text", + "content": "79. Weng, C.Y., Curless, B., Kemelmacher-Shlizerman, I.: Photo wake-up: 3d character animation from a single photo. In: CVPR (2019)" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 130, + 568, + 482, + 590 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 568, + 482, + 590 + ], + "spans": [ + { + "bbox": [ + 130, + 568, + 482, + 590 + ], + "type": "text", + "content": "80. Wu, J., Lim, J.J., Zhang, H., Tenenbaum, J.B., Freeman, W.T.: Physics 101: Learning physical object properties from unlabeled videos. In: BMVC (2016)" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 130, + 590, + 482, + 611 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 590, + 482, + 611 + ], + "spans": [ + { + "bbox": [ + 130, + 590, + 482, + 611 + ], + "type": "text", + "content": "81. Wu, J., Lu, E., Kohli, P., Freeman, W.T., Tenenbaum, J.B.: Learning to see physics via visual de-animation. In: NeurIPS (2017)" + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 130, + 611, + 482, + 643 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 611, + 482, + 643 + ], + "spans": [ + { + "bbox": [ + 130, + 611, + 482, + 643 + ], + "type": "text", + "content": "82. Wu, J., Yildirim, I., Lim, J.J., Freeman, W.T., Tenenbaum, J.B.: Galileo: Perceiving physical object properties by integrating a physics engine with deep learning. In: NeurIPS (2015)" + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 130, + 643, + 482, + 666 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 643, + 482, + 666 + ], + "spans": [ + { + "bbox": [ + 130, + 643, + 482, + 666 + ], + "type": "text", + "content": "83. Wu, R., Chen, L., Yang, T., Guo, C., Li, C., Zhang, X.: Lamp: Learn a motion pattern for few-shot-based video generation. arXiv preprint arXiv:2310.10769 (2023)" + } + ] + } + ], + "index": 23 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 268, + 91, + 447, + 102 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 268, + 91, + 447, + 102 + ], + "spans": [ + { + "bbox": [ + 268, + 91, + 447, + 102 + ], + "type": "text", + "content": "PhysGen: Rigid Physics-Grounded Img2Vid" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 470, + 92, + 481, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 470, + 92, + 481, + 100 + ], + "spans": [ + { + "bbox": [ + 470, + 92, + 481, + 100 + ], + "type": "text", + "content": "19" + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 18 + }, + { + "para_blocks": [ + { + "bbox": [ + 132, + 116, + 480, + 468 + ], + "type": "list", + "angle": 0, + "index": 15, + "blocks": [ + { + "bbox": [ + 132, + 116, + 480, + 138 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 116, + 480, + 138 + ], + "spans": [ + { + "bbox": [ + 132, + 116, + 480, + 138 + ], + "type": "text", + "content": "84. Xia, H., Lin, Z.H., Ma, W.C., Wang, S.: Video2game: Real-time, interactive, realistic and browser-compatible environment from a single video. In: CVPR (2024)" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 132, + 140, + 480, + 160 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 140, + 480, + 160 + ], + "spans": [ + { + "bbox": [ + 132, + 140, + 480, + 160 + ], + "type": "text", + "content": "85. Xia, H., Lin, Z.H., Ma, W.C., Wang, S.: Video2game: Real-time, interactive, realistic and browser-compatible environment from a single video. In: CVPR (2024)" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 133, + 161, + 480, + 182 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 161, + 480, + 182 + ], + "spans": [ + { + "bbox": [ + 133, + 161, + 480, + 182 + ], + "type": "text", + "content": "86. Xie, T., Zong, Z., Qiu, Y., Li, X., Feng, Y., Yang, Y., Jiang, C.: Physgaussian: Physics-integrated 3d gaussians for generative dynamics. In: CVPR (2024)" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 133, + 183, + 480, + 215 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 183, + 480, + 215 + ], + "spans": [ + { + "bbox": [ + 133, + 183, + 480, + 215 + ], + "type": "text", + "content": "87. Xing, J., Xia, M., Zhang, Y., Chen, H., Yu, W., Liu, H., Wang, X., Wong, T.T., Shan, Y.: Dynamiccafter: Animating open-domain images with video diffusion priors. arXiv preprint arXiv:2310.12190 (2023)" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 133, + 216, + 480, + 237 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 216, + 480, + 237 + ], + "spans": [ + { + "bbox": [ + 133, + 216, + 480, + 237 + ], + "type": "text", + "content": "88. Xu, Z., Wu, J., Zeng, A., Tenenbaum, J.B., Song, S.: Densephysnet: Learning dense physical object representations via multi-step dynamic interactions. In: RSS (2019)" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 133, + 238, + 480, + 258 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 238, + 480, + 258 + ], + "spans": [ + { + "bbox": [ + 133, + 238, + 480, + 258 + ], + "type": "text", + "content": "89. Xue, T., Wu, J., Bouman, K.L., Freeman, W.T.: Visual dynamics: Stochastic future generation via layered cross convolutional networks. T-PAMI (2018)" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 133, + 259, + 480, + 291 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 259, + 480, + 291 + ], + "spans": [ + { + "bbox": [ + 133, + 259, + 480, + 291 + ], + "type": "text", + "content": "90. Yang, J., Zhang, H., Li, F., Zou, X., Li, C., Gao, J.: Set-of-mark prompting unleashes extraordinary visual grounding in gpt-4v. arXiv preprint arXiv:2310.11441 (2023)" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 132, + 293, + 480, + 324 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 293, + 480, + 324 + ], + "spans": [ + { + "bbox": [ + 132, + 293, + 480, + 324 + ], + "type": "text", + "content": "91. Yu, J., Cun, X., Qi, C., Zhang, Y., Wang, X., Shan, Y., Zhang, J.: Animatezero: Video diffusion models are zero-shot image animators. arXiv preprint arXiv:2312.03793 (2023)" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 132, + 325, + 480, + 357 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 325, + 480, + 357 + ], + "spans": [ + { + "bbox": [ + 132, + 325, + 480, + 357 + ], + "type": "text", + "content": "92. Yu, L., Lezama, J., Gundavarapu, N.B., Versari, L., Sohn, K., Minnen, D., Cheng, Y., Gupta, A., Gu, X., Hauptmann, A.G., et al.: Language model beats diffusion-tokenizer is key to visual generation. arXiv preprint arXiv:2310.05737 (2023)" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 132, + 358, + 480, + 380 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 358, + 480, + 380 + ], + "spans": [ + { + "bbox": [ + 132, + 358, + 480, + 380 + ], + "type": "text", + "content": "93. Yu, T., Feng, R., Feng, R., Liu, J., Jin, X., Zeng, W., Chen, Z.: Inpaint anything: Segment anything meets image inpainting. arXiv preprint arXiv:2304.06790 (2023)" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 133, + 380, + 480, + 412 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 380, + 480, + 412 + ], + "spans": [ + { + "bbox": [ + 133, + 380, + 480, + 412 + ], + "type": "text", + "content": "94. Zhai, A.J., Shen, Y., Chen, E., Wang, G., Wang, X., Wang, S., Wang, X., Guan, K., Wang, S.: Physical property understanding from language-embedded feature fields. In: CVPR (2024)" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 133, + 413, + 480, + 445 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 413, + 480, + 445 + ], + "spans": [ + { + "bbox": [ + 133, + 413, + 480, + 445 + ], + "type": "text", + "content": "95. Zhang, S., Wang, J., Zhang, Y., Zhao, K., Yuan, H., Qing, Z., Wang, X., Zhao, D., Zhou, J.: I2vgen-xl: High-quality image-to-video synthesis via cascaded diffusion models. arXiv preprint arXiv:2311.04145 (2023)" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 132, + 447, + 480, + 468 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 447, + 480, + 468 + ], + "spans": [ + { + "bbox": [ + 132, + 447, + 480, + 468 + ], + "type": "text", + "content": "96. Zhao, J., Zhang, H.: Thin-plate spline motion model for image animation. In: CVPR (2022)" + } + ] + } + ], + "index": 14 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "spans": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "type": "text", + "content": "20" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 166, + 91, + 215, + 100 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 166, + 91, + 215, + 100 + ], + "spans": [ + { + "bbox": [ + 166, + 91, + 215, + 100 + ], + "type": "text", + "content": "S. Liu et al." + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 19 + } + ], + "_backend": "vlm", + "_version_name": "2.6.4" +} \ No newline at end of file diff --git a/2024/Physical-Based Event Camera Simulator/0986a08f-4884-431b-9c20-c64d0c63357b_content_list.json b/2024/Physical-Based Event Camera Simulator/0986a08f-4884-431b-9c20-c64d0c63357b_content_list.json new file mode 100644 index 0000000000000000000000000000000000000000..afdf8f9172e07309c3127fd91ba139d68eb6f0b0 --- /dev/null +++ b/2024/Physical-Based Event Camera Simulator/0986a08f-4884-431b-9c20-c64d0c63357b_content_list.json @@ -0,0 +1,1792 @@ +[ + { + "type": "text", + "text": "Physical-Based Event Camera Simulator", + "text_level": 1, + "bbox": [ + 269, + 141, + 733, + 162 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Haiqian Han $^{1}$ , Jiacheng Lyu $^{1}$ , Jianing Li $^{1}$ $(\\boxtimes)$ , Henglu Wei $^{1}$ , Cheng Li $^{2}$ , Yajing Wei $^{2}$ , Shu Chen $^{2}$ , and Xiangyang Ji $^{1}$ $(\\boxtimes)$", + "bbox": [ + 222, + 188, + 781, + 220 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "$^{1}$ Tsinghua University, Haidian District, Beijing, 100084, P. R. China", + "bbox": [ + 267, + 231, + 732, + 247 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "2 Beijing Xiaomi Mobile Software Co., Ltd.", + "bbox": [ + 354, + 247, + 645, + 260 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Abstract. Existing event camera simulators primarily focus on the process of generating video events and often overlook the entire optical path in real-world camera systems. To address this limitation, we propose a novel Physical-based Event Camera Simulator (PECS), which is able to generate a high-fidelity realistic event stream by directly interfacing with the 3D scene. Our PECS features a lens simulation block for accurate light-to-sensor chip replication and a multispectral rendering module for precise photocurrent generation. We present two spatiotemporal event metrics to assess the similarity between simulated and actual camera events. Experimental results demonstrate that our PECS outperforms four state-of-the-art simulators by a large margin in terms of event-based signal fidelity. We integrate our PECS into the UE platform to generate extensive multi-task synthetic datasets and evaluate its effectiveness in downstream vision tasks (e.g., video reconstruction). Our open-source code is available at https://github.com/lanpokn/PECS_trail_version.", + "bbox": [ + 261, + 303, + 743, + 525 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Keywords: Event camera simulation $\\cdot$ Physics-based vision $\\cdot$ Hyper-spectral data analysis", + "bbox": [ + 259, + 539, + 740, + 568 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "1 Introduction", + "text_level": 1, + "bbox": [ + 215, + 595, + 375, + 611 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Event cameras [7,43], namely bio-inspired vision sensors, operate fundamentally differently from conventional cameras. Instead of capturing intensity images at a fixed rate, event cameras respond to brightness changes with a stream of asynchronous events. With the advantages of high temporal resolution, high dynamic range, and low power consumption [33], event cameras have found widespread use in various computer vision tasks [4, 13, 17, 18, 21, 22, 25, 29, 39, 40, 49].", + "bbox": [ + 212, + 631, + 787, + 720 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Despite notable advancements in event-based vision, the training of deep learning-based approaches still demands a substantial amount of synthetic event data [19, 53]. It is worth noting that the high cost [43] and deployment challenges [6] associated with event cameras in high-speed or low-light scenes limit the availability of real-world datasets. Thus, several event camera simulators have attempted to generate a large amount of affordable and reliable event data.", + "bbox": [ + 212, + 722, + 787, + 813 + ], + "page_idx": 0 + }, + { + "type": "page_footnote", + "text": "$(\\boxtimes)$ Corresponding author: Xiangyang Ji and Jianing Li.", + "bbox": [ + 230, + 824, + 614, + 839 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/104c9b0d509aa8877ac1f3aaa1c22ea7e83c5a158b9b6bf96fabcbfeb5c63ad9.jpg", + "image_caption": [ + "Fig. 1: Event camera simulators. (a) The existing video to events simulators (e.g., V2E [12]). (b) Our Physical Event Camera Simulator (PECS) is designed to generate high-fidelity realistic events by directly interfacing with the 3D scene." + ], + "image_footnote": [], + "bbox": [ + 217, + 143, + 785, + 224 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "One category refers to optimized-based event camera simulators [5, 12, 14, 15, 20, 31, 38, 41, 57], where the objective is to design hand-crafted modules for generating event data. For example, V2E [12] converts intensity frames into event data via multiple hand-crafted modules. Note that, these simulators take three-channel videos as direct inputs, overlooking the complete optical path in real-world camera systems. It may lead to the loss of optical information, consequently seriously affecting the simulator's accuracy. While some efforts [14, 44] support offline rendering in the 3D scene, they often neglect the consideration of spectrum data in the rendering module. In fact, integrating spectrum information into the event camera simulation process has the potential to enhance the fidelity of synthetic event streams.", + "bbox": [ + 212, + 310, + 787, + 474 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Another category is learning-based event camera simulators [2, 8, 10, 28, 34, 52, 54] that aim to generate event representations, improving the generalization capabilities of deep learning models directly for the target domain rather than relying on raw events. For instance, EventGAN [54] is an end-to-end neural network that directly converts images to event presentations for downstream computer vision tasks. However, these learning-based simulators, requiring retraining for diverse usage cases, may have limited generalization across different scenarios. Moreover, the generated event representations are not in the original signal domain [1], posing potential limitations for broader applications. In other words, physical-based simulators, producing spatiotemporal events, preserve the inherent characteristics of raw camera data, enhancing their versatility compared to event representations by learning-based simulators.", + "bbox": [ + 212, + 477, + 787, + 657 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "To address the aforementioned problems, this paper proposes a novel Physical-based Event Camera Simulator (PECS), which is able to generate a highly realistic event stream by directly interfacing with the 3D scene (see Fig. 1). In fact, the goal of this work is not to optimize hand-crafted event camera simulators (e.g., ESIM or V2E). In contrast, we aim at overcoming the following challenges: (i) How do we model a realistic lens to replicate the light-to-sensor chip process? (ii) How do we design a multispectral renderer to obtain high-precision photocurrents rather than three-channel videos? Towards this end, a realistic lens simulation block using the PBRT renderer [30] is developed to accurately replicate the light-to-sensor chip process. Then, a novel multispectral rendering module is designed to generate high-precision photocurrents by quantum efficiency and Monte Carlo integration. Two asynchronous spatiotemporal event metrics (i.e., Chamfer dis", + "bbox": [ + 212, + 659, + 789, + 840 + ], + "page_idx": 1 + }, + { + "type": "page_number", + "text": "2", + "bbox": [ + 217, + 114, + 228, + 126 + ], + "page_idx": 1 + }, + { + "type": "header", + "text": "Haiqian. Author et al.", + "bbox": [ + 271, + 114, + 419, + 128 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "tance and Gaussian distance) are proposed to measure the similarity between simulated and raw events. Extensive experiments show that our PECS outperforms four state-of-the-art simulators by a large margin in terms of event-based signal fidelity. Moreover, we further deploy real-time versions into the Unreal Engine (UE) platform to generate extensive multi-task synthetic datasets. Typical computer vision tasks (i.e., event-based video reconstruction) are conducted to verify the effectiveness of our PECS.", + "bbox": [ + 212, + 146, + 782, + 251 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Overall, the main contributions of this work are summarized as follows:", + "bbox": [ + 238, + 252, + 751, + 266 + ], + "page_idx": 2 + }, + { + "type": "list", + "sub_type": "text", + "list_items": [ + "- We propose a novel Physical-based Event Camera Simulator (PECS), capable of generating a high-fidelity realistic event stream by directly interfacing with the 3D scene, without relying on video-to-event conversion.", + "- We design a realistic lens simulation block using the PBRT renderer, optimizing pupil sampling to improve the quality and speed of rendering.", + "- We design a novel multispectral rendering module that generates high-precision photocurrents through quantum efficiency and Monte Carlo integration.", + "- We present two asynchronous spatiotemporal metrics to assess the similarity between simulated and raw events. Extensive experiments show that our PECS outperforms four state-of-the-art simulators in terms of signal fidelity." + ], + "bbox": [ + 225, + 277, + 792, + 426 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "To the best of our knowledge, this is the first attempt to explore a physical-based event camera simulator. We believe that our simulator will provide high-fidelity realistic large-scale event data for event-based vision tasks and offer proof of principle for the next generation of neuromorphic cameras.", + "bbox": [ + 212, + 439, + 784, + 500 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "2 Related Work", + "text_level": 1, + "bbox": [ + 214, + 522, + 387, + 537 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "This section will review existing event camera simulators from two perspectives including optimized-based simulators and learning-based simulators.", + "bbox": [ + 212, + 553, + 782, + 582 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Optimized-based Event Camera Simulators. The prior optimized-based simulators include PIX2NVS [5] and ESIM [8, 31], which convert any image or video into asynchronous events. A realistic V2E [12] is developed to model the event camera in low-light or motion blur scenarios. ICNS [14] is presented to simulate noise and estimate the latency by adding the effects of the arbiter. VOLT [20] employs the concept of stochastic processes to model the output process of event data, thereby endowing the simulated event data with good continuity and probabilistic interpretability. Nevertheless, these event camera simulators directly consider three-channel videos as the input and overlook the complete optical path in real-world camera systems. Although Mou et al. [24] first notice the impact of quantum efficiency on event camera simulators and implement video-to-spectrum conversion. However, the lossless conversion from three-channel data to spectral data is nearly impossible. Thus, we design a novel multispectral rendering module that generates high-precision photocurrents through quantum efficiency and Monte Carlo integration.", + "bbox": [ + 212, + 583, + 784, + 808 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Learning-based Event Camera Simulators. With the rapid development of deep learning, some learning-based event camera simulators [2,8,10,28,34,52,54]", + "bbox": [ + 212, + 809, + 785, + 839 + ], + "page_idx": 2 + }, + { + "type": "header", + "text": "Physical-Based Event Camera Simulator", + "bbox": [ + 460, + 114, + 730, + 128 + ], + "page_idx": 2 + }, + { + "type": "page_number", + "text": "3", + "bbox": [ + 774, + 116, + 784, + 126 + ], + "page_idx": 2 + }, + { + "type": "image", + "img_path": "images/2433c22525bbf331f29d2da5f5f86e09c4c54c5163cb502084ab8b57b428d321.jpg", + "image_caption": [ + "Fig. 2: The pipeline of the proposed Physical-based Event Camera Simulator (PECS). It starts with pupil sampling, precalculating the rays expected to pass through the lens. Ray tracing calculates light refraction and reflection within the lens and scene, producing multispectral maps. Incorporating quantum effects, these maps generate accurate photocurrents. Our PECS then leverages logarithmic differences to generate events, enabling the measurement of similarity between simulated and raw camera events in signal fidelity." + ], + "image_footnote": [], + "bbox": [ + 222, + 150, + 772, + 358 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "are developed to generate event representations in the target domain. For instance, Zhu et al. [54] use the modified GAN model to generate synchronous event representations without raw events. Gu et al. [10] propose a domain-adaptive event simulator that can generate both event data and the camera's trajectory. Pantho et al. [28] use a neural network to generate event data represented by regions, which is convenient for use in moving region detection. While these simulators can enhance overall accuracy to a certain degree, the intrinsic challenge lies in the limited interpretability of neural networks. Consequently, these simulators, requiring re-training for diverse usage cases, may have limited generalization across different scenarios. In other words, simulators that rely on learning-based approaches are expected to exhibit lower robustness compared with optimized-based event camera simulators.", + "bbox": [ + 212, + 501, + 787, + 681 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "3 Methods", + "text_level": 1, + "bbox": [ + 215, + 705, + 338, + 722 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "3.1 Framework Overview", + "text_level": 1, + "bbox": [ + 215, + 738, + 437, + 753 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "This work aims at designing a novel Physical-based Event Camera Simulator, termed PECS, which is able to generate a highly realistic event stream by directly interfacing with the 3D scene. As illustrated in Fig. 2, our PECS mainly consists of two modules: a multispectral rendering module and a dynamic event generation module. In the multispectral rendering module, we employ pupil sampling", + "bbox": [ + 212, + 763, + 787, + 840 + ], + "page_idx": 3 + }, + { + "type": "page_number", + "text": "4", + "bbox": [ + 217, + 114, + 228, + 126 + ], + "page_idx": 3 + }, + { + "type": "header", + "text": "Haiqian. Author et al.", + "bbox": [ + 271, + 114, + 419, + 128 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "as a pre-processing step to obtain a set of rays originating from the direction in which the lens may emit. Then, ray tracing is employed to compute the refraction and reflection of light within both the lens and the scene, allowing us to acquire multispectral maps. Meanwhile, multispectral maps with quantum effects generate high-precision photocurrent. Particularly, we design a realistic lens simulation block using the PBRT renderer, which aims at optimizing pupil sampling to improve the quality and speed of rendering. In the dynamic event generation module, we leverage the principle of logarithmic differences in the event camera to generate events. This module tackles the constraints of current simulation systems by incorporating spectrum data and accounting for quantum efficiency. This enhancement contributes to improved accuracy in estimating photocurrent, thereby creating a more realistic depiction of the physical processes involved in the conversion of light signals into photocurrent. Finally, two event metrics (i.e., Chamfer distance and Gaussian distance) are proposed to measure the distance between simulated and raw events in signal fidelity.", + "bbox": [ + 212, + 146, + 787, + 375 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "3.2 Realistic Camera Lenses Simulation", + "text_level": 1, + "bbox": [ + 215, + 396, + 555, + 409 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Realistic camera lens simulation in computer graphics replicates actual camera lens behavior [16] by employing ray tracing to accurately simulate light paths through multiple lens elements. This approach takes into account the intricacies of the lens system, including refraction at interfaces, and improves image quality by estimating incident radiance along arbitrary rays.", + "bbox": [ + 212, + 421, + 784, + 496 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Our realistic camera lens simulation block models a lens system with various elements, tracing rays through them with multispectral information. It considers interactions until rays either exit the optical system or are absorbed by the aperture stop or lens housing. To reduce computational waste, our PECS performs pupil sampling before conducting the actual simulation. In this study, the pupil's definition is a little different from in optics [36]. It refers to all the directions from which rays emit from the lenses at any specific point on the sensor film, not just the center point. Let the sensor plane coordinates be denoted as $(u,v)$ . The pupil is a set of feasible directions represented by points $(x,y)$ on the direction plane. Each point $(x,y)$ in the pupil corresponds to a unique direction. Thus, the mathematical definition of the pupil can be expressed as:", + "bbox": [ + 212, + 497, + 787, + 662 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\nP u p i l (u, v) = \\{(x, y) \\mid R a y (\\{u, v \\}, \\{x, y \\})), \\tag {1}\n$$\n", + "text_format": "latex", + "bbox": [ + 349, + 676, + 784, + 691 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "where $Ray(\\{u, v\\}, \\{x, y\\})$ denotes the ray originating from $(u, v)$ and pointing towards $(x, y)$ passing through the lenses.", + "bbox": [ + 212, + 703, + 784, + 733 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Pupil sampling is the algorithm that identifies the correct pupil for each point on the sensor film. In the case of lenses commonly used in event cameras (e.g., Prophesee Gen4), which are often centrally symmetric with circular apertures, representing the pupil shape as an ellipse simplifies the pupil sampling process. Therefore, the pupil can be approximated as follows:", + "bbox": [ + 212, + 734, + 784, + 809 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\nP u p i l (u, v) = \\{(x, y) \\mid (x, y) \\in E l l i (p, a, b) \\}, \\tag {2}\n$$\n", + "text_format": "latex", + "bbox": [ + 346, + 825, + 784, + 840 + ], + "page_idx": 4 + }, + { + "type": "header", + "text": "Physical-Based Event Camera Simulator", + "bbox": [ + 460, + 114, + 730, + 128 + ], + "page_idx": 4 + }, + { + "type": "page_number", + "text": "5", + "bbox": [ + 774, + 116, + 784, + 126 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "where $Elli(p, a, b)$ is an ellipse centered at point $p$ on the direction plane with semi-major axis length $a$ and semi-minor axis length $b$ .", + "bbox": [ + 212, + 146, + 782, + 176 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "By uniformly sampling points on the beam direction plane, the exit pupil for each sampled point on the sensor film is efficiently computed. For other points, the exit pupil is determined through linear interpolation between the ellipse's center and its major and minor axes. Despite potential errors in the elliptical model for complex lenses, this method still notably decreases the time of ray tracing within the lens, thereby speeding up the lens system simulation.", + "bbox": [ + 212, + 176, + 784, + 267 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "3.3 Multispectral Rendering Module", + "text_level": 1, + "bbox": [ + 214, + 287, + 532, + 303 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "The multispectral rendering module transforms incoming light signals into photocurrent, adjusting for display convenience [12, 14, 31]. An event camera involves detecting photons at various frequencies, initiating photocurrent through the photoelectric effect. After filtering and logarithmic conversion, the currents are converted into photovoltages. The module concentrates on the photon-to-photocurrent conversion, excluding subsequent electronic circuit model steps.", + "bbox": [ + 212, + 310, + 782, + 400 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Most simulation systems estimate photocurrent using the L component of LUV values, but actual photon receivers don't convert different frequency lights into the same number of electrons, as revealed by the Quantum Efficiency (QE) curve. The definition of absolute quantum efficiency [35] can be described as:", + "bbox": [ + 212, + 401, + 784, + 460 + ], + "page_idx": 5 + }, + { + "type": "equation", + "text": "\n$$\n\\mathrm {Q E} _ {\\text {a b s o l u t e}} (\\lambda) = \\frac {N _ {e}}{N _ {v}} = \\frac {R _ {\\lambda}}{\\lambda} \\times \\frac {h c}{e}, \\tag {3}\n$$\n", + "text_format": "latex", + "bbox": [ + 385, + 470, + 784, + 501 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "where $N_{e}$ is the number of electrons, $N_{v}$ is the number of photons, $R_{\\lambda}$ is the current per unit of incident light power (A/W), $\\lambda$ is the wavelength in nm, $h$ is the Planck constant, $c$ is the speed of light in vacuum, and $e$ is the elementary charge. The quantum efficiency is generally given as a relative value ranging from 0 to 1. Its definition can be formulated as follows:", + "bbox": [ + 212, + 507, + 784, + 580 + ], + "page_idx": 5 + }, + { + "type": "equation", + "text": "\n$$\n\\mathrm {Q E} _ {\\text {r e l a t i v e}} (\\lambda) = \\frac {Q E _ {\\text {a b s o l u t e}} (\\lambda)}{\\operatorname* {m a x} (Q E _ {\\text {a b s o l u t e}} (\\lambda))}. \\tag {4}\n$$\n", + "text_format": "latex", + "bbox": [ + 370, + 588, + 784, + 621 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Our PECS overcomes limitations by providing multi-spectral data for precise photocurrent estimation. Leveraging known spectrum information and quantum efficiency, PECS directly computes photocurrent as follows:", + "bbox": [ + 212, + 627, + 782, + 672 + ], + "page_idx": 5 + }, + { + "type": "equation", + "text": "\n$$\ni _ {p} = C \\cdot \\int \\lambda \\cdot Q E (\\lambda) \\cdot L (\\lambda) \\cdot d \\lambda , \\tag {5}\n$$\n", + "text_format": "latex", + "bbox": [ + 388, + 681, + 784, + 712 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "where $i_p$ is the photocurrent after Photoelectric conversion, which is a key parameter in subsequent electronic circuits. C is a constant that depends on the pixel area and units of physical quantities, and $L(\\lambda)$ is the input spectrum. If L is represented by a finite set of discrete channels, the following simple numerical integration method can be used to achieve the aforementioned integration:", + "bbox": [ + 212, + 718, + 782, + 792 + ], + "page_idx": 5 + }, + { + "type": "equation", + "text": "\n$$\ni _ {p} \\approx C \\cdot \\sum_ {k = 1} ^ {N} \\lambda_ {k} \\cdot Q E (\\lambda_ {k}) \\cdot L (\\lambda_ {k}) \\cdot \\left(\\lambda_ {k + 1} - \\lambda_ {k}\\right), \\tag {6}\n$$\n", + "text_format": "latex", + "bbox": [ + 339, + 803, + 784, + 842 + ], + "page_idx": 5 + }, + { + "type": "page_number", + "text": "6", + "bbox": [ + 217, + 114, + 228, + 126 + ], + "page_idx": 5 + }, + { + "type": "header", + "text": "Haiqian. Author et al.", + "bbox": [ + 271, + 114, + 419, + 128 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "where $N$ is the number of known channels of the spectrum. If $N = 3$ , the information carried by the spectrum degrades into a common RGB image.", + "bbox": [ + 212, + 146, + 782, + 176 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "The spectrum representation can be more complex in certain applications, such as using neural networks for implicit continuous representations. Simple data integration may lead to insufficient sampling in regions with larger function values, causing significant errors. To address this, Monte Carlo integration and importance sampling are commonly employed as:", + "bbox": [ + 212, + 176, + 784, + 252 + ], + "page_idx": 6 + }, + { + "type": "equation", + "text": "\n$$\ni _ {p} = \\int_ {a} ^ {b} f (x) d x = \\lim _ {N \\rightarrow \\infty} \\frac {1}{N} \\sum_ {i = 1} ^ {N} \\frac {f \\left(x _ {i}\\right)}{p \\left(x _ {i}\\right)}, \\tag {7}\n$$\n", + "text_format": "latex", + "bbox": [ + 367, + 265, + 785, + 305 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "where $N$ is the number of samples. $p(x)$ is probability density function, which generated samples $x_{i}$ . Applying Eq. 7 to Eq. 5, it can be re-written as:", + "bbox": [ + 214, + 316, + 782, + 347 + ], + "page_idx": 6 + }, + { + "type": "equation", + "text": "\n$$\ni _ {p} \\approx I _ {N} = C \\cdot \\frac {1}{N} \\sum_ {k = 1} ^ {N} \\frac {\\lambda_ {k} \\cdot Q E (\\lambda_ {k}) \\cdot L (\\lambda_ {k})}{p (\\lambda_ {k})}, \\tag {8}\n$$\n", + "text_format": "latex", + "bbox": [ + 352, + 359, + 785, + 400 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "where $I_N$ is an approximation. Choosing a well-behaved $p(\\lambda)$ is crucial to minimize $N$ while maintaining accuracy. In probability theory, it is equal to make the variance small. Assuming that different samples are independent and identically distributed, we can obtain the variance of $I_N$ as:", + "bbox": [ + 214, + 411, + 782, + 472 + ], + "page_idx": 6 + }, + { + "type": "equation", + "text": "\n$$\n\\operatorname {V a r} \\left(I _ {N}\\right) = \\frac {1}{N ^ {2}} \\sum_ {k = 1} ^ {N} \\operatorname {V a r} \\left(\\frac {\\lambda_ {k} \\cdot Q E \\left(\\lambda_ {k}\\right) \\cdot L \\left(\\lambda_ {k}\\right)}{p \\left(\\lambda_ {k}\\right)}\\right) = \\frac {\\sigma_ {N} ^ {2}}{N ^ {2}}, \\tag {9}\n$$\n", + "text_format": "latex", + "bbox": [ + 308, + 484, + 785, + 525 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "In the optimal condition, $\\sigma_N^2 = 0$ , which means $p(\\lambda)$ is totally equal to the original function except for scale. Thus, we can model $p(\\lambda)$ as a piecewise constant function and estimate it by pre-calculating the original function as:", + "bbox": [ + 214, + 537, + 782, + 583 + ], + "page_idx": 6 + }, + { + "type": "equation", + "text": "\n$$\np (\\lambda) = B \\cdot \\lambda_ {j} ^ {\\prime} \\cdot Q E \\left(\\lambda_ {j} ^ {\\prime}\\right) \\cdot L \\left(\\lambda_ {j} ^ {\\prime}\\right), j = \\operatorname {a r g m i n} _ {m} | \\lambda - \\lambda_ {m} ^ {\\prime} |, \\tag {10}\n$$\n", + "text_format": "latex", + "bbox": [ + 308, + 595, + 785, + 614 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "where $m$ is the channel number. For each $m$ , $L(\\lambda_m')$ is pre-calculated through the implicit representation of the spectrum. B is a constant that ensures the integral of $p(\\lambda)$ equals 1. Then, the photocurrent can be obtained as follows:", + "bbox": [ + 214, + 623, + 782, + 669 + ], + "page_idx": 6 + }, + { + "type": "equation", + "text": "\n$$\ni _ {p} \\approx C \\cdot \\frac {1}{N} \\sum_ {k = 1} ^ {N} \\frac {\\lambda_ {k} \\cdot Q E (\\lambda_ {k}) \\cdot L (\\lambda_ {k})}{B \\cdot \\lambda_ {j _ {k}} ^ {\\prime} \\cdot Q E \\left(\\lambda_ {j _ {k}} ^ {\\prime}\\right) \\cdot L \\left(\\lambda_ {j _ {k}} ^ {\\prime}\\right)}, j _ {k} = \\operatorname {a r g m i n} _ {m} | \\lambda_ {k} - \\lambda_ {m} ^ {\\prime} |. \\tag {11}\n$$\n", + "text_format": "latex", + "bbox": [ + 251, + 683, + 785, + 722 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Note that, Monte Carlo integration in Eq. 11 and simple numerical integration in Eq. 6 each have their own advantages and disadvantages. When the spectral data is represented continuously, Monte Carlo integration produces significantly smaller errors with a sharp spectral distribution curve compared to simple integration. However, if the spectrum is only represented by discrete channels, the lack of original information renders Monte Carlo integration somewhat redundant, and simple numerical integration can be used instead.", + "bbox": [ + 214, + 733, + 782, + 839 + ], + "page_idx": 6 + }, + { + "type": "header", + "text": "Physical-Based Event Camera Simulator", + "bbox": [ + 459, + 114, + 732, + 128 + ], + "page_idx": 6 + }, + { + "type": "page_number", + "text": "7", + "bbox": [ + 774, + 114, + 785, + 126 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Note that, our PECS estimates photocurrent by considering spectral data and quantum efficiency, offering more accurate simulations of light signal conversion. Consequently, the accuracy is significantly enhanced compared to the LUV method. Meanwhile, the quantum efficiency of different cameras often varies and can generally be provided by the corresponding sensor manufacturers.", + "bbox": [ + 212, + 146, + 787, + 223 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "3.4 Asynchronous Spatiotemporal Event Metric", + "text_level": 1, + "bbox": [ + 215, + 246, + 625, + 262 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "To measure the similarity between simulated and raw camera events, we present two asynchronous spatiotemporal event metrics (i.e., Chamfer distance and Gaussian distance), which is similar to the metric part in iterative closest point tasks [3,37]. The Chamfer distance metric $C_D$ can be described as:", + "bbox": [ + 212, + 273, + 787, + 335 + ], + "page_idx": 7 + }, + { + "type": "equation", + "text": "\n$$\nC _ {D} (R, Q) = \\frac {1}{| R |} \\sum_ {r \\in R} \\min _ {q \\in Q} d (r, q) + \\frac {1}{| Q |} \\sum_ {q \\in Q} \\min _ {r \\in R} d (r, q), \\tag {12}\n$$\n", + "text_format": "latex", + "bbox": [ + 307, + 345, + 785, + 383 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "where the lower $C_D$ its value, the closer the two events are. $R$ and $Q$ are two point clouds, $r$ or $q$ is a single point in $R$ or $Q$ . Each point can be expressed as a four-dimensional tuple $\\{x,y,p,t\\}$ representing the spatial coordinates, polarity, and temporal information. The distance in the metric can be depicted as:", + "bbox": [ + 214, + 393, + 787, + 455 + ], + "page_idx": 7 + }, + { + "type": "equation", + "text": "\n$$\nd (r, q) = | | r - q | | _ {2}, \\tag {13}\n$$\n", + "text_format": "latex", + "bbox": [ + 434, + 467, + 785, + 486 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "By utilizing KDTREE [9] for acceleration, the time complexity is only $O(n \\log(n))$ . Similarly, the Gaussian distance can be formulated as:", + "bbox": [ + 215, + 494, + 792, + 526 + ], + "page_idx": 7 + }, + { + "type": "equation", + "text": "\n$$\nG _ {D} (R, Q) = \\frac {1}{| R |} \\sum_ {r \\in R} g \\left(\\min _ {q \\in Q} d (r, q)\\right) + \\frac {1}{| Q |} \\sum_ {q \\in Q} g \\left(\\min _ {r \\in R} d (r, q)\\right), \\tag {14}\n$$\n", + "text_format": "latex", + "bbox": [ + 285, + 537, + 785, + 574 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "where the function $g$ makes $G_{D}(R, Q)$ still become lower when the two events get closer while reducing the impact of outliers, and it can be described as:", + "bbox": [ + 214, + 587, + 787, + 619 + ], + "page_idx": 7 + }, + { + "type": "equation", + "text": "\n$$\ng (x) = 1 - \\exp \\left(- \\frac {| | x | | _ {2} ^ {2}}{\\sigma}\\right), \\tag {15}\n$$\n", + "text_format": "latex", + "bbox": [ + 403, + 628, + 785, + 662 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "where $\\sigma$ is a hyperparameter used to adjust the sensitivity range of the original metric. In this study, we set 0.4 as the default.", + "bbox": [ + 214, + 672, + 787, + 702 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "In fact, these two metrics describe the attributes of asynchronous events from different dimensions and can be necessary to assess the similarity between raw and simulated data. The Gaussian distance exhibits relative stability against outliers compared to the Chamfer distance, making it more robust to noise. Nevertheless, its measurement results have a narrower range of variation, making it less distinctive in simultaneous evaluations of multiple simulators.", + "bbox": [ + 212, + 704, + 787, + 794 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Note that, the units of x, y, p, and t in the original data are different, and the range of values for t often far exceeds that of the other components, resulting in Eq. 13 and Eq. 14 almost do not consider information beyond t. To address this", + "bbox": [ + 214, + 795, + 787, + 840 + ], + "page_idx": 7 + }, + { + "type": "page_number", + "text": "8", + "bbox": [ + 217, + 114, + 228, + 126 + ], + "page_idx": 7 + }, + { + "type": "header", + "text": "Haiqian. Author et al.", + "bbox": [ + 271, + 114, + 419, + 128 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "issue, it is necessary to normalize the original measurements first to resolve this scaling problem as follows:", + "bbox": [ + 212, + 146, + 782, + 176 + ], + "page_idx": 8 + }, + { + "type": "equation", + "text": "\n$$\nL _ {\\beta} [ i ] = \\left(\\frac {L _ {\\beta} ^ {\\prime} [ i ] - \\operatorname* {m i n} L _ {\\beta} ^ {\\prime}}{\\operatorname* {m a x} L _ {\\beta} ^ {\\prime} - \\operatorname* {m i n} L _ {\\beta} ^ {\\prime} + \\epsilon}\\right) * \\alpha , \\tag {16}\n$$\n", + "text_format": "latex", + "bbox": [ + 375, + 183, + 785, + 219 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "where $L_{\\beta}^{\\prime}$ is the original measurements, $\\beta$ can be $x, y, p$ and $t$ element. $L_{\\beta}$ is the corresponding processed data. $i$ is the i'th event of the raw data, which means $R[i] = \\{L_x[i], L_y[i], L_p[i], L_t[i]\\}$ . $\\epsilon$ is a small number that is to avoid dividing zero, and $\\alpha$ is a hyperparameter. In this work, we choose $\\alpha = 100$ for space $x, y$ and polarity $p$ , and choose $\\alpha = 1000$ for time $t$ .", + "bbox": [ + 214, + 226, + 787, + 303 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "4 Experiments", + "text_level": 1, + "bbox": [ + 215, + 321, + 375, + 339 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "4.1 Experimental Setting", + "text_level": 1, + "bbox": [ + 215, + 349, + 439, + 366 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "Real-to-Sim Scenario Construction. Spatiotemporal synchronization is essential for quantitatively evaluating, we record the real-world event data using a Prophesee Gen4 camera in a professional optical laboratory. We capture six sequences including rotating disk and translating checkerboards with light changes and various motion speeds. For example, T_0.6_H means translating checkerboards, 0.6 meter-per-seconds, and high light. R_360_L means rotating disk, 360 rpm, and low light. To ensure consistency between the simulated and real-world scenes, we utilize the blender for high-fidelity 3D modeling. We export the scenes to PBRT and simulate a naturally calibrated camera using real lenses [26,42]. This approach enables us to acquire temporally and spatially aligned rendering results for use in subsequent sensor modules. More details of spatiotemporal alignment can be found in the supplementary material. In other methods, as regular RGB cameras struggle to capture clear images in these scenes, we use Blender's rendered RGB video as input. By manually calibrating pinhole camera parameters, we ensure temporally and spatially consistent clear images for simulation inputs, enabling the proper functioning of these simulators. Implementation Details. All experiments are conducted on a GeForce RTX 3050 Ti and 11th Gen Intel(R) Core(TM) i7-11800H CPU. We compare our PECS with four open-source event camera simulators (i.e., ESIM [31], VOLT [20], V2E [12], and ICNS [14]). In our PECS, the hyperparameters in Eq. 11 and Eq. 6 are set to $C = 100$ and $N = 31$ to make an accuracy-speed trade-off. Regarding the ESIM method, we extract its core code from its original ties to Unreal Engine (UE) and replace UE output with video output. For the V2E method, given the clarity and compactness of Blender-generated images, we disable the network interpolation module to prevent extremely slow simulation speed. In the ICNS method, all parameters have been adjusted to achieve optimal results. In the VOLT method, we select a set of parameters for the best performance. We present two event metrics (i.e., Chamfer distance and Gaussian distance) to quantitatively evaluate the similarity between raw events and synthetic data. Besides, we select three representative sequences of the HS-ERGB dataset [45] in the event-based video reconstruction task.", + "bbox": [ + 212, + 371, + 787, + 840 + ], + "page_idx": 8 + }, + { + "type": "header", + "text": "Physical-Based Event Camera Simulator", + "bbox": [ + 460, + 114, + 730, + 128 + ], + "page_idx": 8 + }, + { + "type": "page_number", + "text": "9", + "bbox": [ + 774, + 116, + 784, + 126 + ], + "page_idx": 8 + }, + { + "type": "table", + "img_path": "images/2791daa0e5d1507350134b3563d54e40cd66c4c9df309de2c8ce1e5012159ea6.jpg", + "table_caption": [ + "Table 1: Comparison with state-of-the-art simulators on our six recording sequences. Note that, our Physical Event Camera Simulator (PECS) shows superior performance on both metrics across a diverse range of movement speeds and lighting conditions." + ], + "table_footnote": [], + "table_body": "
SceneGaussian distance ↓Chamfer distance ↓
ESIMVOLTV2EICNSPECSESIMVOLTV2EICNSPECS
Rotate 360 High1.9951.9231.6711.5580.96045.51222.0866.3563.8162.010
Rotate 360 Low1.9651.9311.8311.6811.63135.08225.3917.7294.1313.708
Rotate 60 High1.9951.9111.8161.5741.50439.42520.8149.2974.2373.459
Trans 1mps High1.9631.6551.8271.8061.66762.20610.57315.4509.4533.975
Trans 06mps High1.9671.7231.8391.8371.79863.93910.19818.1549.5884.169
Trans 1mps Low1.9871.8141.7001.6631.71068.11510.95214.0719.1824.467
Average1.9791.8261.7811.6871.54552.38016.66911.8436.7353.631
", + "bbox": [ + 220, + 198, + 777, + 306 + ], + "page_idx": 9 + }, + { + "type": "text", + "text": "4.2 Effective Test", + "text_level": 1, + "bbox": [ + 215, + 333, + 375, + 348 + ], + "page_idx": 9 + }, + { + "type": "text", + "text": "Quantitative Evaluation. To quantitatively evaluate the effectiveness of our PECS, we conduct a comparison with four open-source event camera simulators (i.e., ESIM [31], VOLT [20], V2E [12], and ICNS [14]) on six recording sequences (see Table 1). Note that, our PECS outperforms four state-of-the-art simulators in both two event metrics. More precisely, in terms of the Gaussian distance, our PECS reduces the average by 0.434, 0.281, 0.236, and 0.142 when compared to ESIM, VOLT, V2E, and ICNS, respectively. Meanwhile, in the Chamfer distance, our simulator exhibits an average decrease of 48.749, 13.038, 8.212, and 3.104 compared to ESIM, VOLT, V2E, and ICNS, respectively.", + "bbox": [ + 212, + 357, + 785, + 492 + ], + "page_idx": 9 + }, + { + "type": "text", + "text": "Visualization Evaluation. Some representative visualization results on two motion scenarios (i.e., translating checkerboards and rotation disks) recorded in a professional optical laboratory are illustrated in Fig. 3 and Fig. 4. Obviously, our PECS achieves the best performance against four state-of-the-art simulators including ESIM, VOLT, V2E, and ICNS. We can find that V2E and ESIM simulators are too primitive, resulting in concentrated data distribution in 3D space and a lack of continuity. The event data of ESIM is entirely concentrated at the timestamp of the input frame, lacking continuity, which is the main reason for its poor performance. VOLT performs well in terms of the continuity of event output, but its noise model suffers from significant distortion. Additionally, VOLT simply converts three-channel video to grayscale as photocurrent, without considering quantum efficiency and spectrum. Although the ICNS has an advanced sensor model and supports scene input, the lack of algorithms in the section leads to less accurate event data generation.", + "bbox": [ + 212, + 493, + 785, + 705 + ], + "page_idx": 9 + }, + { + "type": "text", + "text": "4.3 Ablation Study", + "text_level": 1, + "bbox": [ + 215, + 724, + 390, + 742 + ], + "page_idx": 9 + }, + { + "type": "text", + "text": "Influence of the Sampling Number $N$ . To analyze the sampling number $N$ of numerical integration in our PECS, we set the multispectral rendering module with various values of $N$ , and set $C$ as 100. As described in Table 2, we find that the Chamfer distance and Gaussian distance decrease with the increase of N. More precisely, Comparing $N = 31$ and $N = 6$ , the Chamfer distance and Gaussian distance decrease by 6.58 and 0.619, respectively. Overall, this results show that", + "bbox": [ + 212, + 748, + 785, + 839 + ], + "page_idx": 9 + }, + { + "type": "page_number", + "text": "10", + "bbox": [ + 217, + 114, + 235, + 126 + ], + "page_idx": 9 + }, + { + "type": "header", + "text": "Haiqian. Author et al.", + "bbox": [ + 271, + 114, + 419, + 128 + ], + "page_idx": 9 + }, + { + "type": "image", + "img_path": "images/6b8abcfd0b45c52e32bbfc5c50e3ec8642a4f90255d05aebba948309acb0dbc8.jpg", + "image_caption": [ + "Fig. 3: Representative examples of translating checkerboards with various motion speed and light conditions. In each scenario, the first row is a 2D visualization schematic of the event data, and the second row is a 3D visualization schematic of the event data." + ], + "image_footnote": [], + "bbox": [ + 215, + 143, + 785, + 392 + ], + "page_idx": 10 + }, + { + "type": "table", + "img_path": "images/0ce6e6421832c93f9d6a5e2be57e96dcde4ff52c083c81b640701a065692eea2.jpg", + "table_caption": [ + "Table 2: The influence of the sampling number $N$ of Monte-Carlo integration in the multispectral rendering module. It's worth noting that our PECS exhibits improved performance with an increase in $N$ ." + ], + "table_footnote": [], + "table_body": "
N13467111631
Chamfer distance ↓13.968.82616.288.5888.4662.8692.4762.008
Gaussian distance ↓1.5971.6241.5691.5831.5821.1761.0910.964
", + "bbox": [ + 217, + 523, + 785, + 583 + ], + "page_idx": 10 + }, + { + "type": "text", + "text": "the larger the $N$ , the closer the final simulation data is to the real data. Of course, an increase in $N$ also entails a rise in time complexity. Thus, there is often a consideration of the trade-off between accuracy and computational speed.", + "bbox": [ + 212, + 623, + 784, + 670 + ], + "page_idx": 10 + }, + { + "type": "text", + "text": "Influence of the Hyperparameter $C$ . To analyze the effect of the hyperparameter $C$ of numerical integration on the final performance, we set the multispectral rendering module with various values of $C$ , and set $N$ as 31. From Table 3, we can find that there is no simple correlation between the hyperparameter $C$ and the final performance. The distances of the two metrics gradually decrease with an increase in $C$ until they start to increase at nearly 100. Essentially speaking, the hyperparameter $C$ is introduced to align the standard unit system with the unit system used in the actual code. Theoretical analysis indicates the existence of an optimal value, a notion supported by ablation experiment results. The optimal value is observed to be around 100, and the distances of the two metrics decrease as the value of $C$ deviates from this optimal point.", + "bbox": [ + 212, + 672, + 787, + 840 + ], + "page_idx": 10 + }, + { + "type": "header", + "text": "Physical-Based Event Camera Simulator", + "bbox": [ + 460, + 114, + 730, + 128 + ], + "page_idx": 10 + }, + { + "type": "page_number", + "text": "11", + "bbox": [ + 767, + 116, + 782, + 126 + ], + "page_idx": 10 + }, + { + "type": "image", + "img_path": "images/fb5a6526e6be2c8b9b9f46ccfd2a5197331884eaae76e58c664a80883829baac.jpg", + "image_caption": [ + "Fig. 4: Representative visualization results of rotation disks with various motion speed and light conditions. In each scenario, the first row is a 2D visualization schematic of the event data, and the second row is a 3D visualization schematic of the event data." + ], + "image_footnote": [], + "bbox": [ + 212, + 145, + 785, + 392 + ], + "page_idx": 11 + }, + { + "type": "table", + "img_path": "images/5eb0bbb5fda3a0e0a2d31c856366bea114b61a5fb8af7a83f33f8074a594180a.jpg", + "table_caption": [ + "Table 3: The influence of the hyperparameter $C$ of Monte-Carlo integration in the multispectral rendering module. The distances gradually decrease with an increase in $C$ until they start to increase at nearly 100." + ], + "table_footnote": [], + "table_body": "
C10306080100120140160
Chamfer distance ↓6.4937.4452.5085.0662.0082.0112.2642.396
Gaussian distance ↓1.3421.4881.1061.4080.9640.9861.0501.076
", + "bbox": [ + 215, + 512, + 787, + 573 + ], + "page_idx": 11 + }, + { + "type": "text", + "text": "5 Downstream Applications", + "text_level": 1, + "bbox": [ + 215, + 598, + 503, + 616 + ], + "page_idx": 11 + }, + { + "type": "text", + "text": "In this section, we integrate our PECS with the Unreal Engine (UE) platform to generate extensive event-based vision datasets that support multiple tasks. We then proceed to validate the simulator's performance on the event-based video reconstruction task. In contrast to existing simulators, the representative algorithm trained on our synthetic events performs favorably on real event data. Extensive Multi-task Synthetic Datasets. To seamlessly link our PECS with downstream computer vision tasks, we incorporate it into the widely used UE platform, a popular tool for virtual environments. Our PECS, excluding the lens simulation module, is deployed on the UE platform to achieve real-time generation of multi-modal datasets (e.g., RGB frames and events). To facilitate various visual tasks, we also develop an automatic label annotation tool on the UE platform, providing bounding boxes for object detection and scene distance values for depth estimation. As depicted in Fig. 5, we present the synthetic data of a stereo hybrid camera on a drone platform, which mainly includes", + "bbox": [ + 212, + 628, + 787, + 840 + ], + "page_idx": 11 + }, + { + "type": "page_number", + "text": "12", + "bbox": [ + 217, + 114, + 235, + 126 + ], + "page_idx": 11 + }, + { + "type": "header", + "text": "Haiqian. Author et al.", + "bbox": [ + 271, + 114, + 419, + 128 + ], + "page_idx": 11 + }, + { + "type": "image", + "img_path": "images/caa842c3ed36fb0f88478f1f87c390f1a8a0ae322ae0c16a8dca2560d6f13875.jpg", + "image_caption": [ + "Left RGB image" + ], + "image_footnote": [], + "bbox": [ + 220, + 157, + 330, + 222 + ], + "page_idx": 12 + }, + { + "type": "image", + "img_path": "images/27c06e9ff24039c33754a48d733035fa078da15166e2f299225851d7c93d2e6c.jpg", + "image_caption": [ + "Right RGB image" + ], + "image_footnote": [], + "bbox": [ + 334, + 157, + 442, + 222 + ], + "page_idx": 12 + }, + { + "type": "image", + "img_path": "images/3f5501a465017a932a117e8107182a85582cf74e15adecb1f36447518daf233b.jpg", + "image_caption": [ + "Event image" + ], + "image_footnote": [], + "bbox": [ + 444, + 157, + 553, + 222 + ], + "page_idx": 12 + }, + { + "type": "image", + "img_path": "images/a9c6256610cb4901a794d39bfa4122ef72e7a0291be4003bbaa1603014d2ebbe.jpg", + "image_caption": [ + "Object bounding boxes" + ], + "image_footnote": [], + "bbox": [ + 557, + 157, + 669, + 222 + ], + "page_idx": 12 + }, + { + "type": "image", + "img_path": "images/0d56cc1591bb9afd15f027e06cecc87cb584bc438f14d6b464888f5b730e6a7e.jpg", + "image_caption": [ + "Depth label" + ], + "image_footnote": [], + "bbox": [ + 669, + 157, + 781, + 222 + ], + "page_idx": 12 + }, + { + "type": "image", + "img_path": "images/64cb9e86a0805601ed6d072b082da28d3b2114d5bb6219683643d25531590809.jpg", + "image_caption": [ + "Fig. 5: Representative examples of multi-task synthetic datasets. We integrate our PECS into the UE platform to generate extensive synthetic datasets supporting multiple tasks, such as video reconstruction, object detection, and depth estimation." + ], + "image_footnote": [], + "bbox": [ + 220, + 224, + 330, + 287 + ], + "page_idx": 12 + }, + { + "type": "image", + "img_path": "images/7c411251c60e7c11683fafe1f4529cb5d113c1baa96db1b41e416471fe307101.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 334, + 224, + 442, + 287 + ], + "page_idx": 12 + }, + { + "type": "image", + "img_path": "images/3c197aed4f9ec1e4b3734499079ac6a6fbe4a0c1b736eb50f6b0d845310dad06.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 444, + 224, + 555, + 287 + ], + "page_idx": 12 + }, + { + "type": "image", + "img_path": "images/75f834cb943930e68af247341ba0477d90d1ae44bced09ed9dd0041f5b2e17ef.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 558, + 224, + 668, + 287 + ], + "page_idx": 12 + }, + { + "type": "image", + "img_path": "images/2929c1a2d862659b0f84a6a12b17af31a74210fe9779a19a73aec6e3551a9c6e.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 669, + 224, + 782, + 287 + ], + "page_idx": 12 + }, + { + "type": "table", + "img_path": "images/5f8859ab1fe3f0c0fabeff4fc55831732b5b0d21ffb3188655f24026bc61d5e4.jpg", + "table_caption": [ + "Table 4: Performance comparison on event-based video reconstruction. The E2VID network [32] trains on synthetic data from each event camera simulator and tests on three representative sequences of the real-world HS-ERGB dataset [45]." + ], + "table_footnote": [], + "table_body": "
SequencesPSNR↑SSIM↑LPIPS↓
ESIMV2EPECSESIMV2EPECSESIMV2EPECS
spinning_umbrella8.7568.4379.2360.4720.4710.4800.4510.4490.430
fountain_bellevue26.1395.8676.5640.2130.2110.2180.6750.6700.662
bridge_lake_0311.77012.68613.2430.5410.5600.5630.3320.2740.250
Mean8.8898.9979.6780.4090.4140.4200.4860.4640.447
", + "bbox": [ + 220, + 431, + 782, + 520 + ], + "page_idx": 12 + }, + { + "type": "text", + "text": "RGB images, event streams, object detection labels, and scene depth values. Notably, our PECS can generate comprehensive synthetic datasets using both monocular and stereo cameras, supporting a variety of vision tasks, including video reconstruction, object detection, monocular depth estimation, binocular depth estimation, and more. Besides, we also adjust the speed of motion cameras and scene light intensity, generating diverse datasets with high-speed motion and low-light conditions, showcasing the advantages of event cameras.", + "bbox": [ + 212, + 565, + 784, + 671 + ], + "page_idx": 12 + }, + { + "type": "text", + "text": "Validation on Event-based Video Reconstruction. Event-based video reconstruction [32,56] is a typical task that effectively validates the effectiveness of synthetic data from various simulators. Specifically, we first use our PECS and two competitive simulators (e.g., ESIM [31] and V2E [12]) to generate large-scale synthetic datasets in the UE platform. Then, we select a representative event-to-video reconstruction algorithm (i.e., E2VID [32]) to be trained on these simulated datasets respectively. The training details align with those suggested in E2VID [32]. Finally, we test various trained models on the real-world HS-ERGB dataset [45]. We use three metrics (i.e., PSNR, SSIM [46], and LPIPS [51]) to measure the performance of event-based video reconstruction (see Table 4). Obviously, our PECS shows superior generalization capability compared to the top", + "bbox": [ + 212, + 672, + 787, + 840 + ], + "page_idx": 12 + }, + { + "type": "header", + "text": "Physical-Based Event Camera Simulator", + "bbox": [ + 460, + 114, + 730, + 128 + ], + "page_idx": 12 + }, + { + "type": "page_number", + "text": "13", + "bbox": [ + 767, + 114, + 785, + 126 + ], + "page_idx": 12 + }, + { + "type": "image", + "img_path": "images/6929508043559870f592122e13a4601451cdc25859bd1d1f064583f99db2e244.jpg", + "image_caption": [ + "Fig. 6: Visual comparisons on event-based video reconstruction. The E2VID network [32] trains on simulated event data from each simulator and tests on the real-world HS-ERGB dataset [45] recorded by a Propheesee Gen4 camera and a RGB camera." + ], + "image_footnote": [], + "bbox": [ + 218, + 143, + 784, + 390 + ], + "page_idx": 13 + }, + { + "type": "text", + "text": "competitor V2E with an average increase of 0.681 in PSNR, $0.6\\%$ improvement in SSIM, and a $1.7\\%$ decrease in LPIPS. Furthermore, we present some visualization results on the real-world HS-ERGB dataset [45] recorded by a Prophesee Gen4 camera and an RGB camera. As shown in Fig. 6, our PECS-trained network produces visually pleasing images with finer details and fewer artifacts. As our PECS is designed from a physical-based perspective, it naturally promotes reconstructed images with natural image statistics. Hence, the utilization of our simulator proves more effective, enabling the learning of high-quality video reconstruction [11,23,27,47,48,50,55] even with limited or no real training samples.", + "bbox": [ + 212, + 476, + 787, + 613 + ], + "page_idx": 13 + }, + { + "type": "text", + "text": "6 Conclusion", + "text_level": 1, + "bbox": [ + 215, + 638, + 359, + 655 + ], + "page_idx": 13 + }, + { + "type": "text", + "text": "This paper proposes a novel Physical-based Event Camera Simulator (PECS) that can generate a highly realistic event stream by directly interfacing with the 3D scene. To the best of our knowledge, our PECS is the first event camera simulator to design a realistic lens simulation block and a multispectral rendering module. Experiments demonstrate that Our PECS consistently outperforms four state-of-the-art simulators in some scenarios with various motion speeds and light changes, showing that it is currently the most realistic event camera simulator. Besides, integrating our PECS into the Unreal Engine (UE) platform allows the generation of extensive multi-task datasets, showcasing its utility in downstream vision tasks. We also believe that our PECS provides a proof-of-principle tool for the next generation of neuromorphic cameras.", + "bbox": [ + 212, + 672, + 787, + 840 + ], + "page_idx": 13 + }, + { + "type": "page_number", + "text": "14", + "bbox": [ + 217, + 114, + 235, + 126 + ], + "page_idx": 13 + }, + { + "type": "header", + "text": "Haiqian. Author et al.", + "bbox": [ + 271, + 114, + 419, + 128 + ], + "page_idx": 13 + }, + { + "type": "text", + "text": "Acknowledgements", + "text_level": 1, + "bbox": [ + 217, + 143, + 401, + 162 + ], + "page_idx": 14 + }, + { + "type": "text", + "text": "This work is partially supported by the National Natural Science Foundation of China under Grant 20181310845.", + "bbox": [ + 215, + 176, + 787, + 208 + ], + "page_idx": 14 + }, + { + "type": "text", + "text": "References", + "text_level": 1, + "bbox": [ + 217, + 233, + 321, + 250 + ], + "page_idx": 14 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "1. Annamalai, L., Ramanathan, V., Thakur, C.S.: Event-lstm: An unsupervised and asynchronous learning-based representation for event-based data. IEEE RAL 7(2), 4678-4685 (2022)", + "2. Baek, S., Eshraghian, J.K., Thio, W., Sandamirskaya, Y., Iu, H.H., Lu, W.D.: A real-time retinomorphic simulator using a conductance-based discrete neuronal network. In: AICAS. pp. 79-83 (2020)", + "3. Bellekens, B., Spruyt, V., Berkvens, R., Penne, R., Weyn, M.: A benchmark survey of rigid 3d point cloud registration algorithms. Int. J. Adv. Intell. Syst 8(5), 118-127 (2015)", + "4. Bhattacharya, A., Madaan, R., Cladera, F., Vermprala, S., Bonatti, R., Daniilidis, K., Kapoor, A., Kumar, V., Matni, N., Gupta, J.K.: Evidnerf: Reconstructing event data with dynamic neural radiance fields (2023)", + "5. Bi, Y., Andreopoulos, Y.: Pix2nvs: Parameterized conversion of pixel-domain video frames to neuromorphic vision streams. In: ICIP. pp. 1990-1994 (2017)", + "6. Bocca, A., Baek, D.: Automated driving systems: Key advantages, limitations and risks. In: AEIT AUTOMOTIVE. pp. 1-6 (2019)", + "7. Gallego, G., Delbrück, T., Orchard, G., Bartolozzi, C., Taba, B., Censi, A., Leutenegger, S., Davison, A.J., Conradt, J., Daniilidis, K., et al.: Event-based vision: A survey. IEEE TPAMI 44(1), 154-180 (2020)", + "8. Gehrig, D., Gehrig, M., Hidalgo-Carrió, J., Scaramuzza, D.: Video to events: Recyclcng video datasets for event cameras. In: CVPR. pp. 3586-3595 (2020)", + "9. Greenspan, M., Yurick, M.: Approximate kd tree search for efficient icp. In: 3DIM. pp. 442-448 (2003)", + "0. Gu, D., Li, J., Zhang, Y., Tian, Y.: How to learn a domain-adaptive event simulator? In: ACM MM. pp. 1275-1283 (2021)", + "1. He, W., You, K., Qiao, Z., Jia, X., Zhang, Z., Wang, W., Lu, H., Wang, Y., Liao, J.: Timereplayer: Unlocking the potential of event cameras for video interpolation. In: CVPR. pp. 17804-17813 (2022)", + "2. Hu, Y., Liu, S.C., Delbruck, T.: v2e: From video frames to realistic dvs events. In: CVPRW. pp. 1312-1321 (2021)", + "3. Hwang, I., Kim, J., Kim, Y.M.: Ev-nerf: Event based neural radiance field. In: WACV. pp. 837-847 (2023)", + "4. Joubert, D., Marcireau, A., Ralph, N., Jolley, A., van Schaik, A., Cohen, G.: Event camera simulator improvements via characterized parameters. Frontiers in Neuroscience 15, 702765 (2021)", + "5. Kang, Z., Li, J., Zhu, L., Tian, Y.: Retinomorphic sensing: A novel paradigm for future multimedia computing. In: ACM MM. pp. 144-152 (2021)", + "6. Kolb, C., Mitchell, D., Hanrahan, P.: A realistic camera model for computer graphics. In: CGIT. pp. 317-324 (1995)", + "7. Li, D., Li, J., Tian, Y.: Sodformer: Streaming object detection with transformer using events and frames. IEEE TPAMI 45(11), 14020-14037 (2023)" + ], + "bbox": [ + 225, + 266, + 785, + 839 + ], + "page_idx": 14 + }, + { + "type": "header", + "text": "Physical-Based Event Camera Simulator", + "bbox": [ + 460, + 114, + 730, + 128 + ], + "page_idx": 14 + }, + { + "type": "page_number", + "text": "15", + "bbox": [ + 767, + 116, + 785, + 126 + ], + "page_idx": 14 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "18. Li, J., Li, J., Zhu, L., Xiang, X., Huang, T., Tian, Y.: Asynchronous spatiotemporal memory network for continuous event-based object detection. IEEE TIP 31, 2975-2987 (2022)", + "19. Li, W., Saeedi, S., McCormac, J., Clark, R., Tzoumanikas, D., Ye, Q., Huang, Y., Tang, R., Leutenegger, S.: Interiornet: Mega-scale multi-sensor photo-realistic indoor scenes dataset. arXiv (2018)", + "20. Lin, S., Ma, Y., Guo, Z., Wen, B.: Dvs-voltmeter: Stochastic process-based event simulator for dynamic vision sensors. In: ECCV. pp. 578-593 (2022)", + "21. Liu, P., Chen, G., Li, Z., Tang, H., Knoll, A.: Learning local event-based descriptor for patch-based stereo matching. In: ICRA. pp. 412-418 (2022)", + "22. Liu, X., Li, J., Fan, X., Tian, Y.: Event-based monocular dense depth estimation with recurrent transformers. In: arXiv (2022)", + "23. Mei, H., Wang, Z., Yang, X., Wei, X., Delbruck, T.: Deep polarization reconstruction with pdavis events. In: CVPR. pp. 22149-22158 (2023)", + "24. Mou, X., Feng, K., Yi, A., Wang, S., Chen, H., Hu, X., Guo, M., Chen, S., Suess, A.: Accurate event simulation using high-speed videos. Electronic Imaging 34(7), 242-1 (2022)", + "25. Mueggler, E., Rebecq, H., Gallego, G., Delbruck, T., Scaramuzza, D.: The event-camera dataset and simulator: Event-based data for pose estimation, visual odometry, and slam. IJRR 36(2), 142-149 (2017)", + "26. Muglikar, M., Gehrig, M., Gehrig, D., Scaramuzza, D.: How to calibrate your event camera. In: CVPR. pp. 1403-1409 (2021)", + "27. Pan, L., Hartley, R., Scheerlinck, C., Liu, M., Yu, X., Dai, Y.: High frame rate video reconstruction based on an event camera. IEEE TPAMI 44(5), 2519-2533 (2020)", + "28. Pantho, M.J.H., Mbongue, J.M., Bhowmik, P., Bobda, C.: Event camera simulator design for modeling attention-based inference architectures. Journal of Real-Time Image Processing 19(2), 363-374 (2022)", + "29. Peng, X., Wang, Y., Gao, L., Kneip, L.: Globally-optimal event camera motion estimation. In: ECCV. pp. 51-67 (2020)", + "30. Pharr, M., Jakob, W., Humphreys, G.: Physically based rendering: From theory to implementation (2023)", + "31. Rebecq, H., Gehrig, D., Scaramuzza, D.: Esim: An open event camera simulator. In: CoRL. pp. 969-982 (2018)", + "32. Rebecq, H., Ranftl, R., Koltun, V., Scaramuzza, D.: Events-to-video: Bringing modern computer vision to event cameras. In: CVPR. pp. 3857-3866 (2019)", + "33. Rebecq, H., Ranftl, R., Koltun, V., Scaramuzza, D.: High speed and high dynamic range video with an event camera. IEEE TPAMI 43(6), 1964-1980 (2019)", + "34. Rizzo, C., Schuman, C., Plank, J.: Event-based camera simulation wrapper for arcade learning environment. In: ICNS. pp. 1-5 (2022)", + "35. Rogalski, A., Razeghi, M.: Narrow-gap semiconductor photodiodes. In: Photodetectors: Materials and Devices III. vol. 3287, pp. 2-13 (1998)", + "36. Sasián, J.: Introduction to lens design. Cambridge University Press (2019)", + "37. Segal, A., Haehnel, D., Thrun, S.: Generalized-icp. In: RSS. vol. 2, p. 435. Seattle, WA (2009)", + "38. Sengupta, J., Liu, S., Andreou, A.: Retinosim: an event-based data synthesis tool for neuromorphic vision architecture exploration. In: ICNS. pp. 1-9 (2022)", + "39. Su, Q., Chou, Y., Hu, Y., Li, J., Mei, S., Zhang, Z., Li, G.: Deep directly-trained spiking neural networks for object detection. In: ICCV. pp. 6555-6565 (2023)" + ], + "bbox": [ + 215, + 146, + 785, + 839 + ], + "page_idx": 15 + }, + { + "type": "page_number", + "text": "16", + "bbox": [ + 217, + 114, + 235, + 126 + ], + "page_idx": 15 + }, + { + "type": "header", + "text": "Haiqian. Author et al.", + "bbox": [ + 271, + 114, + 419, + 127 + ], + "page_idx": 15 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "40. Sun, L., Sakaridis, C., Liang, J., Jiang, Q., Yang, K., Sun, P., Ye, Y., Wang, K., Gool, L.V.: Event-based fusion for motion deblurring with cross-modal attention. In: ECCV. pp. 412-428 (2022)", + "41. Sundar, V., Ardelean, A., Swedish, T., Bruschini, C., Charbon, E., Gupta, M.: Sodacam: Software-defined cameras via single-photon imaging. In: ICCV. pp. 8165-8176 (2023)", + "42. Ta, K., Bruggemann, D., Brodermann, T., Sakaridis, C., Van Gool, L.: L2e: Lasers to events for 6-dof extrinsic calibration of lidars and event cameras. In: ICRA. pp. 11425-11431 (2023)", + "43. Taverni, G., Moeys, D.P., Li, C., Cavaco, C., Motsnyi, V., Bello, D.S.S., Delbruck, T.: Front and back illuminated dynamic and active pixel vision sensors comparison. IEEE Trans. Circuits Syst. II: Exp. Briefs 65(5), 677-681 (2018)", + "44. Tsuji, Y., Yatagawa, T., Kubo, H., Morishima, S.: Event-based camera simulation using monte carlo path tracing with adaptive denoising. arXiv (2023)", + "45. Tulyakov, S., Bochicchio, A., Gehrig, D., Georgoulis, S., Li, Y., Scaramuzza, D.: Time lens++: Event-based frame interpolation with parametric non-linear flow and multi-scale fusion. In: CVPR. pp. 17755-17764 (2022)", + "46. Wang, Z., Bovik, A.C., Sheikh, H.R., Simoncelli, E.P.: Image quality assessment: From error visibility to structural similarity. IEEE TIP 13(4), 600-612 (2004)", + "47. Weng, W., Zhang, Y., Xiong, Z.: Event-based video reconstruction using transformer. In: ICCV. pp. 2563-2572 (2021)", + "48. Yu, L., Zhang, X., Liao, W., Yang, W., Xia, G.S.: Learning to see through with events. IEEE TPAMI 45(07), 8660-8678 (2023)", + "49. Yu, W., Li, J., Zhang, S., Ji, X.: Learning scale-aware spatio-temporal implicit representation for event-based motion deblurring. In: ICML (2024)", + "50. Zhang, H., Zhang, L., Dai, Y., Li, H., Koniusz, P.: Event-guided multi-patch network with self-supervision for non-uniform motion deblurring. IJCV 131(2), 453-470 (2023)", + "51. Zhang, R., Isola, P., Efros, A.A., Shechtman, E., Wang, O.: The unreasonable effectiveness of deep features as a perceptual metric. In: CVPR. pp. 586-595 (2018)", + "52. Zhang, Z., Cui, S., Chai, K., Yu, H., Dasgupta, S., Mahbub, U., Rahman, T.: V2ce: Video to continuous events simulator. arXiv (2023)", + "53. Zheng, X., Liu, Y., Lu, Y., Hua, T., Pan, T., Zhang, W., Tao, D., Wang, L.: Deep learning for event-based vision: A comprehensive survey and benchmarks. arXiv (2023)", + "54. Zhu, A.Z., Wang, Z., Khant, K., Daniilidis, K.: Eventgan: Leveraging large scale image datasets for event cameras. In: ICCP. pp. 1-11 (2021)", + "55. Zhu, L., Li, J., Wang, X., Huang, T., Tian, Y.: Neuspike-net: High speed video reconstruction via bio-inspired neuromorphic cameras. In: ICCV. pp. 2400-2409 (2021)", + "56. Zhu, L., Wang, X., Chang, Y., Li, J., Huang, T., Tian, Y.: Event-based video reconstruction via potential-assisted spiking neural network. In: CVPR. pp. 3594-3604 (2022)", + "57. Ziegler, A., Teigland, D., Tebbe, J., Gossard, T., Zell, A.: Real-time event simulation with frame-based cameras. In: ICRA. pp. 11669-11675 (2023)" + ], + "bbox": [ + 212, + 146, + 787, + 770 + ], + "page_idx": 16 + }, + { + "type": "header", + "text": "Physical-Based Event Camera Simulator", + "bbox": [ + 460, + 114, + 730, + 127 + ], + "page_idx": 16 + }, + { + "type": "page_number", + "text": "17", + "bbox": [ + 767, + 116, + 784, + 126 + ], + "page_idx": 16 + } +] \ No newline at end of file diff --git a/2024/Physical-Based Event Camera Simulator/0986a08f-4884-431b-9c20-c64d0c63357b_model.json b/2024/Physical-Based Event Camera Simulator/0986a08f-4884-431b-9c20-c64d0c63357b_model.json new file mode 100644 index 0000000000000000000000000000000000000000..5fcbe116cc6a3d0c637acfb22d2a5c144cd088ec --- /dev/null +++ b/2024/Physical-Based Event Camera Simulator/0986a08f-4884-431b-9c20-c64d0c63357b_model.json @@ -0,0 +1,2489 @@ +[ + [ + { + "type": "title", + "bbox": [ + 0.27, + 0.142, + 0.734, + 0.163 + ], + "angle": 0, + "content": "Physical-Based Event Camera Simulator" + }, + { + "type": "text", + "bbox": [ + 0.223, + 0.189, + 0.782, + 0.222 + ], + "angle": 0, + "content": "Haiqian Han\\(^{1}\\), Jiacheng Lyu\\(^{1}\\), Jianing Li\\(^{1}\\) \\((\\boxtimes)\\), Henglu Wei\\(^{1}\\), Cheng Li\\(^{2}\\), Yajing Wei\\(^{2}\\), Shu Chen\\(^{2}\\), and Xiangyang Ji\\(^{1}\\) \\((\\boxtimes)\\)" + }, + { + "type": "text", + "bbox": [ + 0.269, + 0.232, + 0.733, + 0.248 + ], + "angle": 0, + "content": "\\(^{1}\\) Tsinghua University, Haidian District, Beijing, 100084, P. R. China" + }, + { + "type": "text", + "bbox": [ + 0.355, + 0.248, + 0.646, + 0.261 + ], + "angle": 0, + "content": "2 Beijing Xiaomi Mobile Software Co., Ltd." + }, + { + "type": "text", + "bbox": [ + 0.262, + 0.304, + 0.744, + 0.526 + ], + "angle": 0, + "content": "Abstract. Existing event camera simulators primarily focus on the process of generating video events and often overlook the entire optical path in real-world camera systems. To address this limitation, we propose a novel Physical-based Event Camera Simulator (PECS), which is able to generate a high-fidelity realistic event stream by directly interfacing with the 3D scene. Our PECS features a lens simulation block for accurate light-to-sensor chip replication and a multispectral rendering module for precise photocurrent generation. We present two spatiotemporal event metrics to assess the similarity between simulated and actual camera events. Experimental results demonstrate that our PECS outperforms four state-of-the-art simulators by a large margin in terms of event-based signal fidelity. We integrate our PECS into the UE platform to generate extensive multi-task synthetic datasets and evaluate its effectiveness in downstream vision tasks (e.g., video reconstruction). Our open-source code is available at https://github.com/lanpokn/PECS_trail_version." + }, + { + "type": "text", + "bbox": [ + 0.261, + 0.54, + 0.741, + 0.569 + ], + "angle": 0, + "content": "Keywords: Event camera simulation \\(\\cdot\\) Physics-based vision \\(\\cdot\\) Hyper-spectral data analysis" + }, + { + "type": "title", + "bbox": [ + 0.217, + 0.597, + 0.376, + 0.612 + ], + "angle": 0, + "content": "1 Introduction" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.632, + 0.788, + 0.722 + ], + "angle": 0, + "content": "Event cameras [7,43], namely bio-inspired vision sensors, operate fundamentally differently from conventional cameras. Instead of capturing intensity images at a fixed rate, event cameras respond to brightness changes with a stream of asynchronous events. With the advantages of high temporal resolution, high dynamic range, and low power consumption [33], event cameras have found widespread use in various computer vision tasks [4, 13, 17, 18, 21, 22, 25, 29, 39, 40, 49]." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.723, + 0.788, + 0.814 + ], + "angle": 0, + "content": "Despite notable advancements in event-based vision, the training of deep learning-based approaches still demands a substantial amount of synthetic event data [19, 53]. It is worth noting that the high cost [43] and deployment challenges [6] associated with event cameras in high-speed or low-light scenes limit the availability of real-world datasets. Thus, several event camera simulators have attempted to generate a large amount of affordable and reliable event data." + }, + { + "type": "page_footnote", + "bbox": [ + 0.232, + 0.825, + 0.616, + 0.84 + ], + "angle": 0, + "content": "\\((\\boxtimes)\\) Corresponding author: Xiangyang Ji and Jianing Li." + } + ], + [ + { + "type": "page_number", + "bbox": [ + 0.218, + 0.116, + 0.23, + 0.127 + ], + "angle": 0, + "content": "2" + }, + { + "type": "header", + "bbox": [ + 0.272, + 0.115, + 0.42, + 0.129 + ], + "angle": 0, + "content": "Haiqian. Author et al." + }, + { + "type": "image", + "bbox": [ + 0.218, + 0.145, + 0.787, + 0.226 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.214, + 0.236, + 0.788, + 0.279 + ], + "angle": 0, + "content": "Fig. 1: Event camera simulators. (a) The existing video to events simulators (e.g., V2E [12]). (b) Our Physical Event Camera Simulator (PECS) is designed to generate high-fidelity realistic events by directly interfacing with the 3D scene." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.311, + 0.789, + 0.476 + ], + "angle": 0, + "content": "One category refers to optimized-based event camera simulators [5, 12, 14, 15, 20, 31, 38, 41, 57], where the objective is to design hand-crafted modules for generating event data. For example, V2E [12] converts intensity frames into event data via multiple hand-crafted modules. Note that, these simulators take three-channel videos as direct inputs, overlooking the complete optical path in real-world camera systems. It may lead to the loss of optical information, consequently seriously affecting the simulator's accuracy. While some efforts [14, 44] support offline rendering in the 3D scene, they often neglect the consideration of spectrum data in the rendering module. In fact, integrating spectrum information into the event camera simulation process has the potential to enhance the fidelity of synthetic event streams." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.478, + 0.789, + 0.658 + ], + "angle": 0, + "content": "Another category is learning-based event camera simulators [2, 8, 10, 28, 34, 52, 54] that aim to generate event representations, improving the generalization capabilities of deep learning models directly for the target domain rather than relying on raw events. For instance, EventGAN [54] is an end-to-end neural network that directly converts images to event presentations for downstream computer vision tasks. However, these learning-based simulators, requiring retraining for diverse usage cases, may have limited generalization across different scenarios. Moreover, the generated event representations are not in the original signal domain [1], posing potential limitations for broader applications. In other words, physical-based simulators, producing spatiotemporal events, preserve the inherent characteristics of raw camera data, enhancing their versatility compared to event representations by learning-based simulators." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.66, + 0.79, + 0.841 + ], + "angle": 0, + "content": "To address the aforementioned problems, this paper proposes a novel Physical-based Event Camera Simulator (PECS), which is able to generate a highly realistic event stream by directly interfacing with the 3D scene (see Fig. 1). In fact, the goal of this work is not to optimize hand-crafted event camera simulators (e.g., ESIM or V2E). In contrast, we aim at overcoming the following challenges: (i) How do we model a realistic lens to replicate the light-to-sensor chip process? (ii) How do we design a multispectral renderer to obtain high-precision photocurrents rather than three-channel videos? Towards this end, a realistic lens simulation block using the PBRT renderer [30] is developed to accurately replicate the light-to-sensor chip process. Then, a novel multispectral rendering module is designed to generate high-precision photocurrents by quantum efficiency and Monte Carlo integration. Two asynchronous spatiotemporal event metrics (i.e., Chamfer dis" + } + ], + [ + { + "type": "header", + "bbox": [ + 0.461, + 0.115, + 0.732, + 0.129 + ], + "angle": 0, + "content": "Physical-Based Event Camera Simulator" + }, + { + "type": "page_number", + "bbox": [ + 0.775, + 0.117, + 0.785, + 0.127 + ], + "angle": 0, + "content": "3" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.147, + 0.784, + 0.252 + ], + "angle": 0, + "content": "tance and Gaussian distance) are proposed to measure the similarity between simulated and raw events. Extensive experiments show that our PECS outperforms four state-of-the-art simulators by a large margin in terms of event-based signal fidelity. Moreover, we further deploy real-time versions into the Unreal Engine (UE) platform to generate extensive multi-task synthetic datasets. Typical computer vision tasks (i.e., event-based video reconstruction) are conducted to verify the effectiveness of our PECS." + }, + { + "type": "text", + "bbox": [ + 0.239, + 0.253, + 0.753, + 0.267 + ], + "angle": 0, + "content": "Overall, the main contributions of this work are summarized as follows:" + }, + { + "type": "text", + "bbox": [ + 0.226, + 0.279, + 0.786, + 0.323 + ], + "angle": 0, + "content": "- We propose a novel Physical-based Event Camera Simulator (PECS), capable of generating a high-fidelity realistic event stream by directly interfacing with the 3D scene, without relying on video-to-event conversion." + }, + { + "type": "text", + "bbox": [ + 0.226, + 0.324, + 0.786, + 0.353 + ], + "angle": 0, + "content": "- We design a realistic lens simulation block using the PBRT renderer, optimizing pupil sampling to improve the quality and speed of rendering." + }, + { + "type": "text", + "bbox": [ + 0.226, + 0.354, + 0.794, + 0.383 + ], + "angle": 0, + "content": "- We design a novel multispectral rendering module that generates high-precision photocurrents through quantum efficiency and Monte Carlo integration." + }, + { + "type": "text", + "bbox": [ + 0.226, + 0.384, + 0.786, + 0.428 + ], + "angle": 0, + "content": "- We present two asynchronous spatiotemporal metrics to assess the similarity between simulated and raw events. Extensive experiments show that our PECS outperforms four state-of-the-art simulators in terms of signal fidelity." + }, + { + "type": "list", + "bbox": [ + 0.226, + 0.279, + 0.794, + 0.428 + ], + "angle": 0, + "content": null + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.44, + 0.785, + 0.501 + ], + "angle": 0, + "content": "To the best of our knowledge, this is the first attempt to explore a physical-based event camera simulator. We believe that our simulator will provide high-fidelity realistic large-scale event data for event-based vision tasks and offer proof of principle for the next generation of neuromorphic cameras." + }, + { + "type": "title", + "bbox": [ + 0.215, + 0.523, + 0.388, + 0.539 + ], + "angle": 0, + "content": "2 Related Work" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.554, + 0.784, + 0.583 + ], + "angle": 0, + "content": "This section will review existing event camera simulators from two perspectives including optimized-based simulators and learning-based simulators." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.584, + 0.785, + 0.809 + ], + "angle": 0, + "content": "Optimized-based Event Camera Simulators. The prior optimized-based simulators include PIX2NVS [5] and ESIM [8, 31], which convert any image or video into asynchronous events. A realistic V2E [12] is developed to model the event camera in low-light or motion blur scenarios. ICNS [14] is presented to simulate noise and estimate the latency by adding the effects of the arbiter. VOLT [20] employs the concept of stochastic processes to model the output process of event data, thereby endowing the simulated event data with good continuity and probabilistic interpretability. Nevertheless, these event camera simulators directly consider three-channel videos as the input and overlook the complete optical path in real-world camera systems. Although Mou et al. [24] first notice the impact of quantum efficiency on event camera simulators and implement video-to-spectrum conversion. However, the lossless conversion from three-channel data to spectral data is nearly impossible. Thus, we design a novel multispectral rendering module that generates high-precision photocurrents through quantum efficiency and Monte Carlo integration." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.81, + 0.787, + 0.84 + ], + "angle": 0, + "content": "Learning-based Event Camera Simulators. With the rapid development of deep learning, some learning-based event camera simulators [2,8,10,28,34,52,54]" + } + ], + [ + { + "type": "page_number", + "bbox": [ + 0.218, + 0.116, + 0.23, + 0.127 + ], + "angle": 0, + "content": "4" + }, + { + "type": "header", + "bbox": [ + 0.272, + 0.115, + 0.42, + 0.129 + ], + "angle": 0, + "content": "Haiqian. Author et al." + }, + { + "type": "image", + "bbox": [ + 0.223, + 0.151, + 0.773, + 0.359 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.214, + 0.373, + 0.788, + 0.471 + ], + "angle": 0, + "content": "Fig. 2: The pipeline of the proposed Physical-based Event Camera Simulator (PECS). It starts with pupil sampling, precalculating the rays expected to pass through the lens. Ray tracing calculates light refraction and reflection within the lens and scene, producing multispectral maps. Incorporating quantum effects, these maps generate accurate photocurrents. Our PECS then leverages logarithmic differences to generate events, enabling the measurement of similarity between simulated and raw camera events in signal fidelity." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.502, + 0.789, + 0.683 + ], + "angle": 0, + "content": "are developed to generate event representations in the target domain. For instance, Zhu et al. [54] use the modified GAN model to generate synchronous event representations without raw events. Gu et al. [10] propose a domain-adaptive event simulator that can generate both event data and the camera's trajectory. Pantho et al. [28] use a neural network to generate event data represented by regions, which is convenient for use in moving region detection. While these simulators can enhance overall accuracy to a certain degree, the intrinsic challenge lies in the limited interpretability of neural networks. Consequently, these simulators, requiring re-training for diverse usage cases, may have limited generalization across different scenarios. In other words, simulators that rely on learning-based approaches are expected to exhibit lower robustness compared with optimized-based event camera simulators." + }, + { + "type": "title", + "bbox": [ + 0.216, + 0.707, + 0.339, + 0.723 + ], + "angle": 0, + "content": "3 Methods" + }, + { + "type": "title", + "bbox": [ + 0.216, + 0.739, + 0.439, + 0.754 + ], + "angle": 0, + "content": "3.1 Framework Overview" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.765, + 0.788, + 0.842 + ], + "angle": 0, + "content": "This work aims at designing a novel Physical-based Event Camera Simulator, termed PECS, which is able to generate a highly realistic event stream by directly interfacing with the 3D scene. As illustrated in Fig. 2, our PECS mainly consists of two modules: a multispectral rendering module and a dynamic event generation module. In the multispectral rendering module, we employ pupil sampling" + } + ], + [ + { + "type": "header", + "bbox": [ + 0.461, + 0.115, + 0.732, + 0.129 + ], + "angle": 0, + "content": "Physical-Based Event Camera Simulator" + }, + { + "type": "page_number", + "bbox": [ + 0.775, + 0.117, + 0.785, + 0.127 + ], + "angle": 0, + "content": "5" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.147, + 0.788, + 0.375 + ], + "angle": 0, + "content": "as a pre-processing step to obtain a set of rays originating from the direction in which the lens may emit. Then, ray tracing is employed to compute the refraction and reflection of light within both the lens and the scene, allowing us to acquire multispectral maps. Meanwhile, multispectral maps with quantum effects generate high-precision photocurrent. Particularly, we design a realistic lens simulation block using the PBRT renderer, which aims at optimizing pupil sampling to improve the quality and speed of rendering. In the dynamic event generation module, we leverage the principle of logarithmic differences in the event camera to generate events. This module tackles the constraints of current simulation systems by incorporating spectrum data and accounting for quantum efficiency. This enhancement contributes to improved accuracy in estimating photocurrent, thereby creating a more realistic depiction of the physical processes involved in the conversion of light signals into photocurrent. Finally, two event metrics (i.e., Chamfer distance and Gaussian distance) are proposed to measure the distance between simulated and raw events in signal fidelity." + }, + { + "type": "title", + "bbox": [ + 0.216, + 0.397, + 0.557, + 0.41 + ], + "angle": 0, + "content": "3.2 Realistic Camera Lenses Simulation" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.422, + 0.785, + 0.497 + ], + "angle": 0, + "content": "Realistic camera lens simulation in computer graphics replicates actual camera lens behavior [16] by employing ray tracing to accurately simulate light paths through multiple lens elements. This approach takes into account the intricacies of the lens system, including refraction at interfaces, and improves image quality by estimating incident radiance along arbitrary rays." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.498, + 0.788, + 0.664 + ], + "angle": 0, + "content": "Our realistic camera lens simulation block models a lens system with various elements, tracing rays through them with multispectral information. It considers interactions until rays either exit the optical system or are absorbed by the aperture stop or lens housing. To reduce computational waste, our PECS performs pupil sampling before conducting the actual simulation. In this study, the pupil's definition is a little different from in optics [36]. It refers to all the directions from which rays emit from the lenses at any specific point on the sensor film, not just the center point. Let the sensor plane coordinates be denoted as \\((u,v)\\). The pupil is a set of feasible directions represented by points \\((x,y)\\) on the direction plane. Each point \\((x,y)\\) in the pupil corresponds to a unique direction. Thus, the mathematical definition of the pupil can be expressed as:" + }, + { + "type": "equation", + "bbox": [ + 0.351, + 0.677, + 0.785, + 0.693 + ], + "angle": 0, + "content": "\\[\nP u p i l (u, v) = \\{(x, y) \\mid R a y (\\{u, v \\}, \\{x, y \\})), \\tag {1}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.704, + 0.785, + 0.734 + ], + "angle": 0, + "content": "where \\( Ray(\\{u, v\\}, \\{x, y\\}) \\) denotes the ray originating from \\( (u, v) \\) and pointing towards \\( (x, y) \\) passing through the lenses." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.735, + 0.785, + 0.81 + ], + "angle": 0, + "content": "Pupil sampling is the algorithm that identifies the correct pupil for each point on the sensor film. In the case of lenses commonly used in event cameras (e.g., Prophesee Gen4), which are often centrally symmetric with circular apertures, representing the pupil shape as an ellipse simplifies the pupil sampling process. Therefore, the pupil can be approximated as follows:" + }, + { + "type": "equation", + "bbox": [ + 0.347, + 0.826, + 0.785, + 0.842 + ], + "angle": 0, + "content": "\\[\nP u p i l (u, v) = \\{(x, y) \\mid (x, y) \\in E l l i (p, a, b) \\}, \\tag {2}\n\\]" + } + ], + [ + { + "type": "page_number", + "bbox": [ + 0.218, + 0.116, + 0.23, + 0.127 + ], + "angle": 0, + "content": "6" + }, + { + "type": "header", + "bbox": [ + 0.272, + 0.115, + 0.42, + 0.129 + ], + "angle": 0, + "content": "Haiqian. Author et al." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.147, + 0.784, + 0.177 + ], + "angle": 0, + "content": "where \\(Elli(p, a, b)\\) is an ellipse centered at point \\(p\\) on the direction plane with semi-major axis length \\(a\\) and semi-minor axis length \\(b\\)." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.178, + 0.785, + 0.268 + ], + "angle": 0, + "content": "By uniformly sampling points on the beam direction plane, the exit pupil for each sampled point on the sensor film is efficiently computed. For other points, the exit pupil is determined through linear interpolation between the ellipse's center and its major and minor axes. Despite potential errors in the elliptical model for complex lenses, this method still notably decreases the time of ray tracing within the lens, thereby speeding up the lens system simulation." + }, + { + "type": "title", + "bbox": [ + 0.215, + 0.288, + 0.533, + 0.304 + ], + "angle": 0, + "content": "3.3 Multispectral Rendering Module" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.311, + 0.784, + 0.401 + ], + "angle": 0, + "content": "The multispectral rendering module transforms incoming light signals into photocurrent, adjusting for display convenience [12, 14, 31]. An event camera involves detecting photons at various frequencies, initiating photocurrent through the photoelectric effect. After filtering and logarithmic conversion, the currents are converted into photovoltages. The module concentrates on the photon-to-photocurrent conversion, excluding subsequent electronic circuit model steps." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.402, + 0.785, + 0.462 + ], + "angle": 0, + "content": "Most simulation systems estimate photocurrent using the L component of LUV values, but actual photon receivers don't convert different frequency lights into the same number of electrons, as revealed by the Quantum Efficiency (QE) curve. The definition of absolute quantum efficiency [35] can be described as:" + }, + { + "type": "equation", + "bbox": [ + 0.386, + 0.471, + 0.785, + 0.502 + ], + "angle": 0, + "content": "\\[\n\\mathrm {Q E} _ {\\text {a b s o l u t e}} (\\lambda) = \\frac {N _ {e}}{N _ {v}} = \\frac {R _ {\\lambda}}{\\lambda} \\times \\frac {h c}{e}, \\tag {3}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.508, + 0.785, + 0.582 + ], + "angle": 0, + "content": "where \\( N_{e} \\) is the number of electrons, \\( N_{v} \\) is the number of photons, \\( R_{\\lambda} \\) is the current per unit of incident light power (A/W), \\( \\lambda \\) is the wavelength in nm, \\( h \\) is the Planck constant, \\( c \\) is the speed of light in vacuum, and \\( e \\) is the elementary charge. The quantum efficiency is generally given as a relative value ranging from 0 to 1. Its definition can be formulated as follows:" + }, + { + "type": "equation", + "bbox": [ + 0.371, + 0.589, + 0.785, + 0.622 + ], + "angle": 0, + "content": "\\[\n\\mathrm {Q E} _ {\\text {r e l a t i v e}} (\\lambda) = \\frac {Q E _ {\\text {a b s o l u t e}} (\\lambda)}{\\operatorname* {m a x} (Q E _ {\\text {a b s o l u t e}} (\\lambda))}. \\tag {4}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.628, + 0.784, + 0.673 + ], + "angle": 0, + "content": "Our PECS overcomes limitations by providing multi-spectral data for precise photocurrent estimation. Leveraging known spectrum information and quantum efficiency, PECS directly computes photocurrent as follows:" + }, + { + "type": "equation", + "bbox": [ + 0.389, + 0.682, + 0.785, + 0.713 + ], + "angle": 0, + "content": "\\[\ni _ {p} = C \\cdot \\int \\lambda \\cdot Q E (\\lambda) \\cdot L (\\lambda) \\cdot d \\lambda , \\tag {5}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.719, + 0.784, + 0.794 + ], + "angle": 0, + "content": "where \\( i_p \\) is the photocurrent after Photoelectric conversion, which is a key parameter in subsequent electronic circuits. C is a constant that depends on the pixel area and units of physical quantities, and \\( L(\\lambda) \\) is the input spectrum. If L is represented by a finite set of discrete channels, the following simple numerical integration method can be used to achieve the aforementioned integration:" + }, + { + "type": "equation", + "bbox": [ + 0.34, + 0.804, + 0.785, + 0.843 + ], + "angle": 0, + "content": "\\[\ni _ {p} \\approx C \\cdot \\sum_ {k = 1} ^ {N} \\lambda_ {k} \\cdot Q E (\\lambda_ {k}) \\cdot L (\\lambda_ {k}) \\cdot \\left(\\lambda_ {k + 1} - \\lambda_ {k}\\right), \\tag {6}\n\\]" + } + ], + [ + { + "type": "header", + "bbox": [ + 0.46, + 0.115, + 0.733, + 0.129 + ], + "angle": 0, + "content": "Physical-Based Event Camera Simulator" + }, + { + "type": "page_number", + "bbox": [ + 0.775, + 0.116, + 0.787, + 0.127 + ], + "angle": 0, + "content": "7" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.147, + 0.784, + 0.177 + ], + "angle": 0, + "content": "where \\(N\\) is the number of known channels of the spectrum. If \\(N = 3\\), the information carried by the spectrum degrades into a common RGB image." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.178, + 0.785, + 0.253 + ], + "angle": 0, + "content": "The spectrum representation can be more complex in certain applications, such as using neural networks for implicit continuous representations. Simple data integration may lead to insufficient sampling in regions with larger function values, causing significant errors. To address this, Monte Carlo integration and importance sampling are commonly employed as:" + }, + { + "type": "equation", + "bbox": [ + 0.368, + 0.266, + 0.786, + 0.306 + ], + "angle": 0, + "content": "\\[\ni _ {p} = \\int_ {a} ^ {b} f (x) d x = \\lim _ {N \\rightarrow \\infty} \\frac {1}{N} \\sum_ {i = 1} ^ {N} \\frac {f \\left(x _ {i}\\right)}{p \\left(x _ {i}\\right)}, \\tag {7}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.215, + 0.317, + 0.784, + 0.348 + ], + "angle": 0, + "content": "where \\(N\\) is the number of samples. \\(p(x)\\) is probability density function, which generated samples \\(x_{i}\\). Applying Eq. 7 to Eq. 5, it can be re-written as:" + }, + { + "type": "equation", + "bbox": [ + 0.353, + 0.361, + 0.786, + 0.401 + ], + "angle": 0, + "content": "\\[\ni _ {p} \\approx I _ {N} = C \\cdot \\frac {1}{N} \\sum_ {k = 1} ^ {N} \\frac {\\lambda_ {k} \\cdot Q E (\\lambda_ {k}) \\cdot L (\\lambda_ {k})}{p (\\lambda_ {k})}, \\tag {8}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.215, + 0.412, + 0.784, + 0.473 + ], + "angle": 0, + "content": "where \\( I_N \\) is an approximation. Choosing a well-behaved \\( p(\\lambda) \\) is crucial to minimize \\( N \\) while maintaining accuracy. In probability theory, it is equal to make the variance small. Assuming that different samples are independent and identically distributed, we can obtain the variance of \\( I_N \\) as:" + }, + { + "type": "equation", + "bbox": [ + 0.31, + 0.486, + 0.786, + 0.526 + ], + "angle": 0, + "content": "\\[\n\\operatorname {V a r} \\left(I _ {N}\\right) = \\frac {1}{N ^ {2}} \\sum_ {k = 1} ^ {N} \\operatorname {V a r} \\left(\\frac {\\lambda_ {k} \\cdot Q E \\left(\\lambda_ {k}\\right) \\cdot L \\left(\\lambda_ {k}\\right)}{p \\left(\\lambda_ {k}\\right)}\\right) = \\frac {\\sigma_ {N} ^ {2}}{N ^ {2}}, \\tag {9}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.215, + 0.538, + 0.784, + 0.584 + ], + "angle": 0, + "content": "In the optimal condition, \\(\\sigma_N^2 = 0\\), which means \\(p(\\lambda)\\) is totally equal to the original function except for scale. Thus, we can model \\(p(\\lambda)\\) as a piecewise constant function and estimate it by pre-calculating the original function as:" + }, + { + "type": "equation", + "bbox": [ + 0.31, + 0.596, + 0.786, + 0.615 + ], + "angle": 0, + "content": "\\[\np (\\lambda) = B \\cdot \\lambda_ {j} ^ {\\prime} \\cdot Q E \\left(\\lambda_ {j} ^ {\\prime}\\right) \\cdot L \\left(\\lambda_ {j} ^ {\\prime}\\right), j = \\operatorname {a r g m i n} _ {m} | \\lambda - \\lambda_ {m} ^ {\\prime} |, \\tag {10}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.215, + 0.624, + 0.784, + 0.67 + ], + "angle": 0, + "content": "where \\( m \\) is the channel number. For each \\( m \\), \\( L(\\lambda_m') \\) is pre-calculated through the implicit representation of the spectrum. B is a constant that ensures the integral of \\( p(\\lambda) \\) equals 1. Then, the photocurrent can be obtained as follows:" + }, + { + "type": "equation", + "bbox": [ + 0.252, + 0.684, + 0.786, + 0.723 + ], + "angle": 0, + "content": "\\[\ni _ {p} \\approx C \\cdot \\frac {1}{N} \\sum_ {k = 1} ^ {N} \\frac {\\lambda_ {k} \\cdot Q E (\\lambda_ {k}) \\cdot L (\\lambda_ {k})}{B \\cdot \\lambda_ {j _ {k}} ^ {\\prime} \\cdot Q E \\left(\\lambda_ {j _ {k}} ^ {\\prime}\\right) \\cdot L \\left(\\lambda_ {j _ {k}} ^ {\\prime}\\right)}, j _ {k} = \\operatorname {a r g m i n} _ {m} | \\lambda_ {k} - \\lambda_ {m} ^ {\\prime} |. \\tag {11}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.215, + 0.734, + 0.784, + 0.84 + ], + "angle": 0, + "content": "Note that, Monte Carlo integration in Eq. 11 and simple numerical integration in Eq. 6 each have their own advantages and disadvantages. When the spectral data is represented continuously, Monte Carlo integration produces significantly smaller errors with a sharp spectral distribution curve compared to simple integration. However, if the spectrum is only represented by discrete channels, the lack of original information renders Monte Carlo integration somewhat redundant, and simple numerical integration can be used instead." + } + ], + [ + { + "type": "page_number", + "bbox": [ + 0.218, + 0.116, + 0.23, + 0.127 + ], + "angle": 0, + "content": "8" + }, + { + "type": "header", + "bbox": [ + 0.272, + 0.115, + 0.42, + 0.129 + ], + "angle": 0, + "content": "Haiqian. Author et al." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.147, + 0.788, + 0.224 + ], + "angle": 0, + "content": "Note that, our PECS estimates photocurrent by considering spectral data and quantum efficiency, offering more accurate simulations of light signal conversion. Consequently, the accuracy is significantly enhanced compared to the LUV method. Meanwhile, the quantum efficiency of different cameras often varies and can generally be provided by the corresponding sensor manufacturers." + }, + { + "type": "title", + "bbox": [ + 0.216, + 0.247, + 0.627, + 0.263 + ], + "angle": 0, + "content": "3.4 Asynchronous Spatiotemporal Event Metric" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.274, + 0.789, + 0.336 + ], + "angle": 0, + "content": "To measure the similarity between simulated and raw camera events, we present two asynchronous spatiotemporal event metrics (i.e., Chamfer distance and Gaussian distance), which is similar to the metric part in iterative closest point tasks [3,37]. The Chamfer distance metric \\( C_D \\) can be described as:" + }, + { + "type": "equation", + "bbox": [ + 0.308, + 0.346, + 0.787, + 0.385 + ], + "angle": 0, + "content": "\\[\nC _ {D} (R, Q) = \\frac {1}{| R |} \\sum_ {r \\in R} \\min _ {q \\in Q} d (r, q) + \\frac {1}{| Q |} \\sum_ {q \\in Q} \\min _ {r \\in R} d (r, q), \\tag {12}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.215, + 0.395, + 0.789, + 0.457 + ], + "angle": 0, + "content": "where the lower \\( C_D \\) its value, the closer the two events are. \\( R \\) and \\( Q \\) are two point clouds, \\( r \\) or \\( q \\) is a single point in \\( R \\) or \\( Q \\). Each point can be expressed as a four-dimensional tuple \\( \\{x,y,p,t\\} \\) representing the spatial coordinates, polarity, and temporal information. The distance in the metric can be depicted as:" + }, + { + "type": "equation", + "bbox": [ + 0.435, + 0.468, + 0.787, + 0.487 + ], + "angle": 0, + "content": "\\[\nd (r, q) = | | r - q | | _ {2}, \\tag {13}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.216, + 0.496, + 0.793, + 0.527 + ], + "angle": 0, + "content": "By utilizing KDTREE [9] for acceleration, the time complexity is only \\( O(n \\log(n)) \\). Similarly, the Gaussian distance can be formulated as:" + }, + { + "type": "equation", + "bbox": [ + 0.286, + 0.539, + 0.787, + 0.575 + ], + "angle": 0, + "content": "\\[\nG _ {D} (R, Q) = \\frac {1}{| R |} \\sum_ {r \\in R} g \\left(\\min _ {q \\in Q} d (r, q)\\right) + \\frac {1}{| Q |} \\sum_ {q \\in Q} g \\left(\\min _ {r \\in R} d (r, q)\\right), \\tag {14}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.215, + 0.588, + 0.788, + 0.62 + ], + "angle": 0, + "content": "where the function \\( g \\) makes \\( G_{D}(R, Q) \\) still become lower when the two events get closer while reducing the impact of outliers, and it can be described as:" + }, + { + "type": "equation", + "bbox": [ + 0.405, + 0.629, + 0.787, + 0.664 + ], + "angle": 0, + "content": "\\[\ng (x) = 1 - \\exp \\left(- \\frac {| | x | | _ {2} ^ {2}}{\\sigma}\\right), \\tag {15}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.215, + 0.674, + 0.788, + 0.703 + ], + "angle": 0, + "content": "where \\(\\sigma\\) is a hyperparameter used to adjust the sensitivity range of the original metric. In this study, we set 0.4 as the default." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.705, + 0.788, + 0.795 + ], + "angle": 0, + "content": "In fact, these two metrics describe the attributes of asynchronous events from different dimensions and can be necessary to assess the similarity between raw and simulated data. The Gaussian distance exhibits relative stability against outliers compared to the Chamfer distance, making it more robust to noise. Nevertheless, its measurement results have a narrower range of variation, making it less distinctive in simultaneous evaluations of multiple simulators." + }, + { + "type": "text", + "bbox": [ + 0.215, + 0.796, + 0.788, + 0.842 + ], + "angle": 0, + "content": "Note that, the units of x, y, p, and t in the original data are different, and the range of values for t often far exceeds that of the other components, resulting in Eq. 13 and Eq. 14 almost do not consider information beyond t. To address this" + } + ], + [ + { + "type": "header", + "bbox": [ + 0.461, + 0.115, + 0.732, + 0.129 + ], + "angle": 0, + "content": "Physical-Based Event Camera Simulator" + }, + { + "type": "page_number", + "bbox": [ + 0.775, + 0.117, + 0.785, + 0.127 + ], + "angle": 0, + "content": "9" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.147, + 0.784, + 0.178 + ], + "angle": 0, + "content": "issue, it is necessary to normalize the original measurements first to resolve this scaling problem as follows:" + }, + { + "type": "equation", + "bbox": [ + 0.377, + 0.184, + 0.786, + 0.22 + ], + "angle": 0, + "content": "\\[\nL _ {\\beta} [ i ] = \\left(\\frac {L _ {\\beta} ^ {\\prime} [ i ] - \\operatorname* {m i n} L _ {\\beta} ^ {\\prime}}{\\operatorname* {m a x} L _ {\\beta} ^ {\\prime} - \\operatorname* {m i n} L _ {\\beta} ^ {\\prime} + \\epsilon}\\right) * \\alpha , \\tag {16}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.215, + 0.227, + 0.788, + 0.304 + ], + "angle": 0, + "content": "where \\( L_{\\beta}^{\\prime} \\) is the original measurements, \\( \\beta \\) can be \\( x, y, p \\) and \\( t \\) element. \\( L_{\\beta} \\) is the corresponding processed data. \\( i \\) is the i'th event of the raw data, which means \\( R[i] = \\{L_x[i], L_y[i], L_p[i], L_t[i]\\} \\). \\( \\epsilon \\) is a small number that is to avoid dividing zero, and \\( \\alpha \\) is a hyperparameter. In this work, we choose \\( \\alpha = 100 \\) for space \\( x, y \\) and polarity \\( p \\), and choose \\( \\alpha = 1000 \\) for time \\( t \\)." + }, + { + "type": "title", + "bbox": [ + 0.216, + 0.323, + 0.376, + 0.34 + ], + "angle": 0, + "content": "4 Experiments" + }, + { + "type": "title", + "bbox": [ + 0.216, + 0.351, + 0.441, + 0.367 + ], + "angle": 0, + "content": "4.1 Experimental Setting" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.372, + 0.789, + 0.841 + ], + "angle": 0, + "content": "Real-to-Sim Scenario Construction. Spatiotemporal synchronization is essential for quantitatively evaluating, we record the real-world event data using a Prophesee Gen4 camera in a professional optical laboratory. We capture six sequences including rotating disk and translating checkerboards with light changes and various motion speeds. For example, T_0.6_H means translating checkerboards, 0.6 meter-per-seconds, and high light. R_360_L means rotating disk, 360 rpm, and low light. To ensure consistency between the simulated and real-world scenes, we utilize the blender for high-fidelity 3D modeling. We export the scenes to PBRT and simulate a naturally calibrated camera using real lenses [26,42]. This approach enables us to acquire temporally and spatially aligned rendering results for use in subsequent sensor modules. More details of spatiotemporal alignment can be found in the supplementary material. In other methods, as regular RGB cameras struggle to capture clear images in these scenes, we use Blender's rendered RGB video as input. By manually calibrating pinhole camera parameters, we ensure temporally and spatially consistent clear images for simulation inputs, enabling the proper functioning of these simulators. Implementation Details. All experiments are conducted on a GeForce RTX 3050 Ti and 11th Gen Intel(R) Core(TM) i7-11800H CPU. We compare our PECS with four open-source event camera simulators (i.e., ESIM [31], VOLT [20], V2E [12], and ICNS [14]). In our PECS, the hyperparameters in Eq. 11 and Eq. 6 are set to \\( C = 100 \\) and \\( N = 31 \\) to make an accuracy-speed trade-off. Regarding the ESIM method, we extract its core code from its original ties to Unreal Engine (UE) and replace UE output with video output. For the V2E method, given the clarity and compactness of Blender-generated images, we disable the network interpolation module to prevent extremely slow simulation speed. In the ICNS method, all parameters have been adjusted to achieve optimal results. In the VOLT method, we select a set of parameters for the best performance. We present two event metrics (i.e., Chamfer distance and Gaussian distance) to quantitatively evaluate the similarity between raw events and synthetic data. Besides, we select three representative sequences of the HS-ERGB dataset [45] in the event-based video reconstruction task." + } + ], + [ + { + "type": "page_number", + "bbox": [ + 0.218, + 0.116, + 0.236, + 0.127 + ], + "angle": 0, + "content": "10" + }, + { + "type": "header", + "bbox": [ + 0.272, + 0.115, + 0.42, + 0.129 + ], + "angle": 0, + "content": "Haiqian. Author et al." + }, + { + "type": "table_caption", + "bbox": [ + 0.216, + 0.145, + 0.788, + 0.187 + ], + "angle": 0, + "content": "Table 1: Comparison with state-of-the-art simulators on our six recording sequences. Note that, our Physical Event Camera Simulator (PECS) shows superior performance on both metrics across a diverse range of movement speeds and lighting conditions." + }, + { + "type": "table", + "bbox": [ + 0.222, + 0.199, + 0.779, + 0.308 + ], + "angle": 0, + "content": "
SceneGaussian distance ↓Chamfer distance ↓
ESIMVOLTV2EICNSPECSESIMVOLTV2EICNSPECS
Rotate 360 High1.9951.9231.6711.5580.96045.51222.0866.3563.8162.010
Rotate 360 Low1.9651.9311.8311.6811.63135.08225.3917.7294.1313.708
Rotate 60 High1.9951.9111.8161.5741.50439.42520.8149.2974.2373.459
Trans 1mps High1.9631.6551.8271.8061.66762.20610.57315.4509.4533.975
Trans 06mps High1.9671.7231.8391.8371.79863.93910.19818.1549.5884.169
Trans 1mps Low1.9871.8141.7001.6631.71068.11510.95214.0719.1824.467
Average1.9791.8261.7811.6871.54552.38016.66911.8436.7353.631
" + }, + { + "type": "title", + "bbox": [ + 0.216, + 0.334, + 0.377, + 0.349 + ], + "angle": 0, + "content": "4.2 Effective Test" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.358, + 0.787, + 0.493 + ], + "angle": 0, + "content": "Quantitative Evaluation. To quantitatively evaluate the effectiveness of our PECS, we conduct a comparison with four open-source event camera simulators (i.e., ESIM [31], VOLT [20], V2E [12], and ICNS [14]) on six recording sequences (see Table 1). Note that, our PECS outperforms four state-of-the-art simulators in both two event metrics. More precisely, in terms of the Gaussian distance, our PECS reduces the average by 0.434, 0.281, 0.236, and 0.142 when compared to ESIM, VOLT, V2E, and ICNS, respectively. Meanwhile, in the Chamfer distance, our simulator exhibits an average decrease of 48.749, 13.038, 8.212, and 3.104 compared to ESIM, VOLT, V2E, and ICNS, respectively." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.494, + 0.787, + 0.706 + ], + "angle": 0, + "content": "Visualization Evaluation. Some representative visualization results on two motion scenarios (i.e., translating checkerboards and rotation disks) recorded in a professional optical laboratory are illustrated in Fig. 3 and Fig. 4. Obviously, our PECS achieves the best performance against four state-of-the-art simulators including ESIM, VOLT, V2E, and ICNS. We can find that V2E and ESIM simulators are too primitive, resulting in concentrated data distribution in 3D space and a lack of continuity. The event data of ESIM is entirely concentrated at the timestamp of the input frame, lacking continuity, which is the main reason for its poor performance. VOLT performs well in terms of the continuity of event output, but its noise model suffers from significant distortion. Additionally, VOLT simply converts three-channel video to grayscale as photocurrent, without considering quantum efficiency and spectrum. Although the ICNS has an advanced sensor model and supports scene input, the lack of algorithms in the section leads to less accurate event data generation." + }, + { + "type": "title", + "bbox": [ + 0.216, + 0.726, + 0.391, + 0.743 + ], + "angle": 0, + "content": "4.3 Ablation Study" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.749, + 0.787, + 0.84 + ], + "angle": 0, + "content": "Influence of the Sampling Number \\(N\\). To analyze the sampling number \\(N\\) of numerical integration in our PECS, we set the multispectral rendering module with various values of \\(N\\), and set \\(C\\) as 100. As described in Table 2, we find that the Chamfer distance and Gaussian distance decrease with the increase of N. More precisely, Comparing \\(N = 31\\) and \\(N = 6\\), the Chamfer distance and Gaussian distance decrease by 6.58 and 0.619, respectively. Overall, this results show that" + } + ], + [ + { + "type": "header", + "bbox": [ + 0.461, + 0.115, + 0.732, + 0.129 + ], + "angle": 0, + "content": "Physical-Based Event Camera Simulator" + }, + { + "type": "page_number", + "bbox": [ + 0.769, + 0.117, + 0.784, + 0.127 + ], + "angle": 0, + "content": "11" + }, + { + "type": "image", + "bbox": [ + 0.216, + 0.145, + 0.787, + 0.393 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.214, + 0.403, + 0.788, + 0.445 + ], + "angle": 0, + "content": "Fig. 3: Representative examples of translating checkerboards with various motion speed and light conditions. In each scenario, the first row is a 2D visualization schematic of the event data, and the second row is a 3D visualization schematic of the event data." + }, + { + "type": "table_caption", + "bbox": [ + 0.214, + 0.47, + 0.788, + 0.514 + ], + "angle": 0, + "content": "Table 2: The influence of the sampling number \\(N\\) of Monte-Carlo integration in the multispectral rendering module. It's worth noting that our PECS exhibits improved performance with an increase in \\(N\\)." + }, + { + "type": "table", + "bbox": [ + 0.218, + 0.525, + 0.787, + 0.584 + ], + "angle": 0, + "content": "
N13467111631
Chamfer distance ↓13.968.82616.288.5888.4662.8692.4762.008
Gaussian distance ↓1.5971.6241.5691.5831.5821.1761.0910.964
" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.624, + 0.785, + 0.671 + ], + "angle": 0, + "content": "the larger the \\(N\\), the closer the final simulation data is to the real data. Of course, an increase in \\(N\\) also entails a rise in time complexity. Thus, there is often a consideration of the trade-off between accuracy and computational speed." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.674, + 0.788, + 0.841 + ], + "angle": 0, + "content": "Influence of the Hyperparameter \\( C \\). To analyze the effect of the hyperparameter \\( C \\) of numerical integration on the final performance, we set the multispectral rendering module with various values of \\( C \\), and set \\( N \\) as 31. From Table 3, we can find that there is no simple correlation between the hyperparameter \\( C \\) and the final performance. The distances of the two metrics gradually decrease with an increase in \\( C \\) until they start to increase at nearly 100. Essentially speaking, the hyperparameter \\( C \\) is introduced to align the standard unit system with the unit system used in the actual code. Theoretical analysis indicates the existence of an optimal value, a notion supported by ablation experiment results. The optimal value is observed to be around 100, and the distances of the two metrics decrease as the value of \\( C \\) deviates from this optimal point." + } + ], + [ + { + "type": "page_number", + "bbox": [ + 0.218, + 0.116, + 0.236, + 0.127 + ], + "angle": 0, + "content": "12" + }, + { + "type": "header", + "bbox": [ + 0.272, + 0.115, + 0.42, + 0.129 + ], + "angle": 0, + "content": "Haiqian. Author et al." + }, + { + "type": "image", + "bbox": [ + 0.214, + 0.146, + 0.787, + 0.393 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.214, + 0.403, + 0.788, + 0.445 + ], + "angle": 0, + "content": "Fig. 4: Representative visualization results of rotation disks with various motion speed and light conditions. In each scenario, the first row is a 2D visualization schematic of the event data, and the second row is a 3D visualization schematic of the event data." + }, + { + "type": "table_caption", + "bbox": [ + 0.214, + 0.459, + 0.788, + 0.502 + ], + "angle": 0, + "content": "Table 3: The influence of the hyperparameter \\( C \\) of Monte-Carlo integration in the multispectral rendering module. The distances gradually decrease with an increase in \\( C \\) until they start to increase at nearly 100." + }, + { + "type": "table", + "bbox": [ + 0.216, + 0.513, + 0.789, + 0.574 + ], + "angle": 0, + "content": "
C10306080100120140160
Chamfer distance ↓6.4937.4452.5085.0662.0082.0112.2642.396
Gaussian distance ↓1.3421.4881.1061.4080.9640.9861.0501.076
" + }, + { + "type": "title", + "bbox": [ + 0.216, + 0.599, + 0.504, + 0.617 + ], + "angle": 0, + "content": "5 Downstream Applications" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.629, + 0.789, + 0.842 + ], + "angle": 0, + "content": "In this section, we integrate our PECS with the Unreal Engine (UE) platform to generate extensive event-based vision datasets that support multiple tasks. We then proceed to validate the simulator's performance on the event-based video reconstruction task. In contrast to existing simulators, the representative algorithm trained on our synthetic events performs favorably on real event data. Extensive Multi-task Synthetic Datasets. To seamlessly link our PECS with downstream computer vision tasks, we incorporate it into the widely used UE platform, a popular tool for virtual environments. Our PECS, excluding the lens simulation module, is deployed on the UE platform to achieve real-time generation of multi-modal datasets (e.g., RGB frames and events). To facilitate various visual tasks, we also develop an automatic label annotation tool on the UE platform, providing bounding boxes for object detection and scene distance values for depth estimation. As depicted in Fig. 5, we present the synthetic data of a stereo hybrid camera on a drone platform, which mainly includes" + } + ], + [ + { + "type": "header", + "bbox": [ + 0.461, + 0.115, + 0.732, + 0.129 + ], + "angle": 0, + "content": "Physical-Based Event Camera Simulator" + }, + { + "type": "page_number", + "bbox": [ + 0.769, + 0.116, + 0.786, + 0.127 + ], + "angle": 0, + "content": "13" + }, + { + "type": "image_caption", + "bbox": [ + 0.234, + 0.147, + 0.32, + 0.158 + ], + "angle": 0, + "content": "Left RGB image" + }, + { + "type": "image", + "bbox": [ + 0.222, + 0.158, + 0.331, + 0.223 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.344, + 0.147, + 0.435, + 0.158 + ], + "angle": 0, + "content": "Right RGB image" + }, + { + "type": "image", + "bbox": [ + 0.335, + 0.158, + 0.444, + 0.223 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.466, + 0.147, + 0.531, + 0.158 + ], + "angle": 0, + "content": "Event image" + }, + { + "type": "image", + "bbox": [ + 0.446, + 0.158, + 0.555, + 0.223 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.557, + 0.147, + 0.675, + 0.158 + ], + "angle": 0, + "content": "Object bounding boxes" + }, + { + "type": "image", + "bbox": [ + 0.558, + 0.158, + 0.67, + 0.223 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.698, + 0.147, + 0.758, + 0.158 + ], + "angle": 0, + "content": "Depth label" + }, + { + "type": "image", + "bbox": [ + 0.671, + 0.158, + 0.782, + 0.223 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.221, + 0.225, + 0.331, + 0.289 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.335, + 0.225, + 0.444, + 0.289 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.446, + 0.225, + 0.557, + 0.289 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.559, + 0.225, + 0.669, + 0.289 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.671, + 0.225, + 0.783, + 0.289 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.215, + 0.304, + 0.785, + 0.346 + ], + "angle": 0, + "content": "Fig. 5: Representative examples of multi-task synthetic datasets. We integrate our PECS into the UE platform to generate extensive synthetic datasets supporting multiple tasks, such as video reconstruction, object detection, and depth estimation." + }, + { + "type": "table_caption", + "bbox": [ + 0.215, + 0.365, + 0.785, + 0.407 + ], + "angle": 0, + "content": "Table 4: Performance comparison on event-based video reconstruction. The E2VID network [32] trains on synthetic data from each event camera simulator and tests on three representative sequences of the real-world HS-ERGB dataset [45]." + }, + { + "type": "table", + "bbox": [ + 0.222, + 0.432, + 0.783, + 0.521 + ], + "angle": 0, + "content": "
SequencesPSNR↑SSIM↑LPIPS↓
ESIMV2EPECSESIMV2EPECSESIMV2EPECS
spinning_umbrella8.7568.4379.2360.4720.4710.4800.4510.4490.430
fountain_bellevue26.1395.8676.5640.2130.2110.2180.6750.6700.662
bridge_lake_0311.77012.68613.2430.5410.5600.5630.3320.2740.250
Mean8.8898.9979.6780.4090.4140.4200.4860.4640.447
" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.566, + 0.785, + 0.672 + ], + "angle": 0, + "content": "RGB images, event streams, object detection labels, and scene depth values. Notably, our PECS can generate comprehensive synthetic datasets using both monocular and stereo cameras, supporting a variety of vision tasks, including video reconstruction, object detection, monocular depth estimation, binocular depth estimation, and more. Besides, we also adjust the speed of motion cameras and scene light intensity, generating diverse datasets with high-speed motion and low-light conditions, showcasing the advantages of event cameras." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.674, + 0.788, + 0.842 + ], + "angle": 0, + "content": "Validation on Event-based Video Reconstruction. Event-based video reconstruction [32,56] is a typical task that effectively validates the effectiveness of synthetic data from various simulators. Specifically, we first use our PECS and two competitive simulators (e.g., ESIM [31] and V2E [12]) to generate large-scale synthetic datasets in the UE platform. Then, we select a representative event-to-video reconstruction algorithm (i.e., E2VID [32]) to be trained on these simulated datasets respectively. The training details align with those suggested in E2VID [32]. Finally, we test various trained models on the real-world HS-ERGB dataset [45]. We use three metrics (i.e., PSNR, SSIM [46], and LPIPS [51]) to measure the performance of event-based video reconstruction (see Table 4). Obviously, our PECS shows superior generalization capability compared to the top" + } + ], + [ + { + "type": "page_number", + "bbox": [ + 0.218, + 0.116, + 0.236, + 0.127 + ], + "angle": 0, + "content": "14" + }, + { + "type": "header", + "bbox": [ + 0.272, + 0.115, + 0.42, + 0.129 + ], + "angle": 0, + "content": "Haiqian. Author et al." + }, + { + "type": "image", + "bbox": [ + 0.219, + 0.145, + 0.785, + 0.391 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.216, + 0.403, + 0.788, + 0.446 + ], + "angle": 0, + "content": "Fig. 6: Visual comparisons on event-based video reconstruction. The E2VID network [32] trains on simulated event data from each simulator and tests on the real-world HS-ERGB dataset [45] recorded by a Propheesee Gen4 camera and a RGB camera." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.477, + 0.788, + 0.614 + ], + "angle": 0, + "content": "competitor V2E with an average increase of 0.681 in PSNR, \\(0.6\\%\\) improvement in SSIM, and a \\(1.7\\%\\) decrease in LPIPS. Furthermore, we present some visualization results on the real-world HS-ERGB dataset [45] recorded by a Prophesee Gen4 camera and an RGB camera. As shown in Fig. 6, our PECS-trained network produces visually pleasing images with finer details and fewer artifacts. As our PECS is designed from a physical-based perspective, it naturally promotes reconstructed images with natural image statistics. Hence, the utilization of our simulator proves more effective, enabling the learning of high-quality video reconstruction [11,23,27,47,48,50,55] even with limited or no real training samples." + }, + { + "type": "title", + "bbox": [ + 0.216, + 0.639, + 0.36, + 0.656 + ], + "angle": 0, + "content": "6 Conclusion" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.674, + 0.788, + 0.841 + ], + "angle": 0, + "content": "This paper proposes a novel Physical-based Event Camera Simulator (PECS) that can generate a highly realistic event stream by directly interfacing with the 3D scene. To the best of our knowledge, our PECS is the first event camera simulator to design a realistic lens simulation block and a multispectral rendering module. Experiments demonstrate that Our PECS consistently outperforms four state-of-the-art simulators in some scenarios with various motion speeds and light changes, showing that it is currently the most realistic event camera simulator. Besides, integrating our PECS into the Unreal Engine (UE) platform allows the generation of extensive multi-task datasets, showcasing its utility in downstream vision tasks. We also believe that our PECS provides a proof-of-principle tool for the next generation of neuromorphic cameras." + } + ], + [ + { + "type": "header", + "bbox": [ + 0.461, + 0.115, + 0.732, + 0.129 + ], + "angle": 0, + "content": "Physical-Based Event Camera Simulator" + }, + { + "type": "page_number", + "bbox": [ + 0.769, + 0.117, + 0.786, + 0.127 + ], + "angle": 0, + "content": "15" + }, + { + "type": "title", + "bbox": [ + 0.218, + 0.145, + 0.403, + 0.163 + ], + "angle": 0, + "content": "Acknowledgements" + }, + { + "type": "text", + "bbox": [ + 0.217, + 0.178, + 0.788, + 0.209 + ], + "angle": 0, + "content": "This work is partially supported by the National Natural Science Foundation of China under Grant 20181310845." + }, + { + "type": "title", + "bbox": [ + 0.218, + 0.234, + 0.323, + 0.25 + ], + "angle": 0, + "content": "References" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.267, + 0.786, + 0.308 + ], + "angle": 0, + "content": "1. Annamalai, L., Ramanathan, V., Thakur, C.S.: Event-lstm: An unsupervised and asynchronous learning-based representation for event-based data. IEEE RAL 7(2), 4678-4685 (2022)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.31, + 0.787, + 0.351 + ], + "angle": 0, + "content": "2. Baek, S., Eshraghian, J.K., Thio, W., Sandamirskaya, Y., Iu, H.H., Lu, W.D.: A real-time retinomorphic simulator using a conductance-based discrete neuronal network. In: AICAS. pp. 79-83 (2020)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.352, + 0.787, + 0.393 + ], + "angle": 0, + "content": "3. Bellekens, B., Spruyt, V., Berkvens, R., Penne, R., Weyn, M.: A benchmark survey of rigid 3d point cloud registration algorithms. Int. J. Adv. Intell. Syst 8(5), 118-127 (2015)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.394, + 0.787, + 0.434 + ], + "angle": 0, + "content": "4. Bhattacharya, A., Madaan, R., Cladera, F., Vermprala, S., Bonatti, R., Daniilidis, K., Kapoor, A., Kumar, V., Matni, N., Gupta, J.K.: Evidnerf: Reconstructing event data with dynamic neural radiance fields (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.436, + 0.787, + 0.463 + ], + "angle": 0, + "content": "5. Bi, Y., Andreopoulos, Y.: Pix2nvs: Parameterized conversion of pixel-domain video frames to neuromorphic vision streams. In: ICIP. pp. 1990-1994 (2017)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.464, + 0.787, + 0.49 + ], + "angle": 0, + "content": "6. Bocca, A., Baek, D.: Automated driving systems: Key advantages, limitations and risks. In: AEIT AUTOMOTIVE. pp. 1-6 (2019)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.492, + 0.787, + 0.533 + ], + "angle": 0, + "content": "7. Gallego, G., Delbrück, T., Orchard, G., Bartolozzi, C., Taba, B., Censi, A., Leutenegger, S., Davison, A.J., Conradt, J., Daniilidis, K., et al.: Event-based vision: A survey. IEEE TPAMI 44(1), 154-180 (2020)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.534, + 0.787, + 0.561 + ], + "angle": 0, + "content": "8. Gehrig, D., Gehrig, M., Hidalgo-Carrió, J., Scaramuzza, D.: Video to events: Recyclcng video datasets for event cameras. In: CVPR. pp. 3586-3595 (2020)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.562, + 0.787, + 0.589 + ], + "angle": 0, + "content": "9. Greenspan, M., Yurick, M.: Approximate kd tree search for efficient icp. In: 3DIM. pp. 442-448 (2003)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.59, + 0.787, + 0.616 + ], + "angle": 0, + "content": "0. Gu, D., Li, J., Zhang, Y., Tian, Y.: How to learn a domain-adaptive event simulator? In: ACM MM. pp. 1275-1283 (2021)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.617, + 0.787, + 0.658 + ], + "angle": 0, + "content": "1. He, W., You, K., Qiao, Z., Jia, X., Zhang, Z., Wang, W., Lu, H., Wang, Y., Liao, J.: Timereplayer: Unlocking the potential of event cameras for video interpolation. In: CVPR. pp. 17804-17813 (2022)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.66, + 0.787, + 0.687 + ], + "angle": 0, + "content": "2. Hu, Y., Liu, S.C., Delbruck, T.: v2e: From video frames to realistic dvs events. In: CVPRW. pp. 1312-1321 (2021)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.688, + 0.787, + 0.715 + ], + "angle": 0, + "content": "3. Hwang, I., Kim, J., Kim, Y.M.: Ev-nerf: Event based neural radiance field. In: WACV. pp. 837-847 (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.716, + 0.787, + 0.757 + ], + "angle": 0, + "content": "4. Joubert, D., Marcireau, A., Ralph, N., Jolley, A., van Schaik, A., Cohen, G.: Event camera simulator improvements via characterized parameters. Frontiers in Neuroscience 15, 702765 (2021)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.758, + 0.787, + 0.784 + ], + "angle": 0, + "content": "5. Kang, Z., Li, J., Zhu, L., Tian, Y.: Retinomorphic sensing: A novel paradigm for future multimedia computing. In: ACM MM. pp. 144-152 (2021)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.785, + 0.787, + 0.812 + ], + "angle": 0, + "content": "6. Kolb, C., Mitchell, D., Hanrahan, P.: A realistic camera model for computer graphics. In: CGIT. pp. 317-324 (1995)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.813, + 0.787, + 0.84 + ], + "angle": 0, + "content": "7. Li, D., Li, J., Tian, Y.: Sodformer: Streaming object detection with transformer using events and frames. IEEE TPAMI 45(11), 14020-14037 (2023)" + }, + { + "type": "list", + "bbox": [ + 0.226, + 0.267, + 0.787, + 0.84 + ], + "angle": 0, + "content": null + } + ], + [ + { + "type": "page_number", + "bbox": [ + 0.218, + 0.116, + 0.236, + 0.127 + ], + "angle": 0, + "content": "16" + }, + { + "type": "header", + "bbox": [ + 0.272, + 0.115, + 0.42, + 0.128 + ], + "angle": 0, + "content": "Haiqian. Author et al." + }, + { + "type": "ref_text", + "bbox": [ + 0.218, + 0.147, + 0.787, + 0.189 + ], + "angle": 0, + "content": "18. Li, J., Li, J., Zhu, L., Xiang, X., Huang, T., Tian, Y.: Asynchronous spatiotemporal memory network for continuous event-based object detection. IEEE TIP 31, 2975-2987 (2022)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.19, + 0.787, + 0.232 + ], + "angle": 0, + "content": "19. Li, W., Saeedi, S., McCormac, J., Clark, R., Tzoumanikas, D., Ye, Q., Huang, Y., Tang, R., Leutenegger, S.: Interiornet: Mega-scale multi-sensor photo-realistic indoor scenes dataset. arXiv (2018)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.233, + 0.786, + 0.26 + ], + "angle": 0, + "content": "20. Lin, S., Ma, Y., Guo, Z., Wen, B.: Dvs-voltmeter: Stochastic process-based event simulator for dynamic vision sensors. In: ECCV. pp. 578-593 (2022)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.261, + 0.786, + 0.289 + ], + "angle": 0, + "content": "21. Liu, P., Chen, G., Li, Z., Tang, H., Knoll, A.: Learning local event-based descriptor for patch-based stereo matching. In: ICRA. pp. 412-418 (2022)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.29, + 0.786, + 0.317 + ], + "angle": 0, + "content": "22. Liu, X., Li, J., Fan, X., Tian, Y.: Event-based monocular dense depth estimation with recurrent transformers. In: arXiv (2022)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.318, + 0.786, + 0.346 + ], + "angle": 0, + "content": "23. Mei, H., Wang, Z., Yang, X., Wei, X., Delbruck, T.: Deep polarization reconstruction with pdavis events. In: CVPR. pp. 22149-22158 (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.346, + 0.786, + 0.387 + ], + "angle": 0, + "content": "24. Mou, X., Feng, K., Yi, A., Wang, S., Chen, H., Hu, X., Guo, M., Chen, S., Suess, A.: Accurate event simulation using high-speed videos. Electronic Imaging 34(7), 242-1 (2022)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.388, + 0.786, + 0.43 + ], + "angle": 0, + "content": "25. Mueggler, E., Rebecq, H., Gallego, G., Delbruck, T., Scaramuzza, D.: The event-camera dataset and simulator: Event-based data for pose estimation, visual odometry, and slam. IJRR 36(2), 142-149 (2017)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.431, + 0.786, + 0.458 + ], + "angle": 0, + "content": "26. Muglikar, M., Gehrig, M., Gehrig, D., Scaramuzza, D.: How to calibrate your event camera. In: CVPR. pp. 1403-1409 (2021)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.459, + 0.786, + 0.5 + ], + "angle": 0, + "content": "27. Pan, L., Hartley, R., Scheerlinck, C., Liu, M., Yu, X., Dai, Y.: High frame rate video reconstruction based on an event camera. IEEE TPAMI 44(5), 2519-2533 (2020)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.501, + 0.786, + 0.543 + ], + "angle": 0, + "content": "28. Pantho, M.J.H., Mbongue, J.M., Bhowmik, P., Bobda, C.: Event camera simulator design for modeling attention-based inference architectures. Journal of Real-Time Image Processing 19(2), 363-374 (2022)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.544, + 0.786, + 0.571 + ], + "angle": 0, + "content": "29. Peng, X., Wang, Y., Gao, L., Kneip, L.: Globally-optimal event camera motion estimation. In: ECCV. pp. 51-67 (2020)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.571, + 0.786, + 0.599 + ], + "angle": 0, + "content": "30. Pharr, M., Jakob, W., Humphreys, G.: Physically based rendering: From theory to implementation (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.6, + 0.786, + 0.627 + ], + "angle": 0, + "content": "31. Rebecq, H., Gehrig, D., Scaramuzza, D.: Esim: An open event camera simulator. In: CoRL. pp. 969-982 (2018)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.628, + 0.786, + 0.656 + ], + "angle": 0, + "content": "32. Rebecq, H., Ranftl, R., Koltun, V., Scaramuzza, D.: Events-to-video: Bringing modern computer vision to event cameras. In: CVPR. pp. 3857-3866 (2019)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.657, + 0.786, + 0.684 + ], + "angle": 0, + "content": "33. Rebecq, H., Ranftl, R., Koltun, V., Scaramuzza, D.: High speed and high dynamic range video with an event camera. IEEE TPAMI 43(6), 1964-1980 (2019)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.685, + 0.786, + 0.712 + ], + "angle": 0, + "content": "34. Rizzo, C., Schuman, C., Plank, J.: Event-based camera simulation wrapper for arcade learning environment. In: ICNS. pp. 1-5 (2022)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.713, + 0.786, + 0.741 + ], + "angle": 0, + "content": "35. Rogalski, A., Razeghi, M.: Narrow-gap semiconductor photodiodes. In: Photodetectors: Materials and Devices III. vol. 3287, pp. 2-13 (1998)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.741, + 0.737, + 0.755 + ], + "angle": 0, + "content": "36. Sasián, J.: Introduction to lens design. Cambridge University Press (2019)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.756, + 0.786, + 0.783 + ], + "angle": 0, + "content": "37. Segal, A., Haehnel, D., Thrun, S.: Generalized-icp. In: RSS. vol. 2, p. 435. Seattle, WA (2009)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.784, + 0.786, + 0.812 + ], + "angle": 0, + "content": "38. Sengupta, J., Liu, S., Andreou, A.: Retinosim: an event-based data synthesis tool for neuromorphic vision architecture exploration. In: ICNS. pp. 1-9 (2022)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.813, + 0.786, + 0.84 + ], + "angle": 0, + "content": "39. Su, Q., Chou, Y., Hu, Y., Li, J., Mei, S., Zhang, Z., Li, G.: Deep directly-trained spiking neural networks for object detection. In: ICCV. pp. 6555-6565 (2023)" + }, + { + "type": "list", + "bbox": [ + 0.217, + 0.147, + 0.787, + 0.84 + ], + "angle": 0, + "content": null + } + ], + [ + { + "type": "header", + "bbox": [ + 0.461, + 0.115, + 0.732, + 0.128 + ], + "angle": 0, + "content": "Physical-Based Event Camera Simulator" + }, + { + "type": "page_number", + "bbox": [ + 0.769, + 0.117, + 0.785, + 0.127 + ], + "angle": 0, + "content": "17" + }, + { + "type": "ref_text", + "bbox": [ + 0.214, + 0.147, + 0.788, + 0.189 + ], + "angle": 0, + "content": "40. Sun, L., Sakaridis, C., Liang, J., Jiang, Q., Yang, K., Sun, P., Ye, Y., Wang, K., Gool, L.V.: Event-based fusion for motion deblurring with cross-modal attention. In: ECCV. pp. 412-428 (2022)" + }, + { + "type": "ref_text", + "bbox": [ + 0.214, + 0.19, + 0.788, + 0.231 + ], + "angle": 0, + "content": "41. Sundar, V., Ardelean, A., Swedish, T., Bruschini, C., Charbon, E., Gupta, M.: Sodacam: Software-defined cameras via single-photon imaging. In: ICCV. pp. 8165-8176 (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.214, + 0.231, + 0.788, + 0.272 + ], + "angle": 0, + "content": "42. Ta, K., Bruggemann, D., Brodermann, T., Sakaridis, C., Van Gool, L.: L2e: Lasers to events for 6-dof extrinsic calibration of lidars and event cameras. In: ICRA. pp. 11425-11431 (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.214, + 0.272, + 0.788, + 0.314 + ], + "angle": 0, + "content": "43. Taverni, G., Moeys, D.P., Li, C., Cavaco, C., Motsnyi, V., Bello, D.S.S., Delbruck, T.: Front and back illuminated dynamic and active pixel vision sensors comparison. IEEE Trans. Circuits Syst. II: Exp. Briefs 65(5), 677-681 (2018)" + }, + { + "type": "ref_text", + "bbox": [ + 0.214, + 0.314, + 0.788, + 0.342 + ], + "angle": 0, + "content": "44. Tsuji, Y., Yatagawa, T., Kubo, H., Morishima, S.: Event-based camera simulation using monte carlo path tracing with adaptive denoising. arXiv (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.214, + 0.342, + 0.788, + 0.383 + ], + "angle": 0, + "content": "45. Tulyakov, S., Bochicchio, A., Gehrig, D., Georgoulis, S., Li, Y., Scaramuzza, D.: Time lens++: Event-based frame interpolation with parametric non-linear flow and multi-scale fusion. In: CVPR. pp. 17755-17764 (2022)" + }, + { + "type": "ref_text", + "bbox": [ + 0.214, + 0.383, + 0.788, + 0.411 + ], + "angle": 0, + "content": "46. Wang, Z., Bovik, A.C., Sheikh, H.R., Simoncelli, E.P.: Image quality assessment: From error visibility to structural similarity. IEEE TIP 13(4), 600-612 (2004)" + }, + { + "type": "ref_text", + "bbox": [ + 0.214, + 0.411, + 0.788, + 0.439 + ], + "angle": 0, + "content": "47. Weng, W., Zhang, Y., Xiong, Z.: Event-based video reconstruction using transformer. In: ICCV. pp. 2563-2572 (2021)" + }, + { + "type": "ref_text", + "bbox": [ + 0.214, + 0.439, + 0.788, + 0.466 + ], + "angle": 0, + "content": "48. Yu, L., Zhang, X., Liao, W., Yang, W., Xia, G.S.: Learning to see through with events. IEEE TPAMI 45(07), 8660-8678 (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.214, + 0.466, + 0.788, + 0.494 + ], + "angle": 0, + "content": "49. Yu, W., Li, J., Zhang, S., Ji, X.: Learning scale-aware spatio-temporal implicit representation for event-based motion deblurring. In: ICML (2024)" + }, + { + "type": "ref_text", + "bbox": [ + 0.214, + 0.494, + 0.788, + 0.535 + ], + "angle": 0, + "content": "50. Zhang, H., Zhang, L., Dai, Y., Li, H., Koniusz, P.: Event-guided multi-patch network with self-supervision for non-uniform motion deblurring. IJCV 131(2), 453-470 (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.214, + 0.535, + 0.788, + 0.563 + ], + "angle": 0, + "content": "51. Zhang, R., Isola, P., Efros, A.A., Shechtman, E., Wang, O.: The unreasonable effectiveness of deep features as a perceptual metric. In: CVPR. pp. 586-595 (2018)" + }, + { + "type": "ref_text", + "bbox": [ + 0.214, + 0.564, + 0.788, + 0.591 + ], + "angle": 0, + "content": "52. Zhang, Z., Cui, S., Chai, K., Yu, H., Dasgupta, S., Mahbub, U., Rahman, T.: V2ce: Video to continuous events simulator. arXiv (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.214, + 0.591, + 0.788, + 0.632 + ], + "angle": 0, + "content": "53. Zheng, X., Liu, Y., Lu, Y., Hua, T., Pan, T., Zhang, W., Tao, D., Wang, L.: Deep learning for event-based vision: A comprehensive survey and benchmarks. arXiv (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.214, + 0.632, + 0.788, + 0.66 + ], + "angle": 0, + "content": "54. Zhu, A.Z., Wang, Z., Khant, K., Daniilidis, K.: Eventgan: Leveraging large scale image datasets for event cameras. In: ICCP. pp. 1-11 (2021)" + }, + { + "type": "ref_text", + "bbox": [ + 0.214, + 0.66, + 0.788, + 0.701 + ], + "angle": 0, + "content": "55. Zhu, L., Li, J., Wang, X., Huang, T., Tian, Y.: Neuspike-net: High speed video reconstruction via bio-inspired neuromorphic cameras. In: ICCV. pp. 2400-2409 (2021)" + }, + { + "type": "ref_text", + "bbox": [ + 0.214, + 0.701, + 0.788, + 0.742 + ], + "angle": 0, + "content": "56. Zhu, L., Wang, X., Chang, Y., Li, J., Huang, T., Tian, Y.: Event-based video reconstruction via potential-assisted spiking neural network. In: CVPR. pp. 3594-3604 (2022)" + }, + { + "type": "ref_text", + "bbox": [ + 0.214, + 0.742, + 0.788, + 0.771 + ], + "angle": 0, + "content": "57. Ziegler, A., Teigland, D., Tebbe, J., Gossard, T., Zell, A.: Real-time event simulation with frame-based cameras. In: ICRA. pp. 11669-11675 (2023)" + }, + { + "type": "list", + "bbox": [ + 0.214, + 0.147, + 0.788, + 0.771 + ], + "angle": 0, + "content": null + } + ] +] \ No newline at end of file diff --git a/2024/Physical-Based Event Camera Simulator/0986a08f-4884-431b-9c20-c64d0c63357b_origin.pdf b/2024/Physical-Based Event Camera Simulator/0986a08f-4884-431b-9c20-c64d0c63357b_origin.pdf new file mode 100644 index 0000000000000000000000000000000000000000..da311a85b305813596e85e2abb020e64a2c18fe6 --- /dev/null +++ b/2024/Physical-Based Event Camera Simulator/0986a08f-4884-431b-9c20-c64d0c63357b_origin.pdf @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:41ad290616bc36712076b4e08f3de72758f7376f31a90a2ca8f0045d0297f615 +size 1658415 diff --git a/2024/Physical-Based Event Camera Simulator/full.md b/2024/Physical-Based Event Camera Simulator/full.md new file mode 100644 index 0000000000000000000000000000000000000000..3ec547f2c88ba9a6cbf26fa207fbe59e5fe3a399 --- /dev/null +++ b/2024/Physical-Based Event Camera Simulator/full.md @@ -0,0 +1,335 @@ +# Physical-Based Event Camera Simulator + +Haiqian Han $^{1}$ , Jiacheng Lyu $^{1}$ , Jianing Li $^{1}$ $(\boxtimes)$ , Henglu Wei $^{1}$ , Cheng Li $^{2}$ , Yajing Wei $^{2}$ , Shu Chen $^{2}$ , and Xiangyang Ji $^{1}$ $(\boxtimes)$ + +$^{1}$ Tsinghua University, Haidian District, Beijing, 100084, P. R. China + +2 Beijing Xiaomi Mobile Software Co., Ltd. + +Abstract. Existing event camera simulators primarily focus on the process of generating video events and often overlook the entire optical path in real-world camera systems. To address this limitation, we propose a novel Physical-based Event Camera Simulator (PECS), which is able to generate a high-fidelity realistic event stream by directly interfacing with the 3D scene. Our PECS features a lens simulation block for accurate light-to-sensor chip replication and a multispectral rendering module for precise photocurrent generation. We present two spatiotemporal event metrics to assess the similarity between simulated and actual camera events. Experimental results demonstrate that our PECS outperforms four state-of-the-art simulators by a large margin in terms of event-based signal fidelity. We integrate our PECS into the UE platform to generate extensive multi-task synthetic datasets and evaluate its effectiveness in downstream vision tasks (e.g., video reconstruction). Our open-source code is available at https://github.com/lanpokn/PECS_trail_version. + +Keywords: Event camera simulation $\cdot$ Physics-based vision $\cdot$ Hyper-spectral data analysis + +# 1 Introduction + +Event cameras [7,43], namely bio-inspired vision sensors, operate fundamentally differently from conventional cameras. Instead of capturing intensity images at a fixed rate, event cameras respond to brightness changes with a stream of asynchronous events. With the advantages of high temporal resolution, high dynamic range, and low power consumption [33], event cameras have found widespread use in various computer vision tasks [4, 13, 17, 18, 21, 22, 25, 29, 39, 40, 49]. + +Despite notable advancements in event-based vision, the training of deep learning-based approaches still demands a substantial amount of synthetic event data [19, 53]. It is worth noting that the high cost [43] and deployment challenges [6] associated with event cameras in high-speed or low-light scenes limit the availability of real-world datasets. Thus, several event camera simulators have attempted to generate a large amount of affordable and reliable event data. + +![](images/104c9b0d509aa8877ac1f3aaa1c22ea7e83c5a158b9b6bf96fabcbfeb5c63ad9.jpg) +Fig. 1: Event camera simulators. (a) The existing video to events simulators (e.g., V2E [12]). (b) Our Physical Event Camera Simulator (PECS) is designed to generate high-fidelity realistic events by directly interfacing with the 3D scene. + +One category refers to optimized-based event camera simulators [5, 12, 14, 15, 20, 31, 38, 41, 57], where the objective is to design hand-crafted modules for generating event data. For example, V2E [12] converts intensity frames into event data via multiple hand-crafted modules. Note that, these simulators take three-channel videos as direct inputs, overlooking the complete optical path in real-world camera systems. It may lead to the loss of optical information, consequently seriously affecting the simulator's accuracy. While some efforts [14, 44] support offline rendering in the 3D scene, they often neglect the consideration of spectrum data in the rendering module. In fact, integrating spectrum information into the event camera simulation process has the potential to enhance the fidelity of synthetic event streams. + +Another category is learning-based event camera simulators [2, 8, 10, 28, 34, 52, 54] that aim to generate event representations, improving the generalization capabilities of deep learning models directly for the target domain rather than relying on raw events. For instance, EventGAN [54] is an end-to-end neural network that directly converts images to event presentations for downstream computer vision tasks. However, these learning-based simulators, requiring retraining for diverse usage cases, may have limited generalization across different scenarios. Moreover, the generated event representations are not in the original signal domain [1], posing potential limitations for broader applications. In other words, physical-based simulators, producing spatiotemporal events, preserve the inherent characteristics of raw camera data, enhancing their versatility compared to event representations by learning-based simulators. + +To address the aforementioned problems, this paper proposes a novel Physical-based Event Camera Simulator (PECS), which is able to generate a highly realistic event stream by directly interfacing with the 3D scene (see Fig. 1). In fact, the goal of this work is not to optimize hand-crafted event camera simulators (e.g., ESIM or V2E). In contrast, we aim at overcoming the following challenges: (i) How do we model a realistic lens to replicate the light-to-sensor chip process? (ii) How do we design a multispectral renderer to obtain high-precision photocurrents rather than three-channel videos? Towards this end, a realistic lens simulation block using the PBRT renderer [30] is developed to accurately replicate the light-to-sensor chip process. Then, a novel multispectral rendering module is designed to generate high-precision photocurrents by quantum efficiency and Monte Carlo integration. Two asynchronous spatiotemporal event metrics (i.e., Chamfer dis + +tance and Gaussian distance) are proposed to measure the similarity between simulated and raw events. Extensive experiments show that our PECS outperforms four state-of-the-art simulators by a large margin in terms of event-based signal fidelity. Moreover, we further deploy real-time versions into the Unreal Engine (UE) platform to generate extensive multi-task synthetic datasets. Typical computer vision tasks (i.e., event-based video reconstruction) are conducted to verify the effectiveness of our PECS. + +Overall, the main contributions of this work are summarized as follows: + +- We propose a novel Physical-based Event Camera Simulator (PECS), capable of generating a high-fidelity realistic event stream by directly interfacing with the 3D scene, without relying on video-to-event conversion. +- We design a realistic lens simulation block using the PBRT renderer, optimizing pupil sampling to improve the quality and speed of rendering. +- We design a novel multispectral rendering module that generates high-precision photocurrents through quantum efficiency and Monte Carlo integration. +- We present two asynchronous spatiotemporal metrics to assess the similarity between simulated and raw events. Extensive experiments show that our PECS outperforms four state-of-the-art simulators in terms of signal fidelity. + +To the best of our knowledge, this is the first attempt to explore a physical-based event camera simulator. We believe that our simulator will provide high-fidelity realistic large-scale event data for event-based vision tasks and offer proof of principle for the next generation of neuromorphic cameras. + +# 2 Related Work + +This section will review existing event camera simulators from two perspectives including optimized-based simulators and learning-based simulators. + +Optimized-based Event Camera Simulators. The prior optimized-based simulators include PIX2NVS [5] and ESIM [8, 31], which convert any image or video into asynchronous events. A realistic V2E [12] is developed to model the event camera in low-light or motion blur scenarios. ICNS [14] is presented to simulate noise and estimate the latency by adding the effects of the arbiter. VOLT [20] employs the concept of stochastic processes to model the output process of event data, thereby endowing the simulated event data with good continuity and probabilistic interpretability. Nevertheless, these event camera simulators directly consider three-channel videos as the input and overlook the complete optical path in real-world camera systems. Although Mou et al. [24] first notice the impact of quantum efficiency on event camera simulators and implement video-to-spectrum conversion. However, the lossless conversion from three-channel data to spectral data is nearly impossible. Thus, we design a novel multispectral rendering module that generates high-precision photocurrents through quantum efficiency and Monte Carlo integration. + +Learning-based Event Camera Simulators. With the rapid development of deep learning, some learning-based event camera simulators [2,8,10,28,34,52,54] + +![](images/2433c22525bbf331f29d2da5f5f86e09c4c54c5163cb502084ab8b57b428d321.jpg) +Fig. 2: The pipeline of the proposed Physical-based Event Camera Simulator (PECS). It starts with pupil sampling, precalculating the rays expected to pass through the lens. Ray tracing calculates light refraction and reflection within the lens and scene, producing multispectral maps. Incorporating quantum effects, these maps generate accurate photocurrents. Our PECS then leverages logarithmic differences to generate events, enabling the measurement of similarity between simulated and raw camera events in signal fidelity. + +are developed to generate event representations in the target domain. For instance, Zhu et al. [54] use the modified GAN model to generate synchronous event representations without raw events. Gu et al. [10] propose a domain-adaptive event simulator that can generate both event data and the camera's trajectory. Pantho et al. [28] use a neural network to generate event data represented by regions, which is convenient for use in moving region detection. While these simulators can enhance overall accuracy to a certain degree, the intrinsic challenge lies in the limited interpretability of neural networks. Consequently, these simulators, requiring re-training for diverse usage cases, may have limited generalization across different scenarios. In other words, simulators that rely on learning-based approaches are expected to exhibit lower robustness compared with optimized-based event camera simulators. + +# 3 Methods + +# 3.1 Framework Overview + +This work aims at designing a novel Physical-based Event Camera Simulator, termed PECS, which is able to generate a highly realistic event stream by directly interfacing with the 3D scene. As illustrated in Fig. 2, our PECS mainly consists of two modules: a multispectral rendering module and a dynamic event generation module. In the multispectral rendering module, we employ pupil sampling + +as a pre-processing step to obtain a set of rays originating from the direction in which the lens may emit. Then, ray tracing is employed to compute the refraction and reflection of light within both the lens and the scene, allowing us to acquire multispectral maps. Meanwhile, multispectral maps with quantum effects generate high-precision photocurrent. Particularly, we design a realistic lens simulation block using the PBRT renderer, which aims at optimizing pupil sampling to improve the quality and speed of rendering. In the dynamic event generation module, we leverage the principle of logarithmic differences in the event camera to generate events. This module tackles the constraints of current simulation systems by incorporating spectrum data and accounting for quantum efficiency. This enhancement contributes to improved accuracy in estimating photocurrent, thereby creating a more realistic depiction of the physical processes involved in the conversion of light signals into photocurrent. Finally, two event metrics (i.e., Chamfer distance and Gaussian distance) are proposed to measure the distance between simulated and raw events in signal fidelity. + +# 3.2 Realistic Camera Lenses Simulation + +Realistic camera lens simulation in computer graphics replicates actual camera lens behavior [16] by employing ray tracing to accurately simulate light paths through multiple lens elements. This approach takes into account the intricacies of the lens system, including refraction at interfaces, and improves image quality by estimating incident radiance along arbitrary rays. + +Our realistic camera lens simulation block models a lens system with various elements, tracing rays through them with multispectral information. It considers interactions until rays either exit the optical system or are absorbed by the aperture stop or lens housing. To reduce computational waste, our PECS performs pupil sampling before conducting the actual simulation. In this study, the pupil's definition is a little different from in optics [36]. It refers to all the directions from which rays emit from the lenses at any specific point on the sensor film, not just the center point. Let the sensor plane coordinates be denoted as $(u,v)$ . The pupil is a set of feasible directions represented by points $(x,y)$ on the direction plane. Each point $(x,y)$ in the pupil corresponds to a unique direction. Thus, the mathematical definition of the pupil can be expressed as: + +$$ +P u p i l (u, v) = \{(x, y) \mid R a y (\{u, v \}, \{x, y \})), \tag {1} +$$ + +where $Ray(\{u, v\}, \{x, y\})$ denotes the ray originating from $(u, v)$ and pointing towards $(x, y)$ passing through the lenses. + +Pupil sampling is the algorithm that identifies the correct pupil for each point on the sensor film. In the case of lenses commonly used in event cameras (e.g., Prophesee Gen4), which are often centrally symmetric with circular apertures, representing the pupil shape as an ellipse simplifies the pupil sampling process. Therefore, the pupil can be approximated as follows: + +$$ +P u p i l (u, v) = \{(x, y) \mid (x, y) \in E l l i (p, a, b) \}, \tag {2} +$$ + +where $Elli(p, a, b)$ is an ellipse centered at point $p$ on the direction plane with semi-major axis length $a$ and semi-minor axis length $b$ . + +By uniformly sampling points on the beam direction plane, the exit pupil for each sampled point on the sensor film is efficiently computed. For other points, the exit pupil is determined through linear interpolation between the ellipse's center and its major and minor axes. Despite potential errors in the elliptical model for complex lenses, this method still notably decreases the time of ray tracing within the lens, thereby speeding up the lens system simulation. + +# 3.3 Multispectral Rendering Module + +The multispectral rendering module transforms incoming light signals into photocurrent, adjusting for display convenience [12, 14, 31]. An event camera involves detecting photons at various frequencies, initiating photocurrent through the photoelectric effect. After filtering and logarithmic conversion, the currents are converted into photovoltages. The module concentrates on the photon-to-photocurrent conversion, excluding subsequent electronic circuit model steps. + +Most simulation systems estimate photocurrent using the L component of LUV values, but actual photon receivers don't convert different frequency lights into the same number of electrons, as revealed by the Quantum Efficiency (QE) curve. The definition of absolute quantum efficiency [35] can be described as: + +$$ +\mathrm {Q E} _ {\text {a b s o l u t e}} (\lambda) = \frac {N _ {e}}{N _ {v}} = \frac {R _ {\lambda}}{\lambda} \times \frac {h c}{e}, \tag {3} +$$ + +where $N_{e}$ is the number of electrons, $N_{v}$ is the number of photons, $R_{\lambda}$ is the current per unit of incident light power (A/W), $\lambda$ is the wavelength in nm, $h$ is the Planck constant, $c$ is the speed of light in vacuum, and $e$ is the elementary charge. The quantum efficiency is generally given as a relative value ranging from 0 to 1. Its definition can be formulated as follows: + +$$ +\mathrm {Q E} _ {\text {r e l a t i v e}} (\lambda) = \frac {Q E _ {\text {a b s o l u t e}} (\lambda)}{\operatorname* {m a x} (Q E _ {\text {a b s o l u t e}} (\lambda))}. \tag {4} +$$ + +Our PECS overcomes limitations by providing multi-spectral data for precise photocurrent estimation. Leveraging known spectrum information and quantum efficiency, PECS directly computes photocurrent as follows: + +$$ +i _ {p} = C \cdot \int \lambda \cdot Q E (\lambda) \cdot L (\lambda) \cdot d \lambda , \tag {5} +$$ + +where $i_p$ is the photocurrent after Photoelectric conversion, which is a key parameter in subsequent electronic circuits. C is a constant that depends on the pixel area and units of physical quantities, and $L(\lambda)$ is the input spectrum. If L is represented by a finite set of discrete channels, the following simple numerical integration method can be used to achieve the aforementioned integration: + +$$ +i _ {p} \approx C \cdot \sum_ {k = 1} ^ {N} \lambda_ {k} \cdot Q E (\lambda_ {k}) \cdot L (\lambda_ {k}) \cdot \left(\lambda_ {k + 1} - \lambda_ {k}\right), \tag {6} +$$ + +where $N$ is the number of known channels of the spectrum. If $N = 3$ , the information carried by the spectrum degrades into a common RGB image. + +The spectrum representation can be more complex in certain applications, such as using neural networks for implicit continuous representations. Simple data integration may lead to insufficient sampling in regions with larger function values, causing significant errors. To address this, Monte Carlo integration and importance sampling are commonly employed as: + +$$ +i _ {p} = \int_ {a} ^ {b} f (x) d x = \lim _ {N \rightarrow \infty} \frac {1}{N} \sum_ {i = 1} ^ {N} \frac {f \left(x _ {i}\right)}{p \left(x _ {i}\right)}, \tag {7} +$$ + +where $N$ is the number of samples. $p(x)$ is probability density function, which generated samples $x_{i}$ . Applying Eq. 7 to Eq. 5, it can be re-written as: + +$$ +i _ {p} \approx I _ {N} = C \cdot \frac {1}{N} \sum_ {k = 1} ^ {N} \frac {\lambda_ {k} \cdot Q E (\lambda_ {k}) \cdot L (\lambda_ {k})}{p (\lambda_ {k})}, \tag {8} +$$ + +where $I_N$ is an approximation. Choosing a well-behaved $p(\lambda)$ is crucial to minimize $N$ while maintaining accuracy. In probability theory, it is equal to make the variance small. Assuming that different samples are independent and identically distributed, we can obtain the variance of $I_N$ as: + +$$ +\operatorname {V a r} \left(I _ {N}\right) = \frac {1}{N ^ {2}} \sum_ {k = 1} ^ {N} \operatorname {V a r} \left(\frac {\lambda_ {k} \cdot Q E \left(\lambda_ {k}\right) \cdot L \left(\lambda_ {k}\right)}{p \left(\lambda_ {k}\right)}\right) = \frac {\sigma_ {N} ^ {2}}{N ^ {2}}, \tag {9} +$$ + +In the optimal condition, $\sigma_N^2 = 0$ , which means $p(\lambda)$ is totally equal to the original function except for scale. Thus, we can model $p(\lambda)$ as a piecewise constant function and estimate it by pre-calculating the original function as: + +$$ +p (\lambda) = B \cdot \lambda_ {j} ^ {\prime} \cdot Q E \left(\lambda_ {j} ^ {\prime}\right) \cdot L \left(\lambda_ {j} ^ {\prime}\right), j = \operatorname {a r g m i n} _ {m} | \lambda - \lambda_ {m} ^ {\prime} |, \tag {10} +$$ + +where $m$ is the channel number. For each $m$ , $L(\lambda_m')$ is pre-calculated through the implicit representation of the spectrum. B is a constant that ensures the integral of $p(\lambda)$ equals 1. Then, the photocurrent can be obtained as follows: + +$$ +i _ {p} \approx C \cdot \frac {1}{N} \sum_ {k = 1} ^ {N} \frac {\lambda_ {k} \cdot Q E (\lambda_ {k}) \cdot L (\lambda_ {k})}{B \cdot \lambda_ {j _ {k}} ^ {\prime} \cdot Q E \left(\lambda_ {j _ {k}} ^ {\prime}\right) \cdot L \left(\lambda_ {j _ {k}} ^ {\prime}\right)}, j _ {k} = \operatorname {a r g m i n} _ {m} | \lambda_ {k} - \lambda_ {m} ^ {\prime} |. \tag {11} +$$ + +Note that, Monte Carlo integration in Eq. 11 and simple numerical integration in Eq. 6 each have their own advantages and disadvantages. When the spectral data is represented continuously, Monte Carlo integration produces significantly smaller errors with a sharp spectral distribution curve compared to simple integration. However, if the spectrum is only represented by discrete channels, the lack of original information renders Monte Carlo integration somewhat redundant, and simple numerical integration can be used instead. + +Note that, our PECS estimates photocurrent by considering spectral data and quantum efficiency, offering more accurate simulations of light signal conversion. Consequently, the accuracy is significantly enhanced compared to the LUV method. Meanwhile, the quantum efficiency of different cameras often varies and can generally be provided by the corresponding sensor manufacturers. + +# 3.4 Asynchronous Spatiotemporal Event Metric + +To measure the similarity between simulated and raw camera events, we present two asynchronous spatiotemporal event metrics (i.e., Chamfer distance and Gaussian distance), which is similar to the metric part in iterative closest point tasks [3,37]. The Chamfer distance metric $C_D$ can be described as: + +$$ +C _ {D} (R, Q) = \frac {1}{| R |} \sum_ {r \in R} \min _ {q \in Q} d (r, q) + \frac {1}{| Q |} \sum_ {q \in Q} \min _ {r \in R} d (r, q), \tag {12} +$$ + +where the lower $C_D$ its value, the closer the two events are. $R$ and $Q$ are two point clouds, $r$ or $q$ is a single point in $R$ or $Q$ . Each point can be expressed as a four-dimensional tuple $\{x,y,p,t\}$ representing the spatial coordinates, polarity, and temporal information. The distance in the metric can be depicted as: + +$$ +d (r, q) = | | r - q | | _ {2}, \tag {13} +$$ + +By utilizing KDTREE [9] for acceleration, the time complexity is only $O(n \log(n))$ . Similarly, the Gaussian distance can be formulated as: + +$$ +G _ {D} (R, Q) = \frac {1}{| R |} \sum_ {r \in R} g \left(\min _ {q \in Q} d (r, q)\right) + \frac {1}{| Q |} \sum_ {q \in Q} g \left(\min _ {r \in R} d (r, q)\right), \tag {14} +$$ + +where the function $g$ makes $G_{D}(R, Q)$ still become lower when the two events get closer while reducing the impact of outliers, and it can be described as: + +$$ +g (x) = 1 - \exp \left(- \frac {| | x | | _ {2} ^ {2}}{\sigma}\right), \tag {15} +$$ + +where $\sigma$ is a hyperparameter used to adjust the sensitivity range of the original metric. In this study, we set 0.4 as the default. + +In fact, these two metrics describe the attributes of asynchronous events from different dimensions and can be necessary to assess the similarity between raw and simulated data. The Gaussian distance exhibits relative stability against outliers compared to the Chamfer distance, making it more robust to noise. Nevertheless, its measurement results have a narrower range of variation, making it less distinctive in simultaneous evaluations of multiple simulators. + +Note that, the units of x, y, p, and t in the original data are different, and the range of values for t often far exceeds that of the other components, resulting in Eq. 13 and Eq. 14 almost do not consider information beyond t. To address this + +issue, it is necessary to normalize the original measurements first to resolve this scaling problem as follows: + +$$ +L _ {\beta} [ i ] = \left(\frac {L _ {\beta} ^ {\prime} [ i ] - \operatorname* {m i n} L _ {\beta} ^ {\prime}}{\operatorname* {m a x} L _ {\beta} ^ {\prime} - \operatorname* {m i n} L _ {\beta} ^ {\prime} + \epsilon}\right) * \alpha , \tag {16} +$$ + +where $L_{\beta}^{\prime}$ is the original measurements, $\beta$ can be $x, y, p$ and $t$ element. $L_{\beta}$ is the corresponding processed data. $i$ is the i'th event of the raw data, which means $R[i] = \{L_x[i], L_y[i], L_p[i], L_t[i]\}$ . $\epsilon$ is a small number that is to avoid dividing zero, and $\alpha$ is a hyperparameter. In this work, we choose $\alpha = 100$ for space $x, y$ and polarity $p$ , and choose $\alpha = 1000$ for time $t$ . + +# 4 Experiments + +# 4.1 Experimental Setting + +Real-to-Sim Scenario Construction. Spatiotemporal synchronization is essential for quantitatively evaluating, we record the real-world event data using a Prophesee Gen4 camera in a professional optical laboratory. We capture six sequences including rotating disk and translating checkerboards with light changes and various motion speeds. For example, T_0.6_H means translating checkerboards, 0.6 meter-per-seconds, and high light. R_360_L means rotating disk, 360 rpm, and low light. To ensure consistency between the simulated and real-world scenes, we utilize the blender for high-fidelity 3D modeling. We export the scenes to PBRT and simulate a naturally calibrated camera using real lenses [26,42]. This approach enables us to acquire temporally and spatially aligned rendering results for use in subsequent sensor modules. More details of spatiotemporal alignment can be found in the supplementary material. In other methods, as regular RGB cameras struggle to capture clear images in these scenes, we use Blender's rendered RGB video as input. By manually calibrating pinhole camera parameters, we ensure temporally and spatially consistent clear images for simulation inputs, enabling the proper functioning of these simulators. Implementation Details. All experiments are conducted on a GeForce RTX 3050 Ti and 11th Gen Intel(R) Core(TM) i7-11800H CPU. We compare our PECS with four open-source event camera simulators (i.e., ESIM [31], VOLT [20], V2E [12], and ICNS [14]). In our PECS, the hyperparameters in Eq. 11 and Eq. 6 are set to $C = 100$ and $N = 31$ to make an accuracy-speed trade-off. Regarding the ESIM method, we extract its core code from its original ties to Unreal Engine (UE) and replace UE output with video output. For the V2E method, given the clarity and compactness of Blender-generated images, we disable the network interpolation module to prevent extremely slow simulation speed. In the ICNS method, all parameters have been adjusted to achieve optimal results. In the VOLT method, we select a set of parameters for the best performance. We present two event metrics (i.e., Chamfer distance and Gaussian distance) to quantitatively evaluate the similarity between raw events and synthetic data. Besides, we select three representative sequences of the HS-ERGB dataset [45] in the event-based video reconstruction task. + +Table 1: Comparison with state-of-the-art simulators on our six recording sequences. Note that, our Physical Event Camera Simulator (PECS) shows superior performance on both metrics across a diverse range of movement speeds and lighting conditions. + +
SceneGaussian distance ↓Chamfer distance ↓
ESIMVOLTV2EICNSPECSESIMVOLTV2EICNSPECS
Rotate 360 High1.9951.9231.6711.5580.96045.51222.0866.3563.8162.010
Rotate 360 Low1.9651.9311.8311.6811.63135.08225.3917.7294.1313.708
Rotate 60 High1.9951.9111.8161.5741.50439.42520.8149.2974.2373.459
Trans 1mps High1.9631.6551.8271.8061.66762.20610.57315.4509.4533.975
Trans 06mps High1.9671.7231.8391.8371.79863.93910.19818.1549.5884.169
Trans 1mps Low1.9871.8141.7001.6631.71068.11510.95214.0719.1824.467
Average1.9791.8261.7811.6871.54552.38016.66911.8436.7353.631
+ +# 4.2 Effective Test + +Quantitative Evaluation. To quantitatively evaluate the effectiveness of our PECS, we conduct a comparison with four open-source event camera simulators (i.e., ESIM [31], VOLT [20], V2E [12], and ICNS [14]) on six recording sequences (see Table 1). Note that, our PECS outperforms four state-of-the-art simulators in both two event metrics. More precisely, in terms of the Gaussian distance, our PECS reduces the average by 0.434, 0.281, 0.236, and 0.142 when compared to ESIM, VOLT, V2E, and ICNS, respectively. Meanwhile, in the Chamfer distance, our simulator exhibits an average decrease of 48.749, 13.038, 8.212, and 3.104 compared to ESIM, VOLT, V2E, and ICNS, respectively. + +Visualization Evaluation. Some representative visualization results on two motion scenarios (i.e., translating checkerboards and rotation disks) recorded in a professional optical laboratory are illustrated in Fig. 3 and Fig. 4. Obviously, our PECS achieves the best performance against four state-of-the-art simulators including ESIM, VOLT, V2E, and ICNS. We can find that V2E and ESIM simulators are too primitive, resulting in concentrated data distribution in 3D space and a lack of continuity. The event data of ESIM is entirely concentrated at the timestamp of the input frame, lacking continuity, which is the main reason for its poor performance. VOLT performs well in terms of the continuity of event output, but its noise model suffers from significant distortion. Additionally, VOLT simply converts three-channel video to grayscale as photocurrent, without considering quantum efficiency and spectrum. Although the ICNS has an advanced sensor model and supports scene input, the lack of algorithms in the section leads to less accurate event data generation. + +# 4.3 Ablation Study + +Influence of the Sampling Number $N$ . To analyze the sampling number $N$ of numerical integration in our PECS, we set the multispectral rendering module with various values of $N$ , and set $C$ as 100. As described in Table 2, we find that the Chamfer distance and Gaussian distance decrease with the increase of N. More precisely, Comparing $N = 31$ and $N = 6$ , the Chamfer distance and Gaussian distance decrease by 6.58 and 0.619, respectively. Overall, this results show that + +![](images/6b8abcfd0b45c52e32bbfc5c50e3ec8642a4f90255d05aebba948309acb0dbc8.jpg) +Fig. 3: Representative examples of translating checkerboards with various motion speed and light conditions. In each scenario, the first row is a 2D visualization schematic of the event data, and the second row is a 3D visualization schematic of the event data. + +Table 2: The influence of the sampling number $N$ of Monte-Carlo integration in the multispectral rendering module. It's worth noting that our PECS exhibits improved performance with an increase in $N$ . + +
N13467111631
Chamfer distance ↓13.968.82616.288.5888.4662.8692.4762.008
Gaussian distance ↓1.5971.6241.5691.5831.5821.1761.0910.964
+ +the larger the $N$ , the closer the final simulation data is to the real data. Of course, an increase in $N$ also entails a rise in time complexity. Thus, there is often a consideration of the trade-off between accuracy and computational speed. + +Influence of the Hyperparameter $C$ . To analyze the effect of the hyperparameter $C$ of numerical integration on the final performance, we set the multispectral rendering module with various values of $C$ , and set $N$ as 31. From Table 3, we can find that there is no simple correlation between the hyperparameter $C$ and the final performance. The distances of the two metrics gradually decrease with an increase in $C$ until they start to increase at nearly 100. Essentially speaking, the hyperparameter $C$ is introduced to align the standard unit system with the unit system used in the actual code. Theoretical analysis indicates the existence of an optimal value, a notion supported by ablation experiment results. The optimal value is observed to be around 100, and the distances of the two metrics decrease as the value of $C$ deviates from this optimal point. + +![](images/fb5a6526e6be2c8b9b9f46ccfd2a5197331884eaae76e58c664a80883829baac.jpg) +Fig. 4: Representative visualization results of rotation disks with various motion speed and light conditions. In each scenario, the first row is a 2D visualization schematic of the event data, and the second row is a 3D visualization schematic of the event data. + +Table 3: The influence of the hyperparameter $C$ of Monte-Carlo integration in the multispectral rendering module. The distances gradually decrease with an increase in $C$ until they start to increase at nearly 100. + +
C10306080100120140160
Chamfer distance ↓6.4937.4452.5085.0662.0082.0112.2642.396
Gaussian distance ↓1.3421.4881.1061.4080.9640.9861.0501.076
+ +# 5 Downstream Applications + +In this section, we integrate our PECS with the Unreal Engine (UE) platform to generate extensive event-based vision datasets that support multiple tasks. We then proceed to validate the simulator's performance on the event-based video reconstruction task. In contrast to existing simulators, the representative algorithm trained on our synthetic events performs favorably on real event data. Extensive Multi-task Synthetic Datasets. To seamlessly link our PECS with downstream computer vision tasks, we incorporate it into the widely used UE platform, a popular tool for virtual environments. Our PECS, excluding the lens simulation module, is deployed on the UE platform to achieve real-time generation of multi-modal datasets (e.g., RGB frames and events). To facilitate various visual tasks, we also develop an automatic label annotation tool on the UE platform, providing bounding boxes for object detection and scene distance values for depth estimation. As depicted in Fig. 5, we present the synthetic data of a stereo hybrid camera on a drone platform, which mainly includes + +![](images/caa842c3ed36fb0f88478f1f87c390f1a8a0ae322ae0c16a8dca2560d6f13875.jpg) +Left RGB image + +![](images/27c06e9ff24039c33754a48d733035fa078da15166e2f299225851d7c93d2e6c.jpg) +Right RGB image + +![](images/3f5501a465017a932a117e8107182a85582cf74e15adecb1f36447518daf233b.jpg) +Event image + +![](images/a9c6256610cb4901a794d39bfa4122ef72e7a0291be4003bbaa1603014d2ebbe.jpg) +Object bounding boxes + +![](images/0d56cc1591bb9afd15f027e06cecc87cb584bc438f14d6b464888f5b730e6a7e.jpg) +Depth label + +![](images/64cb9e86a0805601ed6d072b082da28d3b2114d5bb6219683643d25531590809.jpg) +Fig. 5: Representative examples of multi-task synthetic datasets. We integrate our PECS into the UE platform to generate extensive synthetic datasets supporting multiple tasks, such as video reconstruction, object detection, and depth estimation. + +![](images/7c411251c60e7c11683fafe1f4529cb5d113c1baa96db1b41e416471fe307101.jpg) + +![](images/3c197aed4f9ec1e4b3734499079ac6a6fbe4a0c1b736eb50f6b0d845310dad06.jpg) + +![](images/75f834cb943930e68af247341ba0477d90d1ae44bced09ed9dd0041f5b2e17ef.jpg) + +![](images/2929c1a2d862659b0f84a6a12b17af31a74210fe9779a19a73aec6e3551a9c6e.jpg) + +Table 4: Performance comparison on event-based video reconstruction. The E2VID network [32] trains on synthetic data from each event camera simulator and tests on three representative sequences of the real-world HS-ERGB dataset [45]. + +
SequencesPSNR↑SSIM↑LPIPS↓
ESIMV2EPECSESIMV2EPECSESIMV2EPECS
spinning_umbrella8.7568.4379.2360.4720.4710.4800.4510.4490.430
fountain_bellevue26.1395.8676.5640.2130.2110.2180.6750.6700.662
bridge_lake_0311.77012.68613.2430.5410.5600.5630.3320.2740.250
Mean8.8898.9979.6780.4090.4140.4200.4860.4640.447
+ +RGB images, event streams, object detection labels, and scene depth values. Notably, our PECS can generate comprehensive synthetic datasets using both monocular and stereo cameras, supporting a variety of vision tasks, including video reconstruction, object detection, monocular depth estimation, binocular depth estimation, and more. Besides, we also adjust the speed of motion cameras and scene light intensity, generating diverse datasets with high-speed motion and low-light conditions, showcasing the advantages of event cameras. + +Validation on Event-based Video Reconstruction. Event-based video reconstruction [32,56] is a typical task that effectively validates the effectiveness of synthetic data from various simulators. Specifically, we first use our PECS and two competitive simulators (e.g., ESIM [31] and V2E [12]) to generate large-scale synthetic datasets in the UE platform. Then, we select a representative event-to-video reconstruction algorithm (i.e., E2VID [32]) to be trained on these simulated datasets respectively. The training details align with those suggested in E2VID [32]. Finally, we test various trained models on the real-world HS-ERGB dataset [45]. We use three metrics (i.e., PSNR, SSIM [46], and LPIPS [51]) to measure the performance of event-based video reconstruction (see Table 4). Obviously, our PECS shows superior generalization capability compared to the top + +![](images/6929508043559870f592122e13a4601451cdc25859bd1d1f064583f99db2e244.jpg) +Fig. 6: Visual comparisons on event-based video reconstruction. The E2VID network [32] trains on simulated event data from each simulator and tests on the real-world HS-ERGB dataset [45] recorded by a Propheesee Gen4 camera and a RGB camera. + +competitor V2E with an average increase of 0.681 in PSNR, $0.6\%$ improvement in SSIM, and a $1.7\%$ decrease in LPIPS. Furthermore, we present some visualization results on the real-world HS-ERGB dataset [45] recorded by a Prophesee Gen4 camera and an RGB camera. As shown in Fig. 6, our PECS-trained network produces visually pleasing images with finer details and fewer artifacts. As our PECS is designed from a physical-based perspective, it naturally promotes reconstructed images with natural image statistics. Hence, the utilization of our simulator proves more effective, enabling the learning of high-quality video reconstruction [11,23,27,47,48,50,55] even with limited or no real training samples. + +# 6 Conclusion + +This paper proposes a novel Physical-based Event Camera Simulator (PECS) that can generate a highly realistic event stream by directly interfacing with the 3D scene. To the best of our knowledge, our PECS is the first event camera simulator to design a realistic lens simulation block and a multispectral rendering module. Experiments demonstrate that Our PECS consistently outperforms four state-of-the-art simulators in some scenarios with various motion speeds and light changes, showing that it is currently the most realistic event camera simulator. Besides, integrating our PECS into the Unreal Engine (UE) platform allows the generation of extensive multi-task datasets, showcasing its utility in downstream vision tasks. We also believe that our PECS provides a proof-of-principle tool for the next generation of neuromorphic cameras. + +# Acknowledgements + +This work is partially supported by the National Natural Science Foundation of China under Grant 20181310845. + +# References + +1. Annamalai, L., Ramanathan, V., Thakur, C.S.: Event-lstm: An unsupervised and asynchronous learning-based representation for event-based data. IEEE RAL 7(2), 4678-4685 (2022) +2. Baek, S., Eshraghian, J.K., Thio, W., Sandamirskaya, Y., Iu, H.H., Lu, W.D.: A real-time retinomorphic simulator using a conductance-based discrete neuronal network. In: AICAS. pp. 79-83 (2020) +3. Bellekens, B., Spruyt, V., Berkvens, R., Penne, R., Weyn, M.: A benchmark survey of rigid 3d point cloud registration algorithms. Int. J. Adv. Intell. Syst 8(5), 118-127 (2015) +4. Bhattacharya, A., Madaan, R., Cladera, F., Vermprala, S., Bonatti, R., Daniilidis, K., Kapoor, A., Kumar, V., Matni, N., Gupta, J.K.: Evidnerf: Reconstructing event data with dynamic neural radiance fields (2023) +5. Bi, Y., Andreopoulos, Y.: Pix2nvs: Parameterized conversion of pixel-domain video frames to neuromorphic vision streams. In: ICIP. pp. 1990-1994 (2017) +6. Bocca, A., Baek, D.: Automated driving systems: Key advantages, limitations and risks. In: AEIT AUTOMOTIVE. pp. 1-6 (2019) +7. Gallego, G., Delbrück, T., Orchard, G., Bartolozzi, C., Taba, B., Censi, A., Leutenegger, S., Davison, A.J., Conradt, J., Daniilidis, K., et al.: Event-based vision: A survey. IEEE TPAMI 44(1), 154-180 (2020) +8. Gehrig, D., Gehrig, M., Hidalgo-Carrió, J., Scaramuzza, D.: Video to events: Recyclcng video datasets for event cameras. In: CVPR. pp. 3586-3595 (2020) +9. Greenspan, M., Yurick, M.: Approximate kd tree search for efficient icp. In: 3DIM. pp. 442-448 (2003) +0. Gu, D., Li, J., Zhang, Y., Tian, Y.: How to learn a domain-adaptive event simulator? In: ACM MM. pp. 1275-1283 (2021) +1. He, W., You, K., Qiao, Z., Jia, X., Zhang, Z., Wang, W., Lu, H., Wang, Y., Liao, J.: Timereplayer: Unlocking the potential of event cameras for video interpolation. In: CVPR. pp. 17804-17813 (2022) +2. Hu, Y., Liu, S.C., Delbruck, T.: v2e: From video frames to realistic dvs events. In: CVPRW. pp. 1312-1321 (2021) +3. Hwang, I., Kim, J., Kim, Y.M.: Ev-nerf: Event based neural radiance field. In: WACV. pp. 837-847 (2023) +4. Joubert, D., Marcireau, A., Ralph, N., Jolley, A., van Schaik, A., Cohen, G.: Event camera simulator improvements via characterized parameters. Frontiers in Neuroscience 15, 702765 (2021) +5. Kang, Z., Li, J., Zhu, L., Tian, Y.: Retinomorphic sensing: A novel paradigm for future multimedia computing. In: ACM MM. pp. 144-152 (2021) +6. Kolb, C., Mitchell, D., Hanrahan, P.: A realistic camera model for computer graphics. In: CGIT. pp. 317-324 (1995) +7. Li, D., Li, J., Tian, Y.: Sodformer: Streaming object detection with transformer using events and frames. IEEE TPAMI 45(11), 14020-14037 (2023) + +18. Li, J., Li, J., Zhu, L., Xiang, X., Huang, T., Tian, Y.: Asynchronous spatiotemporal memory network for continuous event-based object detection. IEEE TIP 31, 2975-2987 (2022) +19. Li, W., Saeedi, S., McCormac, J., Clark, R., Tzoumanikas, D., Ye, Q., Huang, Y., Tang, R., Leutenegger, S.: Interiornet: Mega-scale multi-sensor photo-realistic indoor scenes dataset. arXiv (2018) +20. Lin, S., Ma, Y., Guo, Z., Wen, B.: Dvs-voltmeter: Stochastic process-based event simulator for dynamic vision sensors. In: ECCV. pp. 578-593 (2022) +21. Liu, P., Chen, G., Li, Z., Tang, H., Knoll, A.: Learning local event-based descriptor for patch-based stereo matching. In: ICRA. pp. 412-418 (2022) +22. Liu, X., Li, J., Fan, X., Tian, Y.: Event-based monocular dense depth estimation with recurrent transformers. In: arXiv (2022) +23. Mei, H., Wang, Z., Yang, X., Wei, X., Delbruck, T.: Deep polarization reconstruction with pdavis events. In: CVPR. pp. 22149-22158 (2023) +24. Mou, X., Feng, K., Yi, A., Wang, S., Chen, H., Hu, X., Guo, M., Chen, S., Suess, A.: Accurate event simulation using high-speed videos. Electronic Imaging 34(7), 242-1 (2022) +25. Mueggler, E., Rebecq, H., Gallego, G., Delbruck, T., Scaramuzza, D.: The event-camera dataset and simulator: Event-based data for pose estimation, visual odometry, and slam. IJRR 36(2), 142-149 (2017) +26. Muglikar, M., Gehrig, M., Gehrig, D., Scaramuzza, D.: How to calibrate your event camera. In: CVPR. pp. 1403-1409 (2021) +27. Pan, L., Hartley, R., Scheerlinck, C., Liu, M., Yu, X., Dai, Y.: High frame rate video reconstruction based on an event camera. IEEE TPAMI 44(5), 2519-2533 (2020) +28. Pantho, M.J.H., Mbongue, J.M., Bhowmik, P., Bobda, C.: Event camera simulator design for modeling attention-based inference architectures. Journal of Real-Time Image Processing 19(2), 363-374 (2022) +29. Peng, X., Wang, Y., Gao, L., Kneip, L.: Globally-optimal event camera motion estimation. In: ECCV. pp. 51-67 (2020) +30. Pharr, M., Jakob, W., Humphreys, G.: Physically based rendering: From theory to implementation (2023) +31. Rebecq, H., Gehrig, D., Scaramuzza, D.: Esim: An open event camera simulator. In: CoRL. pp. 969-982 (2018) +32. Rebecq, H., Ranftl, R., Koltun, V., Scaramuzza, D.: Events-to-video: Bringing modern computer vision to event cameras. In: CVPR. pp. 3857-3866 (2019) +33. Rebecq, H., Ranftl, R., Koltun, V., Scaramuzza, D.: High speed and high dynamic range video with an event camera. IEEE TPAMI 43(6), 1964-1980 (2019) +34. Rizzo, C., Schuman, C., Plank, J.: Event-based camera simulation wrapper for arcade learning environment. In: ICNS. pp. 1-5 (2022) +35. Rogalski, A., Razeghi, M.: Narrow-gap semiconductor photodiodes. In: Photodetectors: Materials and Devices III. vol. 3287, pp. 2-13 (1998) +36. Sasián, J.: Introduction to lens design. Cambridge University Press (2019) +37. Segal, A., Haehnel, D., Thrun, S.: Generalized-icp. In: RSS. vol. 2, p. 435. Seattle, WA (2009) +38. Sengupta, J., Liu, S., Andreou, A.: Retinosim: an event-based data synthesis tool for neuromorphic vision architecture exploration. In: ICNS. pp. 1-9 (2022) +39. Su, Q., Chou, Y., Hu, Y., Li, J., Mei, S., Zhang, Z., Li, G.: Deep directly-trained spiking neural networks for object detection. In: ICCV. pp. 6555-6565 (2023) + +40. Sun, L., Sakaridis, C., Liang, J., Jiang, Q., Yang, K., Sun, P., Ye, Y., Wang, K., Gool, L.V.: Event-based fusion for motion deblurring with cross-modal attention. In: ECCV. pp. 412-428 (2022) +41. Sundar, V., Ardelean, A., Swedish, T., Bruschini, C., Charbon, E., Gupta, M.: Sodacam: Software-defined cameras via single-photon imaging. In: ICCV. pp. 8165-8176 (2023) +42. Ta, K., Bruggemann, D., Brodermann, T., Sakaridis, C., Van Gool, L.: L2e: Lasers to events for 6-dof extrinsic calibration of lidars and event cameras. In: ICRA. pp. 11425-11431 (2023) +43. Taverni, G., Moeys, D.P., Li, C., Cavaco, C., Motsnyi, V., Bello, D.S.S., Delbruck, T.: Front and back illuminated dynamic and active pixel vision sensors comparison. IEEE Trans. Circuits Syst. II: Exp. Briefs 65(5), 677-681 (2018) +44. Tsuji, Y., Yatagawa, T., Kubo, H., Morishima, S.: Event-based camera simulation using monte carlo path tracing with adaptive denoising. arXiv (2023) +45. Tulyakov, S., Bochicchio, A., Gehrig, D., Georgoulis, S., Li, Y., Scaramuzza, D.: Time lens++: Event-based frame interpolation with parametric non-linear flow and multi-scale fusion. In: CVPR. pp. 17755-17764 (2022) +46. Wang, Z., Bovik, A.C., Sheikh, H.R., Simoncelli, E.P.: Image quality assessment: From error visibility to structural similarity. IEEE TIP 13(4), 600-612 (2004) +47. Weng, W., Zhang, Y., Xiong, Z.: Event-based video reconstruction using transformer. In: ICCV. pp. 2563-2572 (2021) +48. Yu, L., Zhang, X., Liao, W., Yang, W., Xia, G.S.: Learning to see through with events. IEEE TPAMI 45(07), 8660-8678 (2023) +49. Yu, W., Li, J., Zhang, S., Ji, X.: Learning scale-aware spatio-temporal implicit representation for event-based motion deblurring. In: ICML (2024) +50. Zhang, H., Zhang, L., Dai, Y., Li, H., Koniusz, P.: Event-guided multi-patch network with self-supervision for non-uniform motion deblurring. IJCV 131(2), 453-470 (2023) +51. Zhang, R., Isola, P., Efros, A.A., Shechtman, E., Wang, O.: The unreasonable effectiveness of deep features as a perceptual metric. In: CVPR. pp. 586-595 (2018) +52. Zhang, Z., Cui, S., Chai, K., Yu, H., Dasgupta, S., Mahbub, U., Rahman, T.: V2ce: Video to continuous events simulator. arXiv (2023) +53. Zheng, X., Liu, Y., Lu, Y., Hua, T., Pan, T., Zhang, W., Tao, D., Wang, L.: Deep learning for event-based vision: A comprehensive survey and benchmarks. arXiv (2023) +54. Zhu, A.Z., Wang, Z., Khant, K., Daniilidis, K.: Eventgan: Leveraging large scale image datasets for event cameras. In: ICCP. pp. 1-11 (2021) +55. Zhu, L., Li, J., Wang, X., Huang, T., Tian, Y.: Neuspike-net: High speed video reconstruction via bio-inspired neuromorphic cameras. In: ICCV. pp. 2400-2409 (2021) +56. Zhu, L., Wang, X., Chang, Y., Li, J., Huang, T., Tian, Y.: Event-based video reconstruction via potential-assisted spiking neural network. In: CVPR. pp. 3594-3604 (2022) +57. Ziegler, A., Teigland, D., Tebbe, J., Gossard, T., Zell, A.: Real-time event simulation with frame-based cameras. In: ICRA. pp. 11669-11675 (2023) \ No newline at end of file diff --git a/2024/Physical-Based Event Camera Simulator/images.zip b/2024/Physical-Based Event Camera Simulator/images.zip new file mode 100644 index 0000000000000000000000000000000000000000..a68b76196e9f6e73073c2caffceb70b2e6e5753f --- /dev/null +++ b/2024/Physical-Based Event Camera Simulator/images.zip @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:f4edd8854e6f04f4f55019eebb73e527e10409d0ad533a01083e0546aee83df2 +size 667257 diff --git a/2024/Physical-Based Event Camera Simulator/layout.json b/2024/Physical-Based Event Camera Simulator/layout.json new file mode 100644 index 0000000000000000000000000000000000000000..876737e0e5361b0dbcf0429cfbb7ffca02f3f16f --- /dev/null +++ b/2024/Physical-Based Event Camera Simulator/layout.json @@ -0,0 +1,9674 @@ +{ + "pdf_info": [ + { + "para_blocks": [ + { + "bbox": [ + 165, + 112, + 449, + 129 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 165, + 112, + 449, + 129 + ], + "spans": [ + { + "bbox": [ + 165, + 112, + 449, + 129 + ], + "type": "text", + "content": "Physical-Based Event Camera Simulator" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 136, + 149, + 478, + 175 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 136, + 149, + 478, + 175 + ], + "spans": [ + { + "bbox": [ + 136, + 149, + 478, + 175 + ], + "type": "text", + "content": "Haiqian Han" + }, + { + "bbox": [ + 136, + 149, + 478, + 175 + ], + "type": "inline_equation", + "content": "^{1}" + }, + { + "bbox": [ + 136, + 149, + 478, + 175 + ], + "type": "text", + "content": ", Jiacheng Lyu" + }, + { + "bbox": [ + 136, + 149, + 478, + 175 + ], + "type": "inline_equation", + "content": "^{1}" + }, + { + "bbox": [ + 136, + 149, + 478, + 175 + ], + "type": "text", + "content": ", Jianing Li" + }, + { + "bbox": [ + 136, + 149, + 478, + 175 + ], + "type": "inline_equation", + "content": "^{1}" + }, + { + "bbox": [ + 136, + 149, + 478, + 175 + ], + "type": "inline_equation", + "content": "(\\boxtimes)" + }, + { + "bbox": [ + 136, + 149, + 478, + 175 + ], + "type": "text", + "content": ", Henglu Wei" + }, + { + "bbox": [ + 136, + 149, + 478, + 175 + ], + "type": "inline_equation", + "content": "^{1}" + }, + { + "bbox": [ + 136, + 149, + 478, + 175 + ], + "type": "text", + "content": ", Cheng Li" + }, + { + "bbox": [ + 136, + 149, + 478, + 175 + ], + "type": "inline_equation", + "content": "^{2}" + }, + { + "bbox": [ + 136, + 149, + 478, + 175 + ], + "type": "text", + "content": ", Yajing Wei" + }, + { + "bbox": [ + 136, + 149, + 478, + 175 + ], + "type": "inline_equation", + "content": "^{2}" + }, + { + "bbox": [ + 136, + 149, + 478, + 175 + ], + "type": "text", + "content": ", Shu Chen" + }, + { + "bbox": [ + 136, + 149, + 478, + 175 + ], + "type": "inline_equation", + "content": "^{2}" + }, + { + "bbox": [ + 136, + 149, + 478, + 175 + ], + "type": "text", + "content": ", and Xiangyang Ji" + }, + { + "bbox": [ + 136, + 149, + 478, + 175 + ], + "type": "inline_equation", + "content": "^{1}" + }, + { + "bbox": [ + 136, + 149, + 478, + 175 + ], + "type": "inline_equation", + "content": "(\\boxtimes)" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 164, + 183, + 448, + 196 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 164, + 183, + 448, + 196 + ], + "spans": [ + { + "bbox": [ + 164, + 183, + 448, + 196 + ], + "type": "inline_equation", + "content": "^{1}" + }, + { + "bbox": [ + 164, + 183, + 448, + 196 + ], + "type": "text", + "content": " Tsinghua University, Haidian District, Beijing, 100084, P. R. China" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 217, + 196, + 395, + 206 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 217, + 196, + 395, + 206 + ], + "spans": [ + { + "bbox": [ + 217, + 196, + 395, + 206 + ], + "type": "text", + "content": "2 Beijing Xiaomi Mobile Software Co., Ltd." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 160, + 240, + 455, + 416 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 160, + 240, + 455, + 416 + ], + "spans": [ + { + "bbox": [ + 160, + 240, + 455, + 416 + ], + "type": "text", + "content": "Abstract. Existing event camera simulators primarily focus on the process of generating video events and often overlook the entire optical path in real-world camera systems. To address this limitation, we propose a novel Physical-based Event Camera Simulator (PECS), which is able to generate a high-fidelity realistic event stream by directly interfacing with the 3D scene. Our PECS features a lens simulation block for accurate light-to-sensor chip replication and a multispectral rendering module for precise photocurrent generation. We present two spatiotemporal event metrics to assess the similarity between simulated and actual camera events. Experimental results demonstrate that our PECS outperforms four state-of-the-art simulators by a large margin in terms of event-based signal fidelity. We integrate our PECS into the UE platform to generate extensive multi-task synthetic datasets and evaluate its effectiveness in downstream vision tasks (e.g., video reconstruction). Our open-source code is available at https://github.com/lanpokn/PECS_trail_version." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 159, + 427, + 453, + 450 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 159, + 427, + 453, + 450 + ], + "spans": [ + { + "bbox": [ + 159, + 427, + 453, + 450 + ], + "type": "text", + "content": "Keywords: Event camera simulation " + }, + { + "bbox": [ + 159, + 427, + 453, + 450 + ], + "type": "inline_equation", + "content": "\\cdot" + }, + { + "bbox": [ + 159, + 427, + 453, + 450 + ], + "type": "text", + "content": " Physics-based vision " + }, + { + "bbox": [ + 159, + 427, + 453, + 450 + ], + "type": "inline_equation", + "content": "\\cdot" + }, + { + "bbox": [ + 159, + 427, + 453, + 450 + ], + "type": "text", + "content": " Hyper-spectral data analysis" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 132, + 472, + 230, + 484 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 472, + 230, + 484 + ], + "spans": [ + { + "bbox": [ + 132, + 472, + 230, + 484 + ], + "type": "text", + "content": "1 Introduction" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 130, + 500, + 482, + 571 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 500, + 482, + 571 + ], + "spans": [ + { + "bbox": [ + 130, + 500, + 482, + 571 + ], + "type": "text", + "content": "Event cameras [7,43], namely bio-inspired vision sensors, operate fundamentally differently from conventional cameras. Instead of capturing intensity images at a fixed rate, event cameras respond to brightness changes with a stream of asynchronous events. With the advantages of high temporal resolution, high dynamic range, and low power consumption [33], event cameras have found widespread use in various computer vision tasks [4, 13, 17, 18, 21, 22, 25, 29, 39, 40, 49]." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 130, + 572, + 482, + 644 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 572, + 482, + 644 + ], + "spans": [ + { + "bbox": [ + 130, + 572, + 482, + 644 + ], + "type": "text", + "content": "Despite notable advancements in event-based vision, the training of deep learning-based approaches still demands a substantial amount of synthetic event data [19, 53]. It is worth noting that the high cost [43] and deployment challenges [6] associated with event cameras in high-speed or low-light scenes limit the availability of real-world datasets. Thus, several event camera simulators have attempted to generate a large amount of affordable and reliable event data." + } + ] + } + ], + "index": 8 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 141, + 653, + 376, + 665 + ], + "type": "page_footnote", + "angle": 0, + "lines": [ + { + "bbox": [ + 141, + 653, + 376, + 665 + ], + "spans": [ + { + "bbox": [ + 141, + 653, + 376, + 665 + ], + "type": "inline_equation", + "content": "(\\boxtimes)" + }, + { + "bbox": [ + 141, + 653, + 376, + 665 + ], + "type": "text", + "content": " Corresponding author: Xiangyang Ji and Jianing Li." + } + ] + } + ], + "index": 9 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 0 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 133, + 114, + 481, + 178 + ], + "blocks": [ + { + "bbox": [ + 133, + 114, + 481, + 178 + ], + "lines": [ + { + "bbox": [ + 133, + 114, + 481, + 178 + ], + "spans": [ + { + "bbox": [ + 133, + 114, + 481, + 178 + ], + "type": "image", + "image_path": "104c9b0d509aa8877ac1f3aaa1c22ea7e83c5a158b9b6bf96fabcbfeb5c63ad9.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 130, + 186, + 482, + 220 + ], + "lines": [ + { + "bbox": [ + 130, + 186, + 482, + 220 + ], + "spans": [ + { + "bbox": [ + 130, + 186, + 482, + 220 + ], + "type": "text", + "content": "Fig. 1: Event camera simulators. (a) The existing video to events simulators (e.g., V2E [12]). (b) Our Physical Event Camera Simulator (PECS) is designed to generate high-fidelity realistic events by directly interfacing with the 3D scene." + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_caption" + } + ], + "index": 2 + }, + { + "bbox": [ + 130, + 246, + 482, + 376 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 246, + 482, + 376 + ], + "spans": [ + { + "bbox": [ + 130, + 246, + 482, + 376 + ], + "type": "text", + "content": "One category refers to optimized-based event camera simulators [5, 12, 14, 15, 20, 31, 38, 41, 57], where the objective is to design hand-crafted modules for generating event data. For example, V2E [12] converts intensity frames into event data via multiple hand-crafted modules. Note that, these simulators take three-channel videos as direct inputs, overlooking the complete optical path in real-world camera systems. It may lead to the loss of optical information, consequently seriously affecting the simulator's accuracy. While some efforts [14, 44] support offline rendering in the 3D scene, they often neglect the consideration of spectrum data in the rendering module. In fact, integrating spectrum information into the event camera simulation process has the potential to enhance the fidelity of synthetic event streams." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 130, + 378, + 482, + 521 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 378, + 482, + 521 + ], + "spans": [ + { + "bbox": [ + 130, + 378, + 482, + 521 + ], + "type": "text", + "content": "Another category is learning-based event camera simulators [2, 8, 10, 28, 34, 52, 54] that aim to generate event representations, improving the generalization capabilities of deep learning models directly for the target domain rather than relying on raw events. For instance, EventGAN [54] is an end-to-end neural network that directly converts images to event presentations for downstream computer vision tasks. However, these learning-based simulators, requiring retraining for diverse usage cases, may have limited generalization across different scenarios. Moreover, the generated event representations are not in the original signal domain [1], posing potential limitations for broader applications. In other words, physical-based simulators, producing spatiotemporal events, preserve the inherent characteristics of raw camera data, enhancing their versatility compared to event representations by learning-based simulators." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 130, + 522, + 483, + 666 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 522, + 483, + 666 + ], + "spans": [ + { + "bbox": [ + 130, + 522, + 483, + 666 + ], + "type": "text", + "content": "To address the aforementioned problems, this paper proposes a novel Physical-based Event Camera Simulator (PECS), which is able to generate a highly realistic event stream by directly interfacing with the 3D scene (see Fig. 1). In fact, the goal of this work is not to optimize hand-crafted event camera simulators (e.g., ESIM or V2E). In contrast, we aim at overcoming the following challenges: (i) How do we model a realistic lens to replicate the light-to-sensor chip process? (ii) How do we design a multispectral renderer to obtain high-precision photocurrents rather than three-channel videos? Towards this end, a realistic lens simulation block using the PBRT renderer [30] is developed to accurately replicate the light-to-sensor chip process. Then, a novel multispectral rendering module is designed to generate high-precision photocurrents by quantum efficiency and Monte Carlo integration. Two asynchronous spatiotemporal event metrics (i.e., Chamfer dis" + } + ] + } + ], + "index": 6 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "spans": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "type": "text", + "content": "2" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 166, + 91, + 257, + 102 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 166, + 91, + 257, + 102 + ], + "spans": [ + { + "bbox": [ + 166, + 91, + 257, + 102 + ], + "type": "text", + "content": "Haiqian. Author et al." + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 1 + }, + { + "para_blocks": [ + { + "bbox": [ + 130, + 116, + 479, + 199 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 116, + 479, + 199 + ], + "spans": [ + { + "bbox": [ + 130, + 116, + 479, + 199 + ], + "type": "text", + "content": "tance and Gaussian distance) are proposed to measure the similarity between simulated and raw events. Extensive experiments show that our PECS outperforms four state-of-the-art simulators by a large margin in terms of event-based signal fidelity. Moreover, we further deploy real-time versions into the Unreal Engine (UE) platform to generate extensive multi-task synthetic datasets. Typical computer vision tasks (i.e., event-based video reconstruction) are conducted to verify the effectiveness of our PECS." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 146, + 200, + 460, + 211 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 146, + 200, + 460, + 211 + ], + "spans": [ + { + "bbox": [ + 146, + 200, + 460, + 211 + ], + "type": "text", + "content": "Overall, the main contributions of this work are summarized as follows:" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 138, + 220, + 485, + 338 + ], + "type": "list", + "angle": 0, + "index": 8, + "blocks": [ + { + "bbox": [ + 138, + 220, + 481, + 255 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 220, + 481, + 255 + ], + "spans": [ + { + "bbox": [ + 138, + 220, + 481, + 255 + ], + "type": "text", + "content": "- We propose a novel Physical-based Event Camera Simulator (PECS), capable of generating a high-fidelity realistic event stream by directly interfacing with the 3D scene, without relying on video-to-event conversion." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 138, + 256, + 481, + 279 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 256, + 481, + 279 + ], + "spans": [ + { + "bbox": [ + 138, + 256, + 481, + 279 + ], + "type": "text", + "content": "- We design a realistic lens simulation block using the PBRT renderer, optimizing pupil sampling to improve the quality and speed of rendering." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 138, + 280, + 485, + 303 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 280, + 485, + 303 + ], + "spans": [ + { + "bbox": [ + 138, + 280, + 485, + 303 + ], + "type": "text", + "content": "- We design a novel multispectral rendering module that generates high-precision photocurrents through quantum efficiency and Monte Carlo integration." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 138, + 304, + 481, + 338 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 304, + 481, + 338 + ], + "spans": [ + { + "bbox": [ + 138, + 304, + 481, + 338 + ], + "type": "text", + "content": "- We present two asynchronous spatiotemporal metrics to assess the similarity between simulated and raw events. Extensive experiments show that our PECS outperforms four state-of-the-art simulators in terms of signal fidelity." + } + ] + } + ], + "index": 7 + } + ], + "sub_type": "text" + }, + { + "bbox": [ + 130, + 348, + 480, + 396 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 348, + 480, + 396 + ], + "spans": [ + { + "bbox": [ + 130, + 348, + 480, + 396 + ], + "type": "text", + "content": "To the best of our knowledge, this is the first attempt to explore a physical-based event camera simulator. We believe that our simulator will provide high-fidelity realistic large-scale event data for event-based vision tasks and offer proof of principle for the next generation of neuromorphic cameras." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 131, + 414, + 237, + 426 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 131, + 414, + 237, + 426 + ], + "spans": [ + { + "bbox": [ + 131, + 414, + 237, + 426 + ], + "type": "text", + "content": "2 Related Work" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 130, + 438, + 479, + 461 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 438, + 479, + 461 + ], + "spans": [ + { + "bbox": [ + 130, + 438, + 479, + 461 + ], + "type": "text", + "content": "This section will review existing event camera simulators from two perspectives including optimized-based simulators and learning-based simulators." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 130, + 462, + 480, + 640 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 462, + 480, + 640 + ], + "spans": [ + { + "bbox": [ + 130, + 462, + 480, + 640 + ], + "type": "text", + "content": "Optimized-based Event Camera Simulators. The prior optimized-based simulators include PIX2NVS [5] and ESIM [8, 31], which convert any image or video into asynchronous events. A realistic V2E [12] is developed to model the event camera in low-light or motion blur scenarios. ICNS [14] is presented to simulate noise and estimate the latency by adding the effects of the arbiter. VOLT [20] employs the concept of stochastic processes to model the output process of event data, thereby endowing the simulated event data with good continuity and probabilistic interpretability. Nevertheless, these event camera simulators directly consider three-channel videos as the input and overlook the complete optical path in real-world camera systems. Although Mou et al. [24] first notice the impact of quantum efficiency on event camera simulators and implement video-to-spectrum conversion. However, the lossless conversion from three-channel data to spectral data is nearly impossible. Thus, we design a novel multispectral rendering module that generates high-precision photocurrents through quantum efficiency and Monte Carlo integration." + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 130, + 641, + 481, + 665 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 641, + 481, + 665 + ], + "spans": [ + { + "bbox": [ + 130, + 641, + 481, + 665 + ], + "type": "text", + "content": "Learning-based Event Camera Simulators. With the rapid development of deep learning, some learning-based event camera simulators [2,8,10,28,34,52,54]" + } + ] + } + ], + "index": 13 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 282, + 91, + 447, + 102 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 282, + 91, + 447, + 102 + ], + "spans": [ + { + "bbox": [ + 282, + 91, + 447, + 102 + ], + "type": "text", + "content": "Physical-Based Event Camera Simulator" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 474, + 92, + 480, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 474, + 92, + 480, + 100 + ], + "spans": [ + { + "bbox": [ + 474, + 92, + 480, + 100 + ], + "type": "text", + "content": "3" + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 2 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 136, + 119, + 473, + 284 + ], + "blocks": [ + { + "bbox": [ + 136, + 119, + 473, + 284 + ], + "lines": [ + { + "bbox": [ + 136, + 119, + 473, + 284 + ], + "spans": [ + { + "bbox": [ + 136, + 119, + 473, + 284 + ], + "type": "image", + "image_path": "2433c22525bbf331f29d2da5f5f86e09c4c54c5163cb502084ab8b57b428d321.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 130, + 295, + 482, + 373 + ], + "lines": [ + { + "bbox": [ + 130, + 295, + 482, + 373 + ], + "spans": [ + { + "bbox": [ + 130, + 295, + 482, + 373 + ], + "type": "text", + "content": "Fig. 2: The pipeline of the proposed Physical-based Event Camera Simulator (PECS). It starts with pupil sampling, precalculating the rays expected to pass through the lens. Ray tracing calculates light refraction and reflection within the lens and scene, producing multispectral maps. Incorporating quantum effects, these maps generate accurate photocurrents. Our PECS then leverages logarithmic differences to generate events, enabling the measurement of similarity between simulated and raw camera events in signal fidelity." + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_caption" + } + ], + "index": 2 + }, + { + "bbox": [ + 130, + 397, + 482, + 540 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 397, + 482, + 540 + ], + "spans": [ + { + "bbox": [ + 130, + 397, + 482, + 540 + ], + "type": "text", + "content": "are developed to generate event representations in the target domain. For instance, Zhu et al. [54] use the modified GAN model to generate synchronous event representations without raw events. Gu et al. [10] propose a domain-adaptive event simulator that can generate both event data and the camera's trajectory. Pantho et al. [28] use a neural network to generate event data represented by regions, which is convenient for use in moving region detection. While these simulators can enhance overall accuracy to a certain degree, the intrinsic challenge lies in the limited interpretability of neural networks. Consequently, these simulators, requiring re-training for diverse usage cases, may have limited generalization across different scenarios. In other words, simulators that rely on learning-based approaches are expected to exhibit lower robustness compared with optimized-based event camera simulators." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 132, + 559, + 207, + 572 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 559, + 207, + 572 + ], + "spans": [ + { + "bbox": [ + 132, + 559, + 207, + 572 + ], + "type": "text", + "content": "3 Methods" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 132, + 585, + 268, + 597 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 585, + 268, + 597 + ], + "spans": [ + { + "bbox": [ + 132, + 585, + 268, + 597 + ], + "type": "text", + "content": "3.1 Framework Overview" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 130, + 605, + 482, + 666 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 605, + 482, + 666 + ], + "spans": [ + { + "bbox": [ + 130, + 605, + 482, + 666 + ], + "type": "text", + "content": "This work aims at designing a novel Physical-based Event Camera Simulator, termed PECS, which is able to generate a highly realistic event stream by directly interfacing with the 3D scene. As illustrated in Fig. 2, our PECS mainly consists of two modules: a multispectral rendering module and a dynamic event generation module. In the multispectral rendering module, we employ pupil sampling" + } + ] + } + ], + "index": 7 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "spans": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "type": "text", + "content": "4" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 166, + 91, + 257, + 102 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 166, + 91, + 257, + 102 + ], + "spans": [ + { + "bbox": [ + 166, + 91, + 257, + 102 + ], + "type": "text", + "content": "Haiqian. Author et al." + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 3 + }, + { + "para_blocks": [ + { + "bbox": [ + 130, + 116, + 482, + 297 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 116, + 482, + 297 + ], + "spans": [ + { + "bbox": [ + 130, + 116, + 482, + 297 + ], + "type": "text", + "content": "as a pre-processing step to obtain a set of rays originating from the direction in which the lens may emit. Then, ray tracing is employed to compute the refraction and reflection of light within both the lens and the scene, allowing us to acquire multispectral maps. Meanwhile, multispectral maps with quantum effects generate high-precision photocurrent. Particularly, we design a realistic lens simulation block using the PBRT renderer, which aims at optimizing pupil sampling to improve the quality and speed of rendering. In the dynamic event generation module, we leverage the principle of logarithmic differences in the event camera to generate events. This module tackles the constraints of current simulation systems by incorporating spectrum data and accounting for quantum efficiency. This enhancement contributes to improved accuracy in estimating photocurrent, thereby creating a more realistic depiction of the physical processes involved in the conversion of light signals into photocurrent. Finally, two event metrics (i.e., Chamfer distance and Gaussian distance) are proposed to measure the distance between simulated and raw events in signal fidelity." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 132, + 314, + 340, + 324 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 314, + 340, + 324 + ], + "spans": [ + { + "bbox": [ + 132, + 314, + 340, + 324 + ], + "type": "text", + "content": "3.2 Realistic Camera Lenses Simulation" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 130, + 334, + 480, + 393 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 334, + 480, + 393 + ], + "spans": [ + { + "bbox": [ + 130, + 334, + 480, + 393 + ], + "type": "text", + "content": "Realistic camera lens simulation in computer graphics replicates actual camera lens behavior [16] by employing ray tracing to accurately simulate light paths through multiple lens elements. This approach takes into account the intricacies of the lens system, including refraction at interfaces, and improves image quality by estimating incident radiance along arbitrary rays." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 130, + 394, + 482, + 525 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 394, + 482, + 525 + ], + "spans": [ + { + "bbox": [ + 130, + 394, + 482, + 525 + ], + "type": "text", + "content": "Our realistic camera lens simulation block models a lens system with various elements, tracing rays through them with multispectral information. It considers interactions until rays either exit the optical system or are absorbed by the aperture stop or lens housing. To reduce computational waste, our PECS performs pupil sampling before conducting the actual simulation. In this study, the pupil's definition is a little different from in optics [36]. It refers to all the directions from which rays emit from the lenses at any specific point on the sensor film, not just the center point. Let the sensor plane coordinates be denoted as " + }, + { + "bbox": [ + 130, + 394, + 482, + 525 + ], + "type": "inline_equation", + "content": "(u,v)" + }, + { + "bbox": [ + 130, + 394, + 482, + 525 + ], + "type": "text", + "content": ". The pupil is a set of feasible directions represented by points " + }, + { + "bbox": [ + 130, + 394, + 482, + 525 + ], + "type": "inline_equation", + "content": "(x,y)" + }, + { + "bbox": [ + 130, + 394, + 482, + 525 + ], + "type": "text", + "content": " on the direction plane. Each point " + }, + { + "bbox": [ + 130, + 394, + 482, + 525 + ], + "type": "inline_equation", + "content": "(x,y)" + }, + { + "bbox": [ + 130, + 394, + 482, + 525 + ], + "type": "text", + "content": " in the pupil corresponds to a unique direction. Thus, the mathematical definition of the pupil can be expressed as:" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 214, + 536, + 480, + 548 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 214, + 536, + 480, + 548 + ], + "spans": [ + { + "bbox": [ + 214, + 536, + 480, + 548 + ], + "type": "interline_equation", + "content": "P u p i l (u, v) = \\{(x, y) \\mid R a y (\\{u, v \\}, \\{x, y \\})), \\tag {1}", + "image_path": "e6d9e3527609c0831ada509462ef687c971c8b9b53a0e07dae6e50d05a15f47b.jpg" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 130, + 557, + 480, + 581 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 557, + 480, + 581 + ], + "spans": [ + { + "bbox": [ + 130, + 557, + 480, + 581 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 130, + 557, + 480, + 581 + ], + "type": "inline_equation", + "content": "Ray(\\{u, v\\}, \\{x, y\\})" + }, + { + "bbox": [ + 130, + 557, + 480, + 581 + ], + "type": "text", + "content": " denotes the ray originating from " + }, + { + "bbox": [ + 130, + 557, + 480, + 581 + ], + "type": "inline_equation", + "content": "(u, v)" + }, + { + "bbox": [ + 130, + 557, + 480, + 581 + ], + "type": "text", + "content": " and pointing towards " + }, + { + "bbox": [ + 130, + 557, + 480, + 581 + ], + "type": "inline_equation", + "content": "(x, y)" + }, + { + "bbox": [ + 130, + 557, + 480, + 581 + ], + "type": "text", + "content": " passing through the lenses." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 130, + 582, + 480, + 641 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 582, + 480, + 641 + ], + "spans": [ + { + "bbox": [ + 130, + 582, + 480, + 641 + ], + "type": "text", + "content": "Pupil sampling is the algorithm that identifies the correct pupil for each point on the sensor film. In the case of lenses commonly used in event cameras (e.g., Prophesee Gen4), which are often centrally symmetric with circular apertures, representing the pupil shape as an ellipse simplifies the pupil sampling process. Therefore, the pupil can be approximated as follows:" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 212, + 654, + 480, + 666 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 212, + 654, + 480, + 666 + ], + "spans": [ + { + "bbox": [ + 212, + 654, + 480, + 666 + ], + "type": "interline_equation", + "content": "P u p i l (u, v) = \\{(x, y) \\mid (x, y) \\in E l l i (p, a, b) \\}, \\tag {2}", + "image_path": "e7cb202c84edb90db222cc2a4a9ae75425003b96a62835ebc5cc4f178931c1bc.jpg" + } + ] + } + ], + "index": 9 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 282, + 91, + 447, + 102 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 282, + 91, + 447, + 102 + ], + "spans": [ + { + "bbox": [ + 282, + 91, + 447, + 102 + ], + "type": "text", + "content": "Physical-Based Event Camera Simulator" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 474, + 92, + 480, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 474, + 92, + 480, + 100 + ], + "spans": [ + { + "bbox": [ + 474, + 92, + 480, + 100 + ], + "type": "text", + "content": "5" + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 4 + }, + { + "para_blocks": [ + { + "bbox": [ + 130, + 116, + 479, + 140 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 116, + 479, + 140 + ], + "spans": [ + { + "bbox": [ + 130, + 116, + 479, + 140 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 130, + 116, + 479, + 140 + ], + "type": "inline_equation", + "content": "Elli(p, a, b)" + }, + { + "bbox": [ + 130, + 116, + 479, + 140 + ], + "type": "text", + "content": " is an ellipse centered at point " + }, + { + "bbox": [ + 130, + 116, + 479, + 140 + ], + "type": "inline_equation", + "content": "p" + }, + { + "bbox": [ + 130, + 116, + 479, + 140 + ], + "type": "text", + "content": " on the direction plane with semi-major axis length " + }, + { + "bbox": [ + 130, + 116, + 479, + 140 + ], + "type": "inline_equation", + "content": "a" + }, + { + "bbox": [ + 130, + 116, + 479, + 140 + ], + "type": "text", + "content": " and semi-minor axis length " + }, + { + "bbox": [ + 130, + 116, + 479, + 140 + ], + "type": "inline_equation", + "content": "b" + }, + { + "bbox": [ + 130, + 116, + 479, + 140 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 130, + 140, + 480, + 212 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 140, + 480, + 212 + ], + "spans": [ + { + "bbox": [ + 130, + 140, + 480, + 212 + ], + "type": "text", + "content": "By uniformly sampling points on the beam direction plane, the exit pupil for each sampled point on the sensor film is efficiently computed. For other points, the exit pupil is determined through linear interpolation between the ellipse's center and its major and minor axes. Despite potential errors in the elliptical model for complex lenses, this method still notably decreases the time of ray tracing within the lens, thereby speeding up the lens system simulation." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 131, + 228, + 326, + 240 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 131, + 228, + 326, + 240 + ], + "spans": [ + { + "bbox": [ + 131, + 228, + 326, + 240 + ], + "type": "text", + "content": "3.3 Multispectral Rendering Module" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 130, + 246, + 479, + 317 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 246, + 479, + 317 + ], + "spans": [ + { + "bbox": [ + 130, + 246, + 479, + 317 + ], + "type": "text", + "content": "The multispectral rendering module transforms incoming light signals into photocurrent, adjusting for display convenience [12, 14, 31]. An event camera involves detecting photons at various frequencies, initiating photocurrent through the photoelectric effect. After filtering and logarithmic conversion, the currents are converted into photovoltages. The module concentrates on the photon-to-photocurrent conversion, excluding subsequent electronic circuit model steps." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 130, + 318, + 480, + 365 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 318, + 480, + 365 + ], + "spans": [ + { + "bbox": [ + 130, + 318, + 480, + 365 + ], + "type": "text", + "content": "Most simulation systems estimate photocurrent using the L component of LUV values, but actual photon receivers don't convert different frequency lights into the same number of electrons, as revealed by the Quantum Efficiency (QE) curve. The definition of absolute quantum efficiency [35] can be described as:" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 236, + 373, + 480, + 397 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 236, + 373, + 480, + 397 + ], + "spans": [ + { + "bbox": [ + 236, + 373, + 480, + 397 + ], + "type": "interline_equation", + "content": "\\mathrm {Q E} _ {\\text {a b s o l u t e}} (\\lambda) = \\frac {N _ {e}}{N _ {v}} = \\frac {R _ {\\lambda}}{\\lambda} \\times \\frac {h c}{e}, \\tag {3}", + "image_path": "b49af404faea58e05dc2a697f55de83f4cfc1a7f80eed77d9459acc31a047938.jpg" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 130, + 402, + 480, + 460 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 402, + 480, + 460 + ], + "spans": [ + { + "bbox": [ + 130, + 402, + 480, + 460 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 130, + 402, + 480, + 460 + ], + "type": "inline_equation", + "content": "N_{e}" + }, + { + "bbox": [ + 130, + 402, + 480, + 460 + ], + "type": "text", + "content": " is the number of electrons, " + }, + { + "bbox": [ + 130, + 402, + 480, + 460 + ], + "type": "inline_equation", + "content": "N_{v}" + }, + { + "bbox": [ + 130, + 402, + 480, + 460 + ], + "type": "text", + "content": " is the number of photons, " + }, + { + "bbox": [ + 130, + 402, + 480, + 460 + ], + "type": "inline_equation", + "content": "R_{\\lambda}" + }, + { + "bbox": [ + 130, + 402, + 480, + 460 + ], + "type": "text", + "content": " is the current per unit of incident light power (A/W), " + }, + { + "bbox": [ + 130, + 402, + 480, + 460 + ], + "type": "inline_equation", + "content": "\\lambda" + }, + { + "bbox": [ + 130, + 402, + 480, + 460 + ], + "type": "text", + "content": " is the wavelength in nm, " + }, + { + "bbox": [ + 130, + 402, + 480, + 460 + ], + "type": "inline_equation", + "content": "h" + }, + { + "bbox": [ + 130, + 402, + 480, + 460 + ], + "type": "text", + "content": " is the Planck constant, " + }, + { + "bbox": [ + 130, + 402, + 480, + 460 + ], + "type": "inline_equation", + "content": "c" + }, + { + "bbox": [ + 130, + 402, + 480, + 460 + ], + "type": "text", + "content": " is the speed of light in vacuum, and " + }, + { + "bbox": [ + 130, + 402, + 480, + 460 + ], + "type": "inline_equation", + "content": "e" + }, + { + "bbox": [ + 130, + 402, + 480, + 460 + ], + "type": "text", + "content": " is the elementary charge. The quantum efficiency is generally given as a relative value ranging from 0 to 1. Its definition can be formulated as follows:" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 227, + 466, + 480, + 492 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 227, + 466, + 480, + 492 + ], + "spans": [ + { + "bbox": [ + 227, + 466, + 480, + 492 + ], + "type": "interline_equation", + "content": "\\mathrm {Q E} _ {\\text {r e l a t i v e}} (\\lambda) = \\frac {Q E _ {\\text {a b s o l u t e}} (\\lambda)}{\\operatorname* {m a x} (Q E _ {\\text {a b s o l u t e}} (\\lambda))}. \\tag {4}", + "image_path": "34e9c5082d32db5481f852548835765a38e882a175d4902ade61e621bfbbe7a1.jpg" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 130, + 497, + 479, + 533 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 497, + 479, + 533 + ], + "spans": [ + { + "bbox": [ + 130, + 497, + 479, + 533 + ], + "type": "text", + "content": "Our PECS overcomes limitations by providing multi-spectral data for precise photocurrent estimation. Leveraging known spectrum information and quantum efficiency, PECS directly computes photocurrent as follows:" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 238, + 540, + 480, + 564 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 238, + 540, + 480, + 564 + ], + "spans": [ + { + "bbox": [ + 238, + 540, + 480, + 564 + ], + "type": "interline_equation", + "content": "i _ {p} = C \\cdot \\int \\lambda \\cdot Q E (\\lambda) \\cdot L (\\lambda) \\cdot d \\lambda , \\tag {5}", + "image_path": "01bc09eb6edf007449351d5191c6ad5b1544162fb15edc0b4601354d63c6bb59.jpg" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 130, + 569, + 479, + 628 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 569, + 479, + 628 + ], + "spans": [ + { + "bbox": [ + 130, + 569, + 479, + 628 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 130, + 569, + 479, + 628 + ], + "type": "inline_equation", + "content": "i_p" + }, + { + "bbox": [ + 130, + 569, + 479, + 628 + ], + "type": "text", + "content": " is the photocurrent after Photoelectric conversion, which is a key parameter in subsequent electronic circuits. C is a constant that depends on the pixel area and units of physical quantities, and " + }, + { + "bbox": [ + 130, + 569, + 479, + 628 + ], + "type": "inline_equation", + "content": "L(\\lambda)" + }, + { + "bbox": [ + 130, + 569, + 479, + 628 + ], + "type": "text", + "content": " is the input spectrum. If L is represented by a finite set of discrete channels, the following simple numerical integration method can be used to achieve the aforementioned integration:" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 208, + 636, + 480, + 667 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 208, + 636, + 480, + 667 + ], + "spans": [ + { + "bbox": [ + 208, + 636, + 480, + 667 + ], + "type": "interline_equation", + "content": "i _ {p} \\approx C \\cdot \\sum_ {k = 1} ^ {N} \\lambda_ {k} \\cdot Q E (\\lambda_ {k}) \\cdot L (\\lambda_ {k}) \\cdot \\left(\\lambda_ {k + 1} - \\lambda_ {k}\\right), \\tag {6}", + "image_path": "bad94852880aba3e578e90a89b620bc8d7346ffd81197025340239631d65460a.jpg" + } + ] + } + ], + "index": 13 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "spans": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "type": "text", + "content": "6" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 166, + 91, + 257, + 102 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 166, + 91, + 257, + 102 + ], + "spans": [ + { + "bbox": [ + 166, + 91, + 257, + 102 + ], + "type": "text", + "content": "Haiqian. Author et al." + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 5 + }, + { + "para_blocks": [ + { + "bbox": [ + 130, + 116, + 479, + 140 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 116, + 479, + 140 + ], + "spans": [ + { + "bbox": [ + 130, + 116, + 479, + 140 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 130, + 116, + 479, + 140 + ], + "type": "inline_equation", + "content": "N" + }, + { + "bbox": [ + 130, + 116, + 479, + 140 + ], + "type": "text", + "content": " is the number of known channels of the spectrum. If " + }, + { + "bbox": [ + 130, + 116, + 479, + 140 + ], + "type": "inline_equation", + "content": "N = 3" + }, + { + "bbox": [ + 130, + 116, + 479, + 140 + ], + "type": "text", + "content": ", the information carried by the spectrum degrades into a common RGB image." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 130, + 140, + 480, + 200 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 140, + 480, + 200 + ], + "spans": [ + { + "bbox": [ + 130, + 140, + 480, + 200 + ], + "type": "text", + "content": "The spectrum representation can be more complex in certain applications, such as using neural networks for implicit continuous representations. Simple data integration may lead to insufficient sampling in regions with larger function values, causing significant errors. To address this, Monte Carlo integration and importance sampling are commonly employed as:" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 225, + 210, + 481, + 242 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 225, + 210, + 481, + 242 + ], + "spans": [ + { + "bbox": [ + 225, + 210, + 481, + 242 + ], + "type": "interline_equation", + "content": "i _ {p} = \\int_ {a} ^ {b} f (x) d x = \\lim _ {N \\rightarrow \\infty} \\frac {1}{N} \\sum_ {i = 1} ^ {N} \\frac {f \\left(x _ {i}\\right)}{p \\left(x _ {i}\\right)}, \\tag {7}", + "image_path": "deaf9dab8d2ed5c21a11ef585983a25b369d550bd6ac5c38060f0c6bc1bcab27.jpg" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 131, + 251, + 479, + 275 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 131, + 251, + 479, + 275 + ], + "spans": [ + { + "bbox": [ + 131, + 251, + 479, + 275 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 131, + 251, + 479, + 275 + ], + "type": "inline_equation", + "content": "N" + }, + { + "bbox": [ + 131, + 251, + 479, + 275 + ], + "type": "text", + "content": " is the number of samples. " + }, + { + "bbox": [ + 131, + 251, + 479, + 275 + ], + "type": "inline_equation", + "content": "p(x)" + }, + { + "bbox": [ + 131, + 251, + 479, + 275 + ], + "type": "text", + "content": " is probability density function, which generated samples " + }, + { + "bbox": [ + 131, + 251, + 479, + 275 + ], + "type": "inline_equation", + "content": "x_{i}" + }, + { + "bbox": [ + 131, + 251, + 479, + 275 + ], + "type": "text", + "content": ". Applying Eq. 7 to Eq. 5, it can be re-written as:" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 216, + 285, + 481, + 317 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 216, + 285, + 481, + 317 + ], + "spans": [ + { + "bbox": [ + 216, + 285, + 481, + 317 + ], + "type": "interline_equation", + "content": "i _ {p} \\approx I _ {N} = C \\cdot \\frac {1}{N} \\sum_ {k = 1} ^ {N} \\frac {\\lambda_ {k} \\cdot Q E (\\lambda_ {k}) \\cdot L (\\lambda_ {k})}{p (\\lambda_ {k})}, \\tag {8}", + "image_path": "7ea65122ad3f95ec38157bbd19c02c59201742513bfe5c94db3a9f874f8cf972.jpg" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 131, + 326, + 479, + 374 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 131, + 326, + 479, + 374 + ], + "spans": [ + { + "bbox": [ + 131, + 326, + 479, + 374 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 131, + 326, + 479, + 374 + ], + "type": "inline_equation", + "content": "I_N" + }, + { + "bbox": [ + 131, + 326, + 479, + 374 + ], + "type": "text", + "content": " is an approximation. Choosing a well-behaved " + }, + { + "bbox": [ + 131, + 326, + 479, + 374 + ], + "type": "inline_equation", + "content": "p(\\lambda)" + }, + { + "bbox": [ + 131, + 326, + 479, + 374 + ], + "type": "text", + "content": " is crucial to minimize " + }, + { + "bbox": [ + 131, + 326, + 479, + 374 + ], + "type": "inline_equation", + "content": "N" + }, + { + "bbox": [ + 131, + 326, + 479, + 374 + ], + "type": "text", + "content": " while maintaining accuracy. In probability theory, it is equal to make the variance small. Assuming that different samples are independent and identically distributed, we can obtain the variance of " + }, + { + "bbox": [ + 131, + 326, + 479, + 374 + ], + "type": "inline_equation", + "content": "I_N" + }, + { + "bbox": [ + 131, + 326, + 479, + 374 + ], + "type": "text", + "content": " as:" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 189, + 384, + 481, + 416 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 189, + 384, + 481, + 416 + ], + "spans": [ + { + "bbox": [ + 189, + 384, + 481, + 416 + ], + "type": "interline_equation", + "content": "\\operatorname {V a r} \\left(I _ {N}\\right) = \\frac {1}{N ^ {2}} \\sum_ {k = 1} ^ {N} \\operatorname {V a r} \\left(\\frac {\\lambda_ {k} \\cdot Q E \\left(\\lambda_ {k}\\right) \\cdot L \\left(\\lambda_ {k}\\right)}{p \\left(\\lambda_ {k}\\right)}\\right) = \\frac {\\sigma_ {N} ^ {2}}{N ^ {2}}, \\tag {9}", + "image_path": "9f6dd1bdb47cab6059f15e615ab6f46b24e22f09c0e07d07706fe1cf865416b3.jpg" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 131, + 426, + 479, + 462 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 131, + 426, + 479, + 462 + ], + "spans": [ + { + "bbox": [ + 131, + 426, + 479, + 462 + ], + "type": "text", + "content": "In the optimal condition, " + }, + { + "bbox": [ + 131, + 426, + 479, + 462 + ], + "type": "inline_equation", + "content": "\\sigma_N^2 = 0" + }, + { + "bbox": [ + 131, + 426, + 479, + 462 + ], + "type": "text", + "content": ", which means " + }, + { + "bbox": [ + 131, + 426, + 479, + 462 + ], + "type": "inline_equation", + "content": "p(\\lambda)" + }, + { + "bbox": [ + 131, + 426, + 479, + 462 + ], + "type": "text", + "content": " is totally equal to the original function except for scale. Thus, we can model " + }, + { + "bbox": [ + 131, + 426, + 479, + 462 + ], + "type": "inline_equation", + "content": "p(\\lambda)" + }, + { + "bbox": [ + 131, + 426, + 479, + 462 + ], + "type": "text", + "content": " as a piecewise constant function and estimate it by pre-calculating the original function as:" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 189, + 472, + 481, + 487 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 189, + 472, + 481, + 487 + ], + "spans": [ + { + "bbox": [ + 189, + 472, + 481, + 487 + ], + "type": "interline_equation", + "content": "p (\\lambda) = B \\cdot \\lambda_ {j} ^ {\\prime} \\cdot Q E \\left(\\lambda_ {j} ^ {\\prime}\\right) \\cdot L \\left(\\lambda_ {j} ^ {\\prime}\\right), j = \\operatorname {a r g m i n} _ {m} | \\lambda - \\lambda_ {m} ^ {\\prime} |, \\tag {10}", + "image_path": "a0d6126237f352e6457aed65a4709c13083142d88b79cc5e214c1954b0506361.jpg" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 131, + 494, + 479, + 530 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 131, + 494, + 479, + 530 + ], + "spans": [ + { + "bbox": [ + 131, + 494, + 479, + 530 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 131, + 494, + 479, + 530 + ], + "type": "inline_equation", + "content": "m" + }, + { + "bbox": [ + 131, + 494, + 479, + 530 + ], + "type": "text", + "content": " is the channel number. For each " + }, + { + "bbox": [ + 131, + 494, + 479, + 530 + ], + "type": "inline_equation", + "content": "m" + }, + { + "bbox": [ + 131, + 494, + 479, + 530 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 131, + 494, + 479, + 530 + ], + "type": "inline_equation", + "content": "L(\\lambda_m')" + }, + { + "bbox": [ + 131, + 494, + 479, + 530 + ], + "type": "text", + "content": " is pre-calculated through the implicit representation of the spectrum. B is a constant that ensures the integral of " + }, + { + "bbox": [ + 131, + 494, + 479, + 530 + ], + "type": "inline_equation", + "content": "p(\\lambda)" + }, + { + "bbox": [ + 131, + 494, + 479, + 530 + ], + "type": "text", + "content": " equals 1. Then, the photocurrent can be obtained as follows:" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 154, + 541, + 481, + 572 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 154, + 541, + 481, + 572 + ], + "spans": [ + { + "bbox": [ + 154, + 541, + 481, + 572 + ], + "type": "interline_equation", + "content": "i _ {p} \\approx C \\cdot \\frac {1}{N} \\sum_ {k = 1} ^ {N} \\frac {\\lambda_ {k} \\cdot Q E (\\lambda_ {k}) \\cdot L (\\lambda_ {k})}{B \\cdot \\lambda_ {j _ {k}} ^ {\\prime} \\cdot Q E \\left(\\lambda_ {j _ {k}} ^ {\\prime}\\right) \\cdot L \\left(\\lambda_ {j _ {k}} ^ {\\prime}\\right)}, j _ {k} = \\operatorname {a r g m i n} _ {m} | \\lambda_ {k} - \\lambda_ {m} ^ {\\prime} |. \\tag {11}", + "image_path": "ba60921e71a271334b6dabe92715ba795a4a3f2a9229585f840c1fb272c45cc4.jpg" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 131, + 581, + 479, + 665 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 131, + 581, + 479, + 665 + ], + "spans": [ + { + "bbox": [ + 131, + 581, + 479, + 665 + ], + "type": "text", + "content": "Note that, Monte Carlo integration in Eq. 11 and simple numerical integration in Eq. 6 each have their own advantages and disadvantages. When the spectral data is represented continuously, Monte Carlo integration produces significantly smaller errors with a sharp spectral distribution curve compared to simple integration. However, if the spectrum is only represented by discrete channels, the lack of original information renders Monte Carlo integration somewhat redundant, and simple numerical integration can be used instead." + } + ] + } + ], + "index": 13 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 281, + 91, + 448, + 102 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 281, + 91, + 448, + 102 + ], + "spans": [ + { + "bbox": [ + 281, + 91, + 448, + 102 + ], + "type": "text", + "content": "Physical-Based Event Camera Simulator" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 474, + 91, + 481, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 474, + 91, + 481, + 100 + ], + "spans": [ + { + "bbox": [ + 474, + 91, + 481, + 100 + ], + "type": "text", + "content": "7" + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 6 + }, + { + "para_blocks": [ + { + "bbox": [ + 130, + 116, + 482, + 177 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 116, + 482, + 177 + ], + "spans": [ + { + "bbox": [ + 130, + 116, + 482, + 177 + ], + "type": "text", + "content": "Note that, our PECS estimates photocurrent by considering spectral data and quantum efficiency, offering more accurate simulations of light signal conversion. Consequently, the accuracy is significantly enhanced compared to the LUV method. Meanwhile, the quantum efficiency of different cameras often varies and can generally be provided by the corresponding sensor manufacturers." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 132, + 195, + 383, + 208 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 195, + 383, + 208 + ], + "spans": [ + { + "bbox": [ + 132, + 195, + 383, + 208 + ], + "type": "text", + "content": "3.4 Asynchronous Spatiotemporal Event Metric" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 130, + 217, + 482, + 266 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 217, + 482, + 266 + ], + "spans": [ + { + "bbox": [ + 130, + 217, + 482, + 266 + ], + "type": "text", + "content": "To measure the similarity between simulated and raw camera events, we present two asynchronous spatiotemporal event metrics (i.e., Chamfer distance and Gaussian distance), which is similar to the metric part in iterative closest point tasks [3,37]. The Chamfer distance metric " + }, + { + "bbox": [ + 130, + 217, + 482, + 266 + ], + "type": "inline_equation", + "content": "C_D" + }, + { + "bbox": [ + 130, + 217, + 482, + 266 + ], + "type": "text", + "content": " can be described as:" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 188, + 274, + 481, + 304 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 188, + 274, + 481, + 304 + ], + "spans": [ + { + "bbox": [ + 188, + 274, + 481, + 304 + ], + "type": "interline_equation", + "content": "C _ {D} (R, Q) = \\frac {1}{| R |} \\sum_ {r \\in R} \\min _ {q \\in Q} d (r, q) + \\frac {1}{| Q |} \\sum_ {q \\in Q} \\min _ {r \\in R} d (r, q), \\tag {12}", + "image_path": "0ade8aeef41fec517d2d51d2a27a81e0c2876c5f0297f4ae079e388334dfcd1f.jpg" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 131, + 312, + 482, + 361 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 131, + 312, + 482, + 361 + ], + "spans": [ + { + "bbox": [ + 131, + 312, + 482, + 361 + ], + "type": "text", + "content": "where the lower " + }, + { + "bbox": [ + 131, + 312, + 482, + 361 + ], + "type": "inline_equation", + "content": "C_D" + }, + { + "bbox": [ + 131, + 312, + 482, + 361 + ], + "type": "text", + "content": " its value, the closer the two events are. " + }, + { + "bbox": [ + 131, + 312, + 482, + 361 + ], + "type": "inline_equation", + "content": "R" + }, + { + "bbox": [ + 131, + 312, + 482, + 361 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 131, + 312, + 482, + 361 + ], + "type": "inline_equation", + "content": "Q" + }, + { + "bbox": [ + 131, + 312, + 482, + 361 + ], + "type": "text", + "content": " are two point clouds, " + }, + { + "bbox": [ + 131, + 312, + 482, + 361 + ], + "type": "inline_equation", + "content": "r" + }, + { + "bbox": [ + 131, + 312, + 482, + 361 + ], + "type": "text", + "content": " or " + }, + { + "bbox": [ + 131, + 312, + 482, + 361 + ], + "type": "inline_equation", + "content": "q" + }, + { + "bbox": [ + 131, + 312, + 482, + 361 + ], + "type": "text", + "content": " is a single point in " + }, + { + "bbox": [ + 131, + 312, + 482, + 361 + ], + "type": "inline_equation", + "content": "R" + }, + { + "bbox": [ + 131, + 312, + 482, + 361 + ], + "type": "text", + "content": " or " + }, + { + "bbox": [ + 131, + 312, + 482, + 361 + ], + "type": "inline_equation", + "content": "Q" + }, + { + "bbox": [ + 131, + 312, + 482, + 361 + ], + "type": "text", + "content": ". Each point can be expressed as a four-dimensional tuple " + }, + { + "bbox": [ + 131, + 312, + 482, + 361 + ], + "type": "inline_equation", + "content": "\\{x,y,p,t\\}" + }, + { + "bbox": [ + 131, + 312, + 482, + 361 + ], + "type": "text", + "content": " representing the spatial coordinates, polarity, and temporal information. The distance in the metric can be depicted as:" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 266, + 370, + 481, + 385 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 266, + 370, + 481, + 385 + ], + "spans": [ + { + "bbox": [ + 266, + 370, + 481, + 385 + ], + "type": "interline_equation", + "content": "d (r, q) = | | r - q | | _ {2}, \\tag {13}", + "image_path": "b78fc55983d4e8781e8677a3b21929ae42e12a9d2997544f44dc66e9e00f5e1a.jpg" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 132, + 392, + 485, + 417 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 392, + 485, + 417 + ], + "spans": [ + { + "bbox": [ + 132, + 392, + 485, + 417 + ], + "type": "text", + "content": "By utilizing KDTREE [9] for acceleration, the time complexity is only " + }, + { + "bbox": [ + 132, + 392, + 485, + 417 + ], + "type": "inline_equation", + "content": "O(n \\log(n))" + }, + { + "bbox": [ + 132, + 392, + 485, + 417 + ], + "type": "text", + "content": ". Similarly, the Gaussian distance can be formulated as:" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 175, + 426, + 481, + 455 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 175, + 426, + 481, + 455 + ], + "spans": [ + { + "bbox": [ + 175, + 426, + 481, + 455 + ], + "type": "interline_equation", + "content": "G _ {D} (R, Q) = \\frac {1}{| R |} \\sum_ {r \\in R} g \\left(\\min _ {q \\in Q} d (r, q)\\right) + \\frac {1}{| Q |} \\sum_ {q \\in Q} g \\left(\\min _ {r \\in R} d (r, q)\\right), \\tag {14}", + "image_path": "52f76de842600972db85157e0214487e642952946243dc1788c0dfccdba572d2.jpg" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 131, + 465, + 482, + 491 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 131, + 465, + 482, + 491 + ], + "spans": [ + { + "bbox": [ + 131, + 465, + 482, + 491 + ], + "type": "text", + "content": "where the function " + }, + { + "bbox": [ + 131, + 465, + 482, + 491 + ], + "type": "inline_equation", + "content": "g" + }, + { + "bbox": [ + 131, + 465, + 482, + 491 + ], + "type": "text", + "content": " makes " + }, + { + "bbox": [ + 131, + 465, + 482, + 491 + ], + "type": "inline_equation", + "content": "G_{D}(R, Q)" + }, + { + "bbox": [ + 131, + 465, + 482, + 491 + ], + "type": "text", + "content": " still become lower when the two events get closer while reducing the impact of outliers, and it can be described as:" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 247, + 498, + 481, + 525 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 247, + 498, + 481, + 525 + ], + "spans": [ + { + "bbox": [ + 247, + 498, + 481, + 525 + ], + "type": "interline_equation", + "content": "g (x) = 1 - \\exp \\left(- \\frac {| | x | | _ {2} ^ {2}}{\\sigma}\\right), \\tag {15}", + "image_path": "3eb7a49fe4a6071dc6bab1529ac169623c29986d069cd55070413d1d585980d6.jpg" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 131, + 533, + 482, + 556 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 131, + 533, + 482, + 556 + ], + "spans": [ + { + "bbox": [ + 131, + 533, + 482, + 556 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 131, + 533, + 482, + 556 + ], + "type": "inline_equation", + "content": "\\sigma" + }, + { + "bbox": [ + 131, + 533, + 482, + 556 + ], + "type": "text", + "content": " is a hyperparameter used to adjust the sensitivity range of the original metric. In this study, we set 0.4 as the default." + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 130, + 558, + 482, + 629 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 558, + 482, + 629 + ], + "spans": [ + { + "bbox": [ + 130, + 558, + 482, + 629 + ], + "type": "text", + "content": "In fact, these two metrics describe the attributes of asynchronous events from different dimensions and can be necessary to assess the similarity between raw and simulated data. The Gaussian distance exhibits relative stability against outliers compared to the Chamfer distance, making it more robust to noise. Nevertheless, its measurement results have a narrower range of variation, making it less distinctive in simultaneous evaluations of multiple simulators." + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 131, + 630, + 482, + 666 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 131, + 630, + 482, + 666 + ], + "spans": [ + { + "bbox": [ + 131, + 630, + 482, + 666 + ], + "type": "text", + "content": "Note that, the units of x, y, p, and t in the original data are different, and the range of values for t often far exceeds that of the other components, resulting in Eq. 13 and Eq. 14 almost do not consider information beyond t. To address this" + } + ] + } + ], + "index": 14 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "spans": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "type": "text", + "content": "8" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 166, + 91, + 257, + 102 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 166, + 91, + 257, + 102 + ], + "spans": [ + { + "bbox": [ + 166, + 91, + 257, + 102 + ], + "type": "text", + "content": "Haiqian. Author et al." + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 7 + }, + { + "para_blocks": [ + { + "bbox": [ + 130, + 116, + 479, + 140 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 116, + 479, + 140 + ], + "spans": [ + { + "bbox": [ + 130, + 116, + 479, + 140 + ], + "type": "text", + "content": "issue, it is necessary to normalize the original measurements first to resolve this scaling problem as follows:" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 230, + 145, + 481, + 174 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 230, + 145, + 481, + 174 + ], + "spans": [ + { + "bbox": [ + 230, + 145, + 481, + 174 + ], + "type": "interline_equation", + "content": "L _ {\\beta} [ i ] = \\left(\\frac {L _ {\\beta} ^ {\\prime} [ i ] - \\operatorname* {m i n} L _ {\\beta} ^ {\\prime}}{\\operatorname* {m a x} L _ {\\beta} ^ {\\prime} - \\operatorname* {m i n} L _ {\\beta} ^ {\\prime} + \\epsilon}\\right) * \\alpha , \\tag {16}", + "image_path": "e41226e282cedf1312f0ada17e0bd8d9245f02e3e8efa297bbe31b4676c4389e.jpg" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 131, + 179, + 482, + 240 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 131, + 179, + 482, + 240 + ], + "spans": [ + { + "bbox": [ + 131, + 179, + 482, + 240 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 131, + 179, + 482, + 240 + ], + "type": "inline_equation", + "content": "L_{\\beta}^{\\prime}" + }, + { + "bbox": [ + 131, + 179, + 482, + 240 + ], + "type": "text", + "content": " is the original measurements, " + }, + { + "bbox": [ + 131, + 179, + 482, + 240 + ], + "type": "inline_equation", + "content": "\\beta" + }, + { + "bbox": [ + 131, + 179, + 482, + 240 + ], + "type": "text", + "content": " can be " + }, + { + "bbox": [ + 131, + 179, + 482, + 240 + ], + "type": "inline_equation", + "content": "x, y, p" + }, + { + "bbox": [ + 131, + 179, + 482, + 240 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 131, + 179, + 482, + 240 + ], + "type": "inline_equation", + "content": "t" + }, + { + "bbox": [ + 131, + 179, + 482, + 240 + ], + "type": "text", + "content": " element. " + }, + { + "bbox": [ + 131, + 179, + 482, + 240 + ], + "type": "inline_equation", + "content": "L_{\\beta}" + }, + { + "bbox": [ + 131, + 179, + 482, + 240 + ], + "type": "text", + "content": " is the corresponding processed data. " + }, + { + "bbox": [ + 131, + 179, + 482, + 240 + ], + "type": "inline_equation", + "content": "i" + }, + { + "bbox": [ + 131, + 179, + 482, + 240 + ], + "type": "text", + "content": " is the i'th event of the raw data, which means " + }, + { + "bbox": [ + 131, + 179, + 482, + 240 + ], + "type": "inline_equation", + "content": "R[i] = \\{L_x[i], L_y[i], L_p[i], L_t[i]\\}" + }, + { + "bbox": [ + 131, + 179, + 482, + 240 + ], + "type": "text", + "content": ". " + }, + { + "bbox": [ + 131, + 179, + 482, + 240 + ], + "type": "inline_equation", + "content": "\\epsilon" + }, + { + "bbox": [ + 131, + 179, + 482, + 240 + ], + "type": "text", + "content": " is a small number that is to avoid dividing zero, and " + }, + { + "bbox": [ + 131, + 179, + 482, + 240 + ], + "type": "inline_equation", + "content": "\\alpha" + }, + { + "bbox": [ + 131, + 179, + 482, + 240 + ], + "type": "text", + "content": " is a hyperparameter. In this work, we choose " + }, + { + "bbox": [ + 131, + 179, + 482, + 240 + ], + "type": "inline_equation", + "content": "\\alpha = 100" + }, + { + "bbox": [ + 131, + 179, + 482, + 240 + ], + "type": "text", + "content": " for space " + }, + { + "bbox": [ + 131, + 179, + 482, + 240 + ], + "type": "inline_equation", + "content": "x, y" + }, + { + "bbox": [ + 131, + 179, + 482, + 240 + ], + "type": "text", + "content": " and polarity " + }, + { + "bbox": [ + 131, + 179, + 482, + 240 + ], + "type": "inline_equation", + "content": "p" + }, + { + "bbox": [ + 131, + 179, + 482, + 240 + ], + "type": "text", + "content": ", and choose " + }, + { + "bbox": [ + 131, + 179, + 482, + 240 + ], + "type": "inline_equation", + "content": "\\alpha = 1000" + }, + { + "bbox": [ + 131, + 179, + 482, + 240 + ], + "type": "text", + "content": " for time " + }, + { + "bbox": [ + 131, + 179, + 482, + 240 + ], + "type": "inline_equation", + "content": "t" + }, + { + "bbox": [ + 131, + 179, + 482, + 240 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 132, + 255, + 230, + 269 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 255, + 230, + 269 + ], + "spans": [ + { + "bbox": [ + 132, + 255, + 230, + 269 + ], + "type": "text", + "content": "4 Experiments" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 132, + 277, + 269, + 290 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 277, + 269, + 290 + ], + "spans": [ + { + "bbox": [ + 132, + 277, + 269, + 290 + ], + "type": "text", + "content": "4.1 Experimental Setting" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 130, + 294, + 482, + 666 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 294, + 482, + 666 + ], + "spans": [ + { + "bbox": [ + 130, + 294, + 482, + 666 + ], + "type": "text", + "content": "Real-to-Sim Scenario Construction. Spatiotemporal synchronization is essential for quantitatively evaluating, we record the real-world event data using a Prophesee Gen4 camera in a professional optical laboratory. We capture six sequences including rotating disk and translating checkerboards with light changes and various motion speeds. For example, T_0.6_H means translating checkerboards, 0.6 meter-per-seconds, and high light. R_360_L means rotating disk, 360 rpm, and low light. To ensure consistency between the simulated and real-world scenes, we utilize the blender for high-fidelity 3D modeling. We export the scenes to PBRT and simulate a naturally calibrated camera using real lenses [26,42]. This approach enables us to acquire temporally and spatially aligned rendering results for use in subsequent sensor modules. More details of spatiotemporal alignment can be found in the supplementary material. In other methods, as regular RGB cameras struggle to capture clear images in these scenes, we use Blender's rendered RGB video as input. By manually calibrating pinhole camera parameters, we ensure temporally and spatially consistent clear images for simulation inputs, enabling the proper functioning of these simulators. Implementation Details. All experiments are conducted on a GeForce RTX 3050 Ti and 11th Gen Intel(R) Core(TM) i7-11800H CPU. We compare our PECS with four open-source event camera simulators (i.e., ESIM [31], VOLT [20], V2E [12], and ICNS [14]). In our PECS, the hyperparameters in Eq. 11 and Eq. 6 are set to " + }, + { + "bbox": [ + 130, + 294, + 482, + 666 + ], + "type": "inline_equation", + "content": "C = 100" + }, + { + "bbox": [ + 130, + 294, + 482, + 666 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 130, + 294, + 482, + 666 + ], + "type": "inline_equation", + "content": "N = 31" + }, + { + "bbox": [ + 130, + 294, + 482, + 666 + ], + "type": "text", + "content": " to make an accuracy-speed trade-off. Regarding the ESIM method, we extract its core code from its original ties to Unreal Engine (UE) and replace UE output with video output. For the V2E method, given the clarity and compactness of Blender-generated images, we disable the network interpolation module to prevent extremely slow simulation speed. In the ICNS method, all parameters have been adjusted to achieve optimal results. In the VOLT method, we select a set of parameters for the best performance. We present two event metrics (i.e., Chamfer distance and Gaussian distance) to quantitatively evaluate the similarity between raw events and synthetic data. Besides, we select three representative sequences of the HS-ERGB dataset [45] in the event-based video reconstruction task." + } + ] + } + ], + "index": 7 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 282, + 91, + 447, + 102 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 282, + 91, + 447, + 102 + ], + "spans": [ + { + "bbox": [ + 282, + 91, + 447, + 102 + ], + "type": "text", + "content": "Physical-Based Event Camera Simulator" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 474, + 92, + 480, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 474, + 92, + 480, + 100 + ], + "spans": [ + { + "bbox": [ + 474, + 92, + 480, + 100 + ], + "type": "text", + "content": "9" + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 8 + }, + { + "para_blocks": [ + { + "type": "table", + "bbox": [ + 135, + 157, + 476, + 243 + ], + "blocks": [ + { + "bbox": [ + 132, + 114, + 482, + 148 + ], + "lines": [ + { + "bbox": [ + 132, + 114, + 482, + 148 + ], + "spans": [ + { + "bbox": [ + 132, + 114, + 482, + 148 + ], + "type": "text", + "content": "Table 1: Comparison with state-of-the-art simulators on our six recording sequences. Note that, our Physical Event Camera Simulator (PECS) shows superior performance on both metrics across a diverse range of movement speeds and lighting conditions." + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 135, + 157, + 476, + 243 + ], + "lines": [ + { + "bbox": [ + 135, + 157, + 476, + 243 + ], + "spans": [ + { + "bbox": [ + 135, + 157, + 476, + 243 + ], + "type": "table", + "html": "
SceneGaussian distance ↓Chamfer distance ↓
ESIMVOLTV2EICNSPECSESIMVOLTV2EICNSPECS
Rotate 360 High1.9951.9231.6711.5580.96045.51222.0866.3563.8162.010
Rotate 360 Low1.9651.9311.8311.6811.63135.08225.3917.7294.1313.708
Rotate 60 High1.9951.9111.8161.5741.50439.42520.8149.2974.2373.459
Trans 1mps High1.9631.6551.8271.8061.66762.20610.57315.4509.4533.975
Trans 06mps High1.9671.7231.8391.8371.79863.93910.19818.1549.5884.169
Trans 1mps Low1.9871.8141.7001.6631.71068.11510.95214.0719.1824.467
Average1.9791.8261.7811.6871.54552.38016.66911.8436.7353.631
", + "image_path": "2791daa0e5d1507350134b3563d54e40cd66c4c9df309de2c8ce1e5012159ea6.jpg" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "table_body" + } + ], + "index": 3 + }, + { + "bbox": [ + 132, + 264, + 230, + 276 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 264, + 230, + 276 + ], + "spans": [ + { + "bbox": [ + 132, + 264, + 230, + 276 + ], + "type": "text", + "content": "4.2 Effective Test" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 130, + 283, + 481, + 390 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 283, + 481, + 390 + ], + "spans": [ + { + "bbox": [ + 130, + 283, + 481, + 390 + ], + "type": "text", + "content": "Quantitative Evaluation. To quantitatively evaluate the effectiveness of our PECS, we conduct a comparison with four open-source event camera simulators (i.e., ESIM [31], VOLT [20], V2E [12], and ICNS [14]) on six recording sequences (see Table 1). Note that, our PECS outperforms four state-of-the-art simulators in both two event metrics. More precisely, in terms of the Gaussian distance, our PECS reduces the average by 0.434, 0.281, 0.236, and 0.142 when compared to ESIM, VOLT, V2E, and ICNS, respectively. Meanwhile, in the Chamfer distance, our simulator exhibits an average decrease of 48.749, 13.038, 8.212, and 3.104 compared to ESIM, VOLT, V2E, and ICNS, respectively." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 130, + 391, + 481, + 559 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 391, + 481, + 559 + ], + "spans": [ + { + "bbox": [ + 130, + 391, + 481, + 559 + ], + "type": "text", + "content": "Visualization Evaluation. Some representative visualization results on two motion scenarios (i.e., translating checkerboards and rotation disks) recorded in a professional optical laboratory are illustrated in Fig. 3 and Fig. 4. Obviously, our PECS achieves the best performance against four state-of-the-art simulators including ESIM, VOLT, V2E, and ICNS. We can find that V2E and ESIM simulators are too primitive, resulting in concentrated data distribution in 3D space and a lack of continuity. The event data of ESIM is entirely concentrated at the timestamp of the input frame, lacking continuity, which is the main reason for its poor performance. VOLT performs well in terms of the continuity of event output, but its noise model suffers from significant distortion. Additionally, VOLT simply converts three-channel video to grayscale as photocurrent, without considering quantum efficiency and spectrum. Although the ICNS has an advanced sensor model and supports scene input, the lack of algorithms in the section leads to less accurate event data generation." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 132, + 574, + 239, + 588 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 574, + 239, + 588 + ], + "spans": [ + { + "bbox": [ + 132, + 574, + 239, + 588 + ], + "type": "text", + "content": "4.3 Ablation Study" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 130, + 593, + 481, + 665 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 593, + 481, + 665 + ], + "spans": [ + { + "bbox": [ + 130, + 593, + 481, + 665 + ], + "type": "text", + "content": "Influence of the Sampling Number " + }, + { + "bbox": [ + 130, + 593, + 481, + 665 + ], + "type": "inline_equation", + "content": "N" + }, + { + "bbox": [ + 130, + 593, + 481, + 665 + ], + "type": "text", + "content": ". To analyze the sampling number " + }, + { + "bbox": [ + 130, + 593, + 481, + 665 + ], + "type": "inline_equation", + "content": "N" + }, + { + "bbox": [ + 130, + 593, + 481, + 665 + ], + "type": "text", + "content": " of numerical integration in our PECS, we set the multispectral rendering module with various values of " + }, + { + "bbox": [ + 130, + 593, + 481, + 665 + ], + "type": "inline_equation", + "content": "N" + }, + { + "bbox": [ + 130, + 593, + 481, + 665 + ], + "type": "text", + "content": ", and set " + }, + { + "bbox": [ + 130, + 593, + 481, + 665 + ], + "type": "inline_equation", + "content": "C" + }, + { + "bbox": [ + 130, + 593, + 481, + 665 + ], + "type": "text", + "content": " as 100. As described in Table 2, we find that the Chamfer distance and Gaussian distance decrease with the increase of N. More precisely, Comparing " + }, + { + "bbox": [ + 130, + 593, + 481, + 665 + ], + "type": "inline_equation", + "content": "N = 31" + }, + { + "bbox": [ + 130, + 593, + 481, + 665 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 130, + 593, + 481, + 665 + ], + "type": "inline_equation", + "content": "N = 6" + }, + { + "bbox": [ + 130, + 593, + 481, + 665 + ], + "type": "text", + "content": ", the Chamfer distance and Gaussian distance decrease by 6.58 and 0.619, respectively. Overall, this results show that" + } + ] + } + ], + "index": 8 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "spans": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "type": "text", + "content": "10" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 166, + 91, + 257, + 102 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 166, + 91, + 257, + 102 + ], + "spans": [ + { + "bbox": [ + 166, + 91, + 257, + 102 + ], + "type": "text", + "content": "Haiqian. Author et al." + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 9 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 132, + 114, + 481, + 311 + ], + "blocks": [ + { + "bbox": [ + 132, + 114, + 481, + 311 + ], + "lines": [ + { + "bbox": [ + 132, + 114, + 481, + 311 + ], + "spans": [ + { + "bbox": [ + 132, + 114, + 481, + 311 + ], + "type": "image", + "image_path": "6b8abcfd0b45c52e32bbfc5c50e3ec8642a4f90255d05aebba948309acb0dbc8.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 130, + 319, + 482, + 352 + ], + "lines": [ + { + "bbox": [ + 130, + 319, + 482, + 352 + ], + "spans": [ + { + "bbox": [ + 130, + 319, + 482, + 352 + ], + "type": "text", + "content": "Fig. 3: Representative examples of translating checkerboards with various motion speed and light conditions. In each scenario, the first row is a 2D visualization schematic of the event data, and the second row is a 3D visualization schematic of the event data." + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_caption" + } + ], + "index": 2 + }, + { + "type": "table", + "bbox": [ + 133, + 415, + 481, + 462 + ], + "blocks": [ + { + "bbox": [ + 130, + 372, + 482, + 407 + ], + "lines": [ + { + "bbox": [ + 130, + 372, + 482, + 407 + ], + "spans": [ + { + "bbox": [ + 130, + 372, + 482, + 407 + ], + "type": "text", + "content": "Table 2: The influence of the sampling number " + }, + { + "bbox": [ + 130, + 372, + 482, + 407 + ], + "type": "inline_equation", + "content": "N" + }, + { + "bbox": [ + 130, + 372, + 482, + 407 + ], + "type": "text", + "content": " of Monte-Carlo integration in the multispectral rendering module. It's worth noting that our PECS exhibits improved performance with an increase in " + }, + { + "bbox": [ + 130, + 372, + 482, + 407 + ], + "type": "inline_equation", + "content": "N" + }, + { + "bbox": [ + 130, + 372, + 482, + 407 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 133, + 415, + 481, + 462 + ], + "lines": [ + { + "bbox": [ + 133, + 415, + 481, + 462 + ], + "spans": [ + { + "bbox": [ + 133, + 415, + 481, + 462 + ], + "type": "table", + "html": "
N13467111631
Chamfer distance ↓13.968.82616.288.5888.4662.8692.4762.008
Gaussian distance ↓1.5971.6241.5691.5831.5821.1761.0910.964
", + "image_path": "0ce6e6421832c93f9d6a5e2be57e96dcde4ff52c083c81b640701a065692eea2.jpg" + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "table_body" + } + ], + "index": 5 + }, + { + "bbox": [ + 130, + 494, + 480, + 531 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 494, + 480, + 531 + ], + "spans": [ + { + "bbox": [ + 130, + 494, + 480, + 531 + ], + "type": "text", + "content": "the larger the " + }, + { + "bbox": [ + 130, + 494, + 480, + 531 + ], + "type": "inline_equation", + "content": "N" + }, + { + "bbox": [ + 130, + 494, + 480, + 531 + ], + "type": "text", + "content": ", the closer the final simulation data is to the real data. Of course, an increase in " + }, + { + "bbox": [ + 130, + 494, + 480, + 531 + ], + "type": "inline_equation", + "content": "N" + }, + { + "bbox": [ + 130, + 494, + 480, + 531 + ], + "type": "text", + "content": " also entails a rise in time complexity. Thus, there is often a consideration of the trade-off between accuracy and computational speed." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 130, + 533, + 482, + 666 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 533, + 482, + 666 + ], + "spans": [ + { + "bbox": [ + 130, + 533, + 482, + 666 + ], + "type": "text", + "content": "Influence of the Hyperparameter " + }, + { + "bbox": [ + 130, + 533, + 482, + 666 + ], + "type": "inline_equation", + "content": "C" + }, + { + "bbox": [ + 130, + 533, + 482, + 666 + ], + "type": "text", + "content": ". To analyze the effect of the hyperparameter " + }, + { + "bbox": [ + 130, + 533, + 482, + 666 + ], + "type": "inline_equation", + "content": "C" + }, + { + "bbox": [ + 130, + 533, + 482, + 666 + ], + "type": "text", + "content": " of numerical integration on the final performance, we set the multispectral rendering module with various values of " + }, + { + "bbox": [ + 130, + 533, + 482, + 666 + ], + "type": "inline_equation", + "content": "C" + }, + { + "bbox": [ + 130, + 533, + 482, + 666 + ], + "type": "text", + "content": ", and set " + }, + { + "bbox": [ + 130, + 533, + 482, + 666 + ], + "type": "inline_equation", + "content": "N" + }, + { + "bbox": [ + 130, + 533, + 482, + 666 + ], + "type": "text", + "content": " as 31. From Table 3, we can find that there is no simple correlation between the hyperparameter " + }, + { + "bbox": [ + 130, + 533, + 482, + 666 + ], + "type": "inline_equation", + "content": "C" + }, + { + "bbox": [ + 130, + 533, + 482, + 666 + ], + "type": "text", + "content": " and the final performance. The distances of the two metrics gradually decrease with an increase in " + }, + { + "bbox": [ + 130, + 533, + 482, + 666 + ], + "type": "inline_equation", + "content": "C" + }, + { + "bbox": [ + 130, + 533, + 482, + 666 + ], + "type": "text", + "content": " until they start to increase at nearly 100. Essentially speaking, the hyperparameter " + }, + { + "bbox": [ + 130, + 533, + 482, + 666 + ], + "type": "inline_equation", + "content": "C" + }, + { + "bbox": [ + 130, + 533, + 482, + 666 + ], + "type": "text", + "content": " is introduced to align the standard unit system with the unit system used in the actual code. Theoretical analysis indicates the existence of an optimal value, a notion supported by ablation experiment results. The optimal value is observed to be around 100, and the distances of the two metrics decrease as the value of " + }, + { + "bbox": [ + 130, + 533, + 482, + 666 + ], + "type": "inline_equation", + "content": "C" + }, + { + "bbox": [ + 130, + 533, + 482, + 666 + ], + "type": "text", + "content": " deviates from this optimal point." + } + ] + } + ], + "index": 7 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 282, + 91, + 447, + 102 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 282, + 91, + 447, + 102 + ], + "spans": [ + { + "bbox": [ + 282, + 91, + 447, + 102 + ], + "type": "text", + "content": "Physical-Based Event Camera Simulator" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 470, + 92, + 479, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 470, + 92, + 479, + 100 + ], + "spans": [ + { + "bbox": [ + 470, + 92, + 479, + 100 + ], + "type": "text", + "content": "11" + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 10 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 130, + 115, + 481, + 311 + ], + "blocks": [ + { + "bbox": [ + 130, + 115, + 481, + 311 + ], + "lines": [ + { + "bbox": [ + 130, + 115, + 481, + 311 + ], + "spans": [ + { + "bbox": [ + 130, + 115, + 481, + 311 + ], + "type": "image", + "image_path": "fb5a6526e6be2c8b9b9f46ccfd2a5197331884eaae76e58c664a80883829baac.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 130, + 319, + 482, + 352 + ], + "lines": [ + { + "bbox": [ + 130, + 319, + 482, + 352 + ], + "spans": [ + { + "bbox": [ + 130, + 319, + 482, + 352 + ], + "type": "text", + "content": "Fig. 4: Representative visualization results of rotation disks with various motion speed and light conditions. In each scenario, the first row is a 2D visualization schematic of the event data, and the second row is a 3D visualization schematic of the event data." + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_caption" + } + ], + "index": 2 + }, + { + "type": "table", + "bbox": [ + 132, + 406, + 482, + 454 + ], + "blocks": [ + { + "bbox": [ + 130, + 363, + 482, + 397 + ], + "lines": [ + { + "bbox": [ + 130, + 363, + 482, + 397 + ], + "spans": [ + { + "bbox": [ + 130, + 363, + 482, + 397 + ], + "type": "text", + "content": "Table 3: The influence of the hyperparameter " + }, + { + "bbox": [ + 130, + 363, + 482, + 397 + ], + "type": "inline_equation", + "content": "C" + }, + { + "bbox": [ + 130, + 363, + 482, + 397 + ], + "type": "text", + "content": " of Monte-Carlo integration in the multispectral rendering module. The distances gradually decrease with an increase in " + }, + { + "bbox": [ + 130, + 363, + 482, + 397 + ], + "type": "inline_equation", + "content": "C" + }, + { + "bbox": [ + 130, + 363, + 482, + 397 + ], + "type": "text", + "content": " until they start to increase at nearly 100." + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 132, + 406, + 482, + 454 + ], + "lines": [ + { + "bbox": [ + 132, + 406, + 482, + 454 + ], + "spans": [ + { + "bbox": [ + 132, + 406, + 482, + 454 + ], + "type": "table", + "html": "
C10306080100120140160
Chamfer distance ↓6.4937.4452.5085.0662.0082.0112.2642.396
Gaussian distance ↓1.3421.4881.1061.4080.9640.9861.0501.076
", + "image_path": "5eb0bbb5fda3a0e0a2d31c856366bea114b61a5fb8af7a83f33f8074a594180a.jpg" + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "table_body" + } + ], + "index": 5 + }, + { + "bbox": [ + 132, + 474, + 308, + 488 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 474, + 308, + 488 + ], + "spans": [ + { + "bbox": [ + 132, + 474, + 308, + 488 + ], + "type": "text", + "content": "5 Downstream Applications" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 130, + 498, + 482, + 666 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 498, + 482, + 666 + ], + "spans": [ + { + "bbox": [ + 130, + 498, + 482, + 666 + ], + "type": "text", + "content": "In this section, we integrate our PECS with the Unreal Engine (UE) platform to generate extensive event-based vision datasets that support multiple tasks. We then proceed to validate the simulator's performance on the event-based video reconstruction task. In contrast to existing simulators, the representative algorithm trained on our synthetic events performs favorably on real event data. Extensive Multi-task Synthetic Datasets. To seamlessly link our PECS with downstream computer vision tasks, we incorporate it into the widely used UE platform, a popular tool for virtual environments. Our PECS, excluding the lens simulation module, is deployed on the UE platform to achieve real-time generation of multi-modal datasets (e.g., RGB frames and events). To facilitate various visual tasks, we also develop an automatic label annotation tool on the UE platform, providing bounding boxes for object detection and scene distance values for depth estimation. As depicted in Fig. 5, we present the synthetic data of a stereo hybrid camera on a drone platform, which mainly includes" + } + ] + } + ], + "index": 7 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "spans": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "type": "text", + "content": "12" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 166, + 91, + 257, + 102 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 166, + 91, + 257, + 102 + ], + "spans": [ + { + "bbox": [ + 166, + 91, + 257, + 102 + ], + "type": "text", + "content": "Haiqian. Author et al." + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 11 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 135, + 125, + 202, + 176 + ], + "blocks": [ + { + "bbox": [ + 143, + 116, + 195, + 125 + ], + "lines": [ + { + "bbox": [ + 143, + 116, + 195, + 125 + ], + "spans": [ + { + "bbox": [ + 143, + 116, + 195, + 125 + ], + "type": "text", + "content": "Left RGB image" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 135, + 125, + 202, + 176 + ], + "lines": [ + { + "bbox": [ + 135, + 125, + 202, + 176 + ], + "spans": [ + { + "bbox": [ + 135, + 125, + 202, + 176 + ], + "type": "image", + "image_path": "caa842c3ed36fb0f88478f1f87c390f1a8a0ae322ae0c16a8dca2560d6f13875.jpg" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_body" + } + ], + "index": 3 + }, + { + "type": "image", + "bbox": [ + 205, + 125, + 271, + 176 + ], + "blocks": [ + { + "bbox": [ + 210, + 116, + 266, + 125 + ], + "lines": [ + { + "bbox": [ + 210, + 116, + 266, + 125 + ], + "spans": [ + { + "bbox": [ + 210, + 116, + 266, + 125 + ], + "type": "text", + "content": "Right RGB image" + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 205, + 125, + 271, + 176 + ], + "lines": [ + { + "bbox": [ + 205, + 125, + 271, + 176 + ], + "spans": [ + { + "bbox": [ + 205, + 125, + 271, + 176 + ], + "type": "image", + "image_path": "27c06e9ff24039c33754a48d733035fa078da15166e2f299225851d7c93d2e6c.jpg" + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "image_body" + } + ], + "index": 5 + }, + { + "type": "image", + "bbox": [ + 272, + 125, + 339, + 176 + ], + "blocks": [ + { + "bbox": [ + 285, + 116, + 324, + 125 + ], + "lines": [ + { + "bbox": [ + 285, + 116, + 324, + 125 + ], + "spans": [ + { + "bbox": [ + 285, + 116, + 324, + 125 + ], + "type": "text", + "content": "Event image" + } + ] + } + ], + "index": 6, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 272, + 125, + 339, + 176 + ], + "lines": [ + { + "bbox": [ + 272, + 125, + 339, + 176 + ], + "spans": [ + { + "bbox": [ + 272, + 125, + 339, + 176 + ], + "type": "image", + "image_path": "3f5501a465017a932a117e8107182a85582cf74e15adecb1f36447518daf233b.jpg" + } + ] + } + ], + "index": 7, + "angle": 0, + "type": "image_body" + } + ], + "index": 7 + }, + { + "type": "image", + "bbox": [ + 341, + 125, + 410, + 176 + ], + "blocks": [ + { + "bbox": [ + 340, + 116, + 413, + 125 + ], + "lines": [ + { + "bbox": [ + 340, + 116, + 413, + 125 + ], + "spans": [ + { + "bbox": [ + 340, + 116, + 413, + 125 + ], + "type": "text", + "content": "Object bounding boxes" + } + ] + } + ], + "index": 8, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 341, + 125, + 410, + 176 + ], + "lines": [ + { + "bbox": [ + 341, + 125, + 410, + 176 + ], + "spans": [ + { + "bbox": [ + 341, + 125, + 410, + 176 + ], + "type": "image", + "image_path": "a9c6256610cb4901a794d39bfa4122ef72e7a0291be4003bbaa1603014d2ebbe.jpg" + } + ] + } + ], + "index": 9, + "angle": 0, + "type": "image_body" + } + ], + "index": 9 + }, + { + "type": "image", + "bbox": [ + 410, + 125, + 478, + 176 + ], + "blocks": [ + { + "bbox": [ + 427, + 116, + 463, + 125 + ], + "lines": [ + { + "bbox": [ + 427, + 116, + 463, + 125 + ], + "spans": [ + { + "bbox": [ + 427, + 116, + 463, + 125 + ], + "type": "text", + "content": "Depth label" + } + ] + } + ], + "index": 10, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 410, + 125, + 478, + 176 + ], + "lines": [ + { + "bbox": [ + 410, + 125, + 478, + 176 + ], + "spans": [ + { + "bbox": [ + 410, + 125, + 478, + 176 + ], + "type": "image", + "image_path": "0d56cc1591bb9afd15f027e06cecc87cb584bc438f14d6b464888f5b730e6a7e.jpg" + } + ] + } + ], + "index": 11, + "angle": 0, + "type": "image_body" + } + ], + "index": 11 + }, + { + "type": "image", + "bbox": [ + 135, + 178, + 202, + 228 + ], + "blocks": [ + { + "bbox": [ + 135, + 178, + 202, + 228 + ], + "lines": [ + { + "bbox": [ + 135, + 178, + 202, + 228 + ], + "spans": [ + { + "bbox": [ + 135, + 178, + 202, + 228 + ], + "type": "image", + "image_path": "64cb9e86a0805601ed6d072b082da28d3b2114d5bb6219683643d25531590809.jpg" + } + ] + } + ], + "index": 12, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 131, + 240, + 480, + 274 + ], + "lines": [ + { + "bbox": [ + 131, + 240, + 480, + 274 + ], + "spans": [ + { + "bbox": [ + 131, + 240, + 480, + 274 + ], + "type": "text", + "content": "Fig. 5: Representative examples of multi-task synthetic datasets. We integrate our PECS into the UE platform to generate extensive synthetic datasets supporting multiple tasks, such as video reconstruction, object detection, and depth estimation." + } + ] + } + ], + "index": 17, + "angle": 0, + "type": "image_caption" + } + ], + "index": 12 + }, + { + "type": "image", + "bbox": [ + 205, + 178, + 271, + 228 + ], + "blocks": [ + { + "bbox": [ + 205, + 178, + 271, + 228 + ], + "lines": [ + { + "bbox": [ + 205, + 178, + 271, + 228 + ], + "spans": [ + { + "bbox": [ + 205, + 178, + 271, + 228 + ], + "type": "image", + "image_path": "7c411251c60e7c11683fafe1f4529cb5d113c1baa96db1b41e416471fe307101.jpg" + } + ] + } + ], + "index": 13, + "angle": 0, + "type": "image_body" + } + ], + "index": 13 + }, + { + "type": "image", + "bbox": [ + 272, + 178, + 340, + 228 + ], + "blocks": [ + { + "bbox": [ + 272, + 178, + 340, + 228 + ], + "lines": [ + { + "bbox": [ + 272, + 178, + 340, + 228 + ], + "spans": [ + { + "bbox": [ + 272, + 178, + 340, + 228 + ], + "type": "image", + "image_path": "3c197aed4f9ec1e4b3734499079ac6a6fbe4a0c1b736eb50f6b0d845310dad06.jpg" + } + ] + } + ], + "index": 14, + "angle": 0, + "type": "image_body" + } + ], + "index": 14 + }, + { + "type": "image", + "bbox": [ + 342, + 178, + 409, + 228 + ], + "blocks": [ + { + "bbox": [ + 342, + 178, + 409, + 228 + ], + "lines": [ + { + "bbox": [ + 342, + 178, + 409, + 228 + ], + "spans": [ + { + "bbox": [ + 342, + 178, + 409, + 228 + ], + "type": "image", + "image_path": "75f834cb943930e68af247341ba0477d90d1ae44bced09ed9dd0041f5b2e17ef.jpg" + } + ] + } + ], + "index": 15, + "angle": 0, + "type": "image_body" + } + ], + "index": 15 + }, + { + "type": "image", + "bbox": [ + 410, + 178, + 479, + 228 + ], + "blocks": [ + { + "bbox": [ + 410, + 178, + 479, + 228 + ], + "lines": [ + { + "bbox": [ + 410, + 178, + 479, + 228 + ], + "spans": [ + { + "bbox": [ + 410, + 178, + 479, + 228 + ], + "type": "image", + "image_path": "2929c1a2d862659b0f84a6a12b17af31a74210fe9779a19a73aec6e3551a9c6e.jpg" + } + ] + } + ], + "index": 16, + "angle": 0, + "type": "image_body" + } + ], + "index": 16 + }, + { + "type": "table", + "bbox": [ + 135, + 342, + 479, + 412 + ], + "blocks": [ + { + "bbox": [ + 131, + 289, + 480, + 322 + ], + "lines": [ + { + "bbox": [ + 131, + 289, + 480, + 322 + ], + "spans": [ + { + "bbox": [ + 131, + 289, + 480, + 322 + ], + "type": "text", + "content": "Table 4: Performance comparison on event-based video reconstruction. The E2VID network [32] trains on synthetic data from each event camera simulator and tests on three representative sequences of the real-world HS-ERGB dataset [45]." + } + ] + } + ], + "index": 18, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 135, + 342, + 479, + 412 + ], + "lines": [ + { + "bbox": [ + 135, + 342, + 479, + 412 + ], + "spans": [ + { + "bbox": [ + 135, + 342, + 479, + 412 + ], + "type": "table", + "html": "
SequencesPSNR↑SSIM↑LPIPS↓
ESIMV2EPECSESIMV2EPECSESIMV2EPECS
spinning_umbrella8.7568.4379.2360.4720.4710.4800.4510.4490.430
fountain_bellevue26.1395.8676.5640.2130.2110.2180.6750.6700.662
bridge_lake_0311.77012.68613.2430.5410.5600.5630.3320.2740.250
Mean8.8898.9979.6780.4090.4140.4200.4860.4640.447
", + "image_path": "5f8859ab1fe3f0c0fabeff4fc55831732b5b0d21ffb3188655f24026bc61d5e4.jpg" + } + ] + } + ], + "index": 19, + "angle": 0, + "type": "table_body" + } + ], + "index": 19 + }, + { + "bbox": [ + 130, + 448, + 480, + 532 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 448, + 480, + 532 + ], + "spans": [ + { + "bbox": [ + 130, + 448, + 480, + 532 + ], + "type": "text", + "content": "RGB images, event streams, object detection labels, and scene depth values. Notably, our PECS can generate comprehensive synthetic datasets using both monocular and stereo cameras, supporting a variety of vision tasks, including video reconstruction, object detection, monocular depth estimation, binocular depth estimation, and more. Besides, we also adjust the speed of motion cameras and scene light intensity, generating diverse datasets with high-speed motion and low-light conditions, showcasing the advantages of event cameras." + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 130, + 533, + 482, + 666 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 533, + 482, + 666 + ], + "spans": [ + { + "bbox": [ + 130, + 533, + 482, + 666 + ], + "type": "text", + "content": "Validation on Event-based Video Reconstruction. Event-based video reconstruction [32,56] is a typical task that effectively validates the effectiveness of synthetic data from various simulators. Specifically, we first use our PECS and two competitive simulators (e.g., ESIM [31] and V2E [12]) to generate large-scale synthetic datasets in the UE platform. Then, we select a representative event-to-video reconstruction algorithm (i.e., E2VID [32]) to be trained on these simulated datasets respectively. The training details align with those suggested in E2VID [32]. Finally, we test various trained models on the real-world HS-ERGB dataset [45]. We use three metrics (i.e., PSNR, SSIM [46], and LPIPS [51]) to measure the performance of event-based video reconstruction (see Table 4). Obviously, our PECS shows superior generalization capability compared to the top" + } + ] + } + ], + "index": 21 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 282, + 91, + 447, + 102 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 282, + 91, + 447, + 102 + ], + "spans": [ + { + "bbox": [ + 282, + 91, + 447, + 102 + ], + "type": "text", + "content": "Physical-Based Event Camera Simulator" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 470, + 91, + 481, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 470, + 91, + 481, + 100 + ], + "spans": [ + { + "bbox": [ + 470, + 91, + 481, + 100 + ], + "type": "text", + "content": "13" + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 12 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 134, + 114, + 480, + 309 + ], + "blocks": [ + { + "bbox": [ + 134, + 114, + 480, + 309 + ], + "lines": [ + { + "bbox": [ + 134, + 114, + 480, + 309 + ], + "spans": [ + { + "bbox": [ + 134, + 114, + 480, + 309 + ], + "type": "image", + "image_path": "6929508043559870f592122e13a4601451cdc25859bd1d1f064583f99db2e244.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 132, + 319, + 482, + 353 + ], + "lines": [ + { + "bbox": [ + 132, + 319, + 482, + 353 + ], + "spans": [ + { + "bbox": [ + 132, + 319, + 482, + 353 + ], + "type": "text", + "content": "Fig. 6: Visual comparisons on event-based video reconstruction. The E2VID network [32] trains on simulated event data from each simulator and tests on the real-world HS-ERGB dataset [45] recorded by a Propheesee Gen4 camera and a RGB camera." + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_caption" + } + ], + "index": 2 + }, + { + "bbox": [ + 130, + 377, + 482, + 486 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 377, + 482, + 486 + ], + "spans": [ + { + "bbox": [ + 130, + 377, + 482, + 486 + ], + "type": "text", + "content": "competitor V2E with an average increase of 0.681 in PSNR, " + }, + { + "bbox": [ + 130, + 377, + 482, + 486 + ], + "type": "inline_equation", + "content": "0.6\\%" + }, + { + "bbox": [ + 130, + 377, + 482, + 486 + ], + "type": "text", + "content": " improvement in SSIM, and a " + }, + { + "bbox": [ + 130, + 377, + 482, + 486 + ], + "type": "inline_equation", + "content": "1.7\\%" + }, + { + "bbox": [ + 130, + 377, + 482, + 486 + ], + "type": "text", + "content": " decrease in LPIPS. Furthermore, we present some visualization results on the real-world HS-ERGB dataset [45] recorded by a Prophesee Gen4 camera and an RGB camera. As shown in Fig. 6, our PECS-trained network produces visually pleasing images with finer details and fewer artifacts. As our PECS is designed from a physical-based perspective, it naturally promotes reconstructed images with natural image statistics. Hence, the utilization of our simulator proves more effective, enabling the learning of high-quality video reconstruction [11,23,27,47,48,50,55] even with limited or no real training samples." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 132, + 506, + 220, + 519 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 506, + 220, + 519 + ], + "spans": [ + { + "bbox": [ + 132, + 506, + 220, + 519 + ], + "type": "text", + "content": "6 Conclusion" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 130, + 533, + 482, + 666 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 533, + 482, + 666 + ], + "spans": [ + { + "bbox": [ + 130, + 533, + 482, + 666 + ], + "type": "text", + "content": "This paper proposes a novel Physical-based Event Camera Simulator (PECS) that can generate a highly realistic event stream by directly interfacing with the 3D scene. To the best of our knowledge, our PECS is the first event camera simulator to design a realistic lens simulation block and a multispectral rendering module. Experiments demonstrate that Our PECS consistently outperforms four state-of-the-art simulators in some scenarios with various motion speeds and light changes, showing that it is currently the most realistic event camera simulator. Besides, integrating our PECS into the Unreal Engine (UE) platform allows the generation of extensive multi-task datasets, showcasing its utility in downstream vision tasks. We also believe that our PECS provides a proof-of-principle tool for the next generation of neuromorphic cameras." + } + ] + } + ], + "index": 6 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "spans": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "type": "text", + "content": "14" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 166, + 91, + 257, + 102 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 166, + 91, + 257, + 102 + ], + "spans": [ + { + "bbox": [ + 166, + 91, + 257, + 102 + ], + "type": "text", + "content": "Haiqian. Author et al." + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 13 + }, + { + "para_blocks": [ + { + "bbox": [ + 133, + 114, + 246, + 129 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 114, + 246, + 129 + ], + "spans": [ + { + "bbox": [ + 133, + 114, + 246, + 129 + ], + "type": "text", + "content": "Acknowledgements" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 132, + 140, + 482, + 165 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 140, + 482, + 165 + ], + "spans": [ + { + "bbox": [ + 132, + 140, + 482, + 165 + ], + "type": "text", + "content": "This work is partially supported by the National Natural Science Foundation of China under Grant 20181310845." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 133, + 185, + 197, + 198 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 185, + 197, + 198 + ], + "spans": [ + { + "bbox": [ + 133, + 185, + 197, + 198 + ], + "type": "text", + "content": "References" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 138, + 211, + 481, + 665 + ], + "type": "list", + "angle": 0, + "index": 22, + "blocks": [ + { + "bbox": [ + 138, + 211, + 481, + 243 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 211, + 481, + 243 + ], + "spans": [ + { + "bbox": [ + 138, + 211, + 481, + 243 + ], + "type": "text", + "content": "1. Annamalai, L., Ramanathan, V., Thakur, C.S.: Event-lstm: An unsupervised and asynchronous learning-based representation for event-based data. IEEE RAL 7(2), 4678-4685 (2022)" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 138, + 245, + 481, + 277 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 245, + 481, + 277 + ], + "spans": [ + { + "bbox": [ + 138, + 245, + 481, + 277 + ], + "type": "text", + "content": "2. Baek, S., Eshraghian, J.K., Thio, W., Sandamirskaya, Y., Iu, H.H., Lu, W.D.: A real-time retinomorphic simulator using a conductance-based discrete neuronal network. In: AICAS. pp. 79-83 (2020)" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 138, + 278, + 481, + 311 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 278, + 481, + 311 + ], + "spans": [ + { + "bbox": [ + 138, + 278, + 481, + 311 + ], + "type": "text", + "content": "3. Bellekens, B., Spruyt, V., Berkvens, R., Penne, R., Weyn, M.: A benchmark survey of rigid 3d point cloud registration algorithms. Int. J. Adv. Intell. Syst 8(5), 118-127 (2015)" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 138, + 312, + 481, + 343 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 312, + 481, + 343 + ], + "spans": [ + { + "bbox": [ + 138, + 312, + 481, + 343 + ], + "type": "text", + "content": "4. Bhattacharya, A., Madaan, R., Cladera, F., Vermprala, S., Bonatti, R., Daniilidis, K., Kapoor, A., Kumar, V., Matni, N., Gupta, J.K.: Evidnerf: Reconstructing event data with dynamic neural radiance fields (2023)" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 138, + 345, + 481, + 366 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 345, + 481, + 366 + ], + "spans": [ + { + "bbox": [ + 138, + 345, + 481, + 366 + ], + "type": "text", + "content": "5. Bi, Y., Andreopoulos, Y.: Pix2nvs: Parameterized conversion of pixel-domain video frames to neuromorphic vision streams. In: ICIP. pp. 1990-1994 (2017)" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 138, + 367, + 481, + 388 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 367, + 481, + 388 + ], + "spans": [ + { + "bbox": [ + 138, + 367, + 481, + 388 + ], + "type": "text", + "content": "6. Bocca, A., Baek, D.: Automated driving systems: Key advantages, limitations and risks. In: AEIT AUTOMOTIVE. pp. 1-6 (2019)" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 138, + 389, + 481, + 422 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 389, + 481, + 422 + ], + "spans": [ + { + "bbox": [ + 138, + 389, + 481, + 422 + ], + "type": "text", + "content": "7. Gallego, G., Delbrück, T., Orchard, G., Bartolozzi, C., Taba, B., Censi, A., Leutenegger, S., Davison, A.J., Conradt, J., Daniilidis, K., et al.: Event-based vision: A survey. IEEE TPAMI 44(1), 154-180 (2020)" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 138, + 422, + 481, + 444 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 422, + 481, + 444 + ], + "spans": [ + { + "bbox": [ + 138, + 422, + 481, + 444 + ], + "type": "text", + "content": "8. Gehrig, D., Gehrig, M., Hidalgo-Carrió, J., Scaramuzza, D.: Video to events: Recyclcng video datasets for event cameras. In: CVPR. pp. 3586-3595 (2020)" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 138, + 445, + 481, + 466 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 445, + 481, + 466 + ], + "spans": [ + { + "bbox": [ + 138, + 445, + 481, + 466 + ], + "type": "text", + "content": "9. Greenspan, M., Yurick, M.: Approximate kd tree search for efficient icp. In: 3DIM. pp. 442-448 (2003)" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 138, + 467, + 481, + 487 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 467, + 481, + 487 + ], + "spans": [ + { + "bbox": [ + 138, + 467, + 481, + 487 + ], + "type": "text", + "content": "0. Gu, D., Li, J., Zhang, Y., Tian, Y.: How to learn a domain-adaptive event simulator? In: ACM MM. pp. 1275-1283 (2021)" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 138, + 488, + 481, + 521 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 488, + 481, + 521 + ], + "spans": [ + { + "bbox": [ + 138, + 488, + 481, + 521 + ], + "type": "text", + "content": "1. He, W., You, K., Qiao, Z., Jia, X., Zhang, Z., Wang, W., Lu, H., Wang, Y., Liao, J.: Timereplayer: Unlocking the potential of event cameras for video interpolation. In: CVPR. pp. 17804-17813 (2022)" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 138, + 522, + 481, + 544 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 522, + 481, + 544 + ], + "spans": [ + { + "bbox": [ + 138, + 522, + 481, + 544 + ], + "type": "text", + "content": "2. Hu, Y., Liu, S.C., Delbruck, T.: v2e: From video frames to realistic dvs events. In: CVPRW. pp. 1312-1321 (2021)" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 138, + 544, + 481, + 566 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 544, + 481, + 566 + ], + "spans": [ + { + "bbox": [ + 138, + 544, + 481, + 566 + ], + "type": "text", + "content": "3. Hwang, I., Kim, J., Kim, Y.M.: Ev-nerf: Event based neural radiance field. In: WACV. pp. 837-847 (2023)" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 138, + 567, + 481, + 599 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 567, + 481, + 599 + ], + "spans": [ + { + "bbox": [ + 138, + 567, + 481, + 599 + ], + "type": "text", + "content": "4. Joubert, D., Marcireau, A., Ralph, N., Jolley, A., van Schaik, A., Cohen, G.: Event camera simulator improvements via characterized parameters. Frontiers in Neuroscience 15, 702765 (2021)" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 138, + 600, + 481, + 620 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 600, + 481, + 620 + ], + "spans": [ + { + "bbox": [ + 138, + 600, + 481, + 620 + ], + "type": "text", + "content": "5. Kang, Z., Li, J., Zhu, L., Tian, Y.: Retinomorphic sensing: A novel paradigm for future multimedia computing. In: ACM MM. pp. 144-152 (2021)" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 138, + 621, + 481, + 643 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 621, + 481, + 643 + ], + "spans": [ + { + "bbox": [ + 138, + 621, + 481, + 643 + ], + "type": "text", + "content": "6. Kolb, C., Mitchell, D., Hanrahan, P.: A realistic camera model for computer graphics. In: CGIT. pp. 317-324 (1995)" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 138, + 643, + 481, + 665 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 643, + 481, + 665 + ], + "spans": [ + { + "bbox": [ + 138, + 643, + 481, + 665 + ], + "type": "text", + "content": "7. Li, D., Li, J., Tian, Y.: Sodformer: Streaming object detection with transformer using events and frames. IEEE TPAMI 45(11), 14020-14037 (2023)" + } + ] + } + ], + "index": 21 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 282, + 91, + 447, + 102 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 282, + 91, + 447, + 102 + ], + "spans": [ + { + "bbox": [ + 282, + 91, + 447, + 102 + ], + "type": "text", + "content": "Physical-Based Event Camera Simulator" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 470, + 92, + 481, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 470, + 92, + 481, + 100 + ], + "spans": [ + { + "bbox": [ + 470, + 92, + 481, + 100 + ], + "type": "text", + "content": "15" + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 14 + }, + { + "para_blocks": [ + { + "bbox": [ + 132, + 116, + 481, + 665 + ], + "type": "list", + "angle": 0, + "index": 24, + "blocks": [ + { + "bbox": [ + 133, + 116, + 481, + 149 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 116, + 481, + 149 + ], + "spans": [ + { + "bbox": [ + 133, + 116, + 481, + 149 + ], + "type": "text", + "content": "18. Li, J., Li, J., Zhu, L., Xiang, X., Huang, T., Tian, Y.: Asynchronous spatiotemporal memory network for continuous event-based object detection. IEEE TIP 31, 2975-2987 (2022)" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 132, + 150, + 481, + 183 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 150, + 481, + 183 + ], + "spans": [ + { + "bbox": [ + 132, + 150, + 481, + 183 + ], + "type": "text", + "content": "19. Li, W., Saeedi, S., McCormac, J., Clark, R., Tzoumanikas, D., Ye, Q., Huang, Y., Tang, R., Leutenegger, S.: Interiornet: Mega-scale multi-sensor photo-realistic indoor scenes dataset. arXiv (2018)" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 132, + 184, + 481, + 205 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 184, + 481, + 205 + ], + "spans": [ + { + "bbox": [ + 132, + 184, + 481, + 205 + ], + "type": "text", + "content": "20. Lin, S., Ma, Y., Guo, Z., Wen, B.: Dvs-voltmeter: Stochastic process-based event simulator for dynamic vision sensors. In: ECCV. pp. 578-593 (2022)" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 132, + 206, + 481, + 228 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 206, + 481, + 228 + ], + "spans": [ + { + "bbox": [ + 132, + 206, + 481, + 228 + ], + "type": "text", + "content": "21. Liu, P., Chen, G., Li, Z., Tang, H., Knoll, A.: Learning local event-based descriptor for patch-based stereo matching. In: ICRA. pp. 412-418 (2022)" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 132, + 229, + 481, + 251 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 229, + 481, + 251 + ], + "spans": [ + { + "bbox": [ + 132, + 229, + 481, + 251 + ], + "type": "text", + "content": "22. Liu, X., Li, J., Fan, X., Tian, Y.: Event-based monocular dense depth estimation with recurrent transformers. In: arXiv (2022)" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 132, + 251, + 481, + 274 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 251, + 481, + 274 + ], + "spans": [ + { + "bbox": [ + 132, + 251, + 481, + 274 + ], + "type": "text", + "content": "23. Mei, H., Wang, Z., Yang, X., Wei, X., Delbruck, T.: Deep polarization reconstruction with pdavis events. In: CVPR. pp. 22149-22158 (2023)" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 132, + 274, + 481, + 306 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 274, + 481, + 306 + ], + "spans": [ + { + "bbox": [ + 132, + 274, + 481, + 306 + ], + "type": "text", + "content": "24. Mou, X., Feng, K., Yi, A., Wang, S., Chen, H., Hu, X., Guo, M., Chen, S., Suess, A.: Accurate event simulation using high-speed videos. Electronic Imaging 34(7), 242-1 (2022)" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 132, + 307, + 481, + 340 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 307, + 481, + 340 + ], + "spans": [ + { + "bbox": [ + 132, + 307, + 481, + 340 + ], + "type": "text", + "content": "25. Mueggler, E., Rebecq, H., Gallego, G., Delbruck, T., Scaramuzza, D.: The event-camera dataset and simulator: Event-based data for pose estimation, visual odometry, and slam. IJRR 36(2), 142-149 (2017)" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 132, + 341, + 481, + 362 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 341, + 481, + 362 + ], + "spans": [ + { + "bbox": [ + 132, + 341, + 481, + 362 + ], + "type": "text", + "content": "26. Muglikar, M., Gehrig, M., Gehrig, D., Scaramuzza, D.: How to calibrate your event camera. In: CVPR. pp. 1403-1409 (2021)" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 132, + 363, + 481, + 396 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 363, + 481, + 396 + ], + "spans": [ + { + "bbox": [ + 132, + 363, + 481, + 396 + ], + "type": "text", + "content": "27. Pan, L., Hartley, R., Scheerlinck, C., Liu, M., Yu, X., Dai, Y.: High frame rate video reconstruction based on an event camera. IEEE TPAMI 44(5), 2519-2533 (2020)" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 132, + 396, + 481, + 430 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 396, + 481, + 430 + ], + "spans": [ + { + "bbox": [ + 132, + 396, + 481, + 430 + ], + "type": "text", + "content": "28. Pantho, M.J.H., Mbongue, J.M., Bhowmik, P., Bobda, C.: Event camera simulator design for modeling attention-based inference architectures. Journal of Real-Time Image Processing 19(2), 363-374 (2022)" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 132, + 430, + 481, + 452 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 430, + 481, + 452 + ], + "spans": [ + { + "bbox": [ + 132, + 430, + 481, + 452 + ], + "type": "text", + "content": "29. Peng, X., Wang, Y., Gao, L., Kneip, L.: Globally-optimal event camera motion estimation. In: ECCV. pp. 51-67 (2020)" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 132, + 452, + 481, + 474 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 452, + 481, + 474 + ], + "spans": [ + { + "bbox": [ + 132, + 452, + 481, + 474 + ], + "type": "text", + "content": "30. Pharr, M., Jakob, W., Humphreys, G.: Physically based rendering: From theory to implementation (2023)" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 132, + 475, + 481, + 496 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 475, + 481, + 496 + ], + "spans": [ + { + "bbox": [ + 132, + 475, + 481, + 496 + ], + "type": "text", + "content": "31. Rebecq, H., Gehrig, D., Scaramuzza, D.: Esim: An open event camera simulator. In: CoRL. pp. 969-982 (2018)" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 132, + 497, + 481, + 519 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 497, + 481, + 519 + ], + "spans": [ + { + "bbox": [ + 132, + 497, + 481, + 519 + ], + "type": "text", + "content": "32. Rebecq, H., Ranftl, R., Koltun, V., Scaramuzza, D.: Events-to-video: Bringing modern computer vision to event cameras. In: CVPR. pp. 3857-3866 (2019)" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 132, + 520, + 481, + 541 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 520, + 481, + 541 + ], + "spans": [ + { + "bbox": [ + 132, + 520, + 481, + 541 + ], + "type": "text", + "content": "33. Rebecq, H., Ranftl, R., Koltun, V., Scaramuzza, D.: High speed and high dynamic range video with an event camera. IEEE TPAMI 43(6), 1964-1980 (2019)" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 132, + 542, + 481, + 563 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 542, + 481, + 563 + ], + "spans": [ + { + "bbox": [ + 132, + 542, + 481, + 563 + ], + "type": "text", + "content": "34. Rizzo, C., Schuman, C., Plank, J.: Event-based camera simulation wrapper for arcade learning environment. In: ICNS. pp. 1-5 (2022)" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 132, + 564, + 481, + 586 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 564, + 481, + 586 + ], + "spans": [ + { + "bbox": [ + 132, + 564, + 481, + 586 + ], + "type": "text", + "content": "35. Rogalski, A., Razeghi, M.: Narrow-gap semiconductor photodiodes. In: Photodetectors: Materials and Devices III. vol. 3287, pp. 2-13 (1998)" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 132, + 586, + 451, + 597 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 586, + 451, + 597 + ], + "spans": [ + { + "bbox": [ + 132, + 586, + 451, + 597 + ], + "type": "text", + "content": "36. Sasián, J.: Introduction to lens design. Cambridge University Press (2019)" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 132, + 598, + 481, + 620 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 598, + 481, + 620 + ], + "spans": [ + { + "bbox": [ + 132, + 598, + 481, + 620 + ], + "type": "text", + "content": "37. Segal, A., Haehnel, D., Thrun, S.: Generalized-icp. In: RSS. vol. 2, p. 435. Seattle, WA (2009)" + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 132, + 620, + 481, + 643 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 620, + 481, + 643 + ], + "spans": [ + { + "bbox": [ + 132, + 620, + 481, + 643 + ], + "type": "text", + "content": "38. Sengupta, J., Liu, S., Andreou, A.: Retinosim: an event-based data synthesis tool for neuromorphic vision architecture exploration. In: ICNS. pp. 1-9 (2022)" + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 132, + 643, + 481, + 665 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 643, + 481, + 665 + ], + "spans": [ + { + "bbox": [ + 132, + 643, + 481, + 665 + ], + "type": "text", + "content": "39. Su, Q., Chou, Y., Hu, Y., Li, J., Mei, S., Zhang, Z., Li, G.: Deep directly-trained spiking neural networks for object detection. In: ICCV. pp. 6555-6565 (2023)" + } + ] + } + ], + "index": 23 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "spans": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "type": "text", + "content": "16" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 166, + 91, + 257, + 101 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 166, + 91, + 257, + 101 + ], + "spans": [ + { + "bbox": [ + 166, + 91, + 257, + 101 + ], + "type": "text", + "content": "Haiqian. Author et al." + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 15 + }, + { + "para_blocks": [ + { + "bbox": [ + 130, + 116, + 482, + 610 + ], + "type": "list", + "angle": 0, + "index": 20, + "blocks": [ + { + "bbox": [ + 130, + 116, + 482, + 149 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 116, + 482, + 149 + ], + "spans": [ + { + "bbox": [ + 130, + 116, + 482, + 149 + ], + "type": "text", + "content": "40. Sun, L., Sakaridis, C., Liang, J., Jiang, Q., Yang, K., Sun, P., Ye, Y., Wang, K., Gool, L.V.: Event-based fusion for motion deblurring with cross-modal attention. In: ECCV. pp. 412-428 (2022)" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 130, + 150, + 482, + 182 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 150, + 482, + 182 + ], + "spans": [ + { + "bbox": [ + 130, + 150, + 482, + 182 + ], + "type": "text", + "content": "41. Sundar, V., Ardelean, A., Swedish, T., Bruschini, C., Charbon, E., Gupta, M.: Sodacam: Software-defined cameras via single-photon imaging. In: ICCV. pp. 8165-8176 (2023)" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 130, + 182, + 482, + 215 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 182, + 482, + 215 + ], + "spans": [ + { + "bbox": [ + 130, + 182, + 482, + 215 + ], + "type": "text", + "content": "42. Ta, K., Bruggemann, D., Brodermann, T., Sakaridis, C., Van Gool, L.: L2e: Lasers to events for 6-dof extrinsic calibration of lidars and event cameras. In: ICRA. pp. 11425-11431 (2023)" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 130, + 215, + 482, + 248 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 215, + 482, + 248 + ], + "spans": [ + { + "bbox": [ + 130, + 215, + 482, + 248 + ], + "type": "text", + "content": "43. Taverni, G., Moeys, D.P., Li, C., Cavaco, C., Motsnyi, V., Bello, D.S.S., Delbruck, T.: Front and back illuminated dynamic and active pixel vision sensors comparison. IEEE Trans. Circuits Syst. II: Exp. Briefs 65(5), 677-681 (2018)" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 130, + 248, + 482, + 270 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 248, + 482, + 270 + ], + "spans": [ + { + "bbox": [ + 130, + 248, + 482, + 270 + ], + "type": "text", + "content": "44. Tsuji, Y., Yatagawa, T., Kubo, H., Morishima, S.: Event-based camera simulation using monte carlo path tracing with adaptive denoising. arXiv (2023)" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 130, + 270, + 482, + 303 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 270, + 482, + 303 + ], + "spans": [ + { + "bbox": [ + 130, + 270, + 482, + 303 + ], + "type": "text", + "content": "45. Tulyakov, S., Bochicchio, A., Gehrig, D., Georgoulis, S., Li, Y., Scaramuzza, D.: Time lens++: Event-based frame interpolation with parametric non-linear flow and multi-scale fusion. In: CVPR. pp. 17755-17764 (2022)" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 130, + 303, + 482, + 325 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 303, + 482, + 325 + ], + "spans": [ + { + "bbox": [ + 130, + 303, + 482, + 325 + ], + "type": "text", + "content": "46. Wang, Z., Bovik, A.C., Sheikh, H.R., Simoncelli, E.P.: Image quality assessment: From error visibility to structural similarity. IEEE TIP 13(4), 600-612 (2004)" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 130, + 325, + 482, + 347 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 325, + 482, + 347 + ], + "spans": [ + { + "bbox": [ + 130, + 325, + 482, + 347 + ], + "type": "text", + "content": "47. Weng, W., Zhang, Y., Xiong, Z.: Event-based video reconstruction using transformer. In: ICCV. pp. 2563-2572 (2021)" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 130, + 347, + 482, + 369 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 347, + 482, + 369 + ], + "spans": [ + { + "bbox": [ + 130, + 347, + 482, + 369 + ], + "type": "text", + "content": "48. Yu, L., Zhang, X., Liao, W., Yang, W., Xia, G.S.: Learning to see through with events. IEEE TPAMI 45(07), 8660-8678 (2023)" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 130, + 369, + 482, + 391 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 369, + 482, + 391 + ], + "spans": [ + { + "bbox": [ + 130, + 369, + 482, + 391 + ], + "type": "text", + "content": "49. Yu, W., Li, J., Zhang, S., Ji, X.: Learning scale-aware spatio-temporal implicit representation for event-based motion deblurring. In: ICML (2024)" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 130, + 391, + 482, + 423 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 391, + 482, + 423 + ], + "spans": [ + { + "bbox": [ + 130, + 391, + 482, + 423 + ], + "type": "text", + "content": "50. Zhang, H., Zhang, L., Dai, Y., Li, H., Koniusz, P.: Event-guided multi-patch network with self-supervision for non-uniform motion deblurring. IJCV 131(2), 453-470 (2023)" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 130, + 423, + 482, + 445 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 423, + 482, + 445 + ], + "spans": [ + { + "bbox": [ + 130, + 423, + 482, + 445 + ], + "type": "text", + "content": "51. Zhang, R., Isola, P., Efros, A.A., Shechtman, E., Wang, O.: The unreasonable effectiveness of deep features as a perceptual metric. In: CVPR. pp. 586-595 (2018)" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 130, + 446, + 482, + 468 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 446, + 482, + 468 + ], + "spans": [ + { + "bbox": [ + 130, + 446, + 482, + 468 + ], + "type": "text", + "content": "52. Zhang, Z., Cui, S., Chai, K., Yu, H., Dasgupta, S., Mahbub, U., Rahman, T.: V2ce: Video to continuous events simulator. arXiv (2023)" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 130, + 468, + 482, + 500 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 468, + 482, + 500 + ], + "spans": [ + { + "bbox": [ + 130, + 468, + 482, + 500 + ], + "type": "text", + "content": "53. Zheng, X., Liu, Y., Lu, Y., Hua, T., Pan, T., Zhang, W., Tao, D., Wang, L.: Deep learning for event-based vision: A comprehensive survey and benchmarks. arXiv (2023)" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 130, + 500, + 482, + 522 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 500, + 482, + 522 + ], + "spans": [ + { + "bbox": [ + 130, + 500, + 482, + 522 + ], + "type": "text", + "content": "54. Zhu, A.Z., Wang, Z., Khant, K., Daniilidis, K.: Eventgan: Leveraging large scale image datasets for event cameras. In: ICCP. pp. 1-11 (2021)" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 130, + 522, + 482, + 555 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 522, + 482, + 555 + ], + "spans": [ + { + "bbox": [ + 130, + 522, + 482, + 555 + ], + "type": "text", + "content": "55. Zhu, L., Li, J., Wang, X., Huang, T., Tian, Y.: Neuspike-net: High speed video reconstruction via bio-inspired neuromorphic cameras. In: ICCV. pp. 2400-2409 (2021)" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 130, + 555, + 482, + 587 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 555, + 482, + 587 + ], + "spans": [ + { + "bbox": [ + 130, + 555, + 482, + 587 + ], + "type": "text", + "content": "56. Zhu, L., Wang, X., Chang, Y., Li, J., Huang, T., Tian, Y.: Event-based video reconstruction via potential-assisted spiking neural network. In: CVPR. pp. 3594-3604 (2022)" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 130, + 587, + 482, + 610 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 587, + 482, + 610 + ], + "spans": [ + { + "bbox": [ + 130, + 587, + 482, + 610 + ], + "type": "text", + "content": "57. Ziegler, A., Teigland, D., Tebbe, J., Gossard, T., Zell, A.: Real-time event simulation with frame-based cameras. In: ICRA. pp. 11669-11675 (2023)" + } + ] + } + ], + "index": 19 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 282, + 91, + 447, + 101 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 282, + 91, + 447, + 101 + ], + "spans": [ + { + "bbox": [ + 282, + 91, + 447, + 101 + ], + "type": "text", + "content": "Physical-Based Event Camera Simulator" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 470, + 92, + 480, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 470, + 92, + 480, + 100 + ], + "spans": [ + { + "bbox": [ + 470, + 92, + 480, + 100 + ], + "type": "text", + "content": "17" + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 16 + } + ], + "_backend": "vlm", + "_version_name": "2.6.4" +} \ No newline at end of file diff --git a/2024/Physically Plausible Color Correction for Neural Radiance Fields/88cfdd54-68fa-4f55-bb61-e981b6d01cd5_content_list.json b/2024/Physically Plausible Color Correction for Neural Radiance Fields/88cfdd54-68fa-4f55-bb61-e981b6d01cd5_content_list.json new file mode 100644 index 0000000000000000000000000000000000000000..c1ef4b02f82895d43eec768496cde5e1d6f75a54 --- /dev/null +++ b/2024/Physically Plausible Color Correction for Neural Radiance Fields/88cfdd54-68fa-4f55-bb61-e981b6d01cd5_content_list.json @@ -0,0 +1,1825 @@ +[ + { + "type": "text", + "text": "Physically Plausible Color Correction for Neural Radiance Fields", + "text_level": 1, + "bbox": [ + 223, + 140, + 779, + 186 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Qi Zhang $^{1,\\dagger}$ , Ying Feng $^{1,\\dagger}$ , and Hongdong Li $^{2}$", + "bbox": [ + 318, + 210, + 684, + 228 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "$^{1}$ Independent Researcher {nwpuqzhang, yfeng.von}@gmail.com", + "bbox": [ + 282, + 239, + 720, + 253 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "2 Australian National University hongdong.li@anu.edu.au", + "bbox": [ + 302, + 253, + 699, + 267 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Abstract. Neural Radiance Fields have become the representation of choice for many 3D computer vision and computer graphics applications, e.g., novel view synthesis and 3D reconstruction. Multi-camera systems are commonly used as the image capture setup in NeRF-based multi-view tasks such as dynamic scene acquisition or realistic avatar animation. However, a critical issue that has often been overlooked in this setup is the evident differences in color responses among multiple cameras, which adversely affect the NeRF reconstruction performance. These color discrepancies among multiple input images stem from two aspects: 1) implicit properties of the scenes such as reflections and shadings, and 2) external differences in camera settings and lighting conditions. In this paper, we address this problem by proposing a novel color correction module that simulates the physical color processing in cameras to be embedded in NeRF, enabling the unified color NeRF reconstruction. Besides the view-independent color correction module for external differences, we predict a view-dependent function to minimize the color residual (including, e.g., specular and shading) to eliminate the impact of inherent attributes. We further describe how the method can be extended with a reference image as guidance to achieve aesthetically plausible color consistency and color translation on novel views. Experiments validate that our method is superior to baseline methods in both quantitative and qualitative evaluations of color correction and color consistency.", + "bbox": [ + 261, + 303, + 743, + 623 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Keywords: Neural Radiance Field $\\cdot$ View Synthesis $\\cdot$ Color Translation", + "bbox": [ + 261, + 635, + 740, + 648 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "1 Introduction", + "text_level": 1, + "bbox": [ + 215, + 696, + 375, + 713 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "The synthesis and rendering of novel views using multi-view images has been a long-standing problem in computer vision and graphics, which is also a cornerstone task for augmented reality (AR) and virtual reality (VR) applications. The recent success of Neural Radiance Fields (NeRF) [49] has made the implicit neural representation a powerful approach for novel view synthesis tasks, as", + "bbox": [ + 212, + 727, + 787, + 803 + ], + "page_idx": 0 + }, + { + "type": "page_footnote", + "text": "† These authors contributed equally to this work.", + "bbox": [ + 230, + 811, + 562, + 825 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/fb49fe45339def8d81f97800bf122c05e3117ff359c3d5dc37817359d99b0369.jpg", + "image_caption": [ + "(a) input views", + "Fig. 1: Physically Plausible Color Correction for Neural Radiance Fields. (a) Given a set of images captured under various camera settings and lighting conditions, our method introduces a color correction module into the NeRF-based model, which enables recovering (b) a unified color radiance field with a physically plausible white balance. Unlike prior methods that embed color through features, our color correction module simulates color differences via a nonlinear color translation matrix. Additionally, our method can render (c) novel views translated from the color of a reference image." + ], + "image_footnote": [], + "bbox": [ + 222, + 148, + 421, + 277 + ], + "page_idx": 1 + }, + { + "type": "image", + "img_path": "images/4cc89a1e3946176c2dafe27401ccaa006ca41ef98f98cbffe0ed4f1d7fa52e4c.jpg", + "image_caption": [ + "(b) unified color radiance field" + ], + "image_footnote": [], + "bbox": [ + 428, + 150, + 643, + 276 + ], + "page_idx": 1 + }, + { + "type": "image", + "img_path": "images/ed811e9f51221c5b7e2d444c4f336950faf915831c729ce96247cd8d1550d6c7.jpg", + "image_caption": [ + "(c) color translation" + ], + "image_footnote": [], + "bbox": [ + 651, + 150, + 769, + 273 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "an alternative to traditional image-based reconstruction techniques. The NeRF model employs Multi-Layer Perceptrons (MLPs) to parameterize the mapping from spatial coordinates to corresponding densities and view-dependent colors. This approach has been proven effective with scenes that are consistently well-lit without substantial variations in image brightness.", + "bbox": [ + 212, + 429, + 787, + 505 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "With the increasing prevalence of multi-camera systems [18, 22, 31, 34, 53, 59, 73], NeRF and its variants [4, 5, 50] face challenges in achieving photorealistic novel view synthesis with color consistency. The differences in color response among multiple cameras can significantly impact the accuracy and consistency of synthesized views. Even when capturing with identical camera settings, different types of cameras may present distinct color responses caused by the difference in sensors and ISP algorithms. To address this limitation, NeRF-W [46] and Ha-NeRF [15] incorporate additional latent codes to encode appearances into the NeRF model for consistent view synthesis and appearance control. However, these methods tend to embed the overall appearances as view-dependent features, making it difficult to achieve fine-grained and intuitive color correction. HDR-NeRF [28] attempts to address this issue but primarily focuses on building a physical module for camera exposures. It is necessary to introduce color translation for NeRF-based multi-view tasks, which should involve modeling color correlations across cameras and enabling physically plausible color correction and consistency in NeRF reconstructions.", + "bbox": [ + 212, + 506, + 787, + 748 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "The color difference across multi-view images arises from two aspects: inherent view-dependent scene properties (e.g. reflections and shadings); external differences in camera settings (e.g. white balance, ISO, exposure) and lighting conditions. Naturally, we could calibrate each image via color correction methods [1,3,6,36,37], and then input to NeRF-based method for novel view synthesis. However, these image-based methods struggle to preserve color consistency in", + "bbox": [ + 212, + 750, + 789, + 840 + ], + "page_idx": 1 + }, + { + "type": "page_number", + "text": "2", + "bbox": [ + 217, + 114, + 228, + 126 + ], + "page_idx": 1 + }, + { + "type": "header", + "text": "Q. Zhang, Y. Feng et al.", + "bbox": [ + 271, + 114, + 436, + 128 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "NeRF reconstructions. In addition to faithfully reproducing the captured colors, we consider that color correction process should allow for color adjustment to enhance visual appeal, and even color translation to a given reference image. By incorporating such color correction module, users, including novice ones, can easily recover NeRF and adjust the colors of novel views based on their preferences.", + "bbox": [ + 212, + 146, + 782, + 233 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "In this paper, we propose a novel framework that incorporates a color correction module into the NeRF-based model to achieve unified color radiance field reconstruction from multi-view images with different camera settings and lighting conditions, as shown in Fig. 1. In contrast to the embedding feature for overall color differences in prior NeRF-based methods [15,46], we present a color correction module to predict a nonlinear color translation matrix that simulates the color processing in cameras, such as white balance, nonlinear color enhancement, tone-manipulation, gamma encoding, etc. Additionally, considering the view-dependent factors of inherent properties, we predict a view-dependent function to separate the color residual (e.g. specular and shading of the scene) to improve the accuracy of the color translation matrix. Our technical contributions are summarized as follows:", + "bbox": [ + 212, + 237, + 787, + 417 + ], + "page_idx": 2 + }, + { + "type": "list", + "sub_type": "text", + "list_items": [ + "- We present a novel framework for color correction and translation in NeRF reconstruction, allowing even novice users to perform novel view synthesis under unstable camera settings in an intuitive manner.", + "- We propose a color correction module that defines the non-linear color translation to be embedded in NeRF.", + "- We introduce tailored regularizations to eliminate the impact of inherent scene properties, constrain the scene's NeRF to a white balance color space, and maintain the entire color distribution." + ], + "bbox": [ + 225, + 428, + 782, + 544 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Extensive experiments have shown that our method outperforms baseline methods both quantitatively and qualitatively in terms of color correction.", + "bbox": [ + 212, + 558, + 782, + 588 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "2 Related Works", + "text_level": 1, + "bbox": [ + 214, + 611, + 395, + 628 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Neural Radiance Fields Neural Radiance Fields represent a scene implicitly with multi-layer perceptrons, approximating a continuous function that maps each 3D position and direction to a density and radiance [49]. By incorporating view direction features, NeRF can effectively model view-dependent effects such as reflection and refraction. This allows for photo-realistic synthesis of novel views with volumetric rendering. To alleviate the computational limitation of massive MLP evaluations while maintaining the high-frequency appearances, many extensive works have been proposed to speed up the training and inference of the model [5,13,17,24-26,35,50,54,64,65,74-76] with various geometry proxies and neural features, explore view synthesis for dynamic scenes [12,21,23,43,57], and enhance the generalization of NeRF [14,19,29,33,40,44,63,68]. Many methods [30,41,61] have been proposed to exploit the view-dependent appearance to decouple the physical properties of the scene. NeRF2Mesh [61] proposes a", + "bbox": [ + 212, + 643, + 787, + 840 + ], + "page_idx": 2 + }, + { + "type": "header", + "text": "Color Correction for NeRF", + "bbox": [ + 549, + 114, + 730, + 127 + ], + "page_idx": 2 + }, + { + "type": "page_number", + "text": "3", + "bbox": [ + 774, + 116, + 784, + 126 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "two-stage framework to extract the mesh and texture images by regularizing the view-dependent color. Many inverse rendering methods go a step further in material decoupling [9, 10, 32, 58, 62, 66, 69, 71, 77]. They decompose the radiance into physical properties such as albedo, roughness, metallicity, and lighting by introducing physical-based rendering into NeRF optimization, and enable relighting under novel lighting conditions or material editing. Numerous methods have been proposed to reconstruct scenes from images captured in uncontrolled environments, addressing challenges such as inaccurate poses [7, 16, 42, 51], image blur or noise [45, 48, 52], and inconsistent camera settings [27, 28, 55, 60].", + "bbox": [ + 212, + 146, + 787, + 282 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Appearance control of NeRF Images captured in natural environments or taken by multiple cameras often exhibit inconsistency in brightness and color. These image-dependent appearances may derive from changing lighting conditions, exposure variations and white balancing, or differences in camera sensors and ISP, which are not desired in novel view synthesis. HDR-NeRF [28] focuses on multi-view images of varying exposures. By modeling the simplified physical imaging process with a tone mapper, which maps the radiance of a scene point to a pixel value, HDR-NeRF reconstructs the high dynamic range radiance field, which can generate novel views with consistent exposures. Many methods enable NeRF to handle these inconsistencies by introducing related parameters or embeddings. NeRF-w [46] and Ha-NeRF [15] propose to encode the image-dependent appearance using latent codes, which can then be manipulated to control the overall style of rendered images. Block-NeRF [60] feeds an additional exposure level to the appearance prediction model to deal with the brightness variance and provides the ability to modify the global exposure conditioning during inference. URF [55] predicts a $3 \\times 3$ matrix to restrictively model the white balance and exposure variations. One shortcoming of these methods is that view-dependent appearances, those either caused by changes in camera settings and environment lighting or are intrinsical from scene physical properties, are coupled together and modeled related to both latent codes and view directions. In addition to the aforementioned methods that could control scene appearances by modifying materials or latent codes, some methods [38,39] engage in appearance editing by color post-manipulation and rely heavily on the consistency of views.", + "bbox": [ + 212, + 282, + 787, + 630 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Color Correction Color correction has been extensively studied on single 2D image and multi-camera systems. Single 2D image color correction [1,3,6,36, 37], or white balance correction, cannot guarantee view-consistent results when naively performed on each view. On the other hand, multi-camera color correction [20,56,72] relies on common area or statistical matching, which cannot handle inherent view variance and only matches the global style. To address these limitations, we propose integrating a color correction module into the NeRF optimization process to enable view-consistent color correction with 3D awareness.", + "bbox": [ + 212, + 630, + 787, + 750 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "3 Color Correction for NeRF", + "text_level": 1, + "bbox": [ + 215, + 775, + 516, + 792 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Given multi-view images captured under various camera settings and lighting conditions, our method proposes a color correction module embedded in a NeRF-", + "bbox": [ + 212, + 809, + 785, + 839 + ], + "page_idx": 3 + }, + { + "type": "page_number", + "text": "4", + "bbox": [ + 217, + 114, + 230, + 126 + ], + "page_idx": 3 + }, + { + "type": "header", + "text": "Q. Zhang, Y. Feng et al.", + "bbox": [ + 271, + 114, + 436, + 128 + ], + "page_idx": 3 + }, + { + "type": "image", + "img_path": "images/edbb41236e85ac1f7511e153e8e0fbe243a8187b6e4f0dd02aafb1fdc248f6f7.jpg", + "image_caption": [ + "Fig. 2: Overview of our pipeline. Given multi-view images captured under various camera settings and lighting conditions, our method recovers a unified color radiance field, which enables consistent novel view synthesis. We propose a non-linear color correction module embedded in NeRF-based model to simulate the color processing performed in cameras (e.g. white balance, exposure, nonlinear color enhancement, etc.), which maps the color from unified scene color space to each camera color space. The view-dependent color arises from inherent scene properties like specular and shading is extracted under regularization. We design a robust optimization scheme with tailored losses to achieve physically plausible color correction." + ], + "image_footnote": [], + "bbox": [ + 218, + 148, + 782, + 284 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "based model, enabling the unified color NeRF reconstruction, as shown in Fig. 2. The scene radiance field is decomposed into two components to decouple the inherent color variances like specular and shading. The color correction module is designed as a $3 \\times 7$ translation matrix in nonlinear RGB space. Multiple regularizations on pixels and areas are incorporated to improve color consistency.", + "bbox": [ + 212, + 452, + 787, + 530 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "3.1 Scene Representation", + "text_level": 1, + "bbox": [ + 215, + 550, + 439, + 566 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Neural Radiance Field (NeRF) represents a scene using an implicit neural function, which utilizes MLPs to map the position $\\mathbf{x}$ and view direction $\\mathbf{d}$ to a density $\\sigma(\\mathbf{x})$ and view-dependent color $\\mathbf{c}(\\mathbf{x},\\mathbf{d})$ . To render the color of a pixel in an image from a given camera pose, NeRF casts a ray $\\mathbf{r}(t) = \\mathbf{o} + t\\mathbf{d}$ through the pixel, where $\\mathbf{o}$ and $\\mathbf{d}$ are the origin and direction of the ray, respectively, and $t$ is the parameter that determines the position along the ray. The predicted $\\{(\\sigma,\\mathbf{c})\\}$ of sampled positions on the ray are then accumulated into the color $C(\\mathbf{r})$ of the pixel via volume rendering [47]:", + "bbox": [ + 212, + 575, + 787, + 700 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\nC (\\mathbf {r}) = \\int_ {t _ {n}} ^ {t _ {f}} T (t) \\sigma (\\mathbf {r} (t)) \\mathbf {c} (\\mathbf {r} (t), \\mathbf {d}) d t, \\tag {1}\n$$\n", + "text_format": "latex", + "bbox": [ + 369, + 709, + 785, + 747 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\nT (t) = \\exp \\left(- \\int_ {t _ {n}} ^ {t} \\sigma (\\mathbf {r} (p)) d p\\right), \\tag {2}\n$$\n", + "text_format": "latex", + "bbox": [ + 383, + 753, + 785, + 790 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "where $t_n$ and $t_f$ indicate the near and far boundary of the ray, respectively. $T(t)$ refers to an accumulated transmittance, which quantifies the probability of the ray traveling from $t_n$ to $t$ without encountering other particles.", + "bbox": [ + 212, + 794, + 787, + 840 + ], + "page_idx": 4 + }, + { + "type": "header", + "text": "Color Correction for NeRF", + "bbox": [ + 550, + 114, + 730, + 126 + ], + "page_idx": 4 + }, + { + "type": "page_number", + "text": "5", + "bbox": [ + 774, + 116, + 784, + 126 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "We build our model on the NeRF framework, which is widely used for modeling scenes captured under static camera settings and lighting conditions. To eliminate the influence of inherent color variances, we propose to decompose the appearance into two components: view-independent diffuse color $\\mathbf{c}_d$ and view-dependent color $\\mathbf{c}_s$ . The final color is obtained by summing the two terms:", + "bbox": [ + 212, + 146, + 787, + 223 + ], + "page_idx": 5 + }, + { + "type": "equation", + "text": "\n$$\n\\mathbf {c} = \\mathbf {c} _ {d} (\\mathbf {x}) + \\mathbf {c} _ {s} (\\mathbf {x}, \\mathbf {d}). \\tag {3}\n$$\n", + "text_format": "latex", + "bbox": [ + 426, + 229, + 785, + 244 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "By explicitly considering these two components, we can better model the appearance changes arising from surface properties and these caused by varying lighting conditions and camera settings.", + "bbox": [ + 214, + 250, + 787, + 297 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "3.2 Color Correction Module", + "text_level": 1, + "bbox": [ + 214, + 311, + 472, + 327 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Digital cameras perform a number of processing steps to transform raw images to standard RGB images, including demosaicing, white balance, color correction and etc. A generic model [8] of this pipeline is formulated as:", + "bbox": [ + 212, + 332, + 787, + 378 + ], + "page_idx": 5 + }, + { + "type": "equation", + "text": "\n$$\n\\mathbf {I} _ {s r g b} = f \\left(\\alpha \\mathbf {A} _ {x y z} \\mathbf {D} _ {w b} \\mathbf {I} _ {r a w}\\right), \\tag {4}\n$$\n", + "text_format": "latex", + "bbox": [ + 406, + 385, + 785, + 401 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "where $\\alpha$ is an exposure compensation common gain, $\\mathbf{D}_{wb}$ represents the $3\\times 3$ diagonal matrix for white balance, $\\mathbf{A}_{xyz}$ is a $3\\times 3$ linear transform of color space, and $f(\\cdot)$ is a nonlinear function compounding color enhancement, tonemanipulation, and a final sRGB gamma encoding, etc. It's worth noting that the differences in white balance and lighting conditions become non-linear in the sRGB color space after undergoing these mappings.", + "bbox": [ + 212, + 406, + 787, + 496 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "In our framework, we expect to recover the scene's radiance field with unified color when the input sRGB images exhibit color variances. Inspired by [2], our method introduces a polynomial color-mapping function to model the color deviation between the scene color space $\\mathbf{c}$ and each camera color space $\\mathbf{c}^{cam}$ captured by different cameras, which can be written as:", + "bbox": [ + 212, + 497, + 787, + 571 + ], + "page_idx": 5 + }, + { + "type": "equation", + "text": "\n$$\n\\mathbf {c} ^ {c a m} = \\mathbf {M} \\varPhi (\\mathbf {c}), \\tag {5}\n$$\n", + "text_format": "latex", + "bbox": [ + 446, + 578, + 785, + 594 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "where $\\varPhi$ is the kernel mapping function that transforms the input color into the nonlinear space, specifically, $\\varPhi([R,G,B]^T) \\to [R,G,B,RG,RB,GB,1]^T$ . The matrix $\\mathbf{M}$ is a color translation matrix of shape $3 \\times 7$ . Notably, we utilize an Embedding network $E$ to predict a matrix $\\mathbf{M}$ for each camera and lighting condition, and based on the physical imaging process, the matrix is shared by the diffuse and view-dependent components. The pixel colors of sampled points in the scene color space are first mapped to camera color space, and then accumulated through volume rendering. Combining Eq. (1), Eq. (3) and Eq. (5), the rendering equation can be written as:", + "bbox": [ + 212, + 599, + 787, + 736 + ], + "page_idx": 5 + }, + { + "type": "equation", + "text": "\n$$\nC ^ {c a m} = \\int_ {t _ {n}} ^ {t _ {f}} T (t) \\sigma (\\mathbf {r} (t)) \\left(\\mathbf {M} \\varPhi \\left(\\mathbf {c} _ {d} (\\mathbf {r} (t))\\right) + \\mathbf {M} \\varPhi \\left(\\mathbf {c} _ {s} (\\mathbf {r} (t), \\mathbf {d})\\right)\\right) d t \\tag {6}\n$$\n", + "text_format": "latex", + "bbox": [ + 290, + 741, + 785, + 775 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "where $T(t)$ is defined in Eq. (2). By integrating the color correction module in the rendering pipeline, we can reconstruct the radiance field of the scene while accounting for color variations caused by different cameras and lighting conditions.", + "bbox": [ + 212, + 779, + 787, + 839 + ], + "page_idx": 5 + }, + { + "type": "page_number", + "text": "6", + "bbox": [ + 217, + 114, + 228, + 126 + ], + "page_idx": 5 + }, + { + "type": "header", + "text": "Q. Zhang, Y. Feng et al.", + "bbox": [ + 271, + 114, + 434, + 128 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "3.3 View-consistent Color Distribution", + "text_level": 1, + "bbox": [ + 215, + 146, + 550, + 160 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Considering the color correction matrix is designed to be camera-wise, relying solely on pixel-level color loss hinders the accuracy and efficiency of optimization. Therefore we employ a cross-camera color distribution constraint. Given that digital cameras capture images with three color channels, each with a discrete range of $K$ intensity values, we analyze the intensity distribution in each channel, which is formulated as an intensity histogram to constrain the color consistency.", + "bbox": [ + 212, + 185, + 787, + 276 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Drawing inspiration from the cross-camera color translation task, we transform the RGB color space $C^{cam}$ into YUV color space $Y^{cam} \\in [-1,1]^3$ . The YUV color space comprises one luma component (Y) and two chrominance components, U (blue projection) and V (red projection). We analyze each color channel independently and obtain the intensity histogram by counting the number of pixels of each intensity value. To obtain the intensity histogram as a smooth and differentiable function, we formulate one histogram distribution $\\mathbf{h}(Y^{cam})$ of an image color channel as", + "bbox": [ + 212, + 277, + 789, + 398 + ], + "page_idx": 6 + }, + { + "type": "equation", + "text": "\n$$\n\\mathbf {h} = \\left\\{\\frac {1}{N} \\sum_ {x \\in Y} \\Theta \\left(\\frac {I - \\mu_ {k} + L / 2}{W}\\right) - \\Theta \\left(\\frac {I - \\mu_ {k} - L / 2}{W}\\right) \\right\\} _ {k = 0} ^ {K - 1}, \\tag {7}\n$$\n", + "text_format": "latex", + "bbox": [ + 297, + 416, + 785, + 460 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "where $K$ is the number of bins set in the histogram, and $L$ indicates the length of each bin, which also equals $\\frac{2}{K}$ . Given the $k$ -th bin, its center and bandwidth refer to $\\mu_k = -1 + L(k + \\frac{1}{2})$ and $W = L / 2.5$ , respectively. $\\Theta(x)$ is a Sigmoid function $(1 + \\exp(-x))^{-1}$ defining a differentiable approximation of the histogram distribution.", + "bbox": [ + 214, + 474, + 787, + 551 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Histogram Distance. To incorporate color distribution constraints into our framework and facilitate easy convergence, we propose using the Euclidean distance between the cumulative histograms to measure the distance between two image histograms. Let $\\mathbf{h}_s$ and $\\mathbf{h}_r$ represent the histograms of the rendered and the reference images, respectively. The distance between histograms is defined:", + "bbox": [ + 214, + 554, + 787, + 630 + ], + "page_idx": 6 + }, + { + "type": "equation", + "text": "\n$$\n\\mathcal {D} _ {h} \\left(\\mathbf {h} _ {s}, \\mathbf {h} _ {r}\\right) = \\sum_ {i = 1} ^ {K} \\left(\\mathrm {C D F} _ {i} \\left(\\mathbf {h} _ {s}\\right), \\mathrm {C D F} _ {i} \\left(\\mathbf {h} _ {r}\\right)\\right), \\tag {8}\n$$\n", + "text_format": "latex", + "bbox": [ + 354, + 647, + 785, + 686 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "where $\\mathrm{CDF}_i(\\mathbf{h})$ denotes the $i$ -th element of the cumulative density function of histogram. The histogram distance $\\mathcal{D}_h$ measures the global similarity of color distribution between two images. Unlike NeRF, which focuses on individual pixel values, this approach takes into account the entire color distribution of the images. By considering the cumulative histograms, we can effectively compare and align the color distribution of two different images. This property is particularly useful for color translation tasks with our color correction module. We can achieve consistent color appearance across different images by adapting the color distribution towards the reference image.", + "bbox": [ + 214, + 703, + 787, + 840 + ], + "page_idx": 6 + }, + { + "type": "header", + "text": "Color Correction for NeRF", + "bbox": [ + 549, + 114, + 730, + 127 + ], + "page_idx": 6 + }, + { + "type": "page_number", + "text": "7", + "bbox": [ + 774, + 114, + 785, + 126 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "3.4 Optimization", + "text_level": 1, + "bbox": [ + 215, + 146, + 372, + 161 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "To regularize the reconstructed scene color and reduce ambiguity, we design a series of losses on both pixel and image levels, including color reconstruction loss, specular loss, histogram loss, and gray-world loss.", + "bbox": [ + 212, + 169, + 782, + 213 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Color Reconstruction Loss. To jointly optimize the scene's NeRF and the camera-dependent color correction matrix $\\{\\mathbf{M}_i\\}_{i=1}^N$ , color reconstruction loss is proposed to minimize the mean squared error (MSE) of sampled pixels between the ground truth color $C^{cam}$ and the rendered color $\\hat{C}^{cam}$ which is mapped from scene color by corresponding matrix (see Eq. (5) and Eq. (6)). The color reconstruction loss is formulated as:", + "bbox": [ + 212, + 214, + 784, + 305 + ], + "page_idx": 7 + }, + { + "type": "equation", + "text": "\n$$\n\\mathcal {L} _ {c} = \\sum_ {\\mathbf {r} \\in \\mathcal {R} (\\mathbf {P})} \\left\\| \\widehat {C} ^ {c a m} - C ^ {c a m} \\right\\| _ {2} ^ {2}. \\tag {9}\n$$\n", + "text_format": "latex", + "bbox": [ + 392, + 311, + 785, + 349 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Specular Loss. According to Eq. (3), we decompose the scene's color into diffuse color $\\mathbf{c}_d$ and view-dependent color $\\mathbf{c}_s$ . This view-dependent color accounts for the color variances due to the inherent material properties and geometries. Both components share a common color correction matrix $\\mathbf{M}$ . Ideally, we suppose the cross-image differences are mostly caused by varying light conditions and camera-dependent color translation. Thus, we apply a regularization on $\\mathbf{c}_s$ , which benefits color consistency on the diffuse partition and estimation of color correction matrix, as shown in ablation studies. The L2 regularization on the view-dependent color is formulated as:", + "bbox": [ + 212, + 357, + 787, + 492 + ], + "page_idx": 7 + }, + { + "type": "equation", + "text": "\n$$\n\\mathcal {L} _ {s} = \\sum_ {\\mathbf {r} \\in \\mathcal {R} (\\mathbf {P})} \\left\\| \\widehat {C} _ {s} \\right\\| _ {2} ^ {2} \\tag {10}\n$$\n", + "text_format": "latex", + "bbox": [ + 434, + 501, + 785, + 539 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Gray-world Loss. Due to the correlation between scene color space and color correction matrix, the recovered scene color $\\mathbf{c}$ is under-constrained with only $\\mathcal{L}_g$ . To address this, white balance correction methods often refer to some prior knowledge of the spectral properties of the camera sensors. One commonly used heuristic prior is the \"gray-world\" assumption: the color in the world tends to be natural gray on average [11]. We also define an L2 gray-world regularization for the scene's radiance field:", + "bbox": [ + 214, + 545, + 787, + 650 + ], + "page_idx": 7 + }, + { + "type": "equation", + "text": "\n$$\n\\mathcal {L} _ {g} = \\frac {1}{N _ {\\mathbf {r}}} \\sum_ {\\mathbf {r} \\in \\mathcal {R} (\\mathbf {P})} \\left\\| \\widehat {C} - C _ {0} \\right\\| _ {2} ^ {2}, \\tag {11}\n$$\n", + "text_format": "latex", + "bbox": [ + 401, + 657, + 785, + 694 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "where $C_0$ is a hyperparameter which is generally set as the midway of the color values such as 0.5.", + "bbox": [ + 212, + 700, + 784, + 729 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Histogram Loss. To further constrain the color consistency of reconstructed radiance fields in different views, besides rendering image $\\widehat{C}^{cam}$ from the ground truth view, we render an additional image from random view and obtain its histogram $\\widehat{\\mathbf{h}}_r^{cam}$ during training. Based on $\\widehat{\\mathbf{h}}_r^{cam}$ , $\\widehat{\\mathbf{h}}^{cam}$ of $\\widehat{C}^{cam}$ and $\\mathbf{h}^{cam}$ of ground truth color $C^{cam}$ , we design a cross histogram loss based on Eq. (8),", + "bbox": [ + 212, + 732, + 787, + 808 + ], + "page_idx": 7 + }, + { + "type": "equation", + "text": "\n$$\n\\mathcal {L} _ {h} = \\frac {1}{3} \\left(\\mathcal {D} _ {h} \\left(\\widehat {\\mathbf {h}} ^ {c a m}, \\mathbf {h} ^ {c a m}\\right) + \\mathcal {D} _ {h} \\left(\\widehat {\\mathbf {h}} _ {r} ^ {c a m}, \\mathbf {h} ^ {c a m}\\right) + \\mathcal {D} _ {h} \\left(\\widehat {\\mathbf {h}} ^ {c a m}, \\widehat {\\mathbf {h}} _ {r} ^ {c a m}\\right)\\right). \\tag {12}\n$$\n", + "text_format": "latex", + "bbox": [ + 284, + 816, + 785, + 844 + ], + "page_idx": 7 + }, + { + "type": "page_number", + "text": "8", + "bbox": [ + 217, + 114, + 228, + 126 + ], + "page_idx": 7 + }, + { + "type": "header", + "text": "Q. Zhang, Y. Feng et al.", + "bbox": [ + 271, + 114, + 434, + 128 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Considering that gray-world loss $\\mathcal{L}_g$ and histogram loss $\\mathcal{L}_h$ are defined on color distributions, we sample rays from grids and combine them to form an image instead of rendering the entire image during training for converge efficiency. This is based on the assumption that the color distribution of an image will remain unchanged after the random grid sampling.", + "bbox": [ + 212, + 146, + 782, + 222 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "Total Loss. The complete loss $\\mathcal{L}$ is a weighted sum of the above loss functions from Eq. (9) to Eq. (12):", + "bbox": [ + 212, + 223, + 782, + 255 + ], + "page_idx": 8 + }, + { + "type": "equation", + "text": "\n$$\n\\mathcal {L} = \\mathcal {L} _ {c} + \\lambda_ {s} \\mathcal {L} _ {s} + \\lambda_ {h} \\mathcal {L} _ {h} + \\lambda_ {g} \\mathcal {L} _ {g}, \\tag {13}\n$$\n", + "text_format": "latex", + "bbox": [ + 388, + 271, + 785, + 287 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "where $\\lambda_s = 0.1$ , $\\lambda_h = 0.01$ , and $\\lambda_g = 0.001$ are the weights of different losses.", + "bbox": [ + 214, + 301, + 767, + 316 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "4 Experiments", + "text_level": 1, + "bbox": [ + 214, + 349, + 375, + 366 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "4.1 Implementation Details", + "text_level": 1, + "bbox": [ + 214, + 390, + 455, + 405 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "Our implementation of NeRF follows [67]. Note that our camera correction module could be embedded in different NeRF-based models [4,13,50]. An eight-layer MLP with 256 channels with encoding of position $\\mathbf{x}$ as input is used to generate $\\sigma$ , and followed by one additional 256-channel fully-connected layer to extract a feature. Diffuse color $\\mathbf{c}_d$ is obtained by mapping the feature to RGB channels. The view-dependent color $\\mathbf{c}_s$ is obtained through an additional 128-channel fully connected layer and a mapping layer, with the aforementioned feature and encoding of view direction $\\mathbf{d}$ as input. Both $\\mathbf{c}_d$ and $\\mathbf{c}_s$ are transformed to nonlinear space by the kernel mapping function $\\varPhi$ in Eq. (5). The embedding vector is reshaped to color correction matrix $\\mathbf{M}$ of size $3\\times 7$ . Models are trained over 300,000 iterations with a batch size of 1024.", + "bbox": [ + 212, + 424, + 787, + 590 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "4.2 Evaluation Dataset and Metrics", + "text_level": 1, + "bbox": [ + 214, + 623, + 524, + 637 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "Dataset. We evaluate the proposed method on our collected dataset. The dataset consists of 6 real scenes captured by three digital cameras with different camera settings (e.g. white balance, ISO, exposures, etc.) and light conditions (e.g. non-flashing, and flashing with different lights). Images are collected at 90 different poses for each scene. We select 60 views with different poses as the training dataset. The remaining images are used as test images, corresponding to the same setting with training data. The resolution is $1000 \\times 667$ pixels.", + "bbox": [ + 212, + 656, + 787, + 763 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "Metrics. We report quantitative performance using PSNR (higher is better) and SSIM (higher is better) metrics, as well as the state-of-the-art LPIPS (lower is better) perceptual metric [70]. To better evaluate the color consistency and correction, we propose a $W_{2}$ metric based on Wasserstein distance, which is a distance function defined between probability distributions on the entire color", + "bbox": [ + 212, + 763, + 787, + 840 + ], + "page_idx": 8 + }, + { + "type": "header", + "text": "Color Correction for NeRF", + "bbox": [ + 549, + 114, + 730, + 127 + ], + "page_idx": 8 + }, + { + "type": "page_number", + "text": "9", + "bbox": [ + 774, + 114, + 785, + 127 + ], + "page_idx": 8 + }, + { + "type": "table", + "img_path": "images/485ec5deaca9fdfcce9c44117c822f27ef2178a37be79373eca871e27987b9b4.jpg", + "table_caption": [], + "table_footnote": [], + "table_body": "
ScenePSNR ↑SSIM ↑
NeRFNeRF-WHa-NeRFOursNeRFNeRF-WHa-NeRFOurs
Hamster15.62928.38128.28129.5710.9310.9510.9580.959
Pumpkin17.16230.18930.09631.3590.8780.9430.9420.956
Letters16.78028.04328.38730.4520.7830.9200.9220.955
MailBox18.85325.77225.90328.1280.9000.8950.8960.922
Pipe14.72522.20125.98227.5540.7440.7570.8560.907
Room16.62914.77214.96531.0970.9320.7960.7730.969
Mean16.62924.89325.60229.6940.8610.8770.8910.945
", + "bbox": [ + 218, + 143, + 782, + 239 + ], + "page_idx": 9 + }, + { + "type": "table", + "img_path": "images/edd0d0ea80e67541b105e3d3733c545bfb5835fd034acb6a0bc45a25e18829fe.jpg", + "table_caption": [], + "table_footnote": [], + "table_body": "
SceneLPIPS ↓\\( W_2 \\) (1e-3) ↓
NeRFNeRF-WHa-NeRFOursNeRFNeRF-WHa-NeRFOurs
Hamster0.5080.2190.2250.19974.42.582.592.31
Pumpkin0.5970.2340.2350.18161.11.271.400.53
Letters0.4490.3210.3240.21676.41.110.950.68
MailBox0.4390.2660.2690.20544.95.744.133.71
Pipe0.6210.4760.2790.17771.41.500.600.42
Room0.4960.6650.6680.10389.513.013.56.16
Mean0.5180.3630.3330.18069.64.203.862.30
", + "bbox": [ + 220, + 244, + 782, + 339 + ], + "page_idx": 9 + }, + { + "type": "text", + "text": "Table 1: Quantitative comparisons with baseline methods on our collected scenes, including PSNR, SSIM, LPIPS, and $W_{2}$ metrics. The Wasserstein distance $W_{2}$ measures the difference of entire color distributions.", + "bbox": [ + 215, + 339, + 784, + 380 + ], + "page_idx": 9 + }, + { + "type": "text", + "text": "space. The image-level $W_{2}$ (lower is better) is defined as:", + "bbox": [ + 212, + 420, + 625, + 435 + ], + "page_idx": 9 + }, + { + "type": "equation", + "text": "\n$$\nW _ {2} = \\frac {1}{N} \\sum_ {i} ^ {N} \\mathrm {W a s s e r s t e i n} (\\widehat {C} _ {i} ^ {c a m}, C ^ {t r a i n}),\n$$\n", + "text_format": "latex", + "bbox": [ + 232, + 443, + 493, + 479 + ], + "page_idx": 9 + }, + { + "type": "equation", + "text": "\n$$\n\\text {W a s s e r s t e i n} \\left(\\widehat {C} _ {i} ^ {\\text {c a m}}, C ^ {\\text {t r a i n}}\\right) = \\left\\| \\operatorname {m e a n} \\left(\\widehat {C} _ {i} ^ {\\text {c a m}}\\right) - \\operatorname {m e a n} \\left(C ^ {\\text {t r a i n}}\\right) \\right\\| _ {2} ^ {2} + \\tag {14}\n$$\n", + "text_format": "latex", + "bbox": [ + 232, + 483, + 784, + 508 + ], + "page_idx": 9 + }, + { + "type": "equation", + "text": "\n$$\n\\mathrm {t r a c e} \\left(\\mathrm {c o v} (\\widehat {C} _ {i} ^ {c a m}) + \\mathrm {c o v} (C ^ {t r a i n}) - 2 \\sqrt {\\sqrt {\\mathrm {c o v} (\\widehat {C} _ {i} ^ {c a m})} \\mathrm {c o v} (C ^ {t r a i n}) \\sqrt {\\mathrm {c o v} (\\widehat {C} _ {i} ^ {c a m})}}\\right),\n$$\n", + "text_format": "latex", + "bbox": [ + 232, + 511, + 740, + 549 + ], + "page_idx": 9 + }, + { + "type": "text", + "text": "where $\\mathrm{mean}(\\cdot)$ and $\\mathrm{cov}(\\cdot)$ refer to the mean and covariance of an image. $C^{train}$ denotes the training image. $\\hat{C}^{cam}$ indicates the rendered images those translated with matrix of $C^{train}$ .", + "bbox": [ + 212, + 555, + 784, + 602 + ], + "page_idx": 9 + }, + { + "type": "text", + "text": "4.3 Evaluation", + "text_level": 1, + "bbox": [ + 215, + 622, + 349, + 635 + ], + "page_idx": 9 + }, + { + "type": "text", + "text": "Baselines. We compare our method against the following baseline methods. 1) NeRF [49]: the original NeRF method, also our backbone network. 2) NeRF-W [46]: unofficial implementation where the dimension of latent code is 48. 3) Ha-NeRF [15]: official implementation which can use a reference image to do color translation.", + "bbox": [ + 212, + 643, + 782, + 717 + ], + "page_idx": 9 + }, + { + "type": "text", + "text": "Comparisons. The qualitative results of rendered novel views on our own dataset are shown in Fig. 3. Compared to baseline methods, our method achieves promising results on view rendering quality (see MSE error) and color correction measured by Wasserstein distance between ground truths and rendered images. Specifically, NeRF, which is directly reconstructed from inputs under various conditions, tends to render views with significant artifacts. NeRF-W and Ha-NeRF do ensure color consistency to a certain extent, as shown in Fig. 3. However, these methods can not distinguish the implicit properties (e.g. specular)", + "bbox": [ + 212, + 719, + 784, + 840 + ], + "page_idx": 9 + }, + { + "type": "page_number", + "text": "10", + "bbox": [ + 217, + 114, + 235, + 126 + ], + "page_idx": 9 + }, + { + "type": "header", + "text": "Q. Zhang, Y. Feng et al.", + "bbox": [ + 271, + 114, + 434, + 128 + ], + "page_idx": 9 + }, + { + "type": "image", + "img_path": "images/c4804a127f5b0d36915a8fd06ed8cb16fe45b90d4d9b549a37b0978574db69b0.jpg", + "image_caption": [ + "Fig. 3: Qualitative results of our method compared with baseline methods. We also show the error map and histogram for better visualization. MSE and Wasserstein values are on the top corners of the error map and histogram, respectively. Our method outperforms baselines, providing superior rendering quality and effective color correction. NeRF [49] generates significant artifacts. NeRF-W [46] and Ha-NeRF [15] achieve color correction to a certain extent, but exhibit shortcomings in areas with high illumination (e.g. the leaf of scene Mailbox and 'E' of scene Letters) and generate artifacts (e.g. the window of scene Mailbox and floor of scene Hamster)." + ], + "image_footnote": [], + "bbox": [ + 218, + 143, + 787, + 406 + ], + "page_idx": 10 + }, + { + "type": "text", + "text": "and external differences (e.g. camera settings). Their performance is limited in highlighting regions on Letter 'E' in scene Letters and the leaf in scene Mailbox, and results in artifacts on the window in scene Mailbox and the floor in scene Hamster. In contrast, our method outperforms state-of-the-art methods, providing superior rendering quality and effective color correction. It demonstrates the effectiveness of our color correction module, which physically defines color differences by a nonlinear color translation matrix.", + "bbox": [ + 212, + 564, + 787, + 671 + ], + "page_idx": 10 + }, + { + "type": "text", + "text": "Tab. 1 presents the quantitative results of our method compared to baseline approaches on our collected dataset, showcasing the superior rendering quality achieved by our method. Notably, NeRF-W and Ha-NeRF perform worse than the original NeRF on the scene Room. We analyze this can be attributed to the fact that the scene \"Room\" presents particular challenges with large reflective areas, which can be difficult for these two methods to fit a reasonable solution, and qualitative results are included in the supplementary material. Tab. 1 summarizes the quantitative comparisons on color consistency measured by Eq. (14). Wasserstein distance measures color consistency by calculating the color distribution distance between the input image and the rendered image translated with the color correction matrix of the input. Compared to the pixel-level PSNR", + "bbox": [ + 212, + 674, + 787, + 840 + ], + "page_idx": 10 + }, + { + "type": "header", + "text": "Color Correction for NeRF", + "bbox": [ + 550, + 114, + 730, + 126 + ], + "page_idx": 10 + }, + { + "type": "page_number", + "text": "11", + "bbox": [ + 767, + 114, + 782, + 126 + ], + "page_idx": 10 + }, + { + "type": "image", + "img_path": "images/3f912dbf6e4a7969d3e691498752f6bfc475a1278bf203a59aab85f376f50c96.jpg", + "image_caption": [ + "Fig. 4: Ablation study of gray-world loss $\\mathcal{L}_g$ . Removing $\\mathcal{L}_g$ will be unable to constrain the distribution of the scene color space, resulting in undesirable artifacts or color shifts in the reconstructed scene (e.g. overexposed Hamster and over-saturated Pumpkin)." + ], + "image_footnote": [], + "bbox": [ + 215, + 142, + 787, + 318 + ], + "page_idx": 11 + }, + { + "type": "text", + "text": "metric, the Wasserstein metric measures the entire color distribution. These comparative results demonstrate the effectiveness of our approach in addressing the limitations of existing NeRF-based models, ultimately leading to enhanced color correction and consistency in various camera conditions.", + "bbox": [ + 212, + 401, + 787, + 463 + ], + "page_idx": 11 + }, + { + "type": "text", + "text": "4.4 Ablation Studies", + "text_level": 1, + "bbox": [ + 215, + 483, + 401, + 497 + ], + "page_idx": 11 + }, + { + "type": "text", + "text": "We conduct ablative evaluations on our collected dataset to show the effectiveness of our model design. We compare our model with four variants: model without specular loss (Ours w/o $\\mathcal{L}_s$ ), model without gray-world loss (Ours w/o $\\mathcal{L}_g$ ), model without histogram loss (Ours w/o $\\mathcal{L}_h$ ), and model without nonlinear mapping (Ours w/o $\\varPhi$ ). The quantitative comparisons presented in Fig. 5 indicate that each component is crucial for generating physically plausible color-corrected renderings. Additionally, Fig. 4 displays the qualitative results of the ablation study on view synthesis with the unified color NeRF. This further highlights the importance of gray-world loss in our model. When the gray-world loss is removed, the color distribution of the scene color space becomes unconstrained, and the increased solution space can be detrimental to convergence. While this may not significantly impact the rendering quality and color consistency of novel views, it could introduce undesirable artifacts or color shifts in the reconstructed scene, such as overexposed regions of Hamster shown in Fig. 4.", + "bbox": [ + 212, + 507, + 787, + 718 + ], + "page_idx": 11 + }, + { + "type": "text", + "text": "In addition, we provide qualitative demonstrations of the results with/without specular loss $\\mathcal{L}_s$ , histogram loss $\\mathcal{L}_h$ and nonlinear mapping $\\varPhi$ . The left bottom half of the image represents the view-dependent color $C_s^{cam}$ , and the remaining part displays the entire renderings $C^{cam}$ in camera color space of ground truth. Specifically, removing specular loss will not decompose the view-dependent and view-independent colors of Letters, resulting in a significant break in color consistency (also shown in $W_2$ metric). Eliminating the histogram loss makes it challenging to represent the entire color distribution when relying solely on pixel-level", + "bbox": [ + 212, + 719, + 789, + 840 + ], + "page_idx": 11 + }, + { + "type": "page_number", + "text": "12", + "bbox": [ + 217, + 114, + 235, + 126 + ], + "page_idx": 11 + }, + { + "type": "header", + "text": "Q. Zhang, Y. Feng et al.", + "bbox": [ + 271, + 114, + 436, + 128 + ], + "page_idx": 11 + }, + { + "type": "image", + "img_path": "images/38e0f1468d3da6358ed37a2e4d896e369c89143b5d128ba1c1b06aa249968be9.jpg", + "image_caption": [ + "Ground Truth" + ], + "image_footnote": [], + "bbox": [ + 217, + 143, + 405, + 239 + ], + "page_idx": 12 + }, + { + "type": "image", + "img_path": "images/040a3170eaf9e2b815c43b105a2dab25deef701704885a4ac978be1e68ed6ddf.jpg", + "image_caption": [ + "Ours w/o $\\mathcal{L}_s$" + ], + "image_footnote": [], + "bbox": [ + 405, + 143, + 594, + 239 + ], + "page_idx": 12 + }, + { + "type": "image", + "img_path": "images/f8af0581431f971b575d1fe9eff955a8ff7b31b08f895c94e027407b7e482f6f.jpg", + "image_caption": [ + "Ours w/o $\\mathcal{L}_h$" + ], + "image_footnote": [], + "bbox": [ + 594, + 143, + 784, + 241 + ], + "page_idx": 12 + }, + { + "type": "image", + "img_path": "images/907a351ec30140b91f866a5b6b7ef58984a8e180e3368f9700b6b5b9843fd715.jpg", + "image_caption": [ + "Quantitative Comparisons" + ], + "image_footnote": [], + "bbox": [ + 215, + 250, + 403, + 345 + ], + "page_idx": 12 + }, + { + "type": "image", + "img_path": "images/031f2d3b90f4c0d7bdd144659732162d0012cb5800e9299391c9ab148ef642a6.jpg", + "image_caption": [ + "Ours w/o $\\Phi$", + "Fig. 5: Ablation Study on \"Letter\", including qualitative and quantitative results with/without specular loss $\\mathcal{L}_s$ , histogram loss $\\mathcal{L}_h$ and nonlinear mapping $\\varPhi$ . The left bottom half of the image is the view-dependent color $C_s^{cam}$ , and the other part is the entire color-correct renderings $C^{cam}$ . Removing specular loss struggles to the decomposition of view-dependent and view-independent colors. Removing histogram loss lacks the constraint of the entire color space, making the color offset. Downgrading the color translation matrix suffers from artifacts of the nonlinear color difference." + ], + "image_footnote": [], + "bbox": [ + 405, + 250, + 594, + 347 + ], + "page_idx": 12 + }, + { + "type": "image", + "img_path": "images/ef92f0e33ef81c240635e19ab5811d1ccc8ed9ac0efe23f070b842adb6ab9637.jpg", + "image_caption": [ + "Ours" + ], + "image_footnote": [], + "bbox": [ + 594, + 250, + 784, + 347 + ], + "page_idx": 12 + }, + { + "type": "text", + "text": "MSE error. This may result in color shifts, such as producing darker images than the ground truth. Removing nonlinear mapping and downgrading color translation matrix to $3 \\times 3$ will neglect the nonlinear process during imaging. This forces the network to model more color variances into $\\mathbf{C}_s^{cam}$ , making the color correction module unable to fully perceive the color characteristics of each camera, as shown in Fig. 5. These ablation studies demonstrate the effectiveness of our model and specifically designed losses.", + "bbox": [ + 212, + 496, + 787, + 602 + ], + "page_idx": 12 + }, + { + "type": "text", + "text": "4.5 Color Translation", + "text_level": 1, + "bbox": [ + 215, + 625, + 408, + 638 + ], + "page_idx": 12 + }, + { + "type": "text", + "text": "Fig. 6 presents examples of color translation guided by a reference image. This advantage makes our method particularly suitable for applications where accurate color reproduction and adherence to reference images are crucial, such as in virtual reality, computer graphics, and image synthesis.", + "bbox": [ + 212, + 648, + 787, + 710 + ], + "page_idx": 12 + }, + { + "type": "text", + "text": "5 Conclusion", + "text_level": 1, + "bbox": [ + 215, + 732, + 359, + 750 + ], + "page_idx": 12 + }, + { + "type": "text", + "text": "This paper proposes a framework that incorporates a color correction module within a NeRF to achieve unified color radiance field reconstruction from multiview images captured under different camera settings and lighting conditions. Specifically, we have developed a color correction module to predict a nonlinear color translation matrix that simulates the physical imaging process. Besides the", + "bbox": [ + 212, + 763, + 787, + 840 + ], + "page_idx": 12 + }, + { + "type": "header", + "text": "Color Correction for NeRF", + "bbox": [ + 550, + 114, + 730, + 126 + ], + "page_idx": 12 + }, + { + "type": "page_number", + "text": "13", + "bbox": [ + 767, + 114, + 784, + 126 + ], + "page_idx": 12 + }, + { + "type": "image", + "img_path": "images/77315023113e85d70c608821a9b80273c541a305f6d827f9ba101c3868544750.jpg", + "image_caption": [ + "Fig. 6: Examples of color translation. Our method produces spatially and temporally consistent color transfer results that faithfully match the reference image while maintaining photorealism. Video can also be found in supplementary materials." + ], + "image_footnote": [], + "bbox": [ + 223, + 142, + 785, + 518 + ], + "page_idx": 13 + }, + { + "type": "text", + "text": "color correction module for accommodating external differences across distinct cameras, we predict a view-dependent function to separate inherent differences caused by scene properties. To enhance the accuracy of the color translation, we have designed various regularization strategies and demonstrated their efficacy through experiments. Our approach significantly improves color consistency in NeRF reconstruction, enabling users to render novel views under flexible and unrestricted settings.", + "bbox": [ + 212, + 619, + 787, + 726 + ], + "page_idx": 13 + }, + { + "type": "text", + "text": "Despite its advantages, we acknowledge that our framework has certain limitations. For instance, compared to NeRF, the proposed Histogram loss results in an increased training time of approximately $10\\%$ . Furthermore, the Histogram loss, which computes the difference between 2D images, performs well when the rendered image and reference image are nearby. This limitation could potentially be addressed by introducing a new representation and a 3D Histogram loss. We plan to explore these improvements in future work.", + "bbox": [ + 212, + 734, + 787, + 840 + ], + "page_idx": 13 + }, + { + "type": "page_number", + "text": "14", + "bbox": [ + 217, + 114, + 235, + 126 + ], + "page_idx": 13 + }, + { + "type": "header", + "text": "Q. Zhang, Y. Feng et al.", + "bbox": [ + 271, + 114, + 436, + 128 + ], + "page_idx": 13 + }, + { + "type": "text", + "text": "References", + "text_level": 1, + "bbox": [ + 215, + 143, + 321, + 159 + ], + "page_idx": 14 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "1. Affi, M., Barron, J.T., LeGendre, C., Tsai, Y.T., Bleibel, F.: Cross-camera convolutional color constancy. In: Proceedings of the IEEE/CVF International Conference on Computer Vision. pp. 1981-1990 (2021)", + "2. Affi, M., Brown, M.S.: Interactive white balancing for camera-rendered images. arXiv preprint arXiv:2009.12632 (2020)", + "3. Affifi, M., Brubaker, M.A., Brown, M.S.: Auto white-balance correction for mixed-illuminant scenes. In: Proceedings of the IEEE/CVF Winter Conference on Applications of Computer Vision. pp. 1210–1219 (2022)", + "4. Barron, J.T., Mildenhall, B., Tancik, M., Hedman, P., Martin-Brualla, R., Srinivasan, P.P.: Mip-nerf: A multiscale representation for anti-aliasing neural radiance fields. In: Proceedings of the IEEE/CVF International Conference on Computer Vision. pp. 5855-5864 (2021)", + "5. Barron, J.T., Mildenhall, B., Verbin, D., Srinivasan, P.P., Hedman, P.: Zip-nerf: Anti-aliased grid-based neural radiance fields. arXiv preprint arXiv:2304.06706 (2023)", + "6. Barron, J.T., Tsai, Y.T.: Fast fourier color constancy. In: Proceedings of the IEEE conference on computer vision and pattern recognition. pp. 886-894 (2017)", + "7. Bian, W., Wang, Z., Li, K., Bian, J.W., Prisacariu, V.A.: Nope-nerf: Optimising neural radiance field with no pose prior. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 4160-4169 (2023)", + "8. Bianco, S., Bruna, A., Naccari, F., Schettini, R.: Color space transformations for digital photography exploiting information about the illuminant estimation process. JOSA A 29(3), 374-384 (2012)", + "9. Boss, M., Braun, R., Jampani, V., Barron, J.T., Liu, C., Lensch, H.: Nerd: Neural reflectance decomposition from image collections. In: Proceedings of the IEEE/CVF International Conference on Computer Vision. pp. 12684-12694 (2021)", + "0. Boss, M., Jampani, V., Braun, R., Liu, C., Barron, J., Lensch, H.: Neural-pil: Neural pre-integrated lighting for reflectance decomposition. Advances in Neural Information Processing Systems 34, 10691-10704 (2021)", + "1. Buchsbaum, G.: A spatial processor model for object colour perception. Journal of the Franklin institute 310(1), 1-26 (1980)", + "2. Cao, A., Johnson, J.: Hexplane: A fast representation for dynamic scenes. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 130-141 (2023)", + "3. Chen, A., Xu, Z., Geiger, A., Yu, J., Su, H.: Tensorf: Tensorial radiance fields. In: European Conference on Computer Vision. pp. 333-350. Springer (2022)", + "4. Chen, A., Xu, Z., Zhao, F., Zhang, X., Xiang, F., Yu, J., Su, H.: Mvsnerf: Fast generalizable radiance field reconstruction from multi-view stereo. In: Proceedings of the IEEE/CVF international conference on computer vision. pp. 14124-14133 (2021)", + "5. Chen, X., Zhang, Q., Li, X., Chen, Y., Feng, Y., Wang, X., Wang, J.: Hallucinated neural radiance fields in the wild. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 12943-12952 (2022)", + "6. Chen, Y., Chen, X., Wang, X., Zhang, Q., Guo, Y., Shan, Y., Wang, F.: Local-to-global registration for bundle-adjusting neural radiance fields. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 8264-8273 (2023)" + ], + "bbox": [ + 225, + 178, + 784, + 839 + ], + "page_idx": 14 + }, + { + "type": "header", + "text": "Color Correction for NeRF", + "bbox": [ + 549, + 114, + 730, + 126 + ], + "page_idx": 14 + }, + { + "type": "page_number", + "text": "15", + "bbox": [ + 767, + 116, + 785, + 126 + ], + "page_idx": 14 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "17. Chen, Z., Funkhouser, T., Hedman, P., Tagliasacchi, A.: Mobilenerf: Exploiting the polygon rasterization pipeline for efficient neural field rendering on mobile architectures. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 16569-16578 (2023)", + "18. Cheng, W., Chen, R., Fan, S., Yin, W., Chen, K., Cai, Z., Wang, J., Gao, Y., Yu, Z., Lin, Z., et al.: Dna-rendering: A diverse neural actor repository for high-fidelity human-centric rendering. In: Proceedings of the IEEE/CVF International Conference on Computer Vision. pp. 19982-19993 (2023)", + "19. Chibane, J., Bansal, A., Lazova, V., Pons-Moll, G.: Stereo radiance fields (srf): Learning view synthesis for sparse views of novel scenes. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 7911-7920 (2021)", + "20. Ding, C., Ma, Z.: Multi-camera color correction via hybrid histogram matching. IEEE Transactions on Circuits and Systems for Video Technology 31(9), 3327-3337 (2020)", + "21. Fang, J., Yi, T., Wang, X., Xie, L., Zhang, X., Liu, W., Nießner, M., Tian, Q.: Fast dynamic radiance fields with time-aware neural voxels. In: SIGGRAPH Asia 2022 Conference Papers. pp. 1-9 (2022)", + "22. Fang, Q., Shuai, Q., Dong, J., Bao, H., Zhou, X.: Reconstructing 3d human pose by watching humans in the mirror. In: CVPR (2021)", + "23. Fridovich-Keil, S., Meanti, G., Warburg, F.R., Recht, B., Kanazawa, A.: K-planes: Explicit radiance fields in space, time, and appearance. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 12479-12488 (2023)", + "24. Fridovich-Keil, S., Yu, A., Tancik, M., Chen, Q., Recht, B., Kanazawa, A.: Plenoxels: Radiance fields without neural networks. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 5501-5510 (2022)", + "25. Garbin, S.J., Kowalski, M., Johnson, M., Shotton, J., Valentin, J.: Fastnerf: High-fidelity neural rendering at 200fps. In: Proceedings of the IEEE/CVF International Conference on Computer Vision. pp. 14346-14355 (2021)", + "26. Hedman, P., Srinivasan, P.P., Mildenhall, B., Barron, J.T., Debevec, P.: Baking neural radiance fields for real-time view synthesis. In: Proceedings of the IEEE/CVF International Conference on Computer Vision. pp. 5875-5884 (2021)", + "27. Huang, X., Zhang, Q., Feng, Y., Li, H., Wang, Q.: Inverting the imaging process by learning an implicit camera model. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 21456-21465 (2023)", + "28. Huang, X., Zhang, Q., Feng, Y., Li, H., Wang, X., Wang, Q.: HDR-nerf: High dynamic range neural radiance fields. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 18398-18408 (2022)", + "29. Huang, X., Zhang, Q., Feng, Y., Li, X., Wang, X., Wang, Q.: Local implicit ray function for generalizable radiance field representation. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 97-107 (2023)", + "30. Huang, Z., Liang, Z., Zhang, H., Lin, Y., Jia, K.: Sur2f: A hybrid representation for high-quality and efficient surface reconstruction from multi-view images. arXiv preprint arXiv:2401.03704 (2024)", + "31. Isik, M., Rünz, M., Georgopoulos, M., Khakhulin, T., Starck, J., Agapito, L., Nießner, M.: Humanrf: High-fidelity neural radiance fields for humans in motion. arXiv preprint arXiv:2305.06356 (2023)" + ], + "bbox": [ + 215, + 146, + 785, + 840 + ], + "page_idx": 15 + }, + { + "type": "page_number", + "text": "16", + "bbox": [ + 217, + 114, + 235, + 126 + ], + "page_idx": 15 + }, + { + "type": "header", + "text": "Q. Zhang, Y. Feng et al.", + "bbox": [ + 271, + 114, + 434, + 128 + ], + "page_idx": 15 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "32. Jin, H., Liu, I., Xu, P., Zhang, X., Han, S., Bi, S., Zhou, X., Xu, Z., Su, H.: Tensoroir: Tensorial inverse rendering. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 165-174 (2023)", + "33. Johari, M.M., Lepoittevin, Y., Fleuret, F.: Geonerf: Generalizing nerf with geometry priors. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 18365-18375 (2022)", + "34. Joo, H., Simon, T., Li, X., Liu, H., Tan, L., Gui, L., Banerjee, S., Godisart, T.S., Nabbe, B., Matthews, I., Kanade, T., Nobuhara, S., Sheikh, Y.: Panoptic studio: A massively multiview system for social interaction capture. IEEE Transactions on Pattern Analysis and Machine Intelligence (2017)", + "35. Kerbl, B., Kopanas, G., Leimkuhler, T., Drettakis, G.: 3d gaussian splatting for real-time radiance field rendering. ACM Transactions on Graphics (ToG) 42(4), 1-14 (2023)", + "36. Kinh, F., Yilmaz, D., Özcan, B., Kırac, F.: Deterministic neural illumination mapping for efficient auto-white balance correction. In: Proceedings of the IEEE/CVF International Conference on Computer Vision. pp. 1139-1147 (2023)", + "37. Kinh, F., Yilmaz, D., Ozcan, B., Kirac, F.: Modeling the lighting in scenes as style for auto white-balance correction. In: Proceedings of the IEEE/CVF Winter Conference on Applications of Computer Vision. pp. 4903-4913 (2023)", + "38. Kuang, Z., Luan, F., Bi, S., Shu, Z., Wetzstein, G., Sunkavalli, K.: Palettenerf: Palette-based appearance editing of neural radiance fields. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 20691-20700 (2023)", + "39. Lee, J.H., Kim, D.S.: Ice-nerf: Interactive color editing of nerfs via decomposition-aware weight optimization. In: Proceedings of the IEEE/CVF International Conference on Computer Vision. pp. 3491-3501 (2023)", + "40. Li, J., Feng, Z., She, Q., Ding, H., Wang, C., Lee, G.H.: Mine: Towards continuous depth MPI with nerf for novel view synthesis. In: Proceedings of the IEEE/CVF International Conference on Computer Vision. pp. 12578-12588 (2021)", + "41. Liang, Z., Huang, Z., Ding, C., Jia, K.: Helixsurf: A robust and efficient neural implicit surface learning of indoor scenes with iterative intertwined regularization. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 13165-13174 (2023)", + "42. Lin, C.H., Ma, W.C., Torralba, A., Lucey, S.: Barf: Bundle-adjusting neural radiance fields. In: Proceedings of the IEEE/CVF international conference on computer vision. pp. 5741-5751 (2021)", + "43. Liu, Y.L., Gao, C., Meuleman, A., Tseng, H.Y., Saraf, A., Kim, C., Chuang, Y.Y., Kopf, J., Huang, J.B.: Robust dynamic radiance fields. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 13-23 (2023)", + "44. Liu, Y., Peng, S., Liu, L., Wang, Q., Wang, P., Theobalt, C., Zhou, X., Wang, W.: Neural rays for occlusion-aware image-based rendering. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 7824-7833 (2022)", + "45. Ma, L., Li, X., Liao, J., Zhang, Q., Wang, X., Wang, J., Sander, P.V.: Deblur-nerf: Neural radiance fields from blurry images. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 12861-12870 (2022)", + "46. Martin-Brualla, R., Radwan, N., Sajjadi, M.S., Barron, J.T., Dosovitskiy, A., Duckworth, D.: Nerf in the wild: Neural radiance fields for unconstrained photo collections. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 7210-7219 (2021)" + ], + "bbox": [ + 212, + 146, + 787, + 839 + ], + "page_idx": 16 + }, + { + "type": "header", + "text": "Color Correction for NeRF", + "bbox": [ + 550, + 114, + 730, + 126 + ], + "page_idx": 16 + }, + { + "type": "page_number", + "text": "17", + "bbox": [ + 767, + 116, + 784, + 126 + ], + "page_idx": 16 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "47. Max, N.: Optical models for direct volume rendering. IEEE Transactions on Visualization and Computer Graphics 1(2), 99-108 (1995)", + "48. Mildenhall, B., Hedman, P., Martin-Brualla, R., Srinivasan, P.P., Barron, J.T.: Nerf in the dark: High dynamic range view synthesis from noisy raw images. In: Proceedings of the IEEE/CVF conference on computer vision and pattern recognition. pp. 16190-16199 (2022)", + "49. Mildenhall, B., Srinivasan, P.P., Tancik, M., Barron, J.T., Ramamoorthi, R., Ng, R.: Nerf: Representing scenes as neural radiance fields for view synthesis. In: European Conference on Computer Vision. pp. 405-421 (2020)", + "50. Müller, T., Evans, A., Schied, C., Keller, A.: Instant neural graphics primitives with a multiresolution hash encoding. ACM Transactions on Graphics (ToG) 41(4), 1-15 (2022)", + "51. Park, K., Henzler, P., Mildenhall, B., Barron, J.T., Martin-Brualla, R.: Camp: Camera preconditioning for neural radiance fields. ACM Transactions on Graphics (TOG) 42(6), 1-11 (2023)", + "52. Pearl, N., Treibitz, T., Korman, S.: Nan: Noise-aware nerfs for burst-denoising. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 12672-12681 (2022)", + "53. Peng, S., Zhang, Y., Xu, Y., Wang, Q., Shuai, Q., Bao, H., Zhou, X.: Neural body: Implicit neural representations with structured latent codes for novel view synthesis of dynamic humans. In: CVPR (2021)", + "54. Reiser, C., Peng, S., Liao, Y., Geiger, A.: Kilonerf: Speeding up neural radiance fields with thousands of tiny mlp's. In: Proceedings of the IEEE/CVF International Conference on Computer Vision. pp. 14335-14345 (2021)", + "55. Rematas, K., Liu, A., Srinivasan, P.P., Barron, J.T., Tagliasacchi, A., Funkhouser, T., Ferrari, V.: Urban radiance fields. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 12932-12942 (2022)", + "56. Rodríguez, R.G., Vazquez-Corral, J., Bertalmío, M.: Color matching images with unknown non-linear encodings. IEEE Transactions on Image Processing 29, 4435-4444 (2020)", + "57. Shao, R., Zheng, Z., Tu, H., Liu, B., Zhang, H., Liu, Y.: Tensor4d: Efficient neural 4d decomposition for high-fidelity dynamic reconstruction and rendering. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 16632-16642 (2023)", + "58. Srinivasan, P.P., Deng, B., Zhang, X., Tancik, M., Mildenhall, B., Barron, J.T.: Nerv: Neural reflectance and visibility fields for relighting and view synthesis. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 7495-7504 (2021)", + "59. Sun, G., Chen, X., Chen, Y., Pang, A., Lin, P., Jiang, Y., Xu, L., Yu, J., Wang, J.: Neural free-viewpoint performance rendering under complex human-object interactions. In: Proceedings of the 29th ACM International Conference on Multimedia. pp. 4651-4660 (2021)", + "60. Tancik, M., Casser, V., Yan, X., Pradhan, S., Mildenhall, B., Srinivasan, P.P., Barron, J.T., Kretzschmar, H.: Block-nerf: Scalable large scene neural view synthesis. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 8248-8258 (2022)", + "61. Tang, J., Zhou, H., Chen, X., Hu, T., Ding, E., Wang, J., Zeng, G.: Delicate textured mesh recovery from nef via adaptive surface refinement. arXiv preprint arXiv:2303.02091 (2023)" + ], + "bbox": [ + 215, + 147, + 785, + 839 + ], + "page_idx": 17 + }, + { + "type": "page_number", + "text": "18", + "bbox": [ + 217, + 114, + 235, + 126 + ], + "page_idx": 17 + }, + { + "type": "header", + "text": "Q. Zhang, Y. Feng et al.", + "bbox": [ + 271, + 114, + 434, + 128 + ], + "page_idx": 17 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "62. Verbin, D., Hedman, P., Mildenhall, B., Zickler, T., Barron, J.T., Srinivasan, P.P.: Ref-nerf: Structured view-dependent appearance for neural radiance fields. In: 2022 IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR). pp. 5481-5490. IEEE (2022)", + "63. Wang, Q., Wang, Z., Genova, K., Srinivasan, P.P., Zhou, H., Barron, J.T., Martin-Brualla, R., Snavely, N., Funkhouser, T.: Ibrnet: Learning multi-view image-based rendering. In: Proceedings of the IEEE/CVF conference on computer vision and pattern recognition. pp. 4690-4699 (2021)", + "64. Xie, S., Zhu, H., Liu, Z., Zhang, Q., Zhou, Y., Cao, X., Ma, Z.: Diner: Disorder-invariant implicit neural representation. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 6143-6152 (2023)", + "65. Xu, Q., Xu, Z., Philip, J., Bi, S., Shu, Z., Sunkavalli, K., Neumann, U.: Point-nerf: Point-based neural radiance fields. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 5438-5448 (2022)", + "66. Yao, Y., Zhang, J., Liu, J., Qu, Y., Fang, T., McKinnon, D., Tsin, Y., Quan, L.: Neilf: Neural incident light field for physically-based material estimation. In: European Conference on Computer Vision. pp. 700-716. Springer (2022)", + "67. Yen-Chen, l.: Nerf-pytorch: a pytorch implementation of nerf. URL https://github.com/yenchenlin/nerf-pytorch (2020)", + "68. Yu, A., Ye, V., Tancik, M., Kanazawa, A.: pixelnerf: Neural radiance fields from one or few images. In: Proceedings of the IEEE/CVF conference on computer vision and pattern recognition. pp. 4578-4587 (2021)", + "69. Zhang, K., Luan, F., Li, Z., Snavely, N.: Iron: Inverse rendering by optimizing neural sdfs and materials from photometric images. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 5565-5574 (2022)", + "70. Zhang, R., Isola, P., Efros, A.A., Shechtman, E., Wang, O.: The unreasonable effectiveness of deep features as a perceptual metric. In: Proceedings of the IEEE conference on computer vision and pattern recognition. pp. 586-595 (2018)", + "71. Zhang, X., Srinivasan, P.P., Deng, B., Debevec, P., Freeman, W.T., Barron, J.T.: Nerfactor: Neural factorization of shape and reflectance under an unknown illumination. ACM Transactions on Graphics (ToG) 40(6), 1-18 (2021)", + "72. Zhao, Y., Ferguson, S., Zhou, H., Elliott, C., Rafferty, K.: Color alignment for relative color constancy via non-standard references. IEEE Transactions on Image Processing 31, 6591-6604 (2022)", + "73. Zhou, T., He, K., Wu, D., Xu, T., Zhang, Q., Shao, K., Chen, W., Xu, L., Yu, J.: Relightable neural human assets from multi-view gradient illuminations. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 4315-4327 (2023)", + "74. Zhu, H., Liu, F., Zhang, Q., Cao, X., Ma, Z.: Rhino: Regularizing the hash-based implicit neural representation. arXiv preprint arXiv:2309.12642 (2023)", + "75. Zhu, H., Xie, S., Liu, Z., Liu, F., Zhang, Q., Zhou, Y., Lin, Y., Ma, Z., Cao, X.: Disorder-invariant implicit neural representation. IEEE Transactions on Pattern Analysis and Machine Intelligence (2024)", + "76. Zhu, J., Zhu, H., Zhang, Q., Zhu, F., Ma, Z., Cao, X.: Pyramid nef: Frequency guided fast radiance field optimization. International Journal of Computer Vision 131(10), 2649-2664 (2023)", + "77. Zhuang, Y., Zhang, Q., Wang, X., Zhu, H., Feng, Y., Li, X., Shan, Y., Cao, X.: A pre-convolved representation for plug-and-play neural illumination fields. In: Proceedings of the AAAI Conference on Artificial Intelligence. vol. 38, pp. 7828-7836 (2024)" + ], + "bbox": [ + 212, + 146, + 787, + 830 + ], + "page_idx": 18 + }, + { + "type": "header", + "text": "Color Correction for NeRF", + "bbox": [ + 550, + 114, + 730, + 126 + ], + "page_idx": 18 + }, + { + "type": "page_number", + "text": "19", + "bbox": [ + 767, + 114, + 785, + 126 + ], + "page_idx": 18 + } +] \ No newline at end of file diff --git a/2024/Physically Plausible Color Correction for Neural Radiance Fields/88cfdd54-68fa-4f55-bb61-e981b6d01cd5_model.json b/2024/Physically Plausible Color Correction for Neural Radiance Fields/88cfdd54-68fa-4f55-bb61-e981b6d01cd5_model.json new file mode 100644 index 0000000000000000000000000000000000000000..59f50e6ca0f3ad7d2c18f68d4333a7974b5239d6 --- /dev/null +++ b/2024/Physically Plausible Color Correction for Neural Radiance Fields/88cfdd54-68fa-4f55-bb61-e981b6d01cd5_model.json @@ -0,0 +1,2724 @@ +[ + [ + { + "type": "title", + "bbox": [ + 0.225, + 0.141, + 0.78, + 0.187 + ], + "angle": 0, + "content": "Physically Plausible Color Correction for Neural Radiance Fields" + }, + { + "type": "text", + "bbox": [ + 0.319, + 0.212, + 0.685, + 0.229 + ], + "angle": 0, + "content": "Qi Zhang\\(^{1,\\dagger}\\), Ying Feng\\(^{1,\\dagger}\\), and Hongdong Li\\(^{2}\\)" + }, + { + "type": "text", + "bbox": [ + 0.284, + 0.24, + 0.721, + 0.254 + ], + "angle": 0, + "content": "\\(^{1}\\) Independent Researcher {nwpuqzhang, yfeng.von}@gmail.com" + }, + { + "type": "text", + "bbox": [ + 0.303, + 0.254, + 0.7, + 0.268 + ], + "angle": 0, + "content": "2 Australian National University hongdong.li@anu.edu.au" + }, + { + "type": "text", + "bbox": [ + 0.262, + 0.304, + 0.744, + 0.624 + ], + "angle": 0, + "content": "Abstract. Neural Radiance Fields have become the representation of choice for many 3D computer vision and computer graphics applications, e.g., novel view synthesis and 3D reconstruction. Multi-camera systems are commonly used as the image capture setup in NeRF-based multi-view tasks such as dynamic scene acquisition or realistic avatar animation. However, a critical issue that has often been overlooked in this setup is the evident differences in color responses among multiple cameras, which adversely affect the NeRF reconstruction performance. These color discrepancies among multiple input images stem from two aspects: 1) implicit properties of the scenes such as reflections and shadings, and 2) external differences in camera settings and lighting conditions. In this paper, we address this problem by proposing a novel color correction module that simulates the physical color processing in cameras to be embedded in NeRF, enabling the unified color NeRF reconstruction. Besides the view-independent color correction module for external differences, we predict a view-dependent function to minimize the color residual (including, e.g., specular and shading) to eliminate the impact of inherent attributes. We further describe how the method can be extended with a reference image as guidance to achieve aesthetically plausible color consistency and color translation on novel views. Experiments validate that our method is superior to baseline methods in both quantitative and qualitative evaluations of color correction and color consistency." + }, + { + "type": "text", + "bbox": [ + 0.262, + 0.636, + 0.741, + 0.65 + ], + "angle": 0, + "content": "Keywords: Neural Radiance Field \\(\\cdot\\) View Synthesis \\(\\cdot\\) Color Translation" + }, + { + "type": "title", + "bbox": [ + 0.216, + 0.697, + 0.377, + 0.714 + ], + "angle": 0, + "content": "1 Introduction" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.728, + 0.788, + 0.804 + ], + "angle": 0, + "content": "The synthesis and rendering of novel views using multi-view images has been a long-standing problem in computer vision and graphics, which is also a cornerstone task for augmented reality (AR) and virtual reality (VR) applications. The recent success of Neural Radiance Fields (NeRF) [49] has made the implicit neural representation a powerful approach for novel view synthesis tasks, as" + }, + { + "type": "page_footnote", + "bbox": [ + 0.232, + 0.812, + 0.563, + 0.827 + ], + "angle": 0, + "content": "† These authors contributed equally to this work." + } + ], + [ + { + "type": "page_number", + "bbox": [ + 0.218, + 0.116, + 0.23, + 0.127 + ], + "angle": 0, + "content": "2" + }, + { + "type": "header", + "bbox": [ + 0.272, + 0.115, + 0.437, + 0.13 + ], + "angle": 0, + "content": "Q. Zhang, Y. Feng et al." + }, + { + "type": "image", + "bbox": [ + 0.223, + 0.15, + 0.422, + 0.278 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.288, + 0.278, + 0.357, + 0.288 + ], + "angle": 0, + "content": "(a) input views" + }, + { + "type": "image", + "bbox": [ + 0.429, + 0.151, + 0.644, + 0.277 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.471, + 0.278, + 0.61, + 0.289 + ], + "angle": 0, + "content": "(b) unified color radiance field" + }, + { + "type": "image", + "bbox": [ + 0.653, + 0.151, + 0.77, + 0.274 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.671, + 0.276, + 0.76, + 0.289 + ], + "angle": 0, + "content": "(c) color translation" + }, + { + "type": "image_caption", + "bbox": [ + 0.214, + 0.3, + 0.789, + 0.399 + ], + "angle": 0, + "content": "Fig. 1: Physically Plausible Color Correction for Neural Radiance Fields. (a) Given a set of images captured under various camera settings and lighting conditions, our method introduces a color correction module into the NeRF-based model, which enables recovering (b) a unified color radiance field with a physically plausible white balance. Unlike prior methods that embed color through features, our color correction module simulates color differences via a nonlinear color translation matrix. Additionally, our method can render (c) novel views translated from the color of a reference image." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.43, + 0.789, + 0.506 + ], + "angle": 0, + "content": "an alternative to traditional image-based reconstruction techniques. The NeRF model employs Multi-Layer Perceptrons (MLPs) to parameterize the mapping from spatial coordinates to corresponding densities and view-dependent colors. This approach has been proven effective with scenes that are consistently well-lit without substantial variations in image brightness." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.507, + 0.789, + 0.749 + ], + "angle": 0, + "content": "With the increasing prevalence of multi-camera systems [18, 22, 31, 34, 53, 59, 73], NeRF and its variants [4, 5, 50] face challenges in achieving photorealistic novel view synthesis with color consistency. The differences in color response among multiple cameras can significantly impact the accuracy and consistency of synthesized views. Even when capturing with identical camera settings, different types of cameras may present distinct color responses caused by the difference in sensors and ISP algorithms. To address this limitation, NeRF-W [46] and Ha-NeRF [15] incorporate additional latent codes to encode appearances into the NeRF model for consistent view synthesis and appearance control. However, these methods tend to embed the overall appearances as view-dependent features, making it difficult to achieve fine-grained and intuitive color correction. HDR-NeRF [28] attempts to address this issue but primarily focuses on building a physical module for camera exposures. It is necessary to introduce color translation for NeRF-based multi-view tasks, which should involve modeling color correlations across cameras and enabling physically plausible color correction and consistency in NeRF reconstructions." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.75, + 0.79, + 0.841 + ], + "angle": 0, + "content": "The color difference across multi-view images arises from two aspects: inherent view-dependent scene properties (e.g. reflections and shadings); external differences in camera settings (e.g. white balance, ISO, exposure) and lighting conditions. Naturally, we could calibrate each image via color correction methods [1,3,6,36,37], and then input to NeRF-based method for novel view synthesis. However, these image-based methods struggle to preserve color consistency in" + } + ], + [ + { + "type": "header", + "bbox": [ + 0.55, + 0.115, + 0.732, + 0.128 + ], + "angle": 0, + "content": "Color Correction for NeRF" + }, + { + "type": "page_number", + "bbox": [ + 0.775, + 0.117, + 0.785, + 0.127 + ], + "angle": 0, + "content": "3" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.147, + 0.784, + 0.234 + ], + "angle": 0, + "content": "NeRF reconstructions. In addition to faithfully reproducing the captured colors, we consider that color correction process should allow for color adjustment to enhance visual appeal, and even color translation to a given reference image. By incorporating such color correction module, users, including novice ones, can easily recover NeRF and adjust the colors of novel views based on their preferences." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.238, + 0.788, + 0.419 + ], + "angle": 0, + "content": "In this paper, we propose a novel framework that incorporates a color correction module into the NeRF-based model to achieve unified color radiance field reconstruction from multi-view images with different camera settings and lighting conditions, as shown in Fig. 1. In contrast to the embedding feature for overall color differences in prior NeRF-based methods [15,46], we present a color correction module to predict a nonlinear color translation matrix that simulates the color processing in cameras, such as white balance, nonlinear color enhancement, tone-manipulation, gamma encoding, etc. Additionally, considering the view-dependent factors of inherent properties, we predict a view-dependent function to separate the color residual (e.g. specular and shading of the scene) to improve the accuracy of the color translation matrix. Our technical contributions are summarized as follows:" + }, + { + "type": "text", + "bbox": [ + 0.226, + 0.429, + 0.784, + 0.471 + ], + "angle": 0, + "content": "- We present a novel framework for color correction and translation in NeRF reconstruction, allowing even novice users to perform novel view synthesis under unstable camera settings in an intuitive manner." + }, + { + "type": "text", + "bbox": [ + 0.226, + 0.474, + 0.784, + 0.5 + ], + "angle": 0, + "content": "- We propose a color correction module that defines the non-linear color translation to be embedded in NeRF." + }, + { + "type": "text", + "bbox": [ + 0.226, + 0.504, + 0.784, + 0.545 + ], + "angle": 0, + "content": "- We introduce tailored regularizations to eliminate the impact of inherent scene properties, constrain the scene's NeRF to a white balance color space, and maintain the entire color distribution." + }, + { + "type": "list", + "bbox": [ + 0.226, + 0.429, + 0.784, + 0.545 + ], + "angle": 0, + "content": null + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.559, + 0.784, + 0.589 + ], + "angle": 0, + "content": "Extensive experiments have shown that our method outperforms baseline methods both quantitatively and qualitatively in terms of color correction." + }, + { + "type": "title", + "bbox": [ + 0.215, + 0.612, + 0.396, + 0.629 + ], + "angle": 0, + "content": "2 Related Works" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.644, + 0.788, + 0.841 + ], + "angle": 0, + "content": "Neural Radiance Fields Neural Radiance Fields represent a scene implicitly with multi-layer perceptrons, approximating a continuous function that maps each 3D position and direction to a density and radiance [49]. By incorporating view direction features, NeRF can effectively model view-dependent effects such as reflection and refraction. This allows for photo-realistic synthesis of novel views with volumetric rendering. To alleviate the computational limitation of massive MLP evaluations while maintaining the high-frequency appearances, many extensive works have been proposed to speed up the training and inference of the model [5,13,17,24-26,35,50,54,64,65,74-76] with various geometry proxies and neural features, explore view synthesis for dynamic scenes [12,21,23,43,57], and enhance the generalization of NeRF [14,19,29,33,40,44,63,68]. Many methods [30,41,61] have been proposed to exploit the view-dependent appearance to decouple the physical properties of the scene. NeRF2Mesh [61] proposes a" + } + ], + [ + { + "type": "page_number", + "bbox": [ + 0.218, + 0.116, + 0.232, + 0.127 + ], + "angle": 0, + "content": "4" + }, + { + "type": "header", + "bbox": [ + 0.272, + 0.115, + 0.437, + 0.129 + ], + "angle": 0, + "content": "Q. Zhang, Y. Feng et al." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.147, + 0.788, + 0.283 + ], + "angle": 0, + "content": "two-stage framework to extract the mesh and texture images by regularizing the view-dependent color. Many inverse rendering methods go a step further in material decoupling [9, 10, 32, 58, 62, 66, 69, 71, 77]. They decompose the radiance into physical properties such as albedo, roughness, metallicity, and lighting by introducing physical-based rendering into NeRF optimization, and enable relighting under novel lighting conditions or material editing. Numerous methods have been proposed to reconstruct scenes from images captured in uncontrolled environments, addressing challenges such as inaccurate poses [7, 16, 42, 51], image blur or noise [45, 48, 52], and inconsistent camera settings [27, 28, 55, 60]." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.283, + 0.788, + 0.631 + ], + "angle": 0, + "content": "Appearance control of NeRF Images captured in natural environments or taken by multiple cameras often exhibit inconsistency in brightness and color. These image-dependent appearances may derive from changing lighting conditions, exposure variations and white balancing, or differences in camera sensors and ISP, which are not desired in novel view synthesis. HDR-NeRF [28] focuses on multi-view images of varying exposures. By modeling the simplified physical imaging process with a tone mapper, which maps the radiance of a scene point to a pixel value, HDR-NeRF reconstructs the high dynamic range radiance field, which can generate novel views with consistent exposures. Many methods enable NeRF to handle these inconsistencies by introducing related parameters or embeddings. NeRF-w [46] and Ha-NeRF [15] propose to encode the image-dependent appearance using latent codes, which can then be manipulated to control the overall style of rendered images. Block-NeRF [60] feeds an additional exposure level to the appearance prediction model to deal with the brightness variance and provides the ability to modify the global exposure conditioning during inference. URF [55] predicts a \\(3 \\times 3\\) matrix to restrictively model the white balance and exposure variations. One shortcoming of these methods is that view-dependent appearances, those either caused by changes in camera settings and environment lighting or are intrinsical from scene physical properties, are coupled together and modeled related to both latent codes and view directions. In addition to the aforementioned methods that could control scene appearances by modifying materials or latent codes, some methods [38,39] engage in appearance editing by color post-manipulation and rely heavily on the consistency of views." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.631, + 0.788, + 0.751 + ], + "angle": 0, + "content": "Color Correction Color correction has been extensively studied on single 2D image and multi-camera systems. Single 2D image color correction [1,3,6,36, 37], or white balance correction, cannot guarantee view-consistent results when naively performed on each view. On the other hand, multi-camera color correction [20,56,72] relies on common area or statistical matching, which cannot handle inherent view variance and only matches the global style. To address these limitations, we propose integrating a color correction module into the NeRF optimization process to enable view-consistent color correction with 3D awareness." + }, + { + "type": "title", + "bbox": [ + 0.216, + 0.776, + 0.517, + 0.793 + ], + "angle": 0, + "content": "3 Color Correction for NeRF" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.81, + 0.787, + 0.84 + ], + "angle": 0, + "content": "Given multi-view images captured under various camera settings and lighting conditions, our method proposes a color correction module embedded in a NeRF-" + } + ], + [ + { + "type": "header", + "bbox": [ + 0.551, + 0.115, + 0.732, + 0.127 + ], + "angle": 0, + "content": "Color Correction for NeRF" + }, + { + "type": "page_number", + "bbox": [ + 0.775, + 0.117, + 0.785, + 0.127 + ], + "angle": 0, + "content": "5" + }, + { + "type": "image", + "bbox": [ + 0.219, + 0.149, + 0.784, + 0.285 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.214, + 0.297, + 0.788, + 0.425 + ], + "angle": 0, + "content": "Fig. 2: Overview of our pipeline. Given multi-view images captured under various camera settings and lighting conditions, our method recovers a unified color radiance field, which enables consistent novel view synthesis. We propose a non-linear color correction module embedded in NeRF-based model to simulate the color processing performed in cameras (e.g. white balance, exposure, nonlinear color enhancement, etc.), which maps the color from unified scene color space to each camera color space. The view-dependent color arises from inherent scene properties like specular and shading is extracted under regularization. We design a robust optimization scheme with tailored losses to achieve physically plausible color correction." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.453, + 0.788, + 0.531 + ], + "angle": 0, + "content": "based model, enabling the unified color NeRF reconstruction, as shown in Fig. 2. The scene radiance field is decomposed into two components to decouple the inherent color variances like specular and shading. The color correction module is designed as a \\(3 \\times 7\\) translation matrix in nonlinear RGB space. Multiple regularizations on pixels and areas are incorporated to improve color consistency." + }, + { + "type": "title", + "bbox": [ + 0.216, + 0.551, + 0.441, + 0.567 + ], + "angle": 0, + "content": "3.1 Scene Representation" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.577, + 0.789, + 0.701 + ], + "angle": 0, + "content": "Neural Radiance Field (NeRF) represents a scene using an implicit neural function, which utilizes MLPs to map the position \\(\\mathbf{x}\\) and view direction \\(\\mathbf{d}\\) to a density \\(\\sigma(\\mathbf{x})\\) and view-dependent color \\(\\mathbf{c}(\\mathbf{x},\\mathbf{d})\\). To render the color of a pixel in an image from a given camera pose, NeRF casts a ray \\(\\mathbf{r}(t) = \\mathbf{o} + t\\mathbf{d}\\) through the pixel, where \\(\\mathbf{o}\\) and \\(\\mathbf{d}\\) are the origin and direction of the ray, respectively, and \\(t\\) is the parameter that determines the position along the ray. The predicted \\(\\{(\\sigma,\\mathbf{c})\\}\\) of sampled positions on the ray are then accumulated into the color \\(C(\\mathbf{r})\\) of the pixel via volume rendering [47]:" + }, + { + "type": "equation", + "bbox": [ + 0.37, + 0.71, + 0.787, + 0.748 + ], + "angle": 0, + "content": "\\[\nC (\\mathbf {r}) = \\int_ {t _ {n}} ^ {t _ {f}} T (t) \\sigma (\\mathbf {r} (t)) \\mathbf {c} (\\mathbf {r} (t), \\mathbf {d}) d t, \\tag {1}\n\\]" + }, + { + "type": "equation", + "bbox": [ + 0.385, + 0.754, + 0.787, + 0.791 + ], + "angle": 0, + "content": "\\[\nT (t) = \\exp \\left(- \\int_ {t _ {n}} ^ {t} \\sigma (\\mathbf {r} (p)) d p\\right), \\tag {2}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.795, + 0.788, + 0.842 + ], + "angle": 0, + "content": "where \\( t_n \\) and \\( t_f \\) indicate the near and far boundary of the ray, respectively. \\( T(t) \\) refers to an accumulated transmittance, which quantifies the probability of the ray traveling from \\( t_n \\) to \\( t \\) without encountering other particles." + } + ], + [ + { + "type": "page_number", + "bbox": [ + 0.218, + 0.116, + 0.23, + 0.127 + ], + "angle": 0, + "content": "6" + }, + { + "type": "header", + "bbox": [ + 0.272, + 0.115, + 0.436, + 0.13 + ], + "angle": 0, + "content": "Q. Zhang, Y. Feng et al." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.147, + 0.788, + 0.224 + ], + "angle": 0, + "content": "We build our model on the NeRF framework, which is widely used for modeling scenes captured under static camera settings and lighting conditions. To eliminate the influence of inherent color variances, we propose to decompose the appearance into two components: view-independent diffuse color \\(\\mathbf{c}_d\\) and view-dependent color \\(\\mathbf{c}_s\\). The final color is obtained by summing the two terms:" + }, + { + "type": "equation", + "bbox": [ + 0.428, + 0.23, + 0.786, + 0.246 + ], + "angle": 0, + "content": "\\[\n\\mathbf {c} = \\mathbf {c} _ {d} (\\mathbf {x}) + \\mathbf {c} _ {s} (\\mathbf {x}, \\mathbf {d}). \\tag {3}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.215, + 0.25, + 0.789, + 0.298 + ], + "angle": 0, + "content": "By explicitly considering these two components, we can better model the appearance changes arising from surface properties and these caused by varying lighting conditions and camera settings." + }, + { + "type": "title", + "bbox": [ + 0.215, + 0.313, + 0.473, + 0.328 + ], + "angle": 0, + "content": "3.2 Color Correction Module" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.333, + 0.789, + 0.38 + ], + "angle": 0, + "content": "Digital cameras perform a number of processing steps to transform raw images to standard RGB images, including demosaicing, white balance, color correction and etc. A generic model [8] of this pipeline is formulated as:" + }, + { + "type": "equation", + "bbox": [ + 0.408, + 0.386, + 0.786, + 0.402 + ], + "angle": 0, + "content": "\\[\n\\mathbf {I} _ {s r g b} = f \\left(\\alpha \\mathbf {A} _ {x y z} \\mathbf {D} _ {w b} \\mathbf {I} _ {r a w}\\right), \\tag {4}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.407, + 0.789, + 0.497 + ], + "angle": 0, + "content": "where \\(\\alpha\\) is an exposure compensation common gain, \\(\\mathbf{D}_{wb}\\) represents the \\(3\\times 3\\) diagonal matrix for white balance, \\(\\mathbf{A}_{xyz}\\) is a \\(3\\times 3\\) linear transform of color space, and \\(f(\\cdot)\\) is a nonlinear function compounding color enhancement, tonemanipulation, and a final sRGB gamma encoding, etc. It's worth noting that the differences in white balance and lighting conditions become non-linear in the sRGB color space after undergoing these mappings." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.498, + 0.789, + 0.573 + ], + "angle": 0, + "content": "In our framework, we expect to recover the scene's radiance field with unified color when the input sRGB images exhibit color variances. Inspired by [2], our method introduces a polynomial color-mapping function to model the color deviation between the scene color space \\(\\mathbf{c}\\) and each camera color space \\(\\mathbf{c}^{cam}\\) captured by different cameras, which can be written as:" + }, + { + "type": "equation", + "bbox": [ + 0.447, + 0.579, + 0.786, + 0.595 + ], + "angle": 0, + "content": "\\[\n\\mathbf {c} ^ {c a m} = \\mathbf {M} \\varPhi (\\mathbf {c}), \\tag {5}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.6, + 0.789, + 0.737 + ], + "angle": 0, + "content": "where \\(\\varPhi\\) is the kernel mapping function that transforms the input color into the nonlinear space, specifically, \\(\\varPhi([R,G,B]^T) \\to [R,G,B,RG,RB,GB,1]^T\\). The matrix \\(\\mathbf{M}\\) is a color translation matrix of shape \\(3 \\times 7\\). Notably, we utilize an Embedding network \\(E\\) to predict a matrix \\(\\mathbf{M}\\) for each camera and lighting condition, and based on the physical imaging process, the matrix is shared by the diffuse and view-dependent components. The pixel colors of sampled points in the scene color space are first mapped to camera color space, and then accumulated through volume rendering. Combining Eq. (1), Eq. (3) and Eq. (5), the rendering equation can be written as:" + }, + { + "type": "equation", + "bbox": [ + 0.292, + 0.742, + 0.786, + 0.776 + ], + "angle": 0, + "content": "\\[\nC ^ {c a m} = \\int_ {t _ {n}} ^ {t _ {f}} T (t) \\sigma (\\mathbf {r} (t)) \\left(\\mathbf {M} \\varPhi \\left(\\mathbf {c} _ {d} (\\mathbf {r} (t))\\right) + \\mathbf {M} \\varPhi \\left(\\mathbf {c} _ {s} (\\mathbf {r} (t), \\mathbf {d})\\right)\\right) d t \\tag {6}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.78, + 0.789, + 0.84 + ], + "angle": 0, + "content": "where \\( T(t) \\) is defined in Eq. (2). By integrating the color correction module in the rendering pipeline, we can reconstruct the radiance field of the scene while accounting for color variations caused by different cameras and lighting conditions." + } + ], + [ + { + "type": "header", + "bbox": [ + 0.55, + 0.115, + 0.732, + 0.128 + ], + "angle": 0, + "content": "Color Correction for NeRF" + }, + { + "type": "page_number", + "bbox": [ + 0.775, + 0.116, + 0.787, + 0.127 + ], + "angle": 0, + "content": "7" + }, + { + "type": "title", + "bbox": [ + 0.216, + 0.147, + 0.551, + 0.161 + ], + "angle": 0, + "content": "3.3 View-consistent Color Distribution" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.186, + 0.788, + 0.277 + ], + "angle": 0, + "content": "Considering the color correction matrix is designed to be camera-wise, relying solely on pixel-level color loss hinders the accuracy and efficiency of optimization. Therefore we employ a cross-camera color distribution constraint. Given that digital cameras capture images with three color channels, each with a discrete range of \\( K \\) intensity values, we analyze the intensity distribution in each channel, which is formulated as an intensity histogram to constrain the color consistency." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.279, + 0.79, + 0.4 + ], + "angle": 0, + "content": "Drawing inspiration from the cross-camera color translation task, we transform the RGB color space \\( C^{cam} \\) into YUV color space \\( Y^{cam} \\in [-1,1]^3 \\). The YUV color space comprises one luma component (Y) and two chrominance components, U (blue projection) and V (red projection). We analyze each color channel independently and obtain the intensity histogram by counting the number of pixels of each intensity value. To obtain the intensity histogram as a smooth and differentiable function, we formulate one histogram distribution \\( \\mathbf{h}(Y^{cam}) \\) of an image color channel as" + }, + { + "type": "equation", + "bbox": [ + 0.299, + 0.417, + 0.787, + 0.462 + ], + "angle": 0, + "content": "\\[\n\\mathbf {h} = \\left\\{\\frac {1}{N} \\sum_ {x \\in Y} \\Theta \\left(\\frac {I - \\mu_ {k} + L / 2}{W}\\right) - \\Theta \\left(\\frac {I - \\mu_ {k} - L / 2}{W}\\right) \\right\\} _ {k = 0} ^ {K - 1}, \\tag {7}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.215, + 0.476, + 0.788, + 0.552 + ], + "angle": 0, + "content": "where \\(K\\) is the number of bins set in the histogram, and \\(L\\) indicates the length of each bin, which also equals \\(\\frac{2}{K}\\). Given the \\(k\\)-th bin, its center and bandwidth refer to \\(\\mu_k = -1 + L(k + \\frac{1}{2})\\) and \\(W = L / 2.5\\), respectively. \\(\\Theta(x)\\) is a Sigmoid function \\((1 + \\exp(-x))^{-1}\\) defining a differentiable approximation of the histogram distribution." + }, + { + "type": "text", + "bbox": [ + 0.215, + 0.555, + 0.788, + 0.631 + ], + "angle": 0, + "content": "Histogram Distance. To incorporate color distribution constraints into our framework and facilitate easy convergence, we propose using the Euclidean distance between the cumulative histograms to measure the distance between two image histograms. Let \\(\\mathbf{h}_s\\) and \\(\\mathbf{h}_r\\) represent the histograms of the rendered and the reference images, respectively. The distance between histograms is defined:" + }, + { + "type": "equation", + "bbox": [ + 0.356, + 0.648, + 0.787, + 0.688 + ], + "angle": 0, + "content": "\\[\n\\mathcal {D} _ {h} \\left(\\mathbf {h} _ {s}, \\mathbf {h} _ {r}\\right) = \\sum_ {i = 1} ^ {K} \\left(\\mathrm {C D F} _ {i} \\left(\\mathbf {h} _ {s}\\right), \\mathrm {C D F} _ {i} \\left(\\mathbf {h} _ {r}\\right)\\right), \\tag {8}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.215, + 0.704, + 0.789, + 0.841 + ], + "angle": 0, + "content": "where \\(\\mathrm{CDF}_i(\\mathbf{h})\\) denotes the \\(i\\)-th element of the cumulative density function of histogram. The histogram distance \\(\\mathcal{D}_h\\) measures the global similarity of color distribution between two images. Unlike NeRF, which focuses on individual pixel values, this approach takes into account the entire color distribution of the images. By considering the cumulative histograms, we can effectively compare and align the color distribution of two different images. This property is particularly useful for color translation tasks with our color correction module. We can achieve consistent color appearance across different images by adapting the color distribution towards the reference image." + } + ], + [ + { + "type": "page_number", + "bbox": [ + 0.218, + 0.116, + 0.23, + 0.127 + ], + "angle": 0, + "content": "8" + }, + { + "type": "header", + "bbox": [ + 0.272, + 0.115, + 0.436, + 0.129 + ], + "angle": 0, + "content": "Q. Zhang, Y. Feng et al." + }, + { + "type": "title", + "bbox": [ + 0.216, + 0.147, + 0.373, + 0.162 + ], + "angle": 0, + "content": "3.4 Optimization" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.17, + 0.784, + 0.214 + ], + "angle": 0, + "content": "To regularize the reconstructed scene color and reduce ambiguity, we design a series of losses on both pixel and image levels, including color reconstruction loss, specular loss, histogram loss, and gray-world loss." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.215, + 0.785, + 0.306 + ], + "angle": 0, + "content": "Color Reconstruction Loss. To jointly optimize the scene's NeRF and the camera-dependent color correction matrix \\(\\{\\mathbf{M}_i\\}_{i=1}^N\\), color reconstruction loss is proposed to minimize the mean squared error (MSE) of sampled pixels between the ground truth color \\(C^{cam}\\) and the rendered color \\(\\hat{C}^{cam}\\) which is mapped from scene color by corresponding matrix (see Eq. (5) and Eq. (6)). The color reconstruction loss is formulated as:" + }, + { + "type": "equation", + "bbox": [ + 0.393, + 0.313, + 0.786, + 0.35 + ], + "angle": 0, + "content": "\\[\n\\mathcal {L} _ {c} = \\sum_ {\\mathbf {r} \\in \\mathcal {R} (\\mathbf {P})} \\left\\| \\widehat {C} ^ {c a m} - C ^ {c a m} \\right\\| _ {2} ^ {2}. \\tag {9}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.358, + 0.788, + 0.493 + ], + "angle": 0, + "content": "Specular Loss. According to Eq. (3), we decompose the scene's color into diffuse color \\(\\mathbf{c}_d\\) and view-dependent color \\(\\mathbf{c}_s\\). This view-dependent color accounts for the color variances due to the inherent material properties and geometries. Both components share a common color correction matrix \\(\\mathbf{M}\\). Ideally, we suppose the cross-image differences are mostly caused by varying light conditions and camera-dependent color translation. Thus, we apply a regularization on \\(\\mathbf{c}_s\\), which benefits color consistency on the diffuse partition and estimation of color correction matrix, as shown in ablation studies. The L2 regularization on the view-dependent color is formulated as:" + }, + { + "type": "equation", + "bbox": [ + 0.436, + 0.502, + 0.786, + 0.54 + ], + "angle": 0, + "content": "\\[\n\\mathcal {L} _ {s} = \\sum_ {\\mathbf {r} \\in \\mathcal {R} (\\mathbf {P})} \\left\\| \\widehat {C} _ {s} \\right\\| _ {2} ^ {2} \\tag {10}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.215, + 0.546, + 0.788, + 0.651 + ], + "angle": 0, + "content": "Gray-world Loss. Due to the correlation between scene color space and color correction matrix, the recovered scene color \\(\\mathbf{c}\\) is under-constrained with only \\(\\mathcal{L}_g\\). To address this, white balance correction methods often refer to some prior knowledge of the spectral properties of the camera sensors. One commonly used heuristic prior is the \"gray-world\" assumption: the color in the world tends to be natural gray on average [11]. We also define an L2 gray-world regularization for the scene's radiance field:" + }, + { + "type": "equation", + "bbox": [ + 0.402, + 0.658, + 0.786, + 0.695 + ], + "angle": 0, + "content": "\\[\n\\mathcal {L} _ {g} = \\frac {1}{N _ {\\mathbf {r}}} \\sum_ {\\mathbf {r} \\in \\mathcal {R} (\\mathbf {P})} \\left\\| \\widehat {C} - C _ {0} \\right\\| _ {2} ^ {2}, \\tag {11}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.702, + 0.785, + 0.731 + ], + "angle": 0, + "content": "where \\( C_0 \\) is a hyperparameter which is generally set as the midway of the color values such as 0.5." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.733, + 0.788, + 0.809 + ], + "angle": 0, + "content": "Histogram Loss. To further constrain the color consistency of reconstructed radiance fields in different views, besides rendering image \\(\\widehat{C}^{cam}\\) from the ground truth view, we render an additional image from random view and obtain its histogram \\(\\widehat{\\mathbf{h}}_r^{cam}\\) during training. Based on \\(\\widehat{\\mathbf{h}}_r^{cam}\\), \\(\\widehat{\\mathbf{h}}^{cam}\\) of \\(\\widehat{C}^{cam}\\) and \\(\\mathbf{h}^{cam}\\) of ground truth color \\(C^{cam}\\), we design a cross histogram loss based on Eq. (8)," + }, + { + "type": "equation", + "bbox": [ + 0.285, + 0.817, + 0.786, + 0.845 + ], + "angle": 0, + "content": "\\[\n\\mathcal {L} _ {h} = \\frac {1}{3} \\left(\\mathcal {D} _ {h} \\left(\\widehat {\\mathbf {h}} ^ {c a m}, \\mathbf {h} ^ {c a m}\\right) + \\mathcal {D} _ {h} \\left(\\widehat {\\mathbf {h}} _ {r} ^ {c a m}, \\mathbf {h} ^ {c a m}\\right) + \\mathcal {D} _ {h} \\left(\\widehat {\\mathbf {h}} ^ {c a m}, \\widehat {\\mathbf {h}} _ {r} ^ {c a m}\\right)\\right). \\tag {12}\n\\]" + } + ], + [ + { + "type": "header", + "bbox": [ + 0.55, + 0.115, + 0.732, + 0.128 + ], + "angle": 0, + "content": "Color Correction for NeRF" + }, + { + "type": "page_number", + "bbox": [ + 0.775, + 0.116, + 0.787, + 0.128 + ], + "angle": 0, + "content": "9" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.147, + 0.784, + 0.223 + ], + "angle": 0, + "content": "Considering that gray-world loss \\(\\mathcal{L}_g\\) and histogram loss \\(\\mathcal{L}_h\\) are defined on color distributions, we sample rays from grids and combine them to form an image instead of rendering the entire image during training for converge efficiency. This is based on the assumption that the color distribution of an image will remain unchanged after the random grid sampling." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.224, + 0.784, + 0.256 + ], + "angle": 0, + "content": "Total Loss. The complete loss \\(\\mathcal{L}\\) is a weighted sum of the above loss functions from Eq. (9) to Eq. (12):" + }, + { + "type": "equation", + "bbox": [ + 0.389, + 0.272, + 0.786, + 0.289 + ], + "angle": 0, + "content": "\\[\n\\mathcal {L} = \\mathcal {L} _ {c} + \\lambda_ {s} \\mathcal {L} _ {s} + \\lambda_ {h} \\mathcal {L} _ {h} + \\lambda_ {g} \\mathcal {L} _ {g}, \\tag {13}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.215, + 0.302, + 0.768, + 0.318 + ], + "angle": 0, + "content": "where \\(\\lambda_s = 0.1\\), \\(\\lambda_h = 0.01\\), and \\(\\lambda_g = 0.001\\) are the weights of different losses." + }, + { + "type": "title", + "bbox": [ + 0.215, + 0.35, + 0.376, + 0.367 + ], + "angle": 0, + "content": "4 Experiments" + }, + { + "type": "title", + "bbox": [ + 0.215, + 0.391, + 0.457, + 0.406 + ], + "angle": 0, + "content": "4.1 Implementation Details" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.425, + 0.788, + 0.592 + ], + "angle": 0, + "content": "Our implementation of NeRF follows [67]. Note that our camera correction module could be embedded in different NeRF-based models [4,13,50]. An eight-layer MLP with 256 channels with encoding of position \\(\\mathbf{x}\\) as input is used to generate \\(\\sigma\\), and followed by one additional 256-channel fully-connected layer to extract a feature. Diffuse color \\(\\mathbf{c}_d\\) is obtained by mapping the feature to RGB channels. The view-dependent color \\(\\mathbf{c}_s\\) is obtained through an additional 128-channel fully connected layer and a mapping layer, with the aforementioned feature and encoding of view direction \\(\\mathbf{d}\\) as input. Both \\(\\mathbf{c}_d\\) and \\(\\mathbf{c}_s\\) are transformed to nonlinear space by the kernel mapping function \\(\\varPhi\\) in Eq. (5). The embedding vector is reshaped to color correction matrix \\(\\mathbf{M}\\) of size \\(3\\times 7\\). Models are trained over 300,000 iterations with a batch size of 1024." + }, + { + "type": "title", + "bbox": [ + 0.215, + 0.624, + 0.525, + 0.638 + ], + "angle": 0, + "content": "4.2 Evaluation Dataset and Metrics" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.657, + 0.788, + 0.764 + ], + "angle": 0, + "content": "Dataset. We evaluate the proposed method on our collected dataset. The dataset consists of 6 real scenes captured by three digital cameras with different camera settings (e.g. white balance, ISO, exposures, etc.) and light conditions (e.g. non-flashing, and flashing with different lights). Images are collected at 90 different poses for each scene. We select 60 views with different poses as the training dataset. The remaining images are used as test images, corresponding to the same setting with training data. The resolution is \\(1000 \\times 667\\) pixels." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.765, + 0.788, + 0.841 + ], + "angle": 0, + "content": "Metrics. We report quantitative performance using PSNR (higher is better) and SSIM (higher is better) metrics, as well as the state-of-the-art LPIPS (lower is better) perceptual metric [70]. To better evaluate the color consistency and correction, we propose a \\( W_{2} \\) metric based on Wasserstein distance, which is a distance function defined between probability distributions on the entire color" + } + ], + [ + { + "type": "page_number", + "bbox": [ + 0.218, + 0.116, + 0.236, + 0.127 + ], + "angle": 0, + "content": "10" + }, + { + "type": "header", + "bbox": [ + 0.272, + 0.115, + 0.435, + 0.129 + ], + "angle": 0, + "content": "Q. Zhang, Y. Feng et al." + }, + { + "type": "table", + "bbox": [ + 0.22, + 0.144, + 0.784, + 0.24 + ], + "angle": 0, + "content": "
ScenePSNR ↑SSIM ↑
NeRFNeRF-WHa-NeRFOursNeRFNeRF-WHa-NeRFOurs
Hamster15.62928.38128.28129.5710.9310.9510.9580.959
Pumpkin17.16230.18930.09631.3590.8780.9430.9420.956
Letters16.78028.04328.38730.4520.7830.9200.9220.955
MailBox18.85325.77225.90328.1280.9000.8950.8960.922
Pipe14.72522.20125.98227.5540.7440.7570.8560.907
Room16.62914.77214.96531.0970.9320.7960.7730.969
Mean16.62924.89325.60229.6940.8610.8770.8910.945
" + }, + { + "type": "table", + "bbox": [ + 0.221, + 0.246, + 0.784, + 0.34 + ], + "angle": 0, + "content": "
SceneLPIPS ↓\\( W_2 \\) (1e-3) ↓
NeRFNeRF-WHa-NeRFOursNeRFNeRF-WHa-NeRFOurs
Hamster0.5080.2190.2250.19974.42.582.592.31
Pumpkin0.5970.2340.2350.18161.11.271.400.53
Letters0.4490.3210.3240.21676.41.110.950.68
MailBox0.4390.2660.2690.20544.95.744.133.71
Pipe0.6210.4760.2790.17771.41.500.600.42
Room0.4960.6650.6680.10389.513.013.56.16
Mean0.5180.3630.3330.18069.64.203.862.30
" + }, + { + "type": "table_caption", + "bbox": [ + 0.216, + 0.34, + 0.785, + 0.381 + ], + "angle": 0, + "content": "Table 1: Quantitative comparisons with baseline methods on our collected scenes, including PSNR, SSIM, LPIPS, and \\( W_{2} \\) metrics. The Wasserstein distance \\( W_{2} \\) measures the difference of entire color distributions." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.421, + 0.627, + 0.436 + ], + "angle": 0, + "content": "space. The image-level \\( W_{2} \\) (lower is better) is defined as:" + }, + { + "type": "equation", + "bbox": [ + 0.233, + 0.444, + 0.495, + 0.48 + ], + "angle": 0, + "content": "\\[\nW _ {2} = \\frac {1}{N} \\sum_ {i} ^ {N} \\mathrm {W a s s e r s t e i n} (\\widehat {C} _ {i} ^ {c a m}, C ^ {t r a i n}),\n\\]" + }, + { + "type": "equation", + "bbox": [ + 0.233, + 0.484, + 0.785, + 0.51 + ], + "angle": 0, + "content": "\\[\n\\text {W a s s e r s t e i n} \\left(\\widehat {C} _ {i} ^ {\\text {c a m}}, C ^ {\\text {t r a i n}}\\right) = \\left\\| \\operatorname {m e a n} \\left(\\widehat {C} _ {i} ^ {\\text {c a m}}\\right) - \\operatorname {m e a n} \\left(C ^ {\\text {t r a i n}}\\right) \\right\\| _ {2} ^ {2} + \\tag {14}\n\\]" + }, + { + "type": "equation", + "bbox": [ + 0.233, + 0.512, + 0.741, + 0.55 + ], + "angle": 0, + "content": "\\[\n\\mathrm {t r a c e} \\left(\\mathrm {c o v} (\\widehat {C} _ {i} ^ {c a m}) + \\mathrm {c o v} (C ^ {t r a i n}) - 2 \\sqrt {\\sqrt {\\mathrm {c o v} (\\widehat {C} _ {i} ^ {c a m})} \\mathrm {c o v} (C ^ {t r a i n}) \\sqrt {\\mathrm {c o v} (\\widehat {C} _ {i} ^ {c a m})}}\\right),\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.556, + 0.785, + 0.603 + ], + "angle": 0, + "content": "where \\(\\mathrm{mean}(\\cdot)\\) and \\(\\mathrm{cov}(\\cdot)\\) refer to the mean and covariance of an image. \\(C^{train}\\) denotes the training image. \\(\\hat{C}^{cam}\\) indicates the rendered images those translated with matrix of \\(C^{train}\\)." + }, + { + "type": "title", + "bbox": [ + 0.216, + 0.623, + 0.351, + 0.636 + ], + "angle": 0, + "content": "4.3 Evaluation" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.644, + 0.784, + 0.718 + ], + "angle": 0, + "content": "Baselines. We compare our method against the following baseline methods. 1) NeRF [49]: the original NeRF method, also our backbone network. 2) NeRF-W [46]: unofficial implementation where the dimension of latent code is 48. 3) Ha-NeRF [15]: official implementation which can use a reference image to do color translation." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.72, + 0.785, + 0.841 + ], + "angle": 0, + "content": "Comparisons. The qualitative results of rendered novel views on our own dataset are shown in Fig. 3. Compared to baseline methods, our method achieves promising results on view rendering quality (see MSE error) and color correction measured by Wasserstein distance between ground truths and rendered images. Specifically, NeRF, which is directly reconstructed from inputs under various conditions, tends to render views with significant artifacts. NeRF-W and Ha-NeRF do ensure color consistency to a certain extent, as shown in Fig. 3. However, these methods can not distinguish the implicit properties (e.g. specular)" + } + ], + [ + { + "type": "header", + "bbox": [ + 0.551, + 0.115, + 0.732, + 0.127 + ], + "angle": 0, + "content": "Color Correction for NeRF" + }, + { + "type": "page_number", + "bbox": [ + 0.769, + 0.116, + 0.784, + 0.127 + ], + "angle": 0, + "content": "11" + }, + { + "type": "image", + "bbox": [ + 0.22, + 0.144, + 0.788, + 0.407 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.216, + 0.418, + 0.788, + 0.53 + ], + "angle": 0, + "content": "Fig. 3: Qualitative results of our method compared with baseline methods. We also show the error map and histogram for better visualization. MSE and Wasserstein values are on the top corners of the error map and histogram, respectively. Our method outperforms baselines, providing superior rendering quality and effective color correction. NeRF [49] generates significant artifacts. NeRF-W [46] and Ha-NeRF [15] achieve color correction to a certain extent, but exhibit shortcomings in areas with high illumination (e.g. the leaf of scene Mailbox and 'E' of scene Letters) and generate artifacts (e.g. the window of scene Mailbox and floor of scene Hamster)." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.565, + 0.788, + 0.672 + ], + "angle": 0, + "content": "and external differences (e.g. camera settings). Their performance is limited in highlighting regions on Letter 'E' in scene Letters and the leaf in scene Mailbox, and results in artifacts on the window in scene Mailbox and the floor in scene Hamster. In contrast, our method outperforms state-of-the-art methods, providing superior rendering quality and effective color correction. It demonstrates the effectiveness of our color correction module, which physically defines color differences by a nonlinear color translation matrix." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.675, + 0.789, + 0.842 + ], + "angle": 0, + "content": "Tab. 1 presents the quantitative results of our method compared to baseline approaches on our collected dataset, showcasing the superior rendering quality achieved by our method. Notably, NeRF-W and Ha-NeRF perform worse than the original NeRF on the scene Room. We analyze this can be attributed to the fact that the scene \"Room\" presents particular challenges with large reflective areas, which can be difficult for these two methods to fit a reasonable solution, and qualitative results are included in the supplementary material. Tab. 1 summarizes the quantitative comparisons on color consistency measured by Eq. (14). Wasserstein distance measures color consistency by calculating the color distribution distance between the input image and the rendered image translated with the color correction matrix of the input. Compared to the pixel-level PSNR" + } + ], + [ + { + "type": "page_number", + "bbox": [ + 0.218, + 0.116, + 0.236, + 0.127 + ], + "angle": 0, + "content": "12" + }, + { + "type": "header", + "bbox": [ + 0.272, + 0.115, + 0.437, + 0.129 + ], + "angle": 0, + "content": "Q. Zhang, Y. Feng et al." + }, + { + "type": "image", + "bbox": [ + 0.216, + 0.143, + 0.788, + 0.319 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.214, + 0.33, + 0.788, + 0.375 + ], + "angle": 0, + "content": "Fig. 4: Ablation study of gray-world loss \\(\\mathcal{L}_g\\). Removing \\(\\mathcal{L}_g\\) will be unable to constrain the distribution of the scene color space, resulting in undesirable artifacts or color shifts in the reconstructed scene (e.g. overexposed Hamster and over-saturated Pumpkin)." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.402, + 0.789, + 0.464 + ], + "angle": 0, + "content": "metric, the Wasserstein metric measures the entire color distribution. These comparative results demonstrate the effectiveness of our approach in addressing the limitations of existing NeRF-based models, ultimately leading to enhanced color correction and consistency in various camera conditions." + }, + { + "type": "title", + "bbox": [ + 0.216, + 0.484, + 0.403, + 0.498 + ], + "angle": 0, + "content": "4.4 Ablation Studies" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.508, + 0.789, + 0.719 + ], + "angle": 0, + "content": "We conduct ablative evaluations on our collected dataset to show the effectiveness of our model design. We compare our model with four variants: model without specular loss (Ours w/o \\(\\mathcal{L}_s\\)), model without gray-world loss (Ours w/o \\(\\mathcal{L}_g\\)), model without histogram loss (Ours w/o \\(\\mathcal{L}_h\\)), and model without nonlinear mapping (Ours w/o \\(\\varPhi\\)). The quantitative comparisons presented in Fig. 5 indicate that each component is crucial for generating physically plausible color-corrected renderings. Additionally, Fig. 4 displays the qualitative results of the ablation study on view synthesis with the unified color NeRF. This further highlights the importance of gray-world loss in our model. When the gray-world loss is removed, the color distribution of the scene color space becomes unconstrained, and the increased solution space can be detrimental to convergence. While this may not significantly impact the rendering quality and color consistency of novel views, it could introduce undesirable artifacts or color shifts in the reconstructed scene, such as overexposed regions of Hamster shown in Fig. 4." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.72, + 0.79, + 0.841 + ], + "angle": 0, + "content": "In addition, we provide qualitative demonstrations of the results with/without specular loss \\(\\mathcal{L}_s\\), histogram loss \\(\\mathcal{L}_h\\) and nonlinear mapping \\(\\varPhi\\). The left bottom half of the image represents the view-dependent color \\(C_s^{cam}\\), and the remaining part displays the entire renderings \\(C^{cam}\\) in camera color space of ground truth. Specifically, removing specular loss will not decompose the view-dependent and view-independent colors of Letters, resulting in a significant break in color consistency (also shown in \\(W_2\\) metric). Eliminating the histogram loss makes it challenging to represent the entire color distribution when relying solely on pixel-level" + } + ], + [ + { + "type": "header", + "bbox": [ + 0.551, + 0.115, + 0.732, + 0.127 + ], + "angle": 0, + "content": "Color Correction for NeRF" + }, + { + "type": "page_number", + "bbox": [ + 0.769, + 0.116, + 0.785, + 0.127 + ], + "angle": 0, + "content": "13" + }, + { + "type": "image", + "bbox": [ + 0.218, + 0.144, + 0.406, + 0.241 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.28, + 0.241, + 0.344, + 0.25 + ], + "angle": 0, + "content": "Ground Truth" + }, + { + "type": "image", + "bbox": [ + 0.406, + 0.144, + 0.595, + 0.241 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.472, + 0.241, + 0.53, + 0.25 + ], + "angle": 0, + "content": "Ours w/o \\(\\mathcal{L}_s\\)" + }, + { + "type": "image", + "bbox": [ + 0.596, + 0.144, + 0.785, + 0.242 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.66, + 0.241, + 0.719, + 0.25 + ], + "angle": 0, + "content": "Ours w/o \\(\\mathcal{L}_h\\)" + }, + { + "type": "image", + "bbox": [ + 0.217, + 0.251, + 0.405, + 0.347 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.254, + 0.348, + 0.373, + 0.358 + ], + "angle": 0, + "content": "Quantitative Comparisons" + }, + { + "type": "image", + "bbox": [ + 0.406, + 0.251, + 0.595, + 0.348 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.474, + 0.348, + 0.528, + 0.357 + ], + "angle": 0, + "content": "Ours w/o \\(\\Phi\\)" + }, + { + "type": "image", + "bbox": [ + 0.596, + 0.251, + 0.785, + 0.348 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.678, + 0.348, + 0.702, + 0.357 + ], + "angle": 0, + "content": "Ours" + }, + { + "type": "image_caption", + "bbox": [ + 0.214, + 0.369, + 0.788, + 0.466 + ], + "angle": 0, + "content": "Fig. 5: Ablation Study on \"Letter\", including qualitative and quantitative results with/without specular loss \\(\\mathcal{L}_s\\), histogram loss \\(\\mathcal{L}_h\\) and nonlinear mapping \\(\\varPhi\\). The left bottom half of the image is the view-dependent color \\(C_s^{cam}\\), and the other part is the entire color-correct renderings \\(C^{cam}\\). Removing specular loss struggles to the decomposition of view-dependent and view-independent colors. Removing histogram loss lacks the constraint of the entire color space, making the color offset. Downgrading the color translation matrix suffers from artifacts of the nonlinear color difference." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.497, + 0.788, + 0.603 + ], + "angle": 0, + "content": "MSE error. This may result in color shifts, such as producing darker images than the ground truth. Removing nonlinear mapping and downgrading color translation matrix to \\(3 \\times 3\\) will neglect the nonlinear process during imaging. This forces the network to model more color variances into \\(\\mathbf{C}_s^{cam}\\), making the color correction module unable to fully perceive the color characteristics of each camera, as shown in Fig. 5. These ablation studies demonstrate the effectiveness of our model and specifically designed losses." + }, + { + "type": "title", + "bbox": [ + 0.216, + 0.625, + 0.409, + 0.639 + ], + "angle": 0, + "content": "4.5 Color Translation" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.65, + 0.788, + 0.711 + ], + "angle": 0, + "content": "Fig. 6 presents examples of color translation guided by a reference image. This advantage makes our method particularly suitable for applications where accurate color reproduction and adherence to reference images are crucial, such as in virtual reality, computer graphics, and image synthesis." + }, + { + "type": "title", + "bbox": [ + 0.216, + 0.733, + 0.36, + 0.75 + ], + "angle": 0, + "content": "5 Conclusion" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.765, + 0.788, + 0.841 + ], + "angle": 0, + "content": "This paper proposes a framework that incorporates a color correction module within a NeRF to achieve unified color radiance field reconstruction from multiview images captured under different camera settings and lighting conditions. Specifically, we have developed a color correction module to predict a nonlinear color translation matrix that simulates the physical imaging process. Besides the" + } + ], + [ + { + "type": "page_number", + "bbox": [ + 0.218, + 0.116, + 0.236, + 0.127 + ], + "angle": 0, + "content": "14" + }, + { + "type": "header", + "bbox": [ + 0.272, + 0.115, + 0.437, + 0.129 + ], + "angle": 0, + "content": "Q. Zhang, Y. Feng et al." + }, + { + "type": "image", + "bbox": [ + 0.225, + 0.143, + 0.787, + 0.519 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.214, + 0.529, + 0.788, + 0.573 + ], + "angle": 0, + "content": "Fig. 6: Examples of color translation. Our method produces spatially and temporally consistent color transfer results that faithfully match the reference image while maintaining photorealism. Video can also be found in supplementary materials." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.62, + 0.788, + 0.727 + ], + "angle": 0, + "content": "color correction module for accommodating external differences across distinct cameras, we predict a view-dependent function to separate inherent differences caused by scene properties. To enhance the accuracy of the color translation, we have designed various regularization strategies and demonstrated their efficacy through experiments. Our approach significantly improves color consistency in NeRF reconstruction, enabling users to render novel views under flexible and unrestricted settings." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.735, + 0.788, + 0.842 + ], + "angle": 0, + "content": "Despite its advantages, we acknowledge that our framework has certain limitations. For instance, compared to NeRF, the proposed Histogram loss results in an increased training time of approximately \\(10\\%\\). Furthermore, the Histogram loss, which computes the difference between 2D images, performs well when the rendered image and reference image are nearby. This limitation could potentially be addressed by introducing a new representation and a 3D Histogram loss. We plan to explore these improvements in future work." + } + ], + [ + { + "type": "header", + "bbox": [ + 0.55, + 0.115, + 0.732, + 0.127 + ], + "angle": 0, + "content": "Color Correction for NeRF" + }, + { + "type": "page_number", + "bbox": [ + 0.769, + 0.117, + 0.786, + 0.127 + ], + "angle": 0, + "content": "15" + }, + { + "type": "title", + "bbox": [ + 0.217, + 0.145, + 0.323, + 0.16 + ], + "angle": 0, + "content": "References" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.179, + 0.785, + 0.223 + ], + "angle": 0, + "content": "1. Affi, M., Barron, J.T., LeGendre, C., Tsai, Y.T., Bleibel, F.: Cross-camera convolutional color constancy. In: Proceedings of the IEEE/CVF International Conference on Computer Vision. pp. 1981-1990 (2021)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.223, + 0.785, + 0.25 + ], + "angle": 0, + "content": "2. Affi, M., Brown, M.S.: Interactive white balancing for camera-rendered images. arXiv preprint arXiv:2009.12632 (2020)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.251, + 0.785, + 0.292 + ], + "angle": 0, + "content": "3. Affifi, M., Brubaker, M.A., Brown, M.S.: Auto white-balance correction for mixed-illuminant scenes. In: Proceedings of the IEEE/CVF Winter Conference on Applications of Computer Vision. pp. 1210–1219 (2022)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.293, + 0.785, + 0.348 + ], + "angle": 0, + "content": "4. Barron, J.T., Mildenhall, B., Tancik, M., Hedman, P., Martin-Brualla, R., Srinivasan, P.P.: Mip-nerf: A multiscale representation for anti-aliasing neural radiance fields. In: Proceedings of the IEEE/CVF International Conference on Computer Vision. pp. 5855-5864 (2021)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.349, + 0.785, + 0.39 + ], + "angle": 0, + "content": "5. Barron, J.T., Mildenhall, B., Verbin, D., Srinivasan, P.P., Hedman, P.: Zip-nerf: Anti-aliased grid-based neural radiance fields. arXiv preprint arXiv:2304.06706 (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.391, + 0.785, + 0.419 + ], + "angle": 0, + "content": "6. Barron, J.T., Tsai, Y.T.: Fast fourier color constancy. In: Proceedings of the IEEE conference on computer vision and pattern recognition. pp. 886-894 (2017)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.42, + 0.785, + 0.461 + ], + "angle": 0, + "content": "7. Bian, W., Wang, Z., Li, K., Bian, J.W., Prisacariu, V.A.: Nope-nerf: Optimising neural radiance field with no pose prior. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 4160-4169 (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.462, + 0.785, + 0.503 + ], + "angle": 0, + "content": "8. Bianco, S., Bruna, A., Naccari, F., Schettini, R.: Color space transformations for digital photography exploiting information about the illuminant estimation process. JOSA A 29(3), 374-384 (2012)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.504, + 0.785, + 0.545 + ], + "angle": 0, + "content": "9. Boss, M., Braun, R., Jampani, V., Barron, J.T., Liu, C., Lensch, H.: Nerd: Neural reflectance decomposition from image collections. In: Proceedings of the IEEE/CVF International Conference on Computer Vision. pp. 12684-12694 (2021)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.546, + 0.785, + 0.587 + ], + "angle": 0, + "content": "0. Boss, M., Jampani, V., Braun, R., Liu, C., Barron, J., Lensch, H.: Neural-pil: Neural pre-integrated lighting for reflectance decomposition. Advances in Neural Information Processing Systems 34, 10691-10704 (2021)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.588, + 0.785, + 0.616 + ], + "angle": 0, + "content": "1. Buchsbaum, G.: A spatial processor model for object colour perception. Journal of the Franklin institute 310(1), 1-26 (1980)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.617, + 0.785, + 0.657 + ], + "angle": 0, + "content": "2. Cao, A., Johnson, J.: Hexplane: A fast representation for dynamic scenes. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 130-141 (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.659, + 0.785, + 0.686 + ], + "angle": 0, + "content": "3. Chen, A., Xu, Z., Geiger, A., Yu, J., Su, H.: Tensorf: Tensorial radiance fields. In: European Conference on Computer Vision. pp. 333-350. Springer (2022)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.687, + 0.785, + 0.741 + ], + "angle": 0, + "content": "4. Chen, A., Xu, Z., Zhao, F., Zhang, X., Xiang, F., Yu, J., Su, H.: Mvsnerf: Fast generalizable radiance field reconstruction from multi-view stereo. In: Proceedings of the IEEE/CVF international conference on computer vision. pp. 14124-14133 (2021)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.742, + 0.785, + 0.784 + ], + "angle": 0, + "content": "5. Chen, X., Zhang, Q., Li, X., Chen, Y., Feng, Y., Wang, X., Wang, J.: Hallucinated neural radiance fields in the wild. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 12943-12952 (2022)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.785, + 0.785, + 0.84 + ], + "angle": 0, + "content": "6. Chen, Y., Chen, X., Wang, X., Zhang, Q., Guo, Y., Shan, Y., Wang, F.: Local-to-global registration for bundle-adjusting neural radiance fields. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 8264-8273 (2023)" + }, + { + "type": "list", + "bbox": [ + 0.226, + 0.179, + 0.785, + 0.84 + ], + "angle": 0, + "content": null + } + ], + [ + { + "type": "page_number", + "bbox": [ + 0.218, + 0.116, + 0.236, + 0.127 + ], + "angle": 0, + "content": "16" + }, + { + "type": "header", + "bbox": [ + 0.272, + 0.115, + 0.435, + 0.129 + ], + "angle": 0, + "content": "Q. Zhang, Y. Feng et al." + }, + { + "type": "ref_text", + "bbox": [ + 0.218, + 0.147, + 0.787, + 0.204 + ], + "angle": 0, + "content": "17. Chen, Z., Funkhouser, T., Hedman, P., Tagliasacchi, A.: Mobilenerf: Exploiting the polygon rasterization pipeline for efficient neural field rendering on mobile architectures. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 16569-16578 (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.218, + 0.205, + 0.787, + 0.26 + ], + "angle": 0, + "content": "18. Cheng, W., Chen, R., Fan, S., Yin, W., Chen, K., Cai, Z., Wang, J., Gao, Y., Yu, Z., Lin, Z., et al.: Dna-rendering: A diverse neural actor repository for high-fidelity human-centric rendering. In: Proceedings of the IEEE/CVF International Conference on Computer Vision. pp. 19982-19993 (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.261, + 0.787, + 0.316 + ], + "angle": 0, + "content": "19. Chibane, J., Bansal, A., Lazova, V., Pons-Moll, G.: Stereo radiance fields (srf): Learning view synthesis for sparse views of novel scenes. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 7911-7920 (2021)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.317, + 0.787, + 0.358 + ], + "angle": 0, + "content": "20. Ding, C., Ma, Z.: Multi-camera color correction via hybrid histogram matching. IEEE Transactions on Circuits and Systems for Video Technology 31(9), 3327-3337 (2020)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.36, + 0.787, + 0.402 + ], + "angle": 0, + "content": "21. Fang, J., Yi, T., Wang, X., Xie, L., Zhang, X., Liu, W., Nießner, M., Tian, Q.: Fast dynamic radiance fields with time-aware neural voxels. In: SIGGRAPH Asia 2022 Conference Papers. pp. 1-9 (2022)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.402, + 0.787, + 0.431 + ], + "angle": 0, + "content": "22. Fang, Q., Shuai, Q., Dong, J., Bao, H., Zhou, X.: Reconstructing 3d human pose by watching humans in the mirror. In: CVPR (2021)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.431, + 0.787, + 0.487 + ], + "angle": 0, + "content": "23. Fridovich-Keil, S., Meanti, G., Warburg, F.R., Recht, B., Kanazawa, A.: K-planes: Explicit radiance fields in space, time, and appearance. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 12479-12488 (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.488, + 0.787, + 0.53 + ], + "angle": 0, + "content": "24. Fridovich-Keil, S., Yu, A., Tancik, M., Chen, Q., Recht, B., Kanazawa, A.: Plenoxels: Radiance fields without neural networks. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 5501-5510 (2022)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.53, + 0.787, + 0.572 + ], + "angle": 0, + "content": "25. Garbin, S.J., Kowalski, M., Johnson, M., Shotton, J., Valentin, J.: Fastnerf: High-fidelity neural rendering at 200fps. In: Proceedings of the IEEE/CVF International Conference on Computer Vision. pp. 14346-14355 (2021)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.572, + 0.787, + 0.614 + ], + "angle": 0, + "content": "26. Hedman, P., Srinivasan, P.P., Mildenhall, B., Barron, J.T., Debevec, P.: Baking neural radiance fields for real-time view synthesis. In: Proceedings of the IEEE/CVF International Conference on Computer Vision. pp. 5875-5884 (2021)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.615, + 0.787, + 0.657 + ], + "angle": 0, + "content": "27. Huang, X., Zhang, Q., Feng, Y., Li, H., Wang, Q.: Inverting the imaging process by learning an implicit camera model. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 21456-21465 (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.657, + 0.787, + 0.699 + ], + "angle": 0, + "content": "28. Huang, X., Zhang, Q., Feng, Y., Li, H., Wang, X., Wang, Q.: HDR-nerf: High dynamic range neural radiance fields. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 18398-18408 (2022)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.7, + 0.787, + 0.755 + ], + "angle": 0, + "content": "29. Huang, X., Zhang, Q., Feng, Y., Li, X., Wang, X., Wang, Q.: Local implicit ray function for generalizable radiance field representation. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 97-107 (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.757, + 0.787, + 0.799 + ], + "angle": 0, + "content": "30. Huang, Z., Liang, Z., Zhang, H., Lin, Y., Jia, K.: Sur2f: A hybrid representation for high-quality and efficient surface reconstruction from multi-view images. arXiv preprint arXiv:2401.03704 (2024)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.799, + 0.787, + 0.841 + ], + "angle": 0, + "content": "31. Isik, M., Rünz, M., Georgopoulos, M., Khakhulin, T., Starck, J., Agapito, L., Nießner, M.: Humanrf: High-fidelity neural radiance fields for humans in motion. arXiv preprint arXiv:2305.06356 (2023)" + }, + { + "type": "list", + "bbox": [ + 0.217, + 0.147, + 0.787, + 0.841 + ], + "angle": 0, + "content": null + } + ], + [ + { + "type": "header", + "bbox": [ + 0.551, + 0.115, + 0.732, + 0.127 + ], + "angle": 0, + "content": "Color Correction for NeRF" + }, + { + "type": "page_number", + "bbox": [ + 0.769, + 0.117, + 0.785, + 0.127 + ], + "angle": 0, + "content": "17" + }, + { + "type": "ref_text", + "bbox": [ + 0.214, + 0.147, + 0.788, + 0.189 + ], + "angle": 0, + "content": "32. Jin, H., Liu, I., Xu, P., Zhang, X., Han, S., Bi, S., Zhou, X., Xu, Z., Su, H.: Tensoroir: Tensorial inverse rendering. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 165-174 (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.214, + 0.189, + 0.788, + 0.229 + ], + "angle": 0, + "content": "33. Johari, M.M., Lepoittevin, Y., Fleuret, F.: Geonerf: Generalizing nerf with geometry priors. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 18365-18375 (2022)" + }, + { + "type": "ref_text", + "bbox": [ + 0.214, + 0.229, + 0.788, + 0.284 + ], + "angle": 0, + "content": "34. Joo, H., Simon, T., Li, X., Liu, H., Tan, L., Gui, L., Banerjee, S., Godisart, T.S., Nabbe, B., Matthews, I., Kanade, T., Nobuhara, S., Sheikh, Y.: Panoptic studio: A massively multiview system for social interaction capture. IEEE Transactions on Pattern Analysis and Machine Intelligence (2017)" + }, + { + "type": "ref_text", + "bbox": [ + 0.214, + 0.284, + 0.788, + 0.324 + ], + "angle": 0, + "content": "35. Kerbl, B., Kopanas, G., Leimkuhler, T., Drettakis, G.: 3d gaussian splatting for real-time radiance field rendering. ACM Transactions on Graphics (ToG) 42(4), 1-14 (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.214, + 0.324, + 0.788, + 0.365 + ], + "angle": 0, + "content": "36. Kinh, F., Yilmaz, D., Özcan, B., Kırac, F.: Deterministic neural illumination mapping for efficient auto-white balance correction. In: Proceedings of the IEEE/CVF International Conference on Computer Vision. pp. 1139-1147 (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.214, + 0.365, + 0.788, + 0.405 + ], + "angle": 0, + "content": "37. Kinh, F., Yilmaz, D., Ozcan, B., Kirac, F.: Modeling the lighting in scenes as style for auto white-balance correction. In: Proceedings of the IEEE/CVF Winter Conference on Applications of Computer Vision. pp. 4903-4913 (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.214, + 0.405, + 0.788, + 0.461 + ], + "angle": 0, + "content": "38. Kuang, Z., Luan, F., Bi, S., Shu, Z., Wetzstein, G., Sunkavalli, K.: Palettenerf: Palette-based appearance editing of neural radiance fields. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 20691-20700 (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.214, + 0.461, + 0.788, + 0.5 + ], + "angle": 0, + "content": "39. Lee, J.H., Kim, D.S.: Ice-nerf: Interactive color editing of nerfs via decomposition-aware weight optimization. In: Proceedings of the IEEE/CVF International Conference on Computer Vision. pp. 3491-3501 (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.214, + 0.501, + 0.788, + 0.542 + ], + "angle": 0, + "content": "40. Li, J., Feng, Z., She, Q., Ding, H., Wang, C., Lee, G.H.: Mine: Towards continuous depth MPI with nerf for novel view synthesis. In: Proceedings of the IEEE/CVF International Conference on Computer Vision. pp. 12578-12588 (2021)" + }, + { + "type": "ref_text", + "bbox": [ + 0.214, + 0.542, + 0.788, + 0.595 + ], + "angle": 0, + "content": "41. Liang, Z., Huang, Z., Ding, C., Jia, K.: Helixsurf: A robust and efficient neural implicit surface learning of indoor scenes with iterative intertwined regularization. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 13165-13174 (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.214, + 0.595, + 0.788, + 0.636 + ], + "angle": 0, + "content": "42. Lin, C.H., Ma, W.C., Torralba, A., Lucey, S.: Barf: Bundle-adjusting neural radiance fields. In: Proceedings of the IEEE/CVF international conference on computer vision. pp. 5741-5751 (2021)" + }, + { + "type": "ref_text", + "bbox": [ + 0.214, + 0.636, + 0.788, + 0.69 + ], + "angle": 0, + "content": "43. Liu, Y.L., Gao, C., Meuleman, A., Tseng, H.Y., Saraf, A., Kim, C., Chuang, Y.Y., Kopf, J., Huang, J.B.: Robust dynamic radiance fields. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 13-23 (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.214, + 0.69, + 0.788, + 0.745 + ], + "angle": 0, + "content": "44. Liu, Y., Peng, S., Liu, L., Wang, Q., Wang, P., Theobalt, C., Zhou, X., Wang, W.: Neural rays for occlusion-aware image-based rendering. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 7824-7833 (2022)" + }, + { + "type": "ref_text", + "bbox": [ + 0.214, + 0.745, + 0.788, + 0.786 + ], + "angle": 0, + "content": "45. Ma, L., Li, X., Liao, J., Zhang, Q., Wang, X., Wang, J., Sander, P.V.: Deblur-nerf: Neural radiance fields from blurry images. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 12861-12870 (2022)" + }, + { + "type": "ref_text", + "bbox": [ + 0.214, + 0.786, + 0.788, + 0.84 + ], + "angle": 0, + "content": "46. Martin-Brualla, R., Radwan, N., Sajjadi, M.S., Barron, J.T., Dosovitskiy, A., Duckworth, D.: Nerf in the wild: Neural radiance fields for unconstrained photo collections. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 7210-7219 (2021)" + }, + { + "type": "list", + "bbox": [ + 0.214, + 0.147, + 0.788, + 0.84 + ], + "angle": 0, + "content": null + } + ], + [ + { + "type": "page_number", + "bbox": [ + 0.218, + 0.116, + 0.236, + 0.127 + ], + "angle": 0, + "content": "18" + }, + { + "type": "header", + "bbox": [ + 0.272, + 0.115, + 0.436, + 0.129 + ], + "angle": 0, + "content": "Q. Zhang, Y. Feng et al." + }, + { + "type": "ref_text", + "bbox": [ + 0.218, + 0.148, + 0.785, + 0.175 + ], + "angle": 0, + "content": "47. Max, N.: Optical models for direct volume rendering. IEEE Transactions on Visualization and Computer Graphics 1(2), 99-108 (1995)" + }, + { + "type": "ref_text", + "bbox": [ + 0.218, + 0.178, + 0.787, + 0.232 + ], + "angle": 0, + "content": "48. Mildenhall, B., Hedman, P., Martin-Brualla, R., Srinivasan, P.P., Barron, J.T.: Nerf in the dark: High dynamic range view synthesis from noisy raw images. In: Proceedings of the IEEE/CVF conference on computer vision and pattern recognition. pp. 16190-16199 (2022)" + }, + { + "type": "ref_text", + "bbox": [ + 0.219, + 0.234, + 0.785, + 0.275 + ], + "angle": 0, + "content": "49. Mildenhall, B., Srinivasan, P.P., Tancik, M., Barron, J.T., Ramamoorthi, R., Ng, R.: Nerf: Representing scenes as neural radiance fields for view synthesis. In: European Conference on Computer Vision. pp. 405-421 (2020)" + }, + { + "type": "ref_text", + "bbox": [ + 0.218, + 0.277, + 0.785, + 0.317 + ], + "angle": 0, + "content": "50. Müller, T., Evans, A., Schied, C., Keller, A.: Instant neural graphics primitives with a multiresolution hash encoding. ACM Transactions on Graphics (ToG) 41(4), 1-15 (2022)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.319, + 0.785, + 0.359 + ], + "angle": 0, + "content": "51. Park, K., Henzler, P., Mildenhall, B., Barron, J.T., Martin-Brualla, R.: Camp: Camera preconditioning for neural radiance fields. ACM Transactions on Graphics (TOG) 42(6), 1-11 (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.361, + 0.785, + 0.402 + ], + "angle": 0, + "content": "52. Pearl, N., Treibitz, T., Korman, S.: Nan: Noise-aware nerfs for burst-denoising. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 12672-12681 (2022)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.404, + 0.785, + 0.444 + ], + "angle": 0, + "content": "53. Peng, S., Zhang, Y., Xu, Y., Wang, Q., Shuai, Q., Bao, H., Zhou, X.: Neural body: Implicit neural representations with structured latent codes for novel view synthesis of dynamic humans. In: CVPR (2021)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.446, + 0.785, + 0.487 + ], + "angle": 0, + "content": "54. Reiser, C., Peng, S., Liao, Y., Geiger, A.: Kilonerf: Speeding up neural radiance fields with thousands of tiny mlp's. In: Proceedings of the IEEE/CVF International Conference on Computer Vision. pp. 14335-14345 (2021)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.489, + 0.785, + 0.53 + ], + "angle": 0, + "content": "55. Rematas, K., Liu, A., Srinivasan, P.P., Barron, J.T., Tagliasacchi, A., Funkhouser, T., Ferrari, V.: Urban radiance fields. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 12932-12942 (2022)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.531, + 0.785, + 0.572 + ], + "angle": 0, + "content": "56. Rodríguez, R.G., Vazquez-Corral, J., Bertalmío, M.: Color matching images with unknown non-linear encodings. IEEE Transactions on Image Processing 29, 4435-4444 (2020)" + }, + { + "type": "ref_text", + "bbox": [ + 0.218, + 0.574, + 0.785, + 0.628 + ], + "angle": 0, + "content": "57. Shao, R., Zheng, Z., Tu, H., Liu, B., Zhang, H., Liu, Y.: Tensor4d: Efficient neural 4d decomposition for high-fidelity dynamic reconstruction and rendering. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 16632-16642 (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.218, + 0.63, + 0.785, + 0.685 + ], + "angle": 0, + "content": "58. Srinivasan, P.P., Deng, B., Zhang, X., Tancik, M., Mildenhall, B., Barron, J.T.: Nerv: Neural reflectance and visibility fields for relighting and view synthesis. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 7495-7504 (2021)" + }, + { + "type": "ref_text", + "bbox": [ + 0.218, + 0.687, + 0.785, + 0.741 + ], + "angle": 0, + "content": "59. Sun, G., Chen, X., Chen, Y., Pang, A., Lin, P., Jiang, Y., Xu, L., Yu, J., Wang, J.: Neural free-viewpoint performance rendering under complex human-object interactions. In: Proceedings of the 29th ACM International Conference on Multimedia. pp. 4651-4660 (2021)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.743, + 0.785, + 0.797 + ], + "angle": 0, + "content": "60. Tancik, M., Casser, V., Yan, X., Pradhan, S., Mildenhall, B., Srinivasan, P.P., Barron, J.T., Kretzschmar, H.: Block-nerf: Scalable large scene neural view synthesis. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 8248-8258 (2022)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.799, + 0.785, + 0.84 + ], + "angle": 0, + "content": "61. Tang, J., Zhou, H., Chen, X., Hu, T., Ding, E., Wang, J., Zeng, G.: Delicate textured mesh recovery from nef via adaptive surface refinement. arXiv preprint arXiv:2303.02091 (2023)" + }, + { + "type": "list", + "bbox": [ + 0.217, + 0.148, + 0.787, + 0.84 + ], + "angle": 0, + "content": null + } + ], + [ + { + "type": "header", + "bbox": [ + 0.551, + 0.115, + 0.732, + 0.127 + ], + "angle": 0, + "content": "Color Correction for NeRF" + }, + { + "type": "page_number", + "bbox": [ + 0.769, + 0.116, + 0.786, + 0.127 + ], + "angle": 0, + "content": "19" + }, + { + "type": "ref_text", + "bbox": [ + 0.214, + 0.147, + 0.788, + 0.203 + ], + "angle": 0, + "content": "62. Verbin, D., Hedman, P., Mildenhall, B., Zickler, T., Barron, J.T., Srinivasan, P.P.: Ref-nerf: Structured view-dependent appearance for neural radiance fields. In: 2022 IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR). pp. 5481-5490. IEEE (2022)" + }, + { + "type": "ref_text", + "bbox": [ + 0.214, + 0.203, + 0.787, + 0.258 + ], + "angle": 0, + "content": "63. Wang, Q., Wang, Z., Genova, K., Srinivasan, P.P., Zhou, H., Barron, J.T., Martin-Brualla, R., Snavely, N., Funkhouser, T.: Ibrnet: Learning multi-view image-based rendering. In: Proceedings of the IEEE/CVF conference on computer vision and pattern recognition. pp. 4690-4699 (2021)" + }, + { + "type": "ref_text", + "bbox": [ + 0.214, + 0.258, + 0.787, + 0.299 + ], + "angle": 0, + "content": "64. Xie, S., Zhu, H., Liu, Z., Zhang, Q., Zhou, Y., Cao, X., Ma, Z.: Diner: Disorder-invariant implicit neural representation. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 6143-6152 (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.214, + 0.299, + 0.787, + 0.34 + ], + "angle": 0, + "content": "65. Xu, Q., Xu, Z., Philip, J., Bi, S., Shu, Z., Sunkavalli, K., Neumann, U.: Point-nerf: Point-based neural radiance fields. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 5438-5448 (2022)" + }, + { + "type": "ref_text", + "bbox": [ + 0.214, + 0.34, + 0.787, + 0.381 + ], + "angle": 0, + "content": "66. Yao, Y., Zhang, J., Liu, J., Qu, Y., Fang, T., McKinnon, D., Tsin, Y., Quan, L.: Neilf: Neural incident light field for physically-based material estimation. In: European Conference on Computer Vision. pp. 700-716. Springer (2022)" + }, + { + "type": "ref_text", + "bbox": [ + 0.214, + 0.381, + 0.787, + 0.408 + ], + "angle": 0, + "content": "67. Yen-Chen, l.: Nerf-pytorch: a pytorch implementation of nerf. URL https://github.com/yenchenlin/nerf-pytorch (2020)" + }, + { + "type": "ref_text", + "bbox": [ + 0.214, + 0.408, + 0.787, + 0.449 + ], + "angle": 0, + "content": "68. Yu, A., Ye, V., Tancik, M., Kanazawa, A.: pixelnerf: Neural radiance fields from one or few images. In: Proceedings of the IEEE/CVF conference on computer vision and pattern recognition. pp. 4578-4587 (2021)" + }, + { + "type": "ref_text", + "bbox": [ + 0.214, + 0.449, + 0.787, + 0.49 + ], + "angle": 0, + "content": "69. Zhang, K., Luan, F., Li, Z., Snavely, N.: Iron: Inverse rendering by optimizing neural sdfs and materials from photometric images. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 5565-5574 (2022)" + }, + { + "type": "ref_text", + "bbox": [ + 0.214, + 0.49, + 0.787, + 0.531 + ], + "angle": 0, + "content": "70. Zhang, R., Isola, P., Efros, A.A., Shechtman, E., Wang, O.: The unreasonable effectiveness of deep features as a perceptual metric. In: Proceedings of the IEEE conference on computer vision and pattern recognition. pp. 586-595 (2018)" + }, + { + "type": "ref_text", + "bbox": [ + 0.214, + 0.531, + 0.787, + 0.572 + ], + "angle": 0, + "content": "71. Zhang, X., Srinivasan, P.P., Deng, B., Debevec, P., Freeman, W.T., Barron, J.T.: Nerfactor: Neural factorization of shape and reflectance under an unknown illumination. ACM Transactions on Graphics (ToG) 40(6), 1-18 (2021)" + }, + { + "type": "ref_text", + "bbox": [ + 0.214, + 0.571, + 0.787, + 0.612 + ], + "angle": 0, + "content": "72. Zhao, Y., Ferguson, S., Zhou, H., Elliott, C., Rafferty, K.: Color alignment for relative color constancy via non-standard references. IEEE Transactions on Image Processing 31, 6591-6604 (2022)" + }, + { + "type": "ref_text", + "bbox": [ + 0.214, + 0.612, + 0.787, + 0.667 + ], + "angle": 0, + "content": "73. Zhou, T., He, K., Wu, D., Xu, T., Zhang, Q., Shao, K., Chen, W., Xu, L., Yu, J.: Relightable neural human assets from multi-view gradient illuminations. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 4315-4327 (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.214, + 0.667, + 0.787, + 0.694 + ], + "angle": 0, + "content": "74. Zhu, H., Liu, F., Zhang, Q., Cao, X., Ma, Z.: Rhino: Regularizing the hash-based implicit neural representation. arXiv preprint arXiv:2309.12642 (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.214, + 0.694, + 0.787, + 0.735 + ], + "angle": 0, + "content": "75. Zhu, H., Xie, S., Liu, Z., Liu, F., Zhang, Q., Zhou, Y., Lin, Y., Ma, Z., Cao, X.: Disorder-invariant implicit neural representation. IEEE Transactions on Pattern Analysis and Machine Intelligence (2024)" + }, + { + "type": "ref_text", + "bbox": [ + 0.214, + 0.735, + 0.787, + 0.775 + ], + "angle": 0, + "content": "76. Zhu, J., Zhu, H., Zhang, Q., Zhu, F., Ma, Z., Cao, X.: Pyramid nef: Frequency guided fast radiance field optimization. International Journal of Computer Vision 131(10), 2649-2664 (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.214, + 0.776, + 0.787, + 0.831 + ], + "angle": 0, + "content": "77. Zhuang, Y., Zhang, Q., Wang, X., Zhu, H., Feng, Y., Li, X., Shan, Y., Cao, X.: A pre-convolved representation for plug-and-play neural illumination fields. In: Proceedings of the AAAI Conference on Artificial Intelligence. vol. 38, pp. 7828-7836 (2024)" + }, + { + "type": "list", + "bbox": [ + 0.214, + 0.147, + 0.788, + 0.831 + ], + "angle": 0, + "content": null + } + ] +] \ No newline at end of file diff --git a/2024/Physically Plausible Color Correction for Neural Radiance Fields/88cfdd54-68fa-4f55-bb61-e981b6d01cd5_origin.pdf b/2024/Physically Plausible Color Correction for Neural Radiance Fields/88cfdd54-68fa-4f55-bb61-e981b6d01cd5_origin.pdf new file mode 100644 index 0000000000000000000000000000000000000000..7fa8af6d1d32758eaa7dbdaa6956ef705d1421aa --- /dev/null +++ b/2024/Physically Plausible Color Correction for Neural Radiance Fields/88cfdd54-68fa-4f55-bb61-e981b6d01cd5_origin.pdf @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:d4a7b789549aa003ea8c272bea503f7985edacd9bc8a0af2e1c5f5e7f5709416 +size 6553588 diff --git a/2024/Physically Plausible Color Correction for Neural Radiance Fields/full.md b/2024/Physically Plausible Color Correction for Neural Radiance Fields/full.md new file mode 100644 index 0000000000000000000000000000000000000000..d408fc1b4b39c2f815aeea0716684f530ce051cc --- /dev/null +++ b/2024/Physically Plausible Color Correction for Neural Radiance Fields/full.md @@ -0,0 +1,344 @@ +# Physically Plausible Color Correction for Neural Radiance Fields + +Qi Zhang $^{1,\dagger}$ , Ying Feng $^{1,\dagger}$ , and Hongdong Li $^{2}$ + +$^{1}$ Independent Researcher {nwpuqzhang, yfeng.von}@gmail.com + +2 Australian National University hongdong.li@anu.edu.au + +Abstract. Neural Radiance Fields have become the representation of choice for many 3D computer vision and computer graphics applications, e.g., novel view synthesis and 3D reconstruction. Multi-camera systems are commonly used as the image capture setup in NeRF-based multi-view tasks such as dynamic scene acquisition or realistic avatar animation. However, a critical issue that has often been overlooked in this setup is the evident differences in color responses among multiple cameras, which adversely affect the NeRF reconstruction performance. These color discrepancies among multiple input images stem from two aspects: 1) implicit properties of the scenes such as reflections and shadings, and 2) external differences in camera settings and lighting conditions. In this paper, we address this problem by proposing a novel color correction module that simulates the physical color processing in cameras to be embedded in NeRF, enabling the unified color NeRF reconstruction. Besides the view-independent color correction module for external differences, we predict a view-dependent function to minimize the color residual (including, e.g., specular and shading) to eliminate the impact of inherent attributes. We further describe how the method can be extended with a reference image as guidance to achieve aesthetically plausible color consistency and color translation on novel views. Experiments validate that our method is superior to baseline methods in both quantitative and qualitative evaluations of color correction and color consistency. + +Keywords: Neural Radiance Field $\cdot$ View Synthesis $\cdot$ Color Translation + +# 1 Introduction + +The synthesis and rendering of novel views using multi-view images has been a long-standing problem in computer vision and graphics, which is also a cornerstone task for augmented reality (AR) and virtual reality (VR) applications. The recent success of Neural Radiance Fields (NeRF) [49] has made the implicit neural representation a powerful approach for novel view synthesis tasks, as + +![](images/fb49fe45339def8d81f97800bf122c05e3117ff359c3d5dc37817359d99b0369.jpg) +(a) input views +Fig. 1: Physically Plausible Color Correction for Neural Radiance Fields. (a) Given a set of images captured under various camera settings and lighting conditions, our method introduces a color correction module into the NeRF-based model, which enables recovering (b) a unified color radiance field with a physically plausible white balance. Unlike prior methods that embed color through features, our color correction module simulates color differences via a nonlinear color translation matrix. Additionally, our method can render (c) novel views translated from the color of a reference image. + +![](images/4cc89a1e3946176c2dafe27401ccaa006ca41ef98f98cbffe0ed4f1d7fa52e4c.jpg) +(b) unified color radiance field + +![](images/ed811e9f51221c5b7e2d444c4f336950faf915831c729ce96247cd8d1550d6c7.jpg) +(c) color translation + +an alternative to traditional image-based reconstruction techniques. The NeRF model employs Multi-Layer Perceptrons (MLPs) to parameterize the mapping from spatial coordinates to corresponding densities and view-dependent colors. This approach has been proven effective with scenes that are consistently well-lit without substantial variations in image brightness. + +With the increasing prevalence of multi-camera systems [18, 22, 31, 34, 53, 59, 73], NeRF and its variants [4, 5, 50] face challenges in achieving photorealistic novel view synthesis with color consistency. The differences in color response among multiple cameras can significantly impact the accuracy and consistency of synthesized views. Even when capturing with identical camera settings, different types of cameras may present distinct color responses caused by the difference in sensors and ISP algorithms. To address this limitation, NeRF-W [46] and Ha-NeRF [15] incorporate additional latent codes to encode appearances into the NeRF model for consistent view synthesis and appearance control. However, these methods tend to embed the overall appearances as view-dependent features, making it difficult to achieve fine-grained and intuitive color correction. HDR-NeRF [28] attempts to address this issue but primarily focuses on building a physical module for camera exposures. It is necessary to introduce color translation for NeRF-based multi-view tasks, which should involve modeling color correlations across cameras and enabling physically plausible color correction and consistency in NeRF reconstructions. + +The color difference across multi-view images arises from two aspects: inherent view-dependent scene properties (e.g. reflections and shadings); external differences in camera settings (e.g. white balance, ISO, exposure) and lighting conditions. Naturally, we could calibrate each image via color correction methods [1,3,6,36,37], and then input to NeRF-based method for novel view synthesis. However, these image-based methods struggle to preserve color consistency in + +NeRF reconstructions. In addition to faithfully reproducing the captured colors, we consider that color correction process should allow for color adjustment to enhance visual appeal, and even color translation to a given reference image. By incorporating such color correction module, users, including novice ones, can easily recover NeRF and adjust the colors of novel views based on their preferences. + +In this paper, we propose a novel framework that incorporates a color correction module into the NeRF-based model to achieve unified color radiance field reconstruction from multi-view images with different camera settings and lighting conditions, as shown in Fig. 1. In contrast to the embedding feature for overall color differences in prior NeRF-based methods [15,46], we present a color correction module to predict a nonlinear color translation matrix that simulates the color processing in cameras, such as white balance, nonlinear color enhancement, tone-manipulation, gamma encoding, etc. Additionally, considering the view-dependent factors of inherent properties, we predict a view-dependent function to separate the color residual (e.g. specular and shading of the scene) to improve the accuracy of the color translation matrix. Our technical contributions are summarized as follows: + +- We present a novel framework for color correction and translation in NeRF reconstruction, allowing even novice users to perform novel view synthesis under unstable camera settings in an intuitive manner. +- We propose a color correction module that defines the non-linear color translation to be embedded in NeRF. +- We introduce tailored regularizations to eliminate the impact of inherent scene properties, constrain the scene's NeRF to a white balance color space, and maintain the entire color distribution. + +Extensive experiments have shown that our method outperforms baseline methods both quantitatively and qualitatively in terms of color correction. + +# 2 Related Works + +Neural Radiance Fields Neural Radiance Fields represent a scene implicitly with multi-layer perceptrons, approximating a continuous function that maps each 3D position and direction to a density and radiance [49]. By incorporating view direction features, NeRF can effectively model view-dependent effects such as reflection and refraction. This allows for photo-realistic synthesis of novel views with volumetric rendering. To alleviate the computational limitation of massive MLP evaluations while maintaining the high-frequency appearances, many extensive works have been proposed to speed up the training and inference of the model [5,13,17,24-26,35,50,54,64,65,74-76] with various geometry proxies and neural features, explore view synthesis for dynamic scenes [12,21,23,43,57], and enhance the generalization of NeRF [14,19,29,33,40,44,63,68]. Many methods [30,41,61] have been proposed to exploit the view-dependent appearance to decouple the physical properties of the scene. NeRF2Mesh [61] proposes a + +two-stage framework to extract the mesh and texture images by regularizing the view-dependent color. Many inverse rendering methods go a step further in material decoupling [9, 10, 32, 58, 62, 66, 69, 71, 77]. They decompose the radiance into physical properties such as albedo, roughness, metallicity, and lighting by introducing physical-based rendering into NeRF optimization, and enable relighting under novel lighting conditions or material editing. Numerous methods have been proposed to reconstruct scenes from images captured in uncontrolled environments, addressing challenges such as inaccurate poses [7, 16, 42, 51], image blur or noise [45, 48, 52], and inconsistent camera settings [27, 28, 55, 60]. + +Appearance control of NeRF Images captured in natural environments or taken by multiple cameras often exhibit inconsistency in brightness and color. These image-dependent appearances may derive from changing lighting conditions, exposure variations and white balancing, or differences in camera sensors and ISP, which are not desired in novel view synthesis. HDR-NeRF [28] focuses on multi-view images of varying exposures. By modeling the simplified physical imaging process with a tone mapper, which maps the radiance of a scene point to a pixel value, HDR-NeRF reconstructs the high dynamic range radiance field, which can generate novel views with consistent exposures. Many methods enable NeRF to handle these inconsistencies by introducing related parameters or embeddings. NeRF-w [46] and Ha-NeRF [15] propose to encode the image-dependent appearance using latent codes, which can then be manipulated to control the overall style of rendered images. Block-NeRF [60] feeds an additional exposure level to the appearance prediction model to deal with the brightness variance and provides the ability to modify the global exposure conditioning during inference. URF [55] predicts a $3 \times 3$ matrix to restrictively model the white balance and exposure variations. One shortcoming of these methods is that view-dependent appearances, those either caused by changes in camera settings and environment lighting or are intrinsical from scene physical properties, are coupled together and modeled related to both latent codes and view directions. In addition to the aforementioned methods that could control scene appearances by modifying materials or latent codes, some methods [38,39] engage in appearance editing by color post-manipulation and rely heavily on the consistency of views. + +Color Correction Color correction has been extensively studied on single 2D image and multi-camera systems. Single 2D image color correction [1,3,6,36, 37], or white balance correction, cannot guarantee view-consistent results when naively performed on each view. On the other hand, multi-camera color correction [20,56,72] relies on common area or statistical matching, which cannot handle inherent view variance and only matches the global style. To address these limitations, we propose integrating a color correction module into the NeRF optimization process to enable view-consistent color correction with 3D awareness. + +# 3 Color Correction for NeRF + +Given multi-view images captured under various camera settings and lighting conditions, our method proposes a color correction module embedded in a NeRF- + +![](images/edbb41236e85ac1f7511e153e8e0fbe243a8187b6e4f0dd02aafb1fdc248f6f7.jpg) +Fig. 2: Overview of our pipeline. Given multi-view images captured under various camera settings and lighting conditions, our method recovers a unified color radiance field, which enables consistent novel view synthesis. We propose a non-linear color correction module embedded in NeRF-based model to simulate the color processing performed in cameras (e.g. white balance, exposure, nonlinear color enhancement, etc.), which maps the color from unified scene color space to each camera color space. The view-dependent color arises from inherent scene properties like specular and shading is extracted under regularization. We design a robust optimization scheme with tailored losses to achieve physically plausible color correction. + +based model, enabling the unified color NeRF reconstruction, as shown in Fig. 2. The scene radiance field is decomposed into two components to decouple the inherent color variances like specular and shading. The color correction module is designed as a $3 \times 7$ translation matrix in nonlinear RGB space. Multiple regularizations on pixels and areas are incorporated to improve color consistency. + +# 3.1 Scene Representation + +Neural Radiance Field (NeRF) represents a scene using an implicit neural function, which utilizes MLPs to map the position $\mathbf{x}$ and view direction $\mathbf{d}$ to a density $\sigma(\mathbf{x})$ and view-dependent color $\mathbf{c}(\mathbf{x},\mathbf{d})$ . To render the color of a pixel in an image from a given camera pose, NeRF casts a ray $\mathbf{r}(t) = \mathbf{o} + t\mathbf{d}$ through the pixel, where $\mathbf{o}$ and $\mathbf{d}$ are the origin and direction of the ray, respectively, and $t$ is the parameter that determines the position along the ray. The predicted $\{(\sigma,\mathbf{c})\}$ of sampled positions on the ray are then accumulated into the color $C(\mathbf{r})$ of the pixel via volume rendering [47]: + +$$ +C (\mathbf {r}) = \int_ {t _ {n}} ^ {t _ {f}} T (t) \sigma (\mathbf {r} (t)) \mathbf {c} (\mathbf {r} (t), \mathbf {d}) d t, \tag {1} +$$ + +$$ +T (t) = \exp \left(- \int_ {t _ {n}} ^ {t} \sigma (\mathbf {r} (p)) d p\right), \tag {2} +$$ + +where $t_n$ and $t_f$ indicate the near and far boundary of the ray, respectively. $T(t)$ refers to an accumulated transmittance, which quantifies the probability of the ray traveling from $t_n$ to $t$ without encountering other particles. + +We build our model on the NeRF framework, which is widely used for modeling scenes captured under static camera settings and lighting conditions. To eliminate the influence of inherent color variances, we propose to decompose the appearance into two components: view-independent diffuse color $\mathbf{c}_d$ and view-dependent color $\mathbf{c}_s$ . The final color is obtained by summing the two terms: + +$$ +\mathbf {c} = \mathbf {c} _ {d} (\mathbf {x}) + \mathbf {c} _ {s} (\mathbf {x}, \mathbf {d}). \tag {3} +$$ + +By explicitly considering these two components, we can better model the appearance changes arising from surface properties and these caused by varying lighting conditions and camera settings. + +# 3.2 Color Correction Module + +Digital cameras perform a number of processing steps to transform raw images to standard RGB images, including demosaicing, white balance, color correction and etc. A generic model [8] of this pipeline is formulated as: + +$$ +\mathbf {I} _ {s r g b} = f \left(\alpha \mathbf {A} _ {x y z} \mathbf {D} _ {w b} \mathbf {I} _ {r a w}\right), \tag {4} +$$ + +where $\alpha$ is an exposure compensation common gain, $\mathbf{D}_{wb}$ represents the $3\times 3$ diagonal matrix for white balance, $\mathbf{A}_{xyz}$ is a $3\times 3$ linear transform of color space, and $f(\cdot)$ is a nonlinear function compounding color enhancement, tonemanipulation, and a final sRGB gamma encoding, etc. It's worth noting that the differences in white balance and lighting conditions become non-linear in the sRGB color space after undergoing these mappings. + +In our framework, we expect to recover the scene's radiance field with unified color when the input sRGB images exhibit color variances. Inspired by [2], our method introduces a polynomial color-mapping function to model the color deviation between the scene color space $\mathbf{c}$ and each camera color space $\mathbf{c}^{cam}$ captured by different cameras, which can be written as: + +$$ +\mathbf {c} ^ {c a m} = \mathbf {M} \varPhi (\mathbf {c}), \tag {5} +$$ + +where $\varPhi$ is the kernel mapping function that transforms the input color into the nonlinear space, specifically, $\varPhi([R,G,B]^T) \to [R,G,B,RG,RB,GB,1]^T$ . The matrix $\mathbf{M}$ is a color translation matrix of shape $3 \times 7$ . Notably, we utilize an Embedding network $E$ to predict a matrix $\mathbf{M}$ for each camera and lighting condition, and based on the physical imaging process, the matrix is shared by the diffuse and view-dependent components. The pixel colors of sampled points in the scene color space are first mapped to camera color space, and then accumulated through volume rendering. Combining Eq. (1), Eq. (3) and Eq. (5), the rendering equation can be written as: + +$$ +C ^ {c a m} = \int_ {t _ {n}} ^ {t _ {f}} T (t) \sigma (\mathbf {r} (t)) \left(\mathbf {M} \varPhi \left(\mathbf {c} _ {d} (\mathbf {r} (t))\right) + \mathbf {M} \varPhi \left(\mathbf {c} _ {s} (\mathbf {r} (t), \mathbf {d})\right)\right) d t \tag {6} +$$ + +where $T(t)$ is defined in Eq. (2). By integrating the color correction module in the rendering pipeline, we can reconstruct the radiance field of the scene while accounting for color variations caused by different cameras and lighting conditions. + +# 3.3 View-consistent Color Distribution + +Considering the color correction matrix is designed to be camera-wise, relying solely on pixel-level color loss hinders the accuracy and efficiency of optimization. Therefore we employ a cross-camera color distribution constraint. Given that digital cameras capture images with three color channels, each with a discrete range of $K$ intensity values, we analyze the intensity distribution in each channel, which is formulated as an intensity histogram to constrain the color consistency. + +Drawing inspiration from the cross-camera color translation task, we transform the RGB color space $C^{cam}$ into YUV color space $Y^{cam} \in [-1,1]^3$ . The YUV color space comprises one luma component (Y) and two chrominance components, U (blue projection) and V (red projection). We analyze each color channel independently and obtain the intensity histogram by counting the number of pixels of each intensity value. To obtain the intensity histogram as a smooth and differentiable function, we formulate one histogram distribution $\mathbf{h}(Y^{cam})$ of an image color channel as + +$$ +\mathbf {h} = \left\{\frac {1}{N} \sum_ {x \in Y} \Theta \left(\frac {I - \mu_ {k} + L / 2}{W}\right) - \Theta \left(\frac {I - \mu_ {k} - L / 2}{W}\right) \right\} _ {k = 0} ^ {K - 1}, \tag {7} +$$ + +where $K$ is the number of bins set in the histogram, and $L$ indicates the length of each bin, which also equals $\frac{2}{K}$ . Given the $k$ -th bin, its center and bandwidth refer to $\mu_k = -1 + L(k + \frac{1}{2})$ and $W = L / 2.5$ , respectively. $\Theta(x)$ is a Sigmoid function $(1 + \exp(-x))^{-1}$ defining a differentiable approximation of the histogram distribution. + +Histogram Distance. To incorporate color distribution constraints into our framework and facilitate easy convergence, we propose using the Euclidean distance between the cumulative histograms to measure the distance between two image histograms. Let $\mathbf{h}_s$ and $\mathbf{h}_r$ represent the histograms of the rendered and the reference images, respectively. The distance between histograms is defined: + +$$ +\mathcal {D} _ {h} \left(\mathbf {h} _ {s}, \mathbf {h} _ {r}\right) = \sum_ {i = 1} ^ {K} \left(\mathrm {C D F} _ {i} \left(\mathbf {h} _ {s}\right), \mathrm {C D F} _ {i} \left(\mathbf {h} _ {r}\right)\right), \tag {8} +$$ + +where $\mathrm{CDF}_i(\mathbf{h})$ denotes the $i$ -th element of the cumulative density function of histogram. The histogram distance $\mathcal{D}_h$ measures the global similarity of color distribution between two images. Unlike NeRF, which focuses on individual pixel values, this approach takes into account the entire color distribution of the images. By considering the cumulative histograms, we can effectively compare and align the color distribution of two different images. This property is particularly useful for color translation tasks with our color correction module. We can achieve consistent color appearance across different images by adapting the color distribution towards the reference image. + +# 3.4 Optimization + +To regularize the reconstructed scene color and reduce ambiguity, we design a series of losses on both pixel and image levels, including color reconstruction loss, specular loss, histogram loss, and gray-world loss. + +Color Reconstruction Loss. To jointly optimize the scene's NeRF and the camera-dependent color correction matrix $\{\mathbf{M}_i\}_{i=1}^N$ , color reconstruction loss is proposed to minimize the mean squared error (MSE) of sampled pixels between the ground truth color $C^{cam}$ and the rendered color $\hat{C}^{cam}$ which is mapped from scene color by corresponding matrix (see Eq. (5) and Eq. (6)). The color reconstruction loss is formulated as: + +$$ +\mathcal {L} _ {c} = \sum_ {\mathbf {r} \in \mathcal {R} (\mathbf {P})} \left\| \widehat {C} ^ {c a m} - C ^ {c a m} \right\| _ {2} ^ {2}. \tag {9} +$$ + +Specular Loss. According to Eq. (3), we decompose the scene's color into diffuse color $\mathbf{c}_d$ and view-dependent color $\mathbf{c}_s$ . This view-dependent color accounts for the color variances due to the inherent material properties and geometries. Both components share a common color correction matrix $\mathbf{M}$ . Ideally, we suppose the cross-image differences are mostly caused by varying light conditions and camera-dependent color translation. Thus, we apply a regularization on $\mathbf{c}_s$ , which benefits color consistency on the diffuse partition and estimation of color correction matrix, as shown in ablation studies. The L2 regularization on the view-dependent color is formulated as: + +$$ +\mathcal {L} _ {s} = \sum_ {\mathbf {r} \in \mathcal {R} (\mathbf {P})} \left\| \widehat {C} _ {s} \right\| _ {2} ^ {2} \tag {10} +$$ + +Gray-world Loss. Due to the correlation between scene color space and color correction matrix, the recovered scene color $\mathbf{c}$ is under-constrained with only $\mathcal{L}_g$ . To address this, white balance correction methods often refer to some prior knowledge of the spectral properties of the camera sensors. One commonly used heuristic prior is the "gray-world" assumption: the color in the world tends to be natural gray on average [11]. We also define an L2 gray-world regularization for the scene's radiance field: + +$$ +\mathcal {L} _ {g} = \frac {1}{N _ {\mathbf {r}}} \sum_ {\mathbf {r} \in \mathcal {R} (\mathbf {P})} \left\| \widehat {C} - C _ {0} \right\| _ {2} ^ {2}, \tag {11} +$$ + +where $C_0$ is a hyperparameter which is generally set as the midway of the color values such as 0.5. + +Histogram Loss. To further constrain the color consistency of reconstructed radiance fields in different views, besides rendering image $\widehat{C}^{cam}$ from the ground truth view, we render an additional image from random view and obtain its histogram $\widehat{\mathbf{h}}_r^{cam}$ during training. Based on $\widehat{\mathbf{h}}_r^{cam}$ , $\widehat{\mathbf{h}}^{cam}$ of $\widehat{C}^{cam}$ and $\mathbf{h}^{cam}$ of ground truth color $C^{cam}$ , we design a cross histogram loss based on Eq. (8), + +$$ +\mathcal {L} _ {h} = \frac {1}{3} \left(\mathcal {D} _ {h} \left(\widehat {\mathbf {h}} ^ {c a m}, \mathbf {h} ^ {c a m}\right) + \mathcal {D} _ {h} \left(\widehat {\mathbf {h}} _ {r} ^ {c a m}, \mathbf {h} ^ {c a m}\right) + \mathcal {D} _ {h} \left(\widehat {\mathbf {h}} ^ {c a m}, \widehat {\mathbf {h}} _ {r} ^ {c a m}\right)\right). \tag {12} +$$ + +Considering that gray-world loss $\mathcal{L}_g$ and histogram loss $\mathcal{L}_h$ are defined on color distributions, we sample rays from grids and combine them to form an image instead of rendering the entire image during training for converge efficiency. This is based on the assumption that the color distribution of an image will remain unchanged after the random grid sampling. + +Total Loss. The complete loss $\mathcal{L}$ is a weighted sum of the above loss functions from Eq. (9) to Eq. (12): + +$$ +\mathcal {L} = \mathcal {L} _ {c} + \lambda_ {s} \mathcal {L} _ {s} + \lambda_ {h} \mathcal {L} _ {h} + \lambda_ {g} \mathcal {L} _ {g}, \tag {13} +$$ + +where $\lambda_s = 0.1$ , $\lambda_h = 0.01$ , and $\lambda_g = 0.001$ are the weights of different losses. + +# 4 Experiments + +# 4.1 Implementation Details + +Our implementation of NeRF follows [67]. Note that our camera correction module could be embedded in different NeRF-based models [4,13,50]. An eight-layer MLP with 256 channels with encoding of position $\mathbf{x}$ as input is used to generate $\sigma$ , and followed by one additional 256-channel fully-connected layer to extract a feature. Diffuse color $\mathbf{c}_d$ is obtained by mapping the feature to RGB channels. The view-dependent color $\mathbf{c}_s$ is obtained through an additional 128-channel fully connected layer and a mapping layer, with the aforementioned feature and encoding of view direction $\mathbf{d}$ as input. Both $\mathbf{c}_d$ and $\mathbf{c}_s$ are transformed to nonlinear space by the kernel mapping function $\varPhi$ in Eq. (5). The embedding vector is reshaped to color correction matrix $\mathbf{M}$ of size $3\times 7$ . Models are trained over 300,000 iterations with a batch size of 1024. + +# 4.2 Evaluation Dataset and Metrics + +Dataset. We evaluate the proposed method on our collected dataset. The dataset consists of 6 real scenes captured by three digital cameras with different camera settings (e.g. white balance, ISO, exposures, etc.) and light conditions (e.g. non-flashing, and flashing with different lights). Images are collected at 90 different poses for each scene. We select 60 views with different poses as the training dataset. The remaining images are used as test images, corresponding to the same setting with training data. The resolution is $1000 \times 667$ pixels. + +Metrics. We report quantitative performance using PSNR (higher is better) and SSIM (higher is better) metrics, as well as the state-of-the-art LPIPS (lower is better) perceptual metric [70]. To better evaluate the color consistency and correction, we propose a $W_{2}$ metric based on Wasserstein distance, which is a distance function defined between probability distributions on the entire color + +
ScenePSNR ↑SSIM ↑
NeRFNeRF-WHa-NeRFOursNeRFNeRF-WHa-NeRFOurs
Hamster15.62928.38128.28129.5710.9310.9510.9580.959
Pumpkin17.16230.18930.09631.3590.8780.9430.9420.956
Letters16.78028.04328.38730.4520.7830.9200.9220.955
MailBox18.85325.77225.90328.1280.9000.8950.8960.922
Pipe14.72522.20125.98227.5540.7440.7570.8560.907
Room16.62914.77214.96531.0970.9320.7960.7730.969
Mean16.62924.89325.60229.6940.8610.8770.8910.945
+ +
SceneLPIPS ↓\( W_2 \) (1e-3) ↓
NeRFNeRF-WHa-NeRFOursNeRFNeRF-WHa-NeRFOurs
Hamster0.5080.2190.2250.19974.42.582.592.31
Pumpkin0.5970.2340.2350.18161.11.271.400.53
Letters0.4490.3210.3240.21676.41.110.950.68
MailBox0.4390.2660.2690.20544.95.744.133.71
Pipe0.6210.4760.2790.17771.41.500.600.42
Room0.4960.6650.6680.10389.513.013.56.16
Mean0.5180.3630.3330.18069.64.203.862.30
+ +Table 1: Quantitative comparisons with baseline methods on our collected scenes, including PSNR, SSIM, LPIPS, and $W_{2}$ metrics. The Wasserstein distance $W_{2}$ measures the difference of entire color distributions. + +space. The image-level $W_{2}$ (lower is better) is defined as: + +$$ +W _ {2} = \frac {1}{N} \sum_ {i} ^ {N} \mathrm {W a s s e r s t e i n} (\widehat {C} _ {i} ^ {c a m}, C ^ {t r a i n}), +$$ + +$$ +\text {W a s s e r s t e i n} \left(\widehat {C} _ {i} ^ {\text {c a m}}, C ^ {\text {t r a i n}}\right) = \left\| \operatorname {m e a n} \left(\widehat {C} _ {i} ^ {\text {c a m}}\right) - \operatorname {m e a n} \left(C ^ {\text {t r a i n}}\right) \right\| _ {2} ^ {2} + \tag {14} +$$ + +$$ +\mathrm {t r a c e} \left(\mathrm {c o v} (\widehat {C} _ {i} ^ {c a m}) + \mathrm {c o v} (C ^ {t r a i n}) - 2 \sqrt {\sqrt {\mathrm {c o v} (\widehat {C} _ {i} ^ {c a m})} \mathrm {c o v} (C ^ {t r a i n}) \sqrt {\mathrm {c o v} (\widehat {C} _ {i} ^ {c a m})}}\right), +$$ + +where $\mathrm{mean}(\cdot)$ and $\mathrm{cov}(\cdot)$ refer to the mean and covariance of an image. $C^{train}$ denotes the training image. $\hat{C}^{cam}$ indicates the rendered images those translated with matrix of $C^{train}$ . + +# 4.3 Evaluation + +Baselines. We compare our method against the following baseline methods. 1) NeRF [49]: the original NeRF method, also our backbone network. 2) NeRF-W [46]: unofficial implementation where the dimension of latent code is 48. 3) Ha-NeRF [15]: official implementation which can use a reference image to do color translation. + +Comparisons. The qualitative results of rendered novel views on our own dataset are shown in Fig. 3. Compared to baseline methods, our method achieves promising results on view rendering quality (see MSE error) and color correction measured by Wasserstein distance between ground truths and rendered images. Specifically, NeRF, which is directly reconstructed from inputs under various conditions, tends to render views with significant artifacts. NeRF-W and Ha-NeRF do ensure color consistency to a certain extent, as shown in Fig. 3. However, these methods can not distinguish the implicit properties (e.g. specular) + +![](images/c4804a127f5b0d36915a8fd06ed8cb16fe45b90d4d9b549a37b0978574db69b0.jpg) +Fig. 3: Qualitative results of our method compared with baseline methods. We also show the error map and histogram for better visualization. MSE and Wasserstein values are on the top corners of the error map and histogram, respectively. Our method outperforms baselines, providing superior rendering quality and effective color correction. NeRF [49] generates significant artifacts. NeRF-W [46] and Ha-NeRF [15] achieve color correction to a certain extent, but exhibit shortcomings in areas with high illumination (e.g. the leaf of scene Mailbox and 'E' of scene Letters) and generate artifacts (e.g. the window of scene Mailbox and floor of scene Hamster). + +and external differences (e.g. camera settings). Their performance is limited in highlighting regions on Letter 'E' in scene Letters and the leaf in scene Mailbox, and results in artifacts on the window in scene Mailbox and the floor in scene Hamster. In contrast, our method outperforms state-of-the-art methods, providing superior rendering quality and effective color correction. It demonstrates the effectiveness of our color correction module, which physically defines color differences by a nonlinear color translation matrix. + +Tab. 1 presents the quantitative results of our method compared to baseline approaches on our collected dataset, showcasing the superior rendering quality achieved by our method. Notably, NeRF-W and Ha-NeRF perform worse than the original NeRF on the scene Room. We analyze this can be attributed to the fact that the scene "Room" presents particular challenges with large reflective areas, which can be difficult for these two methods to fit a reasonable solution, and qualitative results are included in the supplementary material. Tab. 1 summarizes the quantitative comparisons on color consistency measured by Eq. (14). Wasserstein distance measures color consistency by calculating the color distribution distance between the input image and the rendered image translated with the color correction matrix of the input. Compared to the pixel-level PSNR + +![](images/3f912dbf6e4a7969d3e691498752f6bfc475a1278bf203a59aab85f376f50c96.jpg) +Fig. 4: Ablation study of gray-world loss $\mathcal{L}_g$ . Removing $\mathcal{L}_g$ will be unable to constrain the distribution of the scene color space, resulting in undesirable artifacts or color shifts in the reconstructed scene (e.g. overexposed Hamster and over-saturated Pumpkin). + +metric, the Wasserstein metric measures the entire color distribution. These comparative results demonstrate the effectiveness of our approach in addressing the limitations of existing NeRF-based models, ultimately leading to enhanced color correction and consistency in various camera conditions. + +# 4.4 Ablation Studies + +We conduct ablative evaluations on our collected dataset to show the effectiveness of our model design. We compare our model with four variants: model without specular loss (Ours w/o $\mathcal{L}_s$ ), model without gray-world loss (Ours w/o $\mathcal{L}_g$ ), model without histogram loss (Ours w/o $\mathcal{L}_h$ ), and model without nonlinear mapping (Ours w/o $\varPhi$ ). The quantitative comparisons presented in Fig. 5 indicate that each component is crucial for generating physically plausible color-corrected renderings. Additionally, Fig. 4 displays the qualitative results of the ablation study on view synthesis with the unified color NeRF. This further highlights the importance of gray-world loss in our model. When the gray-world loss is removed, the color distribution of the scene color space becomes unconstrained, and the increased solution space can be detrimental to convergence. While this may not significantly impact the rendering quality and color consistency of novel views, it could introduce undesirable artifacts or color shifts in the reconstructed scene, such as overexposed regions of Hamster shown in Fig. 4. + +In addition, we provide qualitative demonstrations of the results with/without specular loss $\mathcal{L}_s$ , histogram loss $\mathcal{L}_h$ and nonlinear mapping $\varPhi$ . The left bottom half of the image represents the view-dependent color $C_s^{cam}$ , and the remaining part displays the entire renderings $C^{cam}$ in camera color space of ground truth. Specifically, removing specular loss will not decompose the view-dependent and view-independent colors of Letters, resulting in a significant break in color consistency (also shown in $W_2$ metric). Eliminating the histogram loss makes it challenging to represent the entire color distribution when relying solely on pixel-level + +![](images/38e0f1468d3da6358ed37a2e4d896e369c89143b5d128ba1c1b06aa249968be9.jpg) +Ground Truth + +![](images/040a3170eaf9e2b815c43b105a2dab25deef701704885a4ac978be1e68ed6ddf.jpg) +Ours w/o $\mathcal{L}_s$ + +![](images/f8af0581431f971b575d1fe9eff955a8ff7b31b08f895c94e027407b7e482f6f.jpg) +Ours w/o $\mathcal{L}_h$ + +![](images/907a351ec30140b91f866a5b6b7ef58984a8e180e3368f9700b6b5b9843fd715.jpg) +Quantitative Comparisons + +![](images/031f2d3b90f4c0d7bdd144659732162d0012cb5800e9299391c9ab148ef642a6.jpg) +Ours w/o $\Phi$ +Fig. 5: Ablation Study on "Letter", including qualitative and quantitative results with/without specular loss $\mathcal{L}_s$ , histogram loss $\mathcal{L}_h$ and nonlinear mapping $\varPhi$ . The left bottom half of the image is the view-dependent color $C_s^{cam}$ , and the other part is the entire color-correct renderings $C^{cam}$ . Removing specular loss struggles to the decomposition of view-dependent and view-independent colors. Removing histogram loss lacks the constraint of the entire color space, making the color offset. Downgrading the color translation matrix suffers from artifacts of the nonlinear color difference. + +![](images/ef92f0e33ef81c240635e19ab5811d1ccc8ed9ac0efe23f070b842adb6ab9637.jpg) +Ours + +MSE error. This may result in color shifts, such as producing darker images than the ground truth. Removing nonlinear mapping and downgrading color translation matrix to $3 \times 3$ will neglect the nonlinear process during imaging. This forces the network to model more color variances into $\mathbf{C}_s^{cam}$ , making the color correction module unable to fully perceive the color characteristics of each camera, as shown in Fig. 5. These ablation studies demonstrate the effectiveness of our model and specifically designed losses. + +# 4.5 Color Translation + +Fig. 6 presents examples of color translation guided by a reference image. This advantage makes our method particularly suitable for applications where accurate color reproduction and adherence to reference images are crucial, such as in virtual reality, computer graphics, and image synthesis. + +# 5 Conclusion + +This paper proposes a framework that incorporates a color correction module within a NeRF to achieve unified color radiance field reconstruction from multiview images captured under different camera settings and lighting conditions. Specifically, we have developed a color correction module to predict a nonlinear color translation matrix that simulates the physical imaging process. Besides the + +![](images/77315023113e85d70c608821a9b80273c541a305f6d827f9ba101c3868544750.jpg) +Fig. 6: Examples of color translation. Our method produces spatially and temporally consistent color transfer results that faithfully match the reference image while maintaining photorealism. Video can also be found in supplementary materials. + +color correction module for accommodating external differences across distinct cameras, we predict a view-dependent function to separate inherent differences caused by scene properties. To enhance the accuracy of the color translation, we have designed various regularization strategies and demonstrated their efficacy through experiments. Our approach significantly improves color consistency in NeRF reconstruction, enabling users to render novel views under flexible and unrestricted settings. + +Despite its advantages, we acknowledge that our framework has certain limitations. For instance, compared to NeRF, the proposed Histogram loss results in an increased training time of approximately $10\%$ . Furthermore, the Histogram loss, which computes the difference between 2D images, performs well when the rendered image and reference image are nearby. This limitation could potentially be addressed by introducing a new representation and a 3D Histogram loss. We plan to explore these improvements in future work. + +# References + +1. Affi, M., Barron, J.T., LeGendre, C., Tsai, Y.T., Bleibel, F.: Cross-camera convolutional color constancy. In: Proceedings of the IEEE/CVF International Conference on Computer Vision. pp. 1981-1990 (2021) +2. Affi, M., Brown, M.S.: Interactive white balancing for camera-rendered images. arXiv preprint arXiv:2009.12632 (2020) +3. Affifi, M., Brubaker, M.A., Brown, M.S.: Auto white-balance correction for mixed-illuminant scenes. In: Proceedings of the IEEE/CVF Winter Conference on Applications of Computer Vision. pp. 1210–1219 (2022) +4. Barron, J.T., Mildenhall, B., Tancik, M., Hedman, P., Martin-Brualla, R., Srinivasan, P.P.: Mip-nerf: A multiscale representation for anti-aliasing neural radiance fields. In: Proceedings of the IEEE/CVF International Conference on Computer Vision. pp. 5855-5864 (2021) +5. Barron, J.T., Mildenhall, B., Verbin, D., Srinivasan, P.P., Hedman, P.: Zip-nerf: Anti-aliased grid-based neural radiance fields. arXiv preprint arXiv:2304.06706 (2023) +6. Barron, J.T., Tsai, Y.T.: Fast fourier color constancy. In: Proceedings of the IEEE conference on computer vision and pattern recognition. pp. 886-894 (2017) +7. Bian, W., Wang, Z., Li, K., Bian, J.W., Prisacariu, V.A.: Nope-nerf: Optimising neural radiance field with no pose prior. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 4160-4169 (2023) +8. Bianco, S., Bruna, A., Naccari, F., Schettini, R.: Color space transformations for digital photography exploiting information about the illuminant estimation process. JOSA A 29(3), 374-384 (2012) +9. Boss, M., Braun, R., Jampani, V., Barron, J.T., Liu, C., Lensch, H.: Nerd: Neural reflectance decomposition from image collections. In: Proceedings of the IEEE/CVF International Conference on Computer Vision. pp. 12684-12694 (2021) +0. Boss, M., Jampani, V., Braun, R., Liu, C., Barron, J., Lensch, H.: Neural-pil: Neural pre-integrated lighting for reflectance decomposition. Advances in Neural Information Processing Systems 34, 10691-10704 (2021) +1. Buchsbaum, G.: A spatial processor model for object colour perception. Journal of the Franklin institute 310(1), 1-26 (1980) +2. Cao, A., Johnson, J.: Hexplane: A fast representation for dynamic scenes. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 130-141 (2023) +3. Chen, A., Xu, Z., Geiger, A., Yu, J., Su, H.: Tensorf: Tensorial radiance fields. In: European Conference on Computer Vision. pp. 333-350. Springer (2022) +4. Chen, A., Xu, Z., Zhao, F., Zhang, X., Xiang, F., Yu, J., Su, H.: Mvsnerf: Fast generalizable radiance field reconstruction from multi-view stereo. In: Proceedings of the IEEE/CVF international conference on computer vision. pp. 14124-14133 (2021) +5. Chen, X., Zhang, Q., Li, X., Chen, Y., Feng, Y., Wang, X., Wang, J.: Hallucinated neural radiance fields in the wild. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 12943-12952 (2022) +6. Chen, Y., Chen, X., Wang, X., Zhang, Q., Guo, Y., Shan, Y., Wang, F.: Local-to-global registration for bundle-adjusting neural radiance fields. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 8264-8273 (2023) + +17. Chen, Z., Funkhouser, T., Hedman, P., Tagliasacchi, A.: Mobilenerf: Exploiting the polygon rasterization pipeline for efficient neural field rendering on mobile architectures. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 16569-16578 (2023) +18. Cheng, W., Chen, R., Fan, S., Yin, W., Chen, K., Cai, Z., Wang, J., Gao, Y., Yu, Z., Lin, Z., et al.: Dna-rendering: A diverse neural actor repository for high-fidelity human-centric rendering. In: Proceedings of the IEEE/CVF International Conference on Computer Vision. pp. 19982-19993 (2023) +19. Chibane, J., Bansal, A., Lazova, V., Pons-Moll, G.: Stereo radiance fields (srf): Learning view synthesis for sparse views of novel scenes. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 7911-7920 (2021) +20. Ding, C., Ma, Z.: Multi-camera color correction via hybrid histogram matching. IEEE Transactions on Circuits and Systems for Video Technology 31(9), 3327-3337 (2020) +21. Fang, J., Yi, T., Wang, X., Xie, L., Zhang, X., Liu, W., Nießner, M., Tian, Q.: Fast dynamic radiance fields with time-aware neural voxels. In: SIGGRAPH Asia 2022 Conference Papers. pp. 1-9 (2022) +22. Fang, Q., Shuai, Q., Dong, J., Bao, H., Zhou, X.: Reconstructing 3d human pose by watching humans in the mirror. In: CVPR (2021) +23. Fridovich-Keil, S., Meanti, G., Warburg, F.R., Recht, B., Kanazawa, A.: K-planes: Explicit radiance fields in space, time, and appearance. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 12479-12488 (2023) +24. Fridovich-Keil, S., Yu, A., Tancik, M., Chen, Q., Recht, B., Kanazawa, A.: Plenoxels: Radiance fields without neural networks. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 5501-5510 (2022) +25. Garbin, S.J., Kowalski, M., Johnson, M., Shotton, J., Valentin, J.: Fastnerf: High-fidelity neural rendering at 200fps. In: Proceedings of the IEEE/CVF International Conference on Computer Vision. pp. 14346-14355 (2021) +26. Hedman, P., Srinivasan, P.P., Mildenhall, B., Barron, J.T., Debevec, P.: Baking neural radiance fields for real-time view synthesis. In: Proceedings of the IEEE/CVF International Conference on Computer Vision. pp. 5875-5884 (2021) +27. Huang, X., Zhang, Q., Feng, Y., Li, H., Wang, Q.: Inverting the imaging process by learning an implicit camera model. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 21456-21465 (2023) +28. Huang, X., Zhang, Q., Feng, Y., Li, H., Wang, X., Wang, Q.: HDR-nerf: High dynamic range neural radiance fields. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 18398-18408 (2022) +29. Huang, X., Zhang, Q., Feng, Y., Li, X., Wang, X., Wang, Q.: Local implicit ray function for generalizable radiance field representation. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 97-107 (2023) +30. Huang, Z., Liang, Z., Zhang, H., Lin, Y., Jia, K.: Sur2f: A hybrid representation for high-quality and efficient surface reconstruction from multi-view images. arXiv preprint arXiv:2401.03704 (2024) +31. Isik, M., Rünz, M., Georgopoulos, M., Khakhulin, T., Starck, J., Agapito, L., Nießner, M.: Humanrf: High-fidelity neural radiance fields for humans in motion. arXiv preprint arXiv:2305.06356 (2023) + +32. Jin, H., Liu, I., Xu, P., Zhang, X., Han, S., Bi, S., Zhou, X., Xu, Z., Su, H.: Tensoroir: Tensorial inverse rendering. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 165-174 (2023) +33. Johari, M.M., Lepoittevin, Y., Fleuret, F.: Geonerf: Generalizing nerf with geometry priors. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 18365-18375 (2022) +34. Joo, H., Simon, T., Li, X., Liu, H., Tan, L., Gui, L., Banerjee, S., Godisart, T.S., Nabbe, B., Matthews, I., Kanade, T., Nobuhara, S., Sheikh, Y.: Panoptic studio: A massively multiview system for social interaction capture. IEEE Transactions on Pattern Analysis and Machine Intelligence (2017) +35. Kerbl, B., Kopanas, G., Leimkuhler, T., Drettakis, G.: 3d gaussian splatting for real-time radiance field rendering. ACM Transactions on Graphics (ToG) 42(4), 1-14 (2023) +36. Kinh, F., Yilmaz, D., Özcan, B., Kırac, F.: Deterministic neural illumination mapping for efficient auto-white balance correction. In: Proceedings of the IEEE/CVF International Conference on Computer Vision. pp. 1139-1147 (2023) +37. Kinh, F., Yilmaz, D., Ozcan, B., Kirac, F.: Modeling the lighting in scenes as style for auto white-balance correction. In: Proceedings of the IEEE/CVF Winter Conference on Applications of Computer Vision. pp. 4903-4913 (2023) +38. Kuang, Z., Luan, F., Bi, S., Shu, Z., Wetzstein, G., Sunkavalli, K.: Palettenerf: Palette-based appearance editing of neural radiance fields. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 20691-20700 (2023) +39. Lee, J.H., Kim, D.S.: Ice-nerf: Interactive color editing of nerfs via decomposition-aware weight optimization. In: Proceedings of the IEEE/CVF International Conference on Computer Vision. pp. 3491-3501 (2023) +40. Li, J., Feng, Z., She, Q., Ding, H., Wang, C., Lee, G.H.: Mine: Towards continuous depth MPI with nerf for novel view synthesis. In: Proceedings of the IEEE/CVF International Conference on Computer Vision. pp. 12578-12588 (2021) +41. Liang, Z., Huang, Z., Ding, C., Jia, K.: Helixsurf: A robust and efficient neural implicit surface learning of indoor scenes with iterative intertwined regularization. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 13165-13174 (2023) +42. Lin, C.H., Ma, W.C., Torralba, A., Lucey, S.: Barf: Bundle-adjusting neural radiance fields. In: Proceedings of the IEEE/CVF international conference on computer vision. pp. 5741-5751 (2021) +43. Liu, Y.L., Gao, C., Meuleman, A., Tseng, H.Y., Saraf, A., Kim, C., Chuang, Y.Y., Kopf, J., Huang, J.B.: Robust dynamic radiance fields. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 13-23 (2023) +44. Liu, Y., Peng, S., Liu, L., Wang, Q., Wang, P., Theobalt, C., Zhou, X., Wang, W.: Neural rays for occlusion-aware image-based rendering. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 7824-7833 (2022) +45. Ma, L., Li, X., Liao, J., Zhang, Q., Wang, X., Wang, J., Sander, P.V.: Deblur-nerf: Neural radiance fields from blurry images. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 12861-12870 (2022) +46. Martin-Brualla, R., Radwan, N., Sajjadi, M.S., Barron, J.T., Dosovitskiy, A., Duckworth, D.: Nerf in the wild: Neural radiance fields for unconstrained photo collections. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 7210-7219 (2021) + +47. Max, N.: Optical models for direct volume rendering. IEEE Transactions on Visualization and Computer Graphics 1(2), 99-108 (1995) +48. Mildenhall, B., Hedman, P., Martin-Brualla, R., Srinivasan, P.P., Barron, J.T.: Nerf in the dark: High dynamic range view synthesis from noisy raw images. In: Proceedings of the IEEE/CVF conference on computer vision and pattern recognition. pp. 16190-16199 (2022) +49. Mildenhall, B., Srinivasan, P.P., Tancik, M., Barron, J.T., Ramamoorthi, R., Ng, R.: Nerf: Representing scenes as neural radiance fields for view synthesis. In: European Conference on Computer Vision. pp. 405-421 (2020) +50. Müller, T., Evans, A., Schied, C., Keller, A.: Instant neural graphics primitives with a multiresolution hash encoding. ACM Transactions on Graphics (ToG) 41(4), 1-15 (2022) +51. Park, K., Henzler, P., Mildenhall, B., Barron, J.T., Martin-Brualla, R.: Camp: Camera preconditioning for neural radiance fields. ACM Transactions on Graphics (TOG) 42(6), 1-11 (2023) +52. Pearl, N., Treibitz, T., Korman, S.: Nan: Noise-aware nerfs for burst-denoising. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 12672-12681 (2022) +53. Peng, S., Zhang, Y., Xu, Y., Wang, Q., Shuai, Q., Bao, H., Zhou, X.: Neural body: Implicit neural representations with structured latent codes for novel view synthesis of dynamic humans. In: CVPR (2021) +54. Reiser, C., Peng, S., Liao, Y., Geiger, A.: Kilonerf: Speeding up neural radiance fields with thousands of tiny mlp's. In: Proceedings of the IEEE/CVF International Conference on Computer Vision. pp. 14335-14345 (2021) +55. Rematas, K., Liu, A., Srinivasan, P.P., Barron, J.T., Tagliasacchi, A., Funkhouser, T., Ferrari, V.: Urban radiance fields. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 12932-12942 (2022) +56. Rodríguez, R.G., Vazquez-Corral, J., Bertalmío, M.: Color matching images with unknown non-linear encodings. IEEE Transactions on Image Processing 29, 4435-4444 (2020) +57. Shao, R., Zheng, Z., Tu, H., Liu, B., Zhang, H., Liu, Y.: Tensor4d: Efficient neural 4d decomposition for high-fidelity dynamic reconstruction and rendering. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 16632-16642 (2023) +58. Srinivasan, P.P., Deng, B., Zhang, X., Tancik, M., Mildenhall, B., Barron, J.T.: Nerv: Neural reflectance and visibility fields for relighting and view synthesis. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 7495-7504 (2021) +59. Sun, G., Chen, X., Chen, Y., Pang, A., Lin, P., Jiang, Y., Xu, L., Yu, J., Wang, J.: Neural free-viewpoint performance rendering under complex human-object interactions. In: Proceedings of the 29th ACM International Conference on Multimedia. pp. 4651-4660 (2021) +60. Tancik, M., Casser, V., Yan, X., Pradhan, S., Mildenhall, B., Srinivasan, P.P., Barron, J.T., Kretzschmar, H.: Block-nerf: Scalable large scene neural view synthesis. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 8248-8258 (2022) +61. Tang, J., Zhou, H., Chen, X., Hu, T., Ding, E., Wang, J., Zeng, G.: Delicate textured mesh recovery from nef via adaptive surface refinement. arXiv preprint arXiv:2303.02091 (2023) + +62. Verbin, D., Hedman, P., Mildenhall, B., Zickler, T., Barron, J.T., Srinivasan, P.P.: Ref-nerf: Structured view-dependent appearance for neural radiance fields. In: 2022 IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR). pp. 5481-5490. IEEE (2022) +63. Wang, Q., Wang, Z., Genova, K., Srinivasan, P.P., Zhou, H., Barron, J.T., Martin-Brualla, R., Snavely, N., Funkhouser, T.: Ibrnet: Learning multi-view image-based rendering. In: Proceedings of the IEEE/CVF conference on computer vision and pattern recognition. pp. 4690-4699 (2021) +64. Xie, S., Zhu, H., Liu, Z., Zhang, Q., Zhou, Y., Cao, X., Ma, Z.: Diner: Disorder-invariant implicit neural representation. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 6143-6152 (2023) +65. Xu, Q., Xu, Z., Philip, J., Bi, S., Shu, Z., Sunkavalli, K., Neumann, U.: Point-nerf: Point-based neural radiance fields. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 5438-5448 (2022) +66. Yao, Y., Zhang, J., Liu, J., Qu, Y., Fang, T., McKinnon, D., Tsin, Y., Quan, L.: Neilf: Neural incident light field for physically-based material estimation. In: European Conference on Computer Vision. pp. 700-716. Springer (2022) +67. Yen-Chen, l.: Nerf-pytorch: a pytorch implementation of nerf. URL https://github.com/yenchenlin/nerf-pytorch (2020) +68. Yu, A., Ye, V., Tancik, M., Kanazawa, A.: pixelnerf: Neural radiance fields from one or few images. In: Proceedings of the IEEE/CVF conference on computer vision and pattern recognition. pp. 4578-4587 (2021) +69. Zhang, K., Luan, F., Li, Z., Snavely, N.: Iron: Inverse rendering by optimizing neural sdfs and materials from photometric images. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 5565-5574 (2022) +70. Zhang, R., Isola, P., Efros, A.A., Shechtman, E., Wang, O.: The unreasonable effectiveness of deep features as a perceptual metric. In: Proceedings of the IEEE conference on computer vision and pattern recognition. pp. 586-595 (2018) +71. Zhang, X., Srinivasan, P.P., Deng, B., Debevec, P., Freeman, W.T., Barron, J.T.: Nerfactor: Neural factorization of shape and reflectance under an unknown illumination. ACM Transactions on Graphics (ToG) 40(6), 1-18 (2021) +72. Zhao, Y., Ferguson, S., Zhou, H., Elliott, C., Rafferty, K.: Color alignment for relative color constancy via non-standard references. IEEE Transactions on Image Processing 31, 6591-6604 (2022) +73. Zhou, T., He, K., Wu, D., Xu, T., Zhang, Q., Shao, K., Chen, W., Xu, L., Yu, J.: Relightable neural human assets from multi-view gradient illuminations. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 4315-4327 (2023) +74. Zhu, H., Liu, F., Zhang, Q., Cao, X., Ma, Z.: Rhino: Regularizing the hash-based implicit neural representation. arXiv preprint arXiv:2309.12642 (2023) +75. Zhu, H., Xie, S., Liu, Z., Liu, F., Zhang, Q., Zhou, Y., Lin, Y., Ma, Z., Cao, X.: Disorder-invariant implicit neural representation. IEEE Transactions on Pattern Analysis and Machine Intelligence (2024) +76. Zhu, J., Zhu, H., Zhang, Q., Zhu, F., Ma, Z., Cao, X.: Pyramid nef: Frequency guided fast radiance field optimization. International Journal of Computer Vision 131(10), 2649-2664 (2023) +77. Zhuang, Y., Zhang, Q., Wang, X., Zhu, H., Feng, Y., Li, X., Shan, Y., Cao, X.: A pre-convolved representation for plug-and-play neural illumination fields. In: Proceedings of the AAAI Conference on Artificial Intelligence. vol. 38, pp. 7828-7836 (2024) \ No newline at end of file diff --git a/2024/Physically Plausible Color Correction for Neural Radiance Fields/images.zip b/2024/Physically Plausible Color Correction for Neural Radiance Fields/images.zip new file mode 100644 index 0000000000000000000000000000000000000000..4c15df304592c2c5d9d6c2c994c17da5366a54ff --- /dev/null +++ b/2024/Physically Plausible Color Correction for Neural Radiance Fields/images.zip @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:b1be484e27c0d2d4c143b55e6ec0a7b21920459eadb68545a62995ad7c51c3ec +size 639752 diff --git a/2024/Physically Plausible Color Correction for Neural Radiance Fields/layout.json b/2024/Physically Plausible Color Correction for Neural Radiance Fields/layout.json new file mode 100644 index 0000000000000000000000000000000000000000..ffa44f2fd7b9076d1ec9cbd00623107b964e84b4 --- /dev/null +++ b/2024/Physically Plausible Color Correction for Neural Radiance Fields/layout.json @@ -0,0 +1,10761 @@ +{ + "pdf_info": [ + { + "para_blocks": [ + { + "bbox": [ + 137, + 111, + 477, + 148 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 137, + 111, + 477, + 148 + ], + "spans": [ + { + "bbox": [ + 137, + 111, + 477, + 148 + ], + "type": "text", + "content": "Physically Plausible Color Correction for Neural Radiance Fields" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 195, + 167, + 419, + 181 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 195, + 167, + 419, + 181 + ], + "spans": [ + { + "bbox": [ + 195, + 167, + 419, + 181 + ], + "type": "text", + "content": "Qi Zhang" + }, + { + "bbox": [ + 195, + 167, + 419, + 181 + ], + "type": "inline_equation", + "content": "^{1,\\dagger}" + }, + { + "bbox": [ + 195, + 167, + 419, + 181 + ], + "type": "text", + "content": ", Ying Feng" + }, + { + "bbox": [ + 195, + 167, + 419, + 181 + ], + "type": "inline_equation", + "content": "^{1,\\dagger}" + }, + { + "bbox": [ + 195, + 167, + 419, + 181 + ], + "type": "text", + "content": ", and Hongdong Li" + }, + { + "bbox": [ + 195, + 167, + 419, + 181 + ], + "type": "inline_equation", + "content": "^{2}" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 173, + 190, + 441, + 201 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 173, + 190, + 441, + 201 + ], + "spans": [ + { + "bbox": [ + 173, + 190, + 441, + 201 + ], + "type": "inline_equation", + "content": "^{1}" + }, + { + "bbox": [ + 173, + 190, + 441, + 201 + ], + "type": "text", + "content": " Independent Researcher {nwpuqzhang, yfeng.von}@gmail.com" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 185, + 201, + 428, + 212 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 185, + 201, + 428, + 212 + ], + "spans": [ + { + "bbox": [ + 185, + 201, + 428, + 212 + ], + "type": "text", + "content": "2 Australian National University hongdong.li@anu.edu.au" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 160, + 240, + 455, + 494 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 160, + 240, + 455, + 494 + ], + "spans": [ + { + "bbox": [ + 160, + 240, + 455, + 494 + ], + "type": "text", + "content": "Abstract. Neural Radiance Fields have become the representation of choice for many 3D computer vision and computer graphics applications, e.g., novel view synthesis and 3D reconstruction. Multi-camera systems are commonly used as the image capture setup in NeRF-based multi-view tasks such as dynamic scene acquisition or realistic avatar animation. However, a critical issue that has often been overlooked in this setup is the evident differences in color responses among multiple cameras, which adversely affect the NeRF reconstruction performance. These color discrepancies among multiple input images stem from two aspects: 1) implicit properties of the scenes such as reflections and shadings, and 2) external differences in camera settings and lighting conditions. In this paper, we address this problem by proposing a novel color correction module that simulates the physical color processing in cameras to be embedded in NeRF, enabling the unified color NeRF reconstruction. Besides the view-independent color correction module for external differences, we predict a view-dependent function to minimize the color residual (including, e.g., specular and shading) to eliminate the impact of inherent attributes. We further describe how the method can be extended with a reference image as guidance to achieve aesthetically plausible color consistency and color translation on novel views. Experiments validate that our method is superior to baseline methods in both quantitative and qualitative evaluations of color correction and color consistency." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 160, + 503, + 453, + 514 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 160, + 503, + 453, + 514 + ], + "spans": [ + { + "bbox": [ + 160, + 503, + 453, + 514 + ], + "type": "text", + "content": "Keywords: Neural Radiance Field " + }, + { + "bbox": [ + 160, + 503, + 453, + 514 + ], + "type": "inline_equation", + "content": "\\cdot" + }, + { + "bbox": [ + 160, + 503, + 453, + 514 + ], + "type": "text", + "content": " View Synthesis " + }, + { + "bbox": [ + 160, + 503, + 453, + 514 + ], + "type": "inline_equation", + "content": "\\cdot" + }, + { + "bbox": [ + 160, + 503, + 453, + 514 + ], + "type": "text", + "content": " Color Translation" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 132, + 552, + 230, + 565 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 552, + 230, + 565 + ], + "spans": [ + { + "bbox": [ + 132, + 552, + 230, + 565 + ], + "type": "text", + "content": "1 Introduction" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 130, + 576, + 482, + 636 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 576, + 482, + 636 + ], + "spans": [ + { + "bbox": [ + 130, + 576, + 482, + 636 + ], + "type": "text", + "content": "The synthesis and rendering of novel views using multi-view images has been a long-standing problem in computer vision and graphics, which is also a cornerstone task for augmented reality (AR) and virtual reality (VR) applications. The recent success of Neural Radiance Fields (NeRF) [49] has made the implicit neural representation a powerful approach for novel view synthesis tasks, as" + } + ] + } + ], + "index": 7 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 141, + 643, + 344, + 654 + ], + "type": "page_footnote", + "angle": 0, + "lines": [ + { + "bbox": [ + 141, + 643, + 344, + 654 + ], + "spans": [ + { + "bbox": [ + 141, + 643, + 344, + 654 + ], + "type": "text", + "content": "† These authors contributed equally to this work." + } + ] + } + ], + "index": 8 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 0 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 136, + 118, + 258, + 220 + ], + "blocks": [ + { + "bbox": [ + 136, + 118, + 258, + 220 + ], + "lines": [ + { + "bbox": [ + 136, + 118, + 258, + 220 + ], + "spans": [ + { + "bbox": [ + 136, + 118, + 258, + 220 + ], + "type": "image", + "image_path": "fb49fe45339def8d81f97800bf122c05e3117ff359c3d5dc37817359d99b0369.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 176, + 220, + 218, + 228 + ], + "lines": [ + { + "bbox": [ + 176, + 220, + 218, + 228 + ], + "spans": [ + { + "bbox": [ + 176, + 220, + 218, + 228 + ], + "type": "text", + "content": "(a) input views" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 130, + 237, + 482, + 316 + ], + "lines": [ + { + "bbox": [ + 130, + 237, + 482, + 316 + ], + "spans": [ + { + "bbox": [ + 130, + 237, + 482, + 316 + ], + "type": "text", + "content": "Fig. 1: Physically Plausible Color Correction for Neural Radiance Fields. (a) Given a set of images captured under various camera settings and lighting conditions, our method introduces a color correction module into the NeRF-based model, which enables recovering (b) a unified color radiance field with a physically plausible white balance. Unlike prior methods that embed color through features, our color correction module simulates color differences via a nonlinear color translation matrix. Additionally, our method can render (c) novel views translated from the color of a reference image." + } + ] + } + ], + "index": 8, + "angle": 0, + "type": "image_caption" + } + ], + "index": 2 + }, + { + "type": "image", + "bbox": [ + 262, + 119, + 394, + 219 + ], + "blocks": [ + { + "bbox": [ + 262, + 119, + 394, + 219 + ], + "lines": [ + { + "bbox": [ + 262, + 119, + 394, + 219 + ], + "spans": [ + { + "bbox": [ + 262, + 119, + 394, + 219 + ], + "type": "image", + "image_path": "4cc89a1e3946176c2dafe27401ccaa006ca41ef98f98cbffe0ed4f1d7fa52e4c.jpg" + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 288, + 220, + 373, + 228 + ], + "lines": [ + { + "bbox": [ + 288, + 220, + 373, + 228 + ], + "spans": [ + { + "bbox": [ + 288, + 220, + 373, + 228 + ], + "type": "text", + "content": "(b) unified color radiance field" + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "image_caption" + } + ], + "index": 4 + }, + { + "type": "image", + "bbox": [ + 399, + 119, + 471, + 217 + ], + "blocks": [ + { + "bbox": [ + 399, + 119, + 471, + 217 + ], + "lines": [ + { + "bbox": [ + 399, + 119, + 471, + 217 + ], + "spans": [ + { + "bbox": [ + 399, + 119, + 471, + 217 + ], + "type": "image", + "image_path": "ed811e9f51221c5b7e2d444c4f336950faf915831c729ce96247cd8d1550d6c7.jpg" + } + ] + } + ], + "index": 6, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 410, + 218, + 465, + 228 + ], + "lines": [ + { + "bbox": [ + 410, + 218, + 465, + 228 + ], + "spans": [ + { + "bbox": [ + 410, + 218, + 465, + 228 + ], + "type": "text", + "content": "(c) color translation" + } + ] + } + ], + "index": 7, + "angle": 0, + "type": "image_caption" + } + ], + "index": 6 + }, + { + "bbox": [ + 130, + 340, + 482, + 400 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 340, + 482, + 400 + ], + "spans": [ + { + "bbox": [ + 130, + 340, + 482, + 400 + ], + "type": "text", + "content": "an alternative to traditional image-based reconstruction techniques. The NeRF model employs Multi-Layer Perceptrons (MLPs) to parameterize the mapping from spatial coordinates to corresponding densities and view-dependent colors. This approach has been proven effective with scenes that are consistently well-lit without substantial variations in image brightness." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 130, + 401, + 482, + 593 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 401, + 482, + 593 + ], + "spans": [ + { + "bbox": [ + 130, + 401, + 482, + 593 + ], + "type": "text", + "content": "With the increasing prevalence of multi-camera systems [18, 22, 31, 34, 53, 59, 73], NeRF and its variants [4, 5, 50] face challenges in achieving photorealistic novel view synthesis with color consistency. The differences in color response among multiple cameras can significantly impact the accuracy and consistency of synthesized views. Even when capturing with identical camera settings, different types of cameras may present distinct color responses caused by the difference in sensors and ISP algorithms. To address this limitation, NeRF-W [46] and Ha-NeRF [15] incorporate additional latent codes to encode appearances into the NeRF model for consistent view synthesis and appearance control. However, these methods tend to embed the overall appearances as view-dependent features, making it difficult to achieve fine-grained and intuitive color correction. HDR-NeRF [28] attempts to address this issue but primarily focuses on building a physical module for camera exposures. It is necessary to introduce color translation for NeRF-based multi-view tasks, which should involve modeling color correlations across cameras and enabling physically plausible color correction and consistency in NeRF reconstructions." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 130, + 594, + 483, + 666 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 594, + 483, + 666 + ], + "spans": [ + { + "bbox": [ + 130, + 594, + 483, + 666 + ], + "type": "text", + "content": "The color difference across multi-view images arises from two aspects: inherent view-dependent scene properties (e.g. reflections and shadings); external differences in camera settings (e.g. white balance, ISO, exposure) and lighting conditions. Naturally, we could calibrate each image via color correction methods [1,3,6,36,37], and then input to NeRF-based method for novel view synthesis. However, these image-based methods struggle to preserve color consistency in" + } + ] + } + ], + "index": 11 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "spans": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "type": "text", + "content": "2" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 166, + 91, + 267, + 102 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 166, + 91, + 267, + 102 + ], + "spans": [ + { + "bbox": [ + 166, + 91, + 267, + 102 + ], + "type": "text", + "content": "Q. Zhang, Y. Feng et al." + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 1 + }, + { + "para_blocks": [ + { + "bbox": [ + 130, + 116, + 479, + 185 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 116, + 479, + 185 + ], + "spans": [ + { + "bbox": [ + 130, + 116, + 479, + 185 + ], + "type": "text", + "content": "NeRF reconstructions. In addition to faithfully reproducing the captured colors, we consider that color correction process should allow for color adjustment to enhance visual appeal, and even color translation to a given reference image. By incorporating such color correction module, users, including novice ones, can easily recover NeRF and adjust the colors of novel views based on their preferences." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 130, + 188, + 482, + 331 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 188, + 482, + 331 + ], + "spans": [ + { + "bbox": [ + 130, + 188, + 482, + 331 + ], + "type": "text", + "content": "In this paper, we propose a novel framework that incorporates a color correction module into the NeRF-based model to achieve unified color radiance field reconstruction from multi-view images with different camera settings and lighting conditions, as shown in Fig. 1. In contrast to the embedding feature for overall color differences in prior NeRF-based methods [15,46], we present a color correction module to predict a nonlinear color translation matrix that simulates the color processing in cameras, such as white balance, nonlinear color enhancement, tone-manipulation, gamma encoding, etc. Additionally, considering the view-dependent factors of inherent properties, we predict a view-dependent function to separate the color residual (e.g. specular and shading of the scene) to improve the accuracy of the color translation matrix. Our technical contributions are summarized as follows:" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 138, + 339, + 479, + 431 + ], + "type": "list", + "angle": 0, + "index": 7, + "blocks": [ + { + "bbox": [ + 138, + 339, + 479, + 373 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 339, + 479, + 373 + ], + "spans": [ + { + "bbox": [ + 138, + 339, + 479, + 373 + ], + "type": "text", + "content": "- We present a novel framework for color correction and translation in NeRF reconstruction, allowing even novice users to perform novel view synthesis under unstable camera settings in an intuitive manner." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 138, + 375, + 479, + 396 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 375, + 479, + 396 + ], + "spans": [ + { + "bbox": [ + 138, + 375, + 479, + 396 + ], + "type": "text", + "content": "- We propose a color correction module that defines the non-linear color translation to be embedded in NeRF." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 138, + 399, + 479, + 431 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 399, + 479, + 431 + ], + "spans": [ + { + "bbox": [ + 138, + 399, + 479, + 431 + ], + "type": "text", + "content": "- We introduce tailored regularizations to eliminate the impact of inherent scene properties, constrain the scene's NeRF to a white balance color space, and maintain the entire color distribution." + } + ] + } + ], + "index": 6 + } + ], + "sub_type": "text" + }, + { + "bbox": [ + 130, + 442, + 479, + 466 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 442, + 479, + 466 + ], + "spans": [ + { + "bbox": [ + 130, + 442, + 479, + 466 + ], + "type": "text", + "content": "Extensive experiments have shown that our method outperforms baseline methods both quantitatively and qualitatively in terms of color correction." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 131, + 484, + 242, + 498 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 131, + 484, + 242, + 498 + ], + "spans": [ + { + "bbox": [ + 131, + 484, + 242, + 498 + ], + "type": "text", + "content": "2 Related Works" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 130, + 510, + 482, + 666 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 510, + 482, + 666 + ], + "spans": [ + { + "bbox": [ + 130, + 510, + 482, + 666 + ], + "type": "text", + "content": "Neural Radiance Fields Neural Radiance Fields represent a scene implicitly with multi-layer perceptrons, approximating a continuous function that maps each 3D position and direction to a density and radiance [49]. By incorporating view direction features, NeRF can effectively model view-dependent effects such as reflection and refraction. This allows for photo-realistic synthesis of novel views with volumetric rendering. To alleviate the computational limitation of massive MLP evaluations while maintaining the high-frequency appearances, many extensive works have been proposed to speed up the training and inference of the model [5,13,17,24-26,35,50,54,64,65,74-76] with various geometry proxies and neural features, explore view synthesis for dynamic scenes [12,21,23,43,57], and enhance the generalization of NeRF [14,19,29,33,40,44,63,68]. Many methods [30,41,61] have been proposed to exploit the view-dependent appearance to decouple the physical properties of the scene. NeRF2Mesh [61] proposes a" + } + ] + } + ], + "index": 10 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 336, + 91, + 447, + 101 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 336, + 91, + 447, + 101 + ], + "spans": [ + { + "bbox": [ + 336, + 91, + 447, + 101 + ], + "type": "text", + "content": "Color Correction for NeRF" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 474, + 92, + 480, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 474, + 92, + 480, + 100 + ], + "spans": [ + { + "bbox": [ + 474, + 92, + 480, + 100 + ], + "type": "text", + "content": "3" + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 2 + }, + { + "para_blocks": [ + { + "bbox": [ + 130, + 116, + 482, + 224 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 116, + 482, + 224 + ], + "spans": [ + { + "bbox": [ + 130, + 116, + 482, + 224 + ], + "type": "text", + "content": "two-stage framework to extract the mesh and texture images by regularizing the view-dependent color. Many inverse rendering methods go a step further in material decoupling [9, 10, 32, 58, 62, 66, 69, 71, 77]. They decompose the radiance into physical properties such as albedo, roughness, metallicity, and lighting by introducing physical-based rendering into NeRF optimization, and enable relighting under novel lighting conditions or material editing. Numerous methods have been proposed to reconstruct scenes from images captured in uncontrolled environments, addressing challenges such as inaccurate poses [7, 16, 42, 51], image blur or noise [45, 48, 52], and inconsistent camera settings [27, 28, 55, 60]." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 130, + 224, + 482, + 499 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 224, + 482, + 499 + ], + "spans": [ + { + "bbox": [ + 130, + 224, + 482, + 499 + ], + "type": "text", + "content": "Appearance control of NeRF Images captured in natural environments or taken by multiple cameras often exhibit inconsistency in brightness and color. These image-dependent appearances may derive from changing lighting conditions, exposure variations and white balancing, or differences in camera sensors and ISP, which are not desired in novel view synthesis. HDR-NeRF [28] focuses on multi-view images of varying exposures. By modeling the simplified physical imaging process with a tone mapper, which maps the radiance of a scene point to a pixel value, HDR-NeRF reconstructs the high dynamic range radiance field, which can generate novel views with consistent exposures. Many methods enable NeRF to handle these inconsistencies by introducing related parameters or embeddings. NeRF-w [46] and Ha-NeRF [15] propose to encode the image-dependent appearance using latent codes, which can then be manipulated to control the overall style of rendered images. Block-NeRF [60] feeds an additional exposure level to the appearance prediction model to deal with the brightness variance and provides the ability to modify the global exposure conditioning during inference. URF [55] predicts a " + }, + { + "bbox": [ + 130, + 224, + 482, + 499 + ], + "type": "inline_equation", + "content": "3 \\times 3" + }, + { + "bbox": [ + 130, + 224, + 482, + 499 + ], + "type": "text", + "content": " matrix to restrictively model the white balance and exposure variations. One shortcoming of these methods is that view-dependent appearances, those either caused by changes in camera settings and environment lighting or are intrinsical from scene physical properties, are coupled together and modeled related to both latent codes and view directions. In addition to the aforementioned methods that could control scene appearances by modifying materials or latent codes, some methods [38,39] engage in appearance editing by color post-manipulation and rely heavily on the consistency of views." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 130, + 499, + 482, + 594 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 499, + 482, + 594 + ], + "spans": [ + { + "bbox": [ + 130, + 499, + 482, + 594 + ], + "type": "text", + "content": "Color Correction Color correction has been extensively studied on single 2D image and multi-camera systems. Single 2D image color correction [1,3,6,36, 37], or white balance correction, cannot guarantee view-consistent results when naively performed on each view. On the other hand, multi-camera color correction [20,56,72] relies on common area or statistical matching, which cannot handle inherent view variance and only matches the global style. To address these limitations, we propose integrating a color correction module into the NeRF optimization process to enable view-consistent color correction with 3D awareness." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 132, + 614, + 316, + 628 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 614, + 316, + 628 + ], + "spans": [ + { + "bbox": [ + 132, + 614, + 316, + 628 + ], + "type": "text", + "content": "3 Color Correction for NeRF" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 130, + 641, + 481, + 665 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 641, + 481, + 665 + ], + "spans": [ + { + "bbox": [ + 130, + 641, + 481, + 665 + ], + "type": "text", + "content": "Given multi-view images captured under various camera settings and lighting conditions, our method proposes a color correction module embedded in a NeRF-" + } + ] + } + ], + "index": 6 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 133, + 91, + 141, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 91, + 141, + 100 + ], + "spans": [ + { + "bbox": [ + 133, + 91, + 141, + 100 + ], + "type": "text", + "content": "4" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 166, + 91, + 267, + 102 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 166, + 91, + 267, + 102 + ], + "spans": [ + { + "bbox": [ + 166, + 91, + 267, + 102 + ], + "type": "text", + "content": "Q. Zhang, Y. Feng et al." + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 3 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 134, + 118, + 479, + 225 + ], + "blocks": [ + { + "bbox": [ + 134, + 118, + 479, + 225 + ], + "lines": [ + { + "bbox": [ + 134, + 118, + 479, + 225 + ], + "spans": [ + { + "bbox": [ + 134, + 118, + 479, + 225 + ], + "type": "image", + "image_path": "edbb41236e85ac1f7511e153e8e0fbe243a8187b6e4f0dd02aafb1fdc248f6f7.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 130, + 235, + 482, + 336 + ], + "lines": [ + { + "bbox": [ + 130, + 235, + 482, + 336 + ], + "spans": [ + { + "bbox": [ + 130, + 235, + 482, + 336 + ], + "type": "text", + "content": "Fig. 2: Overview of our pipeline. Given multi-view images captured under various camera settings and lighting conditions, our method recovers a unified color radiance field, which enables consistent novel view synthesis. We propose a non-linear color correction module embedded in NeRF-based model to simulate the color processing performed in cameras (e.g. white balance, exposure, nonlinear color enhancement, etc.), which maps the color from unified scene color space to each camera color space. The view-dependent color arises from inherent scene properties like specular and shading is extracted under regularization. We design a robust optimization scheme with tailored losses to achieve physically plausible color correction." + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_caption" + } + ], + "index": 2 + }, + { + "bbox": [ + 130, + 358, + 482, + 420 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 358, + 482, + 420 + ], + "spans": [ + { + "bbox": [ + 130, + 358, + 482, + 420 + ], + "type": "text", + "content": "based model, enabling the unified color NeRF reconstruction, as shown in Fig. 2. The scene radiance field is decomposed into two components to decouple the inherent color variances like specular and shading. The color correction module is designed as a " + }, + { + "bbox": [ + 130, + 358, + 482, + 420 + ], + "type": "inline_equation", + "content": "3 \\times 7" + }, + { + "bbox": [ + 130, + 358, + 482, + 420 + ], + "type": "text", + "content": " translation matrix in nonlinear RGB space. Multiple regularizations on pixels and areas are incorporated to improve color consistency." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 132, + 436, + 269, + 449 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 436, + 269, + 449 + ], + "spans": [ + { + "bbox": [ + 132, + 436, + 269, + 449 + ], + "type": "text", + "content": "3.1 Scene Representation" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 130, + 456, + 482, + 555 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 456, + 482, + 555 + ], + "spans": [ + { + "bbox": [ + 130, + 456, + 482, + 555 + ], + "type": "text", + "content": "Neural Radiance Field (NeRF) represents a scene using an implicit neural function, which utilizes MLPs to map the position " + }, + { + "bbox": [ + 130, + 456, + 482, + 555 + ], + "type": "inline_equation", + "content": "\\mathbf{x}" + }, + { + "bbox": [ + 130, + 456, + 482, + 555 + ], + "type": "text", + "content": " and view direction " + }, + { + "bbox": [ + 130, + 456, + 482, + 555 + ], + "type": "inline_equation", + "content": "\\mathbf{d}" + }, + { + "bbox": [ + 130, + 456, + 482, + 555 + ], + "type": "text", + "content": " to a density " + }, + { + "bbox": [ + 130, + 456, + 482, + 555 + ], + "type": "inline_equation", + "content": "\\sigma(\\mathbf{x})" + }, + { + "bbox": [ + 130, + 456, + 482, + 555 + ], + "type": "text", + "content": " and view-dependent color " + }, + { + "bbox": [ + 130, + 456, + 482, + 555 + ], + "type": "inline_equation", + "content": "\\mathbf{c}(\\mathbf{x},\\mathbf{d})" + }, + { + "bbox": [ + 130, + 456, + 482, + 555 + ], + "type": "text", + "content": ". To render the color of a pixel in an image from a given camera pose, NeRF casts a ray " + }, + { + "bbox": [ + 130, + 456, + 482, + 555 + ], + "type": "inline_equation", + "content": "\\mathbf{r}(t) = \\mathbf{o} + t\\mathbf{d}" + }, + { + "bbox": [ + 130, + 456, + 482, + 555 + ], + "type": "text", + "content": " through the pixel, where " + }, + { + "bbox": [ + 130, + 456, + 482, + 555 + ], + "type": "inline_equation", + "content": "\\mathbf{o}" + }, + { + "bbox": [ + 130, + 456, + 482, + 555 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 130, + 456, + 482, + 555 + ], + "type": "inline_equation", + "content": "\\mathbf{d}" + }, + { + "bbox": [ + 130, + 456, + 482, + 555 + ], + "type": "text", + "content": " are the origin and direction of the ray, respectively, and " + }, + { + "bbox": [ + 130, + 456, + 482, + 555 + ], + "type": "inline_equation", + "content": "t" + }, + { + "bbox": [ + 130, + 456, + 482, + 555 + ], + "type": "text", + "content": " is the parameter that determines the position along the ray. The predicted " + }, + { + "bbox": [ + 130, + 456, + 482, + 555 + ], + "type": "inline_equation", + "content": "\\{(\\sigma,\\mathbf{c})\\}" + }, + { + "bbox": [ + 130, + 456, + 482, + 555 + ], + "type": "text", + "content": " of sampled positions on the ray are then accumulated into the color " + }, + { + "bbox": [ + 130, + 456, + 482, + 555 + ], + "type": "inline_equation", + "content": "C(\\mathbf{r})" + }, + { + "bbox": [ + 130, + 456, + 482, + 555 + ], + "type": "text", + "content": " of the pixel via volume rendering [47]:" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 226, + 562, + 481, + 592 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 226, + 562, + 481, + 592 + ], + "spans": [ + { + "bbox": [ + 226, + 562, + 481, + 592 + ], + "type": "interline_equation", + "content": "C (\\mathbf {r}) = \\int_ {t _ {n}} ^ {t _ {f}} T (t) \\sigma (\\mathbf {r} (t)) \\mathbf {c} (\\mathbf {r} (t), \\mathbf {d}) d t, \\tag {1}", + "image_path": "85f1cabef74aa0e6fe955d98bdad87a9b99da6be096a9aceba9afea7683b2e1a.jpg" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 235, + 597, + 481, + 626 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 235, + 597, + 481, + 626 + ], + "spans": [ + { + "bbox": [ + 235, + 597, + 481, + 626 + ], + "type": "interline_equation", + "content": "T (t) = \\exp \\left(- \\int_ {t _ {n}} ^ {t} \\sigma (\\mathbf {r} (p)) d p\\right), \\tag {2}", + "image_path": "36342a512d33af59d6881fbdaafde664701af9b583bada2614c03ad277269ae3.jpg" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 130, + 629, + 482, + 666 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 629, + 482, + 666 + ], + "spans": [ + { + "bbox": [ + 130, + 629, + 482, + 666 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 130, + 629, + 482, + 666 + ], + "type": "inline_equation", + "content": "t_n" + }, + { + "bbox": [ + 130, + 629, + 482, + 666 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 130, + 629, + 482, + 666 + ], + "type": "inline_equation", + "content": "t_f" + }, + { + "bbox": [ + 130, + 629, + 482, + 666 + ], + "type": "text", + "content": " indicate the near and far boundary of the ray, respectively. " + }, + { + "bbox": [ + 130, + 629, + 482, + 666 + ], + "type": "inline_equation", + "content": "T(t)" + }, + { + "bbox": [ + 130, + 629, + 482, + 666 + ], + "type": "text", + "content": " refers to an accumulated transmittance, which quantifies the probability of the ray traveling from " + }, + { + "bbox": [ + 130, + 629, + 482, + 666 + ], + "type": "inline_equation", + "content": "t_n" + }, + { + "bbox": [ + 130, + 629, + 482, + 666 + ], + "type": "text", + "content": " to " + }, + { + "bbox": [ + 130, + 629, + 482, + 666 + ], + "type": "inline_equation", + "content": "t" + }, + { + "bbox": [ + 130, + 629, + 482, + 666 + ], + "type": "text", + "content": " without encountering other particles." + } + ] + } + ], + "index": 9 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 337, + 91, + 447, + 100 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 337, + 91, + 447, + 100 + ], + "spans": [ + { + "bbox": [ + 337, + 91, + 447, + 100 + ], + "type": "text", + "content": "Color Correction for NeRF" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 474, + 92, + 480, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 474, + 92, + 480, + 100 + ], + "spans": [ + { + "bbox": [ + 474, + 92, + 480, + 100 + ], + "type": "text", + "content": "5" + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 4 + }, + { + "para_blocks": [ + { + "bbox": [ + 130, + 116, + 482, + 177 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 116, + 482, + 177 + ], + "spans": [ + { + "bbox": [ + 130, + 116, + 482, + 177 + ], + "type": "text", + "content": "We build our model on the NeRF framework, which is widely used for modeling scenes captured under static camera settings and lighting conditions. To eliminate the influence of inherent color variances, we propose to decompose the appearance into two components: view-independent diffuse color " + }, + { + "bbox": [ + 130, + 116, + 482, + 177 + ], + "type": "inline_equation", + "content": "\\mathbf{c}_d" + }, + { + "bbox": [ + 130, + 116, + 482, + 177 + ], + "type": "text", + "content": " and view-dependent color " + }, + { + "bbox": [ + 130, + 116, + 482, + 177 + ], + "type": "inline_equation", + "content": "\\mathbf{c}_s" + }, + { + "bbox": [ + 130, + 116, + 482, + 177 + ], + "type": "text", + "content": ". The final color is obtained by summing the two terms:" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 261, + 182, + 481, + 194 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 261, + 182, + 481, + 194 + ], + "spans": [ + { + "bbox": [ + 261, + 182, + 481, + 194 + ], + "type": "interline_equation", + "content": "\\mathbf {c} = \\mathbf {c} _ {d} (\\mathbf {x}) + \\mathbf {c} _ {s} (\\mathbf {x}, \\mathbf {d}). \\tag {3}", + "image_path": "eca42e909dcc2f8025acbb9b28745117101b404b7bc9661a744cbd53dc1fc25f.jpg" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 131, + 198, + 482, + 236 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 131, + 198, + 482, + 236 + ], + "spans": [ + { + "bbox": [ + 131, + 198, + 482, + 236 + ], + "type": "text", + "content": "By explicitly considering these two components, we can better model the appearance changes arising from surface properties and these caused by varying lighting conditions and camera settings." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 131, + 247, + 289, + 259 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 131, + 247, + 289, + 259 + ], + "spans": [ + { + "bbox": [ + 131, + 247, + 289, + 259 + ], + "type": "text", + "content": "3.2 Color Correction Module" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 130, + 263, + 482, + 300 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 263, + 482, + 300 + ], + "spans": [ + { + "bbox": [ + 130, + 263, + 482, + 300 + ], + "type": "text", + "content": "Digital cameras perform a number of processing steps to transform raw images to standard RGB images, including demosaicing, white balance, color correction and etc. A generic model [8] of this pipeline is formulated as:" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 249, + 305, + 481, + 318 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 249, + 305, + 481, + 318 + ], + "spans": [ + { + "bbox": [ + 249, + 305, + 481, + 318 + ], + "type": "interline_equation", + "content": "\\mathbf {I} _ {s r g b} = f \\left(\\alpha \\mathbf {A} _ {x y z} \\mathbf {D} _ {w b} \\mathbf {I} _ {r a w}\\right), \\tag {4}", + "image_path": "48085b02f89ec22cb5ffff716a17a69eef37cab6b6ce53801d9b7424877bd3cc.jpg" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 130, + 322, + 482, + 393 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 322, + 482, + 393 + ], + "spans": [ + { + "bbox": [ + 130, + 322, + 482, + 393 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 130, + 322, + 482, + 393 + ], + "type": "inline_equation", + "content": "\\alpha" + }, + { + "bbox": [ + 130, + 322, + 482, + 393 + ], + "type": "text", + "content": " is an exposure compensation common gain, " + }, + { + "bbox": [ + 130, + 322, + 482, + 393 + ], + "type": "inline_equation", + "content": "\\mathbf{D}_{wb}" + }, + { + "bbox": [ + 130, + 322, + 482, + 393 + ], + "type": "text", + "content": " represents the " + }, + { + "bbox": [ + 130, + 322, + 482, + 393 + ], + "type": "inline_equation", + "content": "3\\times 3" + }, + { + "bbox": [ + 130, + 322, + 482, + 393 + ], + "type": "text", + "content": " diagonal matrix for white balance, " + }, + { + "bbox": [ + 130, + 322, + 482, + 393 + ], + "type": "inline_equation", + "content": "\\mathbf{A}_{xyz}" + }, + { + "bbox": [ + 130, + 322, + 482, + 393 + ], + "type": "text", + "content": " is a " + }, + { + "bbox": [ + 130, + 322, + 482, + 393 + ], + "type": "inline_equation", + "content": "3\\times 3" + }, + { + "bbox": [ + 130, + 322, + 482, + 393 + ], + "type": "text", + "content": " linear transform of color space, and " + }, + { + "bbox": [ + 130, + 322, + 482, + 393 + ], + "type": "inline_equation", + "content": "f(\\cdot)" + }, + { + "bbox": [ + 130, + 322, + 482, + 393 + ], + "type": "text", + "content": " is a nonlinear function compounding color enhancement, tonemanipulation, and a final sRGB gamma encoding, etc. It's worth noting that the differences in white balance and lighting conditions become non-linear in the sRGB color space after undergoing these mappings." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 130, + 394, + 482, + 453 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 394, + 482, + 453 + ], + "spans": [ + { + "bbox": [ + 130, + 394, + 482, + 453 + ], + "type": "text", + "content": "In our framework, we expect to recover the scene's radiance field with unified color when the input sRGB images exhibit color variances. Inspired by [2], our method introduces a polynomial color-mapping function to model the color deviation between the scene color space " + }, + { + "bbox": [ + 130, + 394, + 482, + 453 + ], + "type": "inline_equation", + "content": "\\mathbf{c}" + }, + { + "bbox": [ + 130, + 394, + 482, + 453 + ], + "type": "text", + "content": " and each camera color space " + }, + { + "bbox": [ + 130, + 394, + 482, + 453 + ], + "type": "inline_equation", + "content": "\\mathbf{c}^{cam}" + }, + { + "bbox": [ + 130, + 394, + 482, + 453 + ], + "type": "text", + "content": " captured by different cameras, which can be written as:" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 273, + 458, + 481, + 471 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 273, + 458, + 481, + 471 + ], + "spans": [ + { + "bbox": [ + 273, + 458, + 481, + 471 + ], + "type": "interline_equation", + "content": "\\mathbf {c} ^ {c a m} = \\mathbf {M} \\varPhi (\\mathbf {c}), \\tag {5}", + "image_path": "8dc2da701aba3e2ed365ce65550d7f792a89777826df614912cd22f8b608b577.jpg" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 130, + 475, + 482, + 583 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 475, + 482, + 583 + ], + "spans": [ + { + "bbox": [ + 130, + 475, + 482, + 583 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 130, + 475, + 482, + 583 + ], + "type": "inline_equation", + "content": "\\varPhi" + }, + { + "bbox": [ + 130, + 475, + 482, + 583 + ], + "type": "text", + "content": " is the kernel mapping function that transforms the input color into the nonlinear space, specifically, " + }, + { + "bbox": [ + 130, + 475, + 482, + 583 + ], + "type": "inline_equation", + "content": "\\varPhi([R,G,B]^T) \\to [R,G,B,RG,RB,GB,1]^T" + }, + { + "bbox": [ + 130, + 475, + 482, + 583 + ], + "type": "text", + "content": ". The matrix " + }, + { + "bbox": [ + 130, + 475, + 482, + 583 + ], + "type": "inline_equation", + "content": "\\mathbf{M}" + }, + { + "bbox": [ + 130, + 475, + 482, + 583 + ], + "type": "text", + "content": " is a color translation matrix of shape " + }, + { + "bbox": [ + 130, + 475, + 482, + 583 + ], + "type": "inline_equation", + "content": "3 \\times 7" + }, + { + "bbox": [ + 130, + 475, + 482, + 583 + ], + "type": "text", + "content": ". Notably, we utilize an Embedding network " + }, + { + "bbox": [ + 130, + 475, + 482, + 583 + ], + "type": "inline_equation", + "content": "E" + }, + { + "bbox": [ + 130, + 475, + 482, + 583 + ], + "type": "text", + "content": " to predict a matrix " + }, + { + "bbox": [ + 130, + 475, + 482, + 583 + ], + "type": "inline_equation", + "content": "\\mathbf{M}" + }, + { + "bbox": [ + 130, + 475, + 482, + 583 + ], + "type": "text", + "content": " for each camera and lighting condition, and based on the physical imaging process, the matrix is shared by the diffuse and view-dependent components. The pixel colors of sampled points in the scene color space are first mapped to camera color space, and then accumulated through volume rendering. Combining Eq. (1), Eq. (3) and Eq. (5), the rendering equation can be written as:" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 178, + 587, + 481, + 614 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 178, + 587, + 481, + 614 + ], + "spans": [ + { + "bbox": [ + 178, + 587, + 481, + 614 + ], + "type": "interline_equation", + "content": "C ^ {c a m} = \\int_ {t _ {n}} ^ {t _ {f}} T (t) \\sigma (\\mathbf {r} (t)) \\left(\\mathbf {M} \\varPhi \\left(\\mathbf {c} _ {d} (\\mathbf {r} (t))\\right) + \\mathbf {M} \\varPhi \\left(\\mathbf {c} _ {s} (\\mathbf {r} (t), \\mathbf {d})\\right)\\right) d t \\tag {6}", + "image_path": "473dd2c69a4fc593a05c40f42e7872f9613db9e8a0026e5bc47539b5d538e51e.jpg" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 130, + 617, + 482, + 665 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 617, + 482, + 665 + ], + "spans": [ + { + "bbox": [ + 130, + 617, + 482, + 665 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 130, + 617, + 482, + 665 + ], + "type": "inline_equation", + "content": "T(t)" + }, + { + "bbox": [ + 130, + 617, + 482, + 665 + ], + "type": "text", + "content": " is defined in Eq. (2). By integrating the color correction module in the rendering pipeline, we can reconstruct the radiance field of the scene while accounting for color variations caused by different cameras and lighting conditions." + } + ] + } + ], + "index": 13 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "spans": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "type": "text", + "content": "6" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 166, + 91, + 266, + 102 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 166, + 91, + 266, + 102 + ], + "spans": [ + { + "bbox": [ + 166, + 91, + 266, + 102 + ], + "type": "text", + "content": "Q. Zhang, Y. Feng et al." + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 5 + }, + { + "para_blocks": [ + { + "bbox": [ + 132, + 116, + 337, + 127 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 116, + 337, + 127 + ], + "spans": [ + { + "bbox": [ + 132, + 116, + 337, + 127 + ], + "type": "text", + "content": "3.3 View-consistent Color Distribution" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 130, + 147, + 482, + 219 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 147, + 482, + 219 + ], + "spans": [ + { + "bbox": [ + 130, + 147, + 482, + 219 + ], + "type": "text", + "content": "Considering the color correction matrix is designed to be camera-wise, relying solely on pixel-level color loss hinders the accuracy and efficiency of optimization. Therefore we employ a cross-camera color distribution constraint. Given that digital cameras capture images with three color channels, each with a discrete range of " + }, + { + "bbox": [ + 130, + 147, + 482, + 219 + ], + "type": "inline_equation", + "content": "K" + }, + { + "bbox": [ + 130, + 147, + 482, + 219 + ], + "type": "text", + "content": " intensity values, we analyze the intensity distribution in each channel, which is formulated as an intensity histogram to constrain the color consistency." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 130, + 220, + 483, + 316 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 220, + 483, + 316 + ], + "spans": [ + { + "bbox": [ + 130, + 220, + 483, + 316 + ], + "type": "text", + "content": "Drawing inspiration from the cross-camera color translation task, we transform the RGB color space " + }, + { + "bbox": [ + 130, + 220, + 483, + 316 + ], + "type": "inline_equation", + "content": "C^{cam}" + }, + { + "bbox": [ + 130, + 220, + 483, + 316 + ], + "type": "text", + "content": " into YUV color space " + }, + { + "bbox": [ + 130, + 220, + 483, + 316 + ], + "type": "inline_equation", + "content": "Y^{cam} \\in [-1,1]^3" + }, + { + "bbox": [ + 130, + 220, + 483, + 316 + ], + "type": "text", + "content": ". The YUV color space comprises one luma component (Y) and two chrominance components, U (blue projection) and V (red projection). We analyze each color channel independently and obtain the intensity histogram by counting the number of pixels of each intensity value. To obtain the intensity histogram as a smooth and differentiable function, we formulate one histogram distribution " + }, + { + "bbox": [ + 130, + 220, + 483, + 316 + ], + "type": "inline_equation", + "content": "\\mathbf{h}(Y^{cam})" + }, + { + "bbox": [ + 130, + 220, + 483, + 316 + ], + "type": "text", + "content": " of an image color channel as" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 182, + 330, + 481, + 365 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 182, + 330, + 481, + 365 + ], + "spans": [ + { + "bbox": [ + 182, + 330, + 481, + 365 + ], + "type": "interline_equation", + "content": "\\mathbf {h} = \\left\\{\\frac {1}{N} \\sum_ {x \\in Y} \\Theta \\left(\\frac {I - \\mu_ {k} + L / 2}{W}\\right) - \\Theta \\left(\\frac {I - \\mu_ {k} - L / 2}{W}\\right) \\right\\} _ {k = 0} ^ {K - 1}, \\tag {7}", + "image_path": "e67150f58b5bb636ff05aa2d9dc6b507cc7ba358769f8f92b870ac47b1bd40ad.jpg" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 131, + 376, + 482, + 437 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 131, + 376, + 482, + 437 + ], + "spans": [ + { + "bbox": [ + 131, + 376, + 482, + 437 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 131, + 376, + 482, + 437 + ], + "type": "inline_equation", + "content": "K" + }, + { + "bbox": [ + 131, + 376, + 482, + 437 + ], + "type": "text", + "content": " is the number of bins set in the histogram, and " + }, + { + "bbox": [ + 131, + 376, + 482, + 437 + ], + "type": "inline_equation", + "content": "L" + }, + { + "bbox": [ + 131, + 376, + 482, + 437 + ], + "type": "text", + "content": " indicates the length of each bin, which also equals " + }, + { + "bbox": [ + 131, + 376, + 482, + 437 + ], + "type": "inline_equation", + "content": "\\frac{2}{K}" + }, + { + "bbox": [ + 131, + 376, + 482, + 437 + ], + "type": "text", + "content": ". Given the " + }, + { + "bbox": [ + 131, + 376, + 482, + 437 + ], + "type": "inline_equation", + "content": "k" + }, + { + "bbox": [ + 131, + 376, + 482, + 437 + ], + "type": "text", + "content": "-th bin, its center and bandwidth refer to " + }, + { + "bbox": [ + 131, + 376, + 482, + 437 + ], + "type": "inline_equation", + "content": "\\mu_k = -1 + L(k + \\frac{1}{2})" + }, + { + "bbox": [ + 131, + 376, + 482, + 437 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 131, + 376, + 482, + 437 + ], + "type": "inline_equation", + "content": "W = L / 2.5" + }, + { + "bbox": [ + 131, + 376, + 482, + 437 + ], + "type": "text", + "content": ", respectively. " + }, + { + "bbox": [ + 131, + 376, + 482, + 437 + ], + "type": "inline_equation", + "content": "\\Theta(x)" + }, + { + "bbox": [ + 131, + 376, + 482, + 437 + ], + "type": "text", + "content": " is a Sigmoid function " + }, + { + "bbox": [ + 131, + 376, + 482, + 437 + ], + "type": "inline_equation", + "content": "(1 + \\exp(-x))^{-1}" + }, + { + "bbox": [ + 131, + 376, + 482, + 437 + ], + "type": "text", + "content": " defining a differentiable approximation of the histogram distribution." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 131, + 439, + 482, + 499 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 131, + 439, + 482, + 499 + ], + "spans": [ + { + "bbox": [ + 131, + 439, + 482, + 499 + ], + "type": "text", + "content": "Histogram Distance. To incorporate color distribution constraints into our framework and facilitate easy convergence, we propose using the Euclidean distance between the cumulative histograms to measure the distance between two image histograms. Let " + }, + { + "bbox": [ + 131, + 439, + 482, + 499 + ], + "type": "inline_equation", + "content": "\\mathbf{h}_s" + }, + { + "bbox": [ + 131, + 439, + 482, + 499 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 131, + 439, + 482, + 499 + ], + "type": "inline_equation", + "content": "\\mathbf{h}_r" + }, + { + "bbox": [ + 131, + 439, + 482, + 499 + ], + "type": "text", + "content": " represent the histograms of the rendered and the reference images, respectively. The distance between histograms is defined:" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 217, + 513, + 481, + 544 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 217, + 513, + 481, + 544 + ], + "spans": [ + { + "bbox": [ + 217, + 513, + 481, + 544 + ], + "type": "interline_equation", + "content": "\\mathcal {D} _ {h} \\left(\\mathbf {h} _ {s}, \\mathbf {h} _ {r}\\right) = \\sum_ {i = 1} ^ {K} \\left(\\mathrm {C D F} _ {i} \\left(\\mathbf {h} _ {s}\\right), \\mathrm {C D F} _ {i} \\left(\\mathbf {h} _ {r}\\right)\\right), \\tag {8}", + "image_path": "f7017df7070557e39c034b206e9f46e049fcbea48b70d7dc3f3e83510d383555.jpg" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 131, + 557, + 482, + 666 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 131, + 557, + 482, + 666 + ], + "spans": [ + { + "bbox": [ + 131, + 557, + 482, + 666 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 131, + 557, + 482, + 666 + ], + "type": "inline_equation", + "content": "\\mathrm{CDF}_i(\\mathbf{h})" + }, + { + "bbox": [ + 131, + 557, + 482, + 666 + ], + "type": "text", + "content": " denotes the " + }, + { + "bbox": [ + 131, + 557, + 482, + 666 + ], + "type": "inline_equation", + "content": "i" + }, + { + "bbox": [ + 131, + 557, + 482, + 666 + ], + "type": "text", + "content": "-th element of the cumulative density function of histogram. The histogram distance " + }, + { + "bbox": [ + 131, + 557, + 482, + 666 + ], + "type": "inline_equation", + "content": "\\mathcal{D}_h" + }, + { + "bbox": [ + 131, + 557, + 482, + 666 + ], + "type": "text", + "content": " measures the global similarity of color distribution between two images. Unlike NeRF, which focuses on individual pixel values, this approach takes into account the entire color distribution of the images. By considering the cumulative histograms, we can effectively compare and align the color distribution of two different images. This property is particularly useful for color translation tasks with our color correction module. We can achieve consistent color appearance across different images by adapting the color distribution towards the reference image." + } + ] + } + ], + "index": 9 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 336, + 91, + 447, + 101 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 336, + 91, + 447, + 101 + ], + "spans": [ + { + "bbox": [ + 336, + 91, + 447, + 101 + ], + "type": "text", + "content": "Color Correction for NeRF" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 474, + 91, + 481, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 474, + 91, + 481, + 100 + ], + "spans": [ + { + "bbox": [ + 474, + 91, + 481, + 100 + ], + "type": "text", + "content": "7" + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 6 + }, + { + "para_blocks": [ + { + "bbox": [ + 132, + 116, + 228, + 128 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 116, + 228, + 128 + ], + "spans": [ + { + "bbox": [ + 132, + 116, + 228, + 128 + ], + "type": "text", + "content": "3.4 Optimization" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 130, + 134, + 479, + 169 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 134, + 479, + 169 + ], + "spans": [ + { + "bbox": [ + 130, + 134, + 479, + 169 + ], + "type": "text", + "content": "To regularize the reconstructed scene color and reduce ambiguity, we design a series of losses on both pixel and image levels, including color reconstruction loss, specular loss, histogram loss, and gray-world loss." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 130, + 170, + 480, + 242 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 170, + 480, + 242 + ], + "spans": [ + { + "bbox": [ + 130, + 170, + 480, + 242 + ], + "type": "text", + "content": "Color Reconstruction Loss. To jointly optimize the scene's NeRF and the camera-dependent color correction matrix " + }, + { + "bbox": [ + 130, + 170, + 480, + 242 + ], + "type": "inline_equation", + "content": "\\{\\mathbf{M}_i\\}_{i=1}^N" + }, + { + "bbox": [ + 130, + 170, + 480, + 242 + ], + "type": "text", + "content": ", color reconstruction loss is proposed to minimize the mean squared error (MSE) of sampled pixels between the ground truth color " + }, + { + "bbox": [ + 130, + 170, + 480, + 242 + ], + "type": "inline_equation", + "content": "C^{cam}" + }, + { + "bbox": [ + 130, + 170, + 480, + 242 + ], + "type": "text", + "content": " and the rendered color " + }, + { + "bbox": [ + 130, + 170, + 480, + 242 + ], + "type": "inline_equation", + "content": "\\hat{C}^{cam}" + }, + { + "bbox": [ + 130, + 170, + 480, + 242 + ], + "type": "text", + "content": " which is mapped from scene color by corresponding matrix (see Eq. (5) and Eq. (6)). The color reconstruction loss is formulated as:" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 240, + 247, + 481, + 277 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 240, + 247, + 481, + 277 + ], + "spans": [ + { + "bbox": [ + 240, + 247, + 481, + 277 + ], + "type": "interline_equation", + "content": "\\mathcal {L} _ {c} = \\sum_ {\\mathbf {r} \\in \\mathcal {R} (\\mathbf {P})} \\left\\| \\widehat {C} ^ {c a m} - C ^ {c a m} \\right\\| _ {2} ^ {2}. \\tag {9}", + "image_path": "1ce866149c08344b07cb47d29b6d2fb8bfedbc1ccbe6c623a2f5c8a4628e2090.jpg" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 130, + 283, + 482, + 390 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 283, + 482, + 390 + ], + "spans": [ + { + "bbox": [ + 130, + 283, + 482, + 390 + ], + "type": "text", + "content": "Specular Loss. According to Eq. (3), we decompose the scene's color into diffuse color " + }, + { + "bbox": [ + 130, + 283, + 482, + 390 + ], + "type": "inline_equation", + "content": "\\mathbf{c}_d" + }, + { + "bbox": [ + 130, + 283, + 482, + 390 + ], + "type": "text", + "content": " and view-dependent color " + }, + { + "bbox": [ + 130, + 283, + 482, + 390 + ], + "type": "inline_equation", + "content": "\\mathbf{c}_s" + }, + { + "bbox": [ + 130, + 283, + 482, + 390 + ], + "type": "text", + "content": ". This view-dependent color accounts for the color variances due to the inherent material properties and geometries. Both components share a common color correction matrix " + }, + { + "bbox": [ + 130, + 283, + 482, + 390 + ], + "type": "inline_equation", + "content": "\\mathbf{M}" + }, + { + "bbox": [ + 130, + 283, + 482, + 390 + ], + "type": "text", + "content": ". Ideally, we suppose the cross-image differences are mostly caused by varying light conditions and camera-dependent color translation. Thus, we apply a regularization on " + }, + { + "bbox": [ + 130, + 283, + 482, + 390 + ], + "type": "inline_equation", + "content": "\\mathbf{c}_s" + }, + { + "bbox": [ + 130, + 283, + 482, + 390 + ], + "type": "text", + "content": ", which benefits color consistency on the diffuse partition and estimation of color correction matrix, as shown in ablation studies. The L2 regularization on the view-dependent color is formulated as:" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 266, + 397, + 481, + 427 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 266, + 397, + 481, + 427 + ], + "spans": [ + { + "bbox": [ + 266, + 397, + 481, + 427 + ], + "type": "interline_equation", + "content": "\\mathcal {L} _ {s} = \\sum_ {\\mathbf {r} \\in \\mathcal {R} (\\mathbf {P})} \\left\\| \\widehat {C} _ {s} \\right\\| _ {2} ^ {2} \\tag {10}", + "image_path": "6079842b59f0fe5334ad3e12834b2bcb146a2e1d4489c8feed1e3d519b2aa20e.jpg" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 131, + 432, + 482, + 515 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 131, + 432, + 482, + 515 + ], + "spans": [ + { + "bbox": [ + 131, + 432, + 482, + 515 + ], + "type": "text", + "content": "Gray-world Loss. Due to the correlation between scene color space and color correction matrix, the recovered scene color " + }, + { + "bbox": [ + 131, + 432, + 482, + 515 + ], + "type": "inline_equation", + "content": "\\mathbf{c}" + }, + { + "bbox": [ + 131, + 432, + 482, + 515 + ], + "type": "text", + "content": " is under-constrained with only " + }, + { + "bbox": [ + 131, + 432, + 482, + 515 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_g" + }, + { + "bbox": [ + 131, + 432, + 482, + 515 + ], + "type": "text", + "content": ". To address this, white balance correction methods often refer to some prior knowledge of the spectral properties of the camera sensors. One commonly used heuristic prior is the \"gray-world\" assumption: the color in the world tends to be natural gray on average [11]. We also define an L2 gray-world regularization for the scene's radiance field:" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 246, + 521, + 481, + 550 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 246, + 521, + 481, + 550 + ], + "spans": [ + { + "bbox": [ + 246, + 521, + 481, + 550 + ], + "type": "interline_equation", + "content": "\\mathcal {L} _ {g} = \\frac {1}{N _ {\\mathbf {r}}} \\sum_ {\\mathbf {r} \\in \\mathcal {R} (\\mathbf {P})} \\left\\| \\widehat {C} - C _ {0} \\right\\| _ {2} ^ {2}, \\tag {11}", + "image_path": "3e53071ae71838ab161c0d9dd91644cb8646bb43cbac9dd6cc12d6486d1b8d13.jpg" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 130, + 555, + 480, + 578 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 555, + 480, + 578 + ], + "spans": [ + { + "bbox": [ + 130, + 555, + 480, + 578 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 130, + 555, + 480, + 578 + ], + "type": "inline_equation", + "content": "C_0" + }, + { + "bbox": [ + 130, + 555, + 480, + 578 + ], + "type": "text", + "content": " is a hyperparameter which is generally set as the midway of the color values such as 0.5." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 130, + 580, + 482, + 640 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 580, + 482, + 640 + ], + "spans": [ + { + "bbox": [ + 130, + 580, + 482, + 640 + ], + "type": "text", + "content": "Histogram Loss. To further constrain the color consistency of reconstructed radiance fields in different views, besides rendering image " + }, + { + "bbox": [ + 130, + 580, + 482, + 640 + ], + "type": "inline_equation", + "content": "\\widehat{C}^{cam}" + }, + { + "bbox": [ + 130, + 580, + 482, + 640 + ], + "type": "text", + "content": " from the ground truth view, we render an additional image from random view and obtain its histogram " + }, + { + "bbox": [ + 130, + 580, + 482, + 640 + ], + "type": "inline_equation", + "content": "\\widehat{\\mathbf{h}}_r^{cam}" + }, + { + "bbox": [ + 130, + 580, + 482, + 640 + ], + "type": "text", + "content": " during training. Based on " + }, + { + "bbox": [ + 130, + 580, + 482, + 640 + ], + "type": "inline_equation", + "content": "\\widehat{\\mathbf{h}}_r^{cam}" + }, + { + "bbox": [ + 130, + 580, + 482, + 640 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 130, + 580, + 482, + 640 + ], + "type": "inline_equation", + "content": "\\widehat{\\mathbf{h}}^{cam}" + }, + { + "bbox": [ + 130, + 580, + 482, + 640 + ], + "type": "text", + "content": " of " + }, + { + "bbox": [ + 130, + 580, + 482, + 640 + ], + "type": "inline_equation", + "content": "\\widehat{C}^{cam}" + }, + { + "bbox": [ + 130, + 580, + 482, + 640 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 130, + 580, + 482, + 640 + ], + "type": "inline_equation", + "content": "\\mathbf{h}^{cam}" + }, + { + "bbox": [ + 130, + 580, + 482, + 640 + ], + "type": "text", + "content": " of ground truth color " + }, + { + "bbox": [ + 130, + 580, + 482, + 640 + ], + "type": "inline_equation", + "content": "C^{cam}" + }, + { + "bbox": [ + 130, + 580, + 482, + 640 + ], + "type": "text", + "content": ", we design a cross histogram loss based on Eq. (8)," + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 174, + 647, + 481, + 669 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 174, + 647, + 481, + 669 + ], + "spans": [ + { + "bbox": [ + 174, + 647, + 481, + 669 + ], + "type": "interline_equation", + "content": "\\mathcal {L} _ {h} = \\frac {1}{3} \\left(\\mathcal {D} _ {h} \\left(\\widehat {\\mathbf {h}} ^ {c a m}, \\mathbf {h} ^ {c a m}\\right) + \\mathcal {D} _ {h} \\left(\\widehat {\\mathbf {h}} _ {r} ^ {c a m}, \\mathbf {h} ^ {c a m}\\right) + \\mathcal {D} _ {h} \\left(\\widehat {\\mathbf {h}} ^ {c a m}, \\widehat {\\mathbf {h}} _ {r} ^ {c a m}\\right)\\right). \\tag {12}", + "image_path": "c11c02c05f924e7c54598bd3395c6475ab310ad5c47cbe8d9a4a31620b36e74e.jpg" + } + ] + } + ], + "index": 12 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "spans": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "type": "text", + "content": "8" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 166, + 91, + 266, + 102 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 166, + 91, + 266, + 102 + ], + "spans": [ + { + "bbox": [ + 166, + 91, + 266, + 102 + ], + "type": "text", + "content": "Q. Zhang, Y. Feng et al." + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 7 + }, + { + "para_blocks": [ + { + "bbox": [ + 130, + 116, + 479, + 176 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 116, + 479, + 176 + ], + "spans": [ + { + "bbox": [ + 130, + 116, + 479, + 176 + ], + "type": "text", + "content": "Considering that gray-world loss " + }, + { + "bbox": [ + 130, + 116, + 479, + 176 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_g" + }, + { + "bbox": [ + 130, + 116, + 479, + 176 + ], + "type": "text", + "content": " and histogram loss " + }, + { + "bbox": [ + 130, + 116, + 479, + 176 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_h" + }, + { + "bbox": [ + 130, + 116, + 479, + 176 + ], + "type": "text", + "content": " are defined on color distributions, we sample rays from grids and combine them to form an image instead of rendering the entire image during training for converge efficiency. This is based on the assumption that the color distribution of an image will remain unchanged after the random grid sampling." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 130, + 177, + 479, + 202 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 177, + 479, + 202 + ], + "spans": [ + { + "bbox": [ + 130, + 177, + 479, + 202 + ], + "type": "text", + "content": "Total Loss. The complete loss " + }, + { + "bbox": [ + 130, + 177, + 479, + 202 + ], + "type": "inline_equation", + "content": "\\mathcal{L}" + }, + { + "bbox": [ + 130, + 177, + 479, + 202 + ], + "type": "text", + "content": " is a weighted sum of the above loss functions from Eq. (9) to Eq. (12):" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 238, + 215, + 481, + 228 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 238, + 215, + 481, + 228 + ], + "spans": [ + { + "bbox": [ + 238, + 215, + 481, + 228 + ], + "type": "interline_equation", + "content": "\\mathcal {L} = \\mathcal {L} _ {c} + \\lambda_ {s} \\mathcal {L} _ {s} + \\lambda_ {h} \\mathcal {L} _ {h} + \\lambda_ {g} \\mathcal {L} _ {g}, \\tag {13}", + "image_path": "0cf01b8f4416f85b01838844ce3a6c5863407098295c109263deee6c9b4614f3.jpg" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 131, + 239, + 470, + 251 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 131, + 239, + 470, + 251 + ], + "spans": [ + { + "bbox": [ + 131, + 239, + 470, + 251 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 131, + 239, + 470, + 251 + ], + "type": "inline_equation", + "content": "\\lambda_s = 0.1" + }, + { + "bbox": [ + 131, + 239, + 470, + 251 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 131, + 239, + 470, + 251 + ], + "type": "inline_equation", + "content": "\\lambda_h = 0.01" + }, + { + "bbox": [ + 131, + 239, + 470, + 251 + ], + "type": "text", + "content": ", and " + }, + { + "bbox": [ + 131, + 239, + 470, + 251 + ], + "type": "inline_equation", + "content": "\\lambda_g = 0.001" + }, + { + "bbox": [ + 131, + 239, + 470, + 251 + ], + "type": "text", + "content": " are the weights of different losses." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 131, + 277, + 230, + 290 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 131, + 277, + 230, + 290 + ], + "spans": [ + { + "bbox": [ + 131, + 277, + 230, + 290 + ], + "type": "text", + "content": "4 Experiments" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 131, + 309, + 279, + 321 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 131, + 309, + 279, + 321 + ], + "spans": [ + { + "bbox": [ + 131, + 309, + 279, + 321 + ], + "type": "text", + "content": "4.1 Implementation Details" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 130, + 336, + 482, + 468 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 336, + 482, + 468 + ], + "spans": [ + { + "bbox": [ + 130, + 336, + 482, + 468 + ], + "type": "text", + "content": "Our implementation of NeRF follows [67]. Note that our camera correction module could be embedded in different NeRF-based models [4,13,50]. An eight-layer MLP with 256 channels with encoding of position " + }, + { + "bbox": [ + 130, + 336, + 482, + 468 + ], + "type": "inline_equation", + "content": "\\mathbf{x}" + }, + { + "bbox": [ + 130, + 336, + 482, + 468 + ], + "type": "text", + "content": " as input is used to generate " + }, + { + "bbox": [ + 130, + 336, + 482, + 468 + ], + "type": "inline_equation", + "content": "\\sigma" + }, + { + "bbox": [ + 130, + 336, + 482, + 468 + ], + "type": "text", + "content": ", and followed by one additional 256-channel fully-connected layer to extract a feature. Diffuse color " + }, + { + "bbox": [ + 130, + 336, + 482, + 468 + ], + "type": "inline_equation", + "content": "\\mathbf{c}_d" + }, + { + "bbox": [ + 130, + 336, + 482, + 468 + ], + "type": "text", + "content": " is obtained by mapping the feature to RGB channels. The view-dependent color " + }, + { + "bbox": [ + 130, + 336, + 482, + 468 + ], + "type": "inline_equation", + "content": "\\mathbf{c}_s" + }, + { + "bbox": [ + 130, + 336, + 482, + 468 + ], + "type": "text", + "content": " is obtained through an additional 128-channel fully connected layer and a mapping layer, with the aforementioned feature and encoding of view direction " + }, + { + "bbox": [ + 130, + 336, + 482, + 468 + ], + "type": "inline_equation", + "content": "\\mathbf{d}" + }, + { + "bbox": [ + 130, + 336, + 482, + 468 + ], + "type": "text", + "content": " as input. Both " + }, + { + "bbox": [ + 130, + 336, + 482, + 468 + ], + "type": "inline_equation", + "content": "\\mathbf{c}_d" + }, + { + "bbox": [ + 130, + 336, + 482, + 468 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 130, + 336, + 482, + 468 + ], + "type": "inline_equation", + "content": "\\mathbf{c}_s" + }, + { + "bbox": [ + 130, + 336, + 482, + 468 + ], + "type": "text", + "content": " are transformed to nonlinear space by the kernel mapping function " + }, + { + "bbox": [ + 130, + 336, + 482, + 468 + ], + "type": "inline_equation", + "content": "\\varPhi" + }, + { + "bbox": [ + 130, + 336, + 482, + 468 + ], + "type": "text", + "content": " in Eq. (5). The embedding vector is reshaped to color correction matrix " + }, + { + "bbox": [ + 130, + 336, + 482, + 468 + ], + "type": "inline_equation", + "content": "\\mathbf{M}" + }, + { + "bbox": [ + 130, + 336, + 482, + 468 + ], + "type": "text", + "content": " of size " + }, + { + "bbox": [ + 130, + 336, + 482, + 468 + ], + "type": "inline_equation", + "content": "3\\times 7" + }, + { + "bbox": [ + 130, + 336, + 482, + 468 + ], + "type": "text", + "content": ". Models are trained over 300,000 iterations with a batch size of 1024." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 131, + 494, + 321, + 505 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 131, + 494, + 321, + 505 + ], + "spans": [ + { + "bbox": [ + 131, + 494, + 321, + 505 + ], + "type": "text", + "content": "4.2 Evaluation Dataset and Metrics" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 130, + 520, + 482, + 605 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 520, + 482, + 605 + ], + "spans": [ + { + "bbox": [ + 130, + 520, + 482, + 605 + ], + "type": "text", + "content": "Dataset. We evaluate the proposed method on our collected dataset. The dataset consists of 6 real scenes captured by three digital cameras with different camera settings (e.g. white balance, ISO, exposures, etc.) and light conditions (e.g. non-flashing, and flashing with different lights). Images are collected at 90 different poses for each scene. We select 60 views with different poses as the training dataset. The remaining images are used as test images, corresponding to the same setting with training data. The resolution is " + }, + { + "bbox": [ + 130, + 520, + 482, + 605 + ], + "type": "inline_equation", + "content": "1000 \\times 667" + }, + { + "bbox": [ + 130, + 520, + 482, + 605 + ], + "type": "text", + "content": " pixels." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 130, + 605, + 482, + 666 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 605, + 482, + 666 + ], + "spans": [ + { + "bbox": [ + 130, + 605, + 482, + 666 + ], + "type": "text", + "content": "Metrics. We report quantitative performance using PSNR (higher is better) and SSIM (higher is better) metrics, as well as the state-of-the-art LPIPS (lower is better) perceptual metric [70]. To better evaluate the color consistency and correction, we propose a " + }, + { + "bbox": [ + 130, + 605, + 482, + 666 + ], + "type": "inline_equation", + "content": "W_{2}" + }, + { + "bbox": [ + 130, + 605, + 482, + 666 + ], + "type": "text", + "content": " metric based on Wasserstein distance, which is a distance function defined between probability distributions on the entire color" + } + ] + } + ], + "index": 11 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 336, + 91, + 447, + 101 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 336, + 91, + 447, + 101 + ], + "spans": [ + { + "bbox": [ + 336, + 91, + 447, + 101 + ], + "type": "text", + "content": "Color Correction for NeRF" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 474, + 91, + 481, + 101 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 474, + 91, + 481, + 101 + ], + "spans": [ + { + "bbox": [ + 474, + 91, + 481, + 101 + ], + "type": "text", + "content": "9" + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 8 + }, + { + "para_blocks": [ + { + "type": "table", + "bbox": [ + 134, + 114, + 479, + 190 + ], + "blocks": [ + { + "bbox": [ + 134, + 114, + 479, + 190 + ], + "lines": [ + { + "bbox": [ + 134, + 114, + 479, + 190 + ], + "spans": [ + { + "bbox": [ + 134, + 114, + 479, + 190 + ], + "type": "table", + "html": "
ScenePSNR ↑SSIM ↑
NeRFNeRF-WHa-NeRFOursNeRFNeRF-WHa-NeRFOurs
Hamster15.62928.38128.28129.5710.9310.9510.9580.959
Pumpkin17.16230.18930.09631.3590.8780.9430.9420.956
Letters16.78028.04328.38730.4520.7830.9200.9220.955
MailBox18.85325.77225.90328.1280.9000.8950.8960.922
Pipe14.72522.20125.98227.5540.7440.7570.8560.907
Room16.62914.77214.96531.0970.9320.7960.7730.969
Mean16.62924.89325.60229.6940.8610.8770.8910.945
", + "image_path": "485ec5deaca9fdfcce9c44117c822f27ef2178a37be79373eca871e27987b9b4.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "table_body" + } + ], + "index": 2 + }, + { + "type": "table", + "bbox": [ + 135, + 194, + 479, + 269 + ], + "blocks": [ + { + "bbox": [ + 135, + 194, + 479, + 269 + ], + "lines": [ + { + "bbox": [ + 135, + 194, + 479, + 269 + ], + "spans": [ + { + "bbox": [ + 135, + 194, + 479, + 269 + ], + "type": "table", + "html": "
SceneLPIPS ↓\\( W_2 \\) (1e-3) ↓
NeRFNeRF-WHa-NeRFOursNeRFNeRF-WHa-NeRFOurs
Hamster0.5080.2190.2250.19974.42.582.592.31
Pumpkin0.5970.2340.2350.18161.11.271.400.53
Letters0.4490.3210.3240.21676.41.110.950.68
MailBox0.4390.2660.2690.20544.95.744.133.71
Pipe0.6210.4760.2790.17771.41.500.600.42
Room0.4960.6650.6680.10389.513.013.56.16
Mean0.5180.3630.3330.18069.64.203.862.30
", + "image_path": "edd0d0ea80e67541b105e3d3733c545bfb5835fd034acb6a0bc45a25e18829fe.jpg" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "table_body" + } + ], + "index": 3 + }, + { + "bbox": [ + 132, + 269, + 480, + 301 + ], + "lines": [ + { + "bbox": [ + 132, + 269, + 480, + 301 + ], + "spans": [ + { + "bbox": [ + 132, + 269, + 480, + 301 + ], + "type": "text", + "content": "Table 1: Quantitative comparisons with baseline methods on our collected scenes, including PSNR, SSIM, LPIPS, and " + }, + { + "bbox": [ + 132, + 269, + 480, + 301 + ], + "type": "inline_equation", + "content": "W_{2}" + }, + { + "bbox": [ + 132, + 269, + 480, + 301 + ], + "type": "text", + "content": " metrics. The Wasserstein distance " + }, + { + "bbox": [ + 132, + 269, + 480, + 301 + ], + "type": "inline_equation", + "content": "W_{2}" + }, + { + "bbox": [ + 132, + 269, + 480, + 301 + ], + "type": "text", + "content": " measures the difference of entire color distributions." + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "text" + }, + { + "bbox": [ + 130, + 333, + 383, + 345 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 333, + 383, + 345 + ], + "spans": [ + { + "bbox": [ + 130, + 333, + 383, + 345 + ], + "type": "text", + "content": "space. The image-level " + }, + { + "bbox": [ + 130, + 333, + 383, + 345 + ], + "type": "inline_equation", + "content": "W_{2}" + }, + { + "bbox": [ + 130, + 333, + 383, + 345 + ], + "type": "text", + "content": " (lower is better) is defined as:" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 142, + 351, + 302, + 380 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 142, + 351, + 302, + 380 + ], + "spans": [ + { + "bbox": [ + 142, + 351, + 302, + 380 + ], + "type": "interline_equation", + "content": "W _ {2} = \\frac {1}{N} \\sum_ {i} ^ {N} \\mathrm {W a s s e r s t e i n} (\\widehat {C} _ {i} ^ {c a m}, C ^ {t r a i n}),", + "image_path": "4ab71d006716826e43ee47d1bead07cf53af7b14244fc5a15e7439b06b54344b.jpg" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 142, + 383, + 480, + 403 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 142, + 383, + 480, + 403 + ], + "spans": [ + { + "bbox": [ + 142, + 383, + 480, + 403 + ], + "type": "interline_equation", + "content": "\\text {W a s s e r s t e i n} \\left(\\widehat {C} _ {i} ^ {\\text {c a m}}, C ^ {\\text {t r a i n}}\\right) = \\left\\| \\operatorname {m e a n} \\left(\\widehat {C} _ {i} ^ {\\text {c a m}}\\right) - \\operatorname {m e a n} \\left(C ^ {\\text {t r a i n}}\\right) \\right\\| _ {2} ^ {2} + \\tag {14}", + "image_path": "583c0fe553034a77eb60787ce127ad0377c5b765dffda2bc7f17e03ad62814f2.jpg" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 142, + 405, + 453, + 435 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 142, + 405, + 453, + 435 + ], + "spans": [ + { + "bbox": [ + 142, + 405, + 453, + 435 + ], + "type": "interline_equation", + "content": "\\mathrm {t r a c e} \\left(\\mathrm {c o v} (\\widehat {C} _ {i} ^ {c a m}) + \\mathrm {c o v} (C ^ {t r a i n}) - 2 \\sqrt {\\sqrt {\\mathrm {c o v} (\\widehat {C} _ {i} ^ {c a m})} \\mathrm {c o v} (C ^ {t r a i n}) \\sqrt {\\mathrm {c o v} (\\widehat {C} _ {i} ^ {c a m})}}\\right),", + "image_path": "bd488c4cb5c737a8c382c53b2ada66d2279a7f69cec73a77f0c5de97f2a9c4cd.jpg" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 130, + 440, + 480, + 477 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 440, + 480, + 477 + ], + "spans": [ + { + "bbox": [ + 130, + 440, + 480, + 477 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 130, + 440, + 480, + 477 + ], + "type": "inline_equation", + "content": "\\mathrm{mean}(\\cdot)" + }, + { + "bbox": [ + 130, + 440, + 480, + 477 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 130, + 440, + 480, + 477 + ], + "type": "inline_equation", + "content": "\\mathrm{cov}(\\cdot)" + }, + { + "bbox": [ + 130, + 440, + 480, + 477 + ], + "type": "text", + "content": " refer to the mean and covariance of an image. " + }, + { + "bbox": [ + 130, + 440, + 480, + 477 + ], + "type": "inline_equation", + "content": "C^{train}" + }, + { + "bbox": [ + 130, + 440, + 480, + 477 + ], + "type": "text", + "content": " denotes the training image. " + }, + { + "bbox": [ + 130, + 440, + 480, + 477 + ], + "type": "inline_equation", + "content": "\\hat{C}^{cam}" + }, + { + "bbox": [ + 130, + 440, + 480, + 477 + ], + "type": "text", + "content": " indicates the rendered images those translated with matrix of " + }, + { + "bbox": [ + 130, + 440, + 480, + 477 + ], + "type": "inline_equation", + "content": "C^{train}" + }, + { + "bbox": [ + 130, + 440, + 480, + 477 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 132, + 493, + 214, + 503 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 493, + 214, + 503 + ], + "spans": [ + { + "bbox": [ + 132, + 493, + 214, + 503 + ], + "type": "text", + "content": "4.3 Evaluation" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 130, + 510, + 479, + 568 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 510, + 479, + 568 + ], + "spans": [ + { + "bbox": [ + 130, + 510, + 479, + 568 + ], + "type": "text", + "content": "Baselines. We compare our method against the following baseline methods. 1) NeRF [49]: the original NeRF method, also our backbone network. 2) NeRF-W [46]: unofficial implementation where the dimension of latent code is 48. 3) Ha-NeRF [15]: official implementation which can use a reference image to do color translation." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 130, + 570, + 480, + 666 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 570, + 480, + 666 + ], + "spans": [ + { + "bbox": [ + 130, + 570, + 480, + 666 + ], + "type": "text", + "content": "Comparisons. The qualitative results of rendered novel views on our own dataset are shown in Fig. 3. Compared to baseline methods, our method achieves promising results on view rendering quality (see MSE error) and color correction measured by Wasserstein distance between ground truths and rendered images. Specifically, NeRF, which is directly reconstructed from inputs under various conditions, tends to render views with significant artifacts. NeRF-W and Ha-NeRF do ensure color consistency to a certain extent, as shown in Fig. 3. However, these methods can not distinguish the implicit properties (e.g. specular)" + } + ] + } + ], + "index": 12 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "spans": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "type": "text", + "content": "10" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 166, + 91, + 266, + 102 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 166, + 91, + 266, + 102 + ], + "spans": [ + { + "bbox": [ + 166, + 91, + 266, + 102 + ], + "type": "text", + "content": "Q. Zhang, Y. Feng et al." + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 9 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 134, + 114, + 482, + 322 + ], + "blocks": [ + { + "bbox": [ + 134, + 114, + 482, + 322 + ], + "lines": [ + { + "bbox": [ + 134, + 114, + 482, + 322 + ], + "spans": [ + { + "bbox": [ + 134, + 114, + 482, + 322 + ], + "type": "image", + "image_path": "c4804a127f5b0d36915a8fd06ed8cb16fe45b90d4d9b549a37b0978574db69b0.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 132, + 331, + 482, + 419 + ], + "lines": [ + { + "bbox": [ + 132, + 331, + 482, + 419 + ], + "spans": [ + { + "bbox": [ + 132, + 331, + 482, + 419 + ], + "type": "text", + "content": "Fig. 3: Qualitative results of our method compared with baseline methods. We also show the error map and histogram for better visualization. MSE and Wasserstein values are on the top corners of the error map and histogram, respectively. Our method outperforms baselines, providing superior rendering quality and effective color correction. NeRF [49] generates significant artifacts. NeRF-W [46] and Ha-NeRF [15] achieve color correction to a certain extent, but exhibit shortcomings in areas with high illumination (e.g. the leaf of scene Mailbox and 'E' of scene Letters) and generate artifacts (e.g. the window of scene Mailbox and floor of scene Hamster)." + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_caption" + } + ], + "index": 2 + }, + { + "bbox": [ + 130, + 447, + 482, + 532 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 447, + 482, + 532 + ], + "spans": [ + { + "bbox": [ + 130, + 447, + 482, + 532 + ], + "type": "text", + "content": "and external differences (e.g. camera settings). Their performance is limited in highlighting regions on Letter 'E' in scene Letters and the leaf in scene Mailbox, and results in artifacts on the window in scene Mailbox and the floor in scene Hamster. In contrast, our method outperforms state-of-the-art methods, providing superior rendering quality and effective color correction. It demonstrates the effectiveness of our color correction module, which physically defines color differences by a nonlinear color translation matrix." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 130, + 534, + 482, + 666 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 534, + 482, + 666 + ], + "spans": [ + { + "bbox": [ + 130, + 534, + 482, + 666 + ], + "type": "text", + "content": "Tab. 1 presents the quantitative results of our method compared to baseline approaches on our collected dataset, showcasing the superior rendering quality achieved by our method. Notably, NeRF-W and Ha-NeRF perform worse than the original NeRF on the scene Room. We analyze this can be attributed to the fact that the scene \"Room\" presents particular challenges with large reflective areas, which can be difficult for these two methods to fit a reasonable solution, and qualitative results are included in the supplementary material. Tab. 1 summarizes the quantitative comparisons on color consistency measured by Eq. (14). Wasserstein distance measures color consistency by calculating the color distribution distance between the input image and the rendered image translated with the color correction matrix of the input. Compared to the pixel-level PSNR" + } + ] + } + ], + "index": 5 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 337, + 91, + 447, + 100 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 337, + 91, + 447, + 100 + ], + "spans": [ + { + "bbox": [ + 337, + 91, + 447, + 100 + ], + "type": "text", + "content": "Color Correction for NeRF" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 470, + 91, + 479, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 470, + 91, + 479, + 100 + ], + "spans": [ + { + "bbox": [ + 470, + 91, + 479, + 100 + ], + "type": "text", + "content": "11" + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 10 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 132, + 113, + 482, + 252 + ], + "blocks": [ + { + "bbox": [ + 132, + 113, + 482, + 252 + ], + "lines": [ + { + "bbox": [ + 132, + 113, + 482, + 252 + ], + "spans": [ + { + "bbox": [ + 132, + 113, + 482, + 252 + ], + "type": "image", + "image_path": "3f912dbf6e4a7969d3e691498752f6bfc475a1278bf203a59aab85f376f50c96.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 130, + 261, + 482, + 297 + ], + "lines": [ + { + "bbox": [ + 130, + 261, + 482, + 297 + ], + "spans": [ + { + "bbox": [ + 130, + 261, + 482, + 297 + ], + "type": "text", + "content": "Fig. 4: Ablation study of gray-world loss " + }, + { + "bbox": [ + 130, + 261, + 482, + 297 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_g" + }, + { + "bbox": [ + 130, + 261, + 482, + 297 + ], + "type": "text", + "content": ". Removing " + }, + { + "bbox": [ + 130, + 261, + 482, + 297 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_g" + }, + { + "bbox": [ + 130, + 261, + 482, + 297 + ], + "type": "text", + "content": " will be unable to constrain the distribution of the scene color space, resulting in undesirable artifacts or color shifts in the reconstructed scene (e.g. overexposed Hamster and over-saturated Pumpkin)." + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_caption" + } + ], + "index": 2 + }, + { + "bbox": [ + 130, + 318, + 482, + 367 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 318, + 482, + 367 + ], + "spans": [ + { + "bbox": [ + 130, + 318, + 482, + 367 + ], + "type": "text", + "content": "metric, the Wasserstein metric measures the entire color distribution. These comparative results demonstrate the effectiveness of our approach in addressing the limitations of existing NeRF-based models, ultimately leading to enhanced color correction and consistency in various camera conditions." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 132, + 383, + 246, + 394 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 383, + 246, + 394 + ], + "spans": [ + { + "bbox": [ + 132, + 383, + 246, + 394 + ], + "type": "text", + "content": "4.4 Ablation Studies" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 130, + 402, + 482, + 569 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 402, + 482, + 569 + ], + "spans": [ + { + "bbox": [ + 130, + 402, + 482, + 569 + ], + "type": "text", + "content": "We conduct ablative evaluations on our collected dataset to show the effectiveness of our model design. We compare our model with four variants: model without specular loss (Ours w/o " + }, + { + "bbox": [ + 130, + 402, + 482, + 569 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_s" + }, + { + "bbox": [ + 130, + 402, + 482, + 569 + ], + "type": "text", + "content": "), model without gray-world loss (Ours w/o " + }, + { + "bbox": [ + 130, + 402, + 482, + 569 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_g" + }, + { + "bbox": [ + 130, + 402, + 482, + 569 + ], + "type": "text", + "content": "), model without histogram loss (Ours w/o " + }, + { + "bbox": [ + 130, + 402, + 482, + 569 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_h" + }, + { + "bbox": [ + 130, + 402, + 482, + 569 + ], + "type": "text", + "content": "), and model without nonlinear mapping (Ours w/o " + }, + { + "bbox": [ + 130, + 402, + 482, + 569 + ], + "type": "inline_equation", + "content": "\\varPhi" + }, + { + "bbox": [ + 130, + 402, + 482, + 569 + ], + "type": "text", + "content": "). The quantitative comparisons presented in Fig. 5 indicate that each component is crucial for generating physically plausible color-corrected renderings. Additionally, Fig. 4 displays the qualitative results of the ablation study on view synthesis with the unified color NeRF. This further highlights the importance of gray-world loss in our model. When the gray-world loss is removed, the color distribution of the scene color space becomes unconstrained, and the increased solution space can be detrimental to convergence. While this may not significantly impact the rendering quality and color consistency of novel views, it could introduce undesirable artifacts or color shifts in the reconstructed scene, such as overexposed regions of Hamster shown in Fig. 4." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 130, + 570, + 483, + 666 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 570, + 483, + 666 + ], + "spans": [ + { + "bbox": [ + 130, + 570, + 483, + 666 + ], + "type": "text", + "content": "In addition, we provide qualitative demonstrations of the results with/without specular loss " + }, + { + "bbox": [ + 130, + 570, + 483, + 666 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_s" + }, + { + "bbox": [ + 130, + 570, + 483, + 666 + ], + "type": "text", + "content": ", histogram loss " + }, + { + "bbox": [ + 130, + 570, + 483, + 666 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_h" + }, + { + "bbox": [ + 130, + 570, + 483, + 666 + ], + "type": "text", + "content": " and nonlinear mapping " + }, + { + "bbox": [ + 130, + 570, + 483, + 666 + ], + "type": "inline_equation", + "content": "\\varPhi" + }, + { + "bbox": [ + 130, + 570, + 483, + 666 + ], + "type": "text", + "content": ". The left bottom half of the image represents the view-dependent color " + }, + { + "bbox": [ + 130, + 570, + 483, + 666 + ], + "type": "inline_equation", + "content": "C_s^{cam}" + }, + { + "bbox": [ + 130, + 570, + 483, + 666 + ], + "type": "text", + "content": ", and the remaining part displays the entire renderings " + }, + { + "bbox": [ + 130, + 570, + 483, + 666 + ], + "type": "inline_equation", + "content": "C^{cam}" + }, + { + "bbox": [ + 130, + 570, + 483, + 666 + ], + "type": "text", + "content": " in camera color space of ground truth. Specifically, removing specular loss will not decompose the view-dependent and view-independent colors of Letters, resulting in a significant break in color consistency (also shown in " + }, + { + "bbox": [ + 130, + 570, + 483, + 666 + ], + "type": "inline_equation", + "content": "W_2" + }, + { + "bbox": [ + 130, + 570, + 483, + 666 + ], + "type": "text", + "content": " metric). Eliminating the histogram loss makes it challenging to represent the entire color distribution when relying solely on pixel-level" + } + ] + } + ], + "index": 7 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "spans": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "type": "text", + "content": "12" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 166, + 91, + 267, + 102 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 166, + 91, + 267, + 102 + ], + "spans": [ + { + "bbox": [ + 166, + 91, + 267, + 102 + ], + "type": "text", + "content": "Q. Zhang, Y. Feng et al." + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 11 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 133, + 114, + 248, + 190 + ], + "blocks": [ + { + "bbox": [ + 133, + 114, + 248, + 190 + ], + "lines": [ + { + "bbox": [ + 133, + 114, + 248, + 190 + ], + "spans": [ + { + "bbox": [ + 133, + 114, + 248, + 190 + ], + "type": "image", + "image_path": "38e0f1468d3da6358ed37a2e4d896e369c89143b5d128ba1c1b06aa249968be9.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 171, + 190, + 210, + 198 + ], + "lines": [ + { + "bbox": [ + 171, + 190, + 210, + 198 + ], + "spans": [ + { + "bbox": [ + 171, + 190, + 210, + 198 + ], + "type": "text", + "content": "Ground Truth" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_caption" + } + ], + "index": 2 + }, + { + "type": "image", + "bbox": [ + 248, + 114, + 364, + 190 + ], + "blocks": [ + { + "bbox": [ + 248, + 114, + 364, + 190 + ], + "lines": [ + { + "bbox": [ + 248, + 114, + 364, + 190 + ], + "spans": [ + { + "bbox": [ + 248, + 114, + 364, + 190 + ], + "type": "image", + "image_path": "040a3170eaf9e2b815c43b105a2dab25deef701704885a4ac978be1e68ed6ddf.jpg" + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 288, + 190, + 324, + 198 + ], + "lines": [ + { + "bbox": [ + 288, + 190, + 324, + 198 + ], + "spans": [ + { + "bbox": [ + 288, + 190, + 324, + 198 + ], + "type": "text", + "content": "Ours w/o " + }, + { + "bbox": [ + 288, + 190, + 324, + 198 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_s" + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "image_caption" + } + ], + "index": 4 + }, + { + "type": "image", + "bbox": [ + 364, + 114, + 480, + 191 + ], + "blocks": [ + { + "bbox": [ + 364, + 114, + 480, + 191 + ], + "lines": [ + { + "bbox": [ + 364, + 114, + 480, + 191 + ], + "spans": [ + { + "bbox": [ + 364, + 114, + 480, + 191 + ], + "type": "image", + "image_path": "f8af0581431f971b575d1fe9eff955a8ff7b31b08f895c94e027407b7e482f6f.jpg" + } + ] + } + ], + "index": 6, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 403, + 190, + 440, + 198 + ], + "lines": [ + { + "bbox": [ + 403, + 190, + 440, + 198 + ], + "spans": [ + { + "bbox": [ + 403, + 190, + 440, + 198 + ], + "type": "text", + "content": "Ours w/o " + }, + { + "bbox": [ + 403, + 190, + 440, + 198 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_h" + } + ] + } + ], + "index": 7, + "angle": 0, + "type": "image_caption" + } + ], + "index": 6 + }, + { + "type": "image", + "bbox": [ + 132, + 198, + 247, + 274 + ], + "blocks": [ + { + "bbox": [ + 132, + 198, + 247, + 274 + ], + "lines": [ + { + "bbox": [ + 132, + 198, + 247, + 274 + ], + "spans": [ + { + "bbox": [ + 132, + 198, + 247, + 274 + ], + "type": "image", + "image_path": "907a351ec30140b91f866a5b6b7ef58984a8e180e3368f9700b6b5b9843fd715.jpg" + } + ] + } + ], + "index": 8, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 155, + 275, + 228, + 283 + ], + "lines": [ + { + "bbox": [ + 155, + 275, + 228, + 283 + ], + "spans": [ + { + "bbox": [ + 155, + 275, + 228, + 283 + ], + "type": "text", + "content": "Quantitative Comparisons" + } + ] + } + ], + "index": 9, + "angle": 0, + "type": "image_caption" + } + ], + "index": 8 + }, + { + "type": "image", + "bbox": [ + 248, + 198, + 364, + 275 + ], + "blocks": [ + { + "bbox": [ + 248, + 198, + 364, + 275 + ], + "lines": [ + { + "bbox": [ + 248, + 198, + 364, + 275 + ], + "spans": [ + { + "bbox": [ + 248, + 198, + 364, + 275 + ], + "type": "image", + "image_path": "031f2d3b90f4c0d7bdd144659732162d0012cb5800e9299391c9ab148ef642a6.jpg" + } + ] + } + ], + "index": 10, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 290, + 275, + 323, + 282 + ], + "lines": [ + { + "bbox": [ + 290, + 275, + 323, + 282 + ], + "spans": [ + { + "bbox": [ + 290, + 275, + 323, + 282 + ], + "type": "text", + "content": "Ours w/o " + }, + { + "bbox": [ + 290, + 275, + 323, + 282 + ], + "type": "inline_equation", + "content": "\\Phi" + } + ] + } + ], + "index": 11, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 130, + 292, + 482, + 369 + ], + "lines": [ + { + "bbox": [ + 130, + 292, + 482, + 369 + ], + "spans": [ + { + "bbox": [ + 130, + 292, + 482, + 369 + ], + "type": "text", + "content": "Fig. 5: Ablation Study on \"Letter\", including qualitative and quantitative results with/without specular loss " + }, + { + "bbox": [ + 130, + 292, + 482, + 369 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_s" + }, + { + "bbox": [ + 130, + 292, + 482, + 369 + ], + "type": "text", + "content": ", histogram loss " + }, + { + "bbox": [ + 130, + 292, + 482, + 369 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_h" + }, + { + "bbox": [ + 130, + 292, + 482, + 369 + ], + "type": "text", + "content": " and nonlinear mapping " + }, + { + "bbox": [ + 130, + 292, + 482, + 369 + ], + "type": "inline_equation", + "content": "\\varPhi" + }, + { + "bbox": [ + 130, + 292, + 482, + 369 + ], + "type": "text", + "content": ". The left bottom half of the image is the view-dependent color " + }, + { + "bbox": [ + 130, + 292, + 482, + 369 + ], + "type": "inline_equation", + "content": "C_s^{cam}" + }, + { + "bbox": [ + 130, + 292, + 482, + 369 + ], + "type": "text", + "content": ", and the other part is the entire color-correct renderings " + }, + { + "bbox": [ + 130, + 292, + 482, + 369 + ], + "type": "inline_equation", + "content": "C^{cam}" + }, + { + "bbox": [ + 130, + 292, + 482, + 369 + ], + "type": "text", + "content": ". Removing specular loss struggles to the decomposition of view-dependent and view-independent colors. Removing histogram loss lacks the constraint of the entire color space, making the color offset. Downgrading the color translation matrix suffers from artifacts of the nonlinear color difference." + } + ] + } + ], + "index": 14, + "angle": 0, + "type": "image_caption" + } + ], + "index": 10 + }, + { + "type": "image", + "bbox": [ + 364, + 198, + 480, + 275 + ], + "blocks": [ + { + "bbox": [ + 364, + 198, + 480, + 275 + ], + "lines": [ + { + "bbox": [ + 364, + 198, + 480, + 275 + ], + "spans": [ + { + "bbox": [ + 364, + 198, + 480, + 275 + ], + "type": "image", + "image_path": "ef92f0e33ef81c240635e19ab5811d1ccc8ed9ac0efe23f070b842adb6ab9637.jpg" + } + ] + } + ], + "index": 12, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 414, + 275, + 429, + 282 + ], + "lines": [ + { + "bbox": [ + 414, + 275, + 429, + 282 + ], + "spans": [ + { + "bbox": [ + 414, + 275, + 429, + 282 + ], + "type": "text", + "content": "Ours" + } + ] + } + ], + "index": 13, + "angle": 0, + "type": "image_caption" + } + ], + "index": 12 + }, + { + "bbox": [ + 130, + 393, + 482, + 477 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 393, + 482, + 477 + ], + "spans": [ + { + "bbox": [ + 130, + 393, + 482, + 477 + ], + "type": "text", + "content": "MSE error. This may result in color shifts, such as producing darker images than the ground truth. Removing nonlinear mapping and downgrading color translation matrix to " + }, + { + "bbox": [ + 130, + 393, + 482, + 477 + ], + "type": "inline_equation", + "content": "3 \\times 3" + }, + { + "bbox": [ + 130, + 393, + 482, + 477 + ], + "type": "text", + "content": " will neglect the nonlinear process during imaging. This forces the network to model more color variances into " + }, + { + "bbox": [ + 130, + 393, + 482, + 477 + ], + "type": "inline_equation", + "content": "\\mathbf{C}_s^{cam}" + }, + { + "bbox": [ + 130, + 393, + 482, + 477 + ], + "type": "text", + "content": ", making the color correction module unable to fully perceive the color characteristics of each camera, as shown in Fig. 5. These ablation studies demonstrate the effectiveness of our model and specifically designed losses." + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 132, + 495, + 250, + 506 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 495, + 250, + 506 + ], + "spans": [ + { + "bbox": [ + 132, + 495, + 250, + 506 + ], + "type": "text", + "content": "4.5 Color Translation" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 130, + 514, + 482, + 563 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 514, + 482, + 563 + ], + "spans": [ + { + "bbox": [ + 130, + 514, + 482, + 563 + ], + "type": "text", + "content": "Fig. 6 presents examples of color translation guided by a reference image. This advantage makes our method particularly suitable for applications where accurate color reproduction and adherence to reference images are crucial, such as in virtual reality, computer graphics, and image synthesis." + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 132, + 580, + 220, + 594 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 580, + 220, + 594 + ], + "spans": [ + { + "bbox": [ + 132, + 580, + 220, + 594 + ], + "type": "text", + "content": "5 Conclusion" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 130, + 605, + 482, + 666 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 605, + 482, + 666 + ], + "spans": [ + { + "bbox": [ + 130, + 605, + 482, + 666 + ], + "type": "text", + "content": "This paper proposes a framework that incorporates a color correction module within a NeRF to achieve unified color radiance field reconstruction from multiview images captured under different camera settings and lighting conditions. Specifically, we have developed a color correction module to predict a nonlinear color translation matrix that simulates the physical imaging process. Besides the" + } + ] + } + ], + "index": 19 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 337, + 91, + 447, + 100 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 337, + 91, + 447, + 100 + ], + "spans": [ + { + "bbox": [ + 337, + 91, + 447, + 100 + ], + "type": "text", + "content": "Color Correction for NeRF" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 470, + 91, + 480, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 470, + 91, + 480, + 100 + ], + "spans": [ + { + "bbox": [ + 470, + 91, + 480, + 100 + ], + "type": "text", + "content": "13" + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 12 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 137, + 113, + 481, + 411 + ], + "blocks": [ + { + "bbox": [ + 137, + 113, + 481, + 411 + ], + "lines": [ + { + "bbox": [ + 137, + 113, + 481, + 411 + ], + "spans": [ + { + "bbox": [ + 137, + 113, + 481, + 411 + ], + "type": "image", + "image_path": "77315023113e85d70c608821a9b80273c541a305f6d827f9ba101c3868544750.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 130, + 418, + 482, + 453 + ], + "lines": [ + { + "bbox": [ + 130, + 418, + 482, + 453 + ], + "spans": [ + { + "bbox": [ + 130, + 418, + 482, + 453 + ], + "type": "text", + "content": "Fig. 6: Examples of color translation. Our method produces spatially and temporally consistent color transfer results that faithfully match the reference image while maintaining photorealism. Video can also be found in supplementary materials." + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_caption" + } + ], + "index": 2 + }, + { + "bbox": [ + 130, + 491, + 482, + 575 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 491, + 482, + 575 + ], + "spans": [ + { + "bbox": [ + 130, + 491, + 482, + 575 + ], + "type": "text", + "content": "color correction module for accommodating external differences across distinct cameras, we predict a view-dependent function to separate inherent differences caused by scene properties. To enhance the accuracy of the color translation, we have designed various regularization strategies and demonstrated their efficacy through experiments. Our approach significantly improves color consistency in NeRF reconstruction, enabling users to render novel views under flexible and unrestricted settings." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 130, + 582, + 482, + 666 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 582, + 482, + 666 + ], + "spans": [ + { + "bbox": [ + 130, + 582, + 482, + 666 + ], + "type": "text", + "content": "Despite its advantages, we acknowledge that our framework has certain limitations. For instance, compared to NeRF, the proposed Histogram loss results in an increased training time of approximately " + }, + { + "bbox": [ + 130, + 582, + 482, + 666 + ], + "type": "inline_equation", + "content": "10\\%" + }, + { + "bbox": [ + 130, + 582, + 482, + 666 + ], + "type": "text", + "content": ". Furthermore, the Histogram loss, which computes the difference between 2D images, performs well when the rendered image and reference image are nearby. This limitation could potentially be addressed by introducing a new representation and a 3D Histogram loss. We plan to explore these improvements in future work." + } + ] + } + ], + "index": 5 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "spans": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "type": "text", + "content": "14" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 166, + 91, + 267, + 102 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 166, + 91, + 267, + 102 + ], + "spans": [ + { + "bbox": [ + 166, + 91, + 267, + 102 + ], + "type": "text", + "content": "Q. Zhang, Y. Feng et al." + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 13 + }, + { + "para_blocks": [ + { + "bbox": [ + 132, + 114, + 197, + 126 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 114, + 197, + 126 + ], + "spans": [ + { + "bbox": [ + 132, + 114, + 197, + 126 + ], + "type": "text", + "content": "References" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 138, + 141, + 480, + 665 + ], + "type": "list", + "angle": 0, + "index": 19, + "blocks": [ + { + "bbox": [ + 138, + 141, + 480, + 176 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 141, + 480, + 176 + ], + "spans": [ + { + "bbox": [ + 138, + 141, + 480, + 176 + ], + "type": "text", + "content": "1. Affi, M., Barron, J.T., LeGendre, C., Tsai, Y.T., Bleibel, F.: Cross-camera convolutional color constancy. In: Proceedings of the IEEE/CVF International Conference on Computer Vision. pp. 1981-1990 (2021)" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 138, + 176, + 480, + 198 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 176, + 480, + 198 + ], + "spans": [ + { + "bbox": [ + 138, + 176, + 480, + 198 + ], + "type": "text", + "content": "2. Affi, M., Brown, M.S.: Interactive white balancing for camera-rendered images. arXiv preprint arXiv:2009.12632 (2020)" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 138, + 198, + 480, + 231 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 198, + 480, + 231 + ], + "spans": [ + { + "bbox": [ + 138, + 198, + 480, + 231 + ], + "type": "text", + "content": "3. Affifi, M., Brubaker, M.A., Brown, M.S.: Auto white-balance correction for mixed-illuminant scenes. In: Proceedings of the IEEE/CVF Winter Conference on Applications of Computer Vision. pp. 1210–1219 (2022)" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 138, + 232, + 480, + 275 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 232, + 480, + 275 + ], + "spans": [ + { + "bbox": [ + 138, + 232, + 480, + 275 + ], + "type": "text", + "content": "4. Barron, J.T., Mildenhall, B., Tancik, M., Hedman, P., Martin-Brualla, R., Srinivasan, P.P.: Mip-nerf: A multiscale representation for anti-aliasing neural radiance fields. In: Proceedings of the IEEE/CVF International Conference on Computer Vision. pp. 5855-5864 (2021)" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 138, + 276, + 480, + 308 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 276, + 480, + 308 + ], + "spans": [ + { + "bbox": [ + 138, + 276, + 480, + 308 + ], + "type": "text", + "content": "5. Barron, J.T., Mildenhall, B., Verbin, D., Srinivasan, P.P., Hedman, P.: Zip-nerf: Anti-aliased grid-based neural radiance fields. arXiv preprint arXiv:2304.06706 (2023)" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 138, + 309, + 480, + 331 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 309, + 480, + 331 + ], + "spans": [ + { + "bbox": [ + 138, + 309, + 480, + 331 + ], + "type": "text", + "content": "6. Barron, J.T., Tsai, Y.T.: Fast fourier color constancy. In: Proceedings of the IEEE conference on computer vision and pattern recognition. pp. 886-894 (2017)" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 138, + 332, + 480, + 365 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 332, + 480, + 365 + ], + "spans": [ + { + "bbox": [ + 138, + 332, + 480, + 365 + ], + "type": "text", + "content": "7. Bian, W., Wang, Z., Li, K., Bian, J.W., Prisacariu, V.A.: Nope-nerf: Optimising neural radiance field with no pose prior. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 4160-4169 (2023)" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 138, + 365, + 480, + 398 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 365, + 480, + 398 + ], + "spans": [ + { + "bbox": [ + 138, + 365, + 480, + 398 + ], + "type": "text", + "content": "8. Bianco, S., Bruna, A., Naccari, F., Schettini, R.: Color space transformations for digital photography exploiting information about the illuminant estimation process. JOSA A 29(3), 374-384 (2012)" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 138, + 399, + 480, + 431 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 399, + 480, + 431 + ], + "spans": [ + { + "bbox": [ + 138, + 399, + 480, + 431 + ], + "type": "text", + "content": "9. Boss, M., Braun, R., Jampani, V., Barron, J.T., Liu, C., Lensch, H.: Nerd: Neural reflectance decomposition from image collections. In: Proceedings of the IEEE/CVF International Conference on Computer Vision. pp. 12684-12694 (2021)" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 138, + 432, + 480, + 464 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 432, + 480, + 464 + ], + "spans": [ + { + "bbox": [ + 138, + 432, + 480, + 464 + ], + "type": "text", + "content": "0. Boss, M., Jampani, V., Braun, R., Liu, C., Barron, J., Lensch, H.: Neural-pil: Neural pre-integrated lighting for reflectance decomposition. Advances in Neural Information Processing Systems 34, 10691-10704 (2021)" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 138, + 465, + 480, + 487 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 465, + 480, + 487 + ], + "spans": [ + { + "bbox": [ + 138, + 465, + 480, + 487 + ], + "type": "text", + "content": "1. Buchsbaum, G.: A spatial processor model for object colour perception. Journal of the Franklin institute 310(1), 1-26 (1980)" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 138, + 488, + 480, + 520 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 488, + 480, + 520 + ], + "spans": [ + { + "bbox": [ + 138, + 488, + 480, + 520 + ], + "type": "text", + "content": "2. Cao, A., Johnson, J.: Hexplane: A fast representation for dynamic scenes. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 130-141 (2023)" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 138, + 521, + 480, + 543 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 521, + 480, + 543 + ], + "spans": [ + { + "bbox": [ + 138, + 521, + 480, + 543 + ], + "type": "text", + "content": "3. Chen, A., Xu, Z., Geiger, A., Yu, J., Su, H.: Tensorf: Tensorial radiance fields. In: European Conference on Computer Vision. pp. 333-350. Springer (2022)" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 138, + 544, + 480, + 586 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 544, + 480, + 586 + ], + "spans": [ + { + "bbox": [ + 138, + 544, + 480, + 586 + ], + "type": "text", + "content": "4. Chen, A., Xu, Z., Zhao, F., Zhang, X., Xiang, F., Yu, J., Su, H.: Mvsnerf: Fast generalizable radiance field reconstruction from multi-view stereo. In: Proceedings of the IEEE/CVF international conference on computer vision. pp. 14124-14133 (2021)" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 138, + 587, + 480, + 620 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 587, + 480, + 620 + ], + "spans": [ + { + "bbox": [ + 138, + 587, + 480, + 620 + ], + "type": "text", + "content": "5. Chen, X., Zhang, Q., Li, X., Chen, Y., Feng, Y., Wang, X., Wang, J.: Hallucinated neural radiance fields in the wild. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 12943-12952 (2022)" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 138, + 621, + 480, + 665 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 621, + 480, + 665 + ], + "spans": [ + { + "bbox": [ + 138, + 621, + 480, + 665 + ], + "type": "text", + "content": "6. Chen, Y., Chen, X., Wang, X., Zhang, Q., Guo, Y., Shan, Y., Wang, F.: Local-to-global registration for bundle-adjusting neural radiance fields. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 8264-8273 (2023)" + } + ] + } + ], + "index": 18 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 336, + 91, + 447, + 100 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 336, + 91, + 447, + 100 + ], + "spans": [ + { + "bbox": [ + 336, + 91, + 447, + 100 + ], + "type": "text", + "content": "Color Correction for NeRF" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 470, + 92, + 481, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 470, + 92, + 481, + 100 + ], + "spans": [ + { + "bbox": [ + 470, + 92, + 481, + 100 + ], + "type": "text", + "content": "15" + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 14 + }, + { + "para_blocks": [ + { + "bbox": [ + 132, + 116, + 481, + 666 + ], + "type": "list", + "angle": 0, + "index": 17, + "blocks": [ + { + "bbox": [ + 133, + 116, + 481, + 161 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 116, + 481, + 161 + ], + "spans": [ + { + "bbox": [ + 133, + 116, + 481, + 161 + ], + "type": "text", + "content": "17. Chen, Z., Funkhouser, T., Hedman, P., Tagliasacchi, A.: Mobilenerf: Exploiting the polygon rasterization pipeline for efficient neural field rendering on mobile architectures. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 16569-16578 (2023)" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 133, + 162, + 481, + 205 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 162, + 481, + 205 + ], + "spans": [ + { + "bbox": [ + 133, + 162, + 481, + 205 + ], + "type": "text", + "content": "18. Cheng, W., Chen, R., Fan, S., Yin, W., Chen, K., Cai, Z., Wang, J., Gao, Y., Yu, Z., Lin, Z., et al.: Dna-rendering: A diverse neural actor repository for high-fidelity human-centric rendering. In: Proceedings of the IEEE/CVF International Conference on Computer Vision. pp. 19982-19993 (2023)" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 132, + 206, + 481, + 250 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 206, + 481, + 250 + ], + "spans": [ + { + "bbox": [ + 132, + 206, + 481, + 250 + ], + "type": "text", + "content": "19. Chibane, J., Bansal, A., Lazova, V., Pons-Moll, G.: Stereo radiance fields (srf): Learning view synthesis for sparse views of novel scenes. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 7911-7920 (2021)" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 132, + 251, + 481, + 283 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 251, + 481, + 283 + ], + "spans": [ + { + "bbox": [ + 132, + 251, + 481, + 283 + ], + "type": "text", + "content": "20. Ding, C., Ma, Z.: Multi-camera color correction via hybrid histogram matching. IEEE Transactions on Circuits and Systems for Video Technology 31(9), 3327-3337 (2020)" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 132, + 285, + 481, + 318 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 285, + 481, + 318 + ], + "spans": [ + { + "bbox": [ + 132, + 285, + 481, + 318 + ], + "type": "text", + "content": "21. Fang, J., Yi, T., Wang, X., Xie, L., Zhang, X., Liu, W., Nießner, M., Tian, Q.: Fast dynamic radiance fields with time-aware neural voxels. In: SIGGRAPH Asia 2022 Conference Papers. pp. 1-9 (2022)" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 132, + 318, + 481, + 341 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 318, + 481, + 341 + ], + "spans": [ + { + "bbox": [ + 132, + 318, + 481, + 341 + ], + "type": "text", + "content": "22. Fang, Q., Shuai, Q., Dong, J., Bao, H., Zhou, X.: Reconstructing 3d human pose by watching humans in the mirror. In: CVPR (2021)" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 132, + 341, + 481, + 385 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 341, + 481, + 385 + ], + "spans": [ + { + "bbox": [ + 132, + 341, + 481, + 385 + ], + "type": "text", + "content": "23. Fridovich-Keil, S., Meanti, G., Warburg, F.R., Recht, B., Kanazawa, A.: K-planes: Explicit radiance fields in space, time, and appearance. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 12479-12488 (2023)" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 132, + 386, + 481, + 419 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 386, + 481, + 419 + ], + "spans": [ + { + "bbox": [ + 132, + 386, + 481, + 419 + ], + "type": "text", + "content": "24. Fridovich-Keil, S., Yu, A., Tancik, M., Chen, Q., Recht, B., Kanazawa, A.: Plenoxels: Radiance fields without neural networks. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 5501-5510 (2022)" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 132, + 419, + 481, + 453 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 419, + 481, + 453 + ], + "spans": [ + { + "bbox": [ + 132, + 419, + 481, + 453 + ], + "type": "text", + "content": "25. Garbin, S.J., Kowalski, M., Johnson, M., Shotton, J., Valentin, J.: Fastnerf: High-fidelity neural rendering at 200fps. In: Proceedings of the IEEE/CVF International Conference on Computer Vision. pp. 14346-14355 (2021)" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 132, + 453, + 481, + 486 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 453, + 481, + 486 + ], + "spans": [ + { + "bbox": [ + 132, + 453, + 481, + 486 + ], + "type": "text", + "content": "26. Hedman, P., Srinivasan, P.P., Mildenhall, B., Barron, J.T., Debevec, P.: Baking neural radiance fields for real-time view synthesis. In: Proceedings of the IEEE/CVF International Conference on Computer Vision. pp. 5875-5884 (2021)" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 132, + 487, + 481, + 520 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 487, + 481, + 520 + ], + "spans": [ + { + "bbox": [ + 132, + 487, + 481, + 520 + ], + "type": "text", + "content": "27. Huang, X., Zhang, Q., Feng, Y., Li, H., Wang, Q.: Inverting the imaging process by learning an implicit camera model. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 21456-21465 (2023)" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 132, + 520, + 481, + 553 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 520, + 481, + 553 + ], + "spans": [ + { + "bbox": [ + 132, + 520, + 481, + 553 + ], + "type": "text", + "content": "28. Huang, X., Zhang, Q., Feng, Y., Li, H., Wang, X., Wang, Q.: HDR-nerf: High dynamic range neural radiance fields. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 18398-18408 (2022)" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 132, + 554, + 481, + 597 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 554, + 481, + 597 + ], + "spans": [ + { + "bbox": [ + 132, + 554, + 481, + 597 + ], + "type": "text", + "content": "29. Huang, X., Zhang, Q., Feng, Y., Li, X., Wang, X., Wang, Q.: Local implicit ray function for generalizable radiance field representation. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 97-107 (2023)" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 132, + 599, + 481, + 632 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 599, + 481, + 632 + ], + "spans": [ + { + "bbox": [ + 132, + 599, + 481, + 632 + ], + "type": "text", + "content": "30. Huang, Z., Liang, Z., Zhang, H., Lin, Y., Jia, K.: Sur2f: A hybrid representation for high-quality and efficient surface reconstruction from multi-view images. arXiv preprint arXiv:2401.03704 (2024)" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 132, + 632, + 481, + 666 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 632, + 481, + 666 + ], + "spans": [ + { + "bbox": [ + 132, + 632, + 481, + 666 + ], + "type": "text", + "content": "31. Isik, M., Rünz, M., Georgopoulos, M., Khakhulin, T., Starck, J., Agapito, L., Nießner, M.: Humanrf: High-fidelity neural radiance fields for humans in motion. arXiv preprint arXiv:2305.06356 (2023)" + } + ] + } + ], + "index": 16 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "spans": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "type": "text", + "content": "16" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 166, + 91, + 266, + 102 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 166, + 91, + 266, + 102 + ], + "spans": [ + { + "bbox": [ + 166, + 91, + 266, + 102 + ], + "type": "text", + "content": "Q. Zhang, Y. Feng et al." + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 15 + }, + { + "para_blocks": [ + { + "bbox": [ + 130, + 116, + 482, + 665 + ], + "type": "list", + "angle": 0, + "index": 17, + "blocks": [ + { + "bbox": [ + 130, + 116, + 482, + 149 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 116, + 482, + 149 + ], + "spans": [ + { + "bbox": [ + 130, + 116, + 482, + 149 + ], + "type": "text", + "content": "32. Jin, H., Liu, I., Xu, P., Zhang, X., Han, S., Bi, S., Zhou, X., Xu, Z., Su, H.: Tensoroir: Tensorial inverse rendering. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 165-174 (2023)" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 130, + 149, + 482, + 181 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 149, + 482, + 181 + ], + "spans": [ + { + "bbox": [ + 130, + 149, + 482, + 181 + ], + "type": "text", + "content": "33. Johari, M.M., Lepoittevin, Y., Fleuret, F.: Geonerf: Generalizing nerf with geometry priors. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 18365-18375 (2022)" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 130, + 181, + 482, + 224 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 181, + 482, + 224 + ], + "spans": [ + { + "bbox": [ + 130, + 181, + 482, + 224 + ], + "type": "text", + "content": "34. Joo, H., Simon, T., Li, X., Liu, H., Tan, L., Gui, L., Banerjee, S., Godisart, T.S., Nabbe, B., Matthews, I., Kanade, T., Nobuhara, S., Sheikh, Y.: Panoptic studio: A massively multiview system for social interaction capture. IEEE Transactions on Pattern Analysis and Machine Intelligence (2017)" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 130, + 224, + 482, + 256 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 224, + 482, + 256 + ], + "spans": [ + { + "bbox": [ + 130, + 224, + 482, + 256 + ], + "type": "text", + "content": "35. Kerbl, B., Kopanas, G., Leimkuhler, T., Drettakis, G.: 3d gaussian splatting for real-time radiance field rendering. ACM Transactions on Graphics (ToG) 42(4), 1-14 (2023)" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 130, + 256, + 482, + 289 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 256, + 482, + 289 + ], + "spans": [ + { + "bbox": [ + 130, + 256, + 482, + 289 + ], + "type": "text", + "content": "36. Kinh, F., Yilmaz, D., Özcan, B., Kırac, F.: Deterministic neural illumination mapping for efficient auto-white balance correction. In: Proceedings of the IEEE/CVF International Conference on Computer Vision. pp. 1139-1147 (2023)" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 130, + 289, + 482, + 320 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 289, + 482, + 320 + ], + "spans": [ + { + "bbox": [ + 130, + 289, + 482, + 320 + ], + "type": "text", + "content": "37. Kinh, F., Yilmaz, D., Ozcan, B., Kirac, F.: Modeling the lighting in scenes as style for auto white-balance correction. In: Proceedings of the IEEE/CVF Winter Conference on Applications of Computer Vision. pp. 4903-4913 (2023)" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 130, + 320, + 482, + 365 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 320, + 482, + 365 + ], + "spans": [ + { + "bbox": [ + 130, + 320, + 482, + 365 + ], + "type": "text", + "content": "38. Kuang, Z., Luan, F., Bi, S., Shu, Z., Wetzstein, G., Sunkavalli, K.: Palettenerf: Palette-based appearance editing of neural radiance fields. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 20691-20700 (2023)" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 130, + 365, + 482, + 396 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 365, + 482, + 396 + ], + "spans": [ + { + "bbox": [ + 130, + 365, + 482, + 396 + ], + "type": "text", + "content": "39. Lee, J.H., Kim, D.S.: Ice-nerf: Interactive color editing of nerfs via decomposition-aware weight optimization. In: Proceedings of the IEEE/CVF International Conference on Computer Vision. pp. 3491-3501 (2023)" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 130, + 396, + 482, + 429 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 396, + 482, + 429 + ], + "spans": [ + { + "bbox": [ + 130, + 396, + 482, + 429 + ], + "type": "text", + "content": "40. Li, J., Feng, Z., She, Q., Ding, H., Wang, C., Lee, G.H.: Mine: Towards continuous depth MPI with nerf for novel view synthesis. In: Proceedings of the IEEE/CVF International Conference on Computer Vision. pp. 12578-12588 (2021)" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 130, + 429, + 482, + 471 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 429, + 482, + 471 + ], + "spans": [ + { + "bbox": [ + 130, + 429, + 482, + 471 + ], + "type": "text", + "content": "41. Liang, Z., Huang, Z., Ding, C., Jia, K.: Helixsurf: A robust and efficient neural implicit surface learning of indoor scenes with iterative intertwined regularization. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 13165-13174 (2023)" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 130, + 471, + 482, + 503 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 471, + 482, + 503 + ], + "spans": [ + { + "bbox": [ + 130, + 471, + 482, + 503 + ], + "type": "text", + "content": "42. Lin, C.H., Ma, W.C., Torralba, A., Lucey, S.: Barf: Bundle-adjusting neural radiance fields. In: Proceedings of the IEEE/CVF international conference on computer vision. pp. 5741-5751 (2021)" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 130, + 503, + 482, + 546 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 503, + 482, + 546 + ], + "spans": [ + { + "bbox": [ + 130, + 503, + 482, + 546 + ], + "type": "text", + "content": "43. Liu, Y.L., Gao, C., Meuleman, A., Tseng, H.Y., Saraf, A., Kim, C., Chuang, Y.Y., Kopf, J., Huang, J.B.: Robust dynamic radiance fields. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 13-23 (2023)" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 130, + 546, + 482, + 590 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 546, + 482, + 590 + ], + "spans": [ + { + "bbox": [ + 130, + 546, + 482, + 590 + ], + "type": "text", + "content": "44. Liu, Y., Peng, S., Liu, L., Wang, Q., Wang, P., Theobalt, C., Zhou, X., Wang, W.: Neural rays for occlusion-aware image-based rendering. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 7824-7833 (2022)" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 130, + 590, + 482, + 622 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 590, + 482, + 622 + ], + "spans": [ + { + "bbox": [ + 130, + 590, + 482, + 622 + ], + "type": "text", + "content": "45. Ma, L., Li, X., Liao, J., Zhang, Q., Wang, X., Wang, J., Sander, P.V.: Deblur-nerf: Neural radiance fields from blurry images. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 12861-12870 (2022)" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 130, + 622, + 482, + 665 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 622, + 482, + 665 + ], + "spans": [ + { + "bbox": [ + 130, + 622, + 482, + 665 + ], + "type": "text", + "content": "46. Martin-Brualla, R., Radwan, N., Sajjadi, M.S., Barron, J.T., Dosovitskiy, A., Duckworth, D.: Nerf in the wild: Neural radiance fields for unconstrained photo collections. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 7210-7219 (2021)" + } + ] + } + ], + "index": 16 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 337, + 91, + 447, + 100 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 337, + 91, + 447, + 100 + ], + "spans": [ + { + "bbox": [ + 337, + 91, + 447, + 100 + ], + "type": "text", + "content": "Color Correction for NeRF" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 470, + 92, + 480, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 470, + 92, + 480, + 100 + ], + "spans": [ + { + "bbox": [ + 470, + 92, + 480, + 100 + ], + "type": "text", + "content": "17" + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 16 + }, + { + "para_blocks": [ + { + "bbox": [ + 132, + 117, + 481, + 665 + ], + "type": "list", + "angle": 0, + "index": 17, + "blocks": [ + { + "bbox": [ + 133, + 117, + 480, + 138 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 117, + 480, + 138 + ], + "spans": [ + { + "bbox": [ + 133, + 117, + 480, + 138 + ], + "type": "text", + "content": "47. Max, N.: Optical models for direct volume rendering. IEEE Transactions on Visualization and Computer Graphics 1(2), 99-108 (1995)" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 133, + 140, + 481, + 183 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 140, + 481, + 183 + ], + "spans": [ + { + "bbox": [ + 133, + 140, + 481, + 183 + ], + "type": "text", + "content": "48. Mildenhall, B., Hedman, P., Martin-Brualla, R., Srinivasan, P.P., Barron, J.T.: Nerf in the dark: High dynamic range view synthesis from noisy raw images. In: Proceedings of the IEEE/CVF conference on computer vision and pattern recognition. pp. 16190-16199 (2022)" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 134, + 185, + 480, + 217 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 134, + 185, + 480, + 217 + ], + "spans": [ + { + "bbox": [ + 134, + 185, + 480, + 217 + ], + "type": "text", + "content": "49. Mildenhall, B., Srinivasan, P.P., Tancik, M., Barron, J.T., Ramamoorthi, R., Ng, R.: Nerf: Representing scenes as neural radiance fields for view synthesis. In: European Conference on Computer Vision. pp. 405-421 (2020)" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 133, + 219, + 480, + 251 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 219, + 480, + 251 + ], + "spans": [ + { + "bbox": [ + 133, + 219, + 480, + 251 + ], + "type": "text", + "content": "50. Müller, T., Evans, A., Schied, C., Keller, A.: Instant neural graphics primitives with a multiresolution hash encoding. ACM Transactions on Graphics (ToG) 41(4), 1-15 (2022)" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 132, + 252, + 480, + 284 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 252, + 480, + 284 + ], + "spans": [ + { + "bbox": [ + 132, + 252, + 480, + 284 + ], + "type": "text", + "content": "51. Park, K., Henzler, P., Mildenhall, B., Barron, J.T., Martin-Brualla, R.: Camp: Camera preconditioning for neural radiance fields. ACM Transactions on Graphics (TOG) 42(6), 1-11 (2023)" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 132, + 285, + 480, + 318 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 285, + 480, + 318 + ], + "spans": [ + { + "bbox": [ + 132, + 285, + 480, + 318 + ], + "type": "text", + "content": "52. Pearl, N., Treibitz, T., Korman, S.: Nan: Noise-aware nerfs for burst-denoising. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 12672-12681 (2022)" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 132, + 319, + 480, + 351 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 319, + 480, + 351 + ], + "spans": [ + { + "bbox": [ + 132, + 319, + 480, + 351 + ], + "type": "text", + "content": "53. Peng, S., Zhang, Y., Xu, Y., Wang, Q., Shuai, Q., Bao, H., Zhou, X.: Neural body: Implicit neural representations with structured latent codes for novel view synthesis of dynamic humans. In: CVPR (2021)" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 132, + 353, + 480, + 385 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 353, + 480, + 385 + ], + "spans": [ + { + "bbox": [ + 132, + 353, + 480, + 385 + ], + "type": "text", + "content": "54. Reiser, C., Peng, S., Liao, Y., Geiger, A.: Kilonerf: Speeding up neural radiance fields with thousands of tiny mlp's. In: Proceedings of the IEEE/CVF International Conference on Computer Vision. pp. 14335-14345 (2021)" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 132, + 387, + 480, + 419 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 387, + 480, + 419 + ], + "spans": [ + { + "bbox": [ + 132, + 387, + 480, + 419 + ], + "type": "text", + "content": "55. Rematas, K., Liu, A., Srinivasan, P.P., Barron, J.T., Tagliasacchi, A., Funkhouser, T., Ferrari, V.: Urban radiance fields. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 12932-12942 (2022)" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 132, + 420, + 480, + 453 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 420, + 480, + 453 + ], + "spans": [ + { + "bbox": [ + 132, + 420, + 480, + 453 + ], + "type": "text", + "content": "56. Rodríguez, R.G., Vazquez-Corral, J., Bertalmío, M.: Color matching images with unknown non-linear encodings. IEEE Transactions on Image Processing 29, 4435-4444 (2020)" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 133, + 454, + 480, + 497 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 454, + 480, + 497 + ], + "spans": [ + { + "bbox": [ + 133, + 454, + 480, + 497 + ], + "type": "text", + "content": "57. Shao, R., Zheng, Z., Tu, H., Liu, B., Zhang, H., Liu, Y.: Tensor4d: Efficient neural 4d decomposition for high-fidelity dynamic reconstruction and rendering. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 16632-16642 (2023)" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 133, + 498, + 480, + 542 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 498, + 480, + 542 + ], + "spans": [ + { + "bbox": [ + 133, + 498, + 480, + 542 + ], + "type": "text", + "content": "58. Srinivasan, P.P., Deng, B., Zhang, X., Tancik, M., Mildenhall, B., Barron, J.T.: Nerv: Neural reflectance and visibility fields for relighting and view synthesis. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 7495-7504 (2021)" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 133, + 544, + 480, + 586 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 544, + 480, + 586 + ], + "spans": [ + { + "bbox": [ + 133, + 544, + 480, + 586 + ], + "type": "text", + "content": "59. Sun, G., Chen, X., Chen, Y., Pang, A., Lin, P., Jiang, Y., Xu, L., Yu, J., Wang, J.: Neural free-viewpoint performance rendering under complex human-object interactions. In: Proceedings of the 29th ACM International Conference on Multimedia. pp. 4651-4660 (2021)" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 132, + 588, + 480, + 631 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 588, + 480, + 631 + ], + "spans": [ + { + "bbox": [ + 132, + 588, + 480, + 631 + ], + "type": "text", + "content": "60. Tancik, M., Casser, V., Yan, X., Pradhan, S., Mildenhall, B., Srinivasan, P.P., Barron, J.T., Kretzschmar, H.: Block-nerf: Scalable large scene neural view synthesis. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 8248-8258 (2022)" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 132, + 632, + 480, + 665 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 632, + 480, + 665 + ], + "spans": [ + { + "bbox": [ + 132, + 632, + 480, + 665 + ], + "type": "text", + "content": "61. Tang, J., Zhou, H., Chen, X., Hu, T., Ding, E., Wang, J., Zeng, G.: Delicate textured mesh recovery from nef via adaptive surface refinement. arXiv preprint arXiv:2303.02091 (2023)" + } + ] + } + ], + "index": 16 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "spans": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "type": "text", + "content": "18" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 166, + 91, + 266, + 102 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 166, + 91, + 266, + 102 + ], + "spans": [ + { + "bbox": [ + 166, + 91, + 266, + 102 + ], + "type": "text", + "content": "Q. Zhang, Y. Feng et al." + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 17 + }, + { + "para_blocks": [ + { + "bbox": [ + 130, + 116, + 482, + 658 + ], + "type": "list", + "angle": 0, + "index": 18, + "blocks": [ + { + "bbox": [ + 130, + 116, + 482, + 160 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 116, + 482, + 160 + ], + "spans": [ + { + "bbox": [ + 130, + 116, + 482, + 160 + ], + "type": "text", + "content": "62. Verbin, D., Hedman, P., Mildenhall, B., Zickler, T., Barron, J.T., Srinivasan, P.P.: Ref-nerf: Structured view-dependent appearance for neural radiance fields. In: 2022 IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR). pp. 5481-5490. IEEE (2022)" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 130, + 160, + 481, + 204 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 160, + 481, + 204 + ], + "spans": [ + { + "bbox": [ + 130, + 160, + 481, + 204 + ], + "type": "text", + "content": "63. Wang, Q., Wang, Z., Genova, K., Srinivasan, P.P., Zhou, H., Barron, J.T., Martin-Brualla, R., Snavely, N., Funkhouser, T.: Ibrnet: Learning multi-view image-based rendering. In: Proceedings of the IEEE/CVF conference on computer vision and pattern recognition. pp. 4690-4699 (2021)" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 130, + 204, + 481, + 236 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 204, + 481, + 236 + ], + "spans": [ + { + "bbox": [ + 130, + 204, + 481, + 236 + ], + "type": "text", + "content": "64. Xie, S., Zhu, H., Liu, Z., Zhang, Q., Zhou, Y., Cao, X., Ma, Z.: Diner: Disorder-invariant implicit neural representation. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 6143-6152 (2023)" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 130, + 236, + 481, + 269 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 236, + 481, + 269 + ], + "spans": [ + { + "bbox": [ + 130, + 236, + 481, + 269 + ], + "type": "text", + "content": "65. Xu, Q., Xu, Z., Philip, J., Bi, S., Shu, Z., Sunkavalli, K., Neumann, U.: Point-nerf: Point-based neural radiance fields. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 5438-5448 (2022)" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 130, + 269, + 481, + 301 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 269, + 481, + 301 + ], + "spans": [ + { + "bbox": [ + 130, + 269, + 481, + 301 + ], + "type": "text", + "content": "66. Yao, Y., Zhang, J., Liu, J., Qu, Y., Fang, T., McKinnon, D., Tsin, Y., Quan, L.: Neilf: Neural incident light field for physically-based material estimation. In: European Conference on Computer Vision. pp. 700-716. Springer (2022)" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 130, + 301, + 481, + 323 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 301, + 481, + 323 + ], + "spans": [ + { + "bbox": [ + 130, + 301, + 481, + 323 + ], + "type": "text", + "content": "67. Yen-Chen, l.: Nerf-pytorch: a pytorch implementation of nerf. URL https://github.com/yenchenlin/nerf-pytorch (2020)" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 130, + 323, + 481, + 355 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 323, + 481, + 355 + ], + "spans": [ + { + "bbox": [ + 130, + 323, + 481, + 355 + ], + "type": "text", + "content": "68. Yu, A., Ye, V., Tancik, M., Kanazawa, A.: pixelnerf: Neural radiance fields from one or few images. In: Proceedings of the IEEE/CVF conference on computer vision and pattern recognition. pp. 4578-4587 (2021)" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 130, + 355, + 481, + 388 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 355, + 481, + 388 + ], + "spans": [ + { + "bbox": [ + 130, + 355, + 481, + 388 + ], + "type": "text", + "content": "69. Zhang, K., Luan, F., Li, Z., Snavely, N.: Iron: Inverse rendering by optimizing neural sdfs and materials from photometric images. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 5565-5574 (2022)" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 130, + 388, + 481, + 420 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 388, + 481, + 420 + ], + "spans": [ + { + "bbox": [ + 130, + 388, + 481, + 420 + ], + "type": "text", + "content": "70. Zhang, R., Isola, P., Efros, A.A., Shechtman, E., Wang, O.: The unreasonable effectiveness of deep features as a perceptual metric. In: Proceedings of the IEEE conference on computer vision and pattern recognition. pp. 586-595 (2018)" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 130, + 420, + 481, + 453 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 420, + 481, + 453 + ], + "spans": [ + { + "bbox": [ + 130, + 420, + 481, + 453 + ], + "type": "text", + "content": "71. Zhang, X., Srinivasan, P.P., Deng, B., Debevec, P., Freeman, W.T., Barron, J.T.: Nerfactor: Neural factorization of shape and reflectance under an unknown illumination. ACM Transactions on Graphics (ToG) 40(6), 1-18 (2021)" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 130, + 452, + 481, + 484 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 452, + 481, + 484 + ], + "spans": [ + { + "bbox": [ + 130, + 452, + 481, + 484 + ], + "type": "text", + "content": "72. Zhao, Y., Ferguson, S., Zhou, H., Elliott, C., Rafferty, K.: Color alignment for relative color constancy via non-standard references. IEEE Transactions on Image Processing 31, 6591-6604 (2022)" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 130, + 484, + 481, + 528 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 484, + 481, + 528 + ], + "spans": [ + { + "bbox": [ + 130, + 484, + 481, + 528 + ], + "type": "text", + "content": "73. Zhou, T., He, K., Wu, D., Xu, T., Zhang, Q., Shao, K., Chen, W., Xu, L., Yu, J.: Relightable neural human assets from multi-view gradient illuminations. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 4315-4327 (2023)" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 130, + 528, + 481, + 549 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 528, + 481, + 549 + ], + "spans": [ + { + "bbox": [ + 130, + 528, + 481, + 549 + ], + "type": "text", + "content": "74. Zhu, H., Liu, F., Zhang, Q., Cao, X., Ma, Z.: Rhino: Regularizing the hash-based implicit neural representation. arXiv preprint arXiv:2309.12642 (2023)" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 130, + 549, + 481, + 582 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 549, + 481, + 582 + ], + "spans": [ + { + "bbox": [ + 130, + 549, + 481, + 582 + ], + "type": "text", + "content": "75. Zhu, H., Xie, S., Liu, Z., Liu, F., Zhang, Q., Zhou, Y., Lin, Y., Ma, Z., Cao, X.: Disorder-invariant implicit neural representation. IEEE Transactions on Pattern Analysis and Machine Intelligence (2024)" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 130, + 582, + 481, + 613 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 582, + 481, + 613 + ], + "spans": [ + { + "bbox": [ + 130, + 582, + 481, + 613 + ], + "type": "text", + "content": "76. Zhu, J., Zhu, H., Zhang, Q., Zhu, F., Ma, Z., Cao, X.: Pyramid nef: Frequency guided fast radiance field optimization. International Journal of Computer Vision 131(10), 2649-2664 (2023)" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 130, + 614, + 481, + 658 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 614, + 481, + 658 + ], + "spans": [ + { + "bbox": [ + 130, + 614, + 481, + 658 + ], + "type": "text", + "content": "77. Zhuang, Y., Zhang, Q., Wang, X., Zhu, H., Feng, Y., Li, X., Shan, Y., Cao, X.: A pre-convolved representation for plug-and-play neural illumination fields. In: Proceedings of the AAAI Conference on Artificial Intelligence. vol. 38, pp. 7828-7836 (2024)" + } + ] + } + ], + "index": 17 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 337, + 91, + 447, + 100 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 337, + 91, + 447, + 100 + ], + "spans": [ + { + "bbox": [ + 337, + 91, + 447, + 100 + ], + "type": "text", + "content": "Color Correction for NeRF" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 470, + 91, + 481, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 470, + 91, + 481, + 100 + ], + "spans": [ + { + "bbox": [ + 470, + 91, + 481, + 100 + ], + "type": "text", + "content": "19" + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 18 + } + ], + "_backend": "vlm", + "_version_name": "2.6.4" +} \ No newline at end of file diff --git a/2024/Physics-Based Interaction with 3D Objects via Video Generation/017e5389-f4df-4a9c-a529-0be7433a235c_content_list.json b/2024/Physics-Based Interaction with 3D Objects via Video Generation/017e5389-f4df-4a9c-a529-0be7433a235c_content_list.json new file mode 100644 index 0000000000000000000000000000000000000000..4250305b3c0c177a7adef8441cf143a7b3f1d09c --- /dev/null +++ b/2024/Physics-Based Interaction with 3D Objects via Video Generation/017e5389-f4df-4a9c-a529-0be7433a235c_content_list.json @@ -0,0 +1,2052 @@ +[ + { + "type": "text", + "text": "Physics-Based Interaction with 3D Objects via Video Generation", + "text_level": 1, + "bbox": [ + 256, + 140, + 750, + 184 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Tianyuan Zhang $^{1}$ , Hong-Xing Yu $^{2}$ , Rundi Wu $^{3}$ , Brandon Y. Feng $^{1}$ , Changxi Zheng $^{3}$ , Noah Snavely $^{4}$ , Jiajun Wu $^{2}$ , and William T. Freeman $^{1}$", + "bbox": [ + 240, + 212, + 759, + 244 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "1 Massachusetts Institute of Technology", + "bbox": [ + 364, + 253, + 635, + 268 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "2 Stanford University", + "bbox": [ + 426, + 268, + 573, + 282 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "3 Columbia University", + "bbox": [ + 426, + 282, + 576, + 296 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "4 Cornell University", + "bbox": [ + 431, + 296, + 568, + 311 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Abstract. Realistic object interactions are crucial for creating immersive virtual experiences, yet synthesizing realistic 3D object dynamics in response to novel interactions remains a significant challenge. Unlike unconditional or text-conditioned dynamics generation, action-conditioned dynamics requires perceiving the physical material properties of objects and grounding the 3D motion prediction on these properties, such as object stiffness. However, estimating physical material properties is an open problem due to the lack of material ground-truth data, as measuring these properties for real objects is highly difficult. We present PhysDreamer, a physics-based approach that endows static 3D objects with interactive dynamics by leveraging the object dynamics priors learned by video generation models. By distilling these priors, PhysDreamer enables the synthesis of realistic object responses to novel interactions, such as external forces or agent manipulations. We demonstrate our approach on diverse examples of elastic objects and evaluate the realism of the synthesized interactions through a user study. PhysDreamer takes a step towards more engaging and realistic virtual experiences by enabling static 3D objects to dynamically respond to interactive stimuli in a physically plausible manner. See our project page at https://physdreamer.github.io/.", + "bbox": [ + 259, + 351, + 743, + 630 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Keywords: Physics-based modeling $\\cdot$ Interactive 3D dynamics", + "bbox": [ + 261, + 643, + 686, + 657 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "1 Introduction", + "text_level": 1, + "bbox": [ + 215, + 684, + 375, + 700 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Realistic object interactions play a pivotal role in creating immersive virtual experiences. Recent advances in 3D vision have enabled the capture and creation of high-quality static 3D assets [34,50], and some methods even extend to 4D assets [46,47,58], generating unconditioned dynamics. However, these methods fail to handle action-conditioned dynamics in response to new physical interactions, such as synthesizing the motion of a rose reacting to a breeze or a touch.", + "bbox": [ + 212, + 718, + 787, + 808 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "The key challenge in synthesizing action-conditioned dynamics lies in understanding the physical material properties of objects. Yet, estimating these", + "bbox": [ + 212, + 809, + 787, + 840 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/b6cf218f3de04570fd2e6afb26ad153d78b79bb65c175233a8d8b1a554215f09.jpg", + "image_caption": [ + "PhysDreamer" + ], + "image_footnote": [], + "bbox": [ + 223, + 143, + 395, + 277 + ], + "page_idx": 1 + }, + { + "type": "image", + "img_path": "images/979f1f2ac25d68411281b0a66c3d37e7f3ba55756d7595f5b67ce4f73d30b6ab.jpg", + "image_caption": [ + "Fig. 1: (Left) Leveraging and distilling dynamics priors from a pre-trained video generation model, we estimate a physical material field for the static 3D object. (Right) The physical material field allows synthesizing interactive 3D dynamics under arbitrary forces. We show rendered sequences from two viewpoints. Red arrows indicate force directions. Please see videos on our project website for better visualization." + ], + "image_footnote": [], + "bbox": [ + 405, + 145, + 460, + 275 + ], + "page_idx": 1 + }, + { + "type": "image", + "img_path": "images/0eab3ab33285313cd3d460ca63d4649625a1da6c22c7bf04b263eb6070414436.jpg", + "image_caption": [ + "Synthesized 3D dynamics under interaction" + ], + "image_footnote": [], + "bbox": [ + 467, + 145, + 782, + 275 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "properties is a challenging task due to the lack of ground-truth data, as measuring these properties for real objects is highly difficult. Real-life objects often exhibit complex, spatially-varying material properties, making the estimation problem even more challenging. Despite the complexity of physical materials, humans can easily imagine how objects would react to external forces, such as the gentle sway of a rose. This ability to imagine object dynamics stems from our physical prior knowledge obtained from observing and interacting with the physical world. This motivates us to distill dynamics priors from video generation models that have been trained on vast, diverse video observations of the physical world.", + "bbox": [ + 212, + 404, + 787, + 556 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "In this work, we focus on synthesizing interactive 3D dynamics. We propose PhysDreamer, a physics-based approach to transforming static 3D objects into interactive ones that can respond to novel interactions. The key idea behind PhysDreamer is to distill dynamics priors learned by video generation models to estimate the physical material properties of static 3D objects. We hypothesize that video generation models, trained on large amounts of video data, implicitly capture the relationship between object appearance and dynamics. By leveraging this learned prior knowledge, PhysDreamer can infer the physical material properties that drive the dynamic behavior of objects, even in the absence of ground-truth material data (Fig. 1).", + "bbox": [ + 212, + 561, + 789, + 714 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "PhysDreamer represents 3D objects using 3D Gaussians [34], models the physical material field with a neural field [71], and simulates 3D dynamics using the differentiable Material Point Method (MPM) [27]. The differentiable simulation and rendering allow for direct optimization of the physical material field and initial velocity field by matching pixel space observations. We focus on elastic dynamics and showcase PhysDreamer through diverse real examples, such as flowers, plants, a beanie hat, and a telephone cord. We evaluate the realism of the synthesized interactive motion through a user study, comparing PhysDreamer", + "bbox": [ + 212, + 719, + 787, + 840 + ], + "page_idx": 1 + }, + { + "type": "page_number", + "text": "2", + "bbox": [ + 217, + 114, + 228, + 126 + ], + "page_idx": 1 + }, + { + "type": "header", + "text": "Zhang et al.", + "bbox": [ + 271, + 114, + 354, + 128 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "to state-of-the-art methods. The results demonstrate that our approach significantly outperforms existing techniques on motion realism and visual quality.", + "bbox": [ + 212, + 146, + 782, + 176 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "In summary, PhysDreamer addresses the challenge of synthesizing interactive 3D dynamics by leveraging the object dynamics priors learned by video generation models. By distilling these priors to estimate the physical material properties of static 3D objects, our approach enables the creation of immersive virtual experiences where objects can respond realistically to novel interactions. The main contributions of our work include enabling static 3D objects to dynamically respond to interactive stimuli in a physically plausible manner and taking a step towards more engaging and realistic virtual experiences. We believe that PhysDreamer has the potential to greatly enhance the realism and interactivity of virtual environments, paving the way for more engaging and lifelike simulations.", + "bbox": [ + 212, + 176, + 787, + 340 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "2 Related work", + "text_level": 1, + "bbox": [ + 214, + 364, + 382, + 380 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "2.1 Dynamic 3D reconstruction", + "text_level": 1, + "bbox": [ + 214, + 393, + 490, + 409 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Dynamic 3D reconstruction methods aim to reconstruct a representation of a dynamic scene from inputs such as depth scans [13,41], RGBD videos [51], or monocular or multi-view videos [1,7,39,45,47,52,53,56,66,69,74,75]. This task is especially challenging in the monocular setting with slow-moving cameras and fast-moving scenes [20]. Novel scene representations are a major driver of recent progress. One prominent approach is to augment a canonical Neural Radiance Fields (NeRF) with a deformation field [56]. This approach can be further improved by incorporating flow supervision [23,66] or as-rigid-as-possible or volume preserving regularization terms [52,53]. Time-modulated NeRFs [8,19,20,43] offer a simpler alternative representation. Due to its Lagrangian nature, 3D Gaussian Splatting [34] is readily adaptable to the task of efficient dynamic scene reconstruction [17,30,39,47,72,74]. Data-driven prior, such as from monocular depth models [38,77] and image diffusion models [67], can also be used to reduce the inherent ambiguity in dynamic reconstruction from monocular videos.", + "bbox": [ + 212, + 417, + 787, + 630 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "2.2 Dynamic 3D generation", + "text_level": 1, + "bbox": [ + 214, + 650, + 459, + 666 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Our work also relates to efforts to synthesize dynamic 3D scenes. A common approach is to integrate a 3D generation pipeline with a video generation model [2,46,58,60]. For instance, Make-A-Video3D begins by creating a static NeRF as per DreamFusion [55], then extending it temporally using Score Distillation Sampling (SDS) [55] derived from a video diffusion model. The approach can be improved with more efficient representations, stronger diffusion priors, and stable training techniques [2, 46]. However, applying SDS with video diffusion models demands significant computational and memory costs. Compact4D [73] and DreamGaussian4D [58] used a more efficient approach, synthesizing 3D dynamics by aligning a reference video from video generation models while employing SDS from image diffusion models to reduce novel view artifacts. These", + "bbox": [ + 212, + 672, + 787, + 840 + ], + "page_idx": 2 + }, + { + "type": "header", + "text": "PhysDreamer", + "bbox": [ + 638, + 114, + 730, + 128 + ], + "page_idx": 2 + }, + { + "type": "page_number", + "text": "3", + "bbox": [ + 774, + 116, + 785, + 126 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "methods are currently limited to producing fixed-length 3D videos. We focus on synthesizing interactive 3D motions under any new physical interactions.", + "bbox": [ + 212, + 146, + 782, + 176 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "2.3 Interactive motion generation", + "text_level": 1, + "bbox": [ + 214, + 199, + 506, + 214 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Interactive motion generation animates still images or 3D contents according to user inputs like text [11, 76], motion fields [21], motion layers [12, 14], or direct manipulation such as dragging and pulling [15, 44]. Early work from Davis et al. [15, 16] demonstrated animating an image using an image-space modal basis extracted from a video of an object undergoing subtle vibrational motions. Building upon this image-space representation [15], Generative Image Dynamics [44] used a diffusion model trained on a dataset with paired image and its modal basis to model scene motion distributions, enabling realistic interaction with still input images. We focus on interacting with 3D objects rather than images.", + "bbox": [ + 212, + 223, + 784, + 359 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "For 3D assets, physics-based approaches enable synthesizing motions under any physical interactions. Virtual Elastic Objects [10] jointly reconstructs the geometry, appearances, and physical parameters of elastic objects in a multiview capture setup with compressed air system. PAC-NeRF [42], DANO [40], and PhysGaussian [18] integrate physics-based simulations with NeRF and 3D Gaussians to generate physically plausible motions. We use the same physics-based approach to generate realistic interactions, but a novel ingredient of our work is to distill the material parameters of the object from pre-trained video generation models.", + "bbox": [ + 212, + 361, + 785, + 494 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "2.4 Video generation models", + "text_level": 1, + "bbox": [ + 214, + 518, + 465, + 532 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Recent progress in video generation is driven by the development of larger autoregressive [26,37,65,68] and diffusion models [3-6,22,24,25,59]. These models, trained on increasingly large datasets, continue to advance the quality and realism of generated video content. The state-of-the-art approach [6] can generate minute-long videos with realistic motions and viewpoint consistency. However, current video generation models cannot support physics-based interactions with objects through external forces.", + "bbox": [ + 212, + 542, + 784, + 648 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "3 Problem formulation", + "text_level": 1, + "bbox": [ + 214, + 672, + 454, + 688 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Given a static object represented by 3D Gaussians $\\{\\mathcal{G}_p\\}_{p=1}^P$ , $\\mathcal{G}_p = \\{\\boldsymbol{x}_p, \\alpha_p, \\boldsymbol{\\Sigma}_p, \\boldsymbol{c}_p\\}$ (where $\\boldsymbol{x}_p$ denotes the position, $\\alpha_p$ denotes the opacity, $\\boldsymbol{\\Sigma}_p$ denotes the covariance matrix, and $\\boldsymbol{c}_p$ denotes the color of the particle), our goal is to estimate physical material property fields for the object to enable realistic interactive motion synthesis. These properties include mass $m$ , Young's modulus $E$ , and Poisson's ratio $\\nu$ . Among these physical properties, Young's modulus $E$ plays a particularly important role in determining the object's motion in response to applied forces. Intuitively, Young's modulus (Eq. 2) measures the material stiffness. A higher Young's modulus results in less deformation and more rigid and", + "bbox": [ + 212, + 703, + 785, + 839 + ], + "page_idx": 3 + }, + { + "type": "page_number", + "text": "4", + "bbox": [ + 217, + 114, + 228, + 126 + ], + "page_idx": 3 + }, + { + "type": "header", + "text": "Zhang et al.", + "bbox": [ + 271, + 114, + 354, + 128 + ], + "page_idx": 3 + }, + { + "type": "image", + "img_path": "images/754050422b7712f0a1809b26eadd9beed8276563972be3ba00dd6bb4484f8f55.jpg", + "image_caption": [ + "3D Gaussians" + ], + "image_footnote": [], + "bbox": [ + 223, + 148, + 367, + 261 + ], + "page_idx": 4 + }, + { + "type": "image", + "img_path": "images/486db178e00153a3fd7a02d3245e4f4fd849c72197c224acfa0dd4cb79cec133.jpg", + "image_caption": [ + "10.30 seconds", + "Fig. 2: Effect of Young's modulus. We depict the motion of a simulated flower under the same external force but with three different Young's moduli, a measure of material stiffness. Flowers with the highest Young's modulus $(100\\times)$ exhibit smaller oscillations and higher frequencies, while the flower with the lowest Young's modulus $(1\\times)$ sways the most and oscillates at the lowest frequency. Time annotations below each image indicate the duration of one complete motion path shown in the figure." + ], + "image_footnote": [], + "bbox": [ + 397, + 150, + 516, + 261 + ], + "page_idx": 4 + }, + { + "type": "image", + "img_path": "images/bc8f8798a4d32ddbd5c4a007f43684398f8c57f00a2e5a1ec14997d6f95e0060.jpg", + "image_caption": [ + "2.30 seconds" + ], + "image_footnote": [], + "bbox": [ + 526, + 150, + 647, + 261 + ], + "page_idx": 4 + }, + { + "type": "image", + "img_path": "images/40aa849ef07be213f45b6ae9c0d6f96148607c9f8299a945c1be340b2750566c.jpg", + "image_caption": [ + "0.87 seconds" + ], + "image_footnote": [], + "bbox": [ + 656, + 150, + 779, + 261 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "higher-frequency motion, while a lower value leads to more flexible and elastic behavior. Fig. 2 illustrates the simulated motion of a flower under the same applied forces but with different Young's modulus.", + "bbox": [ + 212, + 393, + 784, + 441 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Therefore, our problem formulation focuses on estimating the spatially varying Young's modulus field $E(\\pmb{x})$ for the 3D object. To allow particle simulation, we query a particle's Young's modulus by $E_{p} = E(\\pmb{x}_{p})$ . As for other physical properties, the mass for a particle $m_{p}$ can be pre-computed as the product of a constant density $(\\rho)$ and particle volume $V_{p}$ . The particle volume can be estimated [70] by dividing a background cell's volume by the number of particles that cell contains. As for the Poisson's ratio $\\nu_{p}$ , we found that it has negligible impact on object motion in our preliminary experiments (see supplementary materials for details), and so we assume a homogeneous constant Poisson's ratio.", + "bbox": [ + 212, + 444, + 787, + 580 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "4 PhysDreamer", + "text_level": 1, + "bbox": [ + 215, + 622, + 383, + 640 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "PhysDreamer estimates a material field for a static 3D object. Our key idea is to generate a plausible video of the object in motion, and then optimize the material field $E(\\pmb{x})$ to match this synthesized motion. We begin by rendering a static image $(I_0)$ for the 3D scene $\\{\\mathcal{G}_p\\}$ from a certain viewpoint. We then leverage an image-to-video model to generate a short video clip $\\{I_0, I_1, \\dots, I_T\\}$ depicting the object's realistic motion. This generated video serves as our reference video. We then optimize the material field $E(\\pmb{x})$ and an initial velocity field $\\pmb{v}_0(\\pmb{x})$ (both modeled by implicit neural fields [71]) through differentiable simulation and differentiable rendering, such that a rendered video of the simulation matches (from the same viewpoint as $I_0$ ) the reference video. Fig. 3 shows an overview of PhysDreamer.", + "bbox": [ + 212, + 672, + 787, + 840 + ], + "page_idx": 4 + }, + { + "type": "header", + "text": "PhysDreamer", + "bbox": [ + 638, + 114, + 730, + 127 + ], + "page_idx": 4 + }, + { + "type": "page_number", + "text": "5", + "bbox": [ + 774, + 116, + 784, + 126 + ], + "page_idx": 4 + }, + { + "type": "image", + "img_path": "images/7d8c8cc15f188422fc1779b27228e9f37526320ae31318ca7f35803bebc1b59d.jpg", + "image_caption": [ + "Fig. 3: Overview of PhysDreamer. Given an object represented as 3D Gaussians, we first render it (with background) from a viewpoint. Next, we use an image-to-video generation model to produce a reference video of that object in motion. Using differentiable Material Point Methods (MPM) and differentiable rendering, we optimize both a spatially-varying material field and an initial velocity field (not shown in the figure above). This optimization aims to minimize the discrepancy between the rendered video and the reference video. The dashed arrows represent gradient flow." + ], + "image_footnote": [], + "bbox": [ + 222, + 143, + 781, + 354 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "4.1 Preliminaries", + "text_level": 1, + "bbox": [ + 215, + 484, + 372, + 498 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "3D Gaussians [34] adopts a set of anisotropic 3D Gaussian kernels to represent the radiance field of a 3D scene. Although introduced primarily as an efficient method for 3D novel view synthesis, the Lagrangian nature of 3D Gaussians also enables the direct adaptation of particle-based physics simulators. Following PhysGaussian [70], we use the Material Point Method (MPM) to simulate object dynamics directly on these Gaussian particles. Since 3D Gaussians mainly lie on object surfaces, an optional internal filling process can be applied for improved simulation realism [70]. Below, we provide a brief introduction on the underlying physical model and how to integrate MPM into 3D Gaussians. For a more comprehensive introduction of MPM, we refer interested readers to [27,32,33,70].", + "bbox": [ + 212, + 513, + 787, + 667 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Continuum mechanics and elastic materials. Continuum mechanics models material deformation using a map $\\phi$ that transforms points from the undeformed material space $\\mathbf{X}$ to the deformed world space $\\mathbf{x} = \\phi(\\mathbf{X}, t)$ . The Jacobian of the map, $\\pmb{F} = \\nabla_{\\mathbf{X}}\\phi(\\mathbf{X}, t)$ , known as the deformation gradient, measures local rotation and strain. This tensor is crucial in formulating stress-strain relationship. For example, the Cauchy stress in a hyper-elastic material is computed by: $\\pmb{\\sigma} = \\frac{1}{\\operatorname*{det}(\\pmb{F})}\\frac{\\partial\\psi}{\\partial\\pmb{F}}\\pmb{F}^T$ . Here, $\\psi(\\mathbf{F})$ represents the strain energy density function, quantifying the extent of non-rigid deformations. This function is typically designed by experts, to follow principles like material symmetry and rotational invariance while aligning with empirical data. In this work, we use fixed corotated hyperelastic model, whose energy density function can be expressed as:", + "bbox": [ + 212, + 672, + 787, + 840 + ], + "page_idx": 5 + }, + { + "type": "page_number", + "text": "6", + "bbox": [ + 217, + 114, + 228, + 126 + ], + "page_idx": 5 + }, + { + "type": "header", + "text": "Zhang et al.", + "bbox": [ + 271, + 114, + 354, + 128 + ], + "page_idx": 5 + }, + { + "type": "equation", + "text": "\n$$\n\\psi (\\mathbf {F}) = \\mu \\left(\\sum_ {i = 1} ^ {d} \\left(\\sigma_ {i} - 1\\right) ^ {2}\\right) + \\frac {\\lambda}{2} (\\det (\\mathbf {F}) - 1) ^ {2}, \\tag {1}\n$$\n", + "text_format": "latex", + "bbox": [ + 343, + 160, + 784, + 200 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "where $\\sigma_{i}$ denotes a singular value of the deformation gradient. $\\mu$ and $\\lambda$ are related to Young's modulus $E$ and Poisson's ratio $\\nu$ via:", + "bbox": [ + 215, + 205, + 782, + 236 + ], + "page_idx": 6 + }, + { + "type": "equation", + "text": "\n$$\n\\mu = \\frac {E}{2 (1 + \\nu)}, \\quad \\lambda = \\frac {E \\nu}{(1 + \\nu) (1 - 2 \\nu)}. \\tag {2}\n$$\n", + "text_format": "latex", + "bbox": [ + 367, + 250, + 782, + 280 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "The dynamics of an elastic object are governed by the following equations:", + "bbox": [ + 215, + 291, + 751, + 306 + ], + "page_idx": 6 + }, + { + "type": "equation", + "text": "\n$$\n\\rho \\frac {D \\boldsymbol {v}}{D t} = \\nabla \\cdot \\boldsymbol {\\sigma} + \\mathbf {f}, \\quad \\frac {D \\rho}{D t} + \\rho \\nabla \\cdot \\boldsymbol {v} = 0, \\tag {3}\n$$\n", + "text_format": "latex", + "bbox": [ + 364, + 318, + 782, + 345 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "where $\\rho$ denotes density, $\\pmb{v}(\\pmb{x},t)$ denotes the velocity field in world space, and $\\mathbf{f}$ denotes an external force.", + "bbox": [ + 215, + 357, + 782, + 386 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Material Point Method (MPM). We use the Moving Least Squares Material Point Method (MLS-MPM) [27] to solve the above governing equation. MPM is a hybrid Eulerian-Langrangian method widely adopted for simulating dynamics for a wide range of materials, such as solid, fluid, sand, and cloth [31, 35, 57, 62]. MPM offers several advantages, such as easy GPU parallelization [28], handling of topology changes, and the availability of well-documented open-source implementations [29, 48, 49, 70].", + "bbox": [ + 215, + 393, + 784, + 498 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Following PhysGaussian [70], we view the Gaussian particles as the spatial discretization of the object to be simulated, and directly run MPM on these Gaussian particles. Each particle $p$ represents a small volume of the object, and it carries a set of properties including volume $V_{p}$ , mass $m_{p}$ , position $\\boldsymbol{x}_{p}^{t}$ , velocity $\\boldsymbol{v}_{p}^{t}$ , deformation gradient $\\boldsymbol{F}_{p}^{t}$ , and local velocity field gradient $\\boldsymbol{C}_{p}^{t}$ at time step $t$ .", + "bbox": [ + 215, + 500, + 784, + 575 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "MPM operates in a particle-to-grid (P2G) and grid-to-particle (G2P) transfer loop. In the P2G stage, we transfer the momentum from particle to grid by:", + "bbox": [ + 215, + 575, + 782, + 604 + ], + "page_idx": 6 + }, + { + "type": "equation", + "text": "\n$$\nm _ {i} ^ {t} \\boldsymbol {v} _ {i} ^ {t} = \\sum_ {p} N (\\boldsymbol {x} _ {i} - \\boldsymbol {x} _ {p} ^ {t}) [ m _ {p} \\boldsymbol {v} _ {p} ^ {t} + (m _ {p} \\boldsymbol {C} _ {p} ^ {t} - \\frac {4}{(\\Delta x) ^ {2}} \\Delta t V _ {p} \\frac {\\partial \\psi}{\\partial \\boldsymbol {F}} \\boldsymbol {F} _ {p} ^ {t T}) (\\boldsymbol {x} _ {i} - \\boldsymbol {x} _ {p} ^ {t}) ] + \\boldsymbol {f} _ {i} ^ {t}, \\tag {4}\n$$\n", + "text_format": "latex", + "bbox": [ + 225, + 618, + 782, + 652 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "where the mass of the grid node $i$ is $m_i^t = \\sum_p N(\\pmb{x}_i - \\pmb{x}_p^t)m_p$ , $N(\\pmb{x}_i - \\pmb{x}_p^t)$ is the B-spline kernel, $\\Delta x$ is the spatial grid resolution, $\\Delta t$ is the simulation step size, and $\\pmb{v}_i^t$ is the updated velocity on the grid. We then transfer the updated velocity back to the particles and update their positions as:", + "bbox": [ + 215, + 665, + 784, + 726 + ], + "page_idx": 6 + }, + { + "type": "equation", + "text": "\n$$\n\\boldsymbol {v} _ {p} ^ {t + 1} = \\sum_ {i} N (\\boldsymbol {x} _ {i} - \\boldsymbol {x} _ {p} ^ {t}) \\boldsymbol {v} _ {i} ^ {t}, \\quad \\boldsymbol {x} _ {p} ^ {t + 1} = \\boldsymbol {x} _ {p} ^ {t} + \\Delta t \\boldsymbol {v} _ {p} ^ {t + 1}. \\tag {5}\n$$\n", + "text_format": "latex", + "bbox": [ + 326, + 739, + 782, + 768 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Meanwhile, the local velocity gradient and deformation gradient is updated as:", + "bbox": [ + 215, + 781, + 777, + 796 + ], + "page_idx": 6 + }, + { + "type": "equation", + "text": "\n$$\n\\boldsymbol {C} _ {p} ^ {t + 1} = \\frac {4}{(\\Delta x) ^ {2}} \\sum_ {i} N (\\boldsymbol {x} _ {i} - \\boldsymbol {x} _ {p} ^ {t}) \\boldsymbol {v} _ {i} ^ {t} (\\boldsymbol {x} _ {i} - \\boldsymbol {x} _ {p} ^ {t}) ^ {T}, \\quad \\boldsymbol {F} _ {p} ^ {t + 1} = (\\boldsymbol {I} + \\Delta t \\boldsymbol {C} _ {p} ^ {t + 1}) \\boldsymbol {F} _ {p} ^ {t}. \\tag {6}\n$$\n", + "text_format": "latex", + "bbox": [ + 235, + 808, + 782, + 842 + ], + "page_idx": 6 + }, + { + "type": "header", + "text": "PhysDreamer", + "bbox": [ + 638, + 114, + 730, + 127 + ], + "page_idx": 6 + }, + { + "type": "page_number", + "text": "7", + "bbox": [ + 774, + 116, + 785, + 126 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "4.2 Estimating physical properties", + "text_level": 1, + "bbox": [ + 215, + 146, + 513, + 162 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Using MLS-MPM [27, 70] as our physics simulator and the Fixed Corotated hyper-elastic material model for the 3D objects, the simulation process for a single sub-step is formalized as:", + "bbox": [ + 212, + 167, + 787, + 212 + ], + "page_idx": 7 + }, + { + "type": "equation", + "text": "\n$$\n\\boldsymbol {x} ^ {t + 1}, \\boldsymbol {v} ^ {t + 1}, \\boldsymbol {F} ^ {t + 1}, \\boldsymbol {C} ^ {t + 1} = \\mathcal {S} \\left(\\boldsymbol {x} ^ {t}, \\boldsymbol {v} ^ {t}, \\boldsymbol {F} ^ {t}, \\boldsymbol {C} ^ {t}, \\boldsymbol {\\theta}, \\Delta t\\right), \\tag {7}\n$$\n", + "text_format": "latex", + "bbox": [ + 326, + 218, + 785, + 234 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "where $\\pmb{x}^t = [\\pmb{x}_1^t, \\dots, \\pmb{x}_P^t]$ denotes the positions of all particles at time $t$ , and similarly $\\pmb{v}^t = [\\pmb{v}_1^t, \\dots, \\pmb{v}_P^t]$ denotes the velocities of all particles at time $t$ . $\\pmb{F}^t$ and $\\pmb{C}^t$ denote the deformation gradient and the gradient of local velocity fields for all particles, respectively. Both $\\pmb{F}^t$ and $\\pmb{C}^t$ are tracked for simulation purposes, not for rendering. $\\pmb{\\theta}$ denotes the collection of the physical properties of all particles: mass $\\pmb{m} = [m_1, \\dots, m_P]$ , Young's modulus $\\pmb{E} = [E_1, \\dots, E_P]$ , Poisson's ratio $\\pmb{\\nu} = [\\nu_1, \\dots, \\nu_P]$ , and volume $\\pmb{V} = [V_1, \\dots, V_P]$ . $\\Delta t$ is the simulation step size.", + "bbox": [ + 212, + 241, + 782, + 347 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "We use a sub-step size $\\Delta t \\cong 1 \\times 10^{-4}$ for most of our experiments. To simulate dynamics between adjacent video frames, we iterate over hundreds of sub-steps (time interval between frames are typically tens of milliseconds). For simplicity, we abuse notation to express a simulation step with $N$ sub-steps as:", + "bbox": [ + 212, + 347, + 782, + 407 + ], + "page_idx": 7 + }, + { + "type": "equation", + "text": "\n$$\n\\boldsymbol {x} ^ {t + 1}, \\boldsymbol {v} ^ {t + 1}, \\boldsymbol {F} ^ {t + 1}, \\boldsymbol {C} ^ {t + 1} = \\mathcal {S} (\\boldsymbol {x} ^ {t}, \\boldsymbol {v} ^ {t}, \\boldsymbol {F} ^ {t}, \\boldsymbol {C} ^ {t}, \\boldsymbol {\\theta}, \\Delta t, N), \\tag {8}\n$$\n", + "text_format": "latex", + "bbox": [ + 316, + 412, + 784, + 430 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "where the timestamp $t + 1$ is ahead of timestamp $t$ by $N\\Delta t$ . After simulation, we render the Gaussians at each frame:", + "bbox": [ + 214, + 436, + 782, + 465 + ], + "page_idx": 7 + }, + { + "type": "equation", + "text": "\n$$\n\\hat {I} ^ {t} = \\mathcal {F} _ {\\text {r e n d e r}} \\left(\\boldsymbol {x} ^ {t}, \\boldsymbol {\\alpha}, \\boldsymbol {R} ^ {t}, \\Sigma , \\boldsymbol {c}\\right), \\tag {9}\n$$\n", + "text_format": "latex", + "bbox": [ + 401, + 473, + 784, + 489 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "where $\\mathcal{F}_{\\mathrm{render}}$ denotes the differentiable rendering function, and $\\pmb{R}^{t}$ denotes the rotation matrices of all particles obtained from the simulation step.", + "bbox": [ + 212, + 494, + 782, + 525 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Using the generated video as reference, we optimize the spatially-varying Young's modulus $\\pmb{E}$ and an initial velocity $\\pmb{v}^0$ by a per-frame loss function:", + "bbox": [ + 214, + 526, + 782, + 555 + ], + "page_idx": 7 + }, + { + "type": "equation", + "text": "\n$$\nL ^ {t} = \\lambda L _ {1} \\left(\\hat {I} ^ {t}, I ^ {t}\\right) + (1 - \\lambda) L _ {\\mathrm {D} - \\mathrm {S S I M}} \\left(\\hat {I} ^ {t}, I ^ {t}\\right), \\tag {10}\n$$\n", + "text_format": "latex", + "bbox": [ + 352, + 561, + 784, + 579 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "where we set $\\lambda = 0.1$ in our experiments.", + "bbox": [ + 214, + 585, + 513, + 599 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "We parameterize the material field and velocity field by two triplanes [9], each followed by a three-layer MLP. Additionally, we apply a total variation regularization for all spatial planes of both fields to encourage spatial smoothness. Using $\\pmb{u}$ to denote one of the 2D spatial planes, and $\\pmb{u}_{i,j}$ as a feature vector on the 2D plane, we write the total variation regularization term as:", + "bbox": [ + 212, + 599, + 782, + 675 + ], + "page_idx": 7 + }, + { + "type": "equation", + "text": "\n$$\nL _ {\\mathrm {t v}} = \\sum_ {i, j} \\| \\boldsymbol {u} _ {i + 1, j} - \\boldsymbol {u} _ {i, j} \\| _ {2} ^ {2} + \\| \\boldsymbol {u} _ {i, j + 1} - \\boldsymbol {u} _ {i, j} \\| _ {2} ^ {2}. \\tag {11}\n$$\n", + "text_format": "latex", + "bbox": [ + 339, + 681, + 784, + 712 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Rather than optimizing the material parameters and initial velocity jointly, we split the optimization into two stages for better stability and faster convergence. In particular, in the first stage, we randomly initialize the Young's modulus for each Gaussian particle and freeze it. We optimize the initial velocity of each particle using only the first three frames of the reference video. In the second stage, we freeze the initial velocity and optimize the spatially varying Young's modulus. During the second stage, the gradient signal only flows to the previous frame to prevent gradient explosion/vanishing.", + "bbox": [ + 212, + 719, + 787, + 840 + ], + "page_idx": 7 + }, + { + "type": "page_number", + "text": "8", + "bbox": [ + 217, + 114, + 228, + 126 + ], + "page_idx": 7 + }, + { + "type": "header", + "text": "Zhang et al.", + "bbox": [ + 271, + 114, + 354, + 128 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/f27687cab82aa0f9bfca75ad70f1374e2435ca40c028a8b9f0bcd26cd9200ba0.jpg", + "image_caption": [ + "Fig.4: Accelerated MPM with K-Means downsampling. We employ K-Means clustering to create a set of \"driving particles\" (in yellow) at the initial time step $(t = 0)$ . We only simulate these driving particles. When rendering, we obtain each particle's position and rotation by fitting a local rigid body transformation using neighboring driving particles." + ], + "image_footnote": [], + "bbox": [ + 493, + 143, + 756, + 305 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "4.3 Accelerating simulation with subsampling", + "text_level": 1, + "bbox": [ + 215, + 321, + 606, + 338 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "High-fidelity rendering with 3D Gaussians typically requires millions of particles to represent a scene. Running simulations on all the particles poses a significant computational burden. To improve efficiency, we introduce a subsampling procedure for simulation, as illustrated in Fig. 4.", + "bbox": [ + 212, + 349, + 787, + 409 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "Specifically, we apply K-Means clustering to create a set of driving particles $\\{\\mathcal{Q}_q\\}_{q=1}^Q$ at $t = 0$ , where each driving particle is represented by $\\mathcal{Q}_q^0 = \\{\\pmb{x}_q^0, \\pmb{v}_q^0, \\pmb{F}_q^0, \\pmb{C}_q^0, E_q, m_q, \\nu_q, V_q\\}$ . The initial position of a driving particle $\\pmb{x}_q^0$ is computed as the mean of the position $\\pmb{x}_p$ of all cluster members. The number of the driving particles is much smaller than the number of 3D Gaussian particles, $Q \\ll P$ . We run simulations only on the driving particles. During rendering, we compute the position and rotation for each 3D Gaussian particle $\\mathcal{G}_p$ by interpolating the driving particles. In particular, for each 3D Gaussian particle, we find its eight nearest driving particles at $t = 0$ , and we fit a rigid body transformation $\\pmb{T}$ between these eight driving particles at $t = 0$ and at the current timestamp. This rigid body transformation $\\pmb{T}$ is applied to the initial position and rotation of the particle $\\mathcal{G}_p$ to obtain its current position and rotation. We summarize our algorithm with pseudo-code in supplementary materials.", + "bbox": [ + 215, + 410, + 787, + 608 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "5 Experiments", + "text_level": 1, + "bbox": [ + 215, + 632, + 375, + 648 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "5.1 Setup", + "text_level": 1, + "bbox": [ + 215, + 666, + 310, + 681 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "Datasets. We collect eight real-world static scenes by capturing multi-view images. Each scene includes an object and a background. The objects include five flowers (a red rose, a carnation, an orange rose, a tulip, and a white rose), an alocasia plant, a telephone cord, and a beanie hat. For each scene except for the red rose scene, we capture four interaction videos illustrating its natural motion after interaction, such as poking or dragging, and we use the real videos as additional comparison references.", + "bbox": [ + 212, + 698, + 787, + 803 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "Baselines. We compare our approach to two baselines: PhysGaussian [70] and DreamGaussian4D [58]. PhysGaussian [70] integrates MPM simulation to static", + "bbox": [ + 215, + 809, + 785, + 840 + ], + "page_idx": 8 + }, + { + "type": "header", + "text": "PhysDreamer", + "bbox": [ + 638, + 114, + 730, + 127 + ], + "page_idx": 8 + }, + { + "type": "page_number", + "text": "9", + "bbox": [ + 774, + 114, + 785, + 126 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "3D Gaussians to support simulation, but it cannot estimate material properties and relies on manually setting material parameter values. Thus, we use the same initialization strategy as ours to assign material properties for PhysGaussian. DreamGaussian4D [58] generates non-interactive dynamic 3D Gaussians from a static image. It first obtains a static 3D Gaussians using DreamGaussian [63], and then animate it by optimizing a deformation field from a generated driving video. For a fair comparison, we run its deformation field optimization on our reconstructed static 3D Gaussians, and we looped the resulting deformation field when rendering longer videos in later comparison.", + "bbox": [ + 212, + 146, + 787, + 282 + ], + "page_idx": 9 + }, + { + "type": "text", + "text": "Evaluation metrics. We focus on the quality of the synthesized object motion, in particular, visual quality and motion realism. Therefore, we conduct a user study and adopt the Two-alternative Forced Choice (2AFC) protocol: the participants are shown two side-by-side synchronized videos, including one video result from ours and the other one from the competitor's, with a random left-right ordering. The participants are then asked to choose the one with higher visual quality and the one with higher motion realism.", + "bbox": [ + 212, + 287, + 787, + 393 + ], + "page_idx": 9 + }, + { + "type": "text", + "text": "We recruited 100 participants, each asked to judge all 8 scenes, forming a total of 800 2AFC judgement samples for each baseline comparison. For each scene, we create 4 sample video pairs and show participants a random one from the 4 pairs. In particular, we create 4 five-second motion sequences using PhysDreamer with randomized initial conditions (applying an external force to the foreground object or assigning an initial velocity to the object), and render videos from randomly picked viewpoints. For the baseline method, we apply the same initial conditions (for PhysGaussian only) and render videos from the same viewpoint as ours to form the video pairs. Please see supplementary materials for human study details and quantitative metrics for videos (e.g., Fréchet Video Distance [64]).", + "bbox": [ + 212, + 393, + 787, + 546 + ], + "page_idx": 9 + }, + { + "type": "text", + "text": "5.2 Implementation details", + "text_level": 1, + "bbox": [ + 215, + 569, + 454, + 585 + ], + "page_idx": 9 + }, + { + "type": "text", + "text": "Neural material fields. We represent both material field and initial velocity field using triplanes [54] each followed by a three-layer MLP. The triplanes have spatial resolutions of $8^{3}$ and $24^{3}$ for the material field and velocity field, respectively.", + "bbox": [ + 212, + 601, + 785, + 647 + ], + "page_idx": 9 + }, + { + "type": "text", + "text": "3D Gaussian reconstruction. Similar to PhysGaussian [70], we employ anisotropic regularization to reduce skinny artifacts in the reconstruction. Each reconstructed scene contains 0.5 to 1.5 million particles (including foreground and background).", + "bbox": [ + 212, + 652, + 787, + 698 + ], + "page_idx": 9 + }, + { + "type": "text", + "text": "Simulation details. For computational efficiency, we segment the background and keep only foreground object particles for simulation. In our experiments, the foreground object contains around 50 to 300 thousand 3D Gaussian particles. We then discretize the foreground into a $64^3$ grid. The number of driving particles are 10 to 50 times fewer than the number of 3D Gaussian particles, determined by maintaining an average of at least eight particles per occupied voxel. For accurate motion, we use 768 sub-steps between successive video frames, corresponding to a duration of $4.34 \\times 10^{-5}$ second for each sub-step. To address the high memory consumption from large number of steps, we apply simulation", + "bbox": [ + 212, + 704, + 787, + 840 + ], + "page_idx": 9 + }, + { + "type": "page_number", + "text": "10", + "bbox": [ + 217, + 114, + 235, + 126 + ], + "page_idx": 9 + }, + { + "type": "header", + "text": "Zhang et al.", + "bbox": [ + 271, + 114, + 354, + 128 + ], + "page_idx": 9 + }, + { + "type": "image", + "img_path": "images/946f253765b04ae38e1dc2c49e75f817e629d35516977c74d58349625fbfbc08.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 218, + 145, + 330, + 229 + ], + "page_idx": 10 + }, + { + "type": "image", + "img_path": "images/053cf1b7da5d2e83dd5ca37013eec6bda5e658e0a3361f5515c3641af945165e.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 218, + 232, + 328, + 316 + ], + "page_idx": 10 + }, + { + "type": "image", + "img_path": "images/0a09853279d10b314b3874f37167548f2d2f7ffde004df071fe8e6a77f088dc6.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 220, + 318, + 328, + 402 + ], + "page_idx": 10 + }, + { + "type": "image", + "img_path": "images/c3f9ad2754052b701c080b77525609ebe5c885c54afd03d0e843a691542b4c83.jpg", + "image_caption": [ + "Material result" + ], + "image_footnote": [], + "bbox": [ + 220, + 404, + 328, + 488 + ], + "page_idx": 10 + }, + { + "type": "image", + "img_path": "images/f52f355d1637ff52888eca5721f7d66d5b296aef2268ff58c9cee5633c5a7268.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 341, + 145, + 452, + 229 + ], + "page_idx": 10 + }, + { + "type": "image", + "img_path": "images/0f1576aae8aca2b0343ca8dde702b468c47e2557a5d99e8495d763ab01172f63.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 341, + 231, + 452, + 315 + ], + "page_idx": 10 + }, + { + "type": "image", + "img_path": "images/4bab3e82c0c15ae265bb2563b393a2c4a915af1e842bb3c5c0e36a09d628fe71.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 341, + 318, + 452, + 402 + ], + "page_idx": 10 + }, + { + "type": "image", + "img_path": "images/a5fdebb31b0f2ed5b2589b21307e488600fa3e859eec56945a20e8bb7345e7bc.jpg", + "image_caption": [ + "Synthesized 3D dynamics under interaction" + ], + "image_footnote": [], + "bbox": [ + 341, + 404, + 452, + 488 + ], + "page_idx": 10 + }, + { + "type": "image", + "img_path": "images/6c9d4c85d30d4774ce78b6e056bb37c23ba99c8e1f3ae167608bc389055cc4fa.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 454, + 146, + 562, + 229 + ], + "page_idx": 10 + }, + { + "type": "image", + "img_path": "images/337d341034383ccdf26e3a24cade1c7cdad3ec139ade2bab18c182c0d78dc01c.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 454, + 231, + 562, + 315 + ], + "page_idx": 10 + }, + { + "type": "image", + "img_path": "images/5716a94ce533fb9308ccb354cc453be3ef37a13eb0d5e1e3279537dea0674cc7.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 454, + 318, + 562, + 402 + ], + "page_idx": 10 + }, + { + "type": "image", + "img_path": "images/b7b74ebd85c4e1adb9f53a30edfc3c1475bf373e7809fa9944717115d83a8b80.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 454, + 404, + 562, + 488 + ], + "page_idx": 10 + }, + { + "type": "image", + "img_path": "images/02dbaa6ebc3d11669c44674fe95a763c81c3c9e3d83fd92291fbfec7da0929dd.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 563, + 146, + 671, + 229 + ], + "page_idx": 10 + }, + { + "type": "image", + "img_path": "images/a06f62dd64ef27937347af8c8bf97faead16694ec747269be5817bb9a20b0347.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 563, + 231, + 671, + 315 + ], + "page_idx": 10 + }, + { + "type": "image", + "img_path": "images/8d4efb73ec74dbb4774f57ff85479ac556475ea417956382742968b89b8a4675.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 563, + 318, + 671, + 402 + ], + "page_idx": 10 + }, + { + "type": "image", + "img_path": "images/43ca094db8f382473e70d106b35d711f79e879a8f18d9341048d8bd6853880cc.jpg", + "image_caption": [ + "Fig. 5: Interactive 3D dynamics synthesis. (Left) Visualization of the material fields. Brighter color indicates higher Young's modulus within each example. (Right) We apply an external force (red arrow) on each object, and the following columns demonstrate the object dynamics rendered at a static viewpoint." + ], + "image_footnote": [], + "bbox": [ + 563, + 404, + 671, + 488 + ], + "page_idx": 10 + }, + { + "type": "image", + "img_path": "images/7e3125339de5c304c095b6dc7a7bf7109f8a2c0fe45c2161aaeccf906496d144.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 674, + 146, + 782, + 229 + ], + "page_idx": 10 + }, + { + "type": "image", + "img_path": "images/5f10fed9c66aad52ab6aa0224ec2db785d0a30f116f0fabcfde14aa3f884664d.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 674, + 231, + 782, + 315 + ], + "page_idx": 10 + }, + { + "type": "image", + "img_path": "images/311a2cc51be66ebcac18b5ec388b973555d744230d954b2b00a8227b2024959c.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 674, + 318, + 782, + 402 + ], + "page_idx": 10 + }, + { + "type": "image", + "img_path": "images/3c4a3fe009af3da3776db34998d3dc8c17fa89fc447387c6d734ca5f7d9a4656.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 674, + 404, + 782, + 488 + ], + "page_idx": 10 + }, + { + "type": "text", + "text": "state checkpointing and re-computation during gradient back-propagation. We add Dirichlet boundary conditions for stationary grid cells. We fill the internal volumes of certain solid objects to enhance simulation realism [70].", + "bbox": [ + 212, + 588, + 787, + 633 + ], + "page_idx": 10 + }, + { + "type": "text", + "text": "Generating reference videos. We render a 3D object with its background from a viewpoint, and then we use Stable Video Diffusion [4] to animate this rendered image and generate fourteen video frames. We use a small motion bucket number [4] (e.g., 5 or 8) so that the generated video contains mostly object motion and little camera motion. We use rendered images for the video generation, so that our approach can also be used for generated scenes. Also, rendering images directly from 3D Gaussians simplifies later optimization.", + "bbox": [ + 212, + 638, + 787, + 744 + ], + "page_idx": 10 + }, + { + "type": "text", + "text": "5.3 Results", + "text_level": 1, + "bbox": [ + 214, + 753, + 323, + 768 + ], + "page_idx": 10 + }, + { + "type": "text", + "text": "We show our qualitative results of the spatially-varying Young's modulus in Fig. 5 (left), and simulated interactive motion in Fig. 5 (right). Please see our project website videos for a better motion visualization. Tab. 1 presents the user study results in comparison to baseline methods and real captured videos.", + "bbox": [ + 212, + 779, + 787, + 840 + ], + "page_idx": 10 + }, + { + "type": "header", + "text": "PhysDreamer", + "bbox": [ + 638, + 114, + 730, + 127 + ], + "page_idx": 10 + }, + { + "type": "page_number", + "text": "11", + "bbox": [ + 767, + 114, + 784, + 126 + ], + "page_idx": 10 + }, + { + "type": "table", + "img_path": "images/0b4d8778a0a85d5140c8960b026b50b042927ec47a5a1ab956898eb4abfa08a5.jpg", + "table_caption": [ + "Table 1: Human study 2AFC results of PhysDreamer (Ours) over real captured videos and baseline methods (PhysGaussian [70] and DreamGaussian4D [58]) on Motion Realism and overall Visual Quality. \"Rose O\", \"Rose W\", and \"Rose R\" denotes the orange, white, and red roses, respectively." + ], + "table_footnote": [], + "table_body": "
Motion realismAlocasiaCarnationHatRose ORose WRose RCordTulipAvg.
Ours over Real capture86%61%55%63%47%-29%35%53.7%
Ours over PhysGaussian96%89%57%91%93%73%61%86%80.8%
Ours over DreamGaussian75%77%51%78%51%41%71%64%63.5%
Visual quality
Ours over Real capture36%53%28%40%41%-29%34%37.3%
Ours over PhysGaussian67%69%50%75%73%58%58%70%65.0%
Ours over DreamGaussian82%75%74%76%60%47%76%70%70.0%
", + "bbox": [ + 225, + 212, + 779, + 366 + ], + "page_idx": 11 + }, + { + "type": "text", + "text": "Compared to PhysGaussian, $80.8\\%$ of the human participant 2AFC samples prefer PhysDreamer (ours) in motion realism and $65.0\\%$ prefer PhysDreamer in visual quality. Note that since the static scenes are the same, the visual quality also depends on the generated object motion. Fig. 6 shows temporal slices of the motion patterns. We observe that PhysGaussian produces large, unrealistic slow motion due to the lack of a principled estimation of material properties.", + "bbox": [ + 212, + 393, + 785, + 484 + ], + "page_idx": 11 + }, + { + "type": "text", + "text": "Compared to DreamGaussian4D, $70.0\\% / 63.5\\%$ 2AFC samples prefer ours in visual quality and motion realism, respectively. From Fig. 6, we can observe that DreamGaussian4D generates periodic motion with a constant, small magnitude, while PhysDreamer can simulate the damping in motion. This is because DreamGaussian4D does not simulate the physical dynamics but simply distill a motion sequence from a generative model, so it cannot extrapolate to different motion. We further include one more evaluation dimension on \"motion amount\" comparing to DreamGaussian4D, where we ask the participants to judge which video has higher amount of motion, and $73.6\\%$ 2AFC samples prefer PhysDreamer.", + "bbox": [ + 212, + 484, + 785, + 621 + ], + "page_idx": 11 + }, + { + "type": "text", + "text": "Compared to real videos, $53.7\\%$ 2AFC samples favored the motion realism of ours results. Interestingly, under \"Motion Realism\", $86\\%$ of the users indicated that the alocasia outputs were more realistic than real captures. This is surprising, as one would expect a $50\\%$ preference if the videos were indistinguishable. We offer a potential explanation: for thin geometries like alocasia leaves, the Material Point Method tends to produce lower-frequency and slower motions. This can be observed in the video and is evident in the space-time slice visualizations in Fig. 6. Humans are poor at judging the naturalness of motion and may be biased towards smoother and slower motions, as shown in prior studies [36,61].", + "bbox": [ + 212, + 622, + 787, + 758 + ], + "page_idx": 11 + }, + { + "type": "text", + "text": "5.4 Ablation: using multi-view reference videos", + "text_level": 1, + "bbox": [ + 215, + 782, + 619, + 797 + ], + "page_idx": 11 + }, + { + "type": "text", + "text": "For objects with self-occlusion, observing salient motion of all object parts from a single video is challenging (e.g., the alocasia scene where a leaf can occlude", + "bbox": [ + 212, + 809, + 785, + 839 + ], + "page_idx": 11 + }, + { + "type": "page_number", + "text": "12", + "bbox": [ + 217, + 114, + 235, + 126 + ], + "page_idx": 11 + }, + { + "type": "header", + "text": "Zhang et al.", + "bbox": [ + 271, + 114, + 354, + 128 + ], + "page_idx": 11 + }, + { + "type": "image", + "img_path": "images/8417644f2a8bb83365516b07e17e7b2305058377a2f517a2871d2912989ce5cd.jpg", + "image_caption": [ + "Fig. 6: We compare our results with real captured videos, PhysGaussian [70], and DreamGaussian4D [58] using space-time slices. In these slices, the vertical axis represent time, and the horizontal axis shows a spatial slice of the object (denoted by red lines on the \"object\" column). These slices visualize the magnitude and frequencies of these oscillating motions. Results for our PhysDreamer (Ours) and PhysGaussian are simulated with the same initial conditions." + ], + "image_footnote": [], + "bbox": [ + 217, + 143, + 784, + 724 + ], + "page_idx": 12 + }, + { + "type": "header", + "text": "PhysDreamer", + "bbox": [ + 638, + 114, + 730, + 128 + ], + "page_idx": 12 + }, + { + "type": "page_number", + "text": "13", + "bbox": [ + 767, + 114, + 785, + 126 + ], + "page_idx": 12 + }, + { + "type": "image", + "img_path": "images/c0c3bf17c0e32228530761b384ebe146d83190f582b74d2036ea615c1794193c.jpg", + "image_caption": [ + "Fig. 7: Comparison between single-view (top) and two-view (bottom) supervisions. The object (alocasia) exhibits self-occluding structures. We can use generated videos at two views to jointly optimize the material field. In the space-time (X-t) slices, the vertical axis represents time, and the horizontal axis shows a spatial slice of the object." + ], + "image_footnote": [], + "bbox": [ + 217, + 143, + 784, + 281 + ], + "page_idx": 13 + }, + { + "type": "text", + "text": "another leaf). We may alleviate this problem by rendering from multiple viewpoints to provide comprehensive coverage of the object. Here, we use multiple videos in the material estimation, jointly optimizing a video-agnostic, spatially-varying Young's modulus for each particle along with video-specific initial velocities. From the comparison of the alocasia scene in Fig. 7, we can see that using multi-view reference videos (a front view and a back view) helps in such complex self-occluding objects: PhysDreamer benefits significantly from having supervision from two views, while using only a single view leads to artifacts. In our user study, $81.0\\%$ 2AFC samples prefer results with two view supervision in visual quality and $86.0\\%$ in motion realism.", + "bbox": [ + 212, + 375, + 787, + 525 + ], + "page_idx": 13 + }, + { + "type": "text", + "text": "6 Conclusion", + "text_level": 1, + "bbox": [ + 215, + 554, + 359, + 569 + ], + "page_idx": 13 + }, + { + "type": "text", + "text": "In this work, we introduced PhysDreamer, a novel approach to synthesizing interactive 3D dynamics by endowing static 3D objects with physical material properties. Our method distills the object dynamics priors learned by video generation models to estimate the spatially-varying material properties. We showcase dynamics interaction with a diverse set of elastic objects by PhysDreamer. We believe that PhysDreamer takes a significant step towards creating more engaging and immersive virtual environments, opening up a wide range of applications from realistic simulations to interactive virtual experiences.", + "bbox": [ + 212, + 590, + 787, + 712 + ], + "page_idx": 13 + }, + { + "type": "text", + "text": "Limitations. Our approach requires the user to manually specify the object to simulate and separate it from the background, and establish boundary conditions for stationary parts, like the pot of flowers. 3D object discovery may help for simulatable object extraction. In addition, our approach is computationally demanding. Despite our subsampling strategy, our current algorithm takes approximately one minute on a NVIDIA V100 GPU to produce a single second of video. Further improving efficiency remains an important future problem. Finally, in this work, we restrict our scope to elastic objects without collisions.", + "bbox": [ + 212, + 719, + 787, + 839 + ], + "page_idx": 13 + }, + { + "type": "page_number", + "text": "14", + "bbox": [ + 217, + 114, + 235, + 126 + ], + "page_idx": 13 + }, + { + "type": "header", + "text": "Zhang et al.", + "bbox": [ + 271, + 114, + 354, + 128 + ], + "page_idx": 13 + }, + { + "type": "text", + "text": "Acknowledgements. This work is in part supported by the NSF PHY-2019786 (The NSF AI Institute for Artificial Intelligence and Fundamental Interactions, http://iaifi.org/), NSF CIF 1955864 (Occlusion and Directional Resolution in Computational Imaging), RI #2211258, #2338203, ONR MURI N00014-22-1-2740, Quanta Computer, Samsung, and United States Air Force Artificial Intelligence Accelerator and was accomplished under Cooperative Agreement Number FA8750-192-1000. We would like to thank Peter Yichen Chen, Zhengqi Li, Pingchuan Ma, Minghao Guo, Ge Yang, and Shai Avidan for help and insightful discussions.", + "bbox": [ + 212, + 146, + 787, + 282 + ], + "page_idx": 14 + }, + { + "type": "text", + "text": "References", + "text_level": 1, + "bbox": [ + 215, + 310, + 323, + 325 + ], + "page_idx": 14 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "1. Attal, B., Huang, J.B., Richardt, C., Zollhoefer, M., Kopf, J., O'Toole, M., Kim, C.: Hyperreel: High-fidelity 6-dof video with ray-conditioned sampling. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 16610-16620 (2023)", + "2. Bahmani, S., Skorokhodov, I., Rong, V., Wetzstein, G., Guibas, L., Wonka, P., Tulyakov, S., Park, J.J., Tagliasacchi, A., Lindell, D.B.: 4d-fy: Text-to-4d generation using hybrid score distillation sampling. arXiv preprint arXiv:2311.17984 (2023)", + "3. Bar-Tal, O., Chefer, H., Tov, O., Herrmann, C., Paiss, R., Zada, S., Ephrat, A., Hur, J., Li, Y., Michaeli, T., et al.: Lumiere: A space-time diffusion model for video generation. arXiv preprint arXiv:2401.12945 (2024)", + "4. Blattmann, A., Dockhorn, T., Kulal, S., Mendelevitch, D., Kilian, M., Lorenz, D., Levi, Y., English, Z., Voleti, V., Letts, A., et al.: Stable video diffusion: Scaling latent video diffusion models to large datasets. arXiv preprint arXiv:2311.15127 (2023)", + "5. Blattmann, A., Rombach, R., Ling, H., Dockhorn, T., Kim, S.W., Fidler, S., Kreis, K.: Align your latents: High-resolution video synthesis with latent diffusion models. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 22563-22575 (2023)", + "6. Brooks, T., Peebles, B., Homes, C., DePue, W., Guo, Y., Jing, L., Schnurr, D., Taylor, J., Luhman, T., Luhman, E., Ng, C., Wang, R., Ramesh, A.: Video generation models as world simulators (2024), https://openai.com/research/video-generation-models-as-world-simulators", + "7. Cai, Y., Wang, J., Yuille, A., Zhou, Z., Wang, A.: Structure-aware sparse-view x-ray 3d reconstruction. In: CVPR (2024)", + "8. Cao, A., Johnson, J.: Hexplane: A fast representation for dynamic scenes. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 130-141 (2023)", + "9. Chan, E.R., Lin, C.Z., Chan, M.A., Nagano, K., Pan, B., De Mello, S., Gallo, O., Guibas, L.J., Tremblay, J., Khamis, S., et al.: Efficient geometry-aware 3d generative adversarial networks. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 16123-16133 (2022)", + "10. Chen, H.y., Tretschk, E., Stuyck, T., Kadlecek, P., Kavan, L., Vouga, E., Lassner, C.: Virtual elastic objects. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 15827-15837 (2022)" + ], + "bbox": [ + 217, + 345, + 785, + 840 + ], + "page_idx": 14 + }, + { + "type": "header", + "text": "PhysDreamer", + "bbox": [ + 638, + 114, + 730, + 127 + ], + "page_idx": 14 + }, + { + "type": "page_number", + "text": "15", + "bbox": [ + 767, + 114, + 785, + 126 + ], + "page_idx": 14 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "11. Chen, X., Liu, Z., Chen, M., Feng, Y., Liu, Y., Shen, Y., Zhao, H.: Livephoto: Real image animation with text-guided motion control. arXiv preprint arXiv:2312.02928 (2023)", + "12. Chuang, Y.Y., Goldman, D.B., Zheng, K.C., Curless, B., Salesin, D.H., Szeliski, R.: Animating pictures with stochastic motion textures. In: ACM SIGGRAPH 2005 Papers. pp. 853-860 (2005)", + "13. Curless, B., Levoy, M.: A volumetric method for building complex models from range images. In: Proceedings of the 23rd annual conference on Computer graphics and interactive techniques. pp. 303-312 (1996)", + "14. Dai, Z., Zhang, Z., Yao, Y., Qiu, B., Zhu, S., Qin, L., Wang, W.: Animateanything: Fine-grained open domain image animation with motion guidance. arXiv e-prints pp. arXiv-2311 (2023)", + "15. Davis, A., Chen, J.G., Durand, F.: Image-space modal bases for plausible manipulation of objects in video. ACM Transactions on Graphics (TOG) 34(6), 1-7 (2015)", + "16. Davis, M.A.: Visual vibration analysis. Ph.D. thesis, Massachusetts Institute of Technology (2016)", + "17. Duan, Y., Wei, F., Dai, Q., He, Y., Chen, W., Chen, B.: 4d gaussian splatt-ting: Towards efficient novel view synthesis for dynamic scenes. arXiv preprint arXiv:2402.03307 (2024)", + "18. Feng, Y., Shang, Y., Li, X., Shao, T., Jiang, C., Yang, Y.: Pie-nerf: Physics-based interactive elastodynamics with nerf. arXiv preprint arXiv:2311.13099 (2023)", + "19. Fridovich-Keil, S., Meanti, G., Warburg, F.R., Recht, B., Kanazawa, A.: K-planes: Explicit radiance fields in space, time, and appearance. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 12479-12488 (2023)", + "20. Gao, H., Li, R., Tulsiani, S., Russell, B., Kanazawa, A.: Monocular dynamic view synthesis: A reality check. Advances in Neural Information Processing Systems 35, 33768-33780 (2022)", + "21. Geng, D., Owens, A.: Motion guidance: Diffusion-based image editing with differentiable motion estimators. In: The Twelfth International Conference on Learning Representations (2023)", + "22. Girdhar, R., Singh, M., Brown, A., Duval, Q., Azadi, S., Rambhatla, S.S., Shah, A., Yin, X., Parikh, D., Misra, I.: Emu video: Factorizing text-to-video generation by explicit image conditioning. arXiv preprint arXiv:2311.10709 (2023)", + "23. Guo, X., Sun, J., Dai, Y., Chen, G., Ye, X., Tan, X., Ding, E., Zhang, Y., Wang, J.: Forward flow for novel view synthesis of dynamic scenes. In: Proceedings of the IEEE/CVF International Conference on Computer Vision. pp. 16022-16033 (2023)", + "24. Gupta, A., Yu, L., Sohn, K., Gu, X., Hahn, M., Fei-Fei, L., Essa, I., Jiang, L., Lezama, J.: Photorealistic video generation with diffusion models. arXiv preprint arXiv:2312.06662 (2023)", + "25. Ho, J., Chan, W., Sahara, C., Whang, J., Gao, R., Gritsenko, A., Kingma, D.P., Poole, B., Norouzi, M., Fleet, D.J., et al.: Imagen video: High definition video generation with diffusion models. arXiv preprint arXiv:2210.02303 (2022)", + "26. Hong, W., Ding, M., Zheng, W., Liu, X., Tang, J.: Cogvideo: Large-scale pretraining for text-to-video generation via transformers. arXiv preprint arXiv:2205.15868 (2022)", + "27. Hu, Y., Fang, Y., Ge, Z., Qu, Z., Zhu, Y., Pradhana, A., Jiang, C.: A moving least squares material point method with displacement discontinuity and two-way rigid body coupling. ACM Transactions on Graphics (TOG) 37(4), 1-14 (2018)" + ], + "bbox": [ + 215, + 146, + 785, + 840 + ], + "page_idx": 15 + }, + { + "type": "page_number", + "text": "16", + "bbox": [ + 217, + 114, + 235, + 126 + ], + "page_idx": 15 + }, + { + "type": "header", + "text": "Zhang et al.", + "bbox": [ + 271, + 114, + 352, + 128 + ], + "page_idx": 15 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "28. Hu, Y., Li, T.M., Anderson, L., Ragan-Kelley, J., Durand, F.: Taichi: a language for high-performance computation on spatially sparse data structures. ACM Transactions on Graphics (TOG) 38(6), 1-16 (2019)", + "29. Hu, Y., Li, T.M., Anderson, L., Ragan-Kelley, J., Durand, F.: Taichi: a language for high-performance computation on spatially sparse data structures. ACM Transactions on Graphics (TOG) 38(6), 1-16 (2019)", + "30. Huang, Y.H., Sun, Y.T., Yang, Z., Lyu, X., Cao, Y.P., Qi, X.: Sc-gs: Sparse-controlled gaussian splatting for editable dynamic scenes. arXiv preprint arXiv:2312.14937 (2023)", + "31. Jiang, C., Gast, T., Teran, J.: Anisotropic elastoplasticity for cloth, knit and hair frictional contact. ACM Transactions on Graphics (TOG) 36(4), 1-14 (2017)", + "32. Jiang, C., Schroeder, C., Selle, A., Teran, J., Stomakhin, A.: The affine particle-in-cell method. ACM Transactions on Graphics (TOG) 34(4), 1-10 (2015)", + "33. Jiang, C., Schroeder, C., Teran, J., Stomakhin, A., Selle, A.: The material point method for simulating continuum materials. In: ACM SIGGRAPH 2016 courses. pp. 1-52 (2016)", + "34. Kerbl, B., Kopanas, G., Leimkuhler, T., Drettakis, G.: 3d gaussian splatting for real-time radiance field rendering. ACM Transactions on Graphics 42(4) (2023)", + "35. Klár, G., Gast, T., Pradhana, A., Fu, C., Schroeder, C., Jiang, C., Teran, J.: Drucker-prager elastoplasticity for sand animation. ACM Transactions on Graphics (TOG) 35(4), 1-12 (2016)", + "36. Kobayashi, M., Motoyoshi, I.: Perceiving natural speed in natural movies. i-Perception 10(4), 2041669519860544 (2019)", + "37. Kondratyuk, D., Yu, L., Gu, X., Lezama, J., Huang, J., Hornung, R., Adam, H., Akbari, H., Alon, Y., Birodkar, V., et al.: Videopoet: A large language model for zero-shot video generation. arXiv preprint arXiv:2312.14125 (2023)", + "38. Kopf, J., Rong, X., Huang, J.B.: Robust consistent video depth estimation. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 1611-1621 (2021)", + "39. Kratimenos, A., Lei, J., Daniilidis, K.: Dynmf: Neural motion factorization for real-time dynamic view synthesis with 3d gaussian splatting. arXiv preprint arXiv:2312.00112 (2023)", + "40. Le Cleac'h, S., Yu, H.X., Guo, M., Howell, T., Gao, R., Wu, J., Manchester, Z., Schwager, M.: Differentiable physics simulation of dynamics-augmented neural objects. IEEE Robotics and Automation Letters (2023)", + "41. Li, H., Sumner, R.W., Pauly, M.: Global correspondence optimization for non-rigid registration of depth scans. In: Computer graphics forum. vol. 27, pp. 1421-1430. Wiley Online Library (2008)", + "42. Li, X., Qiao, Y.L., Chen, P.Y., Jatavallabhula, K.M., Lin, M., Jiang, C., Gan, C.: Pac-nerf: Physics augmented continuum neural radiance fields for geometry-agnostic system identification. arXiv preprint arXiv:2303.05512 (2023)", + "43. Li, Z., Niklaus, S., Snavely, N., Wang, O.: Neural scene flow fields for space-time view synthesis of dynamic scenes. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 6498-6508 (2021)", + "44. Li, Z., Tucker, R., Snavely, N., Holynski, A.: Generative image dynamics. arXiv preprint arXiv:2309.07906 (2023)", + "45. Li, Z., Wang, Q., Cole, F., Tucker, R., Snavely, N.: Dynibar: Neural dynamic image-based rendering. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 4273-4284 (2023)" + ], + "bbox": [ + 212, + 146, + 787, + 840 + ], + "page_idx": 16 + }, + { + "type": "header", + "text": "PhysDreamer", + "bbox": [ + 638, + 114, + 730, + 127 + ], + "page_idx": 16 + }, + { + "type": "page_number", + "text": "17", + "bbox": [ + 767, + 114, + 785, + 126 + ], + "page_idx": 16 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "46. Ling, H., Kim, S.W., Torralba, A., Fidler, S., Kreis, K.: Align your gaussians: Text-to-4d with dynamic 3d gaussians and composed diffusion models. arXiv preprint arXiv:2312.13763 (2023)", + "47. Luiten, J., Kopanas, G., Leibe, B., Ramanan, D.: Dynamic 3d gaussians: Tracking by persistent dynamic view synthesis. arXiv preprint arXiv:2308.09713 (2023)", + "48. Ma, P., Chen, P.Y., Deng, B., Tenenbaum, J.B., Du, T., Gan, C., Matusik, W.: Learning neural constitutive laws from motion observations for generalizable pde dynamics. In: International Conference on Machine Learning. PMLR (2023)", + "49. Macklin, M.: Warp: A high-performance python framework forgpu simulation and graphics. https://github.com/nvidia/warp (March 2022), nVIDIA GPU Technology Conference (GTC)", + "50. Mildenhall, B., Srinivasan, P.P., Tancik, M., Barron, J.T., Ramamoorthi, R., Ng, R.: Nerf: Representing scenes as neural radiance fields for view synthesis. In: ECCV (2020)", + "51. Newcombe, R.A., Fox, D., Seitz, S.M.: Dynamicfusion: Reconstruction and tracking of non-rigid scenes in real-time. In: Proceedings of the IEEE conference on computer vision and pattern recognition. pp. 343-352 (2015)", + "52. Park, K., Sinha, U., Barron, J.T., Bouaziz, S., Goldman, D.B., Seitz, S.M., Martin-Brualla, R.: Nerfies: Deformable neural radiance fields. In: Proceedings of the IEEE/CVF International Conference on Computer Vision. pp. 5865-5874 (2021)", + "53. Park, K., Sinha, U., Hedman, P., Barron, J.T., Bouaziz, S., Goldman, D.B., Martin-Brualla, R., Seitz, S.M.: Hypernerf: A higher-dimensional representation for topologically varying neural radiance fields. arXiv preprint arXiv:2106.13228 (2021)", + "54. Peng, S., Niemeyer, M., Mescheder, L., Pollefeys, M., Geiger, A.: Convolutional occupancy networks. In: Computer Vision-ECCV 2020: 16th European Conference, Glasgow, UK, August 23-28, 2020, Proceedings, Part III 16. pp. 523-540. Springer (2020)", + "55. Poole, B., Jain, A., Barron, J.T., Mildenhall, B.: Dreamfusion: Text-to-3d using 2d diffusion. In: The Eleventh International Conference on Learning Representations (2022)", + "56. Pumarola, A., Corona, E., Pons-Moll, G., Moreno-Noguer, F.: D-nerf: Neural radiance fields for dynamic scenes. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 10318-10327 (2021)", + "57. Ram, D., Gast, T., Jiang, C., Schroeder, C., Stomakhin, A., Teran, J., Kavehpour, P.: A material point method for viscoelastic fluids, foams and sponges. In: Proceedings of the 14th ACM SIGGRAPH/Eurographics Symposium on Computer Animation. pp. 157-163 (2015)", + "58. Ren, J., Pan, L., Tang, J., Zhang, C., Cao, A., Zeng, G., Liu, Z.: Dreamgaussian4d: Generative 4d gaussian splatting. arXiv preprint arXiv:2312.17142 (2023)", + "59. Singer, U., Polyak, A., Hayes, T., Yin, X., An, J., Zhang, S., Hu, Q., Yang, H., Ashual, O., Gafni, O., et al.: Make-a-video: Text-to-video generation without text-video data. arXiv preprint arXiv:2209.14792 (2022)", + "60. Singer, U., Sheynin, S., Polyak, A., Ashual, O., Makarov, I., Kokkinos, F., Goyal, N., Vedaldi, A., Parikh, D., Johnson, J., et al.: Text-to-4d dynamic scene generation. arXiv preprint arXiv:2301.11280 (2023)", + "61. Stocker, A.A., Simoncelli, E.P.: Noise characteristics and prior expectations in human visual speed perception. Nature neuroscience 9(4), 578-585 (2006)", + "62. Stomakhin, A., Schroeder, C., Chai, L., Teran, J., Selle, A.: A material point method for snow simulation. ACM Transactions on Graphics (TOG) 32(4), 1-10 (2013)" + ], + "bbox": [ + 215, + 146, + 784, + 840 + ], + "page_idx": 17 + }, + { + "type": "page_number", + "text": "18", + "bbox": [ + 217, + 114, + 235, + 126 + ], + "page_idx": 17 + }, + { + "type": "header", + "text": "Zhang et al.", + "bbox": [ + 271, + 114, + 352, + 128 + ], + "page_idx": 17 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "63. Tang, J., Ren, J., Zhou, H., Liu, Z., Zeng, G.: Dreamgaussian: Generative gaussian splatting for efficient 3d content creation. arXiv preprint arXiv:2309.16653 (2023)", + "64. Unterthiner, T., Van Steenkiste, S., Kurach, K., Marinier, R., Michalski, M., Gelly, S.: Towards accurate generative models of video: A new metric & challenges. arXiv preprint arXiv:1812.01717 (2018)", + "65. Villegas, R., Babaeizadeh, M., Kindermans, P.J., Moraldo, H., Zhang, H., Saffar, M.T., Castro, S., Kunze, J., Erhan, D.: Phenaki: Variable length video generation from open domain textual descriptions. In: International Conference on Learning Representations (2022)", + "66. Wang, C., MacDonald, L.E., Jeni, L.A., Lucey, S.: Flow supervision for deformable nerf. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 21128-21137 (2023)", + "67. Wang, C., Zhuang, P., Siarohin, A., Cao, J., Qian, G., Lee, H.Y., Tulyakov, S.: Diffusion priors for dynamic view synthesis from monocular videos. arXiv preprint arXiv:2401.05583 (2024)", + "68. Wu, C., Liang, J., Ji, L., Yang, F., Fang, Y., Jiang, D., Duan, N.: Niwa: Visual synthesis pre-training for neural visual world creation. In: European conference on computer vision. pp. 720-736. Springer (2022)", + "69. Xian, W., Huang, J.B., Kopf, J., Kim, C.: Space-time neural irradiance fields for free-viewpoint video. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 9421-9431 (2021)", + "70. Xie, T., Zong, Z., Qiu, Y., Li, X., Feng, Y., Yang, Y., Jiang, C.: Physgaus-sian: Physics-integrated 3d gaussians for generative dynamics. arXiv preprint arXiv:2311.12198 (2023)", + "71. Xie, Y., Takikawa, T., Saito, S., Litany, O., Yan, S., Khan, N., Tombari, F., Tompkin, J., Sitzmann, V., Sridhar, S.: Neural fields in visual computing and beyond. In: Computer Graphics Forum. vol. 41, pp. 641-676. Wiley Online Library (2022)", + "72. Yang, Z., Gao, X., Zhou, W., Jiao, S., Zhang, Y., Jin, X.: Deformable 3d gaussians for high-fidelity monocular dynamic scene reconstruction. arXiv preprint arXiv:2309.13101 (2023)", + "73. Yin, Y., Xu, D., Wang, Z., Zhao, Y., Wei, Y.: 4dgen: Grounded 4d content generation with spatial-temporal consistency. arXiv preprint arXiv:2312.17225 (2023)", + "74. Yu, H., Julin, J., Milacski, Z.A., Niinuma, K., Jeni, L.A.: Cogs: Controllable gaussian splatting. arXiv preprint arXiv:2312.05664 (2023)", + "75. Yu, H., Julin, J., Milacski, Z.A., Niinuma, K., Jeni, L.A.: Dylin: Making light field networks dynamic. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 12397-12406 (2023)", + "76. Zhang, S., Wang, J., Zhang, Y., Zhao, K., Yuan, H., Qin, Z., Wang, X., Zhao, D., Zhou, J.: I2vgen-xl: High-quality image-to-video synthesis via cascaded diffusion models. arXiv preprint arXiv:2311.04145 (2023)", + "77. Zhang, Z., Cole, F., Tucker, R., Freeman, W.T., Dekel, T.: Consistent depth of moving objects in video. ACM Transactions on Graphics (TOG) 40(4), 1-12 (2021)" + ], + "bbox": [ + 212, + 146, + 787, + 729 + ], + "page_idx": 18 + }, + { + "type": "header", + "text": "PhysDreamer", + "bbox": [ + 638, + 114, + 730, + 127 + ], + "page_idx": 18 + }, + { + "type": "page_number", + "text": "19", + "bbox": [ + 767, + 114, + 785, + 126 + ], + "page_idx": 18 + } +] \ No newline at end of file diff --git a/2024/Physics-Based Interaction with 3D Objects via Video Generation/017e5389-f4df-4a9c-a529-0be7433a235c_model.json b/2024/Physics-Based Interaction with 3D Objects via Video Generation/017e5389-f4df-4a9c-a529-0be7433a235c_model.json new file mode 100644 index 0000000000000000000000000000000000000000..bda183be9920495209905fb74d1eade4ac2ccb25 --- /dev/null +++ b/2024/Physics-Based Interaction with 3D Objects via Video Generation/017e5389-f4df-4a9c-a529-0be7433a235c_model.json @@ -0,0 +1,2900 @@ +[ + [ + { + "type": "title", + "bbox": [ + 0.257, + 0.141, + 0.75, + 0.185 + ], + "angle": 0, + "content": "Physics-Based Interaction with 3D Objects via Video Generation" + }, + { + "type": "text", + "bbox": [ + 0.241, + 0.213, + 0.761, + 0.245 + ], + "angle": 0, + "content": "Tianyuan Zhang\\(^{1}\\), Hong-Xing Yu\\(^{2}\\), Rundi Wu\\(^{3}\\), Brandon Y. Feng\\(^{1}\\), Changxi Zheng\\(^{3}\\), Noah Snavely\\(^{4}\\), Jiajun Wu\\(^{2}\\), and William T. Freeman\\(^{1}\\)" + }, + { + "type": "text", + "bbox": [ + 0.365, + 0.255, + 0.637, + 0.269 + ], + "angle": 0, + "content": "1 Massachusetts Institute of Technology" + }, + { + "type": "text", + "bbox": [ + 0.428, + 0.27, + 0.575, + 0.283 + ], + "angle": 0, + "content": "2 Stanford University" + }, + { + "type": "text", + "bbox": [ + 0.428, + 0.283, + 0.577, + 0.297 + ], + "angle": 0, + "content": "3 Columbia University" + }, + { + "type": "text", + "bbox": [ + 0.432, + 0.297, + 0.57, + 0.312 + ], + "angle": 0, + "content": "4 Cornell University" + }, + { + "type": "text", + "bbox": [ + 0.261, + 0.352, + 0.744, + 0.631 + ], + "angle": 0, + "content": "Abstract. Realistic object interactions are crucial for creating immersive virtual experiences, yet synthesizing realistic 3D object dynamics in response to novel interactions remains a significant challenge. Unlike unconditional or text-conditioned dynamics generation, action-conditioned dynamics requires perceiving the physical material properties of objects and grounding the 3D motion prediction on these properties, such as object stiffness. However, estimating physical material properties is an open problem due to the lack of material ground-truth data, as measuring these properties for real objects is highly difficult. We present PhysDreamer, a physics-based approach that endows static 3D objects with interactive dynamics by leveraging the object dynamics priors learned by video generation models. By distilling these priors, PhysDreamer enables the synthesis of realistic object responses to novel interactions, such as external forces or agent manipulations. We demonstrate our approach on diverse examples of elastic objects and evaluate the realism of the synthesized interactions through a user study. PhysDreamer takes a step towards more engaging and realistic virtual experiences by enabling static 3D objects to dynamically respond to interactive stimuli in a physically plausible manner. See our project page at https://physdreamer.github.io/." + }, + { + "type": "text", + "bbox": [ + 0.262, + 0.644, + 0.687, + 0.658 + ], + "angle": 0, + "content": "Keywords: Physics-based modeling \\(\\cdot\\) Interactive 3D dynamics" + }, + { + "type": "title", + "bbox": [ + 0.216, + 0.685, + 0.377, + 0.701 + ], + "angle": 0, + "content": "1 Introduction" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.719, + 0.788, + 0.809 + ], + "angle": 0, + "content": "Realistic object interactions play a pivotal role in creating immersive virtual experiences. Recent advances in 3D vision have enabled the capture and creation of high-quality static 3D assets [34,50], and some methods even extend to 4D assets [46,47,58], generating unconditioned dynamics. However, these methods fail to handle action-conditioned dynamics in response to new physical interactions, such as synthesizing the motion of a rose reacting to a breeze or a touch." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.81, + 0.788, + 0.841 + ], + "angle": 0, + "content": "The key challenge in synthesizing action-conditioned dynamics lies in understanding the physical material properties of objects. Yet, estimating these" + } + ], + [ + { + "type": "page_number", + "bbox": [ + 0.218, + 0.116, + 0.23, + 0.127 + ], + "angle": 0, + "content": "2" + }, + { + "type": "header", + "bbox": [ + 0.272, + 0.115, + 0.355, + 0.129 + ], + "angle": 0, + "content": "Zhang et al." + }, + { + "type": "image", + "bbox": [ + 0.224, + 0.145, + 0.396, + 0.278 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.271, + 0.28, + 0.346, + 0.291 + ], + "angle": 0, + "content": "PhysDreamer" + }, + { + "type": "image", + "bbox": [ + 0.406, + 0.146, + 0.461, + 0.276 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.468, + 0.146, + 0.784, + 0.276 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.513, + 0.28, + 0.74, + 0.292 + ], + "angle": 0, + "content": "Synthesized 3D dynamics under interaction" + }, + { + "type": "image_caption", + "bbox": [ + 0.214, + 0.302, + 0.789, + 0.373 + ], + "angle": 0, + "content": "Fig. 1: (Left) Leveraging and distilling dynamics priors from a pre-trained video generation model, we estimate a physical material field for the static 3D object. (Right) The physical material field allows synthesizing interactive 3D dynamics under arbitrary forces. We show rendered sequences from two viewpoints. Red arrows indicate force directions. Please see videos on our project website for better visualization." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.405, + 0.789, + 0.557 + ], + "angle": 0, + "content": "properties is a challenging task due to the lack of ground-truth data, as measuring these properties for real objects is highly difficult. Real-life objects often exhibit complex, spatially-varying material properties, making the estimation problem even more challenging. Despite the complexity of physical materials, humans can easily imagine how objects would react to external forces, such as the gentle sway of a rose. This ability to imagine object dynamics stems from our physical prior knowledge obtained from observing and interacting with the physical world. This motivates us to distill dynamics priors from video generation models that have been trained on vast, diverse video observations of the physical world." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.562, + 0.79, + 0.715 + ], + "angle": 0, + "content": "In this work, we focus on synthesizing interactive 3D dynamics. We propose PhysDreamer, a physics-based approach to transforming static 3D objects into interactive ones that can respond to novel interactions. The key idea behind PhysDreamer is to distill dynamics priors learned by video generation models to estimate the physical material properties of static 3D objects. We hypothesize that video generation models, trained on large amounts of video data, implicitly capture the relationship between object appearance and dynamics. By leveraging this learned prior knowledge, PhysDreamer can infer the physical material properties that drive the dynamic behavior of objects, even in the absence of ground-truth material data (Fig. 1)." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.72, + 0.789, + 0.842 + ], + "angle": 0, + "content": "PhysDreamer represents 3D objects using 3D Gaussians [34], models the physical material field with a neural field [71], and simulates 3D dynamics using the differentiable Material Point Method (MPM) [27]. The differentiable simulation and rendering allow for direct optimization of the physical material field and initial velocity field by matching pixel space observations. We focus on elastic dynamics and showcase PhysDreamer through diverse real examples, such as flowers, plants, a beanie hat, and a telephone cord. We evaluate the realism of the synthesized interactive motion through a user study, comparing PhysDreamer" + } + ], + [ + { + "type": "header", + "bbox": [ + 0.64, + 0.115, + 0.732, + 0.129 + ], + "angle": 0, + "content": "PhysDreamer" + }, + { + "type": "page_number", + "bbox": [ + 0.775, + 0.117, + 0.787, + 0.127 + ], + "angle": 0, + "content": "3" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.147, + 0.784, + 0.177 + ], + "angle": 0, + "content": "to state-of-the-art methods. The results demonstrate that our approach significantly outperforms existing techniques on motion realism and visual quality." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.178, + 0.788, + 0.342 + ], + "angle": 0, + "content": "In summary, PhysDreamer addresses the challenge of synthesizing interactive 3D dynamics by leveraging the object dynamics priors learned by video generation models. By distilling these priors to estimate the physical material properties of static 3D objects, our approach enables the creation of immersive virtual experiences where objects can respond realistically to novel interactions. The main contributions of our work include enabling static 3D objects to dynamically respond to interactive stimuli in a physically plausible manner and taking a step towards more engaging and realistic virtual experiences. We believe that PhysDreamer has the potential to greatly enhance the realism and interactivity of virtual environments, paving the way for more engaging and lifelike simulations." + }, + { + "type": "title", + "bbox": [ + 0.215, + 0.365, + 0.383, + 0.381 + ], + "angle": 0, + "content": "2 Related work" + }, + { + "type": "title", + "bbox": [ + 0.215, + 0.395, + 0.491, + 0.41 + ], + "angle": 0, + "content": "2.1 Dynamic 3D reconstruction" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.419, + 0.788, + 0.631 + ], + "angle": 0, + "content": "Dynamic 3D reconstruction methods aim to reconstruct a representation of a dynamic scene from inputs such as depth scans [13,41], RGBD videos [51], or monocular or multi-view videos [1,7,39,45,47,52,53,56,66,69,74,75]. This task is especially challenging in the monocular setting with slow-moving cameras and fast-moving scenes [20]. Novel scene representations are a major driver of recent progress. One prominent approach is to augment a canonical Neural Radiance Fields (NeRF) with a deformation field [56]. This approach can be further improved by incorporating flow supervision [23,66] or as-rigid-as-possible or volume preserving regularization terms [52,53]. Time-modulated NeRFs [8,19,20,43] offer a simpler alternative representation. Due to its Lagrangian nature, 3D Gaussian Splatting [34] is readily adaptable to the task of efficient dynamic scene reconstruction [17,30,39,47,72,74]. Data-driven prior, such as from monocular depth models [38,77] and image diffusion models [67], can also be used to reduce the inherent ambiguity in dynamic reconstruction from monocular videos." + }, + { + "type": "title", + "bbox": [ + 0.215, + 0.651, + 0.46, + 0.667 + ], + "angle": 0, + "content": "2.2 Dynamic 3D generation" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.674, + 0.788, + 0.841 + ], + "angle": 0, + "content": "Our work also relates to efforts to synthesize dynamic 3D scenes. A common approach is to integrate a 3D generation pipeline with a video generation model [2,46,58,60]. For instance, Make-A-Video3D begins by creating a static NeRF as per DreamFusion [55], then extending it temporally using Score Distillation Sampling (SDS) [55] derived from a video diffusion model. The approach can be improved with more efficient representations, stronger diffusion priors, and stable training techniques [2, 46]. However, applying SDS with video diffusion models demands significant computational and memory costs. Compact4D [73] and DreamGaussian4D [58] used a more efficient approach, synthesizing 3D dynamics by aligning a reference video from video generation models while employing SDS from image diffusion models to reduce novel view artifacts. These" + } + ], + [ + { + "type": "page_number", + "bbox": [ + 0.218, + 0.116, + 0.23, + 0.127 + ], + "angle": 0, + "content": "4" + }, + { + "type": "header", + "bbox": [ + 0.272, + 0.115, + 0.355, + 0.129 + ], + "angle": 0, + "content": "Zhang et al." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.147, + 0.784, + 0.178 + ], + "angle": 0, + "content": "methods are currently limited to producing fixed-length 3D videos. We focus on synthesizing interactive 3D motions under any new physical interactions." + }, + { + "type": "title", + "bbox": [ + 0.215, + 0.2, + 0.507, + 0.215 + ], + "angle": 0, + "content": "2.3 Interactive motion generation" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.224, + 0.785, + 0.36 + ], + "angle": 0, + "content": "Interactive motion generation animates still images or 3D contents according to user inputs like text [11, 76], motion fields [21], motion layers [12, 14], or direct manipulation such as dragging and pulling [15, 44]. Early work from Davis et al. [15, 16] demonstrated animating an image using an image-space modal basis extracted from a video of an object undergoing subtle vibrational motions. Building upon this image-space representation [15], Generative Image Dynamics [44] used a diffusion model trained on a dataset with paired image and its modal basis to model scene motion distributions, enabling realistic interaction with still input images. We focus on interacting with 3D objects rather than images." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.362, + 0.787, + 0.496 + ], + "angle": 0, + "content": "For 3D assets, physics-based approaches enable synthesizing motions under any physical interactions. Virtual Elastic Objects [10] jointly reconstructs the geometry, appearances, and physical parameters of elastic objects in a multiview capture setup with compressed air system. PAC-NeRF [42], DANO [40], and PhysGaussian [18] integrate physics-based simulations with NeRF and 3D Gaussians to generate physically plausible motions. We use the same physics-based approach to generate realistic interactions, but a novel ingredient of our work is to distill the material parameters of the object from pre-trained video generation models." + }, + { + "type": "title", + "bbox": [ + 0.215, + 0.519, + 0.467, + 0.534 + ], + "angle": 0, + "content": "2.4 Video generation models" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.544, + 0.785, + 0.65 + ], + "angle": 0, + "content": "Recent progress in video generation is driven by the development of larger autoregressive [26,37,65,68] and diffusion models [3-6,22,24,25,59]. These models, trained on increasingly large datasets, continue to advance the quality and realism of generated video content. The state-of-the-art approach [6] can generate minute-long videos with realistic motions and viewpoint consistency. However, current video generation models cannot support physics-based interactions with objects through external forces." + }, + { + "type": "title", + "bbox": [ + 0.215, + 0.673, + 0.455, + 0.689 + ], + "angle": 0, + "content": "3 Problem formulation" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.704, + 0.787, + 0.84 + ], + "angle": 0, + "content": "Given a static object represented by 3D Gaussians \\(\\{\\mathcal{G}_p\\}_{p=1}^P\\), \\(\\mathcal{G}_p = \\{\\boldsymbol{x}_p, \\alpha_p, \\boldsymbol{\\Sigma}_p, \\boldsymbol{c}_p\\}\\) (where \\(\\boldsymbol{x}_p\\) denotes the position, \\(\\alpha_p\\) denotes the opacity, \\(\\boldsymbol{\\Sigma}_p\\) denotes the covariance matrix, and \\(\\boldsymbol{c}_p\\) denotes the color of the particle), our goal is to estimate physical material property fields for the object to enable realistic interactive motion synthesis. These properties include mass \\(m\\), Young's modulus \\(E\\), and Poisson's ratio \\(\\nu\\). Among these physical properties, Young's modulus \\(E\\) plays a particularly important role in determining the object's motion in response to applied forces. Intuitively, Young's modulus (Eq. 2) measures the material stiffness. A higher Young's modulus results in less deformation and more rigid and" + } + ], + [ + { + "type": "header", + "bbox": [ + 0.64, + 0.115, + 0.732, + 0.128 + ], + "angle": 0, + "content": "PhysDreamer" + }, + { + "type": "page_number", + "bbox": [ + 0.775, + 0.117, + 0.785, + 0.127 + ], + "angle": 0, + "content": "5" + }, + { + "type": "image", + "bbox": [ + 0.225, + 0.149, + 0.368, + 0.262 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.254, + 0.264, + 0.327, + 0.273 + ], + "angle": 0, + "content": "3D Gaussians" + }, + { + "type": "image", + "bbox": [ + 0.398, + 0.151, + 0.517, + 0.262 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.415, + 0.264, + 0.493, + 0.273 + ], + "angle": 0, + "content": "10.30 seconds" + }, + { + "type": "image", + "bbox": [ + 0.527, + 0.151, + 0.648, + 0.262 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.551, + 0.264, + 0.623, + 0.273 + ], + "angle": 0, + "content": "2.30 seconds" + }, + { + "type": "image", + "bbox": [ + 0.658, + 0.151, + 0.78, + 0.262 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.683, + 0.264, + 0.756, + 0.273 + ], + "angle": 0, + "content": "0.87 seconds" + }, + { + "type": "image_caption", + "bbox": [ + 0.214, + 0.284, + 0.788, + 0.368 + ], + "angle": 0, + "content": "Fig. 2: Effect of Young's modulus. We depict the motion of a simulated flower under the same external force but with three different Young's moduli, a measure of material stiffness. Flowers with the highest Young's modulus \\((100\\times)\\) exhibit smaller oscillations and higher frequencies, while the flower with the lowest Young's modulus \\((1\\times)\\) sways the most and oscillates at the lowest frequency. Time annotations below each image indicate the duration of one complete motion path shown in the figure." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.395, + 0.785, + 0.442 + ], + "angle": 0, + "content": "higher-frequency motion, while a lower value leads to more flexible and elastic behavior. Fig. 2 illustrates the simulated motion of a flower under the same applied forces but with different Young's modulus." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.445, + 0.788, + 0.581 + ], + "angle": 0, + "content": "Therefore, our problem formulation focuses on estimating the spatially varying Young's modulus field \\( E(\\pmb{x}) \\) for the 3D object. To allow particle simulation, we query a particle's Young's modulus by \\( E_{p} = E(\\pmb{x}_{p}) \\). As for other physical properties, the mass for a particle \\( m_{p} \\) can be pre-computed as the product of a constant density \\( (\\rho) \\) and particle volume \\( V_{p} \\). The particle volume can be estimated [70] by dividing a background cell's volume by the number of particles that cell contains. As for the Poisson's ratio \\( \\nu_{p} \\), we found that it has negligible impact on object motion in our preliminary experiments (see supplementary materials for details), and so we assume a homogeneous constant Poisson's ratio." + }, + { + "type": "title", + "bbox": [ + 0.216, + 0.623, + 0.384, + 0.641 + ], + "angle": 0, + "content": "4 PhysDreamer" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.674, + 0.788, + 0.841 + ], + "angle": 0, + "content": "PhysDreamer estimates a material field for a static 3D object. Our key idea is to generate a plausible video of the object in motion, and then optimize the material field \\( E(\\pmb{x}) \\) to match this synthesized motion. We begin by rendering a static image \\( (I_0) \\) for the 3D scene \\( \\{\\mathcal{G}_p\\} \\) from a certain viewpoint. We then leverage an image-to-video model to generate a short video clip \\( \\{I_0, I_1, \\dots, I_T\\} \\) depicting the object's realistic motion. This generated video serves as our reference video. We then optimize the material field \\( E(\\pmb{x}) \\) and an initial velocity field \\( \\pmb{v}_0(\\pmb{x}) \\) (both modeled by implicit neural fields [71]) through differentiable simulation and differentiable rendering, such that a rendered video of the simulation matches (from the same viewpoint as \\( I_0 \\)) the reference video. Fig. 3 shows an overview of PhysDreamer." + } + ], + [ + { + "type": "page_number", + "bbox": [ + 0.218, + 0.116, + 0.23, + 0.127 + ], + "angle": 0, + "content": "6" + }, + { + "type": "header", + "bbox": [ + 0.272, + 0.115, + 0.355, + 0.129 + ], + "angle": 0, + "content": "Zhang et al." + }, + { + "type": "image", + "bbox": [ + 0.223, + 0.145, + 0.782, + 0.356 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.214, + 0.365, + 0.788, + 0.464 + ], + "angle": 0, + "content": "Fig. 3: Overview of PhysDreamer. Given an object represented as 3D Gaussians, we first render it (with background) from a viewpoint. Next, we use an image-to-video generation model to produce a reference video of that object in motion. Using differentiable Material Point Methods (MPM) and differentiable rendering, we optimize both a spatially-varying material field and an initial velocity field (not shown in the figure above). This optimization aims to minimize the discrepancy between the rendered video and the reference video. The dashed arrows represent gradient flow." + }, + { + "type": "title", + "bbox": [ + 0.216, + 0.485, + 0.373, + 0.499 + ], + "angle": 0, + "content": "4.1 Preliminaries" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.515, + 0.788, + 0.669 + ], + "angle": 0, + "content": "3D Gaussians [34] adopts a set of anisotropic 3D Gaussian kernels to represent the radiance field of a 3D scene. Although introduced primarily as an efficient method for 3D novel view synthesis, the Lagrangian nature of 3D Gaussians also enables the direct adaptation of particle-based physics simulators. Following PhysGaussian [70], we use the Material Point Method (MPM) to simulate object dynamics directly on these Gaussian particles. Since 3D Gaussians mainly lie on object surfaces, an optional internal filling process can be applied for improved simulation realism [70]. Below, we provide a brief introduction on the underlying physical model and how to integrate MPM into 3D Gaussians. For a more comprehensive introduction of MPM, we refer interested readers to [27,32,33,70]." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.673, + 0.789, + 0.842 + ], + "angle": 0, + "content": "Continuum mechanics and elastic materials. Continuum mechanics models material deformation using a map \\(\\phi\\) that transforms points from the undeformed material space \\(\\mathbf{X}\\) to the deformed world space \\(\\mathbf{x} = \\phi(\\mathbf{X}, t)\\). The Jacobian of the map, \\(\\pmb{F} = \\nabla_{\\mathbf{X}}\\phi(\\mathbf{X}, t)\\), known as the deformation gradient, measures local rotation and strain. This tensor is crucial in formulating stress-strain relationship. For example, the Cauchy stress in a hyper-elastic material is computed by: \\(\\pmb{\\sigma} = \\frac{1}{\\operatorname*{det}(\\pmb{F})}\\frac{\\partial\\psi}{\\partial\\pmb{F}}\\pmb{F}^T\\). Here, \\(\\psi(\\mathbf{F})\\) represents the strain energy density function, quantifying the extent of non-rigid deformations. This function is typically designed by experts, to follow principles like material symmetry and rotational invariance while aligning with empirical data. In this work, we use fixed corotated hyperelastic model, whose energy density function can be expressed as:" + } + ], + [ + { + "type": "header", + "bbox": [ + 0.64, + 0.115, + 0.732, + 0.128 + ], + "angle": 0, + "content": "PhysDreamer" + }, + { + "type": "page_number", + "bbox": [ + 0.776, + 0.117, + 0.786, + 0.127 + ], + "angle": 0, + "content": "7" + }, + { + "type": "equation", + "bbox": [ + 0.344, + 0.161, + 0.785, + 0.201 + ], + "angle": 0, + "content": "\\[\n\\psi (\\mathbf {F}) = \\mu \\left(\\sum_ {i = 1} ^ {d} \\left(\\sigma_ {i} - 1\\right) ^ {2}\\right) + \\frac {\\lambda}{2} (\\det (\\mathbf {F}) - 1) ^ {2}, \\tag {1}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.216, + 0.207, + 0.784, + 0.237 + ], + "angle": 0, + "content": "where \\(\\sigma_{i}\\) denotes a singular value of the deformation gradient. \\(\\mu\\) and \\(\\lambda\\) are related to Young's modulus \\(E\\) and Poisson's ratio \\(\\nu\\) via:" + }, + { + "type": "equation", + "bbox": [ + 0.369, + 0.25, + 0.784, + 0.281 + ], + "angle": 0, + "content": "\\[\n\\mu = \\frac {E}{2 (1 + \\nu)}, \\quad \\lambda = \\frac {E \\nu}{(1 + \\nu) (1 - 2 \\nu)}. \\tag {2}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.216, + 0.292, + 0.752, + 0.307 + ], + "angle": 0, + "content": "The dynamics of an elastic object are governed by the following equations:" + }, + { + "type": "equation", + "bbox": [ + 0.366, + 0.319, + 0.784, + 0.347 + ], + "angle": 0, + "content": "\\[\n\\rho \\frac {D \\boldsymbol {v}}{D t} = \\nabla \\cdot \\boldsymbol {\\sigma} + \\mathbf {f}, \\quad \\frac {D \\rho}{D t} + \\rho \\nabla \\cdot \\boldsymbol {v} = 0, \\tag {3}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.216, + 0.358, + 0.784, + 0.387 + ], + "angle": 0, + "content": "where \\(\\rho\\) denotes density, \\(\\pmb{v}(\\pmb{x},t)\\) denotes the velocity field in world space, and \\(\\mathbf{f}\\) denotes an external force." + }, + { + "type": "text", + "bbox": [ + 0.216, + 0.394, + 0.785, + 0.499 + ], + "angle": 0, + "content": "Material Point Method (MPM). We use the Moving Least Squares Material Point Method (MLS-MPM) [27] to solve the above governing equation. MPM is a hybrid Eulerian-Langrangian method widely adopted for simulating dynamics for a wide range of materials, such as solid, fluid, sand, and cloth [31, 35, 57, 62]. MPM offers several advantages, such as easy GPU parallelization [28], handling of topology changes, and the availability of well-documented open-source implementations [29, 48, 49, 70]." + }, + { + "type": "text", + "bbox": [ + 0.216, + 0.501, + 0.785, + 0.577 + ], + "angle": 0, + "content": "Following PhysGaussian [70], we view the Gaussian particles as the spatial discretization of the object to be simulated, and directly run MPM on these Gaussian particles. Each particle \\( p \\) represents a small volume of the object, and it carries a set of properties including volume \\( V_{p} \\), mass \\( m_{p} \\), position \\( \\boldsymbol{x}_{p}^{t} \\), velocity \\( \\boldsymbol{v}_{p}^{t} \\), deformation gradient \\( \\boldsymbol{F}_{p}^{t} \\), and local velocity field gradient \\( \\boldsymbol{C}_{p}^{t} \\) at time step \\( t \\)." + }, + { + "type": "text", + "bbox": [ + 0.216, + 0.577, + 0.784, + 0.606 + ], + "angle": 0, + "content": "MPM operates in a particle-to-grid (P2G) and grid-to-particle (G2P) transfer loop. In the P2G stage, we transfer the momentum from particle to grid by:" + }, + { + "type": "equation", + "bbox": [ + 0.227, + 0.619, + 0.784, + 0.653 + ], + "angle": 0, + "content": "\\[\nm _ {i} ^ {t} \\boldsymbol {v} _ {i} ^ {t} = \\sum_ {p} N (\\boldsymbol {x} _ {i} - \\boldsymbol {x} _ {p} ^ {t}) [ m _ {p} \\boldsymbol {v} _ {p} ^ {t} + (m _ {p} \\boldsymbol {C} _ {p} ^ {t} - \\frac {4}{(\\Delta x) ^ {2}} \\Delta t V _ {p} \\frac {\\partial \\psi}{\\partial \\boldsymbol {F}} \\boldsymbol {F} _ {p} ^ {t T}) (\\boldsymbol {x} _ {i} - \\boldsymbol {x} _ {p} ^ {t}) ] + \\boldsymbol {f} _ {i} ^ {t}, \\tag {4}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.216, + 0.666, + 0.785, + 0.727 + ], + "angle": 0, + "content": "where the mass of the grid node \\(i\\) is \\(m_i^t = \\sum_p N(\\pmb{x}_i - \\pmb{x}_p^t)m_p\\), \\(N(\\pmb{x}_i - \\pmb{x}_p^t)\\) is the B-spline kernel, \\(\\Delta x\\) is the spatial grid resolution, \\(\\Delta t\\) is the simulation step size, and \\(\\pmb{v}_i^t\\) is the updated velocity on the grid. We then transfer the updated velocity back to the particles and update their positions as:" + }, + { + "type": "equation", + "bbox": [ + 0.327, + 0.741, + 0.784, + 0.769 + ], + "angle": 0, + "content": "\\[\n\\boldsymbol {v} _ {p} ^ {t + 1} = \\sum_ {i} N (\\boldsymbol {x} _ {i} - \\boldsymbol {x} _ {p} ^ {t}) \\boldsymbol {v} _ {i} ^ {t}, \\quad \\boldsymbol {x} _ {p} ^ {t + 1} = \\boldsymbol {x} _ {p} ^ {t} + \\Delta t \\boldsymbol {v} _ {p} ^ {t + 1}. \\tag {5}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.216, + 0.782, + 0.779, + 0.797 + ], + "angle": 0, + "content": "Meanwhile, the local velocity gradient and deformation gradient is updated as:" + }, + { + "type": "equation", + "bbox": [ + 0.236, + 0.809, + 0.784, + 0.843 + ], + "angle": 0, + "content": "\\[\n\\boldsymbol {C} _ {p} ^ {t + 1} = \\frac {4}{(\\Delta x) ^ {2}} \\sum_ {i} N (\\boldsymbol {x} _ {i} - \\boldsymbol {x} _ {p} ^ {t}) \\boldsymbol {v} _ {i} ^ {t} (\\boldsymbol {x} _ {i} - \\boldsymbol {x} _ {p} ^ {t}) ^ {T}, \\quad \\boldsymbol {F} _ {p} ^ {t + 1} = (\\boldsymbol {I} + \\Delta t \\boldsymbol {C} _ {p} ^ {t + 1}) \\boldsymbol {F} _ {p} ^ {t}. \\tag {6}\n\\]" + } + ], + [ + { + "type": "page_number", + "bbox": [ + 0.218, + 0.116, + 0.23, + 0.127 + ], + "angle": 0, + "content": "8" + }, + { + "type": "header", + "bbox": [ + 0.272, + 0.115, + 0.355, + 0.129 + ], + "angle": 0, + "content": "Zhang et al." + }, + { + "type": "title", + "bbox": [ + 0.216, + 0.147, + 0.514, + 0.163 + ], + "angle": 0, + "content": "4.2 Estimating physical properties" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.168, + 0.788, + 0.213 + ], + "angle": 0, + "content": "Using MLS-MPM [27, 70] as our physics simulator and the Fixed Corotated hyper-elastic material model for the 3D objects, the simulation process for a single sub-step is formalized as:" + }, + { + "type": "equation", + "bbox": [ + 0.328, + 0.219, + 0.786, + 0.236 + ], + "angle": 0, + "content": "\\[\n\\boldsymbol {x} ^ {t + 1}, \\boldsymbol {v} ^ {t + 1}, \\boldsymbol {F} ^ {t + 1}, \\boldsymbol {C} ^ {t + 1} = \\mathcal {S} \\left(\\boldsymbol {x} ^ {t}, \\boldsymbol {v} ^ {t}, \\boldsymbol {F} ^ {t}, \\boldsymbol {C} ^ {t}, \\boldsymbol {\\theta}, \\Delta t\\right), \\tag {7}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.242, + 0.784, + 0.348 + ], + "angle": 0, + "content": "where \\(\\pmb{x}^t = [\\pmb{x}_1^t, \\dots, \\pmb{x}_P^t]\\) denotes the positions of all particles at time \\(t\\), and similarly \\(\\pmb{v}^t = [\\pmb{v}_1^t, \\dots, \\pmb{v}_P^t]\\) denotes the velocities of all particles at time \\(t\\). \\(\\pmb{F}^t\\) and \\(\\pmb{C}^t\\) denote the deformation gradient and the gradient of local velocity fields for all particles, respectively. Both \\(\\pmb{F}^t\\) and \\(\\pmb{C}^t\\) are tracked for simulation purposes, not for rendering. \\(\\pmb{\\theta}\\) denotes the collection of the physical properties of all particles: mass \\(\\pmb{m} = [m_1, \\dots, m_P]\\), Young's modulus \\(\\pmb{E} = [E_1, \\dots, E_P]\\), Poisson's ratio \\(\\pmb{\\nu} = [\\nu_1, \\dots, \\nu_P]\\), and volume \\(\\pmb{V} = [V_1, \\dots, V_P]\\). \\(\\Delta t\\) is the simulation step size." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.348, + 0.784, + 0.408 + ], + "angle": 0, + "content": "We use a sub-step size \\(\\Delta t \\cong 1 \\times 10^{-4}\\) for most of our experiments. To simulate dynamics between adjacent video frames, we iterate over hundreds of sub-steps (time interval between frames are typically tens of milliseconds). For simplicity, we abuse notation to express a simulation step with \\(N\\) sub-steps as:" + }, + { + "type": "equation", + "bbox": [ + 0.318, + 0.414, + 0.785, + 0.431 + ], + "angle": 0, + "content": "\\[\n\\boldsymbol {x} ^ {t + 1}, \\boldsymbol {v} ^ {t + 1}, \\boldsymbol {F} ^ {t + 1}, \\boldsymbol {C} ^ {t + 1} = \\mathcal {S} (\\boldsymbol {x} ^ {t}, \\boldsymbol {v} ^ {t}, \\boldsymbol {F} ^ {t}, \\boldsymbol {C} ^ {t}, \\boldsymbol {\\theta}, \\Delta t, N), \\tag {8}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.215, + 0.437, + 0.784, + 0.466 + ], + "angle": 0, + "content": "where the timestamp \\( t + 1 \\) is ahead of timestamp \\( t \\) by \\( N\\Delta t \\). After simulation, we render the Gaussians at each frame:" + }, + { + "type": "equation", + "bbox": [ + 0.402, + 0.474, + 0.785, + 0.49 + ], + "angle": 0, + "content": "\\[\n\\hat {I} ^ {t} = \\mathcal {F} _ {\\text {r e n d e r}} \\left(\\boldsymbol {x} ^ {t}, \\boldsymbol {\\alpha}, \\boldsymbol {R} ^ {t}, \\Sigma , \\boldsymbol {c}\\right), \\tag {9}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.496, + 0.784, + 0.526 + ], + "angle": 0, + "content": "where \\(\\mathcal{F}_{\\mathrm{render}}\\) denotes the differentiable rendering function, and \\(\\pmb{R}^{t}\\) denotes the rotation matrices of all particles obtained from the simulation step." + }, + { + "type": "text", + "bbox": [ + 0.215, + 0.527, + 0.784, + 0.556 + ], + "angle": 0, + "content": "Using the generated video as reference, we optimize the spatially-varying Young's modulus \\( \\pmb{E} \\) and an initial velocity \\( \\pmb{v}^0 \\) by a per-frame loss function:" + }, + { + "type": "equation", + "bbox": [ + 0.353, + 0.563, + 0.785, + 0.58 + ], + "angle": 0, + "content": "\\[\nL ^ {t} = \\lambda L _ {1} \\left(\\hat {I} ^ {t}, I ^ {t}\\right) + (1 - \\lambda) L _ {\\mathrm {D} - \\mathrm {S S I M}} \\left(\\hat {I} ^ {t}, I ^ {t}\\right), \\tag {10}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.215, + 0.586, + 0.514, + 0.6 + ], + "angle": 0, + "content": "where we set \\(\\lambda = 0.1\\) in our experiments." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.601, + 0.784, + 0.676 + ], + "angle": 0, + "content": "We parameterize the material field and velocity field by two triplanes [9], each followed by a three-layer MLP. Additionally, we apply a total variation regularization for all spatial planes of both fields to encourage spatial smoothness. Using \\(\\pmb{u}\\) to denote one of the 2D spatial planes, and \\(\\pmb{u}_{i,j}\\) as a feature vector on the 2D plane, we write the total variation regularization term as:" + }, + { + "type": "equation", + "bbox": [ + 0.341, + 0.683, + 0.785, + 0.713 + ], + "angle": 0, + "content": "\\[\nL _ {\\mathrm {t v}} = \\sum_ {i, j} \\| \\boldsymbol {u} _ {i + 1, j} - \\boldsymbol {u} _ {i, j} \\| _ {2} ^ {2} + \\| \\boldsymbol {u} _ {i, j + 1} - \\boldsymbol {u} _ {i, j} \\| _ {2} ^ {2}. \\tag {11}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.72, + 0.788, + 0.841 + ], + "angle": 0, + "content": "Rather than optimizing the material parameters and initial velocity jointly, we split the optimization into two stages for better stability and faster convergence. In particular, in the first stage, we randomly initialize the Young's modulus for each Gaussian particle and freeze it. We optimize the initial velocity of each particle using only the first three frames of the reference video. In the second stage, we freeze the initial velocity and optimize the spatially varying Young's modulus. During the second stage, the gradient signal only flows to the previous frame to prevent gradient explosion/vanishing." + } + ], + [ + { + "type": "header", + "bbox": [ + 0.64, + 0.115, + 0.732, + 0.128 + ], + "angle": 0, + "content": "PhysDreamer" + }, + { + "type": "page_number", + "bbox": [ + 0.775, + 0.116, + 0.787, + 0.127 + ], + "angle": 0, + "content": "9" + }, + { + "type": "image_caption", + "bbox": [ + 0.216, + 0.149, + 0.461, + 0.289 + ], + "angle": 0, + "content": "Fig.4: Accelerated MPM with K-Means downsampling. We employ K-Means clustering to create a set of \"driving particles\" (in yellow) at the initial time step \\((t = 0)\\). We only simulate these driving particles. When rendering, we obtain each particle's position and rotation by fitting a local rigid body transformation using neighboring driving particles." + }, + { + "type": "image", + "bbox": [ + 0.494, + 0.144, + 0.758, + 0.306 + ], + "angle": 0, + "content": null + }, + { + "type": "title", + "bbox": [ + 0.216, + 0.323, + 0.607, + 0.339 + ], + "angle": 0, + "content": "4.3 Accelerating simulation with subsampling" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.35, + 0.788, + 0.41 + ], + "angle": 0, + "content": "High-fidelity rendering with 3D Gaussians typically requires millions of particles to represent a scene. Running simulations on all the particles poses a significant computational burden. To improve efficiency, we introduce a subsampling procedure for simulation, as illustrated in Fig. 4." + }, + { + "type": "text", + "bbox": [ + 0.216, + 0.411, + 0.788, + 0.609 + ], + "angle": 0, + "content": "Specifically, we apply K-Means clustering to create a set of driving particles \\(\\{\\mathcal{Q}_q\\}_{q=1}^Q\\) at \\(t = 0\\), where each driving particle is represented by \\(\\mathcal{Q}_q^0 = \\{\\pmb{x}_q^0, \\pmb{v}_q^0, \\pmb{F}_q^0, \\pmb{C}_q^0, E_q, m_q, \\nu_q, V_q\\}\\). The initial position of a driving particle \\(\\pmb{x}_q^0\\) is computed as the mean of the position \\(\\pmb{x}_p\\) of all cluster members. The number of the driving particles is much smaller than the number of 3D Gaussian particles, \\(Q \\ll P\\). We run simulations only on the driving particles. During rendering, we compute the position and rotation for each 3D Gaussian particle \\(\\mathcal{G}_p\\) by interpolating the driving particles. In particular, for each 3D Gaussian particle, we find its eight nearest driving particles at \\(t = 0\\), and we fit a rigid body transformation \\(\\pmb{T}\\) between these eight driving particles at \\(t = 0\\) and at the current timestamp. This rigid body transformation \\(\\pmb{T}\\) is applied to the initial position and rotation of the particle \\(\\mathcal{G}_p\\) to obtain its current position and rotation. We summarize our algorithm with pseudo-code in supplementary materials." + }, + { + "type": "title", + "bbox": [ + 0.216, + 0.633, + 0.376, + 0.65 + ], + "angle": 0, + "content": "5 Experiments" + }, + { + "type": "title", + "bbox": [ + 0.216, + 0.667, + 0.312, + 0.683 + ], + "angle": 0, + "content": "5.1 Setup" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.699, + 0.788, + 0.804 + ], + "angle": 0, + "content": "Datasets. We collect eight real-world static scenes by capturing multi-view images. Each scene includes an object and a background. The objects include five flowers (a red rose, a carnation, an orange rose, a tulip, and a white rose), an alocasia plant, a telephone cord, and a beanie hat. For each scene except for the red rose scene, we capture four interaction videos illustrating its natural motion after interaction, such as poking or dragging, and we use the real videos as additional comparison references." + }, + { + "type": "text", + "bbox": [ + 0.216, + 0.81, + 0.787, + 0.841 + ], + "angle": 0, + "content": "Baselines. We compare our approach to two baselines: PhysGaussian [70] and DreamGaussian4D [58]. PhysGaussian [70] integrates MPM simulation to static" + } + ], + [ + { + "type": "page_number", + "bbox": [ + 0.218, + 0.116, + 0.236, + 0.127 + ], + "angle": 0, + "content": "10" + }, + { + "type": "header", + "bbox": [ + 0.272, + 0.115, + 0.355, + 0.129 + ], + "angle": 0, + "content": "Zhang et al." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.147, + 0.788, + 0.284 + ], + "angle": 0, + "content": "3D Gaussians to support simulation, but it cannot estimate material properties and relies on manually setting material parameter values. Thus, we use the same initialization strategy as ours to assign material properties for PhysGaussian. DreamGaussian4D [58] generates non-interactive dynamic 3D Gaussians from a static image. It first obtains a static 3D Gaussians using DreamGaussian [63], and then animate it by optimizing a deformation field from a generated driving video. For a fair comparison, we run its deformation field optimization on our reconstructed static 3D Gaussians, and we looped the resulting deformation field when rendering longer videos in later comparison." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.289, + 0.788, + 0.394 + ], + "angle": 0, + "content": "Evaluation metrics. We focus on the quality of the synthesized object motion, in particular, visual quality and motion realism. Therefore, we conduct a user study and adopt the Two-alternative Forced Choice (2AFC) protocol: the participants are shown two side-by-side synchronized videos, including one video result from ours and the other one from the competitor's, with a random left-right ordering. The participants are then asked to choose the one with higher visual quality and the one with higher motion realism." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.395, + 0.788, + 0.547 + ], + "angle": 0, + "content": "We recruited 100 participants, each asked to judge all 8 scenes, forming a total of 800 2AFC judgement samples for each baseline comparison. For each scene, we create 4 sample video pairs and show participants a random one from the 4 pairs. In particular, we create 4 five-second motion sequences using PhysDreamer with randomized initial conditions (applying an external force to the foreground object or assigning an initial velocity to the object), and render videos from randomly picked viewpoints. For the baseline method, we apply the same initial conditions (for PhysGaussian only) and render videos from the same viewpoint as ours to form the video pairs. Please see supplementary materials for human study details and quantitative metrics for videos (e.g., Fréchet Video Distance [64])." + }, + { + "type": "title", + "bbox": [ + 0.216, + 0.57, + 0.455, + 0.586 + ], + "angle": 0, + "content": "5.2 Implementation details" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.602, + 0.787, + 0.648 + ], + "angle": 0, + "content": "Neural material fields. We represent both material field and initial velocity field using triplanes [54] each followed by a three-layer MLP. The triplanes have spatial resolutions of \\(8^{3}\\) and \\(24^{3}\\) for the material field and velocity field, respectively." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.654, + 0.788, + 0.699 + ], + "angle": 0, + "content": "3D Gaussian reconstruction. Similar to PhysGaussian [70], we employ anisotropic regularization to reduce skinny artifacts in the reconstruction. Each reconstructed scene contains 0.5 to 1.5 million particles (including foreground and background)." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.705, + 0.788, + 0.841 + ], + "angle": 0, + "content": "Simulation details. For computational efficiency, we segment the background and keep only foreground object particles for simulation. In our experiments, the foreground object contains around 50 to 300 thousand 3D Gaussian particles. We then discretize the foreground into a \\(64^3\\) grid. The number of driving particles are 10 to 50 times fewer than the number of 3D Gaussian particles, determined by maintaining an average of at least eight particles per occupied voxel. For accurate motion, we use 768 sub-steps between successive video frames, corresponding to a duration of \\(4.34 \\times 10^{-5}\\) second for each sub-step. To address the high memory consumption from large number of steps, we apply simulation" + } + ], + [ + { + "type": "header", + "bbox": [ + 0.64, + 0.115, + 0.732, + 0.128 + ], + "angle": 0, + "content": "PhysDreamer" + }, + { + "type": "page_number", + "bbox": [ + 0.769, + 0.116, + 0.785, + 0.127 + ], + "angle": 0, + "content": "11" + }, + { + "type": "image", + "bbox": [ + 0.219, + 0.146, + 0.331, + 0.231 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.22, + 0.233, + 0.33, + 0.317 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.221, + 0.319, + 0.33, + 0.403 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.221, + 0.405, + 0.33, + 0.489 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.234, + 0.49, + 0.315, + 0.5 + ], + "angle": 0, + "content": "Material result" + }, + { + "type": "image", + "bbox": [ + 0.342, + 0.146, + 0.453, + 0.23 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.343, + 0.232, + 0.453, + 0.316 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.343, + 0.319, + 0.453, + 0.403 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.343, + 0.405, + 0.453, + 0.489 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.455, + 0.147, + 0.563, + 0.23 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.455, + 0.232, + 0.563, + 0.316 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.455, + 0.319, + 0.563, + 0.403 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.455, + 0.405, + 0.563, + 0.489 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.565, + 0.147, + 0.672, + 0.23 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.565, + 0.232, + 0.672, + 0.316 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.565, + 0.319, + 0.672, + 0.403 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.565, + 0.405, + 0.672, + 0.489 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.675, + 0.147, + 0.783, + 0.23 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.675, + 0.232, + 0.783, + 0.316 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.675, + 0.319, + 0.783, + 0.403 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.675, + 0.405, + 0.783, + 0.489 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.446, + 0.489, + 0.679, + 0.501 + ], + "angle": 0, + "content": "Synthesized 3D dynamics under interaction" + }, + { + "type": "image_caption", + "bbox": [ + 0.214, + 0.513, + 0.788, + 0.57 + ], + "angle": 0, + "content": "Fig. 5: Interactive 3D dynamics synthesis. (Left) Visualization of the material fields. Brighter color indicates higher Young's modulus within each example. (Right) We apply an external force (red arrow) on each object, and the following columns demonstrate the object dynamics rendered at a static viewpoint." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.589, + 0.788, + 0.635 + ], + "angle": 0, + "content": "state checkpointing and re-computation during gradient back-propagation. We add Dirichlet boundary conditions for stationary grid cells. We fill the internal volumes of certain solid objects to enhance simulation realism [70]." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.64, + 0.788, + 0.746 + ], + "angle": 0, + "content": "Generating reference videos. We render a 3D object with its background from a viewpoint, and then we use Stable Video Diffusion [4] to animate this rendered image and generate fourteen video frames. We use a small motion bucket number [4] (e.g., 5 or 8) so that the generated video contains mostly object motion and little camera motion. We use rendered images for the video generation, so that our approach can also be used for generated scenes. Also, rendering images directly from 3D Gaussians simplifies later optimization." + }, + { + "type": "title", + "bbox": [ + 0.215, + 0.755, + 0.325, + 0.769 + ], + "angle": 0, + "content": "5.3 Results" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.78, + 0.789, + 0.842 + ], + "angle": 0, + "content": "We show our qualitative results of the spatially-varying Young's modulus in Fig. 5 (left), and simulated interactive motion in Fig. 5 (right). Please see our project website videos for a better motion visualization. Tab. 1 presents the user study results in comparison to baseline methods and real captured videos." + } + ], + [ + { + "type": "page_number", + "bbox": [ + 0.218, + 0.116, + 0.236, + 0.127 + ], + "angle": 0, + "content": "12" + }, + { + "type": "header", + "bbox": [ + 0.272, + 0.115, + 0.355, + 0.129 + ], + "angle": 0, + "content": "Zhang et al." + }, + { + "type": "table_caption", + "bbox": [ + 0.216, + 0.145, + 0.788, + 0.201 + ], + "angle": 0, + "content": "Table 1: Human study 2AFC results of PhysDreamer (Ours) over real captured videos and baseline methods (PhysGaussian [70] and DreamGaussian4D [58]) on Motion Realism and overall Visual Quality. \"Rose O\", \"Rose W\", and \"Rose R\" denotes the orange, white, and red roses, respectively." + }, + { + "type": "table", + "bbox": [ + 0.226, + 0.213, + 0.78, + 0.367 + ], + "angle": 0, + "content": "
Motion realismAlocasiaCarnationHatRose ORose WRose RCordTulipAvg.
Ours over Real capture86%61%55%63%47%-29%35%53.7%
Ours over PhysGaussian96%89%57%91%93%73%61%86%80.8%
Ours over DreamGaussian75%77%51%78%51%41%71%64%63.5%
Visual quality
Ours over Real capture36%53%28%40%41%-29%34%37.3%
Ours over PhysGaussian67%69%50%75%73%58%58%70%65.0%
Ours over DreamGaussian82%75%74%76%60%47%76%70%70.0%
" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.395, + 0.787, + 0.485 + ], + "angle": 0, + "content": "Compared to PhysGaussian, \\(80.8\\%\\) of the human participant 2AFC samples prefer PhysDreamer (ours) in motion realism and \\(65.0\\%\\) prefer PhysDreamer in visual quality. Note that since the static scenes are the same, the visual quality also depends on the generated object motion. Fig. 6 shows temporal slices of the motion patterns. We observe that PhysGaussian produces large, unrealistic slow motion due to the lack of a principled estimation of material properties." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.486, + 0.787, + 0.622 + ], + "angle": 0, + "content": "Compared to DreamGaussian4D, \\(70.0\\% / 63.5\\%\\) 2AFC samples prefer ours in visual quality and motion realism, respectively. From Fig. 6, we can observe that DreamGaussian4D generates periodic motion with a constant, small magnitude, while PhysDreamer can simulate the damping in motion. This is because DreamGaussian4D does not simulate the physical dynamics but simply distill a motion sequence from a generative model, so it cannot extrapolate to different motion. We further include one more evaluation dimension on \"motion amount\" comparing to DreamGaussian4D, where we ask the participants to judge which video has higher amount of motion, and \\(73.6\\%\\) 2AFC samples prefer PhysDreamer." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.623, + 0.788, + 0.76 + ], + "angle": 0, + "content": "Compared to real videos, \\(53.7\\%\\) 2AFC samples favored the motion realism of ours results. Interestingly, under \"Motion Realism\", \\(86\\%\\) of the users indicated that the alocasia outputs were more realistic than real captures. This is surprising, as one would expect a \\(50\\%\\) preference if the videos were indistinguishable. We offer a potential explanation: for thin geometries like alocasia leaves, the Material Point Method tends to produce lower-frequency and slower motions. This can be observed in the video and is evident in the space-time slice visualizations in Fig. 6. Humans are poor at judging the naturalness of motion and may be biased towards smoother and slower motions, as shown in prior studies [36,61]." + }, + { + "type": "title", + "bbox": [ + 0.216, + 0.783, + 0.62, + 0.799 + ], + "angle": 0, + "content": "5.4 Ablation: using multi-view reference videos" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.81, + 0.787, + 0.84 + ], + "angle": 0, + "content": "For objects with self-occlusion, observing salient motion of all object parts from a single video is challenging (e.g., the alocasia scene where a leaf can occlude" + } + ], + [ + { + "type": "header", + "bbox": [ + 0.64, + 0.115, + 0.732, + 0.129 + ], + "angle": 0, + "content": "PhysDreamer" + }, + { + "type": "page_number", + "bbox": [ + 0.769, + 0.116, + 0.786, + 0.127 + ], + "angle": 0, + "content": "13" + }, + { + "type": "image", + "bbox": [ + 0.218, + 0.145, + 0.785, + 0.725 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.214, + 0.733, + 0.788, + 0.817 + ], + "angle": 0, + "content": "Fig. 6: We compare our results with real captured videos, PhysGaussian [70], and DreamGaussian4D [58] using space-time slices. In these slices, the vertical axis represent time, and the horizontal axis shows a spatial slice of the object (denoted by red lines on the \"object\" column). These slices visualize the magnitude and frequencies of these oscillating motions. Results for our PhysDreamer (Ours) and PhysGaussian are simulated with the same initial conditions." + } + ], + [ + { + "type": "page_number", + "bbox": [ + 0.218, + 0.116, + 0.236, + 0.127 + ], + "angle": 0, + "content": "14" + }, + { + "type": "header", + "bbox": [ + 0.272, + 0.115, + 0.355, + 0.129 + ], + "angle": 0, + "content": "Zhang et al." + }, + { + "type": "image", + "bbox": [ + 0.218, + 0.145, + 0.785, + 0.282 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.214, + 0.296, + 0.788, + 0.354 + ], + "angle": 0, + "content": "Fig. 7: Comparison between single-view (top) and two-view (bottom) supervisions. The object (alocasia) exhibits self-occluding structures. We can use generated videos at two views to jointly optimize the material field. In the space-time (X-t) slices, the vertical axis represents time, and the horizontal axis shows a spatial slice of the object." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.375, + 0.789, + 0.526 + ], + "angle": 0, + "content": "another leaf). We may alleviate this problem by rendering from multiple viewpoints to provide comprehensive coverage of the object. Here, we use multiple videos in the material estimation, jointly optimizing a video-agnostic, spatially-varying Young's modulus for each particle along with video-specific initial velocities. From the comparison of the alocasia scene in Fig. 7, we can see that using multi-view reference videos (a front view and a back view) helps in such complex self-occluding objects: PhysDreamer benefits significantly from having supervision from two views, while using only a single view leads to artifacts. In our user study, \\(81.0\\%\\) 2AFC samples prefer results with two view supervision in visual quality and \\(86.0\\%\\) in motion realism." + }, + { + "type": "title", + "bbox": [ + 0.216, + 0.555, + 0.36, + 0.57 + ], + "angle": 0, + "content": "6 Conclusion" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.592, + 0.788, + 0.713 + ], + "angle": 0, + "content": "In this work, we introduced PhysDreamer, a novel approach to synthesizing interactive 3D dynamics by endowing static 3D objects with physical material properties. Our method distills the object dynamics priors learned by video generation models to estimate the spatially-varying material properties. We showcase dynamics interaction with a diverse set of elastic objects by PhysDreamer. We believe that PhysDreamer takes a significant step towards creating more engaging and immersive virtual environments, opening up a wide range of applications from realistic simulations to interactive virtual experiences." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.72, + 0.788, + 0.84 + ], + "angle": 0, + "content": "Limitations. Our approach requires the user to manually specify the object to simulate and separate it from the background, and establish boundary conditions for stationary parts, like the pot of flowers. 3D object discovery may help for simulatable object extraction. In addition, our approach is computationally demanding. Despite our subsampling strategy, our current algorithm takes approximately one minute on a NVIDIA V100 GPU to produce a single second of video. Further improving efficiency remains an important future problem. Finally, in this work, we restrict our scope to elastic objects without collisions." + } + ], + [ + { + "type": "header", + "bbox": [ + 0.64, + 0.115, + 0.732, + 0.128 + ], + "angle": 0, + "content": "PhysDreamer" + }, + { + "type": "page_number", + "bbox": [ + 0.769, + 0.116, + 0.786, + 0.127 + ], + "angle": 0, + "content": "15" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.147, + 0.788, + 0.283 + ], + "angle": 0, + "content": "Acknowledgements. This work is in part supported by the NSF PHY-2019786 (The NSF AI Institute for Artificial Intelligence and Fundamental Interactions, http://iaifi.org/), NSF CIF 1955864 (Occlusion and Directional Resolution in Computational Imaging), RI #2211258, #2338203, ONR MURI N00014-22-1-2740, Quanta Computer, Samsung, and United States Air Force Artificial Intelligence Accelerator and was accomplished under Cooperative Agreement Number FA8750-192-1000. We would like to thank Peter Yichen Chen, Zhengqi Li, Pingchuan Ma, Minghao Guo, Ge Yang, and Shai Avidan for help and insightful discussions." + }, + { + "type": "title", + "bbox": [ + 0.216, + 0.311, + 0.325, + 0.327 + ], + "angle": 0, + "content": "References" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.346, + 0.787, + 0.403 + ], + "angle": 0, + "content": "1. Attal, B., Huang, J.B., Richardt, C., Zollhoefer, M., Kopf, J., O'Toole, M., Kim, C.: Hyperreel: High-fidelity 6-dof video with ray-conditioned sampling. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 16610-16620 (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.403, + 0.787, + 0.459 + ], + "angle": 0, + "content": "2. Bahmani, S., Skorokhodov, I., Rong, V., Wetzstein, G., Guibas, L., Wonka, P., Tulyakov, S., Park, J.J., Tagliasacchi, A., Lindell, D.B.: 4d-fy: Text-to-4d generation using hybrid score distillation sampling. arXiv preprint arXiv:2311.17984 (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.46, + 0.787, + 0.502 + ], + "angle": 0, + "content": "3. Bar-Tal, O., Chefer, H., Tov, O., Herrmann, C., Paiss, R., Zada, S., Ephrat, A., Hur, J., Li, Y., Michaeli, T., et al.: Lumiere: A space-time diffusion model for video generation. arXiv preprint arXiv:2401.12945 (2024)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.502, + 0.787, + 0.558 + ], + "angle": 0, + "content": "4. Blattmann, A., Dockhorn, T., Kulal, S., Mendelevitch, D., Kilian, M., Lorenz, D., Levi, Y., English, Z., Voleti, V., Letts, A., et al.: Stable video diffusion: Scaling latent video diffusion models to large datasets. arXiv preprint arXiv:2311.15127 (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.558, + 0.787, + 0.614 + ], + "angle": 0, + "content": "5. Blattmann, A., Rombach, R., Ling, H., Dockhorn, T., Kim, S.W., Fidler, S., Kreis, K.: Align your latents: High-resolution video synthesis with latent diffusion models. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 22563-22575 (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.615, + 0.787, + 0.671 + ], + "angle": 0, + "content": "6. Brooks, T., Peebles, B., Homes, C., DePue, W., Guo, Y., Jing, L., Schnurr, D., Taylor, J., Luhman, T., Luhman, E., Ng, C., Wang, R., Ramesh, A.: Video generation models as world simulators (2024), https://openai.com/research/video-generation-models-as-world-simulators" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.671, + 0.787, + 0.699 + ], + "angle": 0, + "content": "7. Cai, Y., Wang, J., Yuille, A., Zhou, Z., Wang, A.: Structure-aware sparse-view x-ray 3d reconstruction. In: CVPR (2024)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.7, + 0.787, + 0.742 + ], + "angle": 0, + "content": "8. Cao, A., Johnson, J.: Hexplane: A fast representation for dynamic scenes. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 130-141 (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.742, + 0.787, + 0.798 + ], + "angle": 0, + "content": "9. Chan, E.R., Lin, C.Z., Chan, M.A., Nagano, K., Pan, B., De Mello, S., Gallo, O., Guibas, L.J., Tremblay, J., Khamis, S., et al.: Efficient geometry-aware 3d generative adversarial networks. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 16123-16133 (2022)" + }, + { + "type": "ref_text", + "bbox": [ + 0.218, + 0.799, + 0.787, + 0.841 + ], + "angle": 0, + "content": "10. Chen, H.y., Tretschk, E., Stuyck, T., Kadlecek, P., Kavan, L., Vouga, E., Lassner, C.: Virtual elastic objects. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 15827-15837 (2022)" + }, + { + "type": "list", + "bbox": [ + 0.218, + 0.346, + 0.787, + 0.841 + ], + "angle": 0, + "content": null + } + ], + [ + { + "type": "page_number", + "bbox": [ + 0.218, + 0.116, + 0.236, + 0.127 + ], + "angle": 0, + "content": "16" + }, + { + "type": "header", + "bbox": [ + 0.272, + 0.115, + 0.354, + 0.129 + ], + "angle": 0, + "content": "Zhang et al." + }, + { + "type": "ref_text", + "bbox": [ + 0.218, + 0.147, + 0.787, + 0.189 + ], + "angle": 0, + "content": "11. Chen, X., Liu, Z., Chen, M., Feng, Y., Liu, Y., Shen, Y., Zhao, H.: Livephoto: Real image animation with text-guided motion control. arXiv preprint arXiv:2312.02928 (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.218, + 0.19, + 0.787, + 0.232 + ], + "angle": 0, + "content": "12. Chuang, Y.Y., Goldman, D.B., Zheng, K.C., Curless, B., Salesin, D.H., Szeliski, R.: Animating pictures with stochastic motion textures. In: ACM SIGGRAPH 2005 Papers. pp. 853-860 (2005)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.232, + 0.787, + 0.273 + ], + "angle": 0, + "content": "13. Curless, B., Levoy, M.: A volumetric method for building complex models from range images. In: Proceedings of the 23rd annual conference on Computer graphics and interactive techniques. pp. 303-312 (1996)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.273, + 0.787, + 0.314 + ], + "angle": 0, + "content": "14. Dai, Z., Zhang, Z., Yao, Y., Qiu, B., Zhu, S., Qin, L., Wang, W.: Animateanything: Fine-grained open domain image animation with motion guidance. arXiv e-prints pp. arXiv-2311 (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.315, + 0.787, + 0.355 + ], + "angle": 0, + "content": "15. Davis, A., Chen, J.G., Durand, F.: Image-space modal bases for plausible manipulation of objects in video. ACM Transactions on Graphics (TOG) 34(6), 1-7 (2015)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.356, + 0.787, + 0.384 + ], + "angle": 0, + "content": "16. Davis, M.A.: Visual vibration analysis. Ph.D. thesis, Massachusetts Institute of Technology (2016)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.384, + 0.787, + 0.424 + ], + "angle": 0, + "content": "17. Duan, Y., Wei, F., Dai, Q., He, Y., Chen, W., Chen, B.: 4d gaussian splatt-ting: Towards efficient novel view synthesis for dynamic scenes. arXiv preprint arXiv:2402.03307 (2024)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.425, + 0.787, + 0.452 + ], + "angle": 0, + "content": "18. Feng, Y., Shang, Y., Li, X., Shao, T., Jiang, C., Yang, Y.: Pie-nerf: Physics-based interactive elastodynamics with nerf. arXiv preprint arXiv:2311.13099 (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.453, + 0.787, + 0.508 + ], + "angle": 0, + "content": "19. Fridovich-Keil, S., Meanti, G., Warburg, F.R., Recht, B., Kanazawa, A.: K-planes: Explicit radiance fields in space, time, and appearance. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 12479-12488 (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.508, + 0.787, + 0.55 + ], + "angle": 0, + "content": "20. Gao, H., Li, R., Tulsiani, S., Russell, B., Kanazawa, A.: Monocular dynamic view synthesis: A reality check. Advances in Neural Information Processing Systems 35, 33768-33780 (2022)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.55, + 0.787, + 0.591 + ], + "angle": 0, + "content": "21. Geng, D., Owens, A.: Motion guidance: Diffusion-based image editing with differentiable motion estimators. In: The Twelfth International Conference on Learning Representations (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.591, + 0.787, + 0.633 + ], + "angle": 0, + "content": "22. Girdhar, R., Singh, M., Brown, A., Duval, Q., Azadi, S., Rambhatla, S.S., Shah, A., Yin, X., Parikh, D., Misra, I.: Emu video: Factorizing text-to-video generation by explicit image conditioning. arXiv preprint arXiv:2311.10709 (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.633, + 0.787, + 0.675 + ], + "angle": 0, + "content": "23. Guo, X., Sun, J., Dai, Y., Chen, G., Ye, X., Tan, X., Ding, E., Zhang, Y., Wang, J.: Forward flow for novel view synthesis of dynamic scenes. In: Proceedings of the IEEE/CVF International Conference on Computer Vision. pp. 16022-16033 (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.675, + 0.787, + 0.715 + ], + "angle": 0, + "content": "24. Gupta, A., Yu, L., Sohn, K., Gu, X., Hahn, M., Fei-Fei, L., Essa, I., Jiang, L., Lezama, J.: Photorealistic video generation with diffusion models. arXiv preprint arXiv:2312.06662 (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.716, + 0.787, + 0.758 + ], + "angle": 0, + "content": "25. Ho, J., Chan, W., Sahara, C., Whang, J., Gao, R., Gritsenko, A., Kingma, D.P., Poole, B., Norouzi, M., Fleet, D.J., et al.: Imagen video: High definition video generation with diffusion models. arXiv preprint arXiv:2210.02303 (2022)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.758, + 0.787, + 0.799 + ], + "angle": 0, + "content": "26. Hong, W., Ding, M., Zheng, W., Liu, X., Tang, J.: Cogvideo: Large-scale pretraining for text-to-video generation via transformers. arXiv preprint arXiv:2205.15868 (2022)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.799, + 0.787, + 0.842 + ], + "angle": 0, + "content": "27. Hu, Y., Fang, Y., Ge, Z., Qu, Z., Zhu, Y., Pradhana, A., Jiang, C.: A moving least squares material point method with displacement discontinuity and two-way rigid body coupling. ACM Transactions on Graphics (TOG) 37(4), 1-14 (2018)" + }, + { + "type": "list", + "bbox": [ + 0.217, + 0.147, + 0.787, + 0.842 + ], + "angle": 0, + "content": null + } + ], + [ + { + "type": "header", + "bbox": [ + 0.64, + 0.115, + 0.732, + 0.128 + ], + "angle": 0, + "content": "PhysDreamer" + }, + { + "type": "page_number", + "bbox": [ + 0.769, + 0.116, + 0.786, + 0.127 + ], + "angle": 0, + "content": "17" + }, + { + "type": "ref_text", + "bbox": [ + 0.214, + 0.147, + 0.788, + 0.189 + ], + "angle": 0, + "content": "28. Hu, Y., Li, T.M., Anderson, L., Ragan-Kelley, J., Durand, F.: Taichi: a language for high-performance computation on spatially sparse data structures. ACM Transactions on Graphics (TOG) 38(6), 1-16 (2019)" + }, + { + "type": "ref_text", + "bbox": [ + 0.214, + 0.19, + 0.787, + 0.232 + ], + "angle": 0, + "content": "29. Hu, Y., Li, T.M., Anderson, L., Ragan-Kelley, J., Durand, F.: Taichi: a language for high-performance computation on spatially sparse data structures. ACM Transactions on Graphics (TOG) 38(6), 1-16 (2019)" + }, + { + "type": "ref_text", + "bbox": [ + 0.216, + 0.233, + 0.786, + 0.274 + ], + "angle": 0, + "content": "30. Huang, Y.H., Sun, Y.T., Yang, Z., Lyu, X., Cao, Y.P., Qi, X.: Sc-gs: Sparse-controlled gaussian splatting for editable dynamic scenes. arXiv preprint arXiv:2312.14937 (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.216, + 0.276, + 0.787, + 0.303 + ], + "angle": 0, + "content": "31. Jiang, C., Gast, T., Teran, J.: Anisotropic elastoplasticity for cloth, knit and hair frictional contact. ACM Transactions on Graphics (TOG) 36(4), 1-14 (2017)" + }, + { + "type": "ref_text", + "bbox": [ + 0.216, + 0.304, + 0.787, + 0.332 + ], + "angle": 0, + "content": "32. Jiang, C., Schroeder, C., Selle, A., Teran, J., Stomakhin, A.: The affine particle-in-cell method. ACM Transactions on Graphics (TOG) 34(4), 1-10 (2015)" + }, + { + "type": "ref_text", + "bbox": [ + 0.216, + 0.333, + 0.787, + 0.374 + ], + "angle": 0, + "content": "33. Jiang, C., Schroeder, C., Teran, J., Stomakhin, A., Selle, A.: The material point method for simulating continuum materials. In: ACM SIGGRAPH 2016 courses. pp. 1-52 (2016)" + }, + { + "type": "ref_text", + "bbox": [ + 0.216, + 0.375, + 0.787, + 0.403 + ], + "angle": 0, + "content": "34. Kerbl, B., Kopanas, G., Leimkuhler, T., Drettakis, G.: 3d gaussian splatting for real-time radiance field rendering. ACM Transactions on Graphics 42(4) (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.216, + 0.403, + 0.787, + 0.445 + ], + "angle": 0, + "content": "35. Klár, G., Gast, T., Pradhana, A., Fu, C., Schroeder, C., Jiang, C., Teran, J.: Drucker-prager elastoplasticity for sand animation. ACM Transactions on Graphics (TOG) 35(4), 1-12 (2016)" + }, + { + "type": "ref_text", + "bbox": [ + 0.216, + 0.446, + 0.787, + 0.473 + ], + "angle": 0, + "content": "36. Kobayashi, M., Motoyoshi, I.: Perceiving natural speed in natural movies. i-Perception 10(4), 2041669519860544 (2019)" + }, + { + "type": "ref_text", + "bbox": [ + 0.216, + 0.474, + 0.787, + 0.515 + ], + "angle": 0, + "content": "37. Kondratyuk, D., Yu, L., Gu, X., Lezama, J., Huang, J., Hornung, R., Adam, H., Akbari, H., Alon, Y., Birodkar, V., et al.: Videopoet: A large language model for zero-shot video generation. arXiv preprint arXiv:2312.14125 (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.216, + 0.517, + 0.787, + 0.558 + ], + "angle": 0, + "content": "38. Kopf, J., Rong, X., Huang, J.B.: Robust consistent video depth estimation. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 1611-1621 (2021)" + }, + { + "type": "ref_text", + "bbox": [ + 0.216, + 0.559, + 0.787, + 0.6 + ], + "angle": 0, + "content": "39. Kratimenos, A., Lei, J., Daniilidis, K.: Dynmf: Neural motion factorization for real-time dynamic view synthesis with 3d gaussian splatting. arXiv preprint arXiv:2312.00112 (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.216, + 0.601, + 0.787, + 0.642 + ], + "angle": 0, + "content": "40. Le Cleac'h, S., Yu, H.X., Guo, M., Howell, T., Gao, R., Wu, J., Manchester, Z., Schwager, M.: Differentiable physics simulation of dynamics-augmented neural objects. IEEE Robotics and Automation Letters (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.216, + 0.643, + 0.787, + 0.685 + ], + "angle": 0, + "content": "41. Li, H., Sumner, R.W., Pauly, M.: Global correspondence optimization for non-rigid registration of depth scans. In: Computer graphics forum. vol. 27, pp. 1421-1430. Wiley Online Library (2008)" + }, + { + "type": "ref_text", + "bbox": [ + 0.216, + 0.686, + 0.787, + 0.728 + ], + "angle": 0, + "content": "42. Li, X., Qiao, Y.L., Chen, P.Y., Jatavallabhula, K.M., Lin, M., Jiang, C., Gan, C.: Pac-nerf: Physics augmented continuum neural radiance fields for geometry-agnostic system identification. arXiv preprint arXiv:2303.05512 (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.216, + 0.729, + 0.787, + 0.77 + ], + "angle": 0, + "content": "43. Li, Z., Niklaus, S., Snavely, N., Wang, O.: Neural scene flow fields for space-time view synthesis of dynamic scenes. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 6498-6508 (2021)" + }, + { + "type": "ref_text", + "bbox": [ + 0.216, + 0.771, + 0.787, + 0.798 + ], + "angle": 0, + "content": "44. Li, Z., Tucker, R., Snavely, N., Holynski, A.: Generative image dynamics. arXiv preprint arXiv:2309.07906 (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.216, + 0.799, + 0.787, + 0.841 + ], + "angle": 0, + "content": "45. Li, Z., Wang, Q., Cole, F., Tucker, R., Snavely, N.: Dynibar: Neural dynamic image-based rendering. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 4273-4284 (2023)" + }, + { + "type": "list", + "bbox": [ + 0.214, + 0.147, + 0.788, + 0.841 + ], + "angle": 0, + "content": null + } + ], + [ + { + "type": "page_number", + "bbox": [ + 0.218, + 0.116, + 0.236, + 0.127 + ], + "angle": 0, + "content": "18" + }, + { + "type": "header", + "bbox": [ + 0.272, + 0.115, + 0.354, + 0.129 + ], + "angle": 0, + "content": "Zhang et al." + }, + { + "type": "ref_text", + "bbox": [ + 0.218, + 0.147, + 0.785, + 0.189 + ], + "angle": 0, + "content": "46. Ling, H., Kim, S.W., Torralba, A., Fidler, S., Kreis, K.: Align your gaussians: Text-to-4d with dynamic 3d gaussians and composed diffusion models. arXiv preprint arXiv:2312.13763 (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.19, + 0.785, + 0.218 + ], + "angle": 0, + "content": "47. Luiten, J., Kopanas, G., Leibe, B., Ramanan, D.: Dynamic 3d gaussians: Tracking by persistent dynamic view synthesis. arXiv preprint arXiv:2308.09713 (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.218, + 0.785, + 0.259 + ], + "angle": 0, + "content": "48. Ma, P., Chen, P.Y., Deng, B., Tenenbaum, J.B., Du, T., Gan, C., Matusik, W.: Learning neural constitutive laws from motion observations for generalizable pde dynamics. In: International Conference on Machine Learning. PMLR (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.259, + 0.785, + 0.3 + ], + "angle": 0, + "content": "49. Macklin, M.: Warp: A high-performance python framework forgpu simulation and graphics. https://github.com/nvidia/warp (March 2022), nVIDIA GPU Technology Conference (GTC)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.3, + 0.785, + 0.342 + ], + "angle": 0, + "content": "50. Mildenhall, B., Srinivasan, P.P., Tancik, M., Barron, J.T., Ramamoorthi, R., Ng, R.: Nerf: Representing scenes as neural radiance fields for view synthesis. In: ECCV (2020)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.342, + 0.785, + 0.383 + ], + "angle": 0, + "content": "51. Newcombe, R.A., Fox, D., Seitz, S.M.: Dynamicfusion: Reconstruction and tracking of non-rigid scenes in real-time. In: Proceedings of the IEEE conference on computer vision and pattern recognition. pp. 343-352 (2015)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.383, + 0.785, + 0.425 + ], + "angle": 0, + "content": "52. Park, K., Sinha, U., Barron, J.T., Bouaziz, S., Goldman, D.B., Seitz, S.M., Martin-Brualla, R.: Nerfies: Deformable neural radiance fields. In: Proceedings of the IEEE/CVF International Conference on Computer Vision. pp. 5865-5874 (2021)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.425, + 0.785, + 0.467 + ], + "angle": 0, + "content": "53. Park, K., Sinha, U., Hedman, P., Barron, J.T., Bouaziz, S., Goldman, D.B., Martin-Brualla, R., Seitz, S.M.: Hypernerf: A higher-dimensional representation for topologically varying neural radiance fields. arXiv preprint arXiv:2106.13228 (2021)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.467, + 0.785, + 0.522 + ], + "angle": 0, + "content": "54. Peng, S., Niemeyer, M., Mescheder, L., Pollefeys, M., Geiger, A.: Convolutional occupancy networks. In: Computer Vision-ECCV 2020: 16th European Conference, Glasgow, UK, August 23-28, 2020, Proceedings, Part III 16. pp. 523-540. Springer (2020)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.522, + 0.785, + 0.563 + ], + "angle": 0, + "content": "55. Poole, B., Jain, A., Barron, J.T., Mildenhall, B.: Dreamfusion: Text-to-3d using 2d diffusion. In: The Eleventh International Conference on Learning Representations (2022)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.564, + 0.785, + 0.605 + ], + "angle": 0, + "content": "56. Pumarola, A., Corona, E., Pons-Moll, G., Moreno-Noguer, F.: D-nerf: Neural radiance fields for dynamic scenes. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 10318-10327 (2021)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.605, + 0.785, + 0.66 + ], + "angle": 0, + "content": "57. Ram, D., Gast, T., Jiang, C., Schroeder, C., Stomakhin, A., Teran, J., Kavehpour, P.: A material point method for viscoelastic fluids, foams and sponges. In: Proceedings of the 14th ACM SIGGRAPH/Eurographics Symposium on Computer Animation. pp. 157-163 (2015)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.66, + 0.785, + 0.688 + ], + "angle": 0, + "content": "58. Ren, J., Pan, L., Tang, J., Zhang, C., Cao, A., Zeng, G., Liu, Z.: Dreamgaussian4d: Generative 4d gaussian splatting. arXiv preprint arXiv:2312.17142 (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.688, + 0.785, + 0.73 + ], + "angle": 0, + "content": "59. Singer, U., Polyak, A., Hayes, T., Yin, X., An, J., Zhang, S., Hu, Q., Yang, H., Ashual, O., Gafni, O., et al.: Make-a-video: Text-to-video generation without text-video data. arXiv preprint arXiv:2209.14792 (2022)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.73, + 0.785, + 0.771 + ], + "angle": 0, + "content": "60. Singer, U., Sheynin, S., Polyak, A., Ashual, O., Makarov, I., Kokkinos, F., Goyal, N., Vedaldi, A., Parikh, D., Johnson, J., et al.: Text-to-4d dynamic scene generation. arXiv preprint arXiv:2301.11280 (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.771, + 0.785, + 0.799 + ], + "angle": 0, + "content": "61. Stocker, A.A., Simoncelli, E.P.: Noise characteristics and prior expectations in human visual speed perception. Nature neuroscience 9(4), 578-585 (2006)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.799, + 0.785, + 0.841 + ], + "angle": 0, + "content": "62. Stomakhin, A., Schroeder, C., Chai, L., Teran, J., Selle, A.: A material point method for snow simulation. ACM Transactions on Graphics (TOG) 32(4), 1-10 (2013)" + }, + { + "type": "list", + "bbox": [ + 0.217, + 0.147, + 0.785, + 0.841 + ], + "angle": 0, + "content": null + } + ], + [ + { + "type": "header", + "bbox": [ + 0.64, + 0.115, + 0.732, + 0.128 + ], + "angle": 0, + "content": "PhysDreamer" + }, + { + "type": "page_number", + "bbox": [ + 0.769, + 0.116, + 0.786, + 0.127 + ], + "angle": 0, + "content": "19" + }, + { + "type": "ref_text", + "bbox": [ + 0.214, + 0.147, + 0.788, + 0.175 + ], + "angle": 0, + "content": "63. Tang, J., Ren, J., Zhou, H., Liu, Z., Zeng, G.: Dreamgaussian: Generative gaussian splatting for efficient 3d content creation. arXiv preprint arXiv:2309.16653 (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.214, + 0.177, + 0.788, + 0.217 + ], + "angle": 0, + "content": "64. Unterthiner, T., Van Steenkiste, S., Kurach, K., Marinier, R., Michalski, M., Gelly, S.: Towards accurate generative models of video: A new metric & challenges. arXiv preprint arXiv:1812.01717 (2018)" + }, + { + "type": "ref_text", + "bbox": [ + 0.216, + 0.218, + 0.787, + 0.272 + ], + "angle": 0, + "content": "65. Villegas, R., Babaeizadeh, M., Kindermans, P.J., Moraldo, H., Zhang, H., Saffar, M.T., Castro, S., Kunze, J., Erhan, D.: Phenaki: Variable length video generation from open domain textual descriptions. In: International Conference on Learning Representations (2022)" + }, + { + "type": "ref_text", + "bbox": [ + 0.216, + 0.273, + 0.787, + 0.314 + ], + "angle": 0, + "content": "66. Wang, C., MacDonald, L.E., Jeni, L.A., Lucey, S.: Flow supervision for deformable nerf. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 21128-21137 (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.216, + 0.315, + 0.787, + 0.355 + ], + "angle": 0, + "content": "67. Wang, C., Zhuang, P., Siarohin, A., Cao, J., Qian, G., Lee, H.Y., Tulyakov, S.: Diffusion priors for dynamic view synthesis from monocular videos. arXiv preprint arXiv:2401.05583 (2024)" + }, + { + "type": "ref_text", + "bbox": [ + 0.216, + 0.357, + 0.787, + 0.397 + ], + "angle": 0, + "content": "68. Wu, C., Liang, J., Ji, L., Yang, F., Fang, Y., Jiang, D., Duan, N.: Niwa: Visual synthesis pre-training for neural visual world creation. In: European conference on computer vision. pp. 720-736. Springer (2022)" + }, + { + "type": "ref_text", + "bbox": [ + 0.216, + 0.398, + 0.787, + 0.438 + ], + "angle": 0, + "content": "69. Xian, W., Huang, J.B., Kopf, J., Kim, C.: Space-time neural irradiance fields for free-viewpoint video. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 9421-9431 (2021)" + }, + { + "type": "ref_text", + "bbox": [ + 0.216, + 0.439, + 0.787, + 0.481 + ], + "angle": 0, + "content": "70. Xie, T., Zong, Z., Qiu, Y., Li, X., Feng, Y., Yang, Y., Jiang, C.: Physgaus-sian: Physics-integrated 3d gaussians for generative dynamics. arXiv preprint arXiv:2311.12198 (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.216, + 0.481, + 0.787, + 0.522 + ], + "angle": 0, + "content": "71. Xie, Y., Takikawa, T., Saito, S., Litany, O., Yan, S., Khan, N., Tombari, F., Tompkin, J., Sitzmann, V., Sridhar, S.: Neural fields in visual computing and beyond. In: Computer Graphics Forum. vol. 41, pp. 641-676. Wiley Online Library (2022)" + }, + { + "type": "ref_text", + "bbox": [ + 0.216, + 0.523, + 0.787, + 0.563 + ], + "angle": 0, + "content": "72. Yang, Z., Gao, X., Zhou, W., Jiao, S., Zhang, Y., Jin, X.: Deformable 3d gaussians for high-fidelity monocular dynamic scene reconstruction. arXiv preprint arXiv:2309.13101 (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.216, + 0.565, + 0.787, + 0.591 + ], + "angle": 0, + "content": "73. Yin, Y., Xu, D., Wang, Z., Zhao, Y., Wei, Y.: 4dgen: Grounded 4d content generation with spatial-temporal consistency. arXiv preprint arXiv:2312.17225 (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.216, + 0.592, + 0.787, + 0.619 + ], + "angle": 0, + "content": "74. Yu, H., Julin, J., Milacski, Z.A., Niinuma, K., Jeni, L.A.: Cogs: Controllable gaussian splatting. arXiv preprint arXiv:2312.05664 (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.216, + 0.62, + 0.787, + 0.66 + ], + "angle": 0, + "content": "75. Yu, H., Julin, J., Milacski, Z.A., Niinuma, K., Jeni, L.A.: Dylin: Making light field networks dynamic. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 12397-12406 (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.216, + 0.661, + 0.787, + 0.702 + ], + "angle": 0, + "content": "76. Zhang, S., Wang, J., Zhang, Y., Zhao, K., Yuan, H., Qin, Z., Wang, X., Zhao, D., Zhou, J.: I2vgen-xl: High-quality image-to-video synthesis via cascaded diffusion models. arXiv preprint arXiv:2311.04145 (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.216, + 0.703, + 0.787, + 0.73 + ], + "angle": 0, + "content": "77. Zhang, Z., Cole, F., Tucker, R., Freeman, W.T., Dekel, T.: Consistent depth of moving objects in video. ACM Transactions on Graphics (TOG) 40(4), 1-12 (2021)" + }, + { + "type": "list", + "bbox": [ + 0.214, + 0.147, + 0.788, + 0.73 + ], + "angle": 0, + "content": null + } + ] +] \ No newline at end of file diff --git a/2024/Physics-Based Interaction with 3D Objects via Video Generation/017e5389-f4df-4a9c-a529-0be7433a235c_origin.pdf b/2024/Physics-Based Interaction with 3D Objects via Video Generation/017e5389-f4df-4a9c-a529-0be7433a235c_origin.pdf new file mode 100644 index 0000000000000000000000000000000000000000..ee9be2434a4fed8bdb1120a1a6189ee6211231fa --- /dev/null +++ b/2024/Physics-Based Interaction with 3D Objects via Video Generation/017e5389-f4df-4a9c-a529-0be7433a235c_origin.pdf @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:8c2c4f64f4e5a2cdac3fa7fb14b99cd55c8fb76e292d744fedbb0700034e963a +size 5286412 diff --git a/2024/Physics-Based Interaction with 3D Objects via Video Generation/full.md b/2024/Physics-Based Interaction with 3D Objects via Video Generation/full.md new file mode 100644 index 0000000000000000000000000000000000000000..ae51cf156d2722562883ededfc10fdcfa0ae0675 --- /dev/null +++ b/2024/Physics-Based Interaction with 3D Objects via Video Generation/full.md @@ -0,0 +1,372 @@ +# Physics-Based Interaction with 3D Objects via Video Generation + +Tianyuan Zhang $^{1}$ , Hong-Xing Yu $^{2}$ , Rundi Wu $^{3}$ , Brandon Y. Feng $^{1}$ , Changxi Zheng $^{3}$ , Noah Snavely $^{4}$ , Jiajun Wu $^{2}$ , and William T. Freeman $^{1}$ + +1 Massachusetts Institute of Technology + +2 Stanford University + +3 Columbia University + +4 Cornell University + +Abstract. Realistic object interactions are crucial for creating immersive virtual experiences, yet synthesizing realistic 3D object dynamics in response to novel interactions remains a significant challenge. Unlike unconditional or text-conditioned dynamics generation, action-conditioned dynamics requires perceiving the physical material properties of objects and grounding the 3D motion prediction on these properties, such as object stiffness. However, estimating physical material properties is an open problem due to the lack of material ground-truth data, as measuring these properties for real objects is highly difficult. We present PhysDreamer, a physics-based approach that endows static 3D objects with interactive dynamics by leveraging the object dynamics priors learned by video generation models. By distilling these priors, PhysDreamer enables the synthesis of realistic object responses to novel interactions, such as external forces or agent manipulations. We demonstrate our approach on diverse examples of elastic objects and evaluate the realism of the synthesized interactions through a user study. PhysDreamer takes a step towards more engaging and realistic virtual experiences by enabling static 3D objects to dynamically respond to interactive stimuli in a physically plausible manner. See our project page at https://physdreamer.github.io/. + +Keywords: Physics-based modeling $\cdot$ Interactive 3D dynamics + +# 1 Introduction + +Realistic object interactions play a pivotal role in creating immersive virtual experiences. Recent advances in 3D vision have enabled the capture and creation of high-quality static 3D assets [34,50], and some methods even extend to 4D assets [46,47,58], generating unconditioned dynamics. However, these methods fail to handle action-conditioned dynamics in response to new physical interactions, such as synthesizing the motion of a rose reacting to a breeze or a touch. + +The key challenge in synthesizing action-conditioned dynamics lies in understanding the physical material properties of objects. Yet, estimating these + +![](images/b6cf218f3de04570fd2e6afb26ad153d78b79bb65c175233a8d8b1a554215f09.jpg) +PhysDreamer + +![](images/979f1f2ac25d68411281b0a66c3d37e7f3ba55756d7595f5b67ce4f73d30b6ab.jpg) +Fig. 1: (Left) Leveraging and distilling dynamics priors from a pre-trained video generation model, we estimate a physical material field for the static 3D object. (Right) The physical material field allows synthesizing interactive 3D dynamics under arbitrary forces. We show rendered sequences from two viewpoints. Red arrows indicate force directions. Please see videos on our project website for better visualization. + +![](images/0eab3ab33285313cd3d460ca63d4649625a1da6c22c7bf04b263eb6070414436.jpg) +Synthesized 3D dynamics under interaction + +properties is a challenging task due to the lack of ground-truth data, as measuring these properties for real objects is highly difficult. Real-life objects often exhibit complex, spatially-varying material properties, making the estimation problem even more challenging. Despite the complexity of physical materials, humans can easily imagine how objects would react to external forces, such as the gentle sway of a rose. This ability to imagine object dynamics stems from our physical prior knowledge obtained from observing and interacting with the physical world. This motivates us to distill dynamics priors from video generation models that have been trained on vast, diverse video observations of the physical world. + +In this work, we focus on synthesizing interactive 3D dynamics. We propose PhysDreamer, a physics-based approach to transforming static 3D objects into interactive ones that can respond to novel interactions. The key idea behind PhysDreamer is to distill dynamics priors learned by video generation models to estimate the physical material properties of static 3D objects. We hypothesize that video generation models, trained on large amounts of video data, implicitly capture the relationship between object appearance and dynamics. By leveraging this learned prior knowledge, PhysDreamer can infer the physical material properties that drive the dynamic behavior of objects, even in the absence of ground-truth material data (Fig. 1). + +PhysDreamer represents 3D objects using 3D Gaussians [34], models the physical material field with a neural field [71], and simulates 3D dynamics using the differentiable Material Point Method (MPM) [27]. The differentiable simulation and rendering allow for direct optimization of the physical material field and initial velocity field by matching pixel space observations. We focus on elastic dynamics and showcase PhysDreamer through diverse real examples, such as flowers, plants, a beanie hat, and a telephone cord. We evaluate the realism of the synthesized interactive motion through a user study, comparing PhysDreamer + +to state-of-the-art methods. The results demonstrate that our approach significantly outperforms existing techniques on motion realism and visual quality. + +In summary, PhysDreamer addresses the challenge of synthesizing interactive 3D dynamics by leveraging the object dynamics priors learned by video generation models. By distilling these priors to estimate the physical material properties of static 3D objects, our approach enables the creation of immersive virtual experiences where objects can respond realistically to novel interactions. The main contributions of our work include enabling static 3D objects to dynamically respond to interactive stimuli in a physically plausible manner and taking a step towards more engaging and realistic virtual experiences. We believe that PhysDreamer has the potential to greatly enhance the realism and interactivity of virtual environments, paving the way for more engaging and lifelike simulations. + +# 2 Related work + +# 2.1 Dynamic 3D reconstruction + +Dynamic 3D reconstruction methods aim to reconstruct a representation of a dynamic scene from inputs such as depth scans [13,41], RGBD videos [51], or monocular or multi-view videos [1,7,39,45,47,52,53,56,66,69,74,75]. This task is especially challenging in the monocular setting with slow-moving cameras and fast-moving scenes [20]. Novel scene representations are a major driver of recent progress. One prominent approach is to augment a canonical Neural Radiance Fields (NeRF) with a deformation field [56]. This approach can be further improved by incorporating flow supervision [23,66] or as-rigid-as-possible or volume preserving regularization terms [52,53]. Time-modulated NeRFs [8,19,20,43] offer a simpler alternative representation. Due to its Lagrangian nature, 3D Gaussian Splatting [34] is readily adaptable to the task of efficient dynamic scene reconstruction [17,30,39,47,72,74]. Data-driven prior, such as from monocular depth models [38,77] and image diffusion models [67], can also be used to reduce the inherent ambiguity in dynamic reconstruction from monocular videos. + +# 2.2 Dynamic 3D generation + +Our work also relates to efforts to synthesize dynamic 3D scenes. A common approach is to integrate a 3D generation pipeline with a video generation model [2,46,58,60]. For instance, Make-A-Video3D begins by creating a static NeRF as per DreamFusion [55], then extending it temporally using Score Distillation Sampling (SDS) [55] derived from a video diffusion model. The approach can be improved with more efficient representations, stronger diffusion priors, and stable training techniques [2, 46]. However, applying SDS with video diffusion models demands significant computational and memory costs. Compact4D [73] and DreamGaussian4D [58] used a more efficient approach, synthesizing 3D dynamics by aligning a reference video from video generation models while employing SDS from image diffusion models to reduce novel view artifacts. These + +methods are currently limited to producing fixed-length 3D videos. We focus on synthesizing interactive 3D motions under any new physical interactions. + +# 2.3 Interactive motion generation + +Interactive motion generation animates still images or 3D contents according to user inputs like text [11, 76], motion fields [21], motion layers [12, 14], or direct manipulation such as dragging and pulling [15, 44]. Early work from Davis et al. [15, 16] demonstrated animating an image using an image-space modal basis extracted from a video of an object undergoing subtle vibrational motions. Building upon this image-space representation [15], Generative Image Dynamics [44] used a diffusion model trained on a dataset with paired image and its modal basis to model scene motion distributions, enabling realistic interaction with still input images. We focus on interacting with 3D objects rather than images. + +For 3D assets, physics-based approaches enable synthesizing motions under any physical interactions. Virtual Elastic Objects [10] jointly reconstructs the geometry, appearances, and physical parameters of elastic objects in a multiview capture setup with compressed air system. PAC-NeRF [42], DANO [40], and PhysGaussian [18] integrate physics-based simulations with NeRF and 3D Gaussians to generate physically plausible motions. We use the same physics-based approach to generate realistic interactions, but a novel ingredient of our work is to distill the material parameters of the object from pre-trained video generation models. + +# 2.4 Video generation models + +Recent progress in video generation is driven by the development of larger autoregressive [26,37,65,68] and diffusion models [3-6,22,24,25,59]. These models, trained on increasingly large datasets, continue to advance the quality and realism of generated video content. The state-of-the-art approach [6] can generate minute-long videos with realistic motions and viewpoint consistency. However, current video generation models cannot support physics-based interactions with objects through external forces. + +# 3 Problem formulation + +Given a static object represented by 3D Gaussians $\{\mathcal{G}_p\}_{p=1}^P$ , $\mathcal{G}_p = \{\boldsymbol{x}_p, \alpha_p, \boldsymbol{\Sigma}_p, \boldsymbol{c}_p\}$ (where $\boldsymbol{x}_p$ denotes the position, $\alpha_p$ denotes the opacity, $\boldsymbol{\Sigma}_p$ denotes the covariance matrix, and $\boldsymbol{c}_p$ denotes the color of the particle), our goal is to estimate physical material property fields for the object to enable realistic interactive motion synthesis. These properties include mass $m$ , Young's modulus $E$ , and Poisson's ratio $\nu$ . Among these physical properties, Young's modulus $E$ plays a particularly important role in determining the object's motion in response to applied forces. Intuitively, Young's modulus (Eq. 2) measures the material stiffness. A higher Young's modulus results in less deformation and more rigid and + +![](images/754050422b7712f0a1809b26eadd9beed8276563972be3ba00dd6bb4484f8f55.jpg) +3D Gaussians + +![](images/486db178e00153a3fd7a02d3245e4f4fd849c72197c224acfa0dd4cb79cec133.jpg) +10.30 seconds +Fig. 2: Effect of Young's modulus. We depict the motion of a simulated flower under the same external force but with three different Young's moduli, a measure of material stiffness. Flowers with the highest Young's modulus $(100\times)$ exhibit smaller oscillations and higher frequencies, while the flower with the lowest Young's modulus $(1\times)$ sways the most and oscillates at the lowest frequency. Time annotations below each image indicate the duration of one complete motion path shown in the figure. + +![](images/bc8f8798a4d32ddbd5c4a007f43684398f8c57f00a2e5a1ec14997d6f95e0060.jpg) +2.30 seconds + +![](images/40aa849ef07be213f45b6ae9c0d6f96148607c9f8299a945c1be340b2750566c.jpg) +0.87 seconds + +higher-frequency motion, while a lower value leads to more flexible and elastic behavior. Fig. 2 illustrates the simulated motion of a flower under the same applied forces but with different Young's modulus. + +Therefore, our problem formulation focuses on estimating the spatially varying Young's modulus field $E(\pmb{x})$ for the 3D object. To allow particle simulation, we query a particle's Young's modulus by $E_{p} = E(\pmb{x}_{p})$ . As for other physical properties, the mass for a particle $m_{p}$ can be pre-computed as the product of a constant density $(\rho)$ and particle volume $V_{p}$ . The particle volume can be estimated [70] by dividing a background cell's volume by the number of particles that cell contains. As for the Poisson's ratio $\nu_{p}$ , we found that it has negligible impact on object motion in our preliminary experiments (see supplementary materials for details), and so we assume a homogeneous constant Poisson's ratio. + +# 4 PhysDreamer + +PhysDreamer estimates a material field for a static 3D object. Our key idea is to generate a plausible video of the object in motion, and then optimize the material field $E(\pmb{x})$ to match this synthesized motion. We begin by rendering a static image $(I_0)$ for the 3D scene $\{\mathcal{G}_p\}$ from a certain viewpoint. We then leverage an image-to-video model to generate a short video clip $\{I_0, I_1, \dots, I_T\}$ depicting the object's realistic motion. This generated video serves as our reference video. We then optimize the material field $E(\pmb{x})$ and an initial velocity field $\pmb{v}_0(\pmb{x})$ (both modeled by implicit neural fields [71]) through differentiable simulation and differentiable rendering, such that a rendered video of the simulation matches (from the same viewpoint as $I_0$ ) the reference video. Fig. 3 shows an overview of PhysDreamer. + +![](images/7d8c8cc15f188422fc1779b27228e9f37526320ae31318ca7f35803bebc1b59d.jpg) +Fig. 3: Overview of PhysDreamer. Given an object represented as 3D Gaussians, we first render it (with background) from a viewpoint. Next, we use an image-to-video generation model to produce a reference video of that object in motion. Using differentiable Material Point Methods (MPM) and differentiable rendering, we optimize both a spatially-varying material field and an initial velocity field (not shown in the figure above). This optimization aims to minimize the discrepancy between the rendered video and the reference video. The dashed arrows represent gradient flow. + +# 4.1 Preliminaries + +3D Gaussians [34] adopts a set of anisotropic 3D Gaussian kernels to represent the radiance field of a 3D scene. Although introduced primarily as an efficient method for 3D novel view synthesis, the Lagrangian nature of 3D Gaussians also enables the direct adaptation of particle-based physics simulators. Following PhysGaussian [70], we use the Material Point Method (MPM) to simulate object dynamics directly on these Gaussian particles. Since 3D Gaussians mainly lie on object surfaces, an optional internal filling process can be applied for improved simulation realism [70]. Below, we provide a brief introduction on the underlying physical model and how to integrate MPM into 3D Gaussians. For a more comprehensive introduction of MPM, we refer interested readers to [27,32,33,70]. + +Continuum mechanics and elastic materials. Continuum mechanics models material deformation using a map $\phi$ that transforms points from the undeformed material space $\mathbf{X}$ to the deformed world space $\mathbf{x} = \phi(\mathbf{X}, t)$ . The Jacobian of the map, $\pmb{F} = \nabla_{\mathbf{X}}\phi(\mathbf{X}, t)$ , known as the deformation gradient, measures local rotation and strain. This tensor is crucial in formulating stress-strain relationship. For example, the Cauchy stress in a hyper-elastic material is computed by: $\pmb{\sigma} = \frac{1}{\operatorname*{det}(\pmb{F})}\frac{\partial\psi}{\partial\pmb{F}}\pmb{F}^T$ . Here, $\psi(\mathbf{F})$ represents the strain energy density function, quantifying the extent of non-rigid deformations. This function is typically designed by experts, to follow principles like material symmetry and rotational invariance while aligning with empirical data. In this work, we use fixed corotated hyperelastic model, whose energy density function can be expressed as: + +$$ +\psi (\mathbf {F}) = \mu \left(\sum_ {i = 1} ^ {d} \left(\sigma_ {i} - 1\right) ^ {2}\right) + \frac {\lambda}{2} (\det (\mathbf {F}) - 1) ^ {2}, \tag {1} +$$ + +where $\sigma_{i}$ denotes a singular value of the deformation gradient. $\mu$ and $\lambda$ are related to Young's modulus $E$ and Poisson's ratio $\nu$ via: + +$$ +\mu = \frac {E}{2 (1 + \nu)}, \quad \lambda = \frac {E \nu}{(1 + \nu) (1 - 2 \nu)}. \tag {2} +$$ + +The dynamics of an elastic object are governed by the following equations: + +$$ +\rho \frac {D \boldsymbol {v}}{D t} = \nabla \cdot \boldsymbol {\sigma} + \mathbf {f}, \quad \frac {D \rho}{D t} + \rho \nabla \cdot \boldsymbol {v} = 0, \tag {3} +$$ + +where $\rho$ denotes density, $\pmb{v}(\pmb{x},t)$ denotes the velocity field in world space, and $\mathbf{f}$ denotes an external force. + +Material Point Method (MPM). We use the Moving Least Squares Material Point Method (MLS-MPM) [27] to solve the above governing equation. MPM is a hybrid Eulerian-Langrangian method widely adopted for simulating dynamics for a wide range of materials, such as solid, fluid, sand, and cloth [31, 35, 57, 62]. MPM offers several advantages, such as easy GPU parallelization [28], handling of topology changes, and the availability of well-documented open-source implementations [29, 48, 49, 70]. + +Following PhysGaussian [70], we view the Gaussian particles as the spatial discretization of the object to be simulated, and directly run MPM on these Gaussian particles. Each particle $p$ represents a small volume of the object, and it carries a set of properties including volume $V_{p}$ , mass $m_{p}$ , position $\boldsymbol{x}_{p}^{t}$ , velocity $\boldsymbol{v}_{p}^{t}$ , deformation gradient $\boldsymbol{F}_{p}^{t}$ , and local velocity field gradient $\boldsymbol{C}_{p}^{t}$ at time step $t$ . + +MPM operates in a particle-to-grid (P2G) and grid-to-particle (G2P) transfer loop. In the P2G stage, we transfer the momentum from particle to grid by: + +$$ +m _ {i} ^ {t} \boldsymbol {v} _ {i} ^ {t} = \sum_ {p} N (\boldsymbol {x} _ {i} - \boldsymbol {x} _ {p} ^ {t}) [ m _ {p} \boldsymbol {v} _ {p} ^ {t} + (m _ {p} \boldsymbol {C} _ {p} ^ {t} - \frac {4}{(\Delta x) ^ {2}} \Delta t V _ {p} \frac {\partial \psi}{\partial \boldsymbol {F}} \boldsymbol {F} _ {p} ^ {t T}) (\boldsymbol {x} _ {i} - \boldsymbol {x} _ {p} ^ {t}) ] + \boldsymbol {f} _ {i} ^ {t}, \tag {4} +$$ + +where the mass of the grid node $i$ is $m_i^t = \sum_p N(\pmb{x}_i - \pmb{x}_p^t)m_p$ , $N(\pmb{x}_i - \pmb{x}_p^t)$ is the B-spline kernel, $\Delta x$ is the spatial grid resolution, $\Delta t$ is the simulation step size, and $\pmb{v}_i^t$ is the updated velocity on the grid. We then transfer the updated velocity back to the particles and update their positions as: + +$$ +\boldsymbol {v} _ {p} ^ {t + 1} = \sum_ {i} N (\boldsymbol {x} _ {i} - \boldsymbol {x} _ {p} ^ {t}) \boldsymbol {v} _ {i} ^ {t}, \quad \boldsymbol {x} _ {p} ^ {t + 1} = \boldsymbol {x} _ {p} ^ {t} + \Delta t \boldsymbol {v} _ {p} ^ {t + 1}. \tag {5} +$$ + +Meanwhile, the local velocity gradient and deformation gradient is updated as: + +$$ +\boldsymbol {C} _ {p} ^ {t + 1} = \frac {4}{(\Delta x) ^ {2}} \sum_ {i} N (\boldsymbol {x} _ {i} - \boldsymbol {x} _ {p} ^ {t}) \boldsymbol {v} _ {i} ^ {t} (\boldsymbol {x} _ {i} - \boldsymbol {x} _ {p} ^ {t}) ^ {T}, \quad \boldsymbol {F} _ {p} ^ {t + 1} = (\boldsymbol {I} + \Delta t \boldsymbol {C} _ {p} ^ {t + 1}) \boldsymbol {F} _ {p} ^ {t}. \tag {6} +$$ + +# 4.2 Estimating physical properties + +Using MLS-MPM [27, 70] as our physics simulator and the Fixed Corotated hyper-elastic material model for the 3D objects, the simulation process for a single sub-step is formalized as: + +$$ +\boldsymbol {x} ^ {t + 1}, \boldsymbol {v} ^ {t + 1}, \boldsymbol {F} ^ {t + 1}, \boldsymbol {C} ^ {t + 1} = \mathcal {S} \left(\boldsymbol {x} ^ {t}, \boldsymbol {v} ^ {t}, \boldsymbol {F} ^ {t}, \boldsymbol {C} ^ {t}, \boldsymbol {\theta}, \Delta t\right), \tag {7} +$$ + +where $\pmb{x}^t = [\pmb{x}_1^t, \dots, \pmb{x}_P^t]$ denotes the positions of all particles at time $t$ , and similarly $\pmb{v}^t = [\pmb{v}_1^t, \dots, \pmb{v}_P^t]$ denotes the velocities of all particles at time $t$ . $\pmb{F}^t$ and $\pmb{C}^t$ denote the deformation gradient and the gradient of local velocity fields for all particles, respectively. Both $\pmb{F}^t$ and $\pmb{C}^t$ are tracked for simulation purposes, not for rendering. $\pmb{\theta}$ denotes the collection of the physical properties of all particles: mass $\pmb{m} = [m_1, \dots, m_P]$ , Young's modulus $\pmb{E} = [E_1, \dots, E_P]$ , Poisson's ratio $\pmb{\nu} = [\nu_1, \dots, \nu_P]$ , and volume $\pmb{V} = [V_1, \dots, V_P]$ . $\Delta t$ is the simulation step size. + +We use a sub-step size $\Delta t \cong 1 \times 10^{-4}$ for most of our experiments. To simulate dynamics between adjacent video frames, we iterate over hundreds of sub-steps (time interval between frames are typically tens of milliseconds). For simplicity, we abuse notation to express a simulation step with $N$ sub-steps as: + +$$ +\boldsymbol {x} ^ {t + 1}, \boldsymbol {v} ^ {t + 1}, \boldsymbol {F} ^ {t + 1}, \boldsymbol {C} ^ {t + 1} = \mathcal {S} (\boldsymbol {x} ^ {t}, \boldsymbol {v} ^ {t}, \boldsymbol {F} ^ {t}, \boldsymbol {C} ^ {t}, \boldsymbol {\theta}, \Delta t, N), \tag {8} +$$ + +where the timestamp $t + 1$ is ahead of timestamp $t$ by $N\Delta t$ . After simulation, we render the Gaussians at each frame: + +$$ +\hat {I} ^ {t} = \mathcal {F} _ {\text {r e n d e r}} \left(\boldsymbol {x} ^ {t}, \boldsymbol {\alpha}, \boldsymbol {R} ^ {t}, \Sigma , \boldsymbol {c}\right), \tag {9} +$$ + +where $\mathcal{F}_{\mathrm{render}}$ denotes the differentiable rendering function, and $\pmb{R}^{t}$ denotes the rotation matrices of all particles obtained from the simulation step. + +Using the generated video as reference, we optimize the spatially-varying Young's modulus $\pmb{E}$ and an initial velocity $\pmb{v}^0$ by a per-frame loss function: + +$$ +L ^ {t} = \lambda L _ {1} \left(\hat {I} ^ {t}, I ^ {t}\right) + (1 - \lambda) L _ {\mathrm {D} - \mathrm {S S I M}} \left(\hat {I} ^ {t}, I ^ {t}\right), \tag {10} +$$ + +where we set $\lambda = 0.1$ in our experiments. + +We parameterize the material field and velocity field by two triplanes [9], each followed by a three-layer MLP. Additionally, we apply a total variation regularization for all spatial planes of both fields to encourage spatial smoothness. Using $\pmb{u}$ to denote one of the 2D spatial planes, and $\pmb{u}_{i,j}$ as a feature vector on the 2D plane, we write the total variation regularization term as: + +$$ +L _ {\mathrm {t v}} = \sum_ {i, j} \| \boldsymbol {u} _ {i + 1, j} - \boldsymbol {u} _ {i, j} \| _ {2} ^ {2} + \| \boldsymbol {u} _ {i, j + 1} - \boldsymbol {u} _ {i, j} \| _ {2} ^ {2}. \tag {11} +$$ + +Rather than optimizing the material parameters and initial velocity jointly, we split the optimization into two stages for better stability and faster convergence. In particular, in the first stage, we randomly initialize the Young's modulus for each Gaussian particle and freeze it. We optimize the initial velocity of each particle using only the first three frames of the reference video. In the second stage, we freeze the initial velocity and optimize the spatially varying Young's modulus. During the second stage, the gradient signal only flows to the previous frame to prevent gradient explosion/vanishing. + +![](images/f27687cab82aa0f9bfca75ad70f1374e2435ca40c028a8b9f0bcd26cd9200ba0.jpg) +Fig.4: Accelerated MPM with K-Means downsampling. We employ K-Means clustering to create a set of "driving particles" (in yellow) at the initial time step $(t = 0)$ . We only simulate these driving particles. When rendering, we obtain each particle's position and rotation by fitting a local rigid body transformation using neighboring driving particles. + +# 4.3 Accelerating simulation with subsampling + +High-fidelity rendering with 3D Gaussians typically requires millions of particles to represent a scene. Running simulations on all the particles poses a significant computational burden. To improve efficiency, we introduce a subsampling procedure for simulation, as illustrated in Fig. 4. + +Specifically, we apply K-Means clustering to create a set of driving particles $\{\mathcal{Q}_q\}_{q=1}^Q$ at $t = 0$ , where each driving particle is represented by $\mathcal{Q}_q^0 = \{\pmb{x}_q^0, \pmb{v}_q^0, \pmb{F}_q^0, \pmb{C}_q^0, E_q, m_q, \nu_q, V_q\}$ . The initial position of a driving particle $\pmb{x}_q^0$ is computed as the mean of the position $\pmb{x}_p$ of all cluster members. The number of the driving particles is much smaller than the number of 3D Gaussian particles, $Q \ll P$ . We run simulations only on the driving particles. During rendering, we compute the position and rotation for each 3D Gaussian particle $\mathcal{G}_p$ by interpolating the driving particles. In particular, for each 3D Gaussian particle, we find its eight nearest driving particles at $t = 0$ , and we fit a rigid body transformation $\pmb{T}$ between these eight driving particles at $t = 0$ and at the current timestamp. This rigid body transformation $\pmb{T}$ is applied to the initial position and rotation of the particle $\mathcal{G}_p$ to obtain its current position and rotation. We summarize our algorithm with pseudo-code in supplementary materials. + +# 5 Experiments + +# 5.1 Setup + +Datasets. We collect eight real-world static scenes by capturing multi-view images. Each scene includes an object and a background. The objects include five flowers (a red rose, a carnation, an orange rose, a tulip, and a white rose), an alocasia plant, a telephone cord, and a beanie hat. For each scene except for the red rose scene, we capture four interaction videos illustrating its natural motion after interaction, such as poking or dragging, and we use the real videos as additional comparison references. + +Baselines. We compare our approach to two baselines: PhysGaussian [70] and DreamGaussian4D [58]. PhysGaussian [70] integrates MPM simulation to static + +3D Gaussians to support simulation, but it cannot estimate material properties and relies on manually setting material parameter values. Thus, we use the same initialization strategy as ours to assign material properties for PhysGaussian. DreamGaussian4D [58] generates non-interactive dynamic 3D Gaussians from a static image. It first obtains a static 3D Gaussians using DreamGaussian [63], and then animate it by optimizing a deformation field from a generated driving video. For a fair comparison, we run its deformation field optimization on our reconstructed static 3D Gaussians, and we looped the resulting deformation field when rendering longer videos in later comparison. + +Evaluation metrics. We focus on the quality of the synthesized object motion, in particular, visual quality and motion realism. Therefore, we conduct a user study and adopt the Two-alternative Forced Choice (2AFC) protocol: the participants are shown two side-by-side synchronized videos, including one video result from ours and the other one from the competitor's, with a random left-right ordering. The participants are then asked to choose the one with higher visual quality and the one with higher motion realism. + +We recruited 100 participants, each asked to judge all 8 scenes, forming a total of 800 2AFC judgement samples for each baseline comparison. For each scene, we create 4 sample video pairs and show participants a random one from the 4 pairs. In particular, we create 4 five-second motion sequences using PhysDreamer with randomized initial conditions (applying an external force to the foreground object or assigning an initial velocity to the object), and render videos from randomly picked viewpoints. For the baseline method, we apply the same initial conditions (for PhysGaussian only) and render videos from the same viewpoint as ours to form the video pairs. Please see supplementary materials for human study details and quantitative metrics for videos (e.g., Fréchet Video Distance [64]). + +# 5.2 Implementation details + +Neural material fields. We represent both material field and initial velocity field using triplanes [54] each followed by a three-layer MLP. The triplanes have spatial resolutions of $8^{3}$ and $24^{3}$ for the material field and velocity field, respectively. + +3D Gaussian reconstruction. Similar to PhysGaussian [70], we employ anisotropic regularization to reduce skinny artifacts in the reconstruction. Each reconstructed scene contains 0.5 to 1.5 million particles (including foreground and background). + +Simulation details. For computational efficiency, we segment the background and keep only foreground object particles for simulation. In our experiments, the foreground object contains around 50 to 300 thousand 3D Gaussian particles. We then discretize the foreground into a $64^3$ grid. The number of driving particles are 10 to 50 times fewer than the number of 3D Gaussian particles, determined by maintaining an average of at least eight particles per occupied voxel. For accurate motion, we use 768 sub-steps between successive video frames, corresponding to a duration of $4.34 \times 10^{-5}$ second for each sub-step. To address the high memory consumption from large number of steps, we apply simulation + +![](images/946f253765b04ae38e1dc2c49e75f817e629d35516977c74d58349625fbfbc08.jpg) + +![](images/053cf1b7da5d2e83dd5ca37013eec6bda5e658e0a3361f5515c3641af945165e.jpg) + +![](images/0a09853279d10b314b3874f37167548f2d2f7ffde004df071fe8e6a77f088dc6.jpg) + +![](images/c3f9ad2754052b701c080b77525609ebe5c885c54afd03d0e843a691542b4c83.jpg) +Material result + +![](images/f52f355d1637ff52888eca5721f7d66d5b296aef2268ff58c9cee5633c5a7268.jpg) + +![](images/0f1576aae8aca2b0343ca8dde702b468c47e2557a5d99e8495d763ab01172f63.jpg) + +![](images/4bab3e82c0c15ae265bb2563b393a2c4a915af1e842bb3c5c0e36a09d628fe71.jpg) + +![](images/a5fdebb31b0f2ed5b2589b21307e488600fa3e859eec56945a20e8bb7345e7bc.jpg) +Synthesized 3D dynamics under interaction + +![](images/6c9d4c85d30d4774ce78b6e056bb37c23ba99c8e1f3ae167608bc389055cc4fa.jpg) + +![](images/337d341034383ccdf26e3a24cade1c7cdad3ec139ade2bab18c182c0d78dc01c.jpg) + +![](images/5716a94ce533fb9308ccb354cc453be3ef37a13eb0d5e1e3279537dea0674cc7.jpg) + +![](images/b7b74ebd85c4e1adb9f53a30edfc3c1475bf373e7809fa9944717115d83a8b80.jpg) + +![](images/02dbaa6ebc3d11669c44674fe95a763c81c3c9e3d83fd92291fbfec7da0929dd.jpg) + +![](images/a06f62dd64ef27937347af8c8bf97faead16694ec747269be5817bb9a20b0347.jpg) + +![](images/8d4efb73ec74dbb4774f57ff85479ac556475ea417956382742968b89b8a4675.jpg) + +![](images/43ca094db8f382473e70d106b35d711f79e879a8f18d9341048d8bd6853880cc.jpg) +Fig. 5: Interactive 3D dynamics synthesis. (Left) Visualization of the material fields. Brighter color indicates higher Young's modulus within each example. (Right) We apply an external force (red arrow) on each object, and the following columns demonstrate the object dynamics rendered at a static viewpoint. + +![](images/7e3125339de5c304c095b6dc7a7bf7109f8a2c0fe45c2161aaeccf906496d144.jpg) + +![](images/5f10fed9c66aad52ab6aa0224ec2db785d0a30f116f0fabcfde14aa3f884664d.jpg) + +![](images/311a2cc51be66ebcac18b5ec388b973555d744230d954b2b00a8227b2024959c.jpg) + +![](images/3c4a3fe009af3da3776db34998d3dc8c17fa89fc447387c6d734ca5f7d9a4656.jpg) + +state checkpointing and re-computation during gradient back-propagation. We add Dirichlet boundary conditions for stationary grid cells. We fill the internal volumes of certain solid objects to enhance simulation realism [70]. + +Generating reference videos. We render a 3D object with its background from a viewpoint, and then we use Stable Video Diffusion [4] to animate this rendered image and generate fourteen video frames. We use a small motion bucket number [4] (e.g., 5 or 8) so that the generated video contains mostly object motion and little camera motion. We use rendered images for the video generation, so that our approach can also be used for generated scenes. Also, rendering images directly from 3D Gaussians simplifies later optimization. + +# 5.3 Results + +We show our qualitative results of the spatially-varying Young's modulus in Fig. 5 (left), and simulated interactive motion in Fig. 5 (right). Please see our project website videos for a better motion visualization. Tab. 1 presents the user study results in comparison to baseline methods and real captured videos. + +Table 1: Human study 2AFC results of PhysDreamer (Ours) over real captured videos and baseline methods (PhysGaussian [70] and DreamGaussian4D [58]) on Motion Realism and overall Visual Quality. "Rose O", "Rose W", and "Rose R" denotes the orange, white, and red roses, respectively. + +
Motion realismAlocasiaCarnationHatRose ORose WRose RCordTulipAvg.
Ours over Real capture86%61%55%63%47%-29%35%53.7%
Ours over PhysGaussian96%89%57%91%93%73%61%86%80.8%
Ours over DreamGaussian75%77%51%78%51%41%71%64%63.5%
Visual quality
Ours over Real capture36%53%28%40%41%-29%34%37.3%
Ours over PhysGaussian67%69%50%75%73%58%58%70%65.0%
Ours over DreamGaussian82%75%74%76%60%47%76%70%70.0%
+ +Compared to PhysGaussian, $80.8\%$ of the human participant 2AFC samples prefer PhysDreamer (ours) in motion realism and $65.0\%$ prefer PhysDreamer in visual quality. Note that since the static scenes are the same, the visual quality also depends on the generated object motion. Fig. 6 shows temporal slices of the motion patterns. We observe that PhysGaussian produces large, unrealistic slow motion due to the lack of a principled estimation of material properties. + +Compared to DreamGaussian4D, $70.0\% / 63.5\%$ 2AFC samples prefer ours in visual quality and motion realism, respectively. From Fig. 6, we can observe that DreamGaussian4D generates periodic motion with a constant, small magnitude, while PhysDreamer can simulate the damping in motion. This is because DreamGaussian4D does not simulate the physical dynamics but simply distill a motion sequence from a generative model, so it cannot extrapolate to different motion. We further include one more evaluation dimension on "motion amount" comparing to DreamGaussian4D, where we ask the participants to judge which video has higher amount of motion, and $73.6\%$ 2AFC samples prefer PhysDreamer. + +Compared to real videos, $53.7\%$ 2AFC samples favored the motion realism of ours results. Interestingly, under "Motion Realism", $86\%$ of the users indicated that the alocasia outputs were more realistic than real captures. This is surprising, as one would expect a $50\%$ preference if the videos were indistinguishable. We offer a potential explanation: for thin geometries like alocasia leaves, the Material Point Method tends to produce lower-frequency and slower motions. This can be observed in the video and is evident in the space-time slice visualizations in Fig. 6. Humans are poor at judging the naturalness of motion and may be biased towards smoother and slower motions, as shown in prior studies [36,61]. + +# 5.4 Ablation: using multi-view reference videos + +For objects with self-occlusion, observing salient motion of all object parts from a single video is challenging (e.g., the alocasia scene where a leaf can occlude + +![](images/8417644f2a8bb83365516b07e17e7b2305058377a2f517a2871d2912989ce5cd.jpg) +Fig. 6: We compare our results with real captured videos, PhysGaussian [70], and DreamGaussian4D [58] using space-time slices. In these slices, the vertical axis represent time, and the horizontal axis shows a spatial slice of the object (denoted by red lines on the "object" column). These slices visualize the magnitude and frequencies of these oscillating motions. Results for our PhysDreamer (Ours) and PhysGaussian are simulated with the same initial conditions. + +![](images/c0c3bf17c0e32228530761b384ebe146d83190f582b74d2036ea615c1794193c.jpg) +Fig. 7: Comparison between single-view (top) and two-view (bottom) supervisions. The object (alocasia) exhibits self-occluding structures. We can use generated videos at two views to jointly optimize the material field. In the space-time (X-t) slices, the vertical axis represents time, and the horizontal axis shows a spatial slice of the object. + +another leaf). We may alleviate this problem by rendering from multiple viewpoints to provide comprehensive coverage of the object. Here, we use multiple videos in the material estimation, jointly optimizing a video-agnostic, spatially-varying Young's modulus for each particle along with video-specific initial velocities. From the comparison of the alocasia scene in Fig. 7, we can see that using multi-view reference videos (a front view and a back view) helps in such complex self-occluding objects: PhysDreamer benefits significantly from having supervision from two views, while using only a single view leads to artifacts. In our user study, $81.0\%$ 2AFC samples prefer results with two view supervision in visual quality and $86.0\%$ in motion realism. + +# 6 Conclusion + +In this work, we introduced PhysDreamer, a novel approach to synthesizing interactive 3D dynamics by endowing static 3D objects with physical material properties. Our method distills the object dynamics priors learned by video generation models to estimate the spatially-varying material properties. We showcase dynamics interaction with a diverse set of elastic objects by PhysDreamer. We believe that PhysDreamer takes a significant step towards creating more engaging and immersive virtual environments, opening up a wide range of applications from realistic simulations to interactive virtual experiences. + +Limitations. Our approach requires the user to manually specify the object to simulate and separate it from the background, and establish boundary conditions for stationary parts, like the pot of flowers. 3D object discovery may help for simulatable object extraction. In addition, our approach is computationally demanding. Despite our subsampling strategy, our current algorithm takes approximately one minute on a NVIDIA V100 GPU to produce a single second of video. Further improving efficiency remains an important future problem. Finally, in this work, we restrict our scope to elastic objects without collisions. + +Acknowledgements. This work is in part supported by the NSF PHY-2019786 (The NSF AI Institute for Artificial Intelligence and Fundamental Interactions, http://iaifi.org/), NSF CIF 1955864 (Occlusion and Directional Resolution in Computational Imaging), RI #2211258, #2338203, ONR MURI N00014-22-1-2740, Quanta Computer, Samsung, and United States Air Force Artificial Intelligence Accelerator and was accomplished under Cooperative Agreement Number FA8750-192-1000. We would like to thank Peter Yichen Chen, Zhengqi Li, Pingchuan Ma, Minghao Guo, Ge Yang, and Shai Avidan for help and insightful discussions. + +# References + +1. Attal, B., Huang, J.B., Richardt, C., Zollhoefer, M., Kopf, J., O'Toole, M., Kim, C.: Hyperreel: High-fidelity 6-dof video with ray-conditioned sampling. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 16610-16620 (2023) +2. Bahmani, S., Skorokhodov, I., Rong, V., Wetzstein, G., Guibas, L., Wonka, P., Tulyakov, S., Park, J.J., Tagliasacchi, A., Lindell, D.B.: 4d-fy: Text-to-4d generation using hybrid score distillation sampling. arXiv preprint arXiv:2311.17984 (2023) +3. Bar-Tal, O., Chefer, H., Tov, O., Herrmann, C., Paiss, R., Zada, S., Ephrat, A., Hur, J., Li, Y., Michaeli, T., et al.: Lumiere: A space-time diffusion model for video generation. arXiv preprint arXiv:2401.12945 (2024) +4. Blattmann, A., Dockhorn, T., Kulal, S., Mendelevitch, D., Kilian, M., Lorenz, D., Levi, Y., English, Z., Voleti, V., Letts, A., et al.: Stable video diffusion: Scaling latent video diffusion models to large datasets. arXiv preprint arXiv:2311.15127 (2023) +5. Blattmann, A., Rombach, R., Ling, H., Dockhorn, T., Kim, S.W., Fidler, S., Kreis, K.: Align your latents: High-resolution video synthesis with latent diffusion models. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 22563-22575 (2023) +6. Brooks, T., Peebles, B., Homes, C., DePue, W., Guo, Y., Jing, L., Schnurr, D., Taylor, J., Luhman, T., Luhman, E., Ng, C., Wang, R., Ramesh, A.: Video generation models as world simulators (2024), https://openai.com/research/video-generation-models-as-world-simulators +7. Cai, Y., Wang, J., Yuille, A., Zhou, Z., Wang, A.: Structure-aware sparse-view x-ray 3d reconstruction. In: CVPR (2024) +8. Cao, A., Johnson, J.: Hexplane: A fast representation for dynamic scenes. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 130-141 (2023) +9. Chan, E.R., Lin, C.Z., Chan, M.A., Nagano, K., Pan, B., De Mello, S., Gallo, O., Guibas, L.J., Tremblay, J., Khamis, S., et al.: Efficient geometry-aware 3d generative adversarial networks. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 16123-16133 (2022) +10. Chen, H.y., Tretschk, E., Stuyck, T., Kadlecek, P., Kavan, L., Vouga, E., Lassner, C.: Virtual elastic objects. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 15827-15837 (2022) + +11. Chen, X., Liu, Z., Chen, M., Feng, Y., Liu, Y., Shen, Y., Zhao, H.: Livephoto: Real image animation with text-guided motion control. arXiv preprint arXiv:2312.02928 (2023) +12. Chuang, Y.Y., Goldman, D.B., Zheng, K.C., Curless, B., Salesin, D.H., Szeliski, R.: Animating pictures with stochastic motion textures. In: ACM SIGGRAPH 2005 Papers. pp. 853-860 (2005) +13. Curless, B., Levoy, M.: A volumetric method for building complex models from range images. In: Proceedings of the 23rd annual conference on Computer graphics and interactive techniques. pp. 303-312 (1996) +14. Dai, Z., Zhang, Z., Yao, Y., Qiu, B., Zhu, S., Qin, L., Wang, W.: Animateanything: Fine-grained open domain image animation with motion guidance. arXiv e-prints pp. arXiv-2311 (2023) +15. Davis, A., Chen, J.G., Durand, F.: Image-space modal bases for plausible manipulation of objects in video. ACM Transactions on Graphics (TOG) 34(6), 1-7 (2015) +16. Davis, M.A.: Visual vibration analysis. Ph.D. thesis, Massachusetts Institute of Technology (2016) +17. Duan, Y., Wei, F., Dai, Q., He, Y., Chen, W., Chen, B.: 4d gaussian splatt-ting: Towards efficient novel view synthesis for dynamic scenes. arXiv preprint arXiv:2402.03307 (2024) +18. Feng, Y., Shang, Y., Li, X., Shao, T., Jiang, C., Yang, Y.: Pie-nerf: Physics-based interactive elastodynamics with nerf. arXiv preprint arXiv:2311.13099 (2023) +19. Fridovich-Keil, S., Meanti, G., Warburg, F.R., Recht, B., Kanazawa, A.: K-planes: Explicit radiance fields in space, time, and appearance. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 12479-12488 (2023) +20. Gao, H., Li, R., Tulsiani, S., Russell, B., Kanazawa, A.: Monocular dynamic view synthesis: A reality check. Advances in Neural Information Processing Systems 35, 33768-33780 (2022) +21. Geng, D., Owens, A.: Motion guidance: Diffusion-based image editing with differentiable motion estimators. In: The Twelfth International Conference on Learning Representations (2023) +22. Girdhar, R., Singh, M., Brown, A., Duval, Q., Azadi, S., Rambhatla, S.S., Shah, A., Yin, X., Parikh, D., Misra, I.: Emu video: Factorizing text-to-video generation by explicit image conditioning. arXiv preprint arXiv:2311.10709 (2023) +23. Guo, X., Sun, J., Dai, Y., Chen, G., Ye, X., Tan, X., Ding, E., Zhang, Y., Wang, J.: Forward flow for novel view synthesis of dynamic scenes. In: Proceedings of the IEEE/CVF International Conference on Computer Vision. pp. 16022-16033 (2023) +24. Gupta, A., Yu, L., Sohn, K., Gu, X., Hahn, M., Fei-Fei, L., Essa, I., Jiang, L., Lezama, J.: Photorealistic video generation with diffusion models. arXiv preprint arXiv:2312.06662 (2023) +25. Ho, J., Chan, W., Sahara, C., Whang, J., Gao, R., Gritsenko, A., Kingma, D.P., Poole, B., Norouzi, M., Fleet, D.J., et al.: Imagen video: High definition video generation with diffusion models. arXiv preprint arXiv:2210.02303 (2022) +26. Hong, W., Ding, M., Zheng, W., Liu, X., Tang, J.: Cogvideo: Large-scale pretraining for text-to-video generation via transformers. arXiv preprint arXiv:2205.15868 (2022) +27. Hu, Y., Fang, Y., Ge, Z., Qu, Z., Zhu, Y., Pradhana, A., Jiang, C.: A moving least squares material point method with displacement discontinuity and two-way rigid body coupling. ACM Transactions on Graphics (TOG) 37(4), 1-14 (2018) + +28. Hu, Y., Li, T.M., Anderson, L., Ragan-Kelley, J., Durand, F.: Taichi: a language for high-performance computation on spatially sparse data structures. ACM Transactions on Graphics (TOG) 38(6), 1-16 (2019) +29. Hu, Y., Li, T.M., Anderson, L., Ragan-Kelley, J., Durand, F.: Taichi: a language for high-performance computation on spatially sparse data structures. ACM Transactions on Graphics (TOG) 38(6), 1-16 (2019) +30. Huang, Y.H., Sun, Y.T., Yang, Z., Lyu, X., Cao, Y.P., Qi, X.: Sc-gs: Sparse-controlled gaussian splatting for editable dynamic scenes. arXiv preprint arXiv:2312.14937 (2023) +31. Jiang, C., Gast, T., Teran, J.: Anisotropic elastoplasticity for cloth, knit and hair frictional contact. ACM Transactions on Graphics (TOG) 36(4), 1-14 (2017) +32. Jiang, C., Schroeder, C., Selle, A., Teran, J., Stomakhin, A.: The affine particle-in-cell method. ACM Transactions on Graphics (TOG) 34(4), 1-10 (2015) +33. Jiang, C., Schroeder, C., Teran, J., Stomakhin, A., Selle, A.: The material point method for simulating continuum materials. In: ACM SIGGRAPH 2016 courses. pp. 1-52 (2016) +34. Kerbl, B., Kopanas, G., Leimkuhler, T., Drettakis, G.: 3d gaussian splatting for real-time radiance field rendering. ACM Transactions on Graphics 42(4) (2023) +35. Klár, G., Gast, T., Pradhana, A., Fu, C., Schroeder, C., Jiang, C., Teran, J.: Drucker-prager elastoplasticity for sand animation. ACM Transactions on Graphics (TOG) 35(4), 1-12 (2016) +36. Kobayashi, M., Motoyoshi, I.: Perceiving natural speed in natural movies. i-Perception 10(4), 2041669519860544 (2019) +37. Kondratyuk, D., Yu, L., Gu, X., Lezama, J., Huang, J., Hornung, R., Adam, H., Akbari, H., Alon, Y., Birodkar, V., et al.: Videopoet: A large language model for zero-shot video generation. arXiv preprint arXiv:2312.14125 (2023) +38. Kopf, J., Rong, X., Huang, J.B.: Robust consistent video depth estimation. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 1611-1621 (2021) +39. Kratimenos, A., Lei, J., Daniilidis, K.: Dynmf: Neural motion factorization for real-time dynamic view synthesis with 3d gaussian splatting. arXiv preprint arXiv:2312.00112 (2023) +40. Le Cleac'h, S., Yu, H.X., Guo, M., Howell, T., Gao, R., Wu, J., Manchester, Z., Schwager, M.: Differentiable physics simulation of dynamics-augmented neural objects. IEEE Robotics and Automation Letters (2023) +41. Li, H., Sumner, R.W., Pauly, M.: Global correspondence optimization for non-rigid registration of depth scans. In: Computer graphics forum. vol. 27, pp. 1421-1430. Wiley Online Library (2008) +42. Li, X., Qiao, Y.L., Chen, P.Y., Jatavallabhula, K.M., Lin, M., Jiang, C., Gan, C.: Pac-nerf: Physics augmented continuum neural radiance fields for geometry-agnostic system identification. arXiv preprint arXiv:2303.05512 (2023) +43. Li, Z., Niklaus, S., Snavely, N., Wang, O.: Neural scene flow fields for space-time view synthesis of dynamic scenes. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 6498-6508 (2021) +44. Li, Z., Tucker, R., Snavely, N., Holynski, A.: Generative image dynamics. arXiv preprint arXiv:2309.07906 (2023) +45. Li, Z., Wang, Q., Cole, F., Tucker, R., Snavely, N.: Dynibar: Neural dynamic image-based rendering. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 4273-4284 (2023) + +46. Ling, H., Kim, S.W., Torralba, A., Fidler, S., Kreis, K.: Align your gaussians: Text-to-4d with dynamic 3d gaussians and composed diffusion models. arXiv preprint arXiv:2312.13763 (2023) +47. Luiten, J., Kopanas, G., Leibe, B., Ramanan, D.: Dynamic 3d gaussians: Tracking by persistent dynamic view synthesis. arXiv preprint arXiv:2308.09713 (2023) +48. Ma, P., Chen, P.Y., Deng, B., Tenenbaum, J.B., Du, T., Gan, C., Matusik, W.: Learning neural constitutive laws from motion observations for generalizable pde dynamics. In: International Conference on Machine Learning. PMLR (2023) +49. Macklin, M.: Warp: A high-performance python framework forgpu simulation and graphics. https://github.com/nvidia/warp (March 2022), nVIDIA GPU Technology Conference (GTC) +50. Mildenhall, B., Srinivasan, P.P., Tancik, M., Barron, J.T., Ramamoorthi, R., Ng, R.: Nerf: Representing scenes as neural radiance fields for view synthesis. In: ECCV (2020) +51. Newcombe, R.A., Fox, D., Seitz, S.M.: Dynamicfusion: Reconstruction and tracking of non-rigid scenes in real-time. In: Proceedings of the IEEE conference on computer vision and pattern recognition. pp. 343-352 (2015) +52. Park, K., Sinha, U., Barron, J.T., Bouaziz, S., Goldman, D.B., Seitz, S.M., Martin-Brualla, R.: Nerfies: Deformable neural radiance fields. In: Proceedings of the IEEE/CVF International Conference on Computer Vision. pp. 5865-5874 (2021) +53. Park, K., Sinha, U., Hedman, P., Barron, J.T., Bouaziz, S., Goldman, D.B., Martin-Brualla, R., Seitz, S.M.: Hypernerf: A higher-dimensional representation for topologically varying neural radiance fields. arXiv preprint arXiv:2106.13228 (2021) +54. Peng, S., Niemeyer, M., Mescheder, L., Pollefeys, M., Geiger, A.: Convolutional occupancy networks. In: Computer Vision-ECCV 2020: 16th European Conference, Glasgow, UK, August 23-28, 2020, Proceedings, Part III 16. pp. 523-540. Springer (2020) +55. Poole, B., Jain, A., Barron, J.T., Mildenhall, B.: Dreamfusion: Text-to-3d using 2d diffusion. In: The Eleventh International Conference on Learning Representations (2022) +56. Pumarola, A., Corona, E., Pons-Moll, G., Moreno-Noguer, F.: D-nerf: Neural radiance fields for dynamic scenes. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 10318-10327 (2021) +57. Ram, D., Gast, T., Jiang, C., Schroeder, C., Stomakhin, A., Teran, J., Kavehpour, P.: A material point method for viscoelastic fluids, foams and sponges. In: Proceedings of the 14th ACM SIGGRAPH/Eurographics Symposium on Computer Animation. pp. 157-163 (2015) +58. Ren, J., Pan, L., Tang, J., Zhang, C., Cao, A., Zeng, G., Liu, Z.: Dreamgaussian4d: Generative 4d gaussian splatting. arXiv preprint arXiv:2312.17142 (2023) +59. Singer, U., Polyak, A., Hayes, T., Yin, X., An, J., Zhang, S., Hu, Q., Yang, H., Ashual, O., Gafni, O., et al.: Make-a-video: Text-to-video generation without text-video data. arXiv preprint arXiv:2209.14792 (2022) +60. Singer, U., Sheynin, S., Polyak, A., Ashual, O., Makarov, I., Kokkinos, F., Goyal, N., Vedaldi, A., Parikh, D., Johnson, J., et al.: Text-to-4d dynamic scene generation. arXiv preprint arXiv:2301.11280 (2023) +61. Stocker, A.A., Simoncelli, E.P.: Noise characteristics and prior expectations in human visual speed perception. Nature neuroscience 9(4), 578-585 (2006) +62. Stomakhin, A., Schroeder, C., Chai, L., Teran, J., Selle, A.: A material point method for snow simulation. ACM Transactions on Graphics (TOG) 32(4), 1-10 (2013) + +63. Tang, J., Ren, J., Zhou, H., Liu, Z., Zeng, G.: Dreamgaussian: Generative gaussian splatting for efficient 3d content creation. arXiv preprint arXiv:2309.16653 (2023) +64. Unterthiner, T., Van Steenkiste, S., Kurach, K., Marinier, R., Michalski, M., Gelly, S.: Towards accurate generative models of video: A new metric & challenges. arXiv preprint arXiv:1812.01717 (2018) +65. Villegas, R., Babaeizadeh, M., Kindermans, P.J., Moraldo, H., Zhang, H., Saffar, M.T., Castro, S., Kunze, J., Erhan, D.: Phenaki: Variable length video generation from open domain textual descriptions. In: International Conference on Learning Representations (2022) +66. Wang, C., MacDonald, L.E., Jeni, L.A., Lucey, S.: Flow supervision for deformable nerf. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 21128-21137 (2023) +67. Wang, C., Zhuang, P., Siarohin, A., Cao, J., Qian, G., Lee, H.Y., Tulyakov, S.: Diffusion priors for dynamic view synthesis from monocular videos. arXiv preprint arXiv:2401.05583 (2024) +68. Wu, C., Liang, J., Ji, L., Yang, F., Fang, Y., Jiang, D., Duan, N.: Niwa: Visual synthesis pre-training for neural visual world creation. In: European conference on computer vision. pp. 720-736. Springer (2022) +69. Xian, W., Huang, J.B., Kopf, J., Kim, C.: Space-time neural irradiance fields for free-viewpoint video. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 9421-9431 (2021) +70. Xie, T., Zong, Z., Qiu, Y., Li, X., Feng, Y., Yang, Y., Jiang, C.: Physgaus-sian: Physics-integrated 3d gaussians for generative dynamics. arXiv preprint arXiv:2311.12198 (2023) +71. Xie, Y., Takikawa, T., Saito, S., Litany, O., Yan, S., Khan, N., Tombari, F., Tompkin, J., Sitzmann, V., Sridhar, S.: Neural fields in visual computing and beyond. In: Computer Graphics Forum. vol. 41, pp. 641-676. Wiley Online Library (2022) +72. Yang, Z., Gao, X., Zhou, W., Jiao, S., Zhang, Y., Jin, X.: Deformable 3d gaussians for high-fidelity monocular dynamic scene reconstruction. arXiv preprint arXiv:2309.13101 (2023) +73. Yin, Y., Xu, D., Wang, Z., Zhao, Y., Wei, Y.: 4dgen: Grounded 4d content generation with spatial-temporal consistency. arXiv preprint arXiv:2312.17225 (2023) +74. Yu, H., Julin, J., Milacski, Z.A., Niinuma, K., Jeni, L.A.: Cogs: Controllable gaussian splatting. arXiv preprint arXiv:2312.05664 (2023) +75. Yu, H., Julin, J., Milacski, Z.A., Niinuma, K., Jeni, L.A.: Dylin: Making light field networks dynamic. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 12397-12406 (2023) +76. Zhang, S., Wang, J., Zhang, Y., Zhao, K., Yuan, H., Qin, Z., Wang, X., Zhao, D., Zhou, J.: I2vgen-xl: High-quality image-to-video synthesis via cascaded diffusion models. arXiv preprint arXiv:2311.04145 (2023) +77. Zhang, Z., Cole, F., Tucker, R., Freeman, W.T., Dekel, T.: Consistent depth of moving objects in video. ACM Transactions on Graphics (TOG) 40(4), 1-12 (2021) \ No newline at end of file diff --git a/2024/Physics-Based Interaction with 3D Objects via Video Generation/images.zip b/2024/Physics-Based Interaction with 3D Objects via Video Generation/images.zip new file mode 100644 index 0000000000000000000000000000000000000000..d2d7bdb2be00e15d122584e315ec07eb0d7180fb --- /dev/null +++ b/2024/Physics-Based Interaction with 3D Objects via Video Generation/images.zip @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:badb46b4fade158d527c5ab189b126631b6c8b3238e1559d956c6a6e29fbcdad +size 651893 diff --git a/2024/Physics-Based Interaction with 3D Objects via Video Generation/layout.json b/2024/Physics-Based Interaction with 3D Objects via Video Generation/layout.json new file mode 100644 index 0000000000000000000000000000000000000000..ad87ee48e6c71c5609266d2fa2173137ab49d023 --- /dev/null +++ b/2024/Physics-Based Interaction with 3D Objects via Video Generation/layout.json @@ -0,0 +1,11365 @@ +{ + "pdf_info": [ + { + "para_blocks": [ + { + "bbox": [ + 157, + 111, + 459, + 146 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 157, + 111, + 459, + 146 + ], + "spans": [ + { + "bbox": [ + 157, + 111, + 459, + 146 + ], + "type": "text", + "content": "Physics-Based Interaction with 3D Objects via Video Generation" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 147, + 168, + 465, + 194 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 147, + 168, + 465, + 194 + ], + "spans": [ + { + "bbox": [ + 147, + 168, + 465, + 194 + ], + "type": "text", + "content": "Tianyuan Zhang" + }, + { + "bbox": [ + 147, + 168, + 465, + 194 + ], + "type": "inline_equation", + "content": "^{1}" + }, + { + "bbox": [ + 147, + 168, + 465, + 194 + ], + "type": "text", + "content": ", Hong-Xing Yu" + }, + { + "bbox": [ + 147, + 168, + 465, + 194 + ], + "type": "inline_equation", + "content": "^{2}" + }, + { + "bbox": [ + 147, + 168, + 465, + 194 + ], + "type": "text", + "content": ", Rundi Wu" + }, + { + "bbox": [ + 147, + 168, + 465, + 194 + ], + "type": "inline_equation", + "content": "^{3}" + }, + { + "bbox": [ + 147, + 168, + 465, + 194 + ], + "type": "text", + "content": ", Brandon Y. Feng" + }, + { + "bbox": [ + 147, + 168, + 465, + 194 + ], + "type": "inline_equation", + "content": "^{1}" + }, + { + "bbox": [ + 147, + 168, + 465, + 194 + ], + "type": "text", + "content": ", Changxi Zheng" + }, + { + "bbox": [ + 147, + 168, + 465, + 194 + ], + "type": "inline_equation", + "content": "^{3}" + }, + { + "bbox": [ + 147, + 168, + 465, + 194 + ], + "type": "text", + "content": ", Noah Snavely" + }, + { + "bbox": [ + 147, + 168, + 465, + 194 + ], + "type": "inline_equation", + "content": "^{4}" + }, + { + "bbox": [ + 147, + 168, + 465, + 194 + ], + "type": "text", + "content": ", Jiajun Wu" + }, + { + "bbox": [ + 147, + 168, + 465, + 194 + ], + "type": "inline_equation", + "content": "^{2}" + }, + { + "bbox": [ + 147, + 168, + 465, + 194 + ], + "type": "text", + "content": ", and William T. Freeman" + }, + { + "bbox": [ + 147, + 168, + 465, + 194 + ], + "type": "inline_equation", + "content": "^{1}" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 223, + 201, + 389, + 213 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 223, + 201, + 389, + 213 + ], + "spans": [ + { + "bbox": [ + 223, + 201, + 389, + 213 + ], + "type": "text", + "content": "1 Massachusetts Institute of Technology" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 261, + 213, + 351, + 224 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 261, + 213, + 351, + 224 + ], + "spans": [ + { + "bbox": [ + 261, + 213, + 351, + 224 + ], + "type": "text", + "content": "2 Stanford University" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 261, + 224, + 353, + 235 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 261, + 224, + 353, + 235 + ], + "spans": [ + { + "bbox": [ + 261, + 224, + 353, + 235 + ], + "type": "text", + "content": "3 Columbia University" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 264, + 235, + 348, + 247 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 264, + 235, + 348, + 247 + ], + "spans": [ + { + "bbox": [ + 264, + 235, + 348, + 247 + ], + "type": "text", + "content": "4 Cornell University" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 159, + 278, + 455, + 499 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 159, + 278, + 455, + 499 + ], + "spans": [ + { + "bbox": [ + 159, + 278, + 455, + 499 + ], + "type": "text", + "content": "Abstract. Realistic object interactions are crucial for creating immersive virtual experiences, yet synthesizing realistic 3D object dynamics in response to novel interactions remains a significant challenge. Unlike unconditional or text-conditioned dynamics generation, action-conditioned dynamics requires perceiving the physical material properties of objects and grounding the 3D motion prediction on these properties, such as object stiffness. However, estimating physical material properties is an open problem due to the lack of material ground-truth data, as measuring these properties for real objects is highly difficult. We present PhysDreamer, a physics-based approach that endows static 3D objects with interactive dynamics by leveraging the object dynamics priors learned by video generation models. By distilling these priors, PhysDreamer enables the synthesis of realistic object responses to novel interactions, such as external forces or agent manipulations. We demonstrate our approach on diverse examples of elastic objects and evaluate the realism of the synthesized interactions through a user study. PhysDreamer takes a step towards more engaging and realistic virtual experiences by enabling static 3D objects to dynamically respond to interactive stimuli in a physically plausible manner. See our project page at https://physdreamer.github.io/." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 160, + 510, + 420, + 521 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 160, + 510, + 420, + 521 + ], + "spans": [ + { + "bbox": [ + 160, + 510, + 420, + 521 + ], + "type": "text", + "content": "Keywords: Physics-based modeling " + }, + { + "bbox": [ + 160, + 510, + 420, + 521 + ], + "type": "inline_equation", + "content": "\\cdot" + }, + { + "bbox": [ + 160, + 510, + 420, + 521 + ], + "type": "text", + "content": " Interactive 3D dynamics" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 132, + 542, + 230, + 555 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 542, + 230, + 555 + ], + "spans": [ + { + "bbox": [ + 132, + 542, + 230, + 555 + ], + "type": "text", + "content": "1 Introduction" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 130, + 569, + 482, + 640 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 569, + 482, + 640 + ], + "spans": [ + { + "bbox": [ + 130, + 569, + 482, + 640 + ], + "type": "text", + "content": "Realistic object interactions play a pivotal role in creating immersive virtual experiences. Recent advances in 3D vision have enabled the capture and creation of high-quality static 3D assets [34,50], and some methods even extend to 4D assets [46,47,58], generating unconditioned dynamics. However, these methods fail to handle action-conditioned dynamics in response to new physical interactions, such as synthesizing the motion of a rose reacting to a breeze or a touch." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 130, + 641, + 482, + 666 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 641, + 482, + 666 + ], + "spans": [ + { + "bbox": [ + 130, + 641, + 482, + 666 + ], + "type": "text", + "content": "The key challenge in synthesizing action-conditioned dynamics lies in understanding the physical material properties of objects. Yet, estimating these" + } + ] + } + ], + "index": 10 + } + ], + "discarded_blocks": [], + "page_size": [ + 612, + 792 + ], + "page_idx": 0 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 137, + 114, + 242, + 220 + ], + "blocks": [ + { + "bbox": [ + 137, + 114, + 242, + 220 + ], + "lines": [ + { + "bbox": [ + 137, + 114, + 242, + 220 + ], + "spans": [ + { + "bbox": [ + 137, + 114, + 242, + 220 + ], + "type": "image", + "image_path": "b6cf218f3de04570fd2e6afb26ad153d78b79bb65c175233a8d8b1a554215f09.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 165, + 221, + 211, + 230 + ], + "lines": [ + { + "bbox": [ + 165, + 221, + 211, + 230 + ], + "spans": [ + { + "bbox": [ + 165, + 221, + 211, + 230 + ], + "type": "text", + "content": "PhysDreamer" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_caption" + } + ], + "index": 2 + }, + { + "type": "image", + "bbox": [ + 248, + 115, + 282, + 218 + ], + "blocks": [ + { + "bbox": [ + 248, + 115, + 282, + 218 + ], + "lines": [ + { + "bbox": [ + 248, + 115, + 282, + 218 + ], + "spans": [ + { + "bbox": [ + 248, + 115, + 282, + 218 + ], + "type": "image", + "image_path": "979f1f2ac25d68411281b0a66c3d37e7f3ba55756d7595f5b67ce4f73d30b6ab.jpg" + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 130, + 239, + 482, + 295 + ], + "lines": [ + { + "bbox": [ + 130, + 239, + 482, + 295 + ], + "spans": [ + { + "bbox": [ + 130, + 239, + 482, + 295 + ], + "type": "text", + "content": "Fig. 1: (Left) Leveraging and distilling dynamics priors from a pre-trained video generation model, we estimate a physical material field for the static 3D object. (Right) The physical material field allows synthesizing interactive 3D dynamics under arbitrary forces. We show rendered sequences from two viewpoints. Red arrows indicate force directions. Please see videos on our project website for better visualization." + } + ] + } + ], + "index": 7, + "angle": 0, + "type": "image_caption" + } + ], + "index": 4 + }, + { + "type": "image", + "bbox": [ + 286, + 115, + 479, + 218 + ], + "blocks": [ + { + "bbox": [ + 286, + 115, + 479, + 218 + ], + "lines": [ + { + "bbox": [ + 286, + 115, + 479, + 218 + ], + "spans": [ + { + "bbox": [ + 286, + 115, + 479, + 218 + ], + "type": "image", + "image_path": "0eab3ab33285313cd3d460ca63d4649625a1da6c22c7bf04b263eb6070414436.jpg" + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 313, + 221, + 452, + 231 + ], + "lines": [ + { + "bbox": [ + 313, + 221, + 452, + 231 + ], + "spans": [ + { + "bbox": [ + 313, + 221, + 452, + 231 + ], + "type": "text", + "content": "Synthesized 3D dynamics under interaction" + } + ] + } + ], + "index": 6, + "angle": 0, + "type": "image_caption" + } + ], + "index": 5 + }, + { + "bbox": [ + 130, + 320, + 482, + 441 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 320, + 482, + 441 + ], + "spans": [ + { + "bbox": [ + 130, + 320, + 482, + 441 + ], + "type": "text", + "content": "properties is a challenging task due to the lack of ground-truth data, as measuring these properties for real objects is highly difficult. Real-life objects often exhibit complex, spatially-varying material properties, making the estimation problem even more challenging. Despite the complexity of physical materials, humans can easily imagine how objects would react to external forces, such as the gentle sway of a rose. This ability to imagine object dynamics stems from our physical prior knowledge obtained from observing and interacting with the physical world. This motivates us to distill dynamics priors from video generation models that have been trained on vast, diverse video observations of the physical world." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 130, + 445, + 483, + 566 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 445, + 483, + 566 + ], + "spans": [ + { + "bbox": [ + 130, + 445, + 483, + 566 + ], + "type": "text", + "content": "In this work, we focus on synthesizing interactive 3D dynamics. We propose PhysDreamer, a physics-based approach to transforming static 3D objects into interactive ones that can respond to novel interactions. The key idea behind PhysDreamer is to distill dynamics priors learned by video generation models to estimate the physical material properties of static 3D objects. We hypothesize that video generation models, trained on large amounts of video data, implicitly capture the relationship between object appearance and dynamics. By leveraging this learned prior knowledge, PhysDreamer can infer the physical material properties that drive the dynamic behavior of objects, even in the absence of ground-truth material data (Fig. 1)." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 130, + 570, + 482, + 666 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 570, + 482, + 666 + ], + "spans": [ + { + "bbox": [ + 130, + 570, + 482, + 666 + ], + "type": "text", + "content": "PhysDreamer represents 3D objects using 3D Gaussians [34], models the physical material field with a neural field [71], and simulates 3D dynamics using the differentiable Material Point Method (MPM) [27]. The differentiable simulation and rendering allow for direct optimization of the physical material field and initial velocity field by matching pixel space observations. We focus on elastic dynamics and showcase PhysDreamer through diverse real examples, such as flowers, plants, a beanie hat, and a telephone cord. We evaluate the realism of the synthesized interactive motion through a user study, comparing PhysDreamer" + } + ] + } + ], + "index": 10 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "spans": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "type": "text", + "content": "2" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 166, + 91, + 217, + 102 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 166, + 91, + 217, + 102 + ], + "spans": [ + { + "bbox": [ + 166, + 91, + 217, + 102 + ], + "type": "text", + "content": "Zhang et al." + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 1 + }, + { + "para_blocks": [ + { + "bbox": [ + 130, + 116, + 479, + 140 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 116, + 479, + 140 + ], + "spans": [ + { + "bbox": [ + 130, + 116, + 479, + 140 + ], + "type": "text", + "content": "to state-of-the-art methods. The results demonstrate that our approach significantly outperforms existing techniques on motion realism and visual quality." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 130, + 140, + 482, + 270 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 140, + 482, + 270 + ], + "spans": [ + { + "bbox": [ + 130, + 140, + 482, + 270 + ], + "type": "text", + "content": "In summary, PhysDreamer addresses the challenge of synthesizing interactive 3D dynamics by leveraging the object dynamics priors learned by video generation models. By distilling these priors to estimate the physical material properties of static 3D objects, our approach enables the creation of immersive virtual experiences where objects can respond realistically to novel interactions. The main contributions of our work include enabling static 3D objects to dynamically respond to interactive stimuli in a physically plausible manner and taking a step towards more engaging and realistic virtual experiences. We believe that PhysDreamer has the potential to greatly enhance the realism and interactivity of virtual environments, paving the way for more engaging and lifelike simulations." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 131, + 289, + 234, + 301 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 131, + 289, + 234, + 301 + ], + "spans": [ + { + "bbox": [ + 131, + 289, + 234, + 301 + ], + "type": "text", + "content": "2 Related work" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 131, + 312, + 300, + 324 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 131, + 312, + 300, + 324 + ], + "spans": [ + { + "bbox": [ + 131, + 312, + 300, + 324 + ], + "type": "text", + "content": "2.1 Dynamic 3D reconstruction" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 130, + 331, + 482, + 499 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 331, + 482, + 499 + ], + "spans": [ + { + "bbox": [ + 130, + 331, + 482, + 499 + ], + "type": "text", + "content": "Dynamic 3D reconstruction methods aim to reconstruct a representation of a dynamic scene from inputs such as depth scans [13,41], RGBD videos [51], or monocular or multi-view videos [1,7,39,45,47,52,53,56,66,69,74,75]. This task is especially challenging in the monocular setting with slow-moving cameras and fast-moving scenes [20]. Novel scene representations are a major driver of recent progress. One prominent approach is to augment a canonical Neural Radiance Fields (NeRF) with a deformation field [56]. This approach can be further improved by incorporating flow supervision [23,66] or as-rigid-as-possible or volume preserving regularization terms [52,53]. Time-modulated NeRFs [8,19,20,43] offer a simpler alternative representation. Due to its Lagrangian nature, 3D Gaussian Splatting [34] is readily adaptable to the task of efficient dynamic scene reconstruction [17,30,39,47,72,74]. Data-driven prior, such as from monocular depth models [38,77] and image diffusion models [67], can also be used to reduce the inherent ambiguity in dynamic reconstruction from monocular videos." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 131, + 515, + 281, + 528 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 131, + 515, + 281, + 528 + ], + "spans": [ + { + "bbox": [ + 131, + 515, + 281, + 528 + ], + "type": "text", + "content": "2.2 Dynamic 3D generation" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 130, + 533, + 482, + 666 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 533, + 482, + 666 + ], + "spans": [ + { + "bbox": [ + 130, + 533, + 482, + 666 + ], + "type": "text", + "content": "Our work also relates to efforts to synthesize dynamic 3D scenes. A common approach is to integrate a 3D generation pipeline with a video generation model [2,46,58,60]. For instance, Make-A-Video3D begins by creating a static NeRF as per DreamFusion [55], then extending it temporally using Score Distillation Sampling (SDS) [55] derived from a video diffusion model. The approach can be improved with more efficient representations, stronger diffusion priors, and stable training techniques [2, 46]. However, applying SDS with video diffusion models demands significant computational and memory costs. Compact4D [73] and DreamGaussian4D [58] used a more efficient approach, synthesizing 3D dynamics by aligning a reference video from video generation models while employing SDS from image diffusion models to reduce novel view artifacts. These" + } + ] + } + ], + "index": 8 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 391, + 91, + 447, + 102 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 391, + 91, + 447, + 102 + ], + "spans": [ + { + "bbox": [ + 391, + 91, + 447, + 102 + ], + "type": "text", + "content": "PhysDreamer" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 474, + 92, + 481, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 474, + 92, + 481, + 100 + ], + "spans": [ + { + "bbox": [ + 474, + 92, + 481, + 100 + ], + "type": "text", + "content": "3" + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 2 + }, + { + "para_blocks": [ + { + "bbox": [ + 130, + 116, + 479, + 140 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 116, + 479, + 140 + ], + "spans": [ + { + "bbox": [ + 130, + 116, + 479, + 140 + ], + "type": "text", + "content": "methods are currently limited to producing fixed-length 3D videos. We focus on synthesizing interactive 3D motions under any new physical interactions." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 131, + 158, + 310, + 170 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 131, + 158, + 310, + 170 + ], + "spans": [ + { + "bbox": [ + 131, + 158, + 310, + 170 + ], + "type": "text", + "content": "2.3 Interactive motion generation" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 130, + 177, + 480, + 285 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 177, + 480, + 285 + ], + "spans": [ + { + "bbox": [ + 130, + 177, + 480, + 285 + ], + "type": "text", + "content": "Interactive motion generation animates still images or 3D contents according to user inputs like text [11, 76], motion fields [21], motion layers [12, 14], or direct manipulation such as dragging and pulling [15, 44]. Early work from Davis et al. [15, 16] demonstrated animating an image using an image-space modal basis extracted from a video of an object undergoing subtle vibrational motions. Building upon this image-space representation [15], Generative Image Dynamics [44] used a diffusion model trained on a dataset with paired image and its modal basis to model scene motion distributions, enabling realistic interaction with still input images. We focus on interacting with 3D objects rather than images." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 130, + 286, + 481, + 392 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 286, + 481, + 392 + ], + "spans": [ + { + "bbox": [ + 130, + 286, + 481, + 392 + ], + "type": "text", + "content": "For 3D assets, physics-based approaches enable synthesizing motions under any physical interactions. Virtual Elastic Objects [10] jointly reconstructs the geometry, appearances, and physical parameters of elastic objects in a multiview capture setup with compressed air system. PAC-NeRF [42], DANO [40], and PhysGaussian [18] integrate physics-based simulations with NeRF and 3D Gaussians to generate physically plausible motions. We use the same physics-based approach to generate realistic interactions, but a novel ingredient of our work is to distill the material parameters of the object from pre-trained video generation models." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 131, + 411, + 285, + 422 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 131, + 411, + 285, + 422 + ], + "spans": [ + { + "bbox": [ + 131, + 411, + 285, + 422 + ], + "type": "text", + "content": "2.4 Video generation models" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 130, + 430, + 480, + 514 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 430, + 480, + 514 + ], + "spans": [ + { + "bbox": [ + 130, + 430, + 480, + 514 + ], + "type": "text", + "content": "Recent progress in video generation is driven by the development of larger autoregressive [26,37,65,68] and diffusion models [3-6,22,24,25,59]. These models, trained on increasingly large datasets, continue to advance the quality and realism of generated video content. The state-of-the-art approach [6] can generate minute-long videos with realistic motions and viewpoint consistency. However, current video generation models cannot support physics-based interactions with objects through external forces." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 131, + 533, + 278, + 545 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 131, + 533, + 278, + 545 + ], + "spans": [ + { + "bbox": [ + 131, + 533, + 278, + 545 + ], + "type": "text", + "content": "3 Problem formulation" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 130, + 557, + 481, + 665 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 557, + 481, + 665 + ], + "spans": [ + { + "bbox": [ + 130, + 557, + 481, + 665 + ], + "type": "text", + "content": "Given a static object represented by 3D Gaussians " + }, + { + "bbox": [ + 130, + 557, + 481, + 665 + ], + "type": "inline_equation", + "content": "\\{\\mathcal{G}_p\\}_{p=1}^P" + }, + { + "bbox": [ + 130, + 557, + 481, + 665 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 130, + 557, + 481, + 665 + ], + "type": "inline_equation", + "content": "\\mathcal{G}_p = \\{\\boldsymbol{x}_p, \\alpha_p, \\boldsymbol{\\Sigma}_p, \\boldsymbol{c}_p\\}" + }, + { + "bbox": [ + 130, + 557, + 481, + 665 + ], + "type": "text", + "content": " (where " + }, + { + "bbox": [ + 130, + 557, + 481, + 665 + ], + "type": "inline_equation", + "content": "\\boldsymbol{x}_p" + }, + { + "bbox": [ + 130, + 557, + 481, + 665 + ], + "type": "text", + "content": " denotes the position, " + }, + { + "bbox": [ + 130, + 557, + 481, + 665 + ], + "type": "inline_equation", + "content": "\\alpha_p" + }, + { + "bbox": [ + 130, + 557, + 481, + 665 + ], + "type": "text", + "content": " denotes the opacity, " + }, + { + "bbox": [ + 130, + 557, + 481, + 665 + ], + "type": "inline_equation", + "content": "\\boldsymbol{\\Sigma}_p" + }, + { + "bbox": [ + 130, + 557, + 481, + 665 + ], + "type": "text", + "content": " denotes the covariance matrix, and " + }, + { + "bbox": [ + 130, + 557, + 481, + 665 + ], + "type": "inline_equation", + "content": "\\boldsymbol{c}_p" + }, + { + "bbox": [ + 130, + 557, + 481, + 665 + ], + "type": "text", + "content": " denotes the color of the particle), our goal is to estimate physical material property fields for the object to enable realistic interactive motion synthesis. These properties include mass " + }, + { + "bbox": [ + 130, + 557, + 481, + 665 + ], + "type": "inline_equation", + "content": "m" + }, + { + "bbox": [ + 130, + 557, + 481, + 665 + ], + "type": "text", + "content": ", Young's modulus " + }, + { + "bbox": [ + 130, + 557, + 481, + 665 + ], + "type": "inline_equation", + "content": "E" + }, + { + "bbox": [ + 130, + 557, + 481, + 665 + ], + "type": "text", + "content": ", and Poisson's ratio " + }, + { + "bbox": [ + 130, + 557, + 481, + 665 + ], + "type": "inline_equation", + "content": "\\nu" + }, + { + "bbox": [ + 130, + 557, + 481, + 665 + ], + "type": "text", + "content": ". Among these physical properties, Young's modulus " + }, + { + "bbox": [ + 130, + 557, + 481, + 665 + ], + "type": "inline_equation", + "content": "E" + }, + { + "bbox": [ + 130, + 557, + 481, + 665 + ], + "type": "text", + "content": " plays a particularly important role in determining the object's motion in response to applied forces. Intuitively, Young's modulus (Eq. 2) measures the material stiffness. A higher Young's modulus results in less deformation and more rigid and" + } + ] + } + ], + "index": 9 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "spans": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "type": "text", + "content": "4" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 166, + 91, + 217, + 102 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 166, + 91, + 217, + 102 + ], + "spans": [ + { + "bbox": [ + 166, + 91, + 217, + 102 + ], + "type": "text", + "content": "Zhang et al." + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 3 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 137, + 118, + 225, + 207 + ], + "blocks": [ + { + "bbox": [ + 137, + 118, + 225, + 207 + ], + "lines": [ + { + "bbox": [ + 137, + 118, + 225, + 207 + ], + "spans": [ + { + "bbox": [ + 137, + 118, + 225, + 207 + ], + "type": "image", + "image_path": "754050422b7712f0a1809b26eadd9beed8276563972be3ba00dd6bb4484f8f55.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 155, + 209, + 200, + 216 + ], + "lines": [ + { + "bbox": [ + 155, + 209, + 200, + 216 + ], + "spans": [ + { + "bbox": [ + 155, + 209, + 200, + 216 + ], + "type": "text", + "content": "3D Gaussians" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_caption" + } + ], + "index": 2 + }, + { + "type": "image", + "bbox": [ + 243, + 119, + 316, + 207 + ], + "blocks": [ + { + "bbox": [ + 243, + 119, + 316, + 207 + ], + "lines": [ + { + "bbox": [ + 243, + 119, + 316, + 207 + ], + "spans": [ + { + "bbox": [ + 243, + 119, + 316, + 207 + ], + "type": "image", + "image_path": "486db178e00153a3fd7a02d3245e4f4fd849c72197c224acfa0dd4cb79cec133.jpg" + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 253, + 209, + 301, + 216 + ], + "lines": [ + { + "bbox": [ + 253, + 209, + 301, + 216 + ], + "spans": [ + { + "bbox": [ + 253, + 209, + 301, + 216 + ], + "type": "text", + "content": "10.30 seconds" + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 130, + 224, + 482, + 291 + ], + "lines": [ + { + "bbox": [ + 130, + 224, + 482, + 291 + ], + "spans": [ + { + "bbox": [ + 130, + 224, + 482, + 291 + ], + "type": "text", + "content": "Fig. 2: Effect of Young's modulus. We depict the motion of a simulated flower under the same external force but with three different Young's moduli, a measure of material stiffness. Flowers with the highest Young's modulus " + }, + { + "bbox": [ + 130, + 224, + 482, + 291 + ], + "type": "inline_equation", + "content": "(100\\times)" + }, + { + "bbox": [ + 130, + 224, + 482, + 291 + ], + "type": "text", + "content": " exhibit smaller oscillations and higher frequencies, while the flower with the lowest Young's modulus " + }, + { + "bbox": [ + 130, + 224, + 482, + 291 + ], + "type": "inline_equation", + "content": "(1\\times)" + }, + { + "bbox": [ + 130, + 224, + 482, + 291 + ], + "type": "text", + "content": " sways the most and oscillates at the lowest frequency. Time annotations below each image indicate the duration of one complete motion path shown in the figure." + } + ] + } + ], + "index": 10, + "angle": 0, + "type": "image_caption" + } + ], + "index": 4 + }, + { + "type": "image", + "bbox": [ + 322, + 119, + 396, + 207 + ], + "blocks": [ + { + "bbox": [ + 322, + 119, + 396, + 207 + ], + "lines": [ + { + "bbox": [ + 322, + 119, + 396, + 207 + ], + "spans": [ + { + "bbox": [ + 322, + 119, + 396, + 207 + ], + "type": "image", + "image_path": "bc8f8798a4d32ddbd5c4a007f43684398f8c57f00a2e5a1ec14997d6f95e0060.jpg" + } + ] + } + ], + "index": 6, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 337, + 209, + 381, + 216 + ], + "lines": [ + { + "bbox": [ + 337, + 209, + 381, + 216 + ], + "spans": [ + { + "bbox": [ + 337, + 209, + 381, + 216 + ], + "type": "text", + "content": "2.30 seconds" + } + ] + } + ], + "index": 7, + "angle": 0, + "type": "image_caption" + } + ], + "index": 6 + }, + { + "type": "image", + "bbox": [ + 402, + 119, + 477, + 207 + ], + "blocks": [ + { + "bbox": [ + 402, + 119, + 477, + 207 + ], + "lines": [ + { + "bbox": [ + 402, + 119, + 477, + 207 + ], + "spans": [ + { + "bbox": [ + 402, + 119, + 477, + 207 + ], + "type": "image", + "image_path": "40aa849ef07be213f45b6ae9c0d6f96148607c9f8299a945c1be340b2750566c.jpg" + } + ] + } + ], + "index": 8, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 417, + 209, + 462, + 216 + ], + "lines": [ + { + "bbox": [ + 417, + 209, + 462, + 216 + ], + "spans": [ + { + "bbox": [ + 417, + 209, + 462, + 216 + ], + "type": "text", + "content": "0.87 seconds" + } + ] + } + ], + "index": 9, + "angle": 0, + "type": "image_caption" + } + ], + "index": 8 + }, + { + "bbox": [ + 130, + 312, + 480, + 350 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 312, + 480, + 350 + ], + "spans": [ + { + "bbox": [ + 130, + 312, + 480, + 350 + ], + "type": "text", + "content": "higher-frequency motion, while a lower value leads to more flexible and elastic behavior. Fig. 2 illustrates the simulated motion of a flower under the same applied forces but with different Young's modulus." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 130, + 352, + 482, + 460 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 352, + 482, + 460 + ], + "spans": [ + { + "bbox": [ + 130, + 352, + 482, + 460 + ], + "type": "text", + "content": "Therefore, our problem formulation focuses on estimating the spatially varying Young's modulus field " + }, + { + "bbox": [ + 130, + 352, + 482, + 460 + ], + "type": "inline_equation", + "content": "E(\\pmb{x})" + }, + { + "bbox": [ + 130, + 352, + 482, + 460 + ], + "type": "text", + "content": " for the 3D object. To allow particle simulation, we query a particle's Young's modulus by " + }, + { + "bbox": [ + 130, + 352, + 482, + 460 + ], + "type": "inline_equation", + "content": "E_{p} = E(\\pmb{x}_{p})" + }, + { + "bbox": [ + 130, + 352, + 482, + 460 + ], + "type": "text", + "content": ". As for other physical properties, the mass for a particle " + }, + { + "bbox": [ + 130, + 352, + 482, + 460 + ], + "type": "inline_equation", + "content": "m_{p}" + }, + { + "bbox": [ + 130, + 352, + 482, + 460 + ], + "type": "text", + "content": " can be pre-computed as the product of a constant density " + }, + { + "bbox": [ + 130, + 352, + 482, + 460 + ], + "type": "inline_equation", + "content": "(\\rho)" + }, + { + "bbox": [ + 130, + 352, + 482, + 460 + ], + "type": "text", + "content": " and particle volume " + }, + { + "bbox": [ + 130, + 352, + 482, + 460 + ], + "type": "inline_equation", + "content": "V_{p}" + }, + { + "bbox": [ + 130, + 352, + 482, + 460 + ], + "type": "text", + "content": ". The particle volume can be estimated [70] by dividing a background cell's volume by the number of particles that cell contains. As for the Poisson's ratio " + }, + { + "bbox": [ + 130, + 352, + 482, + 460 + ], + "type": "inline_equation", + "content": "\\nu_{p}" + }, + { + "bbox": [ + 130, + 352, + 482, + 460 + ], + "type": "text", + "content": ", we found that it has negligible impact on object motion in our preliminary experiments (see supplementary materials for details), and so we assume a homogeneous constant Poisson's ratio." + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 132, + 493, + 235, + 507 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 493, + 235, + 507 + ], + "spans": [ + { + "bbox": [ + 132, + 493, + 235, + 507 + ], + "type": "text", + "content": "4 PhysDreamer" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 130, + 533, + 482, + 666 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 533, + 482, + 666 + ], + "spans": [ + { + "bbox": [ + 130, + 533, + 482, + 666 + ], + "type": "text", + "content": "PhysDreamer estimates a material field for a static 3D object. Our key idea is to generate a plausible video of the object in motion, and then optimize the material field " + }, + { + "bbox": [ + 130, + 533, + 482, + 666 + ], + "type": "inline_equation", + "content": "E(\\pmb{x})" + }, + { + "bbox": [ + 130, + 533, + 482, + 666 + ], + "type": "text", + "content": " to match this synthesized motion. We begin by rendering a static image " + }, + { + "bbox": [ + 130, + 533, + 482, + 666 + ], + "type": "inline_equation", + "content": "(I_0)" + }, + { + "bbox": [ + 130, + 533, + 482, + 666 + ], + "type": "text", + "content": " for the 3D scene " + }, + { + "bbox": [ + 130, + 533, + 482, + 666 + ], + "type": "inline_equation", + "content": "\\{\\mathcal{G}_p\\}" + }, + { + "bbox": [ + 130, + 533, + 482, + 666 + ], + "type": "text", + "content": " from a certain viewpoint. We then leverage an image-to-video model to generate a short video clip " + }, + { + "bbox": [ + 130, + 533, + 482, + 666 + ], + "type": "inline_equation", + "content": "\\{I_0, I_1, \\dots, I_T\\}" + }, + { + "bbox": [ + 130, + 533, + 482, + 666 + ], + "type": "text", + "content": " depicting the object's realistic motion. This generated video serves as our reference video. We then optimize the material field " + }, + { + "bbox": [ + 130, + 533, + 482, + 666 + ], + "type": "inline_equation", + "content": "E(\\pmb{x})" + }, + { + "bbox": [ + 130, + 533, + 482, + 666 + ], + "type": "text", + "content": " and an initial velocity field " + }, + { + "bbox": [ + 130, + 533, + 482, + 666 + ], + "type": "inline_equation", + "content": "\\pmb{v}_0(\\pmb{x})" + }, + { + "bbox": [ + 130, + 533, + 482, + 666 + ], + "type": "text", + "content": " (both modeled by implicit neural fields [71]) through differentiable simulation and differentiable rendering, such that a rendered video of the simulation matches (from the same viewpoint as " + }, + { + "bbox": [ + 130, + 533, + 482, + 666 + ], + "type": "inline_equation", + "content": "I_0" + }, + { + "bbox": [ + 130, + 533, + 482, + 666 + ], + "type": "text", + "content": ") the reference video. Fig. 3 shows an overview of PhysDreamer." + } + ] + } + ], + "index": 14 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 391, + 91, + 447, + 101 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 391, + 91, + 447, + 101 + ], + "spans": [ + { + "bbox": [ + 391, + 91, + 447, + 101 + ], + "type": "text", + "content": "PhysDreamer" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 474, + 92, + 480, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 474, + 92, + 480, + 100 + ], + "spans": [ + { + "bbox": [ + 474, + 92, + 480, + 100 + ], + "type": "text", + "content": "5" + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 4 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 136, + 114, + 478, + 281 + ], + "blocks": [ + { + "bbox": [ + 136, + 114, + 478, + 281 + ], + "lines": [ + { + "bbox": [ + 136, + 114, + 478, + 281 + ], + "spans": [ + { + "bbox": [ + 136, + 114, + 478, + 281 + ], + "type": "image", + "image_path": "7d8c8cc15f188422fc1779b27228e9f37526320ae31318ca7f35803bebc1b59d.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 130, + 289, + 482, + 367 + ], + "lines": [ + { + "bbox": [ + 130, + 289, + 482, + 367 + ], + "spans": [ + { + "bbox": [ + 130, + 289, + 482, + 367 + ], + "type": "text", + "content": "Fig. 3: Overview of PhysDreamer. Given an object represented as 3D Gaussians, we first render it (with background) from a viewpoint. Next, we use an image-to-video generation model to produce a reference video of that object in motion. Using differentiable Material Point Methods (MPM) and differentiable rendering, we optimize both a spatially-varying material field and an initial velocity field (not shown in the figure above). This optimization aims to minimize the discrepancy between the rendered video and the reference video. The dashed arrows represent gradient flow." + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_caption" + } + ], + "index": 2 + }, + { + "bbox": [ + 132, + 384, + 228, + 395 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 384, + 228, + 395 + ], + "spans": [ + { + "bbox": [ + 132, + 384, + 228, + 395 + ], + "type": "text", + "content": "4.1 Preliminaries" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 130, + 407, + 482, + 529 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 407, + 482, + 529 + ], + "spans": [ + { + "bbox": [ + 130, + 407, + 482, + 529 + ], + "type": "text", + "content": "3D Gaussians [34] adopts a set of anisotropic 3D Gaussian kernels to represent the radiance field of a 3D scene. Although introduced primarily as an efficient method for 3D novel view synthesis, the Lagrangian nature of 3D Gaussians also enables the direct adaptation of particle-based physics simulators. Following PhysGaussian [70], we use the Material Point Method (MPM) to simulate object dynamics directly on these Gaussian particles. Since 3D Gaussians mainly lie on object surfaces, an optional internal filling process can be applied for improved simulation realism [70]. Below, we provide a brief introduction on the underlying physical model and how to integrate MPM into 3D Gaussians. For a more comprehensive introduction of MPM, we refer interested readers to [27,32,33,70]." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 130, + 533, + 482, + 666 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 533, + 482, + 666 + ], + "spans": [ + { + "bbox": [ + 130, + 533, + 482, + 666 + ], + "type": "text", + "content": "Continuum mechanics and elastic materials. Continuum mechanics models material deformation using a map " + }, + { + "bbox": [ + 130, + 533, + 482, + 666 + ], + "type": "inline_equation", + "content": "\\phi" + }, + { + "bbox": [ + 130, + 533, + 482, + 666 + ], + "type": "text", + "content": " that transforms points from the undeformed material space " + }, + { + "bbox": [ + 130, + 533, + 482, + 666 + ], + "type": "inline_equation", + "content": "\\mathbf{X}" + }, + { + "bbox": [ + 130, + 533, + 482, + 666 + ], + "type": "text", + "content": " to the deformed world space " + }, + { + "bbox": [ + 130, + 533, + 482, + 666 + ], + "type": "inline_equation", + "content": "\\mathbf{x} = \\phi(\\mathbf{X}, t)" + }, + { + "bbox": [ + 130, + 533, + 482, + 666 + ], + "type": "text", + "content": ". The Jacobian of the map, " + }, + { + "bbox": [ + 130, + 533, + 482, + 666 + ], + "type": "inline_equation", + "content": "\\pmb{F} = \\nabla_{\\mathbf{X}}\\phi(\\mathbf{X}, t)" + }, + { + "bbox": [ + 130, + 533, + 482, + 666 + ], + "type": "text", + "content": ", known as the deformation gradient, measures local rotation and strain. This tensor is crucial in formulating stress-strain relationship. For example, the Cauchy stress in a hyper-elastic material is computed by: " + }, + { + "bbox": [ + 130, + 533, + 482, + 666 + ], + "type": "inline_equation", + "content": "\\pmb{\\sigma} = \\frac{1}{\\operatorname*{det}(\\pmb{F})}\\frac{\\partial\\psi}{\\partial\\pmb{F}}\\pmb{F}^T" + }, + { + "bbox": [ + 130, + 533, + 482, + 666 + ], + "type": "text", + "content": ". Here, " + }, + { + "bbox": [ + 130, + 533, + 482, + 666 + ], + "type": "inline_equation", + "content": "\\psi(\\mathbf{F})" + }, + { + "bbox": [ + 130, + 533, + 482, + 666 + ], + "type": "text", + "content": " represents the strain energy density function, quantifying the extent of non-rigid deformations. This function is typically designed by experts, to follow principles like material symmetry and rotational invariance while aligning with empirical data. In this work, we use fixed corotated hyperelastic model, whose energy density function can be expressed as:" + } + ] + } + ], + "index": 6 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "spans": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "type": "text", + "content": "6" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 166, + 91, + 217, + 102 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 166, + 91, + 217, + 102 + ], + "spans": [ + { + "bbox": [ + 166, + 91, + 217, + 102 + ], + "type": "text", + "content": "Zhang et al." + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 5 + }, + { + "para_blocks": [ + { + "bbox": [ + 210, + 127, + 480, + 159 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 210, + 127, + 480, + 159 + ], + "spans": [ + { + "bbox": [ + 210, + 127, + 480, + 159 + ], + "type": "interline_equation", + "content": "\\psi (\\mathbf {F}) = \\mu \\left(\\sum_ {i = 1} ^ {d} \\left(\\sigma_ {i} - 1\\right) ^ {2}\\right) + \\frac {\\lambda}{2} (\\det (\\mathbf {F}) - 1) ^ {2}, \\tag {1}", + "image_path": "a7c46fc82449ff5cf0a664c1ebf83f2d356dc286cef953aca3eb050a23682be2.jpg" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 132, + 163, + 479, + 187 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 163, + 479, + 187 + ], + "spans": [ + { + "bbox": [ + 132, + 163, + 479, + 187 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 132, + 163, + 479, + 187 + ], + "type": "inline_equation", + "content": "\\sigma_{i}" + }, + { + "bbox": [ + 132, + 163, + 479, + 187 + ], + "type": "text", + "content": " denotes a singular value of the deformation gradient. " + }, + { + "bbox": [ + 132, + 163, + 479, + 187 + ], + "type": "inline_equation", + "content": "\\mu" + }, + { + "bbox": [ + 132, + 163, + 479, + 187 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 132, + 163, + 479, + 187 + ], + "type": "inline_equation", + "content": "\\lambda" + }, + { + "bbox": [ + 132, + 163, + 479, + 187 + ], + "type": "text", + "content": " are related to Young's modulus " + }, + { + "bbox": [ + 132, + 163, + 479, + 187 + ], + "type": "inline_equation", + "content": "E" + }, + { + "bbox": [ + 132, + 163, + 479, + 187 + ], + "type": "text", + "content": " and Poisson's ratio " + }, + { + "bbox": [ + 132, + 163, + 479, + 187 + ], + "type": "inline_equation", + "content": "\\nu" + }, + { + "bbox": [ + 132, + 163, + 479, + 187 + ], + "type": "text", + "content": " via:" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 225, + 198, + 479, + 222 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 225, + 198, + 479, + 222 + ], + "spans": [ + { + "bbox": [ + 225, + 198, + 479, + 222 + ], + "type": "interline_equation", + "content": "\\mu = \\frac {E}{2 (1 + \\nu)}, \\quad \\lambda = \\frac {E \\nu}{(1 + \\nu) (1 - 2 \\nu)}. \\tag {2}", + "image_path": "0b4ce1819fa1d87285cb88c45f3240ad802e86dcb07d864a4552c3eb463fb3d1.jpg" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 132, + 231, + 460, + 243 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 231, + 460, + 243 + ], + "spans": [ + { + "bbox": [ + 132, + 231, + 460, + 243 + ], + "type": "text", + "content": "The dynamics of an elastic object are governed by the following equations:" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 223, + 252, + 479, + 274 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 223, + 252, + 479, + 274 + ], + "spans": [ + { + "bbox": [ + 223, + 252, + 479, + 274 + ], + "type": "interline_equation", + "content": "\\rho \\frac {D \\boldsymbol {v}}{D t} = \\nabla \\cdot \\boldsymbol {\\sigma} + \\mathbf {f}, \\quad \\frac {D \\rho}{D t} + \\rho \\nabla \\cdot \\boldsymbol {v} = 0, \\tag {3}", + "image_path": "5695072448f53be60700734b643b559de073cce6017b59a3044db262998f5be6.jpg" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 132, + 283, + 479, + 306 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 283, + 479, + 306 + ], + "spans": [ + { + "bbox": [ + 132, + 283, + 479, + 306 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 132, + 283, + 479, + 306 + ], + "type": "inline_equation", + "content": "\\rho" + }, + { + "bbox": [ + 132, + 283, + 479, + 306 + ], + "type": "text", + "content": " denotes density, " + }, + { + "bbox": [ + 132, + 283, + 479, + 306 + ], + "type": "inline_equation", + "content": "\\pmb{v}(\\pmb{x},t)" + }, + { + "bbox": [ + 132, + 283, + 479, + 306 + ], + "type": "text", + "content": " denotes the velocity field in world space, and " + }, + { + "bbox": [ + 132, + 283, + 479, + 306 + ], + "type": "inline_equation", + "content": "\\mathbf{f}" + }, + { + "bbox": [ + 132, + 283, + 479, + 306 + ], + "type": "text", + "content": " denotes an external force." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 132, + 312, + 480, + 395 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 312, + 480, + 395 + ], + "spans": [ + { + "bbox": [ + 132, + 312, + 480, + 395 + ], + "type": "text", + "content": "Material Point Method (MPM). We use the Moving Least Squares Material Point Method (MLS-MPM) [27] to solve the above governing equation. MPM is a hybrid Eulerian-Langrangian method widely adopted for simulating dynamics for a wide range of materials, such as solid, fluid, sand, and cloth [31, 35, 57, 62]. MPM offers several advantages, such as easy GPU parallelization [28], handling of topology changes, and the availability of well-documented open-source implementations [29, 48, 49, 70]." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 132, + 396, + 480, + 456 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 396, + 480, + 456 + ], + "spans": [ + { + "bbox": [ + 132, + 396, + 480, + 456 + ], + "type": "text", + "content": "Following PhysGaussian [70], we view the Gaussian particles as the spatial discretization of the object to be simulated, and directly run MPM on these Gaussian particles. Each particle " + }, + { + "bbox": [ + 132, + 396, + 480, + 456 + ], + "type": "inline_equation", + "content": "p" + }, + { + "bbox": [ + 132, + 396, + 480, + 456 + ], + "type": "text", + "content": " represents a small volume of the object, and it carries a set of properties including volume " + }, + { + "bbox": [ + 132, + 396, + 480, + 456 + ], + "type": "inline_equation", + "content": "V_{p}" + }, + { + "bbox": [ + 132, + 396, + 480, + 456 + ], + "type": "text", + "content": ", mass " + }, + { + "bbox": [ + 132, + 396, + 480, + 456 + ], + "type": "inline_equation", + "content": "m_{p}" + }, + { + "bbox": [ + 132, + 396, + 480, + 456 + ], + "type": "text", + "content": ", position " + }, + { + "bbox": [ + 132, + 396, + 480, + 456 + ], + "type": "inline_equation", + "content": "\\boldsymbol{x}_{p}^{t}" + }, + { + "bbox": [ + 132, + 396, + 480, + 456 + ], + "type": "text", + "content": ", velocity " + }, + { + "bbox": [ + 132, + 396, + 480, + 456 + ], + "type": "inline_equation", + "content": "\\boldsymbol{v}_{p}^{t}" + }, + { + "bbox": [ + 132, + 396, + 480, + 456 + ], + "type": "text", + "content": ", deformation gradient " + }, + { + "bbox": [ + 132, + 396, + 480, + 456 + ], + "type": "inline_equation", + "content": "\\boldsymbol{F}_{p}^{t}" + }, + { + "bbox": [ + 132, + 396, + 480, + 456 + ], + "type": "text", + "content": ", and local velocity field gradient " + }, + { + "bbox": [ + 132, + 396, + 480, + 456 + ], + "type": "inline_equation", + "content": "\\boldsymbol{C}_{p}^{t}" + }, + { + "bbox": [ + 132, + 396, + 480, + 456 + ], + "type": "text", + "content": " at time step " + }, + { + "bbox": [ + 132, + 396, + 480, + 456 + ], + "type": "inline_equation", + "content": "t" + }, + { + "bbox": [ + 132, + 396, + 480, + 456 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 132, + 456, + 479, + 479 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 456, + 479, + 479 + ], + "spans": [ + { + "bbox": [ + 132, + 456, + 479, + 479 + ], + "type": "text", + "content": "MPM operates in a particle-to-grid (P2G) and grid-to-particle (G2P) transfer loop. In the P2G stage, we transfer the momentum from particle to grid by:" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 138, + 490, + 479, + 517 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 490, + 479, + 517 + ], + "spans": [ + { + "bbox": [ + 138, + 490, + 479, + 517 + ], + "type": "interline_equation", + "content": "m _ {i} ^ {t} \\boldsymbol {v} _ {i} ^ {t} = \\sum_ {p} N (\\boldsymbol {x} _ {i} - \\boldsymbol {x} _ {p} ^ {t}) [ m _ {p} \\boldsymbol {v} _ {p} ^ {t} + (m _ {p} \\boldsymbol {C} _ {p} ^ {t} - \\frac {4}{(\\Delta x) ^ {2}} \\Delta t V _ {p} \\frac {\\partial \\psi}{\\partial \\boldsymbol {F}} \\boldsymbol {F} _ {p} ^ {t T}) (\\boldsymbol {x} _ {i} - \\boldsymbol {x} _ {p} ^ {t}) ] + \\boldsymbol {f} _ {i} ^ {t}, \\tag {4}", + "image_path": "214794e4e476b808e52d6d3fa2954d9523c62c868a8058738825e73c4a61c200.jpg" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 132, + 527, + 480, + 575 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 527, + 480, + 575 + ], + "spans": [ + { + "bbox": [ + 132, + 527, + 480, + 575 + ], + "type": "text", + "content": "where the mass of the grid node " + }, + { + "bbox": [ + 132, + 527, + 480, + 575 + ], + "type": "inline_equation", + "content": "i" + }, + { + "bbox": [ + 132, + 527, + 480, + 575 + ], + "type": "text", + "content": " is " + }, + { + "bbox": [ + 132, + 527, + 480, + 575 + ], + "type": "inline_equation", + "content": "m_i^t = \\sum_p N(\\pmb{x}_i - \\pmb{x}_p^t)m_p" + }, + { + "bbox": [ + 132, + 527, + 480, + 575 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 132, + 527, + 480, + 575 + ], + "type": "inline_equation", + "content": "N(\\pmb{x}_i - \\pmb{x}_p^t)" + }, + { + "bbox": [ + 132, + 527, + 480, + 575 + ], + "type": "text", + "content": " is the B-spline kernel, " + }, + { + "bbox": [ + 132, + 527, + 480, + 575 + ], + "type": "inline_equation", + "content": "\\Delta x" + }, + { + "bbox": [ + 132, + 527, + 480, + 575 + ], + "type": "text", + "content": " is the spatial grid resolution, " + }, + { + "bbox": [ + 132, + 527, + 480, + 575 + ], + "type": "inline_equation", + "content": "\\Delta t" + }, + { + "bbox": [ + 132, + 527, + 480, + 575 + ], + "type": "text", + "content": " is the simulation step size, and " + }, + { + "bbox": [ + 132, + 527, + 480, + 575 + ], + "type": "inline_equation", + "content": "\\pmb{v}_i^t" + }, + { + "bbox": [ + 132, + 527, + 480, + 575 + ], + "type": "text", + "content": " is the updated velocity on the grid. We then transfer the updated velocity back to the particles and update their positions as:" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 200, + 586, + 479, + 609 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 200, + 586, + 479, + 609 + ], + "spans": [ + { + "bbox": [ + 200, + 586, + 479, + 609 + ], + "type": "interline_equation", + "content": "\\boldsymbol {v} _ {p} ^ {t + 1} = \\sum_ {i} N (\\boldsymbol {x} _ {i} - \\boldsymbol {x} _ {p} ^ {t}) \\boldsymbol {v} _ {i} ^ {t}, \\quad \\boldsymbol {x} _ {p} ^ {t + 1} = \\boldsymbol {x} _ {p} ^ {t} + \\Delta t \\boldsymbol {v} _ {p} ^ {t + 1}. \\tag {5}", + "image_path": "2870bfec5fc0b07ae7bac3d1263f01bd375dbf09bddb14212669a9e66caba29e.jpg" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 132, + 619, + 476, + 631 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 619, + 476, + 631 + ], + "spans": [ + { + "bbox": [ + 132, + 619, + 476, + 631 + ], + "type": "text", + "content": "Meanwhile, the local velocity gradient and deformation gradient is updated as:" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 144, + 640, + 479, + 667 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 144, + 640, + 479, + 667 + ], + "spans": [ + { + "bbox": [ + 144, + 640, + 479, + 667 + ], + "type": "interline_equation", + "content": "\\boldsymbol {C} _ {p} ^ {t + 1} = \\frac {4}{(\\Delta x) ^ {2}} \\sum_ {i} N (\\boldsymbol {x} _ {i} - \\boldsymbol {x} _ {p} ^ {t}) \\boldsymbol {v} _ {i} ^ {t} (\\boldsymbol {x} _ {i} - \\boldsymbol {x} _ {p} ^ {t}) ^ {T}, \\quad \\boldsymbol {F} _ {p} ^ {t + 1} = (\\boldsymbol {I} + \\Delta t \\boldsymbol {C} _ {p} ^ {t + 1}) \\boldsymbol {F} _ {p} ^ {t}. \\tag {6}", + "image_path": "2c61f41a5c8c96c9e06f2d98ce92f1eef15874f6357899415fa40cd083b606e7.jpg" + } + ] + } + ], + "index": 15 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 391, + 91, + 447, + 101 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 391, + 91, + 447, + 101 + ], + "spans": [ + { + "bbox": [ + 391, + 91, + 447, + 101 + ], + "type": "text", + "content": "PhysDreamer" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 474, + 92, + 481, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 474, + 92, + 481, + 100 + ], + "spans": [ + { + "bbox": [ + 474, + 92, + 481, + 100 + ], + "type": "text", + "content": "7" + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 6 + }, + { + "para_blocks": [ + { + "bbox": [ + 132, + 116, + 314, + 129 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 116, + 314, + 129 + ], + "spans": [ + { + "bbox": [ + 132, + 116, + 314, + 129 + ], + "type": "text", + "content": "4.2 Estimating physical properties" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 130, + 133, + 482, + 168 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 133, + 482, + 168 + ], + "spans": [ + { + "bbox": [ + 130, + 133, + 482, + 168 + ], + "type": "text", + "content": "Using MLS-MPM [27, 70] as our physics simulator and the Fixed Corotated hyper-elastic material model for the 3D objects, the simulation process for a single sub-step is formalized as:" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 200, + 173, + 481, + 186 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 200, + 173, + 481, + 186 + ], + "spans": [ + { + "bbox": [ + 200, + 173, + 481, + 186 + ], + "type": "interline_equation", + "content": "\\boldsymbol {x} ^ {t + 1}, \\boldsymbol {v} ^ {t + 1}, \\boldsymbol {F} ^ {t + 1}, \\boldsymbol {C} ^ {t + 1} = \\mathcal {S} \\left(\\boldsymbol {x} ^ {t}, \\boldsymbol {v} ^ {t}, \\boldsymbol {F} ^ {t}, \\boldsymbol {C} ^ {t}, \\boldsymbol {\\theta}, \\Delta t\\right), \\tag {7}", + "image_path": "04061d0794b6758631bcf4fd11b9370f35e5b3de6da06de4a04b2974d88b44d7.jpg" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 130, + 191, + 479, + 275 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 191, + 479, + 275 + ], + "spans": [ + { + "bbox": [ + 130, + 191, + 479, + 275 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 130, + 191, + 479, + 275 + ], + "type": "inline_equation", + "content": "\\pmb{x}^t = [\\pmb{x}_1^t, \\dots, \\pmb{x}_P^t]" + }, + { + "bbox": [ + 130, + 191, + 479, + 275 + ], + "type": "text", + "content": " denotes the positions of all particles at time " + }, + { + "bbox": [ + 130, + 191, + 479, + 275 + ], + "type": "inline_equation", + "content": "t" + }, + { + "bbox": [ + 130, + 191, + 479, + 275 + ], + "type": "text", + "content": ", and similarly " + }, + { + "bbox": [ + 130, + 191, + 479, + 275 + ], + "type": "inline_equation", + "content": "\\pmb{v}^t = [\\pmb{v}_1^t, \\dots, \\pmb{v}_P^t]" + }, + { + "bbox": [ + 130, + 191, + 479, + 275 + ], + "type": "text", + "content": " denotes the velocities of all particles at time " + }, + { + "bbox": [ + 130, + 191, + 479, + 275 + ], + "type": "inline_equation", + "content": "t" + }, + { + "bbox": [ + 130, + 191, + 479, + 275 + ], + "type": "text", + "content": ". " + }, + { + "bbox": [ + 130, + 191, + 479, + 275 + ], + "type": "inline_equation", + "content": "\\pmb{F}^t" + }, + { + "bbox": [ + 130, + 191, + 479, + 275 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 130, + 191, + 479, + 275 + ], + "type": "inline_equation", + "content": "\\pmb{C}^t" + }, + { + "bbox": [ + 130, + 191, + 479, + 275 + ], + "type": "text", + "content": " denote the deformation gradient and the gradient of local velocity fields for all particles, respectively. Both " + }, + { + "bbox": [ + 130, + 191, + 479, + 275 + ], + "type": "inline_equation", + "content": "\\pmb{F}^t" + }, + { + "bbox": [ + 130, + 191, + 479, + 275 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 130, + 191, + 479, + 275 + ], + "type": "inline_equation", + "content": "\\pmb{C}^t" + }, + { + "bbox": [ + 130, + 191, + 479, + 275 + ], + "type": "text", + "content": " are tracked for simulation purposes, not for rendering. " + }, + { + "bbox": [ + 130, + 191, + 479, + 275 + ], + "type": "inline_equation", + "content": "\\pmb{\\theta}" + }, + { + "bbox": [ + 130, + 191, + 479, + 275 + ], + "type": "text", + "content": " denotes the collection of the physical properties of all particles: mass " + }, + { + "bbox": [ + 130, + 191, + 479, + 275 + ], + "type": "inline_equation", + "content": "\\pmb{m} = [m_1, \\dots, m_P]" + }, + { + "bbox": [ + 130, + 191, + 479, + 275 + ], + "type": "text", + "content": ", Young's modulus " + }, + { + "bbox": [ + 130, + 191, + 479, + 275 + ], + "type": "inline_equation", + "content": "\\pmb{E} = [E_1, \\dots, E_P]" + }, + { + "bbox": [ + 130, + 191, + 479, + 275 + ], + "type": "text", + "content": ", Poisson's ratio " + }, + { + "bbox": [ + 130, + 191, + 479, + 275 + ], + "type": "inline_equation", + "content": "\\pmb{\\nu} = [\\nu_1, \\dots, \\nu_P]" + }, + { + "bbox": [ + 130, + 191, + 479, + 275 + ], + "type": "text", + "content": ", and volume " + }, + { + "bbox": [ + 130, + 191, + 479, + 275 + ], + "type": "inline_equation", + "content": "\\pmb{V} = [V_1, \\dots, V_P]" + }, + { + "bbox": [ + 130, + 191, + 479, + 275 + ], + "type": "text", + "content": ". " + }, + { + "bbox": [ + 130, + 191, + 479, + 275 + ], + "type": "inline_equation", + "content": "\\Delta t" + }, + { + "bbox": [ + 130, + 191, + 479, + 275 + ], + "type": "text", + "content": " is the simulation step size." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 130, + 275, + 479, + 323 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 275, + 479, + 323 + ], + "spans": [ + { + "bbox": [ + 130, + 275, + 479, + 323 + ], + "type": "text", + "content": "We use a sub-step size " + }, + { + "bbox": [ + 130, + 275, + 479, + 323 + ], + "type": "inline_equation", + "content": "\\Delta t \\cong 1 \\times 10^{-4}" + }, + { + "bbox": [ + 130, + 275, + 479, + 323 + ], + "type": "text", + "content": " for most of our experiments. To simulate dynamics between adjacent video frames, we iterate over hundreds of sub-steps (time interval between frames are typically tens of milliseconds). For simplicity, we abuse notation to express a simulation step with " + }, + { + "bbox": [ + 130, + 275, + 479, + 323 + ], + "type": "inline_equation", + "content": "N" + }, + { + "bbox": [ + 130, + 275, + 479, + 323 + ], + "type": "text", + "content": " sub-steps as:" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 194, + 327, + 480, + 341 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 194, + 327, + 480, + 341 + ], + "spans": [ + { + "bbox": [ + 194, + 327, + 480, + 341 + ], + "type": "interline_equation", + "content": "\\boldsymbol {x} ^ {t + 1}, \\boldsymbol {v} ^ {t + 1}, \\boldsymbol {F} ^ {t + 1}, \\boldsymbol {C} ^ {t + 1} = \\mathcal {S} (\\boldsymbol {x} ^ {t}, \\boldsymbol {v} ^ {t}, \\boldsymbol {F} ^ {t}, \\boldsymbol {C} ^ {t}, \\boldsymbol {\\theta}, \\Delta t, N), \\tag {8}", + "image_path": "90e6a6e11a50a3c2d2700e75227cadc786640ef3e8175e345a542cd7ad3b56ce.jpg" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 131, + 346, + 479, + 369 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 131, + 346, + 479, + 369 + ], + "spans": [ + { + "bbox": [ + 131, + 346, + 479, + 369 + ], + "type": "text", + "content": "where the timestamp " + }, + { + "bbox": [ + 131, + 346, + 479, + 369 + ], + "type": "inline_equation", + "content": "t + 1" + }, + { + "bbox": [ + 131, + 346, + 479, + 369 + ], + "type": "text", + "content": " is ahead of timestamp " + }, + { + "bbox": [ + 131, + 346, + 479, + 369 + ], + "type": "inline_equation", + "content": "t" + }, + { + "bbox": [ + 131, + 346, + 479, + 369 + ], + "type": "text", + "content": " by " + }, + { + "bbox": [ + 131, + 346, + 479, + 369 + ], + "type": "inline_equation", + "content": "N\\Delta t" + }, + { + "bbox": [ + 131, + 346, + 479, + 369 + ], + "type": "text", + "content": ". After simulation, we render the Gaussians at each frame:" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 246, + 375, + 480, + 388 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 246, + 375, + 480, + 388 + ], + "spans": [ + { + "bbox": [ + 246, + 375, + 480, + 388 + ], + "type": "interline_equation", + "content": "\\hat {I} ^ {t} = \\mathcal {F} _ {\\text {r e n d e r}} \\left(\\boldsymbol {x} ^ {t}, \\boldsymbol {\\alpha}, \\boldsymbol {R} ^ {t}, \\Sigma , \\boldsymbol {c}\\right), \\tag {9}", + "image_path": "0d1bb23aff927f67fe96944f52e0102994a9fc1d81cf70d6559893f2ee7509ac.jpg" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 130, + 392, + 479, + 416 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 392, + 479, + 416 + ], + "spans": [ + { + "bbox": [ + 130, + 392, + 479, + 416 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 130, + 392, + 479, + 416 + ], + "type": "inline_equation", + "content": "\\mathcal{F}_{\\mathrm{render}}" + }, + { + "bbox": [ + 130, + 392, + 479, + 416 + ], + "type": "text", + "content": " denotes the differentiable rendering function, and " + }, + { + "bbox": [ + 130, + 392, + 479, + 416 + ], + "type": "inline_equation", + "content": "\\pmb{R}^{t}" + }, + { + "bbox": [ + 130, + 392, + 479, + 416 + ], + "type": "text", + "content": " denotes the rotation matrices of all particles obtained from the simulation step." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 131, + 417, + 479, + 440 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 131, + 417, + 479, + 440 + ], + "spans": [ + { + "bbox": [ + 131, + 417, + 479, + 440 + ], + "type": "text", + "content": "Using the generated video as reference, we optimize the spatially-varying Young's modulus " + }, + { + "bbox": [ + 131, + 417, + 479, + 440 + ], + "type": "inline_equation", + "content": "\\pmb{E}" + }, + { + "bbox": [ + 131, + 417, + 479, + 440 + ], + "type": "text", + "content": " and an initial velocity " + }, + { + "bbox": [ + 131, + 417, + 479, + 440 + ], + "type": "inline_equation", + "content": "\\pmb{v}^0" + }, + { + "bbox": [ + 131, + 417, + 479, + 440 + ], + "type": "text", + "content": " by a per-frame loss function:" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 216, + 445, + 480, + 459 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 216, + 445, + 480, + 459 + ], + "spans": [ + { + "bbox": [ + 216, + 445, + 480, + 459 + ], + "type": "interline_equation", + "content": "L ^ {t} = \\lambda L _ {1} \\left(\\hat {I} ^ {t}, I ^ {t}\\right) + (1 - \\lambda) L _ {\\mathrm {D} - \\mathrm {S S I M}} \\left(\\hat {I} ^ {t}, I ^ {t}\\right), \\tag {10}", + "image_path": "69f76d2cb7c6b10944e09ddb0ce7cf44c8e109b93b10b9a445146bc4d96c398c.jpg" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 131, + 464, + 314, + 475 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 131, + 464, + 314, + 475 + ], + "spans": [ + { + "bbox": [ + 131, + 464, + 314, + 475 + ], + "type": "text", + "content": "where we set " + }, + { + "bbox": [ + 131, + 464, + 314, + 475 + ], + "type": "inline_equation", + "content": "\\lambda = 0.1" + }, + { + "bbox": [ + 131, + 464, + 314, + 475 + ], + "type": "text", + "content": " in our experiments." + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 130, + 475, + 479, + 535 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 475, + 479, + 535 + ], + "spans": [ + { + "bbox": [ + 130, + 475, + 479, + 535 + ], + "type": "text", + "content": "We parameterize the material field and velocity field by two triplanes [9], each followed by a three-layer MLP. Additionally, we apply a total variation regularization for all spatial planes of both fields to encourage spatial smoothness. Using " + }, + { + "bbox": [ + 130, + 475, + 479, + 535 + ], + "type": "inline_equation", + "content": "\\pmb{u}" + }, + { + "bbox": [ + 130, + 475, + 479, + 535 + ], + "type": "text", + "content": " to denote one of the 2D spatial planes, and " + }, + { + "bbox": [ + 130, + 475, + 479, + 535 + ], + "type": "inline_equation", + "content": "\\pmb{u}_{i,j}" + }, + { + "bbox": [ + 130, + 475, + 479, + 535 + ], + "type": "text", + "content": " as a feature vector on the 2D plane, we write the total variation regularization term as:" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 208, + 540, + 480, + 564 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 208, + 540, + 480, + 564 + ], + "spans": [ + { + "bbox": [ + 208, + 540, + 480, + 564 + ], + "type": "interline_equation", + "content": "L _ {\\mathrm {t v}} = \\sum_ {i, j} \\| \\boldsymbol {u} _ {i + 1, j} - \\boldsymbol {u} _ {i, j} \\| _ {2} ^ {2} + \\| \\boldsymbol {u} _ {i, j + 1} - \\boldsymbol {u} _ {i, j} \\| _ {2} ^ {2}. \\tag {11}", + "image_path": "4250f33422f9fd095ce0871612d2cf67f8b66d8f3e0e142c02a9d0de9d036cf6.jpg" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 130, + 570, + 482, + 666 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 570, + 482, + 666 + ], + "spans": [ + { + "bbox": [ + 130, + 570, + 482, + 666 + ], + "type": "text", + "content": "Rather than optimizing the material parameters and initial velocity jointly, we split the optimization into two stages for better stability and faster convergence. In particular, in the first stage, we randomly initialize the Young's modulus for each Gaussian particle and freeze it. We optimize the initial velocity of each particle using only the first three frames of the reference video. In the second stage, we freeze the initial velocity and optimize the spatially varying Young's modulus. During the second stage, the gradient signal only flows to the previous frame to prevent gradient explosion/vanishing." + } + ] + } + ], + "index": 16 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "spans": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "type": "text", + "content": "8" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 166, + 91, + 217, + 102 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 166, + 91, + 217, + 102 + ], + "spans": [ + { + "bbox": [ + 166, + 91, + 217, + 102 + ], + "type": "text", + "content": "Zhang et al." + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 7 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 302, + 114, + 463, + 242 + ], + "blocks": [ + { + "bbox": [ + 132, + 118, + 282, + 228 + ], + "lines": [ + { + "bbox": [ + 132, + 118, + 282, + 228 + ], + "spans": [ + { + "bbox": [ + 132, + 118, + 282, + 228 + ], + "type": "text", + "content": "Fig.4: Accelerated MPM with K-Means downsampling. We employ K-Means clustering to create a set of \"driving particles\" (in yellow) at the initial time step " + }, + { + "bbox": [ + 132, + 118, + 282, + 228 + ], + "type": "inline_equation", + "content": "(t = 0)" + }, + { + "bbox": [ + 132, + 118, + 282, + 228 + ], + "type": "text", + "content": ". We only simulate these driving particles. When rendering, we obtain each particle's position and rotation by fitting a local rigid body transformation using neighboring driving particles." + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 302, + 114, + 463, + 242 + ], + "lines": [ + { + "bbox": [ + 302, + 114, + 463, + 242 + ], + "spans": [ + { + "bbox": [ + 302, + 114, + 463, + 242 + ], + "type": "image", + "image_path": "f27687cab82aa0f9bfca75ad70f1374e2435ca40c028a8b9f0bcd26cd9200ba0.jpg" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_body" + } + ], + "index": 3 + }, + { + "bbox": [ + 132, + 255, + 371, + 268 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 255, + 371, + 268 + ], + "spans": [ + { + "bbox": [ + 132, + 255, + 371, + 268 + ], + "type": "text", + "content": "4.3 Accelerating simulation with subsampling" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 130, + 277, + 482, + 324 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 277, + 482, + 324 + ], + "spans": [ + { + "bbox": [ + 130, + 277, + 482, + 324 + ], + "type": "text", + "content": "High-fidelity rendering with 3D Gaussians typically requires millions of particles to represent a scene. Running simulations on all the particles poses a significant computational burden. To improve efficiency, we introduce a subsampling procedure for simulation, as illustrated in Fig. 4." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 132, + 325, + 482, + 482 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 325, + 482, + 482 + ], + "spans": [ + { + "bbox": [ + 132, + 325, + 482, + 482 + ], + "type": "text", + "content": "Specifically, we apply K-Means clustering to create a set of driving particles " + }, + { + "bbox": [ + 132, + 325, + 482, + 482 + ], + "type": "inline_equation", + "content": "\\{\\mathcal{Q}_q\\}_{q=1}^Q" + }, + { + "bbox": [ + 132, + 325, + 482, + 482 + ], + "type": "text", + "content": " at " + }, + { + "bbox": [ + 132, + 325, + 482, + 482 + ], + "type": "inline_equation", + "content": "t = 0" + }, + { + "bbox": [ + 132, + 325, + 482, + 482 + ], + "type": "text", + "content": ", where each driving particle is represented by " + }, + { + "bbox": [ + 132, + 325, + 482, + 482 + ], + "type": "inline_equation", + "content": "\\mathcal{Q}_q^0 = \\{\\pmb{x}_q^0, \\pmb{v}_q^0, \\pmb{F}_q^0, \\pmb{C}_q^0, E_q, m_q, \\nu_q, V_q\\}" + }, + { + "bbox": [ + 132, + 325, + 482, + 482 + ], + "type": "text", + "content": ". The initial position of a driving particle " + }, + { + "bbox": [ + 132, + 325, + 482, + 482 + ], + "type": "inline_equation", + "content": "\\pmb{x}_q^0" + }, + { + "bbox": [ + 132, + 325, + 482, + 482 + ], + "type": "text", + "content": " is computed as the mean of the position " + }, + { + "bbox": [ + 132, + 325, + 482, + 482 + ], + "type": "inline_equation", + "content": "\\pmb{x}_p" + }, + { + "bbox": [ + 132, + 325, + 482, + 482 + ], + "type": "text", + "content": " of all cluster members. The number of the driving particles is much smaller than the number of 3D Gaussian particles, " + }, + { + "bbox": [ + 132, + 325, + 482, + 482 + ], + "type": "inline_equation", + "content": "Q \\ll P" + }, + { + "bbox": [ + 132, + 325, + 482, + 482 + ], + "type": "text", + "content": ". We run simulations only on the driving particles. During rendering, we compute the position and rotation for each 3D Gaussian particle " + }, + { + "bbox": [ + 132, + 325, + 482, + 482 + ], + "type": "inline_equation", + "content": "\\mathcal{G}_p" + }, + { + "bbox": [ + 132, + 325, + 482, + 482 + ], + "type": "text", + "content": " by interpolating the driving particles. In particular, for each 3D Gaussian particle, we find its eight nearest driving particles at " + }, + { + "bbox": [ + 132, + 325, + 482, + 482 + ], + "type": "inline_equation", + "content": "t = 0" + }, + { + "bbox": [ + 132, + 325, + 482, + 482 + ], + "type": "text", + "content": ", and we fit a rigid body transformation " + }, + { + "bbox": [ + 132, + 325, + 482, + 482 + ], + "type": "inline_equation", + "content": "\\pmb{T}" + }, + { + "bbox": [ + 132, + 325, + 482, + 482 + ], + "type": "text", + "content": " between these eight driving particles at " + }, + { + "bbox": [ + 132, + 325, + 482, + 482 + ], + "type": "inline_equation", + "content": "t = 0" + }, + { + "bbox": [ + 132, + 325, + 482, + 482 + ], + "type": "text", + "content": " and at the current timestamp. This rigid body transformation " + }, + { + "bbox": [ + 132, + 325, + 482, + 482 + ], + "type": "inline_equation", + "content": "\\pmb{T}" + }, + { + "bbox": [ + 132, + 325, + 482, + 482 + ], + "type": "text", + "content": " is applied to the initial position and rotation of the particle " + }, + { + "bbox": [ + 132, + 325, + 482, + 482 + ], + "type": "inline_equation", + "content": "\\mathcal{G}_p" + }, + { + "bbox": [ + 132, + 325, + 482, + 482 + ], + "type": "text", + "content": " to obtain its current position and rotation. We summarize our algorithm with pseudo-code in supplementary materials." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 132, + 501, + 230, + 514 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 501, + 230, + 514 + ], + "spans": [ + { + "bbox": [ + 132, + 501, + 230, + 514 + ], + "type": "text", + "content": "5 Experiments" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 132, + 528, + 190, + 540 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 528, + 190, + 540 + ], + "spans": [ + { + "bbox": [ + 132, + 528, + 190, + 540 + ], + "type": "text", + "content": "5.1 Setup" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 130, + 553, + 482, + 636 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 553, + 482, + 636 + ], + "spans": [ + { + "bbox": [ + 130, + 553, + 482, + 636 + ], + "type": "text", + "content": "Datasets. We collect eight real-world static scenes by capturing multi-view images. Each scene includes an object and a background. The objects include five flowers (a red rose, a carnation, an orange rose, a tulip, and a white rose), an alocasia plant, a telephone cord, and a beanie hat. For each scene except for the red rose scene, we capture four interaction videos illustrating its natural motion after interaction, such as poking or dragging, and we use the real videos as additional comparison references." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 132, + 641, + 481, + 666 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 641, + 481, + 666 + ], + "spans": [ + { + "bbox": [ + 132, + 641, + 481, + 666 + ], + "type": "text", + "content": "Baselines. We compare our approach to two baselines: PhysGaussian [70] and DreamGaussian4D [58]. PhysGaussian [70] integrates MPM simulation to static" + } + ] + } + ], + "index": 10 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 391, + 91, + 447, + 101 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 391, + 91, + 447, + 101 + ], + "spans": [ + { + "bbox": [ + 391, + 91, + 447, + 101 + ], + "type": "text", + "content": "PhysDreamer" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 474, + 91, + 481, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 474, + 91, + 481, + 100 + ], + "spans": [ + { + "bbox": [ + 474, + 91, + 481, + 100 + ], + "type": "text", + "content": "9" + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 8 + }, + { + "para_blocks": [ + { + "bbox": [ + 130, + 116, + 482, + 224 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 116, + 482, + 224 + ], + "spans": [ + { + "bbox": [ + 130, + 116, + 482, + 224 + ], + "type": "text", + "content": "3D Gaussians to support simulation, but it cannot estimate material properties and relies on manually setting material parameter values. Thus, we use the same initialization strategy as ours to assign material properties for PhysGaussian. DreamGaussian4D [58] generates non-interactive dynamic 3D Gaussians from a static image. It first obtains a static 3D Gaussians using DreamGaussian [63], and then animate it by optimizing a deformation field from a generated driving video. For a fair comparison, we run its deformation field optimization on our reconstructed static 3D Gaussians, and we looped the resulting deformation field when rendering longer videos in later comparison." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 130, + 228, + 482, + 312 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 228, + 482, + 312 + ], + "spans": [ + { + "bbox": [ + 130, + 228, + 482, + 312 + ], + "type": "text", + "content": "Evaluation metrics. We focus on the quality of the synthesized object motion, in particular, visual quality and motion realism. Therefore, we conduct a user study and adopt the Two-alternative Forced Choice (2AFC) protocol: the participants are shown two side-by-side synchronized videos, including one video result from ours and the other one from the competitor's, with a random left-right ordering. The participants are then asked to choose the one with higher visual quality and the one with higher motion realism." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 130, + 312, + 482, + 433 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 312, + 482, + 433 + ], + "spans": [ + { + "bbox": [ + 130, + 312, + 482, + 433 + ], + "type": "text", + "content": "We recruited 100 participants, each asked to judge all 8 scenes, forming a total of 800 2AFC judgement samples for each baseline comparison. For each scene, we create 4 sample video pairs and show participants a random one from the 4 pairs. In particular, we create 4 five-second motion sequences using PhysDreamer with randomized initial conditions (applying an external force to the foreground object or assigning an initial velocity to the object), and render videos from randomly picked viewpoints. For the baseline method, we apply the same initial conditions (for PhysGaussian only) and render videos from the same viewpoint as ours to form the video pairs. Please see supplementary materials for human study details and quantitative metrics for videos (e.g., Fréchet Video Distance [64])." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 132, + 451, + 278, + 464 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 451, + 278, + 464 + ], + "spans": [ + { + "bbox": [ + 132, + 451, + 278, + 464 + ], + "type": "text", + "content": "5.2 Implementation details" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 130, + 476, + 481, + 513 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 476, + 481, + 513 + ], + "spans": [ + { + "bbox": [ + 130, + 476, + 481, + 513 + ], + "type": "text", + "content": "Neural material fields. We represent both material field and initial velocity field using triplanes [54] each followed by a three-layer MLP. The triplanes have spatial resolutions of " + }, + { + "bbox": [ + 130, + 476, + 481, + 513 + ], + "type": "inline_equation", + "content": "8^{3}" + }, + { + "bbox": [ + 130, + 476, + 481, + 513 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 130, + 476, + 481, + 513 + ], + "type": "inline_equation", + "content": "24^{3}" + }, + { + "bbox": [ + 130, + 476, + 481, + 513 + ], + "type": "text", + "content": " for the material field and velocity field, respectively." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 130, + 517, + 482, + 553 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 517, + 482, + 553 + ], + "spans": [ + { + "bbox": [ + 130, + 517, + 482, + 553 + ], + "type": "text", + "content": "3D Gaussian reconstruction. Similar to PhysGaussian [70], we employ anisotropic regularization to reduce skinny artifacts in the reconstruction. Each reconstructed scene contains 0.5 to 1.5 million particles (including foreground and background)." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 130, + 558, + 482, + 666 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 558, + 482, + 666 + ], + "spans": [ + { + "bbox": [ + 130, + 558, + 482, + 666 + ], + "type": "text", + "content": "Simulation details. For computational efficiency, we segment the background and keep only foreground object particles for simulation. In our experiments, the foreground object contains around 50 to 300 thousand 3D Gaussian particles. We then discretize the foreground into a " + }, + { + "bbox": [ + 130, + 558, + 482, + 666 + ], + "type": "inline_equation", + "content": "64^3" + }, + { + "bbox": [ + 130, + 558, + 482, + 666 + ], + "type": "text", + "content": " grid. The number of driving particles are 10 to 50 times fewer than the number of 3D Gaussian particles, determined by maintaining an average of at least eight particles per occupied voxel. For accurate motion, we use 768 sub-steps between successive video frames, corresponding to a duration of " + }, + { + "bbox": [ + 130, + 558, + 482, + 666 + ], + "type": "inline_equation", + "content": "4.34 \\times 10^{-5}" + }, + { + "bbox": [ + 130, + 558, + 482, + 666 + ], + "type": "text", + "content": " second for each sub-step. To address the high memory consumption from large number of steps, we apply simulation" + } + ] + } + ], + "index": 8 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "spans": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "type": "text", + "content": "10" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 166, + 91, + 217, + 102 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 166, + 91, + 217, + 102 + ], + "spans": [ + { + "bbox": [ + 166, + 91, + 217, + 102 + ], + "type": "text", + "content": "Zhang et al." + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 9 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 134, + 115, + 202, + 182 + ], + "blocks": [ + { + "bbox": [ + 134, + 115, + 202, + 182 + ], + "lines": [ + { + "bbox": [ + 134, + 115, + 202, + 182 + ], + "spans": [ + { + "bbox": [ + 134, + 115, + 202, + 182 + ], + "type": "image", + "image_path": "946f253765b04ae38e1dc2c49e75f817e629d35516977c74d58349625fbfbc08.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_body" + } + ], + "index": 2 + }, + { + "type": "image", + "bbox": [ + 134, + 184, + 201, + 251 + ], + "blocks": [ + { + "bbox": [ + 134, + 184, + 201, + 251 + ], + "lines": [ + { + "bbox": [ + 134, + 184, + 201, + 251 + ], + "spans": [ + { + "bbox": [ + 134, + 184, + 201, + 251 + ], + "type": "image", + "image_path": "053cf1b7da5d2e83dd5ca37013eec6bda5e658e0a3361f5515c3641af945165e.jpg" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_body" + } + ], + "index": 3 + }, + { + "type": "image", + "bbox": [ + 135, + 252, + 201, + 319 + ], + "blocks": [ + { + "bbox": [ + 135, + 252, + 201, + 319 + ], + "lines": [ + { + "bbox": [ + 135, + 252, + 201, + 319 + ], + "spans": [ + { + "bbox": [ + 135, + 252, + 201, + 319 + ], + "type": "image", + "image_path": "0a09853279d10b314b3874f37167548f2d2f7ffde004df071fe8e6a77f088dc6.jpg" + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "image_body" + } + ], + "index": 4 + }, + { + "type": "image", + "bbox": [ + 135, + 320, + 201, + 387 + ], + "blocks": [ + { + "bbox": [ + 135, + 320, + 201, + 387 + ], + "lines": [ + { + "bbox": [ + 135, + 320, + 201, + 387 + ], + "spans": [ + { + "bbox": [ + 135, + 320, + 201, + 387 + ], + "type": "image", + "image_path": "c3f9ad2754052b701c080b77525609ebe5c885c54afd03d0e843a691542b4c83.jpg" + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 143, + 388, + 192, + 396 + ], + "lines": [ + { + "bbox": [ + 143, + 388, + 192, + 396 + ], + "spans": [ + { + "bbox": [ + 143, + 388, + 192, + 396 + ], + "type": "text", + "content": "Material result" + } + ] + } + ], + "index": 6, + "angle": 0, + "type": "image_caption" + } + ], + "index": 5 + }, + { + "type": "image", + "bbox": [ + 209, + 115, + 277, + 182 + ], + "blocks": [ + { + "bbox": [ + 209, + 115, + 277, + 182 + ], + "lines": [ + { + "bbox": [ + 209, + 115, + 277, + 182 + ], + "spans": [ + { + "bbox": [ + 209, + 115, + 277, + 182 + ], + "type": "image", + "image_path": "f52f355d1637ff52888eca5721f7d66d5b296aef2268ff58c9cee5633c5a7268.jpg" + } + ] + } + ], + "index": 7, + "angle": 0, + "type": "image_body" + } + ], + "index": 7 + }, + { + "type": "image", + "bbox": [ + 209, + 183, + 277, + 250 + ], + "blocks": [ + { + "bbox": [ + 209, + 183, + 277, + 250 + ], + "lines": [ + { + "bbox": [ + 209, + 183, + 277, + 250 + ], + "spans": [ + { + "bbox": [ + 209, + 183, + 277, + 250 + ], + "type": "image", + "image_path": "0f1576aae8aca2b0343ca8dde702b468c47e2557a5d99e8495d763ab01172f63.jpg" + } + ] + } + ], + "index": 8, + "angle": 0, + "type": "image_body" + } + ], + "index": 8 + }, + { + "type": "image", + "bbox": [ + 209, + 252, + 277, + 319 + ], + "blocks": [ + { + "bbox": [ + 209, + 252, + 277, + 319 + ], + "lines": [ + { + "bbox": [ + 209, + 252, + 277, + 319 + ], + "spans": [ + { + "bbox": [ + 209, + 252, + 277, + 319 + ], + "type": "image", + "image_path": "4bab3e82c0c15ae265bb2563b393a2c4a915af1e842bb3c5c0e36a09d628fe71.jpg" + } + ] + } + ], + "index": 9, + "angle": 0, + "type": "image_body" + } + ], + "index": 9 + }, + { + "type": "image", + "bbox": [ + 209, + 320, + 277, + 387 + ], + "blocks": [ + { + "bbox": [ + 209, + 320, + 277, + 387 + ], + "lines": [ + { + "bbox": [ + 209, + 320, + 277, + 387 + ], + "spans": [ + { + "bbox": [ + 209, + 320, + 277, + 387 + ], + "type": "image", + "image_path": "a5fdebb31b0f2ed5b2589b21307e488600fa3e859eec56945a20e8bb7345e7bc.jpg" + } + ] + } + ], + "index": 10, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 272, + 387, + 415, + 396 + ], + "lines": [ + { + "bbox": [ + 272, + 387, + 415, + 396 + ], + "spans": [ + { + "bbox": [ + 272, + 387, + 415, + 396 + ], + "type": "text", + "content": "Synthesized 3D dynamics under interaction" + } + ] + } + ], + "index": 23, + "angle": 0, + "type": "image_caption" + } + ], + "index": 10 + }, + { + "type": "image", + "bbox": [ + 278, + 116, + 344, + 182 + ], + "blocks": [ + { + "bbox": [ + 278, + 116, + 344, + 182 + ], + "lines": [ + { + "bbox": [ + 278, + 116, + 344, + 182 + ], + "spans": [ + { + "bbox": [ + 278, + 116, + 344, + 182 + ], + "type": "image", + "image_path": "6c9d4c85d30d4774ce78b6e056bb37c23ba99c8e1f3ae167608bc389055cc4fa.jpg" + } + ] + } + ], + "index": 11, + "angle": 0, + "type": "image_body" + } + ], + "index": 11 + }, + { + "type": "image", + "bbox": [ + 278, + 183, + 344, + 250 + ], + "blocks": [ + { + "bbox": [ + 278, + 183, + 344, + 250 + ], + "lines": [ + { + "bbox": [ + 278, + 183, + 344, + 250 + ], + "spans": [ + { + "bbox": [ + 278, + 183, + 344, + 250 + ], + "type": "image", + "image_path": "337d341034383ccdf26e3a24cade1c7cdad3ec139ade2bab18c182c0d78dc01c.jpg" + } + ] + } + ], + "index": 12, + "angle": 0, + "type": "image_body" + } + ], + "index": 12 + }, + { + "type": "image", + "bbox": [ + 278, + 252, + 344, + 319 + ], + "blocks": [ + { + "bbox": [ + 278, + 252, + 344, + 319 + ], + "lines": [ + { + "bbox": [ + 278, + 252, + 344, + 319 + ], + "spans": [ + { + "bbox": [ + 278, + 252, + 344, + 319 + ], + "type": "image", + "image_path": "5716a94ce533fb9308ccb354cc453be3ef37a13eb0d5e1e3279537dea0674cc7.jpg" + } + ] + } + ], + "index": 13, + "angle": 0, + "type": "image_body" + } + ], + "index": 13 + }, + { + "type": "image", + "bbox": [ + 278, + 320, + 344, + 387 + ], + "blocks": [ + { + "bbox": [ + 278, + 320, + 344, + 387 + ], + "lines": [ + { + "bbox": [ + 278, + 320, + 344, + 387 + ], + "spans": [ + { + "bbox": [ + 278, + 320, + 344, + 387 + ], + "type": "image", + "image_path": "b7b74ebd85c4e1adb9f53a30edfc3c1475bf373e7809fa9944717115d83a8b80.jpg" + } + ] + } + ], + "index": 14, + "angle": 0, + "type": "image_body" + } + ], + "index": 14 + }, + { + "type": "image", + "bbox": [ + 345, + 116, + 411, + 182 + ], + "blocks": [ + { + "bbox": [ + 345, + 116, + 411, + 182 + ], + "lines": [ + { + "bbox": [ + 345, + 116, + 411, + 182 + ], + "spans": [ + { + "bbox": [ + 345, + 116, + 411, + 182 + ], + "type": "image", + "image_path": "02dbaa6ebc3d11669c44674fe95a763c81c3c9e3d83fd92291fbfec7da0929dd.jpg" + } + ] + } + ], + "index": 15, + "angle": 0, + "type": "image_body" + } + ], + "index": 15 + }, + { + "type": "image", + "bbox": [ + 345, + 183, + 411, + 250 + ], + "blocks": [ + { + "bbox": [ + 345, + 183, + 411, + 250 + ], + "lines": [ + { + "bbox": [ + 345, + 183, + 411, + 250 + ], + "spans": [ + { + "bbox": [ + 345, + 183, + 411, + 250 + ], + "type": "image", + "image_path": "a06f62dd64ef27937347af8c8bf97faead16694ec747269be5817bb9a20b0347.jpg" + } + ] + } + ], + "index": 16, + "angle": 0, + "type": "image_body" + } + ], + "index": 16 + }, + { + "type": "image", + "bbox": [ + 345, + 252, + 411, + 319 + ], + "blocks": [ + { + "bbox": [ + 345, + 252, + 411, + 319 + ], + "lines": [ + { + "bbox": [ + 345, + 252, + 411, + 319 + ], + "spans": [ + { + "bbox": [ + 345, + 252, + 411, + 319 + ], + "type": "image", + "image_path": "8d4efb73ec74dbb4774f57ff85479ac556475ea417956382742968b89b8a4675.jpg" + } + ] + } + ], + "index": 17, + "angle": 0, + "type": "image_body" + } + ], + "index": 17 + }, + { + "type": "image", + "bbox": [ + 345, + 320, + 411, + 387 + ], + "blocks": [ + { + "bbox": [ + 345, + 320, + 411, + 387 + ], + "lines": [ + { + "bbox": [ + 345, + 320, + 411, + 387 + ], + "spans": [ + { + "bbox": [ + 345, + 320, + 411, + 387 + ], + "type": "image", + "image_path": "43ca094db8f382473e70d106b35d711f79e879a8f18d9341048d8bd6853880cc.jpg" + } + ] + } + ], + "index": 18, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 130, + 406, + 482, + 451 + ], + "lines": [ + { + "bbox": [ + 130, + 406, + 482, + 451 + ], + "spans": [ + { + "bbox": [ + 130, + 406, + 482, + 451 + ], + "type": "text", + "content": "Fig. 5: Interactive 3D dynamics synthesis. (Left) Visualization of the material fields. Brighter color indicates higher Young's modulus within each example. (Right) We apply an external force (red arrow) on each object, and the following columns demonstrate the object dynamics rendered at a static viewpoint." + } + ] + } + ], + "index": 24, + "angle": 0, + "type": "image_caption" + } + ], + "index": 18 + }, + { + "type": "image", + "bbox": [ + 413, + 116, + 479, + 182 + ], + "blocks": [ + { + "bbox": [ + 413, + 116, + 479, + 182 + ], + "lines": [ + { + "bbox": [ + 413, + 116, + 479, + 182 + ], + "spans": [ + { + "bbox": [ + 413, + 116, + 479, + 182 + ], + "type": "image", + "image_path": "7e3125339de5c304c095b6dc7a7bf7109f8a2c0fe45c2161aaeccf906496d144.jpg" + } + ] + } + ], + "index": 19, + "angle": 0, + "type": "image_body" + } + ], + "index": 19 + }, + { + "type": "image", + "bbox": [ + 413, + 183, + 479, + 250 + ], + "blocks": [ + { + "bbox": [ + 413, + 183, + 479, + 250 + ], + "lines": [ + { + "bbox": [ + 413, + 183, + 479, + 250 + ], + "spans": [ + { + "bbox": [ + 413, + 183, + 479, + 250 + ], + "type": "image", + "image_path": "5f10fed9c66aad52ab6aa0224ec2db785d0a30f116f0fabcfde14aa3f884664d.jpg" + } + ] + } + ], + "index": 20, + "angle": 0, + "type": "image_body" + } + ], + "index": 20 + }, + { + "type": "image", + "bbox": [ + 413, + 252, + 479, + 319 + ], + "blocks": [ + { + "bbox": [ + 413, + 252, + 479, + 319 + ], + "lines": [ + { + "bbox": [ + 413, + 252, + 479, + 319 + ], + "spans": [ + { + "bbox": [ + 413, + 252, + 479, + 319 + ], + "type": "image", + "image_path": "311a2cc51be66ebcac18b5ec388b973555d744230d954b2b00a8227b2024959c.jpg" + } + ] + } + ], + "index": 21, + "angle": 0, + "type": "image_body" + } + ], + "index": 21 + }, + { + "type": "image", + "bbox": [ + 413, + 320, + 479, + 387 + ], + "blocks": [ + { + "bbox": [ + 413, + 320, + 479, + 387 + ], + "lines": [ + { + "bbox": [ + 413, + 320, + 479, + 387 + ], + "spans": [ + { + "bbox": [ + 413, + 320, + 479, + 387 + ], + "type": "image", + "image_path": "3c4a3fe009af3da3776db34998d3dc8c17fa89fc447387c6d734ca5f7d9a4656.jpg" + } + ] + } + ], + "index": 22, + "angle": 0, + "type": "image_body" + } + ], + "index": 22 + }, + { + "bbox": [ + 130, + 466, + 482, + 502 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 466, + 482, + 502 + ], + "spans": [ + { + "bbox": [ + 130, + 466, + 482, + 502 + ], + "type": "text", + "content": "state checkpointing and re-computation during gradient back-propagation. We add Dirichlet boundary conditions for stationary grid cells. We fill the internal volumes of certain solid objects to enhance simulation realism [70]." + } + ] + } + ], + "index": 25 + }, + { + "bbox": [ + 130, + 506, + 482, + 590 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 506, + 482, + 590 + ], + "spans": [ + { + "bbox": [ + 130, + 506, + 482, + 590 + ], + "type": "text", + "content": "Generating reference videos. We render a 3D object with its background from a viewpoint, and then we use Stable Video Diffusion [4] to animate this rendered image and generate fourteen video frames. We use a small motion bucket number [4] (e.g., 5 or 8) so that the generated video contains mostly object motion and little camera motion. We use rendered images for the video generation, so that our approach can also be used for generated scenes. Also, rendering images directly from 3D Gaussians simplifies later optimization." + } + ] + } + ], + "index": 26 + }, + { + "bbox": [ + 131, + 597, + 198, + 609 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 131, + 597, + 198, + 609 + ], + "spans": [ + { + "bbox": [ + 131, + 597, + 198, + 609 + ], + "type": "text", + "content": "5.3 Results" + } + ] + } + ], + "index": 27 + }, + { + "bbox": [ + 130, + 617, + 482, + 666 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 617, + 482, + 666 + ], + "spans": [ + { + "bbox": [ + 130, + 617, + 482, + 666 + ], + "type": "text", + "content": "We show our qualitative results of the spatially-varying Young's modulus in Fig. 5 (left), and simulated interactive motion in Fig. 5 (right). Please see our project website videos for a better motion visualization. Tab. 1 presents the user study results in comparison to baseline methods and real captured videos." + } + ] + } + ], + "index": 28 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 391, + 91, + 447, + 101 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 391, + 91, + 447, + 101 + ], + "spans": [ + { + "bbox": [ + 391, + 91, + 447, + 101 + ], + "type": "text", + "content": "PhysDreamer" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 470, + 91, + 480, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 470, + 91, + 480, + 100 + ], + "spans": [ + { + "bbox": [ + 470, + 91, + 480, + 100 + ], + "type": "text", + "content": "11" + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 10 + }, + { + "para_blocks": [ + { + "type": "table", + "bbox": [ + 138, + 168, + 477, + 290 + ], + "blocks": [ + { + "bbox": [ + 132, + 114, + 482, + 159 + ], + "lines": [ + { + "bbox": [ + 132, + 114, + 482, + 159 + ], + "spans": [ + { + "bbox": [ + 132, + 114, + 482, + 159 + ], + "type": "text", + "content": "Table 1: Human study 2AFC results of PhysDreamer (Ours) over real captured videos and baseline methods (PhysGaussian [70] and DreamGaussian4D [58]) on Motion Realism and overall Visual Quality. \"Rose O\", \"Rose W\", and \"Rose R\" denotes the orange, white, and red roses, respectively." + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 138, + 168, + 477, + 290 + ], + "lines": [ + { + "bbox": [ + 138, + 168, + 477, + 290 + ], + "spans": [ + { + "bbox": [ + 138, + 168, + 477, + 290 + ], + "type": "table", + "html": "
Motion realismAlocasiaCarnationHatRose ORose WRose RCordTulipAvg.
Ours over Real capture86%61%55%63%47%-29%35%53.7%
Ours over PhysGaussian96%89%57%91%93%73%61%86%80.8%
Ours over DreamGaussian75%77%51%78%51%41%71%64%63.5%
Visual quality
Ours over Real capture36%53%28%40%41%-29%34%37.3%
Ours over PhysGaussian67%69%50%75%73%58%58%70%65.0%
Ours over DreamGaussian82%75%74%76%60%47%76%70%70.0%
", + "image_path": "0b4d8778a0a85d5140c8960b026b50b042927ec47a5a1ab956898eb4abfa08a5.jpg" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "table_body" + } + ], + "index": 3 + }, + { + "bbox": [ + 130, + 312, + 481, + 384 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 312, + 481, + 384 + ], + "spans": [ + { + "bbox": [ + 130, + 312, + 481, + 384 + ], + "type": "text", + "content": "Compared to PhysGaussian, " + }, + { + "bbox": [ + 130, + 312, + 481, + 384 + ], + "type": "inline_equation", + "content": "80.8\\%" + }, + { + "bbox": [ + 130, + 312, + 481, + 384 + ], + "type": "text", + "content": " of the human participant 2AFC samples prefer PhysDreamer (ours) in motion realism and " + }, + { + "bbox": [ + 130, + 312, + 481, + 384 + ], + "type": "inline_equation", + "content": "65.0\\%" + }, + { + "bbox": [ + 130, + 312, + 481, + 384 + ], + "type": "text", + "content": " prefer PhysDreamer in visual quality. Note that since the static scenes are the same, the visual quality also depends on the generated object motion. Fig. 6 shows temporal slices of the motion patterns. We observe that PhysGaussian produces large, unrealistic slow motion due to the lack of a principled estimation of material properties." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 130, + 384, + 481, + 492 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 384, + 481, + 492 + ], + "spans": [ + { + "bbox": [ + 130, + 384, + 481, + 492 + ], + "type": "text", + "content": "Compared to DreamGaussian4D, " + }, + { + "bbox": [ + 130, + 384, + 481, + 492 + ], + "type": "inline_equation", + "content": "70.0\\% / 63.5\\%" + }, + { + "bbox": [ + 130, + 384, + 481, + 492 + ], + "type": "text", + "content": " 2AFC samples prefer ours in visual quality and motion realism, respectively. From Fig. 6, we can observe that DreamGaussian4D generates periodic motion with a constant, small magnitude, while PhysDreamer can simulate the damping in motion. This is because DreamGaussian4D does not simulate the physical dynamics but simply distill a motion sequence from a generative model, so it cannot extrapolate to different motion. We further include one more evaluation dimension on \"motion amount\" comparing to DreamGaussian4D, where we ask the participants to judge which video has higher amount of motion, and " + }, + { + "bbox": [ + 130, + 384, + 481, + 492 + ], + "type": "inline_equation", + "content": "73.6\\%" + }, + { + "bbox": [ + 130, + 384, + 481, + 492 + ], + "type": "text", + "content": " 2AFC samples prefer PhysDreamer." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 130, + 493, + 482, + 601 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 493, + 482, + 601 + ], + "spans": [ + { + "bbox": [ + 130, + 493, + 482, + 601 + ], + "type": "text", + "content": "Compared to real videos, " + }, + { + "bbox": [ + 130, + 493, + 482, + 601 + ], + "type": "inline_equation", + "content": "53.7\\%" + }, + { + "bbox": [ + 130, + 493, + 482, + 601 + ], + "type": "text", + "content": " 2AFC samples favored the motion realism of ours results. Interestingly, under \"Motion Realism\", " + }, + { + "bbox": [ + 130, + 493, + 482, + 601 + ], + "type": "inline_equation", + "content": "86\\%" + }, + { + "bbox": [ + 130, + 493, + 482, + 601 + ], + "type": "text", + "content": " of the users indicated that the alocasia outputs were more realistic than real captures. This is surprising, as one would expect a " + }, + { + "bbox": [ + 130, + 493, + 482, + 601 + ], + "type": "inline_equation", + "content": "50\\%" + }, + { + "bbox": [ + 130, + 493, + 482, + 601 + ], + "type": "text", + "content": " preference if the videos were indistinguishable. We offer a potential explanation: for thin geometries like alocasia leaves, the Material Point Method tends to produce lower-frequency and slower motions. This can be observed in the video and is evident in the space-time slice visualizations in Fig. 6. Humans are poor at judging the naturalness of motion and may be biased towards smoother and slower motions, as shown in prior studies [36,61]." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 132, + 620, + 379, + 632 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 620, + 379, + 632 + ], + "spans": [ + { + "bbox": [ + 132, + 620, + 379, + 632 + ], + "type": "text", + "content": "5.4 Ablation: using multi-view reference videos" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 130, + 641, + 481, + 665 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 641, + 481, + 665 + ], + "spans": [ + { + "bbox": [ + 130, + 641, + 481, + 665 + ], + "type": "text", + "content": "For objects with self-occlusion, observing salient motion of all object parts from a single video is challenging (e.g., the alocasia scene where a leaf can occlude" + } + ] + } + ], + "index": 8 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "spans": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "type": "text", + "content": "12" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 166, + 91, + 217, + 102 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 166, + 91, + 217, + 102 + ], + "spans": [ + { + "bbox": [ + 166, + 91, + 217, + 102 + ], + "type": "text", + "content": "Zhang et al." + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 11 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 133, + 114, + 480, + 574 + ], + "blocks": [ + { + "bbox": [ + 133, + 114, + 480, + 574 + ], + "lines": [ + { + "bbox": [ + 133, + 114, + 480, + 574 + ], + "spans": [ + { + "bbox": [ + 133, + 114, + 480, + 574 + ], + "type": "image", + "image_path": "8417644f2a8bb83365516b07e17e7b2305058377a2f517a2871d2912989ce5cd.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 130, + 580, + 482, + 647 + ], + "lines": [ + { + "bbox": [ + 130, + 580, + 482, + 647 + ], + "spans": [ + { + "bbox": [ + 130, + 580, + 482, + 647 + ], + "type": "text", + "content": "Fig. 6: We compare our results with real captured videos, PhysGaussian [70], and DreamGaussian4D [58] using space-time slices. In these slices, the vertical axis represent time, and the horizontal axis shows a spatial slice of the object (denoted by red lines on the \"object\" column). These slices visualize the magnitude and frequencies of these oscillating motions. Results for our PhysDreamer (Ours) and PhysGaussian are simulated with the same initial conditions." + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_caption" + } + ], + "index": 2 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 391, + 91, + 447, + 102 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 391, + 91, + 447, + 102 + ], + "spans": [ + { + "bbox": [ + 391, + 91, + 447, + 102 + ], + "type": "text", + "content": "PhysDreamer" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 470, + 91, + 481, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 470, + 91, + 481, + 100 + ], + "spans": [ + { + "bbox": [ + 470, + 91, + 481, + 100 + ], + "type": "text", + "content": "13" + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 12 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 133, + 114, + 480, + 223 + ], + "blocks": [ + { + "bbox": [ + 133, + 114, + 480, + 223 + ], + "lines": [ + { + "bbox": [ + 133, + 114, + 480, + 223 + ], + "spans": [ + { + "bbox": [ + 133, + 114, + 480, + 223 + ], + "type": "image", + "image_path": "c0c3bf17c0e32228530761b384ebe146d83190f582b74d2036ea615c1794193c.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 130, + 234, + 482, + 280 + ], + "lines": [ + { + "bbox": [ + 130, + 234, + 482, + 280 + ], + "spans": [ + { + "bbox": [ + 130, + 234, + 482, + 280 + ], + "type": "text", + "content": "Fig. 7: Comparison between single-view (top) and two-view (bottom) supervisions. The object (alocasia) exhibits self-occluding structures. We can use generated videos at two views to jointly optimize the material field. In the space-time (X-t) slices, the vertical axis represents time, and the horizontal axis shows a spatial slice of the object." + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_caption" + } + ], + "index": 2 + }, + { + "bbox": [ + 130, + 297, + 482, + 416 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 297, + 482, + 416 + ], + "spans": [ + { + "bbox": [ + 130, + 297, + 482, + 416 + ], + "type": "text", + "content": "another leaf). We may alleviate this problem by rendering from multiple viewpoints to provide comprehensive coverage of the object. Here, we use multiple videos in the material estimation, jointly optimizing a video-agnostic, spatially-varying Young's modulus for each particle along with video-specific initial velocities. From the comparison of the alocasia scene in Fig. 7, we can see that using multi-view reference videos (a front view and a back view) helps in such complex self-occluding objects: PhysDreamer benefits significantly from having supervision from two views, while using only a single view leads to artifacts. In our user study, " + }, + { + "bbox": [ + 130, + 297, + 482, + 416 + ], + "type": "inline_equation", + "content": "81.0\\%" + }, + { + "bbox": [ + 130, + 297, + 482, + 416 + ], + "type": "text", + "content": " 2AFC samples prefer results with two view supervision in visual quality and " + }, + { + "bbox": [ + 130, + 297, + 482, + 416 + ], + "type": "inline_equation", + "content": "86.0\\%" + }, + { + "bbox": [ + 130, + 297, + 482, + 416 + ], + "type": "text", + "content": " in motion realism." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 132, + 439, + 220, + 451 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 439, + 220, + 451 + ], + "spans": [ + { + "bbox": [ + 132, + 439, + 220, + 451 + ], + "type": "text", + "content": "6 Conclusion" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 130, + 468, + 482, + 564 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 468, + 482, + 564 + ], + "spans": [ + { + "bbox": [ + 130, + 468, + 482, + 564 + ], + "type": "text", + "content": "In this work, we introduced PhysDreamer, a novel approach to synthesizing interactive 3D dynamics by endowing static 3D objects with physical material properties. Our method distills the object dynamics priors learned by video generation models to estimate the spatially-varying material properties. We showcase dynamics interaction with a diverse set of elastic objects by PhysDreamer. We believe that PhysDreamer takes a significant step towards creating more engaging and immersive virtual environments, opening up a wide range of applications from realistic simulations to interactive virtual experiences." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 130, + 570, + 482, + 665 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 570, + 482, + 665 + ], + "spans": [ + { + "bbox": [ + 130, + 570, + 482, + 665 + ], + "type": "text", + "content": "Limitations. Our approach requires the user to manually specify the object to simulate and separate it from the background, and establish boundary conditions for stationary parts, like the pot of flowers. 3D object discovery may help for simulatable object extraction. In addition, our approach is computationally demanding. Despite our subsampling strategy, our current algorithm takes approximately one minute on a NVIDIA V100 GPU to produce a single second of video. Further improving efficiency remains an important future problem. Finally, in this work, we restrict our scope to elastic objects without collisions." + } + ] + } + ], + "index": 7 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "spans": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "type": "text", + "content": "14" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 166, + 91, + 217, + 102 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 166, + 91, + 217, + 102 + ], + "spans": [ + { + "bbox": [ + 166, + 91, + 217, + 102 + ], + "type": "text", + "content": "Zhang et al." + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 13 + }, + { + "para_blocks": [ + { + "bbox": [ + 130, + 116, + 482, + 224 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 116, + 482, + 224 + ], + "spans": [ + { + "bbox": [ + 130, + 116, + 482, + 224 + ], + "type": "text", + "content": "Acknowledgements. This work is in part supported by the NSF PHY-2019786 (The NSF AI Institute for Artificial Intelligence and Fundamental Interactions, http://iaifi.org/), NSF CIF 1955864 (Occlusion and Directional Resolution in Computational Imaging), RI #2211258, #2338203, ONR MURI N00014-22-1-2740, Quanta Computer, Samsung, and United States Air Force Artificial Intelligence Accelerator and was accomplished under Cooperative Agreement Number FA8750-192-1000. We would like to thank Peter Yichen Chen, Zhengqi Li, Pingchuan Ma, Minghao Guo, Ge Yang, and Shai Avidan for help and insightful discussions." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 132, + 246, + 198, + 258 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 246, + 198, + 258 + ], + "spans": [ + { + "bbox": [ + 132, + 246, + 198, + 258 + ], + "type": "text", + "content": "References" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 133, + 274, + 481, + 666 + ], + "type": "list", + "angle": 0, + "index": 14, + "blocks": [ + { + "bbox": [ + 138, + 274, + 481, + 319 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 274, + 481, + 319 + ], + "spans": [ + { + "bbox": [ + 138, + 274, + 481, + 319 + ], + "type": "text", + "content": "1. Attal, B., Huang, J.B., Richardt, C., Zollhoefer, M., Kopf, J., O'Toole, M., Kim, C.: Hyperreel: High-fidelity 6-dof video with ray-conditioned sampling. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 16610-16620 (2023)" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 138, + 319, + 481, + 363 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 319, + 481, + 363 + ], + "spans": [ + { + "bbox": [ + 138, + 319, + 481, + 363 + ], + "type": "text", + "content": "2. Bahmani, S., Skorokhodov, I., Rong, V., Wetzstein, G., Guibas, L., Wonka, P., Tulyakov, S., Park, J.J., Tagliasacchi, A., Lindell, D.B.: 4d-fy: Text-to-4d generation using hybrid score distillation sampling. arXiv preprint arXiv:2311.17984 (2023)" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 138, + 364, + 481, + 397 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 364, + 481, + 397 + ], + "spans": [ + { + "bbox": [ + 138, + 364, + 481, + 397 + ], + "type": "text", + "content": "3. Bar-Tal, O., Chefer, H., Tov, O., Herrmann, C., Paiss, R., Zada, S., Ephrat, A., Hur, J., Li, Y., Michaeli, T., et al.: Lumiere: A space-time diffusion model for video generation. arXiv preprint arXiv:2401.12945 (2024)" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 138, + 397, + 481, + 441 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 397, + 481, + 441 + ], + "spans": [ + { + "bbox": [ + 138, + 397, + 481, + 441 + ], + "type": "text", + "content": "4. Blattmann, A., Dockhorn, T., Kulal, S., Mendelevitch, D., Kilian, M., Lorenz, D., Levi, Y., English, Z., Voleti, V., Letts, A., et al.: Stable video diffusion: Scaling latent video diffusion models to large datasets. arXiv preprint arXiv:2311.15127 (2023)" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 138, + 441, + 481, + 486 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 441, + 481, + 486 + ], + "spans": [ + { + "bbox": [ + 138, + 441, + 481, + 486 + ], + "type": "text", + "content": "5. Blattmann, A., Rombach, R., Ling, H., Dockhorn, T., Kim, S.W., Fidler, S., Kreis, K.: Align your latents: High-resolution video synthesis with latent diffusion models. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 22563-22575 (2023)" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 138, + 487, + 481, + 531 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 487, + 481, + 531 + ], + "spans": [ + { + "bbox": [ + 138, + 487, + 481, + 531 + ], + "type": "text", + "content": "6. Brooks, T., Peebles, B., Homes, C., DePue, W., Guo, Y., Jing, L., Schnurr, D., Taylor, J., Luhman, T., Luhman, E., Ng, C., Wang, R., Ramesh, A.: Video generation models as world simulators (2024), https://openai.com/research/video-generation-models-as-world-simulators" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 138, + 531, + 481, + 553 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 531, + 481, + 553 + ], + "spans": [ + { + "bbox": [ + 138, + 531, + 481, + 553 + ], + "type": "text", + "content": "7. Cai, Y., Wang, J., Yuille, A., Zhou, Z., Wang, A.: Structure-aware sparse-view x-ray 3d reconstruction. In: CVPR (2024)" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 138, + 554, + 481, + 587 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 554, + 481, + 587 + ], + "spans": [ + { + "bbox": [ + 138, + 554, + 481, + 587 + ], + "type": "text", + "content": "8. Cao, A., Johnson, J.: Hexplane: A fast representation for dynamic scenes. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 130-141 (2023)" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 138, + 587, + 481, + 632 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 587, + 481, + 632 + ], + "spans": [ + { + "bbox": [ + 138, + 587, + 481, + 632 + ], + "type": "text", + "content": "9. Chan, E.R., Lin, C.Z., Chan, M.A., Nagano, K., Pan, B., De Mello, S., Gallo, O., Guibas, L.J., Tremblay, J., Khamis, S., et al.: Efficient geometry-aware 3d generative adversarial networks. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 16123-16133 (2022)" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 133, + 632, + 481, + 666 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 632, + 481, + 666 + ], + "spans": [ + { + "bbox": [ + 133, + 632, + 481, + 666 + ], + "type": "text", + "content": "10. Chen, H.y., Tretschk, E., Stuyck, T., Kadlecek, P., Kavan, L., Vouga, E., Lassner, C.: Virtual elastic objects. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 15827-15837 (2022)" + } + ] + } + ], + "index": 13 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 391, + 91, + 447, + 101 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 391, + 91, + 447, + 101 + ], + "spans": [ + { + "bbox": [ + 391, + 91, + 447, + 101 + ], + "type": "text", + "content": "PhysDreamer" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 470, + 91, + 481, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 470, + 91, + 481, + 100 + ], + "spans": [ + { + "bbox": [ + 470, + 91, + 481, + 100 + ], + "type": "text", + "content": "15" + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 14 + }, + { + "para_blocks": [ + { + "bbox": [ + 132, + 116, + 481, + 666 + ], + "type": "list", + "angle": 0, + "index": 19, + "blocks": [ + { + "bbox": [ + 133, + 116, + 481, + 149 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 116, + 481, + 149 + ], + "spans": [ + { + "bbox": [ + 133, + 116, + 481, + 149 + ], + "type": "text", + "content": "11. Chen, X., Liu, Z., Chen, M., Feng, Y., Liu, Y., Shen, Y., Zhao, H.: Livephoto: Real image animation with text-guided motion control. arXiv preprint arXiv:2312.02928 (2023)" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 133, + 150, + 481, + 183 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 150, + 481, + 183 + ], + "spans": [ + { + "bbox": [ + 133, + 150, + 481, + 183 + ], + "type": "text", + "content": "12. Chuang, Y.Y., Goldman, D.B., Zheng, K.C., Curless, B., Salesin, D.H., Szeliski, R.: Animating pictures with stochastic motion textures. In: ACM SIGGRAPH 2005 Papers. pp. 853-860 (2005)" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 132, + 183, + 481, + 216 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 183, + 481, + 216 + ], + "spans": [ + { + "bbox": [ + 132, + 183, + 481, + 216 + ], + "type": "text", + "content": "13. Curless, B., Levoy, M.: A volumetric method for building complex models from range images. In: Proceedings of the 23rd annual conference on Computer graphics and interactive techniques. pp. 303-312 (1996)" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 132, + 216, + 481, + 248 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 216, + 481, + 248 + ], + "spans": [ + { + "bbox": [ + 132, + 216, + 481, + 248 + ], + "type": "text", + "content": "14. Dai, Z., Zhang, Z., Yao, Y., Qiu, B., Zhu, S., Qin, L., Wang, W.: Animateanything: Fine-grained open domain image animation with motion guidance. arXiv e-prints pp. arXiv-2311 (2023)" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 132, + 249, + 481, + 281 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 249, + 481, + 281 + ], + "spans": [ + { + "bbox": [ + 132, + 249, + 481, + 281 + ], + "type": "text", + "content": "15. Davis, A., Chen, J.G., Durand, F.: Image-space modal bases for plausible manipulation of objects in video. ACM Transactions on Graphics (TOG) 34(6), 1-7 (2015)" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 132, + 281, + 481, + 304 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 281, + 481, + 304 + ], + "spans": [ + { + "bbox": [ + 132, + 281, + 481, + 304 + ], + "type": "text", + "content": "16. Davis, M.A.: Visual vibration analysis. Ph.D. thesis, Massachusetts Institute of Technology (2016)" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 132, + 304, + 481, + 335 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 304, + 481, + 335 + ], + "spans": [ + { + "bbox": [ + 132, + 304, + 481, + 335 + ], + "type": "text", + "content": "17. Duan, Y., Wei, F., Dai, Q., He, Y., Chen, W., Chen, B.: 4d gaussian splatt-ting: Towards efficient novel view synthesis for dynamic scenes. arXiv preprint arXiv:2402.03307 (2024)" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 132, + 336, + 481, + 357 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 336, + 481, + 357 + ], + "spans": [ + { + "bbox": [ + 132, + 336, + 481, + 357 + ], + "type": "text", + "content": "18. Feng, Y., Shang, Y., Li, X., Shao, T., Jiang, C., Yang, Y.: Pie-nerf: Physics-based interactive elastodynamics with nerf. arXiv preprint arXiv:2311.13099 (2023)" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 132, + 358, + 481, + 402 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 358, + 481, + 402 + ], + "spans": [ + { + "bbox": [ + 132, + 358, + 481, + 402 + ], + "type": "text", + "content": "19. Fridovich-Keil, S., Meanti, G., Warburg, F.R., Recht, B., Kanazawa, A.: K-planes: Explicit radiance fields in space, time, and appearance. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 12479-12488 (2023)" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 132, + 402, + 481, + 435 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 402, + 481, + 435 + ], + "spans": [ + { + "bbox": [ + 132, + 402, + 481, + 435 + ], + "type": "text", + "content": "20. Gao, H., Li, R., Tulsiani, S., Russell, B., Kanazawa, A.: Monocular dynamic view synthesis: A reality check. Advances in Neural Information Processing Systems 35, 33768-33780 (2022)" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 132, + 435, + 481, + 468 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 435, + 481, + 468 + ], + "spans": [ + { + "bbox": [ + 132, + 435, + 481, + 468 + ], + "type": "text", + "content": "21. Geng, D., Owens, A.: Motion guidance: Diffusion-based image editing with differentiable motion estimators. In: The Twelfth International Conference on Learning Representations (2023)" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 132, + 468, + 481, + 501 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 468, + 481, + 501 + ], + "spans": [ + { + "bbox": [ + 132, + 468, + 481, + 501 + ], + "type": "text", + "content": "22. Girdhar, R., Singh, M., Brown, A., Duval, Q., Azadi, S., Rambhatla, S.S., Shah, A., Yin, X., Parikh, D., Misra, I.: Emu video: Factorizing text-to-video generation by explicit image conditioning. arXiv preprint arXiv:2311.10709 (2023)" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 132, + 501, + 481, + 534 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 501, + 481, + 534 + ], + "spans": [ + { + "bbox": [ + 132, + 501, + 481, + 534 + ], + "type": "text", + "content": "23. Guo, X., Sun, J., Dai, Y., Chen, G., Ye, X., Tan, X., Ding, E., Zhang, Y., Wang, J.: Forward flow for novel view synthesis of dynamic scenes. In: Proceedings of the IEEE/CVF International Conference on Computer Vision. pp. 16022-16033 (2023)" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 132, + 534, + 481, + 566 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 534, + 481, + 566 + ], + "spans": [ + { + "bbox": [ + 132, + 534, + 481, + 566 + ], + "type": "text", + "content": "24. Gupta, A., Yu, L., Sohn, K., Gu, X., Hahn, M., Fei-Fei, L., Essa, I., Jiang, L., Lezama, J.: Photorealistic video generation with diffusion models. arXiv preprint arXiv:2312.06662 (2023)" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 132, + 567, + 481, + 600 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 567, + 481, + 600 + ], + "spans": [ + { + "bbox": [ + 132, + 567, + 481, + 600 + ], + "type": "text", + "content": "25. Ho, J., Chan, W., Sahara, C., Whang, J., Gao, R., Gritsenko, A., Kingma, D.P., Poole, B., Norouzi, M., Fleet, D.J., et al.: Imagen video: High definition video generation with diffusion models. arXiv preprint arXiv:2210.02303 (2022)" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 132, + 600, + 481, + 632 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 600, + 481, + 632 + ], + "spans": [ + { + "bbox": [ + 132, + 600, + 481, + 632 + ], + "type": "text", + "content": "26. Hong, W., Ding, M., Zheng, W., Liu, X., Tang, J.: Cogvideo: Large-scale pretraining for text-to-video generation via transformers. arXiv preprint arXiv:2205.15868 (2022)" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 132, + 632, + 481, + 666 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 632, + 481, + 666 + ], + "spans": [ + { + "bbox": [ + 132, + 632, + 481, + 666 + ], + "type": "text", + "content": "27. Hu, Y., Fang, Y., Ge, Z., Qu, Z., Zhu, Y., Pradhana, A., Jiang, C.: A moving least squares material point method with displacement discontinuity and two-way rigid body coupling. ACM Transactions on Graphics (TOG) 37(4), 1-14 (2018)" + } + ] + } + ], + "index": 18 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "spans": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "type": "text", + "content": "16" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 166, + 91, + 216, + 102 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 166, + 91, + 216, + 102 + ], + "spans": [ + { + "bbox": [ + 166, + 91, + 216, + 102 + ], + "type": "text", + "content": "Zhang et al." + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 15 + }, + { + "para_blocks": [ + { + "bbox": [ + 130, + 116, + 482, + 666 + ], + "type": "list", + "angle": 0, + "index": 20, + "blocks": [ + { + "bbox": [ + 130, + 116, + 482, + 149 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 116, + 482, + 149 + ], + "spans": [ + { + "bbox": [ + 130, + 116, + 482, + 149 + ], + "type": "text", + "content": "28. Hu, Y., Li, T.M., Anderson, L., Ragan-Kelley, J., Durand, F.: Taichi: a language for high-performance computation on spatially sparse data structures. ACM Transactions on Graphics (TOG) 38(6), 1-16 (2019)" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 130, + 150, + 481, + 183 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 150, + 481, + 183 + ], + "spans": [ + { + "bbox": [ + 130, + 150, + 481, + 183 + ], + "type": "text", + "content": "29. Hu, Y., Li, T.M., Anderson, L., Ragan-Kelley, J., Durand, F.: Taichi: a language for high-performance computation on spatially sparse data structures. ACM Transactions on Graphics (TOG) 38(6), 1-16 (2019)" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 132, + 184, + 481, + 217 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 184, + 481, + 217 + ], + "spans": [ + { + "bbox": [ + 132, + 184, + 481, + 217 + ], + "type": "text", + "content": "30. Huang, Y.H., Sun, Y.T., Yang, Z., Lyu, X., Cao, Y.P., Qi, X.: Sc-gs: Sparse-controlled gaussian splatting for editable dynamic scenes. arXiv preprint arXiv:2312.14937 (2023)" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 132, + 218, + 481, + 239 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 218, + 481, + 239 + ], + "spans": [ + { + "bbox": [ + 132, + 218, + 481, + 239 + ], + "type": "text", + "content": "31. Jiang, C., Gast, T., Teran, J.: Anisotropic elastoplasticity for cloth, knit and hair frictional contact. ACM Transactions on Graphics (TOG) 36(4), 1-14 (2017)" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 132, + 240, + 481, + 262 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 240, + 481, + 262 + ], + "spans": [ + { + "bbox": [ + 132, + 240, + 481, + 262 + ], + "type": "text", + "content": "32. Jiang, C., Schroeder, C., Selle, A., Teran, J., Stomakhin, A.: The affine particle-in-cell method. ACM Transactions on Graphics (TOG) 34(4), 1-10 (2015)" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 132, + 263, + 481, + 296 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 263, + 481, + 296 + ], + "spans": [ + { + "bbox": [ + 132, + 263, + 481, + 296 + ], + "type": "text", + "content": "33. Jiang, C., Schroeder, C., Teran, J., Stomakhin, A., Selle, A.: The material point method for simulating continuum materials. In: ACM SIGGRAPH 2016 courses. pp. 1-52 (2016)" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 132, + 297, + 481, + 319 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 297, + 481, + 319 + ], + "spans": [ + { + "bbox": [ + 132, + 297, + 481, + 319 + ], + "type": "text", + "content": "34. Kerbl, B., Kopanas, G., Leimkuhler, T., Drettakis, G.: 3d gaussian splatting for real-time radiance field rendering. ACM Transactions on Graphics 42(4) (2023)" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 132, + 319, + 481, + 352 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 319, + 481, + 352 + ], + "spans": [ + { + "bbox": [ + 132, + 319, + 481, + 352 + ], + "type": "text", + "content": "35. Klár, G., Gast, T., Pradhana, A., Fu, C., Schroeder, C., Jiang, C., Teran, J.: Drucker-prager elastoplasticity for sand animation. ACM Transactions on Graphics (TOG) 35(4), 1-12 (2016)" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 132, + 353, + 481, + 374 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 353, + 481, + 374 + ], + "spans": [ + { + "bbox": [ + 132, + 353, + 481, + 374 + ], + "type": "text", + "content": "36. Kobayashi, M., Motoyoshi, I.: Perceiving natural speed in natural movies. i-Perception 10(4), 2041669519860544 (2019)" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 132, + 375, + 481, + 407 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 375, + 481, + 407 + ], + "spans": [ + { + "bbox": [ + 132, + 375, + 481, + 407 + ], + "type": "text", + "content": "37. Kondratyuk, D., Yu, L., Gu, X., Lezama, J., Huang, J., Hornung, R., Adam, H., Akbari, H., Alon, Y., Birodkar, V., et al.: Videopoet: A large language model for zero-shot video generation. arXiv preprint arXiv:2312.14125 (2023)" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 132, + 409, + 481, + 441 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 409, + 481, + 441 + ], + "spans": [ + { + "bbox": [ + 132, + 409, + 481, + 441 + ], + "type": "text", + "content": "38. Kopf, J., Rong, X., Huang, J.B.: Robust consistent video depth estimation. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 1611-1621 (2021)" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 132, + 442, + 481, + 475 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 442, + 481, + 475 + ], + "spans": [ + { + "bbox": [ + 132, + 442, + 481, + 475 + ], + "type": "text", + "content": "39. Kratimenos, A., Lei, J., Daniilidis, K.: Dynmf: Neural motion factorization for real-time dynamic view synthesis with 3d gaussian splatting. arXiv preprint arXiv:2312.00112 (2023)" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 132, + 475, + 481, + 508 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 475, + 481, + 508 + ], + "spans": [ + { + "bbox": [ + 132, + 475, + 481, + 508 + ], + "type": "text", + "content": "40. Le Cleac'h, S., Yu, H.X., Guo, M., Howell, T., Gao, R., Wu, J., Manchester, Z., Schwager, M.: Differentiable physics simulation of dynamics-augmented neural objects. IEEE Robotics and Automation Letters (2023)" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 132, + 509, + 481, + 542 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 509, + 481, + 542 + ], + "spans": [ + { + "bbox": [ + 132, + 509, + 481, + 542 + ], + "type": "text", + "content": "41. Li, H., Sumner, R.W., Pauly, M.: Global correspondence optimization for non-rigid registration of depth scans. In: Computer graphics forum. vol. 27, pp. 1421-1430. Wiley Online Library (2008)" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 132, + 543, + 481, + 576 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 543, + 481, + 576 + ], + "spans": [ + { + "bbox": [ + 132, + 543, + 481, + 576 + ], + "type": "text", + "content": "42. Li, X., Qiao, Y.L., Chen, P.Y., Jatavallabhula, K.M., Lin, M., Jiang, C., Gan, C.: Pac-nerf: Physics augmented continuum neural radiance fields for geometry-agnostic system identification. arXiv preprint arXiv:2303.05512 (2023)" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 132, + 577, + 481, + 609 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 577, + 481, + 609 + ], + "spans": [ + { + "bbox": [ + 132, + 577, + 481, + 609 + ], + "type": "text", + "content": "43. Li, Z., Niklaus, S., Snavely, N., Wang, O.: Neural scene flow fields for space-time view synthesis of dynamic scenes. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 6498-6508 (2021)" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 132, + 610, + 481, + 632 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 610, + 481, + 632 + ], + "spans": [ + { + "bbox": [ + 132, + 610, + 481, + 632 + ], + "type": "text", + "content": "44. Li, Z., Tucker, R., Snavely, N., Holynski, A.: Generative image dynamics. arXiv preprint arXiv:2309.07906 (2023)" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 132, + 632, + 481, + 666 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 632, + 481, + 666 + ], + "spans": [ + { + "bbox": [ + 132, + 632, + 481, + 666 + ], + "type": "text", + "content": "45. Li, Z., Wang, Q., Cole, F., Tucker, R., Snavely, N.: Dynibar: Neural dynamic image-based rendering. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 4273-4284 (2023)" + } + ] + } + ], + "index": 19 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 391, + 91, + 447, + 101 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 391, + 91, + 447, + 101 + ], + "spans": [ + { + "bbox": [ + 391, + 91, + 447, + 101 + ], + "type": "text", + "content": "PhysDreamer" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 470, + 91, + 481, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 470, + 91, + 481, + 100 + ], + "spans": [ + { + "bbox": [ + 470, + 91, + 481, + 100 + ], + "type": "text", + "content": "17" + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 16 + }, + { + "para_blocks": [ + { + "bbox": [ + 132, + 116, + 480, + 666 + ], + "type": "list", + "angle": 0, + "index": 19, + "blocks": [ + { + "bbox": [ + 133, + 116, + 480, + 149 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 116, + 480, + 149 + ], + "spans": [ + { + "bbox": [ + 133, + 116, + 480, + 149 + ], + "type": "text", + "content": "46. Ling, H., Kim, S.W., Torralba, A., Fidler, S., Kreis, K.: Align your gaussians: Text-to-4d with dynamic 3d gaussians and composed diffusion models. arXiv preprint arXiv:2312.13763 (2023)" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 132, + 150, + 480, + 172 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 150, + 480, + 172 + ], + "spans": [ + { + "bbox": [ + 132, + 150, + 480, + 172 + ], + "type": "text", + "content": "47. Luiten, J., Kopanas, G., Leibe, B., Ramanan, D.: Dynamic 3d gaussians: Tracking by persistent dynamic view synthesis. arXiv preprint arXiv:2308.09713 (2023)" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 132, + 172, + 480, + 205 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 172, + 480, + 205 + ], + "spans": [ + { + "bbox": [ + 132, + 172, + 480, + 205 + ], + "type": "text", + "content": "48. Ma, P., Chen, P.Y., Deng, B., Tenenbaum, J.B., Du, T., Gan, C., Matusik, W.: Learning neural constitutive laws from motion observations for generalizable pde dynamics. In: International Conference on Machine Learning. PMLR (2023)" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 132, + 205, + 480, + 237 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 205, + 480, + 237 + ], + "spans": [ + { + "bbox": [ + 132, + 205, + 480, + 237 + ], + "type": "text", + "content": "49. Macklin, M.: Warp: A high-performance python framework forgpu simulation and graphics. https://github.com/nvidia/warp (March 2022), nVIDIA GPU Technology Conference (GTC)" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 132, + 237, + 480, + 270 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 237, + 480, + 270 + ], + "spans": [ + { + "bbox": [ + 132, + 237, + 480, + 270 + ], + "type": "text", + "content": "50. Mildenhall, B., Srinivasan, P.P., Tancik, M., Barron, J.T., Ramamoorthi, R., Ng, R.: Nerf: Representing scenes as neural radiance fields for view synthesis. In: ECCV (2020)" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 132, + 270, + 480, + 303 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 270, + 480, + 303 + ], + "spans": [ + { + "bbox": [ + 132, + 270, + 480, + 303 + ], + "type": "text", + "content": "51. Newcombe, R.A., Fox, D., Seitz, S.M.: Dynamicfusion: Reconstruction and tracking of non-rigid scenes in real-time. In: Proceedings of the IEEE conference on computer vision and pattern recognition. pp. 343-352 (2015)" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 132, + 303, + 480, + 336 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 303, + 480, + 336 + ], + "spans": [ + { + "bbox": [ + 132, + 303, + 480, + 336 + ], + "type": "text", + "content": "52. Park, K., Sinha, U., Barron, J.T., Bouaziz, S., Goldman, D.B., Seitz, S.M., Martin-Brualla, R.: Nerfies: Deformable neural radiance fields. In: Proceedings of the IEEE/CVF International Conference on Computer Vision. pp. 5865-5874 (2021)" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 132, + 336, + 480, + 369 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 336, + 480, + 369 + ], + "spans": [ + { + "bbox": [ + 132, + 336, + 480, + 369 + ], + "type": "text", + "content": "53. Park, K., Sinha, U., Hedman, P., Barron, J.T., Bouaziz, S., Goldman, D.B., Martin-Brualla, R., Seitz, S.M.: Hypernerf: A higher-dimensional representation for topologically varying neural radiance fields. arXiv preprint arXiv:2106.13228 (2021)" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 132, + 369, + 480, + 413 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 369, + 480, + 413 + ], + "spans": [ + { + "bbox": [ + 132, + 369, + 480, + 413 + ], + "type": "text", + "content": "54. Peng, S., Niemeyer, M., Mescheder, L., Pollefeys, M., Geiger, A.: Convolutional occupancy networks. In: Computer Vision-ECCV 2020: 16th European Conference, Glasgow, UK, August 23-28, 2020, Proceedings, Part III 16. pp. 523-540. Springer (2020)" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 132, + 413, + 480, + 445 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 413, + 480, + 445 + ], + "spans": [ + { + "bbox": [ + 132, + 413, + 480, + 445 + ], + "type": "text", + "content": "55. Poole, B., Jain, A., Barron, J.T., Mildenhall, B.: Dreamfusion: Text-to-3d using 2d diffusion. In: The Eleventh International Conference on Learning Representations (2022)" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 132, + 446, + 480, + 479 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 446, + 480, + 479 + ], + "spans": [ + { + "bbox": [ + 132, + 446, + 480, + 479 + ], + "type": "text", + "content": "56. Pumarola, A., Corona, E., Pons-Moll, G., Moreno-Noguer, F.: D-nerf: Neural radiance fields for dynamic scenes. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 10318-10327 (2021)" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 132, + 479, + 480, + 522 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 479, + 480, + 522 + ], + "spans": [ + { + "bbox": [ + 132, + 479, + 480, + 522 + ], + "type": "text", + "content": "57. Ram, D., Gast, T., Jiang, C., Schroeder, C., Stomakhin, A., Teran, J., Kavehpour, P.: A material point method for viscoelastic fluids, foams and sponges. In: Proceedings of the 14th ACM SIGGRAPH/Eurographics Symposium on Computer Animation. pp. 157-163 (2015)" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 132, + 522, + 480, + 544 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 522, + 480, + 544 + ], + "spans": [ + { + "bbox": [ + 132, + 522, + 480, + 544 + ], + "type": "text", + "content": "58. Ren, J., Pan, L., Tang, J., Zhang, C., Cao, A., Zeng, G., Liu, Z.: Dreamgaussian4d: Generative 4d gaussian splatting. arXiv preprint arXiv:2312.17142 (2023)" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 132, + 544, + 480, + 578 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 544, + 480, + 578 + ], + "spans": [ + { + "bbox": [ + 132, + 544, + 480, + 578 + ], + "type": "text", + "content": "59. Singer, U., Polyak, A., Hayes, T., Yin, X., An, J., Zhang, S., Hu, Q., Yang, H., Ashual, O., Gafni, O., et al.: Make-a-video: Text-to-video generation without text-video data. arXiv preprint arXiv:2209.14792 (2022)" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 132, + 578, + 480, + 610 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 578, + 480, + 610 + ], + "spans": [ + { + "bbox": [ + 132, + 578, + 480, + 610 + ], + "type": "text", + "content": "60. Singer, U., Sheynin, S., Polyak, A., Ashual, O., Makarov, I., Kokkinos, F., Goyal, N., Vedaldi, A., Parikh, D., Johnson, J., et al.: Text-to-4d dynamic scene generation. arXiv preprint arXiv:2301.11280 (2023)" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 132, + 610, + 480, + 632 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 610, + 480, + 632 + ], + "spans": [ + { + "bbox": [ + 132, + 610, + 480, + 632 + ], + "type": "text", + "content": "61. Stocker, A.A., Simoncelli, E.P.: Noise characteristics and prior expectations in human visual speed perception. Nature neuroscience 9(4), 578-585 (2006)" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 132, + 632, + 480, + 666 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 632, + 480, + 666 + ], + "spans": [ + { + "bbox": [ + 132, + 632, + 480, + 666 + ], + "type": "text", + "content": "62. Stomakhin, A., Schroeder, C., Chai, L., Teran, J., Selle, A.: A material point method for snow simulation. ACM Transactions on Graphics (TOG) 32(4), 1-10 (2013)" + } + ] + } + ], + "index": 18 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "spans": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "type": "text", + "content": "18" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 166, + 91, + 216, + 102 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 166, + 91, + 216, + 102 + ], + "spans": [ + { + "bbox": [ + 166, + 91, + 216, + 102 + ], + "type": "text", + "content": "Zhang et al." + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 17 + }, + { + "para_blocks": [ + { + "bbox": [ + 130, + 116, + 482, + 578 + ], + "type": "list", + "angle": 0, + "index": 17, + "blocks": [ + { + "bbox": [ + 130, + 116, + 482, + 138 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 116, + 482, + 138 + ], + "spans": [ + { + "bbox": [ + 130, + 116, + 482, + 138 + ], + "type": "text", + "content": "63. Tang, J., Ren, J., Zhou, H., Liu, Z., Zeng, G.: Dreamgaussian: Generative gaussian splatting for efficient 3d content creation. arXiv preprint arXiv:2309.16653 (2023)" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 130, + 140, + 482, + 171 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 140, + 482, + 171 + ], + "spans": [ + { + "bbox": [ + 130, + 140, + 482, + 171 + ], + "type": "text", + "content": "64. Unterthiner, T., Van Steenkiste, S., Kurach, K., Marinier, R., Michalski, M., Gelly, S.: Towards accurate generative models of video: A new metric & challenges. arXiv preprint arXiv:1812.01717 (2018)" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 132, + 172, + 481, + 215 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 172, + 481, + 215 + ], + "spans": [ + { + "bbox": [ + 132, + 172, + 481, + 215 + ], + "type": "text", + "content": "65. Villegas, R., Babaeizadeh, M., Kindermans, P.J., Moraldo, H., Zhang, H., Saffar, M.T., Castro, S., Kunze, J., Erhan, D.: Phenaki: Variable length video generation from open domain textual descriptions. In: International Conference on Learning Representations (2022)" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 132, + 216, + 481, + 248 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 216, + 481, + 248 + ], + "spans": [ + { + "bbox": [ + 132, + 216, + 481, + 248 + ], + "type": "text", + "content": "66. Wang, C., MacDonald, L.E., Jeni, L.A., Lucey, S.: Flow supervision for deformable nerf. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 21128-21137 (2023)" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 132, + 249, + 481, + 281 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 249, + 481, + 281 + ], + "spans": [ + { + "bbox": [ + 132, + 249, + 481, + 281 + ], + "type": "text", + "content": "67. Wang, C., Zhuang, P., Siarohin, A., Cao, J., Qian, G., Lee, H.Y., Tulyakov, S.: Diffusion priors for dynamic view synthesis from monocular videos. arXiv preprint arXiv:2401.05583 (2024)" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 132, + 282, + 481, + 314 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 282, + 481, + 314 + ], + "spans": [ + { + "bbox": [ + 132, + 282, + 481, + 314 + ], + "type": "text", + "content": "68. Wu, C., Liang, J., Ji, L., Yang, F., Fang, Y., Jiang, D., Duan, N.: Niwa: Visual synthesis pre-training for neural visual world creation. In: European conference on computer vision. pp. 720-736. Springer (2022)" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 132, + 315, + 481, + 346 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 315, + 481, + 346 + ], + "spans": [ + { + "bbox": [ + 132, + 315, + 481, + 346 + ], + "type": "text", + "content": "69. Xian, W., Huang, J.B., Kopf, J., Kim, C.: Space-time neural irradiance fields for free-viewpoint video. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 9421-9431 (2021)" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 132, + 347, + 481, + 380 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 347, + 481, + 380 + ], + "spans": [ + { + "bbox": [ + 132, + 347, + 481, + 380 + ], + "type": "text", + "content": "70. Xie, T., Zong, Z., Qiu, Y., Li, X., Feng, Y., Yang, Y., Jiang, C.: Physgaus-sian: Physics-integrated 3d gaussians for generative dynamics. arXiv preprint arXiv:2311.12198 (2023)" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 132, + 380, + 481, + 413 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 380, + 481, + 413 + ], + "spans": [ + { + "bbox": [ + 132, + 380, + 481, + 413 + ], + "type": "text", + "content": "71. Xie, Y., Takikawa, T., Saito, S., Litany, O., Yan, S., Khan, N., Tombari, F., Tompkin, J., Sitzmann, V., Sridhar, S.: Neural fields in visual computing and beyond. In: Computer Graphics Forum. vol. 41, pp. 641-676. Wiley Online Library (2022)" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 132, + 414, + 481, + 445 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 414, + 481, + 445 + ], + "spans": [ + { + "bbox": [ + 132, + 414, + 481, + 445 + ], + "type": "text", + "content": "72. Yang, Z., Gao, X., Zhou, W., Jiao, S., Zhang, Y., Jin, X.: Deformable 3d gaussians for high-fidelity monocular dynamic scene reconstruction. arXiv preprint arXiv:2309.13101 (2023)" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 132, + 447, + 481, + 468 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 447, + 481, + 468 + ], + "spans": [ + { + "bbox": [ + 132, + 447, + 481, + 468 + ], + "type": "text", + "content": "73. Yin, Y., Xu, D., Wang, Z., Zhao, Y., Wei, Y.: 4dgen: Grounded 4d content generation with spatial-temporal consistency. arXiv preprint arXiv:2312.17225 (2023)" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 132, + 468, + 481, + 490 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 468, + 481, + 490 + ], + "spans": [ + { + "bbox": [ + 132, + 468, + 481, + 490 + ], + "type": "text", + "content": "74. Yu, H., Julin, J., Milacski, Z.A., Niinuma, K., Jeni, L.A.: Cogs: Controllable gaussian splatting. arXiv preprint arXiv:2312.05664 (2023)" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 132, + 491, + 481, + 522 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 491, + 481, + 522 + ], + "spans": [ + { + "bbox": [ + 132, + 491, + 481, + 522 + ], + "type": "text", + "content": "75. Yu, H., Julin, J., Milacski, Z.A., Niinuma, K., Jeni, L.A.: Dylin: Making light field networks dynamic. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 12397-12406 (2023)" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 132, + 523, + 481, + 555 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 523, + 481, + 555 + ], + "spans": [ + { + "bbox": [ + 132, + 523, + 481, + 555 + ], + "type": "text", + "content": "76. Zhang, S., Wang, J., Zhang, Y., Zhao, K., Yuan, H., Qin, Z., Wang, X., Zhao, D., Zhou, J.: I2vgen-xl: High-quality image-to-video synthesis via cascaded diffusion models. arXiv preprint arXiv:2311.04145 (2023)" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 132, + 556, + 481, + 578 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 556, + 481, + 578 + ], + "spans": [ + { + "bbox": [ + 132, + 556, + 481, + 578 + ], + "type": "text", + "content": "77. Zhang, Z., Cole, F., Tucker, R., Freeman, W.T., Dekel, T.: Consistent depth of moving objects in video. ACM Transactions on Graphics (TOG) 40(4), 1-12 (2021)" + } + ] + } + ], + "index": 16 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 391, + 91, + 447, + 101 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 391, + 91, + 447, + 101 + ], + "spans": [ + { + "bbox": [ + 391, + 91, + 447, + 101 + ], + "type": "text", + "content": "PhysDreamer" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 470, + 91, + 481, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 470, + 91, + 481, + 100 + ], + "spans": [ + { + "bbox": [ + 470, + 91, + 481, + 100 + ], + "type": "text", + "content": "19" + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 18 + } + ], + "_backend": "vlm", + "_version_name": "2.6.4" +} \ No newline at end of file diff --git a/2024/Physics-Free Spectrally Multiplexed Photometric Stereo under Unknown Spectral Composition/09ecf602-4895-4520-8961-a33ebbe3e42d_content_list.json b/2024/Physics-Free Spectrally Multiplexed Photometric Stereo under Unknown Spectral Composition/09ecf602-4895-4520-8961-a33ebbe3e42d_content_list.json new file mode 100644 index 0000000000000000000000000000000000000000..1270bc7918a05a221ebd8af4a2593c3709ccf6f9 --- /dev/null +++ b/2024/Physics-Free Spectrally Multiplexed Photometric Stereo under Unknown Spectral Composition/09ecf602-4895-4520-8961-a33ebbe3e42d_content_list.json @@ -0,0 +1,1539 @@ +[ + { + "type": "text", + "text": "SpectraM-PS: Spectrally Multiplexed Photometric Stereo under Unknown Spectral Composition", + "text_level": 1, + "bbox": [ + 243, + 140, + 759, + 209 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Satoshi Ikehata $^{1,2}$ and Yuta Asano $^{1}$", + "bbox": [ + 357, + 234, + 643, + 250 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "$^{1}$ National Institute of Informatics, Tokyo, Japan", + "bbox": [ + 334, + 261, + 666, + 276 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "2 Tokyo Institute of Technology, Tokyo, Japan", + "bbox": [ + 343, + 277, + 656, + 290 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Abstract. In this paper, we present a groundbreaking spectrally multiplexed photometric stereo approach for recovering surface normals of dynamic surfaces without the need for calibrated lighting or sensors, a notable advancement in the field traditionally hindered by stringent prerequisites and spectral ambiguity. By embracing spectral ambiguity as an advantage, our technique enables the generation of training data without specialized multispectral rendering frameworks. We introduce a unique, physics-free network architecture, SpectraM-PS, that effectively processes multiplexed images to determine surface normals across a wide range of conditions and material types, without relying on specific physically-based knowledge. Additionally, we establish the first benchmark dataset, SpectraM14, for spectrally multiplexed photometric stereo, facilitating comprehensive evaluations against existing calibrated methods. Our contributions significantly enhance the capabilities for dynamic surface recovery, particularly in uncalibrated setups, marking a pivotal step forward in the application of photometric stereo across various domains.", + "bbox": [ + 259, + 329, + 743, + 565 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Keywords: Spectrally Multiplexed Photometric Stereo $\\cdot$ Dynamic Surface Recovery $\\cdot$ Multispectral Photometric Stereo", + "bbox": [ + 259, + 578, + 740, + 606 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "1 Introduction", + "text_level": 1, + "bbox": [ + 215, + 656, + 375, + 672 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Recovering detailed normals of dynamic surfaces is essential for monitoring various processes: in manufacturing, it helps in tracking wear and tear of machine parts; in agriculture, it allows for the observation of crop growth through changes in leaf geometry; and in sports engineering, it aids in improving equipment design and safety by analyzing how surfaces deform upon impact.", + "bbox": [ + 212, + 688, + 787, + 763 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Photometric Stereo (PS) [57,64] derives object surface normals from observations under different lighting conditions at a fixed viewpoint. Despite decades of progress, the requirement for objects to stay stationary during lighting changes challenges the recovery of dynamic surfaces, essential for analyzing temporal surface deformations. PS researches have employed spectral multiplexing for dynamic", + "bbox": [ + 212, + 763, + 787, + 839 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/7abd35d53f8c0a67cc2785e152a28a9a22889d76b27d1c86b437f58a6f97ea06.jpg", + "image_caption": [ + "Fig. 1: (Left) Illustration of our SpectraM-PS. Our method recovers a surface normal map from a spectrally multiplexed image. The spectral/spatial composition for generating the observations is unknown. There is potential for a mismatch between the sensor's spectral sensitivity and the light source's spectral distribution, which may lead to crosstalk. (Right) By applying our method to individual frames of a video, the normal map of dynamic surfaces can be recovered." + ], + "image_footnote": [], + "bbox": [ + 222, + 143, + 464, + 292 + ], + "page_idx": 1 + }, + { + "type": "image", + "img_path": "images/ecdb0a810f486ab59296eff4aed054c1ef32cab1271295e077be17453885b81a.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 465, + 143, + 781, + 292 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "surface recovery [10,20,32,37,51,62]—a technique originally used in telecommunications and spectroscopy [31]. This technique utilizes the varying wavelengths of light to multiplex and subsequently demultiplex signals within a single sensor, thereby increasing the capacity for information transmission.", + "bbox": [ + 212, + 400, + 784, + 460 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Historically, spectrally multiplexed photometric stereo is often referred to as color photometric stereo [7,14,20,32,33], specifically when objects are illuminated with monochromatic red, green, and blue lights from various angles, captured in the camera's RGB channels. Each channel is then treated as an observation under a distinct lighting for PS analysis. This technique has been further extended to not only RGB but also any number of spectral bands and is specifically referred to as multispectral photometric stereo [17,18,46]. These techniques enable dynamic surface recovery by processing each temporal multi-channel frame separately.", + "bbox": [ + 212, + 460, + 787, + 582 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Despite their potential in dynamic surface recovery, current spectrally multiplexed photometric stereo methods face stringent prerequisites that limit their practicality. These include the necessity for precisely calibrated directional lighting in controlled environments [7, 14, 17, 20, 46] and sensors with aligned spectral sensitivities [32, 33]. Furthermore, they make strong assumptions about the surface, requiring it to be convex, integrable, Lambertian, and exhibit uniform chromaticity [20, 37]. By contrast, recent PS methods without spectral multiplexing support non-Lambertian surfaces [29, 55], spatially-varying materials [12, 23], and the use of uncalibrated lighting [11, 25, 26, 54]. This disparity arises from the challenge where identical observations are produced by different spectral compositions of light, surface and sensor [22, 50], a phenomenon absent in conventional PS due to constant spectral compositions of them across images. Recently, Guo et al. [17, 18] thoroughly explored how the spectral ambiguity renders spectrally multiplexed photometric stereo ill-posed, necessitating severely unpractical conditions on light, surface, and sensor to resolve the ambiguity.", + "bbox": [ + 212, + 583, + 787, + 809 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "In this work, we propose a spectrally multiplexed photometric stereo method that recovers normals directly from multiplexed observations produced by $un$", + "bbox": [ + 214, + 809, + 787, + 840 + ], + "page_idx": 1 + }, + { + "type": "page_number", + "text": "2", + "bbox": [ + 217, + 114, + 228, + 126 + ], + "page_idx": 1 + }, + { + "type": "header", + "text": "S. Ikehata and Y. Asano", + "bbox": [ + 271, + 114, + 436, + 127 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "known composition of lights, surface, and sensor (See Fig. 1-left), drawing inspiration from recent data-driven photometric stereo methods [25, 26]. While prior works [18, 46] have considered the spectral ambiguity harmful and something that must be resolved, we demonstrate that it can even be beneficial for a data-driven approach as it compacts the input space and allows for the generation of training data without a multispectral rendering framework. Trained on spectrally composed observations, our generic, physics-free architecture directly maps a single multiplexed image with an order-agnostic, arbitrary number of channels to object surface normals without the need for calibrating lights and sensors, and without imposing severe constraints on surface reflectance and geometry. By applying our method to individual frames of a video, dynamic surface recovery via spectrally multiplexed photometric stereo in uncalibrated, uncontrolled scenarios is achieved as illustrated in Fig. 1 (right).", + "bbox": [ + 212, + 145, + 787, + 343 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "While numerous benchmarks exist for conventional photometric stereo [52,56, 63], not a single benchmark is available for spectrally multiplexed PS. Therefore, we have created the first real benchmark dataset, namely SpectraM, for this task. For comparative evaluations with calibrated methods such as [18,46], we carefully calibrated directional light sources of different wavelengths, including their directions. We implemented five different difficulty settings by varying the type of light sources (RGB vs NIR) and whether individual light sources were illuminated independently or simultaneously, catering to both ideal conditions without channel crosstalk and more realistic conditions with channel crosstalk.", + "bbox": [ + 212, + 343, + 787, + 478 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Our contributions are summarized as follows: (1) We pioneer the use of spectrally multiplexed photometric stereo for recovering dynamic surfaces in uncalibrated setups, employing a data-driven approach to overcome spectral ambiguity, a significant barrier in prior work. (2) We introduce a unique, physics-free neural network, SpectraM-PS (Spectrally Multiplexed PS), that recovers surface normals from a spectrally multiplexed image, capable of handling images with any number of order-agnostic channels. (3) We demonstrate how spectral ambiguity restricts the input space for training data generation, offering a strategy for efficient dataset creation without the need for multispectral rendering. (4) We create the first evaluation benchmark, SpectraM, for this domain, showing our method's superiority over current calibrated spectrally multiplexed photometric stereo techniques.", + "bbox": [ + 212, + 479, + 787, + 660 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "2 Related Work", + "text_level": 1, + "bbox": [ + 214, + 685, + 388, + 702 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Temporally Multiplexed Photometric Stereo (Conventional). From a communication perspective, conventional photometric stereo, as originally proposed by Woodham [64], employs a time multiplexing strategy to recover static surfaces. This method involves temporally varying lighting conditions while capturing images from a fixed viewpoint. Since the same light sources and sensor always provide observations, image differences stem solely from changes in light direction and intensity. This approach simplifies addressing complex conditions such as cast shadows [27, 28], non-Lambertian surfaces [16, 29], non-convex sur", + "bbox": [ + 212, + 718, + 787, + 840 + ], + "page_idx": 2 + }, + { + "type": "header", + "text": "SpectraM-PS", + "bbox": [ + 640, + 114, + 730, + 128 + ], + "page_idx": 2 + }, + { + "type": "page_number", + "text": "3", + "bbox": [ + 774, + 116, + 785, + 126 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "faces [23], and uncalibrated lighting [11, 36, 54]. Recently, learning-based methods [11-13, 23-25, 34, 36, 39-42, 44, 53, 58, 59, 68, 69] have emerged as an effective alternative, addressing challenges faced by traditional, physics-based approaches [6, 8, 16, 19, 21, 28, 45, 65]. These data-driven methods regress normal maps from observations utilizing techniques such as observation map regression [23, 30], set pooling [12, 13], graph neural networks [68], Transformer [24], and neural rendering for inverse rendering optimization [40, 41, 58]. Notably, the introduction of universal photometric stereo methods [25, 26] has enabled the handling of unknown, spatially-varying lighting in a purely data-driven framework. Inspired by these advancements, our work aims to regress normals from observations under unknown light, surface, and sensor conditions.", + "bbox": [ + 212, + 146, + 787, + 313 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Spectrally Multiplexed Photometric Stereo. Despite its potential for dynamic surface recovery, spectrally multiplexed photometric stereo [15, 37] has remained less explored than its mainstream counterpart, primarily due to notable limitations.", + "bbox": [ + 212, + 315, + 787, + 375 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Lighting Constraints: Existing methods necessitate multiple directional lights in controlled settings, contrasting the flexibility of temporally multiplexed techniques that adapt to diverse lighting conditions [26, 47, 49]. They typically require pre-calibrated light directions and distinct light source spectra to prevent channel crosstalk. In contrast, our approach accommodates uncontrolled lighting scenarios without the need for predefined or calibrated setups.", + "bbox": [ + 212, + 376, + 787, + 467 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Surface and Sensor Constraints: Prior works assume significant limitations on surfaces, such as Lambertian, convex, and uniform properties [15,20,37]. Recent advances like Lv et al. [46] extend to non-Lambertian surfaces but still require uniform materials. Sensor requirements typically involve narrow-band spectral responses and a fixed number of channels, limiting flexibility. Our approach leverages a data-driven model, training neural networks on synthetic data to handle complex surfaces and varied spectral sensor responses.", + "bbox": [ + 212, + 467, + 787, + 571 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Data-driven Methods: To our knowledge, there are few data-driven methods for this task [32, 33, 46]. Previous studies, such as those by Ju et al. [32, 33], require identical spectral and spatial lighting conditions during both training and testing, greatly restricting their practicality. ELIE-Net [46] permits variability in training and test setups; however, strong assumptions on both light and surfaces prohibitively limit its applications. Our model, on the other hand, eschews explicit lighting models in favor of learning direct input-output relationships, allowing for accurate predictions under varied and unknown spectral compositions and supporting materials with spatially diverse properties. Furthermore, unlike ELIE-Net's reliance on spectral BRDF datasets, our training approach utilizes assets akin to those employed in conventional photometric stereo.", + "bbox": [ + 212, + 573, + 787, + 739 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "3 Problem Statement", + "text_level": 1, + "bbox": [ + 214, + 762, + 441, + 777 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Given a single image $I \\in \\mathbb{R}^{h \\times w \\times k}$ captured by a static $k$ -channel orthographic sensor, along with an optional object mask $M \\in \\mathbb{R}^{h \\times w}$ , the objective of spectrally multiplexed photometric stereo is to recover the surface normals of the", + "bbox": [ + 212, + 792, + 787, + 839 + ], + "page_idx": 3 + }, + { + "type": "page_number", + "text": "4", + "bbox": [ + 217, + 114, + 228, + 126 + ], + "page_idx": 3 + }, + { + "type": "header", + "text": "S. Ikehata and Y. Asano", + "bbox": [ + 271, + 114, + 436, + 127 + ], + "page_idx": 3 + }, + { + "type": "image", + "img_path": "images/0c3178d821e2637119bc71ebb6a88d3a5c1c483925b28ab8846e50d737096fcb.jpg", + "image_caption": [ + "Fig. 2: SpectraM-PS involves decomposing a spectrally multiplexed image into independent channels. The Global Feature Encoder extracts a feature map from each channel. The surface vector is then recovered by the Dual-scale Surface Normal Decoder at each pixel. We adopt a dual-scale approach to preserve the entire shape, while employing patch-embedding techniques to enhance local surface details." + ], + "image_footnote": [], + "bbox": [ + 230, + 143, + 781, + 333 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "object, $N \\in \\mathbb{R}^{h \\times w \\times 33}$ . The object is supposed to be illuminated by multiple light sources, each with unique spatial and spectral properties. Previous studies have typically assumed an equal number of light sources and sensor channels, with each light source's wavelength precisely matching the spectral response of a single channel, thereby precluding any channel crosstalk, and with the directions of lights predetermined. In contrast, we do not presuppose the spatial distribution of illumination nor require the spectrum of each light source to be exclusively aligned with the spectral responses of the sensor channels, thus permitting channel crosstalk. This distinction is elaborated in subsequent sections.", + "bbox": [ + 212, + 438, + 787, + 575 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "4 Method", + "text_level": 1, + "bbox": [ + 215, + 599, + 328, + 616 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "We propose and tackle the problem of spectrally multiplexed photometric stereo from a single image with multiple channels, produced by an unknown spectral/spatial composition of the sensor, light, and surface. To build such a method, we train neural networks to directly infer the normal map from an image.", + "bbox": [ + 212, + 633, + 785, + 694 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Our method addresses two challenges: (1) a physics-free architecture that accepts a varying number of spectral channels and is agnostic to their order, and (2) an effective approximation of the spectrally multiplexed image for efficient training. We consider (2) to be of significant importance, yet it remains largely unexplored. Synthesizing spectrally multiplexed images in a physically accurate manner is prohibitively challenging, owing to the increased complexity of their parameter spaces and the scarcity of 3D assets with detailed spectral properties,", + "bbox": [ + 212, + 695, + 787, + 801 + ], + "page_idx": 4 + }, + { + "type": "header", + "text": "SpectraM-PS", + "bbox": [ + 640, + 114, + 730, + 128 + ], + "page_idx": 4 + }, + { + "type": "page_number", + "text": "5", + "bbox": [ + 774, + 116, + 784, + 126 + ], + "page_idx": 4 + }, + { + "type": "page_footnote", + "text": "3 It should be noted that unlike conventional PS, reflectance recovery generally falls outside the scope of spectrally multiplexed PS due to its inherently ill-posed nature.", + "bbox": [ + 217, + 810, + 785, + 839 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "as well as the complex nature of light-surface interactions across different wavelengths. To address this issue, developing an efficient approximation method for rendering spectrally multiplexed images using common RGB image rendering techniques is crucial.", + "bbox": [ + 212, + 146, + 787, + 205 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "4.1 Physics-free Spectrally Multiplexed PS Network (SpectraM-PS)", + "text_level": 1, + "bbox": [ + 214, + 234, + 785, + 250 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "The architecture of SpectraM-PS is illustrated in Fig. 2. Drawing inspiration from established Transformer-based photometric stereo networks [24-26], we integrate an encoder to first extract the global features and a decoder to estimate per-pixel surface normals. The architecture derives normals solely from the input image and mask, without prior light information. This indicates that the architecture focuses the network's learning objective on the relationship between input and output without relying on physics-based principles, unlike prior works.", + "bbox": [ + 212, + 263, + 787, + 369 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "In our model, all the interactions among features from different sensor channels are employed by naïve Transformer [61] in similar to [24-26]. Transformer functions by mapping input features to query, key, and value vectors of equal dimensions. These vectors are processed through a multi-head self-attention mechanism, utilizing a softmax layer, followed by a feed-forward network comprising two linear layers. Both the input and output layers maintain identical dimensionality, with the inner layer having twice the dimension of the input. Each layer is surrounded by a residual connection, succeeded by layer normalization [67]. The advantage of employing Transformers in photometric stereo networks lies in their capability to facilitate complex interactions among intermediate features, a task unachievable with simple operations like pooling [11-13, 46] and observation map [23, 30]. Additionally, the token-based attention mechanism allows for different number of input tokens (i.e., sensor channels) between training and test phases and ensures that the results are independent on the order of tokens.", + "bbox": [ + 212, + 371, + 787, + 582 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Building on the established Transformer-based architecture [26] for temporally multiplexed photometric stereo, we extend its scope to a spectrally multiplexed one. To accommodate a variable number of channels and eliminate dependency on their order, an input spectrally multiplexed image is first split into individual channels, each of which is concatenated with an object mask (If no mask is provided, replace with a matrix of ones.) and then input into the same encoder of a neural network. This approach is distinctly different from traditional methods that encode an input image as it is in neural networks [32, 33]. Then, at Preprocessing, we normalize each channel by dividing it by a random value between its maximum and mean. Each channel and mask are resized or cropped to a resolution $(c \\times c)$ that is a multiple of 32 to be input into the multi-scale encoder. Global Feature Encoder first applies a backbone network (i.e., ConvNeXt-T [43]) to individually encode the concatenation of each channel and mask, then uses Transformer layers for channel-axis (i.e., sensor channel) feature communication across scales (the number of Transformer layers is \\{0, 1, 2, 4\\} at \\{1/4, 1/8, 1/16, 1/32\\} scales, hidden dimensions are same with input dimensions), and finally, a feature pyramid network [66] for integrating features", + "bbox": [ + 212, + 583, + 787, + 840 + ], + "page_idx": 5 + }, + { + "type": "page_number", + "text": "6", + "bbox": [ + 217, + 114, + 228, + 126 + ], + "page_idx": 5 + }, + { + "type": "header", + "text": "S. Ikehata and Y. Asano", + "bbox": [ + 271, + 114, + 436, + 127 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "at different levels. Note that the design of encoder is almost the same as [25, 26], except that images are replaced by sensor channels, so details are omitted.", + "bbox": [ + 212, + 146, + 782, + 175 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Given global features $\\in \\mathcal{R}^{k\\times c / 4\\times c / 4\\times 256}$ , our novel Dual-scale Surface Normal Decoder adopts a dual-scale strategy for predicting point-wise surface normals at $m$ (i.e., 2048) sampled locations at the original resolution within the object mask. The first branch recovers low-frequency surface normals at the feature map resolution $(\\frac{c}{4}\\times \\frac{c}{4})$ . Concretely, all global features corresponding to each sample location are processed by five channel-axis Transformer layers (with a 256 hidden dimension) and are pooled via Pooling-by-Multihead-Attention (PMA) [38] using an additional channel-axis Transformer layer (with a 384 hidden dimension). To enhance spatial communication, two spatial-axis Transformer layers (with a 384 hidden dimension) inspired by Ikehata [26] are employed (i.e., Transformer is employed among samples at different locations), with a final MLP $(384\\to 192\\to 3)$ predicting the low-frequency normals at sampled locations. The second branch focuses on high-resolution normal recovery, using patch embedding for local context at the same $m$ locations, with $w\\times w$ patches ( $w = 21$ ) processed by an MLP (with a 256 hidden dimension) and two layer norms. These patches, concatenated with bilinearly interpolated global features, pass through five channel-axis Transformer blocks (with a 256 hidden dimension), PMA (with a 384 hidden dimension), and are merged with the first branch output normals into 387-dimensional vectors. Two additional spatial-axis Transformer layers (with a 384 hidden dimension) enable non-local interactions, culminating in a final MLP $(384\\to 192\\to 3)$ for high-resolution normals, normalized to unit vectors. The complete normal map is formed by merging all the vectors from different sample sets.", + "bbox": [ + 212, + 176, + 787, + 523 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "It should be noted that while SDM-UniPS [26] targets temporally multiplexed PS with tens of images, and its decoder performs normal estimation purely on a pixel basis. In contrast, spectrally multiplexed PS deals with fewer channels (e.g., three with RGB sensors), making a pixel-basis architecture less effective. Therefore, we use patch embedding at the patch-basis decoder to capture fine details with a dual-scale architecture for preserving overall shape. Without a dual-scale design, the recovery of surface normals becomes overly influenced by local image textures captured through patch embedding. This leads to a failure in preserving the entire shape, resulting in a significant reduction in accuracy. Our motivation is supported by Fig. 3 (left), where SDM-UniPS [26] fails to recover fine details with six temporally multiplexed images, while our architecture produces a more plausible normal map.", + "bbox": [ + 212, + 523, + 787, + 705 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "4.2 Efficient Training Strategy Utilizing Spectral Ambiguity", + "text_level": 1, + "bbox": [ + 214, + 724, + 723, + 742 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Aligning the training and test data domains in neural networks is essential for optimal model performance [9, 60]. However, rendering spectrally multiplexed data poses challenges due to the scarcity of multispectral Bidirectional Reflectance Distribution Functions (BRDFs). In reality, ELIE-Net [46] was trained using only 51 measured isotropic spectral BRDFs. On the other hand, given the availability of various large isotropic BRDF databases [1-3, 5, 48], we seek to explore", + "bbox": [ + 212, + 750, + 787, + 840 + ], + "page_idx": 6 + }, + { + "type": "header", + "text": "SpectraM-PS", + "bbox": [ + 638, + 113, + 730, + 128 + ], + "page_idx": 6 + }, + { + "type": "page_number", + "text": "7", + "bbox": [ + 774, + 114, + 785, + 126 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/b695ec0cf40d59275318953451e4682acbb8053d33bf044e8a064af6b2c004d5.jpg", + "image_caption": [ + "SpectraM-PS on Temporally Multiplexed PS" + ], + "image_footnote": [], + "bbox": [ + 222, + 143, + 442, + 314 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/fa05781d3530d106be98b1c1a759a78a3f9e2397822eccebce576fd92afab5f3.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 442, + 143, + 553, + 229 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/141a0aa36e3067d2608f640e4e7407f3b5d8868d002fd3de8643b69849f105f8.jpg", + "image_caption": [ + "Fig. 3: (Left) Comparison of SpectraM-PS and SDM-UniPS [26] on six temporally multiplexed PS images. Due to the patch-wise basis of SpectraM-PS, fine details are better recovered. (Right) Illustration of different lighting conditions in PS-Multiplex." + ], + "image_footnote": [], + "bbox": [ + 442, + 229, + 553, + 314 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/42e96c0f5b3f02f09b242cde36d0d9b9489f1c69ee99d84b5609a77276f0f075.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 558, + 143, + 668, + 229 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/93339ba1241421b894148825b095f4c1f97caed0c5305c7a19dee2a300bb6b99.jpg", + "image_caption": [ + "PS-Multiplex Training Dataset" + ], + "image_footnote": [], + "bbox": [ + 558, + 231, + 668, + 315 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/bbf1bf948329ec33c243adb4595a152a4f695154cc21a5114d95bf908edb8f92.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 669, + 143, + 781, + 229 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/56c548f0928b8d5f0123cbbfb2acef66aa11cac5be906d8cd2ff6d14704e0cc3.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 669, + 231, + 781, + 315 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "utilizing these datasets for training our model, leveraging the fact that our network does not distinguish images based on their physically-based principles. In this section, we highlight how RGB images serve as a practical approximation, simplifying the complexity inherent in multispectral imaging.", + "bbox": [ + 212, + 398, + 782, + 458 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "We begin the discussion by characterizing multispectral imaging. Assuming that the surface doesn't emit light and only reflections on surface are considered, the image formation model is described as follows [35]:", + "bbox": [ + 212, + 459, + 784, + 503 + ], + "page_idx": 7 + }, + { + "type": "equation", + "text": "\n$$\nI _ {(s, p)} = \\int_ {\\Omega} \\left(\\omega_ {i} ^ {\\top} n _ {p}\\right) \\int_ {0} ^ {\\infty} S _ {s} (\\lambda) f _ {p} \\left(\\omega_ {i}, \\omega_ {o}, \\lambda\\right) L _ {p} \\left(\\omega_ {i}, \\lambda\\right) \\mathrm {d} \\lambda \\mathrm {d} \\omega_ {i}. \\tag {1}\n$$\n", + "text_format": "latex", + "bbox": [ + 299, + 512, + 785, + 546 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "In this equation, $I_{(s,p)}$ denotes the incoming spectral radiance at the sensor $s$ (or $s$ -th channel) from a surface point $p$ . The term $f_p$ represents BRDF, $L_p$ the incident light intensity at the surface point, and $\\lambda$ the wavelength of the incident light. The symbols $\\omega_i$ and $\\omega_o$ denote the directions of incident and reflected light, respectively. $S_s(\\lambda)$ refers to the spectral sensitivity of the sensor $s$ at wavelength $\\lambda$ , $n_p$ is the surface normal, and $\\varOmega$ represents the hemisphere over which incident light directions are possible. The integral sums over all incident directions and wavelengths. It is important to note that the incident light intensity $L_p$ depends not only on the direct contribution from light sources but also on the visibility of light (e.g., attached and cast shadows) and indirect illuminations.", + "bbox": [ + 212, + 551, + 787, + 703 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Eq. (1) illustrates the concept of spectral ambiguity, showing that an infinite number of combinations of $S_{s}(\\lambda)$ , $f_{p}(\\omega_{i},\\omega_{o},\\lambda)$ , and $L_{p}(\\omega_{i},\\lambda)$ can result in the same spectral radiance, including narrowband compositions. In other words, with spectral ambiguity, a single observation $I_{(s,p)}$ can encompass the observations for all spectral compositions that satisfy the equation (i.e., metamerism [22, 50]). This perspective justifies the theory of substituting multispectral images, which possess a broad parameter space, with narrowband RGB images. It is worth mentioning that channel crosstalk primarily affects the incident light intensity, consequently distorting the product of $S_{s}(\\lambda)\\cdot f_{p}(\\omega_{i},\\omega_{o},\\lambda)\\cdot L_{p}(\\omega_{i},\\lambda)$ in Eq. (1).", + "bbox": [ + 212, + 704, + 787, + 842 + ], + "page_idx": 7 + }, + { + "type": "page_number", + "text": "8", + "bbox": [ + 217, + 114, + 228, + 126 + ], + "page_idx": 7 + }, + { + "type": "header", + "text": "S. Ikehata and Y. Asano", + "bbox": [ + 271, + 114, + 436, + 127 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "This implies that observations influenced by spectral crosstalk can still be equivalently represented using a narrowband setup under spectral ambiguity. In the experiments, we demonstrate that our model, trained on three narrowband observations can be applied to multiplexed data with channel crosstalk. To realize this approximation, we rendered a large number of three-channel narrowband images using the path-tracing algorithm in Blender [4], where up to 10-bounce reflections are permitted, based on common 3D assets [2] for RGB rendering. Following the rendering pipeline described in [26], we rendered objects by combining three different lighting models: directional, point, and environmental (five combinatorial settings in total as shown in Fig. 3). To simulate spectrally multiplexed images, we defined R, G, and B light sources and illuminated the surface in a multiplexed manner. It is important to note that the rendered RGB images are decomposed into three grayscale images, each of which was independently fed into the network; therefore, any wavelength-dependent information is masked. For material diversity, we adopted the method from [26], categorizing 897 AdobeStock texture maps into three groups: 421 diffuse, 219 specular, and 257 metallic textures. Four objects from a set of 410 3D AdobeStock models were randomly selected and textured with these materials. This structured approach led to the rendering of 106,374 multiplexed images along with their ground truth surface normal maps, forming the 'PS-Multiplex' dataset.", + "bbox": [ + 212, + 145, + 787, + 450 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "5 SpectraM14 Benchmark Dataset", + "text_level": 1, + "bbox": [ + 215, + 469, + 566, + 486 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "Due to the lack of a benchmark for spectrally multiplexed PS, the first comprehensive evaluation dataset, named SpectraM14, is created. This dataset includes 14 objects, each exhibiting a range of optical properties such as monochromatic or multicolored appearances and diffuse or specular reflections, as depicted in Fig. 4. Our benchmark encompasses tasks under five distinct conditions, as described later.", + "bbox": [ + 212, + 500, + 787, + 589 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "Imaging Setup. To acquire our dataset, we utilized a color camera (FLIR GS3-U3-123S6C-C) and an NIR camera (FLIR GS3-U3-41C6NIR-C), both equipped with a $50\\mathrm{mm}$ lens. For the NIR camera, we used narrowband filters with wavelengths of $750\\mathrm{nm}$ , $850\\mathrm{nm}$ , $880\\mathrm{nm}$ , $905\\mathrm{nm}$ , and $940\\mathrm{nm}$ , and the acquired images were manually merged. Objects were placed $0.8\\mathrm{m}$ from the camera to approximate orthographic projection. Following conventional PS benchmarks [52,56,63], data capture occurred in a controlled, dark environment with the scene draped in black cloth to mitigate interreflection. The camera's ISO sensitivity was minimized to enhance image quality. The imaging area was further isolated using low-reflectance cloths to suppress inter-reflection. For each illumination condition, we collected six images under varying exposures to produce HDR input images. For the evaluations throughout this paper, the images are cropped using an object mask and resized to $512\\mathrm{px} \\times 512\\mathrm{px}$ .", + "bbox": [ + 212, + 594, + 787, + 792 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "Lighting. Six LED and three halogen light sources, positioned roughly 1 meter from the object, provided illumination. We used the \"Weeylite S05 RGB Pocket Lamp\" and the \"NPI PIS-UHX-AIR\" for lighting. This setup enabled the use of", + "bbox": [ + 212, + 794, + 787, + 840 + ], + "page_idx": 8 + }, + { + "type": "header", + "text": "SpectraM-PS", + "bbox": [ + 640, + 114, + 730, + 128 + ], + "page_idx": 8 + }, + { + "type": "page_number", + "text": "9", + "bbox": [ + 774, + 116, + 785, + 126 + ], + "page_idx": 8 + }, + { + "type": "image", + "img_path": "images/4df85b5950b772f90a0bd59842858eea97972ceb5bd5cca701904cfbd75c86e8.jpg", + "image_caption": [ + "Fig. 4: Objects in SpectraM14." + ], + "image_footnote": [], + "bbox": [ + 222, + 143, + 781, + 323 + ], + "page_idx": 9 + }, + { + "type": "image", + "img_path": "images/a034460c5348bb0c6d6a9aa4054cb7875debf186dcb33da0567907bcb923e1ea.jpg", + "image_caption": [ + "Fig. 5: Illustration of six conditions in SpectraM14." + ], + "image_footnote": [], + "bbox": [ + 225, + 357, + 776, + 444 + ], + "page_idx": 9 + }, + { + "type": "text", + "text": "red, green, blue, yellow, magenta, cyan, and NIR lighting, with spectra validated using a Hamamatsu Photonics Multichannel Analyzer C10027-01.", + "bbox": [ + 212, + 497, + 784, + 527 + ], + "page_idx": 9 + }, + { + "type": "text", + "text": "Calibration and Ground Truth Data. We measured the directions of lights using specular reflections from a mirror sphere. Light intensity was standardized across the visible spectrum by averaging RGB values from reflected light on a white target. The ground truth normals were captured with a SHINING 3D EinScan-SE scanner.", + "bbox": [ + 212, + 532, + 787, + 607 + ], + "page_idx": 9 + }, + { + "type": "text", + "text": "Evaluation Procedure. The design philosophy of this benchmark is to assess the robustness and adaptability of spectrally multiplexed PS methods under realistic lighting conditions, accounting for variations in channel numbers and the presence of spectral crosstalk. For a comprehensive evaluation, we designed tasks under five distinct conditions as shown in Fig. 5: Condition 1: Color sensor, no crosstalk condition: Six colors of light (red, green, blue, cyan, yellow, magenta) were each independently illuminated and observed with an RGB sensor. Afterward, the channels of RGB were averaged. Condition 2: Color sensor, weak crosstalk condition: Three colors of light (red, green, blue) were simultaneously illuminated and observed through each channel of the RGB sensor. Condition 3: Color sensor, strong crosstalk condition: Three colors of light (cyan, yellow, magenta) were simultaneously illuminated and observed through each channel of the RGB sensor. Condition 4: NIR sensor, no crosstalk condition: Light at wavelengths of $750~\\mathrm{nm}$ , $850~\\mathrm{nm}$ , $880~\\mathrm{nm}$ , $905~\\mathrm{nm}$ , and $940~\\mathrm{nm}$ were each independently illuminated and observed with a monochrome sensor corresponding to", + "bbox": [ + 212, + 613, + 787, + 840 + ], + "page_idx": 9 + }, + { + "type": "page_number", + "text": "10", + "bbox": [ + 217, + 114, + 235, + 126 + ], + "page_idx": 9 + }, + { + "type": "header", + "text": "S. Ikehata and Y. Asano", + "bbox": [ + 271, + 114, + 436, + 127 + ], + "page_idx": 9 + }, + { + "type": "text", + "text": "each wavelength. Condition 5: NIR sensor, spatially-varying lighting condition: New images were created by averaging two images taken under the conditions mentioned above. The combinations were (750 nm, 850 nm), (850 nm, 880 nm), (880 nm, 905 nm), (905 nm, 940 nm), and (940 nm, 750 nm).", + "bbox": [ + 218, + 146, + 784, + 205 + ], + "page_idx": 10 + }, + { + "type": "text", + "text": "6 Experiment", + "text_level": 1, + "bbox": [ + 218, + 233, + 362, + 250 + ], + "page_idx": 10 + }, + { + "type": "text", + "text": "In this section, we evaluate our method on our SpectraM14. Our method is compared with one SOTA optimization-based method [17] and one SOTA learning-based method [46]. The former introduces a closed-form solution for spectrally multiplexed photometric stereo applied to monochromatic surfaces with spatially varying (SV) albedo. The latter presents a Spectral Reflectance Decomposition (SRD) model, which disentangles spectral reflectance into geometric and spectral components for surface normal recovery under non-Lambertian spectral reflectance conditions. Unlike the compared methods, our approach does not assume a specific lighting setup, whereas both methods presume the presence of calibrated single directional light sources.", + "bbox": [ + 218, + 267, + 784, + 417 + ], + "page_idx": 10 + }, + { + "type": "text", + "text": "Training details. SpectraM-PS was trained from scratch on the PS-Multiplex dataset until convergence using the AdamW optimizer, with a step decay learning rate schedule that reduced the learning rate by a factor of 0.8 every ten epochs. We applied learning rate warmup during the first epoch and used a batch size of 16, an initial learning rate of 0.0001, and a weight decay of 0.05. Each batch consisted of three input training multiplexed images with three channels each. The training loss was computed using the Mean Squared Error (MSE) loss function to measure $\\ell_2$ errors between the predicted surface normal vectors and the ground truth surface normal vectors. We measured the reconstruction accuracy of our method by computing the mean angular errors (MAE) between the predicted and true surface normal maps, expressed in degrees.", + "bbox": [ + 218, + 422, + 784, + 589 + ], + "page_idx": 10 + }, + { + "type": "text", + "text": "Computational Cost. The inference time of PS methods varies with the number of pixels and channels in the input image. For Condition 2 and 3 with a $512 \\times 512 \\times 3$ image, the mean and standard deviation of inference times (in sec) over 14 objects in SpectraM14 benchmark were: our method (3.42/0.85), Lv et al. [46] (0.46/0.24) and Guo et al. [17] (2.38/1.10). Our architecture leads to higher computational costs; however, none of the methods were suitable for real-time processing (e.g., 15 fps requires 0.06 sec/frame).", + "bbox": [ + 218, + 593, + 784, + 699 + ], + "page_idx": 10 + }, + { + "type": "text", + "text": "Ablation Study. We firstly validate the individual technical contributions of our training dataset (i.e., PS-Multiplex) and the physics-free architecture (i.e., SpectraM-PS) using a synthetic evaluation dataset. Firstly, we validate the efficacy of our training dataset, PS-Multiplex, by adapting an existing universal photometric stereo architecture designed for the conventional task (i.e., SDM-UniPS [26]) to the spectrally multiplexed photometric stereo task. Since both ours and SDM-UniPS take multiple observations and an object mask as input, this adaptation straightforwardly involves training the model on PS-Multiplex by treating each channel of an image as an individual image. Subsequently, we", + "bbox": [ + 218, + 703, + 784, + 839 + ], + "page_idx": 10 + }, + { + "type": "header", + "text": "SpectraM-PS", + "bbox": [ + 640, + 114, + 730, + 128 + ], + "page_idx": 10 + }, + { + "type": "page_number", + "text": "11", + "bbox": [ + 767, + 116, + 782, + 126 + ], + "page_idx": 10 + }, + { + "type": "table", + "img_path": "images/7d60b929da37779a5e73b98fd496accafaf28eea73ae072dfe6e13321f3d0107.jpg", + "table_caption": [ + "Table 1: Ablation analysis of the contributions of SpectraM-PS and PS-Multiplex." + ], + "table_footnote": [], + "table_body": "
MAE (Uniform)MAE (Piece-wise uniform)MAE (Non-uniform)
MethodNon-LambertianLambertianNon-LambertianLambertianNon-LambertianLambertian
SDM-UniPS [26]12.9 (4.5)12.4 (4.7)15.0 (5.0)21.7 (7.0)14.4 (5.1)15.2 (5.9)
[26] trained on PS-Multiplex11.1 (3.6)11.0 (3.9)10.5 (2.9)12.3 (3.9)10.6 (3.9)11.2 (3.7)
SpectraM-PS (Ours)8.0 (2.7)8.4 (2.7)7.9 (2.4)8.9 (2.9)8.2 (3.2)8.0 (2.5)
", + "bbox": [ + 230, + 164, + 774, + 243 + ], + "page_idx": 11 + }, + { + "type": "text", + "text": "compare this model against our proposed SpectraM-PS to demonstrate the efficacy of our dual-scale design with local patch embedding.", + "bbox": [ + 212, + 258, + 782, + 287 + ], + "page_idx": 11 + }, + { + "type": "text", + "text": "For evaluating the contribution of our architecture (SpectraM-PS) and training dataset (PS-Multiplex), we additionally rendered three-channel spectrally multiplexed images representing six distinct surface material categories: (a) uniform, Lambertian; (b) piece-wise uniform, Lambertian; (c) non-uniform, Lambertian; (d) uniform, non-Lambertian; (e) piece-wise uniform, non-Lambertian; and (f) non-uniform, non-Lambertian. In uniform materials, every point on the surface within a scene exhibits the same material properties. For piece-wise uniform materials, each object in a scene is composed of the same material, yet different objects possess distinct materials. Non-uniform materials feature unique PBR textures assigned to each object. The rendering process for these images was identical to that used for the PS-Multiplex datasets in each category. We generated 100 scenes for each surface material category, and MAEs (stds) are averaged over them. The results are presented in Tab.1. In summary, SDM-UniPS [26] trained on our PS-Multiplex dataset demonstrates proper adaptation to the spectrally multiplexed photometric stereo task. Nonetheless, our SpectraM-PS method significantly enhanced reconstruction accuracy, showcasing an architecture-level improvement over SDM-UniPS for the spectrally multiplexed photometric stereo task, where the number of input channels is typically much fewer than that of input images for conventional PS.", + "bbox": [ + 212, + 290, + 787, + 578 + ], + "page_idx": 11 + }, + { + "type": "text", + "text": "Comparative Evaluation on SpectraM14. The results are illustrated in Tabs. 2", + "text_level": 1, + "bbox": [ + 212, + 583, + 803, + 597 + ], + "page_idx": 11 + }, + { + "type": "text", + "text": "to 6 and Fig. 6. Despite the fact that all existing spectrally multiplexed photometric stereo methods assume calibrated light sources and known directional light source conditions, our proposed method significantly outperformed them. This is because most of the real objects used in our experiment are neither Lambertian nor convex, and do not conform to their assumptions. However, our non-physical-based method successfully restored the normals very stably for these objects. Furthermore, our proposed method enabled robust reconstruction for all objects, despite having been trained only with RGB color images. This result supports the efficacy of our approximation. Furthermore, unlike existing methods that suffer from reduced estimation accuracy with increasing spectral crosstalk, our approach demonstrates only minimal performance degradation. Remarkably, our method excels in recovering a more realistic structure with spatially-varying surface materials. This breakthrough implies that our network can effectively achieve dynamic surface reconstruction across video frames in a universal setting. We will detail this groundbreaking application in the next section. Due to space constraints, not all results can be included here. However, all", + "bbox": [ + 212, + 598, + 787, + 840 + ], + "page_idx": 11 + }, + { + "type": "page_number", + "text": "12", + "bbox": [ + 217, + 114, + 235, + 126 + ], + "page_idx": 11 + }, + { + "type": "header", + "text": "S. Ikehata and Y. Asano", + "bbox": [ + 271, + 114, + 436, + 127 + ], + "page_idx": 11 + }, + { + "type": "table", + "img_path": "images/aff885e0dceeef0084fcc05957627766ce0989c3b587dbef08f05a98dd3d0e47.jpg", + "table_caption": [ + "Table 2: Comparison in condition 1. The values are mean angular errors in degrees." + ], + "table_footnote": [], + "table_body": "
MethodObject ID
1234567891011121314Ave.
Ours8.611.213.07.312.26.111.55.05.77.510.35.16.112.28.9
Lv et al. [46]20.417.021.113.923.110.721.216.610.915.919.016.413.218.617.1
Guo et al. [17]22.615.220.713.427.17.231.324.88.018.224.511.410.229.318.9
", + "bbox": [ + 220, + 160, + 782, + 215 + ], + "page_idx": 12 + }, + { + "type": "table", + "img_path": "images/ab636ee77231ad6daaa269676db9e7373589cf893351c3f50537a1f45e0b13cc.jpg", + "table_caption": [ + "Table 3: Comparison in condition 2. The values are mean angular errors in degrees." + ], + "table_footnote": [], + "table_body": "
MethodObject IDAve.
1234567891011121314
Ours10.111.513.59.012.87.010.95.56.39.312.75.39.814.710.0
Lv et al. [46]22.826.027.019.430.319.522.118.914.319.823.520.819.021.421.7
Guo et al. [17]31.127.529.220.138.019.033.423.513.626.632.714.717.139.325.7
", + "bbox": [ + 218, + 241, + 784, + 301 + ], + "page_idx": 12 + }, + { + "type": "table", + "img_path": "images/e02131d72b8805ec7483481a572a0cbd148e52fa176630fffbee403090147a4d.jpg", + "table_caption": [ + "Table 4: Comparison in condition 3. The values are mean angular errors in degrees." + ], + "table_footnote": [], + "table_body": "
MethodObject ID
1234567891011121314
Ours12.012.815.811.016.55.912.39.86.99.814.66.37.520.1
Lv et al. [46]38.534.138.232.440.138.236.629.638.235.138.236.638.430.9
Guo et al. [17]46.042.656.337.657.145.676.048.029.962.149.850.252.873.7
", + "bbox": [ + 218, + 325, + 784, + 385 + ], + "page_idx": 12 + }, + { + "type": "table", + "img_path": "images/1eed9fd05c367f866fd35c947db5a179ee59b9d48a428b7e75017e8f51e07ee1.jpg", + "table_caption": [ + "Table 5: Comparison in condition 4. The values are mean angular errors in degrees." + ], + "table_footnote": [], + "table_body": "
MethodObject ID
1234567891011121314
Ours10.911.39.76.310.85.312.44.37.86.69.14.07.410.6
Lv et al. [46]24.119.820.712.019.812.415.118.711.417.121.217.519.116.7
Guo et al. [17]30.716.018.39.929.68.423.125.710.214.624.513.929.613.8
", + "bbox": [ + 220, + 409, + 781, + 468 + ], + "page_idx": 12 + }, + { + "type": "table", + "img_path": "images/032798ec3ac385d6807f53395969df4695b0f5d07b2ebef832474a113915ce61.jpg", + "table_caption": [ + "Table 6: Comparison in condition 5. The values are mean angular errors in degrees." + ], + "table_footnote": [], + "table_body": "
MethodObject IDAve.
1234567891011121314
Ours10.911.09.97.311.04.713.34.87.96.59.84.27.410.48.6
Lv et al. [46]29.826.228.922.828.925.821.924.618.724.727.624.822.327.625.0
Guo et al. [17]40.027.329.223.233.325.029.330.120.026.731.626.325.027.227.8
", + "bbox": [ + 222, + 493, + 781, + 551 + ], + "page_idx": 12 + }, + { + "type": "text", + "text": "results are comprehensively presented in the supplementary materials. Additionally, the supplementary materials evaluate the impact of the spatial distribution of light sources on the performance of the proposed method. We also offer an in-depth discussion of each experimental condition therein.", + "bbox": [ + 212, + 569, + 787, + 630 + ], + "page_idx": 12 + }, + { + "type": "text", + "text": "7 Conclusion", + "text_level": 1, + "bbox": [ + 215, + 655, + 359, + 671 + ], + "page_idx": 12 + }, + { + "type": "text", + "text": "In this work, we introduce an innovative approach to spectrally multiplexed photometric stereo under unknown spatial/spectral composition. Turning spectral ambiguity into a benefit, our method allows for the creation of training data without the need for complex multispectral rendering. Our work significantly broadens the scope for dynamic surface analysis, establishing a critical advancement in the utilization of photometric stereo across multiple sectors. Our proposed method exhibits several limitations. Firstly, there is unstable temporal variation in the normal maps reconstructed by our method for dynamic surface reconstruction. This instability arises from factors such as motion blur in certain frames, image noise, or the influence of cast/attached shadows, which become", + "bbox": [ + 212, + 688, + 787, + 840 + ], + "page_idx": 12 + }, + { + "type": "header", + "text": "SpectraM-PS", + "bbox": [ + 640, + 114, + 730, + 128 + ], + "page_idx": 12 + }, + { + "type": "page_number", + "text": "13", + "bbox": [ + 767, + 114, + 785, + 126 + ], + "page_idx": 12 + }, + { + "type": "image", + "img_path": "images/f141a01604beedf219d31153145f94f5220479126cd51437fba78600b9794178.jpg", + "image_caption": [ + "Fig. 6: Evaluation on SpectraM14. Full results are available in the supplementary." + ], + "image_footnote": [], + "bbox": [ + 228, + 146, + 777, + 683 + ], + "page_idx": 13 + }, + { + "type": "text", + "text": "more pronounced compared to conventional photometric stereo methods that utilize numerous images. To recover clean and temporally stable normal maps, we may need to consider temporal consistency and more actively utilize monocular cues. Additionally, while our method targets dynamic surfaces, it currently requires several seconds to up to ten seconds per RGB image, which is far from real-time processing. Considering industrial applications in the future, accelerating the processing speed is a crucial challenge.", + "bbox": [ + 212, + 726, + 787, + 833 + ], + "page_idx": 13 + }, + { + "type": "page_number", + "text": "14", + "bbox": [ + 217, + 114, + 235, + 126 + ], + "page_idx": 13 + }, + { + "type": "header", + "text": "S. Ikehata and Y. Asano", + "bbox": [ + 271, + 114, + 436, + 127 + ], + "page_idx": 13 + }, + { + "type": "text", + "text": "References", + "text_level": 1, + "bbox": [ + 215, + 143, + 323, + 159 + ], + "page_idx": 14 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "1. 3D Textures - Free seamless PBR textures with Diffuse, Normal, Displacement, Occlusion, Specularity and Roughness Maps. https://3dtextures.me/, accessed: 2024-03-07 7", + "2. Adobe Stock. https://stock.adobe.com/ 7,9", + "3. AmbientCG - Free Public Domain PBR Materials. https://ambientcg.com/, accessed: 2024-03-07 7", + "4.Blender.https://www.blender.org/9", + "5. Poliigon - A library of materials, and HDR's for artists including free textures. https://www.poliigon.com/, accessed: 2024-03-07 7", + "6. Alldrin, N., Mallick, S., Kriegman, D.: Resolving the generalized bas-relief ambiguity by entropy minimization. CVPR (2007) 4", + "7. Anderson, R., Stenger, B., Cipolla, R.: Color photometric stereo for multicolored surfaces. ICCV (2011) 2", + "8. Basri, R., Jacobs, D., Kemelmacher, I.: Photometric stereo with general, unknown lighting. International Journal of computer vision 72(3), 239-257 (2007) 4", + "9. Ben-David, S., Blitzer, J., Crammer, K., Kulesza, A., Pereira, F., Vaughan, J.W.: A theory of learning from different domains. Machine learning 79, 151-175 (2010) 7", + "10. Chakrabarti, A., Sunkavalli, K.: Single-image rgb photometric stereo with spatially-varying albedo. In: 2016 Fourth International Conference on 3D Vision (3DV) (2016) 2", + "11. Chen, G., Han, K., Shi, B., Matsushita, Y., Wong, K.K.K.: Self-calibrating deep photometric stereo networks. In: 2019 IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR). pp. 8731-8739 (2019) 2, 4, 6", + "12. Chen, G., Han, K., Wong, K.Y.K.: Ps-fcn: A flexible learning framework for photometric stereo. ECCV (2018) 2, 4, 6", + "13. Chen, G., Waechter, M., Shi, B., Wong, K.Y.K., Matsushita, Y.: What is learned in deep uncalibrated photometric stereo? In: European Conference on Computer Vision. pp. 745-762. Springer (2020) 4, 6", + "14. Decker, B., Kautz, J., Mertens, T., Bekaert, P.: Capturing multiple illumination conditions using time and color multiplexing. CVPR (2009) 2", + "15. Drew, M.S.: Shape from color. Technical Report CSS/LCCR TR 92-07, School of Computing Science, Simon Fraser University, Vancouver, BC (1992) 4", + "16. Goldman, D., Curless, B., Hertzmann, A., Seitz, S.: Shape and spatially-varying brdfs from photometric stereo. In: ICCV (October 2005) 3, 4", + "17. Guo, H., Okura, F., Shi, B.: Multispectral photometric stereo for spatially-varying spectral reflectances. IJCV 130, 2166-2183 (2022) 2, 11, 13", + "18. Guo, H., Okura, F., Shi, B., Funatomi, T., Mukaigawa, Y., Matsushita, Y.: Multispectral photometric stereo for spatially-varying spectral reflectances: A well posed problem? In: CVPR. pp. 963-971 (2021) 2, 3", + "19. Hayakawa, H.: Photometric stereo under a light source with arbitrary motion. JOSA 11(11), 3079-3089 (1994) 4", + "20. Hernández, C., Vogiatzis, G., Brostow, G., Stenger, B., Cipolla, R.: Non-rigid photometric stereo with colored lights. In: ICCV. pp. 1-8 (2007). https://doi.org/10.1109/ICCV.2007.4408939 2, 4", + "21. Hertzmann, A., Seitz, S.: Example-based photometric stereo: shape reconstruction with general, varying brdfs. IEEE TPAMI 27(8), 1254-1264 (2005) 4" + ], + "bbox": [ + 218, + 178, + 785, + 839 + ], + "page_idx": 14 + }, + { + "type": "header", + "text": "SpectraM-PS", + "bbox": [ + 638, + 114, + 730, + 128 + ], + "page_idx": 14 + }, + { + "type": "page_number", + "text": "15", + "bbox": [ + 767, + 114, + 785, + 126 + ], + "page_idx": 14 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "22. Hill, B.: Color capture, color management, and the problem of metamerism: does multispectral imaging offer the solution? In: Proc. SPIE. pp. 2-14. SPIE (1999) 2, 8", + "23. Ikehata, S.: Cnn-ps: Cnn-based photometric stereo for general non-convex surfaces. In: ECCV (2018) 2, 4, 6", + "24. Ikehata, S.: Ps-transformer: Learning sparse photometric stereo network using self-attention mechanism. In: BMVC (2021) 4, 6", + "25. Ikehata, S.: Universal photometric stereo network using global lighting contexts. In: CVPR (2022) 2, 3, 4, 6, 7", + "26. Ikehata, S.: Scalable, detailed and mask-free universal photometric stereo. In: CVPR (2023) 2, 3, 4, 6, 7, 8, 9, 11, 12", + "27. Ikehata, S., Aizawa, K.: Photometric stereo using constrained bivariate regression for general isotropic surfaces. In: CVPR (2014) 3", + "28. Ikehata, S., Wipf, D., Matsushita, Y., Aizawa, K.: Robust photometric stereo using sparse regression. In: CVPR (2012) 3, 4", + "29. Ikehata, S., Wipf, D., Matsushita, Y., Aizawa, K.: Photometric stereo using sparse bayesian regression for general diffuse surfaces. IEEE TPAMI 36(9), 1816-1831 (2014) 2, 3", + "30. Ikehata, S.: Does physical interpretability of observation map improve photometric stereo networks? In: ICIP (2022) 4, 6", + "31. Ishio, H., Minowa, J., Nosu, K.: Review and status of wavelength-division-multiplexing technology and its application. Journal of lightwave technology 2(4), 448-463 (1984) 2", + "32. Ju, Y., Dong, X., Wang, Y., Qi, L., Dong, J.: A dual-cue network for multispectral photometric stereo. Pattern Recognition 100, 107162 (2020) 2, 4, 6", + "33. Ju, Y., Qi, L., Zhou, H., Dong, J., Lu, L.: Demultiplexing colored images for multispectral photometric stereo via deep neural networks. IEEE Access 6, 30804-30818 (2018) 2, 4, 6", + "34. Ju, Y., Dong, J., Chen, S.: Recovering surface normal and arbitrary images: A dual regression network for photometric stereo. IEEE Transactions on Image Processing 30, 3676-3690 (2021) 4", + "35. Kajiya, J.T.: The rendering equation. SIGGRAPH Comput. Graph. 20(4), 143-150 (1986) 8", + "36. Kaya, B., Kumar, S., Oliveira, C., Ferrari, V., Van Gool, L.: Uncalibrated neural inverse rendering for photometric stereo of general surfaces. pp. 3804-3814 (2021) 4", + "37. Kontsevich, L.L., Petrov, A., Vergelskaya, I.: Reconstruction of shape from shading in color images. Journal of the Optical Society of America pp. 1047-1052 (1994) 2, 4", + "38. Lee, J., Lee, Y., Kim, J., Kosiorek, A., Choi, S., Teh, Y.W.: Set transformer: A framework for attention-based permutation-invariant neural networks. In: ICML. pp. 3744-3753 (2019) 7", + "39. Li, J., Robles-Kelly, A., You, S., Matsushita, Y.: Learning to minify photometric stereo. In: CVPR (2019) 4", + "40. Li, J., Li, H.: Neural reflectance for shape recovery with shadow handling. In: CVPR (2022) 4", + "41. Li, J., Li, H.: Self-calibrating photometric stereo by neural inverse rendering. In: ECCV (2022) 4", + "42. Liu, H., Yan, Y., Song, K., Yu, H.: Sps-net: Self-attention photometric stereo network. IEEE Transactions on Instrumentation and Measurement 70, 1-13 (2021) 4" + ], + "bbox": [ + 215, + 147, + 784, + 837 + ], + "page_idx": 15 + }, + { + "type": "page_number", + "text": "16", + "bbox": [ + 217, + 114, + 235, + 126 + ], + "page_idx": 15 + }, + { + "type": "header", + "text": "S. Ikehata and Y. Asano", + "bbox": [ + 271, + 114, + 436, + 127 + ], + "page_idx": 15 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "43. Liu, Z., Mao, H., Chao-Yuan Wu, C.F.: A convnet for the 2020s. In: CVPR (2022) 6", + "44. Logothetis, F., Budvytis, I., Mecca, R., Cipolla, R.: Px-net: Simple and efficient pixel-wise training of photometric stereo networks. In: CVPR. pp. 12757-12766 (2021) 4", + "45. Lu, F., Matsushita, Y., Sato, I., Okabe, T., Sato, Y.: Uncalibrated photometric stereo for unknown isotropic reflectances. In: CVPR. pp. 1490-1497 (2013) 4", + "46. Lv, J., Guo, H., Chen, G., Liang, J., Shi, B.: Non-lambertian multispectral photometric stereo via spectral reflectance decomposition. In: IJCAI (2023) 2, 3, 4, 6, 7, 11, 13", + "47. Mecca, R., Rosman, G., Kimmel, R., Bruckstein, A.: Perspective photometric stereo with shadows. In: Proc. of 4th International Conference on Scale Space and Variational Methods in Computer Vision (2013) 4", + "48. Mitsubishi Electric Research Laboratories (MERL): MERL BRDF Database. http://www.merl.com/brdf/, accessed: 2024-03-07 7", + "49. Mo, Z., Shi, B., Lu, F., Yeung, S.K., Matsushita, Y.: Uncalibrated photometric stereo under natural illumination. pp. 2936-2945. IEEE Computer Society (2018) 4", + "50. Nayatani, Y., Kurioka, Y., Sobagaki, H.: Study on color rendering and metamerism (part 8). Journal of the Illuminating Engineering Institute of Japan 56(9), 529-536 (1972). https://doi.org/10.2150/jieij1917.56.9_529_2, 8", + "51. Ozawa, K., Sato, I., Yamaguchi, M.: Single color image photometric stereo for multi-colored surfaces. Computer Vision and Image Understanding 171, 140-149 (2018) 2", + "52. Ren, J., Wang, F., Zhang, J., Zheng, Q., Ren, M., Shi, B.: Diligent10²: A photometric stereo benchmark dataset with controlled shape and material variation. pp. 12581-12590 (June 2022) 3, 9", + "53. Santo, H., Samejima, M., Sugano, Y., Shi, B., Matsushita, Y.: Deep photometric stereo network. In: International Workshop on Physics Based Vision meets Deep Learning (PBDL) in Conjunction with IEEE International Conference on Computer Vision (ICCV) (2017) 4", + "54. Shi, B., Matsushita, Y., Wei, Y., Xu, C., Tan, P.: Self-calibrating photometric stereo. In: CVPR (2010) 2, 4", + "55. Shi, B., Tan, P., Matsushita, Y., Ikeuchi, K.: A biquadratic reflectance model for radiometric image analysis. In: CVPR (2012) 2", + "56. Shi, B., Wu, Z., Mo, Z., D.Duan, Yeung, S.K., Tan, P.: A benchmark dataset and evaluation for non-lambertian and uncalibrated photometric stereo. In: CVPR (2016) 3, 9", + "57. Silver, W.M.: Determining shape and reflectance using multiple images. Master's thesis, MIT (1980) 1", + "58. Taniai, T., Maehara, T.: Neural Inverse Rendering for General Reflectance Photometric Stereo. In: ICML (2018) 4", + "59. Tiwari, A., Raman, S.: Deepps2: Revisiting photometric stereo using two differently illuminated images. ECCV (2022) 4", + "60. Tzeng, E., Hoffman, J., Saenko, K., Darrell, T.: Adversarial discriminative domain adaptation. In: CVPR (2017) 7", + "61. Vaswani, A., Shazeer, N., Parmar, N., Uszkoreit, J., Jones, L., Gomez, A., Kaiser, L., Polosukhin, I.: Attention is all you need. In: NeurIPS (2017) 6", + "62. Vogiatzis, G., Hernandez, C.: Self-calibrated, multi-spectral photometric stereo for 3d face capture. IJCV 56(97), 91-103 (2012) 2" + ], + "bbox": [ + 215, + 146, + 785, + 839 + ], + "page_idx": 16 + }, + { + "type": "header", + "text": "SpectraM-PS", + "bbox": [ + 638, + 114, + 730, + 128 + ], + "page_idx": 16 + }, + { + "type": "page_number", + "text": "17", + "bbox": [ + 767, + 114, + 785, + 126 + ], + "page_idx": 16 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "63. Wang, F., Ren, J., Guo, H., Ren, M., Shi, B.: Diligent-pi: A photometric stereo benchmark dataset with controlled shape and material variation (October 2023) 3, 9", + "64. Woodham, P.: Photometric method for determining surface orientation from multiple images. Opt. Engg 19(1), 139-144 (1980) 1, 3", + "65. Wu, L., Ganesh, A., Shi, B., Matsushita, Y., Wang, Y., Ma, Y.: Robust photometric stereo via low-rank matrix completion and recovery. In: ACCV (2010) 4", + "66. Xiao, T., Liu, Y., Zhou, B., Jiang, Y., Sun, J.: Unified perceptual parsing for scene understanding. In: ECCV. pp. 418-434 (2018) 6", + "67. Xiong, R., Yang, Y., He, D., Zheng, K., Zheng, S., Xing, C., Zhang, H., Lan, Y., Wang, L., Liu, T.: On layer normalization in the transformer architecture. In: International Conference on Machine Learning. pp. 10524-10533. PMLR (2020) 6", + "68. Yao, Z., Li, K., Fu, Y., Hu, H., Shi, B.: Gps-net: Graph-based photometric stereo network. NeurIPS (2020) 4", + "69. Zheng, Q., Jia, Y., Shi, B., Jiang, X., Duan, L.Y., Kot, A.: Spline-net: Sparse photometric stereo through lighting interpolation and normal estimation networks. ICCV (2019) 4" + ], + "bbox": [ + 215, + 146, + 784, + 383 + ], + "page_idx": 17 + }, + { + "type": "page_number", + "text": "18", + "bbox": [ + 217, + 114, + 235, + 126 + ], + "page_idx": 17 + }, + { + "type": "header", + "text": "S. Ikehata and Y. Asano", + "bbox": [ + 271, + 114, + 436, + 127 + ], + "page_idx": 17 + } +] \ No newline at end of file diff --git a/2024/Physics-Free Spectrally Multiplexed Photometric Stereo under Unknown Spectral Composition/09ecf602-4895-4520-8961-a33ebbe3e42d_model.json b/2024/Physics-Free Spectrally Multiplexed Photometric Stereo under Unknown Spectral Composition/09ecf602-4895-4520-8961-a33ebbe3e42d_model.json new file mode 100644 index 0000000000000000000000000000000000000000..565905cd25716e47b5f0e46335ec71e124901cb1 --- /dev/null +++ b/2024/Physics-Free Spectrally Multiplexed Photometric Stereo under Unknown Spectral Composition/09ecf602-4895-4520-8961-a33ebbe3e42d_model.json @@ -0,0 +1,2326 @@ +[ + [ + { + "type": "title", + "bbox": [ + 0.245, + 0.141, + 0.761, + 0.21 + ], + "angle": 0, + "content": "SpectraM-PS: Spectrally Multiplexed Photometric Stereo under Unknown Spectral Composition" + }, + { + "type": "text", + "bbox": [ + 0.359, + 0.235, + 0.644, + 0.251 + ], + "angle": 0, + "content": "Satoshi Ikehata\\(^{1,2}\\) and Yuta Asano\\(^{1}\\)" + }, + { + "type": "text", + "bbox": [ + 0.336, + 0.262, + 0.668, + 0.277 + ], + "angle": 0, + "content": "\\(^{1}\\) National Institute of Informatics, Tokyo, Japan" + }, + { + "type": "text", + "bbox": [ + 0.344, + 0.278, + 0.657, + 0.291 + ], + "angle": 0, + "content": "2 Tokyo Institute of Technology, Tokyo, Japan" + }, + { + "type": "text", + "bbox": [ + 0.261, + 0.33, + 0.744, + 0.566 + ], + "angle": 0, + "content": "Abstract. In this paper, we present a groundbreaking spectrally multiplexed photometric stereo approach for recovering surface normals of dynamic surfaces without the need for calibrated lighting or sensors, a notable advancement in the field traditionally hindered by stringent prerequisites and spectral ambiguity. By embracing spectral ambiguity as an advantage, our technique enables the generation of training data without specialized multispectral rendering frameworks. We introduce a unique, physics-free network architecture, SpectraM-PS, that effectively processes multiplexed images to determine surface normals across a wide range of conditions and material types, without relying on specific physically-based knowledge. Additionally, we establish the first benchmark dataset, SpectraM14, for spectrally multiplexed photometric stereo, facilitating comprehensive evaluations against existing calibrated methods. Our contributions significantly enhance the capabilities for dynamic surface recovery, particularly in uncalibrated setups, marking a pivotal step forward in the application of photometric stereo across various domains." + }, + { + "type": "text", + "bbox": [ + 0.261, + 0.579, + 0.741, + 0.607 + ], + "angle": 0, + "content": "Keywords: Spectrally Multiplexed Photometric Stereo \\(\\cdot\\) Dynamic Surface Recovery \\(\\cdot\\) Multispectral Photometric Stereo" + }, + { + "type": "title", + "bbox": [ + 0.216, + 0.657, + 0.377, + 0.673 + ], + "angle": 0, + "content": "1 Introduction" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.689, + 0.788, + 0.764 + ], + "angle": 0, + "content": "Recovering detailed normals of dynamic surfaces is essential for monitoring various processes: in manufacturing, it helps in tracking wear and tear of machine parts; in agriculture, it allows for the observation of crop growth through changes in leaf geometry; and in sports engineering, it aids in improving equipment design and safety by analyzing how surfaces deform upon impact." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.765, + 0.788, + 0.84 + ], + "angle": 0, + "content": "Photometric Stereo (PS) [57,64] derives object surface normals from observations under different lighting conditions at a fixed viewpoint. Despite decades of progress, the requirement for objects to stay stationary during lighting changes challenges the recovery of dynamic surfaces, essential for analyzing temporal surface deformations. PS researches have employed spectral multiplexing for dynamic" + } + ], + [ + { + "type": "page_number", + "bbox": [ + 0.218, + 0.116, + 0.23, + 0.127 + ], + "angle": 0, + "content": "2" + }, + { + "type": "header", + "bbox": [ + 0.272, + 0.115, + 0.437, + 0.128 + ], + "angle": 0, + "content": "S. Ikehata and Y. Asano" + }, + { + "type": "image", + "bbox": [ + 0.223, + 0.145, + 0.465, + 0.293 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.467, + 0.145, + 0.782, + 0.293 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.214, + 0.303, + 0.788, + 0.388 + ], + "angle": 0, + "content": "Fig. 1: (Left) Illustration of our SpectraM-PS. Our method recovers a surface normal map from a spectrally multiplexed image. The spectral/spatial composition for generating the observations is unknown. There is potential for a mismatch between the sensor's spectral sensitivity and the light source's spectral distribution, which may lead to crosstalk. (Right) By applying our method to individual frames of a video, the normal map of dynamic surfaces can be recovered." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.401, + 0.785, + 0.461 + ], + "angle": 0, + "content": "surface recovery [10,20,32,37,51,62]—a technique originally used in telecommunications and spectroscopy [31]. This technique utilizes the varying wavelengths of light to multiplex and subsequently demultiplex signals within a single sensor, thereby increasing the capacity for information transmission." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.462, + 0.788, + 0.583 + ], + "angle": 0, + "content": "Historically, spectrally multiplexed photometric stereo is often referred to as color photometric stereo [7,14,20,32,33], specifically when objects are illuminated with monochromatic red, green, and blue lights from various angles, captured in the camera's RGB channels. Each channel is then treated as an observation under a distinct lighting for PS analysis. This technique has been further extended to not only RGB but also any number of spectral bands and is specifically referred to as multispectral photometric stereo [17,18,46]. These techniques enable dynamic surface recovery by processing each temporal multi-channel frame separately." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.584, + 0.788, + 0.81 + ], + "angle": 0, + "content": "Despite their potential in dynamic surface recovery, current spectrally multiplexed photometric stereo methods face stringent prerequisites that limit their practicality. These include the necessity for precisely calibrated directional lighting in controlled environments [7, 14, 17, 20, 46] and sensors with aligned spectral sensitivities [32, 33]. Furthermore, they make strong assumptions about the surface, requiring it to be convex, integrable, Lambertian, and exhibit uniform chromaticity [20, 37]. By contrast, recent PS methods without spectral multiplexing support non-Lambertian surfaces [29, 55], spatially-varying materials [12, 23], and the use of uncalibrated lighting [11, 25, 26, 54]. This disparity arises from the challenge where identical observations are produced by different spectral compositions of light, surface and sensor [22, 50], a phenomenon absent in conventional PS due to constant spectral compositions of them across images. Recently, Guo et al. [17, 18] thoroughly explored how the spectral ambiguity renders spectrally multiplexed photometric stereo ill-posed, necessitating severely unpractical conditions on light, surface, and sensor to resolve the ambiguity." + }, + { + "type": "text", + "bbox": [ + 0.215, + 0.81, + 0.788, + 0.842 + ], + "angle": 0, + "content": "In this work, we propose a spectrally multiplexed photometric stereo method that recovers normals directly from multiplexed observations produced by \\( un \\)" + } + ], + [ + { + "type": "header", + "bbox": [ + 0.641, + 0.115, + 0.732, + 0.129 + ], + "angle": 0, + "content": "SpectraM-PS" + }, + { + "type": "page_number", + "bbox": [ + 0.775, + 0.117, + 0.787, + 0.127 + ], + "angle": 0, + "content": "3" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.146, + 0.788, + 0.344 + ], + "angle": 0, + "content": "known composition of lights, surface, and sensor (See Fig. 1-left), drawing inspiration from recent data-driven photometric stereo methods [25, 26]. While prior works [18, 46] have considered the spectral ambiguity harmful and something that must be resolved, we demonstrate that it can even be beneficial for a data-driven approach as it compacts the input space and allows for the generation of training data without a multispectral rendering framework. Trained on spectrally composed observations, our generic, physics-free architecture directly maps a single multiplexed image with an order-agnostic, arbitrary number of channels to object surface normals without the need for calibrating lights and sensors, and without imposing severe constraints on surface reflectance and geometry. By applying our method to individual frames of a video, dynamic surface recovery via spectrally multiplexed photometric stereo in uncalibrated, uncontrolled scenarios is achieved as illustrated in Fig. 1 (right)." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.344, + 0.788, + 0.479 + ], + "angle": 0, + "content": "While numerous benchmarks exist for conventional photometric stereo [52,56, 63], not a single benchmark is available for spectrally multiplexed PS. Therefore, we have created the first real benchmark dataset, namely SpectraM, for this task. For comparative evaluations with calibrated methods such as [18,46], we carefully calibrated directional light sources of different wavelengths, including their directions. We implemented five different difficulty settings by varying the type of light sources (RGB vs NIR) and whether individual light sources were illuminated independently or simultaneously, catering to both ideal conditions without channel crosstalk and more realistic conditions with channel crosstalk." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.48, + 0.788, + 0.661 + ], + "angle": 0, + "content": "Our contributions are summarized as follows: (1) We pioneer the use of spectrally multiplexed photometric stereo for recovering dynamic surfaces in uncalibrated setups, employing a data-driven approach to overcome spectral ambiguity, a significant barrier in prior work. (2) We introduce a unique, physics-free neural network, SpectraM-PS (Spectrally Multiplexed PS), that recovers surface normals from a spectrally multiplexed image, capable of handling images with any number of order-agnostic channels. (3) We demonstrate how spectral ambiguity restricts the input space for training data generation, offering a strategy for efficient dataset creation without the need for multispectral rendering. (4) We create the first evaluation benchmark, SpectraM, for this domain, showing our method's superiority over current calibrated spectrally multiplexed photometric stereo techniques." + }, + { + "type": "title", + "bbox": [ + 0.215, + 0.686, + 0.389, + 0.703 + ], + "angle": 0, + "content": "2 Related Work" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.719, + 0.788, + 0.842 + ], + "angle": 0, + "content": "Temporally Multiplexed Photometric Stereo (Conventional). From a communication perspective, conventional photometric stereo, as originally proposed by Woodham [64], employs a time multiplexing strategy to recover static surfaces. This method involves temporally varying lighting conditions while capturing images from a fixed viewpoint. Since the same light sources and sensor always provide observations, image differences stem solely from changes in light direction and intensity. This approach simplifies addressing complex conditions such as cast shadows [27, 28], non-Lambertian surfaces [16, 29], non-convex sur" + } + ], + [ + { + "type": "page_number", + "bbox": [ + 0.218, + 0.116, + 0.23, + 0.127 + ], + "angle": 0, + "content": "4" + }, + { + "type": "header", + "bbox": [ + 0.272, + 0.115, + 0.437, + 0.128 + ], + "angle": 0, + "content": "S. Ikehata and Y. Asano" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.147, + 0.788, + 0.314 + ], + "angle": 0, + "content": "faces [23], and uncalibrated lighting [11, 36, 54]. Recently, learning-based methods [11-13, 23-25, 34, 36, 39-42, 44, 53, 58, 59, 68, 69] have emerged as an effective alternative, addressing challenges faced by traditional, physics-based approaches [6, 8, 16, 19, 21, 28, 45, 65]. These data-driven methods regress normal maps from observations utilizing techniques such as observation map regression [23, 30], set pooling [12, 13], graph neural networks [68], Transformer [24], and neural rendering for inverse rendering optimization [40, 41, 58]. Notably, the introduction of universal photometric stereo methods [25, 26] has enabled the handling of unknown, spatially-varying lighting in a purely data-driven framework. Inspired by these advancements, our work aims to regress normals from observations under unknown light, surface, and sensor conditions." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.316, + 0.788, + 0.376 + ], + "angle": 0, + "content": "Spectrally Multiplexed Photometric Stereo. Despite its potential for dynamic surface recovery, spectrally multiplexed photometric stereo [15, 37] has remained less explored than its mainstream counterpart, primarily due to notable limitations." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.377, + 0.788, + 0.468 + ], + "angle": 0, + "content": "Lighting Constraints: Existing methods necessitate multiple directional lights in controlled settings, contrasting the flexibility of temporally multiplexed techniques that adapt to diverse lighting conditions [26, 47, 49]. They typically require pre-calibrated light directions and distinct light source spectra to prevent channel crosstalk. In contrast, our approach accommodates uncontrolled lighting scenarios without the need for predefined or calibrated setups." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.468, + 0.788, + 0.573 + ], + "angle": 0, + "content": "Surface and Sensor Constraints: Prior works assume significant limitations on surfaces, such as Lambertian, convex, and uniform properties [15,20,37]. Recent advances like Lv et al. [46] extend to non-Lambertian surfaces but still require uniform materials. Sensor requirements typically involve narrow-band spectral responses and a fixed number of channels, limiting flexibility. Our approach leverages a data-driven model, training neural networks on synthetic data to handle complex surfaces and varied spectral sensor responses." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.574, + 0.788, + 0.74 + ], + "angle": 0, + "content": "Data-driven Methods: To our knowledge, there are few data-driven methods for this task [32, 33, 46]. Previous studies, such as those by Ju et al. [32, 33], require identical spectral and spatial lighting conditions during both training and testing, greatly restricting their practicality. ELIE-Net [46] permits variability in training and test setups; however, strong assumptions on both light and surfaces prohibitively limit its applications. Our model, on the other hand, eschews explicit lighting models in favor of learning direct input-output relationships, allowing for accurate predictions under varied and unknown spectral compositions and supporting materials with spatially diverse properties. Furthermore, unlike ELIE-Net's reliance on spectral BRDF datasets, our training approach utilizes assets akin to those employed in conventional photometric stereo." + }, + { + "type": "title", + "bbox": [ + 0.215, + 0.763, + 0.442, + 0.779 + ], + "angle": 0, + "content": "3 Problem Statement" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.794, + 0.788, + 0.84 + ], + "angle": 0, + "content": "Given a single image \\(I \\in \\mathbb{R}^{h \\times w \\times k}\\) captured by a static \\(k\\)-channel orthographic sensor, along with an optional object mask \\(M \\in \\mathbb{R}^{h \\times w}\\), the objective of spectrally multiplexed photometric stereo is to recover the surface normals of the" + } + ], + [ + { + "type": "header", + "bbox": [ + 0.641, + 0.115, + 0.732, + 0.129 + ], + "angle": 0, + "content": "SpectraM-PS" + }, + { + "type": "page_number", + "bbox": [ + 0.775, + 0.117, + 0.785, + 0.127 + ], + "angle": 0, + "content": "5" + }, + { + "type": "image", + "bbox": [ + 0.231, + 0.145, + 0.782, + 0.334 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.214, + 0.345, + 0.788, + 0.415 + ], + "angle": 0, + "content": "Fig. 2: SpectraM-PS involves decomposing a spectrally multiplexed image into independent channels. The Global Feature Encoder extracts a feature map from each channel. The surface vector is then recovered by the Dual-scale Surface Normal Decoder at each pixel. We adopt a dual-scale approach to preserve the entire shape, while employing patch-embedding techniques to enhance local surface details." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.439, + 0.788, + 0.577 + ], + "angle": 0, + "content": "object, \\( N \\in \\mathbb{R}^{h \\times w \\times 33} \\). The object is supposed to be illuminated by multiple light sources, each with unique spatial and spectral properties. Previous studies have typically assumed an equal number of light sources and sensor channels, with each light source's wavelength precisely matching the spectral response of a single channel, thereby precluding any channel crosstalk, and with the directions of lights predetermined. In contrast, we do not presuppose the spatial distribution of illumination nor require the spectrum of each light source to be exclusively aligned with the spectral responses of the sensor channels, thus permitting channel crosstalk. This distinction is elaborated in subsequent sections." + }, + { + "type": "title", + "bbox": [ + 0.216, + 0.601, + 0.33, + 0.617 + ], + "angle": 0, + "content": "4 Method" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.635, + 0.787, + 0.695 + ], + "angle": 0, + "content": "We propose and tackle the problem of spectrally multiplexed photometric stereo from a single image with multiple channels, produced by an unknown spectral/spatial composition of the sensor, light, and surface. To build such a method, we train neural networks to directly infer the normal map from an image." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.696, + 0.788, + 0.803 + ], + "angle": 0, + "content": "Our method addresses two challenges: (1) a physics-free architecture that accepts a varying number of spectral channels and is agnostic to their order, and (2) an effective approximation of the spectrally multiplexed image for efficient training. We consider (2) to be of significant importance, yet it remains largely unexplored. Synthesizing spectrally multiplexed images in a physically accurate manner is prohibitively challenging, owing to the increased complexity of their parameter spaces and the scarcity of 3D assets with detailed spectral properties," + }, + { + "type": "page_footnote", + "bbox": [ + 0.218, + 0.811, + 0.787, + 0.84 + ], + "angle": 0, + "content": "3 It should be noted that unlike conventional PS, reflectance recovery generally falls outside the scope of spectrally multiplexed PS due to its inherently ill-posed nature." + } + ], + [ + { + "type": "page_number", + "bbox": [ + 0.218, + 0.116, + 0.23, + 0.127 + ], + "angle": 0, + "content": "6" + }, + { + "type": "header", + "bbox": [ + 0.272, + 0.115, + 0.437, + 0.128 + ], + "angle": 0, + "content": "S. Ikehata and Y. Asano" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.147, + 0.788, + 0.207 + ], + "angle": 0, + "content": "as well as the complex nature of light-surface interactions across different wavelengths. To address this issue, developing an efficient approximation method for rendering spectrally multiplexed images using common RGB image rendering techniques is crucial." + }, + { + "type": "title", + "bbox": [ + 0.215, + 0.235, + 0.786, + 0.251 + ], + "angle": 0, + "content": "4.1 Physics-free Spectrally Multiplexed PS Network (SpectraM-PS)" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.265, + 0.788, + 0.37 + ], + "angle": 0, + "content": "The architecture of SpectraM-PS is illustrated in Fig. 2. Drawing inspiration from established Transformer-based photometric stereo networks [24-26], we integrate an encoder to first extract the global features and a decoder to estimate per-pixel surface normals. The architecture derives normals solely from the input image and mask, without prior light information. This indicates that the architecture focuses the network's learning objective on the relationship between input and output without relying on physics-based principles, unlike prior works." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.372, + 0.788, + 0.583 + ], + "angle": 0, + "content": "In our model, all the interactions among features from different sensor channels are employed by naïve Transformer [61] in similar to [24-26]. Transformer functions by mapping input features to query, key, and value vectors of equal dimensions. These vectors are processed through a multi-head self-attention mechanism, utilizing a softmax layer, followed by a feed-forward network comprising two linear layers. Both the input and output layers maintain identical dimensionality, with the inner layer having twice the dimension of the input. Each layer is surrounded by a residual connection, succeeded by layer normalization [67]. The advantage of employing Transformers in photometric stereo networks lies in their capability to facilitate complex interactions among intermediate features, a task unachievable with simple operations like pooling [11-13, 46] and observation map [23, 30]. Additionally, the token-based attention mechanism allows for different number of input tokens (i.e., sensor channels) between training and test phases and ensures that the results are independent on the order of tokens." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.584, + 0.788, + 0.842 + ], + "angle": 0, + "content": "Building on the established Transformer-based architecture [26] for temporally multiplexed photometric stereo, we extend its scope to a spectrally multiplexed one. To accommodate a variable number of channels and eliminate dependency on their order, an input spectrally multiplexed image is first split into individual channels, each of which is concatenated with an object mask (If no mask is provided, replace with a matrix of ones.) and then input into the same encoder of a neural network. This approach is distinctly different from traditional methods that encode an input image as it is in neural networks [32, 33]. Then, at Preprocessing, we normalize each channel by dividing it by a random value between its maximum and mean. Each channel and mask are resized or cropped to a resolution \\((c \\times c)\\) that is a multiple of 32 to be input into the multi-scale encoder. Global Feature Encoder first applies a backbone network (i.e., ConvNeXt-T [43]) to individually encode the concatenation of each channel and mask, then uses Transformer layers for channel-axis (i.e., sensor channel) feature communication across scales (the number of Transformer layers is \\{0, 1, 2, 4\\} at \\{1/4, 1/8, 1/16, 1/32\\} scales, hidden dimensions are same with input dimensions), and finally, a feature pyramid network [66] for integrating features" + } + ], + [ + { + "type": "header", + "bbox": [ + 0.64, + 0.114, + 0.732, + 0.129 + ], + "angle": 0, + "content": "SpectraM-PS" + }, + { + "type": "page_number", + "bbox": [ + 0.775, + 0.116, + 0.787, + 0.127 + ], + "angle": 0, + "content": "7" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.147, + 0.784, + 0.176 + ], + "angle": 0, + "content": "at different levels. Note that the design of encoder is almost the same as [25, 26], except that images are replaced by sensor channels, so details are omitted." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.177, + 0.788, + 0.524 + ], + "angle": 0, + "content": "Given global features \\(\\in \\mathcal{R}^{k\\times c / 4\\times c / 4\\times 256}\\), our novel Dual-scale Surface Normal Decoder adopts a dual-scale strategy for predicting point-wise surface normals at \\(m\\) (i.e., 2048) sampled locations at the original resolution within the object mask. The first branch recovers low-frequency surface normals at the feature map resolution \\((\\frac{c}{4}\\times \\frac{c}{4})\\). Concretely, all global features corresponding to each sample location are processed by five channel-axis Transformer layers (with a 256 hidden dimension) and are pooled via Pooling-by-Multihead-Attention (PMA) [38] using an additional channel-axis Transformer layer (with a 384 hidden dimension). To enhance spatial communication, two spatial-axis Transformer layers (with a 384 hidden dimension) inspired by Ikehata [26] are employed (i.e., Transformer is employed among samples at different locations), with a final MLP \\((384\\to 192\\to 3)\\) predicting the low-frequency normals at sampled locations. The second branch focuses on high-resolution normal recovery, using patch embedding for local context at the same \\(m\\) locations, with \\(w\\times w\\) patches (\\(w = 21\\)) processed by an MLP (with a 256 hidden dimension) and two layer norms. These patches, concatenated with bilinearly interpolated global features, pass through five channel-axis Transformer blocks (with a 256 hidden dimension), PMA (with a 384 hidden dimension), and are merged with the first branch output normals into 387-dimensional vectors. Two additional spatial-axis Transformer layers (with a 384 hidden dimension) enable non-local interactions, culminating in a final MLP \\((384\\to 192\\to 3)\\) for high-resolution normals, normalized to unit vectors. The complete normal map is formed by merging all the vectors from different sample sets." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.525, + 0.788, + 0.706 + ], + "angle": 0, + "content": "It should be noted that while SDM-UniPS [26] targets temporally multiplexed PS with tens of images, and its decoder performs normal estimation purely on a pixel basis. In contrast, spectrally multiplexed PS deals with fewer channels (e.g., three with RGB sensors), making a pixel-basis architecture less effective. Therefore, we use patch embedding at the patch-basis decoder to capture fine details with a dual-scale architecture for preserving overall shape. Without a dual-scale design, the recovery of surface normals becomes overly influenced by local image textures captured through patch embedding. This leads to a failure in preserving the entire shape, resulting in a significant reduction in accuracy. Our motivation is supported by Fig. 3 (left), where SDM-UniPS [26] fails to recover fine details with six temporally multiplexed images, while our architecture produces a more plausible normal map." + }, + { + "type": "title", + "bbox": [ + 0.215, + 0.726, + 0.725, + 0.743 + ], + "angle": 0, + "content": "4.2 Efficient Training Strategy Utilizing Spectral Ambiguity" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.75, + 0.788, + 0.841 + ], + "angle": 0, + "content": "Aligning the training and test data domains in neural networks is essential for optimal model performance [9, 60]. However, rendering spectrally multiplexed data poses challenges due to the scarcity of multispectral Bidirectional Reflectance Distribution Functions (BRDFs). In reality, ELIE-Net [46] was trained using only 51 measured isotropic spectral BRDFs. On the other hand, given the availability of various large isotropic BRDF databases [1-3, 5, 48], we seek to explore" + } + ], + [ + { + "type": "page_number", + "bbox": [ + 0.218, + 0.116, + 0.23, + 0.127 + ], + "angle": 0, + "content": "8" + }, + { + "type": "header", + "bbox": [ + 0.272, + 0.115, + 0.437, + 0.128 + ], + "angle": 0, + "content": "S. Ikehata and Y. Asano" + }, + { + "type": "image", + "bbox": [ + 0.223, + 0.145, + 0.443, + 0.315 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.261, + 0.316, + 0.513, + 0.329 + ], + "angle": 0, + "content": "SpectraM-PS on Temporally Multiplexed PS" + }, + { + "type": "image", + "bbox": [ + 0.443, + 0.145, + 0.555, + 0.23 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.443, + 0.231, + 0.555, + 0.315 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.559, + 0.144, + 0.669, + 0.231 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.559, + 0.232, + 0.669, + 0.316 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.67, + 0.145, + 0.782, + 0.231 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.67, + 0.232, + 0.782, + 0.316 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.589, + 0.317, + 0.762, + 0.329 + ], + "angle": 0, + "content": "PS-Multiplex Training Dataset" + }, + { + "type": "image_caption", + "bbox": [ + 0.215, + 0.341, + 0.788, + 0.385 + ], + "angle": 0, + "content": "Fig. 3: (Left) Comparison of SpectraM-PS and SDM-UniPS [26] on six temporally multiplexed PS images. Due to the patch-wise basis of SpectraM-PS, fine details are better recovered. (Right) Illustration of different lighting conditions in PS-Multiplex." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.399, + 0.784, + 0.459 + ], + "angle": 0, + "content": "utilizing these datasets for training our model, leveraging the fact that our network does not distinguish images based on their physically-based principles. In this section, we highlight how RGB images serve as a practical approximation, simplifying the complexity inherent in multispectral imaging." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.46, + 0.785, + 0.505 + ], + "angle": 0, + "content": "We begin the discussion by characterizing multispectral imaging. Assuming that the surface doesn't emit light and only reflections on surface are considered, the image formation model is described as follows [35]:" + }, + { + "type": "equation", + "bbox": [ + 0.3, + 0.513, + 0.786, + 0.547 + ], + "angle": 0, + "content": "\\[\nI _ {(s, p)} = \\int_ {\\Omega} \\left(\\omega_ {i} ^ {\\top} n _ {p}\\right) \\int_ {0} ^ {\\infty} S _ {s} (\\lambda) f _ {p} \\left(\\omega_ {i}, \\omega_ {o}, \\lambda\\right) L _ {p} \\left(\\omega_ {i}, \\lambda\\right) \\mathrm {d} \\lambda \\mathrm {d} \\omega_ {i}. \\tag {1}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.553, + 0.788, + 0.704 + ], + "angle": 0, + "content": "In this equation, \\( I_{(s,p)} \\) denotes the incoming spectral radiance at the sensor \\( s \\) (or \\( s \\)-th channel) from a surface point \\( p \\). The term \\( f_p \\) represents BRDF, \\( L_p \\) the incident light intensity at the surface point, and \\( \\lambda \\) the wavelength of the incident light. The symbols \\( \\omega_i \\) and \\( \\omega_o \\) denote the directions of incident and reflected light, respectively. \\( S_s(\\lambda) \\) refers to the spectral sensitivity of the sensor \\( s \\) at wavelength \\( \\lambda \\), \\( n_p \\) is the surface normal, and \\( \\varOmega \\) represents the hemisphere over which incident light directions are possible. The integral sums over all incident directions and wavelengths. It is important to note that the incident light intensity \\( L_p \\) depends not only on the direct contribution from light sources but also on the visibility of light (e.g., attached and cast shadows) and indirect illuminations." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.705, + 0.789, + 0.843 + ], + "angle": 0, + "content": "Eq. (1) illustrates the concept of spectral ambiguity, showing that an infinite number of combinations of \\( S_{s}(\\lambda) \\), \\( f_{p}(\\omega_{i},\\omega_{o},\\lambda) \\), and \\( L_{p}(\\omega_{i},\\lambda) \\) can result in the same spectral radiance, including narrowband compositions. In other words, with spectral ambiguity, a single observation \\( I_{(s,p)} \\) can encompass the observations for all spectral compositions that satisfy the equation (i.e., metamerism [22, 50]). This perspective justifies the theory of substituting multispectral images, which possess a broad parameter space, with narrowband RGB images. It is worth mentioning that channel crosstalk primarily affects the incident light intensity, consequently distorting the product of \\( S_{s}(\\lambda)\\cdot f_{p}(\\omega_{i},\\omega_{o},\\lambda)\\cdot L_{p}(\\omega_{i},\\lambda) \\) in Eq. (1)." + } + ], + [ + { + "type": "header", + "bbox": [ + 0.641, + 0.115, + 0.732, + 0.129 + ], + "angle": 0, + "content": "SpectraM-PS" + }, + { + "type": "page_number", + "bbox": [ + 0.775, + 0.117, + 0.787, + 0.127 + ], + "angle": 0, + "content": "9" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.146, + 0.788, + 0.451 + ], + "angle": 0, + "content": "This implies that observations influenced by spectral crosstalk can still be equivalently represented using a narrowband setup under spectral ambiguity. In the experiments, we demonstrate that our model, trained on three narrowband observations can be applied to multiplexed data with channel crosstalk. To realize this approximation, we rendered a large number of three-channel narrowband images using the path-tracing algorithm in Blender [4], where up to 10-bounce reflections are permitted, based on common 3D assets [2] for RGB rendering. Following the rendering pipeline described in [26], we rendered objects by combining three different lighting models: directional, point, and environmental (five combinatorial settings in total as shown in Fig. 3). To simulate spectrally multiplexed images, we defined R, G, and B light sources and illuminated the surface in a multiplexed manner. It is important to note that the rendered RGB images are decomposed into three grayscale images, each of which was independently fed into the network; therefore, any wavelength-dependent information is masked. For material diversity, we adopted the method from [26], categorizing 897 AdobeStock texture maps into three groups: 421 diffuse, 219 specular, and 257 metallic textures. Four objects from a set of 410 3D AdobeStock models were randomly selected and textured with these materials. This structured approach led to the rendering of 106,374 multiplexed images along with their ground truth surface normal maps, forming the 'PS-Multiplex' dataset." + }, + { + "type": "title", + "bbox": [ + 0.216, + 0.47, + 0.567, + 0.487 + ], + "angle": 0, + "content": "5 SpectraM14 Benchmark Dataset" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.5, + 0.788, + 0.59 + ], + "angle": 0, + "content": "Due to the lack of a benchmark for spectrally multiplexed PS, the first comprehensive evaluation dataset, named SpectraM14, is created. This dataset includes 14 objects, each exhibiting a range of optical properties such as monochromatic or multicolored appearances and diffuse or specular reflections, as depicted in Fig. 4. Our benchmark encompasses tasks under five distinct conditions, as described later." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.595, + 0.788, + 0.793 + ], + "angle": 0, + "content": "Imaging Setup. To acquire our dataset, we utilized a color camera (FLIR GS3-U3-123S6C-C) and an NIR camera (FLIR GS3-U3-41C6NIR-C), both equipped with a \\(50\\mathrm{mm}\\) lens. For the NIR camera, we used narrowband filters with wavelengths of \\(750\\mathrm{nm}\\), \\(850\\mathrm{nm}\\), \\(880\\mathrm{nm}\\), \\(905\\mathrm{nm}\\), and \\(940\\mathrm{nm}\\), and the acquired images were manually merged. Objects were placed \\(0.8\\mathrm{m}\\) from the camera to approximate orthographic projection. Following conventional PS benchmarks [52,56,63], data capture occurred in a controlled, dark environment with the scene draped in black cloth to mitigate interreflection. The camera's ISO sensitivity was minimized to enhance image quality. The imaging area was further isolated using low-reflectance cloths to suppress inter-reflection. For each illumination condition, we collected six images under varying exposures to produce HDR input images. For the evaluations throughout this paper, the images are cropped using an object mask and resized to \\(512\\mathrm{px} \\times 512\\mathrm{px}\\)." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.795, + 0.788, + 0.842 + ], + "angle": 0, + "content": "Lighting. Six LED and three halogen light sources, positioned roughly 1 meter from the object, provided illumination. We used the \"Weeylite S05 RGB Pocket Lamp\" and the \"NPI PIS-UHX-AIR\" for lighting. This setup enabled the use of" + } + ], + [ + { + "type": "page_number", + "bbox": [ + 0.218, + 0.116, + 0.236, + 0.127 + ], + "angle": 0, + "content": "10" + }, + { + "type": "header", + "bbox": [ + 0.272, + 0.115, + 0.437, + 0.128 + ], + "angle": 0, + "content": "S. Ikehata and Y. Asano" + }, + { + "type": "image", + "bbox": [ + 0.223, + 0.145, + 0.782, + 0.324 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.395, + 0.328, + 0.607, + 0.343 + ], + "angle": 0, + "content": "Fig. 4: Objects in SpectraM14." + }, + { + "type": "image", + "bbox": [ + 0.226, + 0.358, + 0.777, + 0.445 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.328, + 0.451, + 0.673, + 0.466 + ], + "angle": 0, + "content": "Fig. 5: Illustration of six conditions in SpectraM14." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.498, + 0.785, + 0.529 + ], + "angle": 0, + "content": "red, green, blue, yellow, magenta, cyan, and NIR lighting, with spectra validated using a Hamamatsu Photonics Multichannel Analyzer C10027-01." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.533, + 0.788, + 0.608 + ], + "angle": 0, + "content": "Calibration and Ground Truth Data. We measured the directions of lights using specular reflections from a mirror sphere. Light intensity was standardized across the visible spectrum by averaging RGB values from reflected light on a white target. The ground truth normals were captured with a SHINING 3D EinScan-SE scanner." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.614, + 0.789, + 0.842 + ], + "angle": 0, + "content": "Evaluation Procedure. The design philosophy of this benchmark is to assess the robustness and adaptability of spectrally multiplexed PS methods under realistic lighting conditions, accounting for variations in channel numbers and the presence of spectral crosstalk. For a comprehensive evaluation, we designed tasks under five distinct conditions as shown in Fig. 5: Condition 1: Color sensor, no crosstalk condition: Six colors of light (red, green, blue, cyan, yellow, magenta) were each independently illuminated and observed with an RGB sensor. Afterward, the channels of RGB were averaged. Condition 2: Color sensor, weak crosstalk condition: Three colors of light (red, green, blue) were simultaneously illuminated and observed through each channel of the RGB sensor. Condition 3: Color sensor, strong crosstalk condition: Three colors of light (cyan, yellow, magenta) were simultaneously illuminated and observed through each channel of the RGB sensor. Condition 4: NIR sensor, no crosstalk condition: Light at wavelengths of \\(750~\\mathrm{nm}\\), \\(850~\\mathrm{nm}\\), \\(880~\\mathrm{nm}\\), \\(905~\\mathrm{nm}\\), and \\(940~\\mathrm{nm}\\) were each independently illuminated and observed with a monochrome sensor corresponding to" + } + ], + [ + { + "type": "header", + "bbox": [ + 0.641, + 0.115, + 0.732, + 0.129 + ], + "angle": 0, + "content": "SpectraM-PS" + }, + { + "type": "page_number", + "bbox": [ + 0.769, + 0.117, + 0.784, + 0.127 + ], + "angle": 0, + "content": "11" + }, + { + "type": "text", + "bbox": [ + 0.219, + 0.147, + 0.785, + 0.207 + ], + "angle": 0, + "content": "each wavelength. Condition 5: NIR sensor, spatially-varying lighting condition: New images were created by averaging two images taken under the conditions mentioned above. The combinations were (750 nm, 850 nm), (850 nm, 880 nm), (880 nm, 905 nm), (905 nm, 940 nm), and (940 nm, 750 nm)." + }, + { + "type": "title", + "bbox": [ + 0.219, + 0.234, + 0.364, + 0.251 + ], + "angle": 0, + "content": "6 Experiment" + }, + { + "type": "text", + "bbox": [ + 0.219, + 0.268, + 0.785, + 0.419 + ], + "angle": 0, + "content": "In this section, we evaluate our method on our SpectraM14. Our method is compared with one SOTA optimization-based method [17] and one SOTA learning-based method [46]. The former introduces a closed-form solution for spectrally multiplexed photometric stereo applied to monochromatic surfaces with spatially varying (SV) albedo. The latter presents a Spectral Reflectance Decomposition (SRD) model, which disentangles spectral reflectance into geometric and spectral components for surface normal recovery under non-Lambertian spectral reflectance conditions. Unlike the compared methods, our approach does not assume a specific lighting setup, whereas both methods presume the presence of calibrated single directional light sources." + }, + { + "type": "text", + "bbox": [ + 0.219, + 0.424, + 0.785, + 0.59 + ], + "angle": 0, + "content": "Training details. SpectraM-PS was trained from scratch on the PS-Multiplex dataset until convergence using the AdamW optimizer, with a step decay learning rate schedule that reduced the learning rate by a factor of 0.8 every ten epochs. We applied learning rate warmup during the first epoch and used a batch size of 16, an initial learning rate of 0.0001, and a weight decay of 0.05. Each batch consisted of three input training multiplexed images with three channels each. The training loss was computed using the Mean Squared Error (MSE) loss function to measure \\(\\ell_2\\) errors between the predicted surface normal vectors and the ground truth surface normal vectors. We measured the reconstruction accuracy of our method by computing the mean angular errors (MAE) between the predicted and true surface normal maps, expressed in degrees." + }, + { + "type": "text", + "bbox": [ + 0.219, + 0.594, + 0.785, + 0.7 + ], + "angle": 0, + "content": "Computational Cost. The inference time of PS methods varies with the number of pixels and channels in the input image. For Condition 2 and 3 with a \\(512 \\times 512 \\times 3\\) image, the mean and standard deviation of inference times (in sec) over 14 objects in SpectraM14 benchmark were: our method (3.42/0.85), Lv et al. [46] (0.46/0.24) and Guo et al. [17] (2.38/1.10). Our architecture leads to higher computational costs; however, none of the methods were suitable for real-time processing (e.g., 15 fps requires 0.06 sec/frame)." + }, + { + "type": "text", + "bbox": [ + 0.219, + 0.704, + 0.785, + 0.84 + ], + "angle": 0, + "content": "Ablation Study. We firstly validate the individual technical contributions of our training dataset (i.e., PS-Multiplex) and the physics-free architecture (i.e., SpectraM-PS) using a synthetic evaluation dataset. Firstly, we validate the efficacy of our training dataset, PS-Multiplex, by adapting an existing universal photometric stereo architecture designed for the conventional task (i.e., SDM-UniPS [26]) to the spectrally multiplexed photometric stereo task. Since both ours and SDM-UniPS take multiple observations and an object mask as input, this adaptation straightforwardly involves training the model on PS-Multiplex by treating each channel of an image as an individual image. Subsequently, we" + } + ], + [ + { + "type": "page_number", + "bbox": [ + 0.218, + 0.116, + 0.236, + 0.127 + ], + "angle": 0, + "content": "12" + }, + { + "type": "header", + "bbox": [ + 0.272, + 0.115, + 0.437, + 0.128 + ], + "angle": 0, + "content": "S. Ikehata and Y. Asano" + }, + { + "type": "table_caption", + "bbox": [ + 0.223, + 0.145, + 0.777, + 0.16 + ], + "angle": 0, + "content": "Table 1: Ablation analysis of the contributions of SpectraM-PS and PS-Multiplex." + }, + { + "type": "table", + "bbox": [ + 0.231, + 0.165, + 0.776, + 0.244 + ], + "angle": 0, + "content": "
MAE (Uniform)MAE (Piece-wise uniform)MAE (Non-uniform)
MethodNon-LambertianLambertianNon-LambertianLambertianNon-LambertianLambertian
SDM-UniPS [26]12.9 (4.5)12.4 (4.7)15.0 (5.0)21.7 (7.0)14.4 (5.1)15.2 (5.9)
[26] trained on PS-Multiplex11.1 (3.6)11.0 (3.9)10.5 (2.9)12.3 (3.9)10.6 (3.9)11.2 (3.7)
SpectraM-PS (Ours)8.0 (2.7)8.4 (2.7)7.9 (2.4)8.9 (2.9)8.2 (3.2)8.0 (2.5)
" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.259, + 0.784, + 0.289 + ], + "angle": 0, + "content": "compare this model against our proposed SpectraM-PS to demonstrate the efficacy of our dual-scale design with local patch embedding." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.291, + 0.788, + 0.579 + ], + "angle": 0, + "content": "For evaluating the contribution of our architecture (SpectraM-PS) and training dataset (PS-Multiplex), we additionally rendered three-channel spectrally multiplexed images representing six distinct surface material categories: (a) uniform, Lambertian; (b) piece-wise uniform, Lambertian; (c) non-uniform, Lambertian; (d) uniform, non-Lambertian; (e) piece-wise uniform, non-Lambertian; and (f) non-uniform, non-Lambertian. In uniform materials, every point on the surface within a scene exhibits the same material properties. For piece-wise uniform materials, each object in a scene is composed of the same material, yet different objects possess distinct materials. Non-uniform materials feature unique PBR textures assigned to each object. The rendering process for these images was identical to that used for the PS-Multiplex datasets in each category. We generated 100 scenes for each surface material category, and MAEs (stds) are averaged over them. The results are presented in Tab.1. In summary, SDM-UniPS [26] trained on our PS-Multiplex dataset demonstrates proper adaptation to the spectrally multiplexed photometric stereo task. Nonetheless, our SpectraM-PS method significantly enhanced reconstruction accuracy, showcasing an architecture-level improvement over SDM-UniPS for the spectrally multiplexed photometric stereo task, where the number of input channels is typically much fewer than that of input images for conventional PS." + }, + { + "type": "title", + "bbox": [ + 0.214, + 0.584, + 0.804, + 0.598 + ], + "angle": 0, + "content": "Comparative Evaluation on SpectraM14. The results are illustrated in Tabs. 2" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.599, + 0.788, + 0.841 + ], + "angle": 0, + "content": "to 6 and Fig. 6. Despite the fact that all existing spectrally multiplexed photometric stereo methods assume calibrated light sources and known directional light source conditions, our proposed method significantly outperformed them. This is because most of the real objects used in our experiment are neither Lambertian nor convex, and do not conform to their assumptions. However, our non-physical-based method successfully restored the normals very stably for these objects. Furthermore, our proposed method enabled robust reconstruction for all objects, despite having been trained only with RGB color images. This result supports the efficacy of our approximation. Furthermore, unlike existing methods that suffer from reduced estimation accuracy with increasing spectral crosstalk, our approach demonstrates only minimal performance degradation. Remarkably, our method excels in recovering a more realistic structure with spatially-varying surface materials. This breakthrough implies that our network can effectively achieve dynamic surface reconstruction across video frames in a universal setting. We will detail this groundbreaking application in the next section. Due to space constraints, not all results can be included here. However, all" + } + ], + [ + { + "type": "header", + "bbox": [ + 0.641, + 0.115, + 0.732, + 0.129 + ], + "angle": 0, + "content": "SpectraM-PS" + }, + { + "type": "page_number", + "bbox": [ + 0.769, + 0.116, + 0.786, + 0.127 + ], + "angle": 0, + "content": "13" + }, + { + "type": "table_caption", + "bbox": [ + 0.22, + 0.145, + 0.782, + 0.159 + ], + "angle": 0, + "content": "Table 2: Comparison in condition 1. The values are mean angular errors in degrees." + }, + { + "type": "table", + "bbox": [ + 0.222, + 0.161, + 0.783, + 0.217 + ], + "angle": 0, + "content": "
MethodObject ID
1234567891011121314Ave.
Ours8.611.213.07.312.26.111.55.05.77.510.35.16.112.28.9
Lv et al. [46]20.417.021.113.923.110.721.216.610.915.919.016.413.218.617.1
Guo et al. [17]22.615.220.713.427.17.231.324.88.018.224.511.410.229.318.9
" + }, + { + "type": "table_caption", + "bbox": [ + 0.221, + 0.229, + 0.782, + 0.242 + ], + "angle": 0, + "content": "Table 3: Comparison in condition 2. The values are mean angular errors in degrees." + }, + { + "type": "table", + "bbox": [ + 0.22, + 0.242, + 0.785, + 0.302 + ], + "angle": 0, + "content": "
MethodObject IDAve.
1234567891011121314
Ours10.111.513.59.012.87.010.95.56.39.312.75.39.814.710.0
Lv et al. [46]22.826.027.019.430.319.522.118.914.319.823.520.819.021.421.7
Guo et al. [17]31.127.529.220.138.019.033.423.513.626.632.714.717.139.325.7
" + }, + { + "type": "table_caption", + "bbox": [ + 0.221, + 0.313, + 0.782, + 0.326 + ], + "angle": 0, + "content": "Table 4: Comparison in condition 3. The values are mean angular errors in degrees." + }, + { + "type": "table", + "bbox": [ + 0.22, + 0.326, + 0.785, + 0.386 + ], + "angle": 0, + "content": "
MethodObject ID
1234567891011121314
Ours12.012.815.811.016.55.912.39.86.99.814.66.37.520.1
Lv et al. [46]38.534.138.232.440.138.236.629.638.235.138.236.638.430.9
Guo et al. [17]46.042.656.337.657.145.676.048.029.962.149.850.252.873.7
" + }, + { + "type": "table_caption", + "bbox": [ + 0.221, + 0.397, + 0.782, + 0.409 + ], + "angle": 0, + "content": "Table 5: Comparison in condition 4. The values are mean angular errors in degrees." + }, + { + "type": "table", + "bbox": [ + 0.222, + 0.41, + 0.782, + 0.469 + ], + "angle": 0, + "content": "
MethodObject ID
1234567891011121314
Ours10.911.39.76.310.85.312.44.37.86.69.14.07.410.6
Lv et al. [46]24.119.820.712.019.812.415.118.711.417.121.217.519.116.7
Guo et al. [17]30.716.018.39.929.68.423.125.710.214.624.513.929.613.8
" + }, + { + "type": "table_caption", + "bbox": [ + 0.221, + 0.48, + 0.782, + 0.493 + ], + "angle": 0, + "content": "Table 6: Comparison in condition 5. The values are mean angular errors in degrees." + }, + { + "type": "table", + "bbox": [ + 0.223, + 0.494, + 0.782, + 0.553 + ], + "angle": 0, + "content": "
MethodObject IDAve.
1234567891011121314
Ours10.911.09.97.311.04.713.34.87.96.59.84.27.410.48.6
Lv et al. [46]29.826.228.922.828.925.821.924.618.724.727.624.822.327.625.0
Guo et al. [17]40.027.329.223.233.325.029.330.120.026.731.626.325.027.227.8
" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.57, + 0.788, + 0.631 + ], + "angle": 0, + "content": "results are comprehensively presented in the supplementary materials. Additionally, the supplementary materials evaluate the impact of the spatial distribution of light sources on the performance of the proposed method. We also offer an in-depth discussion of each experimental condition therein." + }, + { + "type": "title", + "bbox": [ + 0.216, + 0.656, + 0.36, + 0.672 + ], + "angle": 0, + "content": "7 Conclusion" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.689, + 0.789, + 0.842 + ], + "angle": 0, + "content": "In this work, we introduce an innovative approach to spectrally multiplexed photometric stereo under unknown spatial/spectral composition. Turning spectral ambiguity into a benefit, our method allows for the creation of training data without the need for complex multispectral rendering. Our work significantly broadens the scope for dynamic surface analysis, establishing a critical advancement in the utilization of photometric stereo across multiple sectors. Our proposed method exhibits several limitations. Firstly, there is unstable temporal variation in the normal maps reconstructed by our method for dynamic surface reconstruction. This instability arises from factors such as motion blur in certain frames, image noise, or the influence of cast/attached shadows, which become" + } + ], + [ + { + "type": "page_number", + "bbox": [ + 0.218, + 0.116, + 0.236, + 0.127 + ], + "angle": 0, + "content": "14" + }, + { + "type": "header", + "bbox": [ + 0.272, + 0.115, + 0.437, + 0.128 + ], + "angle": 0, + "content": "S. Ikehata and Y. Asano" + }, + { + "type": "image", + "bbox": [ + 0.23, + 0.147, + 0.778, + 0.684 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.226, + 0.694, + 0.776, + 0.71 + ], + "angle": 0, + "content": "Fig. 6: Evaluation on SpectraM14. Full results are available in the supplementary." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.727, + 0.788, + 0.834 + ], + "angle": 0, + "content": "more pronounced compared to conventional photometric stereo methods that utilize numerous images. To recover clean and temporally stable normal maps, we may need to consider temporal consistency and more actively utilize monocular cues. Additionally, while our method targets dynamic surfaces, it currently requires several seconds to up to ten seconds per RGB image, which is far from real-time processing. Considering industrial applications in the future, accelerating the processing speed is a crucial challenge." + } + ], + [ + { + "type": "header", + "bbox": [ + 0.64, + 0.115, + 0.732, + 0.129 + ], + "angle": 0, + "content": "SpectraM-PS" + }, + { + "type": "page_number", + "bbox": [ + 0.769, + 0.116, + 0.786, + 0.127 + ], + "angle": 0, + "content": "15" + }, + { + "type": "title", + "bbox": [ + 0.217, + 0.145, + 0.324, + 0.16 + ], + "angle": 0, + "content": "References" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.179, + 0.786, + 0.22 + ], + "angle": 0, + "content": "1. 3D Textures - Free seamless PBR textures with Diffuse, Normal, Displacement, Occlusion, Specularity and Roughness Maps. https://3dtextures.me/, accessed: 2024-03-07 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.222, + 0.558, + 0.235 + ], + "angle": 0, + "content": "2. Adobe Stock. https://stock.adobe.com/ 7,9" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.236, + 0.786, + 0.262 + ], + "angle": 0, + "content": "3. AmbientCG - Free Public Domain PBR Materials. https://ambientcg.com/, accessed: 2024-03-07 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.264, + 0.507, + 0.277 + ], + "angle": 0, + "content": "4.Blender.https://www.blender.org/9" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.279, + 0.786, + 0.306 + ], + "angle": 0, + "content": "5. Poliigon - A library of materials, and HDR's for artists including free textures. https://www.poliigon.com/, accessed: 2024-03-07 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.307, + 0.786, + 0.334 + ], + "angle": 0, + "content": "6. Alldrin, N., Mallick, S., Kriegman, D.: Resolving the generalized bas-relief ambiguity by entropy minimization. CVPR (2007) 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.335, + 0.786, + 0.362 + ], + "angle": 0, + "content": "7. Anderson, R., Stenger, B., Cipolla, R.: Color photometric stereo for multicolored surfaces. ICCV (2011) 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.363, + 0.786, + 0.391 + ], + "angle": 0, + "content": "8. Basri, R., Jacobs, D., Kemelmacher, I.: Photometric stereo with general, unknown lighting. International Journal of computer vision 72(3), 239-257 (2007) 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.392, + 0.786, + 0.431 + ], + "angle": 0, + "content": "9. Ben-David, S., Blitzer, J., Crammer, K., Kulesza, A., Pereira, F., Vaughan, J.W.: A theory of learning from different domains. Machine learning 79, 151-175 (2010) 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.22, + 0.433, + 0.786, + 0.474 + ], + "angle": 0, + "content": "10. Chakrabarti, A., Sunkavalli, K.: Single-image rgb photometric stereo with spatially-varying albedo. In: 2016 Fourth International Conference on 3D Vision (3DV) (2016) 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.219, + 0.476, + 0.786, + 0.516 + ], + "angle": 0, + "content": "11. Chen, G., Han, K., Shi, B., Matsushita, Y., Wong, K.K.K.: Self-calibrating deep photometric stereo networks. In: 2019 IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR). pp. 8731-8739 (2019) 2, 4, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.219, + 0.518, + 0.786, + 0.544 + ], + "angle": 0, + "content": "12. Chen, G., Han, K., Wong, K.Y.K.: Ps-fcn: A flexible learning framework for photometric stereo. ECCV (2018) 2, 4, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.219, + 0.546, + 0.786, + 0.587 + ], + "angle": 0, + "content": "13. Chen, G., Waechter, M., Shi, B., Wong, K.Y.K., Matsushita, Y.: What is learned in deep uncalibrated photometric stereo? In: European Conference on Computer Vision. pp. 745-762. Springer (2020) 4, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.219, + 0.588, + 0.786, + 0.615 + ], + "angle": 0, + "content": "14. Decker, B., Kautz, J., Mertens, T., Bekaert, P.: Capturing multiple illumination conditions using time and color multiplexing. CVPR (2009) 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.219, + 0.616, + 0.786, + 0.643 + ], + "angle": 0, + "content": "15. Drew, M.S.: Shape from color. Technical Report CSS/LCCR TR 92-07, School of Computing Science, Simon Fraser University, Vancouver, BC (1992) 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.219, + 0.644, + 0.786, + 0.671 + ], + "angle": 0, + "content": "16. Goldman, D., Curless, B., Hertzmann, A., Seitz, S.: Shape and spatially-varying brdfs from photometric stereo. In: ICCV (October 2005) 3, 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.219, + 0.673, + 0.786, + 0.699 + ], + "angle": 0, + "content": "17. Guo, H., Okura, F., Shi, B.: Multispectral photometric stereo for spatially-varying spectral reflectances. IJCV 130, 2166-2183 (2022) 2, 11, 13" + }, + { + "type": "ref_text", + "bbox": [ + 0.219, + 0.7, + 0.786, + 0.741 + ], + "angle": 0, + "content": "18. Guo, H., Okura, F., Shi, B., Funatomi, T., Mukaigawa, Y., Matsushita, Y.: Multispectral photometric stereo for spatially-varying spectral reflectances: A well posed problem? In: CVPR. pp. 963-971 (2021) 2, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.219, + 0.742, + 0.786, + 0.769 + ], + "angle": 0, + "content": "19. Hayakawa, H.: Photometric stereo under a light source with arbitrary motion. JOSA 11(11), 3079-3089 (1994) 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.219, + 0.77, + 0.786, + 0.811 + ], + "angle": 0, + "content": "20. Hernández, C., Vogiatzis, G., Brostow, G., Stenger, B., Cipolla, R.: Non-rigid photometric stereo with colored lights. In: ICCV. pp. 1-8 (2007). https://doi.org/10.1109/ICCV.2007.4408939 2, 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.219, + 0.812, + 0.786, + 0.84 + ], + "angle": 0, + "content": "21. Hertzmann, A., Seitz, S.: Example-based photometric stereo: shape reconstruction with general, varying brdfs. IEEE TPAMI 27(8), 1254-1264 (2005) 4" + }, + { + "type": "list", + "bbox": [ + 0.219, + 0.179, + 0.786, + 0.84 + ], + "angle": 0, + "content": null + } + ], + [ + { + "type": "page_number", + "bbox": [ + 0.218, + 0.116, + 0.236, + 0.127 + ], + "angle": 0, + "content": "16" + }, + { + "type": "header", + "bbox": [ + 0.272, + 0.115, + 0.437, + 0.128 + ], + "angle": 0, + "content": "S. Ikehata and Y. Asano" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.148, + 0.785, + 0.187 + ], + "angle": 0, + "content": "22. Hill, B.: Color capture, color management, and the problem of metamerism: does multispectral imaging offer the solution? In: Proc. SPIE. pp. 2-14. SPIE (1999) 2, 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.189, + 0.785, + 0.217 + ], + "angle": 0, + "content": "23. Ikehata, S.: Cnn-ps: Cnn-based photometric stereo for general non-convex surfaces. In: ECCV (2018) 2, 4, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.218, + 0.785, + 0.243 + ], + "angle": 0, + "content": "24. Ikehata, S.: Ps-transformer: Learning sparse photometric stereo network using self-attention mechanism. In: BMVC (2021) 4, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.244, + 0.785, + 0.27 + ], + "angle": 0, + "content": "25. Ikehata, S.: Universal photometric stereo network using global lighting contexts. In: CVPR (2022) 2, 3, 4, 6, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.271, + 0.785, + 0.297 + ], + "angle": 0, + "content": "26. Ikehata, S.: Scalable, detailed and mask-free universal photometric stereo. In: CVPR (2023) 2, 3, 4, 6, 7, 8, 9, 11, 12" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.298, + 0.785, + 0.324 + ], + "angle": 0, + "content": "27. Ikehata, S., Aizawa, K.: Photometric stereo using constrained bivariate regression for general isotropic surfaces. In: CVPR (2014) 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.325, + 0.785, + 0.352 + ], + "angle": 0, + "content": "28. Ikehata, S., Wipf, D., Matsushita, Y., Aizawa, K.: Robust photometric stereo using sparse regression. In: CVPR (2012) 3, 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.353, + 0.785, + 0.392 + ], + "angle": 0, + "content": "29. Ikehata, S., Wipf, D., Matsushita, Y., Aizawa, K.: Photometric stereo using sparse bayesian regression for general diffuse surfaces. IEEE TPAMI 36(9), 1816-1831 (2014) 2, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.393, + 0.785, + 0.419 + ], + "angle": 0, + "content": "30. Ikehata, S.: Does physical interpretability of observation map improve photometric stereo networks? In: ICIP (2022) 4, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.42, + 0.785, + 0.46 + ], + "angle": 0, + "content": "31. Ishio, H., Minowa, J., Nosu, K.: Review and status of wavelength-division-multiplexing technology and its application. Journal of lightwave technology 2(4), 448-463 (1984) 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.461, + 0.785, + 0.487 + ], + "angle": 0, + "content": "32. Ju, Y., Dong, X., Wang, Y., Qi, L., Dong, J.: A dual-cue network for multispectral photometric stereo. Pattern Recognition 100, 107162 (2020) 2, 4, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.488, + 0.785, + 0.528 + ], + "angle": 0, + "content": "33. Ju, Y., Qi, L., Zhou, H., Dong, J., Lu, L.: Demultiplexing colored images for multispectral photometric stereo via deep neural networks. IEEE Access 6, 30804-30818 (2018) 2, 4, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.529, + 0.785, + 0.569 + ], + "angle": 0, + "content": "34. Ju, Y., Dong, J., Chen, S.: Recovering surface normal and arbitrary images: A dual regression network for photometric stereo. IEEE Transactions on Image Processing 30, 3676-3690 (2021) 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.569, + 0.785, + 0.595 + ], + "angle": 0, + "content": "35. Kajiya, J.T.: The rendering equation. SIGGRAPH Comput. Graph. 20(4), 143-150 (1986) 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.597, + 0.785, + 0.635 + ], + "angle": 0, + "content": "36. Kaya, B., Kumar, S., Oliveira, C., Ferrari, V., Van Gool, L.: Uncalibrated neural inverse rendering for photometric stereo of general surfaces. pp. 3804-3814 (2021) 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.637, + 0.785, + 0.676 + ], + "angle": 0, + "content": "37. Kontsevich, L.L., Petrov, A., Vergelskaya, I.: Reconstruction of shape from shading in color images. Journal of the Optical Society of America pp. 1047-1052 (1994) 2, 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.677, + 0.785, + 0.718 + ], + "angle": 0, + "content": "38. Lee, J., Lee, Y., Kim, J., Kosiorek, A., Choi, S., Teh, Y.W.: Set transformer: A framework for attention-based permutation-invariant neural networks. In: ICML. pp. 3744-3753 (2019) 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.719, + 0.785, + 0.745 + ], + "angle": 0, + "content": "39. Li, J., Robles-Kelly, A., You, S., Matsushita, Y.: Learning to minify photometric stereo. In: CVPR (2019) 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.746, + 0.785, + 0.772 + ], + "angle": 0, + "content": "40. Li, J., Li, H.: Neural reflectance for shape recovery with shadow handling. In: CVPR (2022) 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.773, + 0.785, + 0.799 + ], + "angle": 0, + "content": "41. Li, J., Li, H.: Self-calibrating photometric stereo by neural inverse rendering. In: ECCV (2022) 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.8, + 0.785, + 0.838 + ], + "angle": 0, + "content": "42. Liu, H., Yan, Y., Song, K., Yu, H.: Sps-net: Self-attention photometric stereo network. IEEE Transactions on Instrumentation and Measurement 70, 1-13 (2021) 4" + }, + { + "type": "list", + "bbox": [ + 0.217, + 0.148, + 0.785, + 0.838 + ], + "angle": 0, + "content": null + } + ], + [ + { + "type": "header", + "bbox": [ + 0.64, + 0.115, + 0.732, + 0.129 + ], + "angle": 0, + "content": "SpectraM-PS" + }, + { + "type": "page_number", + "bbox": [ + 0.769, + 0.116, + 0.786, + 0.127 + ], + "angle": 0, + "content": "17" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.147, + 0.785, + 0.174 + ], + "angle": 0, + "content": "43. Liu, Z., Mao, H., Chao-Yuan Wu, C.F.: A convnet for the 2020s. In: CVPR (2022) 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.176, + 0.787, + 0.217 + ], + "angle": 0, + "content": "44. Logothetis, F., Budvytis, I., Mecca, R., Cipolla, R.: Px-net: Simple and efficient pixel-wise training of photometric stereo networks. In: CVPR. pp. 12757-12766 (2021) 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.218, + 0.218, + 0.785, + 0.245 + ], + "angle": 0, + "content": "45. Lu, F., Matsushita, Y., Sato, I., Okabe, T., Sato, Y.: Uncalibrated photometric stereo for unknown isotropic reflectances. In: CVPR. pp. 1490-1497 (2013) 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.218, + 0.246, + 0.785, + 0.285 + ], + "angle": 0, + "content": "46. Lv, J., Guo, H., Chen, G., Liang, J., Shi, B.: Non-lambertian multispectral photometric stereo via spectral reflectance decomposition. In: IJCAI (2023) 2, 3, 4, 6, 7, 11, 13" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.287, + 0.785, + 0.327 + ], + "angle": 0, + "content": "47. Mecca, R., Rosman, G., Kimmel, R., Bruckstein, A.: Perspective photometric stereo with shadows. In: Proc. of 4th International Conference on Scale Space and Variational Methods in Computer Vision (2013) 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.328, + 0.785, + 0.355 + ], + "angle": 0, + "content": "48. Mitsubishi Electric Research Laboratories (MERL): MERL BRDF Database. http://www.merl.com/brdf/, accessed: 2024-03-07 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.356, + 0.785, + 0.395 + ], + "angle": 0, + "content": "49. Mo, Z., Shi, B., Lu, F., Yeung, S.K., Matsushita, Y.: Uncalibrated photometric stereo under natural illumination. pp. 2936-2945. IEEE Computer Society (2018) 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.398, + 0.785, + 0.438 + ], + "angle": 0, + "content": "50. Nayatani, Y., Kurioka, Y., Sobagaki, H.: Study on color rendering and metamerism (part 8). Journal of the Illuminating Engineering Institute of Japan 56(9), 529-536 (1972). https://doi.org/10.2150/jieij1917.56.9_529_2, 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.439, + 0.785, + 0.48 + ], + "angle": 0, + "content": "51. Ozawa, K., Sato, I., Yamaguchi, M.: Single color image photometric stereo for multi-colored surfaces. Computer Vision and Image Understanding 171, 140-149 (2018) 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.481, + 0.785, + 0.522 + ], + "angle": 0, + "content": "52. Ren, J., Wang, F., Zhang, J., Zheng, Q., Ren, M., Shi, B.: Diligent10²: A photometric stereo benchmark dataset with controlled shape and material variation. pp. 12581-12590 (June 2022) 3, 9" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.523, + 0.785, + 0.577 + ], + "angle": 0, + "content": "53. Santo, H., Samejima, M., Sugano, Y., Shi, B., Matsushita, Y.: Deep photometric stereo network. In: International Workshop on Physics Based Vision meets Deep Learning (PBDL) in Conjunction with IEEE International Conference on Computer Vision (ICCV) (2017) 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.578, + 0.785, + 0.604 + ], + "angle": 0, + "content": "54. Shi, B., Matsushita, Y., Wei, Y., Xu, C., Tan, P.: Self-calibrating photometric stereo. In: CVPR (2010) 2, 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.605, + 0.785, + 0.633 + ], + "angle": 0, + "content": "55. Shi, B., Tan, P., Matsushita, Y., Ikeuchi, K.: A biquadratic reflectance model for radiometric image analysis. In: CVPR (2012) 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.634, + 0.785, + 0.673 + ], + "angle": 0, + "content": "56. Shi, B., Wu, Z., Mo, Z., D.Duan, Yeung, S.K., Tan, P.: A benchmark dataset and evaluation for non-lambertian and uncalibrated photometric stereo. In: CVPR (2016) 3, 9" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.674, + 0.785, + 0.702 + ], + "angle": 0, + "content": "57. Silver, W.M.: Determining shape and reflectance using multiple images. Master's thesis, MIT (1980) 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.702, + 0.785, + 0.73 + ], + "angle": 0, + "content": "58. Taniai, T., Maehara, T.: Neural Inverse Rendering for General Reflectance Photometric Stereo. In: ICML (2018) 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.73, + 0.785, + 0.757 + ], + "angle": 0, + "content": "59. Tiwari, A., Raman, S.: Deepps2: Revisiting photometric stereo using two differently illuminated images. ECCV (2022) 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.758, + 0.785, + 0.785 + ], + "angle": 0, + "content": "60. Tzeng, E., Hoffman, J., Saenko, K., Darrell, T.: Adversarial discriminative domain adaptation. In: CVPR (2017) 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.786, + 0.785, + 0.813 + ], + "angle": 0, + "content": "61. Vaswani, A., Shazeer, N., Parmar, N., Uszkoreit, J., Jones, L., Gomez, A., Kaiser, L., Polosukhin, I.: Attention is all you need. In: NeurIPS (2017) 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.814, + 0.785, + 0.84 + ], + "angle": 0, + "content": "62. Vogiatzis, G., Hernandez, C.: Self-calibrated, multi-spectral photometric stereo for 3d face capture. IJCV 56(97), 91-103 (2012) 2" + }, + { + "type": "list", + "bbox": [ + 0.217, + 0.147, + 0.787, + 0.84 + ], + "angle": 0, + "content": null + } + ], + [ + { + "type": "page_number", + "bbox": [ + 0.218, + 0.116, + 0.236, + 0.127 + ], + "angle": 0, + "content": "18" + }, + { + "type": "header", + "bbox": [ + 0.272, + 0.115, + 0.437, + 0.128 + ], + "angle": 0, + "content": "S. Ikehata and Y. Asano" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.147, + 0.785, + 0.188 + ], + "angle": 0, + "content": "63. Wang, F., Ren, J., Guo, H., Ren, M., Shi, B.: Diligent-pi: A photometric stereo benchmark dataset with controlled shape and material variation (October 2023) 3, 9" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.189, + 0.785, + 0.217 + ], + "angle": 0, + "content": "64. Woodham, P.: Photometric method for determining surface orientation from multiple images. Opt. Engg 19(1), 139-144 (1980) 1, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.218, + 0.785, + 0.244 + ], + "angle": 0, + "content": "65. Wu, L., Ganesh, A., Shi, B., Matsushita, Y., Wang, Y., Ma, Y.: Robust photometric stereo via low-rank matrix completion and recovery. In: ACCV (2010) 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.245, + 0.785, + 0.272 + ], + "angle": 0, + "content": "66. Xiao, T., Liu, Y., Zhou, B., Jiang, Y., Sun, J.: Unified perceptual parsing for scene understanding. In: ECCV. pp. 418-434 (2018) 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.273, + 0.785, + 0.313 + ], + "angle": 0, + "content": "67. Xiong, R., Yang, Y., He, D., Zheng, K., Zheng, S., Xing, C., Zhang, H., Lan, Y., Wang, L., Liu, T.: On layer normalization in the transformer architecture. In: International Conference on Machine Learning. pp. 10524-10533. PMLR (2020) 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.315, + 0.785, + 0.342 + ], + "angle": 0, + "content": "68. Yao, Z., Li, K., Fu, Y., Hu, H., Shi, B.: Gps-net: Graph-based photometric stereo network. NeurIPS (2020) 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.343, + 0.785, + 0.384 + ], + "angle": 0, + "content": "69. Zheng, Q., Jia, Y., Shi, B., Jiang, X., Duan, L.Y., Kot, A.: Spline-net: Sparse photometric stereo through lighting interpolation and normal estimation networks. ICCV (2019) 4" + }, + { + "type": "list", + "bbox": [ + 0.217, + 0.147, + 0.785, + 0.384 + ], + "angle": 0, + "content": null + } + ] +] \ No newline at end of file diff --git a/2024/Physics-Free Spectrally Multiplexed Photometric Stereo under Unknown Spectral Composition/09ecf602-4895-4520-8961-a33ebbe3e42d_origin.pdf b/2024/Physics-Free Spectrally Multiplexed Photometric Stereo under Unknown Spectral Composition/09ecf602-4895-4520-8961-a33ebbe3e42d_origin.pdf new file mode 100644 index 0000000000000000000000000000000000000000..feb562a71a02bd5390b67a8869e852fed5c08dce --- /dev/null +++ b/2024/Physics-Free Spectrally Multiplexed Photometric Stereo under Unknown Spectral Composition/09ecf602-4895-4520-8961-a33ebbe3e42d_origin.pdf @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:ffd9a4a7a9e7378e7cad964534af15a7bd870b5489136caaa9bc5a7a30096c70 +size 4507173 diff --git a/2024/Physics-Free Spectrally Multiplexed Photometric Stereo under Unknown Spectral Composition/full.md b/2024/Physics-Free Spectrally Multiplexed Photometric Stereo under Unknown Spectral Composition/full.md new file mode 100644 index 0000000000000000000000000000000000000000..415e978dadb16996c8a1764e0174633aaa6cc5cb --- /dev/null +++ b/2024/Physics-Free Spectrally Multiplexed Photometric Stereo under Unknown Spectral Composition/full.md @@ -0,0 +1,266 @@ +# SpectraM-PS: Spectrally Multiplexed Photometric Stereo under Unknown Spectral Composition + +Satoshi Ikehata $^{1,2}$ and Yuta Asano $^{1}$ + +$^{1}$ National Institute of Informatics, Tokyo, Japan + +2 Tokyo Institute of Technology, Tokyo, Japan + +Abstract. In this paper, we present a groundbreaking spectrally multiplexed photometric stereo approach for recovering surface normals of dynamic surfaces without the need for calibrated lighting or sensors, a notable advancement in the field traditionally hindered by stringent prerequisites and spectral ambiguity. By embracing spectral ambiguity as an advantage, our technique enables the generation of training data without specialized multispectral rendering frameworks. We introduce a unique, physics-free network architecture, SpectraM-PS, that effectively processes multiplexed images to determine surface normals across a wide range of conditions and material types, without relying on specific physically-based knowledge. Additionally, we establish the first benchmark dataset, SpectraM14, for spectrally multiplexed photometric stereo, facilitating comprehensive evaluations against existing calibrated methods. Our contributions significantly enhance the capabilities for dynamic surface recovery, particularly in uncalibrated setups, marking a pivotal step forward in the application of photometric stereo across various domains. + +Keywords: Spectrally Multiplexed Photometric Stereo $\cdot$ Dynamic Surface Recovery $\cdot$ Multispectral Photometric Stereo + +# 1 Introduction + +Recovering detailed normals of dynamic surfaces is essential for monitoring various processes: in manufacturing, it helps in tracking wear and tear of machine parts; in agriculture, it allows for the observation of crop growth through changes in leaf geometry; and in sports engineering, it aids in improving equipment design and safety by analyzing how surfaces deform upon impact. + +Photometric Stereo (PS) [57,64] derives object surface normals from observations under different lighting conditions at a fixed viewpoint. Despite decades of progress, the requirement for objects to stay stationary during lighting changes challenges the recovery of dynamic surfaces, essential for analyzing temporal surface deformations. PS researches have employed spectral multiplexing for dynamic + +![](images/7abd35d53f8c0a67cc2785e152a28a9a22889d76b27d1c86b437f58a6f97ea06.jpg) +Fig. 1: (Left) Illustration of our SpectraM-PS. Our method recovers a surface normal map from a spectrally multiplexed image. The spectral/spatial composition for generating the observations is unknown. There is potential for a mismatch between the sensor's spectral sensitivity and the light source's spectral distribution, which may lead to crosstalk. (Right) By applying our method to individual frames of a video, the normal map of dynamic surfaces can be recovered. + +![](images/ecdb0a810f486ab59296eff4aed054c1ef32cab1271295e077be17453885b81a.jpg) + +surface recovery [10,20,32,37,51,62]—a technique originally used in telecommunications and spectroscopy [31]. This technique utilizes the varying wavelengths of light to multiplex and subsequently demultiplex signals within a single sensor, thereby increasing the capacity for information transmission. + +Historically, spectrally multiplexed photometric stereo is often referred to as color photometric stereo [7,14,20,32,33], specifically when objects are illuminated with monochromatic red, green, and blue lights from various angles, captured in the camera's RGB channels. Each channel is then treated as an observation under a distinct lighting for PS analysis. This technique has been further extended to not only RGB but also any number of spectral bands and is specifically referred to as multispectral photometric stereo [17,18,46]. These techniques enable dynamic surface recovery by processing each temporal multi-channel frame separately. + +Despite their potential in dynamic surface recovery, current spectrally multiplexed photometric stereo methods face stringent prerequisites that limit their practicality. These include the necessity for precisely calibrated directional lighting in controlled environments [7, 14, 17, 20, 46] and sensors with aligned spectral sensitivities [32, 33]. Furthermore, they make strong assumptions about the surface, requiring it to be convex, integrable, Lambertian, and exhibit uniform chromaticity [20, 37]. By contrast, recent PS methods without spectral multiplexing support non-Lambertian surfaces [29, 55], spatially-varying materials [12, 23], and the use of uncalibrated lighting [11, 25, 26, 54]. This disparity arises from the challenge where identical observations are produced by different spectral compositions of light, surface and sensor [22, 50], a phenomenon absent in conventional PS due to constant spectral compositions of them across images. Recently, Guo et al. [17, 18] thoroughly explored how the spectral ambiguity renders spectrally multiplexed photometric stereo ill-posed, necessitating severely unpractical conditions on light, surface, and sensor to resolve the ambiguity. + +In this work, we propose a spectrally multiplexed photometric stereo method that recovers normals directly from multiplexed observations produced by $un$ + +known composition of lights, surface, and sensor (See Fig. 1-left), drawing inspiration from recent data-driven photometric stereo methods [25, 26]. While prior works [18, 46] have considered the spectral ambiguity harmful and something that must be resolved, we demonstrate that it can even be beneficial for a data-driven approach as it compacts the input space and allows for the generation of training data without a multispectral rendering framework. Trained on spectrally composed observations, our generic, physics-free architecture directly maps a single multiplexed image with an order-agnostic, arbitrary number of channels to object surface normals without the need for calibrating lights and sensors, and without imposing severe constraints on surface reflectance and geometry. By applying our method to individual frames of a video, dynamic surface recovery via spectrally multiplexed photometric stereo in uncalibrated, uncontrolled scenarios is achieved as illustrated in Fig. 1 (right). + +While numerous benchmarks exist for conventional photometric stereo [52,56, 63], not a single benchmark is available for spectrally multiplexed PS. Therefore, we have created the first real benchmark dataset, namely SpectraM, for this task. For comparative evaluations with calibrated methods such as [18,46], we carefully calibrated directional light sources of different wavelengths, including their directions. We implemented five different difficulty settings by varying the type of light sources (RGB vs NIR) and whether individual light sources were illuminated independently or simultaneously, catering to both ideal conditions without channel crosstalk and more realistic conditions with channel crosstalk. + +Our contributions are summarized as follows: (1) We pioneer the use of spectrally multiplexed photometric stereo for recovering dynamic surfaces in uncalibrated setups, employing a data-driven approach to overcome spectral ambiguity, a significant barrier in prior work. (2) We introduce a unique, physics-free neural network, SpectraM-PS (Spectrally Multiplexed PS), that recovers surface normals from a spectrally multiplexed image, capable of handling images with any number of order-agnostic channels. (3) We demonstrate how spectral ambiguity restricts the input space for training data generation, offering a strategy for efficient dataset creation without the need for multispectral rendering. (4) We create the first evaluation benchmark, SpectraM, for this domain, showing our method's superiority over current calibrated spectrally multiplexed photometric stereo techniques. + +# 2 Related Work + +Temporally Multiplexed Photometric Stereo (Conventional). From a communication perspective, conventional photometric stereo, as originally proposed by Woodham [64], employs a time multiplexing strategy to recover static surfaces. This method involves temporally varying lighting conditions while capturing images from a fixed viewpoint. Since the same light sources and sensor always provide observations, image differences stem solely from changes in light direction and intensity. This approach simplifies addressing complex conditions such as cast shadows [27, 28], non-Lambertian surfaces [16, 29], non-convex sur + +faces [23], and uncalibrated lighting [11, 36, 54]. Recently, learning-based methods [11-13, 23-25, 34, 36, 39-42, 44, 53, 58, 59, 68, 69] have emerged as an effective alternative, addressing challenges faced by traditional, physics-based approaches [6, 8, 16, 19, 21, 28, 45, 65]. These data-driven methods regress normal maps from observations utilizing techniques such as observation map regression [23, 30], set pooling [12, 13], graph neural networks [68], Transformer [24], and neural rendering for inverse rendering optimization [40, 41, 58]. Notably, the introduction of universal photometric stereo methods [25, 26] has enabled the handling of unknown, spatially-varying lighting in a purely data-driven framework. Inspired by these advancements, our work aims to regress normals from observations under unknown light, surface, and sensor conditions. + +Spectrally Multiplexed Photometric Stereo. Despite its potential for dynamic surface recovery, spectrally multiplexed photometric stereo [15, 37] has remained less explored than its mainstream counterpart, primarily due to notable limitations. + +Lighting Constraints: Existing methods necessitate multiple directional lights in controlled settings, contrasting the flexibility of temporally multiplexed techniques that adapt to diverse lighting conditions [26, 47, 49]. They typically require pre-calibrated light directions and distinct light source spectra to prevent channel crosstalk. In contrast, our approach accommodates uncontrolled lighting scenarios without the need for predefined or calibrated setups. + +Surface and Sensor Constraints: Prior works assume significant limitations on surfaces, such as Lambertian, convex, and uniform properties [15,20,37]. Recent advances like Lv et al. [46] extend to non-Lambertian surfaces but still require uniform materials. Sensor requirements typically involve narrow-band spectral responses and a fixed number of channels, limiting flexibility. Our approach leverages a data-driven model, training neural networks on synthetic data to handle complex surfaces and varied spectral sensor responses. + +Data-driven Methods: To our knowledge, there are few data-driven methods for this task [32, 33, 46]. Previous studies, such as those by Ju et al. [32, 33], require identical spectral and spatial lighting conditions during both training and testing, greatly restricting their practicality. ELIE-Net [46] permits variability in training and test setups; however, strong assumptions on both light and surfaces prohibitively limit its applications. Our model, on the other hand, eschews explicit lighting models in favor of learning direct input-output relationships, allowing for accurate predictions under varied and unknown spectral compositions and supporting materials with spatially diverse properties. Furthermore, unlike ELIE-Net's reliance on spectral BRDF datasets, our training approach utilizes assets akin to those employed in conventional photometric stereo. + +# 3 Problem Statement + +Given a single image $I \in \mathbb{R}^{h \times w \times k}$ captured by a static $k$ -channel orthographic sensor, along with an optional object mask $M \in \mathbb{R}^{h \times w}$ , the objective of spectrally multiplexed photometric stereo is to recover the surface normals of the + +![](images/0c3178d821e2637119bc71ebb6a88d3a5c1c483925b28ab8846e50d737096fcb.jpg) +Fig. 2: SpectraM-PS involves decomposing a spectrally multiplexed image into independent channels. The Global Feature Encoder extracts a feature map from each channel. The surface vector is then recovered by the Dual-scale Surface Normal Decoder at each pixel. We adopt a dual-scale approach to preserve the entire shape, while employing patch-embedding techniques to enhance local surface details. + +object, $N \in \mathbb{R}^{h \times w \times 33}$ . The object is supposed to be illuminated by multiple light sources, each with unique spatial and spectral properties. Previous studies have typically assumed an equal number of light sources and sensor channels, with each light source's wavelength precisely matching the spectral response of a single channel, thereby precluding any channel crosstalk, and with the directions of lights predetermined. In contrast, we do not presuppose the spatial distribution of illumination nor require the spectrum of each light source to be exclusively aligned with the spectral responses of the sensor channels, thus permitting channel crosstalk. This distinction is elaborated in subsequent sections. + +# 4 Method + +We propose and tackle the problem of spectrally multiplexed photometric stereo from a single image with multiple channels, produced by an unknown spectral/spatial composition of the sensor, light, and surface. To build such a method, we train neural networks to directly infer the normal map from an image. + +Our method addresses two challenges: (1) a physics-free architecture that accepts a varying number of spectral channels and is agnostic to their order, and (2) an effective approximation of the spectrally multiplexed image for efficient training. We consider (2) to be of significant importance, yet it remains largely unexplored. Synthesizing spectrally multiplexed images in a physically accurate manner is prohibitively challenging, owing to the increased complexity of their parameter spaces and the scarcity of 3D assets with detailed spectral properties, + +as well as the complex nature of light-surface interactions across different wavelengths. To address this issue, developing an efficient approximation method for rendering spectrally multiplexed images using common RGB image rendering techniques is crucial. + +# 4.1 Physics-free Spectrally Multiplexed PS Network (SpectraM-PS) + +The architecture of SpectraM-PS is illustrated in Fig. 2. Drawing inspiration from established Transformer-based photometric stereo networks [24-26], we integrate an encoder to first extract the global features and a decoder to estimate per-pixel surface normals. The architecture derives normals solely from the input image and mask, without prior light information. This indicates that the architecture focuses the network's learning objective on the relationship between input and output without relying on physics-based principles, unlike prior works. + +In our model, all the interactions among features from different sensor channels are employed by naïve Transformer [61] in similar to [24-26]. Transformer functions by mapping input features to query, key, and value vectors of equal dimensions. These vectors are processed through a multi-head self-attention mechanism, utilizing a softmax layer, followed by a feed-forward network comprising two linear layers. Both the input and output layers maintain identical dimensionality, with the inner layer having twice the dimension of the input. Each layer is surrounded by a residual connection, succeeded by layer normalization [67]. The advantage of employing Transformers in photometric stereo networks lies in their capability to facilitate complex interactions among intermediate features, a task unachievable with simple operations like pooling [11-13, 46] and observation map [23, 30]. Additionally, the token-based attention mechanism allows for different number of input tokens (i.e., sensor channels) between training and test phases and ensures that the results are independent on the order of tokens. + +Building on the established Transformer-based architecture [26] for temporally multiplexed photometric stereo, we extend its scope to a spectrally multiplexed one. To accommodate a variable number of channels and eliminate dependency on their order, an input spectrally multiplexed image is first split into individual channels, each of which is concatenated with an object mask (If no mask is provided, replace with a matrix of ones.) and then input into the same encoder of a neural network. This approach is distinctly different from traditional methods that encode an input image as it is in neural networks [32, 33]. Then, at Preprocessing, we normalize each channel by dividing it by a random value between its maximum and mean. Each channel and mask are resized or cropped to a resolution $(c \times c)$ that is a multiple of 32 to be input into the multi-scale encoder. Global Feature Encoder first applies a backbone network (i.e., ConvNeXt-T [43]) to individually encode the concatenation of each channel and mask, then uses Transformer layers for channel-axis (i.e., sensor channel) feature communication across scales (the number of Transformer layers is \{0, 1, 2, 4\} at \{1/4, 1/8, 1/16, 1/32\} scales, hidden dimensions are same with input dimensions), and finally, a feature pyramid network [66] for integrating features + +at different levels. Note that the design of encoder is almost the same as [25, 26], except that images are replaced by sensor channels, so details are omitted. + +Given global features $\in \mathcal{R}^{k\times c / 4\times c / 4\times 256}$ , our novel Dual-scale Surface Normal Decoder adopts a dual-scale strategy for predicting point-wise surface normals at $m$ (i.e., 2048) sampled locations at the original resolution within the object mask. The first branch recovers low-frequency surface normals at the feature map resolution $(\frac{c}{4}\times \frac{c}{4})$ . Concretely, all global features corresponding to each sample location are processed by five channel-axis Transformer layers (with a 256 hidden dimension) and are pooled via Pooling-by-Multihead-Attention (PMA) [38] using an additional channel-axis Transformer layer (with a 384 hidden dimension). To enhance spatial communication, two spatial-axis Transformer layers (with a 384 hidden dimension) inspired by Ikehata [26] are employed (i.e., Transformer is employed among samples at different locations), with a final MLP $(384\to 192\to 3)$ predicting the low-frequency normals at sampled locations. The second branch focuses on high-resolution normal recovery, using patch embedding for local context at the same $m$ locations, with $w\times w$ patches ( $w = 21$ ) processed by an MLP (with a 256 hidden dimension) and two layer norms. These patches, concatenated with bilinearly interpolated global features, pass through five channel-axis Transformer blocks (with a 256 hidden dimension), PMA (with a 384 hidden dimension), and are merged with the first branch output normals into 387-dimensional vectors. Two additional spatial-axis Transformer layers (with a 384 hidden dimension) enable non-local interactions, culminating in a final MLP $(384\to 192\to 3)$ for high-resolution normals, normalized to unit vectors. The complete normal map is formed by merging all the vectors from different sample sets. + +It should be noted that while SDM-UniPS [26] targets temporally multiplexed PS with tens of images, and its decoder performs normal estimation purely on a pixel basis. In contrast, spectrally multiplexed PS deals with fewer channels (e.g., three with RGB sensors), making a pixel-basis architecture less effective. Therefore, we use patch embedding at the patch-basis decoder to capture fine details with a dual-scale architecture for preserving overall shape. Without a dual-scale design, the recovery of surface normals becomes overly influenced by local image textures captured through patch embedding. This leads to a failure in preserving the entire shape, resulting in a significant reduction in accuracy. Our motivation is supported by Fig. 3 (left), where SDM-UniPS [26] fails to recover fine details with six temporally multiplexed images, while our architecture produces a more plausible normal map. + +# 4.2 Efficient Training Strategy Utilizing Spectral Ambiguity + +Aligning the training and test data domains in neural networks is essential for optimal model performance [9, 60]. However, rendering spectrally multiplexed data poses challenges due to the scarcity of multispectral Bidirectional Reflectance Distribution Functions (BRDFs). In reality, ELIE-Net [46] was trained using only 51 measured isotropic spectral BRDFs. On the other hand, given the availability of various large isotropic BRDF databases [1-3, 5, 48], we seek to explore + +![](images/b695ec0cf40d59275318953451e4682acbb8053d33bf044e8a064af6b2c004d5.jpg) +SpectraM-PS on Temporally Multiplexed PS + +![](images/fa05781d3530d106be98b1c1a759a78a3f9e2397822eccebce576fd92afab5f3.jpg) + +![](images/141a0aa36e3067d2608f640e4e7407f3b5d8868d002fd3de8643b69849f105f8.jpg) +Fig. 3: (Left) Comparison of SpectraM-PS and SDM-UniPS [26] on six temporally multiplexed PS images. Due to the patch-wise basis of SpectraM-PS, fine details are better recovered. (Right) Illustration of different lighting conditions in PS-Multiplex. + +![](images/42e96c0f5b3f02f09b242cde36d0d9b9489f1c69ee99d84b5609a77276f0f075.jpg) + +![](images/93339ba1241421b894148825b095f4c1f97caed0c5305c7a19dee2a300bb6b99.jpg) +PS-Multiplex Training Dataset + +![](images/bbf1bf948329ec33c243adb4595a152a4f695154cc21a5114d95bf908edb8f92.jpg) + +![](images/56c548f0928b8d5f0123cbbfb2acef66aa11cac5be906d8cd2ff6d14704e0cc3.jpg) + +utilizing these datasets for training our model, leveraging the fact that our network does not distinguish images based on their physically-based principles. In this section, we highlight how RGB images serve as a practical approximation, simplifying the complexity inherent in multispectral imaging. + +We begin the discussion by characterizing multispectral imaging. Assuming that the surface doesn't emit light and only reflections on surface are considered, the image formation model is described as follows [35]: + +$$ +I _ {(s, p)} = \int_ {\Omega} \left(\omega_ {i} ^ {\top} n _ {p}\right) \int_ {0} ^ {\infty} S _ {s} (\lambda) f _ {p} \left(\omega_ {i}, \omega_ {o}, \lambda\right) L _ {p} \left(\omega_ {i}, \lambda\right) \mathrm {d} \lambda \mathrm {d} \omega_ {i}. \tag {1} +$$ + +In this equation, $I_{(s,p)}$ denotes the incoming spectral radiance at the sensor $s$ (or $s$ -th channel) from a surface point $p$ . The term $f_p$ represents BRDF, $L_p$ the incident light intensity at the surface point, and $\lambda$ the wavelength of the incident light. The symbols $\omega_i$ and $\omega_o$ denote the directions of incident and reflected light, respectively. $S_s(\lambda)$ refers to the spectral sensitivity of the sensor $s$ at wavelength $\lambda$ , $n_p$ is the surface normal, and $\varOmega$ represents the hemisphere over which incident light directions are possible. The integral sums over all incident directions and wavelengths. It is important to note that the incident light intensity $L_p$ depends not only on the direct contribution from light sources but also on the visibility of light (e.g., attached and cast shadows) and indirect illuminations. + +Eq. (1) illustrates the concept of spectral ambiguity, showing that an infinite number of combinations of $S_{s}(\lambda)$ , $f_{p}(\omega_{i},\omega_{o},\lambda)$ , and $L_{p}(\omega_{i},\lambda)$ can result in the same spectral radiance, including narrowband compositions. In other words, with spectral ambiguity, a single observation $I_{(s,p)}$ can encompass the observations for all spectral compositions that satisfy the equation (i.e., metamerism [22, 50]). This perspective justifies the theory of substituting multispectral images, which possess a broad parameter space, with narrowband RGB images. It is worth mentioning that channel crosstalk primarily affects the incident light intensity, consequently distorting the product of $S_{s}(\lambda)\cdot f_{p}(\omega_{i},\omega_{o},\lambda)\cdot L_{p}(\omega_{i},\lambda)$ in Eq. (1). + +This implies that observations influenced by spectral crosstalk can still be equivalently represented using a narrowband setup under spectral ambiguity. In the experiments, we demonstrate that our model, trained on three narrowband observations can be applied to multiplexed data with channel crosstalk. To realize this approximation, we rendered a large number of three-channel narrowband images using the path-tracing algorithm in Blender [4], where up to 10-bounce reflections are permitted, based on common 3D assets [2] for RGB rendering. Following the rendering pipeline described in [26], we rendered objects by combining three different lighting models: directional, point, and environmental (five combinatorial settings in total as shown in Fig. 3). To simulate spectrally multiplexed images, we defined R, G, and B light sources and illuminated the surface in a multiplexed manner. It is important to note that the rendered RGB images are decomposed into three grayscale images, each of which was independently fed into the network; therefore, any wavelength-dependent information is masked. For material diversity, we adopted the method from [26], categorizing 897 AdobeStock texture maps into three groups: 421 diffuse, 219 specular, and 257 metallic textures. Four objects from a set of 410 3D AdobeStock models were randomly selected and textured with these materials. This structured approach led to the rendering of 106,374 multiplexed images along with their ground truth surface normal maps, forming the 'PS-Multiplex' dataset. + +# 5 SpectraM14 Benchmark Dataset + +Due to the lack of a benchmark for spectrally multiplexed PS, the first comprehensive evaluation dataset, named SpectraM14, is created. This dataset includes 14 objects, each exhibiting a range of optical properties such as monochromatic or multicolored appearances and diffuse or specular reflections, as depicted in Fig. 4. Our benchmark encompasses tasks under five distinct conditions, as described later. + +Imaging Setup. To acquire our dataset, we utilized a color camera (FLIR GS3-U3-123S6C-C) and an NIR camera (FLIR GS3-U3-41C6NIR-C), both equipped with a $50\mathrm{mm}$ lens. For the NIR camera, we used narrowband filters with wavelengths of $750\mathrm{nm}$ , $850\mathrm{nm}$ , $880\mathrm{nm}$ , $905\mathrm{nm}$ , and $940\mathrm{nm}$ , and the acquired images were manually merged. Objects were placed $0.8\mathrm{m}$ from the camera to approximate orthographic projection. Following conventional PS benchmarks [52,56,63], data capture occurred in a controlled, dark environment with the scene draped in black cloth to mitigate interreflection. The camera's ISO sensitivity was minimized to enhance image quality. The imaging area was further isolated using low-reflectance cloths to suppress inter-reflection. For each illumination condition, we collected six images under varying exposures to produce HDR input images. For the evaluations throughout this paper, the images are cropped using an object mask and resized to $512\mathrm{px} \times 512\mathrm{px}$ . + +Lighting. Six LED and three halogen light sources, positioned roughly 1 meter from the object, provided illumination. We used the "Weeylite S05 RGB Pocket Lamp" and the "NPI PIS-UHX-AIR" for lighting. This setup enabled the use of + +![](images/4df85b5950b772f90a0bd59842858eea97972ceb5bd5cca701904cfbd75c86e8.jpg) +Fig. 4: Objects in SpectraM14. + +![](images/a034460c5348bb0c6d6a9aa4054cb7875debf186dcb33da0567907bcb923e1ea.jpg) +Fig. 5: Illustration of six conditions in SpectraM14. + +red, green, blue, yellow, magenta, cyan, and NIR lighting, with spectra validated using a Hamamatsu Photonics Multichannel Analyzer C10027-01. + +Calibration and Ground Truth Data. We measured the directions of lights using specular reflections from a mirror sphere. Light intensity was standardized across the visible spectrum by averaging RGB values from reflected light on a white target. The ground truth normals were captured with a SHINING 3D EinScan-SE scanner. + +Evaluation Procedure. The design philosophy of this benchmark is to assess the robustness and adaptability of spectrally multiplexed PS methods under realistic lighting conditions, accounting for variations in channel numbers and the presence of spectral crosstalk. For a comprehensive evaluation, we designed tasks under five distinct conditions as shown in Fig. 5: Condition 1: Color sensor, no crosstalk condition: Six colors of light (red, green, blue, cyan, yellow, magenta) were each independently illuminated and observed with an RGB sensor. Afterward, the channels of RGB were averaged. Condition 2: Color sensor, weak crosstalk condition: Three colors of light (red, green, blue) were simultaneously illuminated and observed through each channel of the RGB sensor. Condition 3: Color sensor, strong crosstalk condition: Three colors of light (cyan, yellow, magenta) were simultaneously illuminated and observed through each channel of the RGB sensor. Condition 4: NIR sensor, no crosstalk condition: Light at wavelengths of $750~\mathrm{nm}$ , $850~\mathrm{nm}$ , $880~\mathrm{nm}$ , $905~\mathrm{nm}$ , and $940~\mathrm{nm}$ were each independently illuminated and observed with a monochrome sensor corresponding to + +each wavelength. Condition 5: NIR sensor, spatially-varying lighting condition: New images were created by averaging two images taken under the conditions mentioned above. The combinations were (750 nm, 850 nm), (850 nm, 880 nm), (880 nm, 905 nm), (905 nm, 940 nm), and (940 nm, 750 nm). + +# 6 Experiment + +In this section, we evaluate our method on our SpectraM14. Our method is compared with one SOTA optimization-based method [17] and one SOTA learning-based method [46]. The former introduces a closed-form solution for spectrally multiplexed photometric stereo applied to monochromatic surfaces with spatially varying (SV) albedo. The latter presents a Spectral Reflectance Decomposition (SRD) model, which disentangles spectral reflectance into geometric and spectral components for surface normal recovery under non-Lambertian spectral reflectance conditions. Unlike the compared methods, our approach does not assume a specific lighting setup, whereas both methods presume the presence of calibrated single directional light sources. + +Training details. SpectraM-PS was trained from scratch on the PS-Multiplex dataset until convergence using the AdamW optimizer, with a step decay learning rate schedule that reduced the learning rate by a factor of 0.8 every ten epochs. We applied learning rate warmup during the first epoch and used a batch size of 16, an initial learning rate of 0.0001, and a weight decay of 0.05. Each batch consisted of three input training multiplexed images with three channels each. The training loss was computed using the Mean Squared Error (MSE) loss function to measure $\ell_2$ errors between the predicted surface normal vectors and the ground truth surface normal vectors. We measured the reconstruction accuracy of our method by computing the mean angular errors (MAE) between the predicted and true surface normal maps, expressed in degrees. + +Computational Cost. The inference time of PS methods varies with the number of pixels and channels in the input image. For Condition 2 and 3 with a $512 \times 512 \times 3$ image, the mean and standard deviation of inference times (in sec) over 14 objects in SpectraM14 benchmark were: our method (3.42/0.85), Lv et al. [46] (0.46/0.24) and Guo et al. [17] (2.38/1.10). Our architecture leads to higher computational costs; however, none of the methods were suitable for real-time processing (e.g., 15 fps requires 0.06 sec/frame). + +Ablation Study. We firstly validate the individual technical contributions of our training dataset (i.e., PS-Multiplex) and the physics-free architecture (i.e., SpectraM-PS) using a synthetic evaluation dataset. Firstly, we validate the efficacy of our training dataset, PS-Multiplex, by adapting an existing universal photometric stereo architecture designed for the conventional task (i.e., SDM-UniPS [26]) to the spectrally multiplexed photometric stereo task. Since both ours and SDM-UniPS take multiple observations and an object mask as input, this adaptation straightforwardly involves training the model on PS-Multiplex by treating each channel of an image as an individual image. Subsequently, we + +Table 1: Ablation analysis of the contributions of SpectraM-PS and PS-Multiplex. + +
MAE (Uniform)MAE (Piece-wise uniform)MAE (Non-uniform)
MethodNon-LambertianLambertianNon-LambertianLambertianNon-LambertianLambertian
SDM-UniPS [26]12.9 (4.5)12.4 (4.7)15.0 (5.0)21.7 (7.0)14.4 (5.1)15.2 (5.9)
[26] trained on PS-Multiplex11.1 (3.6)11.0 (3.9)10.5 (2.9)12.3 (3.9)10.6 (3.9)11.2 (3.7)
SpectraM-PS (Ours)8.0 (2.7)8.4 (2.7)7.9 (2.4)8.9 (2.9)8.2 (3.2)8.0 (2.5)
+ +compare this model against our proposed SpectraM-PS to demonstrate the efficacy of our dual-scale design with local patch embedding. + +For evaluating the contribution of our architecture (SpectraM-PS) and training dataset (PS-Multiplex), we additionally rendered three-channel spectrally multiplexed images representing six distinct surface material categories: (a) uniform, Lambertian; (b) piece-wise uniform, Lambertian; (c) non-uniform, Lambertian; (d) uniform, non-Lambertian; (e) piece-wise uniform, non-Lambertian; and (f) non-uniform, non-Lambertian. In uniform materials, every point on the surface within a scene exhibits the same material properties. For piece-wise uniform materials, each object in a scene is composed of the same material, yet different objects possess distinct materials. Non-uniform materials feature unique PBR textures assigned to each object. The rendering process for these images was identical to that used for the PS-Multiplex datasets in each category. We generated 100 scenes for each surface material category, and MAEs (stds) are averaged over them. The results are presented in Tab.1. In summary, SDM-UniPS [26] trained on our PS-Multiplex dataset demonstrates proper adaptation to the spectrally multiplexed photometric stereo task. Nonetheless, our SpectraM-PS method significantly enhanced reconstruction accuracy, showcasing an architecture-level improvement over SDM-UniPS for the spectrally multiplexed photometric stereo task, where the number of input channels is typically much fewer than that of input images for conventional PS. + +# Comparative Evaluation on SpectraM14. The results are illustrated in Tabs. 2 + +to 6 and Fig. 6. Despite the fact that all existing spectrally multiplexed photometric stereo methods assume calibrated light sources and known directional light source conditions, our proposed method significantly outperformed them. This is because most of the real objects used in our experiment are neither Lambertian nor convex, and do not conform to their assumptions. However, our non-physical-based method successfully restored the normals very stably for these objects. Furthermore, our proposed method enabled robust reconstruction for all objects, despite having been trained only with RGB color images. This result supports the efficacy of our approximation. Furthermore, unlike existing methods that suffer from reduced estimation accuracy with increasing spectral crosstalk, our approach demonstrates only minimal performance degradation. Remarkably, our method excels in recovering a more realistic structure with spatially-varying surface materials. This breakthrough implies that our network can effectively achieve dynamic surface reconstruction across video frames in a universal setting. We will detail this groundbreaking application in the next section. Due to space constraints, not all results can be included here. However, all + +Table 2: Comparison in condition 1. The values are mean angular errors in degrees. + +
MethodObject ID
1234567891011121314Ave.
Ours8.611.213.07.312.26.111.55.05.77.510.35.16.112.28.9
Lv et al. [46]20.417.021.113.923.110.721.216.610.915.919.016.413.218.617.1
Guo et al. [17]22.615.220.713.427.17.231.324.88.018.224.511.410.229.318.9
+ +Table 3: Comparison in condition 2. The values are mean angular errors in degrees. + +
MethodObject IDAve.
1234567891011121314
Ours10.111.513.59.012.87.010.95.56.39.312.75.39.814.710.0
Lv et al. [46]22.826.027.019.430.319.522.118.914.319.823.520.819.021.421.7
Guo et al. [17]31.127.529.220.138.019.033.423.513.626.632.714.717.139.325.7
+ +Table 4: Comparison in condition 3. The values are mean angular errors in degrees. + +
MethodObject ID
1234567891011121314
Ours12.012.815.811.016.55.912.39.86.99.814.66.37.520.1
Lv et al. [46]38.534.138.232.440.138.236.629.638.235.138.236.638.430.9
Guo et al. [17]46.042.656.337.657.145.676.048.029.962.149.850.252.873.7
+ +Table 5: Comparison in condition 4. The values are mean angular errors in degrees. + +
MethodObject ID
1234567891011121314
Ours10.911.39.76.310.85.312.44.37.86.69.14.07.410.6
Lv et al. [46]24.119.820.712.019.812.415.118.711.417.121.217.519.116.7
Guo et al. [17]30.716.018.39.929.68.423.125.710.214.624.513.929.613.8
+ +Table 6: Comparison in condition 5. The values are mean angular errors in degrees. + +
MethodObject IDAve.
1234567891011121314
Ours10.911.09.97.311.04.713.34.87.96.59.84.27.410.48.6
Lv et al. [46]29.826.228.922.828.925.821.924.618.724.727.624.822.327.625.0
Guo et al. [17]40.027.329.223.233.325.029.330.120.026.731.626.325.027.227.8
+ +results are comprehensively presented in the supplementary materials. Additionally, the supplementary materials evaluate the impact of the spatial distribution of light sources on the performance of the proposed method. We also offer an in-depth discussion of each experimental condition therein. + +# 7 Conclusion + +In this work, we introduce an innovative approach to spectrally multiplexed photometric stereo under unknown spatial/spectral composition. Turning spectral ambiguity into a benefit, our method allows for the creation of training data without the need for complex multispectral rendering. Our work significantly broadens the scope for dynamic surface analysis, establishing a critical advancement in the utilization of photometric stereo across multiple sectors. Our proposed method exhibits several limitations. Firstly, there is unstable temporal variation in the normal maps reconstructed by our method for dynamic surface reconstruction. This instability arises from factors such as motion blur in certain frames, image noise, or the influence of cast/attached shadows, which become + +![](images/f141a01604beedf219d31153145f94f5220479126cd51437fba78600b9794178.jpg) +Fig. 6: Evaluation on SpectraM14. Full results are available in the supplementary. + +more pronounced compared to conventional photometric stereo methods that utilize numerous images. To recover clean and temporally stable normal maps, we may need to consider temporal consistency and more actively utilize monocular cues. Additionally, while our method targets dynamic surfaces, it currently requires several seconds to up to ten seconds per RGB image, which is far from real-time processing. Considering industrial applications in the future, accelerating the processing speed is a crucial challenge. + +# References + +1. 3D Textures - Free seamless PBR textures with Diffuse, Normal, Displacement, Occlusion, Specularity and Roughness Maps. https://3dtextures.me/, accessed: 2024-03-07 7 +2. Adobe Stock. https://stock.adobe.com/ 7,9 +3. AmbientCG - Free Public Domain PBR Materials. https://ambientcg.com/, accessed: 2024-03-07 7 +4.Blender.https://www.blender.org/9 +5. Poliigon - A library of materials, and HDR's for artists including free textures. https://www.poliigon.com/, accessed: 2024-03-07 7 +6. Alldrin, N., Mallick, S., Kriegman, D.: Resolving the generalized bas-relief ambiguity by entropy minimization. CVPR (2007) 4 +7. Anderson, R., Stenger, B., Cipolla, R.: Color photometric stereo for multicolored surfaces. ICCV (2011) 2 +8. Basri, R., Jacobs, D., Kemelmacher, I.: Photometric stereo with general, unknown lighting. International Journal of computer vision 72(3), 239-257 (2007) 4 +9. Ben-David, S., Blitzer, J., Crammer, K., Kulesza, A., Pereira, F., Vaughan, J.W.: A theory of learning from different domains. Machine learning 79, 151-175 (2010) 7 +10. Chakrabarti, A., Sunkavalli, K.: Single-image rgb photometric stereo with spatially-varying albedo. In: 2016 Fourth International Conference on 3D Vision (3DV) (2016) 2 +11. Chen, G., Han, K., Shi, B., Matsushita, Y., Wong, K.K.K.: Self-calibrating deep photometric stereo networks. In: 2019 IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR). pp. 8731-8739 (2019) 2, 4, 6 +12. Chen, G., Han, K., Wong, K.Y.K.: Ps-fcn: A flexible learning framework for photometric stereo. ECCV (2018) 2, 4, 6 +13. Chen, G., Waechter, M., Shi, B., Wong, K.Y.K., Matsushita, Y.: What is learned in deep uncalibrated photometric stereo? In: European Conference on Computer Vision. pp. 745-762. Springer (2020) 4, 6 +14. Decker, B., Kautz, J., Mertens, T., Bekaert, P.: Capturing multiple illumination conditions using time and color multiplexing. CVPR (2009) 2 +15. Drew, M.S.: Shape from color. Technical Report CSS/LCCR TR 92-07, School of Computing Science, Simon Fraser University, Vancouver, BC (1992) 4 +16. Goldman, D., Curless, B., Hertzmann, A., Seitz, S.: Shape and spatially-varying brdfs from photometric stereo. In: ICCV (October 2005) 3, 4 +17. Guo, H., Okura, F., Shi, B.: Multispectral photometric stereo for spatially-varying spectral reflectances. IJCV 130, 2166-2183 (2022) 2, 11, 13 +18. Guo, H., Okura, F., Shi, B., Funatomi, T., Mukaigawa, Y., Matsushita, Y.: Multispectral photometric stereo for spatially-varying spectral reflectances: A well posed problem? In: CVPR. pp. 963-971 (2021) 2, 3 +19. Hayakawa, H.: Photometric stereo under a light source with arbitrary motion. JOSA 11(11), 3079-3089 (1994) 4 +20. Hernández, C., Vogiatzis, G., Brostow, G., Stenger, B., Cipolla, R.: Non-rigid photometric stereo with colored lights. In: ICCV. pp. 1-8 (2007). https://doi.org/10.1109/ICCV.2007.4408939 2, 4 +21. Hertzmann, A., Seitz, S.: Example-based photometric stereo: shape reconstruction with general, varying brdfs. IEEE TPAMI 27(8), 1254-1264 (2005) 4 + +22. Hill, B.: Color capture, color management, and the problem of metamerism: does multispectral imaging offer the solution? In: Proc. SPIE. pp. 2-14. SPIE (1999) 2, 8 +23. Ikehata, S.: Cnn-ps: Cnn-based photometric stereo for general non-convex surfaces. In: ECCV (2018) 2, 4, 6 +24. Ikehata, S.: Ps-transformer: Learning sparse photometric stereo network using self-attention mechanism. In: BMVC (2021) 4, 6 +25. Ikehata, S.: Universal photometric stereo network using global lighting contexts. In: CVPR (2022) 2, 3, 4, 6, 7 +26. Ikehata, S.: Scalable, detailed and mask-free universal photometric stereo. In: CVPR (2023) 2, 3, 4, 6, 7, 8, 9, 11, 12 +27. Ikehata, S., Aizawa, K.: Photometric stereo using constrained bivariate regression for general isotropic surfaces. In: CVPR (2014) 3 +28. Ikehata, S., Wipf, D., Matsushita, Y., Aizawa, K.: Robust photometric stereo using sparse regression. In: CVPR (2012) 3, 4 +29. Ikehata, S., Wipf, D., Matsushita, Y., Aizawa, K.: Photometric stereo using sparse bayesian regression for general diffuse surfaces. IEEE TPAMI 36(9), 1816-1831 (2014) 2, 3 +30. Ikehata, S.: Does physical interpretability of observation map improve photometric stereo networks? In: ICIP (2022) 4, 6 +31. Ishio, H., Minowa, J., Nosu, K.: Review and status of wavelength-division-multiplexing technology and its application. Journal of lightwave technology 2(4), 448-463 (1984) 2 +32. Ju, Y., Dong, X., Wang, Y., Qi, L., Dong, J.: A dual-cue network for multispectral photometric stereo. Pattern Recognition 100, 107162 (2020) 2, 4, 6 +33. Ju, Y., Qi, L., Zhou, H., Dong, J., Lu, L.: Demultiplexing colored images for multispectral photometric stereo via deep neural networks. IEEE Access 6, 30804-30818 (2018) 2, 4, 6 +34. Ju, Y., Dong, J., Chen, S.: Recovering surface normal and arbitrary images: A dual regression network for photometric stereo. IEEE Transactions on Image Processing 30, 3676-3690 (2021) 4 +35. Kajiya, J.T.: The rendering equation. SIGGRAPH Comput. Graph. 20(4), 143-150 (1986) 8 +36. Kaya, B., Kumar, S., Oliveira, C., Ferrari, V., Van Gool, L.: Uncalibrated neural inverse rendering for photometric stereo of general surfaces. pp. 3804-3814 (2021) 4 +37. Kontsevich, L.L., Petrov, A., Vergelskaya, I.: Reconstruction of shape from shading in color images. Journal of the Optical Society of America pp. 1047-1052 (1994) 2, 4 +38. Lee, J., Lee, Y., Kim, J., Kosiorek, A., Choi, S., Teh, Y.W.: Set transformer: A framework for attention-based permutation-invariant neural networks. In: ICML. pp. 3744-3753 (2019) 7 +39. Li, J., Robles-Kelly, A., You, S., Matsushita, Y.: Learning to minify photometric stereo. In: CVPR (2019) 4 +40. Li, J., Li, H.: Neural reflectance for shape recovery with shadow handling. In: CVPR (2022) 4 +41. Li, J., Li, H.: Self-calibrating photometric stereo by neural inverse rendering. In: ECCV (2022) 4 +42. Liu, H., Yan, Y., Song, K., Yu, H.: Sps-net: Self-attention photometric stereo network. IEEE Transactions on Instrumentation and Measurement 70, 1-13 (2021) 4 + +43. Liu, Z., Mao, H., Chao-Yuan Wu, C.F.: A convnet for the 2020s. In: CVPR (2022) 6 +44. Logothetis, F., Budvytis, I., Mecca, R., Cipolla, R.: Px-net: Simple and efficient pixel-wise training of photometric stereo networks. In: CVPR. pp. 12757-12766 (2021) 4 +45. Lu, F., Matsushita, Y., Sato, I., Okabe, T., Sato, Y.: Uncalibrated photometric stereo for unknown isotropic reflectances. In: CVPR. pp. 1490-1497 (2013) 4 +46. Lv, J., Guo, H., Chen, G., Liang, J., Shi, B.: Non-lambertian multispectral photometric stereo via spectral reflectance decomposition. In: IJCAI (2023) 2, 3, 4, 6, 7, 11, 13 +47. Mecca, R., Rosman, G., Kimmel, R., Bruckstein, A.: Perspective photometric stereo with shadows. In: Proc. of 4th International Conference on Scale Space and Variational Methods in Computer Vision (2013) 4 +48. Mitsubishi Electric Research Laboratories (MERL): MERL BRDF Database. http://www.merl.com/brdf/, accessed: 2024-03-07 7 +49. Mo, Z., Shi, B., Lu, F., Yeung, S.K., Matsushita, Y.: Uncalibrated photometric stereo under natural illumination. pp. 2936-2945. IEEE Computer Society (2018) 4 +50. Nayatani, Y., Kurioka, Y., Sobagaki, H.: Study on color rendering and metamerism (part 8). Journal of the Illuminating Engineering Institute of Japan 56(9), 529-536 (1972). https://doi.org/10.2150/jieij1917.56.9_529_2, 8 +51. Ozawa, K., Sato, I., Yamaguchi, M.: Single color image photometric stereo for multi-colored surfaces. Computer Vision and Image Understanding 171, 140-149 (2018) 2 +52. Ren, J., Wang, F., Zhang, J., Zheng, Q., Ren, M., Shi, B.: Diligent10²: A photometric stereo benchmark dataset with controlled shape and material variation. pp. 12581-12590 (June 2022) 3, 9 +53. Santo, H., Samejima, M., Sugano, Y., Shi, B., Matsushita, Y.: Deep photometric stereo network. In: International Workshop on Physics Based Vision meets Deep Learning (PBDL) in Conjunction with IEEE International Conference on Computer Vision (ICCV) (2017) 4 +54. Shi, B., Matsushita, Y., Wei, Y., Xu, C., Tan, P.: Self-calibrating photometric stereo. In: CVPR (2010) 2, 4 +55. Shi, B., Tan, P., Matsushita, Y., Ikeuchi, K.: A biquadratic reflectance model for radiometric image analysis. In: CVPR (2012) 2 +56. Shi, B., Wu, Z., Mo, Z., D.Duan, Yeung, S.K., Tan, P.: A benchmark dataset and evaluation for non-lambertian and uncalibrated photometric stereo. In: CVPR (2016) 3, 9 +57. Silver, W.M.: Determining shape and reflectance using multiple images. Master's thesis, MIT (1980) 1 +58. Taniai, T., Maehara, T.: Neural Inverse Rendering for General Reflectance Photometric Stereo. In: ICML (2018) 4 +59. Tiwari, A., Raman, S.: Deepps2: Revisiting photometric stereo using two differently illuminated images. ECCV (2022) 4 +60. Tzeng, E., Hoffman, J., Saenko, K., Darrell, T.: Adversarial discriminative domain adaptation. In: CVPR (2017) 7 +61. Vaswani, A., Shazeer, N., Parmar, N., Uszkoreit, J., Jones, L., Gomez, A., Kaiser, L., Polosukhin, I.: Attention is all you need. In: NeurIPS (2017) 6 +62. Vogiatzis, G., Hernandez, C.: Self-calibrated, multi-spectral photometric stereo for 3d face capture. IJCV 56(97), 91-103 (2012) 2 + +63. Wang, F., Ren, J., Guo, H., Ren, M., Shi, B.: Diligent-pi: A photometric stereo benchmark dataset with controlled shape and material variation (October 2023) 3, 9 +64. Woodham, P.: Photometric method for determining surface orientation from multiple images. Opt. Engg 19(1), 139-144 (1980) 1, 3 +65. Wu, L., Ganesh, A., Shi, B., Matsushita, Y., Wang, Y., Ma, Y.: Robust photometric stereo via low-rank matrix completion and recovery. In: ACCV (2010) 4 +66. Xiao, T., Liu, Y., Zhou, B., Jiang, Y., Sun, J.: Unified perceptual parsing for scene understanding. In: ECCV. pp. 418-434 (2018) 6 +67. Xiong, R., Yang, Y., He, D., Zheng, K., Zheng, S., Xing, C., Zhang, H., Lan, Y., Wang, L., Liu, T.: On layer normalization in the transformer architecture. In: International Conference on Machine Learning. pp. 10524-10533. PMLR (2020) 6 +68. Yao, Z., Li, K., Fu, Y., Hu, H., Shi, B.: Gps-net: Graph-based photometric stereo network. NeurIPS (2020) 4 +69. Zheng, Q., Jia, Y., Shi, B., Jiang, X., Duan, L.Y., Kot, A.: Spline-net: Sparse photometric stereo through lighting interpolation and normal estimation networks. ICCV (2019) 4 \ No newline at end of file diff --git a/2024/Physics-Free Spectrally Multiplexed Photometric Stereo under Unknown Spectral Composition/images.zip b/2024/Physics-Free Spectrally Multiplexed Photometric Stereo under Unknown Spectral Composition/images.zip new file mode 100644 index 0000000000000000000000000000000000000000..e0d2d004b560e8c5e4ba8c1bbfce43df44d5615c --- /dev/null +++ b/2024/Physics-Free Spectrally Multiplexed Photometric Stereo under Unknown Spectral Composition/images.zip @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:e09cff807d184e13f1e7bf4177a86b39c2adfa0c387e3bf3e36d28255b71223f +size 680603 diff --git a/2024/Physics-Free Spectrally Multiplexed Photometric Stereo under Unknown Spectral Composition/layout.json b/2024/Physics-Free Spectrally Multiplexed Photometric Stereo under Unknown Spectral Composition/layout.json new file mode 100644 index 0000000000000000000000000000000000000000..6d1c339b57ca959ac3a8100ea2c09479420ef5a2 --- /dev/null +++ b/2024/Physics-Free Spectrally Multiplexed Photometric Stereo under Unknown Spectral Composition/layout.json @@ -0,0 +1,8276 @@ +{ + "pdf_info": [ + { + "para_blocks": [ + { + "bbox": [ + 149, + 111, + 465, + 166 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 149, + 111, + 465, + 166 + ], + "spans": [ + { + "bbox": [ + 149, + 111, + 465, + 166 + ], + "type": "text", + "content": "SpectraM-PS: Spectrally Multiplexed Photometric Stereo under Unknown Spectral Composition" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 219, + 186, + 394, + 198 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 219, + 186, + 394, + 198 + ], + "spans": [ + { + "bbox": [ + 219, + 186, + 394, + 198 + ], + "type": "text", + "content": "Satoshi Ikehata" + }, + { + "bbox": [ + 219, + 186, + 394, + 198 + ], + "type": "inline_equation", + "content": "^{1,2}" + }, + { + "bbox": [ + 219, + 186, + 394, + 198 + ], + "type": "text", + "content": " and Yuta Asano" + }, + { + "bbox": [ + 219, + 186, + 394, + 198 + ], + "type": "inline_equation", + "content": "^{1}" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 205, + 207, + 408, + 219 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 205, + 207, + 408, + 219 + ], + "spans": [ + { + "bbox": [ + 205, + 207, + 408, + 219 + ], + "type": "inline_equation", + "content": "^{1}" + }, + { + "bbox": [ + 205, + 207, + 408, + 219 + ], + "type": "text", + "content": " National Institute of Informatics, Tokyo, Japan" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 210, + 220, + 402, + 230 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 210, + 220, + 402, + 230 + ], + "spans": [ + { + "bbox": [ + 210, + 220, + 402, + 230 + ], + "type": "text", + "content": "2 Tokyo Institute of Technology, Tokyo, Japan" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 159, + 261, + 455, + 448 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 159, + 261, + 455, + 448 + ], + "spans": [ + { + "bbox": [ + 159, + 261, + 455, + 448 + ], + "type": "text", + "content": "Abstract. In this paper, we present a groundbreaking spectrally multiplexed photometric stereo approach for recovering surface normals of dynamic surfaces without the need for calibrated lighting or sensors, a notable advancement in the field traditionally hindered by stringent prerequisites and spectral ambiguity. By embracing spectral ambiguity as an advantage, our technique enables the generation of training data without specialized multispectral rendering frameworks. We introduce a unique, physics-free network architecture, SpectraM-PS, that effectively processes multiplexed images to determine surface normals across a wide range of conditions and material types, without relying on specific physically-based knowledge. Additionally, we establish the first benchmark dataset, SpectraM14, for spectrally multiplexed photometric stereo, facilitating comprehensive evaluations against existing calibrated methods. Our contributions significantly enhance the capabilities for dynamic surface recovery, particularly in uncalibrated setups, marking a pivotal step forward in the application of photometric stereo across various domains." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 159, + 458, + 453, + 480 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 159, + 458, + 453, + 480 + ], + "spans": [ + { + "bbox": [ + 159, + 458, + 453, + 480 + ], + "type": "text", + "content": "Keywords: Spectrally Multiplexed Photometric Stereo " + }, + { + "bbox": [ + 159, + 458, + 453, + 480 + ], + "type": "inline_equation", + "content": "\\cdot" + }, + { + "bbox": [ + 159, + 458, + 453, + 480 + ], + "type": "text", + "content": " Dynamic Surface Recovery " + }, + { + "bbox": [ + 159, + 458, + 453, + 480 + ], + "type": "inline_equation", + "content": "\\cdot" + }, + { + "bbox": [ + 159, + 458, + 453, + 480 + ], + "type": "text", + "content": " Multispectral Photometric Stereo" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 132, + 520, + 230, + 533 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 520, + 230, + 533 + ], + "spans": [ + { + "bbox": [ + 132, + 520, + 230, + 533 + ], + "type": "text", + "content": "1 Introduction" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 130, + 545, + 482, + 605 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 545, + 482, + 605 + ], + "spans": [ + { + "bbox": [ + 130, + 545, + 482, + 605 + ], + "type": "text", + "content": "Recovering detailed normals of dynamic surfaces is essential for monitoring various processes: in manufacturing, it helps in tracking wear and tear of machine parts; in agriculture, it allows for the observation of crop growth through changes in leaf geometry; and in sports engineering, it aids in improving equipment design and safety by analyzing how surfaces deform upon impact." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 130, + 605, + 482, + 665 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 605, + 482, + 665 + ], + "spans": [ + { + "bbox": [ + 130, + 605, + 482, + 665 + ], + "type": "text", + "content": "Photometric Stereo (PS) [57,64] derives object surface normals from observations under different lighting conditions at a fixed viewpoint. Despite decades of progress, the requirement for objects to stay stationary during lighting changes challenges the recovery of dynamic surfaces, essential for analyzing temporal surface deformations. PS researches have employed spectral multiplexing for dynamic" + } + ] + } + ], + "index": 8 + } + ], + "discarded_blocks": [], + "page_size": [ + 612, + 792 + ], + "page_idx": 0 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 136, + 114, + 284, + 232 + ], + "blocks": [ + { + "bbox": [ + 136, + 114, + 284, + 232 + ], + "lines": [ + { + "bbox": [ + 136, + 114, + 284, + 232 + ], + "spans": [ + { + "bbox": [ + 136, + 114, + 284, + 232 + ], + "type": "image", + "image_path": "7abd35d53f8c0a67cc2785e152a28a9a22889d76b27d1c86b437f58a6f97ea06.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 130, + 239, + 482, + 307 + ], + "lines": [ + { + "bbox": [ + 130, + 239, + 482, + 307 + ], + "spans": [ + { + "bbox": [ + 130, + 239, + 482, + 307 + ], + "type": "text", + "content": "Fig. 1: (Left) Illustration of our SpectraM-PS. Our method recovers a surface normal map from a spectrally multiplexed image. The spectral/spatial composition for generating the observations is unknown. There is potential for a mismatch between the sensor's spectral sensitivity and the light source's spectral distribution, which may lead to crosstalk. (Right) By applying our method to individual frames of a video, the normal map of dynamic surfaces can be recovered." + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "image_caption" + } + ], + "index": 2 + }, + { + "type": "image", + "bbox": [ + 285, + 114, + 478, + 232 + ], + "blocks": [ + { + "bbox": [ + 285, + 114, + 478, + 232 + ], + "lines": [ + { + "bbox": [ + 285, + 114, + 478, + 232 + ], + "spans": [ + { + "bbox": [ + 285, + 114, + 478, + 232 + ], + "type": "image", + "image_path": "ecdb0a810f486ab59296eff4aed054c1ef32cab1271295e077be17453885b81a.jpg" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_body" + } + ], + "index": 3 + }, + { + "bbox": [ + 130, + 317, + 480, + 365 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 317, + 480, + 365 + ], + "spans": [ + { + "bbox": [ + 130, + 317, + 480, + 365 + ], + "type": "text", + "content": "surface recovery [10,20,32,37,51,62]—a technique originally used in telecommunications and spectroscopy [31]. This technique utilizes the varying wavelengths of light to multiplex and subsequently demultiplex signals within a single sensor, thereby increasing the capacity for information transmission." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 130, + 365, + 482, + 461 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 365, + 482, + 461 + ], + "spans": [ + { + "bbox": [ + 130, + 365, + 482, + 461 + ], + "type": "text", + "content": "Historically, spectrally multiplexed photometric stereo is often referred to as color photometric stereo [7,14,20,32,33], specifically when objects are illuminated with monochromatic red, green, and blue lights from various angles, captured in the camera's RGB channels. Each channel is then treated as an observation under a distinct lighting for PS analysis. This technique has been further extended to not only RGB but also any number of spectral bands and is specifically referred to as multispectral photometric stereo [17,18,46]. These techniques enable dynamic surface recovery by processing each temporal multi-channel frame separately." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 130, + 462, + 482, + 641 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 462, + 482, + 641 + ], + "spans": [ + { + "bbox": [ + 130, + 462, + 482, + 641 + ], + "type": "text", + "content": "Despite their potential in dynamic surface recovery, current spectrally multiplexed photometric stereo methods face stringent prerequisites that limit their practicality. These include the necessity for precisely calibrated directional lighting in controlled environments [7, 14, 17, 20, 46] and sensors with aligned spectral sensitivities [32, 33]. Furthermore, they make strong assumptions about the surface, requiring it to be convex, integrable, Lambertian, and exhibit uniform chromaticity [20, 37]. By contrast, recent PS methods without spectral multiplexing support non-Lambertian surfaces [29, 55], spatially-varying materials [12, 23], and the use of uncalibrated lighting [11, 25, 26, 54]. This disparity arises from the challenge where identical observations are produced by different spectral compositions of light, surface and sensor [22, 50], a phenomenon absent in conventional PS due to constant spectral compositions of them across images. Recently, Guo et al. [17, 18] thoroughly explored how the spectral ambiguity renders spectrally multiplexed photometric stereo ill-posed, necessitating severely unpractical conditions on light, surface, and sensor to resolve the ambiguity." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 131, + 641, + 482, + 666 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 131, + 641, + 482, + 666 + ], + "spans": [ + { + "bbox": [ + 131, + 641, + 482, + 666 + ], + "type": "text", + "content": "In this work, we propose a spectrally multiplexed photometric stereo method that recovers normals directly from multiplexed observations produced by " + }, + { + "bbox": [ + 131, + 641, + 482, + 666 + ], + "type": "inline_equation", + "content": "un" + } + ] + } + ], + "index": 8 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "spans": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "type": "text", + "content": "2" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 166, + 91, + 267, + 101 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 166, + 91, + 267, + 101 + ], + "spans": [ + { + "bbox": [ + 166, + 91, + 267, + 101 + ], + "type": "text", + "content": "S. Ikehata and Y. Asano" + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 1 + }, + { + "para_blocks": [ + { + "bbox": [ + 130, + 115, + 482, + 272 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 115, + 482, + 272 + ], + "spans": [ + { + "bbox": [ + 130, + 115, + 482, + 272 + ], + "type": "text", + "content": "known composition of lights, surface, and sensor (See Fig. 1-left), drawing inspiration from recent data-driven photometric stereo methods [25, 26]. While prior works [18, 46] have considered the spectral ambiguity harmful and something that must be resolved, we demonstrate that it can even be beneficial for a data-driven approach as it compacts the input space and allows for the generation of training data without a multispectral rendering framework. Trained on spectrally composed observations, our generic, physics-free architecture directly maps a single multiplexed image with an order-agnostic, arbitrary number of channels to object surface normals without the need for calibrating lights and sensors, and without imposing severe constraints on surface reflectance and geometry. By applying our method to individual frames of a video, dynamic surface recovery via spectrally multiplexed photometric stereo in uncalibrated, uncontrolled scenarios is achieved as illustrated in Fig. 1 (right)." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 130, + 272, + 482, + 379 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 272, + 482, + 379 + ], + "spans": [ + { + "bbox": [ + 130, + 272, + 482, + 379 + ], + "type": "text", + "content": "While numerous benchmarks exist for conventional photometric stereo [52,56, 63], not a single benchmark is available for spectrally multiplexed PS. Therefore, we have created the first real benchmark dataset, namely SpectraM, for this task. For comparative evaluations with calibrated methods such as [18,46], we carefully calibrated directional light sources of different wavelengths, including their directions. We implemented five different difficulty settings by varying the type of light sources (RGB vs NIR) and whether individual light sources were illuminated independently or simultaneously, catering to both ideal conditions without channel crosstalk and more realistic conditions with channel crosstalk." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 130, + 380, + 482, + 523 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 380, + 482, + 523 + ], + "spans": [ + { + "bbox": [ + 130, + 380, + 482, + 523 + ], + "type": "text", + "content": "Our contributions are summarized as follows: (1) We pioneer the use of spectrally multiplexed photometric stereo for recovering dynamic surfaces in uncalibrated setups, employing a data-driven approach to overcome spectral ambiguity, a significant barrier in prior work. (2) We introduce a unique, physics-free neural network, SpectraM-PS (Spectrally Multiplexed PS), that recovers surface normals from a spectrally multiplexed image, capable of handling images with any number of order-agnostic channels. (3) We demonstrate how spectral ambiguity restricts the input space for training data generation, offering a strategy for efficient dataset creation without the need for multispectral rendering. (4) We create the first evaluation benchmark, SpectraM, for this domain, showing our method's superiority over current calibrated spectrally multiplexed photometric stereo techniques." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 131, + 543, + 238, + 556 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 131, + 543, + 238, + 556 + ], + "spans": [ + { + "bbox": [ + 131, + 543, + 238, + 556 + ], + "type": "text", + "content": "2 Related Work" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 130, + 569, + 482, + 666 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 569, + 482, + 666 + ], + "spans": [ + { + "bbox": [ + 130, + 569, + 482, + 666 + ], + "type": "text", + "content": "Temporally Multiplexed Photometric Stereo (Conventional). From a communication perspective, conventional photometric stereo, as originally proposed by Woodham [64], employs a time multiplexing strategy to recover static surfaces. This method involves temporally varying lighting conditions while capturing images from a fixed viewpoint. Since the same light sources and sensor always provide observations, image differences stem solely from changes in light direction and intensity. This approach simplifies addressing complex conditions such as cast shadows [27, 28], non-Lambertian surfaces [16, 29], non-convex sur" + } + ] + } + ], + "index": 6 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 392, + 91, + 447, + 102 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 392, + 91, + 447, + 102 + ], + "spans": [ + { + "bbox": [ + 392, + 91, + 447, + 102 + ], + "type": "text", + "content": "SpectraM-PS" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 474, + 92, + 481, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 474, + 92, + 481, + 100 + ], + "spans": [ + { + "bbox": [ + 474, + 92, + 481, + 100 + ], + "type": "text", + "content": "3" + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 2 + }, + { + "para_blocks": [ + { + "bbox": [ + 130, + 116, + 482, + 248 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 116, + 482, + 248 + ], + "spans": [ + { + "bbox": [ + 130, + 116, + 482, + 248 + ], + "type": "text", + "content": "faces [23], and uncalibrated lighting [11, 36, 54]. Recently, learning-based methods [11-13, 23-25, 34, 36, 39-42, 44, 53, 58, 59, 68, 69] have emerged as an effective alternative, addressing challenges faced by traditional, physics-based approaches [6, 8, 16, 19, 21, 28, 45, 65]. These data-driven methods regress normal maps from observations utilizing techniques such as observation map regression [23, 30], set pooling [12, 13], graph neural networks [68], Transformer [24], and neural rendering for inverse rendering optimization [40, 41, 58]. Notably, the introduction of universal photometric stereo methods [25, 26] has enabled the handling of unknown, spatially-varying lighting in a purely data-driven framework. Inspired by these advancements, our work aims to regress normals from observations under unknown light, surface, and sensor conditions." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 130, + 250, + 482, + 297 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 250, + 482, + 297 + ], + "spans": [ + { + "bbox": [ + 130, + 250, + 482, + 297 + ], + "type": "text", + "content": "Spectrally Multiplexed Photometric Stereo. Despite its potential for dynamic surface recovery, spectrally multiplexed photometric stereo [15, 37] has remained less explored than its mainstream counterpart, primarily due to notable limitations." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 130, + 298, + 482, + 370 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 298, + 482, + 370 + ], + "spans": [ + { + "bbox": [ + 130, + 298, + 482, + 370 + ], + "type": "text", + "content": "Lighting Constraints: Existing methods necessitate multiple directional lights in controlled settings, contrasting the flexibility of temporally multiplexed techniques that adapt to diverse lighting conditions [26, 47, 49]. They typically require pre-calibrated light directions and distinct light source spectra to prevent channel crosstalk. In contrast, our approach accommodates uncontrolled lighting scenarios without the need for predefined or calibrated setups." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 130, + 370, + 482, + 453 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 370, + 482, + 453 + ], + "spans": [ + { + "bbox": [ + 130, + 370, + 482, + 453 + ], + "type": "text", + "content": "Surface and Sensor Constraints: Prior works assume significant limitations on surfaces, such as Lambertian, convex, and uniform properties [15,20,37]. Recent advances like Lv et al. [46] extend to non-Lambertian surfaces but still require uniform materials. Sensor requirements typically involve narrow-band spectral responses and a fixed number of channels, limiting flexibility. Our approach leverages a data-driven model, training neural networks on synthetic data to handle complex surfaces and varied spectral sensor responses." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 130, + 454, + 482, + 586 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 454, + 482, + 586 + ], + "spans": [ + { + "bbox": [ + 130, + 454, + 482, + 586 + ], + "type": "text", + "content": "Data-driven Methods: To our knowledge, there are few data-driven methods for this task [32, 33, 46]. Previous studies, such as those by Ju et al. [32, 33], require identical spectral and spatial lighting conditions during both training and testing, greatly restricting their practicality. ELIE-Net [46] permits variability in training and test setups; however, strong assumptions on both light and surfaces prohibitively limit its applications. Our model, on the other hand, eschews explicit lighting models in favor of learning direct input-output relationships, allowing for accurate predictions under varied and unknown spectral compositions and supporting materials with spatially diverse properties. Furthermore, unlike ELIE-Net's reliance on spectral BRDF datasets, our training approach utilizes assets akin to those employed in conventional photometric stereo." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 131, + 604, + 270, + 616 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 131, + 604, + 270, + 616 + ], + "spans": [ + { + "bbox": [ + 131, + 604, + 270, + 616 + ], + "type": "text", + "content": "3 Problem Statement" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 130, + 628, + 482, + 665 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 628, + 482, + 665 + ], + "spans": [ + { + "bbox": [ + 130, + 628, + 482, + 665 + ], + "type": "text", + "content": "Given a single image " + }, + { + "bbox": [ + 130, + 628, + 482, + 665 + ], + "type": "inline_equation", + "content": "I \\in \\mathbb{R}^{h \\times w \\times k}" + }, + { + "bbox": [ + 130, + 628, + 482, + 665 + ], + "type": "text", + "content": " captured by a static " + }, + { + "bbox": [ + 130, + 628, + 482, + 665 + ], + "type": "inline_equation", + "content": "k" + }, + { + "bbox": [ + 130, + 628, + 482, + 665 + ], + "type": "text", + "content": "-channel orthographic sensor, along with an optional object mask " + }, + { + "bbox": [ + 130, + 628, + 482, + 665 + ], + "type": "inline_equation", + "content": "M \\in \\mathbb{R}^{h \\times w}" + }, + { + "bbox": [ + 130, + 628, + 482, + 665 + ], + "type": "text", + "content": ", the objective of spectrally multiplexed photometric stereo is to recover the surface normals of the" + } + ] + } + ], + "index": 8 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "spans": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "type": "text", + "content": "4" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 166, + 91, + 267, + 101 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 166, + 91, + 267, + 101 + ], + "spans": [ + { + "bbox": [ + 166, + 91, + 267, + 101 + ], + "type": "text", + "content": "S. Ikehata and Y. Asano" + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 3 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 141, + 114, + 478, + 264 + ], + "blocks": [ + { + "bbox": [ + 141, + 114, + 478, + 264 + ], + "lines": [ + { + "bbox": [ + 141, + 114, + 478, + 264 + ], + "spans": [ + { + "bbox": [ + 141, + 114, + 478, + 264 + ], + "type": "image", + "image_path": "0c3178d821e2637119bc71ebb6a88d3a5c1c483925b28ab8846e50d737096fcb.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 130, + 273, + 482, + 328 + ], + "lines": [ + { + "bbox": [ + 130, + 273, + 482, + 328 + ], + "spans": [ + { + "bbox": [ + 130, + 273, + 482, + 328 + ], + "type": "text", + "content": "Fig. 2: SpectraM-PS involves decomposing a spectrally multiplexed image into independent channels. The Global Feature Encoder extracts a feature map from each channel. The surface vector is then recovered by the Dual-scale Surface Normal Decoder at each pixel. We adopt a dual-scale approach to preserve the entire shape, while employing patch-embedding techniques to enhance local surface details." + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_caption" + } + ], + "index": 2 + }, + { + "bbox": [ + 130, + 347, + 482, + 456 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 347, + 482, + 456 + ], + "spans": [ + { + "bbox": [ + 130, + 347, + 482, + 456 + ], + "type": "text", + "content": "object, " + }, + { + "bbox": [ + 130, + 347, + 482, + 456 + ], + "type": "inline_equation", + "content": "N \\in \\mathbb{R}^{h \\times w \\times 33}" + }, + { + "bbox": [ + 130, + 347, + 482, + 456 + ], + "type": "text", + "content": ". The object is supposed to be illuminated by multiple light sources, each with unique spatial and spectral properties. Previous studies have typically assumed an equal number of light sources and sensor channels, with each light source's wavelength precisely matching the spectral response of a single channel, thereby precluding any channel crosstalk, and with the directions of lights predetermined. In contrast, we do not presuppose the spatial distribution of illumination nor require the spectrum of each light source to be exclusively aligned with the spectral responses of the sensor channels, thus permitting channel crosstalk. This distinction is elaborated in subsequent sections." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 132, + 475, + 201, + 488 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 475, + 201, + 488 + ], + "spans": [ + { + "bbox": [ + 132, + 475, + 201, + 488 + ], + "type": "text", + "content": "4 Method" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 130, + 502, + 481, + 550 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 502, + 481, + 550 + ], + "spans": [ + { + "bbox": [ + 130, + 502, + 481, + 550 + ], + "type": "text", + "content": "We propose and tackle the problem of spectrally multiplexed photometric stereo from a single image with multiple channels, produced by an unknown spectral/spatial composition of the sensor, light, and surface. To build such a method, we train neural networks to directly infer the normal map from an image." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 130, + 551, + 482, + 635 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 551, + 482, + 635 + ], + "spans": [ + { + "bbox": [ + 130, + 551, + 482, + 635 + ], + "type": "text", + "content": "Our method addresses two challenges: (1) a physics-free architecture that accepts a varying number of spectral channels and is agnostic to their order, and (2) an effective approximation of the spectrally multiplexed image for efficient training. We consider (2) to be of significant importance, yet it remains largely unexplored. Synthesizing spectrally multiplexed images in a physically accurate manner is prohibitively challenging, owing to the increased complexity of their parameter spaces and the scarcity of 3D assets with detailed spectral properties," + } + ] + } + ], + "index": 7 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 392, + 91, + 447, + 102 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 392, + 91, + 447, + 102 + ], + "spans": [ + { + "bbox": [ + 392, + 91, + 447, + 102 + ], + "type": "text", + "content": "SpectraM-PS" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 474, + 92, + 480, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 474, + 92, + 480, + 100 + ], + "spans": [ + { + "bbox": [ + 474, + 92, + 480, + 100 + ], + "type": "text", + "content": "5" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 133, + 642, + 481, + 665 + ], + "type": "page_footnote", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 642, + 481, + 665 + ], + "spans": [ + { + "bbox": [ + 133, + 642, + 481, + 665 + ], + "type": "text", + "content": "3 It should be noted that unlike conventional PS, reflectance recovery generally falls outside the scope of spectrally multiplexed PS due to its inherently ill-posed nature." + } + ] + } + ], + "index": 8 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 4 + }, + { + "para_blocks": [ + { + "bbox": [ + 130, + 116, + 482, + 163 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 116, + 482, + 163 + ], + "spans": [ + { + "bbox": [ + 130, + 116, + 482, + 163 + ], + "type": "text", + "content": "as well as the complex nature of light-surface interactions across different wavelengths. To address this issue, developing an efficient approximation method for rendering spectrally multiplexed images using common RGB image rendering techniques is crucial." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 131, + 186, + 481, + 198 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 131, + 186, + 481, + 198 + ], + "spans": [ + { + "bbox": [ + 131, + 186, + 481, + 198 + ], + "type": "text", + "content": "4.1 Physics-free Spectrally Multiplexed PS Network (SpectraM-PS)" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 130, + 209, + 482, + 293 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 209, + 482, + 293 + ], + "spans": [ + { + "bbox": [ + 130, + 209, + 482, + 293 + ], + "type": "text", + "content": "The architecture of SpectraM-PS is illustrated in Fig. 2. Drawing inspiration from established Transformer-based photometric stereo networks [24-26], we integrate an encoder to first extract the global features and a decoder to estimate per-pixel surface normals. The architecture derives normals solely from the input image and mask, without prior light information. This indicates that the architecture focuses the network's learning objective on the relationship between input and output without relying on physics-based principles, unlike prior works." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 130, + 294, + 482, + 461 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 294, + 482, + 461 + ], + "spans": [ + { + "bbox": [ + 130, + 294, + 482, + 461 + ], + "type": "text", + "content": "In our model, all the interactions among features from different sensor channels are employed by naïve Transformer [61] in similar to [24-26]. Transformer functions by mapping input features to query, key, and value vectors of equal dimensions. These vectors are processed through a multi-head self-attention mechanism, utilizing a softmax layer, followed by a feed-forward network comprising two linear layers. Both the input and output layers maintain identical dimensionality, with the inner layer having twice the dimension of the input. Each layer is surrounded by a residual connection, succeeded by layer normalization [67]. The advantage of employing Transformers in photometric stereo networks lies in their capability to facilitate complex interactions among intermediate features, a task unachievable with simple operations like pooling [11-13, 46] and observation map [23, 30]. Additionally, the token-based attention mechanism allows for different number of input tokens (i.e., sensor channels) between training and test phases and ensures that the results are independent on the order of tokens." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 130, + 462, + 482, + 666 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 462, + 482, + 666 + ], + "spans": [ + { + "bbox": [ + 130, + 462, + 482, + 666 + ], + "type": "text", + "content": "Building on the established Transformer-based architecture [26] for temporally multiplexed photometric stereo, we extend its scope to a spectrally multiplexed one. To accommodate a variable number of channels and eliminate dependency on their order, an input spectrally multiplexed image is first split into individual channels, each of which is concatenated with an object mask (If no mask is provided, replace with a matrix of ones.) and then input into the same encoder of a neural network. This approach is distinctly different from traditional methods that encode an input image as it is in neural networks [32, 33]. Then, at Preprocessing, we normalize each channel by dividing it by a random value between its maximum and mean. Each channel and mask are resized or cropped to a resolution " + }, + { + "bbox": [ + 130, + 462, + 482, + 666 + ], + "type": "inline_equation", + "content": "(c \\times c)" + }, + { + "bbox": [ + 130, + 462, + 482, + 666 + ], + "type": "text", + "content": " that is a multiple of 32 to be input into the multi-scale encoder. Global Feature Encoder first applies a backbone network (i.e., ConvNeXt-T [43]) to individually encode the concatenation of each channel and mask, then uses Transformer layers for channel-axis (i.e., sensor channel) feature communication across scales (the number of Transformer layers is \\{0, 1, 2, 4\\} at \\{1/4, 1/8, 1/16, 1/32\\} scales, hidden dimensions are same with input dimensions), and finally, a feature pyramid network [66] for integrating features" + } + ] + } + ], + "index": 6 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "spans": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "type": "text", + "content": "6" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 166, + 91, + 267, + 101 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 166, + 91, + 267, + 101 + ], + "spans": [ + { + "bbox": [ + 166, + 91, + 267, + 101 + ], + "type": "text", + "content": "S. Ikehata and Y. Asano" + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 5 + }, + { + "para_blocks": [ + { + "bbox": [ + 130, + 116, + 479, + 139 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 116, + 479, + 139 + ], + "spans": [ + { + "bbox": [ + 130, + 116, + 479, + 139 + ], + "type": "text", + "content": "at different levels. Note that the design of encoder is almost the same as [25, 26], except that images are replaced by sensor channels, so details are omitted." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 130, + 140, + 482, + 415 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 140, + 482, + 415 + ], + "spans": [ + { + "bbox": [ + 130, + 140, + 482, + 415 + ], + "type": "text", + "content": "Given global features " + }, + { + "bbox": [ + 130, + 140, + 482, + 415 + ], + "type": "inline_equation", + "content": "\\in \\mathcal{R}^{k\\times c / 4\\times c / 4\\times 256}" + }, + { + "bbox": [ + 130, + 140, + 482, + 415 + ], + "type": "text", + "content": ", our novel Dual-scale Surface Normal Decoder adopts a dual-scale strategy for predicting point-wise surface normals at " + }, + { + "bbox": [ + 130, + 140, + 482, + 415 + ], + "type": "inline_equation", + "content": "m" + }, + { + "bbox": [ + 130, + 140, + 482, + 415 + ], + "type": "text", + "content": " (i.e., 2048) sampled locations at the original resolution within the object mask. The first branch recovers low-frequency surface normals at the feature map resolution " + }, + { + "bbox": [ + 130, + 140, + 482, + 415 + ], + "type": "inline_equation", + "content": "(\\frac{c}{4}\\times \\frac{c}{4})" + }, + { + "bbox": [ + 130, + 140, + 482, + 415 + ], + "type": "text", + "content": ". Concretely, all global features corresponding to each sample location are processed by five channel-axis Transformer layers (with a 256 hidden dimension) and are pooled via Pooling-by-Multihead-Attention (PMA) [38] using an additional channel-axis Transformer layer (with a 384 hidden dimension). To enhance spatial communication, two spatial-axis Transformer layers (with a 384 hidden dimension) inspired by Ikehata [26] are employed (i.e., Transformer is employed among samples at different locations), with a final MLP " + }, + { + "bbox": [ + 130, + 140, + 482, + 415 + ], + "type": "inline_equation", + "content": "(384\\to 192\\to 3)" + }, + { + "bbox": [ + 130, + 140, + 482, + 415 + ], + "type": "text", + "content": " predicting the low-frequency normals at sampled locations. The second branch focuses on high-resolution normal recovery, using patch embedding for local context at the same " + }, + { + "bbox": [ + 130, + 140, + 482, + 415 + ], + "type": "inline_equation", + "content": "m" + }, + { + "bbox": [ + 130, + 140, + 482, + 415 + ], + "type": "text", + "content": " locations, with " + }, + { + "bbox": [ + 130, + 140, + 482, + 415 + ], + "type": "inline_equation", + "content": "w\\times w" + }, + { + "bbox": [ + 130, + 140, + 482, + 415 + ], + "type": "text", + "content": " patches (" + }, + { + "bbox": [ + 130, + 140, + 482, + 415 + ], + "type": "inline_equation", + "content": "w = 21" + }, + { + "bbox": [ + 130, + 140, + 482, + 415 + ], + "type": "text", + "content": ") processed by an MLP (with a 256 hidden dimension) and two layer norms. These patches, concatenated with bilinearly interpolated global features, pass through five channel-axis Transformer blocks (with a 256 hidden dimension), PMA (with a 384 hidden dimension), and are merged with the first branch output normals into 387-dimensional vectors. Two additional spatial-axis Transformer layers (with a 384 hidden dimension) enable non-local interactions, culminating in a final MLP " + }, + { + "bbox": [ + 130, + 140, + 482, + 415 + ], + "type": "inline_equation", + "content": "(384\\to 192\\to 3)" + }, + { + "bbox": [ + 130, + 140, + 482, + 415 + ], + "type": "text", + "content": " for high-resolution normals, normalized to unit vectors. The complete normal map is formed by merging all the vectors from different sample sets." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 130, + 415, + 482, + 559 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 415, + 482, + 559 + ], + "spans": [ + { + "bbox": [ + 130, + 415, + 482, + 559 + ], + "type": "text", + "content": "It should be noted that while SDM-UniPS [26] targets temporally multiplexed PS with tens of images, and its decoder performs normal estimation purely on a pixel basis. In contrast, spectrally multiplexed PS deals with fewer channels (e.g., three with RGB sensors), making a pixel-basis architecture less effective. Therefore, we use patch embedding at the patch-basis decoder to capture fine details with a dual-scale architecture for preserving overall shape. Without a dual-scale design, the recovery of surface normals becomes overly influenced by local image textures captured through patch embedding. This leads to a failure in preserving the entire shape, resulting in a significant reduction in accuracy. Our motivation is supported by Fig. 3 (left), where SDM-UniPS [26] fails to recover fine details with six temporally multiplexed images, while our architecture produces a more plausible normal map." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 131, + 574, + 443, + 588 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 131, + 574, + 443, + 588 + ], + "spans": [ + { + "bbox": [ + 131, + 574, + 443, + 588 + ], + "type": "text", + "content": "4.2 Efficient Training Strategy Utilizing Spectral Ambiguity" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 130, + 594, + 482, + 666 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 594, + 482, + 666 + ], + "spans": [ + { + "bbox": [ + 130, + 594, + 482, + 666 + ], + "type": "text", + "content": "Aligning the training and test data domains in neural networks is essential for optimal model performance [9, 60]. However, rendering spectrally multiplexed data poses challenges due to the scarcity of multispectral Bidirectional Reflectance Distribution Functions (BRDFs). In reality, ELIE-Net [46] was trained using only 51 measured isotropic spectral BRDFs. On the other hand, given the availability of various large isotropic BRDF databases [1-3, 5, 48], we seek to explore" + } + ] + } + ], + "index": 6 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 391, + 90, + 447, + 102 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 391, + 90, + 447, + 102 + ], + "spans": [ + { + "bbox": [ + 391, + 90, + 447, + 102 + ], + "type": "text", + "content": "SpectraM-PS" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 474, + 91, + 481, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 474, + 91, + 481, + 100 + ], + "spans": [ + { + "bbox": [ + 474, + 91, + 481, + 100 + ], + "type": "text", + "content": "7" + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 6 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 136, + 114, + 271, + 249 + ], + "blocks": [ + { + "bbox": [ + 136, + 114, + 271, + 249 + ], + "lines": [ + { + "bbox": [ + 136, + 114, + 271, + 249 + ], + "spans": [ + { + "bbox": [ + 136, + 114, + 271, + 249 + ], + "type": "image", + "image_path": "b695ec0cf40d59275318953451e4682acbb8053d33bf044e8a064af6b2c004d5.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 159, + 250, + 313, + 260 + ], + "lines": [ + { + "bbox": [ + 159, + 250, + 313, + 260 + ], + "spans": [ + { + "bbox": [ + 159, + 250, + 313, + 260 + ], + "type": "text", + "content": "SpectraM-PS on Temporally Multiplexed PS" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_caption" + } + ], + "index": 2 + }, + { + "type": "image", + "bbox": [ + 271, + 114, + 339, + 182 + ], + "blocks": [ + { + "bbox": [ + 271, + 114, + 339, + 182 + ], + "lines": [ + { + "bbox": [ + 271, + 114, + 339, + 182 + ], + "spans": [ + { + "bbox": [ + 271, + 114, + 339, + 182 + ], + "type": "image", + "image_path": "fa05781d3530d106be98b1c1a759a78a3f9e2397822eccebce576fd92afab5f3.jpg" + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "image_body" + } + ], + "index": 4 + }, + { + "type": "image", + "bbox": [ + 271, + 182, + 339, + 249 + ], + "blocks": [ + { + "bbox": [ + 271, + 182, + 339, + 249 + ], + "lines": [ + { + "bbox": [ + 271, + 182, + 339, + 249 + ], + "spans": [ + { + "bbox": [ + 271, + 182, + 339, + 249 + ], + "type": "image", + "image_path": "141a0aa36e3067d2608f640e4e7407f3b5d8868d002fd3de8643b69849f105f8.jpg" + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 131, + 270, + 482, + 304 + ], + "lines": [ + { + "bbox": [ + 131, + 270, + 482, + 304 + ], + "spans": [ + { + "bbox": [ + 131, + 270, + 482, + 304 + ], + "type": "text", + "content": "Fig. 3: (Left) Comparison of SpectraM-PS and SDM-UniPS [26] on six temporally multiplexed PS images. Due to the patch-wise basis of SpectraM-PS, fine details are better recovered. (Right) Illustration of different lighting conditions in PS-Multiplex." + } + ] + } + ], + "index": 11, + "angle": 0, + "type": "image_caption" + } + ], + "index": 5 + }, + { + "type": "image", + "bbox": [ + 342, + 114, + 409, + 182 + ], + "blocks": [ + { + "bbox": [ + 342, + 114, + 409, + 182 + ], + "lines": [ + { + "bbox": [ + 342, + 114, + 409, + 182 + ], + "spans": [ + { + "bbox": [ + 342, + 114, + 409, + 182 + ], + "type": "image", + "image_path": "42e96c0f5b3f02f09b242cde36d0d9b9489f1c69ee99d84b5609a77276f0f075.jpg" + } + ] + } + ], + "index": 6, + "angle": 0, + "type": "image_body" + } + ], + "index": 6 + }, + { + "type": "image", + "bbox": [ + 342, + 183, + 409, + 250 + ], + "blocks": [ + { + "bbox": [ + 342, + 183, + 409, + 250 + ], + "lines": [ + { + "bbox": [ + 342, + 183, + 409, + 250 + ], + "spans": [ + { + "bbox": [ + 342, + 183, + 409, + 250 + ], + "type": "image", + "image_path": "93339ba1241421b894148825b095f4c1f97caed0c5305c7a19dee2a300bb6b99.jpg" + } + ] + } + ], + "index": 7, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 360, + 251, + 466, + 260 + ], + "lines": [ + { + "bbox": [ + 360, + 251, + 466, + 260 + ], + "spans": [ + { + "bbox": [ + 360, + 251, + 466, + 260 + ], + "type": "text", + "content": "PS-Multiplex Training Dataset" + } + ] + } + ], + "index": 10, + "angle": 0, + "type": "image_caption" + } + ], + "index": 7 + }, + { + "type": "image", + "bbox": [ + 410, + 114, + 478, + 182 + ], + "blocks": [ + { + "bbox": [ + 410, + 114, + 478, + 182 + ], + "lines": [ + { + "bbox": [ + 410, + 114, + 478, + 182 + ], + "spans": [ + { + "bbox": [ + 410, + 114, + 478, + 182 + ], + "type": "image", + "image_path": "bbf1bf948329ec33c243adb4595a152a4f695154cc21a5114d95bf908edb8f92.jpg" + } + ] + } + ], + "index": 8, + "angle": 0, + "type": "image_body" + } + ], + "index": 8 + }, + { + "type": "image", + "bbox": [ + 410, + 183, + 478, + 250 + ], + "blocks": [ + { + "bbox": [ + 410, + 183, + 478, + 250 + ], + "lines": [ + { + "bbox": [ + 410, + 183, + 478, + 250 + ], + "spans": [ + { + "bbox": [ + 410, + 183, + 478, + 250 + ], + "type": "image", + "image_path": "56c548f0928b8d5f0123cbbfb2acef66aa11cac5be906d8cd2ff6d14704e0cc3.jpg" + } + ] + } + ], + "index": 9, + "angle": 0, + "type": "image_body" + } + ], + "index": 9 + }, + { + "bbox": [ + 130, + 316, + 479, + 363 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 316, + 479, + 363 + ], + "spans": [ + { + "bbox": [ + 130, + 316, + 479, + 363 + ], + "type": "text", + "content": "utilizing these datasets for training our model, leveraging the fact that our network does not distinguish images based on their physically-based principles. In this section, we highlight how RGB images serve as a practical approximation, simplifying the complexity inherent in multispectral imaging." + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 130, + 364, + 480, + 399 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 364, + 480, + 399 + ], + "spans": [ + { + "bbox": [ + 130, + 364, + 480, + 399 + ], + "type": "text", + "content": "We begin the discussion by characterizing multispectral imaging. Assuming that the surface doesn't emit light and only reflections on surface are considered, the image formation model is described as follows [35]:" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 183, + 406, + 481, + 433 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 183, + 406, + 481, + 433 + ], + "spans": [ + { + "bbox": [ + 183, + 406, + 481, + 433 + ], + "type": "interline_equation", + "content": "I _ {(s, p)} = \\int_ {\\Omega} \\left(\\omega_ {i} ^ {\\top} n _ {p}\\right) \\int_ {0} ^ {\\infty} S _ {s} (\\lambda) f _ {p} \\left(\\omega_ {i}, \\omega_ {o}, \\lambda\\right) L _ {p} \\left(\\omega_ {i}, \\lambda\\right) \\mathrm {d} \\lambda \\mathrm {d} \\omega_ {i}. \\tag {1}", + "image_path": "e6c7c1820c0dea63185f7d61c82f06f301e55b440625f03a45b78f2007a21743.jpg" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 130, + 437, + 482, + 557 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 437, + 482, + 557 + ], + "spans": [ + { + "bbox": [ + 130, + 437, + 482, + 557 + ], + "type": "text", + "content": "In this equation, " + }, + { + "bbox": [ + 130, + 437, + 482, + 557 + ], + "type": "inline_equation", + "content": "I_{(s,p)}" + }, + { + "bbox": [ + 130, + 437, + 482, + 557 + ], + "type": "text", + "content": " denotes the incoming spectral radiance at the sensor " + }, + { + "bbox": [ + 130, + 437, + 482, + 557 + ], + "type": "inline_equation", + "content": "s" + }, + { + "bbox": [ + 130, + 437, + 482, + 557 + ], + "type": "text", + "content": " (or " + }, + { + "bbox": [ + 130, + 437, + 482, + 557 + ], + "type": "inline_equation", + "content": "s" + }, + { + "bbox": [ + 130, + 437, + 482, + 557 + ], + "type": "text", + "content": "-th channel) from a surface point " + }, + { + "bbox": [ + 130, + 437, + 482, + 557 + ], + "type": "inline_equation", + "content": "p" + }, + { + "bbox": [ + 130, + 437, + 482, + 557 + ], + "type": "text", + "content": ". The term " + }, + { + "bbox": [ + 130, + 437, + 482, + 557 + ], + "type": "inline_equation", + "content": "f_p" + }, + { + "bbox": [ + 130, + 437, + 482, + 557 + ], + "type": "text", + "content": " represents BRDF, " + }, + { + "bbox": [ + 130, + 437, + 482, + 557 + ], + "type": "inline_equation", + "content": "L_p" + }, + { + "bbox": [ + 130, + 437, + 482, + 557 + ], + "type": "text", + "content": " the incident light intensity at the surface point, and " + }, + { + "bbox": [ + 130, + 437, + 482, + 557 + ], + "type": "inline_equation", + "content": "\\lambda" + }, + { + "bbox": [ + 130, + 437, + 482, + 557 + ], + "type": "text", + "content": " the wavelength of the incident light. The symbols " + }, + { + "bbox": [ + 130, + 437, + 482, + 557 + ], + "type": "inline_equation", + "content": "\\omega_i" + }, + { + "bbox": [ + 130, + 437, + 482, + 557 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 130, + 437, + 482, + 557 + ], + "type": "inline_equation", + "content": "\\omega_o" + }, + { + "bbox": [ + 130, + 437, + 482, + 557 + ], + "type": "text", + "content": " denote the directions of incident and reflected light, respectively. " + }, + { + "bbox": [ + 130, + 437, + 482, + 557 + ], + "type": "inline_equation", + "content": "S_s(\\lambda)" + }, + { + "bbox": [ + 130, + 437, + 482, + 557 + ], + "type": "text", + "content": " refers to the spectral sensitivity of the sensor " + }, + { + "bbox": [ + 130, + 437, + 482, + 557 + ], + "type": "inline_equation", + "content": "s" + }, + { + "bbox": [ + 130, + 437, + 482, + 557 + ], + "type": "text", + "content": " at wavelength " + }, + { + "bbox": [ + 130, + 437, + 482, + 557 + ], + "type": "inline_equation", + "content": "\\lambda" + }, + { + "bbox": [ + 130, + 437, + 482, + 557 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 130, + 437, + 482, + 557 + ], + "type": "inline_equation", + "content": "n_p" + }, + { + "bbox": [ + 130, + 437, + 482, + 557 + ], + "type": "text", + "content": " is the surface normal, and " + }, + { + "bbox": [ + 130, + 437, + 482, + 557 + ], + "type": "inline_equation", + "content": "\\varOmega" + }, + { + "bbox": [ + 130, + 437, + 482, + 557 + ], + "type": "text", + "content": " represents the hemisphere over which incident light directions are possible. The integral sums over all incident directions and wavelengths. It is important to note that the incident light intensity " + }, + { + "bbox": [ + 130, + 437, + 482, + 557 + ], + "type": "inline_equation", + "content": "L_p" + }, + { + "bbox": [ + 130, + 437, + 482, + 557 + ], + "type": "text", + "content": " depends not only on the direct contribution from light sources but also on the visibility of light (e.g., attached and cast shadows) and indirect illuminations." + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 130, + 558, + 482, + 667 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 558, + 482, + 667 + ], + "spans": [ + { + "bbox": [ + 130, + 558, + 482, + 667 + ], + "type": "text", + "content": "Eq. (1) illustrates the concept of spectral ambiguity, showing that an infinite number of combinations of " + }, + { + "bbox": [ + 130, + 558, + 482, + 667 + ], + "type": "inline_equation", + "content": "S_{s}(\\lambda)" + }, + { + "bbox": [ + 130, + 558, + 482, + 667 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 130, + 558, + 482, + 667 + ], + "type": "inline_equation", + "content": "f_{p}(\\omega_{i},\\omega_{o},\\lambda)" + }, + { + "bbox": [ + 130, + 558, + 482, + 667 + ], + "type": "text", + "content": ", and " + }, + { + "bbox": [ + 130, + 558, + 482, + 667 + ], + "type": "inline_equation", + "content": "L_{p}(\\omega_{i},\\lambda)" + }, + { + "bbox": [ + 130, + 558, + 482, + 667 + ], + "type": "text", + "content": " can result in the same spectral radiance, including narrowband compositions. In other words, with spectral ambiguity, a single observation " + }, + { + "bbox": [ + 130, + 558, + 482, + 667 + ], + "type": "inline_equation", + "content": "I_{(s,p)}" + }, + { + "bbox": [ + 130, + 558, + 482, + 667 + ], + "type": "text", + "content": " can encompass the observations for all spectral compositions that satisfy the equation (i.e., metamerism [22, 50]). This perspective justifies the theory of substituting multispectral images, which possess a broad parameter space, with narrowband RGB images. It is worth mentioning that channel crosstalk primarily affects the incident light intensity, consequently distorting the product of " + }, + { + "bbox": [ + 130, + 558, + 482, + 667 + ], + "type": "inline_equation", + "content": "S_{s}(\\lambda)\\cdot f_{p}(\\omega_{i},\\omega_{o},\\lambda)\\cdot L_{p}(\\omega_{i},\\lambda)" + }, + { + "bbox": [ + 130, + 558, + 482, + 667 + ], + "type": "text", + "content": " in Eq. (1)." + } + ] + } + ], + "index": 16 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "spans": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "type": "text", + "content": "8" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 166, + 91, + 267, + 101 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 166, + 91, + 267, + 101 + ], + "spans": [ + { + "bbox": [ + 166, + 91, + 267, + 101 + ], + "type": "text", + "content": "S. Ikehata and Y. Asano" + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 7 + }, + { + "para_blocks": [ + { + "bbox": [ + 130, + 115, + 482, + 357 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 115, + 482, + 357 + ], + "spans": [ + { + "bbox": [ + 130, + 115, + 482, + 357 + ], + "type": "text", + "content": "This implies that observations influenced by spectral crosstalk can still be equivalently represented using a narrowband setup under spectral ambiguity. In the experiments, we demonstrate that our model, trained on three narrowband observations can be applied to multiplexed data with channel crosstalk. To realize this approximation, we rendered a large number of three-channel narrowband images using the path-tracing algorithm in Blender [4], where up to 10-bounce reflections are permitted, based on common 3D assets [2] for RGB rendering. Following the rendering pipeline described in [26], we rendered objects by combining three different lighting models: directional, point, and environmental (five combinatorial settings in total as shown in Fig. 3). To simulate spectrally multiplexed images, we defined R, G, and B light sources and illuminated the surface in a multiplexed manner. It is important to note that the rendered RGB images are decomposed into three grayscale images, each of which was independently fed into the network; therefore, any wavelength-dependent information is masked. For material diversity, we adopted the method from [26], categorizing 897 AdobeStock texture maps into three groups: 421 diffuse, 219 specular, and 257 metallic textures. Four objects from a set of 410 3D AdobeStock models were randomly selected and textured with these materials. This structured approach led to the rendering of 106,374 multiplexed images along with their ground truth surface normal maps, forming the 'PS-Multiplex' dataset." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 132, + 372, + 347, + 385 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 372, + 347, + 385 + ], + "spans": [ + { + "bbox": [ + 132, + 372, + 347, + 385 + ], + "type": "text", + "content": "5 SpectraM14 Benchmark Dataset" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 130, + 396, + 482, + 467 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 396, + 482, + 467 + ], + "spans": [ + { + "bbox": [ + 130, + 396, + 482, + 467 + ], + "type": "text", + "content": "Due to the lack of a benchmark for spectrally multiplexed PS, the first comprehensive evaluation dataset, named SpectraM14, is created. This dataset includes 14 objects, each exhibiting a range of optical properties such as monochromatic or multicolored appearances and diffuse or specular reflections, as depicted in Fig. 4. Our benchmark encompasses tasks under five distinct conditions, as described later." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 130, + 471, + 482, + 628 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 471, + 482, + 628 + ], + "spans": [ + { + "bbox": [ + 130, + 471, + 482, + 628 + ], + "type": "text", + "content": "Imaging Setup. To acquire our dataset, we utilized a color camera (FLIR GS3-U3-123S6C-C) and an NIR camera (FLIR GS3-U3-41C6NIR-C), both equipped with a " + }, + { + "bbox": [ + 130, + 471, + 482, + 628 + ], + "type": "inline_equation", + "content": "50\\mathrm{mm}" + }, + { + "bbox": [ + 130, + 471, + 482, + 628 + ], + "type": "text", + "content": " lens. For the NIR camera, we used narrowband filters with wavelengths of " + }, + { + "bbox": [ + 130, + 471, + 482, + 628 + ], + "type": "inline_equation", + "content": "750\\mathrm{nm}" + }, + { + "bbox": [ + 130, + 471, + 482, + 628 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 130, + 471, + 482, + 628 + ], + "type": "inline_equation", + "content": "850\\mathrm{nm}" + }, + { + "bbox": [ + 130, + 471, + 482, + 628 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 130, + 471, + 482, + 628 + ], + "type": "inline_equation", + "content": "880\\mathrm{nm}" + }, + { + "bbox": [ + 130, + 471, + 482, + 628 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 130, + 471, + 482, + 628 + ], + "type": "inline_equation", + "content": "905\\mathrm{nm}" + }, + { + "bbox": [ + 130, + 471, + 482, + 628 + ], + "type": "text", + "content": ", and " + }, + { + "bbox": [ + 130, + 471, + 482, + 628 + ], + "type": "inline_equation", + "content": "940\\mathrm{nm}" + }, + { + "bbox": [ + 130, + 471, + 482, + 628 + ], + "type": "text", + "content": ", and the acquired images were manually merged. Objects were placed " + }, + { + "bbox": [ + 130, + 471, + 482, + 628 + ], + "type": "inline_equation", + "content": "0.8\\mathrm{m}" + }, + { + "bbox": [ + 130, + 471, + 482, + 628 + ], + "type": "text", + "content": " from the camera to approximate orthographic projection. Following conventional PS benchmarks [52,56,63], data capture occurred in a controlled, dark environment with the scene draped in black cloth to mitigate interreflection. The camera's ISO sensitivity was minimized to enhance image quality. The imaging area was further isolated using low-reflectance cloths to suppress inter-reflection. For each illumination condition, we collected six images under varying exposures to produce HDR input images. For the evaluations throughout this paper, the images are cropped using an object mask and resized to " + }, + { + "bbox": [ + 130, + 471, + 482, + 628 + ], + "type": "inline_equation", + "content": "512\\mathrm{px} \\times 512\\mathrm{px}" + }, + { + "bbox": [ + 130, + 471, + 482, + 628 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 130, + 629, + 482, + 666 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 629, + 482, + 666 + ], + "spans": [ + { + "bbox": [ + 130, + 629, + 482, + 666 + ], + "type": "text", + "content": "Lighting. Six LED and three halogen light sources, positioned roughly 1 meter from the object, provided illumination. We used the \"Weeylite S05 RGB Pocket Lamp\" and the \"NPI PIS-UHX-AIR\" for lighting. This setup enabled the use of" + } + ] + } + ], + "index": 6 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 392, + 91, + 447, + 102 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 392, + 91, + 447, + 102 + ], + "spans": [ + { + "bbox": [ + 392, + 91, + 447, + 102 + ], + "type": "text", + "content": "SpectraM-PS" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 474, + 92, + 481, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 474, + 92, + 481, + 100 + ], + "spans": [ + { + "bbox": [ + 474, + 92, + 481, + 100 + ], + "type": "text", + "content": "9" + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 8 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 136, + 114, + 478, + 256 + ], + "blocks": [ + { + "bbox": [ + 136, + 114, + 478, + 256 + ], + "lines": [ + { + "bbox": [ + 136, + 114, + 478, + 256 + ], + "spans": [ + { + "bbox": [ + 136, + 114, + 478, + 256 + ], + "type": "image", + "image_path": "4df85b5950b772f90a0bd59842858eea97972ceb5bd5cca701904cfbd75c86e8.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 241, + 259, + 371, + 271 + ], + "lines": [ + { + "bbox": [ + 241, + 259, + 371, + 271 + ], + "spans": [ + { + "bbox": [ + 241, + 259, + 371, + 271 + ], + "type": "text", + "content": "Fig. 4: Objects in SpectraM14." + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_caption" + } + ], + "index": 2 + }, + { + "type": "image", + "bbox": [ + 138, + 283, + 475, + 352 + ], + "blocks": [ + { + "bbox": [ + 138, + 283, + 475, + 352 + ], + "lines": [ + { + "bbox": [ + 138, + 283, + 475, + 352 + ], + "spans": [ + { + "bbox": [ + 138, + 283, + 475, + 352 + ], + "type": "image", + "image_path": "a034460c5348bb0c6d6a9aa4054cb7875debf186dcb33da0567907bcb923e1ea.jpg" + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 200, + 357, + 411, + 369 + ], + "lines": [ + { + "bbox": [ + 200, + 357, + 411, + 369 + ], + "spans": [ + { + "bbox": [ + 200, + 357, + 411, + 369 + ], + "type": "text", + "content": "Fig. 5: Illustration of six conditions in SpectraM14." + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "image_caption" + } + ], + "index": 4 + }, + { + "bbox": [ + 130, + 394, + 480, + 418 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 394, + 480, + 418 + ], + "spans": [ + { + "bbox": [ + 130, + 394, + 480, + 418 + ], + "type": "text", + "content": "red, green, blue, yellow, magenta, cyan, and NIR lighting, with spectra validated using a Hamamatsu Photonics Multichannel Analyzer C10027-01." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 130, + 422, + 482, + 481 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 422, + 482, + 481 + ], + "spans": [ + { + "bbox": [ + 130, + 422, + 482, + 481 + ], + "type": "text", + "content": "Calibration and Ground Truth Data. We measured the directions of lights using specular reflections from a mirror sphere. Light intensity was standardized across the visible spectrum by averaging RGB values from reflected light on a white target. The ground truth normals were captured with a SHINING 3D EinScan-SE scanner." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 130, + 486, + 482, + 666 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 486, + 482, + 666 + ], + "spans": [ + { + "bbox": [ + 130, + 486, + 482, + 666 + ], + "type": "text", + "content": "Evaluation Procedure. The design philosophy of this benchmark is to assess the robustness and adaptability of spectrally multiplexed PS methods under realistic lighting conditions, accounting for variations in channel numbers and the presence of spectral crosstalk. For a comprehensive evaluation, we designed tasks under five distinct conditions as shown in Fig. 5: Condition 1: Color sensor, no crosstalk condition: Six colors of light (red, green, blue, cyan, yellow, magenta) were each independently illuminated and observed with an RGB sensor. Afterward, the channels of RGB were averaged. Condition 2: Color sensor, weak crosstalk condition: Three colors of light (red, green, blue) were simultaneously illuminated and observed through each channel of the RGB sensor. Condition 3: Color sensor, strong crosstalk condition: Three colors of light (cyan, yellow, magenta) were simultaneously illuminated and observed through each channel of the RGB sensor. Condition 4: NIR sensor, no crosstalk condition: Light at wavelengths of " + }, + { + "bbox": [ + 130, + 486, + 482, + 666 + ], + "type": "inline_equation", + "content": "750~\\mathrm{nm}" + }, + { + "bbox": [ + 130, + 486, + 482, + 666 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 130, + 486, + 482, + 666 + ], + "type": "inline_equation", + "content": "850~\\mathrm{nm}" + }, + { + "bbox": [ + 130, + 486, + 482, + 666 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 130, + 486, + 482, + 666 + ], + "type": "inline_equation", + "content": "880~\\mathrm{nm}" + }, + { + "bbox": [ + 130, + 486, + 482, + 666 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 130, + 486, + 482, + 666 + ], + "type": "inline_equation", + "content": "905~\\mathrm{nm}" + }, + { + "bbox": [ + 130, + 486, + 482, + 666 + ], + "type": "text", + "content": ", and " + }, + { + "bbox": [ + 130, + 486, + 482, + 666 + ], + "type": "inline_equation", + "content": "940~\\mathrm{nm}" + }, + { + "bbox": [ + 130, + 486, + 482, + 666 + ], + "type": "text", + "content": " were each independently illuminated and observed with a monochrome sensor corresponding to" + } + ] + } + ], + "index": 8 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "spans": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "type": "text", + "content": "10" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 166, + 91, + 267, + 101 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 166, + 91, + 267, + 101 + ], + "spans": [ + { + "bbox": [ + 166, + 91, + 267, + 101 + ], + "type": "text", + "content": "S. Ikehata and Y. Asano" + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 9 + }, + { + "para_blocks": [ + { + "bbox": [ + 134, + 116, + 480, + 163 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 134, + 116, + 480, + 163 + ], + "spans": [ + { + "bbox": [ + 134, + 116, + 480, + 163 + ], + "type": "text", + "content": "each wavelength. Condition 5: NIR sensor, spatially-varying lighting condition: New images were created by averaging two images taken under the conditions mentioned above. The combinations were (750 nm, 850 nm), (850 nm, 880 nm), (880 nm, 905 nm), (905 nm, 940 nm), and (940 nm, 750 nm)." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 134, + 185, + 222, + 198 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 134, + 185, + 222, + 198 + ], + "spans": [ + { + "bbox": [ + 134, + 185, + 222, + 198 + ], + "type": "text", + "content": "6 Experiment" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 134, + 212, + 480, + 331 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 134, + 212, + 480, + 331 + ], + "spans": [ + { + "bbox": [ + 134, + 212, + 480, + 331 + ], + "type": "text", + "content": "In this section, we evaluate our method on our SpectraM14. Our method is compared with one SOTA optimization-based method [17] and one SOTA learning-based method [46]. The former introduces a closed-form solution for spectrally multiplexed photometric stereo applied to monochromatic surfaces with spatially varying (SV) albedo. The latter presents a Spectral Reflectance Decomposition (SRD) model, which disentangles spectral reflectance into geometric and spectral components for surface normal recovery under non-Lambertian spectral reflectance conditions. Unlike the compared methods, our approach does not assume a specific lighting setup, whereas both methods presume the presence of calibrated single directional light sources." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 134, + 335, + 480, + 467 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 134, + 335, + 480, + 467 + ], + "spans": [ + { + "bbox": [ + 134, + 335, + 480, + 467 + ], + "type": "text", + "content": "Training details. SpectraM-PS was trained from scratch on the PS-Multiplex dataset until convergence using the AdamW optimizer, with a step decay learning rate schedule that reduced the learning rate by a factor of 0.8 every ten epochs. We applied learning rate warmup during the first epoch and used a batch size of 16, an initial learning rate of 0.0001, and a weight decay of 0.05. Each batch consisted of three input training multiplexed images with three channels each. The training loss was computed using the Mean Squared Error (MSE) loss function to measure " + }, + { + "bbox": [ + 134, + 335, + 480, + 467 + ], + "type": "inline_equation", + "content": "\\ell_2" + }, + { + "bbox": [ + 134, + 335, + 480, + 467 + ], + "type": "text", + "content": " errors between the predicted surface normal vectors and the ground truth surface normal vectors. We measured the reconstruction accuracy of our method by computing the mean angular errors (MAE) between the predicted and true surface normal maps, expressed in degrees." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 134, + 470, + 480, + 554 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 134, + 470, + 480, + 554 + ], + "spans": [ + { + "bbox": [ + 134, + 470, + 480, + 554 + ], + "type": "text", + "content": "Computational Cost. The inference time of PS methods varies with the number of pixels and channels in the input image. For Condition 2 and 3 with a " + }, + { + "bbox": [ + 134, + 470, + 480, + 554 + ], + "type": "inline_equation", + "content": "512 \\times 512 \\times 3" + }, + { + "bbox": [ + 134, + 470, + 480, + 554 + ], + "type": "text", + "content": " image, the mean and standard deviation of inference times (in sec) over 14 objects in SpectraM14 benchmark were: our method (3.42/0.85), Lv et al. [46] (0.46/0.24) and Guo et al. [17] (2.38/1.10). Our architecture leads to higher computational costs; however, none of the methods were suitable for real-time processing (e.g., 15 fps requires 0.06 sec/frame)." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 134, + 557, + 480, + 665 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 134, + 557, + 480, + 665 + ], + "spans": [ + { + "bbox": [ + 134, + 557, + 480, + 665 + ], + "type": "text", + "content": "Ablation Study. We firstly validate the individual technical contributions of our training dataset (i.e., PS-Multiplex) and the physics-free architecture (i.e., SpectraM-PS) using a synthetic evaluation dataset. Firstly, we validate the efficacy of our training dataset, PS-Multiplex, by adapting an existing universal photometric stereo architecture designed for the conventional task (i.e., SDM-UniPS [26]) to the spectrally multiplexed photometric stereo task. Since both ours and SDM-UniPS take multiple observations and an object mask as input, this adaptation straightforwardly involves training the model on PS-Multiplex by treating each channel of an image as an individual image. Subsequently, we" + } + ] + } + ], + "index": 7 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 392, + 91, + 447, + 102 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 392, + 91, + 447, + 102 + ], + "spans": [ + { + "bbox": [ + 392, + 91, + 447, + 102 + ], + "type": "text", + "content": "SpectraM-PS" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 470, + 92, + 479, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 470, + 92, + 479, + 100 + ], + "spans": [ + { + "bbox": [ + 470, + 92, + 479, + 100 + ], + "type": "text", + "content": "11" + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 10 + }, + { + "para_blocks": [ + { + "type": "table", + "bbox": [ + 141, + 130, + 474, + 193 + ], + "blocks": [ + { + "bbox": [ + 136, + 114, + 475, + 126 + ], + "lines": [ + { + "bbox": [ + 136, + 114, + 475, + 126 + ], + "spans": [ + { + "bbox": [ + 136, + 114, + 475, + 126 + ], + "type": "text", + "content": "Table 1: Ablation analysis of the contributions of SpectraM-PS and PS-Multiplex." + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 141, + 130, + 474, + 193 + ], + "lines": [ + { + "bbox": [ + 141, + 130, + 474, + 193 + ], + "spans": [ + { + "bbox": [ + 141, + 130, + 474, + 193 + ], + "type": "table", + "html": "
MAE (Uniform)MAE (Piece-wise uniform)MAE (Non-uniform)
MethodNon-LambertianLambertianNon-LambertianLambertianNon-LambertianLambertian
SDM-UniPS [26]12.9 (4.5)12.4 (4.7)15.0 (5.0)21.7 (7.0)14.4 (5.1)15.2 (5.9)
[26] trained on PS-Multiplex11.1 (3.6)11.0 (3.9)10.5 (2.9)12.3 (3.9)10.6 (3.9)11.2 (3.7)
SpectraM-PS (Ours)8.0 (2.7)8.4 (2.7)7.9 (2.4)8.9 (2.9)8.2 (3.2)8.0 (2.5)
", + "image_path": "7d60b929da37779a5e73b98fd496accafaf28eea73ae072dfe6e13321f3d0107.jpg" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "table_body" + } + ], + "index": 3 + }, + { + "bbox": [ + 130, + 205, + 479, + 228 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 205, + 479, + 228 + ], + "spans": [ + { + "bbox": [ + 130, + 205, + 479, + 228 + ], + "type": "text", + "content": "compare this model against our proposed SpectraM-PS to demonstrate the efficacy of our dual-scale design with local patch embedding." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 130, + 230, + 482, + 458 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 230, + 482, + 458 + ], + "spans": [ + { + "bbox": [ + 130, + 230, + 482, + 458 + ], + "type": "text", + "content": "For evaluating the contribution of our architecture (SpectraM-PS) and training dataset (PS-Multiplex), we additionally rendered three-channel spectrally multiplexed images representing six distinct surface material categories: (a) uniform, Lambertian; (b) piece-wise uniform, Lambertian; (c) non-uniform, Lambertian; (d) uniform, non-Lambertian; (e) piece-wise uniform, non-Lambertian; and (f) non-uniform, non-Lambertian. In uniform materials, every point on the surface within a scene exhibits the same material properties. For piece-wise uniform materials, each object in a scene is composed of the same material, yet different objects possess distinct materials. Non-uniform materials feature unique PBR textures assigned to each object. The rendering process for these images was identical to that used for the PS-Multiplex datasets in each category. We generated 100 scenes for each surface material category, and MAEs (stds) are averaged over them. The results are presented in Tab.1. In summary, SDM-UniPS [26] trained on our PS-Multiplex dataset demonstrates proper adaptation to the spectrally multiplexed photometric stereo task. Nonetheless, our SpectraM-PS method significantly enhanced reconstruction accuracy, showcasing an architecture-level improvement over SDM-UniPS for the spectrally multiplexed photometric stereo task, where the number of input channels is typically much fewer than that of input images for conventional PS." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 130, + 462, + 492, + 473 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 462, + 492, + 473 + ], + "spans": [ + { + "bbox": [ + 130, + 462, + 492, + 473 + ], + "type": "text", + "content": "Comparative Evaluation on SpectraM14. The results are illustrated in Tabs. 2" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 130, + 474, + 482, + 666 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 474, + 482, + 666 + ], + "spans": [ + { + "bbox": [ + 130, + 474, + 482, + 666 + ], + "type": "text", + "content": "to 6 and Fig. 6. Despite the fact that all existing spectrally multiplexed photometric stereo methods assume calibrated light sources and known directional light source conditions, our proposed method significantly outperformed them. This is because most of the real objects used in our experiment are neither Lambertian nor convex, and do not conform to their assumptions. However, our non-physical-based method successfully restored the normals very stably for these objects. Furthermore, our proposed method enabled robust reconstruction for all objects, despite having been trained only with RGB color images. This result supports the efficacy of our approximation. Furthermore, unlike existing methods that suffer from reduced estimation accuracy with increasing spectral crosstalk, our approach demonstrates only minimal performance degradation. Remarkably, our method excels in recovering a more realistic structure with spatially-varying surface materials. This breakthrough implies that our network can effectively achieve dynamic surface reconstruction across video frames in a universal setting. We will detail this groundbreaking application in the next section. Due to space constraints, not all results can be included here. However, all" + } + ] + } + ], + "index": 7 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "spans": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "type": "text", + "content": "12" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 166, + 91, + 267, + 101 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 166, + 91, + 267, + 101 + ], + "spans": [ + { + "bbox": [ + 166, + 91, + 267, + 101 + ], + "type": "text", + "content": "S. Ikehata and Y. Asano" + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 11 + }, + { + "para_blocks": [ + { + "type": "table", + "bbox": [ + 135, + 127, + 479, + 171 + ], + "blocks": [ + { + "bbox": [ + 134, + 114, + 478, + 125 + ], + "lines": [ + { + "bbox": [ + 134, + 114, + 478, + 125 + ], + "spans": [ + { + "bbox": [ + 134, + 114, + 478, + 125 + ], + "type": "text", + "content": "Table 2: Comparison in condition 1. The values are mean angular errors in degrees." + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 135, + 127, + 479, + 171 + ], + "lines": [ + { + "bbox": [ + 135, + 127, + 479, + 171 + ], + "spans": [ + { + "bbox": [ + 135, + 127, + 479, + 171 + ], + "type": "table", + "html": "
MethodObject ID
1234567891011121314Ave.
Ours8.611.213.07.312.26.111.55.05.77.510.35.16.112.28.9
Lv et al. [46]20.417.021.113.923.110.721.216.610.915.919.016.413.218.617.1
Guo et al. [17]22.615.220.713.427.17.231.324.88.018.224.511.410.229.318.9
", + "image_path": "aff885e0dceeef0084fcc05957627766ce0989c3b587dbef08f05a98dd3d0e47.jpg" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "table_body" + } + ], + "index": 3 + }, + { + "type": "table", + "bbox": [ + 134, + 191, + 480, + 239 + ], + "blocks": [ + { + "bbox": [ + 135, + 181, + 478, + 191 + ], + "lines": [ + { + "bbox": [ + 135, + 181, + 478, + 191 + ], + "spans": [ + { + "bbox": [ + 135, + 181, + 478, + 191 + ], + "type": "text", + "content": "Table 3: Comparison in condition 2. The values are mean angular errors in degrees." + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 134, + 191, + 480, + 239 + ], + "lines": [ + { + "bbox": [ + 134, + 191, + 480, + 239 + ], + "spans": [ + { + "bbox": [ + 134, + 191, + 480, + 239 + ], + "type": "table", + "html": "
MethodObject IDAve.
1234567891011121314
Ours10.111.513.59.012.87.010.95.56.39.312.75.39.814.710.0
Lv et al. [46]22.826.027.019.430.319.522.118.914.319.823.520.819.021.421.7
Guo et al. [17]31.127.529.220.138.019.033.423.513.626.632.714.717.139.325.7
", + "image_path": "ab636ee77231ad6daaa269676db9e7373589cf893351c3f50537a1f45e0b13cc.jpg" + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "table_body" + } + ], + "index": 5 + }, + { + "type": "table", + "bbox": [ + 134, + 258, + 480, + 305 + ], + "blocks": [ + { + "bbox": [ + 135, + 247, + 478, + 258 + ], + "lines": [ + { + "bbox": [ + 135, + 247, + 478, + 258 + ], + "spans": [ + { + "bbox": [ + 135, + 247, + 478, + 258 + ], + "type": "text", + "content": "Table 4: Comparison in condition 3. The values are mean angular errors in degrees." + } + ] + } + ], + "index": 6, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 134, + 258, + 480, + 305 + ], + "lines": [ + { + "bbox": [ + 134, + 258, + 480, + 305 + ], + "spans": [ + { + "bbox": [ + 134, + 258, + 480, + 305 + ], + "type": "table", + "html": "
MethodObject ID
1234567891011121314
Ours12.012.815.811.016.55.912.39.86.99.814.66.37.520.1
Lv et al. [46]38.534.138.232.440.138.236.629.638.235.138.236.638.430.9
Guo et al. [17]46.042.656.337.657.145.676.048.029.962.149.850.252.873.7
", + "image_path": "e02131d72b8805ec7483481a572a0cbd148e52fa176630fffbee403090147a4d.jpg" + } + ] + } + ], + "index": 7, + "angle": 0, + "type": "table_body" + } + ], + "index": 7 + }, + { + "type": "table", + "bbox": [ + 135, + 324, + 478, + 371 + ], + "blocks": [ + { + "bbox": [ + 135, + 314, + 478, + 323 + ], + "lines": [ + { + "bbox": [ + 135, + 314, + 478, + 323 + ], + "spans": [ + { + "bbox": [ + 135, + 314, + 478, + 323 + ], + "type": "text", + "content": "Table 5: Comparison in condition 4. The values are mean angular errors in degrees." + } + ] + } + ], + "index": 8, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 135, + 324, + 478, + 371 + ], + "lines": [ + { + "bbox": [ + 135, + 324, + 478, + 371 + ], + "spans": [ + { + "bbox": [ + 135, + 324, + 478, + 371 + ], + "type": "table", + "html": "
MethodObject ID
1234567891011121314
Ours10.911.39.76.310.85.312.44.37.86.69.14.07.410.6
Lv et al. [46]24.119.820.712.019.812.415.118.711.417.121.217.519.116.7
Guo et al. [17]30.716.018.39.929.68.423.125.710.214.624.513.929.613.8
", + "image_path": "1eed9fd05c367f866fd35c947db5a179ee59b9d48a428b7e75017e8f51e07ee1.jpg" + } + ] + } + ], + "index": 9, + "angle": 0, + "type": "table_body" + } + ], + "index": 9 + }, + { + "type": "table", + "bbox": [ + 136, + 391, + 478, + 437 + ], + "blocks": [ + { + "bbox": [ + 135, + 380, + 478, + 390 + ], + "lines": [ + { + "bbox": [ + 135, + 380, + 478, + 390 + ], + "spans": [ + { + "bbox": [ + 135, + 380, + 478, + 390 + ], + "type": "text", + "content": "Table 6: Comparison in condition 5. The values are mean angular errors in degrees." + } + ] + } + ], + "index": 10, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 136, + 391, + 478, + 437 + ], + "lines": [ + { + "bbox": [ + 136, + 391, + 478, + 437 + ], + "spans": [ + { + "bbox": [ + 136, + 391, + 478, + 437 + ], + "type": "table", + "html": "
MethodObject IDAve.
1234567891011121314
Ours10.911.09.97.311.04.713.34.87.96.59.84.27.410.48.6
Lv et al. [46]29.826.228.922.828.925.821.924.618.724.727.624.822.327.625.0
Guo et al. [17]40.027.329.223.233.325.029.330.120.026.731.626.325.027.227.8
", + "image_path": "032798ec3ac385d6807f53395969df4695b0f5d07b2ebef832474a113915ce61.jpg" + } + ] + } + ], + "index": 11, + "angle": 0, + "type": "table_body" + } + ], + "index": 11 + }, + { + "bbox": [ + 130, + 451, + 482, + 499 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 451, + 482, + 499 + ], + "spans": [ + { + "bbox": [ + 130, + 451, + 482, + 499 + ], + "type": "text", + "content": "results are comprehensively presented in the supplementary materials. Additionally, the supplementary materials evaluate the impact of the spatial distribution of light sources on the performance of the proposed method. We also offer an in-depth discussion of each experimental condition therein." + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 132, + 519, + 220, + 532 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 519, + 220, + 532 + ], + "spans": [ + { + "bbox": [ + 132, + 519, + 220, + 532 + ], + "type": "text", + "content": "7 Conclusion" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 130, + 545, + 482, + 666 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 545, + 482, + 666 + ], + "spans": [ + { + "bbox": [ + 130, + 545, + 482, + 666 + ], + "type": "text", + "content": "In this work, we introduce an innovative approach to spectrally multiplexed photometric stereo under unknown spatial/spectral composition. Turning spectral ambiguity into a benefit, our method allows for the creation of training data without the need for complex multispectral rendering. Our work significantly broadens the scope for dynamic surface analysis, establishing a critical advancement in the utilization of photometric stereo across multiple sectors. Our proposed method exhibits several limitations. Firstly, there is unstable temporal variation in the normal maps reconstructed by our method for dynamic surface reconstruction. This instability arises from factors such as motion blur in certain frames, image noise, or the influence of cast/attached shadows, which become" + } + ] + } + ], + "index": 14 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 392, + 91, + 447, + 102 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 392, + 91, + 447, + 102 + ], + "spans": [ + { + "bbox": [ + 392, + 91, + 447, + 102 + ], + "type": "text", + "content": "SpectraM-PS" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 470, + 91, + 481, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 470, + 91, + 481, + 100 + ], + "spans": [ + { + "bbox": [ + 470, + 91, + 481, + 100 + ], + "type": "text", + "content": "13" + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 12 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 140, + 116, + 476, + 541 + ], + "blocks": [ + { + "bbox": [ + 140, + 116, + 476, + 541 + ], + "lines": [ + { + "bbox": [ + 140, + 116, + 476, + 541 + ], + "spans": [ + { + "bbox": [ + 140, + 116, + 476, + 541 + ], + "type": "image", + "image_path": "f141a01604beedf219d31153145f94f5220479126cd51437fba78600b9794178.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 138, + 549, + 474, + 562 + ], + "lines": [ + { + "bbox": [ + 138, + 549, + 474, + 562 + ], + "spans": [ + { + "bbox": [ + 138, + 549, + 474, + 562 + ], + "type": "text", + "content": "Fig. 6: Evaluation on SpectraM14. Full results are available in the supplementary." + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_caption" + } + ], + "index": 2 + }, + { + "bbox": [ + 130, + 575, + 482, + 660 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 575, + 482, + 660 + ], + "spans": [ + { + "bbox": [ + 130, + 575, + 482, + 660 + ], + "type": "text", + "content": "more pronounced compared to conventional photometric stereo methods that utilize numerous images. To recover clean and temporally stable normal maps, we may need to consider temporal consistency and more actively utilize monocular cues. Additionally, while our method targets dynamic surfaces, it currently requires several seconds to up to ten seconds per RGB image, which is far from real-time processing. Considering industrial applications in the future, accelerating the processing speed is a crucial challenge." + } + ] + } + ], + "index": 4 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "spans": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "type": "text", + "content": "14" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 166, + 91, + 267, + 101 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 166, + 91, + 267, + 101 + ], + "spans": [ + { + "bbox": [ + 166, + 91, + 267, + 101 + ], + "type": "text", + "content": "S. Ikehata and Y. Asano" + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 13 + }, + { + "para_blocks": [ + { + "bbox": [ + 132, + 114, + 198, + 126 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 114, + 198, + 126 + ], + "spans": [ + { + "bbox": [ + 132, + 114, + 198, + 126 + ], + "type": "text", + "content": "References" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 134, + 141, + 481, + 665 + ], + "type": "list", + "angle": 0, + "index": 24, + "blocks": [ + { + "bbox": [ + 138, + 141, + 481, + 174 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 141, + 481, + 174 + ], + "spans": [ + { + "bbox": [ + 138, + 141, + 481, + 174 + ], + "type": "text", + "content": "1. 3D Textures - Free seamless PBR textures with Diffuse, Normal, Displacement, Occlusion, Specularity and Roughness Maps. https://3dtextures.me/, accessed: 2024-03-07 7" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 138, + 175, + 341, + 186 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 175, + 341, + 186 + ], + "spans": [ + { + "bbox": [ + 138, + 175, + 341, + 186 + ], + "type": "text", + "content": "2. Adobe Stock. https://stock.adobe.com/ 7,9" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 138, + 186, + 481, + 207 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 186, + 481, + 207 + ], + "spans": [ + { + "bbox": [ + 138, + 186, + 481, + 207 + ], + "type": "text", + "content": "3. AmbientCG - Free Public Domain PBR Materials. https://ambientcg.com/, accessed: 2024-03-07 7" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 138, + 209, + 310, + 219 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 209, + 310, + 219 + ], + "spans": [ + { + "bbox": [ + 138, + 209, + 310, + 219 + ], + "type": "text", + "content": "4.Blender.https://www.blender.org/9" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 138, + 220, + 481, + 242 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 220, + 481, + 242 + ], + "spans": [ + { + "bbox": [ + 138, + 220, + 481, + 242 + ], + "type": "text", + "content": "5. Poliigon - A library of materials, and HDR's for artists including free textures. https://www.poliigon.com/, accessed: 2024-03-07 7" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 138, + 243, + 481, + 264 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 243, + 481, + 264 + ], + "spans": [ + { + "bbox": [ + 138, + 243, + 481, + 264 + ], + "type": "text", + "content": "6. Alldrin, N., Mallick, S., Kriegman, D.: Resolving the generalized bas-relief ambiguity by entropy minimization. CVPR (2007) 4" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 138, + 265, + 481, + 286 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 265, + 481, + 286 + ], + "spans": [ + { + "bbox": [ + 138, + 265, + 481, + 286 + ], + "type": "text", + "content": "7. Anderson, R., Stenger, B., Cipolla, R.: Color photometric stereo for multicolored surfaces. ICCV (2011) 2" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 138, + 287, + 481, + 309 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 287, + 481, + 309 + ], + "spans": [ + { + "bbox": [ + 138, + 287, + 481, + 309 + ], + "type": "text", + "content": "8. Basri, R., Jacobs, D., Kemelmacher, I.: Photometric stereo with general, unknown lighting. International Journal of computer vision 72(3), 239-257 (2007) 4" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 138, + 310, + 481, + 341 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 310, + 481, + 341 + ], + "spans": [ + { + "bbox": [ + 138, + 310, + 481, + 341 + ], + "type": "text", + "content": "9. Ben-David, S., Blitzer, J., Crammer, K., Kulesza, A., Pereira, F., Vaughan, J.W.: A theory of learning from different domains. Machine learning 79, 151-175 (2010) 7" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 134, + 342, + 481, + 375 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 134, + 342, + 481, + 375 + ], + "spans": [ + { + "bbox": [ + 134, + 342, + 481, + 375 + ], + "type": "text", + "content": "10. Chakrabarti, A., Sunkavalli, K.: Single-image rgb photometric stereo with spatially-varying albedo. In: 2016 Fourth International Conference on 3D Vision (3DV) (2016) 2" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 134, + 376, + 481, + 408 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 134, + 376, + 481, + 408 + ], + "spans": [ + { + "bbox": [ + 134, + 376, + 481, + 408 + ], + "type": "text", + "content": "11. Chen, G., Han, K., Shi, B., Matsushita, Y., Wong, K.K.K.: Self-calibrating deep photometric stereo networks. In: 2019 IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR). pp. 8731-8739 (2019) 2, 4, 6" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 134, + 410, + 481, + 430 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 134, + 410, + 481, + 430 + ], + "spans": [ + { + "bbox": [ + 134, + 410, + 481, + 430 + ], + "type": "text", + "content": "12. Chen, G., Han, K., Wong, K.Y.K.: Ps-fcn: A flexible learning framework for photometric stereo. ECCV (2018) 2, 4, 6" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 134, + 432, + 481, + 464 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 134, + 432, + 481, + 464 + ], + "spans": [ + { + "bbox": [ + 134, + 432, + 481, + 464 + ], + "type": "text", + "content": "13. Chen, G., Waechter, M., Shi, B., Wong, K.Y.K., Matsushita, Y.: What is learned in deep uncalibrated photometric stereo? In: European Conference on Computer Vision. pp. 745-762. Springer (2020) 4, 6" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 134, + 465, + 481, + 487 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 134, + 465, + 481, + 487 + ], + "spans": [ + { + "bbox": [ + 134, + 465, + 481, + 487 + ], + "type": "text", + "content": "14. Decker, B., Kautz, J., Mertens, T., Bekaert, P.: Capturing multiple illumination conditions using time and color multiplexing. CVPR (2009) 2" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 134, + 487, + 481, + 509 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 134, + 487, + 481, + 509 + ], + "spans": [ + { + "bbox": [ + 134, + 487, + 481, + 509 + ], + "type": "text", + "content": "15. Drew, M.S.: Shape from color. Technical Report CSS/LCCR TR 92-07, School of Computing Science, Simon Fraser University, Vancouver, BC (1992) 4" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 134, + 510, + 481, + 531 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 134, + 510, + 481, + 531 + ], + "spans": [ + { + "bbox": [ + 134, + 510, + 481, + 531 + ], + "type": "text", + "content": "16. Goldman, D., Curless, B., Hertzmann, A., Seitz, S.: Shape and spatially-varying brdfs from photometric stereo. In: ICCV (October 2005) 3, 4" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 134, + 533, + 481, + 553 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 134, + 533, + 481, + 553 + ], + "spans": [ + { + "bbox": [ + 134, + 533, + 481, + 553 + ], + "type": "text", + "content": "17. Guo, H., Okura, F., Shi, B.: Multispectral photometric stereo for spatially-varying spectral reflectances. IJCV 130, 2166-2183 (2022) 2, 11, 13" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 134, + 554, + 481, + 586 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 134, + 554, + 481, + 586 + ], + "spans": [ + { + "bbox": [ + 134, + 554, + 481, + 586 + ], + "type": "text", + "content": "18. Guo, H., Okura, F., Shi, B., Funatomi, T., Mukaigawa, Y., Matsushita, Y.: Multispectral photometric stereo for spatially-varying spectral reflectances: A well posed problem? In: CVPR. pp. 963-971 (2021) 2, 3" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 134, + 587, + 481, + 609 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 134, + 587, + 481, + 609 + ], + "spans": [ + { + "bbox": [ + 134, + 587, + 481, + 609 + ], + "type": "text", + "content": "19. Hayakawa, H.: Photometric stereo under a light source with arbitrary motion. JOSA 11(11), 3079-3089 (1994) 4" + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 134, + 609, + 481, + 642 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 134, + 609, + 481, + 642 + ], + "spans": [ + { + "bbox": [ + 134, + 609, + 481, + 642 + ], + "type": "text", + "content": "20. Hernández, C., Vogiatzis, G., Brostow, G., Stenger, B., Cipolla, R.: Non-rigid photometric stereo with colored lights. In: ICCV. pp. 1-8 (2007). https://doi.org/10.1109/ICCV.2007.4408939 2, 4" + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 134, + 643, + 481, + 665 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 134, + 643, + 481, + 665 + ], + "spans": [ + { + "bbox": [ + 134, + 643, + 481, + 665 + ], + "type": "text", + "content": "21. Hertzmann, A., Seitz, S.: Example-based photometric stereo: shape reconstruction with general, varying brdfs. IEEE TPAMI 27(8), 1254-1264 (2005) 4" + } + ] + } + ], + "index": 23 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 391, + 91, + 447, + 102 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 391, + 91, + 447, + 102 + ], + "spans": [ + { + "bbox": [ + 391, + 91, + 447, + 102 + ], + "type": "text", + "content": "SpectraM-PS" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 470, + 91, + 481, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 470, + 91, + 481, + 100 + ], + "spans": [ + { + "bbox": [ + 470, + 91, + 481, + 100 + ], + "type": "text", + "content": "15" + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 14 + }, + { + "para_blocks": [ + { + "bbox": [ + 132, + 117, + 480, + 663 + ], + "type": "list", + "angle": 0, + "index": 23, + "blocks": [ + { + "bbox": [ + 132, + 117, + 480, + 148 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 117, + 480, + 148 + ], + "spans": [ + { + "bbox": [ + 132, + 117, + 480, + 148 + ], + "type": "text", + "content": "22. Hill, B.: Color capture, color management, and the problem of metamerism: does multispectral imaging offer the solution? In: Proc. SPIE. pp. 2-14. SPIE (1999) 2, 8" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 132, + 149, + 480, + 171 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 149, + 480, + 171 + ], + "spans": [ + { + "bbox": [ + 132, + 149, + 480, + 171 + ], + "type": "text", + "content": "23. Ikehata, S.: Cnn-ps: Cnn-based photometric stereo for general non-convex surfaces. In: ECCV (2018) 2, 4, 6" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 132, + 172, + 480, + 192 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 172, + 480, + 192 + ], + "spans": [ + { + "bbox": [ + 132, + 172, + 480, + 192 + ], + "type": "text", + "content": "24. Ikehata, S.: Ps-transformer: Learning sparse photometric stereo network using self-attention mechanism. In: BMVC (2021) 4, 6" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 132, + 193, + 480, + 213 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 193, + 480, + 213 + ], + "spans": [ + { + "bbox": [ + 132, + 193, + 480, + 213 + ], + "type": "text", + "content": "25. Ikehata, S.: Universal photometric stereo network using global lighting contexts. In: CVPR (2022) 2, 3, 4, 6, 7" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 132, + 214, + 480, + 235 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 214, + 480, + 235 + ], + "spans": [ + { + "bbox": [ + 132, + 214, + 480, + 235 + ], + "type": "text", + "content": "26. Ikehata, S.: Scalable, detailed and mask-free universal photometric stereo. In: CVPR (2023) 2, 3, 4, 6, 7, 8, 9, 11, 12" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 132, + 236, + 480, + 256 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 236, + 480, + 256 + ], + "spans": [ + { + "bbox": [ + 132, + 236, + 480, + 256 + ], + "type": "text", + "content": "27. Ikehata, S., Aizawa, K.: Photometric stereo using constrained bivariate regression for general isotropic surfaces. In: CVPR (2014) 3" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 132, + 257, + 480, + 278 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 257, + 480, + 278 + ], + "spans": [ + { + "bbox": [ + 132, + 257, + 480, + 278 + ], + "type": "text", + "content": "28. Ikehata, S., Wipf, D., Matsushita, Y., Aizawa, K.: Robust photometric stereo using sparse regression. In: CVPR (2012) 3, 4" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 132, + 279, + 480, + 310 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 279, + 480, + 310 + ], + "spans": [ + { + "bbox": [ + 132, + 279, + 480, + 310 + ], + "type": "text", + "content": "29. Ikehata, S., Wipf, D., Matsushita, Y., Aizawa, K.: Photometric stereo using sparse bayesian regression for general diffuse surfaces. IEEE TPAMI 36(9), 1816-1831 (2014) 2, 3" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 132, + 311, + 480, + 331 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 311, + 480, + 331 + ], + "spans": [ + { + "bbox": [ + 132, + 311, + 480, + 331 + ], + "type": "text", + "content": "30. Ikehata, S.: Does physical interpretability of observation map improve photometric stereo networks? In: ICIP (2022) 4, 6" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 132, + 332, + 480, + 364 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 332, + 480, + 364 + ], + "spans": [ + { + "bbox": [ + 132, + 332, + 480, + 364 + ], + "type": "text", + "content": "31. Ishio, H., Minowa, J., Nosu, K.: Review and status of wavelength-division-multiplexing technology and its application. Journal of lightwave technology 2(4), 448-463 (1984) 2" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 132, + 365, + 480, + 385 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 365, + 480, + 385 + ], + "spans": [ + { + "bbox": [ + 132, + 365, + 480, + 385 + ], + "type": "text", + "content": "32. Ju, Y., Dong, X., Wang, Y., Qi, L., Dong, J.: A dual-cue network for multispectral photometric stereo. Pattern Recognition 100, 107162 (2020) 2, 4, 6" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 132, + 386, + 480, + 418 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 386, + 480, + 418 + ], + "spans": [ + { + "bbox": [ + 132, + 386, + 480, + 418 + ], + "type": "text", + "content": "33. Ju, Y., Qi, L., Zhou, H., Dong, J., Lu, L.: Demultiplexing colored images for multispectral photometric stereo via deep neural networks. IEEE Access 6, 30804-30818 (2018) 2, 4, 6" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 132, + 418, + 480, + 450 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 418, + 480, + 450 + ], + "spans": [ + { + "bbox": [ + 132, + 418, + 480, + 450 + ], + "type": "text", + "content": "34. Ju, Y., Dong, J., Chen, S.: Recovering surface normal and arbitrary images: A dual regression network for photometric stereo. IEEE Transactions on Image Processing 30, 3676-3690 (2021) 4" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 132, + 450, + 480, + 471 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 450, + 480, + 471 + ], + "spans": [ + { + "bbox": [ + 132, + 450, + 480, + 471 + ], + "type": "text", + "content": "35. Kajiya, J.T.: The rendering equation. SIGGRAPH Comput. Graph. 20(4), 143-150 (1986) 8" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 132, + 472, + 480, + 502 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 472, + 480, + 502 + ], + "spans": [ + { + "bbox": [ + 132, + 472, + 480, + 502 + ], + "type": "text", + "content": "36. Kaya, B., Kumar, S., Oliveira, C., Ferrari, V., Van Gool, L.: Uncalibrated neural inverse rendering for photometric stereo of general surfaces. pp. 3804-3814 (2021) 4" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 132, + 504, + 480, + 535 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 504, + 480, + 535 + ], + "spans": [ + { + "bbox": [ + 132, + 504, + 480, + 535 + ], + "type": "text", + "content": "37. Kontsevich, L.L., Petrov, A., Vergelskaya, I.: Reconstruction of shape from shading in color images. Journal of the Optical Society of America pp. 1047-1052 (1994) 2, 4" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 132, + 536, + 480, + 568 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 536, + 480, + 568 + ], + "spans": [ + { + "bbox": [ + 132, + 536, + 480, + 568 + ], + "type": "text", + "content": "38. Lee, J., Lee, Y., Kim, J., Kosiorek, A., Choi, S., Teh, Y.W.: Set transformer: A framework for attention-based permutation-invariant neural networks. In: ICML. pp. 3744-3753 (2019) 7" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 132, + 569, + 480, + 590 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 569, + 480, + 590 + ], + "spans": [ + { + "bbox": [ + 132, + 569, + 480, + 590 + ], + "type": "text", + "content": "39. Li, J., Robles-Kelly, A., You, S., Matsushita, Y.: Learning to minify photometric stereo. In: CVPR (2019) 4" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 132, + 590, + 480, + 611 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 590, + 480, + 611 + ], + "spans": [ + { + "bbox": [ + 132, + 590, + 480, + 611 + ], + "type": "text", + "content": "40. Li, J., Li, H.: Neural reflectance for shape recovery with shadow handling. In: CVPR (2022) 4" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 132, + 612, + 480, + 632 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 612, + 480, + 632 + ], + "spans": [ + { + "bbox": [ + 132, + 612, + 480, + 632 + ], + "type": "text", + "content": "41. Li, J., Li, H.: Self-calibrating photometric stereo by neural inverse rendering. In: ECCV (2022) 4" + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 132, + 633, + 480, + 663 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 633, + 480, + 663 + ], + "spans": [ + { + "bbox": [ + 132, + 633, + 480, + 663 + ], + "type": "text", + "content": "42. Liu, H., Yan, Y., Song, K., Yu, H.: Sps-net: Self-attention photometric stereo network. IEEE Transactions on Instrumentation and Measurement 70, 1-13 (2021) 4" + } + ] + } + ], + "index": 22 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "spans": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "type": "text", + "content": "16" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 166, + 91, + 267, + 101 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 166, + 91, + 267, + 101 + ], + "spans": [ + { + "bbox": [ + 166, + 91, + 267, + 101 + ], + "type": "text", + "content": "S. Ikehata and Y. Asano" + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 15 + }, + { + "para_blocks": [ + { + "bbox": [ + 132, + 116, + 481, + 665 + ], + "type": "list", + "angle": 0, + "index": 22, + "blocks": [ + { + "bbox": [ + 132, + 116, + 480, + 137 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 116, + 480, + 137 + ], + "spans": [ + { + "bbox": [ + 132, + 116, + 480, + 137 + ], + "type": "text", + "content": "43. Liu, Z., Mao, H., Chao-Yuan Wu, C.F.: A convnet for the 2020s. In: CVPR (2022) 6" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 132, + 139, + 481, + 171 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 139, + 481, + 171 + ], + "spans": [ + { + "bbox": [ + 132, + 139, + 481, + 171 + ], + "type": "text", + "content": "44. Logothetis, F., Budvytis, I., Mecca, R., Cipolla, R.: Px-net: Simple and efficient pixel-wise training of photometric stereo networks. In: CVPR. pp. 12757-12766 (2021) 4" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 133, + 172, + 480, + 194 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 172, + 480, + 194 + ], + "spans": [ + { + "bbox": [ + 133, + 172, + 480, + 194 + ], + "type": "text", + "content": "45. Lu, F., Matsushita, Y., Sato, I., Okabe, T., Sato, Y.: Uncalibrated photometric stereo for unknown isotropic reflectances. In: CVPR. pp. 1490-1497 (2013) 4" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 133, + 194, + 480, + 225 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 194, + 480, + 225 + ], + "spans": [ + { + "bbox": [ + 133, + 194, + 480, + 225 + ], + "type": "text", + "content": "46. Lv, J., Guo, H., Chen, G., Liang, J., Shi, B.: Non-lambertian multispectral photometric stereo via spectral reflectance decomposition. In: IJCAI (2023) 2, 3, 4, 6, 7, 11, 13" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 132, + 227, + 480, + 258 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 227, + 480, + 258 + ], + "spans": [ + { + "bbox": [ + 132, + 227, + 480, + 258 + ], + "type": "text", + "content": "47. Mecca, R., Rosman, G., Kimmel, R., Bruckstein, A.: Perspective photometric stereo with shadows. In: Proc. of 4th International Conference on Scale Space and Variational Methods in Computer Vision (2013) 4" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 132, + 259, + 480, + 281 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 259, + 480, + 281 + ], + "spans": [ + { + "bbox": [ + 132, + 259, + 480, + 281 + ], + "type": "text", + "content": "48. Mitsubishi Electric Research Laboratories (MERL): MERL BRDF Database. http://www.merl.com/brdf/, accessed: 2024-03-07 7" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 132, + 281, + 480, + 312 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 281, + 480, + 312 + ], + "spans": [ + { + "bbox": [ + 132, + 281, + 480, + 312 + ], + "type": "text", + "content": "49. Mo, Z., Shi, B., Lu, F., Yeung, S.K., Matsushita, Y.: Uncalibrated photometric stereo under natural illumination. pp. 2936-2945. IEEE Computer Society (2018) 4" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 132, + 315, + 480, + 346 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 315, + 480, + 346 + ], + "spans": [ + { + "bbox": [ + 132, + 315, + 480, + 346 + ], + "type": "text", + "content": "50. Nayatani, Y., Kurioka, Y., Sobagaki, H.: Study on color rendering and metamerism (part 8). Journal of the Illuminating Engineering Institute of Japan 56(9), 529-536 (1972). https://doi.org/10.2150/jieij1917.56.9_529_2, 8" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 132, + 347, + 480, + 380 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 347, + 480, + 380 + ], + "spans": [ + { + "bbox": [ + 132, + 347, + 480, + 380 + ], + "type": "text", + "content": "51. Ozawa, K., Sato, I., Yamaguchi, M.: Single color image photometric stereo for multi-colored surfaces. Computer Vision and Image Understanding 171, 140-149 (2018) 2" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 132, + 380, + 480, + 413 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 380, + 480, + 413 + ], + "spans": [ + { + "bbox": [ + 132, + 380, + 480, + 413 + ], + "type": "text", + "content": "52. Ren, J., Wang, F., Zhang, J., Zheng, Q., Ren, M., Shi, B.: Diligent10²: A photometric stereo benchmark dataset with controlled shape and material variation. pp. 12581-12590 (June 2022) 3, 9" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 132, + 414, + 480, + 456 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 414, + 480, + 456 + ], + "spans": [ + { + "bbox": [ + 132, + 414, + 480, + 456 + ], + "type": "text", + "content": "53. Santo, H., Samejima, M., Sugano, Y., Shi, B., Matsushita, Y.: Deep photometric stereo network. In: International Workshop on Physics Based Vision meets Deep Learning (PBDL) in Conjunction with IEEE International Conference on Computer Vision (ICCV) (2017) 4" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 132, + 457, + 480, + 478 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 457, + 480, + 478 + ], + "spans": [ + { + "bbox": [ + 132, + 457, + 480, + 478 + ], + "type": "text", + "content": "54. Shi, B., Matsushita, Y., Wei, Y., Xu, C., Tan, P.: Self-calibrating photometric stereo. In: CVPR (2010) 2, 4" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 132, + 479, + 480, + 501 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 479, + 480, + 501 + ], + "spans": [ + { + "bbox": [ + 132, + 479, + 480, + 501 + ], + "type": "text", + "content": "55. Shi, B., Tan, P., Matsushita, Y., Ikeuchi, K.: A biquadratic reflectance model for radiometric image analysis. In: CVPR (2012) 2" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 132, + 502, + 480, + 533 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 502, + 480, + 533 + ], + "spans": [ + { + "bbox": [ + 132, + 502, + 480, + 533 + ], + "type": "text", + "content": "56. Shi, B., Wu, Z., Mo, Z., D.Duan, Yeung, S.K., Tan, P.: A benchmark dataset and evaluation for non-lambertian and uncalibrated photometric stereo. In: CVPR (2016) 3, 9" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 132, + 533, + 480, + 555 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 533, + 480, + 555 + ], + "spans": [ + { + "bbox": [ + 132, + 533, + 480, + 555 + ], + "type": "text", + "content": "57. Silver, W.M.: Determining shape and reflectance using multiple images. Master's thesis, MIT (1980) 1" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 132, + 555, + 480, + 578 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 555, + 480, + 578 + ], + "spans": [ + { + "bbox": [ + 132, + 555, + 480, + 578 + ], + "type": "text", + "content": "58. Taniai, T., Maehara, T.: Neural Inverse Rendering for General Reflectance Photometric Stereo. In: ICML (2018) 4" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 132, + 578, + 480, + 599 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 578, + 480, + 599 + ], + "spans": [ + { + "bbox": [ + 132, + 578, + 480, + 599 + ], + "type": "text", + "content": "59. Tiwari, A., Raman, S.: Deepps2: Revisiting photometric stereo using two differently illuminated images. ECCV (2022) 4" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 132, + 600, + 480, + 621 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 600, + 480, + 621 + ], + "spans": [ + { + "bbox": [ + 132, + 600, + 480, + 621 + ], + "type": "text", + "content": "60. Tzeng, E., Hoffman, J., Saenko, K., Darrell, T.: Adversarial discriminative domain adaptation. In: CVPR (2017) 7" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 132, + 622, + 480, + 643 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 622, + 480, + 643 + ], + "spans": [ + { + "bbox": [ + 132, + 622, + 480, + 643 + ], + "type": "text", + "content": "61. Vaswani, A., Shazeer, N., Parmar, N., Uszkoreit, J., Jones, L., Gomez, A., Kaiser, L., Polosukhin, I.: Attention is all you need. In: NeurIPS (2017) 6" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 132, + 644, + 480, + 665 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 644, + 480, + 665 + ], + "spans": [ + { + "bbox": [ + 132, + 644, + 480, + 665 + ], + "type": "text", + "content": "62. Vogiatzis, G., Hernandez, C.: Self-calibrated, multi-spectral photometric stereo for 3d face capture. IJCV 56(97), 91-103 (2012) 2" + } + ] + } + ], + "index": 21 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 391, + 91, + 447, + 102 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 391, + 91, + 447, + 102 + ], + "spans": [ + { + "bbox": [ + 391, + 91, + 447, + 102 + ], + "type": "text", + "content": "SpectraM-PS" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 470, + 91, + 481, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 470, + 91, + 481, + 100 + ], + "spans": [ + { + "bbox": [ + 470, + 91, + 481, + 100 + ], + "type": "text", + "content": "17" + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 16 + }, + { + "para_blocks": [ + { + "bbox": [ + 132, + 116, + 480, + 304 + ], + "type": "list", + "angle": 0, + "index": 9, + "blocks": [ + { + "bbox": [ + 132, + 116, + 480, + 148 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 116, + 480, + 148 + ], + "spans": [ + { + "bbox": [ + 132, + 116, + 480, + 148 + ], + "type": "text", + "content": "63. Wang, F., Ren, J., Guo, H., Ren, M., Shi, B.: Diligent-pi: A photometric stereo benchmark dataset with controlled shape and material variation (October 2023) 3, 9" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 132, + 149, + 480, + 171 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 149, + 480, + 171 + ], + "spans": [ + { + "bbox": [ + 132, + 149, + 480, + 171 + ], + "type": "text", + "content": "64. Woodham, P.: Photometric method for determining surface orientation from multiple images. Opt. Engg 19(1), 139-144 (1980) 1, 3" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 132, + 172, + 480, + 193 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 172, + 480, + 193 + ], + "spans": [ + { + "bbox": [ + 132, + 172, + 480, + 193 + ], + "type": "text", + "content": "65. Wu, L., Ganesh, A., Shi, B., Matsushita, Y., Wang, Y., Ma, Y.: Robust photometric stereo via low-rank matrix completion and recovery. In: ACCV (2010) 4" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 132, + 194, + 480, + 215 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 194, + 480, + 215 + ], + "spans": [ + { + "bbox": [ + 132, + 194, + 480, + 215 + ], + "type": "text", + "content": "66. Xiao, T., Liu, Y., Zhou, B., Jiang, Y., Sun, J.: Unified perceptual parsing for scene understanding. In: ECCV. pp. 418-434 (2018) 6" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 132, + 216, + 480, + 247 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 216, + 480, + 247 + ], + "spans": [ + { + "bbox": [ + 132, + 216, + 480, + 247 + ], + "type": "text", + "content": "67. Xiong, R., Yang, Y., He, D., Zheng, K., Zheng, S., Xing, C., Zhang, H., Lan, Y., Wang, L., Liu, T.: On layer normalization in the transformer architecture. In: International Conference on Machine Learning. pp. 10524-10533. PMLR (2020) 6" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 132, + 249, + 480, + 270 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 249, + 480, + 270 + ], + "spans": [ + { + "bbox": [ + 132, + 249, + 480, + 270 + ], + "type": "text", + "content": "68. Yao, Z., Li, K., Fu, Y., Hu, H., Shi, B.: Gps-net: Graph-based photometric stereo network. NeurIPS (2020) 4" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 132, + 271, + 480, + 304 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 271, + 480, + 304 + ], + "spans": [ + { + "bbox": [ + 132, + 271, + 480, + 304 + ], + "type": "text", + "content": "69. Zheng, Q., Jia, Y., Shi, B., Jiang, X., Duan, L.Y., Kot, A.: Spline-net: Sparse photometric stereo through lighting interpolation and normal estimation networks. ICCV (2019) 4" + } + ] + } + ], + "index": 8 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "spans": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "type": "text", + "content": "18" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 166, + 91, + 267, + 101 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 166, + 91, + 267, + 101 + ], + "spans": [ + { + "bbox": [ + 166, + 91, + 267, + 101 + ], + "type": "text", + "content": "S. Ikehata and Y. Asano" + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 17 + } + ], + "_backend": "vlm", + "_version_name": "2.6.4" +} \ No newline at end of file diff --git a/2024/Physics-informed Knowledge Transfer for Underwater Monocular Depth Estimation/58639476-9724-4b09-9db4-e436240a5f18_content_list.json b/2024/Physics-informed Knowledge Transfer for Underwater Monocular Depth Estimation/58639476-9724-4b09-9db4-e436240a5f18_content_list.json new file mode 100644 index 0000000000000000000000000000000000000000..9afab1bed4c1ef49dfde461bf64f6e0234972cc0 --- /dev/null +++ b/2024/Physics-informed Knowledge Transfer for Underwater Monocular Depth Estimation/58639476-9724-4b09-9db4-e436240a5f18_content_list.json @@ -0,0 +1,1861 @@ +[ + { + "type": "text", + "text": "Physics-informed Knowledge Transfer for Underwater Monocular Depth Estimation", + "text_level": 1, + "bbox": [ + 263, + 140, + 738, + 185 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Jinghe Yang, Mingming Gong, and Ye Pu", + "bbox": [ + 349, + 212, + 651, + 227 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "The University of Melbourne jinghey@student.unimelb.edu.au, {mingming.gong, ye.pu}@unimelb.edu.au", + "bbox": [ + 235, + 239, + 764, + 268 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Abstract. Compared to the in-air case, underwater depth estimation has its own challenges. For instance, acquiring high-quality training datasets with groundtruth poses difficulties due to sensor limitations in aquatic environments. Additionally, the physics characteristics of underwater imaging diverge significantly from the in-air case, the methods developed for in-air depth estimation underperform when applied underwater, due to the domain gap. To address these challenges, our paper introduces a novel transfer-learning-based method - Physics-informed Underwater Depth Estimation (PUDE). The key idea is to transfer the knowledge of a pre-trained in-air depth estimation model to underwater settings utilizing a small underwater image set without groundtruth measurement, guided by a physical underwater imaging formation model. We propose novel bound losses based on the physical model to rectify the depth estimations to align with actual underwater physical properties. Finally, in the evaluations across multiple datasets, we compare PUDE model with other existing in-air and underwater methods. The results reveal that the PUDE model excels in both quantitative and qualitative comparisons.", + "bbox": [ + 261, + 297, + 754, + 534 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Keywords: Underwater Depth $\\cdot$ Physics-informed $\\cdot$ Knowledge Transfer", + "bbox": [ + 261, + 547, + 740, + 561 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "1 Introduction", + "text_level": 1, + "bbox": [ + 217, + 584, + 375, + 599 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Depth estimation is a pivotal task in computer vision, with substantial implications for robotics applications. Traditional techniques typically employ stereo cameras to estimate depth through disparity calculations. However, these methods are hampered by complex camera system settings and limited by the challenging feature matching in environments with low texture or visibility [23]. In contrast, monocular depth estimation provides an alternative solution, necessitating only a single camera and a less complicated setup. Although monocular depth estimation faces inherent limitations due to the absence of stereoscopic information, many well-performing in-air learning-based depth estimation models have been developed in recent years [5,6,12,21,27-29]. Benefiting from the massive and various training datasets, some of in-air monocular depth estimation models, such as MiDaS [29] and DPT [28], demonstrate powerful generalization performance. However, the pre-trained in-air models are not fully applicable for underwater cases due to the domain gap and the differences in image formation [36]. For example, we applied the in-air trained models, e.g., DPT in [28], to", + "bbox": [ + 212, + 613, + 787, + 843 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "underwater images and observed that DPT has the issues of the overestimation for blurry regions, unclear background region detection and loss of depth details. Additionally, it is challenging to train new models from scratch for underwater settings. Underwater scenarios impose difficulties when capturing a large amount of high-quality images and the groundtruth depth, due to sensor limitations, such as the incapability of LiDAR and the performance drop of stereo matching in water. In contrast, sonar scanning is not affected by water quality and blurriness but suffers from a slow scanning rate and low resolution, lacking detail in scene representation [33]. In response to the distinct challenges of underwater depth estimation, this paper proposes a novel transfer-learning-based approach. Our method transfers knowledge from a pre-trained in-air monocular depth estimation model to underwater settings, guided by the underwater imaging formation model introduced in [9]. Using newly proposed bound losses, which leverage physical information to correct inaccurate depth in distant and background regions, our training method enforces the model to produce estimates aligned with the extracted physical properties. Notably, this adaptation requires minimal training data and no ground truth depth measurements. Our approach benefits from the structural knowledge of the in-air model and integrates underwater physics knowledge, thereby enhancing underwater depth estimation performance. Our contributions are as follows:", + "bbox": [ + 212, + 146, + 787, + 448 + ], + "page_idx": 1 + }, + { + "type": "list", + "sub_type": "text", + "list_items": [ + "- We present experimental observations and limitations when applying an in-air trained model, i.e., DPT in [28], to underwater images. Inspired by these observations, we propose a novel training method to adapt in-air pre-trained depth estimation models for underwater environments. There are two key ingredients in our method: 1) Knowledge transfer using an in-air trained model with good generalisation properties (DPT [28]) to address data scarcity; 2) Integration of the underwater image formation model proposed in [9] to ensure the depth estimates align with underwater physics characteristics.", + "- We propose innovative bound losses to steer the model towards producing depth estimations that adhere to the underwater imaging formation model.", + "- We introduce the Physics-informed Underwater Depth Estimation (PUDE) model, an end-to-end model for underwater monocular depth estimation. In zero-shot evaluations across various datasets, our method demonstrates enhanced performance, both quantitatively and qualitatively, compared to other in-air and underwater methods." + ], + "bbox": [ + 225, + 458, + 787, + 684 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "2 RelatedWork", + "text_level": 1, + "bbox": [ + 215, + 707, + 380, + 723 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "2.1 In-air Monocular Depth Estimation", + "text_level": 1, + "bbox": [ + 215, + 739, + 555, + 753 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Compared to LiDAR, infrared, and stereo vision, monocular depth estimation stands out for its cost-effectiveness and simplicity of configuration, making it appealing for depth perception in autonomous systems. Recent in-air monocular depth estimation methods have exhibited commendable performance, especially supervised methods trained on datasets containing measured groundtruth. They", + "bbox": [ + 212, + 763, + 787, + 840 + ], + "page_idx": 1 + }, + { + "type": "page_number", + "text": "2", + "bbox": [ + 217, + 114, + 228, + 126 + ], + "page_idx": 1 + }, + { + "type": "header", + "text": "J. Yang et al.", + "bbox": [ + 271, + 114, + 362, + 128 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "usually require sensors like LiDAR, RGBD cameras, and stereo matching [11,12, 18,22]. However, these methods are limited by the lack of variety in the datasets, which restricts their generalization across different real-world scenarios.", + "bbox": [ + 212, + 146, + 782, + 191 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "In this context, self-supervised learning using continuous image sequences [14] or calibrated stereo image pairs [13] presents an attractive avenue for exploration but performs less compared to supervised learning [23], and its application to dynamic scenes is challenging. Recognizing this limitation, some studies expand the breadth of training data by harnessing web resources to generate depth groundtruth, thereby enlarging the training dataset and encompassing diverse scenes [8,21,34]. Building further on that, MiDaS [29] extract depth data from 3D movies to enhance dataset diversity and improve generalization. DPT [28] builds upon the MiDaS training dataset, but with supplementary, and implements an encoder-decoder architecture that uses a vision transformer as its backbone, enabling fine-grained and globally coherent estimations and showing strong zero-shot cross-dataset generalization capabilities.", + "bbox": [ + 212, + 191, + 787, + 373 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "2.2 Underwater Monocular Depth Estimation", + "text_level": 1, + "bbox": [ + 214, + 393, + 606, + 407 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Depth measurement in water introduces new challenges due to the unique characteristics of the underwater environment. The limited availability of sensors, such as LiDAR and infrared, makes the direct acquisition of the depth measurement highly challenging. Meanwhile, the image blurriness highly influences the correspondence matching between images [35], thereby the stereo camera can only provide good depth measurement for close regions. The existing underwater depth datasets that rely on stereo matching [1,4], manifest this issue.", + "bbox": [ + 212, + 417, + 782, + 521 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Therefore, monocular depth estimation is a potential solution to tackle these issues. Physics-based methods play a pivotal role in addressing the underwater depth estimation challenge. The underwater image formation model proposed in [9] has been demonstrated to be effective in solving problems in computer vision for underwater scenarios, such as feature extraction [35], neural radiance fields (NeRFs) [19], and depth estimation [10]. Traditional methods utilize the physics-based image formation model to estimate depth directly from underwater images [3,4,7,24-26]. As the foundation of these methods, Dark Channel Prior (DCP) [17], a statistical method initially developed for in-air haze removal, uses pixel intensity to estimate medium transmission. A variant for the underwater environment, UDCP [10], claims that the DCP assumption remains applicable to the green and blue channels of underwater images. However, underwater depth estimation that relies solely on pixel values and the imaging formation model is vulnerable to variable lighting conditions, such as shadows and sunlight flickers.", + "bbox": [ + 212, + 523, + 787, + 734 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Despite the challenges in depth groundtruth acquisition underwater, the learning-based methods show promise. Some works have attempted unsupervised learning, such as UwNET [15] and UW-GAN [16]. They use the GAN method through color restoration to estimate depth. The recent works [2,30] adopt self-supervised learning, utilizing the correspondences between subsequent frames. Notably, they incorporate the physics model into the training of the underwater depth estimation models and have shown to be advantageous.", + "bbox": [ + 212, + 734, + 787, + 839 + ], + "page_idx": 2 + }, + { + "type": "header", + "text": "PUDE", + "bbox": [ + 684, + 113, + 730, + 126 + ], + "page_idx": 2 + }, + { + "type": "page_number", + "text": "3", + "bbox": [ + 774, + 114, + 785, + 126 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "2.3 Underwater Imaging Formation Model", + "text_level": 1, + "bbox": [ + 215, + 146, + 581, + 162 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Light transmission in water is subject to absorption, and minuscule particles cause the backscattering, leading to blur within underwater imagery. A widely used model of underwater image formation is proposed in [9] to describe this physical phenomenon. Consider an underwater image represented by $I_{c} \\in \\mathbb{R}^{HW}$ , where $c$ denotes the color channels $\\{\\mathrm{R}, \\mathrm{G}, \\mathrm{B}\\}$ , and $H$ and $W$ are the image's height and width, respectively. Given an image $I_{c}$ , for each pixel $i$ , $i = 1 \\cdots HW$ , it holds that $I_{c,i} \\in [0,1]$ . The underwater imaging formation model is:", + "bbox": [ + 212, + 171, + 787, + 279 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\nI _ {c, i} = J _ {c, i} \\cdot t _ {c, i} + B _ {c} ^ {\\infty} (1 - t _ {c, i}), \\tag {1}\n$$\n", + "text_format": "latex", + "bbox": [ + 393, + 290, + 785, + 306 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\nt _ {c, i} = e ^ {- \\beta_ {c} \\mathbf {z} _ {i}}. (2)\n$$\n", + "text_format": "latex", + "bbox": [ + 454, + 330, + 785, + 349 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "In this model, $J_{c,i}$ is the original in-air color value at pixel $i$ . The original in-air image is denoted by $J_{c} = [J_{c,1},\\dots,J_{c,HW}] \\in \\mathbb{R}^{HW}$ . The medium transmission rate denoted by $t_{c,i}$ is associated with the beam attenuation coefficient $\\beta_{c}$ and the depth $\\mathbf{z}_{i}$ . Meanwhile, $B_{c}^{\\infty}$ is the coefficient of ambient light causing the backscattering from particles. The captured underwater image $I_{c,i}$ is the aggregate of the attenuation of the original image and the additional backscattering from the water environment. The values of $\\beta_{c}$ and $B_{c}^{\\infty}$ vary depending on different water environments and can be estimated using the depth measurements [1].", + "bbox": [ + 212, + 354, + 787, + 477 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "3 DPT on Underwater Images: Observations", + "text_level": 1, + "bbox": [ + 215, + 500, + 665, + 517 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Existing learning-based monocular depth estimation methods, such as DPT [28] and MiDaS [29], generate the inverse relative depth estimation from a single image. We denote the inverse depth estimation as $\\mathbf{d} \\in \\mathbb{R}^{HW}$ , with $\\mathbf{d}_i$ to be the estimate for pixel $i$ , $i = 1, \\dots, HW$ . The metric depth of the input image is denoted by $\\mathbf{z} \\in \\mathbb{R}^{HW}$ , whose reciprocal ideally has a linear relationship with the inverse relative depth. The relationship between inverse relative depth $\\mathbf{d}_i$ and the metric depth $\\mathbf{z}_i$ at pixel $i$ is:", + "bbox": [ + 212, + 531, + 787, + 638 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\n\\mathbf {z} _ {i} = \\frac {1}{s \\mathbf {d} _ {i} + h}, \\quad \\forall i \\in \\{1, \\dots , H W \\}. \\tag {3}\n$$\n", + "text_format": "latex", + "bbox": [ + 375, + 648, + 785, + 680 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Here, $s$ and $h$ are two unknown scale and shift coefficients. In practice, they can be off-line estimated for specific scenes using sequence images. The overall mapping relationship between the inverse relative depth $\\mathbf{d}_i$ and the inverse of metric depth $\\mathbf{z}_i$ is linear on the in-air images [32]. An example can be seen in the first row of Fig. 1, which is the depth estimation of DPT on an outdoor in-air image from DIODE [31], we can observe that DPT estimation on the in-air image maintains a good linear relationship (the correlation coefficient $r = 0.992$ ). However, when we apply DPT directly to underwater images from the dataset [1,4], in which the depth groundtruth is obtained through the Structure-From-Motion and stereo matching, it is observed that in underwater imagery, the estimated", + "bbox": [ + 212, + 688, + 787, + 840 + ], + "page_idx": 3 + }, + { + "type": "page_number", + "text": "4", + "bbox": [ + 217, + 114, + 228, + 126 + ], + "page_idx": 3 + }, + { + "type": "header", + "text": "J. Yang et al.", + "bbox": [ + 271, + 114, + 362, + 128 + ], + "page_idx": 3 + }, + { + "type": "image", + "img_path": "images/8824e2ba1a6351826f34979e8a17fac4e2e9017600967b24d71b5ee8309f5e69.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 243, + 143, + 415, + 232 + ], + "page_idx": 4 + }, + { + "type": "image", + "img_path": "images/bfbd28f515df07458287e11034a532f3cf58557a3da4626d1d260b559bdd9ada.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 416, + 143, + 586, + 232 + ], + "page_idx": 4 + }, + { + "type": "image", + "img_path": "images/f2910ee34d2f4aebf56aaede5910cfa5207fa1a7b2d273588bb9b6b5d425e791.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 586, + 148, + 746, + 232 + ], + "page_idx": 4 + }, + { + "type": "image", + "img_path": "images/95652f5ed62bf7ea885a6a09c5f4dc3966cf5b6d9b54df8ad997435039bd22e8.jpg", + "image_caption": [ + "Fig. 1: DPT estimation of underwater images shows a curved tendency, with overestimation in distant regions. Additionally, the estimated depth struggles with unclear background detection (blue arrows)." + ], + "image_footnote": [], + "bbox": [ + 245, + 232, + 415, + 321 + ], + "page_idx": 4 + }, + { + "type": "image", + "img_path": "images/91ffd5b80d32babd4b2dffd30e7ac4b2d4dd74ee97c866842af8943309f1d6c8.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 416, + 232, + 586, + 321 + ], + "page_idx": 4 + }, + { + "type": "image", + "img_path": "images/e4f3271b4175f228a5ca1c95387f5af299ca6a8c15de564c09aad55620ece144.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 586, + 237, + 741, + 321 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "inverse depth, denoted by $\\mathbf{d}_i^D$ , and $\\frac{1}{z_i}$ deviates from a linear relationship and present a curve tendency with a weaker correlation, especially when $\\mathbf{d}_i^D$ is small. An example is shown in the second row of Fig. 1.", + "bbox": [ + 212, + 409, + 784, + 458 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "To observe the nonlinear trend caused by image blurriness, we applied the Sea-thru method [1] to calculate the medium transmission rate $t_i$ in Eq. (2) on the grayscale images. We then compared these rates with those obtained from the DPT depth (aligned with groundtruth depth). An example of this procedure on the Sea-thru D3 dataset is shown in Fig. 2a, where we randomly sample 100 points from each image. We observe that the estimated transmission rate from DPT is mostly lower than that calculated by groundtruth depth when the transmission rate is small ( $t_i < 0.2$ ). The results reveal that DPT underestimates the transmission rate in regions with low transmission rates, leading to an overestimation of depth in distant regions. Furthermore, the background edges on the underwater depth image of DPT, such as the blue arrow pointed regions around corals, are less clear, exhibiting fogging compared with the estimation on the in-air image. Based on these experimental observations, we summarise the primary issues of the depth estimation from DPT in water:", + "bbox": [ + 212, + 460, + 787, + 674 + ], + "page_idx": 4 + }, + { + "type": "list", + "sub_type": "text", + "list_items": [ + "O1 : DPT exhibits a propensity for overestimating depth in regions that are relatively far away from the camera. This overestimation is particularly evident in regions with diminished visual clarity and increased blurriness.", + "O2 : Secondly, DPT encounters challenges in accurately discerning background regions in underwater images. This issue is especially pronounced for the region far from the camera. Due to the attenuation of the scenes, it is challenging for DPT to define the background area clearly." + ], + "bbox": [ + 212, + 695, + 784, + 805 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "We propose two bound losses (Sec. 4.3) in our training to address these issues.", + "bbox": [ + 214, + 824, + 772, + 839 + ], + "page_idx": 4 + }, + { + "type": "header", + "text": "PUDE", + "bbox": [ + 684, + 114, + 730, + 126 + ], + "page_idx": 4 + }, + { + "type": "page_number", + "text": "5", + "bbox": [ + 774, + 116, + 784, + 126 + ], + "page_idx": 4 + }, + { + "type": "image", + "img_path": "images/b2317f68ed0d21e33d87d820aff680737832f55f90f08e3ad833b9797c4c8d4f.jpg", + "image_caption": [ + "(a) $t_i$ in Eq. (2) (D3 dataset)" + ], + "image_footnote": [], + "bbox": [ + 233, + 148, + 362, + 244 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/535ee7551ae3b70ee6299384f427519641887e50abc27238662b9da9c87baf56.jpg", + "image_caption": [ + "(b) Point selection and model fitting in Eq. (1), Eq. (3)", + "Fig.2: (a) The comparison between the calculated transmission rate $t_i$ from groundtruth and the estimated $t_i$ from DPT depth, presents an underestimation of the $t_i$ for small values of DPT. (b) The left image shows the points selected in the refined estimation. The right shows the fitting of the estimation." + ], + "image_footnote": [], + "bbox": [ + 424, + 150, + 612, + 247 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/281a5e27ef6c7031f29527ea0313d3ce9de8ecf271bb56aaa1fcbf974be952bf.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 612, + 151, + 754, + 246 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "4 Method", + "text_level": 1, + "bbox": [ + 215, + 354, + 330, + 372 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "4.1 Framework Overview", + "text_level": 1, + "bbox": [ + 215, + 388, + 437, + 402 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Inspired by the observations in Sec. 3, we propose a physics-informed knowledge transfer method for underwater monocular depth estimation, named the Physics-informed Underwater Depth Estimation (PUDE). The main idea of this physics-informed method is to leverage the physical information in underwater imagery to guide knowledge transfer within the neural network, thereby enhancing its adaptation from in-air to underwater environments. Specifically, we adopt the pre-trained DPT-Hybrid model [28], a state-of-the-art (SOTA) model known for its good generalization, as the transferred model. Benefiting from the massive in-air training dataset, the DPT model has powerful spatial structural knowledge. Despite previously noted issues when applied to underwater images, it can still provide a relatively reasonable depth estimation, particularly for regions close to the camera. This estimation serves as our prior depth knowledge for extracting the physical information from the underwater images.", + "bbox": [ + 212, + 416, + 787, + 612 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Most daily images are non-linear images, which are automatically post-processed by nonlinear algorithms integrated into cameras, for example, gamma correction, to approach the non-linear perception of human eyes. By contrast, linear images are images in which pixel values are directly captured by the camera without nonlinear post-processing or only with ratio processing. To ensure the PUDE model's generalizability in reality and keep consistency in training and testing, we utilize post-processed non-linear underwater images as inputs. In contrast, for precise physical information extraction, we utilize the linear underwater images in the estimation of the physics parameters in Eq. (1) and Eq. (3).", + "bbox": [ + 212, + 613, + 787, + 750 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Our training framework consists of two branches shown in Fig. 3. The first branch is the direct knowledge transfer from DPT to PUDE, and we employ a similarity loss to guarantee the similarity between DPT and PUDE in the transfer (in Sec. 4.2). The second branch is the physics-informed knowledge transfer. Inspired by the parameter estimation method in Sea-thru [1], we estimate the parameters in Eq. (1) and Eq. (3) for each input underwater image, utilizing", + "bbox": [ + 212, + 750, + 787, + 840 + ], + "page_idx": 5 + }, + { + "type": "page_number", + "text": "6", + "bbox": [ + 217, + 114, + 228, + 126 + ], + "page_idx": 5 + }, + { + "type": "header", + "text": "J. Yang et al.", + "bbox": [ + 271, + 114, + 362, + 128 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/91205bbd54b8d2398f71974089efaf96b0e84d1b00c1a1d0c8f78f9fc6ea4de1.jpg", + "image_caption": [ + "Fig. 3: Overall training framework for PUDE. The overall training framework comprises two components: the direct knowledge transfer from the DPT model and the physi-informed knowledge transfer based on the DPT estimation. Physical information is extracted from DPT's output to establish the bound losses." + ], + "image_footnote": [], + "bbox": [ + 222, + 145, + 637, + 308 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/f3287d058a9034e99873dc4a7b102847501a7b47962655d1620a4bcf38b317ed.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 694, + 146, + 779, + 297 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "DPT depth estimation. We restrict our usage to the green and blue channels, in line with the conclusions from UDCP [10]. Based on that, we propose two novel bound losses, utilizing the physics information contained in the underwater image formation model in Eq. (1) to enforce the model to produce plausible depth estimations that conform to the physics characteristics of underwater imaging. In Sec. 4.3, we will introduce the second branch and the designed bound losses.", + "bbox": [ + 212, + 402, + 782, + 494 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "4.2 Direct Knowledge Transfer", + "text_level": 1, + "bbox": [ + 215, + 512, + 483, + 527 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "The direct knowledge transfer loop aims to guarantee the similarity between the original DPT and PUDE during the transition to ensure the structural knowledge can be kept in PUDE. We use a similarity loss in the training process to achieve this goal. In the transfer, PUDE is initialized with DPT. We trim the pixels with the largest differences to set the tolerance for the difference between PUDE and the original model DPT during the training. We denote the pixel set after trimming as $M_{trim} \\subseteq \\{1, \\dots, HW\\}$ , and $T$ representing the cardinality of $M_{trim}$ . For each pixel $i \\in M_{trim}$ , we denote the DPT and PUDE estimation as $\\mathbf{d}_i^D$ and $\\mathbf{d}_i^P$ , respectively. The similarity loss $L_s$ is defined as:", + "bbox": [ + 212, + 534, + 787, + 672 + ], + "page_idx": 6 + }, + { + "type": "equation", + "text": "\n$$\nL _ {s} = \\frac {1}{T} \\sum_ {i \\in M _ {\\mathrm {t r i m}}} \\left| \\frac {\\mathbf {d} _ {i} ^ {P} - \\mathbf {d} _ {i} ^ {D}}{\\operatorname {m e d i a n} (\\mathbf {d} ^ {D})} \\right|. \\tag {4}\n$$\n", + "text_format": "latex", + "bbox": [ + 388, + 680, + 785, + 720 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "The trimming sets the tolerance for the changing between PUDE and the original model DPT during the training. In our implementation, we set the trimming rate to be $0.3HW$ .", + "bbox": [ + 212, + 723, + 787, + 768 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "4.3 Physics-informed Knowledge Transfer", + "text_level": 1, + "bbox": [ + 215, + 787, + 571, + 803 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Parameters Estimation: In the physics-informed knowledge transfer loop, we first extract the parameters in Eq. (1) and Eq. (3) using the preliminary depth", + "bbox": [ + 212, + 809, + 785, + 840 + ], + "page_idx": 6 + }, + { + "type": "header", + "text": "PUDE", + "bbox": [ + 684, + 114, + 730, + 126 + ], + "page_idx": 6 + }, + { + "type": "page_number", + "text": "7", + "bbox": [ + 774, + 116, + 785, + 126 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "estimations from DPT. As shown in Fig. 1, DPT's depth estimations on the underwater images tend to overestimate the depth in distant and blurry regions but offer more trustable estimation for the clearer regions, showing a good linear relationship described in Eq. (3). Therefore, we extract the physics information harnessed from DPT's depth estimations for clear regions.", + "bbox": [ + 212, + 146, + 782, + 222 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "The estimation process utilises the DPT depth estimation (on nonlinear post-processed images) and the linear images, consisting of two steps, the rough and refined estimation. The first step is the rough estimation using the whole image to obtain a robust estimation of the background light $\\hat{B}_c^\\infty$ . The rationale for extracting the darkest points in the Sea-thru method is analogous to identifying points where $J_{c,i}$ is assumed to be close to zero and where the color appearance on these pixels is solely due to the backscattering. Therefore, we evenly select the $N_1$ darkest points, on which $J_{c,i} \\approx 0$ , from the whole image and employ these points to estimate the water parameters by solving the least-square problem in Eq. (5). We denote the pixel value on the pixel location $i$ of the color channel $c$ on the input underwater image as $I_{c,i}$ , $i = 1 \\dots HW$ . According to the suggestions in UDCP [10], we only use blue and green channels in this process. We denote the selected point set of each color channel as $M_c$ , $c \\in \\{G, B\\}$ . Substituting (3) into (1), we define the error $\\epsilon_c$ for the color channels as:", + "bbox": [ + 212, + 222, + 787, + 434 + ], + "page_idx": 7 + }, + { + "type": "equation", + "text": "\n$$\n\\epsilon_ {c} := \\sum_ {i \\in M _ {c}} | | I _ {c, i} - B _ {c} ^ {\\infty} (1 - e ^ {- \\frac {\\beta_ {c} / s}{\\mathbf {d} _ {i} ^ {D + h / s}}}) | | _ {2} ^ {2}. \\tag {5}\n$$\n", + "text_format": "latex", + "bbox": [ + 359, + 441, + 785, + 479 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "To simplify the least-square problem, we set two temporary parameters $\\nu_{c} = \\beta_{c} / s$ and $\\mu = h / s$ . Then, we solve the following least-square problem to obtain the estimate of the parameters:", + "bbox": [ + 214, + 488, + 787, + 532 + ], + "page_idx": 7 + }, + { + "type": "equation", + "text": "\n$$\n\\left(\\hat {\\nu} _ {G}, \\hat {\\nu} _ {B}, \\hat {\\mu}, \\hat {B} _ {c} ^ {\\infty}\\right) = \\operatorname * {a r g m i n} _ {\\nu_ {G}, \\nu_ {B}, \\mu , B _ {c} ^ {\\infty}} \\sum_ {c \\in \\{G, B \\}} \\epsilon_ {c}, \\tag {6}\n$$\n", + "text_format": "latex", + "bbox": [ + 359, + 541, + 785, + 575 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Utilizing the preliminary estimates $\\hat{\\nu}_G$ , $\\hat{\\nu}_B$ , and $\\hat{\\mu}$ , we calculate the medium transmission rate $t_{c,i}$ in Eq. (2) for each pixel. Observations from Fig. 2a suggest that depth overestimation of DPT primarily occurs in regions with low transmission rates $t_i$ . Therefore, in our implementation, in the refined estimation step, we exclude the regions where DPT's estimation is inaccurate while maximizing the utilization of the largest point sampling regions, in which the transmission rate is smaller than 0.2. Similar to rough estimation, in refined estimation, we select $N_2$ darkest points from the selected regions to construct the set $M_{c2}$ and address a least-square problem. However, we streamline the process by utilizing the estimated $\\hat{B}_c^\\infty$ in the first step. The error $\\epsilon_{c2}$ in this is:", + "bbox": [ + 214, + 583, + 787, + 737 + ], + "page_idx": 7 + }, + { + "type": "equation", + "text": "\n$$\n\\epsilon_ {c 2} := \\sum_ {i \\in M _ {c 2}} | | I _ {c, i} - \\hat {B} _ {c} ^ {\\infty} (1 - e ^ {- \\frac {\\nu_ {c}}{\\mathbf {d} _ {i} ^ {D} + \\mu}}) | | _ {2} ^ {2}, \\tag {7}\n$$\n", + "text_format": "latex", + "bbox": [ + 359, + 742, + 785, + 779 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "and the least-square problem is:", + "bbox": [ + 215, + 786, + 450, + 801 + ], + "page_idx": 7 + }, + { + "type": "equation", + "text": "\n$$\n\\left(\\hat {\\nu} _ {G}, \\hat {\\nu} _ {B}, \\hat {\\mu}\\right) = \\operatorname * {a r g m i n} _ {\\nu_ {G}, \\nu_ {B}, \\mu} \\sum_ {c \\in \\{G, B \\}} \\epsilon_ {c 2}, \\tag {8}\n$$\n", + "text_format": "latex", + "bbox": [ + 387, + 809, + 785, + 843 + ], + "page_idx": 7 + }, + { + "type": "page_number", + "text": "8", + "bbox": [ + 217, + 114, + 228, + 126 + ], + "page_idx": 7 + }, + { + "type": "header", + "text": "J. Yang et al.", + "bbox": [ + 271, + 114, + 362, + 128 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/ccc1a4dce5fdff632fe807cc9289c1ca417ab43da08a1397123731bf460308d7.jpg", + "image_caption": [ + "(a) RGB" + ], + "image_footnote": [], + "bbox": [ + 218, + 143, + 331, + 204 + ], + "page_idx": 8 + }, + { + "type": "image", + "img_path": "images/a40e9c9e94a086eca498a5972348c4096cd5cfb4a788dfeaaa12c7af29242741.jpg", + "image_caption": [ + "(b) DPT depth", + "Fig. 4: Loss visualization: Depth estimation from DPT loses details due to blurriness. However, our designed bound losses have differences across pixels according to the pixel intensity thereby enhancing the depth estimation details in training." + ], + "image_footnote": [], + "bbox": [ + 334, + 143, + 446, + 203 + ], + "page_idx": 8 + }, + { + "type": "image", + "img_path": "images/a6f11881cbb6b6c58935904100e80e8fe25e1cad02ffb487dfa204a5bf6bbdd2.jpg", + "image_caption": [ + "(c) $\\mathcal{L}_{bu}$" + ], + "image_footnote": [], + "bbox": [ + 450, + 143, + 669, + 203 + ], + "page_idx": 8 + }, + { + "type": "image", + "img_path": "images/81888be583b5065b0ca2874160e14a97bdb7f8193d74836d9d9564dc07707b95.jpg", + "image_caption": [ + "(e) local patches" + ], + "image_footnote": [], + "bbox": [ + 669, + 143, + 782, + 203 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "In our implementation, $N_{1}$ and $N_{2}$ is set to be 500 and 200 respectively. An implementation example of the refined estimation is shown in Fig. 2b.", + "bbox": [ + 214, + 297, + 782, + 328 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "Bound Losses on Medium Transmission Rate: After estimating the parameters of the underwater imaging formation model in Eq. (1), we are ready to present two bound losses to constrain the PUDE model from predicting depths that contravene the underwater characteristics.", + "bbox": [ + 214, + 345, + 782, + 405 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "Lower bound: The first bound is a lower bound on the estimated medium transmission rate $t_{c,i}^{P}$ , $c \\in \\{G, B\\}$ . This bound aids in refining depth estimates in distant regions, preventing from overestimating the depth for the regions that are far from the camera, as described in O1 in Sec. 3. By substituting the obtained parameter $\\hat{\\nu}_{G}, \\hat{\\nu}_{B}$ and $\\hat{\\mu}$ in the parameters estimation step, and using the PUDE depth estimation, we get the estimated $\\hat{t}_{c,i}^{P}$ :", + "bbox": [ + 214, + 407, + 784, + 500 + ], + "page_idx": 8 + }, + { + "type": "equation", + "text": "\n$$\n\\hat {t} _ {c, i} ^ {P} = e ^ {- \\frac {\\hat {\\nu} _ {c}}{\\mathbf {d} _ {i} ^ {P} + \\hat {\\mu}}}. \\tag {9}\n$$\n", + "text_format": "latex", + "bbox": [ + 446, + 507, + 784, + 532 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "As the original color value of each point $J_{c,i}$ is equal to or greater than zero. Based on Eq. (1), we have:", + "bbox": [ + 214, + 537, + 782, + 568 + ], + "page_idx": 8 + }, + { + "type": "equation", + "text": "\n$$\nI _ {c, i} \\geq \\left(1 - t _ {c, i}\\right) B _ {c} ^ {\\infty}, \\tag {10}\n$$\n", + "text_format": "latex", + "bbox": [ + 429, + 575, + 784, + 592 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "substituting the estimated $\\hat{B}_c^\\infty$ , the first lower bound is:", + "bbox": [ + 214, + 601, + 619, + 616 + ], + "page_idx": 8 + }, + { + "type": "equation", + "text": "\n$$\n\\hat {t} _ {c, i} ^ {P} \\geq 1 - \\frac {I _ {c , i}}{\\hat {B} _ {c} ^ {\\infty}}. \\tag {11}\n$$\n", + "text_format": "latex", + "bbox": [ + 444, + 623, + 784, + 656 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "This lower bound on $\\hat{t}_{c,i}^{P}$ is to pull back the depth estimations for the distant regions. When the depth of pixels in these regions is overestimated, it results in an underestimation of the medium transmission rate in Eq. (9). Consequently, the captured pixel value $I_{c,i}$ should not be lower than the estimated backscattering, as shown in Eq. (10).", + "bbox": [ + 214, + 664, + 782, + 739 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "Upper bound: The second bound is an upper bound on the estimated medium transmission rate $t_{c,i}^{P}, c \\in \\{G, B\\}$ to address the challenge of detecting background edges in blurry underwater images, described in O2.", + "bbox": [ + 214, + 741, + 782, + 785 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "Considering that the maximum pixel intensity should not surpass that of pure white, namely $J_{c,i} \\leq 1$ , and referring to Eq. (1), we obtain:", + "bbox": [ + 214, + 786, + 782, + 816 + ], + "page_idx": 8 + }, + { + "type": "equation", + "text": "\n$$\nI _ {c, i} \\leq t _ {c, i} + \\left(1 - t _ {c, i}\\right) B _ {c} ^ {\\infty}. \\tag {12}\n$$\n", + "text_format": "latex", + "bbox": [ + 408, + 824, + 784, + 840 + ], + "page_idx": 8 + }, + { + "type": "header", + "text": "PUDE", + "bbox": [ + 684, + 114, + 730, + 126 + ], + "page_idx": 8 + }, + { + "type": "page_number", + "text": "9", + "bbox": [ + 774, + 114, + 785, + 126 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "In regions corresponding to the pure background area, the distance can be considered infinite. The depth estimation by DPT for these regions is always finite. Therefore, the estimated medium transmission for these background regions should be greater than the true value. We denote the estimated medium transmission rate at each pixel location by DPT as $\\hat{t}_{c,i}^{D}$ , which is defined as:", + "bbox": [ + 212, + 146, + 787, + 223 + ], + "page_idx": 9 + }, + { + "type": "equation", + "text": "\n$$\n\\hat {t} _ {c, i} ^ {D} = e ^ {- \\frac {\\bar {\\nu} _ {c}}{\\mathbf {d} _ {i} ^ {D} + \\bar {\\mu}}}. \\tag {13}\n$$\n", + "text_format": "latex", + "bbox": [ + 450, + 238, + 785, + 262 + ], + "page_idx": 9 + }, + { + "type": "text", + "text": "In distant regions, where the original pixel values $J_{c,i}$ are attenuated, the captured pixel values $I_{c,i}$ are predominantly dictated by backscattering. To select pixel locations likely to be in the background regions, we define the set $M_b$ :", + "bbox": [ + 214, + 272, + 787, + 319 + ], + "page_idx": 9 + }, + { + "type": "equation", + "text": "\n$$\nM _ {b} = \\{i \\mid \\forall c \\in \\{G, B \\}, \\frac {\\hat {B} _ {c} ^ {\\infty} \\left(1 - \\hat {t} _ {c , i} ^ {D}\\right)}{I _ {c , i}} \\geq \\gamma \\}. \\tag {14}\n$$\n", + "text_format": "latex", + "bbox": [ + 352, + 330, + 785, + 367 + ], + "page_idx": 9 + }, + { + "type": "text", + "text": "The parameter $\\gamma$ is utilized to identify regions where pixel values are predominantly determined by backscattering, which are likely to be in the background area. In our experiments, $\\gamma$ is set to be 0.6. As mentioned, for the background regions, $\\hat{t}_{c,i}^{D}$ is greater than the true value. Then based on the Eq. (12), we have,", + "bbox": [ + 214, + 377, + 787, + 439 + ], + "page_idx": 9 + }, + { + "type": "equation", + "text": "\n$$\nI _ {c, i} < \\hat {t} _ {c, i} ^ {D} + \\left(1 - t _ {c, i}\\right) B _ {c} ^ {\\infty}, \\quad i \\in M _ {b}, \\tag {15}\n$$\n", + "text_format": "latex", + "bbox": [ + 377, + 452, + 785, + 470 + ], + "page_idx": 9 + }, + { + "type": "text", + "text": "substituting the estimated parameters, we have the bound:", + "bbox": [ + 214, + 476, + 643, + 491 + ], + "page_idx": 9 + }, + { + "type": "equation", + "text": "\n$$\n\\hat {t} _ {c, i} ^ {P} < \\frac {- I _ {c , i} + \\hat {t} _ {c , i} ^ {D} + \\hat {B} _ {c} ^ {\\infty}}{\\hat {B} _ {c} ^ {\\infty}}, \\quad i \\in M _ {b}. \\tag {16}\n$$\n", + "text_format": "latex", + "bbox": [ + 377, + 503, + 785, + 542 + ], + "page_idx": 9 + }, + { + "type": "text", + "text": "This upper bound on $t_{c,i}^{P}$ in Eq. (16) is to adjust the depth estimation on the background regions. For the pixels on the background regions, the original DPT produce lower depth estimations. Even if the original pixel $J_{c,i}$ is equal to 1 (perfectly white), the captured pixel value $I_{c,i}$ from the background regions should not be greater than the sum of the estimated medium transmission rate from DPT and the backscattering term, as shown in Eq. (15).", + "bbox": [ + 212, + 553, + 787, + 643 + ], + "page_idx": 9 + }, + { + "type": "text", + "text": "Bound losses: We set the bound losses to enforce the estimated transmission rate $\\hat{t}_{c,i}^{P}$ to satisfy the listed two bounds and optimize the estimated depth of PUDE, $d_{i}^{P}$ , in training. To simplify the mathematical expressions in the subsequent losses, we denote the ReLU function as $f:\\mathbb{R}\\to \\mathbb{R}^+$ , where $f(x) = \\max (0,x)$ . The number of elements in $M_b$ denoted as $N_{b}$ . According to the bounds in Eq. (11), Eq. (16), the bound losses are:", + "bbox": [ + 212, + 645, + 787, + 737 + ], + "page_idx": 9 + }, + { + "type": "equation", + "text": "\n$$\n\\mathcal {L} _ {b l} = \\sum_ {c \\in \\{G, B \\}} \\sum_ {i = 1} ^ {H W} f \\left(- I _ {c, i} + \\left(1 - \\hat {t} _ {c, i} ^ {P}\\right) \\hat {B} _ {c} ^ {\\infty}\\right) / H W, \\tag {17}\n$$\n", + "text_format": "latex", + "bbox": [ + 326, + 750, + 785, + 792 + ], + "page_idx": 9 + }, + { + "type": "equation", + "text": "\n$$\n\\mathcal {L} _ {b u} = \\sum_ {c \\in \\{G, B \\}} \\sum_ {i \\in M _ {b}} f \\left(- \\left(1 - \\hat {t} _ {c, i} ^ {P}\\right) \\hat {B} _ {c} ^ {\\infty} - \\hat {t} _ {c, i} ^ {D} + I _ {c, i}\\right) / N _ {b}. \\tag {18}\n$$\n", + "text_format": "latex", + "bbox": [ + 307, + 808, + 785, + 842 + ], + "page_idx": 9 + }, + { + "type": "page_number", + "text": "10", + "bbox": [ + 217, + 114, + 235, + 126 + ], + "page_idx": 9 + }, + { + "type": "header", + "text": "J. Yang et al.", + "bbox": [ + 271, + 114, + 362, + 128 + ], + "page_idx": 9 + }, + { + "type": "table", + "img_path": "images/9fc071227242792028dd303153c50814a6d5dce38c54eef575a38fe24c73f8b0.jpg", + "table_caption": [ + "Table 1: Quantitative results: The test datasets, Sea-thru D3, D5, and SQUID, comprise 43, 48, and 57 samples, respectively, collected from open ocean environment." + ], + "table_footnote": [], + "table_body": "
DatasetsMethodsδ<1.05↑δ<1.052↑δ<1.053↑AbsRel ↓Sq. Rel↓RMSE ↓RMSE log↓
D3MegaDepth [21]0.2750.5000.6590.1710.1121.0190.2250.054
MiDaS [29]0.2280.4290.5700.4080.9531.6350.4250.230
DPT [28]0.4210.6730.7900.1230.0850.8950.1870.038
UWNET [15]0.0970.1920.2900.7742.8292.0270.6500.451
UDepth [37]0.3150.5540.7070.1580.0941.0010.2130.051
Ours (PUDE)0.4430.7090.8260.1010.0590.8630.1610.027
D5MegaDepth [21]0.3520.5800.7220.1300.0431.4670.2020.043
MiDaS [29]0.3400.5730.7140.1310.0431.3050.1970.044
DPT [28]0.3740.6380.7740.1240.0511.4790.2050.045
UWNET [15]0.0750.1580.2400.3950.3342.5760.4090.184
UDepth [37]0.3760.6280.7760.1170.0381.3430.1920.039
Ours (PUDE)0.4350.6780.8030.1050.0341.3430.1860.037
SQUIDMegaDepth [21]0.1850.3520.4890.2530.1742.4190.2860.096
MiDaS [29]0.1960.3790.5070.2960.3472.4440.3090.139
DPT [28]0.3450.5990.7430.1360.0681.6570.1820.040
UWNET [15]0.0820.1670.2550.5100.7413.3630.4780.278
UDepth [37]0.2740.4640.6060.2130.1631.8830.2360.083
Ours (PUDE)0.3870.6320.7720.1190.0431.5240.1630.032
", + "bbox": [ + 217, + 185, + 785, + 402 + ], + "page_idx": 10 + }, + { + "type": "text", + "text": "These two bound losses are related to the pixel-wise intensity on color channel. Especially for the dark and distant regions, the difference in the pixel intensity can reflect the depth differences across pixels, the pixel intensity is mostly decided by the depth-related backscattering. The differences in bound losses caused by pixel intensity can encourage the model to learn the depth difference across pixels in small local blurry regions. In other words, the two bound losses guide the model to learn the nonlinear mechanism of backscattering in water, namely \"how blurriness affects the pixel intensity\". Fig. 4 shows an example. Therefore, the overall performance of the model in terms of detail detection gets improved.", + "bbox": [ + 212, + 433, + 782, + 568 + ], + "page_idx": 10 + }, + { + "type": "text", + "text": "Final loss: We use the parameters $\\beta_0, \\beta_1$ and $\\beta_2$ to adjust the weights between the similarity loss and bound losses, which are set to 1, 10 and 10, respectively. The weights are decided by the metric and visual results on the validation set during training. The total loss $\\mathcal{L}_{total}$ in training is:", + "bbox": [ + 212, + 569, + 784, + 630 + ], + "page_idx": 10 + }, + { + "type": "equation", + "text": "\n$$\n\\mathcal {L} _ {\\text {t o t a l}} = \\beta_ {0} \\mathcal {L} _ {s} + \\beta_ {1} \\mathcal {L} _ {b l} + \\beta_ {2} \\mathcal {L} _ {b u}. \\tag {19}\n$$\n", + "text_format": "latex", + "bbox": [ + 387, + 641, + 784, + 657 + ], + "page_idx": 10 + }, + { + "type": "text", + "text": "5 Experiments", + "text_level": 1, + "bbox": [ + 215, + 707, + 374, + 724 + ], + "page_idx": 10 + }, + { + "type": "text", + "text": "5.1 Implementation Details", + "text_level": 1, + "bbox": [ + 215, + 739, + 455, + 753 + ], + "page_idx": 10 + }, + { + "type": "text", + "text": "The training dataset for the PUDE method contains 76 white-balanced linear underwater images collected in an open ocean environment with continuous depth variation, sourced from the SeaThru-NeRF [19] dataset. This scene structure is beneficial for the accurate estimation of the parameters in the underwater image formation model Eq. (1). We apply gamma correction to these images to", + "bbox": [ + 212, + 763, + 782, + 840 + ], + "page_idx": 10 + }, + { + "type": "header", + "text": "PUDE", + "bbox": [ + 684, + 114, + 730, + 126 + ], + "page_idx": 10 + }, + { + "type": "page_number", + "text": "11", + "bbox": [ + 767, + 116, + 782, + 126 + ], + "page_idx": 10 + }, + { + "type": "image", + "img_path": "images/c9b32c4b244a9104283fb6966adb740200bd240814590b41a5f0ca8065297a7b.jpg", + "image_caption": [ + "Fig. 5: Qualitative comparison on underwater images. Qualitative comparison of PUDE with in-air (MegaDepth, MiDaS, DPT) and underwater methods (UWNET, UDepth) on underwater images. PUDE detects clearer background edges and presents more depth details. (More in Supplementary)" + ], + "image_footnote": [], + "bbox": [ + 218, + 143, + 784, + 287 + ], + "page_idx": 11 + }, + { + "type": "text", + "text": "generate corresponding nonlinear images for training. We randomly sample 20 images with groundtruth from the Sea-thru D3 datasets to construct the validation dataset. During training, the nonlinear images are downsampled to the dimension of $576 \\times 384$ and fed into the neural network. PUDE is initialized with the DPT-Hybrid [28]. The training runs 3 epochs, employing the Adam optimizer with a learning rate of $10^{-5}$ . Notably, an early stopping is applied to prevent overfitting by looking at the validation results.", + "bbox": [ + 212, + 383, + 787, + 488 + ], + "page_idx": 11 + }, + { + "type": "text", + "text": "In testing, we compare the performance of the PUDE model with the underwater depth estimation methods, UDepth [37], UW-Net [15], and in-air methods with good generalization properties MiDaS V2.1 [29], MegaDepth [21] and DPT-Hybrid [28]. The comparison encompasses both quantitative and qualitative evaluations. The quantitative evaluation uses the remaining images from the Sea-thru D3 and the whole D5 datasets [1]. Another zero-shot testing is performed on the SQUID dataset from [4], including groundtruth measurements obtained through stereo-matching. Given the availability of RAW format images in these datasets, we applied gamma correction and brightness enhancement to the images. In the zero-shot qualitative evaluation, the images are sampled from the real underwater dataset [20].", + "bbox": [ + 212, + 489, + 787, + 656 + ], + "page_idx": 11 + }, + { + "type": "text", + "text": "5.2 Evaluations", + "text_level": 1, + "bbox": [ + 215, + 678, + 357, + 691 + ], + "page_idx": 11 + }, + { + "type": "text", + "text": "Quantitative Evaluation: We employ the alignment strategy described in [29] to compare the estimated inverse relative depth with the gorundtruth depth. This alignment operates in the space of inverse ground truth via solving an optimization problem for getting the shift and scale to align the estimated inverse relative depths with the inverse groundtruth. This alignment is subsequently inverted back to the normal depth space. To mitigate the impact of noise and erroneous measurements in groundtruth, we select depth thresholds ranging from $0.1m$ to $15m$ for the Sea-thru dataset [1] and $0.1m$ to $20m$ for the SQUID dataset [4] in our comparisons. We utilize the evaluation metrics described in [11] in our", + "bbox": [ + 212, + 703, + 787, + 840 + ], + "page_idx": 11 + }, + { + "type": "page_number", + "text": "12", + "bbox": [ + 217, + 114, + 235, + 126 + ], + "page_idx": 11 + }, + { + "type": "header", + "text": "J. Yang et al.", + "bbox": [ + 271, + 114, + 362, + 128 + ], + "page_idx": 11 + }, + { + "type": "image", + "img_path": "images/132809b95d12103d908ebd6410aaab47c97a2ff13f9af6164ae736a136607827.jpg", + "image_caption": [ + "Fig. 6: Ablation study: (a) Without using $\\mathcal{L}_{bu}$ , the model cannot clearly detect the background edges near the rocks and has a fogging appearance. (b) Without using $\\mathcal{L}_{bl}$ , the distant rocks cannot be detected, depth of that region is overestimated." + ], + "image_footnote": [], + "bbox": [ + 218, + 143, + 787, + 210 + ], + "page_idx": 12 + }, + { + "type": "text", + "text": "comparison: Mean absolute value of the relative error (AbsRel), Squared relative error (Sq. Rel), Root mean square error (RMSE), Root mean square logarithmic error (RMSE log), Scale-invariant logarithmic error (SILog) and Accuracy with threshold $(\\delta < thr)$ . Note that the thr in the comparison metric is set to be 1.05. The results are shown in Tab. 1. The experimental results demonstrate the superior performance of the PUDE model across all evaluated metrics on three datasets compared to other methods. It achieves lower errors and higher estimation accuracy, substantiating the effectiveness of our proposed approach. Particularly, PUDE demonstrates promising performance in zero-shot testing.", + "bbox": [ + 212, + 290, + 787, + 428 + ], + "page_idx": 12 + }, + { + "type": "text", + "text": "Qualitative Evaluation We also provide qualitative evaluation across underwater images collected from different environments. We sample some underwater images that were collected in a turbid environment from the dataset [20] and do the zero-shot evaluation. The results are shown in Fig. 5. The performance of our PUDE model in underwater depth estimation is superior, particularly in capturing detailed features. It effectively identifies background edges, including the fine contours of distant wires. These outcomes show that PUDE not only inherits robust generalizability from DPT but also incorporates a physics-based understanding of underwater imaging, enabling the PUDE model to discern background regions without requiring labelled training data.", + "bbox": [ + 212, + 449, + 787, + 602 + ], + "page_idx": 12 + }, + { + "type": "text", + "text": "5.3 Ablation Studies", + "text_level": 1, + "bbox": [ + 215, + 622, + 401, + 636 + ], + "page_idx": 12 + }, + { + "type": "text", + "text": "To investigate the influence of the bound losses in training, we retrain the model with and without bound losses and evaluate them on the Sea-thru D5 dataset. The results, shown in Tab. 2, demonstrate both the losses $\\mathcal{L}_{bu}$ and $\\mathcal{L}_{bl}$ contribute to training, which is consistent with our training goals, solving the issues", + "bbox": [ + 212, + 647, + 787, + 708 + ], + "page_idx": 12 + }, + { + "type": "table", + "img_path": "images/8cf5f66185eb1179808892fd904d1379dda65a23eb8fcf1b8b66d5435a1e30f2.jpg", + "table_caption": [ + "Table 2: Ablation study on D5 dataset. All bound losses contribute to the final performance of PUDE. PUDE has the best evaluation results on metrics." + ], + "table_footnote": [], + "table_body": "
LbuLblδ < 1.05δ < 1.052δ < 1.053AbsRelSq. RelRMSERMSE logSILog
✓0.3740.6380.7740.1240.0511.4790.2050.045
0.4260.6720.7980.1160.0481.4560.1990.042
✓0.3980.6570.7950.1090.0341.3830.1920.039
✓✓0.4350.6780.8030.1050.0341.3430.1860.037
", + "bbox": [ + 272, + 773, + 730, + 835 + ], + "page_idx": 12 + }, + { + "type": "header", + "text": "PUDE", + "bbox": [ + 684, + 114, + 730, + 126 + ], + "page_idx": 12 + }, + { + "type": "page_number", + "text": "13", + "bbox": [ + 767, + 114, + 784, + 126 + ], + "page_idx": 12 + }, + { + "type": "image", + "img_path": "images/c297734858541ffa16c4575fe0c884efc816fef65677e0c716be7662f0b9234b.jpg", + "image_caption": [ + "Fig. 7: Method generalization: PUDE(M) also has improved performance in edge detection and detail detection compared with the in-air model, MiDaS." + ], + "image_footnote": [], + "bbox": [ + 245, + 143, + 759, + 213 + ], + "page_idx": 13 + }, + { + "type": "text", + "text": "described in O1 and O2. Moreover, to visually demonstrate the effectiveness of bound losses $\\mathcal{L}_{bl}$ and $\\mathcal{L}_{bu}$ , we conducted a comparative analysis using images from the Sea-thru D5 dataset, as presented in Fig. 6. Without $\\mathcal{L}_{bl}$ , the depth in distant regions is overestimated. By contrast, without $\\mathcal{L}_{bu}$ , the detection of the background edges is not clear, resulting in a foggy appearance near the edges.", + "bbox": [ + 212, + 281, + 787, + 358 + ], + "page_idx": 13 + }, + { + "type": "text", + "text": "5.4 Method Generalization on Other In-air Models", + "text_level": 1, + "bbox": [ + 214, + 378, + 651, + 395 + ], + "page_idx": 13 + }, + { + "type": "text", + "text": "To further explore our method's generalization on different in-air models, we have conducted an additional experiment replacing DPT by MiDaS v2.1 [29] in the training loop, which has the same non-linearity described in Sec. 3 and demonstrate the second-best performance when applied to underwater images directly, and obtained PUDE(M). The quantitative and qualitative results are shown in Fig. 7 and Tab. 3. The results confirm that our method can also enhance the in-water performance of the non-SOTA model, MiDaS. However, as our method relies on knowledge transfer, the efficacy is contingent upon the capability of the original in-air model. A better model yields better transfer results.", + "bbox": [ + 212, + 405, + 787, + 541 + ], + "page_idx": 13 + }, + { + "type": "text", + "text": "6 Conclusion", + "text_level": 1, + "bbox": [ + 214, + 564, + 359, + 580 + ], + "page_idx": 13 + }, + { + "type": "text", + "text": "This paper proposes a novel method for transferring the in-air depth estimation model, DPT, to underwater settings. We observe DPT has issues with overestimating depth for distant regions and unclear background edge detection on underwater images. To tackle these, we use the novel bound losses to enforce the model to produce depth estimation that follows the underwater imaging formation model. The final quantitative and qualitative evaluation results prove the effectiveness of our method. Additionally, the further experiment shows our method's generalization capabilities on other in-air models.", + "bbox": [ + 212, + 595, + 787, + 717 + ], + "page_idx": 13 + }, + { + "type": "table", + "img_path": "images/e795be7e705be203157ac77a59a18c01dc51b7fa73c6b95d8b9cd57966e3ad0f.jpg", + "table_caption": [ + "Table 3: PUDE(M) on SQUID dataset: PUDE(M) shows improved performance on the metrics compared to MiDaS, but PUDE has the best performance." + ], + "table_footnote": [], + "table_body": "
Modelsδ < 1.05δ < 1.052δ < 1.053AbsRelSq. RelRMSERMSE logSILog
MiDaS0.1960.3790.5070.2960.3472.4440.3090.139
PUDE(M)0.2790.4630.6010.1880.1151.8690.2210.067
PUDE0.4350.6780.8030.1050.0341.3430.1860.037
", + "bbox": [ + 266, + 785, + 736, + 837 + ], + "page_idx": 13 + }, + { + "type": "page_number", + "text": "14", + "bbox": [ + 217, + 114, + 235, + 126 + ], + "page_idx": 13 + }, + { + "type": "header", + "text": "J. Yang et al.", + "bbox": [ + 271, + 114, + 362, + 128 + ], + "page_idx": 13 + }, + { + "type": "text", + "text": "Acknowledgements", + "text_level": 1, + "bbox": [ + 217, + 143, + 401, + 162 + ], + "page_idx": 14 + }, + { + "type": "text", + "text": "This work received funding from the Australian Research Council via grant DE220101527, and the Australian Government, via grant AUSMURIB000001.", + "bbox": [ + 215, + 178, + 782, + 208 + ], + "page_idx": 14 + }, + { + "type": "text", + "text": "References", + "text_level": 1, + "bbox": [ + 217, + 233, + 321, + 250 + ], + "page_idx": 14 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "1. Akkaynak, D., Treibitz, T.: Sea-thru: A method for removing water from underwater images. In: Proceedings of the IEEE/CVF conference on computer vision and pattern recognition. pp. 1682-1691 (2019)", + "2. Amitai, S., Klein, I., Treibitz, T.: Self-supervised monocular depth underwater. In: 2023 IEEE International Conference on Robotics and Automation (ICRA). pp. 1098-1104. IEEE (2023)", + "3. Ancuti, C.O., Ancuti, C., De Vleeschouwer, C., Neumann, L., Garcia, R.: Color transfer for underwater dehazing and depth estimation. In: 2017 IEEE International Conference on Image Processing (ICIP). pp. 695-699. IEEE (2017)", + "4. Berman, D., Levy, D., Avidan, S., Treibitz, T.: Underwater single image color restoration using haze-lines and a new quantitative dataset. IEEE Transactions on Pattern Analysis and Machine Intelligence (2020)", + "5. Bhat, S.F., Alhashim, I., Wonka, P.: Adabins: Depth estimation using adaptive bins. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 4009-4018 (2021)", + "6. Bhat, S.F., Birkl, R., Wofk, D., Wonka, P., Müller, M.: Zoedepth: Zero-shot transfer by combining relative and metric depth. arXiv preprint arXiv:2302.12288 (2023)", + "7. Chang, H.H., Cheng, C.Y., Sung, C.C.: Single underwater image restoration based on depth estimation and transmission compensation. IEEE Journal of Oceanic Engineering 44(4), 1130-1149 (2018)", + "8. Chen, W., Qian, S., Deng, J.: Learning single-image depth from videos using quality assessment networks. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 5604-5613 (2019)", + "9. Chiang, J.Y., Chen, Y.C.: Underwater image enhancement by wavelength compensation and dehazing. IEEE transactions on image processing 21(4), 1756-1769 (2011)", + "0. Drews, P.L., Nascimento, E.R., Botelho, S.S., Campos, M.F.M.: Underwater depth estimation and image restoration based on single images. IEEE computer graphics and applications 36(2), 24-35 (2016)", + "1. Eigen, D., Puhrsch, C., Fergus, R.: Depth map prediction from a single image using a multi-scale deep network. Advances in neural information processing systems 27 (2014)", + "2. Fu, H., Gong, M., Wang, C., Batmanghelich, K., Tao, D.: Deep ordinal regression network for monocular depth estimation. In: Proceedings of the IEEE conference on computer vision and pattern recognition. pp. 2002-2011 (2018)", + "3. Godard, C., Mac Aodha, O., Brostow, G.J.: Unsupervised monocular depth estimation with left-right consistency. In: Proceedings of the IEEE conference on computer vision and pattern recognition. pp. 270-279 (2017)", + "4. Godard, C., Mac Aodha, O., Firman, M., Brostow, G.J.: Digging into self-supervised monocular depth estimation. In: Proceedings of the IEEE/CVF international conference on computer vision. pp. 3828-3838 (2019)" + ], + "bbox": [ + 225, + 266, + 784, + 839 + ], + "page_idx": 14 + }, + { + "type": "header", + "text": "PUDE", + "bbox": [ + 684, + 114, + 730, + 126 + ], + "page_idx": 14 + }, + { + "type": "page_number", + "text": "15", + "bbox": [ + 767, + 116, + 784, + 126 + ], + "page_idx": 14 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "15. Gupta, H., Mitra, K.: Unsupervised single image underwater depth estimation. In: 2019 IEEE International Conference on Image Processing (ICIP). pp. 624-628. IEEE (2019)", + "16. Hambarde, P., Murala, S., Dhall, A.: Uw-gan: Single-image depth estimation and image enhancement for underwater images. IEEE Transactions on Instrumentation and Measurement 70, 1-12 (2021)", + "17. He, K., Sun, J., Tang, X.: Single image haze removal using dark channel prior. IEEE transactions on pattern analysis and machine intelligence 33(12), 2341-2353 (2010)", + "18. Laina, I., Rupprecht, C., Belagiannis, V., Tombari, F., Navab, N.: Deeper depth prediction with fully convolutional residual networks. In: 2016 Fourth international conference on 3D vision (3DV). pp. 239-248. IEEE (2016)", + "19. Levy, D., Peleg, A., Pearl, N., Rosenbaum, D., Akkaynak, D., Korman, S., Treibitz, T.: Seathru-nerf: Neural radiance fields in scattering media. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 56-65 (2023)", + "20. Li, C., Guo, C., Ren, W., Cong, R., Hou, J., Kwong, S., Tao, D.: An underwater image enhancement benchmark dataset and beyond. IEEE Transactions on Image Processing 29, 4376-4389 (2019)", + "21. Li, Z., Snavely, N.: Megadepth: Learning single-view depth prediction from internet photos. In: Proceedings of the IEEE conference on computer vision and pattern recognition. pp. 2041-2050 (2018)", + "22. Luo, Y., Ren, J., Lin, M., Pang, J., Sun, W., Li, H., Lin, L.: Single view stereo matching. In: Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition. pp. 155-163 (2018)", + "23. Ming, Y., Meng, X., Fan, C., Yu, H.: Deep learning for monocular depth estimation: A review. Neurocomputing 438, 14-33 (2021)", + "24. Peng, Y.T., Cao, K., Cosman, P.C.: Generalization of the dark channel prior for single image restoration. IEEE Transactions on Image Processing 27(6), 2856-2868 (2018)", + "25. Peng, Y.T., Cosman, P.C.: Underwater image restoration based on image blurriness and light absorption. IEEE transactions on image processing 26(4), 1579-1594 (2017)", + "26. Peng, Y.T., Zhao, X., Cosman, P.C.: Single underwater image enhancement using depth estimation based on blurriness. In: 2015 IEEE International Conference on Image Processing (ICIP). pp. 4952-4956. IEEE (2015)", + "27. Ramamonjisoa, M., Lepetit, V.: Sharpnet: Fast and accurate recovery of occluding contours in monocular depth estimation. In: Proceedings of the IEEE/CVF International Conference on Computer Vision Workshops. pp. 0-0 (2019)", + "28. Ranftl, R., Bochkovskiy, A., Koltun, V.: Vision transformers for dense prediction. In: Proceedings of the IEEE/CVF international conference on computer vision. pp. 12179-12188 (2021)", + "29. Ranftl, R., Lasinger, K., Hafner, D., Schindler, K., Koltun, V.: Towards robust monocular depth estimation: Mixing datasets for zero-shot cross-dataset transfer. IEEE transactions on pattern analysis and machine intelligence 44(3), 1623-1637 (2020)", + "30. Varghese, N., Kumar, A., Rajagopalan, A.: Self-supervised monocular underwater depth recovery, image restoration, and a real-sea video dataset. In: Proceedings of the IEEE/CVF International Conference on Computer Vision. pp. 12248-12258 (2023)" + ], + "bbox": [ + 215, + 146, + 784, + 839 + ], + "page_idx": 15 + }, + { + "type": "page_number", + "text": "16", + "bbox": [ + 217, + 114, + 235, + 126 + ], + "page_idx": 15 + }, + { + "type": "header", + "text": "J. Yang et al.", + "bbox": [ + 271, + 114, + 362, + 128 + ], + "page_idx": 15 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "31. Vasiljevic, I., Kolkin, N., Zhang, S., Luo, R., Wang, H., Dai, F.Z., Daniele, A.F., Mostajabi, M., Basart, S., Walter, M.R., Shakhnarovich, G.: DIODE: A Dense Indoor and Outdoor DEpth Dataset. CoRR abs/1908.00463 (2019), http:// arxiv.org/abs/1908.00463", + "32. Wu, C.Y., Wang, J., Hall, M., Neumann, U., Su, S.: Toward practical monocular indoor depth estimation. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 3814-3824 (2022)", + "33. Wu, Y., Zhou, Y., Chen, S., Ma, Y., Li, Q.: Defect inspection for underwater structures based on line-structured light and binocular vision. Applied Optics 60(25), 7754-7764 (2021)", + "34. Xian, K., Shen, C., Cao, Z., Lu, H., Xiao, Y., Li, R., Luo, Z.: Monocular relative depth perception with web stereo data supervision. In: Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition. pp. 311-320 (2018)", + "35. Yang, J., Gong, M., Nair, G., Lee, J.H., Monty, J., Pu, Y.: Knowledge distillation for feature extraction in underwater vslam. arXiv preprint arXiv:2303.17981 (2023)", + "36. Ye, X., Zhang, J., Yuan, Y., Xu, R., Wang, Z., Li, H.: Underwater depth estimation via stereo adaptation networks. IEEE Transactions on Circuits and Systems for Video Technology (2023)", + "37. Yu, B., Wu, J., Islam, M.J.: Udepth: Fast monocular depth estimation for visually-guided underwater robots. In: 2023 IEEE International Conference on Robotics and Automation (ICRA). pp. 3116-3123. IEEE (2023)" + ], + "bbox": [ + 212, + 146, + 787, + 438 + ], + "page_idx": 16 + }, + { + "type": "header", + "text": "PUDE", + "bbox": [ + 684, + 114, + 730, + 126 + ], + "page_idx": 16 + }, + { + "type": "page_number", + "text": "17", + "bbox": [ + 767, + 114, + 785, + 126 + ], + "page_idx": 16 + } +] \ No newline at end of file diff --git a/2024/Physics-informed Knowledge Transfer for Underwater Monocular Depth Estimation/58639476-9724-4b09-9db4-e436240a5f18_model.json b/2024/Physics-informed Knowledge Transfer for Underwater Monocular Depth Estimation/58639476-9724-4b09-9db4-e436240a5f18_model.json new file mode 100644 index 0000000000000000000000000000000000000000..581b87b377d12c3602069e9e001838ffa7bae632 --- /dev/null +++ b/2024/Physics-informed Knowledge Transfer for Underwater Monocular Depth Estimation/58639476-9724-4b09-9db4-e436240a5f18_model.json @@ -0,0 +1,2368 @@ +[ + [ + { + "type": "title", + "bbox": [ + 0.264, + 0.141, + 0.74, + 0.186 + ], + "angle": 0, + "content": "Physics-informed Knowledge Transfer for Underwater Monocular Depth Estimation" + }, + { + "type": "text", + "bbox": [ + 0.35, + 0.213, + 0.652, + 0.228 + ], + "angle": 0, + "content": "Jinghe Yang, Mingming Gong, and Ye Pu" + }, + { + "type": "text", + "bbox": [ + 0.236, + 0.24, + 0.765, + 0.269 + ], + "angle": 0, + "content": "The University of Melbourne jinghey@student.unimelb.edu.au, {mingming.gong, ye.pu}@unimelb.edu.au" + }, + { + "type": "text", + "bbox": [ + 0.263, + 0.299, + 0.756, + 0.535 + ], + "angle": 0, + "content": "Abstract. Compared to the in-air case, underwater depth estimation has its own challenges. For instance, acquiring high-quality training datasets with groundtruth poses difficulties due to sensor limitations in aquatic environments. Additionally, the physics characteristics of underwater imaging diverge significantly from the in-air case, the methods developed for in-air depth estimation underperform when applied underwater, due to the domain gap. To address these challenges, our paper introduces a novel transfer-learning-based method - Physics-informed Underwater Depth Estimation (PUDE). The key idea is to transfer the knowledge of a pre-trained in-air depth estimation model to underwater settings utilizing a small underwater image set without groundtruth measurement, guided by a physical underwater imaging formation model. We propose novel bound losses based on the physical model to rectify the depth estimations to align with actual underwater physical properties. Finally, in the evaluations across multiple datasets, we compare PUDE model with other existing in-air and underwater methods. The results reveal that the PUDE model excels in both quantitative and qualitative comparisons." + }, + { + "type": "text", + "bbox": [ + 0.263, + 0.548, + 0.741, + 0.562 + ], + "angle": 0, + "content": "Keywords: Underwater Depth \\(\\cdot\\) Physics-informed \\(\\cdot\\) Knowledge Transfer" + }, + { + "type": "title", + "bbox": [ + 0.218, + 0.585, + 0.377, + 0.601 + ], + "angle": 0, + "content": "1 Introduction" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.614, + 0.788, + 0.844 + ], + "angle": 0, + "content": "Depth estimation is a pivotal task in computer vision, with substantial implications for robotics applications. Traditional techniques typically employ stereo cameras to estimate depth through disparity calculations. However, these methods are hampered by complex camera system settings and limited by the challenging feature matching in environments with low texture or visibility [23]. In contrast, monocular depth estimation provides an alternative solution, necessitating only a single camera and a less complicated setup. Although monocular depth estimation faces inherent limitations due to the absence of stereoscopic information, many well-performing in-air learning-based depth estimation models have been developed in recent years [5,6,12,21,27-29]. Benefiting from the massive and various training datasets, some of in-air monocular depth estimation models, such as MiDaS [29] and DPT [28], demonstrate powerful generalization performance. However, the pre-trained in-air models are not fully applicable for underwater cases due to the domain gap and the differences in image formation [36]. For example, we applied the in-air trained models, e.g., DPT in [28], to" + } + ], + [ + { + "type": "page_number", + "bbox": [ + 0.218, + 0.116, + 0.23, + 0.127 + ], + "angle": 0, + "content": "2" + }, + { + "type": "header", + "bbox": [ + 0.272, + 0.115, + 0.364, + 0.129 + ], + "angle": 0, + "content": "J. Yang et al." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.147, + 0.788, + 0.449 + ], + "angle": 0, + "content": "underwater images and observed that DPT has the issues of the overestimation for blurry regions, unclear background region detection and loss of depth details. Additionally, it is challenging to train new models from scratch for underwater settings. Underwater scenarios impose difficulties when capturing a large amount of high-quality images and the groundtruth depth, due to sensor limitations, such as the incapability of LiDAR and the performance drop of stereo matching in water. In contrast, sonar scanning is not affected by water quality and blurriness but suffers from a slow scanning rate and low resolution, lacking detail in scene representation [33]. In response to the distinct challenges of underwater depth estimation, this paper proposes a novel transfer-learning-based approach. Our method transfers knowledge from a pre-trained in-air monocular depth estimation model to underwater settings, guided by the underwater imaging formation model introduced in [9]. Using newly proposed bound losses, which leverage physical information to correct inaccurate depth in distant and background regions, our training method enforces the model to produce estimates aligned with the extracted physical properties. Notably, this adaptation requires minimal training data and no ground truth depth measurements. Our approach benefits from the structural knowledge of the in-air model and integrates underwater physics knowledge, thereby enhancing underwater depth estimation performance. Our contributions are as follows:" + }, + { + "type": "text", + "bbox": [ + 0.226, + 0.459, + 0.788, + 0.579 + ], + "angle": 0, + "content": "- We present experimental observations and limitations when applying an in-air trained model, i.e., DPT in [28], to underwater images. Inspired by these observations, we propose a novel training method to adapt in-air pre-trained depth estimation models for underwater environments. There are two key ingredients in our method: 1) Knowledge transfer using an in-air trained model with good generalisation properties (DPT [28]) to address data scarcity; 2) Integration of the underwater image formation model proposed in [9] to ensure the depth estimates align with underwater physics characteristics." + }, + { + "type": "text", + "bbox": [ + 0.226, + 0.58, + 0.788, + 0.61 + ], + "angle": 0, + "content": "- We propose innovative bound losses to steer the model towards producing depth estimations that adhere to the underwater imaging formation model." + }, + { + "type": "text", + "bbox": [ + 0.226, + 0.61, + 0.788, + 0.685 + ], + "angle": 0, + "content": "- We introduce the Physics-informed Underwater Depth Estimation (PUDE) model, an end-to-end model for underwater monocular depth estimation. In zero-shot evaluations across various datasets, our method demonstrates enhanced performance, both quantitatively and qualitatively, compared to other in-air and underwater methods." + }, + { + "type": "list", + "bbox": [ + 0.226, + 0.459, + 0.788, + 0.685 + ], + "angle": 0, + "content": null + }, + { + "type": "title", + "bbox": [ + 0.216, + 0.708, + 0.381, + 0.724 + ], + "angle": 0, + "content": "2 RelatedWork" + }, + { + "type": "title", + "bbox": [ + 0.216, + 0.74, + 0.557, + 0.755 + ], + "angle": 0, + "content": "2.1 In-air Monocular Depth Estimation" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.765, + 0.788, + 0.841 + ], + "angle": 0, + "content": "Compared to LiDAR, infrared, and stereo vision, monocular depth estimation stands out for its cost-effectiveness and simplicity of configuration, making it appealing for depth perception in autonomous systems. Recent in-air monocular depth estimation methods have exhibited commendable performance, especially supervised methods trained on datasets containing measured groundtruth. They" + } + ], + [ + { + "type": "header", + "bbox": [ + 0.685, + 0.114, + 0.732, + 0.127 + ], + "angle": 0, + "content": "PUDE" + }, + { + "type": "page_number", + "bbox": [ + 0.775, + 0.116, + 0.787, + 0.127 + ], + "angle": 0, + "content": "3" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.147, + 0.784, + 0.192 + ], + "angle": 0, + "content": "usually require sensors like LiDAR, RGBD cameras, and stereo matching [11,12, 18,22]. However, these methods are limited by the lack of variety in the datasets, which restricts their generalization across different real-world scenarios." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.192, + 0.788, + 0.374 + ], + "angle": 0, + "content": "In this context, self-supervised learning using continuous image sequences [14] or calibrated stereo image pairs [13] presents an attractive avenue for exploration but performs less compared to supervised learning [23], and its application to dynamic scenes is challenging. Recognizing this limitation, some studies expand the breadth of training data by harnessing web resources to generate depth groundtruth, thereby enlarging the training dataset and encompassing diverse scenes [8,21,34]. Building further on that, MiDaS [29] extract depth data from 3D movies to enhance dataset diversity and improve generalization. DPT [28] builds upon the MiDaS training dataset, but with supplementary, and implements an encoder-decoder architecture that uses a vision transformer as its backbone, enabling fine-grained and globally coherent estimations and showing strong zero-shot cross-dataset generalization capabilities." + }, + { + "type": "title", + "bbox": [ + 0.215, + 0.394, + 0.607, + 0.409 + ], + "angle": 0, + "content": "2.2 Underwater Monocular Depth Estimation" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.418, + 0.784, + 0.522 + ], + "angle": 0, + "content": "Depth measurement in water introduces new challenges due to the unique characteristics of the underwater environment. The limited availability of sensors, such as LiDAR and infrared, makes the direct acquisition of the depth measurement highly challenging. Meanwhile, the image blurriness highly influences the correspondence matching between images [35], thereby the stereo camera can only provide good depth measurement for close regions. The existing underwater depth datasets that rely on stereo matching [1,4], manifest this issue." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.524, + 0.788, + 0.735 + ], + "angle": 0, + "content": "Therefore, monocular depth estimation is a potential solution to tackle these issues. Physics-based methods play a pivotal role in addressing the underwater depth estimation challenge. The underwater image formation model proposed in [9] has been demonstrated to be effective in solving problems in computer vision for underwater scenarios, such as feature extraction [35], neural radiance fields (NeRFs) [19], and depth estimation [10]. Traditional methods utilize the physics-based image formation model to estimate depth directly from underwater images [3,4,7,24-26]. As the foundation of these methods, Dark Channel Prior (DCP) [17], a statistical method initially developed for in-air haze removal, uses pixel intensity to estimate medium transmission. A variant for the underwater environment, UDCP [10], claims that the DCP assumption remains applicable to the green and blue channels of underwater images. However, underwater depth estimation that relies solely on pixel values and the imaging formation model is vulnerable to variable lighting conditions, such as shadows and sunlight flickers." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.735, + 0.788, + 0.84 + ], + "angle": 0, + "content": "Despite the challenges in depth groundtruth acquisition underwater, the learning-based methods show promise. Some works have attempted unsupervised learning, such as UwNET [15] and UW-GAN [16]. They use the GAN method through color restoration to estimate depth. The recent works [2,30] adopt self-supervised learning, utilizing the correspondences between subsequent frames. Notably, they incorporate the physics model into the training of the underwater depth estimation models and have shown to be advantageous." + } + ], + [ + { + "type": "page_number", + "bbox": [ + 0.218, + 0.116, + 0.23, + 0.127 + ], + "angle": 0, + "content": "4" + }, + { + "type": "header", + "bbox": [ + 0.272, + 0.115, + 0.364, + 0.129 + ], + "angle": 0, + "content": "J. Yang et al." + }, + { + "type": "title", + "bbox": [ + 0.216, + 0.147, + 0.583, + 0.163 + ], + "angle": 0, + "content": "2.3 Underwater Imaging Formation Model" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.172, + 0.788, + 0.28 + ], + "angle": 0, + "content": "Light transmission in water is subject to absorption, and minuscule particles cause the backscattering, leading to blur within underwater imagery. A widely used model of underwater image formation is proposed in [9] to describe this physical phenomenon. Consider an underwater image represented by \\( I_{c} \\in \\mathbb{R}^{HW} \\), where \\( c \\) denotes the color channels \\( \\{\\mathrm{R}, \\mathrm{G}, \\mathrm{B}\\} \\), and \\( H \\) and \\( W \\) are the image's height and width, respectively. Given an image \\( I_{c} \\), for each pixel \\( i \\), \\( i = 1 \\cdots HW \\), it holds that \\( I_{c,i} \\in [0,1] \\). The underwater imaging formation model is:" + }, + { + "type": "equation", + "bbox": [ + 0.395, + 0.291, + 0.786, + 0.308 + ], + "angle": 0, + "content": "\\[\nI _ {c, i} = J _ {c, i} \\cdot t _ {c, i} + B _ {c} ^ {\\infty} (1 - t _ {c, i}), \\tag {1}\n\\]" + }, + { + "type": "equation", + "bbox": [ + 0.455, + 0.332, + 0.786, + 0.351 + ], + "angle": 0, + "content": "\\[\nt _ {c, i} = e ^ {- \\beta_ {c} \\mathbf {z} _ {i}}. (2)\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.356, + 0.788, + 0.478 + ], + "angle": 0, + "content": "In this model, \\( J_{c,i} \\) is the original in-air color value at pixel \\( i \\). The original in-air image is denoted by \\( J_{c} = [J_{c,1},\\dots,J_{c,HW}] \\in \\mathbb{R}^{HW} \\). The medium transmission rate denoted by \\( t_{c,i} \\) is associated with the beam attenuation coefficient \\( \\beta_{c} \\) and the depth \\( \\mathbf{z}_{i} \\). Meanwhile, \\( B_{c}^{\\infty} \\) is the coefficient of ambient light causing the backscattering from particles. The captured underwater image \\( I_{c,i} \\) is the aggregate of the attenuation of the original image and the additional backscattering from the water environment. The values of \\( \\beta_{c} \\) and \\( B_{c}^{\\infty} \\) vary depending on different water environments and can be estimated using the depth measurements [1]." + }, + { + "type": "title", + "bbox": [ + 0.216, + 0.501, + 0.666, + 0.518 + ], + "angle": 0, + "content": "3 DPT on Underwater Images: Observations" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.532, + 0.788, + 0.64 + ], + "angle": 0, + "content": "Existing learning-based monocular depth estimation methods, such as DPT [28] and MiDaS [29], generate the inverse relative depth estimation from a single image. We denote the inverse depth estimation as \\(\\mathbf{d} \\in \\mathbb{R}^{HW}\\), with \\(\\mathbf{d}_i\\) to be the estimate for pixel \\(i\\), \\(i = 1, \\dots, HW\\). The metric depth of the input image is denoted by \\(\\mathbf{z} \\in \\mathbb{R}^{HW}\\), whose reciprocal ideally has a linear relationship with the inverse relative depth. The relationship between inverse relative depth \\(\\mathbf{d}_i\\) and the metric depth \\(\\mathbf{z}_i\\) at pixel \\(i\\) is:" + }, + { + "type": "equation", + "bbox": [ + 0.376, + 0.65, + 0.787, + 0.681 + ], + "angle": 0, + "content": "\\[\n\\mathbf {z} _ {i} = \\frac {1}{s \\mathbf {d} _ {i} + h}, \\quad \\forall i \\in \\{1, \\dots , H W \\}. \\tag {3}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.689, + 0.789, + 0.841 + ], + "angle": 0, + "content": "Here, \\( s \\) and \\( h \\) are two unknown scale and shift coefficients. In practice, they can be off-line estimated for specific scenes using sequence images. The overall mapping relationship between the inverse relative depth \\( \\mathbf{d}_i \\) and the inverse of metric depth \\( \\mathbf{z}_i \\) is linear on the in-air images [32]. An example can be seen in the first row of Fig. 1, which is the depth estimation of DPT on an outdoor in-air image from DIODE [31], we can observe that DPT estimation on the in-air image maintains a good linear relationship (the correlation coefficient \\( r = 0.992 \\)). However, when we apply DPT directly to underwater images from the dataset [1,4], in which the depth groundtruth is obtained through the Structure-From-Motion and stereo matching, it is observed that in underwater imagery, the estimated" + } + ], + [ + { + "type": "header", + "bbox": [ + 0.686, + 0.115, + 0.732, + 0.127 + ], + "angle": 0, + "content": "PUDE" + }, + { + "type": "page_number", + "bbox": [ + 0.775, + 0.117, + 0.785, + 0.127 + ], + "angle": 0, + "content": "5" + }, + { + "type": "image", + "bbox": [ + 0.245, + 0.144, + 0.416, + 0.233 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.417, + 0.145, + 0.588, + 0.233 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.587, + 0.149, + 0.747, + 0.233 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.246, + 0.233, + 0.416, + 0.322 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.418, + 0.233, + 0.588, + 0.322 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.587, + 0.238, + 0.742, + 0.322 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.215, + 0.332, + 0.785, + 0.375 + ], + "angle": 0, + "content": "Fig. 1: DPT estimation of underwater images shows a curved tendency, with overestimation in distant regions. Additionally, the estimated depth struggles with unclear background detection (blue arrows)." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.41, + 0.785, + 0.459 + ], + "angle": 0, + "content": "inverse depth, denoted by \\(\\mathbf{d}_i^D\\), and \\(\\frac{1}{z_i}\\) deviates from a linear relationship and present a curve tendency with a weaker correlation, especially when \\(\\mathbf{d}_i^D\\) is small. An example is shown in the second row of Fig. 1." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.462, + 0.788, + 0.675 + ], + "angle": 0, + "content": "To observe the nonlinear trend caused by image blurriness, we applied the Sea-thru method [1] to calculate the medium transmission rate \\( t_i \\) in Eq. (2) on the grayscale images. We then compared these rates with those obtained from the DPT depth (aligned with groundtruth depth). An example of this procedure on the Sea-thru D3 dataset is shown in Fig. 2a, where we randomly sample 100 points from each image. We observe that the estimated transmission rate from DPT is mostly lower than that calculated by groundtruth depth when the transmission rate is small (\\( t_i < 0.2 \\)). The results reveal that DPT underestimates the transmission rate in regions with low transmission rates, leading to an overestimation of depth in distant regions. Furthermore, the background edges on the underwater depth image of DPT, such as the blue arrow pointed regions around corals, are less clear, exhibiting fogging compared with the estimation on the in-air image. Based on these experimental observations, we summarise the primary issues of the depth estimation from DPT in water:" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.696, + 0.785, + 0.741 + ], + "angle": 0, + "content": "O1 : DPT exhibits a propensity for overestimating depth in regions that are relatively far away from the camera. This overestimation is particularly evident in regions with diminished visual clarity and increased blurriness." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.745, + 0.785, + 0.806 + ], + "angle": 0, + "content": "O2 : Secondly, DPT encounters challenges in accurately discerning background regions in underwater images. This issue is especially pronounced for the region far from the camera. Due to the attenuation of the scenes, it is challenging for DPT to define the background area clearly." + }, + { + "type": "list", + "bbox": [ + 0.214, + 0.696, + 0.785, + 0.806 + ], + "angle": 0, + "content": null + }, + { + "type": "text", + "bbox": [ + 0.215, + 0.825, + 0.774, + 0.84 + ], + "angle": 0, + "content": "We propose two bound losses (Sec. 4.3) in our training to address these issues." + } + ], + [ + { + "type": "page_number", + "bbox": [ + 0.218, + 0.116, + 0.23, + 0.127 + ], + "angle": 0, + "content": "6" + }, + { + "type": "header", + "bbox": [ + 0.272, + 0.115, + 0.364, + 0.129 + ], + "angle": 0, + "content": "J. Yang et al." + }, + { + "type": "image", + "bbox": [ + 0.235, + 0.149, + 0.364, + 0.246 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.218, + 0.248, + 0.39, + 0.261 + ], + "angle": 0, + "content": "(a) \\(t_i\\) in Eq. (2) (D3 dataset)" + }, + { + "type": "image", + "bbox": [ + 0.425, + 0.151, + 0.613, + 0.248 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.434, + 0.248, + 0.756, + 0.261 + ], + "angle": 0, + "content": "(b) Point selection and model fitting in Eq. (1), Eq. (3)" + }, + { + "type": "image", + "bbox": [ + 0.613, + 0.152, + 0.756, + 0.247 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.214, + 0.272, + 0.788, + 0.327 + ], + "angle": 0, + "content": "Fig.2: (a) The comparison between the calculated transmission rate \\( t_i \\) from groundtruth and the estimated \\( t_i \\) from DPT depth, presents an underestimation of the \\( t_i \\) for small values of DPT. (b) The left image shows the points selected in the refined estimation. The right shows the fitting of the estimation." + }, + { + "type": "title", + "bbox": [ + 0.216, + 0.356, + 0.331, + 0.373 + ], + "angle": 0, + "content": "4 Method" + }, + { + "type": "title", + "bbox": [ + 0.216, + 0.39, + 0.439, + 0.404 + ], + "angle": 0, + "content": "4.1 Framework Overview" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.417, + 0.788, + 0.613 + ], + "angle": 0, + "content": "Inspired by the observations in Sec. 3, we propose a physics-informed knowledge transfer method for underwater monocular depth estimation, named the Physics-informed Underwater Depth Estimation (PUDE). The main idea of this physics-informed method is to leverage the physical information in underwater imagery to guide knowledge transfer within the neural network, thereby enhancing its adaptation from in-air to underwater environments. Specifically, we adopt the pre-trained DPT-Hybrid model [28], a state-of-the-art (SOTA) model known for its good generalization, as the transferred model. Benefiting from the massive in-air training dataset, the DPT model has powerful spatial structural knowledge. Despite previously noted issues when applied to underwater images, it can still provide a relatively reasonable depth estimation, particularly for regions close to the camera. This estimation serves as our prior depth knowledge for extracting the physical information from the underwater images." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.614, + 0.788, + 0.75 + ], + "angle": 0, + "content": "Most daily images are non-linear images, which are automatically post-processed by nonlinear algorithms integrated into cameras, for example, gamma correction, to approach the non-linear perception of human eyes. By contrast, linear images are images in which pixel values are directly captured by the camera without nonlinear post-processing or only with ratio processing. To ensure the PUDE model's generalizability in reality and keep consistency in training and testing, we utilize post-processed non-linear underwater images as inputs. In contrast, for precise physical information extraction, we utilize the linear underwater images in the estimation of the physics parameters in Eq. (1) and Eq. (3)." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.75, + 0.788, + 0.842 + ], + "angle": 0, + "content": "Our training framework consists of two branches shown in Fig. 3. The first branch is the direct knowledge transfer from DPT to PUDE, and we employ a similarity loss to guarantee the similarity between DPT and PUDE in the transfer (in Sec. 4.2). The second branch is the physics-informed knowledge transfer. Inspired by the parameter estimation method in Sea-thru [1], we estimate the parameters in Eq. (1) and Eq. (3) for each input underwater image, utilizing" + } + ], + [ + { + "type": "header", + "bbox": [ + 0.686, + 0.115, + 0.732, + 0.127 + ], + "angle": 0, + "content": "PUDE" + }, + { + "type": "page_number", + "bbox": [ + 0.776, + 0.117, + 0.786, + 0.127 + ], + "angle": 0, + "content": "7" + }, + { + "type": "image", + "bbox": [ + 0.223, + 0.146, + 0.638, + 0.309 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.695, + 0.147, + 0.781, + 0.298 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.214, + 0.321, + 0.788, + 0.378 + ], + "angle": 0, + "content": "Fig. 3: Overall training framework for PUDE. The overall training framework comprises two components: the direct knowledge transfer from the DPT model and the physi-informed knowledge transfer based on the DPT estimation. Physical information is extracted from DPT's output to establish the bound losses." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.403, + 0.784, + 0.495 + ], + "angle": 0, + "content": "DPT depth estimation. We restrict our usage to the green and blue channels, in line with the conclusions from UDCP [10]. Based on that, we propose two novel bound losses, utilizing the physics information contained in the underwater image formation model in Eq. (1) to enforce the model to produce plausible depth estimations that conform to the physics characteristics of underwater imaging. In Sec. 4.3, we will introduce the second branch and the designed bound losses." + }, + { + "type": "title", + "bbox": [ + 0.216, + 0.513, + 0.485, + 0.529 + ], + "angle": 0, + "content": "4.2 Direct Knowledge Transfer" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.535, + 0.788, + 0.673 + ], + "angle": 0, + "content": "The direct knowledge transfer loop aims to guarantee the similarity between the original DPT and PUDE during the transition to ensure the structural knowledge can be kept in PUDE. We use a similarity loss in the training process to achieve this goal. In the transfer, PUDE is initialized with DPT. We trim the pixels with the largest differences to set the tolerance for the difference between PUDE and the original model DPT during the training. We denote the pixel set after trimming as \\( M_{trim} \\subseteq \\{1, \\dots, HW\\} \\), and \\( T \\) representing the cardinality of \\( M_{trim} \\). For each pixel \\( i \\in M_{trim} \\), we denote the DPT and PUDE estimation as \\( \\mathbf{d}_i^D \\) and \\( \\mathbf{d}_i^P \\), respectively. The similarity loss \\( L_s \\) is defined as:" + }, + { + "type": "equation", + "bbox": [ + 0.389, + 0.681, + 0.787, + 0.721 + ], + "angle": 0, + "content": "\\[\nL _ {s} = \\frac {1}{T} \\sum_ {i \\in M _ {\\mathrm {t r i m}}} \\left| \\frac {\\mathbf {d} _ {i} ^ {P} - \\mathbf {d} _ {i} ^ {D}}{\\operatorname {m e d i a n} (\\mathbf {d} ^ {D})} \\right|. \\tag {4}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.724, + 0.788, + 0.77 + ], + "angle": 0, + "content": "The trimming sets the tolerance for the changing between PUDE and the original model DPT during the training. In our implementation, we set the trimming rate to be \\(0.3HW\\)." + }, + { + "type": "title", + "bbox": [ + 0.216, + 0.788, + 0.573, + 0.804 + ], + "angle": 0, + "content": "4.3 Physics-informed Knowledge Transfer" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.81, + 0.787, + 0.842 + ], + "angle": 0, + "content": "Parameters Estimation: In the physics-informed knowledge transfer loop, we first extract the parameters in Eq. (1) and Eq. (3) using the preliminary depth" + } + ], + [ + { + "type": "page_number", + "bbox": [ + 0.218, + 0.116, + 0.23, + 0.127 + ], + "angle": 0, + "content": "8" + }, + { + "type": "header", + "bbox": [ + 0.272, + 0.115, + 0.364, + 0.129 + ], + "angle": 0, + "content": "J. Yang et al." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.147, + 0.784, + 0.223 + ], + "angle": 0, + "content": "estimations from DPT. As shown in Fig. 1, DPT's depth estimations on the underwater images tend to overestimate the depth in distant and blurry regions but offer more trustable estimation for the clearer regions, showing a good linear relationship described in Eq. (3). Therefore, we extract the physics information harnessed from DPT's depth estimations for clear regions." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.223, + 0.788, + 0.435 + ], + "angle": 0, + "content": "The estimation process utilises the DPT depth estimation (on nonlinear post-processed images) and the linear images, consisting of two steps, the rough and refined estimation. The first step is the rough estimation using the whole image to obtain a robust estimation of the background light \\(\\hat{B}_c^\\infty\\). The rationale for extracting the darkest points in the Sea-thru method is analogous to identifying points where \\(J_{c,i}\\) is assumed to be close to zero and where the color appearance on these pixels is solely due to the backscattering. Therefore, we evenly select the \\(N_1\\) darkest points, on which \\(J_{c,i} \\approx 0\\), from the whole image and employ these points to estimate the water parameters by solving the least-square problem in Eq. (5). We denote the pixel value on the pixel location \\(i\\) of the color channel \\(c\\) on the input underwater image as \\(I_{c,i}\\), \\(i = 1 \\dots HW\\). According to the suggestions in UDCP [10], we only use blue and green channels in this process. We denote the selected point set of each color channel as \\(M_c\\), \\(c \\in \\{G, B\\}\\). Substituting (3) into (1), we define the error \\(\\epsilon_c\\) for the color channels as:" + }, + { + "type": "equation", + "bbox": [ + 0.361, + 0.442, + 0.786, + 0.481 + ], + "angle": 0, + "content": "\\[\n\\epsilon_ {c} := \\sum_ {i \\in M _ {c}} | | I _ {c, i} - B _ {c} ^ {\\infty} (1 - e ^ {- \\frac {\\beta_ {c} / s}{\\mathbf {d} _ {i} ^ {D + h / s}}}) | | _ {2} ^ {2}. \\tag {5}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.215, + 0.489, + 0.788, + 0.534 + ], + "angle": 0, + "content": "To simplify the least-square problem, we set two temporary parameters \\(\\nu_{c} = \\beta_{c} / s\\) and \\(\\mu = h / s\\). Then, we solve the following least-square problem to obtain the estimate of the parameters:" + }, + { + "type": "equation", + "bbox": [ + 0.361, + 0.542, + 0.786, + 0.576 + ], + "angle": 0, + "content": "\\[\n\\left(\\hat {\\nu} _ {G}, \\hat {\\nu} _ {B}, \\hat {\\mu}, \\hat {B} _ {c} ^ {\\infty}\\right) = \\operatorname * {a r g m i n} _ {\\nu_ {G}, \\nu_ {B}, \\mu , B _ {c} ^ {\\infty}} \\sum_ {c \\in \\{G, B \\}} \\epsilon_ {c}, \\tag {6}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.215, + 0.584, + 0.788, + 0.738 + ], + "angle": 0, + "content": "Utilizing the preliminary estimates \\(\\hat{\\nu}_G\\), \\(\\hat{\\nu}_B\\), and \\(\\hat{\\mu}\\), we calculate the medium transmission rate \\(t_{c,i}\\) in Eq. (2) for each pixel. Observations from Fig. 2a suggest that depth overestimation of DPT primarily occurs in regions with low transmission rates \\(t_i\\). Therefore, in our implementation, in the refined estimation step, we exclude the regions where DPT's estimation is inaccurate while maximizing the utilization of the largest point sampling regions, in which the transmission rate is smaller than 0.2. Similar to rough estimation, in refined estimation, we select \\(N_2\\) darkest points from the selected regions to construct the set \\(M_{c2}\\) and address a least-square problem. However, we streamline the process by utilizing the estimated \\(\\hat{B}_c^\\infty\\) in the first step. The error \\(\\epsilon_{c2}\\) in this is:" + }, + { + "type": "equation", + "bbox": [ + 0.361, + 0.743, + 0.786, + 0.78 + ], + "angle": 0, + "content": "\\[\n\\epsilon_ {c 2} := \\sum_ {i \\in M _ {c 2}} | | I _ {c, i} - \\hat {B} _ {c} ^ {\\infty} (1 - e ^ {- \\frac {\\nu_ {c}}{\\mathbf {d} _ {i} ^ {D} + \\mu}}) | | _ {2} ^ {2}, \\tag {7}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.216, + 0.787, + 0.452, + 0.802 + ], + "angle": 0, + "content": "and the least-square problem is:" + }, + { + "type": "equation", + "bbox": [ + 0.388, + 0.81, + 0.786, + 0.844 + ], + "angle": 0, + "content": "\\[\n\\left(\\hat {\\nu} _ {G}, \\hat {\\nu} _ {B}, \\hat {\\mu}\\right) = \\operatorname * {a r g m i n} _ {\\nu_ {G}, \\nu_ {B}, \\mu} \\sum_ {c \\in \\{G, B \\}} \\epsilon_ {c 2}, \\tag {8}\n\\]" + } + ], + [ + { + "type": "header", + "bbox": [ + 0.686, + 0.115, + 0.732, + 0.127 + ], + "angle": 0, + "content": "PUDE" + }, + { + "type": "page_number", + "bbox": [ + 0.776, + 0.116, + 0.786, + 0.127 + ], + "angle": 0, + "content": "9" + }, + { + "type": "image", + "bbox": [ + 0.219, + 0.144, + 0.332, + 0.205 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.251, + 0.206, + 0.303, + 0.218 + ], + "angle": 0, + "content": "(a) RGB" + }, + { + "type": "image", + "bbox": [ + 0.335, + 0.144, + 0.447, + 0.204 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.345, + 0.206, + 0.435, + 0.218 + ], + "angle": 0, + "content": "(b) DPT depth" + }, + { + "type": "image", + "bbox": [ + 0.452, + 0.144, + 0.671, + 0.204 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.479, + 0.206, + 0.525, + 0.218 + ], + "angle": 0, + "content": "(c) \\(\\mathcal{L}_{bu}\\)" + }, + { + "type": "image", + "bbox": [ + 0.671, + 0.144, + 0.784, + 0.204 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.678, + 0.206, + 0.776, + 0.218 + ], + "angle": 0, + "content": "(e) local patches" + }, + { + "type": "image_caption", + "bbox": [ + 0.215, + 0.229, + 0.784, + 0.272 + ], + "angle": 0, + "content": "Fig. 4: Loss visualization: Depth estimation from DPT loses details due to blurriness. However, our designed bound losses have differences across pixels according to the pixel intensity thereby enhancing the depth estimation details in training." + }, + { + "type": "text", + "bbox": [ + 0.215, + 0.298, + 0.784, + 0.329 + ], + "angle": 0, + "content": "In our implementation, \\( N_{1} \\) and \\( N_{2} \\) is set to be 500 and 200 respectively. An implementation example of the refined estimation is shown in Fig. 2b." + }, + { + "type": "text", + "bbox": [ + 0.215, + 0.347, + 0.784, + 0.406 + ], + "angle": 0, + "content": "Bound Losses on Medium Transmission Rate: After estimating the parameters of the underwater imaging formation model in Eq. (1), we are ready to present two bound losses to constrain the PUDE model from predicting depths that contravene the underwater characteristics." + }, + { + "type": "text", + "bbox": [ + 0.215, + 0.408, + 0.785, + 0.5 + ], + "angle": 0, + "content": "Lower bound: The first bound is a lower bound on the estimated medium transmission rate \\( t_{c,i}^{P} \\), \\( c \\in \\{G, B\\} \\). This bound aids in refining depth estimates in distant regions, preventing from overestimating the depth for the regions that are far from the camera, as described in O1 in Sec. 3. By substituting the obtained parameter \\( \\hat{\\nu}_{G}, \\hat{\\nu}_{B} \\) and \\( \\hat{\\mu} \\) in the parameters estimation step, and using the PUDE depth estimation, we get the estimated \\( \\hat{t}_{c,i}^{P} \\):" + }, + { + "type": "equation", + "bbox": [ + 0.447, + 0.508, + 0.785, + 0.533 + ], + "angle": 0, + "content": "\\[\n\\hat {t} _ {c, i} ^ {P} = e ^ {- \\frac {\\hat {\\nu} _ {c}}{\\mathbf {d} _ {i} ^ {P} + \\hat {\\mu}}}. \\tag {9}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.215, + 0.539, + 0.784, + 0.569 + ], + "angle": 0, + "content": "As the original color value of each point \\( J_{c,i} \\) is equal to or greater than zero. Based on Eq. (1), we have:" + }, + { + "type": "equation", + "bbox": [ + 0.43, + 0.577, + 0.785, + 0.593 + ], + "angle": 0, + "content": "\\[\nI _ {c, i} \\geq \\left(1 - t _ {c, i}\\right) B _ {c} ^ {\\infty}, \\tag {10}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.215, + 0.602, + 0.62, + 0.617 + ], + "angle": 0, + "content": "substituting the estimated \\(\\hat{B}_c^\\infty\\), the first lower bound is:" + }, + { + "type": "equation", + "bbox": [ + 0.445, + 0.624, + 0.785, + 0.657 + ], + "angle": 0, + "content": "\\[\n\\hat {t} _ {c, i} ^ {P} \\geq 1 - \\frac {I _ {c , i}}{\\hat {B} _ {c} ^ {\\infty}}. \\tag {11}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.215, + 0.665, + 0.784, + 0.741 + ], + "angle": 0, + "content": "This lower bound on \\(\\hat{t}_{c,i}^{P}\\) is to pull back the depth estimations for the distant regions. When the depth of pixels in these regions is overestimated, it results in an underestimation of the medium transmission rate in Eq. (9). Consequently, the captured pixel value \\(I_{c,i}\\) should not be lower than the estimated backscattering, as shown in Eq. (10)." + }, + { + "type": "text", + "bbox": [ + 0.215, + 0.742, + 0.784, + 0.786 + ], + "angle": 0, + "content": "Upper bound: The second bound is an upper bound on the estimated medium transmission rate \\( t_{c,i}^{P}, c \\in \\{G, B\\} \\) to address the challenge of detecting background edges in blurry underwater images, described in O2." + }, + { + "type": "text", + "bbox": [ + 0.215, + 0.787, + 0.784, + 0.817 + ], + "angle": 0, + "content": "Considering that the maximum pixel intensity should not surpass that of pure white, namely \\( J_{c,i} \\leq 1 \\), and referring to Eq. (1), we obtain:" + }, + { + "type": "equation", + "bbox": [ + 0.409, + 0.825, + 0.785, + 0.842 + ], + "angle": 0, + "content": "\\[\nI _ {c, i} \\leq t _ {c, i} + \\left(1 - t _ {c, i}\\right) B _ {c} ^ {\\infty}. \\tag {12}\n\\]" + } + ], + [ + { + "type": "page_number", + "bbox": [ + 0.218, + 0.116, + 0.236, + 0.127 + ], + "angle": 0, + "content": "10" + }, + { + "type": "header", + "bbox": [ + 0.272, + 0.115, + 0.364, + 0.129 + ], + "angle": 0, + "content": "J. Yang et al." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.147, + 0.788, + 0.224 + ], + "angle": 0, + "content": "In regions corresponding to the pure background area, the distance can be considered infinite. The depth estimation by DPT for these regions is always finite. Therefore, the estimated medium transmission for these background regions should be greater than the true value. We denote the estimated medium transmission rate at each pixel location by DPT as \\(\\hat{t}_{c,i}^{D}\\), which is defined as:" + }, + { + "type": "equation", + "bbox": [ + 0.451, + 0.239, + 0.786, + 0.263 + ], + "angle": 0, + "content": "\\[\n\\hat {t} _ {c, i} ^ {D} = e ^ {- \\frac {\\bar {\\nu} _ {c}}{\\mathbf {d} _ {i} ^ {D} + \\bar {\\mu}}}. \\tag {13}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.215, + 0.273, + 0.789, + 0.32 + ], + "angle": 0, + "content": "In distant regions, where the original pixel values \\( J_{c,i} \\) are attenuated, the captured pixel values \\( I_{c,i} \\) are predominantly dictated by backscattering. To select pixel locations likely to be in the background regions, we define the set \\( M_b \\):" + }, + { + "type": "equation", + "bbox": [ + 0.353, + 0.331, + 0.786, + 0.368 + ], + "angle": 0, + "content": "\\[\nM _ {b} = \\{i \\mid \\forall c \\in \\{G, B \\}, \\frac {\\hat {B} _ {c} ^ {\\infty} \\left(1 - \\hat {t} _ {c , i} ^ {D}\\right)}{I _ {c , i}} \\geq \\gamma \\}. \\tag {14}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.215, + 0.378, + 0.789, + 0.44 + ], + "angle": 0, + "content": "The parameter \\(\\gamma\\) is utilized to identify regions where pixel values are predominantly determined by backscattering, which are likely to be in the background area. In our experiments, \\(\\gamma\\) is set to be 0.6. As mentioned, for the background regions, \\(\\hat{t}_{c,i}^{D}\\) is greater than the true value. Then based on the Eq. (12), we have," + }, + { + "type": "equation", + "bbox": [ + 0.378, + 0.453, + 0.786, + 0.471 + ], + "angle": 0, + "content": "\\[\nI _ {c, i} < \\hat {t} _ {c, i} ^ {D} + \\left(1 - t _ {c, i}\\right) B _ {c} ^ {\\infty}, \\quad i \\in M _ {b}, \\tag {15}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.215, + 0.477, + 0.644, + 0.492 + ], + "angle": 0, + "content": "substituting the estimated parameters, we have the bound:" + }, + { + "type": "equation", + "bbox": [ + 0.379, + 0.505, + 0.786, + 0.543 + ], + "angle": 0, + "content": "\\[\n\\hat {t} _ {c, i} ^ {P} < \\frac {- I _ {c , i} + \\hat {t} _ {c , i} ^ {D} + \\hat {B} _ {c} ^ {\\infty}}{\\hat {B} _ {c} ^ {\\infty}}, \\quad i \\in M _ {b}. \\tag {16}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.554, + 0.788, + 0.645 + ], + "angle": 0, + "content": "This upper bound on \\( t_{c,i}^{P} \\) in Eq. (16) is to adjust the depth estimation on the background regions. For the pixels on the background regions, the original DPT produce lower depth estimations. Even if the original pixel \\( J_{c,i} \\) is equal to 1 (perfectly white), the captured pixel value \\( I_{c,i} \\) from the background regions should not be greater than the sum of the estimated medium transmission rate from DPT and the backscattering term, as shown in Eq. (15)." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.646, + 0.788, + 0.738 + ], + "angle": 0, + "content": "Bound losses: We set the bound losses to enforce the estimated transmission rate \\(\\hat{t}_{c,i}^{P}\\) to satisfy the listed two bounds and optimize the estimated depth of PUDE, \\(d_{i}^{P}\\), in training. To simplify the mathematical expressions in the subsequent losses, we denote the ReLU function as \\(f:\\mathbb{R}\\to \\mathbb{R}^+\\), where \\(f(x) = \\max (0,x)\\). The number of elements in \\(M_b\\) denoted as \\(N_{b}\\). According to the bounds in Eq. (11), Eq. (16), the bound losses are:" + }, + { + "type": "equation", + "bbox": [ + 0.327, + 0.751, + 0.786, + 0.794 + ], + "angle": 0, + "content": "\\[\n\\mathcal {L} _ {b l} = \\sum_ {c \\in \\{G, B \\}} \\sum_ {i = 1} ^ {H W} f \\left(- I _ {c, i} + \\left(1 - \\hat {t} _ {c, i} ^ {P}\\right) \\hat {B} _ {c} ^ {\\infty}\\right) / H W, \\tag {17}\n\\]" + }, + { + "type": "equation", + "bbox": [ + 0.308, + 0.809, + 0.786, + 0.843 + ], + "angle": 0, + "content": "\\[\n\\mathcal {L} _ {b u} = \\sum_ {c \\in \\{G, B \\}} \\sum_ {i \\in M _ {b}} f \\left(- \\left(1 - \\hat {t} _ {c, i} ^ {P}\\right) \\hat {B} _ {c} ^ {\\infty} - \\hat {t} _ {c, i} ^ {D} + I _ {c, i}\\right) / N _ {b}. \\tag {18}\n\\]" + } + ], + [ + { + "type": "header", + "bbox": [ + 0.685, + 0.115, + 0.732, + 0.127 + ], + "angle": 0, + "content": "PUDE" + }, + { + "type": "page_number", + "bbox": [ + 0.769, + 0.117, + 0.784, + 0.127 + ], + "angle": 0, + "content": "11" + }, + { + "type": "table_caption", + "bbox": [ + 0.216, + 0.145, + 0.784, + 0.174 + ], + "angle": 0, + "content": "Table 1: Quantitative results: The test datasets, Sea-thru D3, D5, and SQUID, comprise 43, 48, and 57 samples, respectively, collected from open ocean environment." + }, + { + "type": "table", + "bbox": [ + 0.218, + 0.186, + 0.787, + 0.404 + ], + "angle": 0, + "content": "
DatasetsMethodsδ<1.05↑δ<1.052↑δ<1.053↑AbsRel ↓Sq. Rel↓RMSE ↓RMSE log↓
D3MegaDepth [21]0.2750.5000.6590.1710.1121.0190.2250.054
MiDaS [29]0.2280.4290.5700.4080.9531.6350.4250.230
DPT [28]0.4210.6730.7900.1230.0850.8950.1870.038
UWNET [15]0.0970.1920.2900.7742.8292.0270.6500.451
UDepth [37]0.3150.5540.7070.1580.0941.0010.2130.051
Ours (PUDE)0.4430.7090.8260.1010.0590.8630.1610.027
D5MegaDepth [21]0.3520.5800.7220.1300.0431.4670.2020.043
MiDaS [29]0.3400.5730.7140.1310.0431.3050.1970.044
DPT [28]0.3740.6380.7740.1240.0511.4790.2050.045
UWNET [15]0.0750.1580.2400.3950.3342.5760.4090.184
UDepth [37]0.3760.6280.7760.1170.0381.3430.1920.039
Ours (PUDE)0.4350.6780.8030.1050.0341.3430.1860.037
SQUIDMegaDepth [21]0.1850.3520.4890.2530.1742.4190.2860.096
MiDaS [29]0.1960.3790.5070.2960.3472.4440.3090.139
DPT [28]0.3450.5990.7430.1360.0681.6570.1820.040
UWNET [15]0.0820.1670.2550.5100.7413.3630.4780.278
UDepth [37]0.2740.4640.6060.2130.1631.8830.2360.083
Ours (PUDE)0.3870.6320.7720.1190.0431.5240.1630.032
" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.434, + 0.784, + 0.569 + ], + "angle": 0, + "content": "These two bound losses are related to the pixel-wise intensity on color channel. Especially for the dark and distant regions, the difference in the pixel intensity can reflect the depth differences across pixels, the pixel intensity is mostly decided by the depth-related backscattering. The differences in bound losses caused by pixel intensity can encourage the model to learn the depth difference across pixels in small local blurry regions. In other words, the two bound losses guide the model to learn the nonlinear mechanism of backscattering in water, namely \"how blurriness affects the pixel intensity\". Fig. 4 shows an example. Therefore, the overall performance of the model in terms of detail detection gets improved." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.57, + 0.785, + 0.631 + ], + "angle": 0, + "content": "Final loss: We use the parameters \\(\\beta_0, \\beta_1\\) and \\(\\beta_2\\) to adjust the weights between the similarity loss and bound losses, which are set to 1, 10 and 10, respectively. The weights are decided by the metric and visual results on the validation set during training. The total loss \\(\\mathcal{L}_{total}\\) in training is:" + }, + { + "type": "equation", + "bbox": [ + 0.388, + 0.642, + 0.785, + 0.658 + ], + "angle": 0, + "content": "\\[\n\\mathcal {L} _ {\\text {t o t a l}} = \\beta_ {0} \\mathcal {L} _ {s} + \\beta_ {1} \\mathcal {L} _ {b l} + \\beta_ {2} \\mathcal {L} _ {b u}. \\tag {19}\n\\]" + }, + { + "type": "title", + "bbox": [ + 0.216, + 0.708, + 0.375, + 0.725 + ], + "angle": 0, + "content": "5 Experiments" + }, + { + "type": "title", + "bbox": [ + 0.216, + 0.74, + 0.456, + 0.755 + ], + "angle": 0, + "content": "5.1 Implementation Details" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.765, + 0.784, + 0.841 + ], + "angle": 0, + "content": "The training dataset for the PUDE method contains 76 white-balanced linear underwater images collected in an open ocean environment with continuous depth variation, sourced from the SeaThru-NeRF [19] dataset. This scene structure is beneficial for the accurate estimation of the parameters in the underwater image formation model Eq. (1). We apply gamma correction to these images to" + } + ], + [ + { + "type": "page_number", + "bbox": [ + 0.218, + 0.116, + 0.236, + 0.127 + ], + "angle": 0, + "content": "12" + }, + { + "type": "header", + "bbox": [ + 0.272, + 0.115, + 0.364, + 0.129 + ], + "angle": 0, + "content": "J. Yang et al." + }, + { + "type": "image", + "bbox": [ + 0.219, + 0.144, + 0.785, + 0.289 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.214, + 0.298, + 0.788, + 0.355 + ], + "angle": 0, + "content": "Fig. 5: Qualitative comparison on underwater images. Qualitative comparison of PUDE with in-air (MegaDepth, MiDaS, DPT) and underwater methods (UWNET, UDepth) on underwater images. PUDE detects clearer background edges and presents more depth details. (More in Supplementary)" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.384, + 0.788, + 0.489 + ], + "angle": 0, + "content": "generate corresponding nonlinear images for training. We randomly sample 20 images with groundtruth from the Sea-thru D3 datasets to construct the validation dataset. During training, the nonlinear images are downsampled to the dimension of \\(576 \\times 384\\) and fed into the neural network. PUDE is initialized with the DPT-Hybrid [28]. The training runs 3 epochs, employing the Adam optimizer with a learning rate of \\(10^{-5}\\). Notably, an early stopping is applied to prevent overfitting by looking at the validation results." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.49, + 0.788, + 0.657 + ], + "angle": 0, + "content": "In testing, we compare the performance of the PUDE model with the underwater depth estimation methods, UDepth [37], UW-Net [15], and in-air methods with good generalization properties MiDaS V2.1 [29], MegaDepth [21] and DPT-Hybrid [28]. The comparison encompasses both quantitative and qualitative evaluations. The quantitative evaluation uses the remaining images from the Sea-thru D3 and the whole D5 datasets [1]. Another zero-shot testing is performed on the SQUID dataset from [4], including groundtruth measurements obtained through stereo-matching. Given the availability of RAW format images in these datasets, we applied gamma correction and brightness enhancement to the images. In the zero-shot qualitative evaluation, the images are sampled from the real underwater dataset [20]." + }, + { + "type": "title", + "bbox": [ + 0.216, + 0.679, + 0.359, + 0.693 + ], + "angle": 0, + "content": "5.2 Evaluations" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.704, + 0.788, + 0.842 + ], + "angle": 0, + "content": "Quantitative Evaluation: We employ the alignment strategy described in [29] to compare the estimated inverse relative depth with the gorundtruth depth. This alignment operates in the space of inverse ground truth via solving an optimization problem for getting the shift and scale to align the estimated inverse relative depths with the inverse groundtruth. This alignment is subsequently inverted back to the normal depth space. To mitigate the impact of noise and erroneous measurements in groundtruth, we select depth thresholds ranging from \\(0.1m\\) to \\(15m\\) for the Sea-thru dataset [1] and \\(0.1m\\) to \\(20m\\) for the SQUID dataset [4] in our comparisons. We utilize the evaluation metrics described in [11] in our" + } + ], + [ + { + "type": "header", + "bbox": [ + 0.686, + 0.115, + 0.732, + 0.127 + ], + "angle": 0, + "content": "PUDE" + }, + { + "type": "page_number", + "bbox": [ + 0.769, + 0.116, + 0.785, + 0.127 + ], + "angle": 0, + "content": "13" + }, + { + "type": "image", + "bbox": [ + 0.219, + 0.144, + 0.788, + 0.212 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.214, + 0.219, + 0.788, + 0.262 + ], + "angle": 0, + "content": "Fig. 6: Ablation study: (a) Without using \\(\\mathcal{L}_{bu}\\), the model cannot clearly detect the background edges near the rocks and has a fogging appearance. (b) Without using \\(\\mathcal{L}_{bl}\\), the distant rocks cannot be detected, depth of that region is overestimated." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.291, + 0.789, + 0.429 + ], + "angle": 0, + "content": "comparison: Mean absolute value of the relative error (AbsRel), Squared relative error (Sq. Rel), Root mean square error (RMSE), Root mean square logarithmic error (RMSE log), Scale-invariant logarithmic error (SILog) and Accuracy with threshold \\((\\delta < thr)\\). Note that the thr in the comparison metric is set to be 1.05. The results are shown in Tab. 1. The experimental results demonstrate the superior performance of the PUDE model across all evaluated metrics on three datasets compared to other methods. It achieves lower errors and higher estimation accuracy, substantiating the effectiveness of our proposed approach. Particularly, PUDE demonstrates promising performance in zero-shot testing." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.45, + 0.789, + 0.603 + ], + "angle": 0, + "content": "Qualitative Evaluation We also provide qualitative evaluation across underwater images collected from different environments. We sample some underwater images that were collected in a turbid environment from the dataset [20] and do the zero-shot evaluation. The results are shown in Fig. 5. The performance of our PUDE model in underwater depth estimation is superior, particularly in capturing detailed features. It effectively identifies background edges, including the fine contours of distant wires. These outcomes show that PUDE not only inherits robust generalizability from DPT but also incorporates a physics-based understanding of underwater imaging, enabling the PUDE model to discern background regions without requiring labelled training data." + }, + { + "type": "title", + "bbox": [ + 0.216, + 0.623, + 0.403, + 0.637 + ], + "angle": 0, + "content": "5.3 Ablation Studies" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.648, + 0.788, + 0.709 + ], + "angle": 0, + "content": "To investigate the influence of the bound losses in training, we retrain the model with and without bound losses and evaluate them on the Sea-thru D5 dataset. The results, shown in Tab. 2, demonstrate both the losses \\(\\mathcal{L}_{bu}\\) and \\(\\mathcal{L}_{bl}\\) contribute to training, which is consistent with our training goals, solving the issues" + }, + { + "type": "table_caption", + "bbox": [ + 0.214, + 0.734, + 0.788, + 0.763 + ], + "angle": 0, + "content": "Table 2: Ablation study on D5 dataset. All bound losses contribute to the final performance of PUDE. PUDE has the best evaluation results on metrics." + }, + { + "type": "table", + "bbox": [ + 0.274, + 0.775, + 0.731, + 0.837 + ], + "angle": 0, + "content": "
LbuLblδ < 1.05δ < 1.052δ < 1.053AbsRelSq. RelRMSERMSE logSILog
✓0.3740.6380.7740.1240.0511.4790.2050.045
0.4260.6720.7980.1160.0481.4560.1990.042
✓0.3980.6570.7950.1090.0341.3830.1920.039
✓✓0.4350.6780.8030.1050.0341.3430.1860.037
" + } + ], + [ + { + "type": "page_number", + "bbox": [ + 0.218, + 0.116, + 0.236, + 0.127 + ], + "angle": 0, + "content": "14" + }, + { + "type": "header", + "bbox": [ + 0.272, + 0.115, + 0.364, + 0.129 + ], + "angle": 0, + "content": "J. Yang et al." + }, + { + "type": "image", + "bbox": [ + 0.246, + 0.144, + 0.761, + 0.214 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.215, + 0.223, + 0.788, + 0.252 + ], + "angle": 0, + "content": "Fig. 7: Method generalization: PUDE(M) also has improved performance in edge detection and detail detection compared with the in-air model, MiDaS." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.282, + 0.789, + 0.359 + ], + "angle": 0, + "content": "described in O1 and O2. Moreover, to visually demonstrate the effectiveness of bound losses \\(\\mathcal{L}_{bl}\\) and \\(\\mathcal{L}_{bu}\\), we conducted a comparative analysis using images from the Sea-thru D5 dataset, as presented in Fig. 6. Without \\(\\mathcal{L}_{bl}\\), the depth in distant regions is overestimated. By contrast, without \\(\\mathcal{L}_{bu}\\), the detection of the background edges is not clear, resulting in a foggy appearance near the edges." + }, + { + "type": "title", + "bbox": [ + 0.215, + 0.38, + 0.652, + 0.396 + ], + "angle": 0, + "content": "5.4 Method Generalization on Other In-air Models" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.406, + 0.789, + 0.542 + ], + "angle": 0, + "content": "To further explore our method's generalization on different in-air models, we have conducted an additional experiment replacing DPT by MiDaS v2.1 [29] in the training loop, which has the same non-linearity described in Sec. 3 and demonstrate the second-best performance when applied to underwater images directly, and obtained PUDE(M). The quantitative and qualitative results are shown in Fig. 7 and Tab. 3. The results confirm that our method can also enhance the in-water performance of the non-SOTA model, MiDaS. However, as our method relies on knowledge transfer, the efficacy is contingent upon the capability of the original in-air model. A better model yields better transfer results." + }, + { + "type": "title", + "bbox": [ + 0.215, + 0.565, + 0.36, + 0.581 + ], + "angle": 0, + "content": "6 Conclusion" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.597, + 0.789, + 0.718 + ], + "angle": 0, + "content": "This paper proposes a novel method for transferring the in-air depth estimation model, DPT, to underwater settings. We observe DPT has issues with overestimating depth for distant regions and unclear background edge detection on underwater images. To tackle these, we use the novel bound losses to enforce the model to produce depth estimation that follows the underwater imaging formation model. The final quantitative and qualitative evaluation results prove the effectiveness of our method. Additionally, the further experiment shows our method's generalization capabilities on other in-air models." + }, + { + "type": "table_caption", + "bbox": [ + 0.215, + 0.744, + 0.788, + 0.773 + ], + "angle": 0, + "content": "Table 3: PUDE(M) on SQUID dataset: PUDE(M) shows improved performance on the metrics compared to MiDaS, but PUDE has the best performance." + }, + { + "type": "table", + "bbox": [ + 0.267, + 0.786, + 0.737, + 0.838 + ], + "angle": 0, + "content": "
Modelsδ < 1.05δ < 1.052δ < 1.053AbsRelSq. RelRMSERMSE logSILog
MiDaS0.1960.3790.5070.2960.3472.4440.3090.139
PUDE(M)0.2790.4630.6010.1880.1151.8690.2210.067
PUDE0.4350.6780.8030.1050.0341.3430.1860.037
" + } + ], + [ + { + "type": "header", + "bbox": [ + 0.686, + 0.115, + 0.732, + 0.127 + ], + "angle": 0, + "content": "PUDE" + }, + { + "type": "page_number", + "bbox": [ + 0.769, + 0.117, + 0.785, + 0.127 + ], + "angle": 0, + "content": "15" + }, + { + "type": "title", + "bbox": [ + 0.218, + 0.145, + 0.403, + 0.163 + ], + "angle": 0, + "content": "Acknowledgements" + }, + { + "type": "text", + "bbox": [ + 0.217, + 0.179, + 0.784, + 0.209 + ], + "angle": 0, + "content": "This work received funding from the Australian Research Council via grant DE220101527, and the Australian Government, via grant AUSMURIB000001." + }, + { + "type": "title", + "bbox": [ + 0.218, + 0.234, + 0.323, + 0.25 + ], + "angle": 0, + "content": "References" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.267, + 0.785, + 0.309 + ], + "angle": 0, + "content": "1. Akkaynak, D., Treibitz, T.: Sea-thru: A method for removing water from underwater images. In: Proceedings of the IEEE/CVF conference on computer vision and pattern recognition. pp. 1682-1691 (2019)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.311, + 0.785, + 0.351 + ], + "angle": 0, + "content": "2. Amitai, S., Klein, I., Treibitz, T.: Self-supervised monocular depth underwater. In: 2023 IEEE International Conference on Robotics and Automation (ICRA). pp. 1098-1104. IEEE (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.353, + 0.785, + 0.393 + ], + "angle": 0, + "content": "3. Ancuti, C.O., Ancuti, C., De Vleeschouwer, C., Neumann, L., Garcia, R.: Color transfer for underwater dehazing and depth estimation. In: 2017 IEEE International Conference on Image Processing (ICIP). pp. 695-699. IEEE (2017)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.395, + 0.785, + 0.435 + ], + "angle": 0, + "content": "4. Berman, D., Levy, D., Avidan, S., Treibitz, T.: Underwater single image color restoration using haze-lines and a new quantitative dataset. IEEE Transactions on Pattern Analysis and Machine Intelligence (2020)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.437, + 0.785, + 0.477 + ], + "angle": 0, + "content": "5. Bhat, S.F., Alhashim, I., Wonka, P.: Adabins: Depth estimation using adaptive bins. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 4009-4018 (2021)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.479, + 0.785, + 0.505 + ], + "angle": 0, + "content": "6. Bhat, S.F., Birkl, R., Wofk, D., Wonka, P., Müller, M.: Zoedepth: Zero-shot transfer by combining relative and metric depth. arXiv preprint arXiv:2302.12288 (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.507, + 0.785, + 0.547 + ], + "angle": 0, + "content": "7. Chang, H.H., Cheng, C.Y., Sung, C.C.: Single underwater image restoration based on depth estimation and transmission compensation. IEEE Journal of Oceanic Engineering 44(4), 1130-1149 (2018)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.549, + 0.785, + 0.589 + ], + "angle": 0, + "content": "8. Chen, W., Qian, S., Deng, J.: Learning single-image depth from videos using quality assessment networks. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 5604-5613 (2019)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.591, + 0.785, + 0.631 + ], + "angle": 0, + "content": "9. Chiang, J.Y., Chen, Y.C.: Underwater image enhancement by wavelength compensation and dehazing. IEEE transactions on image processing 21(4), 1756-1769 (2011)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.633, + 0.785, + 0.673 + ], + "angle": 0, + "content": "0. Drews, P.L., Nascimento, E.R., Botelho, S.S., Campos, M.F.M.: Underwater depth estimation and image restoration based on single images. IEEE computer graphics and applications 36(2), 24-35 (2016)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.675, + 0.785, + 0.714 + ], + "angle": 0, + "content": "1. Eigen, D., Puhrsch, C., Fergus, R.: Depth map prediction from a single image using a multi-scale deep network. Advances in neural information processing systems 27 (2014)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.716, + 0.785, + 0.757 + ], + "angle": 0, + "content": "2. Fu, H., Gong, M., Wang, C., Batmanghelich, K., Tao, D.: Deep ordinal regression network for monocular depth estimation. In: Proceedings of the IEEE conference on computer vision and pattern recognition. pp. 2002-2011 (2018)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.758, + 0.785, + 0.798 + ], + "angle": 0, + "content": "3. Godard, C., Mac Aodha, O., Brostow, G.J.: Unsupervised monocular depth estimation with left-right consistency. In: Proceedings of the IEEE conference on computer vision and pattern recognition. pp. 270-279 (2017)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.8, + 0.785, + 0.84 + ], + "angle": 0, + "content": "4. Godard, C., Mac Aodha, O., Firman, M., Brostow, G.J.: Digging into self-supervised monocular depth estimation. In: Proceedings of the IEEE/CVF international conference on computer vision. pp. 3828-3838 (2019)" + }, + { + "type": "list", + "bbox": [ + 0.226, + 0.267, + 0.785, + 0.84 + ], + "angle": 0, + "content": null + } + ], + [ + { + "type": "page_number", + "bbox": [ + 0.218, + 0.116, + 0.236, + 0.127 + ], + "angle": 0, + "content": "16" + }, + { + "type": "header", + "bbox": [ + 0.272, + 0.115, + 0.364, + 0.129 + ], + "angle": 0, + "content": "J. Yang et al." + }, + { + "type": "ref_text", + "bbox": [ + 0.218, + 0.147, + 0.785, + 0.189 + ], + "angle": 0, + "content": "15. Gupta, H., Mitra, K.: Unsupervised single image underwater depth estimation. In: 2019 IEEE International Conference on Image Processing (ICIP). pp. 624-628. IEEE (2019)" + }, + { + "type": "ref_text", + "bbox": [ + 0.218, + 0.19, + 0.785, + 0.232 + ], + "angle": 0, + "content": "16. Hambarde, P., Murala, S., Dhall, A.: Uw-gan: Single-image depth estimation and image enhancement for underwater images. IEEE Transactions on Instrumentation and Measurement 70, 1-12 (2021)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.232, + 0.785, + 0.272 + ], + "angle": 0, + "content": "17. He, K., Sun, J., Tang, X.: Single image haze removal using dark channel prior. IEEE transactions on pattern analysis and machine intelligence 33(12), 2341-2353 (2010)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.273, + 0.785, + 0.314 + ], + "angle": 0, + "content": "18. Laina, I., Rupprecht, C., Belagiannis, V., Tombari, F., Navab, N.: Deeper depth prediction with fully convolutional residual networks. In: 2016 Fourth international conference on 3D vision (3DV). pp. 239-248. IEEE (2016)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.315, + 0.785, + 0.37 + ], + "angle": 0, + "content": "19. Levy, D., Peleg, A., Pearl, N., Rosenbaum, D., Akkaynak, D., Korman, S., Treibitz, T.: Seathru-nerf: Neural radiance fields in scattering media. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 56-65 (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.371, + 0.785, + 0.412 + ], + "angle": 0, + "content": "20. Li, C., Guo, C., Ren, W., Cong, R., Hou, J., Kwong, S., Tao, D.: An underwater image enhancement benchmark dataset and beyond. IEEE Transactions on Image Processing 29, 4376-4389 (2019)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.412, + 0.785, + 0.453 + ], + "angle": 0, + "content": "21. Li, Z., Snavely, N.: Megadepth: Learning single-view depth prediction from internet photos. In: Proceedings of the IEEE conference on computer vision and pattern recognition. pp. 2041-2050 (2018)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.453, + 0.785, + 0.495 + ], + "angle": 0, + "content": "22. Luo, Y., Ren, J., Lin, M., Pang, J., Sun, W., Li, H., Lin, L.: Single view stereo matching. In: Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition. pp. 155-163 (2018)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.495, + 0.785, + 0.522 + ], + "angle": 0, + "content": "23. Ming, Y., Meng, X., Fan, C., Yu, H.: Deep learning for monocular depth estimation: A review. Neurocomputing 438, 14-33 (2021)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.522, + 0.785, + 0.563 + ], + "angle": 0, + "content": "24. Peng, Y.T., Cao, K., Cosman, P.C.: Generalization of the dark channel prior for single image restoration. IEEE Transactions on Image Processing 27(6), 2856-2868 (2018)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.564, + 0.785, + 0.605 + ], + "angle": 0, + "content": "25. Peng, Y.T., Cosman, P.C.: Underwater image restoration based on image blurriness and light absorption. IEEE transactions on image processing 26(4), 1579-1594 (2017)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.606, + 0.785, + 0.647 + ], + "angle": 0, + "content": "26. Peng, Y.T., Zhao, X., Cosman, P.C.: Single underwater image enhancement using depth estimation based on blurriness. In: 2015 IEEE International Conference on Image Processing (ICIP). pp. 4952-4956. IEEE (2015)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.647, + 0.785, + 0.688 + ], + "angle": 0, + "content": "27. Ramamonjisoa, M., Lepetit, V.: Sharpnet: Fast and accurate recovery of occluding contours in monocular depth estimation. In: Proceedings of the IEEE/CVF International Conference on Computer Vision Workshops. pp. 0-0 (2019)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.689, + 0.785, + 0.731 + ], + "angle": 0, + "content": "28. Ranftl, R., Bochkovskiy, A., Koltun, V.: Vision transformers for dense prediction. In: Proceedings of the IEEE/CVF international conference on computer vision. pp. 12179-12188 (2021)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.731, + 0.785, + 0.785 + ], + "angle": 0, + "content": "29. Ranftl, R., Lasinger, K., Hafner, D., Schindler, K., Koltun, V.: Towards robust monocular depth estimation: Mixing datasets for zero-shot cross-dataset transfer. IEEE transactions on pattern analysis and machine intelligence 44(3), 1623-1637 (2020)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.786, + 0.785, + 0.84 + ], + "angle": 0, + "content": "30. Varghese, N., Kumar, A., Rajagopalan, A.: Self-supervised monocular underwater depth recovery, image restoration, and a real-sea video dataset. In: Proceedings of the IEEE/CVF International Conference on Computer Vision. pp. 12248-12258 (2023)" + }, + { + "type": "list", + "bbox": [ + 0.217, + 0.147, + 0.785, + 0.84 + ], + "angle": 0, + "content": null + } + ], + [ + { + "type": "header", + "bbox": [ + 0.686, + 0.115, + 0.732, + 0.127 + ], + "angle": 0, + "content": "PUDE" + }, + { + "type": "page_number", + "bbox": [ + 0.769, + 0.116, + 0.786, + 0.127 + ], + "angle": 0, + "content": "17" + }, + { + "type": "ref_text", + "bbox": [ + 0.214, + 0.147, + 0.788, + 0.203 + ], + "angle": 0, + "content": "31. Vasiljevic, I., Kolkin, N., Zhang, S., Luo, R., Wang, H., Dai, F.Z., Daniele, A.F., Mostajabi, M., Basart, S., Walter, M.R., Shakhnarovich, G.: DIODE: A Dense Indoor and Outdoor DEpth Dataset. CoRR abs/1908.00463 (2019), http:// arxiv.org/abs/1908.00463" + }, + { + "type": "ref_text", + "bbox": [ + 0.214, + 0.204, + 0.788, + 0.245 + ], + "angle": 0, + "content": "32. Wu, C.Y., Wang, J., Hall, M., Neumann, U., Su, S.: Toward practical monocular indoor depth estimation. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 3814-3824 (2022)" + }, + { + "type": "ref_text", + "bbox": [ + 0.214, + 0.245, + 0.788, + 0.285 + ], + "angle": 0, + "content": "33. Wu, Y., Zhou, Y., Chen, S., Ma, Y., Li, Q.: Defect inspection for underwater structures based on line-structured light and binocular vision. Applied Optics 60(25), 7754-7764 (2021)" + }, + { + "type": "ref_text", + "bbox": [ + 0.214, + 0.287, + 0.788, + 0.328 + ], + "angle": 0, + "content": "34. Xian, K., Shen, C., Cao, Z., Lu, H., Xiao, Y., Li, R., Luo, Z.: Monocular relative depth perception with web stereo data supervision. In: Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition. pp. 311-320 (2018)" + }, + { + "type": "ref_text", + "bbox": [ + 0.214, + 0.328, + 0.788, + 0.356 + ], + "angle": 0, + "content": "35. Yang, J., Gong, M., Nair, G., Lee, J.H., Monty, J., Pu, Y.: Knowledge distillation for feature extraction in underwater vslam. arXiv preprint arXiv:2303.17981 (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.214, + 0.356, + 0.788, + 0.397 + ], + "angle": 0, + "content": "36. Ye, X., Zhang, J., Yuan, Y., Xu, R., Wang, Z., Li, H.: Underwater depth estimation via stereo adaptation networks. IEEE Transactions on Circuits and Systems for Video Technology (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.214, + 0.397, + 0.788, + 0.439 + ], + "angle": 0, + "content": "37. Yu, B., Wu, J., Islam, M.J.: Udepth: Fast monocular depth estimation for visually-guided underwater robots. In: 2023 IEEE International Conference on Robotics and Automation (ICRA). pp. 3116-3123. IEEE (2023)" + }, + { + "type": "list", + "bbox": [ + 0.214, + 0.147, + 0.788, + 0.439 + ], + "angle": 0, + "content": null + } + ] +] \ No newline at end of file diff --git a/2024/Physics-informed Knowledge Transfer for Underwater Monocular Depth Estimation/58639476-9724-4b09-9db4-e436240a5f18_origin.pdf b/2024/Physics-informed Knowledge Transfer for Underwater Monocular Depth Estimation/58639476-9724-4b09-9db4-e436240a5f18_origin.pdf new file mode 100644 index 0000000000000000000000000000000000000000..024ecdfee6d2ab3b8e19b6b65135e78a6f1a0941 --- /dev/null +++ b/2024/Physics-informed Knowledge Transfer for Underwater Monocular Depth Estimation/58639476-9724-4b09-9db4-e436240a5f18_origin.pdf @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:048d3b1aadc4a01d4f54f6d4d7435485d07c549a526dc2adeb414b68aa8e81ae +size 8907967 diff --git a/2024/Physics-informed Knowledge Transfer for Underwater Monocular Depth Estimation/full.md b/2024/Physics-informed Knowledge Transfer for Underwater Monocular Depth Estimation/full.md new file mode 100644 index 0000000000000000000000000000000000000000..2160896589efb18874728956ef75af965fd54c0e --- /dev/null +++ b/2024/Physics-informed Knowledge Transfer for Underwater Monocular Depth Estimation/full.md @@ -0,0 +1,337 @@ +# Physics-informed Knowledge Transfer for Underwater Monocular Depth Estimation + +Jinghe Yang, Mingming Gong, and Ye Pu + +The University of Melbourne jinghey@student.unimelb.edu.au, {mingming.gong, ye.pu}@unimelb.edu.au + +Abstract. Compared to the in-air case, underwater depth estimation has its own challenges. For instance, acquiring high-quality training datasets with groundtruth poses difficulties due to sensor limitations in aquatic environments. Additionally, the physics characteristics of underwater imaging diverge significantly from the in-air case, the methods developed for in-air depth estimation underperform when applied underwater, due to the domain gap. To address these challenges, our paper introduces a novel transfer-learning-based method - Physics-informed Underwater Depth Estimation (PUDE). The key idea is to transfer the knowledge of a pre-trained in-air depth estimation model to underwater settings utilizing a small underwater image set without groundtruth measurement, guided by a physical underwater imaging formation model. We propose novel bound losses based on the physical model to rectify the depth estimations to align with actual underwater physical properties. Finally, in the evaluations across multiple datasets, we compare PUDE model with other existing in-air and underwater methods. The results reveal that the PUDE model excels in both quantitative and qualitative comparisons. + +Keywords: Underwater Depth $\cdot$ Physics-informed $\cdot$ Knowledge Transfer + +# 1 Introduction + +Depth estimation is a pivotal task in computer vision, with substantial implications for robotics applications. Traditional techniques typically employ stereo cameras to estimate depth through disparity calculations. However, these methods are hampered by complex camera system settings and limited by the challenging feature matching in environments with low texture or visibility [23]. In contrast, monocular depth estimation provides an alternative solution, necessitating only a single camera and a less complicated setup. Although monocular depth estimation faces inherent limitations due to the absence of stereoscopic information, many well-performing in-air learning-based depth estimation models have been developed in recent years [5,6,12,21,27-29]. Benefiting from the massive and various training datasets, some of in-air monocular depth estimation models, such as MiDaS [29] and DPT [28], demonstrate powerful generalization performance. However, the pre-trained in-air models are not fully applicable for underwater cases due to the domain gap and the differences in image formation [36]. For example, we applied the in-air trained models, e.g., DPT in [28], to + +underwater images and observed that DPT has the issues of the overestimation for blurry regions, unclear background region detection and loss of depth details. Additionally, it is challenging to train new models from scratch for underwater settings. Underwater scenarios impose difficulties when capturing a large amount of high-quality images and the groundtruth depth, due to sensor limitations, such as the incapability of LiDAR and the performance drop of stereo matching in water. In contrast, sonar scanning is not affected by water quality and blurriness but suffers from a slow scanning rate and low resolution, lacking detail in scene representation [33]. In response to the distinct challenges of underwater depth estimation, this paper proposes a novel transfer-learning-based approach. Our method transfers knowledge from a pre-trained in-air monocular depth estimation model to underwater settings, guided by the underwater imaging formation model introduced in [9]. Using newly proposed bound losses, which leverage physical information to correct inaccurate depth in distant and background regions, our training method enforces the model to produce estimates aligned with the extracted physical properties. Notably, this adaptation requires minimal training data and no ground truth depth measurements. Our approach benefits from the structural knowledge of the in-air model and integrates underwater physics knowledge, thereby enhancing underwater depth estimation performance. Our contributions are as follows: + +- We present experimental observations and limitations when applying an in-air trained model, i.e., DPT in [28], to underwater images. Inspired by these observations, we propose a novel training method to adapt in-air pre-trained depth estimation models for underwater environments. There are two key ingredients in our method: 1) Knowledge transfer using an in-air trained model with good generalisation properties (DPT [28]) to address data scarcity; 2) Integration of the underwater image formation model proposed in [9] to ensure the depth estimates align with underwater physics characteristics. +- We propose innovative bound losses to steer the model towards producing depth estimations that adhere to the underwater imaging formation model. +- We introduce the Physics-informed Underwater Depth Estimation (PUDE) model, an end-to-end model for underwater monocular depth estimation. In zero-shot evaluations across various datasets, our method demonstrates enhanced performance, both quantitatively and qualitatively, compared to other in-air and underwater methods. + +# 2 RelatedWork + +# 2.1 In-air Monocular Depth Estimation + +Compared to LiDAR, infrared, and stereo vision, monocular depth estimation stands out for its cost-effectiveness and simplicity of configuration, making it appealing for depth perception in autonomous systems. Recent in-air monocular depth estimation methods have exhibited commendable performance, especially supervised methods trained on datasets containing measured groundtruth. They + +usually require sensors like LiDAR, RGBD cameras, and stereo matching [11,12, 18,22]. However, these methods are limited by the lack of variety in the datasets, which restricts their generalization across different real-world scenarios. + +In this context, self-supervised learning using continuous image sequences [14] or calibrated stereo image pairs [13] presents an attractive avenue for exploration but performs less compared to supervised learning [23], and its application to dynamic scenes is challenging. Recognizing this limitation, some studies expand the breadth of training data by harnessing web resources to generate depth groundtruth, thereby enlarging the training dataset and encompassing diverse scenes [8,21,34]. Building further on that, MiDaS [29] extract depth data from 3D movies to enhance dataset diversity and improve generalization. DPT [28] builds upon the MiDaS training dataset, but with supplementary, and implements an encoder-decoder architecture that uses a vision transformer as its backbone, enabling fine-grained and globally coherent estimations and showing strong zero-shot cross-dataset generalization capabilities. + +# 2.2 Underwater Monocular Depth Estimation + +Depth measurement in water introduces new challenges due to the unique characteristics of the underwater environment. The limited availability of sensors, such as LiDAR and infrared, makes the direct acquisition of the depth measurement highly challenging. Meanwhile, the image blurriness highly influences the correspondence matching between images [35], thereby the stereo camera can only provide good depth measurement for close regions. The existing underwater depth datasets that rely on stereo matching [1,4], manifest this issue. + +Therefore, monocular depth estimation is a potential solution to tackle these issues. Physics-based methods play a pivotal role in addressing the underwater depth estimation challenge. The underwater image formation model proposed in [9] has been demonstrated to be effective in solving problems in computer vision for underwater scenarios, such as feature extraction [35], neural radiance fields (NeRFs) [19], and depth estimation [10]. Traditional methods utilize the physics-based image formation model to estimate depth directly from underwater images [3,4,7,24-26]. As the foundation of these methods, Dark Channel Prior (DCP) [17], a statistical method initially developed for in-air haze removal, uses pixel intensity to estimate medium transmission. A variant for the underwater environment, UDCP [10], claims that the DCP assumption remains applicable to the green and blue channels of underwater images. However, underwater depth estimation that relies solely on pixel values and the imaging formation model is vulnerable to variable lighting conditions, such as shadows and sunlight flickers. + +Despite the challenges in depth groundtruth acquisition underwater, the learning-based methods show promise. Some works have attempted unsupervised learning, such as UwNET [15] and UW-GAN [16]. They use the GAN method through color restoration to estimate depth. The recent works [2,30] adopt self-supervised learning, utilizing the correspondences between subsequent frames. Notably, they incorporate the physics model into the training of the underwater depth estimation models and have shown to be advantageous. + +# 2.3 Underwater Imaging Formation Model + +Light transmission in water is subject to absorption, and minuscule particles cause the backscattering, leading to blur within underwater imagery. A widely used model of underwater image formation is proposed in [9] to describe this physical phenomenon. Consider an underwater image represented by $I_{c} \in \mathbb{R}^{HW}$ , where $c$ denotes the color channels $\{\mathrm{R}, \mathrm{G}, \mathrm{B}\}$ , and $H$ and $W$ are the image's height and width, respectively. Given an image $I_{c}$ , for each pixel $i$ , $i = 1 \cdots HW$ , it holds that $I_{c,i} \in [0,1]$ . The underwater imaging formation model is: + +$$ +I _ {c, i} = J _ {c, i} \cdot t _ {c, i} + B _ {c} ^ {\infty} (1 - t _ {c, i}), \tag {1} +$$ + +$$ +t _ {c, i} = e ^ {- \beta_ {c} \mathbf {z} _ {i}}. (2) +$$ + +In this model, $J_{c,i}$ is the original in-air color value at pixel $i$ . The original in-air image is denoted by $J_{c} = [J_{c,1},\dots,J_{c,HW}] \in \mathbb{R}^{HW}$ . The medium transmission rate denoted by $t_{c,i}$ is associated with the beam attenuation coefficient $\beta_{c}$ and the depth $\mathbf{z}_{i}$ . Meanwhile, $B_{c}^{\infty}$ is the coefficient of ambient light causing the backscattering from particles. The captured underwater image $I_{c,i}$ is the aggregate of the attenuation of the original image and the additional backscattering from the water environment. The values of $\beta_{c}$ and $B_{c}^{\infty}$ vary depending on different water environments and can be estimated using the depth measurements [1]. + +# 3 DPT on Underwater Images: Observations + +Existing learning-based monocular depth estimation methods, such as DPT [28] and MiDaS [29], generate the inverse relative depth estimation from a single image. We denote the inverse depth estimation as $\mathbf{d} \in \mathbb{R}^{HW}$ , with $\mathbf{d}_i$ to be the estimate for pixel $i$ , $i = 1, \dots, HW$ . The metric depth of the input image is denoted by $\mathbf{z} \in \mathbb{R}^{HW}$ , whose reciprocal ideally has a linear relationship with the inverse relative depth. The relationship between inverse relative depth $\mathbf{d}_i$ and the metric depth $\mathbf{z}_i$ at pixel $i$ is: + +$$ +\mathbf {z} _ {i} = \frac {1}{s \mathbf {d} _ {i} + h}, \quad \forall i \in \{1, \dots , H W \}. \tag {3} +$$ + +Here, $s$ and $h$ are two unknown scale and shift coefficients. In practice, they can be off-line estimated for specific scenes using sequence images. The overall mapping relationship between the inverse relative depth $\mathbf{d}_i$ and the inverse of metric depth $\mathbf{z}_i$ is linear on the in-air images [32]. An example can be seen in the first row of Fig. 1, which is the depth estimation of DPT on an outdoor in-air image from DIODE [31], we can observe that DPT estimation on the in-air image maintains a good linear relationship (the correlation coefficient $r = 0.992$ ). However, when we apply DPT directly to underwater images from the dataset [1,4], in which the depth groundtruth is obtained through the Structure-From-Motion and stereo matching, it is observed that in underwater imagery, the estimated + +![](images/8824e2ba1a6351826f34979e8a17fac4e2e9017600967b24d71b5ee8309f5e69.jpg) + +![](images/bfbd28f515df07458287e11034a532f3cf58557a3da4626d1d260b559bdd9ada.jpg) + +![](images/f2910ee34d2f4aebf56aaede5910cfa5207fa1a7b2d273588bb9b6b5d425e791.jpg) + +![](images/95652f5ed62bf7ea885a6a09c5f4dc3966cf5b6d9b54df8ad997435039bd22e8.jpg) +Fig. 1: DPT estimation of underwater images shows a curved tendency, with overestimation in distant regions. Additionally, the estimated depth struggles with unclear background detection (blue arrows). + +![](images/91ffd5b80d32babd4b2dffd30e7ac4b2d4dd74ee97c866842af8943309f1d6c8.jpg) + +![](images/e4f3271b4175f228a5ca1c95387f5af299ca6a8c15de564c09aad55620ece144.jpg) + +inverse depth, denoted by $\mathbf{d}_i^D$ , and $\frac{1}{z_i}$ deviates from a linear relationship and present a curve tendency with a weaker correlation, especially when $\mathbf{d}_i^D$ is small. An example is shown in the second row of Fig. 1. + +To observe the nonlinear trend caused by image blurriness, we applied the Sea-thru method [1] to calculate the medium transmission rate $t_i$ in Eq. (2) on the grayscale images. We then compared these rates with those obtained from the DPT depth (aligned with groundtruth depth). An example of this procedure on the Sea-thru D3 dataset is shown in Fig. 2a, where we randomly sample 100 points from each image. We observe that the estimated transmission rate from DPT is mostly lower than that calculated by groundtruth depth when the transmission rate is small ( $t_i < 0.2$ ). The results reveal that DPT underestimates the transmission rate in regions with low transmission rates, leading to an overestimation of depth in distant regions. Furthermore, the background edges on the underwater depth image of DPT, such as the blue arrow pointed regions around corals, are less clear, exhibiting fogging compared with the estimation on the in-air image. Based on these experimental observations, we summarise the primary issues of the depth estimation from DPT in water: + +O1 : DPT exhibits a propensity for overestimating depth in regions that are relatively far away from the camera. This overestimation is particularly evident in regions with diminished visual clarity and increased blurriness. +O2 : Secondly, DPT encounters challenges in accurately discerning background regions in underwater images. This issue is especially pronounced for the region far from the camera. Due to the attenuation of the scenes, it is challenging for DPT to define the background area clearly. + +We propose two bound losses (Sec. 4.3) in our training to address these issues. + +![](images/b2317f68ed0d21e33d87d820aff680737832f55f90f08e3ad833b9797c4c8d4f.jpg) +(a) $t_i$ in Eq. (2) (D3 dataset) + +![](images/535ee7551ae3b70ee6299384f427519641887e50abc27238662b9da9c87baf56.jpg) +(b) Point selection and model fitting in Eq. (1), Eq. (3) +Fig.2: (a) The comparison between the calculated transmission rate $t_i$ from groundtruth and the estimated $t_i$ from DPT depth, presents an underestimation of the $t_i$ for small values of DPT. (b) The left image shows the points selected in the refined estimation. The right shows the fitting of the estimation. + +![](images/281a5e27ef6c7031f29527ea0313d3ce9de8ecf271bb56aaa1fcbf974be952bf.jpg) + +# 4 Method + +# 4.1 Framework Overview + +Inspired by the observations in Sec. 3, we propose a physics-informed knowledge transfer method for underwater monocular depth estimation, named the Physics-informed Underwater Depth Estimation (PUDE). The main idea of this physics-informed method is to leverage the physical information in underwater imagery to guide knowledge transfer within the neural network, thereby enhancing its adaptation from in-air to underwater environments. Specifically, we adopt the pre-trained DPT-Hybrid model [28], a state-of-the-art (SOTA) model known for its good generalization, as the transferred model. Benefiting from the massive in-air training dataset, the DPT model has powerful spatial structural knowledge. Despite previously noted issues when applied to underwater images, it can still provide a relatively reasonable depth estimation, particularly for regions close to the camera. This estimation serves as our prior depth knowledge for extracting the physical information from the underwater images. + +Most daily images are non-linear images, which are automatically post-processed by nonlinear algorithms integrated into cameras, for example, gamma correction, to approach the non-linear perception of human eyes. By contrast, linear images are images in which pixel values are directly captured by the camera without nonlinear post-processing or only with ratio processing. To ensure the PUDE model's generalizability in reality and keep consistency in training and testing, we utilize post-processed non-linear underwater images as inputs. In contrast, for precise physical information extraction, we utilize the linear underwater images in the estimation of the physics parameters in Eq. (1) and Eq. (3). + +Our training framework consists of two branches shown in Fig. 3. The first branch is the direct knowledge transfer from DPT to PUDE, and we employ a similarity loss to guarantee the similarity between DPT and PUDE in the transfer (in Sec. 4.2). The second branch is the physics-informed knowledge transfer. Inspired by the parameter estimation method in Sea-thru [1], we estimate the parameters in Eq. (1) and Eq. (3) for each input underwater image, utilizing + +![](images/91205bbd54b8d2398f71974089efaf96b0e84d1b00c1a1d0c8f78f9fc6ea4de1.jpg) +Fig. 3: Overall training framework for PUDE. The overall training framework comprises two components: the direct knowledge transfer from the DPT model and the physi-informed knowledge transfer based on the DPT estimation. Physical information is extracted from DPT's output to establish the bound losses. + +![](images/f3287d058a9034e99873dc4a7b102847501a7b47962655d1620a4bcf38b317ed.jpg) + +DPT depth estimation. We restrict our usage to the green and blue channels, in line with the conclusions from UDCP [10]. Based on that, we propose two novel bound losses, utilizing the physics information contained in the underwater image formation model in Eq. (1) to enforce the model to produce plausible depth estimations that conform to the physics characteristics of underwater imaging. In Sec. 4.3, we will introduce the second branch and the designed bound losses. + +# 4.2 Direct Knowledge Transfer + +The direct knowledge transfer loop aims to guarantee the similarity between the original DPT and PUDE during the transition to ensure the structural knowledge can be kept in PUDE. We use a similarity loss in the training process to achieve this goal. In the transfer, PUDE is initialized with DPT. We trim the pixels with the largest differences to set the tolerance for the difference between PUDE and the original model DPT during the training. We denote the pixel set after trimming as $M_{trim} \subseteq \{1, \dots, HW\}$ , and $T$ representing the cardinality of $M_{trim}$ . For each pixel $i \in M_{trim}$ , we denote the DPT and PUDE estimation as $\mathbf{d}_i^D$ and $\mathbf{d}_i^P$ , respectively. The similarity loss $L_s$ is defined as: + +$$ +L _ {s} = \frac {1}{T} \sum_ {i \in M _ {\mathrm {t r i m}}} \left| \frac {\mathbf {d} _ {i} ^ {P} - \mathbf {d} _ {i} ^ {D}}{\operatorname {m e d i a n} (\mathbf {d} ^ {D})} \right|. \tag {4} +$$ + +The trimming sets the tolerance for the changing between PUDE and the original model DPT during the training. In our implementation, we set the trimming rate to be $0.3HW$ . + +# 4.3 Physics-informed Knowledge Transfer + +Parameters Estimation: In the physics-informed knowledge transfer loop, we first extract the parameters in Eq. (1) and Eq. (3) using the preliminary depth + +estimations from DPT. As shown in Fig. 1, DPT's depth estimations on the underwater images tend to overestimate the depth in distant and blurry regions but offer more trustable estimation for the clearer regions, showing a good linear relationship described in Eq. (3). Therefore, we extract the physics information harnessed from DPT's depth estimations for clear regions. + +The estimation process utilises the DPT depth estimation (on nonlinear post-processed images) and the linear images, consisting of two steps, the rough and refined estimation. The first step is the rough estimation using the whole image to obtain a robust estimation of the background light $\hat{B}_c^\infty$ . The rationale for extracting the darkest points in the Sea-thru method is analogous to identifying points where $J_{c,i}$ is assumed to be close to zero and where the color appearance on these pixels is solely due to the backscattering. Therefore, we evenly select the $N_1$ darkest points, on which $J_{c,i} \approx 0$ , from the whole image and employ these points to estimate the water parameters by solving the least-square problem in Eq. (5). We denote the pixel value on the pixel location $i$ of the color channel $c$ on the input underwater image as $I_{c,i}$ , $i = 1 \dots HW$ . According to the suggestions in UDCP [10], we only use blue and green channels in this process. We denote the selected point set of each color channel as $M_c$ , $c \in \{G, B\}$ . Substituting (3) into (1), we define the error $\epsilon_c$ for the color channels as: + +$$ +\epsilon_ {c} := \sum_ {i \in M _ {c}} | | I _ {c, i} - B _ {c} ^ {\infty} (1 - e ^ {- \frac {\beta_ {c} / s}{\mathbf {d} _ {i} ^ {D + h / s}}}) | | _ {2} ^ {2}. \tag {5} +$$ + +To simplify the least-square problem, we set two temporary parameters $\nu_{c} = \beta_{c} / s$ and $\mu = h / s$ . Then, we solve the following least-square problem to obtain the estimate of the parameters: + +$$ +\left(\hat {\nu} _ {G}, \hat {\nu} _ {B}, \hat {\mu}, \hat {B} _ {c} ^ {\infty}\right) = \operatorname * {a r g m i n} _ {\nu_ {G}, \nu_ {B}, \mu , B _ {c} ^ {\infty}} \sum_ {c \in \{G, B \}} \epsilon_ {c}, \tag {6} +$$ + +Utilizing the preliminary estimates $\hat{\nu}_G$ , $\hat{\nu}_B$ , and $\hat{\mu}$ , we calculate the medium transmission rate $t_{c,i}$ in Eq. (2) for each pixel. Observations from Fig. 2a suggest that depth overestimation of DPT primarily occurs in regions with low transmission rates $t_i$ . Therefore, in our implementation, in the refined estimation step, we exclude the regions where DPT's estimation is inaccurate while maximizing the utilization of the largest point sampling regions, in which the transmission rate is smaller than 0.2. Similar to rough estimation, in refined estimation, we select $N_2$ darkest points from the selected regions to construct the set $M_{c2}$ and address a least-square problem. However, we streamline the process by utilizing the estimated $\hat{B}_c^\infty$ in the first step. The error $\epsilon_{c2}$ in this is: + +$$ +\epsilon_ {c 2} := \sum_ {i \in M _ {c 2}} | | I _ {c, i} - \hat {B} _ {c} ^ {\infty} (1 - e ^ {- \frac {\nu_ {c}}{\mathbf {d} _ {i} ^ {D} + \mu}}) | | _ {2} ^ {2}, \tag {7} +$$ + +and the least-square problem is: + +$$ +\left(\hat {\nu} _ {G}, \hat {\nu} _ {B}, \hat {\mu}\right) = \operatorname * {a r g m i n} _ {\nu_ {G}, \nu_ {B}, \mu} \sum_ {c \in \{G, B \}} \epsilon_ {c 2}, \tag {8} +$$ + +![](images/ccc1a4dce5fdff632fe807cc9289c1ca417ab43da08a1397123731bf460308d7.jpg) +(a) RGB + +![](images/a40e9c9e94a086eca498a5972348c4096cd5cfb4a788dfeaaa12c7af29242741.jpg) +(b) DPT depth +Fig. 4: Loss visualization: Depth estimation from DPT loses details due to blurriness. However, our designed bound losses have differences across pixels according to the pixel intensity thereby enhancing the depth estimation details in training. + +![](images/a6f11881cbb6b6c58935904100e80e8fe25e1cad02ffb487dfa204a5bf6bbdd2.jpg) +(c) $\mathcal{L}_{bu}$ + +![](images/81888be583b5065b0ca2874160e14a97bdb7f8193d74836d9d9564dc07707b95.jpg) +(e) local patches + +In our implementation, $N_{1}$ and $N_{2}$ is set to be 500 and 200 respectively. An implementation example of the refined estimation is shown in Fig. 2b. + +Bound Losses on Medium Transmission Rate: After estimating the parameters of the underwater imaging formation model in Eq. (1), we are ready to present two bound losses to constrain the PUDE model from predicting depths that contravene the underwater characteristics. + +Lower bound: The first bound is a lower bound on the estimated medium transmission rate $t_{c,i}^{P}$ , $c \in \{G, B\}$ . This bound aids in refining depth estimates in distant regions, preventing from overestimating the depth for the regions that are far from the camera, as described in O1 in Sec. 3. By substituting the obtained parameter $\hat{\nu}_{G}, \hat{\nu}_{B}$ and $\hat{\mu}$ in the parameters estimation step, and using the PUDE depth estimation, we get the estimated $\hat{t}_{c,i}^{P}$ : + +$$ +\hat {t} _ {c, i} ^ {P} = e ^ {- \frac {\hat {\nu} _ {c}}{\mathbf {d} _ {i} ^ {P} + \hat {\mu}}}. \tag {9} +$$ + +As the original color value of each point $J_{c,i}$ is equal to or greater than zero. Based on Eq. (1), we have: + +$$ +I _ {c, i} \geq \left(1 - t _ {c, i}\right) B _ {c} ^ {\infty}, \tag {10} +$$ + +substituting the estimated $\hat{B}_c^\infty$ , the first lower bound is: + +$$ +\hat {t} _ {c, i} ^ {P} \geq 1 - \frac {I _ {c , i}}{\hat {B} _ {c} ^ {\infty}}. \tag {11} +$$ + +This lower bound on $\hat{t}_{c,i}^{P}$ is to pull back the depth estimations for the distant regions. When the depth of pixels in these regions is overestimated, it results in an underestimation of the medium transmission rate in Eq. (9). Consequently, the captured pixel value $I_{c,i}$ should not be lower than the estimated backscattering, as shown in Eq. (10). + +Upper bound: The second bound is an upper bound on the estimated medium transmission rate $t_{c,i}^{P}, c \in \{G, B\}$ to address the challenge of detecting background edges in blurry underwater images, described in O2. + +Considering that the maximum pixel intensity should not surpass that of pure white, namely $J_{c,i} \leq 1$ , and referring to Eq. (1), we obtain: + +$$ +I _ {c, i} \leq t _ {c, i} + \left(1 - t _ {c, i}\right) B _ {c} ^ {\infty}. \tag {12} +$$ + +In regions corresponding to the pure background area, the distance can be considered infinite. The depth estimation by DPT for these regions is always finite. Therefore, the estimated medium transmission for these background regions should be greater than the true value. We denote the estimated medium transmission rate at each pixel location by DPT as $\hat{t}_{c,i}^{D}$ , which is defined as: + +$$ +\hat {t} _ {c, i} ^ {D} = e ^ {- \frac {\bar {\nu} _ {c}}{\mathbf {d} _ {i} ^ {D} + \bar {\mu}}}. \tag {13} +$$ + +In distant regions, where the original pixel values $J_{c,i}$ are attenuated, the captured pixel values $I_{c,i}$ are predominantly dictated by backscattering. To select pixel locations likely to be in the background regions, we define the set $M_b$ : + +$$ +M _ {b} = \{i \mid \forall c \in \{G, B \}, \frac {\hat {B} _ {c} ^ {\infty} \left(1 - \hat {t} _ {c , i} ^ {D}\right)}{I _ {c , i}} \geq \gamma \}. \tag {14} +$$ + +The parameter $\gamma$ is utilized to identify regions where pixel values are predominantly determined by backscattering, which are likely to be in the background area. In our experiments, $\gamma$ is set to be 0.6. As mentioned, for the background regions, $\hat{t}_{c,i}^{D}$ is greater than the true value. Then based on the Eq. (12), we have, + +$$ +I _ {c, i} < \hat {t} _ {c, i} ^ {D} + \left(1 - t _ {c, i}\right) B _ {c} ^ {\infty}, \quad i \in M _ {b}, \tag {15} +$$ + +substituting the estimated parameters, we have the bound: + +$$ +\hat {t} _ {c, i} ^ {P} < \frac {- I _ {c , i} + \hat {t} _ {c , i} ^ {D} + \hat {B} _ {c} ^ {\infty}}{\hat {B} _ {c} ^ {\infty}}, \quad i \in M _ {b}. \tag {16} +$$ + +This upper bound on $t_{c,i}^{P}$ in Eq. (16) is to adjust the depth estimation on the background regions. For the pixels on the background regions, the original DPT produce lower depth estimations. Even if the original pixel $J_{c,i}$ is equal to 1 (perfectly white), the captured pixel value $I_{c,i}$ from the background regions should not be greater than the sum of the estimated medium transmission rate from DPT and the backscattering term, as shown in Eq. (15). + +Bound losses: We set the bound losses to enforce the estimated transmission rate $\hat{t}_{c,i}^{P}$ to satisfy the listed two bounds and optimize the estimated depth of PUDE, $d_{i}^{P}$ , in training. To simplify the mathematical expressions in the subsequent losses, we denote the ReLU function as $f:\mathbb{R}\to \mathbb{R}^+$ , where $f(x) = \max (0,x)$ . The number of elements in $M_b$ denoted as $N_{b}$ . According to the bounds in Eq. (11), Eq. (16), the bound losses are: + +$$ +\mathcal {L} _ {b l} = \sum_ {c \in \{G, B \}} \sum_ {i = 1} ^ {H W} f \left(- I _ {c, i} + \left(1 - \hat {t} _ {c, i} ^ {P}\right) \hat {B} _ {c} ^ {\infty}\right) / H W, \tag {17} +$$ + +$$ +\mathcal {L} _ {b u} = \sum_ {c \in \{G, B \}} \sum_ {i \in M _ {b}} f \left(- \left(1 - \hat {t} _ {c, i} ^ {P}\right) \hat {B} _ {c} ^ {\infty} - \hat {t} _ {c, i} ^ {D} + I _ {c, i}\right) / N _ {b}. \tag {18} +$$ + +Table 1: Quantitative results: The test datasets, Sea-thru D3, D5, and SQUID, comprise 43, 48, and 57 samples, respectively, collected from open ocean environment. + +
DatasetsMethodsδ<1.05↑δ<1.052↑δ<1.053↑AbsRel ↓Sq. Rel↓RMSE ↓RMSE log↓
D3MegaDepth [21]0.2750.5000.6590.1710.1121.0190.2250.054
MiDaS [29]0.2280.4290.5700.4080.9531.6350.4250.230
DPT [28]0.4210.6730.7900.1230.0850.8950.1870.038
UWNET [15]0.0970.1920.2900.7742.8292.0270.6500.451
UDepth [37]0.3150.5540.7070.1580.0941.0010.2130.051
Ours (PUDE)0.4430.7090.8260.1010.0590.8630.1610.027
D5MegaDepth [21]0.3520.5800.7220.1300.0431.4670.2020.043
MiDaS [29]0.3400.5730.7140.1310.0431.3050.1970.044
DPT [28]0.3740.6380.7740.1240.0511.4790.2050.045
UWNET [15]0.0750.1580.2400.3950.3342.5760.4090.184
UDepth [37]0.3760.6280.7760.1170.0381.3430.1920.039
Ours (PUDE)0.4350.6780.8030.1050.0341.3430.1860.037
SQUIDMegaDepth [21]0.1850.3520.4890.2530.1742.4190.2860.096
MiDaS [29]0.1960.3790.5070.2960.3472.4440.3090.139
DPT [28]0.3450.5990.7430.1360.0681.6570.1820.040
UWNET [15]0.0820.1670.2550.5100.7413.3630.4780.278
UDepth [37]0.2740.4640.6060.2130.1631.8830.2360.083
Ours (PUDE)0.3870.6320.7720.1190.0431.5240.1630.032
+ +These two bound losses are related to the pixel-wise intensity on color channel. Especially for the dark and distant regions, the difference in the pixel intensity can reflect the depth differences across pixels, the pixel intensity is mostly decided by the depth-related backscattering. The differences in bound losses caused by pixel intensity can encourage the model to learn the depth difference across pixels in small local blurry regions. In other words, the two bound losses guide the model to learn the nonlinear mechanism of backscattering in water, namely "how blurriness affects the pixel intensity". Fig. 4 shows an example. Therefore, the overall performance of the model in terms of detail detection gets improved. + +Final loss: We use the parameters $\beta_0, \beta_1$ and $\beta_2$ to adjust the weights between the similarity loss and bound losses, which are set to 1, 10 and 10, respectively. The weights are decided by the metric and visual results on the validation set during training. The total loss $\mathcal{L}_{total}$ in training is: + +$$ +\mathcal {L} _ {\text {t o t a l}} = \beta_ {0} \mathcal {L} _ {s} + \beta_ {1} \mathcal {L} _ {b l} + \beta_ {2} \mathcal {L} _ {b u}. \tag {19} +$$ + +# 5 Experiments + +# 5.1 Implementation Details + +The training dataset for the PUDE method contains 76 white-balanced linear underwater images collected in an open ocean environment with continuous depth variation, sourced from the SeaThru-NeRF [19] dataset. This scene structure is beneficial for the accurate estimation of the parameters in the underwater image formation model Eq. (1). We apply gamma correction to these images to + +![](images/c9b32c4b244a9104283fb6966adb740200bd240814590b41a5f0ca8065297a7b.jpg) +Fig. 5: Qualitative comparison on underwater images. Qualitative comparison of PUDE with in-air (MegaDepth, MiDaS, DPT) and underwater methods (UWNET, UDepth) on underwater images. PUDE detects clearer background edges and presents more depth details. (More in Supplementary) + +generate corresponding nonlinear images for training. We randomly sample 20 images with groundtruth from the Sea-thru D3 datasets to construct the validation dataset. During training, the nonlinear images are downsampled to the dimension of $576 \times 384$ and fed into the neural network. PUDE is initialized with the DPT-Hybrid [28]. The training runs 3 epochs, employing the Adam optimizer with a learning rate of $10^{-5}$ . Notably, an early stopping is applied to prevent overfitting by looking at the validation results. + +In testing, we compare the performance of the PUDE model with the underwater depth estimation methods, UDepth [37], UW-Net [15], and in-air methods with good generalization properties MiDaS V2.1 [29], MegaDepth [21] and DPT-Hybrid [28]. The comparison encompasses both quantitative and qualitative evaluations. The quantitative evaluation uses the remaining images from the Sea-thru D3 and the whole D5 datasets [1]. Another zero-shot testing is performed on the SQUID dataset from [4], including groundtruth measurements obtained through stereo-matching. Given the availability of RAW format images in these datasets, we applied gamma correction and brightness enhancement to the images. In the zero-shot qualitative evaluation, the images are sampled from the real underwater dataset [20]. + +# 5.2 Evaluations + +Quantitative Evaluation: We employ the alignment strategy described in [29] to compare the estimated inverse relative depth with the gorundtruth depth. This alignment operates in the space of inverse ground truth via solving an optimization problem for getting the shift and scale to align the estimated inverse relative depths with the inverse groundtruth. This alignment is subsequently inverted back to the normal depth space. To mitigate the impact of noise and erroneous measurements in groundtruth, we select depth thresholds ranging from $0.1m$ to $15m$ for the Sea-thru dataset [1] and $0.1m$ to $20m$ for the SQUID dataset [4] in our comparisons. We utilize the evaluation metrics described in [11] in our + +![](images/132809b95d12103d908ebd6410aaab47c97a2ff13f9af6164ae736a136607827.jpg) +Fig. 6: Ablation study: (a) Without using $\mathcal{L}_{bu}$ , the model cannot clearly detect the background edges near the rocks and has a fogging appearance. (b) Without using $\mathcal{L}_{bl}$ , the distant rocks cannot be detected, depth of that region is overestimated. + +comparison: Mean absolute value of the relative error (AbsRel), Squared relative error (Sq. Rel), Root mean square error (RMSE), Root mean square logarithmic error (RMSE log), Scale-invariant logarithmic error (SILog) and Accuracy with threshold $(\delta < thr)$ . Note that the thr in the comparison metric is set to be 1.05. The results are shown in Tab. 1. The experimental results demonstrate the superior performance of the PUDE model across all evaluated metrics on three datasets compared to other methods. It achieves lower errors and higher estimation accuracy, substantiating the effectiveness of our proposed approach. Particularly, PUDE demonstrates promising performance in zero-shot testing. + +Qualitative Evaluation We also provide qualitative evaluation across underwater images collected from different environments. We sample some underwater images that were collected in a turbid environment from the dataset [20] and do the zero-shot evaluation. The results are shown in Fig. 5. The performance of our PUDE model in underwater depth estimation is superior, particularly in capturing detailed features. It effectively identifies background edges, including the fine contours of distant wires. These outcomes show that PUDE not only inherits robust generalizability from DPT but also incorporates a physics-based understanding of underwater imaging, enabling the PUDE model to discern background regions without requiring labelled training data. + +# 5.3 Ablation Studies + +To investigate the influence of the bound losses in training, we retrain the model with and without bound losses and evaluate them on the Sea-thru D5 dataset. The results, shown in Tab. 2, demonstrate both the losses $\mathcal{L}_{bu}$ and $\mathcal{L}_{bl}$ contribute to training, which is consistent with our training goals, solving the issues + +Table 2: Ablation study on D5 dataset. All bound losses contribute to the final performance of PUDE. PUDE has the best evaluation results on metrics. + +
LbuLblδ < 1.05δ < 1.052δ < 1.053AbsRelSq. RelRMSERMSE logSILog
✓0.3740.6380.7740.1240.0511.4790.2050.045
0.4260.6720.7980.1160.0481.4560.1990.042
✓0.3980.6570.7950.1090.0341.3830.1920.039
✓✓0.4350.6780.8030.1050.0341.3430.1860.037
+ +![](images/c297734858541ffa16c4575fe0c884efc816fef65677e0c716be7662f0b9234b.jpg) +Fig. 7: Method generalization: PUDE(M) also has improved performance in edge detection and detail detection compared with the in-air model, MiDaS. + +described in O1 and O2. Moreover, to visually demonstrate the effectiveness of bound losses $\mathcal{L}_{bl}$ and $\mathcal{L}_{bu}$ , we conducted a comparative analysis using images from the Sea-thru D5 dataset, as presented in Fig. 6. Without $\mathcal{L}_{bl}$ , the depth in distant regions is overestimated. By contrast, without $\mathcal{L}_{bu}$ , the detection of the background edges is not clear, resulting in a foggy appearance near the edges. + +# 5.4 Method Generalization on Other In-air Models + +To further explore our method's generalization on different in-air models, we have conducted an additional experiment replacing DPT by MiDaS v2.1 [29] in the training loop, which has the same non-linearity described in Sec. 3 and demonstrate the second-best performance when applied to underwater images directly, and obtained PUDE(M). The quantitative and qualitative results are shown in Fig. 7 and Tab. 3. The results confirm that our method can also enhance the in-water performance of the non-SOTA model, MiDaS. However, as our method relies on knowledge transfer, the efficacy is contingent upon the capability of the original in-air model. A better model yields better transfer results. + +# 6 Conclusion + +This paper proposes a novel method for transferring the in-air depth estimation model, DPT, to underwater settings. We observe DPT has issues with overestimating depth for distant regions and unclear background edge detection on underwater images. To tackle these, we use the novel bound losses to enforce the model to produce depth estimation that follows the underwater imaging formation model. The final quantitative and qualitative evaluation results prove the effectiveness of our method. Additionally, the further experiment shows our method's generalization capabilities on other in-air models. + +Table 3: PUDE(M) on SQUID dataset: PUDE(M) shows improved performance on the metrics compared to MiDaS, but PUDE has the best performance. + +
Modelsδ < 1.05δ < 1.052δ < 1.053AbsRelSq. RelRMSERMSE logSILog
MiDaS0.1960.3790.5070.2960.3472.4440.3090.139
PUDE(M)0.2790.4630.6010.1880.1151.8690.2210.067
PUDE0.4350.6780.8030.1050.0341.3430.1860.037
+ +# Acknowledgements + +This work received funding from the Australian Research Council via grant DE220101527, and the Australian Government, via grant AUSMURIB000001. + +# References + +1. Akkaynak, D., Treibitz, T.: Sea-thru: A method for removing water from underwater images. In: Proceedings of the IEEE/CVF conference on computer vision and pattern recognition. pp. 1682-1691 (2019) +2. Amitai, S., Klein, I., Treibitz, T.: Self-supervised monocular depth underwater. In: 2023 IEEE International Conference on Robotics and Automation (ICRA). pp. 1098-1104. IEEE (2023) +3. Ancuti, C.O., Ancuti, C., De Vleeschouwer, C., Neumann, L., Garcia, R.: Color transfer for underwater dehazing and depth estimation. In: 2017 IEEE International Conference on Image Processing (ICIP). pp. 695-699. IEEE (2017) +4. Berman, D., Levy, D., Avidan, S., Treibitz, T.: Underwater single image color restoration using haze-lines and a new quantitative dataset. IEEE Transactions on Pattern Analysis and Machine Intelligence (2020) +5. Bhat, S.F., Alhashim, I., Wonka, P.: Adabins: Depth estimation using adaptive bins. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 4009-4018 (2021) +6. Bhat, S.F., Birkl, R., Wofk, D., Wonka, P., Müller, M.: Zoedepth: Zero-shot transfer by combining relative and metric depth. arXiv preprint arXiv:2302.12288 (2023) +7. Chang, H.H., Cheng, C.Y., Sung, C.C.: Single underwater image restoration based on depth estimation and transmission compensation. IEEE Journal of Oceanic Engineering 44(4), 1130-1149 (2018) +8. Chen, W., Qian, S., Deng, J.: Learning single-image depth from videos using quality assessment networks. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 5604-5613 (2019) +9. Chiang, J.Y., Chen, Y.C.: Underwater image enhancement by wavelength compensation and dehazing. IEEE transactions on image processing 21(4), 1756-1769 (2011) +0. Drews, P.L., Nascimento, E.R., Botelho, S.S., Campos, M.F.M.: Underwater depth estimation and image restoration based on single images. IEEE computer graphics and applications 36(2), 24-35 (2016) +1. Eigen, D., Puhrsch, C., Fergus, R.: Depth map prediction from a single image using a multi-scale deep network. Advances in neural information processing systems 27 (2014) +2. Fu, H., Gong, M., Wang, C., Batmanghelich, K., Tao, D.: Deep ordinal regression network for monocular depth estimation. In: Proceedings of the IEEE conference on computer vision and pattern recognition. pp. 2002-2011 (2018) +3. Godard, C., Mac Aodha, O., Brostow, G.J.: Unsupervised monocular depth estimation with left-right consistency. In: Proceedings of the IEEE conference on computer vision and pattern recognition. pp. 270-279 (2017) +4. Godard, C., Mac Aodha, O., Firman, M., Brostow, G.J.: Digging into self-supervised monocular depth estimation. In: Proceedings of the IEEE/CVF international conference on computer vision. pp. 3828-3838 (2019) + +15. Gupta, H., Mitra, K.: Unsupervised single image underwater depth estimation. In: 2019 IEEE International Conference on Image Processing (ICIP). pp. 624-628. IEEE (2019) +16. Hambarde, P., Murala, S., Dhall, A.: Uw-gan: Single-image depth estimation and image enhancement for underwater images. IEEE Transactions on Instrumentation and Measurement 70, 1-12 (2021) +17. He, K., Sun, J., Tang, X.: Single image haze removal using dark channel prior. IEEE transactions on pattern analysis and machine intelligence 33(12), 2341-2353 (2010) +18. Laina, I., Rupprecht, C., Belagiannis, V., Tombari, F., Navab, N.: Deeper depth prediction with fully convolutional residual networks. In: 2016 Fourth international conference on 3D vision (3DV). pp. 239-248. IEEE (2016) +19. Levy, D., Peleg, A., Pearl, N., Rosenbaum, D., Akkaynak, D., Korman, S., Treibitz, T.: Seathru-nerf: Neural radiance fields in scattering media. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 56-65 (2023) +20. Li, C., Guo, C., Ren, W., Cong, R., Hou, J., Kwong, S., Tao, D.: An underwater image enhancement benchmark dataset and beyond. IEEE Transactions on Image Processing 29, 4376-4389 (2019) +21. Li, Z., Snavely, N.: Megadepth: Learning single-view depth prediction from internet photos. In: Proceedings of the IEEE conference on computer vision and pattern recognition. pp. 2041-2050 (2018) +22. Luo, Y., Ren, J., Lin, M., Pang, J., Sun, W., Li, H., Lin, L.: Single view stereo matching. In: Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition. pp. 155-163 (2018) +23. Ming, Y., Meng, X., Fan, C., Yu, H.: Deep learning for monocular depth estimation: A review. Neurocomputing 438, 14-33 (2021) +24. Peng, Y.T., Cao, K., Cosman, P.C.: Generalization of the dark channel prior for single image restoration. IEEE Transactions on Image Processing 27(6), 2856-2868 (2018) +25. Peng, Y.T., Cosman, P.C.: Underwater image restoration based on image blurriness and light absorption. IEEE transactions on image processing 26(4), 1579-1594 (2017) +26. Peng, Y.T., Zhao, X., Cosman, P.C.: Single underwater image enhancement using depth estimation based on blurriness. In: 2015 IEEE International Conference on Image Processing (ICIP). pp. 4952-4956. IEEE (2015) +27. Ramamonjisoa, M., Lepetit, V.: Sharpnet: Fast and accurate recovery of occluding contours in monocular depth estimation. In: Proceedings of the IEEE/CVF International Conference on Computer Vision Workshops. pp. 0-0 (2019) +28. Ranftl, R., Bochkovskiy, A., Koltun, V.: Vision transformers for dense prediction. In: Proceedings of the IEEE/CVF international conference on computer vision. pp. 12179-12188 (2021) +29. Ranftl, R., Lasinger, K., Hafner, D., Schindler, K., Koltun, V.: Towards robust monocular depth estimation: Mixing datasets for zero-shot cross-dataset transfer. IEEE transactions on pattern analysis and machine intelligence 44(3), 1623-1637 (2020) +30. Varghese, N., Kumar, A., Rajagopalan, A.: Self-supervised monocular underwater depth recovery, image restoration, and a real-sea video dataset. In: Proceedings of the IEEE/CVF International Conference on Computer Vision. pp. 12248-12258 (2023) + +31. Vasiljevic, I., Kolkin, N., Zhang, S., Luo, R., Wang, H., Dai, F.Z., Daniele, A.F., Mostajabi, M., Basart, S., Walter, M.R., Shakhnarovich, G.: DIODE: A Dense Indoor and Outdoor DEpth Dataset. CoRR abs/1908.00463 (2019), http:// arxiv.org/abs/1908.00463 +32. Wu, C.Y., Wang, J., Hall, M., Neumann, U., Su, S.: Toward practical monocular indoor depth estimation. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 3814-3824 (2022) +33. Wu, Y., Zhou, Y., Chen, S., Ma, Y., Li, Q.: Defect inspection for underwater structures based on line-structured light and binocular vision. Applied Optics 60(25), 7754-7764 (2021) +34. Xian, K., Shen, C., Cao, Z., Lu, H., Xiao, Y., Li, R., Luo, Z.: Monocular relative depth perception with web stereo data supervision. In: Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition. pp. 311-320 (2018) +35. Yang, J., Gong, M., Nair, G., Lee, J.H., Monty, J., Pu, Y.: Knowledge distillation for feature extraction in underwater vslam. arXiv preprint arXiv:2303.17981 (2023) +36. Ye, X., Zhang, J., Yuan, Y., Xu, R., Wang, Z., Li, H.: Underwater depth estimation via stereo adaptation networks. IEEE Transactions on Circuits and Systems for Video Technology (2023) +37. Yu, B., Wu, J., Islam, M.J.: Udepth: Fast monocular depth estimation for visually-guided underwater robots. In: 2023 IEEE International Conference on Robotics and Automation (ICRA). pp. 3116-3123. IEEE (2023) \ No newline at end of file diff --git a/2024/Physics-informed Knowledge Transfer for Underwater Monocular Depth Estimation/images.zip b/2024/Physics-informed Knowledge Transfer for Underwater Monocular Depth Estimation/images.zip new file mode 100644 index 0000000000000000000000000000000000000000..609c02ed0a36b4ef2efb6a61c301409360252913 --- /dev/null +++ b/2024/Physics-informed Knowledge Transfer for Underwater Monocular Depth Estimation/images.zip @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:9fc4f4a2e10e4da1961fa74db3e1917852648abf5df35f7ac918a0cf0446d61c +size 468661 diff --git a/2024/Physics-informed Knowledge Transfer for Underwater Monocular Depth Estimation/layout.json b/2024/Physics-informed Knowledge Transfer for Underwater Monocular Depth Estimation/layout.json new file mode 100644 index 0000000000000000000000000000000000000000..a90f8c6bdebf2d12875bc7633b67f31e20eea3ba --- /dev/null +++ b/2024/Physics-informed Knowledge Transfer for Underwater Monocular Depth Estimation/layout.json @@ -0,0 +1,9777 @@ +{ + "pdf_info": [ + { + "para_blocks": [ + { + "bbox": [ + 161, + 111, + 452, + 147 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 161, + 111, + 452, + 147 + ], + "spans": [ + { + "bbox": [ + 161, + 111, + 452, + 147 + ], + "type": "text", + "content": "Physics-informed Knowledge Transfer for Underwater Monocular Depth Estimation" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 214, + 168, + 399, + 180 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 214, + 168, + 399, + 180 + ], + "spans": [ + { + "bbox": [ + 214, + 168, + 399, + 180 + ], + "type": "text", + "content": "Jinghe Yang, Mingming Gong, and Ye Pu" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 144, + 190, + 468, + 213 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 144, + 190, + 468, + 213 + ], + "spans": [ + { + "bbox": [ + 144, + 190, + 468, + 213 + ], + "type": "text", + "content": "The University of Melbourne jinghey@student.unimelb.edu.au, {mingming.gong, ye.pu}@unimelb.edu.au" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 160, + 236, + 462, + 423 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 160, + 236, + 462, + 423 + ], + "spans": [ + { + "bbox": [ + 160, + 236, + 462, + 423 + ], + "type": "text", + "content": "Abstract. Compared to the in-air case, underwater depth estimation has its own challenges. For instance, acquiring high-quality training datasets with groundtruth poses difficulties due to sensor limitations in aquatic environments. Additionally, the physics characteristics of underwater imaging diverge significantly from the in-air case, the methods developed for in-air depth estimation underperform when applied underwater, due to the domain gap. To address these challenges, our paper introduces a novel transfer-learning-based method - Physics-informed Underwater Depth Estimation (PUDE). The key idea is to transfer the knowledge of a pre-trained in-air depth estimation model to underwater settings utilizing a small underwater image set without groundtruth measurement, guided by a physical underwater imaging formation model. We propose novel bound losses based on the physical model to rectify the depth estimations to align with actual underwater physical properties. Finally, in the evaluations across multiple datasets, we compare PUDE model with other existing in-air and underwater methods. The results reveal that the PUDE model excels in both quantitative and qualitative comparisons." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 160, + 434, + 453, + 445 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 160, + 434, + 453, + 445 + ], + "spans": [ + { + "bbox": [ + 160, + 434, + 453, + 445 + ], + "type": "text", + "content": "Keywords: Underwater Depth " + }, + { + "bbox": [ + 160, + 434, + 453, + 445 + ], + "type": "inline_equation", + "content": "\\cdot" + }, + { + "bbox": [ + 160, + 434, + 453, + 445 + ], + "type": "text", + "content": " Physics-informed " + }, + { + "bbox": [ + 160, + 434, + 453, + 445 + ], + "type": "inline_equation", + "content": "\\cdot" + }, + { + "bbox": [ + 160, + 434, + 453, + 445 + ], + "type": "text", + "content": " Knowledge Transfer" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 133, + 463, + 230, + 475 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 463, + 230, + 475 + ], + "spans": [ + { + "bbox": [ + 133, + 463, + 230, + 475 + ], + "type": "text", + "content": "1 Introduction" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 130, + 486, + 482, + 668 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 486, + 482, + 668 + ], + "spans": [ + { + "bbox": [ + 130, + 486, + 482, + 668 + ], + "type": "text", + "content": "Depth estimation is a pivotal task in computer vision, with substantial implications for robotics applications. Traditional techniques typically employ stereo cameras to estimate depth through disparity calculations. However, these methods are hampered by complex camera system settings and limited by the challenging feature matching in environments with low texture or visibility [23]. In contrast, monocular depth estimation provides an alternative solution, necessitating only a single camera and a less complicated setup. Although monocular depth estimation faces inherent limitations due to the absence of stereoscopic information, many well-performing in-air learning-based depth estimation models have been developed in recent years [5,6,12,21,27-29]. Benefiting from the massive and various training datasets, some of in-air monocular depth estimation models, such as MiDaS [29] and DPT [28], demonstrate powerful generalization performance. However, the pre-trained in-air models are not fully applicable for underwater cases due to the domain gap and the differences in image formation [36]. For example, we applied the in-air trained models, e.g., DPT in [28], to" + } + ] + } + ], + "index": 6 + } + ], + "discarded_blocks": [], + "page_size": [ + 612, + 792 + ], + "page_idx": 0 + }, + { + "para_blocks": [ + { + "bbox": [ + 130, + 116, + 482, + 355 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 116, + 482, + 355 + ], + "spans": [ + { + "bbox": [ + 130, + 116, + 482, + 355 + ], + "type": "text", + "content": "underwater images and observed that DPT has the issues of the overestimation for blurry regions, unclear background region detection and loss of depth details. Additionally, it is challenging to train new models from scratch for underwater settings. Underwater scenarios impose difficulties when capturing a large amount of high-quality images and the groundtruth depth, due to sensor limitations, such as the incapability of LiDAR and the performance drop of stereo matching in water. In contrast, sonar scanning is not affected by water quality and blurriness but suffers from a slow scanning rate and low resolution, lacking detail in scene representation [33]. In response to the distinct challenges of underwater depth estimation, this paper proposes a novel transfer-learning-based approach. Our method transfers knowledge from a pre-trained in-air monocular depth estimation model to underwater settings, guided by the underwater imaging formation model introduced in [9]. Using newly proposed bound losses, which leverage physical information to correct inaccurate depth in distant and background regions, our training method enforces the model to produce estimates aligned with the extracted physical properties. Notably, this adaptation requires minimal training data and no ground truth depth measurements. Our approach benefits from the structural knowledge of the in-air model and integrates underwater physics knowledge, thereby enhancing underwater depth estimation performance. Our contributions are as follows:" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 138, + 363, + 482, + 542 + ], + "type": "list", + "angle": 0, + "index": 6, + "blocks": [ + { + "bbox": [ + 138, + 363, + 482, + 458 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 363, + 482, + 458 + ], + "spans": [ + { + "bbox": [ + 138, + 363, + 482, + 458 + ], + "type": "text", + "content": "- We present experimental observations and limitations when applying an in-air trained model, i.e., DPT in [28], to underwater images. Inspired by these observations, we propose a novel training method to adapt in-air pre-trained depth estimation models for underwater environments. There are two key ingredients in our method: 1) Knowledge transfer using an in-air trained model with good generalisation properties (DPT [28]) to address data scarcity; 2) Integration of the underwater image formation model proposed in [9] to ensure the depth estimates align with underwater physics characteristics." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 138, + 459, + 482, + 483 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 459, + 482, + 483 + ], + "spans": [ + { + "bbox": [ + 138, + 459, + 482, + 483 + ], + "type": "text", + "content": "- We propose innovative bound losses to steer the model towards producing depth estimations that adhere to the underwater imaging formation model." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 138, + 483, + 482, + 542 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 483, + 482, + 542 + ], + "spans": [ + { + "bbox": [ + 138, + 483, + 482, + 542 + ], + "type": "text", + "content": "- We introduce the Physics-informed Underwater Depth Estimation (PUDE) model, an end-to-end model for underwater monocular depth estimation. In zero-shot evaluations across various datasets, our method demonstrates enhanced performance, both quantitatively and qualitatively, compared to other in-air and underwater methods." + } + ] + } + ], + "index": 5 + } + ], + "sub_type": "text" + }, + { + "bbox": [ + 132, + 560, + 233, + 573 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 560, + 233, + 573 + ], + "spans": [ + { + "bbox": [ + 132, + 560, + 233, + 573 + ], + "type": "text", + "content": "2 RelatedWork" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 132, + 586, + 340, + 597 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 586, + 340, + 597 + ], + "spans": [ + { + "bbox": [ + 132, + 586, + 340, + 597 + ], + "type": "text", + "content": "2.1 In-air Monocular Depth Estimation" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 130, + 605, + 482, + 666 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 605, + 482, + 666 + ], + "spans": [ + { + "bbox": [ + 130, + 605, + 482, + 666 + ], + "type": "text", + "content": "Compared to LiDAR, infrared, and stereo vision, monocular depth estimation stands out for its cost-effectiveness and simplicity of configuration, making it appealing for depth perception in autonomous systems. Recent in-air monocular depth estimation methods have exhibited commendable performance, especially supervised methods trained on datasets containing measured groundtruth. They" + } + ] + } + ], + "index": 9 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "spans": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "type": "text", + "content": "2" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 166, + 91, + 222, + 102 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 166, + 91, + 222, + 102 + ], + "spans": [ + { + "bbox": [ + 166, + 91, + 222, + 102 + ], + "type": "text", + "content": "J. Yang et al." + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 1 + }, + { + "para_blocks": [ + { + "bbox": [ + 130, + 116, + 479, + 152 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 116, + 479, + 152 + ], + "spans": [ + { + "bbox": [ + 130, + 116, + 479, + 152 + ], + "type": "text", + "content": "usually require sensors like LiDAR, RGBD cameras, and stereo matching [11,12, 18,22]. However, these methods are limited by the lack of variety in the datasets, which restricts their generalization across different real-world scenarios." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 130, + 152, + 482, + 296 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 152, + 482, + 296 + ], + "spans": [ + { + "bbox": [ + 130, + 152, + 482, + 296 + ], + "type": "text", + "content": "In this context, self-supervised learning using continuous image sequences [14] or calibrated stereo image pairs [13] presents an attractive avenue for exploration but performs less compared to supervised learning [23], and its application to dynamic scenes is challenging. Recognizing this limitation, some studies expand the breadth of training data by harnessing web resources to generate depth groundtruth, thereby enlarging the training dataset and encompassing diverse scenes [8,21,34]. Building further on that, MiDaS [29] extract depth data from 3D movies to enhance dataset diversity and improve generalization. DPT [28] builds upon the MiDaS training dataset, but with supplementary, and implements an encoder-decoder architecture that uses a vision transformer as its backbone, enabling fine-grained and globally coherent estimations and showing strong zero-shot cross-dataset generalization capabilities." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 131, + 312, + 371, + 323 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 131, + 312, + 371, + 323 + ], + "spans": [ + { + "bbox": [ + 131, + 312, + 371, + 323 + ], + "type": "text", + "content": "2.2 Underwater Monocular Depth Estimation" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 130, + 331, + 479, + 413 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 331, + 479, + 413 + ], + "spans": [ + { + "bbox": [ + 130, + 331, + 479, + 413 + ], + "type": "text", + "content": "Depth measurement in water introduces new challenges due to the unique characteristics of the underwater environment. The limited availability of sensors, such as LiDAR and infrared, makes the direct acquisition of the depth measurement highly challenging. Meanwhile, the image blurriness highly influences the correspondence matching between images [35], thereby the stereo camera can only provide good depth measurement for close regions. The existing underwater depth datasets that rely on stereo matching [1,4], manifest this issue." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 130, + 415, + 482, + 582 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 415, + 482, + 582 + ], + "spans": [ + { + "bbox": [ + 130, + 415, + 482, + 582 + ], + "type": "text", + "content": "Therefore, monocular depth estimation is a potential solution to tackle these issues. Physics-based methods play a pivotal role in addressing the underwater depth estimation challenge. The underwater image formation model proposed in [9] has been demonstrated to be effective in solving problems in computer vision for underwater scenarios, such as feature extraction [35], neural radiance fields (NeRFs) [19], and depth estimation [10]. Traditional methods utilize the physics-based image formation model to estimate depth directly from underwater images [3,4,7,24-26]. As the foundation of these methods, Dark Channel Prior (DCP) [17], a statistical method initially developed for in-air haze removal, uses pixel intensity to estimate medium transmission. A variant for the underwater environment, UDCP [10], claims that the DCP assumption remains applicable to the green and blue channels of underwater images. However, underwater depth estimation that relies solely on pixel values and the imaging formation model is vulnerable to variable lighting conditions, such as shadows and sunlight flickers." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 130, + 582, + 482, + 665 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 582, + 482, + 665 + ], + "spans": [ + { + "bbox": [ + 130, + 582, + 482, + 665 + ], + "type": "text", + "content": "Despite the challenges in depth groundtruth acquisition underwater, the learning-based methods show promise. Some works have attempted unsupervised learning, such as UwNET [15] and UW-GAN [16]. They use the GAN method through color restoration to estimate depth. The recent works [2,30] adopt self-supervised learning, utilizing the correspondences between subsequent frames. Notably, they incorporate the physics model into the training of the underwater depth estimation models and have shown to be advantageous." + } + ] + } + ], + "index": 7 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 419, + 90, + 447, + 100 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 419, + 90, + 447, + 100 + ], + "spans": [ + { + "bbox": [ + 419, + 90, + 447, + 100 + ], + "type": "text", + "content": "PUDE" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 474, + 91, + 481, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 474, + 91, + 481, + 100 + ], + "spans": [ + { + "bbox": [ + 474, + 91, + 481, + 100 + ], + "type": "text", + "content": "3" + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 2 + }, + { + "para_blocks": [ + { + "bbox": [ + 132, + 116, + 356, + 129 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 116, + 356, + 129 + ], + "spans": [ + { + "bbox": [ + 132, + 116, + 356, + 129 + ], + "type": "text", + "content": "2.3 Underwater Imaging Formation Model" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 130, + 136, + 482, + 221 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 136, + 482, + 221 + ], + "spans": [ + { + "bbox": [ + 130, + 136, + 482, + 221 + ], + "type": "text", + "content": "Light transmission in water is subject to absorption, and minuscule particles cause the backscattering, leading to blur within underwater imagery. A widely used model of underwater image formation is proposed in [9] to describe this physical phenomenon. Consider an underwater image represented by " + }, + { + "bbox": [ + 130, + 136, + 482, + 221 + ], + "type": "inline_equation", + "content": "I_{c} \\in \\mathbb{R}^{HW}" + }, + { + "bbox": [ + 130, + 136, + 482, + 221 + ], + "type": "text", + "content": ", where " + }, + { + "bbox": [ + 130, + 136, + 482, + 221 + ], + "type": "inline_equation", + "content": "c" + }, + { + "bbox": [ + 130, + 136, + 482, + 221 + ], + "type": "text", + "content": " denotes the color channels " + }, + { + "bbox": [ + 130, + 136, + 482, + 221 + ], + "type": "inline_equation", + "content": "\\{\\mathrm{R}, \\mathrm{G}, \\mathrm{B}\\}" + }, + { + "bbox": [ + 130, + 136, + 482, + 221 + ], + "type": "text", + "content": ", and " + }, + { + "bbox": [ + 130, + 136, + 482, + 221 + ], + "type": "inline_equation", + "content": "H" + }, + { + "bbox": [ + 130, + 136, + 482, + 221 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 130, + 136, + 482, + 221 + ], + "type": "inline_equation", + "content": "W" + }, + { + "bbox": [ + 130, + 136, + 482, + 221 + ], + "type": "text", + "content": " are the image's height and width, respectively. Given an image " + }, + { + "bbox": [ + 130, + 136, + 482, + 221 + ], + "type": "inline_equation", + "content": "I_{c}" + }, + { + "bbox": [ + 130, + 136, + 482, + 221 + ], + "type": "text", + "content": ", for each pixel " + }, + { + "bbox": [ + 130, + 136, + 482, + 221 + ], + "type": "inline_equation", + "content": "i" + }, + { + "bbox": [ + 130, + 136, + 482, + 221 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 130, + 136, + 482, + 221 + ], + "type": "inline_equation", + "content": "i = 1 \\cdots HW" + }, + { + "bbox": [ + 130, + 136, + 482, + 221 + ], + "type": "text", + "content": ", it holds that " + }, + { + "bbox": [ + 130, + 136, + 482, + 221 + ], + "type": "inline_equation", + "content": "I_{c,i} \\in [0,1]" + }, + { + "bbox": [ + 130, + 136, + 482, + 221 + ], + "type": "text", + "content": ". The underwater imaging formation model is:" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 241, + 230, + 481, + 243 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 241, + 230, + 481, + 243 + ], + "spans": [ + { + "bbox": [ + 241, + 230, + 481, + 243 + ], + "type": "interline_equation", + "content": "I _ {c, i} = J _ {c, i} \\cdot t _ {c, i} + B _ {c} ^ {\\infty} (1 - t _ {c, i}), \\tag {1}", + "image_path": "5b887cad38e40c24ed8d3eb106befb20eb74f2f17844586d6765e018bbcf445b.jpg" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 278, + 262, + 481, + 277 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 278, + 262, + 481, + 277 + ], + "spans": [ + { + "bbox": [ + 278, + 262, + 481, + 277 + ], + "type": "interline_equation", + "content": "t _ {c, i} = e ^ {- \\beta_ {c} \\mathbf {z} _ {i}}. (2)", + "image_path": "f5dbfcfaac2a083f8a79af562e630b0a058118d42ec6fc34314eae4112bdfd72.jpg" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 130, + 281, + 482, + 378 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 281, + 482, + 378 + ], + "spans": [ + { + "bbox": [ + 130, + 281, + 482, + 378 + ], + "type": "text", + "content": "In this model, " + }, + { + "bbox": [ + 130, + 281, + 482, + 378 + ], + "type": "inline_equation", + "content": "J_{c,i}" + }, + { + "bbox": [ + 130, + 281, + 482, + 378 + ], + "type": "text", + "content": " is the original in-air color value at pixel " + }, + { + "bbox": [ + 130, + 281, + 482, + 378 + ], + "type": "inline_equation", + "content": "i" + }, + { + "bbox": [ + 130, + 281, + 482, + 378 + ], + "type": "text", + "content": ". The original in-air image is denoted by " + }, + { + "bbox": [ + 130, + 281, + 482, + 378 + ], + "type": "inline_equation", + "content": "J_{c} = [J_{c,1},\\dots,J_{c,HW}] \\in \\mathbb{R}^{HW}" + }, + { + "bbox": [ + 130, + 281, + 482, + 378 + ], + "type": "text", + "content": ". The medium transmission rate denoted by " + }, + { + "bbox": [ + 130, + 281, + 482, + 378 + ], + "type": "inline_equation", + "content": "t_{c,i}" + }, + { + "bbox": [ + 130, + 281, + 482, + 378 + ], + "type": "text", + "content": " is associated with the beam attenuation coefficient " + }, + { + "bbox": [ + 130, + 281, + 482, + 378 + ], + "type": "inline_equation", + "content": "\\beta_{c}" + }, + { + "bbox": [ + 130, + 281, + 482, + 378 + ], + "type": "text", + "content": " and the depth " + }, + { + "bbox": [ + 130, + 281, + 482, + 378 + ], + "type": "inline_equation", + "content": "\\mathbf{z}_{i}" + }, + { + "bbox": [ + 130, + 281, + 482, + 378 + ], + "type": "text", + "content": ". Meanwhile, " + }, + { + "bbox": [ + 130, + 281, + 482, + 378 + ], + "type": "inline_equation", + "content": "B_{c}^{\\infty}" + }, + { + "bbox": [ + 130, + 281, + 482, + 378 + ], + "type": "text", + "content": " is the coefficient of ambient light causing the backscattering from particles. The captured underwater image " + }, + { + "bbox": [ + 130, + 281, + 482, + 378 + ], + "type": "inline_equation", + "content": "I_{c,i}" + }, + { + "bbox": [ + 130, + 281, + 482, + 378 + ], + "type": "text", + "content": " is the aggregate of the attenuation of the original image and the additional backscattering from the water environment. The values of " + }, + { + "bbox": [ + 130, + 281, + 482, + 378 + ], + "type": "inline_equation", + "content": "\\beta_{c}" + }, + { + "bbox": [ + 130, + 281, + 482, + 378 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 130, + 281, + 482, + 378 + ], + "type": "inline_equation", + "content": "B_{c}^{\\infty}" + }, + { + "bbox": [ + 130, + 281, + 482, + 378 + ], + "type": "text", + "content": " vary depending on different water environments and can be estimated using the depth measurements [1]." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 132, + 396, + 407, + 410 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 396, + 407, + 410 + ], + "spans": [ + { + "bbox": [ + 132, + 396, + 407, + 410 + ], + "type": "text", + "content": "3 DPT on Underwater Images: Observations" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 130, + 421, + 482, + 506 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 421, + 482, + 506 + ], + "spans": [ + { + "bbox": [ + 130, + 421, + 482, + 506 + ], + "type": "text", + "content": "Existing learning-based monocular depth estimation methods, such as DPT [28] and MiDaS [29], generate the inverse relative depth estimation from a single image. We denote the inverse depth estimation as " + }, + { + "bbox": [ + 130, + 421, + 482, + 506 + ], + "type": "inline_equation", + "content": "\\mathbf{d} \\in \\mathbb{R}^{HW}" + }, + { + "bbox": [ + 130, + 421, + 482, + 506 + ], + "type": "text", + "content": ", with " + }, + { + "bbox": [ + 130, + 421, + 482, + 506 + ], + "type": "inline_equation", + "content": "\\mathbf{d}_i" + }, + { + "bbox": [ + 130, + 421, + 482, + 506 + ], + "type": "text", + "content": " to be the estimate for pixel " + }, + { + "bbox": [ + 130, + 421, + 482, + 506 + ], + "type": "inline_equation", + "content": "i" + }, + { + "bbox": [ + 130, + 421, + 482, + 506 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 130, + 421, + 482, + 506 + ], + "type": "inline_equation", + "content": "i = 1, \\dots, HW" + }, + { + "bbox": [ + 130, + 421, + 482, + 506 + ], + "type": "text", + "content": ". The metric depth of the input image is denoted by " + }, + { + "bbox": [ + 130, + 421, + 482, + 506 + ], + "type": "inline_equation", + "content": "\\mathbf{z} \\in \\mathbb{R}^{HW}" + }, + { + "bbox": [ + 130, + 421, + 482, + 506 + ], + "type": "text", + "content": ", whose reciprocal ideally has a linear relationship with the inverse relative depth. The relationship between inverse relative depth " + }, + { + "bbox": [ + 130, + 421, + 482, + 506 + ], + "type": "inline_equation", + "content": "\\mathbf{d}_i" + }, + { + "bbox": [ + 130, + 421, + 482, + 506 + ], + "type": "text", + "content": " and the metric depth " + }, + { + "bbox": [ + 130, + 421, + 482, + 506 + ], + "type": "inline_equation", + "content": "\\mathbf{z}_i" + }, + { + "bbox": [ + 130, + 421, + 482, + 506 + ], + "type": "text", + "content": " at pixel " + }, + { + "bbox": [ + 130, + 421, + 482, + 506 + ], + "type": "inline_equation", + "content": "i" + }, + { + "bbox": [ + 130, + 421, + 482, + 506 + ], + "type": "text", + "content": " is:" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 230, + 514, + 481, + 539 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 230, + 514, + 481, + 539 + ], + "spans": [ + { + "bbox": [ + 230, + 514, + 481, + 539 + ], + "type": "interline_equation", + "content": "\\mathbf {z} _ {i} = \\frac {1}{s \\mathbf {d} _ {i} + h}, \\quad \\forall i \\in \\{1, \\dots , H W \\}. \\tag {3}", + "image_path": "cc6914126190541ee7b3403a68f78449b6ac8f83816fabfd3b5c13ec556e2a3e.jpg" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 130, + 545, + 482, + 666 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 545, + 482, + 666 + ], + "spans": [ + { + "bbox": [ + 130, + 545, + 482, + 666 + ], + "type": "text", + "content": "Here, " + }, + { + "bbox": [ + 130, + 545, + 482, + 666 + ], + "type": "inline_equation", + "content": "s" + }, + { + "bbox": [ + 130, + 545, + 482, + 666 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 130, + 545, + 482, + 666 + ], + "type": "inline_equation", + "content": "h" + }, + { + "bbox": [ + 130, + 545, + 482, + 666 + ], + "type": "text", + "content": " are two unknown scale and shift coefficients. In practice, they can be off-line estimated for specific scenes using sequence images. The overall mapping relationship between the inverse relative depth " + }, + { + "bbox": [ + 130, + 545, + 482, + 666 + ], + "type": "inline_equation", + "content": "\\mathbf{d}_i" + }, + { + "bbox": [ + 130, + 545, + 482, + 666 + ], + "type": "text", + "content": " and the inverse of metric depth " + }, + { + "bbox": [ + 130, + 545, + 482, + 666 + ], + "type": "inline_equation", + "content": "\\mathbf{z}_i" + }, + { + "bbox": [ + 130, + 545, + 482, + 666 + ], + "type": "text", + "content": " is linear on the in-air images [32]. An example can be seen in the first row of Fig. 1, which is the depth estimation of DPT on an outdoor in-air image from DIODE [31], we can observe that DPT estimation on the in-air image maintains a good linear relationship (the correlation coefficient " + }, + { + "bbox": [ + 130, + 545, + 482, + 666 + ], + "type": "inline_equation", + "content": "r = 0.992" + }, + { + "bbox": [ + 130, + 545, + 482, + 666 + ], + "type": "text", + "content": "). However, when we apply DPT directly to underwater images from the dataset [1,4], in which the depth groundtruth is obtained through the Structure-From-Motion and stereo matching, it is observed that in underwater imagery, the estimated" + } + ] + } + ], + "index": 10 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "spans": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "type": "text", + "content": "4" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 166, + 91, + 222, + 102 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 166, + 91, + 222, + 102 + ], + "spans": [ + { + "bbox": [ + 166, + 91, + 222, + 102 + ], + "type": "text", + "content": "J. Yang et al." + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 3 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 149, + 114, + 254, + 184 + ], + "blocks": [ + { + "bbox": [ + 149, + 114, + 254, + 184 + ], + "lines": [ + { + "bbox": [ + 149, + 114, + 254, + 184 + ], + "spans": [ + { + "bbox": [ + 149, + 114, + 254, + 184 + ], + "type": "image", + "image_path": "8824e2ba1a6351826f34979e8a17fac4e2e9017600967b24d71b5ee8309f5e69.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_body" + } + ], + "index": 2 + }, + { + "type": "image", + "bbox": [ + 255, + 114, + 359, + 184 + ], + "blocks": [ + { + "bbox": [ + 255, + 114, + 359, + 184 + ], + "lines": [ + { + "bbox": [ + 255, + 114, + 359, + 184 + ], + "spans": [ + { + "bbox": [ + 255, + 114, + 359, + 184 + ], + "type": "image", + "image_path": "bfbd28f515df07458287e11034a532f3cf58557a3da4626d1d260b559bdd9ada.jpg" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_body" + } + ], + "index": 3 + }, + { + "type": "image", + "bbox": [ + 359, + 118, + 457, + 184 + ], + "blocks": [ + { + "bbox": [ + 359, + 118, + 457, + 184 + ], + "lines": [ + { + "bbox": [ + 359, + 118, + 457, + 184 + ], + "spans": [ + { + "bbox": [ + 359, + 118, + 457, + 184 + ], + "type": "image", + "image_path": "f2910ee34d2f4aebf56aaede5910cfa5207fa1a7b2d273588bb9b6b5d425e791.jpg" + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "image_body" + } + ], + "index": 4 + }, + { + "type": "image", + "bbox": [ + 150, + 184, + 254, + 255 + ], + "blocks": [ + { + "bbox": [ + 150, + 184, + 254, + 255 + ], + "lines": [ + { + "bbox": [ + 150, + 184, + 254, + 255 + ], + "spans": [ + { + "bbox": [ + 150, + 184, + 254, + 255 + ], + "type": "image", + "image_path": "95652f5ed62bf7ea885a6a09c5f4dc3966cf5b6d9b54df8ad997435039bd22e8.jpg" + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 131, + 262, + 480, + 297 + ], + "lines": [ + { + "bbox": [ + 131, + 262, + 480, + 297 + ], + "spans": [ + { + "bbox": [ + 131, + 262, + 480, + 297 + ], + "type": "text", + "content": "Fig. 1: DPT estimation of underwater images shows a curved tendency, with overestimation in distant regions. Additionally, the estimated depth struggles with unclear background detection (blue arrows)." + } + ] + } + ], + "index": 8, + "angle": 0, + "type": "image_caption" + } + ], + "index": 5 + }, + { + "type": "image", + "bbox": [ + 255, + 184, + 359, + 255 + ], + "blocks": [ + { + "bbox": [ + 255, + 184, + 359, + 255 + ], + "lines": [ + { + "bbox": [ + 255, + 184, + 359, + 255 + ], + "spans": [ + { + "bbox": [ + 255, + 184, + 359, + 255 + ], + "type": "image", + "image_path": "91ffd5b80d32babd4b2dffd30e7ac4b2d4dd74ee97c866842af8943309f1d6c8.jpg" + } + ] + } + ], + "index": 6, + "angle": 0, + "type": "image_body" + } + ], + "index": 6 + }, + { + "type": "image", + "bbox": [ + 359, + 188, + 454, + 255 + ], + "blocks": [ + { + "bbox": [ + 359, + 188, + 454, + 255 + ], + "lines": [ + { + "bbox": [ + 359, + 188, + 454, + 255 + ], + "spans": [ + { + "bbox": [ + 359, + 188, + 454, + 255 + ], + "type": "image", + "image_path": "e4f3271b4175f228a5ca1c95387f5af299ca6a8c15de564c09aad55620ece144.jpg" + } + ] + } + ], + "index": 7, + "angle": 0, + "type": "image_body" + } + ], + "index": 7 + }, + { + "bbox": [ + 130, + 324, + 480, + 363 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 324, + 480, + 363 + ], + "spans": [ + { + "bbox": [ + 130, + 324, + 480, + 363 + ], + "type": "text", + "content": "inverse depth, denoted by " + }, + { + "bbox": [ + 130, + 324, + 480, + 363 + ], + "type": "inline_equation", + "content": "\\mathbf{d}_i^D" + }, + { + "bbox": [ + 130, + 324, + 480, + 363 + ], + "type": "text", + "content": ", and " + }, + { + "bbox": [ + 130, + 324, + 480, + 363 + ], + "type": "inline_equation", + "content": "\\frac{1}{z_i}" + }, + { + "bbox": [ + 130, + 324, + 480, + 363 + ], + "type": "text", + "content": " deviates from a linear relationship and present a curve tendency with a weaker correlation, especially when " + }, + { + "bbox": [ + 130, + 324, + 480, + 363 + ], + "type": "inline_equation", + "content": "\\mathbf{d}_i^D" + }, + { + "bbox": [ + 130, + 324, + 480, + 363 + ], + "type": "text", + "content": " is small. An example is shown in the second row of Fig. 1." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 130, + 365, + 482, + 534 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 365, + 482, + 534 + ], + "spans": [ + { + "bbox": [ + 130, + 365, + 482, + 534 + ], + "type": "text", + "content": "To observe the nonlinear trend caused by image blurriness, we applied the Sea-thru method [1] to calculate the medium transmission rate " + }, + { + "bbox": [ + 130, + 365, + 482, + 534 + ], + "type": "inline_equation", + "content": "t_i" + }, + { + "bbox": [ + 130, + 365, + 482, + 534 + ], + "type": "text", + "content": " in Eq. (2) on the grayscale images. We then compared these rates with those obtained from the DPT depth (aligned with groundtruth depth). An example of this procedure on the Sea-thru D3 dataset is shown in Fig. 2a, where we randomly sample 100 points from each image. We observe that the estimated transmission rate from DPT is mostly lower than that calculated by groundtruth depth when the transmission rate is small (" + }, + { + "bbox": [ + 130, + 365, + 482, + 534 + ], + "type": "inline_equation", + "content": "t_i < 0.2" + }, + { + "bbox": [ + 130, + 365, + 482, + 534 + ], + "type": "text", + "content": "). The results reveal that DPT underestimates the transmission rate in regions with low transmission rates, leading to an overestimation of depth in distant regions. Furthermore, the background edges on the underwater depth image of DPT, such as the blue arrow pointed regions around corals, are less clear, exhibiting fogging compared with the estimation on the in-air image. Based on these experimental observations, we summarise the primary issues of the depth estimation from DPT in water:" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 130, + 551, + 480, + 638 + ], + "type": "list", + "angle": 0, + "index": 13, + "blocks": [ + { + "bbox": [ + 130, + 551, + 480, + 586 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 551, + 480, + 586 + ], + "spans": [ + { + "bbox": [ + 130, + 551, + 480, + 586 + ], + "type": "text", + "content": "O1 : DPT exhibits a propensity for overestimating depth in regions that are relatively far away from the camera. This overestimation is particularly evident in regions with diminished visual clarity and increased blurriness." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 130, + 590, + 480, + 638 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 590, + 480, + 638 + ], + "spans": [ + { + "bbox": [ + 130, + 590, + 480, + 638 + ], + "type": "text", + "content": "O2 : Secondly, DPT encounters challenges in accurately discerning background regions in underwater images. This issue is especially pronounced for the region far from the camera. Due to the attenuation of the scenes, it is challenging for DPT to define the background area clearly." + } + ] + } + ], + "index": 12 + } + ], + "sub_type": "text" + }, + { + "bbox": [ + 131, + 653, + 473, + 665 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 131, + 653, + 473, + 665 + ], + "spans": [ + { + "bbox": [ + 131, + 653, + 473, + 665 + ], + "type": "text", + "content": "We propose two bound losses (Sec. 4.3) in our training to address these issues." + } + ] + } + ], + "index": 14 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 419, + 91, + 447, + 100 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 419, + 91, + 447, + 100 + ], + "spans": [ + { + "bbox": [ + 419, + 91, + 447, + 100 + ], + "type": "text", + "content": "PUDE" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 474, + 92, + 480, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 474, + 92, + 480, + 100 + ], + "spans": [ + { + "bbox": [ + 474, + 92, + 480, + 100 + ], + "type": "text", + "content": "5" + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 4 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 143, + 118, + 222, + 194 + ], + "blocks": [ + { + "bbox": [ + 143, + 118, + 222, + 194 + ], + "lines": [ + { + "bbox": [ + 143, + 118, + 222, + 194 + ], + "spans": [ + { + "bbox": [ + 143, + 118, + 222, + 194 + ], + "type": "image", + "image_path": "b2317f68ed0d21e33d87d820aff680737832f55f90f08e3ad833b9797c4c8d4f.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 133, + 196, + 238, + 206 + ], + "lines": [ + { + "bbox": [ + 133, + 196, + 238, + 206 + ], + "spans": [ + { + "bbox": [ + 133, + 196, + 238, + 206 + ], + "type": "text", + "content": "(a) " + }, + { + "bbox": [ + 133, + 196, + 238, + 206 + ], + "type": "inline_equation", + "content": "t_i" + }, + { + "bbox": [ + 133, + 196, + 238, + 206 + ], + "type": "text", + "content": " in Eq. (2) (D3 dataset)" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_caption" + } + ], + "index": 2 + }, + { + "type": "image", + "bbox": [ + 260, + 119, + 375, + 196 + ], + "blocks": [ + { + "bbox": [ + 260, + 119, + 375, + 196 + ], + "lines": [ + { + "bbox": [ + 260, + 119, + 375, + 196 + ], + "spans": [ + { + "bbox": [ + 260, + 119, + 375, + 196 + ], + "type": "image", + "image_path": "535ee7551ae3b70ee6299384f427519641887e50abc27238662b9da9c87baf56.jpg" + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 265, + 196, + 462, + 206 + ], + "lines": [ + { + "bbox": [ + 265, + 196, + 462, + 206 + ], + "spans": [ + { + "bbox": [ + 265, + 196, + 462, + 206 + ], + "type": "text", + "content": "(b) Point selection and model fitting in Eq. (1), Eq. (3)" + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 130, + 215, + 482, + 258 + ], + "lines": [ + { + "bbox": [ + 130, + 215, + 482, + 258 + ], + "spans": [ + { + "bbox": [ + 130, + 215, + 482, + 258 + ], + "type": "text", + "content": "Fig.2: (a) The comparison between the calculated transmission rate " + }, + { + "bbox": [ + 130, + 215, + 482, + 258 + ], + "type": "inline_equation", + "content": "t_i" + }, + { + "bbox": [ + 130, + 215, + 482, + 258 + ], + "type": "text", + "content": " from groundtruth and the estimated " + }, + { + "bbox": [ + 130, + 215, + 482, + 258 + ], + "type": "inline_equation", + "content": "t_i" + }, + { + "bbox": [ + 130, + 215, + 482, + 258 + ], + "type": "text", + "content": " from DPT depth, presents an underestimation of the " + }, + { + "bbox": [ + 130, + 215, + 482, + 258 + ], + "type": "inline_equation", + "content": "t_i" + }, + { + "bbox": [ + 130, + 215, + 482, + 258 + ], + "type": "text", + "content": " for small values of DPT. (b) The left image shows the points selected in the refined estimation. The right shows the fitting of the estimation." + } + ] + } + ], + "index": 7, + "angle": 0, + "type": "image_caption" + } + ], + "index": 4 + }, + { + "type": "image", + "bbox": [ + 375, + 120, + 462, + 195 + ], + "blocks": [ + { + "bbox": [ + 375, + 120, + 462, + 195 + ], + "lines": [ + { + "bbox": [ + 375, + 120, + 462, + 195 + ], + "spans": [ + { + "bbox": [ + 375, + 120, + 462, + 195 + ], + "type": "image", + "image_path": "281a5e27ef6c7031f29527ea0313d3ce9de8ecf271bb56aaa1fcbf974be952bf.jpg" + } + ] + } + ], + "index": 6, + "angle": 0, + "type": "image_body" + } + ], + "index": 6 + }, + { + "bbox": [ + 132, + 281, + 202, + 295 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 281, + 202, + 295 + ], + "spans": [ + { + "bbox": [ + 132, + 281, + 202, + 295 + ], + "type": "text", + "content": "4 Method" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 132, + 308, + 268, + 319 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 308, + 268, + 319 + ], + "spans": [ + { + "bbox": [ + 132, + 308, + 268, + 319 + ], + "type": "text", + "content": "4.1 Framework Overview" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 130, + 330, + 482, + 485 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 330, + 482, + 485 + ], + "spans": [ + { + "bbox": [ + 130, + 330, + 482, + 485 + ], + "type": "text", + "content": "Inspired by the observations in Sec. 3, we propose a physics-informed knowledge transfer method for underwater monocular depth estimation, named the Physics-informed Underwater Depth Estimation (PUDE). The main idea of this physics-informed method is to leverage the physical information in underwater imagery to guide knowledge transfer within the neural network, thereby enhancing its adaptation from in-air to underwater environments. Specifically, we adopt the pre-trained DPT-Hybrid model [28], a state-of-the-art (SOTA) model known for its good generalization, as the transferred model. Benefiting from the massive in-air training dataset, the DPT model has powerful spatial structural knowledge. Despite previously noted issues when applied to underwater images, it can still provide a relatively reasonable depth estimation, particularly for regions close to the camera. This estimation serves as our prior depth knowledge for extracting the physical information from the underwater images." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 130, + 486, + 482, + 594 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 486, + 482, + 594 + ], + "spans": [ + { + "bbox": [ + 130, + 486, + 482, + 594 + ], + "type": "text", + "content": "Most daily images are non-linear images, which are automatically post-processed by nonlinear algorithms integrated into cameras, for example, gamma correction, to approach the non-linear perception of human eyes. By contrast, linear images are images in which pixel values are directly captured by the camera without nonlinear post-processing or only with ratio processing. To ensure the PUDE model's generalizability in reality and keep consistency in training and testing, we utilize post-processed non-linear underwater images as inputs. In contrast, for precise physical information extraction, we utilize the linear underwater images in the estimation of the physics parameters in Eq. (1) and Eq. (3)." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 130, + 594, + 482, + 666 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 594, + 482, + 666 + ], + "spans": [ + { + "bbox": [ + 130, + 594, + 482, + 666 + ], + "type": "text", + "content": "Our training framework consists of two branches shown in Fig. 3. The first branch is the direct knowledge transfer from DPT to PUDE, and we employ a similarity loss to guarantee the similarity between DPT and PUDE in the transfer (in Sec. 4.2). The second branch is the physics-informed knowledge transfer. Inspired by the parameter estimation method in Sea-thru [1], we estimate the parameters in Eq. (1) and Eq. (3) for each input underwater image, utilizing" + } + ] + } + ], + "index": 12 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "spans": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "type": "text", + "content": "6" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 166, + 91, + 222, + 102 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 166, + 91, + 222, + 102 + ], + "spans": [ + { + "bbox": [ + 166, + 91, + 222, + 102 + ], + "type": "text", + "content": "J. Yang et al." + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 5 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 136, + 115, + 390, + 244 + ], + "blocks": [ + { + "bbox": [ + 136, + 115, + 390, + 244 + ], + "lines": [ + { + "bbox": [ + 136, + 115, + 390, + 244 + ], + "spans": [ + { + "bbox": [ + 136, + 115, + 390, + 244 + ], + "type": "image", + "image_path": "91205bbd54b8d2398f71974089efaf96b0e84d1b00c1a1d0c8f78f9fc6ea4de1.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 130, + 254, + 482, + 299 + ], + "lines": [ + { + "bbox": [ + 130, + 254, + 482, + 299 + ], + "spans": [ + { + "bbox": [ + 130, + 254, + 482, + 299 + ], + "type": "text", + "content": "Fig. 3: Overall training framework for PUDE. The overall training framework comprises two components: the direct knowledge transfer from the DPT model and the physi-informed knowledge transfer based on the DPT estimation. Physical information is extracted from DPT's output to establish the bound losses." + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "image_caption" + } + ], + "index": 2 + }, + { + "type": "image", + "bbox": [ + 425, + 116, + 477, + 236 + ], + "blocks": [ + { + "bbox": [ + 425, + 116, + 477, + 236 + ], + "lines": [ + { + "bbox": [ + 425, + 116, + 477, + 236 + ], + "spans": [ + { + "bbox": [ + 425, + 116, + 477, + 236 + ], + "type": "image", + "image_path": "f3287d058a9034e99873dc4a7b102847501a7b47962655d1620a4bcf38b317ed.jpg" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_body" + } + ], + "index": 3 + }, + { + "bbox": [ + 130, + 319, + 479, + 392 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 319, + 479, + 392 + ], + "spans": [ + { + "bbox": [ + 130, + 319, + 479, + 392 + ], + "type": "text", + "content": "DPT depth estimation. We restrict our usage to the green and blue channels, in line with the conclusions from UDCP [10]. Based on that, we propose two novel bound losses, utilizing the physics information contained in the underwater image formation model in Eq. (1) to enforce the model to produce plausible depth estimations that conform to the physics characteristics of underwater imaging. In Sec. 4.3, we will introduce the second branch and the designed bound losses." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 132, + 406, + 296, + 418 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 406, + 296, + 418 + ], + "spans": [ + { + "bbox": [ + 132, + 406, + 296, + 418 + ], + "type": "text", + "content": "4.2 Direct Knowledge Transfer" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 130, + 423, + 482, + 533 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 423, + 482, + 533 + ], + "spans": [ + { + "bbox": [ + 130, + 423, + 482, + 533 + ], + "type": "text", + "content": "The direct knowledge transfer loop aims to guarantee the similarity between the original DPT and PUDE during the transition to ensure the structural knowledge can be kept in PUDE. We use a similarity loss in the training process to achieve this goal. In the transfer, PUDE is initialized with DPT. We trim the pixels with the largest differences to set the tolerance for the difference between PUDE and the original model DPT during the training. We denote the pixel set after trimming as " + }, + { + "bbox": [ + 130, + 423, + 482, + 533 + ], + "type": "inline_equation", + "content": "M_{trim} \\subseteq \\{1, \\dots, HW\\}" + }, + { + "bbox": [ + 130, + 423, + 482, + 533 + ], + "type": "text", + "content": ", and " + }, + { + "bbox": [ + 130, + 423, + 482, + 533 + ], + "type": "inline_equation", + "content": "T" + }, + { + "bbox": [ + 130, + 423, + 482, + 533 + ], + "type": "text", + "content": " representing the cardinality of " + }, + { + "bbox": [ + 130, + 423, + 482, + 533 + ], + "type": "inline_equation", + "content": "M_{trim}" + }, + { + "bbox": [ + 130, + 423, + 482, + 533 + ], + "type": "text", + "content": ". For each pixel " + }, + { + "bbox": [ + 130, + 423, + 482, + 533 + ], + "type": "inline_equation", + "content": "i \\in M_{trim}" + }, + { + "bbox": [ + 130, + 423, + 482, + 533 + ], + "type": "text", + "content": ", we denote the DPT and PUDE estimation as " + }, + { + "bbox": [ + 130, + 423, + 482, + 533 + ], + "type": "inline_equation", + "content": "\\mathbf{d}_i^D" + }, + { + "bbox": [ + 130, + 423, + 482, + 533 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 130, + 423, + 482, + 533 + ], + "type": "inline_equation", + "content": "\\mathbf{d}_i^P" + }, + { + "bbox": [ + 130, + 423, + 482, + 533 + ], + "type": "text", + "content": ", respectively. The similarity loss " + }, + { + "bbox": [ + 130, + 423, + 482, + 533 + ], + "type": "inline_equation", + "content": "L_s" + }, + { + "bbox": [ + 130, + 423, + 482, + 533 + ], + "type": "text", + "content": " is defined as:" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 238, + 539, + 481, + 571 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 238, + 539, + 481, + 571 + ], + "spans": [ + { + "bbox": [ + 238, + 539, + 481, + 571 + ], + "type": "interline_equation", + "content": "L _ {s} = \\frac {1}{T} \\sum_ {i \\in M _ {\\mathrm {t r i m}}} \\left| \\frac {\\mathbf {d} _ {i} ^ {P} - \\mathbf {d} _ {i} ^ {D}}{\\operatorname {m e d i a n} (\\mathbf {d} ^ {D})} \\right|. \\tag {4}", + "image_path": "98a3f289eeb8e15a437eef26cdeebeda3523f5df3b8795dec56229e605adbb9a.jpg" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 130, + 573, + 482, + 609 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 573, + 482, + 609 + ], + "spans": [ + { + "bbox": [ + 130, + 573, + 482, + 609 + ], + "type": "text", + "content": "The trimming sets the tolerance for the changing between PUDE and the original model DPT during the training. In our implementation, we set the trimming rate to be " + }, + { + "bbox": [ + 130, + 573, + 482, + 609 + ], + "type": "inline_equation", + "content": "0.3HW" + }, + { + "bbox": [ + 130, + 573, + 482, + 609 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 132, + 624, + 350, + 636 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 624, + 350, + 636 + ], + "spans": [ + { + "bbox": [ + 132, + 624, + 350, + 636 + ], + "type": "text", + "content": "4.3 Physics-informed Knowledge Transfer" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 130, + 641, + 481, + 666 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 641, + 481, + 666 + ], + "spans": [ + { + "bbox": [ + 130, + 641, + 481, + 666 + ], + "type": "text", + "content": "Parameters Estimation: In the physics-informed knowledge transfer loop, we first extract the parameters in Eq. (1) and Eq. (3) using the preliminary depth" + } + ] + } + ], + "index": 11 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 419, + 91, + 447, + 100 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 419, + 91, + 447, + 100 + ], + "spans": [ + { + "bbox": [ + 419, + 91, + 447, + 100 + ], + "type": "text", + "content": "PUDE" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 474, + 92, + 481, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 474, + 92, + 481, + 100 + ], + "spans": [ + { + "bbox": [ + 474, + 92, + 481, + 100 + ], + "type": "text", + "content": "7" + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 6 + }, + { + "para_blocks": [ + { + "bbox": [ + 130, + 116, + 479, + 176 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 116, + 479, + 176 + ], + "spans": [ + { + "bbox": [ + 130, + 116, + 479, + 176 + ], + "type": "text", + "content": "estimations from DPT. As shown in Fig. 1, DPT's depth estimations on the underwater images tend to overestimate the depth in distant and blurry regions but offer more trustable estimation for the clearer regions, showing a good linear relationship described in Eq. (3). Therefore, we extract the physics information harnessed from DPT's depth estimations for clear regions." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 130, + 176, + 482, + 344 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 176, + 482, + 344 + ], + "spans": [ + { + "bbox": [ + 130, + 176, + 482, + 344 + ], + "type": "text", + "content": "The estimation process utilises the DPT depth estimation (on nonlinear post-processed images) and the linear images, consisting of two steps, the rough and refined estimation. The first step is the rough estimation using the whole image to obtain a robust estimation of the background light " + }, + { + "bbox": [ + 130, + 176, + 482, + 344 + ], + "type": "inline_equation", + "content": "\\hat{B}_c^\\infty" + }, + { + "bbox": [ + 130, + 176, + 482, + 344 + ], + "type": "text", + "content": ". The rationale for extracting the darkest points in the Sea-thru method is analogous to identifying points where " + }, + { + "bbox": [ + 130, + 176, + 482, + 344 + ], + "type": "inline_equation", + "content": "J_{c,i}" + }, + { + "bbox": [ + 130, + 176, + 482, + 344 + ], + "type": "text", + "content": " is assumed to be close to zero and where the color appearance on these pixels is solely due to the backscattering. Therefore, we evenly select the " + }, + { + "bbox": [ + 130, + 176, + 482, + 344 + ], + "type": "inline_equation", + "content": "N_1" + }, + { + "bbox": [ + 130, + 176, + 482, + 344 + ], + "type": "text", + "content": " darkest points, on which " + }, + { + "bbox": [ + 130, + 176, + 482, + 344 + ], + "type": "inline_equation", + "content": "J_{c,i} \\approx 0" + }, + { + "bbox": [ + 130, + 176, + 482, + 344 + ], + "type": "text", + "content": ", from the whole image and employ these points to estimate the water parameters by solving the least-square problem in Eq. (5). We denote the pixel value on the pixel location " + }, + { + "bbox": [ + 130, + 176, + 482, + 344 + ], + "type": "inline_equation", + "content": "i" + }, + { + "bbox": [ + 130, + 176, + 482, + 344 + ], + "type": "text", + "content": " of the color channel " + }, + { + "bbox": [ + 130, + 176, + 482, + 344 + ], + "type": "inline_equation", + "content": "c" + }, + { + "bbox": [ + 130, + 176, + 482, + 344 + ], + "type": "text", + "content": " on the input underwater image as " + }, + { + "bbox": [ + 130, + 176, + 482, + 344 + ], + "type": "inline_equation", + "content": "I_{c,i}" + }, + { + "bbox": [ + 130, + 176, + 482, + 344 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 130, + 176, + 482, + 344 + ], + "type": "inline_equation", + "content": "i = 1 \\dots HW" + }, + { + "bbox": [ + 130, + 176, + 482, + 344 + ], + "type": "text", + "content": ". According to the suggestions in UDCP [10], we only use blue and green channels in this process. We denote the selected point set of each color channel as " + }, + { + "bbox": [ + 130, + 176, + 482, + 344 + ], + "type": "inline_equation", + "content": "M_c" + }, + { + "bbox": [ + 130, + 176, + 482, + 344 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 130, + 176, + 482, + 344 + ], + "type": "inline_equation", + "content": "c \\in \\{G, B\\}" + }, + { + "bbox": [ + 130, + 176, + 482, + 344 + ], + "type": "text", + "content": ". Substituting (3) into (1), we define the error " + }, + { + "bbox": [ + 130, + 176, + 482, + 344 + ], + "type": "inline_equation", + "content": "\\epsilon_c" + }, + { + "bbox": [ + 130, + 176, + 482, + 344 + ], + "type": "text", + "content": " for the color channels as:" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 220, + 350, + 481, + 380 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 220, + 350, + 481, + 380 + ], + "spans": [ + { + "bbox": [ + 220, + 350, + 481, + 380 + ], + "type": "interline_equation", + "content": "\\epsilon_ {c} := \\sum_ {i \\in M _ {c}} | | I _ {c, i} - B _ {c} ^ {\\infty} (1 - e ^ {- \\frac {\\beta_ {c} / s}{\\mathbf {d} _ {i} ^ {D + h / s}}}) | | _ {2} ^ {2}. \\tag {5}", + "image_path": "dd181de930aeaa5f7470d7ef8b1f34f71e87bfd67c368d7b3bce5c5a12a53623.jpg" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 131, + 387, + 482, + 422 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 131, + 387, + 482, + 422 + ], + "spans": [ + { + "bbox": [ + 131, + 387, + 482, + 422 + ], + "type": "text", + "content": "To simplify the least-square problem, we set two temporary parameters " + }, + { + "bbox": [ + 131, + 387, + 482, + 422 + ], + "type": "inline_equation", + "content": "\\nu_{c} = \\beta_{c} / s" + }, + { + "bbox": [ + 131, + 387, + 482, + 422 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 131, + 387, + 482, + 422 + ], + "type": "inline_equation", + "content": "\\mu = h / s" + }, + { + "bbox": [ + 131, + 387, + 482, + 422 + ], + "type": "text", + "content": ". Then, we solve the following least-square problem to obtain the estimate of the parameters:" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 220, + 429, + 481, + 456 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 220, + 429, + 481, + 456 + ], + "spans": [ + { + "bbox": [ + 220, + 429, + 481, + 456 + ], + "type": "interline_equation", + "content": "\\left(\\hat {\\nu} _ {G}, \\hat {\\nu} _ {B}, \\hat {\\mu}, \\hat {B} _ {c} ^ {\\infty}\\right) = \\operatorname * {a r g m i n} _ {\\nu_ {G}, \\nu_ {B}, \\mu , B _ {c} ^ {\\infty}} \\sum_ {c \\in \\{G, B \\}} \\epsilon_ {c}, \\tag {6}", + "image_path": "51242f130053e62804f7d6f2f170e7184fe429cb8f8c9aa96e96c98c5b3043bc.jpg" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 131, + 462, + 482, + 584 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 131, + 462, + 482, + 584 + ], + "spans": [ + { + "bbox": [ + 131, + 462, + 482, + 584 + ], + "type": "text", + "content": "Utilizing the preliminary estimates " + }, + { + "bbox": [ + 131, + 462, + 482, + 584 + ], + "type": "inline_equation", + "content": "\\hat{\\nu}_G" + }, + { + "bbox": [ + 131, + 462, + 482, + 584 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 131, + 462, + 482, + 584 + ], + "type": "inline_equation", + "content": "\\hat{\\nu}_B" + }, + { + "bbox": [ + 131, + 462, + 482, + 584 + ], + "type": "text", + "content": ", and " + }, + { + "bbox": [ + 131, + 462, + 482, + 584 + ], + "type": "inline_equation", + "content": "\\hat{\\mu}" + }, + { + "bbox": [ + 131, + 462, + 482, + 584 + ], + "type": "text", + "content": ", we calculate the medium transmission rate " + }, + { + "bbox": [ + 131, + 462, + 482, + 584 + ], + "type": "inline_equation", + "content": "t_{c,i}" + }, + { + "bbox": [ + 131, + 462, + 482, + 584 + ], + "type": "text", + "content": " in Eq. (2) for each pixel. Observations from Fig. 2a suggest that depth overestimation of DPT primarily occurs in regions with low transmission rates " + }, + { + "bbox": [ + 131, + 462, + 482, + 584 + ], + "type": "inline_equation", + "content": "t_i" + }, + { + "bbox": [ + 131, + 462, + 482, + 584 + ], + "type": "text", + "content": ". Therefore, in our implementation, in the refined estimation step, we exclude the regions where DPT's estimation is inaccurate while maximizing the utilization of the largest point sampling regions, in which the transmission rate is smaller than 0.2. Similar to rough estimation, in refined estimation, we select " + }, + { + "bbox": [ + 131, + 462, + 482, + 584 + ], + "type": "inline_equation", + "content": "N_2" + }, + { + "bbox": [ + 131, + 462, + 482, + 584 + ], + "type": "text", + "content": " darkest points from the selected regions to construct the set " + }, + { + "bbox": [ + 131, + 462, + 482, + 584 + ], + "type": "inline_equation", + "content": "M_{c2}" + }, + { + "bbox": [ + 131, + 462, + 482, + 584 + ], + "type": "text", + "content": " and address a least-square problem. However, we streamline the process by utilizing the estimated " + }, + { + "bbox": [ + 131, + 462, + 482, + 584 + ], + "type": "inline_equation", + "content": "\\hat{B}_c^\\infty" + }, + { + "bbox": [ + 131, + 462, + 482, + 584 + ], + "type": "text", + "content": " in the first step. The error " + }, + { + "bbox": [ + 131, + 462, + 482, + 584 + ], + "type": "inline_equation", + "content": "\\epsilon_{c2}" + }, + { + "bbox": [ + 131, + 462, + 482, + 584 + ], + "type": "text", + "content": " in this is:" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 220, + 588, + 481, + 617 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 220, + 588, + 481, + 617 + ], + "spans": [ + { + "bbox": [ + 220, + 588, + 481, + 617 + ], + "type": "interline_equation", + "content": "\\epsilon_ {c 2} := \\sum_ {i \\in M _ {c 2}} | | I _ {c, i} - \\hat {B} _ {c} ^ {\\infty} (1 - e ^ {- \\frac {\\nu_ {c}}{\\mathbf {d} _ {i} ^ {D} + \\mu}}) | | _ {2} ^ {2}, \\tag {7}", + "image_path": "00a7f9b49e6325aad094680cb7c560cf09ac7b5075b2f4b63ad19282c5e198ef.jpg" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 132, + 623, + 276, + 635 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 623, + 276, + 635 + ], + "spans": [ + { + "bbox": [ + 132, + 623, + 276, + 635 + ], + "type": "text", + "content": "and the least-square problem is:" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 237, + 641, + 481, + 668 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 237, + 641, + 481, + 668 + ], + "spans": [ + { + "bbox": [ + 237, + 641, + 481, + 668 + ], + "type": "interline_equation", + "content": "\\left(\\hat {\\nu} _ {G}, \\hat {\\nu} _ {B}, \\hat {\\mu}\\right) = \\operatorname * {a r g m i n} _ {\\nu_ {G}, \\nu_ {B}, \\mu} \\sum_ {c \\in \\{G, B \\}} \\epsilon_ {c 2}, \\tag {8}", + "image_path": "7d2e0dcae2a8cecb54d89bada175b3ba37cff0141a4b81481ab67388978f988a.jpg" + } + ] + } + ], + "index": 10 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "spans": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "type": "text", + "content": "8" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 166, + 91, + 222, + 102 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 166, + 91, + 222, + 102 + ], + "spans": [ + { + "bbox": [ + 166, + 91, + 222, + 102 + ], + "type": "text", + "content": "J. Yang et al." + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 7 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 134, + 114, + 203, + 162 + ], + "blocks": [ + { + "bbox": [ + 134, + 114, + 203, + 162 + ], + "lines": [ + { + "bbox": [ + 134, + 114, + 203, + 162 + ], + "spans": [ + { + "bbox": [ + 134, + 114, + 203, + 162 + ], + "type": "image", + "image_path": "ccc1a4dce5fdff632fe807cc9289c1ca417ab43da08a1397123731bf460308d7.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 153, + 163, + 185, + 172 + ], + "lines": [ + { + "bbox": [ + 153, + 163, + 185, + 172 + ], + "spans": [ + { + "bbox": [ + 153, + 163, + 185, + 172 + ], + "type": "text", + "content": "(a) RGB" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_caption" + } + ], + "index": 2 + }, + { + "type": "image", + "bbox": [ + 205, + 114, + 273, + 161 + ], + "blocks": [ + { + "bbox": [ + 205, + 114, + 273, + 161 + ], + "lines": [ + { + "bbox": [ + 205, + 114, + 273, + 161 + ], + "spans": [ + { + "bbox": [ + 205, + 114, + 273, + 161 + ], + "type": "image", + "image_path": "a40e9c9e94a086eca498a5972348c4096cd5cfb4a788dfeaaa12c7af29242741.jpg" + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 211, + 163, + 266, + 172 + ], + "lines": [ + { + "bbox": [ + 211, + 163, + 266, + 172 + ], + "spans": [ + { + "bbox": [ + 211, + 163, + 266, + 172 + ], + "type": "text", + "content": "(b) DPT depth" + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 131, + 181, + 479, + 215 + ], + "lines": [ + { + "bbox": [ + 131, + 181, + 479, + 215 + ], + "spans": [ + { + "bbox": [ + 131, + 181, + 479, + 215 + ], + "type": "text", + "content": "Fig. 4: Loss visualization: Depth estimation from DPT loses details due to blurriness. However, our designed bound losses have differences across pixels according to the pixel intensity thereby enhancing the depth estimation details in training." + } + ] + } + ], + "index": 10, + "angle": 0, + "type": "image_caption" + } + ], + "index": 4 + }, + { + "type": "image", + "bbox": [ + 276, + 114, + 410, + 161 + ], + "blocks": [ + { + "bbox": [ + 276, + 114, + 410, + 161 + ], + "lines": [ + { + "bbox": [ + 276, + 114, + 410, + 161 + ], + "spans": [ + { + "bbox": [ + 276, + 114, + 410, + 161 + ], + "type": "image", + "image_path": "a6f11881cbb6b6c58935904100e80e8fe25e1cad02ffb487dfa204a5bf6bbdd2.jpg" + } + ] + } + ], + "index": 6, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 293, + 163, + 321, + 172 + ], + "lines": [ + { + "bbox": [ + 293, + 163, + 321, + 172 + ], + "spans": [ + { + "bbox": [ + 293, + 163, + 321, + 172 + ], + "type": "text", + "content": "(c) " + }, + { + "bbox": [ + 293, + 163, + 321, + 172 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_{bu}" + } + ] + } + ], + "index": 7, + "angle": 0, + "type": "image_caption" + } + ], + "index": 6 + }, + { + "type": "image", + "bbox": [ + 410, + 114, + 479, + 161 + ], + "blocks": [ + { + "bbox": [ + 410, + 114, + 479, + 161 + ], + "lines": [ + { + "bbox": [ + 410, + 114, + 479, + 161 + ], + "spans": [ + { + "bbox": [ + 410, + 114, + 479, + 161 + ], + "type": "image", + "image_path": "81888be583b5065b0ca2874160e14a97bdb7f8193d74836d9d9564dc07707b95.jpg" + } + ] + } + ], + "index": 8, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 414, + 163, + 474, + 172 + ], + "lines": [ + { + "bbox": [ + 414, + 163, + 474, + 172 + ], + "spans": [ + { + "bbox": [ + 414, + 163, + 474, + 172 + ], + "type": "text", + "content": "(e) local patches" + } + ] + } + ], + "index": 9, + "angle": 0, + "type": "image_caption" + } + ], + "index": 8 + }, + { + "bbox": [ + 131, + 236, + 479, + 260 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 131, + 236, + 479, + 260 + ], + "spans": [ + { + "bbox": [ + 131, + 236, + 479, + 260 + ], + "type": "text", + "content": "In our implementation, " + }, + { + "bbox": [ + 131, + 236, + 479, + 260 + ], + "type": "inline_equation", + "content": "N_{1}" + }, + { + "bbox": [ + 131, + 236, + 479, + 260 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 131, + 236, + 479, + 260 + ], + "type": "inline_equation", + "content": "N_{2}" + }, + { + "bbox": [ + 131, + 236, + 479, + 260 + ], + "type": "text", + "content": " is set to be 500 and 200 respectively. An implementation example of the refined estimation is shown in Fig. 2b." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 131, + 274, + 479, + 321 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 131, + 274, + 479, + 321 + ], + "spans": [ + { + "bbox": [ + 131, + 274, + 479, + 321 + ], + "type": "text", + "content": "Bound Losses on Medium Transmission Rate: After estimating the parameters of the underwater imaging formation model in Eq. (1), we are ready to present two bound losses to constrain the PUDE model from predicting depths that contravene the underwater characteristics." + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 131, + 323, + 480, + 396 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 131, + 323, + 480, + 396 + ], + "spans": [ + { + "bbox": [ + 131, + 323, + 480, + 396 + ], + "type": "text", + "content": "Lower bound: The first bound is a lower bound on the estimated medium transmission rate " + }, + { + "bbox": [ + 131, + 323, + 480, + 396 + ], + "type": "inline_equation", + "content": "t_{c,i}^{P}" + }, + { + "bbox": [ + 131, + 323, + 480, + 396 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 131, + 323, + 480, + 396 + ], + "type": "inline_equation", + "content": "c \\in \\{G, B\\}" + }, + { + "bbox": [ + 131, + 323, + 480, + 396 + ], + "type": "text", + "content": ". This bound aids in refining depth estimates in distant regions, preventing from overestimating the depth for the regions that are far from the camera, as described in O1 in Sec. 3. By substituting the obtained parameter " + }, + { + "bbox": [ + 131, + 323, + 480, + 396 + ], + "type": "inline_equation", + "content": "\\hat{\\nu}_{G}, \\hat{\\nu}_{B}" + }, + { + "bbox": [ + 131, + 323, + 480, + 396 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 131, + 323, + 480, + 396 + ], + "type": "inline_equation", + "content": "\\hat{\\mu}" + }, + { + "bbox": [ + 131, + 323, + 480, + 396 + ], + "type": "text", + "content": " in the parameters estimation step, and using the PUDE depth estimation, we get the estimated " + }, + { + "bbox": [ + 131, + 323, + 480, + 396 + ], + "type": "inline_equation", + "content": "\\hat{t}_{c,i}^{P}" + }, + { + "bbox": [ + 131, + 323, + 480, + 396 + ], + "type": "text", + "content": ":" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 273, + 402, + 480, + 422 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 273, + 402, + 480, + 422 + ], + "spans": [ + { + "bbox": [ + 273, + 402, + 480, + 422 + ], + "type": "interline_equation", + "content": "\\hat {t} _ {c, i} ^ {P} = e ^ {- \\frac {\\hat {\\nu} _ {c}}{\\mathbf {d} _ {i} ^ {P} + \\hat {\\mu}}}. \\tag {9}", + "image_path": "2f06a0b03c7a7f83512922c9704fe3ac69e492534e1c5bf2cfaa9e38ab7c7458.jpg" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 131, + 426, + 479, + 450 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 131, + 426, + 479, + 450 + ], + "spans": [ + { + "bbox": [ + 131, + 426, + 479, + 450 + ], + "type": "text", + "content": "As the original color value of each point " + }, + { + "bbox": [ + 131, + 426, + 479, + 450 + ], + "type": "inline_equation", + "content": "J_{c,i}" + }, + { + "bbox": [ + 131, + 426, + 479, + 450 + ], + "type": "text", + "content": " is equal to or greater than zero. Based on Eq. (1), we have:" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 263, + 456, + 480, + 469 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 263, + 456, + 480, + 469 + ], + "spans": [ + { + "bbox": [ + 263, + 456, + 480, + 469 + ], + "type": "interline_equation", + "content": "I _ {c, i} \\geq \\left(1 - t _ {c, i}\\right) B _ {c} ^ {\\infty}, \\tag {10}", + "image_path": "70f08701f69d61718d47974d55b2c016ca4101cda00bd0d083f982ecee7167c8.jpg" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 131, + 476, + 379, + 488 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 131, + 476, + 379, + 488 + ], + "spans": [ + { + "bbox": [ + 131, + 476, + 379, + 488 + ], + "type": "text", + "content": "substituting the estimated " + }, + { + "bbox": [ + 131, + 476, + 379, + 488 + ], + "type": "inline_equation", + "content": "\\hat{B}_c^\\infty" + }, + { + "bbox": [ + 131, + 476, + 379, + 488 + ], + "type": "text", + "content": ", the first lower bound is:" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 272, + 494, + 480, + 520 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 272, + 494, + 480, + 520 + ], + "spans": [ + { + "bbox": [ + 272, + 494, + 480, + 520 + ], + "type": "interline_equation", + "content": "\\hat {t} _ {c, i} ^ {P} \\geq 1 - \\frac {I _ {c , i}}{\\hat {B} _ {c} ^ {\\infty}}. \\tag {11}", + "image_path": "3e6be49b97367a5b3183f1b563f47688a3f63a44bac9251a7819018696beb9d3.jpg" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 131, + 526, + 479, + 586 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 131, + 526, + 479, + 586 + ], + "spans": [ + { + "bbox": [ + 131, + 526, + 479, + 586 + ], + "type": "text", + "content": "This lower bound on " + }, + { + "bbox": [ + 131, + 526, + 479, + 586 + ], + "type": "inline_equation", + "content": "\\hat{t}_{c,i}^{P}" + }, + { + "bbox": [ + 131, + 526, + 479, + 586 + ], + "type": "text", + "content": " is to pull back the depth estimations for the distant regions. When the depth of pixels in these regions is overestimated, it results in an underestimation of the medium transmission rate in Eq. (9). Consequently, the captured pixel value " + }, + { + "bbox": [ + 131, + 526, + 479, + 586 + ], + "type": "inline_equation", + "content": "I_{c,i}" + }, + { + "bbox": [ + 131, + 526, + 479, + 586 + ], + "type": "text", + "content": " should not be lower than the estimated backscattering, as shown in Eq. (10)." + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 131, + 587, + 479, + 622 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 131, + 587, + 479, + 622 + ], + "spans": [ + { + "bbox": [ + 131, + 587, + 479, + 622 + ], + "type": "text", + "content": "Upper bound: The second bound is an upper bound on the estimated medium transmission rate " + }, + { + "bbox": [ + 131, + 587, + 479, + 622 + ], + "type": "inline_equation", + "content": "t_{c,i}^{P}, c \\in \\{G, B\\}" + }, + { + "bbox": [ + 131, + 587, + 479, + 622 + ], + "type": "text", + "content": " to address the challenge of detecting background edges in blurry underwater images, described in O2." + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 131, + 623, + 479, + 647 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 131, + 623, + 479, + 647 + ], + "spans": [ + { + "bbox": [ + 131, + 623, + 479, + 647 + ], + "type": "text", + "content": "Considering that the maximum pixel intensity should not surpass that of pure white, namely " + }, + { + "bbox": [ + 131, + 623, + 479, + 647 + ], + "type": "inline_equation", + "content": "J_{c,i} \\leq 1" + }, + { + "bbox": [ + 131, + 623, + 479, + 647 + ], + "type": "text", + "content": ", and referring to Eq. (1), we obtain:" + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 250, + 653, + 480, + 666 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 250, + 653, + 480, + 666 + ], + "spans": [ + { + "bbox": [ + 250, + 653, + 480, + 666 + ], + "type": "interline_equation", + "content": "I _ {c, i} \\leq t _ {c, i} + \\left(1 - t _ {c, i}\\right) B _ {c} ^ {\\infty}. \\tag {12}", + "image_path": "10e252f2bdb094cb72a6895d9fefe8dc32d822126937a164b89207f8a80c5167.jpg" + } + ] + } + ], + "index": 22 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 419, + 91, + 447, + 100 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 419, + 91, + 447, + 100 + ], + "spans": [ + { + "bbox": [ + 419, + 91, + 447, + 100 + ], + "type": "text", + "content": "PUDE" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 474, + 91, + 481, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 474, + 91, + 481, + 100 + ], + "spans": [ + { + "bbox": [ + 474, + 91, + 481, + 100 + ], + "type": "text", + "content": "9" + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 8 + }, + { + "para_blocks": [ + { + "bbox": [ + 130, + 116, + 482, + 177 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 116, + 482, + 177 + ], + "spans": [ + { + "bbox": [ + 130, + 116, + 482, + 177 + ], + "type": "text", + "content": "In regions corresponding to the pure background area, the distance can be considered infinite. The depth estimation by DPT for these regions is always finite. Therefore, the estimated medium transmission for these background regions should be greater than the true value. We denote the estimated medium transmission rate at each pixel location by DPT as " + }, + { + "bbox": [ + 130, + 116, + 482, + 177 + ], + "type": "inline_equation", + "content": "\\hat{t}_{c,i}^{D}" + }, + { + "bbox": [ + 130, + 116, + 482, + 177 + ], + "type": "text", + "content": ", which is defined as:" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 276, + 189, + 481, + 208 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 276, + 189, + 481, + 208 + ], + "spans": [ + { + "bbox": [ + 276, + 189, + 481, + 208 + ], + "type": "interline_equation", + "content": "\\hat {t} _ {c, i} ^ {D} = e ^ {- \\frac {\\bar {\\nu} _ {c}}{\\mathbf {d} _ {i} ^ {D} + \\bar {\\mu}}}. \\tag {13}", + "image_path": "e75e2b0140e99005673c07dc41e8f2e4f95b1fa460ae8b7c9fb8b478bbc63bd4.jpg" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 131, + 216, + 482, + 253 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 131, + 216, + 482, + 253 + ], + "spans": [ + { + "bbox": [ + 131, + 216, + 482, + 253 + ], + "type": "text", + "content": "In distant regions, where the original pixel values " + }, + { + "bbox": [ + 131, + 216, + 482, + 253 + ], + "type": "inline_equation", + "content": "J_{c,i}" + }, + { + "bbox": [ + 131, + 216, + 482, + 253 + ], + "type": "text", + "content": " are attenuated, the captured pixel values " + }, + { + "bbox": [ + 131, + 216, + 482, + 253 + ], + "type": "inline_equation", + "content": "I_{c,i}" + }, + { + "bbox": [ + 131, + 216, + 482, + 253 + ], + "type": "text", + "content": " are predominantly dictated by backscattering. To select pixel locations likely to be in the background regions, we define the set " + }, + { + "bbox": [ + 131, + 216, + 482, + 253 + ], + "type": "inline_equation", + "content": "M_b" + }, + { + "bbox": [ + 131, + 216, + 482, + 253 + ], + "type": "text", + "content": ":" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 216, + 262, + 481, + 291 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 216, + 262, + 481, + 291 + ], + "spans": [ + { + "bbox": [ + 216, + 262, + 481, + 291 + ], + "type": "interline_equation", + "content": "M _ {b} = \\{i \\mid \\forall c \\in \\{G, B \\}, \\frac {\\hat {B} _ {c} ^ {\\infty} \\left(1 - \\hat {t} _ {c , i} ^ {D}\\right)}{I _ {c , i}} \\geq \\gamma \\}. \\tag {14}", + "image_path": "c6ce6cf7ae3bff4a0a9a0111e8a6dced6bc5001e0c68b08465840c380e2b09e2.jpg" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 131, + 299, + 482, + 348 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 131, + 299, + 482, + 348 + ], + "spans": [ + { + "bbox": [ + 131, + 299, + 482, + 348 + ], + "type": "text", + "content": "The parameter " + }, + { + "bbox": [ + 131, + 299, + 482, + 348 + ], + "type": "inline_equation", + "content": "\\gamma" + }, + { + "bbox": [ + 131, + 299, + 482, + 348 + ], + "type": "text", + "content": " is utilized to identify regions where pixel values are predominantly determined by backscattering, which are likely to be in the background area. In our experiments, " + }, + { + "bbox": [ + 131, + 299, + 482, + 348 + ], + "type": "inline_equation", + "content": "\\gamma" + }, + { + "bbox": [ + 131, + 299, + 482, + 348 + ], + "type": "text", + "content": " is set to be 0.6. As mentioned, for the background regions, " + }, + { + "bbox": [ + 131, + 299, + 482, + 348 + ], + "type": "inline_equation", + "content": "\\hat{t}_{c,i}^{D}" + }, + { + "bbox": [ + 131, + 299, + 482, + 348 + ], + "type": "text", + "content": " is greater than the true value. Then based on the Eq. (12), we have," + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 231, + 358, + 481, + 373 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 231, + 358, + 481, + 373 + ], + "spans": [ + { + "bbox": [ + 231, + 358, + 481, + 373 + ], + "type": "interline_equation", + "content": "I _ {c, i} < \\hat {t} _ {c, i} ^ {D} + \\left(1 - t _ {c, i}\\right) B _ {c} ^ {\\infty}, \\quad i \\in M _ {b}, \\tag {15}", + "image_path": "ed3b2cd3b4597da4e9ca0f638ef12c3f68c897fb1c8253fd0bf6af0a0f6b62ba.jpg" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 131, + 377, + 394, + 389 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 131, + 377, + 394, + 389 + ], + "spans": [ + { + "bbox": [ + 131, + 377, + 394, + 389 + ], + "type": "text", + "content": "substituting the estimated parameters, we have the bound:" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 231, + 399, + 481, + 430 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 231, + 399, + 481, + 430 + ], + "spans": [ + { + "bbox": [ + 231, + 399, + 481, + 430 + ], + "type": "interline_equation", + "content": "\\hat {t} _ {c, i} ^ {P} < \\frac {- I _ {c , i} + \\hat {t} _ {c , i} ^ {D} + \\hat {B} _ {c} ^ {\\infty}}{\\hat {B} _ {c} ^ {\\infty}}, \\quad i \\in M _ {b}. \\tag {16}", + "image_path": "d359bd1f784347af7414b775e70f465dde4b05d1caa7685bec4449a5f351a3a8.jpg" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 130, + 438, + 482, + 510 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 438, + 482, + 510 + ], + "spans": [ + { + "bbox": [ + 130, + 438, + 482, + 510 + ], + "type": "text", + "content": "This upper bound on " + }, + { + "bbox": [ + 130, + 438, + 482, + 510 + ], + "type": "inline_equation", + "content": "t_{c,i}^{P}" + }, + { + "bbox": [ + 130, + 438, + 482, + 510 + ], + "type": "text", + "content": " in Eq. (16) is to adjust the depth estimation on the background regions. For the pixels on the background regions, the original DPT produce lower depth estimations. Even if the original pixel " + }, + { + "bbox": [ + 130, + 438, + 482, + 510 + ], + "type": "inline_equation", + "content": "J_{c,i}" + }, + { + "bbox": [ + 130, + 438, + 482, + 510 + ], + "type": "text", + "content": " is equal to 1 (perfectly white), the captured pixel value " + }, + { + "bbox": [ + 130, + 438, + 482, + 510 + ], + "type": "inline_equation", + "content": "I_{c,i}" + }, + { + "bbox": [ + 130, + 438, + 482, + 510 + ], + "type": "text", + "content": " from the background regions should not be greater than the sum of the estimated medium transmission rate from DPT and the backscattering term, as shown in Eq. (15)." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 130, + 511, + 482, + 584 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 511, + 482, + 584 + ], + "spans": [ + { + "bbox": [ + 130, + 511, + 482, + 584 + ], + "type": "text", + "content": "Bound losses: We set the bound losses to enforce the estimated transmission rate " + }, + { + "bbox": [ + 130, + 511, + 482, + 584 + ], + "type": "inline_equation", + "content": "\\hat{t}_{c,i}^{P}" + }, + { + "bbox": [ + 130, + 511, + 482, + 584 + ], + "type": "text", + "content": " to satisfy the listed two bounds and optimize the estimated depth of PUDE, " + }, + { + "bbox": [ + 130, + 511, + 482, + 584 + ], + "type": "inline_equation", + "content": "d_{i}^{P}" + }, + { + "bbox": [ + 130, + 511, + 482, + 584 + ], + "type": "text", + "content": ", in training. To simplify the mathematical expressions in the subsequent losses, we denote the ReLU function as " + }, + { + "bbox": [ + 130, + 511, + 482, + 584 + ], + "type": "inline_equation", + "content": "f:\\mathbb{R}\\to \\mathbb{R}^+" + }, + { + "bbox": [ + 130, + 511, + 482, + 584 + ], + "type": "text", + "content": ", where " + }, + { + "bbox": [ + 130, + 511, + 482, + 584 + ], + "type": "inline_equation", + "content": "f(x) = \\max (0,x)" + }, + { + "bbox": [ + 130, + 511, + 482, + 584 + ], + "type": "text", + "content": ". The number of elements in " + }, + { + "bbox": [ + 130, + 511, + 482, + 584 + ], + "type": "inline_equation", + "content": "M_b" + }, + { + "bbox": [ + 130, + 511, + 482, + 584 + ], + "type": "text", + "content": " denoted as " + }, + { + "bbox": [ + 130, + 511, + 482, + 584 + ], + "type": "inline_equation", + "content": "N_{b}" + }, + { + "bbox": [ + 130, + 511, + 482, + 584 + ], + "type": "text", + "content": ". According to the bounds in Eq. (11), Eq. (16), the bound losses are:" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 200, + 594, + 481, + 628 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 200, + 594, + 481, + 628 + ], + "spans": [ + { + "bbox": [ + 200, + 594, + 481, + 628 + ], + "type": "interline_equation", + "content": "\\mathcal {L} _ {b l} = \\sum_ {c \\in \\{G, B \\}} \\sum_ {i = 1} ^ {H W} f \\left(- I _ {c, i} + \\left(1 - \\hat {t} _ {c, i} ^ {P}\\right) \\hat {B} _ {c} ^ {\\infty}\\right) / H W, \\tag {17}", + "image_path": "c4497daa94919b44dfb45dbfc04044c7097cfcc6133fcac59181950f8bc9278a.jpg" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 188, + 640, + 481, + 667 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 188, + 640, + 481, + 667 + ], + "spans": [ + { + "bbox": [ + 188, + 640, + 481, + 667 + ], + "type": "interline_equation", + "content": "\\mathcal {L} _ {b u} = \\sum_ {c \\in \\{G, B \\}} \\sum_ {i \\in M _ {b}} f \\left(- \\left(1 - \\hat {t} _ {c, i} ^ {P}\\right) \\hat {B} _ {c} ^ {\\infty} - \\hat {t} _ {c, i} ^ {D} + I _ {c, i}\\right) / N _ {b}. \\tag {18}", + "image_path": "56b59a15fd2f6befe2d422748d5bc5d37ef1b99a404aacaf6941438746428b58.jpg" + } + ] + } + ], + "index": 13 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "spans": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "type": "text", + "content": "10" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 166, + 91, + 222, + 102 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 166, + 91, + 222, + 102 + ], + "spans": [ + { + "bbox": [ + 166, + 91, + 222, + 102 + ], + "type": "text", + "content": "J. Yang et al." + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 9 + }, + { + "para_blocks": [ + { + "type": "table", + "bbox": [ + 133, + 147, + 481, + 319 + ], + "blocks": [ + { + "bbox": [ + 132, + 114, + 479, + 137 + ], + "lines": [ + { + "bbox": [ + 132, + 114, + 479, + 137 + ], + "spans": [ + { + "bbox": [ + 132, + 114, + 479, + 137 + ], + "type": "text", + "content": "Table 1: Quantitative results: The test datasets, Sea-thru D3, D5, and SQUID, comprise 43, 48, and 57 samples, respectively, collected from open ocean environment." + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 133, + 147, + 481, + 319 + ], + "lines": [ + { + "bbox": [ + 133, + 147, + 481, + 319 + ], + "spans": [ + { + "bbox": [ + 133, + 147, + 481, + 319 + ], + "type": "table", + "html": "
DatasetsMethodsδ<1.05↑δ<1.052↑δ<1.053↑AbsRel ↓Sq. Rel↓RMSE ↓RMSE log↓
D3MegaDepth [21]0.2750.5000.6590.1710.1121.0190.2250.054
MiDaS [29]0.2280.4290.5700.4080.9531.6350.4250.230
DPT [28]0.4210.6730.7900.1230.0850.8950.1870.038
UWNET [15]0.0970.1920.2900.7742.8292.0270.6500.451
UDepth [37]0.3150.5540.7070.1580.0941.0010.2130.051
Ours (PUDE)0.4430.7090.8260.1010.0590.8630.1610.027
D5MegaDepth [21]0.3520.5800.7220.1300.0431.4670.2020.043
MiDaS [29]0.3400.5730.7140.1310.0431.3050.1970.044
DPT [28]0.3740.6380.7740.1240.0511.4790.2050.045
UWNET [15]0.0750.1580.2400.3950.3342.5760.4090.184
UDepth [37]0.3760.6280.7760.1170.0381.3430.1920.039
Ours (PUDE)0.4350.6780.8030.1050.0341.3430.1860.037
SQUIDMegaDepth [21]0.1850.3520.4890.2530.1742.4190.2860.096
MiDaS [29]0.1960.3790.5070.2960.3472.4440.3090.139
DPT [28]0.3450.5990.7430.1360.0681.6570.1820.040
UWNET [15]0.0820.1670.2550.5100.7413.3630.4780.278
UDepth [37]0.2740.4640.6060.2130.1631.8830.2360.083
Ours (PUDE)0.3870.6320.7720.1190.0431.5240.1630.032
", + "image_path": "9fc071227242792028dd303153c50814a6d5dce38c54eef575a38fe24c73f8b0.jpg" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "table_body" + } + ], + "index": 3 + }, + { + "bbox": [ + 130, + 343, + 479, + 450 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 343, + 479, + 450 + ], + "spans": [ + { + "bbox": [ + 130, + 343, + 479, + 450 + ], + "type": "text", + "content": "These two bound losses are related to the pixel-wise intensity on color channel. Especially for the dark and distant regions, the difference in the pixel intensity can reflect the depth differences across pixels, the pixel intensity is mostly decided by the depth-related backscattering. The differences in bound losses caused by pixel intensity can encourage the model to learn the depth difference across pixels in small local blurry regions. In other words, the two bound losses guide the model to learn the nonlinear mechanism of backscattering in water, namely \"how blurriness affects the pixel intensity\". Fig. 4 shows an example. Therefore, the overall performance of the model in terms of detail detection gets improved." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 130, + 451, + 480, + 499 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 451, + 480, + 499 + ], + "spans": [ + { + "bbox": [ + 130, + 451, + 480, + 499 + ], + "type": "text", + "content": "Final loss: We use the parameters " + }, + { + "bbox": [ + 130, + 451, + 480, + 499 + ], + "type": "inline_equation", + "content": "\\beta_0, \\beta_1" + }, + { + "bbox": [ + 130, + 451, + 480, + 499 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 130, + 451, + 480, + 499 + ], + "type": "inline_equation", + "content": "\\beta_2" + }, + { + "bbox": [ + 130, + 451, + 480, + 499 + ], + "type": "text", + "content": " to adjust the weights between the similarity loss and bound losses, which are set to 1, 10 and 10, respectively. The weights are decided by the metric and visual results on the validation set during training. The total loss " + }, + { + "bbox": [ + 130, + 451, + 480, + 499 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_{total}" + }, + { + "bbox": [ + 130, + 451, + 480, + 499 + ], + "type": "text", + "content": " in training is:" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 237, + 508, + 480, + 521 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 237, + 508, + 480, + 521 + ], + "spans": [ + { + "bbox": [ + 237, + 508, + 480, + 521 + ], + "type": "interline_equation", + "content": "\\mathcal {L} _ {\\text {t o t a l}} = \\beta_ {0} \\mathcal {L} _ {s} + \\beta_ {1} \\mathcal {L} _ {b l} + \\beta_ {2} \\mathcal {L} _ {b u}. \\tag {19}", + "image_path": "36c5e5d446cfaefbd6f7ab4a091df9c39bf79182747661c306115d2ca10b90fc.jpg" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 132, + 560, + 229, + 574 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 560, + 229, + 574 + ], + "spans": [ + { + "bbox": [ + 132, + 560, + 229, + 574 + ], + "type": "text", + "content": "5 Experiments" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 132, + 586, + 279, + 597 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 586, + 279, + 597 + ], + "spans": [ + { + "bbox": [ + 132, + 586, + 279, + 597 + ], + "type": "text", + "content": "5.1 Implementation Details" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 130, + 605, + 479, + 666 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 605, + 479, + 666 + ], + "spans": [ + { + "bbox": [ + 130, + 605, + 479, + 666 + ], + "type": "text", + "content": "The training dataset for the PUDE method contains 76 white-balanced linear underwater images collected in an open ocean environment with continuous depth variation, sourced from the SeaThru-NeRF [19] dataset. This scene structure is beneficial for the accurate estimation of the parameters in the underwater image formation model Eq. (1). We apply gamma correction to these images to" + } + ] + } + ], + "index": 9 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 419, + 91, + 447, + 100 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 419, + 91, + 447, + 100 + ], + "spans": [ + { + "bbox": [ + 419, + 91, + 447, + 100 + ], + "type": "text", + "content": "PUDE" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 470, + 92, + 479, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 470, + 92, + 479, + 100 + ], + "spans": [ + { + "bbox": [ + 470, + 92, + 479, + 100 + ], + "type": "text", + "content": "11" + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 10 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 134, + 114, + 480, + 228 + ], + "blocks": [ + { + "bbox": [ + 134, + 114, + 480, + 228 + ], + "lines": [ + { + "bbox": [ + 134, + 114, + 480, + 228 + ], + "spans": [ + { + "bbox": [ + 134, + 114, + 480, + 228 + ], + "type": "image", + "image_path": "c9b32c4b244a9104283fb6966adb740200bd240814590b41a5f0ca8065297a7b.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 130, + 236, + 482, + 281 + ], + "lines": [ + { + "bbox": [ + 130, + 236, + 482, + 281 + ], + "spans": [ + { + "bbox": [ + 130, + 236, + 482, + 281 + ], + "type": "text", + "content": "Fig. 5: Qualitative comparison on underwater images. Qualitative comparison of PUDE with in-air (MegaDepth, MiDaS, DPT) and underwater methods (UWNET, UDepth) on underwater images. PUDE detects clearer background edges and presents more depth details. (More in Supplementary)" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_caption" + } + ], + "index": 2 + }, + { + "bbox": [ + 130, + 304, + 482, + 387 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 304, + 482, + 387 + ], + "spans": [ + { + "bbox": [ + 130, + 304, + 482, + 387 + ], + "type": "text", + "content": "generate corresponding nonlinear images for training. We randomly sample 20 images with groundtruth from the Sea-thru D3 datasets to construct the validation dataset. During training, the nonlinear images are downsampled to the dimension of " + }, + { + "bbox": [ + 130, + 304, + 482, + 387 + ], + "type": "inline_equation", + "content": "576 \\times 384" + }, + { + "bbox": [ + 130, + 304, + 482, + 387 + ], + "type": "text", + "content": " and fed into the neural network. PUDE is initialized with the DPT-Hybrid [28]. The training runs 3 epochs, employing the Adam optimizer with a learning rate of " + }, + { + "bbox": [ + 130, + 304, + 482, + 387 + ], + "type": "inline_equation", + "content": "10^{-5}" + }, + { + "bbox": [ + 130, + 304, + 482, + 387 + ], + "type": "text", + "content": ". Notably, an early stopping is applied to prevent overfitting by looking at the validation results." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 130, + 388, + 482, + 520 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 388, + 482, + 520 + ], + "spans": [ + { + "bbox": [ + 130, + 388, + 482, + 520 + ], + "type": "text", + "content": "In testing, we compare the performance of the PUDE model with the underwater depth estimation methods, UDepth [37], UW-Net [15], and in-air methods with good generalization properties MiDaS V2.1 [29], MegaDepth [21] and DPT-Hybrid [28]. The comparison encompasses both quantitative and qualitative evaluations. The quantitative evaluation uses the remaining images from the Sea-thru D3 and the whole D5 datasets [1]. Another zero-shot testing is performed on the SQUID dataset from [4], including groundtruth measurements obtained through stereo-matching. Given the availability of RAW format images in these datasets, we applied gamma correction and brightness enhancement to the images. In the zero-shot qualitative evaluation, the images are sampled from the real underwater dataset [20]." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 132, + 537, + 219, + 548 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 537, + 219, + 548 + ], + "spans": [ + { + "bbox": [ + 132, + 537, + 219, + 548 + ], + "type": "text", + "content": "5.2 Evaluations" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 130, + 557, + 482, + 666 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 557, + 482, + 666 + ], + "spans": [ + { + "bbox": [ + 130, + 557, + 482, + 666 + ], + "type": "text", + "content": "Quantitative Evaluation: We employ the alignment strategy described in [29] to compare the estimated inverse relative depth with the gorundtruth depth. This alignment operates in the space of inverse ground truth via solving an optimization problem for getting the shift and scale to align the estimated inverse relative depths with the inverse groundtruth. This alignment is subsequently inverted back to the normal depth space. To mitigate the impact of noise and erroneous measurements in groundtruth, we select depth thresholds ranging from " + }, + { + "bbox": [ + 130, + 557, + 482, + 666 + ], + "type": "inline_equation", + "content": "0.1m" + }, + { + "bbox": [ + 130, + 557, + 482, + 666 + ], + "type": "text", + "content": " to " + }, + { + "bbox": [ + 130, + 557, + 482, + 666 + ], + "type": "inline_equation", + "content": "15m" + }, + { + "bbox": [ + 130, + 557, + 482, + 666 + ], + "type": "text", + "content": " for the Sea-thru dataset [1] and " + }, + { + "bbox": [ + 130, + 557, + 482, + 666 + ], + "type": "inline_equation", + "content": "0.1m" + }, + { + "bbox": [ + 130, + 557, + 482, + 666 + ], + "type": "text", + "content": " to " + }, + { + "bbox": [ + 130, + 557, + 482, + 666 + ], + "type": "inline_equation", + "content": "20m" + }, + { + "bbox": [ + 130, + 557, + 482, + 666 + ], + "type": "text", + "content": " for the SQUID dataset [4] in our comparisons. We utilize the evaluation metrics described in [11] in our" + } + ] + } + ], + "index": 7 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "spans": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "type": "text", + "content": "12" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 166, + 91, + 222, + 102 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 166, + 91, + 222, + 102 + ], + "spans": [ + { + "bbox": [ + 166, + 91, + 222, + 102 + ], + "type": "text", + "content": "J. Yang et al." + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 11 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 134, + 114, + 482, + 167 + ], + "blocks": [ + { + "bbox": [ + 134, + 114, + 482, + 167 + ], + "lines": [ + { + "bbox": [ + 134, + 114, + 482, + 167 + ], + "spans": [ + { + "bbox": [ + 134, + 114, + 482, + 167 + ], + "type": "image", + "image_path": "132809b95d12103d908ebd6410aaab47c97a2ff13f9af6164ae736a136607827.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 130, + 173, + 482, + 207 + ], + "lines": [ + { + "bbox": [ + 130, + 173, + 482, + 207 + ], + "spans": [ + { + "bbox": [ + 130, + 173, + 482, + 207 + ], + "type": "text", + "content": "Fig. 6: Ablation study: (a) Without using " + }, + { + "bbox": [ + 130, + 173, + 482, + 207 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_{bu}" + }, + { + "bbox": [ + 130, + 173, + 482, + 207 + ], + "type": "text", + "content": ", the model cannot clearly detect the background edges near the rocks and has a fogging appearance. (b) Without using " + }, + { + "bbox": [ + 130, + 173, + 482, + 207 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_{bl}" + }, + { + "bbox": [ + 130, + 173, + 482, + 207 + ], + "type": "text", + "content": ", the distant rocks cannot be detected, depth of that region is overestimated." + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_caption" + } + ], + "index": 2 + }, + { + "bbox": [ + 130, + 230, + 482, + 339 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 230, + 482, + 339 + ], + "spans": [ + { + "bbox": [ + 130, + 230, + 482, + 339 + ], + "type": "text", + "content": "comparison: Mean absolute value of the relative error (AbsRel), Squared relative error (Sq. Rel), Root mean square error (RMSE), Root mean square logarithmic error (RMSE log), Scale-invariant logarithmic error (SILog) and Accuracy with threshold " + }, + { + "bbox": [ + 130, + 230, + 482, + 339 + ], + "type": "inline_equation", + "content": "(\\delta < thr)" + }, + { + "bbox": [ + 130, + 230, + 482, + 339 + ], + "type": "text", + "content": ". Note that the thr in the comparison metric is set to be 1.05. The results are shown in Tab. 1. The experimental results demonstrate the superior performance of the PUDE model across all evaluated metrics on three datasets compared to other methods. It achieves lower errors and higher estimation accuracy, substantiating the effectiveness of our proposed approach. Particularly, PUDE demonstrates promising performance in zero-shot testing." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 130, + 356, + 482, + 477 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 356, + 482, + 477 + ], + "spans": [ + { + "bbox": [ + 130, + 356, + 482, + 477 + ], + "type": "text", + "content": "Qualitative Evaluation We also provide qualitative evaluation across underwater images collected from different environments. We sample some underwater images that were collected in a turbid environment from the dataset [20] and do the zero-shot evaluation. The results are shown in Fig. 5. The performance of our PUDE model in underwater depth estimation is superior, particularly in capturing detailed features. It effectively identifies background edges, including the fine contours of distant wires. These outcomes show that PUDE not only inherits robust generalizability from DPT but also incorporates a physics-based understanding of underwater imaging, enabling the PUDE model to discern background regions without requiring labelled training data." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 132, + 493, + 246, + 504 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 493, + 246, + 504 + ], + "spans": [ + { + "bbox": [ + 132, + 493, + 246, + 504 + ], + "type": "text", + "content": "5.3 Ablation Studies" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 130, + 513, + 482, + 561 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 513, + 482, + 561 + ], + "spans": [ + { + "bbox": [ + 130, + 513, + 482, + 561 + ], + "type": "text", + "content": "To investigate the influence of the bound losses in training, we retrain the model with and without bound losses and evaluate them on the Sea-thru D5 dataset. The results, shown in Tab. 2, demonstrate both the losses " + }, + { + "bbox": [ + 130, + 513, + 482, + 561 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_{bu}" + }, + { + "bbox": [ + 130, + 513, + 482, + 561 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 130, + 513, + 482, + 561 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_{bl}" + }, + { + "bbox": [ + 130, + 513, + 482, + 561 + ], + "type": "text", + "content": " contribute to training, which is consistent with our training goals, solving the issues" + } + ] + } + ], + "index": 7 + }, + { + "type": "table", + "bbox": [ + 167, + 613, + 447, + 662 + ], + "blocks": [ + { + "bbox": [ + 130, + 581, + 482, + 604 + ], + "lines": [ + { + "bbox": [ + 130, + 581, + 482, + 604 + ], + "spans": [ + { + "bbox": [ + 130, + 581, + 482, + 604 + ], + "type": "text", + "content": "Table 2: Ablation study on D5 dataset. All bound losses contribute to the final performance of PUDE. PUDE has the best evaluation results on metrics." + } + ] + } + ], + "index": 8, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 167, + 613, + 447, + 662 + ], + "lines": [ + { + "bbox": [ + 167, + 613, + 447, + 662 + ], + "spans": [ + { + "bbox": [ + 167, + 613, + 447, + 662 + ], + "type": "table", + "html": "
LbuLblδ < 1.05δ < 1.052δ < 1.053AbsRelSq. RelRMSERMSE logSILog
✓0.3740.6380.7740.1240.0511.4790.2050.045
0.4260.6720.7980.1160.0481.4560.1990.042
✓0.3980.6570.7950.1090.0341.3830.1920.039
✓✓0.4350.6780.8030.1050.0341.3430.1860.037
", + "image_path": "8cf5f66185eb1179808892fd904d1379dda65a23eb8fcf1b8b66d5435a1e30f2.jpg" + } + ] + } + ], + "index": 9, + "angle": 0, + "type": "table_body" + } + ], + "index": 9 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 419, + 91, + 447, + 100 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 419, + 91, + 447, + 100 + ], + "spans": [ + { + "bbox": [ + 419, + 91, + 447, + 100 + ], + "type": "text", + "content": "PUDE" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 470, + 91, + 480, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 470, + 91, + 480, + 100 + ], + "spans": [ + { + "bbox": [ + 470, + 91, + 480, + 100 + ], + "type": "text", + "content": "13" + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 12 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 150, + 114, + 465, + 169 + ], + "blocks": [ + { + "bbox": [ + 150, + 114, + 465, + 169 + ], + "lines": [ + { + "bbox": [ + 150, + 114, + 465, + 169 + ], + "spans": [ + { + "bbox": [ + 150, + 114, + 465, + 169 + ], + "type": "image", + "image_path": "c297734858541ffa16c4575fe0c884efc816fef65677e0c716be7662f0b9234b.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 131, + 176, + 482, + 199 + ], + "lines": [ + { + "bbox": [ + 131, + 176, + 482, + 199 + ], + "spans": [ + { + "bbox": [ + 131, + 176, + 482, + 199 + ], + "type": "text", + "content": "Fig. 7: Method generalization: PUDE(M) also has improved performance in edge detection and detail detection compared with the in-air model, MiDaS." + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_caption" + } + ], + "index": 2 + }, + { + "bbox": [ + 130, + 223, + 482, + 284 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 223, + 482, + 284 + ], + "spans": [ + { + "bbox": [ + 130, + 223, + 482, + 284 + ], + "type": "text", + "content": "described in O1 and O2. Moreover, to visually demonstrate the effectiveness of bound losses " + }, + { + "bbox": [ + 130, + 223, + 482, + 284 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_{bl}" + }, + { + "bbox": [ + 130, + 223, + 482, + 284 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 130, + 223, + 482, + 284 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_{bu}" + }, + { + "bbox": [ + 130, + 223, + 482, + 284 + ], + "type": "text", + "content": ", we conducted a comparative analysis using images from the Sea-thru D5 dataset, as presented in Fig. 6. Without " + }, + { + "bbox": [ + 130, + 223, + 482, + 284 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_{bl}" + }, + { + "bbox": [ + 130, + 223, + 482, + 284 + ], + "type": "text", + "content": ", the depth in distant regions is overestimated. By contrast, without " + }, + { + "bbox": [ + 130, + 223, + 482, + 284 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_{bu}" + }, + { + "bbox": [ + 130, + 223, + 482, + 284 + ], + "type": "text", + "content": ", the detection of the background edges is not clear, resulting in a foggy appearance near the edges." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 131, + 300, + 399, + 313 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 131, + 300, + 399, + 313 + ], + "spans": [ + { + "bbox": [ + 131, + 300, + 399, + 313 + ], + "type": "text", + "content": "5.4 Method Generalization on Other In-air Models" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 130, + 321, + 482, + 429 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 321, + 482, + 429 + ], + "spans": [ + { + "bbox": [ + 130, + 321, + 482, + 429 + ], + "type": "text", + "content": "To further explore our method's generalization on different in-air models, we have conducted an additional experiment replacing DPT by MiDaS v2.1 [29] in the training loop, which has the same non-linearity described in Sec. 3 and demonstrate the second-best performance when applied to underwater images directly, and obtained PUDE(M). The quantitative and qualitative results are shown in Fig. 7 and Tab. 3. The results confirm that our method can also enhance the in-water performance of the non-SOTA model, MiDaS. However, as our method relies on knowledge transfer, the efficacy is contingent upon the capability of the original in-air model. A better model yields better transfer results." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 131, + 447, + 220, + 460 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 131, + 447, + 220, + 460 + ], + "spans": [ + { + "bbox": [ + 131, + 447, + 220, + 460 + ], + "type": "text", + "content": "6 Conclusion" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 130, + 472, + 482, + 568 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 472, + 482, + 568 + ], + "spans": [ + { + "bbox": [ + 130, + 472, + 482, + 568 + ], + "type": "text", + "content": "This paper proposes a novel method for transferring the in-air depth estimation model, DPT, to underwater settings. We observe DPT has issues with overestimating depth for distant regions and unclear background edge detection on underwater images. To tackle these, we use the novel bound losses to enforce the model to produce depth estimation that follows the underwater imaging formation model. The final quantitative and qualitative evaluation results prove the effectiveness of our method. Additionally, the further experiment shows our method's generalization capabilities on other in-air models." + } + ] + } + ], + "index": 8 + }, + { + "type": "table", + "bbox": [ + 163, + 622, + 451, + 663 + ], + "blocks": [ + { + "bbox": [ + 131, + 589, + 482, + 612 + ], + "lines": [ + { + "bbox": [ + 131, + 589, + 482, + 612 + ], + "spans": [ + { + "bbox": [ + 131, + 589, + 482, + 612 + ], + "type": "text", + "content": "Table 3: PUDE(M) on SQUID dataset: PUDE(M) shows improved performance on the metrics compared to MiDaS, but PUDE has the best performance." + } + ] + } + ], + "index": 9, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 163, + 622, + 451, + 663 + ], + "lines": [ + { + "bbox": [ + 163, + 622, + 451, + 663 + ], + "spans": [ + { + "bbox": [ + 163, + 622, + 451, + 663 + ], + "type": "table", + "html": "
Modelsδ < 1.05δ < 1.052δ < 1.053AbsRelSq. RelRMSERMSE logSILog
MiDaS0.1960.3790.5070.2960.3472.4440.3090.139
PUDE(M)0.2790.4630.6010.1880.1151.8690.2210.067
PUDE0.4350.6780.8030.1050.0341.3430.1860.037
", + "image_path": "e795be7e705be203157ac77a59a18c01dc51b7fa73c6b95d8b9cd57966e3ad0f.jpg" + } + ] + } + ], + "index": 10, + "angle": 0, + "type": "table_body" + } + ], + "index": 10 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "spans": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "type": "text", + "content": "14" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 166, + 91, + 222, + 102 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 166, + 91, + 222, + 102 + ], + "spans": [ + { + "bbox": [ + 166, + 91, + 222, + 102 + ], + "type": "text", + "content": "J. Yang et al." + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 13 + }, + { + "para_blocks": [ + { + "bbox": [ + 133, + 114, + 246, + 129 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 114, + 246, + 129 + ], + "spans": [ + { + "bbox": [ + 133, + 114, + 246, + 129 + ], + "type": "text", + "content": "Acknowledgements" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 132, + 141, + 479, + 165 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 141, + 479, + 165 + ], + "spans": [ + { + "bbox": [ + 132, + 141, + 479, + 165 + ], + "type": "text", + "content": "This work received funding from the Australian Research Council via grant DE220101527, and the Australian Government, via grant AUSMURIB000001." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 133, + 185, + 197, + 198 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 185, + 197, + 198 + ], + "spans": [ + { + "bbox": [ + 133, + 185, + 197, + 198 + ], + "type": "text", + "content": "References" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 138, + 211, + 480, + 665 + ], + "type": "list", + "angle": 0, + "index": 19, + "blocks": [ + { + "bbox": [ + 138, + 211, + 480, + 244 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 211, + 480, + 244 + ], + "spans": [ + { + "bbox": [ + 138, + 211, + 480, + 244 + ], + "type": "text", + "content": "1. Akkaynak, D., Treibitz, T.: Sea-thru: A method for removing water from underwater images. In: Proceedings of the IEEE/CVF conference on computer vision and pattern recognition. pp. 1682-1691 (2019)" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 138, + 246, + 480, + 277 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 246, + 480, + 277 + ], + "spans": [ + { + "bbox": [ + 138, + 246, + 480, + 277 + ], + "type": "text", + "content": "2. Amitai, S., Klein, I., Treibitz, T.: Self-supervised monocular depth underwater. In: 2023 IEEE International Conference on Robotics and Automation (ICRA). pp. 1098-1104. IEEE (2023)" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 138, + 279, + 480, + 311 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 279, + 480, + 311 + ], + "spans": [ + { + "bbox": [ + 138, + 279, + 480, + 311 + ], + "type": "text", + "content": "3. Ancuti, C.O., Ancuti, C., De Vleeschouwer, C., Neumann, L., Garcia, R.: Color transfer for underwater dehazing and depth estimation. In: 2017 IEEE International Conference on Image Processing (ICIP). pp. 695-699. IEEE (2017)" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 138, + 312, + 480, + 344 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 312, + 480, + 344 + ], + "spans": [ + { + "bbox": [ + 138, + 312, + 480, + 344 + ], + "type": "text", + "content": "4. Berman, D., Levy, D., Avidan, S., Treibitz, T.: Underwater single image color restoration using haze-lines and a new quantitative dataset. IEEE Transactions on Pattern Analysis and Machine Intelligence (2020)" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 138, + 346, + 480, + 377 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 346, + 480, + 377 + ], + "spans": [ + { + "bbox": [ + 138, + 346, + 480, + 377 + ], + "type": "text", + "content": "5. Bhat, S.F., Alhashim, I., Wonka, P.: Adabins: Depth estimation using adaptive bins. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 4009-4018 (2021)" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 138, + 379, + 480, + 399 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 379, + 480, + 399 + ], + "spans": [ + { + "bbox": [ + 138, + 379, + 480, + 399 + ], + "type": "text", + "content": "6. Bhat, S.F., Birkl, R., Wofk, D., Wonka, P., Müller, M.: Zoedepth: Zero-shot transfer by combining relative and metric depth. arXiv preprint arXiv:2302.12288 (2023)" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 138, + 401, + 480, + 433 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 401, + 480, + 433 + ], + "spans": [ + { + "bbox": [ + 138, + 401, + 480, + 433 + ], + "type": "text", + "content": "7. Chang, H.H., Cheng, C.Y., Sung, C.C.: Single underwater image restoration based on depth estimation and transmission compensation. IEEE Journal of Oceanic Engineering 44(4), 1130-1149 (2018)" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 138, + 434, + 480, + 466 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 434, + 480, + 466 + ], + "spans": [ + { + "bbox": [ + 138, + 434, + 480, + 466 + ], + "type": "text", + "content": "8. Chen, W., Qian, S., Deng, J.: Learning single-image depth from videos using quality assessment networks. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 5604-5613 (2019)" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 138, + 468, + 480, + 499 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 468, + 480, + 499 + ], + "spans": [ + { + "bbox": [ + 138, + 468, + 480, + 499 + ], + "type": "text", + "content": "9. Chiang, J.Y., Chen, Y.C.: Underwater image enhancement by wavelength compensation and dehazing. IEEE transactions on image processing 21(4), 1756-1769 (2011)" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 138, + 501, + 480, + 533 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 501, + 480, + 533 + ], + "spans": [ + { + "bbox": [ + 138, + 501, + 480, + 533 + ], + "type": "text", + "content": "0. Drews, P.L., Nascimento, E.R., Botelho, S.S., Campos, M.F.M.: Underwater depth estimation and image restoration based on single images. IEEE computer graphics and applications 36(2), 24-35 (2016)" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 138, + 534, + 480, + 565 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 534, + 480, + 565 + ], + "spans": [ + { + "bbox": [ + 138, + 534, + 480, + 565 + ], + "type": "text", + "content": "1. Eigen, D., Puhrsch, C., Fergus, R.: Depth map prediction from a single image using a multi-scale deep network. Advances in neural information processing systems 27 (2014)" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 138, + 567, + 480, + 599 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 567, + 480, + 599 + ], + "spans": [ + { + "bbox": [ + 138, + 567, + 480, + 599 + ], + "type": "text", + "content": "2. Fu, H., Gong, M., Wang, C., Batmanghelich, K., Tao, D.: Deep ordinal regression network for monocular depth estimation. In: Proceedings of the IEEE conference on computer vision and pattern recognition. pp. 2002-2011 (2018)" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 138, + 600, + 480, + 632 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 600, + 480, + 632 + ], + "spans": [ + { + "bbox": [ + 138, + 600, + 480, + 632 + ], + "type": "text", + "content": "3. Godard, C., Mac Aodha, O., Brostow, G.J.: Unsupervised monocular depth estimation with left-right consistency. In: Proceedings of the IEEE conference on computer vision and pattern recognition. pp. 270-279 (2017)" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 138, + 633, + 480, + 665 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 633, + 480, + 665 + ], + "spans": [ + { + "bbox": [ + 138, + 633, + 480, + 665 + ], + "type": "text", + "content": "4. Godard, C., Mac Aodha, O., Firman, M., Brostow, G.J.: Digging into self-supervised monocular depth estimation. In: Proceedings of the IEEE/CVF international conference on computer vision. pp. 3828-3838 (2019)" + } + ] + } + ], + "index": 18 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 419, + 91, + 447, + 100 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 419, + 91, + 447, + 100 + ], + "spans": [ + { + "bbox": [ + 419, + 91, + 447, + 100 + ], + "type": "text", + "content": "PUDE" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 470, + 92, + 480, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 470, + 92, + 480, + 100 + ], + "spans": [ + { + "bbox": [ + 470, + 92, + 480, + 100 + ], + "type": "text", + "content": "15" + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 14 + }, + { + "para_blocks": [ + { + "bbox": [ + 132, + 116, + 480, + 665 + ], + "type": "list", + "angle": 0, + "index": 18, + "blocks": [ + { + "bbox": [ + 133, + 116, + 480, + 149 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 116, + 480, + 149 + ], + "spans": [ + { + "bbox": [ + 133, + 116, + 480, + 149 + ], + "type": "text", + "content": "15. Gupta, H., Mitra, K.: Unsupervised single image underwater depth estimation. In: 2019 IEEE International Conference on Image Processing (ICIP). pp. 624-628. IEEE (2019)" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 133, + 150, + 480, + 183 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 150, + 480, + 183 + ], + "spans": [ + { + "bbox": [ + 133, + 150, + 480, + 183 + ], + "type": "text", + "content": "16. Hambarde, P., Murala, S., Dhall, A.: Uw-gan: Single-image depth estimation and image enhancement for underwater images. IEEE Transactions on Instrumentation and Measurement 70, 1-12 (2021)" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 132, + 183, + 480, + 215 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 183, + 480, + 215 + ], + "spans": [ + { + "bbox": [ + 132, + 183, + 480, + 215 + ], + "type": "text", + "content": "17. He, K., Sun, J., Tang, X.: Single image haze removal using dark channel prior. IEEE transactions on pattern analysis and machine intelligence 33(12), 2341-2353 (2010)" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 132, + 216, + 480, + 248 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 216, + 480, + 248 + ], + "spans": [ + { + "bbox": [ + 132, + 216, + 480, + 248 + ], + "type": "text", + "content": "18. Laina, I., Rupprecht, C., Belagiannis, V., Tombari, F., Navab, N.: Deeper depth prediction with fully convolutional residual networks. In: 2016 Fourth international conference on 3D vision (3DV). pp. 239-248. IEEE (2016)" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 132, + 249, + 480, + 293 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 249, + 480, + 293 + ], + "spans": [ + { + "bbox": [ + 132, + 249, + 480, + 293 + ], + "type": "text", + "content": "19. Levy, D., Peleg, A., Pearl, N., Rosenbaum, D., Akkaynak, D., Korman, S., Treibitz, T.: Seathru-nerf: Neural radiance fields in scattering media. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 56-65 (2023)" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 132, + 293, + 480, + 326 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 293, + 480, + 326 + ], + "spans": [ + { + "bbox": [ + 132, + 293, + 480, + 326 + ], + "type": "text", + "content": "20. Li, C., Guo, C., Ren, W., Cong, R., Hou, J., Kwong, S., Tao, D.: An underwater image enhancement benchmark dataset and beyond. IEEE Transactions on Image Processing 29, 4376-4389 (2019)" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 132, + 326, + 480, + 358 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 326, + 480, + 358 + ], + "spans": [ + { + "bbox": [ + 132, + 326, + 480, + 358 + ], + "type": "text", + "content": "21. Li, Z., Snavely, N.: Megadepth: Learning single-view depth prediction from internet photos. In: Proceedings of the IEEE conference on computer vision and pattern recognition. pp. 2041-2050 (2018)" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 132, + 358, + 480, + 392 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 358, + 480, + 392 + ], + "spans": [ + { + "bbox": [ + 132, + 358, + 480, + 392 + ], + "type": "text", + "content": "22. Luo, Y., Ren, J., Lin, M., Pang, J., Sun, W., Li, H., Lin, L.: Single view stereo matching. In: Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition. pp. 155-163 (2018)" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 132, + 392, + 480, + 413 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 392, + 480, + 413 + ], + "spans": [ + { + "bbox": [ + 132, + 392, + 480, + 413 + ], + "type": "text", + "content": "23. Ming, Y., Meng, X., Fan, C., Yu, H.: Deep learning for monocular depth estimation: A review. Neurocomputing 438, 14-33 (2021)" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 132, + 413, + 480, + 445 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 413, + 480, + 445 + ], + "spans": [ + { + "bbox": [ + 132, + 413, + 480, + 445 + ], + "type": "text", + "content": "24. Peng, Y.T., Cao, K., Cosman, P.C.: Generalization of the dark channel prior for single image restoration. IEEE Transactions on Image Processing 27(6), 2856-2868 (2018)" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 132, + 446, + 480, + 479 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 446, + 480, + 479 + ], + "spans": [ + { + "bbox": [ + 132, + 446, + 480, + 479 + ], + "type": "text", + "content": "25. Peng, Y.T., Cosman, P.C.: Underwater image restoration based on image blurriness and light absorption. IEEE transactions on image processing 26(4), 1579-1594 (2017)" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 132, + 479, + 480, + 512 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 479, + 480, + 512 + ], + "spans": [ + { + "bbox": [ + 132, + 479, + 480, + 512 + ], + "type": "text", + "content": "26. Peng, Y.T., Zhao, X., Cosman, P.C.: Single underwater image enhancement using depth estimation based on blurriness. In: 2015 IEEE International Conference on Image Processing (ICIP). pp. 4952-4956. IEEE (2015)" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 132, + 512, + 480, + 544 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 512, + 480, + 544 + ], + "spans": [ + { + "bbox": [ + 132, + 512, + 480, + 544 + ], + "type": "text", + "content": "27. Ramamonjisoa, M., Lepetit, V.: Sharpnet: Fast and accurate recovery of occluding contours in monocular depth estimation. In: Proceedings of the IEEE/CVF International Conference on Computer Vision Workshops. pp. 0-0 (2019)" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 132, + 545, + 480, + 578 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 545, + 480, + 578 + ], + "spans": [ + { + "bbox": [ + 132, + 545, + 480, + 578 + ], + "type": "text", + "content": "28. Ranftl, R., Bochkovskiy, A., Koltun, V.: Vision transformers for dense prediction. In: Proceedings of the IEEE/CVF international conference on computer vision. pp. 12179-12188 (2021)" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 132, + 578, + 480, + 621 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 578, + 480, + 621 + ], + "spans": [ + { + "bbox": [ + 132, + 578, + 480, + 621 + ], + "type": "text", + "content": "29. Ranftl, R., Lasinger, K., Hafner, D., Schindler, K., Koltun, V.: Towards robust monocular depth estimation: Mixing datasets for zero-shot cross-dataset transfer. IEEE transactions on pattern analysis and machine intelligence 44(3), 1623-1637 (2020)" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 132, + 622, + 480, + 665 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 622, + 480, + 665 + ], + "spans": [ + { + "bbox": [ + 132, + 622, + 480, + 665 + ], + "type": "text", + "content": "30. Varghese, N., Kumar, A., Rajagopalan, A.: Self-supervised monocular underwater depth recovery, image restoration, and a real-sea video dataset. In: Proceedings of the IEEE/CVF International Conference on Computer Vision. pp. 12248-12258 (2023)" + } + ] + } + ], + "index": 17 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "spans": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "type": "text", + "content": "16" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 166, + 91, + 222, + 102 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 166, + 91, + 222, + 102 + ], + "spans": [ + { + "bbox": [ + 166, + 91, + 222, + 102 + ], + "type": "text", + "content": "J. Yang et al." + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 15 + }, + { + "para_blocks": [ + { + "bbox": [ + 130, + 116, + 482, + 347 + ], + "type": "list", + "angle": 0, + "index": 9, + "blocks": [ + { + "bbox": [ + 130, + 116, + 482, + 160 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 116, + 482, + 160 + ], + "spans": [ + { + "bbox": [ + 130, + 116, + 482, + 160 + ], + "type": "text", + "content": "31. Vasiljevic, I., Kolkin, N., Zhang, S., Luo, R., Wang, H., Dai, F.Z., Daniele, A.F., Mostajabi, M., Basart, S., Walter, M.R., Shakhnarovich, G.: DIODE: A Dense Indoor and Outdoor DEpth Dataset. CoRR abs/1908.00463 (2019), http:// arxiv.org/abs/1908.00463" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 130, + 161, + 482, + 194 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 161, + 482, + 194 + ], + "spans": [ + { + "bbox": [ + 130, + 161, + 482, + 194 + ], + "type": "text", + "content": "32. Wu, C.Y., Wang, J., Hall, M., Neumann, U., Su, S.: Toward practical monocular indoor depth estimation. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 3814-3824 (2022)" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 130, + 194, + 482, + 225 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 194, + 482, + 225 + ], + "spans": [ + { + "bbox": [ + 130, + 194, + 482, + 225 + ], + "type": "text", + "content": "33. Wu, Y., Zhou, Y., Chen, S., Ma, Y., Li, Q.: Defect inspection for underwater structures based on line-structured light and binocular vision. Applied Optics 60(25), 7754-7764 (2021)" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 130, + 227, + 482, + 259 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 227, + 482, + 259 + ], + "spans": [ + { + "bbox": [ + 130, + 227, + 482, + 259 + ], + "type": "text", + "content": "34. Xian, K., Shen, C., Cao, Z., Lu, H., Xiao, Y., Li, R., Luo, Z.: Monocular relative depth perception with web stereo data supervision. In: Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition. pp. 311-320 (2018)" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 130, + 259, + 482, + 281 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 259, + 482, + 281 + ], + "spans": [ + { + "bbox": [ + 130, + 259, + 482, + 281 + ], + "type": "text", + "content": "35. Yang, J., Gong, M., Nair, G., Lee, J.H., Monty, J., Pu, Y.: Knowledge distillation for feature extraction in underwater vslam. arXiv preprint arXiv:2303.17981 (2023)" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 130, + 281, + 482, + 314 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 281, + 482, + 314 + ], + "spans": [ + { + "bbox": [ + 130, + 281, + 482, + 314 + ], + "type": "text", + "content": "36. Ye, X., Zhang, J., Yuan, Y., Xu, R., Wang, Z., Li, H.: Underwater depth estimation via stereo adaptation networks. IEEE Transactions on Circuits and Systems for Video Technology (2023)" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 130, + 314, + 482, + 347 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 314, + 482, + 347 + ], + "spans": [ + { + "bbox": [ + 130, + 314, + 482, + 347 + ], + "type": "text", + "content": "37. Yu, B., Wu, J., Islam, M.J.: Udepth: Fast monocular depth estimation for visually-guided underwater robots. In: 2023 IEEE International Conference on Robotics and Automation (ICRA). pp. 3116-3123. IEEE (2023)" + } + ] + } + ], + "index": 8 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 419, + 91, + 447, + 100 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 419, + 91, + 447, + 100 + ], + "spans": [ + { + "bbox": [ + 419, + 91, + 447, + 100 + ], + "type": "text", + "content": "PUDE" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 470, + 91, + 481, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 470, + 91, + 481, + 100 + ], + "spans": [ + { + "bbox": [ + 470, + 91, + 481, + 100 + ], + "type": "text", + "content": "17" + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 16 + } + ], + "_backend": "vlm", + "_version_name": "2.6.4" +} \ No newline at end of file diff --git a/2024/PiTe_ Pixel-Temporal Alignment for Large Video-Language Model/3468332a-054d-4ee9-a3ba-3d30cf6d6594_content_list.json b/2024/PiTe_ Pixel-Temporal Alignment for Large Video-Language Model/3468332a-054d-4ee9-a3ba-3d30cf6d6594_content_list.json new file mode 100644 index 0000000000000000000000000000000000000000..b00af9fae673fe7fa2a15290ae8a17225d74164f --- /dev/null +++ b/2024/PiTe_ Pixel-Temporal Alignment for Large Video-Language Model/3468332a-054d-4ee9-a3ba-3d30cf6d6594_content_list.json @@ -0,0 +1,1744 @@ +[ + { + "type": "text", + "text": "PiTe: Pixel-Temporal Alignment for Large Video-Language Model", + "text_level": 1, + "bbox": [ + 261, + 114, + 673, + 159 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Yang Liu $^{1,2\\star}$ , Pengxiang Ding $^{1\\star}$ , Siteng Huang $^{1}$ , Min Zhang $^{1}$ , Han Zhao $^{1}$ , and Donglin Wang $^{1\\boxtimes}$", + "bbox": [ + 243, + 185, + 689, + 217 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "$^{1}$ Westlake University $^{2}$ Soochow University {liuyang67, dingpengxiang, wangdonglin}@westlake.edu.cn", + "bbox": [ + 243, + 227, + 669, + 257 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Abstract. Fueled by the Large Language Models (LLMs) wave, Large Visual-Language Models (LVLMs) have emerged as a pivotal advancement, bridging the gap between image and text. However, video making it challenging for LVLMs to perform adequately due to the complexity of the relationship between language and spatial-temporal data structure. Recent Large Video-Language Models (LVidLMs) align feature of static visual data like image into latent space of language feature, by general multi-modal tasks to leverage abilities of LLMs sufficiently. In this paper, we explore fine-grained alignment approach via object trajectory for different modalities across both spatial and temporal dimensions simultaneously. Thus, we propose a novel LVidLM by trajectory-guided Pixel-Temporal Alignment, dubbed PiTe, that exhibits promising applicable model property. To achieve fine-grained video-language alignment, we curate a multi-modal pre-training dataset PiTe-143k, the dataset provision of moving trajectories in pixel level for all individual objects, that appear and mention in the video and caption both, by our automatic annotation pipeline. Meanwhile, PiTe demonstrates astounding capabilities on myriad video-related multi-modal tasks through beat the state-of-the-art methods by a large margin.", + "bbox": [ + 163, + 294, + 772, + 501 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Keywords: Large Video-Language Model $\\cdot$ Trajectory-guided Instruction Tuning $\\cdot$ Video Understanding", + "bbox": [ + 163, + 516, + 772, + 544 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "1 Introduction", + "text_level": 1, + "bbox": [ + 117, + 570, + 272, + 585 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Large Language Models (LLMs) have rapidly gained popularity within the AI community, demonstrating astounding capabilities across a wide array of natural language tasks [4, 5, 9, 29, 37, 38]. The powerful language comprehension abilities of LLMs drive researchers to explore their utility in addressing a broader spectrum of tasks across various domains. Consequently, an increasing number of studies are focusing on developing comprehensive Large Visual-Language Models (LVLMs) to tackle vision-related tasks in zero-shot settings [6, 18, 47, 52], particularly in the video understanding [19, 23, 26, 27, 44, 46, 48]. The pursuit of generalist Large Video-Language Models (LVidLMs) will be a perennial challenge. Success in this endeavor hinges on effectively leveraging the exceptional understanding, reasoning, and generative capacities inherently present in LLMs.", + "bbox": [ + 117, + 601, + 818, + 737 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "One potential route towards addressing the issue is aligning visual feature into latent space of language feature. To achieve this, existing LVidLMs apply large-scale vanilla instruction tuning [19, 26,46,48]. However, the conventional question-answering training paradigm primarily assists LLMs in understanding visual data from a spatial perspective, posing challenges in effectively capturing", + "bbox": [ + 117, + 739, + 818, + 799 + ], + "page_idx": 0 + }, + { + "type": "page_footnote", + "text": "* Equal contribution. ☑ Corresponding author.", + "bbox": [ + 133, + 806, + 459, + 821 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/8f14ab64a2c6781696041d2a227162010dd5d83385f41668a3c3e5b215c30bb3.jpg", + "image_caption": [ + "(a) Illustration of training paradigm comparison." + ], + "image_footnote": [], + "bbox": [ + 189, + 118, + 612, + 246 + ], + "page_idx": 1 + }, + { + "type": "image", + "img_path": "images/8640e8608f6e32ca3ab0e52c41c304e58cff15c0ca50863f56c6532418777ba1.jpg", + "image_caption": [ + "(b) Performance comparison.", + "Fig. 1: Comparison with existing LVidLMs in terms of alignment paradigm and performance. For Fig. 1b, $QA$ , $TG$ , $DC$ denote question answering, temporal grounding and dense captioning, respectively." + ], + "image_footnote": [], + "bbox": [ + 642, + 117, + 880, + 247 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "temporal dynamics and spatial consistency relationships. Therefore, relying solely on instruction tuning proves insufficient for achieving comprehensive video comprehension, given the intricate spatial-temporal data structure involved. It is crucial to align different modalities across both spatial and temporal dimensions. Furthermore, offering more fine-grained cross-modal alignment guidance significantly enhances LVidLMs' ability to comprehend videos [24].", + "bbox": [ + 173, + 320, + 880, + 396 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "To bridge the gap, we introduce a novel LVidLM named PiTe, which employs trajectories to intricately align vision and language across both spatial and temporal dimensions at the pixel level, and the distinction from conventional approaches is illustrated in Figure 1a. By requiring the model to forecast the trajectory of individual objects mentioned in the text within the video, it enables the learning of fine-grained text-to-pixel alignment through exploiting the video context along the temporal dimension and enhancing its ability to generate output based on evidence.", + "bbox": [ + 173, + 396, + 880, + 487 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Subsequently, due to there are no ready-made video-language dataset with moving trajectory of objects, we curate a large-scale video-language dataset PiTe-143k through an automated annotation pipeline. Consequently, as shown in Figure 1b, the proposed PiTe significantly augments the LVidLM's capacity to understand videos comprehensively, leading to promising, competitive, and state-of-the-art performance in question-answering, temporal grounding, and dense captioning tasks under zero-shot conditions.", + "bbox": [ + 173, + 487, + 880, + 575 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Overall, our principal contributions in this paper are summarized as follows:", + "bbox": [ + 196, + 577, + 748, + 592 + ], + "page_idx": 1 + }, + { + "type": "list", + "sub_type": "text", + "list_items": [ + "- We curate a large scale video-language dataset PiTe-143k with trajectory for all individual objects by automatic annotation pipeline.", + "- We propose a novel LVidLM PiTe that utilize trajectory to align video and language features across both spatial and temporal dimensions.", + "- Extensive experimental results and analysis on myriad datasets for zero-shot video question answering, temporal grounding, and dense captioning tasks demonstrate the superiority of PiTe." + ], + "bbox": [ + 174, + 592, + 880, + 683 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "2 Related Work", + "text_level": 1, + "bbox": [ + 174, + 705, + 346, + 722 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "2.1 Large Language Models", + "text_level": 1, + "bbox": [ + 174, + 737, + 419, + 753 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Over the last few years, pioneering foundation language models like GPT-1 [32], BERT [7], GPT-2 [33], and T5 [34] laid the groundwork, but GPT-3 [4] groundbreaking model parameters to 175 billion size to achieve remarkable zero-shot performance. Besides, research on scaling law [14] has steered language models to a larger scale. Therefore, driven by the success of InstructGPT [29]", + "bbox": [ + 173, + 761, + 880, + 824 + ], + "page_idx": 1 + }, + { + "type": "page_number", + "text": "2", + "bbox": [ + 176, + 89, + 189, + 99 + ], + "page_idx": 1 + }, + { + "type": "header", + "text": "Y. Liu, P. Ding et al.", + "bbox": [ + 230, + 88, + 374, + 102 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "and ChatGPT [28] which training by reinforcement learning with human feedback (RLHF) based on GPT-3, Large Language Models (LLMs) has made waves in the natural language processing (NLP) community due to its capabilities in language understanding, logical reasoning, and generation. The GPT's success suggest a promising path towards building LLMs, several open-source LLMs have been proposed following it with similar performance including OPT [49], BLOOM [36], GLM [9], LLaMA [37, 38], and Vicuna [5]. Our investigation delves into leveraging the striking language comprehension and zero-shot generalization abilities of LLMs beyond the confines of linguistic modalities. Specifically, we aim to extend these capabilities to multi-modal scenarios, thereby exploring their potential in processing diverse forms of information across different modalities.", + "bbox": [ + 117, + 119, + 820, + 255 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "2.2 Large Visual-Language Models", + "text_level": 1, + "bbox": [ + 117, + 277, + 413, + 292 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "The surge of LLMs has lead to major advancements in NLP tasks, and also has incited interest in developing Large Visual-Language Models (LVLMs). Building a unify LLM with visual inputs for visual language tasks thus remains one of the most important desiderata for LVLMs. Flamingo [1] and OpenFlamingo [3] fuse visual information into intermediate embedding for a frozen LLM by cross-attention mechanism, and train on billions of image-text pairs to align visual and linguistic modalities. Similarly, BLIP-2 [18] introduced the concept of Q-Former to align visual features more effectively with language space. Moreover, MiniGPT-4 [52] enhances its usability significantly by further fine-tuning on more detailed image descriptions with just one projection layer to align a frozen visual encoder with a frozen LLM, and the LLaVA series [21, 22] use simply a multi-layer perception (MLP) in place of the Q-Former and two-stage instruction tuning to enhance this process. Furthermore, PixelLLM [42] leverage the location coordinate of every word in the caption in the image as the connection between different modalities to strengthen the model's performance for the object detection task. Our primary focus lies in transferring the exceptional language comprehension capabilities of LLMs to the analysis of dynamic, continuous visual data found in videos, as opposed to static visual data such as images.", + "bbox": [ + 117, + 301, + 820, + 527 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "2.3 Large Video-Language Models", + "text_level": 1, + "bbox": [ + 117, + 549, + 410, + 565 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Recently, many efforts have been made to transfer the task-handling capability of LVLMs to the video modality, leading to the emergence of Large Video-Language Models (LVidLMs) like VideoChat [19], Video-LLaMA [46], and Video-ChatGPT [26]. Prior researches have demonstrated the capability of LLMs to perform diverse tasks on video content, guided by user instructions through a two-stage training process. These studies align static visual features with LLMs, followed by instruction tuning on datasets annotated either by GPT or humans. Despite being effective in video understanding, the lack of fine-grained spatial-temporal modeling in these models prevents them from understanding or locating object in detail or specific segments. We propose a novel fine-grained alignment strategy at the pixel level across spatial and temporal dimensions to enhance the ability of LLMs to comprehensively analyze video content, thereby facilitating a more detailed understanding of the visual information presented.", + "bbox": [ + 117, + 573, + 820, + 739 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "3 PiTe-143k Dataset", + "text_level": 1, + "bbox": [ + 117, + 761, + 330, + 777 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "To facilitate fine-grained multi-modal alignment research at the pixel level, we introduce a large-scale video-language dataset PiTe-143k. This dataset fills a crucial gap in the existing resources by", + "bbox": [ + 117, + 792, + 818, + 821 + ], + "page_idx": 2 + }, + { + "type": "header", + "text": "PiTe: Pixel-Temporal Alignment for Large Video-Language Model", + "bbox": [ + 325, + 88, + 767, + 103 + ], + "page_idx": 2 + }, + { + "type": "page_number", + "text": "3", + "bbox": [ + 808, + 90, + 818, + 99 + ], + "page_idx": 2 + }, + { + "type": "table", + "img_path": "images/a8e8dbe6ccddeba9a0d935ac24aea60dcbe0bd1d94d4345c50faa729ad337068.jpg", + "table_caption": [ + "Table 1: Comparison between PiTe-143k and existing video instruction datasets." + ], + "table_footnote": [], + "table_body": "
DatasetTotal Dur.Avg. Dur.#Videos#EventsTemporal Localization#Objects Trajectories
VideoChat [19]41h18s8.2k×××
Valley [25]608h40s54.7k×××
Video-ChatGPT [26]432h117s13.3k×××
PiTe-143k (Ours)2086.44h52.18s143.64k343.93k✓1.02M
", + "bbox": [ + 181, + 143, + 875, + 250 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "providing extensive object moving trajectories with video instruction, which were previously unavailable in ready-made datasets. PiTe-143k constructed based on InternVid-10M-FLT [13,40] that each instance contains and entire video and multiple clip captions with start-stop timestamps. As shown in Table 1, PiTe-143k comprises 343.93 thousand event segments and 1.02 million moving trajectory for all individual objects that appear in both visual and textual modalities. To facilitate this objective, we establish an automatic annotation pipeline for PiTe-143k, fostering the advancement of LVidLMs for nuanced pixel-level video comprehension.", + "bbox": [ + 173, + 277, + 880, + 383 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "The automatic annotation pipeline for PiTe-143k comprises two primary stages, as depicted in Fig. 2: (1) Stage 1 involves the noun phrases extraction and referring expressions segmentation, thereby generating object masks within the frame for all individual objects referenced in the event caption; (2) Stage 2 centers on point tracking to capture the moving trajectories corresponding to the masks obtained in Stage 1.", + "bbox": [ + 173, + 383, + 880, + 460 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "3.1 Referring Expression Segmentation", + "text_level": 1, + "bbox": [ + 174, + 479, + 511, + 496 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "In stage 1, we aim to build closely fine-grained connection between video and language. To this end, we extract all noun phrases from caption and find the corresponding objects in the clip.", + "bbox": [ + 173, + 505, + 879, + 535 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "At inception, we leverage constituency parser SuPar [50,51] for language to extract noun phrase as shown in Fig. 3. Notably, in order to pass the simplest and most straightforward language instructions in next step, we only extract noun phrase from the lowest layer. For example in Fig. 3a, we consider two noun phrases $a$ pen and $a$ white table, but the parent node of the former that denotes $a$ pen on a white table not in our consideration because of the complexity of its composition. Following this, we utilize GLaMM [35], the first LVLM that can generate natural language responses seamlessly intertwined with corresponding object segmentation masks, to obtain the corresponding segmentation mask in the first frame of the clip for the text-based referring expression. While certain objects in the video, such as $a$ pen as illustrated in Fig.3a and Fig.2, may be too small to accurate detection. In such challenging cases, we disregard the trajectory information of the noun phrase. Despite this limitation, it has a minimal impact on the overall performance when utilizing extensive pre-training data on a large scale. Meanwhile, leveraging the exceptional language comprehension capabilities of LLMs, GLaMM can effectively filter out invalid referring expressions, those that do not constitute legal object references, such as front as depicted in Fig. 3b.", + "bbox": [ + 173, + 536, + 880, + 748 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "3.2 Point Tracking", + "text_level": 1, + "bbox": [ + 174, + 768, + 346, + 784 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "In stage 2, we aim to transfer the connection constructed in the previous stage to video, expanding out the temporal dimension specific to video compared to image. To this end, we track all individual", + "bbox": [ + 173, + 792, + 880, + 823 + ], + "page_idx": 3 + }, + { + "type": "page_number", + "text": "4", + "bbox": [ + 176, + 90, + 189, + 99 + ], + "page_idx": 3 + }, + { + "type": "header", + "text": "Y. Liu, P. Ding et al.", + "bbox": [ + 230, + 88, + 374, + 102 + ], + "page_idx": 3 + }, + { + "type": "image", + "img_path": "images/f0d3a3cd841ca8cab8227774b97883049cd94b0b152455a8a81e0d75c1ff9e6e.jpg", + "image_caption": [ + "Fig. 2: Automatic annotation pipeline for PiTe-143k. The video sample in the figure showcases two events positioned at the commencement and conclusion of the video. The procedure for extracting noun phrases by SuPar [50,51] is elucidated in Fig. 3." + ], + "image_footnote": [], + "bbox": [ + 127, + 118, + 810, + 542 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "objects in their clip to obtain the trajectory, the trajectory indicates the connection between video and language in both spatial and temporal dimensions.", + "bbox": [ + 112, + 626, + 818, + 655 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "The stage 2 commences when we employ DOT [17], a simple-yet-efficient method for tracking point to recover the trajectory of any scene point, for each clip to capture the trajectories for any point in first frame. According to our observation, the caption of each clip mainly describes simple video content in short sentences, so most of the caption corresponds to just one scene clip, which enables us to track objects that identified in the first frame. Subsequently, filter trajectories according to the segmentation mask of objects obtained in stage 1. So far, we obtain the trajectories for all objects in each clip for each video, we create the connection between video and language from both spatial and temporal through the trajectories, the existence of trajectories in video denotes whether the object exist, and the value of trajectory represent where the object exist in video. Lastly, we utilize the k-means++ [2] clustering algorithm to condense trajectories into three key points, effectively reducing computational demands. This approach is founded on the premise that", + "bbox": [ + 112, + 657, + 821, + 824 + ], + "page_idx": 4 + }, + { + "type": "header", + "text": "PiTe: Pixel-Temporal Alignment for Large Video-Language Model", + "bbox": [ + 325, + 88, + 767, + 103 + ], + "page_idx": 4 + }, + { + "type": "page_number", + "text": "5", + "bbox": [ + 808, + 89, + 821, + 99 + ], + "page_idx": 4 + }, + { + "type": "image", + "img_path": "images/9d8ed1084e159297b7cd070c45409805a38db4947fea84bd61ed6517bdce7c1f.jpg", + "image_caption": [ + "(a) The extraction of the NP for the first event caption of the sample video depicted in Fig. 2." + ], + "image_footnote": [], + "bbox": [ + 176, + 117, + 488, + 281 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/0f8c8bd0de36161106ce4e067048382beb22ce7feaf73171f40b77b56f1969b0.jpg", + "image_caption": [ + "(b) Referring to Fig. 2, the extraction of the NP for the second event caption.", + "Fig. 3: Two samples of constituency parser for Noun Phrase (NP) extraction." + ], + "image_footnote": [], + "bbox": [ + 622, + 117, + 880, + 281 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "three points adequately capture the typical geometric shape of objects, striking a balance between precision and computational efficiency. Furthermore, we conduct a comparative analysis of the performance using various key tracking points, as discussed in Section 5.3.", + "bbox": [ + 173, + 349, + 880, + 396 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "4 PiTe", + "text_level": 1, + "bbox": [ + 174, + 416, + 261, + 431 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "In this section, we propose a novel Large Video-Language Model (LVidLM), PiTe, which align video and language by trajectories across both spatial and temporal dimension. Fig. 4 illustrates an overview of PiTe.", + "bbox": [ + 173, + 446, + 880, + 492 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "4.1 Architecture", + "text_level": 1, + "bbox": [ + 174, + 512, + 328, + 526 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "PiTe is composed of a vision encoder to encode frames from video implemented as ViT [8], a vision adapter to project visual feature to semantic space of LLMs implemented as a linear projection layer, a LLM Vicuna v1.5 [5], and a localization projector or trajectory projector in separate training stage to guide LLMs to understand visual information implemented as a linear projection.", + "bbox": [ + 173, + 535, + 880, + 597 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Vision Encoder. Raw video data can be expressed as multiple frames such as $\\mathbf{v} = \\{f_1, f_2, \\dots, f_N\\} \\in \\mathbb{R}^{N \\times H \\times W \\times C}$ (frames $\\times$ height $\\times$ width $\\times$ channels). Following previous studies [22,26,46], we adopt ViT-L/14 [8] pre-trained from CLIP [31] as the vision encoder ViT to encode visual data. We uniformly sample $N$ frames for video $\\mathbf{v}$ , and encode $i$ -th frame $f_i$ through the vision encoder ViT:", + "bbox": [ + 173, + 614, + 887, + 674 + ], + "page_idx": 5 + }, + { + "type": "equation", + "text": "\n$$\n\\left\\{v _ {i} ^ {\\text {c l s}}, v _ {i} ^ {1}, v _ {i} ^ {2}, \\dots , v _ {i} ^ {P} \\right\\} = \\operatorname {V i T} \\left(f _ {i}\\right), \\tag {1}\n$$\n", + "text_format": "latex", + "bbox": [ + 416, + 680, + 880, + 704 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "where $P$ denotes the number of patches in the vision encoder ViT.", + "bbox": [ + 179, + 705, + 661, + 720 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Visual Adapter. A simpler projector forces the LLMs to learn more on handling visual inputs, leading to better generalization [20]. Hence, we utilize the global feature $v_{i}^{cls}$ from vision encoder ViT as the representation for the $i$ -th frame $f_{i}$ , and we apply a linear projection layer $\\mu(\\cdot)$ to connect the frame feature into the word embedding space of LLMs:", + "bbox": [ + 173, + 739, + 880, + 797 + ], + "page_idx": 5 + }, + { + "type": "equation", + "text": "\n$$\nz _ {i} = \\mu \\left(v _ {i} ^ {\\text {c l s}}\\right). \\tag {2}\n$$\n", + "text_format": "latex", + "bbox": [ + 478, + 803, + 880, + 825 + ], + "page_idx": 5 + }, + { + "type": "page_number", + "text": "6", + "bbox": [ + 176, + 89, + 189, + 99 + ], + "page_idx": 5 + }, + { + "type": "header", + "text": "Y. Liu, P. Ding et al.", + "bbox": [ + 230, + 88, + 374, + 102 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/82e2601f5aebe873d980632c8138ad2c187307a89ed6732422d0813e2a84bf7e.jpg", + "image_caption": [ + "Stage 1: Referring Expression Localization" + ], + "image_footnote": [], + "bbox": [ + 116, + 133, + 818, + 239 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/e4aafee559240a026d41dd990a02c06fff2af3f63260553bcee42929f24a77fc.jpg", + "image_caption": [ + "Stage 2: Pixel-Temporal Alignment" + ], + "image_footnote": [], + "bbox": [ + 116, + 258, + 818, + 561 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/184bfb12f0f5dfe93ddadd8d6b074a66cd13a8b87cabf4922f0da2d30ca1a657.jpg", + "image_caption": [ + "Stage 3: Video Question Answering", + "Fig.4: Schematic of PiTe framework for video-language alignment." + ], + "image_footnote": [], + "bbox": [ + 117, + 584, + 815, + 662 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Subsequently, a sequence of frame tokens $\\mathbf{z} = \\{z_1,z_2,\\dots ,z_N\\} \\in \\mathbb{R}^{N\\times d}$ becomes the input that LLMs can understand, $d$ denotes the hidden dimension of LLMs.", + "bbox": [ + 114, + 718, + 820, + 750 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Large Language Model. After we tokenize and encode video into frame tokens $\\mathbf{z}$ , we concatenate it with textual tokens $\\mathbf{w} = \\{w_{1}, w_{2}, \\ldots, w_{L}\\} \\in \\mathbb{R}^{L}$ and feed as the input to LLMs, we treat visual input as a foreign language in this process. Based on this, LLMs can further encode the input", + "bbox": [ + 112, + 777, + 821, + 824 + ], + "page_idx": 6 + }, + { + "type": "header", + "text": "PiTe: Pixel-Temporal Alignment for Large Video-Language Model", + "bbox": [ + 325, + 88, + 767, + 103 + ], + "page_idx": 6 + }, + { + "type": "page_number", + "text": "7", + "bbox": [ + 808, + 89, + 821, + 99 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "sequence to understand the video and text, then reasoning and generation the response using auto-regressive decoding as follows:", + "bbox": [ + 173, + 119, + 879, + 151 + ], + "page_idx": 7 + }, + { + "type": "equation", + "text": "\n$$\nh _ {i} = \\operatorname {L L M} ^ {-} \\left(\\mathbf {z}, \\mathbf {w} _ {1: i - 1}\\right), \\tag {3}\n$$\n", + "text_format": "latex", + "bbox": [ + 446, + 162, + 880, + 176 + ], + "page_idx": 7 + }, + { + "type": "equation", + "text": "\n$$\nw _ {i} = \\operatorname {a r g m a x} \\left(\\mathbf {m} _ {v} \\cdot h _ {i}\\right), \\tag {4}\n$$\n", + "text_format": "latex", + "bbox": [ + 447, + 181, + 880, + 195 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "where $\\mathrm{LLM}^{-}$ denotes the LLM without the last vocabulary mapping layer, $h_i$ denotes the hidden states of $i$ -th token generated by $\\mathrm{LLM}^{-}$ , $\\mathbf{m}_v \\in \\mathbb{R}^{|V| \\times d}$ denotes the weight of the linear vocabulary mapping layer.", + "bbox": [ + 173, + 207, + 880, + 252 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "4.2 Training Strategy", + "text_level": 1, + "bbox": [ + 174, + 273, + 366, + 287 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "For PiTe model training, we consider a three stage instruction-tuning procedure, as depicted in Fig. 4: (1) Stage 1 centered around training adapters using image-caption pairs; (2) Stage 2 is focuses on aligning video and language features through trajectories; (3) Stage 3 is dedicated to enhancing the model's comprehension by following human instructions through high-quality dialogue instruction tuning.", + "bbox": [ + 173, + 297, + 880, + 372 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Stage 1: Referring Expression Localization. At the initial stage, we aim to train the visual adapter that can align visual features with semantic space of LLMs. To this end, we employ Localized Narratives dataset [30] that contains annotations of human annotators narrating a given image, together with a mouse trajectory of the annotators' attention during the narration. This gives synchronized locations for all words in the narration sentence, the cross-modal attention of human can be used to train our model as condition to bridge vision and language.", + "bbox": [ + 173, + 393, + 880, + 482 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "There is only one visual tokens $\\mathbf{z} = \\{z_1\\} \\in \\mathbb{R}^{1\\times d}$ in this training stage for image instead of video, this can align vision with language in spatial to train adapter without consider temporal information. To use the same language features for localization, we simply add a multi-layer perception (MLP) as localization projector $\\varphi (\\cdot)$ parallel with the vocabulary mapping layer, which maps the language feature to a 2-dimension location:", + "bbox": [ + 173, + 483, + 880, + 551 + ], + "page_idx": 7 + }, + { + "type": "equation", + "text": "\n$$\np _ {i} = \\varphi \\left(h _ {i}\\right), \\tag {5}\n$$\n", + "text_format": "latex", + "bbox": [ + 491, + 554, + 879, + 568 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "where $p_i$ denotes the predicted coordinate for textual token $w_i$ .", + "bbox": [ + 179, + 571, + 635, + 585 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Overall, the leaning objective in the first stage was calculated by standard label-smoothed cross-entropy loss to train captioning output, and $L_{1}$ regression loss to train the localization output:", + "bbox": [ + 173, + 585, + 879, + 614 + ], + "page_idx": 7 + }, + { + "type": "equation", + "text": "\n$$\n\\mathcal {L} _ {1} = \\frac {1}{\\ell} \\sum_ {i = 1} ^ {\\ell} \\left(\\operatorname {C E} \\left(\\operatorname {L L M} \\left(\\mathbf {z}, \\mathbf {w} _ {1: i - 1}\\right), w _ {i}\\right) + \\lambda | \\hat {p} _ {i} - p _ {i} |\\right), \\tag {6}\n$$\n", + "text_format": "latex", + "bbox": [ + 352, + 623, + 880, + 657 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "where $\\ell$ is the length of the generated sequence, $\\hat{p}_i$ represents the ground truth of location, and $\\mathrm{CE}(\\cdot ,\\cdot)$ denotes the cross-entropy function. To enhance training efficiency, we utilize LoRA [12] for fine-tuning the LLM.", + "bbox": [ + 174, + 665, + 880, + 710 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Stage 2: Pixel-Temporal Alignment. After stage 1, the LLM model becomes proficient in understanding visual information. In the stage 2, we aim to train the LLM to understand sequential frames in video. To achieve this target, we curate a detailed object tracking dataset PiTe-143k, as described in Section 3 that uses trajectory as condition to bridge vision and language across both spatial and temporal dimensions. Therefore, the alignment guidance in pixel level improves the model's video fine-grained understanding reliability and overall usability.", + "bbox": [ + 173, + 731, + 880, + 823 + ], + "page_idx": 7 + }, + { + "type": "page_number", + "text": "8", + "bbox": [ + 176, + 89, + 189, + 99 + ], + "page_idx": 7 + }, + { + "type": "header", + "text": "Y. Liu, P. Ding et al.", + "bbox": [ + 230, + 88, + 374, + 102 + ], + "page_idx": 7 + }, + { + "type": "table", + "img_path": "images/080a10cf813006985bd4788c39f7eef82d19b37dcac120f4f7c7ead73b0de48d.jpg", + "table_caption": [ + "Table 2: Training hyper-parameters of PiTe." + ], + "table_footnote": [], + "table_body": "
ConfigurationStage 1 Stage 2 Stage 3ConfigurationStage 1 Stage 2 Stage 3
Vision EncoderOpenAI-CLIP-L/14Learning Rate0.0001
Image/Frame Resolution224×224LoRAr=64 & α=128
Adapter ParameterTunableFrozenFrozenNumerical PrecisionBFloat16
Video Frames100Epoch122
LLMVicuna-7B/13B-v1.5Global Batch Size256
LLM Sequence Length2048Learning Rate ScheduleCosine Decay
OptimizerAdamWWarm-up Ratio0.03
", + "bbox": [ + 120, + 143, + 818, + 272 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "Similar to stage 1, we use the same language features for alignment by a MLP as trajectory projector $\\rho(\\cdot)$ to map the language feature to a 2-dimension location:", + "bbox": [ + 114, + 301, + 820, + 333 + ], + "page_idx": 8 + }, + { + "type": "equation", + "text": "\n$$\n\\mathbf {p} _ {i} = \\rho \\left(h _ {i}\\right), \\tag {7}\n$$\n", + "text_format": "latex", + "bbox": [ + 429, + 344, + 820, + 359 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "where $\\mathbf{p}_i$ denotes the trajectory matrix in $P$ points and $N$ frames for textual token $w_i$ . Here, we define $p_{ijk}$ indicates the coordinate for token $w_i$ for $i$ -th point for model tracking in frame $f_j$ .", + "bbox": [ + 112, + 371, + 820, + 401 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "Overall, the leaning objective in stage 2 was calculated by standard label-smoothed cross-entropy loss to train the generation output, and $L_{1}$ regression loss to train the trajectory output as the condition:", + "bbox": [ + 112, + 401, + 820, + 441 + ], + "page_idx": 8 + }, + { + "type": "equation", + "text": "\n$$\n\\mathcal {L} _ {2} = \\frac {1}{\\ell} \\sum_ {i = 1} ^ {\\ell} \\left(\\operatorname {C E} \\left(\\operatorname {L L M} \\left(\\mathbf {z}, \\mathbf {w} _ {1: i - 1}\\right), w _ {i}\\right) + \\frac {\\lambda}{P \\cdot N} \\sum_ {j = 1} ^ {P} \\sum_ {k = 1} ^ {N} | \\hat {p} _ {i j k} - p _ {i j k} |\\right), \\tag {8}\n$$\n", + "text_format": "latex", + "bbox": [ + 230, + 441, + 820, + 479 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "where $P$ is the number of the points for model tracking to generate trajectory, and $\\mathbf{z} \\in \\mathbb{R}^{N \\times d}$ represents the sequence of visual embedding. We merge the LoRA trained in the stage 1 with the original model and introduce a new LoRA module.", + "bbox": [ + 112, + 486, + 820, + 531 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "It is notable that we use localization projector $\\varphi(\\cdot)$ trained in previous stage to initialize the trajectory projector $\\rho(\\cdot)$ . Specifically, we define the weight of localization projector $\\varphi(\\cdot)$ and trajectory projector $\\rho(\\cdot)$ as $\\mathbf{m}_{\\varphi} \\in \\mathbb{R}^{P \\cdot N \\cdot 2 \\times d}$ and $\\mathbf{m}_{\\rho} \\in \\mathbb{R}^{2 \\times d}$ , respectively. localization projector $\\varphi(\\cdot)$ maps a 2-dimension coordinate on the input image for each token of the LLM output, as for trajectory projector $\\rho(\\cdot)$ , it also output 2-dimension coordinates, but more than $P \\cdot N$ times for $P$ points to tracking for $N$ frames. For each point of each frame, the parameter of trajectory projector $\\rho(\\cdot)$ initialized by the localization projector $\\varphi(\\cdot)$ :", + "bbox": [ + 112, + 532, + 821, + 630 + ], + "page_idx": 8 + }, + { + "type": "equation", + "text": "\n$$\n\\mathbf {m} _ {\\varphi} = \\overbrace {\\mathbf {m} _ {\\rho} \\oplus \\mathbf {m} _ {\\rho} \\oplus \\cdots \\oplus \\mathbf {m} _ {\\rho}} ^ {P \\cdot N}, \\tag {9}\n$$\n", + "text_format": "latex", + "bbox": [ + 372, + 642, + 818, + 672 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "where $\\oplus$ denotes concatenation of matrix in first dimension.", + "bbox": [ + 119, + 685, + 553, + 699 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "Beyond trajectories, our model is attuned to temporal boundaries within the generated text. Specifically, we structure the generation as ..., from s to e or From s to e, ... to facilitate the model to learn in temporal dimension. Here, ... encapsulates the event description, while s and e denote the frame indexes corresponding to the start and end timestamps of event, respectively. This approach further augments the model's understanding of temporal boundaries [13].", + "bbox": [ + 112, + 700, + 818, + 777 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "Dissimilar to the initial training stage, not all generated words are associated with trajectories. In cases where objects lack trajectories or vanish from view over time, we uniformly assign the coordinates of their ground truth as $(-1, -1)$ to signify their absence.", + "bbox": [ + 112, + 777, + 820, + 824 + ], + "page_idx": 8 + }, + { + "type": "header", + "text": "PiTe: Pixel-Temporal Alignment for Large Video-Language Model", + "bbox": [ + 323, + 88, + 767, + 104 + ], + "page_idx": 8 + }, + { + "type": "page_number", + "text": "9", + "bbox": [ + 808, + 89, + 821, + 99 + ], + "page_idx": 8 + }, + { + "type": "table", + "img_path": "images/e423f237267028f84caeff315a6b160852c995baaa6ef36b1fc359d58526d454.jpg", + "table_caption": [ + "Table 3: Comparison between different LVidLMs on zero-shot question-answer." + ], + "table_footnote": [], + "table_body": "
ModelLLM SizeMSVD-QA [41]MSRVTT-QA [43]ActivityNet-QA [45]
Accuracy↑Score↑Accuracy↑Score↑Accuracy↑Score↑
FrozenBiLM [44]1B32.2-16.8-24.7-
LLaMA-Adapter [48]7B54.93.143.82.734.22.7
VideoChat [19]7B56.32.845.02.5-2.2
Video-LLaMA [46]7B51.62.529.61.812.41.1
Video-ChatGPT [26]7B64.93.349.32.835.22.7
PG-Video-LLaVA [27]7B64.13.751.63.339.93.3
PiTe (Ours)7B68.43.956.43.542.03.3
PiTe (Ours)13B71.64.057.73.542.23.4
", + "bbox": [ + 181, + 143, + 875, + 325 + ], + "page_idx": 9 + }, + { + "type": "text", + "text": "Stage 3: Video Question Answering. Following stage 2, we incorporate high-quality dialogue data Valley [25] and Video-ChatGPT [26] in one turn for instruction tuning, enabling the model to follow human instructions for more accurate and generalize capabilities of video understanding.", + "bbox": [ + 173, + 353, + 880, + 398 + ], + "page_idx": 9 + }, + { + "type": "text", + "text": "The leaning objective in third stage was calculated by standard label-smoothed cross-entropy loss for auto-regression generation:", + "bbox": [ + 174, + 398, + 880, + 426 + ], + "page_idx": 9 + }, + { + "type": "equation", + "text": "\n$$\n\\mathcal {L} _ {3} = \\frac {1}{\\ell} \\sum_ {i = 1} ^ {\\ell} \\operatorname {C E} \\left(\\text {L L M} \\left(\\mathbf {z}, \\mathbf {w} _ {1: i - 1}\\right), w _ {i}\\right). \\tag {10}\n$$\n", + "text_format": "latex", + "bbox": [ + 401, + 438, + 880, + 474 + ], + "page_idx": 9 + }, + { + "type": "text", + "text": "Similar to stage 2, we merge the LoRA trained in the stage 1 and stage 2 with the original model and introduce a new LoRA module.", + "bbox": [ + 174, + 482, + 880, + 513 + ], + "page_idx": 9 + }, + { + "type": "text", + "text": "5 Experiments", + "text_level": 1, + "bbox": [ + 174, + 537, + 334, + 555 + ], + "page_idx": 9 + }, + { + "type": "text", + "text": "5.1 Experimental Setup", + "text_level": 1, + "bbox": [ + 174, + 570, + 387, + 585 + ], + "page_idx": 9 + }, + { + "type": "text", + "text": "Tasks, Datasets, and Evaluation Metrics. We conduct a quantitative evaluation of LVidLMs' video understanding capabilities across three tasks: (1) Video Question Answering: This task assesses the comprehensive video comprehension abilities of LVidLMs by requiring the model to answer a variety of questions about the video content based on its understanding. We perform this task on three datasets: MSVD-QA [41], MSRVTT-QA [43], and ActivityNet-QA [45]. The evaluation pipeline for video understanding follows Video-ChatGPT [26], and we report the accuracy and score, which is assessed using GPT-Assistant [28]. (2) Video Temporal Grounding: This task evaluates LVidLMs' capacity to discern the starting and ending timestamps of a segment corresponding to the description of a video clip. This task demands the model to effectively grasp the temporal aspects of the video. We conduct this task on the ActivityNet Captions dataset [15] and calculate Intersection over Union (IoU) between the model-generate time segments and the ground truth time segments. We report mean IoU (mIoU) and Recall@1, IoU $\\geq m$ (R@m) metric, where $m$ values are set at $\\{0.3, 0.5, 0.7\\}$ . (3) Video Dense Captioning: This task requires the model to produce all events depict in the video along with their corresponding start and end timestamps. It necessitates the model to comprehend both the spatial and temporal dimensions of the video simultaneously.", + "bbox": [ + 173, + 595, + 882, + 824 + ], + "page_idx": 9 + }, + { + "type": "page_number", + "text": "10", + "bbox": [ + 178, + 89, + 196, + 99 + ], + "page_idx": 9 + }, + { + "type": "header", + "text": "Y. Liu, P. Ding et al.", + "bbox": [ + 230, + 88, + 374, + 102 + ], + "page_idx": 9 + }, + { + "type": "table", + "img_path": "images/56fcf122d532a45d1d2bf3f98b6624c49b21875e38d0ad1b44ea7f7ee130063b.jpg", + "table_caption": [ + "Table 4: Comparison between different LVidLMs in temporal video grounding and dense video captioning tasks on ActivityNet [11]." + ], + "table_footnote": [], + "table_body": "
ModelLLM SizeTemporal GroundingDense Captioning
R@0.3↑R@0.5↑R@0.7↑mIoU↑SODA_c↑CIDEr↑METEOR↑
VideoChat [19]7B8.83.71.57.20.92.20.9
Video-LLaMA [46]7B6.92.10.86.51.95.81.9
Video-ChatGPT [26]7B26.413.66.118.91.95.82.1
PiTe (Ours)7B30.417.87.822.05.121.75.8
PiTe (Ours)13B37.223.710.926.05.926.56.6
", + "bbox": [ + 120, + 157, + 818, + 296 + ], + "page_idx": 10 + }, + { + "type": "text", + "text": "We conduct this task on the ActivityNet Captions dataset [15]. Initially, we reported SODA_c [10], followed by averages of CIDEr [39] and METEOR [16] under different IoU thresholds of 0.3, 0.5, 0.7, 0.9 based on generate events and ground truth matched pairs to provide a comprehensive analysis. In this paper, all experiments were conducted in a zero-shot setting, and higher values of all evaluation metrics indicate superior performance.", + "bbox": [ + 112, + 321, + 820, + 398 + ], + "page_idx": 10 + }, + { + "type": "text", + "text": "Implementation Details. In this paper, we employ Vicuna v1.5 [5] as LLM to train the PiTe model at two scales: 7B and 13B. Leveraging the efficiency of LoRA [12], the training of the 7B model can be completed in approximately 10 hours using a single Nvidia 8-A100 (80GB VRAM) node, while the 13B model requires around 17 hours. More hyper-parameter settings are shwon in Table 2.", + "bbox": [ + 112, + 419, + 820, + 494 + ], + "page_idx": 10 + }, + { + "type": "text", + "text": "5.2 Main Result", + "text_level": 1, + "bbox": [ + 114, + 515, + 264, + 530 + ], + "page_idx": 10 + }, + { + "type": "text", + "text": "Table 3 and 4 present the comparative performance of the PiTe against state-of-the-art baselines on myriad video understanding datasets.", + "bbox": [ + 112, + 539, + 818, + 570 + ], + "page_idx": 10 + }, + { + "type": "text", + "text": "Question Answering. As illustrated in Table 3, PiTe consistently outperforms the state-of-the-art pure instruction-tuning baselines in terms of all metrics on all datasets. Compared to the top-performing baselines in each dataset, PiTe exhibited notable improvements in the average question answering accuracy, achieving a maximum enhancement of 4.8 and an average improvement of 3.7. For example, PiTe substantially improves accuracy by 64.9 to 68.4 compared to Video-ChatGPT [26] in MSVD-QA dataset [41]. The results showcasing PiTe's proficiency in video comprehension and its capacity to deliver contextually relevant responses according to the given instructions.", + "bbox": [ + 112, + 590, + 820, + 696 + ], + "page_idx": 10 + }, + { + "type": "text", + "text": "Temporal Grounding. As depicted in Table 4, PiTe achieves state-of-the-art performance in the video temporal grounding task across all metrics as well, demonstrating improvements ranging from 18.9 to 22.0 in mIoU compared to Video-ChatGPT [26]. This clearly indicates that trajectory alignment greatly enhances the ability of capture events in temporal dimension for LVidLMs. The incorporation of object trajectories in the temporal dimension of the trajectory matrix equips the model with a precise understanding of temporal event boundaries, thereby establishing a solid foundation for accurate event localization.", + "bbox": [ + 112, + 717, + 821, + 823 + ], + "page_idx": 10 + }, + { + "type": "header", + "text": "PiTe: Pixel-Temporal Alignment for Large Video-Language Model", + "bbox": [ + 325, + 88, + 767, + 103 + ], + "page_idx": 10 + }, + { + "type": "page_number", + "text": "11", + "bbox": [ + 802, + 90, + 818, + 101 + ], + "page_idx": 10 + }, + { + "type": "table", + "img_path": "images/1e48b5bf55078838c965fa8968d872a97b61d88744ea22bc63528999e523e4b4.jpg", + "table_caption": [ + "Table 5: Ablation study of the three-stage training strategy." + ], + "table_footnote": [], + "table_body": "
MethodMSVD-QAActivityNet
Temporal GroundingDense Captioning
Accuracy ScoreR@0.3R@0.5R@0.7mIoUSODA_cCIDErMETEOR
PiTe (Ours)68.43.930.417.87.822.05.121.75.8
w/o initialize68.23.922.810.54.617.15.121.75.8
w/o trajectory68.13.923.912.85.717.45.021.45.8
", + "bbox": [ + 181, + 143, + 875, + 277 + ], + "page_idx": 11 + }, + { + "type": "text", + "text": "Dense Captioning. The outcomes of the dense captioning task, as delineated in Table 4, reveal that PiTe consistent boost compared to all state-of-the-art baselines. Particularly noteworthy is the substantial 15.9 increase in the CIDEr metric [39] when compared to Video-ChatGPT [26]. This underscores the significance of fine-grained alignment in both spatial and temporal dimensions through trajectories, implying that PiTe acquires more generalized and detailed representations to offer more sophisticated event descriptions and accurate event temporal boundaries.", + "bbox": [ + 173, + 306, + 880, + 398 + ], + "page_idx": 11 + }, + { + "type": "text", + "text": "5.3 Analysis", + "text_level": 1, + "bbox": [ + 174, + 422, + 292, + 436 + ], + "page_idx": 11 + }, + { + "type": "text", + "text": "Ablation Study. As reported in Table 5, we conduct ablation experiments on MVSD-QA [41] for question answering and ActivityNet Captions [15] for temporal grounding to verify the individual effects of the proposed contributions under the following settings: (1) w/o initialize: we remove the initialization strategy that use weight of localization projector to initialize trajectory projector; (2) w/o trajectory: we abandon the fine-grained alignment strategy via trajectory.", + "bbox": [ + 173, + 449, + 880, + 525 + ], + "page_idx": 11 + }, + { + "type": "text", + "text": "From the experimental results in Table 5, it can be observed that: (1) Eliminating the initialization strategy for the trajectory projector in PiTe reduces the model's reasoning capabilities and temporal boundary awareness. However, the performance in dense captioning generation remains consistent. This observation suggests that the model maintains its basic ability in comprehending visual content under trajectory-guided training. (2) The removal of the trajectory-guided training diminishes almost all the capability of PiTe, including dense captioning. (3) Without trajectory-guided training, PiTe demonstrates superior performance compared to trajectory-guided training without the initialization strategy for the trajectory projector in temporal grounding. This outcome highlights the difficulty of trajectory-guided training without initialization from a pre-trained localization projector, as the instability of parameters can impede the model's perception to accurately perceive visual temporal information.", + "bbox": [ + 173, + 526, + 880, + 691 + ], + "page_idx": 11 + }, + { + "type": "text", + "text": "Exhibition. To better illustrate the video dialogue performance of PiTe, we present a qualitative example, as shown in Fig. 5a. The illustration from the upper portion of the figure demonstrates PiTe's capability not only to provide precise responses to instruction queries but also to enhance the output with more detailed and accurate video information. The example in the lower segment of the figure highlights the model's proficiency in understanding instruction and capturing event, enabling precise delineation of temporal boundaries within the video, despite the constraint of a 100-frame sampling limit.", + "bbox": [ + 173, + 715, + 880, + 823 + ], + "page_idx": 11 + }, + { + "type": "page_number", + "text": "12", + "bbox": [ + 176, + 89, + 194, + 99 + ], + "page_idx": 11 + }, + { + "type": "header", + "text": "Y. Liu, P. Ding et al.", + "bbox": [ + 230, + 88, + 374, + 102 + ], + "page_idx": 11 + }, + { + "type": "image", + "img_path": "images/e76a39fe0b891717df0d44f7843fa486ca7779662b9d3a4aff474b487a9ebcbd.jpg", + "image_caption": [ + "[Instruction] Who are shown performing impressive feats?" + ], + "image_footnote": [], + "bbox": [ + 133, + 133, + 426, + 188 + ], + "page_idx": 12 + }, + { + "type": "image", + "img_path": "images/b9ee5ca2d267772e80f9143a833f9eb65a4cf5c5520d1a49cb8e1587f2a53ab0.jpg", + "image_caption": [ + "[Response] The video showcases various football players performing impressive feats, including a player jumping over another player and a player making a spectacular catch.", + "[Instruction] During which frames can we see he then shines the shoes of one man who pays him and shakes his hand.", + "[Response] The shoe shining and payment occur from 66 to 99", + "(116.348s ~ 174.522s). Ground truth: 116.06s ~ 172.78s" + ], + "image_footnote": [], + "bbox": [ + 135, + 241, + 426, + 273 + ], + "page_idx": 12 + }, + { + "type": "image", + "img_path": "images/4c6152d1ab3082d20f1d939cae6b4688ccc0a009afb2db382e985e60a717d404.jpg", + "image_caption": [ + "(a) Examples of PiTe's video understanding capabilities.", + "(b) Performance comparison for different tracking point quantity.", + "Fig. 5: PiTe's video understanding capabilities and performance comparison across varying tracking point quantities." + ], + "image_footnote": [], + "bbox": [ + 467, + 123, + 816, + 314 + ], + "page_idx": 12 + }, + { + "type": "text", + "text": "Impact of Tracking Point Quantity. In Fig. 5b, we vary the tracking point quantity $P$ in set of $\\{1,3,5\\}$ . The efficacy of dense captioning tasks tends to improve with an increase in tracking points. However, it is observed that the temporal grounding task undergoes an initial substantial improvement, only to be succeeded by a rapid decline. Less number of tracking points fails to accurately capture the object's geometry, thereby hindering the pixel-level cross-modal alignment guidance for the model. Conversely, a higher quantity of points can enhance the model's comprehension of pure visual information; however, it also introduces noise to make training more challenging. Overall, that the optimal value of $P$ may be different for different tasks, we set $P = 3$ due to its performance maintain stability over multiple tasks.", + "bbox": [ + 112, + 390, + 821, + 526 + ], + "page_idx": 12 + }, + { + "type": "text", + "text": "6 Conclusion", + "text_level": 1, + "bbox": [ + 114, + 546, + 259, + 561 + ], + "page_idx": 12 + }, + { + "type": "text", + "text": "In this paper, we focus on enhancing the performance of Large Video-Language Models (LVidLMs) by incorporating trajectory-based alignment across different modalities. To achieve fine-grained alignment between video and language across spatial and temporal dimensions, we initially curate a comprehensive multi-modal object tracking dataset, PiTe-143k, using a fully automated annotation pipeline. This dataset was developed to address the lack of large-scale video-language datasets that include multi-object moving trajectories. Subsequently, we introduce a novel Pixel-Temporal (PiTe) alignment strategy that leverages trajectory-guided pre-training to address the inherent challenges faced by LVidLMs. Through comparative analyses, we evaluate PiTe against state-of-the-art models and competitive baselines across various tasks in a zero-shot setting, including question-answering, temporal grounding, and dense captioning, showcasing the superior performance of PiTe with more sophisticated event descriptions and accurate event temporal boundaries.", + "bbox": [ + 112, + 575, + 821, + 743 + ], + "page_idx": 12 + }, + { + "type": "text", + "text": "Acknowledgments", + "text_level": 1, + "bbox": [ + 114, + 763, + 292, + 780 + ], + "page_idx": 12 + }, + { + "type": "text", + "text": "This work was supported by the National Science and Technology Innovation 2030 - Major Project (Grant No. 2022ZD0208800), and NSFC General Program (Grant No. 62176215).", + "bbox": [ + 114, + 792, + 820, + 823 + ], + "page_idx": 12 + }, + { + "type": "header", + "text": "PiTe: Pixel-Temporal Alignment for Large Video-Language Model", + "bbox": [ + 325, + 88, + 767, + 103 + ], + "page_idx": 12 + }, + { + "type": "page_number", + "text": "13", + "bbox": [ + 802, + 89, + 821, + 101 + ], + "page_idx": 12 + }, + { + "type": "text", + "text": "References", + "text_level": 1, + "bbox": [ + 176, + 118, + 282, + 133 + ], + "page_idx": 13 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "1. Alayrac, J., Donahue, J., Luc, P., Miech, A., Barr, I., Hasson, Y., Lenc, K., Mensch, A., Millican, K., Reynolds, M., Ring, R., Rutherford, E., Cabi, S., Han, T., Gong, Z., Samangooei, S., Monteiro, M., Menick, J.L., Borgeaud, S., Brock, A., Nematzadeh, A., Sharifzadeh, S., Binkowski, M., Barreira, R., Vinyals, O., Zisserman, A., Simonyan, K.: Flamingo: a visual language model for few-shot learning. In: Proc. of NeurIPS (2022)", + "2. Arthur, D., Vassilvitskii, S.: k-means++: the advantages of careful seeding. In: Proc. of SODA. pp. 1027-1035 (2007)", + "3. Awadalla, A., Gao, I., Gardner, J., Hessel, J., Hanafy, Y., Zhu, W., Marathe, K., Bitton, Y., Gadre, S.Y., Sagawa, S., Jitsev, J., Kornblith, S., Koh, P.W., Ilharco, G., Wortman, M., Schmidt, L.: Openflamingo: An open-source framework for training large autoregressive vision-language models. CoRR (2023)", + "4. Brown, T.B., Mann, B., Ryder, N., Subbiah, M., Kaplan, J., Dhariwal, P., Neelakantan, A., Shyam, P., Sastry, G., Askell, A., Agarwal, S., Herbert-Voss, A., Krueger, G., Henighan, T., Child, R., Ramesh, A., Ziegler, D.M., Wu, J., Winter, C., Hesse, C., Chen, M., Sigler, E., Litwin, M., Gray, S., Chess, B., Clark, J., Berner, C., McCandlish, S., Radford, A., Sutskever, I., Amodei, D.: Language models are few-shot learners. In: Proc. of NeurIPS (2020)", + "5. Chiang, W.L., Li, Z., Lin, Z., Sheng, Y., Wu, Z., Zhang, H., Zheng, L., Zhuang, S., Zhuang, Y., Gonzalez, J.E., Stoica, I., Xing, E.P.: Vicuna: An open-source chatbot impressing gpt-4 with $90\\%$ * chatgpt quality (2023)", + "6. Dai, W., Li, J., Li, D., Tiong, A.M.H., Zhao, J., Wang, W., Li, B., Fung, P., Hoi, S.C.H.: Instructblip: Towards general-purpose vision-language models with instruction tuning. In: Proc. of NeurIPS (2023)", + "7. Devlin, J., Chang, M., Lee, K., Toutanova, K.: BERT: pre-training of deep bidirectional transformers for language understanding. In: Proc. of AACL. pp. 4171-4186 (2019)", + "8. Dosovitskiy, A., Beyer, L., Kolesnikov, A., Weissenborn, D., Zhai, X., Unterthiner, T., Dehghani, M., Minderer, M., Heigold, G., Gelly, S., Uszkoreit, J., Houlsby, N.: An image is worth 16x16 words: Transformers for image recognition at scale. In: Proc. of ICLR (2021)", + "9. Du, Z., Qian, Y., Liu, X., Ding, M., Qiu, J., Yang, Z., Tang, J.: GLM: general language model pretraining with autoregressive blank infilling. In: Proc. of ACL. pp. 320-335 (2022)", + "10. Fujita, S., Hirao, T., Kamigaito, H., Okumura, M., Nagata, M.: SODA: story oriented dense video captioning evaluation framework. In: Proc. of ECCV. pp. 517-531 (2020)", + "11. Heilbron, F.C., Escorcia, V., Ghanem, B., Niebles, J.C.: Activitynet: A large-scale video benchmark for human activity understanding. In: Proc. of CVPR. pp. 961-970 (2015)", + "12. Hu, E.J., Shen, Y., Wallis, P., Allen-Zhu, Z., Li, Y., Wang, S., Wang, L., Chen, W.: Lora: Low-rank adaptation of large language models. In: Proc. of ICLR (2022)", + "13. Huang, B., Wang, X., Chen, H., Song, Z., Zhu, W.: Vtimellm: Empower llm to grasp video moments. In: Proc. of CVPR. pp. 14271-14280 (2024)", + "14. Kaplan, J., McCandlish, S., Henighan, T., Brown, T.B., Chess, B., Child, R., Gray, S., Radford, A., Wu, J., Amodei, D.: Scaling laws for neural language models. CoRR (2020)", + "15. Krishna, R., Hata, K., Ren, F., Fei-Fei, L., Niebles, J.C.: Dense-c captioning events in videos. In: Proc. of ICCV. pp. 706-715 (2017)", + "16. Lavie, A., Agarwal, A.: METEOR: an automatic metric for MT evaluation with high levels of correlation with human judgments. In: Proc. of WMT@ACL. pp. 228-231 (2007)", + "17. Le Moing, G., Ponce, J., Schmid, C.: Dense optical tracking: Connecting the dots. In: Proc. of CVPR. pp. 19187-19197 (2024)", + "18. Li, J., Li, D., Savarese, S., Hoi, S.C.H.: BLIP-2: bootstrapping language-image pre-training with frozen image encoders and large language models. In: Proc. of ICML. pp. 19730-19742 (2023)", + "19. Li, K., He, Y., Wang, Y., Li, Y., Wang, W., Luo, P., Wang, Y., Wang, L., Qiao, Y.: Videochat: Chatcentric video understanding. CoRR (2023)", + "20. Lin, J., Yin, H., Ping, W., Molchanov, P., Shoeybi, M., Han, S.: Vila: On pre-training for visual language models. In: Proc. of CVPR. pp. 26689-26699 (2024)" + ], + "bbox": [ + 176, + 148, + 880, + 821 + ], + "page_idx": 13 + }, + { + "type": "page_number", + "text": "14", + "bbox": [ + 176, + 89, + 196, + 99 + ], + "page_idx": 13 + }, + { + "type": "header", + "text": "Y. Liu, P. Ding et al.", + "bbox": [ + 230, + 88, + 374, + 102 + ], + "page_idx": 13 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "21. Liu, H., Li, C., Li, Y., Lee, Y.J.: Improved baselines with visual instruction tuning. In: Proc. of CVPR. pp. 26296-26306 (2024)", + "22. Liu, H., Li, C., Wu, Q., Lee, Y.J.: Visual instruction tuning. In: Proc. of NeurIPS (2023)", + "23. Liu, Y., Ping, H., Zhang, D., Sun, Q., Li, S., Zhou, G.: Comment-aware multi-modal heterogeneous pre-training for humor detection in short-form videos. In: Proc. of ECAI. pp. 1568-1575 (2023)", + "24. Liu, Y., Shen, T., Zhang, D., Sun, Q., Li, S., Zhou, G.: Comment-aided video-language alignment via contrastive pre-training for short-form video humor detection. In: Proc. of ICMR. pp. 442-450 (2024)", + "25. Luo, R., Zhao, Z., Yang, M., Dong, J., Qiu, M., Lu, P., Wang, T., Wei, Z.: Valley: Video assistant with large language model enhanced ability. CoRR (2023)", + "26. Maaz, M., Rasheed, H.A., Khan, S.H., Khan, F.S.: Video-chatgpt: Towards detailed video understanding via large vision and language models (2024)", + "27. Munasinghe, S., Thushara, R., Maaz, M., Rasheed, H.A., Khan, S., Shah, M., Khan, F.S.: Pg-video-llava: Pixel grounding large video-language models. CoRR (2023)", + "28. OpenAI: Chatgpt: Optimizing language models for dialogue (2023)", + "29. Ouyang, L., Wu, J., Jiang, X., Almeida, D., Wainwright, C.L., Mishkin, P., Zhang, C., Agarwal, S., Slama, K., Ray, A., Schulman, J., Hilton, J., Kelton, F., Miller, L., Simens, M., Askell, A., Welinder, P., Christiano, P.F., Leike, J., Lowe, R.: Training language models to follow instructions with human feedback. In: Proc. of NeurIPS (2022)", + "30. Pont-Tuset, J., Uijlings, J.R.R., Changpinyo, S., Soricut, R., Ferrari, V.: Connecting vision and language with localized narratives. In: Proc. of ECCV. pp. 647-664 (2020)", + "31. Radford, A., Kim, J.W., Hallacy, C., Ramesh, A., Goh, G., Agarwal, S., Sastry, G., Askell, A., Mishkin, P., Clark, J., Krueger, G., Sutskever, I.: Learning transferable visual models from natural language supervision. In: Proc. of ICML. pp. 8748-8763 (2021)", + "32. Radford, A., Narasimhan, K.: Improving language understanding by generative pre-training (2018)", + "33. Radford, A., Wu, J., Child, R., Luan, D., Amodei, D., Sutskever, I.: Language models are unsupervised multitask learners (2019)", + "34. Raffel, C., Shazeer, N., Roberts, A., Lee, K., Narang, S., Matena, M., Zhou, Y., Li, W., Liu, P.J.: Exploring the limits of transfer learning with a unified text-to-text transformer. J. Mach. Learn. Res. pp. 140:1-140:67 (2020)", + "35. Rasheed, H., Maaz, M., Shaji, S., Shaker, A., Khan, S., Cholakkal, H., Anwer, R.M., Xing, E., Yang, M.H., Khan, F.S.: Glamm: Pixel grounding large multimodal model. In: Proc. of CVPR. pp. 13009-13018 (2024)", + "36. Scao, T.L., Fan, A., Akiki, C., Pavlick, E., Ilic, S., Hesslow, D., Castagné, R., Luccioni, A.S., Yvon, F., Gallé, M., Tow, J., Rush, A.M., Biderman, S., Webson, A., Ammanamanchi, P.S., Wang, T., Sagot, B., Muennighoff, N., del Moral, A.V., Ruwase, O., Bawden, R., Bekman, S., McMillan-Major, A., Beltagy, I., Nguyen, H., Saulnier, L., Tan, S., Suarez, P.O., Sanh, V., Laurençon, H., Jernite, Y., Launay, J., Mitchell, M., Raffel, C., Gokaslan, A., Simhi, A., Soroa, A., Aji, A.F., Alfassy, A., Rogers, A., Nitzav, A.K., Xu, C., Mou, C., Emezue, C., Klamm, C., Leong, C., van Strien, D., Adelani, D.I., et al.: BLOOM: A 176b-parameter open-access multilingual language model. CoRR (2022)", + "37. Touvron, H., Lavril, T., Izacard, G., Martinet, X., Lachaux, M., Lacroix, T., Rozière, B., Goyal, N., Hambro, E., Azhar, F., Rodriguez, A., Joulin, A., Grave, E., Lample, G.: Llama: Open and efficient foundation language models. CoRR (2023)", + "38. Touvron, H., Martin, L., Stone, K., Albert, P., Almahairi, A., Babaei, Y., Bashlykov, N., Batra, S., Bhargava, P., Bhosale, S., Bikel, D., Blecher, L., Canton-Ferrer, C., Chen, M., Cucurull, G., Esiobu, D., Fernandes, J., Fu, J., Fu, W., Fuller, B., Gao, C., Goswami, V., Goyal, N., Hartshorn, A., Hosseini, S., Hou, R., Inan, H., Kardas, M., Kerkez, V., Khabsa, M., Kloumann, I., Korenev, A., Koura, P.S., Lachaux, M., Lavril, T., Lee, J., Liskovich, D., Lu, Y., Mao, Y., Martinet, X., Mihaylov, T., Mishra, P., Molybog, I., Nie, Y., Poulton, A., Reizenstein, J., Rungta, R., Saladi, K., Schelten, A., Silva, R., Smith, E.M., Subramanian, R., Tan, X.E., Tang, B., Taylor, R., Williams, A., Kuan, J.X., Xu, P., Yan, Z., Zarov, I., Zhang, Y., Fan, A., Kambadur, M., Narang, S., Rodriguez, A., Stojnic, R., Edunov, S., Scialom, T.: Llama 2: Open foundation and fine-tuned chat models. CoRR (2023)" + ], + "bbox": [ + 116, + 119, + 820, + 823 + ], + "page_idx": 14 + }, + { + "type": "header", + "text": "PiTe: Pixel-Temporal Alignment for Large Video-Language Model", + "bbox": [ + 325, + 88, + 767, + 103 + ], + "page_idx": 14 + }, + { + "type": "page_number", + "text": "15", + "bbox": [ + 802, + 89, + 818, + 99 + ], + "page_idx": 14 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "39. Vedantam, R., Zitnick, C.L., Parikh, D.: Cider: Consensus-based image description evaluation. In: Proc. of CVPR. pp. 4566-4575 (2015)", + "40. Wang, Y., He, Y., Li, Y., Li, K., Yu, J., Ma, X., Li, X., Chen, G., Chen, X., Wang, Y., Luo, P., Liu, Z., Wang, Y., Wang, L., Qiao, Y.: Internvid: A large-scale video-text dataset for multimodal understanding and generation. In: Proc. of ICLR (2024)", + "41. Xu, D., Zhao, Z., Xiao, J., Wu, F., Zhang, H., He, X., Zhuang, Y.: Video question answering via gradually refined attention over appearance and motion. In: Proceedings of the 2017 ACM on Multimedia Conference, MM 2017, Mountain View, CA, USA, October 23-27, 2017. pp. 1645-1653 (2017)", + "42. Xu, J., Zhou, X., Yan, S., Gu, X., Arnab, A., Sun, C., Wang, X., Schmid, C.: Pixel-aligned language model. In: Proc. of CVPR. pp. 13030-13039 (2024)", + "43. Xu, J., Mei, T., Yao, T., Rui, Y.: MSR-VTT: A large video description dataset for bridging video and language. In: Proc. of CVPR. pp. 5288-5296 (2016)", + "44. Yang, A., Miech, A., Sivic, J., Laptev, I., Schmid, C.: Zero-shot video question answering via frozen bidirectional language models. In: Proc. of NeurIPS (2022)", + "45. Yu, Z., Xu, D., Yu, J., Yu, T., Zhao, Z., Zhuang, Y., Tao, D.: Activitynet-qa: A dataset for understanding complex web videos via question answering. In: Proc. of AAAI. pp. 9127-9134 (2019)", + "46. Zhang, H., Li, X., Bing, L.: Video-llama: An instruction-tuned audio-visual language model for video understanding. In: Proc. of EMNLP. pp. 543-553 (2023)", + "47. Zhang, M., Huang, S., Li, W., Wang, D.: Tree structure-aware few-shot image classification via hierarchical aggregation. In: Proc. of ECCV. pp. 453-470 (2022)", + "48. Zhang, R., Han, J., Liu, C., Zhou, A., Lu, P., Qiao, Y., Li, H., Gao, P.: LLaMA-adapter: Efficient fine-tuning of large language models with zero-initialized attention. In: Proc. of ICLR (2024)", + "49. Zhang, S., Roller, S., Goyal, N., Artetxe, M., Chen, M., Chen, S., Dewan, C., Diab, M.T., Li, X., Lin, X.V., Mihaylov, T., Ott, M., Shleifer, S., Shuster, K., Simig, D., Koura, P.S., Sridhar, A., Wang, T., Zettlemoyer, L.: OPT: open pre-trained transformer language models. CoRR (2022)", + "50. Zhang, Y., Li, Z., Zhang, M.: Efficient second-order treecrf for neural dependency parsing. In: Proc. of ACL. pp. 3295-3305 (2020)", + "51. Zhang, Y., Zhou, H., Li, Z.: Fast and accurate neural CRF constituency parsing. In: Proc. of IJCAI. pp. 4046-4053 (2020)", + "52. Zhu, D., Chen, J., Shen, X., Li, X., Elhoseiny, M.: MiniGPT-4: Enhancing vision-language understanding with advanced large language models. In: Proc. of ICLR (2024)" + ], + "bbox": [ + 176, + 119, + 880, + 550 + ], + "page_idx": 15 + }, + { + "type": "page_number", + "text": "16", + "bbox": [ + 176, + 89, + 196, + 99 + ], + "page_idx": 15 + }, + { + "type": "header", + "text": "Y. Liu, P. Ding et al.", + "bbox": [ + 230, + 88, + 374, + 102 + ], + "page_idx": 15 + } +] \ No newline at end of file diff --git a/2024/PiTe_ Pixel-Temporal Alignment for Large Video-Language Model/3468332a-054d-4ee9-a3ba-3d30cf6d6594_model.json b/2024/PiTe_ Pixel-Temporal Alignment for Large Video-Language Model/3468332a-054d-4ee9-a3ba-3d30cf6d6594_model.json new file mode 100644 index 0000000000000000000000000000000000000000..408561af20d657b9575ff860076b9433243f6630 --- /dev/null +++ b/2024/PiTe_ Pixel-Temporal Alignment for Large Video-Language Model/3468332a-054d-4ee9-a3ba-3d30cf6d6594_model.json @@ -0,0 +1,2476 @@ +[ + [ + { + "type": "title", + "bbox": [ + 0.263, + 0.115, + 0.674, + 0.16 + ], + "angle": 0, + "content": "PiTe: Pixel-Temporal Alignment for Large Video-Language Model" + }, + { + "type": "text", + "bbox": [ + 0.245, + 0.186, + 0.691, + 0.218 + ], + "angle": 0, + "content": "Yang Liu\\(^{1,2\\star}\\), Pengxiang Ding\\(^{1\\star}\\), Siteng Huang\\(^{1}\\), Min Zhang\\(^{1}\\), Han Zhao\\(^{1}\\), and Donglin Wang\\(^{1\\boxtimes}\\)" + }, + { + "type": "text", + "bbox": [ + 0.244, + 0.228, + 0.671, + 0.258 + ], + "angle": 0, + "content": "\\(^{1}\\)Westlake University \\(^{2}\\)Soochow University {liuyang67, dingpengxiang, wangdonglin}@westlake.edu.cn" + }, + { + "type": "text", + "bbox": [ + 0.165, + 0.295, + 0.774, + 0.502 + ], + "angle": 0, + "content": "Abstract. Fueled by the Large Language Models (LLMs) wave, Large Visual-Language Models (LVLMs) have emerged as a pivotal advancement, bridging the gap between image and text. However, video making it challenging for LVLMs to perform adequately due to the complexity of the relationship between language and spatial-temporal data structure. Recent Large Video-Language Models (LVidLMs) align feature of static visual data like image into latent space of language feature, by general multi-modal tasks to leverage abilities of LLMs sufficiently. In this paper, we explore fine-grained alignment approach via object trajectory for different modalities across both spatial and temporal dimensions simultaneously. Thus, we propose a novel LVidLM by trajectory-guided Pixel-Temporal Alignment, dubbed PiTe, that exhibits promising applicable model property. To achieve fine-grained video-language alignment, we curate a multi-modal pre-training dataset PiTe-143k, the dataset provision of moving trajectories in pixel level for all individual objects, that appear and mention in the video and caption both, by our automatic annotation pipeline. Meanwhile, PiTe demonstrates astounding capabilities on myriad video-related multi-modal tasks through beat the state-of-the-art methods by a large margin." + }, + { + "type": "text", + "bbox": [ + 0.165, + 0.517, + 0.774, + 0.545 + ], + "angle": 0, + "content": "Keywords: Large Video-Language Model \\(\\cdot\\) Trajectory-guided Instruction Tuning \\(\\cdot\\) Video Understanding" + }, + { + "type": "title", + "bbox": [ + 0.119, + 0.571, + 0.273, + 0.587 + ], + "angle": 0, + "content": "1 Introduction" + }, + { + "type": "text", + "bbox": [ + 0.118, + 0.602, + 0.82, + 0.738 + ], + "angle": 0, + "content": "Large Language Models (LLMs) have rapidly gained popularity within the AI community, demonstrating astounding capabilities across a wide array of natural language tasks [4, 5, 9, 29, 37, 38]. The powerful language comprehension abilities of LLMs drive researchers to explore their utility in addressing a broader spectrum of tasks across various domains. Consequently, an increasing number of studies are focusing on developing comprehensive Large Visual-Language Models (LVLMs) to tackle vision-related tasks in zero-shot settings [6, 18, 47, 52], particularly in the video understanding [19, 23, 26, 27, 44, 46, 48]. The pursuit of generalist Large Video-Language Models (LVidLMs) will be a perennial challenge. Success in this endeavor hinges on effectively leveraging the exceptional understanding, reasoning, and generative capacities inherently present in LLMs." + }, + { + "type": "text", + "bbox": [ + 0.118, + 0.74, + 0.82, + 0.8 + ], + "angle": 0, + "content": "One potential route towards addressing the issue is aligning visual feature into latent space of language feature. To achieve this, existing LVidLMs apply large-scale vanilla instruction tuning [19, 26,46,48]. However, the conventional question-answering training paradigm primarily assists LLMs in understanding visual data from a spatial perspective, posing challenges in effectively capturing" + }, + { + "type": "page_footnote", + "bbox": [ + 0.135, + 0.808, + 0.46, + 0.823 + ], + "angle": 0, + "content": "* Equal contribution. ☑ Corresponding author." + } + ], + [ + { + "type": "page_number", + "bbox": [ + 0.177, + 0.09, + 0.19, + 0.101 + ], + "angle": 0, + "content": "2" + }, + { + "type": "header", + "bbox": [ + 0.231, + 0.089, + 0.375, + 0.103 + ], + "angle": 0, + "content": "Y. Liu, P. Ding et al." + }, + { + "type": "image", + "bbox": [ + 0.19, + 0.119, + 0.614, + 0.247 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.255, + 0.25, + 0.538, + 0.262 + ], + "angle": 0, + "content": "(a) Illustration of training paradigm comparison." + }, + { + "type": "image", + "bbox": [ + 0.643, + 0.118, + 0.882, + 0.248 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.669, + 0.25, + 0.839, + 0.262 + ], + "angle": 0, + "content": "(b) Performance comparison." + }, + { + "type": "image_caption", + "bbox": [ + 0.175, + 0.273, + 0.88, + 0.302 + ], + "angle": 0, + "content": "Fig. 1: Comparison with existing LVidLMs in terms of alignment paradigm and performance. For Fig. 1b, \\(QA\\), \\(TG\\), \\(DC\\) denote question answering, temporal grounding and dense captioning, respectively." + }, + { + "type": "text", + "bbox": [ + 0.174, + 0.321, + 0.881, + 0.397 + ], + "angle": 0, + "content": "temporal dynamics and spatial consistency relationships. Therefore, relying solely on instruction tuning proves insufficient for achieving comprehensive video comprehension, given the intricate spatial-temporal data structure involved. It is crucial to align different modalities across both spatial and temporal dimensions. Furthermore, offering more fine-grained cross-modal alignment guidance significantly enhances LVidLMs' ability to comprehend videos [24]." + }, + { + "type": "text", + "bbox": [ + 0.174, + 0.397, + 0.882, + 0.488 + ], + "angle": 0, + "content": "To bridge the gap, we introduce a novel LVidLM named PiTe, which employs trajectories to intricately align vision and language across both spatial and temporal dimensions at the pixel level, and the distinction from conventional approaches is illustrated in Figure 1a. By requiring the model to forecast the trajectory of individual objects mentioned in the text within the video, it enables the learning of fine-grained text-to-pixel alignment through exploiting the video context along the temporal dimension and enhancing its ability to generate output based on evidence." + }, + { + "type": "text", + "bbox": [ + 0.174, + 0.488, + 0.882, + 0.577 + ], + "angle": 0, + "content": "Subsequently, due to there are no ready-made video-language dataset with moving trajectory of objects, we curate a large-scale video-language dataset PiTe-143k through an automated annotation pipeline. Consequently, as shown in Figure 1b, the proposed PiTe significantly augments the LVidLM's capacity to understand videos comprehensively, leading to promising, competitive, and state-of-the-art performance in question-answering, temporal grounding, and dense captioning tasks under zero-shot conditions." + }, + { + "type": "text", + "bbox": [ + 0.197, + 0.578, + 0.749, + 0.593 + ], + "angle": 0, + "content": "Overall, our principal contributions in this paper are summarized as follows:" + }, + { + "type": "text", + "bbox": [ + 0.175, + 0.593, + 0.88, + 0.623 + ], + "angle": 0, + "content": "- We curate a large scale video-language dataset PiTe-143k with trajectory for all individual objects by automatic annotation pipeline." + }, + { + "type": "text", + "bbox": [ + 0.175, + 0.623, + 0.881, + 0.653 + ], + "angle": 0, + "content": "- We propose a novel LVidLM PiTe that utilize trajectory to align video and language features across both spatial and temporal dimensions." + }, + { + "type": "text", + "bbox": [ + 0.175, + 0.654, + 0.881, + 0.684 + ], + "angle": 0, + "content": "- Extensive experimental results and analysis on myriad datasets for zero-shot video question answering, temporal grounding, and dense captioning tasks demonstrate the superiority of PiTe." + }, + { + "type": "list", + "bbox": [ + 0.175, + 0.593, + 0.881, + 0.684 + ], + "angle": 0, + "content": null + }, + { + "type": "title", + "bbox": [ + 0.176, + 0.706, + 0.348, + 0.723 + ], + "angle": 0, + "content": "2 Related Work" + }, + { + "type": "title", + "bbox": [ + 0.176, + 0.738, + 0.42, + 0.754 + ], + "angle": 0, + "content": "2.1 Large Language Models" + }, + { + "type": "text", + "bbox": [ + 0.174, + 0.762, + 0.882, + 0.825 + ], + "angle": 0, + "content": "Over the last few years, pioneering foundation language models like GPT-1 [32], BERT [7], GPT-2 [33], and T5 [34] laid the groundwork, but GPT-3 [4] groundbreaking model parameters to 175 billion size to achieve remarkable zero-shot performance. Besides, research on scaling law [14] has steered language models to a larger scale. Therefore, driven by the success of InstructGPT [29]" + } + ], + [ + { + "type": "header", + "bbox": [ + 0.326, + 0.089, + 0.768, + 0.104 + ], + "angle": 0, + "content": "PiTe: Pixel-Temporal Alignment for Large Video-Language Model" + }, + { + "type": "page_number", + "bbox": [ + 0.81, + 0.091, + 0.82, + 0.101 + ], + "angle": 0, + "content": "3" + }, + { + "type": "text", + "bbox": [ + 0.118, + 0.12, + 0.821, + 0.256 + ], + "angle": 0, + "content": "and ChatGPT [28] which training by reinforcement learning with human feedback (RLHF) based on GPT-3, Large Language Models (LLMs) has made waves in the natural language processing (NLP) community due to its capabilities in language understanding, logical reasoning, and generation. The GPT's success suggest a promising path towards building LLMs, several open-source LLMs have been proposed following it with similar performance including OPT [49], BLOOM [36], GLM [9], LLaMA [37, 38], and Vicuna [5]. Our investigation delves into leveraging the striking language comprehension and zero-shot generalization abilities of LLMs beyond the confines of linguistic modalities. Specifically, we aim to extend these capabilities to multi-modal scenarios, thereby exploring their potential in processing diverse forms of information across different modalities." + }, + { + "type": "title", + "bbox": [ + 0.119, + 0.278, + 0.415, + 0.294 + ], + "angle": 0, + "content": "2.2 Large Visual-Language Models" + }, + { + "type": "text", + "bbox": [ + 0.118, + 0.302, + 0.821, + 0.528 + ], + "angle": 0, + "content": "The surge of LLMs has lead to major advancements in NLP tasks, and also has incited interest in developing Large Visual-Language Models (LVLMs). Building a unify LLM with visual inputs for visual language tasks thus remains one of the most important desiderata for LVLMs. Flamingo [1] and OpenFlamingo [3] fuse visual information into intermediate embedding for a frozen LLM by cross-attention mechanism, and train on billions of image-text pairs to align visual and linguistic modalities. Similarly, BLIP-2 [18] introduced the concept of Q-Former to align visual features more effectively with language space. Moreover, MiniGPT-4 [52] enhances its usability significantly by further fine-tuning on more detailed image descriptions with just one projection layer to align a frozen visual encoder with a frozen LLM, and the LLaVA series [21, 22] use simply a multi-layer perception (MLP) in place of the Q-Former and two-stage instruction tuning to enhance this process. Furthermore, PixelLLM [42] leverage the location coordinate of every word in the caption in the image as the connection between different modalities to strengthen the model's performance for the object detection task. Our primary focus lies in transferring the exceptional language comprehension capabilities of LLMs to the analysis of dynamic, continuous visual data found in videos, as opposed to static visual data such as images." + }, + { + "type": "title", + "bbox": [ + 0.119, + 0.55, + 0.411, + 0.566 + ], + "angle": 0, + "content": "2.3 Large Video-Language Models" + }, + { + "type": "text", + "bbox": [ + 0.118, + 0.574, + 0.821, + 0.74 + ], + "angle": 0, + "content": "Recently, many efforts have been made to transfer the task-handling capability of LVLMs to the video modality, leading to the emergence of Large Video-Language Models (LVidLMs) like VideoChat [19], Video-LLaMA [46], and Video-ChatGPT [26]. Prior researches have demonstrated the capability of LLMs to perform diverse tasks on video content, guided by user instructions through a two-stage training process. These studies align static visual features with LLMs, followed by instruction tuning on datasets annotated either by GPT or humans. Despite being effective in video understanding, the lack of fine-grained spatial-temporal modeling in these models prevents them from understanding or locating object in detail or specific segments. We propose a novel fine-grained alignment strategy at the pixel level across spatial and temporal dimensions to enhance the ability of LLMs to comprehensively analyze video content, thereby facilitating a more detailed understanding of the visual information presented." + }, + { + "type": "title", + "bbox": [ + 0.119, + 0.762, + 0.331, + 0.779 + ], + "angle": 0, + "content": "3 PiTe-143k Dataset" + }, + { + "type": "text", + "bbox": [ + 0.119, + 0.793, + 0.82, + 0.823 + ], + "angle": 0, + "content": "To facilitate fine-grained multi-modal alignment research at the pixel level, we introduce a large-scale video-language dataset PiTe-143k. This dataset fills a crucial gap in the existing resources by" + } + ], + [ + { + "type": "page_number", + "bbox": [ + 0.178, + 0.091, + 0.19, + 0.101 + ], + "angle": 0, + "content": "4" + }, + { + "type": "header", + "bbox": [ + 0.231, + 0.089, + 0.375, + 0.103 + ], + "angle": 0, + "content": "Y. Liu, P. Ding et al." + }, + { + "type": "table_caption", + "bbox": [ + 0.257, + 0.119, + 0.799, + 0.133 + ], + "angle": 0, + "content": "Table 1: Comparison between PiTe-143k and existing video instruction datasets." + }, + { + "type": "table", + "bbox": [ + 0.182, + 0.145, + 0.877, + 0.251 + ], + "angle": 0, + "content": "
DatasetTotal Dur.Avg. Dur.#Videos#EventsTemporal Localization#Objects Trajectories
VideoChat [19]41h18s8.2k×××
Valley [25]608h40s54.7k×××
Video-ChatGPT [26]432h117s13.3k×××
PiTe-143k (Ours)2086.44h52.18s143.64k343.93k✓1.02M
" + }, + { + "type": "text", + "bbox": [ + 0.174, + 0.279, + 0.881, + 0.384 + ], + "angle": 0, + "content": "providing extensive object moving trajectories with video instruction, which were previously unavailable in ready-made datasets. PiTe-143k constructed based on InternVid-10M-FLT [13,40] that each instance contains and entire video and multiple clip captions with start-stop timestamps. As shown in Table 1, PiTe-143k comprises 343.93 thousand event segments and 1.02 million moving trajectory for all individual objects that appear in both visual and textual modalities. To facilitate this objective, we establish an automatic annotation pipeline for PiTe-143k, fostering the advancement of LVidLMs for nuanced pixel-level video comprehension." + }, + { + "type": "text", + "bbox": [ + 0.174, + 0.385, + 0.882, + 0.461 + ], + "angle": 0, + "content": "The automatic annotation pipeline for PiTe-143k comprises two primary stages, as depicted in Fig. 2: (1) Stage 1 involves the noun phrases extraction and referring expressions segmentation, thereby generating object masks within the frame for all individual objects referenced in the event caption; (2) Stage 2 centers on point tracking to capture the moving trajectories corresponding to the masks obtained in Stage 1." + }, + { + "type": "title", + "bbox": [ + 0.175, + 0.481, + 0.513, + 0.497 + ], + "angle": 0, + "content": "3.1 Referring Expression Segmentation" + }, + { + "type": "text", + "bbox": [ + 0.174, + 0.506, + 0.88, + 0.536 + ], + "angle": 0, + "content": "In stage 1, we aim to build closely fine-grained connection between video and language. To this end, we extract all noun phrases from caption and find the corresponding objects in the clip." + }, + { + "type": "text", + "bbox": [ + 0.174, + 0.537, + 0.882, + 0.749 + ], + "angle": 0, + "content": "At inception, we leverage constituency parser SuPar [50,51] for language to extract noun phrase as shown in Fig. 3. Notably, in order to pass the simplest and most straightforward language instructions in next step, we only extract noun phrase from the lowest layer. For example in Fig. 3a, we consider two noun phrases \\( a \\) pen and \\( a \\) white table, but the parent node of the former that denotes \\( a \\) pen on a white table not in our consideration because of the complexity of its composition. Following this, we utilize GLaMM [35], the first LVLM that can generate natural language responses seamlessly intertwined with corresponding object segmentation masks, to obtain the corresponding segmentation mask in the first frame of the clip for the text-based referring expression. While certain objects in the video, such as \\( a \\) pen as illustrated in Fig.3a and Fig.2, may be too small to accurate detection. In such challenging cases, we disregard the trajectory information of the noun phrase. Despite this limitation, it has a minimal impact on the overall performance when utilizing extensive pre-training data on a large scale. Meanwhile, leveraging the exceptional language comprehension capabilities of LLMs, GLaMM can effectively filter out invalid referring expressions, those that do not constitute legal object references, such as front as depicted in Fig. 3b." + }, + { + "type": "title", + "bbox": [ + 0.175, + 0.769, + 0.347, + 0.785 + ], + "angle": 0, + "content": "3.2 Point Tracking" + }, + { + "type": "text", + "bbox": [ + 0.174, + 0.793, + 0.882, + 0.824 + ], + "angle": 0, + "content": "In stage 2, we aim to transfer the connection constructed in the previous stage to video, expanding out the temporal dimension specific to video compared to image. To this end, we track all individual" + } + ], + [ + { + "type": "header", + "bbox": [ + 0.326, + 0.089, + 0.769, + 0.104 + ], + "angle": 0, + "content": "PiTe: Pixel-Temporal Alignment for Large Video-Language Model" + }, + { + "type": "page_number", + "bbox": [ + 0.81, + 0.09, + 0.822, + 0.101 + ], + "angle": 0, + "content": "5" + }, + { + "type": "image", + "bbox": [ + 0.129, + 0.119, + 0.812, + 0.543 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.114, + 0.554, + 0.823, + 0.597 + ], + "angle": 0, + "content": "Fig. 2: Automatic annotation pipeline for PiTe-143k. The video sample in the figure showcases two events positioned at the commencement and conclusion of the video. The procedure for extracting noun phrases by SuPar [50,51] is elucidated in Fig. 3." + }, + { + "type": "text", + "bbox": [ + 0.114, + 0.627, + 0.82, + 0.656 + ], + "angle": 0, + "content": "objects in their clip to obtain the trajectory, the trajectory indicates the connection between video and language in both spatial and temporal dimensions." + }, + { + "type": "text", + "bbox": [ + 0.114, + 0.658, + 0.822, + 0.825 + ], + "angle": 0, + "content": "The stage 2 commences when we employ DOT [17], a simple-yet-efficient method for tracking point to recover the trajectory of any scene point, for each clip to capture the trajectories for any point in first frame. According to our observation, the caption of each clip mainly describes simple video content in short sentences, so most of the caption corresponds to just one scene clip, which enables us to track objects that identified in the first frame. Subsequently, filter trajectories according to the segmentation mask of objects obtained in stage 1. So far, we obtain the trajectories for all objects in each clip for each video, we create the connection between video and language from both spatial and temporal through the trajectories, the existence of trajectories in video denotes whether the object exist, and the value of trajectory represent where the object exist in video. Lastly, we utilize the k-means++ [2] clustering algorithm to condense trajectories into three key points, effectively reducing computational demands. This approach is founded on the premise that" + } + ], + [ + { + "type": "page_number", + "bbox": [ + 0.177, + 0.09, + 0.19, + 0.101 + ], + "angle": 0, + "content": "6" + }, + { + "type": "header", + "bbox": [ + 0.231, + 0.089, + 0.375, + 0.103 + ], + "angle": 0, + "content": "Y. Liu, P. Ding et al." + }, + { + "type": "image", + "bbox": [ + 0.177, + 0.118, + 0.49, + 0.282 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.177, + 0.284, + 0.49, + 0.307 + ], + "angle": 0, + "content": "(a) The extraction of the NP for the first event caption of the sample video depicted in Fig. 2." + }, + { + "type": "image", + "bbox": [ + 0.624, + 0.118, + 0.882, + 0.282 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.625, + 0.284, + 0.882, + 0.307 + ], + "angle": 0, + "content": "(b) Referring to Fig. 2, the extraction of the NP for the second event caption." + }, + { + "type": "image_caption", + "bbox": [ + 0.269, + 0.318, + 0.786, + 0.333 + ], + "angle": 0, + "content": "Fig. 3: Two samples of constituency parser for Noun Phrase (NP) extraction." + }, + { + "type": "text", + "bbox": [ + 0.174, + 0.351, + 0.882, + 0.397 + ], + "angle": 0, + "content": "three points adequately capture the typical geometric shape of objects, striking a balance between precision and computational efficiency. Furthermore, we conduct a comparative analysis of the performance using various key tracking points, as discussed in Section 5.3." + }, + { + "type": "title", + "bbox": [ + 0.176, + 0.417, + 0.262, + 0.433 + ], + "angle": 0, + "content": "4 PiTe" + }, + { + "type": "text", + "bbox": [ + 0.174, + 0.447, + 0.882, + 0.493 + ], + "angle": 0, + "content": "In this section, we propose a novel Large Video-Language Model (LVidLM), PiTe, which align video and language by trajectories across both spatial and temporal dimension. Fig. 4 illustrates an overview of PiTe." + }, + { + "type": "title", + "bbox": [ + 0.176, + 0.513, + 0.329, + 0.527 + ], + "angle": 0, + "content": "4.1 Architecture" + }, + { + "type": "text", + "bbox": [ + 0.174, + 0.536, + 0.882, + 0.598 + ], + "angle": 0, + "content": "PiTe is composed of a vision encoder to encode frames from video implemented as ViT [8], a vision adapter to project visual feature to semantic space of LLMs implemented as a linear projection layer, a LLM Vicuna v1.5 [5], and a localization projector or trajectory projector in separate training stage to guide LLMs to understand visual information implemented as a linear projection." + }, + { + "type": "text", + "bbox": [ + 0.174, + 0.615, + 0.888, + 0.675 + ], + "angle": 0, + "content": "Vision Encoder. Raw video data can be expressed as multiple frames such as \\(\\mathbf{v} = \\{f_1, f_2, \\dots, f_N\\} \\in \\mathbb{R}^{N \\times H \\times W \\times C}\\) (frames \\(\\times\\) height \\(\\times\\) width \\(\\times\\) channels). Following previous studies [22,26,46], we adopt ViT-L/14 [8] pre-trained from CLIP [31] as the vision encoder ViT to encode visual data. We uniformly sample \\(N\\) frames for video \\(\\mathbf{v}\\), and encode \\(i\\)-th frame \\(f_i\\) through the vision encoder ViT:" + }, + { + "type": "equation", + "bbox": [ + 0.417, + 0.681, + 0.882, + 0.705 + ], + "angle": 0, + "content": "\\[\n\\left\\{v _ {i} ^ {\\text {c l s}}, v _ {i} ^ {1}, v _ {i} ^ {2}, \\dots , v _ {i} ^ {P} \\right\\} = \\operatorname {V i T} \\left(f _ {i}\\right), \\tag {1}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.18, + 0.706, + 0.663, + 0.722 + ], + "angle": 0, + "content": "where \\(P\\) denotes the number of patches in the vision encoder ViT." + }, + { + "type": "text", + "bbox": [ + 0.174, + 0.74, + 0.882, + 0.798 + ], + "angle": 0, + "content": "Visual Adapter. A simpler projector forces the LLMs to learn more on handling visual inputs, leading to better generalization [20]. Hence, we utilize the global feature \\( v_{i}^{cls} \\) from vision encoder ViT as the representation for the \\( i \\)-th frame \\( f_{i} \\), and we apply a linear projection layer \\( \\mu(\\cdot) \\) to connect the frame feature into the word embedding space of LLMs:" + }, + { + "type": "equation", + "bbox": [ + 0.479, + 0.804, + 0.881, + 0.827 + ], + "angle": 0, + "content": "\\[\nz _ {i} = \\mu \\left(v _ {i} ^ {\\text {c l s}}\\right). \\tag {2}\n\\]" + } + ], + [ + { + "type": "header", + "bbox": [ + 0.326, + 0.089, + 0.769, + 0.104 + ], + "angle": 0, + "content": "PiTe: Pixel-Temporal Alignment for Large Video-Language Model" + }, + { + "type": "page_number", + "bbox": [ + 0.81, + 0.09, + 0.822, + 0.101 + ], + "angle": 0, + "content": "7" + }, + { + "type": "image_caption", + "bbox": [ + 0.117, + 0.117, + 0.406, + 0.132 + ], + "angle": 0, + "content": "Stage 1: Referring Expression Localization" + }, + { + "type": "image", + "bbox": [ + 0.117, + 0.134, + 0.819, + 0.241 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.117, + 0.249, + 0.356, + 0.264 + ], + "angle": 0, + "content": "Stage 2: Pixel-Temporal Alignment" + }, + { + "type": "image", + "bbox": [ + 0.117, + 0.26, + 0.819, + 0.563 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.117, + 0.568, + 0.356, + 0.582 + ], + "angle": 0, + "content": "Stage 3: Video Question Answering" + }, + { + "type": "image", + "bbox": [ + 0.118, + 0.585, + 0.816, + 0.664 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.239, + 0.672, + 0.696, + 0.687 + ], + "angle": 0, + "content": "Fig.4: Schematic of PiTe framework for video-language alignment." + }, + { + "type": "text", + "bbox": [ + 0.115, + 0.719, + 0.821, + 0.75 + ], + "angle": 0, + "content": "Subsequently, a sequence of frame tokens \\(\\mathbf{z} = \\{z_1,z_2,\\dots ,z_N\\} \\in \\mathbb{R}^{N\\times d}\\) becomes the input that LLMs can understand, \\(d\\) denotes the hidden dimension of LLMs." + }, + { + "type": "text", + "bbox": [ + 0.114, + 0.778, + 0.823, + 0.825 + ], + "angle": 0, + "content": "Large Language Model. After we tokenize and encode video into frame tokens \\(\\mathbf{z}\\), we concatenate it with textual tokens \\(\\mathbf{w} = \\{w_{1}, w_{2}, \\ldots, w_{L}\\} \\in \\mathbb{R}^{L}\\) and feed as the input to LLMs, we treat visual input as a foreign language in this process. Based on this, LLMs can further encode the input" + } + ], + [ + { + "type": "page_number", + "bbox": [ + 0.177, + 0.09, + 0.19, + 0.101 + ], + "angle": 0, + "content": "8" + }, + { + "type": "header", + "bbox": [ + 0.231, + 0.089, + 0.375, + 0.103 + ], + "angle": 0, + "content": "Y. Liu, P. Ding et al." + }, + { + "type": "text", + "bbox": [ + 0.174, + 0.121, + 0.88, + 0.152 + ], + "angle": 0, + "content": "sequence to understand the video and text, then reasoning and generation the response using auto-regressive decoding as follows:" + }, + { + "type": "equation", + "bbox": [ + 0.447, + 0.163, + 0.881, + 0.177 + ], + "angle": 0, + "content": "\\[\nh _ {i} = \\operatorname {L L M} ^ {-} \\left(\\mathbf {z}, \\mathbf {w} _ {1: i - 1}\\right), \\tag {3}\n\\]" + }, + { + "type": "equation", + "bbox": [ + 0.448, + 0.182, + 0.881, + 0.196 + ], + "angle": 0, + "content": "\\[\nw _ {i} = \\operatorname {a r g m a x} \\left(\\mathbf {m} _ {v} \\cdot h _ {i}\\right), \\tag {4}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.174, + 0.208, + 0.881, + 0.253 + ], + "angle": 0, + "content": "where \\(\\mathrm{LLM}^{-}\\) denotes the LLM without the last vocabulary mapping layer, \\(h_i\\) denotes the hidden states of \\(i\\)-th token generated by \\(\\mathrm{LLM}^{-}\\), \\(\\mathbf{m}_v \\in \\mathbb{R}^{|V| \\times d}\\) denotes the weight of the linear vocabulary mapping layer." + }, + { + "type": "title", + "bbox": [ + 0.176, + 0.274, + 0.367, + 0.289 + ], + "angle": 0, + "content": "4.2 Training Strategy" + }, + { + "type": "text", + "bbox": [ + 0.174, + 0.298, + 0.881, + 0.373 + ], + "angle": 0, + "content": "For PiTe model training, we consider a three stage instruction-tuning procedure, as depicted in Fig. 4: (1) Stage 1 centered around training adapters using image-caption pairs; (2) Stage 2 is focuses on aligning video and language features through trajectories; (3) Stage 3 is dedicated to enhancing the model's comprehension by following human instructions through high-quality dialogue instruction tuning." + }, + { + "type": "text", + "bbox": [ + 0.174, + 0.394, + 0.881, + 0.483 + ], + "angle": 0, + "content": "Stage 1: Referring Expression Localization. At the initial stage, we aim to train the visual adapter that can align visual features with semantic space of LLMs. To this end, we employ Localized Narratives dataset [30] that contains annotations of human annotators narrating a given image, together with a mouse trajectory of the annotators' attention during the narration. This gives synchronized locations for all words in the narration sentence, the cross-modal attention of human can be used to train our model as condition to bridge vision and language." + }, + { + "type": "text", + "bbox": [ + 0.174, + 0.484, + 0.881, + 0.553 + ], + "angle": 0, + "content": "There is only one visual tokens \\(\\mathbf{z} = \\{z_1\\} \\in \\mathbb{R}^{1\\times d}\\) in this training stage for image instead of video, this can align vision with language in spatial to train adapter without consider temporal information. To use the same language features for localization, we simply add a multi-layer perception (MLP) as localization projector \\(\\varphi (\\cdot)\\) parallel with the vocabulary mapping layer, which maps the language feature to a 2-dimension location:" + }, + { + "type": "equation", + "bbox": [ + 0.493, + 0.555, + 0.88, + 0.569 + ], + "angle": 0, + "content": "\\[\np _ {i} = \\varphi \\left(h _ {i}\\right), \\tag {5}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.18, + 0.572, + 0.637, + 0.587 + ], + "angle": 0, + "content": "where \\( p_i \\) denotes the predicted coordinate for textual token \\( w_i \\)." + }, + { + "type": "text", + "bbox": [ + 0.174, + 0.587, + 0.88, + 0.615 + ], + "angle": 0, + "content": "Overall, the leaning objective in the first stage was calculated by standard label-smoothed cross-entropy loss to train captioning output, and \\( L_{1} \\) regression loss to train the localization output:" + }, + { + "type": "equation", + "bbox": [ + 0.354, + 0.624, + 0.881, + 0.658 + ], + "angle": 0, + "content": "\\[\n\\mathcal {L} _ {1} = \\frac {1}{\\ell} \\sum_ {i = 1} ^ {\\ell} \\left(\\operatorname {C E} \\left(\\operatorname {L L M} \\left(\\mathbf {z}, \\mathbf {w} _ {1: i - 1}\\right), w _ {i}\\right) + \\lambda | \\hat {p} _ {i} - p _ {i} |\\right), \\tag {6}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.175, + 0.666, + 0.881, + 0.711 + ], + "angle": 0, + "content": "where \\(\\ell\\) is the length of the generated sequence, \\(\\hat{p}_i\\) represents the ground truth of location, and \\(\\mathrm{CE}(\\cdot ,\\cdot)\\) denotes the cross-entropy function. To enhance training efficiency, we utilize LoRA [12] for fine-tuning the LLM." + }, + { + "type": "text", + "bbox": [ + 0.174, + 0.732, + 0.882, + 0.824 + ], + "angle": 0, + "content": "Stage 2: Pixel-Temporal Alignment. After stage 1, the LLM model becomes proficient in understanding visual information. In the stage 2, we aim to train the LLM to understand sequential frames in video. To achieve this target, we curate a detailed object tracking dataset PiTe-143k, as described in Section 3 that uses trajectory as condition to bridge vision and language across both spatial and temporal dimensions. Therefore, the alignment guidance in pixel level improves the model's video fine-grained understanding reliability and overall usability." + } + ], + [ + { + "type": "header", + "bbox": [ + 0.325, + 0.089, + 0.769, + 0.105 + ], + "angle": 0, + "content": "PiTe: Pixel-Temporal Alignment for Large Video-Language Model" + }, + { + "type": "page_number", + "bbox": [ + 0.809, + 0.09, + 0.822, + 0.101 + ], + "angle": 0, + "content": "9" + }, + { + "type": "table_caption", + "bbox": [ + 0.316, + 0.118, + 0.622, + 0.133 + ], + "angle": 0, + "content": "Table 2: Training hyper-parameters of PiTe." + }, + { + "type": "table", + "bbox": [ + 0.121, + 0.145, + 0.819, + 0.273 + ], + "angle": 0, + "content": "
ConfigurationStage 1 Stage 2 Stage 3ConfigurationStage 1 Stage 2 Stage 3
Vision EncoderOpenAI-CLIP-L/14Learning Rate0.0001
Image/Frame Resolution224×224LoRAr=64 & α=128
Adapter ParameterTunableFrozenFrozenNumerical PrecisionBFloat16
Video Frames100Epoch122
LLMVicuna-7B/13B-v1.5Global Batch Size256
LLM Sequence Length2048Learning Rate ScheduleCosine Decay
OptimizerAdamWWarm-up Ratio0.03
" + }, + { + "type": "text", + "bbox": [ + 0.115, + 0.303, + 0.821, + 0.334 + ], + "angle": 0, + "content": "Similar to stage 1, we use the same language features for alignment by a MLP as trajectory projector \\(\\rho(\\cdot)\\) to map the language feature to a 2-dimension location:" + }, + { + "type": "equation", + "bbox": [ + 0.43, + 0.345, + 0.821, + 0.361 + ], + "angle": 0, + "content": "\\[\n\\mathbf {p} _ {i} = \\rho \\left(h _ {i}\\right), \\tag {7}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.114, + 0.372, + 0.821, + 0.402 + ], + "angle": 0, + "content": "where \\(\\mathbf{p}_i\\) denotes the trajectory matrix in \\(P\\) points and \\(N\\) frames for textual token \\(w_i\\). Here, we define \\(p_{ijk}\\) indicates the coordinate for token \\(w_i\\) for \\(i\\)-th point for model tracking in frame \\(f_j\\)." + }, + { + "type": "text", + "bbox": [ + 0.114, + 0.402, + 0.821, + 0.442 + ], + "angle": 0, + "content": "Overall, the leaning objective in stage 2 was calculated by standard label-smoothed cross-entropy loss to train the generation output, and \\( L_{1} \\) regression loss to train the trajectory output as the condition:" + }, + { + "type": "equation", + "bbox": [ + 0.232, + 0.443, + 0.821, + 0.481 + ], + "angle": 0, + "content": "\\[\n\\mathcal {L} _ {2} = \\frac {1}{\\ell} \\sum_ {i = 1} ^ {\\ell} \\left(\\operatorname {C E} \\left(\\operatorname {L L M} \\left(\\mathbf {z}, \\mathbf {w} _ {1: i - 1}\\right), w _ {i}\\right) + \\frac {\\lambda}{P \\cdot N} \\sum_ {j = 1} ^ {P} \\sum_ {k = 1} ^ {N} | \\hat {p} _ {i j k} - p _ {i j k} |\\right), \\tag {8}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.114, + 0.487, + 0.821, + 0.532 + ], + "angle": 0, + "content": "where \\(P\\) is the number of the points for model tracking to generate trajectory, and \\(\\mathbf{z} \\in \\mathbb{R}^{N \\times d}\\) represents the sequence of visual embedding. We merge the LoRA trained in the stage 1 with the original model and introduce a new LoRA module." + }, + { + "type": "text", + "bbox": [ + 0.114, + 0.533, + 0.822, + 0.631 + ], + "angle": 0, + "content": "It is notable that we use localization projector \\(\\varphi(\\cdot)\\) trained in previous stage to initialize the trajectory projector \\(\\rho(\\cdot)\\). Specifically, we define the weight of localization projector \\(\\varphi(\\cdot)\\) and trajectory projector \\(\\rho(\\cdot)\\) as \\(\\mathbf{m}_{\\varphi} \\in \\mathbb{R}^{P \\cdot N \\cdot 2 \\times d}\\) and \\(\\mathbf{m}_{\\rho} \\in \\mathbb{R}^{2 \\times d}\\), respectively. localization projector \\(\\varphi(\\cdot)\\) maps a 2-dimension coordinate on the input image for each token of the LLM output, as for trajectory projector \\(\\rho(\\cdot)\\), it also output 2-dimension coordinates, but more than \\(P \\cdot N\\) times for \\(P\\) points to tracking for \\(N\\) frames. For each point of each frame, the parameter of trajectory projector \\(\\rho(\\cdot)\\) initialized by the localization projector \\(\\varphi(\\cdot)\\):" + }, + { + "type": "equation", + "bbox": [ + 0.373, + 0.643, + 0.82, + 0.674 + ], + "angle": 0, + "content": "\\[\n\\mathbf {m} _ {\\varphi} = \\overbrace {\\mathbf {m} _ {\\rho} \\oplus \\mathbf {m} _ {\\rho} \\oplus \\cdots \\oplus \\mathbf {m} _ {\\rho}} ^ {P \\cdot N}, \\tag {9}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.12, + 0.686, + 0.554, + 0.7 + ], + "angle": 0, + "content": "where \\(\\oplus\\) denotes concatenation of matrix in first dimension." + }, + { + "type": "text", + "bbox": [ + 0.114, + 0.702, + 0.82, + 0.778 + ], + "angle": 0, + "content": "Beyond trajectories, our model is attuned to temporal boundaries within the generated text. Specifically, we structure the generation as ..., from s to e or From s to e, ... to facilitate the model to learn in temporal dimension. Here, ... encapsulates the event description, while s and e denote the frame indexes corresponding to the start and end timestamps of event, respectively. This approach further augments the model's understanding of temporal boundaries [13]." + }, + { + "type": "text", + "bbox": [ + 0.114, + 0.779, + 0.821, + 0.825 + ], + "angle": 0, + "content": "Dissimilar to the initial training stage, not all generated words are associated with trajectories. In cases where objects lack trajectories or vanish from view over time, we uniformly assign the coordinates of their ground truth as \\((-1, -1)\\) to signify their absence." + } + ], + [ + { + "type": "page_number", + "bbox": [ + 0.179, + 0.09, + 0.197, + 0.101 + ], + "angle": 0, + "content": "10" + }, + { + "type": "header", + "bbox": [ + 0.231, + 0.089, + 0.375, + 0.103 + ], + "angle": 0, + "content": "Y. Liu, P. Ding et al." + }, + { + "type": "table_caption", + "bbox": [ + 0.262, + 0.118, + 0.794, + 0.133 + ], + "angle": 0, + "content": "Table 3: Comparison between different LVidLMs on zero-shot question-answer." + }, + { + "type": "table", + "bbox": [ + 0.182, + 0.145, + 0.877, + 0.326 + ], + "angle": 0, + "content": "
ModelLLM SizeMSVD-QA [41]MSRVTT-QA [43]ActivityNet-QA [45]
Accuracy↑Score↑Accuracy↑Score↑Accuracy↑Score↑
FrozenBiLM [44]1B32.2-16.8-24.7-
LLaMA-Adapter [48]7B54.93.143.82.734.22.7
VideoChat [19]7B56.32.845.02.5-2.2
Video-LLaMA [46]7B51.62.529.61.812.41.1
Video-ChatGPT [26]7B64.93.349.32.835.22.7
PG-Video-LLaVA [27]7B64.13.751.63.339.93.3
PiTe (Ours)7B68.43.956.43.542.03.3
PiTe (Ours)13B71.64.057.73.542.23.4
" + }, + { + "type": "text", + "bbox": [ + 0.174, + 0.354, + 0.882, + 0.399 + ], + "angle": 0, + "content": "Stage 3: Video Question Answering. Following stage 2, we incorporate high-quality dialogue data Valley [25] and Video-ChatGPT [26] in one turn for instruction tuning, enabling the model to follow human instructions for more accurate and generalize capabilities of video understanding." + }, + { + "type": "text", + "bbox": [ + 0.175, + 0.4, + 0.882, + 0.427 + ], + "angle": 0, + "content": "The leaning objective in third stage was calculated by standard label-smoothed cross-entropy loss for auto-regression generation:" + }, + { + "type": "equation", + "bbox": [ + 0.402, + 0.439, + 0.881, + 0.475 + ], + "angle": 0, + "content": "\\[\n\\mathcal {L} _ {3} = \\frac {1}{\\ell} \\sum_ {i = 1} ^ {\\ell} \\operatorname {C E} \\left(\\text {L L M} \\left(\\mathbf {z}, \\mathbf {w} _ {1: i - 1}\\right), w _ {i}\\right). \\tag {10}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.175, + 0.483, + 0.882, + 0.514 + ], + "angle": 0, + "content": "Similar to stage 2, we merge the LoRA trained in the stage 1 and stage 2 with the original model and introduce a new LoRA module." + }, + { + "type": "title", + "bbox": [ + 0.176, + 0.538, + 0.336, + 0.556 + ], + "angle": 0, + "content": "5 Experiments" + }, + { + "type": "title", + "bbox": [ + 0.176, + 0.571, + 0.388, + 0.587 + ], + "angle": 0, + "content": "5.1 Experimental Setup" + }, + { + "type": "text", + "bbox": [ + 0.174, + 0.596, + 0.883, + 0.825 + ], + "angle": 0, + "content": "Tasks, Datasets, and Evaluation Metrics. We conduct a quantitative evaluation of LVidLMs' video understanding capabilities across three tasks: (1) Video Question Answering: This task assesses the comprehensive video comprehension abilities of LVidLMs by requiring the model to answer a variety of questions about the video content based on its understanding. We perform this task on three datasets: MSVD-QA [41], MSRVTT-QA [43], and ActivityNet-QA [45]. The evaluation pipeline for video understanding follows Video-ChatGPT [26], and we report the accuracy and score, which is assessed using GPT-Assistant [28]. (2) Video Temporal Grounding: This task evaluates LVidLMs' capacity to discern the starting and ending timestamps of a segment corresponding to the description of a video clip. This task demands the model to effectively grasp the temporal aspects of the video. We conduct this task on the ActivityNet Captions dataset [15] and calculate Intersection over Union (IoU) between the model-generate time segments and the ground truth time segments. We report mean IoU (mIoU) and Recall@1, IoU \\(\\geq m\\) (R@m) metric, where \\(m\\) values are set at \\(\\{0.3, 0.5, 0.7\\}\\). (3) Video Dense Captioning: This task requires the model to produce all events depict in the video along with their corresponding start and end timestamps. It necessitates the model to comprehend both the spatial and temporal dimensions of the video simultaneously." + } + ], + [ + { + "type": "header", + "bbox": [ + 0.326, + 0.089, + 0.768, + 0.104 + ], + "angle": 0, + "content": "PiTe: Pixel-Temporal Alignment for Large Video-Language Model" + }, + { + "type": "page_number", + "bbox": [ + 0.803, + 0.091, + 0.819, + 0.102 + ], + "angle": 0, + "content": "11" + }, + { + "type": "table_caption", + "bbox": [ + 0.115, + 0.118, + 0.823, + 0.148 + ], + "angle": 0, + "content": "Table 4: Comparison between different LVidLMs in temporal video grounding and dense video captioning tasks on ActivityNet [11]." + }, + { + "type": "table", + "bbox": [ + 0.122, + 0.159, + 0.819, + 0.297 + ], + "angle": 0, + "content": "
ModelLLM SizeTemporal GroundingDense Captioning
R@0.3↑R@0.5↑R@0.7↑mIoU↑SODA_c↑CIDEr↑METEOR↑
VideoChat [19]7B8.83.71.57.20.92.20.9
Video-LLaMA [46]7B6.92.10.86.51.95.81.9
Video-ChatGPT [26]7B26.413.66.118.91.95.82.1
PiTe (Ours)7B30.417.87.822.05.121.75.8
PiTe (Ours)13B37.223.710.926.05.926.56.6
" + }, + { + "type": "text", + "bbox": [ + 0.114, + 0.323, + 0.821, + 0.4 + ], + "angle": 0, + "content": "We conduct this task on the ActivityNet Captions dataset [15]. Initially, we reported SODA_c [10], followed by averages of CIDEr [39] and METEOR [16] under different IoU thresholds of 0.3, 0.5, 0.7, 0.9 based on generate events and ground truth matched pairs to provide a comprehensive analysis. In this paper, all experiments were conducted in a zero-shot setting, and higher values of all evaluation metrics indicate superior performance." + }, + { + "type": "text", + "bbox": [ + 0.114, + 0.42, + 0.821, + 0.496 + ], + "angle": 0, + "content": "Implementation Details. In this paper, we employ Vicuna v1.5 [5] as LLM to train the PiTe model at two scales: 7B and 13B. Leveraging the efficiency of LoRA [12], the training of the 7B model can be completed in approximately 10 hours using a single Nvidia 8-A100 (80GB VRAM) node, while the 13B model requires around 17 hours. More hyper-parameter settings are shwon in Table 2." + }, + { + "type": "title", + "bbox": [ + 0.115, + 0.516, + 0.265, + 0.531 + ], + "angle": 0, + "content": "5.2 Main Result" + }, + { + "type": "text", + "bbox": [ + 0.114, + 0.54, + 0.82, + 0.571 + ], + "angle": 0, + "content": "Table 3 and 4 present the comparative performance of the PiTe against state-of-the-art baselines on myriad video understanding datasets." + }, + { + "type": "text", + "bbox": [ + 0.114, + 0.591, + 0.821, + 0.698 + ], + "angle": 0, + "content": "Question Answering. As illustrated in Table 3, PiTe consistently outperforms the state-of-the-art pure instruction-tuning baselines in terms of all metrics on all datasets. Compared to the top-performing baselines in each dataset, PiTe exhibited notable improvements in the average question answering accuracy, achieving a maximum enhancement of 4.8 and an average improvement of 3.7. For example, PiTe substantially improves accuracy by 64.9 to 68.4 compared to Video-ChatGPT [26] in MSVD-QA dataset [41]. The results showcasing PiTe's proficiency in video comprehension and its capacity to deliver contextually relevant responses according to the given instructions." + }, + { + "type": "text", + "bbox": [ + 0.114, + 0.718, + 0.822, + 0.824 + ], + "angle": 0, + "content": "Temporal Grounding. As depicted in Table 4, PiTe achieves state-of-the-art performance in the video temporal grounding task across all metrics as well, demonstrating improvements ranging from 18.9 to 22.0 in mIoU compared to Video-ChatGPT [26]. This clearly indicates that trajectory alignment greatly enhances the ability of capture events in temporal dimension for LVidLMs. The incorporation of object trajectories in the temporal dimension of the trajectory matrix equips the model with a precise understanding of temporal event boundaries, thereby establishing a solid foundation for accurate event localization." + } + ], + [ + { + "type": "page_number", + "bbox": [ + 0.178, + 0.09, + 0.196, + 0.101 + ], + "angle": 0, + "content": "12" + }, + { + "type": "header", + "bbox": [ + 0.231, + 0.089, + 0.375, + 0.103 + ], + "angle": 0, + "content": "Y. Liu, P. Ding et al." + }, + { + "type": "table_caption", + "bbox": [ + 0.325, + 0.119, + 0.731, + 0.133 + ], + "angle": 0, + "content": "Table 5: Ablation study of the three-stage training strategy." + }, + { + "type": "table", + "bbox": [ + 0.182, + 0.144, + 0.877, + 0.279 + ], + "angle": 0, + "content": "
MethodMSVD-QAActivityNet
Temporal GroundingDense Captioning
Accuracy ScoreR@0.3R@0.5R@0.7mIoUSODA_cCIDErMETEOR
PiTe (Ours)68.43.930.417.87.822.05.121.75.8
w/o initialize68.23.922.810.54.617.15.121.75.8
w/o trajectory68.13.923.912.85.717.45.021.45.8
" + }, + { + "type": "text", + "bbox": [ + 0.174, + 0.307, + 0.882, + 0.399 + ], + "angle": 0, + "content": "Dense Captioning. The outcomes of the dense captioning task, as delineated in Table 4, reveal that PiTe consistent boost compared to all state-of-the-art baselines. Particularly noteworthy is the substantial 15.9 increase in the CIDEr metric [39] when compared to Video-ChatGPT [26]. This underscores the significance of fine-grained alignment in both spatial and temporal dimensions through trajectories, implying that PiTe acquires more generalized and detailed representations to offer more sophisticated event descriptions and accurate event temporal boundaries." + }, + { + "type": "title", + "bbox": [ + 0.176, + 0.423, + 0.293, + 0.438 + ], + "angle": 0, + "content": "5.3 Analysis" + }, + { + "type": "text", + "bbox": [ + 0.174, + 0.45, + 0.881, + 0.526 + ], + "angle": 0, + "content": "Ablation Study. As reported in Table 5, we conduct ablation experiments on MVSD-QA [41] for question answering and ActivityNet Captions [15] for temporal grounding to verify the individual effects of the proposed contributions under the following settings: (1) w/o initialize: we remove the initialization strategy that use weight of localization projector to initialize trajectory projector; (2) w/o trajectory: we abandon the fine-grained alignment strategy via trajectory." + }, + { + "type": "text", + "bbox": [ + 0.174, + 0.527, + 0.882, + 0.693 + ], + "angle": 0, + "content": "From the experimental results in Table 5, it can be observed that: (1) Eliminating the initialization strategy for the trajectory projector in PiTe reduces the model's reasoning capabilities and temporal boundary awareness. However, the performance in dense captioning generation remains consistent. This observation suggests that the model maintains its basic ability in comprehending visual content under trajectory-guided training. (2) The removal of the trajectory-guided training diminishes almost all the capability of PiTe, including dense captioning. (3) Without trajectory-guided training, PiTe demonstrates superior performance compared to trajectory-guided training without the initialization strategy for the trajectory projector in temporal grounding. This outcome highlights the difficulty of trajectory-guided training without initialization from a pre-trained localization projector, as the instability of parameters can impede the model's perception to accurately perceive visual temporal information." + }, + { + "type": "text", + "bbox": [ + 0.174, + 0.717, + 0.882, + 0.824 + ], + "angle": 0, + "content": "Exhibition. To better illustrate the video dialogue performance of PiTe, we present a qualitative example, as shown in Fig. 5a. The illustration from the upper portion of the figure demonstrates PiTe's capability not only to provide precise responses to instruction queries but also to enhance the output with more detailed and accurate video information. The example in the lower segment of the figure highlights the model's proficiency in understanding instruction and capturing event, enabling precise delineation of temporal boundaries within the video, despite the constraint of a 100-frame sampling limit." + } + ], + [ + { + "type": "header", + "bbox": [ + 0.326, + 0.089, + 0.769, + 0.104 + ], + "angle": 0, + "content": "PiTe: Pixel-Temporal Alignment for Large Video-Language Model" + }, + { + "type": "page_number", + "bbox": [ + 0.803, + 0.09, + 0.822, + 0.102 + ], + "angle": 0, + "content": "13" + }, + { + "type": "image", + "bbox": [ + 0.135, + 0.135, + 0.428, + 0.189 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.136, + 0.189, + 0.417, + 0.2 + ], + "angle": 0, + "content": "[Instruction] Who are shown performing impressive feats?" + }, + { + "type": "image_caption", + "bbox": [ + 0.136, + 0.206, + 0.401, + 0.237 + ], + "angle": 0, + "content": "[Response] The video showcases various football players performing impressive feats, including a player jumping over another player and a player making a spectacular catch." + }, + { + "type": "image", + "bbox": [ + 0.137, + 0.242, + 0.428, + 0.274 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.136, + 0.275, + 0.429, + 0.294 + ], + "angle": 0, + "content": "[Instruction] During which frames can we see he then shines the shoes of one man who pays him and shakes his hand." + }, + { + "type": "image_caption", + "bbox": [ + 0.137, + 0.296, + 0.412, + 0.306 + ], + "angle": 0, + "content": "[Response] The shoe shining and payment occur from 66 to 99" + }, + { + "type": "image_caption", + "bbox": [ + 0.137, + 0.306, + 0.379, + 0.316 + ], + "angle": 0, + "content": "(116.348s ~ 174.522s). Ground truth: 116.06s ~ 172.78s" + }, + { + "type": "image_caption", + "bbox": [ + 0.117, + 0.319, + 0.436, + 0.331 + ], + "angle": 0, + "content": "(a) Examples of PiTe's video understanding capabilities." + }, + { + "type": "image", + "bbox": [ + 0.468, + 0.124, + 0.817, + 0.315 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.453, + 0.319, + 0.82, + 0.332 + ], + "angle": 0, + "content": "(b) Performance comparison for different tracking point quantity." + }, + { + "type": "image_caption", + "bbox": [ + 0.114, + 0.343, + 0.821, + 0.37 + ], + "angle": 0, + "content": "Fig. 5: PiTe's video understanding capabilities and performance comparison across varying tracking point quantities." + }, + { + "type": "text", + "bbox": [ + 0.114, + 0.391, + 0.822, + 0.527 + ], + "angle": 0, + "content": "Impact of Tracking Point Quantity. In Fig. 5b, we vary the tracking point quantity \\( P \\) in set of \\( \\{1,3,5\\} \\). The efficacy of dense captioning tasks tends to improve with an increase in tracking points. However, it is observed that the temporal grounding task undergoes an initial substantial improvement, only to be succeeded by a rapid decline. Less number of tracking points fails to accurately capture the object's geometry, thereby hindering the pixel-level cross-modal alignment guidance for the model. Conversely, a higher quantity of points can enhance the model's comprehension of pure visual information; however, it also introduces noise to make training more challenging. Overall, that the optimal value of \\( P \\) may be different for different tasks, we set \\( P = 3 \\) due to its performance maintain stability over multiple tasks." + }, + { + "type": "title", + "bbox": [ + 0.115, + 0.547, + 0.26, + 0.563 + ], + "angle": 0, + "content": "6 Conclusion" + }, + { + "type": "text", + "bbox": [ + 0.114, + 0.577, + 0.823, + 0.744 + ], + "angle": 0, + "content": "In this paper, we focus on enhancing the performance of Large Video-Language Models (LVidLMs) by incorporating trajectory-based alignment across different modalities. To achieve fine-grained alignment between video and language across spatial and temporal dimensions, we initially curate a comprehensive multi-modal object tracking dataset, PiTe-143k, using a fully automated annotation pipeline. This dataset was developed to address the lack of large-scale video-language datasets that include multi-object moving trajectories. Subsequently, we introduce a novel Pixel-Temporal (PiTe) alignment strategy that leverages trajectory-guided pre-training to address the inherent challenges faced by LVidLMs. Through comparative analyses, we evaluate PiTe against state-of-the-art models and competitive baselines across various tasks in a zero-shot setting, including question-answering, temporal grounding, and dense captioning, showcasing the superior performance of PiTe with more sophisticated event descriptions and accurate event temporal boundaries." + }, + { + "type": "title", + "bbox": [ + 0.116, + 0.764, + 0.294, + 0.781 + ], + "angle": 0, + "content": "Acknowledgments" + }, + { + "type": "text", + "bbox": [ + 0.115, + 0.793, + 0.821, + 0.824 + ], + "angle": 0, + "content": "This work was supported by the National Science and Technology Innovation 2030 - Major Project (Grant No. 2022ZD0208800), and NSFC General Program (Grant No. 62176215)." + } + ], + [ + { + "type": "page_number", + "bbox": [ + 0.178, + 0.09, + 0.197, + 0.101 + ], + "angle": 0, + "content": "14" + }, + { + "type": "header", + "bbox": [ + 0.231, + 0.089, + 0.375, + 0.103 + ], + "angle": 0, + "content": "Y. Liu, P. Ding et al." + }, + { + "type": "title", + "bbox": [ + 0.177, + 0.119, + 0.284, + 0.135 + ], + "angle": 0, + "content": "References" + }, + { + "type": "ref_text", + "bbox": [ + 0.184, + 0.149, + 0.882, + 0.218 + ], + "angle": 0, + "content": "1. Alayrac, J., Donahue, J., Luc, P., Miech, A., Barr, I., Hasson, Y., Lenc, K., Mensch, A., Millican, K., Reynolds, M., Ring, R., Rutherford, E., Cabi, S., Han, T., Gong, Z., Samangooei, S., Monteiro, M., Menick, J.L., Borgeaud, S., Brock, A., Nematzadeh, A., Sharifzadeh, S., Binkowski, M., Barreira, R., Vinyals, O., Zisserman, A., Simonyan, K.: Flamingo: a visual language model for few-shot learning. In: Proc. of NeurIPS (2022)" + }, + { + "type": "ref_text", + "bbox": [ + 0.185, + 0.219, + 0.88, + 0.245 + ], + "angle": 0, + "content": "2. Arthur, D., Vassilvitskii, S.: k-means++: the advantages of careful seeding. In: Proc. of SODA. pp. 1027-1035 (2007)" + }, + { + "type": "ref_text", + "bbox": [ + 0.185, + 0.246, + 0.881, + 0.287 + ], + "angle": 0, + "content": "3. Awadalla, A., Gao, I., Gardner, J., Hessel, J., Hanafy, Y., Zhu, W., Marathe, K., Bitton, Y., Gadre, S.Y., Sagawa, S., Jitsev, J., Kornblith, S., Koh, P.W., Ilharco, G., Wortman, M., Schmidt, L.: Openflamingo: An open-source framework for training large autoregressive vision-language models. CoRR (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.184, + 0.288, + 0.882, + 0.355 + ], + "angle": 0, + "content": "4. Brown, T.B., Mann, B., Ryder, N., Subbiah, M., Kaplan, J., Dhariwal, P., Neelakantan, A., Shyam, P., Sastry, G., Askell, A., Agarwal, S., Herbert-Voss, A., Krueger, G., Henighan, T., Child, R., Ramesh, A., Ziegler, D.M., Wu, J., Winter, C., Hesse, C., Chen, M., Sigler, E., Litwin, M., Gray, S., Chess, B., Clark, J., Berner, C., McCandlish, S., Radford, A., Sutskever, I., Amodei, D.: Language models are few-shot learners. In: Proc. of NeurIPS (2020)" + }, + { + "type": "ref_text", + "bbox": [ + 0.185, + 0.357, + 0.881, + 0.396 + ], + "angle": 0, + "content": "5. Chiang, W.L., Li, Z., Lin, Z., Sheng, Y., Wu, Z., Zhang, H., Zheng, L., Zhuang, S., Zhuang, Y., Gonzalez, J.E., Stoica, I., Xing, E.P.: Vicuna: An open-source chatbot impressing gpt-4 with \\(90\\%\\) * chatgpt quality (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.185, + 0.398, + 0.881, + 0.424 + ], + "angle": 0, + "content": "6. Dai, W., Li, J., Li, D., Tiong, A.M.H., Zhao, J., Wang, W., Li, B., Fung, P., Hoi, S.C.H.: Instructblip: Towards general-purpose vision-language models with instruction tuning. In: Proc. of NeurIPS (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.185, + 0.425, + 0.881, + 0.452 + ], + "angle": 0, + "content": "7. Devlin, J., Chang, M., Lee, K., Toutanova, K.: BERT: pre-training of deep bidirectional transformers for language understanding. In: Proc. of AACL. pp. 4171-4186 (2019)" + }, + { + "type": "ref_text", + "bbox": [ + 0.185, + 0.453, + 0.881, + 0.493 + ], + "angle": 0, + "content": "8. Dosovitskiy, A., Beyer, L., Kolesnikov, A., Weissenborn, D., Zhai, X., Unterthiner, T., Dehghani, M., Minderer, M., Heigold, G., Gelly, S., Uszkoreit, J., Houlsby, N.: An image is worth 16x16 words: Transformers for image recognition at scale. In: Proc. of ICLR (2021)" + }, + { + "type": "ref_text", + "bbox": [ + 0.185, + 0.494, + 0.881, + 0.522 + ], + "angle": 0, + "content": "9. Du, Z., Qian, Y., Liu, X., Ding, M., Qiu, J., Yang, Z., Tang, J.: GLM: general language model pretraining with autoregressive blank infilling. In: Proc. of ACL. pp. 320-335 (2022)" + }, + { + "type": "ref_text", + "bbox": [ + 0.18, + 0.522, + 0.881, + 0.548 + ], + "angle": 0, + "content": "10. Fujita, S., Hirao, T., Kamigaito, H., Okumura, M., Nagata, M.: SODA: story oriented dense video captioning evaluation framework. In: Proc. of ECCV. pp. 517-531 (2020)" + }, + { + "type": "ref_text", + "bbox": [ + 0.179, + 0.549, + 0.881, + 0.576 + ], + "angle": 0, + "content": "11. Heilbron, F.C., Escorcia, V., Ghanem, B., Niebles, J.C.: Activitynet: A large-scale video benchmark for human activity understanding. In: Proc. of CVPR. pp. 961-970 (2015)" + }, + { + "type": "ref_text", + "bbox": [ + 0.179, + 0.577, + 0.881, + 0.603 + ], + "angle": 0, + "content": "12. Hu, E.J., Shen, Y., Wallis, P., Allen-Zhu, Z., Li, Y., Wang, S., Wang, L., Chen, W.: Lora: Low-rank adaptation of large language models. In: Proc. of ICLR (2022)" + }, + { + "type": "ref_text", + "bbox": [ + 0.179, + 0.604, + 0.881, + 0.631 + ], + "angle": 0, + "content": "13. Huang, B., Wang, X., Chen, H., Song, Z., Zhu, W.: Vtimellm: Empower llm to grasp video moments. In: Proc. of CVPR. pp. 14271-14280 (2024)" + }, + { + "type": "ref_text", + "bbox": [ + 0.179, + 0.632, + 0.881, + 0.658 + ], + "angle": 0, + "content": "14. Kaplan, J., McCandlish, S., Henighan, T., Brown, T.B., Chess, B., Child, R., Gray, S., Radford, A., Wu, J., Amodei, D.: Scaling laws for neural language models. CoRR (2020)" + }, + { + "type": "ref_text", + "bbox": [ + 0.179, + 0.659, + 0.881, + 0.685 + ], + "angle": 0, + "content": "15. Krishna, R., Hata, K., Ren, F., Fei-Fei, L., Niebles, J.C.: Dense-c captioning events in videos. In: Proc. of ICCV. pp. 706-715 (2017)" + }, + { + "type": "ref_text", + "bbox": [ + 0.179, + 0.687, + 0.881, + 0.713 + ], + "angle": 0, + "content": "16. Lavie, A., Agarwal, A.: METEOR: an automatic metric for MT evaluation with high levels of correlation with human judgments. In: Proc. of WMT@ACL. pp. 228-231 (2007)" + }, + { + "type": "ref_text", + "bbox": [ + 0.179, + 0.714, + 0.881, + 0.741 + ], + "angle": 0, + "content": "17. Le Moing, G., Ponce, J., Schmid, C.: Dense optical tracking: Connecting the dots. In: Proc. of CVPR. pp. 19187-19197 (2024)" + }, + { + "type": "ref_text", + "bbox": [ + 0.179, + 0.741, + 0.881, + 0.768 + ], + "angle": 0, + "content": "18. Li, J., Li, D., Savarese, S., Hoi, S.C.H.: BLIP-2: bootstrapping language-image pre-training with frozen image encoders and large language models. In: Proc. of ICML. pp. 19730-19742 (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.179, + 0.769, + 0.881, + 0.795 + ], + "angle": 0, + "content": "19. Li, K., He, Y., Wang, Y., Li, Y., Wang, W., Luo, P., Wang, Y., Wang, L., Qiao, Y.: Videochat: Chatcentric video understanding. CoRR (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.178, + 0.796, + 0.881, + 0.823 + ], + "angle": 0, + "content": "20. Lin, J., Yin, H., Ping, W., Molchanov, P., Shoeybi, M., Han, S.: Vila: On pre-training for visual language models. In: Proc. of CVPR. pp. 26689-26699 (2024)" + }, + { + "type": "list", + "bbox": [ + 0.177, + 0.149, + 0.882, + 0.823 + ], + "angle": 0, + "content": null + } + ], + [ + { + "type": "header", + "bbox": [ + 0.326, + 0.089, + 0.768, + 0.104 + ], + "angle": 0, + "content": "PiTe: Pixel-Temporal Alignment for Large Video-Language Model" + }, + { + "type": "page_number", + "bbox": [ + 0.803, + 0.09, + 0.82, + 0.101 + ], + "angle": 0, + "content": "15" + }, + { + "type": "ref_text", + "bbox": [ + 0.117, + 0.121, + 0.82, + 0.149 + ], + "angle": 0, + "content": "21. Liu, H., Li, C., Li, Y., Lee, Y.J.: Improved baselines with visual instruction tuning. In: Proc. of CVPR. pp. 26296-26306 (2024)" + }, + { + "type": "ref_text", + "bbox": [ + 0.118, + 0.149, + 0.734, + 0.163 + ], + "angle": 0, + "content": "22. Liu, H., Li, C., Wu, Q., Lee, Y.J.: Visual instruction tuning. In: Proc. of NeurIPS (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.118, + 0.163, + 0.821, + 0.19 + ], + "angle": 0, + "content": "23. Liu, Y., Ping, H., Zhang, D., Sun, Q., Li, S., Zhou, G.: Comment-aware multi-modal heterogeneous pre-training for humor detection in short-form videos. In: Proc. of ECAI. pp. 1568-1575 (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.117, + 0.19, + 0.821, + 0.218 + ], + "angle": 0, + "content": "24. Liu, Y., Shen, T., Zhang, D., Sun, Q., Li, S., Zhou, G.: Comment-aided video-language alignment via contrastive pre-training for short-form video humor detection. In: Proc. of ICMR. pp. 442-450 (2024)" + }, + { + "type": "ref_text", + "bbox": [ + 0.117, + 0.218, + 0.821, + 0.246 + ], + "angle": 0, + "content": "25. Luo, R., Zhao, Z., Yang, M., Dong, J., Qiu, M., Lu, P., Wang, T., Wei, Z.: Valley: Video assistant with large language model enhanced ability. CoRR (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.117, + 0.246, + 0.821, + 0.273 + ], + "angle": 0, + "content": "26. Maaz, M., Rasheed, H.A., Khan, S.H., Khan, F.S.: Video-chatgpt: Towards detailed video understanding via large vision and language models (2024)" + }, + { + "type": "ref_text", + "bbox": [ + 0.117, + 0.273, + 0.821, + 0.3 + ], + "angle": 0, + "content": "27. Munasinghe, S., Thushara, R., Maaz, M., Rasheed, H.A., Khan, S., Shah, M., Khan, F.S.: Pg-video-llava: Pixel grounding large video-language models. CoRR (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.117, + 0.3, + 0.59, + 0.314 + ], + "angle": 0, + "content": "28. OpenAI: Chatgpt: Optimizing language models for dialogue (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.117, + 0.314, + 0.821, + 0.369 + ], + "angle": 0, + "content": "29. Ouyang, L., Wu, J., Jiang, X., Almeida, D., Wainwright, C.L., Mishkin, P., Zhang, C., Agarwal, S., Slama, K., Ray, A., Schulman, J., Hilton, J., Kelton, F., Miller, L., Simens, M., Askell, A., Welinder, P., Christiano, P.F., Leike, J., Lowe, R.: Training language models to follow instructions with human feedback. In: Proc. of NeurIPS (2022)" + }, + { + "type": "ref_text", + "bbox": [ + 0.117, + 0.369, + 0.821, + 0.397 + ], + "angle": 0, + "content": "30. Pont-Tuset, J., Uijlings, J.R.R., Changpinyo, S., Soricut, R., Ferrari, V.: Connecting vision and language with localized narratives. In: Proc. of ECCV. pp. 647-664 (2020)" + }, + { + "type": "ref_text", + "bbox": [ + 0.117, + 0.397, + 0.821, + 0.438 + ], + "angle": 0, + "content": "31. Radford, A., Kim, J.W., Hallacy, C., Ramesh, A., Goh, G., Agarwal, S., Sastry, G., Askell, A., Mishkin, P., Clark, J., Krueger, G., Sutskever, I.: Learning transferable visual models from natural language supervision. In: Proc. of ICML. pp. 8748-8763 (2021)" + }, + { + "type": "ref_text", + "bbox": [ + 0.117, + 0.438, + 0.801, + 0.452 + ], + "angle": 0, + "content": "32. Radford, A., Narasimhan, K.: Improving language understanding by generative pre-training (2018)" + }, + { + "type": "ref_text", + "bbox": [ + 0.117, + 0.452, + 0.821, + 0.478 + ], + "angle": 0, + "content": "33. Radford, A., Wu, J., Child, R., Luan, D., Amodei, D., Sutskever, I.: Language models are unsupervised multitask learners (2019)" + }, + { + "type": "ref_text", + "bbox": [ + 0.117, + 0.479, + 0.821, + 0.52 + ], + "angle": 0, + "content": "34. Raffel, C., Shazeer, N., Roberts, A., Lee, K., Narang, S., Matena, M., Zhou, Y., Li, W., Liu, P.J.: Exploring the limits of transfer learning with a unified text-to-text transformer. J. Mach. Learn. Res. pp. 140:1-140:67 (2020)" + }, + { + "type": "ref_text", + "bbox": [ + 0.117, + 0.52, + 0.821, + 0.56 + ], + "angle": 0, + "content": "35. Rasheed, H., Maaz, M., Shaji, S., Shaker, A., Khan, S., Cholakkal, H., Anwer, R.M., Xing, E., Yang, M.H., Khan, F.S.: Glamm: Pixel grounding large multimodal model. In: Proc. of CVPR. pp. 13009-13018 (2024)" + }, + { + "type": "ref_text", + "bbox": [ + 0.117, + 0.561, + 0.821, + 0.658 + ], + "angle": 0, + "content": "36. Scao, T.L., Fan, A., Akiki, C., Pavlick, E., Ilic, S., Hesslow, D., Castagné, R., Luccioni, A.S., Yvon, F., Gallé, M., Tow, J., Rush, A.M., Biderman, S., Webson, A., Ammanamanchi, P.S., Wang, T., Sagot, B., Muennighoff, N., del Moral, A.V., Ruwase, O., Bawden, R., Bekman, S., McMillan-Major, A., Beltagy, I., Nguyen, H., Saulnier, L., Tan, S., Suarez, P.O., Sanh, V., Laurençon, H., Jernite, Y., Launay, J., Mitchell, M., Raffel, C., Gokaslan, A., Simhi, A., Soroa, A., Aji, A.F., Alfassy, A., Rogers, A., Nitzav, A.K., Xu, C., Mou, C., Emezue, C., Klamm, C., Leong, C., van Strien, D., Adelani, D.I., et al.: BLOOM: A 176b-parameter open-access multilingual language model. CoRR (2022)" + }, + { + "type": "ref_text", + "bbox": [ + 0.117, + 0.658, + 0.821, + 0.699 + ], + "angle": 0, + "content": "37. Touvron, H., Lavril, T., Izacard, G., Martinet, X., Lachaux, M., Lacroix, T., Rozière, B., Goyal, N., Hambro, E., Azhar, F., Rodriguez, A., Joulin, A., Grave, E., Lample, G.: Llama: Open and efficient foundation language models. CoRR (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.117, + 0.699, + 0.821, + 0.824 + ], + "angle": 0, + "content": "38. Touvron, H., Martin, L., Stone, K., Albert, P., Almahairi, A., Babaei, Y., Bashlykov, N., Batra, S., Bhargava, P., Bhosale, S., Bikel, D., Blecher, L., Canton-Ferrer, C., Chen, M., Cucurull, G., Esiobu, D., Fernandes, J., Fu, J., Fu, W., Fuller, B., Gao, C., Goswami, V., Goyal, N., Hartshorn, A., Hosseini, S., Hou, R., Inan, H., Kardas, M., Kerkez, V., Khabsa, M., Kloumann, I., Korenev, A., Koura, P.S., Lachaux, M., Lavril, T., Lee, J., Liskovich, D., Lu, Y., Mao, Y., Martinet, X., Mihaylov, T., Mishra, P., Molybog, I., Nie, Y., Poulton, A., Reizenstein, J., Rungta, R., Saladi, K., Schelten, A., Silva, R., Smith, E.M., Subramanian, R., Tan, X.E., Tang, B., Taylor, R., Williams, A., Kuan, J.X., Xu, P., Yan, Z., Zarov, I., Zhang, Y., Fan, A., Kambadur, M., Narang, S., Rodriguez, A., Stojnic, R., Edunov, S., Scialom, T.: Llama 2: Open foundation and fine-tuned chat models. CoRR (2023)" + }, + { + "type": "list", + "bbox": [ + 0.117, + 0.121, + 0.821, + 0.824 + ], + "angle": 0, + "content": null + } + ], + [ + { + "type": "page_number", + "bbox": [ + 0.178, + 0.09, + 0.197, + 0.101 + ], + "angle": 0, + "content": "16" + }, + { + "type": "header", + "bbox": [ + 0.231, + 0.089, + 0.375, + 0.103 + ], + "angle": 0, + "content": "Y. Liu, P. Ding et al." + }, + { + "type": "ref_text", + "bbox": [ + 0.177, + 0.121, + 0.882, + 0.149 + ], + "angle": 0, + "content": "39. Vedantam, R., Zitnick, C.L., Parikh, D.: Cider: Consensus-based image description evaluation. In: Proc. of CVPR. pp. 4566-4575 (2015)" + }, + { + "type": "ref_text", + "bbox": [ + 0.177, + 0.149, + 0.882, + 0.19 + ], + "angle": 0, + "content": "40. Wang, Y., He, Y., Li, Y., Li, K., Yu, J., Ma, X., Li, X., Chen, G., Chen, X., Wang, Y., Luo, P., Liu, Z., Wang, Y., Wang, L., Qiao, Y.: Internvid: A large-scale video-text dataset for multimodal understanding and generation. In: Proc. of ICLR (2024)" + }, + { + "type": "ref_text", + "bbox": [ + 0.177, + 0.19, + 0.882, + 0.232 + ], + "angle": 0, + "content": "41. Xu, D., Zhao, Z., Xiao, J., Wu, F., Zhang, H., He, X., Zhuang, Y.: Video question answering via gradually refined attention over appearance and motion. In: Proceedings of the 2017 ACM on Multimedia Conference, MM 2017, Mountain View, CA, USA, October 23-27, 2017. pp. 1645-1653 (2017)" + }, + { + "type": "ref_text", + "bbox": [ + 0.177, + 0.232, + 0.882, + 0.26 + ], + "angle": 0, + "content": "42. Xu, J., Zhou, X., Yan, S., Gu, X., Arnab, A., Sun, C., Wang, X., Schmid, C.: Pixel-aligned language model. In: Proc. of CVPR. pp. 13030-13039 (2024)" + }, + { + "type": "ref_text", + "bbox": [ + 0.177, + 0.26, + 0.882, + 0.288 + ], + "angle": 0, + "content": "43. Xu, J., Mei, T., Yao, T., Rui, Y.: MSR-VTT: A large video description dataset for bridging video and language. In: Proc. of CVPR. pp. 5288-5296 (2016)" + }, + { + "type": "ref_text", + "bbox": [ + 0.177, + 0.288, + 0.882, + 0.316 + ], + "angle": 0, + "content": "44. Yang, A., Miech, A., Sivic, J., Laptev, I., Schmid, C.: Zero-shot video question answering via frozen bidirectional language models. In: Proc. of NeurIPS (2022)" + }, + { + "type": "ref_text", + "bbox": [ + 0.177, + 0.316, + 0.882, + 0.343 + ], + "angle": 0, + "content": "45. Yu, Z., Xu, D., Yu, J., Yu, T., Zhao, Z., Zhuang, Y., Tao, D.: Activitynet-qa: A dataset for understanding complex web videos via question answering. In: Proc. of AAAI. pp. 9127-9134 (2019)" + }, + { + "type": "ref_text", + "bbox": [ + 0.177, + 0.343, + 0.882, + 0.371 + ], + "angle": 0, + "content": "46. Zhang, H., Li, X., Bing, L.: Video-llama: An instruction-tuned audio-visual language model for video understanding. In: Proc. of EMNLP. pp. 543-553 (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.177, + 0.371, + 0.882, + 0.399 + ], + "angle": 0, + "content": "47. Zhang, M., Huang, S., Li, W., Wang, D.: Tree structure-aware few-shot image classification via hierarchical aggregation. In: Proc. of ECCV. pp. 453-470 (2022)" + }, + { + "type": "ref_text", + "bbox": [ + 0.177, + 0.399, + 0.882, + 0.426 + ], + "angle": 0, + "content": "48. Zhang, R., Han, J., Liu, C., Zhou, A., Lu, P., Qiao, Y., Li, H., Gao, P.: LLaMA-adapter: Efficient fine-tuning of large language models with zero-initialized attention. In: Proc. of ICLR (2024)" + }, + { + "type": "ref_text", + "bbox": [ + 0.177, + 0.426, + 0.882, + 0.468 + ], + "angle": 0, + "content": "49. Zhang, S., Roller, S., Goyal, N., Artetxe, M., Chen, M., Chen, S., Dewan, C., Diab, M.T., Li, X., Lin, X.V., Mihaylov, T., Ott, M., Shleifer, S., Shuster, K., Simig, D., Koura, P.S., Sridhar, A., Wang, T., Zettlemoyer, L.: OPT: open pre-trained transformer language models. CoRR (2022)" + }, + { + "type": "ref_text", + "bbox": [ + 0.177, + 0.468, + 0.882, + 0.496 + ], + "angle": 0, + "content": "50. Zhang, Y., Li, Z., Zhang, M.: Efficient second-order treecrf for neural dependency parsing. In: Proc. of ACL. pp. 3295-3305 (2020)" + }, + { + "type": "ref_text", + "bbox": [ + 0.177, + 0.496, + 0.882, + 0.523 + ], + "angle": 0, + "content": "51. Zhang, Y., Zhou, H., Li, Z.: Fast and accurate neural CRF constituency parsing. In: Proc. of IJCAI. pp. 4046-4053 (2020)" + }, + { + "type": "ref_text", + "bbox": [ + 0.177, + 0.523, + 0.882, + 0.551 + ], + "angle": 0, + "content": "52. Zhu, D., Chen, J., Shen, X., Li, X., Elhoseiny, M.: MiniGPT-4: Enhancing vision-language understanding with advanced large language models. In: Proc. of ICLR (2024)" + }, + { + "type": "list", + "bbox": [ + 0.177, + 0.121, + 0.882, + 0.551 + ], + "angle": 0, + "content": null + } + ] +] \ No newline at end of file diff --git a/2024/PiTe_ Pixel-Temporal Alignment for Large Video-Language Model/3468332a-054d-4ee9-a3ba-3d30cf6d6594_origin.pdf b/2024/PiTe_ Pixel-Temporal Alignment for Large Video-Language Model/3468332a-054d-4ee9-a3ba-3d30cf6d6594_origin.pdf new file mode 100644 index 0000000000000000000000000000000000000000..dd75739b4670d860b9f159cf6f40233837ef0f6f --- /dev/null +++ b/2024/PiTe_ Pixel-Temporal Alignment for Large Video-Language Model/3468332a-054d-4ee9-a3ba-3d30cf6d6594_origin.pdf @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:cfeeac76b0e3d57544b32f83cdd2823f872edaebc3720d82616d2a2ddfef2f94 +size 9212883 diff --git a/2024/PiTe_ Pixel-Temporal Alignment for Large Video-Language Model/full.md b/2024/PiTe_ Pixel-Temporal Alignment for Large Video-Language Model/full.md new file mode 100644 index 0000000000000000000000000000000000000000..45d50f71a10d05e4883a771b451f38527263a9be --- /dev/null +++ b/2024/PiTe_ Pixel-Temporal Alignment for Large Video-Language Model/full.md @@ -0,0 +1,323 @@ +# PiTe: Pixel-Temporal Alignment for Large Video-Language Model + +Yang Liu $^{1,2\star}$ , Pengxiang Ding $^{1\star}$ , Siteng Huang $^{1}$ , Min Zhang $^{1}$ , Han Zhao $^{1}$ , and Donglin Wang $^{1\boxtimes}$ + +$^{1}$ Westlake University $^{2}$ Soochow University {liuyang67, dingpengxiang, wangdonglin}@westlake.edu.cn + +Abstract. Fueled by the Large Language Models (LLMs) wave, Large Visual-Language Models (LVLMs) have emerged as a pivotal advancement, bridging the gap between image and text. However, video making it challenging for LVLMs to perform adequately due to the complexity of the relationship between language and spatial-temporal data structure. Recent Large Video-Language Models (LVidLMs) align feature of static visual data like image into latent space of language feature, by general multi-modal tasks to leverage abilities of LLMs sufficiently. In this paper, we explore fine-grained alignment approach via object trajectory for different modalities across both spatial and temporal dimensions simultaneously. Thus, we propose a novel LVidLM by trajectory-guided Pixel-Temporal Alignment, dubbed PiTe, that exhibits promising applicable model property. To achieve fine-grained video-language alignment, we curate a multi-modal pre-training dataset PiTe-143k, the dataset provision of moving trajectories in pixel level for all individual objects, that appear and mention in the video and caption both, by our automatic annotation pipeline. Meanwhile, PiTe demonstrates astounding capabilities on myriad video-related multi-modal tasks through beat the state-of-the-art methods by a large margin. + +Keywords: Large Video-Language Model $\cdot$ Trajectory-guided Instruction Tuning $\cdot$ Video Understanding + +# 1 Introduction + +Large Language Models (LLMs) have rapidly gained popularity within the AI community, demonstrating astounding capabilities across a wide array of natural language tasks [4, 5, 9, 29, 37, 38]. The powerful language comprehension abilities of LLMs drive researchers to explore their utility in addressing a broader spectrum of tasks across various domains. Consequently, an increasing number of studies are focusing on developing comprehensive Large Visual-Language Models (LVLMs) to tackle vision-related tasks in zero-shot settings [6, 18, 47, 52], particularly in the video understanding [19, 23, 26, 27, 44, 46, 48]. The pursuit of generalist Large Video-Language Models (LVidLMs) will be a perennial challenge. Success in this endeavor hinges on effectively leveraging the exceptional understanding, reasoning, and generative capacities inherently present in LLMs. + +One potential route towards addressing the issue is aligning visual feature into latent space of language feature. To achieve this, existing LVidLMs apply large-scale vanilla instruction tuning [19, 26,46,48]. However, the conventional question-answering training paradigm primarily assists LLMs in understanding visual data from a spatial perspective, posing challenges in effectively capturing + +![](images/8f14ab64a2c6781696041d2a227162010dd5d83385f41668a3c3e5b215c30bb3.jpg) +(a) Illustration of training paradigm comparison. + +![](images/8640e8608f6e32ca3ab0e52c41c304e58cff15c0ca50863f56c6532418777ba1.jpg) +(b) Performance comparison. +Fig. 1: Comparison with existing LVidLMs in terms of alignment paradigm and performance. For Fig. 1b, $QA$ , $TG$ , $DC$ denote question answering, temporal grounding and dense captioning, respectively. + +temporal dynamics and spatial consistency relationships. Therefore, relying solely on instruction tuning proves insufficient for achieving comprehensive video comprehension, given the intricate spatial-temporal data structure involved. It is crucial to align different modalities across both spatial and temporal dimensions. Furthermore, offering more fine-grained cross-modal alignment guidance significantly enhances LVidLMs' ability to comprehend videos [24]. + +To bridge the gap, we introduce a novel LVidLM named PiTe, which employs trajectories to intricately align vision and language across both spatial and temporal dimensions at the pixel level, and the distinction from conventional approaches is illustrated in Figure 1a. By requiring the model to forecast the trajectory of individual objects mentioned in the text within the video, it enables the learning of fine-grained text-to-pixel alignment through exploiting the video context along the temporal dimension and enhancing its ability to generate output based on evidence. + +Subsequently, due to there are no ready-made video-language dataset with moving trajectory of objects, we curate a large-scale video-language dataset PiTe-143k through an automated annotation pipeline. Consequently, as shown in Figure 1b, the proposed PiTe significantly augments the LVidLM's capacity to understand videos comprehensively, leading to promising, competitive, and state-of-the-art performance in question-answering, temporal grounding, and dense captioning tasks under zero-shot conditions. + +Overall, our principal contributions in this paper are summarized as follows: + +- We curate a large scale video-language dataset PiTe-143k with trajectory for all individual objects by automatic annotation pipeline. +- We propose a novel LVidLM PiTe that utilize trajectory to align video and language features across both spatial and temporal dimensions. +- Extensive experimental results and analysis on myriad datasets for zero-shot video question answering, temporal grounding, and dense captioning tasks demonstrate the superiority of PiTe. + +# 2 Related Work + +# 2.1 Large Language Models + +Over the last few years, pioneering foundation language models like GPT-1 [32], BERT [7], GPT-2 [33], and T5 [34] laid the groundwork, but GPT-3 [4] groundbreaking model parameters to 175 billion size to achieve remarkable zero-shot performance. Besides, research on scaling law [14] has steered language models to a larger scale. Therefore, driven by the success of InstructGPT [29] + +and ChatGPT [28] which training by reinforcement learning with human feedback (RLHF) based on GPT-3, Large Language Models (LLMs) has made waves in the natural language processing (NLP) community due to its capabilities in language understanding, logical reasoning, and generation. The GPT's success suggest a promising path towards building LLMs, several open-source LLMs have been proposed following it with similar performance including OPT [49], BLOOM [36], GLM [9], LLaMA [37, 38], and Vicuna [5]. Our investigation delves into leveraging the striking language comprehension and zero-shot generalization abilities of LLMs beyond the confines of linguistic modalities. Specifically, we aim to extend these capabilities to multi-modal scenarios, thereby exploring their potential in processing diverse forms of information across different modalities. + +# 2.2 Large Visual-Language Models + +The surge of LLMs has lead to major advancements in NLP tasks, and also has incited interest in developing Large Visual-Language Models (LVLMs). Building a unify LLM with visual inputs for visual language tasks thus remains one of the most important desiderata for LVLMs. Flamingo [1] and OpenFlamingo [3] fuse visual information into intermediate embedding for a frozen LLM by cross-attention mechanism, and train on billions of image-text pairs to align visual and linguistic modalities. Similarly, BLIP-2 [18] introduced the concept of Q-Former to align visual features more effectively with language space. Moreover, MiniGPT-4 [52] enhances its usability significantly by further fine-tuning on more detailed image descriptions with just one projection layer to align a frozen visual encoder with a frozen LLM, and the LLaVA series [21, 22] use simply a multi-layer perception (MLP) in place of the Q-Former and two-stage instruction tuning to enhance this process. Furthermore, PixelLLM [42] leverage the location coordinate of every word in the caption in the image as the connection between different modalities to strengthen the model's performance for the object detection task. Our primary focus lies in transferring the exceptional language comprehension capabilities of LLMs to the analysis of dynamic, continuous visual data found in videos, as opposed to static visual data such as images. + +# 2.3 Large Video-Language Models + +Recently, many efforts have been made to transfer the task-handling capability of LVLMs to the video modality, leading to the emergence of Large Video-Language Models (LVidLMs) like VideoChat [19], Video-LLaMA [46], and Video-ChatGPT [26]. Prior researches have demonstrated the capability of LLMs to perform diverse tasks on video content, guided by user instructions through a two-stage training process. These studies align static visual features with LLMs, followed by instruction tuning on datasets annotated either by GPT or humans. Despite being effective in video understanding, the lack of fine-grained spatial-temporal modeling in these models prevents them from understanding or locating object in detail or specific segments. We propose a novel fine-grained alignment strategy at the pixel level across spatial and temporal dimensions to enhance the ability of LLMs to comprehensively analyze video content, thereby facilitating a more detailed understanding of the visual information presented. + +# 3 PiTe-143k Dataset + +To facilitate fine-grained multi-modal alignment research at the pixel level, we introduce a large-scale video-language dataset PiTe-143k. This dataset fills a crucial gap in the existing resources by + +Table 1: Comparison between PiTe-143k and existing video instruction datasets. + +
DatasetTotal Dur.Avg. Dur.#Videos#EventsTemporal Localization#Objects Trajectories
VideoChat [19]41h18s8.2k×××
Valley [25]608h40s54.7k×××
Video-ChatGPT [26]432h117s13.3k×××
PiTe-143k (Ours)2086.44h52.18s143.64k343.93k✓1.02M
+ +providing extensive object moving trajectories with video instruction, which were previously unavailable in ready-made datasets. PiTe-143k constructed based on InternVid-10M-FLT [13,40] that each instance contains and entire video and multiple clip captions with start-stop timestamps. As shown in Table 1, PiTe-143k comprises 343.93 thousand event segments and 1.02 million moving trajectory for all individual objects that appear in both visual and textual modalities. To facilitate this objective, we establish an automatic annotation pipeline for PiTe-143k, fostering the advancement of LVidLMs for nuanced pixel-level video comprehension. + +The automatic annotation pipeline for PiTe-143k comprises two primary stages, as depicted in Fig. 2: (1) Stage 1 involves the noun phrases extraction and referring expressions segmentation, thereby generating object masks within the frame for all individual objects referenced in the event caption; (2) Stage 2 centers on point tracking to capture the moving trajectories corresponding to the masks obtained in Stage 1. + +# 3.1 Referring Expression Segmentation + +In stage 1, we aim to build closely fine-grained connection between video and language. To this end, we extract all noun phrases from caption and find the corresponding objects in the clip. + +At inception, we leverage constituency parser SuPar [50,51] for language to extract noun phrase as shown in Fig. 3. Notably, in order to pass the simplest and most straightforward language instructions in next step, we only extract noun phrase from the lowest layer. For example in Fig. 3a, we consider two noun phrases $a$ pen and $a$ white table, but the parent node of the former that denotes $a$ pen on a white table not in our consideration because of the complexity of its composition. Following this, we utilize GLaMM [35], the first LVLM that can generate natural language responses seamlessly intertwined with corresponding object segmentation masks, to obtain the corresponding segmentation mask in the first frame of the clip for the text-based referring expression. While certain objects in the video, such as $a$ pen as illustrated in Fig.3a and Fig.2, may be too small to accurate detection. In such challenging cases, we disregard the trajectory information of the noun phrase. Despite this limitation, it has a minimal impact on the overall performance when utilizing extensive pre-training data on a large scale. Meanwhile, leveraging the exceptional language comprehension capabilities of LLMs, GLaMM can effectively filter out invalid referring expressions, those that do not constitute legal object references, such as front as depicted in Fig. 3b. + +# 3.2 Point Tracking + +In stage 2, we aim to transfer the connection constructed in the previous stage to video, expanding out the temporal dimension specific to video compared to image. To this end, we track all individual + +![](images/f0d3a3cd841ca8cab8227774b97883049cd94b0b152455a8a81e0d75c1ff9e6e.jpg) +Fig. 2: Automatic annotation pipeline for PiTe-143k. The video sample in the figure showcases two events positioned at the commencement and conclusion of the video. The procedure for extracting noun phrases by SuPar [50,51] is elucidated in Fig. 3. + +objects in their clip to obtain the trajectory, the trajectory indicates the connection between video and language in both spatial and temporal dimensions. + +The stage 2 commences when we employ DOT [17], a simple-yet-efficient method for tracking point to recover the trajectory of any scene point, for each clip to capture the trajectories for any point in first frame. According to our observation, the caption of each clip mainly describes simple video content in short sentences, so most of the caption corresponds to just one scene clip, which enables us to track objects that identified in the first frame. Subsequently, filter trajectories according to the segmentation mask of objects obtained in stage 1. So far, we obtain the trajectories for all objects in each clip for each video, we create the connection between video and language from both spatial and temporal through the trajectories, the existence of trajectories in video denotes whether the object exist, and the value of trajectory represent where the object exist in video. Lastly, we utilize the k-means++ [2] clustering algorithm to condense trajectories into three key points, effectively reducing computational demands. This approach is founded on the premise that + +![](images/9d8ed1084e159297b7cd070c45409805a38db4947fea84bd61ed6517bdce7c1f.jpg) +(a) The extraction of the NP for the first event caption of the sample video depicted in Fig. 2. + +![](images/0f8c8bd0de36161106ce4e067048382beb22ce7feaf73171f40b77b56f1969b0.jpg) +(b) Referring to Fig. 2, the extraction of the NP for the second event caption. +Fig. 3: Two samples of constituency parser for Noun Phrase (NP) extraction. + +three points adequately capture the typical geometric shape of objects, striking a balance between precision and computational efficiency. Furthermore, we conduct a comparative analysis of the performance using various key tracking points, as discussed in Section 5.3. + +# 4 PiTe + +In this section, we propose a novel Large Video-Language Model (LVidLM), PiTe, which align video and language by trajectories across both spatial and temporal dimension. Fig. 4 illustrates an overview of PiTe. + +# 4.1 Architecture + +PiTe is composed of a vision encoder to encode frames from video implemented as ViT [8], a vision adapter to project visual feature to semantic space of LLMs implemented as a linear projection layer, a LLM Vicuna v1.5 [5], and a localization projector or trajectory projector in separate training stage to guide LLMs to understand visual information implemented as a linear projection. + +Vision Encoder. Raw video data can be expressed as multiple frames such as $\mathbf{v} = \{f_1, f_2, \dots, f_N\} \in \mathbb{R}^{N \times H \times W \times C}$ (frames $\times$ height $\times$ width $\times$ channels). Following previous studies [22,26,46], we adopt ViT-L/14 [8] pre-trained from CLIP [31] as the vision encoder ViT to encode visual data. We uniformly sample $N$ frames for video $\mathbf{v}$ , and encode $i$ -th frame $f_i$ through the vision encoder ViT: + +$$ +\left\{v _ {i} ^ {\text {c l s}}, v _ {i} ^ {1}, v _ {i} ^ {2}, \dots , v _ {i} ^ {P} \right\} = \operatorname {V i T} \left(f _ {i}\right), \tag {1} +$$ + +where $P$ denotes the number of patches in the vision encoder ViT. + +Visual Adapter. A simpler projector forces the LLMs to learn more on handling visual inputs, leading to better generalization [20]. Hence, we utilize the global feature $v_{i}^{cls}$ from vision encoder ViT as the representation for the $i$ -th frame $f_{i}$ , and we apply a linear projection layer $\mu(\cdot)$ to connect the frame feature into the word embedding space of LLMs: + +$$ +z _ {i} = \mu \left(v _ {i} ^ {\text {c l s}}\right). \tag {2} +$$ + +![](images/82e2601f5aebe873d980632c8138ad2c187307a89ed6732422d0813e2a84bf7e.jpg) +Stage 1: Referring Expression Localization + +![](images/e4aafee559240a026d41dd990a02c06fff2af3f63260553bcee42929f24a77fc.jpg) +Stage 2: Pixel-Temporal Alignment + +![](images/184bfb12f0f5dfe93ddadd8d6b074a66cd13a8b87cabf4922f0da2d30ca1a657.jpg) +Stage 3: Video Question Answering +Fig.4: Schematic of PiTe framework for video-language alignment. + +Subsequently, a sequence of frame tokens $\mathbf{z} = \{z_1,z_2,\dots ,z_N\} \in \mathbb{R}^{N\times d}$ becomes the input that LLMs can understand, $d$ denotes the hidden dimension of LLMs. + +Large Language Model. After we tokenize and encode video into frame tokens $\mathbf{z}$ , we concatenate it with textual tokens $\mathbf{w} = \{w_{1}, w_{2}, \ldots, w_{L}\} \in \mathbb{R}^{L}$ and feed as the input to LLMs, we treat visual input as a foreign language in this process. Based on this, LLMs can further encode the input + +sequence to understand the video and text, then reasoning and generation the response using auto-regressive decoding as follows: + +$$ +h _ {i} = \operatorname {L L M} ^ {-} \left(\mathbf {z}, \mathbf {w} _ {1: i - 1}\right), \tag {3} +$$ + +$$ +w _ {i} = \operatorname {a r g m a x} \left(\mathbf {m} _ {v} \cdot h _ {i}\right), \tag {4} +$$ + +where $\mathrm{LLM}^{-}$ denotes the LLM without the last vocabulary mapping layer, $h_i$ denotes the hidden states of $i$ -th token generated by $\mathrm{LLM}^{-}$ , $\mathbf{m}_v \in \mathbb{R}^{|V| \times d}$ denotes the weight of the linear vocabulary mapping layer. + +# 4.2 Training Strategy + +For PiTe model training, we consider a three stage instruction-tuning procedure, as depicted in Fig. 4: (1) Stage 1 centered around training adapters using image-caption pairs; (2) Stage 2 is focuses on aligning video and language features through trajectories; (3) Stage 3 is dedicated to enhancing the model's comprehension by following human instructions through high-quality dialogue instruction tuning. + +Stage 1: Referring Expression Localization. At the initial stage, we aim to train the visual adapter that can align visual features with semantic space of LLMs. To this end, we employ Localized Narratives dataset [30] that contains annotations of human annotators narrating a given image, together with a mouse trajectory of the annotators' attention during the narration. This gives synchronized locations for all words in the narration sentence, the cross-modal attention of human can be used to train our model as condition to bridge vision and language. + +There is only one visual tokens $\mathbf{z} = \{z_1\} \in \mathbb{R}^{1\times d}$ in this training stage for image instead of video, this can align vision with language in spatial to train adapter without consider temporal information. To use the same language features for localization, we simply add a multi-layer perception (MLP) as localization projector $\varphi (\cdot)$ parallel with the vocabulary mapping layer, which maps the language feature to a 2-dimension location: + +$$ +p _ {i} = \varphi \left(h _ {i}\right), \tag {5} +$$ + +where $p_i$ denotes the predicted coordinate for textual token $w_i$ . + +Overall, the leaning objective in the first stage was calculated by standard label-smoothed cross-entropy loss to train captioning output, and $L_{1}$ regression loss to train the localization output: + +$$ +\mathcal {L} _ {1} = \frac {1}{\ell} \sum_ {i = 1} ^ {\ell} \left(\operatorname {C E} \left(\operatorname {L L M} \left(\mathbf {z}, \mathbf {w} _ {1: i - 1}\right), w _ {i}\right) + \lambda | \hat {p} _ {i} - p _ {i} |\right), \tag {6} +$$ + +where $\ell$ is the length of the generated sequence, $\hat{p}_i$ represents the ground truth of location, and $\mathrm{CE}(\cdot ,\cdot)$ denotes the cross-entropy function. To enhance training efficiency, we utilize LoRA [12] for fine-tuning the LLM. + +Stage 2: Pixel-Temporal Alignment. After stage 1, the LLM model becomes proficient in understanding visual information. In the stage 2, we aim to train the LLM to understand sequential frames in video. To achieve this target, we curate a detailed object tracking dataset PiTe-143k, as described in Section 3 that uses trajectory as condition to bridge vision and language across both spatial and temporal dimensions. Therefore, the alignment guidance in pixel level improves the model's video fine-grained understanding reliability and overall usability. + +Table 2: Training hyper-parameters of PiTe. + +
ConfigurationStage 1 Stage 2 Stage 3ConfigurationStage 1 Stage 2 Stage 3
Vision EncoderOpenAI-CLIP-L/14Learning Rate0.0001
Image/Frame Resolution224×224LoRAr=64 & α=128
Adapter ParameterTunableFrozenFrozenNumerical PrecisionBFloat16
Video Frames100Epoch122
LLMVicuna-7B/13B-v1.5Global Batch Size256
LLM Sequence Length2048Learning Rate ScheduleCosine Decay
OptimizerAdamWWarm-up Ratio0.03
+ +Similar to stage 1, we use the same language features for alignment by a MLP as trajectory projector $\rho(\cdot)$ to map the language feature to a 2-dimension location: + +$$ +\mathbf {p} _ {i} = \rho \left(h _ {i}\right), \tag {7} +$$ + +where $\mathbf{p}_i$ denotes the trajectory matrix in $P$ points and $N$ frames for textual token $w_i$ . Here, we define $p_{ijk}$ indicates the coordinate for token $w_i$ for $i$ -th point for model tracking in frame $f_j$ . + +Overall, the leaning objective in stage 2 was calculated by standard label-smoothed cross-entropy loss to train the generation output, and $L_{1}$ regression loss to train the trajectory output as the condition: + +$$ +\mathcal {L} _ {2} = \frac {1}{\ell} \sum_ {i = 1} ^ {\ell} \left(\operatorname {C E} \left(\operatorname {L L M} \left(\mathbf {z}, \mathbf {w} _ {1: i - 1}\right), w _ {i}\right) + \frac {\lambda}{P \cdot N} \sum_ {j = 1} ^ {P} \sum_ {k = 1} ^ {N} | \hat {p} _ {i j k} - p _ {i j k} |\right), \tag {8} +$$ + +where $P$ is the number of the points for model tracking to generate trajectory, and $\mathbf{z} \in \mathbb{R}^{N \times d}$ represents the sequence of visual embedding. We merge the LoRA trained in the stage 1 with the original model and introduce a new LoRA module. + +It is notable that we use localization projector $\varphi(\cdot)$ trained in previous stage to initialize the trajectory projector $\rho(\cdot)$ . Specifically, we define the weight of localization projector $\varphi(\cdot)$ and trajectory projector $\rho(\cdot)$ as $\mathbf{m}_{\varphi} \in \mathbb{R}^{P \cdot N \cdot 2 \times d}$ and $\mathbf{m}_{\rho} \in \mathbb{R}^{2 \times d}$ , respectively. localization projector $\varphi(\cdot)$ maps a 2-dimension coordinate on the input image for each token of the LLM output, as for trajectory projector $\rho(\cdot)$ , it also output 2-dimension coordinates, but more than $P \cdot N$ times for $P$ points to tracking for $N$ frames. For each point of each frame, the parameter of trajectory projector $\rho(\cdot)$ initialized by the localization projector $\varphi(\cdot)$ : + +$$ +\mathbf {m} _ {\varphi} = \overbrace {\mathbf {m} _ {\rho} \oplus \mathbf {m} _ {\rho} \oplus \cdots \oplus \mathbf {m} _ {\rho}} ^ {P \cdot N}, \tag {9} +$$ + +where $\oplus$ denotes concatenation of matrix in first dimension. + +Beyond trajectories, our model is attuned to temporal boundaries within the generated text. Specifically, we structure the generation as ..., from s to e or From s to e, ... to facilitate the model to learn in temporal dimension. Here, ... encapsulates the event description, while s and e denote the frame indexes corresponding to the start and end timestamps of event, respectively. This approach further augments the model's understanding of temporal boundaries [13]. + +Dissimilar to the initial training stage, not all generated words are associated with trajectories. In cases where objects lack trajectories or vanish from view over time, we uniformly assign the coordinates of their ground truth as $(-1, -1)$ to signify their absence. + +Table 3: Comparison between different LVidLMs on zero-shot question-answer. + +
ModelLLM SizeMSVD-QA [41]MSRVTT-QA [43]ActivityNet-QA [45]
Accuracy↑Score↑Accuracy↑Score↑Accuracy↑Score↑
FrozenBiLM [44]1B32.2-16.8-24.7-
LLaMA-Adapter [48]7B54.93.143.82.734.22.7
VideoChat [19]7B56.32.845.02.5-2.2
Video-LLaMA [46]7B51.62.529.61.812.41.1
Video-ChatGPT [26]7B64.93.349.32.835.22.7
PG-Video-LLaVA [27]7B64.13.751.63.339.93.3
PiTe (Ours)7B68.43.956.43.542.03.3
PiTe (Ours)13B71.64.057.73.542.23.4
+ +Stage 3: Video Question Answering. Following stage 2, we incorporate high-quality dialogue data Valley [25] and Video-ChatGPT [26] in one turn for instruction tuning, enabling the model to follow human instructions for more accurate and generalize capabilities of video understanding. + +The leaning objective in third stage was calculated by standard label-smoothed cross-entropy loss for auto-regression generation: + +$$ +\mathcal {L} _ {3} = \frac {1}{\ell} \sum_ {i = 1} ^ {\ell} \operatorname {C E} \left(\text {L L M} \left(\mathbf {z}, \mathbf {w} _ {1: i - 1}\right), w _ {i}\right). \tag {10} +$$ + +Similar to stage 2, we merge the LoRA trained in the stage 1 and stage 2 with the original model and introduce a new LoRA module. + +# 5 Experiments + +# 5.1 Experimental Setup + +Tasks, Datasets, and Evaluation Metrics. We conduct a quantitative evaluation of LVidLMs' video understanding capabilities across three tasks: (1) Video Question Answering: This task assesses the comprehensive video comprehension abilities of LVidLMs by requiring the model to answer a variety of questions about the video content based on its understanding. We perform this task on three datasets: MSVD-QA [41], MSRVTT-QA [43], and ActivityNet-QA [45]. The evaluation pipeline for video understanding follows Video-ChatGPT [26], and we report the accuracy and score, which is assessed using GPT-Assistant [28]. (2) Video Temporal Grounding: This task evaluates LVidLMs' capacity to discern the starting and ending timestamps of a segment corresponding to the description of a video clip. This task demands the model to effectively grasp the temporal aspects of the video. We conduct this task on the ActivityNet Captions dataset [15] and calculate Intersection over Union (IoU) between the model-generate time segments and the ground truth time segments. We report mean IoU (mIoU) and Recall@1, IoU $\geq m$ (R@m) metric, where $m$ values are set at $\{0.3, 0.5, 0.7\}$ . (3) Video Dense Captioning: This task requires the model to produce all events depict in the video along with their corresponding start and end timestamps. It necessitates the model to comprehend both the spatial and temporal dimensions of the video simultaneously. + +Table 4: Comparison between different LVidLMs in temporal video grounding and dense video captioning tasks on ActivityNet [11]. + +
ModelLLM SizeTemporal GroundingDense Captioning
R@0.3↑R@0.5↑R@0.7↑mIoU↑SODA_c↑CIDEr↑METEOR↑
VideoChat [19]7B8.83.71.57.20.92.20.9
Video-LLaMA [46]7B6.92.10.86.51.95.81.9
Video-ChatGPT [26]7B26.413.66.118.91.95.82.1
PiTe (Ours)7B30.417.87.822.05.121.75.8
PiTe (Ours)13B37.223.710.926.05.926.56.6
+ +We conduct this task on the ActivityNet Captions dataset [15]. Initially, we reported SODA_c [10], followed by averages of CIDEr [39] and METEOR [16] under different IoU thresholds of 0.3, 0.5, 0.7, 0.9 based on generate events and ground truth matched pairs to provide a comprehensive analysis. In this paper, all experiments were conducted in a zero-shot setting, and higher values of all evaluation metrics indicate superior performance. + +Implementation Details. In this paper, we employ Vicuna v1.5 [5] as LLM to train the PiTe model at two scales: 7B and 13B. Leveraging the efficiency of LoRA [12], the training of the 7B model can be completed in approximately 10 hours using a single Nvidia 8-A100 (80GB VRAM) node, while the 13B model requires around 17 hours. More hyper-parameter settings are shwon in Table 2. + +# 5.2 Main Result + +Table 3 and 4 present the comparative performance of the PiTe against state-of-the-art baselines on myriad video understanding datasets. + +Question Answering. As illustrated in Table 3, PiTe consistently outperforms the state-of-the-art pure instruction-tuning baselines in terms of all metrics on all datasets. Compared to the top-performing baselines in each dataset, PiTe exhibited notable improvements in the average question answering accuracy, achieving a maximum enhancement of 4.8 and an average improvement of 3.7. For example, PiTe substantially improves accuracy by 64.9 to 68.4 compared to Video-ChatGPT [26] in MSVD-QA dataset [41]. The results showcasing PiTe's proficiency in video comprehension and its capacity to deliver contextually relevant responses according to the given instructions. + +Temporal Grounding. As depicted in Table 4, PiTe achieves state-of-the-art performance in the video temporal grounding task across all metrics as well, demonstrating improvements ranging from 18.9 to 22.0 in mIoU compared to Video-ChatGPT [26]. This clearly indicates that trajectory alignment greatly enhances the ability of capture events in temporal dimension for LVidLMs. The incorporation of object trajectories in the temporal dimension of the trajectory matrix equips the model with a precise understanding of temporal event boundaries, thereby establishing a solid foundation for accurate event localization. + +Table 5: Ablation study of the three-stage training strategy. + +
MethodMSVD-QAActivityNet
Temporal GroundingDense Captioning
Accuracy ScoreR@0.3R@0.5R@0.7mIoUSODA_cCIDErMETEOR
PiTe (Ours)68.43.930.417.87.822.05.121.75.8
w/o initialize68.23.922.810.54.617.15.121.75.8
w/o trajectory68.13.923.912.85.717.45.021.45.8
+ +Dense Captioning. The outcomes of the dense captioning task, as delineated in Table 4, reveal that PiTe consistent boost compared to all state-of-the-art baselines. Particularly noteworthy is the substantial 15.9 increase in the CIDEr metric [39] when compared to Video-ChatGPT [26]. This underscores the significance of fine-grained alignment in both spatial and temporal dimensions through trajectories, implying that PiTe acquires more generalized and detailed representations to offer more sophisticated event descriptions and accurate event temporal boundaries. + +# 5.3 Analysis + +Ablation Study. As reported in Table 5, we conduct ablation experiments on MVSD-QA [41] for question answering and ActivityNet Captions [15] for temporal grounding to verify the individual effects of the proposed contributions under the following settings: (1) w/o initialize: we remove the initialization strategy that use weight of localization projector to initialize trajectory projector; (2) w/o trajectory: we abandon the fine-grained alignment strategy via trajectory. + +From the experimental results in Table 5, it can be observed that: (1) Eliminating the initialization strategy for the trajectory projector in PiTe reduces the model's reasoning capabilities and temporal boundary awareness. However, the performance in dense captioning generation remains consistent. This observation suggests that the model maintains its basic ability in comprehending visual content under trajectory-guided training. (2) The removal of the trajectory-guided training diminishes almost all the capability of PiTe, including dense captioning. (3) Without trajectory-guided training, PiTe demonstrates superior performance compared to trajectory-guided training without the initialization strategy for the trajectory projector in temporal grounding. This outcome highlights the difficulty of trajectory-guided training without initialization from a pre-trained localization projector, as the instability of parameters can impede the model's perception to accurately perceive visual temporal information. + +Exhibition. To better illustrate the video dialogue performance of PiTe, we present a qualitative example, as shown in Fig. 5a. The illustration from the upper portion of the figure demonstrates PiTe's capability not only to provide precise responses to instruction queries but also to enhance the output with more detailed and accurate video information. The example in the lower segment of the figure highlights the model's proficiency in understanding instruction and capturing event, enabling precise delineation of temporal boundaries within the video, despite the constraint of a 100-frame sampling limit. + +![](images/e76a39fe0b891717df0d44f7843fa486ca7779662b9d3a4aff474b487a9ebcbd.jpg) +[Instruction] Who are shown performing impressive feats? + +![](images/b9ee5ca2d267772e80f9143a833f9eb65a4cf5c5520d1a49cb8e1587f2a53ab0.jpg) +[Response] The video showcases various football players performing impressive feats, including a player jumping over another player and a player making a spectacular catch. +[Instruction] During which frames can we see he then shines the shoes of one man who pays him and shakes his hand. +[Response] The shoe shining and payment occur from 66 to 99 +(116.348s ~ 174.522s). Ground truth: 116.06s ~ 172.78s + +![](images/4c6152d1ab3082d20f1d939cae6b4688ccc0a009afb2db382e985e60a717d404.jpg) +(a) Examples of PiTe's video understanding capabilities. +(b) Performance comparison for different tracking point quantity. +Fig. 5: PiTe's video understanding capabilities and performance comparison across varying tracking point quantities. + +Impact of Tracking Point Quantity. In Fig. 5b, we vary the tracking point quantity $P$ in set of $\{1,3,5\}$ . The efficacy of dense captioning tasks tends to improve with an increase in tracking points. However, it is observed that the temporal grounding task undergoes an initial substantial improvement, only to be succeeded by a rapid decline. Less number of tracking points fails to accurately capture the object's geometry, thereby hindering the pixel-level cross-modal alignment guidance for the model. Conversely, a higher quantity of points can enhance the model's comprehension of pure visual information; however, it also introduces noise to make training more challenging. Overall, that the optimal value of $P$ may be different for different tasks, we set $P = 3$ due to its performance maintain stability over multiple tasks. + +# 6 Conclusion + +In this paper, we focus on enhancing the performance of Large Video-Language Models (LVidLMs) by incorporating trajectory-based alignment across different modalities. To achieve fine-grained alignment between video and language across spatial and temporal dimensions, we initially curate a comprehensive multi-modal object tracking dataset, PiTe-143k, using a fully automated annotation pipeline. This dataset was developed to address the lack of large-scale video-language datasets that include multi-object moving trajectories. Subsequently, we introduce a novel Pixel-Temporal (PiTe) alignment strategy that leverages trajectory-guided pre-training to address the inherent challenges faced by LVidLMs. Through comparative analyses, we evaluate PiTe against state-of-the-art models and competitive baselines across various tasks in a zero-shot setting, including question-answering, temporal grounding, and dense captioning, showcasing the superior performance of PiTe with more sophisticated event descriptions and accurate event temporal boundaries. + +# Acknowledgments + +This work was supported by the National Science and Technology Innovation 2030 - Major Project (Grant No. 2022ZD0208800), and NSFC General Program (Grant No. 62176215). + +# References + +1. Alayrac, J., Donahue, J., Luc, P., Miech, A., Barr, I., Hasson, Y., Lenc, K., Mensch, A., Millican, K., Reynolds, M., Ring, R., Rutherford, E., Cabi, S., Han, T., Gong, Z., Samangooei, S., Monteiro, M., Menick, J.L., Borgeaud, S., Brock, A., Nematzadeh, A., Sharifzadeh, S., Binkowski, M., Barreira, R., Vinyals, O., Zisserman, A., Simonyan, K.: Flamingo: a visual language model for few-shot learning. In: Proc. of NeurIPS (2022) +2. Arthur, D., Vassilvitskii, S.: k-means++: the advantages of careful seeding. In: Proc. of SODA. pp. 1027-1035 (2007) +3. Awadalla, A., Gao, I., Gardner, J., Hessel, J., Hanafy, Y., Zhu, W., Marathe, K., Bitton, Y., Gadre, S.Y., Sagawa, S., Jitsev, J., Kornblith, S., Koh, P.W., Ilharco, G., Wortman, M., Schmidt, L.: Openflamingo: An open-source framework for training large autoregressive vision-language models. CoRR (2023) +4. Brown, T.B., Mann, B., Ryder, N., Subbiah, M., Kaplan, J., Dhariwal, P., Neelakantan, A., Shyam, P., Sastry, G., Askell, A., Agarwal, S., Herbert-Voss, A., Krueger, G., Henighan, T., Child, R., Ramesh, A., Ziegler, D.M., Wu, J., Winter, C., Hesse, C., Chen, M., Sigler, E., Litwin, M., Gray, S., Chess, B., Clark, J., Berner, C., McCandlish, S., Radford, A., Sutskever, I., Amodei, D.: Language models are few-shot learners. In: Proc. of NeurIPS (2020) +5. Chiang, W.L., Li, Z., Lin, Z., Sheng, Y., Wu, Z., Zhang, H., Zheng, L., Zhuang, S., Zhuang, Y., Gonzalez, J.E., Stoica, I., Xing, E.P.: Vicuna: An open-source chatbot impressing gpt-4 with $90\%$ * chatgpt quality (2023) +6. Dai, W., Li, J., Li, D., Tiong, A.M.H., Zhao, J., Wang, W., Li, B., Fung, P., Hoi, S.C.H.: Instructblip: Towards general-purpose vision-language models with instruction tuning. In: Proc. of NeurIPS (2023) +7. Devlin, J., Chang, M., Lee, K., Toutanova, K.: BERT: pre-training of deep bidirectional transformers for language understanding. In: Proc. of AACL. pp. 4171-4186 (2019) +8. Dosovitskiy, A., Beyer, L., Kolesnikov, A., Weissenborn, D., Zhai, X., Unterthiner, T., Dehghani, M., Minderer, M., Heigold, G., Gelly, S., Uszkoreit, J., Houlsby, N.: An image is worth 16x16 words: Transformers for image recognition at scale. In: Proc. of ICLR (2021) +9. Du, Z., Qian, Y., Liu, X., Ding, M., Qiu, J., Yang, Z., Tang, J.: GLM: general language model pretraining with autoregressive blank infilling. In: Proc. of ACL. pp. 320-335 (2022) +10. Fujita, S., Hirao, T., Kamigaito, H., Okumura, M., Nagata, M.: SODA: story oriented dense video captioning evaluation framework. In: Proc. of ECCV. pp. 517-531 (2020) +11. Heilbron, F.C., Escorcia, V., Ghanem, B., Niebles, J.C.: Activitynet: A large-scale video benchmark for human activity understanding. In: Proc. of CVPR. pp. 961-970 (2015) +12. Hu, E.J., Shen, Y., Wallis, P., Allen-Zhu, Z., Li, Y., Wang, S., Wang, L., Chen, W.: Lora: Low-rank adaptation of large language models. In: Proc. of ICLR (2022) +13. Huang, B., Wang, X., Chen, H., Song, Z., Zhu, W.: Vtimellm: Empower llm to grasp video moments. In: Proc. of CVPR. pp. 14271-14280 (2024) +14. Kaplan, J., McCandlish, S., Henighan, T., Brown, T.B., Chess, B., Child, R., Gray, S., Radford, A., Wu, J., Amodei, D.: Scaling laws for neural language models. CoRR (2020) +15. Krishna, R., Hata, K., Ren, F., Fei-Fei, L., Niebles, J.C.: Dense-c captioning events in videos. In: Proc. of ICCV. pp. 706-715 (2017) +16. Lavie, A., Agarwal, A.: METEOR: an automatic metric for MT evaluation with high levels of correlation with human judgments. In: Proc. of WMT@ACL. pp. 228-231 (2007) +17. Le Moing, G., Ponce, J., Schmid, C.: Dense optical tracking: Connecting the dots. In: Proc. of CVPR. pp. 19187-19197 (2024) +18. Li, J., Li, D., Savarese, S., Hoi, S.C.H.: BLIP-2: bootstrapping language-image pre-training with frozen image encoders and large language models. In: Proc. of ICML. pp. 19730-19742 (2023) +19. Li, K., He, Y., Wang, Y., Li, Y., Wang, W., Luo, P., Wang, Y., Wang, L., Qiao, Y.: Videochat: Chatcentric video understanding. CoRR (2023) +20. Lin, J., Yin, H., Ping, W., Molchanov, P., Shoeybi, M., Han, S.: Vila: On pre-training for visual language models. In: Proc. of CVPR. pp. 26689-26699 (2024) + +21. Liu, H., Li, C., Li, Y., Lee, Y.J.: Improved baselines with visual instruction tuning. In: Proc. of CVPR. pp. 26296-26306 (2024) +22. Liu, H., Li, C., Wu, Q., Lee, Y.J.: Visual instruction tuning. In: Proc. of NeurIPS (2023) +23. Liu, Y., Ping, H., Zhang, D., Sun, Q., Li, S., Zhou, G.: Comment-aware multi-modal heterogeneous pre-training for humor detection in short-form videos. In: Proc. of ECAI. pp. 1568-1575 (2023) +24. Liu, Y., Shen, T., Zhang, D., Sun, Q., Li, S., Zhou, G.: Comment-aided video-language alignment via contrastive pre-training for short-form video humor detection. In: Proc. of ICMR. pp. 442-450 (2024) +25. Luo, R., Zhao, Z., Yang, M., Dong, J., Qiu, M., Lu, P., Wang, T., Wei, Z.: Valley: Video assistant with large language model enhanced ability. CoRR (2023) +26. Maaz, M., Rasheed, H.A., Khan, S.H., Khan, F.S.: Video-chatgpt: Towards detailed video understanding via large vision and language models (2024) +27. Munasinghe, S., Thushara, R., Maaz, M., Rasheed, H.A., Khan, S., Shah, M., Khan, F.S.: Pg-video-llava: Pixel grounding large video-language models. CoRR (2023) +28. OpenAI: Chatgpt: Optimizing language models for dialogue (2023) +29. Ouyang, L., Wu, J., Jiang, X., Almeida, D., Wainwright, C.L., Mishkin, P., Zhang, C., Agarwal, S., Slama, K., Ray, A., Schulman, J., Hilton, J., Kelton, F., Miller, L., Simens, M., Askell, A., Welinder, P., Christiano, P.F., Leike, J., Lowe, R.: Training language models to follow instructions with human feedback. In: Proc. of NeurIPS (2022) +30. Pont-Tuset, J., Uijlings, J.R.R., Changpinyo, S., Soricut, R., Ferrari, V.: Connecting vision and language with localized narratives. In: Proc. of ECCV. pp. 647-664 (2020) +31. Radford, A., Kim, J.W., Hallacy, C., Ramesh, A., Goh, G., Agarwal, S., Sastry, G., Askell, A., Mishkin, P., Clark, J., Krueger, G., Sutskever, I.: Learning transferable visual models from natural language supervision. In: Proc. of ICML. pp. 8748-8763 (2021) +32. Radford, A., Narasimhan, K.: Improving language understanding by generative pre-training (2018) +33. Radford, A., Wu, J., Child, R., Luan, D., Amodei, D., Sutskever, I.: Language models are unsupervised multitask learners (2019) +34. Raffel, C., Shazeer, N., Roberts, A., Lee, K., Narang, S., Matena, M., Zhou, Y., Li, W., Liu, P.J.: Exploring the limits of transfer learning with a unified text-to-text transformer. J. Mach. Learn. Res. pp. 140:1-140:67 (2020) +35. Rasheed, H., Maaz, M., Shaji, S., Shaker, A., Khan, S., Cholakkal, H., Anwer, R.M., Xing, E., Yang, M.H., Khan, F.S.: Glamm: Pixel grounding large multimodal model. In: Proc. of CVPR. pp. 13009-13018 (2024) +36. Scao, T.L., Fan, A., Akiki, C., Pavlick, E., Ilic, S., Hesslow, D., Castagné, R., Luccioni, A.S., Yvon, F., Gallé, M., Tow, J., Rush, A.M., Biderman, S., Webson, A., Ammanamanchi, P.S., Wang, T., Sagot, B., Muennighoff, N., del Moral, A.V., Ruwase, O., Bawden, R., Bekman, S., McMillan-Major, A., Beltagy, I., Nguyen, H., Saulnier, L., Tan, S., Suarez, P.O., Sanh, V., Laurençon, H., Jernite, Y., Launay, J., Mitchell, M., Raffel, C., Gokaslan, A., Simhi, A., Soroa, A., Aji, A.F., Alfassy, A., Rogers, A., Nitzav, A.K., Xu, C., Mou, C., Emezue, C., Klamm, C., Leong, C., van Strien, D., Adelani, D.I., et al.: BLOOM: A 176b-parameter open-access multilingual language model. CoRR (2022) +37. Touvron, H., Lavril, T., Izacard, G., Martinet, X., Lachaux, M., Lacroix, T., Rozière, B., Goyal, N., Hambro, E., Azhar, F., Rodriguez, A., Joulin, A., Grave, E., Lample, G.: Llama: Open and efficient foundation language models. CoRR (2023) +38. Touvron, H., Martin, L., Stone, K., Albert, P., Almahairi, A., Babaei, Y., Bashlykov, N., Batra, S., Bhargava, P., Bhosale, S., Bikel, D., Blecher, L., Canton-Ferrer, C., Chen, M., Cucurull, G., Esiobu, D., Fernandes, J., Fu, J., Fu, W., Fuller, B., Gao, C., Goswami, V., Goyal, N., Hartshorn, A., Hosseini, S., Hou, R., Inan, H., Kardas, M., Kerkez, V., Khabsa, M., Kloumann, I., Korenev, A., Koura, P.S., Lachaux, M., Lavril, T., Lee, J., Liskovich, D., Lu, Y., Mao, Y., Martinet, X., Mihaylov, T., Mishra, P., Molybog, I., Nie, Y., Poulton, A., Reizenstein, J., Rungta, R., Saladi, K., Schelten, A., Silva, R., Smith, E.M., Subramanian, R., Tan, X.E., Tang, B., Taylor, R., Williams, A., Kuan, J.X., Xu, P., Yan, Z., Zarov, I., Zhang, Y., Fan, A., Kambadur, M., Narang, S., Rodriguez, A., Stojnic, R., Edunov, S., Scialom, T.: Llama 2: Open foundation and fine-tuned chat models. CoRR (2023) + +39. Vedantam, R., Zitnick, C.L., Parikh, D.: Cider: Consensus-based image description evaluation. In: Proc. of CVPR. pp. 4566-4575 (2015) +40. Wang, Y., He, Y., Li, Y., Li, K., Yu, J., Ma, X., Li, X., Chen, G., Chen, X., Wang, Y., Luo, P., Liu, Z., Wang, Y., Wang, L., Qiao, Y.: Internvid: A large-scale video-text dataset for multimodal understanding and generation. In: Proc. of ICLR (2024) +41. Xu, D., Zhao, Z., Xiao, J., Wu, F., Zhang, H., He, X., Zhuang, Y.: Video question answering via gradually refined attention over appearance and motion. In: Proceedings of the 2017 ACM on Multimedia Conference, MM 2017, Mountain View, CA, USA, October 23-27, 2017. pp. 1645-1653 (2017) +42. Xu, J., Zhou, X., Yan, S., Gu, X., Arnab, A., Sun, C., Wang, X., Schmid, C.: Pixel-aligned language model. In: Proc. of CVPR. pp. 13030-13039 (2024) +43. Xu, J., Mei, T., Yao, T., Rui, Y.: MSR-VTT: A large video description dataset for bridging video and language. In: Proc. of CVPR. pp. 5288-5296 (2016) +44. Yang, A., Miech, A., Sivic, J., Laptev, I., Schmid, C.: Zero-shot video question answering via frozen bidirectional language models. In: Proc. of NeurIPS (2022) +45. Yu, Z., Xu, D., Yu, J., Yu, T., Zhao, Z., Zhuang, Y., Tao, D.: Activitynet-qa: A dataset for understanding complex web videos via question answering. In: Proc. of AAAI. pp. 9127-9134 (2019) +46. Zhang, H., Li, X., Bing, L.: Video-llama: An instruction-tuned audio-visual language model for video understanding. In: Proc. of EMNLP. pp. 543-553 (2023) +47. Zhang, M., Huang, S., Li, W., Wang, D.: Tree structure-aware few-shot image classification via hierarchical aggregation. In: Proc. of ECCV. pp. 453-470 (2022) +48. Zhang, R., Han, J., Liu, C., Zhou, A., Lu, P., Qiao, Y., Li, H., Gao, P.: LLaMA-adapter: Efficient fine-tuning of large language models with zero-initialized attention. In: Proc. of ICLR (2024) +49. Zhang, S., Roller, S., Goyal, N., Artetxe, M., Chen, M., Chen, S., Dewan, C., Diab, M.T., Li, X., Lin, X.V., Mihaylov, T., Ott, M., Shleifer, S., Shuster, K., Simig, D., Koura, P.S., Sridhar, A., Wang, T., Zettlemoyer, L.: OPT: open pre-trained transformer language models. CoRR (2022) +50. Zhang, Y., Li, Z., Zhang, M.: Efficient second-order treecrf for neural dependency parsing. In: Proc. of ACL. pp. 3295-3305 (2020) +51. Zhang, Y., Zhou, H., Li, Z.: Fast and accurate neural CRF constituency parsing. In: Proc. of IJCAI. pp. 4046-4053 (2020) +52. Zhu, D., Chen, J., Shen, X., Li, X., Elhoseiny, M.: MiniGPT-4: Enhancing vision-language understanding with advanced large language models. In: Proc. of ICLR (2024) \ No newline at end of file diff --git a/2024/PiTe_ Pixel-Temporal Alignment for Large Video-Language Model/images.zip b/2024/PiTe_ Pixel-Temporal Alignment for Large Video-Language Model/images.zip new file mode 100644 index 0000000000000000000000000000000000000000..b9f5be53b1b541b4240abdc3705142d30c512f28 --- /dev/null +++ b/2024/PiTe_ Pixel-Temporal Alignment for Large Video-Language Model/images.zip @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:9511d2d98dc2901d1dc05dde4779166f8a7d087f1e00eb01e110933fc38fc80a +size 780097 diff --git a/2024/PiTe_ Pixel-Temporal Alignment for Large Video-Language Model/layout.json b/2024/PiTe_ Pixel-Temporal Alignment for Large Video-Language Model/layout.json new file mode 100644 index 0000000000000000000000000000000000000000..c2868834348f1f9cea4f4e39fab6807b10747652 --- /dev/null +++ b/2024/PiTe_ Pixel-Temporal Alignment for Large Video-Language Model/layout.json @@ -0,0 +1,9259 @@ +{ + "pdf_info": [ + { + "para_blocks": [ + { + "bbox": [ + 160, + 91, + 412, + 126 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 160, + 91, + 412, + 126 + ], + "spans": [ + { + "bbox": [ + 160, + 91, + 412, + 126 + ], + "type": "text", + "content": "PiTe: Pixel-Temporal Alignment for Large Video-Language Model" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 149, + 147, + 422, + 172 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 149, + 147, + 422, + 172 + ], + "spans": [ + { + "bbox": [ + 149, + 147, + 422, + 172 + ], + "type": "text", + "content": "Yang Liu" + }, + { + "bbox": [ + 149, + 147, + 422, + 172 + ], + "type": "inline_equation", + "content": "^{1,2\\star}" + }, + { + "bbox": [ + 149, + 147, + 422, + 172 + ], + "type": "text", + "content": ", Pengxiang Ding" + }, + { + "bbox": [ + 149, + 147, + 422, + 172 + ], + "type": "inline_equation", + "content": "^{1\\star}" + }, + { + "bbox": [ + 149, + 147, + 422, + 172 + ], + "type": "text", + "content": ", Siteng Huang" + }, + { + "bbox": [ + 149, + 147, + 422, + 172 + ], + "type": "inline_equation", + "content": "^{1}" + }, + { + "bbox": [ + 149, + 147, + 422, + 172 + ], + "type": "text", + "content": ", Min Zhang" + }, + { + "bbox": [ + 149, + 147, + 422, + 172 + ], + "type": "inline_equation", + "content": "^{1}" + }, + { + "bbox": [ + 149, + 147, + 422, + 172 + ], + "type": "text", + "content": ", Han Zhao" + }, + { + "bbox": [ + 149, + 147, + 422, + 172 + ], + "type": "inline_equation", + "content": "^{1}" + }, + { + "bbox": [ + 149, + 147, + 422, + 172 + ], + "type": "text", + "content": ", and Donglin Wang" + }, + { + "bbox": [ + 149, + 147, + 422, + 172 + ], + "type": "inline_equation", + "content": "^{1\\boxtimes}" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 149, + 180, + 410, + 204 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 149, + 180, + 410, + 204 + ], + "spans": [ + { + "bbox": [ + 149, + 180, + 410, + 204 + ], + "type": "inline_equation", + "content": "^{1}" + }, + { + "bbox": [ + 149, + 180, + 410, + 204 + ], + "type": "text", + "content": "Westlake University " + }, + { + "bbox": [ + 149, + 180, + 410, + 204 + ], + "type": "inline_equation", + "content": "^{2}" + }, + { + "bbox": [ + 149, + 180, + 410, + 204 + ], + "type": "text", + "content": "Soochow University {liuyang67, dingpengxiang, wangdonglin}@westlake.edu.cn" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 100, + 233, + 473, + 397 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 100, + 233, + 473, + 397 + ], + "spans": [ + { + "bbox": [ + 100, + 233, + 473, + 397 + ], + "type": "text", + "content": "Abstract. Fueled by the Large Language Models (LLMs) wave, Large Visual-Language Models (LVLMs) have emerged as a pivotal advancement, bridging the gap between image and text. However, video making it challenging for LVLMs to perform adequately due to the complexity of the relationship between language and spatial-temporal data structure. Recent Large Video-Language Models (LVidLMs) align feature of static visual data like image into latent space of language feature, by general multi-modal tasks to leverage abilities of LLMs sufficiently. In this paper, we explore fine-grained alignment approach via object trajectory for different modalities across both spatial and temporal dimensions simultaneously. Thus, we propose a novel LVidLM by trajectory-guided Pixel-Temporal Alignment, dubbed PiTe, that exhibits promising applicable model property. To achieve fine-grained video-language alignment, we curate a multi-modal pre-training dataset PiTe-143k, the dataset provision of moving trajectories in pixel level for all individual objects, that appear and mention in the video and caption both, by our automatic annotation pipeline. Meanwhile, PiTe demonstrates astounding capabilities on myriad video-related multi-modal tasks through beat the state-of-the-art methods by a large margin." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 100, + 409, + 473, + 431 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 100, + 409, + 473, + 431 + ], + "spans": [ + { + "bbox": [ + 100, + 409, + 473, + 431 + ], + "type": "text", + "content": "Keywords: Large Video-Language Model " + }, + { + "bbox": [ + 100, + 409, + 473, + 431 + ], + "type": "inline_equation", + "content": "\\cdot" + }, + { + "bbox": [ + 100, + 409, + 473, + 431 + ], + "type": "text", + "content": " Trajectory-guided Instruction Tuning " + }, + { + "bbox": [ + 100, + 409, + 473, + 431 + ], + "type": "inline_equation", + "content": "\\cdot" + }, + { + "bbox": [ + 100, + 409, + 473, + 431 + ], + "type": "text", + "content": " Video Understanding" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 72, + 452, + 167, + 464 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 72, + 452, + 167, + 464 + ], + "spans": [ + { + "bbox": [ + 72, + 452, + 167, + 464 + ], + "type": "text", + "content": "1 Introduction" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 72, + 476, + 501, + 584 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 72, + 476, + 501, + 584 + ], + "spans": [ + { + "bbox": [ + 72, + 476, + 501, + 584 + ], + "type": "text", + "content": "Large Language Models (LLMs) have rapidly gained popularity within the AI community, demonstrating astounding capabilities across a wide array of natural language tasks [4, 5, 9, 29, 37, 38]. The powerful language comprehension abilities of LLMs drive researchers to explore their utility in addressing a broader spectrum of tasks across various domains. Consequently, an increasing number of studies are focusing on developing comprehensive Large Visual-Language Models (LVLMs) to tackle vision-related tasks in zero-shot settings [6, 18, 47, 52], particularly in the video understanding [19, 23, 26, 27, 44, 46, 48]. The pursuit of generalist Large Video-Language Models (LVidLMs) will be a perennial challenge. Success in this endeavor hinges on effectively leveraging the exceptional understanding, reasoning, and generative capacities inherently present in LLMs." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 72, + 586, + 501, + 633 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 72, + 586, + 501, + 633 + ], + "spans": [ + { + "bbox": [ + 72, + 586, + 501, + 633 + ], + "type": "text", + "content": "One potential route towards addressing the issue is aligning visual feature into latent space of language feature. To achieve this, existing LVidLMs apply large-scale vanilla instruction tuning [19, 26,46,48]. However, the conventional question-answering training paradigm primarily assists LLMs in understanding visual data from a spatial perspective, posing challenges in effectively capturing" + } + ] + } + ], + "index": 7 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 82, + 639, + 281, + 651 + ], + "type": "page_footnote", + "angle": 0, + "lines": [ + { + "bbox": [ + 82, + 639, + 281, + 651 + ], + "spans": [ + { + "bbox": [ + 82, + 639, + 281, + 651 + ], + "type": "text", + "content": "* Equal contribution. ☑ Corresponding author." + } + ] + } + ], + "index": 8 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 0 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 116, + 94, + 375, + 195 + ], + "blocks": [ + { + "bbox": [ + 116, + 94, + 375, + 195 + ], + "lines": [ + { + "bbox": [ + 116, + 94, + 375, + 195 + ], + "spans": [ + { + "bbox": [ + 116, + 94, + 375, + 195 + ], + "type": "image", + "image_path": "8f14ab64a2c6781696041d2a227162010dd5d83385f41668a3c3e5b215c30bb3.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 156, + 198, + 329, + 207 + ], + "lines": [ + { + "bbox": [ + 156, + 198, + 329, + 207 + ], + "spans": [ + { + "bbox": [ + 156, + 198, + 329, + 207 + ], + "type": "text", + "content": "(a) Illustration of training paradigm comparison." + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_caption" + } + ], + "index": 2 + }, + { + "type": "image", + "bbox": [ + 393, + 93, + 539, + 196 + ], + "blocks": [ + { + "bbox": [ + 393, + 93, + 539, + 196 + ], + "lines": [ + { + "bbox": [ + 393, + 93, + 539, + 196 + ], + "spans": [ + { + "bbox": [ + 393, + 93, + 539, + 196 + ], + "type": "image", + "image_path": "8640e8608f6e32ca3ab0e52c41c304e58cff15c0ca50863f56c6532418777ba1.jpg" + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 409, + 198, + 513, + 207 + ], + "lines": [ + { + "bbox": [ + 409, + 198, + 513, + 207 + ], + "spans": [ + { + "bbox": [ + 409, + 198, + 513, + 207 + ], + "type": "text", + "content": "(b) Performance comparison." + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 107, + 216, + 538, + 239 + ], + "lines": [ + { + "bbox": [ + 107, + 216, + 538, + 239 + ], + "spans": [ + { + "bbox": [ + 107, + 216, + 538, + 239 + ], + "type": "text", + "content": "Fig. 1: Comparison with existing LVidLMs in terms of alignment paradigm and performance. For Fig. 1b, " + }, + { + "bbox": [ + 107, + 216, + 538, + 239 + ], + "type": "inline_equation", + "content": "QA" + }, + { + "bbox": [ + 107, + 216, + 538, + 239 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 107, + 216, + 538, + 239 + ], + "type": "inline_equation", + "content": "TG" + }, + { + "bbox": [ + 107, + 216, + 538, + 239 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 107, + 216, + 538, + 239 + ], + "type": "inline_equation", + "content": "DC" + }, + { + "bbox": [ + 107, + 216, + 538, + 239 + ], + "type": "text", + "content": " denote question answering, temporal grounding and dense captioning, respectively." + } + ] + } + ], + "index": 6, + "angle": 0, + "type": "image_caption" + } + ], + "index": 4 + }, + { + "bbox": [ + 106, + 254, + 539, + 314 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 106, + 254, + 539, + 314 + ], + "spans": [ + { + "bbox": [ + 106, + 254, + 539, + 314 + ], + "type": "text", + "content": "temporal dynamics and spatial consistency relationships. Therefore, relying solely on instruction tuning proves insufficient for achieving comprehensive video comprehension, given the intricate spatial-temporal data structure involved. It is crucial to align different modalities across both spatial and temporal dimensions. Furthermore, offering more fine-grained cross-modal alignment guidance significantly enhances LVidLMs' ability to comprehend videos [24]." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 106, + 314, + 539, + 386 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 106, + 314, + 539, + 386 + ], + "spans": [ + { + "bbox": [ + 106, + 314, + 539, + 386 + ], + "type": "text", + "content": "To bridge the gap, we introduce a novel LVidLM named PiTe, which employs trajectories to intricately align vision and language across both spatial and temporal dimensions at the pixel level, and the distinction from conventional approaches is illustrated in Figure 1a. By requiring the model to forecast the trajectory of individual objects mentioned in the text within the video, it enables the learning of fine-grained text-to-pixel alignment through exploiting the video context along the temporal dimension and enhancing its ability to generate output based on evidence." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 106, + 386, + 539, + 456 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 106, + 386, + 539, + 456 + ], + "spans": [ + { + "bbox": [ + 106, + 386, + 539, + 456 + ], + "type": "text", + "content": "Subsequently, due to there are no ready-made video-language dataset with moving trajectory of objects, we curate a large-scale video-language dataset PiTe-143k through an automated annotation pipeline. Consequently, as shown in Figure 1b, the proposed PiTe significantly augments the LVidLM's capacity to understand videos comprehensively, leading to promising, competitive, and state-of-the-art performance in question-answering, temporal grounding, and dense captioning tasks under zero-shot conditions." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 120, + 457, + 458, + 469 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 120, + 457, + 458, + 469 + ], + "spans": [ + { + "bbox": [ + 120, + 457, + 458, + 469 + ], + "type": "text", + "content": "Overall, our principal contributions in this paper are summarized as follows:" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 107, + 469, + 539, + 541 + ], + "type": "list", + "angle": 0, + "index": 14, + "blocks": [ + { + "bbox": [ + 107, + 469, + 538, + 493 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 107, + 469, + 538, + 493 + ], + "spans": [ + { + "bbox": [ + 107, + 469, + 538, + 493 + ], + "type": "text", + "content": "- We curate a large scale video-language dataset PiTe-143k with trajectory for all individual objects by automatic annotation pipeline." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 107, + 493, + 539, + 517 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 107, + 493, + 539, + 517 + ], + "spans": [ + { + "bbox": [ + 107, + 493, + 539, + 517 + ], + "type": "text", + "content": "- We propose a novel LVidLM PiTe that utilize trajectory to align video and language features across both spatial and temporal dimensions." + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 107, + 517, + 539, + 541 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 107, + 517, + 539, + 541 + ], + "spans": [ + { + "bbox": [ + 107, + 517, + 539, + 541 + ], + "type": "text", + "content": "- Extensive experimental results and analysis on myriad datasets for zero-shot video question answering, temporal grounding, and dense captioning tasks demonstrate the superiority of PiTe." + } + ] + } + ], + "index": 13 + } + ], + "sub_type": "text" + }, + { + "bbox": [ + 107, + 559, + 212, + 572 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 107, + 559, + 212, + 572 + ], + "spans": [ + { + "bbox": [ + 107, + 559, + 212, + 572 + ], + "type": "text", + "content": "2 Related Work" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 107, + 584, + 257, + 597 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 107, + 584, + 257, + 597 + ], + "spans": [ + { + "bbox": [ + 107, + 584, + 257, + 597 + ], + "type": "text", + "content": "2.1 Large Language Models" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 106, + 603, + 539, + 653 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 106, + 603, + 539, + 653 + ], + "spans": [ + { + "bbox": [ + 106, + 603, + 539, + 653 + ], + "type": "text", + "content": "Over the last few years, pioneering foundation language models like GPT-1 [32], BERT [7], GPT-2 [33], and T5 [34] laid the groundwork, but GPT-3 [4] groundbreaking model parameters to 175 billion size to achieve remarkable zero-shot performance. Besides, research on scaling law [14] has steered language models to a larger scale. Therefore, driven by the success of InstructGPT [29]" + } + ] + } + ], + "index": 17 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 108, + 71, + 116, + 79 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 108, + 71, + 116, + 79 + ], + "spans": [ + { + "bbox": [ + 108, + 71, + 116, + 79 + ], + "type": "text", + "content": "2" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 141, + 70, + 229, + 81 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 141, + 70, + 229, + 81 + ], + "spans": [ + { + "bbox": [ + 141, + 70, + 229, + 81 + ], + "type": "text", + "content": "Y. Liu, P. Ding et al." + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 1 + }, + { + "para_blocks": [ + { + "bbox": [ + 72, + 95, + 502, + 202 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 72, + 95, + 502, + 202 + ], + "spans": [ + { + "bbox": [ + 72, + 95, + 502, + 202 + ], + "type": "text", + "content": "and ChatGPT [28] which training by reinforcement learning with human feedback (RLHF) based on GPT-3, Large Language Models (LLMs) has made waves in the natural language processing (NLP) community due to its capabilities in language understanding, logical reasoning, and generation. The GPT's success suggest a promising path towards building LLMs, several open-source LLMs have been proposed following it with similar performance including OPT [49], BLOOM [36], GLM [9], LLaMA [37, 38], and Vicuna [5]. Our investigation delves into leveraging the striking language comprehension and zero-shot generalization abilities of LLMs beyond the confines of linguistic modalities. Specifically, we aim to extend these capabilities to multi-modal scenarios, thereby exploring their potential in processing diverse forms of information across different modalities." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 72, + 220, + 253, + 232 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 72, + 220, + 253, + 232 + ], + "spans": [ + { + "bbox": [ + 72, + 220, + 253, + 232 + ], + "type": "text", + "content": "2.2 Large Visual-Language Models" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 72, + 239, + 502, + 418 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 72, + 239, + 502, + 418 + ], + "spans": [ + { + "bbox": [ + 72, + 239, + 502, + 418 + ], + "type": "text", + "content": "The surge of LLMs has lead to major advancements in NLP tasks, and also has incited interest in developing Large Visual-Language Models (LVLMs). Building a unify LLM with visual inputs for visual language tasks thus remains one of the most important desiderata for LVLMs. Flamingo [1] and OpenFlamingo [3] fuse visual information into intermediate embedding for a frozen LLM by cross-attention mechanism, and train on billions of image-text pairs to align visual and linguistic modalities. Similarly, BLIP-2 [18] introduced the concept of Q-Former to align visual features more effectively with language space. Moreover, MiniGPT-4 [52] enhances its usability significantly by further fine-tuning on more detailed image descriptions with just one projection layer to align a frozen visual encoder with a frozen LLM, and the LLaVA series [21, 22] use simply a multi-layer perception (MLP) in place of the Q-Former and two-stage instruction tuning to enhance this process. Furthermore, PixelLLM [42] leverage the location coordinate of every word in the caption in the image as the connection between different modalities to strengthen the model's performance for the object detection task. Our primary focus lies in transferring the exceptional language comprehension capabilities of LLMs to the analysis of dynamic, continuous visual data found in videos, as opposed to static visual data such as images." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 72, + 435, + 251, + 448 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 72, + 435, + 251, + 448 + ], + "spans": [ + { + "bbox": [ + 72, + 435, + 251, + 448 + ], + "type": "text", + "content": "2.3 Large Video-Language Models" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 72, + 454, + 502, + 586 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 72, + 454, + 502, + 586 + ], + "spans": [ + { + "bbox": [ + 72, + 454, + 502, + 586 + ], + "type": "text", + "content": "Recently, many efforts have been made to transfer the task-handling capability of LVLMs to the video modality, leading to the emergence of Large Video-Language Models (LVidLMs) like VideoChat [19], Video-LLaMA [46], and Video-ChatGPT [26]. Prior researches have demonstrated the capability of LLMs to perform diverse tasks on video content, guided by user instructions through a two-stage training process. These studies align static visual features with LLMs, followed by instruction tuning on datasets annotated either by GPT or humans. Despite being effective in video understanding, the lack of fine-grained spatial-temporal modeling in these models prevents them from understanding or locating object in detail or specific segments. We propose a novel fine-grained alignment strategy at the pixel level across spatial and temporal dimensions to enhance the ability of LLMs to comprehensively analyze video content, thereby facilitating a more detailed understanding of the visual information presented." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 72, + 603, + 202, + 616 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 72, + 603, + 202, + 616 + ], + "spans": [ + { + "bbox": [ + 72, + 603, + 202, + 616 + ], + "type": "text", + "content": "3 PiTe-143k Dataset" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 72, + 628, + 501, + 651 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 72, + 628, + 501, + 651 + ], + "spans": [ + { + "bbox": [ + 72, + 628, + 501, + 651 + ], + "type": "text", + "content": "To facilitate fine-grained multi-modal alignment research at the pixel level, we introduce a large-scale video-language dataset PiTe-143k. This dataset fills a crucial gap in the existing resources by" + } + ] + } + ], + "index": 8 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 199, + 70, + 470, + 82 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 199, + 70, + 470, + 82 + ], + "spans": [ + { + "bbox": [ + 199, + 70, + 470, + 82 + ], + "type": "text", + "content": "PiTe: Pixel-Temporal Alignment for Large Video-Language Model" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 495, + 72, + 501, + 79 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 495, + 72, + 501, + 79 + ], + "spans": [ + { + "bbox": [ + 495, + 72, + 501, + 79 + ], + "type": "text", + "content": "3" + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 2 + }, + { + "para_blocks": [ + { + "type": "table", + "bbox": [ + 111, + 114, + 536, + 198 + ], + "blocks": [ + { + "bbox": [ + 157, + 94, + 488, + 105 + ], + "lines": [ + { + "bbox": [ + 157, + 94, + 488, + 105 + ], + "spans": [ + { + "bbox": [ + 157, + 94, + 488, + 105 + ], + "type": "text", + "content": "Table 1: Comparison between PiTe-143k and existing video instruction datasets." + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 111, + 114, + 536, + 198 + ], + "lines": [ + { + "bbox": [ + 111, + 114, + 536, + 198 + ], + "spans": [ + { + "bbox": [ + 111, + 114, + 536, + 198 + ], + "type": "table", + "html": "
DatasetTotal Dur.Avg. Dur.#Videos#EventsTemporal Localization#Objects Trajectories
VideoChat [19]41h18s8.2k×××
Valley [25]608h40s54.7k×××
Video-ChatGPT [26]432h117s13.3k×××
PiTe-143k (Ours)2086.44h52.18s143.64k343.93k✓1.02M
", + "image_path": "a8e8dbe6ccddeba9a0d935ac24aea60dcbe0bd1d94d4345c50faa729ad337068.jpg" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "table_body" + } + ], + "index": 3 + }, + { + "bbox": [ + 106, + 220, + 539, + 304 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 106, + 220, + 539, + 304 + ], + "spans": [ + { + "bbox": [ + 106, + 220, + 539, + 304 + ], + "type": "text", + "content": "providing extensive object moving trajectories with video instruction, which were previously unavailable in ready-made datasets. PiTe-143k constructed based on InternVid-10M-FLT [13,40] that each instance contains and entire video and multiple clip captions with start-stop timestamps. As shown in Table 1, PiTe-143k comprises 343.93 thousand event segments and 1.02 million moving trajectory for all individual objects that appear in both visual and textual modalities. To facilitate this objective, we establish an automatic annotation pipeline for PiTe-143k, fostering the advancement of LVidLMs for nuanced pixel-level video comprehension." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 106, + 304, + 539, + 365 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 106, + 304, + 539, + 365 + ], + "spans": [ + { + "bbox": [ + 106, + 304, + 539, + 365 + ], + "type": "text", + "content": "The automatic annotation pipeline for PiTe-143k comprises two primary stages, as depicted in Fig. 2: (1) Stage 1 involves the noun phrases extraction and referring expressions segmentation, thereby generating object masks within the frame for all individual objects referenced in the event caption; (2) Stage 2 centers on point tracking to capture the moving trajectories corresponding to the masks obtained in Stage 1." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 107, + 380, + 313, + 393 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 107, + 380, + 313, + 393 + ], + "spans": [ + { + "bbox": [ + 107, + 380, + 313, + 393 + ], + "type": "text", + "content": "3.1 Referring Expression Segmentation" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 106, + 400, + 538, + 424 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 106, + 400, + 538, + 424 + ], + "spans": [ + { + "bbox": [ + 106, + 400, + 538, + 424 + ], + "type": "text", + "content": "In stage 1, we aim to build closely fine-grained connection between video and language. To this end, we extract all noun phrases from caption and find the corresponding objects in the clip." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 106, + 425, + 539, + 593 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 106, + 425, + 539, + 593 + ], + "spans": [ + { + "bbox": [ + 106, + 425, + 539, + 593 + ], + "type": "text", + "content": "At inception, we leverage constituency parser SuPar [50,51] for language to extract noun phrase as shown in Fig. 3. Notably, in order to pass the simplest and most straightforward language instructions in next step, we only extract noun phrase from the lowest layer. For example in Fig. 3a, we consider two noun phrases " + }, + { + "bbox": [ + 106, + 425, + 539, + 593 + ], + "type": "inline_equation", + "content": "a" + }, + { + "bbox": [ + 106, + 425, + 539, + 593 + ], + "type": "text", + "content": " pen and " + }, + { + "bbox": [ + 106, + 425, + 539, + 593 + ], + "type": "inline_equation", + "content": "a" + }, + { + "bbox": [ + 106, + 425, + 539, + 593 + ], + "type": "text", + "content": " white table, but the parent node of the former that denotes " + }, + { + "bbox": [ + 106, + 425, + 539, + 593 + ], + "type": "inline_equation", + "content": "a" + }, + { + "bbox": [ + 106, + 425, + 539, + 593 + ], + "type": "text", + "content": " pen on a white table not in our consideration because of the complexity of its composition. Following this, we utilize GLaMM [35], the first LVLM that can generate natural language responses seamlessly intertwined with corresponding object segmentation masks, to obtain the corresponding segmentation mask in the first frame of the clip for the text-based referring expression. While certain objects in the video, such as " + }, + { + "bbox": [ + 106, + 425, + 539, + 593 + ], + "type": "inline_equation", + "content": "a" + }, + { + "bbox": [ + 106, + 425, + 539, + 593 + ], + "type": "text", + "content": " pen as illustrated in Fig.3a and Fig.2, may be too small to accurate detection. In such challenging cases, we disregard the trajectory information of the noun phrase. Despite this limitation, it has a minimal impact on the overall performance when utilizing extensive pre-training data on a large scale. Meanwhile, leveraging the exceptional language comprehension capabilities of LLMs, GLaMM can effectively filter out invalid referring expressions, those that do not constitute legal object references, such as front as depicted in Fig. 3b." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 107, + 609, + 212, + 621 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 107, + 609, + 212, + 621 + ], + "spans": [ + { + "bbox": [ + 107, + 609, + 212, + 621 + ], + "type": "text", + "content": "3.2 Point Tracking" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 106, + 628, + 539, + 652 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 106, + 628, + 539, + 652 + ], + "spans": [ + { + "bbox": [ + 106, + 628, + 539, + 652 + ], + "type": "text", + "content": "In stage 2, we aim to transfer the connection constructed in the previous stage to video, expanding out the temporal dimension specific to video compared to image. To this end, we track all individual" + } + ] + } + ], + "index": 10 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 108, + 72, + 116, + 79 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 108, + 72, + 116, + 79 + ], + "spans": [ + { + "bbox": [ + 108, + 72, + 116, + 79 + ], + "type": "text", + "content": "4" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 141, + 70, + 229, + 81 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 141, + 70, + 229, + 81 + ], + "spans": [ + { + "bbox": [ + 141, + 70, + 229, + 81 + ], + "type": "text", + "content": "Y. Liu, P. Ding et al." + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 3 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 78, + 94, + 496, + 430 + ], + "blocks": [ + { + "bbox": [ + 78, + 94, + 496, + 430 + ], + "lines": [ + { + "bbox": [ + 78, + 94, + 496, + 430 + ], + "spans": [ + { + "bbox": [ + 78, + 94, + 496, + 430 + ], + "type": "image", + "image_path": "f0d3a3cd841ca8cab8227774b97883049cd94b0b152455a8a81e0d75c1ff9e6e.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 69, + 438, + 503, + 472 + ], + "lines": [ + { + "bbox": [ + 69, + 438, + 503, + 472 + ], + "spans": [ + { + "bbox": [ + 69, + 438, + 503, + 472 + ], + "type": "text", + "content": "Fig. 2: Automatic annotation pipeline for PiTe-143k. The video sample in the figure showcases two events positioned at the commencement and conclusion of the video. The procedure for extracting noun phrases by SuPar [50,51] is elucidated in Fig. 3." + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_caption" + } + ], + "index": 2 + }, + { + "bbox": [ + 69, + 496, + 501, + 519 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 69, + 496, + 501, + 519 + ], + "spans": [ + { + "bbox": [ + 69, + 496, + 501, + 519 + ], + "type": "text", + "content": "objects in their clip to obtain the trajectory, the trajectory indicates the connection between video and language in both spatial and temporal dimensions." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 69, + 521, + 503, + 653 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 69, + 521, + 503, + 653 + ], + "spans": [ + { + "bbox": [ + 69, + 521, + 503, + 653 + ], + "type": "text", + "content": "The stage 2 commences when we employ DOT [17], a simple-yet-efficient method for tracking point to recover the trajectory of any scene point, for each clip to capture the trajectories for any point in first frame. According to our observation, the caption of each clip mainly describes simple video content in short sentences, so most of the caption corresponds to just one scene clip, which enables us to track objects that identified in the first frame. Subsequently, filter trajectories according to the segmentation mask of objects obtained in stage 1. So far, we obtain the trajectories for all objects in each clip for each video, we create the connection between video and language from both spatial and temporal through the trajectories, the existence of trajectories in video denotes whether the object exist, and the value of trajectory represent where the object exist in video. Lastly, we utilize the k-means++ [2] clustering algorithm to condense trajectories into three key points, effectively reducing computational demands. This approach is founded on the premise that" + } + ] + } + ], + "index": 5 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 199, + 70, + 470, + 82 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 199, + 70, + 470, + 82 + ], + "spans": [ + { + "bbox": [ + 199, + 70, + 470, + 82 + ], + "type": "text", + "content": "PiTe: Pixel-Temporal Alignment for Large Video-Language Model" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 495, + 71, + 503, + 79 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 495, + 71, + 503, + 79 + ], + "spans": [ + { + "bbox": [ + 495, + 71, + 503, + 79 + ], + "type": "text", + "content": "5" + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 4 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 108, + 93, + 299, + 223 + ], + "blocks": [ + { + "bbox": [ + 108, + 93, + 299, + 223 + ], + "lines": [ + { + "bbox": [ + 108, + 93, + 299, + 223 + ], + "spans": [ + { + "bbox": [ + 108, + 93, + 299, + 223 + ], + "type": "image", + "image_path": "9d8ed1084e159297b7cd070c45409805a38db4947fea84bd61ed6517bdce7c1f.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 108, + 224, + 299, + 243 + ], + "lines": [ + { + "bbox": [ + 108, + 224, + 299, + 243 + ], + "spans": [ + { + "bbox": [ + 108, + 224, + 299, + 243 + ], + "type": "text", + "content": "(a) The extraction of the NP for the first event caption of the sample video depicted in Fig. 2." + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_caption" + } + ], + "index": 2 + }, + { + "type": "image", + "bbox": [ + 381, + 93, + 539, + 223 + ], + "blocks": [ + { + "bbox": [ + 381, + 93, + 539, + 223 + ], + "lines": [ + { + "bbox": [ + 381, + 93, + 539, + 223 + ], + "spans": [ + { + "bbox": [ + 381, + 93, + 539, + 223 + ], + "type": "image", + "image_path": "0f8c8bd0de36161106ce4e067048382beb22ce7feaf73171f40b77b56f1969b0.jpg" + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 382, + 224, + 539, + 243 + ], + "lines": [ + { + "bbox": [ + 382, + 224, + 539, + 243 + ], + "spans": [ + { + "bbox": [ + 382, + 224, + 539, + 243 + ], + "type": "text", + "content": "(b) Referring to Fig. 2, the extraction of the NP for the second event caption." + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 164, + 251, + 481, + 263 + ], + "lines": [ + { + "bbox": [ + 164, + 251, + 481, + 263 + ], + "spans": [ + { + "bbox": [ + 164, + 251, + 481, + 263 + ], + "type": "text", + "content": "Fig. 3: Two samples of constituency parser for Noun Phrase (NP) extraction." + } + ] + } + ], + "index": 6, + "angle": 0, + "type": "image_caption" + } + ], + "index": 4 + }, + { + "bbox": [ + 106, + 277, + 539, + 314 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 106, + 277, + 539, + 314 + ], + "spans": [ + { + "bbox": [ + 106, + 277, + 539, + 314 + ], + "type": "text", + "content": "three points adequately capture the typical geometric shape of objects, striking a balance between precision and computational efficiency. Furthermore, we conduct a comparative analysis of the performance using various key tracking points, as discussed in Section 5.3." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 107, + 330, + 160, + 342 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 107, + 330, + 160, + 342 + ], + "spans": [ + { + "bbox": [ + 107, + 330, + 160, + 342 + ], + "type": "text", + "content": "4 PiTe" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 106, + 354, + 539, + 390 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 106, + 354, + 539, + 390 + ], + "spans": [ + { + "bbox": [ + 106, + 354, + 539, + 390 + ], + "type": "text", + "content": "In this section, we propose a novel Large Video-Language Model (LVidLM), PiTe, which align video and language by trajectories across both spatial and temporal dimension. Fig. 4 illustrates an overview of PiTe." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 107, + 406, + 201, + 417 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 107, + 406, + 201, + 417 + ], + "spans": [ + { + "bbox": [ + 107, + 406, + 201, + 417 + ], + "type": "text", + "content": "4.1 Architecture" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 106, + 424, + 539, + 473 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 106, + 424, + 539, + 473 + ], + "spans": [ + { + "bbox": [ + 106, + 424, + 539, + 473 + ], + "type": "text", + "content": "PiTe is composed of a vision encoder to encode frames from video implemented as ViT [8], a vision adapter to project visual feature to semantic space of LLMs implemented as a linear projection layer, a LLM Vicuna v1.5 [5], and a localization projector or trajectory projector in separate training stage to guide LLMs to understand visual information implemented as a linear projection." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 106, + 487, + 543, + 534 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 106, + 487, + 543, + 534 + ], + "spans": [ + { + "bbox": [ + 106, + 487, + 543, + 534 + ], + "type": "text", + "content": "Vision Encoder. Raw video data can be expressed as multiple frames such as " + }, + { + "bbox": [ + 106, + 487, + 543, + 534 + ], + "type": "inline_equation", + "content": "\\mathbf{v} = \\{f_1, f_2, \\dots, f_N\\} \\in \\mathbb{R}^{N \\times H \\times W \\times C}" + }, + { + "bbox": [ + 106, + 487, + 543, + 534 + ], + "type": "text", + "content": " (frames " + }, + { + "bbox": [ + 106, + 487, + 543, + 534 + ], + "type": "inline_equation", + "content": "\\times" + }, + { + "bbox": [ + 106, + 487, + 543, + 534 + ], + "type": "text", + "content": " height " + }, + { + "bbox": [ + 106, + 487, + 543, + 534 + ], + "type": "inline_equation", + "content": "\\times" + }, + { + "bbox": [ + 106, + 487, + 543, + 534 + ], + "type": "text", + "content": " width " + }, + { + "bbox": [ + 106, + 487, + 543, + 534 + ], + "type": "inline_equation", + "content": "\\times" + }, + { + "bbox": [ + 106, + 487, + 543, + 534 + ], + "type": "text", + "content": " channels). Following previous studies [22,26,46], we adopt ViT-L/14 [8] pre-trained from CLIP [31] as the vision encoder ViT to encode visual data. We uniformly sample " + }, + { + "bbox": [ + 106, + 487, + 543, + 534 + ], + "type": "inline_equation", + "content": "N" + }, + { + "bbox": [ + 106, + 487, + 543, + 534 + ], + "type": "text", + "content": " frames for video " + }, + { + "bbox": [ + 106, + 487, + 543, + 534 + ], + "type": "inline_equation", + "content": "\\mathbf{v}" + }, + { + "bbox": [ + 106, + 487, + 543, + 534 + ], + "type": "text", + "content": ", and encode " + }, + { + "bbox": [ + 106, + 487, + 543, + 534 + ], + "type": "inline_equation", + "content": "i" + }, + { + "bbox": [ + 106, + 487, + 543, + 534 + ], + "type": "text", + "content": "-th frame " + }, + { + "bbox": [ + 106, + 487, + 543, + 534 + ], + "type": "inline_equation", + "content": "f_i" + }, + { + "bbox": [ + 106, + 487, + 543, + 534 + ], + "type": "text", + "content": " through the vision encoder ViT:" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 255, + 539, + 539, + 558 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 255, + 539, + 539, + 558 + ], + "spans": [ + { + "bbox": [ + 255, + 539, + 539, + 558 + ], + "type": "interline_equation", + "content": "\\left\\{v _ {i} ^ {\\text {c l s}}, v _ {i} ^ {1}, v _ {i} ^ {2}, \\dots , v _ {i} ^ {P} \\right\\} = \\operatorname {V i T} \\left(f _ {i}\\right), \\tag {1}", + "image_path": "46e1bbde1786f3710ead6aa7a30483c04734b287a1870f37ed5c093d07309fe3.jpg" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 110, + 559, + 405, + 571 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 110, + 559, + 405, + 571 + ], + "spans": [ + { + "bbox": [ + 110, + 559, + 405, + 571 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 110, + 559, + 405, + 571 + ], + "type": "inline_equation", + "content": "P" + }, + { + "bbox": [ + 110, + 559, + 405, + 571 + ], + "type": "text", + "content": " denotes the number of patches in the vision encoder ViT." + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 106, + 586, + 539, + 632 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 106, + 586, + 539, + 632 + ], + "spans": [ + { + "bbox": [ + 106, + 586, + 539, + 632 + ], + "type": "text", + "content": "Visual Adapter. A simpler projector forces the LLMs to learn more on handling visual inputs, leading to better generalization [20]. Hence, we utilize the global feature " + }, + { + "bbox": [ + 106, + 586, + 539, + 632 + ], + "type": "inline_equation", + "content": "v_{i}^{cls}" + }, + { + "bbox": [ + 106, + 586, + 539, + 632 + ], + "type": "text", + "content": " from vision encoder ViT as the representation for the " + }, + { + "bbox": [ + 106, + 586, + 539, + 632 + ], + "type": "inline_equation", + "content": "i" + }, + { + "bbox": [ + 106, + 586, + 539, + 632 + ], + "type": "text", + "content": "-th frame " + }, + { + "bbox": [ + 106, + 586, + 539, + 632 + ], + "type": "inline_equation", + "content": "f_{i}" + }, + { + "bbox": [ + 106, + 586, + 539, + 632 + ], + "type": "text", + "content": ", and we apply a linear projection layer " + }, + { + "bbox": [ + 106, + 586, + 539, + 632 + ], + "type": "inline_equation", + "content": "\\mu(\\cdot)" + }, + { + "bbox": [ + 106, + 586, + 539, + 632 + ], + "type": "text", + "content": " to connect the frame feature into the word embedding space of LLMs:" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 293, + 636, + 539, + 654 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 636, + 539, + 654 + ], + "spans": [ + { + "bbox": [ + 293, + 636, + 539, + 654 + ], + "type": "interline_equation", + "content": "z _ {i} = \\mu \\left(v _ {i} ^ {\\text {c l s}}\\right). \\tag {2}", + "image_path": "f69297f498c65f309c038621faf048b71a4aa14ab0a21eef139526a9631e27e0.jpg" + } + ] + } + ], + "index": 16 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 108, + 71, + 116, + 79 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 108, + 71, + 116, + 79 + ], + "spans": [ + { + "bbox": [ + 108, + 71, + 116, + 79 + ], + "type": "text", + "content": "6" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 141, + 70, + 229, + 81 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 141, + 70, + 229, + 81 + ], + "spans": [ + { + "bbox": [ + 141, + 70, + 229, + 81 + ], + "type": "text", + "content": "Y. Liu, P. Ding et al." + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 5 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 71, + 106, + 501, + 190 + ], + "blocks": [ + { + "bbox": [ + 71, + 92, + 248, + 104 + ], + "lines": [ + { + "bbox": [ + 71, + 92, + 248, + 104 + ], + "spans": [ + { + "bbox": [ + 71, + 92, + 248, + 104 + ], + "type": "text", + "content": "Stage 1: Referring Expression Localization" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 71, + 106, + 501, + 190 + ], + "lines": [ + { + "bbox": [ + 71, + 106, + 501, + 190 + ], + "spans": [ + { + "bbox": [ + 71, + 106, + 501, + 190 + ], + "type": "image", + "image_path": "82e2601f5aebe873d980632c8138ad2c187307a89ed6732422d0813e2a84bf7e.jpg" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_body" + } + ], + "index": 3 + }, + { + "type": "image", + "bbox": [ + 71, + 205, + 501, + 445 + ], + "blocks": [ + { + "bbox": [ + 71, + 197, + 217, + 209 + ], + "lines": [ + { + "bbox": [ + 71, + 197, + 217, + 209 + ], + "spans": [ + { + "bbox": [ + 71, + 197, + 217, + 209 + ], + "type": "text", + "content": "Stage 2: Pixel-Temporal Alignment" + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 71, + 205, + 501, + 445 + ], + "lines": [ + { + "bbox": [ + 71, + 205, + 501, + 445 + ], + "spans": [ + { + "bbox": [ + 71, + 205, + 501, + 445 + ], + "type": "image", + "image_path": "e4aafee559240a026d41dd990a02c06fff2af3f63260553bcee42929f24a77fc.jpg" + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "image_body" + } + ], + "index": 5 + }, + { + "type": "image", + "bbox": [ + 72, + 463, + 499, + 525 + ], + "blocks": [ + { + "bbox": [ + 71, + 449, + 217, + 460 + ], + "lines": [ + { + "bbox": [ + 71, + 449, + 217, + 460 + ], + "spans": [ + { + "bbox": [ + 71, + 449, + 217, + 460 + ], + "type": "text", + "content": "Stage 3: Video Question Answering" + } + ] + } + ], + "index": 6, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 72, + 463, + 499, + 525 + ], + "lines": [ + { + "bbox": [ + 72, + 463, + 499, + 525 + ], + "spans": [ + { + "bbox": [ + 72, + 463, + 499, + 525 + ], + "type": "image", + "image_path": "184bfb12f0f5dfe93ddadd8d6b074a66cd13a8b87cabf4922f0da2d30ca1a657.jpg" + } + ] + } + ], + "index": 7, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 146, + 532, + 425, + 544 + ], + "lines": [ + { + "bbox": [ + 146, + 532, + 425, + 544 + ], + "spans": [ + { + "bbox": [ + 146, + 532, + 425, + 544 + ], + "type": "text", + "content": "Fig.4: Schematic of PiTe framework for video-language alignment." + } + ] + } + ], + "index": 8, + "angle": 0, + "type": "image_caption" + } + ], + "index": 7 + }, + { + "bbox": [ + 70, + 569, + 502, + 594 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 70, + 569, + 502, + 594 + ], + "spans": [ + { + "bbox": [ + 70, + 569, + 502, + 594 + ], + "type": "text", + "content": "Subsequently, a sequence of frame tokens " + }, + { + "bbox": [ + 70, + 569, + 502, + 594 + ], + "type": "inline_equation", + "content": "\\mathbf{z} = \\{z_1,z_2,\\dots ,z_N\\} \\in \\mathbb{R}^{N\\times d}" + }, + { + "bbox": [ + 70, + 569, + 502, + 594 + ], + "type": "text", + "content": " becomes the input that LLMs can understand, " + }, + { + "bbox": [ + 70, + 569, + 502, + 594 + ], + "type": "inline_equation", + "content": "d" + }, + { + "bbox": [ + 70, + 569, + 502, + 594 + ], + "type": "text", + "content": " denotes the hidden dimension of LLMs." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 69, + 616, + 503, + 653 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 69, + 616, + 503, + 653 + ], + "spans": [ + { + "bbox": [ + 69, + 616, + 503, + 653 + ], + "type": "text", + "content": "Large Language Model. After we tokenize and encode video into frame tokens " + }, + { + "bbox": [ + 69, + 616, + 503, + 653 + ], + "type": "inline_equation", + "content": "\\mathbf{z}" + }, + { + "bbox": [ + 69, + 616, + 503, + 653 + ], + "type": "text", + "content": ", we concatenate it with textual tokens " + }, + { + "bbox": [ + 69, + 616, + 503, + 653 + ], + "type": "inline_equation", + "content": "\\mathbf{w} = \\{w_{1}, w_{2}, \\ldots, w_{L}\\} \\in \\mathbb{R}^{L}" + }, + { + "bbox": [ + 69, + 616, + 503, + 653 + ], + "type": "text", + "content": " and feed as the input to LLMs, we treat visual input as a foreign language in this process. Based on this, LLMs can further encode the input" + } + ] + } + ], + "index": 10 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 199, + 70, + 470, + 82 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 199, + 70, + 470, + 82 + ], + "spans": [ + { + "bbox": [ + 199, + 70, + 470, + 82 + ], + "type": "text", + "content": "PiTe: Pixel-Temporal Alignment for Large Video-Language Model" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 495, + 71, + 503, + 79 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 495, + 71, + 503, + 79 + ], + "spans": [ + { + "bbox": [ + 495, + 71, + 503, + 79 + ], + "type": "text", + "content": "7" + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 6 + }, + { + "para_blocks": [ + { + "bbox": [ + 106, + 95, + 538, + 120 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 106, + 95, + 538, + 120 + ], + "spans": [ + { + "bbox": [ + 106, + 95, + 538, + 120 + ], + "type": "text", + "content": "sequence to understand the video and text, then reasoning and generation the response using auto-regressive decoding as follows:" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 273, + 129, + 539, + 140 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 273, + 129, + 539, + 140 + ], + "spans": [ + { + "bbox": [ + 273, + 129, + 539, + 140 + ], + "type": "interline_equation", + "content": "h _ {i} = \\operatorname {L L M} ^ {-} \\left(\\mathbf {z}, \\mathbf {w} _ {1: i - 1}\\right), \\tag {3}", + "image_path": "b9959e15853b2fa41265d58b1ddaca4afa447a751a230e7f40251de4d58263a7.jpg" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 274, + 144, + 539, + 155 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 274, + 144, + 539, + 155 + ], + "spans": [ + { + "bbox": [ + 274, + 144, + 539, + 155 + ], + "type": "interline_equation", + "content": "w _ {i} = \\operatorname {a r g m a x} \\left(\\mathbf {m} _ {v} \\cdot h _ {i}\\right), \\tag {4}", + "image_path": "47ef6940877bf1a8f8499943a933976f6ae45ad3028e13c1287897ecfee7a4ee.jpg" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 106, + 164, + 539, + 200 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 106, + 164, + 539, + 200 + ], + "spans": [ + { + "bbox": [ + 106, + 164, + 539, + 200 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 106, + 164, + 539, + 200 + ], + "type": "inline_equation", + "content": "\\mathrm{LLM}^{-}" + }, + { + "bbox": [ + 106, + 164, + 539, + 200 + ], + "type": "text", + "content": " denotes the LLM without the last vocabulary mapping layer, " + }, + { + "bbox": [ + 106, + 164, + 539, + 200 + ], + "type": "inline_equation", + "content": "h_i" + }, + { + "bbox": [ + 106, + 164, + 539, + 200 + ], + "type": "text", + "content": " denotes the hidden states of " + }, + { + "bbox": [ + 106, + 164, + 539, + 200 + ], + "type": "inline_equation", + "content": "i" + }, + { + "bbox": [ + 106, + 164, + 539, + 200 + ], + "type": "text", + "content": "-th token generated by " + }, + { + "bbox": [ + 106, + 164, + 539, + 200 + ], + "type": "inline_equation", + "content": "\\mathrm{LLM}^{-}" + }, + { + "bbox": [ + 106, + 164, + 539, + 200 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 106, + 164, + 539, + 200 + ], + "type": "inline_equation", + "content": "\\mathbf{m}_v \\in \\mathbb{R}^{|V| \\times d}" + }, + { + "bbox": [ + 106, + 164, + 539, + 200 + ], + "type": "text", + "content": " denotes the weight of the linear vocabulary mapping layer." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 107, + 217, + 224, + 228 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 107, + 217, + 224, + 228 + ], + "spans": [ + { + "bbox": [ + 107, + 217, + 224, + 228 + ], + "type": "text", + "content": "4.2 Training Strategy" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 106, + 236, + 539, + 295 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 106, + 236, + 539, + 295 + ], + "spans": [ + { + "bbox": [ + 106, + 236, + 539, + 295 + ], + "type": "text", + "content": "For PiTe model training, we consider a three stage instruction-tuning procedure, as depicted in Fig. 4: (1) Stage 1 centered around training adapters using image-caption pairs; (2) Stage 2 is focuses on aligning video and language features through trajectories; (3) Stage 3 is dedicated to enhancing the model's comprehension by following human instructions through high-quality dialogue instruction tuning." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 106, + 312, + 539, + 382 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 106, + 312, + 539, + 382 + ], + "spans": [ + { + "bbox": [ + 106, + 312, + 539, + 382 + ], + "type": "text", + "content": "Stage 1: Referring Expression Localization. At the initial stage, we aim to train the visual adapter that can align visual features with semantic space of LLMs. To this end, we employ Localized Narratives dataset [30] that contains annotations of human annotators narrating a given image, together with a mouse trajectory of the annotators' attention during the narration. This gives synchronized locations for all words in the narration sentence, the cross-modal attention of human can be used to train our model as condition to bridge vision and language." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 106, + 383, + 539, + 437 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 106, + 383, + 539, + 437 + ], + "spans": [ + { + "bbox": [ + 106, + 383, + 539, + 437 + ], + "type": "text", + "content": "There is only one visual tokens " + }, + { + "bbox": [ + 106, + 383, + 539, + 437 + ], + "type": "inline_equation", + "content": "\\mathbf{z} = \\{z_1\\} \\in \\mathbb{R}^{1\\times d}" + }, + { + "bbox": [ + 106, + 383, + 539, + 437 + ], + "type": "text", + "content": " in this training stage for image instead of video, this can align vision with language in spatial to train adapter without consider temporal information. To use the same language features for localization, we simply add a multi-layer perception (MLP) as localization projector " + }, + { + "bbox": [ + 106, + 383, + 539, + 437 + ], + "type": "inline_equation", + "content": "\\varphi (\\cdot)" + }, + { + "bbox": [ + 106, + 383, + 539, + 437 + ], + "type": "text", + "content": " parallel with the vocabulary mapping layer, which maps the language feature to a 2-dimension location:" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 301, + 439, + 538, + 450 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 301, + 439, + 538, + 450 + ], + "spans": [ + { + "bbox": [ + 301, + 439, + 538, + 450 + ], + "type": "interline_equation", + "content": "p _ {i} = \\varphi \\left(h _ {i}\\right), \\tag {5}", + "image_path": "cbb38e89b6dc16f9998161e319751ae6b743a1377136ed8f5b75bdb4f249641f.jpg" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 110, + 453, + 389, + 464 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 110, + 453, + 389, + 464 + ], + "spans": [ + { + "bbox": [ + 110, + 453, + 389, + 464 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 110, + 453, + 389, + 464 + ], + "type": "inline_equation", + "content": "p_i" + }, + { + "bbox": [ + 110, + 453, + 389, + 464 + ], + "type": "text", + "content": " denotes the predicted coordinate for textual token " + }, + { + "bbox": [ + 110, + 453, + 389, + 464 + ], + "type": "inline_equation", + "content": "w_i" + }, + { + "bbox": [ + 110, + 453, + 389, + 464 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 106, + 464, + 538, + 487 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 106, + 464, + 538, + 487 + ], + "spans": [ + { + "bbox": [ + 106, + 464, + 538, + 487 + ], + "type": "text", + "content": "Overall, the leaning objective in the first stage was calculated by standard label-smoothed cross-entropy loss to train captioning output, and " + }, + { + "bbox": [ + 106, + 464, + 538, + 487 + ], + "type": "inline_equation", + "content": "L_{1}" + }, + { + "bbox": [ + 106, + 464, + 538, + 487 + ], + "type": "text", + "content": " regression loss to train the localization output:" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 216, + 494, + 539, + 521 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 216, + 494, + 539, + 521 + ], + "spans": [ + { + "bbox": [ + 216, + 494, + 539, + 521 + ], + "type": "interline_equation", + "content": "\\mathcal {L} _ {1} = \\frac {1}{\\ell} \\sum_ {i = 1} ^ {\\ell} \\left(\\operatorname {C E} \\left(\\operatorname {L L M} \\left(\\mathbf {z}, \\mathbf {w} _ {1: i - 1}\\right), w _ {i}\\right) + \\lambda | \\hat {p} _ {i} - p _ {i} |\\right), \\tag {6}", + "image_path": "0c88620ecb1187a8760da04cdf182451ece113ef7cc762e1a60d5397ac05fc88.jpg" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 107, + 527, + 539, + 563 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 107, + 527, + 539, + 563 + ], + "spans": [ + { + "bbox": [ + 107, + 527, + 539, + 563 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 107, + 527, + 539, + 563 + ], + "type": "inline_equation", + "content": "\\ell" + }, + { + "bbox": [ + 107, + 527, + 539, + 563 + ], + "type": "text", + "content": " is the length of the generated sequence, " + }, + { + "bbox": [ + 107, + 527, + 539, + 563 + ], + "type": "inline_equation", + "content": "\\hat{p}_i" + }, + { + "bbox": [ + 107, + 527, + 539, + 563 + ], + "type": "text", + "content": " represents the ground truth of location, and " + }, + { + "bbox": [ + 107, + 527, + 539, + 563 + ], + "type": "inline_equation", + "content": "\\mathrm{CE}(\\cdot ,\\cdot)" + }, + { + "bbox": [ + 107, + 527, + 539, + 563 + ], + "type": "text", + "content": " denotes the cross-entropy function. To enhance training efficiency, we utilize LoRA [12] for fine-tuning the LLM." + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 106, + 579, + 539, + 652 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 106, + 579, + 539, + 652 + ], + "spans": [ + { + "bbox": [ + 106, + 579, + 539, + 652 + ], + "type": "text", + "content": "Stage 2: Pixel-Temporal Alignment. After stage 1, the LLM model becomes proficient in understanding visual information. In the stage 2, we aim to train the LLM to understand sequential frames in video. To achieve this target, we curate a detailed object tracking dataset PiTe-143k, as described in Section 3 that uses trajectory as condition to bridge vision and language across both spatial and temporal dimensions. Therefore, the alignment guidance in pixel level improves the model's video fine-grained understanding reliability and overall usability." + } + ] + } + ], + "index": 15 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 108, + 71, + 116, + 79 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 108, + 71, + 116, + 79 + ], + "spans": [ + { + "bbox": [ + 108, + 71, + 116, + 79 + ], + "type": "text", + "content": "8" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 141, + 70, + 229, + 81 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 141, + 70, + 229, + 81 + ], + "spans": [ + { + "bbox": [ + 141, + 70, + 229, + 81 + ], + "type": "text", + "content": "Y. Liu, P. Ding et al." + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 7 + }, + { + "para_blocks": [ + { + "type": "table", + "bbox": [ + 74, + 114, + 501, + 216 + ], + "blocks": [ + { + "bbox": [ + 193, + 93, + 380, + 105 + ], + "lines": [ + { + "bbox": [ + 193, + 93, + 380, + 105 + ], + "spans": [ + { + "bbox": [ + 193, + 93, + 380, + 105 + ], + "type": "text", + "content": "Table 2: Training hyper-parameters of PiTe." + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 74, + 114, + 501, + 216 + ], + "lines": [ + { + "bbox": [ + 74, + 114, + 501, + 216 + ], + "spans": [ + { + "bbox": [ + 74, + 114, + 501, + 216 + ], + "type": "table", + "html": "
ConfigurationStage 1 Stage 2 Stage 3ConfigurationStage 1 Stage 2 Stage 3
Vision EncoderOpenAI-CLIP-L/14Learning Rate0.0001
Image/Frame Resolution224×224LoRAr=64 & α=128
Adapter ParameterTunableFrozenFrozenNumerical PrecisionBFloat16
Video Frames100Epoch122
LLMVicuna-7B/13B-v1.5Global Batch Size256
LLM Sequence Length2048Learning Rate ScheduleCosine Decay
OptimizerAdamWWarm-up Ratio0.03
", + "image_path": "080a10cf813006985bd4788c39f7eef82d19b37dcac120f4f7c7ead73b0de48d.jpg" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "table_body" + } + ], + "index": 3 + }, + { + "bbox": [ + 70, + 239, + 502, + 264 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 70, + 239, + 502, + 264 + ], + "spans": [ + { + "bbox": [ + 70, + 239, + 502, + 264 + ], + "type": "text", + "content": "Similar to stage 1, we use the same language features for alignment by a MLP as trajectory projector " + }, + { + "bbox": [ + 70, + 239, + 502, + 264 + ], + "type": "inline_equation", + "content": "\\rho(\\cdot)" + }, + { + "bbox": [ + 70, + 239, + 502, + 264 + ], + "type": "text", + "content": " to map the language feature to a 2-dimension location:" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 263, + 273, + 502, + 285 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 263, + 273, + 502, + 285 + ], + "spans": [ + { + "bbox": [ + 263, + 273, + 502, + 285 + ], + "type": "interline_equation", + "content": "\\mathbf {p} _ {i} = \\rho \\left(h _ {i}\\right), \\tag {7}", + "image_path": "1bf7c3723f6148117316cfa59c156a4956c0d1392770cf27da8e3d287636a5e1.jpg" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 69, + 294, + 502, + 318 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 69, + 294, + 502, + 318 + ], + "spans": [ + { + "bbox": [ + 69, + 294, + 502, + 318 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 69, + 294, + 502, + 318 + ], + "type": "inline_equation", + "content": "\\mathbf{p}_i" + }, + { + "bbox": [ + 69, + 294, + 502, + 318 + ], + "type": "text", + "content": " denotes the trajectory matrix in " + }, + { + "bbox": [ + 69, + 294, + 502, + 318 + ], + "type": "inline_equation", + "content": "P" + }, + { + "bbox": [ + 69, + 294, + 502, + 318 + ], + "type": "text", + "content": " points and " + }, + { + "bbox": [ + 69, + 294, + 502, + 318 + ], + "type": "inline_equation", + "content": "N" + }, + { + "bbox": [ + 69, + 294, + 502, + 318 + ], + "type": "text", + "content": " frames for textual token " + }, + { + "bbox": [ + 69, + 294, + 502, + 318 + ], + "type": "inline_equation", + "content": "w_i" + }, + { + "bbox": [ + 69, + 294, + 502, + 318 + ], + "type": "text", + "content": ". Here, we define " + }, + { + "bbox": [ + 69, + 294, + 502, + 318 + ], + "type": "inline_equation", + "content": "p_{ijk}" + }, + { + "bbox": [ + 69, + 294, + 502, + 318 + ], + "type": "text", + "content": " indicates the coordinate for token " + }, + { + "bbox": [ + 69, + 294, + 502, + 318 + ], + "type": "inline_equation", + "content": "w_i" + }, + { + "bbox": [ + 69, + 294, + 502, + 318 + ], + "type": "text", + "content": " for " + }, + { + "bbox": [ + 69, + 294, + 502, + 318 + ], + "type": "inline_equation", + "content": "i" + }, + { + "bbox": [ + 69, + 294, + 502, + 318 + ], + "type": "text", + "content": "-th point for model tracking in frame " + }, + { + "bbox": [ + 69, + 294, + 502, + 318 + ], + "type": "inline_equation", + "content": "f_j" + }, + { + "bbox": [ + 69, + 294, + 502, + 318 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 69, + 318, + 502, + 350 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 69, + 318, + 502, + 350 + ], + "spans": [ + { + "bbox": [ + 69, + 318, + 502, + 350 + ], + "type": "text", + "content": "Overall, the leaning objective in stage 2 was calculated by standard label-smoothed cross-entropy loss to train the generation output, and " + }, + { + "bbox": [ + 69, + 318, + 502, + 350 + ], + "type": "inline_equation", + "content": "L_{1}" + }, + { + "bbox": [ + 69, + 318, + 502, + 350 + ], + "type": "text", + "content": " regression loss to train the trajectory output as the condition:" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 141, + 350, + 502, + 380 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 141, + 350, + 502, + 380 + ], + "spans": [ + { + "bbox": [ + 141, + 350, + 502, + 380 + ], + "type": "interline_equation", + "content": "\\mathcal {L} _ {2} = \\frac {1}{\\ell} \\sum_ {i = 1} ^ {\\ell} \\left(\\operatorname {C E} \\left(\\operatorname {L L M} \\left(\\mathbf {z}, \\mathbf {w} _ {1: i - 1}\\right), w _ {i}\\right) + \\frac {\\lambda}{P \\cdot N} \\sum_ {j = 1} ^ {P} \\sum_ {k = 1} ^ {N} | \\hat {p} _ {i j k} - p _ {i j k} |\\right), \\tag {8}", + "image_path": "a1aabb1f1e9510e00b7a38c7df36e578fc8ee22b342327d94d36cf6e0442bc57.jpg" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 69, + 385, + 502, + 421 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 69, + 385, + 502, + 421 + ], + "spans": [ + { + "bbox": [ + 69, + 385, + 502, + 421 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 69, + 385, + 502, + 421 + ], + "type": "inline_equation", + "content": "P" + }, + { + "bbox": [ + 69, + 385, + 502, + 421 + ], + "type": "text", + "content": " is the number of the points for model tracking to generate trajectory, and " + }, + { + "bbox": [ + 69, + 385, + 502, + 421 + ], + "type": "inline_equation", + "content": "\\mathbf{z} \\in \\mathbb{R}^{N \\times d}" + }, + { + "bbox": [ + 69, + 385, + 502, + 421 + ], + "type": "text", + "content": " represents the sequence of visual embedding. We merge the LoRA trained in the stage 1 with the original model and introduce a new LoRA module." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 69, + 422, + 503, + 499 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 69, + 422, + 503, + 499 + ], + "spans": [ + { + "bbox": [ + 69, + 422, + 503, + 499 + ], + "type": "text", + "content": "It is notable that we use localization projector " + }, + { + "bbox": [ + 69, + 422, + 503, + 499 + ], + "type": "inline_equation", + "content": "\\varphi(\\cdot)" + }, + { + "bbox": [ + 69, + 422, + 503, + 499 + ], + "type": "text", + "content": " trained in previous stage to initialize the trajectory projector " + }, + { + "bbox": [ + 69, + 422, + 503, + 499 + ], + "type": "inline_equation", + "content": "\\rho(\\cdot)" + }, + { + "bbox": [ + 69, + 422, + 503, + 499 + ], + "type": "text", + "content": ". Specifically, we define the weight of localization projector " + }, + { + "bbox": [ + 69, + 422, + 503, + 499 + ], + "type": "inline_equation", + "content": "\\varphi(\\cdot)" + }, + { + "bbox": [ + 69, + 422, + 503, + 499 + ], + "type": "text", + "content": " and trajectory projector " + }, + { + "bbox": [ + 69, + 422, + 503, + 499 + ], + "type": "inline_equation", + "content": "\\rho(\\cdot)" + }, + { + "bbox": [ + 69, + 422, + 503, + 499 + ], + "type": "text", + "content": " as " + }, + { + "bbox": [ + 69, + 422, + 503, + 499 + ], + "type": "inline_equation", + "content": "\\mathbf{m}_{\\varphi} \\in \\mathbb{R}^{P \\cdot N \\cdot 2 \\times d}" + }, + { + "bbox": [ + 69, + 422, + 503, + 499 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 69, + 422, + 503, + 499 + ], + "type": "inline_equation", + "content": "\\mathbf{m}_{\\rho} \\in \\mathbb{R}^{2 \\times d}" + }, + { + "bbox": [ + 69, + 422, + 503, + 499 + ], + "type": "text", + "content": ", respectively. localization projector " + }, + { + "bbox": [ + 69, + 422, + 503, + 499 + ], + "type": "inline_equation", + "content": "\\varphi(\\cdot)" + }, + { + "bbox": [ + 69, + 422, + 503, + 499 + ], + "type": "text", + "content": " maps a 2-dimension coordinate on the input image for each token of the LLM output, as for trajectory projector " + }, + { + "bbox": [ + 69, + 422, + 503, + 499 + ], + "type": "inline_equation", + "content": "\\rho(\\cdot)" + }, + { + "bbox": [ + 69, + 422, + 503, + 499 + ], + "type": "text", + "content": ", it also output 2-dimension coordinates, but more than " + }, + { + "bbox": [ + 69, + 422, + 503, + 499 + ], + "type": "inline_equation", + "content": "P \\cdot N" + }, + { + "bbox": [ + 69, + 422, + 503, + 499 + ], + "type": "text", + "content": " times for " + }, + { + "bbox": [ + 69, + 422, + 503, + 499 + ], + "type": "inline_equation", + "content": "P" + }, + { + "bbox": [ + 69, + 422, + 503, + 499 + ], + "type": "text", + "content": " points to tracking for " + }, + { + "bbox": [ + 69, + 422, + 503, + 499 + ], + "type": "inline_equation", + "content": "N" + }, + { + "bbox": [ + 69, + 422, + 503, + 499 + ], + "type": "text", + "content": " frames. For each point of each frame, the parameter of trajectory projector " + }, + { + "bbox": [ + 69, + 422, + 503, + 499 + ], + "type": "inline_equation", + "content": "\\rho(\\cdot)" + }, + { + "bbox": [ + 69, + 422, + 503, + 499 + ], + "type": "text", + "content": " initialized by the localization projector " + }, + { + "bbox": [ + 69, + 422, + 503, + 499 + ], + "type": "inline_equation", + "content": "\\varphi(\\cdot)" + }, + { + "bbox": [ + 69, + 422, + 503, + 499 + ], + "type": "text", + "content": ":" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 228, + 509, + 501, + 533 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 228, + 509, + 501, + 533 + ], + "spans": [ + { + "bbox": [ + 228, + 509, + 501, + 533 + ], + "type": "interline_equation", + "content": "\\mathbf {m} _ {\\varphi} = \\overbrace {\\mathbf {m} _ {\\rho} \\oplus \\mathbf {m} _ {\\rho} \\oplus \\cdots \\oplus \\mathbf {m} _ {\\rho}} ^ {P \\cdot N}, \\tag {9}", + "image_path": "9ce5aacd6e701b37f5186496c45e0d60dd38814fde3288959cd63eca087859f7.jpg" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 73, + 543, + 339, + 554 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 73, + 543, + 339, + 554 + ], + "spans": [ + { + "bbox": [ + 73, + 543, + 339, + 554 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 73, + 543, + 339, + 554 + ], + "type": "inline_equation", + "content": "\\oplus" + }, + { + "bbox": [ + 73, + 543, + 339, + 554 + ], + "type": "text", + "content": " denotes concatenation of matrix in first dimension." + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 69, + 555, + 501, + 616 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 69, + 555, + 501, + 616 + ], + "spans": [ + { + "bbox": [ + 69, + 555, + 501, + 616 + ], + "type": "text", + "content": "Beyond trajectories, our model is attuned to temporal boundaries within the generated text. Specifically, we structure the generation as ..., from s to e or From s to e, ... to facilitate the model to learn in temporal dimension. Here, ... encapsulates the event description, while s and e denote the frame indexes corresponding to the start and end timestamps of event, respectively. This approach further augments the model's understanding of temporal boundaries [13]." + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 69, + 616, + 502, + 653 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 69, + 616, + 502, + 653 + ], + "spans": [ + { + "bbox": [ + 69, + 616, + 502, + 653 + ], + "type": "text", + "content": "Dissimilar to the initial training stage, not all generated words are associated with trajectories. In cases where objects lack trajectories or vanish from view over time, we uniformly assign the coordinates of their ground truth as " + }, + { + "bbox": [ + 69, + 616, + 502, + 653 + ], + "type": "inline_equation", + "content": "(-1, -1)" + }, + { + "bbox": [ + 69, + 616, + 502, + 653 + ], + "type": "text", + "content": " to signify their absence." + } + ] + } + ], + "index": 14 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 198, + 70, + 470, + 83 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 198, + 70, + 470, + 83 + ], + "spans": [ + { + "bbox": [ + 198, + 70, + 470, + 83 + ], + "type": "text", + "content": "PiTe: Pixel-Temporal Alignment for Large Video-Language Model" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 495, + 71, + 503, + 79 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 495, + 71, + 503, + 79 + ], + "spans": [ + { + "bbox": [ + 495, + 71, + 503, + 79 + ], + "type": "text", + "content": "9" + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 8 + }, + { + "para_blocks": [ + { + "type": "table", + "bbox": [ + 111, + 114, + 536, + 258 + ], + "blocks": [ + { + "bbox": [ + 160, + 93, + 485, + 105 + ], + "lines": [ + { + "bbox": [ + 160, + 93, + 485, + 105 + ], + "spans": [ + { + "bbox": [ + 160, + 93, + 485, + 105 + ], + "type": "text", + "content": "Table 3: Comparison between different LVidLMs on zero-shot question-answer." + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 111, + 114, + 536, + 258 + ], + "lines": [ + { + "bbox": [ + 111, + 114, + 536, + 258 + ], + "spans": [ + { + "bbox": [ + 111, + 114, + 536, + 258 + ], + "type": "table", + "html": "
ModelLLM SizeMSVD-QA [41]MSRVTT-QA [43]ActivityNet-QA [45]
Accuracy↑Score↑Accuracy↑Score↑Accuracy↑Score↑
FrozenBiLM [44]1B32.2-16.8-24.7-
LLaMA-Adapter [48]7B54.93.143.82.734.22.7
VideoChat [19]7B56.32.845.02.5-2.2
Video-LLaMA [46]7B51.62.529.61.812.41.1
Video-ChatGPT [26]7B64.93.349.32.835.22.7
PG-Video-LLaVA [27]7B64.13.751.63.339.93.3
PiTe (Ours)7B68.43.956.43.542.03.3
PiTe (Ours)13B71.64.057.73.542.23.4
", + "image_path": "e423f237267028f84caeff315a6b160852c995baaa6ef36b1fc359d58526d454.jpg" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "table_body" + } + ], + "index": 3 + }, + { + "bbox": [ + 106, + 280, + 539, + 316 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 106, + 280, + 539, + 316 + ], + "spans": [ + { + "bbox": [ + 106, + 280, + 539, + 316 + ], + "type": "text", + "content": "Stage 3: Video Question Answering. Following stage 2, we incorporate high-quality dialogue data Valley [25] and Video-ChatGPT [26] in one turn for instruction tuning, enabling the model to follow human instructions for more accurate and generalize capabilities of video understanding." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 107, + 316, + 539, + 338 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 107, + 316, + 539, + 338 + ], + "spans": [ + { + "bbox": [ + 107, + 316, + 539, + 338 + ], + "type": "text", + "content": "The leaning objective in third stage was calculated by standard label-smoothed cross-entropy loss for auto-regression generation:" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 246, + 347, + 539, + 376 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 246, + 347, + 539, + 376 + ], + "spans": [ + { + "bbox": [ + 246, + 347, + 539, + 376 + ], + "type": "interline_equation", + "content": "\\mathcal {L} _ {3} = \\frac {1}{\\ell} \\sum_ {i = 1} ^ {\\ell} \\operatorname {C E} \\left(\\text {L L M} \\left(\\mathbf {z}, \\mathbf {w} _ {1: i - 1}\\right), w _ {i}\\right). \\tag {10}", + "image_path": "eb18912e2e7f32d4cb93297cbdbb0907745ef67a537abda747098ed6cde332b5.jpg" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 107, + 382, + 539, + 407 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 107, + 382, + 539, + 407 + ], + "spans": [ + { + "bbox": [ + 107, + 382, + 539, + 407 + ], + "type": "text", + "content": "Similar to stage 2, we merge the LoRA trained in the stage 1 and stage 2 with the original model and introduce a new LoRA module." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 107, + 426, + 205, + 440 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 107, + 426, + 205, + 440 + ], + "spans": [ + { + "bbox": [ + 107, + 426, + 205, + 440 + ], + "type": "text", + "content": "5 Experiments" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 107, + 452, + 237, + 464 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 107, + 452, + 237, + 464 + ], + "spans": [ + { + "bbox": [ + 107, + 452, + 237, + 464 + ], + "type": "text", + "content": "5.1 Experimental Setup" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 106, + 472, + 540, + 653 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 106, + 472, + 540, + 653 + ], + "spans": [ + { + "bbox": [ + 106, + 472, + 540, + 653 + ], + "type": "text", + "content": "Tasks, Datasets, and Evaluation Metrics. We conduct a quantitative evaluation of LVidLMs' video understanding capabilities across three tasks: (1) Video Question Answering: This task assesses the comprehensive video comprehension abilities of LVidLMs by requiring the model to answer a variety of questions about the video content based on its understanding. We perform this task on three datasets: MSVD-QA [41], MSRVTT-QA [43], and ActivityNet-QA [45]. The evaluation pipeline for video understanding follows Video-ChatGPT [26], and we report the accuracy and score, which is assessed using GPT-Assistant [28]. (2) Video Temporal Grounding: This task evaluates LVidLMs' capacity to discern the starting and ending timestamps of a segment corresponding to the description of a video clip. This task demands the model to effectively grasp the temporal aspects of the video. We conduct this task on the ActivityNet Captions dataset [15] and calculate Intersection over Union (IoU) between the model-generate time segments and the ground truth time segments. We report mean IoU (mIoU) and Recall@1, IoU " + }, + { + "bbox": [ + 106, + 472, + 540, + 653 + ], + "type": "inline_equation", + "content": "\\geq m" + }, + { + "bbox": [ + 106, + 472, + 540, + 653 + ], + "type": "text", + "content": " (R@m) metric, where " + }, + { + "bbox": [ + 106, + 472, + 540, + 653 + ], + "type": "inline_equation", + "content": "m" + }, + { + "bbox": [ + 106, + 472, + 540, + 653 + ], + "type": "text", + "content": " values are set at " + }, + { + "bbox": [ + 106, + 472, + 540, + 653 + ], + "type": "inline_equation", + "content": "\\{0.3, 0.5, 0.7\\}" + }, + { + "bbox": [ + 106, + 472, + 540, + 653 + ], + "type": "text", + "content": ". (3) Video Dense Captioning: This task requires the model to produce all events depict in the video along with their corresponding start and end timestamps. It necessitates the model to comprehend both the spatial and temporal dimensions of the video simultaneously." + } + ] + } + ], + "index": 10 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 109, + 71, + 120, + 79 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 109, + 71, + 120, + 79 + ], + "spans": [ + { + "bbox": [ + 109, + 71, + 120, + 79 + ], + "type": "text", + "content": "10" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 141, + 70, + 229, + 81 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 141, + 70, + 229, + 81 + ], + "spans": [ + { + "bbox": [ + 141, + 70, + 229, + 81 + ], + "type": "text", + "content": "Y. Liu, P. Ding et al." + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 9 + }, + { + "para_blocks": [ + { + "type": "table", + "bbox": [ + 74, + 125, + 501, + 235 + ], + "blocks": [ + { + "bbox": [ + 70, + 93, + 503, + 117 + ], + "lines": [ + { + "bbox": [ + 70, + 93, + 503, + 117 + ], + "spans": [ + { + "bbox": [ + 70, + 93, + 503, + 117 + ], + "type": "text", + "content": "Table 4: Comparison between different LVidLMs in temporal video grounding and dense video captioning tasks on ActivityNet [11]." + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 74, + 125, + 501, + 235 + ], + "lines": [ + { + "bbox": [ + 74, + 125, + 501, + 235 + ], + "spans": [ + { + "bbox": [ + 74, + 125, + 501, + 235 + ], + "type": "table", + "html": "
ModelLLM SizeTemporal GroundingDense Captioning
R@0.3↑R@0.5↑R@0.7↑mIoU↑SODA_c↑CIDEr↑METEOR↑
VideoChat [19]7B8.83.71.57.20.92.20.9
Video-LLaMA [46]7B6.92.10.86.51.95.81.9
Video-ChatGPT [26]7B26.413.66.118.91.95.82.1
PiTe (Ours)7B30.417.87.822.05.121.75.8
PiTe (Ours)13B37.223.710.926.05.926.56.6
", + "image_path": "56fcf122d532a45d1d2bf3f98b6624c49b21875e38d0ad1b44ea7f7ee130063b.jpg" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "table_body" + } + ], + "index": 3 + }, + { + "bbox": [ + 69, + 255, + 502, + 316 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 69, + 255, + 502, + 316 + ], + "spans": [ + { + "bbox": [ + 69, + 255, + 502, + 316 + ], + "type": "text", + "content": "We conduct this task on the ActivityNet Captions dataset [15]. Initially, we reported SODA_c [10], followed by averages of CIDEr [39] and METEOR [16] under different IoU thresholds of 0.3, 0.5, 0.7, 0.9 based on generate events and ground truth matched pairs to provide a comprehensive analysis. In this paper, all experiments were conducted in a zero-shot setting, and higher values of all evaluation metrics indicate superior performance." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 69, + 332, + 502, + 392 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 69, + 332, + 502, + 392 + ], + "spans": [ + { + "bbox": [ + 69, + 332, + 502, + 392 + ], + "type": "text", + "content": "Implementation Details. In this paper, we employ Vicuna v1.5 [5] as LLM to train the PiTe model at two scales: 7B and 13B. Leveraging the efficiency of LoRA [12], the training of the 7B model can be completed in approximately 10 hours using a single Nvidia 8-A100 (80GB VRAM) node, while the 13B model requires around 17 hours. More hyper-parameter settings are shwon in Table 2." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 70, + 408, + 162, + 420 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 70, + 408, + 162, + 420 + ], + "spans": [ + { + "bbox": [ + 70, + 408, + 162, + 420 + ], + "type": "text", + "content": "5.2 Main Result" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 69, + 427, + 501, + 452 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 69, + 427, + 501, + 452 + ], + "spans": [ + { + "bbox": [ + 69, + 427, + 501, + 452 + ], + "type": "text", + "content": "Table 3 and 4 present the comparative performance of the PiTe against state-of-the-art baselines on myriad video understanding datasets." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 69, + 468, + 502, + 552 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 69, + 468, + 502, + 552 + ], + "spans": [ + { + "bbox": [ + 69, + 468, + 502, + 552 + ], + "type": "text", + "content": "Question Answering. As illustrated in Table 3, PiTe consistently outperforms the state-of-the-art pure instruction-tuning baselines in terms of all metrics on all datasets. Compared to the top-performing baselines in each dataset, PiTe exhibited notable improvements in the average question answering accuracy, achieving a maximum enhancement of 4.8 and an average improvement of 3.7. For example, PiTe substantially improves accuracy by 64.9 to 68.4 compared to Video-ChatGPT [26] in MSVD-QA dataset [41]. The results showcasing PiTe's proficiency in video comprehension and its capacity to deliver contextually relevant responses according to the given instructions." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 69, + 568, + 503, + 652 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 69, + 568, + 503, + 652 + ], + "spans": [ + { + "bbox": [ + 69, + 568, + 503, + 652 + ], + "type": "text", + "content": "Temporal Grounding. As depicted in Table 4, PiTe achieves state-of-the-art performance in the video temporal grounding task across all metrics as well, demonstrating improvements ranging from 18.9 to 22.0 in mIoU compared to Video-ChatGPT [26]. This clearly indicates that trajectory alignment greatly enhances the ability of capture events in temporal dimension for LVidLMs. The incorporation of object trajectories in the temporal dimension of the trajectory matrix equips the model with a precise understanding of temporal event boundaries, thereby establishing a solid foundation for accurate event localization." + } + ] + } + ], + "index": 9 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 199, + 70, + 470, + 82 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 199, + 70, + 470, + 82 + ], + "spans": [ + { + "bbox": [ + 199, + 70, + 470, + 82 + ], + "type": "text", + "content": "PiTe: Pixel-Temporal Alignment for Large Video-Language Model" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 491, + 72, + 501, + 80 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 491, + 72, + 501, + 80 + ], + "spans": [ + { + "bbox": [ + 491, + 72, + 501, + 80 + ], + "type": "text", + "content": "11" + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 10 + }, + { + "para_blocks": [ + { + "type": "table", + "bbox": [ + 111, + 114, + 536, + 220 + ], + "blocks": [ + { + "bbox": [ + 198, + 94, + 447, + 105 + ], + "lines": [ + { + "bbox": [ + 198, + 94, + 447, + 105 + ], + "spans": [ + { + "bbox": [ + 198, + 94, + 447, + 105 + ], + "type": "text", + "content": "Table 5: Ablation study of the three-stage training strategy." + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 111, + 114, + 536, + 220 + ], + "lines": [ + { + "bbox": [ + 111, + 114, + 536, + 220 + ], + "spans": [ + { + "bbox": [ + 111, + 114, + 536, + 220 + ], + "type": "table", + "html": "
MethodMSVD-QAActivityNet
Temporal GroundingDense Captioning
Accuracy ScoreR@0.3R@0.5R@0.7mIoUSODA_cCIDErMETEOR
PiTe (Ours)68.43.930.417.87.822.05.121.75.8
w/o initialize68.23.922.810.54.617.15.121.75.8
w/o trajectory68.13.923.912.85.717.45.021.45.8
", + "image_path": "1e48b5bf55078838c965fa8968d872a97b61d88744ea22bc63528999e523e4b4.jpg" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "table_body" + } + ], + "index": 3 + }, + { + "bbox": [ + 106, + 243, + 539, + 316 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 106, + 243, + 539, + 316 + ], + "spans": [ + { + "bbox": [ + 106, + 243, + 539, + 316 + ], + "type": "text", + "content": "Dense Captioning. The outcomes of the dense captioning task, as delineated in Table 4, reveal that PiTe consistent boost compared to all state-of-the-art baselines. Particularly noteworthy is the substantial 15.9 increase in the CIDEr metric [39] when compared to Video-ChatGPT [26]. This underscores the significance of fine-grained alignment in both spatial and temporal dimensions through trajectories, implying that PiTe acquires more generalized and detailed representations to offer more sophisticated event descriptions and accurate event temporal boundaries." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 107, + 335, + 179, + 346 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 107, + 335, + 179, + 346 + ], + "spans": [ + { + "bbox": [ + 107, + 335, + 179, + 346 + ], + "type": "text", + "content": "5.3 Analysis" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 106, + 356, + 539, + 416 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 106, + 356, + 539, + 416 + ], + "spans": [ + { + "bbox": [ + 106, + 356, + 539, + 416 + ], + "type": "text", + "content": "Ablation Study. As reported in Table 5, we conduct ablation experiments on MVSD-QA [41] for question answering and ActivityNet Captions [15] for temporal grounding to verify the individual effects of the proposed contributions under the following settings: (1) w/o initialize: we remove the initialization strategy that use weight of localization projector to initialize trajectory projector; (2) w/o trajectory: we abandon the fine-grained alignment strategy via trajectory." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 106, + 417, + 539, + 548 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 106, + 417, + 539, + 548 + ], + "spans": [ + { + "bbox": [ + 106, + 417, + 539, + 548 + ], + "type": "text", + "content": "From the experimental results in Table 5, it can be observed that: (1) Eliminating the initialization strategy for the trajectory projector in PiTe reduces the model's reasoning capabilities and temporal boundary awareness. However, the performance in dense captioning generation remains consistent. This observation suggests that the model maintains its basic ability in comprehending visual content under trajectory-guided training. (2) The removal of the trajectory-guided training diminishes almost all the capability of PiTe, including dense captioning. (3) Without trajectory-guided training, PiTe demonstrates superior performance compared to trajectory-guided training without the initialization strategy for the trajectory projector in temporal grounding. This outcome highlights the difficulty of trajectory-guided training without initialization from a pre-trained localization projector, as the instability of parameters can impede the model's perception to accurately perceive visual temporal information." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 106, + 567, + 539, + 652 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 106, + 567, + 539, + 652 + ], + "spans": [ + { + "bbox": [ + 106, + 567, + 539, + 652 + ], + "type": "text", + "content": "Exhibition. To better illustrate the video dialogue performance of PiTe, we present a qualitative example, as shown in Fig. 5a. The illustration from the upper portion of the figure demonstrates PiTe's capability not only to provide precise responses to instruction queries but also to enhance the output with more detailed and accurate video information. The example in the lower segment of the figure highlights the model's proficiency in understanding instruction and capturing event, enabling precise delineation of temporal boundaries within the video, despite the constraint of a 100-frame sampling limit." + } + ] + } + ], + "index": 8 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 108, + 71, + 119, + 79 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 108, + 71, + 119, + 79 + ], + "spans": [ + { + "bbox": [ + 108, + 71, + 119, + 79 + ], + "type": "text", + "content": "12" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 141, + 70, + 229, + 81 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 141, + 70, + 229, + 81 + ], + "spans": [ + { + "bbox": [ + 141, + 70, + 229, + 81 + ], + "type": "text", + "content": "Y. Liu, P. Ding et al." + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 11 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 82, + 106, + 261, + 149 + ], + "blocks": [ + { + "bbox": [ + 82, + 106, + 261, + 149 + ], + "lines": [ + { + "bbox": [ + 82, + 106, + 261, + 149 + ], + "spans": [ + { + "bbox": [ + 82, + 106, + 261, + 149 + ], + "type": "image", + "image_path": "e76a39fe0b891717df0d44f7843fa486ca7779662b9d3a4aff474b487a9ebcbd.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 83, + 149, + 255, + 158 + ], + "lines": [ + { + "bbox": [ + 83, + 149, + 255, + 158 + ], + "spans": [ + { + "bbox": [ + 83, + 149, + 255, + 158 + ], + "type": "text", + "content": "[Instruction] Who are shown performing impressive feats?" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_caption" + } + ], + "index": 2 + }, + { + "type": "image", + "bbox": [ + 83, + 191, + 261, + 217 + ], + "blocks": [ + { + "bbox": [ + 83, + 163, + 245, + 187 + ], + "lines": [ + { + "bbox": [ + 83, + 163, + 245, + 187 + ], + "spans": [ + { + "bbox": [ + 83, + 163, + 245, + 187 + ], + "type": "text", + "content": "[Response] The video showcases various football players performing impressive feats, including a player jumping over another player and a player making a spectacular catch." + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 83, + 191, + 261, + 217 + ], + "lines": [ + { + "bbox": [ + 83, + 191, + 261, + 217 + ], + "spans": [ + { + "bbox": [ + 83, + 191, + 261, + 217 + ], + "type": "image", + "image_path": "b9ee5ca2d267772e80f9143a833f9eb65a4cf5c5520d1a49cb8e1587f2a53ab0.jpg" + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 83, + 217, + 262, + 232 + ], + "lines": [ + { + "bbox": [ + 83, + 217, + 262, + 232 + ], + "spans": [ + { + "bbox": [ + 83, + 217, + 262, + 232 + ], + "type": "text", + "content": "[Instruction] During which frames can we see he then shines the shoes of one man who pays him and shakes his hand." + } + ] + } + ], + "index": 6, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 83, + 234, + 252, + 242 + ], + "lines": [ + { + "bbox": [ + 83, + 234, + 252, + 242 + ], + "spans": [ + { + "bbox": [ + 83, + 234, + 252, + 242 + ], + "type": "text", + "content": "[Response] The shoe shining and payment occur from 66 to 99" + } + ] + } + ], + "index": 7, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 83, + 242, + 231, + 250 + ], + "lines": [ + { + "bbox": [ + 83, + 242, + 231, + 250 + ], + "spans": [ + { + "bbox": [ + 83, + 242, + 231, + 250 + ], + "type": "text", + "content": "(116.348s ~ 174.522s). Ground truth: 116.06s ~ 172.78s" + } + ] + } + ], + "index": 8, + "angle": 0, + "type": "image_caption" + } + ], + "index": 5 + }, + { + "type": "image", + "bbox": [ + 286, + 98, + 500, + 249 + ], + "blocks": [ + { + "bbox": [ + 71, + 252, + 266, + 262 + ], + "lines": [ + { + "bbox": [ + 71, + 252, + 266, + 262 + ], + "spans": [ + { + "bbox": [ + 71, + 252, + 266, + 262 + ], + "type": "text", + "content": "(a) Examples of PiTe's video understanding capabilities." + } + ] + } + ], + "index": 9, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 286, + 98, + 500, + 249 + ], + "lines": [ + { + "bbox": [ + 286, + 98, + 500, + 249 + ], + "spans": [ + { + "bbox": [ + 286, + 98, + 500, + 249 + ], + "type": "image", + "image_path": "4c6152d1ab3082d20f1d939cae6b4688ccc0a009afb2db382e985e60a717d404.jpg" + } + ] + } + ], + "index": 10, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 277, + 252, + 501, + 262 + ], + "lines": [ + { + "bbox": [ + 277, + 252, + 501, + 262 + ], + "spans": [ + { + "bbox": [ + 277, + 252, + 501, + 262 + ], + "type": "text", + "content": "(b) Performance comparison for different tracking point quantity." + } + ] + } + ], + "index": 11, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 69, + 271, + 502, + 293 + ], + "lines": [ + { + "bbox": [ + 69, + 271, + 502, + 293 + ], + "spans": [ + { + "bbox": [ + 69, + 271, + 502, + 293 + ], + "type": "text", + "content": "Fig. 5: PiTe's video understanding capabilities and performance comparison across varying tracking point quantities." + } + ] + } + ], + "index": 12, + "angle": 0, + "type": "image_caption" + } + ], + "index": 10 + }, + { + "bbox": [ + 69, + 309, + 503, + 417 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 69, + 309, + 503, + 417 + ], + "spans": [ + { + "bbox": [ + 69, + 309, + 503, + 417 + ], + "type": "text", + "content": "Impact of Tracking Point Quantity. In Fig. 5b, we vary the tracking point quantity " + }, + { + "bbox": [ + 69, + 309, + 503, + 417 + ], + "type": "inline_equation", + "content": "P" + }, + { + "bbox": [ + 69, + 309, + 503, + 417 + ], + "type": "text", + "content": " in set of " + }, + { + "bbox": [ + 69, + 309, + 503, + 417 + ], + "type": "inline_equation", + "content": "\\{1,3,5\\}" + }, + { + "bbox": [ + 69, + 309, + 503, + 417 + ], + "type": "text", + "content": ". The efficacy of dense captioning tasks tends to improve with an increase in tracking points. However, it is observed that the temporal grounding task undergoes an initial substantial improvement, only to be succeeded by a rapid decline. Less number of tracking points fails to accurately capture the object's geometry, thereby hindering the pixel-level cross-modal alignment guidance for the model. Conversely, a higher quantity of points can enhance the model's comprehension of pure visual information; however, it also introduces noise to make training more challenging. Overall, that the optimal value of " + }, + { + "bbox": [ + 69, + 309, + 503, + 417 + ], + "type": "inline_equation", + "content": "P" + }, + { + "bbox": [ + 69, + 309, + 503, + 417 + ], + "type": "text", + "content": " may be different for different tasks, we set " + }, + { + "bbox": [ + 69, + 309, + 503, + 417 + ], + "type": "inline_equation", + "content": "P = 3" + }, + { + "bbox": [ + 69, + 309, + 503, + 417 + ], + "type": "text", + "content": " due to its performance maintain stability over multiple tasks." + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 70, + 433, + 159, + 445 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 70, + 433, + 159, + 445 + ], + "spans": [ + { + "bbox": [ + 70, + 433, + 159, + 445 + ], + "type": "text", + "content": "6 Conclusion" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 69, + 456, + 503, + 589 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 69, + 456, + 503, + 589 + ], + "spans": [ + { + "bbox": [ + 69, + 456, + 503, + 589 + ], + "type": "text", + "content": "In this paper, we focus on enhancing the performance of Large Video-Language Models (LVidLMs) by incorporating trajectory-based alignment across different modalities. To achieve fine-grained alignment between video and language across spatial and temporal dimensions, we initially curate a comprehensive multi-modal object tracking dataset, PiTe-143k, using a fully automated annotation pipeline. This dataset was developed to address the lack of large-scale video-language datasets that include multi-object moving trajectories. Subsequently, we introduce a novel Pixel-Temporal (PiTe) alignment strategy that leverages trajectory-guided pre-training to address the inherent challenges faced by LVidLMs. Through comparative analyses, we evaluate PiTe against state-of-the-art models and competitive baselines across various tasks in a zero-shot setting, including question-answering, temporal grounding, and dense captioning, showcasing the superior performance of PiTe with more sophisticated event descriptions and accurate event temporal boundaries." + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 70, + 605, + 179, + 618 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 70, + 605, + 179, + 618 + ], + "spans": [ + { + "bbox": [ + 70, + 605, + 179, + 618 + ], + "type": "text", + "content": "Acknowledgments" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 70, + 628, + 502, + 652 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 70, + 628, + 502, + 652 + ], + "spans": [ + { + "bbox": [ + 70, + 628, + 502, + 652 + ], + "type": "text", + "content": "This work was supported by the National Science and Technology Innovation 2030 - Major Project (Grant No. 2022ZD0208800), and NSFC General Program (Grant No. 62176215)." + } + ] + } + ], + "index": 17 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 199, + 70, + 470, + 82 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 199, + 70, + 470, + 82 + ], + "spans": [ + { + "bbox": [ + 199, + 70, + 470, + 82 + ], + "type": "text", + "content": "PiTe: Pixel-Temporal Alignment for Large Video-Language Model" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 491, + 71, + 503, + 80 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 491, + 71, + 503, + 80 + ], + "spans": [ + { + "bbox": [ + 491, + 71, + 503, + 80 + ], + "type": "text", + "content": "13" + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 12 + }, + { + "para_blocks": [ + { + "bbox": [ + 108, + 94, + 173, + 106 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 108, + 94, + 173, + 106 + ], + "spans": [ + { + "bbox": [ + 108, + 94, + 173, + 106 + ], + "type": "text", + "content": "References" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 108, + 118, + 539, + 651 + ], + "type": "list", + "angle": 0, + "index": 23, + "blocks": [ + { + "bbox": [ + 112, + 118, + 539, + 172 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 112, + 118, + 539, + 172 + ], + "spans": [ + { + "bbox": [ + 112, + 118, + 539, + 172 + ], + "type": "text", + "content": "1. Alayrac, J., Donahue, J., Luc, P., Miech, A., Barr, I., Hasson, Y., Lenc, K., Mensch, A., Millican, K., Reynolds, M., Ring, R., Rutherford, E., Cabi, S., Han, T., Gong, Z., Samangooei, S., Monteiro, M., Menick, J.L., Borgeaud, S., Brock, A., Nematzadeh, A., Sharifzadeh, S., Binkowski, M., Barreira, R., Vinyals, O., Zisserman, A., Simonyan, K.: Flamingo: a visual language model for few-shot learning. In: Proc. of NeurIPS (2022)" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 113, + 173, + 538, + 194 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 113, + 173, + 538, + 194 + ], + "spans": [ + { + "bbox": [ + 113, + 173, + 538, + 194 + ], + "type": "text", + "content": "2. Arthur, D., Vassilvitskii, S.: k-means++: the advantages of careful seeding. In: Proc. of SODA. pp. 1027-1035 (2007)" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 113, + 194, + 539, + 227 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 113, + 194, + 539, + 227 + ], + "spans": [ + { + "bbox": [ + 113, + 194, + 539, + 227 + ], + "type": "text", + "content": "3. Awadalla, A., Gao, I., Gardner, J., Hessel, J., Hanafy, Y., Zhu, W., Marathe, K., Bitton, Y., Gadre, S.Y., Sagawa, S., Jitsev, J., Kornblith, S., Koh, P.W., Ilharco, G., Wortman, M., Schmidt, L.: Openflamingo: An open-source framework for training large autoregressive vision-language models. CoRR (2023)" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 112, + 228, + 539, + 281 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 112, + 228, + 539, + 281 + ], + "spans": [ + { + "bbox": [ + 112, + 228, + 539, + 281 + ], + "type": "text", + "content": "4. Brown, T.B., Mann, B., Ryder, N., Subbiah, M., Kaplan, J., Dhariwal, P., Neelakantan, A., Shyam, P., Sastry, G., Askell, A., Agarwal, S., Herbert-Voss, A., Krueger, G., Henighan, T., Child, R., Ramesh, A., Ziegler, D.M., Wu, J., Winter, C., Hesse, C., Chen, M., Sigler, E., Litwin, M., Gray, S., Chess, B., Clark, J., Berner, C., McCandlish, S., Radford, A., Sutskever, I., Amodei, D.: Language models are few-shot learners. In: Proc. of NeurIPS (2020)" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 113, + 282, + 539, + 313 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 113, + 282, + 539, + 313 + ], + "spans": [ + { + "bbox": [ + 113, + 282, + 539, + 313 + ], + "type": "text", + "content": "5. Chiang, W.L., Li, Z., Lin, Z., Sheng, Y., Wu, Z., Zhang, H., Zheng, L., Zhuang, S., Zhuang, Y., Gonzalez, J.E., Stoica, I., Xing, E.P.: Vicuna: An open-source chatbot impressing gpt-4 with " + }, + { + "bbox": [ + 113, + 282, + 539, + 313 + ], + "type": "inline_equation", + "content": "90\\%" + }, + { + "bbox": [ + 113, + 282, + 539, + 313 + ], + "type": "text", + "content": " * chatgpt quality (2023)" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 113, + 315, + 539, + 335 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 113, + 315, + 539, + 335 + ], + "spans": [ + { + "bbox": [ + 113, + 315, + 539, + 335 + ], + "type": "text", + "content": "6. Dai, W., Li, J., Li, D., Tiong, A.M.H., Zhao, J., Wang, W., Li, B., Fung, P., Hoi, S.C.H.: Instructblip: Towards general-purpose vision-language models with instruction tuning. In: Proc. of NeurIPS (2023)" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 113, + 336, + 539, + 357 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 113, + 336, + 539, + 357 + ], + "spans": [ + { + "bbox": [ + 113, + 336, + 539, + 357 + ], + "type": "text", + "content": "7. Devlin, J., Chang, M., Lee, K., Toutanova, K.: BERT: pre-training of deep bidirectional transformers for language understanding. In: Proc. of AACL. pp. 4171-4186 (2019)" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 113, + 358, + 539, + 390 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 113, + 358, + 539, + 390 + ], + "spans": [ + { + "bbox": [ + 113, + 358, + 539, + 390 + ], + "type": "text", + "content": "8. Dosovitskiy, A., Beyer, L., Kolesnikov, A., Weissenborn, D., Zhai, X., Unterthiner, T., Dehghani, M., Minderer, M., Heigold, G., Gelly, S., Uszkoreit, J., Houlsby, N.: An image is worth 16x16 words: Transformers for image recognition at scale. In: Proc. of ICLR (2021)" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 113, + 391, + 539, + 413 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 113, + 391, + 539, + 413 + ], + "spans": [ + { + "bbox": [ + 113, + 391, + 539, + 413 + ], + "type": "text", + "content": "9. Du, Z., Qian, Y., Liu, X., Ding, M., Qiu, J., Yang, Z., Tang, J.: GLM: general language model pretraining with autoregressive blank infilling. In: Proc. of ACL. pp. 320-335 (2022)" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 110, + 413, + 539, + 434 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 110, + 413, + 539, + 434 + ], + "spans": [ + { + "bbox": [ + 110, + 413, + 539, + 434 + ], + "type": "text", + "content": "10. Fujita, S., Hirao, T., Kamigaito, H., Okumura, M., Nagata, M.: SODA: story oriented dense video captioning evaluation framework. In: Proc. of ECCV. pp. 517-531 (2020)" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 109, + 434, + 539, + 456 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 109, + 434, + 539, + 456 + ], + "spans": [ + { + "bbox": [ + 109, + 434, + 539, + 456 + ], + "type": "text", + "content": "11. Heilbron, F.C., Escorcia, V., Ghanem, B., Niebles, J.C.: Activitynet: A large-scale video benchmark for human activity understanding. In: Proc. of CVPR. pp. 961-970 (2015)" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 109, + 456, + 539, + 477 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 109, + 456, + 539, + 477 + ], + "spans": [ + { + "bbox": [ + 109, + 456, + 539, + 477 + ], + "type": "text", + "content": "12. Hu, E.J., Shen, Y., Wallis, P., Allen-Zhu, Z., Li, Y., Wang, S., Wang, L., Chen, W.: Lora: Low-rank adaptation of large language models. In: Proc. of ICLR (2022)" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 109, + 478, + 539, + 499 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 109, + 478, + 539, + 499 + ], + "spans": [ + { + "bbox": [ + 109, + 478, + 539, + 499 + ], + "type": "text", + "content": "13. Huang, B., Wang, X., Chen, H., Song, Z., Zhu, W.: Vtimellm: Empower llm to grasp video moments. In: Proc. of CVPR. pp. 14271-14280 (2024)" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 109, + 500, + 539, + 521 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 109, + 500, + 539, + 521 + ], + "spans": [ + { + "bbox": [ + 109, + 500, + 539, + 521 + ], + "type": "text", + "content": "14. Kaplan, J., McCandlish, S., Henighan, T., Brown, T.B., Chess, B., Child, R., Gray, S., Radford, A., Wu, J., Amodei, D.: Scaling laws for neural language models. CoRR (2020)" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 109, + 521, + 539, + 542 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 109, + 521, + 539, + 542 + ], + "spans": [ + { + "bbox": [ + 109, + 521, + 539, + 542 + ], + "type": "text", + "content": "15. Krishna, R., Hata, K., Ren, F., Fei-Fei, L., Niebles, J.C.: Dense-c captioning events in videos. In: Proc. of ICCV. pp. 706-715 (2017)" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 109, + 544, + 539, + 564 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 109, + 544, + 539, + 564 + ], + "spans": [ + { + "bbox": [ + 109, + 544, + 539, + 564 + ], + "type": "text", + "content": "16. Lavie, A., Agarwal, A.: METEOR: an automatic metric for MT evaluation with high levels of correlation with human judgments. In: Proc. of WMT@ACL. pp. 228-231 (2007)" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 109, + 565, + 539, + 586 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 109, + 565, + 539, + 586 + ], + "spans": [ + { + "bbox": [ + 109, + 565, + 539, + 586 + ], + "type": "text", + "content": "17. Le Moing, G., Ponce, J., Schmid, C.: Dense optical tracking: Connecting the dots. In: Proc. of CVPR. pp. 19187-19197 (2024)" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 109, + 586, + 539, + 608 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 109, + 586, + 539, + 608 + ], + "spans": [ + { + "bbox": [ + 109, + 586, + 539, + 608 + ], + "type": "text", + "content": "18. Li, J., Li, D., Savarese, S., Hoi, S.C.H.: BLIP-2: bootstrapping language-image pre-training with frozen image encoders and large language models. In: Proc. of ICML. pp. 19730-19742 (2023)" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 109, + 609, + 539, + 629 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 109, + 609, + 539, + 629 + ], + "spans": [ + { + "bbox": [ + 109, + 609, + 539, + 629 + ], + "type": "text", + "content": "19. Li, K., He, Y., Wang, Y., Li, Y., Wang, W., Luo, P., Wang, Y., Wang, L., Qiao, Y.: Videochat: Chatcentric video understanding. CoRR (2023)" + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 108, + 630, + 539, + 651 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 108, + 630, + 539, + 651 + ], + "spans": [ + { + "bbox": [ + 108, + 630, + 539, + 651 + ], + "type": "text", + "content": "20. Lin, J., Yin, H., Ping, W., Molchanov, P., Shoeybi, M., Han, S.: Vila: On pre-training for visual language models. In: Proc. of CVPR. pp. 26689-26699 (2024)" + } + ] + } + ], + "index": 22 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 108, + 71, + 120, + 79 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 108, + 71, + 120, + 79 + ], + "spans": [ + { + "bbox": [ + 108, + 71, + 120, + 79 + ], + "type": "text", + "content": "14" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 141, + 70, + 229, + 81 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 141, + 70, + 229, + 81 + ], + "spans": [ + { + "bbox": [ + 141, + 70, + 229, + 81 + ], + "type": "text", + "content": "Y. Liu, P. Ding et al." + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 13 + }, + { + "para_blocks": [ + { + "bbox": [ + 71, + 95, + 502, + 652 + ], + "type": "list", + "angle": 0, + "index": 20, + "blocks": [ + { + "bbox": [ + 71, + 95, + 501, + 118 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 71, + 95, + 501, + 118 + ], + "spans": [ + { + "bbox": [ + 71, + 95, + 501, + 118 + ], + "type": "text", + "content": "21. Liu, H., Li, C., Li, Y., Lee, Y.J.: Improved baselines with visual instruction tuning. In: Proc. of CVPR. pp. 26296-26306 (2024)" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 72, + 118, + 449, + 129 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 72, + 118, + 449, + 129 + ], + "spans": [ + { + "bbox": [ + 72, + 118, + 449, + 129 + ], + "type": "text", + "content": "22. Liu, H., Li, C., Wu, Q., Lee, Y.J.: Visual instruction tuning. In: Proc. of NeurIPS (2023)" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 72, + 129, + 502, + 150 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 72, + 129, + 502, + 150 + ], + "spans": [ + { + "bbox": [ + 72, + 129, + 502, + 150 + ], + "type": "text", + "content": "23. Liu, Y., Ping, H., Zhang, D., Sun, Q., Li, S., Zhou, G.: Comment-aware multi-modal heterogeneous pre-training for humor detection in short-form videos. In: Proc. of ECAI. pp. 1568-1575 (2023)" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 71, + 150, + 502, + 172 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 71, + 150, + 502, + 172 + ], + "spans": [ + { + "bbox": [ + 71, + 150, + 502, + 172 + ], + "type": "text", + "content": "24. Liu, Y., Shen, T., Zhang, D., Sun, Q., Li, S., Zhou, G.: Comment-aided video-language alignment via contrastive pre-training for short-form video humor detection. In: Proc. of ICMR. pp. 442-450 (2024)" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 71, + 172, + 502, + 194 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 71, + 172, + 502, + 194 + ], + "spans": [ + { + "bbox": [ + 71, + 172, + 502, + 194 + ], + "type": "text", + "content": "25. Luo, R., Zhao, Z., Yang, M., Dong, J., Qiu, M., Lu, P., Wang, T., Wei, Z.: Valley: Video assistant with large language model enhanced ability. CoRR (2023)" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 71, + 194, + 502, + 216 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 71, + 194, + 502, + 216 + ], + "spans": [ + { + "bbox": [ + 71, + 194, + 502, + 216 + ], + "type": "text", + "content": "26. Maaz, M., Rasheed, H.A., Khan, S.H., Khan, F.S.: Video-chatgpt: Towards detailed video understanding via large vision and language models (2024)" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 71, + 216, + 502, + 237 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 71, + 216, + 502, + 237 + ], + "spans": [ + { + "bbox": [ + 71, + 216, + 502, + 237 + ], + "type": "text", + "content": "27. Munasinghe, S., Thushara, R., Maaz, M., Rasheed, H.A., Khan, S., Shah, M., Khan, F.S.: Pg-video-llava: Pixel grounding large video-language models. CoRR (2023)" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 71, + 237, + 361, + 248 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 71, + 237, + 361, + 248 + ], + "spans": [ + { + "bbox": [ + 71, + 237, + 361, + 248 + ], + "type": "text", + "content": "28. OpenAI: Chatgpt: Optimizing language models for dialogue (2023)" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 71, + 248, + 502, + 292 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 71, + 248, + 502, + 292 + ], + "spans": [ + { + "bbox": [ + 71, + 248, + 502, + 292 + ], + "type": "text", + "content": "29. Ouyang, L., Wu, J., Jiang, X., Almeida, D., Wainwright, C.L., Mishkin, P., Zhang, C., Agarwal, S., Slama, K., Ray, A., Schulman, J., Hilton, J., Kelton, F., Miller, L., Simens, M., Askell, A., Welinder, P., Christiano, P.F., Leike, J., Lowe, R.: Training language models to follow instructions with human feedback. In: Proc. of NeurIPS (2022)" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 71, + 292, + 502, + 314 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 71, + 292, + 502, + 314 + ], + "spans": [ + { + "bbox": [ + 71, + 292, + 502, + 314 + ], + "type": "text", + "content": "30. Pont-Tuset, J., Uijlings, J.R.R., Changpinyo, S., Soricut, R., Ferrari, V.: Connecting vision and language with localized narratives. In: Proc. of ECCV. pp. 647-664 (2020)" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 71, + 314, + 502, + 346 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 71, + 314, + 502, + 346 + ], + "spans": [ + { + "bbox": [ + 71, + 314, + 502, + 346 + ], + "type": "text", + "content": "31. Radford, A., Kim, J.W., Hallacy, C., Ramesh, A., Goh, G., Agarwal, S., Sastry, G., Askell, A., Mishkin, P., Clark, J., Krueger, G., Sutskever, I.: Learning transferable visual models from natural language supervision. In: Proc. of ICML. pp. 8748-8763 (2021)" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 71, + 346, + 490, + 357 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 71, + 346, + 490, + 357 + ], + "spans": [ + { + "bbox": [ + 71, + 346, + 490, + 357 + ], + "type": "text", + "content": "32. Radford, A., Narasimhan, K.: Improving language understanding by generative pre-training (2018)" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 71, + 357, + 502, + 378 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 71, + 357, + 502, + 378 + ], + "spans": [ + { + "bbox": [ + 71, + 357, + 502, + 378 + ], + "type": "text", + "content": "33. Radford, A., Wu, J., Child, R., Luan, D., Amodei, D., Sutskever, I.: Language models are unsupervised multitask learners (2019)" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 71, + 379, + 502, + 411 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 71, + 379, + 502, + 411 + ], + "spans": [ + { + "bbox": [ + 71, + 379, + 502, + 411 + ], + "type": "text", + "content": "34. Raffel, C., Shazeer, N., Roberts, A., Lee, K., Narang, S., Matena, M., Zhou, Y., Li, W., Liu, P.J.: Exploring the limits of transfer learning with a unified text-to-text transformer. J. Mach. Learn. Res. pp. 140:1-140:67 (2020)" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 71, + 411, + 502, + 443 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 71, + 411, + 502, + 443 + ], + "spans": [ + { + "bbox": [ + 71, + 411, + 502, + 443 + ], + "type": "text", + "content": "35. Rasheed, H., Maaz, M., Shaji, S., Shaker, A., Khan, S., Cholakkal, H., Anwer, R.M., Xing, E., Yang, M.H., Khan, F.S.: Glamm: Pixel grounding large multimodal model. In: Proc. of CVPR. pp. 13009-13018 (2024)" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 71, + 444, + 502, + 521 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 71, + 444, + 502, + 521 + ], + "spans": [ + { + "bbox": [ + 71, + 444, + 502, + 521 + ], + "type": "text", + "content": "36. Scao, T.L., Fan, A., Akiki, C., Pavlick, E., Ilic, S., Hesslow, D., Castagné, R., Luccioni, A.S., Yvon, F., Gallé, M., Tow, J., Rush, A.M., Biderman, S., Webson, A., Ammanamanchi, P.S., Wang, T., Sagot, B., Muennighoff, N., del Moral, A.V., Ruwase, O., Bawden, R., Bekman, S., McMillan-Major, A., Beltagy, I., Nguyen, H., Saulnier, L., Tan, S., Suarez, P.O., Sanh, V., Laurençon, H., Jernite, Y., Launay, J., Mitchell, M., Raffel, C., Gokaslan, A., Simhi, A., Soroa, A., Aji, A.F., Alfassy, A., Rogers, A., Nitzav, A.K., Xu, C., Mou, C., Emezue, C., Klamm, C., Leong, C., van Strien, D., Adelani, D.I., et al.: BLOOM: A 176b-parameter open-access multilingual language model. CoRR (2022)" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 71, + 521, + 502, + 553 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 71, + 521, + 502, + 553 + ], + "spans": [ + { + "bbox": [ + 71, + 521, + 502, + 553 + ], + "type": "text", + "content": "37. Touvron, H., Lavril, T., Izacard, G., Martinet, X., Lachaux, M., Lacroix, T., Rozière, B., Goyal, N., Hambro, E., Azhar, F., Rodriguez, A., Joulin, A., Grave, E., Lample, G.: Llama: Open and efficient foundation language models. CoRR (2023)" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 71, + 553, + 502, + 652 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 71, + 553, + 502, + 652 + ], + "spans": [ + { + "bbox": [ + 71, + 553, + 502, + 652 + ], + "type": "text", + "content": "38. Touvron, H., Martin, L., Stone, K., Albert, P., Almahairi, A., Babaei, Y., Bashlykov, N., Batra, S., Bhargava, P., Bhosale, S., Bikel, D., Blecher, L., Canton-Ferrer, C., Chen, M., Cucurull, G., Esiobu, D., Fernandes, J., Fu, J., Fu, W., Fuller, B., Gao, C., Goswami, V., Goyal, N., Hartshorn, A., Hosseini, S., Hou, R., Inan, H., Kardas, M., Kerkez, V., Khabsa, M., Kloumann, I., Korenev, A., Koura, P.S., Lachaux, M., Lavril, T., Lee, J., Liskovich, D., Lu, Y., Mao, Y., Martinet, X., Mihaylov, T., Mishra, P., Molybog, I., Nie, Y., Poulton, A., Reizenstein, J., Rungta, R., Saladi, K., Schelten, A., Silva, R., Smith, E.M., Subramanian, R., Tan, X.E., Tang, B., Taylor, R., Williams, A., Kuan, J.X., Xu, P., Yan, Z., Zarov, I., Zhang, Y., Fan, A., Kambadur, M., Narang, S., Rodriguez, A., Stojnic, R., Edunov, S., Scialom, T.: Llama 2: Open foundation and fine-tuned chat models. CoRR (2023)" + } + ] + } + ], + "index": 19 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 199, + 70, + 470, + 82 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 199, + 70, + 470, + 82 + ], + "spans": [ + { + "bbox": [ + 199, + 70, + 470, + 82 + ], + "type": "text", + "content": "PiTe: Pixel-Temporal Alignment for Large Video-Language Model" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 491, + 71, + 501, + 79 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 491, + 71, + 501, + 79 + ], + "spans": [ + { + "bbox": [ + 491, + 71, + 501, + 79 + ], + "type": "text", + "content": "15" + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 14 + }, + { + "para_blocks": [ + { + "bbox": [ + 108, + 95, + 539, + 436 + ], + "type": "list", + "angle": 0, + "index": 16, + "blocks": [ + { + "bbox": [ + 108, + 95, + 539, + 118 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 108, + 95, + 539, + 118 + ], + "spans": [ + { + "bbox": [ + 108, + 95, + 539, + 118 + ], + "type": "text", + "content": "39. Vedantam, R., Zitnick, C.L., Parikh, D.: Cider: Consensus-based image description evaluation. In: Proc. of CVPR. pp. 4566-4575 (2015)" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 108, + 118, + 539, + 150 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 108, + 118, + 539, + 150 + ], + "spans": [ + { + "bbox": [ + 108, + 118, + 539, + 150 + ], + "type": "text", + "content": "40. Wang, Y., He, Y., Li, Y., Li, K., Yu, J., Ma, X., Li, X., Chen, G., Chen, X., Wang, Y., Luo, P., Liu, Z., Wang, Y., Wang, L., Qiao, Y.: Internvid: A large-scale video-text dataset for multimodal understanding and generation. In: Proc. of ICLR (2024)" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 108, + 150, + 539, + 183 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 108, + 150, + 539, + 183 + ], + "spans": [ + { + "bbox": [ + 108, + 150, + 539, + 183 + ], + "type": "text", + "content": "41. Xu, D., Zhao, Z., Xiao, J., Wu, F., Zhang, H., He, X., Zhuang, Y.: Video question answering via gradually refined attention over appearance and motion. In: Proceedings of the 2017 ACM on Multimedia Conference, MM 2017, Mountain View, CA, USA, October 23-27, 2017. pp. 1645-1653 (2017)" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 108, + 183, + 539, + 205 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 108, + 183, + 539, + 205 + ], + "spans": [ + { + "bbox": [ + 108, + 183, + 539, + 205 + ], + "type": "text", + "content": "42. Xu, J., Zhou, X., Yan, S., Gu, X., Arnab, A., Sun, C., Wang, X., Schmid, C.: Pixel-aligned language model. In: Proc. of CVPR. pp. 13030-13039 (2024)" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 108, + 205, + 539, + 228 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 108, + 205, + 539, + 228 + ], + "spans": [ + { + "bbox": [ + 108, + 205, + 539, + 228 + ], + "type": "text", + "content": "43. Xu, J., Mei, T., Yao, T., Rui, Y.: MSR-VTT: A large video description dataset for bridging video and language. In: Proc. of CVPR. pp. 5288-5296 (2016)" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 108, + 228, + 539, + 250 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 108, + 228, + 539, + 250 + ], + "spans": [ + { + "bbox": [ + 108, + 228, + 539, + 250 + ], + "type": "text", + "content": "44. Yang, A., Miech, A., Sivic, J., Laptev, I., Schmid, C.: Zero-shot video question answering via frozen bidirectional language models. In: Proc. of NeurIPS (2022)" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 108, + 250, + 539, + 271 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 108, + 250, + 539, + 271 + ], + "spans": [ + { + "bbox": [ + 108, + 250, + 539, + 271 + ], + "type": "text", + "content": "45. Yu, Z., Xu, D., Yu, J., Yu, T., Zhao, Z., Zhuang, Y., Tao, D.: Activitynet-qa: A dataset for understanding complex web videos via question answering. In: Proc. of AAAI. pp. 9127-9134 (2019)" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 108, + 271, + 539, + 293 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 108, + 271, + 539, + 293 + ], + "spans": [ + { + "bbox": [ + 108, + 271, + 539, + 293 + ], + "type": "text", + "content": "46. Zhang, H., Li, X., Bing, L.: Video-llama: An instruction-tuned audio-visual language model for video understanding. In: Proc. of EMNLP. pp. 543-553 (2023)" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 108, + 293, + 539, + 316 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 108, + 293, + 539, + 316 + ], + "spans": [ + { + "bbox": [ + 108, + 293, + 539, + 316 + ], + "type": "text", + "content": "47. Zhang, M., Huang, S., Li, W., Wang, D.: Tree structure-aware few-shot image classification via hierarchical aggregation. In: Proc. of ECCV. pp. 453-470 (2022)" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 108, + 316, + 539, + 337 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 108, + 316, + 539, + 337 + ], + "spans": [ + { + "bbox": [ + 108, + 316, + 539, + 337 + ], + "type": "text", + "content": "48. Zhang, R., Han, J., Liu, C., Zhou, A., Lu, P., Qiao, Y., Li, H., Gao, P.: LLaMA-adapter: Efficient fine-tuning of large language models with zero-initialized attention. In: Proc. of ICLR (2024)" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 108, + 337, + 539, + 370 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 108, + 337, + 539, + 370 + ], + "spans": [ + { + "bbox": [ + 108, + 337, + 539, + 370 + ], + "type": "text", + "content": "49. Zhang, S., Roller, S., Goyal, N., Artetxe, M., Chen, M., Chen, S., Dewan, C., Diab, M.T., Li, X., Lin, X.V., Mihaylov, T., Ott, M., Shleifer, S., Shuster, K., Simig, D., Koura, P.S., Sridhar, A., Wang, T., Zettlemoyer, L.: OPT: open pre-trained transformer language models. CoRR (2022)" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 108, + 370, + 539, + 392 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 108, + 370, + 539, + 392 + ], + "spans": [ + { + "bbox": [ + 108, + 370, + 539, + 392 + ], + "type": "text", + "content": "50. Zhang, Y., Li, Z., Zhang, M.: Efficient second-order treecrf for neural dependency parsing. In: Proc. of ACL. pp. 3295-3305 (2020)" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 108, + 392, + 539, + 414 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 108, + 392, + 539, + 414 + ], + "spans": [ + { + "bbox": [ + 108, + 392, + 539, + 414 + ], + "type": "text", + "content": "51. Zhang, Y., Zhou, H., Li, Z.: Fast and accurate neural CRF constituency parsing. In: Proc. of IJCAI. pp. 4046-4053 (2020)" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 108, + 414, + 539, + 436 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 108, + 414, + 539, + 436 + ], + "spans": [ + { + "bbox": [ + 108, + 414, + 539, + 436 + ], + "type": "text", + "content": "52. Zhu, D., Chen, J., Shen, X., Li, X., Elhoseiny, M.: MiniGPT-4: Enhancing vision-language understanding with advanced large language models. In: Proc. of ICLR (2024)" + } + ] + } + ], + "index": 15 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 108, + 71, + 120, + 79 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 108, + 71, + 120, + 79 + ], + "spans": [ + { + "bbox": [ + 108, + 71, + 120, + 79 + ], + "type": "text", + "content": "16" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 141, + 70, + 229, + 81 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 141, + 70, + 229, + 81 + ], + "spans": [ + { + "bbox": [ + 141, + 70, + 229, + 81 + ], + "type": "text", + "content": "Y. Liu, P. Ding et al." + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 15 + } + ], + "_backend": "vlm", + "_version_name": "2.6.4" +} \ No newline at end of file diff --git a/2024/Pick-a-back_ Selective Device-to-Device Knowledge Transfer in Federated Continual Learning/ecda4e75-a956-4af4-a55c-23aa50728c18_content_list.json b/2024/Pick-a-back_ Selective Device-to-Device Knowledge Transfer in Federated Continual Learning/ecda4e75-a956-4af4-a55c-23aa50728c18_content_list.json new file mode 100644 index 0000000000000000000000000000000000000000..8ad686f716841f067f1355e4d6423ec3a9f2dc75 --- /dev/null +++ b/2024/Pick-a-back_ Selective Device-to-Device Knowledge Transfer in Federated Continual Learning/ecda4e75-a956-4af4-a55c-23aa50728c18_content_list.json @@ -0,0 +1,1508 @@ +[ + { + "type": "text", + "text": "Pick-a-back: Selective Device-to-Device Knowledge Transfer in Federated Continual Learning", + "text_level": 1, + "bbox": [ + 217, + 140, + 785, + 186 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "JinYi Yoon1 and HyungJune Lee2", + "bbox": [ + 361, + 210, + 638, + 229 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "$^{1}$ Virginia Tech, Blacksburg, VA, USA jinyioon@vt.edu \n $^{2}$ Ewha Womans University, Seoul, South Korea hyungjune.lee@ewha.ac.kr", + "bbox": [ + 339, + 238, + 663, + 295 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Abstract. With the explosion of edge intelligence, leveraging federated indirect knowledge has become crucial for boosting the tasks of individual learners. However, the conventional approach to knowledge reuse often leads to catastrophic forgetting issues. In this paper, we revisit the concept of continual learning in the context of edge intelligence and address the knowledge transfer problem to enhance federated continual learning. Since each learner processes private heterogeneous data, we propose Pick-a-back, a device-to-device knowledge federation framework by selectively reusing the external knowledge with similar behavioral patterns. By borrowing indirect experiences, an edge device can initiate learning from useful knowledge and thus achieve faster yet more generalized knowledge acquisition. Using continual tasks consisting of various datasets on lightweight architectures, we have validated that Pick-a-back provides a significant inference improvement of up to $8.0\\%$ via selective knowledge federation. Our codes are available at https://github.com/jinyi-yoon/Pick-a-back.git", + "bbox": [ + 261, + 330, + 743, + 539 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Keywords: Federated continual learning, edge intelligence, device-to-device knowledge federation, selective knowledge transfer", + "bbox": [ + 259, + 551, + 743, + 580 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "1 Introduction", + "text_level": 1, + "bbox": [ + 215, + 604, + 375, + 619 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "As artificial intelligence (AI) permeates nearly every field, data has emerged as a pivotal source driving remarkable advancements. Most of the data originates from the end users or systems, stimulating a need for distributed learning on low-end edge devices to address privacy and latency concerns [2,31]. However, individual low-end edge devices often lack the computational capability to generate advanced knowledge [6]. In the context of human learning paradigms, where any individuals become educated and exchange helpful knowledge with one another, the emergence of edge intelligence and its widespread availability allow external knowledge to serve as a valuable resource in supporting individual learners. By facilitating some sharable insights [28,59] and harnessing collective intelligence [14], it has widely revolutionized learning paradigms across various fields, even where numerous data have non-identical forms and heterogeneous distributions.", + "bbox": [ + 212, + 635, + 787, + 816 + ], + "page_idx": 0 + }, + { + "type": "page_footnote", + "text": "* Corresponding author", + "bbox": [ + 217, + 824, + 382, + 839 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "While leveraging external knowledge to enhance individual tasks, conventional learners often encounter the challenge of catastrophic forgetting, wherein they adapt only to the assigned tasks, and consequently lose previously acquired knowledge in some degree over time. In contrast, to more effectively retain and reinforce pre-trained knowledge together with new knowledge, the learners would rather adopt a continual learning approach for ensuring both robustness and adaptability. To enhance the efficacy of individual continual learning, it is not imperative for learners to confine themselves to their own knowledge. The prevalence of edge intelligence may facilitate knowledge federation, enabling individual learners to leverage a broader or deeper understanding of data, which can be possibly heterogeneous. While recent works have devised federated continual learning frameworks, they typically take a centralized approach for assistance [3,18,50], rather than direct device-to-device knowledge transfer in a distributed manner.", + "bbox": [ + 212, + 146, + 787, + 342 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "As for edge AI, individual learners might be confronted with a series of diverse heterogeneous tasks. At the edge side accompanied with its own private data, each learner attempts to form its own distinct knowledge. Then, a question to raise is that how can an individual edge learner effectively harness externally available prior knowledge? It has been reported that there exists unique positive and negative knowledge at each model [43,57]. Maybe the answer would be straightforward yet meaningful – the learners would selectively reuse or customize some knowledge that aligns well with their specific behavioral patterns. Rather than starting from scratch, learners can benefit from the indirectly attained knowledge formed by others who have already seen similar patterns. In that the decision patterns can not be exactly the same, indirect knowledge transfer (with follow-up training) may offer a rich source of otherwise untrainable and unattainable knowledge from private data. This can lead to performance improvement via selective acquisition of some common yet distinct knowledge.", + "bbox": [ + 212, + 347, + 787, + 559 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "In this work, we revisit the knowledge reuse problem. We consider three challenges in edge intelligence with knowledge transfer: 1) edge devices have limited computational and storage resources; 2) learners are not willing to disclose their raw data; and 3) edge devices should take into account the reduction of communication costs. To overcome these concerns, our approach aims to enhance the performance of lightweight learning models, while maintaining data privacy and avoiding additional communication overhead.", + "bbox": [ + 212, + 563, + 787, + 667 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "We propose a selective and customized knowledge federation framework from one device to another, based on heterogeneous continual learning, called Pick-a-back, which picks a relevant backbone knowledge from other devices. Beyond the achievable knowledge of each individual learner, we strategically search for some effective knowledge sources selectively and then customize the commonly shared knowledge to closely align with the local task. By leveraging decision similarity based on mutated data to assess the efficacy of the backbone among the neighboring models spread in the form of physically separated edge intelligence, we selectively fetch and customize constructive knowledge. This approach not only enables decentralized knowledge exchange among the learners at devices, but also facilitates a more efficient and scalable learning paradigm.", + "bbox": [ + 212, + 672, + 787, + 840 + ], + "page_idx": 1 + }, + { + "type": "page_number", + "text": "2", + "bbox": [ + 217, + 114, + 228, + 126 + ], + "page_idx": 1 + }, + { + "type": "header", + "text": "J. Yoon and H. Lee", + "bbox": [ + 271, + 114, + 403, + 127 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "This work contributes to advancing the field of deep learning in edge intelligence by providing selective knowledge transfer and fostering collaboration among local learners. It can thereby enhance the efficiency of the individual continual learning process. The main contributions can be summarized as follows:", + "bbox": [ + 212, + 146, + 787, + 207 + ], + "page_idx": 2 + }, + { + "type": "list", + "sub_type": "text", + "list_items": [ + "1) We address the problem of device-to-device knowledge federation in the continual learning. We demonstrate the potential of leveraging indirect knowledge extracted by the neighboring edge learners to obtain otherwise untrainable and unattainable insights. This work provides a way to maintain a model with heterogeneous data collected across a network of edge devices.", + "2) As the diverse tasks imply their own unique knowledge, we propose selective and customized knowledge transfer by capturing the decision pattern similarity relevant to a target task running on a device. Pick-a-back adopts each relevant knowledge for a specific task, achieving high fidelity through the acquisition of some common yet distinct knowledge from unseen examples, while preserving data privacy and communication cost at the edge.", + "3) Extensive experimental results on various datasets across different architectures have shown that Pick-a-back provides a significant inference improvement up to $8.0\\%$ via selective knowledge federation." + ], + "bbox": [ + 220, + 215, + 787, + 425 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "2 Related Work", + "text_level": 1, + "bbox": [ + 215, + 448, + 387, + 464 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Typical continual learning refers to training on a series of multiple tasks [22] or incremental classes [13]. The primary objective is to address the catastrophic forgetting problem, where previously learned knowledge is decaying over time as adapting to new tasks, using memorization [40,52] or regularization [9,23].", + "bbox": [ + 212, + 478, + 787, + 542 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Knowledge-reusable continual learning. More closely related to our problem, there has been a growing interest in reusing existing knowledge for its subsequent tasks [30, 58]. Unlike conventional approaches, these methods have focused on concatenating a new model to its previous model, allowing input data to propagate through the previous model [46]. Rather than reusing the whole model, researchers have proposed to activate only the essential parts of the model [19,29]. In *Piggyback* [29], they leverage a common model and mask out unnecessary neurons to extract relevant knowledge for a target task. It effectively reuses existing knowledge with only a single sharable model, but it is less able to accept incremental tasks. *CPG* [19] combines neuron expansion with neuron pruning and masking techniques to address the challenges of forgetting, model compactness, and achieving higher fidelity in knowledge reuse. We adopt these knowledge reusable continual learning approaches to retain and extract valuable insights from previous tasks, while aiming to find useful knowledge (for reuse) based on model similarity. Although some works have applied some similarity metrics such as cosine similarity to federated learning scenarios [20], it is limited in applicability when dealing with clients having a different number of outputs (classes). In contrast, our approach provides flexibility in output heterogeneity, ensuring broader applicability in real-world scenarios, where most clients may", + "bbox": [ + 212, + 551, + 787, + 843 + ], + "page_idx": 2 + }, + { + "type": "header", + "text": "Pick-a-back", + "bbox": [ + 651, + 114, + 730, + 126 + ], + "page_idx": 2 + }, + { + "type": "page_number", + "text": "3", + "bbox": [ + 774, + 116, + 784, + 126 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "have different shapes of outputs. Moreover, our decision pattern-based similarity is novel in federated continual learning, as the knowledge itself is not exactly the same even with the similar decision patterns. This enables private learners to be exposed to similar yet unseen broader knowledge, leading to improved performance and efficiency in subsequent learning tasks.", + "bbox": [ + 212, + 146, + 787, + 223 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Federated continual learning. With widespread accessibility of edge AI, learners do not necessarily be confined to their own private knowledge, as a standalone learner is permitted to only access its own local information and thus suffer from restricted experiences to some extent. Although recent works tackle the problem of federated continual learning [50,61], they focus on class-incremental learning, whereas we tackle task-incremental learning. In task-incremental learning, transfer learning turns out to be very effective, where a model is pretrained based on the source domain and then finetuned for the target domain. This selective device-to-device knowledge transfer can be a significant impetus to enhance model fidelity, also in continual learning. It has been commonly addressed in the field of federated continual learning but only in a centralized manner [18,50]. It has been also observed that there is both positive and negative knowledge related to a specific target task [43,57], highlighting the potential harm of static knowledge selection. To address the two-sided nature of knowledge reuse, FedWeIT [57] aggregates the parameters from clients as global knowledge, and then the clients mask the federated parameters adaptive to the specific task. However, previous works including FedWeIT rely on a central server, causing communication bottlenecks and privacy limitations. Furthermore, ImageNet-pretrained models have widely been utilized in transfer learning. However, it is not feasible for edge-side federated continual learning, as pretrained models are specifically designed for well-known large-scale architectures. The architectures in continual learning are dynamically changing, and edge AI cannot be tailored to large-scale models. In Pick-a-back, we rather explore decentralized, communication-efficient, and privacy-preserving knowledge transfer in lightweight on-device models.", + "bbox": [ + 212, + 243, + 789, + 606 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "3 Selective Knowledge Transfer", + "text_level": 1, + "bbox": [ + 215, + 635, + 537, + 652 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "We consider the problem of device-to-device knowledge federation to enhance individual continual learning in a distributed manner. We propose Pick-a-back, a simple yet effective knowledge federation approach by discovering a relevant neighboring device who has acquired the knowledge of relevant tasks and then selectively infusing it into the model via continual learning. By leveraging a federated learning framework, which allows collaborative learning across devices, we aim to address the challenge of distributed continual learning, while also facilitating the reuse of previously acquired knowledge at each different device. The integration of these two fields of study enables efficient knowledge selection, sharing, and expansion among devices (with heterogeneous data), yielding continuous learning improvement and adaptation across tasks or learning scenarios.", + "bbox": [ + 212, + 672, + 789, + 840 + ], + "page_idx": 3 + }, + { + "type": "page_number", + "text": "4", + "bbox": [ + 217, + 114, + 230, + 126 + ], + "page_idx": 3 + }, + { + "type": "header", + "text": "J. Yoon and H. Lee", + "bbox": [ + 271, + 114, + 403, + 127 + ], + "page_idx": 3 + }, + { + "type": "image", + "img_path": "images/9d33422866b30461ddd6c6b21b8acaa4f2eb16d106e360a8320233fc715468c0.jpg", + "image_caption": [ + "Fig. 1: Overall architecture of Pick-a-back: (1) Each device trains an individual continual learning model on a series of tasks. (2) Given these available neighboring knowledge, a device selects a relevant neighboring model that can improve learning performance by acquiring indirect knowledge from the federated but physically separated devices. (3) On top of the retrieved knowledge, (4) the continual learner stacks new local tasks and resumes learning and federation. By fetching the generalized knowledge with relevant decision patterns, Pick-a-back offers a cumulative yet selective knowledge federation with heterogeneous data at a distinct device" + ], + "image_footnote": [], + "bbox": [ + 248, + 145, + 754, + 305 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "3.1 Problem Formulation", + "text_level": 1, + "bbox": [ + 215, + 460, + 437, + 474 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Given $L$ devices, each learner $l \\in \\{1, 2, \\dots, L\\}$ at device $d_l$ processes its own sequential tasks (i.e., training data) of $T^{(l)} = \\{T_1^{(l)}, T_2^{(l)}, \\dots \\}$ . A scratch model $M_i^{(l)}$ is constructed by training only with private data $T_i^{(l)} = \\{(X_i^{(l)}, Y_i^{(l)})\\}$ , where $T_i^{(l)}$ for task $i$ and device $d_l$ is heterogeneous across other tasks and devices. We train the model $\\hat{M}_i^{(l)}$ with target data $T_i^{(l)}$ on top of its previous version $\\hat{M}_{i-1}^{(l)}$ with some learning parameters $\\theta_i^{(l)}$ , i.e., $\\hat{M}_i^{(l)} = \\hat{M}_{i-1}^{(l)}(T_i^{(l)}, \\theta_i^{(l)})$ in a continual manner. Here, at every task $i$ , the goal is to find a good reference model that can maximize the overall fidelity of model $\\hat{M}_i^{(l)}$ .", + "bbox": [ + 212, + 491, + 787, + 630 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "3.2 Architecture", + "text_level": 1, + "bbox": [ + 215, + 657, + 367, + 671 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Our learning architecture incorporates two distinct modes (Fig. 1): i) continual learner for local training; and ii) federated learner for knowledge transfer. Initially, multiple learners train independently on a specific private task. Subsequently, one learner at a device transitions to a federated learning mode to explore effective indirect knowledge from a neighboring relevant model. Once a neighboring model is selected, we use the selected model as a backbone, the main target task is then executed in a continual manner. By incorporating the retrieved knowledge together with its existing own knowledge, the learner can construct cumulative knowledge in a continual manner without forgetting. This process of continual learning and federated learning is repeated for each task.", + "bbox": [ + 212, + 688, + 787, + 840 + ], + "page_idx": 4 + }, + { + "type": "header", + "text": "#", + "bbox": [ + 633, + 59, + 656, + 75 + ], + "page_idx": 4 + }, + { + "type": "header", + "text": "EWHA WOMANS UNIVERSITY", + "bbox": [ + 658, + 61, + 800, + 71 + ], + "page_idx": 4 + }, + { + "type": "header", + "text": "Pick-a-back", + "bbox": [ + 651, + 114, + 730, + 126 + ], + "page_idx": 4 + }, + { + "type": "page_number", + "text": "5", + "bbox": [ + 774, + 114, + 784, + 126 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Continual learner. In our approach to continual learning, we adopt task-incremental learning, where a series of tasks is processed sequentially. Each task has a distinct learning objective, and a local model is trained independently on each task with freezing or isolating the models for previous knowledge. By employing task-oriented learning, we aim to mitigate the catastrophic forgetting problem, where previously learned knowledge vanishes as training on new tasks or incorporating new data over time. Our approach can be combined with any continual learner that has gradually been constructing knowledge. It is fundamentally different from conventional continual learning that simply memorizes multiple models or data, which do not leverage the given knowledge. We suppose that a continual learner concatenates a new model on top of the previously acquired knowledge, and thus it can reduce memory usage and also enhance the model fidelity. This sustainability of the continual model is particularly valuable in the context of edge AI and aligns with the goals of the recent learning paradigm.", + "bbox": [ + 212, + 146, + 787, + 357 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "We pursue reusing the pre-configured knowledge by using the model. We mainly use the latest continual learning effort, $CPG$ [19], which expands the model when necessary, prunes it for compactness, and reuses the network by masking out specific ineffective neurons. Inspired by the masking strategy of Piggyback [29], it samples some necessary neurons for the target task to allow the extraction of common knowledge. $CPG$ only trains the learnable binary mask whether to activate each neuron for the target task. As it ensures the retention of previous knowledge via freezing, the results remain exactly the same even when training new tasks. However, without loss of generality, this masking strategy can be applied to any continual learner that holds multiple knowledge.", + "bbox": [ + 212, + 358, + 787, + 508 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Federated learner. Due to the privacy and communication overhead, distributed learners are unwilling to share private raw data. We thus adopt a model-sharing approach rather than data-sharing. Here, each client is regarded as an independent learner. When each learner at a distinct device is working on a distinct task but lacks knowledge, there might be some other knowledge that has a similar, broader, or deeper experience beyond the learner based on its own local data. We assume that clients are knowledgeable with their own currently available neighboring models via existing neighbor selection methods [38, 53] or graph construction approaches [33, 62]. These neighboring diverse knowledge sources have been trained on different datasets, capturing various perspectives and feature extraction capabilities. By exposing the learner to external knowledge, the model is capable of leveraging enhanced feature representations, enabling it to capture a broader set of patterns and information that may be beneficial for its specific learning objective. The learner can also develop a more comprehensive understanding of the data, generalize better to unseen examples by mitigating overfitting, and thus improve the fidelity of the model on a variety of tasks and data. By leveraging a diverse ensemble of knowledge sources, the learner can benefit from its complementary expertise, leading to accuracy improvement.", + "bbox": [ + 212, + 522, + 787, + 792 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Each learner at a device seeks relevant knowledge that can contribute to further improving its performance based on local continual learning. To acquire necessary constructive knowledge, we identify effective indirect experiences based", + "bbox": [ + 212, + 795, + 787, + 839 + ], + "page_idx": 5 + }, + { + "type": "page_number", + "text": "6", + "bbox": [ + 217, + 114, + 228, + 126 + ], + "page_idx": 5 + }, + { + "type": "header", + "text": "J. Yoon and H. Lee", + "bbox": [ + 271, + 114, + 403, + 127 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "on decision pattern similarities. Once we have obtained knowledge from other learners, we accumulate ongoing new local tasks on top of the model in a continual manner, ensuring that the retrieved knowledge becomes infused with local knowledge. Our main focus is on the current target task, and thus, if we want to retain the knowledge from our previous tasks, we have to either store the existing knowledge separately or concatenate it with the fetched model. If the currently constructed knowledge turns out to be sufficiently effective, there may be no further need to incorporate further knowledge from other learners.", + "bbox": [ + 212, + 146, + 787, + 268 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "3.3 Selective Knowledge Federation", + "text_level": 1, + "bbox": [ + 214, + 287, + 524, + 303 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Each learner is responsible for distinct tasks, and tasks may exhibit unique strengths, weaknesses, or biases. Consequently, each task at a local learner plays its own unique different role and forms distinct knowledge. Some external models may have learned other representations that have captured specialized knowledge relevant to specific tasks. Therefore, selective knowledge transfer can be a very effective way for enhancing the performance of each student model at an edge device. Our approach is simple and straightforward – if models have similar or synergetic capabilities in classifying different samples, they are likely to have potential sources of common but more generalized knowledge. Since it has the possibility to attain broader intuition but also inexperienced knowledge, this indirect knowledge characterized by similar decision patterns may greatly accelerate and improve the learning for individual learners.", + "bbox": [ + 212, + 311, + 787, + 492 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "As the similarity of model do not straightly incur similar knowledge, we focus on the decision pattern similarity rather than the model itself. We borrow an idea to calculate the model similarity from ModelDiff [27], which calculates the similarity of decision boundaries by comparing the reaction to the input data. Initially, we sample input data $x_{i}$ from the requester and adversarial data $x_{i}'$ from a neighbor. To address privacy concerns, an alternative approach may sample data only from the requester, or the participants may opt to exchange mutated data, thus avoiding the sharing of private data. Then, we inject a pair-wise set of input data, $\\{(x_{1},x_{1}'),(x_{2},x_{2}'),\\dots ,(x_{n},x_{n}')\\}$ , into each model $f$ and $g$ and then computes the decision distance vector (DDV), $DDV_{f} = \\{d_{1},d_{2},\\dots ,d_{n}\\}$ , which is a distance vector for the output pair, where $d_{i} = \\text{distance}(f(x_{i}),f(x_{i}'))$ , $i\\in \\{1,2,\\ldots ,n\\}$ . Finally, it compares the cosine similarity between the distance vectors of models, $sim(DDV_{f},DDV_{g})$ . Since it focuses the decision similarity instead of directly comparing the outputs, it is applicable to any heterogeneous task with a different number of output classes. To search a pair-wise set that can approximate the decision boundary most effectively, the input data is mutated, and similarities are iteratively computed, similar to genetic algorithms.", + "bbox": [ + 212, + 493, + 787, + 750 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "By incorporating it in Pick-a-back, when we search for a similar model for task $T_{i}^{(l)}$ of learner $l$ among $\\hat{M}_{i-1}^{k}$ of all neighboring learner $k$ . We sample an input pair-wise data $X_{i}^{(l+k)}$ and $X_{i}^{\\prime(l+k)}$ from $X_{i}^{(l)} \\cup X_{i-1}^{(k)}$ (or just from $X_{i}^{(l)}$ ). By injecting $X_{i}^{(l+k)}$ and $X_{i}^{\\prime(l+k)}$ into the learner model $M_{i}^{(l)}$ and the neighboring models $\\hat{M}_{i-1}^{(k)}$ , we produce the distance vector of outputs $D_{M_{i}^{(l)}}$ and $D_{\\hat{M}_{i-1}^{(k)}}$ . Then,", + "bbox": [ + 212, + 750, + 789, + 843 + ], + "page_idx": 6 + }, + { + "type": "header", + "text": "Pick-a-back", + "bbox": [ + 651, + 114, + 730, + 126 + ], + "page_idx": 6 + }, + { + "type": "page_number", + "text": "7", + "bbox": [ + 774, + 114, + 784, + 125 + ], + "page_idx": 6 + }, + { + "type": "code", + "sub_type": "algorithm", + "code_caption": [ + "Algorithm 1 Federated Continual Learning with Pick-a-back" + ], + "code_body": "1: Input: $L$ : # of devices, (for each device $d_{l}$ ) $M_{0}^{(l)}$ : random initial model, $T^{(d_{l})}$ : a sequence of tasks, $\\theta_{i}^{(l)}$ : learning parameters \n2: for task $i = 1: length(T^{(l)})$ do \n3: Device $d_{l}$ trains task $i$ such that $M_{i}^{(l)} = M_{i-1}^{(l)}(T_{i}^{(l)}, \\theta_{i}^{(l)})$ ; $\\triangleright$ Continual Learner \n4: for device $k = 1: L$ do $\\triangleright$ Federated Learner \n5: Compute the similarity, $sim(M_{i}^{(l)}, \\hat{M}_{i-1}^{(k)})$ ; \n6: end for \n7: Select and retrieve knowledge $\\hat{M}_{i-1}^{(l)} \\gets \\hat{M}_{i-1}^{(k)}$ such that $\\arg \\max_{k} sim(M_{i}^{(l)}, \\hat{M}_{i-1}^{(k)})$ ; \n8: Train task $i$ on the fetched model, $\\hat{M}_{i}^{(l)} = \\hat{M}_{i-1}^{(l)}(T_{i}^{(l)}, \\theta_{i}^{(l)})$ ; $\\triangleright$ Continual Learner \n9: end for", + "bbox": [ + 222, + 166, + 785, + 313 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "we select the most similar model as follows:", + "bbox": [ + 215, + 340, + 529, + 354 + ], + "page_idx": 7 + }, + { + "type": "equation", + "text": "\n$$\n\\underset {k} {\\operatorname {a r g m a x}} \\operatorname {s i m} \\left(D D V _ {M _ {i} ^ {(l)}}, D D V _ {\\hat {M} _ {i - 1} ^ {(k)}}\\right) \\tag {1}\n$$\n", + "text_format": "latex", + "bbox": [ + 375, + 363, + 785, + 387 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "We fetch the indirect knowledge that has already experienced the target decision pattern. After selecting the exterior knowledge, we continue to learn $T_{i}^{l}$ on $\\hat{M}_{i - 1}^{(k)}$ .", + "bbox": [ + 214, + 393, + 787, + 428 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Training process. Each learner $l$ trains $i$ -th task $T_{i}^{(l)}$ and constructs a local model $M_{i}^{(l)}$ with continual learning. Then, the learner $l$ checks if there exists any helpful exterior knowledge by comparing $M_{i}^{(l)}$ with $M_{i-1}^{(k)}$ for all neighboring edge learners $k$ . After selecting relevant knowledge for knowledge transfer, device $d_{l}$ processes the $i$ -th task $T_{i}^{(l)}$ over $\\hat{M}_{i-1}^{(k)}$ and constructs $\\hat{M}_{i}^{(l)} = \\hat{M}_{i-1}^{(k)}(T_{i}^{(l)}, \\theta_{i}^{(l)})$ .", + "bbox": [ + 214, + 438, + 787, + 523 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "The overall procedure is described in Algorithm 1.", + "bbox": [ + 238, + 523, + 602, + 539 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "4 Experiments", + "text_level": 1, + "bbox": [ + 215, + 556, + 375, + 574 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "We have validated Pick-a-back extensively on four different classification task groups with 13 public data using four different well-known architectures. Each group is considered as one experimental environment where knowledge can be shared, and one superclass group is a single task.", + "bbox": [ + 212, + 584, + 787, + 646 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Datasets. We used four different groups of CIFAR-100 [25], Classified-ImageNet [8], Fine-Grained [10,24,37,48,55], and Digits [7,11,35,36,39,45]. The organization of the dataset and the details are provided in Appendix Sec. A1.", + "bbox": [ + 214, + 655, + 799, + 703 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Architectures. We have presented the results using both lightweight models and large-scale models. For the lightweight models, we used LeNet-5 [26] and MobileNetV1 [17], where ImageNet-pretrained models are not publicly available. On the other hand, VGG16-BN [51] and EfficientNet-B0 [54] that have publicly available ImageNet-pretrained models were implemented using PyTorch [42]. Details of model configurations and learning parameters are in Appendix Sec. A2. We used CIFAR-100 on LeNet-5 as default, unless otherwise noted.", + "bbox": [ + 214, + 712, + 787, + 818 + ], + "page_idx": 7 + }, + { + "type": "page_number", + "text": "8", + "bbox": [ + 217, + 114, + 228, + 126 + ], + "page_idx": 7 + }, + { + "type": "header", + "text": "J. Yoon and H. Lee", + "bbox": [ + 271, + 114, + 403, + 126 + ], + "page_idx": 7 + }, + { + "type": "page_footnote", + "text": "1 Our codes are available at https://github.com/jinyi-yoon/Pick-a-back.git", + "bbox": [ + 217, + 824, + 764, + 840 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/a659d8c9b10ab0113ab026bad83487f231c600742db27fbf8aecc938d719c511.jpg", + "image_caption": [ + "(a) Absolute Acc." + ], + "image_footnote": [], + "bbox": [ + 217, + 143, + 334, + 224 + ], + "page_idx": 8 + }, + { + "type": "image", + "img_path": "images/fd37e59ee1bbd2997524d31a6964ca28ab86a0c0fc8db682f94784de713d3ab9.jpg", + "image_caption": [ + "(b) Improved Acc." + ], + "image_footnote": [], + "bbox": [ + 336, + 143, + 455, + 224 + ], + "page_idx": 8 + }, + { + "type": "image", + "img_path": "images/85cab9c58ed596d781e76927c2745f641eda0efdd4b085cf3447279db225d8b2.jpg", + "image_caption": [ + "(a) Training dynamics", + "Fig.2: Heatmap of the accuracy for various backbone and target tasks sorted by the accuracy of $w / o$ Backbone. The first row represents the averaged values for all tasks, while the first column shows the results of $w / o$ Backbone. Absolute accuracy is scaled per each row, and accuracy improvement is scaled across all results", + "Fig. 3: Training dynamics for the most effective case of the target task 14 using the backbone task of 4 and accuracy improvement with respect to the smoothed accuracy of $w / o$ Backbone. The smoothed accuracy refers to the averaged accuracy for a subset of tasks with lower accuracy. For example, the second x-value represents the average of two tasks with the lowest performance, while the last value is the average across all tasks" + ], + "image_footnote": [], + "bbox": [ + 465, + 152, + 617, + 224 + ], + "page_idx": 8 + }, + { + "type": "image", + "img_path": "images/9cd885a93119920e92c63bdbe08a52b893464870a5afabfbc16386bb638054b3.jpg", + "image_caption": [ + "(b) Accuracy improvement" + ], + "image_footnote": [], + "bbox": [ + 620, + 154, + 782, + 224 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "Baselines. We compare Pick-a-back with diverse learning approaches of federated learning, continual learning, and continual federated learning.", + "bbox": [ + 214, + 402, + 787, + 431 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "First, we let our Pick-a-back run on top of a certain continual learning scheme: 1) ProgressiveNet [47]: continual training only with expansion by concatenating scratch models for sequential tasks; 2) PackNet [30]: continual training only with pruning by allocating partial neurons exclusively to each task; 3) CPG [19]: continual training with expansion if necessary and pruning for the model compactness; and 4) WSN [21]: continual learning with reusing the masked weights of the prior subnetworks.", + "bbox": [ + 212, + 433, + 787, + 537 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "We further compare across existing federated and continual federated learning approaches: 1) FedAVG [32]: one of the pioneering works for model regularization in federated learning; 2) FedWeIT [57]: a federated continual learning paradigm by introducing sparse task-specific parameters along with global federated parameters; and 3) FCCL [18]: a federated continual learning approach to tackle the catastrophic forgetting problem within a heterogeneous setting.", + "bbox": [ + 212, + 539, + 787, + 630 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "We also compare Pick-a-back with different knowledge selection strategies: 1) $w / o$ Backbone: intact continual learning without using any selective knowledge transfer, such as Progressive Networks, PackNet, or CPG; 2) Uniform Knowledge: applying the same task to its backbone model and averaging all of the results for each task; 3) Internal Knowledge: training the target task iteratively as a backbone model; 4) Random Knowledge: randomly selecting a backbone model; and 5) Upper: selecting an ideal backbone with a certain task that shows the highest accuracy improvement, being served as an unachievable upper bound.", + "bbox": [ + 212, + 630, + 787, + 752 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "Measures. We have investigated 1) Avg: the averaged accuracy of all tasks and 2) Top: the accuracy of the most effective task that has gained the best accuracy improvement via selective knowledge. To compare the feasibility of models for edge AI, we have shown 1) Model size: the actual storage overhead of a model; 2) Expansion: the expansion degree where the scratch model has the value of 1;", + "bbox": [ + 212, + 763, + 789, + 840 + ], + "page_idx": 8 + }, + { + "type": "header", + "text": "Pick-a-back", + "bbox": [ + 651, + 114, + 730, + 126 + ], + "page_idx": 8 + }, + { + "type": "page_number", + "text": "9", + "bbox": [ + 774, + 114, + 784, + 126 + ], + "page_idx": 8 + }, + { + "type": "table", + "img_path": "images/dc7eee3fcac86e583cc01e8ad75e0025996a92987e4366e0556be8908bd596c9.jpg", + "table_caption": [ + "Table 1: Performance comparison. The bold text highlights the best performance per client, excluding Upper as an ideal upper-bound due to its infeasible brute-force manner" + ], + "table_footnote": [], + "table_body": "
Client ID1511716181231476824131951092011Avg
w/o Backbone40.055.057.260.451.866.067.467.669.270.872.673.074.275.275.876.477.077.481.284.069.1
FedAVG24.431.634.638.032.039.631.832.049.037.829.834.026.430.835.845.427.223.435.639.033.9
ProgressiveNet44.658.854.463.064.468.071.268.269.270.673.872.470.473.074.078.877.273.684.683.869.7
PackNet38.055.056.459.463.869.068.272.271.272.874.472.070.274.879.678.078.676.882.680.269.7
WSN34.253.849.453.459.860.263.663.066.865.064.665.466.662.661.267.666.658.674.079.861.8
FedWeIT40.052.555.056.056.260.556.760.567.065.260.265.066.064.265.065.766.751.068.777.861.0
FCCL37.054.452.460.662.268.267.264.468.270.672.872.468.477.073.875.078.668.879.278.267.5
Uniform41.558.657.662.165.670.769.771.671.272.574.773.073.776.477.178.779.475.283.583.070.8
Internal43.255.456.862.267.867.069.071.473.874.876.871.873.476.279.277.681.079.282.083.871.1
Random39.859.056.659.267.672.668.273.071.471.475.672.874.476.478.078.277.874.684.283.870.7
Upper46.462.460.064.267.874.872.275.673.875.677.075.076.079.480.080.681.679.286.285.473.7
Pick-a-back46.459.060.063.264.472.671.875.672.472.277.073.072.876.276.080.278.275.884.285.471.8
", + "bbox": [ + 218, + 196, + 784, + 338 + ], + "page_idx": 9 + }, + { + "type": "text", + "text": "3) Sparsity: the released ratio of a model, where higher sparsity indicates more space for next tasks; 4) Training time: the execution time to run for one epoch; and 5) Used data: the number of data samples used for training.", + "bbox": [ + 212, + 378, + 787, + 425 + ], + "page_idx": 9 + }, + { + "type": "text", + "text": "4.1 Overall Performance and Comparisons", + "text_level": 1, + "bbox": [ + 215, + 449, + 580, + 465 + ], + "page_idx": 9 + }, + { + "type": "text", + "text": "Motivations for selective knowledge federation. We first examined the impact of different knowledge on various target tasks. As in Fig. 2a, a significant portion of the results and the average accuracy achieved through backbone usage are represented by a blue color, indicating that the continual learner benefited from the knowledge federation. However, there exist not only positive backbones but also negative backbones, suggesting a need for selective knowledge to assist a continual learner specific to its target task. Taking a closer look at the accuracy improvement relative to $w / o$ Backbone in Fig. 2b, a well-constructed model such as task 9 or 11 is more likely disturbed by external knowledge. On the other hand, a target task with initially poor performance tends to gain a greater improvement in accuracy. It implies the importance of selectively incorporating constructive knowledge to enhance the learning process for a specific target task.", + "bbox": [ + 212, + 474, + 787, + 657 + ], + "page_idx": 9 + }, + { + "type": "text", + "text": "Ablation study on selective knowledge federation. Inspired by the Fig. 2 we verified the effectiveness, compared to diverse learning approaches, including FedAVG, ProgressiveNet, PackNet, WSN, FedWeIT, and FCCL, along with various backbone selection strategies of Uniform, Internal, Random, and Upper, as in Table 1 and in Table A5. We consistently demonstrate superior fidelity on average. By employing our selection approach, we observed a significant accuracy improvement of up to $8.0\\%$ , with only two negative cases out of 20 tasks. Since FedAVG simply averages the model parameters without considering the continual tasks, it suffers from losing the knowledge of trained tasks. Notably, although Pick-a-back with CPG shows the best performance, Pick-a-back with other continual learning approaches of ProgressiveNet and PackNet also outperforms", + "bbox": [ + 212, + 672, + 787, + 840 + ], + "page_idx": 9 + }, + { + "type": "page_number", + "text": "10", + "bbox": [ + 217, + 114, + 235, + 126 + ], + "page_idx": 9 + }, + { + "type": "header", + "text": "J. Yoon and H. Lee", + "bbox": [ + 271, + 114, + 403, + 127 + ], + "page_idx": 9 + }, + { + "type": "text", + "text": "$w / o$ Backbone. Even Uniform, Internal, and Random selection strategies exhibit enhanced performance, suggesting the efficacy of inter-device knowledge transfer based on Pick-a-back primitive. We also compared with other recent (federated) continual learning approaches (e.g., WSN, FedWeIT, and FCCL), showing that Pick-a-back outperforms these methods. This is due to the effectiveness of our knowledge reuse approach along with our backbone selection strategy.", + "bbox": [ + 212, + 146, + 787, + 238 + ], + "page_idx": 10 + }, + { + "type": "text", + "text": "Benefit of backbone knowledge. Taking a closer look at how Pick-a-back benefits from backbone knowledge, in Fig. 3a, it is evident that the test accuracy of Pick-a-back consistently surpasses the training accuracy of $w / o$ Backbone, starting from the beginning until the saturation. It indicates that indirect knowledge assists individual learners to be exposed to inexperienced knowledge and adapt to unseen data. Even within Pick-a-back itself, the test accuracy outperforms the training accuracy in the early stages of training, suggesting that the transferred knowledge has relevance to the target knowledge and drives toward high performance even earlier. Furthermore, $w / o$ Backbone achieves an accuracy of $67.4\\%$ , which can be attained only within approximately 20 epochs using Pick-a-back, whereas it takes over 50 epochs without knowledge transfer. This demonstrates that Pick-a-back not only accelerates the learning, but also enhances the overall performance. Please refer to the empirical convergence analysis in Appendix Fig. A1.", + "bbox": [ + 212, + 253, + 789, + 450 + ], + "page_idx": 10 + }, + { + "type": "text", + "text": "Possibility of similar but generalized knowledge. To justify the task-specific selection, we compared to uniform knowledge, where the same backbone is applied to all tasks. As in Fig. 3b, the results showed that Upper exhibits better improvement in tasks with poor initial performance. However, the uniform knowledge ignores task-specific improvements, leading to similar improvements across all target tasks. Interestingly, we also observed that only using internal knowledge, where the target task is trained twice, allowed further improvement. It means that even with a single task iteratively training by accumulating knowledge on top of an identical model turns out to induce better knowledge. It also implies that a backbone task with a similar, ideally the same in personal knowledge case, the decision boundary can further improve the knowledge. While the same task aims for identical knowledge, using similar but distinct knowledge enables generalization and better improvement beyond individual tasks. As expected, Pick-a-back outperformed the one only with internal knowledge, highlighting the advantage of leveraging similar indirect experiences. This means that we provide a more opportunity to acquire superior knowledge by mitigating the overfitting problem and extending to the unseen data beyond local knowledge.", + "bbox": [ + 212, + 465, + 789, + 723 + ], + "page_idx": 10 + }, + { + "type": "text", + "text": "4.2 Scalability Analysis", + "text_level": 1, + "bbox": [ + 214, + 744, + 423, + 760 + ], + "page_idx": 10 + }, + { + "type": "text", + "text": "Lightweight models for edge AI. We evaluated Pick-a-back in various architectures from the lightweight models of LeNet-5, MobileNet-V1, and EfficientNet-B0[2]", + "bbox": [ + 212, + 771, + 787, + 804 + ], + "page_idx": 10 + }, + { + "type": "header", + "text": "Pick-a-back", + "bbox": [ + 651, + 114, + 730, + 126 + ], + "page_idx": 10 + }, + { + "type": "page_number", + "text": "11", + "bbox": [ + 767, + 116, + 782, + 126 + ], + "page_idx": 10 + }, + { + "type": "page_footnote", + "text": "2 For EfficientNet-B0, we enlarge the images to $224 \\times 224$ pixels for evaluation, but a model size is measured with the original size of $32 \\times 32$ for a fair comparison.", + "bbox": [ + 217, + 810, + 787, + 839 + ], + "page_idx": 10 + }, + { + "type": "table", + "img_path": "images/9f2cb4cd3203fb9b619b390a98a668f8972af5179d67706c06e7f9240dd4a11b.jpg", + "table_caption": [ + "Table 2: Accuracy, model size, and averaged training time per one epoch with respect to different architectures on CIFAR-100. We averaged the training time over 10 epochs" + ], + "table_footnote": [], + "table_body": "
AccuracyModel SizeTraining Time
AvgTop
VGG16-BNw/o Backbone78.968.2164 MB4.10 s
Pick-a-back81.0 (+2.1)76.6 (+8.4)
EfficientNet-B0w/o Backbone77.844.620 MB9.29 s
Pick-a-back81.8 (+4.0)54.4 (+9.8)
MobileNet-V1w/o Backbone65.459.816 MB4.95 s
Pick-a-back66.7 (+1.3)65.4 (+5.6)
LeNet-5w/o Backbone69.167.60.3 MB2.18 s
Pick-a-back71.8 (+2.7)75.6 (+8.0)
", + "bbox": [ + 238, + 184, + 766, + 328 + ], + "page_idx": 11 + }, + { + "type": "table", + "img_path": "images/47948256f40d0203e94fd6d1832fd13d0996d882a571da41471a35c6b4a6ce23.jpg", + "table_caption": [ + "Table 3: Various datasets with LeNet-5" + ], + "table_footnote": [], + "table_body": "
DigitsCIFAR-100Classified-ImageNetFine-Grained
AvgTopAvgTopAvgTopAvgTop
Upper95.889.473.775.661.471.833.948.5
w/o Backbone93.880.769.167.656.159.225.021.0
Pick-a-back95.186.471.875.658.167.628.327.5
(+1.3)(+5.7)(+2.7)(+8.0)(+2.0)(+8.4)(+3.3)(+6.5)
", + "bbox": [ + 223, + 367, + 779, + 450 + ], + "page_idx": 11 + }, + { + "type": "text", + "text": "to the large-scale models of VGG16-BN in Table 2. It indicates that all models benefit from selectively transferred knowledge, leading to accuracy improvement of up to $8.4\\%$ , $9.8\\%$ , $5.6\\%$ , and $8.0\\%$ , respectively. However, in resource-constrained environments, the limitations of memory and processing power make it challenging to deploy a large-scale model at the edge. To make learning practical in the real-world, lightweight models can be flexibly deployed across multiple edge devices, reducing dependence on remote servers. The lightweight LeNet-5 has a much smaller size of 0.18 times compared to VGG16-BN, enabling faster training and inference time. Although the lightweight models may not match closely to the large-scale models, Pick-a-back can still achieve improvement, comparable to the large models, while greatly reducing the storage requirement and run-time.", + "bbox": [ + 212, + 477, + 787, + 645 + ], + "page_idx": 11 + }, + { + "type": "text", + "text": "Scalability across diverse datasets. We assessed the scalability of Pick-a-back across various datasets in Table 3. We stably achieve improvements across different datasets, even with varying degrees of class complexity. In Digits, where most tasks except for the Roman task are generally easy to classify, the average accuracy improvements may appear modest, but Pick-a-back approach achieved a significant accuracy improvement of up to $5.72\\%$ in the relatively challenging Roman task. Similarly, in the classified-ImageNet dataset, where there exist enough neighboring tasks to provide additional knowledge, Pick-a-back shows substantial improvements even for the underperforming tasks. However, when considering fine-grained datasets with a large number of classes per task (ranging $102\\sim 250$ classes), classification becomes inherently more difficult with simple models, resulting in relatively lower accuracy. Nonetheless, by applying Pick", + "bbox": [ + 212, + 657, + 789, + 840 + ], + "page_idx": 11 + }, + { + "type": "page_number", + "text": "12", + "bbox": [ + 217, + 114, + 235, + 126 + ], + "page_idx": 11 + }, + { + "type": "header", + "text": "J. Yoon and H. Lee", + "bbox": [ + 271, + 114, + 403, + 126 + ], + "page_idx": 11 + }, + { + "type": "table", + "img_path": "images/28a0dd5caee43129cd069c7e3d5a9c931f1d797aa28383069aeb081bb9cb8aea.jpg", + "table_caption": [ + "Table 4: Performance on Fine-Grained Table 5: Heterogeneous data sampling datasets using EfficientNet-B0 with $\\alpha = 0.5$" + ], + "table_footnote": [], + "table_body": "
Client ID12345AvgHeteroBalancedw/o Sampling
AvgTopAvgTopAvgTop
Pretrained w/o Backbone28.990.542.449.937.149.8Upper59.360.269.684.473.775.6
66.37.428.212.724.6w/o Backbone52.640.264.664.171.869.167.6
Pick-a-back49.5 (+41.2)77.4 (+11.1)44.0 (+36.6)46.4 (+18.2)64.6 (+51.9)56.4 (+31.8)Pick-a-back56.1 (+3.5)58.6 (+18.4)67.3 (+3.2)80.0 (+8.2)71.8 (+2.7)75.6 (+8.0)
", + "bbox": [ + 220, + 184, + 782, + 247 + ], + "page_idx": 12 + }, + { + "type": "text", + "text": "$a$ -back, it is possible to advance the knowledge and achieve performance gains. While there is a noticeable performance gap of $0.7\\%$ , $1.9\\%$ , and $3.3\\%$ , respectively, between Pick-a-back and Upper in other datasets, the gap in the fine-grained dataset is quite large with $5.6\\%$ on average and up to $21.0\\%$ for the task of flowers. This implies the potential of Pick-a-back to drive further improvements in complex tasks that require more sophisticated classification. We further explore the performance with the different number of clients (i.e., classes) and the different learning seeds in Appendix Table A9 and in Appendix Table A10, respectively.", + "bbox": [ + 212, + 277, + 787, + 398 + ], + "page_idx": 12 + }, + { + "type": "text", + "text": "Impact on the most complicated task. We investigate the most complicated tasks of CUB 55, Stanford Cars 24, Flower 37, WikiArt 48, and Sketch 10, with each featuring a diverse range of class numbers of 200, 196, 102, 195, and 250 and running at each separate client. As shown in Table 4 with the adoption of Pick-a-back, we observe remarkable improvements in overall accuracy, averaging at $31.8\\%$ , and peaking at an impressive $51.9\\%$ . Although these results might appear surprising, they align with the findings previously established by 19 and 29, where scratch accuracy on ImageNet-pretrained models of ResNet50 and VGG16-BN, achieved $58.83\\%$ and $43.84\\%$ , respectively. Incorporating Pick-a-back not only surpasses the average accuracy of ImageNet-pretrained models, but does so through a judicious infusion of tailored knowledge. It is worth noting that ImageNet-pretrained models often necessitate extensive prior training data. However, in many instances, the efficacy diminishes due to the lack of distinct knowledge selection, showing some critical drawbacks in clients #1, #3, and #5.", + "bbox": [ + 212, + 414, + 787, + 627 + ], + "page_idx": 12 + }, + { + "type": "text", + "text": "Superiority in heterogeneous data scenarios. In real-world scenarios, there are many unique challenges such as personalization [5] or heterogeneity [4]. Here, we explored the effectiveness of Pick-a-back under data heterogeneity. Leveraging the heterogeneity settings from federated learning researches [16, 60], we applied $\\alpha$ of 0.5 based on the Dirichlet distribution $Dir(\\alpha)$ . As in Table [5] we observed a degradation in the accuracy of scratch models as the data heterogeneity increases. It implies that the accuracy is achievable, and the external knowledge can be particularly beneficial for weaker classes in the presence of heterogeneous data. While the average accuracy improvement remained consistent, we achieved significant enhancements in heterogeneous data settings. It is essential to address unbalanced data distributions, as certain classes may receive less training attention. By leveraging Pick-a-back, we are able to achieve substantial performance gains, with improvements of up to $18.4\\%$ .", + "bbox": [ + 212, + 643, + 787, + 840 + ], + "page_idx": 12 + }, + { + "type": "header", + "text": "Pick-a-back", + "bbox": [ + 651, + 114, + 730, + 126 + ], + "page_idx": 12 + }, + { + "type": "page_number", + "text": "13", + "bbox": [ + 767, + 114, + 784, + 126 + ], + "page_idx": 12 + }, + { + "type": "table", + "img_path": "images/7defe012871b18bc7b914b931c7ed6e9a09e75a42ab83dc8a59849c1053f5977.jpg", + "table_caption": [ + "Table 6: Performance and data used for training in ImageNet-pretrained VGG16-BN" + ], + "table_footnote": [], + "table_body": "
w/o Backbone → Pick-a-backSize of Training Data
AvgTop
w/ Pretrained84.9 → 83.882.8 → 84.492.8 GB
w/o Pretrained78.9 → 81.068.2 → 76.60.3 GB
", + "bbox": [ + 285, + 183, + 718, + 247 + ], + "page_idx": 13 + }, + { + "type": "text", + "text": "4.3 Discussions and Limitations", + "text_level": 1, + "bbox": [ + 215, + 285, + 491, + 299 + ], + "page_idx": 13 + }, + { + "type": "text", + "text": "ImageNet-pretrained models. It is important to note that a lightweight model such as LeNet-5 may not outperform a large-scale model even with Pick-a-back. Also, VGG16-BN models initialized with ImageNet-pretrained weights demonstrate strong performance generally for visual tasks as in Table 6 and in Appendix Table A6. However, their effectiveness heavily relies on the availability of large-scale labeled datasets that need to be uploaded to a central server. Additionally, the pretrained models are specifically designed for image classification, and their performance may vary when applied to specific or specialized tasks. Furthermore, the availability of pretrained models is limited only to well-known large-scale architectures, making their deployment usually infeasible to edge learning. Without pretrained model, selective knowledge federation through Pick-a-back achieves similar performance on average with a small gap of $3.9\\%$ compared to the original pretrained model (w/o Backbone), but only with $0.3\\%$ of training data.", + "bbox": [ + 212, + 306, + 787, + 503 + ], + "page_idx": 13 + }, + { + "type": "text", + "text": "We further provided the orthogonality of continual learner in Pick-a-back and catastrophic forgetting prevention across different continual learning approaches in Appendix Table A11. Lastly, we need image sharing among clients and model sharing from a selected neighboring client, which may incur transmission overhead and privacy issues. We thus further assess the communication and computation time and how we can defend against attacks in Appendix Table A12 and Fig. A2.", + "bbox": [ + 212, + 503, + 787, + 595 + ], + "page_idx": 13 + }, + { + "type": "text", + "text": "5 Conclusion", + "text_level": 1, + "bbox": [ + 215, + 614, + 359, + 630 + ], + "page_idx": 13 + }, + { + "type": "text", + "text": "We propose Pick-a-back that picks a task-specific backbone knowledge and fosters the collective intelligence in federated continual learning. As the individual learner pursues each distinct goal with private heterogeneous data, we selectively transfer the relevant external knowledge to the target task between neighboring learners.", + "bbox": [ + 212, + 643, + 787, + 703 + ], + "page_idx": 13 + }, + { + "type": "text", + "text": "As AI expands towards edges for scalability, low-end devices tend to be trained toward personalized and specialized but somewhat partially limited intelligence. Here, in case of a scarcity of data or with limited performance, it becomes essential to seek assistance from external sources. The question - how can a device make a local model more effectively trained with exterior knowledge? - can be answered: Pick-a-back offers an effective way to selective knowledge transfer from one device to another towards distributed edge intelligence. By harnessing the power of neighboring devices, individual learners can gain access to valuable insights and expertise via collective intelligence, augmenting their own capabilities.", + "bbox": [ + 212, + 704, + 787, + 840 + ], + "page_idx": 13 + }, + { + "type": "page_number", + "text": "14", + "bbox": [ + 217, + 114, + 235, + 126 + ], + "page_idx": 13 + }, + { + "type": "header", + "text": "J. Yoon and H. Lee", + "bbox": [ + 271, + 114, + 403, + 126 + ], + "page_idx": 13 + }, + { + "type": "text", + "text": "Acknowledgments", + "text_level": 1, + "bbox": [ + 217, + 143, + 392, + 162 + ], + "page_idx": 14 + }, + { + "type": "text", + "text": "We thank the Virginia Tech Presidential Postdoctoral Fellowship Program for supporting JinYi Yoon's postdoctoral associate position. This work was supported by the Institute of Information & communications Technology Planning and Evaluation (IITP) grant funded by the Korea government (MSIT) (No. RS-2022-00155966, Artificial Intelligence Convergence Innovation Human Resources Development (Ewha Womans University) and No. RS-2021-II212068, Artificial Intelligence Innovation Hub).", + "bbox": [ + 212, + 174, + 787, + 282 + ], + "page_idx": 14 + }, + { + "type": "text", + "text": "References", + "text_level": 1, + "bbox": [ + 217, + 301, + 321, + 318 + ], + "page_idx": 14 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "1. Arani, E., et al.: Learning fast, learning slow: A general continual learning method based on complementary learning system. ICLR (2022)", + "2. Cao, L.: Decentralized ai: Edge intelligence and smart blockchain, metaverse, web3, and desci. IEEE Intelligent Systems 37(3), 6-19 (2022)", + "3. Casado, F.E., Lema, D., Criado, M.F., Iglesias, R., Regueiro, C.V., Barro, S.: Concept drift detection and adaptation for federated and continual learning. Multimedia Tools and Applications pp. 1-23 (2022)", + "4. Chen, D., Gao, D., Kuang, W., Li, Y., Ding, B.: pfl-bench: A comprehensive benchmark for personalized federated learning. Advances in Neural Information Processing Systems 35, 9344-9360 (2022)", + "5. Chen, D., Gao, D., Xie, Y., Pan, X., Li, Z., Li, Y., Ding, B., Zhou, J.: Fs-real: Towards real-world cross-device federated learning. In: Proceedings of the 29th ACM SIGKDD Conference on Knowledge Discovery and Data Mining. pp. 3829-3841 (2023)", + "6. Chen, Y., Inaltekin, H., Gorlatova, M.: Adaptslam: Edge-assisted adaptive slam with resource constraints via uncertainty minimization. arXiv preprint arXiv:2301.04620 (2023)", + "7. Colianni, S.: Mnist as .jpg (2017), https://www.kaggle.com/scolianni/ mnistas.jpg", + "8. Deng, J., Dong, W., Socher, R., Li, L.J., Li, K., Fei-Fei, L.: Imagenet: A large-scale hierarchical image database. In: 2009 IEEE Conference on Computer Vision and Pattern Recognition. pp. 248-255 (2009). https://doi.org/10.1109/CVPR.2009.5206848", + "9. Dhar, P., Singh, R.V., Peng, K.C., Wu, Z., Chellappa, R.: Learning without memorizing. In: Proceedings of the IEEE/CVF conference on computer vision and pattern recognition. pp. 5138-5146 (2019)", + "10. Eitz, M., Hays, J., Alexa, M.: How do humans sketch objects? ACM Transactions on graphics (TOG) 31(4), 1-10 (2012)", + "11. El-Sawy, A., El-Bakry, H., Loey, M.: Cnn for handwritten arabic digits recognition based on lenet-5. In: Proceedings of the International Conference on Advanced Intelligent Systems and Informatics 2016 2. pp. 566-575. Springer (2017)", + "12. Goodfellow, I., Bengio, Y., Courville, A.: Deep Learning. MIT Press (2016), http://www.deeplearningbook.org", + "13. Gu, Y., Yang, X., Wei, K., Deng, C.: Not just selection, but exploration: Online class-incremental continual learning via dual view consistency. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 7442-7451 (2022)" + ], + "bbox": [ + 218, + 330, + 787, + 839 + ], + "page_idx": 14 + }, + { + "type": "header", + "text": "Pick-a-back", + "bbox": [ + 651, + 114, + 730, + 126 + ], + "page_idx": 14 + }, + { + "type": "page_number", + "text": "15", + "bbox": [ + 767, + 116, + 785, + 126 + ], + "page_idx": 14 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "14. Ha, D., Tang, Y.: Collective intelligence for deep learning: A survey of recent developments. Collective Intelligence 1(1), 263391372211114874 (2022)", + "15. Haim, N., Vardi, G., Yehudai, G., Shamir, O., Irani, M.: Reconstructing training data from trained neural networks. Advances in Neural Information Processing Systems 35, 22911-22924 (2022)", + "16. He, C., Li, S., So, J., Zeng, X., Zhang, M., Wang, H., Wang, X., Vepakomma, P., Singh, A., Qiu, H., et al.: Fedml: A research library and benchmark for federated machine learning. arXiv preprint arXiv:2007.13518 (2020)", + "17. Howard, A.G., Zhu, M., Chen, B., Kalenichenko, D., Wang, W., Weyand, T., Andreetto, M., Adam, H.: Mobilenets: Efficient convolutional neural networks for mobile vision applications. arXiv preprint arXiv:1704.04861 (2017)", + "18. Huang, W., Ye, M., Du, B.: Learn from others and be yourself in heterogeneous federated learning. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 10143-10153 (2022)", + "19. Hung, C.Y., Tu, C.H., Wu, C.E., Chen, C.H., Chan, Y.M., Chen, C.S.: Compacting, picking and growing for unforgetting continual learning. Advances in Neural Information Processing Systems 32 (2019)", + "20. Jeong, W., Hwang, S.J.: Factorized-fl: Personalized federated learning with parameter factorization & similarity matching. Advances in Neural Information Processing Systems 35, 35684-35695 (2022)", + "21. Kang, H., Mina, R.J.L., Madjid, S.R.H., Yoon, J., Hasegawa-Johnson, M., Hwang, S.J., Yoo, C.D.: Forget-free continual learning with winning subnetworks. In: International Conference on Machine Learning. pp. 10734-10750. PMLR (2022)", + "22. Kim, G., Esmaeilpour, S., Xiao, C., Liu, B.: Continual learning based on ood detection and task masking. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 3856-3866 (2022)", + "23. Kirkpatrick, J., Pascanu, R., Rabinowitz, N., Veness, J., Desjardins, G., Rusu, A.A., Milan, K., Quan, J., Ramalho, T., Grabska-Barwinska, A., et al.: Overcoming catastrophic forgetting in neural networks. Proceedings of the national academy of sciences 114(13), 3521-3526 (2017)", + "24. Krause, J., Stark, M., Deng, J., Fei-Fei, L.: 3d object representations for fine-grained categorization. In: Proceedings of the IEEE international conference on computer vision workshops. pp. 554-561 (2013)", + "25. Krizhevsky, A., Hinton, G., et al.: Learning multiple layers of features from tiny images (2009)", + "26. LeCun, Y., et al.: Lenet-5, convolutional neural networks. URL: http://yann.lecun.com/exdb/lenet 20(5), 14 (2015)", + "27. Li, Y., Zhang, Z., Liu, B., Yang, Z., Liu, Y.: Modeldiff: testing-based dnn similarity comparison for model reuse detection. In: Proceedings of the 30th ACM SIGSOFT International Symposium on Software Testing and Analysis. pp. 139-151 (2021)", + "28. Liu, F., Wu, X., Ge, S., Fan, W., Zou, Y.: Federated learning for vision-and-language grounding problems. In: Proceedings of the AAAI Conference on Artificial Intelligence. vol. 34, pp. 11572-11579 (2020)", + "29. Mallya, A., Davis, D., Lazebnik, S.: Piggyback: Adapting a single network to multiple tasks by learning to mask weights. In: Proceedings of the European Conference on Computer Vision (ECCV). pp. 67-82 (2018)", + "30. Mallya, A., Lazebnik, S.: Packnet: Adding multiple tasks to a single network by iterative pruning. In: Proceedings of the IEEE conference on Computer Vision and Pattern Recognition. pp. 7765-7773 (2018)", + "31. Mao, B., Liu, J., Wu, Y., Kato, N.: Security and privacy on 6g network edge: A survey. IEEE Communications Surveys & Tutorials (2023)" + ], + "bbox": [ + 215, + 146, + 785, + 839 + ], + "page_idx": 15 + }, + { + "type": "page_number", + "text": "16", + "bbox": [ + 217, + 114, + 235, + 126 + ], + "page_idx": 15 + }, + { + "type": "header", + "text": "J. Yoon and H. Lee", + "bbox": [ + 271, + 114, + 403, + 126 + ], + "page_idx": 15 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "32. McMahan, B., Moore, E., Ramage, D., Hampson, S., y Arcas, B.A.: Communication-efficient learning of deep networks from decentralized data. In: Artificial intelligence and statistics. pp. 1273-1282. PMLR (2017)", + "33. Meng, C., Tang, M., Setayesh, M., Wong, V.W.: Gnn-based neighbor selection and resource allocation for decentralized federated learning. In: GLOBECOM 2023-2023 IEEE Global Communications Conference. pp. 1223-1228. IEEE (2023)", + "34. Miller, G.A.: Wordnet: a lexical database for english. Communications of the ACM 38(11), 39-41 (1995)", + "35. Nazarpour, K., Chen, M.: Handwritten Chinese Numbers (1 2017). https://doi.org/10.17634/137930-3, https://data.ncl.ac.uk/articles/dataset/Handwritten_Chinese_Numbers/10280831", + "36. Netzer, Y., Wang, T., Coates, A., Bissacco, A., Wu, B., Ng, A.Y.: Reading digits in natural images with unsupervised feature learning (2011)", + "37. Nilsback, M.E., Zisserman, A.: Automated flower classification over a large number of classes. In: 2008 Sixth Indian Conference on Computer Vision, Graphics & Image Processing. pp. 722-729. IEEE (2008)", + "38. Nishio, T., Yonetani, R.: Client selection for federated learning with heterogeneous resources in mobile edge. In: ICC 2019-2019 IEEE international conference on communications (ICC). pp. 1-7. IEEE (2019)", + "39. Noyan, M.A.: Hand written roman numerals dataset (Sep 2021). https://doi.org/10.5281/zenodo.5385144, https://doi.org/10.5281/zenodo.5385144", + "40. Ostapenko, O., Puscas, M., Klein, T., Jahnichen, P., Nabi, M.: Learning to remember: A synaptic plasticity driven framework for continual learning. In: Proceedings of the IEEE/CVF conference on computer vision and pattern recognition. pp. 11321-11329 (2019)", + "41. Parks Associates: At CES 2024, Parks Associates Announces New Research Showing Average Number of Connected Devices per US Internet Household Reached 17 in 2023. https://www.parksassociates.com/blogs/press-releases/at-ces-2024-parks-associates-announces-new-research-showing-average-number-of-connected-devices-per-us-internet-household-reached-17-in-2023 (2024), accessed: 2024-05-15", + "42. PyTorch: Retrieved from python model zoo (2017), https://pytorch.org/vision/stable/models.html", + "43. Radford, A., Kim, J.W., Hallacy, C., Ramesh, A., Goh, G., Agarwal, S., Sastry, G., Askell, A., Mishkin, P., Clark, J., et al.: Learning transferable visual models from natural language supervision. In: International conference on machine learning. pp. 8748-8763. PMLR (2021)", + "44. Roy, A.G., Siddiqui, S., Pölsterl, S., Navab, N., Wachinger, C.: Brainstorm: A peer-to-peer environment for decentralized federated learning. arXiv preprint arXiv:1905.06731 (2019)", + "45. Roy, P., Ghosh, S., Bhattacharya, S., Pal, U.: Effects of degradations on deep neural network architectures. arXiv preprint arXiv:1807.10108 (2018)", + "46. Rusu, A.A., Rabinowitz, N.C., Desjardins, G., Soyer, H., Kirkpatrick, J., Kavukcuoglu, K., Pascanu, R., Hadsell, R.: Progressive neural networks. arXiv preprint arXiv:1606.04671 (2016)", + "47. Rusu, A.A., Rabinowitz, N.C., Desjardins, G., Soyer, H., Kirkpatrick, J., Kavukcuoglu, K., Pascanu, R., Hadsell, R.: Progressive neural networks. arXiv preprint arXiv:1606.04671 (2016)", + "48. Saleh, B., Elgammal, A.: Large-scale classification of fine-art paintings: Learning the right metric on the right feature. arXiv preprint arXiv:1505.00855 (2015)" + ], + "bbox": [ + 212, + 146, + 787, + 840 + ], + "page_idx": 16 + }, + { + "type": "header", + "text": "Pick-a-back", + "bbox": [ + 651, + 114, + 730, + 126 + ], + "page_idx": 16 + }, + { + "type": "page_number", + "text": "17", + "bbox": [ + 767, + 114, + 785, + 126 + ], + "page_idx": 16 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "49. Shayan, M., Fung, C., Yoon, C.J., Beschastnikh, I.: Biscotti: A blockchain system for private and secure federated learning. IEEE Transactions on Parallel and Distributed Systems 32(7), 1513-1525 (2020)", + "50. Shenaj, D., Toldo, M., Rigon, A., Zanuttigh, P.: Asynchronous federated continual learning. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 5054-5062 (2023)", + "51. Simonyan, K., Zisserman, A.: Very deep convolutional networks for large-scale image recognition. arXiv preprint arXiv:1409.1556 (2014)", + "52. Stoychev, S., Churamani, N., Gunes, H.: Latent generative replay for resource-efficient continual learning of facial expressions. In: 2023 IEEE 17th International Conference on Automatic Face and Gesture Recognition (FG). pp. 1-8. IEEE (2023)", + "53. Sun, W., Yang, Z., Zhang, X., Liu, Y.: Energy-efficient neighbor discovery in mobile ad hoc and wireless sensor networks: A survey. IEEE Communications Surveys & Tutorials 16(3), 1448-1459 (2014)", + "54. Tan, M., Le, Q.: Efficientnet: Rethinking model scaling for convolutional neural networks. In: International conference on machine learning. pp. 6105-6114. PMLR (2019)", + "55. Wah, C., Branson, S., Welinder, P., Perona, P., Belongie, S.: The caltech-ucsd birds-200-2011 dataset (2011)", + "56. Wei, K., Li, J., Ding, M., Ma, C., Yang, H.H., Farokhi, F., Jin, S., Quek, T.Q., Poor, H.V.: Federated learning with differential privacy: Algorithms and performance analysis. IEEE transactions on information forensics and security 15, 3454-3469 (2020)", + "57. Yoon, J., Jeong, W., Lee, G., Yang, E., Hwang, S.J.: Federated continual learning with weighted inter-client transfer. In: International Conference on Machine Learning. pp. 12073-12086. PMLR (2021)", + "58. Yoon, J., Yang, E., Lee, J., Hwang, S.J.: Lifelong learning with dynamically expandable networks. arXiv preprint arXiv:1708.01547 (2017)", + "59. Yu, Q., Liu, Y., Wang, Y., Xu, K., Liu, J.: Multimodal federated learning via contrastive representation ensemble. arXiv preprint arXiv:2302.08888 (2023)", + "60. Yurochkin, M., Agarwal, M., Ghosh, S., Greenewald, K., Hoang, N., Khazaeni, Y. Bayesian nonparametric federated learning of neural networks. In: International conference on machine learning. pp. 7252-7261. PMLR (2019)", + "61. Zhang, J., Chen, C., Zhuang, W., Lyu, L.: Target: Federated class-continual learning via exemplar-free distillation. In: Proceedings of the IEEE/CVF International Conference on Computer Vision. pp. 4782-4793 (2023)", + "62. Zhang, T., Mai, C., Chang, Y., Chen, C., Shu, L., Zheng, Z.: Fedego: privacy-preserving personalized federated graph learning with ego-graphs. ACM Transactions on Knowledge Discovery from Data 18(2), 1-27 (2023)", + "63. Zhu, L., Liu, Z., Han, S.: Deep leakage from gradients. Advances in neural information processing systems 32 (2019)" + ], + "bbox": [ + 215, + 147, + 785, + 715 + ], + "page_idx": 17 + }, + { + "type": "page_number", + "text": "18", + "bbox": [ + 217, + 114, + 235, + 126 + ], + "page_idx": 17 + }, + { + "type": "header", + "text": "J. Yoon and H. Lee", + "bbox": [ + 271, + 114, + 403, + 126 + ], + "page_idx": 17 + } +] \ No newline at end of file diff --git a/2024/Pick-a-back_ Selective Device-to-Device Knowledge Transfer in Federated Continual Learning/ecda4e75-a956-4af4-a55c-23aa50728c18_model.json b/2024/Pick-a-back_ Selective Device-to-Device Knowledge Transfer in Federated Continual Learning/ecda4e75-a956-4af4-a55c-23aa50728c18_model.json new file mode 100644 index 0000000000000000000000000000000000000000..b3aa17a8e4261cecacfa207b8707371c7ec837d3 --- /dev/null +++ b/2024/Pick-a-back_ Selective Device-to-Device Knowledge Transfer in Federated Continual Learning/ecda4e75-a956-4af4-a55c-23aa50728c18_model.json @@ -0,0 +1,2271 @@ +[ + [ + { + "type": "title", + "bbox": [ + 0.218, + 0.141, + 0.787, + 0.187 + ], + "angle": 0, + "content": "Pick-a-back: Selective Device-to-Device Knowledge Transfer in Federated Continual Learning" + }, + { + "type": "text", + "bbox": [ + 0.362, + 0.212, + 0.639, + 0.23 + ], + "angle": 0, + "content": "JinYi Yoon1 and HyungJune Lee2" + }, + { + "type": "text", + "bbox": [ + 0.34, + 0.239, + 0.664, + 0.296 + ], + "angle": 0, + "content": "\\(^{1}\\) Virginia Tech, Blacksburg, VA, USA jinyioon@vt.edu \n\\(^{2}\\) Ewha Womans University, Seoul, South Korea hyungjune.lee@ewha.ac.kr" + }, + { + "type": "text", + "bbox": [ + 0.262, + 0.331, + 0.744, + 0.54 + ], + "angle": 0, + "content": "Abstract. With the explosion of edge intelligence, leveraging federated indirect knowledge has become crucial for boosting the tasks of individual learners. However, the conventional approach to knowledge reuse often leads to catastrophic forgetting issues. In this paper, we revisit the concept of continual learning in the context of edge intelligence and address the knowledge transfer problem to enhance federated continual learning. Since each learner processes private heterogeneous data, we propose Pick-a-back, a device-to-device knowledge federation framework by selectively reusing the external knowledge with similar behavioral patterns. By borrowing indirect experiences, an edge device can initiate learning from useful knowledge and thus achieve faster yet more generalized knowledge acquisition. Using continual tasks consisting of various datasets on lightweight architectures, we have validated that Pick-a-back provides a significant inference improvement of up to \\(8.0\\%\\) via selective knowledge federation. Our codes are available at https://github.com/jinyi-yoon/Pick-a-back.git" + }, + { + "type": "text", + "bbox": [ + 0.261, + 0.552, + 0.744, + 0.581 + ], + "angle": 0, + "content": "Keywords: Federated continual learning, edge intelligence, device-to-device knowledge federation, selective knowledge transfer" + }, + { + "type": "title", + "bbox": [ + 0.216, + 0.605, + 0.377, + 0.621 + ], + "angle": 0, + "content": "1 Introduction" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.636, + 0.788, + 0.818 + ], + "angle": 0, + "content": "As artificial intelligence (AI) permeates nearly every field, data has emerged as a pivotal source driving remarkable advancements. Most of the data originates from the end users or systems, stimulating a need for distributed learning on low-end edge devices to address privacy and latency concerns [2,31]. However, individual low-end edge devices often lack the computational capability to generate advanced knowledge [6]. In the context of human learning paradigms, where any individuals become educated and exchange helpful knowledge with one another, the emergence of edge intelligence and its widespread availability allow external knowledge to serve as a valuable resource in supporting individual learners. By facilitating some sharable insights [28,59] and harnessing collective intelligence [14], it has widely revolutionized learning paradigms across various fields, even where numerous data have non-identical forms and heterogeneous distributions." + }, + { + "type": "page_footnote", + "bbox": [ + 0.218, + 0.825, + 0.383, + 0.84 + ], + "angle": 0, + "content": "* Corresponding author" + } + ], + [ + { + "type": "page_number", + "bbox": [ + 0.218, + 0.116, + 0.23, + 0.127 + ], + "angle": 0, + "content": "2" + }, + { + "type": "header", + "bbox": [ + 0.272, + 0.115, + 0.404, + 0.128 + ], + "angle": 0, + "content": "J. Yoon and H. Lee" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.147, + 0.788, + 0.343 + ], + "angle": 0, + "content": "While leveraging external knowledge to enhance individual tasks, conventional learners often encounter the challenge of catastrophic forgetting, wherein they adapt only to the assigned tasks, and consequently lose previously acquired knowledge in some degree over time. In contrast, to more effectively retain and reinforce pre-trained knowledge together with new knowledge, the learners would rather adopt a continual learning approach for ensuring both robustness and adaptability. To enhance the efficacy of individual continual learning, it is not imperative for learners to confine themselves to their own knowledge. The prevalence of edge intelligence may facilitate knowledge federation, enabling individual learners to leverage a broader or deeper understanding of data, which can be possibly heterogeneous. While recent works have devised federated continual learning frameworks, they typically take a centralized approach for assistance [3,18,50], rather than direct device-to-device knowledge transfer in a distributed manner." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.348, + 0.788, + 0.56 + ], + "angle": 0, + "content": "As for edge AI, individual learners might be confronted with a series of diverse heterogeneous tasks. At the edge side accompanied with its own private data, each learner attempts to form its own distinct knowledge. Then, a question to raise is that how can an individual edge learner effectively harness externally available prior knowledge? It has been reported that there exists unique positive and negative knowledge at each model [43,57]. Maybe the answer would be straightforward yet meaningful – the learners would selectively reuse or customize some knowledge that aligns well with their specific behavioral patterns. Rather than starting from scratch, learners can benefit from the indirectly attained knowledge formed by others who have already seen similar patterns. In that the decision patterns can not be exactly the same, indirect knowledge transfer (with follow-up training) may offer a rich source of otherwise untrainable and unattainable knowledge from private data. This can lead to performance improvement via selective acquisition of some common yet distinct knowledge." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.564, + 0.788, + 0.669 + ], + "angle": 0, + "content": "In this work, we revisit the knowledge reuse problem. We consider three challenges in edge intelligence with knowledge transfer: 1) edge devices have limited computational and storage resources; 2) learners are not willing to disclose their raw data; and 3) edge devices should take into account the reduction of communication costs. To overcome these concerns, our approach aims to enhance the performance of lightweight learning models, while maintaining data privacy and avoiding additional communication overhead." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.674, + 0.788, + 0.841 + ], + "angle": 0, + "content": "We propose a selective and customized knowledge federation framework from one device to another, based on heterogeneous continual learning, called Pick-a-back, which picks a relevant backbone knowledge from other devices. Beyond the achievable knowledge of each individual learner, we strategically search for some effective knowledge sources selectively and then customize the commonly shared knowledge to closely align with the local task. By leveraging decision similarity based on mutated data to assess the efficacy of the backbone among the neighboring models spread in the form of physically separated edge intelligence, we selectively fetch and customize constructive knowledge. This approach not only enables decentralized knowledge exchange among the learners at devices, but also facilitates a more efficient and scalable learning paradigm." + } + ], + [ + { + "type": "header", + "bbox": [ + 0.652, + 0.115, + 0.732, + 0.127 + ], + "angle": 0, + "content": "Pick-a-back" + }, + { + "type": "page_number", + "bbox": [ + 0.775, + 0.117, + 0.785, + 0.127 + ], + "angle": 0, + "content": "3" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.147, + 0.788, + 0.208 + ], + "angle": 0, + "content": "This work contributes to advancing the field of deep learning in edge intelligence by providing selective knowledge transfer and fostering collaboration among local learners. It can thereby enhance the efficiency of the individual continual learning process. The main contributions can be summarized as follows:" + }, + { + "type": "text", + "bbox": [ + 0.221, + 0.216, + 0.787, + 0.29 + ], + "angle": 0, + "content": "1) We address the problem of device-to-device knowledge federation in the continual learning. We demonstrate the potential of leveraging indirect knowledge extracted by the neighboring edge learners to obtain otherwise untrainable and unattainable insights. This work provides a way to maintain a model with heterogeneous data collected across a network of edge devices." + }, + { + "type": "text", + "bbox": [ + 0.221, + 0.291, + 0.788, + 0.381 + ], + "angle": 0, + "content": "2) As the diverse tasks imply their own unique knowledge, we propose selective and customized knowledge transfer by capturing the decision pattern similarity relevant to a target task running on a device. Pick-a-back adopts each relevant knowledge for a specific task, achieving high fidelity through the acquisition of some common yet distinct knowledge from unseen examples, while preserving data privacy and communication cost at the edge." + }, + { + "type": "text", + "bbox": [ + 0.221, + 0.381, + 0.788, + 0.426 + ], + "angle": 0, + "content": "3) Extensive experimental results on various datasets across different architectures have shown that Pick-a-back provides a significant inference improvement up to \\(8.0\\%\\) via selective knowledge federation." + }, + { + "type": "list", + "bbox": [ + 0.221, + 0.216, + 0.788, + 0.426 + ], + "angle": 0, + "content": null + }, + { + "type": "title", + "bbox": [ + 0.216, + 0.449, + 0.388, + 0.465 + ], + "angle": 0, + "content": "2 Related Work" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.479, + 0.788, + 0.543 + ], + "angle": 0, + "content": "Typical continual learning refers to training on a series of multiple tasks [22] or incremental classes [13]. The primary objective is to address the catastrophic forgetting problem, where previously learned knowledge is decaying over time as adapting to new tasks, using memorization [40,52] or regularization [9,23]." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.553, + 0.789, + 0.844 + ], + "angle": 0, + "content": "Knowledge-reusable continual learning. More closely related to our problem, there has been a growing interest in reusing existing knowledge for its subsequent tasks [30, 58]. Unlike conventional approaches, these methods have focused on concatenating a new model to its previous model, allowing input data to propagate through the previous model [46]. Rather than reusing the whole model, researchers have proposed to activate only the essential parts of the model [19,29]. In *Piggyback* [29], they leverage a common model and mask out unnecessary neurons to extract relevant knowledge for a target task. It effectively reuses existing knowledge with only a single sharable model, but it is less able to accept incremental tasks. *CPG* [19] combines neuron expansion with neuron pruning and masking techniques to address the challenges of forgetting, model compactness, and achieving higher fidelity in knowledge reuse. We adopt these knowledge reusable continual learning approaches to retain and extract valuable insights from previous tasks, while aiming to find useful knowledge (for reuse) based on model similarity. Although some works have applied some similarity metrics such as cosine similarity to federated learning scenarios [20], it is limited in applicability when dealing with clients having a different number of outputs (classes). In contrast, our approach provides flexibility in output heterogeneity, ensuring broader applicability in real-world scenarios, where most clients may" + } + ], + [ + { + "type": "page_number", + "bbox": [ + 0.218, + 0.116, + 0.232, + 0.127 + ], + "angle": 0, + "content": "4" + }, + { + "type": "header", + "bbox": [ + 0.272, + 0.115, + 0.404, + 0.128 + ], + "angle": 0, + "content": "J. Yoon and H. Lee" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.147, + 0.788, + 0.224 + ], + "angle": 0, + "content": "have different shapes of outputs. Moreover, our decision pattern-based similarity is novel in federated continual learning, as the knowledge itself is not exactly the same even with the similar decision patterns. This enables private learners to be exposed to similar yet unseen broader knowledge, leading to improved performance and efficiency in subsequent learning tasks." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.244, + 0.79, + 0.607 + ], + "angle": 0, + "content": "Federated continual learning. With widespread accessibility of edge AI, learners do not necessarily be confined to their own private knowledge, as a standalone learner is permitted to only access its own local information and thus suffer from restricted experiences to some extent. Although recent works tackle the problem of federated continual learning [50,61], they focus on class-incremental learning, whereas we tackle task-incremental learning. In task-incremental learning, transfer learning turns out to be very effective, where a model is pretrained based on the source domain and then finetuned for the target domain. This selective device-to-device knowledge transfer can be a significant impetus to enhance model fidelity, also in continual learning. It has been commonly addressed in the field of federated continual learning but only in a centralized manner [18,50]. It has been also observed that there is both positive and negative knowledge related to a specific target task [43,57], highlighting the potential harm of static knowledge selection. To address the two-sided nature of knowledge reuse, FedWeIT [57] aggregates the parameters from clients as global knowledge, and then the clients mask the federated parameters adaptive to the specific task. However, previous works including FedWeIT rely on a central server, causing communication bottlenecks and privacy limitations. Furthermore, ImageNet-pretrained models have widely been utilized in transfer learning. However, it is not feasible for edge-side federated continual learning, as pretrained models are specifically designed for well-known large-scale architectures. The architectures in continual learning are dynamically changing, and edge AI cannot be tailored to large-scale models. In Pick-a-back, we rather explore decentralized, communication-efficient, and privacy-preserving knowledge transfer in lightweight on-device models." + }, + { + "type": "title", + "bbox": [ + 0.216, + 0.636, + 0.538, + 0.654 + ], + "angle": 0, + "content": "3 Selective Knowledge Transfer" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.674, + 0.79, + 0.841 + ], + "angle": 0, + "content": "We consider the problem of device-to-device knowledge federation to enhance individual continual learning in a distributed manner. We propose Pick-a-back, a simple yet effective knowledge federation approach by discovering a relevant neighboring device who has acquired the knowledge of relevant tasks and then selectively infusing it into the model via continual learning. By leveraging a federated learning framework, which allows collaborative learning across devices, we aim to address the challenge of distributed continual learning, while also facilitating the reuse of previously acquired knowledge at each different device. The integration of these two fields of study enables efficient knowledge selection, sharing, and expansion among devices (with heterogeneous data), yielding continuous learning improvement and adaptation across tasks or learning scenarios." + } + ], + [ + { + "type": "header", + "bbox": [ + 0.635, + 0.06, + 0.657, + 0.077 + ], + "angle": 0, + "content": "#" + }, + { + "type": "header", + "bbox": [ + 0.66, + 0.063, + 0.801, + 0.073 + ], + "angle": 0, + "content": "EWHA WOMANS UNIVERSITY" + }, + { + "type": "header", + "bbox": [ + 0.652, + 0.115, + 0.732, + 0.127 + ], + "angle": 0, + "content": "Pick-a-back" + }, + { + "type": "page_number", + "bbox": [ + 0.775, + 0.116, + 0.785, + 0.127 + ], + "angle": 0, + "content": "5" + }, + { + "type": "image", + "bbox": [ + 0.249, + 0.146, + 0.756, + 0.306 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.214, + 0.318, + 0.788, + 0.429 + ], + "angle": 0, + "content": "Fig. 1: Overall architecture of Pick-a-back: (1) Each device trains an individual continual learning model on a series of tasks. (2) Given these available neighboring knowledge, a device selects a relevant neighboring model that can improve learning performance by acquiring indirect knowledge from the federated but physically separated devices. (3) On top of the retrieved knowledge, (4) the continual learner stacks new local tasks and resumes learning and federation. By fetching the generalized knowledge with relevant decision patterns, Pick-a-back offers a cumulative yet selective knowledge federation with heterogeneous data at a distinct device" + }, + { + "type": "title", + "bbox": [ + 0.216, + 0.461, + 0.439, + 0.475 + ], + "angle": 0, + "content": "3.1 Problem Formulation" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.492, + 0.789, + 0.631 + ], + "angle": 0, + "content": "Given \\( L \\) devices, each learner \\( l \\in \\{1, 2, \\dots, L\\} \\) at device \\( d_l \\) processes its own sequential tasks (i.e., training data) of \\( T^{(l)} = \\{T_1^{(l)}, T_2^{(l)}, \\dots \\} \\). A scratch model \\( M_i^{(l)} \\) is constructed by training only with private data \\( T_i^{(l)} = \\{(X_i^{(l)}, Y_i^{(l)})\\} \\), where \\( T_i^{(l)} \\) for task \\( i \\) and device \\( d_l \\) is heterogeneous across other tasks and devices. We train the model \\( \\hat{M}_i^{(l)} \\) with target data \\( T_i^{(l)} \\) on top of its previous version \\( \\hat{M}_{i-1}^{(l)} \\) with some learning parameters \\( \\theta_i^{(l)} \\), i.e., \\( \\hat{M}_i^{(l)} = \\hat{M}_{i-1}^{(l)}(T_i^{(l)}, \\theta_i^{(l)}) \\) in a continual manner. Here, at every task \\( i \\), the goal is to find a good reference model that can maximize the overall fidelity of model \\( \\hat{M}_i^{(l)} \\)." + }, + { + "type": "title", + "bbox": [ + 0.216, + 0.658, + 0.368, + 0.672 + ], + "angle": 0, + "content": "3.2 Architecture" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.689, + 0.789, + 0.841 + ], + "angle": 0, + "content": "Our learning architecture incorporates two distinct modes (Fig. 1): i) continual learner for local training; and ii) federated learner for knowledge transfer. Initially, multiple learners train independently on a specific private task. Subsequently, one learner at a device transitions to a federated learning mode to explore effective indirect knowledge from a neighboring relevant model. Once a neighboring model is selected, we use the selected model as a backbone, the main target task is then executed in a continual manner. By incorporating the retrieved knowledge together with its existing own knowledge, the learner can construct cumulative knowledge in a continual manner without forgetting. This process of continual learning and federated learning is repeated for each task." + } + ], + [ + { + "type": "page_number", + "bbox": [ + 0.218, + 0.116, + 0.23, + 0.127 + ], + "angle": 0, + "content": "6" + }, + { + "type": "header", + "bbox": [ + 0.272, + 0.115, + 0.404, + 0.128 + ], + "angle": 0, + "content": "J. Yoon and H. Lee" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.147, + 0.788, + 0.358 + ], + "angle": 0, + "content": "Continual learner. In our approach to continual learning, we adopt task-incremental learning, where a series of tasks is processed sequentially. Each task has a distinct learning objective, and a local model is trained independently on each task with freezing or isolating the models for previous knowledge. By employing task-oriented learning, we aim to mitigate the catastrophic forgetting problem, where previously learned knowledge vanishes as training on new tasks or incorporating new data over time. Our approach can be combined with any continual learner that has gradually been constructing knowledge. It is fundamentally different from conventional continual learning that simply memorizes multiple models or data, which do not leverage the given knowledge. We suppose that a continual learner concatenates a new model on top of the previously acquired knowledge, and thus it can reduce memory usage and also enhance the model fidelity. This sustainability of the continual model is particularly valuable in the context of edge AI and aligns with the goals of the recent learning paradigm." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.359, + 0.788, + 0.51 + ], + "angle": 0, + "content": "We pursue reusing the pre-configured knowledge by using the model. We mainly use the latest continual learning effort, \\( CPG \\) [19], which expands the model when necessary, prunes it for compactness, and reuses the network by masking out specific ineffective neurons. Inspired by the masking strategy of Piggyback [29], it samples some necessary neurons for the target task to allow the extraction of common knowledge. \\( CPG \\) only trains the learnable binary mask whether to activate each neuron for the target task. As it ensures the retention of previous knowledge via freezing, the results remain exactly the same even when training new tasks. However, without loss of generality, this masking strategy can be applied to any continual learner that holds multiple knowledge." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.523, + 0.788, + 0.794 + ], + "angle": 0, + "content": "Federated learner. Due to the privacy and communication overhead, distributed learners are unwilling to share private raw data. We thus adopt a model-sharing approach rather than data-sharing. Here, each client is regarded as an independent learner. When each learner at a distinct device is working on a distinct task but lacks knowledge, there might be some other knowledge that has a similar, broader, or deeper experience beyond the learner based on its own local data. We assume that clients are knowledgeable with their own currently available neighboring models via existing neighbor selection methods [38, 53] or graph construction approaches [33, 62]. These neighboring diverse knowledge sources have been trained on different datasets, capturing various perspectives and feature extraction capabilities. By exposing the learner to external knowledge, the model is capable of leveraging enhanced feature representations, enabling it to capture a broader set of patterns and information that may be beneficial for its specific learning objective. The learner can also develop a more comprehensive understanding of the data, generalize better to unseen examples by mitigating overfitting, and thus improve the fidelity of the model on a variety of tasks and data. By leveraging a diverse ensemble of knowledge sources, the learner can benefit from its complementary expertise, leading to accuracy improvement." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.796, + 0.788, + 0.84 + ], + "angle": 0, + "content": "Each learner at a device seeks relevant knowledge that can contribute to further improving its performance based on local continual learning. To acquire necessary constructive knowledge, we identify effective indirect experiences based" + } + ], + [ + { + "type": "header", + "bbox": [ + 0.652, + 0.115, + 0.732, + 0.127 + ], + "angle": 0, + "content": "Pick-a-back" + }, + { + "type": "page_number", + "bbox": [ + 0.775, + 0.116, + 0.785, + 0.126 + ], + "angle": 0, + "content": "7" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.147, + 0.788, + 0.269 + ], + "angle": 0, + "content": "on decision pattern similarities. Once we have obtained knowledge from other learners, we accumulate ongoing new local tasks on top of the model in a continual manner, ensuring that the retrieved knowledge becomes infused with local knowledge. Our main focus is on the current target task, and thus, if we want to retain the knowledge from our previous tasks, we have to either store the existing knowledge separately or concatenate it with the fetched model. If the currently constructed knowledge turns out to be sufficiently effective, there may be no further need to incorporate further knowledge from other learners." + }, + { + "type": "title", + "bbox": [ + 0.215, + 0.289, + 0.525, + 0.304 + ], + "angle": 0, + "content": "3.3 Selective Knowledge Federation" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.312, + 0.789, + 0.493 + ], + "angle": 0, + "content": "Each learner is responsible for distinct tasks, and tasks may exhibit unique strengths, weaknesses, or biases. Consequently, each task at a local learner plays its own unique different role and forms distinct knowledge. Some external models may have learned other representations that have captured specialized knowledge relevant to specific tasks. Therefore, selective knowledge transfer can be a very effective way for enhancing the performance of each student model at an edge device. Our approach is simple and straightforward – if models have similar or synergetic capabilities in classifying different samples, they are likely to have potential sources of common but more generalized knowledge. Since it has the possibility to attain broader intuition but also inexperienced knowledge, this indirect knowledge characterized by similar decision patterns may greatly accelerate and improve the learning for individual learners." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.494, + 0.789, + 0.75 + ], + "angle": 0, + "content": "As the similarity of model do not straightly incur similar knowledge, we focus on the decision pattern similarity rather than the model itself. We borrow an idea to calculate the model similarity from ModelDiff [27], which calculates the similarity of decision boundaries by comparing the reaction to the input data. Initially, we sample input data \\( x_{i} \\) from the requester and adversarial data \\( x_{i}' \\) from a neighbor. To address privacy concerns, an alternative approach may sample data only from the requester, or the participants may opt to exchange mutated data, thus avoiding the sharing of private data. Then, we inject a pair-wise set of input data, \\( \\{(x_{1},x_{1}'),(x_{2},x_{2}'),\\dots ,(x_{n},x_{n}')\\} \\), into each model \\( f \\) and \\( g \\) and then computes the decision distance vector (DDV), \\( DDV_{f} = \\{d_{1},d_{2},\\dots ,d_{n}\\} \\), which is a distance vector for the output pair, where \\( d_{i} = \\text{distance}(f(x_{i}),f(x_{i}')) \\), \\( i\\in \\{1,2,\\ldots ,n\\} \\). Finally, it compares the cosine similarity between the distance vectors of models, \\( sim(DDV_{f},DDV_{g}) \\). Since it focuses the decision similarity instead of directly comparing the outputs, it is applicable to any heterogeneous task with a different number of output classes. To search a pair-wise set that can approximate the decision boundary most effectively, the input data is mutated, and similarities are iteratively computed, similar to genetic algorithms." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.75, + 0.79, + 0.844 + ], + "angle": 0, + "content": "By incorporating it in Pick-a-back, when we search for a similar model for task \\( T_{i}^{(l)} \\) of learner \\( l \\) among \\( \\hat{M}_{i-1}^{k} \\) of all neighboring learner \\( k \\). We sample an input pair-wise data \\( X_{i}^{(l+k)} \\) and \\( X_{i}^{\\prime(l+k)} \\) from \\( X_{i}^{(l)} \\cup X_{i-1}^{(k)} \\) (or just from \\( X_{i}^{(l)} \\)). By injecting \\( X_{i}^{(l+k)} \\) and \\( X_{i}^{\\prime(l+k)} \\) into the learner model \\( M_{i}^{(l)} \\) and the neighboring models \\( \\hat{M}_{i-1}^{(k)} \\), we produce the distance vector of outputs \\( D_{M_{i}^{(l)}} \\) and \\( D_{\\hat{M}_{i-1}^{(k)}} \\). Then," + } + ], + [ + { + "type": "page_number", + "bbox": [ + 0.218, + 0.116, + 0.23, + 0.127 + ], + "angle": 0, + "content": "8" + }, + { + "type": "header", + "bbox": [ + 0.272, + 0.115, + 0.404, + 0.127 + ], + "angle": 0, + "content": "J. Yoon and H. Lee" + }, + { + "type": "code_caption", + "bbox": [ + 0.218, + 0.146, + 0.665, + 0.162 + ], + "angle": 0, + "content": "Algorithm 1 Federated Continual Learning with Pick-a-back" + }, + { + "type": "algorithm", + "bbox": [ + 0.223, + 0.167, + 0.787, + 0.314 + ], + "angle": 0, + "content": "1: Input: \\(L\\): # of devices, (for each device \\(d_{l}\\)) \\(M_{0}^{(l)}\\): random initial model, \\(T^{(d_{l})}\\): a sequence of tasks, \\(\\theta_{i}^{(l)}\\): learning parameters \n2: for task \\(i = 1: length(T^{(l)})\\) do \n3: Device \\(d_{l}\\) trains task \\(i\\) such that \\(M_{i}^{(l)} = M_{i-1}^{(l)}(T_{i}^{(l)}, \\theta_{i}^{(l)})\\); \\(\\triangleright\\) Continual Learner \n4: for device \\(k = 1: L\\) do \\(\\triangleright\\) Federated Learner \n5: Compute the similarity, \\(sim(M_{i}^{(l)}, \\hat{M}_{i-1}^{(k)})\\); \n6: end for \n7: Select and retrieve knowledge \\(\\hat{M}_{i-1}^{(l)} \\gets \\hat{M}_{i-1}^{(k)}\\) such that \\(\\arg \\max_{k} sim(M_{i}^{(l)}, \\hat{M}_{i-1}^{(k)})\\); \n8: Train task \\(i\\) on the fetched model, \\(\\hat{M}_{i}^{(l)} = \\hat{M}_{i-1}^{(l)}(T_{i}^{(l)}, \\theta_{i}^{(l)})\\); \\(\\triangleright\\) Continual Learner \n9: end for" + }, + { + "type": "text", + "bbox": [ + 0.216, + 0.342, + 0.531, + 0.356 + ], + "angle": 0, + "content": "we select the most similar model as follows:" + }, + { + "type": "equation", + "bbox": [ + 0.377, + 0.364, + 0.787, + 0.388 + ], + "angle": 0, + "content": "\\[\n\\underset {k} {\\operatorname {a r g m a x}} \\operatorname {s i m} \\left(D D V _ {M _ {i} ^ {(l)}}, D D V _ {\\hat {M} _ {i - 1} ^ {(k)}}\\right) \\tag {1}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.215, + 0.394, + 0.789, + 0.429 + ], + "angle": 0, + "content": "We fetch the indirect knowledge that has already experienced the target decision pattern. After selecting the exterior knowledge, we continue to learn \\( T_{i}^{l} \\) on \\( \\hat{M}_{i - 1}^{(k)} \\)." + }, + { + "type": "text", + "bbox": [ + 0.215, + 0.439, + 0.789, + 0.524 + ], + "angle": 0, + "content": "Training process. Each learner \\( l \\) trains \\( i \\)-th task \\( T_{i}^{(l)} \\) and constructs a local model \\( M_{i}^{(l)} \\) with continual learning. Then, the learner \\( l \\) checks if there exists any helpful exterior knowledge by comparing \\( M_{i}^{(l)} \\) with \\( M_{i-1}^{(k)} \\) for all neighboring edge learners \\( k \\). After selecting relevant knowledge for knowledge transfer, device \\( d_{l} \\) processes the \\( i \\)-th task \\( T_{i}^{(l)} \\) over \\( \\hat{M}_{i-1}^{(k)} \\) and constructs \\( \\hat{M}_{i}^{(l)} = \\hat{M}_{i-1}^{(k)}(T_{i}^{(l)}, \\theta_{i}^{(l)}) \\)." + }, + { + "type": "text", + "bbox": [ + 0.24, + 0.524, + 0.604, + 0.54 + ], + "angle": 0, + "content": "The overall procedure is described in Algorithm 1." + }, + { + "type": "title", + "bbox": [ + 0.216, + 0.558, + 0.376, + 0.575 + ], + "angle": 0, + "content": "4 Experiments" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.585, + 0.789, + 0.647 + ], + "angle": 0, + "content": "We have validated Pick-a-back extensively on four different classification task groups with 13 public data using four different well-known architectures. Each group is considered as one experimental environment where knowledge can be shared, and one superclass group is a single task." + }, + { + "type": "text", + "bbox": [ + 0.215, + 0.656, + 0.8, + 0.704 + ], + "angle": 0, + "content": "Datasets. We used four different groups of CIFAR-100 [25], Classified-ImageNet [8], Fine-Grained [10,24,37,48,55], and Digits [7,11,35,36,39,45]. The organization of the dataset and the details are provided in Appendix Sec. A1." + }, + { + "type": "text", + "bbox": [ + 0.215, + 0.713, + 0.789, + 0.819 + ], + "angle": 0, + "content": "Architectures. We have presented the results using both lightweight models and large-scale models. For the lightweight models, we used LeNet-5 [26] and MobileNetV1 [17], where ImageNet-pretrained models are not publicly available. On the other hand, VGG16-BN [51] and EfficientNet-B0 [54] that have publicly available ImageNet-pretrained models were implemented using PyTorch [42]. Details of model configurations and learning parameters are in Appendix Sec. A2. We used CIFAR-100 on LeNet-5 as default, unless otherwise noted." + }, + { + "type": "page_footnote", + "bbox": [ + 0.218, + 0.825, + 0.765, + 0.842 + ], + "angle": 0, + "content": "1 Our codes are available at https://github.com/jinyi-yoon/Pick-a-back.git" + } + ], + [ + { + "type": "header", + "bbox": [ + 0.652, + 0.115, + 0.732, + 0.127 + ], + "angle": 0, + "content": "Pick-a-back" + }, + { + "type": "page_number", + "bbox": [ + 0.775, + 0.116, + 0.785, + 0.127 + ], + "angle": 0, + "content": "9" + }, + { + "type": "image", + "bbox": [ + 0.218, + 0.144, + 0.336, + 0.226 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.225, + 0.227, + 0.328, + 0.238 + ], + "angle": 0, + "content": "(a) Absolute Acc." + }, + { + "type": "image", + "bbox": [ + 0.338, + 0.144, + 0.457, + 0.226 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.345, + 0.227, + 0.452, + 0.238 + ], + "angle": 0, + "content": "(b) Improved Acc." + }, + { + "type": "image", + "bbox": [ + 0.466, + 0.154, + 0.619, + 0.226 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.477, + 0.228, + 0.607, + 0.239 + ], + "angle": 0, + "content": "(a) Training dynamics" + }, + { + "type": "image", + "bbox": [ + 0.622, + 0.155, + 0.784, + 0.226 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.624, + 0.228, + 0.78, + 0.239 + ], + "angle": 0, + "content": "(b) Accuracy improvement" + }, + { + "type": "image_caption", + "bbox": [ + 0.214, + 0.25, + 0.46, + 0.375 + ], + "angle": 0, + "content": "Fig.2: Heatmap of the accuracy for various backbone and target tasks sorted by the accuracy of \\( w / o \\) Backbone. The first row represents the averaged values for all tasks, while the first column shows the results of \\( w / o \\) Backbone. Absolute accuracy is scaled per each row, and accuracy improvement is scaled across all results" + }, + { + "type": "image_caption", + "bbox": [ + 0.463, + 0.251, + 0.788, + 0.376 + ], + "angle": 0, + "content": "Fig. 3: Training dynamics for the most effective case of the target task 14 using the backbone task of 4 and accuracy improvement with respect to the smoothed accuracy of \\( w / o \\) Backbone. The smoothed accuracy refers to the averaged accuracy for a subset of tasks with lower accuracy. For example, the second x-value represents the average of two tasks with the lowest performance, while the last value is the average across all tasks" + }, + { + "type": "text", + "bbox": [ + 0.215, + 0.404, + 0.789, + 0.433 + ], + "angle": 0, + "content": "Baselines. We compare Pick-a-back with diverse learning approaches of federated learning, continual learning, and continual federated learning." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.434, + 0.789, + 0.539 + ], + "angle": 0, + "content": "First, we let our Pick-a-back run on top of a certain continual learning scheme: 1) ProgressiveNet [47]: continual training only with expansion by concatenating scratch models for sequential tasks; 2) PackNet [30]: continual training only with pruning by allocating partial neurons exclusively to each task; 3) CPG [19]: continual training with expansion if necessary and pruning for the model compactness; and 4) WSN [21]: continual learning with reusing the masked weights of the prior subnetworks." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.54, + 0.789, + 0.631 + ], + "angle": 0, + "content": "We further compare across existing federated and continual federated learning approaches: 1) FedAVG [32]: one of the pioneering works for model regularization in federated learning; 2) FedWeIT [57]: a federated continual learning paradigm by introducing sparse task-specific parameters along with global federated parameters; and 3) FCCL [18]: a federated continual learning approach to tackle the catastrophic forgetting problem within a heterogeneous setting." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.631, + 0.789, + 0.753 + ], + "angle": 0, + "content": "We also compare Pick-a-back with different knowledge selection strategies: 1) \\( w / o \\) Backbone: intact continual learning without using any selective knowledge transfer, such as Progressive Networks, PackNet, or CPG; 2) Uniform Knowledge: applying the same task to its backbone model and averaging all of the results for each task; 3) Internal Knowledge: training the target task iteratively as a backbone model; 4) Random Knowledge: randomly selecting a backbone model; and 5) Upper: selecting an ideal backbone with a certain task that shows the highest accuracy improvement, being served as an unachievable upper bound." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.765, + 0.79, + 0.842 + ], + "angle": 0, + "content": "Measures. We have investigated 1) Avg: the averaged accuracy of all tasks and 2) Top: the accuracy of the most effective task that has gained the best accuracy improvement via selective knowledge. To compare the feasibility of models for edge AI, we have shown 1) Model size: the actual storage overhead of a model; 2) Expansion: the expansion degree where the scratch model has the value of 1;" + } + ], + [ + { + "type": "page_number", + "bbox": [ + 0.218, + 0.116, + 0.236, + 0.127 + ], + "angle": 0, + "content": "10" + }, + { + "type": "header", + "bbox": [ + 0.272, + 0.115, + 0.404, + 0.128 + ], + "angle": 0, + "content": "J. Yoon and H. Lee" + }, + { + "type": "table_caption", + "bbox": [ + 0.216, + 0.145, + 0.788, + 0.174 + ], + "angle": 0, + "content": "Table 1: Performance comparison. The bold text highlights the best performance per client, excluding Upper as an ideal upper-bound due to its infeasible brute-force manner" + }, + { + "type": "table", + "bbox": [ + 0.219, + 0.198, + 0.785, + 0.339 + ], + "angle": 0, + "content": "
Client ID1511716181231476824131951092011Avg
w/o Backbone40.055.057.260.451.866.067.467.669.270.872.673.074.275.275.876.477.077.481.284.069.1
FedAVG24.431.634.638.032.039.631.832.049.037.829.834.026.430.835.845.427.223.435.639.033.9
ProgressiveNet44.658.854.463.064.468.071.268.269.270.673.872.470.473.074.078.877.273.684.683.869.7
PackNet38.055.056.459.463.869.068.272.271.272.874.472.070.274.879.678.078.676.882.680.269.7
WSN34.253.849.453.459.860.263.663.066.865.064.665.466.662.661.267.666.658.674.079.861.8
FedWeIT40.052.555.056.056.260.556.760.567.065.260.265.066.064.265.065.766.751.068.777.861.0
FCCL37.054.452.460.662.268.267.264.468.270.672.872.468.477.073.875.078.668.879.278.267.5
Uniform41.558.657.662.165.670.769.771.671.272.574.773.073.776.477.178.779.475.283.583.070.8
Internal43.255.456.862.267.867.069.071.473.874.876.871.873.476.279.277.681.079.282.083.871.1
Random39.859.056.659.267.672.668.273.071.471.475.672.874.476.478.078.277.874.684.283.870.7
Upper46.462.460.064.267.874.872.275.673.875.677.075.076.079.480.080.681.679.286.285.473.7
Pick-a-back46.459.060.063.264.472.671.875.672.472.277.073.072.876.276.080.278.275.884.285.471.8
" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.38, + 0.789, + 0.426 + ], + "angle": 0, + "content": "3) Sparsity: the released ratio of a model, where higher sparsity indicates more space for next tasks; 4) Training time: the execution time to run for one epoch; and 5) Used data: the number of data samples used for training." + }, + { + "type": "title", + "bbox": [ + 0.216, + 0.45, + 0.581, + 0.466 + ], + "angle": 0, + "content": "4.1 Overall Performance and Comparisons" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.476, + 0.789, + 0.658 + ], + "angle": 0, + "content": "Motivations for selective knowledge federation. We first examined the impact of different knowledge on various target tasks. As in Fig. 2a, a significant portion of the results and the average accuracy achieved through backbone usage are represented by a blue color, indicating that the continual learner benefited from the knowledge federation. However, there exist not only positive backbones but also negative backbones, suggesting a need for selective knowledge to assist a continual learner specific to its target task. Taking a closer look at the accuracy improvement relative to \\( w / o \\) Backbone in Fig. 2b, a well-constructed model such as task 9 or 11 is more likely disturbed by external knowledge. On the other hand, a target task with initially poor performance tends to gain a greater improvement in accuracy. It implies the importance of selectively incorporating constructive knowledge to enhance the learning process for a specific target task." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.674, + 0.789, + 0.842 + ], + "angle": 0, + "content": "Ablation study on selective knowledge federation. Inspired by the Fig. 2 we verified the effectiveness, compared to diverse learning approaches, including FedAVG, ProgressiveNet, PackNet, WSN, FedWeIT, and FCCL, along with various backbone selection strategies of Uniform, Internal, Random, and Upper, as in Table 1 and in Table A5. We consistently demonstrate superior fidelity on average. By employing our selection approach, we observed a significant accuracy improvement of up to \\(8.0\\%\\), with only two negative cases out of 20 tasks. Since FedAVG simply averages the model parameters without considering the continual tasks, it suffers from losing the knowledge of trained tasks. Notably, although Pick-a-back with CPG shows the best performance, Pick-a-back with other continual learning approaches of ProgressiveNet and PackNet also outperforms" + } + ], + [ + { + "type": "header", + "bbox": [ + 0.652, + 0.115, + 0.732, + 0.127 + ], + "angle": 0, + "content": "Pick-a-back" + }, + { + "type": "page_number", + "bbox": [ + 0.769, + 0.117, + 0.784, + 0.127 + ], + "angle": 0, + "content": "11" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.147, + 0.788, + 0.239 + ], + "angle": 0, + "content": "\\(w / o\\) Backbone. Even Uniform, Internal, and Random selection strategies exhibit enhanced performance, suggesting the efficacy of inter-device knowledge transfer based on Pick-a-back primitive. We also compared with other recent (federated) continual learning approaches (e.g., WSN, FedWeIT, and FCCL), showing that Pick-a-back outperforms these methods. This is due to the effectiveness of our knowledge reuse approach along with our backbone selection strategy." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.254, + 0.79, + 0.452 + ], + "angle": 0, + "content": "Benefit of backbone knowledge. Taking a closer look at how Pick-a-back benefits from backbone knowledge, in Fig. 3a, it is evident that the test accuracy of Pick-a-back consistently surpasses the training accuracy of \\( w / o \\) Backbone, starting from the beginning until the saturation. It indicates that indirect knowledge assists individual learners to be exposed to inexperienced knowledge and adapt to unseen data. Even within Pick-a-back itself, the test accuracy outperforms the training accuracy in the early stages of training, suggesting that the transferred knowledge has relevance to the target knowledge and drives toward high performance even earlier. Furthermore, \\( w / o \\) Backbone achieves an accuracy of \\( 67.4\\% \\), which can be attained only within approximately 20 epochs using Pick-a-back, whereas it takes over 50 epochs without knowledge transfer. This demonstrates that Pick-a-back not only accelerates the learning, but also enhances the overall performance. Please refer to the empirical convergence analysis in Appendix Fig. A1." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.466, + 0.79, + 0.724 + ], + "angle": 0, + "content": "Possibility of similar but generalized knowledge. To justify the task-specific selection, we compared to uniform knowledge, where the same backbone is applied to all tasks. As in Fig. 3b, the results showed that Upper exhibits better improvement in tasks with poor initial performance. However, the uniform knowledge ignores task-specific improvements, leading to similar improvements across all target tasks. Interestingly, we also observed that only using internal knowledge, where the target task is trained twice, allowed further improvement. It means that even with a single task iteratively training by accumulating knowledge on top of an identical model turns out to induce better knowledge. It also implies that a backbone task with a similar, ideally the same in personal knowledge case, the decision boundary can further improve the knowledge. While the same task aims for identical knowledge, using similar but distinct knowledge enables generalization and better improvement beyond individual tasks. As expected, Pick-a-back outperformed the one only with internal knowledge, highlighting the advantage of leveraging similar indirect experiences. This means that we provide a more opportunity to acquire superior knowledge by mitigating the overfitting problem and extending to the unseen data beyond local knowledge." + }, + { + "type": "title", + "bbox": [ + 0.215, + 0.746, + 0.424, + 0.761 + ], + "angle": 0, + "content": "4.2 Scalability Analysis" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.772, + 0.789, + 0.805 + ], + "angle": 0, + "content": "Lightweight models for edge AI. We evaluated Pick-a-back in various architectures from the lightweight models of LeNet-5, MobileNet-V1, and EfficientNet-B0[2]" + }, + { + "type": "page_footnote", + "bbox": [ + 0.218, + 0.811, + 0.788, + 0.84 + ], + "angle": 0, + "content": "2 For EfficientNet-B0, we enlarge the images to \\( 224 \\times 224 \\) pixels for evaluation, but a model size is measured with the original size of \\( 32 \\times 32 \\) for a fair comparison." + } + ], + [ + { + "type": "page_number", + "bbox": [ + 0.218, + 0.116, + 0.236, + 0.127 + ], + "angle": 0, + "content": "12" + }, + { + "type": "header", + "bbox": [ + 0.272, + 0.115, + 0.404, + 0.127 + ], + "angle": 0, + "content": "J. Yoon and H. Lee" + }, + { + "type": "table_caption", + "bbox": [ + 0.216, + 0.145, + 0.788, + 0.174 + ], + "angle": 0, + "content": "Table 2: Accuracy, model size, and averaged training time per one epoch with respect to different architectures on CIFAR-100. We averaged the training time over 10 epochs" + }, + { + "type": "table", + "bbox": [ + 0.239, + 0.185, + 0.767, + 0.329 + ], + "angle": 0, + "content": "
AccuracyModel SizeTraining Time
AvgTop
VGG16-BNw/o Backbone78.968.2164 MB4.10 s
Pick-a-back81.0 (+2.1)76.6 (+8.4)
EfficientNet-B0w/o Backbone77.844.620 MB9.29 s
Pick-a-back81.8 (+4.0)54.4 (+9.8)
MobileNet-V1w/o Backbone65.459.816 MB4.95 s
Pick-a-back66.7 (+1.3)65.4 (+5.6)
LeNet-5w/o Backbone69.167.60.3 MB2.18 s
Pick-a-back71.8 (+2.7)75.6 (+8.0)
" + }, + { + "type": "table_caption", + "bbox": [ + 0.365, + 0.341, + 0.637, + 0.354 + ], + "angle": 0, + "content": "Table 3: Various datasets with LeNet-5" + }, + { + "type": "table", + "bbox": [ + 0.224, + 0.368, + 0.78, + 0.452 + ], + "angle": 0, + "content": "
DigitsCIFAR-100Classified-ImageNetFine-Grained
AvgTopAvgTopAvgTopAvgTop
Upper95.889.473.775.661.471.833.948.5
w/o Backbone93.880.769.167.656.159.225.021.0
Pick-a-back95.186.471.875.658.167.628.327.5
(+1.3)(+5.7)(+2.7)(+8.0)(+2.0)(+8.4)(+3.3)(+6.5)
" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.478, + 0.789, + 0.646 + ], + "angle": 0, + "content": "to the large-scale models of VGG16-BN in Table 2. It indicates that all models benefit from selectively transferred knowledge, leading to accuracy improvement of up to \\(8.4\\%\\), \\(9.8\\%\\), \\(5.6\\%\\), and \\(8.0\\%\\), respectively. However, in resource-constrained environments, the limitations of memory and processing power make it challenging to deploy a large-scale model at the edge. To make learning practical in the real-world, lightweight models can be flexibly deployed across multiple edge devices, reducing dependence on remote servers. The lightweight LeNet-5 has a much smaller size of 0.18 times compared to VGG16-BN, enabling faster training and inference time. Although the lightweight models may not match closely to the large-scale models, Pick-a-back can still achieve improvement, comparable to the large models, while greatly reducing the storage requirement and run-time." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.659, + 0.79, + 0.842 + ], + "angle": 0, + "content": "Scalability across diverse datasets. We assessed the scalability of Pick-a-back across various datasets in Table 3. We stably achieve improvements across different datasets, even with varying degrees of class complexity. In Digits, where most tasks except for the Roman task are generally easy to classify, the average accuracy improvements may appear modest, but Pick-a-back approach achieved a significant accuracy improvement of up to \\(5.72\\%\\) in the relatively challenging Roman task. Similarly, in the classified-ImageNet dataset, where there exist enough neighboring tasks to provide additional knowledge, Pick-a-back shows substantial improvements even for the underperforming tasks. However, when considering fine-grained datasets with a large number of classes per task (ranging \\(102\\sim 250\\) classes), classification becomes inherently more difficult with simple models, resulting in relatively lower accuracy. Nonetheless, by applying Pick" + } + ], + [ + { + "type": "header", + "bbox": [ + 0.652, + 0.115, + 0.732, + 0.127 + ], + "angle": 0, + "content": "Pick-a-back" + }, + { + "type": "page_number", + "bbox": [ + 0.769, + 0.116, + 0.785, + 0.127 + ], + "angle": 0, + "content": "13" + }, + { + "type": "table_caption", + "bbox": [ + 0.216, + 0.145, + 0.785, + 0.173 + ], + "angle": 0, + "content": "Table 4: Performance on Fine-Grained Table 5: Heterogeneous data sampling datasets using EfficientNet-B0 with \\(\\alpha = 0.5\\)" + }, + { + "type": "table", + "bbox": [ + 0.221, + 0.185, + 0.784, + 0.248 + ], + "angle": 0, + "content": "
Client ID12345AvgHeteroBalancedw/o Sampling
AvgTopAvgTopAvgTop
Pretrained w/o Backbone28.990.542.449.937.149.8Upper59.360.269.684.473.775.6
66.37.428.212.724.6w/o Backbone52.640.264.664.171.869.167.6
Pick-a-back49.5 (+41.2)77.4 (+11.1)44.0 (+36.6)46.4 (+18.2)64.6 (+51.9)56.4 (+31.8)Pick-a-back56.1 (+3.5)58.6 (+18.4)67.3 (+3.2)80.0 (+8.2)71.8 (+2.7)75.6 (+8.0)
" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.278, + 0.788, + 0.4 + ], + "angle": 0, + "content": "\\(a\\)-back, it is possible to advance the knowledge and achieve performance gains. While there is a noticeable performance gap of \\(0.7\\%\\), \\(1.9\\%\\), and \\(3.3\\%\\), respectively, between Pick-a-back and Upper in other datasets, the gap in the fine-grained dataset is quite large with \\(5.6\\%\\) on average and up to \\(21.0\\%\\) for the task of flowers. This implies the potential of Pick-a-back to drive further improvements in complex tasks that require more sophisticated classification. We further explore the performance with the different number of clients (i.e., classes) and the different learning seeds in Appendix Table A9 and in Appendix Table A10, respectively." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.415, + 0.789, + 0.628 + ], + "angle": 0, + "content": "Impact on the most complicated task. We investigate the most complicated tasks of CUB 55, Stanford Cars 24, Flower 37, WikiArt 48, and Sketch 10, with each featuring a diverse range of class numbers of 200, 196, 102, 195, and 250 and running at each separate client. As shown in Table 4 with the adoption of Pick-a-back, we observe remarkable improvements in overall accuracy, averaging at \\(31.8\\%\\), and peaking at an impressive \\(51.9\\%\\). Although these results might appear surprising, they align with the findings previously established by 19 and 29, where scratch accuracy on ImageNet-pretrained models of ResNet50 and VGG16-BN, achieved \\(58.83\\%\\) and \\(43.84\\%\\), respectively. Incorporating Pick-a-back not only surpasses the average accuracy of ImageNet-pretrained models, but does so through a judicious infusion of tailored knowledge. It is worth noting that ImageNet-pretrained models often necessitate extensive prior training data. However, in many instances, the efficacy diminishes due to the lack of distinct knowledge selection, showing some critical drawbacks in clients #1, #3, and #5." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.644, + 0.789, + 0.841 + ], + "angle": 0, + "content": "Superiority in heterogeneous data scenarios. In real-world scenarios, there are many unique challenges such as personalization [5] or heterogeneity [4]. Here, we explored the effectiveness of Pick-a-back under data heterogeneity. Leveraging the heterogeneity settings from federated learning researches [16, 60], we applied \\(\\alpha\\) of 0.5 based on the Dirichlet distribution \\(Dir(\\alpha)\\). As in Table [5] we observed a degradation in the accuracy of scratch models as the data heterogeneity increases. It implies that the accuracy is achievable, and the external knowledge can be particularly beneficial for weaker classes in the presence of heterogeneous data. While the average accuracy improvement remained consistent, we achieved significant enhancements in heterogeneous data settings. It is essential to address unbalanced data distributions, as certain classes may receive less training attention. By leveraging Pick-a-back, we are able to achieve substantial performance gains, with improvements of up to \\(18.4\\%\\)." + } + ], + [ + { + "type": "page_number", + "bbox": [ + 0.218, + 0.116, + 0.236, + 0.127 + ], + "angle": 0, + "content": "14" + }, + { + "type": "header", + "bbox": [ + 0.272, + 0.115, + 0.404, + 0.127 + ], + "angle": 0, + "content": "J. Yoon and H. Lee" + }, + { + "type": "table_caption", + "bbox": [ + 0.216, + 0.145, + 0.787, + 0.16 + ], + "angle": 0, + "content": "Table 6: Performance and data used for training in ImageNet-pretrained VGG16-BN" + }, + { + "type": "table", + "bbox": [ + 0.286, + 0.184, + 0.72, + 0.248 + ], + "angle": 0, + "content": "
w/o Backbone → Pick-a-backSize of Training Data
AvgTop
w/ Pretrained84.9 → 83.882.8 → 84.492.8 GB
w/o Pretrained78.9 → 81.068.2 → 76.60.3 GB
" + }, + { + "type": "title", + "bbox": [ + 0.216, + 0.286, + 0.492, + 0.3 + ], + "angle": 0, + "content": "4.3 Discussions and Limitations" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.308, + 0.789, + 0.504 + ], + "angle": 0, + "content": "ImageNet-pretrained models. It is important to note that a lightweight model such as LeNet-5 may not outperform a large-scale model even with Pick-a-back. Also, VGG16-BN models initialized with ImageNet-pretrained weights demonstrate strong performance generally for visual tasks as in Table 6 and in Appendix Table A6. However, their effectiveness heavily relies on the availability of large-scale labeled datasets that need to be uploaded to a central server. Additionally, the pretrained models are specifically designed for image classification, and their performance may vary when applied to specific or specialized tasks. Furthermore, the availability of pretrained models is limited only to well-known large-scale architectures, making their deployment usually infeasible to edge learning. Without pretrained model, selective knowledge federation through Pick-a-back achieves similar performance on average with a small gap of \\(3.9\\%\\) compared to the original pretrained model (w/o Backbone), but only with \\(0.3\\%\\) of training data." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.505, + 0.788, + 0.597 + ], + "angle": 0, + "content": "We further provided the orthogonality of continual learner in Pick-a-back and catastrophic forgetting prevention across different continual learning approaches in Appendix Table A11. Lastly, we need image sharing among clients and model sharing from a selected neighboring client, which may incur transmission overhead and privacy issues. We thus further assess the communication and computation time and how we can defend against attacks in Appendix Table A12 and Fig. A2." + }, + { + "type": "title", + "bbox": [ + 0.216, + 0.615, + 0.36, + 0.631 + ], + "angle": 0, + "content": "5 Conclusion" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.644, + 0.788, + 0.704 + ], + "angle": 0, + "content": "We propose Pick-a-back that picks a task-specific backbone knowledge and fosters the collective intelligence in federated continual learning. As the individual learner pursues each distinct goal with private heterogeneous data, we selectively transfer the relevant external knowledge to the target task between neighboring learners." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.705, + 0.789, + 0.841 + ], + "angle": 0, + "content": "As AI expands towards edges for scalability, low-end devices tend to be trained toward personalized and specialized but somewhat partially limited intelligence. Here, in case of a scarcity of data or with limited performance, it becomes essential to seek assistance from external sources. The question - how can a device make a local model more effectively trained with exterior knowledge? - can be answered: Pick-a-back offers an effective way to selective knowledge transfer from one device to another towards distributed edge intelligence. By harnessing the power of neighboring devices, individual learners can gain access to valuable insights and expertise via collective intelligence, augmenting their own capabilities." + } + ], + [ + { + "type": "header", + "bbox": [ + 0.652, + 0.115, + 0.732, + 0.127 + ], + "angle": 0, + "content": "Pick-a-back" + }, + { + "type": "page_number", + "bbox": [ + 0.769, + 0.117, + 0.786, + 0.127 + ], + "angle": 0, + "content": "15" + }, + { + "type": "title", + "bbox": [ + 0.218, + 0.145, + 0.393, + 0.163 + ], + "angle": 0, + "content": "Acknowledgments" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.175, + 0.788, + 0.283 + ], + "angle": 0, + "content": "We thank the Virginia Tech Presidential Postdoctoral Fellowship Program for supporting JinYi Yoon's postdoctoral associate position. This work was supported by the Institute of Information & communications Technology Planning and Evaluation (IITP) grant funded by the Korea government (MSIT) (No. RS-2022-00155966, Artificial Intelligence Convergence Innovation Human Resources Development (Ewha Womans University) and No. RS-2021-II212068, Artificial Intelligence Innovation Hub)." + }, + { + "type": "title", + "bbox": [ + 0.218, + 0.303, + 0.323, + 0.319 + ], + "angle": 0, + "content": "References" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.332, + 0.787, + 0.36 + ], + "angle": 0, + "content": "1. Arani, E., et al.: Learning fast, learning slow: A general continual learning method based on complementary learning system. ICLR (2022)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.361, + 0.788, + 0.387 + ], + "angle": 0, + "content": "2. Cao, L.: Decentralized ai: Edge intelligence and smart blockchain, metaverse, web3, and desci. IEEE Intelligent Systems 37(3), 6-19 (2022)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.388, + 0.788, + 0.429 + ], + "angle": 0, + "content": "3. Casado, F.E., Lema, D., Criado, M.F., Iglesias, R., Regueiro, C.V., Barro, S.: Concept drift detection and adaptation for federated and continual learning. Multimedia Tools and Applications pp. 1-23 (2022)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.429, + 0.787, + 0.47 + ], + "angle": 0, + "content": "4. Chen, D., Gao, D., Kuang, W., Li, Y., Ding, B.: pfl-bench: A comprehensive benchmark for personalized federated learning. Advances in Neural Information Processing Systems 35, 9344-9360 (2022)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.471, + 0.787, + 0.525 + ], + "angle": 0, + "content": "5. Chen, D., Gao, D., Xie, Y., Pan, X., Li, Z., Li, Y., Ding, B., Zhou, J.: Fs-real: Towards real-world cross-device federated learning. In: Proceedings of the 29th ACM SIGKDD Conference on Knowledge Discovery and Data Mining. pp. 3829-3841 (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.526, + 0.787, + 0.566 + ], + "angle": 0, + "content": "6. Chen, Y., Inaltekin, H., Gorlatova, M.: Adaptslam: Edge-assisted adaptive slam with resource constraints via uncertainty minimization. arXiv preprint arXiv:2301.04620 (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.567, + 0.787, + 0.594 + ], + "angle": 0, + "content": "7. Colianni, S.: Mnist as .jpg (2017), https://www.kaggle.com/scolianni/ mnistas.jpg" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.595, + 0.787, + 0.648 + ], + "angle": 0, + "content": "8. Deng, J., Dong, W., Socher, R., Li, L.J., Li, K., Fei-Fei, L.: Imagenet: A large-scale hierarchical image database. In: 2009 IEEE Conference on Computer Vision and Pattern Recognition. pp. 248-255 (2009). https://doi.org/10.1109/CVPR.2009.5206848" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.649, + 0.787, + 0.69 + ], + "angle": 0, + "content": "9. Dhar, P., Singh, R.V., Peng, K.C., Wu, Z., Chellappa, R.: Learning without memorizing. In: Proceedings of the IEEE/CVF conference on computer vision and pattern recognition. pp. 5138-5146 (2019)" + }, + { + "type": "ref_text", + "bbox": [ + 0.22, + 0.691, + 0.787, + 0.718 + ], + "angle": 0, + "content": "10. Eitz, M., Hays, J., Alexa, M.: How do humans sketch objects? ACM Transactions on graphics (TOG) 31(4), 1-10 (2012)" + }, + { + "type": "ref_text", + "bbox": [ + 0.22, + 0.718, + 0.787, + 0.758 + ], + "angle": 0, + "content": "11. El-Sawy, A., El-Bakry, H., Loey, M.: Cnn for handwritten arabic digits recognition based on lenet-5. In: Proceedings of the International Conference on Advanced Intelligent Systems and Informatics 2016 2. pp. 566-575. Springer (2017)" + }, + { + "type": "ref_text", + "bbox": [ + 0.22, + 0.759, + 0.787, + 0.786 + ], + "angle": 0, + "content": "12. Goodfellow, I., Bengio, Y., Courville, A.: Deep Learning. MIT Press (2016), http://www.deeplearningbook.org" + }, + { + "type": "ref_text", + "bbox": [ + 0.22, + 0.786, + 0.787, + 0.84 + ], + "angle": 0, + "content": "13. Gu, Y., Yang, X., Wei, K., Deng, C.: Not just selection, but exploration: Online class-incremental continual learning via dual view consistency. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 7442-7451 (2022)" + }, + { + "type": "list", + "bbox": [ + 0.22, + 0.332, + 0.788, + 0.84 + ], + "angle": 0, + "content": null + } + ], + [ + { + "type": "page_number", + "bbox": [ + 0.218, + 0.116, + 0.236, + 0.127 + ], + "angle": 0, + "content": "16" + }, + { + "type": "header", + "bbox": [ + 0.272, + 0.115, + 0.404, + 0.127 + ], + "angle": 0, + "content": "J. Yoon and H. Lee" + }, + { + "type": "ref_text", + "bbox": [ + 0.218, + 0.147, + 0.785, + 0.175 + ], + "angle": 0, + "content": "14. Ha, D., Tang, Y.: Collective intelligence for deep learning: A survey of recent developments. Collective Intelligence 1(1), 263391372211114874 (2022)" + }, + { + "type": "ref_text", + "bbox": [ + 0.218, + 0.175, + 0.787, + 0.215 + ], + "angle": 0, + "content": "15. Haim, N., Vardi, G., Yehudai, G., Shamir, O., Irani, M.: Reconstructing training data from trained neural networks. Advances in Neural Information Processing Systems 35, 22911-22924 (2022)" + }, + { + "type": "ref_text", + "bbox": [ + 0.218, + 0.215, + 0.787, + 0.257 + ], + "angle": 0, + "content": "16. He, C., Li, S., So, J., Zeng, X., Zhang, M., Wang, H., Wang, X., Vepakomma, P., Singh, A., Qiu, H., et al.: Fedml: A research library and benchmark for federated machine learning. arXiv preprint arXiv:2007.13518 (2020)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.257, + 0.787, + 0.298 + ], + "angle": 0, + "content": "17. Howard, A.G., Zhu, M., Chen, B., Kalenichenko, D., Wang, W., Weyand, T., Andreetto, M., Adam, H.: Mobilenets: Efficient convolutional neural networks for mobile vision applications. arXiv preprint arXiv:1704.04861 (2017)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.298, + 0.787, + 0.338 + ], + "angle": 0, + "content": "18. Huang, W., Ye, M., Du, B.: Learn from others and be yourself in heterogeneous federated learning. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 10143-10153 (2022)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.338, + 0.787, + 0.379 + ], + "angle": 0, + "content": "19. Hung, C.Y., Tu, C.H., Wu, C.E., Chen, C.H., Chan, Y.M., Chen, C.S.: Compacting, picking and growing for unforgetting continual learning. Advances in Neural Information Processing Systems 32 (2019)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.379, + 0.787, + 0.419 + ], + "angle": 0, + "content": "20. Jeong, W., Hwang, S.J.: Factorized-fl: Personalized federated learning with parameter factorization & similarity matching. Advances in Neural Information Processing Systems 35, 35684-35695 (2022)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.42, + 0.787, + 0.461 + ], + "angle": 0, + "content": "21. Kang, H., Mina, R.J.L., Madjid, S.R.H., Yoon, J., Hasegawa-Johnson, M., Hwang, S.J., Yoo, C.D.: Forget-free continual learning with winning subnetworks. In: International Conference on Machine Learning. pp. 10734-10750. PMLR (2022)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.461, + 0.787, + 0.501 + ], + "angle": 0, + "content": "22. Kim, G., Esmaeilpour, S., Xiao, C., Liu, B.: Continual learning based on ood detection and task masking. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 3856-3866 (2022)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.501, + 0.787, + 0.556 + ], + "angle": 0, + "content": "23. Kirkpatrick, J., Pascanu, R., Rabinowitz, N., Veness, J., Desjardins, G., Rusu, A.A., Milan, K., Quan, J., Ramalho, T., Grabska-Barwinska, A., et al.: Overcoming catastrophic forgetting in neural networks. Proceedings of the national academy of sciences 114(13), 3521-3526 (2017)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.556, + 0.787, + 0.597 + ], + "angle": 0, + "content": "24. Krause, J., Stark, M., Deng, J., Fei-Fei, L.: 3d object representations for fine-grained categorization. In: Proceedings of the IEEE international conference on computer vision workshops. pp. 554-561 (2013)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.597, + 0.787, + 0.623 + ], + "angle": 0, + "content": "25. Krizhevsky, A., Hinton, G., et al.: Learning multiple layers of features from tiny images (2009)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.623, + 0.787, + 0.651 + ], + "angle": 0, + "content": "26. LeCun, Y., et al.: Lenet-5, convolutional neural networks. URL: http://yann.lecun.com/exdb/lenet 20(5), 14 (2015)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.651, + 0.787, + 0.691 + ], + "angle": 0, + "content": "27. Li, Y., Zhang, Z., Liu, B., Yang, Z., Liu, Y.: Modeldiff: testing-based dnn similarity comparison for model reuse detection. In: Proceedings of the 30th ACM SIGSOFT International Symposium on Software Testing and Analysis. pp. 139-151 (2021)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.691, + 0.787, + 0.732 + ], + "angle": 0, + "content": "28. Liu, F., Wu, X., Ge, S., Fan, W., Zou, Y.: Federated learning for vision-and-language grounding problems. In: Proceedings of the AAAI Conference on Artificial Intelligence. vol. 34, pp. 11572-11579 (2020)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.732, + 0.787, + 0.772 + ], + "angle": 0, + "content": "29. Mallya, A., Davis, D., Lazebnik, S.: Piggyback: Adapting a single network to multiple tasks by learning to mask weights. In: Proceedings of the European Conference on Computer Vision (ECCV). pp. 67-82 (2018)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.772, + 0.787, + 0.813 + ], + "angle": 0, + "content": "30. Mallya, A., Lazebnik, S.: Packnet: Adding multiple tasks to a single network by iterative pruning. In: Proceedings of the IEEE conference on Computer Vision and Pattern Recognition. pp. 7765-7773 (2018)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.813, + 0.787, + 0.84 + ], + "angle": 0, + "content": "31. Mao, B., Liu, J., Wu, Y., Kato, N.: Security and privacy on 6g network edge: A survey. IEEE Communications Surveys & Tutorials (2023)" + }, + { + "type": "list", + "bbox": [ + 0.217, + 0.147, + 0.787, + 0.84 + ], + "angle": 0, + "content": null + } + ], + [ + { + "type": "header", + "bbox": [ + 0.652, + 0.115, + 0.732, + 0.127 + ], + "angle": 0, + "content": "Pick-a-back" + }, + { + "type": "page_number", + "bbox": [ + 0.769, + 0.116, + 0.786, + 0.127 + ], + "angle": 0, + "content": "17" + }, + { + "type": "ref_text", + "bbox": [ + 0.214, + 0.147, + 0.788, + 0.189 + ], + "angle": 0, + "content": "32. McMahan, B., Moore, E., Ramage, D., Hampson, S., y Arcas, B.A.: Communication-efficient learning of deep networks from decentralized data. In: Artificial intelligence and statistics. pp. 1273-1282. PMLR (2017)" + }, + { + "type": "ref_text", + "bbox": [ + 0.214, + 0.19, + 0.788, + 0.232 + ], + "angle": 0, + "content": "33. Meng, C., Tang, M., Setayesh, M., Wong, V.W.: Gnn-based neighbor selection and resource allocation for decentralized federated learning. In: GLOBECOM 2023-2023 IEEE Global Communications Conference. pp. 1223-1228. IEEE (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.214, + 0.232, + 0.788, + 0.259 + ], + "angle": 0, + "content": "34. Miller, G.A.: Wordnet: a lexical database for english. Communications of the ACM 38(11), 39-41 (1995)" + }, + { + "type": "ref_text", + "bbox": [ + 0.214, + 0.259, + 0.788, + 0.3 + ], + "angle": 0, + "content": "35. Nazarpour, K., Chen, M.: Handwritten Chinese Numbers (1 2017). https://doi.org/10.17634/137930-3, https://data.ncl.ac.uk/articles/dataset/Handwritten_Chinese_Numbers/10280831" + }, + { + "type": "ref_text", + "bbox": [ + 0.214, + 0.3, + 0.788, + 0.328 + ], + "angle": 0, + "content": "36. Netzer, Y., Wang, T., Coates, A., Bissacco, A., Wu, B., Ng, A.Y.: Reading digits in natural images with unsupervised feature learning (2011)" + }, + { + "type": "ref_text", + "bbox": [ + 0.214, + 0.328, + 0.788, + 0.37 + ], + "angle": 0, + "content": "37. Nilsback, M.E., Zisserman, A.: Automated flower classification over a large number of classes. In: 2008 Sixth Indian Conference on Computer Vision, Graphics & Image Processing. pp. 722-729. IEEE (2008)" + }, + { + "type": "ref_text", + "bbox": [ + 0.214, + 0.371, + 0.788, + 0.412 + ], + "angle": 0, + "content": "38. Nishio, T., Yonetani, R.: Client selection for federated learning with heterogeneous resources in mobile edge. In: ICC 2019-2019 IEEE international conference on communications (ICC). pp. 1-7. IEEE (2019)" + }, + { + "type": "ref_text", + "bbox": [ + 0.214, + 0.412, + 0.788, + 0.439 + ], + "angle": 0, + "content": "39. Noyan, M.A.: Hand written roman numerals dataset (Sep 2021). https://doi.org/10.5281/zenodo.5385144, https://doi.org/10.5281/zenodo.5385144" + }, + { + "type": "ref_text", + "bbox": [ + 0.214, + 0.439, + 0.788, + 0.495 + ], + "angle": 0, + "content": "40. Ostapenko, O., Puscas, M., Klein, T., Jahnichen, P., Nabi, M.: Learning to remember: A synaptic plasticity driven framework for continual learning. In: Proceedings of the IEEE/CVF conference on computer vision and pattern recognition. pp. 11321-11329 (2019)" + }, + { + "type": "ref_text", + "bbox": [ + 0.214, + 0.495, + 0.788, + 0.577 + ], + "angle": 0, + "content": "41. Parks Associates: At CES 2024, Parks Associates Announces New Research Showing Average Number of Connected Devices per US Internet Household Reached 17 in 2023. https://www.parksassociates.com/blogs/press-releases/at-ces-2024-parks-associates-announces-new-research-showing-average-number-of-connected-devices-per-us-internet-household-reached-17-in-2023 (2024), accessed: 2024-05-15" + }, + { + "type": "ref_text", + "bbox": [ + 0.214, + 0.577, + 0.788, + 0.604 + ], + "angle": 0, + "content": "42. PyTorch: Retrieved from python model zoo (2017), https://pytorch.org/vision/stable/models.html" + }, + { + "type": "ref_text", + "bbox": [ + 0.214, + 0.605, + 0.788, + 0.66 + ], + "angle": 0, + "content": "43. Radford, A., Kim, J.W., Hallacy, C., Ramesh, A., Goh, G., Agarwal, S., Sastry, G., Askell, A., Mishkin, P., Clark, J., et al.: Learning transferable visual models from natural language supervision. In: International conference on machine learning. pp. 8748-8763. PMLR (2021)" + }, + { + "type": "ref_text", + "bbox": [ + 0.214, + 0.66, + 0.788, + 0.702 + ], + "angle": 0, + "content": "44. Roy, A.G., Siddiqui, S., Pölsterl, S., Navab, N., Wachinger, C.: Brainstorm: A peer-to-peer environment for decentralized federated learning. arXiv preprint arXiv:1905.06731 (2019)" + }, + { + "type": "ref_text", + "bbox": [ + 0.214, + 0.702, + 0.788, + 0.73 + ], + "angle": 0, + "content": "45. Roy, P., Ghosh, S., Bhattacharya, S., Pal, U.: Effects of degradations on deep neural network architectures. arXiv preprint arXiv:1807.10108 (2018)" + }, + { + "type": "ref_text", + "bbox": [ + 0.214, + 0.73, + 0.788, + 0.772 + ], + "angle": 0, + "content": "46. Rusu, A.A., Rabinowitz, N.C., Desjardins, G., Soyer, H., Kirkpatrick, J., Kavukcuoglu, K., Pascanu, R., Hadsell, R.: Progressive neural networks. arXiv preprint arXiv:1606.04671 (2016)" + }, + { + "type": "ref_text", + "bbox": [ + 0.214, + 0.772, + 0.788, + 0.813 + ], + "angle": 0, + "content": "47. Rusu, A.A., Rabinowitz, N.C., Desjardins, G., Soyer, H., Kirkpatrick, J., Kavukcuoglu, K., Pascanu, R., Hadsell, R.: Progressive neural networks. arXiv preprint arXiv:1606.04671 (2016)" + }, + { + "type": "ref_text", + "bbox": [ + 0.214, + 0.813, + 0.788, + 0.841 + ], + "angle": 0, + "content": "48. Saleh, B., Elgammal, A.: Large-scale classification of fine-art paintings: Learning the right metric on the right feature. arXiv preprint arXiv:1505.00855 (2015)" + }, + { + "type": "list", + "bbox": [ + 0.214, + 0.147, + 0.788, + 0.841 + ], + "angle": 0, + "content": null + } + ], + [ + { + "type": "page_number", + "bbox": [ + 0.218, + 0.116, + 0.236, + 0.127 + ], + "angle": 0, + "content": "18" + }, + { + "type": "header", + "bbox": [ + 0.272, + 0.115, + 0.404, + 0.127 + ], + "angle": 0, + "content": "J. Yoon and H. Lee" + }, + { + "type": "ref_text", + "bbox": [ + 0.218, + 0.148, + 0.785, + 0.189 + ], + "angle": 0, + "content": "49. Shayan, M., Fung, C., Yoon, C.J., Beschastnikh, I.: Biscotti: A blockchain system for private and secure federated learning. IEEE Transactions on Parallel and Distributed Systems 32(7), 1513-1525 (2020)" + }, + { + "type": "ref_text", + "bbox": [ + 0.218, + 0.19, + 0.785, + 0.232 + ], + "angle": 0, + "content": "50. Shenaj, D., Toldo, M., Rigon, A., Zanuttigh, P.: Asynchronous federated continual learning. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 5054-5062 (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.232, + 0.785, + 0.259 + ], + "angle": 0, + "content": "51. Simonyan, K., Zisserman, A.: Very deep convolutional networks for large-scale image recognition. arXiv preprint arXiv:1409.1556 (2014)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.26, + 0.787, + 0.3 + ], + "angle": 0, + "content": "52. Stoychev, S., Churamani, N., Gunes, H.: Latent generative replay for resource-efficient continual learning of facial expressions. In: 2023 IEEE 17th International Conference on Automatic Face and Gesture Recognition (FG). pp. 1-8. IEEE (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.301, + 0.785, + 0.342 + ], + "angle": 0, + "content": "53. Sun, W., Yang, Z., Zhang, X., Liu, Y.: Energy-efficient neighbor discovery in mobile ad hoc and wireless sensor networks: A survey. IEEE Communications Surveys & Tutorials 16(3), 1448-1459 (2014)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.343, + 0.785, + 0.383 + ], + "angle": 0, + "content": "54. Tan, M., Le, Q.: Efficientnet: Rethinking model scaling for convolutional neural networks. In: International conference on machine learning. pp. 6105-6114. PMLR (2019)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.384, + 0.785, + 0.411 + ], + "angle": 0, + "content": "55. Wah, C., Branson, S., Welinder, P., Perona, P., Belongie, S.: The caltech-ucsd birds-200-2011 dataset (2011)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.412, + 0.785, + 0.467 + ], + "angle": 0, + "content": "56. Wei, K., Li, J., Ding, M., Ma, C., Yang, H.H., Farokhi, F., Jin, S., Quek, T.Q., Poor, H.V.: Federated learning with differential privacy: Algorithms and performance analysis. IEEE transactions on information forensics and security 15, 3454-3469 (2020)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.468, + 0.785, + 0.508 + ], + "angle": 0, + "content": "57. Yoon, J., Jeong, W., Lee, G., Yang, E., Hwang, S.J.: Federated continual learning with weighted inter-client transfer. In: International Conference on Machine Learning. pp. 12073-12086. PMLR (2021)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.509, + 0.785, + 0.536 + ], + "angle": 0, + "content": "58. Yoon, J., Yang, E., Lee, J., Hwang, S.J.: Lifelong learning with dynamically expandable networks. arXiv preprint arXiv:1708.01547 (2017)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.537, + 0.785, + 0.563 + ], + "angle": 0, + "content": "59. Yu, Q., Liu, Y., Wang, Y., Xu, K., Liu, J.: Multimodal federated learning via contrastive representation ensemble. arXiv preprint arXiv:2302.08888 (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.565, + 0.785, + 0.605 + ], + "angle": 0, + "content": "60. Yurochkin, M., Agarwal, M., Ghosh, S., Greenewald, K., Hoang, N., Khazaeni, Y. Bayesian nonparametric federated learning of neural networks. In: International conference on machine learning. pp. 7252-7261. PMLR (2019)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.606, + 0.785, + 0.646 + ], + "angle": 0, + "content": "61. Zhang, J., Chen, C., Zhuang, W., Lyu, L.: Target: Federated class-continual learning via exemplar-free distillation. In: Proceedings of the IEEE/CVF International Conference on Computer Vision. pp. 4782-4793 (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.647, + 0.785, + 0.688 + ], + "angle": 0, + "content": "62. Zhang, T., Mai, C., Chang, Y., Chen, C., Shu, L., Zheng, Z.: Fedego: privacy-preserving personalized federated graph learning with ego-graphs. ACM Transactions on Knowledge Discovery from Data 18(2), 1-27 (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.689, + 0.785, + 0.716 + ], + "angle": 0, + "content": "63. Zhu, L., Liu, Z., Han, S.: Deep leakage from gradients. Advances in neural information processing systems 32 (2019)" + }, + { + "type": "list", + "bbox": [ + 0.217, + 0.148, + 0.787, + 0.716 + ], + "angle": 0, + "content": null + } + ] +] \ No newline at end of file diff --git a/2024/Pick-a-back_ Selective Device-to-Device Knowledge Transfer in Federated Continual Learning/ecda4e75-a956-4af4-a55c-23aa50728c18_origin.pdf b/2024/Pick-a-back_ Selective Device-to-Device Knowledge Transfer in Federated Continual Learning/ecda4e75-a956-4af4-a55c-23aa50728c18_origin.pdf new file mode 100644 index 0000000000000000000000000000000000000000..3f66767c8cef1094eb79d0d88f54bc4f394b0c61 --- /dev/null +++ b/2024/Pick-a-back_ Selective Device-to-Device Knowledge Transfer in Federated Continual Learning/ecda4e75-a956-4af4-a55c-23aa50728c18_origin.pdf @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:bffc3fccb9a1cf71b747ac6b3eab3a0f615bf54ea75c1257c86b26d677c8c277 +size 2050103 diff --git a/2024/Pick-a-back_ Selective Device-to-Device Knowledge Transfer in Federated Continual Learning/full.md b/2024/Pick-a-back_ Selective Device-to-Device Knowledge Transfer in Federated Continual Learning/full.md new file mode 100644 index 0000000000000000000000000000000000000000..be142c46115af50bd4136992df933711f7f6d6d0 --- /dev/null +++ b/2024/Pick-a-back_ Selective Device-to-Device Knowledge Transfer in Federated Continual Learning/full.md @@ -0,0 +1,259 @@ +# Pick-a-back: Selective Device-to-Device Knowledge Transfer in Federated Continual Learning + +JinYi Yoon1 and HyungJune Lee2 + +$^{1}$ Virginia Tech, Blacksburg, VA, USA jinyioon@vt.edu + $^{2}$ Ewha Womans University, Seoul, South Korea hyungjune.lee@ewha.ac.kr + +Abstract. With the explosion of edge intelligence, leveraging federated indirect knowledge has become crucial for boosting the tasks of individual learners. However, the conventional approach to knowledge reuse often leads to catastrophic forgetting issues. In this paper, we revisit the concept of continual learning in the context of edge intelligence and address the knowledge transfer problem to enhance federated continual learning. Since each learner processes private heterogeneous data, we propose Pick-a-back, a device-to-device knowledge federation framework by selectively reusing the external knowledge with similar behavioral patterns. By borrowing indirect experiences, an edge device can initiate learning from useful knowledge and thus achieve faster yet more generalized knowledge acquisition. Using continual tasks consisting of various datasets on lightweight architectures, we have validated that Pick-a-back provides a significant inference improvement of up to $8.0\%$ via selective knowledge federation. Our codes are available at https://github.com/jinyi-yoon/Pick-a-back.git + +Keywords: Federated continual learning, edge intelligence, device-to-device knowledge federation, selective knowledge transfer + +# 1 Introduction + +As artificial intelligence (AI) permeates nearly every field, data has emerged as a pivotal source driving remarkable advancements. Most of the data originates from the end users or systems, stimulating a need for distributed learning on low-end edge devices to address privacy and latency concerns [2,31]. However, individual low-end edge devices often lack the computational capability to generate advanced knowledge [6]. In the context of human learning paradigms, where any individuals become educated and exchange helpful knowledge with one another, the emergence of edge intelligence and its widespread availability allow external knowledge to serve as a valuable resource in supporting individual learners. By facilitating some sharable insights [28,59] and harnessing collective intelligence [14], it has widely revolutionized learning paradigms across various fields, even where numerous data have non-identical forms and heterogeneous distributions. + +While leveraging external knowledge to enhance individual tasks, conventional learners often encounter the challenge of catastrophic forgetting, wherein they adapt only to the assigned tasks, and consequently lose previously acquired knowledge in some degree over time. In contrast, to more effectively retain and reinforce pre-trained knowledge together with new knowledge, the learners would rather adopt a continual learning approach for ensuring both robustness and adaptability. To enhance the efficacy of individual continual learning, it is not imperative for learners to confine themselves to their own knowledge. The prevalence of edge intelligence may facilitate knowledge federation, enabling individual learners to leverage a broader or deeper understanding of data, which can be possibly heterogeneous. While recent works have devised federated continual learning frameworks, they typically take a centralized approach for assistance [3,18,50], rather than direct device-to-device knowledge transfer in a distributed manner. + +As for edge AI, individual learners might be confronted with a series of diverse heterogeneous tasks. At the edge side accompanied with its own private data, each learner attempts to form its own distinct knowledge. Then, a question to raise is that how can an individual edge learner effectively harness externally available prior knowledge? It has been reported that there exists unique positive and negative knowledge at each model [43,57]. Maybe the answer would be straightforward yet meaningful – the learners would selectively reuse or customize some knowledge that aligns well with their specific behavioral patterns. Rather than starting from scratch, learners can benefit from the indirectly attained knowledge formed by others who have already seen similar patterns. In that the decision patterns can not be exactly the same, indirect knowledge transfer (with follow-up training) may offer a rich source of otherwise untrainable and unattainable knowledge from private data. This can lead to performance improvement via selective acquisition of some common yet distinct knowledge. + +In this work, we revisit the knowledge reuse problem. We consider three challenges in edge intelligence with knowledge transfer: 1) edge devices have limited computational and storage resources; 2) learners are not willing to disclose their raw data; and 3) edge devices should take into account the reduction of communication costs. To overcome these concerns, our approach aims to enhance the performance of lightweight learning models, while maintaining data privacy and avoiding additional communication overhead. + +We propose a selective and customized knowledge federation framework from one device to another, based on heterogeneous continual learning, called Pick-a-back, which picks a relevant backbone knowledge from other devices. Beyond the achievable knowledge of each individual learner, we strategically search for some effective knowledge sources selectively and then customize the commonly shared knowledge to closely align with the local task. By leveraging decision similarity based on mutated data to assess the efficacy of the backbone among the neighboring models spread in the form of physically separated edge intelligence, we selectively fetch and customize constructive knowledge. This approach not only enables decentralized knowledge exchange among the learners at devices, but also facilitates a more efficient and scalable learning paradigm. + +This work contributes to advancing the field of deep learning in edge intelligence by providing selective knowledge transfer and fostering collaboration among local learners. It can thereby enhance the efficiency of the individual continual learning process. The main contributions can be summarized as follows: + +1) We address the problem of device-to-device knowledge federation in the continual learning. We demonstrate the potential of leveraging indirect knowledge extracted by the neighboring edge learners to obtain otherwise untrainable and unattainable insights. This work provides a way to maintain a model with heterogeneous data collected across a network of edge devices. +2) As the diverse tasks imply their own unique knowledge, we propose selective and customized knowledge transfer by capturing the decision pattern similarity relevant to a target task running on a device. Pick-a-back adopts each relevant knowledge for a specific task, achieving high fidelity through the acquisition of some common yet distinct knowledge from unseen examples, while preserving data privacy and communication cost at the edge. +3) Extensive experimental results on various datasets across different architectures have shown that Pick-a-back provides a significant inference improvement up to $8.0\%$ via selective knowledge federation. + +# 2 Related Work + +Typical continual learning refers to training on a series of multiple tasks [22] or incremental classes [13]. The primary objective is to address the catastrophic forgetting problem, where previously learned knowledge is decaying over time as adapting to new tasks, using memorization [40,52] or regularization [9,23]. + +Knowledge-reusable continual learning. More closely related to our problem, there has been a growing interest in reusing existing knowledge for its subsequent tasks [30, 58]. Unlike conventional approaches, these methods have focused on concatenating a new model to its previous model, allowing input data to propagate through the previous model [46]. Rather than reusing the whole model, researchers have proposed to activate only the essential parts of the model [19,29]. In *Piggyback* [29], they leverage a common model and mask out unnecessary neurons to extract relevant knowledge for a target task. It effectively reuses existing knowledge with only a single sharable model, but it is less able to accept incremental tasks. *CPG* [19] combines neuron expansion with neuron pruning and masking techniques to address the challenges of forgetting, model compactness, and achieving higher fidelity in knowledge reuse. We adopt these knowledge reusable continual learning approaches to retain and extract valuable insights from previous tasks, while aiming to find useful knowledge (for reuse) based on model similarity. Although some works have applied some similarity metrics such as cosine similarity to federated learning scenarios [20], it is limited in applicability when dealing with clients having a different number of outputs (classes). In contrast, our approach provides flexibility in output heterogeneity, ensuring broader applicability in real-world scenarios, where most clients may + +have different shapes of outputs. Moreover, our decision pattern-based similarity is novel in federated continual learning, as the knowledge itself is not exactly the same even with the similar decision patterns. This enables private learners to be exposed to similar yet unseen broader knowledge, leading to improved performance and efficiency in subsequent learning tasks. + +Federated continual learning. With widespread accessibility of edge AI, learners do not necessarily be confined to their own private knowledge, as a standalone learner is permitted to only access its own local information and thus suffer from restricted experiences to some extent. Although recent works tackle the problem of federated continual learning [50,61], they focus on class-incremental learning, whereas we tackle task-incremental learning. In task-incremental learning, transfer learning turns out to be very effective, where a model is pretrained based on the source domain and then finetuned for the target domain. This selective device-to-device knowledge transfer can be a significant impetus to enhance model fidelity, also in continual learning. It has been commonly addressed in the field of federated continual learning but only in a centralized manner [18,50]. It has been also observed that there is both positive and negative knowledge related to a specific target task [43,57], highlighting the potential harm of static knowledge selection. To address the two-sided nature of knowledge reuse, FedWeIT [57] aggregates the parameters from clients as global knowledge, and then the clients mask the federated parameters adaptive to the specific task. However, previous works including FedWeIT rely on a central server, causing communication bottlenecks and privacy limitations. Furthermore, ImageNet-pretrained models have widely been utilized in transfer learning. However, it is not feasible for edge-side federated continual learning, as pretrained models are specifically designed for well-known large-scale architectures. The architectures in continual learning are dynamically changing, and edge AI cannot be tailored to large-scale models. In Pick-a-back, we rather explore decentralized, communication-efficient, and privacy-preserving knowledge transfer in lightweight on-device models. + +# 3 Selective Knowledge Transfer + +We consider the problem of device-to-device knowledge federation to enhance individual continual learning in a distributed manner. We propose Pick-a-back, a simple yet effective knowledge federation approach by discovering a relevant neighboring device who has acquired the knowledge of relevant tasks and then selectively infusing it into the model via continual learning. By leveraging a federated learning framework, which allows collaborative learning across devices, we aim to address the challenge of distributed continual learning, while also facilitating the reuse of previously acquired knowledge at each different device. The integration of these two fields of study enables efficient knowledge selection, sharing, and expansion among devices (with heterogeneous data), yielding continuous learning improvement and adaptation across tasks or learning scenarios. + +![](images/9d33422866b30461ddd6c6b21b8acaa4f2eb16d106e360a8320233fc715468c0.jpg) +Fig. 1: Overall architecture of Pick-a-back: (1) Each device trains an individual continual learning model on a series of tasks. (2) Given these available neighboring knowledge, a device selects a relevant neighboring model that can improve learning performance by acquiring indirect knowledge from the federated but physically separated devices. (3) On top of the retrieved knowledge, (4) the continual learner stacks new local tasks and resumes learning and federation. By fetching the generalized knowledge with relevant decision patterns, Pick-a-back offers a cumulative yet selective knowledge federation with heterogeneous data at a distinct device + +# 3.1 Problem Formulation + +Given $L$ devices, each learner $l \in \{1, 2, \dots, L\}$ at device $d_l$ processes its own sequential tasks (i.e., training data) of $T^{(l)} = \{T_1^{(l)}, T_2^{(l)}, \dots \}$ . A scratch model $M_i^{(l)}$ is constructed by training only with private data $T_i^{(l)} = \{(X_i^{(l)}, Y_i^{(l)})\}$ , where $T_i^{(l)}$ for task $i$ and device $d_l$ is heterogeneous across other tasks and devices. We train the model $\hat{M}_i^{(l)}$ with target data $T_i^{(l)}$ on top of its previous version $\hat{M}_{i-1}^{(l)}$ with some learning parameters $\theta_i^{(l)}$ , i.e., $\hat{M}_i^{(l)} = \hat{M}_{i-1}^{(l)}(T_i^{(l)}, \theta_i^{(l)})$ in a continual manner. Here, at every task $i$ , the goal is to find a good reference model that can maximize the overall fidelity of model $\hat{M}_i^{(l)}$ . + +# 3.2 Architecture + +Our learning architecture incorporates two distinct modes (Fig. 1): i) continual learner for local training; and ii) federated learner for knowledge transfer. Initially, multiple learners train independently on a specific private task. Subsequently, one learner at a device transitions to a federated learning mode to explore effective indirect knowledge from a neighboring relevant model. Once a neighboring model is selected, we use the selected model as a backbone, the main target task is then executed in a continual manner. By incorporating the retrieved knowledge together with its existing own knowledge, the learner can construct cumulative knowledge in a continual manner without forgetting. This process of continual learning and federated learning is repeated for each task. + +Continual learner. In our approach to continual learning, we adopt task-incremental learning, where a series of tasks is processed sequentially. Each task has a distinct learning objective, and a local model is trained independently on each task with freezing or isolating the models for previous knowledge. By employing task-oriented learning, we aim to mitigate the catastrophic forgetting problem, where previously learned knowledge vanishes as training on new tasks or incorporating new data over time. Our approach can be combined with any continual learner that has gradually been constructing knowledge. It is fundamentally different from conventional continual learning that simply memorizes multiple models or data, which do not leverage the given knowledge. We suppose that a continual learner concatenates a new model on top of the previously acquired knowledge, and thus it can reduce memory usage and also enhance the model fidelity. This sustainability of the continual model is particularly valuable in the context of edge AI and aligns with the goals of the recent learning paradigm. + +We pursue reusing the pre-configured knowledge by using the model. We mainly use the latest continual learning effort, $CPG$ [19], which expands the model when necessary, prunes it for compactness, and reuses the network by masking out specific ineffective neurons. Inspired by the masking strategy of Piggyback [29], it samples some necessary neurons for the target task to allow the extraction of common knowledge. $CPG$ only trains the learnable binary mask whether to activate each neuron for the target task. As it ensures the retention of previous knowledge via freezing, the results remain exactly the same even when training new tasks. However, without loss of generality, this masking strategy can be applied to any continual learner that holds multiple knowledge. + +Federated learner. Due to the privacy and communication overhead, distributed learners are unwilling to share private raw data. We thus adopt a model-sharing approach rather than data-sharing. Here, each client is regarded as an independent learner. When each learner at a distinct device is working on a distinct task but lacks knowledge, there might be some other knowledge that has a similar, broader, or deeper experience beyond the learner based on its own local data. We assume that clients are knowledgeable with their own currently available neighboring models via existing neighbor selection methods [38, 53] or graph construction approaches [33, 62]. These neighboring diverse knowledge sources have been trained on different datasets, capturing various perspectives and feature extraction capabilities. By exposing the learner to external knowledge, the model is capable of leveraging enhanced feature representations, enabling it to capture a broader set of patterns and information that may be beneficial for its specific learning objective. The learner can also develop a more comprehensive understanding of the data, generalize better to unseen examples by mitigating overfitting, and thus improve the fidelity of the model on a variety of tasks and data. By leveraging a diverse ensemble of knowledge sources, the learner can benefit from its complementary expertise, leading to accuracy improvement. + +Each learner at a device seeks relevant knowledge that can contribute to further improving its performance based on local continual learning. To acquire necessary constructive knowledge, we identify effective indirect experiences based + +on decision pattern similarities. Once we have obtained knowledge from other learners, we accumulate ongoing new local tasks on top of the model in a continual manner, ensuring that the retrieved knowledge becomes infused with local knowledge. Our main focus is on the current target task, and thus, if we want to retain the knowledge from our previous tasks, we have to either store the existing knowledge separately or concatenate it with the fetched model. If the currently constructed knowledge turns out to be sufficiently effective, there may be no further need to incorporate further knowledge from other learners. + +# 3.3 Selective Knowledge Federation + +Each learner is responsible for distinct tasks, and tasks may exhibit unique strengths, weaknesses, or biases. Consequently, each task at a local learner plays its own unique different role and forms distinct knowledge. Some external models may have learned other representations that have captured specialized knowledge relevant to specific tasks. Therefore, selective knowledge transfer can be a very effective way for enhancing the performance of each student model at an edge device. Our approach is simple and straightforward – if models have similar or synergetic capabilities in classifying different samples, they are likely to have potential sources of common but more generalized knowledge. Since it has the possibility to attain broader intuition but also inexperienced knowledge, this indirect knowledge characterized by similar decision patterns may greatly accelerate and improve the learning for individual learners. + +As the similarity of model do not straightly incur similar knowledge, we focus on the decision pattern similarity rather than the model itself. We borrow an idea to calculate the model similarity from ModelDiff [27], which calculates the similarity of decision boundaries by comparing the reaction to the input data. Initially, we sample input data $x_{i}$ from the requester and adversarial data $x_{i}'$ from a neighbor. To address privacy concerns, an alternative approach may sample data only from the requester, or the participants may opt to exchange mutated data, thus avoiding the sharing of private data. Then, we inject a pair-wise set of input data, $\{(x_{1},x_{1}'),(x_{2},x_{2}'),\dots ,(x_{n},x_{n}')\}$ , into each model $f$ and $g$ and then computes the decision distance vector (DDV), $DDV_{f} = \{d_{1},d_{2},\dots ,d_{n}\}$ , which is a distance vector for the output pair, where $d_{i} = \text{distance}(f(x_{i}),f(x_{i}'))$ , $i\in \{1,2,\ldots ,n\}$ . Finally, it compares the cosine similarity between the distance vectors of models, $sim(DDV_{f},DDV_{g})$ . Since it focuses the decision similarity instead of directly comparing the outputs, it is applicable to any heterogeneous task with a different number of output classes. To search a pair-wise set that can approximate the decision boundary most effectively, the input data is mutated, and similarities are iteratively computed, similar to genetic algorithms. + +By incorporating it in Pick-a-back, when we search for a similar model for task $T_{i}^{(l)}$ of learner $l$ among $\hat{M}_{i-1}^{k}$ of all neighboring learner $k$ . We sample an input pair-wise data $X_{i}^{(l+k)}$ and $X_{i}^{\prime(l+k)}$ from $X_{i}^{(l)} \cup X_{i-1}^{(k)}$ (or just from $X_{i}^{(l)}$ ). By injecting $X_{i}^{(l+k)}$ and $X_{i}^{\prime(l+k)}$ into the learner model $M_{i}^{(l)}$ and the neighboring models $\hat{M}_{i-1}^{(k)}$ , we produce the distance vector of outputs $D_{M_{i}^{(l)}}$ and $D_{\hat{M}_{i-1}^{(k)}}$ . Then, + +Algorithm 1 Federated Continual Learning with Pick-a-back +1: Input: $L$ : # of devices, (for each device $d_{l}$ ) $M_{0}^{(l)}$ : random initial model, $T^{(d_{l})}$ : a sequence of tasks, $\theta_{i}^{(l)}$ : learning parameters +2: for task $i = 1: length(T^{(l)})$ do +3: Device $d_{l}$ trains task $i$ such that $M_{i}^{(l)} = M_{i-1}^{(l)}(T_{i}^{(l)}, \theta_{i}^{(l)})$ ; $\triangleright$ Continual Learner +4: for device $k = 1: L$ do $\triangleright$ Federated Learner +5: Compute the similarity, $sim(M_{i}^{(l)}, \hat{M}_{i-1}^{(k)})$ ; +6: end for +7: Select and retrieve knowledge $\hat{M}_{i-1}^{(l)} \gets \hat{M}_{i-1}^{(k)}$ such that $\arg \max_{k} sim(M_{i}^{(l)}, \hat{M}_{i-1}^{(k)})$ ; +8: Train task $i$ on the fetched model, $\hat{M}_{i}^{(l)} = \hat{M}_{i-1}^{(l)}(T_{i}^{(l)}, \theta_{i}^{(l)})$ ; $\triangleright$ Continual Learner +9: end for + +we select the most similar model as follows: + +$$ +\underset {k} {\operatorname {a r g m a x}} \operatorname {s i m} \left(D D V _ {M _ {i} ^ {(l)}}, D D V _ {\hat {M} _ {i - 1} ^ {(k)}}\right) \tag {1} +$$ + +We fetch the indirect knowledge that has already experienced the target decision pattern. After selecting the exterior knowledge, we continue to learn $T_{i}^{l}$ on $\hat{M}_{i - 1}^{(k)}$ . + +Training process. Each learner $l$ trains $i$ -th task $T_{i}^{(l)}$ and constructs a local model $M_{i}^{(l)}$ with continual learning. Then, the learner $l$ checks if there exists any helpful exterior knowledge by comparing $M_{i}^{(l)}$ with $M_{i-1}^{(k)}$ for all neighboring edge learners $k$ . After selecting relevant knowledge for knowledge transfer, device $d_{l}$ processes the $i$ -th task $T_{i}^{(l)}$ over $\hat{M}_{i-1}^{(k)}$ and constructs $\hat{M}_{i}^{(l)} = \hat{M}_{i-1}^{(k)}(T_{i}^{(l)}, \theta_{i}^{(l)})$ . + +The overall procedure is described in Algorithm 1. + +# 4 Experiments + +We have validated Pick-a-back extensively on four different classification task groups with 13 public data using four different well-known architectures. Each group is considered as one experimental environment where knowledge can be shared, and one superclass group is a single task. + +Datasets. We used four different groups of CIFAR-100 [25], Classified-ImageNet [8], Fine-Grained [10,24,37,48,55], and Digits [7,11,35,36,39,45]. The organization of the dataset and the details are provided in Appendix Sec. A1. + +Architectures. We have presented the results using both lightweight models and large-scale models. For the lightweight models, we used LeNet-5 [26] and MobileNetV1 [17], where ImageNet-pretrained models are not publicly available. On the other hand, VGG16-BN [51] and EfficientNet-B0 [54] that have publicly available ImageNet-pretrained models were implemented using PyTorch [42]. Details of model configurations and learning parameters are in Appendix Sec. A2. We used CIFAR-100 on LeNet-5 as default, unless otherwise noted. + +![](images/a659d8c9b10ab0113ab026bad83487f231c600742db27fbf8aecc938d719c511.jpg) +(a) Absolute Acc. + +![](images/fd37e59ee1bbd2997524d31a6964ca28ab86a0c0fc8db682f94784de713d3ab9.jpg) +(b) Improved Acc. + +![](images/85cab9c58ed596d781e76927c2745f641eda0efdd4b085cf3447279db225d8b2.jpg) +(a) Training dynamics +Fig.2: Heatmap of the accuracy for various backbone and target tasks sorted by the accuracy of $w / o$ Backbone. The first row represents the averaged values for all tasks, while the first column shows the results of $w / o$ Backbone. Absolute accuracy is scaled per each row, and accuracy improvement is scaled across all results +Fig. 3: Training dynamics for the most effective case of the target task 14 using the backbone task of 4 and accuracy improvement with respect to the smoothed accuracy of $w / o$ Backbone. The smoothed accuracy refers to the averaged accuracy for a subset of tasks with lower accuracy. For example, the second x-value represents the average of two tasks with the lowest performance, while the last value is the average across all tasks + +![](images/9cd885a93119920e92c63bdbe08a52b893464870a5afabfbc16386bb638054b3.jpg) +(b) Accuracy improvement + +Baselines. We compare Pick-a-back with diverse learning approaches of federated learning, continual learning, and continual federated learning. + +First, we let our Pick-a-back run on top of a certain continual learning scheme: 1) ProgressiveNet [47]: continual training only with expansion by concatenating scratch models for sequential tasks; 2) PackNet [30]: continual training only with pruning by allocating partial neurons exclusively to each task; 3) CPG [19]: continual training with expansion if necessary and pruning for the model compactness; and 4) WSN [21]: continual learning with reusing the masked weights of the prior subnetworks. + +We further compare across existing federated and continual federated learning approaches: 1) FedAVG [32]: one of the pioneering works for model regularization in federated learning; 2) FedWeIT [57]: a federated continual learning paradigm by introducing sparse task-specific parameters along with global federated parameters; and 3) FCCL [18]: a federated continual learning approach to tackle the catastrophic forgetting problem within a heterogeneous setting. + +We also compare Pick-a-back with different knowledge selection strategies: 1) $w / o$ Backbone: intact continual learning without using any selective knowledge transfer, such as Progressive Networks, PackNet, or CPG; 2) Uniform Knowledge: applying the same task to its backbone model and averaging all of the results for each task; 3) Internal Knowledge: training the target task iteratively as a backbone model; 4) Random Knowledge: randomly selecting a backbone model; and 5) Upper: selecting an ideal backbone with a certain task that shows the highest accuracy improvement, being served as an unachievable upper bound. + +Measures. We have investigated 1) Avg: the averaged accuracy of all tasks and 2) Top: the accuracy of the most effective task that has gained the best accuracy improvement via selective knowledge. To compare the feasibility of models for edge AI, we have shown 1) Model size: the actual storage overhead of a model; 2) Expansion: the expansion degree where the scratch model has the value of 1; + +Table 1: Performance comparison. The bold text highlights the best performance per client, excluding Upper as an ideal upper-bound due to its infeasible brute-force manner + +
Client ID1511716181231476824131951092011Avg
w/o Backbone40.055.057.260.451.866.067.467.669.270.872.673.074.275.275.876.477.077.481.284.069.1
FedAVG24.431.634.638.032.039.631.832.049.037.829.834.026.430.835.845.427.223.435.639.033.9
ProgressiveNet44.658.854.463.064.468.071.268.269.270.673.872.470.473.074.078.877.273.684.683.869.7
PackNet38.055.056.459.463.869.068.272.271.272.874.472.070.274.879.678.078.676.882.680.269.7
WSN34.253.849.453.459.860.263.663.066.865.064.665.466.662.661.267.666.658.674.079.861.8
FedWeIT40.052.555.056.056.260.556.760.567.065.260.265.066.064.265.065.766.751.068.777.861.0
FCCL37.054.452.460.662.268.267.264.468.270.672.872.468.477.073.875.078.668.879.278.267.5
Uniform41.558.657.662.165.670.769.771.671.272.574.773.073.776.477.178.779.475.283.583.070.8
Internal43.255.456.862.267.867.069.071.473.874.876.871.873.476.279.277.681.079.282.083.871.1
Random39.859.056.659.267.672.668.273.071.471.475.672.874.476.478.078.277.874.684.283.870.7
Upper46.462.460.064.267.874.872.275.673.875.677.075.076.079.480.080.681.679.286.285.473.7
Pick-a-back46.459.060.063.264.472.671.875.672.472.277.073.072.876.276.080.278.275.884.285.471.8
+ +3) Sparsity: the released ratio of a model, where higher sparsity indicates more space for next tasks; 4) Training time: the execution time to run for one epoch; and 5) Used data: the number of data samples used for training. + +# 4.1 Overall Performance and Comparisons + +Motivations for selective knowledge federation. We first examined the impact of different knowledge on various target tasks. As in Fig. 2a, a significant portion of the results and the average accuracy achieved through backbone usage are represented by a blue color, indicating that the continual learner benefited from the knowledge federation. However, there exist not only positive backbones but also negative backbones, suggesting a need for selective knowledge to assist a continual learner specific to its target task. Taking a closer look at the accuracy improvement relative to $w / o$ Backbone in Fig. 2b, a well-constructed model such as task 9 or 11 is more likely disturbed by external knowledge. On the other hand, a target task with initially poor performance tends to gain a greater improvement in accuracy. It implies the importance of selectively incorporating constructive knowledge to enhance the learning process for a specific target task. + +Ablation study on selective knowledge federation. Inspired by the Fig. 2 we verified the effectiveness, compared to diverse learning approaches, including FedAVG, ProgressiveNet, PackNet, WSN, FedWeIT, and FCCL, along with various backbone selection strategies of Uniform, Internal, Random, and Upper, as in Table 1 and in Table A5. We consistently demonstrate superior fidelity on average. By employing our selection approach, we observed a significant accuracy improvement of up to $8.0\%$ , with only two negative cases out of 20 tasks. Since FedAVG simply averages the model parameters without considering the continual tasks, it suffers from losing the knowledge of trained tasks. Notably, although Pick-a-back with CPG shows the best performance, Pick-a-back with other continual learning approaches of ProgressiveNet and PackNet also outperforms + +$w / o$ Backbone. Even Uniform, Internal, and Random selection strategies exhibit enhanced performance, suggesting the efficacy of inter-device knowledge transfer based on Pick-a-back primitive. We also compared with other recent (federated) continual learning approaches (e.g., WSN, FedWeIT, and FCCL), showing that Pick-a-back outperforms these methods. This is due to the effectiveness of our knowledge reuse approach along with our backbone selection strategy. + +Benefit of backbone knowledge. Taking a closer look at how Pick-a-back benefits from backbone knowledge, in Fig. 3a, it is evident that the test accuracy of Pick-a-back consistently surpasses the training accuracy of $w / o$ Backbone, starting from the beginning until the saturation. It indicates that indirect knowledge assists individual learners to be exposed to inexperienced knowledge and adapt to unseen data. Even within Pick-a-back itself, the test accuracy outperforms the training accuracy in the early stages of training, suggesting that the transferred knowledge has relevance to the target knowledge and drives toward high performance even earlier. Furthermore, $w / o$ Backbone achieves an accuracy of $67.4\%$ , which can be attained only within approximately 20 epochs using Pick-a-back, whereas it takes over 50 epochs without knowledge transfer. This demonstrates that Pick-a-back not only accelerates the learning, but also enhances the overall performance. Please refer to the empirical convergence analysis in Appendix Fig. A1. + +Possibility of similar but generalized knowledge. To justify the task-specific selection, we compared to uniform knowledge, where the same backbone is applied to all tasks. As in Fig. 3b, the results showed that Upper exhibits better improvement in tasks with poor initial performance. However, the uniform knowledge ignores task-specific improvements, leading to similar improvements across all target tasks. Interestingly, we also observed that only using internal knowledge, where the target task is trained twice, allowed further improvement. It means that even with a single task iteratively training by accumulating knowledge on top of an identical model turns out to induce better knowledge. It also implies that a backbone task with a similar, ideally the same in personal knowledge case, the decision boundary can further improve the knowledge. While the same task aims for identical knowledge, using similar but distinct knowledge enables generalization and better improvement beyond individual tasks. As expected, Pick-a-back outperformed the one only with internal knowledge, highlighting the advantage of leveraging similar indirect experiences. This means that we provide a more opportunity to acquire superior knowledge by mitigating the overfitting problem and extending to the unseen data beyond local knowledge. + +# 4.2 Scalability Analysis + +Lightweight models for edge AI. We evaluated Pick-a-back in various architectures from the lightweight models of LeNet-5, MobileNet-V1, and EfficientNet-B0[2] + +Table 2: Accuracy, model size, and averaged training time per one epoch with respect to different architectures on CIFAR-100. We averaged the training time over 10 epochs + +
AccuracyModel SizeTraining Time
AvgTop
VGG16-BNw/o Backbone78.968.2164 MB4.10 s
Pick-a-back81.0 (+2.1)76.6 (+8.4)
EfficientNet-B0w/o Backbone77.844.620 MB9.29 s
Pick-a-back81.8 (+4.0)54.4 (+9.8)
MobileNet-V1w/o Backbone65.459.816 MB4.95 s
Pick-a-back66.7 (+1.3)65.4 (+5.6)
LeNet-5w/o Backbone69.167.60.3 MB2.18 s
Pick-a-back71.8 (+2.7)75.6 (+8.0)
+ +Table 3: Various datasets with LeNet-5 + +
DigitsCIFAR-100Classified-ImageNetFine-Grained
AvgTopAvgTopAvgTopAvgTop
Upper95.889.473.775.661.471.833.948.5
w/o Backbone93.880.769.167.656.159.225.021.0
Pick-a-back95.186.471.875.658.167.628.327.5
(+1.3)(+5.7)(+2.7)(+8.0)(+2.0)(+8.4)(+3.3)(+6.5)
+ +to the large-scale models of VGG16-BN in Table 2. It indicates that all models benefit from selectively transferred knowledge, leading to accuracy improvement of up to $8.4\%$ , $9.8\%$ , $5.6\%$ , and $8.0\%$ , respectively. However, in resource-constrained environments, the limitations of memory and processing power make it challenging to deploy a large-scale model at the edge. To make learning practical in the real-world, lightweight models can be flexibly deployed across multiple edge devices, reducing dependence on remote servers. The lightweight LeNet-5 has a much smaller size of 0.18 times compared to VGG16-BN, enabling faster training and inference time. Although the lightweight models may not match closely to the large-scale models, Pick-a-back can still achieve improvement, comparable to the large models, while greatly reducing the storage requirement and run-time. + +Scalability across diverse datasets. We assessed the scalability of Pick-a-back across various datasets in Table 3. We stably achieve improvements across different datasets, even with varying degrees of class complexity. In Digits, where most tasks except for the Roman task are generally easy to classify, the average accuracy improvements may appear modest, but Pick-a-back approach achieved a significant accuracy improvement of up to $5.72\%$ in the relatively challenging Roman task. Similarly, in the classified-ImageNet dataset, where there exist enough neighboring tasks to provide additional knowledge, Pick-a-back shows substantial improvements even for the underperforming tasks. However, when considering fine-grained datasets with a large number of classes per task (ranging $102\sim 250$ classes), classification becomes inherently more difficult with simple models, resulting in relatively lower accuracy. Nonetheless, by applying Pick + +Table 4: Performance on Fine-Grained Table 5: Heterogeneous data sampling datasets using EfficientNet-B0 with $\alpha = 0.5$ + +
Client ID12345AvgHeteroBalancedw/o Sampling
AvgTopAvgTopAvgTop
Pretrained w/o Backbone28.990.542.449.937.149.8Upper59.360.269.684.473.775.6
66.37.428.212.724.6w/o Backbone52.640.264.664.171.869.167.6
Pick-a-back49.5 (+41.2)77.4 (+11.1)44.0 (+36.6)46.4 (+18.2)64.6 (+51.9)56.4 (+31.8)Pick-a-back56.1 (+3.5)58.6 (+18.4)67.3 (+3.2)80.0 (+8.2)71.8 (+2.7)75.6 (+8.0)
+ +$a$ -back, it is possible to advance the knowledge and achieve performance gains. While there is a noticeable performance gap of $0.7\%$ , $1.9\%$ , and $3.3\%$ , respectively, between Pick-a-back and Upper in other datasets, the gap in the fine-grained dataset is quite large with $5.6\%$ on average and up to $21.0\%$ for the task of flowers. This implies the potential of Pick-a-back to drive further improvements in complex tasks that require more sophisticated classification. We further explore the performance with the different number of clients (i.e., classes) and the different learning seeds in Appendix Table A9 and in Appendix Table A10, respectively. + +Impact on the most complicated task. We investigate the most complicated tasks of CUB 55, Stanford Cars 24, Flower 37, WikiArt 48, and Sketch 10, with each featuring a diverse range of class numbers of 200, 196, 102, 195, and 250 and running at each separate client. As shown in Table 4 with the adoption of Pick-a-back, we observe remarkable improvements in overall accuracy, averaging at $31.8\%$ , and peaking at an impressive $51.9\%$ . Although these results might appear surprising, they align with the findings previously established by 19 and 29, where scratch accuracy on ImageNet-pretrained models of ResNet50 and VGG16-BN, achieved $58.83\%$ and $43.84\%$ , respectively. Incorporating Pick-a-back not only surpasses the average accuracy of ImageNet-pretrained models, but does so through a judicious infusion of tailored knowledge. It is worth noting that ImageNet-pretrained models often necessitate extensive prior training data. However, in many instances, the efficacy diminishes due to the lack of distinct knowledge selection, showing some critical drawbacks in clients #1, #3, and #5. + +Superiority in heterogeneous data scenarios. In real-world scenarios, there are many unique challenges such as personalization [5] or heterogeneity [4]. Here, we explored the effectiveness of Pick-a-back under data heterogeneity. Leveraging the heterogeneity settings from federated learning researches [16, 60], we applied $\alpha$ of 0.5 based on the Dirichlet distribution $Dir(\alpha)$ . As in Table [5] we observed a degradation in the accuracy of scratch models as the data heterogeneity increases. It implies that the accuracy is achievable, and the external knowledge can be particularly beneficial for weaker classes in the presence of heterogeneous data. While the average accuracy improvement remained consistent, we achieved significant enhancements in heterogeneous data settings. It is essential to address unbalanced data distributions, as certain classes may receive less training attention. By leveraging Pick-a-back, we are able to achieve substantial performance gains, with improvements of up to $18.4\%$ . + +Table 6: Performance and data used for training in ImageNet-pretrained VGG16-BN + +
w/o Backbone → Pick-a-backSize of Training Data
AvgTop
w/ Pretrained84.9 → 83.882.8 → 84.492.8 GB
w/o Pretrained78.9 → 81.068.2 → 76.60.3 GB
+ +# 4.3 Discussions and Limitations + +ImageNet-pretrained models. It is important to note that a lightweight model such as LeNet-5 may not outperform a large-scale model even with Pick-a-back. Also, VGG16-BN models initialized with ImageNet-pretrained weights demonstrate strong performance generally for visual tasks as in Table 6 and in Appendix Table A6. However, their effectiveness heavily relies on the availability of large-scale labeled datasets that need to be uploaded to a central server. Additionally, the pretrained models are specifically designed for image classification, and their performance may vary when applied to specific or specialized tasks. Furthermore, the availability of pretrained models is limited only to well-known large-scale architectures, making their deployment usually infeasible to edge learning. Without pretrained model, selective knowledge federation through Pick-a-back achieves similar performance on average with a small gap of $3.9\%$ compared to the original pretrained model (w/o Backbone), but only with $0.3\%$ of training data. + +We further provided the orthogonality of continual learner in Pick-a-back and catastrophic forgetting prevention across different continual learning approaches in Appendix Table A11. Lastly, we need image sharing among clients and model sharing from a selected neighboring client, which may incur transmission overhead and privacy issues. We thus further assess the communication and computation time and how we can defend against attacks in Appendix Table A12 and Fig. A2. + +# 5 Conclusion + +We propose Pick-a-back that picks a task-specific backbone knowledge and fosters the collective intelligence in federated continual learning. As the individual learner pursues each distinct goal with private heterogeneous data, we selectively transfer the relevant external knowledge to the target task between neighboring learners. + +As AI expands towards edges for scalability, low-end devices tend to be trained toward personalized and specialized but somewhat partially limited intelligence. Here, in case of a scarcity of data or with limited performance, it becomes essential to seek assistance from external sources. The question - how can a device make a local model more effectively trained with exterior knowledge? - can be answered: Pick-a-back offers an effective way to selective knowledge transfer from one device to another towards distributed edge intelligence. By harnessing the power of neighboring devices, individual learners can gain access to valuable insights and expertise via collective intelligence, augmenting their own capabilities. + +# Acknowledgments + +We thank the Virginia Tech Presidential Postdoctoral Fellowship Program for supporting JinYi Yoon's postdoctoral associate position. This work was supported by the Institute of Information & communications Technology Planning and Evaluation (IITP) grant funded by the Korea government (MSIT) (No. RS-2022-00155966, Artificial Intelligence Convergence Innovation Human Resources Development (Ewha Womans University) and No. RS-2021-II212068, Artificial Intelligence Innovation Hub). + +# References + +1. Arani, E., et al.: Learning fast, learning slow: A general continual learning method based on complementary learning system. ICLR (2022) +2. Cao, L.: Decentralized ai: Edge intelligence and smart blockchain, metaverse, web3, and desci. IEEE Intelligent Systems 37(3), 6-19 (2022) +3. Casado, F.E., Lema, D., Criado, M.F., Iglesias, R., Regueiro, C.V., Barro, S.: Concept drift detection and adaptation for federated and continual learning. Multimedia Tools and Applications pp. 1-23 (2022) +4. Chen, D., Gao, D., Kuang, W., Li, Y., Ding, B.: pfl-bench: A comprehensive benchmark for personalized federated learning. Advances in Neural Information Processing Systems 35, 9344-9360 (2022) +5. Chen, D., Gao, D., Xie, Y., Pan, X., Li, Z., Li, Y., Ding, B., Zhou, J.: Fs-real: Towards real-world cross-device federated learning. In: Proceedings of the 29th ACM SIGKDD Conference on Knowledge Discovery and Data Mining. pp. 3829-3841 (2023) +6. Chen, Y., Inaltekin, H., Gorlatova, M.: Adaptslam: Edge-assisted adaptive slam with resource constraints via uncertainty minimization. arXiv preprint arXiv:2301.04620 (2023) +7. Colianni, S.: Mnist as .jpg (2017), https://www.kaggle.com/scolianni/ mnistas.jpg +8. Deng, J., Dong, W., Socher, R., Li, L.J., Li, K., Fei-Fei, L.: Imagenet: A large-scale hierarchical image database. In: 2009 IEEE Conference on Computer Vision and Pattern Recognition. pp. 248-255 (2009). https://doi.org/10.1109/CVPR.2009.5206848 +9. Dhar, P., Singh, R.V., Peng, K.C., Wu, Z., Chellappa, R.: Learning without memorizing. In: Proceedings of the IEEE/CVF conference on computer vision and pattern recognition. pp. 5138-5146 (2019) +10. Eitz, M., Hays, J., Alexa, M.: How do humans sketch objects? ACM Transactions on graphics (TOG) 31(4), 1-10 (2012) +11. El-Sawy, A., El-Bakry, H., Loey, M.: Cnn for handwritten arabic digits recognition based on lenet-5. In: Proceedings of the International Conference on Advanced Intelligent Systems and Informatics 2016 2. pp. 566-575. Springer (2017) +12. Goodfellow, I., Bengio, Y., Courville, A.: Deep Learning. MIT Press (2016), http://www.deeplearningbook.org +13. Gu, Y., Yang, X., Wei, K., Deng, C.: Not just selection, but exploration: Online class-incremental continual learning via dual view consistency. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 7442-7451 (2022) + +14. Ha, D., Tang, Y.: Collective intelligence for deep learning: A survey of recent developments. Collective Intelligence 1(1), 263391372211114874 (2022) +15. Haim, N., Vardi, G., Yehudai, G., Shamir, O., Irani, M.: Reconstructing training data from trained neural networks. Advances in Neural Information Processing Systems 35, 22911-22924 (2022) +16. He, C., Li, S., So, J., Zeng, X., Zhang, M., Wang, H., Wang, X., Vepakomma, P., Singh, A., Qiu, H., et al.: Fedml: A research library and benchmark for federated machine learning. arXiv preprint arXiv:2007.13518 (2020) +17. Howard, A.G., Zhu, M., Chen, B., Kalenichenko, D., Wang, W., Weyand, T., Andreetto, M., Adam, H.: Mobilenets: Efficient convolutional neural networks for mobile vision applications. arXiv preprint arXiv:1704.04861 (2017) +18. Huang, W., Ye, M., Du, B.: Learn from others and be yourself in heterogeneous federated learning. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 10143-10153 (2022) +19. Hung, C.Y., Tu, C.H., Wu, C.E., Chen, C.H., Chan, Y.M., Chen, C.S.: Compacting, picking and growing for unforgetting continual learning. Advances in Neural Information Processing Systems 32 (2019) +20. Jeong, W., Hwang, S.J.: Factorized-fl: Personalized federated learning with parameter factorization & similarity matching. Advances in Neural Information Processing Systems 35, 35684-35695 (2022) +21. Kang, H., Mina, R.J.L., Madjid, S.R.H., Yoon, J., Hasegawa-Johnson, M., Hwang, S.J., Yoo, C.D.: Forget-free continual learning with winning subnetworks. In: International Conference on Machine Learning. pp. 10734-10750. PMLR (2022) +22. Kim, G., Esmaeilpour, S., Xiao, C., Liu, B.: Continual learning based on ood detection and task masking. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 3856-3866 (2022) +23. Kirkpatrick, J., Pascanu, R., Rabinowitz, N., Veness, J., Desjardins, G., Rusu, A.A., Milan, K., Quan, J., Ramalho, T., Grabska-Barwinska, A., et al.: Overcoming catastrophic forgetting in neural networks. Proceedings of the national academy of sciences 114(13), 3521-3526 (2017) +24. Krause, J., Stark, M., Deng, J., Fei-Fei, L.: 3d object representations for fine-grained categorization. In: Proceedings of the IEEE international conference on computer vision workshops. pp. 554-561 (2013) +25. Krizhevsky, A., Hinton, G., et al.: Learning multiple layers of features from tiny images (2009) +26. LeCun, Y., et al.: Lenet-5, convolutional neural networks. URL: http://yann.lecun.com/exdb/lenet 20(5), 14 (2015) +27. Li, Y., Zhang, Z., Liu, B., Yang, Z., Liu, Y.: Modeldiff: testing-based dnn similarity comparison for model reuse detection. In: Proceedings of the 30th ACM SIGSOFT International Symposium on Software Testing and Analysis. pp. 139-151 (2021) +28. Liu, F., Wu, X., Ge, S., Fan, W., Zou, Y.: Federated learning for vision-and-language grounding problems. In: Proceedings of the AAAI Conference on Artificial Intelligence. vol. 34, pp. 11572-11579 (2020) +29. Mallya, A., Davis, D., Lazebnik, S.: Piggyback: Adapting a single network to multiple tasks by learning to mask weights. In: Proceedings of the European Conference on Computer Vision (ECCV). pp. 67-82 (2018) +30. Mallya, A., Lazebnik, S.: Packnet: Adding multiple tasks to a single network by iterative pruning. In: Proceedings of the IEEE conference on Computer Vision and Pattern Recognition. pp. 7765-7773 (2018) +31. Mao, B., Liu, J., Wu, Y., Kato, N.: Security and privacy on 6g network edge: A survey. IEEE Communications Surveys & Tutorials (2023) + +32. McMahan, B., Moore, E., Ramage, D., Hampson, S., y Arcas, B.A.: Communication-efficient learning of deep networks from decentralized data. In: Artificial intelligence and statistics. pp. 1273-1282. PMLR (2017) +33. Meng, C., Tang, M., Setayesh, M., Wong, V.W.: Gnn-based neighbor selection and resource allocation for decentralized federated learning. In: GLOBECOM 2023-2023 IEEE Global Communications Conference. pp. 1223-1228. IEEE (2023) +34. Miller, G.A.: Wordnet: a lexical database for english. Communications of the ACM 38(11), 39-41 (1995) +35. Nazarpour, K., Chen, M.: Handwritten Chinese Numbers (1 2017). https://doi.org/10.17634/137930-3, https://data.ncl.ac.uk/articles/dataset/Handwritten_Chinese_Numbers/10280831 +36. Netzer, Y., Wang, T., Coates, A., Bissacco, A., Wu, B., Ng, A.Y.: Reading digits in natural images with unsupervised feature learning (2011) +37. Nilsback, M.E., Zisserman, A.: Automated flower classification over a large number of classes. In: 2008 Sixth Indian Conference on Computer Vision, Graphics & Image Processing. pp. 722-729. IEEE (2008) +38. Nishio, T., Yonetani, R.: Client selection for federated learning with heterogeneous resources in mobile edge. In: ICC 2019-2019 IEEE international conference on communications (ICC). pp. 1-7. IEEE (2019) +39. Noyan, M.A.: Hand written roman numerals dataset (Sep 2021). https://doi.org/10.5281/zenodo.5385144, https://doi.org/10.5281/zenodo.5385144 +40. Ostapenko, O., Puscas, M., Klein, T., Jahnichen, P., Nabi, M.: Learning to remember: A synaptic plasticity driven framework for continual learning. In: Proceedings of the IEEE/CVF conference on computer vision and pattern recognition. pp. 11321-11329 (2019) +41. Parks Associates: At CES 2024, Parks Associates Announces New Research Showing Average Number of Connected Devices per US Internet Household Reached 17 in 2023. https://www.parksassociates.com/blogs/press-releases/at-ces-2024-parks-associates-announces-new-research-showing-average-number-of-connected-devices-per-us-internet-household-reached-17-in-2023 (2024), accessed: 2024-05-15 +42. PyTorch: Retrieved from python model zoo (2017), https://pytorch.org/vision/stable/models.html +43. Radford, A., Kim, J.W., Hallacy, C., Ramesh, A., Goh, G., Agarwal, S., Sastry, G., Askell, A., Mishkin, P., Clark, J., et al.: Learning transferable visual models from natural language supervision. In: International conference on machine learning. pp. 8748-8763. PMLR (2021) +44. Roy, A.G., Siddiqui, S., Pölsterl, S., Navab, N., Wachinger, C.: Brainstorm: A peer-to-peer environment for decentralized federated learning. arXiv preprint arXiv:1905.06731 (2019) +45. Roy, P., Ghosh, S., Bhattacharya, S., Pal, U.: Effects of degradations on deep neural network architectures. arXiv preprint arXiv:1807.10108 (2018) +46. Rusu, A.A., Rabinowitz, N.C., Desjardins, G., Soyer, H., Kirkpatrick, J., Kavukcuoglu, K., Pascanu, R., Hadsell, R.: Progressive neural networks. arXiv preprint arXiv:1606.04671 (2016) +47. Rusu, A.A., Rabinowitz, N.C., Desjardins, G., Soyer, H., Kirkpatrick, J., Kavukcuoglu, K., Pascanu, R., Hadsell, R.: Progressive neural networks. arXiv preprint arXiv:1606.04671 (2016) +48. Saleh, B., Elgammal, A.: Large-scale classification of fine-art paintings: Learning the right metric on the right feature. arXiv preprint arXiv:1505.00855 (2015) + +49. Shayan, M., Fung, C., Yoon, C.J., Beschastnikh, I.: Biscotti: A blockchain system for private and secure federated learning. IEEE Transactions on Parallel and Distributed Systems 32(7), 1513-1525 (2020) +50. Shenaj, D., Toldo, M., Rigon, A., Zanuttigh, P.: Asynchronous federated continual learning. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 5054-5062 (2023) +51. Simonyan, K., Zisserman, A.: Very deep convolutional networks for large-scale image recognition. arXiv preprint arXiv:1409.1556 (2014) +52. Stoychev, S., Churamani, N., Gunes, H.: Latent generative replay for resource-efficient continual learning of facial expressions. In: 2023 IEEE 17th International Conference on Automatic Face and Gesture Recognition (FG). pp. 1-8. IEEE (2023) +53. Sun, W., Yang, Z., Zhang, X., Liu, Y.: Energy-efficient neighbor discovery in mobile ad hoc and wireless sensor networks: A survey. IEEE Communications Surveys & Tutorials 16(3), 1448-1459 (2014) +54. Tan, M., Le, Q.: Efficientnet: Rethinking model scaling for convolutional neural networks. In: International conference on machine learning. pp. 6105-6114. PMLR (2019) +55. Wah, C., Branson, S., Welinder, P., Perona, P., Belongie, S.: The caltech-ucsd birds-200-2011 dataset (2011) +56. Wei, K., Li, J., Ding, M., Ma, C., Yang, H.H., Farokhi, F., Jin, S., Quek, T.Q., Poor, H.V.: Federated learning with differential privacy: Algorithms and performance analysis. IEEE transactions on information forensics and security 15, 3454-3469 (2020) +57. Yoon, J., Jeong, W., Lee, G., Yang, E., Hwang, S.J.: Federated continual learning with weighted inter-client transfer. In: International Conference on Machine Learning. pp. 12073-12086. PMLR (2021) +58. Yoon, J., Yang, E., Lee, J., Hwang, S.J.: Lifelong learning with dynamically expandable networks. arXiv preprint arXiv:1708.01547 (2017) +59. Yu, Q., Liu, Y., Wang, Y., Xu, K., Liu, J.: Multimodal federated learning via contrastive representation ensemble. arXiv preprint arXiv:2302.08888 (2023) +60. Yurochkin, M., Agarwal, M., Ghosh, S., Greenewald, K., Hoang, N., Khazaeni, Y. Bayesian nonparametric federated learning of neural networks. In: International conference on machine learning. pp. 7252-7261. PMLR (2019) +61. Zhang, J., Chen, C., Zhuang, W., Lyu, L.: Target: Federated class-continual learning via exemplar-free distillation. In: Proceedings of the IEEE/CVF International Conference on Computer Vision. pp. 4782-4793 (2023) +62. Zhang, T., Mai, C., Chang, Y., Chen, C., Shu, L., Zheng, Z.: Fedego: privacy-preserving personalized federated graph learning with ego-graphs. ACM Transactions on Knowledge Discovery from Data 18(2), 1-27 (2023) +63. Zhu, L., Liu, Z., Han, S.: Deep leakage from gradients. Advances in neural information processing systems 32 (2019) \ No newline at end of file diff --git a/2024/Pick-a-back_ Selective Device-to-Device Knowledge Transfer in Federated Continual Learning/images.zip b/2024/Pick-a-back_ Selective Device-to-Device Knowledge Transfer in Federated Continual Learning/images.zip new file mode 100644 index 0000000000000000000000000000000000000000..12d38b46debdffe24bcadb2364a863c27d8e262f --- /dev/null +++ b/2024/Pick-a-back_ Selective Device-to-Device Knowledge Transfer in Federated Continual Learning/images.zip @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:6d91ff1c78e1dddd1819fe1fdde2d67cbf794bb408584a3b7a86a5ef41da9e9b +size 311407 diff --git a/2024/Pick-a-back_ Selective Device-to-Device Knowledge Transfer in Federated Continual Learning/layout.json b/2024/Pick-a-back_ Selective Device-to-Device Knowledge Transfer in Federated Continual Learning/layout.json new file mode 100644 index 0000000000000000000000000000000000000000..854e15695fafffa9cbb4dba919bee6aeb064f277 --- /dev/null +++ b/2024/Pick-a-back_ Selective Device-to-Device Knowledge Transfer in Federated Continual Learning/layout.json @@ -0,0 +1,9147 @@ +{ + "pdf_info": [ + { + "para_blocks": [ + { + "bbox": [ + 133, + 111, + 481, + 148 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 111, + 481, + 148 + ], + "spans": [ + { + "bbox": [ + 133, + 111, + 481, + 148 + ], + "type": "text", + "content": "Pick-a-back: Selective Device-to-Device Knowledge Transfer in Federated Continual Learning" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 221, + 167, + 391, + 182 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 221, + 167, + 391, + 182 + ], + "spans": [ + { + "bbox": [ + 221, + 167, + 391, + 182 + ], + "type": "text", + "content": "JinYi Yoon1 and HyungJune Lee2" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 208, + 189, + 406, + 234 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 208, + 189, + 406, + 234 + ], + "spans": [ + { + "bbox": [ + 208, + 189, + 406, + 234 + ], + "type": "inline_equation", + "content": "^{1}" + }, + { + "bbox": [ + 208, + 189, + 406, + 234 + ], + "type": "text", + "content": " Virginia Tech, Blacksburg, VA, USA jinyioon@vt.edu \n" + }, + { + "bbox": [ + 208, + 189, + 406, + 234 + ], + "type": "inline_equation", + "content": "^{2}" + }, + { + "bbox": [ + 208, + 189, + 406, + 234 + ], + "type": "text", + "content": " Ewha Womans University, Seoul, South Korea hyungjune.lee@ewha.ac.kr" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 160, + 262, + 455, + 427 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 160, + 262, + 455, + 427 + ], + "spans": [ + { + "bbox": [ + 160, + 262, + 455, + 427 + ], + "type": "text", + "content": "Abstract. With the explosion of edge intelligence, leveraging federated indirect knowledge has become crucial for boosting the tasks of individual learners. However, the conventional approach to knowledge reuse often leads to catastrophic forgetting issues. In this paper, we revisit the concept of continual learning in the context of edge intelligence and address the knowledge transfer problem to enhance federated continual learning. Since each learner processes private heterogeneous data, we propose Pick-a-back, a device-to-device knowledge federation framework by selectively reusing the external knowledge with similar behavioral patterns. By borrowing indirect experiences, an edge device can initiate learning from useful knowledge and thus achieve faster yet more generalized knowledge acquisition. Using continual tasks consisting of various datasets on lightweight architectures, we have validated that Pick-a-back provides a significant inference improvement of up to " + }, + { + "bbox": [ + 160, + 262, + 455, + 427 + ], + "type": "inline_equation", + "content": "8.0\\%" + }, + { + "bbox": [ + 160, + 262, + 455, + 427 + ], + "type": "text", + "content": " via selective knowledge federation. Our codes are available at https://github.com/jinyi-yoon/Pick-a-back.git" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 159, + 437, + 455, + 460 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 159, + 437, + 455, + 460 + ], + "spans": [ + { + "bbox": [ + 159, + 437, + 455, + 460 + ], + "type": "text", + "content": "Keywords: Federated continual learning, edge intelligence, device-to-device knowledge federation, selective knowledge transfer" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 132, + 479, + 230, + 491 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 479, + 230, + 491 + ], + "spans": [ + { + "bbox": [ + 132, + 479, + 230, + 491 + ], + "type": "text", + "content": "1 Introduction" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 130, + 503, + 482, + 647 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 503, + 482, + 647 + ], + "spans": [ + { + "bbox": [ + 130, + 503, + 482, + 647 + ], + "type": "text", + "content": "As artificial intelligence (AI) permeates nearly every field, data has emerged as a pivotal source driving remarkable advancements. Most of the data originates from the end users or systems, stimulating a need for distributed learning on low-end edge devices to address privacy and latency concerns [2,31]. However, individual low-end edge devices often lack the computational capability to generate advanced knowledge [6]. In the context of human learning paradigms, where any individuals become educated and exchange helpful knowledge with one another, the emergence of edge intelligence and its widespread availability allow external knowledge to serve as a valuable resource in supporting individual learners. By facilitating some sharable insights [28,59] and harnessing collective intelligence [14], it has widely revolutionized learning paradigms across various fields, even where numerous data have non-identical forms and heterogeneous distributions." + } + ] + } + ], + "index": 6 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 133, + 653, + 234, + 665 + ], + "type": "page_footnote", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 653, + 234, + 665 + ], + "spans": [ + { + "bbox": [ + 133, + 653, + 234, + 665 + ], + "type": "text", + "content": "* Corresponding author" + } + ] + } + ], + "index": 7 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 0 + }, + { + "para_blocks": [ + { + "bbox": [ + 130, + 116, + 482, + 271 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 116, + 482, + 271 + ], + "spans": [ + { + "bbox": [ + 130, + 116, + 482, + 271 + ], + "type": "text", + "content": "While leveraging external knowledge to enhance individual tasks, conventional learners often encounter the challenge of catastrophic forgetting, wherein they adapt only to the assigned tasks, and consequently lose previously acquired knowledge in some degree over time. In contrast, to more effectively retain and reinforce pre-trained knowledge together with new knowledge, the learners would rather adopt a continual learning approach for ensuring both robustness and adaptability. To enhance the efficacy of individual continual learning, it is not imperative for learners to confine themselves to their own knowledge. The prevalence of edge intelligence may facilitate knowledge federation, enabling individual learners to leverage a broader or deeper understanding of data, which can be possibly heterogeneous. While recent works have devised federated continual learning frameworks, they typically take a centralized approach for assistance [3,18,50], rather than direct device-to-device knowledge transfer in a distributed manner." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 130, + 275, + 482, + 443 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 275, + 482, + 443 + ], + "spans": [ + { + "bbox": [ + 130, + 275, + 482, + 443 + ], + "type": "text", + "content": "As for edge AI, individual learners might be confronted with a series of diverse heterogeneous tasks. At the edge side accompanied with its own private data, each learner attempts to form its own distinct knowledge. Then, a question to raise is that how can an individual edge learner effectively harness externally available prior knowledge? It has been reported that there exists unique positive and negative knowledge at each model [43,57]. Maybe the answer would be straightforward yet meaningful – the learners would selectively reuse or customize some knowledge that aligns well with their specific behavioral patterns. Rather than starting from scratch, learners can benefit from the indirectly attained knowledge formed by others who have already seen similar patterns. In that the decision patterns can not be exactly the same, indirect knowledge transfer (with follow-up training) may offer a rich source of otherwise untrainable and unattainable knowledge from private data. This can lead to performance improvement via selective acquisition of some common yet distinct knowledge." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 130, + 446, + 482, + 529 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 446, + 482, + 529 + ], + "spans": [ + { + "bbox": [ + 130, + 446, + 482, + 529 + ], + "type": "text", + "content": "In this work, we revisit the knowledge reuse problem. We consider three challenges in edge intelligence with knowledge transfer: 1) edge devices have limited computational and storage resources; 2) learners are not willing to disclose their raw data; and 3) edge devices should take into account the reduction of communication costs. To overcome these concerns, our approach aims to enhance the performance of lightweight learning models, while maintaining data privacy and avoiding additional communication overhead." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 130, + 533, + 482, + 666 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 533, + 482, + 666 + ], + "spans": [ + { + "bbox": [ + 130, + 533, + 482, + 666 + ], + "type": "text", + "content": "We propose a selective and customized knowledge federation framework from one device to another, based on heterogeneous continual learning, called Pick-a-back, which picks a relevant backbone knowledge from other devices. Beyond the achievable knowledge of each individual learner, we strategically search for some effective knowledge sources selectively and then customize the commonly shared knowledge to closely align with the local task. By leveraging decision similarity based on mutated data to assess the efficacy of the backbone among the neighboring models spread in the form of physically separated edge intelligence, we selectively fetch and customize constructive knowledge. This approach not only enables decentralized knowledge exchange among the learners at devices, but also facilitates a more efficient and scalable learning paradigm." + } + ] + } + ], + "index": 5 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "spans": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "type": "text", + "content": "2" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 166, + 91, + 247, + 101 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 166, + 91, + 247, + 101 + ], + "spans": [ + { + "bbox": [ + 166, + 91, + 247, + 101 + ], + "type": "text", + "content": "J. Yoon and H. Lee" + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 1 + }, + { + "para_blocks": [ + { + "bbox": [ + 130, + 116, + 482, + 164 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 116, + 482, + 164 + ], + "spans": [ + { + "bbox": [ + 130, + 116, + 482, + 164 + ], + "type": "text", + "content": "This work contributes to advancing the field of deep learning in edge intelligence by providing selective knowledge transfer and fostering collaboration among local learners. It can thereby enhance the efficiency of the individual continual learning process. The main contributions can be summarized as follows:" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 135, + 171, + 482, + 337 + ], + "type": "list", + "angle": 0, + "index": 6, + "blocks": [ + { + "bbox": [ + 135, + 171, + 481, + 229 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 135, + 171, + 481, + 229 + ], + "spans": [ + { + "bbox": [ + 135, + 171, + 481, + 229 + ], + "type": "text", + "content": "1) We address the problem of device-to-device knowledge federation in the continual learning. We demonstrate the potential of leveraging indirect knowledge extracted by the neighboring edge learners to obtain otherwise untrainable and unattainable insights. This work provides a way to maintain a model with heterogeneous data collected across a network of edge devices." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 135, + 230, + 482, + 301 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 135, + 230, + 482, + 301 + ], + "spans": [ + { + "bbox": [ + 135, + 230, + 482, + 301 + ], + "type": "text", + "content": "2) As the diverse tasks imply their own unique knowledge, we propose selective and customized knowledge transfer by capturing the decision pattern similarity relevant to a target task running on a device. Pick-a-back adopts each relevant knowledge for a specific task, achieving high fidelity through the acquisition of some common yet distinct knowledge from unseen examples, while preserving data privacy and communication cost at the edge." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 135, + 301, + 482, + 337 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 135, + 301, + 482, + 337 + ], + "spans": [ + { + "bbox": [ + 135, + 301, + 482, + 337 + ], + "type": "text", + "content": "3) Extensive experimental results on various datasets across different architectures have shown that Pick-a-back provides a significant inference improvement up to " + }, + { + "bbox": [ + 135, + 301, + 482, + 337 + ], + "type": "inline_equation", + "content": "8.0\\%" + }, + { + "bbox": [ + 135, + 301, + 482, + 337 + ], + "type": "text", + "content": " via selective knowledge federation." + } + ] + } + ], + "index": 5 + } + ], + "sub_type": "text" + }, + { + "bbox": [ + 132, + 355, + 237, + 368 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 355, + 237, + 368 + ], + "spans": [ + { + "bbox": [ + 132, + 355, + 237, + 368 + ], + "type": "text", + "content": "2 Related Work" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 130, + 379, + 482, + 430 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 379, + 482, + 430 + ], + "spans": [ + { + "bbox": [ + 130, + 379, + 482, + 430 + ], + "type": "text", + "content": "Typical continual learning refers to training on a series of multiple tasks [22] or incremental classes [13]. The primary objective is to address the catastrophic forgetting problem, where previously learned knowledge is decaying over time as adapting to new tasks, using memorization [40,52] or regularization [9,23]." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 130, + 437, + 482, + 668 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 437, + 482, + 668 + ], + "spans": [ + { + "bbox": [ + 130, + 437, + 482, + 668 + ], + "type": "text", + "content": "Knowledge-reusable continual learning. More closely related to our problem, there has been a growing interest in reusing existing knowledge for its subsequent tasks [30, 58]. Unlike conventional approaches, these methods have focused on concatenating a new model to its previous model, allowing input data to propagate through the previous model [46]. Rather than reusing the whole model, researchers have proposed to activate only the essential parts of the model [19,29]. In *Piggyback* [29], they leverage a common model and mask out unnecessary neurons to extract relevant knowledge for a target task. It effectively reuses existing knowledge with only a single sharable model, but it is less able to accept incremental tasks. *CPG* [19] combines neuron expansion with neuron pruning and masking techniques to address the challenges of forgetting, model compactness, and achieving higher fidelity in knowledge reuse. We adopt these knowledge reusable continual learning approaches to retain and extract valuable insights from previous tasks, while aiming to find useful knowledge (for reuse) based on model similarity. Although some works have applied some similarity metrics such as cosine similarity to federated learning scenarios [20], it is limited in applicability when dealing with clients having a different number of outputs (classes). In contrast, our approach provides flexibility in output heterogeneity, ensuring broader applicability in real-world scenarios, where most clients may" + } + ] + } + ], + "index": 9 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 399, + 91, + 447, + 100 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 399, + 91, + 447, + 100 + ], + "spans": [ + { + "bbox": [ + 399, + 91, + 447, + 100 + ], + "type": "text", + "content": "Pick-a-back" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 474, + 92, + 480, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 474, + 92, + 480, + 100 + ], + "spans": [ + { + "bbox": [ + 474, + 92, + 480, + 100 + ], + "type": "text", + "content": "3" + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 2 + }, + { + "para_blocks": [ + { + "bbox": [ + 130, + 116, + 482, + 177 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 116, + 482, + 177 + ], + "spans": [ + { + "bbox": [ + 130, + 116, + 482, + 177 + ], + "type": "text", + "content": "have different shapes of outputs. Moreover, our decision pattern-based similarity is novel in federated continual learning, as the knowledge itself is not exactly the same even with the similar decision patterns. This enables private learners to be exposed to similar yet unseen broader knowledge, leading to improved performance and efficiency in subsequent learning tasks." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 130, + 193, + 483, + 480 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 193, + 483, + 480 + ], + "spans": [ + { + "bbox": [ + 130, + 193, + 483, + 480 + ], + "type": "text", + "content": "Federated continual learning. With widespread accessibility of edge AI, learners do not necessarily be confined to their own private knowledge, as a standalone learner is permitted to only access its own local information and thus suffer from restricted experiences to some extent. Although recent works tackle the problem of federated continual learning [50,61], they focus on class-incremental learning, whereas we tackle task-incremental learning. In task-incremental learning, transfer learning turns out to be very effective, where a model is pretrained based on the source domain and then finetuned for the target domain. This selective device-to-device knowledge transfer can be a significant impetus to enhance model fidelity, also in continual learning. It has been commonly addressed in the field of federated continual learning but only in a centralized manner [18,50]. It has been also observed that there is both positive and negative knowledge related to a specific target task [43,57], highlighting the potential harm of static knowledge selection. To address the two-sided nature of knowledge reuse, FedWeIT [57] aggregates the parameters from clients as global knowledge, and then the clients mask the federated parameters adaptive to the specific task. However, previous works including FedWeIT rely on a central server, causing communication bottlenecks and privacy limitations. Furthermore, ImageNet-pretrained models have widely been utilized in transfer learning. However, it is not feasible for edge-side federated continual learning, as pretrained models are specifically designed for well-known large-scale architectures. The architectures in continual learning are dynamically changing, and edge AI cannot be tailored to large-scale models. In Pick-a-back, we rather explore decentralized, communication-efficient, and privacy-preserving knowledge transfer in lightweight on-device models." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 132, + 503, + 329, + 517 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 503, + 329, + 517 + ], + "spans": [ + { + "bbox": [ + 132, + 503, + 329, + 517 + ], + "type": "text", + "content": "3 Selective Knowledge Transfer" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 130, + 533, + 483, + 666 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 533, + 483, + 666 + ], + "spans": [ + { + "bbox": [ + 130, + 533, + 483, + 666 + ], + "type": "text", + "content": "We consider the problem of device-to-device knowledge federation to enhance individual continual learning in a distributed manner. We propose Pick-a-back, a simple yet effective knowledge federation approach by discovering a relevant neighboring device who has acquired the knowledge of relevant tasks and then selectively infusing it into the model via continual learning. By leveraging a federated learning framework, which allows collaborative learning across devices, we aim to address the challenge of distributed continual learning, while also facilitating the reuse of previously acquired knowledge at each different device. The integration of these two fields of study enables efficient knowledge selection, sharing, and expansion among devices (with heterogeneous data), yielding continuous learning improvement and adaptation across tasks or learning scenarios." + } + ] + } + ], + "index": 5 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 133, + 91, + 141, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 91, + 141, + 100 + ], + "spans": [ + { + "bbox": [ + 133, + 91, + 141, + 100 + ], + "type": "text", + "content": "4" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 166, + 91, + 247, + 101 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 166, + 91, + 247, + 101 + ], + "spans": [ + { + "bbox": [ + 166, + 91, + 247, + 101 + ], + "type": "text", + "content": "J. Yoon and H. Lee" + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 3 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 152, + 115, + 462, + 242 + ], + "blocks": [ + { + "bbox": [ + 152, + 115, + 462, + 242 + ], + "lines": [ + { + "bbox": [ + 152, + 115, + 462, + 242 + ], + "spans": [ + { + "bbox": [ + 152, + 115, + 462, + 242 + ], + "type": "image", + "image_path": "9d33422866b30461ddd6c6b21b8acaa4f2eb16d106e360a8320233fc715468c0.jpg" + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 130, + 251, + 482, + 339 + ], + "lines": [ + { + "bbox": [ + 130, + 251, + 482, + 339 + ], + "spans": [ + { + "bbox": [ + 130, + 251, + 482, + 339 + ], + "type": "text", + "content": "Fig. 1: Overall architecture of Pick-a-back: (1) Each device trains an individual continual learning model on a series of tasks. (2) Given these available neighboring knowledge, a device selects a relevant neighboring model that can improve learning performance by acquiring indirect knowledge from the federated but physically separated devices. (3) On top of the retrieved knowledge, (4) the continual learner stacks new local tasks and resumes learning and federation. By fetching the generalized knowledge with relevant decision patterns, Pick-a-back offers a cumulative yet selective knowledge federation with heterogeneous data at a distinct device" + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "image_caption" + } + ], + "index": 4 + }, + { + "bbox": [ + 132, + 365, + 268, + 376 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 365, + 268, + 376 + ], + "spans": [ + { + "bbox": [ + 132, + 365, + 268, + 376 + ], + "type": "text", + "content": "3.1 Problem Formulation" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 130, + 389, + 482, + 499 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 389, + 482, + 499 + ], + "spans": [ + { + "bbox": [ + 130, + 389, + 482, + 499 + ], + "type": "text", + "content": "Given " + }, + { + "bbox": [ + 130, + 389, + 482, + 499 + ], + "type": "inline_equation", + "content": "L" + }, + { + "bbox": [ + 130, + 389, + 482, + 499 + ], + "type": "text", + "content": " devices, each learner " + }, + { + "bbox": [ + 130, + 389, + 482, + 499 + ], + "type": "inline_equation", + "content": "l \\in \\{1, 2, \\dots, L\\}" + }, + { + "bbox": [ + 130, + 389, + 482, + 499 + ], + "type": "text", + "content": " at device " + }, + { + "bbox": [ + 130, + 389, + 482, + 499 + ], + "type": "inline_equation", + "content": "d_l" + }, + { + "bbox": [ + 130, + 389, + 482, + 499 + ], + "type": "text", + "content": " processes its own sequential tasks (i.e., training data) of " + }, + { + "bbox": [ + 130, + 389, + 482, + 499 + ], + "type": "inline_equation", + "content": "T^{(l)} = \\{T_1^{(l)}, T_2^{(l)}, \\dots \\}" + }, + { + "bbox": [ + 130, + 389, + 482, + 499 + ], + "type": "text", + "content": ". A scratch model " + }, + { + "bbox": [ + 130, + 389, + 482, + 499 + ], + "type": "inline_equation", + "content": "M_i^{(l)}" + }, + { + "bbox": [ + 130, + 389, + 482, + 499 + ], + "type": "text", + "content": " is constructed by training only with private data " + }, + { + "bbox": [ + 130, + 389, + 482, + 499 + ], + "type": "inline_equation", + "content": "T_i^{(l)} = \\{(X_i^{(l)}, Y_i^{(l)})\\}" + }, + { + "bbox": [ + 130, + 389, + 482, + 499 + ], + "type": "text", + "content": ", where " + }, + { + "bbox": [ + 130, + 389, + 482, + 499 + ], + "type": "inline_equation", + "content": "T_i^{(l)}" + }, + { + "bbox": [ + 130, + 389, + 482, + 499 + ], + "type": "text", + "content": " for task " + }, + { + "bbox": [ + 130, + 389, + 482, + 499 + ], + "type": "inline_equation", + "content": "i" + }, + { + "bbox": [ + 130, + 389, + 482, + 499 + ], + "type": "text", + "content": " and device " + }, + { + "bbox": [ + 130, + 389, + 482, + 499 + ], + "type": "inline_equation", + "content": "d_l" + }, + { + "bbox": [ + 130, + 389, + 482, + 499 + ], + "type": "text", + "content": " is heterogeneous across other tasks and devices. We train the model " + }, + { + "bbox": [ + 130, + 389, + 482, + 499 + ], + "type": "inline_equation", + "content": "\\hat{M}_i^{(l)}" + }, + { + "bbox": [ + 130, + 389, + 482, + 499 + ], + "type": "text", + "content": " with target data " + }, + { + "bbox": [ + 130, + 389, + 482, + 499 + ], + "type": "inline_equation", + "content": "T_i^{(l)}" + }, + { + "bbox": [ + 130, + 389, + 482, + 499 + ], + "type": "text", + "content": " on top of its previous version " + }, + { + "bbox": [ + 130, + 389, + 482, + 499 + ], + "type": "inline_equation", + "content": "\\hat{M}_{i-1}^{(l)}" + }, + { + "bbox": [ + 130, + 389, + 482, + 499 + ], + "type": "text", + "content": " with some learning parameters " + }, + { + "bbox": [ + 130, + 389, + 482, + 499 + ], + "type": "inline_equation", + "content": "\\theta_i^{(l)}" + }, + { + "bbox": [ + 130, + 389, + 482, + 499 + ], + "type": "text", + "content": ", i.e., " + }, + { + "bbox": [ + 130, + 389, + 482, + 499 + ], + "type": "inline_equation", + "content": "\\hat{M}_i^{(l)} = \\hat{M}_{i-1}^{(l)}(T_i^{(l)}, \\theta_i^{(l)})" + }, + { + "bbox": [ + 130, + 389, + 482, + 499 + ], + "type": "text", + "content": " in a continual manner. Here, at every task " + }, + { + "bbox": [ + 130, + 389, + 482, + 499 + ], + "type": "inline_equation", + "content": "i" + }, + { + "bbox": [ + 130, + 389, + 482, + 499 + ], + "type": "text", + "content": ", the goal is to find a good reference model that can maximize the overall fidelity of model " + }, + { + "bbox": [ + 130, + 389, + 482, + 499 + ], + "type": "inline_equation", + "content": "\\hat{M}_i^{(l)}" + }, + { + "bbox": [ + 130, + 389, + 482, + 499 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 132, + 521, + 225, + 532 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 521, + 225, + 532 + ], + "spans": [ + { + "bbox": [ + 132, + 521, + 225, + 532 + ], + "type": "text", + "content": "3.2 Architecture" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 130, + 545, + 482, + 666 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 545, + 482, + 666 + ], + "spans": [ + { + "bbox": [ + 130, + 545, + 482, + 666 + ], + "type": "text", + "content": "Our learning architecture incorporates two distinct modes (Fig. 1): i) continual learner for local training; and ii) federated learner for knowledge transfer. Initially, multiple learners train independently on a specific private task. Subsequently, one learner at a device transitions to a federated learning mode to explore effective indirect knowledge from a neighboring relevant model. Once a neighboring model is selected, we use the selected model as a backbone, the main target task is then executed in a continual manner. By incorporating the retrieved knowledge together with its existing own knowledge, the learner can construct cumulative knowledge in a continual manner without forgetting. This process of continual learning and federated learning is repeated for each task." + } + ] + } + ], + "index": 9 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 388, + 47, + 402, + 60 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 388, + 47, + 402, + 60 + ], + "spans": [ + { + "bbox": [ + 388, + 47, + 402, + 60 + ], + "type": "text", + "content": "#" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 403, + 49, + 490, + 57 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 403, + 49, + 490, + 57 + ], + "spans": [ + { + "bbox": [ + 403, + 49, + 490, + 57 + ], + "type": "text", + "content": "EWHA WOMANS UNIVERSITY" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 399, + 91, + 447, + 100 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 399, + 91, + 447, + 100 + ], + "spans": [ + { + "bbox": [ + 399, + 91, + 447, + 100 + ], + "type": "text", + "content": "Pick-a-back" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 474, + 91, + 480, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 474, + 91, + 480, + 100 + ], + "spans": [ + { + "bbox": [ + 474, + 91, + 480, + 100 + ], + "type": "text", + "content": "5" + } + ] + } + ], + "index": 3 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 4 + }, + { + "para_blocks": [ + { + "bbox": [ + 130, + 116, + 482, + 283 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 116, + 482, + 283 + ], + "spans": [ + { + "bbox": [ + 130, + 116, + 482, + 283 + ], + "type": "text", + "content": "Continual learner. In our approach to continual learning, we adopt task-incremental learning, where a series of tasks is processed sequentially. Each task has a distinct learning objective, and a local model is trained independently on each task with freezing or isolating the models for previous knowledge. By employing task-oriented learning, we aim to mitigate the catastrophic forgetting problem, where previously learned knowledge vanishes as training on new tasks or incorporating new data over time. Our approach can be combined with any continual learner that has gradually been constructing knowledge. It is fundamentally different from conventional continual learning that simply memorizes multiple models or data, which do not leverage the given knowledge. We suppose that a continual learner concatenates a new model on top of the previously acquired knowledge, and thus it can reduce memory usage and also enhance the model fidelity. This sustainability of the continual model is particularly valuable in the context of edge AI and aligns with the goals of the recent learning paradigm." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 130, + 284, + 482, + 403 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 284, + 482, + 403 + ], + "spans": [ + { + "bbox": [ + 130, + 284, + 482, + 403 + ], + "type": "text", + "content": "We pursue reusing the pre-configured knowledge by using the model. We mainly use the latest continual learning effort, " + }, + { + "bbox": [ + 130, + 284, + 482, + 403 + ], + "type": "inline_equation", + "content": "CPG" + }, + { + "bbox": [ + 130, + 284, + 482, + 403 + ], + "type": "text", + "content": " [19], which expands the model when necessary, prunes it for compactness, and reuses the network by masking out specific ineffective neurons. Inspired by the masking strategy of Piggyback [29], it samples some necessary neurons for the target task to allow the extraction of common knowledge. " + }, + { + "bbox": [ + 130, + 284, + 482, + 403 + ], + "type": "inline_equation", + "content": "CPG" + }, + { + "bbox": [ + 130, + 284, + 482, + 403 + ], + "type": "text", + "content": " only trains the learnable binary mask whether to activate each neuron for the target task. As it ensures the retention of previous knowledge via freezing, the results remain exactly the same even when training new tasks. However, without loss of generality, this masking strategy can be applied to any continual learner that holds multiple knowledge." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 130, + 414, + 482, + 628 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 414, + 482, + 628 + ], + "spans": [ + { + "bbox": [ + 130, + 414, + 482, + 628 + ], + "type": "text", + "content": "Federated learner. Due to the privacy and communication overhead, distributed learners are unwilling to share private raw data. We thus adopt a model-sharing approach rather than data-sharing. Here, each client is regarded as an independent learner. When each learner at a distinct device is working on a distinct task but lacks knowledge, there might be some other knowledge that has a similar, broader, or deeper experience beyond the learner based on its own local data. We assume that clients are knowledgeable with their own currently available neighboring models via existing neighbor selection methods [38, 53] or graph construction approaches [33, 62]. These neighboring diverse knowledge sources have been trained on different datasets, capturing various perspectives and feature extraction capabilities. By exposing the learner to external knowledge, the model is capable of leveraging enhanced feature representations, enabling it to capture a broader set of patterns and information that may be beneficial for its specific learning objective. The learner can also develop a more comprehensive understanding of the data, generalize better to unseen examples by mitigating overfitting, and thus improve the fidelity of the model on a variety of tasks and data. By leveraging a diverse ensemble of knowledge sources, the learner can benefit from its complementary expertise, leading to accuracy improvement." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 130, + 630, + 482, + 665 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 630, + 482, + 665 + ], + "spans": [ + { + "bbox": [ + 130, + 630, + 482, + 665 + ], + "type": "text", + "content": "Each learner at a device seeks relevant knowledge that can contribute to further improving its performance based on local continual learning. To acquire necessary constructive knowledge, we identify effective indirect experiences based" + } + ] + } + ], + "index": 5 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "spans": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "type": "text", + "content": "6" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 166, + 91, + 247, + 101 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 166, + 91, + 247, + 101 + ], + "spans": [ + { + "bbox": [ + 166, + 91, + 247, + 101 + ], + "type": "text", + "content": "J. Yoon and H. Lee" + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 5 + }, + { + "para_blocks": [ + { + "bbox": [ + 130, + 116, + 482, + 213 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 116, + 482, + 213 + ], + "spans": [ + { + "bbox": [ + 130, + 116, + 482, + 213 + ], + "type": "text", + "content": "on decision pattern similarities. Once we have obtained knowledge from other learners, we accumulate ongoing new local tasks on top of the model in a continual manner, ensuring that the retrieved knowledge becomes infused with local knowledge. Our main focus is on the current target task, and thus, if we want to retain the knowledge from our previous tasks, we have to either store the existing knowledge separately or concatenate it with the fetched model. If the currently constructed knowledge turns out to be sufficiently effective, there may be no further need to incorporate further knowledge from other learners." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 131, + 228, + 321, + 240 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 131, + 228, + 321, + 240 + ], + "spans": [ + { + "bbox": [ + 131, + 228, + 321, + 240 + ], + "type": "text", + "content": "3.3 Selective Knowledge Federation" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 130, + 247, + 482, + 390 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 247, + 482, + 390 + ], + "spans": [ + { + "bbox": [ + 130, + 247, + 482, + 390 + ], + "type": "text", + "content": "Each learner is responsible for distinct tasks, and tasks may exhibit unique strengths, weaknesses, or biases. Consequently, each task at a local learner plays its own unique different role and forms distinct knowledge. Some external models may have learned other representations that have captured specialized knowledge relevant to specific tasks. Therefore, selective knowledge transfer can be a very effective way for enhancing the performance of each student model at an edge device. Our approach is simple and straightforward – if models have similar or synergetic capabilities in classifying different samples, they are likely to have potential sources of common but more generalized knowledge. Since it has the possibility to attain broader intuition but also inexperienced knowledge, this indirect knowledge characterized by similar decision patterns may greatly accelerate and improve the learning for individual learners." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 130, + 391, + 482, + 594 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 391, + 482, + 594 + ], + "spans": [ + { + "bbox": [ + 130, + 391, + 482, + 594 + ], + "type": "text", + "content": "As the similarity of model do not straightly incur similar knowledge, we focus on the decision pattern similarity rather than the model itself. We borrow an idea to calculate the model similarity from ModelDiff [27], which calculates the similarity of decision boundaries by comparing the reaction to the input data. Initially, we sample input data " + }, + { + "bbox": [ + 130, + 391, + 482, + 594 + ], + "type": "inline_equation", + "content": "x_{i}" + }, + { + "bbox": [ + 130, + 391, + 482, + 594 + ], + "type": "text", + "content": " from the requester and adversarial data " + }, + { + "bbox": [ + 130, + 391, + 482, + 594 + ], + "type": "inline_equation", + "content": "x_{i}'" + }, + { + "bbox": [ + 130, + 391, + 482, + 594 + ], + "type": "text", + "content": " from a neighbor. To address privacy concerns, an alternative approach may sample data only from the requester, or the participants may opt to exchange mutated data, thus avoiding the sharing of private data. Then, we inject a pair-wise set of input data, " + }, + { + "bbox": [ + 130, + 391, + 482, + 594 + ], + "type": "inline_equation", + "content": "\\{(x_{1},x_{1}'),(x_{2},x_{2}'),\\dots ,(x_{n},x_{n}')\\}" + }, + { + "bbox": [ + 130, + 391, + 482, + 594 + ], + "type": "text", + "content": ", into each model " + }, + { + "bbox": [ + 130, + 391, + 482, + 594 + ], + "type": "inline_equation", + "content": "f" + }, + { + "bbox": [ + 130, + 391, + 482, + 594 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 130, + 391, + 482, + 594 + ], + "type": "inline_equation", + "content": "g" + }, + { + "bbox": [ + 130, + 391, + 482, + 594 + ], + "type": "text", + "content": " and then computes the decision distance vector (DDV), " + }, + { + "bbox": [ + 130, + 391, + 482, + 594 + ], + "type": "inline_equation", + "content": "DDV_{f} = \\{d_{1},d_{2},\\dots ,d_{n}\\}" + }, + { + "bbox": [ + 130, + 391, + 482, + 594 + ], + "type": "text", + "content": ", which is a distance vector for the output pair, where " + }, + { + "bbox": [ + 130, + 391, + 482, + 594 + ], + "type": "inline_equation", + "content": "d_{i} = \\text{distance}(f(x_{i}),f(x_{i}'))" + }, + { + "bbox": [ + 130, + 391, + 482, + 594 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 130, + 391, + 482, + 594 + ], + "type": "inline_equation", + "content": "i\\in \\{1,2,\\ldots ,n\\}" + }, + { + "bbox": [ + 130, + 391, + 482, + 594 + ], + "type": "text", + "content": ". Finally, it compares the cosine similarity between the distance vectors of models, " + }, + { + "bbox": [ + 130, + 391, + 482, + 594 + ], + "type": "inline_equation", + "content": "sim(DDV_{f},DDV_{g})" + }, + { + "bbox": [ + 130, + 391, + 482, + 594 + ], + "type": "text", + "content": ". Since it focuses the decision similarity instead of directly comparing the outputs, it is applicable to any heterogeneous task with a different number of output classes. To search a pair-wise set that can approximate the decision boundary most effectively, the input data is mutated, and similarities are iteratively computed, similar to genetic algorithms." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 130, + 594, + 483, + 668 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 594, + 483, + 668 + ], + "spans": [ + { + "bbox": [ + 130, + 594, + 483, + 668 + ], + "type": "text", + "content": "By incorporating it in Pick-a-back, when we search for a similar model for task " + }, + { + "bbox": [ + 130, + 594, + 483, + 668 + ], + "type": "inline_equation", + "content": "T_{i}^{(l)}" + }, + { + "bbox": [ + 130, + 594, + 483, + 668 + ], + "type": "text", + "content": " of learner " + }, + { + "bbox": [ + 130, + 594, + 483, + 668 + ], + "type": "inline_equation", + "content": "l" + }, + { + "bbox": [ + 130, + 594, + 483, + 668 + ], + "type": "text", + "content": " among " + }, + { + "bbox": [ + 130, + 594, + 483, + 668 + ], + "type": "inline_equation", + "content": "\\hat{M}_{i-1}^{k}" + }, + { + "bbox": [ + 130, + 594, + 483, + 668 + ], + "type": "text", + "content": " of all neighboring learner " + }, + { + "bbox": [ + 130, + 594, + 483, + 668 + ], + "type": "inline_equation", + "content": "k" + }, + { + "bbox": [ + 130, + 594, + 483, + 668 + ], + "type": "text", + "content": ". We sample an input pair-wise data " + }, + { + "bbox": [ + 130, + 594, + 483, + 668 + ], + "type": "inline_equation", + "content": "X_{i}^{(l+k)}" + }, + { + "bbox": [ + 130, + 594, + 483, + 668 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 130, + 594, + 483, + 668 + ], + "type": "inline_equation", + "content": "X_{i}^{\\prime(l+k)}" + }, + { + "bbox": [ + 130, + 594, + 483, + 668 + ], + "type": "text", + "content": " from " + }, + { + "bbox": [ + 130, + 594, + 483, + 668 + ], + "type": "inline_equation", + "content": "X_{i}^{(l)} \\cup X_{i-1}^{(k)}" + }, + { + "bbox": [ + 130, + 594, + 483, + 668 + ], + "type": "text", + "content": " (or just from " + }, + { + "bbox": [ + 130, + 594, + 483, + 668 + ], + "type": "inline_equation", + "content": "X_{i}^{(l)}" + }, + { + "bbox": [ + 130, + 594, + 483, + 668 + ], + "type": "text", + "content": "). By injecting " + }, + { + "bbox": [ + 130, + 594, + 483, + 668 + ], + "type": "inline_equation", + "content": "X_{i}^{(l+k)}" + }, + { + "bbox": [ + 130, + 594, + 483, + 668 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 130, + 594, + 483, + 668 + ], + "type": "inline_equation", + "content": "X_{i}^{\\prime(l+k)}" + }, + { + "bbox": [ + 130, + 594, + 483, + 668 + ], + "type": "text", + "content": " into the learner model " + }, + { + "bbox": [ + 130, + 594, + 483, + 668 + ], + "type": "inline_equation", + "content": "M_{i}^{(l)}" + }, + { + "bbox": [ + 130, + 594, + 483, + 668 + ], + "type": "text", + "content": " and the neighboring models " + }, + { + "bbox": [ + 130, + 594, + 483, + 668 + ], + "type": "inline_equation", + "content": "\\hat{M}_{i-1}^{(k)}" + }, + { + "bbox": [ + 130, + 594, + 483, + 668 + ], + "type": "text", + "content": ", we produce the distance vector of outputs " + }, + { + "bbox": [ + 130, + 594, + 483, + 668 + ], + "type": "inline_equation", + "content": "D_{M_{i}^{(l)}}" + }, + { + "bbox": [ + 130, + 594, + 483, + 668 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 130, + 594, + 483, + 668 + ], + "type": "inline_equation", + "content": "D_{\\hat{M}_{i-1}^{(k)}}" + }, + { + "bbox": [ + 130, + 594, + 483, + 668 + ], + "type": "text", + "content": ". Then," + } + ] + } + ], + "index": 6 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 399, + 91, + 447, + 100 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 399, + 91, + 447, + 100 + ], + "spans": [ + { + "bbox": [ + 399, + 91, + 447, + 100 + ], + "type": "text", + "content": "Pick-a-back" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 474, + 91, + 480, + 99 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 474, + 91, + 480, + 99 + ], + "spans": [ + { + "bbox": [ + 474, + 91, + 480, + 99 + ], + "type": "text", + "content": "7" + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 6 + }, + { + "para_blocks": [ + { + "type": "code", + "bbox": [ + 136, + 132, + 481, + 248 + ], + "blocks": [ + { + "bbox": [ + 133, + 115, + 406, + 128 + ], + "lines": [ + { + "bbox": [ + 133, + 115, + 406, + 128 + ], + "spans": [ + { + "bbox": [ + 133, + 115, + 406, + 128 + ], + "type": "text", + "content": "Algorithm 1 Federated Continual Learning with Pick-a-back" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "code_caption" + }, + { + "bbox": [ + 136, + 132, + 481, + 248 + ], + "lines": [ + { + "bbox": [ + 136, + 132, + 481, + 248 + ], + "spans": [ + { + "bbox": [ + 136, + 132, + 481, + 248 + ], + "type": "text", + "content": "1: Input: " + }, + { + "bbox": [ + 136, + 132, + 481, + 248 + ], + "type": "inline_equation", + "content": "L" + }, + { + "bbox": [ + 136, + 132, + 481, + 248 + ], + "type": "text", + "content": ": # of devices, (for each device " + }, + { + "bbox": [ + 136, + 132, + 481, + 248 + ], + "type": "inline_equation", + "content": "d_{l}" + }, + { + "bbox": [ + 136, + 132, + 481, + 248 + ], + "type": "text", + "content": ") " + }, + { + "bbox": [ + 136, + 132, + 481, + 248 + ], + "type": "inline_equation", + "content": "M_{0}^{(l)}" + }, + { + "bbox": [ + 136, + 132, + 481, + 248 + ], + "type": "text", + "content": ": random initial model, " + }, + { + "bbox": [ + 136, + 132, + 481, + 248 + ], + "type": "inline_equation", + "content": "T^{(d_{l})}" + }, + { + "bbox": [ + 136, + 132, + 481, + 248 + ], + "type": "text", + "content": ": a sequence of tasks, " + }, + { + "bbox": [ + 136, + 132, + 481, + 248 + ], + "type": "inline_equation", + "content": "\\theta_{i}^{(l)}" + }, + { + "bbox": [ + 136, + 132, + 481, + 248 + ], + "type": "text", + "content": ": learning parameters \n2: for task " + }, + { + "bbox": [ + 136, + 132, + 481, + 248 + ], + "type": "inline_equation", + "content": "i = 1: length(T^{(l)})" + }, + { + "bbox": [ + 136, + 132, + 481, + 248 + ], + "type": "text", + "content": " do \n3: Device " + }, + { + "bbox": [ + 136, + 132, + 481, + 248 + ], + "type": "inline_equation", + "content": "d_{l}" + }, + { + "bbox": [ + 136, + 132, + 481, + 248 + ], + "type": "text", + "content": " trains task " + }, + { + "bbox": [ + 136, + 132, + 481, + 248 + ], + "type": "inline_equation", + "content": "i" + }, + { + "bbox": [ + 136, + 132, + 481, + 248 + ], + "type": "text", + "content": " such that " + }, + { + "bbox": [ + 136, + 132, + 481, + 248 + ], + "type": "inline_equation", + "content": "M_{i}^{(l)} = M_{i-1}^{(l)}(T_{i}^{(l)}, \\theta_{i}^{(l)})" + }, + { + "bbox": [ + 136, + 132, + 481, + 248 + ], + "type": "text", + "content": "; " + }, + { + "bbox": [ + 136, + 132, + 481, + 248 + ], + "type": "inline_equation", + "content": "\\triangleright" + }, + { + "bbox": [ + 136, + 132, + 481, + 248 + ], + "type": "text", + "content": " Continual Learner \n4: for device " + }, + { + "bbox": [ + 136, + 132, + 481, + 248 + ], + "type": "inline_equation", + "content": "k = 1: L" + }, + { + "bbox": [ + 136, + 132, + 481, + 248 + ], + "type": "text", + "content": " do " + }, + { + "bbox": [ + 136, + 132, + 481, + 248 + ], + "type": "inline_equation", + "content": "\\triangleright" + }, + { + "bbox": [ + 136, + 132, + 481, + 248 + ], + "type": "text", + "content": " Federated Learner \n5: Compute the similarity, " + }, + { + "bbox": [ + 136, + 132, + 481, + 248 + ], + "type": "inline_equation", + "content": "sim(M_{i}^{(l)}, \\hat{M}_{i-1}^{(k)})" + }, + { + "bbox": [ + 136, + 132, + 481, + 248 + ], + "type": "text", + "content": "; \n6: end for \n7: Select and retrieve knowledge " + }, + { + "bbox": [ + 136, + 132, + 481, + 248 + ], + "type": "inline_equation", + "content": "\\hat{M}_{i-1}^{(l)} \\gets \\hat{M}_{i-1}^{(k)}" + }, + { + "bbox": [ + 136, + 132, + 481, + 248 + ], + "type": "text", + "content": " such that " + }, + { + "bbox": [ + 136, + 132, + 481, + 248 + ], + "type": "inline_equation", + "content": "\\arg \\max_{k} sim(M_{i}^{(l)}, \\hat{M}_{i-1}^{(k)})" + }, + { + "bbox": [ + 136, + 132, + 481, + 248 + ], + "type": "text", + "content": "; \n8: Train task " + }, + { + "bbox": [ + 136, + 132, + 481, + 248 + ], + "type": "inline_equation", + "content": "i" + }, + { + "bbox": [ + 136, + 132, + 481, + 248 + ], + "type": "text", + "content": " on the fetched model, " + }, + { + "bbox": [ + 136, + 132, + 481, + 248 + ], + "type": "inline_equation", + "content": "\\hat{M}_{i}^{(l)} = \\hat{M}_{i-1}^{(l)}(T_{i}^{(l)}, \\theta_{i}^{(l)})" + }, + { + "bbox": [ + 136, + 132, + 481, + 248 + ], + "type": "text", + "content": "; " + }, + { + "bbox": [ + 136, + 132, + 481, + 248 + ], + "type": "inline_equation", + "content": "\\triangleright" + }, + { + "bbox": [ + 136, + 132, + 481, + 248 + ], + "type": "text", + "content": " Continual Learner \n9: end for" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "code_body" + } + ], + "index": 3, + "sub_type": "algorithm" + }, + { + "bbox": [ + 132, + 270, + 324, + 281 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 270, + 324, + 281 + ], + "spans": [ + { + "bbox": [ + 132, + 270, + 324, + 281 + ], + "type": "text", + "content": "we select the most similar model as follows:" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 230, + 288, + 481, + 307 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 230, + 288, + 481, + 307 + ], + "spans": [ + { + "bbox": [ + 230, + 288, + 481, + 307 + ], + "type": "interline_equation", + "content": "\\underset {k} {\\operatorname {a r g m a x}} \\operatorname {s i m} \\left(D D V _ {M _ {i} ^ {(l)}}, D D V _ {\\hat {M} _ {i - 1} ^ {(k)}}\\right) \\tag {1}", + "image_path": "5f04f2ce78e1bdbd412d63fa5f7fa5e27653624365d1afd9bd557fe46339514a.jpg" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 131, + 312, + 482, + 339 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 131, + 312, + 482, + 339 + ], + "spans": [ + { + "bbox": [ + 131, + 312, + 482, + 339 + ], + "type": "text", + "content": "We fetch the indirect knowledge that has already experienced the target decision pattern. After selecting the exterior knowledge, we continue to learn " + }, + { + "bbox": [ + 131, + 312, + 482, + 339 + ], + "type": "inline_equation", + "content": "T_{i}^{l}" + }, + { + "bbox": [ + 131, + 312, + 482, + 339 + ], + "type": "text", + "content": " on " + }, + { + "bbox": [ + 131, + 312, + 482, + 339 + ], + "type": "inline_equation", + "content": "\\hat{M}_{i - 1}^{(k)}" + }, + { + "bbox": [ + 131, + 312, + 482, + 339 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 131, + 347, + 482, + 415 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 131, + 347, + 482, + 415 + ], + "spans": [ + { + "bbox": [ + 131, + 347, + 482, + 415 + ], + "type": "text", + "content": "Training process. Each learner " + }, + { + "bbox": [ + 131, + 347, + 482, + 415 + ], + "type": "inline_equation", + "content": "l" + }, + { + "bbox": [ + 131, + 347, + 482, + 415 + ], + "type": "text", + "content": " trains " + }, + { + "bbox": [ + 131, + 347, + 482, + 415 + ], + "type": "inline_equation", + "content": "i" + }, + { + "bbox": [ + 131, + 347, + 482, + 415 + ], + "type": "text", + "content": "-th task " + }, + { + "bbox": [ + 131, + 347, + 482, + 415 + ], + "type": "inline_equation", + "content": "T_{i}^{(l)}" + }, + { + "bbox": [ + 131, + 347, + 482, + 415 + ], + "type": "text", + "content": " and constructs a local model " + }, + { + "bbox": [ + 131, + 347, + 482, + 415 + ], + "type": "inline_equation", + "content": "M_{i}^{(l)}" + }, + { + "bbox": [ + 131, + 347, + 482, + 415 + ], + "type": "text", + "content": " with continual learning. Then, the learner " + }, + { + "bbox": [ + 131, + 347, + 482, + 415 + ], + "type": "inline_equation", + "content": "l" + }, + { + "bbox": [ + 131, + 347, + 482, + 415 + ], + "type": "text", + "content": " checks if there exists any helpful exterior knowledge by comparing " + }, + { + "bbox": [ + 131, + 347, + 482, + 415 + ], + "type": "inline_equation", + "content": "M_{i}^{(l)}" + }, + { + "bbox": [ + 131, + 347, + 482, + 415 + ], + "type": "text", + "content": " with " + }, + { + "bbox": [ + 131, + 347, + 482, + 415 + ], + "type": "inline_equation", + "content": "M_{i-1}^{(k)}" + }, + { + "bbox": [ + 131, + 347, + 482, + 415 + ], + "type": "text", + "content": " for all neighboring edge learners " + }, + { + "bbox": [ + 131, + 347, + 482, + 415 + ], + "type": "inline_equation", + "content": "k" + }, + { + "bbox": [ + 131, + 347, + 482, + 415 + ], + "type": "text", + "content": ". After selecting relevant knowledge for knowledge transfer, device " + }, + { + "bbox": [ + 131, + 347, + 482, + 415 + ], + "type": "inline_equation", + "content": "d_{l}" + }, + { + "bbox": [ + 131, + 347, + 482, + 415 + ], + "type": "text", + "content": " processes the " + }, + { + "bbox": [ + 131, + 347, + 482, + 415 + ], + "type": "inline_equation", + "content": "i" + }, + { + "bbox": [ + 131, + 347, + 482, + 415 + ], + "type": "text", + "content": "-th task " + }, + { + "bbox": [ + 131, + 347, + 482, + 415 + ], + "type": "inline_equation", + "content": "T_{i}^{(l)}" + }, + { + "bbox": [ + 131, + 347, + 482, + 415 + ], + "type": "text", + "content": " over " + }, + { + "bbox": [ + 131, + 347, + 482, + 415 + ], + "type": "inline_equation", + "content": "\\hat{M}_{i-1}^{(k)}" + }, + { + "bbox": [ + 131, + 347, + 482, + 415 + ], + "type": "text", + "content": " and constructs " + }, + { + "bbox": [ + 131, + 347, + 482, + 415 + ], + "type": "inline_equation", + "content": "\\hat{M}_{i}^{(l)} = \\hat{M}_{i-1}^{(k)}(T_{i}^{(l)}, \\theta_{i}^{(l)})" + }, + { + "bbox": [ + 131, + 347, + 482, + 415 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 146, + 415, + 369, + 427 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 146, + 415, + 369, + 427 + ], + "spans": [ + { + "bbox": [ + 146, + 415, + 369, + 427 + ], + "type": "text", + "content": "The overall procedure is described in Algorithm 1." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 132, + 441, + 230, + 455 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 441, + 230, + 455 + ], + "spans": [ + { + "bbox": [ + 132, + 441, + 230, + 455 + ], + "type": "text", + "content": "4 Experiments" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 130, + 463, + 482, + 512 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 463, + 482, + 512 + ], + "spans": [ + { + "bbox": [ + 130, + 463, + 482, + 512 + ], + "type": "text", + "content": "We have validated Pick-a-back extensively on four different classification task groups with 13 public data using four different well-known architectures. Each group is considered as one experimental environment where knowledge can be shared, and one superclass group is a single task." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 131, + 519, + 489, + 557 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 131, + 519, + 489, + 557 + ], + "spans": [ + { + "bbox": [ + 131, + 519, + 489, + 557 + ], + "type": "text", + "content": "Datasets. We used four different groups of CIFAR-100 [25], Classified-ImageNet [8], Fine-Grained [10,24,37,48,55], and Digits [7,11,35,36,39,45]. The organization of the dataset and the details are provided in Appendix Sec. A1." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 131, + 564, + 482, + 648 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 131, + 564, + 482, + 648 + ], + "spans": [ + { + "bbox": [ + 131, + 564, + 482, + 648 + ], + "type": "text", + "content": "Architectures. We have presented the results using both lightweight models and large-scale models. For the lightweight models, we used LeNet-5 [26] and MobileNetV1 [17], where ImageNet-pretrained models are not publicly available. On the other hand, VGG16-BN [51] and EfficientNet-B0 [54] that have publicly available ImageNet-pretrained models were implemented using PyTorch [42]. Details of model configurations and learning parameters are in Appendix Sec. A2. We used CIFAR-100 on LeNet-5 as default, unless otherwise noted." + } + ] + } + ], + "index": 12 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "spans": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "type": "text", + "content": "8" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 166, + 91, + 247, + 100 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 166, + 91, + 247, + 100 + ], + "spans": [ + { + "bbox": [ + 166, + 91, + 247, + 100 + ], + "type": "text", + "content": "J. Yoon and H. Lee" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 133, + 653, + 468, + 666 + ], + "type": "page_footnote", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 653, + 468, + 666 + ], + "spans": [ + { + "bbox": [ + 133, + 653, + 468, + 666 + ], + "type": "text", + "content": "1 Our codes are available at https://github.com/jinyi-yoon/Pick-a-back.git" + } + ] + } + ], + "index": 13 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 7 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 133, + 114, + 205, + 178 + ], + "blocks": [ + { + "bbox": [ + 133, + 114, + 205, + 178 + ], + "lines": [ + { + "bbox": [ + 133, + 114, + 205, + 178 + ], + "spans": [ + { + "bbox": [ + 133, + 114, + 205, + 178 + ], + "type": "image", + "image_path": "a659d8c9b10ab0113ab026bad83487f231c600742db27fbf8aecc938d719c511.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 137, + 179, + 200, + 188 + ], + "lines": [ + { + "bbox": [ + 137, + 179, + 200, + 188 + ], + "spans": [ + { + "bbox": [ + 137, + 179, + 200, + 188 + ], + "type": "text", + "content": "(a) Absolute Acc." + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_caption" + } + ], + "index": 2 + }, + { + "type": "image", + "bbox": [ + 206, + 114, + 279, + 178 + ], + "blocks": [ + { + "bbox": [ + 206, + 114, + 279, + 178 + ], + "lines": [ + { + "bbox": [ + 206, + 114, + 279, + 178 + ], + "spans": [ + { + "bbox": [ + 206, + 114, + 279, + 178 + ], + "type": "image", + "image_path": "fd37e59ee1bbd2997524d31a6964ca28ab86a0c0fc8db682f94784de713d3ab9.jpg" + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 211, + 179, + 276, + 188 + ], + "lines": [ + { + "bbox": [ + 211, + 179, + 276, + 188 + ], + "spans": [ + { + "bbox": [ + 211, + 179, + 276, + 188 + ], + "type": "text", + "content": "(b) Improved Acc." + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "image_caption" + } + ], + "index": 4 + }, + { + "type": "image", + "bbox": [ + 285, + 121, + 378, + 178 + ], + "blocks": [ + { + "bbox": [ + 285, + 121, + 378, + 178 + ], + "lines": [ + { + "bbox": [ + 285, + 121, + 378, + 178 + ], + "spans": [ + { + "bbox": [ + 285, + 121, + 378, + 178 + ], + "type": "image", + "image_path": "85cab9c58ed596d781e76927c2745f641eda0efdd4b085cf3447279db225d8b2.jpg" + } + ] + } + ], + "index": 6, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 291, + 180, + 371, + 189 + ], + "lines": [ + { + "bbox": [ + 291, + 180, + 371, + 189 + ], + "spans": [ + { + "bbox": [ + 291, + 180, + 371, + 189 + ], + "type": "text", + "content": "(a) Training dynamics" + } + ] + } + ], + "index": 7, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 130, + 198, + 281, + 297 + ], + "lines": [ + { + "bbox": [ + 130, + 198, + 281, + 297 + ], + "spans": [ + { + "bbox": [ + 130, + 198, + 281, + 297 + ], + "type": "text", + "content": "Fig.2: Heatmap of the accuracy for various backbone and target tasks sorted by the accuracy of " + }, + { + "bbox": [ + 130, + 198, + 281, + 297 + ], + "type": "inline_equation", + "content": "w / o" + }, + { + "bbox": [ + 130, + 198, + 281, + 297 + ], + "type": "text", + "content": " Backbone. The first row represents the averaged values for all tasks, while the first column shows the results of " + }, + { + "bbox": [ + 130, + 198, + 281, + 297 + ], + "type": "inline_equation", + "content": "w / o" + }, + { + "bbox": [ + 130, + 198, + 281, + 297 + ], + "type": "text", + "content": " Backbone. Absolute accuracy is scaled per each row, and accuracy improvement is scaled across all results" + } + ] + } + ], + "index": 10, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 283, + 198, + 482, + 297 + ], + "lines": [ + { + "bbox": [ + 283, + 198, + 482, + 297 + ], + "spans": [ + { + "bbox": [ + 283, + 198, + 482, + 297 + ], + "type": "text", + "content": "Fig. 3: Training dynamics for the most effective case of the target task 14 using the backbone task of 4 and accuracy improvement with respect to the smoothed accuracy of " + }, + { + "bbox": [ + 283, + 198, + 482, + 297 + ], + "type": "inline_equation", + "content": "w / o" + }, + { + "bbox": [ + 283, + 198, + 482, + 297 + ], + "type": "text", + "content": " Backbone. The smoothed accuracy refers to the averaged accuracy for a subset of tasks with lower accuracy. For example, the second x-value represents the average of two tasks with the lowest performance, while the last value is the average across all tasks" + } + ] + } + ], + "index": 11, + "angle": 0, + "type": "image_caption" + } + ], + "index": 6 + }, + { + "type": "image", + "bbox": [ + 380, + 122, + 479, + 178 + ], + "blocks": [ + { + "bbox": [ + 380, + 122, + 479, + 178 + ], + "lines": [ + { + "bbox": [ + 380, + 122, + 479, + 178 + ], + "spans": [ + { + "bbox": [ + 380, + 122, + 479, + 178 + ], + "type": "image", + "image_path": "9cd885a93119920e92c63bdbe08a52b893464870a5afabfbc16386bb638054b3.jpg" + } + ] + } + ], + "index": 8, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 381, + 180, + 477, + 189 + ], + "lines": [ + { + "bbox": [ + 381, + 180, + 477, + 189 + ], + "spans": [ + { + "bbox": [ + 381, + 180, + 477, + 189 + ], + "type": "text", + "content": "(b) Accuracy improvement" + } + ] + } + ], + "index": 9, + "angle": 0, + "type": "image_caption" + } + ], + "index": 8 + }, + { + "bbox": [ + 131, + 319, + 482, + 342 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 131, + 319, + 482, + 342 + ], + "spans": [ + { + "bbox": [ + 131, + 319, + 482, + 342 + ], + "type": "text", + "content": "Baselines. We compare Pick-a-back with diverse learning approaches of federated learning, continual learning, and continual federated learning." + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 130, + 343, + 482, + 426 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 343, + 482, + 426 + ], + "spans": [ + { + "bbox": [ + 130, + 343, + 482, + 426 + ], + "type": "text", + "content": "First, we let our Pick-a-back run on top of a certain continual learning scheme: 1) ProgressiveNet [47]: continual training only with expansion by concatenating scratch models for sequential tasks; 2) PackNet [30]: continual training only with pruning by allocating partial neurons exclusively to each task; 3) CPG [19]: continual training with expansion if necessary and pruning for the model compactness; and 4) WSN [21]: continual learning with reusing the masked weights of the prior subnetworks." + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 130, + 427, + 482, + 499 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 427, + 482, + 499 + ], + "spans": [ + { + "bbox": [ + 130, + 427, + 482, + 499 + ], + "type": "text", + "content": "We further compare across existing federated and continual federated learning approaches: 1) FedAVG [32]: one of the pioneering works for model regularization in federated learning; 2) FedWeIT [57]: a federated continual learning paradigm by introducing sparse task-specific parameters along with global federated parameters; and 3) FCCL [18]: a federated continual learning approach to tackle the catastrophic forgetting problem within a heterogeneous setting." + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 130, + 499, + 482, + 596 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 499, + 482, + 596 + ], + "spans": [ + { + "bbox": [ + 130, + 499, + 482, + 596 + ], + "type": "text", + "content": "We also compare Pick-a-back with different knowledge selection strategies: 1) " + }, + { + "bbox": [ + 130, + 499, + 482, + 596 + ], + "type": "inline_equation", + "content": "w / o" + }, + { + "bbox": [ + 130, + 499, + 482, + 596 + ], + "type": "text", + "content": " Backbone: intact continual learning without using any selective knowledge transfer, such as Progressive Networks, PackNet, or CPG; 2) Uniform Knowledge: applying the same task to its backbone model and averaging all of the results for each task; 3) Internal Knowledge: training the target task iteratively as a backbone model; 4) Random Knowledge: randomly selecting a backbone model; and 5) Upper: selecting an ideal backbone with a certain task that shows the highest accuracy improvement, being served as an unachievable upper bound." + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 130, + 605, + 483, + 666 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 605, + 483, + 666 + ], + "spans": [ + { + "bbox": [ + 130, + 605, + 483, + 666 + ], + "type": "text", + "content": "Measures. We have investigated 1) Avg: the averaged accuracy of all tasks and 2) Top: the accuracy of the most effective task that has gained the best accuracy improvement via selective knowledge. To compare the feasibility of models for edge AI, we have shown 1) Model size: the actual storage overhead of a model; 2) Expansion: the expansion degree where the scratch model has the value of 1;" + } + ] + } + ], + "index": 16 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 399, + 91, + 447, + 100 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 399, + 91, + 447, + 100 + ], + "spans": [ + { + "bbox": [ + 399, + 91, + 447, + 100 + ], + "type": "text", + "content": "Pick-a-back" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 474, + 91, + 480, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 474, + 91, + 480, + 100 + ], + "spans": [ + { + "bbox": [ + 474, + 91, + 480, + 100 + ], + "type": "text", + "content": "9" + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 8 + }, + { + "para_blocks": [ + { + "type": "table", + "bbox": [ + 134, + 156, + 480, + 268 + ], + "blocks": [ + { + "bbox": [ + 132, + 114, + 482, + 137 + ], + "lines": [ + { + "bbox": [ + 132, + 114, + 482, + 137 + ], + "spans": [ + { + "bbox": [ + 132, + 114, + 482, + 137 + ], + "type": "text", + "content": "Table 1: Performance comparison. The bold text highlights the best performance per client, excluding Upper as an ideal upper-bound due to its infeasible brute-force manner" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 134, + 156, + 480, + 268 + ], + "lines": [ + { + "bbox": [ + 134, + 156, + 480, + 268 + ], + "spans": [ + { + "bbox": [ + 134, + 156, + 480, + 268 + ], + "type": "table", + "html": "
Client ID1511716181231476824131951092011Avg
w/o Backbone40.055.057.260.451.866.067.467.669.270.872.673.074.275.275.876.477.077.481.284.069.1
FedAVG24.431.634.638.032.039.631.832.049.037.829.834.026.430.835.845.427.223.435.639.033.9
ProgressiveNet44.658.854.463.064.468.071.268.269.270.673.872.470.473.074.078.877.273.684.683.869.7
PackNet38.055.056.459.463.869.068.272.271.272.874.472.070.274.879.678.078.676.882.680.269.7
WSN34.253.849.453.459.860.263.663.066.865.064.665.466.662.661.267.666.658.674.079.861.8
FedWeIT40.052.555.056.056.260.556.760.567.065.260.265.066.064.265.065.766.751.068.777.861.0
FCCL37.054.452.460.662.268.267.264.468.270.672.872.468.477.073.875.078.668.879.278.267.5
Uniform41.558.657.662.165.670.769.771.671.272.574.773.073.776.477.178.779.475.283.583.070.8
Internal43.255.456.862.267.867.069.071.473.874.876.871.873.476.279.277.681.079.282.083.871.1
Random39.859.056.659.267.672.668.273.071.471.475.672.874.476.478.078.277.874.684.283.870.7
Upper46.462.460.064.267.874.872.275.673.875.677.075.076.079.480.080.681.679.286.285.473.7
Pick-a-back46.459.060.063.264.472.671.875.672.472.277.073.072.876.276.080.278.275.884.285.471.8
", + "image_path": "dc7eee3fcac86e583cc01e8ad75e0025996a92987e4366e0556be8908bd596c9.jpg" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "table_body" + } + ], + "index": 3 + }, + { + "bbox": [ + 130, + 300, + 482, + 337 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 300, + 482, + 337 + ], + "spans": [ + { + "bbox": [ + 130, + 300, + 482, + 337 + ], + "type": "text", + "content": "3) Sparsity: the released ratio of a model, where higher sparsity indicates more space for next tasks; 4) Training time: the execution time to run for one epoch; and 5) Used data: the number of data samples used for training." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 132, + 356, + 355, + 369 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 356, + 355, + 369 + ], + "spans": [ + { + "bbox": [ + 132, + 356, + 355, + 369 + ], + "type": "text", + "content": "4.1 Overall Performance and Comparisons" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 130, + 376, + 482, + 521 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 376, + 482, + 521 + ], + "spans": [ + { + "bbox": [ + 130, + 376, + 482, + 521 + ], + "type": "text", + "content": "Motivations for selective knowledge federation. We first examined the impact of different knowledge on various target tasks. As in Fig. 2a, a significant portion of the results and the average accuracy achieved through backbone usage are represented by a blue color, indicating that the continual learner benefited from the knowledge federation. However, there exist not only positive backbones but also negative backbones, suggesting a need for selective knowledge to assist a continual learner specific to its target task. Taking a closer look at the accuracy improvement relative to " + }, + { + "bbox": [ + 130, + 376, + 482, + 521 + ], + "type": "inline_equation", + "content": "w / o" + }, + { + "bbox": [ + 130, + 376, + 482, + 521 + ], + "type": "text", + "content": " Backbone in Fig. 2b, a well-constructed model such as task 9 or 11 is more likely disturbed by external knowledge. On the other hand, a target task with initially poor performance tends to gain a greater improvement in accuracy. It implies the importance of selectively incorporating constructive knowledge to enhance the learning process for a specific target task." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 130, + 533, + 482, + 666 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 533, + 482, + 666 + ], + "spans": [ + { + "bbox": [ + 130, + 533, + 482, + 666 + ], + "type": "text", + "content": "Ablation study on selective knowledge federation. Inspired by the Fig. 2 we verified the effectiveness, compared to diverse learning approaches, including FedAVG, ProgressiveNet, PackNet, WSN, FedWeIT, and FCCL, along with various backbone selection strategies of Uniform, Internal, Random, and Upper, as in Table 1 and in Table A5. We consistently demonstrate superior fidelity on average. By employing our selection approach, we observed a significant accuracy improvement of up to " + }, + { + "bbox": [ + 130, + 533, + 482, + 666 + ], + "type": "inline_equation", + "content": "8.0\\%" + }, + { + "bbox": [ + 130, + 533, + 482, + 666 + ], + "type": "text", + "content": ", with only two negative cases out of 20 tasks. Since FedAVG simply averages the model parameters without considering the continual tasks, it suffers from losing the knowledge of trained tasks. Notably, although Pick-a-back with CPG shows the best performance, Pick-a-back with other continual learning approaches of ProgressiveNet and PackNet also outperforms" + } + ] + } + ], + "index": 7 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "spans": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "type": "text", + "content": "10" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 166, + 91, + 247, + 101 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 166, + 91, + 247, + 101 + ], + "spans": [ + { + "bbox": [ + 166, + 91, + 247, + 101 + ], + "type": "text", + "content": "J. Yoon and H. Lee" + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 9 + }, + { + "para_blocks": [ + { + "bbox": [ + 130, + 116, + 482, + 189 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 116, + 482, + 189 + ], + "spans": [ + { + "bbox": [ + 130, + 116, + 482, + 189 + ], + "type": "inline_equation", + "content": "w / o" + }, + { + "bbox": [ + 130, + 116, + 482, + 189 + ], + "type": "text", + "content": " Backbone. Even Uniform, Internal, and Random selection strategies exhibit enhanced performance, suggesting the efficacy of inter-device knowledge transfer based on Pick-a-back primitive. We also compared with other recent (federated) continual learning approaches (e.g., WSN, FedWeIT, and FCCL), showing that Pick-a-back outperforms these methods. This is due to the effectiveness of our knowledge reuse approach along with our backbone selection strategy." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 130, + 201, + 483, + 357 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 201, + 483, + 357 + ], + "spans": [ + { + "bbox": [ + 130, + 201, + 483, + 357 + ], + "type": "text", + "content": "Benefit of backbone knowledge. Taking a closer look at how Pick-a-back benefits from backbone knowledge, in Fig. 3a, it is evident that the test accuracy of Pick-a-back consistently surpasses the training accuracy of " + }, + { + "bbox": [ + 130, + 201, + 483, + 357 + ], + "type": "inline_equation", + "content": "w / o" + }, + { + "bbox": [ + 130, + 201, + 483, + 357 + ], + "type": "text", + "content": " Backbone, starting from the beginning until the saturation. It indicates that indirect knowledge assists individual learners to be exposed to inexperienced knowledge and adapt to unseen data. Even within Pick-a-back itself, the test accuracy outperforms the training accuracy in the early stages of training, suggesting that the transferred knowledge has relevance to the target knowledge and drives toward high performance even earlier. Furthermore, " + }, + { + "bbox": [ + 130, + 201, + 483, + 357 + ], + "type": "inline_equation", + "content": "w / o" + }, + { + "bbox": [ + 130, + 201, + 483, + 357 + ], + "type": "text", + "content": " Backbone achieves an accuracy of " + }, + { + "bbox": [ + 130, + 201, + 483, + 357 + ], + "type": "inline_equation", + "content": "67.4\\%" + }, + { + "bbox": [ + 130, + 201, + 483, + 357 + ], + "type": "text", + "content": ", which can be attained only within approximately 20 epochs using Pick-a-back, whereas it takes over 50 epochs without knowledge transfer. This demonstrates that Pick-a-back not only accelerates the learning, but also enhances the overall performance. Please refer to the empirical convergence analysis in Appendix Fig. A1." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 130, + 369, + 483, + 573 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 369, + 483, + 573 + ], + "spans": [ + { + "bbox": [ + 130, + 369, + 483, + 573 + ], + "type": "text", + "content": "Possibility of similar but generalized knowledge. To justify the task-specific selection, we compared to uniform knowledge, where the same backbone is applied to all tasks. As in Fig. 3b, the results showed that Upper exhibits better improvement in tasks with poor initial performance. However, the uniform knowledge ignores task-specific improvements, leading to similar improvements across all target tasks. Interestingly, we also observed that only using internal knowledge, where the target task is trained twice, allowed further improvement. It means that even with a single task iteratively training by accumulating knowledge on top of an identical model turns out to induce better knowledge. It also implies that a backbone task with a similar, ideally the same in personal knowledge case, the decision boundary can further improve the knowledge. While the same task aims for identical knowledge, using similar but distinct knowledge enables generalization and better improvement beyond individual tasks. As expected, Pick-a-back outperformed the one only with internal knowledge, highlighting the advantage of leveraging similar indirect experiences. This means that we provide a more opportunity to acquire superior knowledge by mitigating the overfitting problem and extending to the unseen data beyond local knowledge." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 131, + 590, + 259, + 602 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 131, + 590, + 259, + 602 + ], + "spans": [ + { + "bbox": [ + 131, + 590, + 259, + 602 + ], + "type": "text", + "content": "4.2 Scalability Analysis" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 130, + 611, + 482, + 637 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 611, + 482, + 637 + ], + "spans": [ + { + "bbox": [ + 130, + 611, + 482, + 637 + ], + "type": "text", + "content": "Lightweight models for edge AI. We evaluated Pick-a-back in various architectures from the lightweight models of LeNet-5, MobileNet-V1, and EfficientNet-B0[2]" + } + ] + } + ], + "index": 6 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 399, + 91, + 447, + 100 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 399, + 91, + 447, + 100 + ], + "spans": [ + { + "bbox": [ + 399, + 91, + 447, + 100 + ], + "type": "text", + "content": "Pick-a-back" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 470, + 92, + 479, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 470, + 92, + 479, + 100 + ], + "spans": [ + { + "bbox": [ + 470, + 92, + 479, + 100 + ], + "type": "text", + "content": "11" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 133, + 642, + 482, + 665 + ], + "type": "page_footnote", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 642, + 482, + 665 + ], + "spans": [ + { + "bbox": [ + 133, + 642, + 482, + 665 + ], + "type": "text", + "content": "2 For EfficientNet-B0, we enlarge the images to " + }, + { + "bbox": [ + 133, + 642, + 482, + 665 + ], + "type": "inline_equation", + "content": "224 \\times 224" + }, + { + "bbox": [ + 133, + 642, + 482, + 665 + ], + "type": "text", + "content": " pixels for evaluation, but a model size is measured with the original size of " + }, + { + "bbox": [ + 133, + 642, + 482, + 665 + ], + "type": "inline_equation", + "content": "32 \\times 32" + }, + { + "bbox": [ + 133, + 642, + 482, + 665 + ], + "type": "text", + "content": " for a fair comparison." + } + ] + } + ], + "index": 7 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 10 + }, + { + "para_blocks": [ + { + "type": "table", + "bbox": [ + 146, + 146, + 469, + 260 + ], + "blocks": [ + { + "bbox": [ + 132, + 114, + 482, + 137 + ], + "lines": [ + { + "bbox": [ + 132, + 114, + 482, + 137 + ], + "spans": [ + { + "bbox": [ + 132, + 114, + 482, + 137 + ], + "type": "text", + "content": "Table 2: Accuracy, model size, and averaged training time per one epoch with respect to different architectures on CIFAR-100. We averaged the training time over 10 epochs" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 146, + 146, + 469, + 260 + ], + "lines": [ + { + "bbox": [ + 146, + 146, + 469, + 260 + ], + "spans": [ + { + "bbox": [ + 146, + 146, + 469, + 260 + ], + "type": "table", + "html": "
AccuracyModel SizeTraining Time
AvgTop
VGG16-BNw/o Backbone78.968.2164 MB4.10 s
Pick-a-back81.0 (+2.1)76.6 (+8.4)
EfficientNet-B0w/o Backbone77.844.620 MB9.29 s
Pick-a-back81.8 (+4.0)54.4 (+9.8)
MobileNet-V1w/o Backbone65.459.816 MB4.95 s
Pick-a-back66.7 (+1.3)65.4 (+5.6)
LeNet-5w/o Backbone69.167.60.3 MB2.18 s
Pick-a-back71.8 (+2.7)75.6 (+8.0)
", + "image_path": "9f2cb4cd3203fb9b619b390a98a668f8972af5179d67706c06e7f9240dd4a11b.jpg" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "table_body" + } + ], + "index": 3 + }, + { + "type": "table", + "bbox": [ + 137, + 291, + 477, + 357 + ], + "blocks": [ + { + "bbox": [ + 223, + 270, + 389, + 280 + ], + "lines": [ + { + "bbox": [ + 223, + 270, + 389, + 280 + ], + "spans": [ + { + "bbox": [ + 223, + 270, + 389, + 280 + ], + "type": "text", + "content": "Table 3: Various datasets with LeNet-5" + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 137, + 291, + 477, + 357 + ], + "lines": [ + { + "bbox": [ + 137, + 291, + 477, + 357 + ], + "spans": [ + { + "bbox": [ + 137, + 291, + 477, + 357 + ], + "type": "table", + "html": "
DigitsCIFAR-100Classified-ImageNetFine-Grained
AvgTopAvgTopAvgTopAvgTop
Upper95.889.473.775.661.471.833.948.5
w/o Backbone93.880.769.167.656.159.225.021.0
Pick-a-back95.186.471.875.658.167.628.327.5
(+1.3)(+5.7)(+2.7)(+8.0)(+2.0)(+8.4)(+3.3)(+6.5)
", + "image_path": "47948256f40d0203e94fd6d1832fd13d0996d882a571da41471a35c6b4a6ce23.jpg" + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "table_body" + } + ], + "index": 5 + }, + { + "bbox": [ + 130, + 378, + 482, + 511 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 378, + 482, + 511 + ], + "spans": [ + { + "bbox": [ + 130, + 378, + 482, + 511 + ], + "type": "text", + "content": "to the large-scale models of VGG16-BN in Table 2. It indicates that all models benefit from selectively transferred knowledge, leading to accuracy improvement of up to " + }, + { + "bbox": [ + 130, + 378, + 482, + 511 + ], + "type": "inline_equation", + "content": "8.4\\%" + }, + { + "bbox": [ + 130, + 378, + 482, + 511 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 130, + 378, + 482, + 511 + ], + "type": "inline_equation", + "content": "9.8\\%" + }, + { + "bbox": [ + 130, + 378, + 482, + 511 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 130, + 378, + 482, + 511 + ], + "type": "inline_equation", + "content": "5.6\\%" + }, + { + "bbox": [ + 130, + 378, + 482, + 511 + ], + "type": "text", + "content": ", and " + }, + { + "bbox": [ + 130, + 378, + 482, + 511 + ], + "type": "inline_equation", + "content": "8.0\\%" + }, + { + "bbox": [ + 130, + 378, + 482, + 511 + ], + "type": "text", + "content": ", respectively. However, in resource-constrained environments, the limitations of memory and processing power make it challenging to deploy a large-scale model at the edge. To make learning practical in the real-world, lightweight models can be flexibly deployed across multiple edge devices, reducing dependence on remote servers. The lightweight LeNet-5 has a much smaller size of 0.18 times compared to VGG16-BN, enabling faster training and inference time. Although the lightweight models may not match closely to the large-scale models, Pick-a-back can still achieve improvement, comparable to the large models, while greatly reducing the storage requirement and run-time." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 130, + 521, + 483, + 666 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 521, + 483, + 666 + ], + "spans": [ + { + "bbox": [ + 130, + 521, + 483, + 666 + ], + "type": "text", + "content": "Scalability across diverse datasets. We assessed the scalability of Pick-a-back across various datasets in Table 3. We stably achieve improvements across different datasets, even with varying degrees of class complexity. In Digits, where most tasks except for the Roman task are generally easy to classify, the average accuracy improvements may appear modest, but Pick-a-back approach achieved a significant accuracy improvement of up to " + }, + { + "bbox": [ + 130, + 521, + 483, + 666 + ], + "type": "inline_equation", + "content": "5.72\\%" + }, + { + "bbox": [ + 130, + 521, + 483, + 666 + ], + "type": "text", + "content": " in the relatively challenging Roman task. Similarly, in the classified-ImageNet dataset, where there exist enough neighboring tasks to provide additional knowledge, Pick-a-back shows substantial improvements even for the underperforming tasks. However, when considering fine-grained datasets with a large number of classes per task (ranging " + }, + { + "bbox": [ + 130, + 521, + 483, + 666 + ], + "type": "inline_equation", + "content": "102\\sim 250" + }, + { + "bbox": [ + 130, + 521, + 483, + 666 + ], + "type": "text", + "content": " classes), classification becomes inherently more difficult with simple models, resulting in relatively lower accuracy. Nonetheless, by applying Pick" + } + ] + } + ], + "index": 7 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "spans": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "type": "text", + "content": "12" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 166, + 91, + 247, + 100 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 166, + 91, + 247, + 100 + ], + "spans": [ + { + "bbox": [ + 166, + 91, + 247, + 100 + ], + "type": "text", + "content": "J. Yoon and H. Lee" + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 11 + }, + { + "para_blocks": [ + { + "type": "table", + "bbox": [ + 135, + 146, + 479, + 196 + ], + "blocks": [ + { + "bbox": [ + 132, + 114, + 480, + 137 + ], + "lines": [ + { + "bbox": [ + 132, + 114, + 480, + 137 + ], + "spans": [ + { + "bbox": [ + 132, + 114, + 480, + 137 + ], + "type": "text", + "content": "Table 4: Performance on Fine-Grained Table 5: Heterogeneous data sampling datasets using EfficientNet-B0 with " + }, + { + "bbox": [ + 132, + 114, + 480, + 137 + ], + "type": "inline_equation", + "content": "\\alpha = 0.5" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 135, + 146, + 479, + 196 + ], + "lines": [ + { + "bbox": [ + 135, + 146, + 479, + 196 + ], + "spans": [ + { + "bbox": [ + 135, + 146, + 479, + 196 + ], + "type": "table", + "html": "
Client ID12345AvgHeteroBalancedw/o Sampling
AvgTopAvgTopAvgTop
Pretrained w/o Backbone28.990.542.449.937.149.8Upper59.360.269.684.473.775.6
66.37.428.212.724.6w/o Backbone52.640.264.664.171.869.167.6
Pick-a-back49.5 (+41.2)77.4 (+11.1)44.0 (+36.6)46.4 (+18.2)64.6 (+51.9)56.4 (+31.8)Pick-a-back56.1 (+3.5)58.6 (+18.4)67.3 (+3.2)80.0 (+8.2)71.8 (+2.7)75.6 (+8.0)
", + "image_path": "28a0dd5caee43129cd069c7e3d5a9c931f1d797aa28383069aeb081bb9cb8aea.jpg" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "table_body" + } + ], + "index": 3 + }, + { + "bbox": [ + 130, + 220, + 482, + 316 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 220, + 482, + 316 + ], + "spans": [ + { + "bbox": [ + 130, + 220, + 482, + 316 + ], + "type": "inline_equation", + "content": "a" + }, + { + "bbox": [ + 130, + 220, + 482, + 316 + ], + "type": "text", + "content": "-back, it is possible to advance the knowledge and achieve performance gains. While there is a noticeable performance gap of " + }, + { + "bbox": [ + 130, + 220, + 482, + 316 + ], + "type": "inline_equation", + "content": "0.7\\%" + }, + { + "bbox": [ + 130, + 220, + 482, + 316 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 130, + 220, + 482, + 316 + ], + "type": "inline_equation", + "content": "1.9\\%" + }, + { + "bbox": [ + 130, + 220, + 482, + 316 + ], + "type": "text", + "content": ", and " + }, + { + "bbox": [ + 130, + 220, + 482, + 316 + ], + "type": "inline_equation", + "content": "3.3\\%" + }, + { + "bbox": [ + 130, + 220, + 482, + 316 + ], + "type": "text", + "content": ", respectively, between Pick-a-back and Upper in other datasets, the gap in the fine-grained dataset is quite large with " + }, + { + "bbox": [ + 130, + 220, + 482, + 316 + ], + "type": "inline_equation", + "content": "5.6\\%" + }, + { + "bbox": [ + 130, + 220, + 482, + 316 + ], + "type": "text", + "content": " on average and up to " + }, + { + "bbox": [ + 130, + 220, + 482, + 316 + ], + "type": "inline_equation", + "content": "21.0\\%" + }, + { + "bbox": [ + 130, + 220, + 482, + 316 + ], + "type": "text", + "content": " for the task of flowers. This implies the potential of Pick-a-back to drive further improvements in complex tasks that require more sophisticated classification. We further explore the performance with the different number of clients (i.e., classes) and the different learning seeds in Appendix Table A9 and in Appendix Table A10, respectively." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 130, + 328, + 482, + 497 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 328, + 482, + 497 + ], + "spans": [ + { + "bbox": [ + 130, + 328, + 482, + 497 + ], + "type": "text", + "content": "Impact on the most complicated task. We investigate the most complicated tasks of CUB 55, Stanford Cars 24, Flower 37, WikiArt 48, and Sketch 10, with each featuring a diverse range of class numbers of 200, 196, 102, 195, and 250 and running at each separate client. As shown in Table 4 with the adoption of Pick-a-back, we observe remarkable improvements in overall accuracy, averaging at " + }, + { + "bbox": [ + 130, + 328, + 482, + 497 + ], + "type": "inline_equation", + "content": "31.8\\%" + }, + { + "bbox": [ + 130, + 328, + 482, + 497 + ], + "type": "text", + "content": ", and peaking at an impressive " + }, + { + "bbox": [ + 130, + 328, + 482, + 497 + ], + "type": "inline_equation", + "content": "51.9\\%" + }, + { + "bbox": [ + 130, + 328, + 482, + 497 + ], + "type": "text", + "content": ". Although these results might appear surprising, they align with the findings previously established by 19 and 29, where scratch accuracy on ImageNet-pretrained models of ResNet50 and VGG16-BN, achieved " + }, + { + "bbox": [ + 130, + 328, + 482, + 497 + ], + "type": "inline_equation", + "content": "58.83\\%" + }, + { + "bbox": [ + 130, + 328, + 482, + 497 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 130, + 328, + 482, + 497 + ], + "type": "inline_equation", + "content": "43.84\\%" + }, + { + "bbox": [ + 130, + 328, + 482, + 497 + ], + "type": "text", + "content": ", respectively. Incorporating Pick-a-back not only surpasses the average accuracy of ImageNet-pretrained models, but does so through a judicious infusion of tailored knowledge. It is worth noting that ImageNet-pretrained models often necessitate extensive prior training data. However, in many instances, the efficacy diminishes due to the lack of distinct knowledge selection, showing some critical drawbacks in clients #1, #3, and #5." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 130, + 510, + 482, + 666 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 510, + 482, + 666 + ], + "spans": [ + { + "bbox": [ + 130, + 510, + 482, + 666 + ], + "type": "text", + "content": "Superiority in heterogeneous data scenarios. In real-world scenarios, there are many unique challenges such as personalization [5] or heterogeneity [4]. Here, we explored the effectiveness of Pick-a-back under data heterogeneity. Leveraging the heterogeneity settings from federated learning researches [16, 60], we applied " + }, + { + "bbox": [ + 130, + 510, + 482, + 666 + ], + "type": "inline_equation", + "content": "\\alpha" + }, + { + "bbox": [ + 130, + 510, + 482, + 666 + ], + "type": "text", + "content": " of 0.5 based on the Dirichlet distribution " + }, + { + "bbox": [ + 130, + 510, + 482, + 666 + ], + "type": "inline_equation", + "content": "Dir(\\alpha)" + }, + { + "bbox": [ + 130, + 510, + 482, + 666 + ], + "type": "text", + "content": ". As in Table [5] we observed a degradation in the accuracy of scratch models as the data heterogeneity increases. It implies that the accuracy is achievable, and the external knowledge can be particularly beneficial for weaker classes in the presence of heterogeneous data. While the average accuracy improvement remained consistent, we achieved significant enhancements in heterogeneous data settings. It is essential to address unbalanced data distributions, as certain classes may receive less training attention. By leveraging Pick-a-back, we are able to achieve substantial performance gains, with improvements of up to " + }, + { + "bbox": [ + 130, + 510, + 482, + 666 + ], + "type": "inline_equation", + "content": "18.4\\%" + }, + { + "bbox": [ + 130, + 510, + 482, + 666 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 6 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 399, + 91, + 447, + 100 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 399, + 91, + 447, + 100 + ], + "spans": [ + { + "bbox": [ + 399, + 91, + 447, + 100 + ], + "type": "text", + "content": "Pick-a-back" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 470, + 91, + 480, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 470, + 91, + 480, + 100 + ], + "spans": [ + { + "bbox": [ + 470, + 91, + 480, + 100 + ], + "type": "text", + "content": "13" + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 12 + }, + { + "para_blocks": [ + { + "type": "table", + "bbox": [ + 175, + 145, + 440, + 196 + ], + "blocks": [ + { + "bbox": [ + 132, + 114, + 481, + 126 + ], + "lines": [ + { + "bbox": [ + 132, + 114, + 481, + 126 + ], + "spans": [ + { + "bbox": [ + 132, + 114, + 481, + 126 + ], + "type": "text", + "content": "Table 6: Performance and data used for training in ImageNet-pretrained VGG16-BN" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 175, + 145, + 440, + 196 + ], + "lines": [ + { + "bbox": [ + 175, + 145, + 440, + 196 + ], + "spans": [ + { + "bbox": [ + 175, + 145, + 440, + 196 + ], + "type": "table", + "html": "
w/o Backbone → Pick-a-backSize of Training Data
AvgTop
w/ Pretrained84.9 → 83.882.8 → 84.492.8 GB
w/o Pretrained78.9 → 81.068.2 → 76.60.3 GB
", + "image_path": "7defe012871b18bc7b914b931c7ed6e9a09e75a42ab83dc8a59849c1053f5977.jpg" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "table_body" + } + ], + "index": 3 + }, + { + "bbox": [ + 132, + 226, + 301, + 237 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 226, + 301, + 237 + ], + "spans": [ + { + "bbox": [ + 132, + 226, + 301, + 237 + ], + "type": "text", + "content": "4.3 Discussions and Limitations" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 130, + 243, + 482, + 399 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 243, + 482, + 399 + ], + "spans": [ + { + "bbox": [ + 130, + 243, + 482, + 399 + ], + "type": "text", + "content": "ImageNet-pretrained models. It is important to note that a lightweight model such as LeNet-5 may not outperform a large-scale model even with Pick-a-back. Also, VGG16-BN models initialized with ImageNet-pretrained weights demonstrate strong performance generally for visual tasks as in Table 6 and in Appendix Table A6. However, their effectiveness heavily relies on the availability of large-scale labeled datasets that need to be uploaded to a central server. Additionally, the pretrained models are specifically designed for image classification, and their performance may vary when applied to specific or specialized tasks. Furthermore, the availability of pretrained models is limited only to well-known large-scale architectures, making their deployment usually infeasible to edge learning. Without pretrained model, selective knowledge federation through Pick-a-back achieves similar performance on average with a small gap of " + }, + { + "bbox": [ + 130, + 243, + 482, + 399 + ], + "type": "inline_equation", + "content": "3.9\\%" + }, + { + "bbox": [ + 130, + 243, + 482, + 399 + ], + "type": "text", + "content": " compared to the original pretrained model (w/o Backbone), but only with " + }, + { + "bbox": [ + 130, + 243, + 482, + 399 + ], + "type": "inline_equation", + "content": "0.3\\%" + }, + { + "bbox": [ + 130, + 243, + 482, + 399 + ], + "type": "text", + "content": " of training data." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 130, + 399, + 482, + 472 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 399, + 482, + 472 + ], + "spans": [ + { + "bbox": [ + 130, + 399, + 482, + 472 + ], + "type": "text", + "content": "We further provided the orthogonality of continual learner in Pick-a-back and catastrophic forgetting prevention across different continual learning approaches in Appendix Table A11. Lastly, we need image sharing among clients and model sharing from a selected neighboring client, which may incur transmission overhead and privacy issues. We thus further assess the communication and computation time and how we can defend against attacks in Appendix Table A12 and Fig. A2." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 132, + 487, + 220, + 499 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 487, + 220, + 499 + ], + "spans": [ + { + "bbox": [ + 132, + 487, + 220, + 499 + ], + "type": "text", + "content": "5 Conclusion" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 130, + 510, + 482, + 557 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 510, + 482, + 557 + ], + "spans": [ + { + "bbox": [ + 130, + 510, + 482, + 557 + ], + "type": "text", + "content": "We propose Pick-a-back that picks a task-specific backbone knowledge and fosters the collective intelligence in federated continual learning. As the individual learner pursues each distinct goal with private heterogeneous data, we selectively transfer the relevant external knowledge to the target task between neighboring learners." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 130, + 558, + 482, + 666 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 558, + 482, + 666 + ], + "spans": [ + { + "bbox": [ + 130, + 558, + 482, + 666 + ], + "type": "text", + "content": "As AI expands towards edges for scalability, low-end devices tend to be trained toward personalized and specialized but somewhat partially limited intelligence. Here, in case of a scarcity of data or with limited performance, it becomes essential to seek assistance from external sources. The question - how can a device make a local model more effectively trained with exterior knowledge? - can be answered: Pick-a-back offers an effective way to selective knowledge transfer from one device to another towards distributed edge intelligence. By harnessing the power of neighboring devices, individual learners can gain access to valuable insights and expertise via collective intelligence, augmenting their own capabilities." + } + ] + } + ], + "index": 9 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "spans": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "type": "text", + "content": "14" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 166, + 91, + 247, + 100 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 166, + 91, + 247, + 100 + ], + "spans": [ + { + "bbox": [ + 166, + 91, + 247, + 100 + ], + "type": "text", + "content": "J. Yoon and H. Lee" + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 13 + }, + { + "para_blocks": [ + { + "bbox": [ + 133, + 114, + 240, + 129 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 114, + 240, + 129 + ], + "spans": [ + { + "bbox": [ + 133, + 114, + 240, + 129 + ], + "type": "text", + "content": "Acknowledgments" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 130, + 138, + 482, + 224 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 138, + 482, + 224 + ], + "spans": [ + { + "bbox": [ + 130, + 138, + 482, + 224 + ], + "type": "text", + "content": "We thank the Virginia Tech Presidential Postdoctoral Fellowship Program for supporting JinYi Yoon's postdoctoral associate position. This work was supported by the Institute of Information & communications Technology Planning and Evaluation (IITP) grant funded by the Korea government (MSIT) (No. RS-2022-00155966, Artificial Intelligence Convergence Innovation Human Resources Development (Ewha Womans University) and No. RS-2021-II212068, Artificial Intelligence Innovation Hub)." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 133, + 239, + 197, + 252 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 239, + 197, + 252 + ], + "spans": [ + { + "bbox": [ + 133, + 239, + 197, + 252 + ], + "type": "text", + "content": "References" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 134, + 262, + 482, + 665 + ], + "type": "list", + "angle": 0, + "index": 18, + "blocks": [ + { + "bbox": [ + 138, + 262, + 481, + 285 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 262, + 481, + 285 + ], + "spans": [ + { + "bbox": [ + 138, + 262, + 481, + 285 + ], + "type": "text", + "content": "1. Arani, E., et al.: Learning fast, learning slow: A general continual learning method based on complementary learning system. ICLR (2022)" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 138, + 285, + 482, + 306 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 285, + 482, + 306 + ], + "spans": [ + { + "bbox": [ + 138, + 285, + 482, + 306 + ], + "type": "text", + "content": "2. Cao, L.: Decentralized ai: Edge intelligence and smart blockchain, metaverse, web3, and desci. IEEE Intelligent Systems 37(3), 6-19 (2022)" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 138, + 307, + 482, + 339 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 307, + 482, + 339 + ], + "spans": [ + { + "bbox": [ + 138, + 307, + 482, + 339 + ], + "type": "text", + "content": "3. Casado, F.E., Lema, D., Criado, M.F., Iglesias, R., Regueiro, C.V., Barro, S.: Concept drift detection and adaptation for federated and continual learning. Multimedia Tools and Applications pp. 1-23 (2022)" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 138, + 339, + 481, + 372 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 339, + 481, + 372 + ], + "spans": [ + { + "bbox": [ + 138, + 339, + 481, + 372 + ], + "type": "text", + "content": "4. Chen, D., Gao, D., Kuang, W., Li, Y., Ding, B.: pfl-bench: A comprehensive benchmark for personalized federated learning. Advances in Neural Information Processing Systems 35, 9344-9360 (2022)" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 138, + 373, + 481, + 415 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 373, + 481, + 415 + ], + "spans": [ + { + "bbox": [ + 138, + 373, + 481, + 415 + ], + "type": "text", + "content": "5. Chen, D., Gao, D., Xie, Y., Pan, X., Li, Z., Li, Y., Ding, B., Zhou, J.: Fs-real: Towards real-world cross-device federated learning. In: Proceedings of the 29th ACM SIGKDD Conference on Knowledge Discovery and Data Mining. pp. 3829-3841 (2023)" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 138, + 416, + 481, + 448 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 416, + 481, + 448 + ], + "spans": [ + { + "bbox": [ + 138, + 416, + 481, + 448 + ], + "type": "text", + "content": "6. Chen, Y., Inaltekin, H., Gorlatova, M.: Adaptslam: Edge-assisted adaptive slam with resource constraints via uncertainty minimization. arXiv preprint arXiv:2301.04620 (2023)" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 138, + 449, + 481, + 470 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 449, + 481, + 470 + ], + "spans": [ + { + "bbox": [ + 138, + 449, + 481, + 470 + ], + "type": "text", + "content": "7. Colianni, S.: Mnist as .jpg (2017), https://www.kaggle.com/scolianni/ mnistas.jpg" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 138, + 471, + 481, + 513 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 471, + 481, + 513 + ], + "spans": [ + { + "bbox": [ + 138, + 471, + 481, + 513 + ], + "type": "text", + "content": "8. Deng, J., Dong, W., Socher, R., Li, L.J., Li, K., Fei-Fei, L.: Imagenet: A large-scale hierarchical image database. In: 2009 IEEE Conference on Computer Vision and Pattern Recognition. pp. 248-255 (2009). https://doi.org/10.1109/CVPR.2009.5206848" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 138, + 514, + 481, + 546 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 514, + 481, + 546 + ], + "spans": [ + { + "bbox": [ + 138, + 514, + 481, + 546 + ], + "type": "text", + "content": "9. Dhar, P., Singh, R.V., Peng, K.C., Wu, Z., Chellappa, R.: Learning without memorizing. In: Proceedings of the IEEE/CVF conference on computer vision and pattern recognition. pp. 5138-5146 (2019)" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 134, + 547, + 481, + 568 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 134, + 547, + 481, + 568 + ], + "spans": [ + { + "bbox": [ + 134, + 547, + 481, + 568 + ], + "type": "text", + "content": "10. Eitz, M., Hays, J., Alexa, M.: How do humans sketch objects? ACM Transactions on graphics (TOG) 31(4), 1-10 (2012)" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 134, + 568, + 481, + 600 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 134, + 568, + 481, + 600 + ], + "spans": [ + { + "bbox": [ + 134, + 568, + 481, + 600 + ], + "type": "text", + "content": "11. El-Sawy, A., El-Bakry, H., Loey, M.: Cnn for handwritten arabic digits recognition based on lenet-5. In: Proceedings of the International Conference on Advanced Intelligent Systems and Informatics 2016 2. pp. 566-575. Springer (2017)" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 134, + 601, + 481, + 622 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 134, + 601, + 481, + 622 + ], + "spans": [ + { + "bbox": [ + 134, + 601, + 481, + 622 + ], + "type": "text", + "content": "12. Goodfellow, I., Bengio, Y., Courville, A.: Deep Learning. MIT Press (2016), http://www.deeplearningbook.org" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 134, + 622, + 481, + 665 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 134, + 622, + 481, + 665 + ], + "spans": [ + { + "bbox": [ + 134, + 622, + 481, + 665 + ], + "type": "text", + "content": "13. Gu, Y., Yang, X., Wei, K., Deng, C.: Not just selection, but exploration: Online class-incremental continual learning via dual view consistency. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 7442-7451 (2022)" + } + ] + } + ], + "index": 17 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 399, + 91, + 447, + 100 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 399, + 91, + 447, + 100 + ], + "spans": [ + { + "bbox": [ + 399, + 91, + 447, + 100 + ], + "type": "text", + "content": "Pick-a-back" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 470, + 92, + 481, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 470, + 92, + 481, + 100 + ], + "spans": [ + { + "bbox": [ + 470, + 92, + 481, + 100 + ], + "type": "text", + "content": "15" + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 14 + }, + { + "para_blocks": [ + { + "bbox": [ + 132, + 116, + 481, + 665 + ], + "type": "list", + "angle": 0, + "index": 20, + "blocks": [ + { + "bbox": [ + 133, + 116, + 480, + 138 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 116, + 480, + 138 + ], + "spans": [ + { + "bbox": [ + 133, + 116, + 480, + 138 + ], + "type": "text", + "content": "14. Ha, D., Tang, Y.: Collective intelligence for deep learning: A survey of recent developments. Collective Intelligence 1(1), 263391372211114874 (2022)" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 133, + 138, + 481, + 170 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 138, + 481, + 170 + ], + "spans": [ + { + "bbox": [ + 133, + 138, + 481, + 170 + ], + "type": "text", + "content": "15. Haim, N., Vardi, G., Yehudai, G., Shamir, O., Irani, M.: Reconstructing training data from trained neural networks. Advances in Neural Information Processing Systems 35, 22911-22924 (2022)" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 133, + 170, + 481, + 203 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 170, + 481, + 203 + ], + "spans": [ + { + "bbox": [ + 133, + 170, + 481, + 203 + ], + "type": "text", + "content": "16. He, C., Li, S., So, J., Zeng, X., Zhang, M., Wang, H., Wang, X., Vepakomma, P., Singh, A., Qiu, H., et al.: Fedml: A research library and benchmark for federated machine learning. arXiv preprint arXiv:2007.13518 (2020)" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 132, + 203, + 481, + 236 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 203, + 481, + 236 + ], + "spans": [ + { + "bbox": [ + 132, + 203, + 481, + 236 + ], + "type": "text", + "content": "17. Howard, A.G., Zhu, M., Chen, B., Kalenichenko, D., Wang, W., Weyand, T., Andreetto, M., Adam, H.: Mobilenets: Efficient convolutional neural networks for mobile vision applications. arXiv preprint arXiv:1704.04861 (2017)" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 132, + 236, + 481, + 267 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 236, + 481, + 267 + ], + "spans": [ + { + "bbox": [ + 132, + 236, + 481, + 267 + ], + "type": "text", + "content": "18. Huang, W., Ye, M., Du, B.: Learn from others and be yourself in heterogeneous federated learning. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 10143-10153 (2022)" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 132, + 267, + 481, + 300 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 267, + 481, + 300 + ], + "spans": [ + { + "bbox": [ + 132, + 267, + 481, + 300 + ], + "type": "text", + "content": "19. Hung, C.Y., Tu, C.H., Wu, C.E., Chen, C.H., Chan, Y.M., Chen, C.S.: Compacting, picking and growing for unforgetting continual learning. Advances in Neural Information Processing Systems 32 (2019)" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 132, + 300, + 481, + 331 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 300, + 481, + 331 + ], + "spans": [ + { + "bbox": [ + 132, + 300, + 481, + 331 + ], + "type": "text", + "content": "20. Jeong, W., Hwang, S.J.: Factorized-fl: Personalized federated learning with parameter factorization & similarity matching. Advances in Neural Information Processing Systems 35, 35684-35695 (2022)" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 132, + 332, + 481, + 365 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 332, + 481, + 365 + ], + "spans": [ + { + "bbox": [ + 132, + 332, + 481, + 365 + ], + "type": "text", + "content": "21. Kang, H., Mina, R.J.L., Madjid, S.R.H., Yoon, J., Hasegawa-Johnson, M., Hwang, S.J., Yoo, C.D.: Forget-free continual learning with winning subnetworks. In: International Conference on Machine Learning. pp. 10734-10750. PMLR (2022)" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 132, + 365, + 481, + 396 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 365, + 481, + 396 + ], + "spans": [ + { + "bbox": [ + 132, + 365, + 481, + 396 + ], + "type": "text", + "content": "22. Kim, G., Esmaeilpour, S., Xiao, C., Liu, B.: Continual learning based on ood detection and task masking. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 3856-3866 (2022)" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 132, + 396, + 481, + 440 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 396, + 481, + 440 + ], + "spans": [ + { + "bbox": [ + 132, + 396, + 481, + 440 + ], + "type": "text", + "content": "23. Kirkpatrick, J., Pascanu, R., Rabinowitz, N., Veness, J., Desjardins, G., Rusu, A.A., Milan, K., Quan, J., Ramalho, T., Grabska-Barwinska, A., et al.: Overcoming catastrophic forgetting in neural networks. Proceedings of the national academy of sciences 114(13), 3521-3526 (2017)" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 132, + 440, + 481, + 472 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 440, + 481, + 472 + ], + "spans": [ + { + "bbox": [ + 132, + 440, + 481, + 472 + ], + "type": "text", + "content": "24. Krause, J., Stark, M., Deng, J., Fei-Fei, L.: 3d object representations for fine-grained categorization. In: Proceedings of the IEEE international conference on computer vision workshops. pp. 554-561 (2013)" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 132, + 472, + 481, + 493 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 472, + 481, + 493 + ], + "spans": [ + { + "bbox": [ + 132, + 472, + 481, + 493 + ], + "type": "text", + "content": "25. Krizhevsky, A., Hinton, G., et al.: Learning multiple layers of features from tiny images (2009)" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 132, + 493, + 481, + 515 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 493, + 481, + 515 + ], + "spans": [ + { + "bbox": [ + 132, + 493, + 481, + 515 + ], + "type": "text", + "content": "26. LeCun, Y., et al.: Lenet-5, convolutional neural networks. URL: http://yann.lecun.com/exdb/lenet 20(5), 14 (2015)" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 132, + 515, + 481, + 547 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 515, + 481, + 547 + ], + "spans": [ + { + "bbox": [ + 132, + 515, + 481, + 547 + ], + "type": "text", + "content": "27. Li, Y., Zhang, Z., Liu, B., Yang, Z., Liu, Y.: Modeldiff: testing-based dnn similarity comparison for model reuse detection. In: Proceedings of the 30th ACM SIGSOFT International Symposium on Software Testing and Analysis. pp. 139-151 (2021)" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 132, + 547, + 481, + 579 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 547, + 481, + 579 + ], + "spans": [ + { + "bbox": [ + 132, + 547, + 481, + 579 + ], + "type": "text", + "content": "28. Liu, F., Wu, X., Ge, S., Fan, W., Zou, Y.: Federated learning for vision-and-language grounding problems. In: Proceedings of the AAAI Conference on Artificial Intelligence. vol. 34, pp. 11572-11579 (2020)" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 132, + 579, + 481, + 611 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 579, + 481, + 611 + ], + "spans": [ + { + "bbox": [ + 132, + 579, + 481, + 611 + ], + "type": "text", + "content": "29. Mallya, A., Davis, D., Lazebnik, S.: Piggyback: Adapting a single network to multiple tasks by learning to mask weights. In: Proceedings of the European Conference on Computer Vision (ECCV). pp. 67-82 (2018)" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 132, + 611, + 481, + 643 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 611, + 481, + 643 + ], + "spans": [ + { + "bbox": [ + 132, + 611, + 481, + 643 + ], + "type": "text", + "content": "30. Mallya, A., Lazebnik, S.: Packnet: Adding multiple tasks to a single network by iterative pruning. In: Proceedings of the IEEE conference on Computer Vision and Pattern Recognition. pp. 7765-7773 (2018)" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 132, + 643, + 481, + 665 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 643, + 481, + 665 + ], + "spans": [ + { + "bbox": [ + 132, + 643, + 481, + 665 + ], + "type": "text", + "content": "31. Mao, B., Liu, J., Wu, Y., Kato, N.: Security and privacy on 6g network edge: A survey. IEEE Communications Surveys & Tutorials (2023)" + } + ] + } + ], + "index": 19 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "spans": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "type": "text", + "content": "16" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 166, + 91, + 247, + 100 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 166, + 91, + 247, + 100 + ], + "spans": [ + { + "bbox": [ + 166, + 91, + 247, + 100 + ], + "type": "text", + "content": "J. Yoon and H. Lee" + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 15 + }, + { + "para_blocks": [ + { + "bbox": [ + 130, + 116, + 482, + 666 + ], + "type": "list", + "angle": 0, + "index": 19, + "blocks": [ + { + "bbox": [ + 130, + 116, + 482, + 149 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 116, + 482, + 149 + ], + "spans": [ + { + "bbox": [ + 130, + 116, + 482, + 149 + ], + "type": "text", + "content": "32. McMahan, B., Moore, E., Ramage, D., Hampson, S., y Arcas, B.A.: Communication-efficient learning of deep networks from decentralized data. In: Artificial intelligence and statistics. pp. 1273-1282. PMLR (2017)" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 130, + 150, + 482, + 183 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 150, + 482, + 183 + ], + "spans": [ + { + "bbox": [ + 130, + 150, + 482, + 183 + ], + "type": "text", + "content": "33. Meng, C., Tang, M., Setayesh, M., Wong, V.W.: Gnn-based neighbor selection and resource allocation for decentralized federated learning. In: GLOBECOM 2023-2023 IEEE Global Communications Conference. pp. 1223-1228. IEEE (2023)" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 130, + 183, + 482, + 205 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 183, + 482, + 205 + ], + "spans": [ + { + "bbox": [ + 130, + 183, + 482, + 205 + ], + "type": "text", + "content": "34. Miller, G.A.: Wordnet: a lexical database for english. Communications of the ACM 38(11), 39-41 (1995)" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 130, + 205, + 482, + 237 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 205, + 482, + 237 + ], + "spans": [ + { + "bbox": [ + 130, + 205, + 482, + 237 + ], + "type": "text", + "content": "35. Nazarpour, K., Chen, M.: Handwritten Chinese Numbers (1 2017). https://doi.org/10.17634/137930-3, https://data.ncl.ac.uk/articles/dataset/Handwritten_Chinese_Numbers/10280831" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 130, + 237, + 482, + 259 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 237, + 482, + 259 + ], + "spans": [ + { + "bbox": [ + 130, + 237, + 482, + 259 + ], + "type": "text", + "content": "36. Netzer, Y., Wang, T., Coates, A., Bissacco, A., Wu, B., Ng, A.Y.: Reading digits in natural images with unsupervised feature learning (2011)" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 130, + 259, + 482, + 293 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 259, + 482, + 293 + ], + "spans": [ + { + "bbox": [ + 130, + 259, + 482, + 293 + ], + "type": "text", + "content": "37. Nilsback, M.E., Zisserman, A.: Automated flower classification over a large number of classes. In: 2008 Sixth Indian Conference on Computer Vision, Graphics & Image Processing. pp. 722-729. IEEE (2008)" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 130, + 293, + 482, + 326 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 293, + 482, + 326 + ], + "spans": [ + { + "bbox": [ + 130, + 293, + 482, + 326 + ], + "type": "text", + "content": "38. Nishio, T., Yonetani, R.: Client selection for federated learning with heterogeneous resources in mobile edge. In: ICC 2019-2019 IEEE international conference on communications (ICC). pp. 1-7. IEEE (2019)" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 130, + 326, + 482, + 347 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 326, + 482, + 347 + ], + "spans": [ + { + "bbox": [ + 130, + 326, + 482, + 347 + ], + "type": "text", + "content": "39. Noyan, M.A.: Hand written roman numerals dataset (Sep 2021). https://doi.org/10.5281/zenodo.5385144, https://doi.org/10.5281/zenodo.5385144" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 130, + 347, + 482, + 392 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 347, + 482, + 392 + ], + "spans": [ + { + "bbox": [ + 130, + 347, + 482, + 392 + ], + "type": "text", + "content": "40. Ostapenko, O., Puscas, M., Klein, T., Jahnichen, P., Nabi, M.: Learning to remember: A synaptic plasticity driven framework for continual learning. In: Proceedings of the IEEE/CVF conference on computer vision and pattern recognition. pp. 11321-11329 (2019)" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 130, + 392, + 482, + 456 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 392, + 482, + 456 + ], + "spans": [ + { + "bbox": [ + 130, + 392, + 482, + 456 + ], + "type": "text", + "content": "41. Parks Associates: At CES 2024, Parks Associates Announces New Research Showing Average Number of Connected Devices per US Internet Household Reached 17 in 2023. https://www.parksassociates.com/blogs/press-releases/at-ces-2024-parks-associates-announces-new-research-showing-average-number-of-connected-devices-per-us-internet-household-reached-17-in-2023 (2024), accessed: 2024-05-15" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 130, + 456, + 482, + 478 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 456, + 482, + 478 + ], + "spans": [ + { + "bbox": [ + 130, + 456, + 482, + 478 + ], + "type": "text", + "content": "42. PyTorch: Retrieved from python model zoo (2017), https://pytorch.org/vision/stable/models.html" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 130, + 479, + 482, + 522 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 479, + 482, + 522 + ], + "spans": [ + { + "bbox": [ + 130, + 479, + 482, + 522 + ], + "type": "text", + "content": "43. Radford, A., Kim, J.W., Hallacy, C., Ramesh, A., Goh, G., Agarwal, S., Sastry, G., Askell, A., Mishkin, P., Clark, J., et al.: Learning transferable visual models from natural language supervision. In: International conference on machine learning. pp. 8748-8763. PMLR (2021)" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 130, + 522, + 482, + 555 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 522, + 482, + 555 + ], + "spans": [ + { + "bbox": [ + 130, + 522, + 482, + 555 + ], + "type": "text", + "content": "44. Roy, A.G., Siddiqui, S., Pölsterl, S., Navab, N., Wachinger, C.: Brainstorm: A peer-to-peer environment for decentralized federated learning. arXiv preprint arXiv:1905.06731 (2019)" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 130, + 555, + 482, + 578 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 555, + 482, + 578 + ], + "spans": [ + { + "bbox": [ + 130, + 555, + 482, + 578 + ], + "type": "text", + "content": "45. Roy, P., Ghosh, S., Bhattacharya, S., Pal, U.: Effects of degradations on deep neural network architectures. arXiv preprint arXiv:1807.10108 (2018)" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 130, + 578, + 482, + 611 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 578, + 482, + 611 + ], + "spans": [ + { + "bbox": [ + 130, + 578, + 482, + 611 + ], + "type": "text", + "content": "46. Rusu, A.A., Rabinowitz, N.C., Desjardins, G., Soyer, H., Kirkpatrick, J., Kavukcuoglu, K., Pascanu, R., Hadsell, R.: Progressive neural networks. arXiv preprint arXiv:1606.04671 (2016)" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 130, + 611, + 482, + 643 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 611, + 482, + 643 + ], + "spans": [ + { + "bbox": [ + 130, + 611, + 482, + 643 + ], + "type": "text", + "content": "47. Rusu, A.A., Rabinowitz, N.C., Desjardins, G., Soyer, H., Kirkpatrick, J., Kavukcuoglu, K., Pascanu, R., Hadsell, R.: Progressive neural networks. arXiv preprint arXiv:1606.04671 (2016)" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 130, + 643, + 482, + 666 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 643, + 482, + 666 + ], + "spans": [ + { + "bbox": [ + 130, + 643, + 482, + 666 + ], + "type": "text", + "content": "48. Saleh, B., Elgammal, A.: Large-scale classification of fine-art paintings: Learning the right metric on the right feature. arXiv preprint arXiv:1505.00855 (2015)" + } + ] + } + ], + "index": 18 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 399, + 91, + 447, + 100 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 399, + 91, + 447, + 100 + ], + "spans": [ + { + "bbox": [ + 399, + 91, + 447, + 100 + ], + "type": "text", + "content": "Pick-a-back" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 470, + 91, + 481, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 470, + 91, + 481, + 100 + ], + "spans": [ + { + "bbox": [ + 470, + 91, + 481, + 100 + ], + "type": "text", + "content": "17" + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 16 + }, + { + "para_blocks": [ + { + "bbox": [ + 132, + 117, + 481, + 567 + ], + "type": "list", + "angle": 0, + "index": 17, + "blocks": [ + { + "bbox": [ + 133, + 117, + 480, + 149 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 117, + 480, + 149 + ], + "spans": [ + { + "bbox": [ + 133, + 117, + 480, + 149 + ], + "type": "text", + "content": "49. Shayan, M., Fung, C., Yoon, C.J., Beschastnikh, I.: Biscotti: A blockchain system for private and secure federated learning. IEEE Transactions on Parallel and Distributed Systems 32(7), 1513-1525 (2020)" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 133, + 150, + 480, + 183 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 150, + 480, + 183 + ], + "spans": [ + { + "bbox": [ + 133, + 150, + 480, + 183 + ], + "type": "text", + "content": "50. Shenaj, D., Toldo, M., Rigon, A., Zanuttigh, P.: Asynchronous federated continual learning. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 5054-5062 (2023)" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 132, + 183, + 480, + 205 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 183, + 480, + 205 + ], + "spans": [ + { + "bbox": [ + 132, + 183, + 480, + 205 + ], + "type": "text", + "content": "51. Simonyan, K., Zisserman, A.: Very deep convolutional networks for large-scale image recognition. arXiv preprint arXiv:1409.1556 (2014)" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 132, + 205, + 481, + 237 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 205, + 481, + 237 + ], + "spans": [ + { + "bbox": [ + 132, + 205, + 481, + 237 + ], + "type": "text", + "content": "52. Stoychev, S., Churamani, N., Gunes, H.: Latent generative replay for resource-efficient continual learning of facial expressions. In: 2023 IEEE 17th International Conference on Automatic Face and Gesture Recognition (FG). pp. 1-8. IEEE (2023)" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 132, + 238, + 480, + 270 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 238, + 480, + 270 + ], + "spans": [ + { + "bbox": [ + 132, + 238, + 480, + 270 + ], + "type": "text", + "content": "53. Sun, W., Yang, Z., Zhang, X., Liu, Y.: Energy-efficient neighbor discovery in mobile ad hoc and wireless sensor networks: A survey. IEEE Communications Surveys & Tutorials 16(3), 1448-1459 (2014)" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 132, + 271, + 480, + 303 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 271, + 480, + 303 + ], + "spans": [ + { + "bbox": [ + 132, + 271, + 480, + 303 + ], + "type": "text", + "content": "54. Tan, M., Le, Q.: Efficientnet: Rethinking model scaling for convolutional neural networks. In: International conference on machine learning. pp. 6105-6114. PMLR (2019)" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 132, + 304, + 480, + 325 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 304, + 480, + 325 + ], + "spans": [ + { + "bbox": [ + 132, + 304, + 480, + 325 + ], + "type": "text", + "content": "55. Wah, C., Branson, S., Welinder, P., Perona, P., Belongie, S.: The caltech-ucsd birds-200-2011 dataset (2011)" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 132, + 326, + 480, + 369 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 326, + 480, + 369 + ], + "spans": [ + { + "bbox": [ + 132, + 326, + 480, + 369 + ], + "type": "text", + "content": "56. Wei, K., Li, J., Ding, M., Ma, C., Yang, H.H., Farokhi, F., Jin, S., Quek, T.Q., Poor, H.V.: Federated learning with differential privacy: Algorithms and performance analysis. IEEE transactions on information forensics and security 15, 3454-3469 (2020)" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 132, + 370, + 480, + 402 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 370, + 480, + 402 + ], + "spans": [ + { + "bbox": [ + 132, + 370, + 480, + 402 + ], + "type": "text", + "content": "57. Yoon, J., Jeong, W., Lee, G., Yang, E., Hwang, S.J.: Federated continual learning with weighted inter-client transfer. In: International Conference on Machine Learning. pp. 12073-12086. PMLR (2021)" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 132, + 403, + 480, + 424 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 403, + 480, + 424 + ], + "spans": [ + { + "bbox": [ + 132, + 403, + 480, + 424 + ], + "type": "text", + "content": "58. Yoon, J., Yang, E., Lee, J., Hwang, S.J.: Lifelong learning with dynamically expandable networks. arXiv preprint arXiv:1708.01547 (2017)" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 132, + 425, + 480, + 445 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 425, + 480, + 445 + ], + "spans": [ + { + "bbox": [ + 132, + 425, + 480, + 445 + ], + "type": "text", + "content": "59. Yu, Q., Liu, Y., Wang, Y., Xu, K., Liu, J.: Multimodal federated learning via contrastive representation ensemble. arXiv preprint arXiv:2302.08888 (2023)" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 132, + 447, + 480, + 479 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 447, + 480, + 479 + ], + "spans": [ + { + "bbox": [ + 132, + 447, + 480, + 479 + ], + "type": "text", + "content": "60. Yurochkin, M., Agarwal, M., Ghosh, S., Greenewald, K., Hoang, N., Khazaeni, Y. Bayesian nonparametric federated learning of neural networks. In: International conference on machine learning. pp. 7252-7261. PMLR (2019)" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 132, + 479, + 480, + 511 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 479, + 480, + 511 + ], + "spans": [ + { + "bbox": [ + 132, + 479, + 480, + 511 + ], + "type": "text", + "content": "61. Zhang, J., Chen, C., Zhuang, W., Lyu, L.: Target: Federated class-continual learning via exemplar-free distillation. In: Proceedings of the IEEE/CVF International Conference on Computer Vision. pp. 4782-4793 (2023)" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 132, + 512, + 480, + 544 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 512, + 480, + 544 + ], + "spans": [ + { + "bbox": [ + 132, + 512, + 480, + 544 + ], + "type": "text", + "content": "62. Zhang, T., Mai, C., Chang, Y., Chen, C., Shu, L., Zheng, Z.: Fedego: privacy-preserving personalized federated graph learning with ego-graphs. ACM Transactions on Knowledge Discovery from Data 18(2), 1-27 (2023)" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 132, + 545, + 480, + 567 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 545, + 480, + 567 + ], + "spans": [ + { + "bbox": [ + 132, + 545, + 480, + 567 + ], + "type": "text", + "content": "63. Zhu, L., Liu, Z., Han, S.: Deep leakage from gradients. Advances in neural information processing systems 32 (2019)" + } + ] + } + ], + "index": 16 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "spans": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "type": "text", + "content": "18" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 166, + 91, + 247, + 100 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 166, + 91, + 247, + 100 + ], + "spans": [ + { + "bbox": [ + 166, + 91, + 247, + 100 + ], + "type": "text", + "content": "J. Yoon and H. Lee" + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 17 + } + ], + "_backend": "vlm", + "_version_name": "2.6.4" +} \ No newline at end of file diff --git a/2024/Pix2Gif_ Motion-Guided Diffusion for GIF Generation/b48548a6-50ce-46d1-9de4-333ad980d37c_content_list.json b/2024/Pix2Gif_ Motion-Guided Diffusion for GIF Generation/b48548a6-50ce-46d1-9de4-333ad980d37c_content_list.json new file mode 100644 index 0000000000000000000000000000000000000000..288e204f03d992dd7e122c53d7f2dd15853efa4f --- /dev/null +++ b/2024/Pix2Gif_ Motion-Guided Diffusion for GIF Generation/b48548a6-50ce-46d1-9de4-333ad980d37c_content_list.json @@ -0,0 +1,1578 @@ +[ + { + "type": "text", + "text": "Pix2Gif: Motion-Guided Diffusion for GIF Generation", + "text_level": 1, + "bbox": [ + 259, + 140, + 743, + 184 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Hitsh Kandala1, Jianfeng Gao1, and Jianwei Yang1", + "bbox": [ + 300, + 213, + 702, + 229 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Microsoft Research \nkhitesh2000@gmail.com,{jfgao, jianwei.yang}@microsoft.com", + "bbox": [ + 282, + 241, + 718, + 268 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/931773cd42f45d093782adbb62d46c47f9065e3b654c6ab8741a95653ba3bc9c.jpg", + "image_caption": [ + "Fig. 1: Our model creates distinct frames based on the provided source image and caption, adjusting according to different levels of motion magnitude (optical flow magnitude) specified in the input conditions. It stands well for both high spatial quality and temporal consistency." + ], + "image_footnote": [], + "bbox": [ + 287, + 313, + 715, + 460 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Abstract. We present *Pix2Gif*, a motion-guided diffusion model for image-to-GIF (video) generation. We tackle this problem differently by formulating the task as an image translation problem steered by text and motion magnitude prompts, as shown in Fig. 1. To ensure that the model adheres to motion guidance, we propose a new motion-guided warping module to spatially transform the features of the source image conditioned on the two types of prompts. Furthermore, we introduce a perceptual loss to ensure the transformed feature map remains within the same space as the target image, ensuring content consistency and coherence. In preparation for the model training, we meticulously curated data by extracting coherent image frames from the TGIF video-caption dataset, which provides rich information about the temporal changes of subjects. After pretraining, we apply our model in a zero-shot manner to a number of video datasets. Extensive qualitative and quantitative experiments demonstrate the effectiveness of our model - it not only captures the semantic prompt from text but also the spatial ones from motion guidance. We train all our models using a single node of $16 \\times \\mathrm{V}100$ GPUs.", + "bbox": [ + 261, + 590, + 743, + 839 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Keywords: Diffusion $\\cdot$ image-to-video generation $\\cdot$ image-to-image translation", + "bbox": [ + 261, + 147, + 743, + 174 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "1 Introduction", + "text_level": 1, + "bbox": [ + 215, + 202, + 375, + 218 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Visual content generation has been significantly advanced by the huge progress of diffusion models [21, 22, 35, 51]. Recently, the development of latent diffusion models (LDMs) [45] has led us to a new quality level of generated images. It has inspired a lot of works for customized and controllable image generation [32, 46, 65, 69], and fine-grained image editing [6, 20, 28, 36].", + "bbox": [ + 212, + 234, + 782, + 310 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "In this work, we focus on converting a single image to an animated Graphics Interchange Format (GIF), which is valuable for design yet under-explored. Despite the absence of image-to-GIF generation models, diffusion-based video generation has emerged as a hot topic recently. Compared with text-to-image generation, however, text-to-video generation requires not only high quality for individual frames but also visual consistency and temporal coherence across frames. To achieve this goal, existing works expand the LDMs to video diffusion models (VDMs) by either inflating the 2D CNNs in LDMs to 3D ones [23] or introducing an additional temporal attention layer to bridge the diffusion for each frame [5,15,19,50,59]. In addition to text prompts, a few recent works also explored the way of using images or other prompts to make the video generation model more customizable and controllable [11,38]. However, due to the high cost of VDMs to generate a sequence of video frames in one run, most (if not all) of these works require a compromise of reducing the resolution of generated frames $(64 \\times 64$ typically), and the usage of extra super-resolution diffusion models for upscaling [31,47]. Moreover, since these methods use the temporal attention layers to model the cross-frame dependency implicitly, it is quite hard for them to preserve good controllability of the frame-to-frame temporal dynamics in a fine-grained manner.", + "bbox": [ + 212, + 311, + 784, + 595 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Given that animated GIF usually contains less number of frames and requires more specializations, we take a different strategy and formulate the image-to-GIF generation as an image translation process. To decouple the generation of visual contents and temporal dynamics, we further introduce a motion flow magnitude as extra guidance in addition to image and text prompts. Unlike the aforementioned works, our model takes one or more history frames as the condition and produces only one future frame at once. This brings some unique advantages: (i) simplicity - our model can be purely built on top of LDMs and trained end-to-end with high resolution, without any cascaded diffusion processes for upscaling. (ii) controllability - we could inject detailed and different text and motion prompts at each time step for generating a frame, which gains much better controllability of the model. Our work is inspired by a line of canonical works for future frame prediction [40, 48, 56]. However, due to the lack of a powerful image-generation engine, these works fail to produce high-quality results and can only be applied to specific video domains [16, 30, 52]. Moreover, they cannot support other types of prompts or conditions than the history frames. To address this problem,", + "bbox": [ + 212, + 598, + 785, + 839 + ], + "page_idx": 1 + }, + { + "type": "page_number", + "text": "2", + "bbox": [ + 217, + 114, + 228, + 126 + ], + "page_idx": 1 + }, + { + "type": "header", + "text": "H. Kandala et al.", + "bbox": [ + 271, + 114, + 387, + 126 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "we exploit a modern diffusion-based pipeline. More specifically, we follow text-conditioned image editing approaches (e.g., InstructPix2Pix [6]), and propose a new temporal image editing to produce future frames given history frames. To train the model, we curate a new training dataset based on TGIF [33] by extracting frames and calculating the magnitude of optical flow between them. We then selected an appropriate range of the optical flow magnitude and sampled frame pairs from each GIF in a manner that ensures diversity. In the end, we train our diffusion model called Pix2Gif, which can generate high-quality animated GIF consisting of multiple frames, given a single image and text and motion magnitude prompts. In summary, our main contributions are:", + "bbox": [ + 212, + 146, + 787, + 297 + ], + "page_idx": 2 + }, + { + "type": "list", + "sub_type": "text", + "list_items": [ + "- We are the first to explore an image-to-image translation formula for generating animated GIFs from an image, guided by a text prompt and motion magnitude.", + "- We propose a flow-based warping module with a perceptual loss in the diffusion process that takes motion magnitude as input and controls the temporal dynamics and consistency between future frames and the initial ones.", + "- We curate a new dataset, comprised of 78,692 short GIF clips for training, and 10,546 for evaluation. The new dataset covers a variety of visual domains.", + "- Quantitative and qualitative results demonstrate the effectiveness of our proposed method for generating visually consistent coherent GIFs from a single image, and it can be generalized to a wide range of visual domains." + ], + "bbox": [ + 225, + 315, + 782, + 483 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "2 Prior Work", + "text_level": 1, + "bbox": [ + 215, + 513, + 362, + 529 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Image and video generation has been a long-standing problem in the community. It can be tackled by different approaches, which can be categorized into four groups: generative adversarial networks (GANs) [17,26,27,44], transformer-based autoregressive decoding [10,12,41,42,60,66], masked image modeling [7,8,55,67]. Most of the recent works exploited diffusion models for image generation given their high-quality outputs and huge open-source supports [45, 47]. Recently, a number of works have extended the text-to-image generation model into image translation or editing models [6,20,28,36,69] or video generation models [5,15, 23,50,57]. Below we provide a brief overview of the related diffusion-based image and video generation methods.", + "bbox": [ + 212, + 551, + 787, + 703 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Image-to-Image Translation. Diffusion-based image-to-image generation has drawn increasing attention. Different from text-to-image generation, it takes an image as input and edits its contents following the text instructions while keeping the unrelated parts unchanged. SDEdit [36] and ILVR [9] are two pioneering works that impose reference image conditions to an existing latent diffusion model for controllable image generation. Later on, to conduct local edits, the authors in [2] proposed blended latent diffusion to steer the diffusion process with a user-specified mask, where the pixels out of the mask remain the same", + "bbox": [ + 212, + 719, + 787, + 839 + ], + "page_idx": 2 + }, + { + "type": "header", + "text": "Pix2Gif: Motion-Guided Diffusion for GIF Generation", + "bbox": [ + 370, + 114, + 730, + 127 + ], + "page_idx": 2 + }, + { + "type": "page_number", + "text": "3", + "bbox": [ + 774, + 116, + 784, + 126 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "as the input image while the region inside is edited following the textual description. Instead of manipulating the image space, Prompt2Prompt [20] proposed to edit the image by manipulating the textual context (e.g., swapping or adding words.) to which the latent diffusion model cross-attends. However, this method requires forwarding a text-to-image generation process to obtain the cross-attention maps, and thus cannot be applied to real images. Imagic [28] proposed to blend the embeddings of a real image with the textual context embedding so that the generated image obeys both the image and text conditions.", + "bbox": [ + 212, + 146, + 787, + 268 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "All the aforementioned works leverage a frozen latent diffusion model and control the generation with modified text or image prompts. To enable arbitrary image editing, InstructPix2Pix [6] proposed to finetune the LDM to follow user instructions that precisely convey the user intents, e.g., \"change the cat to dog\". The model is trained by a synthetic dataset consisting of triplets $\\langle \\text{image}_{\\text{src}}, \\text{instruction}, \\text{image}_{\\text{tgt}} \\rangle$ . The resulting model could allow both realistic and generated images and support arbitrary language instructions. Some other works also exploit a similar way to train the model to follow instructions [18,68]. To further enhance the language understanding, MGIE [13] exploited a large multimodal model to produce a more comprehensive textual context for the instructed image editing.", + "bbox": [ + 212, + 270, + 787, + 436 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "In this work, we employ the image-to-image translation pipeline and are the first to formulate a GIF generation as an image translation problem. Given a reference image, the goal is to generate a realistic future frame following a textual instruction. Therefore, the focus is on how to perform temporal rather than spatial editing on a source image. When the process rolls out, it gradually gives a sequence of frames.", + "bbox": [ + 212, + 440, + 787, + 531 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Conditioned Video Generation. Speaking of the high-level goal, our work resembles conditioned video generation. For video generation, a conventional way is inflating the 2D U-Net used in LDM to 3D U-Net [71] by replacing the 2D convolution layers with 3D ones. Likewise, a similar strategy is taken in [5,15,19,50,59], but with a slight difference in that they use interleaved spatial and temporal attention layers in the U-Net. Due to the high cost of generating a sequence of video frames in one shot, the output videos usually have a resolution as low as $64 \\times 64$ . To attain high-resolution videos, these methods need to use one or more super-resolution diffusion models [31,47] to upscale the resolution by 4 or 8 times. To accelerate the training, a pre-trained text-to-image LDM is usually used to initialize these models. Adding spatial-temporal modules is also a commonly used strategy for autoregressive models [24, 55, 61, 62]. Similarly, both [62] and [24] exploit a pretrained autoregressive image generation model as the starting point. In [55], however, the authors introduce and pretrained a new video encoder, which is then used to train a masked video decoder.", + "bbox": [ + 212, + 549, + 787, + 776 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Besides text-to-video generation, using images as the condition for video generation draws increasing attention. On one hand, once a text-to-video generation is trained, it can be further finetuned for image-to-video generation [4]. In [11], the authors introduce additional structural conditions (e.g., depth maps) for", + "bbox": [ + 212, + 779, + 787, + 840 + ], + "page_idx": 3 + }, + { + "type": "page_number", + "text": "4", + "bbox": [ + 217, + 114, + 228, + 126 + ], + "page_idx": 3 + }, + { + "type": "header", + "text": "H. Kandala et al.", + "bbox": [ + 271, + 114, + 387, + 127 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "more controllable video generation. Alternatively in [38], a latent flow diffusion model is introduced for image-conditioned video generation by explicitly generating a sequence of optimal flows and masks as the guidance. On the other hand, a few concurrent works to ours directly approach image-to-video generation on top of video diffusion models [58, 63, 70]. All these works share a similar spirit to text-to-video generation models but add additional images as the reference.", + "bbox": [ + 212, + 146, + 782, + 236 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Our method uses a diffusion model but differs from all the aforementioned methods in that we reformulate video generation as a frame-to-frame translation problem based on the history frames. As [6] suggests image-to-image translation can maintain a decent visual consistency. In addition, we also introduce a motion flow magnitude as another condition to explicitly control the temporal dynamics.", + "bbox": [ + 212, + 239, + 784, + 315 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Future Frame Prediction. Future frame prediction or forecasting [40, 48, 56] has been a long-standing problem before the prevalence of diffusion models. It has been used as an anomaly detection approach by comparing the observed frame and the predicted ones [3, 34] and video representation learning for various downstream tasks [14]. For these problems, a recurrent network such as LSTM [53], ConvLSTM [39, 56] or 3D-CNN [1] is usually used as the model architecture, and GAN [17] or Variational Autoencoder (VAE) [29] is used as the learning objectives. With the emergence of VQ-VAE [43], the authors in [25] exploited axial transformer blocks to chain the encoder-decoder for autoregressive next-frame prediction. In [57], the authors proposed masked conditional video diffusion to unify different tasks of video prediction, generation and interpolation. Nevertheless, all of these models are trained on domain-specific video datasets such as MovingMNIST [30], CATER [16] and UCF-101 [52], etc, far from being a generic video generation model.", + "bbox": [ + 212, + 333, + 785, + 544 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Our work takes inspiration from future frame prediction methods but proposes a simpler yet effective strategy by formulating it as an image-to-image translation problem. Furthermore, our model simultaneously takes image, text and motion magnitude as the guidance for better controllability. To attain a model as general as possible, we curate a new training dataset covering a wide range of domains. Without any further dataset-specific finetuning, our model achieves plausible video generation results as shown in Fig. 1.", + "bbox": [ + 212, + 547, + 787, + 652 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "3 Method", + "text_level": 1, + "bbox": [ + 215, + 689, + 328, + 704 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Our goal is to generate GIFs, given an initial frame, a motion description, and a measure of optical flow. We frame this as an image-to-image translation problem based on latent diffusion. We first explain how we created our training dataset in Sec. 3.1, then outline our model's principles and training strategy in Sec. 3.2. Next, we delve into the specifics of our proposed model, explaining its various components in Sec. 3.3. Finally, we focus on the loss functions used to train our model in Sec. 3.4.", + "bbox": [ + 212, + 734, + 787, + 839 + ], + "page_idx": 4 + }, + { + "type": "header", + "text": "Pix2Gif: Motion-Guided Diffusion for GIF Generation", + "bbox": [ + 370, + 114, + 730, + 128 + ], + "page_idx": 4 + }, + { + "type": "page_number", + "text": "5", + "bbox": [ + 774, + 116, + 784, + 126 + ], + "page_idx": 4 + }, + { + "type": "image", + "img_path": "images/8d0c160964518d0d443d4947b2090224c98840295ce1819c05b58b0e6d0e1ce5.jpg", + "image_caption": [ + "Fig. 2: The three step process of curating the TGIF dataset. Starting from extracting frames $A$ to restricting the range of optical flow $B$ and then maintaining the diversity of pairs $C$ ." + ], + "image_footnote": [], + "bbox": [ + 287, + 143, + 715, + 268 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/ea8b16389739e85bdecfa9ac8571d8be904b50cfadaf607d471bc3376e4ff0c5.jpg", + "image_caption": [ + "Fig.3: $\\text{Pix2Gif}$ model pipeline. We propose an end-to-end network where the inputs are encoded by $\\mathcal{E}$ , CLIP and $\\mathcal{M}$ to output $\\mathcal{E}(c_L), c_T$ and $c_M$ respectively, which then goes into $\\mathcal{W}$ to form the conditioning input for LDM." + ], + "image_footnote": [], + "bbox": [ + 287, + 338, + 718, + 481 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "3.1 Dataset", + "text_level": 1, + "bbox": [ + 215, + 566, + 326, + 579 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "We used the Tumblr GIF (TGIF) dataset [33], which predominantly consists of human-centric animated GIFs described by captions. The dataset features a range of GIFs with varied movements over 1-3 seconds.", + "bbox": [ + 212, + 595, + 787, + 641 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "The curation process, as shown in Fig. 2, involved extracting frames from all GIFs and calculating the optical flow between all possible frame pairs. The number of extracted frames from each GIF varied, with an average of about 41 frames. The optical flow histogram calculated between all frames ranged from [0, 200]. We selected the [2, 20] range, capturing small yet significant motion and excluding pairs with drastic changes. Despite the restricted range, we still had a significant number of training pairs. To avoid model overfitting and maintain diversity, we randomly selected a minimum of 10 pairs or the number of pairs within the restricted range from each GIF. This approach resulted in a final dataset with nearly equal representation of all values within the selected range. The final dataset contained 783,184 training pairs and 105,041 validation pairs. Each data point consisted of a pair of frames from the same GIF, the corresponding caption, and the calculated optical flow between the frames.", + "bbox": [ + 212, + 643, + 787, + 840 + ], + "page_idx": 5 + }, + { + "type": "page_number", + "text": "6", + "bbox": [ + 217, + 114, + 228, + 126 + ], + "page_idx": 5 + }, + { + "type": "header", + "text": "H. Kandala et al.", + "bbox": [ + 271, + 114, + 387, + 126 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "3.2 Preliminary: Instructed Image Editing", + "text_level": 1, + "bbox": [ + 215, + 146, + 578, + 162 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Our model is fundamentally grounded in the latent diffusion models (LDMs) for image generation and editing [6, 45]. More specifically, we build upon InstructPix2Pix [6] by framing our objective in the context of an instructed image-to-image translation task. Given an image $x$ , the forward diffusion procedure introduces noise to the encoded latent $z$ , thereby producing a noisy latent vector $z_{t}$ . This process is carried out over $T$ timesteps, with each timestep $t \\in \\{1, \\dots, T\\}$ seeing an increment in the noise level until it culminates into a random noise $n$ . A network $e_{\\theta}$ is trained by minimizing the following latent diffusion objective to predict noise existing in the noisy latent $z_{t}$ , considering factors in image conditioning $c_{I}$ and textual instruction $c_{T}$ :", + "bbox": [ + 212, + 178, + 787, + 330 + ], + "page_idx": 6 + }, + { + "type": "equation", + "text": "\n$$\nL _ {L D M} = \\mathbb {E} _ {\\mathcal {E} (x), \\mathcal {E} (c _ {I}), c _ {T}, \\epsilon \\sim \\mathcal {N} (0, 1), t} \\left[ | | \\epsilon - \\epsilon_ {\\theta} \\left(z _ {t}, t, \\mathcal {E} \\left(c _ {I}\\right), c _ {T}\\right) | | _ {2} ^ {2} \\right] \\tag {1}\n$$\n", + "text_format": "latex", + "bbox": [ + 294, + 343, + 785, + 362 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "where $\\mathcal{E}$ is the VQ-VAE encoder that transforms the images from pixel space to discrete latent space. To facilitate image conditioning, $z_{t}$ and $\\mathcal{E}(c_I)$ are concatenated and then fed into a convolutional layer. The model is trained for conditional and unconditional denoising, given the image and caption condition individually or collectively.", + "bbox": [ + 212, + 375, + 787, + 450 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "3.3 Our Model: Pix2Gif", + "text_level": 1, + "bbox": [ + 215, + 478, + 429, + 493 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "We build our model similar to InstructPix2Pix and frame our objective in the context of a text-instructed and motion-guided temporal editing problem. Compared with the original InstructPix2Pix pipeline, the main innovation is the newly introduced motion-based warping module. The overall model pipeline is shown in Fig. 3.", + "bbox": [ + 212, + 510, + 785, + 585 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Our model takes three inputs: an image, a text instruction, and a motion magnitude. These inputs are fed into the model through two pathways - once through the diffusion model directly, and again through the warping module, which will be discussed in Sec. 3.3 and Sec. 3.3. When passed through the caption, we add the phrase \"The optical flow is _. to the original caption. The flow input is then appended at the end in a word format rather than a numerical one, as the CLIP model tends to assign higher similarity scores to word forms than to numerical representations of numbers for the same image. Finally, our model is trained by minimizing the following loss function:", + "bbox": [ + 212, + 587, + 787, + 723 + ], + "page_idx": 6 + }, + { + "type": "equation", + "text": "\n$$\nL _ {L D M} ^ {\\prime} = \\mathbb {E} _ {\\mathcal {E} (x), \\mathcal {E} (c _ {I}), c _ {T}, c _ {M}, \\epsilon \\sim \\mathcal {N} (0, 1), t} \\left[ | | \\epsilon - \\epsilon_ {\\theta} (z _ {t}, t, \\mathcal {E} (c _ {I}), c _ {T}, c _ {M}) | | _ {2} ^ {2} \\right] \\tag {2}\n$$\n", + "text_format": "latex", + "bbox": [ + 269, + 738, + 785, + 756 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "where $c_{M}$ is the motion condition. The altered caption is processed via the pretrained CLIP model to yield $c_{T}$ , while the output of $\\mathcal{M}$ gives us $c_{M}$ . These two conditions are then added linearly, serving as the conditioning input for both the Warping Module $\\mathcal{W}$ (discussed in Sec. 3.3) and the Latent Diffusion Model LDM (referenced in Sec. 3.2).", + "bbox": [ + 212, + 768, + 787, + 844 + ], + "page_idx": 6 + }, + { + "type": "header", + "text": "Pix2Gif: Motion-Guided Diffusion for GIF Generation", + "bbox": [ + 370, + 114, + 730, + 127 + ], + "page_idx": 6 + }, + { + "type": "page_number", + "text": "7", + "bbox": [ + 774, + 116, + 784, + 126 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/d5aa299c7c9f03e600a0952a7ba0d5b45d37c8c7c14f76f9a2734e53197ddd82.jpg", + "image_caption": [ + "Fig. 4: Deep dive into the Warping Module $\\mathcal{W}$ . It comprises of three units: $\\mathcal{I}_M$ , $\\mathcal{F}_{Net}$ and $\\mathcal{W}_{Net}$ ." + ], + "image_footnote": [], + "bbox": [ + 303, + 145, + 702, + 311 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Motion Embedding Layer In traditional conditional diffusion models like [6], text prompts are often enough for image generation. We initially passed the motion input indirectly through the prompt. However, this divided the model's attention on a single caption token, which is problematic when the main caption stays the same but the motion input changes. To let the model focus on the motion input, we included an embedding layer that converts the motion input into an integer and selects an embedding vector. This vector is then duplicated and concatenated with itself to generate $c_{M}$ , which when combined with the caption embedding $c_{T}$ , provides the conditioning input $c_{L} = c_{T} + c_{M}$ for both warping module $\\mathcal{W}$ and LDM.", + "bbox": [ + 212, + 388, + 787, + 541 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Warping Module One of the main components of Pix2Gif is the Warping Module $\\mathcal{W}$ . As illustrated in Fig. 4, it technically comprises two networks: the FlowNet $(\\mathcal{F}_{Net})$ and the WarpNet $(\\mathcal{W}_{Net})$ . Ordinarily, the computation of optical flow involves two images. However, in this case, we initially have only one image - the source image - and that too in the latent domain. Thus, our goal is to learn the optical flow utilizing just one latent image. This is achieved via $\\mathcal{F}_{Net}$ , conditioned on $c_{L}$ , which guides it to generate a flow feature map in the intended direction with the hint of text and motion prompts. This condition is processed by the Injection Module $(\\mathcal{I}_{\\mathcal{M}})$ , a compact encoder designed to make $c_{L}$ compatible for concatenation with one of the intermediate feature maps near the end of the network. This configuration enables $\\mathcal{F}_{Net}$ to independently learn high-level features, which are then guided in the desired direction with the introduction of $c_{L}$ . The architecture of $\\mathcal{F}_{Net}$ resembles that of UNet, producing an output with a fixed channel of 2 to capture changes in the horizontal and vertical components. This optical flow feature map $F = \\mathcal{F}_{Net}(\\mathcal{E}(c_{I}), \\mathcal{I}_{\\mathcal{M}}(c_{L}))$ , along with the source latent $(\\mathcal{E}(c_{I}))$ , is then processed through $\\mathcal{W}_{Net}$ to yield a Fischer map $z_{W} = \\mathcal{W}_{Net}(\\mathcal{E}(c_{I}), F)$ . This transformation is learned more efficiently and abstractly in the latent space than in the pixel space.", + "bbox": [ + 212, + 566, + 787, + 843 + ], + "page_idx": 7 + }, + { + "type": "page_number", + "text": "8", + "bbox": [ + 217, + 114, + 228, + 126 + ], + "page_idx": 7 + }, + { + "type": "header", + "text": "H. Kandala et al.", + "bbox": [ + 271, + 114, + 387, + 126 + ], + "page_idx": 7 + }, + { + "type": "table", + "img_path": "images/0fd4bb3a5606307054604849bb0be542a56b53e6c09a2fe793d7e1c5b7838c89.jpg", + "table_caption": [ + "Table 1: Quantitative comparison with state-of-the-art image-text-to-video generation models for the zero-shot setting." + ], + "table_footnote": [], + "table_body": "
MethodUCF-101 [52]MSR-VTT [64]
FVD ↓CLIPSim ↑PIC ↑FVD ↓CLIPSim ↑PIC ↑
I2VGen-XL [70]563.120.28650.6329278.620.22720.6018
DynamiCrafter [63]527.060.27960.6307271.630.26020.6135
Pix2Gif (Ours)285.020.28150.8763168.690.25730.8521
", + "bbox": [ + 277, + 184, + 723, + 265 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "3.4 Losses", + "text_level": 1, + "bbox": [ + 215, + 301, + 313, + 315 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "Our model incorporates two different types of losses. The first type is the standard L2 loss Eq. (2), which is utilized by the stable diffusion model and talked about in Sec. 3.2", + "bbox": [ + 212, + 325, + 782, + 369 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "The second loss type in our model is the perceptual loss, which is calculated by comparing the latent features of the image condition $\\mathcal{E}(c_I)$ and the warped image $z_W$ . This is implemented using a pre-trained VGG network [49], modified to accommodate 4 channels instead of the standard 3. The modification involves averaging the weights from the first three channels to initialize the fourth. The perceptual loss, $L_{p}$ , is defined as:", + "bbox": [ + 212, + 371, + 784, + 460 + ], + "page_idx": 8 + }, + { + "type": "equation", + "text": "\n$$\nL _ {p} \\left(\\mathcal {E} \\left(c _ {I}\\right), z _ {W}\\right) = \\sum_ {k} \\lambda_ {k} \\left\\| \\phi_ {k} \\left(\\mathcal {E} \\left(c _ {I}\\right)\\right) - \\phi_ {k} \\left(z _ {W}\\right) \\right\\| ^ {2} \\tag {3}\n$$\n", + "text_format": "latex", + "bbox": [ + 334, + 473, + 784, + 503 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "Here, $\\phi_k(.)$ be the feature map of the k-th layer of the VGG network, $||.||$ denotes the Frobenius norm, and $\\lambda_{k}$ is a weighting factor. This loss ensures that the warped image maintains high-level features like edges, textures, and object types, making the images more perceptually and semantically similar and preserving the overall structure of the source image.", + "bbox": [ + 212, + 513, + 784, + 589 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "In conclusion, the total loss function, denoted as $L_{T}$ , for our objective is computed by a weighted sum of the two individual losses.", + "bbox": [ + 212, + 590, + 784, + 619 + ], + "page_idx": 8 + }, + { + "type": "equation", + "text": "\n$$\nL _ {T} = L _ {L D M} ^ {\\prime} + \\lambda_ {P} L _ {P} \\tag {4}\n$$\n", + "text_format": "latex", + "bbox": [ + 424, + 635, + 784, + 652 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "Here, $\\lambda_{P}$ is the weighting factor for perceptual loss. These two losses together provide a holistic framework to train our model by ensuring pixel-level accuracy, preservation of high-level features, and smooth motion transitions.", + "bbox": [ + 212, + 657, + 782, + 702 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "4 Experiments", + "text_level": 1, + "bbox": [ + 215, + 724, + 375, + 742 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "4.1 Setup", + "text_level": 1, + "bbox": [ + 215, + 755, + 310, + 770 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "Datasets We utilize the Tumbled GIF (TGIF) dataset for our training and validation purposes as discussed in Sec. 3.1. We evaluate our model on two datasets: MSR-VTT [64] and UCF-101 [52], following the common practice. For these datasets, we follow the sampling strategy as outlined in [63].", + "bbox": [ + 212, + 779, + 784, + 840 + ], + "page_idx": 8 + }, + { + "type": "header", + "text": "Pix2Gif: Motion-Guided Diffusion for GIF Generation", + "bbox": [ + 370, + 114, + 730, + 127 + ], + "page_idx": 8 + }, + { + "type": "page_number", + "text": "9", + "bbox": [ + 774, + 116, + 784, + 126 + ], + "page_idx": 8 + }, + { + "type": "image", + "img_path": "images/55bcb091bbb48343b53c9a576aa3723491ca2d0e3e4ee295e700cb9bbe4c368e.jpg", + "image_caption": [ + "(a)" + ], + "image_footnote": [], + "bbox": [ + 289, + 143, + 715, + 304 + ], + "page_idx": 9 + }, + { + "type": "image", + "img_path": "images/efed11d51d9f0a373847c70abec6cd96a80958ee630596d275f4c359dff3a39e.jpg", + "image_caption": [ + "(b)", + "Fig. 5: Comparison studies with other image-text to video models. Given a source image and a caption, frames are extracted from the generated 16-frame video at 256x256 resolution." + ], + "image_footnote": [], + "bbox": [ + 303, + 318, + 702, + 486 + ], + "page_idx": 9 + }, + { + "type": "text", + "text": "**Implementations** Our model is initialized with the exponential moving average (EMA) weights of the Stable Diffusion v1.5 checkpoint and the improved ftMSE autoencoder weights $^2$ . We trained the model at 256x256 resolution for 7 epochs on a single node of 16 V100 GPUs for 25k steps. We used the AdamW optimizer with a learning rate of $10^{-4}$ . We set the weighting factor for perceptual loss $(\\lambda_P)$ as $10^{-2}$ .", + "bbox": [ + 212, + 584, + 787, + 676 + ], + "page_idx": 9 + }, + { + "type": "text", + "text": "Metrics We report Frechet Video Distance (FVD) [54], CLIP Similarity (CLIP-Sim) which is the average similarity calculated for all the generated frames with the input caption and Perceptual Input Conformity (PIC) as described in [63] for all methods. For comparison, we assess the zero-shot generation performance on I2VGen-XL [70] and DynamiCrafter [63].", + "bbox": [ + 212, + 696, + 787, + 776 + ], + "page_idx": 9 + }, + { + "type": "page_number", + "text": "10", + "bbox": [ + 217, + 114, + 235, + 126 + ], + "page_idx": 9 + }, + { + "type": "header", + "text": "H. Kandala et al.", + "bbox": [ + 271, + 114, + 387, + 126 + ], + "page_idx": 9 + }, + { + "type": "page_footnote", + "text": "1 https://huggingface.co/runwayml/stable-diffusion-v1-5/blob/main/v1-5-pruned-emaonly.ckpt", + "bbox": [ + 217, + 782, + 764, + 811 + ], + "page_idx": 9 + }, + { + "type": "page_footnote", + "text": "$^{2}$ https://huggingface.co/stabilityai/sd-vae-ft-mse-original/blob/main/vae-ft-mse-840000-ema-pruned.ckpt", + "bbox": [ + 218, + 811, + 761, + 839 + ], + "page_idx": 9 + }, + { + "type": "image", + "img_path": "images/a6d06ceb5f7ff589d8201b917c88fd5558c9f8e007881344ecd2ad3cc1365967.jpg", + "image_caption": [ + "(a) A1: Cat is playing with wool", + "Fig. 6: Pix2Gif showing composition capabilities for different types of motions. [GIFs best viewed in Adobe Acrobat Reader]" + ], + "image_footnote": [], + "bbox": [ + 225, + 271, + 400, + 311 + ], + "page_idx": 10 + }, + { + "type": "image", + "img_path": "images/6d1df9846d32c3c417e770db8893e2c700ccac1972cad12bd3931a448c233f9b.jpg", + "image_caption": [ + "(b) A2: Cat is dancing" + ], + "image_footnote": [], + "bbox": [ + 413, + 271, + 588, + 311 + ], + "page_idx": 10 + }, + { + "type": "image", + "img_path": "images/6521da9d68a5bcc4e12d333381ba148d366c1bb0578f7e2a2db390f83c915c1b.jpg", + "image_caption": [ + "(c) A1+2: Cat dancing and playing with wool" + ], + "image_footnote": [], + "bbox": [ + 602, + 271, + 777, + 311 + ], + "page_idx": 10 + }, + { + "type": "table", + "img_path": "images/fa3da07e371c75d1c5989644694901d6ebc782a3e98b4674677da903453eaea1.jpg", + "table_caption": [ + "Table 2: Ablation study comparing image translation methods with a focus on motion coherency at varyingcfg_img values." + ], + "table_footnote": [], + "table_body": "
Method /cfgimg1.41.61.82.0
L2 ↓PCC ↑L2 ↓PCC ↑L2 ↓PCC ↑L2 ↓PCC ↑
InstructPix2Pix [6]23.429-0.22925.492-0.02827.037-0.42327.5300.139
Pix2Gif-Base7.5800.9895.1880.9875.5950.9927.0290.991
Pix2Gif1.7460.9951.9720.9952.9440.9974.0760.997
", + "bbox": [ + 264, + 406, + 738, + 486 + ], + "page_idx": 10 + }, + { + "type": "text", + "text": "4.2 Results", + "text_level": 1, + "bbox": [ + 215, + 527, + 323, + 540 + ], + "page_idx": 10 + }, + { + "type": "text", + "text": "Comparisons with previous works Fig. 5 and Tab. 1 provide a qualitative and quantitative comparison of three image-text to video models: I2VGenXL [70], DynamiCrafter [63], and our $Pix2Gif$ .", + "bbox": [ + 212, + 551, + 784, + 598 + ], + "page_idx": 10 + }, + { + "type": "text", + "text": "In Fig. 5a, the I2VGen-XL model misshapes the dog's face and generates it sideways in a nonsensical manner. DynamiCrafter appears to disregard the input parameters, as the initial frame differs significantly in position, color, and texture. It is also challenging to discern whether the dog is eating or merely moving its mouth. Our model, Pix2Gif, accurately retains all the dog's details and successfully depicts it eating from a plate. In Fig. 5b, we assess the models' capabilities by generating a video from a relatively dark image. Once again, I2VGen-XL starts strong, producing some impressive frames, but these soon turn into highly stylized and improbable images. DynamiCrafter appears to misinterpret the input image, generating something significantly different, although it seems to adhere to the caption. Conversely, Pix2Gif comprehends the inputs effectively and produces corresponding motion while preserving the overall integrity of the source image.", + "bbox": [ + 212, + 598, + 785, + 794 + ], + "page_idx": 10 + }, + { + "type": "text", + "text": "Quantitatively, $\\text{Pix2Gif}$ excels in both the FVD and PIC metrics shown in Tab. 1, which aligns with our observations of the frames generated in Fig. 5. These frames effectively preserve the structure and closely adhere to the input", + "bbox": [ + 212, + 794, + 785, + 839 + ], + "page_idx": 10 + }, + { + "type": "header", + "text": "Pix2Gif: Motion-Guided Diffusion for GIF Generation", + "bbox": [ + 370, + 114, + 730, + 127 + ], + "page_idx": 10 + }, + { + "type": "page_number", + "text": "11", + "bbox": [ + 767, + 114, + 782, + 126 + ], + "page_idx": 10 + }, + { + "type": "image", + "img_path": "images/20380b7b9b5f5c420b65199024486c10deadb7700bbedbc0f1e37c1fda8ee609.jpg", + "image_caption": [ + "(a) source frame score" + ], + "image_footnote": [], + "bbox": [ + 267, + 150, + 442, + 255 + ], + "page_idx": 11 + }, + { + "type": "image", + "img_path": "images/a809d8a20838bf48e9a5d83ec4872530839b98b33196e04079032818808d3ed3.jpg", + "image_caption": [ + "(b) target frame score", + "Fig. 7: Ablation study between the earlier variants of our model by comparing average similarity score for 100 samples." + ], + "image_footnote": [], + "bbox": [ + 549, + 150, + 725, + 255 + ], + "page_idx": 11 + }, + { + "type": "text", + "text": "prompts (source image and caption). However, Pix2Gif does not perform as well in the CLIPSim metric, despite accurately following the caption. The other two models as seen in Fig. 5 do follow the caption, but they fail to adhere to the input image and produce plausible temporal transitions. This is partially attributed to the inherent model design in these two methods. Both methods attempt to generate a full sequence of frames at once using the 3D diffusion network, which inevitably compounds the spatial and temporal dimensions. Moreover, the results indicate that they function more as text-to-video models than image-text-to-video models, especially DynamiCrafter. This discrepancy also raises questions about the effectiveness of the CLIPSim metric for evaluating image-text-to-video models and calls for more sophisticated metrics for evaluating video generation.", + "bbox": [ + 212, + 343, + 787, + 510 + ], + "page_idx": 11 + }, + { + "type": "text", + "text": "Compositionality of actions Fig. 6 illustrates an intriguing emerging capability of $\\text{Pix2Gif}$ : the ability to combine actions. In Fig. 6a, we see a cat playing with wool, with only the cat's paws and the wool moving. In Fig. 6b, we instruct the cat to dance, resulting in the cat moving its body but the wool remaining still. Finally, in Fig. 6c, we provide a caption that blends the actions from Fig. 6a and Fig. 6b. The result is a scene where the cat is both moving the wool and its body. This demonstrates $\\text{Pix2Gif}$ 's ability to comprehend the caption and its associated motion, and to convert that understanding into a GIF. Such compositional capability significantly increases user controllability, a crucial aspect for practical applications.", + "bbox": [ + 212, + 523, + 787, + 676 + ], + "page_idx": 11 + }, + { + "type": "text", + "text": "4.3 Ablations", + "text_level": 1, + "bbox": [ + 215, + 698, + 341, + 712 + ], + "page_idx": 11 + }, + { + "type": "text", + "text": "We design a few variants of $\\text{Pix2Gif}$ for our ablation studies:", + "bbox": [ + 215, + 724, + 656, + 739 + ], + "page_idx": 11 + }, + { + "type": "list", + "sub_type": "text", + "list_items": [ + "- Pix2Gif-Base: We train InstructPix2Pix with our data, and append the text prompt with \"The optical flow is \".", + "- Pix2Gif-Motion-embed: The motion embedding layer is added to encode the motion magnitude and combined with textual embedding.", + "- Pix2Gif-Warp: We further add the warping module into the model but differently only use the warped feature for the LDM." + ], + "bbox": [ + 223, + 748, + 784, + 838 + ], + "page_idx": 11 + }, + { + "type": "page_number", + "text": "12", + "bbox": [ + 217, + 114, + 235, + 126 + ], + "page_idx": 11 + }, + { + "type": "header", + "text": "H. Kandala et al.", + "bbox": [ + 271, + 114, + 387, + 126 + ], + "page_idx": 11 + }, + { + "type": "image", + "img_path": "images/5649367366230c207fa30b68016064fa7f2d2b333cfb6e6a9fc335676ce9c9e8.jpg", + "image_caption": [ + "(a) source frame score" + ], + "image_footnote": [], + "bbox": [ + 269, + 150, + 442, + 255 + ], + "page_idx": 12 + }, + { + "type": "image", + "img_path": "images/9283edcfb9834d1b90281ccb209697d04fd0c3a04a613e2ff8f8034374c42533.jpg", + "image_caption": [ + "(b) target frame score", + "Fig. 8: Ablation study on ways to input $z_{W}$ to LDM by comparing average similarity score for 100 samples." + ], + "image_footnote": [], + "bbox": [ + 552, + 151, + 725, + 255 + ], + "page_idx": 12 + }, + { + "type": "list", + "sub_type": "text", + "list_items": [ + "- Pix2Gif-Warp-add: Different from Pix2Gif-Warp, we instead add the warped feature and source image feature as input to the LDM.", + "- Pix2Gif-Warp-concat: Instead of adding in Pix2Gif-Warp-add, we concatenate the warped feature and source image feature as the input, but do not include the perceptual loss." + ], + "bbox": [ + 223, + 344, + 784, + 420 + ], + "page_idx": 12 + }, + { + "type": "text", + "text": "For comparative studies with our model's variants, we generate an 8-frame video. We use the X-CLIP model [37] to extract features from our generated video, and CLIP to extract features from the source, target, and generated frames. The optimal range of CFG.img for best results is considered to be [1.6, 2.2]. Throughout this discussion, we evaluate our model's performance using four metrics, which we believe effectively measure the different aspects of generating motion through the image translation framework used in Pix2Gif.", + "bbox": [ + 212, + 431, + 787, + 537 + ], + "page_idx": 12 + }, + { + "type": "text", + "text": "Motion Coherency Our task is framed as an image translation problem with motion magnitude as a guide. We assess the motion quality or temporal coherence in the GIFs using L2 loss and Pearson Correlation Coefficient (PCC) for InstructPix2Pix, Pix2Gif-Base, and Pix2Gif. The L2 loss evaluates the match between the motion values of the generated frames and the actual inputs, while the PCC checks if they follow the same trend. These metrics are calculated between the input motion magnitude values and the optical flow values, derived from the source image and the generated frames. Our model exhibits the highest correlation and the lowest L2 loss across all CFG.img values as seen in Tab. 2, proving its efficacy and controllability in generating GIFs with specific motions. The Pix2Gif-Base outperforms the original InstructPix2Pix, emphasizing the importance of our new dataset.", + "bbox": [ + 212, + 551, + 787, + 733 + ], + "page_idx": 12 + }, + { + "type": "text", + "text": "Image-Video Similarity Score To evaluate the semantic properties of the video produced by our model, we created two similarity scores: a source frame score and a target frame score. The source frame score quantifies how well the video retains the primary attributes of the source frame, essentially measuring the accuracy of the source image portrayal throughout the video. The target frame score indicates the precision of the scene or subject development from", + "bbox": [ + 212, + 750, + 787, + 840 + ], + "page_idx": 12 + }, + { + "type": "header", + "text": "Pix2Gif: Motion-Guided Diffusion for GIF Generation", + "bbox": [ + 370, + 114, + 730, + 127 + ], + "page_idx": 12 + }, + { + "type": "page_number", + "text": "13", + "bbox": [ + 767, + 114, + 784, + 126 + ], + "page_idx": 12 + }, + { + "type": "text", + "text": "the source frame in the video. It also underscores the model's ability to handle uncertainty and potential changes, as the target frame represents a possible state that the video might reach.", + "bbox": [ + 212, + 146, + 782, + 191 + ], + "page_idx": 13 + }, + { + "type": "text", + "text": "We began our experiments with $\\text{Pix2Gif-Base}$ and then enhanced it incrementally, first with $\\mathcal{M}$ (Pix2Gif-Motion-embed), then by adding $\\mathcal{W}$ (Pix2Gif-Warp). Here, we conducted an ablation study shown in Fig. 8 to understand the best way to feed $z_{W}$ into LDM. Our experiments included feeding only $z_{W}$ (Pix2Gif-Warp), adding it with the $\\mathcal{E}(c_{I})$ (Pix2Gif-Warp-add), and concatenating both (Pix2Gif-Warp-concat). Both Pix2Gif-Warp-add and Pix2Gif-Warp-concat had higher average similarity scores than Pix2Gif-Warp, and Pix2Gif-Warp-concat performed the best. This can be attributed to the fact that in the addition process, $\\mathcal{E}(c_{I})$ loses its unique characteristics, which are required by the diffusion model for effective unconditional denoising. Therefore, to achieve the best results, we combined $\\mathcal{E}(c_{I})$ and $z_{W}$ before inputting them into the concat attention layer of the LDM. Now finally we integrate $L_{P}$ (Pix2Gif) and compare them in Fig. 7. As expected, Pix2Gif-Base was outperformed by the other versions. Pix2Gif generated more coherent and controlled motion, albeit often with limited extent. All models began to converge outside the optimal range, producing similar frames. The Pix2Gif-Motion-embed model creates a significant amount of one-directional motion, which can sometimes be nonsensical, and hence the addition of $\\mathcal{W}$ helps to mitigate this issue.", + "bbox": [ + 212, + 191, + 787, + 465 + ], + "page_idx": 13 + }, + { + "type": "text", + "text": "5 Limitations and Future Work", + "text_level": 1, + "bbox": [ + 215, + 484, + 535, + 501 + ], + "page_idx": 13 + }, + { + "type": "text", + "text": "The current Pix2Gif model is our initial attempt to generate videos by treating it as an image translation task. However, this method has some limitations that prevent us from generating high-quality and long GIFs or videos. Firstly, the model generates images with a resolution of 256x256 pixels. If these images are used to generate subsequent frames, the quality of the frames deteriorates further. Secondly, due to limitations in computational power, we are only able to use a small portion of a larger, curated dataset for training our model. Our primary objective now is to improve the quality of the generated frames, as this could significantly enhance the effectiveness of this method.", + "bbox": [ + 212, + 515, + 787, + 652 + ], + "page_idx": 13 + }, + { + "type": "text", + "text": "6 Conclusion", + "text_level": 1, + "bbox": [ + 215, + 672, + 359, + 690 + ], + "page_idx": 13 + }, + { + "type": "text", + "text": "In this work, we proposed Pix2Gif, an image-to-GIF (video) generation model based on an image-to-image translation paradigm. To ensure temporal coherence across frames, we proposed a motion-guided warping module that learns to spatially warp the source image feature into the target one while maintaining visual consistency via a perceptual loss. Starting from TGIF, we curated a new dataset specifically used for training our model. The experimental results demonstrated the effectiveness of our model to generate GIFs with better temporal coherence compared with current state-of-the-art methods. Interestingly, the model also exhibits better controllability and some emerging action compositionality.", + "bbox": [ + 212, + 703, + 787, + 840 + ], + "page_idx": 13 + }, + { + "type": "page_number", + "text": "14", + "bbox": [ + 217, + 114, + 235, + 126 + ], + "page_idx": 13 + }, + { + "type": "header", + "text": "H. Kandala et al.", + "bbox": [ + 271, + 114, + 387, + 126 + ], + "page_idx": 13 + }, + { + "type": "text", + "text": "References", + "text_level": 1, + "bbox": [ + 217, + 143, + 321, + 159 + ], + "page_idx": 14 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "1. Aigner, S., Körner, M.: Futuregan: Anticipating the future frames of video sequences using spatio-temporal 3d convolutions in progressively growing gans (2018)", + "2. Avrahami, O., Fried, O., Lischinski, D.: Blended latent diffusion. ACM Transactions on Graphics 42(4), 1-11 (Jul 2023). https://doi.org/10.1145/3592450, http://dx.doi.org/10.1145/3592450", + "3. Baradaran, M., Bergevin, R.: Future video prediction from a single frame for video anomaly detection (2023)", + "4. Blattmann, A., Dockhorn, T., Kulal, S., Mendelevitch, D., Kilian, M., Lorenz, D., Levi, Y., English, Z., Voleti, V., Letts, A., Jampani, V., Rombach, R.: Stable video diffusion: Scaling latent video diffusion models to large datasets (2023)", + "5. Blattmann, A., Rombach, R., Ling, H., Dockhorn, T., Kim, S.W., Fidler, S., Kreis, K.: Align your latents: High-resolution video synthesis with latent diffusion models (2023)", + "6. Brooks, T., Holynski, A., Efros, A.A.: Instructpix2pix: Learning to follow image editing instructions. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 18392-18402 (2023)", + "7. Chang, H., Zhang, H., Barber, J., Maschinot, A., Lezama, J., Jiang, L., Yang, M.H., Murphy, K., Freeman, W.T., Rubinstein, M., Li, Y., Krishnan, D.: Muse: Text-to-image generation via masked generative transformers (2023)", + "8. Chang, H., Zhang, H., Jiang, L., Liu, C., Freeman, W.T.: Maskgit: Masked generative image transformer (2022)", + "9. Choi, J., Kim, S., Jeong, Y., Gwon, Y., Yoon, S.: Ilvr: Conditioning method for denoising diffusion probabilistic models (2021)", + "10. Ding, M., Yang, Z., Hong, W., Zheng, W., Zhou, C., Yin, D., Lin, J., Zou, X., Shao, Z., Yang, H., et al.: Cogview: Mastering text-to-image generation via transformers. Advances in Neural Information Processing Systems 34, 19822-19835 (2021)", + "11. Esser, P., Chiu, J., Atighehchian, P., Granskog, J., Germanidis, A.: Structure and content-guided video synthesis with diffusion models. In: Proceedings of the IEEE/CVF International Conference on Computer Vision. pp. 7346-7356 (2023)", + "12. Esser, P., Rombach, R., Ommer, B.: Taming transformers for high-resolution image synthesis (2021)", + "13. Fu, T.J., Hu, W., Du, X., Wang, W.Y., Yang, Y., Gan, Z.: Guiding instruction-based image editing via multimodal large language models (2023)", + "14. Fujitake, M., Sugimoto, A.: Video representation learning through prediction for online object detection. In: Proceedings of the IEEE/CVF Winter Conference on Applications of Computer Vision. pp. 530-539 (2022)", + "15. Ge, S., Nah, S., Liu, G., Poon, T., Tao, A., Catanzaro, B., Jacobs, D., Huang, J.B., Liu, M.Y., Balaji, Y.: Preserve your own correlation: A noise prior for video diffusion models (2023)", + "16. Girdhar, R., Ramanan, D.: Cater: A diagnostic dataset for compositional actions and temporal reasoning (2020)", + "17. Goodfellow, I., Pouget-Abadie, J., Mirza, M., Xu, B., Warde-Farley, D., Ozair, S., Courville, A., Bengio, Y.: Generative adversarial networks. Communications of the ACM 63(11), 139-144 (2020)", + "18. Gu, Y., Yang, J., Usuyama, N., Li, C., Zhang, S., Lungren, M.P., Gao, J., Poon, H.: Biomedjourney: Counterfactual biomedical image generation by instruction-learning from multimodal patient journeys. arXiv preprint arXiv:2310.10765 (2023)" + ], + "bbox": [ + 218, + 178, + 784, + 839 + ], + "page_idx": 14 + }, + { + "type": "header", + "text": "Pix2Gif: Motion-Guided Diffusion for GIF Generation", + "bbox": [ + 370, + 114, + 730, + 127 + ], + "page_idx": 14 + }, + { + "type": "page_number", + "text": "15", + "bbox": [ + 767, + 116, + 784, + 126 + ], + "page_idx": 14 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "19. He, Y., Yang, T., Zhang, Y., Shan, Y., Chen, Q.: Latent video diffusion models for high-fidelity long video generation (2023)", + "20. Hertz, A., Mokady, R., Tenenbaum, J., Aberman, K., Pritch, Y., Cohen-Or, D.: Prompt-to-prompt image editing with cross attention control (2022)", + "21. Ho, J., Jain, A., Abbeel, P.: Denoising diffusion probabilistic models (2020)", + "22. Ho, J., Salimans, T.: Classifier-free diffusion guidance. arXiv preprint arXiv:2207.12598 (2022)", + "23. Ho, J., Salimans, T., Gritsenko, A., Chan, W., Norouzi, M., Fleet, D.J.: Video diffusion models (2022)", + "24. Hong, W., Ding, M., Zheng, W., Liu, X., Tang, J.: Cogvideo: Large-scale pretraining for text-to-video generation via transformers (2022)", + "25. Hu, Y., Luo, C., Chen, Z.: Make it move: Controllable image-to-video generation with text descriptions (2022)", + "26. Kang, M., Zhu, J.Y., Zhang, R., Park, J., Shechtman, E., Paris, S., Park, T.: Scaling up gans for text-to-image synthesis. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 10124-10134 (2023)", + "27. Karras, T., Laine, S., Aila, T.: A style-based generator architecture for generative adversarial networks. In: Proceedings of the IEEE/CVF conference on computer vision and pattern recognition. pp. 4401-4410 (2019)", + "28. Kawar, B., Zada, S., Lang, O., Tov, O., Chang, H., Dekel, T., Mosseri, I., Irani, M.: Imagic: Text-based real image editing with diffusion models. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 6007-6017 (2023)", + "29. Kingma, D.P., Welling, M.: Auto-encoding variational bayes (2022)", + "30. Lee, J., Lee, J., Lee, S., Yoon, S.: Mutual suppression network for video prediction using disentangled features (2019)", + "31. Li, H., Yang, Y., Chang, M., Chen, S., Feng, H., Xu, Z., Li, Q., Chen, Y.: Srdiff: Single image super-resolution with diffusion probabilistic models. Neurocomputing 479, 47-59 (2022)", + "32. Li, Y., Liu, H., Wu, Q., Mu, F., Yang, J., Gao, J., Li, C., Lee, Y.J.: Gligen: Open-set grounded text-to-image generation (2023)", + "33. Li, Y., Song, Y., Cao, L., Tetreault, J., Goldberg, L., Jaimes, A., Luo, J.: Tgif: A new dataset and benchmark on animated gif description (2016)", + "34. Liu, W., Luo, W., Lian, D., Gao, S.: Future frame prediction for anomaly detection - a new baseline (2018)", + "35. Lu, C., Zhou, Y., Bao, F., Chen, J., Li, C., Zhu, J.: Dpm-solver: A fast ode solver for diffusion probabilistic model sampling in around 10 steps (2022)", + "36. Meng, C., He, Y., Song, Y., Song, J., Wu, J., Zhu, J.Y., Ermon, S.: Sdedit: Guided image synthesis and editing with stochastic differential equations (2022)", + "37. Ni, B., Peng, H., Chen, M., Zhang, S., Meng, G., Fu, J., Xiang, S., Ling, H.: Expanding language-image pretrained models for general video recognition (2022)", + "38. Ni, H., Shi, C., Li, K., Huang, S.X., Min, M.R.: Conditional image-to-video generation with latent flow diffusion models (2023)", + "39. Oh, J., Guo, X., Lee, H., Lewis, R., Singh, S.: Action-conditional video prediction using deep networks in atari games (2015)", + "40. Oliu, M., Selva, J., Escalera, S.: Folded recurrent neural networks for future video prediction (2018)", + "41. van den Oord, A., Kalchbrenner, N., Kavukcuoglu, K.: Pixel recurrent neural networks (2016)", + "42. van den Oord, A., Kalchbrenner, N., Vinyals, O., Espeholt, L., Graves, A., Kavukcuoglu, K.: Conditional image generation with pixelCNN decoders (2016)" + ], + "bbox": [ + 215, + 146, + 784, + 839 + ], + "page_idx": 15 + }, + { + "type": "page_number", + "text": "16", + "bbox": [ + 217, + 114, + 235, + 126 + ], + "page_idx": 15 + }, + { + "type": "header", + "text": "H. Kandala et al.", + "bbox": [ + 271, + 114, + 387, + 126 + ], + "page_idx": 15 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "43. van den Oord, A., Vinyals, O., Kavukcuoglu, K.: Neural discrete representation learning (2018)", + "44. Radford, A., Metz, L., Chintala, S.: Unsupervised representation learning with deep convolutional generative adversarial networks. arXiv preprint arXiv:1511.06434 (2015)", + "45. Rombach, R., Blattmann, A., Lorenz, D., Esser, P., Ommer, B.: High-resolution image synthesis with latent diffusion models (2022)", + "46. Ruiz, N., Li, Y., Jampani, V., Pritch, Y., Rubinstein, M., Aberman, K.: Dreambooth: Fine tuning text-to-image diffusion models for subject-driven generation (2023)", + "47. Sahara, C., Chan, W., Saxena, S., Li, L., Whang, J., Denton, E.L., Ghasemipour, K., Gontijo Lopes, R., Karagol Ayan, B., Salimans, T., et al.: Photorealistic text-to-image diffusion models with deep language understanding. Advances in Neural Information Processing Systems 35, 36479-36494 (2022)", + "48. Shi, X., Chen, Z., Wang, H., Yeung, D.Y., kin Wong, W., chun Woo, W.: Convolutional LSTM network: A machine learning approach for precipitation nowcasting (2015)", + "49. Simonyan, K., Zisserman, A.: Very deep convolutional networks for large-scale image recognition (2015)", + "50. Singer, U., Polyak, A., Hayes, T., Yin, X., An, J., Zhang, S., Hu, Q., Yang, H., Ashual, O., Gafni, O., Parikh, D., Gupta, S., Taigman, Y.: Make-a-video: Text-to-video generation without text-video data (2022)", + "51. Song, J., Meng, C., Ermon, S.: Denoising diffusion implicit models (2022)", + "52. Soomro, K., Zamir, A.R., Shah, M.: Ucf101: A dataset of 101 human actions classes from videos in the wild (2012)", + "53. Srivastava, N., Mansimov, E., Salakhutdinov, R.: Unsupervised learning of video representations using lstms (2016)", + "54. Unterthiner, T., van Steenkiste, S., Kurach, K., Marinier, R., Michalski, M., Gelly, S.: Fvd: A new metric for video generation (2019)", + "55. Villegas, R., Babaeizadeh, M., Kindermans, P.J., Moraldo, H., Zhang, H., Saffar, M.T., Castro, S., Kunze, J., Erhan, D.: Phenaki: Variable length video generation from open domain textual description (2022)", + "56. Villegas, R., Yang, J., Hong, S., Lin, X., Lee, H.: Decomposing motion and content for natural video sequence prediction (2018)", + "57. Voleti, V., Jolicoeur-Martineau, A., Pal, C.: Mcvd: Masked conditional video diffusion for prediction, generation, and interpolation (2022)", + "58. Wang, C., Gu, J., Hu, P., Xu, S., Xu, H., Liang, X.: Dreamvideo: High-fidelity image-to-video generation with image retention and text guidance. arXiv preprint arXiv:2312.03018 (2023)", + "59. Wang, Y., Chen, X., Ma, X., Zhou, S., Huang, Z., Wang, Y., Yang, C., He, Y., Yu, J., Yang, P., et al.: Lavie: High-quality video generation with cascaded latent diffusion models. arXiv preprint arXiv:2309.15103 (2023)", + "60. Weissenborn, D., Täckström, O., Uszkoreit, J.: Scaling autoregressive video models (2020)", + "61. Wu, C., Huang, L., Zhang, Q., Li, B., Ji, L., Yang, F., Sapiro, G., Duan, N.: Godiva: Generating open-domain videos from natural descriptions (2021)", + "62. Wu, C., Liang, J., Ji, L., Yang, F., Fang, Y., Jiang, D., Duan, N.: Niwa: Visual synthesis pre-training for neural visual world creation (2021)", + "63. Xing, J., Xia, M., Zhang, Y., Chen, H., Wang, X., Wong, T.T., Shan, Y.: Dynamicultural: Animating open-domain images with video diffusion priors. arXiv preprint arXiv:2310.12190 (2023)" + ], + "bbox": [ + 215, + 146, + 784, + 839 + ], + "page_idx": 16 + }, + { + "type": "header", + "text": "Pix2Gif: Motion-Guided Diffusion for GIF Generation", + "bbox": [ + 370, + 114, + 730, + 127 + ], + "page_idx": 16 + }, + { + "type": "page_number", + "text": "17", + "bbox": [ + 767, + 114, + 785, + 126 + ], + "page_idx": 16 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "64. Xu, J., Mei, T., Yao, T., Rui, Y.: Msr-vtt: A large video description dataset for bridging video and language. In: Proceedings of the IEEE conference on computer vision and pattern recognition. pp. 5288-5296 (2016)", + "65. Yang, Z., Wang, J., Gan, Z., Li, L., Lin, K., Wu, C., Duan, N., Liu, Z., Liu, C., Zeng, M., Wang, L.: Reco: Region-controlled text-to-image generation (2022)", + "66. Yu, J., Xu, Y., Koh, J.Y., Luong, T., Baid, G., Wang, Z., Vasudevan, V., Ku, A., et al.: Scaling autoregressive models for content-rich text-to-image generation. arXiv preprint arXiv:2206.10789 2(3), 5 (2022)", + "67. Yu, L., Cheng, Y., Sohn, K., Lezama, J., Zhang, H., Chang, H., Hauptmann, A.G., Yang, M.H., Hao, Y., Essa, I., et al.: Magvit: Masked generative video transformer. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 10459-10469 (2023)", + "68. Yu, L., Shi, B., Pasunuru, R., Muller, B., Golovneva, O., Wang, T., Babu, A., Tang, B., Karrer, B., Sheynin, S., Ross, C., Polyak, A., Howes, R., Sharma, V., Xu, P., Tamoyan, H., Ashual, O., Singer, U., Li, S.W., Zhang, S., James, R., Ghosh, G., Taigman, Y., Fazel-Zarandi, M., Celikyilmaz, A., Zettlemoyer, L., Aghajanyan, A.: Scaling autoregressive multi-modal models: Pretraining and instruction tuning (2023)", + "69. Zhang, L., Rao, A., Agrawala, M.: Adding conditional control to text-to-image diffusion models. In: Proceedings of the IEEE/CVF International Conference on Computer Vision. pp. 3836-3847 (2023)", + "70. Zhang, S., Wang, J., Zhang, Y., Zhao, K., Yuan, H., Qin, Z., Wang, X., Zhao, D., Zhou, J.: I2vgen-xl: High-quality image-to-video synthesis via cascaded diffusion models. arXiv preprint arXiv:2311.04145 (2023)", + "71. Özgün Çiçek, Abdulkadir, A., Lienkamp, S.S., Brox, T., Ronneberger, O.: 3d u-net: Learning dense volumetric segmentation from sparse annotation (2016)" + ], + "bbox": [ + 215, + 146, + 784, + 507 + ], + "page_idx": 17 + }, + { + "type": "page_number", + "text": "18", + "bbox": [ + 217, + 114, + 235, + 126 + ], + "page_idx": 17 + }, + { + "type": "header", + "text": "H. Kandala et al.", + "bbox": [ + 271, + 114, + 387, + 126 + ], + "page_idx": 17 + } +] \ No newline at end of file diff --git a/2024/Pix2Gif_ Motion-Guided Diffusion for GIF Generation/b48548a6-50ce-46d1-9de4-333ad980d37c_model.json b/2024/Pix2Gif_ Motion-Guided Diffusion for GIF Generation/b48548a6-50ce-46d1-9de4-333ad980d37c_model.json new file mode 100644 index 0000000000000000000000000000000000000000..3aae79d470483edb0cc4fa37dbb542d4554cec97 --- /dev/null +++ b/2024/Pix2Gif_ Motion-Guided Diffusion for GIF Generation/b48548a6-50ce-46d1-9de4-333ad980d37c_model.json @@ -0,0 +1,2535 @@ +[ + [ + { + "type": "title", + "bbox": [ + 0.261, + 0.141, + 0.744, + 0.185 + ], + "angle": 0, + "content": "Pix2Gif: Motion-Guided Diffusion for GIF Generation" + }, + { + "type": "text", + "bbox": [ + 0.301, + 0.214, + 0.703, + 0.23 + ], + "angle": 0, + "content": "Hitsh Kandala1, Jianfeng Gao1, and Jianwei Yang1" + }, + { + "type": "text", + "bbox": [ + 0.283, + 0.242, + 0.72, + 0.27 + ], + "angle": 0, + "content": "Microsoft Research \nkhitesh2000@gmail.com,{jfgao, jianwei.yang}@microsoft.com" + }, + { + "type": "image", + "bbox": [ + 0.289, + 0.314, + 0.716, + 0.462 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.214, + 0.474, + 0.788, + 0.53 + ], + "angle": 0, + "content": "Fig. 1: Our model creates distinct frames based on the provided source image and caption, adjusting according to different levels of motion magnitude (optical flow magnitude) specified in the input conditions. It stands well for both high spatial quality and temporal consistency." + }, + { + "type": "text", + "bbox": [ + 0.262, + 0.591, + 0.744, + 0.84 + ], + "angle": 0, + "content": "Abstract. We present *Pix2Gif*, a motion-guided diffusion model for image-to-GIF (video) generation. We tackle this problem differently by formulating the task as an image translation problem steered by text and motion magnitude prompts, as shown in Fig. 1. To ensure that the model adheres to motion guidance, we propose a new motion-guided warping module to spatially transform the features of the source image conditioned on the two types of prompts. Furthermore, we introduce a perceptual loss to ensure the transformed feature map remains within the same space as the target image, ensuring content consistency and coherence. In preparation for the model training, we meticulously curated data by extracting coherent image frames from the TGIF video-caption dataset, which provides rich information about the temporal changes of subjects. After pretraining, we apply our model in a zero-shot manner to a number of video datasets. Extensive qualitative and quantitative experiments demonstrate the effectiveness of our model - it not only captures the semantic prompt from text but also the spatial ones from motion guidance. We train all our models using a single node of \\(16 \\times \\mathrm{V}100\\) GPUs." + } + ], + [ + { + "type": "page_number", + "bbox": [ + 0.218, + 0.116, + 0.23, + 0.127 + ], + "angle": 0, + "content": "2" + }, + { + "type": "header", + "bbox": [ + 0.272, + 0.115, + 0.388, + 0.127 + ], + "angle": 0, + "content": "H. Kandala et al." + }, + { + "type": "text", + "bbox": [ + 0.262, + 0.148, + 0.745, + 0.175 + ], + "angle": 0, + "content": "Keywords: Diffusion \\(\\cdot\\) image-to-video generation \\(\\cdot\\) image-to-image translation" + }, + { + "type": "title", + "bbox": [ + 0.217, + 0.203, + 0.376, + 0.219 + ], + "angle": 0, + "content": "1 Introduction" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.236, + 0.784, + 0.311 + ], + "angle": 0, + "content": "Visual content generation has been significantly advanced by the huge progress of diffusion models [21, 22, 35, 51]. Recently, the development of latent diffusion models (LDMs) [45] has led us to a new quality level of generated images. It has inspired a lot of works for customized and controllable image generation [32, 46, 65, 69], and fine-grained image editing [6, 20, 28, 36]." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.312, + 0.785, + 0.597 + ], + "angle": 0, + "content": "In this work, we focus on converting a single image to an animated Graphics Interchange Format (GIF), which is valuable for design yet under-explored. Despite the absence of image-to-GIF generation models, diffusion-based video generation has emerged as a hot topic recently. Compared with text-to-image generation, however, text-to-video generation requires not only high quality for individual frames but also visual consistency and temporal coherence across frames. To achieve this goal, existing works expand the LDMs to video diffusion models (VDMs) by either inflating the 2D CNNs in LDMs to 3D ones [23] or introducing an additional temporal attention layer to bridge the diffusion for each frame [5,15,19,50,59]. In addition to text prompts, a few recent works also explored the way of using images or other prompts to make the video generation model more customizable and controllable [11,38]. However, due to the high cost of VDMs to generate a sequence of video frames in one run, most (if not all) of these works require a compromise of reducing the resolution of generated frames \\((64 \\times 64\\) typically), and the usage of extra super-resolution diffusion models for upscaling [31,47]. Moreover, since these methods use the temporal attention layers to model the cross-frame dependency implicitly, it is quite hard for them to preserve good controllability of the frame-to-frame temporal dynamics in a fine-grained manner." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.599, + 0.787, + 0.84 + ], + "angle": 0, + "content": "Given that animated GIF usually contains less number of frames and requires more specializations, we take a different strategy and formulate the image-to-GIF generation as an image translation process. To decouple the generation of visual contents and temporal dynamics, we further introduce a motion flow magnitude as extra guidance in addition to image and text prompts. Unlike the aforementioned works, our model takes one or more history frames as the condition and produces only one future frame at once. This brings some unique advantages: (i) simplicity - our model can be purely built on top of LDMs and trained end-to-end with high resolution, without any cascaded diffusion processes for upscaling. (ii) controllability - we could inject detailed and different text and motion prompts at each time step for generating a frame, which gains much better controllability of the model. Our work is inspired by a line of canonical works for future frame prediction [40, 48, 56]. However, due to the lack of a powerful image-generation engine, these works fail to produce high-quality results and can only be applied to specific video domains [16, 30, 52]. Moreover, they cannot support other types of prompts or conditions than the history frames. To address this problem," + } + ], + [ + { + "type": "header", + "bbox": [ + 0.371, + 0.115, + 0.732, + 0.128 + ], + "angle": 0, + "content": "Pix2Gif: Motion-Guided Diffusion for GIF Generation" + }, + { + "type": "page_number", + "bbox": [ + 0.775, + 0.117, + 0.785, + 0.127 + ], + "angle": 0, + "content": "3" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.147, + 0.788, + 0.298 + ], + "angle": 0, + "content": "we exploit a modern diffusion-based pipeline. More specifically, we follow text-conditioned image editing approaches (e.g., InstructPix2Pix [6]), and propose a new temporal image editing to produce future frames given history frames. To train the model, we curate a new training dataset based on TGIF [33] by extracting frames and calculating the magnitude of optical flow between them. We then selected an appropriate range of the optical flow magnitude and sampled frame pairs from each GIF in a manner that ensures diversity. In the end, we train our diffusion model called Pix2Gif, which can generate high-quality animated GIF consisting of multiple frames, given a single image and text and motion magnitude prompts. In summary, our main contributions are:" + }, + { + "type": "text", + "bbox": [ + 0.226, + 0.316, + 0.784, + 0.36 + ], + "angle": 0, + "content": "- We are the first to explore an image-to-image translation formula for generating animated GIFs from an image, guided by a text prompt and motion magnitude." + }, + { + "type": "text", + "bbox": [ + 0.226, + 0.363, + 0.784, + 0.406 + ], + "angle": 0, + "content": "- We propose a flow-based warping module with a perceptual loss in the diffusion process that takes motion magnitude as input and controls the temporal dynamics and consistency between future frames and the initial ones." + }, + { + "type": "text", + "bbox": [ + 0.226, + 0.409, + 0.784, + 0.437 + ], + "angle": 0, + "content": "- We curate a new dataset, comprised of 78,692 short GIF clips for training, and 10,546 for evaluation. The new dataset covers a variety of visual domains." + }, + { + "type": "text", + "bbox": [ + 0.226, + 0.441, + 0.784, + 0.484 + ], + "angle": 0, + "content": "- Quantitative and qualitative results demonstrate the effectiveness of our proposed method for generating visually consistent coherent GIFs from a single image, and it can be generalized to a wide range of visual domains." + }, + { + "type": "list", + "bbox": [ + 0.226, + 0.316, + 0.784, + 0.484 + ], + "angle": 0, + "content": null + }, + { + "type": "title", + "bbox": [ + 0.216, + 0.514, + 0.364, + 0.53 + ], + "angle": 0, + "content": "2 Prior Work" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.552, + 0.788, + 0.704 + ], + "angle": 0, + "content": "Image and video generation has been a long-standing problem in the community. It can be tackled by different approaches, which can be categorized into four groups: generative adversarial networks (GANs) [17,26,27,44], transformer-based autoregressive decoding [10,12,41,42,60,66], masked image modeling [7,8,55,67]. Most of the recent works exploited diffusion models for image generation given their high-quality outputs and huge open-source supports [45, 47]. Recently, a number of works have extended the text-to-image generation model into image translation or editing models [6,20,28,36,69] or video generation models [5,15, 23,50,57]. Below we provide a brief overview of the related diffusion-based image and video generation methods." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.72, + 0.788, + 0.84 + ], + "angle": 0, + "content": "Image-to-Image Translation. Diffusion-based image-to-image generation has drawn increasing attention. Different from text-to-image generation, it takes an image as input and edits its contents following the text instructions while keeping the unrelated parts unchanged. SDEdit [36] and ILVR [9] are two pioneering works that impose reference image conditions to an existing latent diffusion model for controllable image generation. Later on, to conduct local edits, the authors in [2] proposed blended latent diffusion to steer the diffusion process with a user-specified mask, where the pixels out of the mask remain the same" + } + ], + [ + { + "type": "page_number", + "bbox": [ + 0.218, + 0.116, + 0.23, + 0.127 + ], + "angle": 0, + "content": "4" + }, + { + "type": "header", + "bbox": [ + 0.272, + 0.115, + 0.388, + 0.128 + ], + "angle": 0, + "content": "H. Kandala et al." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.147, + 0.788, + 0.269 + ], + "angle": 0, + "content": "as the input image while the region inside is edited following the textual description. Instead of manipulating the image space, Prompt2Prompt [20] proposed to edit the image by manipulating the textual context (e.g., swapping or adding words.) to which the latent diffusion model cross-attends. However, this method requires forwarding a text-to-image generation process to obtain the cross-attention maps, and thus cannot be applied to real images. Imagic [28] proposed to blend the embeddings of a real image with the textual context embedding so that the generated image obeys both the image and text conditions." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.271, + 0.788, + 0.438 + ], + "angle": 0, + "content": "All the aforementioned works leverage a frozen latent diffusion model and control the generation with modified text or image prompts. To enable arbitrary image editing, InstructPix2Pix [6] proposed to finetune the LDM to follow user instructions that precisely convey the user intents, e.g., \"change the cat to dog\". The model is trained by a synthetic dataset consisting of triplets \\(\\langle \\text{image}_{\\text{src}}, \\text{instruction}, \\text{image}_{\\text{tgt}} \\rangle\\). The resulting model could allow both realistic and generated images and support arbitrary language instructions. Some other works also exploit a similar way to train the model to follow instructions [18,68]. To further enhance the language understanding, MGIE [13] exploited a large multimodal model to produce a more comprehensive textual context for the instructed image editing." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.441, + 0.788, + 0.532 + ], + "angle": 0, + "content": "In this work, we employ the image-to-image translation pipeline and are the first to formulate a GIF generation as an image translation problem. Given a reference image, the goal is to generate a realistic future frame following a textual instruction. Therefore, the focus is on how to perform temporal rather than spatial editing on a source image. When the process rolls out, it gradually gives a sequence of frames." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.55, + 0.788, + 0.777 + ], + "angle": 0, + "content": "Conditioned Video Generation. Speaking of the high-level goal, our work resembles conditioned video generation. For video generation, a conventional way is inflating the 2D U-Net used in LDM to 3D U-Net [71] by replacing the 2D convolution layers with 3D ones. Likewise, a similar strategy is taken in [5,15,19,50,59], but with a slight difference in that they use interleaved spatial and temporal attention layers in the U-Net. Due to the high cost of generating a sequence of video frames in one shot, the output videos usually have a resolution as low as \\(64 \\times 64\\). To attain high-resolution videos, these methods need to use one or more super-resolution diffusion models [31,47] to upscale the resolution by 4 or 8 times. To accelerate the training, a pre-trained text-to-image LDM is usually used to initialize these models. Adding spatial-temporal modules is also a commonly used strategy for autoregressive models [24, 55, 61, 62]. Similarly, both [62] and [24] exploit a pretrained autoregressive image generation model as the starting point. In [55], however, the authors introduce and pretrained a new video encoder, which is then used to train a masked video decoder." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.78, + 0.788, + 0.842 + ], + "angle": 0, + "content": "Besides text-to-video generation, using images as the condition for video generation draws increasing attention. On one hand, once a text-to-video generation is trained, it can be further finetuned for image-to-video generation [4]. In [11], the authors introduce additional structural conditions (e.g., depth maps) for" + } + ], + [ + { + "type": "header", + "bbox": [ + 0.371, + 0.115, + 0.732, + 0.129 + ], + "angle": 0, + "content": "Pix2Gif: Motion-Guided Diffusion for GIF Generation" + }, + { + "type": "page_number", + "bbox": [ + 0.775, + 0.117, + 0.785, + 0.127 + ], + "angle": 0, + "content": "5" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.147, + 0.784, + 0.237 + ], + "angle": 0, + "content": "more controllable video generation. Alternatively in [38], a latent flow diffusion model is introduced for image-conditioned video generation by explicitly generating a sequence of optimal flows and masks as the guidance. On the other hand, a few concurrent works to ours directly approach image-to-video generation on top of video diffusion models [58, 63, 70]. All these works share a similar spirit to text-to-video generation models but add additional images as the reference." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.24, + 0.785, + 0.316 + ], + "angle": 0, + "content": "Our method uses a diffusion model but differs from all the aforementioned methods in that we reformulate video generation as a frame-to-frame translation problem based on the history frames. As [6] suggests image-to-image translation can maintain a decent visual consistency. In addition, we also introduce a motion flow magnitude as another condition to explicitly control the temporal dynamics." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.334, + 0.787, + 0.545 + ], + "angle": 0, + "content": "Future Frame Prediction. Future frame prediction or forecasting [40, 48, 56] has been a long-standing problem before the prevalence of diffusion models. It has been used as an anomaly detection approach by comparing the observed frame and the predicted ones [3, 34] and video representation learning for various downstream tasks [14]. For these problems, a recurrent network such as LSTM [53], ConvLSTM [39, 56] or 3D-CNN [1] is usually used as the model architecture, and GAN [17] or Variational Autoencoder (VAE) [29] is used as the learning objectives. With the emergence of VQ-VAE [43], the authors in [25] exploited axial transformer blocks to chain the encoder-decoder for autoregressive next-frame prediction. In [57], the authors proposed masked conditional video diffusion to unify different tasks of video prediction, generation and interpolation. Nevertheless, all of these models are trained on domain-specific video datasets such as MovingMNIST [30], CATER [16] and UCF-101 [52], etc, far from being a generic video generation model." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.548, + 0.788, + 0.653 + ], + "angle": 0, + "content": "Our work takes inspiration from future frame prediction methods but proposes a simpler yet effective strategy by formulating it as an image-to-image translation problem. Furthermore, our model simultaneously takes image, text and motion magnitude as the guidance for better controllability. To attain a model as general as possible, we curate a new training dataset covering a wide range of domains. Without any further dataset-specific finetuning, our model achieves plausible video generation results as shown in Fig. 1." + }, + { + "type": "title", + "bbox": [ + 0.216, + 0.69, + 0.33, + 0.705 + ], + "angle": 0, + "content": "3 Method" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.735, + 0.788, + 0.84 + ], + "angle": 0, + "content": "Our goal is to generate GIFs, given an initial frame, a motion description, and a measure of optical flow. We frame this as an image-to-image translation problem based on latent diffusion. We first explain how we created our training dataset in Sec. 3.1, then outline our model's principles and training strategy in Sec. 3.2. Next, we delve into the specifics of our proposed model, explaining its various components in Sec. 3.3. Finally, we focus on the loss functions used to train our model in Sec. 3.4." + } + ], + [ + { + "type": "page_number", + "bbox": [ + 0.218, + 0.116, + 0.23, + 0.127 + ], + "angle": 0, + "content": "6" + }, + { + "type": "header", + "bbox": [ + 0.272, + 0.115, + 0.388, + 0.127 + ], + "angle": 0, + "content": "H. Kandala et al." + }, + { + "type": "image", + "bbox": [ + 0.288, + 0.144, + 0.716, + 0.27 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.214, + 0.28, + 0.789, + 0.322 + ], + "angle": 0, + "content": "Fig. 2: The three step process of curating the TGIF dataset. Starting from extracting frames \\( A \\) to restricting the range of optical flow \\( B \\) and then maintaining the diversity of pairs \\( C \\)." + }, + { + "type": "image", + "bbox": [ + 0.288, + 0.339, + 0.72, + 0.482 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.214, + 0.492, + 0.789, + 0.534 + ], + "angle": 0, + "content": "Fig.3: \\( \\text{Pix2Gif} \\) model pipeline. We propose an end-to-end network where the inputs are encoded by \\( \\mathcal{E} \\), CLIP and \\( \\mathcal{M} \\) to output \\( \\mathcal{E}(c_L), c_T \\) and \\( c_M \\) respectively, which then goes into \\( \\mathcal{W} \\) to form the conditioning input for LDM." + }, + { + "type": "title", + "bbox": [ + 0.216, + 0.567, + 0.327, + 0.58 + ], + "angle": 0, + "content": "3.1 Dataset" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.597, + 0.789, + 0.642 + ], + "angle": 0, + "content": "We used the Tumblr GIF (TGIF) dataset [33], which predominantly consists of human-centric animated GIFs described by captions. The dataset features a range of GIFs with varied movements over 1-3 seconds." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.644, + 0.789, + 0.841 + ], + "angle": 0, + "content": "The curation process, as shown in Fig. 2, involved extracting frames from all GIFs and calculating the optical flow between all possible frame pairs. The number of extracted frames from each GIF varied, with an average of about 41 frames. The optical flow histogram calculated between all frames ranged from [0, 200]. We selected the [2, 20] range, capturing small yet significant motion and excluding pairs with drastic changes. Despite the restricted range, we still had a significant number of training pairs. To avoid model overfitting and maintain diversity, we randomly selected a minimum of 10 pairs or the number of pairs within the restricted range from each GIF. This approach resulted in a final dataset with nearly equal representation of all values within the selected range. The final dataset contained 783,184 training pairs and 105,041 validation pairs. Each data point consisted of a pair of frames from the same GIF, the corresponding caption, and the calculated optical flow between the frames." + } + ], + [ + { + "type": "header", + "bbox": [ + 0.371, + 0.115, + 0.731, + 0.128 + ], + "angle": 0, + "content": "Pix2Gif: Motion-Guided Diffusion for GIF Generation" + }, + { + "type": "page_number", + "bbox": [ + 0.775, + 0.117, + 0.785, + 0.127 + ], + "angle": 0, + "content": "7" + }, + { + "type": "title", + "bbox": [ + 0.216, + 0.147, + 0.58, + 0.163 + ], + "angle": 0, + "content": "3.2 Preliminary: Instructed Image Editing" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.179, + 0.788, + 0.332 + ], + "angle": 0, + "content": "Our model is fundamentally grounded in the latent diffusion models (LDMs) for image generation and editing [6, 45]. More specifically, we build upon InstructPix2Pix [6] by framing our objective in the context of an instructed image-to-image translation task. Given an image \\( x \\), the forward diffusion procedure introduces noise to the encoded latent \\( z \\), thereby producing a noisy latent vector \\( z_{t} \\). This process is carried out over \\( T \\) timesteps, with each timestep \\( t \\in \\{1, \\dots, T\\} \\) seeing an increment in the noise level until it culminates into a random noise \\( n \\). A network \\( e_{\\theta} \\) is trained by minimizing the following latent diffusion objective to predict noise existing in the noisy latent \\( z_{t} \\), considering factors in image conditioning \\( c_{I} \\) and textual instruction \\( c_{T} \\):" + }, + { + "type": "equation", + "bbox": [ + 0.295, + 0.344, + 0.786, + 0.363 + ], + "angle": 0, + "content": "\\[\nL _ {L D M} = \\mathbb {E} _ {\\mathcal {E} (x), \\mathcal {E} (c _ {I}), c _ {T}, \\epsilon \\sim \\mathcal {N} (0, 1), t} \\left[ | | \\epsilon - \\epsilon_ {\\theta} \\left(z _ {t}, t, \\mathcal {E} \\left(c _ {I}\\right), c _ {T}\\right) | | _ {2} ^ {2} \\right] \\tag {1}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.375, + 0.788, + 0.452 + ], + "angle": 0, + "content": "where \\(\\mathcal{E}\\) is the VQ-VAE encoder that transforms the images from pixel space to discrete latent space. To facilitate image conditioning, \\(z_{t}\\) and \\(\\mathcal{E}(c_I)\\) are concatenated and then fed into a convolutional layer. The model is trained for conditional and unconditional denoising, given the image and caption condition individually or collectively." + }, + { + "type": "title", + "bbox": [ + 0.216, + 0.479, + 0.431, + 0.494 + ], + "angle": 0, + "content": "3.3 Our Model: Pix2Gif" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.511, + 0.787, + 0.587 + ], + "angle": 0, + "content": "We build our model similar to InstructPix2Pix and frame our objective in the context of a text-instructed and motion-guided temporal editing problem. Compared with the original InstructPix2Pix pipeline, the main innovation is the newly introduced motion-based warping module. The overall model pipeline is shown in Fig. 3." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.588, + 0.788, + 0.724 + ], + "angle": 0, + "content": "Our model takes three inputs: an image, a text instruction, and a motion magnitude. These inputs are fed into the model through two pathways - once through the diffusion model directly, and again through the warping module, which will be discussed in Sec. 3.3 and Sec. 3.3. When passed through the caption, we add the phrase \"The optical flow is _. to the original caption. The flow input is then appended at the end in a word format rather than a numerical one, as the CLIP model tends to assign higher similarity scores to word forms than to numerical representations of numbers for the same image. Finally, our model is trained by minimizing the following loss function:" + }, + { + "type": "equation", + "bbox": [ + 0.27, + 0.739, + 0.786, + 0.757 + ], + "angle": 0, + "content": "\\[\nL _ {L D M} ^ {\\prime} = \\mathbb {E} _ {\\mathcal {E} (x), \\mathcal {E} (c _ {I}), c _ {T}, c _ {M}, \\epsilon \\sim \\mathcal {N} (0, 1), t} \\left[ | | \\epsilon - \\epsilon_ {\\theta} (z _ {t}, t, \\mathcal {E} (c _ {I}), c _ {T}, c _ {M}) | | _ {2} ^ {2} \\right] \\tag {2}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.769, + 0.788, + 0.845 + ], + "angle": 0, + "content": "where \\( c_{M} \\) is the motion condition. The altered caption is processed via the pretrained CLIP model to yield \\( c_{T} \\), while the output of \\( \\mathcal{M} \\) gives us \\( c_{M} \\). These two conditions are then added linearly, serving as the conditioning input for both the Warping Module \\( \\mathcal{W} \\) (discussed in Sec. 3.3) and the Latent Diffusion Model LDM (referenced in Sec. 3.2)." + } + ], + [ + { + "type": "page_number", + "bbox": [ + 0.218, + 0.116, + 0.23, + 0.127 + ], + "angle": 0, + "content": "8" + }, + { + "type": "header", + "bbox": [ + 0.272, + 0.115, + 0.388, + 0.127 + ], + "angle": 0, + "content": "H. Kandala et al." + }, + { + "type": "image", + "bbox": [ + 0.304, + 0.146, + 0.703, + 0.313 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.215, + 0.324, + 0.785, + 0.352 + ], + "angle": 0, + "content": "Fig. 4: Deep dive into the Warping Module \\(\\mathcal{W}\\). It comprises of three units: \\(\\mathcal{I}_M\\), \\(\\mathcal{F}_{Net}\\) and \\(\\mathcal{W}_{Net}\\)." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.389, + 0.788, + 0.542 + ], + "angle": 0, + "content": "Motion Embedding Layer In traditional conditional diffusion models like [6], text prompts are often enough for image generation. We initially passed the motion input indirectly through the prompt. However, this divided the model's attention on a single caption token, which is problematic when the main caption stays the same but the motion input changes. To let the model focus on the motion input, we included an embedding layer that converts the motion input into an integer and selects an embedding vector. This vector is then duplicated and concatenated with itself to generate \\( c_{M} \\), which when combined with the caption embedding \\( c_{T} \\), provides the conditioning input \\( c_{L} = c_{T} + c_{M} \\) for both warping module \\( \\mathcal{W} \\) and LDM." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.568, + 0.789, + 0.844 + ], + "angle": 0, + "content": "Warping Module One of the main components of Pix2Gif is the Warping Module \\(\\mathcal{W}\\). As illustrated in Fig. 4, it technically comprises two networks: the FlowNet \\((\\mathcal{F}_{Net})\\) and the WarpNet \\((\\mathcal{W}_{Net})\\). Ordinarily, the computation of optical flow involves two images. However, in this case, we initially have only one image - the source image - and that too in the latent domain. Thus, our goal is to learn the optical flow utilizing just one latent image. This is achieved via \\(\\mathcal{F}_{Net}\\), conditioned on \\(c_{L}\\), which guides it to generate a flow feature map in the intended direction with the hint of text and motion prompts. This condition is processed by the Injection Module \\((\\mathcal{I}_{\\mathcal{M}})\\), a compact encoder designed to make \\(c_{L}\\) compatible for concatenation with one of the intermediate feature maps near the end of the network. This configuration enables \\(\\mathcal{F}_{Net}\\) to independently learn high-level features, which are then guided in the desired direction with the introduction of \\(c_{L}\\). The architecture of \\(\\mathcal{F}_{Net}\\) resembles that of UNet, producing an output with a fixed channel of 2 to capture changes in the horizontal and vertical components. This optical flow feature map \\(F = \\mathcal{F}_{Net}(\\mathcal{E}(c_{I}), \\mathcal{I}_{\\mathcal{M}}(c_{L}))\\), along with the source latent \\((\\mathcal{E}(c_{I}))\\), is then processed through \\(\\mathcal{W}_{Net}\\) to yield a Fischer map \\(z_{W} = \\mathcal{W}_{Net}(\\mathcal{E}(c_{I}), F)\\). This transformation is learned more efficiently and abstractly in the latent space than in the pixel space." + } + ], + [ + { + "type": "header", + "bbox": [ + 0.371, + 0.115, + 0.732, + 0.128 + ], + "angle": 0, + "content": "Pix2Gif: Motion-Guided Diffusion for GIF Generation" + }, + { + "type": "page_number", + "bbox": [ + 0.775, + 0.117, + 0.785, + 0.127 + ], + "angle": 0, + "content": "9" + }, + { + "type": "table_caption", + "bbox": [ + 0.216, + 0.145, + 0.784, + 0.173 + ], + "angle": 0, + "content": "Table 1: Quantitative comparison with state-of-the-art image-text-to-video generation models for the zero-shot setting." + }, + { + "type": "table", + "bbox": [ + 0.278, + 0.185, + 0.725, + 0.266 + ], + "angle": 0, + "content": "
MethodUCF-101 [52]MSR-VTT [64]
FVD ↓CLIPSim ↑PIC ↑FVD ↓CLIPSim ↑PIC ↑
I2VGen-XL [70]563.120.28650.6329278.620.22720.6018
DynamiCrafter [63]527.060.27960.6307271.630.26020.6135
Pix2Gif (Ours)285.020.28150.8763168.690.25730.8521
" + }, + { + "type": "title", + "bbox": [ + 0.216, + 0.303, + 0.315, + 0.316 + ], + "angle": 0, + "content": "3.4 Losses" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.327, + 0.784, + 0.37 + ], + "angle": 0, + "content": "Our model incorporates two different types of losses. The first type is the standard L2 loss Eq. (2), which is utilized by the stable diffusion model and talked about in Sec. 3.2" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.372, + 0.785, + 0.462 + ], + "angle": 0, + "content": "The second loss type in our model is the perceptual loss, which is calculated by comparing the latent features of the image condition \\(\\mathcal{E}(c_I)\\) and the warped image \\(z_W\\). This is implemented using a pre-trained VGG network [49], modified to accommodate 4 channels instead of the standard 3. The modification involves averaging the weights from the first three channels to initialize the fourth. The perceptual loss, \\(L_{p}\\), is defined as:" + }, + { + "type": "equation", + "bbox": [ + 0.335, + 0.474, + 0.785, + 0.504 + ], + "angle": 0, + "content": "\\[\nL _ {p} \\left(\\mathcal {E} \\left(c _ {I}\\right), z _ {W}\\right) = \\sum_ {k} \\lambda_ {k} \\left\\| \\phi_ {k} \\left(\\mathcal {E} \\left(c _ {I}\\right)\\right) - \\phi_ {k} \\left(z _ {W}\\right) \\right\\| ^ {2} \\tag {3}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.515, + 0.785, + 0.59 + ], + "angle": 0, + "content": "Here, \\(\\phi_k(.)\\) be the feature map of the k-th layer of the VGG network, \\(||.||\\) denotes the Frobenius norm, and \\(\\lambda_{k}\\) is a weighting factor. This loss ensures that the warped image maintains high-level features like edges, textures, and object types, making the images more perceptually and semantically similar and preserving the overall structure of the source image." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.591, + 0.785, + 0.621 + ], + "angle": 0, + "content": "In conclusion, the total loss function, denoted as \\( L_{T} \\), for our objective is computed by a weighted sum of the two individual losses." + }, + { + "type": "equation", + "bbox": [ + 0.426, + 0.636, + 0.785, + 0.653 + ], + "angle": 0, + "content": "\\[\nL _ {T} = L _ {L D M} ^ {\\prime} + \\lambda_ {P} L _ {P} \\tag {4}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.658, + 0.784, + 0.703 + ], + "angle": 0, + "content": "Here, \\(\\lambda_{P}\\) is the weighting factor for perceptual loss. These two losses together provide a holistic framework to train our model by ensuring pixel-level accuracy, preservation of high-level features, and smooth motion transitions." + }, + { + "type": "title", + "bbox": [ + 0.216, + 0.725, + 0.376, + 0.743 + ], + "angle": 0, + "content": "4 Experiments" + }, + { + "type": "title", + "bbox": [ + 0.216, + 0.756, + 0.312, + 0.771 + ], + "angle": 0, + "content": "4.1 Setup" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.78, + 0.785, + 0.841 + ], + "angle": 0, + "content": "Datasets We utilize the Tumbled GIF (TGIF) dataset for our training and validation purposes as discussed in Sec. 3.1. We evaluate our model on two datasets: MSR-VTT [64] and UCF-101 [52], following the common practice. For these datasets, we follow the sampling strategy as outlined in [63]." + } + ], + [ + { + "type": "page_number", + "bbox": [ + 0.218, + 0.116, + 0.236, + 0.127 + ], + "angle": 0, + "content": "10" + }, + { + "type": "header", + "bbox": [ + 0.272, + 0.115, + 0.388, + 0.127 + ], + "angle": 0, + "content": "H. Kandala et al." + }, + { + "type": "image", + "bbox": [ + 0.29, + 0.145, + 0.716, + 0.305 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.492, + 0.307, + 0.512, + 0.318 + ], + "angle": 0, + "content": "(a)" + }, + { + "type": "image", + "bbox": [ + 0.304, + 0.319, + 0.703, + 0.487 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.492, + 0.489, + 0.513, + 0.502 + ], + "angle": 0, + "content": "(b)" + }, + { + "type": "image_caption", + "bbox": [ + 0.215, + 0.513, + 0.788, + 0.555 + ], + "angle": 0, + "content": "Fig. 5: Comparison studies with other image-text to video models. Given a source image and a caption, frames are extracted from the generated 16-frame video at 256x256 resolution." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.585, + 0.788, + 0.678 + ], + "angle": 0, + "content": "**Implementations** Our model is initialized with the exponential moving average (EMA) weights of the Stable Diffusion v1.5 checkpoint and the improved ftMSE autoencoder weights\\(^2\\). We trained the model at 256x256 resolution for 7 epochs on a single node of 16 V100 GPUs for 25k steps. We used the AdamW optimizer with a learning rate of \\(10^{-4}\\). We set the weighting factor for perceptual loss \\((\\lambda_P)\\) as \\(10^{-2}\\)." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.698, + 0.788, + 0.777 + ], + "angle": 0, + "content": "Metrics We report Frechet Video Distance (FVD) [54], CLIP Similarity (CLIP-Sim) which is the average similarity calculated for all the generated frames with the input caption and Perceptual Input Conformity (PIC) as described in [63] for all methods. For comparison, we assess the zero-shot generation performance on I2VGen-XL [70] and DynamiCrafter [63]." + }, + { + "type": "page_footnote", + "bbox": [ + 0.218, + 0.783, + 0.765, + 0.812 + ], + "angle": 0, + "content": "1 https://huggingface.co/runwayml/stable-diffusion-v1-5/blob/main/v1-5-pruned-emaonly.ckpt" + }, + { + "type": "page_footnote", + "bbox": [ + 0.22, + 0.812, + 0.763, + 0.84 + ], + "angle": 0, + "content": "\\(^{2}\\) https://huggingface.co/stabilityai/sd-vae-ft-mse-original/blob/main/vae-ft-mse-840000-ema-pruned.ckpt" + }, + { + "type": "list", + "bbox": [ + 0.218, + 0.783, + 0.765, + 0.84 + ], + "angle": 0, + "content": null + } + ], + [ + { + "type": "header", + "bbox": [ + 0.371, + 0.115, + 0.732, + 0.128 + ], + "angle": 0, + "content": "Pix2Gif: Motion-Guided Diffusion for GIF Generation" + }, + { + "type": "page_number", + "bbox": [ + 0.768, + 0.116, + 0.784, + 0.127 + ], + "angle": 0, + "content": "11" + }, + { + "type": "image_caption", + "bbox": [ + 0.219, + 0.26, + 0.404, + 0.272 + ], + "angle": 0, + "content": "(a) A1: Cat is playing with wool" + }, + { + "type": "image", + "bbox": [ + 0.226, + 0.272, + 0.401, + 0.313 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.435, + 0.26, + 0.57, + 0.272 + ], + "angle": 0, + "content": "(b) A2: Cat is dancing" + }, + { + "type": "image", + "bbox": [ + 0.415, + 0.272, + 0.589, + 0.313 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.599, + 0.25, + 0.785, + 0.272 + ], + "angle": 0, + "content": "(c) A1+2: Cat dancing and playing with wool" + }, + { + "type": "image", + "bbox": [ + 0.603, + 0.272, + 0.778, + 0.313 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.215, + 0.324, + 0.785, + 0.353 + ], + "angle": 0, + "content": "Fig. 6: Pix2Gif showing composition capabilities for different types of motions. [GIFs best viewed in Adobe Acrobat Reader]" + }, + { + "type": "table_caption", + "bbox": [ + 0.215, + 0.367, + 0.785, + 0.396 + ], + "angle": 0, + "content": "Table 2: Ablation study comparing image translation methods with a focus on motion coherency at varyingcfg_img values." + }, + { + "type": "table", + "bbox": [ + 0.265, + 0.407, + 0.739, + 0.487 + ], + "angle": 0, + "content": "
Method /cfgimg1.41.61.82.0
L2 ↓PCC ↑L2 ↓PCC ↑L2 ↓PCC ↑L2 ↓PCC ↑
InstructPix2Pix [6]23.429-0.22925.492-0.02827.037-0.42327.5300.139
Pix2Gif-Base7.5800.9895.1880.9875.5950.9927.0290.991
Pix2Gif1.7460.9951.9720.9952.9440.9974.0760.997
" + }, + { + "type": "title", + "bbox": [ + 0.216, + 0.528, + 0.324, + 0.541 + ], + "angle": 0, + "content": "4.2 Results" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.553, + 0.785, + 0.599 + ], + "angle": 0, + "content": "Comparisons with previous works Fig. 5 and Tab. 1 provide a qualitative and quantitative comparison of three image-text to video models: I2VGenXL [70], DynamiCrafter [63], and our \\(Pix2Gif\\)." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.599, + 0.787, + 0.795 + ], + "angle": 0, + "content": "In Fig. 5a, the I2VGen-XL model misshapes the dog's face and generates it sideways in a nonsensical manner. DynamiCrafter appears to disregard the input parameters, as the initial frame differs significantly in position, color, and texture. It is also challenging to discern whether the dog is eating or merely moving its mouth. Our model, Pix2Gif, accurately retains all the dog's details and successfully depicts it eating from a plate. In Fig. 5b, we assess the models' capabilities by generating a video from a relatively dark image. Once again, I2VGen-XL starts strong, producing some impressive frames, but these soon turn into highly stylized and improbable images. DynamiCrafter appears to misinterpret the input image, generating something significantly different, although it seems to adhere to the caption. Conversely, Pix2Gif comprehends the inputs effectively and produces corresponding motion while preserving the overall integrity of the source image." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.795, + 0.787, + 0.84 + ], + "angle": 0, + "content": "Quantitatively, \\( \\text{Pix2Gif} \\) excels in both the FVD and PIC metrics shown in Tab. 1, which aligns with our observations of the frames generated in Fig. 5. These frames effectively preserve the structure and closely adhere to the input" + } + ], + [ + { + "type": "page_number", + "bbox": [ + 0.218, + 0.116, + 0.236, + 0.127 + ], + "angle": 0, + "content": "12" + }, + { + "type": "header", + "bbox": [ + 0.272, + 0.115, + 0.388, + 0.127 + ], + "angle": 0, + "content": "H. Kandala et al." + }, + { + "type": "image", + "bbox": [ + 0.268, + 0.151, + 0.444, + 0.256 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.296, + 0.263, + 0.427, + 0.275 + ], + "angle": 0, + "content": "(a) source frame score" + }, + { + "type": "image", + "bbox": [ + 0.55, + 0.151, + 0.727, + 0.256 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.578, + 0.263, + 0.708, + 0.275 + ], + "angle": 0, + "content": "(b) target frame score" + }, + { + "type": "image_caption", + "bbox": [ + 0.215, + 0.286, + 0.785, + 0.315 + ], + "angle": 0, + "content": "Fig. 7: Ablation study between the earlier variants of our model by comparing average similarity score for 100 samples." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.344, + 0.788, + 0.511 + ], + "angle": 0, + "content": "prompts (source image and caption). However, Pix2Gif does not perform as well in the CLIPSim metric, despite accurately following the caption. The other two models as seen in Fig. 5 do follow the caption, but they fail to adhere to the input image and produce plausible temporal transitions. This is partially attributed to the inherent model design in these two methods. Both methods attempt to generate a full sequence of frames at once using the 3D diffusion network, which inevitably compounds the spatial and temporal dimensions. Moreover, the results indicate that they function more as text-to-video models than image-text-to-video models, especially DynamiCrafter. This discrepancy also raises questions about the effectiveness of the CLIPSim metric for evaluating image-text-to-video models and calls for more sophisticated metrics for evaluating video generation." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.525, + 0.788, + 0.678 + ], + "angle": 0, + "content": "Compositionality of actions Fig. 6 illustrates an intriguing emerging capability of \\( \\text{Pix2Gif} \\): the ability to combine actions. In Fig. 6a, we see a cat playing with wool, with only the cat's paws and the wool moving. In Fig. 6b, we instruct the cat to dance, resulting in the cat moving its body but the wool remaining still. Finally, in Fig. 6c, we provide a caption that blends the actions from Fig. 6a and Fig. 6b. The result is a scene where the cat is both moving the wool and its body. This demonstrates \\( \\text{Pix2Gif} \\)'s ability to comprehend the caption and its associated motion, and to convert that understanding into a GIF. Such compositional capability significantly increases user controllability, a crucial aspect for practical applications." + }, + { + "type": "title", + "bbox": [ + 0.216, + 0.699, + 0.342, + 0.713 + ], + "angle": 0, + "content": "4.3 Ablations" + }, + { + "type": "text", + "bbox": [ + 0.216, + 0.725, + 0.658, + 0.74 + ], + "angle": 0, + "content": "We design a few variants of \\( \\text{Pix2Gif} \\) for our ablation studies:" + }, + { + "type": "text", + "bbox": [ + 0.225, + 0.749, + 0.784, + 0.779 + ], + "angle": 0, + "content": "- Pix2Gif-Base: We train InstructPix2Pix with our data, and append the text prompt with \"The optical flow is \"." + }, + { + "type": "text", + "bbox": [ + 0.225, + 0.78, + 0.785, + 0.809 + ], + "angle": 0, + "content": "- Pix2Gif-Motion-embed: The motion embedding layer is added to encode the motion magnitude and combined with textual embedding." + }, + { + "type": "text", + "bbox": [ + 0.225, + 0.81, + 0.784, + 0.839 + ], + "angle": 0, + "content": "- Pix2Gif-Warp: We further add the warping module into the model but differently only use the warped feature for the LDM." + }, + { + "type": "list", + "bbox": [ + 0.225, + 0.749, + 0.785, + 0.839 + ], + "angle": 0, + "content": null + } + ], + [ + { + "type": "header", + "bbox": [ + 0.371, + 0.115, + 0.732, + 0.128 + ], + "angle": 0, + "content": "Pix2Gif: Motion-Guided Diffusion for GIF Generation" + }, + { + "type": "page_number", + "bbox": [ + 0.769, + 0.116, + 0.785, + 0.127 + ], + "angle": 0, + "content": "13" + }, + { + "type": "image", + "bbox": [ + 0.27, + 0.151, + 0.444, + 0.256 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.295, + 0.263, + 0.427, + 0.275 + ], + "angle": 0, + "content": "(a) source frame score" + }, + { + "type": "image", + "bbox": [ + 0.553, + 0.152, + 0.726, + 0.256 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.578, + 0.263, + 0.708, + 0.275 + ], + "angle": 0, + "content": "(b) target frame score" + }, + { + "type": "image_caption", + "bbox": [ + 0.215, + 0.286, + 0.785, + 0.315 + ], + "angle": 0, + "content": "Fig. 8: Ablation study on ways to input \\( z_{W} \\) to LDM by comparing average similarity score for 100 samples." + }, + { + "type": "text", + "bbox": [ + 0.225, + 0.345, + 0.784, + 0.375 + ], + "angle": 0, + "content": "- Pix2Gif-Warp-add: Different from Pix2Gif-Warp, we instead add the warped feature and source image feature as input to the LDM." + }, + { + "type": "text", + "bbox": [ + 0.226, + 0.376, + 0.785, + 0.421 + ], + "angle": 0, + "content": "- Pix2Gif-Warp-concat: Instead of adding in Pix2Gif-Warp-add, we concatenate the warped feature and source image feature as the input, but do not include the perceptual loss." + }, + { + "type": "list", + "bbox": [ + 0.225, + 0.345, + 0.785, + 0.421 + ], + "angle": 0, + "content": null + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.432, + 0.788, + 0.538 + ], + "angle": 0, + "content": "For comparative studies with our model's variants, we generate an 8-frame video. We use the X-CLIP model [37] to extract features from our generated video, and CLIP to extract features from the source, target, and generated frames. The optimal range of CFG.img for best results is considered to be [1.6, 2.2]. Throughout this discussion, we evaluate our model's performance using four metrics, which we believe effectively measure the different aspects of generating motion through the image translation framework used in Pix2Gif." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.553, + 0.788, + 0.734 + ], + "angle": 0, + "content": "Motion Coherency Our task is framed as an image translation problem with motion magnitude as a guide. We assess the motion quality or temporal coherence in the GIFs using L2 loss and Pearson Correlation Coefficient (PCC) for InstructPix2Pix, Pix2Gif-Base, and Pix2Gif. The L2 loss evaluates the match between the motion values of the generated frames and the actual inputs, while the PCC checks if they follow the same trend. These metrics are calculated between the input motion magnitude values and the optical flow values, derived from the source image and the generated frames. Our model exhibits the highest correlation and the lowest L2 loss across all CFG.img values as seen in Tab. 2, proving its efficacy and controllability in generating GIFs with specific motions. The Pix2Gif-Base outperforms the original InstructPix2Pix, emphasizing the importance of our new dataset." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.75, + 0.788, + 0.841 + ], + "angle": 0, + "content": "Image-Video Similarity Score To evaluate the semantic properties of the video produced by our model, we created two similarity scores: a source frame score and a target frame score. The source frame score quantifies how well the video retains the primary attributes of the source frame, essentially measuring the accuracy of the source image portrayal throughout the video. The target frame score indicates the precision of the scene or subject development from" + } + ], + [ + { + "type": "page_number", + "bbox": [ + 0.218, + 0.116, + 0.236, + 0.127 + ], + "angle": 0, + "content": "14" + }, + { + "type": "header", + "bbox": [ + 0.272, + 0.115, + 0.388, + 0.127 + ], + "angle": 0, + "content": "H. Kandala et al." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.147, + 0.784, + 0.192 + ], + "angle": 0, + "content": "the source frame in the video. It also underscores the model's ability to handle uncertainty and potential changes, as the target frame represents a possible state that the video might reach." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.192, + 0.788, + 0.466 + ], + "angle": 0, + "content": "We began our experiments with \\( \\text{Pix2Gif-Base} \\) and then enhanced it incrementally, first with \\( \\mathcal{M} \\) (Pix2Gif-Motion-embed), then by adding \\( \\mathcal{W} \\) (Pix2Gif-Warp). Here, we conducted an ablation study shown in Fig. 8 to understand the best way to feed \\( z_{W} \\) into LDM. Our experiments included feeding only \\( z_{W} \\) (Pix2Gif-Warp), adding it with the \\( \\mathcal{E}(c_{I}) \\) (Pix2Gif-Warp-add), and concatenating both (Pix2Gif-Warp-concat). Both Pix2Gif-Warp-add and Pix2Gif-Warp-concat had higher average similarity scores than Pix2Gif-Warp, and Pix2Gif-Warp-concat performed the best. This can be attributed to the fact that in the addition process, \\( \\mathcal{E}(c_{I}) \\) loses its unique characteristics, which are required by the diffusion model for effective unconditional denoising. Therefore, to achieve the best results, we combined \\( \\mathcal{E}(c_{I}) \\) and \\( z_{W} \\) before inputting them into the concat attention layer of the LDM. Now finally we integrate \\( L_{P} \\) (Pix2Gif) and compare them in Fig. 7. As expected, Pix2Gif-Base was outperformed by the other versions. Pix2Gif generated more coherent and controlled motion, albeit often with limited extent. All models began to converge outside the optimal range, producing similar frames. The Pix2Gif-Motion-embed model creates a significant amount of one-directional motion, which can sometimes be nonsensical, and hence the addition of \\( \\mathcal{W} \\) helps to mitigate this issue." + }, + { + "type": "title", + "bbox": [ + 0.216, + 0.486, + 0.536, + 0.502 + ], + "angle": 0, + "content": "5 Limitations and Future Work" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.516, + 0.788, + 0.653 + ], + "angle": 0, + "content": "The current Pix2Gif model is our initial attempt to generate videos by treating it as an image translation task. However, this method has some limitations that prevent us from generating high-quality and long GIFs or videos. Firstly, the model generates images with a resolution of 256x256 pixels. If these images are used to generate subsequent frames, the quality of the frames deteriorates further. Secondly, due to limitations in computational power, we are only able to use a small portion of a larger, curated dataset for training our model. Our primary objective now is to improve the quality of the generated frames, as this could significantly enhance the effectiveness of this method." + }, + { + "type": "title", + "bbox": [ + 0.216, + 0.674, + 0.36, + 0.691 + ], + "angle": 0, + "content": "6 Conclusion" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.704, + 0.788, + 0.841 + ], + "angle": 0, + "content": "In this work, we proposed Pix2Gif, an image-to-GIF (video) generation model based on an image-to-image translation paradigm. To ensure temporal coherence across frames, we proposed a motion-guided warping module that learns to spatially warp the source image feature into the target one while maintaining visual consistency via a perceptual loss. Starting from TGIF, we curated a new dataset specifically used for training our model. The experimental results demonstrated the effectiveness of our model to generate GIFs with better temporal coherence compared with current state-of-the-art methods. Interestingly, the model also exhibits better controllability and some emerging action compositionality." + } + ], + [ + { + "type": "header", + "bbox": [ + 0.371, + 0.115, + 0.732, + 0.128 + ], + "angle": 0, + "content": "Pix2Gif: Motion-Guided Diffusion for GIF Generation" + }, + { + "type": "page_number", + "bbox": [ + 0.769, + 0.117, + 0.785, + 0.127 + ], + "angle": 0, + "content": "15" + }, + { + "type": "title", + "bbox": [ + 0.218, + 0.145, + 0.323, + 0.16 + ], + "angle": 0, + "content": "References" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.179, + 0.785, + 0.208 + ], + "angle": 0, + "content": "1. Aigner, S., Körner, M.: Futuregan: Anticipating the future frames of video sequences using spatio-temporal 3d convolutions in progressively growing gans (2018)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.209, + 0.785, + 0.25 + ], + "angle": 0, + "content": "2. Avrahami, O., Fried, O., Lischinski, D.: Blended latent diffusion. ACM Transactions on Graphics 42(4), 1-11 (Jul 2023). https://doi.org/10.1145/3592450, http://dx.doi.org/10.1145/3592450" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.251, + 0.785, + 0.278 + ], + "angle": 0, + "content": "3. Baradaran, M., Bergevin, R.: Future video prediction from a single frame for video anomaly detection (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.279, + 0.785, + 0.32 + ], + "angle": 0, + "content": "4. Blattmann, A., Dockhorn, T., Kulal, S., Mendelevitch, D., Kilian, M., Lorenz, D., Levi, Y., English, Z., Voleti, V., Letts, A., Jampani, V., Rombach, R.: Stable video diffusion: Scaling latent video diffusion models to large datasets (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.321, + 0.785, + 0.362 + ], + "angle": 0, + "content": "5. Blattmann, A., Rombach, R., Ling, H., Dockhorn, T., Kim, S.W., Fidler, S., Kreis, K.: Align your latents: High-resolution video synthesis with latent diffusion models (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.363, + 0.785, + 0.404 + ], + "angle": 0, + "content": "6. Brooks, T., Holynski, A., Efros, A.A.: Instructpix2pix: Learning to follow image editing instructions. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 18392-18402 (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.405, + 0.785, + 0.447 + ], + "angle": 0, + "content": "7. Chang, H., Zhang, H., Barber, J., Maschinot, A., Lezama, J., Jiang, L., Yang, M.H., Murphy, K., Freeman, W.T., Rubinstein, M., Li, Y., Krishnan, D.: Muse: Text-to-image generation via masked generative transformers (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.448, + 0.785, + 0.475 + ], + "angle": 0, + "content": "8. Chang, H., Zhang, H., Jiang, L., Liu, C., Freeman, W.T.: Maskgit: Masked generative image transformer (2022)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.476, + 0.785, + 0.503 + ], + "angle": 0, + "content": "9. Choi, J., Kim, S., Jeong, Y., Gwon, Y., Yoon, S.: Ilvr: Conditioning method for denoising diffusion probabilistic models (2021)" + }, + { + "type": "ref_text", + "bbox": [ + 0.22, + 0.504, + 0.785, + 0.545 + ], + "angle": 0, + "content": "10. Ding, M., Yang, Z., Hong, W., Zheng, W., Zhou, C., Yin, D., Lin, J., Zou, X., Shao, Z., Yang, H., et al.: Cogview: Mastering text-to-image generation via transformers. Advances in Neural Information Processing Systems 34, 19822-19835 (2021)" + }, + { + "type": "ref_text", + "bbox": [ + 0.219, + 0.546, + 0.785, + 0.587 + ], + "angle": 0, + "content": "11. Esser, P., Chiu, J., Atighehchian, P., Granskog, J., Germanidis, A.: Structure and content-guided video synthesis with diffusion models. In: Proceedings of the IEEE/CVF International Conference on Computer Vision. pp. 7346-7356 (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.219, + 0.588, + 0.785, + 0.615 + ], + "angle": 0, + "content": "12. Esser, P., Rombach, R., Ommer, B.: Taming transformers for high-resolution image synthesis (2021)" + }, + { + "type": "ref_text", + "bbox": [ + 0.219, + 0.616, + 0.785, + 0.644 + ], + "angle": 0, + "content": "13. Fu, T.J., Hu, W., Du, X., Wang, W.Y., Yang, Y., Gan, Z.: Guiding instruction-based image editing via multimodal large language models (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.219, + 0.645, + 0.785, + 0.685 + ], + "angle": 0, + "content": "14. Fujitake, M., Sugimoto, A.: Video representation learning through prediction for online object detection. In: Proceedings of the IEEE/CVF Winter Conference on Applications of Computer Vision. pp. 530-539 (2022)" + }, + { + "type": "ref_text", + "bbox": [ + 0.219, + 0.686, + 0.785, + 0.728 + ], + "angle": 0, + "content": "15. Ge, S., Nah, S., Liu, G., Poon, T., Tao, A., Catanzaro, B., Jacobs, D., Huang, J.B., Liu, M.Y., Balaji, Y.: Preserve your own correlation: A noise prior for video diffusion models (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.219, + 0.729, + 0.785, + 0.756 + ], + "angle": 0, + "content": "16. Girdhar, R., Ramanan, D.: Cater: A diagnostic dataset for compositional actions and temporal reasoning (2020)" + }, + { + "type": "ref_text", + "bbox": [ + 0.219, + 0.757, + 0.785, + 0.798 + ], + "angle": 0, + "content": "17. Goodfellow, I., Pouget-Abadie, J., Mirza, M., Xu, B., Warde-Farley, D., Ozair, S., Courville, A., Bengio, Y.: Generative adversarial networks. Communications of the ACM 63(11), 139-144 (2020)" + }, + { + "type": "ref_text", + "bbox": [ + 0.219, + 0.799, + 0.785, + 0.84 + ], + "angle": 0, + "content": "18. Gu, Y., Yang, J., Usuyama, N., Li, C., Zhang, S., Lungren, M.P., Gao, J., Poon, H.: Biomedjourney: Counterfactual biomedical image generation by instruction-learning from multimodal patient journeys. arXiv preprint arXiv:2310.10765 (2023)" + }, + { + "type": "list", + "bbox": [ + 0.219, + 0.179, + 0.785, + 0.84 + ], + "angle": 0, + "content": null + } + ], + [ + { + "type": "page_number", + "bbox": [ + 0.218, + 0.116, + 0.236, + 0.127 + ], + "angle": 0, + "content": "16" + }, + { + "type": "header", + "bbox": [ + 0.272, + 0.115, + 0.388, + 0.127 + ], + "angle": 0, + "content": "H. Kandala et al." + }, + { + "type": "ref_text", + "bbox": [ + 0.218, + 0.147, + 0.785, + 0.175 + ], + "angle": 0, + "content": "19. He, Y., Yang, T., Zhang, Y., Shan, Y., Chen, Q.: Latent video diffusion models for high-fidelity long video generation (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.218, + 0.176, + 0.785, + 0.202 + ], + "angle": 0, + "content": "20. Hertz, A., Mokady, R., Tenenbaum, J., Aberman, K., Pritch, Y., Cohen-Or, D.: Prompt-to-prompt image editing with cross attention control (2022)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.203, + 0.743, + 0.216 + ], + "angle": 0, + "content": "21. Ho, J., Jain, A., Abbeel, P.: Denoising diffusion probabilistic models (2020)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.217, + 0.785, + 0.242 + ], + "angle": 0, + "content": "22. Ho, J., Salimans, T.: Classifier-free diffusion guidance. arXiv preprint arXiv:2207.12598 (2022)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.243, + 0.785, + 0.27 + ], + "angle": 0, + "content": "23. Ho, J., Salimans, T., Gritsenko, A., Chan, W., Norouzi, M., Fleet, D.J.: Video diffusion models (2022)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.271, + 0.785, + 0.297 + ], + "angle": 0, + "content": "24. Hong, W., Ding, M., Zheng, W., Liu, X., Tang, J.: Cogvideo: Large-scale pretraining for text-to-video generation via transformers (2022)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.298, + 0.785, + 0.324 + ], + "angle": 0, + "content": "25. Hu, Y., Luo, C., Chen, Z.: Make it move: Controllable image-to-video generation with text descriptions (2022)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.325, + 0.785, + 0.365 + ], + "angle": 0, + "content": "26. Kang, M., Zhu, J.Y., Zhang, R., Park, J., Shechtman, E., Paris, S., Park, T.: Scaling up gans for text-to-image synthesis. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 10124-10134 (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.366, + 0.785, + 0.406 + ], + "angle": 0, + "content": "27. Karras, T., Laine, S., Aila, T.: A style-based generator architecture for generative adversarial networks. In: Proceedings of the IEEE/CVF conference on computer vision and pattern recognition. pp. 4401-4410 (2019)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.407, + 0.785, + 0.46 + ], + "angle": 0, + "content": "28. Kawar, B., Zada, S., Lang, O., Tov, O., Chang, H., Dekel, T., Mosseri, I., Irani, M.: Imagic: Text-based real image editing with diffusion models. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 6007-6017 (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.461, + 0.69, + 0.474 + ], + "angle": 0, + "content": "29. Kingma, D.P., Welling, M.: Auto-encoding variational bayes (2022)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.475, + 0.785, + 0.5 + ], + "angle": 0, + "content": "30. Lee, J., Lee, J., Lee, S., Yoon, S.: Mutual suppression network for video prediction using disentangled features (2019)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.501, + 0.785, + 0.542 + ], + "angle": 0, + "content": "31. Li, H., Yang, Y., Chang, M., Chen, S., Feng, H., Xu, Z., Li, Q., Chen, Y.: Srdiff: Single image super-resolution with diffusion probabilistic models. Neurocomputing 479, 47-59 (2022)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.543, + 0.785, + 0.569 + ], + "angle": 0, + "content": "32. Li, Y., Liu, H., Wu, Q., Mu, F., Yang, J., Gao, J., Li, C., Lee, Y.J.: Gligen: Open-set grounded text-to-image generation (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.57, + 0.785, + 0.596 + ], + "angle": 0, + "content": "33. Li, Y., Song, Y., Cao, L., Tetreault, J., Goldberg, L., Jaimes, A., Luo, J.: Tgif: A new dataset and benchmark on animated gif description (2016)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.597, + 0.785, + 0.623 + ], + "angle": 0, + "content": "34. Liu, W., Luo, W., Lian, D., Gao, S.: Future frame prediction for anomaly detection - a new baseline (2018)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.624, + 0.785, + 0.65 + ], + "angle": 0, + "content": "35. Lu, C., Zhou, Y., Bao, F., Chen, J., Li, C., Zhu, J.: Dpm-solver: A fast ode solver for diffusion probabilistic model sampling in around 10 steps (2022)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.651, + 0.785, + 0.677 + ], + "angle": 0, + "content": "36. Meng, C., He, Y., Song, Y., Song, J., Wu, J., Zhu, J.Y., Ermon, S.: Sdedit: Guided image synthesis and editing with stochastic differential equations (2022)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.678, + 0.785, + 0.704 + ], + "angle": 0, + "content": "37. Ni, B., Peng, H., Chen, M., Zhang, S., Meng, G., Fu, J., Xiang, S., Ling, H.: Expanding language-image pretrained models for general video recognition (2022)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.705, + 0.785, + 0.731 + ], + "angle": 0, + "content": "38. Ni, H., Shi, C., Li, K., Huang, S.X., Min, M.R.: Conditional image-to-video generation with latent flow diffusion models (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.732, + 0.785, + 0.759 + ], + "angle": 0, + "content": "39. Oh, J., Guo, X., Lee, H., Lewis, R., Singh, S.: Action-conditional video prediction using deep networks in atari games (2015)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.76, + 0.785, + 0.785 + ], + "angle": 0, + "content": "40. Oliu, M., Selva, J., Escalera, S.: Folded recurrent neural networks for future video prediction (2018)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.786, + 0.785, + 0.812 + ], + "angle": 0, + "content": "41. van den Oord, A., Kalchbrenner, N., Kavukcuoglu, K.: Pixel recurrent neural networks (2016)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.813, + 0.785, + 0.84 + ], + "angle": 0, + "content": "42. van den Oord, A., Kalchbrenner, N., Vinyals, O., Espeholt, L., Graves, A., Kavukcuoglu, K.: Conditional image generation with pixelCNN decoders (2016)" + }, + { + "type": "list", + "bbox": [ + 0.217, + 0.147, + 0.785, + 0.84 + ], + "angle": 0, + "content": null + } + ], + [ + { + "type": "header", + "bbox": [ + 0.371, + 0.115, + 0.732, + 0.128 + ], + "angle": 0, + "content": "Pix2Gif: Motion-Guided Diffusion for GIF Generation" + }, + { + "type": "page_number", + "bbox": [ + 0.768, + 0.116, + 0.786, + 0.127 + ], + "angle": 0, + "content": "17" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.147, + 0.785, + 0.175 + ], + "angle": 0, + "content": "43. van den Oord, A., Vinyals, O., Kavukcuoglu, K.: Neural discrete representation learning (2018)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.175, + 0.785, + 0.215 + ], + "angle": 0, + "content": "44. Radford, A., Metz, L., Chintala, S.: Unsupervised representation learning with deep convolutional generative adversarial networks. arXiv preprint arXiv:1511.06434 (2015)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.217, + 0.785, + 0.242 + ], + "angle": 0, + "content": "45. Rombach, R., Blattmann, A., Lorenz, D., Esser, P., Ommer, B.: High-resolution image synthesis with latent diffusion models (2022)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.243, + 0.785, + 0.284 + ], + "angle": 0, + "content": "46. Ruiz, N., Li, Y., Jampani, V., Pritch, Y., Rubinstein, M., Aberman, K.: Dreambooth: Fine tuning text-to-image diffusion models for subject-driven generation (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.285, + 0.785, + 0.338 + ], + "angle": 0, + "content": "47. Sahara, C., Chan, W., Saxena, S., Li, L., Whang, J., Denton, E.L., Ghasemipour, K., Gontijo Lopes, R., Karagol Ayan, B., Salimans, T., et al.: Photorealistic text-to-image diffusion models with deep language understanding. Advances in Neural Information Processing Systems 35, 36479-36494 (2022)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.339, + 0.785, + 0.379 + ], + "angle": 0, + "content": "48. Shi, X., Chen, Z., Wang, H., Yeung, D.Y., kin Wong, W., chun Woo, W.: Convolutional LSTM network: A machine learning approach for precipitation nowcasting (2015)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.38, + 0.785, + 0.406 + ], + "angle": 0, + "content": "49. Simonyan, K., Zisserman, A.: Very deep convolutional networks for large-scale image recognition (2015)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.407, + 0.785, + 0.447 + ], + "angle": 0, + "content": "50. Singer, U., Polyak, A., Hayes, T., Yin, X., An, J., Zhang, S., Hu, Q., Yang, H., Ashual, O., Gafni, O., Parikh, D., Gupta, S., Taigman, Y.: Make-a-video: Text-to-video generation without text-video data (2022)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.448, + 0.732, + 0.46 + ], + "angle": 0, + "content": "51. Song, J., Meng, C., Ermon, S.: Denoising diffusion implicit models (2022)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.461, + 0.785, + 0.487 + ], + "angle": 0, + "content": "52. Soomro, K., Zamir, A.R., Shah, M.: Ucf101: A dataset of 101 human actions classes from videos in the wild (2012)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.488, + 0.785, + 0.514 + ], + "angle": 0, + "content": "53. Srivastava, N., Mansimov, E., Salakhutdinov, R.: Unsupervised learning of video representations using lstms (2016)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.515, + 0.785, + 0.541 + ], + "angle": 0, + "content": "54. Unterthiner, T., van Steenkiste, S., Kurach, K., Marinier, R., Michalski, M., Gelly, S.: Fvd: A new metric for video generation (2019)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.542, + 0.785, + 0.582 + ], + "angle": 0, + "content": "55. Villegas, R., Babaeizadeh, M., Kindermans, P.J., Moraldo, H., Zhang, H., Saffar, M.T., Castro, S., Kunze, J., Erhan, D.: Phenaki: Variable length video generation from open domain textual description (2022)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.583, + 0.785, + 0.609 + ], + "angle": 0, + "content": "56. Villegas, R., Yang, J., Hong, S., Lin, X., Lee, H.: Decomposing motion and content for natural video sequence prediction (2018)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.61, + 0.785, + 0.637 + ], + "angle": 0, + "content": "57. Voleti, V., Jolicoeur-Martineau, A., Pal, C.: Mcvd: Masked conditional video diffusion for prediction, generation, and interpolation (2022)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.637, + 0.785, + 0.677 + ], + "angle": 0, + "content": "58. Wang, C., Gu, J., Hu, P., Xu, S., Xu, H., Liang, X.: Dreamvideo: High-fidelity image-to-video generation with image retention and text guidance. arXiv preprint arXiv:2312.03018 (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.678, + 0.785, + 0.718 + ], + "angle": 0, + "content": "59. Wang, Y., Chen, X., Ma, X., Zhou, S., Huang, Z., Wang, Y., Yang, C., He, Y., Yu, J., Yang, P., et al.: Lavie: High-quality video generation with cascaded latent diffusion models. arXiv preprint arXiv:2309.15103 (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.719, + 0.785, + 0.746 + ], + "angle": 0, + "content": "60. Weissenborn, D., Täckström, O., Uszkoreit, J.: Scaling autoregressive video models (2020)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.746, + 0.785, + 0.772 + ], + "angle": 0, + "content": "61. Wu, C., Huang, L., Zhang, Q., Li, B., Ji, L., Yang, F., Sapiro, G., Duan, N.: Godiva: Generating open-domain videos from natural descriptions (2021)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.773, + 0.785, + 0.8 + ], + "angle": 0, + "content": "62. Wu, C., Liang, J., Ji, L., Yang, F., Fang, Y., Jiang, D., Duan, N.: Niwa: Visual synthesis pre-training for neural visual world creation (2021)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.8, + 0.785, + 0.84 + ], + "angle": 0, + "content": "63. Xing, J., Xia, M., Zhang, Y., Chen, H., Wang, X., Wong, T.T., Shan, Y.: Dynamicultural: Animating open-domain images with video diffusion priors. arXiv preprint arXiv:2310.12190 (2023)" + }, + { + "type": "list", + "bbox": [ + 0.217, + 0.147, + 0.785, + 0.84 + ], + "angle": 0, + "content": null + } + ], + [ + { + "type": "page_number", + "bbox": [ + 0.218, + 0.116, + 0.236, + 0.127 + ], + "angle": 0, + "content": "18" + }, + { + "type": "header", + "bbox": [ + 0.272, + 0.115, + 0.388, + 0.127 + ], + "angle": 0, + "content": "H. Kandala et al." + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.147, + 0.785, + 0.189 + ], + "angle": 0, + "content": "64. Xu, J., Mei, T., Yao, T., Rui, Y.: Msr-vtt: A large video description dataset for bridging video and language. In: Proceedings of the IEEE conference on computer vision and pattern recognition. pp. 5288-5296 (2016)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.19, + 0.785, + 0.217 + ], + "angle": 0, + "content": "65. Yang, Z., Wang, J., Gan, Z., Li, L., Lin, K., Wu, C., Duan, N., Liu, Z., Liu, C., Zeng, M., Wang, L.: Reco: Region-controlled text-to-image generation (2022)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.218, + 0.785, + 0.259 + ], + "angle": 0, + "content": "66. Yu, J., Xu, Y., Koh, J.Y., Luong, T., Baid, G., Wang, Z., Vasudevan, V., Ku, A., et al.: Scaling autoregressive models for content-rich text-to-image generation. arXiv preprint arXiv:2206.10789 2(3), 5 (2022)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.26, + 0.785, + 0.313 + ], + "angle": 0, + "content": "67. Yu, L., Cheng, Y., Sohn, K., Lezama, J., Zhang, H., Chang, H., Hauptmann, A.G., Yang, M.H., Hao, Y., Essa, I., et al.: Magvit: Masked generative video transformer. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 10459-10469 (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.315, + 0.785, + 0.396 + ], + "angle": 0, + "content": "68. Yu, L., Shi, B., Pasunuru, R., Muller, B., Golovneva, O., Wang, T., Babu, A., Tang, B., Karrer, B., Sheynin, S., Ross, C., Polyak, A., Howes, R., Sharma, V., Xu, P., Tamoyan, H., Ashual, O., Singer, U., Li, S.W., Zhang, S., James, R., Ghosh, G., Taigman, Y., Fazel-Zarandi, M., Celikyilmaz, A., Zettlemoyer, L., Aghajanyan, A.: Scaling autoregressive multi-modal models: Pretraining and instruction tuning (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.398, + 0.785, + 0.438 + ], + "angle": 0, + "content": "69. Zhang, L., Rao, A., Agrawala, M.: Adding conditional control to text-to-image diffusion models. In: Proceedings of the IEEE/CVF International Conference on Computer Vision. pp. 3836-3847 (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.439, + 0.785, + 0.481 + ], + "angle": 0, + "content": "70. Zhang, S., Wang, J., Zhang, Y., Zhao, K., Yuan, H., Qin, Z., Wang, X., Zhao, D., Zhou, J.: I2vgen-xl: High-quality image-to-video synthesis via cascaded diffusion models. arXiv preprint arXiv:2311.04145 (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.481, + 0.785, + 0.508 + ], + "angle": 0, + "content": "71. Özgün Çiçek, Abdulkadir, A., Lienkamp, S.S., Brox, T., Ronneberger, O.: 3d u-net: Learning dense volumetric segmentation from sparse annotation (2016)" + }, + { + "type": "list", + "bbox": [ + 0.217, + 0.147, + 0.785, + 0.508 + ], + "angle": 0, + "content": null + } + ] +] \ No newline at end of file diff --git a/2024/Pix2Gif_ Motion-Guided Diffusion for GIF Generation/b48548a6-50ce-46d1-9de4-333ad980d37c_origin.pdf b/2024/Pix2Gif_ Motion-Guided Diffusion for GIF Generation/b48548a6-50ce-46d1-9de4-333ad980d37c_origin.pdf new file mode 100644 index 0000000000000000000000000000000000000000..95fe67172a1b55be75acd8ee7dbdc973f84d52e6 --- /dev/null +++ b/2024/Pix2Gif_ Motion-Guided Diffusion for GIF Generation/b48548a6-50ce-46d1-9de4-333ad980d37c_origin.pdf @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:9a6751faacc45a0fb7757816f79e1fe7a763a3e2a063646e06d50ca21629de38 +size 14104376 diff --git a/2024/Pix2Gif_ Motion-Guided Diffusion for GIF Generation/full.md b/2024/Pix2Gif_ Motion-Guided Diffusion for GIF Generation/full.md new file mode 100644 index 0000000000000000000000000000000000000000..8f58afebd89ccc54bfda3966ec42d5b1ddc56378 --- /dev/null +++ b/2024/Pix2Gif_ Motion-Guided Diffusion for GIF Generation/full.md @@ -0,0 +1,284 @@ +# Pix2Gif: Motion-Guided Diffusion for GIF Generation + +Hitsh Kandala1, Jianfeng Gao1, and Jianwei Yang1 + +Microsoft Research +khitesh2000@gmail.com,{jfgao, jianwei.yang}@microsoft.com + +![](images/931773cd42f45d093782adbb62d46c47f9065e3b654c6ab8741a95653ba3bc9c.jpg) +Fig. 1: Our model creates distinct frames based on the provided source image and caption, adjusting according to different levels of motion magnitude (optical flow magnitude) specified in the input conditions. It stands well for both high spatial quality and temporal consistency. + +Abstract. We present *Pix2Gif*, a motion-guided diffusion model for image-to-GIF (video) generation. We tackle this problem differently by formulating the task as an image translation problem steered by text and motion magnitude prompts, as shown in Fig. 1. To ensure that the model adheres to motion guidance, we propose a new motion-guided warping module to spatially transform the features of the source image conditioned on the two types of prompts. Furthermore, we introduce a perceptual loss to ensure the transformed feature map remains within the same space as the target image, ensuring content consistency and coherence. In preparation for the model training, we meticulously curated data by extracting coherent image frames from the TGIF video-caption dataset, which provides rich information about the temporal changes of subjects. After pretraining, we apply our model in a zero-shot manner to a number of video datasets. Extensive qualitative and quantitative experiments demonstrate the effectiveness of our model - it not only captures the semantic prompt from text but also the spatial ones from motion guidance. We train all our models using a single node of $16 \times \mathrm{V}100$ GPUs. + +Keywords: Diffusion $\cdot$ image-to-video generation $\cdot$ image-to-image translation + +# 1 Introduction + +Visual content generation has been significantly advanced by the huge progress of diffusion models [21, 22, 35, 51]. Recently, the development of latent diffusion models (LDMs) [45] has led us to a new quality level of generated images. It has inspired a lot of works for customized and controllable image generation [32, 46, 65, 69], and fine-grained image editing [6, 20, 28, 36]. + +In this work, we focus on converting a single image to an animated Graphics Interchange Format (GIF), which is valuable for design yet under-explored. Despite the absence of image-to-GIF generation models, diffusion-based video generation has emerged as a hot topic recently. Compared with text-to-image generation, however, text-to-video generation requires not only high quality for individual frames but also visual consistency and temporal coherence across frames. To achieve this goal, existing works expand the LDMs to video diffusion models (VDMs) by either inflating the 2D CNNs in LDMs to 3D ones [23] or introducing an additional temporal attention layer to bridge the diffusion for each frame [5,15,19,50,59]. In addition to text prompts, a few recent works also explored the way of using images or other prompts to make the video generation model more customizable and controllable [11,38]. However, due to the high cost of VDMs to generate a sequence of video frames in one run, most (if not all) of these works require a compromise of reducing the resolution of generated frames $(64 \times 64$ typically), and the usage of extra super-resolution diffusion models for upscaling [31,47]. Moreover, since these methods use the temporal attention layers to model the cross-frame dependency implicitly, it is quite hard for them to preserve good controllability of the frame-to-frame temporal dynamics in a fine-grained manner. + +Given that animated GIF usually contains less number of frames and requires more specializations, we take a different strategy and formulate the image-to-GIF generation as an image translation process. To decouple the generation of visual contents and temporal dynamics, we further introduce a motion flow magnitude as extra guidance in addition to image and text prompts. Unlike the aforementioned works, our model takes one or more history frames as the condition and produces only one future frame at once. This brings some unique advantages: (i) simplicity - our model can be purely built on top of LDMs and trained end-to-end with high resolution, without any cascaded diffusion processes for upscaling. (ii) controllability - we could inject detailed and different text and motion prompts at each time step for generating a frame, which gains much better controllability of the model. Our work is inspired by a line of canonical works for future frame prediction [40, 48, 56]. However, due to the lack of a powerful image-generation engine, these works fail to produce high-quality results and can only be applied to specific video domains [16, 30, 52]. Moreover, they cannot support other types of prompts or conditions than the history frames. To address this problem, + +we exploit a modern diffusion-based pipeline. More specifically, we follow text-conditioned image editing approaches (e.g., InstructPix2Pix [6]), and propose a new temporal image editing to produce future frames given history frames. To train the model, we curate a new training dataset based on TGIF [33] by extracting frames and calculating the magnitude of optical flow between them. We then selected an appropriate range of the optical flow magnitude and sampled frame pairs from each GIF in a manner that ensures diversity. In the end, we train our diffusion model called Pix2Gif, which can generate high-quality animated GIF consisting of multiple frames, given a single image and text and motion magnitude prompts. In summary, our main contributions are: + +- We are the first to explore an image-to-image translation formula for generating animated GIFs from an image, guided by a text prompt and motion magnitude. +- We propose a flow-based warping module with a perceptual loss in the diffusion process that takes motion magnitude as input and controls the temporal dynamics and consistency between future frames and the initial ones. +- We curate a new dataset, comprised of 78,692 short GIF clips for training, and 10,546 for evaluation. The new dataset covers a variety of visual domains. +- Quantitative and qualitative results demonstrate the effectiveness of our proposed method for generating visually consistent coherent GIFs from a single image, and it can be generalized to a wide range of visual domains. + +# 2 Prior Work + +Image and video generation has been a long-standing problem in the community. It can be tackled by different approaches, which can be categorized into four groups: generative adversarial networks (GANs) [17,26,27,44], transformer-based autoregressive decoding [10,12,41,42,60,66], masked image modeling [7,8,55,67]. Most of the recent works exploited diffusion models for image generation given their high-quality outputs and huge open-source supports [45, 47]. Recently, a number of works have extended the text-to-image generation model into image translation or editing models [6,20,28,36,69] or video generation models [5,15, 23,50,57]. Below we provide a brief overview of the related diffusion-based image and video generation methods. + +Image-to-Image Translation. Diffusion-based image-to-image generation has drawn increasing attention. Different from text-to-image generation, it takes an image as input and edits its contents following the text instructions while keeping the unrelated parts unchanged. SDEdit [36] and ILVR [9] are two pioneering works that impose reference image conditions to an existing latent diffusion model for controllable image generation. Later on, to conduct local edits, the authors in [2] proposed blended latent diffusion to steer the diffusion process with a user-specified mask, where the pixels out of the mask remain the same + +as the input image while the region inside is edited following the textual description. Instead of manipulating the image space, Prompt2Prompt [20] proposed to edit the image by manipulating the textual context (e.g., swapping or adding words.) to which the latent diffusion model cross-attends. However, this method requires forwarding a text-to-image generation process to obtain the cross-attention maps, and thus cannot be applied to real images. Imagic [28] proposed to blend the embeddings of a real image with the textual context embedding so that the generated image obeys both the image and text conditions. + +All the aforementioned works leverage a frozen latent diffusion model and control the generation with modified text or image prompts. To enable arbitrary image editing, InstructPix2Pix [6] proposed to finetune the LDM to follow user instructions that precisely convey the user intents, e.g., "change the cat to dog". The model is trained by a synthetic dataset consisting of triplets $\langle \text{image}_{\text{src}}, \text{instruction}, \text{image}_{\text{tgt}} \rangle$ . The resulting model could allow both realistic and generated images and support arbitrary language instructions. Some other works also exploit a similar way to train the model to follow instructions [18,68]. To further enhance the language understanding, MGIE [13] exploited a large multimodal model to produce a more comprehensive textual context for the instructed image editing. + +In this work, we employ the image-to-image translation pipeline and are the first to formulate a GIF generation as an image translation problem. Given a reference image, the goal is to generate a realistic future frame following a textual instruction. Therefore, the focus is on how to perform temporal rather than spatial editing on a source image. When the process rolls out, it gradually gives a sequence of frames. + +Conditioned Video Generation. Speaking of the high-level goal, our work resembles conditioned video generation. For video generation, a conventional way is inflating the 2D U-Net used in LDM to 3D U-Net [71] by replacing the 2D convolution layers with 3D ones. Likewise, a similar strategy is taken in [5,15,19,50,59], but with a slight difference in that they use interleaved spatial and temporal attention layers in the U-Net. Due to the high cost of generating a sequence of video frames in one shot, the output videos usually have a resolution as low as $64 \times 64$ . To attain high-resolution videos, these methods need to use one or more super-resolution diffusion models [31,47] to upscale the resolution by 4 or 8 times. To accelerate the training, a pre-trained text-to-image LDM is usually used to initialize these models. Adding spatial-temporal modules is also a commonly used strategy for autoregressive models [24, 55, 61, 62]. Similarly, both [62] and [24] exploit a pretrained autoregressive image generation model as the starting point. In [55], however, the authors introduce and pretrained a new video encoder, which is then used to train a masked video decoder. + +Besides text-to-video generation, using images as the condition for video generation draws increasing attention. On one hand, once a text-to-video generation is trained, it can be further finetuned for image-to-video generation [4]. In [11], the authors introduce additional structural conditions (e.g., depth maps) for + +more controllable video generation. Alternatively in [38], a latent flow diffusion model is introduced for image-conditioned video generation by explicitly generating a sequence of optimal flows and masks as the guidance. On the other hand, a few concurrent works to ours directly approach image-to-video generation on top of video diffusion models [58, 63, 70]. All these works share a similar spirit to text-to-video generation models but add additional images as the reference. + +Our method uses a diffusion model but differs from all the aforementioned methods in that we reformulate video generation as a frame-to-frame translation problem based on the history frames. As [6] suggests image-to-image translation can maintain a decent visual consistency. In addition, we also introduce a motion flow magnitude as another condition to explicitly control the temporal dynamics. + +Future Frame Prediction. Future frame prediction or forecasting [40, 48, 56] has been a long-standing problem before the prevalence of diffusion models. It has been used as an anomaly detection approach by comparing the observed frame and the predicted ones [3, 34] and video representation learning for various downstream tasks [14]. For these problems, a recurrent network such as LSTM [53], ConvLSTM [39, 56] or 3D-CNN [1] is usually used as the model architecture, and GAN [17] or Variational Autoencoder (VAE) [29] is used as the learning objectives. With the emergence of VQ-VAE [43], the authors in [25] exploited axial transformer blocks to chain the encoder-decoder for autoregressive next-frame prediction. In [57], the authors proposed masked conditional video diffusion to unify different tasks of video prediction, generation and interpolation. Nevertheless, all of these models are trained on domain-specific video datasets such as MovingMNIST [30], CATER [16] and UCF-101 [52], etc, far from being a generic video generation model. + +Our work takes inspiration from future frame prediction methods but proposes a simpler yet effective strategy by formulating it as an image-to-image translation problem. Furthermore, our model simultaneously takes image, text and motion magnitude as the guidance for better controllability. To attain a model as general as possible, we curate a new training dataset covering a wide range of domains. Without any further dataset-specific finetuning, our model achieves plausible video generation results as shown in Fig. 1. + +# 3 Method + +Our goal is to generate GIFs, given an initial frame, a motion description, and a measure of optical flow. We frame this as an image-to-image translation problem based on latent diffusion. We first explain how we created our training dataset in Sec. 3.1, then outline our model's principles and training strategy in Sec. 3.2. Next, we delve into the specifics of our proposed model, explaining its various components in Sec. 3.3. Finally, we focus on the loss functions used to train our model in Sec. 3.4. + +![](images/8d0c160964518d0d443d4947b2090224c98840295ce1819c05b58b0e6d0e1ce5.jpg) +Fig. 2: The three step process of curating the TGIF dataset. Starting from extracting frames $A$ to restricting the range of optical flow $B$ and then maintaining the diversity of pairs $C$ . + +![](images/ea8b16389739e85bdecfa9ac8571d8be904b50cfadaf607d471bc3376e4ff0c5.jpg) +Fig.3: $\text{Pix2Gif}$ model pipeline. We propose an end-to-end network where the inputs are encoded by $\mathcal{E}$ , CLIP and $\mathcal{M}$ to output $\mathcal{E}(c_L), c_T$ and $c_M$ respectively, which then goes into $\mathcal{W}$ to form the conditioning input for LDM. + +# 3.1 Dataset + +We used the Tumblr GIF (TGIF) dataset [33], which predominantly consists of human-centric animated GIFs described by captions. The dataset features a range of GIFs with varied movements over 1-3 seconds. + +The curation process, as shown in Fig. 2, involved extracting frames from all GIFs and calculating the optical flow between all possible frame pairs. The number of extracted frames from each GIF varied, with an average of about 41 frames. The optical flow histogram calculated between all frames ranged from [0, 200]. We selected the [2, 20] range, capturing small yet significant motion and excluding pairs with drastic changes. Despite the restricted range, we still had a significant number of training pairs. To avoid model overfitting and maintain diversity, we randomly selected a minimum of 10 pairs or the number of pairs within the restricted range from each GIF. This approach resulted in a final dataset with nearly equal representation of all values within the selected range. The final dataset contained 783,184 training pairs and 105,041 validation pairs. Each data point consisted of a pair of frames from the same GIF, the corresponding caption, and the calculated optical flow between the frames. + +# 3.2 Preliminary: Instructed Image Editing + +Our model is fundamentally grounded in the latent diffusion models (LDMs) for image generation and editing [6, 45]. More specifically, we build upon InstructPix2Pix [6] by framing our objective in the context of an instructed image-to-image translation task. Given an image $x$ , the forward diffusion procedure introduces noise to the encoded latent $z$ , thereby producing a noisy latent vector $z_{t}$ . This process is carried out over $T$ timesteps, with each timestep $t \in \{1, \dots, T\}$ seeing an increment in the noise level until it culminates into a random noise $n$ . A network $e_{\theta}$ is trained by minimizing the following latent diffusion objective to predict noise existing in the noisy latent $z_{t}$ , considering factors in image conditioning $c_{I}$ and textual instruction $c_{T}$ : + +$$ +L _ {L D M} = \mathbb {E} _ {\mathcal {E} (x), \mathcal {E} (c _ {I}), c _ {T}, \epsilon \sim \mathcal {N} (0, 1), t} \left[ | | \epsilon - \epsilon_ {\theta} \left(z _ {t}, t, \mathcal {E} \left(c _ {I}\right), c _ {T}\right) | | _ {2} ^ {2} \right] \tag {1} +$$ + +where $\mathcal{E}$ is the VQ-VAE encoder that transforms the images from pixel space to discrete latent space. To facilitate image conditioning, $z_{t}$ and $\mathcal{E}(c_I)$ are concatenated and then fed into a convolutional layer. The model is trained for conditional and unconditional denoising, given the image and caption condition individually or collectively. + +# 3.3 Our Model: Pix2Gif + +We build our model similar to InstructPix2Pix and frame our objective in the context of a text-instructed and motion-guided temporal editing problem. Compared with the original InstructPix2Pix pipeline, the main innovation is the newly introduced motion-based warping module. The overall model pipeline is shown in Fig. 3. + +Our model takes three inputs: an image, a text instruction, and a motion magnitude. These inputs are fed into the model through two pathways - once through the diffusion model directly, and again through the warping module, which will be discussed in Sec. 3.3 and Sec. 3.3. When passed through the caption, we add the phrase "The optical flow is _. to the original caption. The flow input is then appended at the end in a word format rather than a numerical one, as the CLIP model tends to assign higher similarity scores to word forms than to numerical representations of numbers for the same image. Finally, our model is trained by minimizing the following loss function: + +$$ +L _ {L D M} ^ {\prime} = \mathbb {E} _ {\mathcal {E} (x), \mathcal {E} (c _ {I}), c _ {T}, c _ {M}, \epsilon \sim \mathcal {N} (0, 1), t} \left[ | | \epsilon - \epsilon_ {\theta} (z _ {t}, t, \mathcal {E} (c _ {I}), c _ {T}, c _ {M}) | | _ {2} ^ {2} \right] \tag {2} +$$ + +where $c_{M}$ is the motion condition. The altered caption is processed via the pretrained CLIP model to yield $c_{T}$ , while the output of $\mathcal{M}$ gives us $c_{M}$ . These two conditions are then added linearly, serving as the conditioning input for both the Warping Module $\mathcal{W}$ (discussed in Sec. 3.3) and the Latent Diffusion Model LDM (referenced in Sec. 3.2). + +![](images/d5aa299c7c9f03e600a0952a7ba0d5b45d37c8c7c14f76f9a2734e53197ddd82.jpg) +Fig. 4: Deep dive into the Warping Module $\mathcal{W}$ . It comprises of three units: $\mathcal{I}_M$ , $\mathcal{F}_{Net}$ and $\mathcal{W}_{Net}$ . + +Motion Embedding Layer In traditional conditional diffusion models like [6], text prompts are often enough for image generation. We initially passed the motion input indirectly through the prompt. However, this divided the model's attention on a single caption token, which is problematic when the main caption stays the same but the motion input changes. To let the model focus on the motion input, we included an embedding layer that converts the motion input into an integer and selects an embedding vector. This vector is then duplicated and concatenated with itself to generate $c_{M}$ , which when combined with the caption embedding $c_{T}$ , provides the conditioning input $c_{L} = c_{T} + c_{M}$ for both warping module $\mathcal{W}$ and LDM. + +Warping Module One of the main components of Pix2Gif is the Warping Module $\mathcal{W}$ . As illustrated in Fig. 4, it technically comprises two networks: the FlowNet $(\mathcal{F}_{Net})$ and the WarpNet $(\mathcal{W}_{Net})$ . Ordinarily, the computation of optical flow involves two images. However, in this case, we initially have only one image - the source image - and that too in the latent domain. Thus, our goal is to learn the optical flow utilizing just one latent image. This is achieved via $\mathcal{F}_{Net}$ , conditioned on $c_{L}$ , which guides it to generate a flow feature map in the intended direction with the hint of text and motion prompts. This condition is processed by the Injection Module $(\mathcal{I}_{\mathcal{M}})$ , a compact encoder designed to make $c_{L}$ compatible for concatenation with one of the intermediate feature maps near the end of the network. This configuration enables $\mathcal{F}_{Net}$ to independently learn high-level features, which are then guided in the desired direction with the introduction of $c_{L}$ . The architecture of $\mathcal{F}_{Net}$ resembles that of UNet, producing an output with a fixed channel of 2 to capture changes in the horizontal and vertical components. This optical flow feature map $F = \mathcal{F}_{Net}(\mathcal{E}(c_{I}), \mathcal{I}_{\mathcal{M}}(c_{L}))$ , along with the source latent $(\mathcal{E}(c_{I}))$ , is then processed through $\mathcal{W}_{Net}$ to yield a Fischer map $z_{W} = \mathcal{W}_{Net}(\mathcal{E}(c_{I}), F)$ . This transformation is learned more efficiently and abstractly in the latent space than in the pixel space. + +Table 1: Quantitative comparison with state-of-the-art image-text-to-video generation models for the zero-shot setting. + +
MethodUCF-101 [52]MSR-VTT [64]
FVD ↓CLIPSim ↑PIC ↑FVD ↓CLIPSim ↑PIC ↑
I2VGen-XL [70]563.120.28650.6329278.620.22720.6018
DynamiCrafter [63]527.060.27960.6307271.630.26020.6135
Pix2Gif (Ours)285.020.28150.8763168.690.25730.8521
+ +# 3.4 Losses + +Our model incorporates two different types of losses. The first type is the standard L2 loss Eq. (2), which is utilized by the stable diffusion model and talked about in Sec. 3.2 + +The second loss type in our model is the perceptual loss, which is calculated by comparing the latent features of the image condition $\mathcal{E}(c_I)$ and the warped image $z_W$ . This is implemented using a pre-trained VGG network [49], modified to accommodate 4 channels instead of the standard 3. The modification involves averaging the weights from the first three channels to initialize the fourth. The perceptual loss, $L_{p}$ , is defined as: + +$$ +L _ {p} \left(\mathcal {E} \left(c _ {I}\right), z _ {W}\right) = \sum_ {k} \lambda_ {k} \left\| \phi_ {k} \left(\mathcal {E} \left(c _ {I}\right)\right) - \phi_ {k} \left(z _ {W}\right) \right\| ^ {2} \tag {3} +$$ + +Here, $\phi_k(.)$ be the feature map of the k-th layer of the VGG network, $||.||$ denotes the Frobenius norm, and $\lambda_{k}$ is a weighting factor. This loss ensures that the warped image maintains high-level features like edges, textures, and object types, making the images more perceptually and semantically similar and preserving the overall structure of the source image. + +In conclusion, the total loss function, denoted as $L_{T}$ , for our objective is computed by a weighted sum of the two individual losses. + +$$ +L _ {T} = L _ {L D M} ^ {\prime} + \lambda_ {P} L _ {P} \tag {4} +$$ + +Here, $\lambda_{P}$ is the weighting factor for perceptual loss. These two losses together provide a holistic framework to train our model by ensuring pixel-level accuracy, preservation of high-level features, and smooth motion transitions. + +# 4 Experiments + +# 4.1 Setup + +Datasets We utilize the Tumbled GIF (TGIF) dataset for our training and validation purposes as discussed in Sec. 3.1. We evaluate our model on two datasets: MSR-VTT [64] and UCF-101 [52], following the common practice. For these datasets, we follow the sampling strategy as outlined in [63]. + +![](images/55bcb091bbb48343b53c9a576aa3723491ca2d0e3e4ee295e700cb9bbe4c368e.jpg) +(a) + +![](images/efed11d51d9f0a373847c70abec6cd96a80958ee630596d275f4c359dff3a39e.jpg) +(b) +Fig. 5: Comparison studies with other image-text to video models. Given a source image and a caption, frames are extracted from the generated 16-frame video at 256x256 resolution. + +**Implementations** Our model is initialized with the exponential moving average (EMA) weights of the Stable Diffusion v1.5 checkpoint and the improved ftMSE autoencoder weights $^2$ . We trained the model at 256x256 resolution for 7 epochs on a single node of 16 V100 GPUs for 25k steps. We used the AdamW optimizer with a learning rate of $10^{-4}$ . We set the weighting factor for perceptual loss $(\lambda_P)$ as $10^{-2}$ . + +Metrics We report Frechet Video Distance (FVD) [54], CLIP Similarity (CLIP-Sim) which is the average similarity calculated for all the generated frames with the input caption and Perceptual Input Conformity (PIC) as described in [63] for all methods. For comparison, we assess the zero-shot generation performance on I2VGen-XL [70] and DynamiCrafter [63]. + +![](images/a6d06ceb5f7ff589d8201b917c88fd5558c9f8e007881344ecd2ad3cc1365967.jpg) +(a) A1: Cat is playing with wool +Fig. 6: Pix2Gif showing composition capabilities for different types of motions. [GIFs best viewed in Adobe Acrobat Reader] + +![](images/6d1df9846d32c3c417e770db8893e2c700ccac1972cad12bd3931a448c233f9b.jpg) +(b) A2: Cat is dancing + +![](images/6521da9d68a5bcc4e12d333381ba148d366c1bb0578f7e2a2db390f83c915c1b.jpg) +(c) A1+2: Cat dancing and playing with wool + +Table 2: Ablation study comparing image translation methods with a focus on motion coherency at varyingcfg_img values. + +
Method /cfgimg1.41.61.82.0
L2 ↓PCC ↑L2 ↓PCC ↑L2 ↓PCC ↑L2 ↓PCC ↑
InstructPix2Pix [6]23.429-0.22925.492-0.02827.037-0.42327.5300.139
Pix2Gif-Base7.5800.9895.1880.9875.5950.9927.0290.991
Pix2Gif1.7460.9951.9720.9952.9440.9974.0760.997
+ +# 4.2 Results + +Comparisons with previous works Fig. 5 and Tab. 1 provide a qualitative and quantitative comparison of three image-text to video models: I2VGenXL [70], DynamiCrafter [63], and our $Pix2Gif$ . + +In Fig. 5a, the I2VGen-XL model misshapes the dog's face and generates it sideways in a nonsensical manner. DynamiCrafter appears to disregard the input parameters, as the initial frame differs significantly in position, color, and texture. It is also challenging to discern whether the dog is eating or merely moving its mouth. Our model, Pix2Gif, accurately retains all the dog's details and successfully depicts it eating from a plate. In Fig. 5b, we assess the models' capabilities by generating a video from a relatively dark image. Once again, I2VGen-XL starts strong, producing some impressive frames, but these soon turn into highly stylized and improbable images. DynamiCrafter appears to misinterpret the input image, generating something significantly different, although it seems to adhere to the caption. Conversely, Pix2Gif comprehends the inputs effectively and produces corresponding motion while preserving the overall integrity of the source image. + +Quantitatively, $\text{Pix2Gif}$ excels in both the FVD and PIC metrics shown in Tab. 1, which aligns with our observations of the frames generated in Fig. 5. These frames effectively preserve the structure and closely adhere to the input + +![](images/20380b7b9b5f5c420b65199024486c10deadb7700bbedbc0f1e37c1fda8ee609.jpg) +(a) source frame score + +![](images/a809d8a20838bf48e9a5d83ec4872530839b98b33196e04079032818808d3ed3.jpg) +(b) target frame score +Fig. 7: Ablation study between the earlier variants of our model by comparing average similarity score for 100 samples. + +prompts (source image and caption). However, Pix2Gif does not perform as well in the CLIPSim metric, despite accurately following the caption. The other two models as seen in Fig. 5 do follow the caption, but they fail to adhere to the input image and produce plausible temporal transitions. This is partially attributed to the inherent model design in these two methods. Both methods attempt to generate a full sequence of frames at once using the 3D diffusion network, which inevitably compounds the spatial and temporal dimensions. Moreover, the results indicate that they function more as text-to-video models than image-text-to-video models, especially DynamiCrafter. This discrepancy also raises questions about the effectiveness of the CLIPSim metric for evaluating image-text-to-video models and calls for more sophisticated metrics for evaluating video generation. + +Compositionality of actions Fig. 6 illustrates an intriguing emerging capability of $\text{Pix2Gif}$ : the ability to combine actions. In Fig. 6a, we see a cat playing with wool, with only the cat's paws and the wool moving. In Fig. 6b, we instruct the cat to dance, resulting in the cat moving its body but the wool remaining still. Finally, in Fig. 6c, we provide a caption that blends the actions from Fig. 6a and Fig. 6b. The result is a scene where the cat is both moving the wool and its body. This demonstrates $\text{Pix2Gif}$ 's ability to comprehend the caption and its associated motion, and to convert that understanding into a GIF. Such compositional capability significantly increases user controllability, a crucial aspect for practical applications. + +# 4.3 Ablations + +We design a few variants of $\text{Pix2Gif}$ for our ablation studies: + +- Pix2Gif-Base: We train InstructPix2Pix with our data, and append the text prompt with "The optical flow is ". +- Pix2Gif-Motion-embed: The motion embedding layer is added to encode the motion magnitude and combined with textual embedding. +- Pix2Gif-Warp: We further add the warping module into the model but differently only use the warped feature for the LDM. + +![](images/5649367366230c207fa30b68016064fa7f2d2b333cfb6e6a9fc335676ce9c9e8.jpg) +(a) source frame score + +![](images/9283edcfb9834d1b90281ccb209697d04fd0c3a04a613e2ff8f8034374c42533.jpg) +(b) target frame score +Fig. 8: Ablation study on ways to input $z_{W}$ to LDM by comparing average similarity score for 100 samples. + +- Pix2Gif-Warp-add: Different from Pix2Gif-Warp, we instead add the warped feature and source image feature as input to the LDM. +- Pix2Gif-Warp-concat: Instead of adding in Pix2Gif-Warp-add, we concatenate the warped feature and source image feature as the input, but do not include the perceptual loss. + +For comparative studies with our model's variants, we generate an 8-frame video. We use the X-CLIP model [37] to extract features from our generated video, and CLIP to extract features from the source, target, and generated frames. The optimal range of CFG.img for best results is considered to be [1.6, 2.2]. Throughout this discussion, we evaluate our model's performance using four metrics, which we believe effectively measure the different aspects of generating motion through the image translation framework used in Pix2Gif. + +Motion Coherency Our task is framed as an image translation problem with motion magnitude as a guide. We assess the motion quality or temporal coherence in the GIFs using L2 loss and Pearson Correlation Coefficient (PCC) for InstructPix2Pix, Pix2Gif-Base, and Pix2Gif. The L2 loss evaluates the match between the motion values of the generated frames and the actual inputs, while the PCC checks if they follow the same trend. These metrics are calculated between the input motion magnitude values and the optical flow values, derived from the source image and the generated frames. Our model exhibits the highest correlation and the lowest L2 loss across all CFG.img values as seen in Tab. 2, proving its efficacy and controllability in generating GIFs with specific motions. The Pix2Gif-Base outperforms the original InstructPix2Pix, emphasizing the importance of our new dataset. + +Image-Video Similarity Score To evaluate the semantic properties of the video produced by our model, we created two similarity scores: a source frame score and a target frame score. The source frame score quantifies how well the video retains the primary attributes of the source frame, essentially measuring the accuracy of the source image portrayal throughout the video. The target frame score indicates the precision of the scene or subject development from + +the source frame in the video. It also underscores the model's ability to handle uncertainty and potential changes, as the target frame represents a possible state that the video might reach. + +We began our experiments with $\text{Pix2Gif-Base}$ and then enhanced it incrementally, first with $\mathcal{M}$ (Pix2Gif-Motion-embed), then by adding $\mathcal{W}$ (Pix2Gif-Warp). Here, we conducted an ablation study shown in Fig. 8 to understand the best way to feed $z_{W}$ into LDM. Our experiments included feeding only $z_{W}$ (Pix2Gif-Warp), adding it with the $\mathcal{E}(c_{I})$ (Pix2Gif-Warp-add), and concatenating both (Pix2Gif-Warp-concat). Both Pix2Gif-Warp-add and Pix2Gif-Warp-concat had higher average similarity scores than Pix2Gif-Warp, and Pix2Gif-Warp-concat performed the best. This can be attributed to the fact that in the addition process, $\mathcal{E}(c_{I})$ loses its unique characteristics, which are required by the diffusion model for effective unconditional denoising. Therefore, to achieve the best results, we combined $\mathcal{E}(c_{I})$ and $z_{W}$ before inputting them into the concat attention layer of the LDM. Now finally we integrate $L_{P}$ (Pix2Gif) and compare them in Fig. 7. As expected, Pix2Gif-Base was outperformed by the other versions. Pix2Gif generated more coherent and controlled motion, albeit often with limited extent. All models began to converge outside the optimal range, producing similar frames. The Pix2Gif-Motion-embed model creates a significant amount of one-directional motion, which can sometimes be nonsensical, and hence the addition of $\mathcal{W}$ helps to mitigate this issue. + +# 5 Limitations and Future Work + +The current Pix2Gif model is our initial attempt to generate videos by treating it as an image translation task. However, this method has some limitations that prevent us from generating high-quality and long GIFs or videos. Firstly, the model generates images with a resolution of 256x256 pixels. If these images are used to generate subsequent frames, the quality of the frames deteriorates further. Secondly, due to limitations in computational power, we are only able to use a small portion of a larger, curated dataset for training our model. Our primary objective now is to improve the quality of the generated frames, as this could significantly enhance the effectiveness of this method. + +# 6 Conclusion + +In this work, we proposed Pix2Gif, an image-to-GIF (video) generation model based on an image-to-image translation paradigm. To ensure temporal coherence across frames, we proposed a motion-guided warping module that learns to spatially warp the source image feature into the target one while maintaining visual consistency via a perceptual loss. Starting from TGIF, we curated a new dataset specifically used for training our model. The experimental results demonstrated the effectiveness of our model to generate GIFs with better temporal coherence compared with current state-of-the-art methods. Interestingly, the model also exhibits better controllability and some emerging action compositionality. + +# References + +1. Aigner, S., Körner, M.: Futuregan: Anticipating the future frames of video sequences using spatio-temporal 3d convolutions in progressively growing gans (2018) +2. Avrahami, O., Fried, O., Lischinski, D.: Blended latent diffusion. ACM Transactions on Graphics 42(4), 1-11 (Jul 2023). https://doi.org/10.1145/3592450, http://dx.doi.org/10.1145/3592450 +3. Baradaran, M., Bergevin, R.: Future video prediction from a single frame for video anomaly detection (2023) +4. Blattmann, A., Dockhorn, T., Kulal, S., Mendelevitch, D., Kilian, M., Lorenz, D., Levi, Y., English, Z., Voleti, V., Letts, A., Jampani, V., Rombach, R.: Stable video diffusion: Scaling latent video diffusion models to large datasets (2023) +5. Blattmann, A., Rombach, R., Ling, H., Dockhorn, T., Kim, S.W., Fidler, S., Kreis, K.: Align your latents: High-resolution video synthesis with latent diffusion models (2023) +6. Brooks, T., Holynski, A., Efros, A.A.: Instructpix2pix: Learning to follow image editing instructions. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 18392-18402 (2023) +7. Chang, H., Zhang, H., Barber, J., Maschinot, A., Lezama, J., Jiang, L., Yang, M.H., Murphy, K., Freeman, W.T., Rubinstein, M., Li, Y., Krishnan, D.: Muse: Text-to-image generation via masked generative transformers (2023) +8. Chang, H., Zhang, H., Jiang, L., Liu, C., Freeman, W.T.: Maskgit: Masked generative image transformer (2022) +9. Choi, J., Kim, S., Jeong, Y., Gwon, Y., Yoon, S.: Ilvr: Conditioning method for denoising diffusion probabilistic models (2021) +10. Ding, M., Yang, Z., Hong, W., Zheng, W., Zhou, C., Yin, D., Lin, J., Zou, X., Shao, Z., Yang, H., et al.: Cogview: Mastering text-to-image generation via transformers. Advances in Neural Information Processing Systems 34, 19822-19835 (2021) +11. Esser, P., Chiu, J., Atighehchian, P., Granskog, J., Germanidis, A.: Structure and content-guided video synthesis with diffusion models. In: Proceedings of the IEEE/CVF International Conference on Computer Vision. pp. 7346-7356 (2023) +12. Esser, P., Rombach, R., Ommer, B.: Taming transformers for high-resolution image synthesis (2021) +13. Fu, T.J., Hu, W., Du, X., Wang, W.Y., Yang, Y., Gan, Z.: Guiding instruction-based image editing via multimodal large language models (2023) +14. Fujitake, M., Sugimoto, A.: Video representation learning through prediction for online object detection. In: Proceedings of the IEEE/CVF Winter Conference on Applications of Computer Vision. pp. 530-539 (2022) +15. Ge, S., Nah, S., Liu, G., Poon, T., Tao, A., Catanzaro, B., Jacobs, D., Huang, J.B., Liu, M.Y., Balaji, Y.: Preserve your own correlation: A noise prior for video diffusion models (2023) +16. Girdhar, R., Ramanan, D.: Cater: A diagnostic dataset for compositional actions and temporal reasoning (2020) +17. Goodfellow, I., Pouget-Abadie, J., Mirza, M., Xu, B., Warde-Farley, D., Ozair, S., Courville, A., Bengio, Y.: Generative adversarial networks. Communications of the ACM 63(11), 139-144 (2020) +18. Gu, Y., Yang, J., Usuyama, N., Li, C., Zhang, S., Lungren, M.P., Gao, J., Poon, H.: Biomedjourney: Counterfactual biomedical image generation by instruction-learning from multimodal patient journeys. arXiv preprint arXiv:2310.10765 (2023) + +19. He, Y., Yang, T., Zhang, Y., Shan, Y., Chen, Q.: Latent video diffusion models for high-fidelity long video generation (2023) +20. Hertz, A., Mokady, R., Tenenbaum, J., Aberman, K., Pritch, Y., Cohen-Or, D.: Prompt-to-prompt image editing with cross attention control (2022) +21. Ho, J., Jain, A., Abbeel, P.: Denoising diffusion probabilistic models (2020) +22. Ho, J., Salimans, T.: Classifier-free diffusion guidance. arXiv preprint arXiv:2207.12598 (2022) +23. Ho, J., Salimans, T., Gritsenko, A., Chan, W., Norouzi, M., Fleet, D.J.: Video diffusion models (2022) +24. Hong, W., Ding, M., Zheng, W., Liu, X., Tang, J.: Cogvideo: Large-scale pretraining for text-to-video generation via transformers (2022) +25. Hu, Y., Luo, C., Chen, Z.: Make it move: Controllable image-to-video generation with text descriptions (2022) +26. Kang, M., Zhu, J.Y., Zhang, R., Park, J., Shechtman, E., Paris, S., Park, T.: Scaling up gans for text-to-image synthesis. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 10124-10134 (2023) +27. Karras, T., Laine, S., Aila, T.: A style-based generator architecture for generative adversarial networks. In: Proceedings of the IEEE/CVF conference on computer vision and pattern recognition. pp. 4401-4410 (2019) +28. Kawar, B., Zada, S., Lang, O., Tov, O., Chang, H., Dekel, T., Mosseri, I., Irani, M.: Imagic: Text-based real image editing with diffusion models. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 6007-6017 (2023) +29. Kingma, D.P., Welling, M.: Auto-encoding variational bayes (2022) +30. Lee, J., Lee, J., Lee, S., Yoon, S.: Mutual suppression network for video prediction using disentangled features (2019) +31. Li, H., Yang, Y., Chang, M., Chen, S., Feng, H., Xu, Z., Li, Q., Chen, Y.: Srdiff: Single image super-resolution with diffusion probabilistic models. Neurocomputing 479, 47-59 (2022) +32. Li, Y., Liu, H., Wu, Q., Mu, F., Yang, J., Gao, J., Li, C., Lee, Y.J.: Gligen: Open-set grounded text-to-image generation (2023) +33. Li, Y., Song, Y., Cao, L., Tetreault, J., Goldberg, L., Jaimes, A., Luo, J.: Tgif: A new dataset and benchmark on animated gif description (2016) +34. Liu, W., Luo, W., Lian, D., Gao, S.: Future frame prediction for anomaly detection - a new baseline (2018) +35. Lu, C., Zhou, Y., Bao, F., Chen, J., Li, C., Zhu, J.: Dpm-solver: A fast ode solver for diffusion probabilistic model sampling in around 10 steps (2022) +36. Meng, C., He, Y., Song, Y., Song, J., Wu, J., Zhu, J.Y., Ermon, S.: Sdedit: Guided image synthesis and editing with stochastic differential equations (2022) +37. Ni, B., Peng, H., Chen, M., Zhang, S., Meng, G., Fu, J., Xiang, S., Ling, H.: Expanding language-image pretrained models for general video recognition (2022) +38. Ni, H., Shi, C., Li, K., Huang, S.X., Min, M.R.: Conditional image-to-video generation with latent flow diffusion models (2023) +39. Oh, J., Guo, X., Lee, H., Lewis, R., Singh, S.: Action-conditional video prediction using deep networks in atari games (2015) +40. Oliu, M., Selva, J., Escalera, S.: Folded recurrent neural networks for future video prediction (2018) +41. van den Oord, A., Kalchbrenner, N., Kavukcuoglu, K.: Pixel recurrent neural networks (2016) +42. van den Oord, A., Kalchbrenner, N., Vinyals, O., Espeholt, L., Graves, A., Kavukcuoglu, K.: Conditional image generation with pixelCNN decoders (2016) + +43. van den Oord, A., Vinyals, O., Kavukcuoglu, K.: Neural discrete representation learning (2018) +44. Radford, A., Metz, L., Chintala, S.: Unsupervised representation learning with deep convolutional generative adversarial networks. arXiv preprint arXiv:1511.06434 (2015) +45. Rombach, R., Blattmann, A., Lorenz, D., Esser, P., Ommer, B.: High-resolution image synthesis with latent diffusion models (2022) +46. Ruiz, N., Li, Y., Jampani, V., Pritch, Y., Rubinstein, M., Aberman, K.: Dreambooth: Fine tuning text-to-image diffusion models for subject-driven generation (2023) +47. Sahara, C., Chan, W., Saxena, S., Li, L., Whang, J., Denton, E.L., Ghasemipour, K., Gontijo Lopes, R., Karagol Ayan, B., Salimans, T., et al.: Photorealistic text-to-image diffusion models with deep language understanding. Advances in Neural Information Processing Systems 35, 36479-36494 (2022) +48. Shi, X., Chen, Z., Wang, H., Yeung, D.Y., kin Wong, W., chun Woo, W.: Convolutional LSTM network: A machine learning approach for precipitation nowcasting (2015) +49. Simonyan, K., Zisserman, A.: Very deep convolutional networks for large-scale image recognition (2015) +50. Singer, U., Polyak, A., Hayes, T., Yin, X., An, J., Zhang, S., Hu, Q., Yang, H., Ashual, O., Gafni, O., Parikh, D., Gupta, S., Taigman, Y.: Make-a-video: Text-to-video generation without text-video data (2022) +51. Song, J., Meng, C., Ermon, S.: Denoising diffusion implicit models (2022) +52. Soomro, K., Zamir, A.R., Shah, M.: Ucf101: A dataset of 101 human actions classes from videos in the wild (2012) +53. Srivastava, N., Mansimov, E., Salakhutdinov, R.: Unsupervised learning of video representations using lstms (2016) +54. Unterthiner, T., van Steenkiste, S., Kurach, K., Marinier, R., Michalski, M., Gelly, S.: Fvd: A new metric for video generation (2019) +55. Villegas, R., Babaeizadeh, M., Kindermans, P.J., Moraldo, H., Zhang, H., Saffar, M.T., Castro, S., Kunze, J., Erhan, D.: Phenaki: Variable length video generation from open domain textual description (2022) +56. Villegas, R., Yang, J., Hong, S., Lin, X., Lee, H.: Decomposing motion and content for natural video sequence prediction (2018) +57. Voleti, V., Jolicoeur-Martineau, A., Pal, C.: Mcvd: Masked conditional video diffusion for prediction, generation, and interpolation (2022) +58. Wang, C., Gu, J., Hu, P., Xu, S., Xu, H., Liang, X.: Dreamvideo: High-fidelity image-to-video generation with image retention and text guidance. arXiv preprint arXiv:2312.03018 (2023) +59. Wang, Y., Chen, X., Ma, X., Zhou, S., Huang, Z., Wang, Y., Yang, C., He, Y., Yu, J., Yang, P., et al.: Lavie: High-quality video generation with cascaded latent diffusion models. arXiv preprint arXiv:2309.15103 (2023) +60. Weissenborn, D., Täckström, O., Uszkoreit, J.: Scaling autoregressive video models (2020) +61. Wu, C., Huang, L., Zhang, Q., Li, B., Ji, L., Yang, F., Sapiro, G., Duan, N.: Godiva: Generating open-domain videos from natural descriptions (2021) +62. Wu, C., Liang, J., Ji, L., Yang, F., Fang, Y., Jiang, D., Duan, N.: Niwa: Visual synthesis pre-training for neural visual world creation (2021) +63. Xing, J., Xia, M., Zhang, Y., Chen, H., Wang, X., Wong, T.T., Shan, Y.: Dynamicultural: Animating open-domain images with video diffusion priors. arXiv preprint arXiv:2310.12190 (2023) + +64. Xu, J., Mei, T., Yao, T., Rui, Y.: Msr-vtt: A large video description dataset for bridging video and language. In: Proceedings of the IEEE conference on computer vision and pattern recognition. pp. 5288-5296 (2016) +65. Yang, Z., Wang, J., Gan, Z., Li, L., Lin, K., Wu, C., Duan, N., Liu, Z., Liu, C., Zeng, M., Wang, L.: Reco: Region-controlled text-to-image generation (2022) +66. Yu, J., Xu, Y., Koh, J.Y., Luong, T., Baid, G., Wang, Z., Vasudevan, V., Ku, A., et al.: Scaling autoregressive models for content-rich text-to-image generation. arXiv preprint arXiv:2206.10789 2(3), 5 (2022) +67. Yu, L., Cheng, Y., Sohn, K., Lezama, J., Zhang, H., Chang, H., Hauptmann, A.G., Yang, M.H., Hao, Y., Essa, I., et al.: Magvit: Masked generative video transformer. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 10459-10469 (2023) +68. Yu, L., Shi, B., Pasunuru, R., Muller, B., Golovneva, O., Wang, T., Babu, A., Tang, B., Karrer, B., Sheynin, S., Ross, C., Polyak, A., Howes, R., Sharma, V., Xu, P., Tamoyan, H., Ashual, O., Singer, U., Li, S.W., Zhang, S., James, R., Ghosh, G., Taigman, Y., Fazel-Zarandi, M., Celikyilmaz, A., Zettlemoyer, L., Aghajanyan, A.: Scaling autoregressive multi-modal models: Pretraining and instruction tuning (2023) +69. Zhang, L., Rao, A., Agrawala, M.: Adding conditional control to text-to-image diffusion models. In: Proceedings of the IEEE/CVF International Conference on Computer Vision. pp. 3836-3847 (2023) +70. Zhang, S., Wang, J., Zhang, Y., Zhao, K., Yuan, H., Qin, Z., Wang, X., Zhao, D., Zhou, J.: I2vgen-xl: High-quality image-to-video synthesis via cascaded diffusion models. arXiv preprint arXiv:2311.04145 (2023) +71. Özgün Çiçek, Abdulkadir, A., Lienkamp, S.S., Brox, T., Ronneberger, O.: 3d u-net: Learning dense volumetric segmentation from sparse annotation (2016) \ No newline at end of file diff --git a/2024/Pix2Gif_ Motion-Guided Diffusion for GIF Generation/images.zip b/2024/Pix2Gif_ Motion-Guided Diffusion for GIF Generation/images.zip new file mode 100644 index 0000000000000000000000000000000000000000..bbaab7666fbba7a39c835bfa13d8ce5bea7c853b --- /dev/null +++ b/2024/Pix2Gif_ Motion-Guided Diffusion for GIF Generation/images.zip @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:46f15c75746af71d55f4a3ed0ad479f9028888718bb36abfb865a23cd56e2475 +size 347124 diff --git a/2024/Pix2Gif_ Motion-Guided Diffusion for GIF Generation/layout.json b/2024/Pix2Gif_ Motion-Guided Diffusion for GIF Generation/layout.json new file mode 100644 index 0000000000000000000000000000000000000000..c298a79048f515c6f00c38108efb08ec19bd04d9 --- /dev/null +++ b/2024/Pix2Gif_ Motion-Guided Diffusion for GIF Generation/layout.json @@ -0,0 +1,9404 @@ +{ + "pdf_info": [ + { + "para_blocks": [ + { + "bbox": [ + 159, + 111, + 455, + 146 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 159, + 111, + 455, + 146 + ], + "spans": [ + { + "bbox": [ + 159, + 111, + 455, + 146 + ], + "type": "text", + "content": "Pix2Gif: Motion-Guided Diffusion for GIF Generation" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 184, + 169, + 430, + 182 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 184, + 169, + 430, + 182 + ], + "spans": [ + { + "bbox": [ + 184, + 169, + 430, + 182 + ], + "type": "text", + "content": "Hitsh Kandala1, Jianfeng Gao1, and Jianwei Yang1" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 173, + 191, + 440, + 213 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 173, + 191, + 440, + 213 + ], + "spans": [ + { + "bbox": [ + 173, + 191, + 440, + 213 + ], + "type": "text", + "content": "Microsoft Research \nkhitesh2000@gmail.com,{jfgao, jianwei.yang}@microsoft.com" + } + ] + } + ], + "index": 2 + }, + { + "type": "image", + "bbox": [ + 176, + 248, + 438, + 365 + ], + "blocks": [ + { + "bbox": [ + 176, + 248, + 438, + 365 + ], + "lines": [ + { + "bbox": [ + 176, + 248, + 438, + 365 + ], + "spans": [ + { + "bbox": [ + 176, + 248, + 438, + 365 + ], + "type": "image", + "image_path": "931773cd42f45d093782adbb62d46c47f9065e3b654c6ab8741a95653ba3bc9c.jpg" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 130, + 375, + 482, + 419 + ], + "lines": [ + { + "bbox": [ + 130, + 375, + 482, + 419 + ], + "spans": [ + { + "bbox": [ + 130, + 375, + 482, + 419 + ], + "type": "text", + "content": "Fig. 1: Our model creates distinct frames based on the provided source image and caption, adjusting according to different levels of motion magnitude (optical flow magnitude) specified in the input conditions. It stands well for both high spatial quality and temporal consistency." + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "image_caption" + } + ], + "index": 3 + }, + { + "bbox": [ + 160, + 468, + 455, + 665 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 160, + 468, + 455, + 665 + ], + "spans": [ + { + "bbox": [ + 160, + 468, + 455, + 665 + ], + "type": "text", + "content": "Abstract. We present *Pix2Gif*, a motion-guided diffusion model for image-to-GIF (video) generation. We tackle this problem differently by formulating the task as an image translation problem steered by text and motion magnitude prompts, as shown in Fig. 1. To ensure that the model adheres to motion guidance, we propose a new motion-guided warping module to spatially transform the features of the source image conditioned on the two types of prompts. Furthermore, we introduce a perceptual loss to ensure the transformed feature map remains within the same space as the target image, ensuring content consistency and coherence. In preparation for the model training, we meticulously curated data by extracting coherent image frames from the TGIF video-caption dataset, which provides rich information about the temporal changes of subjects. After pretraining, we apply our model in a zero-shot manner to a number of video datasets. Extensive qualitative and quantitative experiments demonstrate the effectiveness of our model - it not only captures the semantic prompt from text but also the spatial ones from motion guidance. We train all our models using a single node of " + }, + { + "bbox": [ + 160, + 468, + 455, + 665 + ], + "type": "inline_equation", + "content": "16 \\times \\mathrm{V}100" + }, + { + "bbox": [ + 160, + 468, + 455, + 665 + ], + "type": "text", + "content": " GPUs." + } + ] + } + ], + "index": 5 + } + ], + "discarded_blocks": [], + "page_size": [ + 612, + 792 + ], + "page_idx": 0 + }, + { + "para_blocks": [ + { + "bbox": [ + 160, + 117, + 455, + 138 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 160, + 117, + 455, + 138 + ], + "spans": [ + { + "bbox": [ + 160, + 117, + 455, + 138 + ], + "type": "text", + "content": "Keywords: Diffusion " + }, + { + "bbox": [ + 160, + 117, + 455, + 138 + ], + "type": "inline_equation", + "content": "\\cdot" + }, + { + "bbox": [ + 160, + 117, + 455, + 138 + ], + "type": "text", + "content": " image-to-video generation " + }, + { + "bbox": [ + 160, + 117, + 455, + 138 + ], + "type": "inline_equation", + "content": "\\cdot" + }, + { + "bbox": [ + 160, + 117, + 455, + 138 + ], + "type": "text", + "content": " image-to-image translation" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 132, + 160, + 230, + 173 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 160, + 230, + 173 + ], + "spans": [ + { + "bbox": [ + 132, + 160, + 230, + 173 + ], + "type": "text", + "content": "1 Introduction" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 130, + 186, + 479, + 246 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 186, + 479, + 246 + ], + "spans": [ + { + "bbox": [ + 130, + 186, + 479, + 246 + ], + "type": "text", + "content": "Visual content generation has been significantly advanced by the huge progress of diffusion models [21, 22, 35, 51]. Recently, the development of latent diffusion models (LDMs) [45] has led us to a new quality level of generated images. It has inspired a lot of works for customized and controllable image generation [32, 46, 65, 69], and fine-grained image editing [6, 20, 28, 36]." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 130, + 247, + 480, + 472 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 247, + 480, + 472 + ], + "spans": [ + { + "bbox": [ + 130, + 247, + 480, + 472 + ], + "type": "text", + "content": "In this work, we focus on converting a single image to an animated Graphics Interchange Format (GIF), which is valuable for design yet under-explored. Despite the absence of image-to-GIF generation models, diffusion-based video generation has emerged as a hot topic recently. Compared with text-to-image generation, however, text-to-video generation requires not only high quality for individual frames but also visual consistency and temporal coherence across frames. To achieve this goal, existing works expand the LDMs to video diffusion models (VDMs) by either inflating the 2D CNNs in LDMs to 3D ones [23] or introducing an additional temporal attention layer to bridge the diffusion for each frame [5,15,19,50,59]. In addition to text prompts, a few recent works also explored the way of using images or other prompts to make the video generation model more customizable and controllable [11,38]. However, due to the high cost of VDMs to generate a sequence of video frames in one run, most (if not all) of these works require a compromise of reducing the resolution of generated frames " + }, + { + "bbox": [ + 130, + 247, + 480, + 472 + ], + "type": "inline_equation", + "content": "(64 \\times 64" + }, + { + "bbox": [ + 130, + 247, + 480, + 472 + ], + "type": "text", + "content": " typically), and the usage of extra super-resolution diffusion models for upscaling [31,47]. Moreover, since these methods use the temporal attention layers to model the cross-frame dependency implicitly, it is quite hard for them to preserve good controllability of the frame-to-frame temporal dynamics in a fine-grained manner." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 130, + 474, + 481, + 665 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 474, + 481, + 665 + ], + "spans": [ + { + "bbox": [ + 130, + 474, + 481, + 665 + ], + "type": "text", + "content": "Given that animated GIF usually contains less number of frames and requires more specializations, we take a different strategy and formulate the image-to-GIF generation as an image translation process. To decouple the generation of visual contents and temporal dynamics, we further introduce a motion flow magnitude as extra guidance in addition to image and text prompts. Unlike the aforementioned works, our model takes one or more history frames as the condition and produces only one future frame at once. This brings some unique advantages: (i) simplicity - our model can be purely built on top of LDMs and trained end-to-end with high resolution, without any cascaded diffusion processes for upscaling. (ii) controllability - we could inject detailed and different text and motion prompts at each time step for generating a frame, which gains much better controllability of the model. Our work is inspired by a line of canonical works for future frame prediction [40, 48, 56]. However, due to the lack of a powerful image-generation engine, these works fail to produce high-quality results and can only be applied to specific video domains [16, 30, 52]. Moreover, they cannot support other types of prompts or conditions than the history frames. To address this problem," + } + ] + } + ], + "index": 6 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "spans": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "type": "text", + "content": "2" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 166, + 91, + 237, + 100 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 166, + 91, + 237, + 100 + ], + "spans": [ + { + "bbox": [ + 166, + 91, + 237, + 100 + ], + "type": "text", + "content": "H. Kandala et al." + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 1 + }, + { + "para_blocks": [ + { + "bbox": [ + 130, + 116, + 482, + 236 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 116, + 482, + 236 + ], + "spans": [ + { + "bbox": [ + 130, + 116, + 482, + 236 + ], + "type": "text", + "content": "we exploit a modern diffusion-based pipeline. More specifically, we follow text-conditioned image editing approaches (e.g., InstructPix2Pix [6]), and propose a new temporal image editing to produce future frames given history frames. To train the model, we curate a new training dataset based on TGIF [33] by extracting frames and calculating the magnitude of optical flow between them. We then selected an appropriate range of the optical flow magnitude and sampled frame pairs from each GIF in a manner that ensures diversity. In the end, we train our diffusion model called Pix2Gif, which can generate high-quality animated GIF consisting of multiple frames, given a single image and text and motion magnitude prompts. In summary, our main contributions are:" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 138, + 250, + 479, + 383 + ], + "type": "list", + "angle": 0, + "index": 7, + "blocks": [ + { + "bbox": [ + 138, + 250, + 479, + 285 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 250, + 479, + 285 + ], + "spans": [ + { + "bbox": [ + 138, + 250, + 479, + 285 + ], + "type": "text", + "content": "- We are the first to explore an image-to-image translation formula for generating animated GIFs from an image, guided by a text prompt and motion magnitude." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 138, + 287, + 479, + 321 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 287, + 479, + 321 + ], + "spans": [ + { + "bbox": [ + 138, + 287, + 479, + 321 + ], + "type": "text", + "content": "- We propose a flow-based warping module with a perceptual loss in the diffusion process that takes motion magnitude as input and controls the temporal dynamics and consistency between future frames and the initial ones." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 138, + 323, + 479, + 346 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 323, + 479, + 346 + ], + "spans": [ + { + "bbox": [ + 138, + 323, + 479, + 346 + ], + "type": "text", + "content": "- We curate a new dataset, comprised of 78,692 short GIF clips for training, and 10,546 for evaluation. The new dataset covers a variety of visual domains." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 138, + 349, + 479, + 383 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 349, + 479, + 383 + ], + "spans": [ + { + "bbox": [ + 138, + 349, + 479, + 383 + ], + "type": "text", + "content": "- Quantitative and qualitative results demonstrate the effectiveness of our proposed method for generating visually consistent coherent GIFs from a single image, and it can be generalized to a wide range of visual domains." + } + ] + } + ], + "index": 6 + } + ], + "sub_type": "text" + }, + { + "bbox": [ + 132, + 407, + 222, + 419 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 407, + 222, + 419 + ], + "spans": [ + { + "bbox": [ + 132, + 407, + 222, + 419 + ], + "type": "text", + "content": "2 Prior Work" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 130, + 437, + 482, + 557 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 437, + 482, + 557 + ], + "spans": [ + { + "bbox": [ + 130, + 437, + 482, + 557 + ], + "type": "text", + "content": "Image and video generation has been a long-standing problem in the community. It can be tackled by different approaches, which can be categorized into four groups: generative adversarial networks (GANs) [17,26,27,44], transformer-based autoregressive decoding [10,12,41,42,60,66], masked image modeling [7,8,55,67]. Most of the recent works exploited diffusion models for image generation given their high-quality outputs and huge open-source supports [45, 47]. Recently, a number of works have extended the text-to-image generation model into image translation or editing models [6,20,28,36,69] or video generation models [5,15, 23,50,57]. Below we provide a brief overview of the related diffusion-based image and video generation methods." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 130, + 570, + 482, + 665 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 570, + 482, + 665 + ], + "spans": [ + { + "bbox": [ + 130, + 570, + 482, + 665 + ], + "type": "text", + "content": "Image-to-Image Translation. Diffusion-based image-to-image generation has drawn increasing attention. Different from text-to-image generation, it takes an image as input and edits its contents following the text instructions while keeping the unrelated parts unchanged. SDEdit [36] and ILVR [9] are two pioneering works that impose reference image conditions to an existing latent diffusion model for controllable image generation. Later on, to conduct local edits, the authors in [2] proposed blended latent diffusion to steer the diffusion process with a user-specified mask, where the pixels out of the mask remain the same" + } + ] + } + ], + "index": 10 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 227, + 91, + 447, + 101 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 227, + 91, + 447, + 101 + ], + "spans": [ + { + "bbox": [ + 227, + 91, + 447, + 101 + ], + "type": "text", + "content": "Pix2Gif: Motion-Guided Diffusion for GIF Generation" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 474, + 92, + 480, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 474, + 92, + 480, + 100 + ], + "spans": [ + { + "bbox": [ + 474, + 92, + 480, + 100 + ], + "type": "text", + "content": "3" + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 2 + }, + { + "para_blocks": [ + { + "bbox": [ + 130, + 116, + 482, + 213 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 116, + 482, + 213 + ], + "spans": [ + { + "bbox": [ + 130, + 116, + 482, + 213 + ], + "type": "text", + "content": "as the input image while the region inside is edited following the textual description. Instead of manipulating the image space, Prompt2Prompt [20] proposed to edit the image by manipulating the textual context (e.g., swapping or adding words.) to which the latent diffusion model cross-attends. However, this method requires forwarding a text-to-image generation process to obtain the cross-attention maps, and thus cannot be applied to real images. Imagic [28] proposed to blend the embeddings of a real image with the textual context embedding so that the generated image obeys both the image and text conditions." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 130, + 214, + 482, + 346 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 214, + 482, + 346 + ], + "spans": [ + { + "bbox": [ + 130, + 214, + 482, + 346 + ], + "type": "text", + "content": "All the aforementioned works leverage a frozen latent diffusion model and control the generation with modified text or image prompts. To enable arbitrary image editing, InstructPix2Pix [6] proposed to finetune the LDM to follow user instructions that precisely convey the user intents, e.g., \"change the cat to dog\". The model is trained by a synthetic dataset consisting of triplets " + }, + { + "bbox": [ + 130, + 214, + 482, + 346 + ], + "type": "inline_equation", + "content": "\\langle \\text{image}_{\\text{src}}, \\text{instruction}, \\text{image}_{\\text{tgt}} \\rangle" + }, + { + "bbox": [ + 130, + 214, + 482, + 346 + ], + "type": "text", + "content": ". The resulting model could allow both realistic and generated images and support arbitrary language instructions. Some other works also exploit a similar way to train the model to follow instructions [18,68]. To further enhance the language understanding, MGIE [13] exploited a large multimodal model to produce a more comprehensive textual context for the instructed image editing." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 130, + 349, + 482, + 421 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 349, + 482, + 421 + ], + "spans": [ + { + "bbox": [ + 130, + 349, + 482, + 421 + ], + "type": "text", + "content": "In this work, we employ the image-to-image translation pipeline and are the first to formulate a GIF generation as an image translation problem. Given a reference image, the goal is to generate a realistic future frame following a textual instruction. Therefore, the focus is on how to perform temporal rather than spatial editing on a source image. When the process rolls out, it gradually gives a sequence of frames." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 130, + 435, + 482, + 615 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 435, + 482, + 615 + ], + "spans": [ + { + "bbox": [ + 130, + 435, + 482, + 615 + ], + "type": "text", + "content": "Conditioned Video Generation. Speaking of the high-level goal, our work resembles conditioned video generation. For video generation, a conventional way is inflating the 2D U-Net used in LDM to 3D U-Net [71] by replacing the 2D convolution layers with 3D ones. Likewise, a similar strategy is taken in [5,15,19,50,59], but with a slight difference in that they use interleaved spatial and temporal attention layers in the U-Net. Due to the high cost of generating a sequence of video frames in one shot, the output videos usually have a resolution as low as " + }, + { + "bbox": [ + 130, + 435, + 482, + 615 + ], + "type": "inline_equation", + "content": "64 \\times 64" + }, + { + "bbox": [ + 130, + 435, + 482, + 615 + ], + "type": "text", + "content": ". To attain high-resolution videos, these methods need to use one or more super-resolution diffusion models [31,47] to upscale the resolution by 4 or 8 times. To accelerate the training, a pre-trained text-to-image LDM is usually used to initialize these models. Adding spatial-temporal modules is also a commonly used strategy for autoregressive models [24, 55, 61, 62]. Similarly, both [62] and [24] exploit a pretrained autoregressive image generation model as the starting point. In [55], however, the authors introduce and pretrained a new video encoder, which is then used to train a masked video decoder." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 130, + 617, + 482, + 666 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 617, + 482, + 666 + ], + "spans": [ + { + "bbox": [ + 130, + 617, + 482, + 666 + ], + "type": "text", + "content": "Besides text-to-video generation, using images as the condition for video generation draws increasing attention. On one hand, once a text-to-video generation is trained, it can be further finetuned for image-to-video generation [4]. In [11], the authors introduce additional structural conditions (e.g., depth maps) for" + } + ] + } + ], + "index": 6 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "spans": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "type": "text", + "content": "4" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 166, + 91, + 237, + 101 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 166, + 91, + 237, + 101 + ], + "spans": [ + { + "bbox": [ + 166, + 91, + 237, + 101 + ], + "type": "text", + "content": "H. Kandala et al." + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 3 + }, + { + "para_blocks": [ + { + "bbox": [ + 130, + 116, + 479, + 187 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 116, + 479, + 187 + ], + "spans": [ + { + "bbox": [ + 130, + 116, + 479, + 187 + ], + "type": "text", + "content": "more controllable video generation. Alternatively in [38], a latent flow diffusion model is introduced for image-conditioned video generation by explicitly generating a sequence of optimal flows and masks as the guidance. On the other hand, a few concurrent works to ours directly approach image-to-video generation on top of video diffusion models [58, 63, 70]. All these works share a similar spirit to text-to-video generation models but add additional images as the reference." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 130, + 190, + 480, + 250 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 190, + 480, + 250 + ], + "spans": [ + { + "bbox": [ + 130, + 190, + 480, + 250 + ], + "type": "text", + "content": "Our method uses a diffusion model but differs from all the aforementioned methods in that we reformulate video generation as a frame-to-frame translation problem based on the history frames. As [6] suggests image-to-image translation can maintain a decent visual consistency. In addition, we also introduce a motion flow magnitude as another condition to explicitly control the temporal dynamics." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 130, + 264, + 481, + 431 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 264, + 481, + 431 + ], + "spans": [ + { + "bbox": [ + 130, + 264, + 481, + 431 + ], + "type": "text", + "content": "Future Frame Prediction. Future frame prediction or forecasting [40, 48, 56] has been a long-standing problem before the prevalence of diffusion models. It has been used as an anomaly detection approach by comparing the observed frame and the predicted ones [3, 34] and video representation learning for various downstream tasks [14]. For these problems, a recurrent network such as LSTM [53], ConvLSTM [39, 56] or 3D-CNN [1] is usually used as the model architecture, and GAN [17] or Variational Autoencoder (VAE) [29] is used as the learning objectives. With the emergence of VQ-VAE [43], the authors in [25] exploited axial transformer blocks to chain the encoder-decoder for autoregressive next-frame prediction. In [57], the authors proposed masked conditional video diffusion to unify different tasks of video prediction, generation and interpolation. Nevertheless, all of these models are trained on domain-specific video datasets such as MovingMNIST [30], CATER [16] and UCF-101 [52], etc, far from being a generic video generation model." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 130, + 434, + 482, + 517 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 434, + 482, + 517 + ], + "spans": [ + { + "bbox": [ + 130, + 434, + 482, + 517 + ], + "type": "text", + "content": "Our work takes inspiration from future frame prediction methods but proposes a simpler yet effective strategy by formulating it as an image-to-image translation problem. Furthermore, our model simultaneously takes image, text and motion magnitude as the guidance for better controllability. To attain a model as general as possible, we curate a new training dataset covering a wide range of domains. Without any further dataset-specific finetuning, our model achieves plausible video generation results as shown in Fig. 1." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 132, + 546, + 201, + 558 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 546, + 201, + 558 + ], + "spans": [ + { + "bbox": [ + 132, + 546, + 201, + 558 + ], + "type": "text", + "content": "3 Method" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 130, + 582, + 482, + 665 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 582, + 482, + 665 + ], + "spans": [ + { + "bbox": [ + 130, + 582, + 482, + 665 + ], + "type": "text", + "content": "Our goal is to generate GIFs, given an initial frame, a motion description, and a measure of optical flow. We frame this as an image-to-image translation problem based on latent diffusion. We first explain how we created our training dataset in Sec. 3.1, then outline our model's principles and training strategy in Sec. 3.2. Next, we delve into the specifics of our proposed model, explaining its various components in Sec. 3.3. Finally, we focus on the loss functions used to train our model in Sec. 3.4." + } + ] + } + ], + "index": 7 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 227, + 91, + 447, + 102 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 227, + 91, + 447, + 102 + ], + "spans": [ + { + "bbox": [ + 227, + 91, + 447, + 102 + ], + "type": "text", + "content": "Pix2Gif: Motion-Guided Diffusion for GIF Generation" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 474, + 92, + 480, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 474, + 92, + 480, + 100 + ], + "spans": [ + { + "bbox": [ + 474, + 92, + 480, + 100 + ], + "type": "text", + "content": "5" + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 4 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 176, + 114, + 438, + 213 + ], + "blocks": [ + { + "bbox": [ + 176, + 114, + 438, + 213 + ], + "lines": [ + { + "bbox": [ + 176, + 114, + 438, + 213 + ], + "spans": [ + { + "bbox": [ + 176, + 114, + 438, + 213 + ], + "type": "image", + "image_path": "8d0c160964518d0d443d4947b2090224c98840295ce1819c05b58b0e6d0e1ce5.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 130, + 221, + 482, + 255 + ], + "lines": [ + { + "bbox": [ + 130, + 221, + 482, + 255 + ], + "spans": [ + { + "bbox": [ + 130, + 221, + 482, + 255 + ], + "type": "text", + "content": "Fig. 2: The three step process of curating the TGIF dataset. Starting from extracting frames " + }, + { + "bbox": [ + 130, + 221, + 482, + 255 + ], + "type": "inline_equation", + "content": "A" + }, + { + "bbox": [ + 130, + 221, + 482, + 255 + ], + "type": "text", + "content": " to restricting the range of optical flow " + }, + { + "bbox": [ + 130, + 221, + 482, + 255 + ], + "type": "inline_equation", + "content": "B" + }, + { + "bbox": [ + 130, + 221, + 482, + 255 + ], + "type": "text", + "content": " and then maintaining the diversity of pairs " + }, + { + "bbox": [ + 130, + 221, + 482, + 255 + ], + "type": "inline_equation", + "content": "C" + }, + { + "bbox": [ + 130, + 221, + 482, + 255 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_caption" + } + ], + "index": 2 + }, + { + "type": "image", + "bbox": [ + 176, + 268, + 440, + 381 + ], + "blocks": [ + { + "bbox": [ + 176, + 268, + 440, + 381 + ], + "lines": [ + { + "bbox": [ + 176, + 268, + 440, + 381 + ], + "spans": [ + { + "bbox": [ + 176, + 268, + 440, + 381 + ], + "type": "image", + "image_path": "ea8b16389739e85bdecfa9ac8571d8be904b50cfadaf607d471bc3376e4ff0c5.jpg" + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 130, + 389, + 482, + 422 + ], + "lines": [ + { + "bbox": [ + 130, + 389, + 482, + 422 + ], + "spans": [ + { + "bbox": [ + 130, + 389, + 482, + 422 + ], + "type": "text", + "content": "Fig.3: " + }, + { + "bbox": [ + 130, + 389, + 482, + 422 + ], + "type": "inline_equation", + "content": "\\text{Pix2Gif}" + }, + { + "bbox": [ + 130, + 389, + 482, + 422 + ], + "type": "text", + "content": " model pipeline. We propose an end-to-end network where the inputs are encoded by " + }, + { + "bbox": [ + 130, + 389, + 482, + 422 + ], + "type": "inline_equation", + "content": "\\mathcal{E}" + }, + { + "bbox": [ + 130, + 389, + 482, + 422 + ], + "type": "text", + "content": ", CLIP and " + }, + { + "bbox": [ + 130, + 389, + 482, + 422 + ], + "type": "inline_equation", + "content": "\\mathcal{M}" + }, + { + "bbox": [ + 130, + 389, + 482, + 422 + ], + "type": "text", + "content": " to output " + }, + { + "bbox": [ + 130, + 389, + 482, + 422 + ], + "type": "inline_equation", + "content": "\\mathcal{E}(c_L), c_T" + }, + { + "bbox": [ + 130, + 389, + 482, + 422 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 130, + 389, + 482, + 422 + ], + "type": "inline_equation", + "content": "c_M" + }, + { + "bbox": [ + 130, + 389, + 482, + 422 + ], + "type": "text", + "content": " respectively, which then goes into " + }, + { + "bbox": [ + 130, + 389, + 482, + 422 + ], + "type": "inline_equation", + "content": "\\mathcal{W}" + }, + { + "bbox": [ + 130, + 389, + 482, + 422 + ], + "type": "text", + "content": " to form the conditioning input for LDM." + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "image_caption" + } + ], + "index": 4 + }, + { + "bbox": [ + 132, + 449, + 200, + 459 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 449, + 200, + 459 + ], + "spans": [ + { + "bbox": [ + 132, + 449, + 200, + 459 + ], + "type": "text", + "content": "3.1 Dataset" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 130, + 472, + 482, + 508 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 472, + 482, + 508 + ], + "spans": [ + { + "bbox": [ + 130, + 472, + 482, + 508 + ], + "type": "text", + "content": "We used the Tumblr GIF (TGIF) dataset [33], which predominantly consists of human-centric animated GIFs described by captions. The dataset features a range of GIFs with varied movements over 1-3 seconds." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 130, + 510, + 482, + 666 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 510, + 482, + 666 + ], + "spans": [ + { + "bbox": [ + 130, + 510, + 482, + 666 + ], + "type": "text", + "content": "The curation process, as shown in Fig. 2, involved extracting frames from all GIFs and calculating the optical flow between all possible frame pairs. The number of extracted frames from each GIF varied, with an average of about 41 frames. The optical flow histogram calculated between all frames ranged from [0, 200]. We selected the [2, 20] range, capturing small yet significant motion and excluding pairs with drastic changes. Despite the restricted range, we still had a significant number of training pairs. To avoid model overfitting and maintain diversity, we randomly selected a minimum of 10 pairs or the number of pairs within the restricted range from each GIF. This approach resulted in a final dataset with nearly equal representation of all values within the selected range. The final dataset contained 783,184 training pairs and 105,041 validation pairs. Each data point consisted of a pair of frames from the same GIF, the corresponding caption, and the calculated optical flow between the frames." + } + ] + } + ], + "index": 8 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "spans": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "type": "text", + "content": "6" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 166, + 91, + 237, + 100 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 166, + 91, + 237, + 100 + ], + "spans": [ + { + "bbox": [ + 166, + 91, + 237, + 100 + ], + "type": "text", + "content": "H. Kandala et al." + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 5 + }, + { + "para_blocks": [ + { + "bbox": [ + 132, + 116, + 354, + 129 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 116, + 354, + 129 + ], + "spans": [ + { + "bbox": [ + 132, + 116, + 354, + 129 + ], + "type": "text", + "content": "3.2 Preliminary: Instructed Image Editing" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 130, + 141, + 482, + 262 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 141, + 482, + 262 + ], + "spans": [ + { + "bbox": [ + 130, + 141, + 482, + 262 + ], + "type": "text", + "content": "Our model is fundamentally grounded in the latent diffusion models (LDMs) for image generation and editing [6, 45]. More specifically, we build upon InstructPix2Pix [6] by framing our objective in the context of an instructed image-to-image translation task. Given an image " + }, + { + "bbox": [ + 130, + 141, + 482, + 262 + ], + "type": "inline_equation", + "content": "x" + }, + { + "bbox": [ + 130, + 141, + 482, + 262 + ], + "type": "text", + "content": ", the forward diffusion procedure introduces noise to the encoded latent " + }, + { + "bbox": [ + 130, + 141, + 482, + 262 + ], + "type": "inline_equation", + "content": "z" + }, + { + "bbox": [ + 130, + 141, + 482, + 262 + ], + "type": "text", + "content": ", thereby producing a noisy latent vector " + }, + { + "bbox": [ + 130, + 141, + 482, + 262 + ], + "type": "inline_equation", + "content": "z_{t}" + }, + { + "bbox": [ + 130, + 141, + 482, + 262 + ], + "type": "text", + "content": ". This process is carried out over " + }, + { + "bbox": [ + 130, + 141, + 482, + 262 + ], + "type": "inline_equation", + "content": "T" + }, + { + "bbox": [ + 130, + 141, + 482, + 262 + ], + "type": "text", + "content": " timesteps, with each timestep " + }, + { + "bbox": [ + 130, + 141, + 482, + 262 + ], + "type": "inline_equation", + "content": "t \\in \\{1, \\dots, T\\}" + }, + { + "bbox": [ + 130, + 141, + 482, + 262 + ], + "type": "text", + "content": " seeing an increment in the noise level until it culminates into a random noise " + }, + { + "bbox": [ + 130, + 141, + 482, + 262 + ], + "type": "inline_equation", + "content": "n" + }, + { + "bbox": [ + 130, + 141, + 482, + 262 + ], + "type": "text", + "content": ". A network " + }, + { + "bbox": [ + 130, + 141, + 482, + 262 + ], + "type": "inline_equation", + "content": "e_{\\theta}" + }, + { + "bbox": [ + 130, + 141, + 482, + 262 + ], + "type": "text", + "content": " is trained by minimizing the following latent diffusion objective to predict noise existing in the noisy latent " + }, + { + "bbox": [ + 130, + 141, + 482, + 262 + ], + "type": "inline_equation", + "content": "z_{t}" + }, + { + "bbox": [ + 130, + 141, + 482, + 262 + ], + "type": "text", + "content": ", considering factors in image conditioning " + }, + { + "bbox": [ + 130, + 141, + 482, + 262 + ], + "type": "inline_equation", + "content": "c_{I}" + }, + { + "bbox": [ + 130, + 141, + 482, + 262 + ], + "type": "text", + "content": " and textual instruction " + }, + { + "bbox": [ + 130, + 141, + 482, + 262 + ], + "type": "inline_equation", + "content": "c_{T}" + }, + { + "bbox": [ + 130, + 141, + 482, + 262 + ], + "type": "text", + "content": ":" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 180, + 272, + 481, + 287 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 180, + 272, + 481, + 287 + ], + "spans": [ + { + "bbox": [ + 180, + 272, + 481, + 287 + ], + "type": "interline_equation", + "content": "L _ {L D M} = \\mathbb {E} _ {\\mathcal {E} (x), \\mathcal {E} (c _ {I}), c _ {T}, \\epsilon \\sim \\mathcal {N} (0, 1), t} \\left[ | | \\epsilon - \\epsilon_ {\\theta} \\left(z _ {t}, t, \\mathcal {E} \\left(c _ {I}\\right), c _ {T}\\right) | | _ {2} ^ {2} \\right] \\tag {1}", + "image_path": "89f42740d47906c0066ac54036297a313b98fb44e459f6788d567120ce9b29a6.jpg" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 130, + 297, + 482, + 357 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 297, + 482, + 357 + ], + "spans": [ + { + "bbox": [ + 130, + 297, + 482, + 357 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 130, + 297, + 482, + 357 + ], + "type": "inline_equation", + "content": "\\mathcal{E}" + }, + { + "bbox": [ + 130, + 297, + 482, + 357 + ], + "type": "text", + "content": " is the VQ-VAE encoder that transforms the images from pixel space to discrete latent space. To facilitate image conditioning, " + }, + { + "bbox": [ + 130, + 297, + 482, + 357 + ], + "type": "inline_equation", + "content": "z_{t}" + }, + { + "bbox": [ + 130, + 297, + 482, + 357 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 130, + 297, + 482, + 357 + ], + "type": "inline_equation", + "content": "\\mathcal{E}(c_I)" + }, + { + "bbox": [ + 130, + 297, + 482, + 357 + ], + "type": "text", + "content": " are concatenated and then fed into a convolutional layer. The model is trained for conditional and unconditional denoising, given the image and caption condition individually or collectively." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 132, + 379, + 263, + 391 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 379, + 263, + 391 + ], + "spans": [ + { + "bbox": [ + 132, + 379, + 263, + 391 + ], + "type": "text", + "content": "3.3 Our Model: Pix2Gif" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 130, + 404, + 481, + 464 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 404, + 481, + 464 + ], + "spans": [ + { + "bbox": [ + 130, + 404, + 481, + 464 + ], + "type": "text", + "content": "We build our model similar to InstructPix2Pix and frame our objective in the context of a text-instructed and motion-guided temporal editing problem. Compared with the original InstructPix2Pix pipeline, the main innovation is the newly introduced motion-based warping module. The overall model pipeline is shown in Fig. 3." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 130, + 465, + 482, + 573 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 465, + 482, + 573 + ], + "spans": [ + { + "bbox": [ + 130, + 465, + 482, + 573 + ], + "type": "text", + "content": "Our model takes three inputs: an image, a text instruction, and a motion magnitude. These inputs are fed into the model through two pathways - once through the diffusion model directly, and again through the warping module, which will be discussed in Sec. 3.3 and Sec. 3.3. When passed through the caption, we add the phrase \"The optical flow is _. to the original caption. The flow input is then appended at the end in a word format rather than a numerical one, as the CLIP model tends to assign higher similarity scores to word forms than to numerical representations of numbers for the same image. Finally, our model is trained by minimizing the following loss function:" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 165, + 585, + 481, + 599 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 165, + 585, + 481, + 599 + ], + "spans": [ + { + "bbox": [ + 165, + 585, + 481, + 599 + ], + "type": "interline_equation", + "content": "L _ {L D M} ^ {\\prime} = \\mathbb {E} _ {\\mathcal {E} (x), \\mathcal {E} (c _ {I}), c _ {T}, c _ {M}, \\epsilon \\sim \\mathcal {N} (0, 1), t} \\left[ | | \\epsilon - \\epsilon_ {\\theta} (z _ {t}, t, \\mathcal {E} (c _ {I}), c _ {T}, c _ {M}) | | _ {2} ^ {2} \\right] \\tag {2}", + "image_path": "e6435a73030bd90cd3897a07b1ad649f071cc631b459195342115d4aa9f82992.jpg" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 130, + 609, + 482, + 669 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 609, + 482, + 669 + ], + "spans": [ + { + "bbox": [ + 130, + 609, + 482, + 669 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 130, + 609, + 482, + 669 + ], + "type": "inline_equation", + "content": "c_{M}" + }, + { + "bbox": [ + 130, + 609, + 482, + 669 + ], + "type": "text", + "content": " is the motion condition. The altered caption is processed via the pretrained CLIP model to yield " + }, + { + "bbox": [ + 130, + 609, + 482, + 669 + ], + "type": "inline_equation", + "content": "c_{T}" + }, + { + "bbox": [ + 130, + 609, + 482, + 669 + ], + "type": "text", + "content": ", while the output of " + }, + { + "bbox": [ + 130, + 609, + 482, + 669 + ], + "type": "inline_equation", + "content": "\\mathcal{M}" + }, + { + "bbox": [ + 130, + 609, + 482, + 669 + ], + "type": "text", + "content": " gives us " + }, + { + "bbox": [ + 130, + 609, + 482, + 669 + ], + "type": "inline_equation", + "content": "c_{M}" + }, + { + "bbox": [ + 130, + 609, + 482, + 669 + ], + "type": "text", + "content": ". These two conditions are then added linearly, serving as the conditioning input for both the Warping Module " + }, + { + "bbox": [ + 130, + 609, + 482, + 669 + ], + "type": "inline_equation", + "content": "\\mathcal{W}" + }, + { + "bbox": [ + 130, + 609, + 482, + 669 + ], + "type": "text", + "content": " (discussed in Sec. 3.3) and the Latent Diffusion Model LDM (referenced in Sec. 3.2)." + } + ] + } + ], + "index": 10 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 227, + 91, + 447, + 101 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 227, + 91, + 447, + 101 + ], + "spans": [ + { + "bbox": [ + 227, + 91, + 447, + 101 + ], + "type": "text", + "content": "Pix2Gif: Motion-Guided Diffusion for GIF Generation" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 474, + 92, + 480, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 474, + 92, + 480, + 100 + ], + "spans": [ + { + "bbox": [ + 474, + 92, + 480, + 100 + ], + "type": "text", + "content": "7" + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 6 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 186, + 115, + 430, + 247 + ], + "blocks": [ + { + "bbox": [ + 186, + 115, + 430, + 247 + ], + "lines": [ + { + "bbox": [ + 186, + 115, + 430, + 247 + ], + "spans": [ + { + "bbox": [ + 186, + 115, + 430, + 247 + ], + "type": "image", + "image_path": "d5aa299c7c9f03e600a0952a7ba0d5b45d37c8c7c14f76f9a2734e53197ddd82.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 131, + 256, + 480, + 278 + ], + "lines": [ + { + "bbox": [ + 131, + 256, + 480, + 278 + ], + "spans": [ + { + "bbox": [ + 131, + 256, + 480, + 278 + ], + "type": "text", + "content": "Fig. 4: Deep dive into the Warping Module " + }, + { + "bbox": [ + 131, + 256, + 480, + 278 + ], + "type": "inline_equation", + "content": "\\mathcal{W}" + }, + { + "bbox": [ + 131, + 256, + 480, + 278 + ], + "type": "text", + "content": ". It comprises of three units: " + }, + { + "bbox": [ + 131, + 256, + 480, + 278 + ], + "type": "inline_equation", + "content": "\\mathcal{I}_M" + }, + { + "bbox": [ + 131, + 256, + 480, + 278 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 131, + 256, + 480, + 278 + ], + "type": "inline_equation", + "content": "\\mathcal{F}_{Net}" + }, + { + "bbox": [ + 131, + 256, + 480, + 278 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 131, + 256, + 480, + 278 + ], + "type": "inline_equation", + "content": "\\mathcal{W}_{Net}" + }, + { + "bbox": [ + 131, + 256, + 480, + 278 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_caption" + } + ], + "index": 2 + }, + { + "bbox": [ + 130, + 308, + 482, + 429 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 308, + 482, + 429 + ], + "spans": [ + { + "bbox": [ + 130, + 308, + 482, + 429 + ], + "type": "text", + "content": "Motion Embedding Layer In traditional conditional diffusion models like [6], text prompts are often enough for image generation. We initially passed the motion input indirectly through the prompt. However, this divided the model's attention on a single caption token, which is problematic when the main caption stays the same but the motion input changes. To let the model focus on the motion input, we included an embedding layer that converts the motion input into an integer and selects an embedding vector. This vector is then duplicated and concatenated with itself to generate " + }, + { + "bbox": [ + 130, + 308, + 482, + 429 + ], + "type": "inline_equation", + "content": "c_{M}" + }, + { + "bbox": [ + 130, + 308, + 482, + 429 + ], + "type": "text", + "content": ", which when combined with the caption embedding " + }, + { + "bbox": [ + 130, + 308, + 482, + 429 + ], + "type": "inline_equation", + "content": "c_{T}" + }, + { + "bbox": [ + 130, + 308, + 482, + 429 + ], + "type": "text", + "content": ", provides the conditioning input " + }, + { + "bbox": [ + 130, + 308, + 482, + 429 + ], + "type": "inline_equation", + "content": "c_{L} = c_{T} + c_{M}" + }, + { + "bbox": [ + 130, + 308, + 482, + 429 + ], + "type": "text", + "content": " for both warping module " + }, + { + "bbox": [ + 130, + 308, + 482, + 429 + ], + "type": "inline_equation", + "content": "\\mathcal{W}" + }, + { + "bbox": [ + 130, + 308, + 482, + 429 + ], + "type": "text", + "content": " and LDM." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 130, + 449, + 482, + 668 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 449, + 482, + 668 + ], + "spans": [ + { + "bbox": [ + 130, + 449, + 482, + 668 + ], + "type": "text", + "content": "Warping Module One of the main components of Pix2Gif is the Warping Module " + }, + { + "bbox": [ + 130, + 449, + 482, + 668 + ], + "type": "inline_equation", + "content": "\\mathcal{W}" + }, + { + "bbox": [ + 130, + 449, + 482, + 668 + ], + "type": "text", + "content": ". As illustrated in Fig. 4, it technically comprises two networks: the FlowNet " + }, + { + "bbox": [ + 130, + 449, + 482, + 668 + ], + "type": "inline_equation", + "content": "(\\mathcal{F}_{Net})" + }, + { + "bbox": [ + 130, + 449, + 482, + 668 + ], + "type": "text", + "content": " and the WarpNet " + }, + { + "bbox": [ + 130, + 449, + 482, + 668 + ], + "type": "inline_equation", + "content": "(\\mathcal{W}_{Net})" + }, + { + "bbox": [ + 130, + 449, + 482, + 668 + ], + "type": "text", + "content": ". Ordinarily, the computation of optical flow involves two images. However, in this case, we initially have only one image - the source image - and that too in the latent domain. Thus, our goal is to learn the optical flow utilizing just one latent image. This is achieved via " + }, + { + "bbox": [ + 130, + 449, + 482, + 668 + ], + "type": "inline_equation", + "content": "\\mathcal{F}_{Net}" + }, + { + "bbox": [ + 130, + 449, + 482, + 668 + ], + "type": "text", + "content": ", conditioned on " + }, + { + "bbox": [ + 130, + 449, + 482, + 668 + ], + "type": "inline_equation", + "content": "c_{L}" + }, + { + "bbox": [ + 130, + 449, + 482, + 668 + ], + "type": "text", + "content": ", which guides it to generate a flow feature map in the intended direction with the hint of text and motion prompts. This condition is processed by the Injection Module " + }, + { + "bbox": [ + 130, + 449, + 482, + 668 + ], + "type": "inline_equation", + "content": "(\\mathcal{I}_{\\mathcal{M}})" + }, + { + "bbox": [ + 130, + 449, + 482, + 668 + ], + "type": "text", + "content": ", a compact encoder designed to make " + }, + { + "bbox": [ + 130, + 449, + 482, + 668 + ], + "type": "inline_equation", + "content": "c_{L}" + }, + { + "bbox": [ + 130, + 449, + 482, + 668 + ], + "type": "text", + "content": " compatible for concatenation with one of the intermediate feature maps near the end of the network. This configuration enables " + }, + { + "bbox": [ + 130, + 449, + 482, + 668 + ], + "type": "inline_equation", + "content": "\\mathcal{F}_{Net}" + }, + { + "bbox": [ + 130, + 449, + 482, + 668 + ], + "type": "text", + "content": " to independently learn high-level features, which are then guided in the desired direction with the introduction of " + }, + { + "bbox": [ + 130, + 449, + 482, + 668 + ], + "type": "inline_equation", + "content": "c_{L}" + }, + { + "bbox": [ + 130, + 449, + 482, + 668 + ], + "type": "text", + "content": ". The architecture of " + }, + { + "bbox": [ + 130, + 449, + 482, + 668 + ], + "type": "inline_equation", + "content": "\\mathcal{F}_{Net}" + }, + { + "bbox": [ + 130, + 449, + 482, + 668 + ], + "type": "text", + "content": " resembles that of UNet, producing an output with a fixed channel of 2 to capture changes in the horizontal and vertical components. This optical flow feature map " + }, + { + "bbox": [ + 130, + 449, + 482, + 668 + ], + "type": "inline_equation", + "content": "F = \\mathcal{F}_{Net}(\\mathcal{E}(c_{I}), \\mathcal{I}_{\\mathcal{M}}(c_{L}))" + }, + { + "bbox": [ + 130, + 449, + 482, + 668 + ], + "type": "text", + "content": ", along with the source latent " + }, + { + "bbox": [ + 130, + 449, + 482, + 668 + ], + "type": "inline_equation", + "content": "(\\mathcal{E}(c_{I}))" + }, + { + "bbox": [ + 130, + 449, + 482, + 668 + ], + "type": "text", + "content": ", is then processed through " + }, + { + "bbox": [ + 130, + 449, + 482, + 668 + ], + "type": "inline_equation", + "content": "\\mathcal{W}_{Net}" + }, + { + "bbox": [ + 130, + 449, + 482, + 668 + ], + "type": "text", + "content": " to yield a Fischer map " + }, + { + "bbox": [ + 130, + 449, + 482, + 668 + ], + "type": "inline_equation", + "content": "z_{W} = \\mathcal{W}_{Net}(\\mathcal{E}(c_{I}), F)" + }, + { + "bbox": [ + 130, + 449, + 482, + 668 + ], + "type": "text", + "content": ". This transformation is learned more efficiently and abstractly in the latent space than in the pixel space." + } + ] + } + ], + "index": 5 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "spans": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "type": "text", + "content": "8" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 166, + 91, + 237, + 100 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 166, + 91, + 237, + 100 + ], + "spans": [ + { + "bbox": [ + 166, + 91, + 237, + 100 + ], + "type": "text", + "content": "H. Kandala et al." + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 7 + }, + { + "para_blocks": [ + { + "type": "table", + "bbox": [ + 170, + 146, + 443, + 210 + ], + "blocks": [ + { + "bbox": [ + 132, + 114, + 479, + 137 + ], + "lines": [ + { + "bbox": [ + 132, + 114, + 479, + 137 + ], + "spans": [ + { + "bbox": [ + 132, + 114, + 479, + 137 + ], + "type": "text", + "content": "Table 1: Quantitative comparison with state-of-the-art image-text-to-video generation models for the zero-shot setting." + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 170, + 146, + 443, + 210 + ], + "lines": [ + { + "bbox": [ + 170, + 146, + 443, + 210 + ], + "spans": [ + { + "bbox": [ + 170, + 146, + 443, + 210 + ], + "type": "table", + "html": "
MethodUCF-101 [52]MSR-VTT [64]
FVD ↓CLIPSim ↑PIC ↑FVD ↓CLIPSim ↑PIC ↑
I2VGen-XL [70]563.120.28650.6329278.620.22720.6018
DynamiCrafter [63]527.060.27960.6307271.630.26020.6135
Pix2Gif (Ours)285.020.28150.8763168.690.25730.8521
", + "image_path": "0fd4bb3a5606307054604849bb0be542a56b53e6c09a2fe793d7e1c5b7838c89.jpg" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "table_body" + } + ], + "index": 3 + }, + { + "bbox": [ + 132, + 239, + 192, + 250 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 239, + 192, + 250 + ], + "spans": [ + { + "bbox": [ + 132, + 239, + 192, + 250 + ], + "type": "text", + "content": "3.4 Losses" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 130, + 258, + 479, + 293 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 258, + 479, + 293 + ], + "spans": [ + { + "bbox": [ + 130, + 258, + 479, + 293 + ], + "type": "text", + "content": "Our model incorporates two different types of losses. The first type is the standard L2 loss Eq. (2), which is utilized by the stable diffusion model and talked about in Sec. 3.2" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 130, + 294, + 480, + 365 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 294, + 480, + 365 + ], + "spans": [ + { + "bbox": [ + 130, + 294, + 480, + 365 + ], + "type": "text", + "content": "The second loss type in our model is the perceptual loss, which is calculated by comparing the latent features of the image condition " + }, + { + "bbox": [ + 130, + 294, + 480, + 365 + ], + "type": "inline_equation", + "content": "\\mathcal{E}(c_I)" + }, + { + "bbox": [ + 130, + 294, + 480, + 365 + ], + "type": "text", + "content": " and the warped image " + }, + { + "bbox": [ + 130, + 294, + 480, + 365 + ], + "type": "inline_equation", + "content": "z_W" + }, + { + "bbox": [ + 130, + 294, + 480, + 365 + ], + "type": "text", + "content": ". This is implemented using a pre-trained VGG network [49], modified to accommodate 4 channels instead of the standard 3. The modification involves averaging the weights from the first three channels to initialize the fourth. The perceptual loss, " + }, + { + "bbox": [ + 130, + 294, + 480, + 365 + ], + "type": "inline_equation", + "content": "L_{p}" + }, + { + "bbox": [ + 130, + 294, + 480, + 365 + ], + "type": "text", + "content": ", is defined as:" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 205, + 375, + 480, + 399 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 205, + 375, + 480, + 399 + ], + "spans": [ + { + "bbox": [ + 205, + 375, + 480, + 399 + ], + "type": "interline_equation", + "content": "L _ {p} \\left(\\mathcal {E} \\left(c _ {I}\\right), z _ {W}\\right) = \\sum_ {k} \\lambda_ {k} \\left\\| \\phi_ {k} \\left(\\mathcal {E} \\left(c _ {I}\\right)\\right) - \\phi_ {k} \\left(z _ {W}\\right) \\right\\| ^ {2} \\tag {3}", + "image_path": "69a545f3f8cc50c2dde747adda450c3b731651c72158ad4648f3f3e3994e8e6e.jpg" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 130, + 407, + 480, + 467 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 407, + 480, + 467 + ], + "spans": [ + { + "bbox": [ + 130, + 407, + 480, + 467 + ], + "type": "text", + "content": "Here, " + }, + { + "bbox": [ + 130, + 407, + 480, + 467 + ], + "type": "inline_equation", + "content": "\\phi_k(.)" + }, + { + "bbox": [ + 130, + 407, + 480, + 467 + ], + "type": "text", + "content": " be the feature map of the k-th layer of the VGG network, " + }, + { + "bbox": [ + 130, + 407, + 480, + 467 + ], + "type": "inline_equation", + "content": "||.||" + }, + { + "bbox": [ + 130, + 407, + 480, + 467 + ], + "type": "text", + "content": " denotes the Frobenius norm, and " + }, + { + "bbox": [ + 130, + 407, + 480, + 467 + ], + "type": "inline_equation", + "content": "\\lambda_{k}" + }, + { + "bbox": [ + 130, + 407, + 480, + 467 + ], + "type": "text", + "content": " is a weighting factor. This loss ensures that the warped image maintains high-level features like edges, textures, and object types, making the images more perceptually and semantically similar and preserving the overall structure of the source image." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 130, + 468, + 480, + 491 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 468, + 480, + 491 + ], + "spans": [ + { + "bbox": [ + 130, + 468, + 480, + 491 + ], + "type": "text", + "content": "In conclusion, the total loss function, denoted as " + }, + { + "bbox": [ + 130, + 468, + 480, + 491 + ], + "type": "inline_equation", + "content": "L_{T}" + }, + { + "bbox": [ + 130, + 468, + 480, + 491 + ], + "type": "text", + "content": ", for our objective is computed by a weighted sum of the two individual losses." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 260, + 503, + 480, + 517 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 260, + 503, + 480, + 517 + ], + "spans": [ + { + "bbox": [ + 260, + 503, + 480, + 517 + ], + "type": "interline_equation", + "content": "L _ {T} = L _ {L D M} ^ {\\prime} + \\lambda_ {P} L _ {P} \\tag {4}", + "image_path": "777aee7d13b7745dc173f8fc68d7f959e815144cfe4cbf7dd3bbf08dc97f43e9.jpg" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 130, + 521, + 479, + 556 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 521, + 479, + 556 + ], + "spans": [ + { + "bbox": [ + 130, + 521, + 479, + 556 + ], + "type": "text", + "content": "Here, " + }, + { + "bbox": [ + 130, + 521, + 479, + 556 + ], + "type": "inline_equation", + "content": "\\lambda_{P}" + }, + { + "bbox": [ + 130, + 521, + 479, + 556 + ], + "type": "text", + "content": " is the weighting factor for perceptual loss. These two losses together provide a holistic framework to train our model by ensuring pixel-level accuracy, preservation of high-level features, and smooth motion transitions." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 132, + 574, + 230, + 588 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 574, + 230, + 588 + ], + "spans": [ + { + "bbox": [ + 132, + 574, + 230, + 588 + ], + "type": "text", + "content": "4 Experiments" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 132, + 598, + 190, + 610 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 598, + 190, + 610 + ], + "spans": [ + { + "bbox": [ + 132, + 598, + 190, + 610 + ], + "type": "text", + "content": "4.1 Setup" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 130, + 617, + 480, + 666 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 617, + 480, + 666 + ], + "spans": [ + { + "bbox": [ + 130, + 617, + 480, + 666 + ], + "type": "text", + "content": "Datasets We utilize the Tumbled GIF (TGIF) dataset for our training and validation purposes as discussed in Sec. 3.1. We evaluate our model on two datasets: MSR-VTT [64] and UCF-101 [52], following the common practice. For these datasets, we follow the sampling strategy as outlined in [63]." + } + ] + } + ], + "index": 14 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 227, + 91, + 447, + 101 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 227, + 91, + 447, + 101 + ], + "spans": [ + { + "bbox": [ + 227, + 91, + 447, + 101 + ], + "type": "text", + "content": "Pix2Gif: Motion-Guided Diffusion for GIF Generation" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 474, + 92, + 480, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 474, + 92, + 480, + 100 + ], + "spans": [ + { + "bbox": [ + 474, + 92, + 480, + 100 + ], + "type": "text", + "content": "9" + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 8 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 177, + 114, + 438, + 241 + ], + "blocks": [ + { + "bbox": [ + 177, + 114, + 438, + 241 + ], + "lines": [ + { + "bbox": [ + 177, + 114, + 438, + 241 + ], + "spans": [ + { + "bbox": [ + 177, + 114, + 438, + 241 + ], + "type": "image", + "image_path": "55bcb091bbb48343b53c9a576aa3723491ca2d0e3e4ee295e700cb9bbe4c368e.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 301, + 243, + 313, + 251 + ], + "lines": [ + { + "bbox": [ + 301, + 243, + 313, + 251 + ], + "spans": [ + { + "bbox": [ + 301, + 243, + 313, + 251 + ], + "type": "text", + "content": "(a)" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_caption" + } + ], + "index": 2 + }, + { + "type": "image", + "bbox": [ + 186, + 252, + 430, + 385 + ], + "blocks": [ + { + "bbox": [ + 186, + 252, + 430, + 385 + ], + "lines": [ + { + "bbox": [ + 186, + 252, + 430, + 385 + ], + "spans": [ + { + "bbox": [ + 186, + 252, + 430, + 385 + ], + "type": "image", + "image_path": "efed11d51d9f0a373847c70abec6cd96a80958ee630596d275f4c359dff3a39e.jpg" + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 301, + 387, + 313, + 397 + ], + "lines": [ + { + "bbox": [ + 301, + 387, + 313, + 397 + ], + "spans": [ + { + "bbox": [ + 301, + 387, + 313, + 397 + ], + "type": "text", + "content": "(b)" + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 131, + 406, + 482, + 439 + ], + "lines": [ + { + "bbox": [ + 131, + 406, + 482, + 439 + ], + "spans": [ + { + "bbox": [ + 131, + 406, + 482, + 439 + ], + "type": "text", + "content": "Fig. 5: Comparison studies with other image-text to video models. Given a source image and a caption, frames are extracted from the generated 16-frame video at 256x256 resolution." + } + ] + } + ], + "index": 6, + "angle": 0, + "type": "image_caption" + } + ], + "index": 4 + }, + { + "bbox": [ + 130, + 463, + 482, + 536 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 463, + 482, + 536 + ], + "spans": [ + { + "bbox": [ + 130, + 463, + 482, + 536 + ], + "type": "text", + "content": "**Implementations** Our model is initialized with the exponential moving average (EMA) weights of the Stable Diffusion v1.5 checkpoint and the improved ftMSE autoencoder weights" + }, + { + "bbox": [ + 130, + 463, + 482, + 536 + ], + "type": "inline_equation", + "content": "^2" + }, + { + "bbox": [ + 130, + 463, + 482, + 536 + ], + "type": "text", + "content": ". We trained the model at 256x256 resolution for 7 epochs on a single node of 16 V100 GPUs for 25k steps. We used the AdamW optimizer with a learning rate of " + }, + { + "bbox": [ + 130, + 463, + 482, + 536 + ], + "type": "inline_equation", + "content": "10^{-4}" + }, + { + "bbox": [ + 130, + 463, + 482, + 536 + ], + "type": "text", + "content": ". We set the weighting factor for perceptual loss " + }, + { + "bbox": [ + 130, + 463, + 482, + 536 + ], + "type": "inline_equation", + "content": "(\\lambda_P)" + }, + { + "bbox": [ + 130, + 463, + 482, + 536 + ], + "type": "text", + "content": " as " + }, + { + "bbox": [ + 130, + 463, + 482, + 536 + ], + "type": "inline_equation", + "content": "10^{-2}" + }, + { + "bbox": [ + 130, + 463, + 482, + 536 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 130, + 552, + 482, + 615 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 552, + 482, + 615 + ], + "spans": [ + { + "bbox": [ + 130, + 552, + 482, + 615 + ], + "type": "text", + "content": "Metrics We report Frechet Video Distance (FVD) [54], CLIP Similarity (CLIP-Sim) which is the average similarity calculated for all the generated frames with the input caption and Perceptual Input Conformity (PIC) as described in [63] for all methods. For comparison, we assess the zero-shot generation performance on I2VGen-XL [70] and DynamiCrafter [63]." + } + ] + } + ], + "index": 8 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "spans": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "type": "text", + "content": "10" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 166, + 91, + 237, + 100 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 166, + 91, + 237, + 100 + ], + "spans": [ + { + "bbox": [ + 166, + 91, + 237, + 100 + ], + "type": "text", + "content": "H. Kandala et al." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 133, + 620, + 468, + 643 + ], + "type": "page_footnote", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 620, + 468, + 643 + ], + "spans": [ + { + "bbox": [ + 133, + 620, + 468, + 643 + ], + "type": "text", + "content": "1 https://huggingface.co/runwayml/stable-diffusion-v1-5/blob/main/v1-5-pruned-emaonly.ckpt" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 134, + 643, + 466, + 665 + ], + "type": "page_footnote", + "angle": 0, + "lines": [ + { + "bbox": [ + 134, + 643, + 466, + 665 + ], + "spans": [ + { + "bbox": [ + 134, + 643, + 466, + 665 + ], + "type": "inline_equation", + "content": "^{2}" + }, + { + "bbox": [ + 134, + 643, + 466, + 665 + ], + "type": "text", + "content": " https://huggingface.co/stabilityai/sd-vae-ft-mse-original/blob/main/vae-ft-mse-840000-ema-pruned.ckpt" + } + ] + } + ], + "index": 10 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 9 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 138, + 215, + 245, + 247 + ], + "blocks": [ + { + "bbox": [ + 134, + 205, + 247, + 215 + ], + "lines": [ + { + "bbox": [ + 134, + 205, + 247, + 215 + ], + "spans": [ + { + "bbox": [ + 134, + 205, + 247, + 215 + ], + "type": "text", + "content": "(a) A1: Cat is playing with wool" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 138, + 215, + 245, + 247 + ], + "lines": [ + { + "bbox": [ + 138, + 215, + 245, + 247 + ], + "spans": [ + { + "bbox": [ + 138, + 215, + 245, + 247 + ], + "type": "image", + "image_path": "a6d06ceb5f7ff589d8201b917c88fd5558c9f8e007881344ecd2ad3cc1365967.jpg" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 131, + 256, + 480, + 279 + ], + "lines": [ + { + "bbox": [ + 131, + 256, + 480, + 279 + ], + "spans": [ + { + "bbox": [ + 131, + 256, + 480, + 279 + ], + "type": "text", + "content": "Fig. 6: Pix2Gif showing composition capabilities for different types of motions. [GIFs best viewed in Adobe Acrobat Reader]" + } + ] + } + ], + "index": 8, + "angle": 0, + "type": "image_caption" + } + ], + "index": 3 + }, + { + "type": "image", + "bbox": [ + 253, + 215, + 360, + 247 + ], + "blocks": [ + { + "bbox": [ + 266, + 205, + 348, + 215 + ], + "lines": [ + { + "bbox": [ + 266, + 205, + 348, + 215 + ], + "spans": [ + { + "bbox": [ + 266, + 205, + 348, + 215 + ], + "type": "text", + "content": "(b) A2: Cat is dancing" + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 253, + 215, + 360, + 247 + ], + "lines": [ + { + "bbox": [ + 253, + 215, + 360, + 247 + ], + "spans": [ + { + "bbox": [ + 253, + 215, + 360, + 247 + ], + "type": "image", + "image_path": "6d1df9846d32c3c417e770db8893e2c700ccac1972cad12bd3931a448c233f9b.jpg" + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "image_body" + } + ], + "index": 5 + }, + { + "type": "image", + "bbox": [ + 369, + 215, + 476, + 247 + ], + "blocks": [ + { + "bbox": [ + 366, + 198, + 480, + 215 + ], + "lines": [ + { + "bbox": [ + 366, + 198, + 480, + 215 + ], + "spans": [ + { + "bbox": [ + 366, + 198, + 480, + 215 + ], + "type": "text", + "content": "(c) A1+2: Cat dancing and playing with wool" + } + ] + } + ], + "index": 6, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 369, + 215, + 476, + 247 + ], + "lines": [ + { + "bbox": [ + 369, + 215, + 476, + 247 + ], + "spans": [ + { + "bbox": [ + 369, + 215, + 476, + 247 + ], + "type": "image", + "image_path": "6521da9d68a5bcc4e12d333381ba148d366c1bb0578f7e2a2db390f83c915c1b.jpg" + } + ] + } + ], + "index": 7, + "angle": 0, + "type": "image_body" + } + ], + "index": 7 + }, + { + "type": "table", + "bbox": [ + 162, + 322, + 452, + 385 + ], + "blocks": [ + { + "bbox": [ + 131, + 290, + 480, + 313 + ], + "lines": [ + { + "bbox": [ + 131, + 290, + 480, + 313 + ], + "spans": [ + { + "bbox": [ + 131, + 290, + 480, + 313 + ], + "type": "text", + "content": "Table 2: Ablation study comparing image translation methods with a focus on motion coherency at varyingcfg_img values." + } + ] + } + ], + "index": 9, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 162, + 322, + 452, + 385 + ], + "lines": [ + { + "bbox": [ + 162, + 322, + 452, + 385 + ], + "spans": [ + { + "bbox": [ + 162, + 322, + 452, + 385 + ], + "type": "table", + "html": "
Method /cfgimg1.41.61.82.0
L2 ↓PCC ↑L2 ↓PCC ↑L2 ↓PCC ↑L2 ↓PCC ↑
InstructPix2Pix [6]23.429-0.22925.492-0.02827.037-0.42327.5300.139
Pix2Gif-Base7.5800.9895.1880.9875.5950.9927.0290.991
Pix2Gif1.7460.9951.9720.9952.9440.9974.0760.997
", + "image_path": "fa3da07e371c75d1c5989644694901d6ebc782a3e98b4674677da903453eaea1.jpg" + } + ] + } + ], + "index": 10, + "angle": 0, + "type": "table_body" + } + ], + "index": 10 + }, + { + "bbox": [ + 132, + 418, + 198, + 428 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 418, + 198, + 428 + ], + "spans": [ + { + "bbox": [ + 132, + 418, + 198, + 428 + ], + "type": "text", + "content": "4.2 Results" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 130, + 437, + 480, + 474 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 437, + 480, + 474 + ], + "spans": [ + { + "bbox": [ + 130, + 437, + 480, + 474 + ], + "type": "text", + "content": "Comparisons with previous works Fig. 5 and Tab. 1 provide a qualitative and quantitative comparison of three image-text to video models: I2VGenXL [70], DynamiCrafter [63], and our " + }, + { + "bbox": [ + 130, + 437, + 480, + 474 + ], + "type": "inline_equation", + "content": "Pix2Gif" + }, + { + "bbox": [ + 130, + 437, + 480, + 474 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 130, + 474, + 481, + 629 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 474, + 481, + 629 + ], + "spans": [ + { + "bbox": [ + 130, + 474, + 481, + 629 + ], + "type": "text", + "content": "In Fig. 5a, the I2VGen-XL model misshapes the dog's face and generates it sideways in a nonsensical manner. DynamiCrafter appears to disregard the input parameters, as the initial frame differs significantly in position, color, and texture. It is also challenging to discern whether the dog is eating or merely moving its mouth. Our model, Pix2Gif, accurately retains all the dog's details and successfully depicts it eating from a plate. In Fig. 5b, we assess the models' capabilities by generating a video from a relatively dark image. Once again, I2VGen-XL starts strong, producing some impressive frames, but these soon turn into highly stylized and improbable images. DynamiCrafter appears to misinterpret the input image, generating something significantly different, although it seems to adhere to the caption. Conversely, Pix2Gif comprehends the inputs effectively and produces corresponding motion while preserving the overall integrity of the source image." + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 130, + 629, + 481, + 665 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 629, + 481, + 665 + ], + "spans": [ + { + "bbox": [ + 130, + 629, + 481, + 665 + ], + "type": "text", + "content": "Quantitatively, " + }, + { + "bbox": [ + 130, + 629, + 481, + 665 + ], + "type": "inline_equation", + "content": "\\text{Pix2Gif}" + }, + { + "bbox": [ + 130, + 629, + 481, + 665 + ], + "type": "text", + "content": " excels in both the FVD and PIC metrics shown in Tab. 1, which aligns with our observations of the frames generated in Fig. 5. These frames effectively preserve the structure and closely adhere to the input" + } + ] + } + ], + "index": 14 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 227, + 91, + 447, + 101 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 227, + 91, + 447, + 101 + ], + "spans": [ + { + "bbox": [ + 227, + 91, + 447, + 101 + ], + "type": "text", + "content": "Pix2Gif: Motion-Guided Diffusion for GIF Generation" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 470, + 91, + 479, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 470, + 91, + 479, + 100 + ], + "spans": [ + { + "bbox": [ + 470, + 91, + 479, + 100 + ], + "type": "text", + "content": "11" + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 10 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 164, + 119, + 271, + 202 + ], + "blocks": [ + { + "bbox": [ + 164, + 119, + 271, + 202 + ], + "lines": [ + { + "bbox": [ + 164, + 119, + 271, + 202 + ], + "spans": [ + { + "bbox": [ + 164, + 119, + 271, + 202 + ], + "type": "image", + "image_path": "20380b7b9b5f5c420b65199024486c10deadb7700bbedbc0f1e37c1fda8ee609.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 181, + 208, + 261, + 217 + ], + "lines": [ + { + "bbox": [ + 181, + 208, + 261, + 217 + ], + "spans": [ + { + "bbox": [ + 181, + 208, + 261, + 217 + ], + "type": "text", + "content": "(a) source frame score" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_caption" + } + ], + "index": 2 + }, + { + "type": "image", + "bbox": [ + 336, + 119, + 444, + 202 + ], + "blocks": [ + { + "bbox": [ + 336, + 119, + 444, + 202 + ], + "lines": [ + { + "bbox": [ + 336, + 119, + 444, + 202 + ], + "spans": [ + { + "bbox": [ + 336, + 119, + 444, + 202 + ], + "type": "image", + "image_path": "a809d8a20838bf48e9a5d83ec4872530839b98b33196e04079032818808d3ed3.jpg" + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 353, + 208, + 433, + 217 + ], + "lines": [ + { + "bbox": [ + 353, + 208, + 433, + 217 + ], + "spans": [ + { + "bbox": [ + 353, + 208, + 433, + 217 + ], + "type": "text", + "content": "(b) target frame score" + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 131, + 226, + 480, + 249 + ], + "lines": [ + { + "bbox": [ + 131, + 226, + 480, + 249 + ], + "spans": [ + { + "bbox": [ + 131, + 226, + 480, + 249 + ], + "type": "text", + "content": "Fig. 7: Ablation study between the earlier variants of our model by comparing average similarity score for 100 samples." + } + ] + } + ], + "index": 6, + "angle": 0, + "type": "image_caption" + } + ], + "index": 4 + }, + { + "bbox": [ + 130, + 272, + 482, + 404 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 272, + 482, + 404 + ], + "spans": [ + { + "bbox": [ + 130, + 272, + 482, + 404 + ], + "type": "text", + "content": "prompts (source image and caption). However, Pix2Gif does not perform as well in the CLIPSim metric, despite accurately following the caption. The other two models as seen in Fig. 5 do follow the caption, but they fail to adhere to the input image and produce plausible temporal transitions. This is partially attributed to the inherent model design in these two methods. Both methods attempt to generate a full sequence of frames at once using the 3D diffusion network, which inevitably compounds the spatial and temporal dimensions. Moreover, the results indicate that they function more as text-to-video models than image-text-to-video models, especially DynamiCrafter. This discrepancy also raises questions about the effectiveness of the CLIPSim metric for evaluating image-text-to-video models and calls for more sophisticated metrics for evaluating video generation." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 130, + 415, + 482, + 536 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 415, + 482, + 536 + ], + "spans": [ + { + "bbox": [ + 130, + 415, + 482, + 536 + ], + "type": "text", + "content": "Compositionality of actions Fig. 6 illustrates an intriguing emerging capability of " + }, + { + "bbox": [ + 130, + 415, + 482, + 536 + ], + "type": "inline_equation", + "content": "\\text{Pix2Gif}" + }, + { + "bbox": [ + 130, + 415, + 482, + 536 + ], + "type": "text", + "content": ": the ability to combine actions. In Fig. 6a, we see a cat playing with wool, with only the cat's paws and the wool moving. In Fig. 6b, we instruct the cat to dance, resulting in the cat moving its body but the wool remaining still. Finally, in Fig. 6c, we provide a caption that blends the actions from Fig. 6a and Fig. 6b. The result is a scene where the cat is both moving the wool and its body. This demonstrates " + }, + { + "bbox": [ + 130, + 415, + 482, + 536 + ], + "type": "inline_equation", + "content": "\\text{Pix2Gif}" + }, + { + "bbox": [ + 130, + 415, + 482, + 536 + ], + "type": "text", + "content": "'s ability to comprehend the caption and its associated motion, and to convert that understanding into a GIF. Such compositional capability significantly increases user controllability, a crucial aspect for practical applications." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 132, + 553, + 209, + 564 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 553, + 209, + 564 + ], + "spans": [ + { + "bbox": [ + 132, + 553, + 209, + 564 + ], + "type": "text", + "content": "4.3 Ablations" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 132, + 574, + 402, + 586 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 574, + 402, + 586 + ], + "spans": [ + { + "bbox": [ + 132, + 574, + 402, + 586 + ], + "type": "text", + "content": "We design a few variants of " + }, + { + "bbox": [ + 132, + 574, + 402, + 586 + ], + "type": "inline_equation", + "content": "\\text{Pix2Gif}" + }, + { + "bbox": [ + 132, + 574, + 402, + 586 + ], + "type": "text", + "content": " for our ablation studies:" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 137, + 593, + 480, + 664 + ], + "type": "list", + "angle": 0, + "index": 14, + "blocks": [ + { + "bbox": [ + 137, + 593, + 479, + 616 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 137, + 593, + 479, + 616 + ], + "spans": [ + { + "bbox": [ + 137, + 593, + 479, + 616 + ], + "type": "text", + "content": "- Pix2Gif-Base: We train InstructPix2Pix with our data, and append the text prompt with \"The optical flow is \"." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 137, + 617, + 480, + 640 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 137, + 617, + 480, + 640 + ], + "spans": [ + { + "bbox": [ + 137, + 617, + 480, + 640 + ], + "type": "text", + "content": "- Pix2Gif-Motion-embed: The motion embedding layer is added to encode the motion magnitude and combined with textual embedding." + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 137, + 641, + 479, + 664 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 137, + 641, + 479, + 664 + ], + "spans": [ + { + "bbox": [ + 137, + 641, + 479, + 664 + ], + "type": "text", + "content": "- Pix2Gif-Warp: We further add the warping module into the model but differently only use the warped feature for the LDM." + } + ] + } + ], + "index": 13 + } + ], + "sub_type": "text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "spans": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "type": "text", + "content": "12" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 166, + 91, + 237, + 100 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 166, + 91, + 237, + 100 + ], + "spans": [ + { + "bbox": [ + 166, + 91, + 237, + 100 + ], + "type": "text", + "content": "H. Kandala et al." + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 11 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 165, + 119, + 271, + 202 + ], + "blocks": [ + { + "bbox": [ + 165, + 119, + 271, + 202 + ], + "lines": [ + { + "bbox": [ + 165, + 119, + 271, + 202 + ], + "spans": [ + { + "bbox": [ + 165, + 119, + 271, + 202 + ], + "type": "image", + "image_path": "5649367366230c207fa30b68016064fa7f2d2b333cfb6e6a9fc335676ce9c9e8.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 180, + 208, + 261, + 217 + ], + "lines": [ + { + "bbox": [ + 180, + 208, + 261, + 217 + ], + "spans": [ + { + "bbox": [ + 180, + 208, + 261, + 217 + ], + "type": "text", + "content": "(a) source frame score" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_caption" + } + ], + "index": 2 + }, + { + "type": "image", + "bbox": [ + 338, + 120, + 444, + 202 + ], + "blocks": [ + { + "bbox": [ + 338, + 120, + 444, + 202 + ], + "lines": [ + { + "bbox": [ + 338, + 120, + 444, + 202 + ], + "spans": [ + { + "bbox": [ + 338, + 120, + 444, + 202 + ], + "type": "image", + "image_path": "9283edcfb9834d1b90281ccb209697d04fd0c3a04a613e2ff8f8034374c42533.jpg" + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 353, + 208, + 433, + 217 + ], + "lines": [ + { + "bbox": [ + 353, + 208, + 433, + 217 + ], + "spans": [ + { + "bbox": [ + 353, + 208, + 433, + 217 + ], + "type": "text", + "content": "(b) target frame score" + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 131, + 226, + 480, + 249 + ], + "lines": [ + { + "bbox": [ + 131, + 226, + 480, + 249 + ], + "spans": [ + { + "bbox": [ + 131, + 226, + 480, + 249 + ], + "type": "text", + "content": "Fig. 8: Ablation study on ways to input " + }, + { + "bbox": [ + 131, + 226, + 480, + 249 + ], + "type": "inline_equation", + "content": "z_{W}" + }, + { + "bbox": [ + 131, + 226, + 480, + 249 + ], + "type": "text", + "content": " to LDM by comparing average similarity score for 100 samples." + } + ] + } + ], + "index": 6, + "angle": 0, + "type": "image_caption" + } + ], + "index": 4 + }, + { + "bbox": [ + 137, + 273, + 480, + 333 + ], + "type": "list", + "angle": 0, + "index": 9, + "blocks": [ + { + "bbox": [ + 137, + 273, + 479, + 297 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 137, + 273, + 479, + 297 + ], + "spans": [ + { + "bbox": [ + 137, + 273, + 479, + 297 + ], + "type": "text", + "content": "- Pix2Gif-Warp-add: Different from Pix2Gif-Warp, we instead add the warped feature and source image feature as input to the LDM." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 138, + 297, + 480, + 333 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 297, + 480, + 333 + ], + "spans": [ + { + "bbox": [ + 138, + 297, + 480, + 333 + ], + "type": "text", + "content": "- Pix2Gif-Warp-concat: Instead of adding in Pix2Gif-Warp-add, we concatenate the warped feature and source image feature as the input, but do not include the perceptual loss." + } + ] + } + ], + "index": 8 + } + ], + "sub_type": "text" + }, + { + "bbox": [ + 130, + 342, + 482, + 426 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 342, + 482, + 426 + ], + "spans": [ + { + "bbox": [ + 130, + 342, + 482, + 426 + ], + "type": "text", + "content": "For comparative studies with our model's variants, we generate an 8-frame video. We use the X-CLIP model [37] to extract features from our generated video, and CLIP to extract features from the source, target, and generated frames. The optimal range of CFG.img for best results is considered to be [1.6, 2.2]. Throughout this discussion, we evaluate our model's performance using four metrics, which we believe effectively measure the different aspects of generating motion through the image translation framework used in Pix2Gif." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 130, + 437, + 482, + 581 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 437, + 482, + 581 + ], + "spans": [ + { + "bbox": [ + 130, + 437, + 482, + 581 + ], + "type": "text", + "content": "Motion Coherency Our task is framed as an image translation problem with motion magnitude as a guide. We assess the motion quality or temporal coherence in the GIFs using L2 loss and Pearson Correlation Coefficient (PCC) for InstructPix2Pix, Pix2Gif-Base, and Pix2Gif. The L2 loss evaluates the match between the motion values of the generated frames and the actual inputs, while the PCC checks if they follow the same trend. These metrics are calculated between the input motion magnitude values and the optical flow values, derived from the source image and the generated frames. Our model exhibits the highest correlation and the lowest L2 loss across all CFG.img values as seen in Tab. 2, proving its efficacy and controllability in generating GIFs with specific motions. The Pix2Gif-Base outperforms the original InstructPix2Pix, emphasizing the importance of our new dataset." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 130, + 594, + 482, + 666 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 594, + 482, + 666 + ], + "spans": [ + { + "bbox": [ + 130, + 594, + 482, + 666 + ], + "type": "text", + "content": "Image-Video Similarity Score To evaluate the semantic properties of the video produced by our model, we created two similarity scores: a source frame score and a target frame score. The source frame score quantifies how well the video retains the primary attributes of the source frame, essentially measuring the accuracy of the source image portrayal throughout the video. The target frame score indicates the precision of the scene or subject development from" + } + ] + } + ], + "index": 12 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 227, + 91, + 447, + 101 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 227, + 91, + 447, + 101 + ], + "spans": [ + { + "bbox": [ + 227, + 91, + 447, + 101 + ], + "type": "text", + "content": "Pix2Gif: Motion-Guided Diffusion for GIF Generation" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 470, + 91, + 480, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 470, + 91, + 480, + 100 + ], + "spans": [ + { + "bbox": [ + 470, + 91, + 480, + 100 + ], + "type": "text", + "content": "13" + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 12 + }, + { + "para_blocks": [ + { + "bbox": [ + 130, + 116, + 479, + 152 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 116, + 479, + 152 + ], + "spans": [ + { + "bbox": [ + 130, + 116, + 479, + 152 + ], + "type": "text", + "content": "the source frame in the video. It also underscores the model's ability to handle uncertainty and potential changes, as the target frame represents a possible state that the video might reach." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 130, + 152, + 482, + 369 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 152, + 482, + 369 + ], + "spans": [ + { + "bbox": [ + 130, + 152, + 482, + 369 + ], + "type": "text", + "content": "We began our experiments with " + }, + { + "bbox": [ + 130, + 152, + 482, + 369 + ], + "type": "inline_equation", + "content": "\\text{Pix2Gif-Base}" + }, + { + "bbox": [ + 130, + 152, + 482, + 369 + ], + "type": "text", + "content": " and then enhanced it incrementally, first with " + }, + { + "bbox": [ + 130, + 152, + 482, + 369 + ], + "type": "inline_equation", + "content": "\\mathcal{M}" + }, + { + "bbox": [ + 130, + 152, + 482, + 369 + ], + "type": "text", + "content": " (Pix2Gif-Motion-embed), then by adding " + }, + { + "bbox": [ + 130, + 152, + 482, + 369 + ], + "type": "inline_equation", + "content": "\\mathcal{W}" + }, + { + "bbox": [ + 130, + 152, + 482, + 369 + ], + "type": "text", + "content": " (Pix2Gif-Warp). Here, we conducted an ablation study shown in Fig. 8 to understand the best way to feed " + }, + { + "bbox": [ + 130, + 152, + 482, + 369 + ], + "type": "inline_equation", + "content": "z_{W}" + }, + { + "bbox": [ + 130, + 152, + 482, + 369 + ], + "type": "text", + "content": " into LDM. Our experiments included feeding only " + }, + { + "bbox": [ + 130, + 152, + 482, + 369 + ], + "type": "inline_equation", + "content": "z_{W}" + }, + { + "bbox": [ + 130, + 152, + 482, + 369 + ], + "type": "text", + "content": " (Pix2Gif-Warp), adding it with the " + }, + { + "bbox": [ + 130, + 152, + 482, + 369 + ], + "type": "inline_equation", + "content": "\\mathcal{E}(c_{I})" + }, + { + "bbox": [ + 130, + 152, + 482, + 369 + ], + "type": "text", + "content": " (Pix2Gif-Warp-add), and concatenating both (Pix2Gif-Warp-concat). Both Pix2Gif-Warp-add and Pix2Gif-Warp-concat had higher average similarity scores than Pix2Gif-Warp, and Pix2Gif-Warp-concat performed the best. This can be attributed to the fact that in the addition process, " + }, + { + "bbox": [ + 130, + 152, + 482, + 369 + ], + "type": "inline_equation", + "content": "\\mathcal{E}(c_{I})" + }, + { + "bbox": [ + 130, + 152, + 482, + 369 + ], + "type": "text", + "content": " loses its unique characteristics, which are required by the diffusion model for effective unconditional denoising. Therefore, to achieve the best results, we combined " + }, + { + "bbox": [ + 130, + 152, + 482, + 369 + ], + "type": "inline_equation", + "content": "\\mathcal{E}(c_{I})" + }, + { + "bbox": [ + 130, + 152, + 482, + 369 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 130, + 152, + 482, + 369 + ], + "type": "inline_equation", + "content": "z_{W}" + }, + { + "bbox": [ + 130, + 152, + 482, + 369 + ], + "type": "text", + "content": " before inputting them into the concat attention layer of the LDM. Now finally we integrate " + }, + { + "bbox": [ + 130, + 152, + 482, + 369 + ], + "type": "inline_equation", + "content": "L_{P}" + }, + { + "bbox": [ + 130, + 152, + 482, + 369 + ], + "type": "text", + "content": " (Pix2Gif) and compare them in Fig. 7. As expected, Pix2Gif-Base was outperformed by the other versions. Pix2Gif generated more coherent and controlled motion, albeit often with limited extent. All models began to converge outside the optimal range, producing similar frames. The Pix2Gif-Motion-embed model creates a significant amount of one-directional motion, which can sometimes be nonsensical, and hence the addition of " + }, + { + "bbox": [ + 130, + 152, + 482, + 369 + ], + "type": "inline_equation", + "content": "\\mathcal{W}" + }, + { + "bbox": [ + 130, + 152, + 482, + 369 + ], + "type": "text", + "content": " helps to mitigate this issue." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 132, + 384, + 328, + 397 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 384, + 328, + 397 + ], + "spans": [ + { + "bbox": [ + 132, + 384, + 328, + 397 + ], + "type": "text", + "content": "5 Limitations and Future Work" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 130, + 408, + 482, + 517 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 408, + 482, + 517 + ], + "spans": [ + { + "bbox": [ + 130, + 408, + 482, + 517 + ], + "type": "text", + "content": "The current Pix2Gif model is our initial attempt to generate videos by treating it as an image translation task. However, this method has some limitations that prevent us from generating high-quality and long GIFs or videos. Firstly, the model generates images with a resolution of 256x256 pixels. If these images are used to generate subsequent frames, the quality of the frames deteriorates further. Secondly, due to limitations in computational power, we are only able to use a small portion of a larger, curated dataset for training our model. Our primary objective now is to improve the quality of the generated frames, as this could significantly enhance the effectiveness of this method." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 132, + 533, + 220, + 547 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 533, + 220, + 547 + ], + "spans": [ + { + "bbox": [ + 132, + 533, + 220, + 547 + ], + "type": "text", + "content": "6 Conclusion" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 130, + 557, + 482, + 666 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 557, + 482, + 666 + ], + "spans": [ + { + "bbox": [ + 130, + 557, + 482, + 666 + ], + "type": "text", + "content": "In this work, we proposed Pix2Gif, an image-to-GIF (video) generation model based on an image-to-image translation paradigm. To ensure temporal coherence across frames, we proposed a motion-guided warping module that learns to spatially warp the source image feature into the target one while maintaining visual consistency via a perceptual loss. Starting from TGIF, we curated a new dataset specifically used for training our model. The experimental results demonstrated the effectiveness of our model to generate GIFs with better temporal coherence compared with current state-of-the-art methods. Interestingly, the model also exhibits better controllability and some emerging action compositionality." + } + ] + } + ], + "index": 7 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "spans": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "type": "text", + "content": "14" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 166, + 91, + 237, + 100 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 166, + 91, + 237, + 100 + ], + "spans": [ + { + "bbox": [ + 166, + 91, + 237, + 100 + ], + "type": "text", + "content": "H. Kandala et al." + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 13 + }, + { + "para_blocks": [ + { + "bbox": [ + 133, + 114, + 197, + 126 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 114, + 197, + 126 + ], + "spans": [ + { + "bbox": [ + 133, + 114, + 197, + 126 + ], + "type": "text", + "content": "References" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 134, + 141, + 480, + 665 + ], + "type": "list", + "angle": 0, + "index": 21, + "blocks": [ + { + "bbox": [ + 138, + 141, + 480, + 164 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 141, + 480, + 164 + ], + "spans": [ + { + "bbox": [ + 138, + 141, + 480, + 164 + ], + "type": "text", + "content": "1. Aigner, S., Körner, M.: Futuregan: Anticipating the future frames of video sequences using spatio-temporal 3d convolutions in progressively growing gans (2018)" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 138, + 165, + 480, + 198 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 165, + 480, + 198 + ], + "spans": [ + { + "bbox": [ + 138, + 165, + 480, + 198 + ], + "type": "text", + "content": "2. Avrahami, O., Fried, O., Lischinski, D.: Blended latent diffusion. ACM Transactions on Graphics 42(4), 1-11 (Jul 2023). https://doi.org/10.1145/3592450, http://dx.doi.org/10.1145/3592450" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 138, + 198, + 480, + 220 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 198, + 480, + 220 + ], + "spans": [ + { + "bbox": [ + 138, + 198, + 480, + 220 + ], + "type": "text", + "content": "3. Baradaran, M., Bergevin, R.: Future video prediction from a single frame for video anomaly detection (2023)" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 138, + 220, + 480, + 253 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 220, + 480, + 253 + ], + "spans": [ + { + "bbox": [ + 138, + 220, + 480, + 253 + ], + "type": "text", + "content": "4. Blattmann, A., Dockhorn, T., Kulal, S., Mendelevitch, D., Kilian, M., Lorenz, D., Levi, Y., English, Z., Voleti, V., Letts, A., Jampani, V., Rombach, R.: Stable video diffusion: Scaling latent video diffusion models to large datasets (2023)" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 138, + 254, + 480, + 286 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 254, + 480, + 286 + ], + "spans": [ + { + "bbox": [ + 138, + 254, + 480, + 286 + ], + "type": "text", + "content": "5. Blattmann, A., Rombach, R., Ling, H., Dockhorn, T., Kim, S.W., Fidler, S., Kreis, K.: Align your latents: High-resolution video synthesis with latent diffusion models (2023)" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 138, + 287, + 480, + 319 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 287, + 480, + 319 + ], + "spans": [ + { + "bbox": [ + 138, + 287, + 480, + 319 + ], + "type": "text", + "content": "6. Brooks, T., Holynski, A., Efros, A.A.: Instructpix2pix: Learning to follow image editing instructions. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 18392-18402 (2023)" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 138, + 320, + 480, + 354 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 320, + 480, + 354 + ], + "spans": [ + { + "bbox": [ + 138, + 320, + 480, + 354 + ], + "type": "text", + "content": "7. Chang, H., Zhang, H., Barber, J., Maschinot, A., Lezama, J., Jiang, L., Yang, M.H., Murphy, K., Freeman, W.T., Rubinstein, M., Li, Y., Krishnan, D.: Muse: Text-to-image generation via masked generative transformers (2023)" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 138, + 354, + 480, + 376 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 354, + 480, + 376 + ], + "spans": [ + { + "bbox": [ + 138, + 354, + 480, + 376 + ], + "type": "text", + "content": "8. Chang, H., Zhang, H., Jiang, L., Liu, C., Freeman, W.T.: Maskgit: Masked generative image transformer (2022)" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 138, + 376, + 480, + 398 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 376, + 480, + 398 + ], + "spans": [ + { + "bbox": [ + 138, + 376, + 480, + 398 + ], + "type": "text", + "content": "9. Choi, J., Kim, S., Jeong, Y., Gwon, Y., Yoon, S.: Ilvr: Conditioning method for denoising diffusion probabilistic models (2021)" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 134, + 399, + 480, + 431 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 134, + 399, + 480, + 431 + ], + "spans": [ + { + "bbox": [ + 134, + 399, + 480, + 431 + ], + "type": "text", + "content": "10. Ding, M., Yang, Z., Hong, W., Zheng, W., Zhou, C., Yin, D., Lin, J., Zou, X., Shao, Z., Yang, H., et al.: Cogview: Mastering text-to-image generation via transformers. Advances in Neural Information Processing Systems 34, 19822-19835 (2021)" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 134, + 432, + 480, + 464 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 134, + 432, + 480, + 464 + ], + "spans": [ + { + "bbox": [ + 134, + 432, + 480, + 464 + ], + "type": "text", + "content": "11. Esser, P., Chiu, J., Atighehchian, P., Granskog, J., Germanidis, A.: Structure and content-guided video synthesis with diffusion models. In: Proceedings of the IEEE/CVF International Conference on Computer Vision. pp. 7346-7356 (2023)" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 134, + 465, + 480, + 487 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 134, + 465, + 480, + 487 + ], + "spans": [ + { + "bbox": [ + 134, + 465, + 480, + 487 + ], + "type": "text", + "content": "12. Esser, P., Rombach, R., Ommer, B.: Taming transformers for high-resolution image synthesis (2021)" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 134, + 487, + 480, + 510 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 134, + 487, + 480, + 510 + ], + "spans": [ + { + "bbox": [ + 134, + 487, + 480, + 510 + ], + "type": "text", + "content": "13. Fu, T.J., Hu, W., Du, X., Wang, W.Y., Yang, Y., Gan, Z.: Guiding instruction-based image editing via multimodal large language models (2023)" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 134, + 510, + 480, + 542 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 134, + 510, + 480, + 542 + ], + "spans": [ + { + "bbox": [ + 134, + 510, + 480, + 542 + ], + "type": "text", + "content": "14. Fujitake, M., Sugimoto, A.: Video representation learning through prediction for online object detection. In: Proceedings of the IEEE/CVF Winter Conference on Applications of Computer Vision. pp. 530-539 (2022)" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 134, + 543, + 480, + 576 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 134, + 543, + 480, + 576 + ], + "spans": [ + { + "bbox": [ + 134, + 543, + 480, + 576 + ], + "type": "text", + "content": "15. Ge, S., Nah, S., Liu, G., Poon, T., Tao, A., Catanzaro, B., Jacobs, D., Huang, J.B., Liu, M.Y., Balaji, Y.: Preserve your own correlation: A noise prior for video diffusion models (2023)" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 134, + 577, + 480, + 598 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 134, + 577, + 480, + 598 + ], + "spans": [ + { + "bbox": [ + 134, + 577, + 480, + 598 + ], + "type": "text", + "content": "16. Girdhar, R., Ramanan, D.: Cater: A diagnostic dataset for compositional actions and temporal reasoning (2020)" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 134, + 599, + 480, + 632 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 134, + 599, + 480, + 632 + ], + "spans": [ + { + "bbox": [ + 134, + 599, + 480, + 632 + ], + "type": "text", + "content": "17. Goodfellow, I., Pouget-Abadie, J., Mirza, M., Xu, B., Warde-Farley, D., Ozair, S., Courville, A., Bengio, Y.: Generative adversarial networks. Communications of the ACM 63(11), 139-144 (2020)" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 134, + 632, + 480, + 665 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 134, + 632, + 480, + 665 + ], + "spans": [ + { + "bbox": [ + 134, + 632, + 480, + 665 + ], + "type": "text", + "content": "18. Gu, Y., Yang, J., Usuyama, N., Li, C., Zhang, S., Lungren, M.P., Gao, J., Poon, H.: Biomedjourney: Counterfactual biomedical image generation by instruction-learning from multimodal patient journeys. arXiv preprint arXiv:2310.10765 (2023)" + } + ] + } + ], + "index": 20 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 227, + 91, + 447, + 101 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 227, + 91, + 447, + 101 + ], + "spans": [ + { + "bbox": [ + 227, + 91, + 447, + 101 + ], + "type": "text", + "content": "Pix2Gif: Motion-Guided Diffusion for GIF Generation" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 470, + 92, + 480, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 470, + 92, + 480, + 100 + ], + "spans": [ + { + "bbox": [ + 470, + 92, + 480, + 100 + ], + "type": "text", + "content": "15" + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 14 + }, + { + "para_blocks": [ + { + "bbox": [ + 132, + 116, + 480, + 665 + ], + "type": "list", + "angle": 0, + "index": 26, + "blocks": [ + { + "bbox": [ + 133, + 116, + 480, + 138 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 116, + 480, + 138 + ], + "spans": [ + { + "bbox": [ + 133, + 116, + 480, + 138 + ], + "type": "text", + "content": "19. He, Y., Yang, T., Zhang, Y., Shan, Y., Chen, Q.: Latent video diffusion models for high-fidelity long video generation (2023)" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 133, + 139, + 480, + 159 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 139, + 480, + 159 + ], + "spans": [ + { + "bbox": [ + 133, + 139, + 480, + 159 + ], + "type": "text", + "content": "20. Hertz, A., Mokady, R., Tenenbaum, J., Aberman, K., Pritch, Y., Cohen-Or, D.: Prompt-to-prompt image editing with cross attention control (2022)" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 132, + 160, + 454, + 171 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 160, + 454, + 171 + ], + "spans": [ + { + "bbox": [ + 132, + 160, + 454, + 171 + ], + "type": "text", + "content": "21. Ho, J., Jain, A., Abbeel, P.: Denoising diffusion probabilistic models (2020)" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 132, + 171, + 480, + 191 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 171, + 480, + 191 + ], + "spans": [ + { + "bbox": [ + 132, + 171, + 480, + 191 + ], + "type": "text", + "content": "22. Ho, J., Salimans, T.: Classifier-free diffusion guidance. arXiv preprint arXiv:2207.12598 (2022)" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 132, + 192, + 480, + 213 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 192, + 480, + 213 + ], + "spans": [ + { + "bbox": [ + 132, + 192, + 480, + 213 + ], + "type": "text", + "content": "23. Ho, J., Salimans, T., Gritsenko, A., Chan, W., Norouzi, M., Fleet, D.J.: Video diffusion models (2022)" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 132, + 214, + 480, + 235 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 214, + 480, + 235 + ], + "spans": [ + { + "bbox": [ + 132, + 214, + 480, + 235 + ], + "type": "text", + "content": "24. Hong, W., Ding, M., Zheng, W., Liu, X., Tang, J.: Cogvideo: Large-scale pretraining for text-to-video generation via transformers (2022)" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 132, + 236, + 480, + 256 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 236, + 480, + 256 + ], + "spans": [ + { + "bbox": [ + 132, + 236, + 480, + 256 + ], + "type": "text", + "content": "25. Hu, Y., Luo, C., Chen, Z.: Make it move: Controllable image-to-video generation with text descriptions (2022)" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 132, + 257, + 480, + 289 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 257, + 480, + 289 + ], + "spans": [ + { + "bbox": [ + 132, + 257, + 480, + 289 + ], + "type": "text", + "content": "26. Kang, M., Zhu, J.Y., Zhang, R., Park, J., Shechtman, E., Paris, S., Park, T.: Scaling up gans for text-to-image synthesis. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 10124-10134 (2023)" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 132, + 289, + 480, + 321 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 289, + 480, + 321 + ], + "spans": [ + { + "bbox": [ + 132, + 289, + 480, + 321 + ], + "type": "text", + "content": "27. Karras, T., Laine, S., Aila, T.: A style-based generator architecture for generative adversarial networks. In: Proceedings of the IEEE/CVF conference on computer vision and pattern recognition. pp. 4401-4410 (2019)" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 132, + 322, + 480, + 364 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 322, + 480, + 364 + ], + "spans": [ + { + "bbox": [ + 132, + 322, + 480, + 364 + ], + "type": "text", + "content": "28. Kawar, B., Zada, S., Lang, O., Tov, O., Chang, H., Dekel, T., Mosseri, I., Irani, M.: Imagic: Text-based real image editing with diffusion models. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 6007-6017 (2023)" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 132, + 365, + 422, + 375 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 365, + 422, + 375 + ], + "spans": [ + { + "bbox": [ + 132, + 365, + 422, + 375 + ], + "type": "text", + "content": "29. Kingma, D.P., Welling, M.: Auto-encoding variational bayes (2022)" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 132, + 376, + 480, + 396 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 376, + 480, + 396 + ], + "spans": [ + { + "bbox": [ + 132, + 376, + 480, + 396 + ], + "type": "text", + "content": "30. Lee, J., Lee, J., Lee, S., Yoon, S.: Mutual suppression network for video prediction using disentangled features (2019)" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 132, + 396, + 480, + 429 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 396, + 480, + 429 + ], + "spans": [ + { + "bbox": [ + 132, + 396, + 480, + 429 + ], + "type": "text", + "content": "31. Li, H., Yang, Y., Chang, M., Chen, S., Feng, H., Xu, Z., Li, Q., Chen, Y.: Srdiff: Single image super-resolution with diffusion probabilistic models. Neurocomputing 479, 47-59 (2022)" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 132, + 430, + 480, + 450 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 430, + 480, + 450 + ], + "spans": [ + { + "bbox": [ + 132, + 430, + 480, + 450 + ], + "type": "text", + "content": "32. Li, Y., Liu, H., Wu, Q., Mu, F., Yang, J., Gao, J., Li, C., Lee, Y.J.: Gligen: Open-set grounded text-to-image generation (2023)" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 132, + 451, + 480, + 472 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 451, + 480, + 472 + ], + "spans": [ + { + "bbox": [ + 132, + 451, + 480, + 472 + ], + "type": "text", + "content": "33. Li, Y., Song, Y., Cao, L., Tetreault, J., Goldberg, L., Jaimes, A., Luo, J.: Tgif: A new dataset and benchmark on animated gif description (2016)" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 132, + 472, + 480, + 493 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 472, + 480, + 493 + ], + "spans": [ + { + "bbox": [ + 132, + 472, + 480, + 493 + ], + "type": "text", + "content": "34. Liu, W., Luo, W., Lian, D., Gao, S.: Future frame prediction for anomaly detection - a new baseline (2018)" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 132, + 494, + 480, + 514 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 494, + 480, + 514 + ], + "spans": [ + { + "bbox": [ + 132, + 494, + 480, + 514 + ], + "type": "text", + "content": "35. Lu, C., Zhou, Y., Bao, F., Chen, J., Li, C., Zhu, J.: Dpm-solver: A fast ode solver for diffusion probabilistic model sampling in around 10 steps (2022)" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 132, + 515, + 480, + 536 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 515, + 480, + 536 + ], + "spans": [ + { + "bbox": [ + 132, + 515, + 480, + 536 + ], + "type": "text", + "content": "36. Meng, C., He, Y., Song, Y., Song, J., Wu, J., Zhu, J.Y., Ermon, S.: Sdedit: Guided image synthesis and editing with stochastic differential equations (2022)" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 132, + 536, + 480, + 557 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 536, + 480, + 557 + ], + "spans": [ + { + "bbox": [ + 132, + 536, + 480, + 557 + ], + "type": "text", + "content": "37. Ni, B., Peng, H., Chen, M., Zhang, S., Meng, G., Fu, J., Xiang, S., Ling, H.: Expanding language-image pretrained models for general video recognition (2022)" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 132, + 558, + 480, + 578 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 558, + 480, + 578 + ], + "spans": [ + { + "bbox": [ + 132, + 558, + 480, + 578 + ], + "type": "text", + "content": "38. Ni, H., Shi, C., Li, K., Huang, S.X., Min, M.R.: Conditional image-to-video generation with latent flow diffusion models (2023)" + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 132, + 579, + 480, + 601 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 579, + 480, + 601 + ], + "spans": [ + { + "bbox": [ + 132, + 579, + 480, + 601 + ], + "type": "text", + "content": "39. Oh, J., Guo, X., Lee, H., Lewis, R., Singh, S.: Action-conditional video prediction using deep networks in atari games (2015)" + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 132, + 601, + 480, + 621 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 601, + 480, + 621 + ], + "spans": [ + { + "bbox": [ + 132, + 601, + 480, + 621 + ], + "type": "text", + "content": "40. Oliu, M., Selva, J., Escalera, S.: Folded recurrent neural networks for future video prediction (2018)" + } + ] + } + ], + "index": 23 + }, + { + "bbox": [ + 132, + 622, + 480, + 643 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 622, + 480, + 643 + ], + "spans": [ + { + "bbox": [ + 132, + 622, + 480, + 643 + ], + "type": "text", + "content": "41. van den Oord, A., Kalchbrenner, N., Kavukcuoglu, K.: Pixel recurrent neural networks (2016)" + } + ] + } + ], + "index": 24 + }, + { + "bbox": [ + 132, + 643, + 480, + 665 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 643, + 480, + 665 + ], + "spans": [ + { + "bbox": [ + 132, + 643, + 480, + 665 + ], + "type": "text", + "content": "42. van den Oord, A., Kalchbrenner, N., Vinyals, O., Espeholt, L., Graves, A., Kavukcuoglu, K.: Conditional image generation with pixelCNN decoders (2016)" + } + ] + } + ], + "index": 25 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "spans": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "type": "text", + "content": "16" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 166, + 91, + 237, + 100 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 166, + 91, + 237, + 100 + ], + "spans": [ + { + "bbox": [ + 166, + 91, + 237, + 100 + ], + "type": "text", + "content": "H. Kandala et al." + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 15 + }, + { + "para_blocks": [ + { + "bbox": [ + 132, + 116, + 480, + 665 + ], + "type": "list", + "angle": 0, + "index": 23, + "blocks": [ + { + "bbox": [ + 132, + 116, + 480, + 138 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 116, + 480, + 138 + ], + "spans": [ + { + "bbox": [ + 132, + 116, + 480, + 138 + ], + "type": "text", + "content": "43. van den Oord, A., Vinyals, O., Kavukcuoglu, K.: Neural discrete representation learning (2018)" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 132, + 138, + 480, + 170 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 138, + 480, + 170 + ], + "spans": [ + { + "bbox": [ + 132, + 138, + 480, + 170 + ], + "type": "text", + "content": "44. Radford, A., Metz, L., Chintala, S.: Unsupervised representation learning with deep convolutional generative adversarial networks. arXiv preprint arXiv:1511.06434 (2015)" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 132, + 171, + 480, + 191 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 171, + 480, + 191 + ], + "spans": [ + { + "bbox": [ + 132, + 171, + 480, + 191 + ], + "type": "text", + "content": "45. Rombach, R., Blattmann, A., Lorenz, D., Esser, P., Ommer, B.: High-resolution image synthesis with latent diffusion models (2022)" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 132, + 192, + 480, + 224 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 192, + 480, + 224 + ], + "spans": [ + { + "bbox": [ + 132, + 192, + 480, + 224 + ], + "type": "text", + "content": "46. Ruiz, N., Li, Y., Jampani, V., Pritch, Y., Rubinstein, M., Aberman, K.: Dreambooth: Fine tuning text-to-image diffusion models for subject-driven generation (2023)" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 132, + 225, + 480, + 267 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 225, + 480, + 267 + ], + "spans": [ + { + "bbox": [ + 132, + 225, + 480, + 267 + ], + "type": "text", + "content": "47. Sahara, C., Chan, W., Saxena, S., Li, L., Whang, J., Denton, E.L., Ghasemipour, K., Gontijo Lopes, R., Karagol Ayan, B., Salimans, T., et al.: Photorealistic text-to-image diffusion models with deep language understanding. Advances in Neural Information Processing Systems 35, 36479-36494 (2022)" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 132, + 268, + 480, + 300 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 268, + 480, + 300 + ], + "spans": [ + { + "bbox": [ + 132, + 268, + 480, + 300 + ], + "type": "text", + "content": "48. Shi, X., Chen, Z., Wang, H., Yeung, D.Y., kin Wong, W., chun Woo, W.: Convolutional LSTM network: A machine learning approach for precipitation nowcasting (2015)" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 132, + 300, + 480, + 321 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 300, + 480, + 321 + ], + "spans": [ + { + "bbox": [ + 132, + 300, + 480, + 321 + ], + "type": "text", + "content": "49. Simonyan, K., Zisserman, A.: Very deep convolutional networks for large-scale image recognition (2015)" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 132, + 322, + 480, + 354 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 322, + 480, + 354 + ], + "spans": [ + { + "bbox": [ + 132, + 322, + 480, + 354 + ], + "type": "text", + "content": "50. Singer, U., Polyak, A., Hayes, T., Yin, X., An, J., Zhang, S., Hu, Q., Yang, H., Ashual, O., Gafni, O., Parikh, D., Gupta, S., Taigman, Y.: Make-a-video: Text-to-video generation without text-video data (2022)" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 132, + 354, + 447, + 364 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 354, + 447, + 364 + ], + "spans": [ + { + "bbox": [ + 132, + 354, + 447, + 364 + ], + "type": "text", + "content": "51. Song, J., Meng, C., Ermon, S.: Denoising diffusion implicit models (2022)" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 132, + 365, + 480, + 385 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 365, + 480, + 385 + ], + "spans": [ + { + "bbox": [ + 132, + 365, + 480, + 385 + ], + "type": "text", + "content": "52. Soomro, K., Zamir, A.R., Shah, M.: Ucf101: A dataset of 101 human actions classes from videos in the wild (2012)" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 132, + 386, + 480, + 407 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 386, + 480, + 407 + ], + "spans": [ + { + "bbox": [ + 132, + 386, + 480, + 407 + ], + "type": "text", + "content": "53. Srivastava, N., Mansimov, E., Salakhutdinov, R.: Unsupervised learning of video representations using lstms (2016)" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 132, + 407, + 480, + 428 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 407, + 480, + 428 + ], + "spans": [ + { + "bbox": [ + 132, + 407, + 480, + 428 + ], + "type": "text", + "content": "54. Unterthiner, T., van Steenkiste, S., Kurach, K., Marinier, R., Michalski, M., Gelly, S.: Fvd: A new metric for video generation (2019)" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 132, + 429, + 480, + 460 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 429, + 480, + 460 + ], + "spans": [ + { + "bbox": [ + 132, + 429, + 480, + 460 + ], + "type": "text", + "content": "55. Villegas, R., Babaeizadeh, M., Kindermans, P.J., Moraldo, H., Zhang, H., Saffar, M.T., Castro, S., Kunze, J., Erhan, D.: Phenaki: Variable length video generation from open domain textual description (2022)" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 132, + 461, + 480, + 482 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 461, + 480, + 482 + ], + "spans": [ + { + "bbox": [ + 132, + 461, + 480, + 482 + ], + "type": "text", + "content": "56. Villegas, R., Yang, J., Hong, S., Lin, X., Lee, H.: Decomposing motion and content for natural video sequence prediction (2018)" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 132, + 483, + 480, + 504 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 483, + 480, + 504 + ], + "spans": [ + { + "bbox": [ + 132, + 483, + 480, + 504 + ], + "type": "text", + "content": "57. Voleti, V., Jolicoeur-Martineau, A., Pal, C.: Mcvd: Masked conditional video diffusion for prediction, generation, and interpolation (2022)" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 132, + 504, + 480, + 536 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 504, + 480, + 536 + ], + "spans": [ + { + "bbox": [ + 132, + 504, + 480, + 536 + ], + "type": "text", + "content": "58. Wang, C., Gu, J., Hu, P., Xu, S., Xu, H., Liang, X.: Dreamvideo: High-fidelity image-to-video generation with image retention and text guidance. arXiv preprint arXiv:2312.03018 (2023)" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 132, + 536, + 480, + 568 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 536, + 480, + 568 + ], + "spans": [ + { + "bbox": [ + 132, + 536, + 480, + 568 + ], + "type": "text", + "content": "59. Wang, Y., Chen, X., Ma, X., Zhou, S., Huang, Z., Wang, Y., Yang, C., He, Y., Yu, J., Yang, P., et al.: Lavie: High-quality video generation with cascaded latent diffusion models. arXiv preprint arXiv:2309.15103 (2023)" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 132, + 569, + 480, + 590 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 569, + 480, + 590 + ], + "spans": [ + { + "bbox": [ + 132, + 569, + 480, + 590 + ], + "type": "text", + "content": "60. Weissenborn, D., Täckström, O., Uszkoreit, J.: Scaling autoregressive video models (2020)" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 132, + 590, + 480, + 611 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 590, + 480, + 611 + ], + "spans": [ + { + "bbox": [ + 132, + 590, + 480, + 611 + ], + "type": "text", + "content": "61. Wu, C., Huang, L., Zhang, Q., Li, B., Ji, L., Yang, F., Sapiro, G., Duan, N.: Godiva: Generating open-domain videos from natural descriptions (2021)" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 132, + 612, + 480, + 633 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 612, + 480, + 633 + ], + "spans": [ + { + "bbox": [ + 132, + 612, + 480, + 633 + ], + "type": "text", + "content": "62. Wu, C., Liang, J., Ji, L., Yang, F., Fang, Y., Jiang, D., Duan, N.: Niwa: Visual synthesis pre-training for neural visual world creation (2021)" + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 132, + 633, + 480, + 665 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 633, + 480, + 665 + ], + "spans": [ + { + "bbox": [ + 132, + 633, + 480, + 665 + ], + "type": "text", + "content": "63. Xing, J., Xia, M., Zhang, Y., Chen, H., Wang, X., Wong, T.T., Shan, Y.: Dynamicultural: Animating open-domain images with video diffusion priors. arXiv preprint arXiv:2310.12190 (2023)" + } + ] + } + ], + "index": 22 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 227, + 91, + 447, + 101 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 227, + 91, + 447, + 101 + ], + "spans": [ + { + "bbox": [ + 227, + 91, + 447, + 101 + ], + "type": "text", + "content": "Pix2Gif: Motion-Guided Diffusion for GIF Generation" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 470, + 91, + 481, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 470, + 91, + 481, + 100 + ], + "spans": [ + { + "bbox": [ + 470, + 91, + 481, + 100 + ], + "type": "text", + "content": "17" + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 16 + }, + { + "para_blocks": [ + { + "bbox": [ + 132, + 116, + 480, + 402 + ], + "type": "list", + "angle": 0, + "index": 10, + "blocks": [ + { + "bbox": [ + 132, + 116, + 480, + 149 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 116, + 480, + 149 + ], + "spans": [ + { + "bbox": [ + 132, + 116, + 480, + 149 + ], + "type": "text", + "content": "64. Xu, J., Mei, T., Yao, T., Rui, Y.: Msr-vtt: A large video description dataset for bridging video and language. In: Proceedings of the IEEE conference on computer vision and pattern recognition. pp. 5288-5296 (2016)" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 132, + 150, + 480, + 171 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 150, + 480, + 171 + ], + "spans": [ + { + "bbox": [ + 132, + 150, + 480, + 171 + ], + "type": "text", + "content": "65. Yang, Z., Wang, J., Gan, Z., Li, L., Lin, K., Wu, C., Duan, N., Liu, Z., Liu, C., Zeng, M., Wang, L.: Reco: Region-controlled text-to-image generation (2022)" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 132, + 172, + 480, + 205 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 172, + 480, + 205 + ], + "spans": [ + { + "bbox": [ + 132, + 172, + 480, + 205 + ], + "type": "text", + "content": "66. Yu, J., Xu, Y., Koh, J.Y., Luong, T., Baid, G., Wang, Z., Vasudevan, V., Ku, A., et al.: Scaling autoregressive models for content-rich text-to-image generation. arXiv preprint arXiv:2206.10789 2(3), 5 (2022)" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 132, + 205, + 480, + 247 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 205, + 480, + 247 + ], + "spans": [ + { + "bbox": [ + 132, + 205, + 480, + 247 + ], + "type": "text", + "content": "67. Yu, L., Cheng, Y., Sohn, K., Lezama, J., Zhang, H., Chang, H., Hauptmann, A.G., Yang, M.H., Hao, Y., Essa, I., et al.: Magvit: Masked generative video transformer. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 10459-10469 (2023)" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 132, + 249, + 480, + 313 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 249, + 480, + 313 + ], + "spans": [ + { + "bbox": [ + 132, + 249, + 480, + 313 + ], + "type": "text", + "content": "68. Yu, L., Shi, B., Pasunuru, R., Muller, B., Golovneva, O., Wang, T., Babu, A., Tang, B., Karrer, B., Sheynin, S., Ross, C., Polyak, A., Howes, R., Sharma, V., Xu, P., Tamoyan, H., Ashual, O., Singer, U., Li, S.W., Zhang, S., James, R., Ghosh, G., Taigman, Y., Fazel-Zarandi, M., Celikyilmaz, A., Zettlemoyer, L., Aghajanyan, A.: Scaling autoregressive multi-modal models: Pretraining and instruction tuning (2023)" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 132, + 315, + 480, + 346 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 315, + 480, + 346 + ], + "spans": [ + { + "bbox": [ + 132, + 315, + 480, + 346 + ], + "type": "text", + "content": "69. Zhang, L., Rao, A., Agrawala, M.: Adding conditional control to text-to-image diffusion models. In: Proceedings of the IEEE/CVF International Conference on Computer Vision. pp. 3836-3847 (2023)" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 132, + 347, + 480, + 380 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 347, + 480, + 380 + ], + "spans": [ + { + "bbox": [ + 132, + 347, + 480, + 380 + ], + "type": "text", + "content": "70. Zhang, S., Wang, J., Zhang, Y., Zhao, K., Yuan, H., Qin, Z., Wang, X., Zhao, D., Zhou, J.: I2vgen-xl: High-quality image-to-video synthesis via cascaded diffusion models. arXiv preprint arXiv:2311.04145 (2023)" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 132, + 380, + 480, + 402 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 380, + 480, + 402 + ], + "spans": [ + { + "bbox": [ + 132, + 380, + 480, + 402 + ], + "type": "text", + "content": "71. Özgün Çiçek, Abdulkadir, A., Lienkamp, S.S., Brox, T., Ronneberger, O.: 3d u-net: Learning dense volumetric segmentation from sparse annotation (2016)" + } + ] + } + ], + "index": 9 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "spans": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "type": "text", + "content": "18" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 166, + 91, + 237, + 100 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 166, + 91, + 237, + 100 + ], + "spans": [ + { + "bbox": [ + 166, + 91, + 237, + 100 + ], + "type": "text", + "content": "H. Kandala et al." + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 17 + } + ], + "_backend": "vlm", + "_version_name": "2.6.4" +} \ No newline at end of file diff --git a/2024/PixArt-Sigma_ Weak-to-Strong Training of Diffusion Transformer for 4K Text-to-Image Generation/56a1c2d9-7461-45b0-bc11-505740d1999b_content_list.json b/2024/PixArt-Sigma_ Weak-to-Strong Training of Diffusion Transformer for 4K Text-to-Image Generation/56a1c2d9-7461-45b0-bc11-505740d1999b_content_list.json new file mode 100644 index 0000000000000000000000000000000000000000..444a74cd21e8f594b5358f9e45db776cfebe44c4 --- /dev/null +++ b/2024/PixArt-Sigma_ Weak-to-Strong Training of Diffusion Transformer for 4K Text-to-Image Generation/56a1c2d9-7461-45b0-bc11-505740d1999b_content_list.json @@ -0,0 +1,1759 @@ +[ + { + "type": "text", + "text": "PIXART- $\\boldsymbol{\\Sigma}$ : Weak-to-Strong Training of Diffusion Transformer for 4K Text-to-Image Generation", + "text_level": 1, + "bbox": [ + 222, + 140, + 781, + 186 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Junsong Chen $^{1,2,3*}$ , Chongjian Ge $^{1,3*}$ , Enze Xie $^{1*†}$ , Yue Wu $^{1*}$ , Lewei Yao $^{1,4}$ , Xiaozhe Ren $^{1}$ , Zhongdao Wang $^{1}$ , Ping Luo $^{3}$ , Huchuan Lu $^{2}$ , and Zhenguo Li $^{1}$", + "bbox": [ + 222, + 210, + 779, + 244 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "1 Huawei Noah's Ark Lab 2 Dalian University of Technology 3 HKU 4 HKUST Project Page: https://pixart-alpha.github.io/PixArt-sigma-project/", + "bbox": [ + 222, + 253, + 779, + 284 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Abstract. In this paper, we introduce $\\mathrm{PIXART - \\Sigma}$ , a Diffusion Transformer model (DiT) capable of directly generating images at 4K resolution. $\\mathrm{PIXART - \\Sigma}$ represents a significant advancement over its predecessor, $\\mathrm{PIXART - \\alpha}$ , offering images of markedly higher fidelity and improved alignment with text prompts. A key feature of $\\mathrm{PIXART - \\Sigma}$ is its training efficiency. Leveraging the foundational pre-training of $\\mathrm{PIXART - \\alpha}$ , it evolves from the 'weaker' baseline to a 'stronger' model via incorporating higher quality data, a process we term \"weak-to-strong training\". The advancements in $\\mathrm{PIXART - \\Sigma}$ are twofold: (1) High-Quality Training Data: $\\mathrm{PIXART - \\Sigma}$ incorporates superior-quality image data, paired with more precise and detailed image captions. (2) Efficient Token Compression: we propose a novel attention module within the DiT framework that compresses both keys and values, significantly improving efficiency and facilitating ultra-high-resolution image generation. Thanks to these improvements, $\\mathrm{PIXART - \\Sigma}$ achieves superior image quality and user prompt adherence capabilities with significantly smaller model size (0.6B parameters) than existing text-to-image diffusion models, such as SDXL (2.6B parameters) and SD Cascade (5.1B parameters). Moreover, $\\mathrm{PIXART - \\Sigma}$ 's capability to generate 4K images supports the creation of high-resolution posters and wallpapers, efficiently bolstering the production of high-quality visual content in industries such as film and gaming.", + "bbox": [ + 259, + 321, + 743, + 613 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Keywords: T2I Synthesis, Diffusion Transformer, Efficient Model", + "bbox": [ + 261, + 626, + 709, + 641 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "1 Introduction", + "text_level": 1, + "bbox": [ + 215, + 667, + 375, + 683 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "The recent emergence of high-quality Text-to-Image (T2I) models has profoundly impacted the AI Generated Content (AIGC) community. This includes both proprietary models such as DALL-E 3 [31], Midjourney [29], as well as open-source models like Stable Diffusion [36] and PixART- $\\alpha$ [4]. Nonetheless, developing a top-tier T2I model involves considerable resources; for instance, training SD1.5 from scratch necessitates about 6000 A100 GPU days [36], posing a substantial barrier to individual researchers with limited resources and impeding innovation", + "bbox": [ + 212, + 699, + 787, + 806 + ], + "page_idx": 0 + }, + { + "type": "page_footnote", + "text": "*Equal contribution. Work done during the students' internships at Huawei Noah's Ark Lab. †Project lead and corresponding author. Thanks to Charan for the 4K dataset collection.", + "bbox": [ + 215, + 816, + 781, + 839 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/2eae751a45613fdced33b20a28013da2d82387456ecdc459652b02b3a308e744.jpg", + "image_caption": [ + "Fig. 1: Images generated by $\\mathbf{P}\\mathbf{I}\\mathbf{x}\\mathbf{A}\\mathbf{R}\\mathbf{T} - \\boldsymbol{\\Sigma}$ . The model can output photo-realistic, high aesthetic, extreme aspect ratio, multi-style images, and follow user instructions." + ], + "image_footnote": [], + "bbox": [ + 243, + 143, + 759, + 763 + ], + "page_idx": 1 + }, + { + "type": "page_number", + "text": "2", + "bbox": [ + 217, + 114, + 228, + 126 + ], + "page_idx": 1 + }, + { + "type": "header", + "text": "J. Chen et al.", + "bbox": [ + 271, + 114, + 364, + 126 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "within the AIGC community. Over time, the AIGC community will gain access to continuously updated, higher-quality datasets and more advanced algorithms. A pivotal question is: how can we efficiently integrate these new elements into an existing model, achieving a more powerful version within the constraints of limited resources?", + "bbox": [ + 218, + 145, + 787, + 220 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "To explore this issue, our research focuses on enhancing $\\mathrm{PIXART - \\alpha}$ , an efficient T2I training method. $\\mathrm{PIXART - \\alpha}$ represents an early venture within the DiT framework, a model structure with significant potential, as evidenced by works such as GenTron [6], Sora [33] and Stable Diffusion 3 [42]. To maximize this potential, we build upon the pre-trained foundation of $\\mathrm{PIXART - \\alpha}$ , integrating advanced elements to facilitate its continuous improvement, resulting in a more powerful model, $\\mathrm{PIXART - \\Sigma}$ . We refer to this process of evolving from a relatively weaker baseline to a stronger model through efficient training as \"weak-to-strong training\". Specifically, to achieve \"weak-to-strong training\", we introduce the following enhancements:", + "bbox": [ + 218, + 229, + 787, + 378 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Higher-Quality Training Data: We collect a high-quality dataset superior to that used in $\\mathrm{PIXART - \\alpha}$ , focusing on two key aspects: (i) High-quality images: The dataset comprises 33M high-resolution images sourced from the Internet, all exceeding 1K resolution, including 2.3M images with resolutions around 4K. These images are predominantly characterized by their high aesthetic and encompass a wide range of artistic styles. (ii) Dense and accurate captions: To provide more precise and detailed captions for the aforementioned images, we replace the LLaVA [22] used in $\\mathrm{PIXART - \\alpha}$ with a more powerful image captioner, Share-Captioner [5]. Furthermore, to improve the model's alignment capacity between the textual and visual concepts, we extend the token length of the text encoder (i.e., Flan-T5 [10]) to approximately 300 words. We observe these improvements effectively eliminate the model's tendency for hallucination, leading to higher-quality text-image alignment.", + "bbox": [ + 218, + 387, + 787, + 583 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Efficient Token Compression: To enhance $\\mathrm{PIXART - \\alpha}$ , we expand its generation resolution from 1K to 4K. Generating images at ultra-high resolutions (e.g., 2K/4K) introduces a significant increase in the number of tokens, leading to a substantial rise in computational demand. To address this challenge, we introduced a self-attention module with key and value token compression tailored to the DiT framework. Specifically, we utilize group convolutions with a stride of 2 for local aggregation of keys and values. Additionally, we employ a specialized weight initialization scheme, allowing for a smooth adaptation from a pre-trained model without KV compression. This design effectively reduces training and inference time by $\\sim 34\\%$ for high-resolution image generation.", + "bbox": [ + 218, + 590, + 787, + 741 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Weak-to-Strong Training Strategy: we propose several fine-tuning techniques to rapidly adapt from a weak model to a strong model efficiently. That includes (1) replacing with a more powerful Variational Autoencoder (VAE) [36], (2) scaling from low to high resolution, and (3) evolving from a model without Key-Value (KV) compression to one with KV compression. These outcomes confirm the validity and effectiveness of the \"weak-to-strong training\" approach.", + "bbox": [ + 218, + 750, + 787, + 839 + ], + "page_idx": 2 + }, + { + "type": "header", + "text": "PixART- $\\boldsymbol{\\Sigma}$ : Weak-to-Strong Training of DiT for 4K T2I Generation", + "bbox": [ + 282, + 113, + 730, + 128 + ], + "page_idx": 2 + }, + { + "type": "page_number", + "text": "3", + "bbox": [ + 774, + 116, + 784, + 126 + ], + "page_idx": 2 + }, + { + "type": "image", + "img_path": "images/7e35c89910d2f12dd3e67db0e4bd640a959fe122d35481f09a9a261e1d50789e.jpg", + "image_caption": [ + "Fig. 2: 4K image generation with complex dense instructions. $\\mathrm{PIXART - \\Sigma}$ can directly generate 4K resolution images without post-processing, and accurately respond to the given prompt." + ], + "image_footnote": [], + "bbox": [ + 218, + 143, + 779, + 415 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Through the proposed improvements, $\\mathrm{PIXART - \\Sigma}$ achieves high-quality 4K resolution image generation at a minimal training cost and model parameters. Specifically, fine-tuning from a pre-trained model, we additionally utilize only $9\\%$ of the GPU days required by $\\mathrm{PIXART - \\alpha}$ to achieve a strong 1K high-resolution image generation model, which is impressive considering replacing with new training data and a more powerful VAE. Moreover, we use only 0.6B parameters while SDXL [36] and SD Cascade [35] use 2.6B and 5.1B parameters respectively. Images generated by $\\mathrm{PIXART - \\Sigma}$ possess an aesthetic quality comparable to current top-tier T2I products, such as DALL-E 3 [31] and MJV6 [29] (as in Fig. 4). Additionally, $\\mathrm{PIXART - \\Sigma}$ also demonstrates exceptional capability for fine-grained alignment with textual prompts (as shown in Fig. 2 and 3).", + "bbox": [ + 212, + 467, + 787, + 633 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "2 Related Work", + "text_level": 1, + "bbox": [ + 215, + 656, + 387, + 672 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Diffusion Transformers. The Transformer architecture has achieved remarkable success across various domains, such as language modeling [37,38], computer vision [23, 43, 51, 53], and other areas [3, 14]. In the realm of diffusion models, DiT [34] and UViT [2] pioneer the use of Transformer architecture. Subsequent works, including DiffiT [15], SiT [28], and FiT [26], have improved upon DiT's architecture, while [12, 52] enhance training efficiency through masked modeling techniques. For Text-to-Image (T2I) synthesis, PixART- $\\alpha$ [4] explore efficient T2I training schemes, achieving the first Transformer-based T2I model capable of generating 1024px high-quality images. GenTron [6] explores the flexibility and scalability of diffusion Transformers in both image and video generation. The re", + "bbox": [ + 212, + 688, + 787, + 840 + ], + "page_idx": 3 + }, + { + "type": "page_number", + "text": "4", + "bbox": [ + 217, + 114, + 228, + 126 + ], + "page_idx": 3 + }, + { + "type": "header", + "text": "J. Chen et al.", + "bbox": [ + 271, + 114, + 362, + 126 + ], + "page_idx": 3 + }, + { + "type": "image", + "img_path": "images/fb9c7c710fe0e9d2ee06c3f7f4d3cf783b86318f0d6b42b108580da8e3edc19c.jpg", + "image_caption": [], + "image_footnote": [ + "Prompt: A close-up photo of a person. The subject is a woman. She wore a blue coat with a gray dress underneath. She has blue eyes and blond hair, and wears a pair of earrings. Behind are blurred city buildings and streets." + ], + "bbox": [ + 218, + 143, + 401, + 300 + ], + "page_idx": 4 + }, + { + "type": "image", + "img_path": "images/12e79cda23078bf0271910267e946835d7e8f827a93e81cc47ed7c78fbbc42e9.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 405, + 143, + 589, + 300 + ], + "page_idx": 4 + }, + { + "type": "image", + "img_path": "images/e41f0f585330f34bde73a7d6b4aacb5c0c71eccf8581ed83c3162caa14c4a68d.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 591, + 143, + 777, + 300 + ], + "page_idx": 4 + }, + { + "type": "image", + "img_path": "images/ea7a0a13f17110ee9ad20ffc30c8e4306c16b7efceab99065c96d8c050e7ad46.jpg", + "image_caption": [], + "image_footnote": [ + "Prompt: half a solid black background and half a solid white background" + ], + "bbox": [ + 220, + 325, + 403, + 467 + ], + "page_idx": 4 + }, + { + "type": "image", + "img_path": "images/2a9f33ba284496177d5d438d058378e702d205d9b8133067ed20332253946dd5.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 413, + 325, + 589, + 465 + ], + "page_idx": 4 + }, + { + "type": "image", + "img_path": "images/aa2f55b578413185cd7fffa3afbcab97e71b7f0f458f495054ee6eb8905ee330.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 594, + 325, + 687, + 467 + ], + "page_idx": 4 + }, + { + "type": "image", + "img_path": "images/39caa60e2a5cfb34fb011d38b26399a54a7ce7c4d76b21e0709d0b22a2a4aea0.jpg", + "image_caption": [ + "Fig. 3: Comparison of $\\mathbf{P}\\mathbf{i}\\mathbf{x}\\mathbf{A}\\mathbf{r}\\mathbf{T} - \\boldsymbol{\\Sigma}$ with open-source models, e.g., $\\mathbf{P}\\mathbf{i}\\mathbf{x}\\mathbf{A}\\mathbf{r}\\mathbf{T} - \\alpha$ and SDXL: Compared with $\\mathbf{P}\\mathbf{i}\\mathbf{x}\\mathbf{A}\\mathbf{r}\\mathbf{T} - \\alpha$ , $\\mathbf{P}\\mathbf{i}\\mathbf{x}\\mathbf{A}\\mathbf{r}\\mathbf{T} - \\boldsymbol{\\Sigma}$ improves the realism of portraits and the capability of semantic analysis. Compared with SDXL, our method has a better ability to follow user instructions. The keywords are highlighted as blue." + ], + "image_footnote": [ + "Prompt: Pixel art style of a snowboarder in mid-air performs a trick on a black rail, wearing a blue sweatshirt and black pants, with arms outstretched. The serene snowy landscape background, dotted with trees, complements the scene. The low-angle perspective emphasizes the trick's height and skill." + ], + "bbox": [ + 220, + 481, + 403, + 621 + ], + "page_idx": 4 + }, + { + "type": "image", + "img_path": "images/d55309bb589a3373b0becb9b973a4e2557629f448d5ee2d29acb693587b4d085.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 406, + 479, + 589, + 621 + ], + "page_idx": 4 + }, + { + "type": "image", + "img_path": "images/c526f9445e51f205cd246014870e00ff4bdc8233f908d7f120ef10d62e7c5143.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 593, + 479, + 777, + 621 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "cent advent of the powerful video generation model Sora [33] has further underscored the potential of Diffusion Transformers. In this work, for the first time, we explore using the Transformer architecture to generate 4K ultra-high-resolution images directly, tackling the computational complexity challenges posed by involving long-sequence tokens.", + "bbox": [ + 212, + 729, + 787, + 805 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "High Resolution Image Generation greatly enhances visual quality and is important in various industries such as film and gaming. However, increasing", + "bbox": [ + 212, + 809, + 787, + 840 + ], + "page_idx": 4 + }, + { + "type": "header", + "text": "PixART- $\\boldsymbol{\\Sigma}$ : Weak-to-Strong Training of DiT for 4K T2I Generation", + "bbox": [ + 282, + 114, + 730, + 128 + ], + "page_idx": 4 + }, + { + "type": "page_number", + "text": "5", + "bbox": [ + 774, + 116, + 784, + 126 + ], + "page_idx": 4 + }, + { + "type": "image", + "img_path": "images/14c5470177a9e12fbbd690e435925b8cf2b8786a6527e154d55c523f3dce123d.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 217, + 143, + 496, + 361 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/2d57c3884bd71131bb8388954711c28bdd9a712c0033176cb3944a4d2f656fef.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 496, + 143, + 638, + 361 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/d556f7bae82ad9665e6187894d569497c6b51ec9c6a93ba90053c62cd9c38460.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 638, + 143, + 779, + 361 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Prompt: a small cactus with a happy face in the Sahara desert", + "bbox": [ + 218, + 364, + 511, + 373 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/d180f59fd0eb73be17a0325501067e112bcbcab0ac6b83eb0a16027ac91e92e6.jpg", + "image_caption": [ + "Fig. 4: Compare $\\mathbf{P}\\mathbf{i}\\mathbf{x}\\mathbf{A}\\mathbf{r}\\mathbf{T} - \\boldsymbol{\\Sigma}$ and four other T2I products: Firefly 2, Imagen 2, Dalle 3, and Midjourney 6. Images generated by $\\mathbf{P}\\mathbf{i}\\mathbf{x}\\mathbf{A}\\mathbf{r}\\mathbf{T} - \\boldsymbol{\\Sigma}$ are very competitive with these commercial products." + ], + "image_footnote": [], + "bbox": [ + 217, + 375, + 496, + 592 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/8fa7178d9227a3b60727c4ce8ce09e21c5a52f26a58a40283722b2a94a6a2561.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 496, + 375, + 638, + 592 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/6de248d743f167195a1999991156d028247a18fa3f01e7cf191918951ed62eac.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 638, + 375, + 779, + 592 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Prompt: a cherry pick scientist", + "bbox": [ + 218, + 593, + 364, + 604 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "image resolution introduces challenges due to the substantial increase in computational demands. Numerous methods have been explored in this direction. For instance, Imagen [40], GigaGAN [18] and Stable Diffusion [39] introduce an additional super-resolution network, while Stable Cascade [35] employs multiple diffusion networks to increase resolution progressively. These combined-model solutions, however, can introduce cumulative errors. On the other hand, works like SDXL [36], DALL-E 2 [30], Playground [19] and $\\mathrm{PIXART - \\alpha}$ [4] have attempted to generate high-resolution images using diffusion models directly. Nevertheless, these efforts are capped at generating images with resolutions up to $1024\\mathrm{px}$ due to computational complexity. In this paper, we push this boundary to $4\\mathrm{K}$ resolution, significantly enhancing the visual quality of the generated content.", + "bbox": [ + 212, + 672, + 787, + 840 + ], + "page_idx": 5 + }, + { + "type": "page_number", + "text": "6", + "bbox": [ + 217, + 114, + 228, + 126 + ], + "page_idx": 5 + }, + { + "type": "header", + "text": "J. Chen et al.", + "bbox": [ + 271, + 114, + 362, + 126 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Efficient Transformer architecture. The self-attention mechanism in Transformer suffers from quadratic computational complexity with respect to the number of tokens, which hinders the scaling up of token quantity. Many works have sought improvements in this area: (1) Sparse Attention [7, 8, 45, 46, 48], which reduces the overall computational load by selectively processing a subset of tokens. For instance, PVT v2 [46] employs a convolutional kernel to condense the space of the key and value, thus lowering the complexity involved in computing the attention. (2) Local Attention [13, 23, 47, 54] focuses attention within nearby regions; notably, Swin Transformer [23] utilizes window-based attention to limit computations to within a specified window size. (3) Low-rank/Linear Attention [9, 25, 44]. The Linformer [44] reduces the computational complexity of the self-attention mechanism through low-rank approximations. In this paper, inspired by PVT v2 [46], we employ a self-attention mechanism based on key/value compression to mitigate the high complexity of processing 4K images.", + "bbox": [ + 212, + 146, + 787, + 359 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "3 Framework", + "text_level": 1, + "bbox": [ + 215, + 380, + 361, + 396 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "3.1 Data Analysis", + "text_level": 1, + "bbox": [ + 215, + 411, + 377, + 426 + ], + "page_idx": 6 + }, + { + "type": "table", + "img_path": "images/419a573d179dfc5658c9ac8963a6cf5f15a7cd350fe87f0b24c3816cc8859715.jpg", + "table_caption": [], + "table_footnote": [], + "table_body": "
LLaVA HallucinationsThe image features a large, ornate church with a tall, pointed roof and a large stained-glass window. The church is situated in a city, and it has a white and gray color scheme. The style of the church is Gothic, which is evident from the pointed roof and the intricate details of the stained-glass window. The presence of statues and a clock on the building further emphasizes the grand and historical nature of the structure.
Share-Captioner CorrectnessThe image captures the grandeur of a cathedral, painted in a pristine white, stands majestically against the backdrop of a clear blue sky. The entrance to the cathedral is adorned with a large rose window. Flanking the entrance are statues of saints, standing as silent guardians of the sacred space within. The perspective of the photo, taken from a low angle. The lush green trees in the background adding serenity to the scene.
LLaVA HallucinationsThe image features a woman and a man sitting on a brick walkway near a body of water, which could be a river or a lake. They are both wearing head coverings, and the woman is holding a handbag. The scene is set during the day, with the sun shining brightly, creating a warm and inviting atmosphere. The style of the image is a black and white photo, which adds a timeless and classic feel to the scene.
Share-Captioner CorrectnessThe image captures a serene scene at a harbor. Two individuals are seated on a bench, their backs to the camera, engrossed in the view of the water. The water, a deep shade of blue, is dotted with boats of various sizes and colors, including a white boat with a green stripe and a red boat. The sky above is a light blue.
", + "bbox": [ + 222, + 428, + 779, + 571 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Fig. 5: Comparative illustration of hallucinations: Contrasting differences in hallucination occurrences between LLaVA and Share-Captioner, with red indicating hallucinations and green denoting correctness.", + "bbox": [ + 212, + 579, + 787, + 622 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Higher Aesthetic and higher Resolution. To enhance the aesthetic quality of our dataset, we expand our internal data from 14M to 33M. For clarity, we name the two datasets Internal- $\\alpha$ and Internal- $\\Sigma$ , respectively. Note that this expansion still falls short compared to the vast images utilized by currently available open-source models like SD v1.5, which uses 2B data. We demonstrate that effective training strategies with limited data amount can still obtain a strong T2I model.", + "bbox": [ + 212, + 628, + 787, + 733 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "The images within Internal- $\\Sigma$ are above 1K resolution. To facilitate 4K resolution generation, we additionally collect a dataset of 8M real photographic images at 4K resolution. To ensure aesthetic quality, we employ an aesthetic scoring model (AES) [1] to filter these 4K images. This process yields a highly refined dataset of 2M ultra-high-resolution and high-quality images.", + "bbox": [ + 212, + 734, + 787, + 809 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Interestingly, we have observed that as the resolution of the images increases, there is an improvement in the model's fidelity (Fréchet Inception Dis-", + "bbox": [ + 212, + 809, + 787, + 839 + ], + "page_idx": 6 + }, + { + "type": "header", + "text": "PIXART- $\\boldsymbol{\\Sigma}$ : Weak-to-Strong Training of DiT for 4K T2I Generation", + "bbox": [ + 281, + 113, + 730, + 128 + ], + "page_idx": 6 + }, + { + "type": "page_number", + "text": "7", + "bbox": [ + 774, + 114, + 784, + 126 + ], + "page_idx": 6 + }, + { + "type": "table", + "img_path": "images/0364c1206f08d202fc8bbafc49beec55b269b47752aa00663ce43cb832701365.jpg", + "table_caption": [ + "Table 1: Statistics of noun concepts for different datasets. VN: valid distinct nouns (appearing more than 10 times); DN: total distinct nouns; Average: average noun count per image; ACL: Average Caption length." + ], + "table_footnote": [], + "table_body": "
DatasetVolumeCaptionVN/DNTotal NounACLAverage
Internal-α14MRaw187K/931K175M2511.7/Img
Internal-α14MLLaVA28K/215K536M9829.3/Img
Internal-α14MShare-Captioner51K/420K815M18454.4/Img
Internal-Σ33MRaw294K/1512K485M3514.4/Img
Internal-Σ33MShare-Captioner77K/714K1804M18053.6/Img
4K-Σ2.3MShare-Captioner24K/96K115M16349.5/Img
", + "bbox": [ + 230, + 176, + 763, + 295 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "tance (FID) [17]) and semantic alignment (CLIP Score), which underscores the importance of the capabilities of generating high-resolution images.", + "bbox": [ + 212, + 315, + 782, + 345 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Better Text-Image Alignment.", + "text_level": 1, + "bbox": [ + 215, + 348, + 485, + 362 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Recent works such as $\\mathrm{PIXART - }\\alpha$ [4] and DALL-E 3 [31] emphasize the significance of text-image alignment. Strengthening this alignment is crucial for boosting model capabilities. To refine our collected \"raw\" descriptions further, we focus on improving both the length and accuracy of our captions. Notably, our captions (Internal- $\\Sigma$ ) show several advantages over the one used in $\\mathrm{PIXART - }\\alpha$ (Internal- $\\alpha$ ) in the following aspects: 1. Enhanced caption accuracy: As depicted in Fig. 5, LLaVa used in $\\mathrm{PIXART - }\\alpha$ encounters certain hallucination. We leverage a more powerful model, i.e., Share-Captioner [5], to", + "bbox": [ + 212, + 363, + 486, + 621 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/2779e528b6630f430c51a8bf54e45109d8b21c4d9a8462cb49a0b90232c2228d.jpg", + "image_caption": [ + "Fig.6: Histogram Visualization of the Caption Length. We randomly select 1M captions from the raw captions, Internal- $\\alpha$ , and Internal- $\\Sigma$ to draw the corresponding histogram. ACL denotes the average caption length." + ], + "image_footnote": [], + "bbox": [ + 503, + 354, + 782, + 511 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "generate detailed and correct captions, augmenting the collected raw prompts.", + "bbox": [ + 212, + 622, + 779, + 638 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "2. Increased caption length: As shown in Tab. 1 and Fig. 6, the average caption length increased significantly to 180 words, highly enhancing the descriptive power of the captions. Additionally, we extend the token processing length of the text encoder from 120 tokens (as in Internal- $\\alpha$ ) to 300 tokens. Our model is trained on a mix of long (Share-Captioner) and short (raw) captions with a ratio of $60\\%$ and $40\\%$ , respectively. This approach enhances the diversity of textual descriptions and mitigates potential biases that might arise from solely relying on generative captions.", + "bbox": [ + 212, + 638, + 787, + 760 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Tab. 1 demonstrates a summary for both Internal- $\\alpha$ and $-\\Sigma$ , where we assess the diversity of the datasets through various metrics, including the noun variety, total noun count, average caption length, and average nouns per image.", + "bbox": [ + 212, + 762, + 787, + 806 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "High-Quality Evaluation Dataset. Most SoTA T2I models chose MSCOCO [21] as the evaluation set to assess the FID and CLIP Scores. However, we observe", + "bbox": [ + 212, + 809, + 802, + 839 + ], + "page_idx": 7 + }, + { + "type": "page_number", + "text": "8", + "bbox": [ + 217, + 114, + 228, + 126 + ], + "page_idx": 7 + }, + { + "type": "header", + "text": "J. Chen et al.", + "bbox": [ + 271, + 114, + 362, + 126 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "evaluations conducted on the MSCOCO dataset may not adequately reflect a model's capabilities in aesthetics and text-image alignment. Therefore, we propose a curated set comprising 30,000 high-quality, aesthetically pleasing text-image pairs to facilitate the assessment. The selected samples of the dataset are presented in the appendix. This dataset is designed to provide a more comprehensive evaluation of a model's performance, particularly in capturing the intricacies of aesthetic appeal and the fidelity of alignment between textual descriptions and visual content. Unless otherwise specified, the evaluation experiments in the paper are conducted in the collected High-Quality Evaluation Dataset.", + "bbox": [ + 212, + 146, + 787, + 282 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "3.2 Efficient DiT Design", + "text_level": 1, + "bbox": [ + 215, + 313, + 429, + 329 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "An efficient DiT network is essential since the computational demand significantly increases when generating images at ultra-high resolutions. The attention mechanism plays a pivotal role in the efficacy of Diffusion Transformers, yet its quadratic computational demands significantly limit model scalability, particularly at higher resolutions e.g., 2K and 4K. Inspired by PVT v2 [46], we incorporate KV compression within the original PixART- $\\alpha$ s framework to address the computational challenges. This design adds a mere $0.018\\%$ to the total parameters yet achieves efficient reduction in computational costs via token compression, while still preserving both spatial and semantic information.", + "bbox": [ + 212, + 347, + 787, + 482 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "Key-Value (KV) Token Compression. Our motivation stems from the intriguing observation that applying key-value (KV) token compression directly to the pre-trained $\\mathrm{PIXART - \\alpha}$ can still generate reasonable images. This suggests a redundancy in the features. Considering the high similarity within adjacent $R\\times R$ patches, we assume that feature semantics within a window are redundant and can be compressed reasonably. We propose KV token compression, which is denoted as $f_{c}(\\cdot)$ , to compress token features within a $R\\times R$ window through a compression operator, as depicted in Fig. 7.", + "bbox": [ + 212, + 484, + 787, + 604 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "Furthermore, to mitigate the potential information loss caused by KV compression in self-attention computation, we opt to retain all the tokens of queries (Q). This strategy allows us to utilize KV compression effectively while mitigating the risk of losing crucial information. By employing KV compression, we enhance the efficiency of attention computations and reduce the computation complexity from $O(N^2)$ to $O\\left(\\frac{N^2}{R^2}\\right)$ , thereby making the computational cost of directly generating high-resolution images manageable.", + "bbox": [ + 212, + 607, + 493, + 840 + ], + "page_idx": 8 + }, + { + "type": "image", + "img_path": "images/fb5547b66eff4ce085940a61aa3d0b749e09e7cae296ff5305df0ec9bbb92298.jpg", + "image_caption": [ + "Fig. 7: Design of KV Token Compression. We merge KV tokens in spatial space to reduce the computation complexity." + ], + "image_footnote": [], + "bbox": [ + 504, + 638, + 787, + 758 + ], + "page_idx": 8 + }, + { + "type": "header", + "text": "PIXART- $\\Sigma$ : Weak-to-Strong Training of DiT for 4K T2I Generation", + "bbox": [ + 281, + 113, + 730, + 128 + ], + "page_idx": 8 + }, + { + "type": "page_number", + "text": "9", + "bbox": [ + 774, + 114, + 784, + 126 + ], + "page_idx": 8 + }, + { + "type": "table", + "img_path": "images/de22726a3d6bc016081655eec88453d29fc0f3f7ac466701f877ca0615df3dc7.jpg", + "table_caption": [], + "table_footnote": [], + "table_body": "
StrategiesVAE Change512 →1024 RA512 →1024 RA + PE Interp.RA + PE Adjust + KV CompressRA + PE Interp. + KV Compress + Conv Avg Init
0 step
Fast \nAdapt2K step100 step100 step100 step100 step
", + "bbox": [ + 217, + 133, + 781, + 321 + ], + "page_idx": 9 + }, + { + "type": "text", + "text": "Fig. 8: This illustration demonstrates how our training strategy designs expedite the model's convergence during the transition to VAEs, adjustment to higher resolutions, and the KV compression process, facilitating rapid learning from weak to strong.", + "bbox": [ + 215, + 333, + 784, + 375 + ], + "page_idx": 9 + }, + { + "type": "equation", + "text": "\n$$\n\\operatorname {A t t e n t i o n} (Q, K, V) = \\operatorname {s o f t m a x} \\left(\\frac {Q \\cdot f _ {c} (K) ^ {T}}{\\sqrt {d _ {k}}}\\right) f _ {c} (V) \\tag {1}\n$$\n", + "text_format": "latex", + "bbox": [ + 315, + 387, + 784, + 422 + ], + "page_idx": 9 + }, + { + "type": "text", + "text": "We compress deep layers using the convolution operator \"Conv2×2\" with specific initialization. Detailed experiments on other design variants are discussed in Sec. 5. Specifically, we design a specialized convolution kernel initialization \"Conv Avg Init\" that utilizes group convolution and initializes the weights $w = \\frac{1}{R^2}$ , equivalent to an average operator. This initialization strategy can initially produce coarse results, accelerating the fine-tuning process while only introducing 0.018% additional parameters.", + "bbox": [ + 212, + 426, + 784, + 532 + ], + "page_idx": 9 + }, + { + "type": "text", + "text": "3.3 Weak-to-Strong Training Strategy", + "text_level": 1, + "bbox": [ + 215, + 556, + 542, + 571 + ], + "page_idx": 9 + }, + { + "type": "text", + "text": "We propose several efficient training strategies to enhance the transition from a \"weak\" model to a \"strong\" model. These strategies encompass VAE rapid adaptation, high-resolution fine-tuning, and KV Token compression.", + "bbox": [ + 212, + 582, + 784, + 628 + ], + "page_idx": 9 + }, + { + "type": "text", + "text": "Adapting model to new VAEs. As VAEs develop, training T2I models from scratch is resource-intensive. We replace PixART- $\\alpha$ s VAE with SDXL's VAE and continue fine-tuning the diffusion model. We observe a rapid convergence phenomenon that fine-tuning quickly converges at 2K training steps as shown in Fig 8 (a). Fine-tuning is more efficient when dealing with VAE", + "bbox": [ + 215, + 628, + 493, + 777 + ], + "page_idx": 9 + }, + { + "type": "text", + "text": "Table 2: We fine-tune a high-resolution model from a low-resolution model and observe that even fine-tuning for a relatively short duration, such as 1K steps, can still yield high-quality results.", + "bbox": [ + 501, + 628, + 785, + 700 + ], + "page_idx": 9 + }, + { + "type": "table", + "img_path": "images/b70faae669c303f485345053405a2b65ba540daa1a7162668fe9d23847c1eb5d.jpg", + "table_caption": [ + "(a)", + "(b)", + "(c)" + ], + "table_footnote": [], + "table_body": "
ResolutionIterationsFID ↓CLIP ↑
25620K16.560.270
256 → 5121K9.750.272
256 → 512100K8.910.276
", + "bbox": [ + 531, + 705, + 748, + 770 + ], + "page_idx": 9 + }, + { + "type": "text", + "text": "model transferring and negates the necessity of training from scratch.", + "bbox": [ + 215, + 779, + 712, + 792 + ], + "page_idx": 9 + }, + { + "type": "text", + "text": "Adapting to Higher-Resolution. When we fine-tune from a low-resolution (LR) model to a high-resolution (HR) model, we observe a performance degradation in Fig. 8 (b), which we attribute to discrepancies in positional embeddings", + "bbox": [ + 215, + 794, + 785, + 840 + ], + "page_idx": 9 + }, + { + "type": "page_number", + "text": "10", + "bbox": [ + 217, + 114, + 235, + 126 + ], + "page_idx": 9 + }, + { + "type": "header", + "text": "J. Chen et al.", + "bbox": [ + 271, + 114, + 362, + 126 + ], + "page_idx": 9 + }, + { + "type": "text", + "text": "(PE) between different resolutions. To address this, we utilize the \"PE Interpolation\" trick [4, 49]: initializing the HR model's PE by interpolating the LR model's PE, significantly enhancing the HR model's initial status and expediting fine-tuning. We can obtain visually pleasing images even within only 100 training iterations. Besides, we quantitatively compare the model's performance as illustrated in Tab. 2. The fine-tuning quickly converges at 1K steps, and further training slightly improves the performance. This illustrates that using the \"PE Interpolation\" trick enables rapid convergence of higher resolution generation, obviating the need for training from scratch for generating at higher resolutions.", + "bbox": [ + 212, + 145, + 787, + 282 + ], + "page_idx": 10 + }, + { + "type": "text", + "text": "Adapting model to KV compression. We can use KV compression directly when fine-tuning from LR pre-trained models without KV compression. As shown in Fig. 8 (c), with our \"Conv Avg Init.\" strategy, $\\mathrm{PIXART - \\Sigma}$ starts from a better initial state, making converging easier and faster. Notably, $\\mathrm{PIXART - \\Sigma}$ performs satisfied visual results even within 100 training steps. Finally, through the KV compression operators and compression layers design in Sec 3.2, we can reduce $\\sim 34\\%$ of the training and inference time.", + "bbox": [ + 212, + 282, + 787, + 388 + ], + "page_idx": 10 + }, + { + "type": "text", + "text": "4 Experiment", + "text_level": 1, + "bbox": [ + 215, + 407, + 366, + 424 + ], + "page_idx": 10 + }, + { + "type": "text", + "text": "4.1 Implementation Details", + "text_level": 1, + "bbox": [ + 215, + 433, + 455, + 449 + ], + "page_idx": 10 + }, + { + "type": "text", + "text": "Training Details. We follow Imagen [40] and $\\mathrm{PIXART - \\alpha}$ [4] to employ the T5 [10]'s encoder (i.e., Flan-T5-XXL) as the text encoder for conditional feature extraction, and use $\\mathrm{PIXART - \\alpha}$ [4] as our base diffusion model. Unlike most works that extract fixed 77 text tokens, we adjust the length of text tokens from $\\mathrm{PIXART - \\alpha}$ 's 120 to 300, as the caption curated in Internal- $\\Sigma$ is much denser to provide highly fine-grained details. To capture the latent features of input images, we employ a pre-trained and frozen VAE from SDXL [36]. Other implementation details are the same as $\\mathrm{PIXART - \\alpha}$ . Models are finetuned on the $\\mathrm{PIXART - \\alpha}$ 's 256px pre-trained checkpoint with the position embedding interpolation trick [4]. Our final models, including 1K resolution, are trained on 32 V100 GPUs. We additionally use 16 A100 GPUs to train the 2K and 4K image generation models. For further information, please refer to the appendix.", + "bbox": [ + 212, + 454, + 787, + 635 + ], + "page_idx": 10 + }, + { + "type": "text", + "text": "Note that we use CAME optimizer [27] with a weight decay of 0 and a constant learning rate of 2e-5, instead of the regular AdamW [24] optimizer. This helps us reduce the dimension of the optimizer's state, leading to lower GPU memory without performance degradation.", + "bbox": [ + 212, + 635, + 787, + 696 + ], + "page_idx": 10 + }, + { + "type": "text", + "text": "Evaluation Metrics. To better illustrate aesthetics and semantic ability, we collect 30K high-quality text-image pairs (as mentioned in Sec. 3.1) to benchmark T2I models. We mainly evaluate $\\mathsf{PIXART - \\Sigma}$ via human and AI preference study since FID [39] may not adequately reflect the generation quality. However, we still provide the FID results on the collected dataset in the appendix.", + "bbox": [ + 212, + 696, + 787, + 772 + ], + "page_idx": 10 + }, + { + "type": "text", + "text": "4.2 Performance Comparisons", + "text_level": 1, + "bbox": [ + 215, + 789, + 478, + 804 + ], + "page_idx": 10 + }, + { + "type": "text", + "text": "Image Quality Assessment. We qualitatively evaluated our methodology against both closed-source text-to-image products and open-source models. As", + "bbox": [ + 212, + 809, + 785, + 840 + ], + "page_idx": 10 + }, + { + "type": "header", + "text": "PIXART- $\\Sigma$ : Weak-to-Strong Training of DiT for 4K T2I Generation", + "bbox": [ + 281, + 113, + 730, + 128 + ], + "page_idx": 10 + }, + { + "type": "page_number", + "text": "11", + "bbox": [ + 767, + 114, + 782, + 126 + ], + "page_idx": 10 + }, + { + "type": "text", + "text": "illustrated in Fig. 1, our model can produce high-quality, photo-realistic images with intricate details over diverse aspect ratios and styles. This capability underscores the superior performance of our approach in generating visually compelling content from textual descriptions. In Fig. 3, we compare $\\mathsf{PIXART - \\Sigma}$ with open-source models SDXL [36] and $\\mathsf{PIXART - \\alpha}$ [4], our method enhances the realism of portraits and boosts the capacity for semantic analysis. In contrast to SDXL, our approach demonstrates superior proficiency in following instructions.", + "bbox": [ + 212, + 146, + 787, + 253 + ], + "page_idx": 11 + }, + { + "type": "text", + "text": "Not only superior to open-source models, but our method is also very competitive with current T2I closed-source products, as depicted in Fig. 4. PixART- $\\Sigma$ produces photo-realistic results and adheres closely to user instructions, which is on par with contemporary commercial products.", + "bbox": [ + 212, + 255, + 787, + 316 + ], + "page_idx": 11 + }, + { + "type": "text", + "text": "High-resolution Generation. Our method is capable of directly generating 4K resolution images without the need for any post-processing. Additionally, it excels in accurately following complex, detailed, and long text provided by users, as demonstrated in Fig. 2. Thus, users do not need prompt engineering to achieve satisfactory results.", + "bbox": [ + 212, + 319, + 431, + 500 + ], + "page_idx": 11 + }, + { + "type": "text", + "text": "Our approach enables direct 4K image generation. In parallel, studies [11, 16] have", + "bbox": [ + 212, + 503, + 431, + 549 + ], + "page_idx": 11 + }, + { + "type": "image", + "img_path": "images/8537c6c5f4e0fabc7b967cfdf7be066099cb65f6eb850b73531d752d75e72409.jpg", + "image_caption": [ + "Fig. 9: Comparisons on FID and Clip-Score with Open-sourced T2I Models. PixArt- $\\Sigma$ demonstrates enhanced performance in terms of FID and Clip-Score on the curated 30K High-Quality Evaluation Dataset." + ], + "image_footnote": [], + "bbox": [ + 475, + 325, + 750, + 474 + ], + "page_idx": 11 + }, + { + "type": "text", + "text": "introduced tuning-free post-processing techniques aimed at generating HR images from LR models or employing super-resolution models [50] to produce HR images. However, their corresponding results often exhibit artifacts for two primary reasons: (1) Accumulative error may arise due to the cascade pipeline. (2) These methods do not capture the true distribution of 4K images nor learn the alignment between text and 4K images. We argue that our method might be a more promising way to generate high-resolution images. Our method yields superior results, and more visual comparison is included in the supplement.", + "bbox": [ + 212, + 549, + 787, + 671 + ], + "page_idx": 11 + }, + { + "type": "text", + "text": "Human/AI (GPT4V) Preference Study. We evaluate the well-trained model in both the human and AI preference study using a subset of 300 captions randomly collected from the High-Quality Evaluation Dataset mentioned in Sec. 3.1. We collect images generated by overall six open-source models, including PixART- $\\alpha$ , PixART- $\\Sigma$ , SD1.5 [39], Stable Turbo [41], Stable XL [36], Stable Cascade [35] and Playground-V2.0 [20]. We develop a website for the human preference study to display the prompts and their corresponding images. This website was distributed to trained evaluators, who were asked to assess the images, ranking them according to quality and how well they matched the text prompts. The results, illustrated by the blue bar in Fig. 9, indicate a marked preference for PixART- $\\Sigma$ over the other six T2I generators. PixART- $\\Sigma$ genera", + "bbox": [ + 212, + 672, + 787, + 840 + ], + "page_idx": 11 + }, + { + "type": "page_number", + "text": "12", + "bbox": [ + 217, + 114, + 235, + 126 + ], + "page_idx": 11 + }, + { + "type": "header", + "text": "J. Chen et al.", + "bbox": [ + 271, + 114, + 362, + 126 + ], + "page_idx": 11 + }, + { + "type": "table", + "img_path": "images/0715374b530774f0d28375ee990b978a13468839c19bff4b754e0b5b3f3f9365.jpg", + "table_caption": [], + "table_footnote": [], + "table_body": "
LayersFID ↓CLIP-Score ↑
N/A8.2440.276
Shallow (1-14)9.2780.275
Middle (7-20)9.0630.276
Deep (14-27)8.5320.275
", + "bbox": [ + 281, + 143, + 468, + 210 + ], + "page_idx": 12 + }, + { + "type": "table", + "img_path": "images/a3c63d2a9a9015a022e2fdb560f796d1150d3f3d862e490c5ca0ceb4254ece5a.jpg", + "table_caption": [ + "(a) Compression layers." + ], + "table_footnote": [], + "table_body": "
Res.RatioFID ↓CLIP-Score ↑Train Latency ↓
51218.2440.2762.3
51229.0630.2762.2 (-4%)
51249.6060.2762.1 (-9%)
102415.6850.27727.5
102425.5120.27322.5 (-18%)
102445.6440.27620.0 (-27%)
102495.7120.27517.8 (-35%)
", + "bbox": [ + 248, + 234, + 488, + 333 + ], + "page_idx": 12 + }, + { + "type": "table", + "img_path": "images/52d8bf0029703224b190897cef333a295345da29422a5f6e5de41a4ac8119ae7.jpg", + "table_caption": [ + "(c) Compression rations on different resolutions." + ], + "table_footnote": [], + "table_body": "
OperatorFID ↓CLIP-Score ↑
N/A8.2440.276
Token Discarding8.9180.275
Token Pooling9.4150.275
Conv2×28.5050.274
", + "bbox": [ + 524, + 143, + 720, + 210 + ], + "page_idx": 12 + }, + { + "type": "table", + "img_path": "images/103d6ecf87e07692ade7f7b6476f8be22c960876e7e52ba091e2e0bcebd9284d.jpg", + "table_caption": [ + "(b) Compression operators." + ], + "table_footnote": [], + "table_body": "
Res.RatioTrain Latency (s/Iter@32BS)↓ Test Latency ↓ (s/Img)
2K15658
2K437 (-34%)38 (-34%)
4K119191
4K4125 (-35%)60 (-34%)
", + "bbox": [ + 524, + 243, + 750, + 325 + ], + "page_idx": 12 + }, + { + "type": "text", + "text": "(d) Speed of different resolutions.", + "bbox": [ + 537, + 325, + 733, + 337 + ], + "page_idx": 12 + }, + { + "type": "text", + "text": "Table 3: KV-Token Compression Settings in Image Generation. This study employs FID, CMMD, and CLIP-Score metrics to assess the impact of various token compression components, such as compression ratio, positions, operators, and varying resolutions. Speed calculation in Tab. 3c is Second/Iteration/384 Batch-size.", + "bbox": [ + 212, + 349, + 784, + 406 + ], + "page_idx": 12 + }, + { + "type": "text", + "text": "ates superior high-quality images that closely follow user prompts, using a much smaller size (0.6B parameters) compared to existing T2I diffusion models like SDXL (2.6B parameters) and SD Cascade (5.1B parameters).", + "bbox": [ + 212, + 416, + 784, + 462 + ], + "page_idx": 12 + }, + { + "type": "text", + "text": "For AI preference study, we employ the advanced multimodal model, GPT-4 V [32], as the evaluator. For each trial, we supply GPT-4V with two images: one from $\\mathrm{PIXART - \\Sigma}$ and another from a competing T2I model. We craft distinct prompts guiding GPT-4V to vote based on image quality and image-and-text alignment. The results, represented by orange and green bars in Fig. 9, demonstrate consistent outcomes in both human and AI preference studies. Specifically, $\\mathrm{PIXART - \\Sigma}$ surpasses the baseline, $\\mathrm{PIXART - \\alpha}$ , in effectiveness. Compared to advanced models such as Stable Cascaded, $\\mathrm{PIXART - \\Sigma}$ exhibits competitive or superior performance on image quality and instruction-following abilities.", + "bbox": [ + 212, + 463, + 785, + 601 + ], + "page_idx": 12 + }, + { + "type": "text", + "text": "5 Ablation Studies", + "text_level": 1, + "bbox": [ + 215, + 631, + 415, + 645 + ], + "page_idx": 12 + }, + { + "type": "text", + "text": "We conduct ablation studies on generation performance on various KV compression designs. Unless specified, the experiments are conducted on 512px generation. The detailed settings of each ablation trial are included in the appendix.", + "bbox": [ + 212, + 670, + 784, + 715 + ], + "page_idx": 12 + }, + { + "type": "text", + "text": "5.1 Experimental settings", + "text_level": 1, + "bbox": [ + 215, + 746, + 442, + 762 + ], + "page_idx": 12 + }, + { + "type": "text", + "text": "We use the test set described in Sec. 3.1 for evaluation. We employ FID to compute the distributional difference between the collected and generated data for comparative metrics. Furthermore, we utilize CLIP-Score to assess the alignment between prompts and the generated images.", + "bbox": [ + 212, + 779, + 784, + 840 + ], + "page_idx": 12 + }, + { + "type": "header", + "text": "PIXART- $\\Sigma$ : Weak-to-Strong Training of DiT for 4K T2I Generation", + "bbox": [ + 281, + 113, + 730, + 128 + ], + "page_idx": 12 + }, + { + "type": "page_number", + "text": "13", + "bbox": [ + 767, + 114, + 784, + 126 + ], + "page_idx": 12 + }, + { + "type": "text", + "text": "5.2 Compression Designs", + "text_level": 1, + "bbox": [ + 215, + 146, + 437, + 162 + ], + "page_idx": 13 + }, + { + "type": "text", + "text": "Compression positions. We implemented KV compression at different depths within the Transformer structure: in the shallow layers (1~14), the intermediate layers (7~20), and the deep layers (14~27). As indicated in Tab. 3a, employing KV compression on deep layers notably achieves superior performance. We speculate this is because shallow layers typically encode detailed texture content, while deep layers abstract high-level semantic content. Since compression affects image quality rather than semantic information, compressing deep layers achieves the least information loss, making it practical for accelerating training without compromising generation quality.", + "bbox": [ + 212, + 169, + 787, + 305 + ], + "page_idx": 13 + }, + { + "type": "text", + "text": "Compression operators. We explored the impact of different compression operators. We employed three techniques, random discarding, average pooling, and parametric convolution, to compress $2 \\times 2$ tokens into a single token. As illustrated in Table 3b, the \"Conv $2 \\times 2$ \" method outperforms the others, underscoring the advantage of using a learnable kernel to more effectively reduce redundant features than simple discarding methods.", + "bbox": [ + 212, + 305, + 787, + 393 + ], + "page_idx": 13 + }, + { + "type": "text", + "text": "Compression ratios on different resolutions. We investigated the influence of varying compression ratios on different resolutions. As shown in Tab. 3c, remarkably, we find that token compression does not affect the alignment between textual and generated images (CLIP Score) but influences the image quality (FID) across resolutions. Although there is a slight degradation in image quality with increasing compression ratios, our strategy brings a training speedup of $18\\%$ to $35\\%$ . This suggests that our proposed KV compression is both effective and efficient for achieving high-resolution T2I generation.", + "bbox": [ + 212, + 396, + 787, + 516 + ], + "page_idx": 13 + }, + { + "type": "text", + "text": "Speed comparisons on different resolutions. We further comprehensively validate the speed acceleration in both training and inference in Tab. 3d. By 'latency', we refer to the duration required for inference on a single sample. Our method speeds up training and inference by about $35\\%$ in the 4K generation. Notably, training acceleration increases with resolution, accelerating from $18\\%$ to $35\\%$ as the resolution increases from 1K to 4K. This indicates the effectiveness of our method with increasing resolution, demonstrating its potential applicability to even higher-resolution image generation tasks.", + "bbox": [ + 212, + 517, + 787, + 638 + ], + "page_idx": 13 + }, + { + "type": "text", + "text": "6 Conclusion", + "text_level": 1, + "bbox": [ + 215, + 659, + 359, + 674 + ], + "page_idx": 13 + }, + { + "type": "text", + "text": "In this paper, we introduce $\\mathrm{PIXART - \\Sigma}$ , a Text-to-Image (T2I) diffusion model capable of directly generating high-quality images at 4K resolution. Building upon the pre-trained foundation of $\\mathrm{PIXART - \\alpha}$ , $\\mathrm{PIXART - \\Sigma}$ achieves efficient training through a novel \"weak-to-strong training\" methodology. This approach is characterized by the incorporation of higher-quality data and the integration of efficient token compression. $\\mathrm{PIXART - \\Sigma}$ excels at producing high-fidelity images while adhering closely to textual prompts, surpassing the high standards set by its predecessor, $\\mathrm{PIXART - \\alpha}$ . We believe that the innovations presented in $\\mathrm{PIXART - \\Sigma}$ will not only contribute to advancements in the AIGC community but also pave the way for entities to access more efficient, and high-quality generative models.", + "bbox": [ + 212, + 688, + 787, + 840 + ], + "page_idx": 13 + }, + { + "type": "page_number", + "text": "14", + "bbox": [ + 217, + 114, + 235, + 126 + ], + "page_idx": 13 + }, + { + "type": "header", + "text": "J. Chen et al.", + "bbox": [ + 271, + 114, + 362, + 126 + ], + "page_idx": 13 + }, + { + "type": "text", + "text": "Acknowledgements. We would like to thank Zeqiang Lai, Fei Chen, and Shuchen Xue for discussing and helping implement the DMD part.", + "bbox": [ + 215, + 146, + 795, + 176 + ], + "page_idx": 14 + }, + { + "type": "text", + "text": "References", + "text_level": 1, + "bbox": [ + 215, + 202, + 321, + 218 + ], + "page_idx": 14 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "1. Aesthetic predictor (2023), https://github.com/christophschuhmann/improved-aesthetic-predictor7", + "2. Bao, F., Nie, S., Xue, K., Cao, Y., Li, C., Su, H., Zhu, J.: All are worth words: A vit backbone for diffusion models. In: CVPR (2023) 4", + "3. Chandra, A., Tünnermann, L., Lofstedt, T., Gratz, R.: Transformer-based deep learning for predicting protein properties in the life sciences. eLife Sciences Publications, Ltd (2023) 4", + "4. Chen, J., Yu, J., Ge, C., Yao, L., Xie, E., Wu, Y., Wang, Z., Kwok, J., Luo, P., Lu, H., Li, Z.: PixArt-α: Fast training of diffusion transformer for photorealistic text-to-image synthesis. In: ICLR (2024) 1, 4, 6, 8, 11, 12", + "5. Chen, L., Li, J., Dong, X., Zhang, P., He, C., Wang, J., Zhao, F., Lin, D.: ShareGPT4V: Improving large multi-modal models with better captions. In: arXiv (2023) 3, 8", + "6. Chen, S., Xu, M., Ren, J., Cong, Y., He, S., Xie, Y., Sinha, A., Luo, P., Xiang, T., Perez-Rua, J.M.: Gentron: Delving deep into diffusion transformers for image and video generation. arXiv preprint arXiv:2312.04557 (2023) 3, 4", + "7. Chen, T., Cheng, Y., Gan, Z., Yuan, L., Zhang, L., Wang, Z.: Chasing sparsity in vision transformers: An end-to-end exploration. In: NeurIPS (2021) 7", + "8. Chen, X., Liu, Z., Tang, H., Yi, L., Zhao, H., Han, S.: SparseViT: Revisiting activation sparsity for efficient high-resolution vision transformer. In: CVPR (2023) 7", + "9. Choromanski, K., Likhosherstov, V., Dohan, D., Song, X., Gane, A., Sarlos, T., Hawkins, P., Davis, J., Mohiuddin, A., Kaiser, L., et al.: Rethinking attention with performers. In: ICLR (2021) 7", + "0. Chung, H.W., Hou, L., Longpre, S., Zoph, B., Tay, Y., Fedus, W., Li, Y., Wang, X., Dehghani, M., Brahma, S., et al.: Scaling instruction-finetuned language models. In: arXiv (2022) 3, 11", + " 1. Du, R., Chang, D., Hospedales, T., Song, Y.Z., Ma, Z.: Demofusion: Democratising high-resolution image generation with no $$#. In: CVPR (2024) 12", + "2. Gao, S., Zhou, P., Cheng, M.M., Yan, S.: Masked diffusion transformer is a strong image synthesizer. In: ICCV (2023) 4", + "3. Ge, C., Ding, X., Tong, Z., Yuan, L., Wang, J., Song, Y., Luo, P.: Advancing vision transformers with group-mix attention. In: arXiv (2023) 7", + "4. von Glehn, I., Spencer, J.S., P fauna, D.: A self-attention ansatz for ab-initio quantum chemistry. In: ICLR (2023) 4", + "5. Hatamizadeh, A., Song, J., Liu, G., Kautz, J., Vahdat, A.: Diffit: Diffusion vision transformers for image generation. In: arXiv (2023) 4", + "6. He, Y., Yang, S., Chen, H., Cun, X., Xia, M., Zhang, Y., Wang, X., He, R., Chen, Q., Shan, Y.: ScaleCrafter: Tuning-free higher-resolution visual generation with diffusion models. In: ICLR (2024) 12", + "7. Heusel, M., Ramsauer, H., Unterthiner, T., Nessler, B., Hochreiter, S.: GANs trained by a two time-scale update rule converge to a local nash equilibrium. In: NeurIPS (2017) 8" + ], + "bbox": [ + 225, + 237, + 784, + 839 + ], + "page_idx": 14 + }, + { + "type": "header", + "text": "PIXART- $\\boldsymbol{\\Sigma}$ : Weak-to-Strong Training of DiT for 4K T2I Generation", + "bbox": [ + 281, + 114, + 730, + 128 + ], + "page_idx": 14 + }, + { + "type": "page_number", + "text": "15", + "bbox": [ + 767, + 116, + 784, + 126 + ], + "page_idx": 14 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "18. Kang, M., Zhu, J.Y., Zhang, R., Park, J., Shechtman, E., Paris, S., Park, T.: Scaling up gans for text-to-image synthesis. In: CVPR (2023) 6", + "19. Li, D., Kamko, A., Akhgari, E., Sabet, A., Xu, L., Doshi, S.: Playground v2.5: Three insights towards enhancing aesthetic quality in text-to-image generation. In: arXiv (2024) 6", + "20. Li, D., Kamko, A., Sabet, A., Akhgari, E., Xu, L., Doshi, S.: Playground v2, https://huggingface.co/playgroundai/playground-v2-1024px-aesthetic 12", + "21. Lin, T.Y., Maire, M., Belongie, S., Hays, J., Perona, P., Ramanan, D., Dólar, P., Zitnick, C.L.: Microsoft COCO: Common objects in context. In: ECCV (2014) 8", + "22. Liu, H., Li, C., Wu, Q., Lee, Y.J.: Visual instruction tuning. In: arXiv (2023) 3", + "23. Liu, Z., Lin, Y., Cao, Y., Hu, H., Wei, Y., Zhang, Z., Lin, S., Guo, B.: Swin transformer: Hierarchical vision transformer using shifted windows. In: ICCV (2021) 4, 7", + "24. Loshchilov, I., Hutter, F.: Decoupled weight decay regularization. In: arXiv (2017) 11", + "25. Lu, J., Yao, J., Zhang, J., Zhu, X., Xu, H., Gao, W., Xu, C., Xiang, T., Zhang, L.: Soft: Softmax-free transformer with linear complexity. In: NeurIPS (2021) 7", + "26. Lu, Z., Wang, Z., Huang, D., Wu, C., Liu, X., Ouyang, W., Bai, L.: Fit: Flexible vision transformer for diffusion model. In: arXiv (2024) 4", + "27. Luo, Y., Ren, X., Zheng, Z., Jiang, Z., Jiang, X., You, Y.: CAME: Confidence-guided adaptive memory efficient optimization. In: ACL (2023) 11", + "28. Ma, N., Goldstein, M., Albergo, M.S., Boffi, N.M., Vanden-Eijnden, E., Xie, S.: SiT: Exploring flow and diffusion-based generative models with scalable interpolant transformers. In: arXiv (2024) 4", + "29. Midjourney: Midjourney (2023), https://www.midjourney.com 1, 4", + "30. OpenAI: Dalle-2 (2023), https://openai.com/dall-e-2 6", + "31. OpenAI: Dalle-3 (2023), https://openai.com/dall-e-3 1, 4, 8", + "32. OpenAI: Gpt-4v(ison) system card. In: OpenAI (2023) 13", + "33. OpenAI: Sora (2024), https://openai.com/sora 3, 5", + "34. Peebles, W., Xie, S.: Scalable diffusion models with transformers. In: ICCV (2023) 4", + "35. Pernias, P., Rampas, D., Richter, M.L., Pal, C., Aubreville, M.: Würstchen: An efficient architecture for large-scale text-to-image diffusion models. In: ICLR (2023) 4, 6, 12", + "36. Podell, D., English, Z., Lacey, K., Blattmann, A., Dockhorn, T., Müller, J., Penna, J., Rombach, R.: SDXL: Improving latent diffusion models for high-resolution image synthesis. In: arXiv (2023) 1, 3, 4, 6, 11, 12", + "37. Radford, A., Narasimhan, K., Salimans, T., Sutskever, I., et al.: Improving language understanding by generative pre-training. OpenAI blog (2018) 4", + "38. Radford, A., Wu, J., Child, R., Luan, D., Amodei, D., Sutskever, I., et al.: Language models are unsupervised multitask learners. OpenAI blog (2019) 4", + "39. Rombach, R., Blattmann, A., Lorenz, D., Esser, P., Ommer, B.: High-resolution image synthesis with latent diffusion models. In: CVPR (2022) 6, 11, 12", + "40. Sahara, C., Chan, W., Saxena, S., Li, L., Whang, J., Denton, E.L., Ghasemipour, K., Gontijo Lopes, R., Karagol Ayan, B., Salimans, T., et al.: Photorealistic text-to-image diffusion models with deep language understanding. In: NeurIPS (2022) 6, 11", + "41. Sauer, A., Lorenz, D., Blattmann, A., Rombach, R.: Adversarial diffusion distillation. In: arXiv (2023) 12", + "42. Stability.AI: Stable diffusion 3 (2024), https://stability.ai/news/stable-diffusion-3 3" + ], + "bbox": [ + 215, + 146, + 784, + 837 + ], + "page_idx": 15 + }, + { + "type": "page_number", + "text": "16", + "bbox": [ + 217, + 114, + 235, + 126 + ], + "page_idx": 15 + }, + { + "type": "header", + "text": "J. Chen et al.", + "bbox": [ + 271, + 114, + 362, + 126 + ], + "page_idx": 15 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "43. Touvron, H., Cord, M., Douze, M., Massa, F., Sablayrolles, A., Jégou, H.: Training data-efficient image transformers & distillation through attention. In: ICML (2021) 4", + "44. Wang, S., Li, B.Z., Khabsa, M., Fang, H., Ma, H.: Linformer: Self-attention with linear complexity. In: arXiv (2020) 7", + "45. Wang, W., Xie, E., Li, X., Fan, D.P., Song, K., Liang, D., Lu, T., Luo, P., Shao, L.: Pyramid vision transformer: A versatile backbone for dense prediction without convolutions. In: ICCV (2021) 7", + "46. Wang, W., Xie, E., Li, X., Fan, D.P., Song, K., Liang, D., Lu, T., Luo, P., Shao, L.: Pvt v2: Improved baselines with pyramid vision transformer. Computational Visual Media (2022) 7, 9", + "47. Xia, Z., Pan, X., Song, S., Li, L.E., Huang, G.: Vision transformer with deformable attention. In: CVPR (2022) 7", + "48. Xie, E., Wang, W., Yu, Z., Anandkumar, A., Alvarez, J.M., Luo, P.: SegFormer: Simple and efficient design for semantic segmentation with transformers. In: NeurIPS (2021) 7", + "49. Xie, E., Yao, L., Shi, H., Liu, Z., Zhou, D., Liu, Z., Li, J., Li, Z.: DiffFit: Unlocking transferability of large diffusion models via simple parameter-efficient fine-tuning. In: ICCV (2023) 11", + "50. Xue, Z., Song, G., Guo, Q., Liu, B., Zong, Z., Liu, Y., Luo, P.: Raphael: Text-to-image generation via large mixture of diffusion paths. In: NeurIPS (2023) 12", + "51. Yuan, L., Chen, Y., Wang, T., Yu, W., Shi, Y., Jiang, Z.H., Tay, F.E., Feng, J., Yan, S.: Tokens-to-token ViT: Training vision transformers from scratch onImagenet. In: ICCV (2021) 4", + "52. Zheng, H., Nie, W., Vahdat, A., Anandkumar, A.: Fast training of diffusion models with masked transformers. In: arXiv (2023) 4", + "53. Zheng, S., Lu, J., Zhao, H., Zhu, X., Luo, Z., Wang, Y., Fu, Y., Feng, J., Xiang, T., Torr, P.H., et al.: Rethinking semantic segmentation from a sequence-to-sequence perspective with transformers. In: CVPR (2021) 4", + "54. Zhu, X., Su, W., Lu, L., Li, B., Wang, X., Dai, J.: Deformable DETR: Deformable transformers for end-to-end object detection. ICLR (2021) 7" + ], + "bbox": [ + 215, + 146, + 784, + 575 + ], + "page_idx": 16 + }, + { + "type": "header", + "text": "PixART- $\\boldsymbol{\\Sigma}$ : Weak-to-Strong Training of DiT for 4K T2I Generation", + "bbox": [ + 282, + 114, + 730, + 128 + ], + "page_idx": 16 + }, + { + "type": "page_number", + "text": "17", + "bbox": [ + 767, + 116, + 784, + 126 + ], + "page_idx": 16 + } +] \ No newline at end of file diff --git a/2024/PixArt-Sigma_ Weak-to-Strong Training of Diffusion Transformer for 4K Text-to-Image Generation/56a1c2d9-7461-45b0-bc11-505740d1999b_model.json b/2024/PixArt-Sigma_ Weak-to-Strong Training of Diffusion Transformer for 4K Text-to-Image Generation/56a1c2d9-7461-45b0-bc11-505740d1999b_model.json new file mode 100644 index 0000000000000000000000000000000000000000..0ab0055724645d693735e38b29294d9f174c7fb5 --- /dev/null +++ b/2024/PixArt-Sigma_ Weak-to-Strong Training of Diffusion Transformer for 4K Text-to-Image Generation/56a1c2d9-7461-45b0-bc11-505740d1999b_model.json @@ -0,0 +1,2401 @@ +[ + [ + { + "type": "title", + "bbox": [ + 0.223, + 0.141, + 0.782, + 0.187 + ], + "angle": 0, + "content": "PIXART- \\(\\boldsymbol{\\Sigma}\\): Weak-to-Strong Training of Diffusion Transformer for 4K Text-to-Image Generation" + }, + { + "type": "text", + "bbox": [ + 0.223, + 0.212, + 0.78, + 0.245 + ], + "angle": 0, + "content": "Junsong Chen\\(^{1,2,3*}\\), Chongjian Ge\\(^{1,3*}\\), Enze Xie\\(^{1*†}\\), Yue Wu\\(^{1*}\\), Lewei Yao\\(^{1,4}\\), Xiaozhe Ren\\(^{1}\\), Zhongdao Wang\\(^{1}\\), Ping Luo\\(^{3}\\), Huchuan Lu\\(^{2}\\), and Zhenguo Li\\(^{1}\\)" + }, + { + "type": "text", + "bbox": [ + 0.223, + 0.254, + 0.78, + 0.285 + ], + "angle": 0, + "content": "1 Huawei Noah's Ark Lab 2 Dalian University of Technology 3 HKU 4 HKUST Project Page: https://pixart-alpha.github.io/PixArt-sigma-project/" + }, + { + "type": "text", + "bbox": [ + 0.261, + 0.322, + 0.744, + 0.614 + ], + "angle": 0, + "content": "Abstract. In this paper, we introduce \\(\\mathrm{PIXART - \\Sigma}\\), a Diffusion Transformer model (DiT) capable of directly generating images at 4K resolution. \\(\\mathrm{PIXART - \\Sigma}\\) represents a significant advancement over its predecessor, \\(\\mathrm{PIXART - \\alpha}\\), offering images of markedly higher fidelity and improved alignment with text prompts. A key feature of \\(\\mathrm{PIXART - \\Sigma}\\) is its training efficiency. Leveraging the foundational pre-training of \\(\\mathrm{PIXART - \\alpha}\\), it evolves from the 'weaker' baseline to a 'stronger' model via incorporating higher quality data, a process we term \"weak-to-strong training\". The advancements in \\(\\mathrm{PIXART - \\Sigma}\\) are twofold: (1) High-Quality Training Data: \\(\\mathrm{PIXART - \\Sigma}\\) incorporates superior-quality image data, paired with more precise and detailed image captions. (2) Efficient Token Compression: we propose a novel attention module within the DiT framework that compresses both keys and values, significantly improving efficiency and facilitating ultra-high-resolution image generation. Thanks to these improvements, \\(\\mathrm{PIXART - \\Sigma}\\) achieves superior image quality and user prompt adherence capabilities with significantly smaller model size (0.6B parameters) than existing text-to-image diffusion models, such as SDXL (2.6B parameters) and SD Cascade (5.1B parameters). Moreover, \\(\\mathrm{PIXART - \\Sigma}\\)'s capability to generate 4K images supports the creation of high-resolution posters and wallpapers, efficiently bolstering the production of high-quality visual content in industries such as film and gaming." + }, + { + "type": "text", + "bbox": [ + 0.262, + 0.627, + 0.71, + 0.642 + ], + "angle": 0, + "content": "Keywords: T2I Synthesis, Diffusion Transformer, Efficient Model" + }, + { + "type": "title", + "bbox": [ + 0.216, + 0.668, + 0.377, + 0.684 + ], + "angle": 0, + "content": "1 Introduction" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.7, + 0.788, + 0.807 + ], + "angle": 0, + "content": "The recent emergence of high-quality Text-to-Image (T2I) models has profoundly impacted the AI Generated Content (AIGC) community. This includes both proprietary models such as DALL-E 3 [31], Midjourney [29], as well as open-source models like Stable Diffusion [36] and PixART-\\(\\alpha\\) [4]. Nonetheless, developing a top-tier T2I model involves considerable resources; for instance, training SD1.5 from scratch necessitates about 6000 A100 GPU days [36], posing a substantial barrier to individual researchers with limited resources and impeding innovation" + }, + { + "type": "page_footnote", + "bbox": [ + 0.216, + 0.817, + 0.782, + 0.84 + ], + "angle": 0, + "content": "*Equal contribution. Work done during the students' internships at Huawei Noah's Ark Lab. †Project lead and corresponding author. Thanks to Charan for the 4K dataset collection." + } + ], + [ + { + "type": "page_number", + "bbox": [ + 0.218, + 0.116, + 0.23, + 0.127 + ], + "angle": 0, + "content": "2" + }, + { + "type": "header", + "bbox": [ + 0.272, + 0.115, + 0.365, + 0.127 + ], + "angle": 0, + "content": "J. Chen et al." + }, + { + "type": "image", + "bbox": [ + 0.244, + 0.145, + 0.761, + 0.764 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.216, + 0.773, + 0.788, + 0.803 + ], + "angle": 0, + "content": "Fig. 1: Images generated by \\(\\mathbf{P}\\mathbf{I}\\mathbf{x}\\mathbf{A}\\mathbf{R}\\mathbf{T} - \\boldsymbol{\\Sigma}\\). The model can output photo-realistic, high aesthetic, extreme aspect ratio, multi-style images, and follow user instructions." + } + ], + [ + { + "type": "header", + "bbox": [ + 0.283, + 0.114, + 0.732, + 0.129 + ], + "angle": 0, + "content": "PixART- \\(\\boldsymbol{\\Sigma}\\) : Weak-to-Strong Training of DiT for 4K T2I Generation" + }, + { + "type": "page_number", + "bbox": [ + 0.775, + 0.117, + 0.785, + 0.127 + ], + "angle": 0, + "content": "3" + }, + { + "type": "text", + "bbox": [ + 0.219, + 0.146, + 0.788, + 0.221 + ], + "angle": 0, + "content": "within the AIGC community. Over time, the AIGC community will gain access to continuously updated, higher-quality datasets and more advanced algorithms. A pivotal question is: how can we efficiently integrate these new elements into an existing model, achieving a more powerful version within the constraints of limited resources?" + }, + { + "type": "text", + "bbox": [ + 0.219, + 0.23, + 0.788, + 0.38 + ], + "angle": 0, + "content": "To explore this issue, our research focuses on enhancing \\(\\mathrm{PIXART - \\alpha}\\), an efficient T2I training method. \\(\\mathrm{PIXART - \\alpha}\\) represents an early venture within the DiT framework, a model structure with significant potential, as evidenced by works such as GenTron [6], Sora [33] and Stable Diffusion 3 [42]. To maximize this potential, we build upon the pre-trained foundation of \\(\\mathrm{PIXART - \\alpha}\\), integrating advanced elements to facilitate its continuous improvement, resulting in a more powerful model, \\(\\mathrm{PIXART - \\Sigma}\\). We refer to this process of evolving from a relatively weaker baseline to a stronger model through efficient training as \"weak-to-strong training\". Specifically, to achieve \"weak-to-strong training\", we introduce the following enhancements:" + }, + { + "type": "text", + "bbox": [ + 0.219, + 0.388, + 0.788, + 0.584 + ], + "angle": 0, + "content": "Higher-Quality Training Data: We collect a high-quality dataset superior to that used in \\(\\mathrm{PIXART - \\alpha}\\), focusing on two key aspects: (i) High-quality images: The dataset comprises 33M high-resolution images sourced from the Internet, all exceeding 1K resolution, including 2.3M images with resolutions around 4K. These images are predominantly characterized by their high aesthetic and encompass a wide range of artistic styles. (ii) Dense and accurate captions: To provide more precise and detailed captions for the aforementioned images, we replace the LLaVA [22] used in \\(\\mathrm{PIXART - \\alpha}\\) with a more powerful image captioner, Share-Captioner [5]. Furthermore, to improve the model's alignment capacity between the textual and visual concepts, we extend the token length of the text encoder (i.e., Flan-T5 [10]) to approximately 300 words. We observe these improvements effectively eliminate the model's tendency for hallucination, leading to higher-quality text-image alignment." + }, + { + "type": "text", + "bbox": [ + 0.219, + 0.592, + 0.788, + 0.742 + ], + "angle": 0, + "content": "Efficient Token Compression: To enhance \\(\\mathrm{PIXART - \\alpha}\\), we expand its generation resolution from 1K to 4K. Generating images at ultra-high resolutions (e.g., 2K/4K) introduces a significant increase in the number of tokens, leading to a substantial rise in computational demand. To address this challenge, we introduced a self-attention module with key and value token compression tailored to the DiT framework. Specifically, we utilize group convolutions with a stride of 2 for local aggregation of keys and values. Additionally, we employ a specialized weight initialization scheme, allowing for a smooth adaptation from a pre-trained model without KV compression. This design effectively reduces training and inference time by \\(\\sim 34\\%\\) for high-resolution image generation." + }, + { + "type": "text", + "bbox": [ + 0.219, + 0.75, + 0.788, + 0.84 + ], + "angle": 0, + "content": "Weak-to-Strong Training Strategy: we propose several fine-tuning techniques to rapidly adapt from a weak model to a strong model efficiently. That includes (1) replacing with a more powerful Variational Autoencoder (VAE) [36], (2) scaling from low to high resolution, and (3) evolving from a model without Key-Value (KV) compression to one with KV compression. These outcomes confirm the validity and effectiveness of the \"weak-to-strong training\" approach." + } + ], + [ + { + "type": "page_number", + "bbox": [ + 0.218, + 0.116, + 0.23, + 0.127 + ], + "angle": 0, + "content": "4" + }, + { + "type": "header", + "bbox": [ + 0.272, + 0.115, + 0.364, + 0.127 + ], + "angle": 0, + "content": "J. Chen et al." + }, + { + "type": "image", + "bbox": [ + 0.219, + 0.145, + 0.78, + 0.416 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.214, + 0.424, + 0.788, + 0.466 + ], + "angle": 0, + "content": "Fig. 2: 4K image generation with complex dense instructions. \\(\\mathrm{PIXART - \\Sigma}\\) can directly generate 4K resolution images without post-processing, and accurately respond to the given prompt." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.468, + 0.788, + 0.635 + ], + "angle": 0, + "content": "Through the proposed improvements, \\(\\mathrm{PIXART - \\Sigma}\\) achieves high-quality 4K resolution image generation at a minimal training cost and model parameters. Specifically, fine-tuning from a pre-trained model, we additionally utilize only \\(9\\%\\) of the GPU days required by \\(\\mathrm{PIXART - \\alpha}\\) to achieve a strong 1K high-resolution image generation model, which is impressive considering replacing with new training data and a more powerful VAE. Moreover, we use only 0.6B parameters while SDXL [36] and SD Cascade [35] use 2.6B and 5.1B parameters respectively. Images generated by \\(\\mathrm{PIXART - \\Sigma}\\) possess an aesthetic quality comparable to current top-tier T2I products, such as DALL-E 3 [31] and MJV6 [29] (as in Fig. 4). Additionally, \\(\\mathrm{PIXART - \\Sigma}\\) also demonstrates exceptional capability for fine-grained alignment with textual prompts (as shown in Fig. 2 and 3)." + }, + { + "type": "title", + "bbox": [ + 0.216, + 0.657, + 0.388, + 0.673 + ], + "angle": 0, + "content": "2 Related Work" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.689, + 0.789, + 0.841 + ], + "angle": 0, + "content": "Diffusion Transformers. The Transformer architecture has achieved remarkable success across various domains, such as language modeling [37,38], computer vision [23, 43, 51, 53], and other areas [3, 14]. In the realm of diffusion models, DiT [34] and UViT [2] pioneer the use of Transformer architecture. Subsequent works, including DiffiT [15], SiT [28], and FiT [26], have improved upon DiT's architecture, while [12, 52] enhance training efficiency through masked modeling techniques. For Text-to-Image (T2I) synthesis, PixART-\\(\\alpha\\) [4] explore efficient T2I training schemes, achieving the first Transformer-based T2I model capable of generating 1024px high-quality images. GenTron [6] explores the flexibility and scalability of diffusion Transformers in both image and video generation. The re" + } + ], + [ + { + "type": "header", + "bbox": [ + 0.283, + 0.115, + 0.731, + 0.129 + ], + "angle": 0, + "content": "PixART- \\(\\boldsymbol{\\Sigma}\\) : Weak-to-Strong Training of DiT for 4K T2I Generation" + }, + { + "type": "page_number", + "bbox": [ + 0.775, + 0.117, + 0.785, + 0.127 + ], + "angle": 0, + "content": "5" + }, + { + "type": "image", + "bbox": [ + 0.22, + 0.145, + 0.403, + 0.301 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.406, + 0.145, + 0.591, + 0.301 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.593, + 0.145, + 0.779, + 0.301 + ], + "angle": 0, + "content": null + }, + { + "type": "image_footnote", + "bbox": [ + 0.223, + 0.304, + 0.776, + 0.325 + ], + "angle": 0, + "content": "Prompt: A close-up photo of a person. The subject is a woman. She wore a blue coat with a gray dress underneath. She has blue eyes and blond hair, and wears a pair of earrings. Behind are blurred city buildings and streets." + }, + { + "type": "image", + "bbox": [ + 0.221, + 0.327, + 0.404, + 0.468 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.414, + 0.327, + 0.591, + 0.467 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.595, + 0.327, + 0.689, + 0.468 + ], + "angle": 0, + "content": null + }, + { + "type": "image_footnote", + "bbox": [ + 0.223, + 0.47, + 0.562, + 0.479 + ], + "angle": 0, + "content": "Prompt: half a solid black background and half a solid white background" + }, + { + "type": "image", + "bbox": [ + 0.221, + 0.482, + 0.404, + 0.622 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.408, + 0.481, + 0.591, + 0.622 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.594, + 0.481, + 0.778, + 0.622 + ], + "angle": 0, + "content": null + }, + { + "type": "image_footnote", + "bbox": [ + 0.221, + 0.624, + 0.758, + 0.656 + ], + "angle": 0, + "content": "Prompt: Pixel art style of a snowboarder in mid-air performs a trick on a black rail, wearing a blue sweatshirt and black pants, with arms outstretched. The serene snowy landscape background, dotted with trees, complements the scene. The low-angle perspective emphasizes the trick's height and skill." + }, + { + "type": "image_caption", + "bbox": [ + 0.214, + 0.663, + 0.788, + 0.72 + ], + "angle": 0, + "content": "Fig. 3: Comparison of \\(\\mathbf{P}\\mathbf{i}\\mathbf{x}\\mathbf{A}\\mathbf{r}\\mathbf{T} - \\boldsymbol{\\Sigma}\\) with open-source models, e.g., \\(\\mathbf{P}\\mathbf{i}\\mathbf{x}\\mathbf{A}\\mathbf{r}\\mathbf{T} - \\alpha\\) and SDXL: Compared with \\(\\mathbf{P}\\mathbf{i}\\mathbf{x}\\mathbf{A}\\mathbf{r}\\mathbf{T} - \\alpha\\), \\(\\mathbf{P}\\mathbf{i}\\mathbf{x}\\mathbf{A}\\mathbf{r}\\mathbf{T} - \\boldsymbol{\\Sigma}\\) improves the realism of portraits and the capability of semantic analysis. Compared with SDXL, our method has a better ability to follow user instructions. The keywords are highlighted as blue." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.73, + 0.788, + 0.806 + ], + "angle": 0, + "content": "cent advent of the powerful video generation model Sora [33] has further underscored the potential of Diffusion Transformers. In this work, for the first time, we explore using the Transformer architecture to generate 4K ultra-high-resolution images directly, tackling the computational complexity challenges posed by involving long-sequence tokens." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.81, + 0.788, + 0.841 + ], + "angle": 0, + "content": "High Resolution Image Generation greatly enhances visual quality and is important in various industries such as film and gaming. However, increasing" + } + ], + [ + { + "type": "page_number", + "bbox": [ + 0.218, + 0.116, + 0.23, + 0.127 + ], + "angle": 0, + "content": "6" + }, + { + "type": "header", + "bbox": [ + 0.272, + 0.115, + 0.364, + 0.127 + ], + "angle": 0, + "content": "J. Chen et al." + }, + { + "type": "image", + "bbox": [ + 0.218, + 0.144, + 0.498, + 0.362 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.498, + 0.144, + 0.639, + 0.362 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.64, + 0.144, + 0.78, + 0.362 + ], + "angle": 0, + "content": null + }, + { + "type": "text", + "bbox": [ + 0.219, + 0.365, + 0.512, + 0.374 + ], + "angle": 0, + "content": "Prompt: a small cactus with a happy face in the Sahara desert" + }, + { + "type": "image", + "bbox": [ + 0.218, + 0.376, + 0.498, + 0.593 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.498, + 0.376, + 0.639, + 0.593 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.64, + 0.376, + 0.78, + 0.593 + ], + "angle": 0, + "content": null + }, + { + "type": "text", + "bbox": [ + 0.219, + 0.594, + 0.365, + 0.605 + ], + "angle": 0, + "content": "Prompt: a cherry pick scientist" + }, + { + "type": "image_caption", + "bbox": [ + 0.216, + 0.61, + 0.788, + 0.653 + ], + "angle": 0, + "content": "Fig. 4: Compare \\(\\mathbf{P}\\mathbf{i}\\mathbf{x}\\mathbf{A}\\mathbf{r}\\mathbf{T} - \\boldsymbol{\\Sigma}\\) and four other T2I products: Firefly 2, Imagen 2, Dalle 3, and Midjourney 6. Images generated by \\(\\mathbf{P}\\mathbf{i}\\mathbf{x}\\mathbf{A}\\mathbf{r}\\mathbf{T} - \\boldsymbol{\\Sigma}\\) are very competitive with these commercial products." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.674, + 0.789, + 0.842 + ], + "angle": 0, + "content": "image resolution introduces challenges due to the substantial increase in computational demands. Numerous methods have been explored in this direction. For instance, Imagen [40], GigaGAN [18] and Stable Diffusion [39] introduce an additional super-resolution network, while Stable Cascade [35] employs multiple diffusion networks to increase resolution progressively. These combined-model solutions, however, can introduce cumulative errors. On the other hand, works like SDXL [36], DALL-E 2 [30], Playground [19] and \\(\\mathrm{PIXART - \\alpha}\\) [4] have attempted to generate high-resolution images using diffusion models directly. Nevertheless, these efforts are capped at generating images with resolutions up to \\(1024\\mathrm{px}\\) due to computational complexity. In this paper, we push this boundary to \\(4\\mathrm{K}\\) resolution, significantly enhancing the visual quality of the generated content." + } + ], + [ + { + "type": "header", + "bbox": [ + 0.282, + 0.114, + 0.732, + 0.129 + ], + "angle": 0, + "content": "PIXART- \\(\\boldsymbol{\\Sigma}\\) : Weak-to-Strong Training of DiT for 4K T2I Generation" + }, + { + "type": "page_number", + "bbox": [ + 0.775, + 0.116, + 0.785, + 0.127 + ], + "angle": 0, + "content": "7" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.147, + 0.788, + 0.36 + ], + "angle": 0, + "content": "Efficient Transformer architecture. The self-attention mechanism in Transformer suffers from quadratic computational complexity with respect to the number of tokens, which hinders the scaling up of token quantity. Many works have sought improvements in this area: (1) Sparse Attention [7, 8, 45, 46, 48], which reduces the overall computational load by selectively processing a subset of tokens. For instance, PVT v2 [46] employs a convolutional kernel to condense the space of the key and value, thus lowering the complexity involved in computing the attention. (2) Local Attention [13, 23, 47, 54] focuses attention within nearby regions; notably, Swin Transformer [23] utilizes window-based attention to limit computations to within a specified window size. (3) Low-rank/Linear Attention [9, 25, 44]. The Linformer [44] reduces the computational complexity of the self-attention mechanism through low-rank approximations. In this paper, inspired by PVT v2 [46], we employ a self-attention mechanism based on key/value compression to mitigate the high complexity of processing 4K images." + }, + { + "type": "title", + "bbox": [ + 0.216, + 0.381, + 0.362, + 0.397 + ], + "angle": 0, + "content": "3 Framework" + }, + { + "type": "title", + "bbox": [ + 0.216, + 0.412, + 0.379, + 0.427 + ], + "angle": 0, + "content": "3.1 Data Analysis" + }, + { + "type": "table", + "bbox": [ + 0.223, + 0.429, + 0.78, + 0.572 + ], + "angle": 0, + "content": "
LLaVA HallucinationsThe image features a large, ornate church with a tall, pointed roof and a large stained-glass window. The church is situated in a city, and it has a white and gray color scheme. The style of the church is Gothic, which is evident from the pointed roof and the intricate details of the stained-glass window. The presence of statues and a clock on the building further emphasizes the grand and historical nature of the structure.
Share-Captioner CorrectnessThe image captures the grandeur of a cathedral, painted in a pristine white, stands majestically against the backdrop of a clear blue sky. The entrance to the cathedral is adorned with a large rose window. Flanking the entrance are statues of saints, standing as silent guardians of the sacred space within. The perspective of the photo, taken from a low angle. The lush green trees in the background adding serenity to the scene.
LLaVA HallucinationsThe image features a woman and a man sitting on a brick walkway near a body of water, which could be a river or a lake. They are both wearing head coverings, and the woman is holding a handbag. The scene is set during the day, with the sun shining brightly, creating a warm and inviting atmosphere. The style of the image is a black and white photo, which adds a timeless and classic feel to the scene.
Share-Captioner CorrectnessThe image captures a serene scene at a harbor. Two individuals are seated on a bench, their backs to the camera, engrossed in the view of the water. The water, a deep shade of blue, is dotted with boats of various sizes and colors, including a white boat with a green stripe and a red boat. The sky above is a light blue.
" + }, + { + "type": "image_caption", + "bbox": [ + 0.214, + 0.58, + 0.788, + 0.623 + ], + "angle": 0, + "content": "Fig. 5: Comparative illustration of hallucinations: Contrasting differences in hallucination occurrences between LLaVA and Share-Captioner, with red indicating hallucinations and green denoting correctness." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.629, + 0.788, + 0.734 + ], + "angle": 0, + "content": "Higher Aesthetic and higher Resolution. To enhance the aesthetic quality of our dataset, we expand our internal data from 14M to 33M. For clarity, we name the two datasets Internal- \\(\\alpha\\) and Internal- \\(\\Sigma\\), respectively. Note that this expansion still falls short compared to the vast images utilized by currently available open-source models like SD v1.5, which uses 2B data. We demonstrate that effective training strategies with limited data amount can still obtain a strong T2I model." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.735, + 0.788, + 0.81 + ], + "angle": 0, + "content": "The images within Internal- \\(\\Sigma\\) are above 1K resolution. To facilitate 4K resolution generation, we additionally collect a dataset of 8M real photographic images at 4K resolution. To ensure aesthetic quality, we employ an aesthetic scoring model (AES) [1] to filter these 4K images. This process yields a highly refined dataset of 2M ultra-high-resolution and high-quality images." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.81, + 0.788, + 0.84 + ], + "angle": 0, + "content": "Interestingly, we have observed that as the resolution of the images increases, there is an improvement in the model's fidelity (Fréchet Inception Dis-" + } + ], + [ + { + "type": "page_number", + "bbox": [ + 0.218, + 0.116, + 0.23, + 0.127 + ], + "angle": 0, + "content": "8" + }, + { + "type": "header", + "bbox": [ + 0.272, + 0.115, + 0.364, + 0.127 + ], + "angle": 0, + "content": "J. Chen et al." + }, + { + "type": "table_caption", + "bbox": [ + 0.216, + 0.133, + 0.788, + 0.175 + ], + "angle": 0, + "content": "Table 1: Statistics of noun concepts for different datasets. VN: valid distinct nouns (appearing more than 10 times); DN: total distinct nouns; Average: average noun count per image; ACL: Average Caption length." + }, + { + "type": "table", + "bbox": [ + 0.231, + 0.177, + 0.764, + 0.296 + ], + "angle": 0, + "content": "
DatasetVolumeCaptionVN/DNTotal NounACLAverage
Internal-α14MRaw187K/931K175M2511.7/Img
Internal-α14MLLaVA28K/215K536M9829.3/Img
Internal-α14MShare-Captioner51K/420K815M18454.4/Img
Internal-Σ33MRaw294K/1512K485M3514.4/Img
Internal-Σ33MShare-Captioner77K/714K1804M18053.6/Img
4K-Σ2.3MShare-Captioner24K/96K115M16349.5/Img
" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.316, + 0.784, + 0.347 + ], + "angle": 0, + "content": "tance (FID) [17]) and semantic alignment (CLIP Score), which underscores the importance of the capabilities of generating high-resolution images." + }, + { + "type": "title", + "bbox": [ + 0.216, + 0.349, + 0.486, + 0.363 + ], + "angle": 0, + "content": "Better Text-Image Alignment." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.364, + 0.487, + 0.622 + ], + "angle": 0, + "content": "Recent works such as \\(\\mathrm{PIXART - }\\alpha\\) [4] and DALL-E 3 [31] emphasize the significance of text-image alignment. Strengthening this alignment is crucial for boosting model capabilities. To refine our collected \"raw\" descriptions further, we focus on improving both the length and accuracy of our captions. Notably, our captions (Internal- \\(\\Sigma\\) ) show several advantages over the one used in \\(\\mathrm{PIXART - }\\alpha\\) (Internal- \\(\\alpha\\) ) in the following aspects: 1. Enhanced caption accuracy: As depicted in Fig. 5, LLaVa used in \\(\\mathrm{PIXART - }\\alpha\\) encounters certain hallucination. We leverage a more powerful model, i.e., Share-Captioner [5], to" + }, + { + "type": "image", + "bbox": [ + 0.504, + 0.355, + 0.783, + 0.512 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.497, + 0.518, + 0.788, + 0.602 + ], + "angle": 0, + "content": "Fig.6: Histogram Visualization of the Caption Length. We randomly select 1M captions from the raw captions, Internal- \\(\\alpha\\) , and Internal- \\(\\Sigma\\) to draw the corresponding histogram. ACL denotes the average caption length." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.623, + 0.78, + 0.639 + ], + "angle": 0, + "content": "generate detailed and correct captions, augmenting the collected raw prompts." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.64, + 0.788, + 0.761 + ], + "angle": 0, + "content": "2. Increased caption length: As shown in Tab. 1 and Fig. 6, the average caption length increased significantly to 180 words, highly enhancing the descriptive power of the captions. Additionally, we extend the token processing length of the text encoder from 120 tokens (as in Internal-\\(\\alpha\\)) to 300 tokens. Our model is trained on a mix of long (Share-Captioner) and short (raw) captions with a ratio of \\(60\\%\\) and \\(40\\%\\), respectively. This approach enhances the diversity of textual descriptions and mitigates potential biases that might arise from solely relying on generative captions." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.763, + 0.788, + 0.808 + ], + "angle": 0, + "content": "Tab. 1 demonstrates a summary for both Internal-\\(\\alpha\\) and \\(-\\Sigma\\), where we assess the diversity of the datasets through various metrics, including the noun variety, total noun count, average caption length, and average nouns per image." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.81, + 0.803, + 0.84 + ], + "angle": 0, + "content": "High-Quality Evaluation Dataset. Most SoTA T2I models chose MSCOCO [21] as the evaluation set to assess the FID and CLIP Scores. However, we observe" + } + ], + [ + { + "type": "header", + "bbox": [ + 0.282, + 0.114, + 0.732, + 0.129 + ], + "angle": 0, + "content": "PIXART- \\(\\Sigma\\) : Weak-to-Strong Training of DiT for 4K T2I Generation" + }, + { + "type": "page_number", + "bbox": [ + 0.775, + 0.116, + 0.785, + 0.127 + ], + "angle": 0, + "content": "9" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.147, + 0.788, + 0.284 + ], + "angle": 0, + "content": "evaluations conducted on the MSCOCO dataset may not adequately reflect a model's capabilities in aesthetics and text-image alignment. Therefore, we propose a curated set comprising 30,000 high-quality, aesthetically pleasing text-image pairs to facilitate the assessment. The selected samples of the dataset are presented in the appendix. This dataset is designed to provide a more comprehensive evaluation of a model's performance, particularly in capturing the intricacies of aesthetic appeal and the fidelity of alignment between textual descriptions and visual content. Unless otherwise specified, the evaluation experiments in the paper are conducted in the collected High-Quality Evaluation Dataset." + }, + { + "type": "title", + "bbox": [ + 0.216, + 0.314, + 0.431, + 0.33 + ], + "angle": 0, + "content": "3.2 Efficient DiT Design" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.348, + 0.788, + 0.483 + ], + "angle": 0, + "content": "An efficient DiT network is essential since the computational demand significantly increases when generating images at ultra-high resolutions. The attention mechanism plays a pivotal role in the efficacy of Diffusion Transformers, yet its quadratic computational demands significantly limit model scalability, particularly at higher resolutions e.g., 2K and 4K. Inspired by PVT v2 [46], we incorporate KV compression within the original PixART-\\(\\alpha\\)s framework to address the computational challenges. This design adds a mere \\(0.018\\%\\) to the total parameters yet achieves efficient reduction in computational costs via token compression, while still preserving both spatial and semantic information." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.485, + 0.788, + 0.606 + ], + "angle": 0, + "content": "Key-Value (KV) Token Compression. Our motivation stems from the intriguing observation that applying key-value (KV) token compression directly to the pre-trained \\(\\mathrm{PIXART - \\alpha}\\) can still generate reasonable images. This suggests a redundancy in the features. Considering the high similarity within adjacent \\(R\\times R\\) patches, we assume that feature semantics within a window are redundant and can be compressed reasonably. We propose KV token compression, which is denoted as \\(f_{c}(\\cdot)\\), to compress token features within a \\(R\\times R\\) window through a compression operator, as depicted in Fig. 7." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.608, + 0.494, + 0.841 + ], + "angle": 0, + "content": "Furthermore, to mitigate the potential information loss caused by KV compression in self-attention computation, we opt to retain all the tokens of queries (Q). This strategy allows us to utilize KV compression effectively while mitigating the risk of losing crucial information. By employing KV compression, we enhance the efficiency of attention computations and reduce the computation complexity from \\( O(N^2) \\) to \\( O\\left(\\frac{N^2}{R^2}\\right) \\), thereby making the computational cost of directly generating high-resolution images manageable." + }, + { + "type": "image", + "bbox": [ + 0.506, + 0.64, + 0.788, + 0.759 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.503, + 0.769, + 0.788, + 0.813 + ], + "angle": 0, + "content": "Fig. 7: Design of KV Token Compression. We merge KV tokens in spatial space to reduce the computation complexity." + } + ], + [ + { + "type": "page_number", + "bbox": [ + 0.218, + 0.116, + 0.236, + 0.127 + ], + "angle": 0, + "content": "10" + }, + { + "type": "header", + "bbox": [ + 0.272, + 0.115, + 0.364, + 0.127 + ], + "angle": 0, + "content": "J. Chen et al." + }, + { + "type": "table", + "bbox": [ + 0.218, + 0.135, + 0.782, + 0.322 + ], + "angle": 0, + "content": "
StrategiesVAE Change512 →1024 RA512 →1024 RA + PE Interp.RA + PE Adjust + KV CompressRA + PE Interp. + KV Compress + Conv Avg Init
0 step
Fast \nAdapt2K step100 step100 step100 step100 step
" + }, + { + "type": "table_caption", + "bbox": [ + 0.324, + 0.323, + 0.339, + 0.333 + ], + "angle": 0, + "content": "(a)" + }, + { + "type": "table_caption", + "bbox": [ + 0.468, + 0.323, + 0.484, + 0.333 + ], + "angle": 0, + "content": "(b)" + }, + { + "type": "table_caption", + "bbox": [ + 0.673, + 0.324, + 0.689, + 0.333 + ], + "angle": 0, + "content": "(c)" + }, + { + "type": "image_caption", + "bbox": [ + 0.216, + 0.334, + 0.785, + 0.375 + ], + "angle": 0, + "content": "Fig. 8: This illustration demonstrates how our training strategy designs expedite the model's convergence during the transition to VAEs, adjustment to higher resolutions, and the KV compression process, facilitating rapid learning from weak to strong." + }, + { + "type": "equation", + "bbox": [ + 0.316, + 0.388, + 0.785, + 0.423 + ], + "angle": 0, + "content": "\\[\n\\operatorname {A t t e n t i o n} (Q, K, V) = \\operatorname {s o f t m a x} \\left(\\frac {Q \\cdot f _ {c} (K) ^ {T}}{\\sqrt {d _ {k}}}\\right) f _ {c} (V) \\tag {1}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.428, + 0.785, + 0.534 + ], + "angle": 0, + "content": "We compress deep layers using the convolution operator \"Conv2×2\" with specific initialization. Detailed experiments on other design variants are discussed in Sec. 5. Specifically, we design a specialized convolution kernel initialization \"Conv Avg Init\" that utilizes group convolution and initializes the weights \\( w = \\frac{1}{R^2} \\), equivalent to an average operator. This initialization strategy can initially produce coarse results, accelerating the fine-tuning process while only introducing 0.018% additional parameters." + }, + { + "type": "title", + "bbox": [ + 0.216, + 0.557, + 0.544, + 0.573 + ], + "angle": 0, + "content": "3.3 Weak-to-Strong Training Strategy" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.583, + 0.785, + 0.629 + ], + "angle": 0, + "content": "We propose several efficient training strategies to enhance the transition from a \"weak\" model to a \"strong\" model. These strategies encompass VAE rapid adaptation, high-resolution fine-tuning, and KV Token compression." + }, + { + "type": "text", + "bbox": [ + 0.216, + 0.63, + 0.494, + 0.779 + ], + "angle": 0, + "content": "Adapting model to new VAEs. As VAEs develop, training T2I models from scratch is resource-intensive. We replace PixART- \\(\\alpha\\) s VAE with SDXL's VAE and continue fine-tuning the diffusion model. We observe a rapid convergence phenomenon that fine-tuning quickly converges at 2K training steps as shown in Fig 8 (a). Fine-tuning is more efficient when dealing with VAE" + }, + { + "type": "text", + "bbox": [ + 0.503, + 0.63, + 0.787, + 0.701 + ], + "angle": 0, + "content": "Table 2: We fine-tune a high-resolution model from a low-resolution model and observe that even fine-tuning for a relatively short duration, such as 1K steps, can still yield high-quality results." + }, + { + "type": "table", + "bbox": [ + 0.532, + 0.707, + 0.749, + 0.771 + ], + "angle": 0, + "content": "
ResolutionIterationsFID ↓CLIP ↑
25620K16.560.270
256 → 5121K9.750.272
256 → 512100K8.910.276
" + }, + { + "type": "text", + "bbox": [ + 0.216, + 0.78, + 0.714, + 0.794 + ], + "angle": 0, + "content": "model transferring and negates the necessity of training from scratch." + }, + { + "type": "text", + "bbox": [ + 0.216, + 0.795, + 0.787, + 0.841 + ], + "angle": 0, + "content": "Adapting to Higher-Resolution. When we fine-tune from a low-resolution (LR) model to a high-resolution (HR) model, we observe a performance degradation in Fig. 8 (b), which we attribute to discrepancies in positional embeddings" + } + ], + [ + { + "type": "header", + "bbox": [ + 0.282, + 0.114, + 0.732, + 0.129 + ], + "angle": 0, + "content": "PIXART- \\(\\Sigma\\) : Weak-to-Strong Training of DiT for 4K T2I Generation" + }, + { + "type": "page_number", + "bbox": [ + 0.768, + 0.116, + 0.784, + 0.127 + ], + "angle": 0, + "content": "11" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.146, + 0.788, + 0.283 + ], + "angle": 0, + "content": "(PE) between different resolutions. To address this, we utilize the \"PE Interpolation\" trick [4, 49]: initializing the HR model's PE by interpolating the LR model's PE, significantly enhancing the HR model's initial status and expediting fine-tuning. We can obtain visually pleasing images even within only 100 training iterations. Besides, we quantitatively compare the model's performance as illustrated in Tab. 2. The fine-tuning quickly converges at 1K steps, and further training slightly improves the performance. This illustrates that using the \"PE Interpolation\" trick enables rapid convergence of higher resolution generation, obviating the need for training from scratch for generating at higher resolutions." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.283, + 0.788, + 0.39 + ], + "angle": 0, + "content": "Adapting model to KV compression. We can use KV compression directly when fine-tuning from LR pre-trained models without KV compression. As shown in Fig. 8 (c), with our \"Conv Avg Init.\" strategy, \\(\\mathrm{PIXART - \\Sigma}\\) starts from a better initial state, making converging easier and faster. Notably, \\(\\mathrm{PIXART - \\Sigma}\\) performs satisfied visual results even within 100 training steps. Finally, through the KV compression operators and compression layers design in Sec 3.2, we can reduce \\(\\sim 34\\%\\) of the training and inference time." + }, + { + "type": "title", + "bbox": [ + 0.216, + 0.408, + 0.367, + 0.425 + ], + "angle": 0, + "content": "4 Experiment" + }, + { + "type": "title", + "bbox": [ + 0.216, + 0.434, + 0.457, + 0.45 + ], + "angle": 0, + "content": "4.1 Implementation Details" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.455, + 0.788, + 0.636 + ], + "angle": 0, + "content": "Training Details. We follow Imagen [40] and \\(\\mathrm{PIXART - \\alpha}\\) [4] to employ the T5 [10]'s encoder (i.e., Flan-T5-XXL) as the text encoder for conditional feature extraction, and use \\(\\mathrm{PIXART - \\alpha}\\) [4] as our base diffusion model. Unlike most works that extract fixed 77 text tokens, we adjust the length of text tokens from \\(\\mathrm{PIXART - \\alpha}\\)'s 120 to 300, as the caption curated in Internal-\\(\\Sigma\\) is much denser to provide highly fine-grained details. To capture the latent features of input images, we employ a pre-trained and frozen VAE from SDXL [36]. Other implementation details are the same as \\(\\mathrm{PIXART - \\alpha}\\). Models are finetuned on the \\(\\mathrm{PIXART - \\alpha}\\)'s 256px pre-trained checkpoint with the position embedding interpolation trick [4]. Our final models, including 1K resolution, are trained on 32 V100 GPUs. We additionally use 16 A100 GPUs to train the 2K and 4K image generation models. For further information, please refer to the appendix." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.636, + 0.788, + 0.697 + ], + "angle": 0, + "content": "Note that we use CAME optimizer [27] with a weight decay of 0 and a constant learning rate of 2e-5, instead of the regular AdamW [24] optimizer. This helps us reduce the dimension of the optimizer's state, leading to lower GPU memory without performance degradation." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.697, + 0.788, + 0.773 + ], + "angle": 0, + "content": "Evaluation Metrics. To better illustrate aesthetics and semantic ability, we collect 30K high-quality text-image pairs (as mentioned in Sec. 3.1) to benchmark T2I models. We mainly evaluate \\(\\mathsf{PIXART - \\Sigma}\\) via human and AI preference study since FID [39] may not adequately reflect the generation quality. However, we still provide the FID results on the collected dataset in the appendix." + }, + { + "type": "title", + "bbox": [ + 0.216, + 0.79, + 0.48, + 0.805 + ], + "angle": 0, + "content": "4.2 Performance Comparisons" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.81, + 0.787, + 0.842 + ], + "angle": 0, + "content": "Image Quality Assessment. We qualitatively evaluated our methodology against both closed-source text-to-image products and open-source models. As" + } + ], + [ + { + "type": "page_number", + "bbox": [ + 0.218, + 0.116, + 0.236, + 0.127 + ], + "angle": 0, + "content": "12" + }, + { + "type": "header", + "bbox": [ + 0.272, + 0.115, + 0.364, + 0.127 + ], + "angle": 0, + "content": "J. Chen et al." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.147, + 0.788, + 0.254 + ], + "angle": 0, + "content": "illustrated in Fig. 1, our model can produce high-quality, photo-realistic images with intricate details over diverse aspect ratios and styles. This capability underscores the superior performance of our approach in generating visually compelling content from textual descriptions. In Fig. 3, we compare \\(\\mathsf{PIXART - \\Sigma}\\) with open-source models SDXL [36] and \\(\\mathsf{PIXART - \\alpha}\\) [4], our method enhances the realism of portraits and boosts the capacity for semantic analysis. In contrast to SDXL, our approach demonstrates superior proficiency in following instructions." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.256, + 0.788, + 0.318 + ], + "angle": 0, + "content": "Not only superior to open-source models, but our method is also very competitive with current T2I closed-source products, as depicted in Fig. 4. PixART- \\(\\Sigma\\) produces photo-realistic results and adheres closely to user instructions, which is on par with contemporary commercial products." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.32, + 0.432, + 0.501 + ], + "angle": 0, + "content": "High-resolution Generation. Our method is capable of directly generating 4K resolution images without the need for any post-processing. Additionally, it excels in accurately following complex, detailed, and long text provided by users, as demonstrated in Fig. 2. Thus, users do not need prompt engineering to achieve satisfactory results." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.505, + 0.432, + 0.55 + ], + "angle": 0, + "content": "Our approach enables direct 4K image generation. In parallel, studies [11, 16] have" + }, + { + "type": "image", + "bbox": [ + 0.476, + 0.327, + 0.75, + 0.476 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.441, + 0.478, + 0.788, + 0.547 + ], + "angle": 0, + "content": "Fig. 9: Comparisons on FID and Clip-Score with Open-sourced T2I Models. PixArt- \\(\\Sigma\\) demonstrates enhanced performance in terms of FID and Clip-Score on the curated 30K High-Quality Evaluation Dataset." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.55, + 0.788, + 0.672 + ], + "angle": 0, + "content": "introduced tuning-free post-processing techniques aimed at generating HR images from LR models or employing super-resolution models [50] to produce HR images. However, their corresponding results often exhibit artifacts for two primary reasons: (1) Accumulative error may arise due to the cascade pipeline. (2) These methods do not capture the true distribution of 4K images nor learn the alignment between text and 4K images. We argue that our method might be a more promising way to generate high-resolution images. Our method yields superior results, and more visual comparison is included in the supplement." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.673, + 0.789, + 0.841 + ], + "angle": 0, + "content": "Human/AI (GPT4V) Preference Study. We evaluate the well-trained model in both the human and AI preference study using a subset of 300 captions randomly collected from the High-Quality Evaluation Dataset mentioned in Sec. 3.1. We collect images generated by overall six open-source models, including PixART-\\(\\alpha\\), PixART-\\(\\Sigma\\), SD1.5 [39], Stable Turbo [41], Stable XL [36], Stable Cascade [35] and Playground-V2.0 [20]. We develop a website for the human preference study to display the prompts and their corresponding images. This website was distributed to trained evaluators, who were asked to assess the images, ranking them according to quality and how well they matched the text prompts. The results, illustrated by the blue bar in Fig. 9, indicate a marked preference for PixART-\\(\\Sigma\\) over the other six T2I generators. PixART-\\(\\Sigma\\) genera" + } + ], + [ + { + "type": "header", + "bbox": [ + 0.282, + 0.114, + 0.732, + 0.129 + ], + "angle": 0, + "content": "PIXART- \\(\\Sigma\\) : Weak-to-Strong Training of DiT for 4K T2I Generation" + }, + { + "type": "page_number", + "bbox": [ + 0.769, + 0.116, + 0.785, + 0.127 + ], + "angle": 0, + "content": "13" + }, + { + "type": "table", + "bbox": [ + 0.282, + 0.144, + 0.47, + 0.211 + ], + "angle": 0, + "content": "
LayersFID ↓CLIP-Score ↑
N/A8.2440.276
Shallow (1-14)9.2780.275
Middle (7-20)9.0630.276
Deep (14-27)8.5320.275
" + }, + { + "type": "table_caption", + "bbox": [ + 0.306, + 0.213, + 0.444, + 0.224 + ], + "angle": 0, + "content": "(a) Compression layers." + }, + { + "type": "table", + "bbox": [ + 0.249, + 0.236, + 0.49, + 0.334 + ], + "angle": 0, + "content": "
Res.RatioFID ↓CLIP-Score ↑Train Latency ↓
51218.2440.2762.3
51229.0630.2762.2 (-4%)
51249.6060.2762.1 (-9%)
102415.6850.27727.5
102425.5120.27322.5 (-18%)
102445.6440.27620.0 (-27%)
102495.7120.27517.8 (-35%)
" + }, + { + "type": "table_caption", + "bbox": [ + 0.232, + 0.335, + 0.506, + 0.347 + ], + "angle": 0, + "content": "(c) Compression rations on different resolutions." + }, + { + "type": "table", + "bbox": [ + 0.525, + 0.145, + 0.722, + 0.211 + ], + "angle": 0, + "content": "
OperatorFID ↓CLIP-Score ↑
N/A8.2440.276
Token Discarding8.9180.275
Token Pooling9.4150.275
Conv2×28.5050.274
" + }, + { + "type": "table_caption", + "bbox": [ + 0.542, + 0.212, + 0.703, + 0.223 + ], + "angle": 0, + "content": "(b) Compression operators." + }, + { + "type": "table", + "bbox": [ + 0.525, + 0.244, + 0.75, + 0.326 + ], + "angle": 0, + "content": "
Res.RatioTrain Latency (s/Iter@32BS)↓ Test Latency ↓ (s/Img)
2K15658
2K437 (-34%)38 (-34%)
4K119191
4K4125 (-35%)60 (-34%)
" + }, + { + "type": "table_caption", + "bbox": [ + 0.539, + 0.327, + 0.735, + 0.338 + ], + "angle": 0, + "content": "(d) Speed of different resolutions." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.351, + 0.785, + 0.407 + ], + "angle": 0, + "content": "Table 3: KV-Token Compression Settings in Image Generation. This study employs FID, CMMD, and CLIP-Score metrics to assess the impact of various token compression components, such as compression ratio, positions, operators, and varying resolutions. Speed calculation in Tab. 3c is Second/Iteration/384 Batch-size." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.417, + 0.785, + 0.463 + ], + "angle": 0, + "content": "ates superior high-quality images that closely follow user prompts, using a much smaller size (0.6B parameters) compared to existing T2I diffusion models like SDXL (2.6B parameters) and SD Cascade (5.1B parameters)." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.464, + 0.787, + 0.602 + ], + "angle": 0, + "content": "For AI preference study, we employ the advanced multimodal model, GPT-4 V [32], as the evaluator. For each trial, we supply GPT-4V with two images: one from \\(\\mathrm{PIXART - \\Sigma}\\) and another from a competing T2I model. We craft distinct prompts guiding GPT-4V to vote based on image quality and image-and-text alignment. The results, represented by orange and green bars in Fig. 9, demonstrate consistent outcomes in both human and AI preference studies. Specifically, \\(\\mathrm{PIXART - \\Sigma}\\) surpasses the baseline, \\(\\mathrm{PIXART - \\alpha}\\), in effectiveness. Compared to advanced models such as Stable Cascaded, \\(\\mathrm{PIXART - \\Sigma}\\) exhibits competitive or superior performance on image quality and instruction-following abilities." + }, + { + "type": "title", + "bbox": [ + 0.216, + 0.632, + 0.416, + 0.646 + ], + "angle": 0, + "content": "5 Ablation Studies" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.671, + 0.785, + 0.717 + ], + "angle": 0, + "content": "We conduct ablation studies on generation performance on various KV compression designs. Unless specified, the experiments are conducted on 512px generation. The detailed settings of each ablation trial are included in the appendix." + }, + { + "type": "title", + "bbox": [ + 0.216, + 0.747, + 0.444, + 0.763 + ], + "angle": 0, + "content": "5.1 Experimental settings" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.78, + 0.785, + 0.841 + ], + "angle": 0, + "content": "We use the test set described in Sec. 3.1 for evaluation. We employ FID to compute the distributional difference between the collected and generated data for comparative metrics. Furthermore, we utilize CLIP-Score to assess the alignment between prompts and the generated images." + } + ], + [ + { + "type": "page_number", + "bbox": [ + 0.218, + 0.116, + 0.236, + 0.127 + ], + "angle": 0, + "content": "14" + }, + { + "type": "header", + "bbox": [ + 0.272, + 0.115, + 0.364, + 0.127 + ], + "angle": 0, + "content": "J. Chen et al." + }, + { + "type": "title", + "bbox": [ + 0.216, + 0.147, + 0.439, + 0.163 + ], + "angle": 0, + "content": "5.2 Compression Designs" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.17, + 0.788, + 0.306 + ], + "angle": 0, + "content": "Compression positions. We implemented KV compression at different depths within the Transformer structure: in the shallow layers (1~14), the intermediate layers (7~20), and the deep layers (14~27). As indicated in Tab. 3a, employing KV compression on deep layers notably achieves superior performance. We speculate this is because shallow layers typically encode detailed texture content, while deep layers abstract high-level semantic content. Since compression affects image quality rather than semantic information, compressing deep layers achieves the least information loss, making it practical for accelerating training without compromising generation quality." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.306, + 0.788, + 0.395 + ], + "angle": 0, + "content": "Compression operators. We explored the impact of different compression operators. We employed three techniques, random discarding, average pooling, and parametric convolution, to compress \\(2 \\times 2\\) tokens into a single token. As illustrated in Table 3b, the \"Conv \\(2 \\times 2\\)\" method outperforms the others, underscoring the advantage of using a learnable kernel to more effectively reduce redundant features than simple discarding methods." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.397, + 0.788, + 0.517 + ], + "angle": 0, + "content": "Compression ratios on different resolutions. We investigated the influence of varying compression ratios on different resolutions. As shown in Tab. 3c, remarkably, we find that token compression does not affect the alignment between textual and generated images (CLIP Score) but influences the image quality (FID) across resolutions. Although there is a slight degradation in image quality with increasing compression ratios, our strategy brings a training speedup of \\(18\\%\\) to \\(35\\%\\). This suggests that our proposed KV compression is both effective and efficient for achieving high-resolution T2I generation." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.518, + 0.788, + 0.639 + ], + "angle": 0, + "content": "Speed comparisons on different resolutions. We further comprehensively validate the speed acceleration in both training and inference in Tab. 3d. By 'latency', we refer to the duration required for inference on a single sample. Our method speeds up training and inference by about \\(35\\%\\) in the 4K generation. Notably, training acceleration increases with resolution, accelerating from \\(18\\%\\) to \\(35\\%\\) as the resolution increases from 1K to 4K. This indicates the effectiveness of our method with increasing resolution, demonstrating its potential applicability to even higher-resolution image generation tasks." + }, + { + "type": "title", + "bbox": [ + 0.216, + 0.66, + 0.36, + 0.675 + ], + "angle": 0, + "content": "6 Conclusion" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.689, + 0.788, + 0.841 + ], + "angle": 0, + "content": "In this paper, we introduce \\(\\mathrm{PIXART - \\Sigma}\\), a Text-to-Image (T2I) diffusion model capable of directly generating high-quality images at 4K resolution. Building upon the pre-trained foundation of \\(\\mathrm{PIXART - \\alpha}\\), \\(\\mathrm{PIXART - \\Sigma}\\) achieves efficient training through a novel \"weak-to-strong training\" methodology. This approach is characterized by the incorporation of higher-quality data and the integration of efficient token compression. \\(\\mathrm{PIXART - \\Sigma}\\) excels at producing high-fidelity images while adhering closely to textual prompts, surpassing the high standards set by its predecessor, \\(\\mathrm{PIXART - \\alpha}\\). We believe that the innovations presented in \\(\\mathrm{PIXART - \\Sigma}\\) will not only contribute to advancements in the AIGC community but also pave the way for entities to access more efficient, and high-quality generative models." + } + ], + [ + { + "type": "header", + "bbox": [ + 0.282, + 0.115, + 0.732, + 0.129 + ], + "angle": 0, + "content": "PIXART- \\(\\boldsymbol{\\Sigma}\\) : Weak-to-Strong Training of DiT for 4K T2I Generation" + }, + { + "type": "page_number", + "bbox": [ + 0.769, + 0.117, + 0.785, + 0.127 + ], + "angle": 0, + "content": "15" + }, + { + "type": "text", + "bbox": [ + 0.216, + 0.147, + 0.796, + 0.178 + ], + "angle": 0, + "content": "Acknowledgements. We would like to thank Zeqiang Lai, Fei Chen, and Shuchen Xue for discussing and helping implement the DMD part." + }, + { + "type": "title", + "bbox": [ + 0.217, + 0.203, + 0.323, + 0.219 + ], + "angle": 0, + "content": "References" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.238, + 0.785, + 0.265 + ], + "angle": 0, + "content": "1. Aesthetic predictor (2023), https://github.com/christophschuhmann/improved-aesthetic-predictor7" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.266, + 0.785, + 0.293 + ], + "angle": 0, + "content": "2. Bao, F., Nie, S., Xue, K., Cao, Y., Li, C., Su, H., Zhu, J.: All are worth words: A vit backbone for diffusion models. In: CVPR (2023) 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.295, + 0.785, + 0.335 + ], + "angle": 0, + "content": "3. Chandra, A., Tünnermann, L., Lofstedt, T., Gratz, R.: Transformer-based deep learning for predicting protein properties in the life sciences. eLife Sciences Publications, Ltd (2023) 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.336, + 0.785, + 0.378 + ], + "angle": 0, + "content": "4. Chen, J., Yu, J., Ge, C., Yao, L., Xie, E., Wu, Y., Wang, Z., Kwok, J., Luo, P., Lu, H., Li, Z.: PixArt-α: Fast training of diffusion transformer for photorealistic text-to-image synthesis. In: ICLR (2024) 1, 4, 6, 8, 11, 12" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.379, + 0.785, + 0.419 + ], + "angle": 0, + "content": "5. Chen, L., Li, J., Dong, X., Zhang, P., He, C., Wang, J., Zhao, F., Lin, D.: ShareGPT4V: Improving large multi-modal models with better captions. In: arXiv (2023) 3, 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.42, + 0.785, + 0.462 + ], + "angle": 0, + "content": "6. Chen, S., Xu, M., Ren, J., Cong, Y., He, S., Xie, Y., Sinha, A., Luo, P., Xiang, T., Perez-Rua, J.M.: Gentron: Delving deep into diffusion transformers for image and video generation. arXiv preprint arXiv:2312.04557 (2023) 3, 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.463, + 0.785, + 0.49 + ], + "angle": 0, + "content": "7. Chen, T., Cheng, Y., Gan, Z., Yuan, L., Zhang, L., Wang, Z.: Chasing sparsity in vision transformers: An end-to-end exploration. In: NeurIPS (2021) 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.491, + 0.785, + 0.53 + ], + "angle": 0, + "content": "8. Chen, X., Liu, Z., Tang, H., Yi, L., Zhao, H., Han, S.: SparseViT: Revisiting activation sparsity for efficient high-resolution vision transformer. In: CVPR (2023) 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.532, + 0.785, + 0.573 + ], + "angle": 0, + "content": "9. Choromanski, K., Likhosherstov, V., Dohan, D., Song, X., Gane, A., Sarlos, T., Hawkins, P., Davis, J., Mohiuddin, A., Kaiser, L., et al.: Rethinking attention with performers. In: ICLR (2021) 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.574, + 0.785, + 0.615 + ], + "angle": 0, + "content": "0. Chung, H.W., Hou, L., Longpre, S., Zoph, B., Tay, Y., Fedus, W., Li, Y., Wang, X., Dehghani, M., Brahma, S., et al.: Scaling instruction-finetuned language models. In: arXiv (2022) 3, 11" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.616, + 0.785, + 0.644 + ], + "angle": 0, + "content": " 1. Du, R., Chang, D., Hospedales, T., Song, Y.Z., Ma, Z.: Demofusion: Democratising high-resolution image generation with no $$#. In: CVPR (2024) 12" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.645, + 0.785, + 0.672 + ], + "angle": 0, + "content": "2. Gao, S., Zhou, P., Cheng, M.M., Yan, S.: Masked diffusion transformer is a strong image synthesizer. In: ICCV (2023) 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.673, + 0.785, + 0.7 + ], + "angle": 0, + "content": "3. Ge, C., Ding, X., Tong, Z., Yuan, L., Wang, J., Song, Y., Luo, P.: Advancing vision transformers with group-mix attention. In: arXiv (2023) 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.701, + 0.785, + 0.728 + ], + "angle": 0, + "content": "4. von Glehn, I., Spencer, J.S., P fauna, D.: A self-attention ansatz for ab-initio quantum chemistry. In: ICLR (2023) 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.729, + 0.785, + 0.756 + ], + "angle": 0, + "content": "5. Hatamizadeh, A., Song, J., Liu, G., Kautz, J., Vahdat, A.: Diffit: Diffusion vision transformers for image generation. In: arXiv (2023) 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.757, + 0.785, + 0.798 + ], + "angle": 0, + "content": "6. He, Y., Yang, S., Chen, H., Cun, X., Xia, M., Zhang, Y., Wang, X., He, R., Chen, Q., Shan, Y.: ScaleCrafter: Tuning-free higher-resolution visual generation with diffusion models. In: ICLR (2024) 12" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.799, + 0.785, + 0.84 + ], + "angle": 0, + "content": "7. Heusel, M., Ramsauer, H., Unterthiner, T., Nessler, B., Hochreiter, S.: GANs trained by a two time-scale update rule converge to a local nash equilibrium. In: NeurIPS (2017) 8" + }, + { + "type": "list", + "bbox": [ + 0.226, + 0.238, + 0.785, + 0.84 + ], + "angle": 0, + "content": null + } + ], + [ + { + "type": "page_number", + "bbox": [ + 0.218, + 0.116, + 0.236, + 0.127 + ], + "angle": 0, + "content": "16" + }, + { + "type": "header", + "bbox": [ + 0.272, + 0.115, + 0.364, + 0.127 + ], + "angle": 0, + "content": "J. Chen et al." + }, + { + "type": "ref_text", + "bbox": [ + 0.218, + 0.147, + 0.785, + 0.175 + ], + "angle": 0, + "content": "18. Kang, M., Zhu, J.Y., Zhang, R., Park, J., Shechtman, E., Paris, S., Park, T.: Scaling up gans for text-to-image synthesis. In: CVPR (2023) 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.218, + 0.176, + 0.785, + 0.215 + ], + "angle": 0, + "content": "19. Li, D., Kamko, A., Akhgari, E., Sabet, A., Xu, L., Doshi, S.: Playground v2.5: Three insights towards enhancing aesthetic quality in text-to-image generation. In: arXiv (2024) 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.217, + 0.785, + 0.243 + ], + "angle": 0, + "content": "20. Li, D., Kamko, A., Sabet, A., Akhgari, E., Xu, L., Doshi, S.: Playground v2, https://huggingface.co/playgroundai/playground-v2-1024px-aesthetic 12" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.244, + 0.785, + 0.27 + ], + "angle": 0, + "content": "21. Lin, T.Y., Maire, M., Belongie, S., Hays, J., Perona, P., Ramanan, D., Dólar, P., Zitnick, C.L.: Microsoft COCO: Common objects in context. In: ECCV (2014) 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.271, + 0.772, + 0.284 + ], + "angle": 0, + "content": "22. Liu, H., Li, C., Wu, Q., Lee, Y.J.: Visual instruction tuning. In: arXiv (2023) 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.285, + 0.785, + 0.323 + ], + "angle": 0, + "content": "23. Liu, Z., Lin, Y., Cao, Y., Hu, H., Wei, Y., Zhang, Z., Lin, S., Guo, B.: Swin transformer: Hierarchical vision transformer using shifted windows. In: ICCV (2021) 4, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.325, + 0.785, + 0.35 + ], + "angle": 0, + "content": "24. Loshchilov, I., Hutter, F.: Decoupled weight decay regularization. In: arXiv (2017) 11" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.351, + 0.785, + 0.379 + ], + "angle": 0, + "content": "25. Lu, J., Yao, J., Zhang, J., Zhu, X., Xu, H., Gao, W., Xu, C., Xiang, T., Zhang, L.: Soft: Softmax-free transformer with linear complexity. In: NeurIPS (2021) 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.379, + 0.785, + 0.405 + ], + "angle": 0, + "content": "26. Lu, Z., Wang, Z., Huang, D., Wu, C., Liu, X., Ouyang, W., Bai, L.: Fit: Flexible vision transformer for diffusion model. In: arXiv (2024) 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.407, + 0.785, + 0.433 + ], + "angle": 0, + "content": "27. Luo, Y., Ren, X., Zheng, Z., Jiang, Z., Jiang, X., You, Y.: CAME: Confidence-guided adaptive memory efficient optimization. In: ACL (2023) 11" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.434, + 0.785, + 0.473 + ], + "angle": 0, + "content": "28. Ma, N., Goldstein, M., Albergo, M.S., Boffi, N.M., Vanden-Eijnden, E., Xie, S.: SiT: Exploring flow and diffusion-based generative models with scalable interpolant transformers. In: arXiv (2024) 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.474, + 0.696, + 0.487 + ], + "angle": 0, + "content": "29. Midjourney: Midjourney (2023), https://www.midjourney.com 1, 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.488, + 0.636, + 0.5 + ], + "angle": 0, + "content": "30. OpenAI: Dalle-2 (2023), https://openai.com/dall-e-2 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.501, + 0.669, + 0.514 + ], + "angle": 0, + "content": "31. OpenAI: Dalle-3 (2023), https://openai.com/dall-e-3 1, 4, 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.515, + 0.635, + 0.528 + ], + "angle": 0, + "content": "32. OpenAI: Gpt-4v(ison) system card. In: OpenAI (2023) 13" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.528, + 0.602, + 0.54 + ], + "angle": 0, + "content": "33. OpenAI: Sora (2024), https://openai.com/sora 3, 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.541, + 0.785, + 0.565 + ], + "angle": 0, + "content": "34. Peebles, W., Xie, S.: Scalable diffusion models with transformers. In: ICCV (2023) 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.567, + 0.785, + 0.607 + ], + "angle": 0, + "content": "35. Pernias, P., Rampas, D., Richter, M.L., Pal, C., Aubreville, M.: Würstchen: An efficient architecture for large-scale text-to-image diffusion models. In: ICLR (2023) 4, 6, 12" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.608, + 0.785, + 0.649 + ], + "angle": 0, + "content": "36. Podell, D., English, Z., Lacey, K., Blattmann, A., Dockhorn, T., Müller, J., Penna, J., Rombach, R.: SDXL: Improving latent diffusion models for high-resolution image synthesis. In: arXiv (2023) 1, 3, 4, 6, 11, 12" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.65, + 0.785, + 0.676 + ], + "angle": 0, + "content": "37. Radford, A., Narasimhan, K., Salimans, T., Sutskever, I., et al.: Improving language understanding by generative pre-training. OpenAI blog (2018) 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.677, + 0.785, + 0.703 + ], + "angle": 0, + "content": "38. Radford, A., Wu, J., Child, R., Luan, D., Amodei, D., Sutskever, I., et al.: Language models are unsupervised multitask learners. OpenAI blog (2019) 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.704, + 0.785, + 0.731 + ], + "angle": 0, + "content": "39. Rombach, R., Blattmann, A., Lorenz, D., Esser, P., Ommer, B.: High-resolution image synthesis with latent diffusion models. In: CVPR (2022) 6, 11, 12" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.732, + 0.785, + 0.784 + ], + "angle": 0, + "content": "40. Sahara, C., Chan, W., Saxena, S., Li, L., Whang, J., Denton, E.L., Ghasemipour, K., Gontijo Lopes, R., Karagol Ayan, B., Salimans, T., et al.: Photorealistic text-to-image diffusion models with deep language understanding. In: NeurIPS (2022) 6, 11" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.785, + 0.785, + 0.812 + ], + "angle": 0, + "content": "41. Sauer, A., Lorenz, D., Blattmann, A., Rombach, R.: Adversarial diffusion distillation. In: arXiv (2023) 12" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.813, + 0.785, + 0.838 + ], + "angle": 0, + "content": "42. Stability.AI: Stable diffusion 3 (2024), https://stability.ai/news/stable-diffusion-3 3" + }, + { + "type": "list", + "bbox": [ + 0.217, + 0.147, + 0.785, + 0.838 + ], + "angle": 0, + "content": null + } + ], + [ + { + "type": "header", + "bbox": [ + 0.283, + 0.115, + 0.732, + 0.129 + ], + "angle": 0, + "content": "PixART- \\(\\boldsymbol{\\Sigma}\\) : Weak-to-Strong Training of DiT for 4K T2I Generation" + }, + { + "type": "page_number", + "bbox": [ + 0.769, + 0.117, + 0.785, + 0.127 + ], + "angle": 0, + "content": "17" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.147, + 0.785, + 0.188 + ], + "angle": 0, + "content": "43. Touvron, H., Cord, M., Douze, M., Massa, F., Sablayrolles, A., Jégou, H.: Training data-efficient image transformers & distillation through attention. In: ICML (2021) 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.189, + 0.785, + 0.217 + ], + "angle": 0, + "content": "44. Wang, S., Li, B.Z., Khabsa, M., Fang, H., Ma, H.: Linformer: Self-attention with linear complexity. In: arXiv (2020) 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.218, + 0.785, + 0.258 + ], + "angle": 0, + "content": "45. Wang, W., Xie, E., Li, X., Fan, D.P., Song, K., Liang, D., Lu, T., Luo, P., Shao, L.: Pyramid vision transformer: A versatile backbone for dense prediction without convolutions. In: ICCV (2021) 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.259, + 0.785, + 0.3 + ], + "angle": 0, + "content": "46. Wang, W., Xie, E., Li, X., Fan, D.P., Song, K., Liang, D., Lu, T., Luo, P., Shao, L.: Pvt v2: Improved baselines with pyramid vision transformer. Computational Visual Media (2022) 7, 9" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.3, + 0.785, + 0.328 + ], + "angle": 0, + "content": "47. Xia, Z., Pan, X., Song, S., Li, L.E., Huang, G.: Vision transformer with deformable attention. In: CVPR (2022) 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.328, + 0.785, + 0.37 + ], + "angle": 0, + "content": "48. Xie, E., Wang, W., Yu, Z., Anandkumar, A., Alvarez, J.M., Luo, P.: SegFormer: Simple and efficient design for semantic segmentation with transformers. In: NeurIPS (2021) 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.37, + 0.785, + 0.411 + ], + "angle": 0, + "content": "49. Xie, E., Yao, L., Shi, H., Liu, Z., Zhou, D., Liu, Z., Li, J., Li, Z.: DiffFit: Unlocking transferability of large diffusion models via simple parameter-efficient fine-tuning. In: ICCV (2023) 11" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.411, + 0.785, + 0.439 + ], + "angle": 0, + "content": "50. Xue, Z., Song, G., Guo, Q., Liu, B., Zong, Z., Liu, Y., Luo, P.: Raphael: Text-to-image generation via large mixture of diffusion paths. In: NeurIPS (2023) 12" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.439, + 0.785, + 0.481 + ], + "angle": 0, + "content": "51. Yuan, L., Chen, Y., Wang, T., Yu, W., Shi, Y., Jiang, Z.H., Tay, F.E., Feng, J., Yan, S.: Tokens-to-token ViT: Training vision transformers from scratch onImagenet. In: ICCV (2021) 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.481, + 0.785, + 0.508 + ], + "angle": 0, + "content": "52. Zheng, H., Nie, W., Vahdat, A., Anandkumar, A.: Fast training of diffusion models with masked transformers. In: arXiv (2023) 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.508, + 0.785, + 0.55 + ], + "angle": 0, + "content": "53. Zheng, S., Lu, J., Zhao, H., Zhu, X., Luo, Z., Wang, Y., Fu, Y., Feng, J., Xiang, T., Torr, P.H., et al.: Rethinking semantic segmentation from a sequence-to-sequence perspective with transformers. In: CVPR (2021) 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.55, + 0.785, + 0.577 + ], + "angle": 0, + "content": "54. Zhu, X., Su, W., Lu, L., Li, B., Wang, X., Dai, J.: Deformable DETR: Deformable transformers for end-to-end object detection. ICLR (2021) 7" + }, + { + "type": "list", + "bbox": [ + 0.217, + 0.147, + 0.785, + 0.577 + ], + "angle": 0, + "content": null + } + ] +] \ No newline at end of file diff --git a/2024/PixArt-Sigma_ Weak-to-Strong Training of Diffusion Transformer for 4K Text-to-Image Generation/56a1c2d9-7461-45b0-bc11-505740d1999b_origin.pdf b/2024/PixArt-Sigma_ Weak-to-Strong Training of Diffusion Transformer for 4K Text-to-Image Generation/56a1c2d9-7461-45b0-bc11-505740d1999b_origin.pdf new file mode 100644 index 0000000000000000000000000000000000000000..e3be8304b349f4c6b6f0b1bfed4ea67a23dfcdae --- /dev/null +++ b/2024/PixArt-Sigma_ Weak-to-Strong Training of Diffusion Transformer for 4K Text-to-Image Generation/56a1c2d9-7461-45b0-bc11-505740d1999b_origin.pdf @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:500befa8db43b2470a28e0f1d8bc0b3d24f8b295402903b3e7bf257044c8a0ec +size 10517874 diff --git a/2024/PixArt-Sigma_ Weak-to-Strong Training of Diffusion Transformer for 4K Text-to-Image Generation/full.md b/2024/PixArt-Sigma_ Weak-to-Strong Training of Diffusion Transformer for 4K Text-to-Image Generation/full.md new file mode 100644 index 0000000000000000000000000000000000000000..c1c323ebae989db75799246262188f8a44c1fd6f --- /dev/null +++ b/2024/PixArt-Sigma_ Weak-to-Strong Training of Diffusion Transformer for 4K Text-to-Image Generation/full.md @@ -0,0 +1,297 @@ +# PIXART- $\boldsymbol{\Sigma}$ : Weak-to-Strong Training of Diffusion Transformer for 4K Text-to-Image Generation + +Junsong Chen $^{1,2,3*}$ , Chongjian Ge $^{1,3*}$ , Enze Xie $^{1*†}$ , Yue Wu $^{1*}$ , Lewei Yao $^{1,4}$ , Xiaozhe Ren $^{1}$ , Zhongdao Wang $^{1}$ , Ping Luo $^{3}$ , Huchuan Lu $^{2}$ , and Zhenguo Li $^{1}$ + +1 Huawei Noah's Ark Lab 2 Dalian University of Technology 3 HKU 4 HKUST Project Page: https://pixart-alpha.github.io/PixArt-sigma-project/ + +Abstract. In this paper, we introduce $\mathrm{PIXART - \Sigma}$ , a Diffusion Transformer model (DiT) capable of directly generating images at 4K resolution. $\mathrm{PIXART - \Sigma}$ represents a significant advancement over its predecessor, $\mathrm{PIXART - \alpha}$ , offering images of markedly higher fidelity and improved alignment with text prompts. A key feature of $\mathrm{PIXART - \Sigma}$ is its training efficiency. Leveraging the foundational pre-training of $\mathrm{PIXART - \alpha}$ , it evolves from the 'weaker' baseline to a 'stronger' model via incorporating higher quality data, a process we term "weak-to-strong training". The advancements in $\mathrm{PIXART - \Sigma}$ are twofold: (1) High-Quality Training Data: $\mathrm{PIXART - \Sigma}$ incorporates superior-quality image data, paired with more precise and detailed image captions. (2) Efficient Token Compression: we propose a novel attention module within the DiT framework that compresses both keys and values, significantly improving efficiency and facilitating ultra-high-resolution image generation. Thanks to these improvements, $\mathrm{PIXART - \Sigma}$ achieves superior image quality and user prompt adherence capabilities with significantly smaller model size (0.6B parameters) than existing text-to-image diffusion models, such as SDXL (2.6B parameters) and SD Cascade (5.1B parameters). Moreover, $\mathrm{PIXART - \Sigma}$ 's capability to generate 4K images supports the creation of high-resolution posters and wallpapers, efficiently bolstering the production of high-quality visual content in industries such as film and gaming. + +Keywords: T2I Synthesis, Diffusion Transformer, Efficient Model + +# 1 Introduction + +The recent emergence of high-quality Text-to-Image (T2I) models has profoundly impacted the AI Generated Content (AIGC) community. This includes both proprietary models such as DALL-E 3 [31], Midjourney [29], as well as open-source models like Stable Diffusion [36] and PixART- $\alpha$ [4]. Nonetheless, developing a top-tier T2I model involves considerable resources; for instance, training SD1.5 from scratch necessitates about 6000 A100 GPU days [36], posing a substantial barrier to individual researchers with limited resources and impeding innovation + +![](images/2eae751a45613fdced33b20a28013da2d82387456ecdc459652b02b3a308e744.jpg) +Fig. 1: Images generated by $\mathbf{P}\mathbf{I}\mathbf{x}\mathbf{A}\mathbf{R}\mathbf{T} - \boldsymbol{\Sigma}$ . The model can output photo-realistic, high aesthetic, extreme aspect ratio, multi-style images, and follow user instructions. + +within the AIGC community. Over time, the AIGC community will gain access to continuously updated, higher-quality datasets and more advanced algorithms. A pivotal question is: how can we efficiently integrate these new elements into an existing model, achieving a more powerful version within the constraints of limited resources? + +To explore this issue, our research focuses on enhancing $\mathrm{PIXART - \alpha}$ , an efficient T2I training method. $\mathrm{PIXART - \alpha}$ represents an early venture within the DiT framework, a model structure with significant potential, as evidenced by works such as GenTron [6], Sora [33] and Stable Diffusion 3 [42]. To maximize this potential, we build upon the pre-trained foundation of $\mathrm{PIXART - \alpha}$ , integrating advanced elements to facilitate its continuous improvement, resulting in a more powerful model, $\mathrm{PIXART - \Sigma}$ . We refer to this process of evolving from a relatively weaker baseline to a stronger model through efficient training as "weak-to-strong training". Specifically, to achieve "weak-to-strong training", we introduce the following enhancements: + +Higher-Quality Training Data: We collect a high-quality dataset superior to that used in $\mathrm{PIXART - \alpha}$ , focusing on two key aspects: (i) High-quality images: The dataset comprises 33M high-resolution images sourced from the Internet, all exceeding 1K resolution, including 2.3M images with resolutions around 4K. These images are predominantly characterized by their high aesthetic and encompass a wide range of artistic styles. (ii) Dense and accurate captions: To provide more precise and detailed captions for the aforementioned images, we replace the LLaVA [22] used in $\mathrm{PIXART - \alpha}$ with a more powerful image captioner, Share-Captioner [5]. Furthermore, to improve the model's alignment capacity between the textual and visual concepts, we extend the token length of the text encoder (i.e., Flan-T5 [10]) to approximately 300 words. We observe these improvements effectively eliminate the model's tendency for hallucination, leading to higher-quality text-image alignment. + +Efficient Token Compression: To enhance $\mathrm{PIXART - \alpha}$ , we expand its generation resolution from 1K to 4K. Generating images at ultra-high resolutions (e.g., 2K/4K) introduces a significant increase in the number of tokens, leading to a substantial rise in computational demand. To address this challenge, we introduced a self-attention module with key and value token compression tailored to the DiT framework. Specifically, we utilize group convolutions with a stride of 2 for local aggregation of keys and values. Additionally, we employ a specialized weight initialization scheme, allowing for a smooth adaptation from a pre-trained model without KV compression. This design effectively reduces training and inference time by $\sim 34\%$ for high-resolution image generation. + +Weak-to-Strong Training Strategy: we propose several fine-tuning techniques to rapidly adapt from a weak model to a strong model efficiently. That includes (1) replacing with a more powerful Variational Autoencoder (VAE) [36], (2) scaling from low to high resolution, and (3) evolving from a model without Key-Value (KV) compression to one with KV compression. These outcomes confirm the validity and effectiveness of the "weak-to-strong training" approach. + +![](images/7e35c89910d2f12dd3e67db0e4bd640a959fe122d35481f09a9a261e1d50789e.jpg) +Fig. 2: 4K image generation with complex dense instructions. $\mathrm{PIXART - \Sigma}$ can directly generate 4K resolution images without post-processing, and accurately respond to the given prompt. + +Through the proposed improvements, $\mathrm{PIXART - \Sigma}$ achieves high-quality 4K resolution image generation at a minimal training cost and model parameters. Specifically, fine-tuning from a pre-trained model, we additionally utilize only $9\%$ of the GPU days required by $\mathrm{PIXART - \alpha}$ to achieve a strong 1K high-resolution image generation model, which is impressive considering replacing with new training data and a more powerful VAE. Moreover, we use only 0.6B parameters while SDXL [36] and SD Cascade [35] use 2.6B and 5.1B parameters respectively. Images generated by $\mathrm{PIXART - \Sigma}$ possess an aesthetic quality comparable to current top-tier T2I products, such as DALL-E 3 [31] and MJV6 [29] (as in Fig. 4). Additionally, $\mathrm{PIXART - \Sigma}$ also demonstrates exceptional capability for fine-grained alignment with textual prompts (as shown in Fig. 2 and 3). + +# 2 Related Work + +Diffusion Transformers. The Transformer architecture has achieved remarkable success across various domains, such as language modeling [37,38], computer vision [23, 43, 51, 53], and other areas [3, 14]. In the realm of diffusion models, DiT [34] and UViT [2] pioneer the use of Transformer architecture. Subsequent works, including DiffiT [15], SiT [28], and FiT [26], have improved upon DiT's architecture, while [12, 52] enhance training efficiency through masked modeling techniques. For Text-to-Image (T2I) synthesis, PixART- $\alpha$ [4] explore efficient T2I training schemes, achieving the first Transformer-based T2I model capable of generating 1024px high-quality images. GenTron [6] explores the flexibility and scalability of diffusion Transformers in both image and video generation. The re + +![](images/fb9c7c710fe0e9d2ee06c3f7f4d3cf783b86318f0d6b42b108580da8e3edc19c.jpg) +Prompt: A close-up photo of a person. The subject is a woman. She wore a blue coat with a gray dress underneath. She has blue eyes and blond hair, and wears a pair of earrings. Behind are blurred city buildings and streets. + +![](images/12e79cda23078bf0271910267e946835d7e8f827a93e81cc47ed7c78fbbc42e9.jpg) + +![](images/e41f0f585330f34bde73a7d6b4aacb5c0c71eccf8581ed83c3162caa14c4a68d.jpg) + +![](images/ea7a0a13f17110ee9ad20ffc30c8e4306c16b7efceab99065c96d8c050e7ad46.jpg) +Prompt: half a solid black background and half a solid white background + +![](images/2a9f33ba284496177d5d438d058378e702d205d9b8133067ed20332253946dd5.jpg) + +![](images/aa2f55b578413185cd7fffa3afbcab97e71b7f0f458f495054ee6eb8905ee330.jpg) + +Fig. 3: Comparison of $\mathbf{P}\mathbf{i}\mathbf{x}\mathbf{A}\mathbf{r}\mathbf{T} - \boldsymbol{\Sigma}$ with open-source models, e.g., $\mathbf{P}\mathbf{i}\mathbf{x}\mathbf{A}\mathbf{r}\mathbf{T} - \alpha$ and SDXL: Compared with $\mathbf{P}\mathbf{i}\mathbf{x}\mathbf{A}\mathbf{r}\mathbf{T} - \alpha$ , $\mathbf{P}\mathbf{i}\mathbf{x}\mathbf{A}\mathbf{r}\mathbf{T} - \boldsymbol{\Sigma}$ improves the realism of portraits and the capability of semantic analysis. Compared with SDXL, our method has a better ability to follow user instructions. The keywords are highlighted as blue. +![](images/39caa60e2a5cfb34fb011d38b26399a54a7ce7c4d76b21e0709d0b22a2a4aea0.jpg) +Prompt: Pixel art style of a snowboarder in mid-air performs a trick on a black rail, wearing a blue sweatshirt and black pants, with arms outstretched. The serene snowy landscape background, dotted with trees, complements the scene. The low-angle perspective emphasizes the trick's height and skill. + +![](images/d55309bb589a3373b0becb9b973a4e2557629f448d5ee2d29acb693587b4d085.jpg) + +![](images/c526f9445e51f205cd246014870e00ff4bdc8233f908d7f120ef10d62e7c5143.jpg) + +cent advent of the powerful video generation model Sora [33] has further underscored the potential of Diffusion Transformers. In this work, for the first time, we explore using the Transformer architecture to generate 4K ultra-high-resolution images directly, tackling the computational complexity challenges posed by involving long-sequence tokens. + +High Resolution Image Generation greatly enhances visual quality and is important in various industries such as film and gaming. However, increasing + +![](images/14c5470177a9e12fbbd690e435925b8cf2b8786a6527e154d55c523f3dce123d.jpg) + +![](images/2d57c3884bd71131bb8388954711c28bdd9a712c0033176cb3944a4d2f656fef.jpg) + +![](images/d556f7bae82ad9665e6187894d569497c6b51ec9c6a93ba90053c62cd9c38460.jpg) + +Prompt: a small cactus with a happy face in the Sahara desert + +![](images/d180f59fd0eb73be17a0325501067e112bcbcab0ac6b83eb0a16027ac91e92e6.jpg) +Fig. 4: Compare $\mathbf{P}\mathbf{i}\mathbf{x}\mathbf{A}\mathbf{r}\mathbf{T} - \boldsymbol{\Sigma}$ and four other T2I products: Firefly 2, Imagen 2, Dalle 3, and Midjourney 6. Images generated by $\mathbf{P}\mathbf{i}\mathbf{x}\mathbf{A}\mathbf{r}\mathbf{T} - \boldsymbol{\Sigma}$ are very competitive with these commercial products. + +![](images/8fa7178d9227a3b60727c4ce8ce09e21c5a52f26a58a40283722b2a94a6a2561.jpg) + +![](images/6de248d743f167195a1999991156d028247a18fa3f01e7cf191918951ed62eac.jpg) + +Prompt: a cherry pick scientist + +image resolution introduces challenges due to the substantial increase in computational demands. Numerous methods have been explored in this direction. For instance, Imagen [40], GigaGAN [18] and Stable Diffusion [39] introduce an additional super-resolution network, while Stable Cascade [35] employs multiple diffusion networks to increase resolution progressively. These combined-model solutions, however, can introduce cumulative errors. On the other hand, works like SDXL [36], DALL-E 2 [30], Playground [19] and $\mathrm{PIXART - \alpha}$ [4] have attempted to generate high-resolution images using diffusion models directly. Nevertheless, these efforts are capped at generating images with resolutions up to $1024\mathrm{px}$ due to computational complexity. In this paper, we push this boundary to $4\mathrm{K}$ resolution, significantly enhancing the visual quality of the generated content. + +Efficient Transformer architecture. The self-attention mechanism in Transformer suffers from quadratic computational complexity with respect to the number of tokens, which hinders the scaling up of token quantity. Many works have sought improvements in this area: (1) Sparse Attention [7, 8, 45, 46, 48], which reduces the overall computational load by selectively processing a subset of tokens. For instance, PVT v2 [46] employs a convolutional kernel to condense the space of the key and value, thus lowering the complexity involved in computing the attention. (2) Local Attention [13, 23, 47, 54] focuses attention within nearby regions; notably, Swin Transformer [23] utilizes window-based attention to limit computations to within a specified window size. (3) Low-rank/Linear Attention [9, 25, 44]. The Linformer [44] reduces the computational complexity of the self-attention mechanism through low-rank approximations. In this paper, inspired by PVT v2 [46], we employ a self-attention mechanism based on key/value compression to mitigate the high complexity of processing 4K images. + +# 3 Framework + +# 3.1 Data Analysis + +
LLaVA HallucinationsThe image features a large, ornate church with a tall, pointed roof and a large stained-glass window. The church is situated in a city, and it has a white and gray color scheme. The style of the church is Gothic, which is evident from the pointed roof and the intricate details of the stained-glass window. The presence of statues and a clock on the building further emphasizes the grand and historical nature of the structure.
Share-Captioner CorrectnessThe image captures the grandeur of a cathedral, painted in a pristine white, stands majestically against the backdrop of a clear blue sky. The entrance to the cathedral is adorned with a large rose window. Flanking the entrance are statues of saints, standing as silent guardians of the sacred space within. The perspective of the photo, taken from a low angle. The lush green trees in the background adding serenity to the scene.
LLaVA HallucinationsThe image features a woman and a man sitting on a brick walkway near a body of water, which could be a river or a lake. They are both wearing head coverings, and the woman is holding a handbag. The scene is set during the day, with the sun shining brightly, creating a warm and inviting atmosphere. The style of the image is a black and white photo, which adds a timeless and classic feel to the scene.
Share-Captioner CorrectnessThe image captures a serene scene at a harbor. Two individuals are seated on a bench, their backs to the camera, engrossed in the view of the water. The water, a deep shade of blue, is dotted with boats of various sizes and colors, including a white boat with a green stripe and a red boat. The sky above is a light blue.
+ +Fig. 5: Comparative illustration of hallucinations: Contrasting differences in hallucination occurrences between LLaVA and Share-Captioner, with red indicating hallucinations and green denoting correctness. + +Higher Aesthetic and higher Resolution. To enhance the aesthetic quality of our dataset, we expand our internal data from 14M to 33M. For clarity, we name the two datasets Internal- $\alpha$ and Internal- $\Sigma$ , respectively. Note that this expansion still falls short compared to the vast images utilized by currently available open-source models like SD v1.5, which uses 2B data. We demonstrate that effective training strategies with limited data amount can still obtain a strong T2I model. + +The images within Internal- $\Sigma$ are above 1K resolution. To facilitate 4K resolution generation, we additionally collect a dataset of 8M real photographic images at 4K resolution. To ensure aesthetic quality, we employ an aesthetic scoring model (AES) [1] to filter these 4K images. This process yields a highly refined dataset of 2M ultra-high-resolution and high-quality images. + +Interestingly, we have observed that as the resolution of the images increases, there is an improvement in the model's fidelity (Fréchet Inception Dis- + +Table 1: Statistics of noun concepts for different datasets. VN: valid distinct nouns (appearing more than 10 times); DN: total distinct nouns; Average: average noun count per image; ACL: Average Caption length. + +
DatasetVolumeCaptionVN/DNTotal NounACLAverage
Internal-α14MRaw187K/931K175M2511.7/Img
Internal-α14MLLaVA28K/215K536M9829.3/Img
Internal-α14MShare-Captioner51K/420K815M18454.4/Img
Internal-Σ33MRaw294K/1512K485M3514.4/Img
Internal-Σ33MShare-Captioner77K/714K1804M18053.6/Img
4K-Σ2.3MShare-Captioner24K/96K115M16349.5/Img
+ +tance (FID) [17]) and semantic alignment (CLIP Score), which underscores the importance of the capabilities of generating high-resolution images. + +# Better Text-Image Alignment. + +Recent works such as $\mathrm{PIXART - }\alpha$ [4] and DALL-E 3 [31] emphasize the significance of text-image alignment. Strengthening this alignment is crucial for boosting model capabilities. To refine our collected "raw" descriptions further, we focus on improving both the length and accuracy of our captions. Notably, our captions (Internal- $\Sigma$ ) show several advantages over the one used in $\mathrm{PIXART - }\alpha$ (Internal- $\alpha$ ) in the following aspects: 1. Enhanced caption accuracy: As depicted in Fig. 5, LLaVa used in $\mathrm{PIXART - }\alpha$ encounters certain hallucination. We leverage a more powerful model, i.e., Share-Captioner [5], to + +![](images/2779e528b6630f430c51a8bf54e45109d8b21c4d9a8462cb49a0b90232c2228d.jpg) +Fig.6: Histogram Visualization of the Caption Length. We randomly select 1M captions from the raw captions, Internal- $\alpha$ , and Internal- $\Sigma$ to draw the corresponding histogram. ACL denotes the average caption length. + +generate detailed and correct captions, augmenting the collected raw prompts. + +2. Increased caption length: As shown in Tab. 1 and Fig. 6, the average caption length increased significantly to 180 words, highly enhancing the descriptive power of the captions. Additionally, we extend the token processing length of the text encoder from 120 tokens (as in Internal- $\alpha$ ) to 300 tokens. Our model is trained on a mix of long (Share-Captioner) and short (raw) captions with a ratio of $60\%$ and $40\%$ , respectively. This approach enhances the diversity of textual descriptions and mitigates potential biases that might arise from solely relying on generative captions. + +Tab. 1 demonstrates a summary for both Internal- $\alpha$ and $-\Sigma$ , where we assess the diversity of the datasets through various metrics, including the noun variety, total noun count, average caption length, and average nouns per image. + +High-Quality Evaluation Dataset. Most SoTA T2I models chose MSCOCO [21] as the evaluation set to assess the FID and CLIP Scores. However, we observe + +evaluations conducted on the MSCOCO dataset may not adequately reflect a model's capabilities in aesthetics and text-image alignment. Therefore, we propose a curated set comprising 30,000 high-quality, aesthetically pleasing text-image pairs to facilitate the assessment. The selected samples of the dataset are presented in the appendix. This dataset is designed to provide a more comprehensive evaluation of a model's performance, particularly in capturing the intricacies of aesthetic appeal and the fidelity of alignment between textual descriptions and visual content. Unless otherwise specified, the evaluation experiments in the paper are conducted in the collected High-Quality Evaluation Dataset. + +# 3.2 Efficient DiT Design + +An efficient DiT network is essential since the computational demand significantly increases when generating images at ultra-high resolutions. The attention mechanism plays a pivotal role in the efficacy of Diffusion Transformers, yet its quadratic computational demands significantly limit model scalability, particularly at higher resolutions e.g., 2K and 4K. Inspired by PVT v2 [46], we incorporate KV compression within the original PixART- $\alpha$ s framework to address the computational challenges. This design adds a mere $0.018\%$ to the total parameters yet achieves efficient reduction in computational costs via token compression, while still preserving both spatial and semantic information. + +Key-Value (KV) Token Compression. Our motivation stems from the intriguing observation that applying key-value (KV) token compression directly to the pre-trained $\mathrm{PIXART - \alpha}$ can still generate reasonable images. This suggests a redundancy in the features. Considering the high similarity within adjacent $R\times R$ patches, we assume that feature semantics within a window are redundant and can be compressed reasonably. We propose KV token compression, which is denoted as $f_{c}(\cdot)$ , to compress token features within a $R\times R$ window through a compression operator, as depicted in Fig. 7. + +Furthermore, to mitigate the potential information loss caused by KV compression in self-attention computation, we opt to retain all the tokens of queries (Q). This strategy allows us to utilize KV compression effectively while mitigating the risk of losing crucial information. By employing KV compression, we enhance the efficiency of attention computations and reduce the computation complexity from $O(N^2)$ to $O\left(\frac{N^2}{R^2}\right)$ , thereby making the computational cost of directly generating high-resolution images manageable. + +![](images/fb5547b66eff4ce085940a61aa3d0b749e09e7cae296ff5305df0ec9bbb92298.jpg) +Fig. 7: Design of KV Token Compression. We merge KV tokens in spatial space to reduce the computation complexity. + +
StrategiesVAE Change512 →1024 RA512 →1024 RA + PE Interp.RA + PE Adjust + KV CompressRA + PE Interp. + KV Compress + Conv Avg Init
0 step
Fast +Adapt2K step100 step100 step100 step100 step
+ +Fig. 8: This illustration demonstrates how our training strategy designs expedite the model's convergence during the transition to VAEs, adjustment to higher resolutions, and the KV compression process, facilitating rapid learning from weak to strong. + +$$ +\operatorname {A t t e n t i o n} (Q, K, V) = \operatorname {s o f t m a x} \left(\frac {Q \cdot f _ {c} (K) ^ {T}}{\sqrt {d _ {k}}}\right) f _ {c} (V) \tag {1} +$$ + +We compress deep layers using the convolution operator "Conv2×2" with specific initialization. Detailed experiments on other design variants are discussed in Sec. 5. Specifically, we design a specialized convolution kernel initialization "Conv Avg Init" that utilizes group convolution and initializes the weights $w = \frac{1}{R^2}$ , equivalent to an average operator. This initialization strategy can initially produce coarse results, accelerating the fine-tuning process while only introducing 0.018% additional parameters. + +# 3.3 Weak-to-Strong Training Strategy + +We propose several efficient training strategies to enhance the transition from a "weak" model to a "strong" model. These strategies encompass VAE rapid adaptation, high-resolution fine-tuning, and KV Token compression. + +Adapting model to new VAEs. As VAEs develop, training T2I models from scratch is resource-intensive. We replace PixART- $\alpha$ s VAE with SDXL's VAE and continue fine-tuning the diffusion model. We observe a rapid convergence phenomenon that fine-tuning quickly converges at 2K training steps as shown in Fig 8 (a). Fine-tuning is more efficient when dealing with VAE + +Table 2: We fine-tune a high-resolution model from a low-resolution model and observe that even fine-tuning for a relatively short duration, such as 1K steps, can still yield high-quality results. + +(a) +(b) +(c) + +
ResolutionIterationsFID ↓CLIP ↑
25620K16.560.270
256 → 5121K9.750.272
256 → 512100K8.910.276
+ +model transferring and negates the necessity of training from scratch. + +Adapting to Higher-Resolution. When we fine-tune from a low-resolution (LR) model to a high-resolution (HR) model, we observe a performance degradation in Fig. 8 (b), which we attribute to discrepancies in positional embeddings + +(PE) between different resolutions. To address this, we utilize the "PE Interpolation" trick [4, 49]: initializing the HR model's PE by interpolating the LR model's PE, significantly enhancing the HR model's initial status and expediting fine-tuning. We can obtain visually pleasing images even within only 100 training iterations. Besides, we quantitatively compare the model's performance as illustrated in Tab. 2. The fine-tuning quickly converges at 1K steps, and further training slightly improves the performance. This illustrates that using the "PE Interpolation" trick enables rapid convergence of higher resolution generation, obviating the need for training from scratch for generating at higher resolutions. + +Adapting model to KV compression. We can use KV compression directly when fine-tuning from LR pre-trained models without KV compression. As shown in Fig. 8 (c), with our "Conv Avg Init." strategy, $\mathrm{PIXART - \Sigma}$ starts from a better initial state, making converging easier and faster. Notably, $\mathrm{PIXART - \Sigma}$ performs satisfied visual results even within 100 training steps. Finally, through the KV compression operators and compression layers design in Sec 3.2, we can reduce $\sim 34\%$ of the training and inference time. + +# 4 Experiment + +# 4.1 Implementation Details + +Training Details. We follow Imagen [40] and $\mathrm{PIXART - \alpha}$ [4] to employ the T5 [10]'s encoder (i.e., Flan-T5-XXL) as the text encoder for conditional feature extraction, and use $\mathrm{PIXART - \alpha}$ [4] as our base diffusion model. Unlike most works that extract fixed 77 text tokens, we adjust the length of text tokens from $\mathrm{PIXART - \alpha}$ 's 120 to 300, as the caption curated in Internal- $\Sigma$ is much denser to provide highly fine-grained details. To capture the latent features of input images, we employ a pre-trained and frozen VAE from SDXL [36]. Other implementation details are the same as $\mathrm{PIXART - \alpha}$ . Models are finetuned on the $\mathrm{PIXART - \alpha}$ 's 256px pre-trained checkpoint with the position embedding interpolation trick [4]. Our final models, including 1K resolution, are trained on 32 V100 GPUs. We additionally use 16 A100 GPUs to train the 2K and 4K image generation models. For further information, please refer to the appendix. + +Note that we use CAME optimizer [27] with a weight decay of 0 and a constant learning rate of 2e-5, instead of the regular AdamW [24] optimizer. This helps us reduce the dimension of the optimizer's state, leading to lower GPU memory without performance degradation. + +Evaluation Metrics. To better illustrate aesthetics and semantic ability, we collect 30K high-quality text-image pairs (as mentioned in Sec. 3.1) to benchmark T2I models. We mainly evaluate $\mathsf{PIXART - \Sigma}$ via human and AI preference study since FID [39] may not adequately reflect the generation quality. However, we still provide the FID results on the collected dataset in the appendix. + +# 4.2 Performance Comparisons + +Image Quality Assessment. We qualitatively evaluated our methodology against both closed-source text-to-image products and open-source models. As + +illustrated in Fig. 1, our model can produce high-quality, photo-realistic images with intricate details over diverse aspect ratios and styles. This capability underscores the superior performance of our approach in generating visually compelling content from textual descriptions. In Fig. 3, we compare $\mathsf{PIXART - \Sigma}$ with open-source models SDXL [36] and $\mathsf{PIXART - \alpha}$ [4], our method enhances the realism of portraits and boosts the capacity for semantic analysis. In contrast to SDXL, our approach demonstrates superior proficiency in following instructions. + +Not only superior to open-source models, but our method is also very competitive with current T2I closed-source products, as depicted in Fig. 4. PixART- $\Sigma$ produces photo-realistic results and adheres closely to user instructions, which is on par with contemporary commercial products. + +High-resolution Generation. Our method is capable of directly generating 4K resolution images without the need for any post-processing. Additionally, it excels in accurately following complex, detailed, and long text provided by users, as demonstrated in Fig. 2. Thus, users do not need prompt engineering to achieve satisfactory results. + +Our approach enables direct 4K image generation. In parallel, studies [11, 16] have + +![](images/8537c6c5f4e0fabc7b967cfdf7be066099cb65f6eb850b73531d752d75e72409.jpg) +Fig. 9: Comparisons on FID and Clip-Score with Open-sourced T2I Models. PixArt- $\Sigma$ demonstrates enhanced performance in terms of FID and Clip-Score on the curated 30K High-Quality Evaluation Dataset. + +introduced tuning-free post-processing techniques aimed at generating HR images from LR models or employing super-resolution models [50] to produce HR images. However, their corresponding results often exhibit artifacts for two primary reasons: (1) Accumulative error may arise due to the cascade pipeline. (2) These methods do not capture the true distribution of 4K images nor learn the alignment between text and 4K images. We argue that our method might be a more promising way to generate high-resolution images. Our method yields superior results, and more visual comparison is included in the supplement. + +Human/AI (GPT4V) Preference Study. We evaluate the well-trained model in both the human and AI preference study using a subset of 300 captions randomly collected from the High-Quality Evaluation Dataset mentioned in Sec. 3.1. We collect images generated by overall six open-source models, including PixART- $\alpha$ , PixART- $\Sigma$ , SD1.5 [39], Stable Turbo [41], Stable XL [36], Stable Cascade [35] and Playground-V2.0 [20]. We develop a website for the human preference study to display the prompts and their corresponding images. This website was distributed to trained evaluators, who were asked to assess the images, ranking them according to quality and how well they matched the text prompts. The results, illustrated by the blue bar in Fig. 9, indicate a marked preference for PixART- $\Sigma$ over the other six T2I generators. PixART- $\Sigma$ genera + +
LayersFID ↓CLIP-Score ↑
N/A8.2440.276
Shallow (1-14)9.2780.275
Middle (7-20)9.0630.276
Deep (14-27)8.5320.275
+ +(a) Compression layers. + +
Res.RatioFID ↓CLIP-Score ↑Train Latency ↓
51218.2440.2762.3
51229.0630.2762.2 (-4%)
51249.6060.2762.1 (-9%)
102415.6850.27727.5
102425.5120.27322.5 (-18%)
102445.6440.27620.0 (-27%)
102495.7120.27517.8 (-35%)
+ +(c) Compression rations on different resolutions. + +
OperatorFID ↓CLIP-Score ↑
N/A8.2440.276
Token Discarding8.9180.275
Token Pooling9.4150.275
Conv2×28.5050.274
+ +(b) Compression operators. + +
Res.RatioTrain Latency (s/Iter@32BS)↓ Test Latency ↓ (s/Img)
2K15658
2K437 (-34%)38 (-34%)
4K119191
4K4125 (-35%)60 (-34%)
+ +(d) Speed of different resolutions. + +Table 3: KV-Token Compression Settings in Image Generation. This study employs FID, CMMD, and CLIP-Score metrics to assess the impact of various token compression components, such as compression ratio, positions, operators, and varying resolutions. Speed calculation in Tab. 3c is Second/Iteration/384 Batch-size. + +ates superior high-quality images that closely follow user prompts, using a much smaller size (0.6B parameters) compared to existing T2I diffusion models like SDXL (2.6B parameters) and SD Cascade (5.1B parameters). + +For AI preference study, we employ the advanced multimodal model, GPT-4 V [32], as the evaluator. For each trial, we supply GPT-4V with two images: one from $\mathrm{PIXART - \Sigma}$ and another from a competing T2I model. We craft distinct prompts guiding GPT-4V to vote based on image quality and image-and-text alignment. The results, represented by orange and green bars in Fig. 9, demonstrate consistent outcomes in both human and AI preference studies. Specifically, $\mathrm{PIXART - \Sigma}$ surpasses the baseline, $\mathrm{PIXART - \alpha}$ , in effectiveness. Compared to advanced models such as Stable Cascaded, $\mathrm{PIXART - \Sigma}$ exhibits competitive or superior performance on image quality and instruction-following abilities. + +# 5 Ablation Studies + +We conduct ablation studies on generation performance on various KV compression designs. Unless specified, the experiments are conducted on 512px generation. The detailed settings of each ablation trial are included in the appendix. + +# 5.1 Experimental settings + +We use the test set described in Sec. 3.1 for evaluation. We employ FID to compute the distributional difference between the collected and generated data for comparative metrics. Furthermore, we utilize CLIP-Score to assess the alignment between prompts and the generated images. + +# 5.2 Compression Designs + +Compression positions. We implemented KV compression at different depths within the Transformer structure: in the shallow layers (1~14), the intermediate layers (7~20), and the deep layers (14~27). As indicated in Tab. 3a, employing KV compression on deep layers notably achieves superior performance. We speculate this is because shallow layers typically encode detailed texture content, while deep layers abstract high-level semantic content. Since compression affects image quality rather than semantic information, compressing deep layers achieves the least information loss, making it practical for accelerating training without compromising generation quality. + +Compression operators. We explored the impact of different compression operators. We employed three techniques, random discarding, average pooling, and parametric convolution, to compress $2 \times 2$ tokens into a single token. As illustrated in Table 3b, the "Conv $2 \times 2$ " method outperforms the others, underscoring the advantage of using a learnable kernel to more effectively reduce redundant features than simple discarding methods. + +Compression ratios on different resolutions. We investigated the influence of varying compression ratios on different resolutions. As shown in Tab. 3c, remarkably, we find that token compression does not affect the alignment between textual and generated images (CLIP Score) but influences the image quality (FID) across resolutions. Although there is a slight degradation in image quality with increasing compression ratios, our strategy brings a training speedup of $18\%$ to $35\%$ . This suggests that our proposed KV compression is both effective and efficient for achieving high-resolution T2I generation. + +Speed comparisons on different resolutions. We further comprehensively validate the speed acceleration in both training and inference in Tab. 3d. By 'latency', we refer to the duration required for inference on a single sample. Our method speeds up training and inference by about $35\%$ in the 4K generation. Notably, training acceleration increases with resolution, accelerating from $18\%$ to $35\%$ as the resolution increases from 1K to 4K. This indicates the effectiveness of our method with increasing resolution, demonstrating its potential applicability to even higher-resolution image generation tasks. + +# 6 Conclusion + +In this paper, we introduce $\mathrm{PIXART - \Sigma}$ , a Text-to-Image (T2I) diffusion model capable of directly generating high-quality images at 4K resolution. Building upon the pre-trained foundation of $\mathrm{PIXART - \alpha}$ , $\mathrm{PIXART - \Sigma}$ achieves efficient training through a novel "weak-to-strong training" methodology. This approach is characterized by the incorporation of higher-quality data and the integration of efficient token compression. $\mathrm{PIXART - \Sigma}$ excels at producing high-fidelity images while adhering closely to textual prompts, surpassing the high standards set by its predecessor, $\mathrm{PIXART - \alpha}$ . We believe that the innovations presented in $\mathrm{PIXART - \Sigma}$ will not only contribute to advancements in the AIGC community but also pave the way for entities to access more efficient, and high-quality generative models. + +Acknowledgements. We would like to thank Zeqiang Lai, Fei Chen, and Shuchen Xue for discussing and helping implement the DMD part. + +# References + +1. Aesthetic predictor (2023), https://github.com/christophschuhmann/improved-aesthetic-predictor7 +2. Bao, F., Nie, S., Xue, K., Cao, Y., Li, C., Su, H., Zhu, J.: All are worth words: A vit backbone for diffusion models. In: CVPR (2023) 4 +3. Chandra, A., Tünnermann, L., Lofstedt, T., Gratz, R.: Transformer-based deep learning for predicting protein properties in the life sciences. eLife Sciences Publications, Ltd (2023) 4 +4. Chen, J., Yu, J., Ge, C., Yao, L., Xie, E., Wu, Y., Wang, Z., Kwok, J., Luo, P., Lu, H., Li, Z.: PixArt-α: Fast training of diffusion transformer for photorealistic text-to-image synthesis. In: ICLR (2024) 1, 4, 6, 8, 11, 12 +5. Chen, L., Li, J., Dong, X., Zhang, P., He, C., Wang, J., Zhao, F., Lin, D.: ShareGPT4V: Improving large multi-modal models with better captions. In: arXiv (2023) 3, 8 +6. Chen, S., Xu, M., Ren, J., Cong, Y., He, S., Xie, Y., Sinha, A., Luo, P., Xiang, T., Perez-Rua, J.M.: Gentron: Delving deep into diffusion transformers for image and video generation. arXiv preprint arXiv:2312.04557 (2023) 3, 4 +7. Chen, T., Cheng, Y., Gan, Z., Yuan, L., Zhang, L., Wang, Z.: Chasing sparsity in vision transformers: An end-to-end exploration. In: NeurIPS (2021) 7 +8. Chen, X., Liu, Z., Tang, H., Yi, L., Zhao, H., Han, S.: SparseViT: Revisiting activation sparsity for efficient high-resolution vision transformer. In: CVPR (2023) 7 +9. Choromanski, K., Likhosherstov, V., Dohan, D., Song, X., Gane, A., Sarlos, T., Hawkins, P., Davis, J., Mohiuddin, A., Kaiser, L., et al.: Rethinking attention with performers. In: ICLR (2021) 7 +0. Chung, H.W., Hou, L., Longpre, S., Zoph, B., Tay, Y., Fedus, W., Li, Y., Wang, X., Dehghani, M., Brahma, S., et al.: Scaling instruction-finetuned language models. In: arXiv (2022) 3, 11 + 1. Du, R., Chang, D., Hospedales, T., Song, Y.Z., Ma, Z.: Demofusion: Democratising high-resolution image generation with no $$#. In: CVPR (2024) 12 +2. Gao, S., Zhou, P., Cheng, M.M., Yan, S.: Masked diffusion transformer is a strong image synthesizer. In: ICCV (2023) 4 +3. Ge, C., Ding, X., Tong, Z., Yuan, L., Wang, J., Song, Y., Luo, P.: Advancing vision transformers with group-mix attention. In: arXiv (2023) 7 +4. von Glehn, I., Spencer, J.S., P fauna, D.: A self-attention ansatz for ab-initio quantum chemistry. In: ICLR (2023) 4 +5. Hatamizadeh, A., Song, J., Liu, G., Kautz, J., Vahdat, A.: Diffit: Diffusion vision transformers for image generation. In: arXiv (2023) 4 +6. He, Y., Yang, S., Chen, H., Cun, X., Xia, M., Zhang, Y., Wang, X., He, R., Chen, Q., Shan, Y.: ScaleCrafter: Tuning-free higher-resolution visual generation with diffusion models. In: ICLR (2024) 12 +7. Heusel, M., Ramsauer, H., Unterthiner, T., Nessler, B., Hochreiter, S.: GANs trained by a two time-scale update rule converge to a local nash equilibrium. In: NeurIPS (2017) 8 + +18. Kang, M., Zhu, J.Y., Zhang, R., Park, J., Shechtman, E., Paris, S., Park, T.: Scaling up gans for text-to-image synthesis. In: CVPR (2023) 6 +19. Li, D., Kamko, A., Akhgari, E., Sabet, A., Xu, L., Doshi, S.: Playground v2.5: Three insights towards enhancing aesthetic quality in text-to-image generation. In: arXiv (2024) 6 +20. Li, D., Kamko, A., Sabet, A., Akhgari, E., Xu, L., Doshi, S.: Playground v2, https://huggingface.co/playgroundai/playground-v2-1024px-aesthetic 12 +21. Lin, T.Y., Maire, M., Belongie, S., Hays, J., Perona, P., Ramanan, D., Dólar, P., Zitnick, C.L.: Microsoft COCO: Common objects in context. In: ECCV (2014) 8 +22. Liu, H., Li, C., Wu, Q., Lee, Y.J.: Visual instruction tuning. In: arXiv (2023) 3 +23. Liu, Z., Lin, Y., Cao, Y., Hu, H., Wei, Y., Zhang, Z., Lin, S., Guo, B.: Swin transformer: Hierarchical vision transformer using shifted windows. In: ICCV (2021) 4, 7 +24. Loshchilov, I., Hutter, F.: Decoupled weight decay regularization. In: arXiv (2017) 11 +25. Lu, J., Yao, J., Zhang, J., Zhu, X., Xu, H., Gao, W., Xu, C., Xiang, T., Zhang, L.: Soft: Softmax-free transformer with linear complexity. In: NeurIPS (2021) 7 +26. Lu, Z., Wang, Z., Huang, D., Wu, C., Liu, X., Ouyang, W., Bai, L.: Fit: Flexible vision transformer for diffusion model. In: arXiv (2024) 4 +27. Luo, Y., Ren, X., Zheng, Z., Jiang, Z., Jiang, X., You, Y.: CAME: Confidence-guided adaptive memory efficient optimization. In: ACL (2023) 11 +28. Ma, N., Goldstein, M., Albergo, M.S., Boffi, N.M., Vanden-Eijnden, E., Xie, S.: SiT: Exploring flow and diffusion-based generative models with scalable interpolant transformers. In: arXiv (2024) 4 +29. Midjourney: Midjourney (2023), https://www.midjourney.com 1, 4 +30. OpenAI: Dalle-2 (2023), https://openai.com/dall-e-2 6 +31. OpenAI: Dalle-3 (2023), https://openai.com/dall-e-3 1, 4, 8 +32. OpenAI: Gpt-4v(ison) system card. In: OpenAI (2023) 13 +33. OpenAI: Sora (2024), https://openai.com/sora 3, 5 +34. Peebles, W., Xie, S.: Scalable diffusion models with transformers. In: ICCV (2023) 4 +35. Pernias, P., Rampas, D., Richter, M.L., Pal, C., Aubreville, M.: Würstchen: An efficient architecture for large-scale text-to-image diffusion models. In: ICLR (2023) 4, 6, 12 +36. Podell, D., English, Z., Lacey, K., Blattmann, A., Dockhorn, T., Müller, J., Penna, J., Rombach, R.: SDXL: Improving latent diffusion models for high-resolution image synthesis. In: arXiv (2023) 1, 3, 4, 6, 11, 12 +37. Radford, A., Narasimhan, K., Salimans, T., Sutskever, I., et al.: Improving language understanding by generative pre-training. OpenAI blog (2018) 4 +38. Radford, A., Wu, J., Child, R., Luan, D., Amodei, D., Sutskever, I., et al.: Language models are unsupervised multitask learners. OpenAI blog (2019) 4 +39. Rombach, R., Blattmann, A., Lorenz, D., Esser, P., Ommer, B.: High-resolution image synthesis with latent diffusion models. In: CVPR (2022) 6, 11, 12 +40. Sahara, C., Chan, W., Saxena, S., Li, L., Whang, J., Denton, E.L., Ghasemipour, K., Gontijo Lopes, R., Karagol Ayan, B., Salimans, T., et al.: Photorealistic text-to-image diffusion models with deep language understanding. In: NeurIPS (2022) 6, 11 +41. Sauer, A., Lorenz, D., Blattmann, A., Rombach, R.: Adversarial diffusion distillation. In: arXiv (2023) 12 +42. Stability.AI: Stable diffusion 3 (2024), https://stability.ai/news/stable-diffusion-3 3 + +43. Touvron, H., Cord, M., Douze, M., Massa, F., Sablayrolles, A., Jégou, H.: Training data-efficient image transformers & distillation through attention. In: ICML (2021) 4 +44. Wang, S., Li, B.Z., Khabsa, M., Fang, H., Ma, H.: Linformer: Self-attention with linear complexity. In: arXiv (2020) 7 +45. Wang, W., Xie, E., Li, X., Fan, D.P., Song, K., Liang, D., Lu, T., Luo, P., Shao, L.: Pyramid vision transformer: A versatile backbone for dense prediction without convolutions. In: ICCV (2021) 7 +46. Wang, W., Xie, E., Li, X., Fan, D.P., Song, K., Liang, D., Lu, T., Luo, P., Shao, L.: Pvt v2: Improved baselines with pyramid vision transformer. Computational Visual Media (2022) 7, 9 +47. Xia, Z., Pan, X., Song, S., Li, L.E., Huang, G.: Vision transformer with deformable attention. In: CVPR (2022) 7 +48. Xie, E., Wang, W., Yu, Z., Anandkumar, A., Alvarez, J.M., Luo, P.: SegFormer: Simple and efficient design for semantic segmentation with transformers. In: NeurIPS (2021) 7 +49. Xie, E., Yao, L., Shi, H., Liu, Z., Zhou, D., Liu, Z., Li, J., Li, Z.: DiffFit: Unlocking transferability of large diffusion models via simple parameter-efficient fine-tuning. In: ICCV (2023) 11 +50. Xue, Z., Song, G., Guo, Q., Liu, B., Zong, Z., Liu, Y., Luo, P.: Raphael: Text-to-image generation via large mixture of diffusion paths. In: NeurIPS (2023) 12 +51. Yuan, L., Chen, Y., Wang, T., Yu, W., Shi, Y., Jiang, Z.H., Tay, F.E., Feng, J., Yan, S.: Tokens-to-token ViT: Training vision transformers from scratch onImagenet. In: ICCV (2021) 4 +52. Zheng, H., Nie, W., Vahdat, A., Anandkumar, A.: Fast training of diffusion models with masked transformers. In: arXiv (2023) 4 +53. Zheng, S., Lu, J., Zhao, H., Zhu, X., Luo, Z., Wang, Y., Fu, Y., Feng, J., Xiang, T., Torr, P.H., et al.: Rethinking semantic segmentation from a sequence-to-sequence perspective with transformers. In: CVPR (2021) 4 +54. Zhu, X., Su, W., Lu, L., Li, B., Wang, X., Dai, J.: Deformable DETR: Deformable transformers for end-to-end object detection. ICLR (2021) 7 \ No newline at end of file diff --git a/2024/PixArt-Sigma_ Weak-to-Strong Training of Diffusion Transformer for 4K Text-to-Image Generation/images.zip b/2024/PixArt-Sigma_ Weak-to-Strong Training of Diffusion Transformer for 4K Text-to-Image Generation/images.zip new file mode 100644 index 0000000000000000000000000000000000000000..63f8c6746ef902072e73de6b7138989668cd7cae --- /dev/null +++ b/2024/PixArt-Sigma_ Weak-to-Strong Training of Diffusion Transformer for 4K Text-to-Image Generation/images.zip @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:8532542c0fbbf5e544493abcbdc4fffa14a2c86e22cb620125c7677b6827a098 +size 950271 diff --git a/2024/PixArt-Sigma_ Weak-to-Strong Training of Diffusion Transformer for 4K Text-to-Image Generation/layout.json b/2024/PixArt-Sigma_ Weak-to-Strong Training of Diffusion Transformer for 4K Text-to-Image Generation/layout.json new file mode 100644 index 0000000000000000000000000000000000000000..bdcec661ae004d7ad15d6f1a8a992b89d25a3a8b --- /dev/null +++ b/2024/PixArt-Sigma_ Weak-to-Strong Training of Diffusion Transformer for 4K Text-to-Image Generation/layout.json @@ -0,0 +1,9846 @@ +{ + "pdf_info": [ + { + "para_blocks": [ + { + "bbox": [ + 136, + 111, + 478, + 148 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 136, + 111, + 478, + 148 + ], + "spans": [ + { + "bbox": [ + 136, + 111, + 478, + 148 + ], + "type": "text", + "content": "PIXART- " + }, + { + "bbox": [ + 136, + 111, + 478, + 148 + ], + "type": "inline_equation", + "content": "\\boldsymbol{\\Sigma}" + }, + { + "bbox": [ + 136, + 111, + 478, + 148 + ], + "type": "text", + "content": ": Weak-to-Strong Training of Diffusion Transformer for 4K Text-to-Image Generation" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 136, + 167, + 477, + 194 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 136, + 167, + 477, + 194 + ], + "spans": [ + { + "bbox": [ + 136, + 167, + 477, + 194 + ], + "type": "text", + "content": "Junsong Chen" + }, + { + "bbox": [ + 136, + 167, + 477, + 194 + ], + "type": "inline_equation", + "content": "^{1,2,3*}" + }, + { + "bbox": [ + 136, + 167, + 477, + 194 + ], + "type": "text", + "content": ", Chongjian Ge" + }, + { + "bbox": [ + 136, + 167, + 477, + 194 + ], + "type": "inline_equation", + "content": "^{1,3*}" + }, + { + "bbox": [ + 136, + 167, + 477, + 194 + ], + "type": "text", + "content": ", Enze Xie" + }, + { + "bbox": [ + 136, + 167, + 477, + 194 + ], + "type": "inline_equation", + "content": "^{1*†}" + }, + { + "bbox": [ + 136, + 167, + 477, + 194 + ], + "type": "text", + "content": ", Yue Wu" + }, + { + "bbox": [ + 136, + 167, + 477, + 194 + ], + "type": "inline_equation", + "content": "^{1*}" + }, + { + "bbox": [ + 136, + 167, + 477, + 194 + ], + "type": "text", + "content": ", Lewei Yao" + }, + { + "bbox": [ + 136, + 167, + 477, + 194 + ], + "type": "inline_equation", + "content": "^{1,4}" + }, + { + "bbox": [ + 136, + 167, + 477, + 194 + ], + "type": "text", + "content": ", Xiaozhe Ren" + }, + { + "bbox": [ + 136, + 167, + 477, + 194 + ], + "type": "inline_equation", + "content": "^{1}" + }, + { + "bbox": [ + 136, + 167, + 477, + 194 + ], + "type": "text", + "content": ", Zhongdao Wang" + }, + { + "bbox": [ + 136, + 167, + 477, + 194 + ], + "type": "inline_equation", + "content": "^{1}" + }, + { + "bbox": [ + 136, + 167, + 477, + 194 + ], + "type": "text", + "content": ", Ping Luo" + }, + { + "bbox": [ + 136, + 167, + 477, + 194 + ], + "type": "inline_equation", + "content": "^{3}" + }, + { + "bbox": [ + 136, + 167, + 477, + 194 + ], + "type": "text", + "content": ", Huchuan Lu" + }, + { + "bbox": [ + 136, + 167, + 477, + 194 + ], + "type": "inline_equation", + "content": "^{2}" + }, + { + "bbox": [ + 136, + 167, + 477, + 194 + ], + "type": "text", + "content": ", and Zhenguo Li" + }, + { + "bbox": [ + 136, + 167, + 477, + 194 + ], + "type": "inline_equation", + "content": "^{1}" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 136, + 201, + 477, + 225 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 136, + 201, + 477, + 225 + ], + "spans": [ + { + "bbox": [ + 136, + 201, + 477, + 225 + ], + "type": "text", + "content": "1 Huawei Noah's Ark Lab 2 Dalian University of Technology 3 HKU 4 HKUST Project Page: https://pixart-alpha.github.io/PixArt-sigma-project/" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 159, + 255, + 455, + 486 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 159, + 255, + 455, + 486 + ], + "spans": [ + { + "bbox": [ + 159, + 255, + 455, + 486 + ], + "type": "text", + "content": "Abstract. In this paper, we introduce " + }, + { + "bbox": [ + 159, + 255, + 455, + 486 + ], + "type": "inline_equation", + "content": "\\mathrm{PIXART - \\Sigma}" + }, + { + "bbox": [ + 159, + 255, + 455, + 486 + ], + "type": "text", + "content": ", a Diffusion Transformer model (DiT) capable of directly generating images at 4K resolution. " + }, + { + "bbox": [ + 159, + 255, + 455, + 486 + ], + "type": "inline_equation", + "content": "\\mathrm{PIXART - \\Sigma}" + }, + { + "bbox": [ + 159, + 255, + 455, + 486 + ], + "type": "text", + "content": " represents a significant advancement over its predecessor, " + }, + { + "bbox": [ + 159, + 255, + 455, + 486 + ], + "type": "inline_equation", + "content": "\\mathrm{PIXART - \\alpha}" + }, + { + "bbox": [ + 159, + 255, + 455, + 486 + ], + "type": "text", + "content": ", offering images of markedly higher fidelity and improved alignment with text prompts. A key feature of " + }, + { + "bbox": [ + 159, + 255, + 455, + 486 + ], + "type": "inline_equation", + "content": "\\mathrm{PIXART - \\Sigma}" + }, + { + "bbox": [ + 159, + 255, + 455, + 486 + ], + "type": "text", + "content": " is its training efficiency. Leveraging the foundational pre-training of " + }, + { + "bbox": [ + 159, + 255, + 455, + 486 + ], + "type": "inline_equation", + "content": "\\mathrm{PIXART - \\alpha}" + }, + { + "bbox": [ + 159, + 255, + 455, + 486 + ], + "type": "text", + "content": ", it evolves from the 'weaker' baseline to a 'stronger' model via incorporating higher quality data, a process we term \"weak-to-strong training\". The advancements in " + }, + { + "bbox": [ + 159, + 255, + 455, + 486 + ], + "type": "inline_equation", + "content": "\\mathrm{PIXART - \\Sigma}" + }, + { + "bbox": [ + 159, + 255, + 455, + 486 + ], + "type": "text", + "content": " are twofold: (1) High-Quality Training Data: " + }, + { + "bbox": [ + 159, + 255, + 455, + 486 + ], + "type": "inline_equation", + "content": "\\mathrm{PIXART - \\Sigma}" + }, + { + "bbox": [ + 159, + 255, + 455, + 486 + ], + "type": "text", + "content": " incorporates superior-quality image data, paired with more precise and detailed image captions. (2) Efficient Token Compression: we propose a novel attention module within the DiT framework that compresses both keys and values, significantly improving efficiency and facilitating ultra-high-resolution image generation. Thanks to these improvements, " + }, + { + "bbox": [ + 159, + 255, + 455, + 486 + ], + "type": "inline_equation", + "content": "\\mathrm{PIXART - \\Sigma}" + }, + { + "bbox": [ + 159, + 255, + 455, + 486 + ], + "type": "text", + "content": " achieves superior image quality and user prompt adherence capabilities with significantly smaller model size (0.6B parameters) than existing text-to-image diffusion models, such as SDXL (2.6B parameters) and SD Cascade (5.1B parameters). Moreover, " + }, + { + "bbox": [ + 159, + 255, + 455, + 486 + ], + "type": "inline_equation", + "content": "\\mathrm{PIXART - \\Sigma}" + }, + { + "bbox": [ + 159, + 255, + 455, + 486 + ], + "type": "text", + "content": "'s capability to generate 4K images supports the creation of high-resolution posters and wallpapers, efficiently bolstering the production of high-quality visual content in industries such as film and gaming." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 160, + 496, + 434, + 508 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 160, + 496, + 434, + 508 + ], + "spans": [ + { + "bbox": [ + 160, + 496, + 434, + 508 + ], + "type": "text", + "content": "Keywords: T2I Synthesis, Diffusion Transformer, Efficient Model" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 132, + 529, + 230, + 541 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 529, + 230, + 541 + ], + "spans": [ + { + "bbox": [ + 132, + 529, + 230, + 541 + ], + "type": "text", + "content": "1 Introduction" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 130, + 554, + 482, + 639 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 554, + 482, + 639 + ], + "spans": [ + { + "bbox": [ + 130, + 554, + 482, + 639 + ], + "type": "text", + "content": "The recent emergence of high-quality Text-to-Image (T2I) models has profoundly impacted the AI Generated Content (AIGC) community. This includes both proprietary models such as DALL-E 3 [31], Midjourney [29], as well as open-source models like Stable Diffusion [36] and PixART-" + }, + { + "bbox": [ + 130, + 554, + 482, + 639 + ], + "type": "inline_equation", + "content": "\\alpha" + }, + { + "bbox": [ + 130, + 554, + 482, + 639 + ], + "type": "text", + "content": " [4]. Nonetheless, developing a top-tier T2I model involves considerable resources; for instance, training SD1.5 from scratch necessitates about 6000 A100 GPU days [36], posing a substantial barrier to individual researchers with limited resources and impeding innovation" + } + ] + } + ], + "index": 6 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 132, + 647, + 478, + 665 + ], + "type": "page_footnote", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 647, + 478, + 665 + ], + "spans": [ + { + "bbox": [ + 132, + 647, + 478, + 665 + ], + "type": "text", + "content": "*Equal contribution. Work done during the students' internships at Huawei Noah's Ark Lab. †Project lead and corresponding author. Thanks to Charan for the 4K dataset collection." + } + ] + } + ], + "index": 7 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 0 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 149, + 114, + 465, + 605 + ], + "blocks": [ + { + "bbox": [ + 149, + 114, + 465, + 605 + ], + "lines": [ + { + "bbox": [ + 149, + 114, + 465, + 605 + ], + "spans": [ + { + "bbox": [ + 149, + 114, + 465, + 605 + ], + "type": "image", + "image_path": "2eae751a45613fdced33b20a28013da2d82387456ecdc459652b02b3a308e744.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 132, + 612, + 482, + 635 + ], + "lines": [ + { + "bbox": [ + 132, + 612, + 482, + 635 + ], + "spans": [ + { + "bbox": [ + 132, + 612, + 482, + 635 + ], + "type": "text", + "content": "Fig. 1: Images generated by " + }, + { + "bbox": [ + 132, + 612, + 482, + 635 + ], + "type": "inline_equation", + "content": "\\mathbf{P}\\mathbf{I}\\mathbf{x}\\mathbf{A}\\mathbf{R}\\mathbf{T} - \\boldsymbol{\\Sigma}" + }, + { + "bbox": [ + 132, + 612, + 482, + 635 + ], + "type": "text", + "content": ". The model can output photo-realistic, high aesthetic, extreme aspect ratio, multi-style images, and follow user instructions." + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_caption" + } + ], + "index": 2 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "spans": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "type": "text", + "content": "2" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 166, + 91, + 223, + 100 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 166, + 91, + 223, + 100 + ], + "spans": [ + { + "bbox": [ + 166, + 91, + 223, + 100 + ], + "type": "text", + "content": "J. Chen et al." + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 1 + }, + { + "para_blocks": [ + { + "bbox": [ + 134, + 115, + 482, + 175 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 134, + 115, + 482, + 175 + ], + "spans": [ + { + "bbox": [ + 134, + 115, + 482, + 175 + ], + "type": "text", + "content": "within the AIGC community. Over time, the AIGC community will gain access to continuously updated, higher-quality datasets and more advanced algorithms. A pivotal question is: how can we efficiently integrate these new elements into an existing model, achieving a more powerful version within the constraints of limited resources?" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 134, + 182, + 482, + 300 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 134, + 182, + 482, + 300 + ], + "spans": [ + { + "bbox": [ + 134, + 182, + 482, + 300 + ], + "type": "text", + "content": "To explore this issue, our research focuses on enhancing " + }, + { + "bbox": [ + 134, + 182, + 482, + 300 + ], + "type": "inline_equation", + "content": "\\mathrm{PIXART - \\alpha}" + }, + { + "bbox": [ + 134, + 182, + 482, + 300 + ], + "type": "text", + "content": ", an efficient T2I training method. " + }, + { + "bbox": [ + 134, + 182, + 482, + 300 + ], + "type": "inline_equation", + "content": "\\mathrm{PIXART - \\alpha}" + }, + { + "bbox": [ + 134, + 182, + 482, + 300 + ], + "type": "text", + "content": " represents an early venture within the DiT framework, a model structure with significant potential, as evidenced by works such as GenTron [6], Sora [33] and Stable Diffusion 3 [42]. To maximize this potential, we build upon the pre-trained foundation of " + }, + { + "bbox": [ + 134, + 182, + 482, + 300 + ], + "type": "inline_equation", + "content": "\\mathrm{PIXART - \\alpha}" + }, + { + "bbox": [ + 134, + 182, + 482, + 300 + ], + "type": "text", + "content": ", integrating advanced elements to facilitate its continuous improvement, resulting in a more powerful model, " + }, + { + "bbox": [ + 134, + 182, + 482, + 300 + ], + "type": "inline_equation", + "content": "\\mathrm{PIXART - \\Sigma}" + }, + { + "bbox": [ + 134, + 182, + 482, + 300 + ], + "type": "text", + "content": ". We refer to this process of evolving from a relatively weaker baseline to a stronger model through efficient training as \"weak-to-strong training\". Specifically, to achieve \"weak-to-strong training\", we introduce the following enhancements:" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 134, + 307, + 482, + 462 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 134, + 307, + 482, + 462 + ], + "spans": [ + { + "bbox": [ + 134, + 307, + 482, + 462 + ], + "type": "text", + "content": "Higher-Quality Training Data: We collect a high-quality dataset superior to that used in " + }, + { + "bbox": [ + 134, + 307, + 482, + 462 + ], + "type": "inline_equation", + "content": "\\mathrm{PIXART - \\alpha}" + }, + { + "bbox": [ + 134, + 307, + 482, + 462 + ], + "type": "text", + "content": ", focusing on two key aspects: (i) High-quality images: The dataset comprises 33M high-resolution images sourced from the Internet, all exceeding 1K resolution, including 2.3M images with resolutions around 4K. These images are predominantly characterized by their high aesthetic and encompass a wide range of artistic styles. (ii) Dense and accurate captions: To provide more precise and detailed captions for the aforementioned images, we replace the LLaVA [22] used in " + }, + { + "bbox": [ + 134, + 307, + 482, + 462 + ], + "type": "inline_equation", + "content": "\\mathrm{PIXART - \\alpha}" + }, + { + "bbox": [ + 134, + 307, + 482, + 462 + ], + "type": "text", + "content": " with a more powerful image captioner, Share-Captioner [5]. Furthermore, to improve the model's alignment capacity between the textual and visual concepts, we extend the token length of the text encoder (i.e., Flan-T5 [10]) to approximately 300 words. We observe these improvements effectively eliminate the model's tendency for hallucination, leading to higher-quality text-image alignment." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 134, + 468, + 482, + 587 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 134, + 468, + 482, + 587 + ], + "spans": [ + { + "bbox": [ + 134, + 468, + 482, + 587 + ], + "type": "text", + "content": "Efficient Token Compression: To enhance " + }, + { + "bbox": [ + 134, + 468, + 482, + 587 + ], + "type": "inline_equation", + "content": "\\mathrm{PIXART - \\alpha}" + }, + { + "bbox": [ + 134, + 468, + 482, + 587 + ], + "type": "text", + "content": ", we expand its generation resolution from 1K to 4K. Generating images at ultra-high resolutions (e.g., 2K/4K) introduces a significant increase in the number of tokens, leading to a substantial rise in computational demand. To address this challenge, we introduced a self-attention module with key and value token compression tailored to the DiT framework. Specifically, we utilize group convolutions with a stride of 2 for local aggregation of keys and values. Additionally, we employ a specialized weight initialization scheme, allowing for a smooth adaptation from a pre-trained model without KV compression. This design effectively reduces training and inference time by " + }, + { + "bbox": [ + 134, + 468, + 482, + 587 + ], + "type": "inline_equation", + "content": "\\sim 34\\%" + }, + { + "bbox": [ + 134, + 468, + 482, + 587 + ], + "type": "text", + "content": " for high-resolution image generation." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 134, + 594, + 482, + 665 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 134, + 594, + 482, + 665 + ], + "spans": [ + { + "bbox": [ + 134, + 594, + 482, + 665 + ], + "type": "text", + "content": "Weak-to-Strong Training Strategy: we propose several fine-tuning techniques to rapidly adapt from a weak model to a strong model efficiently. That includes (1) replacing with a more powerful Variational Autoencoder (VAE) [36], (2) scaling from low to high resolution, and (3) evolving from a model without Key-Value (KV) compression to one with KV compression. These outcomes confirm the validity and effectiveness of the \"weak-to-strong training\" approach." + } + ] + } + ], + "index": 6 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 173, + 90, + 447, + 102 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 173, + 90, + 447, + 102 + ], + "spans": [ + { + "bbox": [ + 173, + 90, + 447, + 102 + ], + "type": "text", + "content": "PixART- " + }, + { + "bbox": [ + 173, + 90, + 447, + 102 + ], + "type": "inline_equation", + "content": "\\boldsymbol{\\Sigma}" + }, + { + "bbox": [ + 173, + 90, + 447, + 102 + ], + "type": "text", + "content": " : Weak-to-Strong Training of DiT for 4K T2I Generation" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 474, + 92, + 480, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 474, + 92, + 480, + 100 + ], + "spans": [ + { + "bbox": [ + 474, + 92, + 480, + 100 + ], + "type": "text", + "content": "3" + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 2 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 134, + 114, + 477, + 329 + ], + "blocks": [ + { + "bbox": [ + 134, + 114, + 477, + 329 + ], + "lines": [ + { + "bbox": [ + 134, + 114, + 477, + 329 + ], + "spans": [ + { + "bbox": [ + 134, + 114, + 477, + 329 + ], + "type": "image", + "image_path": "7e35c89910d2f12dd3e67db0e4bd640a959fe122d35481f09a9a261e1d50789e.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 130, + 335, + 482, + 369 + ], + "lines": [ + { + "bbox": [ + 130, + 335, + 482, + 369 + ], + "spans": [ + { + "bbox": [ + 130, + 335, + 482, + 369 + ], + "type": "text", + "content": "Fig. 2: 4K image generation with complex dense instructions. " + }, + { + "bbox": [ + 130, + 335, + 482, + 369 + ], + "type": "inline_equation", + "content": "\\mathrm{PIXART - \\Sigma}" + }, + { + "bbox": [ + 130, + 335, + 482, + 369 + ], + "type": "text", + "content": " can directly generate 4K resolution images without post-processing, and accurately respond to the given prompt." + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_caption" + } + ], + "index": 2 + }, + { + "bbox": [ + 130, + 370, + 482, + 502 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 370, + 482, + 502 + ], + "spans": [ + { + "bbox": [ + 130, + 370, + 482, + 502 + ], + "type": "text", + "content": "Through the proposed improvements, " + }, + { + "bbox": [ + 130, + 370, + 482, + 502 + ], + "type": "inline_equation", + "content": "\\mathrm{PIXART - \\Sigma}" + }, + { + "bbox": [ + 130, + 370, + 482, + 502 + ], + "type": "text", + "content": " achieves high-quality 4K resolution image generation at a minimal training cost and model parameters. Specifically, fine-tuning from a pre-trained model, we additionally utilize only " + }, + { + "bbox": [ + 130, + 370, + 482, + 502 + ], + "type": "inline_equation", + "content": "9\\%" + }, + { + "bbox": [ + 130, + 370, + 482, + 502 + ], + "type": "text", + "content": " of the GPU days required by " + }, + { + "bbox": [ + 130, + 370, + 482, + 502 + ], + "type": "inline_equation", + "content": "\\mathrm{PIXART - \\alpha}" + }, + { + "bbox": [ + 130, + 370, + 482, + 502 + ], + "type": "text", + "content": " to achieve a strong 1K high-resolution image generation model, which is impressive considering replacing with new training data and a more powerful VAE. Moreover, we use only 0.6B parameters while SDXL [36] and SD Cascade [35] use 2.6B and 5.1B parameters respectively. Images generated by " + }, + { + "bbox": [ + 130, + 370, + 482, + 502 + ], + "type": "inline_equation", + "content": "\\mathrm{PIXART - \\Sigma}" + }, + { + "bbox": [ + 130, + 370, + 482, + 502 + ], + "type": "text", + "content": " possess an aesthetic quality comparable to current top-tier T2I products, such as DALL-E 3 [31] and MJV6 [29] (as in Fig. 4). Additionally, " + }, + { + "bbox": [ + 130, + 370, + 482, + 502 + ], + "type": "inline_equation", + "content": "\\mathrm{PIXART - \\Sigma}" + }, + { + "bbox": [ + 130, + 370, + 482, + 502 + ], + "type": "text", + "content": " also demonstrates exceptional capability for fine-grained alignment with textual prompts (as shown in Fig. 2 and 3)." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 132, + 520, + 237, + 533 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 520, + 237, + 533 + ], + "spans": [ + { + "bbox": [ + 132, + 520, + 237, + 533 + ], + "type": "text", + "content": "2 Related Work" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 130, + 545, + 482, + 666 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 545, + 482, + 666 + ], + "spans": [ + { + "bbox": [ + 130, + 545, + 482, + 666 + ], + "type": "text", + "content": "Diffusion Transformers. The Transformer architecture has achieved remarkable success across various domains, such as language modeling [37,38], computer vision [23, 43, 51, 53], and other areas [3, 14]. In the realm of diffusion models, DiT [34] and UViT [2] pioneer the use of Transformer architecture. Subsequent works, including DiffiT [15], SiT [28], and FiT [26], have improved upon DiT's architecture, while [12, 52] enhance training efficiency through masked modeling techniques. For Text-to-Image (T2I) synthesis, PixART-" + }, + { + "bbox": [ + 130, + 545, + 482, + 666 + ], + "type": "inline_equation", + "content": "\\alpha" + }, + { + "bbox": [ + 130, + 545, + 482, + 666 + ], + "type": "text", + "content": " [4] explore efficient T2I training schemes, achieving the first Transformer-based T2I model capable of generating 1024px high-quality images. GenTron [6] explores the flexibility and scalability of diffusion Transformers in both image and video generation. The re" + } + ] + } + ], + "index": 6 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "spans": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "type": "text", + "content": "4" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 166, + 91, + 222, + 100 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 166, + 91, + 222, + 100 + ], + "spans": [ + { + "bbox": [ + 166, + 91, + 222, + 100 + ], + "type": "text", + "content": "J. Chen et al." + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 3 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 134, + 114, + 246, + 238 + ], + "blocks": [ + { + "bbox": [ + 134, + 114, + 246, + 238 + ], + "lines": [ + { + "bbox": [ + 134, + 114, + 246, + 238 + ], + "spans": [ + { + "bbox": [ + 134, + 114, + 246, + 238 + ], + "type": "image", + "image_path": "fb9c7c710fe0e9d2ee06c3f7f4d3cf783b86318f0d6b42b108580da8e3edc19c.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 136, + 240, + 474, + 257 + ], + "lines": [ + { + "bbox": [ + 136, + 240, + 474, + 257 + ], + "spans": [ + { + "bbox": [ + 136, + 240, + 474, + 257 + ], + "type": "text", + "content": "Prompt: A close-up photo of a person. The subject is a woman. She wore a blue coat with a gray dress underneath. She has blue eyes and blond hair, and wears a pair of earrings. Behind are blurred city buildings and streets." + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "image_footnote" + } + ], + "index": 2 + }, + { + "type": "image", + "bbox": [ + 248, + 114, + 361, + 238 + ], + "blocks": [ + { + "bbox": [ + 248, + 114, + 361, + 238 + ], + "lines": [ + { + "bbox": [ + 248, + 114, + 361, + 238 + ], + "spans": [ + { + "bbox": [ + 248, + 114, + 361, + 238 + ], + "type": "image", + "image_path": "12e79cda23078bf0271910267e946835d7e8f827a93e81cc47ed7c78fbbc42e9.jpg" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_body" + } + ], + "index": 3 + }, + { + "type": "image", + "bbox": [ + 362, + 114, + 476, + 238 + ], + "blocks": [ + { + "bbox": [ + 362, + 114, + 476, + 238 + ], + "lines": [ + { + "bbox": [ + 362, + 114, + 476, + 238 + ], + "spans": [ + { + "bbox": [ + 362, + 114, + 476, + 238 + ], + "type": "image", + "image_path": "e41f0f585330f34bde73a7d6b4aacb5c0c71eccf8581ed83c3162caa14c4a68d.jpg" + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "image_body" + } + ], + "index": 4 + }, + { + "type": "image", + "bbox": [ + 135, + 258, + 247, + 370 + ], + "blocks": [ + { + "bbox": [ + 135, + 258, + 247, + 370 + ], + "lines": [ + { + "bbox": [ + 135, + 258, + 247, + 370 + ], + "spans": [ + { + "bbox": [ + 135, + 258, + 247, + 370 + ], + "type": "image", + "image_path": "ea7a0a13f17110ee9ad20ffc30c8e4306c16b7efceab99065c96d8c050e7ad46.jpg" + } + ] + } + ], + "index": 6, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 136, + 372, + 343, + 379 + ], + "lines": [ + { + "bbox": [ + 136, + 372, + 343, + 379 + ], + "spans": [ + { + "bbox": [ + 136, + 372, + 343, + 379 + ], + "type": "text", + "content": "Prompt: half a solid black background and half a solid white background" + } + ] + } + ], + "index": 9, + "angle": 0, + "type": "image_footnote" + } + ], + "index": 6 + }, + { + "type": "image", + "bbox": [ + 253, + 258, + 361, + 369 + ], + "blocks": [ + { + "bbox": [ + 253, + 258, + 361, + 369 + ], + "lines": [ + { + "bbox": [ + 253, + 258, + 361, + 369 + ], + "spans": [ + { + "bbox": [ + 253, + 258, + 361, + 369 + ], + "type": "image", + "image_path": "2a9f33ba284496177d5d438d058378e702d205d9b8133067ed20332253946dd5.jpg" + } + ] + } + ], + "index": 7, + "angle": 0, + "type": "image_body" + } + ], + "index": 7 + }, + { + "type": "image", + "bbox": [ + 364, + 258, + 421, + 370 + ], + "blocks": [ + { + "bbox": [ + 364, + 258, + 421, + 370 + ], + "lines": [ + { + "bbox": [ + 364, + 258, + 421, + 370 + ], + "spans": [ + { + "bbox": [ + 364, + 258, + 421, + 370 + ], + "type": "image", + "image_path": "aa2f55b578413185cd7fffa3afbcab97e71b7f0f458f495054ee6eb8905ee330.jpg" + } + ] + } + ], + "index": 8, + "angle": 0, + "type": "image_body" + } + ], + "index": 8 + }, + { + "type": "image", + "bbox": [ + 135, + 381, + 247, + 492 + ], + "blocks": [ + { + "bbox": [ + 135, + 381, + 247, + 492 + ], + "lines": [ + { + "bbox": [ + 135, + 381, + 247, + 492 + ], + "spans": [ + { + "bbox": [ + 135, + 381, + 247, + 492 + ], + "type": "image", + "image_path": "39caa60e2a5cfb34fb011d38b26399a54a7ce7c4d76b21e0709d0b22a2a4aea0.jpg" + } + ] + } + ], + "index": 10, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 135, + 494, + 463, + 519 + ], + "lines": [ + { + "bbox": [ + 135, + 494, + 463, + 519 + ], + "spans": [ + { + "bbox": [ + 135, + 494, + 463, + 519 + ], + "type": "text", + "content": "Prompt: Pixel art style of a snowboarder in mid-air performs a trick on a black rail, wearing a blue sweatshirt and black pants, with arms outstretched. The serene snowy landscape background, dotted with trees, complements the scene. The low-angle perspective emphasizes the trick's height and skill." + } + ] + } + ], + "index": 13, + "angle": 0, + "type": "image_footnote" + }, + { + "bbox": [ + 130, + 525, + 482, + 570 + ], + "lines": [ + { + "bbox": [ + 130, + 525, + 482, + 570 + ], + "spans": [ + { + "bbox": [ + 130, + 525, + 482, + 570 + ], + "type": "text", + "content": "Fig. 3: Comparison of " + }, + { + "bbox": [ + 130, + 525, + 482, + 570 + ], + "type": "inline_equation", + "content": "\\mathbf{P}\\mathbf{i}\\mathbf{x}\\mathbf{A}\\mathbf{r}\\mathbf{T} - \\boldsymbol{\\Sigma}" + }, + { + "bbox": [ + 130, + 525, + 482, + 570 + ], + "type": "text", + "content": " with open-source models, e.g., " + }, + { + "bbox": [ + 130, + 525, + 482, + 570 + ], + "type": "inline_equation", + "content": "\\mathbf{P}\\mathbf{i}\\mathbf{x}\\mathbf{A}\\mathbf{r}\\mathbf{T} - \\alpha" + }, + { + "bbox": [ + 130, + 525, + 482, + 570 + ], + "type": "text", + "content": " and SDXL: Compared with " + }, + { + "bbox": [ + 130, + 525, + 482, + 570 + ], + "type": "inline_equation", + "content": "\\mathbf{P}\\mathbf{i}\\mathbf{x}\\mathbf{A}\\mathbf{r}\\mathbf{T} - \\alpha" + }, + { + "bbox": [ + 130, + 525, + 482, + 570 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 130, + 525, + 482, + 570 + ], + "type": "inline_equation", + "content": "\\mathbf{P}\\mathbf{i}\\mathbf{x}\\mathbf{A}\\mathbf{r}\\mathbf{T} - \\boldsymbol{\\Sigma}" + }, + { + "bbox": [ + 130, + 525, + 482, + 570 + ], + "type": "text", + "content": " improves the realism of portraits and the capability of semantic analysis. Compared with SDXL, our method has a better ability to follow user instructions. The keywords are highlighted as blue." + } + ] + } + ], + "index": 14, + "angle": 0, + "type": "image_caption" + } + ], + "index": 10 + }, + { + "type": "image", + "bbox": [ + 249, + 380, + 361, + 492 + ], + "blocks": [ + { + "bbox": [ + 249, + 380, + 361, + 492 + ], + "lines": [ + { + "bbox": [ + 249, + 380, + 361, + 492 + ], + "spans": [ + { + "bbox": [ + 249, + 380, + 361, + 492 + ], + "type": "image", + "image_path": "d55309bb589a3373b0becb9b973a4e2557629f448d5ee2d29acb693587b4d085.jpg" + } + ] + } + ], + "index": 11, + "angle": 0, + "type": "image_body" + } + ], + "index": 11 + }, + { + "type": "image", + "bbox": [ + 363, + 380, + 476, + 492 + ], + "blocks": [ + { + "bbox": [ + 363, + 380, + 476, + 492 + ], + "lines": [ + { + "bbox": [ + 363, + 380, + 476, + 492 + ], + "spans": [ + { + "bbox": [ + 363, + 380, + 476, + 492 + ], + "type": "image", + "image_path": "c526f9445e51f205cd246014870e00ff4bdc8233f908d7f120ef10d62e7c5143.jpg" + } + ] + } + ], + "index": 12, + "angle": 0, + "type": "image_body" + } + ], + "index": 12 + }, + { + "bbox": [ + 130, + 578, + 482, + 638 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 578, + 482, + 638 + ], + "spans": [ + { + "bbox": [ + 130, + 578, + 482, + 638 + ], + "type": "text", + "content": "cent advent of the powerful video generation model Sora [33] has further underscored the potential of Diffusion Transformers. In this work, for the first time, we explore using the Transformer architecture to generate 4K ultra-high-resolution images directly, tackling the computational complexity challenges posed by involving long-sequence tokens." + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 130, + 641, + 482, + 666 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 641, + 482, + 666 + ], + "spans": [ + { + "bbox": [ + 130, + 641, + 482, + 666 + ], + "type": "text", + "content": "High Resolution Image Generation greatly enhances visual quality and is important in various industries such as film and gaming. However, increasing" + } + ] + } + ], + "index": 16 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 173, + 91, + 447, + 102 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 173, + 91, + 447, + 102 + ], + "spans": [ + { + "bbox": [ + 173, + 91, + 447, + 102 + ], + "type": "text", + "content": "PixART- " + }, + { + "bbox": [ + 173, + 91, + 447, + 102 + ], + "type": "inline_equation", + "content": "\\boldsymbol{\\Sigma}" + }, + { + "bbox": [ + 173, + 91, + 447, + 102 + ], + "type": "text", + "content": " : Weak-to-Strong Training of DiT for 4K T2I Generation" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 474, + 92, + 480, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 474, + 92, + 480, + 100 + ], + "spans": [ + { + "bbox": [ + 474, + 92, + 480, + 100 + ], + "type": "text", + "content": "5" + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 4 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 133, + 114, + 304, + 286 + ], + "blocks": [ + { + "bbox": [ + 133, + 114, + 304, + 286 + ], + "lines": [ + { + "bbox": [ + 133, + 114, + 304, + 286 + ], + "spans": [ + { + "bbox": [ + 133, + 114, + 304, + 286 + ], + "type": "image", + "image_path": "14c5470177a9e12fbbd690e435925b8cf2b8786a6527e154d55c523f3dce123d.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_body" + } + ], + "index": 2 + }, + { + "type": "image", + "bbox": [ + 304, + 114, + 391, + 286 + ], + "blocks": [ + { + "bbox": [ + 304, + 114, + 391, + 286 + ], + "lines": [ + { + "bbox": [ + 304, + 114, + 391, + 286 + ], + "spans": [ + { + "bbox": [ + 304, + 114, + 391, + 286 + ], + "type": "image", + "image_path": "2d57c3884bd71131bb8388954711c28bdd9a712c0033176cb3944a4d2f656fef.jpg" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_body" + } + ], + "index": 3 + }, + { + "type": "image", + "bbox": [ + 391, + 114, + 477, + 286 + ], + "blocks": [ + { + "bbox": [ + 391, + 114, + 477, + 286 + ], + "lines": [ + { + "bbox": [ + 391, + 114, + 477, + 286 + ], + "spans": [ + { + "bbox": [ + 391, + 114, + 477, + 286 + ], + "type": "image", + "image_path": "d556f7bae82ad9665e6187894d569497c6b51ec9c6a93ba90053c62cd9c38460.jpg" + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "image_body" + } + ], + "index": 4 + }, + { + "bbox": [ + 134, + 289, + 313, + 296 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 134, + 289, + 313, + 296 + ], + "spans": [ + { + "bbox": [ + 134, + 289, + 313, + 296 + ], + "type": "text", + "content": "Prompt: a small cactus with a happy face in the Sahara desert" + } + ] + } + ], + "index": 5 + }, + { + "type": "image", + "bbox": [ + 133, + 297, + 304, + 469 + ], + "blocks": [ + { + "bbox": [ + 133, + 297, + 304, + 469 + ], + "lines": [ + { + "bbox": [ + 133, + 297, + 304, + 469 + ], + "spans": [ + { + "bbox": [ + 133, + 297, + 304, + 469 + ], + "type": "image", + "image_path": "d180f59fd0eb73be17a0325501067e112bcbcab0ac6b83eb0a16027ac91e92e6.jpg" + } + ] + } + ], + "index": 6, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 132, + 483, + 482, + 517 + ], + "lines": [ + { + "bbox": [ + 132, + 483, + 482, + 517 + ], + "spans": [ + { + "bbox": [ + 132, + 483, + 482, + 517 + ], + "type": "text", + "content": "Fig. 4: Compare " + }, + { + "bbox": [ + 132, + 483, + 482, + 517 + ], + "type": "inline_equation", + "content": "\\mathbf{P}\\mathbf{i}\\mathbf{x}\\mathbf{A}\\mathbf{r}\\mathbf{T} - \\boldsymbol{\\Sigma}" + }, + { + "bbox": [ + 132, + 483, + 482, + 517 + ], + "type": "text", + "content": " and four other T2I products: Firefly 2, Imagen 2, Dalle 3, and Midjourney 6. Images generated by " + }, + { + "bbox": [ + 132, + 483, + 482, + 517 + ], + "type": "inline_equation", + "content": "\\mathbf{P}\\mathbf{i}\\mathbf{x}\\mathbf{A}\\mathbf{r}\\mathbf{T} - \\boldsymbol{\\Sigma}" + }, + { + "bbox": [ + 132, + 483, + 482, + 517 + ], + "type": "text", + "content": " are very competitive with these commercial products." + } + ] + } + ], + "index": 10, + "angle": 0, + "type": "image_caption" + } + ], + "index": 6 + }, + { + "type": "image", + "bbox": [ + 304, + 297, + 391, + 469 + ], + "blocks": [ + { + "bbox": [ + 304, + 297, + 391, + 469 + ], + "lines": [ + { + "bbox": [ + 304, + 297, + 391, + 469 + ], + "spans": [ + { + "bbox": [ + 304, + 297, + 391, + 469 + ], + "type": "image", + "image_path": "8fa7178d9227a3b60727c4ce8ce09e21c5a52f26a58a40283722b2a94a6a2561.jpg" + } + ] + } + ], + "index": 7, + "angle": 0, + "type": "image_body" + } + ], + "index": 7 + }, + { + "type": "image", + "bbox": [ + 391, + 297, + 477, + 469 + ], + "blocks": [ + { + "bbox": [ + 391, + 297, + 477, + 469 + ], + "lines": [ + { + "bbox": [ + 391, + 297, + 477, + 469 + ], + "spans": [ + { + "bbox": [ + 391, + 297, + 477, + 469 + ], + "type": "image", + "image_path": "6de248d743f167195a1999991156d028247a18fa3f01e7cf191918951ed62eac.jpg" + } + ] + } + ], + "index": 8, + "angle": 0, + "type": "image_body" + } + ], + "index": 8 + }, + { + "bbox": [ + 134, + 470, + 223, + 479 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 134, + 470, + 223, + 479 + ], + "spans": [ + { + "bbox": [ + 134, + 470, + 223, + 479 + ], + "type": "text", + "content": "Prompt: a cherry pick scientist" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 130, + 533, + 482, + 666 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 533, + 482, + 666 + ], + "spans": [ + { + "bbox": [ + 130, + 533, + 482, + 666 + ], + "type": "text", + "content": "image resolution introduces challenges due to the substantial increase in computational demands. Numerous methods have been explored in this direction. For instance, Imagen [40], GigaGAN [18] and Stable Diffusion [39] introduce an additional super-resolution network, while Stable Cascade [35] employs multiple diffusion networks to increase resolution progressively. These combined-model solutions, however, can introduce cumulative errors. On the other hand, works like SDXL [36], DALL-E 2 [30], Playground [19] and " + }, + { + "bbox": [ + 130, + 533, + 482, + 666 + ], + "type": "inline_equation", + "content": "\\mathrm{PIXART - \\alpha}" + }, + { + "bbox": [ + 130, + 533, + 482, + 666 + ], + "type": "text", + "content": " [4] have attempted to generate high-resolution images using diffusion models directly. Nevertheless, these efforts are capped at generating images with resolutions up to " + }, + { + "bbox": [ + 130, + 533, + 482, + 666 + ], + "type": "inline_equation", + "content": "1024\\mathrm{px}" + }, + { + "bbox": [ + 130, + 533, + 482, + 666 + ], + "type": "text", + "content": " due to computational complexity. In this paper, we push this boundary to " + }, + { + "bbox": [ + 130, + 533, + 482, + 666 + ], + "type": "inline_equation", + "content": "4\\mathrm{K}" + }, + { + "bbox": [ + 130, + 533, + 482, + 666 + ], + "type": "text", + "content": " resolution, significantly enhancing the visual quality of the generated content." + } + ] + } + ], + "index": 11 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "spans": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "type": "text", + "content": "6" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 166, + 91, + 222, + 100 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 166, + 91, + 222, + 100 + ], + "spans": [ + { + "bbox": [ + 166, + 91, + 222, + 100 + ], + "type": "text", + "content": "J. Chen et al." + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 5 + }, + { + "para_blocks": [ + { + "bbox": [ + 130, + 116, + 482, + 285 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 116, + 482, + 285 + ], + "spans": [ + { + "bbox": [ + 130, + 116, + 482, + 285 + ], + "type": "text", + "content": "Efficient Transformer architecture. The self-attention mechanism in Transformer suffers from quadratic computational complexity with respect to the number of tokens, which hinders the scaling up of token quantity. Many works have sought improvements in this area: (1) Sparse Attention [7, 8, 45, 46, 48], which reduces the overall computational load by selectively processing a subset of tokens. For instance, PVT v2 [46] employs a convolutional kernel to condense the space of the key and value, thus lowering the complexity involved in computing the attention. (2) Local Attention [13, 23, 47, 54] focuses attention within nearby regions; notably, Swin Transformer [23] utilizes window-based attention to limit computations to within a specified window size. (3) Low-rank/Linear Attention [9, 25, 44]. The Linformer [44] reduces the computational complexity of the self-attention mechanism through low-rank approximations. In this paper, inspired by PVT v2 [46], we employ a self-attention mechanism based on key/value compression to mitigate the high complexity of processing 4K images." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 132, + 301, + 221, + 314 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 301, + 221, + 314 + ], + "spans": [ + { + "bbox": [ + 132, + 301, + 221, + 314 + ], + "type": "text", + "content": "3 Framework" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 132, + 326, + 231, + 338 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 326, + 231, + 338 + ], + "spans": [ + { + "bbox": [ + 132, + 326, + 231, + 338 + ], + "type": "text", + "content": "3.1 Data Analysis" + } + ] + } + ], + "index": 4 + }, + { + "type": "table", + "bbox": [ + 136, + 339, + 477, + 453 + ], + "blocks": [ + { + "bbox": [ + 136, + 339, + 477, + 453 + ], + "lines": [ + { + "bbox": [ + 136, + 339, + 477, + 453 + ], + "spans": [ + { + "bbox": [ + 136, + 339, + 477, + 453 + ], + "type": "table", + "html": "
LLaVA HallucinationsThe image features a large, ornate church with a tall, pointed roof and a large stained-glass window. The church is situated in a city, and it has a white and gray color scheme. The style of the church is Gothic, which is evident from the pointed roof and the intricate details of the stained-glass window. The presence of statues and a clock on the building further emphasizes the grand and historical nature of the structure.
Share-Captioner CorrectnessThe image captures the grandeur of a cathedral, painted in a pristine white, stands majestically against the backdrop of a clear blue sky. The entrance to the cathedral is adorned with a large rose window. Flanking the entrance are statues of saints, standing as silent guardians of the sacred space within. The perspective of the photo, taken from a low angle. The lush green trees in the background adding serenity to the scene.
LLaVA HallucinationsThe image features a woman and a man sitting on a brick walkway near a body of water, which could be a river or a lake. They are both wearing head coverings, and the woman is holding a handbag. The scene is set during the day, with the sun shining brightly, creating a warm and inviting atmosphere. The style of the image is a black and white photo, which adds a timeless and classic feel to the scene.
Share-Captioner CorrectnessThe image captures a serene scene at a harbor. Two individuals are seated on a bench, their backs to the camera, engrossed in the view of the water. The water, a deep shade of blue, is dotted with boats of various sizes and colors, including a white boat with a green stripe and a red boat. The sky above is a light blue.
", + "image_path": "419a573d179dfc5658c9ac8963a6cf5f15a7cd350fe87f0b24c3816cc8859715.jpg" + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "table_body" + } + ], + "index": 5 + }, + { + "bbox": [ + 130, + 459, + 482, + 493 + ], + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 459, + 482, + 493 + ], + "spans": [ + { + "bbox": [ + 130, + 459, + 482, + 493 + ], + "type": "text", + "content": "Fig. 5: Comparative illustration of hallucinations: Contrasting differences in hallucination occurrences between LLaVA and Share-Captioner, with red indicating hallucinations and green denoting correctness." + } + ] + } + ], + "index": 6, + "type": "text" + }, + { + "bbox": [ + 130, + 498, + 482, + 581 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 498, + 482, + 581 + ], + "spans": [ + { + "bbox": [ + 130, + 498, + 482, + 581 + ], + "type": "text", + "content": "Higher Aesthetic and higher Resolution. To enhance the aesthetic quality of our dataset, we expand our internal data from 14M to 33M. For clarity, we name the two datasets Internal- " + }, + { + "bbox": [ + 130, + 498, + 482, + 581 + ], + "type": "inline_equation", + "content": "\\alpha" + }, + { + "bbox": [ + 130, + 498, + 482, + 581 + ], + "type": "text", + "content": " and Internal- " + }, + { + "bbox": [ + 130, + 498, + 482, + 581 + ], + "type": "inline_equation", + "content": "\\Sigma" + }, + { + "bbox": [ + 130, + 498, + 482, + 581 + ], + "type": "text", + "content": ", respectively. Note that this expansion still falls short compared to the vast images utilized by currently available open-source models like SD v1.5, which uses 2B data. We demonstrate that effective training strategies with limited data amount can still obtain a strong T2I model." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 130, + 582, + 482, + 641 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 582, + 482, + 641 + ], + "spans": [ + { + "bbox": [ + 130, + 582, + 482, + 641 + ], + "type": "text", + "content": "The images within Internal- " + }, + { + "bbox": [ + 130, + 582, + 482, + 641 + ], + "type": "inline_equation", + "content": "\\Sigma" + }, + { + "bbox": [ + 130, + 582, + 482, + 641 + ], + "type": "text", + "content": " are above 1K resolution. To facilitate 4K resolution generation, we additionally collect a dataset of 8M real photographic images at 4K resolution. To ensure aesthetic quality, we employ an aesthetic scoring model (AES) [1] to filter these 4K images. This process yields a highly refined dataset of 2M ultra-high-resolution and high-quality images." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 130, + 641, + 482, + 665 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 641, + 482, + 665 + ], + "spans": [ + { + "bbox": [ + 130, + 641, + 482, + 665 + ], + "type": "text", + "content": "Interestingly, we have observed that as the resolution of the images increases, there is an improvement in the model's fidelity (Fréchet Inception Dis-" + } + ] + } + ], + "index": 9 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 172, + 90, + 447, + 102 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 172, + 90, + 447, + 102 + ], + "spans": [ + { + "bbox": [ + 172, + 90, + 447, + 102 + ], + "type": "text", + "content": "PIXART- " + }, + { + "bbox": [ + 172, + 90, + 447, + 102 + ], + "type": "inline_equation", + "content": "\\boldsymbol{\\Sigma}" + }, + { + "bbox": [ + 172, + 90, + 447, + 102 + ], + "type": "text", + "content": " : Weak-to-Strong Training of DiT for 4K T2I Generation" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 474, + 91, + 480, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 474, + 91, + 480, + 100 + ], + "spans": [ + { + "bbox": [ + 474, + 91, + 480, + 100 + ], + "type": "text", + "content": "7" + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 6 + }, + { + "para_blocks": [ + { + "type": "table", + "bbox": [ + 141, + 140, + 467, + 234 + ], + "blocks": [ + { + "bbox": [ + 132, + 105, + 482, + 138 + ], + "lines": [ + { + "bbox": [ + 132, + 105, + 482, + 138 + ], + "spans": [ + { + "bbox": [ + 132, + 105, + 482, + 138 + ], + "type": "text", + "content": "Table 1: Statistics of noun concepts for different datasets. VN: valid distinct nouns (appearing more than 10 times); DN: total distinct nouns; Average: average noun count per image; ACL: Average Caption length." + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 141, + 140, + 467, + 234 + ], + "lines": [ + { + "bbox": [ + 141, + 140, + 467, + 234 + ], + "spans": [ + { + "bbox": [ + 141, + 140, + 467, + 234 + ], + "type": "table", + "html": "
DatasetVolumeCaptionVN/DNTotal NounACLAverage
Internal-α14MRaw187K/931K175M2511.7/Img
Internal-α14MLLaVA28K/215K536M9829.3/Img
Internal-α14MShare-Captioner51K/420K815M18454.4/Img
Internal-Σ33MRaw294K/1512K485M3514.4/Img
Internal-Σ33MShare-Captioner77K/714K1804M18053.6/Img
4K-Σ2.3MShare-Captioner24K/96K115M16349.5/Img
", + "image_path": "0364c1206f08d202fc8bbafc49beec55b269b47752aa00663ce43cb832701365.jpg" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "table_body" + } + ], + "index": 3 + }, + { + "bbox": [ + 130, + 250, + 479, + 274 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 250, + 479, + 274 + ], + "spans": [ + { + "bbox": [ + 130, + 250, + 479, + 274 + ], + "type": "text", + "content": "tance (FID) [17]) and semantic alignment (CLIP Score), which underscores the importance of the capabilities of generating high-resolution images." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 132, + 276, + 297, + 287 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 276, + 297, + 287 + ], + "spans": [ + { + "bbox": [ + 132, + 276, + 297, + 287 + ], + "type": "text", + "content": "Better Text-Image Alignment." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 130, + 288, + 298, + 492 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 288, + 298, + 492 + ], + "spans": [ + { + "bbox": [ + 130, + 288, + 298, + 492 + ], + "type": "text", + "content": "Recent works such as " + }, + { + "bbox": [ + 130, + 288, + 298, + 492 + ], + "type": "inline_equation", + "content": "\\mathrm{PIXART - }\\alpha" + }, + { + "bbox": [ + 130, + 288, + 298, + 492 + ], + "type": "text", + "content": " [4] and DALL-E 3 [31] emphasize the significance of text-image alignment. Strengthening this alignment is crucial for boosting model capabilities. To refine our collected \"raw\" descriptions further, we focus on improving both the length and accuracy of our captions. Notably, our captions (Internal- " + }, + { + "bbox": [ + 130, + 288, + 298, + 492 + ], + "type": "inline_equation", + "content": "\\Sigma" + }, + { + "bbox": [ + 130, + 288, + 298, + 492 + ], + "type": "text", + "content": " ) show several advantages over the one used in " + }, + { + "bbox": [ + 130, + 288, + 298, + 492 + ], + "type": "inline_equation", + "content": "\\mathrm{PIXART - }\\alpha" + }, + { + "bbox": [ + 130, + 288, + 298, + 492 + ], + "type": "text", + "content": " (Internal- " + }, + { + "bbox": [ + 130, + 288, + 298, + 492 + ], + "type": "inline_equation", + "content": "\\alpha" + }, + { + "bbox": [ + 130, + 288, + 298, + 492 + ], + "type": "text", + "content": " ) in the following aspects: 1. Enhanced caption accuracy: As depicted in Fig. 5, LLaVa used in " + }, + { + "bbox": [ + 130, + 288, + 298, + 492 + ], + "type": "inline_equation", + "content": "\\mathrm{PIXART - }\\alpha" + }, + { + "bbox": [ + 130, + 288, + 298, + 492 + ], + "type": "text", + "content": " encounters certain hallucination. We leverage a more powerful model, i.e., Share-Captioner [5], to" + } + ] + } + ], + "index": 6 + }, + { + "type": "image", + "bbox": [ + 308, + 281, + 479, + 405 + ], + "blocks": [ + { + "bbox": [ + 308, + 281, + 479, + 405 + ], + "lines": [ + { + "bbox": [ + 308, + 281, + 479, + 405 + ], + "spans": [ + { + "bbox": [ + 308, + 281, + 479, + 405 + ], + "type": "image", + "image_path": "2779e528b6630f430c51a8bf54e45109d8b21c4d9a8462cb49a0b90232c2228d.jpg" + } + ] + } + ], + "index": 7, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 304, + 410, + 482, + 476 + ], + "lines": [ + { + "bbox": [ + 304, + 410, + 482, + 476 + ], + "spans": [ + { + "bbox": [ + 304, + 410, + 482, + 476 + ], + "type": "text", + "content": "Fig.6: Histogram Visualization of the Caption Length. We randomly select 1M captions from the raw captions, Internal- " + }, + { + "bbox": [ + 304, + 410, + 482, + 476 + ], + "type": "inline_equation", + "content": "\\alpha" + }, + { + "bbox": [ + 304, + 410, + 482, + 476 + ], + "type": "text", + "content": " , and Internal- " + }, + { + "bbox": [ + 304, + 410, + 482, + 476 + ], + "type": "inline_equation", + "content": "\\Sigma" + }, + { + "bbox": [ + 304, + 410, + 482, + 476 + ], + "type": "text", + "content": " to draw the corresponding histogram. ACL denotes the average caption length." + } + ] + } + ], + "index": 8, + "angle": 0, + "type": "image_caption" + } + ], + "index": 7 + }, + { + "bbox": [ + 130, + 493, + 477, + 506 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 493, + 477, + 506 + ], + "spans": [ + { + "bbox": [ + 130, + 493, + 477, + 506 + ], + "type": "text", + "content": "generate detailed and correct captions, augmenting the collected raw prompts." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 130, + 506, + 482, + 602 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 506, + 482, + 602 + ], + "spans": [ + { + "bbox": [ + 130, + 506, + 482, + 602 + ], + "type": "text", + "content": "2. Increased caption length: As shown in Tab. 1 and Fig. 6, the average caption length increased significantly to 180 words, highly enhancing the descriptive power of the captions. Additionally, we extend the token processing length of the text encoder from 120 tokens (as in Internal-" + }, + { + "bbox": [ + 130, + 506, + 482, + 602 + ], + "type": "inline_equation", + "content": "\\alpha" + }, + { + "bbox": [ + 130, + 506, + 482, + 602 + ], + "type": "text", + "content": ") to 300 tokens. Our model is trained on a mix of long (Share-Captioner) and short (raw) captions with a ratio of " + }, + { + "bbox": [ + 130, + 506, + 482, + 602 + ], + "type": "inline_equation", + "content": "60\\%" + }, + { + "bbox": [ + 130, + 506, + 482, + 602 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 130, + 506, + 482, + 602 + ], + "type": "inline_equation", + "content": "40\\%" + }, + { + "bbox": [ + 130, + 506, + 482, + 602 + ], + "type": "text", + "content": ", respectively. This approach enhances the diversity of textual descriptions and mitigates potential biases that might arise from solely relying on generative captions." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 130, + 604, + 482, + 639 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 604, + 482, + 639 + ], + "spans": [ + { + "bbox": [ + 130, + 604, + 482, + 639 + ], + "type": "text", + "content": "Tab. 1 demonstrates a summary for both Internal-" + }, + { + "bbox": [ + 130, + 604, + 482, + 639 + ], + "type": "inline_equation", + "content": "\\alpha" + }, + { + "bbox": [ + 130, + 604, + 482, + 639 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 130, + 604, + 482, + 639 + ], + "type": "inline_equation", + "content": "-\\Sigma" + }, + { + "bbox": [ + 130, + 604, + 482, + 639 + ], + "type": "text", + "content": ", where we assess the diversity of the datasets through various metrics, including the noun variety, total noun count, average caption length, and average nouns per image." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 130, + 641, + 491, + 665 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 641, + 491, + 665 + ], + "spans": [ + { + "bbox": [ + 130, + 641, + 491, + 665 + ], + "type": "text", + "content": "High-Quality Evaluation Dataset. Most SoTA T2I models chose MSCOCO [21] as the evaluation set to assess the FID and CLIP Scores. However, we observe" + } + ] + } + ], + "index": 12 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "spans": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "type": "text", + "content": "8" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 166, + 91, + 222, + 100 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 166, + 91, + 222, + 100 + ], + "spans": [ + { + "bbox": [ + 166, + 91, + 222, + 100 + ], + "type": "text", + "content": "J. Chen et al." + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 7 + }, + { + "para_blocks": [ + { + "bbox": [ + 130, + 116, + 482, + 224 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 116, + 482, + 224 + ], + "spans": [ + { + "bbox": [ + 130, + 116, + 482, + 224 + ], + "type": "text", + "content": "evaluations conducted on the MSCOCO dataset may not adequately reflect a model's capabilities in aesthetics and text-image alignment. Therefore, we propose a curated set comprising 30,000 high-quality, aesthetically pleasing text-image pairs to facilitate the assessment. The selected samples of the dataset are presented in the appendix. This dataset is designed to provide a more comprehensive evaluation of a model's performance, particularly in capturing the intricacies of aesthetic appeal and the fidelity of alignment between textual descriptions and visual content. Unless otherwise specified, the evaluation experiments in the paper are conducted in the collected High-Quality Evaluation Dataset." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 132, + 248, + 263, + 261 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 248, + 263, + 261 + ], + "spans": [ + { + "bbox": [ + 132, + 248, + 263, + 261 + ], + "type": "text", + "content": "3.2 Efficient DiT Design" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 130, + 275, + 482, + 382 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 275, + 482, + 382 + ], + "spans": [ + { + "bbox": [ + 130, + 275, + 482, + 382 + ], + "type": "text", + "content": "An efficient DiT network is essential since the computational demand significantly increases when generating images at ultra-high resolutions. The attention mechanism plays a pivotal role in the efficacy of Diffusion Transformers, yet its quadratic computational demands significantly limit model scalability, particularly at higher resolutions e.g., 2K and 4K. Inspired by PVT v2 [46], we incorporate KV compression within the original PixART-" + }, + { + "bbox": [ + 130, + 275, + 482, + 382 + ], + "type": "inline_equation", + "content": "\\alpha" + }, + { + "bbox": [ + 130, + 275, + 482, + 382 + ], + "type": "text", + "content": "s framework to address the computational challenges. This design adds a mere " + }, + { + "bbox": [ + 130, + 275, + 482, + 382 + ], + "type": "inline_equation", + "content": "0.018\\%" + }, + { + "bbox": [ + 130, + 275, + 482, + 382 + ], + "type": "text", + "content": " to the total parameters yet achieves efficient reduction in computational costs via token compression, while still preserving both spatial and semantic information." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 130, + 384, + 482, + 479 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 384, + 482, + 479 + ], + "spans": [ + { + "bbox": [ + 130, + 384, + 482, + 479 + ], + "type": "text", + "content": "Key-Value (KV) Token Compression. Our motivation stems from the intriguing observation that applying key-value (KV) token compression directly to the pre-trained " + }, + { + "bbox": [ + 130, + 384, + 482, + 479 + ], + "type": "inline_equation", + "content": "\\mathrm{PIXART - \\alpha}" + }, + { + "bbox": [ + 130, + 384, + 482, + 479 + ], + "type": "text", + "content": " can still generate reasonable images. This suggests a redundancy in the features. Considering the high similarity within adjacent " + }, + { + "bbox": [ + 130, + 384, + 482, + 479 + ], + "type": "inline_equation", + "content": "R\\times R" + }, + { + "bbox": [ + 130, + 384, + 482, + 479 + ], + "type": "text", + "content": " patches, we assume that feature semantics within a window are redundant and can be compressed reasonably. We propose KV token compression, which is denoted as " + }, + { + "bbox": [ + 130, + 384, + 482, + 479 + ], + "type": "inline_equation", + "content": "f_{c}(\\cdot)" + }, + { + "bbox": [ + 130, + 384, + 482, + 479 + ], + "type": "text", + "content": ", to compress token features within a " + }, + { + "bbox": [ + 130, + 384, + 482, + 479 + ], + "type": "inline_equation", + "content": "R\\times R" + }, + { + "bbox": [ + 130, + 384, + 482, + 479 + ], + "type": "text", + "content": " window through a compression operator, as depicted in Fig. 7." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 130, + 481, + 302, + 666 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 481, + 302, + 666 + ], + "spans": [ + { + "bbox": [ + 130, + 481, + 302, + 666 + ], + "type": "text", + "content": "Furthermore, to mitigate the potential information loss caused by KV compression in self-attention computation, we opt to retain all the tokens of queries (Q). This strategy allows us to utilize KV compression effectively while mitigating the risk of losing crucial information. By employing KV compression, we enhance the efficiency of attention computations and reduce the computation complexity from " + }, + { + "bbox": [ + 130, + 481, + 302, + 666 + ], + "type": "inline_equation", + "content": "O(N^2)" + }, + { + "bbox": [ + 130, + 481, + 302, + 666 + ], + "type": "text", + "content": " to " + }, + { + "bbox": [ + 130, + 481, + 302, + 666 + ], + "type": "inline_equation", + "content": "O\\left(\\frac{N^2}{R^2}\\right)" + }, + { + "bbox": [ + 130, + 481, + 302, + 666 + ], + "type": "text", + "content": ", thereby making the computational cost of directly generating high-resolution images manageable." + } + ] + } + ], + "index": 6 + }, + { + "type": "image", + "bbox": [ + 309, + 506, + 482, + 601 + ], + "blocks": [ + { + "bbox": [ + 309, + 506, + 482, + 601 + ], + "lines": [ + { + "bbox": [ + 309, + 506, + 482, + 601 + ], + "spans": [ + { + "bbox": [ + 309, + 506, + 482, + 601 + ], + "type": "image", + "image_path": "fb5547b66eff4ce085940a61aa3d0b749e09e7cae296ff5305df0ec9bbb92298.jpg" + } + ] + } + ], + "index": 7, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 307, + 609, + 482, + 643 + ], + "lines": [ + { + "bbox": [ + 307, + 609, + 482, + 643 + ], + "spans": [ + { + "bbox": [ + 307, + 609, + 482, + 643 + ], + "type": "text", + "content": "Fig. 7: Design of KV Token Compression. We merge KV tokens in spatial space to reduce the computation complexity." + } + ] + } + ], + "index": 8, + "angle": 0, + "type": "image_caption" + } + ], + "index": 7 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 172, + 90, + 447, + 102 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 172, + 90, + 447, + 102 + ], + "spans": [ + { + "bbox": [ + 172, + 90, + 447, + 102 + ], + "type": "text", + "content": "PIXART- " + }, + { + "bbox": [ + 172, + 90, + 447, + 102 + ], + "type": "inline_equation", + "content": "\\Sigma" + }, + { + "bbox": [ + 172, + 90, + 447, + 102 + ], + "type": "text", + "content": " : Weak-to-Strong Training of DiT for 4K T2I Generation" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 474, + 91, + 480, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 474, + 91, + 480, + 100 + ], + "spans": [ + { + "bbox": [ + 474, + 91, + 480, + 100 + ], + "type": "text", + "content": "9" + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 8 + }, + { + "para_blocks": [ + { + "type": "table", + "bbox": [ + 133, + 106, + 478, + 255 + ], + "blocks": [ + { + "bbox": [ + 133, + 106, + 478, + 255 + ], + "lines": [ + { + "bbox": [ + 133, + 106, + 478, + 255 + ], + "spans": [ + { + "bbox": [ + 133, + 106, + 478, + 255 + ], + "type": "table", + "html": "
StrategiesVAE Change512 →1024 RA512 →1024 RA + PE Interp.RA + PE Adjust + KV CompressRA + PE Interp. + KV Compress + Conv Avg Init
0 step
Fast \nAdapt2K step100 step100 step100 step100 step
", + "image_path": "de22726a3d6bc016081655eec88453d29fc0f3f7ac466701f877ca0615df3dc7.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "table_body" + } + ], + "index": 2 + }, + { + "bbox": [ + 132, + 264, + 480, + 297 + ], + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 264, + 480, + 297 + ], + "spans": [ + { + "bbox": [ + 132, + 264, + 480, + 297 + ], + "type": "text", + "content": "Fig. 8: This illustration demonstrates how our training strategy designs expedite the model's convergence during the transition to VAEs, adjustment to higher resolutions, and the KV compression process, facilitating rapid learning from weak to strong." + } + ] + } + ], + "index": 6, + "type": "text" + }, + { + "bbox": [ + 193, + 307, + 480, + 335 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 193, + 307, + 480, + 335 + ], + "spans": [ + { + "bbox": [ + 193, + 307, + 480, + 335 + ], + "type": "interline_equation", + "content": "\\operatorname {A t t e n t i o n} (Q, K, V) = \\operatorname {s o f t m a x} \\left(\\frac {Q \\cdot f _ {c} (K) ^ {T}}{\\sqrt {d _ {k}}}\\right) f _ {c} (V) \\tag {1}", + "image_path": "d5304977d601bd83de71dee1fff784f62bbd7a44f8dc92c01209ede6412cde78.jpg" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 130, + 338, + 480, + 422 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 338, + 480, + 422 + ], + "spans": [ + { + "bbox": [ + 130, + 338, + 480, + 422 + ], + "type": "text", + "content": "We compress deep layers using the convolution operator \"Conv2×2\" with specific initialization. Detailed experiments on other design variants are discussed in Sec. 5. Specifically, we design a specialized convolution kernel initialization \"Conv Avg Init\" that utilizes group convolution and initializes the weights " + }, + { + "bbox": [ + 130, + 338, + 480, + 422 + ], + "type": "inline_equation", + "content": "w = \\frac{1}{R^2}" + }, + { + "bbox": [ + 130, + 338, + 480, + 422 + ], + "type": "text", + "content": ", equivalent to an average operator. This initialization strategy can initially produce coarse results, accelerating the fine-tuning process while only introducing 0.018% additional parameters." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 132, + 441, + 332, + 453 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 441, + 332, + 453 + ], + "spans": [ + { + "bbox": [ + 132, + 441, + 332, + 453 + ], + "type": "text", + "content": "3.3 Weak-to-Strong Training Strategy" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 130, + 461, + 480, + 498 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 461, + 480, + 498 + ], + "spans": [ + { + "bbox": [ + 130, + 461, + 480, + 498 + ], + "type": "text", + "content": "We propose several efficient training strategies to enhance the transition from a \"weak\" model to a \"strong\" model. These strategies encompass VAE rapid adaptation, high-resolution fine-tuning, and KV Token compression." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 132, + 498, + 302, + 616 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 498, + 302, + 616 + ], + "spans": [ + { + "bbox": [ + 132, + 498, + 302, + 616 + ], + "type": "text", + "content": "Adapting model to new VAEs. As VAEs develop, training T2I models from scratch is resource-intensive. We replace PixART- " + }, + { + "bbox": [ + 132, + 498, + 302, + 616 + ], + "type": "inline_equation", + "content": "\\alpha" + }, + { + "bbox": [ + 132, + 498, + 302, + 616 + ], + "type": "text", + "content": " s VAE with SDXL's VAE and continue fine-tuning the diffusion model. We observe a rapid convergence phenomenon that fine-tuning quickly converges at 2K training steps as shown in Fig 8 (a). Fine-tuning is more efficient when dealing with VAE" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 307, + 498, + 481, + 555 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 498, + 481, + 555 + ], + "spans": [ + { + "bbox": [ + 307, + 498, + 481, + 555 + ], + "type": "text", + "content": "Table 2: We fine-tune a high-resolution model from a low-resolution model and observe that even fine-tuning for a relatively short duration, such as 1K steps, can still yield high-quality results." + } + ] + } + ], + "index": 12 + }, + { + "type": "table", + "bbox": [ + 325, + 559, + 458, + 610 + ], + "blocks": [ + { + "bbox": [ + 198, + 255, + 207, + 263 + ], + "lines": [ + { + "bbox": [ + 198, + 255, + 207, + 263 + ], + "spans": [ + { + "bbox": [ + 198, + 255, + 207, + 263 + ], + "type": "text", + "content": "(a)" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 286, + 255, + 296, + 263 + ], + "lines": [ + { + "bbox": [ + 286, + 255, + 296, + 263 + ], + "spans": [ + { + "bbox": [ + 286, + 255, + 296, + 263 + ], + "type": "text", + "content": "(b)" + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 411, + 256, + 421, + 263 + ], + "lines": [ + { + "bbox": [ + 411, + 256, + 421, + 263 + ], + "spans": [ + { + "bbox": [ + 411, + 256, + 421, + 263 + ], + "type": "text", + "content": "(c)" + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 325, + 559, + 458, + 610 + ], + "lines": [ + { + "bbox": [ + 325, + 559, + 458, + 610 + ], + "spans": [ + { + "bbox": [ + 325, + 559, + 458, + 610 + ], + "type": "table", + "html": "
ResolutionIterationsFID ↓CLIP ↑
25620K16.560.270
256 → 5121K9.750.272
256 → 512100K8.910.276
", + "image_path": "b70faae669c303f485345053405a2b65ba540daa1a7162668fe9d23847c1eb5d.jpg" + } + ] + } + ], + "index": 13, + "angle": 0, + "type": "table_body" + } + ], + "index": 13 + }, + { + "bbox": [ + 132, + 617, + 436, + 628 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 617, + 436, + 628 + ], + "spans": [ + { + "bbox": [ + 132, + 617, + 436, + 628 + ], + "type": "text", + "content": "model transferring and negates the necessity of training from scratch." + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 132, + 629, + 481, + 666 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 629, + 481, + 666 + ], + "spans": [ + { + "bbox": [ + 132, + 629, + 481, + 666 + ], + "type": "text", + "content": "Adapting to Higher-Resolution. When we fine-tune from a low-resolution (LR) model to a high-resolution (HR) model, we observe a performance degradation in Fig. 8 (b), which we attribute to discrepancies in positional embeddings" + } + ] + } + ], + "index": 15 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "spans": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "type": "text", + "content": "10" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 166, + 91, + 222, + 100 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 166, + 91, + 222, + 100 + ], + "spans": [ + { + "bbox": [ + 166, + 91, + 222, + 100 + ], + "type": "text", + "content": "J. Chen et al." + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 9 + }, + { + "para_blocks": [ + { + "bbox": [ + 130, + 115, + 482, + 224 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 115, + 482, + 224 + ], + "spans": [ + { + "bbox": [ + 130, + 115, + 482, + 224 + ], + "type": "text", + "content": "(PE) between different resolutions. To address this, we utilize the \"PE Interpolation\" trick [4, 49]: initializing the HR model's PE by interpolating the LR model's PE, significantly enhancing the HR model's initial status and expediting fine-tuning. We can obtain visually pleasing images even within only 100 training iterations. Besides, we quantitatively compare the model's performance as illustrated in Tab. 2. The fine-tuning quickly converges at 1K steps, and further training slightly improves the performance. This illustrates that using the \"PE Interpolation\" trick enables rapid convergence of higher resolution generation, obviating the need for training from scratch for generating at higher resolutions." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 130, + 224, + 482, + 308 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 224, + 482, + 308 + ], + "spans": [ + { + "bbox": [ + 130, + 224, + 482, + 308 + ], + "type": "text", + "content": "Adapting model to KV compression. We can use KV compression directly when fine-tuning from LR pre-trained models without KV compression. As shown in Fig. 8 (c), with our \"Conv Avg Init.\" strategy, " + }, + { + "bbox": [ + 130, + 224, + 482, + 308 + ], + "type": "inline_equation", + "content": "\\mathrm{PIXART - \\Sigma}" + }, + { + "bbox": [ + 130, + 224, + 482, + 308 + ], + "type": "text", + "content": " starts from a better initial state, making converging easier and faster. Notably, " + }, + { + "bbox": [ + 130, + 224, + 482, + 308 + ], + "type": "inline_equation", + "content": "\\mathrm{PIXART - \\Sigma}" + }, + { + "bbox": [ + 130, + 224, + 482, + 308 + ], + "type": "text", + "content": " performs satisfied visual results even within 100 training steps. Finally, through the KV compression operators and compression layers design in Sec 3.2, we can reduce " + }, + { + "bbox": [ + 130, + 224, + 482, + 308 + ], + "type": "inline_equation", + "content": "\\sim 34\\%" + }, + { + "bbox": [ + 130, + 224, + 482, + 308 + ], + "type": "text", + "content": " of the training and inference time." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 132, + 323, + 224, + 336 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 323, + 224, + 336 + ], + "spans": [ + { + "bbox": [ + 132, + 323, + 224, + 336 + ], + "type": "text", + "content": "4 Experiment" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 132, + 343, + 279, + 356 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 343, + 279, + 356 + ], + "spans": [ + { + "bbox": [ + 132, + 343, + 279, + 356 + ], + "type": "text", + "content": "4.1 Implementation Details" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 130, + 360, + 482, + 503 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 360, + 482, + 503 + ], + "spans": [ + { + "bbox": [ + 130, + 360, + 482, + 503 + ], + "type": "text", + "content": "Training Details. We follow Imagen [40] and " + }, + { + "bbox": [ + 130, + 360, + 482, + 503 + ], + "type": "inline_equation", + "content": "\\mathrm{PIXART - \\alpha}" + }, + { + "bbox": [ + 130, + 360, + 482, + 503 + ], + "type": "text", + "content": " [4] to employ the T5 [10]'s encoder (i.e., Flan-T5-XXL) as the text encoder for conditional feature extraction, and use " + }, + { + "bbox": [ + 130, + 360, + 482, + 503 + ], + "type": "inline_equation", + "content": "\\mathrm{PIXART - \\alpha}" + }, + { + "bbox": [ + 130, + 360, + 482, + 503 + ], + "type": "text", + "content": " [4] as our base diffusion model. Unlike most works that extract fixed 77 text tokens, we adjust the length of text tokens from " + }, + { + "bbox": [ + 130, + 360, + 482, + 503 + ], + "type": "inline_equation", + "content": "\\mathrm{PIXART - \\alpha}" + }, + { + "bbox": [ + 130, + 360, + 482, + 503 + ], + "type": "text", + "content": "'s 120 to 300, as the caption curated in Internal-" + }, + { + "bbox": [ + 130, + 360, + 482, + 503 + ], + "type": "inline_equation", + "content": "\\Sigma" + }, + { + "bbox": [ + 130, + 360, + 482, + 503 + ], + "type": "text", + "content": " is much denser to provide highly fine-grained details. To capture the latent features of input images, we employ a pre-trained and frozen VAE from SDXL [36]. Other implementation details are the same as " + }, + { + "bbox": [ + 130, + 360, + 482, + 503 + ], + "type": "inline_equation", + "content": "\\mathrm{PIXART - \\alpha}" + }, + { + "bbox": [ + 130, + 360, + 482, + 503 + ], + "type": "text", + "content": ". Models are finetuned on the " + }, + { + "bbox": [ + 130, + 360, + 482, + 503 + ], + "type": "inline_equation", + "content": "\\mathrm{PIXART - \\alpha}" + }, + { + "bbox": [ + 130, + 360, + 482, + 503 + ], + "type": "text", + "content": "'s 256px pre-trained checkpoint with the position embedding interpolation trick [4]. Our final models, including 1K resolution, are trained on 32 V100 GPUs. We additionally use 16 A100 GPUs to train the 2K and 4K image generation models. For further information, please refer to the appendix." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 130, + 503, + 482, + 552 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 503, + 482, + 552 + ], + "spans": [ + { + "bbox": [ + 130, + 503, + 482, + 552 + ], + "type": "text", + "content": "Note that we use CAME optimizer [27] with a weight decay of 0 and a constant learning rate of 2e-5, instead of the regular AdamW [24] optimizer. This helps us reduce the dimension of the optimizer's state, leading to lower GPU memory without performance degradation." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 130, + 552, + 482, + 612 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 552, + 482, + 612 + ], + "spans": [ + { + "bbox": [ + 130, + 552, + 482, + 612 + ], + "type": "text", + "content": "Evaluation Metrics. To better illustrate aesthetics and semantic ability, we collect 30K high-quality text-image pairs (as mentioned in Sec. 3.1) to benchmark T2I models. We mainly evaluate " + }, + { + "bbox": [ + 130, + 552, + 482, + 612 + ], + "type": "inline_equation", + "content": "\\mathsf{PIXART - \\Sigma}" + }, + { + "bbox": [ + 130, + 552, + 482, + 612 + ], + "type": "text", + "content": " via human and AI preference study since FID [39] may not adequately reflect the generation quality. However, we still provide the FID results on the collected dataset in the appendix." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 132, + 625, + 293, + 637 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 625, + 293, + 637 + ], + "spans": [ + { + "bbox": [ + 132, + 625, + 293, + 637 + ], + "type": "text", + "content": "4.2 Performance Comparisons" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 130, + 641, + 481, + 666 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 641, + 481, + 666 + ], + "spans": [ + { + "bbox": [ + 130, + 641, + 481, + 666 + ], + "type": "text", + "content": "Image Quality Assessment. We qualitatively evaluated our methodology against both closed-source text-to-image products and open-source models. As" + } + ] + } + ], + "index": 10 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 172, + 90, + 447, + 102 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 172, + 90, + 447, + 102 + ], + "spans": [ + { + "bbox": [ + 172, + 90, + 447, + 102 + ], + "type": "text", + "content": "PIXART- " + }, + { + "bbox": [ + 172, + 90, + 447, + 102 + ], + "type": "inline_equation", + "content": "\\Sigma" + }, + { + "bbox": [ + 172, + 90, + 447, + 102 + ], + "type": "text", + "content": " : Weak-to-Strong Training of DiT for 4K T2I Generation" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 470, + 91, + 479, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 470, + 91, + 479, + 100 + ], + "spans": [ + { + "bbox": [ + 470, + 91, + 479, + 100 + ], + "type": "text", + "content": "11" + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 10 + }, + { + "para_blocks": [ + { + "bbox": [ + 130, + 116, + 482, + 201 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 116, + 482, + 201 + ], + "spans": [ + { + "bbox": [ + 130, + 116, + 482, + 201 + ], + "type": "text", + "content": "illustrated in Fig. 1, our model can produce high-quality, photo-realistic images with intricate details over diverse aspect ratios and styles. This capability underscores the superior performance of our approach in generating visually compelling content from textual descriptions. In Fig. 3, we compare " + }, + { + "bbox": [ + 130, + 116, + 482, + 201 + ], + "type": "inline_equation", + "content": "\\mathsf{PIXART - \\Sigma}" + }, + { + "bbox": [ + 130, + 116, + 482, + 201 + ], + "type": "text", + "content": " with open-source models SDXL [36] and " + }, + { + "bbox": [ + 130, + 116, + 482, + 201 + ], + "type": "inline_equation", + "content": "\\mathsf{PIXART - \\alpha}" + }, + { + "bbox": [ + 130, + 116, + 482, + 201 + ], + "type": "text", + "content": " [4], our method enhances the realism of portraits and boosts the capacity for semantic analysis. In contrast to SDXL, our approach demonstrates superior proficiency in following instructions." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 130, + 202, + 482, + 251 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 202, + 482, + 251 + ], + "spans": [ + { + "bbox": [ + 130, + 202, + 482, + 251 + ], + "type": "text", + "content": "Not only superior to open-source models, but our method is also very competitive with current T2I closed-source products, as depicted in Fig. 4. PixART- " + }, + { + "bbox": [ + 130, + 202, + 482, + 251 + ], + "type": "inline_equation", + "content": "\\Sigma" + }, + { + "bbox": [ + 130, + 202, + 482, + 251 + ], + "type": "text", + "content": " produces photo-realistic results and adheres closely to user instructions, which is on par with contemporary commercial products." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 130, + 253, + 264, + 396 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 253, + 264, + 396 + ], + "spans": [ + { + "bbox": [ + 130, + 253, + 264, + 396 + ], + "type": "text", + "content": "High-resolution Generation. Our method is capable of directly generating 4K resolution images without the need for any post-processing. Additionally, it excels in accurately following complex, detailed, and long text provided by users, as demonstrated in Fig. 2. Thus, users do not need prompt engineering to achieve satisfactory results." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 130, + 399, + 264, + 435 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 399, + 264, + 435 + ], + "spans": [ + { + "bbox": [ + 130, + 399, + 264, + 435 + ], + "type": "text", + "content": "Our approach enables direct 4K image generation. In parallel, studies [11, 16] have" + } + ] + } + ], + "index": 5 + }, + { + "type": "image", + "bbox": [ + 291, + 258, + 459, + 376 + ], + "blocks": [ + { + "bbox": [ + 291, + 258, + 459, + 376 + ], + "lines": [ + { + "bbox": [ + 291, + 258, + 459, + 376 + ], + "spans": [ + { + "bbox": [ + 291, + 258, + 459, + 376 + ], + "type": "image", + "image_path": "8537c6c5f4e0fabc7b967cfdf7be066099cb65f6eb850b73531d752d75e72409.jpg" + } + ] + } + ], + "index": 6, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 269, + 378, + 482, + 433 + ], + "lines": [ + { + "bbox": [ + 269, + 378, + 482, + 433 + ], + "spans": [ + { + "bbox": [ + 269, + 378, + 482, + 433 + ], + "type": "text", + "content": "Fig. 9: Comparisons on FID and Clip-Score with Open-sourced T2I Models. PixArt- " + }, + { + "bbox": [ + 269, + 378, + 482, + 433 + ], + "type": "inline_equation", + "content": "\\Sigma" + }, + { + "bbox": [ + 269, + 378, + 482, + 433 + ], + "type": "text", + "content": " demonstrates enhanced performance in terms of FID and Clip-Score on the curated 30K High-Quality Evaluation Dataset." + } + ] + } + ], + "index": 7, + "angle": 0, + "type": "image_caption" + } + ], + "index": 6 + }, + { + "bbox": [ + 130, + 435, + 482, + 532 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 435, + 482, + 532 + ], + "spans": [ + { + "bbox": [ + 130, + 435, + 482, + 532 + ], + "type": "text", + "content": "introduced tuning-free post-processing techniques aimed at generating HR images from LR models or employing super-resolution models [50] to produce HR images. However, their corresponding results often exhibit artifacts for two primary reasons: (1) Accumulative error may arise due to the cascade pipeline. (2) These methods do not capture the true distribution of 4K images nor learn the alignment between text and 4K images. We argue that our method might be a more promising way to generate high-resolution images. Our method yields superior results, and more visual comparison is included in the supplement." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 130, + 533, + 482, + 666 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 533, + 482, + 666 + ], + "spans": [ + { + "bbox": [ + 130, + 533, + 482, + 666 + ], + "type": "text", + "content": "Human/AI (GPT4V) Preference Study. We evaluate the well-trained model in both the human and AI preference study using a subset of 300 captions randomly collected from the High-Quality Evaluation Dataset mentioned in Sec. 3.1. We collect images generated by overall six open-source models, including PixART-" + }, + { + "bbox": [ + 130, + 533, + 482, + 666 + ], + "type": "inline_equation", + "content": "\\alpha" + }, + { + "bbox": [ + 130, + 533, + 482, + 666 + ], + "type": "text", + "content": ", PixART-" + }, + { + "bbox": [ + 130, + 533, + 482, + 666 + ], + "type": "inline_equation", + "content": "\\Sigma" + }, + { + "bbox": [ + 130, + 533, + 482, + 666 + ], + "type": "text", + "content": ", SD1.5 [39], Stable Turbo [41], Stable XL [36], Stable Cascade [35] and Playground-V2.0 [20]. We develop a website for the human preference study to display the prompts and their corresponding images. This website was distributed to trained evaluators, who were asked to assess the images, ranking them according to quality and how well they matched the text prompts. The results, illustrated by the blue bar in Fig. 9, indicate a marked preference for PixART-" + }, + { + "bbox": [ + 130, + 533, + 482, + 666 + ], + "type": "inline_equation", + "content": "\\Sigma" + }, + { + "bbox": [ + 130, + 533, + 482, + 666 + ], + "type": "text", + "content": " over the other six T2I generators. PixART-" + }, + { + "bbox": [ + 130, + 533, + 482, + 666 + ], + "type": "inline_equation", + "content": "\\Sigma" + }, + { + "bbox": [ + 130, + 533, + 482, + 666 + ], + "type": "text", + "content": " genera" + } + ] + } + ], + "index": 9 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "spans": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "type": "text", + "content": "12" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 166, + 91, + 222, + 100 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 166, + 91, + 222, + 100 + ], + "spans": [ + { + "bbox": [ + 166, + 91, + 222, + 100 + ], + "type": "text", + "content": "J. Chen et al." + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 11 + }, + { + "para_blocks": [ + { + "type": "table", + "bbox": [ + 172, + 114, + 287, + 167 + ], + "blocks": [ + { + "bbox": [ + 172, + 114, + 287, + 167 + ], + "lines": [ + { + "bbox": [ + 172, + 114, + 287, + 167 + ], + "spans": [ + { + "bbox": [ + 172, + 114, + 287, + 167 + ], + "type": "table", + "html": "
LayersFID ↓CLIP-Score ↑
N/A8.2440.276
Shallow (1-14)9.2780.275
Middle (7-20)9.0630.276
Deep (14-27)8.5320.275
", + "image_path": "0715374b530774f0d28375ee990b978a13468839c19bff4b754e0b5b3f3f9365.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "table_body" + } + ], + "index": 2 + }, + { + "type": "table", + "bbox": [ + 152, + 186, + 299, + 264 + ], + "blocks": [ + { + "bbox": [ + 187, + 168, + 271, + 177 + ], + "lines": [ + { + "bbox": [ + 187, + 168, + 271, + 177 + ], + "spans": [ + { + "bbox": [ + 187, + 168, + 271, + 177 + ], + "type": "text", + "content": "(a) Compression layers." + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 152, + 186, + 299, + 264 + ], + "lines": [ + { + "bbox": [ + 152, + 186, + 299, + 264 + ], + "spans": [ + { + "bbox": [ + 152, + 186, + 299, + 264 + ], + "type": "table", + "html": "
Res.RatioFID ↓CLIP-Score ↑Train Latency ↓
51218.2440.2762.3
51229.0630.2762.2 (-4%)
51249.6060.2762.1 (-9%)
102415.6850.27727.5
102425.5120.27322.5 (-18%)
102445.6440.27620.0 (-27%)
102495.7120.27517.8 (-35%)
", + "image_path": "a3c63d2a9a9015a022e2fdb560f796d1150d3f3d862e490c5ca0ceb4254ece5a.jpg" + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "table_body" + } + ], + "index": 4 + }, + { + "type": "table", + "bbox": [ + 321, + 114, + 441, + 167 + ], + "blocks": [ + { + "bbox": [ + 141, + 265, + 309, + 274 + ], + "lines": [ + { + "bbox": [ + 141, + 265, + 309, + 274 + ], + "spans": [ + { + "bbox": [ + 141, + 265, + 309, + 274 + ], + "type": "text", + "content": "(c) Compression rations on different resolutions." + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 321, + 114, + 441, + 167 + ], + "lines": [ + { + "bbox": [ + 321, + 114, + 441, + 167 + ], + "spans": [ + { + "bbox": [ + 321, + 114, + 441, + 167 + ], + "type": "table", + "html": "
OperatorFID ↓CLIP-Score ↑
N/A8.2440.276
Token Discarding8.9180.275
Token Pooling9.4150.275
Conv2×28.5050.274
", + "image_path": "52d8bf0029703224b190897cef333a295345da29422a5f6e5de41a4ac8119ae7.jpg" + } + ] + } + ], + "index": 6, + "angle": 0, + "type": "table_body" + } + ], + "index": 6 + }, + { + "type": "table", + "bbox": [ + 321, + 193, + 459, + 258 + ], + "blocks": [ + { + "bbox": [ + 331, + 167, + 430, + 176 + ], + "lines": [ + { + "bbox": [ + 331, + 167, + 430, + 176 + ], + "spans": [ + { + "bbox": [ + 331, + 167, + 430, + 176 + ], + "type": "text", + "content": "(b) Compression operators." + } + ] + } + ], + "index": 7, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 321, + 193, + 459, + 258 + ], + "lines": [ + { + "bbox": [ + 321, + 193, + 459, + 258 + ], + "spans": [ + { + "bbox": [ + 321, + 193, + 459, + 258 + ], + "type": "table", + "html": "
Res.RatioTrain Latency (s/Iter@32BS)↓ Test Latency ↓ (s/Img)
2K15658
2K437 (-34%)38 (-34%)
4K119191
4K4125 (-35%)60 (-34%)
", + "image_path": "103d6ecf87e07692ade7f7b6476f8be22c960876e7e52ba091e2e0bcebd9284d.jpg" + } + ] + } + ], + "index": 8, + "angle": 0, + "type": "table_body" + } + ], + "index": 8 + }, + { + "bbox": [ + 329, + 258, + 449, + 267 + ], + "lines": [ + { + "bbox": [ + 329, + 258, + 449, + 267 + ], + "spans": [ + { + "bbox": [ + 329, + 258, + 449, + 267 + ], + "type": "text", + "content": "(d) Speed of different resolutions." + } + ] + } + ], + "index": 9, + "angle": 0, + "type": "text" + }, + { + "bbox": [ + 130, + 277, + 480, + 322 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 277, + 480, + 322 + ], + "spans": [ + { + "bbox": [ + 130, + 277, + 480, + 322 + ], + "type": "text", + "content": "Table 3: KV-Token Compression Settings in Image Generation. This study employs FID, CMMD, and CLIP-Score metrics to assess the impact of various token compression components, such as compression ratio, positions, operators, and varying resolutions. Speed calculation in Tab. 3c is Second/Iteration/384 Batch-size." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 130, + 330, + 480, + 366 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 330, + 480, + 366 + ], + "spans": [ + { + "bbox": [ + 130, + 330, + 480, + 366 + ], + "type": "text", + "content": "ates superior high-quality images that closely follow user prompts, using a much smaller size (0.6B parameters) compared to existing T2I diffusion models like SDXL (2.6B parameters) and SD Cascade (5.1B parameters)." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 130, + 367, + 481, + 476 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 367, + 481, + 476 + ], + "spans": [ + { + "bbox": [ + 130, + 367, + 481, + 476 + ], + "type": "text", + "content": "For AI preference study, we employ the advanced multimodal model, GPT-4 V [32], as the evaluator. For each trial, we supply GPT-4V with two images: one from " + }, + { + "bbox": [ + 130, + 367, + 481, + 476 + ], + "type": "inline_equation", + "content": "\\mathrm{PIXART - \\Sigma}" + }, + { + "bbox": [ + 130, + 367, + 481, + 476 + ], + "type": "text", + "content": " and another from a competing T2I model. We craft distinct prompts guiding GPT-4V to vote based on image quality and image-and-text alignment. The results, represented by orange and green bars in Fig. 9, demonstrate consistent outcomes in both human and AI preference studies. Specifically, " + }, + { + "bbox": [ + 130, + 367, + 481, + 476 + ], + "type": "inline_equation", + "content": "\\mathrm{PIXART - \\Sigma}" + }, + { + "bbox": [ + 130, + 367, + 481, + 476 + ], + "type": "text", + "content": " surpasses the baseline, " + }, + { + "bbox": [ + 130, + 367, + 481, + 476 + ], + "type": "inline_equation", + "content": "\\mathrm{PIXART - \\alpha}" + }, + { + "bbox": [ + 130, + 367, + 481, + 476 + ], + "type": "text", + "content": ", in effectiveness. Compared to advanced models such as Stable Cascaded, " + }, + { + "bbox": [ + 130, + 367, + 481, + 476 + ], + "type": "inline_equation", + "content": "\\mathrm{PIXART - \\Sigma}" + }, + { + "bbox": [ + 130, + 367, + 481, + 476 + ], + "type": "text", + "content": " exhibits competitive or superior performance on image quality and instruction-following abilities." + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 132, + 500, + 254, + 511 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 500, + 254, + 511 + ], + "spans": [ + { + "bbox": [ + 132, + 500, + 254, + 511 + ], + "type": "text", + "content": "5 Ablation Studies" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 130, + 531, + 480, + 567 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 531, + 480, + 567 + ], + "spans": [ + { + "bbox": [ + 130, + 531, + 480, + 567 + ], + "type": "text", + "content": "We conduct ablation studies on generation performance on various KV compression designs. Unless specified, the experiments are conducted on 512px generation. The detailed settings of each ablation trial are included in the appendix." + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 132, + 591, + 271, + 604 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 591, + 271, + 604 + ], + "spans": [ + { + "bbox": [ + 132, + 591, + 271, + 604 + ], + "type": "text", + "content": "5.1 Experimental settings" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 130, + 617, + 480, + 666 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 617, + 480, + 666 + ], + "spans": [ + { + "bbox": [ + 130, + 617, + 480, + 666 + ], + "type": "text", + "content": "We use the test set described in Sec. 3.1 for evaluation. We employ FID to compute the distributional difference between the collected and generated data for comparative metrics. Furthermore, we utilize CLIP-Score to assess the alignment between prompts and the generated images." + } + ] + } + ], + "index": 16 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 172, + 90, + 447, + 102 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 172, + 90, + 447, + 102 + ], + "spans": [ + { + "bbox": [ + 172, + 90, + 447, + 102 + ], + "type": "text", + "content": "PIXART- " + }, + { + "bbox": [ + 172, + 90, + 447, + 102 + ], + "type": "inline_equation", + "content": "\\Sigma" + }, + { + "bbox": [ + 172, + 90, + 447, + 102 + ], + "type": "text", + "content": " : Weak-to-Strong Training of DiT for 4K T2I Generation" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 470, + 91, + 480, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 470, + 91, + 480, + 100 + ], + "spans": [ + { + "bbox": [ + 470, + 91, + 480, + 100 + ], + "type": "text", + "content": "13" + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 12 + }, + { + "para_blocks": [ + { + "bbox": [ + 132, + 116, + 268, + 129 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 116, + 268, + 129 + ], + "spans": [ + { + "bbox": [ + 132, + 116, + 268, + 129 + ], + "type": "text", + "content": "5.2 Compression Designs" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 130, + 134, + 482, + 242 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 134, + 482, + 242 + ], + "spans": [ + { + "bbox": [ + 130, + 134, + 482, + 242 + ], + "type": "text", + "content": "Compression positions. We implemented KV compression at different depths within the Transformer structure: in the shallow layers (1~14), the intermediate layers (7~20), and the deep layers (14~27). As indicated in Tab. 3a, employing KV compression on deep layers notably achieves superior performance. We speculate this is because shallow layers typically encode detailed texture content, while deep layers abstract high-level semantic content. Since compression affects image quality rather than semantic information, compressing deep layers achieves the least information loss, making it practical for accelerating training without compromising generation quality." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 130, + 242, + 482, + 312 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 242, + 482, + 312 + ], + "spans": [ + { + "bbox": [ + 130, + 242, + 482, + 312 + ], + "type": "text", + "content": "Compression operators. We explored the impact of different compression operators. We employed three techniques, random discarding, average pooling, and parametric convolution, to compress " + }, + { + "bbox": [ + 130, + 242, + 482, + 312 + ], + "type": "inline_equation", + "content": "2 \\times 2" + }, + { + "bbox": [ + 130, + 242, + 482, + 312 + ], + "type": "text", + "content": " tokens into a single token. As illustrated in Table 3b, the \"Conv " + }, + { + "bbox": [ + 130, + 242, + 482, + 312 + ], + "type": "inline_equation", + "content": "2 \\times 2" + }, + { + "bbox": [ + 130, + 242, + 482, + 312 + ], + "type": "text", + "content": "\" method outperforms the others, underscoring the advantage of using a learnable kernel to more effectively reduce redundant features than simple discarding methods." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 130, + 314, + 482, + 409 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 314, + 482, + 409 + ], + "spans": [ + { + "bbox": [ + 130, + 314, + 482, + 409 + ], + "type": "text", + "content": "Compression ratios on different resolutions. We investigated the influence of varying compression ratios on different resolutions. As shown in Tab. 3c, remarkably, we find that token compression does not affect the alignment between textual and generated images (CLIP Score) but influences the image quality (FID) across resolutions. Although there is a slight degradation in image quality with increasing compression ratios, our strategy brings a training speedup of " + }, + { + "bbox": [ + 130, + 314, + 482, + 409 + ], + "type": "inline_equation", + "content": "18\\%" + }, + { + "bbox": [ + 130, + 314, + 482, + 409 + ], + "type": "text", + "content": " to " + }, + { + "bbox": [ + 130, + 314, + 482, + 409 + ], + "type": "inline_equation", + "content": "35\\%" + }, + { + "bbox": [ + 130, + 314, + 482, + 409 + ], + "type": "text", + "content": ". This suggests that our proposed KV compression is both effective and efficient for achieving high-resolution T2I generation." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 130, + 410, + 482, + 506 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 410, + 482, + 506 + ], + "spans": [ + { + "bbox": [ + 130, + 410, + 482, + 506 + ], + "type": "text", + "content": "Speed comparisons on different resolutions. We further comprehensively validate the speed acceleration in both training and inference in Tab. 3d. By 'latency', we refer to the duration required for inference on a single sample. Our method speeds up training and inference by about " + }, + { + "bbox": [ + 130, + 410, + 482, + 506 + ], + "type": "inline_equation", + "content": "35\\%" + }, + { + "bbox": [ + 130, + 410, + 482, + 506 + ], + "type": "text", + "content": " in the 4K generation. Notably, training acceleration increases with resolution, accelerating from " + }, + { + "bbox": [ + 130, + 410, + 482, + 506 + ], + "type": "inline_equation", + "content": "18\\%" + }, + { + "bbox": [ + 130, + 410, + 482, + 506 + ], + "type": "text", + "content": " to " + }, + { + "bbox": [ + 130, + 410, + 482, + 506 + ], + "type": "inline_equation", + "content": "35\\%" + }, + { + "bbox": [ + 130, + 410, + 482, + 506 + ], + "type": "text", + "content": " as the resolution increases from 1K to 4K. This indicates the effectiveness of our method with increasing resolution, demonstrating its potential applicability to even higher-resolution image generation tasks." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 132, + 522, + 220, + 534 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 522, + 220, + 534 + ], + "spans": [ + { + "bbox": [ + 132, + 522, + 220, + 534 + ], + "type": "text", + "content": "6 Conclusion" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 130, + 545, + 482, + 666 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 545, + 482, + 666 + ], + "spans": [ + { + "bbox": [ + 130, + 545, + 482, + 666 + ], + "type": "text", + "content": "In this paper, we introduce " + }, + { + "bbox": [ + 130, + 545, + 482, + 666 + ], + "type": "inline_equation", + "content": "\\mathrm{PIXART - \\Sigma}" + }, + { + "bbox": [ + 130, + 545, + 482, + 666 + ], + "type": "text", + "content": ", a Text-to-Image (T2I) diffusion model capable of directly generating high-quality images at 4K resolution. Building upon the pre-trained foundation of " + }, + { + "bbox": [ + 130, + 545, + 482, + 666 + ], + "type": "inline_equation", + "content": "\\mathrm{PIXART - \\alpha}" + }, + { + "bbox": [ + 130, + 545, + 482, + 666 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 130, + 545, + 482, + 666 + ], + "type": "inline_equation", + "content": "\\mathrm{PIXART - \\Sigma}" + }, + { + "bbox": [ + 130, + 545, + 482, + 666 + ], + "type": "text", + "content": " achieves efficient training through a novel \"weak-to-strong training\" methodology. This approach is characterized by the incorporation of higher-quality data and the integration of efficient token compression. " + }, + { + "bbox": [ + 130, + 545, + 482, + 666 + ], + "type": "inline_equation", + "content": "\\mathrm{PIXART - \\Sigma}" + }, + { + "bbox": [ + 130, + 545, + 482, + 666 + ], + "type": "text", + "content": " excels at producing high-fidelity images while adhering closely to textual prompts, surpassing the high standards set by its predecessor, " + }, + { + "bbox": [ + 130, + 545, + 482, + 666 + ], + "type": "inline_equation", + "content": "\\mathrm{PIXART - \\alpha}" + }, + { + "bbox": [ + 130, + 545, + 482, + 666 + ], + "type": "text", + "content": ". We believe that the innovations presented in " + }, + { + "bbox": [ + 130, + 545, + 482, + 666 + ], + "type": "inline_equation", + "content": "\\mathrm{PIXART - \\Sigma}" + }, + { + "bbox": [ + 130, + 545, + 482, + 666 + ], + "type": "text", + "content": " will not only contribute to advancements in the AIGC community but also pave the way for entities to access more efficient, and high-quality generative models." + } + ] + } + ], + "index": 8 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "spans": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "type": "text", + "content": "14" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 166, + 91, + 222, + 100 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 166, + 91, + 222, + 100 + ], + "spans": [ + { + "bbox": [ + 166, + 91, + 222, + 100 + ], + "type": "text", + "content": "J. Chen et al." + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 13 + }, + { + "para_blocks": [ + { + "bbox": [ + 132, + 116, + 487, + 140 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 116, + 487, + 140 + ], + "spans": [ + { + "bbox": [ + 132, + 116, + 487, + 140 + ], + "type": "text", + "content": "Acknowledgements. We would like to thank Zeqiang Lai, Fei Chen, and Shuchen Xue for discussing and helping implement the DMD part." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 132, + 160, + 197, + 173 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 160, + 197, + 173 + ], + "spans": [ + { + "bbox": [ + 132, + 160, + 197, + 173 + ], + "type": "text", + "content": "References" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 138, + 188, + 480, + 665 + ], + "type": "list", + "angle": 0, + "index": 21, + "blocks": [ + { + "bbox": [ + 138, + 188, + 480, + 209 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 188, + 480, + 209 + ], + "spans": [ + { + "bbox": [ + 138, + 188, + 480, + 209 + ], + "type": "text", + "content": "1. Aesthetic predictor (2023), https://github.com/christophschuhmann/improved-aesthetic-predictor7" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 138, + 210, + 480, + 232 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 210, + 480, + 232 + ], + "spans": [ + { + "bbox": [ + 138, + 210, + 480, + 232 + ], + "type": "text", + "content": "2. Bao, F., Nie, S., Xue, K., Cao, Y., Li, C., Su, H., Zhu, J.: All are worth words: A vit backbone for diffusion models. In: CVPR (2023) 4" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 138, + 233, + 480, + 265 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 233, + 480, + 265 + ], + "spans": [ + { + "bbox": [ + 138, + 233, + 480, + 265 + ], + "type": "text", + "content": "3. Chandra, A., Tünnermann, L., Lofstedt, T., Gratz, R.: Transformer-based deep learning for predicting protein properties in the life sciences. eLife Sciences Publications, Ltd (2023) 4" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 138, + 266, + 480, + 299 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 266, + 480, + 299 + ], + "spans": [ + { + "bbox": [ + 138, + 266, + 480, + 299 + ], + "type": "text", + "content": "4. Chen, J., Yu, J., Ge, C., Yao, L., Xie, E., Wu, Y., Wang, Z., Kwok, J., Luo, P., Lu, H., Li, Z.: PixArt-α: Fast training of diffusion transformer for photorealistic text-to-image synthesis. In: ICLR (2024) 1, 4, 6, 8, 11, 12" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 138, + 300, + 480, + 331 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 300, + 480, + 331 + ], + "spans": [ + { + "bbox": [ + 138, + 300, + 480, + 331 + ], + "type": "text", + "content": "5. Chen, L., Li, J., Dong, X., Zhang, P., He, C., Wang, J., Zhao, F., Lin, D.: ShareGPT4V: Improving large multi-modal models with better captions. In: arXiv (2023) 3, 8" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 138, + 332, + 480, + 365 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 332, + 480, + 365 + ], + "spans": [ + { + "bbox": [ + 138, + 332, + 480, + 365 + ], + "type": "text", + "content": "6. Chen, S., Xu, M., Ren, J., Cong, Y., He, S., Xie, Y., Sinha, A., Luo, P., Xiang, T., Perez-Rua, J.M.: Gentron: Delving deep into diffusion transformers for image and video generation. arXiv preprint arXiv:2312.04557 (2023) 3, 4" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 138, + 366, + 480, + 388 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 366, + 480, + 388 + ], + "spans": [ + { + "bbox": [ + 138, + 366, + 480, + 388 + ], + "type": "text", + "content": "7. Chen, T., Cheng, Y., Gan, Z., Yuan, L., Zhang, L., Wang, Z.: Chasing sparsity in vision transformers: An end-to-end exploration. In: NeurIPS (2021) 7" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 138, + 388, + 480, + 419 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 388, + 480, + 419 + ], + "spans": [ + { + "bbox": [ + 138, + 388, + 480, + 419 + ], + "type": "text", + "content": "8. Chen, X., Liu, Z., Tang, H., Yi, L., Zhao, H., Han, S.: SparseViT: Revisiting activation sparsity for efficient high-resolution vision transformer. In: CVPR (2023) 7" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 138, + 421, + 480, + 453 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 421, + 480, + 453 + ], + "spans": [ + { + "bbox": [ + 138, + 421, + 480, + 453 + ], + "type": "text", + "content": "9. Choromanski, K., Likhosherstov, V., Dohan, D., Song, X., Gane, A., Sarlos, T., Hawkins, P., Davis, J., Mohiuddin, A., Kaiser, L., et al.: Rethinking attention with performers. In: ICLR (2021) 7" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 138, + 454, + 480, + 487 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 454, + 480, + 487 + ], + "spans": [ + { + "bbox": [ + 138, + 454, + 480, + 487 + ], + "type": "text", + "content": "0. Chung, H.W., Hou, L., Longpre, S., Zoph, B., Tay, Y., Fedus, W., Li, Y., Wang, X., Dehghani, M., Brahma, S., et al.: Scaling instruction-finetuned language models. In: arXiv (2022) 3, 11" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 138, + 487, + 480, + 510 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 487, + 480, + 510 + ], + "spans": [ + { + "bbox": [ + 138, + 487, + 480, + 510 + ], + "type": "text", + "content": " 1. Du, R., Chang, D., Hospedales, T., Song, Y.Z., Ma, Z.: Demofusion: Democratising high-resolution image generation with no $$#. In: CVPR (2024) 12" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 138, + 510, + 480, + 532 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 510, + 480, + 532 + ], + "spans": [ + { + "bbox": [ + 138, + 510, + 480, + 532 + ], + "type": "text", + "content": "2. Gao, S., Zhou, P., Cheng, M.M., Yan, S.: Masked diffusion transformer is a strong image synthesizer. In: ICCV (2023) 4" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 138, + 533, + 480, + 554 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 533, + 480, + 554 + ], + "spans": [ + { + "bbox": [ + 138, + 533, + 480, + 554 + ], + "type": "text", + "content": "3. Ge, C., Ding, X., Tong, Z., Yuan, L., Wang, J., Song, Y., Luo, P.: Advancing vision transformers with group-mix attention. In: arXiv (2023) 7" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 138, + 555, + 480, + 576 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 555, + 480, + 576 + ], + "spans": [ + { + "bbox": [ + 138, + 555, + 480, + 576 + ], + "type": "text", + "content": "4. von Glehn, I., Spencer, J.S., P fauna, D.: A self-attention ansatz for ab-initio quantum chemistry. In: ICLR (2023) 4" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 138, + 577, + 480, + 598 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 577, + 480, + 598 + ], + "spans": [ + { + "bbox": [ + 138, + 577, + 480, + 598 + ], + "type": "text", + "content": "5. Hatamizadeh, A., Song, J., Liu, G., Kautz, J., Vahdat, A.: Diffit: Diffusion vision transformers for image generation. In: arXiv (2023) 4" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 138, + 599, + 480, + 632 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 599, + 480, + 632 + ], + "spans": [ + { + "bbox": [ + 138, + 599, + 480, + 632 + ], + "type": "text", + "content": "6. He, Y., Yang, S., Chen, H., Cun, X., Xia, M., Zhang, Y., Wang, X., He, R., Chen, Q., Shan, Y.: ScaleCrafter: Tuning-free higher-resolution visual generation with diffusion models. In: ICLR (2024) 12" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 138, + 632, + 480, + 665 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 632, + 480, + 665 + ], + "spans": [ + { + "bbox": [ + 138, + 632, + 480, + 665 + ], + "type": "text", + "content": "7. Heusel, M., Ramsauer, H., Unterthiner, T., Nessler, B., Hochreiter, S.: GANs trained by a two time-scale update rule converge to a local nash equilibrium. In: NeurIPS (2017) 8" + } + ] + } + ], + "index": 20 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 172, + 91, + 447, + 102 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 172, + 91, + 447, + 102 + ], + "spans": [ + { + "bbox": [ + 172, + 91, + 447, + 102 + ], + "type": "text", + "content": "PIXART- " + }, + { + "bbox": [ + 172, + 91, + 447, + 102 + ], + "type": "inline_equation", + "content": "\\boldsymbol{\\Sigma}" + }, + { + "bbox": [ + 172, + 91, + 447, + 102 + ], + "type": "text", + "content": " : Weak-to-Strong Training of DiT for 4K T2I Generation" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 470, + 92, + 480, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 470, + 92, + 480, + 100 + ], + "spans": [ + { + "bbox": [ + 470, + 92, + 480, + 100 + ], + "type": "text", + "content": "15" + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 14 + }, + { + "para_blocks": [ + { + "bbox": [ + 132, + 116, + 480, + 663 + ], + "type": "list", + "angle": 0, + "index": 27, + "blocks": [ + { + "bbox": [ + 133, + 116, + 480, + 138 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 116, + 480, + 138 + ], + "spans": [ + { + "bbox": [ + 133, + 116, + 480, + 138 + ], + "type": "text", + "content": "18. Kang, M., Zhu, J.Y., Zhang, R., Park, J., Shechtman, E., Paris, S., Park, T.: Scaling up gans for text-to-image synthesis. In: CVPR (2023) 6" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 133, + 139, + 480, + 170 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 139, + 480, + 170 + ], + "spans": [ + { + "bbox": [ + 133, + 139, + 480, + 170 + ], + "type": "text", + "content": "19. Li, D., Kamko, A., Akhgari, E., Sabet, A., Xu, L., Doshi, S.: Playground v2.5: Three insights towards enhancing aesthetic quality in text-to-image generation. In: arXiv (2024) 6" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 132, + 171, + 480, + 192 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 171, + 480, + 192 + ], + "spans": [ + { + "bbox": [ + 132, + 171, + 480, + 192 + ], + "type": "text", + "content": "20. Li, D., Kamko, A., Sabet, A., Akhgari, E., Xu, L., Doshi, S.: Playground v2, https://huggingface.co/playgroundai/playground-v2-1024px-aesthetic 12" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 132, + 193, + 480, + 213 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 193, + 480, + 213 + ], + "spans": [ + { + "bbox": [ + 132, + 193, + 480, + 213 + ], + "type": "text", + "content": "21. Lin, T.Y., Maire, M., Belongie, S., Hays, J., Perona, P., Ramanan, D., Dólar, P., Zitnick, C.L.: Microsoft COCO: Common objects in context. In: ECCV (2014) 8" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 132, + 214, + 472, + 224 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 214, + 472, + 224 + ], + "spans": [ + { + "bbox": [ + 132, + 214, + 472, + 224 + ], + "type": "text", + "content": "22. Liu, H., Li, C., Wu, Q., Lee, Y.J.: Visual instruction tuning. In: arXiv (2023) 3" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 132, + 225, + 480, + 255 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 225, + 480, + 255 + ], + "spans": [ + { + "bbox": [ + 132, + 225, + 480, + 255 + ], + "type": "text", + "content": "23. Liu, Z., Lin, Y., Cao, Y., Hu, H., Wei, Y., Zhang, Z., Lin, S., Guo, B.: Swin transformer: Hierarchical vision transformer using shifted windows. In: ICCV (2021) 4, 7" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 132, + 257, + 480, + 277 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 257, + 480, + 277 + ], + "spans": [ + { + "bbox": [ + 132, + 257, + 480, + 277 + ], + "type": "text", + "content": "24. Loshchilov, I., Hutter, F.: Decoupled weight decay regularization. In: arXiv (2017) 11" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 132, + 277, + 480, + 300 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 277, + 480, + 300 + ], + "spans": [ + { + "bbox": [ + 132, + 277, + 480, + 300 + ], + "type": "text", + "content": "25. Lu, J., Yao, J., Zhang, J., Zhu, X., Xu, H., Gao, W., Xu, C., Xiang, T., Zhang, L.: Soft: Softmax-free transformer with linear complexity. In: NeurIPS (2021) 7" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 132, + 300, + 480, + 320 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 300, + 480, + 320 + ], + "spans": [ + { + "bbox": [ + 132, + 300, + 480, + 320 + ], + "type": "text", + "content": "26. Lu, Z., Wang, Z., Huang, D., Wu, C., Liu, X., Ouyang, W., Bai, L.: Fit: Flexible vision transformer for diffusion model. In: arXiv (2024) 4" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 132, + 322, + 480, + 342 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 322, + 480, + 342 + ], + "spans": [ + { + "bbox": [ + 132, + 322, + 480, + 342 + ], + "type": "text", + "content": "27. Luo, Y., Ren, X., Zheng, Z., Jiang, Z., Jiang, X., You, Y.: CAME: Confidence-guided adaptive memory efficient optimization. In: ACL (2023) 11" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 132, + 343, + 480, + 374 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 343, + 480, + 374 + ], + "spans": [ + { + "bbox": [ + 132, + 343, + 480, + 374 + ], + "type": "text", + "content": "28. Ma, N., Goldstein, M., Albergo, M.S., Boffi, N.M., Vanden-Eijnden, E., Xie, S.: SiT: Exploring flow and diffusion-based generative models with scalable interpolant transformers. In: arXiv (2024) 4" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 132, + 375, + 425, + 385 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 375, + 425, + 385 + ], + "spans": [ + { + "bbox": [ + 132, + 375, + 425, + 385 + ], + "type": "text", + "content": "29. Midjourney: Midjourney (2023), https://www.midjourney.com 1, 4" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 132, + 386, + 389, + 396 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 386, + 389, + 396 + ], + "spans": [ + { + "bbox": [ + 132, + 386, + 389, + 396 + ], + "type": "text", + "content": "30. OpenAI: Dalle-2 (2023), https://openai.com/dall-e-2 6" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 132, + 396, + 409, + 407 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 396, + 409, + 407 + ], + "spans": [ + { + "bbox": [ + 132, + 396, + 409, + 407 + ], + "type": "text", + "content": "31. OpenAI: Dalle-3 (2023), https://openai.com/dall-e-3 1, 4, 8" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 132, + 407, + 388, + 418 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 407, + 388, + 418 + ], + "spans": [ + { + "bbox": [ + 132, + 407, + 388, + 418 + ], + "type": "text", + "content": "32. OpenAI: Gpt-4v(ison) system card. In: OpenAI (2023) 13" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 132, + 418, + 368, + 427 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 418, + 368, + 427 + ], + "spans": [ + { + "bbox": [ + 132, + 418, + 368, + 427 + ], + "type": "text", + "content": "33. OpenAI: Sora (2024), https://openai.com/sora 3, 5" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 132, + 428, + 480, + 447 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 428, + 480, + 447 + ], + "spans": [ + { + "bbox": [ + 132, + 428, + 480, + 447 + ], + "type": "text", + "content": "34. Peebles, W., Xie, S.: Scalable diffusion models with transformers. In: ICCV (2023) 4" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 132, + 449, + 480, + 480 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 449, + 480, + 480 + ], + "spans": [ + { + "bbox": [ + 132, + 449, + 480, + 480 + ], + "type": "text", + "content": "35. Pernias, P., Rampas, D., Richter, M.L., Pal, C., Aubreville, M.: Würstchen: An efficient architecture for large-scale text-to-image diffusion models. In: ICLR (2023) 4, 6, 12" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 132, + 481, + 480, + 514 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 481, + 480, + 514 + ], + "spans": [ + { + "bbox": [ + 132, + 481, + 480, + 514 + ], + "type": "text", + "content": "36. Podell, D., English, Z., Lacey, K., Blattmann, A., Dockhorn, T., Müller, J., Penna, J., Rombach, R.: SDXL: Improving latent diffusion models for high-resolution image synthesis. In: arXiv (2023) 1, 3, 4, 6, 11, 12" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 132, + 514, + 480, + 535 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 514, + 480, + 535 + ], + "spans": [ + { + "bbox": [ + 132, + 514, + 480, + 535 + ], + "type": "text", + "content": "37. Radford, A., Narasimhan, K., Salimans, T., Sutskever, I., et al.: Improving language understanding by generative pre-training. OpenAI blog (2018) 4" + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 132, + 536, + 480, + 556 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 536, + 480, + 556 + ], + "spans": [ + { + "bbox": [ + 132, + 536, + 480, + 556 + ], + "type": "text", + "content": "38. Radford, A., Wu, J., Child, R., Luan, D., Amodei, D., Sutskever, I., et al.: Language models are unsupervised multitask learners. OpenAI blog (2019) 4" + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 132, + 557, + 480, + 578 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 557, + 480, + 578 + ], + "spans": [ + { + "bbox": [ + 132, + 557, + 480, + 578 + ], + "type": "text", + "content": "39. Rombach, R., Blattmann, A., Lorenz, D., Esser, P., Ommer, B.: High-resolution image synthesis with latent diffusion models. In: CVPR (2022) 6, 11, 12" + } + ] + } + ], + "index": 23 + }, + { + "bbox": [ + 132, + 579, + 480, + 620 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 579, + 480, + 620 + ], + "spans": [ + { + "bbox": [ + 132, + 579, + 480, + 620 + ], + "type": "text", + "content": "40. Sahara, C., Chan, W., Saxena, S., Li, L., Whang, J., Denton, E.L., Ghasemipour, K., Gontijo Lopes, R., Karagol Ayan, B., Salimans, T., et al.: Photorealistic text-to-image diffusion models with deep language understanding. In: NeurIPS (2022) 6, 11" + } + ] + } + ], + "index": 24 + }, + { + "bbox": [ + 132, + 621, + 480, + 643 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 621, + 480, + 643 + ], + "spans": [ + { + "bbox": [ + 132, + 621, + 480, + 643 + ], + "type": "text", + "content": "41. Sauer, A., Lorenz, D., Blattmann, A., Rombach, R.: Adversarial diffusion distillation. In: arXiv (2023) 12" + } + ] + } + ], + "index": 25 + }, + { + "bbox": [ + 132, + 643, + 480, + 663 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 643, + 480, + 663 + ], + "spans": [ + { + "bbox": [ + 132, + 643, + 480, + 663 + ], + "type": "text", + "content": "42. Stability.AI: Stable diffusion 3 (2024), https://stability.ai/news/stable-diffusion-3 3" + } + ] + } + ], + "index": 26 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "spans": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "type": "text", + "content": "16" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 166, + 91, + 222, + 100 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 166, + 91, + 222, + 100 + ], + "spans": [ + { + "bbox": [ + 166, + 91, + 222, + 100 + ], + "type": "text", + "content": "J. Chen et al." + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 15 + }, + { + "para_blocks": [ + { + "bbox": [ + 132, + 116, + 480, + 456 + ], + "type": "list", + "angle": 0, + "index": 14, + "blocks": [ + { + "bbox": [ + 132, + 116, + 480, + 148 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 116, + 480, + 148 + ], + "spans": [ + { + "bbox": [ + 132, + 116, + 480, + 148 + ], + "type": "text", + "content": "43. Touvron, H., Cord, M., Douze, M., Massa, F., Sablayrolles, A., Jégou, H.: Training data-efficient image transformers & distillation through attention. In: ICML (2021) 4" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 132, + 149, + 480, + 171 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 149, + 480, + 171 + ], + "spans": [ + { + "bbox": [ + 132, + 149, + 480, + 171 + ], + "type": "text", + "content": "44. Wang, S., Li, B.Z., Khabsa, M., Fang, H., Ma, H.: Linformer: Self-attention with linear complexity. In: arXiv (2020) 7" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 132, + 172, + 480, + 204 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 172, + 480, + 204 + ], + "spans": [ + { + "bbox": [ + 132, + 172, + 480, + 204 + ], + "type": "text", + "content": "45. Wang, W., Xie, E., Li, X., Fan, D.P., Song, K., Liang, D., Lu, T., Luo, P., Shao, L.: Pyramid vision transformer: A versatile backbone for dense prediction without convolutions. In: ICCV (2021) 7" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 132, + 205, + 480, + 237 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 205, + 480, + 237 + ], + "spans": [ + { + "bbox": [ + 132, + 205, + 480, + 237 + ], + "type": "text", + "content": "46. Wang, W., Xie, E., Li, X., Fan, D.P., Song, K., Liang, D., Lu, T., Luo, P., Shao, L.: Pvt v2: Improved baselines with pyramid vision transformer. Computational Visual Media (2022) 7, 9" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 132, + 237, + 480, + 259 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 237, + 480, + 259 + ], + "spans": [ + { + "bbox": [ + 132, + 237, + 480, + 259 + ], + "type": "text", + "content": "47. Xia, Z., Pan, X., Song, S., Li, L.E., Huang, G.: Vision transformer with deformable attention. In: CVPR (2022) 7" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 132, + 259, + 480, + 293 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 259, + 480, + 293 + ], + "spans": [ + { + "bbox": [ + 132, + 259, + 480, + 293 + ], + "type": "text", + "content": "48. Xie, E., Wang, W., Yu, Z., Anandkumar, A., Alvarez, J.M., Luo, P.: SegFormer: Simple and efficient design for semantic segmentation with transformers. In: NeurIPS (2021) 7" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 132, + 293, + 480, + 325 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 293, + 480, + 325 + ], + "spans": [ + { + "bbox": [ + 132, + 293, + 480, + 325 + ], + "type": "text", + "content": "49. Xie, E., Yao, L., Shi, H., Liu, Z., Zhou, D., Liu, Z., Li, J., Li, Z.: DiffFit: Unlocking transferability of large diffusion models via simple parameter-efficient fine-tuning. In: ICCV (2023) 11" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 132, + 325, + 480, + 347 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 325, + 480, + 347 + ], + "spans": [ + { + "bbox": [ + 132, + 325, + 480, + 347 + ], + "type": "text", + "content": "50. Xue, Z., Song, G., Guo, Q., Liu, B., Zong, Z., Liu, Y., Luo, P.: Raphael: Text-to-image generation via large mixture of diffusion paths. In: NeurIPS (2023) 12" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 132, + 347, + 480, + 380 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 347, + 480, + 380 + ], + "spans": [ + { + "bbox": [ + 132, + 347, + 480, + 380 + ], + "type": "text", + "content": "51. Yuan, L., Chen, Y., Wang, T., Yu, W., Shi, Y., Jiang, Z.H., Tay, F.E., Feng, J., Yan, S.: Tokens-to-token ViT: Training vision transformers from scratch onImagenet. In: ICCV (2021) 4" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 132, + 380, + 480, + 402 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 380, + 480, + 402 + ], + "spans": [ + { + "bbox": [ + 132, + 380, + 480, + 402 + ], + "type": "text", + "content": "52. Zheng, H., Nie, W., Vahdat, A., Anandkumar, A.: Fast training of diffusion models with masked transformers. In: arXiv (2023) 4" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 132, + 402, + 480, + 435 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 402, + 480, + 435 + ], + "spans": [ + { + "bbox": [ + 132, + 402, + 480, + 435 + ], + "type": "text", + "content": "53. Zheng, S., Lu, J., Zhao, H., Zhu, X., Luo, Z., Wang, Y., Fu, Y., Feng, J., Xiang, T., Torr, P.H., et al.: Rethinking semantic segmentation from a sequence-to-sequence perspective with transformers. In: CVPR (2021) 4" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 132, + 435, + 480, + 456 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 435, + 480, + 456 + ], + "spans": [ + { + "bbox": [ + 132, + 435, + 480, + 456 + ], + "type": "text", + "content": "54. Zhu, X., Su, W., Lu, L., Li, B., Wang, X., Dai, J.: Deformable DETR: Deformable transformers for end-to-end object detection. ICLR (2021) 7" + } + ] + } + ], + "index": 13 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 173, + 91, + 447, + 102 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 173, + 91, + 447, + 102 + ], + "spans": [ + { + "bbox": [ + 173, + 91, + 447, + 102 + ], + "type": "text", + "content": "PixART- " + }, + { + "bbox": [ + 173, + 91, + 447, + 102 + ], + "type": "inline_equation", + "content": "\\boldsymbol{\\Sigma}" + }, + { + "bbox": [ + 173, + 91, + 447, + 102 + ], + "type": "text", + "content": " : Weak-to-Strong Training of DiT for 4K T2I Generation" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 470, + 92, + 480, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 470, + 92, + 480, + 100 + ], + "spans": [ + { + "bbox": [ + 470, + 92, + 480, + 100 + ], + "type": "text", + "content": "17" + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 16 + } + ], + "_backend": "vlm", + "_version_name": "2.6.4" +} \ No newline at end of file